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