projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
anv: Add support for new MMAP_OFFSET ioctl.
[mesa.git] / src / intel / vulkan / anv_device.c
1 /*
2 * Copyright © 2015 Intel Corporation
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 */
23
24 #include <assert.h>
25 #include <stdbool.h>
26 #include <string.h>
27 #include <sys/mman.h>
28 #include <sys/sysinfo.h>
29 #include <unistd.h>
30 #include <fcntl.h>
31 #include <xf86drm.h>
32 #include "drm-uapi/drm_fourcc.h"
33
34 #include "anv_private.h"
35 #include "util/debug.h"
36 #include "util/build_id.h"
37 #include "util/disk_cache.h"
38 #include "util/mesa-sha1.h"
39 #include "util/os_file.h"
40 #include "util/u_atomic.h"
41 #include "util/u_string.h"
42 #include "util/xmlpool.h"
43 #include "git_sha1.h"
44 #include "vk_util.h"
45 #include "common/gen_aux_map.h"
46 #include "common/gen_defines.h"
47 #include "compiler/glsl_types.h"
48
49 #include "genxml/gen7_pack.h"
50
51 static const char anv_dri_options_xml[] =
52 DRI_CONF_BEGIN
53 DRI_CONF_SECTION_PERFORMANCE
54 DRI_CONF_VK_X11_OVERRIDE_MIN_IMAGE_COUNT(0)
55 DRI_CONF_VK_X11_STRICT_IMAGE_COUNT("false")
56 DRI_CONF_SECTION_END
57
58 DRI_CONF_SECTION_DEBUG
59 DRI_CONF_ALWAYS_FLUSH_CACHE("false")
60 DRI_CONF_VK_WSI_FORCE_BGRA8_UNORM_FIRST("false")
61 DRI_CONF_SECTION_END
62 DRI_CONF_END;
63
64 /* This is probably far to big but it reflects the max size used for messages
65 * in OpenGLs KHR_debug.
66 */
67 #define MAX_DEBUG_MESSAGE_LENGTH 4096
68
69 static void
70 compiler_debug_log(void *data, const char *fmt, ...)
71 {
72 char str[MAX_DEBUG_MESSAGE_LENGTH];
73 struct anv_device *device = (struct anv_device *)data;
74 struct anv_instance *instance = device->physical->instance;
75
76 if (list_is_empty(&instance->debug_report_callbacks.callbacks))
77 return;
78
79 va_list args;
80 va_start(args, fmt);
81 (void) vsnprintf(str, MAX_DEBUG_MESSAGE_LENGTH, fmt, args);
82 va_end(args);
83
84 vk_debug_report(&instance->debug_report_callbacks,
85 VK_DEBUG_REPORT_DEBUG_BIT_EXT,
86 VK_DEBUG_REPORT_OBJECT_TYPE_UNKNOWN_EXT,
87 0, 0, 0, "anv", str);
88 }
89
90 static void
91 compiler_perf_log(void *data, const char *fmt, ...)
92 {
93 va_list args;
94 va_start(args, fmt);
95
96 if (unlikely(INTEL_DEBUG & DEBUG_PERF))
97 intel_logd_v(fmt, args);
98
99 va_end(args);
100 }
101
102 static uint64_t
103 anv_compute_heap_size(int fd, uint64_t gtt_size)
104 {
105 /* Query the total ram from the system */
106 struct sysinfo info;
107 sysinfo(&info);
108
109 uint64_t total_ram = (uint64_t)info.totalram * (uint64_t)info.mem_unit;
110
111 /* We don't want to burn too much ram with the GPU. If the user has 4GiB
112 * or less, we use at most half. If they have more than 4GiB, we use 3/4.
113 */
114 uint64_t available_ram;
115 if (total_ram <= 4ull * 1024ull * 1024ull * 1024ull)
116 available_ram = total_ram / 2;
117 else
118 available_ram = total_ram * 3 / 4;
119
120 /* We also want to leave some padding for things we allocate in the driver,
121 * so don't go over 3/4 of the GTT either.
122 */
123 uint64_t available_gtt = gtt_size * 3 / 4;
124
125 return MIN2(available_ram, available_gtt);
126 }
127
128 static VkResult
129 anv_physical_device_init_heaps(struct anv_physical_device *device, int fd)
130 {
131 if (anv_gem_get_context_param(fd, 0, I915_CONTEXT_PARAM_GTT_SIZE,
132 &device->gtt_size) == -1) {
133 /* If, for whatever reason, we can't actually get the GTT size from the
134 * kernel (too old?) fall back to the aperture size.
135 */
136 anv_perf_warn(NULL, NULL,
137 "Failed to get I915_CONTEXT_PARAM_GTT_SIZE: %m");
138
139 if (anv_gem_get_aperture(fd, &device->gtt_size) == -1) {
140 return vk_errorfi(device->instance, NULL,
141 VK_ERROR_INITIALIZATION_FAILED,
142 "failed to get aperture size: %m");
143 }
144 }
145
146 /* We only allow 48-bit addresses with softpin because knowing the actual
147 * address is required for the vertex cache flush workaround.
