projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
vulkan/wsi: Add a driconf option to force WSI to advertise BGRA8_UNORM first
[mesa.git] / src / intel / vulkan / anv_device.c
1 /*
2 * Copyright © 2015 Intel Corporation
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 */
23
24 #include <assert.h>
25 #include <stdbool.h>
26 #include <string.h>
27 #include <sys/mman.h>
28 #include <sys/sysinfo.h>
29 #include <unistd.h>
30 #include <fcntl.h>
31 #include <xf86drm.h>
32 #include "drm-uapi/drm_fourcc.h"
33
34 #include "anv_private.h"
35 #include "util/debug.h"
36 #include "util/build_id.h"
37 #include "util/disk_cache.h"
38 #include "util/mesa-sha1.h"
39 #include "util/os_file.h"
40 #include "util/u_atomic.h"
41 #include "util/u_string.h"
42 #include "util/xmlpool.h"
43 #include "git_sha1.h"
44 #include "vk_util.h"
45 #include "common/gen_aux_map.h"
46 #include "common/gen_defines.h"
47 #include "compiler/glsl_types.h"
48
49 #include "genxml/gen7_pack.h"
50
51 static const char anv_dri_options_xml[] =
52 DRI_CONF_BEGIN
53 DRI_CONF_SECTION_PERFORMANCE
54 DRI_CONF_VK_X11_OVERRIDE_MIN_IMAGE_COUNT(0)
55 DRI_CONF_VK_X11_STRICT_IMAGE_COUNT("false")
56 DRI_CONF_SECTION_END
57
58 DRI_CONF_SECTION_DEBUG
59 DRI_CONF_ALWAYS_FLUSH_CACHE("false")
60 DRI_CONF_VK_WSI_FORCE_BGRA8_UNORM_FIRST("false")
61 DRI_CONF_SECTION_END
62 DRI_CONF_END;
63
64 /* This is probably far to big but it reflects the max size used for messages
65 * in OpenGLs KHR_debug.
66 */
67 #define MAX_DEBUG_MESSAGE_LENGTH 4096
68
69 static void
70 compiler_debug_log(void *data, const char *fmt, ...)
71 {
72 char str[MAX_DEBUG_MESSAGE_LENGTH];
73 struct anv_device *device = (struct anv_device *)data;
74
75 if (list_is_empty(&device->instance->debug_report_callbacks.callbacks))
76 return;
77
78 va_list args;
79 va_start(args, fmt);
80 (void) vsnprintf(str, MAX_DEBUG_MESSAGE_LENGTH, fmt, args);
81 va_end(args);
82
83 vk_debug_report(&device->instance->debug_report_callbacks,
84 VK_DEBUG_REPORT_DEBUG_BIT_EXT,
85 VK_DEBUG_REPORT_OBJECT_TYPE_UNKNOWN_EXT,
86 0, 0, 0, "anv", str);
87 }
88
89 static void
90 compiler_perf_log(void *data, const char *fmt, ...)
91 {
92 va_list args;
93 va_start(args, fmt);
94
95 if (unlikely(INTEL_DEBUG & DEBUG_PERF))
96 intel_logd_v(fmt, args);
97
98 va_end(args);
99 }
100
101 static uint64_t
102 anv_compute_heap_size(int fd, uint64_t gtt_size)
103 {
104 /* Query the total ram from the system */
105 struct sysinfo info;
106 sysinfo(&info);
107
108 uint64_t total_ram = (uint64_t)info.totalram * (uint64_t)info.mem_unit;
109
110 /* We don't want to burn too much ram with the GPU. If the user has 4GiB
111 * or less, we use at most half. If they have more than 4GiB, we use 3/4.
112 */
113 uint64_t available_ram;
114 if (total_ram <= 4ull * 1024ull * 1024ull * 1024ull)
115 available_ram = total_ram / 2;
116 else
117 available_ram = total_ram * 3 / 4;
118
119 /* We also want to leave some padding for things we allocate in the driver,
120 * so don't go over 3/4 of the GTT either.
121 */
122 uint64_t available_gtt = gtt_size * 3 / 4;
123
124 return MIN2(available_ram, available_gtt);
125 }
126
127 static VkResult
128 anv_physical_device_init_heaps(struct anv_physical_device *device, int fd)
129 {
130 if (anv_gem_get_context_param(fd, 0, I915_CONTEXT_PARAM_GTT_SIZE,
131 &device->gtt_size) == -1) {
132 /* If, for whatever reason, we can't actually get the GTT size from the
133 * kernel (too old?) fall back to the aperture size.
134 */
135 anv_perf_warn(NULL, NULL,
136 "Failed to get I915_CONTEXT_PARAM_GTT_SIZE: %m");
137
138 if (anv_gem_get_aperture(fd, &device->gtt_size) == -1) {
139 return vk_errorf(NULL, NULL, VK_ERROR_INITIALIZATION_FAILED,
140 "failed to get aperture size: %m");
141 }
142 }
143
144 /* We only allow 48-bit addresses with softpin because knowing the actual
145 * address is required for the vertex cache flush workaround.
146 */
147 device->supports_48bit_addresses = (device->info.gen >= 8) &&
148 device->has_softpin &&
149 device->gtt_size > (4ULL << 30 /* GiB */);
150
151 uint64_t heap_size = anv_compute_heap_size(fd, device->gtt_size);
152
153 if (heap_size > (2ull << 30) && !device->supports_48bit_addresses) {
154 /* When running with an overridden PCI ID, we may get a GTT size from
155 * the kernel that is greater than 2 GiB but the execbuf check for 48bit
156 * address support can still fail. Just clamp the address space size to
157 * 2 GiB if we don't have 48-bit support.
158 */
159 intel_logw("%s:%d: The kernel reported a GTT size larger than 2 GiB but "
160 "not support for 48-bit addresses",
161 __FILE__, __LINE__);
162 heap_size = 2ull << 30;
163 }
164
165 device->memory.heap_count = 1;
166 device->memory.heaps[0] = (struct anv_memory_heap) {
167 .size = heap_size,
168 .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
169 };
170
171 uint32_t type_count = 0;
172 for (uint32_t heap = 0; heap < device->memory.heap_count; heap++) {
173 if (device->info.has_llc) {
174 /* Big core GPUs share LLC with the CPU and thus one memory type can be
175 * both cached and coherent at the same time.
176 */
177 device->memory.types[type_count++] = (struct anv_memory_type) {
178 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
179 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
180 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
181 VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
182 .heapIndex = heap,
183 };
184 } else {
185 /* The spec requires that we expose a host-visible, coherent memory
186 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
187 * to give the application a choice between cached, but not coherent and
188 * coherent but uncached (WC though).
189 */
190 device->memory.types[type_count++] = (struct anv_memory_type) {
191 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
192 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
193 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
194 .heapIndex = heap,
195 };
196 device->memory.types[type_count++] = (struct anv_memory_type) {
197 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
198 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
199 VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
200 .heapIndex = heap,
201 };
202 }
203 }
204 device->memory.type_count = type_count;
205
206 return VK_SUCCESS;
207 }
208
209 static VkResult
210 anv_physical_device_init_uuids(struct anv_physical_device *device)
211 {
212 const struct build_id_note *note =
213 build_id_find_nhdr_for_addr(anv_physical_device_init_uuids);
214 if (!note) {
215 return vk_errorf(device->instance, device,
216 VK_ERROR_INITIALIZATION_FAILED,
217 "Failed to find build-id");
218 }
219
220 unsigned build_id_len = build_id_length(note);
221 if (build_id_len < 20) {
222 return vk_errorf(device->instance, device,
223 VK_ERROR_INITIALIZATION_FAILED,
224 "build-id too short. It needs to be a SHA");
225 }
226
227 memcpy(device->driver_build_sha1, build_id_data(note), 20);
228
229 struct mesa_sha1 sha1_ctx;
230 uint8_t sha1[20];
231 STATIC_ASSERT(VK_UUID_SIZE <= sizeof(sha1));
232
233 /* The pipeline cache UUID is used for determining when a pipeline cache is
234 * invalid. It needs both a driver build and the PCI ID of the device.
235 */
236 _mesa_sha1_init(&sha1_ctx);
237 _mesa_sha1_update(&sha1_ctx, build_id_data(note), build_id_len);
238 _mesa_sha1_update(&sha1_ctx, &device->chipset_id,
239 sizeof(device->chipset_id));
240 _mesa_sha1_update(&sha1_ctx, &device->always_use_bindless,
241 sizeof(device->always_use_bindless));
242 _mesa_sha1_update(&sha1_ctx, &device->has_a64_buffer_access,
243 sizeof(device->has_a64_buffer_access));
244 _mesa_sha1_update(&sha1_ctx, &device->has_bindless_images,
245 sizeof(device->has_bindless_images));
246 _mesa_sha1_update(&sha1_ctx, &device->has_bindless_samplers,
247 sizeof(device->has_bindless_samplers));
248 _mesa_sha1_final(&sha1_ctx, sha1);
249 memcpy(device->pipeline_cache_uuid, sha1, VK_UUID_SIZE);
250
251 /* The driver UUID is used for determining sharability of images and memory
252 * between two Vulkan instances in separate processes. People who want to
253 * share memory need to also check the device UUID (below) so all this
254 * needs to be is the build-id.
255 */
256 memcpy(device->driver_uuid, build_id_data(note), VK_UUID_SIZE);
257
258 /* The device UUID uniquely identifies the given device within the machine.
259 * Since we never have more than one device, this doesn't need to be a real
260 * UUID. However, on the off-chance that someone tries to use this to
261 * cache pre-tiled images or something of the like, we use the PCI ID and
262 * some bits of ISL info to ensure that this is safe.
263 */
264 _mesa_sha1_init(&sha1_ctx);
265 _mesa_sha1_update(&sha1_ctx, &device->chipset_id,
266 sizeof(device->chipset_id));
267 _mesa_sha1_update(&sha1_ctx, &device->isl_dev.has_bit6_swizzling,
268 sizeof(device->isl_dev.has_bit6_swizzling));
269 _mesa_sha1_final(&sha1_ctx, sha1);
270 memcpy(device->device_uuid, sha1, VK_UUID_SIZE);
271
272 return VK_SUCCESS;
273 }
274
275 static void
276 anv_physical_device_init_disk_cache(struct anv_physical_device *device)
277 {
278 #ifdef ENABLE_SHADER_CACHE
279 char renderer[10];
280 ASSERTED int len = snprintf(renderer, sizeof(renderer), "anv_%04x",
281 device->chipset_id);
282 assert(len == sizeof(renderer) - 2);
283
284 char timestamp[41];
285 _mesa_sha1_format(timestamp, device->driver_build_sha1);
286
287 const uint64_t driver_flags =
288 brw_get_compiler_config_value(device->compiler);
289 device->disk_cache = disk_cache_create(renderer, timestamp, driver_flags);
290 #else
291 device->disk_cache = NULL;
292 #endif
293 }
294
295 static void
296 anv_physical_device_free_disk_cache(struct anv_physical_device *device)
297 {
298 #ifdef ENABLE_SHADER_CACHE
299 if (device->disk_cache)
300 disk_cache_destroy(device->disk_cache);
301 #else
302 assert(device->disk_cache == NULL);
303 #endif
304 }
305
306 static uint64_t
307 get_available_system_memory()
308 {
309 char *meminfo = os_read_file("/proc/meminfo");
310 if (!meminfo)
311 return 0;
312
313 char *str = strstr(meminfo, "MemAvailable:");
314 if (!str) {
315 free(meminfo);
316 return 0;
317 }
318
319 uint64_t kb_mem_available;
320 if (sscanf(str, "MemAvailable: %" PRIx64, &kb_mem_available) == 1) {
321 free(meminfo);
322 return kb_mem_available << 10;
323 }
324
325 free(meminfo);
326 return 0;
327 }
328
329 static VkResult
330 anv_physical_device_init(struct anv_physical_device *device,
331 struct anv_instance *instance,
332 drmDevicePtr drm_device)
333 {
334 const char *primary_path = drm_device->nodes[DRM_NODE_PRIMARY];
335 const char *path = drm_device->nodes[DRM_NODE_RENDER];
336 VkResult result;
337 int fd;
338 int master_fd = -1;
339
340 brw_process_intel_debug_variable();
341
342 fd = open(path, O_RDWR | O_CLOEXEC);
343 if (fd < 0)
344 return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
345
346 device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
347 device->instance = instance;
348
349 assert(strlen(path) < ARRAY_SIZE(device->path));
350 snprintf(device->path, ARRAY_SIZE(device->path), "%s", path);
351
352 if (!gen_get_device_info_from_fd(fd, &device->info)) {
353 result = vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
354 goto fail;
355 }
356 device->chipset_id = device->info.chipset_id;
357 device->no_hw = device->info.no_hw;
358
359 if (getenv("INTEL_NO_HW") != NULL)
360 device->no_hw = true;
361
362 device->pci_info.domain = drm_device->businfo.pci->domain;
363 device->pci_info.bus = drm_device->businfo.pci->bus;
364 device->pci_info.device = drm_device->businfo.pci->dev;
365 device->pci_info.function = drm_device->businfo.pci->func;
366
367 device->name = gen_get_device_name(device->chipset_id);
368
369 if (device->info.is_haswell) {
370 intel_logw("Haswell Vulkan support is incomplete");
371 } else if (device->info.gen == 7 && !device->info.is_baytrail) {
372 intel_logw("Ivy Bridge Vulkan support is incomplete");
373 } else if (device->info.gen == 7 && device->info.is_baytrail) {
374 intel_logw("Bay Trail Vulkan support is incomplete");
375 } else if (device->info.gen >= 8 && device->info.gen <= 11) {
376 /* Gen8-11 fully supported */
377 } else if (device->info.gen == 12) {
378 intel_logw("Vulkan is not yet fully supported on gen12");
379 } else {
380 result = vk_errorf(device->instance, device,
381 VK_ERROR_INCOMPATIBLE_DRIVER,
382 "Vulkan not yet supported on %s", device->name);
383 goto fail;
384 }
385
386 device->cmd_parser_version = -1;
387 if (device->info.gen == 7) {
388 device->cmd_parser_version =
389 anv_gem_get_param(fd, I915_PARAM_CMD_PARSER_VERSION);
390 if (device->cmd_parser_version == -1) {
391 result = vk_errorf(device->instance, device,
392 VK_ERROR_INITIALIZATION_FAILED,
393 "failed to get command parser version");
394 goto fail;
395 }
396 }
397
398 if (!anv_gem_get_param(fd, I915_PARAM_HAS_WAIT_TIMEOUT)) {
399 result = vk_errorf(device->instance, device,
400 VK_ERROR_INITIALIZATION_FAILED,
401 "kernel missing gem wait");
402 goto fail;
403 }
404
405 if (!anv_gem_get_param(fd, I915_PARAM_HAS_EXECBUF2)) {
406 result = vk_errorf(device->instance, device,
407 VK_ERROR_INITIALIZATION_FAILED,
408 "kernel missing execbuf2");
409 goto fail;
410 }
411
412 if (!device->info.has_llc &&
413 anv_gem_get_param(fd, I915_PARAM_MMAP_VERSION) < 1) {
414 result = vk_errorf(device->instance, device,
415 VK_ERROR_INITIALIZATION_FAILED,
416 "kernel missing wc mmap");
417 goto fail;
418 }
419
420 device->has_softpin = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_SOFTPIN);
421 device->has_exec_async = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_ASYNC);
422 device->has_exec_capture = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_CAPTURE);
423 device->has_exec_fence = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_FENCE);
424 device->has_syncobj = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_FENCE_ARRAY);
425 device->has_syncobj_wait = device->has_syncobj &&
426 anv_gem_supports_syncobj_wait(fd);
427 device->has_context_priority = anv_gem_has_context_priority(fd);
428
429 result = anv_physical_device_init_heaps(device, fd);
430 if (result != VK_SUCCESS)
431 goto fail;
432
433 device->use_softpin = device->has_softpin &&
434 device->supports_48bit_addresses;
435
436 device->has_context_isolation =
437 anv_gem_get_param(fd, I915_PARAM_HAS_CONTEXT_ISOLATION);
438
439 device->always_use_bindless =
440 env_var_as_boolean("ANV_ALWAYS_BINDLESS", false);
441
442 /* We first got the A64 messages on broadwell and we can only use them if
443 * we can pass addresses directly into the shader which requires softpin.
444 */
445 device->has_a64_buffer_access = device->info.gen >= 8 &&
446 device->use_softpin;
447
448 /* We first get bindless image access on Skylake and we can only really do
449 * it if we don't have any relocations so we need softpin.
450 */
451 device->has_bindless_images = device->info.gen >= 9 &&
452 device->use_softpin;
453
454 /* We've had bindless samplers since Ivy Bridge (forever in Vulkan terms)
455 * because it's just a matter of setting the sampler address in the sample
456 * message header. However, we've not bothered to wire it up for vec4 so
457 * we leave it disabled on gen7.
458 */
459 device->has_bindless_samplers = device->info.gen >= 8;
460
461 device->has_mem_available = get_available_system_memory() != 0;
462
463 device->always_flush_cache =
464 driQueryOptionb(&instance->dri_options, "always_flush_cache");
465
466 /* Starting with Gen10, the timestamp frequency of the command streamer may
467 * vary from one part to another. We can query the value from the kernel.
468 */
469 if (device->info.gen >= 10) {
470 int timestamp_frequency =
471 anv_gem_get_param(fd, I915_PARAM_CS_TIMESTAMP_FREQUENCY);
472
473 if (timestamp_frequency < 0)
474 intel_logw("Kernel 4.16-rc1+ required to properly query CS timestamp frequency");
475 else
476 device->info.timestamp_frequency = timestamp_frequency;
477 }
478
479 /* GENs prior to 8 do not support EU/Subslice info */
480 if (device->info.gen >= 8) {
481 device->subslice_total = anv_gem_get_param(fd, I915_PARAM_SUBSLICE_TOTAL);
482 device->eu_total = anv_gem_get_param(fd, I915_PARAM_EU_TOTAL);
483
484 /* Without this information, we cannot get the right Braswell
485 * brandstrings, and we have to use conservative numbers for GPGPU on
486 * many platforms, but otherwise, things will just work.
487 */
488 if (device->subslice_total < 1 || device->eu_total < 1) {
489 intel_logw("Kernel 4.1 required to properly query GPU properties");
490 }
491 } else if (device->info.gen == 7) {
492 device->subslice_total = 1 << (device->info.gt - 1);
493 }
494
495 if (device->info.is_cherryview &&
496 device->subslice_total > 0 && device->eu_total > 0) {
497 /* Logical CS threads = EUs per subslice * num threads per EU */
498 uint32_t max_cs_threads =
499 device->eu_total / device->subslice_total * device->info.num_thread_per_eu;
500
501 /* Fuse configurations may give more threads than expected, never less. */
502 if (max_cs_threads > device->info.max_cs_threads)
503 device->info.max_cs_threads = max_cs_threads;
504 }
505
506 device->compiler = brw_compiler_create(NULL, &device->info);
507 if (device->compiler == NULL) {
508 result = vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
509 goto fail;
510 }
511 device->compiler->shader_debug_log = compiler_debug_log;
512 device->compiler->shader_perf_log = compiler_perf_log;
513 device->compiler->supports_pull_constants = false;
514 device->compiler->constant_buffer_0_is_relative =
515 device->info.gen < 8 || !device->has_context_isolation;
516 device->compiler->supports_shader_constants = true;
517 device->compiler->compact_params = false;
518
519 /* Broadwell PRM says:
520 *
521 * "Before Gen8, there was a historical configuration control field to
522 * swizzle address bit[6] for in X/Y tiling modes. This was set in three
523 * different places: TILECTL[1:0], ARB_MODE[5:4], and
524 * DISP_ARB_CTL[14:13].
525 *
526 * For Gen8 and subsequent generations, the swizzle fields are all
527 * reserved, and the CPU's memory controller performs all address
528 * swizzling modifications."
