projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
anv/android: disable shared representable image support explicitly
[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/strtod.h"
36 #include "util/debug.h"
37 #include "util/build_id.h"
38 #include "util/disk_cache.h"
39 #include "util/mesa-sha1.h"
40 #include "util/os_file.h"
41 #include "util/u_atomic.h"
42 #include "util/u_string.h"
43 #include "git_sha1.h"
44 #include "vk_util.h"
45 #include "common/gen_defines.h"
46 #include "compiler/glsl_types.h"
47
48 #include "genxml/gen7_pack.h"
49
50 /* This is probably far to big but it reflects the max size used for messages
51 * in OpenGLs KHR_debug.
52 */
53 #define MAX_DEBUG_MESSAGE_LENGTH 4096
54
55 static void
56 compiler_debug_log(void *data, const char *fmt, ...)
57 {
58 char str[MAX_DEBUG_MESSAGE_LENGTH];
59 struct anv_device *device = (struct anv_device *)data;
60
61 if (list_empty(&device->instance->debug_report_callbacks.callbacks))
62 return;
63
64 va_list args;
65 va_start(args, fmt);
66 (void) vsnprintf(str, MAX_DEBUG_MESSAGE_LENGTH, fmt, args);
67 va_end(args);
68
69 vk_debug_report(&device->instance->debug_report_callbacks,
70 VK_DEBUG_REPORT_DEBUG_BIT_EXT,
71 VK_DEBUG_REPORT_OBJECT_TYPE_UNKNOWN_EXT,
72 0, 0, 0, "anv", str);
73 }
74
75 static void
76 compiler_perf_log(void *data, const char *fmt, ...)
77 {
78 va_list args;
79 va_start(args, fmt);
80
81 if (unlikely(INTEL_DEBUG & DEBUG_PERF))
82 intel_logd_v(fmt, args);
83
84 va_end(args);
85 }
86
87 static uint64_t
88 anv_compute_heap_size(int fd, uint64_t gtt_size)
89 {
90 /* Query the total ram from the system */
91 struct sysinfo info;
92 sysinfo(&info);
93
94 uint64_t total_ram = (uint64_t)info.totalram * (uint64_t)info.mem_unit;
95
96 /* We don't want to burn too much ram with the GPU. If the user has 4GiB
97 * or less, we use at most half. If they have more than 4GiB, we use 3/4.
98 */
99 uint64_t available_ram;
100 if (total_ram <= 4ull * 1024ull * 1024ull * 1024ull)
101 available_ram = total_ram / 2;
102 else
103 available_ram = total_ram * 3 / 4;
104
105 /* We also want to leave some padding for things we allocate in the driver,
106 * so don't go over 3/4 of the GTT either.
107 */
108 uint64_t available_gtt = gtt_size * 3 / 4;
109
110 return MIN2(available_ram, available_gtt);
111 }
112
113 static VkResult
114 anv_physical_device_init_heaps(struct anv_physical_device *device, int fd)
115 {
116 uint64_t gtt_size;
117 if (anv_gem_get_context_param(fd, 0, I915_CONTEXT_PARAM_GTT_SIZE,
118 &gtt_size) == -1) {
119 /* If, for whatever reason, we can't actually get the GTT size from the
120 * kernel (too old?) fall back to the aperture size.
121 */
122 anv_perf_warn(NULL, NULL,
123 "Failed to get I915_CONTEXT_PARAM_GTT_SIZE: %m");
124
125 if (anv_gem_get_aperture(fd, &gtt_size) == -1) {
126 return vk_errorf(NULL, NULL, VK_ERROR_INITIALIZATION_FAILED,
127 "failed to get aperture size: %m");
128 }
129 }
130
131 device->supports_48bit_addresses = (device->info.gen >= 8) &&
132 gtt_size > (4ULL << 30 /* GiB */);
133
134 uint64_t heap_size = anv_compute_heap_size(fd, gtt_size);
135
136 if (heap_size > (2ull << 30) && !device->supports_48bit_addresses) {
137 /* When running with an overridden PCI ID, we may get a GTT size from
138 * the kernel that is greater than 2 GiB but the execbuf check for 48bit
139 * address support can still fail. Just clamp the address space size to
140 * 2 GiB if we don't have 48-bit support.
141 */
142 intel_logw("%s:%d: The kernel reported a GTT size larger than 2 GiB but "
143 "not support for 48-bit addresses",
144 __FILE__, __LINE__);
145 heap_size = 2ull << 30;
146 }
147
148 if (heap_size <= 3ull * (1ull << 30)) {
149 /* In this case, everything fits nicely into the 32-bit address space,
150 * so there's no need for supporting 48bit addresses on client-allocated
151 * memory objects.
152 */
153 device->memory.heap_count = 1;
154 device->memory.heaps[0] = (struct anv_memory_heap) {
155 .vma_start = LOW_HEAP_MIN_ADDRESS,
156 .vma_size = LOW_HEAP_SIZE,
157 .size = heap_size,
158 .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
159 .supports_48bit_addresses = false,
160 };
161 } else {
162 /* Not everything will fit nicely into a 32-bit address space. In this
163 * case we need a 64-bit heap. Advertise a small 32-bit heap and a
164 * larger 48-bit heap. If we're in this case, then we have a total heap
165 * size larger than 3GiB which most likely means they have 8 GiB of
166 * video memory and so carving off 1 GiB for the 32-bit heap should be
167 * reasonable.
168 */
169 const uint64_t heap_size_32bit = 1ull << 30;
170 const uint64_t heap_size_48bit = heap_size - heap_size_32bit;
171
172 assert(device->supports_48bit_addresses);
173
174 device->memory.heap_count = 2;
175 device->memory.heaps[0] = (struct anv_memory_heap) {
176 .vma_start = HIGH_HEAP_MIN_ADDRESS,
177 /* Leave the last 4GiB out of the high vma range, so that no state
178 * base address + size can overflow 48 bits. For more information see
179 * the comment about Wa32bitGeneralStateOffset in anv_allocator.c
180 */
181 .vma_size = gtt_size - (1ull << 32) - HIGH_HEAP_MIN_ADDRESS,
182 .size = heap_size_48bit,
183 .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
184 .supports_48bit_addresses = true,
185 };
186 device->memory.heaps[1] = (struct anv_memory_heap) {
187 .vma_start = LOW_HEAP_MIN_ADDRESS,
188 .vma_size = LOW_HEAP_SIZE,
189 .size = heap_size_32bit,
190 .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
191 .supports_48bit_addresses = false,
192 };
193 }
194
195 uint32_t type_count = 0;
196 for (uint32_t heap = 0; heap < device->memory.heap_count; heap++) {
197 uint32_t valid_buffer_usage = ~0;
198
199 /* There appears to be a hardware issue in the VF cache where it only
200 * considers the bottom 32 bits of memory addresses. If you happen to
201 * have two vertex buffers which get placed exactly 4 GiB apart and use
202 * them in back-to-back draw calls, you can get collisions. In order to
203 * solve this problem, we require vertex and index buffers be bound to
204 * memory allocated out of the 32-bit heap.
205 */
206 if (device->memory.heaps[heap].supports_48bit_addresses) {
207 valid_buffer_usage &= ~(VK_BUFFER_USAGE_INDEX_BUFFER_BIT |
208 VK_BUFFER_USAGE_VERTEX_BUFFER_BIT);
209 }
210
211 if (device->info.has_llc) {
212 /* Big core GPUs share LLC with the CPU and thus one memory type can be
213 * both cached and coherent at the same time.
214 */
215 device->memory.types[type_count++] = (struct anv_memory_type) {
216 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
217 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
218 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
219 VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
220 .heapIndex = heap,
221 .valid_buffer_usage = valid_buffer_usage,
222 };
223 } else {
224 /* The spec requires that we expose a host-visible, coherent memory
225 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
226 * to give the application a choice between cached, but not coherent and
227 * coherent but uncached (WC though).
228 */
229 device->memory.types[type_count++] = (struct anv_memory_type) {
230 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
231 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
232 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
233 .heapIndex = heap,
234 .valid_buffer_usage = valid_buffer_usage,
235 };
236 device->memory.types[type_count++] = (struct anv_memory_type) {
237 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
238 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
239 VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
240 .heapIndex = heap,
241 .valid_buffer_usage = valid_buffer_usage,
242 };
243 }
244 }
245 device->memory.type_count = type_count;
246
247 return VK_SUCCESS;
248 }
249
250 static VkResult
251 anv_physical_device_init_uuids(struct anv_physical_device *device)
252 {
253 const struct build_id_note *note =
254 build_id_find_nhdr_for_addr(anv_physical_device_init_uuids);
255 if (!note) {
256 return vk_errorf(device->instance, device,
257 VK_ERROR_INITIALIZATION_FAILED,
258 "Failed to find build-id");
259 }
260
261 unsigned build_id_len = build_id_length(note);
262 if (build_id_len < 20) {
263 return vk_errorf(device->instance, device,
264 VK_ERROR_INITIALIZATION_FAILED,
265 "build-id too short. It needs to be a SHA");
266 }
267
268 memcpy(device->driver_build_sha1, build_id_data(note), 20);
269
270 struct mesa_sha1 sha1_ctx;
271 uint8_t sha1[20];
272 STATIC_ASSERT(VK_UUID_SIZE <= sizeof(sha1));
273
274 /* The pipeline cache UUID is used for determining when a pipeline cache is
275 * invalid. It needs both a driver build and the PCI ID of the device.
276 */
277 _mesa_sha1_init(&sha1_ctx);
278 _mesa_sha1_update(&sha1_ctx, build_id_data(note), build_id_len);
279 _mesa_sha1_update(&sha1_ctx, &device->chipset_id,
280 sizeof(device->chipset_id));
281 _mesa_sha1_update(&sha1_ctx, &device->always_use_bindless,
282 sizeof(device->always_use_bindless));
283 _mesa_sha1_update(&sha1_ctx, &device->has_a64_buffer_access,
284 sizeof(device->has_a64_buffer_access));
285 _mesa_sha1_update(&sha1_ctx, &device->has_bindless_images,
286 sizeof(device->has_bindless_images));
287 _mesa_sha1_update(&sha1_ctx, &device->has_bindless_samplers,
288 sizeof(device->has_bindless_samplers));
289 _mesa_sha1_final(&sha1_ctx, sha1);
290 memcpy(device->pipeline_cache_uuid, sha1, VK_UUID_SIZE);
291
292 /* The driver UUID is used for determining sharability of images and memory
293 * between two Vulkan instances in separate processes. People who want to
294 * share memory need to also check the device UUID (below) so all this
295 * needs to be is the build-id.
296 */
297 memcpy(device->driver_uuid, build_id_data(note), VK_UUID_SIZE);
298
299 /* The device UUID uniquely identifies the given device within the machine.
300 * Since we never have more than one device, this doesn't need to be a real
301 * UUID. However, on the off-chance that someone tries to use this to
302 * cache pre-tiled images or something of the like, we use the PCI ID and
303 * some bits of ISL info to ensure that this is safe.
304 */
305 _mesa_sha1_init(&sha1_ctx);
306 _mesa_sha1_update(&sha1_ctx, &device->chipset_id,
307 sizeof(device->chipset_id));
308 _mesa_sha1_update(&sha1_ctx, &device->isl_dev.has_bit6_swizzling,
309 sizeof(device->isl_dev.has_bit6_swizzling));
310 _mesa_sha1_final(&sha1_ctx, sha1);
311 memcpy(device->device_uuid, sha1, VK_UUID_SIZE);
312
313 return VK_SUCCESS;
314 }
315
316 static void
317 anv_physical_device_init_disk_cache(struct anv_physical_device *device)
318 {
319 #ifdef ENABLE_SHADER_CACHE
320 char renderer[10];
321 ASSERTED int len = snprintf(renderer, sizeof(renderer), "anv_%04x",
322 device->chipset_id);
323 assert(len == sizeof(renderer) - 2);
324
325 char timestamp[41];
326 _mesa_sha1_format(timestamp, device->driver_build_sha1);
327
328 const uint64_t driver_flags =
329 brw_get_compiler_config_value(device->compiler);
330 device->disk_cache = disk_cache_create(renderer, timestamp, driver_flags);
331 #else
332 device->disk_cache = NULL;
333 #endif
334 }
335
336 static void
337 anv_physical_device_free_disk_cache(struct anv_physical_device *device)
338 {
339 #ifdef ENABLE_SHADER_CACHE
340 if (device->disk_cache)
341 disk_cache_destroy(device->disk_cache);
342 #else
343 assert(device->disk_cache == NULL);
344 #endif
345 }
346
347 static uint64_t
348 get_available_system_memory()
349 {
350 char *meminfo = os_read_file("/proc/meminfo");
351 if (!meminfo)
352 return 0;
353
354 char *str = strstr(meminfo, "MemAvailable:");
355 if (!str) {
356 free(meminfo);
357 return 0;
358 }
359
360 uint64_t kb_mem_available;
361 if (sscanf(str, "MemAvailable: %" PRIx64, &kb_mem_available) == 1) {
362 free(meminfo);
363 return kb_mem_available << 10;
364 }
365
366 free(meminfo);
367 return 0;
368 }
369
370 static VkResult
371 anv_physical_device_init(struct anv_physical_device *device,
372 struct anv_instance *instance,
373 drmDevicePtr drm_device)
374 {
375 const char *primary_path = drm_device->nodes[DRM_NODE_PRIMARY];
376 const char *path = drm_device->nodes[DRM_NODE_RENDER];
377 VkResult result;
378 int fd;
379 int master_fd = -1;
380
381 brw_process_intel_debug_variable();
382
383 fd = open(path, O_RDWR | O_CLOEXEC);
384 if (fd < 0)
385 return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
386
387 device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
388 device->instance = instance;
389
390 assert(strlen(path) < ARRAY_SIZE(device->path));
391 snprintf(device->path, ARRAY_SIZE(device->path), "%s", path);
392
393 if (!gen_get_device_info_from_fd(fd, &device->info)) {
394 result = vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
395 goto fail;
396 }
397 device->chipset_id = device->info.chipset_id;
398 device->no_hw = device->info.no_hw;
399
400 if (getenv("INTEL_NO_HW") != NULL)
401 device->no_hw = true;
402
403 device->pci_info.domain = drm_device->businfo.pci->domain;
404 device->pci_info.bus = drm_device->businfo.pci->bus;
405 device->pci_info.device = drm_device->businfo.pci->dev;
406 device->pci_info.function = drm_device->businfo.pci->func;
407
408 device->name = gen_get_device_name(device->chipset_id);
409
410 if (device->info.is_haswell) {
411 intel_logw("Haswell Vulkan support is incomplete");
412 } else if (device->info.gen == 7 && !device->info.is_baytrail) {
413 intel_logw("Ivy Bridge Vulkan support is incomplete");
414 } else if (device->info.gen == 7 && device->info.is_baytrail) {
415 intel_logw("Bay Trail Vulkan support is incomplete");
416 } else if (device->info.gen >= 8 && device->info.gen <= 11) {
417 /* Gen8-11 fully supported */
418 } else {
419 result = vk_errorf(device->instance, device,
420 VK_ERROR_INCOMPATIBLE_DRIVER,
421 "Vulkan not yet supported on %s", device->name);
422 goto fail;
423 }
424
425 device->cmd_parser_version = -1;
426 if (device->info.gen == 7) {
427 device->cmd_parser_version =
428 anv_gem_get_param(fd, I915_PARAM_CMD_PARSER_VERSION);
429 if (device->cmd_parser_version == -1) {
430 result = vk_errorf(device->instance, device,
431 VK_ERROR_INITIALIZATION_FAILED,
432 "failed to get command parser version");
433 goto fail;
434 }
435 }
436
437 if (!anv_gem_get_param(fd, I915_PARAM_HAS_WAIT_TIMEOUT)) {
438 result = vk_errorf(device->instance, device,
439 VK_ERROR_INITIALIZATION_FAILED,
440 "kernel missing gem wait");
441 goto fail;
442 }
443
444 if (!anv_gem_get_param(fd, I915_PARAM_HAS_EXECBUF2)) {
445 result = vk_errorf(device->instance, device,
446 VK_ERROR_INITIALIZATION_FAILED,
447 "kernel missing execbuf2");
448 goto fail;
449 }
450
451 if (!device->info.has_llc &&
452 anv_gem_get_param(fd, I915_PARAM_MMAP_VERSION) < 1) {
453 result = vk_errorf(device->instance, device,
454 VK_ERROR_INITIALIZATION_FAILED,
455 "kernel missing wc mmap");
456 goto fail;
457 }
458
459 result = anv_physical_device_init_heaps(device, fd);
460 if (result != VK_SUCCESS)
461 goto fail;
462
463 device->has_exec_async = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_ASYNC);
464 device->has_exec_capture = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_CAPTURE);
465 device->has_exec_fence = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_FENCE);
466 device->has_syncobj = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_FENCE_ARRAY);
467 device->has_syncobj_wait = device->has_syncobj &&
468 anv_gem_supports_syncobj_wait(fd);
469 device->has_context_priority = anv_gem_has_context_priority(fd);
470
471 device->use_softpin = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_SOFTPIN)
472 && device->supports_48bit_addresses;
473
474 device->has_context_isolation =
475 anv_gem_get_param(fd, I915_PARAM_HAS_CONTEXT_ISOLATION);
476
477 device->always_use_bindless =
478 env_var_as_boolean("ANV_ALWAYS_BINDLESS", false);
479
480 /* We first got the A64 messages on broadwell and we can only use them if
481 * we can pass addresses directly into the shader which requires softpin.
482 */
483 device->has_a64_buffer_access = device->info.gen >= 8 &&
484 device->use_softpin;
485
486 /* We first get bindless image access on Skylake and we can only really do
487 * it if we don't have any relocations so we need softpin.
488 */
489 device->has_bindless_images = device->info.gen >= 9 &&
490 device->use_softpin;
491
492 /* We've had bindless samplers since Ivy Bridge (forever in Vulkan terms)
493 * because it's just a matter of setting the sampler address in the sample
494 * message header. However, we've not bothered to wire it up for vec4 so
495 * we leave it disabled on gen7.
496 */
497 device->has_bindless_samplers = device->info.gen >= 8;
498
499 device->has_mem_available = get_available_system_memory() != 0;
500
501 /* Starting with Gen10, the timestamp frequency of the command streamer may
502 * vary from one part to another. We can query the value from the kernel.
503 */
504 if (device->info.gen >= 10) {
505 int timestamp_frequency =
506 anv_gem_get_param(fd, I915_PARAM_CS_TIMESTAMP_FREQUENCY);
507
508 if (timestamp_frequency < 0)
509 intel_logw("Kernel 4.16-rc1+ required to properly query CS timestamp frequency");
510 else
511 device->info.timestamp_frequency = timestamp_frequency;
512 }
513
514 /* GENs prior to 8 do not support EU/Subslice info */
515 if (device->info.gen >= 8) {
516 device->subslice_total = anv_gem_get_param(fd, I915_PARAM_SUBSLICE_TOTAL);
517 device->eu_total = anv_gem_get_param(fd, I915_PARAM_EU_TOTAL);
518
519 /* Without this information, we cannot get the right Braswell
520 * brandstrings, and we have to use conservative numbers for GPGPU on
521 * many platforms, but otherwise, things will just work.
