projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
anv: Return VK_ERROR_OUT_OF_DEVICE_MEMORY for too-large buffers
[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_SECTION_END
61 DRI_CONF_END;
62
63 /* This is probably far to big but it reflects the max size used for messages
64 * in OpenGLs KHR_debug.
65 */
66 #define MAX_DEBUG_MESSAGE_LENGTH 4096
67
68 static void
69 compiler_debug_log(void *data, const char *fmt, ...)
70 {
71 char str[MAX_DEBUG_MESSAGE_LENGTH];
72 struct anv_device *device = (struct anv_device *)data;
73
74 if (list_is_empty(&device->instance->debug_report_callbacks.callbacks))
75 return;
76
77 va_list args;
78 va_start(args, fmt);
79 (void) vsnprintf(str, MAX_DEBUG_MESSAGE_LENGTH, fmt, args);
80 va_end(args);
81
82 vk_debug_report(&device->instance->debug_report_callbacks,
83 VK_DEBUG_REPORT_DEBUG_BIT_EXT,
84 VK_DEBUG_REPORT_OBJECT_TYPE_UNKNOWN_EXT,
85 0, 0, 0, "anv", str);
86 }
87
88 static void
89 compiler_perf_log(void *data, const char *fmt, ...)
90 {
91 va_list args;
92 va_start(args, fmt);
93
94 if (unlikely(INTEL_DEBUG & DEBUG_PERF))
95 intel_logd_v(fmt, args);
96
97 va_end(args);
98 }
99
100 static uint64_t
101 anv_compute_heap_size(int fd, uint64_t gtt_size)
102 {
103 /* Query the total ram from the system */
104 struct sysinfo info;
105 sysinfo(&info);
106
107 uint64_t total_ram = (uint64_t)info.totalram * (uint64_t)info.mem_unit;
108
109 /* We don't want to burn too much ram with the GPU. If the user has 4GiB
110 * or less, we use at most half. If they have more than 4GiB, we use 3/4.
111 */
112 uint64_t available_ram;
113 if (total_ram <= 4ull * 1024ull * 1024ull * 1024ull)
114 available_ram = total_ram / 2;
115 else
116 available_ram = total_ram * 3 / 4;
117
118 /* We also want to leave some padding for things we allocate in the driver,
119 * so don't go over 3/4 of the GTT either.
120 */
121 uint64_t available_gtt = gtt_size * 3 / 4;
122
123 return MIN2(available_ram, available_gtt);
124 }
125
126 static VkResult
127 anv_physical_device_init_heaps(struct anv_physical_device *device, int fd)
128 {
129 if (anv_gem_get_context_param(fd, 0, I915_CONTEXT_PARAM_GTT_SIZE,
130 &device->gtt_size) == -1) {
131 /* If, for whatever reason, we can't actually get the GTT size from the
132 * kernel (too old?) fall back to the aperture size.
133 */
134 anv_perf_warn(NULL, NULL,
135 "Failed to get I915_CONTEXT_PARAM_GTT_SIZE: %m");
136
137 if (anv_gem_get_aperture(fd, &device->gtt_size) == -1) {
138 return vk_errorf(NULL, NULL, VK_ERROR_INITIALIZATION_FAILED,
139 "failed to get aperture size: %m");
140 }
141 }
142
143 /* We only allow 48-bit addresses with softpin because knowing the actual
144 * address is required for the vertex cache flush workaround.
145 */
146 device->supports_48bit_addresses = (device->info.gen >= 8) &&
147 device->has_softpin &&
148 device->gtt_size > (4ULL << 30 /* GiB */);
149
150 uint64_t heap_size = anv_compute_heap_size(fd, device->gtt_size);
151
152 if (heap_size > (2ull << 30) && !device->supports_48bit_addresses) {
153 /* When running with an overridden PCI ID, we may get a GTT size from
154 * the kernel that is greater than 2 GiB but the execbuf check for 48bit
155 * address support can still fail. Just clamp the address space size to
156 * 2 GiB if we don't have 48-bit support.
157 */
158 intel_logw("%s:%d: The kernel reported a GTT size larger than 2 GiB but "
159 "not support for 48-bit addresses",
160 __FILE__, __LINE__);
161 heap_size = 2ull << 30;
162 }
163
164 device->memory.heap_count = 1;
165 device->memory.heaps[0] = (struct anv_memory_heap) {
166 .size = heap_size,
167 .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
168 };
169
170 uint32_t type_count = 0;
171 for (uint32_t heap = 0; heap < device->memory.heap_count; heap++) {
172 if (device->info.has_llc) {
173 /* Big core GPUs share LLC with the CPU and thus one memory type can be
174 * both cached and coherent at the same time.
175 */
176 device->memory.types[type_count++] = (struct anv_memory_type) {
177 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
178 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
179 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
180 VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
181 .heapIndex = heap,
182 };
183 } else {
184 /* The spec requires that we expose a host-visible, coherent memory
185 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
186 * to give the application a choice between cached, but not coherent and
187 * coherent but uncached (WC though).
188 */
189 device->memory.types[type_count++] = (struct anv_memory_type) {
190 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
191 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
192 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
193 .heapIndex = heap,
194 };
195 device->memory.types[type_count++] = (struct anv_memory_type) {
196 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
197 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
198 VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
199 .heapIndex = heap,
200 };
201 }
202 }
203 device->memory.type_count = type_count;
204
205 return VK_SUCCESS;
206 }
207
208 static VkResult
209 anv_physical_device_init_uuids(struct anv_physical_device *device)
210 {
211 const struct build_id_note *note =
212 build_id_find_nhdr_for_addr(anv_physical_device_init_uuids);
213 if (!note) {
214 return vk_errorf(device->instance, device,
215 VK_ERROR_INITIALIZATION_FAILED,
216 "Failed to find build-id");
217 }
218
219 unsigned build_id_len = build_id_length(note);
220 if (build_id_len < 20) {
221 return vk_errorf(device->instance, device,
222 VK_ERROR_INITIALIZATION_FAILED,
223 "build-id too short. It needs to be a SHA");
224 }
225
226 memcpy(device->driver_build_sha1, build_id_data(note), 20);
227
228 struct mesa_sha1 sha1_ctx;
229 uint8_t sha1[20];
230 STATIC_ASSERT(VK_UUID_SIZE <= sizeof(sha1));
231
232 /* The pipeline cache UUID is used for determining when a pipeline cache is
233 * invalid. It needs both a driver build and the PCI ID of the device.
234 */
235 _mesa_sha1_init(&sha1_ctx);
236 _mesa_sha1_update(&sha1_ctx, build_id_data(note), build_id_len);
237 _mesa_sha1_update(&sha1_ctx, &device->chipset_id,
238 sizeof(device->chipset_id));
239 _mesa_sha1_update(&sha1_ctx, &device->always_use_bindless,
240 sizeof(device->always_use_bindless));
241 _mesa_sha1_update(&sha1_ctx, &device->has_a64_buffer_access,
242 sizeof(device->has_a64_buffer_access));
243 _mesa_sha1_update(&sha1_ctx, &device->has_bindless_images,
244 sizeof(device->has_bindless_images));
245 _mesa_sha1_update(&sha1_ctx, &device->has_bindless_samplers,
246 sizeof(device->has_bindless_samplers));
247 _mesa_sha1_final(&sha1_ctx, sha1);
248 memcpy(device->pipeline_cache_uuid, sha1, VK_UUID_SIZE);
249
250 /* The driver UUID is used for determining sharability of images and memory
251 * between two Vulkan instances in separate processes. People who want to
252 * share memory need to also check the device UUID (below) so all this
253 * needs to be is the build-id.
254 */
255 memcpy(device->driver_uuid, build_id_data(note), VK_UUID_SIZE);
256
257 /* The device UUID uniquely identifies the given device within the machine.
258 * Since we never have more than one device, this doesn't need to be a real
259 * UUID. However, on the off-chance that someone tries to use this to
260 * cache pre-tiled images or something of the like, we use the PCI ID and
261 * some bits of ISL info to ensure that this is safe.
262 */
263 _mesa_sha1_init(&sha1_ctx);
264 _mesa_sha1_update(&sha1_ctx, &device->chipset_id,
265 sizeof(device->chipset_id));
266 _mesa_sha1_update(&sha1_ctx, &device->isl_dev.has_bit6_swizzling,
267 sizeof(device->isl_dev.has_bit6_swizzling));
268 _mesa_sha1_final(&sha1_ctx, sha1);
269 memcpy(device->device_uuid, sha1, VK_UUID_SIZE);
270
271 return VK_SUCCESS;
272 }
273
274 static void
275 anv_physical_device_init_disk_cache(struct anv_physical_device *device)
276 {
277 #ifdef ENABLE_SHADER_CACHE
278 char renderer[10];
279 ASSERTED int len = snprintf(renderer, sizeof(renderer), "anv_%04x",
280 device->chipset_id);
281 assert(len == sizeof(renderer) - 2);
282
283 char timestamp[41];
284 _mesa_sha1_format(timestamp, device->driver_build_sha1);
285
286 const uint64_t driver_flags =
287 brw_get_compiler_config_value(device->compiler);
288 device->disk_cache = disk_cache_create(renderer, timestamp, driver_flags);
289 #else
290 device->disk_cache = NULL;
291 #endif
292 }
293
294 static void
295 anv_physical_device_free_disk_cache(struct anv_physical_device *device)
296 {
297 #ifdef ENABLE_SHADER_CACHE
298 if (device->disk_cache)
299 disk_cache_destroy(device->disk_cache);
300 #else
301 assert(device->disk_cache == NULL);
302 #endif
303 }
304
305 static uint64_t
306 get_available_system_memory()
307 {
308 char *meminfo = os_read_file("/proc/meminfo");
309 if (!meminfo)
310 return 0;
311
312 char *str = strstr(meminfo, "MemAvailable:");
313 if (!str) {
314 free(meminfo);
315 return 0;
316 }
317
318 uint64_t kb_mem_available;
319 if (sscanf(str, "MemAvailable: %" PRIx64, &kb_mem_available) == 1) {
320 free(meminfo);
321 return kb_mem_available << 10;
322 }
323
324 free(meminfo);
325 return 0;
326 }
327
328 static VkResult
329 anv_physical_device_init(struct anv_physical_device *device,
330 struct anv_instance *instance,
331 drmDevicePtr drm_device)
332 {
333 const char *primary_path = drm_device->nodes[DRM_NODE_PRIMARY];
334 const char *path = drm_device->nodes[DRM_NODE_RENDER];
335 VkResult result;
336 int fd;
337 int master_fd = -1;
338
339 brw_process_intel_debug_variable();
340
341 fd = open(path, O_RDWR | O_CLOEXEC);
342 if (fd < 0)
343 return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
344
345 device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
346 device->instance = instance;
347
348 assert(strlen(path) < ARRAY_SIZE(device->path));
349 snprintf(device->path, ARRAY_SIZE(device->path), "%s", path);
350
351 if (!gen_get_device_info_from_fd(fd, &device->info)) {
352 result = vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
353 goto fail;
354 }
355 device->chipset_id = device->info.chipset_id;
356 device->no_hw = device->info.no_hw;
357
358 if (getenv("INTEL_NO_HW") != NULL)
359 device->no_hw = true;
360
361 device->pci_info.domain = drm_device->businfo.pci->domain;
362 device->pci_info.bus = drm_device->businfo.pci->bus;
363 device->pci_info.device = drm_device->businfo.pci->dev;
364 device->pci_info.function = drm_device->businfo.pci->func;
365
366 device->name = gen_get_device_name(device->chipset_id);
367
368 if (device->info.is_haswell) {
369 intel_logw("Haswell Vulkan support is incomplete");
370 } else if (device->info.gen == 7 && !device->info.is_baytrail) {
371 intel_logw("Ivy Bridge Vulkan support is incomplete");
372 } else if (device->info.gen == 7 && device->info.is_baytrail) {
373 intel_logw("Bay Trail Vulkan support is incomplete");
374 } else if (device->info.gen >= 8 && device->info.gen <= 11) {
375 /* Gen8-11 fully supported */
376 } else if (device->info.gen == 12) {
377 intel_logw("Vulkan is not yet fully supported on gen12");
378 } else {
379 result = vk_errorf(device->instance, device,
380 VK_ERROR_INCOMPATIBLE_DRIVER,
381 "Vulkan not yet supported on %s", device->name);
382 goto fail;
383 }
384
385 device->cmd_parser_version = -1;
386 if (device->info.gen == 7) {
387 device->cmd_parser_version =
388 anv_gem_get_param(fd, I915_PARAM_CMD_PARSER_VERSION);
389 if (device->cmd_parser_version == -1) {
390 result = vk_errorf(device->instance, device,
391 VK_ERROR_INITIALIZATION_FAILED,
392 "failed to get command parser version");
393 goto fail;
394 }
395 }
396
397 if (!anv_gem_get_param(fd, I915_PARAM_HAS_WAIT_TIMEOUT)) {
398 result = vk_errorf(device->instance, device,
399 VK_ERROR_INITIALIZATION_FAILED,
400 "kernel missing gem wait");
401 goto fail;
402 }
403
404 if (!anv_gem_get_param(fd, I915_PARAM_HAS_EXECBUF2)) {
405 result = vk_errorf(device->instance, device,
406 VK_ERROR_INITIALIZATION_FAILED,
407 "kernel missing execbuf2");
408 goto fail;
409 }
410
411 if (!device->info.has_llc &&
412 anv_gem_get_param(fd, I915_PARAM_MMAP_VERSION) < 1) {
413 result = vk_errorf(device->instance, device,
414 VK_ERROR_INITIALIZATION_FAILED,
415 "kernel missing wc mmap");
416 goto fail;
417 }
418
419 device->has_softpin = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_SOFTPIN);
420 device->has_exec_async = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_ASYNC);
421 device->has_exec_capture = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_CAPTURE);
422 device->has_exec_fence = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_FENCE);
423 device->has_syncobj = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_FENCE_ARRAY);
424 device->has_syncobj_wait = device->has_syncobj &&
425 anv_gem_supports_syncobj_wait(fd);
426 device->has_context_priority = anv_gem_has_context_priority(fd);
427
428 result = anv_physical_device_init_heaps(device, fd);
429 if (result != VK_SUCCESS)
430 goto fail;
431
432 device->use_softpin = device->has_softpin &&
433 device->supports_48bit_addresses;
434
435 device->has_context_isolation =
436 anv_gem_get_param(fd, I915_PARAM_HAS_CONTEXT_ISOLATION);
437
438 device->always_use_bindless =
439 env_var_as_boolean("ANV_ALWAYS_BINDLESS", false);
440
441 /* We first got the A64 messages on broadwell and we can only use them if
442 * we can pass addresses directly into the shader which requires softpin.
443 */
444 device->has_a64_buffer_access = device->info.gen >= 8 &&
445 device->use_softpin;
446
447 /* We first get bindless image access on Skylake and we can only really do
448 * it if we don't have any relocations so we need softpin.
449 */
450 device->has_bindless_images = device->info.gen >= 9 &&
451 device->use_softpin;
452
453 /* We've had bindless samplers since Ivy Bridge (forever in Vulkan terms)
454 * because it's just a matter of setting the sampler address in the sample
455 * message header. However, we've not bothered to wire it up for vec4 so
456 * we leave it disabled on gen7.
457 */
458 device->has_bindless_samplers = device->info.gen >= 8;
459
460 device->has_mem_available = get_available_system_memory() != 0;
461
462 device->always_flush_cache =
463 driQueryOptionb(&instance->dri_options, "always_flush_cache");
464
465 /* Starting with Gen10, the timestamp frequency of the command streamer may
466 * vary from one part to another. We can query the value from the kernel.
467 */
468 if (device->info.gen >= 10) {
469 int timestamp_frequency =
470 anv_gem_get_param(fd, I915_PARAM_CS_TIMESTAMP_FREQUENCY);
471
472 if (timestamp_frequency < 0)
473 intel_logw("Kernel 4.16-rc1+ required to properly query CS timestamp frequency");
474 else
475 device->info.timestamp_frequency = timestamp_frequency;
476 }
477
478 /* GENs prior to 8 do not support EU/Subslice info */
479 if (device->info.gen >= 8) {
480 device->subslice_total = anv_gem_get_param(fd, I915_PARAM_SUBSLICE_TOTAL);
481 device->eu_total = anv_gem_get_param(fd, I915_PARAM_EU_TOTAL);
482
483 /* Without this information, we cannot get the right Braswell
484 * brandstrings, and we have to use conservative numbers for GPGPU on
485 * many platforms, but otherwise, things will just work.
486 */
487 if (device->subslice_total < 1 || device->eu_total < 1) {
488 intel_logw("Kernel 4.1 required to properly query GPU properties");
489 }
490 } else if (device->info.gen == 7) {
491 device->subslice_total = 1 << (device->info.gt - 1);
492 }
493
494 if (device->info.is_cherryview &&
495 device->subslice_total > 0 && device->eu_total > 0) {
496 /* Logical CS threads = EUs per subslice * num threads per EU */
497 uint32_t max_cs_threads =
498 device->eu_total / device->subslice_total * device->info.num_thread_per_eu;
499
500 /* Fuse configurations may give more threads than expected, never less. */
501 if (max_cs_threads > device->info.max_cs_threads)
502 device->info.max_cs_threads = max_cs_threads;
503 }
504
505 device->compiler = brw_compiler_create(NULL, &device->info);
506 if (device->compiler == NULL) {
507 result = vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
508 goto fail;
509 }
510 device->compiler->shader_debug_log = compiler_debug_log;
511 device->compiler->shader_perf_log = compiler_perf_log;
512 device->compiler->supports_pull_constants = false;
513 device->compiler->constant_buffer_0_is_relative =
514 device->info.gen < 8 || !device->has_context_isolation;
515 device->compiler->supports_shader_constants = true;
516 device->compiler->compact_params = false;
517
518 /* Broadwell PRM says:
519 *
520 * "Before Gen8, there was a historical configuration control field to
521 * swizzle address bit[6] for in X/Y tiling modes. This was set in three
522 * different places: TILECTL[1:0], ARB_MODE[5:4], and
523 * DISP_ARB_CTL[14:13].
524 *
525 * For Gen8 and subsequent generations, the swizzle fields are all
526 * reserved, and the CPU's memory controller performs all address
527 * swizzling modifications."