529 */
530 bool swizzled =
531 device->info.gen < 8 && anv_gem_get_bit6_swizzle(fd, I915_TILING_X);
532
533 isl_device_init(&device->isl_dev, &device->info, swizzled);
534
535 result = anv_physical_device_init_uuids(device);
536 if (result != VK_SUCCESS)
537 goto fail;
538
539 anv_physical_device_init_disk_cache(device);
540
541 if (instance->enabled_extensions.KHR_display) {
542 master_fd = open(primary_path, O_RDWR | O_CLOEXEC);
543 if (master_fd >= 0) {
544 /* prod the device with a GETPARAM call which will fail if
545 * we don't have permission to even render on this device
546 */
547 if (anv_gem_get_param(master_fd, I915_PARAM_CHIPSET_ID) == 0) {
548 close(master_fd);
549 master_fd = -1;
550 }
551 }
552 }
553 device->master_fd = master_fd;
554
555 result = anv_init_wsi(device);
556 if (result != VK_SUCCESS) {
557 ralloc_free(device->compiler);
558 anv_physical_device_free_disk_cache(device);
559 goto fail;
560 }
561
562 device->perf = anv_get_perf(&device->info, fd);
563
564 anv_physical_device_get_supported_extensions(device,
565 &device->supported_extensions);
566
567
568 device->local_fd = fd;
569
570 return VK_SUCCESS;
571
572 fail:
573 close(fd);
574 if (master_fd != -1)
575 close(master_fd);
576 return result;
577 }
578
579 static void
580 anv_physical_device_finish(struct anv_physical_device *device)
581 {
582 anv_finish_wsi(device);
583 anv_physical_device_free_disk_cache(device);
584 ralloc_free(device->compiler);
585 ralloc_free(device->perf);
586 close(device->local_fd);
587 if (device->master_fd >= 0)
588 close(device->master_fd);
589 }
590
591 static void *
592 default_alloc_func(void *pUserData, size_t size, size_t align,
593 VkSystemAllocationScope allocationScope)
594 {
595 return malloc(size);
596 }
597
598 static void *
599 default_realloc_func(void *pUserData, void *pOriginal, size_t size,
600 size_t align, VkSystemAllocationScope allocationScope)
601 {
602 return realloc(pOriginal, size);
603 }
604
605 static void
606 default_free_func(void *pUserData, void *pMemory)
607 {
608 free(pMemory);
609 }
610
611 static const VkAllocationCallbacks default_alloc = {
612 .pUserData = NULL,
613 .pfnAllocation = default_alloc_func,
614 .pfnReallocation = default_realloc_func,
615 .pfnFree = default_free_func,
616 };
617
618 VkResult anv_EnumerateInstanceExtensionProperties(
619 const char* pLayerName,
620 uint32_t* pPropertyCount,
621 VkExtensionProperties* pProperties)
622 {
623 VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
624
625 for (int i = 0; i < ANV_INSTANCE_EXTENSION_COUNT; i++) {
626 if (anv_instance_extensions_supported.extensions[i]) {
627 vk_outarray_append(&out, prop) {
628 *prop = anv_instance_extensions[i];
629 }
630 }
631 }
632
633 return vk_outarray_status(&out);
634 }
635
636 VkResult anv_CreateInstance(
637 const VkInstanceCreateInfo* pCreateInfo,
638 const VkAllocationCallbacks* pAllocator,
639 VkInstance* pInstance)
640 {
641 struct anv_instance *instance;
642 VkResult result;
643
644 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO);
645
646 struct anv_instance_extension_table enabled_extensions = {};
647 for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
648 int idx;
649 for (idx = 0; idx < ANV_INSTANCE_EXTENSION_COUNT; idx++) {
650 if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
651 anv_instance_extensions[idx].extensionName) == 0)
652 break;
653 }
654
655 if (idx >= ANV_INSTANCE_EXTENSION_COUNT)
656 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
657
658 if (!anv_instance_extensions_supported.extensions[idx])
659 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
660
661 enabled_extensions.extensions[idx] = true;
662 }
663
664 instance = vk_alloc2(&default_alloc, pAllocator, sizeof(*instance), 8,
665 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
666 if (!instance)
667 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
668
669 instance->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
670
671 if (pAllocator)
672 instance->alloc = *pAllocator;
673 else
674 instance->alloc = default_alloc;
675
676 instance->app_info = (struct anv_app_info) { .api_version = 0 };
677 if (pCreateInfo->pApplicationInfo) {
678 const VkApplicationInfo *app = pCreateInfo->pApplicationInfo;
679
680 instance->app_info.app_name =
681 vk_strdup(&instance->alloc, app->pApplicationName,
682 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
683 instance->app_info.app_version = app->applicationVersion;
684
685 instance->app_info.engine_name =
686 vk_strdup(&instance->alloc, app->pEngineName,
687 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
688 instance->app_info.engine_version = app->engineVersion;
689
690 instance->app_info.api_version = app->apiVersion;
691 }
692
693 if (instance->app_info.api_version == 0)
694 instance->app_info.api_version = VK_API_VERSION_1_0;
695
696 instance->enabled_extensions = enabled_extensions;
697
698 for (unsigned i = 0; i < ARRAY_SIZE(instance->dispatch.entrypoints); i++) {
699 /* Vulkan requires that entrypoints for extensions which have not been
700 * enabled must not be advertised.
701 */
702 if (!anv_instance_entrypoint_is_enabled(i, instance->app_info.api_version,
703 &instance->enabled_extensions)) {
704 instance->dispatch.entrypoints[i] = NULL;
705 } else {
706 instance->dispatch.entrypoints[i] =
707 anv_instance_dispatch_table.entrypoints[i];
708 }
709 }
710
711 struct anv_physical_device *pdevice = &instance->physicalDevice;
712 for (unsigned i = 0; i < ARRAY_SIZE(pdevice->dispatch.entrypoints); i++) {
713 /* Vulkan requires that entrypoints for extensions which have not been
714 * enabled must not be advertised.
715 */
716 if (!anv_physical_device_entrypoint_is_enabled(i, instance->app_info.api_version,
717 &instance->enabled_extensions)) {
718 pdevice->dispatch.entrypoints[i] = NULL;
719 } else {
720 pdevice->dispatch.entrypoints[i] =
721 anv_physical_device_dispatch_table.entrypoints[i];
722 }
723 }
724
725 for (unsigned i = 0; i < ARRAY_SIZE(instance->device_dispatch.entrypoints); i++) {
726 /* Vulkan requires that entrypoints for extensions which have not been
727 * enabled must not be advertised.
728 */
729 if (!anv_device_entrypoint_is_enabled(i, instance->app_info.api_version,
730 &instance->enabled_extensions, NULL)) {
731 instance->device_dispatch.entrypoints[i] = NULL;
732 } else {
733 instance->device_dispatch.entrypoints[i] =
734 anv_device_dispatch_table.entrypoints[i];
735 }
736 }
737
738 instance->physicalDeviceCount = -1;
739
740 result = vk_debug_report_instance_init(&instance->debug_report_callbacks);
741 if (result != VK_SUCCESS) {
742 vk_free2(&default_alloc, pAllocator, instance);
743 return vk_error(result);
744 }
745
746 instance->pipeline_cache_enabled =
747 env_var_as_boolean("ANV_ENABLE_PIPELINE_CACHE", true);
748
749 glsl_type_singleton_init_or_ref();
750
751 VG(VALGRIND_CREATE_MEMPOOL(instance, 0, false));
752
753 driParseOptionInfo(&instance->available_dri_options, anv_dri_options_xml);
754 driParseConfigFiles(&instance->dri_options, &instance->available_dri_options,
755 0, "anv", NULL,
756 instance->app_info.engine_name,
757 instance->app_info.engine_version);
758
759 *pInstance = anv_instance_to_handle(instance);
760
761 return VK_SUCCESS;
762 }
763
764 void anv_DestroyInstance(
765 VkInstance _instance,
766 const VkAllocationCallbacks* pAllocator)
767 {
768 ANV_FROM_HANDLE(anv_instance, instance, _instance);
769
770 if (!instance)
771 return;
772
773 if (instance->physicalDeviceCount > 0) {
774 /* We support at most one physical device. */
775 assert(instance->physicalDeviceCount == 1);
776 anv_physical_device_finish(&instance->physicalDevice);
777 }
778
779 vk_free(&instance->alloc, (char *)instance->app_info.app_name);
780 vk_free(&instance->alloc, (char *)instance->app_info.engine_name);
781
782 VG(VALGRIND_DESTROY_MEMPOOL(instance));
783
784 vk_debug_report_instance_destroy(&instance->debug_report_callbacks);
785
786 glsl_type_singleton_decref();
787
788 driDestroyOptionCache(&instance->dri_options);
789 driDestroyOptionInfo(&instance->available_dri_options);
790
791 vk_free(&instance->alloc, instance);
792 }
793
794 static VkResult
795 anv_enumerate_devices(struct anv_instance *instance)
796 {
797 /* TODO: Check for more devices ? */
798 drmDevicePtr devices[8];
799 VkResult result = VK_ERROR_INCOMPATIBLE_DRIVER;
800 int max_devices;
801
802 instance->physicalDeviceCount = 0;
803
804 max_devices = drmGetDevices2(0, devices, ARRAY_SIZE(devices));
805 if (max_devices < 1)
806 return VK_ERROR_INCOMPATIBLE_DRIVER;
807
808 for (unsigned i = 0; i < (unsigned)max_devices; i++) {
809 if (devices[i]->available_nodes & 1 << DRM_NODE_RENDER &&
810 devices[i]->bustype == DRM_BUS_PCI &&
811 devices[i]->deviceinfo.pci->vendor_id == 0x8086) {
812
813 result = anv_physical_device_init(&instance->physicalDevice,
814 instance, devices[i]);
815 if (result != VK_ERROR_INCOMPATIBLE_DRIVER)
816 break;
817 }
818 }
819 drmFreeDevices(devices, max_devices);
820
821 if (result == VK_SUCCESS)
822 instance->physicalDeviceCount = 1;
823
824 return result;
825 }
826
827 static VkResult
828 anv_instance_ensure_physical_device(struct anv_instance *instance)
829 {
830 if (instance->physicalDeviceCount < 0) {
831 VkResult result = anv_enumerate_devices(instance);
832 if (result != VK_SUCCESS &&
833 result != VK_ERROR_INCOMPATIBLE_DRIVER)
834 return result;
835 }
836
837 return VK_SUCCESS;
838 }
839
840 VkResult anv_EnumeratePhysicalDevices(
841 VkInstance _instance,
842 uint32_t* pPhysicalDeviceCount,
843 VkPhysicalDevice* pPhysicalDevices)
844 {
845 ANV_FROM_HANDLE(anv_instance, instance, _instance);
846 VK_OUTARRAY_MAKE(out, pPhysicalDevices, pPhysicalDeviceCount);
847
848 VkResult result = anv_instance_ensure_physical_device(instance);
849 if (result != VK_SUCCESS)
850 return result;
851
852 if (instance->physicalDeviceCount == 0)
853 return VK_SUCCESS;
854
855 assert(instance->physicalDeviceCount == 1);
856 vk_outarray_append(&out, i) {
857 *i = anv_physical_device_to_handle(&instance->physicalDevice);
858 }
859
860 return vk_outarray_status(&out);
861 }
862
863 VkResult anv_EnumeratePhysicalDeviceGroups(
864 VkInstance _instance,
865 uint32_t* pPhysicalDeviceGroupCount,
866 VkPhysicalDeviceGroupProperties* pPhysicalDeviceGroupProperties)
867 {
868 ANV_FROM_HANDLE(anv_instance, instance, _instance);
869 VK_OUTARRAY_MAKE(out, pPhysicalDeviceGroupProperties,
870 pPhysicalDeviceGroupCount);
871
872 VkResult result = anv_instance_ensure_physical_device(instance);
873 if (result != VK_SUCCESS)
874 return result;
875
876 if (instance->physicalDeviceCount == 0)
877 return VK_SUCCESS;
878
879 assert(instance->physicalDeviceCount == 1);
880
881 vk_outarray_append(&out, p) {
882 p->physicalDeviceCount = 1;
883 memset(p->physicalDevices, 0, sizeof(p->physicalDevices));
884 p->physicalDevices[0] =
885 anv_physical_device_to_handle(&instance->physicalDevice);
886 p->subsetAllocation = false;
887
888 vk_foreach_struct(ext, p->pNext)
889 anv_debug_ignored_stype(ext->sType);
890 }
891
892 return vk_outarray_status(&out);
893 }
894
895 void anv_GetPhysicalDeviceFeatures(
896 VkPhysicalDevice physicalDevice,
897 VkPhysicalDeviceFeatures* pFeatures)
898 {
899 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
900
901 *pFeatures = (VkPhysicalDeviceFeatures) {
902 .robustBufferAccess = true,
903 .fullDrawIndexUint32 = true,
904 .imageCubeArray = true,
905 .independentBlend = true,
906 .geometryShader = true,
907 .tessellationShader = true,
908 .sampleRateShading = true,
909 .dualSrcBlend = true,
910 .logicOp = true,
911 .multiDrawIndirect = true,
912 .drawIndirectFirstInstance = true,
913 .depthClamp = true,
914 .depthBiasClamp = true,
915 .fillModeNonSolid = true,
916 .depthBounds = pdevice->info.gen >= 12,
917 .wideLines = true,
918 .largePoints = true,
919 .alphaToOne = true,
920 .multiViewport = true,
921 .samplerAnisotropy = true,
922 .textureCompressionETC2 = pdevice->info.gen >= 8 ||
923 pdevice->info.is_baytrail,
924 .textureCompressionASTC_LDR = pdevice->info.gen >= 9, /* FINISHME CHV */
925 .textureCompressionBC = true,
926 .occlusionQueryPrecise = true,
927 .pipelineStatisticsQuery = true,
928 .fragmentStoresAndAtomics = true,
929 .shaderTessellationAndGeometryPointSize = true,
930 .shaderImageGatherExtended = true,
931 .shaderStorageImageExtendedFormats = true,
932 .shaderStorageImageMultisample = false,
933 .shaderStorageImageReadWithoutFormat = false,
934 .shaderStorageImageWriteWithoutFormat = true,
935 .shaderUniformBufferArrayDynamicIndexing = true,
936 .shaderSampledImageArrayDynamicIndexing = true,
937 .shaderStorageBufferArrayDynamicIndexing = true,
938 .shaderStorageImageArrayDynamicIndexing = true,
939 .shaderClipDistance = true,
940 .shaderCullDistance = true,
941 .shaderFloat64 = pdevice->info.gen >= 8 &&
942 pdevice->info.has_64bit_types,
943 .shaderInt64 = pdevice->info.gen >= 8 &&
944 pdevice->info.has_64bit_types,
945 .shaderInt16 = pdevice->info.gen >= 8,
946 .shaderResourceMinLod = pdevice->info.gen >= 9,
947 .variableMultisampleRate = true,
948 .inheritedQueries = true,
949 };
950
951 /* We can't do image stores in vec4 shaders */
952 pFeatures->vertexPipelineStoresAndAtomics =
953 pdevice->compiler->scalar_stage[MESA_SHADER_VERTEX] &&
954 pdevice->compiler->scalar_stage[MESA_SHADER_GEOMETRY];
955
956 struct anv_app_info *app_info = &pdevice->instance->app_info;
957
958 /* The new DOOM and Wolfenstein games require depthBounds without
959 * checking for it. They seem to run fine without it so just claim it's
960 * there and accept the consequences.
961 */
962 if (app_info->engine_name && strcmp(app_info->engine_name, "idTech") == 0)
963 pFeatures->depthBounds = true;
964 }
965
966 void anv_GetPhysicalDeviceFeatures2(
967 VkPhysicalDevice physicalDevice,
968 VkPhysicalDeviceFeatures2* pFeatures)
969 {
970 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
971 anv_GetPhysicalDeviceFeatures(physicalDevice, &pFeatures->features);
972
973 vk_foreach_struct(ext, pFeatures->pNext) {
974 switch (ext->sType) {
975 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES_KHR: {
976 VkPhysicalDevice8BitStorageFeaturesKHR *features =
977 (VkPhysicalDevice8BitStorageFeaturesKHR *)ext;
978 features->storageBuffer8BitAccess = pdevice->info.gen >= 8;
979 features->uniformAndStorageBuffer8BitAccess = pdevice->info.gen >= 8;
980 features->storagePushConstant8 = pdevice->info.gen >= 8;
981 break;
982 }
983
984 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES: {
985 VkPhysicalDevice16BitStorageFeatures *features =
986 (VkPhysicalDevice16BitStorageFeatures *)ext;
987 features->storageBuffer16BitAccess = pdevice->info.gen >= 8;
988 features->uniformAndStorageBuffer16BitAccess = pdevice->info.gen >= 8;
989 features->storagePushConstant16 = pdevice->info.gen >= 8;
990 features->storageInputOutput16 = false;
991 break;
992 }
993
994 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_EXT: {
995 VkPhysicalDeviceBufferDeviceAddressFeaturesEXT *features = (void *)ext;
996 features->bufferDeviceAddress = pdevice->has_a64_buffer_access;
997 features->bufferDeviceAddressCaptureReplay = false;
998 features->bufferDeviceAddressMultiDevice = false;
999 break;
1000 }
1001
1002 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_KHR: {
1003 VkPhysicalDeviceBufferDeviceAddressFeaturesKHR *features = (void *)ext;
1004 features->bufferDeviceAddress = pdevice->has_a64_buffer_access;
1005 features->bufferDeviceAddressCaptureReplay =
1006 pdevice->has_a64_buffer_access;
1007 features->bufferDeviceAddressMultiDevice = false;
1008 break;
1009 }
1010
1011 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COMPUTE_SHADER_DERIVATIVES_FEATURES_NV: {
1012 VkPhysicalDeviceComputeShaderDerivativesFeaturesNV *features =
1013 (VkPhysicalDeviceComputeShaderDerivativesFeaturesNV *)ext;
1014 features->computeDerivativeGroupQuads = true;
1015 features->computeDerivativeGroupLinear = true;
1016 break;
1017 }
1018
1019 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT: {
1020 VkPhysicalDeviceConditionalRenderingFeaturesEXT *features =
1021 (VkPhysicalDeviceConditionalRenderingFeaturesEXT*)ext;
1022 features->conditionalRendering = pdevice->info.gen >= 8 ||
1023 pdevice->info.is_haswell;
1024 features->inheritedConditionalRendering = pdevice->info.gen >= 8 ||
1025 pdevice->info.is_haswell;
1026 break;
1027 }
1028
1029 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_CLIP_ENABLE_FEATURES_EXT: {
1030 VkPhysicalDeviceDepthClipEnableFeaturesEXT *features =
1031 (VkPhysicalDeviceDepthClipEnableFeaturesEXT *)ext;
1032 features->depthClipEnable = true;
1033 break;
1034 }
1035
1036 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR: {
1037 VkPhysicalDeviceFloat16Int8FeaturesKHR *features = (void *)ext;
1038 features->shaderFloat16 = pdevice->info.gen >= 8;
1039 features->shaderInt8 = pdevice->info.gen >= 8;
1040 break;
1041 }
1042
1043 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FRAGMENT_SHADER_INTERLOCK_FEATURES_EXT: {
1044 VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT *features =
1045 (VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT *)ext;
1046 features->fragmentShaderSampleInterlock = pdevice->info.gen >= 9;
1047 features->fragmentShaderPixelInterlock = pdevice->info.gen >= 9;
1048 features->fragmentShaderShadingRateInterlock = false;
1049 break;
1050 }
1051
1052 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_QUERY_RESET_FEATURES_EXT: {
1053 VkPhysicalDeviceHostQueryResetFeaturesEXT *features =
1054 (VkPhysicalDeviceHostQueryResetFeaturesEXT *)ext;
1055 features->hostQueryReset = true;
1056 break;
1057 }
1058
1059 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT: {
1060 VkPhysicalDeviceDescriptorIndexingFeaturesEXT *features =
1061 (VkPhysicalDeviceDescriptorIndexingFeaturesEXT *)ext;
1062 bool descIndexing = pdevice->has_a64_buffer_access &&
1063 pdevice->has_bindless_images;
1064 features->shaderInputAttachmentArrayDynamicIndexing = false;
1065 features->shaderUniformTexelBufferArrayDynamicIndexing = descIndexing;
1066 features->shaderStorageTexelBufferArrayDynamicIndexing = descIndexing;
1067 features->shaderUniformBufferArrayNonUniformIndexing = false;
1068 features->shaderSampledImageArrayNonUniformIndexing = descIndexing;
1069 features->shaderStorageBufferArrayNonUniformIndexing = descIndexing;
1070 features->shaderStorageImageArrayNonUniformIndexing = descIndexing;
1071 features->shaderInputAttachmentArrayNonUniformIndexing = false;
1072 features->shaderUniformTexelBufferArrayNonUniformIndexing = descIndexing;
1073 features->shaderStorageTexelBufferArrayNonUniformIndexing = descIndexing;
1074 features->descriptorBindingUniformBufferUpdateAfterBind = false;
1075 features->descriptorBindingSampledImageUpdateAfterBind = descIndexing;
1076 features->descriptorBindingStorageImageUpdateAfterBind = descIndexing;
1077 features->descriptorBindingStorageBufferUpdateAfterBind = descIndexing;
1078 features->descriptorBindingUniformTexelBufferUpdateAfterBind = descIndexing;
1079 features->descriptorBindingStorageTexelBufferUpdateAfterBind = descIndexing;
1080 features->descriptorBindingUpdateUnusedWhilePending = descIndexing;
1081 features->descriptorBindingPartiallyBound = descIndexing;
1082 features->descriptorBindingVariableDescriptorCount = false;
1083 features->runtimeDescriptorArray = descIndexing;
1084 break;
1085 }
1086
1087 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INDEX_TYPE_UINT8_FEATURES_EXT: {
1088 VkPhysicalDeviceIndexTypeUint8FeaturesEXT *features =
1089 (VkPhysicalDeviceIndexTypeUint8FeaturesEXT *)ext;
1090 features->indexTypeUint8 = true;
1091 break;
1092 }
1093
1094 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_FEATURES_EXT: {
1095 VkPhysicalDeviceInlineUniformBlockFeaturesEXT *features =
1096 (VkPhysicalDeviceInlineUniformBlockFeaturesEXT *)ext;
1097 features->inlineUniformBlock = true;
1098 features->descriptorBindingInlineUniformBlockUpdateAfterBind = true;
1099 break;
1100 }
1101
1102 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_FEATURES_EXT: {
1103 VkPhysicalDeviceLineRasterizationFeaturesEXT *features =
1104 (VkPhysicalDeviceLineRasterizationFeaturesEXT *)ext;
1105 features->rectangularLines = true;
1106 features->bresenhamLines = true;
1107 /* Support for Smooth lines with MSAA was removed on gen11. From the
1108 * BSpec section "Multisample ModesState" table for "AA Line Support
1109 * Requirements":
1110 *
1111 * GEN10:BUG:######## NUM_MULTISAMPLES == 1
1112 *
1113 * Fortunately, this isn't a case most people care about.