522 */
523 if (device->subslice_total < 1 || device->eu_total < 1) {
524 intel_logw("Kernel 4.1 required to properly query GPU properties");
525 }
526 } else if (device->info.gen == 7) {
527 device->subslice_total = 1 << (device->info.gt - 1);
528 }
529
530 if (device->info.is_cherryview &&
531 device->subslice_total > 0 && device->eu_total > 0) {
532 /* Logical CS threads = EUs per subslice * num threads per EU */
533 uint32_t max_cs_threads =
534 device->eu_total / device->subslice_total * device->info.num_thread_per_eu;
535
536 /* Fuse configurations may give more threads than expected, never less. */
537 if (max_cs_threads > device->info.max_cs_threads)
538 device->info.max_cs_threads = max_cs_threads;
539 }
540
541 device->compiler = brw_compiler_create(NULL, &device->info);
542 if (device->compiler == NULL) {
543 result = vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
544 goto fail;
545 }
546 device->compiler->shader_debug_log = compiler_debug_log;
547 device->compiler->shader_perf_log = compiler_perf_log;
548 device->compiler->supports_pull_constants = false;
549 device->compiler->constant_buffer_0_is_relative =
550 device->info.gen < 8 || !device->has_context_isolation;
551 device->compiler->supports_shader_constants = true;
552
553 /* Broadwell PRM says:
554 *
555 * "Before Gen8, there was a historical configuration control field to
556 * swizzle address bit[6] for in X/Y tiling modes. This was set in three
557 * different places: TILECTL[1:0], ARB_MODE[5:4], and
558 * DISP_ARB_CTL[14:13].
559 *
560 * For Gen8 and subsequent generations, the swizzle fields are all
561 * reserved, and the CPU's memory controller performs all address
562 * swizzling modifications."
563 */
564 bool swizzled =
565 device->info.gen < 8 && anv_gem_get_bit6_swizzle(fd, I915_TILING_X);
566
567 isl_device_init(&device->isl_dev, &device->info, swizzled);
568
569 result = anv_physical_device_init_uuids(device);
570 if (result != VK_SUCCESS)
571 goto fail;
572
573 anv_physical_device_init_disk_cache(device);
574
575 if (instance->enabled_extensions.KHR_display) {
576 master_fd = open(primary_path, O_RDWR | O_CLOEXEC);
577 if (master_fd >= 0) {
578 /* prod the device with a GETPARAM call which will fail if
579 * we don't have permission to even render on this device
580 */
581 if (anv_gem_get_param(master_fd, I915_PARAM_CHIPSET_ID) == 0) {
582 close(master_fd);
583 master_fd = -1;
584 }
585 }
586 }
587 device->master_fd = master_fd;
588
589 result = anv_init_wsi(device);
590 if (result != VK_SUCCESS) {
591 ralloc_free(device->compiler);
592 anv_physical_device_free_disk_cache(device);
593 goto fail;
594 }
595
596 anv_physical_device_get_supported_extensions(device,
597 &device->supported_extensions);
598
599
600 device->local_fd = fd;
601
602 return VK_SUCCESS;
603
604 fail:
605 close(fd);
606 if (master_fd != -1)
607 close(master_fd);
608 return result;
609 }
610
611 static void
612 anv_physical_device_finish(struct anv_physical_device *device)
613 {
614 anv_finish_wsi(device);
615 anv_physical_device_free_disk_cache(device);
616 ralloc_free(device->compiler);
617 close(device->local_fd);
618 if (device->master_fd >= 0)
619 close(device->master_fd);
620 }
621
622 static void *
623 default_alloc_func(void *pUserData, size_t size, size_t align,
624 VkSystemAllocationScope allocationScope)
625 {
626 return malloc(size);
627 }
628
629 static void *
630 default_realloc_func(void *pUserData, void *pOriginal, size_t size,
631 size_t align, VkSystemAllocationScope allocationScope)
632 {
633 return realloc(pOriginal, size);
634 }
635
636 static void
637 default_free_func(void *pUserData, void *pMemory)
638 {
639 free(pMemory);
640 }
641
642 static const VkAllocationCallbacks default_alloc = {
643 .pUserData = NULL,
644 .pfnAllocation = default_alloc_func,
645 .pfnReallocation = default_realloc_func,
646 .pfnFree = default_free_func,
647 };
648
649 VkResult anv_EnumerateInstanceExtensionProperties(
650 const char* pLayerName,
651 uint32_t* pPropertyCount,
652 VkExtensionProperties* pProperties)
653 {
654 VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
655
656 for (int i = 0; i < ANV_INSTANCE_EXTENSION_COUNT; i++) {
657 if (anv_instance_extensions_supported.extensions[i]) {
658 vk_outarray_append(&out, prop) {
659 *prop = anv_instance_extensions[i];
660 }
661 }
662 }
663
664 return vk_outarray_status(&out);
665 }
666
667 VkResult anv_CreateInstance(
668 const VkInstanceCreateInfo* pCreateInfo,
669 const VkAllocationCallbacks* pAllocator,
670 VkInstance* pInstance)
671 {
672 struct anv_instance *instance;
673 VkResult result;
674
675 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO);
676
677 struct anv_instance_extension_table enabled_extensions = {};
678 for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
679 int idx;
680 for (idx = 0; idx < ANV_INSTANCE_EXTENSION_COUNT; idx++) {
681 if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
682 anv_instance_extensions[idx].extensionName) == 0)
683 break;
684 }
685
686 if (idx >= ANV_INSTANCE_EXTENSION_COUNT)
687 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
688
689 if (!anv_instance_extensions_supported.extensions[idx])
690 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
691
692 enabled_extensions.extensions[idx] = true;
693 }
694
695 instance = vk_alloc2(&default_alloc, pAllocator, sizeof(*instance), 8,
696 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
697 if (!instance)
698 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
699
700 instance->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
701
702 if (pAllocator)
703 instance->alloc = *pAllocator;
704 else
705 instance->alloc = default_alloc;
706
707 instance->app_info = (struct anv_app_info) { .api_version = 0 };
708 if (pCreateInfo->pApplicationInfo) {
709 const VkApplicationInfo *app = pCreateInfo->pApplicationInfo;
710
711 instance->app_info.app_name =
712 vk_strdup(&instance->alloc, app->pApplicationName,
713 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
714 instance->app_info.app_version = app->applicationVersion;
715
716 instance->app_info.engine_name =
717 vk_strdup(&instance->alloc, app->pEngineName,
718 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
719 instance->app_info.engine_version = app->engineVersion;
720
721 instance->app_info.api_version = app->apiVersion;
722 }
723
724 if (instance->app_info.api_version == 0)
725 instance->app_info.api_version = VK_API_VERSION_1_0;
726
727 instance->enabled_extensions = enabled_extensions;
728
729 for (unsigned i = 0; i < ARRAY_SIZE(instance->dispatch.entrypoints); i++) {
730 /* Vulkan requires that entrypoints for extensions which have not been
731 * enabled must not be advertised.
732 */
733 if (!anv_instance_entrypoint_is_enabled(i, instance->app_info.api_version,
734 &instance->enabled_extensions)) {
735 instance->dispatch.entrypoints[i] = NULL;
736 } else {
737 instance->dispatch.entrypoints[i] =
738 anv_instance_dispatch_table.entrypoints[i];
739 }
740 }
741
742 for (unsigned i = 0; i < ARRAY_SIZE(instance->device_dispatch.entrypoints); i++) {
743 /* Vulkan requires that entrypoints for extensions which have not been
744 * enabled must not be advertised.
745 */
746 if (!anv_device_entrypoint_is_enabled(i, instance->app_info.api_version,
747 &instance->enabled_extensions, NULL)) {
748 instance->device_dispatch.entrypoints[i] = NULL;
749 } else {
750 instance->device_dispatch.entrypoints[i] =
751 anv_device_dispatch_table.entrypoints[i];
752 }
753 }
754
755 instance->physicalDeviceCount = -1;
756
757 result = vk_debug_report_instance_init(&instance->debug_report_callbacks);
758 if (result != VK_SUCCESS) {
759 vk_free2(&default_alloc, pAllocator, instance);
760 return vk_error(result);
761 }
762
763 instance->pipeline_cache_enabled =
764 env_var_as_boolean("ANV_ENABLE_PIPELINE_CACHE", true);
765
766 _mesa_locale_init();
767 glsl_type_singleton_init_or_ref();
768
769 VG(VALGRIND_CREATE_MEMPOOL(instance, 0, false));
770
771 *pInstance = anv_instance_to_handle(instance);
772
773 return VK_SUCCESS;
774 }
775
776 void anv_DestroyInstance(
777 VkInstance _instance,
778 const VkAllocationCallbacks* pAllocator)
779 {
780 ANV_FROM_HANDLE(anv_instance, instance, _instance);
781
782 if (!instance)
783 return;
784
785 if (instance->physicalDeviceCount > 0) {
786 /* We support at most one physical device. */
787 assert(instance->physicalDeviceCount == 1);
788 anv_physical_device_finish(&instance->physicalDevice);
789 }
790
791 vk_free(&instance->alloc, (char *)instance->app_info.app_name);
792 vk_free(&instance->alloc, (char *)instance->app_info.engine_name);
793
794 VG(VALGRIND_DESTROY_MEMPOOL(instance));
795
796 vk_debug_report_instance_destroy(&instance->debug_report_callbacks);
797
798 glsl_type_singleton_decref();
799 _mesa_locale_fini();
800
801 vk_free(&instance->alloc, instance);
802 }
803
804 static VkResult
805 anv_enumerate_devices(struct anv_instance *instance)
806 {
807 /* TODO: Check for more devices ? */
808 drmDevicePtr devices[8];
809 VkResult result = VK_ERROR_INCOMPATIBLE_DRIVER;
810 int max_devices;
811
812 instance->physicalDeviceCount = 0;
813
814 max_devices = drmGetDevices2(0, devices, ARRAY_SIZE(devices));
815 if (max_devices < 1)
816 return VK_ERROR_INCOMPATIBLE_DRIVER;
817
818 for (unsigned i = 0; i < (unsigned)max_devices; i++) {
819 if (devices[i]->available_nodes & 1 << DRM_NODE_RENDER &&
820 devices[i]->bustype == DRM_BUS_PCI &&
821 devices[i]->deviceinfo.pci->vendor_id == 0x8086) {
822
823 result = anv_physical_device_init(&instance->physicalDevice,
824 instance, devices[i]);
825 if (result != VK_ERROR_INCOMPATIBLE_DRIVER)
826 break;
827 }
828 }
829 drmFreeDevices(devices, max_devices);
830
831 if (result == VK_SUCCESS)
832 instance->physicalDeviceCount = 1;
833
834 return result;
835 }
836
837 static VkResult
838 anv_instance_ensure_physical_device(struct anv_instance *instance)
839 {
840 if (instance->physicalDeviceCount < 0) {
841 VkResult result = anv_enumerate_devices(instance);
842 if (result != VK_SUCCESS &&
843 result != VK_ERROR_INCOMPATIBLE_DRIVER)
844 return result;
845 }
846
847 return VK_SUCCESS;
848 }
849
850 VkResult anv_EnumeratePhysicalDevices(
851 VkInstance _instance,
852 uint32_t* pPhysicalDeviceCount,
853 VkPhysicalDevice* pPhysicalDevices)
854 {
855 ANV_FROM_HANDLE(anv_instance, instance, _instance);
856 VK_OUTARRAY_MAKE(out, pPhysicalDevices, pPhysicalDeviceCount);
857
858 VkResult result = anv_instance_ensure_physical_device(instance);
859 if (result != VK_SUCCESS)
860 return result;
861
862 if (instance->physicalDeviceCount == 0)
863 return VK_SUCCESS;
864
865 assert(instance->physicalDeviceCount == 1);
866 vk_outarray_append(&out, i) {
867 *i = anv_physical_device_to_handle(&instance->physicalDevice);
868 }
869
870 return vk_outarray_status(&out);
871 }
872
873 VkResult anv_EnumeratePhysicalDeviceGroups(
874 VkInstance _instance,
875 uint32_t* pPhysicalDeviceGroupCount,
876 VkPhysicalDeviceGroupProperties* pPhysicalDeviceGroupProperties)
877 {
878 ANV_FROM_HANDLE(anv_instance, instance, _instance);
879 VK_OUTARRAY_MAKE(out, pPhysicalDeviceGroupProperties,
880 pPhysicalDeviceGroupCount);
881
882 VkResult result = anv_instance_ensure_physical_device(instance);
883 if (result != VK_SUCCESS)
884 return result;
885
886 if (instance->physicalDeviceCount == 0)
887 return VK_SUCCESS;
888
889 assert(instance->physicalDeviceCount == 1);
890
891 vk_outarray_append(&out, p) {
892 p->physicalDeviceCount = 1;
893 memset(p->physicalDevices, 0, sizeof(p->physicalDevices));
894 p->physicalDevices[0] =
895 anv_physical_device_to_handle(&instance->physicalDevice);
896 p->subsetAllocation = false;
897
898 vk_foreach_struct(ext, p->pNext)
899 anv_debug_ignored_stype(ext->sType);
900 }
901
902 return vk_outarray_status(&out);
903 }
904
905 void anv_GetPhysicalDeviceFeatures(
906 VkPhysicalDevice physicalDevice,
907 VkPhysicalDeviceFeatures* pFeatures)
908 {
909 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
910
911 *pFeatures = (VkPhysicalDeviceFeatures) {
912 .robustBufferAccess = true,
913 .fullDrawIndexUint32 = true,
914 .imageCubeArray = true,
915 .independentBlend = true,
916 .geometryShader = true,
917 .tessellationShader = true,
918 .sampleRateShading = true,
919 .dualSrcBlend = true,
920 .logicOp = true,
921 .multiDrawIndirect = true,
922 .drawIndirectFirstInstance = true,
923 .depthClamp = true,
924 .depthBiasClamp = true,
925 .fillModeNonSolid = true,
926 .depthBounds = false,
927 .wideLines = true,
928 .largePoints = true,
929 .alphaToOne = true,
930 .multiViewport = true,
931 .samplerAnisotropy = true,
932 .textureCompressionETC2 = pdevice->info.gen >= 8 ||
933 pdevice->info.is_baytrail,
934 .textureCompressionASTC_LDR = pdevice->info.gen >= 9, /* FINISHME CHV */
935 .textureCompressionBC = true,
936 .occlusionQueryPrecise = true,
937 .pipelineStatisticsQuery = true,
938 .fragmentStoresAndAtomics = true,
939 .shaderTessellationAndGeometryPointSize = true,
940 .shaderImageGatherExtended = true,
941 .shaderStorageImageExtendedFormats = true,
942 .shaderStorageImageMultisample = false,
943 .shaderStorageImageReadWithoutFormat = false,
944 .shaderStorageImageWriteWithoutFormat = true,
945 .shaderUniformBufferArrayDynamicIndexing = true,
946 .shaderSampledImageArrayDynamicIndexing = true,
947 .shaderStorageBufferArrayDynamicIndexing = true,
948 .shaderStorageImageArrayDynamicIndexing = true,
949 .shaderClipDistance = true,
950 .shaderCullDistance = true,
951 .shaderFloat64 = pdevice->info.gen >= 8 &&
952 pdevice->info.has_64bit_types,
953 .shaderInt64 = pdevice->info.gen >= 8 &&
954 pdevice->info.has_64bit_types,
955 .shaderInt16 = pdevice->info.gen >= 8,
956 .shaderResourceMinLod = pdevice->info.gen >= 9,
957 .variableMultisampleRate = true,
958 .inheritedQueries = true,
959 };
960
961 /* We can't do image stores in vec4 shaders */
962 pFeatures->vertexPipelineStoresAndAtomics =
963 pdevice->compiler->scalar_stage[MESA_SHADER_VERTEX] &&
964 pdevice->compiler->scalar_stage[MESA_SHADER_GEOMETRY];
965
966 struct anv_app_info *app_info = &pdevice->instance->app_info;
967
968 /* The new DOOM and Wolfenstein games require depthBounds without
969 * checking for it. They seem to run fine without it so just claim it's
970 * there and accept the consequences.
