projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
anv: Use a pNext loop in AllocateMemory
[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_COMPUTE_SHADER_DERIVATIVES_FEATURES_NV: {
1002 VkPhysicalDeviceComputeShaderDerivativesFeaturesNV *features =
1003 (VkPhysicalDeviceComputeShaderDerivativesFeaturesNV *)ext;
1004 features->computeDerivativeGroupQuads = true;
1005 features->computeDerivativeGroupLinear = true;
1006 break;
1007 }
1008
1009 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT: {
1010 VkPhysicalDeviceConditionalRenderingFeaturesEXT *features =
1011 (VkPhysicalDeviceConditionalRenderingFeaturesEXT*)ext;
1012 features->conditionalRendering = pdevice->info.gen >= 8 ||
1013 pdevice->info.is_haswell;
1014 features->inheritedConditionalRendering = pdevice->info.gen >= 8 ||
1015 pdevice->info.is_haswell;
1016 break;
1017 }
1018
1019 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_CLIP_ENABLE_FEATURES_EXT: {
1020 VkPhysicalDeviceDepthClipEnableFeaturesEXT *features =
1021 (VkPhysicalDeviceDepthClipEnableFeaturesEXT *)ext;
1022 features->depthClipEnable = true;
1023 break;
1024 }
1025
1026 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR: {
1027 VkPhysicalDeviceFloat16Int8FeaturesKHR *features = (void *)ext;
1028 features->shaderFloat16 = pdevice->info.gen >= 8;
1029 features->shaderInt8 = pdevice->info.gen >= 8;
1030 break;
1031 }
1032
1033 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FRAGMENT_SHADER_INTERLOCK_FEATURES_EXT: {
1034 VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT *features =
1035 (VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT *)ext;
1036 features->fragmentShaderSampleInterlock = pdevice->info.gen >= 9;
1037 features->fragmentShaderPixelInterlock = pdevice->info.gen >= 9;
1038 features->fragmentShaderShadingRateInterlock = false;
1039 break;
1040 }
1041
1042 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_QUERY_RESET_FEATURES_EXT: {
1043 VkPhysicalDeviceHostQueryResetFeaturesEXT *features =
1044 (VkPhysicalDeviceHostQueryResetFeaturesEXT *)ext;
1045 features->hostQueryReset = true;
1046 break;
1047 }
1048
1049 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT: {
1050 VkPhysicalDeviceDescriptorIndexingFeaturesEXT *features =
1051 (VkPhysicalDeviceDescriptorIndexingFeaturesEXT *)ext;
1052 features->shaderInputAttachmentArrayDynamicIndexing = false;
1053 features->shaderUniformTexelBufferArrayDynamicIndexing = true;
1054 features->shaderStorageTexelBufferArrayDynamicIndexing = true;
1055 features->shaderUniformBufferArrayNonUniformIndexing = false;
1056 features->shaderSampledImageArrayNonUniformIndexing = true;
1057 features->shaderStorageBufferArrayNonUniformIndexing = true;
1058 features->shaderStorageImageArrayNonUniformIndexing = true;
1059 features->shaderInputAttachmentArrayNonUniformIndexing = false;
1060 features->shaderUniformTexelBufferArrayNonUniformIndexing = true;
1061 features->shaderStorageTexelBufferArrayNonUniformIndexing = true;
1062 features->descriptorBindingUniformBufferUpdateAfterBind = false;
1063 features->descriptorBindingSampledImageUpdateAfterBind = true;
1064 features->descriptorBindingStorageImageUpdateAfterBind = true;
1065 features->descriptorBindingStorageBufferUpdateAfterBind = true;
1066 features->descriptorBindingUniformTexelBufferUpdateAfterBind = true;
1067 features->descriptorBindingStorageTexelBufferUpdateAfterBind = true;
1068 features->descriptorBindingUpdateUnusedWhilePending = true;
1069 features->descriptorBindingPartiallyBound = true;
1070 features->descriptorBindingVariableDescriptorCount = false;
1071 features->runtimeDescriptorArray = true;
1072 break;
1073 }
1074
1075 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INDEX_TYPE_UINT8_FEATURES_EXT: {
1076 VkPhysicalDeviceIndexTypeUint8FeaturesEXT *features =
1077 (VkPhysicalDeviceIndexTypeUint8FeaturesEXT *)ext;
1078 features->indexTypeUint8 = true;
1079 break;
1080 }
1081
1082 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_FEATURES_EXT: {
1083 VkPhysicalDeviceInlineUniformBlockFeaturesEXT *features =
1084 (VkPhysicalDeviceInlineUniformBlockFeaturesEXT *)ext;
1085 features->inlineUniformBlock = true;
1086 features->descriptorBindingInlineUniformBlockUpdateAfterBind = true;
1087 break;
1088 }
1089
1090 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_FEATURES_EXT: {
1091 VkPhysicalDeviceLineRasterizationFeaturesEXT *features =
1092 (VkPhysicalDeviceLineRasterizationFeaturesEXT *)ext;
1093 features->rectangularLines = true;
1094 features->bresenhamLines = true;
1095 features->smoothLines = true;
1096 features->stippledRectangularLines = false;
1097 features->stippledBresenhamLines = true;
1098 features->stippledSmoothLines = false;
1099 break;
1100 }
1101
1102 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES: {
1103 VkPhysicalDeviceMultiviewFeatures *features =
1104 (VkPhysicalDeviceMultiviewFeatures *)ext;
1105 features->multiview = true;
1106 features->multiviewGeometryShader = true;
1107 features->multiviewTessellationShader = true;
1108 break;
1109 }
1110
1111 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGELESS_FRAMEBUFFER_FEATURES_KHR: {
1112 VkPhysicalDeviceImagelessFramebufferFeaturesKHR *features =
1113 (VkPhysicalDeviceImagelessFramebufferFeaturesKHR *)ext;
1114 features->imagelessFramebuffer = true;
1115 break;
1116 }
1117
1118 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR: {
1119 VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR *features =
1120 (VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR *)ext;
1121 features->pipelineExecutableInfo = true;
1122 break;
1123 }
1124
1125 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES: {
1126 VkPhysicalDeviceProtectedMemoryFeatures *features = (void *)ext;
1127 features->protectedMemory = false;
1128 break;
1129 }
1130
1131 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES: {
1132 VkPhysicalDeviceSamplerYcbcrConversionFeatures *features =
1133 (VkPhysicalDeviceSamplerYcbcrConversionFeatures *) ext;
1134 features->samplerYcbcrConversion = true;
1135 break;
1136 }
1137
1138 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES_EXT: {
1139 VkPhysicalDeviceScalarBlockLayoutFeaturesEXT *features =
1140 (VkPhysicalDeviceScalarBlockLayoutFeaturesEXT *)ext;
1141 features->scalarBlockLayout = true;
1142 break;
1143 }
1144
1145 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SEPARATE_DEPTH_STENCIL_LAYOUTS_FEATURES_KHR: {
1146 VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR *features =
1147 (VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR *)ext;
1148 features->separateDepthStencilLayouts = true;
1149 break;
1150 }
1151
1152 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_INT64_FEATURES_KHR: {
1153 VkPhysicalDeviceShaderAtomicInt64FeaturesKHR *features = (void *)ext;
1154 features->shaderBufferInt64Atomics =
1155 pdevice->info.gen >= 9 && pdevice->use_softpin;
1156 features->shaderSharedInt64Atomics = VK_FALSE;
1157 break;
1158 }
1159
1160 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DEMOTE_TO_HELPER_INVOCATION_FEATURES_EXT: {
1161 VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT *features = (void *)ext;
1162 features->shaderDemoteToHelperInvocation = true;
1163 break;
1164 }
1165
1166 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CLOCK_FEATURES_KHR: {
1167 VkPhysicalDeviceShaderClockFeaturesKHR *features =
1168 (VkPhysicalDeviceShaderClockFeaturesKHR *)ext;
1169 features->shaderSubgroupClock = true;
1170 features->shaderDeviceClock = false;
1171 break;
1172 }
1173
1174 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETERS_FEATURES: {
1175 VkPhysicalDeviceShaderDrawParametersFeatures *features = (void *)ext;
1176 features->shaderDrawParameters = true;
1177 break;
1178 }
1179
1180 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_SUBGROUP_EXTENDED_TYPES_FEATURES_KHR: {
1181 VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR *features =
1182 (VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR *)ext;
1183 features->shaderSubgroupExtendedTypes = true;
1184 break;
1185 }
1186
1187 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_FEATURES_EXT: {
1188 VkPhysicalDeviceSubgroupSizeControlFeaturesEXT *features =
1189 (VkPhysicalDeviceSubgroupSizeControlFeaturesEXT *)ext;
1190 features->subgroupSizeControl = true;
1191 features->computeFullSubgroups = true;
1192 break;
1193 }
1194
1195 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_FEATURES_EXT: {
1196 VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT *features =
1197 (VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT *)ext;
1198 features->texelBufferAlignment = true;
1199 break;
1200 }
1201
1202 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_FEATURES_KHR: {
1203 VkPhysicalDeviceTimelineSemaphoreFeaturesKHR *features =
1204 (VkPhysicalDeviceTimelineSemaphoreFeaturesKHR *) ext;
1205 features->timelineSemaphore = true;
1206 break;
1207 }
1208
1209 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES: {
1210 VkPhysicalDeviceVariablePointersFeatures *features = (void *)ext;
1211 features->variablePointersStorageBuffer = true;
1212 features->variablePointers = true;
1213 break;
1214 }
1215
1216 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT: {
1217 VkPhysicalDeviceTransformFeedbackFeaturesEXT *features =
1218 (VkPhysicalDeviceTransformFeedbackFeaturesEXT *)ext;
1219 features->transformFeedback = true;
1220 features->geometryStreams = true;
1221 break;
1222 }
1223
1224 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_UNIFORM_BUFFER_STANDARD_LAYOUT_FEATURES_KHR: {
1225 VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR *features =
1226 (VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR *)ext;
1227 features->uniformBufferStandardLayout = true;
1228 break;
1229 }
1230
1231 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT: {
1232 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *features =
1233 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *)ext;
1234 features->vertexAttributeInstanceRateDivisor = true;
1235 features->vertexAttributeInstanceRateZeroDivisor = true;
1236 break;
1237 }
1238
1239 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_MEMORY_MODEL_FEATURES_KHR: {
1240 VkPhysicalDeviceVulkanMemoryModelFeaturesKHR *features = (void *)ext;
1241 features->vulkanMemoryModel = true;
1242 features->vulkanMemoryModelDeviceScope = true;
1243 features->vulkanMemoryModelAvailabilityVisibilityChains = true;
1244 break;
1245 }
1246
1247 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_YCBCR_IMAGE_ARRAYS_FEATURES_EXT: {
1248 VkPhysicalDeviceYcbcrImageArraysFeaturesEXT *features =
1249 (VkPhysicalDeviceYcbcrImageArraysFeaturesEXT *)ext;
1250 features->ycbcrImageArrays = true;
1251 break;
1252 }
1253
1254 default:
1255 anv_debug_ignored_stype(ext->sType);
1256 break;
1257 }
1258 }
1259 }
1260
1261 #define MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS 64
1262
1263 #define MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS 64
1264 #define MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS 256
1265
1266 void anv_GetPhysicalDeviceProperties(
1267 VkPhysicalDevice physicalDevice,
1268 VkPhysicalDeviceProperties* pProperties)
1269 {
1270 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
1271 const struct gen_device_info *devinfo = &pdevice->info;
1272
1273 /* See assertions made when programming the buffer surface state. */
1274 const uint32_t max_raw_buffer_sz = devinfo->gen >= 7 ?
1275 (1ul << 30) : (1ul << 27);
1276
1277 const uint32_t max_ssbos = pdevice->has_a64_buffer_access ? UINT16_MAX : 64;
1278 const uint32_t max_textures =
1279 pdevice->has_bindless_images ? UINT16_MAX : 128;
1280 const uint32_t max_samplers =
1281 pdevice->has_bindless_samplers ? UINT16_MAX :
1282 (devinfo->gen >= 8 || devinfo->is_haswell) ? 128 : 16;
1283 const uint32_t max_images =
1284 pdevice->has_bindless_images ? UINT16_MAX : MAX_IMAGES;
1285
1286 /* If we can use bindless for everything, claim a high per-stage limit,
1287 * otherwise use the binding table size, minus the slots reserved for
1288 * render targets and one slot for the descriptor buffer. */
1289 const uint32_t max_per_stage =
1290 pdevice->has_bindless_images && pdevice->has_a64_buffer_access
1291 ? UINT32_MAX : MAX_BINDING_TABLE_SIZE - MAX_RTS - 1;
1292
1293 const uint32_t max_workgroup_size = 32 * devinfo->max_cs_threads;
1294
1295 VkSampleCountFlags sample_counts =
1296 isl_device_get_sample_counts(&pdevice->isl_dev);
1297
1298
1299 VkPhysicalDeviceLimits limits = {
1300 .maxImageDimension1D = (1 << 14),
1301 .maxImageDimension2D = (1 << 14),
1302 .maxImageDimension3D = (1 << 11),
1303 .maxImageDimensionCube = (1 << 14),
1304 .maxImageArrayLayers = (1 << 11),
1305 .maxTexelBufferElements = 128 * 1024 * 1024,
1306 .maxUniformBufferRange = (1ul << 27),
1307 .maxStorageBufferRange = max_raw_buffer_sz,
1308 .maxPushConstantsSize = MAX_PUSH_CONSTANTS_SIZE,
1309 .maxMemoryAllocationCount = UINT32_MAX,
1310 .maxSamplerAllocationCount = 64 * 1024,
1311 .bufferImageGranularity = 64, /* A cache line */
1312 .sparseAddressSpaceSize = 0,
1313 .maxBoundDescriptorSets = MAX_SETS,
1314 .maxPerStageDescriptorSamplers = max_samplers,
1315 .maxPerStageDescriptorUniformBuffers = MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS,
1316 .maxPerStageDescriptorStorageBuffers = max_ssbos,
1317 .maxPerStageDescriptorSampledImages = max_textures,
1318 .maxPerStageDescriptorStorageImages = max_images,
1319 .maxPerStageDescriptorInputAttachments = MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS,
1320 .maxPerStageResources = max_per_stage,
1321 .maxDescriptorSetSamplers = 6 * max_samplers, /* number of stages * maxPerStageDescriptorSamplers */
1322 .maxDescriptorSetUniformBuffers = 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS, /* number of stages * maxPerStageDescriptorUniformBuffers */
1323 .maxDescriptorSetUniformBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2,
1324 .maxDescriptorSetStorageBuffers = 6 * max_ssbos, /* number of stages * maxPerStageDescriptorStorageBuffers */
1325 .maxDescriptorSetStorageBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2,
1326 .maxDescriptorSetSampledImages = 6 * max_textures, /* number of stages * maxPerStageDescriptorSampledImages */
1327 .maxDescriptorSetStorageImages = 6 * max_images, /* number of stages * maxPerStageDescriptorStorageImages */
1328 .maxDescriptorSetInputAttachments = MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS,
1329 .maxVertexInputAttributes = MAX_VBS,
1330 .maxVertexInputBindings = MAX_VBS,
1331 .maxVertexInputAttributeOffset = 2047,
1332 .maxVertexInputBindingStride = 2048,
1333 .maxVertexOutputComponents = 128,
1334 .maxTessellationGenerationLevel = 64,
1335 .maxTessellationPatchSize = 32,
1336 .maxTessellationControlPerVertexInputComponents = 128,
1337 .maxTessellationControlPerVertexOutputComponents = 128,
1338 .maxTessellationControlPerPatchOutputComponents = 128,
1339 .maxTessellationControlTotalOutputComponents = 2048,
1340 .maxTessellationEvaluationInputComponents = 128,
1341 .maxTessellationEvaluationOutputComponents = 128,
1342 .maxGeometryShaderInvocations = 32,
1343 .maxGeometryInputComponents = 64,
1344 .maxGeometryOutputComponents = 128,
1345 .maxGeometryOutputVertices = 256,
1346 .maxGeometryTotalOutputComponents = 1024,
1347 .maxFragmentInputComponents = 116, /* 128 components - (PSIZ, CLIP_DIST0, CLIP_DIST1) */
1348 .maxFragmentOutputAttachments = 8,
1349 .maxFragmentDualSrcAttachments = 1,
1350 .maxFragmentCombinedOutputResources = 8,
1351 .maxComputeSharedMemorySize = 64 * 1024,
1352 .maxComputeWorkGroupCount = { 65535, 65535, 65535 },
1353 .maxComputeWorkGroupInvocations = max_workgroup_size,
1354 .maxComputeWorkGroupSize = {
1355 max_workgroup_size,
1356 max_workgroup_size,
1357 max_workgroup_size,
1358 },
1359 .subPixelPrecisionBits = 8,
1360 .subTexelPrecisionBits = 8,
1361 .mipmapPrecisionBits = 8,
1362 .maxDrawIndexedIndexValue = UINT32_MAX,
1363 .maxDrawIndirectCount = UINT32_MAX,
1364 .maxSamplerLodBias = 16,
1365 .maxSamplerAnisotropy = 16,
1366 .maxViewports = MAX_VIEWPORTS,
1367 .maxViewportDimensions = { (1 << 14), (1 << 14) },
1368 .viewportBoundsRange = { INT16_MIN, INT16_MAX },
1369 .viewportSubPixelBits = 13, /* We take a float? */
1370 .minMemoryMapAlignment = 4096, /* A page */
1371 /* The dataport requires texel alignment so we need to assume a worst
1372 * case of R32G32B32A32 which is 16 bytes.