148 */
149 device->supports_48bit_addresses = (device->info.gen >= 8) &&
150 device->has_softpin &&
151 device->gtt_size > (4ULL << 30 /* GiB */);
152
153 uint64_t heap_size = anv_compute_heap_size(fd, device->gtt_size);
154
155 if (heap_size > (2ull << 30) && !device->supports_48bit_addresses) {
156 /* When running with an overridden PCI ID, we may get a GTT size from
157 * the kernel that is greater than 2 GiB but the execbuf check for 48bit
158 * address support can still fail. Just clamp the address space size to
159 * 2 GiB if we don't have 48-bit support.
160 */
161 intel_logw("%s:%d: The kernel reported a GTT size larger than 2 GiB but "
162 "not support for 48-bit addresses",
163 __FILE__, __LINE__);
164 heap_size = 2ull << 30;
165 }
166
167 device->memory.heap_count = 1;
168 device->memory.heaps[0] = (struct anv_memory_heap) {
169 .size = heap_size,
170 .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
171 };
172
173 uint32_t type_count = 0;
174 for (uint32_t heap = 0; heap < device->memory.heap_count; heap++) {
175 if (device->info.has_llc) {
176 /* Big core GPUs share LLC with the CPU and thus one memory type can be
177 * both cached and coherent at the same time.
178 */
179 device->memory.types[type_count++] = (struct anv_memory_type) {
180 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
181 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
182 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
183 VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
184 .heapIndex = heap,
185 };
186 } else {
187 /* The spec requires that we expose a host-visible, coherent memory
188 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
189 * to give the application a choice between cached, but not coherent and
190 * coherent but uncached (WC though).
191 */
192 device->memory.types[type_count++] = (struct anv_memory_type) {
193 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
194 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
195 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
196 .heapIndex = heap,
197 };
198 device->memory.types[type_count++] = (struct anv_memory_type) {
199 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
200 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
201 VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
202 .heapIndex = heap,
203 };
204 }
205 }
206 device->memory.type_count = type_count;
207
208 return VK_SUCCESS;
209 }
210
211 static VkResult
212 anv_physical_device_init_uuids(struct anv_physical_device *device)
213 {
214 const struct build_id_note *note =
215 build_id_find_nhdr_for_addr(anv_physical_device_init_uuids);
216 if (!note) {
217 return vk_errorfi(device->instance, NULL,
218 VK_ERROR_INITIALIZATION_FAILED,
219 "Failed to find build-id");
220 }
221
222 unsigned build_id_len = build_id_length(note);
223 if (build_id_len < 20) {
224 return vk_errorfi(device->instance, NULL,
225 VK_ERROR_INITIALIZATION_FAILED,
226 "build-id too short. It needs to be a SHA");
227 }
228
229 memcpy(device->driver_build_sha1, build_id_data(note), 20);
230
231 struct mesa_sha1 sha1_ctx;
232 uint8_t sha1[20];
233 STATIC_ASSERT(VK_UUID_SIZE <= sizeof(sha1));
234
235 /* The pipeline cache UUID is used for determining when a pipeline cache is
236 * invalid. It needs both a driver build and the PCI ID of the device.
237 */
238 _mesa_sha1_init(&sha1_ctx);
239 _mesa_sha1_update(&sha1_ctx, build_id_data(note), build_id_len);
240 _mesa_sha1_update(&sha1_ctx, &device->info.chipset_id,
241 sizeof(device->info.chipset_id));
242 _mesa_sha1_update(&sha1_ctx, &device->always_use_bindless,
243 sizeof(device->always_use_bindless));
244 _mesa_sha1_update(&sha1_ctx, &device->has_a64_buffer_access,
245 sizeof(device->has_a64_buffer_access));
246 _mesa_sha1_update(&sha1_ctx, &device->has_bindless_images,
247 sizeof(device->has_bindless_images));
248 _mesa_sha1_update(&sha1_ctx, &device->has_bindless_samplers,
249 sizeof(device->has_bindless_samplers));
250 _mesa_sha1_final(&sha1_ctx, sha1);
251 memcpy(device->pipeline_cache_uuid, sha1, VK_UUID_SIZE);
252
253 /* The driver UUID is used for determining sharability of images and memory
254 * between two Vulkan instances in separate processes. People who want to
255 * share memory need to also check the device UUID (below) so all this
256 * needs to be is the build-id.
257 */
258 memcpy(device->driver_uuid, build_id_data(note), VK_UUID_SIZE);
259
260 /* The device UUID uniquely identifies the given device within the machine.
261 * Since we never have more than one device, this doesn't need to be a real
262 * UUID. However, on the off-chance that someone tries to use this to
263 * cache pre-tiled images or something of the like, we use the PCI ID and
264 * some bits of ISL info to ensure that this is safe.