1114 */
1115 features->smoothLines = pdevice->info.gen < 10;
1116 features->stippledRectangularLines = false;
1117 features->stippledBresenhamLines = true;
1118 features->stippledSmoothLines = false;
1119 break;
1120 }
1121
1122 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES: {
1123 VkPhysicalDeviceMultiviewFeatures *features =
1124 (VkPhysicalDeviceMultiviewFeatures *)ext;
1125 features->multiview = true;
1126 features->multiviewGeometryShader = true;
1127 features->multiviewTessellationShader = true;
1128 break;
1129 }
1130
1131 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGELESS_FRAMEBUFFER_FEATURES_KHR: {
1132 VkPhysicalDeviceImagelessFramebufferFeaturesKHR *features =
1133 (VkPhysicalDeviceImagelessFramebufferFeaturesKHR *)ext;
1134 features->imagelessFramebuffer = true;
1135 break;
1136 }
1137
1138 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR: {
1139 VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR *features =
1140 (VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR *)ext;
1141 features->pipelineExecutableInfo = true;
1142 break;
1143 }
1144
1145 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES: {
1146 VkPhysicalDeviceProtectedMemoryFeatures *features = (void *)ext;
1147 features->protectedMemory = false;
1148 break;
1149 }
1150
1151 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES: {
1152 VkPhysicalDeviceSamplerYcbcrConversionFeatures *features =
1153 (VkPhysicalDeviceSamplerYcbcrConversionFeatures *) ext;
1154 features->samplerYcbcrConversion = true;
1155 break;
1156 }
1157
1158 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES_EXT: {
1159 VkPhysicalDeviceScalarBlockLayoutFeaturesEXT *features =
1160 (VkPhysicalDeviceScalarBlockLayoutFeaturesEXT *)ext;
1161 features->scalarBlockLayout = true;
1162 break;
1163 }
1164
1165 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SEPARATE_DEPTH_STENCIL_LAYOUTS_FEATURES_KHR: {
1166 VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR *features =
1167 (VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR *)ext;
1168 features->separateDepthStencilLayouts = true;
1169 break;
1170 }
1171
1172 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_INT64_FEATURES_KHR: {
1173 VkPhysicalDeviceShaderAtomicInt64FeaturesKHR *features = (void *)ext;
1174 features->shaderBufferInt64Atomics =
1175 pdevice->info.gen >= 9 && pdevice->use_softpin;
1176 features->shaderSharedInt64Atomics = VK_FALSE;
1177 break;
1178 }
1179
1180 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DEMOTE_TO_HELPER_INVOCATION_FEATURES_EXT: {
1181 VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT *features = (void *)ext;
1182 features->shaderDemoteToHelperInvocation = true;
1183 break;
1184 }
1185
1186 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CLOCK_FEATURES_KHR: {
1187 VkPhysicalDeviceShaderClockFeaturesKHR *features =
1188 (VkPhysicalDeviceShaderClockFeaturesKHR *)ext;
1189 features->shaderSubgroupClock = true;
1190 features->shaderDeviceClock = false;
1191 break;
1192 }
1193
1194 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETERS_FEATURES: {
1195 VkPhysicalDeviceShaderDrawParametersFeatures *features = (void *)ext;
1196 features->shaderDrawParameters = true;
1197 break;
1198 }
1199
1200 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_SUBGROUP_EXTENDED_TYPES_FEATURES_KHR: {
1201 VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR *features =
1202 (VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR *)ext;
1203 features->shaderSubgroupExtendedTypes = true;
1204 break;
1205 }
1206
1207 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_FEATURES_EXT: {
1208 VkPhysicalDeviceSubgroupSizeControlFeaturesEXT *features =
1209 (VkPhysicalDeviceSubgroupSizeControlFeaturesEXT *)ext;
1210 features->subgroupSizeControl = true;
1211 features->computeFullSubgroups = true;
1212 break;
1213 }
1214
1215 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_FEATURES_EXT: {
1216 VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT *features =
1217 (VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT *)ext;
1218 features->texelBufferAlignment = true;
1219 break;
1220 }
1221
1222 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_FEATURES_KHR: {
1223 VkPhysicalDeviceTimelineSemaphoreFeaturesKHR *features =
1224 (VkPhysicalDeviceTimelineSemaphoreFeaturesKHR *) ext;
1225 features->timelineSemaphore = true;
1226 break;
1227 }
1228
1229 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES: {
1230 VkPhysicalDeviceVariablePointersFeatures *features = (void *)ext;
1231 features->variablePointersStorageBuffer = true;
1232 features->variablePointers = true;
1233 break;
1234 }
1235
1236 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT: {
1237 VkPhysicalDeviceTransformFeedbackFeaturesEXT *features =
1238 (VkPhysicalDeviceTransformFeedbackFeaturesEXT *)ext;
1239 features->transformFeedback = true;
1240 features->geometryStreams = true;
1241 break;
1242 }
1243
1244 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_UNIFORM_BUFFER_STANDARD_LAYOUT_FEATURES_KHR: {
1245 VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR *features =
1246 (VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR *)ext;
1247 features->uniformBufferStandardLayout = true;
1248 break;
1249 }
1250
1251 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT: {
1252 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *features =
1253 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *)ext;
1254 features->vertexAttributeInstanceRateDivisor = true;
1255 features->vertexAttributeInstanceRateZeroDivisor = true;
1256 break;
1257 }
1258
1259 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_MEMORY_MODEL_FEATURES_KHR: {
1260 VkPhysicalDeviceVulkanMemoryModelFeaturesKHR *features = (void *)ext;
1261 features->vulkanMemoryModel = true;
1262 features->vulkanMemoryModelDeviceScope = true;
1263 features->vulkanMemoryModelAvailabilityVisibilityChains = true;
1264 break;
1265 }
1266
1267 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_YCBCR_IMAGE_ARRAYS_FEATURES_EXT: {
1268 VkPhysicalDeviceYcbcrImageArraysFeaturesEXT *features =
1269 (VkPhysicalDeviceYcbcrImageArraysFeaturesEXT *)ext;
1270 features->ycbcrImageArrays = true;
1271 break;
1272 }
1273
1274 default:
1275 anv_debug_ignored_stype(ext->sType);
1276 break;
1277 }
1278 }
1279 }
1280
1281 #define MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS 64
1282
1283 #define MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS 64
1284 #define MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS 256
1285
1286 void anv_GetPhysicalDeviceProperties(
1287 VkPhysicalDevice physicalDevice,
1288 VkPhysicalDeviceProperties* pProperties)
1289 {
1290 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
1291 const struct gen_device_info *devinfo = &pdevice->info;
1292
1293 /* See assertions made when programming the buffer surface state. */
1294 const uint32_t max_raw_buffer_sz = devinfo->gen >= 7 ?
1295 (1ul << 30) : (1ul << 27);
1296
1297 const uint32_t max_ssbos = pdevice->has_a64_buffer_access ? UINT16_MAX : 64;
1298 const uint32_t max_textures =
1299 pdevice->has_bindless_images ? UINT16_MAX : 128;
1300 const uint32_t max_samplers =
1301 pdevice->has_bindless_samplers ? UINT16_MAX :
1302 (devinfo->gen >= 8 || devinfo->is_haswell) ? 128 : 16;
1303 const uint32_t max_images =
1304 pdevice->has_bindless_images ? UINT16_MAX : MAX_IMAGES;
1305
1306 /* If we can use bindless for everything, claim a high per-stage limit,
1307 * otherwise use the binding table size, minus the slots reserved for
1308 * render targets and one slot for the descriptor buffer. */
1309 const uint32_t max_per_stage =
1310 pdevice->has_bindless_images && pdevice->has_a64_buffer_access
1311 ? UINT32_MAX : MAX_BINDING_TABLE_SIZE - MAX_RTS - 1;
1312
1313 const uint32_t max_workgroup_size = 32 * devinfo->max_cs_threads;
1314
1315 VkSampleCountFlags sample_counts =
1316 isl_device_get_sample_counts(&pdevice->isl_dev);
1317
1318
1319 VkPhysicalDeviceLimits limits = {
1320 .maxImageDimension1D = (1 << 14),
1321 .maxImageDimension2D = (1 << 14),
1322 .maxImageDimension3D = (1 << 11),
1323 .maxImageDimensionCube = (1 << 14),
1324 .maxImageArrayLayers = (1 << 11),
1325 .maxTexelBufferElements = 128 * 1024 * 1024,
1326 .maxUniformBufferRange = (1ul << 27),
1327 .maxStorageBufferRange = max_raw_buffer_sz,
1328 .maxPushConstantsSize = MAX_PUSH_CONSTANTS_SIZE,
1329 .maxMemoryAllocationCount = UINT32_MAX,
1330 .maxSamplerAllocationCount = 64 * 1024,
1331 .bufferImageGranularity = 64, /* A cache line */
1332 .sparseAddressSpaceSize = 0,
1333 .maxBoundDescriptorSets = MAX_SETS,
1334 .maxPerStageDescriptorSamplers = max_samplers,
1335 .maxPerStageDescriptorUniformBuffers = MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS,
1336 .maxPerStageDescriptorStorageBuffers = max_ssbos,
1337 .maxPerStageDescriptorSampledImages = max_textures,
1338 .maxPerStageDescriptorStorageImages = max_images,
1339 .maxPerStageDescriptorInputAttachments = MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS,
1340 .maxPerStageResources = max_per_stage,
1341 .maxDescriptorSetSamplers = 6 * max_samplers, /* number of stages * maxPerStageDescriptorSamplers */
1342 .maxDescriptorSetUniformBuffers = 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS, /* number of stages * maxPerStageDescriptorUniformBuffers */
1343 .maxDescriptorSetUniformBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2,
1344 .maxDescriptorSetStorageBuffers = 6 * max_ssbos, /* number of stages * maxPerStageDescriptorStorageBuffers */
1345 .maxDescriptorSetStorageBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2,
1346 .maxDescriptorSetSampledImages = 6 * max_textures, /* number of stages * maxPerStageDescriptorSampledImages */
1347 .maxDescriptorSetStorageImages = 6 * max_images, /* number of stages * maxPerStageDescriptorStorageImages */
1348 .maxDescriptorSetInputAttachments = MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS,
1349 .maxVertexInputAttributes = MAX_VBS,
1350 .maxVertexInputBindings = MAX_VBS,
1351 .maxVertexInputAttributeOffset = 2047,
1352 .maxVertexInputBindingStride = 2048,
1353 .maxVertexOutputComponents = 128,
1354 .maxTessellationGenerationLevel = 64,
1355 .maxTessellationPatchSize = 32,
1356 .maxTessellationControlPerVertexInputComponents = 128,
1357 .maxTessellationControlPerVertexOutputComponents = 128,
1358 .maxTessellationControlPerPatchOutputComponents = 128,
1359 .maxTessellationControlTotalOutputComponents = 2048,
1360 .maxTessellationEvaluationInputComponents = 128,
1361 .maxTessellationEvaluationOutputComponents = 128,
1362 .maxGeometryShaderInvocations = 32,
1363 .maxGeometryInputComponents = 64,
1364 .maxGeometryOutputComponents = 128,
1365 .maxGeometryOutputVertices = 256,
1366 .maxGeometryTotalOutputComponents = 1024,
1367 .maxFragmentInputComponents = 116, /* 128 components - (PSIZ, CLIP_DIST0, CLIP_DIST1) */
1368 .maxFragmentOutputAttachments = 8,
1369 .maxFragmentDualSrcAttachments = 1,
1370 .maxFragmentCombinedOutputResources = 8,
1371 .maxComputeSharedMemorySize = 64 * 1024,
1372 .maxComputeWorkGroupCount = { 65535, 65535, 65535 },
1373 .maxComputeWorkGroupInvocations = max_workgroup_size,
1374 .maxComputeWorkGroupSize = {
1375 max_workgroup_size,
1376 max_workgroup_size,
1377 max_workgroup_size,
1378 },
1379 .subPixelPrecisionBits = 8,
1380 .subTexelPrecisionBits = 8,
1381 .mipmapPrecisionBits = 8,
1382 .maxDrawIndexedIndexValue = UINT32_MAX,
1383 .maxDrawIndirectCount = UINT32_MAX,
1384 .maxSamplerLodBias = 16,
1385 .maxSamplerAnisotropy = 16,
1386 .maxViewports = MAX_VIEWPORTS,
1387 .maxViewportDimensions = { (1 << 14), (1 << 14) },
1388 .viewportBoundsRange = { INT16_MIN, INT16_MAX },
1389 .viewportSubPixelBits = 13, /* We take a float? */
1390 .minMemoryMapAlignment = 4096, /* A page */
1391 /* The dataport requires texel alignment so we need to assume a worst
1392 * case of R32G32B32A32 which is 16 bytes.
1393 */
1394 .minTexelBufferOffsetAlignment = 16,
1395 /* We need 16 for UBO block reads to work and 32 for push UBOs */
1396 .minUniformBufferOffsetAlignment = 32,
1397 .minStorageBufferOffsetAlignment = 4,
1398 .minTexelOffset = -8,
1399 .maxTexelOffset = 7,
1400 .minTexelGatherOffset = -32,
1401 .maxTexelGatherOffset = 31,
1402 .minInterpolationOffset = -0.5,
1403 .maxInterpolationOffset = 0.4375,
1404 .subPixelInterpolationOffsetBits = 4,
1405 .maxFramebufferWidth = (1 << 14),
1406 .maxFramebufferHeight = (1 << 14),
1407 .maxFramebufferLayers = (1 << 11),
1408 .framebufferColorSampleCounts = sample_counts,
1409 .framebufferDepthSampleCounts = sample_counts,
1410 .framebufferStencilSampleCounts = sample_counts,
1411 .framebufferNoAttachmentsSampleCounts = sample_counts,
1412 .maxColorAttachments = MAX_RTS,
1413 .sampledImageColorSampleCounts = sample_counts,
1414 .sampledImageIntegerSampleCounts = sample_counts,
1415 .sampledImageDepthSampleCounts = sample_counts,
1416 .sampledImageStencilSampleCounts = sample_counts,
1417 .storageImageSampleCounts = VK_SAMPLE_COUNT_1_BIT,
1418 .maxSampleMaskWords = 1,
1419 .timestampComputeAndGraphics = true,
1420 .timestampPeriod = 1000000000.0 / devinfo->timestamp_frequency,
1421 .maxClipDistances = 8,
1422 .maxCullDistances = 8,
1423 .maxCombinedClipAndCullDistances = 8,
1424 .discreteQueuePriorities = 2,
1425 .pointSizeRange = { 0.125, 255.875 },
1426 .lineWidthRange = {
1427 0.0,
1428 (devinfo->gen >= 9 || devinfo->is_cherryview) ?
1429 2047.9921875 : 7.9921875,
1430 },
1431 .pointSizeGranularity = (1.0 / 8.0),
1432 .lineWidthGranularity = (1.0 / 128.0),
1433 .strictLines = false,
1434 .standardSampleLocations = true,
1435 .optimalBufferCopyOffsetAlignment = 128,
1436 .optimalBufferCopyRowPitchAlignment = 128,
1437 .nonCoherentAtomSize = 64,
1438 };
1439
1440 *pProperties = (VkPhysicalDeviceProperties) {
1441 .apiVersion = anv_physical_device_api_version(pdevice),
1442 .driverVersion = vk_get_driver_version(),
1443 .vendorID = 0x8086,
1444 .deviceID = pdevice->chipset_id,
1445 .deviceType = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
1446 .limits = limits,
1447 .sparseProperties = {0}, /* Broadwell doesn't do sparse. */
1448 };
1449
1450 snprintf(pProperties->deviceName, sizeof(pProperties->deviceName),
1451 "%s", pdevice->name);
1452 memcpy(pProperties->pipelineCacheUUID,
1453 pdevice->pipeline_cache_uuid, VK_UUID_SIZE);
1454 }
1455
1456 void anv_GetPhysicalDeviceProperties2(
1457 VkPhysicalDevice physicalDevice,
1458 VkPhysicalDeviceProperties2* pProperties)
1459 {
1460 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
1461
1462 anv_GetPhysicalDeviceProperties(physicalDevice, &pProperties->properties);
1463
1464 vk_foreach_struct(ext, pProperties->pNext) {
1465 switch (ext->sType) {
1466 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_STENCIL_RESOLVE_PROPERTIES_KHR: {
1467 VkPhysicalDeviceDepthStencilResolvePropertiesKHR *props =
1468 (VkPhysicalDeviceDepthStencilResolvePropertiesKHR *)ext;
1469
1470 /* We support all of the depth resolve modes */
1471 props->supportedDepthResolveModes =
1472 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR |
1473 VK_RESOLVE_MODE_AVERAGE_BIT_KHR |
1474 VK_RESOLVE_MODE_MIN_BIT_KHR |
1475 VK_RESOLVE_MODE_MAX_BIT_KHR;
1476
1477 /* Average doesn't make sense for stencil so we don't support that */
1478 props->supportedStencilResolveModes =
1479 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR;
1480 if (pdevice->info.gen >= 8) {
1481 /* The advanced stencil resolve modes currently require stencil
1482 * sampling be supported by the hardware.
1483 */
1484 props->supportedStencilResolveModes |=
1485 VK_RESOLVE_MODE_MIN_BIT_KHR |
1486 VK_RESOLVE_MODE_MAX_BIT_KHR;
1487 }
1488
1489 props->independentResolveNone = true;
1490 props->independentResolve = true;
1491 break;
1492 }
1493
1494 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES_EXT: {
1495 VkPhysicalDeviceDescriptorIndexingPropertiesEXT *props =
1496 (VkPhysicalDeviceDescriptorIndexingPropertiesEXT *)ext;
1497
1498 /* It's a bit hard to exactly map our implementation to the limits
1499 * described here. The bindless surface handle in the extended
1500 * message descriptors is 20 bits and it's an index into the table of
1501 * RENDER_SURFACE_STATE structs that starts at bindless surface base
1502 * address. Given that most things consume two surface states per
1503 * view (general/sampled for textures and write-only/read-write for
1504 * images), we claim 2^19 things.
1505 *
1506 * For SSBOs, we just use A64 messages so there is no real limit
1507 * there beyond the limit on the total size of a descriptor set.