1373 */
1374 .minTexelBufferOffsetAlignment = 16,
1375 /* We need 16 for UBO block reads to work and 32 for push UBOs */
1376 .minUniformBufferOffsetAlignment = 32,
1377 .minStorageBufferOffsetAlignment = 4,
1378 .minTexelOffset = -8,
1379 .maxTexelOffset = 7,
1380 .minTexelGatherOffset = -32,
1381 .maxTexelGatherOffset = 31,
1382 .minInterpolationOffset = -0.5,
1383 .maxInterpolationOffset = 0.4375,
1384 .subPixelInterpolationOffsetBits = 4,
1385 .maxFramebufferWidth = (1 << 14),
1386 .maxFramebufferHeight = (1 << 14),
1387 .maxFramebufferLayers = (1 << 11),
1388 .framebufferColorSampleCounts = sample_counts,
1389 .framebufferDepthSampleCounts = sample_counts,
1390 .framebufferStencilSampleCounts = sample_counts,
1391 .framebufferNoAttachmentsSampleCounts = sample_counts,
1392 .maxColorAttachments = MAX_RTS,
1393 .sampledImageColorSampleCounts = sample_counts,
1394 .sampledImageIntegerSampleCounts = VK_SAMPLE_COUNT_1_BIT,
1395 .sampledImageDepthSampleCounts = sample_counts,
1396 .sampledImageStencilSampleCounts = sample_counts,
1397 .storageImageSampleCounts = VK_SAMPLE_COUNT_1_BIT,
1398 .maxSampleMaskWords = 1,
1399 .timestampComputeAndGraphics = true,
1400 .timestampPeriod = 1000000000.0 / devinfo->timestamp_frequency,
1401 .maxClipDistances = 8,
1402 .maxCullDistances = 8,
1403 .maxCombinedClipAndCullDistances = 8,
1404 .discreteQueuePriorities = 2,
1405 .pointSizeRange = { 0.125, 255.875 },
1406 .lineWidthRange = {
1407 0.0,
1408 (devinfo->gen >= 9 || devinfo->is_cherryview) ?
1409 2047.9921875 : 7.9921875,
1410 },
1411 .pointSizeGranularity = (1.0 / 8.0),
1412 .lineWidthGranularity = (1.0 / 128.0),
1413 .strictLines = false,
1414 .standardSampleLocations = true,
1415 .optimalBufferCopyOffsetAlignment = 128,
1416 .optimalBufferCopyRowPitchAlignment = 128,
1417 .nonCoherentAtomSize = 64,
1418 };
1419
1420 *pProperties = (VkPhysicalDeviceProperties) {
1421 .apiVersion = anv_physical_device_api_version(pdevice),
1422 .driverVersion = vk_get_driver_version(),
1423 .vendorID = 0x8086,
1424 .deviceID = pdevice->chipset_id,
1425 .deviceType = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
1426 .limits = limits,
1427 .sparseProperties = {0}, /* Broadwell doesn't do sparse. */
1428 };
1429
1430 snprintf(pProperties->deviceName, sizeof(pProperties->deviceName),
1431 "%s", pdevice->name);
1432 memcpy(pProperties->pipelineCacheUUID,
1433 pdevice->pipeline_cache_uuid, VK_UUID_SIZE);
1434 }
1435
1436 void anv_GetPhysicalDeviceProperties2(
1437 VkPhysicalDevice physicalDevice,
1438 VkPhysicalDeviceProperties2* pProperties)
1439 {
1440 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
1441
1442 anv_GetPhysicalDeviceProperties(physicalDevice, &pProperties->properties);
1443
1444 vk_foreach_struct(ext, pProperties->pNext) {
1445 switch (ext->sType) {
1446 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_STENCIL_RESOLVE_PROPERTIES_KHR: {
1447 VkPhysicalDeviceDepthStencilResolvePropertiesKHR *props =
1448 (VkPhysicalDeviceDepthStencilResolvePropertiesKHR *)ext;
1449
1450 /* We support all of the depth resolve modes */
1451 props->supportedDepthResolveModes =
1452 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR |
1453 VK_RESOLVE_MODE_AVERAGE_BIT_KHR |
1454 VK_RESOLVE_MODE_MIN_BIT_KHR |
1455 VK_RESOLVE_MODE_MAX_BIT_KHR;
1456
1457 /* Average doesn't make sense for stencil so we don't support that */
1458 props->supportedStencilResolveModes =
1459 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR;
1460 if (pdevice->info.gen >= 8) {
1461 /* The advanced stencil resolve modes currently require stencil
1462 * sampling be supported by the hardware.
1463 */
1464 props->supportedStencilResolveModes |=
1465 VK_RESOLVE_MODE_MIN_BIT_KHR |
1466 VK_RESOLVE_MODE_MAX_BIT_KHR;
1467 }
1468
1469 props->independentResolveNone = true;
1470 props->independentResolve = true;
1471 break;
1472 }
1473
1474 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES_EXT: {
1475 VkPhysicalDeviceDescriptorIndexingPropertiesEXT *props =
1476 (VkPhysicalDeviceDescriptorIndexingPropertiesEXT *)ext;
1477
1478 /* It's a bit hard to exactly map our implementation to the limits
1479 * described here. The bindless surface handle in the extended
1480 * message descriptors is 20 bits and it's an index into the table of
1481 * RENDER_SURFACE_STATE structs that starts at bindless surface base
1482 * address. Given that most things consume two surface states per
1483 * view (general/sampled for textures and write-only/read-write for
1484 * images), we claim 2^19 things.
1485 *
1486 * For SSBOs, we just use A64 messages so there is no real limit
1487 * there beyond the limit on the total size of a descriptor set.
1488 */
1489 const unsigned max_bindless_views = 1 << 19;
1490
1491 props->maxUpdateAfterBindDescriptorsInAllPools = max_bindless_views;
1492 props->shaderUniformBufferArrayNonUniformIndexingNative = false;
1493 props->shaderSampledImageArrayNonUniformIndexingNative = false;
1494 props->shaderStorageBufferArrayNonUniformIndexingNative = true;
1495 props->shaderStorageImageArrayNonUniformIndexingNative = false;
1496 props->shaderInputAttachmentArrayNonUniformIndexingNative = false;
1497 props->robustBufferAccessUpdateAfterBind = true;
1498 props->quadDivergentImplicitLod = false;
1499 props->maxPerStageDescriptorUpdateAfterBindSamplers = max_bindless_views;
1500 props->maxPerStageDescriptorUpdateAfterBindUniformBuffers = MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS;
1501 props->maxPerStageDescriptorUpdateAfterBindStorageBuffers = UINT32_MAX;
1502 props->maxPerStageDescriptorUpdateAfterBindSampledImages = max_bindless_views;
1503 props->maxPerStageDescriptorUpdateAfterBindStorageImages = max_bindless_views;
1504 props->maxPerStageDescriptorUpdateAfterBindInputAttachments = MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS;
1505 props->maxPerStageUpdateAfterBindResources = UINT32_MAX;
1506 props->maxDescriptorSetUpdateAfterBindSamplers = max_bindless_views;
1507 props->maxDescriptorSetUpdateAfterBindUniformBuffers = 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS;
1508 props->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2;
1509 props->maxDescriptorSetUpdateAfterBindStorageBuffers = UINT32_MAX;
1510 props->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2;
1511 props->maxDescriptorSetUpdateAfterBindSampledImages = max_bindless_views;
1512 props->maxDescriptorSetUpdateAfterBindStorageImages = max_bindless_views;
1513 props->maxDescriptorSetUpdateAfterBindInputAttachments = MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS;
1514 break;
1515 }
1516
1517 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR: {
1518 VkPhysicalDeviceDriverPropertiesKHR *driver_props =
1519 (VkPhysicalDeviceDriverPropertiesKHR *) ext;
1520
1521 driver_props->driverID = VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA_KHR;
1522 snprintf(driver_props->driverName, VK_MAX_DRIVER_NAME_SIZE_KHR,
1523 "Intel open-source Mesa driver");
1524
1525 snprintf(driver_props->driverInfo, VK_MAX_DRIVER_INFO_SIZE_KHR,
1526 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1);
1527
1528 driver_props->conformanceVersion = (VkConformanceVersionKHR) {
1529 .major = 1,
1530 .minor = 1,
1531 .subminor = 2,
1532 .patch = 0,
1533 };
1534 break;
1535 }
1536
1537 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT: {
1538 VkPhysicalDeviceExternalMemoryHostPropertiesEXT *props =
1539 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT *) ext;
1540 /* Userptr needs page aligned memory. */
1541 props->minImportedHostPointerAlignment = 4096;
1542 break;
1543 }
1544
1545 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES: {
1546 VkPhysicalDeviceIDProperties *id_props =
1547 (VkPhysicalDeviceIDProperties *)ext;
1548 memcpy(id_props->deviceUUID, pdevice->device_uuid, VK_UUID_SIZE);
1549 memcpy(id_props->driverUUID, pdevice->driver_uuid, VK_UUID_SIZE);
1550 /* The LUID is for Windows. */
1551 id_props->deviceLUIDValid = false;
1552 break;
1553 }
1554
1555 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_PROPERTIES_EXT: {
1556 VkPhysicalDeviceInlineUniformBlockPropertiesEXT *props =
1557 (VkPhysicalDeviceInlineUniformBlockPropertiesEXT *)ext;
1558 props->maxInlineUniformBlockSize = MAX_INLINE_UNIFORM_BLOCK_SIZE;
1559 props->maxPerStageDescriptorInlineUniformBlocks =
1560 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1561 props->maxPerStageDescriptorUpdateAfterBindInlineUniformBlocks =
1562 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1563 props->maxDescriptorSetInlineUniformBlocks =
1564 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1565 props->maxDescriptorSetUpdateAfterBindInlineUniformBlocks =
1566 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1567 break;
1568 }
1569
1570 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_PROPERTIES_EXT: {
1571 VkPhysicalDeviceLineRasterizationPropertiesEXT *props =
1572 (VkPhysicalDeviceLineRasterizationPropertiesEXT *)ext;
1573 /* In the Skylake PRM Vol. 7, subsection titled "GIQ (Diamond)
1574 * Sampling Rules - Legacy Mode", it says the following:
1575 *
1576 * "Note that the device divides a pixel into a 16x16 array of
1577 * subpixels, referenced by their upper left corners."
1578 *
1579 * This is the only known reference in the PRMs to the subpixel
1580 * precision of line rasterization and a "16x16 array of subpixels"
1581 * implies 4 subpixel precision bits. Empirical testing has shown
1582 * that 4 subpixel precision bits applies to all line rasterization
1583 * types.
1584 */
1585 props->lineSubPixelPrecisionBits = 4;
1586 break;
1587 }
1588
1589 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES: {
1590 VkPhysicalDeviceMaintenance3Properties *props =
1591 (VkPhysicalDeviceMaintenance3Properties *)ext;
1592 /* This value doesn't matter for us today as our per-stage
1593 * descriptors are the real limit.
1594 */
1595 props->maxPerSetDescriptors = 1024;
1596 props->maxMemoryAllocationSize = MAX_MEMORY_ALLOCATION_SIZE;
1597 break;
1598 }
1599
1600 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES: {
1601 VkPhysicalDeviceMultiviewProperties *properties =
1602 (VkPhysicalDeviceMultiviewProperties *)ext;
1603 properties->maxMultiviewViewCount = 16;
1604 properties->maxMultiviewInstanceIndex = UINT32_MAX / 16;
1605 break;
1606 }
1607
1608 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT: {
1609 VkPhysicalDevicePCIBusInfoPropertiesEXT *properties =
1610 (VkPhysicalDevicePCIBusInfoPropertiesEXT *)ext;
1611 properties->pciDomain = pdevice->pci_info.domain;
1612 properties->pciBus = pdevice->pci_info.bus;
1613 properties->pciDevice = pdevice->pci_info.device;
1614 properties->pciFunction = pdevice->pci_info.function;
1615 break;
1616 }
1617
1618 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES: {
1619 VkPhysicalDevicePointClippingProperties *properties =
1620 (VkPhysicalDevicePointClippingProperties *) ext;
1621 properties->pointClippingBehavior = VK_POINT_CLIPPING_BEHAVIOR_USER_CLIP_PLANES_ONLY;
1622 break;
1623 }
1624
1625 #pragma GCC diagnostic push
1626 #pragma GCC diagnostic ignored "-Wswitch"
1627 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PRESENTATION_PROPERTIES_ANDROID: {
1628 VkPhysicalDevicePresentationPropertiesANDROID *props =
1629 (VkPhysicalDevicePresentationPropertiesANDROID *)ext;
1630 props->sharedImage = VK_FALSE;
1631 break;
1632 }
1633 #pragma GCC diagnostic pop
1634
1635 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES: {
1636 VkPhysicalDeviceProtectedMemoryProperties *props =
1637 (VkPhysicalDeviceProtectedMemoryProperties *)ext;
1638 props->protectedNoFault = false;
1639 break;
1640 }
1641
1642 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR: {
1643 VkPhysicalDevicePushDescriptorPropertiesKHR *properties =
1644 (VkPhysicalDevicePushDescriptorPropertiesKHR *) ext;
1645
1646 properties->maxPushDescriptors = MAX_PUSH_DESCRIPTORS;
1647 break;
1648 }
1649
1650 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT: {
1651 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT *properties =
1652 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT *)ext;
1653 properties->filterMinmaxImageComponentMapping = pdevice->info.gen >= 9;
1654 properties->filterMinmaxSingleComponentFormats = true;
1655 break;
1656 }
1657
1658 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES: {
1659 VkPhysicalDeviceSubgroupProperties *properties = (void *)ext;
1660
1661 properties->subgroupSize = BRW_SUBGROUP_SIZE;
1662
1663 VkShaderStageFlags scalar_stages = 0;
1664 for (unsigned stage = 0; stage < MESA_SHADER_STAGES; stage++) {
1665 if (pdevice->compiler->scalar_stage[stage])
1666 scalar_stages |= mesa_to_vk_shader_stage(stage);
1667 }
1668 properties->supportedStages = scalar_stages;
1669
1670 properties->supportedOperations = VK_SUBGROUP_FEATURE_BASIC_BIT |
1671 VK_SUBGROUP_FEATURE_VOTE_BIT |
1672 VK_SUBGROUP_FEATURE_BALLOT_BIT |
1673 VK_SUBGROUP_FEATURE_SHUFFLE_BIT |
1674 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT |
1675 VK_SUBGROUP_FEATURE_QUAD_BIT;
1676 if (pdevice->info.gen >= 8) {
1677 /* TODO: There's no technical reason why these can't be made to
1678 * work on gen7 but they don't at the moment so it's best to leave
1679 * the feature disabled than enabled and broken.