265 */
266 _mesa_sha1_init(&sha1_ctx);
267 _mesa_sha1_update(&sha1_ctx, &device->info.chipset_id,
268 sizeof(device->info.chipset_id));
269 _mesa_sha1_update(&sha1_ctx, &device->isl_dev.has_bit6_swizzling,
270 sizeof(device->isl_dev.has_bit6_swizzling));
271 _mesa_sha1_final(&sha1_ctx, sha1);
272 memcpy(device->device_uuid, sha1, VK_UUID_SIZE);
273
274 return VK_SUCCESS;
275 }
276
277 static void
278 anv_physical_device_init_disk_cache(struct anv_physical_device *device)
279 {
280 #ifdef ENABLE_SHADER_CACHE
281 char renderer[10];
282 ASSERTED int len = snprintf(renderer, sizeof(renderer), "anv_%04x",
283 device->info.chipset_id);
284 assert(len == sizeof(renderer) - 2);
285
286 char timestamp[41];
287 _mesa_sha1_format(timestamp, device->driver_build_sha1);
288
289 const uint64_t driver_flags =
290 brw_get_compiler_config_value(device->compiler);
291 device->disk_cache = disk_cache_create(renderer, timestamp, driver_flags);
292 #else
293 device->disk_cache = NULL;
294 #endif
295 }
296
297 static void
298 anv_physical_device_free_disk_cache(struct anv_physical_device *device)
299 {
300 #ifdef ENABLE_SHADER_CACHE
301 if (device->disk_cache)
302 disk_cache_destroy(device->disk_cache);
303 #else
304 assert(device->disk_cache == NULL);
305 #endif
306 }
307
308 static uint64_t
309 get_available_system_memory()
310 {
311 char *meminfo = os_read_file("/proc/meminfo", NULL);
312 if (!meminfo)
313 return 0;
314
315 char *str = strstr(meminfo, "MemAvailable:");
316 if (!str) {
317 free(meminfo);
318 return 0;
319 }
320
321 uint64_t kb_mem_available;
322 if (sscanf(str, "MemAvailable: %" PRIx64, &kb_mem_available) == 1) {
323 free(meminfo);
324 return kb_mem_available << 10;
325 }
326
327 free(meminfo);
328 return 0;
329 }
330
331 static VkResult
332 anv_physical_device_try_create(struct anv_instance *instance,
333 drmDevicePtr drm_device,
334 struct anv_physical_device **device_out)
335 {
336 const char *primary_path = drm_device->nodes[DRM_NODE_PRIMARY];
337 const char *path = drm_device->nodes[DRM_NODE_RENDER];
338 VkResult result;
339 int fd;
340 int master_fd = -1;
341
342 brw_process_intel_debug_variable();
343
344 fd = open(path, O_RDWR | O_CLOEXEC);
345 if (fd < 0)
346 return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
347
348 struct gen_device_info devinfo;
349 if (!gen_get_device_info_from_fd(fd, &devinfo)) {
350 result = vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
351 goto fail_fd;
352 }
353
354 const char *device_name = gen_get_device_name(devinfo.chipset_id);
355
356 if (devinfo.is_haswell) {
357 intel_logw("Haswell Vulkan support is incomplete");
358 } else if (devinfo.gen == 7 && !devinfo.is_baytrail) {
359 intel_logw("Ivy Bridge Vulkan support is incomplete");
360 } else if (devinfo.gen == 7 && devinfo.is_baytrail) {
361 intel_logw("Bay Trail Vulkan support is incomplete");
362 } else if (devinfo.gen >= 8 && devinfo.gen <= 11) {
363 /* Gen8-11 fully supported */
364 } else if (devinfo.gen == 12) {
365 intel_logw("Vulkan is not yet fully supported on gen12");
366 } else {
367 result = vk_errorfi(instance, NULL, VK_ERROR_INCOMPATIBLE_DRIVER,
368 "Vulkan not yet supported on %s", device_name);
369 goto fail_fd;
370 }
371
372 struct anv_physical_device *device =
373 vk_alloc(&instance->alloc, sizeof(*device), 8,
374 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
375 if (device == NULL) {
376 result = vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
377 goto fail_fd;
378 }
379
380 device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
381 device->instance = instance;
382
383 assert(strlen(path) < ARRAY_SIZE(device->path));
384 snprintf(device->path, ARRAY_SIZE(device->path), "%s", path);
385
386 device->info = devinfo;
387 device->name = device_name;
388
389 device->no_hw = device->info.no_hw;
390 if (getenv("INTEL_NO_HW") != NULL)
391 device->no_hw = true;
392