1508 */
1509 const unsigned max_bindless_views = 1 << 19;
1510
1511 props->maxUpdateAfterBindDescriptorsInAllPools = max_bindless_views;
1512 props->shaderUniformBufferArrayNonUniformIndexingNative = false;
1513 props->shaderSampledImageArrayNonUniformIndexingNative = false;
1514 props->shaderStorageBufferArrayNonUniformIndexingNative = true;
1515 props->shaderStorageImageArrayNonUniformIndexingNative = false;
1516 props->shaderInputAttachmentArrayNonUniformIndexingNative = false;
1517 props->robustBufferAccessUpdateAfterBind = true;
1518 props->quadDivergentImplicitLod = false;
1519 props->maxPerStageDescriptorUpdateAfterBindSamplers = max_bindless_views;
1520 props->maxPerStageDescriptorUpdateAfterBindUniformBuffers = MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS;
1521 props->maxPerStageDescriptorUpdateAfterBindStorageBuffers = UINT32_MAX;
1522 props->maxPerStageDescriptorUpdateAfterBindSampledImages = max_bindless_views;
1523 props->maxPerStageDescriptorUpdateAfterBindStorageImages = max_bindless_views;
1524 props->maxPerStageDescriptorUpdateAfterBindInputAttachments = MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS;
1525 props->maxPerStageUpdateAfterBindResources = UINT32_MAX;
1526 props->maxDescriptorSetUpdateAfterBindSamplers = max_bindless_views;
1527 props->maxDescriptorSetUpdateAfterBindUniformBuffers = 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS;
1528 props->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2;
1529 props->maxDescriptorSetUpdateAfterBindStorageBuffers = UINT32_MAX;
1530 props->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2;
1531 props->maxDescriptorSetUpdateAfterBindSampledImages = max_bindless_views;
1532 props->maxDescriptorSetUpdateAfterBindStorageImages = max_bindless_views;
1533 props->maxDescriptorSetUpdateAfterBindInputAttachments = MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS;
1534 break;
1535 }
1536
1537 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR: {
1538 VkPhysicalDeviceDriverPropertiesKHR *driver_props =
1539 (VkPhysicalDeviceDriverPropertiesKHR *) ext;
1540
1541 driver_props->driverID = VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA_KHR;
1542 memset(driver_props->driverName, 0,
1543 sizeof(driver_props->driverName));
1544 snprintf(driver_props->driverName, VK_MAX_DRIVER_NAME_SIZE_KHR,
1545 "Intel open-source Mesa driver");
1546
1547 memset(driver_props->driverInfo, 0,
1548 sizeof(driver_props->driverInfo));
1549 snprintf(driver_props->driverInfo, VK_MAX_DRIVER_INFO_SIZE_KHR,
1550 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1);
1551
1552 driver_props->conformanceVersion = (VkConformanceVersionKHR) {
1553 .major = 1,
1554 .minor = 1,
1555 .subminor = 2,
1556 .patch = 0,
1557 };
1558 break;
1559 }
1560
1561 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT: {
1562 VkPhysicalDeviceExternalMemoryHostPropertiesEXT *props =
1563 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT *) ext;
1564 /* Userptr needs page aligned memory. */
1565 props->minImportedHostPointerAlignment = 4096;
1566 break;
1567 }
1568
1569 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES: {
1570 VkPhysicalDeviceIDProperties *id_props =
1571 (VkPhysicalDeviceIDProperties *)ext;
1572 memcpy(id_props->deviceUUID, pdevice->device_uuid, VK_UUID_SIZE);
1573 memcpy(id_props->driverUUID, pdevice->driver_uuid, VK_UUID_SIZE);
1574 /* The LUID is for Windows. */
1575 memset(id_props->deviceLUID, 0, VK_UUID_SIZE);
1576 id_props->deviceLUIDValid = false;
1577 break;
1578 }
1579
1580 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_PROPERTIES_EXT: {
1581 VkPhysicalDeviceInlineUniformBlockPropertiesEXT *props =
1582 (VkPhysicalDeviceInlineUniformBlockPropertiesEXT *)ext;
1583 props->maxInlineUniformBlockSize = MAX_INLINE_UNIFORM_BLOCK_SIZE;
1584 props->maxPerStageDescriptorInlineUniformBlocks =
1585 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1586 props->maxPerStageDescriptorUpdateAfterBindInlineUniformBlocks =
1587 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1588 props->maxDescriptorSetInlineUniformBlocks =
1589 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1590 props->maxDescriptorSetUpdateAfterBindInlineUniformBlocks =
1591 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1592 break;
1593 }
1594
1595 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_PROPERTIES_EXT: {
1596 VkPhysicalDeviceLineRasterizationPropertiesEXT *props =
1597 (VkPhysicalDeviceLineRasterizationPropertiesEXT *)ext;
1598 /* In the Skylake PRM Vol. 7, subsection titled "GIQ (Diamond)
1599 * Sampling Rules - Legacy Mode", it says the following:
1600 *
1601 * "Note that the device divides a pixel into a 16x16 array of
1602 * subpixels, referenced by their upper left corners."
1603 *
1604 * This is the only known reference in the PRMs to the subpixel
1605 * precision of line rasterization and a "16x16 array of subpixels"
1606 * implies 4 subpixel precision bits. Empirical testing has shown
1607 * that 4 subpixel precision bits applies to all line rasterization
1608 * types.
1609 */
1610 props->lineSubPixelPrecisionBits = 4;
1611 break;
1612 }
1613
1614 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES: {
1615 VkPhysicalDeviceMaintenance3Properties *props =
1616 (VkPhysicalDeviceMaintenance3Properties *)ext;
1617 /* This value doesn't matter for us today as our per-stage
1618 * descriptors are the real limit.
1619 */
1620 props->maxPerSetDescriptors = 1024;
1621 props->maxMemoryAllocationSize = MAX_MEMORY_ALLOCATION_SIZE;
1622 break;
1623 }
1624
1625 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES: {
1626 VkPhysicalDeviceMultiviewProperties *properties =
1627 (VkPhysicalDeviceMultiviewProperties *)ext;
1628 properties->maxMultiviewViewCount = 16;
1629 properties->maxMultiviewInstanceIndex = UINT32_MAX / 16;
1630 break;
1631 }
1632
1633 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT: {
1634 VkPhysicalDevicePCIBusInfoPropertiesEXT *properties =
1635 (VkPhysicalDevicePCIBusInfoPropertiesEXT *)ext;
1636 properties->pciDomain = pdevice->pci_info.domain;
1637 properties->pciBus = pdevice->pci_info.bus;
1638 properties->pciDevice = pdevice->pci_info.device;
1639 properties->pciFunction = pdevice->pci_info.function;
1640 break;
1641 }
1642
1643 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES: {
1644 VkPhysicalDevicePointClippingProperties *properties =
1645 (VkPhysicalDevicePointClippingProperties *) ext;
1646 properties->pointClippingBehavior = VK_POINT_CLIPPING_BEHAVIOR_USER_CLIP_PLANES_ONLY;
1647 break;
1648 }
1649
1650 #pragma GCC diagnostic push
1651 #pragma GCC diagnostic ignored "-Wswitch"
1652 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PRESENTATION_PROPERTIES_ANDROID: {
1653 VkPhysicalDevicePresentationPropertiesANDROID *props =
1654 (VkPhysicalDevicePresentationPropertiesANDROID *)ext;
1655 props->sharedImage = VK_FALSE;
1656 break;
1657 }
1658 #pragma GCC diagnostic pop
1659
1660 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES: {
1661 VkPhysicalDeviceProtectedMemoryProperties *props =
1662 (VkPhysicalDeviceProtectedMemoryProperties *)ext;
1663 props->protectedNoFault = false;
1664 break;
1665 }
1666
1667 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR: {
1668 VkPhysicalDevicePushDescriptorPropertiesKHR *properties =
1669 (VkPhysicalDevicePushDescriptorPropertiesKHR *) ext;
1670
1671 properties->maxPushDescriptors = MAX_PUSH_DESCRIPTORS;
1672 break;
1673 }
1674
1675 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT: {
1676 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT *properties =
1677 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT *)ext;
1678 properties->filterMinmaxImageComponentMapping = pdevice->info.gen >= 9;
1679 properties->filterMinmaxSingleComponentFormats = pdevice->info.gen >= 9;
1680 break;
1681 }
1682
1683 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES: {
1684 VkPhysicalDeviceSubgroupProperties *properties = (void *)ext;
1685
1686 properties->subgroupSize = BRW_SUBGROUP_SIZE;
1687
1688 VkShaderStageFlags scalar_stages = 0;
1689 for (unsigned stage = 0; stage < MESA_SHADER_STAGES; stage++) {
1690 if (pdevice->compiler->scalar_stage[stage])
1691 scalar_stages |= mesa_to_vk_shader_stage(stage);
1692 }
1693 properties->supportedStages = scalar_stages;
1694
1695 properties->supportedOperations = VK_SUBGROUP_FEATURE_BASIC_BIT |
1696 VK_SUBGROUP_FEATURE_VOTE_BIT |
1697 VK_SUBGROUP_FEATURE_BALLOT_BIT |
1698 VK_SUBGROUP_FEATURE_SHUFFLE_BIT |
1699 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT |
1700 VK_SUBGROUP_FEATURE_QUAD_BIT;
1701 if (pdevice->info.gen >= 8) {
1702 /* TODO: There's no technical reason why these can't be made to
1703 * work on gen7 but they don't at the moment so it's best to leave
1704 * the feature disabled than enabled and broken.
1705 */
1706 properties->supportedOperations |=
1707 VK_SUBGROUP_FEATURE_ARITHMETIC_BIT |
1708 VK_SUBGROUP_FEATURE_CLUSTERED_BIT;
1709 }
1710 properties->quadOperationsInAllStages = pdevice->info.gen >= 8;
1711 break;
1712 }
1713
1714 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_PROPERTIES_EXT: {
1715 VkPhysicalDeviceSubgroupSizeControlPropertiesEXT *props =
1716 (VkPhysicalDeviceSubgroupSizeControlPropertiesEXT *)ext;
1717 STATIC_ASSERT(8 <= BRW_SUBGROUP_SIZE && BRW_SUBGROUP_SIZE <= 32);
1718 props->minSubgroupSize = 8;
1719 props->maxSubgroupSize = 32;
1720 props->maxComputeWorkgroupSubgroups = pdevice->info.max_cs_threads;
1721 props->requiredSubgroupSizeStages = VK_SHADER_STAGE_COMPUTE_BIT;
1722 break;
1723 }
1724 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT_CONTROLS_PROPERTIES_KHR : {
1725 VkPhysicalDeviceFloatControlsPropertiesKHR *properties = (void *)ext;
1726 properties->denormBehaviorIndependence = VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_ALL_KHR;
1727 properties->roundingModeIndependence = VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_NONE_KHR;
1728
1729 /* Broadwell does not support HF denorms and there are restrictions
1730 * other gens. According to Kabylake's PRM:
1731 *
1732 * "math - Extended Math Function
1733 * [...]
1734 * Restriction : Half-float denorms are always retained."
1735 */
1736 properties->shaderDenormFlushToZeroFloat16 = false;
1737 properties->shaderDenormPreserveFloat16 = pdevice->info.gen > 8;
1738 properties->shaderRoundingModeRTEFloat16 = true;
1739 properties->shaderRoundingModeRTZFloat16 = true;
1740 properties->shaderSignedZeroInfNanPreserveFloat16 = true;
1741
1742 properties->shaderDenormFlushToZeroFloat32 = true;
1743 properties->shaderDenormPreserveFloat32 = true;
1744 properties->shaderRoundingModeRTEFloat32 = true;
1745 properties->shaderRoundingModeRTZFloat32 = true;
1746 properties->shaderSignedZeroInfNanPreserveFloat32 = true;
1747
1748 properties->shaderDenormFlushToZeroFloat64 = true;
1749 properties->shaderDenormPreserveFloat64 = true;
1750 properties->shaderRoundingModeRTEFloat64 = true;
1751 properties->shaderRoundingModeRTZFloat64 = true;
1752 properties->shaderSignedZeroInfNanPreserveFloat64 = true;
1753 break;
1754 }
1755
1756 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_PROPERTIES_EXT: {
1757 VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT *props =
1758 (VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT *)ext;
1759
1760 /* From the SKL PRM Vol. 2d, docs for RENDER_SURFACE_STATE::Surface
1761 * Base Address:
1762 *
1763 * "For SURFTYPE_BUFFER non-rendertarget surfaces, this field
1764 * specifies the base address of the first element of the surface,
1765 * computed in software by adding the surface base address to the
1766 * byte offset of the element in the buffer. The base address must
1767 * be aligned to element size."
1768 *
1769 * The typed dataport messages require that things be texel aligned.
1770 * Otherwise, we may just load/store the wrong data or, in the worst
1771 * case, there may be hangs.
1772 */
1773 props->storageTexelBufferOffsetAlignmentBytes = 16;
1774 props->storageTexelBufferOffsetSingleTexelAlignment = true;
1775
1776 /* The sampler, however, is much more forgiving and it can handle
1777 * arbitrary byte alignment for linear and buffer surfaces. It's
1778 * hard to find a good PRM citation for this but years of empirical
1779 * experience demonstrate that this is true.
1780 */
1781 props->uniformTexelBufferOffsetAlignmentBytes = 1;
1782 props->uniformTexelBufferOffsetSingleTexelAlignment = false;
1783 break;
1784 }
1785
1786 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_PROPERTIES_KHR: {
1787 VkPhysicalDeviceTimelineSemaphorePropertiesKHR *props =
1788 (VkPhysicalDeviceTimelineSemaphorePropertiesKHR *) ext;
1789 props->maxTimelineSemaphoreValueDifference = UINT64_MAX;
1790 break;
1791 }
1792
1793 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT: {
1794 VkPhysicalDeviceTransformFeedbackPropertiesEXT *props =
1795 (VkPhysicalDeviceTransformFeedbackPropertiesEXT *)ext;
1796
1797 props->maxTransformFeedbackStreams = MAX_XFB_STREAMS;
1798 props->maxTransformFeedbackBuffers = MAX_XFB_BUFFERS;
1799 props->maxTransformFeedbackBufferSize = (1ull << 32);
1800 props->maxTransformFeedbackStreamDataSize = 128 * 4;
1801 props->maxTransformFeedbackBufferDataSize = 128 * 4;
1802 props->maxTransformFeedbackBufferDataStride = 2048;
1803 props->transformFeedbackQueries = true;
1804 props->transformFeedbackStreamsLinesTriangles = false;
1805 props->transformFeedbackRasterizationStreamSelect = false;
1806 props->transformFeedbackDraw = true;
1807 break;
1808 }
1809
1810 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT: {
1811 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *props =
1812 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *)ext;
1813 /* We have to restrict this a bit for multiview */
1814 props->maxVertexAttribDivisor = UINT32_MAX / 16;
1815 break;
1816 }
1817
1818 default:
1819 anv_debug_ignored_stype(ext->sType);
1820 break;
1821 }
1822 }
1823 }
1824
1825 /* We support exactly one queue family. */
1826 static const VkQueueFamilyProperties
1827 anv_queue_family_properties = {
1828 .queueFlags = VK_QUEUE_GRAPHICS_BIT |
1829 VK_QUEUE_COMPUTE_BIT |
1830 VK_QUEUE_TRANSFER_BIT,
1831 .queueCount = 1,
1832 .timestampValidBits = 36, /* XXX: Real value here */
1833 .minImageTransferGranularity = { 1, 1, 1 },
1834 };
1835
1836 void anv_GetPhysicalDeviceQueueFamilyProperties(
1837 VkPhysicalDevice physicalDevice,
1838 uint32_t* pCount,
1839 VkQueueFamilyProperties* pQueueFamilyProperties)
1840 {
1841 VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pCount);
1842
1843 vk_outarray_append(&out, p) {
1844 *p = anv_queue_family_properties;
1845 }
1846 }
1847
1848 void anv_GetPhysicalDeviceQueueFamilyProperties2(
1849 VkPhysicalDevice physicalDevice,
1850 uint32_t* pQueueFamilyPropertyCount,
1851 VkQueueFamilyProperties2* pQueueFamilyProperties)
1852 {
1853
1854 VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pQueueFamilyPropertyCount);
1855
1856 vk_outarray_append(&out, p) {
1857 p->queueFamilyProperties = anv_queue_family_properties;
1858
1859 vk_foreach_struct(s, p->pNext) {
1860 anv_debug_ignored_stype(s->sType);
1861 }
1862 }
1863 }
1864
1865 void anv_GetPhysicalDeviceMemoryProperties(
1866 VkPhysicalDevice physicalDevice,
1867 VkPhysicalDeviceMemoryProperties* pMemoryProperties)
1868 {
1869 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
1870
1871 pMemoryProperties->memoryTypeCount = physical_device->memory.type_count;
1872 for (uint32_t i = 0; i < physical_device->memory.type_count; i++) {
1873 pMemoryProperties->memoryTypes[i] = (VkMemoryType) {
1874 .propertyFlags = physical_device->memory.types[i].propertyFlags,
1875 .heapIndex = physical_device->memory.types[i].heapIndex,
1876 };
1877 }
1878
1879 pMemoryProperties->memoryHeapCount = physical_device->memory.heap_count;
1880 for (uint32_t i = 0; i < physical_device->memory.heap_count; i++) {
1881 pMemoryProperties->memoryHeaps[i] = (VkMemoryHeap) {
1882 .size = physical_device->memory.heaps[i].size,
1883 .flags = physical_device->memory.heaps[i].flags,
1884 };
1885 }
1886 }
1887
1888 static void
1889 anv_get_memory_budget(VkPhysicalDevice physicalDevice,
1890 VkPhysicalDeviceMemoryBudgetPropertiesEXT *memoryBudget)
1891 {
1892 ANV_FROM_HANDLE(anv_physical_device, device, physicalDevice);
1893 uint64_t sys_available = get_available_system_memory();
1894 assert(sys_available > 0);
1895
1896 VkDeviceSize total_heaps_size = 0;
1897 for (size_t i = 0; i < device->memory.heap_count; i++)
1898 total_heaps_size += device->memory.heaps[i].size;
1899
1900 for (size_t i = 0; i < device->memory.heap_count; i++) {
1901 VkDeviceSize heap_size = device->memory.heaps[i].size;
1902 VkDeviceSize heap_used = device->memory.heaps[i].used;
1903 VkDeviceSize heap_budget;
1904
1905 double heap_proportion = (double) heap_size / total_heaps_size;
1906 VkDeviceSize sys_available_prop = sys_available * heap_proportion;
1907
1908 /*
1909 * Let's not incite the app to starve the system: report at most 90% of
1910 * available system memory.
1911 */
1912 uint64_t heap_available = sys_available_prop * 9 / 10;
1913 heap_budget = MIN2(heap_size, heap_used + heap_available);
1914
1915 /*
1916 * Round down to the nearest MB
1917 */
1918 heap_budget &= ~((1ull << 20) - 1);
1919
1920 /*
1921 * The heapBudget value must be non-zero for array elements less than
1922 * VkPhysicalDeviceMemoryProperties::memoryHeapCount. The heapBudget
1923 * value must be less than or equal to VkMemoryHeap::size for each heap.
1924 */
1925 assert(0 < heap_budget && heap_budget <= heap_size);
1926
1927 memoryBudget->heapUsage[i] = heap_used;
1928 memoryBudget->heapBudget[i] = heap_budget;
1929 }
1930
1931 /* The heapBudget and heapUsage values must be zero for array elements
1932 * greater than or equal to VkPhysicalDeviceMemoryProperties::memoryHeapCount
1933 */
1934 for (uint32_t i = device->memory.heap_count; i < VK_MAX_MEMORY_HEAPS; i++) {
1935 memoryBudget->heapBudget[i] = 0;
1936 memoryBudget->heapUsage[i] = 0;
1937 }
1938 }
1939
1940 void anv_GetPhysicalDeviceMemoryProperties2(
1941 VkPhysicalDevice physicalDevice,
1942 VkPhysicalDeviceMemoryProperties2* pMemoryProperties)
1943 {
1944 anv_GetPhysicalDeviceMemoryProperties(physicalDevice,
1945 &pMemoryProperties->memoryProperties);
1946
1947 vk_foreach_struct(ext, pMemoryProperties->pNext) {
1948 switch (ext->sType) {
1949 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT:
1950 anv_get_memory_budget(physicalDevice, (void*)ext);
1951 break;
1952 default:
1953 anv_debug_ignored_stype(ext->sType);
1954 break;
1955 }
1956 }
1957 }
1958
1959 void
1960 anv_GetDeviceGroupPeerMemoryFeatures(
1961 VkDevice device,
1962 uint32_t heapIndex,
1963 uint32_t localDeviceIndex,
1964 uint32_t remoteDeviceIndex,
1965 VkPeerMemoryFeatureFlags* pPeerMemoryFeatures)
1966 {
1967 assert(localDeviceIndex == 0 && remoteDeviceIndex == 0);
1968 *pPeerMemoryFeatures = VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT |
1969 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT |
1970 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT |
1971 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT;
1972 }
1973
1974 PFN_vkVoidFunction anv_GetInstanceProcAddr(
1975 VkInstance _instance,
1976 const char* pName)
1977 {
1978 ANV_FROM_HANDLE(anv_instance, instance, _instance);
1979
1980 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
1981 * when we have to return valid function pointers, NULL, or it's left
1982 * undefined. See the table for exact details.