1680 */
1681 properties->supportedOperations |=
1682 VK_SUBGROUP_FEATURE_ARITHMETIC_BIT |
1683 VK_SUBGROUP_FEATURE_CLUSTERED_BIT;
1684 }
1685 properties->quadOperationsInAllStages = pdevice->info.gen >= 8;
1686 break;
1687 }
1688
1689 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_PROPERTIES_EXT: {
1690 VkPhysicalDeviceSubgroupSizeControlPropertiesEXT *props =
1691 (VkPhysicalDeviceSubgroupSizeControlPropertiesEXT *)ext;
1692 STATIC_ASSERT(8 <= BRW_SUBGROUP_SIZE && BRW_SUBGROUP_SIZE <= 32);
1693 props->minSubgroupSize = 8;
1694 props->maxSubgroupSize = 32;
1695 props->maxComputeWorkgroupSubgroups = pdevice->info.max_cs_threads;
1696 props->requiredSubgroupSizeStages = VK_SHADER_STAGE_COMPUTE_BIT;
1697 break;
1698 }
1699 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT_CONTROLS_PROPERTIES_KHR : {
1700 VkPhysicalDeviceFloatControlsPropertiesKHR *properties = (void *)ext;
1701 properties->denormBehaviorIndependence = VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_ALL_KHR;
1702 properties->roundingModeIndependence = VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_NONE_KHR;
1703
1704 /* Broadwell does not support HF denorms and there are restrictions
1705 * other gens. According to Kabylake's PRM:
1706 *
1707 * "math - Extended Math Function
1708 * [...]
1709 * Restriction : Half-float denorms are always retained."
1710 */
1711 properties->shaderDenormFlushToZeroFloat16 = false;
1712 properties->shaderDenormPreserveFloat16 = pdevice->info.gen > 8;
1713 properties->shaderRoundingModeRTEFloat16 = true;
1714 properties->shaderRoundingModeRTZFloat16 = true;
1715 properties->shaderSignedZeroInfNanPreserveFloat16 = true;
1716
1717 properties->shaderDenormFlushToZeroFloat32 = true;
1718 properties->shaderDenormPreserveFloat32 = true;
1719 properties->shaderRoundingModeRTEFloat32 = true;
1720 properties->shaderRoundingModeRTZFloat32 = true;
1721 properties->shaderSignedZeroInfNanPreserveFloat32 = true;
1722
1723 properties->shaderDenormFlushToZeroFloat64 = true;
1724 properties->shaderDenormPreserveFloat64 = true;
1725 properties->shaderRoundingModeRTEFloat64 = true;
1726 properties->shaderRoundingModeRTZFloat64 = true;
1727 properties->shaderSignedZeroInfNanPreserveFloat64 = true;
1728 break;
1729 }
1730
1731 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_PROPERTIES_EXT: {
1732 VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT *props =
1733 (VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT *)ext;
1734
1735 /* From the SKL PRM Vol. 2d, docs for RENDER_SURFACE_STATE::Surface
1736 * Base Address:
1737 *
1738 * "For SURFTYPE_BUFFER non-rendertarget surfaces, this field
1739 * specifies the base address of the first element of the surface,
1740 * computed in software by adding the surface base address to the
1741 * byte offset of the element in the buffer. The base address must
1742 * be aligned to element size."
1743 *
1744 * The typed dataport messages require that things be texel aligned.
1745 * Otherwise, we may just load/store the wrong data or, in the worst
1746 * case, there may be hangs.
1747 */
1748 props->storageTexelBufferOffsetAlignmentBytes = 16;
1749 props->storageTexelBufferOffsetSingleTexelAlignment = true;
1750
1751 /* The sampler, however, is much more forgiving and it can handle
1752 * arbitrary byte alignment for linear and buffer surfaces. It's
1753 * hard to find a good PRM citation for this but years of empirical
1754 * experience demonstrate that this is true.
1755 */
1756 props->uniformTexelBufferOffsetAlignmentBytes = 1;
1757 props->uniformTexelBufferOffsetSingleTexelAlignment = false;
1758 break;
1759 }
1760
1761 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_PROPERTIES_KHR: {
1762 VkPhysicalDeviceTimelineSemaphorePropertiesKHR *props =
1763 (VkPhysicalDeviceTimelineSemaphorePropertiesKHR *) ext;
1764 props->maxTimelineSemaphoreValueDifference = UINT64_MAX;
1765 break;
1766 }
1767
1768 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT: {
1769 VkPhysicalDeviceTransformFeedbackPropertiesEXT *props =
1770 (VkPhysicalDeviceTransformFeedbackPropertiesEXT *)ext;
1771
1772 props->maxTransformFeedbackStreams = MAX_XFB_STREAMS;
1773 props->maxTransformFeedbackBuffers = MAX_XFB_BUFFERS;
1774 props->maxTransformFeedbackBufferSize = (1ull << 32);
1775 props->maxTransformFeedbackStreamDataSize = 128 * 4;
1776 props->maxTransformFeedbackBufferDataSize = 128 * 4;
1777 props->maxTransformFeedbackBufferDataStride = 2048;
1778 props->transformFeedbackQueries = true;
1779 props->transformFeedbackStreamsLinesTriangles = false;
1780 props->transformFeedbackRasterizationStreamSelect = false;
1781 props->transformFeedbackDraw = true;
1782 break;
1783 }
1784
1785 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT: {
1786 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *props =
1787 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *)ext;
1788 /* We have to restrict this a bit for multiview */
1789 props->maxVertexAttribDivisor = UINT32_MAX / 16;
1790 break;
1791 }
1792
1793 default:
1794 anv_debug_ignored_stype(ext->sType);
1795 break;
1796 }
1797 }
1798 }
1799
1800 /* We support exactly one queue family. */
1801 static const VkQueueFamilyProperties
1802 anv_queue_family_properties = {
1803 .queueFlags = VK_QUEUE_GRAPHICS_BIT |
1804 VK_QUEUE_COMPUTE_BIT |
1805 VK_QUEUE_TRANSFER_BIT,
1806 .queueCount = 1,
1807 .timestampValidBits = 36, /* XXX: Real value here */
1808 .minImageTransferGranularity = { 1, 1, 1 },
1809 };
1810
1811 void anv_GetPhysicalDeviceQueueFamilyProperties(
1812 VkPhysicalDevice physicalDevice,
1813 uint32_t* pCount,
1814 VkQueueFamilyProperties* pQueueFamilyProperties)
1815 {
1816 VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pCount);
1817
1818 vk_outarray_append(&out, p) {
1819 *p = anv_queue_family_properties;
1820 }
1821 }
1822
1823 void anv_GetPhysicalDeviceQueueFamilyProperties2(
1824 VkPhysicalDevice physicalDevice,
1825 uint32_t* pQueueFamilyPropertyCount,
1826 VkQueueFamilyProperties2* pQueueFamilyProperties)
1827 {
1828
1829 VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pQueueFamilyPropertyCount);
1830
1831 vk_outarray_append(&out, p) {
1832 p->queueFamilyProperties = anv_queue_family_properties;
1833
1834 vk_foreach_struct(s, p->pNext) {
1835 anv_debug_ignored_stype(s->sType);
1836 }
1837 }
1838 }
1839
1840 void anv_GetPhysicalDeviceMemoryProperties(
1841 VkPhysicalDevice physicalDevice,
1842 VkPhysicalDeviceMemoryProperties* pMemoryProperties)
1843 {
1844 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
1845
1846 pMemoryProperties->memoryTypeCount = physical_device->memory.type_count;
1847 for (uint32_t i = 0; i < physical_device->memory.type_count; i++) {
1848 pMemoryProperties->memoryTypes[i] = (VkMemoryType) {
1849 .propertyFlags = physical_device->memory.types[i].propertyFlags,
1850 .heapIndex = physical_device->memory.types[i].heapIndex,
1851 };
1852 }
1853
1854 pMemoryProperties->memoryHeapCount = physical_device->memory.heap_count;
1855 for (uint32_t i = 0; i < physical_device->memory.heap_count; i++) {
1856 pMemoryProperties->memoryHeaps[i] = (VkMemoryHeap) {
1857 .size = physical_device->memory.heaps[i].size,
1858 .flags = physical_device->memory.heaps[i].flags,
1859 };
1860 }
1861 }
1862
1863 static void
1864 anv_get_memory_budget(VkPhysicalDevice physicalDevice,
1865 VkPhysicalDeviceMemoryBudgetPropertiesEXT *memoryBudget)
1866 {
1867 ANV_FROM_HANDLE(anv_physical_device, device, physicalDevice);
1868 uint64_t sys_available = get_available_system_memory();
1869 assert(sys_available > 0);
1870
1871 VkDeviceSize total_heaps_size = 0;
1872 for (size_t i = 0; i < device->memory.heap_count; i++)
1873 total_heaps_size += device->memory.heaps[i].size;
1874
1875 for (size_t i = 0; i < device->memory.heap_count; i++) {
1876 VkDeviceSize heap_size = device->memory.heaps[i].size;
1877 VkDeviceSize heap_used = device->memory.heaps[i].used;
1878 VkDeviceSize heap_budget;
1879
1880 double heap_proportion = (double) heap_size / total_heaps_size;
1881 VkDeviceSize sys_available_prop = sys_available * heap_proportion;
1882
1883 /*
1884 * Let's not incite the app to starve the system: report at most 90% of
1885 * available system memory.
1886 */
1887 uint64_t heap_available = sys_available_prop * 9 / 10;
1888 heap_budget = MIN2(heap_size, heap_used + heap_available);
1889
1890 /*
1891 * Round down to the nearest MB
1892 */
1893 heap_budget &= ~((1ull << 20) - 1);
1894
1895 /*
1896 * The heapBudget value must be non-zero for array elements less than
1897 * VkPhysicalDeviceMemoryProperties::memoryHeapCount. The heapBudget
1898 * value must be less than or equal to VkMemoryHeap::size for each heap.
1899 */
1900 assert(0 < heap_budget && heap_budget <= heap_size);
1901
1902 memoryBudget->heapUsage[i] = heap_used;
1903 memoryBudget->heapBudget[i] = heap_budget;
1904 }
1905
1906 /* The heapBudget and heapUsage values must be zero for array elements
1907 * greater than or equal to VkPhysicalDeviceMemoryProperties::memoryHeapCount
1908 */
1909 for (uint32_t i = device->memory.heap_count; i < VK_MAX_MEMORY_HEAPS; i++) {
1910 memoryBudget->heapBudget[i] = 0;
1911 memoryBudget->heapUsage[i] = 0;
1912 }
1913 }
1914
1915 void anv_GetPhysicalDeviceMemoryProperties2(
1916 VkPhysicalDevice physicalDevice,
1917 VkPhysicalDeviceMemoryProperties2* pMemoryProperties)
1918 {
1919 anv_GetPhysicalDeviceMemoryProperties(physicalDevice,
1920 &pMemoryProperties->memoryProperties);
1921
1922 vk_foreach_struct(ext, pMemoryProperties->pNext) {
1923 switch (ext->sType) {
1924 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT:
1925 anv_get_memory_budget(physicalDevice, (void*)ext);
1926 break;
1927 default:
1928 anv_debug_ignored_stype(ext->sType);
1929 break;
1930 }
1931 }
1932 }
1933
1934 void
1935 anv_GetDeviceGroupPeerMemoryFeatures(
1936 VkDevice device,
1937 uint32_t heapIndex,
1938 uint32_t localDeviceIndex,
1939 uint32_t remoteDeviceIndex,
1940 VkPeerMemoryFeatureFlags* pPeerMemoryFeatures)
1941 {
1942 assert(localDeviceIndex == 0 && remoteDeviceIndex == 0);
1943 *pPeerMemoryFeatures = VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT |
1944 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT |
1945 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT |
1946 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT;
1947 }
1948
1949 PFN_vkVoidFunction anv_GetInstanceProcAddr(
1950 VkInstance _instance,
1951 const char* pName)
1952 {
1953 ANV_FROM_HANDLE(anv_instance, instance, _instance);
1954
1955 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
1956 * when we have to return valid function pointers, NULL, or it's left
1957 * undefined. See the table for exact details.
1958 */
1959 if (pName == NULL)
1960 return NULL;
1961
1962 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
1963 if (strcmp(pName, "vk" #entrypoint) == 0) \
1964 return (PFN_vkVoidFunction)anv_##entrypoint
1965
1966 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties);
1967 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties);
1968 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceVersion);
1969 LOOKUP_ANV_ENTRYPOINT(CreateInstance);
1970
1971 #undef LOOKUP_ANV_ENTRYPOINT
1972
1973 if (instance == NULL)
1974 return NULL;
1975
1976 int idx = anv_get_instance_entrypoint_index(pName);
1977 if (idx >= 0)
1978 return instance->dispatch.entrypoints[idx];
1979
1980 idx = anv_get_physical_device_entrypoint_index(pName);
1981 if (idx >= 0)
1982 return instance->physicalDevice.dispatch.entrypoints[idx];
1983
1984 idx = anv_get_device_entrypoint_index(pName);
1985 if (idx >= 0)
1986 return instance->device_dispatch.entrypoints[idx];
1987
1988 return NULL;
1989 }
1990
1991 /* With version 1+ of the loader interface the ICD should expose
1992 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
1993 */
1994 PUBLIC
1995 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
1996 VkInstance instance,
1997 const char* pName);
1998
1999 PUBLIC
2000 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
2001 VkInstance instance,
2002 const char* pName)
2003 {
2004 return anv_GetInstanceProcAddr(instance, pName);
2005 }
2006
2007 PFN_vkVoidFunction anv_GetDeviceProcAddr(
2008 VkDevice _device,
2009 const char* pName)
2010 {
2011 ANV_FROM_HANDLE(anv_device, device, _device);
2012
2013 if (!device || !pName)
2014 return NULL;
2015
2016 int idx = anv_get_device_entrypoint_index(pName);
2017 if (idx < 0)
2018 return NULL;
2019
2020 return device->dispatch.entrypoints[idx];
2021 }
2022
2023 /* With version 4+ of the loader interface the ICD should expose
2024 * vk_icdGetPhysicalDeviceProcAddr()
2025 */
2026 PUBLIC
2027 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetPhysicalDeviceProcAddr(
2028 VkInstance _instance,
2029 const char* pName);
2030
2031 PFN_vkVoidFunction vk_icdGetPhysicalDeviceProcAddr(
2032 VkInstance _instance,
2033 const char* pName)
2034 {
2035 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2036
2037 if (!pName || !instance)
2038 return NULL;
2039
2040 int idx = anv_get_physical_device_entrypoint_index(pName);
2041 if (idx < 0)
2042 return NULL;
2043
2044 return instance->physicalDevice.dispatch.entrypoints[idx];
2045 }
2046
2047
2048 VkResult
2049 anv_CreateDebugReportCallbackEXT(VkInstance _instance,
2050 const VkDebugReportCallbackCreateInfoEXT* pCreateInfo,
2051 const VkAllocationCallbacks* pAllocator,
2052 VkDebugReportCallbackEXT* pCallback)
2053 {
2054 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2055 return vk_create_debug_report_callback(&instance->debug_report_callbacks,
2056 pCreateInfo, pAllocator, &instance->alloc,
2057 pCallback);
2058 }
2059
2060 void
2061 anv_DestroyDebugReportCallbackEXT(VkInstance _instance,
2062 VkDebugReportCallbackEXT _callback,
2063 const VkAllocationCallbacks* pAllocator)
2064 {
2065 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2066 vk_destroy_debug_report_callback(&instance->debug_report_callbacks,
2067 _callback, pAllocator, &instance->alloc);
2068 }
2069
2070 void
2071 anv_DebugReportMessageEXT(VkInstance _instance,
2072 VkDebugReportFlagsEXT flags,
2073 VkDebugReportObjectTypeEXT objectType,
2074 uint64_t object,
2075 size_t location,
2076 int32_t messageCode,
2077 const char* pLayerPrefix,
2078 const char* pMessage)
2079 {
2080 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2081 vk_debug_report(&instance->debug_report_callbacks, flags, objectType,
2082 object, location, messageCode, pLayerPrefix, pMessage);
2083 }
2084
2085 static struct anv_state
2086 anv_state_pool_emit_data(struct anv_state_pool *pool, size_t size, size_t align, const void *p)
2087 {
2088 struct anv_state state;
2089
2090 state = anv_state_pool_alloc(pool, size, align);
2091 memcpy(state.map, p, size);
2092
2093 return state;
2094 }
2095
2096 /* Haswell border color is a bit of a disaster. Float and unorm formats use a
2097 * straightforward 32-bit float color in the first 64 bytes. Instead of using
2098 * a nice float/integer union like Gen8+, Haswell specifies the integer border
2099 * color as a separate entry /after/ the float color. The layout of this entry
2100 * also depends on the format's bpp (with extra hacks for RG32), and overlaps.