971 */
972 if (app_info->engine_name && strcmp(app_info->engine_name, "idTech") == 0)
973 pFeatures->depthBounds = true;
974 }
975
976 void anv_GetPhysicalDeviceFeatures2(
977 VkPhysicalDevice physicalDevice,
978 VkPhysicalDeviceFeatures2* pFeatures)
979 {
980 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
981 anv_GetPhysicalDeviceFeatures(physicalDevice, &pFeatures->features);
982
983 vk_foreach_struct(ext, pFeatures->pNext) {
984 switch (ext->sType) {
985 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES_KHR: {
986 VkPhysicalDevice8BitStorageFeaturesKHR *features =
987 (VkPhysicalDevice8BitStorageFeaturesKHR *)ext;
988 features->storageBuffer8BitAccess = pdevice->info.gen >= 8;
989 features->uniformAndStorageBuffer8BitAccess = pdevice->info.gen >= 8;
990 features->storagePushConstant8 = pdevice->info.gen >= 8;
991 break;
992 }
993
994 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES: {
995 VkPhysicalDevice16BitStorageFeatures *features =
996 (VkPhysicalDevice16BitStorageFeatures *)ext;
997 features->storageBuffer16BitAccess = pdevice->info.gen >= 8;
998 features->uniformAndStorageBuffer16BitAccess = pdevice->info.gen >= 8;
999 features->storagePushConstant16 = pdevice->info.gen >= 8;
1000 features->storageInputOutput16 = false;
1001 break;
1002 }
1003
1004 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_EXT: {
1005 VkPhysicalDeviceBufferDeviceAddressFeaturesEXT *features = (void *)ext;
1006 features->bufferDeviceAddress = pdevice->has_a64_buffer_access;
1007 features->bufferDeviceAddressCaptureReplay = false;
1008 features->bufferDeviceAddressMultiDevice = false;
1009 break;
1010 }
1011
1012 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COMPUTE_SHADER_DERIVATIVES_FEATURES_NV: {
1013 VkPhysicalDeviceComputeShaderDerivativesFeaturesNV *features =
1014 (VkPhysicalDeviceComputeShaderDerivativesFeaturesNV *)ext;
1015 features->computeDerivativeGroupQuads = true;
1016 features->computeDerivativeGroupLinear = true;
1017 break;
1018 }
1019
1020 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT: {
1021 VkPhysicalDeviceConditionalRenderingFeaturesEXT *features =
1022 (VkPhysicalDeviceConditionalRenderingFeaturesEXT*)ext;
1023 features->conditionalRendering = pdevice->info.gen >= 8 ||
1024 pdevice->info.is_haswell;
1025 features->inheritedConditionalRendering = pdevice->info.gen >= 8 ||
1026 pdevice->info.is_haswell;
1027 break;
1028 }
1029
1030 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_CLIP_ENABLE_FEATURES_EXT: {
1031 VkPhysicalDeviceDepthClipEnableFeaturesEXT *features =
1032 (VkPhysicalDeviceDepthClipEnableFeaturesEXT *)ext;
1033 features->depthClipEnable = true;
1034 break;
1035 }
1036
1037 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR: {
1038 VkPhysicalDeviceFloat16Int8FeaturesKHR *features = (void *)ext;
1039 features->shaderFloat16 = pdevice->info.gen >= 8;
1040 features->shaderInt8 = pdevice->info.gen >= 8;
1041 break;
1042 }
1043
1044 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FRAGMENT_SHADER_INTERLOCK_FEATURES_EXT: {
1045 VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT *features =
1046 (VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT *)ext;
1047 features->fragmentShaderSampleInterlock = pdevice->info.gen >= 9;
1048 features->fragmentShaderPixelInterlock = pdevice->info.gen >= 9;
1049 features->fragmentShaderShadingRateInterlock = false;
1050 break;
1051 }
1052
1053 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_QUERY_RESET_FEATURES_EXT: {
1054 VkPhysicalDeviceHostQueryResetFeaturesEXT *features =
1055 (VkPhysicalDeviceHostQueryResetFeaturesEXT *)ext;
1056 features->hostQueryReset = true;
1057 break;
1058 }
1059
1060 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT: {
1061 VkPhysicalDeviceDescriptorIndexingFeaturesEXT *features =
1062 (VkPhysicalDeviceDescriptorIndexingFeaturesEXT *)ext;
1063 features->shaderInputAttachmentArrayDynamicIndexing = false;
1064 features->shaderUniformTexelBufferArrayDynamicIndexing = true;
1065 features->shaderStorageTexelBufferArrayDynamicIndexing = true;
1066 features->shaderUniformBufferArrayNonUniformIndexing = false;
1067 features->shaderSampledImageArrayNonUniformIndexing = true;
1068 features->shaderStorageBufferArrayNonUniformIndexing = true;
1069 features->shaderStorageImageArrayNonUniformIndexing = true;
1070 features->shaderInputAttachmentArrayNonUniformIndexing = false;
1071 features->shaderUniformTexelBufferArrayNonUniformIndexing = true;
1072 features->shaderStorageTexelBufferArrayNonUniformIndexing = true;
1073 features->descriptorBindingUniformBufferUpdateAfterBind = false;
1074 features->descriptorBindingSampledImageUpdateAfterBind = true;
1075 features->descriptorBindingStorageImageUpdateAfterBind = true;
1076 features->descriptorBindingStorageBufferUpdateAfterBind = true;
1077 features->descriptorBindingUniformTexelBufferUpdateAfterBind = true;
1078 features->descriptorBindingStorageTexelBufferUpdateAfterBind = true;
1079 features->descriptorBindingUpdateUnusedWhilePending = true;
1080 features->descriptorBindingPartiallyBound = true;
1081 features->descriptorBindingVariableDescriptorCount = false;
1082 features->runtimeDescriptorArray = true;
1083 break;
1084 }
1085
1086 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INDEX_TYPE_UINT8_FEATURES_EXT: {
1087 VkPhysicalDeviceIndexTypeUint8FeaturesEXT *features =
1088 (VkPhysicalDeviceIndexTypeUint8FeaturesEXT *)ext;
1089 features->indexTypeUint8 = true;
1090 break;
1091 }
1092
1093 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_FEATURES_EXT: {
1094 VkPhysicalDeviceInlineUniformBlockFeaturesEXT *features =
1095 (VkPhysicalDeviceInlineUniformBlockFeaturesEXT *)ext;
1096 features->inlineUniformBlock = true;
1097 features->descriptorBindingInlineUniformBlockUpdateAfterBind = true;
1098 break;
1099 }
1100
1101 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_FEATURES_EXT: {
1102 VkPhysicalDeviceLineRasterizationFeaturesEXT *features =
1103 (VkPhysicalDeviceLineRasterizationFeaturesEXT *)ext;
1104 features->rectangularLines = true;
1105 features->bresenhamLines = true;
1106 features->smoothLines = true;
1107 features->stippledRectangularLines = false;
1108 features->stippledBresenhamLines = true;
1109 features->stippledSmoothLines = false;
1110 }
1111
1112 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES: {
1113 VkPhysicalDeviceMultiviewFeatures *features =
1114 (VkPhysicalDeviceMultiviewFeatures *)ext;
1115 features->multiview = true;
1116 features->multiviewGeometryShader = true;
1117 features->multiviewTessellationShader = true;
1118 break;
1119 }
1120
1121 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGELESS_FRAMEBUFFER_FEATURES_KHR: {
1122 VkPhysicalDeviceImagelessFramebufferFeaturesKHR *features =
1123 (VkPhysicalDeviceImagelessFramebufferFeaturesKHR *)ext;
1124 features->imagelessFramebuffer = true;
1125 break;
1126 }
1127
1128 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES: {
1129 VkPhysicalDeviceProtectedMemoryFeatures *features = (void *)ext;
1130 features->protectedMemory = false;
1131 break;
1132 }
1133
1134 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES: {
1135 VkPhysicalDeviceSamplerYcbcrConversionFeatures *features =
1136 (VkPhysicalDeviceSamplerYcbcrConversionFeatures *) ext;
1137 features->samplerYcbcrConversion = true;
1138 break;
1139 }
1140
1141 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES_EXT: {
1142 VkPhysicalDeviceScalarBlockLayoutFeaturesEXT *features =
1143 (VkPhysicalDeviceScalarBlockLayoutFeaturesEXT *)ext;
1144 features->scalarBlockLayout = true;
1145 break;
1146 }
1147
1148 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_INT64_FEATURES_KHR: {
1149 VkPhysicalDeviceShaderAtomicInt64FeaturesKHR *features = (void *)ext;
1150 features->shaderBufferInt64Atomics =
1151 pdevice->info.gen >= 9 && pdevice->use_softpin;
1152 features->shaderSharedInt64Atomics = VK_FALSE;
1153 break;
1154 }
1155
1156 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DEMOTE_TO_HELPER_INVOCATION_FEATURES_EXT: {
1157 VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT *features = (void *)ext;
1158 features->shaderDemoteToHelperInvocation = true;
1159 break;
1160 }
1161
1162 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETERS_FEATURES: {
1163 VkPhysicalDeviceShaderDrawParametersFeatures *features = (void *)ext;
1164 features->shaderDrawParameters = true;
1165 break;
1166 }
1167
1168 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_FEATURES_EXT: {
1169 VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT *features =
1170 (VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT *)ext;
1171 features->texelBufferAlignment = true;
1172 break;
1173 }
1174
1175 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES: {
1176 VkPhysicalDeviceVariablePointersFeatures *features = (void *)ext;
1177 features->variablePointersStorageBuffer = true;
1178 features->variablePointers = true;
1179 break;
1180 }
1181
1182 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT: {
1183 VkPhysicalDeviceTransformFeedbackFeaturesEXT *features =
1184 (VkPhysicalDeviceTransformFeedbackFeaturesEXT *)ext;
1185 features->transformFeedback = true;
1186 features->geometryStreams = true;
1187 break;
1188 }
1189
1190 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_UNIFORM_BUFFER_STANDARD_LAYOUT_FEATURES_KHR: {
1191 VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR *features =
1192 (VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR *)ext;
1193 features->uniformBufferStandardLayout = true;
1194 break;
1195 }
1196
1197 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT: {
1198 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *features =
1199 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *)ext;
1200 features->vertexAttributeInstanceRateDivisor = true;
1201 features->vertexAttributeInstanceRateZeroDivisor = true;
1202 break;
1203 }
1204
1205 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_YCBCR_IMAGE_ARRAYS_FEATURES_EXT: {
1206 VkPhysicalDeviceYcbcrImageArraysFeaturesEXT *features =
1207 (VkPhysicalDeviceYcbcrImageArraysFeaturesEXT *)ext;
1208 features->ycbcrImageArrays = true;
1209 break;
1210 }
1211
1212 default:
1213 anv_debug_ignored_stype(ext->sType);
1214 break;
1215 }
1216 }
1217 }
1218
1219 #define MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS 64
1220
1221 #define MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS 64
1222 #define MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS 256
1223
1224 void anv_GetPhysicalDeviceProperties(
1225 VkPhysicalDevice physicalDevice,
1226 VkPhysicalDeviceProperties* pProperties)
1227 {
1228 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
1229 const struct gen_device_info *devinfo = &pdevice->info;
1230
1231 /* See assertions made when programming the buffer surface state. */
1232 const uint32_t max_raw_buffer_sz = devinfo->gen >= 7 ?
1233 (1ul << 30) : (1ul << 27);
1234
1235 const uint32_t max_ssbos = pdevice->has_a64_buffer_access ? UINT16_MAX : 64;
1236 const uint32_t max_textures =
1237 pdevice->has_bindless_images ? UINT16_MAX : 128;
1238 const uint32_t max_samplers =
1239 pdevice->has_bindless_samplers ? UINT16_MAX :
1240 (devinfo->gen >= 8 || devinfo->is_haswell) ? 128 : 16;
1241 const uint32_t max_images =
1242 pdevice->has_bindless_images ? UINT16_MAX : MAX_IMAGES;
1243
1244 /* The moment we have anything bindless, claim a high per-stage limit */
1245 const uint32_t max_per_stage =
1246 pdevice->has_a64_buffer_access ? UINT32_MAX :
1247 MAX_BINDING_TABLE_SIZE - MAX_RTS;
1248
1249 VkSampleCountFlags sample_counts =
1250 isl_device_get_sample_counts(&pdevice->isl_dev);
1251
1252
1253 VkPhysicalDeviceLimits limits = {
1254 .maxImageDimension1D = (1 << 14),
1255 .maxImageDimension2D = (1 << 14),
1256 .maxImageDimension3D = (1 << 11),
1257 .maxImageDimensionCube = (1 << 14),
1258 .maxImageArrayLayers = (1 << 11),
1259 .maxTexelBufferElements = 128 * 1024 * 1024,
1260 .maxUniformBufferRange = (1ul << 27),
1261 .maxStorageBufferRange = max_raw_buffer_sz,
1262 .maxPushConstantsSize = MAX_PUSH_CONSTANTS_SIZE,
1263 .maxMemoryAllocationCount = UINT32_MAX,
1264 .maxSamplerAllocationCount = 64 * 1024,
1265 .bufferImageGranularity = 64, /* A cache line */
1266 .sparseAddressSpaceSize = 0,
1267 .maxBoundDescriptorSets = MAX_SETS,
1268 .maxPerStageDescriptorSamplers = max_samplers,
1269 .maxPerStageDescriptorUniformBuffers = MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS,
1270 .maxPerStageDescriptorStorageBuffers = max_ssbos,
1271 .maxPerStageDescriptorSampledImages = max_textures,
1272 .maxPerStageDescriptorStorageImages = max_images,
1273 .maxPerStageDescriptorInputAttachments = MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS,
1274 .maxPerStageResources = max_per_stage,
1275 .maxDescriptorSetSamplers = 6 * max_samplers, /* number of stages * maxPerStageDescriptorSamplers */
1276 .maxDescriptorSetUniformBuffers = 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS, /* number of stages * maxPerStageDescriptorUniformBuffers */
1277 .maxDescriptorSetUniformBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2,
1278 .maxDescriptorSetStorageBuffers = 6 * max_ssbos, /* number of stages * maxPerStageDescriptorStorageBuffers */
1279 .maxDescriptorSetStorageBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2,
1280 .maxDescriptorSetSampledImages = 6 * max_textures, /* number of stages * maxPerStageDescriptorSampledImages */
1281 .maxDescriptorSetStorageImages = 6 * max_images, /* number of stages * maxPerStageDescriptorStorageImages */
1282 .maxDescriptorSetInputAttachments = MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS,
1283 .maxVertexInputAttributes = MAX_VBS,
1284 .maxVertexInputBindings = MAX_VBS,
1285 .maxVertexInputAttributeOffset = 2047,
1286 .maxVertexInputBindingStride = 2048,
1287 .maxVertexOutputComponents = 128,
1288 .maxTessellationGenerationLevel = 64,
1289 .maxTessellationPatchSize = 32,
1290 .maxTessellationControlPerVertexInputComponents = 128,
1291 .maxTessellationControlPerVertexOutputComponents = 128,
1292 .maxTessellationControlPerPatchOutputComponents = 128,
1293 .maxTessellationControlTotalOutputComponents = 2048,
1294 .maxTessellationEvaluationInputComponents = 128,
1295 .maxTessellationEvaluationOutputComponents = 128,
1296 .maxGeometryShaderInvocations = 32,
1297 .maxGeometryInputComponents = 64,
1298 .maxGeometryOutputComponents = 128,
1299 .maxGeometryOutputVertices = 256,
1300 .maxGeometryTotalOutputComponents = 1024,
1301 .maxFragmentInputComponents = 116, /* 128 components - (PSIZ, CLIP_DIST0, CLIP_DIST1) */
1302 .maxFragmentOutputAttachments = 8,
1303 .maxFragmentDualSrcAttachments = 1,
1304 .maxFragmentCombinedOutputResources = 8,
1305 .maxComputeSharedMemorySize = 64 * 1024,
1306 .maxComputeWorkGroupCount = { 65535, 65535, 65535 },
1307 .maxComputeWorkGroupInvocations = 32 * devinfo->max_cs_threads,
1308 .maxComputeWorkGroupSize = {
1309 16 * devinfo->max_cs_threads,
1310 16 * devinfo->max_cs_threads,
1311 16 * devinfo->max_cs_threads,
1312 },
1313 .subPixelPrecisionBits = 8,
1314 .subTexelPrecisionBits = 8,
1315 .mipmapPrecisionBits = 8,
1316 .maxDrawIndexedIndexValue = UINT32_MAX,
1317 .maxDrawIndirectCount = UINT32_MAX,
1318 .maxSamplerLodBias = 16,
1319 .maxSamplerAnisotropy = 16,
1320 .maxViewports = MAX_VIEWPORTS,
1321 .maxViewportDimensions = { (1 << 14), (1 << 14) },
1322 .viewportBoundsRange = { INT16_MIN, INT16_MAX },
1323 .viewportSubPixelBits = 13, /* We take a float? */
1324 .minMemoryMapAlignment = 4096, /* A page */
1325 /* The dataport requires texel alignment so we need to assume a worst
1326 * case of R32G32B32A32 which is 16 bytes.
1327 */
1328 .minTexelBufferOffsetAlignment = 16,
1329 /* We need 16 for UBO block reads to work and 32 for push UBOs */
1330 .minUniformBufferOffsetAlignment = 32,
1331 .minStorageBufferOffsetAlignment = 4,
1332 .minTexelOffset = -8,
1333 .maxTexelOffset = 7,
1334 .minTexelGatherOffset = -32,
1335 .maxTexelGatherOffset = 31,
1336 .minInterpolationOffset = -0.5,
1337 .maxInterpolationOffset = 0.4375,
1338 .subPixelInterpolationOffsetBits = 4,
1339 .maxFramebufferWidth = (1 << 14),
1340 .maxFramebufferHeight = (1 << 14),
1341 .maxFramebufferLayers = (1 << 11),
1342 .framebufferColorSampleCounts = sample_counts,
1343 .framebufferDepthSampleCounts = sample_counts,
1344 .framebufferStencilSampleCounts = sample_counts,
1345 .framebufferNoAttachmentsSampleCounts = sample_counts,
1346 .maxColorAttachments = MAX_RTS,
1347 .sampledImageColorSampleCounts = sample_counts,
1348 .sampledImageIntegerSampleCounts = VK_SAMPLE_COUNT_1_BIT,
1349 .sampledImageDepthSampleCounts = sample_counts,
1350 .sampledImageStencilSampleCounts = sample_counts,
1351 .storageImageSampleCounts = VK_SAMPLE_COUNT_1_BIT,
1352 .maxSampleMaskWords = 1,
1353 .timestampComputeAndGraphics = true,
1354 .timestampPeriod = 1000000000.0 / devinfo->timestamp_frequency,
1355 .maxClipDistances = 8,
1356 .maxCullDistances = 8,
1357 .maxCombinedClipAndCullDistances = 8,
1358 .discreteQueuePriorities = 2,
1359 .pointSizeRange = { 0.125, 255.875 },
1360 .lineWidthRange = {
1361 0.0,
1362 (devinfo->gen >= 9 || devinfo->is_cherryview) ?
1363 2047.9921875 : 7.9921875,
1364 },
1365 .pointSizeGranularity = (1.0 / 8.0),
1366 .lineWidthGranularity = (1.0 / 128.0),
1367 .strictLines = false,
1368 .standardSampleLocations = true,
1369 .optimalBufferCopyOffsetAlignment = 128,
1370 .optimalBufferCopyRowPitchAlignment = 128,
1371 .nonCoherentAtomSize = 64,
1372 };
1373
1374 *pProperties = (VkPhysicalDeviceProperties) {
1375 .apiVersion = anv_physical_device_api_version(pdevice),
1376 .driverVersion = vk_get_driver_version(),
1377 .vendorID = 0x8086,
1378 .deviceID = pdevice->chipset_id,
1379 .deviceType = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
1380 .limits = limits,
1381 .sparseProperties = {0}, /* Broadwell doesn't do sparse. */
1382 };
1383
1384 snprintf(pProperties->deviceName, sizeof(pProperties->deviceName),
1385 "%s", pdevice->name);
1386 memcpy(pProperties->pipelineCacheUUID,
1387 pdevice->pipeline_cache_uuid, VK_UUID_SIZE);
1388 }
1389
1390 void anv_GetPhysicalDeviceProperties2(
1391 VkPhysicalDevice physicalDevice,
1392 VkPhysicalDeviceProperties2* pProperties)
1393 {
1394 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
1395
1396 anv_GetPhysicalDeviceProperties(physicalDevice, &pProperties->properties);
1397
1398 vk_foreach_struct(ext, pProperties->pNext) {
1399 switch (ext->sType) {
1400 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_STENCIL_RESOLVE_PROPERTIES_KHR: {
1401 VkPhysicalDeviceDepthStencilResolvePropertiesKHR *props =
1402 (VkPhysicalDeviceDepthStencilResolvePropertiesKHR *)ext;
1403
1404 /* We support all of the depth resolve modes */
1405 props->supportedDepthResolveModes =
1406 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR |
1407 VK_RESOLVE_MODE_AVERAGE_BIT_KHR |
1408 VK_RESOLVE_MODE_MIN_BIT_KHR |
1409 VK_RESOLVE_MODE_MAX_BIT_KHR;
1410
1411 /* Average doesn't make sense for stencil so we don't support that */
1412 props->supportedStencilResolveModes =
1413 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR;
1414 if (pdevice->info.gen >= 8) {
1415 /* The advanced stencil resolve modes currently require stencil
1416 * sampling be supported by the hardware.
1417 */
1418 props->supportedStencilResolveModes |=
1419 VK_RESOLVE_MODE_MIN_BIT_KHR |
1420 VK_RESOLVE_MODE_MAX_BIT_KHR;
1421 }
1422
1423 props->independentResolveNone = true;
1424 props->independentResolve = true;
1425 break;
1426 }
1427
1428 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES_EXT: {
1429 VkPhysicalDeviceDescriptorIndexingPropertiesEXT *props =
1430 (VkPhysicalDeviceDescriptorIndexingPropertiesEXT *)ext;
1431
1432 /* It's a bit hard to exactly map our implementation to the limits
1433 * described here. The bindless surface handle in the extended
1434 * message descriptors is 20 bits and it's an index into the table of
1435 * RENDER_SURFACE_STATE structs that starts at bindless surface base
1436 * address. Given that most things consume two surface states per
1437 * view (general/sampled for textures and write-only/read-write for
1438 * images), we claim 2^19 things.