393 device->pci_info.domain = drm_device->businfo.pci->domain;
394 device->pci_info.bus = drm_device->businfo.pci->bus;
395 device->pci_info.device = drm_device->businfo.pci->dev;
396 device->pci_info.function = drm_device->businfo.pci->func;
397
398 device->cmd_parser_version = -1;
399 if (device->info.gen == 7) {
400 device->cmd_parser_version =
401 anv_gem_get_param(fd, I915_PARAM_CMD_PARSER_VERSION);
402 if (device->cmd_parser_version == -1) {
403 result = vk_errorfi(device->instance, NULL,
404 VK_ERROR_INITIALIZATION_FAILED,
405 "failed to get command parser version");
406 goto fail_alloc;
407 }
408 }
409
410 if (!anv_gem_get_param(fd, I915_PARAM_HAS_WAIT_TIMEOUT)) {
411 result = vk_errorfi(device->instance, NULL,
412 VK_ERROR_INITIALIZATION_FAILED,
413 "kernel missing gem wait");
414 goto fail_alloc;
415 }
416
417 if (!anv_gem_get_param(fd, I915_PARAM_HAS_EXECBUF2)) {
418 result = vk_errorfi(device->instance, NULL,
419 VK_ERROR_INITIALIZATION_FAILED,
420 "kernel missing execbuf2");
421 goto fail_alloc;
422 }
423
424 if (!device->info.has_llc &&
425 anv_gem_get_param(fd, I915_PARAM_MMAP_VERSION) < 1) {
426 result = vk_errorfi(device->instance, NULL,
427 VK_ERROR_INITIALIZATION_FAILED,
428 "kernel missing wc mmap");
429 goto fail_alloc;
430 }
431
432 device->has_softpin = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_SOFTPIN);
433 device->has_exec_async = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_ASYNC);
434 device->has_exec_capture = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_CAPTURE);
435 device->has_exec_fence = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_FENCE);
436 device->has_syncobj = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_FENCE_ARRAY);
437 device->has_syncobj_wait = device->has_syncobj &&
438 anv_gem_supports_syncobj_wait(fd);
439 device->has_context_priority = anv_gem_has_context_priority(fd);
440
441 result = anv_physical_device_init_heaps(device, fd);
442 if (result != VK_SUCCESS)
443 goto fail_alloc;
444
445 device->use_softpin = device->has_softpin &&
446 device->supports_48bit_addresses;
447
448 device->has_context_isolation =
449 anv_gem_get_param(fd, I915_PARAM_HAS_CONTEXT_ISOLATION);
450
451 device->always_use_bindless =
452 env_var_as_boolean("ANV_ALWAYS_BINDLESS", false);
453
454 /* We first got the A64 messages on broadwell and we can only use them if
455 * we can pass addresses directly into the shader which requires softpin.
456 */
457 device->has_a64_buffer_access = device->info.gen >= 8 &&
458 device->use_softpin;
459
460 /* We first get bindless image access on Skylake and we can only really do
461 * it if we don't have any relocations so we need softpin.
462 */
463 device->has_bindless_images = device->info.gen >= 9 &&
464 device->use_softpin;
465
466 /* We've had bindless samplers since Ivy Bridge (forever in Vulkan terms)
467 * because it's just a matter of setting the sampler address in the sample
468 * message header. However, we've not bothered to wire it up for vec4 so
469 * we leave it disabled on gen7.
470 */
471 device->has_bindless_samplers = device->info.gen >= 8;
472
473 device->has_implicit_ccs = device->info.has_aux_map;
474
475 device->has_mem_available = get_available_system_memory() != 0;
476
477 device->always_flush_cache =
478 driQueryOptionb(&instance->dri_options, "always_flush_cache");
479
480 device->has_mmap_offset =
481 anv_gem_get_param(fd, I915_PARAM_MMAP_GTT_VERSION) >= 4;
482
483 /* GENs prior to 8 do not support EU/Subslice info */
484 if (device->info.gen >= 8) {
485 device->subslice_total = anv_gem_get_param(fd, I915_PARAM_SUBSLICE_TOTAL);
486 device->eu_total = anv_gem_get_param(fd, I915_PARAM_EU_TOTAL);
487
488 /* Without this information, we cannot get the right Braswell
489 * brandstrings, and we have to use conservative numbers for GPGPU on
490 * many platforms, but otherwise, things will just work.