2101 *
2102 * Since we don't know the format/bpp, we can't make any of the border colors
2103 * containing '1' work for all formats, as it would be in the wrong place for
2104 * some of them. We opt to make 32-bit integers work as this seems like the
2105 * most common option. Fortunately, transparent black works regardless, as
2106 * all zeroes is the same in every bit-size.
2107 */
2108 struct hsw_border_color {
2109 float float32[4];
2110 uint32_t _pad0[12];
2111 uint32_t uint32[4];
2112 uint32_t _pad1[108];
2113 };
2114
2115 struct gen8_border_color {
2116 union {
2117 float float32[4];
2118 uint32_t uint32[4];
2119 };
2120 /* Pad out to 64 bytes */
2121 uint32_t _pad[12];
2122 };
2123
2124 static void
2125 anv_device_init_border_colors(struct anv_device *device)
2126 {
2127 if (device->info.is_haswell) {
2128 static const struct hsw_border_color border_colors[] = {
2129 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
2130 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
2131 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
2132 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = { .uint32 = { 0, 0, 0, 0 } },
2133 [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = { .uint32 = { 0, 0, 0, 1 } },
2134 [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = { .uint32 = { 1, 1, 1, 1 } },
2135 };
2136
2137 device->border_colors =
2138 anv_state_pool_emit_data(&device->dynamic_state_pool,
2139 sizeof(border_colors), 512, border_colors);
2140 } else {
2141 static const struct gen8_border_color border_colors[] = {
2142 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
2143 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
2144 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
2145 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = { .uint32 = { 0, 0, 0, 0 } },
2146 [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = { .uint32 = { 0, 0, 0, 1 } },
2147 [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = { .uint32 = { 1, 1, 1, 1 } },
2148 };
2149
2150 device->border_colors =
2151 anv_state_pool_emit_data(&device->dynamic_state_pool,
2152 sizeof(border_colors), 64, border_colors);
2153 }
2154 }
2155
2156 static VkResult
2157 anv_device_init_trivial_batch(struct anv_device *device)
2158 {
2159 VkResult result = anv_device_alloc_bo(device, 4096,
2160 ANV_BO_ALLOC_MAPPED,
2161 0 /* explicit_address */,
2162 &device->trivial_batch_bo);
2163 if (result != VK_SUCCESS)
2164 return result;
2165
2166 struct anv_batch batch = {
2167 .start = device->trivial_batch_bo->map,
2168 .next = device->trivial_batch_bo->map,
2169 .end = device->trivial_batch_bo->map + 4096,
2170 };
2171
2172 anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
2173 anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
2174
2175 if (!device->info.has_llc)
2176 gen_clflush_range(batch.start, batch.next - batch.start);
2177
2178 return VK_SUCCESS;
2179 }
2180
2181 VkResult anv_EnumerateDeviceExtensionProperties(
2182 VkPhysicalDevice physicalDevice,
2183 const char* pLayerName,
2184 uint32_t* pPropertyCount,
2185 VkExtensionProperties* pProperties)
2186 {
2187 ANV_FROM_HANDLE(anv_physical_device, device, physicalDevice);
2188 VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
2189
2190 for (int i = 0; i < ANV_DEVICE_EXTENSION_COUNT; i++) {
2191 if (device->supported_extensions.extensions[i]) {
2192 vk_outarray_append(&out, prop) {
2193 *prop = anv_device_extensions[i];
2194 }
2195 }
2196 }
2197
2198 return vk_outarray_status(&out);
2199 }
2200
2201 static void
2202 anv_device_init_dispatch(struct anv_device *device)
2203 {
2204 const struct anv_device_dispatch_table *genX_table;
2205 switch (device->info.gen) {
2206 case 12:
2207 genX_table = &gen12_device_dispatch_table;
2208 break;
2209 case 11:
2210 genX_table = &gen11_device_dispatch_table;
2211 break;
2212 case 10:
2213 genX_table = &gen10_device_dispatch_table;
2214 break;
2215 case 9:
2216 genX_table = &gen9_device_dispatch_table;
2217 break;
2218 case 8:
2219 genX_table = &gen8_device_dispatch_table;
2220 break;
2221 case 7:
2222 if (device->info.is_haswell)
2223 genX_table = &gen75_device_dispatch_table;
2224 else
2225 genX_table = &gen7_device_dispatch_table;
2226 break;
2227 default:
2228 unreachable("unsupported gen\n");
2229 }
2230
2231 for (unsigned i = 0; i < ARRAY_SIZE(device->dispatch.entrypoints); i++) {
2232 /* Vulkan requires that entrypoints for extensions which have not been
2233 * enabled must not be advertised.
2234 */
2235 if (!anv_device_entrypoint_is_enabled(i, device->instance->app_info.api_version,
2236 &device->instance->enabled_extensions,
2237 &device->enabled_extensions)) {
2238 device->dispatch.entrypoints[i] = NULL;
2239 } else if (genX_table->entrypoints[i]) {
2240 device->dispatch.entrypoints[i] = genX_table->entrypoints[i];
2241 } else {
2242 device->dispatch.entrypoints[i] =
2243 anv_device_dispatch_table.entrypoints[i];
2244 }
2245 }
2246 }
2247
2248 static int
2249 vk_priority_to_gen(int priority)
2250 {
2251 switch (priority) {
2252 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT:
2253 return GEN_CONTEXT_LOW_PRIORITY;
2254 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT:
2255 return GEN_CONTEXT_MEDIUM_PRIORITY;
2256 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT:
2257 return GEN_CONTEXT_HIGH_PRIORITY;
2258 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT:
2259 return GEN_CONTEXT_REALTIME_PRIORITY;
2260 default:
2261 unreachable("Invalid priority");
2262 }
2263 }
2264
2265 static VkResult
2266 anv_device_init_hiz_clear_value_bo(struct anv_device *device)
2267 {
2268 VkResult result = anv_device_alloc_bo(device, 4096,
2269 ANV_BO_ALLOC_MAPPED,
2270 0 /* explicit_address */,
2271 &device->hiz_clear_bo);
2272 if (result != VK_SUCCESS)
2273 return result;
2274
2275 union isl_color_value hiz_clear = { .u32 = { 0, } };
2276 hiz_clear.f32[0] = ANV_HZ_FC_VAL;
2277
2278 memcpy(device->hiz_clear_bo->map, hiz_clear.u32, sizeof(hiz_clear.u32));
2279
2280 if (!device->info.has_llc)
2281 gen_clflush_range(device->hiz_clear_bo->map, sizeof(hiz_clear.u32));
2282
2283 return VK_SUCCESS;
2284 }
2285
2286 static bool
2287 get_bo_from_pool(struct gen_batch_decode_bo *ret,
2288 struct anv_block_pool *pool,
2289 uint64_t address)
2290 {
2291 anv_block_pool_foreach_bo(bo, pool) {
2292 uint64_t bo_address = gen_48b_address(bo->offset);
2293 if (address >= bo_address && address < (bo_address + bo->size)) {
2294 *ret = (struct gen_batch_decode_bo) {
2295 .addr = bo_address,
2296 .size = bo->size,
2297 .map = bo->map,
2298 };
2299 return true;
2300 }
2301 }
2302 return false;
2303 }
2304
2305 /* Finding a buffer for batch decoding */
2306 static struct gen_batch_decode_bo
2307 decode_get_bo(void *v_batch, bool ppgtt, uint64_t address)
2308 {
2309 struct anv_device *device = v_batch;
2310 struct gen_batch_decode_bo ret_bo = {};
2311
2312 assert(ppgtt);
2313
2314 if (get_bo_from_pool(&ret_bo, &device->dynamic_state_pool.block_pool, address))
2315 return ret_bo;
2316 if (get_bo_from_pool(&ret_bo, &device->instruction_state_pool.block_pool, address))
2317 return ret_bo;
2318 if (get_bo_from_pool(&ret_bo, &device->binding_table_pool.block_pool, address))
2319 return ret_bo;
2320 if (get_bo_from_pool(&ret_bo, &device->surface_state_pool.block_pool, address))
2321 return ret_bo;
2322
2323 if (!device->cmd_buffer_being_decoded)
2324 return (struct gen_batch_decode_bo) { };
2325
2326 struct anv_batch_bo **bo;
2327
2328 u_vector_foreach(bo, &device->cmd_buffer_being_decoded->seen_bbos) {
2329 /* The decoder zeroes out the top 16 bits, so we need to as well */
2330 uint64_t bo_address = (*bo)->bo->offset & (~0ull >> 16);
2331
2332 if (address >= bo_address && address < bo_address + (*bo)->bo->size) {
2333 return (struct gen_batch_decode_bo) {
2334 .addr = bo_address,
2335 .size = (*bo)->bo->size,
2336 .map = (*bo)->bo->map,
2337 };
2338 }
2339 }
2340
2341 return (struct gen_batch_decode_bo) { };
2342 }
2343
2344 struct gen_aux_map_buffer {
2345 struct gen_buffer base;
2346 struct anv_state state;
2347 };
2348
2349 static struct gen_buffer *
2350 gen_aux_map_buffer_alloc(void *driver_ctx, uint32_t size)
2351 {
2352 struct gen_aux_map_buffer *buf = malloc(sizeof(struct gen_aux_map_buffer));
2353 if (!buf)
2354 return NULL;
2355
2356 struct anv_device *device = (struct anv_device*)driver_ctx;
2357 assert(device->instance->physicalDevice.supports_48bit_addresses &&
2358 device->instance->physicalDevice.use_softpin);
2359
2360 struct anv_state_pool *pool = &device->dynamic_state_pool;
2361 buf->state = anv_state_pool_alloc(pool, size, size);
2362
2363 buf->base.gpu = pool->block_pool.bo->offset + buf->state.offset;
2364 buf->base.gpu_end = buf->base.gpu + buf->state.alloc_size;
2365 buf->base.map = buf->state.map;
2366 buf->base.driver_bo = &buf->state;
2367 return &buf->base;
2368 }
2369
2370 static void
2371 gen_aux_map_buffer_free(void *driver_ctx, struct gen_buffer *buffer)
2372 {
2373 struct gen_aux_map_buffer *buf = (struct gen_aux_map_buffer*)buffer;
2374 struct anv_device *device = (struct anv_device*)driver_ctx;
2375 struct anv_state_pool *pool = &device->dynamic_state_pool;
2376 anv_state_pool_free(pool, buf->state);
2377 free(buf);
2378 }
2379
2380 static struct gen_mapped_pinned_buffer_alloc aux_map_allocator = {
2381 .alloc = gen_aux_map_buffer_alloc,
2382 .free = gen_aux_map_buffer_free,
2383 };
2384
2385 VkResult anv_CreateDevice(
2386 VkPhysicalDevice physicalDevice,
2387 const VkDeviceCreateInfo* pCreateInfo,
2388 const VkAllocationCallbacks* pAllocator,
2389 VkDevice* pDevice)
2390 {
2391 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
2392 VkResult result;
2393 struct anv_device *device;
2394
2395 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO);
2396
2397 struct anv_device_extension_table enabled_extensions = { };
2398 for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
2399 int idx;
2400 for (idx = 0; idx < ANV_DEVICE_EXTENSION_COUNT; idx++) {
2401 if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
2402 anv_device_extensions[idx].extensionName) == 0)
2403 break;
2404 }
2405
2406 if (idx >= ANV_DEVICE_EXTENSION_COUNT)
2407 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
2408
2409 if (!physical_device->supported_extensions.extensions[idx])
2410 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
2411
2412 enabled_extensions.extensions[idx] = true;
2413 }
2414
2415 /* Check enabled features */
2416 if (pCreateInfo->pEnabledFeatures) {
2417 VkPhysicalDeviceFeatures supported_features;
2418 anv_GetPhysicalDeviceFeatures(physicalDevice, &supported_features);
2419 VkBool32 *supported_feature = (VkBool32 *)&supported_features;
2420 VkBool32 *enabled_feature = (VkBool32 *)pCreateInfo->pEnabledFeatures;
2421 unsigned num_features = sizeof(VkPhysicalDeviceFeatures) / sizeof(VkBool32);
2422 for (uint32_t i = 0; i < num_features; i++) {
2423 if (enabled_feature[i] && !supported_feature[i])
2424 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
2425 }
2426 }
2427
2428 /* Check requested queues and fail if we are requested to create any
2429 * queues with flags we don't support.