1439 *
1440 * For SSBOs, we just use A64 messages so there is no real limit
1441 * there beyond the limit on the total size of a descriptor set.
1442 */
1443 const unsigned max_bindless_views = 1 << 19;
1444
1445 props->maxUpdateAfterBindDescriptorsInAllPools = max_bindless_views;
1446 props->shaderUniformBufferArrayNonUniformIndexingNative = false;
1447 props->shaderSampledImageArrayNonUniformIndexingNative = false;
1448 props->shaderStorageBufferArrayNonUniformIndexingNative = true;
1449 props->shaderStorageImageArrayNonUniformIndexingNative = false;
1450 props->shaderInputAttachmentArrayNonUniformIndexingNative = false;
1451 props->robustBufferAccessUpdateAfterBind = true;
1452 props->quadDivergentImplicitLod = false;
1453 props->maxPerStageDescriptorUpdateAfterBindSamplers = max_bindless_views;
1454 props->maxPerStageDescriptorUpdateAfterBindUniformBuffers = MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS;
1455 props->maxPerStageDescriptorUpdateAfterBindStorageBuffers = UINT32_MAX;
1456 props->maxPerStageDescriptorUpdateAfterBindSampledImages = max_bindless_views;
1457 props->maxPerStageDescriptorUpdateAfterBindStorageImages = max_bindless_views;
1458 props->maxPerStageDescriptorUpdateAfterBindInputAttachments = MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS;
1459 props->maxPerStageUpdateAfterBindResources = UINT32_MAX;
1460 props->maxDescriptorSetUpdateAfterBindSamplers = max_bindless_views;
1461 props->maxDescriptorSetUpdateAfterBindUniformBuffers = 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS;
1462 props->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2;
1463 props->maxDescriptorSetUpdateAfterBindStorageBuffers = UINT32_MAX;
1464 props->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2;
1465 props->maxDescriptorSetUpdateAfterBindSampledImages = max_bindless_views;
1466 props->maxDescriptorSetUpdateAfterBindStorageImages = max_bindless_views;
1467 props->maxDescriptorSetUpdateAfterBindInputAttachments = MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS;
1468 break;
1469 }
1470
1471 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR: {
1472 VkPhysicalDeviceDriverPropertiesKHR *driver_props =
1473 (VkPhysicalDeviceDriverPropertiesKHR *) ext;
1474
1475 driver_props->driverID = VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA_KHR;
1476 snprintf(driver_props->driverName, VK_MAX_DRIVER_NAME_SIZE_KHR,
1477 "Intel open-source Mesa driver");
1478
1479 snprintf(driver_props->driverInfo, VK_MAX_DRIVER_INFO_SIZE_KHR,
1480 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1);
1481
1482 driver_props->conformanceVersion = (VkConformanceVersionKHR) {
1483 .major = 1,
1484 .minor = 1,
1485 .subminor = 2,
1486 .patch = 0,
1487 };
1488 break;
1489 }
1490
1491 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT: {
1492 VkPhysicalDeviceExternalMemoryHostPropertiesEXT *props =
1493 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT *) ext;
1494 /* Userptr needs page aligned memory. */
1495 props->minImportedHostPointerAlignment = 4096;
1496 break;
1497 }
1498
1499 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES: {
1500 VkPhysicalDeviceIDProperties *id_props =
1501 (VkPhysicalDeviceIDProperties *)ext;
1502 memcpy(id_props->deviceUUID, pdevice->device_uuid, VK_UUID_SIZE);
1503 memcpy(id_props->driverUUID, pdevice->driver_uuid, VK_UUID_SIZE);
1504 /* The LUID is for Windows. */
1505 id_props->deviceLUIDValid = false;
1506 break;
1507 }
1508
1509 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_PROPERTIES_EXT: {
1510 VkPhysicalDeviceInlineUniformBlockPropertiesEXT *props =
1511 (VkPhysicalDeviceInlineUniformBlockPropertiesEXT *)ext;
1512 props->maxInlineUniformBlockSize = MAX_INLINE_UNIFORM_BLOCK_SIZE;
1513 props->maxPerStageDescriptorInlineUniformBlocks =
1514 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1515 props->maxPerStageDescriptorUpdateAfterBindInlineUniformBlocks =
1516 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1517 props->maxDescriptorSetInlineUniformBlocks =
1518 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1519 props->maxDescriptorSetUpdateAfterBindInlineUniformBlocks =
1520 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1521 break;
1522 }
1523
1524 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_PROPERTIES_EXT: {
1525 VkPhysicalDeviceLineRasterizationPropertiesEXT *props =
1526 (VkPhysicalDeviceLineRasterizationPropertiesEXT *)ext;
1527 /* In the Skylake PRM Vol. 7, subsection titled "GIQ (Diamond)
1528 * Sampling Rules - Legacy Mode", it says the following:
1529 *
1530 * "Note that the device divides a pixel into a 16x16 array of
1531 * subpixels, referenced by their upper left corners."
1532 *
1533 * This is the only known reference in the PRMs to the subpixel
1534 * precision of line rasterization and a "16x16 array of subpixels"
1535 * implies 4 subpixel precision bits. Empirical testing has shown
1536 * that 4 subpixel precision bits applies to all line rasterization
1537 * types.
1538 */
1539 props->lineSubPixelPrecisionBits = 4;
1540 break;
1541 }
1542
1543 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES: {
1544 VkPhysicalDeviceMaintenance3Properties *props =
1545 (VkPhysicalDeviceMaintenance3Properties *)ext;
1546 /* This value doesn't matter for us today as our per-stage
1547 * descriptors are the real limit.
1548 */
1549 props->maxPerSetDescriptors = 1024;
1550 props->maxMemoryAllocationSize = MAX_MEMORY_ALLOCATION_SIZE;
1551 break;
1552 }
1553
1554 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES: {
1555 VkPhysicalDeviceMultiviewProperties *properties =
1556 (VkPhysicalDeviceMultiviewProperties *)ext;
1557 properties->maxMultiviewViewCount = 16;
1558 properties->maxMultiviewInstanceIndex = UINT32_MAX / 16;
1559 break;
1560 }
1561
1562 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT: {
1563 VkPhysicalDevicePCIBusInfoPropertiesEXT *properties =
1564 (VkPhysicalDevicePCIBusInfoPropertiesEXT *)ext;
1565 properties->pciDomain = pdevice->pci_info.domain;
1566 properties->pciBus = pdevice->pci_info.bus;
1567 properties->pciDevice = pdevice->pci_info.device;
1568 properties->pciFunction = pdevice->pci_info.function;
1569 break;
1570 }
1571
1572 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES: {
1573 VkPhysicalDevicePointClippingProperties *properties =
1574 (VkPhysicalDevicePointClippingProperties *) ext;
1575 properties->pointClippingBehavior = VK_POINT_CLIPPING_BEHAVIOR_USER_CLIP_PLANES_ONLY;
1576 break;
1577 }
1578
1579 #pragma GCC diagnostic push
1580 #pragma GCC diagnostic ignored "-Wswitch"
1581 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PRESENTATION_PROPERTIES_ANDROID: {
1582 VkPhysicalDevicePresentationPropertiesANDROID *props =
1583 (VkPhysicalDevicePresentationPropertiesANDROID *)ext;
1584 props->sharedImage = VK_FALSE;
1585 break;
1586 }
1587 #pragma GCC diagnostic pop
1588
1589 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES: {
1590 VkPhysicalDeviceProtectedMemoryProperties *props =
1591 (VkPhysicalDeviceProtectedMemoryProperties *)ext;
1592 props->protectedNoFault = false;
1593 break;
1594 }
1595
1596 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR: {
1597 VkPhysicalDevicePushDescriptorPropertiesKHR *properties =
1598 (VkPhysicalDevicePushDescriptorPropertiesKHR *) ext;
1599
1600 properties->maxPushDescriptors = MAX_PUSH_DESCRIPTORS;
1601 break;
1602 }
1603
1604 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT: {
1605 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT *properties =
1606 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT *)ext;
1607 properties->filterMinmaxImageComponentMapping = pdevice->info.gen >= 9;
1608 properties->filterMinmaxSingleComponentFormats = true;
1609 break;
1610 }
1611
1612 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES: {
1613 VkPhysicalDeviceSubgroupProperties *properties = (void *)ext;
1614
1615 properties->subgroupSize = BRW_SUBGROUP_SIZE;
1616
1617 VkShaderStageFlags scalar_stages = 0;
1618 for (unsigned stage = 0; stage < MESA_SHADER_STAGES; stage++) {
1619 if (pdevice->compiler->scalar_stage[stage])
1620 scalar_stages |= mesa_to_vk_shader_stage(stage);
1621 }
1622 properties->supportedStages = scalar_stages;
1623
1624 properties->supportedOperations = VK_SUBGROUP_FEATURE_BASIC_BIT |
1625 VK_SUBGROUP_FEATURE_VOTE_BIT |
1626 VK_SUBGROUP_FEATURE_BALLOT_BIT |
1627 VK_SUBGROUP_FEATURE_SHUFFLE_BIT |
1628 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT |
1629 VK_SUBGROUP_FEATURE_QUAD_BIT;
1630 if (pdevice->info.gen >= 8) {
1631 /* TODO: There's no technical reason why these can't be made to
1632 * work on gen7 but they don't at the moment so it's best to leave
1633 * the feature disabled than enabled and broken.
1634 */
1635 properties->supportedOperations |=
1636 VK_SUBGROUP_FEATURE_ARITHMETIC_BIT |
1637 VK_SUBGROUP_FEATURE_CLUSTERED_BIT;
1638 }
1639 properties->quadOperationsInAllStages = pdevice->info.gen >= 8;
1640 break;
1641 }
1642
1643 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_PROPERTIES_EXT: {
1644 VkPhysicalDeviceSubgroupSizeControlPropertiesEXT *props =
1645 (VkPhysicalDeviceSubgroupSizeControlPropertiesEXT *)ext;
1646 STATIC_ASSERT(8 <= BRW_SUBGROUP_SIZE && BRW_SUBGROUP_SIZE <= 32);
1647 props->minSubgroupSize = 8;
1648 props->maxSubgroupSize = 32;
1649 props->maxComputeWorkgroupSubgroups = pdevice->info.max_cs_threads;
1650 props->requiredSubgroupSizeStages = VK_SHADER_STAGE_COMPUTE_BIT;
1651 break;
1652 }
1653
1654 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_PROPERTIES_EXT: {
1655 VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT *props =
1656 (VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT *)ext;
1657
1658 /* From the SKL PRM Vol. 2d, docs for RENDER_SURFACE_STATE::Surface
1659 * Base Address:
1660 *
1661 * "For SURFTYPE_BUFFER non-rendertarget surfaces, this field
1662 * specifies the base address of the first element of the surface,
1663 * computed in software by adding the surface base address to the
1664 * byte offset of the element in the buffer. The base address must
1665 * be aligned to element size."
1666 *
1667 * The typed dataport messages require that things be texel aligned.
1668 * Otherwise, we may just load/store the wrong data or, in the worst
1669 * case, there may be hangs.
1670 */
1671 props->storageTexelBufferOffsetAlignmentBytes = 16;
1672 props->storageTexelBufferOffsetSingleTexelAlignment = true;
1673
1674 /* The sampler, however, is much more forgiving and it can handle
1675 * arbitrary byte alignment for linear and buffer surfaces. It's
1676 * hard to find a good PRM citation for this but years of empirical
1677 * experience demonstrate that this is true.
1678 */
1679 props->uniformTexelBufferOffsetAlignmentBytes = 1;
1680 props->uniformTexelBufferOffsetSingleTexelAlignment = false;
1681 break;
1682 }
1683
1684 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT: {
1685 VkPhysicalDeviceTransformFeedbackPropertiesEXT *props =
1686 (VkPhysicalDeviceTransformFeedbackPropertiesEXT *)ext;
1687
1688 props->maxTransformFeedbackStreams = MAX_XFB_STREAMS;
1689 props->maxTransformFeedbackBuffers = MAX_XFB_BUFFERS;
1690 props->maxTransformFeedbackBufferSize = (1ull << 32);
1691 props->maxTransformFeedbackStreamDataSize = 128 * 4;
1692 props->maxTransformFeedbackBufferDataSize = 128 * 4;
1693 props->maxTransformFeedbackBufferDataStride = 2048;
1694 props->transformFeedbackQueries = true;
1695 props->transformFeedbackStreamsLinesTriangles = false;
1696 props->transformFeedbackRasterizationStreamSelect = false;
1697 props->transformFeedbackDraw = true;
1698 break;
1699 }
1700
1701 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT: {
1702 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *props =
1703 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *)ext;
1704 /* We have to restrict this a bit for multiview */
1705 props->maxVertexAttribDivisor = UINT32_MAX / 16;
1706 break;
1707 }
1708
1709 default:
1710 anv_debug_ignored_stype(ext->sType);
1711 break;
1712 }
1713 }
1714 }
1715
1716 /* We support exactly one queue family. */
1717 static const VkQueueFamilyProperties
1718 anv_queue_family_properties = {
1719 .queueFlags = VK_QUEUE_GRAPHICS_BIT |
1720 VK_QUEUE_COMPUTE_BIT |
1721 VK_QUEUE_TRANSFER_BIT,
1722 .queueCount = 1,
1723 .timestampValidBits = 36, /* XXX: Real value here */
1724 .minImageTransferGranularity = { 1, 1, 1 },
1725 };
1726
1727 void anv_GetPhysicalDeviceQueueFamilyProperties(
1728 VkPhysicalDevice physicalDevice,
1729 uint32_t* pCount,
1730 VkQueueFamilyProperties* pQueueFamilyProperties)
1731 {
1732 VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pCount);
1733
1734 vk_outarray_append(&out, p) {
1735 *p = anv_queue_family_properties;
1736 }
1737 }
1738
1739 void anv_GetPhysicalDeviceQueueFamilyProperties2(
1740 VkPhysicalDevice physicalDevice,
1741 uint32_t* pQueueFamilyPropertyCount,
1742 VkQueueFamilyProperties2* pQueueFamilyProperties)
1743 {
1744
1745 VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pQueueFamilyPropertyCount);
1746
1747 vk_outarray_append(&out, p) {
1748 p->queueFamilyProperties = anv_queue_family_properties;
1749
1750 vk_foreach_struct(s, p->pNext) {
1751 anv_debug_ignored_stype(s->sType);
1752 }
1753 }
1754 }
1755
1756 void anv_GetPhysicalDeviceMemoryProperties(
1757 VkPhysicalDevice physicalDevice,
1758 VkPhysicalDeviceMemoryProperties* pMemoryProperties)
1759 {
1760 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
1761
1762 pMemoryProperties->memoryTypeCount = physical_device->memory.type_count;
1763 for (uint32_t i = 0; i < physical_device->memory.type_count; i++) {
1764 pMemoryProperties->memoryTypes[i] = (VkMemoryType) {
1765 .propertyFlags = physical_device->memory.types[i].propertyFlags,
1766 .heapIndex = physical_device->memory.types[i].heapIndex,
1767 };
1768 }
1769
1770 pMemoryProperties->memoryHeapCount = physical_device->memory.heap_count;
1771 for (uint32_t i = 0; i < physical_device->memory.heap_count; i++) {
1772 pMemoryProperties->memoryHeaps[i] = (VkMemoryHeap) {
1773 .size = physical_device->memory.heaps[i].size,
1774 .flags = physical_device->memory.heaps[i].flags,
1775 };
1776 }
1777 }
1778
1779 static void
1780 anv_get_memory_budget(VkPhysicalDevice physicalDevice,
1781 VkPhysicalDeviceMemoryBudgetPropertiesEXT *memoryBudget)
1782 {
1783 ANV_FROM_HANDLE(anv_physical_device, device, physicalDevice);
1784 uint64_t sys_available = get_available_system_memory();
1785 assert(sys_available > 0);
1786
1787 VkDeviceSize total_heaps_size = 0;
1788 for (size_t i = 0; i < device->memory.heap_count; i++)
1789 total_heaps_size += device->memory.heaps[i].size;
1790
1791 for (size_t i = 0; i < device->memory.heap_count; i++) {
1792 VkDeviceSize heap_size = device->memory.heaps[i].size;
1793 VkDeviceSize heap_used = device->memory.heaps[i].used;
1794 VkDeviceSize heap_budget;
1795
1796 double heap_proportion = (double) heap_size / total_heaps_size;
1797 VkDeviceSize sys_available_prop = sys_available * heap_proportion;
1798
1799 /*
1800 * Let's not incite the app to starve the system: report at most 90% of
1801 * available system memory.
1802 */
1803 uint64_t heap_available = sys_available_prop * 9 / 10;
1804 heap_budget = MIN2(heap_size, heap_used + heap_available);
1805
1806 /*
1807 * Round down to the nearest MB
1808 */
1809 heap_budget &= ~((1ull << 20) - 1);
1810
1811 /*
1812 * The heapBudget value must be non-zero for array elements less than
1813 * VkPhysicalDeviceMemoryProperties::memoryHeapCount. The heapBudget
1814 * value must be less than or equal to VkMemoryHeap::size for each heap.
1815 */
1816 assert(0 < heap_budget && heap_budget <= heap_size);
1817
1818 memoryBudget->heapUsage[i] = heap_used;
1819 memoryBudget->heapBudget[i] = heap_budget;
1820 }
1821
1822 /* The heapBudget and heapUsage values must be zero for array elements
1823 * greater than or equal to VkPhysicalDeviceMemoryProperties::memoryHeapCount
1824 */
1825 for (uint32_t i = device->memory.heap_count; i < VK_MAX_MEMORY_HEAPS; i++) {
1826 memoryBudget->heapBudget[i] = 0;
1827 memoryBudget->heapUsage[i] = 0;
1828 }
1829 }
1830
1831 void anv_GetPhysicalDeviceMemoryProperties2(
1832 VkPhysicalDevice physicalDevice,
1833 VkPhysicalDeviceMemoryProperties2* pMemoryProperties)
1834 {
1835 anv_GetPhysicalDeviceMemoryProperties(physicalDevice,
1836 &pMemoryProperties->memoryProperties);
1837
1838 vk_foreach_struct(ext, pMemoryProperties->pNext) {
1839 switch (ext->sType) {
1840 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT:
1841 anv_get_memory_budget(physicalDevice, (void*)ext);
1842 break;
1843 default:
1844 anv_debug_ignored_stype(ext->sType);
1845 break;
1846 }
1847 }
1848 }
1849
1850 void
1851 anv_GetDeviceGroupPeerMemoryFeatures(
1852 VkDevice device,
1853 uint32_t heapIndex,
1854 uint32_t localDeviceIndex,
1855 uint32_t remoteDeviceIndex,
1856 VkPeerMemoryFeatureFlags* pPeerMemoryFeatures)
1857 {
1858 assert(localDeviceIndex == 0 && remoteDeviceIndex == 0);
1859 *pPeerMemoryFeatures = VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT |
1860 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT |
1861 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT |
1862 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT;
1863 }
1864
1865 PFN_vkVoidFunction anv_GetInstanceProcAddr(
1866 VkInstance _instance,
1867 const char* pName)
1868 {
1869 ANV_FROM_HANDLE(anv_instance, instance, _instance);
1870
1871 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
1872 * when we have to return valid function pointers, NULL, or it's left
1873 * undefined. See the table for exact details.
