anv/radv: Resolving 'GetInstanceProcAddr' should not require a valid instance
[mesa.git] / src / intel / vulkan / anv_device.c
1 /*
2 * Copyright © 2015 Intel Corporation
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 */
23
24 #include <assert.h>
25 #include <stdbool.h>
26 #include <string.h>
27 #include <sys/mman.h>
28 #include <sys/sysinfo.h>
29 #include <unistd.h>
30 #include <fcntl.h>
31 #include <xf86drm.h>
32 #include "drm-uapi/drm_fourcc.h"
33
34 #include "anv_private.h"
35 #include "util/debug.h"
36 #include "util/build_id.h"
37 #include "util/disk_cache.h"
38 #include "util/mesa-sha1.h"
39 #include "util/os_file.h"
40 #include "util/u_atomic.h"
41 #include "util/u_string.h"
42 #include "util/xmlpool.h"
43 #include "git_sha1.h"
44 #include "vk_util.h"
45 #include "common/gen_aux_map.h"
46 #include "common/gen_defines.h"
47 #include "compiler/glsl_types.h"
48
49 #include "genxml/gen7_pack.h"
50
51 static const char anv_dri_options_xml[] =
52 DRI_CONF_BEGIN
53 DRI_CONF_SECTION_PERFORMANCE
54 DRI_CONF_VK_X11_OVERRIDE_MIN_IMAGE_COUNT(0)
55 DRI_CONF_VK_X11_STRICT_IMAGE_COUNT("false")
56 DRI_CONF_SECTION_END
57
58 DRI_CONF_SECTION_DEBUG
59 DRI_CONF_ALWAYS_FLUSH_CACHE("false")
60 DRI_CONF_VK_WSI_FORCE_BGRA8_UNORM_FIRST("false")
61 DRI_CONF_SECTION_END
62 DRI_CONF_END;
63
64 /* This is probably far to big but it reflects the max size used for messages
65 * in OpenGLs KHR_debug.
66 */
67 #define MAX_DEBUG_MESSAGE_LENGTH 4096
68
69 static void
70 compiler_debug_log(void *data, const char *fmt, ...)
71 {
72 char str[MAX_DEBUG_MESSAGE_LENGTH];
73 struct anv_device *device = (struct anv_device *)data;
74 struct anv_instance *instance = device->physical->instance;
75
76 if (list_is_empty(&instance->debug_report_callbacks.callbacks))
77 return;
78
79 va_list args;
80 va_start(args, fmt);
81 (void) vsnprintf(str, MAX_DEBUG_MESSAGE_LENGTH, fmt, args);
82 va_end(args);
83
84 vk_debug_report(&instance->debug_report_callbacks,
85 VK_DEBUG_REPORT_DEBUG_BIT_EXT,
86 VK_DEBUG_REPORT_OBJECT_TYPE_UNKNOWN_EXT,
87 0, 0, 0, "anv", str);
88 }
89
90 static void
91 compiler_perf_log(void *data, const char *fmt, ...)
92 {
93 va_list args;
94 va_start(args, fmt);
95
96 if (unlikely(INTEL_DEBUG & DEBUG_PERF))
97 intel_logd_v(fmt, args);
98
99 va_end(args);
100 }
101
102 static uint64_t
103 anv_compute_heap_size(int fd, uint64_t gtt_size)
104 {
105 /* Query the total ram from the system */
106 struct sysinfo info;
107 sysinfo(&info);
108
109 uint64_t total_ram = (uint64_t)info.totalram * (uint64_t)info.mem_unit;
110
111 /* We don't want to burn too much ram with the GPU. If the user has 4GiB
112 * or less, we use at most half. If they have more than 4GiB, we use 3/4.
113 */
114 uint64_t available_ram;
115 if (total_ram <= 4ull * 1024ull * 1024ull * 1024ull)
116 available_ram = total_ram / 2;
117 else
118 available_ram = total_ram * 3 / 4;
119
120 /* We also want to leave some padding for things we allocate in the driver,
121 * so don't go over 3/4 of the GTT either.
122 */
123 uint64_t available_gtt = gtt_size * 3 / 4;
124
125 return MIN2(available_ram, available_gtt);
126 }
127
128 static VkResult
129 anv_physical_device_init_heaps(struct anv_physical_device *device, int fd)
130 {
131 if (anv_gem_get_context_param(fd, 0, I915_CONTEXT_PARAM_GTT_SIZE,
132 &device->gtt_size) == -1) {
133 /* If, for whatever reason, we can't actually get the GTT size from the
134 * kernel (too old?) fall back to the aperture size.
135 */
136 anv_perf_warn(NULL, NULL,
137 "Failed to get I915_CONTEXT_PARAM_GTT_SIZE: %m");
138
139 if (anv_gem_get_aperture(fd, &device->gtt_size) == -1) {
140 return vk_errorfi(device->instance, NULL,
141 VK_ERROR_INITIALIZATION_FAILED,
142 "failed to get aperture size: %m");
143 }
144 }
145
146 /* We only allow 48-bit addresses with softpin because knowing the actual
147 * address is required for the vertex cache flush workaround.
148 */
149 device->supports_48bit_addresses = (device->info.gen >= 8) &&
150 device->has_softpin &&
151 device->gtt_size > (4ULL << 30 /* GiB */);
152
153 uint64_t heap_size = anv_compute_heap_size(fd, device->gtt_size);
154
155 if (heap_size > (2ull << 30) && !device->supports_48bit_addresses) {
156 /* When running with an overridden PCI ID, we may get a GTT size from
157 * the kernel that is greater than 2 GiB but the execbuf check for 48bit
158 * address support can still fail. Just clamp the address space size to
159 * 2 GiB if we don't have 48-bit support.
160 */
161 intel_logw("%s:%d: The kernel reported a GTT size larger than 2 GiB but "
162 "not support for 48-bit addresses",
163 __FILE__, __LINE__);
164 heap_size = 2ull << 30;
165 }
166
167 device->memory.heap_count = 1;
168 device->memory.heaps[0] = (struct anv_memory_heap) {
169 .size = heap_size,
170 .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
171 };
172
173 uint32_t type_count = 0;
174 for (uint32_t heap = 0; heap < device->memory.heap_count; heap++) {
175 if (device->info.has_llc) {
176 /* Big core GPUs share LLC with the CPU and thus one memory type can be
177 * both cached and coherent at the same time.
178 */
179 device->memory.types[type_count++] = (struct anv_memory_type) {
180 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
181 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
182 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
183 VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
184 .heapIndex = heap,
185 };
186 } else {
187 /* The spec requires that we expose a host-visible, coherent memory
188 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
189 * to give the application a choice between cached, but not coherent and
190 * coherent but uncached (WC though).
191 */
192 device->memory.types[type_count++] = (struct anv_memory_type) {
193 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
194 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
195 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
196 .heapIndex = heap,
197 };
198 device->memory.types[type_count++] = (struct anv_memory_type) {
199 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
200 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
201 VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
202 .heapIndex = heap,
203 };
204 }
205 }
206 device->memory.type_count = type_count;
207
208 return VK_SUCCESS;
209 }
210
211 static VkResult
212 anv_physical_device_init_uuids(struct anv_physical_device *device)
213 {
214 const struct build_id_note *note =
215 build_id_find_nhdr_for_addr(anv_physical_device_init_uuids);
216 if (!note) {
217 return vk_errorfi(device->instance, NULL,
218 VK_ERROR_INITIALIZATION_FAILED,
219 "Failed to find build-id");
220 }
221
222 unsigned build_id_len = build_id_length(note);
223 if (build_id_len < 20) {
224 return vk_errorfi(device->instance, NULL,
225 VK_ERROR_INITIALIZATION_FAILED,
226 "build-id too short. It needs to be a SHA");
227 }
228
229 memcpy(device->driver_build_sha1, build_id_data(note), 20);
230
231 struct mesa_sha1 sha1_ctx;
232 uint8_t sha1[20];
233 STATIC_ASSERT(VK_UUID_SIZE <= sizeof(sha1));
234
235 /* The pipeline cache UUID is used for determining when a pipeline cache is
236 * invalid. It needs both a driver build and the PCI ID of the device.
237 */
238 _mesa_sha1_init(&sha1_ctx);
239 _mesa_sha1_update(&sha1_ctx, build_id_data(note), build_id_len);
240 _mesa_sha1_update(&sha1_ctx, &device->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 device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
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_free(&device->instance->alloc, device);
599 }
600
601 static void *
602 default_alloc_func(void *pUserData, size_t size, size_t align,
603 VkSystemAllocationScope allocationScope)
604 {
605 return malloc(size);
606 }
607
608 static void *
609 default_realloc_func(void *pUserData, void *pOriginal, size_t size,
610 size_t align, VkSystemAllocationScope allocationScope)
611 {
612 return realloc(pOriginal, size);
613 }
614
615 static void
616 default_free_func(void *pUserData, void *pMemory)
617 {
618 free(pMemory);
619 }
620
621 static const VkAllocationCallbacks default_alloc = {
622 .pUserData = NULL,
623 .pfnAllocation = default_alloc_func,
624 .pfnReallocation = default_realloc_func,
625 .pfnFree = default_free_func,
626 };
627
628 VkResult anv_EnumerateInstanceExtensionProperties(
629 const char* pLayerName,
630 uint32_t* pPropertyCount,
631 VkExtensionProperties* pProperties)
632 {
633 VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
634
635 for (int i = 0; i < ANV_INSTANCE_EXTENSION_COUNT; i++) {
636 if (anv_instance_extensions_supported.extensions[i]) {
637 vk_outarray_append(&out, prop) {
638 *prop = anv_instance_extensions[i];
639 }
640 }
641 }
642
643 return vk_outarray_status(&out);
644 }
645
646 VkResult anv_CreateInstance(
647 const VkInstanceCreateInfo* pCreateInfo,
648 const VkAllocationCallbacks* pAllocator,
649 VkInstance* pInstance)
650 {
651 struct anv_instance *instance;
652 VkResult result;
653
654 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO);
655
656 struct anv_instance_extension_table enabled_extensions = {};
657 for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
658 int idx;
659 for (idx = 0; idx < ANV_INSTANCE_EXTENSION_COUNT; idx++) {
660 if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
661 anv_instance_extensions[idx].extensionName) == 0)
662 break;
663 }
664
665 if (idx >= ANV_INSTANCE_EXTENSION_COUNT)
666 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
667
668 if (!anv_instance_extensions_supported.extensions[idx])
669 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
670
671 enabled_extensions.extensions[idx] = true;
672 }
673
674 instance = vk_alloc2(&default_alloc, pAllocator, sizeof(*instance), 8,
675 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
676 if (!instance)
677 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
678
679 instance->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
680
681 if (pAllocator)
682 instance->alloc = *pAllocator;
683 else
684 instance->alloc = default_alloc;
685
686 instance->app_info = (struct anv_app_info) { .api_version = 0 };
687 if (pCreateInfo->pApplicationInfo) {
688 const VkApplicationInfo *app = pCreateInfo->pApplicationInfo;
689
690 instance->app_info.app_name =
691 vk_strdup(&instance->alloc, app->pApplicationName,
692 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
693 instance->app_info.app_version = app->applicationVersion;
694
695 instance->app_info.engine_name =
696 vk_strdup(&instance->alloc, app->pEngineName,
697 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
698 instance->app_info.engine_version = app->engineVersion;
699
700 instance->app_info.api_version = app->apiVersion;
701 }
702
703 if (instance->app_info.api_version == 0)
704 instance->app_info.api_version = VK_API_VERSION_1_0;
705
706 instance->enabled_extensions = enabled_extensions;
707
708 for (unsigned i = 0; i < ARRAY_SIZE(instance->dispatch.entrypoints); i++) {
709 /* Vulkan requires that entrypoints for extensions which have not been
710 * enabled must not be advertised.
711 */
712 if (!anv_instance_entrypoint_is_enabled(i, instance->app_info.api_version,
713 &instance->enabled_extensions)) {
714 instance->dispatch.entrypoints[i] = NULL;
715 } else {
716 instance->dispatch.entrypoints[i] =
717 anv_instance_dispatch_table.entrypoints[i];
718 }
719 }
720
721 for (unsigned i = 0; i < ARRAY_SIZE(instance->physical_device_dispatch.entrypoints); i++) {
722 /* Vulkan requires that entrypoints for extensions which have not been
723 * enabled must not be advertised.
724 */
725 if (!anv_physical_device_entrypoint_is_enabled(i, instance->app_info.api_version,
726 &instance->enabled_extensions)) {
727 instance->physical_device_dispatch.entrypoints[i] = NULL;
728 } else {
729 instance->physical_device_dispatch.entrypoints[i] =
730 anv_physical_device_dispatch_table.entrypoints[i];
731 }
732 }
733
734 for (unsigned i = 0; i < ARRAY_SIZE(instance->device_dispatch.entrypoints); i++) {
735 /* Vulkan requires that entrypoints for extensions which have not been
736 * enabled must not be advertised.
737 */
738 if (!anv_device_entrypoint_is_enabled(i, instance->app_info.api_version,
739 &instance->enabled_extensions, NULL)) {
740 instance->device_dispatch.entrypoints[i] = NULL;
741 } else {
742 instance->device_dispatch.entrypoints[i] =
743 anv_device_dispatch_table.entrypoints[i];
744 }
745 }
746
747 instance->physical_devices_enumerated = false;
748 list_inithead(&instance->physical_devices);
749
750 result = vk_debug_report_instance_init(&instance->debug_report_callbacks);
751 if (result != VK_SUCCESS) {
752 vk_free2(&default_alloc, pAllocator, instance);
753 return vk_error(result);
754 }
755
756 instance->pipeline_cache_enabled =
757 env_var_as_boolean("ANV_ENABLE_PIPELINE_CACHE", true);
758
759 glsl_type_singleton_init_or_ref();
760
761 VG(VALGRIND_CREATE_MEMPOOL(instance, 0, false));
762
763 driParseOptionInfo(&instance->available_dri_options, anv_dri_options_xml);
764 driParseConfigFiles(&instance->dri_options, &instance->available_dri_options,
765 0, "anv", NULL,
766 instance->app_info.engine_name,
767 instance->app_info.engine_version);
768
769 *pInstance = anv_instance_to_handle(instance);
770
771 return VK_SUCCESS;
772 }
773
774 void anv_DestroyInstance(
775 VkInstance _instance,
776 const VkAllocationCallbacks* pAllocator)
777 {
778 ANV_FROM_HANDLE(anv_instance, instance, _instance);
779
780 if (!instance)
781 return;
782
783 list_for_each_entry_safe(struct anv_physical_device, pdevice,
784 &instance->physical_devices, link)
785 anv_physical_device_destroy(pdevice);
786
787 vk_free(&instance->alloc, (char *)instance->app_info.app_name);
788 vk_free(&instance->alloc, (char *)instance->app_info.engine_name);
789
790 VG(VALGRIND_DESTROY_MEMPOOL(instance));
791
792 vk_debug_report_instance_destroy(&instance->debug_report_callbacks);
793
794 glsl_type_singleton_decref();
795
796 driDestroyOptionCache(&instance->dri_options);
797 driDestroyOptionInfo(&instance->available_dri_options);
798
799 vk_free(&instance->alloc, instance);
800 }
801
802 static VkResult
803 anv_enumerate_physical_devices(struct anv_instance *instance)
804 {
805 if (instance->physical_devices_enumerated)
806 return VK_SUCCESS;
807
808 instance->physical_devices_enumerated = true;
809
810 /* TODO: Check for more devices ? */
811 drmDevicePtr devices[8];
812 int max_devices;
813
814 max_devices = drmGetDevices2(0, devices, ARRAY_SIZE(devices));
815 if (max_devices < 1)
816 return VK_SUCCESS;
817
818 VkResult result = VK_SUCCESS;
819 for (unsigned i = 0; i < (unsigned)max_devices; i++) {
820 if (devices[i]->available_nodes & 1 << DRM_NODE_RENDER &&
821 devices[i]->bustype == DRM_BUS_PCI &&
822 devices[i]->deviceinfo.pci->vendor_id == 0x8086) {
823
824 struct anv_physical_device *pdevice;
825 result = anv_physical_device_try_create(instance, devices[i],
826 &pdevice);
827 /* Incompatible DRM device, skip. */
828 if (result == VK_ERROR_INCOMPATIBLE_DRIVER) {
829 result = VK_SUCCESS;
830 continue;
831 }
832
833 /* Error creating the physical device, report the error. */
834 if (result != VK_SUCCESS)
835 break;
836
837 list_addtail(&pdevice->link, &instance->physical_devices);
838 }
839 }
840 drmFreeDevices(devices, max_devices);
841
842 /* If we successfully enumerated any devices, call it success */
843 return result;
844 }
845
846 VkResult anv_EnumeratePhysicalDevices(
847 VkInstance _instance,
848 uint32_t* pPhysicalDeviceCount,
849 VkPhysicalDevice* pPhysicalDevices)
850 {
851 ANV_FROM_HANDLE(anv_instance, instance, _instance);
852 VK_OUTARRAY_MAKE(out, pPhysicalDevices, pPhysicalDeviceCount);
853
854 VkResult result = anv_enumerate_physical_devices(instance);
855 if (result != VK_SUCCESS)
856 return result;
857
858 list_for_each_entry(struct anv_physical_device, pdevice,
859 &instance->physical_devices, link) {
860 vk_outarray_append(&out, i) {
861 *i = anv_physical_device_to_handle(pdevice);
862 }
863 }
864
865 return vk_outarray_status(&out);
866 }
867
868 VkResult anv_EnumeratePhysicalDeviceGroups(
869 VkInstance _instance,
870 uint32_t* pPhysicalDeviceGroupCount,
871 VkPhysicalDeviceGroupProperties* pPhysicalDeviceGroupProperties)
872 {
873 ANV_FROM_HANDLE(anv_instance, instance, _instance);
874 VK_OUTARRAY_MAKE(out, pPhysicalDeviceGroupProperties,
875 pPhysicalDeviceGroupCount);
876
877 VkResult result = anv_enumerate_physical_devices(instance);
878 if (result != VK_SUCCESS)
879 return result;
880
881 list_for_each_entry(struct anv_physical_device, pdevice,
882 &instance->physical_devices, link) {
883 vk_outarray_append(&out, p) {
884 p->physicalDeviceCount = 1;
885 memset(p->physicalDevices, 0, sizeof(p->physicalDevices));
886 p->physicalDevices[0] = anv_physical_device_to_handle(pdevice);
887 p->subsetAllocation = false;
888
889 vk_foreach_struct(ext, p->pNext)
890 anv_debug_ignored_stype(ext->sType);
891 }
892 }
893
894 return vk_outarray_status(&out);
895 }
896
897 void anv_GetPhysicalDeviceFeatures(
898 VkPhysicalDevice physicalDevice,
899 VkPhysicalDeviceFeatures* pFeatures)
900 {
901 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
902
903 *pFeatures = (VkPhysicalDeviceFeatures) {
904 .robustBufferAccess = true,
905 .fullDrawIndexUint32 = true,
906 .imageCubeArray = true,
907 .independentBlend = true,
908 .geometryShader = true,
909 .tessellationShader = true,
910 .sampleRateShading = true,
911 .dualSrcBlend = true,
912 .logicOp = true,
913 .multiDrawIndirect = true,
914 .drawIndirectFirstInstance = true,
915 .depthClamp = true,
916 .depthBiasClamp = true,
917 .fillModeNonSolid = true,
918 .depthBounds = pdevice->info.gen >= 12,
919 .wideLines = true,
920 .largePoints = true,
921 .alphaToOne = true,
922 .multiViewport = true,
923 .samplerAnisotropy = true,
924 .textureCompressionETC2 = pdevice->info.gen >= 8 ||
925 pdevice->info.is_baytrail,
926 .textureCompressionASTC_LDR = pdevice->info.gen >= 9, /* FINISHME CHV */
927 .textureCompressionBC = true,
928 .occlusionQueryPrecise = true,
929 .pipelineStatisticsQuery = true,
930 .fragmentStoresAndAtomics = true,
931 .shaderTessellationAndGeometryPointSize = true,
932 .shaderImageGatherExtended = true,
933 .shaderStorageImageExtendedFormats = true,
934 .shaderStorageImageMultisample = false,
935 .shaderStorageImageReadWithoutFormat = false,
936 .shaderStorageImageWriteWithoutFormat = true,
937 .shaderUniformBufferArrayDynamicIndexing = true,
938 .shaderSampledImageArrayDynamicIndexing = true,
939 .shaderStorageBufferArrayDynamicIndexing = true,
940 .shaderStorageImageArrayDynamicIndexing = true,
941 .shaderClipDistance = true,
942 .shaderCullDistance = true,
943 .shaderFloat64 = pdevice->info.gen >= 8 &&
944 pdevice->info.has_64bit_float,
945 .shaderInt64 = pdevice->info.gen >= 8 &&
946 pdevice->info.has_64bit_int,
947 .shaderInt16 = pdevice->info.gen >= 8,
948 .shaderResourceMinLod = pdevice->info.gen >= 9,
949 .variableMultisampleRate = true,
950 .inheritedQueries = true,
951 };
952
953 /* We can't do image stores in vec4 shaders */
954 pFeatures->vertexPipelineStoresAndAtomics =
955 pdevice->compiler->scalar_stage[MESA_SHADER_VERTEX] &&
956 pdevice->compiler->scalar_stage[MESA_SHADER_GEOMETRY];
957
958 struct anv_app_info *app_info = &pdevice->instance->app_info;
959
960 /* The new DOOM and Wolfenstein games require depthBounds without
961 * checking for it. They seem to run fine without it so just claim it's
962 * there and accept the consequences.