2430 */
2431 assert(pCreateInfo->queueCreateInfoCount > 0);
2432 for (uint32_t i = 0; i < pCreateInfo->queueCreateInfoCount; i++) {
2433 if (pCreateInfo->pQueueCreateInfos[i].flags != 0)
2434 return vk_error(VK_ERROR_INITIALIZATION_FAILED);
2435 }
2436
2437 /* Check if client specified queue priority. */
2438 const VkDeviceQueueGlobalPriorityCreateInfoEXT *queue_priority =
2439 vk_find_struct_const(pCreateInfo->pQueueCreateInfos[0].pNext,
2440 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT);
2441
2442 VkQueueGlobalPriorityEXT priority =
2443 queue_priority ? queue_priority->globalPriority :
2444 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT;
2445
2446 device = vk_alloc2(&physical_device->instance->alloc, pAllocator,
2447 sizeof(*device), 8,
2448 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
2449 if (!device)
2450 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2451
2452 if (INTEL_DEBUG & DEBUG_BATCH) {
2453 const unsigned decode_flags =
2454 GEN_BATCH_DECODE_FULL |
2455 ((INTEL_DEBUG & DEBUG_COLOR) ? GEN_BATCH_DECODE_IN_COLOR : 0) |
2456 GEN_BATCH_DECODE_OFFSETS |
2457 GEN_BATCH_DECODE_FLOATS;
2458
2459 gen_batch_decode_ctx_init(&device->decoder_ctx,
2460 &physical_device->info,
2461 stderr, decode_flags, NULL,
2462 decode_get_bo, NULL, device);
2463 }
2464
2465 device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
2466 device->instance = physical_device->instance;
2467 device->chipset_id = physical_device->chipset_id;
2468 device->no_hw = physical_device->no_hw;
2469 device->_lost = false;
2470
2471 if (pAllocator)
2472 device->alloc = *pAllocator;
2473 else
2474 device->alloc = physical_device->instance->alloc;
2475
2476 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
2477 device->fd = open(physical_device->path, O_RDWR | O_CLOEXEC);
2478 if (device->fd == -1) {
2479 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2480 goto fail_device;
2481 }
2482
2483 device->context_id = anv_gem_create_context(device);
2484 if (device->context_id == -1) {
2485 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2486 goto fail_fd;
2487 }
2488
2489 result = anv_queue_init(device, &device->queue);
2490 if (result != VK_SUCCESS)
2491 goto fail_context_id;
2492
2493 if (physical_device->use_softpin) {
2494 if (pthread_mutex_init(&device->vma_mutex, NULL) != 0) {
2495 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2496 goto fail_queue;
2497 }
2498
2499 /* keep the page with address zero out of the allocator */
2500 util_vma_heap_init(&device->vma_lo,
2501 LOW_HEAP_MIN_ADDRESS, LOW_HEAP_SIZE);
2502
2503 util_vma_heap_init(&device->vma_cva, CLIENT_VISIBLE_HEAP_MIN_ADDRESS,
2504 CLIENT_VISIBLE_HEAP_SIZE);
2505
2506 /* Leave the last 4GiB out of the high vma range, so that no state
2507 * base address + size can overflow 48 bits. For more information see
2508 * the comment about Wa32bitGeneralStateOffset in anv_allocator.c
2509 */
2510 util_vma_heap_init(&device->vma_hi, HIGH_HEAP_MIN_ADDRESS,
2511 physical_device->gtt_size - (1ull << 32) -
2512 HIGH_HEAP_MIN_ADDRESS);
2513 }
2514
2515 list_inithead(&device->memory_objects);
2516
2517 /* As per spec, the driver implementation may deny requests to acquire
2518 * a priority above the default priority (MEDIUM) if the caller does not
2519 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
2520 * is returned.
2521 */
2522 if (physical_device->has_context_priority) {
2523 int err = anv_gem_set_context_param(device->fd, device->context_id,
2524 I915_CONTEXT_PARAM_PRIORITY,
2525 vk_priority_to_gen(priority));
2526 if (err != 0 && priority > VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT) {
2527 result = vk_error(VK_ERROR_NOT_PERMITTED_EXT);
2528 goto fail_vmas;
2529 }
2530 }
2531
2532 device->info = physical_device->info;
2533 device->isl_dev = physical_device->isl_dev;
2534
2535 /* On Broadwell and later, we can use batch chaining to more efficiently
2536 * implement growing command buffers. Prior to Haswell, the kernel
2537 * command parser gets in the way and we have to fall back to growing
2538 * the batch.
2539 */
2540 device->can_chain_batches = device->info.gen >= 8;
2541
2542 device->robust_buffer_access = pCreateInfo->pEnabledFeatures &&
2543 pCreateInfo->pEnabledFeatures->robustBufferAccess;
2544 device->enabled_extensions = enabled_extensions;
2545
2546 anv_device_init_dispatch(device);
2547
2548 if (pthread_mutex_init(&device->mutex, NULL) != 0) {
2549 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2550 goto fail_queue;
2551 }
2552
2553 pthread_condattr_t condattr;
2554 if (pthread_condattr_init(&condattr) != 0) {
2555 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2556 goto fail_mutex;
2557 }
2558 if (pthread_condattr_setclock(&condattr, CLOCK_MONOTONIC) != 0) {
2559 pthread_condattr_destroy(&condattr);
2560 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2561 goto fail_mutex;
2562 }
2563 if (pthread_cond_init(&device->queue_submit, &condattr) != 0) {
2564 pthread_condattr_destroy(&condattr);
2565 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2566 goto fail_mutex;
2567 }
2568 pthread_condattr_destroy(&condattr);
2569
2570 result = anv_bo_cache_init(&device->bo_cache);
2571 if (result != VK_SUCCESS)
2572 goto fail_queue_cond;
2573
2574 anv_bo_pool_init(&device->batch_bo_pool, device);
2575
2576 result = anv_state_pool_init(&device->dynamic_state_pool, device,
2577 DYNAMIC_STATE_POOL_MIN_ADDRESS, 16384);
2578 if (result != VK_SUCCESS)
2579 goto fail_batch_bo_pool;
2580
2581 result = anv_state_pool_init(&device->instruction_state_pool, device,
2582 INSTRUCTION_STATE_POOL_MIN_ADDRESS, 16384);
2583 if (result != VK_SUCCESS)
2584 goto fail_dynamic_state_pool;
2585
2586 result = anv_state_pool_init(&device->surface_state_pool, device,
2587 SURFACE_STATE_POOL_MIN_ADDRESS, 4096);
2588 if (result != VK_SUCCESS)
2589 goto fail_instruction_state_pool;
2590
2591 if (physical_device->use_softpin) {
2592 result = anv_state_pool_init(&device->binding_table_pool, device,
2593 BINDING_TABLE_POOL_MIN_ADDRESS, 4096);
2594 if (result != VK_SUCCESS)
2595 goto fail_surface_state_pool;
2596 }
2597
2598 if (device->info.gen >= 12) {
2599 device->aux_map_ctx = gen_aux_map_init(device, &aux_map_allocator,
2600 &physical_device->info);
2601 if (!device->aux_map_ctx)
2602 goto fail_binding_table_pool;
2603 }
2604
2605 result = anv_device_alloc_bo(device, 4096, 0 /* flags */,
2606 0 /* explicit_address */,
2607 &device->workaround_bo);
2608 if (result != VK_SUCCESS)
2609 goto fail_surface_aux_map_pool;
2610
2611 result = anv_device_init_trivial_batch(device);
2612 if (result != VK_SUCCESS)
2613 goto fail_workaround_bo;
2614
2615 if (device->info.gen >= 10) {
2616 result = anv_device_init_hiz_clear_value_bo(device);
2617 if (result != VK_SUCCESS)
2618 goto fail_trivial_batch_bo;
2619 }
2620
2621 anv_scratch_pool_init(device, &device->scratch_pool);
2622
2623 switch (device->info.gen) {
2624 case 7:
2625 if (!device->info.is_haswell)
2626 result = gen7_init_device_state(device);
2627 else
2628 result = gen75_init_device_state(device);
2629 break;
2630 case 8:
2631 result = gen8_init_device_state(device);
2632 break;
2633 case 9:
2634 result = gen9_init_device_state(device);
2635 break;
2636 case 10:
2637 result = gen10_init_device_state(device);
2638 break;
2639 case 11:
2640 result = gen11_init_device_state(device);
2641 break;
2642 case 12:
2643 result = gen12_init_device_state(device);
2644 break;
2645 default:
2646 /* Shouldn't get here as we don't create physical devices for any other
2647 * gens. */
2648 unreachable("unhandled gen");
2649 }
2650 if (result != VK_SUCCESS)
2651 goto fail_workaround_bo;
2652
2653 anv_pipeline_cache_init(&device->default_pipeline_cache, device, true);
2654
2655 anv_device_init_blorp(device);
2656
2657 anv_device_init_border_colors(device);
2658
2659 anv_device_perf_init(device);
2660
2661 *pDevice = anv_device_to_handle(device);
2662
2663 return VK_SUCCESS;
2664
2665 fail_workaround_bo:
2666 anv_scratch_pool_finish(device, &device->scratch_pool);
2667 if (device->info.gen >= 10)
2668 anv_device_release_bo(device, device->hiz_clear_bo);
2669 anv_device_release_bo(device, device->workaround_bo);
2670 fail_trivial_batch_bo:
2671 anv_device_release_bo(device, device->trivial_batch_bo);
2672 fail_surface_aux_map_pool:
2673 if (device->info.gen >= 12) {
2674 gen_aux_map_finish(device->aux_map_ctx);
2675 device->aux_map_ctx = NULL;
2676 }
2677 fail_binding_table_pool:
2678 if (physical_device->use_softpin)
2679 anv_state_pool_finish(&device->binding_table_pool);
2680 fail_surface_state_pool:
2681 anv_state_pool_finish(&device->surface_state_pool);
2682 fail_instruction_state_pool:
2683 anv_state_pool_finish(&device->instruction_state_pool);
2684 fail_dynamic_state_pool:
2685 anv_state_pool_finish(&device->dynamic_state_pool);
2686 fail_batch_bo_pool:
2687 anv_bo_pool_finish(&device->batch_bo_pool);
2688 anv_bo_cache_finish(&device->bo_cache);
2689 fail_queue_cond:
2690 pthread_cond_destroy(&device->queue_submit);
2691 fail_mutex:
2692 pthread_mutex_destroy(&device->mutex);
2693 fail_vmas:
2694 if (physical_device->use_softpin) {
2695 util_vma_heap_finish(&device->vma_hi);
2696 util_vma_heap_finish(&device->vma_cva);
2697 util_vma_heap_finish(&device->vma_lo);
2698 }
2699 fail_queue:
2700 anv_queue_finish(&device->queue);
2701 fail_context_id:
2702 anv_gem_destroy_context(device, device->context_id);
2703 fail_fd:
2704 close(device->fd);
2705 fail_device:
2706 vk_free(&device->alloc, device);
2707
2708 return result;
2709 }
2710
2711 void anv_DestroyDevice(
2712 VkDevice _device,
2713 const VkAllocationCallbacks* pAllocator)
2714 {
2715 ANV_FROM_HANDLE(anv_device, device, _device);
2716 struct anv_physical_device *physical_device;
2717
2718 if (!device)
2719 return;
2720
2721 physical_device = &device->instance->physicalDevice;
2722
2723 anv_device_finish_blorp(device);
2724
2725 anv_pipeline_cache_finish(&device->default_pipeline_cache);
2726
2727 anv_queue_finish(&device->queue);
2728
2729 #ifdef HAVE_VALGRIND
2730 /* We only need to free these to prevent valgrind errors. The backing
2731 * BO will go away in a couple of lines so we don't actually leak.
2732 */
2733 anv_state_pool_free(&device->dynamic_state_pool, device->border_colors);
2734 anv_state_pool_free(&device->dynamic_state_pool, device->slice_hash);
2735 #endif
2736
2737 anv_scratch_pool_finish(device, &device->scratch_pool);
2738
2739 anv_device_release_bo(device, device->workaround_bo);
2740 anv_device_release_bo(device, device->trivial_batch_bo);
2741 if (device->info.gen >= 10)
2742 anv_device_release_bo(device, device->hiz_clear_bo);
2743
2744 if (device->info.gen >= 12) {
2745 gen_aux_map_finish(device->aux_map_ctx);
2746 device->aux_map_ctx = NULL;
2747 }
2748
2749 if (physical_device->use_softpin)
2750 anv_state_pool_finish(&device->binding_table_pool);
2751 anv_state_pool_finish(&device->surface_state_pool);
2752 anv_state_pool_finish(&device->instruction_state_pool);
2753 anv_state_pool_finish(&device->dynamic_state_pool);
2754
2755 anv_bo_pool_finish(&device->batch_bo_pool);
2756
2757 anv_bo_cache_finish(&device->bo_cache);
2758
2759 if (physical_device->use_softpin) {
2760 util_vma_heap_finish(&device->vma_hi);
2761 util_vma_heap_finish(&device->vma_cva);
2762 util_vma_heap_finish(&device->vma_lo);
2763 }
2764
2765 pthread_cond_destroy(&device->queue_submit);
2766 pthread_mutex_destroy(&device->mutex);
2767
2768 anv_gem_destroy_context(device, device->context_id);
2769
2770 if (INTEL_DEBUG & DEBUG_BATCH)
2771 gen_batch_decode_ctx_finish(&device->decoder_ctx);
2772
2773 close(device->fd);
2774
2775 vk_free(&device->alloc, device);
2776 }
2777
2778 VkResult anv_EnumerateInstanceLayerProperties(
2779 uint32_t* pPropertyCount,
2780 VkLayerProperties* pProperties)
2781 {
2782 if (pProperties == NULL) {
2783 *pPropertyCount = 0;
2784 return VK_SUCCESS;
2785 }
2786
2787 /* None supported at this time */
2788 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
2789 }
2790
2791 VkResult anv_EnumerateDeviceLayerProperties(
2792 VkPhysicalDevice physicalDevice,
2793 uint32_t* pPropertyCount,
2794 VkLayerProperties* pProperties)
2795 {
2796 if (pProperties == NULL) {
2797 *pPropertyCount = 0;
2798 return VK_SUCCESS;
2799 }
2800
2801 /* None supported at this time */
2802 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
2803 }
2804
2805 void anv_GetDeviceQueue(
2806 VkDevice _device,
2807 uint32_t queueNodeIndex,
2808 uint32_t queueIndex,
2809 VkQueue* pQueue)
2810 {
2811 const VkDeviceQueueInfo2 info = {
2812 .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_INFO_2,
2813 .pNext = NULL,
2814 .flags = 0,
2815 .queueFamilyIndex = queueNodeIndex,
2816 .queueIndex = queueIndex,
2817 };
2818
2819 anv_GetDeviceQueue2(_device, &info, pQueue);
2820 }
2821
2822 void anv_GetDeviceQueue2(
2823 VkDevice _device,
2824 const VkDeviceQueueInfo2* pQueueInfo,
2825 VkQueue* pQueue)
2826 {
2827 ANV_FROM_HANDLE(anv_device, device, _device);
2828
2829 assert(pQueueInfo->queueIndex == 0);
2830
2831 if (pQueueInfo->flags == device->queue.flags)
2832 *pQueue = anv_queue_to_handle(&device->queue);
2833 else
2834 *pQueue = NULL;
2835 }
2836
2837 VkResult
2838 _anv_device_set_lost(struct anv_device *device,
2839 const char *file, int line,
2840 const char *msg, ...)
2841 {
2842 VkResult err;
2843 va_list ap;
2844
2845 p_atomic_inc(&device->_lost);
2846
2847 va_start(ap, msg);
2848 err = __vk_errorv(device->instance, device,
2849 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT,
2850 VK_ERROR_DEVICE_LOST, file, line, msg, ap);
2851 va_end(ap);
2852
2853 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
2854 abort();
2855
2856 return err;
2857 }
2858
2859 VkResult
2860 _anv_queue_set_lost(struct anv_queue *queue,
2861 const char *file, int line,
2862 const char *msg, ...)
2863 {
2864 VkResult err;
2865 va_list ap;
2866
2867 p_atomic_inc(&queue->device->_lost);
2868
2869 va_start(ap, msg);
2870 err = __vk_errorv(queue->device->instance, queue->device,
2871 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT,
2872 VK_ERROR_DEVICE_LOST, file, line, msg, ap);
2873 va_end(ap);
2874
2875 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
2876 abort();
2877
2878 return err;
2879 }
2880
2881 VkResult
2882 anv_device_query_status(struct anv_device *device)
2883 {
2884 /* This isn't likely as most of the callers of this function already check
2885 * for it. However, it doesn't hurt to check and it potentially lets us
2886 * avoid an ioctl.