1874 */
1875 if (pName == NULL)
1876 return NULL;
1877
1878 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
1879 if (strcmp(pName, "vk" #entrypoint) == 0) \
1880 return (PFN_vkVoidFunction)anv_##entrypoint
1881
1882 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties);
1883 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties);
1884 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceVersion);
1885 LOOKUP_ANV_ENTRYPOINT(CreateInstance);
1886
1887 #undef LOOKUP_ANV_ENTRYPOINT
1888
1889 if (instance == NULL)
1890 return NULL;
1891
1892 int idx = anv_get_instance_entrypoint_index(pName);
1893 if (idx >= 0)
1894 return instance->dispatch.entrypoints[idx];
1895
1896 idx = anv_get_device_entrypoint_index(pName);
1897 if (idx >= 0)
1898 return instance->device_dispatch.entrypoints[idx];
1899
1900 return NULL;
1901 }
1902
1903 /* With version 1+ of the loader interface the ICD should expose
1904 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
1905 */
1906 PUBLIC
1907 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
1908 VkInstance instance,
1909 const char* pName);
1910
1911 PUBLIC
1912 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
1913 VkInstance instance,
1914 const char* pName)
1915 {
1916 return anv_GetInstanceProcAddr(instance, pName);
1917 }
1918
1919 PFN_vkVoidFunction anv_GetDeviceProcAddr(
1920 VkDevice _device,
1921 const char* pName)
1922 {
1923 ANV_FROM_HANDLE(anv_device, device, _device);
1924
1925 if (!device || !pName)
1926 return NULL;
1927
1928 int idx = anv_get_device_entrypoint_index(pName);
1929 if (idx < 0)
1930 return NULL;
1931
1932 return device->dispatch.entrypoints[idx];
1933 }
1934
1935 VkResult
1936 anv_CreateDebugReportCallbackEXT(VkInstance _instance,
1937 const VkDebugReportCallbackCreateInfoEXT* pCreateInfo,
1938 const VkAllocationCallbacks* pAllocator,
1939 VkDebugReportCallbackEXT* pCallback)
1940 {
1941 ANV_FROM_HANDLE(anv_instance, instance, _instance);
1942 return vk_create_debug_report_callback(&instance->debug_report_callbacks,
1943 pCreateInfo, pAllocator, &instance->alloc,
1944 pCallback);
1945 }
1946
1947 void
1948 anv_DestroyDebugReportCallbackEXT(VkInstance _instance,
1949 VkDebugReportCallbackEXT _callback,
1950 const VkAllocationCallbacks* pAllocator)
1951 {
1952 ANV_FROM_HANDLE(anv_instance, instance, _instance);
1953 vk_destroy_debug_report_callback(&instance->debug_report_callbacks,
1954 _callback, pAllocator, &instance->alloc);
1955 }
1956
1957 void
1958 anv_DebugReportMessageEXT(VkInstance _instance,
1959 VkDebugReportFlagsEXT flags,
1960 VkDebugReportObjectTypeEXT objectType,
1961 uint64_t object,
1962 size_t location,
1963 int32_t messageCode,
1964 const char* pLayerPrefix,
1965 const char* pMessage)
1966 {
1967 ANV_FROM_HANDLE(anv_instance, instance, _instance);
1968 vk_debug_report(&instance->debug_report_callbacks, flags, objectType,
1969 object, location, messageCode, pLayerPrefix, pMessage);
1970 }
1971
1972 static void
1973 anv_queue_init(struct anv_device *device, struct anv_queue *queue)
1974 {
1975 queue->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
1976 queue->device = device;
1977 queue->flags = 0;
1978 }
1979
1980 static void
1981 anv_queue_finish(struct anv_queue *queue)
1982 {
1983 }
1984
1985 static struct anv_state
1986 anv_state_pool_emit_data(struct anv_state_pool *pool, size_t size, size_t align, const void *p)
1987 {
1988 struct anv_state state;
1989
1990 state = anv_state_pool_alloc(pool, size, align);
1991 memcpy(state.map, p, size);
1992
1993 return state;
1994 }
1995
1996 /* Haswell border color is a bit of a disaster. Float and unorm formats use a
1997 * straightforward 32-bit float color in the first 64 bytes. Instead of using
1998 * a nice float/integer union like Gen8+, Haswell specifies the integer border
1999 * color as a separate entry /after/ the float color. The layout of this entry
2000 * also depends on the format's bpp (with extra hacks for RG32), and overlaps.
2001 *
2002 * Since we don't know the format/bpp, we can't make any of the border colors
2003 * containing '1' work for all formats, as it would be in the wrong place for
2004 * some of them. We opt to make 32-bit integers work as this seems like the
2005 * most common option. Fortunately, transparent black works regardless, as
2006 * all zeroes is the same in every bit-size.
2007 */
2008 struct hsw_border_color {
2009 float float32[4];
2010 uint32_t _pad0[12];
2011 uint32_t uint32[4];
2012 uint32_t _pad1[108];
2013 };
2014
2015 struct gen8_border_color {
2016 union {
2017 float float32[4];
2018 uint32_t uint32[4];
2019 };
2020 /* Pad out to 64 bytes */
2021 uint32_t _pad[12];
2022 };
2023
2024 static void
2025 anv_device_init_border_colors(struct anv_device *device)
2026 {
2027 if (device->info.is_haswell) {
2028 static const struct hsw_border_color border_colors[] = {
2029 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
2030 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
2031 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
2032 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = { .uint32 = { 0, 0, 0, 0 } },
2033 [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = { .uint32 = { 0, 0, 0, 1 } },
2034 [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = { .uint32 = { 1, 1, 1, 1 } },
2035 };
2036
2037 device->border_colors =
2038 anv_state_pool_emit_data(&device->dynamic_state_pool,
2039 sizeof(border_colors), 512, border_colors);
2040 } else {
2041 static const struct gen8_border_color border_colors[] = {
2042 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
2043 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
2044 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
2045 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = { .uint32 = { 0, 0, 0, 0 } },
2046 [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = { .uint32 = { 0, 0, 0, 1 } },
2047 [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = { .uint32 = { 1, 1, 1, 1 } },
2048 };
2049
2050 device->border_colors =
2051 anv_state_pool_emit_data(&device->dynamic_state_pool,
2052 sizeof(border_colors), 64, border_colors);
2053 }
2054 }
2055
2056 static void
2057 anv_device_init_trivial_batch(struct anv_device *device)
2058 {
2059 anv_bo_init_new(&device->trivial_batch_bo, device, 4096);
2060
2061 if (device->instance->physicalDevice.has_exec_async)
2062 device->trivial_batch_bo.flags |= EXEC_OBJECT_ASYNC;
2063
2064 if (device->instance->physicalDevice.use_softpin)
2065 device->trivial_batch_bo.flags |= EXEC_OBJECT_PINNED;
2066
2067 anv_vma_alloc(device, &device->trivial_batch_bo);
2068
2069 void *map = anv_gem_mmap(device, device->trivial_batch_bo.gem_handle,
2070 0, 4096, 0);
2071
2072 struct anv_batch batch = {
2073 .start = map,
2074 .next = map,
2075 .end = map + 4096,
2076 };
2077
2078 anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
2079 anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
2080
2081 if (!device->info.has_llc)
2082 gen_clflush_range(map, batch.next - map);
2083
2084 anv_gem_munmap(map, device->trivial_batch_bo.size);
2085 }
2086
2087 VkResult anv_EnumerateDeviceExtensionProperties(
2088 VkPhysicalDevice physicalDevice,
2089 const char* pLayerName,
2090 uint32_t* pPropertyCount,
2091 VkExtensionProperties* pProperties)
2092 {
2093 ANV_FROM_HANDLE(anv_physical_device, device, physicalDevice);
2094 VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
2095
2096 for (int i = 0; i < ANV_DEVICE_EXTENSION_COUNT; i++) {
2097 if (device->supported_extensions.extensions[i]) {
2098 vk_outarray_append(&out, prop) {
2099 *prop = anv_device_extensions[i];
2100 }
2101 }
2102 }
2103
2104 return vk_outarray_status(&out);
2105 }
2106
2107 static void
2108 anv_device_init_dispatch(struct anv_device *device)
2109 {
2110 const struct anv_device_dispatch_table *genX_table;
2111 switch (device->info.gen) {
2112 case 11:
2113 genX_table = &gen11_device_dispatch_table;
2114 break;
2115 case 10:
2116 genX_table = &gen10_device_dispatch_table;
2117 break;
2118 case 9:
2119 genX_table = &gen9_device_dispatch_table;
2120 break;
2121 case 8:
2122 genX_table = &gen8_device_dispatch_table;
2123 break;
2124 case 7:
2125 if (device->info.is_haswell)
2126 genX_table = &gen75_device_dispatch_table;
2127 else
2128 genX_table = &gen7_device_dispatch_table;
2129 break;
2130 default:
2131 unreachable("unsupported gen\n");
2132 }
2133
2134 for (unsigned i = 0; i < ARRAY_SIZE(device->dispatch.entrypoints); i++) {
2135 /* Vulkan requires that entrypoints for extensions which have not been
2136 * enabled must not be advertised.
2137 */
2138 if (!anv_device_entrypoint_is_enabled(i, device->instance->app_info.api_version,
2139 &device->instance->enabled_extensions,
2140 &device->enabled_extensions)) {
2141 device->dispatch.entrypoints[i] = NULL;
2142 } else if (genX_table->entrypoints[i]) {
2143 device->dispatch.entrypoints[i] = genX_table->entrypoints[i];
2144 } else {
2145 device->dispatch.entrypoints[i] =
2146 anv_device_dispatch_table.entrypoints[i];
2147 }
2148 }
2149 }
2150
2151 static int
2152 vk_priority_to_gen(int priority)
2153 {
2154 switch (priority) {
2155 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT:
2156 return GEN_CONTEXT_LOW_PRIORITY;
2157 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT:
2158 return GEN_CONTEXT_MEDIUM_PRIORITY;
2159 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT:
2160 return GEN_CONTEXT_HIGH_PRIORITY;
2161 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT:
2162 return GEN_CONTEXT_REALTIME_PRIORITY;
2163 default:
2164 unreachable("Invalid priority");
2165 }
2166 }
2167
2168 static void
2169 anv_device_init_hiz_clear_value_bo(struct anv_device *device)
2170 {
2171 anv_bo_init_new(&device->hiz_clear_bo, device, 4096);
2172
2173 if (device->instance->physicalDevice.has_exec_async)
2174 device->hiz_clear_bo.flags |= EXEC_OBJECT_ASYNC;
2175
2176 if (device->instance->physicalDevice.use_softpin)
2177 device->hiz_clear_bo.flags |= EXEC_OBJECT_PINNED;
2178
2179 anv_vma_alloc(device, &device->hiz_clear_bo);
2180
2181 uint32_t *map = anv_gem_mmap(device, device->hiz_clear_bo.gem_handle,
2182 0, 4096, 0);
2183
2184 union isl_color_value hiz_clear = { .u32 = { 0, } };
2185 hiz_clear.f32[0] = ANV_HZ_FC_VAL;
2186
2187 memcpy(map, hiz_clear.u32, sizeof(hiz_clear.u32));
2188 anv_gem_munmap(map, device->hiz_clear_bo.size);
2189 }
2190
2191 static bool
2192 get_bo_from_pool(struct gen_batch_decode_bo *ret,
2193 struct anv_block_pool *pool,
2194 uint64_t address)
2195 {
2196 for (uint32_t i = 0; i < pool->nbos; i++) {
2197 uint64_t bo_address = pool->bos[i].offset & (~0ull >> 16);
2198 uint32_t bo_size = pool->bos[i].size;
2199 if (address >= bo_address && address < (bo_address + bo_size)) {
2200 *ret = (struct gen_batch_decode_bo) {
2201 .addr = bo_address,
2202 .size = bo_size,
2203 .map = pool->bos[i].map,
2204 };
2205 return true;
2206 }
2207 }
2208 return false;
2209 }
2210
2211 /* Finding a buffer for batch decoding */
2212 static struct gen_batch_decode_bo
2213 decode_get_bo(void *v_batch, bool ppgtt, uint64_t address)
2214 {
2215 struct anv_device *device = v_batch;
2216 struct gen_batch_decode_bo ret_bo = {};
2217
2218 assert(ppgtt);
2219
2220 if (get_bo_from_pool(&ret_bo, &device->dynamic_state_pool.block_pool, address))
2221 return ret_bo;
2222 if (get_bo_from_pool(&ret_bo, &device->instruction_state_pool.block_pool, address))
2223 return ret_bo;
2224 if (get_bo_from_pool(&ret_bo, &device->binding_table_pool.block_pool, address))
2225 return ret_bo;
2226 if (get_bo_from_pool(&ret_bo, &device->surface_state_pool.block_pool, address))
2227 return ret_bo;
2228
2229 if (!device->cmd_buffer_being_decoded)
2230 return (struct gen_batch_decode_bo) { };
2231
2232 struct anv_batch_bo **bo;
2233
2234 u_vector_foreach(bo, &device->cmd_buffer_being_decoded->seen_bbos) {
2235 /* The decoder zeroes out the top 16 bits, so we need to as well */
2236 uint64_t bo_address = (*bo)->bo.offset & (~0ull >> 16);
2237
2238 if (address >= bo_address && address < bo_address + (*bo)->bo.size) {
2239 return (struct gen_batch_decode_bo) {
2240 .addr = bo_address,
2241 .size = (*bo)->bo.size,
2242 .map = (*bo)->bo.map,
2243 };
2244 }
2245 }
2246
2247 return (struct gen_batch_decode_bo) { };
2248 }
2249
2250 VkResult anv_CreateDevice(
2251 VkPhysicalDevice physicalDevice,
2252 const VkDeviceCreateInfo* pCreateInfo,
2253 const VkAllocationCallbacks* pAllocator,
2254 VkDevice* pDevice)
2255 {
2256 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
2257 VkResult result;
2258 struct anv_device *device;
2259
2260 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO);
2261
2262 struct anv_device_extension_table enabled_extensions = { };
2263 for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
2264 int idx;
2265 for (idx = 0; idx < ANV_DEVICE_EXTENSION_COUNT; idx++) {
2266 if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
2267 anv_device_extensions[idx].extensionName) == 0)
2268 break;
2269 }
2270
2271 if (idx >= ANV_DEVICE_EXTENSION_COUNT)
2272 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
2273
2274 if (!physical_device->supported_extensions.extensions[idx])
2275 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
2276
2277 enabled_extensions.extensions[idx] = true;
2278 }
2279
2280 /* Check enabled features */
2281 if (pCreateInfo->pEnabledFeatures) {
2282 VkPhysicalDeviceFeatures supported_features;
2283 anv_GetPhysicalDeviceFeatures(physicalDevice, &supported_features);
2284 VkBool32 *supported_feature = (VkBool32 *)&supported_features;
2285 VkBool32 *enabled_feature = (VkBool32 *)pCreateInfo->pEnabledFeatures;
2286 unsigned num_features = sizeof(VkPhysicalDeviceFeatures) / sizeof(VkBool32);
2287 for (uint32_t i = 0; i < num_features; i++) {
2288 if (enabled_feature[i] && !supported_feature[i])
2289 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
2290 }
2291 }
2292
2293 /* Check requested queues and fail if we are requested to create any
2294 * queues with flags we don't support.
2295 */
2296 assert(pCreateInfo->queueCreateInfoCount > 0);
2297 for (uint32_t i = 0; i < pCreateInfo->queueCreateInfoCount; i++) {
2298 if (pCreateInfo->pQueueCreateInfos[i].flags != 0)
2299 return vk_error(VK_ERROR_INITIALIZATION_FAILED);
2300 }
2301
2302 /* Check if client specified queue priority. */
2303 const VkDeviceQueueGlobalPriorityCreateInfoEXT *queue_priority =
2304 vk_find_struct_const(pCreateInfo->pQueueCreateInfos[0].pNext,
2305 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT);
2306
2307 VkQueueGlobalPriorityEXT priority =
2308 queue_priority ? queue_priority->globalPriority :
2309 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT;
2310
2311 device = vk_alloc2(&physical_device->instance->alloc, pAllocator,
2312 sizeof(*device), 8,
2313 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
2314 if (!device)
2315 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2316
2317 if (INTEL_DEBUG & DEBUG_BATCH) {
2318 const unsigned decode_flags =
2319 GEN_BATCH_DECODE_FULL |
2320 ((INTEL_DEBUG & DEBUG_COLOR) ? GEN_BATCH_DECODE_IN_COLOR : 0) |
2321 GEN_BATCH_DECODE_OFFSETS |
2322 GEN_BATCH_DECODE_FLOATS;
2323
2324 gen_batch_decode_ctx_init(&device->decoder_ctx,
2325 &physical_device->info,
2326 stderr, decode_flags, NULL,
2327 decode_get_bo, NULL, device);
2328 }
2329
2330 device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
2331 device->instance = physical_device->instance;
2332 device->chipset_id = physical_device->chipset_id;
2333 device->no_hw = physical_device->no_hw;
2334 device->_lost = false;
2335
2336 if (pAllocator)
2337 device->alloc = *pAllocator;
2338 else
2339 device->alloc = physical_device->instance->alloc;
2340
2341 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
2342 device->fd = open(physical_device->path, O_RDWR | O_CLOEXEC);
2343 if (device->fd == -1) {
2344 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2345 goto fail_device;
2346 }
2347
2348 device->context_id = anv_gem_create_context(device);
2349 if (device->context_id == -1) {
2350 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2351 goto fail_fd;
2352 }
2353
2354 if (physical_device->use_softpin) {
2355 if (pthread_mutex_init(&device->vma_mutex, NULL) != 0) {
2356 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2357 goto fail_fd;
2358 }
2359
2360 /* keep the page with address zero out of the allocator */
2361 struct anv_memory_heap *low_heap =
2362 &physical_device->memory.heaps[physical_device->memory.heap_count - 1];
2363 util_vma_heap_init(&device->vma_lo, low_heap->vma_start, low_heap->vma_size);
2364 device->vma_lo_available = low_heap->size;
2365
2366 struct anv_memory_heap *high_heap =
2367 &physical_device->memory.heaps[0];
2368 util_vma_heap_init(&device->vma_hi, high_heap->vma_start, high_heap->vma_size);
2369 device->vma_hi_available = physical_device->memory.heap_count == 1 ? 0 :
2370 high_heap->size;
2371 }
2372
2373 list_inithead(&device->memory_objects);
2374
2375 /* As per spec, the driver implementation may deny requests to acquire
2376 * a priority above the default priority (MEDIUM) if the caller does not
2377 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
2378 * is returned.
2379 */
2380 if (physical_device->has_context_priority) {
2381 int err = anv_gem_set_context_param(device->fd, device->context_id,
2382 I915_CONTEXT_PARAM_PRIORITY,
2383 vk_priority_to_gen(priority));
2384 if (err != 0 && priority > VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT) {
2385 result = vk_error(VK_ERROR_NOT_PERMITTED_EXT);
2386 goto fail_fd;
2387 }
2388 }
2389
2390 device->info = physical_device->info;
2391 device->isl_dev = physical_device->isl_dev;
2392
2393 /* On Broadwell and later, we can use batch chaining to more efficiently
2394 * implement growing command buffers. Prior to Haswell, the kernel
2395 * command parser gets in the way and we have to fall back to growing
2396 * the batch.