528 */
529 bool swizzled =
530 device->info.gen < 8 && anv_gem_get_bit6_swizzle(fd, I915_TILING_X);
531
532 isl_device_init(&device->isl_dev, &device->info, swizzled);
533
534 result = anv_physical_device_init_uuids(device);
535 if (result != VK_SUCCESS)
536 goto fail;
537
538 anv_physical_device_init_disk_cache(device);
539
540 if (instance->enabled_extensions.KHR_display) {
541 master_fd = open(primary_path, O_RDWR | O_CLOEXEC);
542 if (master_fd >= 0) {
543 /* prod the device with a GETPARAM call which will fail if
544 * we don't have permission to even render on this device
545 */
546 if (anv_gem_get_param(master_fd, I915_PARAM_CHIPSET_ID) == 0) {
547 close(master_fd);
548 master_fd = -1;
549 }
550 }
551 }
552 device->master_fd = master_fd;
553
554 result = anv_init_wsi(device);
555 if (result != VK_SUCCESS) {
556 ralloc_free(device->compiler);
557 anv_physical_device_free_disk_cache(device);
558 goto fail;
559 }
560
561 device->perf = anv_get_perf(&device->info, fd);
562
563 anv_physical_device_get_supported_extensions(device,
564 &device->supported_extensions);
565
566
567 device->local_fd = fd;
568
569 return VK_SUCCESS;
570
571 fail:
572 close(fd);
573 if (master_fd != -1)
574 close(master_fd);
575 return result;
576 }
577
578 static void
579 anv_physical_device_finish(struct anv_physical_device *device)
580 {
581 anv_finish_wsi(device);
582 anv_physical_device_free_disk_cache(device);
583 ralloc_free(device->compiler);
584 ralloc_free(device->perf);
585 close(device->local_fd);
586 if (device->master_fd >= 0)
587 close(device->master_fd);
588 }
589
590 static void *
591 default_alloc_func(void *pUserData, size_t size, size_t align,
592 VkSystemAllocationScope allocationScope)
593 {
594 return malloc(size);
595 }
596
597 static void *
598 default_realloc_func(void *pUserData, void *pOriginal, size_t size,
599 size_t align, VkSystemAllocationScope allocationScope)
600 {
601 return realloc(pOriginal, size);
602 }
603
604 static void
605 default_free_func(void *pUserData, void *pMemory)
606 {
607 free(pMemory);
608 }
609
610 static const VkAllocationCallbacks default_alloc = {
611 .pUserData = NULL,
612 .pfnAllocation = default_alloc_func,
613 .pfnReallocation = default_realloc_func,
614 .pfnFree = default_free_func,
615 };
616
617 VkResult anv_EnumerateInstanceExtensionProperties(
618 const char* pLayerName,
619 uint32_t* pPropertyCount,
620 VkExtensionProperties* pProperties)
621 {
622 VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
623
624 for (int i = 0; i < ANV_INSTANCE_EXTENSION_COUNT; i++) {
625 if (anv_instance_extensions_supported.extensions[i]) {
626 vk_outarray_append(&out, prop) {
627 *prop = anv_instance_extensions[i];
628 }
629 }
630 }
631
632 return vk_outarray_status(&out);
633 }
634
635 VkResult anv_CreateInstance(
636 const VkInstanceCreateInfo* pCreateInfo,
637 const VkAllocationCallbacks* pAllocator,
638 VkInstance* pInstance)
639 {
640 struct anv_instance *instance;
641 VkResult result;
642
643 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO);
644
645 struct anv_instance_extension_table enabled_extensions = {};
646 for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
647 int idx;
648 for (idx = 0; idx < ANV_INSTANCE_EXTENSION_COUNT; idx++) {
649 if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
650 anv_instance_extensions[idx].extensionName) == 0)
651 break;
652 }
653
654 if (idx >= ANV_INSTANCE_EXTENSION_COUNT)
655 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
656
657 if (!anv_instance_extensions_supported.extensions[idx])
658 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
659
660 enabled_extensions.extensions[idx] = true;
661 }
662
663 instance = vk_alloc2(&default_alloc, pAllocator, sizeof(*instance), 8,
664 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
665 if (!instance)
666 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
667
668 instance->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
669
670 if (pAllocator)
671 instance->alloc = *pAllocator;
672 else
673 instance->alloc = default_alloc;
674
675 instance->app_info = (struct anv_app_info) { .api_version = 0 };
676 if (pCreateInfo->pApplicationInfo) {
677 const VkApplicationInfo *app = pCreateInfo->pApplicationInfo;
678
679 instance->app_info.app_name =
680 vk_strdup(&instance->alloc, app->pApplicationName,
681 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
682 instance->app_info.app_version = app->applicationVersion;
683
684 instance->app_info.engine_name =
685 vk_strdup(&instance->alloc, app->pEngineName,
686 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
687 instance->app_info.engine_version = app->engineVersion;
688
689 instance->app_info.api_version = app->apiVersion;
690 }
691
692 if (instance->app_info.api_version == 0)
693 instance->app_info.api_version = VK_API_VERSION_1_0;
694
695 instance->enabled_extensions = enabled_extensions;
696
697 for (unsigned i = 0; i < ARRAY_SIZE(instance->dispatch.entrypoints); i++) {
698 /* Vulkan requires that entrypoints for extensions which have not been
699 * enabled must not be advertised.
700 */
701 if (!anv_instance_entrypoint_is_enabled(i, instance->app_info.api_version,
702 &instance->enabled_extensions)) {
703 instance->dispatch.entrypoints[i] = NULL;
704 } else {
705 instance->dispatch.entrypoints[i] =
706 anv_instance_dispatch_table.entrypoints[i];
707 }
708 }
709
710 struct anv_physical_device *pdevice = &instance->physicalDevice;
711 for (unsigned i = 0; i < ARRAY_SIZE(pdevice->dispatch.entrypoints); i++) {
712 /* Vulkan requires that entrypoints for extensions which have not been
713 * enabled must not be advertised.
714 */
715 if (!anv_physical_device_entrypoint_is_enabled(i, instance->app_info.api_version,
716 &instance->enabled_extensions)) {
717 pdevice->dispatch.entrypoints[i] = NULL;
718 } else {
719 pdevice->dispatch.entrypoints[i] =
720 anv_physical_device_dispatch_table.entrypoints[i];
721 }
722 }
723
724 for (unsigned i = 0; i < ARRAY_SIZE(instance->device_dispatch.entrypoints); i++) {
725 /* Vulkan requires that entrypoints for extensions which have not been
726 * enabled must not be advertised.
727 */
728 if (!anv_device_entrypoint_is_enabled(i, instance->app_info.api_version,
729 &instance->enabled_extensions, NULL)) {
730 instance->device_dispatch.entrypoints[i] = NULL;
731 } else {
732 instance->device_dispatch.entrypoints[i] =
733 anv_device_dispatch_table.entrypoints[i];
734 }
735 }
736
737 instance->physicalDeviceCount = -1;
738
739 result = vk_debug_report_instance_init(&instance->debug_report_callbacks);
740 if (result != VK_SUCCESS) {
741 vk_free2(&default_alloc, pAllocator, instance);
742 return vk_error(result);
743 }
744
745 instance->pipeline_cache_enabled =
746 env_var_as_boolean("ANV_ENABLE_PIPELINE_CACHE", true);
747
748 glsl_type_singleton_init_or_ref();
749
750 VG(VALGRIND_CREATE_MEMPOOL(instance, 0, false));
751
752 driParseOptionInfo(&instance->available_dri_options, anv_dri_options_xml);
753 driParseConfigFiles(&instance->dri_options, &instance->available_dri_options,
754 0, "anv", NULL,
755 instance->app_info.engine_name,
756 instance->app_info.engine_version);
757
758 *pInstance = anv_instance_to_handle(instance);
759
760 return VK_SUCCESS;
761 }
762
763 void anv_DestroyInstance(
764 VkInstance _instance,
765 const VkAllocationCallbacks* pAllocator)
766 {
767 ANV_FROM_HANDLE(anv_instance, instance, _instance);
768
769 if (!instance)
770 return;
771
772 if (instance->physicalDeviceCount > 0) {
773 /* We support at most one physical device. */
774 assert(instance->physicalDeviceCount == 1);
775 anv_physical_device_finish(&instance->physicalDevice);
776 }
777
778 vk_free(&instance->alloc, (char *)instance->app_info.app_name);
779 vk_free(&instance->alloc, (char *)instance->app_info.engine_name);
780
781 VG(VALGRIND_DESTROY_MEMPOOL(instance));
782
783 vk_debug_report_instance_destroy(&instance->debug_report_callbacks);
784
785 glsl_type_singleton_decref();
786
787 driDestroyOptionCache(&instance->dri_options);
788 driDestroyOptionInfo(&instance->available_dri_options);
789
790 vk_free(&instance->alloc, instance);
791 }
792
793 static VkResult
794 anv_enumerate_devices(struct anv_instance *instance)
795 {
796 /* TODO: Check for more devices ? */
797 drmDevicePtr devices[8];
798 VkResult result = VK_ERROR_INCOMPATIBLE_DRIVER;
799 int max_devices;
800
801 instance->physicalDeviceCount = 0;
802
803 max_devices = drmGetDevices2(0, devices, ARRAY_SIZE(devices));
804 if (max_devices < 1)
805 return VK_ERROR_INCOMPATIBLE_DRIVER;
806
807 for (unsigned i = 0; i < (unsigned)max_devices; i++) {
808 if (devices[i]->available_nodes & 1 << DRM_NODE_RENDER &&
809 devices[i]->bustype == DRM_BUS_PCI &&
810 devices[i]->deviceinfo.pci->vendor_id == 0x8086) {
811
812 result = anv_physical_device_init(&instance->physicalDevice,
813 instance, devices[i]);
814 if (result != VK_ERROR_INCOMPATIBLE_DRIVER)
815 break;
816 }
817 }
818 drmFreeDevices(devices, max_devices);
819
820 if (result == VK_SUCCESS)
821 instance->physicalDeviceCount = 1;
822
823 return result;
824 }
825
826 static VkResult
827 anv_instance_ensure_physical_device(struct anv_instance *instance)
828 {
829 if (instance->physicalDeviceCount < 0) {
830 VkResult result = anv_enumerate_devices(instance);
831 if (result != VK_SUCCESS &&
832 result != VK_ERROR_INCOMPATIBLE_DRIVER)
833 return result;
834 }
835
836 return VK_SUCCESS;
837 }
838
839 VkResult anv_EnumeratePhysicalDevices(
840 VkInstance _instance,
841 uint32_t* pPhysicalDeviceCount,
842 VkPhysicalDevice* pPhysicalDevices)
843 {
844 ANV_FROM_HANDLE(anv_instance, instance, _instance);
845 VK_OUTARRAY_MAKE(out, pPhysicalDevices, pPhysicalDeviceCount);
846
847 VkResult result = anv_instance_ensure_physical_device(instance);
848 if (result != VK_SUCCESS)
849 return result;
850
851 if (instance->physicalDeviceCount == 0)
852 return VK_SUCCESS;
853
854 assert(instance->physicalDeviceCount == 1);
855 vk_outarray_append(&out, i) {
856 *i = anv_physical_device_to_handle(&instance->physicalDevice);
857 }
858
859 return vk_outarray_status(&out);
860 }
861
862 VkResult anv_EnumeratePhysicalDeviceGroups(
863 VkInstance _instance,
864 uint32_t* pPhysicalDeviceGroupCount,
865 VkPhysicalDeviceGroupProperties* pPhysicalDeviceGroupProperties)
866 {
867 ANV_FROM_HANDLE(anv_instance, instance, _instance);
868 VK_OUTARRAY_MAKE(out, pPhysicalDeviceGroupProperties,
869 pPhysicalDeviceGroupCount);
870
871 VkResult result = anv_instance_ensure_physical_device(instance);
872 if (result != VK_SUCCESS)
873 return result;
874
875 if (instance->physicalDeviceCount == 0)
876 return VK_SUCCESS;
877
878 assert(instance->physicalDeviceCount == 1);
879
880 vk_outarray_append(&out, p) {
881 p->physicalDeviceCount = 1;
882 memset(p->physicalDevices, 0, sizeof(p->physicalDevices));
883 p->physicalDevices[0] =
884 anv_physical_device_to_handle(&instance->physicalDevice);
885 p->subsetAllocation = false;
886
887 vk_foreach_struct(ext, p->pNext)
888 anv_debug_ignored_stype(ext->sType);
889 }
890
891 return vk_outarray_status(&out);
892 }
893
894 void anv_GetPhysicalDeviceFeatures(
895 VkPhysicalDevice physicalDevice,
896 VkPhysicalDeviceFeatures* pFeatures)
897 {
898 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
899
900 *pFeatures = (VkPhysicalDeviceFeatures) {
901 .robustBufferAccess = true,
902 .fullDrawIndexUint32 = true,
903 .imageCubeArray = true,
904 .independentBlend = true,
905 .geometryShader = true,
906 .tessellationShader = true,
907 .sampleRateShading = true,
908 .dualSrcBlend = true,
909 .logicOp = true,
910 .multiDrawIndirect = true,
911 .drawIndirectFirstInstance = true,
912 .depthClamp = true,
913 .depthBiasClamp = true,
914 .fillModeNonSolid = true,
915 .depthBounds = pdevice->info.gen >= 12,
916 .wideLines = true,
917 .largePoints = true,
918 .alphaToOne = true,
919 .multiViewport = true,
920 .samplerAnisotropy = true,
921 .textureCompressionETC2 = pdevice->info.gen >= 8 ||
922 pdevice->info.is_baytrail,
923 .textureCompressionASTC_LDR = pdevice->info.gen >= 9, /* FINISHME CHV */
924 .textureCompressionBC = true,
925 .occlusionQueryPrecise = true,
926 .pipelineStatisticsQuery = true,
927 .fragmentStoresAndAtomics = true,
928 .shaderTessellationAndGeometryPointSize = true,
929 .shaderImageGatherExtended = true,
930 .shaderStorageImageExtendedFormats = true,
931 .shaderStorageImageMultisample = false,
932 .shaderStorageImageReadWithoutFormat = false,
933 .shaderStorageImageWriteWithoutFormat = true,
934 .shaderUniformBufferArrayDynamicIndexing = true,
935 .shaderSampledImageArrayDynamicIndexing = true,
936 .shaderStorageBufferArrayDynamicIndexing = true,
937 .shaderStorageImageArrayDynamicIndexing = true,
938 .shaderClipDistance = true,
939 .shaderCullDistance = true,
940 .shaderFloat64 = pdevice->info.gen >= 8 &&
941 pdevice->info.has_64bit_types,
942 .shaderInt64 = pdevice->info.gen >= 8 &&
943 pdevice->info.has_64bit_types,
944 .shaderInt16 = pdevice->info.gen >= 8,
945 .shaderResourceMinLod = pdevice->info.gen >= 9,
946 .variableMultisampleRate = true,
947 .inheritedQueries = true,
948 };
949
950 /* We can't do image stores in vec4 shaders */
951 pFeatures->vertexPipelineStoresAndAtomics =
952 pdevice->compiler->scalar_stage[MESA_SHADER_VERTEX] &&
953 pdevice->compiler->scalar_stage[MESA_SHADER_GEOMETRY];
954
955 struct anv_app_info *app_info = &pdevice->instance->app_info;
956
957 /* The new DOOM and Wolfenstein games require depthBounds without
958 * checking for it. They seem to run fine without it so just claim it's
959 * there and accept the consequences.
960 */
961 if (app_info->engine_name && strcmp(app_info->engine_name, "idTech") == 0)
962 pFeatures->depthBounds = true;
963 }
964
965 void anv_GetPhysicalDeviceFeatures2(
966 VkPhysicalDevice physicalDevice,
967 VkPhysicalDeviceFeatures2* pFeatures)
968 {
969 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
970 anv_GetPhysicalDeviceFeatures(physicalDevice, &pFeatures->features);
971
972 vk_foreach_struct(ext, pFeatures->pNext) {
973 switch (ext->sType) {
974 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES_KHR: {
975 VkPhysicalDevice8BitStorageFeaturesKHR *features =
976 (VkPhysicalDevice8BitStorageFeaturesKHR *)ext;
977 features->storageBuffer8BitAccess = pdevice->info.gen >= 8;
978 features->uniformAndStorageBuffer8BitAccess = pdevice->info.gen >= 8;
979 features->storagePushConstant8 = pdevice->info.gen >= 8;
980 break;
981 }
982
983 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES: {
984 VkPhysicalDevice16BitStorageFeatures *features =
985 (VkPhysicalDevice16BitStorageFeatures *)ext;
986 features->storageBuffer16BitAccess = pdevice->info.gen >= 8;
987 features->uniformAndStorageBuffer16BitAccess = pdevice->info.gen >= 8;
988 features->storagePushConstant16 = pdevice->info.gen >= 8;
989 features->storageInputOutput16 = false;
990 break;
991 }
992
993 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_EXT: {
994 VkPhysicalDeviceBufferDeviceAddressFeaturesEXT *features = (void *)ext;
995 features->bufferDeviceAddress = pdevice->has_a64_buffer_access;
996 features->bufferDeviceAddressCaptureReplay = false;
997 features->bufferDeviceAddressMultiDevice = false;
998 break;
999 }
1000
1001 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_KHR: {
1002 VkPhysicalDeviceBufferDeviceAddressFeaturesKHR *features = (void *)ext;
1003 features->bufferDeviceAddress = pdevice->has_a64_buffer_access;
1004 features->bufferDeviceAddressCaptureReplay =
1005 pdevice->has_a64_buffer_access;
1006 features->bufferDeviceAddressMultiDevice = false;
1007 break;
1008 }
1009
1010 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COMPUTE_SHADER_DERIVATIVES_FEATURES_NV: {
1011 VkPhysicalDeviceComputeShaderDerivativesFeaturesNV *features =
1012 (VkPhysicalDeviceComputeShaderDerivativesFeaturesNV *)ext;
1013 features->computeDerivativeGroupQuads = true;
1014 features->computeDerivativeGroupLinear = true;
1015 break;
1016 }
1017
1018 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT: {
1019 VkPhysicalDeviceConditionalRenderingFeaturesEXT *features =
1020 (VkPhysicalDeviceConditionalRenderingFeaturesEXT*)ext;
1021 features->conditionalRendering = pdevice->info.gen >= 8 ||
1022 pdevice->info.is_haswell;
1023 features->inheritedConditionalRendering = pdevice->info.gen >= 8 ||
1024 pdevice->info.is_haswell;
1025 break;
1026 }
1027
1028 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_CLIP_ENABLE_FEATURES_EXT: {
1029 VkPhysicalDeviceDepthClipEnableFeaturesEXT *features =
1030 (VkPhysicalDeviceDepthClipEnableFeaturesEXT *)ext;
1031 features->depthClipEnable = true;
1032 break;
1033 }
1034
1035 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR: {
1036 VkPhysicalDeviceFloat16Int8FeaturesKHR *features = (void *)ext;
1037 features->shaderFloat16 = pdevice->info.gen >= 8;
1038 features->shaderInt8 = pdevice->info.gen >= 8;
1039 break;
1040 }
1041
1042 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FRAGMENT_SHADER_INTERLOCK_FEATURES_EXT: {
1043 VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT *features =
1044 (VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT *)ext;
1045 features->fragmentShaderSampleInterlock = pdevice->info.gen >= 9;
1046 features->fragmentShaderPixelInterlock = pdevice->info.gen >= 9;
1047 features->fragmentShaderShadingRateInterlock = false;
1048 break;
1049 }
1050
1051 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_QUERY_RESET_FEATURES_EXT: {
1052 VkPhysicalDeviceHostQueryResetFeaturesEXT *features =
1053 (VkPhysicalDeviceHostQueryResetFeaturesEXT *)ext;
1054 features->hostQueryReset = true;
1055 break;
1056 }
1057
1058 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT: {
1059 VkPhysicalDeviceDescriptorIndexingFeaturesEXT *features =
1060 (VkPhysicalDeviceDescriptorIndexingFeaturesEXT *)ext;
1061 features->shaderInputAttachmentArrayDynamicIndexing = false;
1062 features->shaderUniformTexelBufferArrayDynamicIndexing = true;
1063 features->shaderStorageTexelBufferArrayDynamicIndexing = true;
1064 features->shaderUniformBufferArrayNonUniformIndexing = false;
1065 features->shaderSampledImageArrayNonUniformIndexing = true;
1066 features->shaderStorageBufferArrayNonUniformIndexing = true;
1067 features->shaderStorageImageArrayNonUniformIndexing = true;
1068 features->shaderInputAttachmentArrayNonUniformIndexing = false;