963 */
964 if (app_info->engine_name && strcmp(app_info->engine_name, "idTech") == 0)
965 pFeatures->depthBounds = true;
966 }
967
968 static void
969 anv_get_physical_device_features_1_1(struct anv_physical_device *pdevice,
970 VkPhysicalDeviceVulkan11Features *f)
971 {
972 assert(f->sType == VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES);
973
974 f->storageBuffer16BitAccess = pdevice->info.gen >= 8;
975 f->uniformAndStorageBuffer16BitAccess = pdevice->info.gen >= 8;
976 f->storagePushConstant16 = pdevice->info.gen >= 8;
977 f->storageInputOutput16 = false;
978 f->multiview = true;
979 f->multiviewGeometryShader = true;
980 f->multiviewTessellationShader = true;
981 f->variablePointersStorageBuffer = true;
982 f->variablePointers = true;
983 f->protectedMemory = false;
984 f->samplerYcbcrConversion = true;
985 f->shaderDrawParameters = true;
986 }
987
988 static void
989 anv_get_physical_device_features_1_2(struct anv_physical_device *pdevice,
990 VkPhysicalDeviceVulkan12Features *f)
991 {
992 assert(f->sType == VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES);
993
994 f->samplerMirrorClampToEdge = true;
995 f->drawIndirectCount = true;
996 f->storageBuffer8BitAccess = pdevice->info.gen >= 8;
997 f->uniformAndStorageBuffer8BitAccess = pdevice->info.gen >= 8;
998 f->storagePushConstant8 = pdevice->info.gen >= 8;
999 f->shaderBufferInt64Atomics = pdevice->info.gen >= 9 &&
1000 pdevice->use_softpin;
1001 f->shaderSharedInt64Atomics = false;
1002 f->shaderFloat16 = pdevice->info.gen >= 8;
1003 f->shaderInt8 = pdevice->info.gen >= 8;
1004
1005 bool descIndexing = pdevice->has_a64_buffer_access &&
1006 pdevice->has_bindless_images;
1007 f->descriptorIndexing = descIndexing;
1008 f->shaderInputAttachmentArrayDynamicIndexing = false;
1009 f->shaderUniformTexelBufferArrayDynamicIndexing = descIndexing;
1010 f->shaderStorageTexelBufferArrayDynamicIndexing = descIndexing;
1011 f->shaderUniformBufferArrayNonUniformIndexing = false;
1012 f->shaderSampledImageArrayNonUniformIndexing = descIndexing;
1013 f->shaderStorageBufferArrayNonUniformIndexing = descIndexing;
1014 f->shaderStorageImageArrayNonUniformIndexing = descIndexing;
1015 f->shaderInputAttachmentArrayNonUniformIndexing = false;
1016 f->shaderUniformTexelBufferArrayNonUniformIndexing = descIndexing;
1017 f->shaderStorageTexelBufferArrayNonUniformIndexing = descIndexing;
1018 f->descriptorBindingUniformBufferUpdateAfterBind = false;
1019 f->descriptorBindingSampledImageUpdateAfterBind = descIndexing;
1020 f->descriptorBindingStorageImageUpdateAfterBind = descIndexing;
1021 f->descriptorBindingStorageBufferUpdateAfterBind = descIndexing;
1022 f->descriptorBindingUniformTexelBufferUpdateAfterBind = descIndexing;
1023 f->descriptorBindingStorageTexelBufferUpdateAfterBind = descIndexing;
1024 f->descriptorBindingUpdateUnusedWhilePending = descIndexing;
1025 f->descriptorBindingPartiallyBound = descIndexing;
1026 f->descriptorBindingVariableDescriptorCount = false;
1027 f->runtimeDescriptorArray = descIndexing;
1028
1029 f->samplerFilterMinmax = pdevice->info.gen >= 9;
1030 f->scalarBlockLayout = true;
1031 f->imagelessFramebuffer = true;
1032 f->uniformBufferStandardLayout = true;
1033 f->shaderSubgroupExtendedTypes = true;
1034 f->separateDepthStencilLayouts = true;
1035 f->hostQueryReset = true;
1036 f->timelineSemaphore = true;
1037 f->bufferDeviceAddress = pdevice->has_a64_buffer_access;
1038 f->bufferDeviceAddressCaptureReplay = pdevice->has_a64_buffer_access;
1039 f->bufferDeviceAddressMultiDevice = false;
1040 f->vulkanMemoryModel = true;
1041 f->vulkanMemoryModelDeviceScope = true;
1042 f->vulkanMemoryModelAvailabilityVisibilityChains = true;
1043 f->shaderOutputViewportIndex = true;
1044 f->shaderOutputLayer = true;
1045 f->subgroupBroadcastDynamicId = true;
1046 }
1047
1048 void anv_GetPhysicalDeviceFeatures2(
1049 VkPhysicalDevice physicalDevice,
1050 VkPhysicalDeviceFeatures2* pFeatures)
1051 {
1052 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
1053 anv_GetPhysicalDeviceFeatures(physicalDevice, &pFeatures->features);
1054
1055 VkPhysicalDeviceVulkan11Features core_1_1 = {
1056 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES,
1057 };
1058 anv_get_physical_device_features_1_1(pdevice, &core_1_1);
1059
1060 VkPhysicalDeviceVulkan12Features core_1_2 = {
1061 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES,
1062 };
1063 anv_get_physical_device_features_1_2(pdevice, &core_1_2);
1064
1065 #define CORE_FEATURE(major, minor, feature) \
1066 features->feature = core_##major##_##minor.feature
1067
1068
1069 vk_foreach_struct(ext, pFeatures->pNext) {
1070 switch (ext->sType) {
1071 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES_KHR: {
1072 VkPhysicalDevice8BitStorageFeaturesKHR *features =
1073 (VkPhysicalDevice8BitStorageFeaturesKHR *)ext;
1074 CORE_FEATURE(1, 2, storageBuffer8BitAccess);
1075 CORE_FEATURE(1, 2, uniformAndStorageBuffer8BitAccess);
1076 CORE_FEATURE(1, 2, storagePushConstant8);
1077 break;
1078 }
1079
1080 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES: {
1081 VkPhysicalDevice16BitStorageFeatures *features =
1082 (VkPhysicalDevice16BitStorageFeatures *)ext;
1083 CORE_FEATURE(1, 1, storageBuffer16BitAccess);
1084 CORE_FEATURE(1, 1, uniformAndStorageBuffer16BitAccess);
1085 CORE_FEATURE(1, 1, storagePushConstant16);
1086 CORE_FEATURE(1, 1, storageInputOutput16);
1087 break;
1088 }
1089
1090 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_EXT: {
1091 VkPhysicalDeviceBufferDeviceAddressFeaturesEXT *features = (void *)ext;
1092 features->bufferDeviceAddress = pdevice->has_a64_buffer_access;
1093 features->bufferDeviceAddressCaptureReplay = false;
1094 features->bufferDeviceAddressMultiDevice = false;
1095 break;
1096 }
1097
1098 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_KHR: {
1099 VkPhysicalDeviceBufferDeviceAddressFeaturesKHR *features = (void *)ext;
1100 CORE_FEATURE(1, 2, bufferDeviceAddress);
1101 CORE_FEATURE(1, 2, bufferDeviceAddressCaptureReplay);
1102 CORE_FEATURE(1, 2, bufferDeviceAddressMultiDevice);
1103 break;
1104 }
1105
1106 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COMPUTE_SHADER_DERIVATIVES_FEATURES_NV: {
1107 VkPhysicalDeviceComputeShaderDerivativesFeaturesNV *features =
1108 (VkPhysicalDeviceComputeShaderDerivativesFeaturesNV *)ext;
1109 features->computeDerivativeGroupQuads = true;
1110 features->computeDerivativeGroupLinear = true;
1111 break;
1112 }
1113
1114 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT: {
1115 VkPhysicalDeviceConditionalRenderingFeaturesEXT *features =
1116 (VkPhysicalDeviceConditionalRenderingFeaturesEXT*)ext;
1117 features->conditionalRendering = pdevice->info.gen >= 8 ||
1118 pdevice->info.is_haswell;
1119 features->inheritedConditionalRendering = pdevice->info.gen >= 8 ||
1120 pdevice->info.is_haswell;
1121 break;
1122 }
1123
1124 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_CLIP_ENABLE_FEATURES_EXT: {
1125 VkPhysicalDeviceDepthClipEnableFeaturesEXT *features =
1126 (VkPhysicalDeviceDepthClipEnableFeaturesEXT *)ext;
1127 features->depthClipEnable = true;
1128 break;
1129 }
1130
1131 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR: {
1132 VkPhysicalDeviceFloat16Int8FeaturesKHR *features = (void *)ext;
1133 CORE_FEATURE(1, 2, shaderFloat16);
1134 CORE_FEATURE(1, 2, shaderInt8);
1135 break;
1136 }
1137
1138 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FRAGMENT_SHADER_INTERLOCK_FEATURES_EXT: {
1139 VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT *features =
1140 (VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT *)ext;
1141 features->fragmentShaderSampleInterlock = pdevice->info.gen >= 9;
1142 features->fragmentShaderPixelInterlock = pdevice->info.gen >= 9;
1143 features->fragmentShaderShadingRateInterlock = false;
1144 break;
1145 }
1146
1147 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_QUERY_RESET_FEATURES_EXT: {
1148 VkPhysicalDeviceHostQueryResetFeaturesEXT *features =
1149 (VkPhysicalDeviceHostQueryResetFeaturesEXT *)ext;
1150 CORE_FEATURE(1, 2, hostQueryReset);
1151 break;
1152 }
1153
1154 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT: {
1155 VkPhysicalDeviceDescriptorIndexingFeaturesEXT *features =
1156 (VkPhysicalDeviceDescriptorIndexingFeaturesEXT *)ext;
1157 CORE_FEATURE(1, 2, shaderInputAttachmentArrayDynamicIndexing);
1158 CORE_FEATURE(1, 2, shaderUniformTexelBufferArrayDynamicIndexing);
1159 CORE_FEATURE(1, 2, shaderStorageTexelBufferArrayDynamicIndexing);
1160 CORE_FEATURE(1, 2, shaderUniformBufferArrayNonUniformIndexing);
1161 CORE_FEATURE(1, 2, shaderSampledImageArrayNonUniformIndexing);
1162 CORE_FEATURE(1, 2, shaderStorageBufferArrayNonUniformIndexing);
1163 CORE_FEATURE(1, 2, shaderStorageImageArrayNonUniformIndexing);
1164 CORE_FEATURE(1, 2, shaderInputAttachmentArrayNonUniformIndexing);
1165 CORE_FEATURE(1, 2, shaderUniformTexelBufferArrayNonUniformIndexing);
1166 CORE_FEATURE(1, 2, shaderStorageTexelBufferArrayNonUniformIndexing);
1167 CORE_FEATURE(1, 2, descriptorBindingUniformBufferUpdateAfterBind);
1168 CORE_FEATURE(1, 2, descriptorBindingSampledImageUpdateAfterBind);
1169 CORE_FEATURE(1, 2, descriptorBindingStorageImageUpdateAfterBind);
1170 CORE_FEATURE(1, 2, descriptorBindingStorageBufferUpdateAfterBind);
1171 CORE_FEATURE(1, 2, descriptorBindingUniformTexelBufferUpdateAfterBind);
1172 CORE_FEATURE(1, 2, descriptorBindingStorageTexelBufferUpdateAfterBind);
1173 CORE_FEATURE(1, 2, descriptorBindingUpdateUnusedWhilePending);
1174 CORE_FEATURE(1, 2, descriptorBindingPartiallyBound);
1175 CORE_FEATURE(1, 2, descriptorBindingVariableDescriptorCount);
1176 CORE_FEATURE(1, 2, runtimeDescriptorArray);
1177 break;
1178 }
1179
1180 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INDEX_TYPE_UINT8_FEATURES_EXT: {
1181 VkPhysicalDeviceIndexTypeUint8FeaturesEXT *features =
1182 (VkPhysicalDeviceIndexTypeUint8FeaturesEXT *)ext;
1183 features->indexTypeUint8 = true;
1184 break;
1185 }
1186
1187 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_FEATURES_EXT: {
1188 VkPhysicalDeviceInlineUniformBlockFeaturesEXT *features =
1189 (VkPhysicalDeviceInlineUniformBlockFeaturesEXT *)ext;
1190 features->inlineUniformBlock = true;
1191 features->descriptorBindingInlineUniformBlockUpdateAfterBind = true;
1192 break;
1193 }
1194
1195 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_FEATURES_EXT: {
1196 VkPhysicalDeviceLineRasterizationFeaturesEXT *features =
1197 (VkPhysicalDeviceLineRasterizationFeaturesEXT *)ext;
1198 features->rectangularLines = true;
1199 features->bresenhamLines = true;
1200 /* Support for Smooth lines with MSAA was removed on gen11. From the
1201 * BSpec section "Multisample ModesState" table for "AA Line Support
1202 * Requirements":
1203 *
1204 * GEN10:BUG:######## NUM_MULTISAMPLES == 1
1205 *
1206 * Fortunately, this isn't a case most people care about.
1207 */
1208 features->smoothLines = pdevice->info.gen < 10;
1209 features->stippledRectangularLines = false;
1210 features->stippledBresenhamLines = true;
1211 features->stippledSmoothLines = false;
1212 break;
1213 }
1214
1215 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES: {
1216 VkPhysicalDeviceMultiviewFeatures *features =
1217 (VkPhysicalDeviceMultiviewFeatures *)ext;
1218 CORE_FEATURE(1, 1, multiview);
1219 CORE_FEATURE(1, 1, multiviewGeometryShader);
1220 CORE_FEATURE(1, 1, multiviewTessellationShader);
1221 break;
1222 }
1223
1224 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGELESS_FRAMEBUFFER_FEATURES_KHR: {
1225 VkPhysicalDeviceImagelessFramebufferFeaturesKHR *features =
1226 (VkPhysicalDeviceImagelessFramebufferFeaturesKHR *)ext;
1227 CORE_FEATURE(1, 2, imagelessFramebuffer);
1228 break;
1229 }
1230
1231 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR: {
1232 VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR *features =
1233 (VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR *)ext;
1234 features->pipelineExecutableInfo = true;
1235 break;
1236 }
1237
1238 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES: {
1239 VkPhysicalDeviceProtectedMemoryFeatures *features = (void *)ext;
1240 CORE_FEATURE(1, 1, protectedMemory);
1241 break;
1242 }
1243
1244 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES: {
1245 VkPhysicalDeviceSamplerYcbcrConversionFeatures *features =
1246 (VkPhysicalDeviceSamplerYcbcrConversionFeatures *) ext;
1247 CORE_FEATURE(1, 1, samplerYcbcrConversion);
1248 break;
1249 }
1250
1251 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES_EXT: {
1252 VkPhysicalDeviceScalarBlockLayoutFeaturesEXT *features =
1253 (VkPhysicalDeviceScalarBlockLayoutFeaturesEXT *)ext;
1254 CORE_FEATURE(1, 2, scalarBlockLayout);
1255 break;
1256 }
1257
1258 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SEPARATE_DEPTH_STENCIL_LAYOUTS_FEATURES_KHR: {
1259 VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR *features =
1260 (VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR *)ext;
1261 CORE_FEATURE(1, 2, separateDepthStencilLayouts);
1262 break;
1263 }
1264
1265 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_INT64_FEATURES_KHR: {
1266 VkPhysicalDeviceShaderAtomicInt64FeaturesKHR *features = (void *)ext;
1267 CORE_FEATURE(1, 2, shaderBufferInt64Atomics);
1268 CORE_FEATURE(1, 2, shaderSharedInt64Atomics);
1269 break;
1270 }
1271
1272 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DEMOTE_TO_HELPER_INVOCATION_FEATURES_EXT: {
1273 VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT *features = (void *)ext;
1274 features->shaderDemoteToHelperInvocation = true;
1275 break;
1276 }
1277
1278 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CLOCK_FEATURES_KHR: {
1279 VkPhysicalDeviceShaderClockFeaturesKHR *features =
1280 (VkPhysicalDeviceShaderClockFeaturesKHR *)ext;
1281 features->shaderSubgroupClock = true;
1282 features->shaderDeviceClock = false;
1283 break;
1284 }
1285
1286 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETERS_FEATURES: {
1287 VkPhysicalDeviceShaderDrawParametersFeatures *features = (void *)ext;
1288 CORE_FEATURE(1, 1, shaderDrawParameters);
1289 break;
1290 }
1291
1292 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_SUBGROUP_EXTENDED_TYPES_FEATURES_KHR: {
1293 VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR *features =
1294 (VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR *)ext;
1295 CORE_FEATURE(1, 2, shaderSubgroupExtendedTypes);
1296 break;
1297 }
1298
1299 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_FEATURES_EXT: {
1300 VkPhysicalDeviceSubgroupSizeControlFeaturesEXT *features =
1301 (VkPhysicalDeviceSubgroupSizeControlFeaturesEXT *)ext;
1302 features->subgroupSizeControl = true;
1303 features->computeFullSubgroups = true;
1304 break;
1305 }
1306
1307 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_FEATURES_EXT: {
1308 VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT *features =
1309 (VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT *)ext;
1310 features->texelBufferAlignment = true;
1311 break;
1312 }
1313
1314 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_FEATURES_KHR: {
1315 VkPhysicalDeviceTimelineSemaphoreFeaturesKHR *features =
1316 (VkPhysicalDeviceTimelineSemaphoreFeaturesKHR *) ext;
1317 CORE_FEATURE(1, 2, timelineSemaphore);
1318 break;
1319 }
1320
1321 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES: {
1322 VkPhysicalDeviceVariablePointersFeatures *features = (void *)ext;
1323 CORE_FEATURE(1, 1, variablePointersStorageBuffer);
1324 CORE_FEATURE(1, 1, variablePointers);
1325 break;
1326 }
1327
1328 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT: {
1329 VkPhysicalDeviceTransformFeedbackFeaturesEXT *features =
1330 (VkPhysicalDeviceTransformFeedbackFeaturesEXT *)ext;
1331 features->transformFeedback = true;
1332 features->geometryStreams = true;
1333 break;
1334 }
1335
1336 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_UNIFORM_BUFFER_STANDARD_LAYOUT_FEATURES_KHR: {
1337 VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR *features =
1338 (VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR *)ext;
1339 CORE_FEATURE(1, 2, uniformBufferStandardLayout);
1340 break;
1341 }
1342
1343 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT: {
1344 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *features =
1345 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *)ext;
1346 features->vertexAttributeInstanceRateDivisor = true;
1347 features->vertexAttributeInstanceRateZeroDivisor = true;
1348 break;
1349 }
1350
1351 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES:
1352 anv_get_physical_device_features_1_1(pdevice, (void *)ext);
1353 break;
1354
1355 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES:
1356 anv_get_physical_device_features_1_2(pdevice, (void *)ext);
1357 break;
1358
1359 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_MEMORY_MODEL_FEATURES_KHR: {
1360 VkPhysicalDeviceVulkanMemoryModelFeaturesKHR *features = (void *)ext;
1361 CORE_FEATURE(1, 2, vulkanMemoryModel);
1362 CORE_FEATURE(1, 2, vulkanMemoryModelDeviceScope);
1363 CORE_FEATURE(1, 2, vulkanMemoryModelAvailabilityVisibilityChains);
1364 break;
1365 }
1366
1367 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_YCBCR_IMAGE_ARRAYS_FEATURES_EXT: {
1368 VkPhysicalDeviceYcbcrImageArraysFeaturesEXT *features =
1369 (VkPhysicalDeviceYcbcrImageArraysFeaturesEXT *)ext;
1370 features->ycbcrImageArrays = true;
1371 break;
1372 }
1373
1374 default:
1375 anv_debug_ignored_stype(ext->sType);
1376 break;
1377 }
1378 }
1379
1380 #undef CORE_FEATURE
1381 }
1382
1383 #define MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS 64
1384
1385 #define MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS 64
1386 #define MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS 256
1387
1388 void anv_GetPhysicalDeviceProperties(
1389 VkPhysicalDevice physicalDevice,
1390 VkPhysicalDeviceProperties* pProperties)
1391 {
1392 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
1393 const struct gen_device_info *devinfo = &pdevice->info;
1394
1395 /* See assertions made when programming the buffer surface state. */
1396 const uint32_t max_raw_buffer_sz = devinfo->gen >= 7 ?