2397 */
2398 device->can_chain_batches = device->info.gen >= 8;
2399
2400 device->robust_buffer_access = pCreateInfo->pEnabledFeatures &&
2401 pCreateInfo->pEnabledFeatures->robustBufferAccess;
2402 device->enabled_extensions = enabled_extensions;
2403
2404 anv_device_init_dispatch(device);
2405
2406 if (pthread_mutex_init(&device->mutex, NULL) != 0) {
2407 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2408 goto fail_context_id;
2409 }
2410
2411 pthread_condattr_t condattr;
2412 if (pthread_condattr_init(&condattr) != 0) {
2413 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2414 goto fail_mutex;
2415 }
2416 if (pthread_condattr_setclock(&condattr, CLOCK_MONOTONIC) != 0) {
2417 pthread_condattr_destroy(&condattr);
2418 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2419 goto fail_mutex;
2420 }
2421 if (pthread_cond_init(&device->queue_submit, &condattr) != 0) {
2422 pthread_condattr_destroy(&condattr);
2423 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2424 goto fail_mutex;
2425 }
2426 pthread_condattr_destroy(&condattr);
2427
2428 uint64_t bo_flags =
2429 (physical_device->supports_48bit_addresses ? EXEC_OBJECT_SUPPORTS_48B_ADDRESS : 0) |
2430 (physical_device->has_exec_async ? EXEC_OBJECT_ASYNC : 0) |
2431 (physical_device->has_exec_capture ? EXEC_OBJECT_CAPTURE : 0) |
2432 (physical_device->use_softpin ? EXEC_OBJECT_PINNED : 0);
2433
2434 anv_bo_pool_init(&device->batch_bo_pool, device, bo_flags);
2435
2436 result = anv_bo_cache_init(&device->bo_cache);
2437 if (result != VK_SUCCESS)
2438 goto fail_batch_bo_pool;
2439
2440 if (!physical_device->use_softpin)
2441 bo_flags &= ~EXEC_OBJECT_SUPPORTS_48B_ADDRESS;
2442
2443 result = anv_state_pool_init(&device->dynamic_state_pool, device,
2444 DYNAMIC_STATE_POOL_MIN_ADDRESS,
2445 16384,
2446 bo_flags);
2447 if (result != VK_SUCCESS)
2448 goto fail_bo_cache;
2449
2450 result = anv_state_pool_init(&device->instruction_state_pool, device,
2451 INSTRUCTION_STATE_POOL_MIN_ADDRESS,
2452 16384,
2453 bo_flags);
2454 if (result != VK_SUCCESS)
2455 goto fail_dynamic_state_pool;
2456
2457 result = anv_state_pool_init(&device->surface_state_pool, device,
2458 SURFACE_STATE_POOL_MIN_ADDRESS,
2459 4096,
2460 bo_flags);
2461 if (result != VK_SUCCESS)
2462 goto fail_instruction_state_pool;
2463
2464 if (physical_device->use_softpin) {
2465 result = anv_state_pool_init(&device->binding_table_pool, device,
2466 BINDING_TABLE_POOL_MIN_ADDRESS,
2467 4096,
2468 bo_flags);
2469 if (result != VK_SUCCESS)
2470 goto fail_surface_state_pool;
2471 }
2472
2473 result = anv_bo_init_new(&device->workaround_bo, device, 4096);
2474 if (result != VK_SUCCESS)
2475 goto fail_binding_table_pool;
2476
2477 if (physical_device->use_softpin)
2478 device->workaround_bo.flags |= EXEC_OBJECT_PINNED;
2479
2480 if (!anv_vma_alloc(device, &device->workaround_bo))
2481 goto fail_workaround_bo;
2482
2483 anv_device_init_trivial_batch(device);
2484
2485 if (device->info.gen >= 10)
2486 anv_device_init_hiz_clear_value_bo(device);
2487
2488 anv_scratch_pool_init(device, &device->scratch_pool);
2489
2490 anv_queue_init(device, &device->queue);
2491
2492 switch (device->info.gen) {
2493 case 7:
2494 if (!device->info.is_haswell)
2495 result = gen7_init_device_state(device);
2496 else
2497 result = gen75_init_device_state(device);
2498 break;
2499 case 8:
2500 result = gen8_init_device_state(device);
2501 break;
2502 case 9:
2503 result = gen9_init_device_state(device);
2504 break;
2505 case 10:
2506 result = gen10_init_device_state(device);
2507 break;
2508 case 11:
2509 result = gen11_init_device_state(device);
2510 break;
2511 default:
2512 /* Shouldn't get here as we don't create physical devices for any other
2513 * gens. */
2514 unreachable("unhandled gen");
2515 }
2516 if (result != VK_SUCCESS)
2517 goto fail_workaround_bo;
2518
2519 anv_pipeline_cache_init(&device->default_pipeline_cache, device, true);
2520
2521 anv_device_init_blorp(device);
2522
2523 anv_device_init_border_colors(device);
2524
2525 *pDevice = anv_device_to_handle(device);
2526
2527 return VK_SUCCESS;
2528
2529 fail_workaround_bo:
2530 anv_queue_finish(&device->queue);
2531 anv_scratch_pool_finish(device, &device->scratch_pool);
2532 anv_gem_munmap(device->workaround_bo.map, device->workaround_bo.size);
2533 anv_gem_close(device, device->workaround_bo.gem_handle);
2534 fail_binding_table_pool:
2535 if (physical_device->use_softpin)
2536 anv_state_pool_finish(&device->binding_table_pool);
2537 fail_surface_state_pool:
2538 anv_state_pool_finish(&device->surface_state_pool);
2539 fail_instruction_state_pool:
2540 anv_state_pool_finish(&device->instruction_state_pool);
2541 fail_dynamic_state_pool:
2542 anv_state_pool_finish(&device->dynamic_state_pool);
2543 fail_bo_cache:
2544 anv_bo_cache_finish(&device->bo_cache);
2545 fail_batch_bo_pool:
2546 anv_bo_pool_finish(&device->batch_bo_pool);
2547 pthread_cond_destroy(&device->queue_submit);
2548 fail_mutex:
2549 pthread_mutex_destroy(&device->mutex);
2550 fail_context_id:
2551 anv_gem_destroy_context(device, device->context_id);
2552 fail_fd:
2553 close(device->fd);
2554 fail_device:
2555 vk_free(&device->alloc, device);
2556
2557 return result;
2558 }
2559
2560 void anv_DestroyDevice(
2561 VkDevice _device,
2562 const VkAllocationCallbacks* pAllocator)
2563 {
2564 ANV_FROM_HANDLE(anv_device, device, _device);
2565 struct anv_physical_device *physical_device;
2566
2567 if (!device)
2568 return;
2569
2570 physical_device = &device->instance->physicalDevice;
2571
2572 anv_device_finish_blorp(device);
2573
2574 anv_pipeline_cache_finish(&device->default_pipeline_cache);
2575
2576 anv_queue_finish(&device->queue);
2577
2578 #ifdef HAVE_VALGRIND
2579 /* We only need to free these to prevent valgrind errors. The backing
2580 * BO will go away in a couple of lines so we don't actually leak.
2581 */
2582 anv_state_pool_free(&device->dynamic_state_pool, device->border_colors);
2583 #endif
2584
2585 anv_scratch_pool_finish(device, &device->scratch_pool);
2586
2587 anv_gem_munmap(device->workaround_bo.map, device->workaround_bo.size);
2588 anv_vma_free(device, &device->workaround_bo);
2589 anv_gem_close(device, device->workaround_bo.gem_handle);
2590
2591 anv_vma_free(device, &device->trivial_batch_bo);
2592 anv_gem_close(device, device->trivial_batch_bo.gem_handle);
2593 if (device->info.gen >= 10)
2594 anv_gem_close(device, device->hiz_clear_bo.gem_handle);
2595
2596 if (physical_device->use_softpin)
2597 anv_state_pool_finish(&device->binding_table_pool);
2598 anv_state_pool_finish(&device->surface_state_pool);
2599 anv_state_pool_finish(&device->instruction_state_pool);
2600 anv_state_pool_finish(&device->dynamic_state_pool);
2601
2602 anv_bo_cache_finish(&device->bo_cache);
2603
2604 anv_bo_pool_finish(&device->batch_bo_pool);
2605
2606 pthread_cond_destroy(&device->queue_submit);
2607 pthread_mutex_destroy(&device->mutex);
2608
2609 anv_gem_destroy_context(device, device->context_id);
2610
2611 if (INTEL_DEBUG & DEBUG_BATCH)
2612 gen_batch_decode_ctx_finish(&device->decoder_ctx);
2613
2614 close(device->fd);
2615
2616 vk_free(&device->alloc, device);
2617 }
2618
2619 VkResult anv_EnumerateInstanceLayerProperties(
2620 uint32_t* pPropertyCount,
2621 VkLayerProperties* pProperties)
2622 {
2623 if (pProperties == NULL) {
2624 *pPropertyCount = 0;
2625 return VK_SUCCESS;
2626 }
2627
2628 /* None supported at this time */
2629 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
2630 }
2631
2632 VkResult anv_EnumerateDeviceLayerProperties(
2633 VkPhysicalDevice physicalDevice,
2634 uint32_t* pPropertyCount,
2635 VkLayerProperties* pProperties)
2636 {
2637 if (pProperties == NULL) {
2638 *pPropertyCount = 0;
2639 return VK_SUCCESS;
2640 }
2641
2642 /* None supported at this time */
2643 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
2644 }
2645
2646 void anv_GetDeviceQueue(
2647 VkDevice _device,
2648 uint32_t queueNodeIndex,
2649 uint32_t queueIndex,
2650 VkQueue* pQueue)
2651 {
2652 ANV_FROM_HANDLE(anv_device, device, _device);
2653
2654 assert(queueIndex == 0);
2655
2656 *pQueue = anv_queue_to_handle(&device->queue);
2657 }
2658
2659 void anv_GetDeviceQueue2(
2660 VkDevice _device,
2661 const VkDeviceQueueInfo2* pQueueInfo,
2662 VkQueue* pQueue)
2663 {
2664 ANV_FROM_HANDLE(anv_device, device, _device);
2665
2666 assert(pQueueInfo->queueIndex == 0);
2667
2668 if (pQueueInfo->flags == device->queue.flags)
2669 *pQueue = anv_queue_to_handle(&device->queue);
2670 else
2671 *pQueue = NULL;
2672 }
2673
2674 VkResult
2675 _anv_device_set_lost(struct anv_device *device,
2676 const char *file, int line,
2677 const char *msg, ...)
2678 {
2679 VkResult err;
2680 va_list ap;
2681
2682 device->_lost = true;
2683
2684 va_start(ap, msg);
2685 err = __vk_errorv(device->instance, device,
2686 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT,
2687 VK_ERROR_DEVICE_LOST, file, line, msg, ap);
2688 va_end(ap);
2689
2690 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
2691 abort();
2692
2693 return err;
2694 }
2695
2696 VkResult
2697 anv_device_query_status(struct anv_device *device)
2698 {
2699 /* This isn't likely as most of the callers of this function already check
2700 * for it. However, it doesn't hurt to check and it potentially lets us
2701 * avoid an ioctl.
2702 */
2703 if (anv_device_is_lost(device))
2704 return VK_ERROR_DEVICE_LOST;
2705
2706 uint32_t active, pending;
2707 int ret = anv_gem_gpu_get_reset_stats(device, &active, &pending);
2708 if (ret == -1) {
2709 /* We don't know the real error. */
2710 return anv_device_set_lost(device, "get_reset_stats failed: %m");
2711 }
2712
2713 if (active) {
2714 return anv_device_set_lost(device, "GPU hung on one of our command buffers");
2715 } else if (pending) {
2716 return anv_device_set_lost(device, "GPU hung with commands in-flight");
2717 }
2718
2719 return VK_SUCCESS;
2720 }
2721
2722 VkResult
2723 anv_device_bo_busy(struct anv_device *device, struct anv_bo *bo)
2724 {
2725 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
2726 * Other usages of the BO (such as on different hardware) will not be
2727 * flagged as "busy" by this ioctl. Use with care.
2728 */
2729 int ret = anv_gem_busy(device, bo->gem_handle);
2730 if (ret == 1) {
2731 return VK_NOT_READY;
2732 } else if (ret == -1) {
2733 /* We don't know the real error. */
2734 return anv_device_set_lost(device, "gem wait failed: %m");
2735 }
2736
2737 /* Query for device status after the busy call. If the BO we're checking
2738 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
2739 * client because it clearly doesn't have valid data. Yes, this most
2740 * likely means an ioctl, but we just did an ioctl to query the busy status
2741 * so it's no great loss.
2742 */
2743 return anv_device_query_status(device);
2744 }
2745
2746 VkResult
2747 anv_device_wait(struct anv_device *device, struct anv_bo *bo,
2748 int64_t timeout)
2749 {
2750 int ret = anv_gem_wait(device, bo->gem_handle, &timeout);
2751 if (ret == -1 && errno == ETIME) {
2752 return VK_TIMEOUT;
2753 } else if (ret == -1) {
2754 /* We don't know the real error. */
2755 return anv_device_set_lost(device, "gem wait failed: %m");
2756 }
2757
2758 /* Query for device status after the wait. If the BO we're waiting on got
2759 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
2760 * because it clearly doesn't have valid data. Yes, this most likely means
2761 * an ioctl, but we just did an ioctl to wait so it's no great loss.
2762 */
2763 return anv_device_query_status(device);
2764 }
2765
2766 VkResult anv_DeviceWaitIdle(
2767 VkDevice _device)
2768 {
2769 ANV_FROM_HANDLE(anv_device, device, _device);
2770 if (anv_device_is_lost(device))
2771 return VK_ERROR_DEVICE_LOST;
2772
2773 struct anv_batch batch;
2774
2775 uint32_t cmds[8];
2776 batch.start = batch.next = cmds;
2777 batch.end = (void *) cmds + sizeof(cmds);
2778
2779 anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
2780 anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
2781
2782 return anv_device_submit_simple_batch(device, &batch);
2783 }
2784
2785 bool
2786 anv_vma_alloc(struct anv_device *device, struct anv_bo *bo)
2787 {
2788 if (!(bo->flags & EXEC_OBJECT_PINNED))
2789 return true;
2790
2791 pthread_mutex_lock(&device->vma_mutex);
2792
2793 bo->offset = 0;
2794
2795 if (bo->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS &&
2796 device->vma_hi_available >= bo->size) {
2797 uint64_t addr = util_vma_heap_alloc(&device->vma_hi, bo->size, 4096);
2798 if (addr) {
2799 bo->offset = gen_canonical_address(addr);
2800 assert(addr == gen_48b_address(bo->offset));
2801 device->vma_hi_available -= bo->size;
2802 }
2803 }
2804
2805 if (bo->offset == 0 && device->vma_lo_available >= bo->size) {
2806 uint64_t addr = util_vma_heap_alloc(&device->vma_lo, bo->size, 4096);
2807 if (addr) {
2808 bo->offset = gen_canonical_address(addr);
2809 assert(addr == gen_48b_address(bo->offset));
2810 device->vma_lo_available -= bo->size;
2811 }
2812 }
2813
2814 pthread_mutex_unlock(&device->vma_mutex);
2815
2816 return bo->offset != 0;
2817 }
2818
2819 void
2820 anv_vma_free(struct anv_device *device, struct anv_bo *bo)
2821 {
2822 if (!(bo->flags & EXEC_OBJECT_PINNED))
2823 return;
2824
2825 const uint64_t addr_48b = gen_48b_address(bo->offset);
2826
2827 pthread_mutex_lock(&device->vma_mutex);
2828
2829 if (addr_48b >= LOW_HEAP_MIN_ADDRESS &&
2830 addr_48b <= LOW_HEAP_MAX_ADDRESS) {
2831 util_vma_heap_free(&device->vma_lo, addr_48b, bo->size);
2832 device->vma_lo_available += bo->size;
2833 } else {
2834 ASSERTED const struct anv_physical_device *physical_device =
2835 &device->instance->physicalDevice;
2836 assert(addr_48b >= physical_device->memory.heaps[0].vma_start &&
2837 addr_48b < (physical_device->memory.heaps[0].vma_start +
2838 physical_device->memory.heaps[0].vma_size));
2839 util_vma_heap_free(&device->vma_hi, addr_48b, bo->size);
2840 device->vma_hi_available += bo->size;
2841 }
2842
2843 pthread_mutex_unlock(&device->vma_mutex);
2844
2845 bo->offset = 0;
2846 }
2847
2848 VkResult
2849 anv_bo_init_new(struct anv_bo *bo, struct anv_device *device, uint64_t size)
2850 {
2851 uint32_t gem_handle = anv_gem_create(device, size);
2852 if (!gem_handle)
2853 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
2854
2855 anv_bo_init(bo, gem_handle, size);
2856
2857 return VK_SUCCESS;
2858 }
2859
2860 VkResult anv_AllocateMemory(
2861 VkDevice _device,
2862 const VkMemoryAllocateInfo* pAllocateInfo,
2863 const VkAllocationCallbacks* pAllocator,
2864 VkDeviceMemory* pMem)
2865 {
2866 ANV_FROM_HANDLE(anv_device, device, _device);
2867 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
2868 struct anv_device_memory *mem;
2869 VkResult result = VK_SUCCESS;
2870
2871 assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
2872
2873 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
2874 assert(pAllocateInfo->allocationSize > 0);
2875
2876 if (pAllocateInfo->allocationSize > MAX_MEMORY_ALLOCATION_SIZE)
2877 return VK_ERROR_OUT_OF_DEVICE_MEMORY;
2878
2879 /* FINISHME: Fail if allocation request exceeds heap size. */
2880
2881 mem = vk_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
2882 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2883 if (mem == NULL)
2884 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2885
2886 assert(pAllocateInfo->memoryTypeIndex < pdevice->memory.type_count);
2887 mem->type = &pdevice->memory.types[pAllocateInfo->memoryTypeIndex];
2888 mem->map = NULL;
2889 mem->map_size = 0;
2890 mem->ahw = NULL;
2891 mem->host_ptr = NULL;
2892
2893 uint64_t bo_flags = 0;
2894
2895 assert(mem->type->heapIndex < pdevice->memory.heap_count);
2896 if (pdevice->memory.heaps[mem->type->heapIndex].supports_48bit_addresses)
2897 bo_flags |= EXEC_OBJECT_SUPPORTS_48B_ADDRESS;
2898
2899 const struct wsi_memory_allocate_info *wsi_info =
2900 vk_find_struct_const(pAllocateInfo->pNext, WSI_MEMORY_ALLOCATE_INFO_MESA);
2901 if (wsi_info && wsi_info->implicit_sync) {
2902 /* We need to set the WRITE flag on window system buffers so that GEM
2903 * will know we're writing to them and synchronize uses on other rings
2904 * (eg if the display server uses the blitter ring).