1069 features->shaderUniformTexelBufferArrayNonUniformIndexing = true;
1070 features->shaderStorageTexelBufferArrayNonUniformIndexing = true;
1071 features->descriptorBindingUniformBufferUpdateAfterBind = false;
1072 features->descriptorBindingSampledImageUpdateAfterBind = true;
1073 features->descriptorBindingStorageImageUpdateAfterBind = true;
1074 features->descriptorBindingStorageBufferUpdateAfterBind = true;
1075 features->descriptorBindingUniformTexelBufferUpdateAfterBind = true;
1076 features->descriptorBindingStorageTexelBufferUpdateAfterBind = true;
1077 features->descriptorBindingUpdateUnusedWhilePending = true;
1078 features->descriptorBindingPartiallyBound = true;
1079 features->descriptorBindingVariableDescriptorCount = false;
1080 features->runtimeDescriptorArray = true;
1081 break;
1082 }
1083
1084 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INDEX_TYPE_UINT8_FEATURES_EXT: {
1085 VkPhysicalDeviceIndexTypeUint8FeaturesEXT *features =
1086 (VkPhysicalDeviceIndexTypeUint8FeaturesEXT *)ext;
1087 features->indexTypeUint8 = true;
1088 break;
1089 }
1090
1091 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_FEATURES_EXT: {
1092 VkPhysicalDeviceInlineUniformBlockFeaturesEXT *features =
1093 (VkPhysicalDeviceInlineUniformBlockFeaturesEXT *)ext;
1094 features->inlineUniformBlock = true;
1095 features->descriptorBindingInlineUniformBlockUpdateAfterBind = true;
1096 break;
1097 }
1098
1099 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_FEATURES_EXT: {
1100 VkPhysicalDeviceLineRasterizationFeaturesEXT *features =
1101 (VkPhysicalDeviceLineRasterizationFeaturesEXT *)ext;
1102 features->rectangularLines = true;
1103 features->bresenhamLines = true;
1104 features->smoothLines = true;
1105 features->stippledRectangularLines = false;
1106 features->stippledBresenhamLines = true;
1107 features->stippledSmoothLines = false;
1108 break;
1109 }
1110
1111 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES: {
1112 VkPhysicalDeviceMultiviewFeatures *features =
1113 (VkPhysicalDeviceMultiviewFeatures *)ext;
1114 features->multiview = true;
1115 features->multiviewGeometryShader = true;
1116 features->multiviewTessellationShader = true;
1117 break;
1118 }
1119
1120 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGELESS_FRAMEBUFFER_FEATURES_KHR: {
1121 VkPhysicalDeviceImagelessFramebufferFeaturesKHR *features =
1122 (VkPhysicalDeviceImagelessFramebufferFeaturesKHR *)ext;
1123 features->imagelessFramebuffer = true;
1124 break;
1125 }
1126
1127 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR: {
1128 VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR *features =
1129 (VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR *)ext;
1130 features->pipelineExecutableInfo = true;
1131 break;
1132 }
1133
1134 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES: {
1135 VkPhysicalDeviceProtectedMemoryFeatures *features = (void *)ext;
1136 features->protectedMemory = false;
1137 break;
1138 }
1139
1140 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES: {
1141 VkPhysicalDeviceSamplerYcbcrConversionFeatures *features =
1142 (VkPhysicalDeviceSamplerYcbcrConversionFeatures *) ext;
1143 features->samplerYcbcrConversion = true;
1144 break;
1145 }
1146
1147 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES_EXT: {
1148 VkPhysicalDeviceScalarBlockLayoutFeaturesEXT *features =
1149 (VkPhysicalDeviceScalarBlockLayoutFeaturesEXT *)ext;
1150 features->scalarBlockLayout = true;
1151 break;
1152 }
1153
1154 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SEPARATE_DEPTH_STENCIL_LAYOUTS_FEATURES_KHR: {
1155 VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR *features =
1156 (VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR *)ext;
1157 features->separateDepthStencilLayouts = true;
1158 break;
1159 }
1160
1161 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_INT64_FEATURES_KHR: {
1162 VkPhysicalDeviceShaderAtomicInt64FeaturesKHR *features = (void *)ext;
1163 features->shaderBufferInt64Atomics =
1164 pdevice->info.gen >= 9 && pdevice->use_softpin;
1165 features->shaderSharedInt64Atomics = VK_FALSE;
1166 break;
1167 }
1168
1169 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DEMOTE_TO_HELPER_INVOCATION_FEATURES_EXT: {
1170 VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT *features = (void *)ext;
1171 features->shaderDemoteToHelperInvocation = true;
1172 break;
1173 }
1174
1175 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CLOCK_FEATURES_KHR: {
1176 VkPhysicalDeviceShaderClockFeaturesKHR *features =
1177 (VkPhysicalDeviceShaderClockFeaturesKHR *)ext;
1178 features->shaderSubgroupClock = true;
1179 features->shaderDeviceClock = false;
1180 break;
1181 }
1182
1183 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETERS_FEATURES: {
1184 VkPhysicalDeviceShaderDrawParametersFeatures *features = (void *)ext;
1185 features->shaderDrawParameters = true;
1186 break;
1187 }
1188
1189 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_SUBGROUP_EXTENDED_TYPES_FEATURES_KHR: {
1190 VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR *features =
1191 (VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR *)ext;
1192 features->shaderSubgroupExtendedTypes = true;
1193 break;
1194 }
1195
1196 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_FEATURES_EXT: {
1197 VkPhysicalDeviceSubgroupSizeControlFeaturesEXT *features =
1198 (VkPhysicalDeviceSubgroupSizeControlFeaturesEXT *)ext;
1199 features->subgroupSizeControl = true;
1200 features->computeFullSubgroups = true;
1201 break;
1202 }
1203
1204 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_FEATURES_EXT: {
1205 VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT *features =
1206 (VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT *)ext;
1207 features->texelBufferAlignment = true;
1208 break;
1209 }
1210
1211 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_FEATURES_KHR: {
1212 VkPhysicalDeviceTimelineSemaphoreFeaturesKHR *features =
1213 (VkPhysicalDeviceTimelineSemaphoreFeaturesKHR *) ext;
1214 features->timelineSemaphore = true;
1215 break;
1216 }
1217
1218 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES: {
1219 VkPhysicalDeviceVariablePointersFeatures *features = (void *)ext;
1220 features->variablePointersStorageBuffer = true;
1221 features->variablePointers = true;
1222 break;
1223 }
1224
1225 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT: {
1226 VkPhysicalDeviceTransformFeedbackFeaturesEXT *features =
1227 (VkPhysicalDeviceTransformFeedbackFeaturesEXT *)ext;
1228 features->transformFeedback = true;
1229 features->geometryStreams = true;
1230 break;
1231 }
1232
1233 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_UNIFORM_BUFFER_STANDARD_LAYOUT_FEATURES_KHR: {
1234 VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR *features =
1235 (VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR *)ext;
1236 features->uniformBufferStandardLayout = true;
1237 break;
1238 }
1239
1240 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT: {
1241 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *features =
1242 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *)ext;
1243 features->vertexAttributeInstanceRateDivisor = true;
1244 features->vertexAttributeInstanceRateZeroDivisor = true;
1245 break;
1246 }
1247
1248 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_MEMORY_MODEL_FEATURES_KHR: {
1249 VkPhysicalDeviceVulkanMemoryModelFeaturesKHR *features = (void *)ext;
1250 features->vulkanMemoryModel = true;
1251 features->vulkanMemoryModelDeviceScope = true;
1252 features->vulkanMemoryModelAvailabilityVisibilityChains = true;
1253 break;
1254 }
1255
1256 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_YCBCR_IMAGE_ARRAYS_FEATURES_EXT: {
1257 VkPhysicalDeviceYcbcrImageArraysFeaturesEXT *features =
1258 (VkPhysicalDeviceYcbcrImageArraysFeaturesEXT *)ext;
1259 features->ycbcrImageArrays = true;
1260 break;
1261 }
1262
1263 default:
1264 anv_debug_ignored_stype(ext->sType);
1265 break;
1266 }
1267 }
1268 }
1269
1270 #define MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS 64
1271
1272 #define MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS 64
1273 #define MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS 256
1274
1275 void anv_GetPhysicalDeviceProperties(
1276 VkPhysicalDevice physicalDevice,
1277 VkPhysicalDeviceProperties* pProperties)
1278 {
1279 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
1280 const struct gen_device_info *devinfo = &pdevice->info;
1281
1282 /* See assertions made when programming the buffer surface state. */
1283 const uint32_t max_raw_buffer_sz = devinfo->gen >= 7 ?
1284 (1ul << 30) : (1ul << 27);
1285
1286 const uint32_t max_ssbos = pdevice->has_a64_buffer_access ? UINT16_MAX : 64;
1287 const uint32_t max_textures =
1288 pdevice->has_bindless_images ? UINT16_MAX : 128;
1289 const uint32_t max_samplers =
1290 pdevice->has_bindless_samplers ? UINT16_MAX :
1291 (devinfo->gen >= 8 || devinfo->is_haswell) ? 128 : 16;
1292 const uint32_t max_images =
1293 pdevice->has_bindless_images ? UINT16_MAX : MAX_IMAGES;
1294
1295 /* If we can use bindless for everything, claim a high per-stage limit,
1296 * otherwise use the binding table size, minus the slots reserved for
1297 * render targets and one slot for the descriptor buffer. */
1298 const uint32_t max_per_stage =
1299 pdevice->has_bindless_images && pdevice->has_a64_buffer_access
1300 ? UINT32_MAX : MAX_BINDING_TABLE_SIZE - MAX_RTS - 1;
1301
1302 const uint32_t max_workgroup_size = 32 * devinfo->max_cs_threads;
1303
1304 VkSampleCountFlags sample_counts =
1305 isl_device_get_sample_counts(&pdevice->isl_dev);
1306
1307
1308 VkPhysicalDeviceLimits limits = {
1309 .maxImageDimension1D = (1 << 14),
1310 .maxImageDimension2D = (1 << 14),
1311 .maxImageDimension3D = (1 << 11),
1312 .maxImageDimensionCube = (1 << 14),
1313 .maxImageArrayLayers = (1 << 11),
1314 .maxTexelBufferElements = 128 * 1024 * 1024,
1315 .maxUniformBufferRange = (1ul << 27),
1316 .maxStorageBufferRange = max_raw_buffer_sz,
1317 .maxPushConstantsSize = MAX_PUSH_CONSTANTS_SIZE,
1318 .maxMemoryAllocationCount = UINT32_MAX,
1319 .maxSamplerAllocationCount = 64 * 1024,
1320 .bufferImageGranularity = 64, /* A cache line */
1321 .sparseAddressSpaceSize = 0,
1322 .maxBoundDescriptorSets = MAX_SETS,
1323 .maxPerStageDescriptorSamplers = max_samplers,
1324 .maxPerStageDescriptorUniformBuffers = MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS,
1325 .maxPerStageDescriptorStorageBuffers = max_ssbos,
1326 .maxPerStageDescriptorSampledImages = max_textures,
1327 .maxPerStageDescriptorStorageImages = max_images,
1328 .maxPerStageDescriptorInputAttachments = MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS,
1329 .maxPerStageResources = max_per_stage,
1330 .maxDescriptorSetSamplers = 6 * max_samplers, /* number of stages * maxPerStageDescriptorSamplers */
1331 .maxDescriptorSetUniformBuffers = 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS, /* number of stages * maxPerStageDescriptorUniformBuffers */
1332 .maxDescriptorSetUniformBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2,
1333 .maxDescriptorSetStorageBuffers = 6 * max_ssbos, /* number of stages * maxPerStageDescriptorStorageBuffers */
1334 .maxDescriptorSetStorageBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2,
1335 .maxDescriptorSetSampledImages = 6 * max_textures, /* number of stages * maxPerStageDescriptorSampledImages */
1336 .maxDescriptorSetStorageImages = 6 * max_images, /* number of stages * maxPerStageDescriptorStorageImages */
1337 .maxDescriptorSetInputAttachments = MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS,
1338 .maxVertexInputAttributes = MAX_VBS,
1339 .maxVertexInputBindings = MAX_VBS,
1340 .maxVertexInputAttributeOffset = 2047,
1341 .maxVertexInputBindingStride = 2048,
1342 .maxVertexOutputComponents = 128,
1343 .maxTessellationGenerationLevel = 64,
1344 .maxTessellationPatchSize = 32,
1345 .maxTessellationControlPerVertexInputComponents = 128,
1346 .maxTessellationControlPerVertexOutputComponents = 128,
1347 .maxTessellationControlPerPatchOutputComponents = 128,
1348 .maxTessellationControlTotalOutputComponents = 2048,
1349 .maxTessellationEvaluationInputComponents = 128,
1350 .maxTessellationEvaluationOutputComponents = 128,
1351 .maxGeometryShaderInvocations = 32,
1352 .maxGeometryInputComponents = 64,
1353 .maxGeometryOutputComponents = 128,
1354 .maxGeometryOutputVertices = 256,
1355 .maxGeometryTotalOutputComponents = 1024,
1356 .maxFragmentInputComponents = 116, /* 128 components - (PSIZ, CLIP_DIST0, CLIP_DIST1) */
1357 .maxFragmentOutputAttachments = 8,
1358 .maxFragmentDualSrcAttachments = 1,
1359 .maxFragmentCombinedOutputResources = 8,
1360 .maxComputeSharedMemorySize = 64 * 1024,
1361 .maxComputeWorkGroupCount = { 65535, 65535, 65535 },
1362 .maxComputeWorkGroupInvocations = max_workgroup_size,
1363 .maxComputeWorkGroupSize = {
1364 max_workgroup_size,
1365 max_workgroup_size,
1366 max_workgroup_size,
1367 },
1368 .subPixelPrecisionBits = 8,
1369 .subTexelPrecisionBits = 8,
1370 .mipmapPrecisionBits = 8,
1371 .maxDrawIndexedIndexValue = UINT32_MAX,
1372 .maxDrawIndirectCount = UINT32_MAX,
1373 .maxSamplerLodBias = 16,
1374 .maxSamplerAnisotropy = 16,
1375 .maxViewports = MAX_VIEWPORTS,
1376 .maxViewportDimensions = { (1 << 14), (1 << 14) },
1377 .viewportBoundsRange = { INT16_MIN, INT16_MAX },
1378 .viewportSubPixelBits = 13, /* We take a float? */
1379 .minMemoryMapAlignment = 4096, /* A page */
1380 /* The dataport requires texel alignment so we need to assume a worst
1381 * case of R32G32B32A32 which is 16 bytes.
1382 */
1383 .minTexelBufferOffsetAlignment = 16,
1384 /* We need 16 for UBO block reads to work and 32 for push UBOs */
1385 .minUniformBufferOffsetAlignment = 32,
1386 .minStorageBufferOffsetAlignment = 4,
1387 .minTexelOffset = -8,
1388 .maxTexelOffset = 7,
1389 .minTexelGatherOffset = -32,
1390 .maxTexelGatherOffset = 31,
1391 .minInterpolationOffset = -0.5,
1392 .maxInterpolationOffset = 0.4375,
1393 .subPixelInterpolationOffsetBits = 4,
1394 .maxFramebufferWidth = (1 << 14),
1395 .maxFramebufferHeight = (1 << 14),
1396 .maxFramebufferLayers = (1 << 11),
1397 .framebufferColorSampleCounts = sample_counts,
1398 .framebufferDepthSampleCounts = sample_counts,
1399 .framebufferStencilSampleCounts = sample_counts,
1400 .framebufferNoAttachmentsSampleCounts = sample_counts,
1401 .maxColorAttachments = MAX_RTS,
1402 .sampledImageColorSampleCounts = sample_counts,
1403 .sampledImageIntegerSampleCounts = VK_SAMPLE_COUNT_1_BIT,
1404 .sampledImageDepthSampleCounts = sample_counts,
1405 .sampledImageStencilSampleCounts = sample_counts,
1406 .storageImageSampleCounts = VK_SAMPLE_COUNT_1_BIT,
1407 .maxSampleMaskWords = 1,
1408 .timestampComputeAndGraphics = true,
1409 .timestampPeriod = 1000000000.0 / devinfo->timestamp_frequency,
1410 .maxClipDistances = 8,
1411 .maxCullDistances = 8,
1412 .maxCombinedClipAndCullDistances = 8,
1413 .discreteQueuePriorities = 2,
1414 .pointSizeRange = { 0.125, 255.875 },
1415 .lineWidthRange = {
1416 0.0,
1417 (devinfo->gen >= 9 || devinfo->is_cherryview) ?
1418 2047.9921875 : 7.9921875,
1419 },
1420 .pointSizeGranularity = (1.0 / 8.0),
1421 .lineWidthGranularity = (1.0 / 128.0),
1422 .strictLines = false,
1423 .standardSampleLocations = true,
1424 .optimalBufferCopyOffsetAlignment = 128,
1425 .optimalBufferCopyRowPitchAlignment = 128,
1426 .nonCoherentAtomSize = 64,
1427 };
1428
1429 *pProperties = (VkPhysicalDeviceProperties) {
1430 .apiVersion = anv_physical_device_api_version(pdevice),
1431 .driverVersion = vk_get_driver_version(),
1432 .vendorID = 0x8086,
1433 .deviceID = pdevice->chipset_id,
1434 .deviceType = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
1435 .limits = limits,
1436 .sparseProperties = {0}, /* Broadwell doesn't do sparse. */
1437 };
1438
1439 snprintf(pProperties->deviceName, sizeof(pProperties->deviceName),
1440 "%s", pdevice->name);
1441 memcpy(pProperties->pipelineCacheUUID,
1442 pdevice->pipeline_cache_uuid, VK_UUID_SIZE);
1443 }
1444
1445 void anv_GetPhysicalDeviceProperties2(
1446 VkPhysicalDevice physicalDevice,
1447 VkPhysicalDeviceProperties2* pProperties)
1448 {
1449 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
1450
1451 anv_GetPhysicalDeviceProperties(physicalDevice, &pProperties->properties);
1452
1453 vk_foreach_struct(ext, pProperties->pNext) {
1454 switch (ext->sType) {
1455 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_STENCIL_RESOLVE_PROPERTIES_KHR: {
1456 VkPhysicalDeviceDepthStencilResolvePropertiesKHR *props =
1457 (VkPhysicalDeviceDepthStencilResolvePropertiesKHR *)ext;
1458
1459 /* We support all of the depth resolve modes */
1460 props->supportedDepthResolveModes =
1461 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR |
1462 VK_RESOLVE_MODE_AVERAGE_BIT_KHR |
1463 VK_RESOLVE_MODE_MIN_BIT_KHR |
1464 VK_RESOLVE_MODE_MAX_BIT_KHR;
1465
1466 /* Average doesn't make sense for stencil so we don't support that */
1467 props->supportedStencilResolveModes =
1468 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR;
1469 if (pdevice->info.gen >= 8) {
1470 /* The advanced stencil resolve modes currently require stencil
1471 * sampling be supported by the hardware.
1472 */
1473 props->supportedStencilResolveModes |=
1474 VK_RESOLVE_MODE_MIN_BIT_KHR |
1475 VK_RESOLVE_MODE_MAX_BIT_KHR;
1476 }
1477
1478 props->independentResolveNone = true;
1479 props->independentResolve = true;
1480 break;
1481 }
1482
1483 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES_EXT: {
1484 VkPhysicalDeviceDescriptorIndexingPropertiesEXT *props =
1485 (VkPhysicalDeviceDescriptorIndexingPropertiesEXT *)ext;
1486
1487 /* It's a bit hard to exactly map our implementation to the limits
1488 * described here. The bindless surface handle in the extended
1489 * message descriptors is 20 bits and it's an index into the table of
1490 * RENDER_SURFACE_STATE structs that starts at bindless surface base
1491 * address. Given that most things consume two surface states per
1492 * view (general/sampled for textures and write-only/read-write for
1493 * images), we claim 2^19 things.
1494 *
1495 * For SSBOs, we just use A64 messages so there is no real limit
1496 * there beyond the limit on the total size of a descriptor set.