1983 */
1984 if (pName == NULL)
1985 return NULL;
1986
1987 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
1988 if (strcmp(pName, "vk" #entrypoint) == 0) \
1989 return (PFN_vkVoidFunction)anv_##entrypoint
1990
1991 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties);
1992 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties);
1993 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceVersion);
1994 LOOKUP_ANV_ENTRYPOINT(CreateInstance);
1995
1996 #undef LOOKUP_ANV_ENTRYPOINT
1997
1998 if (instance == NULL)
1999 return NULL;
2000
2001 int idx = anv_get_instance_entrypoint_index(pName);
2002 if (idx >= 0)
2003 return instance->dispatch.entrypoints[idx];
2004
2005 idx = anv_get_physical_device_entrypoint_index(pName);
2006 if (idx >= 0)
2007 return instance->physicalDevice.dispatch.entrypoints[idx];
2008
2009 idx = anv_get_device_entrypoint_index(pName);
2010 if (idx >= 0)
2011 return instance->device_dispatch.entrypoints[idx];
2012
2013 return NULL;
2014 }
2015
2016 /* With version 1+ of the loader interface the ICD should expose
2017 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
2018 */
2019 PUBLIC
2020 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
2021 VkInstance instance,
2022 const char* pName);
2023
2024 PUBLIC
2025 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
2026 VkInstance instance,
2027 const char* pName)
2028 {
2029 return anv_GetInstanceProcAddr(instance, pName);
2030 }
2031
2032 PFN_vkVoidFunction anv_GetDeviceProcAddr(
2033 VkDevice _device,
2034 const char* pName)
2035 {
2036 ANV_FROM_HANDLE(anv_device, device, _device);
2037
2038 if (!device || !pName)
2039 return NULL;
2040
2041 int idx = anv_get_device_entrypoint_index(pName);
2042 if (idx < 0)
2043 return NULL;
2044
2045 return device->dispatch.entrypoints[idx];
2046 }
2047
2048 /* With version 4+ of the loader interface the ICD should expose
2049 * vk_icdGetPhysicalDeviceProcAddr()
2050 */
2051 PUBLIC
2052 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetPhysicalDeviceProcAddr(
2053 VkInstance _instance,
2054 const char* pName);
2055
2056 PFN_vkVoidFunction vk_icdGetPhysicalDeviceProcAddr(
2057 VkInstance _instance,
2058 const char* pName)
2059 {
2060 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2061
2062 if (!pName || !instance)
2063 return NULL;
2064
2065 int idx = anv_get_physical_device_entrypoint_index(pName);
2066 if (idx < 0)
2067 return NULL;
2068
2069 return instance->physicalDevice.dispatch.entrypoints[idx];
2070 }
2071
2072
2073 VkResult
2074 anv_CreateDebugReportCallbackEXT(VkInstance _instance,
2075 const VkDebugReportCallbackCreateInfoEXT* pCreateInfo,
2076 const VkAllocationCallbacks* pAllocator,
2077 VkDebugReportCallbackEXT* pCallback)
2078 {
2079 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2080 return vk_create_debug_report_callback(&instance->debug_report_callbacks,
2081 pCreateInfo, pAllocator, &instance->alloc,
2082 pCallback);
2083 }
2084
2085 void
2086 anv_DestroyDebugReportCallbackEXT(VkInstance _instance,
2087 VkDebugReportCallbackEXT _callback,
2088 const VkAllocationCallbacks* pAllocator)
2089 {
2090 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2091 vk_destroy_debug_report_callback(&instance->debug_report_callbacks,
2092 _callback, pAllocator, &instance->alloc);
2093 }
2094
2095 void
2096 anv_DebugReportMessageEXT(VkInstance _instance,
2097 VkDebugReportFlagsEXT flags,
2098 VkDebugReportObjectTypeEXT objectType,
2099 uint64_t object,
2100 size_t location,
2101 int32_t messageCode,
2102 const char* pLayerPrefix,
2103 const char* pMessage)
2104 {
2105 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2106 vk_debug_report(&instance->debug_report_callbacks, flags, objectType,
2107 object, location, messageCode, pLayerPrefix, pMessage);
2108 }
2109
2110 static struct anv_state
2111 anv_state_pool_emit_data(struct anv_state_pool *pool, size_t size, size_t align, const void *p)
2112 {
2113 struct anv_state state;
2114
2115 state = anv_state_pool_alloc(pool, size, align);
2116 memcpy(state.map, p, size);
2117
2118 return state;
2119 }
2120
2121 /* Haswell border color is a bit of a disaster. Float and unorm formats use a
2122 * straightforward 32-bit float color in the first 64 bytes. Instead of using
2123 * a nice float/integer union like Gen8+, Haswell specifies the integer border
2124 * color as a separate entry /after/ the float color. The layout of this entry
2125 * also depends on the format's bpp (with extra hacks for RG32), and overlaps.
2126 *
2127 * Since we don't know the format/bpp, we can't make any of the border colors
2128 * containing '1' work for all formats, as it would be in the wrong place for
2129 * some of them. We opt to make 32-bit integers work as this seems like the
2130 * most common option. Fortunately, transparent black works regardless, as
2131 * all zeroes is the same in every bit-size.
2132 */
2133 struct hsw_border_color {
2134 float float32[4];
2135 uint32_t _pad0[12];
2136 uint32_t uint32[4];
2137 uint32_t _pad1[108];
2138 };
2139
2140 struct gen8_border_color {
2141 union {
2142 float float32[4];
2143 uint32_t uint32[4];
2144 };
2145 /* Pad out to 64 bytes */
2146 uint32_t _pad[12];
2147 };
2148
2149 static void
2150 anv_device_init_border_colors(struct anv_device *device)
2151 {
2152 if (device->info.is_haswell) {
2153 static const struct hsw_border_color border_colors[] = {
2154 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
2155 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
2156 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
2157 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = { .uint32 = { 0, 0, 0, 0 } },
2158 [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = { .uint32 = { 0, 0, 0, 1 } },
2159 [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = { .uint32 = { 1, 1, 1, 1 } },
2160 };
2161
2162 device->border_colors =
2163 anv_state_pool_emit_data(&device->dynamic_state_pool,
2164 sizeof(border_colors), 512, border_colors);
2165 } else {
2166 static const struct gen8_border_color border_colors[] = {
2167 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
2168 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
2169 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
2170 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = { .uint32 = { 0, 0, 0, 0 } },
2171 [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = { .uint32 = { 0, 0, 0, 1 } },
2172 [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = { .uint32 = { 1, 1, 1, 1 } },
2173 };
2174
2175 device->border_colors =
2176 anv_state_pool_emit_data(&device->dynamic_state_pool,
2177 sizeof(border_colors), 64, border_colors);
2178 }
2179 }
2180
2181 static VkResult
2182 anv_device_init_trivial_batch(struct anv_device *device)
2183 {
2184 VkResult result = anv_device_alloc_bo(device, 4096,
2185 ANV_BO_ALLOC_MAPPED,
2186 0 /* explicit_address */,
2187 &device->trivial_batch_bo);
2188 if (result != VK_SUCCESS)
2189 return result;
2190
2191 struct anv_batch batch = {
2192 .start = device->trivial_batch_bo->map,
2193 .next = device->trivial_batch_bo->map,
2194 .end = device->trivial_batch_bo->map + 4096,
2195 };
2196
2197 anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
2198 anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
2199
2200 if (!device->info.has_llc)
2201 gen_clflush_range(batch.start, batch.next - batch.start);
2202
2203 return VK_SUCCESS;
2204 }
2205
2206 VkResult anv_EnumerateDeviceExtensionProperties(
2207 VkPhysicalDevice physicalDevice,
2208 const char* pLayerName,
2209 uint32_t* pPropertyCount,
2210 VkExtensionProperties* pProperties)
2211 {
2212 ANV_FROM_HANDLE(anv_physical_device, device, physicalDevice);
2213 VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
2214
2215 for (int i = 0; i < ANV_DEVICE_EXTENSION_COUNT; i++) {
2216 if (device->supported_extensions.extensions[i]) {
2217 vk_outarray_append(&out, prop) {
2218 *prop = anv_device_extensions[i];
2219 }
2220 }
2221 }
2222
2223 return vk_outarray_status(&out);
2224 }
2225
2226 static void
2227 anv_device_init_dispatch(struct anv_device *device)
2228 {
2229 const struct anv_device_dispatch_table *genX_table;
2230 switch (device->info.gen) {
2231 case 12:
2232 genX_table = &gen12_device_dispatch_table;
2233 break;
2234 case 11:
2235 genX_table = &gen11_device_dispatch_table;
2236 break;
2237 case 10:
2238 genX_table = &gen10_device_dispatch_table;
2239 break;
2240 case 9:
2241 genX_table = &gen9_device_dispatch_table;
2242 break;
2243 case 8:
2244 genX_table = &gen8_device_dispatch_table;
2245 break;
2246 case 7:
2247 if (device->info.is_haswell)
2248 genX_table = &gen75_device_dispatch_table;
2249 else
2250 genX_table = &gen7_device_dispatch_table;
2251 break;
2252 default:
2253 unreachable("unsupported gen\n");
2254 }
2255
2256 for (unsigned i = 0; i < ARRAY_SIZE(device->dispatch.entrypoints); i++) {
2257 /* Vulkan requires that entrypoints for extensions which have not been
2258 * enabled must not be advertised.
2259 */
2260 if (!anv_device_entrypoint_is_enabled(i, device->instance->app_info.api_version,
2261 &device->instance->enabled_extensions,
2262 &device->enabled_extensions)) {
2263 device->dispatch.entrypoints[i] = NULL;
2264 } else if (genX_table->entrypoints[i]) {
2265 device->dispatch.entrypoints[i] = genX_table->entrypoints[i];
2266 } else {
2267 device->dispatch.entrypoints[i] =
2268 anv_device_dispatch_table.entrypoints[i];
2269 }
2270 }
2271 }
2272
2273 static int
2274 vk_priority_to_gen(int priority)
2275 {
2276 switch (priority) {
2277 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT:
2278 return GEN_CONTEXT_LOW_PRIORITY;
2279 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT:
2280 return GEN_CONTEXT_MEDIUM_PRIORITY;
2281 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT:
2282 return GEN_CONTEXT_HIGH_PRIORITY;
2283 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT:
2284 return GEN_CONTEXT_REALTIME_PRIORITY;
2285 default:
2286 unreachable("Invalid priority");
2287 }
2288 }
2289
2290 static VkResult
2291 anv_device_init_hiz_clear_value_bo(struct anv_device *device)
2292 {
2293 VkResult result = anv_device_alloc_bo(device, 4096,
2294 ANV_BO_ALLOC_MAPPED,
2295 0 /* explicit_address */,
2296 &device->hiz_clear_bo);
2297 if (result != VK_SUCCESS)
2298 return result;
2299
2300 union isl_color_value hiz_clear = { .u32 = { 0, } };
2301 hiz_clear.f32[0] = ANV_HZ_FC_VAL;
2302
2303 memcpy(device->hiz_clear_bo->map, hiz_clear.u32, sizeof(hiz_clear.u32));
2304
2305 if (!device->info.has_llc)
2306 gen_clflush_range(device->hiz_clear_bo->map, sizeof(hiz_clear.u32));
2307
2308 return VK_SUCCESS;
2309 }
2310
2311 static bool
2312 get_bo_from_pool(struct gen_batch_decode_bo *ret,
2313 struct anv_block_pool *pool,
2314 uint64_t address)
2315 {
2316 anv_block_pool_foreach_bo(bo, pool) {
2317 uint64_t bo_address = gen_48b_address(bo->offset);
2318 if (address >= bo_address && address < (bo_address + bo->size)) {
2319 *ret = (struct gen_batch_decode_bo) {
2320 .addr = bo_address,
2321 .size = bo->size,
2322 .map = bo->map,
2323 };
2324 return true;
2325 }
2326 }
2327 return false;
2328 }
2329
2330 /* Finding a buffer for batch decoding */
2331 static struct gen_batch_decode_bo
2332 decode_get_bo(void *v_batch, bool ppgtt, uint64_t address)
2333 {
2334 struct anv_device *device = v_batch;
2335 struct gen_batch_decode_bo ret_bo = {};
2336
2337 assert(ppgtt);
2338
2339 if (get_bo_from_pool(&ret_bo, &device->dynamic_state_pool.block_pool, address))
2340 return ret_bo;
2341 if (get_bo_from_pool(&ret_bo, &device->instruction_state_pool.block_pool, address))
2342 return ret_bo;
2343 if (get_bo_from_pool(&ret_bo, &device->binding_table_pool.block_pool, address))
2344 return ret_bo;
2345 if (get_bo_from_pool(&ret_bo, &device->surface_state_pool.block_pool, address))
2346 return ret_bo;
2347
2348 if (!device->cmd_buffer_being_decoded)
2349 return (struct gen_batch_decode_bo) { };
2350
2351 struct anv_batch_bo **bo;
2352
2353 u_vector_foreach(bo, &device->cmd_buffer_being_decoded->seen_bbos) {
2354 /* The decoder zeroes out the top 16 bits, so we need to as well */
2355 uint64_t bo_address = (*bo)->bo->offset & (~0ull >> 16);
2356
2357 if (address >= bo_address && address < bo_address + (*bo)->bo->size) {
2358 return (struct gen_batch_decode_bo) {
2359 .addr = bo_address,
2360 .size = (*bo)->bo->size,
2361 .map = (*bo)->bo->map,
2362 };
2363 }
2364 }
2365
2366 return (struct gen_batch_decode_bo) { };
2367 }
2368
2369 struct gen_aux_map_buffer {
2370 struct gen_buffer base;
2371 struct anv_state state;
2372 };
2373
2374 static struct gen_buffer *
2375 gen_aux_map_buffer_alloc(void *driver_ctx, uint32_t size)
2376 {
2377 struct gen_aux_map_buffer *buf = malloc(sizeof(struct gen_aux_map_buffer));
2378 if (!buf)
2379 return NULL;
2380
2381 struct anv_device *device = (struct anv_device*)driver_ctx;
2382 assert(device->instance->physicalDevice.supports_48bit_addresses &&
2383 device->instance->physicalDevice.use_softpin);
2384
2385 struct anv_state_pool *pool = &device->dynamic_state_pool;
2386 buf->state = anv_state_pool_alloc(pool, size, size);
2387
2388 buf->base.gpu = pool->block_pool.bo->offset + buf->state.offset;
2389 buf->base.gpu_end = buf->base.gpu + buf->state.alloc_size;
2390 buf->base.map = buf->state.map;
2391 buf->base.driver_bo = &buf->state;
2392 return &buf->base;
2393 }
2394
2395 static void
2396 gen_aux_map_buffer_free(void *driver_ctx, struct gen_buffer *buffer)
2397 {
2398 struct gen_aux_map_buffer *buf = (struct gen_aux_map_buffer*)buffer;
2399 struct anv_device *device = (struct anv_device*)driver_ctx;
2400 struct anv_state_pool *pool = &device->dynamic_state_pool;
2401 anv_state_pool_free(pool, buf->state);
2402 free(buf);
2403 }
2404
2405 static struct gen_mapped_pinned_buffer_alloc aux_map_allocator = {
2406 .alloc = gen_aux_map_buffer_alloc,
2407 .free = gen_aux_map_buffer_free,
2408 };
2409
2410 VkResult anv_CreateDevice(
2411 VkPhysicalDevice physicalDevice,
2412 const VkDeviceCreateInfo* pCreateInfo,
2413 const VkAllocationCallbacks* pAllocator,
2414 VkDevice* pDevice)
2415 {
2416 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
2417 VkResult result;
2418 struct anv_device *device;
2419
2420 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO);
2421
2422 struct anv_device_extension_table enabled_extensions = { };
2423 for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
2424 int idx;
2425 for (idx = 0; idx < ANV_DEVICE_EXTENSION_COUNT; idx++) {
2426 if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
2427 anv_device_extensions[idx].extensionName) == 0)
2428 break;
2429 }
2430
2431 if (idx >= ANV_DEVICE_EXTENSION_COUNT)
2432 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
2433
2434 if (!physical_device->supported_extensions.extensions[idx])
2435 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
2436
2437 enabled_extensions.extensions[idx] = true;
2438 }
2439
2440 /* Check enabled features */
2441 if (pCreateInfo->pEnabledFeatures) {
2442 VkPhysicalDeviceFeatures supported_features;
2443 anv_GetPhysicalDeviceFeatures(physicalDevice, &supported_features);
2444 VkBool32 *supported_feature = (VkBool32 *)&supported_features;
2445 VkBool32 *enabled_feature = (VkBool32 *)pCreateInfo->pEnabledFeatures;
2446 unsigned num_features = sizeof(VkPhysicalDeviceFeatures) / sizeof(VkBool32);
2447 for (uint32_t i = 0; i < num_features; i++) {
2448 if (enabled_feature[i] && !supported_feature[i])
2449 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
2450 }
2451 }
2452
2453 /* Check requested queues and fail if we are requested to create any
2454 * queues with flags we don't support.
2455 */
2456 assert(pCreateInfo->queueCreateInfoCount > 0);
2457 for (uint32_t i = 0; i < pCreateInfo->queueCreateInfoCount; i++) {
2458 if (pCreateInfo->pQueueCreateInfos[i].flags != 0)
2459 return vk_error(VK_ERROR_INITIALIZATION_FAILED);
2460 }
2461
2462 /* Check if client specified queue priority. */
2463 const VkDeviceQueueGlobalPriorityCreateInfoEXT *queue_priority =
2464 vk_find_struct_const(pCreateInfo->pQueueCreateInfos[0].pNext,
2465 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT);
2466
2467 VkQueueGlobalPriorityEXT priority =
2468 queue_priority ? queue_priority->globalPriority :
2469 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT;
2470
2471 device = vk_alloc2(&physical_device->instance->alloc, pAllocator,
2472 sizeof(*device), 8,
2473 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
2474 if (!device)
2475 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2476
2477 if (INTEL_DEBUG & DEBUG_BATCH) {
2478 const unsigned decode_flags =
2479 GEN_BATCH_DECODE_FULL |
2480 ((INTEL_DEBUG & DEBUG_COLOR) ? GEN_BATCH_DECODE_IN_COLOR : 0) |
2481 GEN_BATCH_DECODE_OFFSETS |
2482 GEN_BATCH_DECODE_FLOATS;
2483
2484 gen_batch_decode_ctx_init(&device->decoder_ctx,
2485 &physical_device->info,
2486 stderr, decode_flags, NULL,
2487 decode_get_bo, NULL, device);
2488 }
2489
2490 device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
2491 device->instance = physical_device->instance;
2492 device->chipset_id = physical_device->chipset_id;
2493 device->no_hw = physical_device->no_hw;
2494 device->_lost = false;
2495
2496 if (pAllocator)
2497 device->alloc = *pAllocator;
2498 else
2499 device->alloc = physical_device->instance->alloc;
2500
2501 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
2502 device->fd = open(physical_device->path, O_RDWR | O_CLOEXEC);
2503 if (device->fd == -1) {
2504 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2505 goto fail_device;
2506 }
2507
2508 device->context_id = anv_gem_create_context(device);
2509 if (device->context_id == -1) {
2510 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2511 goto fail_fd;
2512 }
2513
2514 result = anv_queue_init(device, &device->queue);
2515 if (result != VK_SUCCESS)
2516 goto fail_context_id;
2517
2518 if (physical_device->use_softpin) {
2519 if (pthread_mutex_init(&device->vma_mutex, NULL) != 0) {
2520 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2521 goto fail_queue;
2522 }
2523
2524 /* keep the page with address zero out of the allocator */
2525 util_vma_heap_init(&device->vma_lo,
2526 LOW_HEAP_MIN_ADDRESS, LOW_HEAP_SIZE);
2527
2528 util_vma_heap_init(&device->vma_cva, CLIENT_VISIBLE_HEAP_MIN_ADDRESS,
2529 CLIENT_VISIBLE_HEAP_SIZE);
2530
2531 /* Leave the last 4GiB out of the high vma range, so that no state
2532 * base address + size can overflow 48 bits. For more information see
2533 * the comment about Wa32bitGeneralStateOffset in anv_allocator.c
2534 */
2535 util_vma_heap_init(&device->vma_hi, HIGH_HEAP_MIN_ADDRESS,
2536 physical_device->gtt_size - (1ull << 32) -
2537 HIGH_HEAP_MIN_ADDRESS);
2538 }
2539
2540 list_inithead(&device->memory_objects);
2541
2542 /* As per spec, the driver implementation may deny requests to acquire
2543 * a priority above the default priority (MEDIUM) if the caller does not
2544 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
2545 * is returned.