2905 */
2906 bo_flags |= EXEC_OBJECT_WRITE;
2907 } else if (pdevice->has_exec_async) {
2908 bo_flags |= EXEC_OBJECT_ASYNC;
2909 }
2910
2911 if (pdevice->use_softpin)
2912 bo_flags |= EXEC_OBJECT_PINNED;
2913
2914 const VkExportMemoryAllocateInfo *export_info =
2915 vk_find_struct_const(pAllocateInfo->pNext, EXPORT_MEMORY_ALLOCATE_INFO);
2916
2917 /* Check if we need to support Android HW buffer export. If so,
2918 * create AHardwareBuffer and import memory from it.
2919 */
2920 bool android_export = false;
2921 if (export_info && export_info->handleTypes &
2922 VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID)
2923 android_export = true;
2924
2925 /* Android memory import. */
2926 const struct VkImportAndroidHardwareBufferInfoANDROID *ahw_import_info =
2927 vk_find_struct_const(pAllocateInfo->pNext,
2928 IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID);
2929
2930 if (ahw_import_info) {
2931 result = anv_import_ahw_memory(_device, mem, ahw_import_info);
2932 if (result != VK_SUCCESS)
2933 goto fail;
2934
2935 goto success;
2936 } else if (android_export) {
2937 result = anv_create_ahw_memory(_device, mem, pAllocateInfo);
2938 if (result != VK_SUCCESS)
2939 goto fail;
2940
2941 const struct VkImportAndroidHardwareBufferInfoANDROID import_info = {
2942 .buffer = mem->ahw,
2943 };
2944 result = anv_import_ahw_memory(_device, mem, &import_info);
2945 if (result != VK_SUCCESS)
2946 goto fail;
2947
2948 goto success;
2949 }
2950
2951 const VkImportMemoryFdInfoKHR *fd_info =
2952 vk_find_struct_const(pAllocateInfo->pNext, IMPORT_MEMORY_FD_INFO_KHR);
2953
2954 /* The Vulkan spec permits handleType to be 0, in which case the struct is
2955 * ignored.
2956 */
2957 if (fd_info && fd_info->handleType) {
2958 /* At the moment, we support only the below handle types. */
2959 assert(fd_info->handleType ==
2960 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
2961 fd_info->handleType ==
2962 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
2963
2964 result = anv_bo_cache_import(device, &device->bo_cache, fd_info->fd,
2965 bo_flags | ANV_BO_EXTERNAL, &mem->bo);
2966 if (result != VK_SUCCESS)
2967 goto fail;
2968
2969 VkDeviceSize aligned_alloc_size =
2970 align_u64(pAllocateInfo->allocationSize, 4096);
2971
2972 /* For security purposes, we reject importing the bo if it's smaller
2973 * than the requested allocation size. This prevents a malicious client
2974 * from passing a buffer to a trusted client, lying about the size, and
2975 * telling the trusted client to try and texture from an image that goes
2976 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
2977 * in the trusted client. The trusted client can protect itself against
2978 * this sort of attack but only if it can trust the buffer size.
2979 */
2980 if (mem->bo->size < aligned_alloc_size) {
2981 result = vk_errorf(device->instance, device,
2982 VK_ERROR_INVALID_EXTERNAL_HANDLE,
2983 "aligned allocationSize too large for "
2984 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT: "
2985 "%"PRIu64"B > %"PRIu64"B",
2986 aligned_alloc_size, mem->bo->size);
2987 anv_bo_cache_release(device, &device->bo_cache, mem->bo);
2988 goto fail;
2989 }
2990
2991 /* From the Vulkan spec:
2992 *
2993 * "Importing memory from a file descriptor transfers ownership of
2994 * the file descriptor from the application to the Vulkan
2995 * implementation. The application must not perform any operations on
2996 * the file descriptor after a successful import."
2997 *
2998 * If the import fails, we leave the file descriptor open.
2999 */
3000 close(fd_info->fd);
3001 goto success;
3002 }
3003
3004 const VkImportMemoryHostPointerInfoEXT *host_ptr_info =
3005 vk_find_struct_const(pAllocateInfo->pNext,
3006 IMPORT_MEMORY_HOST_POINTER_INFO_EXT);
3007 if (host_ptr_info && host_ptr_info->handleType) {
3008 if (host_ptr_info->handleType ==
3009 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_MAPPED_FOREIGN_MEMORY_BIT_EXT) {
3010 result = vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE);
3011 goto fail;
3012 }
3013
3014 assert(host_ptr_info->handleType ==
3015 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT);
3016
3017 result = anv_bo_cache_import_host_ptr(
3018 device, &device->bo_cache, host_ptr_info->pHostPointer,
3019 pAllocateInfo->allocationSize, bo_flags, &mem->bo);
3020
3021 if (result != VK_SUCCESS)
3022 goto fail;
3023
3024 mem->host_ptr = host_ptr_info->pHostPointer;
3025 goto success;
3026 }
3027
3028 /* Regular allocate (not importing memory). */
3029
3030 if (export_info && export_info->handleTypes)
3031 bo_flags |= ANV_BO_EXTERNAL;
3032
3033 result = anv_bo_cache_alloc(device, &device->bo_cache,
3034 pAllocateInfo->allocationSize, bo_flags,
3035 &mem->bo);
3036 if (result != VK_SUCCESS)
3037 goto fail;
3038
3039 const VkMemoryDedicatedAllocateInfo *dedicated_info =
3040 vk_find_struct_const(pAllocateInfo->pNext, MEMORY_DEDICATED_ALLOCATE_INFO);
3041 if (dedicated_info && dedicated_info->image != VK_NULL_HANDLE) {
3042 ANV_FROM_HANDLE(anv_image, image, dedicated_info->image);
3043
3044 /* Some legacy (non-modifiers) consumers need the tiling to be set on
3045 * the BO. In this case, we have a dedicated allocation.
3046 */
3047 if (image->needs_set_tiling) {
3048 const uint32_t i915_tiling =
3049 isl_tiling_to_i915_tiling(image->planes[0].surface.isl.tiling);
3050 int ret = anv_gem_set_tiling(device, mem->bo->gem_handle,
3051 image->planes[0].surface.isl.row_pitch_B,
3052 i915_tiling);
3053 if (ret) {
3054 anv_bo_cache_release(device, &device->bo_cache, mem->bo);
3055 return vk_errorf(device->instance, NULL,
3056 VK_ERROR_OUT_OF_DEVICE_MEMORY,
3057 "failed to set BO tiling: %m");
3058 }
3059 }
3060 }
3061
3062 success:
3063 pthread_mutex_lock(&device->mutex);
3064 list_addtail(&mem->link, &device->memory_objects);
3065 pthread_mutex_unlock(&device->mutex);
3066
3067 *pMem = anv_device_memory_to_handle(mem);
3068
3069 p_atomic_add(&pdevice->memory.heaps[mem->type->heapIndex].used,
3070 mem->bo->size);
3071
3072 return VK_SUCCESS;
3073
3074 fail:
3075 vk_free2(&device->alloc, pAllocator, mem);
3076
3077 return result;
3078 }
3079
3080 VkResult anv_GetMemoryFdKHR(
3081 VkDevice device_h,
3082 const VkMemoryGetFdInfoKHR* pGetFdInfo,
3083 int* pFd)
3084 {
3085 ANV_FROM_HANDLE(anv_device, dev, device_h);
3086 ANV_FROM_HANDLE(anv_device_memory, mem, pGetFdInfo->memory);
3087
3088 assert(pGetFdInfo->sType == VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR);
3089
3090 assert(pGetFdInfo->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
3091 pGetFdInfo->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
3092
3093 return anv_bo_cache_export(dev, &dev->bo_cache, mem->bo, pFd);
3094 }
3095
3096 VkResult anv_GetMemoryFdPropertiesKHR(
3097 VkDevice _device,
3098 VkExternalMemoryHandleTypeFlagBits handleType,
3099 int fd,
3100 VkMemoryFdPropertiesKHR* pMemoryFdProperties)
3101 {
3102 ANV_FROM_HANDLE(anv_device, device, _device);
3103 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3104
3105 switch (handleType) {
3106 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT:
3107 /* dma-buf can be imported as any memory type */
3108 pMemoryFdProperties->memoryTypeBits =
3109 (1 << pdevice->memory.type_count) - 1;
3110 return VK_SUCCESS;
3111
3112 default:
3113 /* The valid usage section for this function says:
3114 *
3115 * "handleType must not be one of the handle types defined as
3116 * opaque."
3117 *
3118 * So opaque handle types fall into the default "unsupported" case.
3119 */
3120 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE);
3121 }
3122 }
3123
3124 VkResult anv_GetMemoryHostPointerPropertiesEXT(
3125 VkDevice _device,
3126 VkExternalMemoryHandleTypeFlagBits handleType,
3127 const void* pHostPointer,
3128 VkMemoryHostPointerPropertiesEXT* pMemoryHostPointerProperties)
3129 {
3130 ANV_FROM_HANDLE(anv_device, device, _device);
3131
3132 assert(pMemoryHostPointerProperties->sType ==
3133 VK_STRUCTURE_TYPE_MEMORY_HOST_POINTER_PROPERTIES_EXT);
3134
3135 switch (handleType) {
3136 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT: {
3137 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3138
3139 /* Host memory can be imported as any memory type. */
3140 pMemoryHostPointerProperties->memoryTypeBits =
3141 (1ull << pdevice->memory.type_count) - 1;
3142
3143 return VK_SUCCESS;
3144 }
3145 default:
3146 return VK_ERROR_INVALID_EXTERNAL_HANDLE;
3147 }
3148 }
3149
3150 void anv_FreeMemory(
3151 VkDevice _device,
3152 VkDeviceMemory _mem,
3153 const VkAllocationCallbacks* pAllocator)
3154 {
3155 ANV_FROM_HANDLE(anv_device, device, _device);
3156 ANV_FROM_HANDLE(anv_device_memory, mem, _mem);
3157 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3158
3159 if (mem == NULL)
3160 return;
3161
3162 pthread_mutex_lock(&device->mutex);
3163 list_del(&mem->link);
3164 pthread_mutex_unlock(&device->mutex);
3165
3166 if (mem->map)
3167 anv_UnmapMemory(_device, _mem);
3168
3169 p_atomic_add(&pdevice->memory.heaps[mem->type->heapIndex].used,
3170 -mem->bo->size);
3171
3172 anv_bo_cache_release(device, &device->bo_cache, mem->bo);
3173
3174 #if defined(ANDROID) && ANDROID_API_LEVEL >= 26
3175 if (mem->ahw)
3176 AHardwareBuffer_release(mem->ahw);
3177 #endif
3178
3179 vk_free2(&device->alloc, pAllocator, mem);
3180 }
3181
3182 VkResult anv_MapMemory(
3183 VkDevice _device,
3184 VkDeviceMemory _memory,
3185 VkDeviceSize offset,
3186 VkDeviceSize size,
3187 VkMemoryMapFlags flags,
3188 void** ppData)
3189 {
3190 ANV_FROM_HANDLE(anv_device, device, _device);
3191 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
3192
3193 if (mem == NULL) {
3194 *ppData = NULL;
3195 return VK_SUCCESS;
3196 }
3197
3198 if (mem->host_ptr) {
3199 *ppData = mem->host_ptr + offset;
3200 return VK_SUCCESS;
3201 }
3202
3203 if (size == VK_WHOLE_SIZE)
3204 size = mem->bo->size - offset;
3205
3206 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
3207 *
3208 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
3209 * assert(size != 0);
3210 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
3211 * equal to the size of the memory minus offset
3212 */
3213 assert(size > 0);
3214 assert(offset + size <= mem->bo->size);
3215
3216 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
3217 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
3218 * at a time is valid. We could just mmap up front and return an offset
3219 * pointer here, but that may exhaust virtual memory on 32 bit
3220 * userspace. */
3221
3222 uint32_t gem_flags = 0;
3223
3224 if (!device->info.has_llc &&
3225 (mem->type->propertyFlags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT))
3226 gem_flags |= I915_MMAP_WC;
3227
3228 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
3229 uint64_t map_offset = offset & ~4095ull;
3230 assert(offset >= map_offset);
3231 uint64_t map_size = (offset + size) - map_offset;
3232
3233 /* Let's map whole pages */
3234 map_size = align_u64(map_size, 4096);
3235
3236 void *map = anv_gem_mmap(device, mem->bo->gem_handle,
3237 map_offset, map_size, gem_flags);
3238 if (map == MAP_FAILED)
3239 return vk_error(VK_ERROR_MEMORY_MAP_FAILED);
3240
3241 mem->map = map;
3242 mem->map_size = map_size;
3243
3244 *ppData = mem->map + (offset - map_offset);
3245
3246 return VK_SUCCESS;
3247 }
3248
3249 void anv_UnmapMemory(
3250 VkDevice _device,
3251 VkDeviceMemory _memory)
3252 {
3253 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
3254
3255 if (mem == NULL || mem->host_ptr)
3256 return;
3257
3258 anv_gem_munmap(mem->map, mem->map_size);
3259
3260 mem->map = NULL;
3261 mem->map_size = 0;
3262 }
3263
3264 static void
3265 clflush_mapped_ranges(struct anv_device *device,
3266 uint32_t count,
3267 const VkMappedMemoryRange *ranges)
3268 {
3269 for (uint32_t i = 0; i < count; i++) {
3270 ANV_FROM_HANDLE(anv_device_memory, mem, ranges[i].memory);
3271 if (ranges[i].offset >= mem->map_size)
3272 continue;
3273
3274 gen_clflush_range(mem->map + ranges[i].offset,
3275 MIN2(ranges[i].size, mem->map_size - ranges[i].offset));
3276 }
3277 }
3278
3279 VkResult anv_FlushMappedMemoryRanges(
3280 VkDevice _device,
3281 uint32_t memoryRangeCount,
3282 const VkMappedMemoryRange* pMemoryRanges)
3283 {
3284 ANV_FROM_HANDLE(anv_device, device, _device);
3285
3286 if (device->info.has_llc)
3287 return VK_SUCCESS;
3288
3289 /* Make sure the writes we're flushing have landed. */
3290 __builtin_ia32_mfence();
3291
3292 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
3293
3294 return VK_SUCCESS;
3295 }
3296
3297 VkResult anv_InvalidateMappedMemoryRanges(
3298 VkDevice _device,
3299 uint32_t memoryRangeCount,
3300 const VkMappedMemoryRange* pMemoryRanges)
3301 {
3302 ANV_FROM_HANDLE(anv_device, device, _device);
3303
3304 if (device->info.has_llc)
3305 return VK_SUCCESS;
3306
3307 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
3308
3309 /* Make sure no reads get moved up above the invalidate. */
3310 __builtin_ia32_mfence();
3311
3312 return VK_SUCCESS;
3313 }
3314
3315 void anv_GetBufferMemoryRequirements(
3316 VkDevice _device,
3317 VkBuffer _buffer,
3318 VkMemoryRequirements* pMemoryRequirements)
3319 {
3320 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
3321 ANV_FROM_HANDLE(anv_device, device, _device);
3322 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3323
3324 /* The Vulkan spec (git aaed022) says:
3325 *
3326 * memoryTypeBits is a bitfield and contains one bit set for every
3327 * supported memory type for the resource. The bit `1<<i` is set if and
3328 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3329 * structure for the physical device is supported.
3330 */
3331 uint32_t memory_types = 0;
3332 for (uint32_t i = 0; i < pdevice->memory.type_count; i++) {
3333 uint32_t valid_usage = pdevice->memory.types[i].valid_buffer_usage;
3334 if ((valid_usage & buffer->usage) == buffer->usage)
3335 memory_types |= (1u << i);
3336 }
3337
3338 /* Base alignment requirement of a cache line */
3339 uint32_t alignment = 16;
3340
3341 /* We need an alignment of 32 for pushing UBOs */
3342 if (buffer->usage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT)
3343 alignment = MAX2(alignment, 32);
3344
3345 pMemoryRequirements->size = buffer->size;
3346 pMemoryRequirements->alignment = alignment;
3347
3348 /* Storage and Uniform buffers should have their size aligned to
3349 * 32-bits to avoid boundary checks when last DWord is not complete.
3350 * This would ensure that not internal padding would be needed for
3351 * 16-bit types.
3352 */
3353 if (device->robust_buffer_access &&
3354 (buffer->usage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT ||
3355 buffer->usage & VK_BUFFER_USAGE_STORAGE_BUFFER_BIT))
3356 pMemoryRequirements->size = align_u64(buffer->size, 4);
3357
3358 pMemoryRequirements->memoryTypeBits = memory_types;
3359 }
3360
3361 void anv_GetBufferMemoryRequirements2(
3362 VkDevice _device,
3363 const VkBufferMemoryRequirementsInfo2* pInfo,
3364 VkMemoryRequirements2* pMemoryRequirements)
3365 {
3366 anv_GetBufferMemoryRequirements(_device, pInfo->buffer,
3367 &pMemoryRequirements->memoryRequirements);
3368
3369 vk_foreach_struct(ext, pMemoryRequirements->pNext) {
3370 switch (ext->sType) {
3371 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
3372 VkMemoryDedicatedRequirements *requirements = (void *)ext;
3373 requirements->prefersDedicatedAllocation = false;
3374 requirements->requiresDedicatedAllocation = false;
3375 break;
3376 }
3377
3378 default:
3379 anv_debug_ignored_stype(ext->sType);
3380 break;
3381 }
3382 }
3383 }
3384
3385 void anv_GetImageMemoryRequirements(
3386 VkDevice _device,
3387 VkImage _image,
3388 VkMemoryRequirements* pMemoryRequirements)
3389 {
3390 ANV_FROM_HANDLE(anv_image, image, _image);
3391 ANV_FROM_HANDLE(anv_device, device, _device);
3392 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3393
3394 /* The Vulkan spec (git aaed022) says:
3395 *
3396 * memoryTypeBits is a bitfield and contains one bit set for every
3397 * supported memory type for the resource. The bit `1<<i` is set if and
3398 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3399 * structure for the physical device is supported.
3400 *
3401 * All types are currently supported for images.
3402 */
3403 uint32_t memory_types = (1ull << pdevice->memory.type_count) - 1;
3404
3405 /* We must have image allocated or imported at this point. According to the
3406 * specification, external images must have been bound to memory before
3407 * calling GetImageMemoryRequirements.
3408 */
3409 assert(image->size > 0);
3410
3411 pMemoryRequirements->size = image->size;
3412 pMemoryRequirements->alignment = image->alignment;
3413 pMemoryRequirements->memoryTypeBits = memory_types;
3414 }
3415
3416 void anv_GetImageMemoryRequirements2(
3417 VkDevice _device,
3418 const VkImageMemoryRequirementsInfo2* pInfo,
3419 VkMemoryRequirements2* pMemoryRequirements)
3420 {
3421 ANV_FROM_HANDLE(anv_device, device, _device);
3422 ANV_FROM_HANDLE(anv_image, image, pInfo->image);
3423
3424 anv_GetImageMemoryRequirements(_device, pInfo->image,
3425 &pMemoryRequirements->memoryRequirements);
3426
3427 vk_foreach_struct_const(ext, pInfo->pNext) {
3428 switch (ext->sType) {
3429 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO: {
3430 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3431 const VkImagePlaneMemoryRequirementsInfo *plane_reqs =
3432 (const VkImagePlaneMemoryRequirementsInfo *) ext;
3433 uint32_t plane = anv_image_aspect_to_plane(image->aspects,
3434 plane_reqs->planeAspect);
3435
3436 assert(image->planes[plane].offset == 0);
3437
3438 /* The Vulkan spec (git aaed022) says:
3439 *
3440 * memoryTypeBits is a bitfield and contains one bit set for every
3441 * supported memory type for the resource. The bit `1<<i` is set
3442 * if and only if the memory type `i` in the
3443 * VkPhysicalDeviceMemoryProperties structure for the physical
3444 * device is supported.