1497 */
1498 const unsigned max_bindless_views = 1 << 19;
1499
1500 props->maxUpdateAfterBindDescriptorsInAllPools = max_bindless_views;
1501 props->shaderUniformBufferArrayNonUniformIndexingNative = false;
1502 props->shaderSampledImageArrayNonUniformIndexingNative = false;
1503 props->shaderStorageBufferArrayNonUniformIndexingNative = true;
1504 props->shaderStorageImageArrayNonUniformIndexingNative = false;
1505 props->shaderInputAttachmentArrayNonUniformIndexingNative = false;
1506 props->robustBufferAccessUpdateAfterBind = true;
1507 props->quadDivergentImplicitLod = false;
1508 props->maxPerStageDescriptorUpdateAfterBindSamplers = max_bindless_views;
1509 props->maxPerStageDescriptorUpdateAfterBindUniformBuffers = MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS;
1510 props->maxPerStageDescriptorUpdateAfterBindStorageBuffers = UINT32_MAX;
1511 props->maxPerStageDescriptorUpdateAfterBindSampledImages = max_bindless_views;
1512 props->maxPerStageDescriptorUpdateAfterBindStorageImages = max_bindless_views;
1513 props->maxPerStageDescriptorUpdateAfterBindInputAttachments = MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS;
1514 props->maxPerStageUpdateAfterBindResources = UINT32_MAX;
1515 props->maxDescriptorSetUpdateAfterBindSamplers = max_bindless_views;
1516 props->maxDescriptorSetUpdateAfterBindUniformBuffers = 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS;
1517 props->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2;
1518 props->maxDescriptorSetUpdateAfterBindStorageBuffers = UINT32_MAX;
1519 props->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2;
1520 props->maxDescriptorSetUpdateAfterBindSampledImages = max_bindless_views;
1521 props->maxDescriptorSetUpdateAfterBindStorageImages = max_bindless_views;
1522 props->maxDescriptorSetUpdateAfterBindInputAttachments = MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS;
1523 break;
1524 }
1525
1526 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR: {
1527 VkPhysicalDeviceDriverPropertiesKHR *driver_props =
1528 (VkPhysicalDeviceDriverPropertiesKHR *) ext;
1529
1530 driver_props->driverID = VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA_KHR;
1531 snprintf(driver_props->driverName, VK_MAX_DRIVER_NAME_SIZE_KHR,
1532 "Intel open-source Mesa driver");
1533
1534 snprintf(driver_props->driverInfo, VK_MAX_DRIVER_INFO_SIZE_KHR,
1535 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1);
1536
1537 driver_props->conformanceVersion = (VkConformanceVersionKHR) {
1538 .major = 1,
1539 .minor = 1,
1540 .subminor = 2,
1541 .patch = 0,
1542 };
1543 break;
1544 }
1545
1546 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT: {
1547 VkPhysicalDeviceExternalMemoryHostPropertiesEXT *props =
1548 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT *) ext;
1549 /* Userptr needs page aligned memory. */
1550 props->minImportedHostPointerAlignment = 4096;
1551 break;
1552 }
1553
1554 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES: {
1555 VkPhysicalDeviceIDProperties *id_props =
1556 (VkPhysicalDeviceIDProperties *)ext;
1557 memcpy(id_props->deviceUUID, pdevice->device_uuid, VK_UUID_SIZE);
1558 memcpy(id_props->driverUUID, pdevice->driver_uuid, VK_UUID_SIZE);
1559 /* The LUID is for Windows. */
1560 id_props->deviceLUIDValid = false;
1561 break;
1562 }
1563
1564 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_PROPERTIES_EXT: {
1565 VkPhysicalDeviceInlineUniformBlockPropertiesEXT *props =
1566 (VkPhysicalDeviceInlineUniformBlockPropertiesEXT *)ext;
1567 props->maxInlineUniformBlockSize = MAX_INLINE_UNIFORM_BLOCK_SIZE;
1568 props->maxPerStageDescriptorInlineUniformBlocks =
1569 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1570 props->maxPerStageDescriptorUpdateAfterBindInlineUniformBlocks =
1571 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1572 props->maxDescriptorSetInlineUniformBlocks =
1573 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1574 props->maxDescriptorSetUpdateAfterBindInlineUniformBlocks =
1575 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1576 break;
1577 }
1578
1579 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_PROPERTIES_EXT: {
1580 VkPhysicalDeviceLineRasterizationPropertiesEXT *props =
1581 (VkPhysicalDeviceLineRasterizationPropertiesEXT *)ext;
1582 /* In the Skylake PRM Vol. 7, subsection titled "GIQ (Diamond)
1583 * Sampling Rules - Legacy Mode", it says the following:
1584 *
1585 * "Note that the device divides a pixel into a 16x16 array of
1586 * subpixels, referenced by their upper left corners."
1587 *
1588 * This is the only known reference in the PRMs to the subpixel
1589 * precision of line rasterization and a "16x16 array of subpixels"
1590 * implies 4 subpixel precision bits. Empirical testing has shown
1591 * that 4 subpixel precision bits applies to all line rasterization
1592 * types.
1593 */
1594 props->lineSubPixelPrecisionBits = 4;
1595 break;
1596 }
1597
1598 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES: {
1599 VkPhysicalDeviceMaintenance3Properties *props =
1600 (VkPhysicalDeviceMaintenance3Properties *)ext;
1601 /* This value doesn't matter for us today as our per-stage
1602 * descriptors are the real limit.
1603 */
1604 props->maxPerSetDescriptors = 1024;
1605 props->maxMemoryAllocationSize = MAX_MEMORY_ALLOCATION_SIZE;
1606 break;
1607 }
1608
1609 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES: {
1610 VkPhysicalDeviceMultiviewProperties *properties =
1611 (VkPhysicalDeviceMultiviewProperties *)ext;
1612 properties->maxMultiviewViewCount = 16;
1613 properties->maxMultiviewInstanceIndex = UINT32_MAX / 16;
1614 break;
1615 }
1616
1617 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT: {
1618 VkPhysicalDevicePCIBusInfoPropertiesEXT *properties =
1619 (VkPhysicalDevicePCIBusInfoPropertiesEXT *)ext;
1620 properties->pciDomain = pdevice->pci_info.domain;
1621 properties->pciBus = pdevice->pci_info.bus;
1622 properties->pciDevice = pdevice->pci_info.device;
1623 properties->pciFunction = pdevice->pci_info.function;
1624 break;
1625 }
1626
1627 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES: {
1628 VkPhysicalDevicePointClippingProperties *properties =
1629 (VkPhysicalDevicePointClippingProperties *) ext;
1630 properties->pointClippingBehavior = VK_POINT_CLIPPING_BEHAVIOR_USER_CLIP_PLANES_ONLY;
1631 break;
1632 }
1633
1634 #pragma GCC diagnostic push
1635 #pragma GCC diagnostic ignored "-Wswitch"
1636 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PRESENTATION_PROPERTIES_ANDROID: {
1637 VkPhysicalDevicePresentationPropertiesANDROID *props =
1638 (VkPhysicalDevicePresentationPropertiesANDROID *)ext;
1639 props->sharedImage = VK_FALSE;
1640 break;
1641 }
1642 #pragma GCC diagnostic pop
1643
1644 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES: {
1645 VkPhysicalDeviceProtectedMemoryProperties *props =
1646 (VkPhysicalDeviceProtectedMemoryProperties *)ext;
1647 props->protectedNoFault = false;
1648 break;
1649 }
1650
1651 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR: {
1652 VkPhysicalDevicePushDescriptorPropertiesKHR *properties =
1653 (VkPhysicalDevicePushDescriptorPropertiesKHR *) ext;
1654
1655 properties->maxPushDescriptors = MAX_PUSH_DESCRIPTORS;
1656 break;
1657 }
1658
1659 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT: {
1660 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT *properties =
1661 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT *)ext;
1662 properties->filterMinmaxImageComponentMapping = pdevice->info.gen >= 9;
1663 properties->filterMinmaxSingleComponentFormats = true;
1664 break;
1665 }
1666
1667 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES: {
1668 VkPhysicalDeviceSubgroupProperties *properties = (void *)ext;
1669
1670 properties->subgroupSize = BRW_SUBGROUP_SIZE;
1671
1672 VkShaderStageFlags scalar_stages = 0;
1673 for (unsigned stage = 0; stage < MESA_SHADER_STAGES; stage++) {
1674 if (pdevice->compiler->scalar_stage[stage])
1675 scalar_stages |= mesa_to_vk_shader_stage(stage);
1676 }
1677 properties->supportedStages = scalar_stages;
1678
1679 properties->supportedOperations = VK_SUBGROUP_FEATURE_BASIC_BIT |
1680 VK_SUBGROUP_FEATURE_VOTE_BIT |
1681 VK_SUBGROUP_FEATURE_BALLOT_BIT |
1682 VK_SUBGROUP_FEATURE_SHUFFLE_BIT |
1683 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT |
1684 VK_SUBGROUP_FEATURE_QUAD_BIT;
1685 if (pdevice->info.gen >= 8) {
1686 /* TODO: There's no technical reason why these can't be made to
1687 * work on gen7 but they don't at the moment so it's best to leave
1688 * the feature disabled than enabled and broken.
1689 */
1690 properties->supportedOperations |=
1691 VK_SUBGROUP_FEATURE_ARITHMETIC_BIT |
1692 VK_SUBGROUP_FEATURE_CLUSTERED_BIT;
1693 }
1694 properties->quadOperationsInAllStages = pdevice->info.gen >= 8;
1695 break;
1696 }
1697
1698 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_PROPERTIES_EXT: {
1699 VkPhysicalDeviceSubgroupSizeControlPropertiesEXT *props =
1700 (VkPhysicalDeviceSubgroupSizeControlPropertiesEXT *)ext;
1701 STATIC_ASSERT(8 <= BRW_SUBGROUP_SIZE && BRW_SUBGROUP_SIZE <= 32);
1702 props->minSubgroupSize = 8;
1703 props->maxSubgroupSize = 32;
1704 props->maxComputeWorkgroupSubgroups = pdevice->info.max_cs_threads;
1705 props->requiredSubgroupSizeStages = VK_SHADER_STAGE_COMPUTE_BIT;
1706 break;
1707 }
1708 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT_CONTROLS_PROPERTIES_KHR : {
1709 VkPhysicalDeviceFloatControlsPropertiesKHR *properties = (void *)ext;
1710 properties->denormBehaviorIndependence = VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_ALL_KHR;
1711 properties->roundingModeIndependence = VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_NONE_KHR;
1712
1713 /* Broadwell does not support HF denorms and there are restrictions
1714 * other gens. According to Kabylake's PRM:
1715 *
1716 * "math - Extended Math Function
1717 * [...]
1718 * Restriction : Half-float denorms are always retained."
1719 */
1720 properties->shaderDenormFlushToZeroFloat16 = false;
1721 properties->shaderDenormPreserveFloat16 = pdevice->info.gen > 8;
1722 properties->shaderRoundingModeRTEFloat16 = true;
1723 properties->shaderRoundingModeRTZFloat16 = true;
1724 properties->shaderSignedZeroInfNanPreserveFloat16 = true;
1725
1726 properties->shaderDenormFlushToZeroFloat32 = true;
1727 properties->shaderDenormPreserveFloat32 = true;
1728 properties->shaderRoundingModeRTEFloat32 = true;
1729 properties->shaderRoundingModeRTZFloat32 = true;
1730 properties->shaderSignedZeroInfNanPreserveFloat32 = true;
1731
1732 properties->shaderDenormFlushToZeroFloat64 = true;
1733 properties->shaderDenormPreserveFloat64 = true;
1734 properties->shaderRoundingModeRTEFloat64 = true;
1735 properties->shaderRoundingModeRTZFloat64 = true;
1736 properties->shaderSignedZeroInfNanPreserveFloat64 = true;
1737 break;
1738 }
1739
1740 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_PROPERTIES_EXT: {
1741 VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT *props =
1742 (VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT *)ext;
1743
1744 /* From the SKL PRM Vol. 2d, docs for RENDER_SURFACE_STATE::Surface
1745 * Base Address:
1746 *
1747 * "For SURFTYPE_BUFFER non-rendertarget surfaces, this field
1748 * specifies the base address of the first element of the surface,
1749 * computed in software by adding the surface base address to the
1750 * byte offset of the element in the buffer. The base address must
1751 * be aligned to element size."
1752 *
1753 * The typed dataport messages require that things be texel aligned.
1754 * Otherwise, we may just load/store the wrong data or, in the worst
1755 * case, there may be hangs.
1756 */
1757 props->storageTexelBufferOffsetAlignmentBytes = 16;
1758 props->storageTexelBufferOffsetSingleTexelAlignment = true;
1759
1760 /* The sampler, however, is much more forgiving and it can handle
1761 * arbitrary byte alignment for linear and buffer surfaces. It's
1762 * hard to find a good PRM citation for this but years of empirical
1763 * experience demonstrate that this is true.
1764 */
1765 props->uniformTexelBufferOffsetAlignmentBytes = 1;
1766 props->uniformTexelBufferOffsetSingleTexelAlignment = false;
1767 break;
1768 }
1769
1770 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_PROPERTIES_KHR: {
1771 VkPhysicalDeviceTimelineSemaphorePropertiesKHR *props =
1772 (VkPhysicalDeviceTimelineSemaphorePropertiesKHR *) ext;
1773 props->maxTimelineSemaphoreValueDifference = UINT64_MAX;
1774 break;
1775 }
1776
1777 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT: {
1778 VkPhysicalDeviceTransformFeedbackPropertiesEXT *props =
1779 (VkPhysicalDeviceTransformFeedbackPropertiesEXT *)ext;
1780
1781 props->maxTransformFeedbackStreams = MAX_XFB_STREAMS;
1782 props->maxTransformFeedbackBuffers = MAX_XFB_BUFFERS;
1783 props->maxTransformFeedbackBufferSize = (1ull << 32);
1784 props->maxTransformFeedbackStreamDataSize = 128 * 4;
1785 props->maxTransformFeedbackBufferDataSize = 128 * 4;
1786 props->maxTransformFeedbackBufferDataStride = 2048;
1787 props->transformFeedbackQueries = true;
1788 props->transformFeedbackStreamsLinesTriangles = false;
1789 props->transformFeedbackRasterizationStreamSelect = false;
1790 props->transformFeedbackDraw = true;
1791 break;
1792 }
1793
1794 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT: {
1795 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *props =
1796 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *)ext;
1797 /* We have to restrict this a bit for multiview */
1798 props->maxVertexAttribDivisor = UINT32_MAX / 16;
1799 break;
1800 }
1801
1802 default:
1803 anv_debug_ignored_stype(ext->sType);
1804 break;
1805 }
1806 }
1807 }
1808
1809 /* We support exactly one queue family. */
1810 static const VkQueueFamilyProperties
1811 anv_queue_family_properties = {
1812 .queueFlags = VK_QUEUE_GRAPHICS_BIT |
1813 VK_QUEUE_COMPUTE_BIT |
1814 VK_QUEUE_TRANSFER_BIT,
1815 .queueCount = 1,
1816 .timestampValidBits = 36, /* XXX: Real value here */
1817 .minImageTransferGranularity = { 1, 1, 1 },
1818 };
1819
1820 void anv_GetPhysicalDeviceQueueFamilyProperties(
1821 VkPhysicalDevice physicalDevice,
1822 uint32_t* pCount,
1823 VkQueueFamilyProperties* pQueueFamilyProperties)
1824 {
1825 VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pCount);
1826
1827 vk_outarray_append(&out, p) {
1828 *p = anv_queue_family_properties;
1829 }
1830 }
1831
1832 void anv_GetPhysicalDeviceQueueFamilyProperties2(
1833 VkPhysicalDevice physicalDevice,
1834 uint32_t* pQueueFamilyPropertyCount,
1835 VkQueueFamilyProperties2* pQueueFamilyProperties)
1836 {
1837
1838 VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pQueueFamilyPropertyCount);
1839
1840 vk_outarray_append(&out, p) {
1841 p->queueFamilyProperties = anv_queue_family_properties;
1842
1843 vk_foreach_struct(s, p->pNext) {
1844 anv_debug_ignored_stype(s->sType);
1845 }
1846 }
1847 }
1848
1849 void anv_GetPhysicalDeviceMemoryProperties(
1850 VkPhysicalDevice physicalDevice,
1851 VkPhysicalDeviceMemoryProperties* pMemoryProperties)
1852 {
1853 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
1854
1855 pMemoryProperties->memoryTypeCount = physical_device->memory.type_count;
1856 for (uint32_t i = 0; i < physical_device->memory.type_count; i++) {
1857 pMemoryProperties->memoryTypes[i] = (VkMemoryType) {
1858 .propertyFlags = physical_device->memory.types[i].propertyFlags,
1859 .heapIndex = physical_device->memory.types[i].heapIndex,
1860 };
1861 }
1862
1863 pMemoryProperties->memoryHeapCount = physical_device->memory.heap_count;
1864 for (uint32_t i = 0; i < physical_device->memory.heap_count; i++) {
1865 pMemoryProperties->memoryHeaps[i] = (VkMemoryHeap) {
1866 .size = physical_device->memory.heaps[i].size,
1867 .flags = physical_device->memory.heaps[i].flags,
1868 };
1869 }
1870 }
1871
1872 static void
1873 anv_get_memory_budget(VkPhysicalDevice physicalDevice,
1874 VkPhysicalDeviceMemoryBudgetPropertiesEXT *memoryBudget)
1875 {
1876 ANV_FROM_HANDLE(anv_physical_device, device, physicalDevice);
1877 uint64_t sys_available = get_available_system_memory();
1878 assert(sys_available > 0);
1879
1880 VkDeviceSize total_heaps_size = 0;
1881 for (size_t i = 0; i < device->memory.heap_count; i++)
1882 total_heaps_size += device->memory.heaps[i].size;
1883
1884 for (size_t i = 0; i < device->memory.heap_count; i++) {
1885 VkDeviceSize heap_size = device->memory.heaps[i].size;
1886 VkDeviceSize heap_used = device->memory.heaps[i].used;
1887 VkDeviceSize heap_budget;
1888
1889 double heap_proportion = (double) heap_size / total_heaps_size;
1890 VkDeviceSize sys_available_prop = sys_available * heap_proportion;
1891
1892 /*
1893 * Let's not incite the app to starve the system: report at most 90% of
1894 * available system memory.
1895 */
1896 uint64_t heap_available = sys_available_prop * 9 / 10;
1897 heap_budget = MIN2(heap_size, heap_used + heap_available);
1898
1899 /*
1900 * Round down to the nearest MB
1901 */
1902 heap_budget &= ~((1ull << 20) - 1);
1903
1904 /*
1905 * The heapBudget value must be non-zero for array elements less than
1906 * VkPhysicalDeviceMemoryProperties::memoryHeapCount. The heapBudget
1907 * value must be less than or equal to VkMemoryHeap::size for each heap.
1908 */
1909 assert(0 < heap_budget && heap_budget <= heap_size);
1910
1911 memoryBudget->heapUsage[i] = heap_used;
1912 memoryBudget->heapBudget[i] = heap_budget;
1913 }
1914
1915 /* The heapBudget and heapUsage values must be zero for array elements
1916 * greater than or equal to VkPhysicalDeviceMemoryProperties::memoryHeapCount
1917 */
1918 for (uint32_t i = device->memory.heap_count; i < VK_MAX_MEMORY_HEAPS; i++) {
1919 memoryBudget->heapBudget[i] = 0;
1920 memoryBudget->heapUsage[i] = 0;
1921 }
1922 }
1923
1924 void anv_GetPhysicalDeviceMemoryProperties2(
1925 VkPhysicalDevice physicalDevice,
1926 VkPhysicalDeviceMemoryProperties2* pMemoryProperties)
1927 {
1928 anv_GetPhysicalDeviceMemoryProperties(physicalDevice,
1929 &pMemoryProperties->memoryProperties);
1930
1931 vk_foreach_struct(ext, pMemoryProperties->pNext) {
1932 switch (ext->sType) {
1933 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT:
1934 anv_get_memory_budget(physicalDevice, (void*)ext);
1935 break;
1936 default:
1937 anv_debug_ignored_stype(ext->sType);
1938 break;
1939 }
1940 }
1941 }
1942
1943 void
1944 anv_GetDeviceGroupPeerMemoryFeatures(
1945 VkDevice device,
1946 uint32_t heapIndex,
1947 uint32_t localDeviceIndex,
1948 uint32_t remoteDeviceIndex,
1949 VkPeerMemoryFeatureFlags* pPeerMemoryFeatures)
1950 {
1951 assert(localDeviceIndex == 0 && remoteDeviceIndex == 0);
1952 *pPeerMemoryFeatures = VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT |
1953 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT |
1954 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT |
1955 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT;
1956 }
1957
1958 PFN_vkVoidFunction anv_GetInstanceProcAddr(
1959 VkInstance _instance,
1960 const char* pName)
1961 {
1962 ANV_FROM_HANDLE(anv_instance, instance, _instance);
1963
1964 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
1965 * when we have to return valid function pointers, NULL, or it's left
1966 * undefined. See the table for exact details.