2887 */
2888 if (anv_device_is_lost(device))
2889 return VK_ERROR_DEVICE_LOST;
2890
2891 uint32_t active, pending;
2892 int ret = anv_gem_gpu_get_reset_stats(device, &active, &pending);
2893 if (ret == -1) {
2894 /* We don't know the real error. */
2895 return anv_device_set_lost(device, "get_reset_stats failed: %m");
2896 }
2897
2898 if (active) {
2899 return anv_device_set_lost(device, "GPU hung on one of our command buffers");
2900 } else if (pending) {
2901 return anv_device_set_lost(device, "GPU hung with commands in-flight");
2902 }
2903
2904 return VK_SUCCESS;
2905 }
2906
2907 VkResult
2908 anv_device_bo_busy(struct anv_device *device, struct anv_bo *bo)
2909 {
2910 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
2911 * Other usages of the BO (such as on different hardware) will not be
2912 * flagged as "busy" by this ioctl. Use with care.
2913 */
2914 int ret = anv_gem_busy(device, bo->gem_handle);
2915 if (ret == 1) {
2916 return VK_NOT_READY;
2917 } else if (ret == -1) {
2918 /* We don't know the real error. */
2919 return anv_device_set_lost(device, "gem wait failed: %m");
2920 }
2921
2922 /* Query for device status after the busy call. If the BO we're checking
2923 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
2924 * client because it clearly doesn't have valid data. Yes, this most
2925 * likely means an ioctl, but we just did an ioctl to query the busy status
2926 * so it's no great loss.
2927 */
2928 return anv_device_query_status(device);
2929 }
2930
2931 VkResult
2932 anv_device_wait(struct anv_device *device, struct anv_bo *bo,
2933 int64_t timeout)
2934 {
2935 int ret = anv_gem_wait(device, bo->gem_handle, &timeout);
2936 if (ret == -1 && errno == ETIME) {
2937 return VK_TIMEOUT;
2938 } else if (ret == -1) {
2939 /* We don't know the real error. */
2940 return anv_device_set_lost(device, "gem wait failed: %m");
2941 }
2942
2943 /* Query for device status after the wait. If the BO we're waiting on got
2944 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
2945 * because it clearly doesn't have valid data. Yes, this most likely means
2946 * an ioctl, but we just did an ioctl to wait so it's no great loss.
2947 */
2948 return anv_device_query_status(device);
2949 }
2950
2951 VkResult anv_DeviceWaitIdle(
2952 VkDevice _device)
2953 {
2954 ANV_FROM_HANDLE(anv_device, device, _device);
2955
2956 if (anv_device_is_lost(device))
2957 return VK_ERROR_DEVICE_LOST;
2958
2959 return anv_queue_submit_simple_batch(&device->queue, NULL);
2960 }
2961
2962 bool
2963 anv_vma_alloc(struct anv_device *device, struct anv_bo *bo,
2964 uint64_t client_address)
2965 {
2966 if (!(bo->flags & EXEC_OBJECT_PINNED)) {
2967 assert(!(bo->has_client_visible_address));
2968 return true;
2969 }
2970
2971 pthread_mutex_lock(&device->vma_mutex);
2972
2973 bo->offset = 0;
2974
2975 if (bo->has_client_visible_address) {
2976 assert(bo->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS);
2977 if (client_address) {
2978 if (util_vma_heap_alloc_addr(&device->vma_cva,
2979 client_address, bo->size)) {
2980 bo->offset = gen_canonical_address(client_address);
2981 }
2982 } else {
2983 uint64_t addr = util_vma_heap_alloc(&device->vma_cva, bo->size, 4096);
2984 if (addr) {
2985 bo->offset = gen_canonical_address(addr);
2986 assert(addr == gen_48b_address(bo->offset));
2987 }
2988 }
2989 /* We don't want to fall back to other heaps */
2990 goto done;
2991 }
2992
2993 assert(client_address == 0);
2994
2995 if (bo->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) {
2996 uint64_t addr = util_vma_heap_alloc(&device->vma_hi, bo->size, 4096);
2997 if (addr) {
2998 bo->offset = gen_canonical_address(addr);
2999 assert(addr == gen_48b_address(bo->offset));
3000 }
3001 }
3002
3003 if (bo->offset == 0) {
3004 uint64_t addr = util_vma_heap_alloc(&device->vma_lo, bo->size, 4096);
3005 if (addr) {
3006 bo->offset = gen_canonical_address(addr);
3007 assert(addr == gen_48b_address(bo->offset));
3008 }
3009 }
3010
3011 done:
3012 pthread_mutex_unlock(&device->vma_mutex);
3013
3014 return bo->offset != 0;
3015 }
3016
3017 void
3018 anv_vma_free(struct anv_device *device, struct anv_bo *bo)
3019 {
3020 if (!(bo->flags & EXEC_OBJECT_PINNED))
3021 return;
3022
3023 const uint64_t addr_48b = gen_48b_address(bo->offset);
3024
3025 pthread_mutex_lock(&device->vma_mutex);
3026
3027 if (addr_48b >= LOW_HEAP_MIN_ADDRESS &&
3028 addr_48b <= LOW_HEAP_MAX_ADDRESS) {
3029 util_vma_heap_free(&device->vma_lo, addr_48b, bo->size);
3030 } else if (addr_48b >= CLIENT_VISIBLE_HEAP_MIN_ADDRESS &&
3031 addr_48b <= CLIENT_VISIBLE_HEAP_MAX_ADDRESS) {
3032 util_vma_heap_free(&device->vma_cva, addr_48b, bo->size);
3033 } else {
3034 assert(addr_48b >= HIGH_HEAP_MIN_ADDRESS);
3035 util_vma_heap_free(&device->vma_hi, addr_48b, bo->size);
3036 }
3037
3038 pthread_mutex_unlock(&device->vma_mutex);
3039
3040 bo->offset = 0;
3041 }
3042
3043 VkResult anv_AllocateMemory(
3044 VkDevice _device,
3045 const VkMemoryAllocateInfo* pAllocateInfo,
3046 const VkAllocationCallbacks* pAllocator,
3047 VkDeviceMemory* pMem)
3048 {
3049 ANV_FROM_HANDLE(anv_device, device, _device);
3050 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3051 struct anv_device_memory *mem;
3052 VkResult result = VK_SUCCESS;
3053
3054 assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
3055
3056 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
3057 assert(pAllocateInfo->allocationSize > 0);
3058
3059 VkDeviceSize aligned_alloc_size =
3060 align_u64(pAllocateInfo->allocationSize, 4096);
3061
3062 if (aligned_alloc_size > MAX_MEMORY_ALLOCATION_SIZE)
3063 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
3064
3065 assert(pAllocateInfo->memoryTypeIndex < pdevice->memory.type_count);
3066 struct anv_memory_type *mem_type =
3067 &pdevice->memory.types[pAllocateInfo->memoryTypeIndex];
3068 assert(mem_type->heapIndex < pdevice->memory.heap_count);
3069 struct anv_memory_heap *mem_heap =
3070 &pdevice->memory.heaps[mem_type->heapIndex];
3071
3072 uint64_t mem_heap_used = p_atomic_read(&mem_heap->used);
3073 if (mem_heap_used + aligned_alloc_size > mem_heap->size)
3074 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
3075
3076 mem = vk_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
3077 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
3078 if (mem == NULL)
3079 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
3080
3081 assert(pAllocateInfo->memoryTypeIndex < pdevice->memory.type_count);
3082 mem->type = mem_type;
3083 mem->map = NULL;
3084 mem->map_size = 0;
3085 mem->ahw = NULL;
3086 mem->host_ptr = NULL;
3087
3088 enum anv_bo_alloc_flags alloc_flags = 0;
3089
3090 const VkExportMemoryAllocateInfo *export_info = NULL;
3091 const VkImportAndroidHardwareBufferInfoANDROID *ahw_import_info = NULL;
3092 const VkImportMemoryFdInfoKHR *fd_info = NULL;
3093 const VkImportMemoryHostPointerInfoEXT *host_ptr_info = NULL;
3094 const VkMemoryDedicatedAllocateInfo *dedicated_info = NULL;
3095
3096 vk_foreach_struct_const(ext, pAllocateInfo->pNext) {
3097 switch (ext->sType) {
3098 case VK_STRUCTURE_TYPE_EXPORT_MEMORY_ALLOCATE_INFO:
3099 export_info = (void *)ext;
3100 break;
3101
3102 case VK_STRUCTURE_TYPE_IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID:
3103 ahw_import_info = (void *)ext;
3104 break;
3105
3106 case VK_STRUCTURE_TYPE_IMPORT_MEMORY_FD_INFO_KHR:
3107 fd_info = (void *)ext;
3108 break;
3109
3110 case VK_STRUCTURE_TYPE_IMPORT_MEMORY_HOST_POINTER_INFO_EXT:
3111 host_ptr_info = (void *)ext;
3112 break;
3113
3114 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_ALLOCATE_INFO:
3115 dedicated_info = (void *)ext;
3116 break;
3117
3118 case VK_STRUCTURE_TYPE_WSI_MEMORY_ALLOCATE_INFO_MESA: {
3119 const struct wsi_memory_allocate_info *wsi_info = (void *)ext;
3120 if (wsi_info->implicit_sync) {
3121 /* We need to set the WRITE flag on window system buffers so that
3122 * GEM will know we're writing to them and synchronize uses on
3123 * other rings (eg if the display server uses the blitter ring).
3124 */
3125 alloc_flags |= ANV_BO_ALLOC_IMPLICIT_SYNC |
3126 ANV_BO_ALLOC_IMPLICIT_WRITE;
3127 }
3128 break;
3129 }
3130
3131 default:
3132 anv_debug_ignored_stype(ext->sType);
3133 break;
3134 }
3135 }
3136
3137 /* Check if we need to support Android HW buffer export. If so,
3138 * create AHardwareBuffer and import memory from it.
3139 */
3140 bool android_export = false;
3141 if (export_info && export_info->handleTypes &
3142 VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID)
3143 android_export = true;
3144
3145 if (ahw_import_info) {
3146 result = anv_import_ahw_memory(_device, mem, ahw_import_info);
3147 if (result != VK_SUCCESS)
3148 goto fail;
3149
3150 goto success;
3151 } else if (android_export) {
3152 result = anv_create_ahw_memory(_device, mem, pAllocateInfo);
3153 if (result != VK_SUCCESS)
3154 goto fail;
3155
3156 const struct VkImportAndroidHardwareBufferInfoANDROID import_info = {
3157 .buffer = mem->ahw,
3158 };
3159 result = anv_import_ahw_memory(_device, mem, &import_info);
3160 if (result != VK_SUCCESS)
3161 goto fail;
3162
3163 goto success;
3164 }
3165
3166 /* The Vulkan spec permits handleType to be 0, in which case the struct is
3167 * ignored.
3168 */
3169 if (fd_info && fd_info->handleType) {
3170 /* At the moment, we support only the below handle types. */
3171 assert(fd_info->handleType ==
3172 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
3173 fd_info->handleType ==
3174 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
3175
3176 result = anv_device_import_bo(device, fd_info->fd, alloc_flags,
3177 0 /* client_address */, &mem->bo);
3178 if (result != VK_SUCCESS)
3179 goto fail;
3180
3181 VkDeviceSize aligned_alloc_size =
3182 align_u64(pAllocateInfo->allocationSize, 4096);
3183
3184 /* For security purposes, we reject importing the bo if it's smaller
3185 * than the requested allocation size. This prevents a malicious client
3186 * from passing a buffer to a trusted client, lying about the size, and
3187 * telling the trusted client to try and texture from an image that goes
3188 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
3189 * in the trusted client. The trusted client can protect itself against
3190 * this sort of attack but only if it can trust the buffer size.
3191 */
3192 if (mem->bo->size < aligned_alloc_size) {
3193 result = vk_errorf(device->instance, device,
3194 VK_ERROR_INVALID_EXTERNAL_HANDLE,
3195 "aligned allocationSize too large for "
3196 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT: "
3197 "%"PRIu64"B > %"PRIu64"B",
3198 aligned_alloc_size, mem->bo->size);
3199 anv_device_release_bo(device, mem->bo);
3200 goto fail;
3201 }
3202
3203 /* From the Vulkan spec:
3204 *
3205 * "Importing memory from a file descriptor transfers ownership of
3206 * the file descriptor from the application to the Vulkan
3207 * implementation. The application must not perform any operations on
3208 * the file descriptor after a successful import."
3209 *
3210 * If the import fails, we leave the file descriptor open.
3211 */
3212 close(fd_info->fd);
3213 goto success;
3214 }
3215
3216 if (host_ptr_info && host_ptr_info->handleType) {
3217 if (host_ptr_info->handleType ==
3218 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_MAPPED_FOREIGN_MEMORY_BIT_EXT) {
3219 result = vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE);
3220 goto fail;
3221 }
3222
3223 assert(host_ptr_info->handleType ==
3224 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT);
3225
3226 result = anv_device_import_bo_from_host_ptr(device,
3227 host_ptr_info->pHostPointer,
3228 pAllocateInfo->allocationSize,
3229 alloc_flags,
3230 0 /* client_address */,
3231 &mem->bo);
3232
3233 if (result != VK_SUCCESS)
3234 goto fail;
3235
3236 mem->host_ptr = host_ptr_info->pHostPointer;
3237 goto success;
3238 }
3239
3240 /* Regular allocate (not importing memory). */
3241
3242 if (export_info && export_info->handleTypes)
3243 alloc_flags |= ANV_BO_ALLOC_EXTERNAL;
3244
3245 result = anv_device_alloc_bo(device, pAllocateInfo->allocationSize,
3246 alloc_flags, 0 /* explicit_address */,
3247 &mem->bo);
3248 if (result != VK_SUCCESS)
3249 goto fail;
3250
3251 if (dedicated_info && dedicated_info->image != VK_NULL_HANDLE) {
3252 ANV_FROM_HANDLE(anv_image, image, dedicated_info->image);
3253
3254 /* Some legacy (non-modifiers) consumers need the tiling to be set on
3255 * the BO. In this case, we have a dedicated allocation.
3256 */
3257 if (image->needs_set_tiling) {
3258 const uint32_t i915_tiling =
3259 isl_tiling_to_i915_tiling(image->planes[0].surface.isl.tiling);
3260 int ret = anv_gem_set_tiling(device, mem->bo->gem_handle,
3261 image->planes[0].surface.isl.row_pitch_B,
3262 i915_tiling);
3263 if (ret) {
3264 anv_device_release_bo(device, mem->bo);
3265 result = vk_errorf(device->instance, NULL,
3266 VK_ERROR_OUT_OF_DEVICE_MEMORY,
3267 "failed to set BO tiling: %m");
3268 goto fail;
3269 }
3270 }
3271 }
3272
3273 success:
3274 mem_heap_used = p_atomic_add_return(&mem_heap->used, mem->bo->size);
3275 if (mem_heap_used > mem_heap->size) {
3276 p_atomic_add(&mem_heap->used, -mem->bo->size);
3277 anv_device_release_bo(device, mem->bo);
3278 result = vk_errorf(device->instance, NULL,
3279 VK_ERROR_OUT_OF_DEVICE_MEMORY,
3280 "Out of heap memory");
3281 goto fail;
3282 }
3283
3284 pthread_mutex_lock(&device->mutex);
3285 list_addtail(&mem->link, &device->memory_objects);
3286 pthread_mutex_unlock(&device->mutex);
3287
3288 *pMem = anv_device_memory_to_handle(mem);
3289
3290 return VK_SUCCESS;
3291
3292 fail:
3293 vk_free2(&device->alloc, pAllocator, mem);
3294
3295 return result;
3296 }
3297
3298 VkResult anv_GetMemoryFdKHR(
3299 VkDevice device_h,
3300 const VkMemoryGetFdInfoKHR* pGetFdInfo,
3301 int* pFd)
3302 {
3303 ANV_FROM_HANDLE(anv_device, dev, device_h);
3304 ANV_FROM_HANDLE(anv_device_memory, mem, pGetFdInfo->memory);
3305
3306 assert(pGetFdInfo->sType == VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR);
3307
3308 assert(pGetFdInfo->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
3309 pGetFdInfo->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
3310
3311 return anv_device_export_bo(dev, mem->bo, pFd);
3312 }
3313
3314 VkResult anv_GetMemoryFdPropertiesKHR(
3315 VkDevice _device,
3316 VkExternalMemoryHandleTypeFlagBits handleType,
3317 int fd,
3318 VkMemoryFdPropertiesKHR* pMemoryFdProperties)
3319 {
3320 ANV_FROM_HANDLE(anv_device, device, _device);
3321 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3322
3323 switch (handleType) {
3324 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT:
3325 /* dma-buf can be imported as any memory type */
3326 pMemoryFdProperties->memoryTypeBits =
3327 (1 << pdevice->memory.type_count) - 1;
3328 return VK_SUCCESS;
3329
3330 default:
3331 /* The valid usage section for this function says:
3332 *
3333 * "handleType must not be one of the handle types defined as
3334 * opaque."