1397 (1ul << 30) : (1ul << 27);
1398
1399 const uint32_t max_ssbos = pdevice->has_a64_buffer_access ? UINT16_MAX : 64;
1400 const uint32_t max_textures =
1401 pdevice->has_bindless_images ? UINT16_MAX : 128;
1402 const uint32_t max_samplers =
1403 pdevice->has_bindless_samplers ? UINT16_MAX :
1404 (devinfo->gen >= 8 || devinfo->is_haswell) ? 128 : 16;
1405 const uint32_t max_images =
1406 pdevice->has_bindless_images ? UINT16_MAX : MAX_IMAGES;
1407
1408 /* If we can use bindless for everything, claim a high per-stage limit,
1409 * otherwise use the binding table size, minus the slots reserved for
1410 * render targets and one slot for the descriptor buffer. */
1411 const uint32_t max_per_stage =
1412 pdevice->has_bindless_images && pdevice->has_a64_buffer_access
1413 ? UINT32_MAX : MAX_BINDING_TABLE_SIZE - MAX_RTS - 1;
1414
1415 const uint32_t max_workgroup_size = 32 * devinfo->max_cs_threads;
1416
1417 VkSampleCountFlags sample_counts =
1418 isl_device_get_sample_counts(&pdevice->isl_dev);
1419
1420
1421 VkPhysicalDeviceLimits limits = {
1422 .maxImageDimension1D = (1 << 14),
1423 .maxImageDimension2D = (1 << 14),
1424 .maxImageDimension3D = (1 << 11),
1425 .maxImageDimensionCube = (1 << 14),
1426 .maxImageArrayLayers = (1 << 11),
1427 .maxTexelBufferElements = 128 * 1024 * 1024,
1428 .maxUniformBufferRange = (1ul << 27),
1429 .maxStorageBufferRange = max_raw_buffer_sz,
1430 .maxPushConstantsSize = MAX_PUSH_CONSTANTS_SIZE,
1431 .maxMemoryAllocationCount = UINT32_MAX,
1432 .maxSamplerAllocationCount = 64 * 1024,
1433 .bufferImageGranularity = 64, /* A cache line */
1434 .sparseAddressSpaceSize = 0,
1435 .maxBoundDescriptorSets = MAX_SETS,
1436 .maxPerStageDescriptorSamplers = max_samplers,
1437 .maxPerStageDescriptorUniformBuffers = MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS,
1438 .maxPerStageDescriptorStorageBuffers = max_ssbos,
1439 .maxPerStageDescriptorSampledImages = max_textures,
1440 .maxPerStageDescriptorStorageImages = max_images,
1441 .maxPerStageDescriptorInputAttachments = MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS,
1442 .maxPerStageResources = max_per_stage,
1443 .maxDescriptorSetSamplers = 6 * max_samplers, /* number of stages * maxPerStageDescriptorSamplers */
1444 .maxDescriptorSetUniformBuffers = 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS, /* number of stages * maxPerStageDescriptorUniformBuffers */
1445 .maxDescriptorSetUniformBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2,
1446 .maxDescriptorSetStorageBuffers = 6 * max_ssbos, /* number of stages * maxPerStageDescriptorStorageBuffers */
1447 .maxDescriptorSetStorageBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2,
1448 .maxDescriptorSetSampledImages = 6 * max_textures, /* number of stages * maxPerStageDescriptorSampledImages */
1449 .maxDescriptorSetStorageImages = 6 * max_images, /* number of stages * maxPerStageDescriptorStorageImages */
1450 .maxDescriptorSetInputAttachments = MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS,
1451 .maxVertexInputAttributes = MAX_VBS,
1452 .maxVertexInputBindings = MAX_VBS,
1453 .maxVertexInputAttributeOffset = 2047,
1454 .maxVertexInputBindingStride = 2048,
1455 .maxVertexOutputComponents = 128,
1456 .maxTessellationGenerationLevel = 64,
1457 .maxTessellationPatchSize = 32,
1458 .maxTessellationControlPerVertexInputComponents = 128,
1459 .maxTessellationControlPerVertexOutputComponents = 128,
1460 .maxTessellationControlPerPatchOutputComponents = 128,
1461 .maxTessellationControlTotalOutputComponents = 2048,
1462 .maxTessellationEvaluationInputComponents = 128,
1463 .maxTessellationEvaluationOutputComponents = 128,
1464 .maxGeometryShaderInvocations = 32,
1465 .maxGeometryInputComponents = 64,
1466 .maxGeometryOutputComponents = 128,
1467 .maxGeometryOutputVertices = 256,
1468 .maxGeometryTotalOutputComponents = 1024,
1469 .maxFragmentInputComponents = 116, /* 128 components - (PSIZ, CLIP_DIST0, CLIP_DIST1) */
1470 .maxFragmentOutputAttachments = 8,
1471 .maxFragmentDualSrcAttachments = 1,
1472 .maxFragmentCombinedOutputResources = 8,
1473 .maxComputeSharedMemorySize = 64 * 1024,
1474 .maxComputeWorkGroupCount = { 65535, 65535, 65535 },
1475 .maxComputeWorkGroupInvocations = max_workgroup_size,
1476 .maxComputeWorkGroupSize = {
1477 max_workgroup_size,
1478 max_workgroup_size,
1479 max_workgroup_size,
1480 },
1481 .subPixelPrecisionBits = 8,
1482 .subTexelPrecisionBits = 8,
1483 .mipmapPrecisionBits = 8,
1484 .maxDrawIndexedIndexValue = UINT32_MAX,
1485 .maxDrawIndirectCount = UINT32_MAX,
1486 .maxSamplerLodBias = 16,
1487 .maxSamplerAnisotropy = 16,
1488 .maxViewports = MAX_VIEWPORTS,
1489 .maxViewportDimensions = { (1 << 14), (1 << 14) },
1490 .viewportBoundsRange = { INT16_MIN, INT16_MAX },
1491 .viewportSubPixelBits = 13, /* We take a float? */
1492 .minMemoryMapAlignment = 4096, /* A page */
1493 /* The dataport requires texel alignment so we need to assume a worst
1494 * case of R32G32B32A32 which is 16 bytes.
1495 */
1496 .minTexelBufferOffsetAlignment = 16,
1497 /* We need 16 for UBO block reads to work and 32 for push UBOs */
1498 .minUniformBufferOffsetAlignment = 32,
1499 .minStorageBufferOffsetAlignment = 4,
1500 .minTexelOffset = -8,
1501 .maxTexelOffset = 7,
1502 .minTexelGatherOffset = -32,
1503 .maxTexelGatherOffset = 31,
1504 .minInterpolationOffset = -0.5,
1505 .maxInterpolationOffset = 0.4375,
1506 .subPixelInterpolationOffsetBits = 4,
1507 .maxFramebufferWidth = (1 << 14),
1508 .maxFramebufferHeight = (1 << 14),
1509 .maxFramebufferLayers = (1 << 11),
1510 .framebufferColorSampleCounts = sample_counts,
1511 .framebufferDepthSampleCounts = sample_counts,
1512 .framebufferStencilSampleCounts = sample_counts,
1513 .framebufferNoAttachmentsSampleCounts = sample_counts,
1514 .maxColorAttachments = MAX_RTS,
1515 .sampledImageColorSampleCounts = sample_counts,
1516 .sampledImageIntegerSampleCounts = sample_counts,
1517 .sampledImageDepthSampleCounts = sample_counts,
1518 .sampledImageStencilSampleCounts = sample_counts,
1519 .storageImageSampleCounts = VK_SAMPLE_COUNT_1_BIT,
1520 .maxSampleMaskWords = 1,
1521 .timestampComputeAndGraphics = true,
1522 .timestampPeriod = 1000000000.0 / devinfo->timestamp_frequency,
1523 .maxClipDistances = 8,
1524 .maxCullDistances = 8,
1525 .maxCombinedClipAndCullDistances = 8,
1526 .discreteQueuePriorities = 2,
1527 .pointSizeRange = { 0.125, 255.875 },
1528 .lineWidthRange = {
1529 0.0,
1530 (devinfo->gen >= 9 || devinfo->is_cherryview) ?
1531 2047.9921875 : 7.9921875,
1532 },
1533 .pointSizeGranularity = (1.0 / 8.0),
1534 .lineWidthGranularity = (1.0 / 128.0),
1535 .strictLines = false,
1536 .standardSampleLocations = true,
1537 .optimalBufferCopyOffsetAlignment = 128,
1538 .optimalBufferCopyRowPitchAlignment = 128,
1539 .nonCoherentAtomSize = 64,
1540 };
1541
1542 *pProperties = (VkPhysicalDeviceProperties) {
1543 .apiVersion = anv_physical_device_api_version(pdevice),
1544 .driverVersion = vk_get_driver_version(),
1545 .vendorID = 0x8086,
1546 .deviceID = pdevice->info.chipset_id,
1547 .deviceType = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
1548 .limits = limits,
1549 .sparseProperties = {0}, /* Broadwell doesn't do sparse. */
1550 };
1551
1552 snprintf(pProperties->deviceName, sizeof(pProperties->deviceName),
1553 "%s", pdevice->name);
1554 memcpy(pProperties->pipelineCacheUUID,
1555 pdevice->pipeline_cache_uuid, VK_UUID_SIZE);
1556 }
1557
1558 static void
1559 anv_get_physical_device_properties_1_1(struct anv_physical_device *pdevice,
1560 VkPhysicalDeviceVulkan11Properties *p)
1561 {
1562 assert(p->sType == VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES);
1563
1564 memcpy(p->deviceUUID, pdevice->device_uuid, VK_UUID_SIZE);
1565 memcpy(p->driverUUID, pdevice->driver_uuid, VK_UUID_SIZE);
1566 memset(p->deviceLUID, 0, VK_LUID_SIZE);
1567 p->deviceNodeMask = 0;
1568 p->deviceLUIDValid = false;
1569
1570 p->subgroupSize = BRW_SUBGROUP_SIZE;
1571 VkShaderStageFlags scalar_stages = 0;
1572 for (unsigned stage = 0; stage < MESA_SHADER_STAGES; stage++) {
1573 if (pdevice->compiler->scalar_stage[stage])
1574 scalar_stages |= mesa_to_vk_shader_stage(stage);
1575 }
1576 p->subgroupSupportedStages = scalar_stages;
1577 p->subgroupSupportedOperations = VK_SUBGROUP_FEATURE_BASIC_BIT |
1578 VK_SUBGROUP_FEATURE_VOTE_BIT |
1579 VK_SUBGROUP_FEATURE_BALLOT_BIT |
1580 VK_SUBGROUP_FEATURE_SHUFFLE_BIT |
1581 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT |
1582 VK_SUBGROUP_FEATURE_QUAD_BIT;
1583 if (pdevice->info.gen >= 8) {
1584 /* TODO: There's no technical reason why these can't be made to
1585 * work on gen7 but they don't at the moment so it's best to leave
1586 * the feature disabled than enabled and broken.
1587 */
1588 p->subgroupSupportedOperations |= VK_SUBGROUP_FEATURE_ARITHMETIC_BIT |
1589 VK_SUBGROUP_FEATURE_CLUSTERED_BIT;
1590 }
1591 p->subgroupQuadOperationsInAllStages = pdevice->info.gen >= 8;
1592
1593 p->pointClippingBehavior = VK_POINT_CLIPPING_BEHAVIOR_USER_CLIP_PLANES_ONLY;
1594 p->maxMultiviewViewCount = 16;
1595 p->maxMultiviewInstanceIndex = UINT32_MAX / 16;
1596 p->protectedNoFault = false;
1597 /* This value doesn't matter for us today as our per-stage descriptors are
1598 * the real limit.
1599 */
1600 p->maxPerSetDescriptors = 1024;
1601 p->maxMemoryAllocationSize = MAX_MEMORY_ALLOCATION_SIZE;
1602 }
1603
1604 static void
1605 anv_get_physical_device_properties_1_2(struct anv_physical_device *pdevice,
1606 VkPhysicalDeviceVulkan12Properties *p)
1607 {
1608 assert(p->sType == VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES);
1609
1610 p->driverID = VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA_KHR;
1611 memset(p->driverName, 0, sizeof(p->driverName));
1612 snprintf(p->driverName, VK_MAX_DRIVER_NAME_SIZE_KHR,
1613 "Intel open-source Mesa driver");
1614 memset(p->driverInfo, 0, sizeof(p->driverInfo));
1615 snprintf(p->driverInfo, VK_MAX_DRIVER_INFO_SIZE_KHR,
1616 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1);
1617 p->conformanceVersion = (VkConformanceVersionKHR) {
1618 .major = 1,
1619 .minor = 2,
1620 .subminor = 0,
1621 .patch = 0,
1622 };
1623
1624 p->denormBehaviorIndependence =
1625 VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_ALL_KHR;
1626 p->roundingModeIndependence =
1627 VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_NONE_KHR;
1628
1629 /* Broadwell does not support HF denorms and there are restrictions
1630 * other gens. According to Kabylake's PRM:
1631 *
1632 * "math - Extended Math Function
1633 * [...]
1634 * Restriction : Half-float denorms are always retained."
1635 */
1636 p->shaderDenormFlushToZeroFloat16 = false;
1637 p->shaderDenormPreserveFloat16 = pdevice->info.gen > 8;
1638 p->shaderRoundingModeRTEFloat16 = true;
1639 p->shaderRoundingModeRTZFloat16 = true;
1640 p->shaderSignedZeroInfNanPreserveFloat16 = true;
1641
1642 p->shaderDenormFlushToZeroFloat32 = true;
1643 p->shaderDenormPreserveFloat32 = true;
1644 p->shaderRoundingModeRTEFloat32 = true;
1645 p->shaderRoundingModeRTZFloat32 = true;
1646 p->shaderSignedZeroInfNanPreserveFloat32 = true;
1647
1648 p->shaderDenormFlushToZeroFloat64 = true;
1649 p->shaderDenormPreserveFloat64 = true;
1650 p->shaderRoundingModeRTEFloat64 = true;
1651 p->shaderRoundingModeRTZFloat64 = true;
1652 p->shaderSignedZeroInfNanPreserveFloat64 = true;
1653
1654 /* It's a bit hard to exactly map our implementation to the limits
1655 * described here. The bindless surface handle in the extended
1656 * message descriptors is 20 bits and it's an index into the table of
1657 * RENDER_SURFACE_STATE structs that starts at bindless surface base
1658 * address. Given that most things consume two surface states per
1659 * view (general/sampled for textures and write-only/read-write for
1660 * images), we claim 2^19 things.
1661 *
1662 * For SSBOs, we just use A64 messages so there is no real limit
1663 * there beyond the limit on the total size of a descriptor set.
1664 */
1665 const unsigned max_bindless_views = 1 << 19;
1666 p->maxUpdateAfterBindDescriptorsInAllPools = max_bindless_views;
1667 p->shaderUniformBufferArrayNonUniformIndexingNative = false;
1668 p->shaderSampledImageArrayNonUniformIndexingNative = false;
1669 p->shaderStorageBufferArrayNonUniformIndexingNative = true;
1670 p->shaderStorageImageArrayNonUniformIndexingNative = false;
1671 p->shaderInputAttachmentArrayNonUniformIndexingNative = false;
1672 p->robustBufferAccessUpdateAfterBind = true;
1673 p->quadDivergentImplicitLod = false;
1674 p->maxPerStageDescriptorUpdateAfterBindSamplers = max_bindless_views;
1675 p->maxPerStageDescriptorUpdateAfterBindUniformBuffers = MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS;
1676 p->maxPerStageDescriptorUpdateAfterBindStorageBuffers = UINT32_MAX;
1677 p->maxPerStageDescriptorUpdateAfterBindSampledImages = max_bindless_views;
1678 p->maxPerStageDescriptorUpdateAfterBindStorageImages = max_bindless_views;
1679 p->maxPerStageDescriptorUpdateAfterBindInputAttachments = MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS;
1680 p->maxPerStageUpdateAfterBindResources = UINT32_MAX;
1681 p->maxDescriptorSetUpdateAfterBindSamplers = max_bindless_views;
1682 p->maxDescriptorSetUpdateAfterBindUniformBuffers = 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS;
1683 p->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2;
1684 p->maxDescriptorSetUpdateAfterBindStorageBuffers = UINT32_MAX;
1685 p->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2;
1686 p->maxDescriptorSetUpdateAfterBindSampledImages = max_bindless_views;
1687 p->maxDescriptorSetUpdateAfterBindStorageImages = max_bindless_views;
1688 p->maxDescriptorSetUpdateAfterBindInputAttachments = MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS;
1689
1690 /* We support all of the depth resolve modes */
1691 p->supportedDepthResolveModes = VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR |
1692 VK_RESOLVE_MODE_AVERAGE_BIT_KHR |
1693 VK_RESOLVE_MODE_MIN_BIT_KHR |
1694 VK_RESOLVE_MODE_MAX_BIT_KHR;
1695 /* Average doesn't make sense for stencil so we don't support that */
1696 p->supportedStencilResolveModes = VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR;
1697 if (pdevice->info.gen >= 8) {
1698 /* The advanced stencil resolve modes currently require stencil
1699 * sampling be supported by the hardware.