1967 */
1968 if (pName == NULL)
1969 return NULL;
1970
1971 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
1972 if (strcmp(pName, "vk" #entrypoint) == 0) \
1973 return (PFN_vkVoidFunction)anv_##entrypoint
1974
1975 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties);
1976 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties);
1977 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceVersion);
1978 LOOKUP_ANV_ENTRYPOINT(CreateInstance);
1979
1980 #undef LOOKUP_ANV_ENTRYPOINT
1981
1982 if (instance == NULL)
1983 return NULL;
1984
1985 int idx = anv_get_instance_entrypoint_index(pName);
1986 if (idx >= 0)
1987 return instance->dispatch.entrypoints[idx];
1988
1989 idx = anv_get_physical_device_entrypoint_index(pName);
1990 if (idx >= 0)
1991 return instance->physicalDevice.dispatch.entrypoints[idx];
1992
1993 idx = anv_get_device_entrypoint_index(pName);
1994 if (idx >= 0)
1995 return instance->device_dispatch.entrypoints[idx];
1996
1997 return NULL;
1998 }
1999
2000 /* With version 1+ of the loader interface the ICD should expose
2001 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
2002 */
2003 PUBLIC
2004 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
2005 VkInstance instance,
2006 const char* pName);
2007
2008 PUBLIC
2009 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
2010 VkInstance instance,
2011 const char* pName)
2012 {
2013 return anv_GetInstanceProcAddr(instance, pName);
2014 }
2015
2016 PFN_vkVoidFunction anv_GetDeviceProcAddr(
2017 VkDevice _device,
2018 const char* pName)
2019 {
2020 ANV_FROM_HANDLE(anv_device, device, _device);
2021
2022 if (!device || !pName)
2023 return NULL;
2024
2025 int idx = anv_get_device_entrypoint_index(pName);
2026 if (idx < 0)
2027 return NULL;
2028
2029 return device->dispatch.entrypoints[idx];
2030 }
2031
2032 /* With version 4+ of the loader interface the ICD should expose
2033 * vk_icdGetPhysicalDeviceProcAddr()
2034 */
2035 PUBLIC
2036 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetPhysicalDeviceProcAddr(
2037 VkInstance _instance,
2038 const char* pName);
2039
2040 PFN_vkVoidFunction vk_icdGetPhysicalDeviceProcAddr(
2041 VkInstance _instance,
2042 const char* pName)
2043 {
2044 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2045
2046 if (!pName || !instance)
2047 return NULL;
2048
2049 int idx = anv_get_physical_device_entrypoint_index(pName);
2050 if (idx < 0)
2051 return NULL;
2052
2053 return instance->physicalDevice.dispatch.entrypoints[idx];
2054 }
2055
2056
2057 VkResult
2058 anv_CreateDebugReportCallbackEXT(VkInstance _instance,
2059 const VkDebugReportCallbackCreateInfoEXT* pCreateInfo,
2060 const VkAllocationCallbacks* pAllocator,
2061 VkDebugReportCallbackEXT* pCallback)
2062 {
2063 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2064 return vk_create_debug_report_callback(&instance->debug_report_callbacks,
2065 pCreateInfo, pAllocator, &instance->alloc,
2066 pCallback);
2067 }
2068
2069 void
2070 anv_DestroyDebugReportCallbackEXT(VkInstance _instance,
2071 VkDebugReportCallbackEXT _callback,
2072 const VkAllocationCallbacks* pAllocator)
2073 {
2074 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2075 vk_destroy_debug_report_callback(&instance->debug_report_callbacks,
2076 _callback, pAllocator, &instance->alloc);
2077 }
2078
2079 void
2080 anv_DebugReportMessageEXT(VkInstance _instance,
2081 VkDebugReportFlagsEXT flags,
2082 VkDebugReportObjectTypeEXT objectType,
2083 uint64_t object,
2084 size_t location,
2085 int32_t messageCode,
2086 const char* pLayerPrefix,
2087 const char* pMessage)
2088 {
2089 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2090 vk_debug_report(&instance->debug_report_callbacks, flags, objectType,
2091 object, location, messageCode, pLayerPrefix, pMessage);
2092 }
2093
2094 static struct anv_state
2095 anv_state_pool_emit_data(struct anv_state_pool *pool, size_t size, size_t align, const void *p)
2096 {
2097 struct anv_state state;
2098
2099 state = anv_state_pool_alloc(pool, size, align);
2100 memcpy(state.map, p, size);
2101
2102 return state;
2103 }
2104
2105 /* Haswell border color is a bit of a disaster. Float and unorm formats use a
2106 * straightforward 32-bit float color in the first 64 bytes. Instead of using
2107 * a nice float/integer union like Gen8+, Haswell specifies the integer border
2108 * color as a separate entry /after/ the float color. The layout of this entry
2109 * also depends on the format's bpp (with extra hacks for RG32), and overlaps.
2110 *
2111 * Since we don't know the format/bpp, we can't make any of the border colors
2112 * containing '1' work for all formats, as it would be in the wrong place for
2113 * some of them. We opt to make 32-bit integers work as this seems like the
2114 * most common option. Fortunately, transparent black works regardless, as
2115 * all zeroes is the same in every bit-size.
2116 */
2117 struct hsw_border_color {
2118 float float32[4];
2119 uint32_t _pad0[12];
2120 uint32_t uint32[4];
2121 uint32_t _pad1[108];
2122 };
2123
2124 struct gen8_border_color {
2125 union {
2126 float float32[4];
2127 uint32_t uint32[4];
2128 };
2129 /* Pad out to 64 bytes */
2130 uint32_t _pad[12];
2131 };
2132
2133 static void
2134 anv_device_init_border_colors(struct anv_device *device)
2135 {
2136 if (device->info.is_haswell) {
2137 static const struct hsw_border_color border_colors[] = {
2138 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
2139 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
2140 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
2141 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = { .uint32 = { 0, 0, 0, 0 } },
2142 [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = { .uint32 = { 0, 0, 0, 1 } },
2143 [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = { .uint32 = { 1, 1, 1, 1 } },
2144 };
2145
2146 device->border_colors =
2147 anv_state_pool_emit_data(&device->dynamic_state_pool,
2148 sizeof(border_colors), 512, border_colors);
2149 } else {
2150 static const struct gen8_border_color border_colors[] = {
2151 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
2152 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
2153 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
2154 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = { .uint32 = { 0, 0, 0, 0 } },
2155 [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = { .uint32 = { 0, 0, 0, 1 } },
2156 [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = { .uint32 = { 1, 1, 1, 1 } },
2157 };
2158
2159 device->border_colors =
2160 anv_state_pool_emit_data(&device->dynamic_state_pool,
2161 sizeof(border_colors), 64, border_colors);
2162 }
2163 }
2164
2165 static VkResult
2166 anv_device_init_trivial_batch(struct anv_device *device)
2167 {
2168 VkResult result = anv_device_alloc_bo(device, 4096,
2169 ANV_BO_ALLOC_MAPPED,
2170 0 /* explicit_address */,
2171 &device->trivial_batch_bo);
2172 if (result != VK_SUCCESS)
2173 return result;
2174
2175 struct anv_batch batch = {
2176 .start = device->trivial_batch_bo->map,
2177 .next = device->trivial_batch_bo->map,
2178 .end = device->trivial_batch_bo->map + 4096,
2179 };
2180
2181 anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
2182 anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
2183
2184 if (!device->info.has_llc)
2185 gen_clflush_range(batch.start, batch.next - batch.start);
2186
2187 return VK_SUCCESS;
2188 }
2189
2190 VkResult anv_EnumerateDeviceExtensionProperties(
2191 VkPhysicalDevice physicalDevice,
2192 const char* pLayerName,
2193 uint32_t* pPropertyCount,
2194 VkExtensionProperties* pProperties)
2195 {
2196 ANV_FROM_HANDLE(anv_physical_device, device, physicalDevice);
2197 VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
2198
2199 for (int i = 0; i < ANV_DEVICE_EXTENSION_COUNT; i++) {
2200 if (device->supported_extensions.extensions[i]) {
2201 vk_outarray_append(&out, prop) {
2202 *prop = anv_device_extensions[i];
2203 }
2204 }
2205 }
2206
2207 return vk_outarray_status(&out);
2208 }
2209
2210 static void
2211 anv_device_init_dispatch(struct anv_device *device)
2212 {
2213 const struct anv_device_dispatch_table *genX_table;
2214 switch (device->info.gen) {
2215 case 12:
2216 genX_table = &gen12_device_dispatch_table;
2217 break;
2218 case 11:
2219 genX_table = &gen11_device_dispatch_table;
2220 break;
2221 case 10:
2222 genX_table = &gen10_device_dispatch_table;
2223 break;
2224 case 9:
2225 genX_table = &gen9_device_dispatch_table;
2226 break;
2227 case 8:
2228 genX_table = &gen8_device_dispatch_table;
2229 break;
2230 case 7:
2231 if (device->info.is_haswell)
2232 genX_table = &gen75_device_dispatch_table;
2233 else
2234 genX_table = &gen7_device_dispatch_table;
2235 break;
2236 default:
2237 unreachable("unsupported gen\n");
2238 }
2239
2240 for (unsigned i = 0; i < ARRAY_SIZE(device->dispatch.entrypoints); i++) {
2241 /* Vulkan requires that entrypoints for extensions which have not been
2242 * enabled must not be advertised.
2243 */
2244 if (!anv_device_entrypoint_is_enabled(i, device->instance->app_info.api_version,
2245 &device->instance->enabled_extensions,
2246 &device->enabled_extensions)) {
2247 device->dispatch.entrypoints[i] = NULL;
2248 } else if (genX_table->entrypoints[i]) {
2249 device->dispatch.entrypoints[i] = genX_table->entrypoints[i];
2250 } else {
2251 device->dispatch.entrypoints[i] =
2252 anv_device_dispatch_table.entrypoints[i];
2253 }
2254 }
2255 }
2256
2257 static int
2258 vk_priority_to_gen(int priority)
2259 {
2260 switch (priority) {
2261 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT:
2262 return GEN_CONTEXT_LOW_PRIORITY;
2263 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT:
2264 return GEN_CONTEXT_MEDIUM_PRIORITY;
2265 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT:
2266 return GEN_CONTEXT_HIGH_PRIORITY;
2267 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT:
2268 return GEN_CONTEXT_REALTIME_PRIORITY;
2269 default:
2270 unreachable("Invalid priority");
2271 }
2272 }
2273
2274 static VkResult
2275 anv_device_init_hiz_clear_value_bo(struct anv_device *device)
2276 {
2277 VkResult result = anv_device_alloc_bo(device, 4096,
2278 ANV_BO_ALLOC_MAPPED,
2279 0 /* explicit_address */,
2280 &device->hiz_clear_bo);
2281 if (result != VK_SUCCESS)
2282 return result;
2283
2284 union isl_color_value hiz_clear = { .u32 = { 0, } };
2285 hiz_clear.f32[0] = ANV_HZ_FC_VAL;
2286
2287 memcpy(device->hiz_clear_bo->map, hiz_clear.u32, sizeof(hiz_clear.u32));
2288
2289 if (!device->info.has_llc)
2290 gen_clflush_range(device->hiz_clear_bo->map, sizeof(hiz_clear.u32));
2291
2292 return VK_SUCCESS;
2293 }
2294
2295 static bool
2296 get_bo_from_pool(struct gen_batch_decode_bo *ret,
2297 struct anv_block_pool *pool,
2298 uint64_t address)
2299 {
2300 anv_block_pool_foreach_bo(bo, pool) {
2301 uint64_t bo_address = gen_48b_address(bo->offset);
2302 if (address >= bo_address && address < (bo_address + bo->size)) {
2303 *ret = (struct gen_batch_decode_bo) {
2304 .addr = bo_address,
2305 .size = bo->size,
2306 .map = bo->map,
2307 };
2308 return true;
2309 }
2310 }
2311 return false;
2312 }
2313
2314 /* Finding a buffer for batch decoding */
2315 static struct gen_batch_decode_bo
2316 decode_get_bo(void *v_batch, bool ppgtt, uint64_t address)
2317 {
2318 struct anv_device *device = v_batch;
2319 struct gen_batch_decode_bo ret_bo = {};
2320
2321 assert(ppgtt);
2322
2323 if (get_bo_from_pool(&ret_bo, &device->dynamic_state_pool.block_pool, address))
2324 return ret_bo;
2325 if (get_bo_from_pool(&ret_bo, &device->instruction_state_pool.block_pool, address))
2326 return ret_bo;
2327 if (get_bo_from_pool(&ret_bo, &device->binding_table_pool.block_pool, address))
2328 return ret_bo;
2329 if (get_bo_from_pool(&ret_bo, &device->surface_state_pool.block_pool, address))
2330 return ret_bo;
2331
2332 if (!device->cmd_buffer_being_decoded)
2333 return (struct gen_batch_decode_bo) { };
2334
2335 struct anv_batch_bo **bo;
2336
2337 u_vector_foreach(bo, &device->cmd_buffer_being_decoded->seen_bbos) {
2338 /* The decoder zeroes out the top 16 bits, so we need to as well */
2339 uint64_t bo_address = (*bo)->bo->offset & (~0ull >> 16);
2340
2341 if (address >= bo_address && address < bo_address + (*bo)->bo->size) {
2342 return (struct gen_batch_decode_bo) {
2343 .addr = bo_address,
2344 .size = (*bo)->bo->size,
2345 .map = (*bo)->bo->map,
2346 };
2347 }
2348 }
2349
2350 return (struct gen_batch_decode_bo) { };
2351 }
2352
2353 struct gen_aux_map_buffer {
2354 struct gen_buffer base;
2355 struct anv_state state;
2356 };
2357
2358 static struct gen_buffer *
2359 gen_aux_map_buffer_alloc(void *driver_ctx, uint32_t size)
2360 {
2361 struct gen_aux_map_buffer *buf = malloc(sizeof(struct gen_aux_map_buffer));
2362 if (!buf)
2363 return NULL;
2364
2365 struct anv_device *device = (struct anv_device*)driver_ctx;
2366 assert(device->instance->physicalDevice.supports_48bit_addresses &&
2367 device->instance->physicalDevice.use_softpin);
2368
2369 struct anv_state_pool *pool = &device->dynamic_state_pool;
2370 buf->state = anv_state_pool_alloc(pool, size, size);
2371
2372 buf->base.gpu = pool->block_pool.bo->offset + buf->state.offset;
2373 buf->base.gpu_end = buf->base.gpu + buf->state.alloc_size;
2374 buf->base.map = buf->state.map;
2375 buf->base.driver_bo = &buf->state;
2376 return &buf->base;
2377 }
2378
2379 static void
2380 gen_aux_map_buffer_free(void *driver_ctx, struct gen_buffer *buffer)
2381 {
2382 struct gen_aux_map_buffer *buf = (struct gen_aux_map_buffer*)buffer;
2383 struct anv_device *device = (struct anv_device*)driver_ctx;
2384 struct anv_state_pool *pool = &device->dynamic_state_pool;
2385 anv_state_pool_free(pool, buf->state);
2386 free(buf);
2387 }
2388
2389 static struct gen_mapped_pinned_buffer_alloc aux_map_allocator = {
2390 .alloc = gen_aux_map_buffer_alloc,
2391 .free = gen_aux_map_buffer_free,
2392 };
2393
2394 VkResult anv_CreateDevice(
2395 VkPhysicalDevice physicalDevice,
2396 const VkDeviceCreateInfo* pCreateInfo,
2397 const VkAllocationCallbacks* pAllocator,
2398 VkDevice* pDevice)
2399 {
2400 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
2401 VkResult result;
2402 struct anv_device *device;
2403
2404 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO);
2405
2406 struct anv_device_extension_table enabled_extensions = { };
2407 for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
2408 int idx;
2409 for (idx = 0; idx < ANV_DEVICE_EXTENSION_COUNT; idx++) {
2410 if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
2411 anv_device_extensions[idx].extensionName) == 0)
2412 break;
2413 }
2414
2415 if (idx >= ANV_DEVICE_EXTENSION_COUNT)
2416 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
2417
2418 if (!physical_device->supported_extensions.extensions[idx])
2419 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
2420
2421 enabled_extensions.extensions[idx] = true;
2422 }
2423
2424 /* Check enabled features */
2425 if (pCreateInfo->pEnabledFeatures) {
2426 VkPhysicalDeviceFeatures supported_features;
2427 anv_GetPhysicalDeviceFeatures(physicalDevice, &supported_features);
2428 VkBool32 *supported_feature = (VkBool32 *)&supported_features;
2429 VkBool32 *enabled_feature = (VkBool32 *)pCreateInfo->pEnabledFeatures;
2430 unsigned num_features = sizeof(VkPhysicalDeviceFeatures) / sizeof(VkBool32);
2431 for (uint32_t i = 0; i < num_features; i++) {
2432 if (enabled_feature[i] && !supported_feature[i])
2433 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
2434 }
2435 }
2436
2437 /* Check requested queues and fail if we are requested to create any
2438 * queues with flags we don't support.
2439 */
2440 assert(pCreateInfo->queueCreateInfoCount > 0);
2441 for (uint32_t i = 0; i < pCreateInfo->queueCreateInfoCount; i++) {
2442 if (pCreateInfo->pQueueCreateInfos[i].flags != 0)
2443 return vk_error(VK_ERROR_INITIALIZATION_FAILED);
2444 }
2445
2446 /* Check if client specified queue priority. */
2447 const VkDeviceQueueGlobalPriorityCreateInfoEXT *queue_priority =
2448 vk_find_struct_const(pCreateInfo->pQueueCreateInfos[0].pNext,
2449 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT);
2450
2451 VkQueueGlobalPriorityEXT priority =
2452 queue_priority ? queue_priority->globalPriority :
2453 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT;
2454
2455 device = vk_alloc2(&physical_device->instance->alloc, pAllocator,
2456 sizeof(*device), 8,
2457 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
2458 if (!device)
2459 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2460
2461 if (INTEL_DEBUG & DEBUG_BATCH) {
2462 const unsigned decode_flags =
2463 GEN_BATCH_DECODE_FULL |
2464 ((INTEL_DEBUG & DEBUG_COLOR) ? GEN_BATCH_DECODE_IN_COLOR : 0) |
2465 GEN_BATCH_DECODE_OFFSETS |
2466 GEN_BATCH_DECODE_FLOATS;
2467
2468 gen_batch_decode_ctx_init(&device->decoder_ctx,
2469 &physical_device->info,
2470 stderr, decode_flags, NULL,
2471 decode_get_bo, NULL, device);
2472 }
2473
2474 device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
2475 device->instance = physical_device->instance;
2476 device->chipset_id = physical_device->chipset_id;
2477 device->no_hw = physical_device->no_hw;
2478 device->_lost = false;
2479
2480 if (pAllocator)
2481 device->alloc = *pAllocator;
2482 else
2483 device->alloc = physical_device->instance->alloc;
2484
2485 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
2486 device->fd = open(physical_device->path, O_RDWR | O_CLOEXEC);
2487 if (device->fd == -1) {
2488 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2489 goto fail_device;
2490 }
2491
2492 device->context_id = anv_gem_create_context(device);
2493 if (device->context_id == -1) {
2494 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2495 goto fail_fd;
2496 }
2497
2498 result = anv_queue_init(device, &device->queue);
2499 if (result != VK_SUCCESS)
2500 goto fail_context_id;
2501
2502 if (physical_device->use_softpin) {
2503 if (pthread_mutex_init(&device->vma_mutex, NULL) != 0) {
2504 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2505 goto fail_queue;
2506 }
2507
2508 /* keep the page with address zero out of the allocator */
2509 util_vma_heap_init(&device->vma_lo,
2510 LOW_HEAP_MIN_ADDRESS, LOW_HEAP_SIZE);
2511
2512 util_vma_heap_init(&device->vma_cva, CLIENT_VISIBLE_HEAP_MIN_ADDRESS,
2513 CLIENT_VISIBLE_HEAP_SIZE);
2514
2515 /* Leave the last 4GiB out of the high vma range, so that no state
2516 * base address + size can overflow 48 bits. For more information see
2517 * the comment about Wa32bitGeneralStateOffset in anv_allocator.c
2518 */
2519 util_vma_heap_init(&device->vma_hi, HIGH_HEAP_MIN_ADDRESS,
2520 physical_device->gtt_size - (1ull << 32) -
2521 HIGH_HEAP_MIN_ADDRESS);
2522 }
2523
2524 list_inithead(&device->memory_objects);
2525
2526 /* As per spec, the driver implementation may deny requests to acquire
2527 * a priority above the default priority (MEDIUM) if the caller does not
2528 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
2529 * is returned.
2530 */
2531 if (physical_device->has_context_priority) {
2532 int err = anv_gem_set_context_param(device->fd, device->context_id,
2533 I915_CONTEXT_PARAM_PRIORITY,
2534 vk_priority_to_gen(priority));
2535 if (err != 0 && priority > VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT) {
2536 result = vk_error(VK_ERROR_NOT_PERMITTED_EXT);
2537 goto fail_vmas;
2538 }
2539 }
2540
2541 device->info = physical_device->info;
2542 device->isl_dev = physical_device->isl_dev;
2543
2544 /* On Broadwell and later, we can use batch chaining to more efficiently
2545 * implement growing command buffers. Prior to Haswell, the kernel
2546 * command parser gets in the way and we have to fall back to growing
2547 * the batch.
