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