1700 */
1701 p->supportedStencilResolveModes |= VK_RESOLVE_MODE_MIN_BIT_KHR |
1702 VK_RESOLVE_MODE_MAX_BIT_KHR;
1703 }
1704 p->independentResolveNone = true;
1705 p->independentResolve = true;
1706
1707 p->filterMinmaxSingleComponentFormats = pdevice->info.gen >= 9;
1708 p->filterMinmaxImageComponentMapping = pdevice->info.gen >= 9;
1709
1710 p->maxTimelineSemaphoreValueDifference = UINT64_MAX;
1711
1712 p->framebufferIntegerColorSampleCounts =
1713 isl_device_get_sample_counts(&pdevice->isl_dev);
1714 }
1715
1716 void anv_GetPhysicalDeviceProperties2(
1717 VkPhysicalDevice physicalDevice,
1718 VkPhysicalDeviceProperties2* pProperties)
1719 {
1720 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
1721
1722 anv_GetPhysicalDeviceProperties(physicalDevice, &pProperties->properties);
1723
1724 VkPhysicalDeviceVulkan11Properties core_1_1 = {
1725 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES,
1726 };
1727 anv_get_physical_device_properties_1_1(pdevice, &core_1_1);
1728
1729 VkPhysicalDeviceVulkan12Properties core_1_2 = {
1730 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES,
1731 };
1732 anv_get_physical_device_properties_1_2(pdevice, &core_1_2);
1733
1734 #define CORE_RENAMED_PROPERTY(major, minor, ext_property, core_property) \
1735 memcpy(&properties->ext_property, &core_##major##_##minor.core_property, \
1736 sizeof(core_##major##_##minor.core_property))
1737
1738 #define CORE_PROPERTY(major, minor, property) \
1739 CORE_RENAMED_PROPERTY(major, minor, property, property)
1740
1741 vk_foreach_struct(ext, pProperties->pNext) {
1742 switch (ext->sType) {
1743 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_STENCIL_RESOLVE_PROPERTIES_KHR: {
1744 VkPhysicalDeviceDepthStencilResolvePropertiesKHR *properties =
1745 (VkPhysicalDeviceDepthStencilResolvePropertiesKHR *)ext;
1746 CORE_PROPERTY(1, 2, supportedDepthResolveModes);
1747 CORE_PROPERTY(1, 2, supportedStencilResolveModes);
1748 CORE_PROPERTY(1, 2, independentResolveNone);
1749 CORE_PROPERTY(1, 2, independentResolve);
1750 break;
1751 }
1752
1753 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES_EXT: {
1754 VkPhysicalDeviceDescriptorIndexingPropertiesEXT *properties =
1755 (VkPhysicalDeviceDescriptorIndexingPropertiesEXT *)ext;
1756 CORE_PROPERTY(1, 2, maxUpdateAfterBindDescriptorsInAllPools);
1757 CORE_PROPERTY(1, 2, shaderUniformBufferArrayNonUniformIndexingNative);
1758 CORE_PROPERTY(1, 2, shaderSampledImageArrayNonUniformIndexingNative);
1759 CORE_PROPERTY(1, 2, shaderStorageBufferArrayNonUniformIndexingNative);
1760 CORE_PROPERTY(1, 2, shaderStorageImageArrayNonUniformIndexingNative);
1761 CORE_PROPERTY(1, 2, shaderInputAttachmentArrayNonUniformIndexingNative);
1762 CORE_PROPERTY(1, 2, robustBufferAccessUpdateAfterBind);
1763 CORE_PROPERTY(1, 2, quadDivergentImplicitLod);
1764 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindSamplers);
1765 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindUniformBuffers);
1766 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindStorageBuffers);
1767 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindSampledImages);
1768 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindStorageImages);
1769 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindInputAttachments);
1770 CORE_PROPERTY(1, 2, maxPerStageUpdateAfterBindResources);
1771 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindSamplers);
1772 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindUniformBuffers);
1773 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindUniformBuffersDynamic);
1774 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindStorageBuffers);
1775 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindStorageBuffersDynamic);
1776 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindSampledImages);
1777 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindStorageImages);
1778 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindInputAttachments);
1779 break;
1780 }
1781
1782 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR: {
1783 VkPhysicalDeviceDriverPropertiesKHR *properties =
1784 (VkPhysicalDeviceDriverPropertiesKHR *) ext;
1785 CORE_PROPERTY(1, 2, driverID);
1786 CORE_PROPERTY(1, 2, driverName);
1787 CORE_PROPERTY(1, 2, driverInfo);
1788 CORE_PROPERTY(1, 2, conformanceVersion);
1789 break;
1790 }
1791
1792 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT: {
1793 VkPhysicalDeviceExternalMemoryHostPropertiesEXT *props =
1794 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT *) ext;
1795 /* Userptr needs page aligned memory. */
1796 props->minImportedHostPointerAlignment = 4096;
1797 break;
1798 }
1799
1800 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES: {
1801 VkPhysicalDeviceIDProperties *properties =
1802 (VkPhysicalDeviceIDProperties *)ext;
1803 CORE_PROPERTY(1, 1, deviceUUID);
1804 CORE_PROPERTY(1, 1, driverUUID);
1805 CORE_PROPERTY(1, 1, deviceLUID);
1806 CORE_PROPERTY(1, 1, deviceLUIDValid);
1807 break;
1808 }
1809
1810 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_PROPERTIES_EXT: {
1811 VkPhysicalDeviceInlineUniformBlockPropertiesEXT *props =
1812 (VkPhysicalDeviceInlineUniformBlockPropertiesEXT *)ext;
1813 props->maxInlineUniformBlockSize = MAX_INLINE_UNIFORM_BLOCK_SIZE;
1814 props->maxPerStageDescriptorInlineUniformBlocks =
1815 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1816 props->maxPerStageDescriptorUpdateAfterBindInlineUniformBlocks =
1817 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1818 props->maxDescriptorSetInlineUniformBlocks =
1819 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1820 props->maxDescriptorSetUpdateAfterBindInlineUniformBlocks =
1821 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1822 break;
1823 }
1824
1825 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_PROPERTIES_EXT: {
1826 VkPhysicalDeviceLineRasterizationPropertiesEXT *props =
1827 (VkPhysicalDeviceLineRasterizationPropertiesEXT *)ext;
1828 /* In the Skylake PRM Vol. 7, subsection titled "GIQ (Diamond)
1829 * Sampling Rules - Legacy Mode", it says the following:
1830 *
1831 * "Note that the device divides a pixel into a 16x16 array of
1832 * subpixels, referenced by their upper left corners."
1833 *
1834 * This is the only known reference in the PRMs to the subpixel
1835 * precision of line rasterization and a "16x16 array of subpixels"
1836 * implies 4 subpixel precision bits. Empirical testing has shown
1837 * that 4 subpixel precision bits applies to all line rasterization
1838 * types.
1839 */
1840 props->lineSubPixelPrecisionBits = 4;
1841 break;
1842 }
1843
1844 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES: {
1845 VkPhysicalDeviceMaintenance3Properties *properties =
1846 (VkPhysicalDeviceMaintenance3Properties *)ext;
1847 /* This value doesn't matter for us today as our per-stage
1848 * descriptors are the real limit.
1849 */
1850 CORE_PROPERTY(1, 1, maxPerSetDescriptors);
1851 CORE_PROPERTY(1, 1, maxMemoryAllocationSize);
1852 break;
1853 }
1854
1855 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES: {
1856 VkPhysicalDeviceMultiviewProperties *properties =
1857 (VkPhysicalDeviceMultiviewProperties *)ext;
1858 CORE_PROPERTY(1, 1, maxMultiviewViewCount);
1859 CORE_PROPERTY(1, 1, maxMultiviewInstanceIndex);
1860 break;
1861 }
1862
1863 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT: {
1864 VkPhysicalDevicePCIBusInfoPropertiesEXT *properties =
1865 (VkPhysicalDevicePCIBusInfoPropertiesEXT *)ext;
1866 properties->pciDomain = pdevice->pci_info.domain;
1867 properties->pciBus = pdevice->pci_info.bus;
1868 properties->pciDevice = pdevice->pci_info.device;
1869 properties->pciFunction = pdevice->pci_info.function;
1870 break;
1871 }
1872
1873 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES: {
1874 VkPhysicalDevicePointClippingProperties *properties =
1875 (VkPhysicalDevicePointClippingProperties *) ext;
1876 CORE_PROPERTY(1, 1, pointClippingBehavior);
1877 break;
1878 }
1879
1880 #pragma GCC diagnostic push
1881 #pragma GCC diagnostic ignored "-Wswitch"
1882 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PRESENTATION_PROPERTIES_ANDROID: {
1883 VkPhysicalDevicePresentationPropertiesANDROID *props =
1884 (VkPhysicalDevicePresentationPropertiesANDROID *)ext;
1885 props->sharedImage = VK_FALSE;
1886 break;
1887 }
1888 #pragma GCC diagnostic pop
1889
1890 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES: {
1891 VkPhysicalDeviceProtectedMemoryProperties *properties =
1892 (VkPhysicalDeviceProtectedMemoryProperties *)ext;
1893 CORE_PROPERTY(1, 1, protectedNoFault);
1894 break;
1895 }
1896
1897 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR: {
1898 VkPhysicalDevicePushDescriptorPropertiesKHR *properties =
1899 (VkPhysicalDevicePushDescriptorPropertiesKHR *) ext;
1900 properties->maxPushDescriptors = MAX_PUSH_DESCRIPTORS;
1901 break;
1902 }
1903
1904 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT: {
1905 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT *properties =
1906 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT *)ext;
1907 CORE_PROPERTY(1, 2, filterMinmaxImageComponentMapping);
1908 CORE_PROPERTY(1, 2, filterMinmaxSingleComponentFormats);
1909 break;
1910 }
1911
1912 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES: {
1913 VkPhysicalDeviceSubgroupProperties *properties = (void *)ext;
1914 CORE_PROPERTY(1, 1, subgroupSize);
1915 CORE_RENAMED_PROPERTY(1, 1, supportedStages,
1916 subgroupSupportedStages);
1917 CORE_RENAMED_PROPERTY(1, 1, supportedOperations,
1918 subgroupSupportedOperations);
1919 CORE_RENAMED_PROPERTY(1, 1, quadOperationsInAllStages,
1920 subgroupQuadOperationsInAllStages);
1921 break;
1922 }
1923
1924 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_PROPERTIES_EXT: {
1925 VkPhysicalDeviceSubgroupSizeControlPropertiesEXT *props =
1926 (VkPhysicalDeviceSubgroupSizeControlPropertiesEXT *)ext;
1927 STATIC_ASSERT(8 <= BRW_SUBGROUP_SIZE && BRW_SUBGROUP_SIZE <= 32);
1928 props->minSubgroupSize = 8;
1929 props->maxSubgroupSize = 32;
1930 props->maxComputeWorkgroupSubgroups = pdevice->info.max_cs_threads;
1931 props->requiredSubgroupSizeStages = VK_SHADER_STAGE_COMPUTE_BIT;
1932 break;
1933 }
1934 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT_CONTROLS_PROPERTIES_KHR : {
1935 VkPhysicalDeviceFloatControlsPropertiesKHR *properties = (void *)ext;
1936 CORE_PROPERTY(1, 2, denormBehaviorIndependence);
1937 CORE_PROPERTY(1, 2, roundingModeIndependence);
1938 CORE_PROPERTY(1, 2, shaderDenormFlushToZeroFloat16);
1939 CORE_PROPERTY(1, 2, shaderDenormPreserveFloat16);
1940 CORE_PROPERTY(1, 2, shaderRoundingModeRTEFloat16);
1941 CORE_PROPERTY(1, 2, shaderRoundingModeRTZFloat16);
1942 CORE_PROPERTY(1, 2, shaderSignedZeroInfNanPreserveFloat16);
1943 CORE_PROPERTY(1, 2, shaderDenormFlushToZeroFloat32);
1944 CORE_PROPERTY(1, 2, shaderDenormPreserveFloat32);
1945 CORE_PROPERTY(1, 2, shaderRoundingModeRTEFloat32);
1946 CORE_PROPERTY(1, 2, shaderRoundingModeRTZFloat32);
1947 CORE_PROPERTY(1, 2, shaderSignedZeroInfNanPreserveFloat32);
1948 CORE_PROPERTY(1, 2, shaderDenormFlushToZeroFloat64);
1949 CORE_PROPERTY(1, 2, shaderDenormPreserveFloat64);
1950 CORE_PROPERTY(1, 2, shaderRoundingModeRTEFloat64);
1951 CORE_PROPERTY(1, 2, shaderRoundingModeRTZFloat64);
1952 CORE_PROPERTY(1, 2, shaderSignedZeroInfNanPreserveFloat64);
1953 break;
1954 }
1955
1956 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_PROPERTIES_EXT: {
1957 VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT *props =
1958 (VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT *)ext;
1959
1960 /* From the SKL PRM Vol. 2d, docs for RENDER_SURFACE_STATE::Surface
1961 * Base Address:
1962 *
1963 * "For SURFTYPE_BUFFER non-rendertarget surfaces, this field
1964 * specifies the base address of the first element of the surface,
1965 * computed in software by adding the surface base address to the
1966 * byte offset of the element in the buffer. The base address must
1967 * be aligned to element size."
1968 *
1969 * The typed dataport messages require that things be texel aligned.
1970 * Otherwise, we may just load/store the wrong data or, in the worst
1971 * case, there may be hangs.
1972 */
1973 props->storageTexelBufferOffsetAlignmentBytes = 16;
1974 props->storageTexelBufferOffsetSingleTexelAlignment = true;
1975
1976 /* The sampler, however, is much more forgiving and it can handle
1977 * arbitrary byte alignment for linear and buffer surfaces. It's
1978 * hard to find a good PRM citation for this but years of empirical
1979 * experience demonstrate that this is true.
1980 */
1981 props->uniformTexelBufferOffsetAlignmentBytes = 1;
1982 props->uniformTexelBufferOffsetSingleTexelAlignment = false;
1983 break;
1984 }
1985
1986 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_PROPERTIES_KHR: {
1987 VkPhysicalDeviceTimelineSemaphorePropertiesKHR *properties =
1988 (VkPhysicalDeviceTimelineSemaphorePropertiesKHR *) ext;
1989 CORE_PROPERTY(1, 2, maxTimelineSemaphoreValueDifference);
1990 break;
1991 }
1992
1993 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT: {
1994 VkPhysicalDeviceTransformFeedbackPropertiesEXT *props =
1995 (VkPhysicalDeviceTransformFeedbackPropertiesEXT *)ext;
1996
1997 props->maxTransformFeedbackStreams = MAX_XFB_STREAMS;
1998 props->maxTransformFeedbackBuffers = MAX_XFB_BUFFERS;
1999 props->maxTransformFeedbackBufferSize = (1ull << 32);
2000 props->maxTransformFeedbackStreamDataSize = 128 * 4;
2001 props->maxTransformFeedbackBufferDataSize = 128 * 4;
2002 props->maxTransformFeedbackBufferDataStride = 2048;
2003 props->transformFeedbackQueries = true;
2004 props->transformFeedbackStreamsLinesTriangles = false;
2005 props->transformFeedbackRasterizationStreamSelect = false;
2006 props->transformFeedbackDraw = true;
2007 break;
2008 }
2009
2010 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT: {
2011 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *props =
2012 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *)ext;
2013 /* We have to restrict this a bit for multiview */
2014 props->maxVertexAttribDivisor = UINT32_MAX / 16;
2015 break;
2016 }
2017
2018 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES:
2019 anv_get_physical_device_properties_1_1(pdevice, (void *)ext);
2020 break;
2021
2022 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES:
2023 anv_get_physical_device_properties_1_2(pdevice, (void *)ext);
2024 break;
2025
2026 default:
2027 anv_debug_ignored_stype(ext->sType);
2028 break;
2029 }
2030 }
2031
2032 #undef CORE_RENAMED_PROPERTY
2033 #undef CORE_PROPERTY
2034 }
2035
2036 /* We support exactly one queue family. */
2037 static const VkQueueFamilyProperties
2038 anv_queue_family_properties = {
2039 .queueFlags = VK_QUEUE_GRAPHICS_BIT |
2040 VK_QUEUE_COMPUTE_BIT |
2041 VK_QUEUE_TRANSFER_BIT,
2042 .queueCount = 1,
2043 .timestampValidBits = 36, /* XXX: Real value here */
2044 .minImageTransferGranularity = { 1, 1, 1 },
2045 };
2046
2047 void anv_GetPhysicalDeviceQueueFamilyProperties(
2048 VkPhysicalDevice physicalDevice,
2049 uint32_t* pCount,
2050 VkQueueFamilyProperties* pQueueFamilyProperties)
2051 {
2052 VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pCount);
2053
2054 vk_outarray_append(&out, p) {
2055 *p = anv_queue_family_properties;
2056 }
2057 }
2058
2059 void anv_GetPhysicalDeviceQueueFamilyProperties2(
2060 VkPhysicalDevice physicalDevice,
2061 uint32_t* pQueueFamilyPropertyCount,
2062 VkQueueFamilyProperties2* pQueueFamilyProperties)
2063 {
2064
2065 VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pQueueFamilyPropertyCount);
2066
2067 vk_outarray_append(&out, p) {
2068 p->queueFamilyProperties = anv_queue_family_properties;
2069
2070 vk_foreach_struct(s, p->pNext) {
2071 anv_debug_ignored_stype(s->sType);
2072 }
2073 }
2074 }
2075
2076 void anv_GetPhysicalDeviceMemoryProperties(
2077 VkPhysicalDevice physicalDevice,
2078 VkPhysicalDeviceMemoryProperties* pMemoryProperties)
2079 {
2080 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
2081
2082 pMemoryProperties->memoryTypeCount = physical_device->memory.type_count;
2083 for (uint32_t i = 0; i < physical_device->memory.type_count; i++) {
2084 pMemoryProperties->memoryTypes[i] = (VkMemoryType) {
2085 .propertyFlags = physical_device->memory.types[i].propertyFlags,
2086 .heapIndex = physical_device->memory.types[i].heapIndex,
2087 };
2088 }
2089
2090 pMemoryProperties->memoryHeapCount = physical_device->memory.heap_count;
2091 for (uint32_t i = 0; i < physical_device->memory.heap_count; i++) {
2092 pMemoryProperties->memoryHeaps[i] = (VkMemoryHeap) {
2093 .size = physical_device->memory.heaps[i].size,
2094 .flags = physical_device->memory.heaps[i].flags,
2095 };
2096 }
2097 }
2098
2099 static void
2100 anv_get_memory_budget(VkPhysicalDevice physicalDevice,
2101 VkPhysicalDeviceMemoryBudgetPropertiesEXT *memoryBudget)
2102 {
2103 ANV_FROM_HANDLE(anv_physical_device, device, physicalDevice);
2104 uint64_t sys_available = get_available_system_memory();
2105 assert(sys_available > 0);
2106
2107 VkDeviceSize total_heaps_size = 0;
2108 for (size_t i = 0; i < device->memory.heap_count; i++)
2109 total_heaps_size += device->memory.heaps[i].size;
2110
2111 for (size_t i = 0; i < device->memory.heap_count; i++) {
2112 VkDeviceSize heap_size = device->memory.heaps[i].size;
2113 VkDeviceSize heap_used = device->memory.heaps[i].used;
2114 VkDeviceSize heap_budget;
2115
2116 double heap_proportion = (double) heap_size / total_heaps_size;
2117 VkDeviceSize sys_available_prop = sys_available * heap_proportion;
2118
2119 /*
2120 * Let's not incite the app to starve the system: report at most 90% of
2121 * available system memory.
2122 */
2123 uint64_t heap_available = sys_available_prop * 9 / 10;
2124 heap_budget = MIN2(heap_size, heap_used + heap_available);
2125
2126 /*
2127 * Round down to the nearest MB
2128 */
2129 heap_budget &= ~((1ull << 20) - 1);
2130
2131 /*
2132 * The heapBudget value must be non-zero for array elements less than
2133 * VkPhysicalDeviceMemoryProperties::memoryHeapCount. The heapBudget
2134 * value must be less than or equal to VkMemoryHeap::size for each heap.
2135 */
2136 assert(0 < heap_budget && heap_budget <= heap_size);
2137
2138 memoryBudget->heapUsage[i] = heap_used;
2139 memoryBudget->heapBudget[i] = heap_budget;
2140 }
2141
2142 /* The heapBudget and heapUsage values must be zero for array elements
2143 * greater than or equal to VkPhysicalDeviceMemoryProperties::memoryHeapCount
2144 */
2145 for (uint32_t i = device->memory.heap_count; i < VK_MAX_MEMORY_HEAPS; i++) {
2146 memoryBudget->heapBudget[i] = 0;
2147 memoryBudget->heapUsage[i] = 0;
2148 }
2149 }
2150
2151 void anv_GetPhysicalDeviceMemoryProperties2(
2152 VkPhysicalDevice physicalDevice,
2153 VkPhysicalDeviceMemoryProperties2* pMemoryProperties)
2154 {
2155 anv_GetPhysicalDeviceMemoryProperties(physicalDevice,
2156 &pMemoryProperties->memoryProperties);
2157
2158 vk_foreach_struct(ext, pMemoryProperties->pNext) {
2159 switch (ext->sType) {
2160 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT:
2161 anv_get_memory_budget(physicalDevice, (void*)ext);
2162 break;
2163 default:
2164 anv_debug_ignored_stype(ext->sType);
2165 break;
2166 }
2167 }
2168 }
2169
2170 void
2171 anv_GetDeviceGroupPeerMemoryFeatures(
2172 VkDevice device,
2173 uint32_t heapIndex,
2174 uint32_t localDeviceIndex,
2175 uint32_t remoteDeviceIndex,
2176 VkPeerMemoryFeatureFlags* pPeerMemoryFeatures)
2177 {
2178 assert(localDeviceIndex == 0 && remoteDeviceIndex == 0);
2179 *pPeerMemoryFeatures = VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT |
2180 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT |
2181 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT |
2182 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT;
2183 }
2184
2185 PFN_vkVoidFunction anv_GetInstanceProcAddr(
2186 VkInstance _instance,
2187 const char* pName)
2188 {
2189 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2190
2191 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
2192 * when we have to return valid function pointers, NULL, or it's left
2193 * undefined. See the table for exact details.
2194 */
2195 if (pName == NULL)
2196 return NULL;
2197
2198 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
2199 if (strcmp(pName, "vk" #entrypoint) == 0) \
2200 return (PFN_vkVoidFunction)anv_##entrypoint
2201
2202 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties);
2203 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties);
2204 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceVersion);
2205 LOOKUP_ANV_ENTRYPOINT(CreateInstance);
2206
2207 /* GetInstanceProcAddr() can also be called with a NULL instance.