2548 */
2549 device->can_chain_batches = device->info.gen >= 8;
2550
2551 device->robust_buffer_access = pCreateInfo->pEnabledFeatures &&
2552 pCreateInfo->pEnabledFeatures->robustBufferAccess;
2553 device->enabled_extensions = enabled_extensions;
2554
2555 anv_device_init_dispatch(device);
2556
2557 if (pthread_mutex_init(&device->mutex, NULL) != 0) {
2558 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2559 goto fail_queue;
2560 }
2561
2562 pthread_condattr_t condattr;
2563 if (pthread_condattr_init(&condattr) != 0) {
2564 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2565 goto fail_mutex;
2566 }
2567 if (pthread_condattr_setclock(&condattr, CLOCK_MONOTONIC) != 0) {
2568 pthread_condattr_destroy(&condattr);
2569 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2570 goto fail_mutex;
2571 }
2572 if (pthread_cond_init(&device->queue_submit, &condattr) != 0) {
2573 pthread_condattr_destroy(&condattr);
2574 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2575 goto fail_mutex;
2576 }
2577 pthread_condattr_destroy(&condattr);
2578
2579 result = anv_bo_cache_init(&device->bo_cache);
2580 if (result != VK_SUCCESS)
2581 goto fail_queue_cond;
2582
2583 anv_bo_pool_init(&device->batch_bo_pool, device);
2584
2585 result = anv_state_pool_init(&device->dynamic_state_pool, device,
2586 DYNAMIC_STATE_POOL_MIN_ADDRESS, 16384);
2587 if (result != VK_SUCCESS)
2588 goto fail_batch_bo_pool;
2589
2590 result = anv_state_pool_init(&device->instruction_state_pool, device,
2591 INSTRUCTION_STATE_POOL_MIN_ADDRESS, 16384);
2592 if (result != VK_SUCCESS)
2593 goto fail_dynamic_state_pool;
2594
2595 result = anv_state_pool_init(&device->surface_state_pool, device,
2596 SURFACE_STATE_POOL_MIN_ADDRESS, 4096);
2597 if (result != VK_SUCCESS)
2598 goto fail_instruction_state_pool;
2599
2600 if (physical_device->use_softpin) {
2601 result = anv_state_pool_init(&device->binding_table_pool, device,
2602 BINDING_TABLE_POOL_MIN_ADDRESS, 4096);
2603 if (result != VK_SUCCESS)
2604 goto fail_surface_state_pool;
2605 }
2606
2607 if (device->info.gen >= 12) {
2608 device->aux_map_ctx = gen_aux_map_init(device, &aux_map_allocator,
2609 &physical_device->info);
2610 if (!device->aux_map_ctx)
2611 goto fail_binding_table_pool;
2612 }
2613
2614 result = anv_device_alloc_bo(device, 4096, 0 /* flags */,
2615 0 /* explicit_address */,
2616 &device->workaround_bo);
2617 if (result != VK_SUCCESS)
2618 goto fail_surface_aux_map_pool;
2619
2620 result = anv_device_init_trivial_batch(device);
2621 if (result != VK_SUCCESS)
2622 goto fail_workaround_bo;
2623
2624 if (device->info.gen >= 10) {
2625 result = anv_device_init_hiz_clear_value_bo(device);
2626 if (result != VK_SUCCESS)
2627 goto fail_trivial_batch_bo;
2628 }
2629
2630 anv_scratch_pool_init(device, &device->scratch_pool);
2631
2632 switch (device->info.gen) {
2633 case 7:
2634 if (!device->info.is_haswell)
2635 result = gen7_init_device_state(device);
2636 else
2637 result = gen75_init_device_state(device);
2638 break;
2639 case 8:
2640 result = gen8_init_device_state(device);
2641 break;
2642 case 9:
2643 result = gen9_init_device_state(device);
2644 break;
2645 case 10:
2646 result = gen10_init_device_state(device);
2647 break;
2648 case 11:
2649 result = gen11_init_device_state(device);
2650 break;
2651 case 12:
2652 result = gen12_init_device_state(device);
2653 break;
2654 default:
2655 /* Shouldn't get here as we don't create physical devices for any other
2656 * gens. */
2657 unreachable("unhandled gen");
2658 }
2659 if (result != VK_SUCCESS)
2660 goto fail_workaround_bo;
2661
2662 anv_pipeline_cache_init(&device->default_pipeline_cache, device, true);
2663
2664 anv_device_init_blorp(device);
2665
2666 anv_device_init_border_colors(device);
2667
2668 anv_device_perf_init(device);
2669
2670 *pDevice = anv_device_to_handle(device);
2671
2672 return VK_SUCCESS;
2673
2674 fail_workaround_bo:
2675 anv_scratch_pool_finish(device, &device->scratch_pool);
2676 if (device->info.gen >= 10)
2677 anv_device_release_bo(device, device->hiz_clear_bo);
2678 anv_device_release_bo(device, device->workaround_bo);
2679 fail_trivial_batch_bo:
2680 anv_device_release_bo(device, device->trivial_batch_bo);
2681 fail_surface_aux_map_pool:
2682 if (device->info.gen >= 12) {
2683 gen_aux_map_finish(device->aux_map_ctx);
2684 device->aux_map_ctx = NULL;
2685 }
2686 fail_binding_table_pool:
2687 if (physical_device->use_softpin)
2688 anv_state_pool_finish(&device->binding_table_pool);
2689 fail_surface_state_pool:
2690 anv_state_pool_finish(&device->surface_state_pool);
2691 fail_instruction_state_pool:
2692 anv_state_pool_finish(&device->instruction_state_pool);
2693 fail_dynamic_state_pool:
2694 anv_state_pool_finish(&device->dynamic_state_pool);
2695 fail_batch_bo_pool:
2696 anv_bo_pool_finish(&device->batch_bo_pool);
2697 anv_bo_cache_finish(&device->bo_cache);
2698 fail_queue_cond:
2699 pthread_cond_destroy(&device->queue_submit);
2700 fail_mutex:
2701 pthread_mutex_destroy(&device->mutex);
2702 fail_vmas:
2703 if (physical_device->use_softpin) {
2704 util_vma_heap_finish(&device->vma_hi);
2705 util_vma_heap_finish(&device->vma_cva);
2706 util_vma_heap_finish(&device->vma_lo);
2707 }
2708 fail_queue:
2709 anv_queue_finish(&device->queue);
2710 fail_context_id:
2711 anv_gem_destroy_context(device, device->context_id);
2712 fail_fd:
2713 close(device->fd);
2714 fail_device:
2715 vk_free(&device->alloc, device);
2716
2717 return result;
2718 }
2719
2720 void anv_DestroyDevice(
2721 VkDevice _device,
2722 const VkAllocationCallbacks* pAllocator)
2723 {
2724 ANV_FROM_HANDLE(anv_device, device, _device);
2725 struct anv_physical_device *physical_device;
2726
2727 if (!device)
2728 return;
2729
2730 physical_device = &device->instance->physicalDevice;
2731
2732 anv_device_finish_blorp(device);
2733
2734 anv_pipeline_cache_finish(&device->default_pipeline_cache);
2735
2736 anv_queue_finish(&device->queue);
2737
2738 #ifdef HAVE_VALGRIND
2739 /* We only need to free these to prevent valgrind errors. The backing
2740 * BO will go away in a couple of lines so we don't actually leak.
2741 */
2742 anv_state_pool_free(&device->dynamic_state_pool, device->border_colors);
2743 anv_state_pool_free(&device->dynamic_state_pool, device->slice_hash);
2744 #endif
2745
2746 anv_scratch_pool_finish(device, &device->scratch_pool);
2747
2748 anv_device_release_bo(device, device->workaround_bo);
2749 anv_device_release_bo(device, device->trivial_batch_bo);
2750 if (device->info.gen >= 10)
2751 anv_device_release_bo(device, device->hiz_clear_bo);
2752
2753 if (device->info.gen >= 12) {
2754 gen_aux_map_finish(device->aux_map_ctx);
2755 device->aux_map_ctx = NULL;
2756 }
2757
2758 if (physical_device->use_softpin)
2759 anv_state_pool_finish(&device->binding_table_pool);
2760 anv_state_pool_finish(&device->surface_state_pool);
2761 anv_state_pool_finish(&device->instruction_state_pool);
2762 anv_state_pool_finish(&device->dynamic_state_pool);
2763
2764 anv_bo_pool_finish(&device->batch_bo_pool);
2765
2766 anv_bo_cache_finish(&device->bo_cache);
2767
2768 if (physical_device->use_softpin) {
2769 util_vma_heap_finish(&device->vma_hi);
2770 util_vma_heap_finish(&device->vma_cva);
2771 util_vma_heap_finish(&device->vma_lo);
2772 }
2773
2774 pthread_cond_destroy(&device->queue_submit);
2775 pthread_mutex_destroy(&device->mutex);
2776
2777 anv_gem_destroy_context(device, device->context_id);
2778
2779 if (INTEL_DEBUG & DEBUG_BATCH)
2780 gen_batch_decode_ctx_finish(&device->decoder_ctx);
2781
2782 close(device->fd);
2783
2784 vk_free(&device->alloc, device);
2785 }
2786
2787 VkResult anv_EnumerateInstanceLayerProperties(
2788 uint32_t* pPropertyCount,
2789 VkLayerProperties* pProperties)
2790 {
2791 if (pProperties == NULL) {
2792 *pPropertyCount = 0;
2793 return VK_SUCCESS;
2794 }
2795
2796 /* None supported at this time */
2797 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
2798 }
2799
2800 VkResult anv_EnumerateDeviceLayerProperties(
2801 VkPhysicalDevice physicalDevice,
2802 uint32_t* pPropertyCount,
2803 VkLayerProperties* pProperties)
2804 {
2805 if (pProperties == NULL) {
2806 *pPropertyCount = 0;
2807 return VK_SUCCESS;
2808 }
2809
2810 /* None supported at this time */
2811 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
2812 }
2813
2814 void anv_GetDeviceQueue(
2815 VkDevice _device,
2816 uint32_t queueNodeIndex,
2817 uint32_t queueIndex,
2818 VkQueue* pQueue)
2819 {
2820 const VkDeviceQueueInfo2 info = {
2821 .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_INFO_2,
2822 .pNext = NULL,
2823 .flags = 0,
2824 .queueFamilyIndex = queueNodeIndex,
2825 .queueIndex = queueIndex,
2826 };
2827
2828 anv_GetDeviceQueue2(_device, &info, pQueue);
2829 }
2830
2831 void anv_GetDeviceQueue2(
2832 VkDevice _device,
2833 const VkDeviceQueueInfo2* pQueueInfo,
2834 VkQueue* pQueue)
2835 {
2836 ANV_FROM_HANDLE(anv_device, device, _device);
2837
2838 assert(pQueueInfo->queueIndex == 0);
2839
2840 if (pQueueInfo->flags == device->queue.flags)
2841 *pQueue = anv_queue_to_handle(&device->queue);
2842 else
2843 *pQueue = NULL;
2844 }
2845
2846 VkResult
2847 _anv_device_set_lost(struct anv_device *device,
2848 const char *file, int line,
2849 const char *msg, ...)
2850 {
2851 VkResult err;
2852 va_list ap;
2853
2854 p_atomic_inc(&device->_lost);
2855
2856 va_start(ap, msg);
2857 err = __vk_errorv(device->instance, device,
2858 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT,
2859 VK_ERROR_DEVICE_LOST, file, line, msg, ap);
2860 va_end(ap);
2861
2862 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
2863 abort();
2864
2865 return err;
2866 }
2867
2868 VkResult
2869 _anv_queue_set_lost(struct anv_queue *queue,
2870 const char *file, int line,
2871 const char *msg, ...)
2872 {
2873 VkResult err;
2874 va_list ap;
2875
2876 p_atomic_inc(&queue->device->_lost);
2877
2878 va_start(ap, msg);
2879 err = __vk_errorv(queue->device->instance, queue->device,
2880 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT,
2881 VK_ERROR_DEVICE_LOST, file, line, msg, ap);
2882 va_end(ap);
2883
2884 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
2885 abort();
2886
2887 return err;
2888 }
2889
2890 VkResult
2891 anv_device_query_status(struct anv_device *device)
2892 {
2893 /* This isn't likely as most of the callers of this function already check
2894 * for it. However, it doesn't hurt to check and it potentially lets us
2895 * avoid an ioctl.
2896 */
2897 if (anv_device_is_lost(device))
2898 return VK_ERROR_DEVICE_LOST;
2899
2900 uint32_t active, pending;
2901 int ret = anv_gem_gpu_get_reset_stats(device, &active, &pending);
2902 if (ret == -1) {
2903 /* We don't know the real error. */
2904 return anv_device_set_lost(device, "get_reset_stats failed: %m");
2905 }
2906
2907 if (active) {
2908 return anv_device_set_lost(device, "GPU hung on one of our command buffers");
2909 } else if (pending) {
2910 return anv_device_set_lost(device, "GPU hung with commands in-flight");
2911 }
2912
2913 return VK_SUCCESS;
2914 }
2915
2916 VkResult
2917 anv_device_bo_busy(struct anv_device *device, struct anv_bo *bo)
2918 {
2919 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
2920 * Other usages of the BO (such as on different hardware) will not be
2921 * flagged as "busy" by this ioctl. Use with care.
2922 */
2923 int ret = anv_gem_busy(device, bo->gem_handle);
2924 if (ret == 1) {
2925 return VK_NOT_READY;
2926 } else if (ret == -1) {
2927 /* We don't know the real error. */
2928 return anv_device_set_lost(device, "gem wait failed: %m");
2929 }
2930
2931 /* Query for device status after the busy call. If the BO we're checking
2932 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
2933 * client because it clearly doesn't have valid data. Yes, this most
2934 * likely means an ioctl, but we just did an ioctl to query the busy status
2935 * so it's no great loss.
2936 */
2937 return anv_device_query_status(device);
2938 }
2939
2940 VkResult
2941 anv_device_wait(struct anv_device *device, struct anv_bo *bo,
2942 int64_t timeout)
2943 {
2944 int ret = anv_gem_wait(device, bo->gem_handle, &timeout);
2945 if (ret == -1 && errno == ETIME) {
2946 return VK_TIMEOUT;
2947 } else if (ret == -1) {
2948 /* We don't know the real error. */
2949 return anv_device_set_lost(device, "gem wait failed: %m");
2950 }
2951
2952 /* Query for device status after the wait. If the BO we're waiting on got
2953 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
2954 * because it clearly doesn't have valid data. Yes, this most likely means
2955 * an ioctl, but we just did an ioctl to wait so it's no great loss.
2956 */
2957 return anv_device_query_status(device);
2958 }
2959
2960 VkResult anv_DeviceWaitIdle(
2961 VkDevice _device)
2962 {
2963 ANV_FROM_HANDLE(anv_device, device, _device);
2964
2965 if (anv_device_is_lost(device))
2966 return VK_ERROR_DEVICE_LOST;
2967
2968 return anv_queue_submit_simple_batch(&device->queue, NULL);
2969 }
2970
2971 bool
2972 anv_vma_alloc(struct anv_device *device, struct anv_bo *bo,
2973 uint64_t client_address)
2974 {
2975 if (!(bo->flags & EXEC_OBJECT_PINNED)) {
2976 assert(!(bo->has_client_visible_address));
2977 return true;
2978 }
2979
2980 pthread_mutex_lock(&device->vma_mutex);
2981
2982 bo->offset = 0;
2983
2984 if (bo->has_client_visible_address) {
2985 assert(bo->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS);
2986 if (client_address) {
2987 if (util_vma_heap_alloc_addr(&device->vma_cva,
2988 client_address, bo->size)) {
2989 bo->offset = gen_canonical_address(client_address);
2990 }
2991 } else {
2992 uint64_t addr = util_vma_heap_alloc(&device->vma_cva, bo->size, 4096);
2993 if (addr) {
2994 bo->offset = gen_canonical_address(addr);
2995 assert(addr == gen_48b_address(bo->offset));
2996 }
2997 }
2998 /* We don't want to fall back to other heaps */
2999 goto done;
3000 }
3001
3002 assert(client_address == 0);
3003
3004 if (bo->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) {
3005 uint64_t addr = util_vma_heap_alloc(&device->vma_hi, bo->size, 4096);
3006 if (addr) {
3007 bo->offset = gen_canonical_address(addr);
3008 assert(addr == gen_48b_address(bo->offset));
3009 }
3010 }
3011
3012 if (bo->offset == 0) {
3013 uint64_t addr = util_vma_heap_alloc(&device->vma_lo, bo->size, 4096);
3014 if (addr) {
3015 bo->offset = gen_canonical_address(addr);
3016 assert(addr == gen_48b_address(bo->offset));
3017 }
3018 }
3019
3020 done:
3021 pthread_mutex_unlock(&device->vma_mutex);
3022
3023 return bo->offset != 0;
3024 }
3025
3026 void
3027 anv_vma_free(struct anv_device *device, struct anv_bo *bo)
3028 {
3029 if (!(bo->flags & EXEC_OBJECT_PINNED))
3030 return;
3031
3032 const uint64_t addr_48b = gen_48b_address(bo->offset);
3033
3034 pthread_mutex_lock(&device->vma_mutex);
3035
3036 if (addr_48b >= LOW_HEAP_MIN_ADDRESS &&
3037 addr_48b <= LOW_HEAP_MAX_ADDRESS) {
3038 util_vma_heap_free(&device->vma_lo, addr_48b, bo->size);
3039 } else if (addr_48b >= CLIENT_VISIBLE_HEAP_MIN_ADDRESS &&
3040 addr_48b <= CLIENT_VISIBLE_HEAP_MAX_ADDRESS) {
3041 util_vma_heap_free(&device->vma_cva, addr_48b, bo->size);
3042 } else {
3043 assert(addr_48b >= HIGH_HEAP_MIN_ADDRESS);
3044 util_vma_heap_free(&device->vma_hi, addr_48b, bo->size);
3045 }
3046
3047 pthread_mutex_unlock(&device->vma_mutex);
3048
3049 bo->offset = 0;
3050 }
3051
3052 VkResult anv_AllocateMemory(
3053 VkDevice _device,
3054 const VkMemoryAllocateInfo* pAllocateInfo,
3055 const VkAllocationCallbacks* pAllocator,
3056 VkDeviceMemory* pMem)
3057 {
3058 ANV_FROM_HANDLE(anv_device, device, _device);
3059 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3060 struct anv_device_memory *mem;
3061 VkResult result = VK_SUCCESS;
3062
3063 assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
3064
3065 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
3066 assert(pAllocateInfo->allocationSize > 0);
3067
3068 VkDeviceSize aligned_alloc_size =
3069 align_u64(pAllocateInfo->allocationSize, 4096);
3070
3071 if (aligned_alloc_size > MAX_MEMORY_ALLOCATION_SIZE)
3072 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
3073
3074 assert(pAllocateInfo->memoryTypeIndex < pdevice->memory.type_count);
3075 struct anv_memory_type *mem_type =
3076 &pdevice->memory.types[pAllocateInfo->memoryTypeIndex];
3077 assert(mem_type->heapIndex < pdevice->memory.heap_count);
3078 struct anv_memory_heap *mem_heap =
3079 &pdevice->memory.heaps[mem_type->heapIndex];
3080
3081 uint64_t mem_heap_used = p_atomic_read(&mem_heap->used);
3082 if (mem_heap_used + aligned_alloc_size > mem_heap->size)
3083 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
3084
3085 mem = vk_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
3086 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
3087 if (mem == NULL)
3088 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
3089
3090 assert(pAllocateInfo->memoryTypeIndex < pdevice->memory.type_count);
3091 mem->type = mem_type;
3092 mem->map = NULL;
3093 mem->map_size = 0;
3094 mem->ahw = NULL;
3095 mem->host_ptr = NULL;
3096
3097 enum anv_bo_alloc_flags alloc_flags = 0;
3098
3099 const VkExportMemoryAllocateInfo *export_info = NULL;
3100 const VkImportAndroidHardwareBufferInfoANDROID *ahw_import_info = NULL;
3101 const VkImportMemoryFdInfoKHR *fd_info = NULL;
3102 const VkImportMemoryHostPointerInfoEXT *host_ptr_info = NULL;
3103 const VkMemoryDedicatedAllocateInfo *dedicated_info = NULL;
3104 VkMemoryAllocateFlags vk_flags = 0;
3105 uint64_t client_address = 0;
3106
3107 vk_foreach_struct_const(ext, pAllocateInfo->pNext) {
3108 switch (ext->sType) {
3109 case VK_STRUCTURE_TYPE_EXPORT_MEMORY_ALLOCATE_INFO:
3110 export_info = (void *)ext;
3111 break;
3112
3113 case VK_STRUCTURE_TYPE_IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID:
3114 ahw_import_info = (void *)ext;
3115 break;
3116
3117 case VK_STRUCTURE_TYPE_IMPORT_MEMORY_FD_INFO_KHR:
3118 fd_info = (void *)ext;
3119 break;
3120
3121 case VK_STRUCTURE_TYPE_IMPORT_MEMORY_HOST_POINTER_INFO_EXT:
3122 host_ptr_info = (void *)ext;
3123 break;
3124
3125 case VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_FLAGS_INFO: {
3126 const VkMemoryAllocateFlagsInfo *flags_info = (void *)ext;
3127 vk_flags = flags_info->flags;
3128 break;
3129 }
3130
3131 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_ALLOCATE_INFO:
3132 dedicated_info = (void *)ext;
3133 break;
3134
3135 case VK_STRUCTURE_TYPE_MEMORY_OPAQUE_CAPTURE_ADDRESS_ALLOCATE_INFO_KHR: {
3136 const VkMemoryOpaqueCaptureAddressAllocateInfoKHR *addr_info =
3137 (const VkMemoryOpaqueCaptureAddressAllocateInfoKHR *)ext;
3138 client_address = addr_info->opaqueCaptureAddress;
3139 break;
3140 }
3141
3142 default:
3143 anv_debug_ignored_stype(ext->sType);
3144 break;
3145 }
3146 }
3147
3148 if (vk_flags & VK_MEMORY_ALLOCATE_DEVICE_ADDRESS_BIT_KHR)
3149 alloc_flags |= ANV_BO_ALLOC_CLIENT_VISIBLE_ADDRESS;
3150
3151 /* Check if we need to support Android HW buffer export. If so,
3152 * create AHardwareBuffer and import memory from it.