3445 *
3446 * All types are currently supported for images.
3447 */
3448 pMemoryRequirements->memoryRequirements.memoryTypeBits =
3449 (1ull << pdevice->memory.type_count) - 1;
3450
3451 /* We must have image allocated or imported at this point. According to the
3452 * specification, external images must have been bound to memory before
3453 * calling GetImageMemoryRequirements.
3454 */
3455 assert(image->planes[plane].size > 0);
3456
3457 pMemoryRequirements->memoryRequirements.size = image->planes[plane].size;
3458 pMemoryRequirements->memoryRequirements.alignment =
3459 image->planes[plane].alignment;
3460 break;
3461 }
3462
3463 default:
3464 anv_debug_ignored_stype(ext->sType);
3465 break;
3466 }
3467 }
3468
3469 vk_foreach_struct(ext, pMemoryRequirements->pNext) {
3470 switch (ext->sType) {
3471 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
3472 VkMemoryDedicatedRequirements *requirements = (void *)ext;
3473 if (image->needs_set_tiling || image->external_format) {
3474 /* If we need to set the tiling for external consumers, we need a
3475 * dedicated allocation.
3476 *
3477 * See also anv_AllocateMemory.
3478 */
3479 requirements->prefersDedicatedAllocation = true;
3480 requirements->requiresDedicatedAllocation = true;
3481 } else {
3482 requirements->prefersDedicatedAllocation = false;
3483 requirements->requiresDedicatedAllocation = false;
3484 }
3485 break;
3486 }
3487
3488 default:
3489 anv_debug_ignored_stype(ext->sType);
3490 break;
3491 }
3492 }
3493 }
3494
3495 void anv_GetImageSparseMemoryRequirements(
3496 VkDevice device,
3497 VkImage image,
3498 uint32_t* pSparseMemoryRequirementCount,
3499 VkSparseImageMemoryRequirements* pSparseMemoryRequirements)
3500 {
3501 *pSparseMemoryRequirementCount = 0;
3502 }
3503
3504 void anv_GetImageSparseMemoryRequirements2(
3505 VkDevice device,
3506 const VkImageSparseMemoryRequirementsInfo2* pInfo,
3507 uint32_t* pSparseMemoryRequirementCount,
3508 VkSparseImageMemoryRequirements2* pSparseMemoryRequirements)
3509 {
3510 *pSparseMemoryRequirementCount = 0;
3511 }
3512
3513 void anv_GetDeviceMemoryCommitment(
3514 VkDevice device,
3515 VkDeviceMemory memory,
3516 VkDeviceSize* pCommittedMemoryInBytes)
3517 {
3518 *pCommittedMemoryInBytes = 0;
3519 }
3520
3521 static void
3522 anv_bind_buffer_memory(const VkBindBufferMemoryInfo *pBindInfo)
3523 {
3524 ANV_FROM_HANDLE(anv_device_memory, mem, pBindInfo->memory);
3525 ANV_FROM_HANDLE(anv_buffer, buffer, pBindInfo->buffer);
3526
3527 assert(pBindInfo->sType == VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO);
3528
3529 if (mem) {
3530 assert((buffer->usage & mem->type->valid_buffer_usage) == buffer->usage);
3531 buffer->address = (struct anv_address) {
3532 .bo = mem->bo,
3533 .offset = pBindInfo->memoryOffset,
3534 };
3535 } else {
3536 buffer->address = ANV_NULL_ADDRESS;
3537 }
3538 }
3539
3540 VkResult anv_BindBufferMemory(
3541 VkDevice device,
3542 VkBuffer buffer,
3543 VkDeviceMemory memory,
3544 VkDeviceSize memoryOffset)
3545 {
3546 anv_bind_buffer_memory(
3547 &(VkBindBufferMemoryInfo) {
3548 .sType = VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO,
3549 .buffer = buffer,
3550 .memory = memory,
3551 .memoryOffset = memoryOffset,
3552 });
3553
3554 return VK_SUCCESS;
3555 }
3556
3557 VkResult anv_BindBufferMemory2(
3558 VkDevice device,
3559 uint32_t bindInfoCount,
3560 const VkBindBufferMemoryInfo* pBindInfos)
3561 {
3562 for (uint32_t i = 0; i < bindInfoCount; i++)
3563 anv_bind_buffer_memory(&pBindInfos[i]);
3564
3565 return VK_SUCCESS;
3566 }
3567
3568 VkResult anv_QueueBindSparse(
3569 VkQueue _queue,
3570 uint32_t bindInfoCount,
3571 const VkBindSparseInfo* pBindInfo,
3572 VkFence fence)
3573 {
3574 ANV_FROM_HANDLE(anv_queue, queue, _queue);
3575 if (anv_device_is_lost(queue->device))
3576 return VK_ERROR_DEVICE_LOST;
3577
3578 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
3579 }
3580
3581 // Event functions
3582
3583 VkResult anv_CreateEvent(
3584 VkDevice _device,
3585 const VkEventCreateInfo* pCreateInfo,
3586 const VkAllocationCallbacks* pAllocator,
3587 VkEvent* pEvent)
3588 {
3589 ANV_FROM_HANDLE(anv_device, device, _device);
3590 struct anv_state state;
3591 struct anv_event *event;
3592
3593 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_EVENT_CREATE_INFO);
3594
3595 state = anv_state_pool_alloc(&device->dynamic_state_pool,
3596 sizeof(*event), 8);
3597 event = state.map;
3598 event->state = state;
3599 event->semaphore = VK_EVENT_RESET;
3600
3601 if (!device->info.has_llc) {
3602 /* Make sure the writes we're flushing have landed. */
3603 __builtin_ia32_mfence();
3604 __builtin_ia32_clflush(event);
3605 }
3606
3607 *pEvent = anv_event_to_handle(event);
3608
3609 return VK_SUCCESS;
3610 }
3611
3612 void anv_DestroyEvent(
3613 VkDevice _device,
3614 VkEvent _event,
3615 const VkAllocationCallbacks* pAllocator)
3616 {
3617 ANV_FROM_HANDLE(anv_device, device, _device);
3618 ANV_FROM_HANDLE(anv_event, event, _event);
3619
3620 if (!event)
3621 return;
3622
3623 anv_state_pool_free(&device->dynamic_state_pool, event->state);
3624 }
3625
3626 VkResult anv_GetEventStatus(
3627 VkDevice _device,
3628 VkEvent _event)
3629 {
3630 ANV_FROM_HANDLE(anv_device, device, _device);
3631 ANV_FROM_HANDLE(anv_event, event, _event);
3632
3633 if (anv_device_is_lost(device))
3634 return VK_ERROR_DEVICE_LOST;
3635
3636 if (!device->info.has_llc) {
3637 /* Invalidate read cache before reading event written by GPU. */
3638 __builtin_ia32_clflush(event);
3639 __builtin_ia32_mfence();
3640
3641 }
3642
3643 return event->semaphore;
3644 }
3645
3646 VkResult anv_SetEvent(
3647 VkDevice _device,
3648 VkEvent _event)
3649 {
3650 ANV_FROM_HANDLE(anv_device, device, _device);
3651 ANV_FROM_HANDLE(anv_event, event, _event);
3652
3653 event->semaphore = VK_EVENT_SET;
3654
3655 if (!device->info.has_llc) {
3656 /* Make sure the writes we're flushing have landed. */
3657 __builtin_ia32_mfence();
3658 __builtin_ia32_clflush(event);
3659 }
3660
3661 return VK_SUCCESS;
3662 }
3663
3664 VkResult anv_ResetEvent(
3665 VkDevice _device,
3666 VkEvent _event)
3667 {
3668 ANV_FROM_HANDLE(anv_device, device, _device);
3669 ANV_FROM_HANDLE(anv_event, event, _event);
3670
3671 event->semaphore = VK_EVENT_RESET;
3672
3673 if (!device->info.has_llc) {
3674 /* Make sure the writes we're flushing have landed. */
3675 __builtin_ia32_mfence();
3676 __builtin_ia32_clflush(event);
3677 }
3678
3679 return VK_SUCCESS;
3680 }
3681
3682 // Buffer functions
3683
3684 VkResult anv_CreateBuffer(
3685 VkDevice _device,
3686 const VkBufferCreateInfo* pCreateInfo,
3687 const VkAllocationCallbacks* pAllocator,
3688 VkBuffer* pBuffer)
3689 {
3690 ANV_FROM_HANDLE(anv_device, device, _device);
3691 struct anv_buffer *buffer;
3692
3693 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
3694
3695 buffer = vk_alloc2(&device->alloc, pAllocator, sizeof(*buffer), 8,
3696 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
3697 if (buffer == NULL)
3698 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
3699
3700 buffer->size = pCreateInfo->size;
3701 buffer->usage = pCreateInfo->usage;
3702 buffer->address = ANV_NULL_ADDRESS;
3703
3704 *pBuffer = anv_buffer_to_handle(buffer);
3705
3706 return VK_SUCCESS;
3707 }
3708
3709 void anv_DestroyBuffer(
3710 VkDevice _device,
3711 VkBuffer _buffer,
3712 const VkAllocationCallbacks* pAllocator)
3713 {
3714 ANV_FROM_HANDLE(anv_device, device, _device);
3715 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
3716
3717 if (!buffer)
3718 return;
3719
3720 vk_free2(&device->alloc, pAllocator, buffer);
3721 }
3722
3723 VkDeviceAddress anv_GetBufferDeviceAddressEXT(
3724 VkDevice device,
3725 const VkBufferDeviceAddressInfoEXT* pInfo)
3726 {
3727 ANV_FROM_HANDLE(anv_buffer, buffer, pInfo->buffer);
3728
3729 assert(buffer->address.bo->flags & EXEC_OBJECT_PINNED);
3730
3731 return anv_address_physical(buffer->address);
3732 }
3733
3734 void
3735 anv_fill_buffer_surface_state(struct anv_device *device, struct anv_state state,
3736 enum isl_format format,
3737 struct anv_address address,
3738 uint32_t range, uint32_t stride)
3739 {
3740 isl_buffer_fill_state(&device->isl_dev, state.map,
3741 .address = anv_address_physical(address),
3742 .mocs = device->default_mocs,
3743 .size_B = range,
3744 .format = format,
3745 .swizzle = ISL_SWIZZLE_IDENTITY,
3746 .stride_B = stride);
3747 }
3748
3749 void anv_DestroySampler(
3750 VkDevice _device,
3751 VkSampler _sampler,
3752 const VkAllocationCallbacks* pAllocator)
3753 {
3754 ANV_FROM_HANDLE(anv_device, device, _device);
3755 ANV_FROM_HANDLE(anv_sampler, sampler, _sampler);
3756
3757 if (!sampler)
3758 return;
3759
3760 if (sampler->bindless_state.map) {
3761 anv_state_pool_free(&device->dynamic_state_pool,
3762 sampler->bindless_state);
3763 }
3764
3765 vk_free2(&device->alloc, pAllocator, sampler);
3766 }
3767
3768 VkResult anv_CreateFramebuffer(
3769 VkDevice _device,
3770 const VkFramebufferCreateInfo* pCreateInfo,
3771 const VkAllocationCallbacks* pAllocator,
3772 VkFramebuffer* pFramebuffer)
3773 {
3774 ANV_FROM_HANDLE(anv_device, device, _device);
3775 struct anv_framebuffer *framebuffer;
3776
3777 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
3778
3779 size_t size = sizeof(*framebuffer);
3780
3781 /* VK_KHR_imageless_framebuffer extension says:
3782 *
3783 * If flags includes VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR,
3784 * parameter pAttachments is ignored.
3785 */
3786 if (!(pCreateInfo->flags & VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR)) {
3787 size += sizeof(struct anv_image_view *) * pCreateInfo->attachmentCount;
3788 framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
3789 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
3790 if (framebuffer == NULL)
3791 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
3792
3793 for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
3794 ANV_FROM_HANDLE(anv_image_view, iview, pCreateInfo->pAttachments[i]);
3795 framebuffer->attachments[i] = iview;
3796 }
3797 framebuffer->attachment_count = pCreateInfo->attachmentCount;
3798 } else {
3799 assert(device->enabled_extensions.KHR_imageless_framebuffer);
3800 framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
3801 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
3802 if (framebuffer == NULL)
3803 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
3804
3805 framebuffer->attachment_count = 0;
3806 }
3807
3808 framebuffer->width = pCreateInfo->width;
3809 framebuffer->height = pCreateInfo->height;
3810 framebuffer->layers = pCreateInfo->layers;
3811
3812 *pFramebuffer = anv_framebuffer_to_handle(framebuffer);
3813
3814 return VK_SUCCESS;
3815 }
3816
3817 void anv_DestroyFramebuffer(
3818 VkDevice _device,
3819 VkFramebuffer _fb,
3820 const VkAllocationCallbacks* pAllocator)
3821 {
3822 ANV_FROM_HANDLE(anv_device, device, _device);
3823 ANV_FROM_HANDLE(anv_framebuffer, fb, _fb);
3824
3825 if (!fb)
3826 return;
3827
3828 vk_free2(&device->alloc, pAllocator, fb);
3829 }
3830
3831 static const VkTimeDomainEXT anv_time_domains[] = {
3832 VK_TIME_DOMAIN_DEVICE_EXT,
3833 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT,
3834 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT,
3835 };
3836
3837 VkResult anv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
3838 VkPhysicalDevice physicalDevice,
3839 uint32_t *pTimeDomainCount,
3840 VkTimeDomainEXT *pTimeDomains)
3841 {
3842 int d;
3843 VK_OUTARRAY_MAKE(out, pTimeDomains, pTimeDomainCount);
3844
3845 for (d = 0; d < ARRAY_SIZE(anv_time_domains); d++) {
3846 vk_outarray_append(&out, i) {
3847 *i = anv_time_domains[d];
3848 }
3849 }
3850
3851 return vk_outarray_status(&out);
3852 }
3853
3854 static uint64_t
3855 anv_clock_gettime(clockid_t clock_id)
3856 {
3857 struct timespec current;
3858 int ret;
3859
3860 ret = clock_gettime(clock_id, &current);
3861 if (ret < 0 && clock_id == CLOCK_MONOTONIC_RAW)
3862 ret = clock_gettime(CLOCK_MONOTONIC, &current);
3863 if (ret < 0)
3864 return 0;
3865
3866 return (uint64_t) current.tv_sec * 1000000000ULL + current.tv_nsec;
3867 }
3868
3869 #define TIMESTAMP 0x2358
3870
3871 VkResult anv_GetCalibratedTimestampsEXT(
3872 VkDevice _device,
3873 uint32_t timestampCount,
3874 const VkCalibratedTimestampInfoEXT *pTimestampInfos,
3875 uint64_t *pTimestamps,
3876 uint64_t *pMaxDeviation)
3877 {
3878 ANV_FROM_HANDLE(anv_device, device, _device);
3879 uint64_t timestamp_frequency = device->info.timestamp_frequency;
3880 int ret;
3881 int d;
3882 uint64_t begin, end;
3883 uint64_t max_clock_period = 0;
3884
3885 begin = anv_clock_gettime(CLOCK_MONOTONIC_RAW);
3886
3887 for (d = 0; d < timestampCount; d++) {
3888 switch (pTimestampInfos[d].timeDomain) {
3889 case VK_TIME_DOMAIN_DEVICE_EXT:
3890 ret = anv_gem_reg_read(device, TIMESTAMP | 1,
3891 &pTimestamps[d]);
3892
3893 if (ret != 0) {
3894 return anv_device_set_lost(device, "Failed to read the TIMESTAMP "
3895 "register: %m");
3896 }
3897 uint64_t device_period = DIV_ROUND_UP(1000000000, timestamp_frequency);
3898 max_clock_period = MAX2(max_clock_period, device_period);
3899 break;
3900 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT:
3901 pTimestamps[d] = anv_clock_gettime(CLOCK_MONOTONIC);
3902 max_clock_period = MAX2(max_clock_period, 1);
3903 break;
3904
3905 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT:
3906 pTimestamps[d] = begin;
3907 break;
3908 default:
3909 pTimestamps[d] = 0;
3910 break;
3911 }
3912 }
3913
3914 end = anv_clock_gettime(CLOCK_MONOTONIC_RAW);
3915
3916 /*
3917 * The maximum deviation is the sum of the interval over which we
3918 * perform the sampling and the maximum period of any sampled
3919 * clock. That's because the maximum skew between any two sampled
3920 * clock edges is when the sampled clock with the largest period is
3921 * sampled at the end of that period but right at the beginning of the
3922 * sampling interval and some other clock is sampled right at the
3923 * begining of its sampling period and right at the end of the
3924 * sampling interval. Let's assume the GPU has the longest clock
3925 * period and that the application is sampling GPU and monotonic:
3926 *
3927 * s e
3928 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
3929 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
3930 *
3931 * g
3932 * 0 1 2 3
3933 * GPU -----_____-----_____-----_____-----_____
3934 *
3935 * m
3936 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
3937 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
3938 *
3939 * Interval <----------------->
3940 * Deviation <-------------------------->
3941 *
3942 * s = read(raw) 2
3943 * g = read(GPU) 1
3944 * m = read(monotonic) 2
3945 * e = read(raw) b
3946 *
3947 * We round the sample interval up by one tick to cover sampling error
3948 * in the interval clock
3949 */
3950
3951 uint64_t sample_interval = end - begin + 1;
3952
3953 *pMaxDeviation = sample_interval + max_clock_period;
3954
3955 return VK_SUCCESS;
3956 }
3957
3958 /* vk_icd.h does not declare this function, so we declare it here to
3959 * suppress Wmissing-prototypes.
3960 */
3961 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
3962 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion);
3963
3964 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
3965 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion)
3966 {
3967 /* For the full details on loader interface versioning, see
3968 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
3969 * What follows is a condensed summary, to help you navigate the large and
3970 * confusing official doc.
3971 *
3972 * - Loader interface v0 is incompatible with later versions. We don't
3973 * support it.
3974 *
3975 * - In loader interface v1:
3976 * - The first ICD entrypoint called by the loader is
3977 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
3978 * entrypoint.
3979 * - The ICD must statically expose no other Vulkan symbol unless it is
3980 * linked with -Bsymbolic.
3981 * - Each dispatchable Vulkan handle created by the ICD must be
3982 * a pointer to a struct whose first member is VK_LOADER_DATA. The
3983 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
3984 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
3985 * vkDestroySurfaceKHR(). The ICD must be capable of working with
3986 * such loader-managed surfaces.
3987 *
3988 * - Loader interface v2 differs from v1 in:
3989 * - The first ICD entrypoint called by the loader is
3990 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
3991 * statically expose this entrypoint.
3992 *
3993 * - Loader interface v3 differs from v2 in:
3994 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
3995 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
3996 * because the loader no longer does so.
3997 */
3998 *pSupportedVersion = MIN2(*pSupportedVersion, 3u);
3999 return VK_SUCCESS;
4000 }