3335 *
3336 * So opaque handle types fall into the default "unsupported" case.
3337 */
3338 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE);
3339 }
3340 }
3341
3342 VkResult anv_GetMemoryHostPointerPropertiesEXT(
3343 VkDevice _device,
3344 VkExternalMemoryHandleTypeFlagBits handleType,
3345 const void* pHostPointer,
3346 VkMemoryHostPointerPropertiesEXT* pMemoryHostPointerProperties)
3347 {
3348 ANV_FROM_HANDLE(anv_device, device, _device);
3349
3350 assert(pMemoryHostPointerProperties->sType ==
3351 VK_STRUCTURE_TYPE_MEMORY_HOST_POINTER_PROPERTIES_EXT);
3352
3353 switch (handleType) {
3354 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT: {
3355 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3356
3357 /* Host memory can be imported as any memory type. */
3358 pMemoryHostPointerProperties->memoryTypeBits =
3359 (1ull << pdevice->memory.type_count) - 1;
3360
3361 return VK_SUCCESS;
3362 }
3363 default:
3364 return VK_ERROR_INVALID_EXTERNAL_HANDLE;
3365 }
3366 }
3367
3368 void anv_FreeMemory(
3369 VkDevice _device,
3370 VkDeviceMemory _mem,
3371 const VkAllocationCallbacks* pAllocator)
3372 {
3373 ANV_FROM_HANDLE(anv_device, device, _device);
3374 ANV_FROM_HANDLE(anv_device_memory, mem, _mem);
3375 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3376
3377 if (mem == NULL)
3378 return;
3379
3380 pthread_mutex_lock(&device->mutex);
3381 list_del(&mem->link);
3382 pthread_mutex_unlock(&device->mutex);
3383
3384 if (mem->map)
3385 anv_UnmapMemory(_device, _mem);
3386
3387 p_atomic_add(&pdevice->memory.heaps[mem->type->heapIndex].used,
3388 -mem->bo->size);
3389
3390 anv_device_release_bo(device, mem->bo);
3391
3392 #if defined(ANDROID) && ANDROID_API_LEVEL >= 26
3393 if (mem->ahw)
3394 AHardwareBuffer_release(mem->ahw);
3395 #endif
3396
3397 vk_free2(&device->alloc, pAllocator, mem);
3398 }
3399
3400 VkResult anv_MapMemory(
3401 VkDevice _device,
3402 VkDeviceMemory _memory,
3403 VkDeviceSize offset,
3404 VkDeviceSize size,
3405 VkMemoryMapFlags flags,
3406 void** ppData)
3407 {
3408 ANV_FROM_HANDLE(anv_device, device, _device);
3409 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
3410
3411 if (mem == NULL) {
3412 *ppData = NULL;
3413 return VK_SUCCESS;
3414 }
3415
3416 if (mem->host_ptr) {
3417 *ppData = mem->host_ptr + offset;
3418 return VK_SUCCESS;
3419 }
3420
3421 if (size == VK_WHOLE_SIZE)
3422 size = mem->bo->size - offset;
3423
3424 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
3425 *
3426 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
3427 * assert(size != 0);
3428 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
3429 * equal to the size of the memory minus offset
3430 */
3431 assert(size > 0);
3432 assert(offset + size <= mem->bo->size);
3433
3434 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
3435 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
3436 * at a time is valid. We could just mmap up front and return an offset
3437 * pointer here, but that may exhaust virtual memory on 32 bit
3438 * userspace. */
3439
3440 uint32_t gem_flags = 0;
3441
3442 if (!device->info.has_llc &&
3443 (mem->type->propertyFlags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT))
3444 gem_flags |= I915_MMAP_WC;
3445
3446 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
3447 uint64_t map_offset = offset & ~4095ull;
3448 assert(offset >= map_offset);
3449 uint64_t map_size = (offset + size) - map_offset;
3450
3451 /* Let's map whole pages */
3452 map_size = align_u64(map_size, 4096);
3453
3454 void *map = anv_gem_mmap(device, mem->bo->gem_handle,
3455 map_offset, map_size, gem_flags);
3456 if (map == MAP_FAILED)
3457 return vk_error(VK_ERROR_MEMORY_MAP_FAILED);
3458
3459 mem->map = map;
3460 mem->map_size = map_size;
3461
3462 *ppData = mem->map + (offset - map_offset);
3463
3464 return VK_SUCCESS;
3465 }
3466
3467 void anv_UnmapMemory(
3468 VkDevice _device,
3469 VkDeviceMemory _memory)
3470 {
3471 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
3472
3473 if (mem == NULL || mem->host_ptr)
3474 return;
3475
3476 anv_gem_munmap(mem->map, mem->map_size);
3477
3478 mem->map = NULL;
3479 mem->map_size = 0;
3480 }
3481
3482 static void
3483 clflush_mapped_ranges(struct anv_device *device,
3484 uint32_t count,
3485 const VkMappedMemoryRange *ranges)
3486 {
3487 for (uint32_t i = 0; i < count; i++) {
3488 ANV_FROM_HANDLE(anv_device_memory, mem, ranges[i].memory);
3489 if (ranges[i].offset >= mem->map_size)
3490 continue;
3491
3492 gen_clflush_range(mem->map + ranges[i].offset,
3493 MIN2(ranges[i].size, mem->map_size - ranges[i].offset));
3494 }
3495 }
3496
3497 VkResult anv_FlushMappedMemoryRanges(
3498 VkDevice _device,
3499 uint32_t memoryRangeCount,
3500 const VkMappedMemoryRange* pMemoryRanges)
3501 {
3502 ANV_FROM_HANDLE(anv_device, device, _device);
3503
3504 if (device->info.has_llc)
3505 return VK_SUCCESS;
3506
3507 /* Make sure the writes we're flushing have landed. */
3508 __builtin_ia32_mfence();
3509
3510 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
3511
3512 return VK_SUCCESS;
3513 }
3514
3515 VkResult anv_InvalidateMappedMemoryRanges(
3516 VkDevice _device,
3517 uint32_t memoryRangeCount,
3518 const VkMappedMemoryRange* pMemoryRanges)
3519 {
3520 ANV_FROM_HANDLE(anv_device, device, _device);
3521
3522 if (device->info.has_llc)
3523 return VK_SUCCESS;
3524
3525 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
3526
3527 /* Make sure no reads get moved up above the invalidate. */
3528 __builtin_ia32_mfence();
3529
3530 return VK_SUCCESS;
3531 }
3532
3533 void anv_GetBufferMemoryRequirements(
3534 VkDevice _device,
3535 VkBuffer _buffer,
3536 VkMemoryRequirements* pMemoryRequirements)
3537 {
3538 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
3539 ANV_FROM_HANDLE(anv_device, device, _device);
3540 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3541
3542 /* The Vulkan spec (git aaed022) says:
3543 *
3544 * memoryTypeBits is a bitfield and contains one bit set for every
3545 * supported memory type for the resource. The bit `1<<i` is set if and
3546 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3547 * structure for the physical device is supported.
3548 */
3549 uint32_t memory_types = (1ull << pdevice->memory.type_count) - 1;
3550
3551 /* Base alignment requirement of a cache line */
3552 uint32_t alignment = 16;
3553
3554 /* We need an alignment of 32 for pushing UBOs */
3555 if (buffer->usage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT)
3556 alignment = MAX2(alignment, 32);
3557
3558 pMemoryRequirements->size = buffer->size;
3559 pMemoryRequirements->alignment = alignment;
3560
3561 /* Storage and Uniform buffers should have their size aligned to
3562 * 32-bits to avoid boundary checks when last DWord is not complete.
3563 * This would ensure that not internal padding would be needed for
3564 * 16-bit types.
3565 */
3566 if (device->robust_buffer_access &&
3567 (buffer->usage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT ||
3568 buffer->usage & VK_BUFFER_USAGE_STORAGE_BUFFER_BIT))
3569 pMemoryRequirements->size = align_u64(buffer->size, 4);
3570
3571 pMemoryRequirements->memoryTypeBits = memory_types;
3572 }
3573
3574 void anv_GetBufferMemoryRequirements2(
3575 VkDevice _device,
3576 const VkBufferMemoryRequirementsInfo2* pInfo,
3577 VkMemoryRequirements2* pMemoryRequirements)
3578 {
3579 anv_GetBufferMemoryRequirements(_device, pInfo->buffer,
3580 &pMemoryRequirements->memoryRequirements);
3581
3582 vk_foreach_struct(ext, pMemoryRequirements->pNext) {
3583 switch (ext->sType) {
3584 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
3585 VkMemoryDedicatedRequirements *requirements = (void *)ext;
3586 requirements->prefersDedicatedAllocation = false;
3587 requirements->requiresDedicatedAllocation = false;
3588 break;
3589 }
3590
3591 default:
3592 anv_debug_ignored_stype(ext->sType);
3593 break;
3594 }
3595 }
3596 }
3597
3598 void anv_GetImageMemoryRequirements(
3599 VkDevice _device,
3600 VkImage _image,
3601 VkMemoryRequirements* pMemoryRequirements)
3602 {
3603 ANV_FROM_HANDLE(anv_image, image, _image);
3604 ANV_FROM_HANDLE(anv_device, device, _device);
3605 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3606
3607 /* The Vulkan spec (git aaed022) says:
3608 *
3609 * memoryTypeBits is a bitfield and contains one bit set for every
3610 * supported memory type for the resource. The bit `1<<i` is set if and
3611 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3612 * structure for the physical device is supported.
3613 *
3614 * All types are currently supported for images.
3615 */
3616 uint32_t memory_types = (1ull << pdevice->memory.type_count) - 1;
3617
3618 /* We must have image allocated or imported at this point. According to the
3619 * specification, external images must have been bound to memory before
3620 * calling GetImageMemoryRequirements.
3621 */
3622 assert(image->size > 0);
3623
3624 pMemoryRequirements->size = image->size;
3625 pMemoryRequirements->alignment = image->alignment;
3626 pMemoryRequirements->memoryTypeBits = memory_types;
3627 }
3628
3629 void anv_GetImageMemoryRequirements2(
3630 VkDevice _device,
3631 const VkImageMemoryRequirementsInfo2* pInfo,
3632 VkMemoryRequirements2* pMemoryRequirements)
3633 {
3634 ANV_FROM_HANDLE(anv_device, device, _device);
3635 ANV_FROM_HANDLE(anv_image, image, pInfo->image);
3636
3637 anv_GetImageMemoryRequirements(_device, pInfo->image,
3638 &pMemoryRequirements->memoryRequirements);
3639
3640 vk_foreach_struct_const(ext, pInfo->pNext) {
3641 switch (ext->sType) {
3642 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO: {
3643 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3644 const VkImagePlaneMemoryRequirementsInfo *plane_reqs =
3645 (const VkImagePlaneMemoryRequirementsInfo *) ext;
3646 uint32_t plane = anv_image_aspect_to_plane(image->aspects,
3647 plane_reqs->planeAspect);
3648
3649 assert(image->planes[plane].offset == 0);
3650
3651 /* The Vulkan spec (git aaed022) says:
3652 *
3653 * memoryTypeBits is a bitfield and contains one bit set for every
3654 * supported memory type for the resource. The bit `1<<i` is set
3655 * if and only if the memory type `i` in the
3656 * VkPhysicalDeviceMemoryProperties structure for the physical
3657 * device is supported.
3658 *
3659 * All types are currently supported for images.
3660 */
3661 pMemoryRequirements->memoryRequirements.memoryTypeBits =
3662 (1ull << pdevice->memory.type_count) - 1;
3663
3664 /* We must have image allocated or imported at this point. According to the
3665 * specification, external images must have been bound to memory before
3666 * calling GetImageMemoryRequirements.
3667 */
3668 assert(image->planes[plane].size > 0);
3669
3670 pMemoryRequirements->memoryRequirements.size = image->planes[plane].size;
3671 pMemoryRequirements->memoryRequirements.alignment =
3672 image->planes[plane].alignment;
3673 break;
3674 }
3675
3676 default:
3677 anv_debug_ignored_stype(ext->sType);
3678 break;
3679 }
3680 }
3681
3682 vk_foreach_struct(ext, pMemoryRequirements->pNext) {
3683 switch (ext->sType) {
3684 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
3685 VkMemoryDedicatedRequirements *requirements = (void *)ext;
3686 if (image->needs_set_tiling || image->external_format) {
3687 /* If we need to set the tiling for external consumers, we need a
3688 * dedicated allocation.
3689 *
3690 * See also anv_AllocateMemory.