2208 * See https://gitlab.khronos.org/vulkan/vulkan/issues/2057
2209 */
2210 LOOKUP_ANV_ENTRYPOINT(GetInstanceProcAddr);
2211
2212 #undef LOOKUP_ANV_ENTRYPOINT
2213
2214 if (instance == NULL)
2215 return NULL;
2216
2217 int idx = anv_get_instance_entrypoint_index(pName);
2218 if (idx >= 0)
2219 return instance->dispatch.entrypoints[idx];
2220
2221 idx = anv_get_physical_device_entrypoint_index(pName);
2222 if (idx >= 0)
2223 return instance->physical_device_dispatch.entrypoints[idx];
2224
2225 idx = anv_get_device_entrypoint_index(pName);
2226 if (idx >= 0)
2227 return instance->device_dispatch.entrypoints[idx];
2228
2229 return NULL;
2230 }
2231
2232 /* With version 1+ of the loader interface the ICD should expose
2233 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
2234 */
2235 PUBLIC
2236 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
2237 VkInstance instance,
2238 const char* pName);
2239
2240 PUBLIC
2241 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
2242 VkInstance instance,
2243 const char* pName)
2244 {
2245 return anv_GetInstanceProcAddr(instance, pName);
2246 }
2247
2248 PFN_vkVoidFunction anv_GetDeviceProcAddr(
2249 VkDevice _device,
2250 const char* pName)
2251 {
2252 ANV_FROM_HANDLE(anv_device, device, _device);
2253
2254 if (!device || !pName)
2255 return NULL;
2256
2257 int idx = anv_get_device_entrypoint_index(pName);
2258 if (idx < 0)
2259 return NULL;
2260
2261 return device->dispatch.entrypoints[idx];
2262 }
2263
2264 /* With version 4+ of the loader interface the ICD should expose
2265 * vk_icdGetPhysicalDeviceProcAddr()
2266 */
2267 PUBLIC
2268 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetPhysicalDeviceProcAddr(
2269 VkInstance _instance,
2270 const char* pName);
2271
2272 PFN_vkVoidFunction vk_icdGetPhysicalDeviceProcAddr(
2273 VkInstance _instance,
2274 const char* pName)
2275 {
2276 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2277
2278 if (!pName || !instance)
2279 return NULL;
2280
2281 int idx = anv_get_physical_device_entrypoint_index(pName);
2282 if (idx < 0)
2283 return NULL;
2284
2285 return instance->physical_device_dispatch.entrypoints[idx];
2286 }
2287
2288
2289 VkResult
2290 anv_CreateDebugReportCallbackEXT(VkInstance _instance,
2291 const VkDebugReportCallbackCreateInfoEXT* pCreateInfo,
2292 const VkAllocationCallbacks* pAllocator,
2293 VkDebugReportCallbackEXT* pCallback)
2294 {
2295 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2296 return vk_create_debug_report_callback(&instance->debug_report_callbacks,
2297 pCreateInfo, pAllocator, &instance->alloc,
2298 pCallback);
2299 }
2300
2301 void
2302 anv_DestroyDebugReportCallbackEXT(VkInstance _instance,
2303 VkDebugReportCallbackEXT _callback,
2304 const VkAllocationCallbacks* pAllocator)
2305 {
2306 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2307 vk_destroy_debug_report_callback(&instance->debug_report_callbacks,
2308 _callback, pAllocator, &instance->alloc);
2309 }
2310
2311 void
2312 anv_DebugReportMessageEXT(VkInstance _instance,
2313 VkDebugReportFlagsEXT flags,
2314 VkDebugReportObjectTypeEXT objectType,
2315 uint64_t object,
2316 size_t location,
2317 int32_t messageCode,
2318 const char* pLayerPrefix,
2319 const char* pMessage)
2320 {
2321 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2322 vk_debug_report(&instance->debug_report_callbacks, flags, objectType,
2323 object, location, messageCode, pLayerPrefix, pMessage);
2324 }
2325
2326 static struct anv_state
2327 anv_state_pool_emit_data(struct anv_state_pool *pool, size_t size, size_t align, const void *p)
2328 {
2329 struct anv_state state;
2330
2331 state = anv_state_pool_alloc(pool, size, align);
2332 memcpy(state.map, p, size);
2333
2334 return state;
2335 }
2336
2337 /* Haswell border color is a bit of a disaster. Float and unorm formats use a
2338 * straightforward 32-bit float color in the first 64 bytes. Instead of using
2339 * a nice float/integer union like Gen8+, Haswell specifies the integer border
2340 * color as a separate entry /after/ the float color. The layout of this entry
2341 * also depends on the format's bpp (with extra hacks for RG32), and overlaps.
2342 *
2343 * Since we don't know the format/bpp, we can't make any of the border colors
2344 * containing '1' work for all formats, as it would be in the wrong place for
2345 * some of them. We opt to make 32-bit integers work as this seems like the
2346 * most common option. Fortunately, transparent black works regardless, as
2347 * all zeroes is the same in every bit-size.
2348 */
2349 struct hsw_border_color {
2350 float float32[4];
2351 uint32_t _pad0[12];
2352 uint32_t uint32[4];
2353 uint32_t _pad1[108];
2354 };
2355
2356 struct gen8_border_color {
2357 union {
2358 float float32[4];
2359 uint32_t uint32[4];
2360 };
2361 /* Pad out to 64 bytes */
2362 uint32_t _pad[12];
2363 };
2364
2365 static void
2366 anv_device_init_border_colors(struct anv_device *device)
2367 {
2368 if (device->info.is_haswell) {
2369 static const struct hsw_border_color border_colors[] = {
2370 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
2371 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
2372 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
2373 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = { .uint32 = { 0, 0, 0, 0 } },
2374 [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = { .uint32 = { 0, 0, 0, 1 } },
2375 [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = { .uint32 = { 1, 1, 1, 1 } },
2376 };
2377
2378 device->border_colors =
2379 anv_state_pool_emit_data(&device->dynamic_state_pool,
2380 sizeof(border_colors), 512, border_colors);
2381 } else {
2382 static const struct gen8_border_color border_colors[] = {
2383 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
2384 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
2385 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
2386 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = { .uint32 = { 0, 0, 0, 0 } },
2387 [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = { .uint32 = { 0, 0, 0, 1 } },
2388 [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = { .uint32 = { 1, 1, 1, 1 } },
2389 };
2390
2391 device->border_colors =
2392 anv_state_pool_emit_data(&device->dynamic_state_pool,
2393 sizeof(border_colors), 64, border_colors);
2394 }
2395 }
2396
2397 static VkResult
2398 anv_device_init_trivial_batch(struct anv_device *device)
2399 {
2400 VkResult result = anv_device_alloc_bo(device, 4096,
2401 ANV_BO_ALLOC_MAPPED,
2402 0 /* explicit_address */,
2403 &device->trivial_batch_bo);
2404 if (result != VK_SUCCESS)
2405 return result;
2406
2407 struct anv_batch batch = {
2408 .start = device->trivial_batch_bo->map,
2409 .next = device->trivial_batch_bo->map,
2410 .end = device->trivial_batch_bo->map + 4096,
2411 };
2412
2413 anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
2414 anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
2415
2416 if (!device->info.has_llc)
2417 gen_clflush_range(batch.start, batch.next - batch.start);
2418
2419 return VK_SUCCESS;
2420 }
2421
2422 VkResult anv_EnumerateDeviceExtensionProperties(
2423 VkPhysicalDevice physicalDevice,
2424 const char* pLayerName,
2425 uint32_t* pPropertyCount,
2426 VkExtensionProperties* pProperties)
2427 {
2428 ANV_FROM_HANDLE(anv_physical_device, device, physicalDevice);
2429 VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
2430
2431 for (int i = 0; i < ANV_DEVICE_EXTENSION_COUNT; i++) {
2432 if (device->supported_extensions.extensions[i]) {
2433 vk_outarray_append(&out, prop) {
2434 *prop = anv_device_extensions[i];
2435 }
2436 }
2437 }
2438
2439 return vk_outarray_status(&out);
2440 }
2441
2442 static void
2443 anv_device_init_dispatch(struct anv_device *device)
2444 {
2445 const struct anv_instance *instance = device->physical->instance;
2446
2447 const struct anv_device_dispatch_table *genX_table;
2448 switch (device->info.gen) {
2449 case 12:
2450 genX_table = &gen12_device_dispatch_table;
2451 break;
2452 case 11:
2453 genX_table = &gen11_device_dispatch_table;
2454 break;
2455 case 10:
2456 genX_table = &gen10_device_dispatch_table;
2457 break;
2458 case 9:
2459 genX_table = &gen9_device_dispatch_table;
2460 break;
2461 case 8:
2462 genX_table = &gen8_device_dispatch_table;
2463 break;
2464 case 7:
2465 if (device->info.is_haswell)
2466 genX_table = &gen75_device_dispatch_table;
2467 else
2468 genX_table = &gen7_device_dispatch_table;
2469 break;
2470 default:
2471 unreachable("unsupported gen\n");
2472 }
2473
2474 for (unsigned i = 0; i < ARRAY_SIZE(device->dispatch.entrypoints); i++) {
2475 /* Vulkan requires that entrypoints for extensions which have not been
2476 * enabled must not be advertised.
2477 */
2478 if (!anv_device_entrypoint_is_enabled(i, instance->app_info.api_version,
2479 &instance->enabled_extensions,
2480 &device->enabled_extensions)) {
2481 device->dispatch.entrypoints[i] = NULL;
2482 } else if (genX_table->entrypoints[i]) {
2483 device->dispatch.entrypoints[i] = genX_table->entrypoints[i];
2484 } else {
2485 device->dispatch.entrypoints[i] =
2486 anv_device_dispatch_table.entrypoints[i];
2487 }
2488 }
2489 }
2490
2491 static int
2492 vk_priority_to_gen(int priority)
2493 {
2494 switch (priority) {
2495 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT:
2496 return GEN_CONTEXT_LOW_PRIORITY;
2497 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT:
2498 return GEN_CONTEXT_MEDIUM_PRIORITY;
2499 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT:
2500 return GEN_CONTEXT_HIGH_PRIORITY;
2501 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT:
2502 return GEN_CONTEXT_REALTIME_PRIORITY;
2503 default:
2504 unreachable("Invalid priority");
2505 }
2506 }
2507
2508 static VkResult
2509 anv_device_init_hiz_clear_value_bo(struct anv_device *device)
2510 {
2511 VkResult result = anv_device_alloc_bo(device, 4096,
2512 ANV_BO_ALLOC_MAPPED,
2513 0 /* explicit_address */,
2514 &device->hiz_clear_bo);
2515 if (result != VK_SUCCESS)
2516 return result;
2517
2518 union isl_color_value hiz_clear = { .u32 = { 0, } };
2519 hiz_clear.f32[0] = ANV_HZ_FC_VAL;
2520
2521 memcpy(device->hiz_clear_bo->map, hiz_clear.u32, sizeof(hiz_clear.u32));
2522
2523 if (!device->info.has_llc)
2524 gen_clflush_range(device->hiz_clear_bo->map, sizeof(hiz_clear.u32));
2525
2526 return VK_SUCCESS;
2527 }
2528
2529 static bool
2530 get_bo_from_pool(struct gen_batch_decode_bo *ret,
2531 struct anv_block_pool *pool,
2532 uint64_t address)
2533 {
2534 anv_block_pool_foreach_bo(bo, pool) {
2535 uint64_t bo_address = gen_48b_address(bo->offset);
2536 if (address >= bo_address && address < (bo_address + bo->size)) {
2537 *ret = (struct gen_batch_decode_bo) {
2538 .addr = bo_address,
2539 .size = bo->size,
2540 .map = bo->map,
2541 };
2542 return true;
2543 }
2544 }
2545 return false;
2546 }
2547
2548 /* Finding a buffer for batch decoding */
2549 static struct gen_batch_decode_bo
2550 decode_get_bo(void *v_batch, bool ppgtt, uint64_t address)
2551 {
2552 struct anv_device *device = v_batch;
2553 struct gen_batch_decode_bo ret_bo = {};
2554
2555 assert(ppgtt);
2556
2557 if (get_bo_from_pool(&ret_bo, &device->dynamic_state_pool.block_pool, address))
2558 return ret_bo;
2559 if (get_bo_from_pool(&ret_bo, &device->instruction_state_pool.block_pool, address))
2560 return ret_bo;
2561 if (get_bo_from_pool(&ret_bo, &device->binding_table_pool.block_pool, address))
2562 return ret_bo;
2563 if (get_bo_from_pool(&ret_bo, &device->surface_state_pool.block_pool, address))
2564 return ret_bo;
2565
2566 if (!device->cmd_buffer_being_decoded)
2567 return (struct gen_batch_decode_bo) { };
2568
2569 struct anv_batch_bo **bo;
2570
2571 u_vector_foreach(bo, &device->cmd_buffer_being_decoded->seen_bbos) {
2572 /* The decoder zeroes out the top 16 bits, so we need to as well */
2573 uint64_t bo_address = (*bo)->bo->offset & (~0ull >> 16);
2574
2575 if (address >= bo_address && address < bo_address + (*bo)->bo->size) {
2576 return (struct gen_batch_decode_bo) {
2577 .addr = bo_address,
2578 .size = (*bo)->bo->size,
2579 .map = (*bo)->bo->map,
2580 };
2581 }
2582 }
2583
2584 return (struct gen_batch_decode_bo) { };
2585 }
2586
2587 struct gen_aux_map_buffer {
2588 struct gen_buffer base;
2589 struct anv_state state;
2590 };
2591
2592 static struct gen_buffer *
2593 gen_aux_map_buffer_alloc(void *driver_ctx, uint32_t size)
2594 {
2595 struct gen_aux_map_buffer *buf = malloc(sizeof(struct gen_aux_map_buffer));
2596 if (!buf)
2597 return NULL;
2598
2599 struct anv_device *device = (struct anv_device*)driver_ctx;
2600 assert(device->physical->supports_48bit_addresses &&
2601 device->physical->use_softpin);
2602
2603 struct anv_state_pool *pool = &device->dynamic_state_pool;
2604 buf->state = anv_state_pool_alloc(pool, size, size);
2605
2606 buf->base.gpu = pool->block_pool.bo->offset + buf->state.offset;
2607 buf->base.gpu_end = buf->base.gpu + buf->state.alloc_size;
2608 buf->base.map = buf->state.map;
2609 buf->base.driver_bo = &buf->state;
2610 return &buf->base;
2611 }
2612
2613 static void
2614 gen_aux_map_buffer_free(void *driver_ctx, struct gen_buffer *buffer)
2615 {
2616 struct gen_aux_map_buffer *buf = (struct gen_aux_map_buffer*)buffer;
2617 struct anv_device *device = (struct anv_device*)driver_ctx;
2618 struct anv_state_pool *pool = &device->dynamic_state_pool;
2619 anv_state_pool_free(pool, buf->state);
2620 free(buf);
2621 }
2622
2623 static struct gen_mapped_pinned_buffer_alloc aux_map_allocator = {
2624 .alloc = gen_aux_map_buffer_alloc,
2625 .free = gen_aux_map_buffer_free,
2626 };
2627
2628 static VkResult
2629 check_physical_device_features(VkPhysicalDevice physicalDevice,
2630 const VkPhysicalDeviceFeatures *features)
2631 {
2632 VkPhysicalDeviceFeatures supported_features;
2633 anv_GetPhysicalDeviceFeatures(physicalDevice, &supported_features);
2634 VkBool32 *supported_feature = (VkBool32 *)&supported_features;
2635 VkBool32 *enabled_feature = (VkBool32 *)features;
2636 unsigned num_features = sizeof(VkPhysicalDeviceFeatures) / sizeof(VkBool32);
2637 for (uint32_t i = 0; i < num_features; i++) {
2638 if (enabled_feature[i] && !supported_feature[i])
2639 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
2640 }
2641
2642 return VK_SUCCESS;
2643 }
2644
2645 VkResult anv_CreateDevice(
2646 VkPhysicalDevice physicalDevice,
2647 const VkDeviceCreateInfo* pCreateInfo,
2648 const VkAllocationCallbacks* pAllocator,
2649 VkDevice* pDevice)
2650 {
2651 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
2652 VkResult result;
2653 struct anv_device *device;
2654
2655 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO);
2656
2657 struct anv_device_extension_table enabled_extensions = { };
2658 for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
2659 int idx;
2660 for (idx = 0; idx < ANV_DEVICE_EXTENSION_COUNT; idx++) {
2661 if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
2662 anv_device_extensions[idx].extensionName) == 0)
2663 break;
2664 }
2665
2666 if (idx >= ANV_DEVICE_EXTENSION_COUNT)
2667 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
2668
2669 if (!physical_device->supported_extensions.extensions[idx])
2670 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
2671
2672 enabled_extensions.extensions[idx] = true;
2673 }
2674
2675 /* Check enabled features */
2676 bool robust_buffer_access = false;
2677 if (pCreateInfo->pEnabledFeatures) {
2678 result = check_physical_device_features(physicalDevice,
2679 pCreateInfo->pEnabledFeatures);
2680 if (result != VK_SUCCESS)
2681 return result;
2682
2683 if (pCreateInfo->pEnabledFeatures->robustBufferAccess)
2684 robust_buffer_access = true;
2685 }
2686
2687 vk_foreach_struct_const(ext, pCreateInfo->pNext) {
2688 switch (ext->sType) {
2689 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2: {
2690 const VkPhysicalDeviceFeatures2 *features = (const void *)ext;
2691 result = check_physical_device_features(physicalDevice,
2692 &features->features);
2693 if (result != VK_SUCCESS)
2694 return result;
2695
2696 if (features->features.robustBufferAccess)
2697 robust_buffer_access = true;
2698 break;
2699 }
2700
2701 default:
2702 /* Don't warn */
2703 break;
2704 }
2705 }
2706
2707 /* Check requested queues and fail if we are requested to create any
2708 * queues with flags we don't support.
2709 */
2710 assert(pCreateInfo->queueCreateInfoCount > 0);
2711 for (uint32_t i = 0; i < pCreateInfo->queueCreateInfoCount; i++) {
2712 if (pCreateInfo->pQueueCreateInfos[i].flags != 0)
2713 return vk_error(VK_ERROR_INITIALIZATION_FAILED);
2714 }
2715
2716 /* Check if client specified queue priority. */
2717 const VkDeviceQueueGlobalPriorityCreateInfoEXT *queue_priority =
2718 vk_find_struct_const(pCreateInfo->pQueueCreateInfos[0].pNext,
2719 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT);
2720
2721 VkQueueGlobalPriorityEXT priority =
2722 queue_priority ? queue_priority->globalPriority :
2723 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT;
2724
2725 device = vk_alloc2(&physical_device->instance->alloc, pAllocator,
2726 sizeof(*device), 8,
2727 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
2728 if (!device)
2729 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2730
2731 if (INTEL_DEBUG & DEBUG_BATCH) {
2732 const unsigned decode_flags =
2733 GEN_BATCH_DECODE_FULL |
2734 ((INTEL_DEBUG & DEBUG_COLOR) ? GEN_BATCH_DECODE_IN_COLOR : 0) |
2735 GEN_BATCH_DECODE_OFFSETS |
2736 GEN_BATCH_DECODE_FLOATS;
2737
2738 gen_batch_decode_ctx_init(&device->decoder_ctx,
2739 &physical_device->info,
2740 stderr, decode_flags, NULL,
2741 decode_get_bo, NULL, device);
2742 }
2743
2744 device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
2745 device->physical = physical_device;
2746 device->no_hw = physical_device->no_hw;
2747 device->_lost = false;
2748
2749 if (pAllocator)
2750 device->alloc = *pAllocator;
2751 else
2752 device->alloc = physical_device->instance->alloc;
2753
2754 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
2755 device->fd = open(physical_device->path, O_RDWR | O_CLOEXEC);
2756 if (device->fd == -1) {
2757 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2758 goto fail_device;
2759 }
2760
2761 device->context_id = anv_gem_create_context(device);
2762 if (device->context_id == -1) {
2763 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2764 goto fail_fd;
2765 }
2766
2767 result = anv_queue_init(device, &device->queue);
2768 if (result != VK_SUCCESS)
2769 goto fail_context_id;
2770
2771 if (physical_device->use_softpin) {
2772 if (pthread_mutex_init(&device->vma_mutex, NULL) != 0) {
2773 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2774 goto fail_queue;
2775 }
2776
2777 /* keep the page with address zero out of the allocator */
2778 util_vma_heap_init(&device->vma_lo,
2779 LOW_HEAP_MIN_ADDRESS, LOW_HEAP_SIZE);
2780
2781 util_vma_heap_init(&device->vma_cva, CLIENT_VISIBLE_HEAP_MIN_ADDRESS,
2782 CLIENT_VISIBLE_HEAP_SIZE);
2783
2784 /* Leave the last 4GiB out of the high vma range, so that no state
2785 * base address + size can overflow 48 bits. For more information see
2786 * the comment about Wa32bitGeneralStateOffset in anv_allocator.c
2787 */
2788 util_vma_heap_init(&device->vma_hi, HIGH_HEAP_MIN_ADDRESS,
2789 physical_device->gtt_size - (1ull << 32) -
2790 HIGH_HEAP_MIN_ADDRESS);
2791 }
2792
2793 list_inithead(&device->memory_objects);
2794
2795 /* As per spec, the driver implementation may deny requests to acquire
2796 * a priority above the default priority (MEDIUM) if the caller does not
2797 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
2798 * is returned.