2546 */
2547 if (physical_device->has_context_priority) {
2548 int err = anv_gem_set_context_param(device->fd, device->context_id,
2549 I915_CONTEXT_PARAM_PRIORITY,
2550 vk_priority_to_gen(priority));
2551 if (err != 0 && priority > VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT) {
2552 result = vk_error(VK_ERROR_NOT_PERMITTED_EXT);
2553 goto fail_vmas;
2554 }
2555 }
2556
2557 device->info = physical_device->info;
2558 device->isl_dev = physical_device->isl_dev;
2559
2560 /* On Broadwell and later, we can use batch chaining to more efficiently
2561 * implement growing command buffers. Prior to Haswell, the kernel
2562 * command parser gets in the way and we have to fall back to growing
2563 * the batch.
2564 */
2565 device->can_chain_batches = device->info.gen >= 8;
2566
2567 device->robust_buffer_access = pCreateInfo->pEnabledFeatures &&
2568 pCreateInfo->pEnabledFeatures->robustBufferAccess;
2569 device->enabled_extensions = enabled_extensions;
2570
2571 anv_device_init_dispatch(device);
2572
2573 if (pthread_mutex_init(&device->mutex, NULL) != 0) {
2574 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2575 goto fail_queue;
2576 }
2577
2578 pthread_condattr_t condattr;
2579 if (pthread_condattr_init(&condattr) != 0) {
2580 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2581 goto fail_mutex;
2582 }
2583 if (pthread_condattr_setclock(&condattr, CLOCK_MONOTONIC) != 0) {
2584 pthread_condattr_destroy(&condattr);
2585 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2586 goto fail_mutex;
2587 }
2588 if (pthread_cond_init(&device->queue_submit, &condattr) != 0) {
2589 pthread_condattr_destroy(&condattr);
2590 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2591 goto fail_mutex;
2592 }
2593 pthread_condattr_destroy(&condattr);
2594
2595 result = anv_bo_cache_init(&device->bo_cache);
2596 if (result != VK_SUCCESS)
2597 goto fail_queue_cond;
2598
2599 anv_bo_pool_init(&device->batch_bo_pool, device);
2600
2601 result = anv_state_pool_init(&device->dynamic_state_pool, device,
2602 DYNAMIC_STATE_POOL_MIN_ADDRESS, 16384);
2603 if (result != VK_SUCCESS)
2604 goto fail_batch_bo_pool;
2605
2606 result = anv_state_pool_init(&device->instruction_state_pool, device,
2607 INSTRUCTION_STATE_POOL_MIN_ADDRESS, 16384);
2608 if (result != VK_SUCCESS)
2609 goto fail_dynamic_state_pool;
2610
2611 result = anv_state_pool_init(&device->surface_state_pool, device,
2612 SURFACE_STATE_POOL_MIN_ADDRESS, 4096);
2613 if (result != VK_SUCCESS)
2614 goto fail_instruction_state_pool;
2615
2616 if (physical_device->use_softpin) {
2617 result = anv_state_pool_init(&device->binding_table_pool, device,
2618 BINDING_TABLE_POOL_MIN_ADDRESS, 4096);
2619 if (result != VK_SUCCESS)
2620 goto fail_surface_state_pool;
2621 }
2622
2623 if (device->info.gen >= 12) {
2624 device->aux_map_ctx = gen_aux_map_init(device, &aux_map_allocator,
2625 &physical_device->info);
2626 if (!device->aux_map_ctx)
2627 goto fail_binding_table_pool;
2628 }
2629
2630 result = anv_device_alloc_bo(device, 4096, 0 /* flags */,
2631 0 /* explicit_address */,
2632 &device->workaround_bo);
2633 if (result != VK_SUCCESS)
2634 goto fail_surface_aux_map_pool;
2635
2636 result = anv_device_init_trivial_batch(device);
2637 if (result != VK_SUCCESS)
2638 goto fail_workaround_bo;
2639
2640 if (device->info.gen >= 10) {
2641 result = anv_device_init_hiz_clear_value_bo(device);
2642 if (result != VK_SUCCESS)
2643 goto fail_trivial_batch_bo;
2644 }
2645
2646 anv_scratch_pool_init(device, &device->scratch_pool);
2647
2648 switch (device->info.gen) {
2649 case 7:
2650 if (!device->info.is_haswell)
2651 result = gen7_init_device_state(device);
2652 else
2653 result = gen75_init_device_state(device);
2654 break;
2655 case 8:
2656 result = gen8_init_device_state(device);
2657 break;
2658 case 9:
2659 result = gen9_init_device_state(device);
2660 break;
2661 case 10:
2662 result = gen10_init_device_state(device);
2663 break;
2664 case 11:
2665 result = gen11_init_device_state(device);
2666 break;
2667 case 12:
2668 result = gen12_init_device_state(device);
2669 break;
2670 default:
2671 /* Shouldn't get here as we don't create physical devices for any other
2672 * gens. */
2673 unreachable("unhandled gen");
2674 }
2675 if (result != VK_SUCCESS)
2676 goto fail_workaround_bo;
2677
2678 anv_pipeline_cache_init(&device->default_pipeline_cache, device, true);
2679
2680 anv_device_init_blorp(device);
2681
2682 anv_device_init_border_colors(device);
2683
2684 anv_device_perf_init(device);
2685
2686 *pDevice = anv_device_to_handle(device);
2687
2688 return VK_SUCCESS;
2689
2690 fail_workaround_bo:
2691 anv_scratch_pool_finish(device, &device->scratch_pool);
2692 if (device->info.gen >= 10)
2693 anv_device_release_bo(device, device->hiz_clear_bo);
2694 anv_device_release_bo(device, device->workaround_bo);
2695 fail_trivial_batch_bo:
2696 anv_device_release_bo(device, device->trivial_batch_bo);
2697 fail_surface_aux_map_pool:
2698 if (device->info.gen >= 12) {
2699 gen_aux_map_finish(device->aux_map_ctx);
2700 device->aux_map_ctx = NULL;
2701 }
2702 fail_binding_table_pool:
2703 if (physical_device->use_softpin)
2704 anv_state_pool_finish(&device->binding_table_pool);
2705 fail_surface_state_pool:
2706 anv_state_pool_finish(&device->surface_state_pool);
2707 fail_instruction_state_pool:
2708 anv_state_pool_finish(&device->instruction_state_pool);
2709 fail_dynamic_state_pool:
2710 anv_state_pool_finish(&device->dynamic_state_pool);
2711 fail_batch_bo_pool:
2712 anv_bo_pool_finish(&device->batch_bo_pool);
2713 anv_bo_cache_finish(&device->bo_cache);
2714 fail_queue_cond:
2715 pthread_cond_destroy(&device->queue_submit);
2716 fail_mutex:
2717 pthread_mutex_destroy(&device->mutex);
2718 fail_vmas:
2719 if (physical_device->use_softpin) {
2720 util_vma_heap_finish(&device->vma_hi);
2721 util_vma_heap_finish(&device->vma_cva);
2722 util_vma_heap_finish(&device->vma_lo);
2723 }
2724 fail_queue:
2725 anv_queue_finish(&device->queue);
2726 fail_context_id:
2727 anv_gem_destroy_context(device, device->context_id);
2728 fail_fd:
2729 close(device->fd);
2730 fail_device:
2731 vk_free(&device->alloc, device);
2732
2733 return result;
2734 }
2735
2736 void anv_DestroyDevice(
2737 VkDevice _device,
2738 const VkAllocationCallbacks* pAllocator)
2739 {
2740 ANV_FROM_HANDLE(anv_device, device, _device);
2741 struct anv_physical_device *physical_device;
2742
2743 if (!device)
2744 return;
2745
2746 physical_device = &device->instance->physicalDevice;
2747
2748 anv_device_finish_blorp(device);
2749
2750 anv_pipeline_cache_finish(&device->default_pipeline_cache);
2751
2752 anv_queue_finish(&device->queue);
2753
2754 #ifdef HAVE_VALGRIND
2755 /* We only need to free these to prevent valgrind errors. The backing
2756 * BO will go away in a couple of lines so we don't actually leak.
2757 */
2758 anv_state_pool_free(&device->dynamic_state_pool, device->border_colors);
2759 anv_state_pool_free(&device->dynamic_state_pool, device->slice_hash);
2760 #endif
2761
2762 anv_scratch_pool_finish(device, &device->scratch_pool);
2763
2764 anv_device_release_bo(device, device->workaround_bo);
2765 anv_device_release_bo(device, device->trivial_batch_bo);
2766 if (device->info.gen >= 10)
2767 anv_device_release_bo(device, device->hiz_clear_bo);
2768
2769 if (device->info.gen >= 12) {
2770 gen_aux_map_finish(device->aux_map_ctx);
2771 device->aux_map_ctx = NULL;
2772 }
2773
2774 if (physical_device->use_softpin)
2775 anv_state_pool_finish(&device->binding_table_pool);
2776 anv_state_pool_finish(&device->surface_state_pool);
2777 anv_state_pool_finish(&device->instruction_state_pool);
2778 anv_state_pool_finish(&device->dynamic_state_pool);
2779
2780 anv_bo_pool_finish(&device->batch_bo_pool);
2781
2782 anv_bo_cache_finish(&device->bo_cache);
2783
2784 if (physical_device->use_softpin) {
2785 util_vma_heap_finish(&device->vma_hi);
2786 util_vma_heap_finish(&device->vma_cva);
2787 util_vma_heap_finish(&device->vma_lo);
2788 }
2789
2790 pthread_cond_destroy(&device->queue_submit);
2791 pthread_mutex_destroy(&device->mutex);
2792
2793 anv_gem_destroy_context(device, device->context_id);
2794
2795 if (INTEL_DEBUG & DEBUG_BATCH)
2796 gen_batch_decode_ctx_finish(&device->decoder_ctx);
2797
2798 close(device->fd);
2799
2800 vk_free(&device->alloc, device);
2801 }
2802
2803 VkResult anv_EnumerateInstanceLayerProperties(
2804 uint32_t* pPropertyCount,
2805 VkLayerProperties* pProperties)
2806 {
2807 if (pProperties == NULL) {
2808 *pPropertyCount = 0;
2809 return VK_SUCCESS;
2810 }
2811
2812 /* None supported at this time */
2813 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
2814 }
2815
2816 VkResult anv_EnumerateDeviceLayerProperties(
2817 VkPhysicalDevice physicalDevice,
2818 uint32_t* pPropertyCount,
2819 VkLayerProperties* pProperties)
2820 {
2821 if (pProperties == NULL) {
2822 *pPropertyCount = 0;
2823 return VK_SUCCESS;
2824 }
2825
2826 /* None supported at this time */
2827 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
2828 }
2829
2830 void anv_GetDeviceQueue(
2831 VkDevice _device,
2832 uint32_t queueNodeIndex,
2833 uint32_t queueIndex,
2834 VkQueue* pQueue)
2835 {
2836 const VkDeviceQueueInfo2 info = {
2837 .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_INFO_2,
2838 .pNext = NULL,
2839 .flags = 0,
2840 .queueFamilyIndex = queueNodeIndex,
2841 .queueIndex = queueIndex,
2842 };
2843
2844 anv_GetDeviceQueue2(_device, &info, pQueue);
2845 }
2846
2847 void anv_GetDeviceQueue2(
2848 VkDevice _device,
2849 const VkDeviceQueueInfo2* pQueueInfo,
2850 VkQueue* pQueue)
2851 {
2852 ANV_FROM_HANDLE(anv_device, device, _device);
2853
2854 assert(pQueueInfo->queueIndex == 0);
2855
2856 if (pQueueInfo->flags == device->queue.flags)
2857 *pQueue = anv_queue_to_handle(&device->queue);
2858 else
2859 *pQueue = NULL;
2860 }
2861
2862 VkResult
2863 _anv_device_set_lost(struct anv_device *device,
2864 const char *file, int line,
2865 const char *msg, ...)
2866 {
2867 VkResult err;
2868 va_list ap;
2869
2870 p_atomic_inc(&device->_lost);
2871
2872 va_start(ap, msg);
2873 err = __vk_errorv(device->instance, device,
2874 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT,
2875 VK_ERROR_DEVICE_LOST, file, line, msg, ap);
2876 va_end(ap);
2877
2878 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
2879 abort();
2880
2881 return err;
2882 }
2883
2884 VkResult
2885 _anv_queue_set_lost(struct anv_queue *queue,
2886 const char *file, int line,
2887 const char *msg, ...)
2888 {
2889 VkResult err;
2890 va_list ap;
2891
2892 p_atomic_inc(&queue->device->_lost);
2893
2894 va_start(ap, msg);
2895 err = __vk_errorv(queue->device->instance, queue->device,
2896 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT,
2897 VK_ERROR_DEVICE_LOST, file, line, msg, ap);
2898 va_end(ap);
2899
2900 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
2901 abort();
2902
2903 return err;
2904 }
2905
2906 VkResult
2907 anv_device_query_status(struct anv_device *device)
2908 {
2909 /* This isn't likely as most of the callers of this function already check
2910 * for it. However, it doesn't hurt to check and it potentially lets us
2911 * avoid an ioctl.
2912 */
2913 if (anv_device_is_lost(device))
2914 return VK_ERROR_DEVICE_LOST;
2915
2916 uint32_t active, pending;
2917 int ret = anv_gem_gpu_get_reset_stats(device, &active, &pending);
2918 if (ret == -1) {
2919 /* We don't know the real error. */
2920 return anv_device_set_lost(device, "get_reset_stats failed: %m");
2921 }
2922
2923 if (active) {
2924 return anv_device_set_lost(device, "GPU hung on one of our command buffers");
2925 } else if (pending) {
2926 return anv_device_set_lost(device, "GPU hung with commands in-flight");
2927 }
2928
2929 return VK_SUCCESS;
2930 }
2931
2932 VkResult
2933 anv_device_bo_busy(struct anv_device *device, struct anv_bo *bo)
2934 {
2935 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
2936 * Other usages of the BO (such as on different hardware) will not be
2937 * flagged as "busy" by this ioctl. Use with care.
2938 */
2939 int ret = anv_gem_busy(device, bo->gem_handle);
2940 if (ret == 1) {
2941 return VK_NOT_READY;
2942 } else if (ret == -1) {
2943 /* We don't know the real error. */
2944 return anv_device_set_lost(device, "gem wait failed: %m");
2945 }
2946
2947 /* Query for device status after the busy call. If the BO we're checking
2948 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
2949 * client because it clearly doesn't have valid data. Yes, this most
2950 * likely means an ioctl, but we just did an ioctl to query the busy status
2951 * so it's no great loss.
2952 */
2953 return anv_device_query_status(device);
2954 }
2955
2956 VkResult
2957 anv_device_wait(struct anv_device *device, struct anv_bo *bo,
2958 int64_t timeout)
2959 {
2960 int ret = anv_gem_wait(device, bo->gem_handle, &timeout);
2961 if (ret == -1 && errno == ETIME) {
2962 return VK_TIMEOUT;
2963 } else if (ret == -1) {
2964 /* We don't know the real error. */
2965 return anv_device_set_lost(device, "gem wait failed: %m");
2966 }
2967
2968 /* Query for device status after the wait. If the BO we're waiting on got
2969 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
2970 * because it clearly doesn't have valid data. Yes, this most likely means
2971 * an ioctl, but we just did an ioctl to wait so it's no great loss.
2972 */
2973 return anv_device_query_status(device);
2974 }
2975
2976 VkResult anv_DeviceWaitIdle(
2977 VkDevice _device)
2978 {
2979 ANV_FROM_HANDLE(anv_device, device, _device);
2980
2981 if (anv_device_is_lost(device))
2982 return VK_ERROR_DEVICE_LOST;
2983
2984 return anv_queue_submit_simple_batch(&device->queue, NULL);
2985 }
2986
2987 bool
2988 anv_vma_alloc(struct anv_device *device, struct anv_bo *bo,
2989 uint64_t client_address)
2990 {
2991 const struct anv_physical_device *pdevice = &device->instance->physicalDevice;
2992 const struct gen_device_info *devinfo = &pdevice->info;
2993 /* Gen12 CCS surface addresses need to be 64K aligned. We have no way of
2994 * telling what this allocation is for so pick the largest alignment.
2995 */
2996 const uint32_t vma_alignment =
2997 devinfo->gen >= 12 ? (64 * 1024) : (4 * 1024);
2998
2999 if (!(bo->flags & EXEC_OBJECT_PINNED)) {
3000 assert(!(bo->has_client_visible_address));
3001 return true;
3002 }
3003
3004 pthread_mutex_lock(&device->vma_mutex);
3005
3006 bo->offset = 0;
3007
3008 if (bo->has_client_visible_address) {
3009 assert(bo->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS);
3010 if (client_address) {
3011 if (util_vma_heap_alloc_addr(&device->vma_cva,
3012 client_address, bo->size)) {
3013 bo->offset = gen_canonical_address(client_address);
3014 }
3015 } else {
3016 uint64_t addr =
3017 util_vma_heap_alloc(&device->vma_cva, bo->size, vma_alignment);
3018 if (addr) {
3019 bo->offset = gen_canonical_address(addr);
3020 assert(addr == gen_48b_address(bo->offset));
3021 }
3022 }
3023 /* We don't want to fall back to other heaps */
3024 goto done;
3025 }
3026
3027 assert(client_address == 0);
3028
3029 if (bo->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) {
3030 uint64_t addr =
3031 util_vma_heap_alloc(&device->vma_hi, bo->size, vma_alignment);
3032 if (addr) {
3033 bo->offset = gen_canonical_address(addr);
3034 assert(addr == gen_48b_address(bo->offset));
3035 }
3036 }
3037
3038 if (bo->offset == 0) {
3039 uint64_t addr =
3040 util_vma_heap_alloc(&device->vma_lo, bo->size, vma_alignment);
3041 if (addr) {
3042 bo->offset = gen_canonical_address(addr);
3043 assert(addr == gen_48b_address(bo->offset));
3044 }
3045 }
3046
3047 done:
3048 pthread_mutex_unlock(&device->vma_mutex);
3049
3050 return bo->offset != 0;
3051 }
3052
3053 void
3054 anv_vma_free(struct anv_device *device, struct anv_bo *bo)
3055 {
3056 if (!(bo->flags & EXEC_OBJECT_PINNED))
3057 return;
3058
3059 const uint64_t addr_48b = gen_48b_address(bo->offset);
3060
3061 pthread_mutex_lock(&device->vma_mutex);
3062
3063 if (addr_48b >= LOW_HEAP_MIN_ADDRESS &&
3064 addr_48b <= LOW_HEAP_MAX_ADDRESS) {
3065 util_vma_heap_free(&device->vma_lo, addr_48b, bo->size);
3066 } else if (addr_48b >= CLIENT_VISIBLE_HEAP_MIN_ADDRESS &&
3067 addr_48b <= CLIENT_VISIBLE_HEAP_MAX_ADDRESS) {
3068 util_vma_heap_free(&device->vma_cva, addr_48b, bo->size);
3069 } else {
3070 assert(addr_48b >= HIGH_HEAP_MIN_ADDRESS);
3071 util_vma_heap_free(&device->vma_hi, addr_48b, bo->size);
3072 }
3073
3074 pthread_mutex_unlock(&device->vma_mutex);
3075
3076 bo->offset = 0;
3077 }
3078
3079 VkResult anv_AllocateMemory(
3080 VkDevice _device,
3081 const VkMemoryAllocateInfo* pAllocateInfo,
3082 const VkAllocationCallbacks* pAllocator,
3083 VkDeviceMemory* pMem)
3084 {
3085 ANV_FROM_HANDLE(anv_device, device, _device);
3086 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3087 struct anv_device_memory *mem;
3088 VkResult result = VK_SUCCESS;
3089
3090 assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
3091
3092 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
3093 assert(pAllocateInfo->allocationSize > 0);
3094
3095 VkDeviceSize aligned_alloc_size =
3096 align_u64(pAllocateInfo->allocationSize, 4096);
3097
3098 if (aligned_alloc_size > MAX_MEMORY_ALLOCATION_SIZE)
3099 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
3100
3101 assert(pAllocateInfo->memoryTypeIndex < pdevice->memory.type_count);
3102 struct anv_memory_type *mem_type =
3103 &pdevice->memory.types[pAllocateInfo->memoryTypeIndex];
3104 assert(mem_type->heapIndex < pdevice->memory.heap_count);
3105 struct anv_memory_heap *mem_heap =
3106 &pdevice->memory.heaps[mem_type->heapIndex];
3107
3108 uint64_t mem_heap_used = p_atomic_read(&mem_heap->used);
3109 if (mem_heap_used + aligned_alloc_size > mem_heap->size)
3110 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
3111
3112 mem = vk_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
3113 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
3114 if (mem == NULL)
3115 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
3116
3117 assert(pAllocateInfo->memoryTypeIndex < pdevice->memory.type_count);
3118 mem->type = mem_type;
3119 mem->map = NULL;
3120 mem->map_size = 0;
3121 mem->ahw = NULL;
3122 mem->host_ptr = NULL;
3123
3124 enum anv_bo_alloc_flags alloc_flags = 0;
3125
3126 const VkExportMemoryAllocateInfo *export_info = NULL;
3127 const VkImportAndroidHardwareBufferInfoANDROID *ahw_import_info = NULL;
3128 const VkImportMemoryFdInfoKHR *fd_info = NULL;
3129 const VkImportMemoryHostPointerInfoEXT *host_ptr_info = NULL;
3130 const VkMemoryDedicatedAllocateInfo *dedicated_info = NULL;
3131 VkMemoryAllocateFlags vk_flags = 0;
3132 uint64_t client_address = 0;
3133
3134 vk_foreach_struct_const(ext, pAllocateInfo->pNext) {
3135 switch (ext->sType) {
3136 case VK_STRUCTURE_TYPE_EXPORT_MEMORY_ALLOCATE_INFO:
3137 export_info = (void *)ext;
3138 break;
3139
3140 case VK_STRUCTURE_TYPE_IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID:
3141 ahw_import_info = (void *)ext;
3142 break;
3143
3144 case VK_STRUCTURE_TYPE_IMPORT_MEMORY_FD_INFO_KHR:
3145 fd_info = (void *)ext;
3146 break;
3147
3148 case VK_STRUCTURE_TYPE_IMPORT_MEMORY_HOST_POINTER_INFO_EXT:
3149 host_ptr_info = (void *)ext;
3150 break;
3151
3152 case VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_FLAGS_INFO: {
3153 const VkMemoryAllocateFlagsInfo *flags_info = (void *)ext;
3154 vk_flags = flags_info->flags;
3155 break;
3156 }
3157
3158 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_ALLOCATE_INFO:
3159 dedicated_info = (void *)ext;
3160 break;
3161
3162 case VK_STRUCTURE_TYPE_MEMORY_OPAQUE_CAPTURE_ADDRESS_ALLOCATE_INFO_KHR: {
3163 const VkMemoryOpaqueCaptureAddressAllocateInfoKHR *addr_info =
3164 (const VkMemoryOpaqueCaptureAddressAllocateInfoKHR *)ext;
3165 client_address = addr_info->opaqueCaptureAddress;
3166 break;
3167 }
3168
3169 default:
3170 anv_debug_ignored_stype(ext->sType);
3171 break;
3172 }
3173 }
3174
3175 if (vk_flags & VK_MEMORY_ALLOCATE_DEVICE_ADDRESS_BIT_KHR)
3176 alloc_flags |= ANV_BO_ALLOC_CLIENT_VISIBLE_ADDRESS;
3177
3178 /* Check if we need to support Android HW buffer export. If so,
3179 * create AHardwareBuffer and import memory from it.