3153 */
3154 bool android_export = false;
3155 if (export_info && export_info->handleTypes &
3156 VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID)
3157 android_export = true;
3158
3159 if (ahw_import_info) {
3160 result = anv_import_ahw_memory(_device, mem, ahw_import_info);
3161 if (result != VK_SUCCESS)
3162 goto fail;
3163
3164 goto success;
3165 } else if (android_export) {
3166 result = anv_create_ahw_memory(_device, mem, pAllocateInfo);
3167 if (result != VK_SUCCESS)
3168 goto fail;
3169
3170 const struct VkImportAndroidHardwareBufferInfoANDROID import_info = {
3171 .buffer = mem->ahw,
3172 };
3173 result = anv_import_ahw_memory(_device, mem, &import_info);
3174 if (result != VK_SUCCESS)
3175 goto fail;
3176
3177 goto success;
3178 }
3179
3180 /* The Vulkan spec permits handleType to be 0, in which case the struct is
3181 * ignored.
3182 */
3183 if (fd_info && fd_info->handleType) {
3184 /* At the moment, we support only the below handle types. */
3185 assert(fd_info->handleType ==
3186 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
3187 fd_info->handleType ==
3188 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
3189
3190 result = anv_device_import_bo(device, fd_info->fd, alloc_flags,
3191 client_address, &mem->bo);
3192 if (result != VK_SUCCESS)
3193 goto fail;
3194
3195 VkDeviceSize aligned_alloc_size =
3196 align_u64(pAllocateInfo->allocationSize, 4096);
3197
3198 /* For security purposes, we reject importing the bo if it's smaller
3199 * than the requested allocation size. This prevents a malicious client
3200 * from passing a buffer to a trusted client, lying about the size, and
3201 * telling the trusted client to try and texture from an image that goes
3202 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
3203 * in the trusted client. The trusted client can protect itself against
3204 * this sort of attack but only if it can trust the buffer size.
3205 */
3206 if (mem->bo->size < aligned_alloc_size) {
3207 result = vk_errorf(device->instance, device,
3208 VK_ERROR_INVALID_EXTERNAL_HANDLE,
3209 "aligned allocationSize too large for "
3210 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT: "
3211 "%"PRIu64"B > %"PRIu64"B",
3212 aligned_alloc_size, mem->bo->size);
3213 anv_device_release_bo(device, mem->bo);
3214 goto fail;
3215 }
3216
3217 /* From the Vulkan spec:
3218 *
3219 * "Importing memory from a file descriptor transfers ownership of
3220 * the file descriptor from the application to the Vulkan
3221 * implementation. The application must not perform any operations on
3222 * the file descriptor after a successful import."
3223 *
3224 * If the import fails, we leave the file descriptor open.
3225 */
3226 close(fd_info->fd);
3227 goto success;
3228 }
3229
3230 if (host_ptr_info && host_ptr_info->handleType) {
3231 if (host_ptr_info->handleType ==
3232 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_MAPPED_FOREIGN_MEMORY_BIT_EXT) {
3233 result = vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE);
3234 goto fail;
3235 }
3236
3237 assert(host_ptr_info->handleType ==
3238 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT);
3239
3240 result = anv_device_import_bo_from_host_ptr(device,
3241 host_ptr_info->pHostPointer,
3242 pAllocateInfo->allocationSize,
3243 alloc_flags,
3244 client_address,
3245 &mem->bo);
3246 if (result != VK_SUCCESS)
3247 goto fail;
3248
3249 mem->host_ptr = host_ptr_info->pHostPointer;
3250 goto success;
3251 }
3252
3253 /* Regular allocate (not importing memory). */
3254
3255 if (export_info && export_info->handleTypes)
3256 alloc_flags |= ANV_BO_ALLOC_EXTERNAL;
3257
3258 result = anv_device_alloc_bo(device, pAllocateInfo->allocationSize,
3259 alloc_flags, client_address, &mem->bo);
3260 if (result != VK_SUCCESS)
3261 goto fail;
3262
3263 if (dedicated_info && dedicated_info->image != VK_NULL_HANDLE) {
3264 ANV_FROM_HANDLE(anv_image, image, dedicated_info->image);
3265
3266 /* Some legacy (non-modifiers) consumers need the tiling to be set on
3267 * the BO. In this case, we have a dedicated allocation.
3268 */
3269 if (image->needs_set_tiling) {
3270 const uint32_t i915_tiling =
3271 isl_tiling_to_i915_tiling(image->planes[0].surface.isl.tiling);
3272 int ret = anv_gem_set_tiling(device, mem->bo->gem_handle,
3273 image->planes[0].surface.isl.row_pitch_B,
3274 i915_tiling);
3275 if (ret) {
3276 anv_device_release_bo(device, mem->bo);
3277 result = vk_errorf(device->instance, NULL,
3278 VK_ERROR_OUT_OF_DEVICE_MEMORY,
3279 "failed to set BO tiling: %m");
3280 goto fail;
3281 }
3282 }
3283 }
3284
3285 success:
3286 mem_heap_used = p_atomic_add_return(&mem_heap->used, mem->bo->size);
3287 if (mem_heap_used > mem_heap->size) {
3288 p_atomic_add(&mem_heap->used, -mem->bo->size);
3289 anv_device_release_bo(device, mem->bo);
3290 result = vk_errorf(device->instance, NULL,
3291 VK_ERROR_OUT_OF_DEVICE_MEMORY,
3292 "Out of heap memory");
3293 goto fail;
3294 }
3295
3296 pthread_mutex_lock(&device->mutex);
3297 list_addtail(&mem->link, &device->memory_objects);
3298 pthread_mutex_unlock(&device->mutex);
3299
3300 *pMem = anv_device_memory_to_handle(mem);
3301
3302 return VK_SUCCESS;
3303
3304 fail:
3305 vk_free2(&device->alloc, pAllocator, mem);
3306
3307 return result;
3308 }
3309
3310 VkResult anv_GetMemoryFdKHR(
3311 VkDevice device_h,
3312 const VkMemoryGetFdInfoKHR* pGetFdInfo,
3313 int* pFd)
3314 {
3315 ANV_FROM_HANDLE(anv_device, dev, device_h);
3316 ANV_FROM_HANDLE(anv_device_memory, mem, pGetFdInfo->memory);
3317
3318 assert(pGetFdInfo->sType == VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR);
3319
3320 assert(pGetFdInfo->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
3321 pGetFdInfo->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
3322
3323 return anv_device_export_bo(dev, mem->bo, pFd);
3324 }
3325
3326 VkResult anv_GetMemoryFdPropertiesKHR(
3327 VkDevice _device,
3328 VkExternalMemoryHandleTypeFlagBits handleType,
3329 int fd,
3330 VkMemoryFdPropertiesKHR* pMemoryFdProperties)
3331 {
3332 ANV_FROM_HANDLE(anv_device, device, _device);
3333 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3334
3335 switch (handleType) {
3336 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT:
3337 /* dma-buf can be imported as any memory type */
3338 pMemoryFdProperties->memoryTypeBits =
3339 (1 << pdevice->memory.type_count) - 1;
3340 return VK_SUCCESS;
3341
3342 default:
3343 /* The valid usage section for this function says:
3344 *
3345 * "handleType must not be one of the handle types defined as
3346 * opaque."
3347 *
3348 * So opaque handle types fall into the default "unsupported" case.
3349 */
3350 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE);
3351 }
3352 }
3353
3354 VkResult anv_GetMemoryHostPointerPropertiesEXT(
3355 VkDevice _device,
3356 VkExternalMemoryHandleTypeFlagBits handleType,
3357 const void* pHostPointer,
3358 VkMemoryHostPointerPropertiesEXT* pMemoryHostPointerProperties)
3359 {
3360 ANV_FROM_HANDLE(anv_device, device, _device);
3361
3362 assert(pMemoryHostPointerProperties->sType ==
3363 VK_STRUCTURE_TYPE_MEMORY_HOST_POINTER_PROPERTIES_EXT);
3364
3365 switch (handleType) {
3366 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT: {
3367 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3368
3369 /* Host memory can be imported as any memory type. */
3370 pMemoryHostPointerProperties->memoryTypeBits =
3371 (1ull << pdevice->memory.type_count) - 1;
3372
3373 return VK_SUCCESS;
3374 }
3375 default:
3376 return VK_ERROR_INVALID_EXTERNAL_HANDLE;
3377 }
3378 }
3379
3380 void anv_FreeMemory(
3381 VkDevice _device,
3382 VkDeviceMemory _mem,
3383 const VkAllocationCallbacks* pAllocator)
3384 {
3385 ANV_FROM_HANDLE(anv_device, device, _device);
3386 ANV_FROM_HANDLE(anv_device_memory, mem, _mem);
3387 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3388
3389 if (mem == NULL)
3390 return;
3391
3392 pthread_mutex_lock(&device->mutex);
3393 list_del(&mem->link);
3394 pthread_mutex_unlock(&device->mutex);
3395
3396 if (mem->map)
3397 anv_UnmapMemory(_device, _mem);
3398
3399 p_atomic_add(&pdevice->memory.heaps[mem->type->heapIndex].used,
3400 -mem->bo->size);
3401
3402 anv_device_release_bo(device, mem->bo);
3403
3404 #if defined(ANDROID) && ANDROID_API_LEVEL >= 26
3405 if (mem->ahw)
3406 AHardwareBuffer_release(mem->ahw);
3407 #endif
3408
3409 vk_free2(&device->alloc, pAllocator, mem);
3410 }
3411
3412 VkResult anv_MapMemory(
3413 VkDevice _device,
3414 VkDeviceMemory _memory,
3415 VkDeviceSize offset,
3416 VkDeviceSize size,
3417 VkMemoryMapFlags flags,
3418 void** ppData)
3419 {
3420 ANV_FROM_HANDLE(anv_device, device, _device);
3421 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
3422
3423 if (mem == NULL) {
3424 *ppData = NULL;
3425 return VK_SUCCESS;
3426 }
3427
3428 if (mem->host_ptr) {
3429 *ppData = mem->host_ptr + offset;
3430 return VK_SUCCESS;
3431 }
3432
3433 if (size == VK_WHOLE_SIZE)
3434 size = mem->bo->size - offset;
3435
3436 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
3437 *
3438 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
3439 * assert(size != 0);
3440 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
3441 * equal to the size of the memory minus offset
3442 */
3443 assert(size > 0);
3444 assert(offset + size <= mem->bo->size);
3445
3446 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
3447 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
3448 * at a time is valid. We could just mmap up front and return an offset
3449 * pointer here, but that may exhaust virtual memory on 32 bit
3450 * userspace. */
3451
3452 uint32_t gem_flags = 0;
3453
3454 if (!device->info.has_llc &&
3455 (mem->type->propertyFlags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT))
3456 gem_flags |= I915_MMAP_WC;
3457
3458 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
3459 uint64_t map_offset = offset & ~4095ull;
3460 assert(offset >= map_offset);
3461 uint64_t map_size = (offset + size) - map_offset;
3462
3463 /* Let's map whole pages */
3464 map_size = align_u64(map_size, 4096);
3465
3466 void *map = anv_gem_mmap(device, mem->bo->gem_handle,
3467 map_offset, map_size, gem_flags);
3468 if (map == MAP_FAILED)
3469 return vk_error(VK_ERROR_MEMORY_MAP_FAILED);
3470
3471 mem->map = map;
3472 mem->map_size = map_size;
3473
3474 *ppData = mem->map + (offset - map_offset);
3475
3476 return VK_SUCCESS;
3477 }
3478
3479 void anv_UnmapMemory(
3480 VkDevice _device,
3481 VkDeviceMemory _memory)
3482 {
3483 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
3484
3485 if (mem == NULL || mem->host_ptr)
3486 return;
3487
3488 anv_gem_munmap(mem->map, mem->map_size);
3489
3490 mem->map = NULL;
3491 mem->map_size = 0;
3492 }
3493
3494 static void
3495 clflush_mapped_ranges(struct anv_device *device,
3496 uint32_t count,
3497 const VkMappedMemoryRange *ranges)
3498 {
3499 for (uint32_t i = 0; i < count; i++) {
3500 ANV_FROM_HANDLE(anv_device_memory, mem, ranges[i].memory);
3501 if (ranges[i].offset >= mem->map_size)
3502 continue;
3503
3504 gen_clflush_range(mem->map + ranges[i].offset,
3505 MIN2(ranges[i].size, mem->map_size - ranges[i].offset));
3506 }
3507 }
3508
3509 VkResult anv_FlushMappedMemoryRanges(
3510 VkDevice _device,
3511 uint32_t memoryRangeCount,
3512 const VkMappedMemoryRange* pMemoryRanges)
3513 {
3514 ANV_FROM_HANDLE(anv_device, device, _device);
3515
3516 if (device->info.has_llc)
3517 return VK_SUCCESS;
3518
3519 /* Make sure the writes we're flushing have landed. */
3520 __builtin_ia32_mfence();
3521
3522 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
3523
3524 return VK_SUCCESS;
3525 }
3526
3527 VkResult anv_InvalidateMappedMemoryRanges(
3528 VkDevice _device,
3529 uint32_t memoryRangeCount,
3530 const VkMappedMemoryRange* pMemoryRanges)
3531 {
3532 ANV_FROM_HANDLE(anv_device, device, _device);
3533
3534 if (device->info.has_llc)
3535 return VK_SUCCESS;
3536
3537 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
3538
3539 /* Make sure no reads get moved up above the invalidate. */
3540 __builtin_ia32_mfence();
3541
3542 return VK_SUCCESS;
3543 }
3544
3545 void anv_GetBufferMemoryRequirements(
3546 VkDevice _device,
3547 VkBuffer _buffer,
3548 VkMemoryRequirements* pMemoryRequirements)
3549 {
3550 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
3551 ANV_FROM_HANDLE(anv_device, device, _device);
3552 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3553
3554 /* The Vulkan spec (git aaed022) says:
3555 *
3556 * memoryTypeBits is a bitfield and contains one bit set for every
3557 * supported memory type for the resource. The bit `1<<i` is set if and
3558 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3559 * structure for the physical device is supported.
3560 */
3561 uint32_t memory_types = (1ull << pdevice->memory.type_count) - 1;
3562
3563 /* Base alignment requirement of a cache line */
3564 uint32_t alignment = 16;
3565
3566 /* We need an alignment of 32 for pushing UBOs */
3567 if (buffer->usage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT)
3568 alignment = MAX2(alignment, 32);
3569
3570 pMemoryRequirements->size = buffer->size;
3571 pMemoryRequirements->alignment = alignment;
3572
3573 /* Storage and Uniform buffers should have their size aligned to
3574 * 32-bits to avoid boundary checks when last DWord is not complete.
3575 * This would ensure that not internal padding would be needed for
3576 * 16-bit types.
3577 */
3578 if (device->robust_buffer_access &&
3579 (buffer->usage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT ||
3580 buffer->usage & VK_BUFFER_USAGE_STORAGE_BUFFER_BIT))
3581 pMemoryRequirements->size = align_u64(buffer->size, 4);
3582
3583 pMemoryRequirements->memoryTypeBits = memory_types;
3584 }
3585
3586 void anv_GetBufferMemoryRequirements2(
3587 VkDevice _device,
3588 const VkBufferMemoryRequirementsInfo2* pInfo,
3589 VkMemoryRequirements2* pMemoryRequirements)
3590 {
3591 anv_GetBufferMemoryRequirements(_device, pInfo->buffer,
3592 &pMemoryRequirements->memoryRequirements);
3593
3594 vk_foreach_struct(ext, pMemoryRequirements->pNext) {
3595 switch (ext->sType) {
3596 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
3597 VkMemoryDedicatedRequirements *requirements = (void *)ext;
3598 requirements->prefersDedicatedAllocation = false;
3599 requirements->requiresDedicatedAllocation = false;
3600 break;
3601 }
3602
3603 default:
3604 anv_debug_ignored_stype(ext->sType);
3605 break;
3606 }
3607 }
3608 }
3609
3610 void anv_GetImageMemoryRequirements(
3611 VkDevice _device,
3612 VkImage _image,
3613 VkMemoryRequirements* pMemoryRequirements)
3614 {
3615 ANV_FROM_HANDLE(anv_image, image, _image);
3616 ANV_FROM_HANDLE(anv_device, device, _device);
3617 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3618
3619 /* The Vulkan spec (git aaed022) says:
3620 *
3621 * memoryTypeBits is a bitfield and contains one bit set for every
3622 * supported memory type for the resource. The bit `1<<i` is set if and
3623 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3624 * structure for the physical device is supported.
3625 *
3626 * All types are currently supported for images.
3627 */
3628 uint32_t memory_types = (1ull << pdevice->memory.type_count) - 1;
3629
3630 /* We must have image allocated or imported at this point. According to the
3631 * specification, external images must have been bound to memory before
3632 * calling GetImageMemoryRequirements.
3633 */
3634 assert(image->size > 0);
3635
3636 pMemoryRequirements->size = image->size;
3637 pMemoryRequirements->alignment = image->alignment;
3638 pMemoryRequirements->memoryTypeBits = memory_types;
3639 }
3640
3641 void anv_GetImageMemoryRequirements2(
3642 VkDevice _device,
3643 const VkImageMemoryRequirementsInfo2* pInfo,
3644 VkMemoryRequirements2* pMemoryRequirements)
3645 {
3646 ANV_FROM_HANDLE(anv_device, device, _device);
3647 ANV_FROM_HANDLE(anv_image, image, pInfo->image);
3648
3649 anv_GetImageMemoryRequirements(_device, pInfo->image,
3650 &pMemoryRequirements->memoryRequirements);
3651
3652 vk_foreach_struct_const(ext, pInfo->pNext) {
3653 switch (ext->sType) {
3654 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO: {
3655 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3656 const VkImagePlaneMemoryRequirementsInfo *plane_reqs =
3657 (const VkImagePlaneMemoryRequirementsInfo *) ext;
3658 uint32_t plane = anv_image_aspect_to_plane(image->aspects,
3659 plane_reqs->planeAspect);
3660
3661 assert(image->planes[plane].offset == 0);
3662
3663 /* The Vulkan spec (git aaed022) says:
3664 *
3665 * memoryTypeBits is a bitfield and contains one bit set for every
3666 * supported memory type for the resource. The bit `1<<i` is set
3667 * if and only if the memory type `i` in the
3668 * VkPhysicalDeviceMemoryProperties structure for the physical
3669 * device is supported.