491 */
492 if (device->subslice_total < 1 || device->eu_total < 1) {
493 intel_logw("Kernel 4.1 required to properly query GPU properties");
494 }
495 } else if (device->info.gen == 7) {
496 device->subslice_total = 1 << (device->info.gt - 1);
497 }
498
499 if (device->info.is_cherryview &&
500 device->subslice_total > 0 && device->eu_total > 0) {
501 /* Logical CS threads = EUs per subslice * num threads per EU */
502 uint32_t max_cs_threads =
503 device->eu_total / device->subslice_total * device->info.num_thread_per_eu;
504
505 /* Fuse configurations may give more threads than expected, never less. */
506 if (max_cs_threads > device->info.max_cs_threads)
507 device->info.max_cs_threads = max_cs_threads;
508 }
509
510 device->compiler = brw_compiler_create(NULL, &device->info);
511 if (device->compiler == NULL) {
512 result = vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
513 goto fail_alloc;
514 }
515 device->compiler->shader_debug_log = compiler_debug_log;
516 device->compiler->shader_perf_log = compiler_perf_log;
517 device->compiler->supports_pull_constants = false;
518 device->compiler->constant_buffer_0_is_relative =
519 device->info.gen < 8 || !device->has_context_isolation;
520 device->compiler->supports_shader_constants = true;
521 device->compiler->compact_params = false;
522
523 /* Broadwell PRM says:
524 *
525 * "Before Gen8, there was a historical configuration control field to
526 * swizzle address bit[6] for in X/Y tiling modes. This was set in three
527 * different places: TILECTL[1:0], ARB_MODE[5:4], and
528 * DISP_ARB_CTL[14:13].
529 *
530 * For Gen8 and subsequent generations, the swizzle fields are all
531 * reserved, and the CPU's memory controller performs all address
532 * swizzling modifications."
533 */
534 bool swizzled =
535 device->info.gen < 8 && anv_gem_get_bit6_swizzle(fd, I915_TILING_X);
536
537 isl_device_init(&device->isl_dev, &device->info, swizzled);
538
539 result = anv_physical_device_init_uuids(device);
540 if (result != VK_SUCCESS)
541 goto fail_compiler;
542
543 anv_physical_device_init_disk_cache(device);
544
545 if (instance->enabled_extensions.KHR_display) {
546 master_fd = open(primary_path, O_RDWR | O_CLOEXEC);
547 if (master_fd >= 0) {
548 /* prod the device with a GETPARAM call which will fail if
549 * we don't have permission to even render on this device
550 */
551 if (anv_gem_get_param(master_fd, I915_PARAM_CHIPSET_ID) == 0) {
552 close(master_fd);
553 master_fd = -1;
554 }
555 }
556 }
557 device->master_fd = master_fd;
558
559 result = anv_init_wsi(device);
560 if (result != VK_SUCCESS)
561 goto fail_disk_cache;
562
563 device->perf = anv_get_perf(&device->info, fd);
564
565 anv_physical_device_get_supported_extensions(device,
566 &device->supported_extensions);
567
568
569 device->local_fd = fd;
570
571 *device_out = device;
572
573 return VK_SUCCESS;
574
575 fail_disk_cache:
576 anv_physical_device_free_disk_cache(device);
577 fail_compiler:
578 ralloc_free(device->compiler);
579 fail_alloc:
580 vk_free(&instance->alloc, device);
581 fail_fd:
582 close(fd);
583 if (master_fd != -1)
584 close(master_fd);
585 return result;
586 }
587
588 static void
589 anv_physical_device_destroy(struct anv_physical_device *device)
590 {
591 anv_finish_wsi(device);
592 anv_physical_device_free_disk_cache(device);
593 ralloc_free(device->compiler);
594 ralloc_free(device->perf);
595 close(device->local_fd);
596 if (device->master_fd >= 0)
597 close(device->master_fd);
598 vk_free(&device->instance->alloc, device);
599 }
600
601 static void *
602 default_alloc_func(void *pUserData, size_t size, size_t align,
603 VkSystemAllocationScope allocationScope)
604 {
605 return malloc(size);
606 }
607
608 static void *
609 default_realloc_func(void *pUserData, void *pOriginal, size_t size,
610 size_t align, VkSystemAllocationScope allocationScope)
611 {
612 return realloc(pOriginal, size);
613 }
614
615 static void
616 default_free_func(void *pUserData, void *pMemory)
617 {
618 free(pMemory);
619 }
620
621 static const VkAllocationCallbacks default_alloc = {
622 .pUserData = NULL,
623 .pfnAllocation = default_alloc_func,
624 .pfnReallocation = default_realloc_func,
625 .pfnFree = default_free_func,
626 };
627
628 VkResult anv_EnumerateInstanceExtensionProperties(
629 const char* pLayerName,
630 uint32_t* pPropertyCount,
631 VkExtensionProperties* pProperties)
632 {
633 VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