3180 */
3181 bool android_export = false;
3182 if (export_info && export_info->handleTypes &
3183 VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID)
3184 android_export = true;
3185
3186 if (ahw_import_info) {
3187 result = anv_import_ahw_memory(_device, mem, ahw_import_info);
3188 if (result != VK_SUCCESS)
3189 goto fail;
3190
3191 goto success;
3192 } else if (android_export) {
3193 result = anv_create_ahw_memory(_device, mem, pAllocateInfo);
3194 if (result != VK_SUCCESS)
3195 goto fail;
3196
3197 const VkImportAndroidHardwareBufferInfoANDROID import_info = {
3198 .buffer = mem->ahw,
3199 };
3200 result = anv_import_ahw_memory(_device, mem, &import_info);
3201 if (result != VK_SUCCESS)
3202 goto fail;
3203
3204 goto success;
3205 }
3206
3207 /* The Vulkan spec permits handleType to be 0, in which case the struct is
3208 * ignored.
3209 */
3210 if (fd_info && fd_info->handleType) {
3211 /* At the moment, we support only the below handle types. */
3212 assert(fd_info->handleType ==
3213 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
3214 fd_info->handleType ==
3215 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
3216
3217 result = anv_device_import_bo(device, fd_info->fd, alloc_flags,
3218 client_address, &mem->bo);
3219 if (result != VK_SUCCESS)
3220 goto fail;
3221
3222 VkDeviceSize aligned_alloc_size =
3223 align_u64(pAllocateInfo->allocationSize, 4096);
3224
3225 /* For security purposes, we reject importing the bo if it's smaller
3226 * than the requested allocation size. This prevents a malicious client
3227 * from passing a buffer to a trusted client, lying about the size, and
3228 * telling the trusted client to try and texture from an image that goes
3229 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
3230 * in the trusted client. The trusted client can protect itself against
3231 * this sort of attack but only if it can trust the buffer size.
3232 */
3233 if (mem->bo->size < aligned_alloc_size) {
3234 result = vk_errorf(device->instance, device,
3235 VK_ERROR_INVALID_EXTERNAL_HANDLE,
3236 "aligned allocationSize too large for "
3237 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT: "
3238 "%"PRIu64"B > %"PRIu64"B",
3239 aligned_alloc_size, mem->bo->size);
3240 anv_device_release_bo(device, mem->bo);
3241 goto fail;
3242 }
3243
3244 /* From the Vulkan spec:
3245 *
3246 * "Importing memory from a file descriptor transfers ownership of
3247 * the file descriptor from the application to the Vulkan
3248 * implementation. The application must not perform any operations on
3249 * the file descriptor after a successful import."
3250 *
3251 * If the import fails, we leave the file descriptor open.
3252 */
3253 close(fd_info->fd);
3254 goto success;
3255 }
3256
3257 if (host_ptr_info && host_ptr_info->handleType) {
3258 if (host_ptr_info->handleType ==
3259 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_MAPPED_FOREIGN_MEMORY_BIT_EXT) {
3260 result = vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE);
3261 goto fail;
3262 }
3263
3264 assert(host_ptr_info->handleType ==
3265 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT);
3266
3267 result = anv_device_import_bo_from_host_ptr(device,
3268 host_ptr_info->pHostPointer,
3269 pAllocateInfo->allocationSize,
3270 alloc_flags,
3271 client_address,
3272 &mem->bo);
3273 if (result != VK_SUCCESS)
3274 goto fail;
3275
3276 mem->host_ptr = host_ptr_info->pHostPointer;
3277 goto success;
3278 }
3279
3280 /* Regular allocate (not importing memory). */
3281
3282 if (export_info && export_info->handleTypes)
3283 alloc_flags |= ANV_BO_ALLOC_EXTERNAL;
3284
3285 result = anv_device_alloc_bo(device, pAllocateInfo->allocationSize,
3286 alloc_flags, client_address, &mem->bo);
3287 if (result != VK_SUCCESS)
3288 goto fail;
3289
3290 if (dedicated_info && dedicated_info->image != VK_NULL_HANDLE) {
3291 ANV_FROM_HANDLE(anv_image, image, dedicated_info->image);
3292
3293 /* Some legacy (non-modifiers) consumers need the tiling to be set on
3294 * the BO. In this case, we have a dedicated allocation.
3295 */
3296 if (image->needs_set_tiling) {
3297 const uint32_t i915_tiling =
3298 isl_tiling_to_i915_tiling(image->planes[0].surface.isl.tiling);
3299 int ret = anv_gem_set_tiling(device, mem->bo->gem_handle,
3300 image->planes[0].surface.isl.row_pitch_B,
3301 i915_tiling);
3302 if (ret) {
3303 anv_device_release_bo(device, mem->bo);
3304 result = vk_errorf(device->instance, NULL,
3305 VK_ERROR_OUT_OF_DEVICE_MEMORY,
3306 "failed to set BO tiling: %m");
3307 goto fail;
3308 }
3309 }
3310 }
3311
3312 success:
3313 mem_heap_used = p_atomic_add_return(&mem_heap->used, mem->bo->size);
3314 if (mem_heap_used > mem_heap->size) {
3315 p_atomic_add(&mem_heap->used, -mem->bo->size);
3316 anv_device_release_bo(device, mem->bo);
3317 result = vk_errorf(device->instance, NULL,
3318 VK_ERROR_OUT_OF_DEVICE_MEMORY,
3319 "Out of heap memory");
3320 goto fail;
3321 }
3322
3323 pthread_mutex_lock(&device->mutex);
3324 list_addtail(&mem->link, &device->memory_objects);
3325 pthread_mutex_unlock(&device->mutex);
3326
3327 *pMem = anv_device_memory_to_handle(mem);
3328
3329 return VK_SUCCESS;
3330
3331 fail:
3332 vk_free2(&device->alloc, pAllocator, mem);
3333
3334 return result;
3335 }
3336
3337 VkResult anv_GetMemoryFdKHR(
3338 VkDevice device_h,
3339 const VkMemoryGetFdInfoKHR* pGetFdInfo,
3340 int* pFd)
3341 {
3342 ANV_FROM_HANDLE(anv_device, dev, device_h);
3343 ANV_FROM_HANDLE(anv_device_memory, mem, pGetFdInfo->memory);
3344
3345 assert(pGetFdInfo->sType == VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR);
3346
3347 assert(pGetFdInfo->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
3348 pGetFdInfo->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
3349
3350 return anv_device_export_bo(dev, mem->bo, pFd);
3351 }
3352
3353 VkResult anv_GetMemoryFdPropertiesKHR(
3354 VkDevice _device,
3355 VkExternalMemoryHandleTypeFlagBits handleType,
3356 int fd,
3357 VkMemoryFdPropertiesKHR* pMemoryFdProperties)
3358 {
3359 ANV_FROM_HANDLE(anv_device, device, _device);
3360 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3361
3362 switch (handleType) {
3363 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT:
3364 /* dma-buf can be imported as any memory type */
3365 pMemoryFdProperties->memoryTypeBits =
3366 (1 << pdevice->memory.type_count) - 1;
3367 return VK_SUCCESS;
3368
3369 default:
3370 /* The valid usage section for this function says:
3371 *
3372 * "handleType must not be one of the handle types defined as
3373 * opaque."
3374 *
3375 * So opaque handle types fall into the default "unsupported" case.
3376 */
3377 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE);
3378 }
3379 }
3380
3381 VkResult anv_GetMemoryHostPointerPropertiesEXT(
3382 VkDevice _device,
3383 VkExternalMemoryHandleTypeFlagBits handleType,
3384 const void* pHostPointer,
3385 VkMemoryHostPointerPropertiesEXT* pMemoryHostPointerProperties)
3386 {
3387 ANV_FROM_HANDLE(anv_device, device, _device);
3388
3389 assert(pMemoryHostPointerProperties->sType ==
3390 VK_STRUCTURE_TYPE_MEMORY_HOST_POINTER_PROPERTIES_EXT);
3391
3392 switch (handleType) {
3393 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT: {
3394 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3395
3396 /* Host memory can be imported as any memory type. */
3397 pMemoryHostPointerProperties->memoryTypeBits =
3398 (1ull << pdevice->memory.type_count) - 1;
3399
3400 return VK_SUCCESS;
3401 }
3402 default:
3403 return VK_ERROR_INVALID_EXTERNAL_HANDLE;
3404 }
3405 }
3406
3407 void anv_FreeMemory(
3408 VkDevice _device,
3409 VkDeviceMemory _mem,
3410 const VkAllocationCallbacks* pAllocator)
3411 {
3412 ANV_FROM_HANDLE(anv_device, device, _device);
3413 ANV_FROM_HANDLE(anv_device_memory, mem, _mem);
3414 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3415
3416 if (mem == NULL)
3417 return;
3418
3419 pthread_mutex_lock(&device->mutex);
3420 list_del(&mem->link);
3421 pthread_mutex_unlock(&device->mutex);
3422
3423 if (mem->map)
3424 anv_UnmapMemory(_device, _mem);
3425
3426 p_atomic_add(&pdevice->memory.heaps[mem->type->heapIndex].used,
3427 -mem->bo->size);
3428
3429 anv_device_release_bo(device, mem->bo);
3430
3431 #if defined(ANDROID) && ANDROID_API_LEVEL >= 26
3432 if (mem->ahw)
3433 AHardwareBuffer_release(mem->ahw);
3434 #endif
3435
3436 vk_free2(&device->alloc, pAllocator, mem);
3437 }
3438
3439 VkResult anv_MapMemory(
3440 VkDevice _device,
3441 VkDeviceMemory _memory,
3442 VkDeviceSize offset,
3443 VkDeviceSize size,
3444 VkMemoryMapFlags flags,
3445 void** ppData)
3446 {
3447 ANV_FROM_HANDLE(anv_device, device, _device);
3448 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
3449
3450 if (mem == NULL) {
3451 *ppData = NULL;
3452 return VK_SUCCESS;
3453 }
3454
3455 if (mem->host_ptr) {
3456 *ppData = mem->host_ptr + offset;
3457 return VK_SUCCESS;
3458 }
3459
3460 if (size == VK_WHOLE_SIZE)
3461 size = mem->bo->size - offset;
3462
3463 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
3464 *
3465 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
3466 * assert(size != 0);
3467 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
3468 * equal to the size of the memory minus offset
3469 */
3470 assert(size > 0);
3471 assert(offset + size <= mem->bo->size);
3472
3473 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
3474 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
3475 * at a time is valid. We could just mmap up front and return an offset
3476 * pointer here, but that may exhaust virtual memory on 32 bit
3477 * userspace. */
3478
3479 uint32_t gem_flags = 0;
3480
3481 if (!device->info.has_llc &&
3482 (mem->type->propertyFlags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT))
3483 gem_flags |= I915_MMAP_WC;
3484
3485 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
3486 uint64_t map_offset = offset & ~4095ull;
3487 assert(offset >= map_offset);
3488 uint64_t map_size = (offset + size) - map_offset;
3489
3490 /* Let's map whole pages */
3491 map_size = align_u64(map_size, 4096);
3492
3493 void *map = anv_gem_mmap(device, mem->bo->gem_handle,
3494 map_offset, map_size, gem_flags);
3495 if (map == MAP_FAILED)
3496 return vk_error(VK_ERROR_MEMORY_MAP_FAILED);
3497
3498 mem->map = map;
3499 mem->map_size = map_size;
3500
3501 *ppData = mem->map + (offset - map_offset);
3502
3503 return VK_SUCCESS;
3504 }
3505
3506 void anv_UnmapMemory(
3507 VkDevice _device,
3508 VkDeviceMemory _memory)
3509 {
3510 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
3511
3512 if (mem == NULL || mem->host_ptr)
3513 return;
3514
3515 anv_gem_munmap(mem->map, mem->map_size);
3516
3517 mem->map = NULL;
3518 mem->map_size = 0;
3519 }
3520
3521 static void
3522 clflush_mapped_ranges(struct anv_device *device,
3523 uint32_t count,
3524 const VkMappedMemoryRange *ranges)
3525 {
3526 for (uint32_t i = 0; i < count; i++) {
3527 ANV_FROM_HANDLE(anv_device_memory, mem, ranges[i].memory);
3528 if (ranges[i].offset >= mem->map_size)
3529 continue;
3530
3531 gen_clflush_range(mem->map + ranges[i].offset,
3532 MIN2(ranges[i].size, mem->map_size - ranges[i].offset));
3533 }
3534 }
3535
3536 VkResult anv_FlushMappedMemoryRanges(
3537 VkDevice _device,
3538 uint32_t memoryRangeCount,
3539 const VkMappedMemoryRange* pMemoryRanges)
3540 {
3541 ANV_FROM_HANDLE(anv_device, device, _device);
3542
3543 if (device->info.has_llc)
3544 return VK_SUCCESS;
3545
3546 /* Make sure the writes we're flushing have landed. */
3547 __builtin_ia32_mfence();
3548
3549 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
3550
3551 return VK_SUCCESS;
3552 }
3553
3554 VkResult anv_InvalidateMappedMemoryRanges(
3555 VkDevice _device,
3556 uint32_t memoryRangeCount,
3557 const VkMappedMemoryRange* pMemoryRanges)
3558 {
3559 ANV_FROM_HANDLE(anv_device, device, _device);
3560
3561 if (device->info.has_llc)
3562 return VK_SUCCESS;
3563
3564 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
3565
3566 /* Make sure no reads get moved up above the invalidate. */
3567 __builtin_ia32_mfence();
3568
3569 return VK_SUCCESS;
3570 }
3571
3572 void anv_GetBufferMemoryRequirements(
3573 VkDevice _device,
3574 VkBuffer _buffer,
3575 VkMemoryRequirements* pMemoryRequirements)
3576 {
3577 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
3578 ANV_FROM_HANDLE(anv_device, device, _device);
3579 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3580
3581 /* The Vulkan spec (git aaed022) says:
3582 *
3583 * memoryTypeBits is a bitfield and contains one bit set for every
3584 * supported memory type for the resource. The bit `1<<i` is set if and
3585 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3586 * structure for the physical device is supported.
3587 */
3588 uint32_t memory_types = (1ull << pdevice->memory.type_count) - 1;
3589
3590 /* Base alignment requirement of a cache line */
3591 uint32_t alignment = 16;
3592
3593 /* We need an alignment of 32 for pushing UBOs */
3594 if (buffer->usage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT)
3595 alignment = MAX2(alignment, 32);
3596
3597 pMemoryRequirements->size = buffer->size;
3598 pMemoryRequirements->alignment = alignment;
3599
3600 /* Storage and Uniform buffers should have their size aligned to
3601 * 32-bits to avoid boundary checks when last DWord is not complete.
3602 * This would ensure that not internal padding would be needed for
3603 * 16-bit types.
3604 */
3605 if (device->robust_buffer_access &&
3606 (buffer->usage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT ||
3607 buffer->usage & VK_BUFFER_USAGE_STORAGE_BUFFER_BIT))
3608 pMemoryRequirements->size = align_u64(buffer->size, 4);
3609
3610 pMemoryRequirements->memoryTypeBits = memory_types;
3611 }
3612
3613 void anv_GetBufferMemoryRequirements2(
3614 VkDevice _device,
3615 const VkBufferMemoryRequirementsInfo2* pInfo,
3616 VkMemoryRequirements2* pMemoryRequirements)
3617 {
3618 anv_GetBufferMemoryRequirements(_device, pInfo->buffer,
3619 &pMemoryRequirements->memoryRequirements);
3620
3621 vk_foreach_struct(ext, pMemoryRequirements->pNext) {
3622 switch (ext->sType) {
3623 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
3624 VkMemoryDedicatedRequirements *requirements = (void *)ext;
3625 requirements->prefersDedicatedAllocation = false;
3626 requirements->requiresDedicatedAllocation = false;
3627 break;
3628 }
3629
3630 default:
3631 anv_debug_ignored_stype(ext->sType);
3632 break;
3633 }
3634 }
3635 }
3636
3637 void anv_GetImageMemoryRequirements(
3638 VkDevice _device,
3639 VkImage _image,
3640 VkMemoryRequirements* pMemoryRequirements)
3641 {
3642 ANV_FROM_HANDLE(anv_image, image, _image);
3643 ANV_FROM_HANDLE(anv_device, device, _device);
3644 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3645
3646 /* The Vulkan spec (git aaed022) says:
3647 *
3648 * memoryTypeBits is a bitfield and contains one bit set for every
3649 * supported memory type for the resource. The bit `1<<i` is set if and
3650 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3651 * structure for the physical device is supported.
3652 *
3653 * All types are currently supported for images.
3654 */
3655 uint32_t memory_types = (1ull << pdevice->memory.type_count) - 1;
3656
3657 /* We must have image allocated or imported at this point. According to the
3658 * specification, external images must have been bound to memory before
3659 * calling GetImageMemoryRequirements.
3660 */
3661 assert(image->size > 0);
3662
3663 pMemoryRequirements->size = image->size;
3664 pMemoryRequirements->alignment = image->alignment;
3665 pMemoryRequirements->memoryTypeBits = memory_types;
3666 }
3667
3668 void anv_GetImageMemoryRequirements2(
3669 VkDevice _device,
3670 const VkImageMemoryRequirementsInfo2* pInfo,
3671 VkMemoryRequirements2* pMemoryRequirements)
3672 {
3673 ANV_FROM_HANDLE(anv_device, device, _device);
3674 ANV_FROM_HANDLE(anv_image, image, pInfo->image);
3675
3676 anv_GetImageMemoryRequirements(_device, pInfo->image,
3677 &pMemoryRequirements->memoryRequirements);
3678
3679 vk_foreach_struct_const(ext, pInfo->pNext) {
3680 switch (ext->sType) {
3681 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO: {
3682 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3683 const VkImagePlaneMemoryRequirementsInfo *plane_reqs =
3684 (const VkImagePlaneMemoryRequirementsInfo *) ext;
3685 uint32_t plane = anv_image_aspect_to_plane(image->aspects,
3686 plane_reqs->planeAspect);
3687
3688 assert(image->planes[plane].offset == 0);
3689
3690 /* The Vulkan spec (git aaed022) says:
3691 *
3692 * memoryTypeBits is a bitfield and contains one bit set for every
3693 * supported memory type for the resource. The bit `1<<i` is set
3694 * if and only if the memory type `i` in the
3695 * VkPhysicalDeviceMemoryProperties structure for the physical
3696 * device is supported.