2799 */
2800 if (physical_device->has_context_priority) {
2801 int err = anv_gem_set_context_param(device->fd, device->context_id,
2802 I915_CONTEXT_PARAM_PRIORITY,
2803 vk_priority_to_gen(priority));
2804 if (err != 0 && priority > VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT) {
2805 result = vk_error(VK_ERROR_NOT_PERMITTED_EXT);
2806 goto fail_vmas;
2807 }
2808 }
2809
2810 device->info = physical_device->info;
2811 device->isl_dev = physical_device->isl_dev;
2812
2813 /* On Broadwell and later, we can use batch chaining to more efficiently
2814 * implement growing command buffers. Prior to Haswell, the kernel
2815 * command parser gets in the way and we have to fall back to growing
2816 * the batch.
2817 */
2818 device->can_chain_batches = device->info.gen >= 8;
2819
2820 device->robust_buffer_access = robust_buffer_access;
2821 device->enabled_extensions = enabled_extensions;
2822
2823 anv_device_init_dispatch(device);
2824
2825 if (pthread_mutex_init(&device->mutex, NULL) != 0) {
2826 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2827 goto fail_queue;
2828 }
2829
2830 pthread_condattr_t condattr;
2831 if (pthread_condattr_init(&condattr) != 0) {
2832 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2833 goto fail_mutex;
2834 }
2835 if (pthread_condattr_setclock(&condattr, CLOCK_MONOTONIC) != 0) {
2836 pthread_condattr_destroy(&condattr);
2837 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2838 goto fail_mutex;
2839 }
2840 if (pthread_cond_init(&device->queue_submit, &condattr) != 0) {
2841 pthread_condattr_destroy(&condattr);
2842 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2843 goto fail_mutex;
2844 }
2845 pthread_condattr_destroy(&condattr);
2846
2847 result = anv_bo_cache_init(&device->bo_cache);
2848 if (result != VK_SUCCESS)
2849 goto fail_queue_cond;
2850
2851 anv_bo_pool_init(&device->batch_bo_pool, device);
2852
2853 result = anv_state_pool_init(&device->dynamic_state_pool, device,
2854 DYNAMIC_STATE_POOL_MIN_ADDRESS, 16384);
2855 if (result != VK_SUCCESS)
2856 goto fail_batch_bo_pool;
2857
2858 result = anv_state_pool_init(&device->instruction_state_pool, device,
2859 INSTRUCTION_STATE_POOL_MIN_ADDRESS, 16384);
2860 if (result != VK_SUCCESS)
2861 goto fail_dynamic_state_pool;
2862
2863 result = anv_state_pool_init(&device->surface_state_pool, device,
2864 SURFACE_STATE_POOL_MIN_ADDRESS, 4096);
2865 if (result != VK_SUCCESS)
2866 goto fail_instruction_state_pool;
2867
2868 if (physical_device->use_softpin) {
2869 result = anv_state_pool_init(&device->binding_table_pool, device,
2870 BINDING_TABLE_POOL_MIN_ADDRESS, 4096);
2871 if (result != VK_SUCCESS)
2872 goto fail_surface_state_pool;
2873 }
2874
2875 if (device->info.gen >= 12) {
2876 device->aux_map_ctx = gen_aux_map_init(device, &aux_map_allocator,
2877 &physical_device->info);
2878 if (!device->aux_map_ctx)
2879 goto fail_binding_table_pool;
2880 }
2881
2882 result = anv_device_alloc_bo(device, 4096, 0 /* flags */,
2883 0 /* explicit_address */,
2884 &device->workaround_bo);
2885 if (result != VK_SUCCESS)
2886 goto fail_surface_aux_map_pool;
2887
2888 result = anv_device_init_trivial_batch(device);
2889 if (result != VK_SUCCESS)
2890 goto fail_workaround_bo;
2891
2892 if (device->info.gen >= 10) {
2893 result = anv_device_init_hiz_clear_value_bo(device);
2894 if (result != VK_SUCCESS)
2895 goto fail_trivial_batch_bo;
2896 }
2897
2898 anv_scratch_pool_init(device, &device->scratch_pool);
2899
2900 switch (device->info.gen) {
2901 case 7:
2902 if (!device->info.is_haswell)
2903 result = gen7_init_device_state(device);
2904 else
2905 result = gen75_init_device_state(device);
2906 break;
2907 case 8:
2908 result = gen8_init_device_state(device);
2909 break;
2910 case 9:
2911 result = gen9_init_device_state(device);
2912 break;
2913 case 10:
2914 result = gen10_init_device_state(device);
2915 break;
2916 case 11:
2917 result = gen11_init_device_state(device);
2918 break;
2919 case 12:
2920 result = gen12_init_device_state(device);
2921 break;
2922 default:
2923 /* Shouldn't get here as we don't create physical devices for any other
2924 * gens. */
2925 unreachable("unhandled gen");
2926 }
2927 if (result != VK_SUCCESS)
2928 goto fail_workaround_bo;
2929
2930 anv_pipeline_cache_init(&device->default_pipeline_cache, device, true);
2931
2932 anv_device_init_blorp(device);
2933
2934 anv_device_init_border_colors(device);
2935
2936 anv_device_perf_init(device);
2937
2938 *pDevice = anv_device_to_handle(device);
2939
2940 return VK_SUCCESS;
2941
2942 fail_workaround_bo:
2943 anv_scratch_pool_finish(device, &device->scratch_pool);
2944 if (device->info.gen >= 10)
2945 anv_device_release_bo(device, device->hiz_clear_bo);
2946 anv_device_release_bo(device, device->workaround_bo);
2947 fail_trivial_batch_bo:
2948 anv_device_release_bo(device, device->trivial_batch_bo);
2949 fail_surface_aux_map_pool:
2950 if (device->info.gen >= 12) {
2951 gen_aux_map_finish(device->aux_map_ctx);
2952 device->aux_map_ctx = NULL;
2953 }
2954 fail_binding_table_pool:
2955 if (physical_device->use_softpin)
2956 anv_state_pool_finish(&device->binding_table_pool);
2957 fail_surface_state_pool:
2958 anv_state_pool_finish(&device->surface_state_pool);
2959 fail_instruction_state_pool:
2960 anv_state_pool_finish(&device->instruction_state_pool);
2961 fail_dynamic_state_pool:
2962 anv_state_pool_finish(&device->dynamic_state_pool);
2963 fail_batch_bo_pool:
2964 anv_bo_pool_finish(&device->batch_bo_pool);
2965 anv_bo_cache_finish(&device->bo_cache);
2966 fail_queue_cond:
2967 pthread_cond_destroy(&device->queue_submit);
2968 fail_mutex:
2969 pthread_mutex_destroy(&device->mutex);
2970 fail_vmas:
2971 if (physical_device->use_softpin) {
2972 util_vma_heap_finish(&device->vma_hi);
2973 util_vma_heap_finish(&device->vma_cva);
2974 util_vma_heap_finish(&device->vma_lo);
2975 }
2976 fail_queue:
2977 anv_queue_finish(&device->queue);
2978 fail_context_id:
2979 anv_gem_destroy_context(device, device->context_id);
2980 fail_fd:
2981 close(device->fd);
2982 fail_device:
2983 vk_free(&device->alloc, device);
2984
2985 return result;
2986 }
2987
2988 void anv_DestroyDevice(
2989 VkDevice _device,
2990 const VkAllocationCallbacks* pAllocator)
2991 {
2992 ANV_FROM_HANDLE(anv_device, device, _device);
2993
2994 if (!device)
2995 return;
2996
2997 anv_device_finish_blorp(device);
2998
2999 anv_pipeline_cache_finish(&device->default_pipeline_cache);
3000
3001 anv_queue_finish(&device->queue);
3002
3003 #ifdef HAVE_VALGRIND
3004 /* We only need to free these to prevent valgrind errors. The backing
3005 * BO will go away in a couple of lines so we don't actually leak.
3006 */
3007 anv_state_pool_free(&device->dynamic_state_pool, device->border_colors);
3008 anv_state_pool_free(&device->dynamic_state_pool, device->slice_hash);
3009 #endif
3010
3011 anv_scratch_pool_finish(device, &device->scratch_pool);
3012
3013 anv_device_release_bo(device, device->workaround_bo);
3014 anv_device_release_bo(device, device->trivial_batch_bo);
3015 if (device->info.gen >= 10)
3016 anv_device_release_bo(device, device->hiz_clear_bo);
3017
3018 if (device->info.gen >= 12) {
3019 gen_aux_map_finish(device->aux_map_ctx);
3020 device->aux_map_ctx = NULL;
3021 }
3022
3023 if (device->physical->use_softpin)
3024 anv_state_pool_finish(&device->binding_table_pool);
3025 anv_state_pool_finish(&device->surface_state_pool);
3026 anv_state_pool_finish(&device->instruction_state_pool);
3027 anv_state_pool_finish(&device->dynamic_state_pool);
3028
3029 anv_bo_pool_finish(&device->batch_bo_pool);
3030
3031 anv_bo_cache_finish(&device->bo_cache);
3032
3033 if (device->physical->use_softpin) {
3034 util_vma_heap_finish(&device->vma_hi);
3035 util_vma_heap_finish(&device->vma_cva);
3036 util_vma_heap_finish(&device->vma_lo);
3037 }
3038
3039 pthread_cond_destroy(&device->queue_submit);
3040 pthread_mutex_destroy(&device->mutex);
3041
3042 anv_gem_destroy_context(device, device->context_id);
3043
3044 if (INTEL_DEBUG & DEBUG_BATCH)
3045 gen_batch_decode_ctx_finish(&device->decoder_ctx);
3046
3047 close(device->fd);
3048
3049 vk_free(&device->alloc, device);
3050 }
3051
3052 VkResult anv_EnumerateInstanceLayerProperties(
3053 uint32_t* pPropertyCount,
3054 VkLayerProperties* pProperties)
3055 {
3056 if (pProperties == NULL) {
3057 *pPropertyCount = 0;
3058 return VK_SUCCESS;
3059 }
3060
3061 /* None supported at this time */
3062 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
3063 }
3064
3065 VkResult anv_EnumerateDeviceLayerProperties(
3066 VkPhysicalDevice physicalDevice,
3067 uint32_t* pPropertyCount,
3068 VkLayerProperties* pProperties)
3069 {
3070 if (pProperties == NULL) {
3071 *pPropertyCount = 0;
3072 return VK_SUCCESS;
3073 }
3074
3075 /* None supported at this time */
3076 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
3077 }
3078
3079 void anv_GetDeviceQueue(
3080 VkDevice _device,
3081 uint32_t queueNodeIndex,
3082 uint32_t queueIndex,
3083 VkQueue* pQueue)
3084 {
3085 const VkDeviceQueueInfo2 info = {
3086 .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_INFO_2,
3087 .pNext = NULL,
3088 .flags = 0,
3089 .queueFamilyIndex = queueNodeIndex,
3090 .queueIndex = queueIndex,
3091 };
3092
3093 anv_GetDeviceQueue2(_device, &info, pQueue);
3094 }
3095
3096 void anv_GetDeviceQueue2(
3097 VkDevice _device,
3098 const VkDeviceQueueInfo2* pQueueInfo,
3099 VkQueue* pQueue)
3100 {
3101 ANV_FROM_HANDLE(anv_device, device, _device);
3102
3103 assert(pQueueInfo->queueIndex == 0);
3104
3105 if (pQueueInfo->flags == device->queue.flags)
3106 *pQueue = anv_queue_to_handle(&device->queue);
3107 else
3108 *pQueue = NULL;
3109 }
3110
3111 VkResult
3112 _anv_device_set_lost(struct anv_device *device,
3113 const char *file, int line,
3114 const char *msg, ...)
3115 {
3116 VkResult err;
3117 va_list ap;
3118
3119 p_atomic_inc(&device->_lost);
3120
3121 va_start(ap, msg);
3122 err = __vk_errorv(device->physical->instance, device,
3123 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT,
3124 VK_ERROR_DEVICE_LOST, file, line, msg, ap);
3125 va_end(ap);
3126
3127 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
3128 abort();
3129
3130 return err;
3131 }
3132
3133 VkResult
3134 _anv_queue_set_lost(struct anv_queue *queue,
3135 const char *file, int line,
3136 const char *msg, ...)
3137 {
3138 VkResult err;
3139 va_list ap;
3140
3141 p_atomic_inc(&queue->device->_lost);
3142
3143 va_start(ap, msg);
3144 err = __vk_errorv(queue->device->physical->instance, queue->device,
3145 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT,
3146 VK_ERROR_DEVICE_LOST, file, line, msg, ap);
3147 va_end(ap);
3148
3149 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
3150 abort();
3151
3152 return err;
3153 }
3154
3155 VkResult
3156 anv_device_query_status(struct anv_device *device)
3157 {
3158 /* This isn't likely as most of the callers of this function already check
3159 * for it. However, it doesn't hurt to check and it potentially lets us
3160 * avoid an ioctl.
3161 */
3162 if (anv_device_is_lost(device))
3163 return VK_ERROR_DEVICE_LOST;
3164
3165 uint32_t active, pending;
3166 int ret = anv_gem_gpu_get_reset_stats(device, &active, &pending);
3167 if (ret == -1) {
3168 /* We don't know the real error. */
3169 return anv_device_set_lost(device, "get_reset_stats failed: %m");
3170 }
3171
3172 if (active) {
3173 return anv_device_set_lost(device, "GPU hung on one of our command buffers");
3174 } else if (pending) {
3175 return anv_device_set_lost(device, "GPU hung with commands in-flight");
3176 }
3177
3178 return VK_SUCCESS;
3179 }
3180
3181 VkResult
3182 anv_device_bo_busy(struct anv_device *device, struct anv_bo *bo)
3183 {
3184 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
3185 * Other usages of the BO (such as on different hardware) will not be
3186 * flagged as "busy" by this ioctl. Use with care.
3187 */
3188 int ret = anv_gem_busy(device, bo->gem_handle);
3189 if (ret == 1) {
3190 return VK_NOT_READY;
3191 } else if (ret == -1) {
3192 /* We don't know the real error. */
3193 return anv_device_set_lost(device, "gem wait failed: %m");
3194 }
3195
3196 /* Query for device status after the busy call. If the BO we're checking
3197 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
3198 * client because it clearly doesn't have valid data. Yes, this most
3199 * likely means an ioctl, but we just did an ioctl to query the busy status
3200 * so it's no great loss.
3201 */
3202 return anv_device_query_status(device);
3203 }
3204
3205 VkResult
3206 anv_device_wait(struct anv_device *device, struct anv_bo *bo,
3207 int64_t timeout)
3208 {
3209 int ret = anv_gem_wait(device, bo->gem_handle, &timeout);
3210 if (ret == -1 && errno == ETIME) {
3211 return VK_TIMEOUT;
3212 } else if (ret == -1) {
3213 /* We don't know the real error. */
3214 return anv_device_set_lost(device, "gem wait failed: %m");
3215 }
3216
3217 /* Query for device status after the wait. If the BO we're waiting on got
3218 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
3219 * because it clearly doesn't have valid data. Yes, this most likely means
3220 * an ioctl, but we just did an ioctl to wait so it's no great loss.
3221 */
3222 return anv_device_query_status(device);
3223 }
3224
3225 VkResult anv_DeviceWaitIdle(
3226 VkDevice _device)
3227 {
3228 ANV_FROM_HANDLE(anv_device, device, _device);
3229
3230 if (anv_device_is_lost(device))
3231 return VK_ERROR_DEVICE_LOST;
3232
3233 return anv_queue_submit_simple_batch(&device->queue, NULL);
3234 }
3235
3236 uint64_t
3237 anv_vma_alloc(struct anv_device *device,
3238 uint64_t size, uint64_t align,
3239 enum anv_bo_alloc_flags alloc_flags,
3240 uint64_t client_address)
3241 {
3242 pthread_mutex_lock(&device->vma_mutex);
3243
3244 uint64_t addr = 0;
3245
3246 if (alloc_flags & ANV_BO_ALLOC_CLIENT_VISIBLE_ADDRESS) {
3247 if (client_address) {
3248 if (util_vma_heap_alloc_addr(&device->vma_cva,
3249 client_address, size)) {
3250 addr = client_address;
3251 }
3252 } else {
3253 addr = util_vma_heap_alloc(&device->vma_cva, size, align);
3254 }
3255 /* We don't want to fall back to other heaps */
3256 goto done;
3257 }
3258
3259 assert(client_address == 0);
3260
3261 if (!(alloc_flags & ANV_BO_ALLOC_32BIT_ADDRESS))
3262 addr = util_vma_heap_alloc(&device->vma_hi, size, align);
3263
3264 if (addr == 0)
3265 addr = util_vma_heap_alloc(&device->vma_lo, size, align);
3266
3267 done:
3268 pthread_mutex_unlock(&device->vma_mutex);
3269
3270 assert(addr == gen_48b_address(addr));
3271 return gen_canonical_address(addr);
3272 }
3273
3274 void
3275 anv_vma_free(struct anv_device *device,
3276 uint64_t address, uint64_t size)
3277 {
3278 const uint64_t addr_48b = gen_48b_address(address);
3279
3280 pthread_mutex_lock(&device->vma_mutex);
3281
3282 if (addr_48b >= LOW_HEAP_MIN_ADDRESS &&
3283 addr_48b <= LOW_HEAP_MAX_ADDRESS) {
3284 util_vma_heap_free(&device->vma_lo, addr_48b, size);
3285 } else if (addr_48b >= CLIENT_VISIBLE_HEAP_MIN_ADDRESS &&
3286 addr_48b <= CLIENT_VISIBLE_HEAP_MAX_ADDRESS) {
3287 util_vma_heap_free(&device->vma_cva, addr_48b, size);
3288 } else {
3289 assert(addr_48b >= HIGH_HEAP_MIN_ADDRESS);
3290 util_vma_heap_free(&device->vma_hi, addr_48b, size);
3291 }
3292
3293 pthread_mutex_unlock(&device->vma_mutex);
3294 }
3295
3296 VkResult anv_AllocateMemory(
3297 VkDevice _device,
3298 const VkMemoryAllocateInfo* pAllocateInfo,
3299 const VkAllocationCallbacks* pAllocator,
3300 VkDeviceMemory* pMem)
3301 {
3302 ANV_FROM_HANDLE(anv_device, device, _device);
3303 struct anv_physical_device *pdevice = device->physical;
3304 struct anv_device_memory *mem;
3305 VkResult result = VK_SUCCESS;
3306
3307 assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
3308
3309 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
3310 assert(pAllocateInfo->allocationSize > 0);
3311
3312 VkDeviceSize aligned_alloc_size =
3313 align_u64(pAllocateInfo->allocationSize, 4096);
3314
3315 if (aligned_alloc_size > MAX_MEMORY_ALLOCATION_SIZE)
3316 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
3317
3318 assert(pAllocateInfo->memoryTypeIndex < pdevice->memory.type_count);
3319 struct anv_memory_type *mem_type =
3320 &pdevice->memory.types[pAllocateInfo->memoryTypeIndex];
3321 assert(mem_type->heapIndex < pdevice->memory.heap_count);
3322 struct anv_memory_heap *mem_heap =
3323 &pdevice->memory.heaps[mem_type->heapIndex];
3324
3325 uint64_t mem_heap_used = p_atomic_read(&mem_heap->used);
3326 if (mem_heap_used + aligned_alloc_size > mem_heap->size)
3327 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
3328
3329 mem = vk_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
3330 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
3331 if (mem == NULL)
3332 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
3333
3334 assert(pAllocateInfo->memoryTypeIndex < pdevice->memory.type_count);
3335 mem->type = mem_type;
3336 mem->map = NULL;
3337 mem->map_size = 0;
3338 mem->ahw = NULL;
3339 mem->host_ptr = NULL;
3340
3341 enum anv_bo_alloc_flags alloc_flags = 0;
3342
3343 const VkExportMemoryAllocateInfo *export_info = NULL;
3344 const VkImportAndroidHardwareBufferInfoANDROID *ahw_import_info = NULL;
3345 const VkImportMemoryFdInfoKHR *fd_info = NULL;
3346 const VkImportMemoryHostPointerInfoEXT *host_ptr_info = NULL;
3347 const VkMemoryDedicatedAllocateInfo *dedicated_info = NULL;
3348 VkMemoryAllocateFlags vk_flags = 0;
3349 uint64_t client_address = 0;
3350
3351 vk_foreach_struct_const(ext, pAllocateInfo->pNext) {
3352 switch (ext->sType) {
3353 case VK_STRUCTURE_TYPE_EXPORT_MEMORY_ALLOCATE_INFO:
3354 export_info = (void *)ext;
3355 break;
3356
3357 case VK_STRUCTURE_TYPE_IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID:
3358 ahw_import_info = (void *)ext;
3359 break;
3360
3361 case VK_STRUCTURE_TYPE_IMPORT_MEMORY_FD_INFO_KHR:
3362 fd_info = (void *)ext;
3363 break;
3364
3365 case VK_STRUCTURE_TYPE_IMPORT_MEMORY_HOST_POINTER_INFO_EXT:
3366 host_ptr_info = (void *)ext;
3367 break;
3368
3369 case VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_FLAGS_INFO: {
3370 const VkMemoryAllocateFlagsInfo *flags_info = (void *)ext;
3371 vk_flags = flags_info->flags;
3372 break;
3373 }
3374
3375 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_ALLOCATE_INFO:
3376 dedicated_info = (void *)ext;
3377 break;
3378
3379 case VK_STRUCTURE_TYPE_MEMORY_OPAQUE_CAPTURE_ADDRESS_ALLOCATE_INFO_KHR: {
3380 const VkMemoryOpaqueCaptureAddressAllocateInfoKHR *addr_info =
3381 (const VkMemoryOpaqueCaptureAddressAllocateInfoKHR *)ext;
3382 client_address = addr_info->opaqueCaptureAddress;
3383 break;
3384 }
3385
3386 default:
3387 anv_debug_ignored_stype(ext->sType);
3388 break;
3389 }
3390 }
3391
3392 /* By default, we want all VkDeviceMemory objects to support CCS */
3393 if (device->physical->has_implicit_ccs)
3394 alloc_flags |= ANV_BO_ALLOC_IMPLICIT_CCS;
3395
3396 if (vk_flags & VK_MEMORY_ALLOCATE_DEVICE_ADDRESS_BIT_KHR)
3397 alloc_flags |= ANV_BO_ALLOC_CLIENT_VISIBLE_ADDRESS;
3398
3399 if ((export_info && export_info->handleTypes) ||
3400 (fd_info && fd_info->handleType) ||
3401 (host_ptr_info && host_ptr_info->handleType)) {
3402 /* Anything imported or exported is EXTERNAL */
3403 alloc_flags |= ANV_BO_ALLOC_EXTERNAL;
3404
3405 /* We can't have implicit CCS on external memory with an AUX-table.