634
635 for (int i = 0; i < ANV_INSTANCE_EXTENSION_COUNT; i++) {
636 if (anv_instance_extensions_supported.extensions[i]) {
637 vk_outarray_append(&out, prop) {
638 *prop = anv_instance_extensions[i];
639 }
640 }
641 }
642
643 return vk_outarray_status(&out);
644 }
645
646 VkResult anv_CreateInstance(
647 const VkInstanceCreateInfo* pCreateInfo,
648 const VkAllocationCallbacks* pAllocator,
649 VkInstance* pInstance)
650 {
651 struct anv_instance *instance;
652 VkResult result;
653
654 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO);
655
656 struct anv_instance_extension_table enabled_extensions = {};
657 for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
658 int idx;
659 for (idx = 0; idx < ANV_INSTANCE_EXTENSION_COUNT; idx++) {
660 if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
661 anv_instance_extensions[idx].extensionName) == 0)
662 break;
663 }
664
665 if (idx >= ANV_INSTANCE_EXTENSION_COUNT)
666 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
667
668 if (!anv_instance_extensions_supported.extensions[idx])
669 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
670
671 enabled_extensions.extensions[idx] = true;
672 }
673
674 instance = vk_alloc2(&default_alloc, pAllocator, sizeof(*instance), 8,
675 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
676 if (!instance)
677 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
678
679 instance->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
680
681 if (pAllocator)
682 instance->alloc = *pAllocator;
683 else
684 instance->alloc = default_alloc;
685
686 instance->app_info = (struct anv_app_info) { .api_version = 0 };
687 if (pCreateInfo->pApplicationInfo) {
688 const VkApplicationInfo *app = pCreateInfo->pApplicationInfo;
689
690 instance->app_info.app_name =
691 vk_strdup(&instance->alloc, app->pApplicationName,
692 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
693 instance->app_info.app_version = app->applicationVersion;
694
695 instance->app_info.engine_name =
696 vk_strdup(&instance->alloc, app->pEngineName,
697 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
698 instance->app_info.engine_version = app->engineVersion;
699
700 instance->app_info.api_version = app->apiVersion;
701 }
702
703 if (instance->app_info.api_version == 0)
704 instance->app_info.api_version = VK_API_VERSION_1_0;
705
706 instance->enabled_extensions = enabled_extensions;
707
708 for (unsigned i = 0; i < ARRAY_SIZE(instance->dispatch.entrypoints); i++) {
709 /* Vulkan requires that entrypoints for extensions which have not been
710 * enabled must not be advertised.
711 */
712 if (!anv_instance_entrypoint_is_enabled(i, instance->app_info.api_version,
713 &instance->enabled_extensions)) {
714 instance->dispatch.entrypoints[i] = NULL;
715 } else {
716 instance->dispatch.entrypoints[i] =
717 anv_instance_dispatch_table.entrypoints[i];
718 }
719 }
720
721 for (unsigned i = 0; i < ARRAY_SIZE(instance->physical_device_dispatch.entrypoints); i++) {
722 /* Vulkan requires that entrypoints for extensions which have not been
723 * enabled must not be advertised.
724 */
725 if (!anv_physical_device_entrypoint_is_enabled(i, instance->app_info.api_version,
726 &instance->enabled_extensions)) {
727 instance->physical_device_dispatch.entrypoints[i] = NULL;
728 } else {
729 instance->physical_device_dispatch.entrypoints[i] =
730 anv_physical_device_dispatch_table.entrypoints[i];
731 }
732 }
733
734 for (unsigned i = 0; i < ARRAY_SIZE(instance->device_dispatch.entrypoints); i++) {
735 /* Vulkan requires that entrypoints for extensions which have not been
736 * enabled must not be advertised.
737 */
738 if (!anv_device_entrypoint_is_enabled(i, instance->app_info.api_version,
739 &instance->enabled_extensions, NULL)) {
740 instance->device_dispatch.entrypoints[i] = NULL;
741 } else {
742 instance->device_dispatch.entrypoints[i] =
743 anv_device_dispatch_table.entrypoints[i];
744 }
745 }
746
747 instance->physical_devices_enumerated = false;
748 list_inithead(&instance->physical_devices);
749
750 result = vk_debug_report_instance_init(&instance->debug_report_callbacks);
751 if (result != VK_SUCCESS) {
752 vk_free2(&default_alloc, pAllocator, instance);
753 return vk_error(result);
754 }
755
756 instance->pipeline_cache_enabled =
757 env_var_as_boolean("ANV_ENABLE_PIPELINE_CACHE", true);
758
759 glsl_type_singleton_init_or_ref();
760
761 VG(VALGRIND_CREATE_MEMPOOL(instance, 0, false));