3691 */
3692 requirements->prefersDedicatedAllocation = true;
3693 requirements->requiresDedicatedAllocation = true;
3694 } else {
3695 requirements->prefersDedicatedAllocation = false;
3696 requirements->requiresDedicatedAllocation = false;
3697 }
3698 break;
3699 }
3700
3701 default:
3702 anv_debug_ignored_stype(ext->sType);
3703 break;
3704 }
3705 }
3706 }
3707
3708 void anv_GetImageSparseMemoryRequirements(
3709 VkDevice device,
3710 VkImage image,
3711 uint32_t* pSparseMemoryRequirementCount,
3712 VkSparseImageMemoryRequirements* pSparseMemoryRequirements)
3713 {
3714 *pSparseMemoryRequirementCount = 0;
3715 }
3716
3717 void anv_GetImageSparseMemoryRequirements2(
3718 VkDevice device,
3719 const VkImageSparseMemoryRequirementsInfo2* pInfo,
3720 uint32_t* pSparseMemoryRequirementCount,
3721 VkSparseImageMemoryRequirements2* pSparseMemoryRequirements)
3722 {
3723 *pSparseMemoryRequirementCount = 0;
3724 }
3725
3726 void anv_GetDeviceMemoryCommitment(
3727 VkDevice device,
3728 VkDeviceMemory memory,
3729 VkDeviceSize* pCommittedMemoryInBytes)
3730 {
3731 *pCommittedMemoryInBytes = 0;
3732 }
3733
3734 static void
3735 anv_bind_buffer_memory(const VkBindBufferMemoryInfo *pBindInfo)
3736 {
3737 ANV_FROM_HANDLE(anv_device_memory, mem, pBindInfo->memory);
3738 ANV_FROM_HANDLE(anv_buffer, buffer, pBindInfo->buffer);
3739
3740 assert(pBindInfo->sType == VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO);
3741
3742 if (mem) {
3743 buffer->address = (struct anv_address) {
3744 .bo = mem->bo,
3745 .offset = pBindInfo->memoryOffset,
3746 };
3747 } else {
3748 buffer->address = ANV_NULL_ADDRESS;
3749 }
3750 }
3751
3752 VkResult anv_BindBufferMemory(
3753 VkDevice device,
3754 VkBuffer buffer,
3755 VkDeviceMemory memory,
3756 VkDeviceSize memoryOffset)
3757 {
3758 anv_bind_buffer_memory(
3759 &(VkBindBufferMemoryInfo) {
3760 .sType = VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO,
3761 .buffer = buffer,
3762 .memory = memory,
3763 .memoryOffset = memoryOffset,
3764 });
3765
3766 return VK_SUCCESS;
3767 }
3768
3769 VkResult anv_BindBufferMemory2(
3770 VkDevice device,
3771 uint32_t bindInfoCount,
3772 const VkBindBufferMemoryInfo* pBindInfos)
3773 {
3774 for (uint32_t i = 0; i < bindInfoCount; i++)
3775 anv_bind_buffer_memory(&pBindInfos[i]);
3776
3777 return VK_SUCCESS;
3778 }
3779
3780 VkResult anv_QueueBindSparse(
3781 VkQueue _queue,
3782 uint32_t bindInfoCount,
3783 const VkBindSparseInfo* pBindInfo,
3784 VkFence fence)
3785 {
3786 ANV_FROM_HANDLE(anv_queue, queue, _queue);
3787 if (anv_device_is_lost(queue->device))
3788 return VK_ERROR_DEVICE_LOST;
3789
3790 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
3791 }
3792
3793 // Event functions
3794
3795 VkResult anv_CreateEvent(
3796 VkDevice _device,
3797 const VkEventCreateInfo* pCreateInfo,
3798 const VkAllocationCallbacks* pAllocator,
3799 VkEvent* pEvent)
3800 {
3801 ANV_FROM_HANDLE(anv_device, device, _device);
3802 struct anv_state state;
3803 struct anv_event *event;
3804
3805 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_EVENT_CREATE_INFO);
3806
3807 state = anv_state_pool_alloc(&device->dynamic_state_pool,
3808 sizeof(*event), 8);
3809 event = state.map;
3810 event->state = state;
3811 event->semaphore = VK_EVENT_RESET;
3812
3813 if (!device->info.has_llc) {
3814 /* Make sure the writes we're flushing have landed. */
3815 __builtin_ia32_mfence();
3816 __builtin_ia32_clflush(event);
3817 }
3818
3819 *pEvent = anv_event_to_handle(event);
3820
3821 return VK_SUCCESS;
3822 }
3823
3824 void anv_DestroyEvent(
3825 VkDevice _device,
3826 VkEvent _event,
3827 const VkAllocationCallbacks* pAllocator)
3828 {
3829 ANV_FROM_HANDLE(anv_device, device, _device);
3830 ANV_FROM_HANDLE(anv_event, event, _event);
3831
3832 if (!event)
3833 return;
3834
3835 anv_state_pool_free(&device->dynamic_state_pool, event->state);
3836 }
3837
3838 VkResult anv_GetEventStatus(
3839 VkDevice _device,
3840 VkEvent _event)
3841 {
3842 ANV_FROM_HANDLE(anv_device, device, _device);
3843 ANV_FROM_HANDLE(anv_event, event, _event);
3844
3845 if (anv_device_is_lost(device))
3846 return VK_ERROR_DEVICE_LOST;
3847
3848 if (!device->info.has_llc) {
3849 /* Invalidate read cache before reading event written by GPU. */
3850 __builtin_ia32_clflush(event);
3851 __builtin_ia32_mfence();
3852
3853 }
3854
3855 return event->semaphore;
3856 }
3857
3858 VkResult anv_SetEvent(
3859 VkDevice _device,
3860 VkEvent _event)
3861 {
3862 ANV_FROM_HANDLE(anv_device, device, _device);
3863 ANV_FROM_HANDLE(anv_event, event, _event);
3864
3865 event->semaphore = VK_EVENT_SET;
3866
3867 if (!device->info.has_llc) {
3868 /* Make sure the writes we're flushing have landed. */
3869 __builtin_ia32_mfence();
3870 __builtin_ia32_clflush(event);
3871 }
3872
3873 return VK_SUCCESS;
3874 }
3875
3876 VkResult anv_ResetEvent(
3877 VkDevice _device,
3878 VkEvent _event)
3879 {
3880 ANV_FROM_HANDLE(anv_device, device, _device);
3881 ANV_FROM_HANDLE(anv_event, event, _event);
3882
3883 event->semaphore = VK_EVENT_RESET;
3884
3885 if (!device->info.has_llc) {
3886 /* Make sure the writes we're flushing have landed. */
3887 __builtin_ia32_mfence();
3888 __builtin_ia32_clflush(event);
3889 }
3890
3891 return VK_SUCCESS;
3892 }
3893
3894 // Buffer functions
3895
3896 VkResult anv_CreateBuffer(
3897 VkDevice _device,
3898 const VkBufferCreateInfo* pCreateInfo,
3899 const VkAllocationCallbacks* pAllocator,
3900 VkBuffer* pBuffer)
3901 {
3902 ANV_FROM_HANDLE(anv_device, device, _device);
3903 struct anv_buffer *buffer;
3904
3905 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
3906
3907 buffer = vk_alloc2(&device->alloc, pAllocator, sizeof(*buffer), 8,
3908 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
3909 if (buffer == NULL)
3910 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
3911
3912 buffer->size = pCreateInfo->size;
3913 buffer->usage = pCreateInfo->usage;
3914 buffer->address = ANV_NULL_ADDRESS;
3915
3916 *pBuffer = anv_buffer_to_handle(buffer);
3917
3918 return VK_SUCCESS;
3919 }
3920
3921 void anv_DestroyBuffer(
3922 VkDevice _device,
3923 VkBuffer _buffer,
3924 const VkAllocationCallbacks* pAllocator)
3925 {
3926 ANV_FROM_HANDLE(anv_device, device, _device);
3927 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
3928
3929 if (!buffer)
3930 return;
3931
3932 vk_free2(&device->alloc, pAllocator, buffer);
3933 }
3934
3935 VkDeviceAddress anv_GetBufferDeviceAddressEXT(
3936 VkDevice device,
3937 const VkBufferDeviceAddressInfoEXT* pInfo)
3938 {
3939 ANV_FROM_HANDLE(anv_buffer, buffer, pInfo->buffer);
3940
3941 assert(buffer->address.bo->flags & EXEC_OBJECT_PINNED);
3942
3943 return anv_address_physical(buffer->address);
3944 }
3945
3946 void
3947 anv_fill_buffer_surface_state(struct anv_device *device, struct anv_state state,
3948 enum isl_format format,
3949 struct anv_address address,
3950 uint32_t range, uint32_t stride)
3951 {
3952 isl_buffer_fill_state(&device->isl_dev, state.map,
3953 .address = anv_address_physical(address),
3954 .mocs = device->isl_dev.mocs.internal,
3955 .size_B = range,
3956 .format = format,
3957 .swizzle = ISL_SWIZZLE_IDENTITY,
3958 .stride_B = stride);
3959 }
3960
3961 void anv_DestroySampler(
3962 VkDevice _device,
3963 VkSampler _sampler,
3964 const VkAllocationCallbacks* pAllocator)
3965 {
3966 ANV_FROM_HANDLE(anv_device, device, _device);
3967 ANV_FROM_HANDLE(anv_sampler, sampler, _sampler);
3968
3969 if (!sampler)
3970 return;
3971
3972 if (sampler->bindless_state.map) {
3973 anv_state_pool_free(&device->dynamic_state_pool,
3974 sampler->bindless_state);
3975 }
3976
3977 vk_free2(&device->alloc, pAllocator, sampler);
3978 }
3979
3980 VkResult anv_CreateFramebuffer(
3981 VkDevice _device,
3982 const VkFramebufferCreateInfo* pCreateInfo,
3983 const VkAllocationCallbacks* pAllocator,
3984 VkFramebuffer* pFramebuffer)
3985 {
3986 ANV_FROM_HANDLE(anv_device, device, _device);
3987 struct anv_framebuffer *framebuffer;
3988
3989 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
3990
3991 size_t size = sizeof(*framebuffer);
3992
3993 /* VK_KHR_imageless_framebuffer extension says:
3994 *
3995 * If flags includes VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR,
3996 * parameter pAttachments is ignored.
3997 */
3998 if (!(pCreateInfo->flags & VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR)) {
3999 size += sizeof(struct anv_image_view *) * pCreateInfo->attachmentCount;
4000 framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
4001 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
4002 if (framebuffer == NULL)
4003 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
4004
4005 for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
4006 ANV_FROM_HANDLE(anv_image_view, iview, pCreateInfo->pAttachments[i]);
4007 framebuffer->attachments[i] = iview;
4008 }
4009 framebuffer->attachment_count = pCreateInfo->attachmentCount;
4010 } else {
4011 assert(device->enabled_extensions.KHR_imageless_framebuffer);
4012 framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
4013 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
4014 if (framebuffer == NULL)
4015 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
4016
4017 framebuffer->attachment_count = 0;
4018 }
4019
4020 framebuffer->width = pCreateInfo->width;
4021 framebuffer->height = pCreateInfo->height;
4022 framebuffer->layers = pCreateInfo->layers;
4023
4024 *pFramebuffer = anv_framebuffer_to_handle(framebuffer);
4025
4026 return VK_SUCCESS;
4027 }
4028
4029 void anv_DestroyFramebuffer(
4030 VkDevice _device,
4031 VkFramebuffer _fb,
4032 const VkAllocationCallbacks* pAllocator)
4033 {
4034 ANV_FROM_HANDLE(anv_device, device, _device);
4035 ANV_FROM_HANDLE(anv_framebuffer, fb, _fb);
4036
4037 if (!fb)
4038 return;
4039
4040 vk_free2(&device->alloc, pAllocator, fb);
4041 }
4042
4043 static const VkTimeDomainEXT anv_time_domains[] = {
4044 VK_TIME_DOMAIN_DEVICE_EXT,
4045 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT,
4046 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT,
4047 };
4048
4049 VkResult anv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
4050 VkPhysicalDevice physicalDevice,
4051 uint32_t *pTimeDomainCount,
4052 VkTimeDomainEXT *pTimeDomains)
4053 {
4054 int d;
4055 VK_OUTARRAY_MAKE(out, pTimeDomains, pTimeDomainCount);
4056
4057 for (d = 0; d < ARRAY_SIZE(anv_time_domains); d++) {
4058 vk_outarray_append(&out, i) {
4059 *i = anv_time_domains[d];
4060 }
4061 }
4062
4063 return vk_outarray_status(&out);
4064 }
4065
4066 static uint64_t
4067 anv_clock_gettime(clockid_t clock_id)
4068 {
4069 struct timespec current;
4070 int ret;
4071
4072 ret = clock_gettime(clock_id, &current);
4073 if (ret < 0 && clock_id == CLOCK_MONOTONIC_RAW)
4074 ret = clock_gettime(CLOCK_MONOTONIC, &current);
4075 if (ret < 0)
4076 return 0;
4077
4078 return (uint64_t) current.tv_sec * 1000000000ULL + current.tv_nsec;
4079 }
4080
4081 #define TIMESTAMP 0x2358
4082
4083 VkResult anv_GetCalibratedTimestampsEXT(
4084 VkDevice _device,
4085 uint32_t timestampCount,
4086 const VkCalibratedTimestampInfoEXT *pTimestampInfos,
4087 uint64_t *pTimestamps,
4088 uint64_t *pMaxDeviation)
4089 {
4090 ANV_FROM_HANDLE(anv_device, device, _device);
4091 uint64_t timestamp_frequency = device->info.timestamp_frequency;
4092 int ret;
4093 int d;
4094 uint64_t begin, end;
4095 uint64_t max_clock_period = 0;
4096
4097 begin = anv_clock_gettime(CLOCK_MONOTONIC_RAW);
4098
4099 for (d = 0; d < timestampCount; d++) {
4100 switch (pTimestampInfos[d].timeDomain) {
4101 case VK_TIME_DOMAIN_DEVICE_EXT:
4102 ret = anv_gem_reg_read(device, TIMESTAMP | 1,
4103 &pTimestamps[d]);
4104
4105 if (ret != 0) {
4106 return anv_device_set_lost(device, "Failed to read the TIMESTAMP "
4107 "register: %m");
4108 }
4109 uint64_t device_period = DIV_ROUND_UP(1000000000, timestamp_frequency);
4110 max_clock_period = MAX2(max_clock_period, device_period);
4111 break;
4112 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT:
4113 pTimestamps[d] = anv_clock_gettime(CLOCK_MONOTONIC);
4114 max_clock_period = MAX2(max_clock_period, 1);
4115 break;
4116
4117 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT:
4118 pTimestamps[d] = begin;
4119 break;
4120 default:
4121 pTimestamps[d] = 0;
4122 break;
4123 }
4124 }
4125
4126 end = anv_clock_gettime(CLOCK_MONOTONIC_RAW);
4127
4128 /*
4129 * The maximum deviation is the sum of the interval over which we
4130 * perform the sampling and the maximum period of any sampled
4131 * clock. That's because the maximum skew between any two sampled
4132 * clock edges is when the sampled clock with the largest period is
4133 * sampled at the end of that period but right at the beginning of the
4134 * sampling interval and some other clock is sampled right at the
4135 * begining of its sampling period and right at the end of the
4136 * sampling interval. Let's assume the GPU has the longest clock
4137 * period and that the application is sampling GPU and monotonic:
4138 *
4139 * s e
4140 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
4141 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4142 *
4143 * g
4144 * 0 1 2 3
4145 * GPU -----_____-----_____-----_____-----_____
4146 *
4147 * m
4148 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
4149 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4150 *
4151 * Interval <----------------->
4152 * Deviation <-------------------------->
4153 *
4154 * s = read(raw) 2
4155 * g = read(GPU) 1
4156 * m = read(monotonic) 2
4157 * e = read(raw) b
4158 *
4159 * We round the sample interval up by one tick to cover sampling error
4160 * in the interval clock
4161 */
4162
4163 uint64_t sample_interval = end - begin + 1;
4164
4165 *pMaxDeviation = sample_interval + max_clock_period;
4166
4167 return VK_SUCCESS;
4168 }
4169
4170 /* vk_icd.h does not declare this function, so we declare it here to
4171 * suppress Wmissing-prototypes.
4172 */
4173 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4174 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion);
4175
4176 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4177 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion)
4178 {
4179 /* For the full details on loader interface versioning, see
4180 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
4181 * What follows is a condensed summary, to help you navigate the large and
4182 * confusing official doc.
4183 *
4184 * - Loader interface v0 is incompatible with later versions. We don't
4185 * support it.
4186 *
4187 * - In loader interface v1:
4188 * - The first ICD entrypoint called by the loader is
4189 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
4190 * entrypoint.
4191 * - The ICD must statically expose no other Vulkan symbol unless it is
4192 * linked with -Bsymbolic.
4193 * - Each dispatchable Vulkan handle created by the ICD must be
4194 * a pointer to a struct whose first member is VK_LOADER_DATA. The
4195 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
4196 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
4197 * vkDestroySurfaceKHR(). The ICD must be capable of working with
4198 * such loader-managed surfaces.
4199 *
4200 * - Loader interface v2 differs from v1 in:
4201 * - The first ICD entrypoint called by the loader is
4202 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
4203 * statically expose this entrypoint.
4204 *
4205 * - Loader interface v3 differs from v2 in:
4206 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
4207 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
4208 * because the loader no longer does so.
4209 *
4210 * - Loader interface v4 differs from v3 in:
4211 * - The ICD must implement vk_icdGetPhysicalDeviceProcAddr().
4212 */
4213 *pSupportedVersion = MIN2(*pSupportedVersion, 4u);
4214 return VK_SUCCESS;
4215 }