3406 * Doing so would require us to sync the aux tables across processes
3407 * which is impractical.
3408 */
3409 if (device->info.has_aux_map)
3410 alloc_flags &= ~ANV_BO_ALLOC_IMPLICIT_CCS;
3411 }
3412
3413 /* Check if we need to support Android HW buffer export. If so,
3414 * create AHardwareBuffer and import memory from it.
3415 */
3416 bool android_export = false;
3417 if (export_info && export_info->handleTypes &
3418 VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID)
3419 android_export = true;
3420
3421 if (ahw_import_info) {
3422 result = anv_import_ahw_memory(_device, mem, ahw_import_info);
3423 if (result != VK_SUCCESS)
3424 goto fail;
3425
3426 goto success;
3427 } else if (android_export) {
3428 result = anv_create_ahw_memory(_device, mem, pAllocateInfo);
3429 if (result != VK_SUCCESS)
3430 goto fail;
3431
3432 const VkImportAndroidHardwareBufferInfoANDROID import_info = {
3433 .buffer = mem->ahw,
3434 };
3435 result = anv_import_ahw_memory(_device, mem, &import_info);
3436 if (result != VK_SUCCESS)
3437 goto fail;
3438
3439 goto success;
3440 }
3441
3442 /* The Vulkan spec permits handleType to be 0, in which case the struct is
3443 * ignored.
3444 */
3445 if (fd_info && fd_info->handleType) {
3446 /* At the moment, we support only the below handle types. */
3447 assert(fd_info->handleType ==
3448 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
3449 fd_info->handleType ==
3450 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
3451
3452 result = anv_device_import_bo(device, fd_info->fd, alloc_flags,
3453 client_address, &mem->bo);
3454 if (result != VK_SUCCESS)
3455 goto fail;
3456
3457 /* For security purposes, we reject importing the bo if it's smaller
3458 * than the requested allocation size. This prevents a malicious client
3459 * from passing a buffer to a trusted client, lying about the size, and
3460 * telling the trusted client to try and texture from an image that goes
3461 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
3462 * in the trusted client. The trusted client can protect itself against
3463 * this sort of attack but only if it can trust the buffer size.
3464 */
3465 if (mem->bo->size < aligned_alloc_size) {
3466 result = vk_errorf(device, device, VK_ERROR_INVALID_EXTERNAL_HANDLE,
3467 "aligned allocationSize too large for "
3468 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT: "
3469 "%"PRIu64"B > %"PRIu64"B",
3470 aligned_alloc_size, mem->bo->size);
3471 anv_device_release_bo(device, mem->bo);
3472 goto fail;
3473 }
3474
3475 /* From the Vulkan spec:
3476 *
3477 * "Importing memory from a file descriptor transfers ownership of
3478 * the file descriptor from the application to the Vulkan
3479 * implementation. The application must not perform any operations on
3480 * the file descriptor after a successful import."
3481 *
3482 * If the import fails, we leave the file descriptor open.
3483 */
3484 close(fd_info->fd);
3485 goto success;
3486 }
3487
3488 if (host_ptr_info && host_ptr_info->handleType) {
3489 if (host_ptr_info->handleType ==
3490 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_MAPPED_FOREIGN_MEMORY_BIT_EXT) {
3491 result = vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE);
3492 goto fail;
3493 }
3494
3495 assert(host_ptr_info->handleType ==
3496 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT);
3497
3498 result = anv_device_import_bo_from_host_ptr(device,
3499 host_ptr_info->pHostPointer,
3500 pAllocateInfo->allocationSize,
3501 alloc_flags,
3502 client_address,
3503 &mem->bo);
3504 if (result != VK_SUCCESS)
3505 goto fail;
3506
3507 mem->host_ptr = host_ptr_info->pHostPointer;
3508 goto success;
3509 }
3510
3511 /* Regular allocate (not importing memory). */
3512
3513 result = anv_device_alloc_bo(device, pAllocateInfo->allocationSize,
3514 alloc_flags, client_address, &mem->bo);
3515 if (result != VK_SUCCESS)
3516 goto fail;
3517
3518 if (dedicated_info && dedicated_info->image != VK_NULL_HANDLE) {
3519 ANV_FROM_HANDLE(anv_image, image, dedicated_info->image);
3520
3521 /* Some legacy (non-modifiers) consumers need the tiling to be set on
3522 * the BO. In this case, we have a dedicated allocation.
3523 */
3524 if (image->needs_set_tiling) {
3525 const uint32_t i915_tiling =
3526 isl_tiling_to_i915_tiling(image->planes[0].surface.isl.tiling);
3527 int ret = anv_gem_set_tiling(device, mem->bo->gem_handle,
3528 image->planes[0].surface.isl.row_pitch_B,
3529 i915_tiling);
3530 if (ret) {
3531 anv_device_release_bo(device, mem->bo);
3532 result = vk_errorf(device, device, VK_ERROR_OUT_OF_DEVICE_MEMORY,
3533 "failed to set BO tiling: %m");
3534 goto fail;
3535 }
3536 }
3537 }
3538
3539 success:
3540 mem_heap_used = p_atomic_add_return(&mem_heap->used, mem->bo->size);
3541 if (mem_heap_used > mem_heap->size) {
3542 p_atomic_add(&mem_heap->used, -mem->bo->size);
3543 anv_device_release_bo(device, mem->bo);
3544 result = vk_errorf(device, device, VK_ERROR_OUT_OF_DEVICE_MEMORY,
3545 "Out of heap memory");
3546 goto fail;
3547 }
3548
3549 pthread_mutex_lock(&device->mutex);
3550 list_addtail(&mem->link, &device->memory_objects);
3551 pthread_mutex_unlock(&device->mutex);
3552
3553 *pMem = anv_device_memory_to_handle(mem);
3554
3555 return VK_SUCCESS;
3556
3557 fail:
3558 vk_free2(&device->alloc, pAllocator, mem);
3559
3560 return result;
3561 }
3562
3563 VkResult anv_GetMemoryFdKHR(
3564 VkDevice device_h,
3565 const VkMemoryGetFdInfoKHR* pGetFdInfo,
3566 int* pFd)
3567 {
3568 ANV_FROM_HANDLE(anv_device, dev, device_h);
3569 ANV_FROM_HANDLE(anv_device_memory, mem, pGetFdInfo->memory);
3570
3571 assert(pGetFdInfo->sType == VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR);
3572
3573 assert(pGetFdInfo->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
3574 pGetFdInfo->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
3575
3576 return anv_device_export_bo(dev, mem->bo, pFd);
3577 }
3578
3579 VkResult anv_GetMemoryFdPropertiesKHR(
3580 VkDevice _device,
3581 VkExternalMemoryHandleTypeFlagBits handleType,
3582 int fd,
3583 VkMemoryFdPropertiesKHR* pMemoryFdProperties)
3584 {
3585 ANV_FROM_HANDLE(anv_device, device, _device);
3586
3587 switch (handleType) {
3588 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT:
3589 /* dma-buf can be imported as any memory type */
3590 pMemoryFdProperties->memoryTypeBits =
3591 (1 << device->physical->memory.type_count) - 1;
3592 return VK_SUCCESS;
3593
3594 default:
3595 /* The valid usage section for this function says:
3596 *
3597 * "handleType must not be one of the handle types defined as
3598 * opaque."
3599 *
3600 * So opaque handle types fall into the default "unsupported" case.
3601 */
3602 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE);
3603 }
3604 }
3605
3606 VkResult anv_GetMemoryHostPointerPropertiesEXT(
3607 VkDevice _device,
3608 VkExternalMemoryHandleTypeFlagBits handleType,
3609 const void* pHostPointer,
3610 VkMemoryHostPointerPropertiesEXT* pMemoryHostPointerProperties)
3611 {
3612 ANV_FROM_HANDLE(anv_device, device, _device);
3613
3614 assert(pMemoryHostPointerProperties->sType ==
3615 VK_STRUCTURE_TYPE_MEMORY_HOST_POINTER_PROPERTIES_EXT);
3616
3617 switch (handleType) {
3618 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT:
3619 /* Host memory can be imported as any memory type. */
3620 pMemoryHostPointerProperties->memoryTypeBits =
3621 (1ull << device->physical->memory.type_count) - 1;
3622
3623 return VK_SUCCESS;
3624
3625 default:
3626 return VK_ERROR_INVALID_EXTERNAL_HANDLE;
3627 }
3628 }
3629
3630 void anv_FreeMemory(
3631 VkDevice _device,
3632 VkDeviceMemory _mem,
3633 const VkAllocationCallbacks* pAllocator)
3634 {
3635 ANV_FROM_HANDLE(anv_device, device, _device);
3636 ANV_FROM_HANDLE(anv_device_memory, mem, _mem);
3637
3638 if (mem == NULL)
3639 return;
3640
3641 pthread_mutex_lock(&device->mutex);
3642 list_del(&mem->link);
3643 pthread_mutex_unlock(&device->mutex);
3644
3645 if (mem->map)
3646 anv_UnmapMemory(_device, _mem);
3647
3648 p_atomic_add(&device->physical->memory.heaps[mem->type->heapIndex].used,
3649 -mem->bo->size);
3650
3651 anv_device_release_bo(device, mem->bo);
3652
3653 #if defined(ANDROID) && ANDROID_API_LEVEL >= 26
3654 if (mem->ahw)
3655 AHardwareBuffer_release(mem->ahw);
3656 #endif
3657
3658 vk_free2(&device->alloc, pAllocator, mem);
3659 }
3660
3661 VkResult anv_MapMemory(
3662 VkDevice _device,
3663 VkDeviceMemory _memory,
3664 VkDeviceSize offset,
3665 VkDeviceSize size,
3666 VkMemoryMapFlags flags,
3667 void** ppData)
3668 {
3669 ANV_FROM_HANDLE(anv_device, device, _device);
3670 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
3671
3672 if (mem == NULL) {
3673 *ppData = NULL;
3674 return VK_SUCCESS;
3675 }
3676
3677 if (mem->host_ptr) {
3678 *ppData = mem->host_ptr + offset;
3679 return VK_SUCCESS;
3680 }
3681
3682 if (size == VK_WHOLE_SIZE)
3683 size = mem->bo->size - offset;
3684
3685 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
3686 *
3687 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
3688 * assert(size != 0);
3689 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
3690 * equal to the size of the memory minus offset
3691 */
3692 assert(size > 0);
3693 assert(offset + size <= mem->bo->size);
3694
3695 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
3696 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
3697 * at a time is valid. We could just mmap up front and return an offset
3698 * pointer here, but that may exhaust virtual memory on 32 bit
3699 * userspace. */
3700
3701 uint32_t gem_flags = 0;
3702
3703 if (!device->info.has_llc &&
3704 (mem->type->propertyFlags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT))
3705 gem_flags |= I915_MMAP_WC;
3706
3707 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
3708 uint64_t map_offset;
3709 if (!device->physical->has_mmap_offset)
3710 map_offset = offset & ~4095ull;
3711 else
3712 map_offset = 0;
3713 assert(offset >= map_offset);
3714 uint64_t map_size = (offset + size) - map_offset;
3715
3716 /* Let's map whole pages */
3717 map_size = align_u64(map_size, 4096);
3718
3719 void *map = anv_gem_mmap(device, mem->bo->gem_handle,
3720 map_offset, map_size, gem_flags);
3721 if (map == MAP_FAILED)
3722 return vk_error(VK_ERROR_MEMORY_MAP_FAILED);
3723
3724 mem->map = map;
3725 mem->map_size = map_size;
3726
3727 *ppData = mem->map + (offset - map_offset);
3728
3729 return VK_SUCCESS;
3730 }
3731
3732 void anv_UnmapMemory(
3733 VkDevice _device,
3734 VkDeviceMemory _memory)
3735 {
3736 ANV_FROM_HANDLE(anv_device, device, _device);
3737 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
3738
3739 if (mem == NULL || mem->host_ptr)
3740 return;
3741
3742 anv_gem_munmap(device, mem->map, mem->map_size);
3743
3744 mem->map = NULL;
3745 mem->map_size = 0;
3746 }
3747
3748 static void
3749 clflush_mapped_ranges(struct anv_device *device,
3750 uint32_t count,
3751 const VkMappedMemoryRange *ranges)
3752 {
3753 for (uint32_t i = 0; i < count; i++) {
3754 ANV_FROM_HANDLE(anv_device_memory, mem, ranges[i].memory);
3755 if (ranges[i].offset >= mem->map_size)
3756 continue;
3757
3758 gen_clflush_range(mem->map + ranges[i].offset,
3759 MIN2(ranges[i].size, mem->map_size - ranges[i].offset));
3760 }
3761 }
3762
3763 VkResult anv_FlushMappedMemoryRanges(
3764 VkDevice _device,
3765 uint32_t memoryRangeCount,
3766 const VkMappedMemoryRange* pMemoryRanges)
3767 {
3768 ANV_FROM_HANDLE(anv_device, device, _device);
3769
3770 if (device->info.has_llc)
3771 return VK_SUCCESS;
3772
3773 /* Make sure the writes we're flushing have landed. */
3774 __builtin_ia32_mfence();
3775
3776 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
3777
3778 return VK_SUCCESS;
3779 }
3780
3781 VkResult anv_InvalidateMappedMemoryRanges(
3782 VkDevice _device,
3783 uint32_t memoryRangeCount,
3784 const VkMappedMemoryRange* pMemoryRanges)
3785 {
3786 ANV_FROM_HANDLE(anv_device, device, _device);
3787
3788 if (device->info.has_llc)
3789 return VK_SUCCESS;
3790
3791 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
3792
3793 /* Make sure no reads get moved up above the invalidate. */
3794 __builtin_ia32_mfence();
3795
3796 return VK_SUCCESS;
3797 }
3798
3799 void anv_GetBufferMemoryRequirements(
3800 VkDevice _device,
3801 VkBuffer _buffer,
3802 VkMemoryRequirements* pMemoryRequirements)
3803 {
3804 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
3805 ANV_FROM_HANDLE(anv_device, device, _device);
3806
3807 /* The Vulkan spec (git aaed022) says:
3808 *
3809 * memoryTypeBits is a bitfield and contains one bit set for every
3810 * supported memory type for the resource. The bit `1<<i` is set if and
3811 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3812 * structure for the physical device is supported.
3813 */
3814 uint32_t memory_types = (1ull << device->physical->memory.type_count) - 1;
3815
3816 /* Base alignment requirement of a cache line */
3817 uint32_t alignment = 16;
3818
3819 /* We need an alignment of 32 for pushing UBOs */
3820 if (buffer->usage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT)
3821 alignment = MAX2(alignment, 32);
3822
3823 pMemoryRequirements->size = buffer->size;
3824 pMemoryRequirements->alignment = alignment;
3825
3826 /* Storage and Uniform buffers should have their size aligned to
3827 * 32-bits to avoid boundary checks when last DWord is not complete.
3828 * This would ensure that not internal padding would be needed for
3829 * 16-bit types.
3830 */
3831 if (device->robust_buffer_access &&
3832 (buffer->usage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT ||
3833 buffer->usage & VK_BUFFER_USAGE_STORAGE_BUFFER_BIT))
3834 pMemoryRequirements->size = align_u64(buffer->size, 4);
3835
3836 pMemoryRequirements->memoryTypeBits = memory_types;
3837 }
3838
3839 void anv_GetBufferMemoryRequirements2(
3840 VkDevice _device,
3841 const VkBufferMemoryRequirementsInfo2* pInfo,
3842 VkMemoryRequirements2* pMemoryRequirements)
3843 {
3844 anv_GetBufferMemoryRequirements(_device, pInfo->buffer,
3845 &pMemoryRequirements->memoryRequirements);
3846
3847 vk_foreach_struct(ext, pMemoryRequirements->pNext) {
3848 switch (ext->sType) {
3849 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
3850 VkMemoryDedicatedRequirements *requirements = (void *)ext;
3851 requirements->prefersDedicatedAllocation = false;
3852 requirements->requiresDedicatedAllocation = false;
3853 break;
3854 }
3855
3856 default:
3857 anv_debug_ignored_stype(ext->sType);
3858 break;
3859 }
3860 }
3861 }
3862
3863 void anv_GetImageMemoryRequirements(
3864 VkDevice _device,
3865 VkImage _image,
3866 VkMemoryRequirements* pMemoryRequirements)
3867 {
3868 ANV_FROM_HANDLE(anv_image, image, _image);
3869 ANV_FROM_HANDLE(anv_device, device, _device);
3870
3871 /* The Vulkan spec (git aaed022) says:
3872 *
3873 * memoryTypeBits is a bitfield and contains one bit set for every
3874 * supported memory type for the resource. The bit `1<<i` is set if and
3875 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3876 * structure for the physical device is supported.
3877 *
3878 * All types are currently supported for images.
3879 */
3880 uint32_t memory_types = (1ull << device->physical->memory.type_count) - 1;
3881
3882 /* We must have image allocated or imported at this point. According to the
3883 * specification, external images must have been bound to memory before
3884 * calling GetImageMemoryRequirements.
3885 */
3886 assert(image->size > 0);
3887
3888 pMemoryRequirements->size = image->size;
3889 pMemoryRequirements->alignment = image->alignment;
3890 pMemoryRequirements->memoryTypeBits = memory_types;
3891 }
3892
3893 void anv_GetImageMemoryRequirements2(
3894 VkDevice _device,
3895 const VkImageMemoryRequirementsInfo2* pInfo,
3896 VkMemoryRequirements2* pMemoryRequirements)
3897 {
3898 ANV_FROM_HANDLE(anv_device, device, _device);
3899 ANV_FROM_HANDLE(anv_image, image, pInfo->image);
3900
3901 anv_GetImageMemoryRequirements(_device, pInfo->image,
3902 &pMemoryRequirements->memoryRequirements);
3903
3904 vk_foreach_struct_const(ext, pInfo->pNext) {
3905 switch (ext->sType) {
3906 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO: {
3907 const VkImagePlaneMemoryRequirementsInfo *plane_reqs =
3908 (const VkImagePlaneMemoryRequirementsInfo *) ext;
3909 uint32_t plane = anv_image_aspect_to_plane(image->aspects,
3910 plane_reqs->planeAspect);
3911
3912 assert(image->planes[plane].offset == 0);
3913
3914 /* The Vulkan spec (git aaed022) says:
3915 *
3916 * memoryTypeBits is a bitfield and contains one bit set for every
3917 * supported memory type for the resource. The bit `1<<i` is set
3918 * if and only if the memory type `i` in the
3919 * VkPhysicalDeviceMemoryProperties structure for the physical
3920 * device is supported.