762
763 driParseOptionInfo(&instance->available_dri_options, anv_dri_options_xml);
764 driParseConfigFiles(&instance->dri_options, &instance->available_dri_options,
765 0, "anv", NULL,
766 instance->app_info.engine_name,
767 instance->app_info.engine_version);
768
769 *pInstance = anv_instance_to_handle(instance);
770
771 return VK_SUCCESS;
772 }
773
774 void anv_DestroyInstance(
775 VkInstance _instance,
776 const VkAllocationCallbacks* pAllocator)
777 {
778 ANV_FROM_HANDLE(anv_instance, instance, _instance);
779
780 if (!instance)
781 return;
782
783 list_for_each_entry_safe(struct anv_physical_device, pdevice,
784 &instance->physical_devices, link)
785 anv_physical_device_destroy(pdevice);
786
787 vk_free(&instance->alloc, (char *)instance->app_info.app_name);
788 vk_free(&instance->alloc, (char *)instance->app_info.engine_name);
789
790 VG(VALGRIND_DESTROY_MEMPOOL(instance));
791
792 vk_debug_report_instance_destroy(&instance->debug_report_callbacks);
793
794 glsl_type_singleton_decref();
795
796 driDestroyOptionCache(&instance->dri_options);
797 driDestroyOptionInfo(&instance->available_dri_options);
798
799 vk_free(&instance->alloc, instance);
800 }
801
802 static VkResult
803 anv_enumerate_physical_devices(struct anv_instance *instance)
804 {
805 if (instance->physical_devices_enumerated)
806 return VK_SUCCESS;
807
808 instance->physical_devices_enumerated = true;
809
810 /* TODO: Check for more devices ? */
811 drmDevicePtr devices[8];
812 int max_devices;
813
814 max_devices = drmGetDevices2(0, devices, ARRAY_SIZE(devices));
815 if (max_devices < 1)
816 return VK_SUCCESS;
817
818 VkResult result = VK_SUCCESS;
819 for (unsigned i = 0; i < (unsigned)max_devices; i++) {
820 if (devices[i]->available_nodes & 1 << DRM_NODE_RENDER &&
821 devices[i]->bustype == DRM_BUS_PCI &&
822 devices[i]->deviceinfo.pci->vendor_id == 0x8086) {
823
824 struct anv_physical_device *pdevice;
825 result = anv_physical_device_try_create(instance, devices[i],
826 &pdevice);
827 /* Incompatible DRM device, skip. */
828 if (result == VK_ERROR_INCOMPATIBLE_DRIVER) {
829 result = VK_SUCCESS;
830 continue;
831 }
832
833 /* Error creating the physical device, report the error. */
834 if (result != VK_SUCCESS)
835 break;
836
837 list_addtail(&pdevice->link, &instance->physical_devices);
838 }
839 }
840 drmFreeDevices(devices, max_devices);
841
842 /* If we successfully enumerated any devices, call it success */
843 return result;
844 }
845
846 VkResult anv_EnumeratePhysicalDevices(
847 VkInstance _instance,
848 uint32_t* pPhysicalDeviceCount,
849 VkPhysicalDevice* pPhysicalDevices)
850 {
851 ANV_FROM_HANDLE(anv_instance, instance, _instance);
852 VK_OUTARRAY_MAKE(out, pPhysicalDevices, pPhysicalDeviceCount);
853
854 VkResult result = anv_enumerate_physical_devices(instance);
855 if (result != VK_SUCCESS)
856 return result;
857
858 list_for_each_entry(struct anv_physical_device, pdevice,
859 &instance->physical_devices, link) {
860 vk_outarray_append(&out, i) {
861 *i = anv_physical_device_to_handle(pdevice);
862 }
863 }
864
865 return vk_outarray_status(&out);
866 }
867
868 VkResult anv_EnumeratePhysicalDeviceGroups(
869 VkInstance _instance,
870 uint32_t* pPhysicalDeviceGroupCount,
871 VkPhysicalDeviceGroupProperties* pPhysicalDeviceGroupProperties)
872 {
873 ANV_FROM_HANDLE(anv_instance, instance, _instance);
874 VK_OUTARRAY_MAKE(out, pPhysicalDeviceGroupProperties,
875 pPhysicalDeviceGroupCount);
876
877 VkResult result = anv_enumerate_physical_devices(instance);
878 if (result != VK_SUCCESS)
879 return result;
880
881 list_for_each_entry(struct anv_physical_device, pdevice,
882 &instance->physical_devices, link) {
883 vk_outarray_append(&out, p) {
884 p->physicalDeviceCount = 1;
885 memset(p->physicalDevices, 0, sizeof(p->physicalDevices));
886 p->physicalDevices[0] = anv_physical_device_to_handle(pdevice);
887 p->subsetAllocation = false;
888
889 vk_foreach_struct(ext, p->pNext)
890 anv_debug_ignored_stype(ext->sType);
891 }
892 }
893
894 return vk_outarray_status(&out);
895 }
896
897 void anv_GetPhysicalDeviceFeatures(
898 VkPhysicalDevice physicalDevice,
899 VkPhysicalDeviceFeatures* pFeatures)
900 {
901 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
902
903 *pFeatures = (VkPhysicalDeviceFeatures) {
904 .robustBufferAccess = true,
905 .fullDrawIndexUint32 = true,
906 .imageCubeArray = true,
907 .independentBlend = true,
908 .geometryShader = true,