3670 *
3671 * All types are currently supported for images.
3672 */
3673 pMemoryRequirements->memoryRequirements.memoryTypeBits =
3674 (1ull << pdevice->memory.type_count) - 1;
3675
3676 /* We must have image allocated or imported at this point. According to the
3677 * specification, external images must have been bound to memory before
3678 * calling GetImageMemoryRequirements.
3679 */
3680 assert(image->planes[plane].size > 0);
3681
3682 pMemoryRequirements->memoryRequirements.size = image->planes[plane].size;
3683 pMemoryRequirements->memoryRequirements.alignment =
3684 image->planes[plane].alignment;
3685 break;
3686 }
3687
3688 default:
3689 anv_debug_ignored_stype(ext->sType);
3690 break;
3691 }
3692 }
3693
3694 vk_foreach_struct(ext, pMemoryRequirements->pNext) {
3695 switch (ext->sType) {
3696 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
3697 VkMemoryDedicatedRequirements *requirements = (void *)ext;
3698 if (image->needs_set_tiling || image->external_format) {
3699 /* If we need to set the tiling for external consumers, we need a
3700 * dedicated allocation.
3701 *
3702 * See also anv_AllocateMemory.
3703 */
3704 requirements->prefersDedicatedAllocation = true;
3705 requirements->requiresDedicatedAllocation = true;
3706 } else {
3707 requirements->prefersDedicatedAllocation = false;
3708 requirements->requiresDedicatedAllocation = false;
3709 }
3710 break;
3711 }
3712
3713 default:
3714 anv_debug_ignored_stype(ext->sType);
3715 break;
3716 }
3717 }
3718 }
3719
3720 void anv_GetImageSparseMemoryRequirements(
3721 VkDevice device,
3722 VkImage image,
3723 uint32_t* pSparseMemoryRequirementCount,
3724 VkSparseImageMemoryRequirements* pSparseMemoryRequirements)
3725 {
3726 *pSparseMemoryRequirementCount = 0;
3727 }
3728
3729 void anv_GetImageSparseMemoryRequirements2(
3730 VkDevice device,
3731 const VkImageSparseMemoryRequirementsInfo2* pInfo,
3732 uint32_t* pSparseMemoryRequirementCount,
3733 VkSparseImageMemoryRequirements2* pSparseMemoryRequirements)
3734 {
3735 *pSparseMemoryRequirementCount = 0;
3736 }
3737
3738 void anv_GetDeviceMemoryCommitment(
3739 VkDevice device,
3740 VkDeviceMemory memory,
3741 VkDeviceSize* pCommittedMemoryInBytes)
3742 {
3743 *pCommittedMemoryInBytes = 0;
3744 }
3745
3746 static void
3747 anv_bind_buffer_memory(const VkBindBufferMemoryInfo *pBindInfo)
3748 {
3749 ANV_FROM_HANDLE(anv_device_memory, mem, pBindInfo->memory);
3750 ANV_FROM_HANDLE(anv_buffer, buffer, pBindInfo->buffer);
3751
3752 assert(pBindInfo->sType == VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO);
3753
3754 if (mem) {
3755 buffer->address = (struct anv_address) {
3756 .bo = mem->bo,
3757 .offset = pBindInfo->memoryOffset,
3758 };
3759 } else {
3760 buffer->address = ANV_NULL_ADDRESS;
3761 }
3762 }
3763
3764 VkResult anv_BindBufferMemory(
3765 VkDevice device,
3766 VkBuffer buffer,
3767 VkDeviceMemory memory,
3768 VkDeviceSize memoryOffset)
3769 {
3770 anv_bind_buffer_memory(
3771 &(VkBindBufferMemoryInfo) {
3772 .sType = VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO,
3773 .buffer = buffer,
3774 .memory = memory,
3775 .memoryOffset = memoryOffset,
3776 });
3777
3778 return VK_SUCCESS;
3779 }
3780
3781 VkResult anv_BindBufferMemory2(
3782 VkDevice device,
3783 uint32_t bindInfoCount,
3784 const VkBindBufferMemoryInfo* pBindInfos)
3785 {
3786 for (uint32_t i = 0; i < bindInfoCount; i++)
3787 anv_bind_buffer_memory(&pBindInfos[i]);
3788
3789 return VK_SUCCESS;
3790 }
3791
3792 VkResult anv_QueueBindSparse(
3793 VkQueue _queue,
3794 uint32_t bindInfoCount,
3795 const VkBindSparseInfo* pBindInfo,
3796 VkFence fence)
3797 {
3798 ANV_FROM_HANDLE(anv_queue, queue, _queue);
3799 if (anv_device_is_lost(queue->device))
3800 return VK_ERROR_DEVICE_LOST;
3801
3802 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
3803 }
3804
3805 // Event functions
3806
3807 VkResult anv_CreateEvent(
3808 VkDevice _device,
3809 const VkEventCreateInfo* pCreateInfo,
3810 const VkAllocationCallbacks* pAllocator,
3811 VkEvent* pEvent)
3812 {
3813 ANV_FROM_HANDLE(anv_device, device, _device);
3814 struct anv_state state;
3815 struct anv_event *event;
3816
3817 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_EVENT_CREATE_INFO);
3818
3819 state = anv_state_pool_alloc(&device->dynamic_state_pool,
3820 sizeof(*event), 8);
3821 event = state.map;
3822 event->state = state;
3823 event->semaphore = VK_EVENT_RESET;
3824
3825 if (!device->info.has_llc) {
3826 /* Make sure the writes we're flushing have landed. */
3827 __builtin_ia32_mfence();
3828 __builtin_ia32_clflush(event);
3829 }
3830
3831 *pEvent = anv_event_to_handle(event);
3832
3833 return VK_SUCCESS;
3834 }
3835
3836 void anv_DestroyEvent(
3837 VkDevice _device,
3838 VkEvent _event,
3839 const VkAllocationCallbacks* pAllocator)
3840 {
3841 ANV_FROM_HANDLE(anv_device, device, _device);
3842 ANV_FROM_HANDLE(anv_event, event, _event);
3843
3844 if (!event)
3845 return;
3846
3847 anv_state_pool_free(&device->dynamic_state_pool, event->state);
3848 }
3849
3850 VkResult anv_GetEventStatus(
3851 VkDevice _device,
3852 VkEvent _event)
3853 {
3854 ANV_FROM_HANDLE(anv_device, device, _device);
3855 ANV_FROM_HANDLE(anv_event, event, _event);
3856
3857 if (anv_device_is_lost(device))
3858 return VK_ERROR_DEVICE_LOST;
3859
3860 if (!device->info.has_llc) {
3861 /* Invalidate read cache before reading event written by GPU. */
3862 __builtin_ia32_clflush(event);
3863 __builtin_ia32_mfence();
3864
3865 }
3866
3867 return event->semaphore;
3868 }
3869
3870 VkResult anv_SetEvent(
3871 VkDevice _device,
3872 VkEvent _event)
3873 {
3874 ANV_FROM_HANDLE(anv_device, device, _device);
3875 ANV_FROM_HANDLE(anv_event, event, _event);
3876
3877 event->semaphore = VK_EVENT_SET;
3878
3879 if (!device->info.has_llc) {
3880 /* Make sure the writes we're flushing have landed. */
3881 __builtin_ia32_mfence();
3882 __builtin_ia32_clflush(event);
3883 }
3884
3885 return VK_SUCCESS;
3886 }
3887
3888 VkResult anv_ResetEvent(
3889 VkDevice _device,
3890 VkEvent _event)
3891 {
3892 ANV_FROM_HANDLE(anv_device, device, _device);
3893 ANV_FROM_HANDLE(anv_event, event, _event);
3894
3895 event->semaphore = VK_EVENT_RESET;
3896
3897 if (!device->info.has_llc) {
3898 /* Make sure the writes we're flushing have landed. */
3899 __builtin_ia32_mfence();
3900 __builtin_ia32_clflush(event);
3901 }
3902
3903 return VK_SUCCESS;
3904 }
3905
3906 // Buffer functions
3907
3908 VkResult anv_CreateBuffer(
3909 VkDevice _device,
3910 const VkBufferCreateInfo* pCreateInfo,
3911 const VkAllocationCallbacks* pAllocator,
3912 VkBuffer* pBuffer)
3913 {
3914 ANV_FROM_HANDLE(anv_device, device, _device);
3915 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3916 struct anv_buffer *buffer;
3917
3918 /* Don't allow creating buffers bigger than our address space. The real
3919 * issue here is that we may align up the buffer size and we don't want
3920 * doing so to cause roll-over. However, no one has any business
3921 * allocating a buffer larger than our GTT size.
3922 */
3923 if (pCreateInfo->size > pdevice->gtt_size)
3924 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
3925
3926 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
3927
3928 buffer = vk_alloc2(&device->alloc, pAllocator, sizeof(*buffer), 8,
3929 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
3930 if (buffer == NULL)
3931 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
3932
3933 buffer->size = pCreateInfo->size;
3934 buffer->usage = pCreateInfo->usage;
3935 buffer->address = ANV_NULL_ADDRESS;
3936
3937 *pBuffer = anv_buffer_to_handle(buffer);
3938
3939 return VK_SUCCESS;
3940 }
3941
3942 void anv_DestroyBuffer(
3943 VkDevice _device,
3944 VkBuffer _buffer,
3945 const VkAllocationCallbacks* pAllocator)
3946 {
3947 ANV_FROM_HANDLE(anv_device, device, _device);
3948 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
3949
3950 if (!buffer)
3951 return;
3952
3953 vk_free2(&device->alloc, pAllocator, buffer);
3954 }
3955
3956 VkDeviceAddress anv_GetBufferDeviceAddressKHR(
3957 VkDevice device,
3958 const VkBufferDeviceAddressInfoKHR* pInfo)
3959 {
3960 ANV_FROM_HANDLE(anv_buffer, buffer, pInfo->buffer);
3961
3962 assert(!anv_address_is_null(buffer->address));
3963 assert(buffer->address.bo->flags & EXEC_OBJECT_PINNED);
3964
3965 return anv_address_physical(buffer->address);
3966 }
3967
3968 uint64_t anv_GetBufferOpaqueCaptureAddressKHR(
3969 VkDevice device,
3970 const VkBufferDeviceAddressInfoKHR* pInfo)
3971 {
3972 return 0;
3973 }
3974
3975 uint64_t anv_GetDeviceMemoryOpaqueCaptureAddressKHR(
3976 VkDevice device,
3977 const VkDeviceMemoryOpaqueCaptureAddressInfoKHR* pInfo)
3978 {
3979 ANV_FROM_HANDLE(anv_device_memory, memory, pInfo->memory);
3980
3981 assert(memory->bo->flags & EXEC_OBJECT_PINNED);
3982 assert(memory->bo->has_client_visible_address);
3983
3984 return gen_48b_address(memory->bo->offset);
3985 }
3986
3987 void
3988 anv_fill_buffer_surface_state(struct anv_device *device, struct anv_state state,
3989 enum isl_format format,
3990 struct anv_address address,
3991 uint32_t range, uint32_t stride)
3992 {
3993 isl_buffer_fill_state(&device->isl_dev, state.map,
3994 .address = anv_address_physical(address),
3995 .mocs = device->isl_dev.mocs.internal,
3996 .size_B = range,
3997 .format = format,
3998 .swizzle = ISL_SWIZZLE_IDENTITY,
3999 .stride_B = stride);
4000 }
4001
4002 void anv_DestroySampler(
4003 VkDevice _device,
4004 VkSampler _sampler,
4005 const VkAllocationCallbacks* pAllocator)
4006 {
4007 ANV_FROM_HANDLE(anv_device, device, _device);
4008 ANV_FROM_HANDLE(anv_sampler, sampler, _sampler);
4009
4010 if (!sampler)
4011 return;
4012
4013 if (sampler->bindless_state.map) {
4014 anv_state_pool_free(&device->dynamic_state_pool,
4015 sampler->bindless_state);
4016 }
4017
4018 vk_free2(&device->alloc, pAllocator, sampler);
4019 }
4020
4021 VkResult anv_CreateFramebuffer(
4022 VkDevice _device,
4023 const VkFramebufferCreateInfo* pCreateInfo,
4024 const VkAllocationCallbacks* pAllocator,
4025 VkFramebuffer* pFramebuffer)
4026 {
4027 ANV_FROM_HANDLE(anv_device, device, _device);
4028 struct anv_framebuffer *framebuffer;
4029
4030 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
4031
4032 size_t size = sizeof(*framebuffer);
4033
4034 /* VK_KHR_imageless_framebuffer extension says:
4035 *
4036 * If flags includes VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR,
4037 * parameter pAttachments is ignored.
4038 */
4039 if (!(pCreateInfo->flags & VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR)) {
4040 size += sizeof(struct anv_image_view *) * pCreateInfo->attachmentCount;
4041 framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
4042 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
4043 if (framebuffer == NULL)
4044 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
4045
4046 for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
4047 ANV_FROM_HANDLE(anv_image_view, iview, pCreateInfo->pAttachments[i]);
4048 framebuffer->attachments[i] = iview;
4049 }
4050 framebuffer->attachment_count = pCreateInfo->attachmentCount;
4051 } else {
4052 assert(device->enabled_extensions.KHR_imageless_framebuffer);
4053 framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
4054 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
4055 if (framebuffer == NULL)
4056 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
4057
4058 framebuffer->attachment_count = 0;
4059 }
4060
4061 framebuffer->width = pCreateInfo->width;
4062 framebuffer->height = pCreateInfo->height;
4063 framebuffer->layers = pCreateInfo->layers;
4064
4065 *pFramebuffer = anv_framebuffer_to_handle(framebuffer);
4066
4067 return VK_SUCCESS;
4068 }
4069
4070 void anv_DestroyFramebuffer(
4071 VkDevice _device,
4072 VkFramebuffer _fb,
4073 const VkAllocationCallbacks* pAllocator)
4074 {
4075 ANV_FROM_HANDLE(anv_device, device, _device);
4076 ANV_FROM_HANDLE(anv_framebuffer, fb, _fb);
4077
4078 if (!fb)
4079 return;
4080
4081 vk_free2(&device->alloc, pAllocator, fb);
4082 }
4083
4084 static const VkTimeDomainEXT anv_time_domains[] = {
4085 VK_TIME_DOMAIN_DEVICE_EXT,
4086 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT,
4087 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT,
4088 };
4089
4090 VkResult anv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
4091 VkPhysicalDevice physicalDevice,
4092 uint32_t *pTimeDomainCount,
4093 VkTimeDomainEXT *pTimeDomains)
4094 {
4095 int d;
4096 VK_OUTARRAY_MAKE(out, pTimeDomains, pTimeDomainCount);
4097
4098 for (d = 0; d < ARRAY_SIZE(anv_time_domains); d++) {
4099 vk_outarray_append(&out, i) {
4100 *i = anv_time_domains[d];
4101 }
4102 }
4103
4104 return vk_outarray_status(&out);
4105 }
4106
4107 static uint64_t
4108 anv_clock_gettime(clockid_t clock_id)
4109 {
4110 struct timespec current;
4111 int ret;
4112
4113 ret = clock_gettime(clock_id, &current);
4114 if (ret < 0 && clock_id == CLOCK_MONOTONIC_RAW)
4115 ret = clock_gettime(CLOCK_MONOTONIC, &current);
4116 if (ret < 0)
4117 return 0;
4118
4119 return (uint64_t) current.tv_sec * 1000000000ULL + current.tv_nsec;
4120 }
4121
4122 #define TIMESTAMP 0x2358
4123
4124 VkResult anv_GetCalibratedTimestampsEXT(
4125 VkDevice _device,
4126 uint32_t timestampCount,
4127 const VkCalibratedTimestampInfoEXT *pTimestampInfos,
4128 uint64_t *pTimestamps,
4129 uint64_t *pMaxDeviation)
4130 {
4131 ANV_FROM_HANDLE(anv_device, device, _device);
4132 uint64_t timestamp_frequency = device->info.timestamp_frequency;
4133 int ret;
4134 int d;
4135 uint64_t begin, end;
4136 uint64_t max_clock_period = 0;
4137
4138 begin = anv_clock_gettime(CLOCK_MONOTONIC_RAW);
4139
4140 for (d = 0; d < timestampCount; d++) {
4141 switch (pTimestampInfos[d].timeDomain) {
4142 case VK_TIME_DOMAIN_DEVICE_EXT:
4143 ret = anv_gem_reg_read(device, TIMESTAMP | 1,
4144 &pTimestamps[d]);
4145
4146 if (ret != 0) {
4147 return anv_device_set_lost(device, "Failed to read the TIMESTAMP "
4148 "register: %m");
4149 }
4150 uint64_t device_period = DIV_ROUND_UP(1000000000, timestamp_frequency);
4151 max_clock_period = MAX2(max_clock_period, device_period);
4152 break;
4153 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT:
4154 pTimestamps[d] = anv_clock_gettime(CLOCK_MONOTONIC);
4155 max_clock_period = MAX2(max_clock_period, 1);
4156 break;
4157
4158 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT:
4159 pTimestamps[d] = begin;
4160 break;
4161 default:
4162 pTimestamps[d] = 0;
4163 break;
4164 }
4165 }
4166
4167 end = anv_clock_gettime(CLOCK_MONOTONIC_RAW);
4168
4169 /*
4170 * The maximum deviation is the sum of the interval over which we
4171 * perform the sampling and the maximum period of any sampled
4172 * clock. That's because the maximum skew between any two sampled
4173 * clock edges is when the sampled clock with the largest period is
4174 * sampled at the end of that period but right at the beginning of the
4175 * sampling interval and some other clock is sampled right at the
4176 * begining of its sampling period and right at the end of the
4177 * sampling interval. Let's assume the GPU has the longest clock
4178 * period and that the application is sampling GPU and monotonic:
4179 *
4180 * s e
4181 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
4182 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4183 *
4184 * g
4185 * 0 1 2 3
4186 * GPU -----_____-----_____-----_____-----_____
4187 *
4188 * m
4189 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
4190 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4191 *
4192 * Interval <----------------->
4193 * Deviation <-------------------------->
4194 *
4195 * s = read(raw) 2
4196 * g = read(GPU) 1
4197 * m = read(monotonic) 2
4198 * e = read(raw) b
4199 *
4200 * We round the sample interval up by one tick to cover sampling error
4201 * in the interval clock
4202 */
4203
4204 uint64_t sample_interval = end - begin + 1;
4205
4206 *pMaxDeviation = sample_interval + max_clock_period;
4207
4208 return VK_SUCCESS;
4209 }
4210
4211 /* vk_icd.h does not declare this function, so we declare it here to
4212 * suppress Wmissing-prototypes.
4213 */
4214 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4215 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion);
4216
4217 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4218 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion)
4219 {
4220 /* For the full details on loader interface versioning, see
4221 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
4222 * What follows is a condensed summary, to help you navigate the large and
4223 * confusing official doc.
4224 *
4225 * - Loader interface v0 is incompatible with later versions. We don't
4226 * support it.
4227 *
4228 * - In loader interface v1:
4229 * - The first ICD entrypoint called by the loader is
4230 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
4231 * entrypoint.
4232 * - The ICD must statically expose no other Vulkan symbol unless it is
4233 * linked with -Bsymbolic.
4234 * - Each dispatchable Vulkan handle created by the ICD must be
4235 * a pointer to a struct whose first member is VK_LOADER_DATA. The
4236 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
4237 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
4238 * vkDestroySurfaceKHR(). The ICD must be capable of working with
4239 * such loader-managed surfaces.
4240 *
4241 * - Loader interface v2 differs from v1 in:
4242 * - The first ICD entrypoint called by the loader is
4243 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
4244 * statically expose this entrypoint.
4245 *
4246 * - Loader interface v3 differs from v2 in:
4247 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
4248 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
4249 * because the loader no longer does so.
4250 *
4251 * - Loader interface v4 differs from v3 in:
4252 * - The ICD must implement vk_icdGetPhysicalDeviceProcAddr().
4253 */
4254 *pSupportedVersion = MIN2(*pSupportedVersion, 4u);
4255 return VK_SUCCESS;
4256 }