3921 *
3922 * All types are currently supported for images.
3923 */
3924 pMemoryRequirements->memoryRequirements.memoryTypeBits =
3925 (1ull << device->physical->memory.type_count) - 1;
3926
3927 /* We must have image allocated or imported at this point. According to the
3928 * specification, external images must have been bound to memory before
3929 * calling GetImageMemoryRequirements.
3930 */
3931 assert(image->planes[plane].size > 0);
3932
3933 pMemoryRequirements->memoryRequirements.size = image->planes[plane].size;
3934 pMemoryRequirements->memoryRequirements.alignment =
3935 image->planes[plane].alignment;
3936 break;
3937 }
3938
3939 default:
3940 anv_debug_ignored_stype(ext->sType);
3941 break;
3942 }
3943 }
3944
3945 vk_foreach_struct(ext, pMemoryRequirements->pNext) {
3946 switch (ext->sType) {
3947 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
3948 VkMemoryDedicatedRequirements *requirements = (void *)ext;
3949 if (image->needs_set_tiling || image->external_format) {
3950 /* If we need to set the tiling for external consumers, we need a
3951 * dedicated allocation.
3952 *
3953 * See also anv_AllocateMemory.
3954 */
3955 requirements->prefersDedicatedAllocation = true;
3956 requirements->requiresDedicatedAllocation = true;
3957 } else {
3958 requirements->prefersDedicatedAllocation = false;
3959 requirements->requiresDedicatedAllocation = false;
3960 }
3961 break;
3962 }
3963
3964 default:
3965 anv_debug_ignored_stype(ext->sType);
3966 break;
3967 }
3968 }
3969 }
3970
3971 void anv_GetImageSparseMemoryRequirements(
3972 VkDevice device,
3973 VkImage image,
3974 uint32_t* pSparseMemoryRequirementCount,
3975 VkSparseImageMemoryRequirements* pSparseMemoryRequirements)
3976 {
3977 *pSparseMemoryRequirementCount = 0;
3978 }
3979
3980 void anv_GetImageSparseMemoryRequirements2(
3981 VkDevice device,
3982 const VkImageSparseMemoryRequirementsInfo2* pInfo,
3983 uint32_t* pSparseMemoryRequirementCount,
3984 VkSparseImageMemoryRequirements2* pSparseMemoryRequirements)
3985 {
3986 *pSparseMemoryRequirementCount = 0;
3987 }
3988
3989 void anv_GetDeviceMemoryCommitment(
3990 VkDevice device,
3991 VkDeviceMemory memory,
3992 VkDeviceSize* pCommittedMemoryInBytes)
3993 {
3994 *pCommittedMemoryInBytes = 0;
3995 }
3996
3997 static void
3998 anv_bind_buffer_memory(const VkBindBufferMemoryInfo *pBindInfo)
3999 {
4000 ANV_FROM_HANDLE(anv_device_memory, mem, pBindInfo->memory);
4001 ANV_FROM_HANDLE(anv_buffer, buffer, pBindInfo->buffer);
4002
4003 assert(pBindInfo->sType == VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO);
4004
4005 if (mem) {
4006 buffer->address = (struct anv_address) {
4007 .bo = mem->bo,
4008 .offset = pBindInfo->memoryOffset,
4009 };
4010 } else {
4011 buffer->address = ANV_NULL_ADDRESS;
4012 }
4013 }
4014
4015 VkResult anv_BindBufferMemory(
4016 VkDevice device,
4017 VkBuffer buffer,
4018 VkDeviceMemory memory,
4019 VkDeviceSize memoryOffset)
4020 {
4021 anv_bind_buffer_memory(
4022 &(VkBindBufferMemoryInfo) {
4023 .sType = VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO,
4024 .buffer = buffer,
4025 .memory = memory,
4026 .memoryOffset = memoryOffset,
4027 });
4028
4029 return VK_SUCCESS;
4030 }
4031
4032 VkResult anv_BindBufferMemory2(
4033 VkDevice device,
4034 uint32_t bindInfoCount,
4035 const VkBindBufferMemoryInfo* pBindInfos)
4036 {
4037 for (uint32_t i = 0; i < bindInfoCount; i++)
4038 anv_bind_buffer_memory(&pBindInfos[i]);
4039
4040 return VK_SUCCESS;
4041 }
4042
4043 VkResult anv_QueueBindSparse(
4044 VkQueue _queue,
4045 uint32_t bindInfoCount,
4046 const VkBindSparseInfo* pBindInfo,
4047 VkFence fence)
4048 {
4049 ANV_FROM_HANDLE(anv_queue, queue, _queue);
4050 if (anv_device_is_lost(queue->device))
4051 return VK_ERROR_DEVICE_LOST;
4052
4053 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
4054 }
4055
4056 // Event functions
4057
4058 VkResult anv_CreateEvent(
4059 VkDevice _device,
4060 const VkEventCreateInfo* pCreateInfo,
4061 const VkAllocationCallbacks* pAllocator,
4062 VkEvent* pEvent)
4063 {
4064 ANV_FROM_HANDLE(anv_device, device, _device);
4065 struct anv_state state;
4066 struct anv_event *event;
4067
4068 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_EVENT_CREATE_INFO);
4069
4070 state = anv_state_pool_alloc(&device->dynamic_state_pool,
4071 sizeof(*event), 8);
4072 event = state.map;
4073 event->state = state;
4074 event->semaphore = VK_EVENT_RESET;
4075
4076 if (!device->info.has_llc) {
4077 /* Make sure the writes we're flushing have landed. */
4078 __builtin_ia32_mfence();
4079 __builtin_ia32_clflush(event);
4080 }
4081
4082 *pEvent = anv_event_to_handle(event);
4083
4084 return VK_SUCCESS;
4085 }
4086
4087 void anv_DestroyEvent(
4088 VkDevice _device,
4089 VkEvent _event,
4090 const VkAllocationCallbacks* pAllocator)
4091 {
4092 ANV_FROM_HANDLE(anv_device, device, _device);
4093 ANV_FROM_HANDLE(anv_event, event, _event);
4094
4095 if (!event)
4096 return;
4097
4098 anv_state_pool_free(&device->dynamic_state_pool, event->state);
4099 }
4100
4101 VkResult anv_GetEventStatus(
4102 VkDevice _device,
4103 VkEvent _event)
4104 {
4105 ANV_FROM_HANDLE(anv_device, device, _device);
4106 ANV_FROM_HANDLE(anv_event, event, _event);
4107
4108 if (anv_device_is_lost(device))
4109 return VK_ERROR_DEVICE_LOST;
4110
4111 if (!device->info.has_llc) {
4112 /* Invalidate read cache before reading event written by GPU. */
4113 __builtin_ia32_clflush(event);
4114 __builtin_ia32_mfence();
4115
4116 }
4117
4118 return event->semaphore;
4119 }
4120
4121 VkResult anv_SetEvent(
4122 VkDevice _device,
4123 VkEvent _event)
4124 {
4125 ANV_FROM_HANDLE(anv_device, device, _device);
4126 ANV_FROM_HANDLE(anv_event, event, _event);
4127
4128 event->semaphore = VK_EVENT_SET;
4129
4130 if (!device->info.has_llc) {
4131 /* Make sure the writes we're flushing have landed. */
4132 __builtin_ia32_mfence();
4133 __builtin_ia32_clflush(event);
4134 }
4135
4136 return VK_SUCCESS;
4137 }
4138
4139 VkResult anv_ResetEvent(
4140 VkDevice _device,
4141 VkEvent _event)
4142 {
4143 ANV_FROM_HANDLE(anv_device, device, _device);
4144 ANV_FROM_HANDLE(anv_event, event, _event);
4145
4146 event->semaphore = VK_EVENT_RESET;
4147
4148 if (!device->info.has_llc) {
4149 /* Make sure the writes we're flushing have landed. */
4150 __builtin_ia32_mfence();
4151 __builtin_ia32_clflush(event);
4152 }
4153
4154 return VK_SUCCESS;
4155 }
4156
4157 // Buffer functions
4158
4159 VkResult anv_CreateBuffer(
4160 VkDevice _device,
4161 const VkBufferCreateInfo* pCreateInfo,
4162 const VkAllocationCallbacks* pAllocator,
4163 VkBuffer* pBuffer)
4164 {
4165 ANV_FROM_HANDLE(anv_device, device, _device);
4166 struct anv_buffer *buffer;
4167
4168 /* Don't allow creating buffers bigger than our address space. The real
4169 * issue here is that we may align up the buffer size and we don't want
4170 * doing so to cause roll-over. However, no one has any business
4171 * allocating a buffer larger than our GTT size.
4172 */
4173 if (pCreateInfo->size > device->physical->gtt_size)
4174 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
4175
4176 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
4177
4178 buffer = vk_alloc2(&device->alloc, pAllocator, sizeof(*buffer), 8,
4179 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
4180 if (buffer == NULL)
4181 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
4182
4183 buffer->size = pCreateInfo->size;
4184 buffer->usage = pCreateInfo->usage;
4185 buffer->address = ANV_NULL_ADDRESS;
4186
4187 *pBuffer = anv_buffer_to_handle(buffer);
4188
4189 return VK_SUCCESS;
4190 }
4191
4192 void anv_DestroyBuffer(
4193 VkDevice _device,
4194 VkBuffer _buffer,
4195 const VkAllocationCallbacks* pAllocator)
4196 {
4197 ANV_FROM_HANDLE(anv_device, device, _device);
4198 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
4199
4200 if (!buffer)
4201 return;
4202
4203 vk_free2(&device->alloc, pAllocator, buffer);
4204 }
4205
4206 VkDeviceAddress anv_GetBufferDeviceAddress(
4207 VkDevice device,
4208 const VkBufferDeviceAddressInfoKHR* pInfo)
4209 {
4210 ANV_FROM_HANDLE(anv_buffer, buffer, pInfo->buffer);
4211
4212 assert(!anv_address_is_null(buffer->address));
4213 assert(buffer->address.bo->flags & EXEC_OBJECT_PINNED);
4214
4215 return anv_address_physical(buffer->address);
4216 }
4217
4218 uint64_t anv_GetBufferOpaqueCaptureAddress(
4219 VkDevice device,
4220 const VkBufferDeviceAddressInfoKHR* pInfo)
4221 {
4222 return 0;
4223 }
4224
4225 uint64_t anv_GetDeviceMemoryOpaqueCaptureAddress(
4226 VkDevice device,
4227 const VkDeviceMemoryOpaqueCaptureAddressInfoKHR* pInfo)
4228 {
4229 ANV_FROM_HANDLE(anv_device_memory, memory, pInfo->memory);
4230
4231 assert(memory->bo->flags & EXEC_OBJECT_PINNED);
4232 assert(memory->bo->has_client_visible_address);
4233
4234 return gen_48b_address(memory->bo->offset);
4235 }
4236
4237 void
4238 anv_fill_buffer_surface_state(struct anv_device *device, struct anv_state state,
4239 enum isl_format format,
4240 struct anv_address address,
4241 uint32_t range, uint32_t stride)
4242 {
4243 isl_buffer_fill_state(&device->isl_dev, state.map,
4244 .address = anv_address_physical(address),
4245 .mocs = device->isl_dev.mocs.internal,
4246 .size_B = range,
4247 .format = format,
4248 .swizzle = ISL_SWIZZLE_IDENTITY,
4249 .stride_B = stride);
4250 }
4251
4252 void anv_DestroySampler(
4253 VkDevice _device,
4254 VkSampler _sampler,
4255 const VkAllocationCallbacks* pAllocator)
4256 {
4257 ANV_FROM_HANDLE(anv_device, device, _device);
4258 ANV_FROM_HANDLE(anv_sampler, sampler, _sampler);
4259
4260 if (!sampler)
4261 return;
4262
4263 if (sampler->bindless_state.map) {
4264 anv_state_pool_free(&device->dynamic_state_pool,
4265 sampler->bindless_state);
4266 }
4267
4268 vk_free2(&device->alloc, pAllocator, sampler);
4269 }
4270
4271 VkResult anv_CreateFramebuffer(
4272 VkDevice _device,
4273 const VkFramebufferCreateInfo* pCreateInfo,
4274 const VkAllocationCallbacks* pAllocator,
4275 VkFramebuffer* pFramebuffer)
4276 {
4277 ANV_FROM_HANDLE(anv_device, device, _device);
4278 struct anv_framebuffer *framebuffer;
4279
4280 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
4281
4282 size_t size = sizeof(*framebuffer);
4283
4284 /* VK_KHR_imageless_framebuffer extension says:
4285 *
4286 * If flags includes VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR,
4287 * parameter pAttachments is ignored.
4288 */
4289 if (!(pCreateInfo->flags & VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR)) {
4290 size += sizeof(struct anv_image_view *) * pCreateInfo->attachmentCount;
4291 framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
4292 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
4293 if (framebuffer == NULL)
4294 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
4295
4296 for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
4297 ANV_FROM_HANDLE(anv_image_view, iview, pCreateInfo->pAttachments[i]);
4298 framebuffer->attachments[i] = iview;
4299 }
4300 framebuffer->attachment_count = pCreateInfo->attachmentCount;
4301 } else {
4302 assert(device->enabled_extensions.KHR_imageless_framebuffer);
4303 framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
4304 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
4305 if (framebuffer == NULL)
4306 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
4307
4308 framebuffer->attachment_count = 0;
4309 }
4310
4311 framebuffer->width = pCreateInfo->width;
4312 framebuffer->height = pCreateInfo->height;
4313 framebuffer->layers = pCreateInfo->layers;
4314
4315 *pFramebuffer = anv_framebuffer_to_handle(framebuffer);
4316
4317 return VK_SUCCESS;
4318 }
4319
4320 void anv_DestroyFramebuffer(
4321 VkDevice _device,
4322 VkFramebuffer _fb,
4323 const VkAllocationCallbacks* pAllocator)
4324 {
4325 ANV_FROM_HANDLE(anv_device, device, _device);
4326 ANV_FROM_HANDLE(anv_framebuffer, fb, _fb);
4327
4328 if (!fb)
4329 return;
4330
4331 vk_free2(&device->alloc, pAllocator, fb);
4332 }
4333
4334 static const VkTimeDomainEXT anv_time_domains[] = {
4335 VK_TIME_DOMAIN_DEVICE_EXT,
4336 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT,
4337 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT,
4338 };
4339
4340 VkResult anv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
4341 VkPhysicalDevice physicalDevice,
4342 uint32_t *pTimeDomainCount,
4343 VkTimeDomainEXT *pTimeDomains)
4344 {
4345 int d;
4346 VK_OUTARRAY_MAKE(out, pTimeDomains, pTimeDomainCount);
4347
4348 for (d = 0; d < ARRAY_SIZE(anv_time_domains); d++) {
4349 vk_outarray_append(&out, i) {
4350 *i = anv_time_domains[d];
4351 }
4352 }
4353
4354 return vk_outarray_status(&out);
4355 }
4356
4357 static uint64_t
4358 anv_clock_gettime(clockid_t clock_id)
4359 {
4360 struct timespec current;
4361 int ret;
4362
4363 ret = clock_gettime(clock_id, &current);
4364 if (ret < 0 && clock_id == CLOCK_MONOTONIC_RAW)
4365 ret = clock_gettime(CLOCK_MONOTONIC, &current);
4366 if (ret < 0)
4367 return 0;
4368
4369 return (uint64_t) current.tv_sec * 1000000000ULL + current.tv_nsec;
4370 }
4371
4372 #define TIMESTAMP 0x2358
4373
4374 VkResult anv_GetCalibratedTimestampsEXT(
4375 VkDevice _device,
4376 uint32_t timestampCount,
4377 const VkCalibratedTimestampInfoEXT *pTimestampInfos,
4378 uint64_t *pTimestamps,
4379 uint64_t *pMaxDeviation)
4380 {
4381 ANV_FROM_HANDLE(anv_device, device, _device);
4382 uint64_t timestamp_frequency = device->info.timestamp_frequency;
4383 int ret;
4384 int d;
4385 uint64_t begin, end;
4386 uint64_t max_clock_period = 0;
4387
4388 begin = anv_clock_gettime(CLOCK_MONOTONIC_RAW);
4389
4390 for (d = 0; d < timestampCount; d++) {
4391 switch (pTimestampInfos[d].timeDomain) {
4392 case VK_TIME_DOMAIN_DEVICE_EXT:
4393 ret = anv_gem_reg_read(device, TIMESTAMP | 1,
4394 &pTimestamps[d]);
4395
4396 if (ret != 0) {
4397 return anv_device_set_lost(device, "Failed to read the TIMESTAMP "
4398 "register: %m");
4399 }
4400 uint64_t device_period = DIV_ROUND_UP(1000000000, timestamp_frequency);
4401 max_clock_period = MAX2(max_clock_period, device_period);
4402 break;
4403 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT:
4404 pTimestamps[d] = anv_clock_gettime(CLOCK_MONOTONIC);
4405 max_clock_period = MAX2(max_clock_period, 1);
4406 break;
4407
4408 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT:
4409 pTimestamps[d] = begin;
4410 break;
4411 default:
4412 pTimestamps[d] = 0;
4413 break;
4414 }
4415 }
4416
4417 end = anv_clock_gettime(CLOCK_MONOTONIC_RAW);
4418
4419 /*
4420 * The maximum deviation is the sum of the interval over which we
4421 * perform the sampling and the maximum period of any sampled
4422 * clock. That's because the maximum skew between any two sampled
4423 * clock edges is when the sampled clock with the largest period is
4424 * sampled at the end of that period but right at the beginning of the
4425 * sampling interval and some other clock is sampled right at the
4426 * begining of its sampling period and right at the end of the
4427 * sampling interval. Let's assume the GPU has the longest clock
4428 * period and that the application is sampling GPU and monotonic:
4429 *
4430 * s e
4431 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
4432 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4433 *
4434 * g
4435 * 0 1 2 3
4436 * GPU -----_____-----_____-----_____-----_____
4437 *
4438 * m
4439 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
4440 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4441 *
4442 * Interval <----------------->
4443 * Deviation <-------------------------->
4444 *
4445 * s = read(raw) 2
4446 * g = read(GPU) 1
4447 * m = read(monotonic) 2
4448 * e = read(raw) b
4449 *
4450 * We round the sample interval up by one tick to cover sampling error
4451 * in the interval clock
4452 */
4453
4454 uint64_t sample_interval = end - begin + 1;
4455
4456 *pMaxDeviation = sample_interval + max_clock_period;
4457
4458 return VK_SUCCESS;
4459 }
4460
4461 /* vk_icd.h does not declare this function, so we declare it here to
4462 * suppress Wmissing-prototypes.
4463 */
4464 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4465 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion);
4466
4467 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4468 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion)
4469 {
4470 /* For the full details on loader interface versioning, see
4471 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
4472 * What follows is a condensed summary, to help you navigate the large and
4473 * confusing official doc.
4474 *
4475 * - Loader interface v0 is incompatible with later versions. We don't
4476 * support it.
4477 *
4478 * - In loader interface v1:
4479 * - The first ICD entrypoint called by the loader is
4480 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
4481 * entrypoint.
4482 * - The ICD must statically expose no other Vulkan symbol unless it is
4483 * linked with -Bsymbolic.
4484 * - Each dispatchable Vulkan handle created by the ICD must be
4485 * a pointer to a struct whose first member is VK_LOADER_DATA. The
4486 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
4487 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
4488 * vkDestroySurfaceKHR(). The ICD must be capable of working with
4489 * such loader-managed surfaces.
4490 *
4491 * - Loader interface v2 differs from v1 in:
4492 * - The first ICD entrypoint called by the loader is
4493 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
4494 * statically expose this entrypoint.
4495 *
4496 * - Loader interface v3 differs from v2 in:
4497 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
4498 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
4499 * because the loader no longer does so.
4500 *
4501 * - Loader interface v4 differs from v3 in:
4502 * - The ICD must implement vk_icdGetPhysicalDeviceProcAddr().
4503 */
4504 *pSupportedVersion = MIN2(*pSupportedVersion, 4u);
4505 return VK_SUCCESS;
4506 }