3697 *
3698 * All types are currently supported for images.
3699 */
3700 pMemoryRequirements->memoryRequirements.memoryTypeBits =
3701 (1ull << pdevice->memory.type_count) - 1;
3702
3703 /* We must have image allocated or imported at this point. According to the
3704 * specification, external images must have been bound to memory before
3705 * calling GetImageMemoryRequirements.
3706 */
3707 assert(image->planes[plane].size > 0);
3708
3709 pMemoryRequirements->memoryRequirements.size = image->planes[plane].size;
3710 pMemoryRequirements->memoryRequirements.alignment =
3711 image->planes[plane].alignment;
3712 break;
3713 }
3714
3715 default:
3716 anv_debug_ignored_stype(ext->sType);
3717 break;
3718 }
3719 }
3720
3721 vk_foreach_struct(ext, pMemoryRequirements->pNext) {
3722 switch (ext->sType) {
3723 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
3724 VkMemoryDedicatedRequirements *requirements = (void *)ext;
3725 if (image->needs_set_tiling || image->external_format) {
3726 /* If we need to set the tiling for external consumers, we need a
3727 * dedicated allocation.
3728 *
3729 * See also anv_AllocateMemory.
3730 */
3731 requirements->prefersDedicatedAllocation = true;
3732 requirements->requiresDedicatedAllocation = true;
3733 } else {
3734 requirements->prefersDedicatedAllocation = false;
3735 requirements->requiresDedicatedAllocation = false;
3736 }
3737 break;
3738 }
3739
3740 default:
3741 anv_debug_ignored_stype(ext->sType);
3742 break;
3743 }
3744 }
3745 }
3746
3747 void anv_GetImageSparseMemoryRequirements(
3748 VkDevice device,
3749 VkImage image,
3750 uint32_t* pSparseMemoryRequirementCount,
3751 VkSparseImageMemoryRequirements* pSparseMemoryRequirements)
3752 {
3753 *pSparseMemoryRequirementCount = 0;
3754 }
3755
3756 void anv_GetImageSparseMemoryRequirements2(
3757 VkDevice device,
3758 const VkImageSparseMemoryRequirementsInfo2* pInfo,
3759 uint32_t* pSparseMemoryRequirementCount,
3760 VkSparseImageMemoryRequirements2* pSparseMemoryRequirements)
3761 {
3762 *pSparseMemoryRequirementCount = 0;
3763 }
3764
3765 void anv_GetDeviceMemoryCommitment(
3766 VkDevice device,
3767 VkDeviceMemory memory,
3768 VkDeviceSize* pCommittedMemoryInBytes)
3769 {
3770 *pCommittedMemoryInBytes = 0;
3771 }
3772
3773 static void
3774 anv_bind_buffer_memory(const VkBindBufferMemoryInfo *pBindInfo)
3775 {
3776 ANV_FROM_HANDLE(anv_device_memory, mem, pBindInfo->memory);
3777 ANV_FROM_HANDLE(anv_buffer, buffer, pBindInfo->buffer);
3778
3779 assert(pBindInfo->sType == VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO);
3780
3781 if (mem) {
3782 buffer->address = (struct anv_address) {
3783 .bo = mem->bo,
3784 .offset = pBindInfo->memoryOffset,
3785 };
3786 } else {
3787 buffer->address = ANV_NULL_ADDRESS;
3788 }
3789 }
3790
3791 VkResult anv_BindBufferMemory(
3792 VkDevice device,
3793 VkBuffer buffer,
3794 VkDeviceMemory memory,
3795 VkDeviceSize memoryOffset)
3796 {
3797 anv_bind_buffer_memory(
3798 &(VkBindBufferMemoryInfo) {
3799 .sType = VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO,
3800 .buffer = buffer,
3801 .memory = memory,
3802 .memoryOffset = memoryOffset,
3803 });
3804
3805 return VK_SUCCESS;
3806 }
3807
3808 VkResult anv_BindBufferMemory2(
3809 VkDevice device,
3810 uint32_t bindInfoCount,
3811 const VkBindBufferMemoryInfo* pBindInfos)
3812 {
3813 for (uint32_t i = 0; i < bindInfoCount; i++)
3814 anv_bind_buffer_memory(&pBindInfos[i]);
3815
3816 return VK_SUCCESS;
3817 }
3818
3819 VkResult anv_QueueBindSparse(
3820 VkQueue _queue,
3821 uint32_t bindInfoCount,
3822 const VkBindSparseInfo* pBindInfo,
3823 VkFence fence)
3824 {
3825 ANV_FROM_HANDLE(anv_queue, queue, _queue);
3826 if (anv_device_is_lost(queue->device))
3827 return VK_ERROR_DEVICE_LOST;
3828
3829 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
3830 }
3831
3832 // Event functions
3833
3834 VkResult anv_CreateEvent(
3835 VkDevice _device,
3836 const VkEventCreateInfo* pCreateInfo,
3837 const VkAllocationCallbacks* pAllocator,
3838 VkEvent* pEvent)
3839 {
3840 ANV_FROM_HANDLE(anv_device, device, _device);
3841 struct anv_state state;
3842 struct anv_event *event;
3843
3844 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_EVENT_CREATE_INFO);
3845
3846 state = anv_state_pool_alloc(&device->dynamic_state_pool,
3847 sizeof(*event), 8);
3848 event = state.map;
3849 event->state = state;
3850 event->semaphore = VK_EVENT_RESET;
3851
3852 if (!device->info.has_llc) {
3853 /* Make sure the writes we're flushing have landed. */
3854 __builtin_ia32_mfence();
3855 __builtin_ia32_clflush(event);
3856 }
3857
3858 *pEvent = anv_event_to_handle(event);
3859
3860 return VK_SUCCESS;
3861 }
3862
3863 void anv_DestroyEvent(
3864 VkDevice _device,
3865 VkEvent _event,
3866 const VkAllocationCallbacks* pAllocator)
3867 {
3868 ANV_FROM_HANDLE(anv_device, device, _device);
3869 ANV_FROM_HANDLE(anv_event, event, _event);
3870
3871 if (!event)
3872 return;
3873
3874 anv_state_pool_free(&device->dynamic_state_pool, event->state);
3875 }
3876
3877 VkResult anv_GetEventStatus(
3878 VkDevice _device,
3879 VkEvent _event)
3880 {
3881 ANV_FROM_HANDLE(anv_device, device, _device);
3882 ANV_FROM_HANDLE(anv_event, event, _event);
3883
3884 if (anv_device_is_lost(device))
3885 return VK_ERROR_DEVICE_LOST;
3886
3887 if (!device->info.has_llc) {
3888 /* Invalidate read cache before reading event written by GPU. */
3889 __builtin_ia32_clflush(event);
3890 __builtin_ia32_mfence();
3891
3892 }
3893
3894 return event->semaphore;
3895 }
3896
3897 VkResult anv_SetEvent(
3898 VkDevice _device,
3899 VkEvent _event)
3900 {
3901 ANV_FROM_HANDLE(anv_device, device, _device);
3902 ANV_FROM_HANDLE(anv_event, event, _event);
3903
3904 event->semaphore = VK_EVENT_SET;
3905
3906 if (!device->info.has_llc) {
3907 /* Make sure the writes we're flushing have landed. */
3908 __builtin_ia32_mfence();
3909 __builtin_ia32_clflush(event);
3910 }
3911
3912 return VK_SUCCESS;
3913 }
3914
3915 VkResult anv_ResetEvent(
3916 VkDevice _device,
3917 VkEvent _event)
3918 {
3919 ANV_FROM_HANDLE(anv_device, device, _device);
3920 ANV_FROM_HANDLE(anv_event, event, _event);
3921
3922 event->semaphore = VK_EVENT_RESET;
3923
3924 if (!device->info.has_llc) {
3925 /* Make sure the writes we're flushing have landed. */
3926 __builtin_ia32_mfence();
3927 __builtin_ia32_clflush(event);
3928 }
3929
3930 return VK_SUCCESS;
3931 }
3932
3933 // Buffer functions
3934
3935 VkResult anv_CreateBuffer(
3936 VkDevice _device,
3937 const VkBufferCreateInfo* pCreateInfo,
3938 const VkAllocationCallbacks* pAllocator,
3939 VkBuffer* pBuffer)
3940 {
3941 ANV_FROM_HANDLE(anv_device, device, _device);
3942 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3943 struct anv_buffer *buffer;
3944
3945 /* Don't allow creating buffers bigger than our address space. The real
3946 * issue here is that we may align up the buffer size and we don't want
3947 * doing so to cause roll-over. However, no one has any business
3948 * allocating a buffer larger than our GTT size.
3949 */
3950 if (pCreateInfo->size > pdevice->gtt_size)
3951 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
3952
3953 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
3954
3955 buffer = vk_alloc2(&device->alloc, pAllocator, sizeof(*buffer), 8,
3956 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
3957 if (buffer == NULL)
3958 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
3959
3960 buffer->size = pCreateInfo->size;
3961 buffer->usage = pCreateInfo->usage;
3962 buffer->address = ANV_NULL_ADDRESS;
3963
3964 *pBuffer = anv_buffer_to_handle(buffer);
3965
3966 return VK_SUCCESS;
3967 }
3968
3969 void anv_DestroyBuffer(
3970 VkDevice _device,
3971 VkBuffer _buffer,
3972 const VkAllocationCallbacks* pAllocator)
3973 {
3974 ANV_FROM_HANDLE(anv_device, device, _device);
3975 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
3976
3977 if (!buffer)
3978 return;
3979
3980 vk_free2(&device->alloc, pAllocator, buffer);
3981 }
3982
3983 VkDeviceAddress anv_GetBufferDeviceAddressKHR(
3984 VkDevice device,
3985 const VkBufferDeviceAddressInfoKHR* pInfo)
3986 {
3987 ANV_FROM_HANDLE(anv_buffer, buffer, pInfo->buffer);
3988
3989 assert(!anv_address_is_null(buffer->address));
3990 assert(buffer->address.bo->flags & EXEC_OBJECT_PINNED);
3991
3992 return anv_address_physical(buffer->address);
3993 }
3994
3995 uint64_t anv_GetBufferOpaqueCaptureAddressKHR(
3996 VkDevice device,
3997 const VkBufferDeviceAddressInfoKHR* pInfo)
3998 {
3999 return 0;
4000 }
4001
4002 uint64_t anv_GetDeviceMemoryOpaqueCaptureAddressKHR(
4003 VkDevice device,
4004 const VkDeviceMemoryOpaqueCaptureAddressInfoKHR* pInfo)
4005 {
4006 ANV_FROM_HANDLE(anv_device_memory, memory, pInfo->memory);
4007
4008 assert(memory->bo->flags & EXEC_OBJECT_PINNED);
4009 assert(memory->bo->has_client_visible_address);
4010
4011 return gen_48b_address(memory->bo->offset);
4012 }
4013
4014 void
4015 anv_fill_buffer_surface_state(struct anv_device *device, struct anv_state state,
4016 enum isl_format format,
4017 struct anv_address address,
4018 uint32_t range, uint32_t stride)
4019 {
4020 isl_buffer_fill_state(&device->isl_dev, state.map,
4021 .address = anv_address_physical(address),
4022 .mocs = device->isl_dev.mocs.internal,
4023 .size_B = range,
4024 .format = format,
4025 .swizzle = ISL_SWIZZLE_IDENTITY,
4026 .stride_B = stride);
4027 }
4028
4029 void anv_DestroySampler(
4030 VkDevice _device,
4031 VkSampler _sampler,
4032 const VkAllocationCallbacks* pAllocator)
4033 {
4034 ANV_FROM_HANDLE(anv_device, device, _device);
4035 ANV_FROM_HANDLE(anv_sampler, sampler, _sampler);
4036
4037 if (!sampler)
4038 return;
4039
4040 if (sampler->bindless_state.map) {
4041 anv_state_pool_free(&device->dynamic_state_pool,
4042 sampler->bindless_state);
4043 }
4044
4045 vk_free2(&device->alloc, pAllocator, sampler);
4046 }
4047
4048 VkResult anv_CreateFramebuffer(
4049 VkDevice _device,
4050 const VkFramebufferCreateInfo* pCreateInfo,
4051 const VkAllocationCallbacks* pAllocator,
4052 VkFramebuffer* pFramebuffer)
4053 {
4054 ANV_FROM_HANDLE(anv_device, device, _device);
4055 struct anv_framebuffer *framebuffer;
4056
4057 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
4058
4059 size_t size = sizeof(*framebuffer);
4060
4061 /* VK_KHR_imageless_framebuffer extension says:
4062 *
4063 * If flags includes VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR,
4064 * parameter pAttachments is ignored.
4065 */
4066 if (!(pCreateInfo->flags & VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR)) {
4067 size += sizeof(struct anv_image_view *) * pCreateInfo->attachmentCount;
4068 framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
4069 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
4070 if (framebuffer == NULL)
4071 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
4072
4073 for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
4074 ANV_FROM_HANDLE(anv_image_view, iview, pCreateInfo->pAttachments[i]);
4075 framebuffer->attachments[i] = iview;
4076 }
4077 framebuffer->attachment_count = pCreateInfo->attachmentCount;
4078 } else {
4079 assert(device->enabled_extensions.KHR_imageless_framebuffer);
4080 framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
4081 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
4082 if (framebuffer == NULL)
4083 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
4084
4085 framebuffer->attachment_count = 0;
4086 }
4087
4088 framebuffer->width = pCreateInfo->width;
4089 framebuffer->height = pCreateInfo->height;
4090 framebuffer->layers = pCreateInfo->layers;
4091
4092 *pFramebuffer = anv_framebuffer_to_handle(framebuffer);
4093
4094 return VK_SUCCESS;
4095 }
4096
4097 void anv_DestroyFramebuffer(
4098 VkDevice _device,
4099 VkFramebuffer _fb,
4100 const VkAllocationCallbacks* pAllocator)
4101 {
4102 ANV_FROM_HANDLE(anv_device, device, _device);
4103 ANV_FROM_HANDLE(anv_framebuffer, fb, _fb);
4104
4105 if (!fb)
4106 return;
4107
4108 vk_free2(&device->alloc, pAllocator, fb);
4109 }
4110
4111 static const VkTimeDomainEXT anv_time_domains[] = {
4112 VK_TIME_DOMAIN_DEVICE_EXT,
4113 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT,
4114 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT,
4115 };
4116
4117 VkResult anv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
4118 VkPhysicalDevice physicalDevice,
4119 uint32_t *pTimeDomainCount,
4120 VkTimeDomainEXT *pTimeDomains)
4121 {
4122 int d;
4123 VK_OUTARRAY_MAKE(out, pTimeDomains, pTimeDomainCount);
4124
4125 for (d = 0; d < ARRAY_SIZE(anv_time_domains); d++) {
4126 vk_outarray_append(&out, i) {
4127 *i = anv_time_domains[d];
4128 }
4129 }
4130
4131 return vk_outarray_status(&out);
4132 }
4133
4134 static uint64_t
4135 anv_clock_gettime(clockid_t clock_id)
4136 {
4137 struct timespec current;
4138 int ret;
4139
4140 ret = clock_gettime(clock_id, &current);
4141 if (ret < 0 && clock_id == CLOCK_MONOTONIC_RAW)
4142 ret = clock_gettime(CLOCK_MONOTONIC, &current);
4143 if (ret < 0)
4144 return 0;
4145
4146 return (uint64_t) current.tv_sec * 1000000000ULL + current.tv_nsec;
4147 }
4148
4149 #define TIMESTAMP 0x2358
4150
4151 VkResult anv_GetCalibratedTimestampsEXT(
4152 VkDevice _device,
4153 uint32_t timestampCount,
4154 const VkCalibratedTimestampInfoEXT *pTimestampInfos,
4155 uint64_t *pTimestamps,
4156 uint64_t *pMaxDeviation)
4157 {
4158 ANV_FROM_HANDLE(anv_device, device, _device);
4159 uint64_t timestamp_frequency = device->info.timestamp_frequency;
4160 int ret;
4161 int d;
4162 uint64_t begin, end;
4163 uint64_t max_clock_period = 0;
4164
4165 begin = anv_clock_gettime(CLOCK_MONOTONIC_RAW);
4166
4167 for (d = 0; d < timestampCount; d++) {
4168 switch (pTimestampInfos[d].timeDomain) {
4169 case VK_TIME_DOMAIN_DEVICE_EXT:
4170 ret = anv_gem_reg_read(device, TIMESTAMP | 1,
4171 &pTimestamps[d]);
4172
4173 if (ret != 0) {
4174 return anv_device_set_lost(device, "Failed to read the TIMESTAMP "
4175 "register: %m");
4176 }
4177 uint64_t device_period = DIV_ROUND_UP(1000000000, timestamp_frequency);
4178 max_clock_period = MAX2(max_clock_period, device_period);
4179 break;
4180 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT:
4181 pTimestamps[d] = anv_clock_gettime(CLOCK_MONOTONIC);
4182 max_clock_period = MAX2(max_clock_period, 1);
4183 break;
4184
4185 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT:
4186 pTimestamps[d] = begin;
4187 break;
4188 default:
4189 pTimestamps[d] = 0;
4190 break;
4191 }
4192 }
4193
4194 end = anv_clock_gettime(CLOCK_MONOTONIC_RAW);
4195
4196 /*
4197 * The maximum deviation is the sum of the interval over which we
4198 * perform the sampling and the maximum period of any sampled
4199 * clock. That's because the maximum skew between any two sampled
4200 * clock edges is when the sampled clock with the largest period is
4201 * sampled at the end of that period but right at the beginning of the
4202 * sampling interval and some other clock is sampled right at the
4203 * begining of its sampling period and right at the end of the
4204 * sampling interval. Let's assume the GPU has the longest clock
4205 * period and that the application is sampling GPU and monotonic:
4206 *
4207 * s e
4208 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
4209 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4210 *
4211 * g
4212 * 0 1 2 3
4213 * GPU -----_____-----_____-----_____-----_____
4214 *
4215 * m
4216 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
4217 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4218 *
4219 * Interval <----------------->
4220 * Deviation <-------------------------->
4221 *
4222 * s = read(raw) 2
4223 * g = read(GPU) 1
4224 * m = read(monotonic) 2
4225 * e = read(raw) b
4226 *
4227 * We round the sample interval up by one tick to cover sampling error
4228 * in the interval clock
4229 */
4230
4231 uint64_t sample_interval = end - begin + 1;
4232
4233 *pMaxDeviation = sample_interval + max_clock_period;
4234
4235 return VK_SUCCESS;
4236 }
4237
4238 /* vk_icd.h does not declare this function, so we declare it here to
4239 * suppress Wmissing-prototypes.
4240 */
4241 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4242 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion);
4243
4244 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4245 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion)
4246 {
4247 /* For the full details on loader interface versioning, see
4248 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
4249 * What follows is a condensed summary, to help you navigate the large and
4250 * confusing official doc.
4251 *
4252 * - Loader interface v0 is incompatible with later versions. We don't
4253 * support it.
4254 *
4255 * - In loader interface v1:
4256 * - The first ICD entrypoint called by the loader is
4257 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
4258 * entrypoint.
4259 * - The ICD must statically expose no other Vulkan symbol unless it is
4260 * linked with -Bsymbolic.
4261 * - Each dispatchable Vulkan handle created by the ICD must be
4262 * a pointer to a struct whose first member is VK_LOADER_DATA. The
4263 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
4264 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
4265 * vkDestroySurfaceKHR(). The ICD must be capable of working with
4266 * such loader-managed surfaces.
4267 *
4268 * - Loader interface v2 differs from v1 in:
4269 * - The first ICD entrypoint called by the loader is
4270 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
4271 * statically expose this entrypoint.
4272 *
4273 * - Loader interface v3 differs from v2 in:
4274 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
4275 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
4276 * because the loader no longer does so.
4277 *
4278 * - Loader interface v4 differs from v3 in:
4279 * - The ICD must implement vk_icdGetPhysicalDeviceProcAddr().
4280 */
4281 *pSupportedVersion = MIN2(*pSupportedVersion, 4u);
4282 return VK_SUCCESS;
4283 }