909 .tessellationShader = true,
910 .sampleRateShading = true,
911 .dualSrcBlend = true,
912 .logicOp = true,
913 .multiDrawIndirect = true,
914 .drawIndirectFirstInstance = true,
915 .depthClamp = true,
916 .depthBiasClamp = true,
917 .fillModeNonSolid = true,
918 .depthBounds = pdevice->info.gen >= 12,
919 .wideLines = true,
920 .largePoints = true,
921 .alphaToOne = true,
922 .multiViewport = true,
923 .samplerAnisotropy = true,
924 .textureCompressionETC2 = pdevice->info.gen >= 8 ||
925 pdevice->info.is_baytrail,
926 .textureCompressionASTC_LDR = pdevice->info.gen >= 9, /* FINISHME CHV */
927 .textureCompressionBC = true,
928 .occlusionQueryPrecise = true,
929 .pipelineStatisticsQuery = true,
930 .fragmentStoresAndAtomics = true,
931 .shaderTessellationAndGeometryPointSize = true,
932 .shaderImageGatherExtended = true,
933 .shaderStorageImageExtendedFormats = true,
934 .shaderStorageImageMultisample = false,
935 .shaderStorageImageReadWithoutFormat = false,
936 .shaderStorageImageWriteWithoutFormat = true,
937 .shaderUniformBufferArrayDynamicIndexing = true,
938 .shaderSampledImageArrayDynamicIndexing = true,
939 .shaderStorageBufferArrayDynamicIndexing = true,
940 .shaderStorageImageArrayDynamicIndexing = true,
941 .shaderClipDistance = true,
942 .shaderCullDistance = true,
943 .shaderFloat64 = pdevice->info.gen >= 8 &&
944 pdevice->info.has_64bit_float,
945 .shaderInt64 = pdevice->info.gen >= 8 &&
946 pdevice->info.has_64bit_int,
947 .shaderInt16 = pdevice->info.gen >= 8,
948 .shaderResourceMinLod = pdevice->info.gen >= 9,
949 .variableMultisampleRate = true,
950 .inheritedQueries = true,
951 };
952
953 /* We can't do image stores in vec4 shaders */
954 pFeatures->vertexPipelineStoresAndAtomics =
955 pdevice->compiler->scalar_stage[MESA_SHADER_VERTEX] &&
956 pdevice->compiler->scalar_stage[MESA_SHADER_GEOMETRY];
957
958 struct anv_app_info *app_info = &pdevice->instance->app_info;
959
960 /* The new DOOM and Wolfenstein games require depthBounds without
961 * checking for it. They seem to run fine without it so just claim it's
962 * there and accept the consequences.
963 */
964 if (app_info->engine_name && strcmp(app_info->engine_name, "idTech") == 0)
965 pFeatures->depthBounds = true;
966 }
967
968 static void
969 anv_get_physical_device_features_1_1(struct anv_physical_device *pdevice,
970 VkPhysicalDeviceVulkan11Features *f)
971 {
972 assert(f->sType == VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES);
973
974 f->storageBuffer16BitAccess = pdevice->info.gen >= 8;
975 f->uniformAndStorageBuffer16BitAccess = pdevice->info.gen >= 8;
976 f->storagePushConstant16 = pdevice->info.gen >= 8;
977 f->storageInputOutput16 = false;
978 f->multiview = true;
979 f->multiviewGeometryShader = true;
980 f->multiviewTessellationShader = true;
981 f->variablePointersStorageBuffer = true;
982 f->variablePointers = true;
983 f->protectedMemory = false;
984 f->samplerYcbcrConversion = true;
985 f->shaderDrawParameters = true;
986 }
987
988 static void
989 anv_get_physical_device_features_1_2(struct anv_physical_device *pdevice,
990 VkPhysicalDeviceVulkan12Features *f)
991 {
992 assert(f->sType == VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES);
993
994 f->samplerMirrorClampToEdge = true;
995 f->drawIndirectCount = true;
996 f->storageBuffer8BitAccess = pdevice->info.gen >= 8;
997 f->uniformAndStorageBuffer8BitAccess = pdevice->info.gen >= 8;
998 f->storagePushConstant8 = pdevice->info.gen >= 8;
999 f->shaderBufferInt64Atomics = pdevice->info.gen >= 9 &&
1000 pdevice->use_softpin;
1001 f->shaderSharedInt64Atomics = false;
1002 f->shaderFloat16 = pdevice->info.gen >= 8;
1003 f->shaderInt8 = pdevice->info.gen >= 8;
1004
1005 bool descIndexing = pdevice->has_a64_buffer_access &&
1006 pdevice->has_bindless_images;
1007 f->descriptorIndexing = descIndexing;
1008 f->shaderInputAttachmentArrayDynamicIndexing = false;
1009 f->shaderUniformTexelBufferArrayDynamicIndexing = descIndexing;
1010 f->shaderStorageTexelBufferArrayDynamicIndexing = descIndexing;
1011 f->shaderUniformBufferArrayNonUniformIndexing = false;
1012 f->shaderSampledImageArrayNonUniformIndexing = descIndexing;
1013 f->shaderStorageBufferArrayNonUniformIndexing = descIndexing;
1014 f->shaderStorageImageArrayNonUniformIndexing = descIndexing;
1015 f->shaderInputAttachmentArrayNonUniformIndexing = false;
1016 f->shaderUniformTexelBufferArrayNonUniformIndexing = descIndexing;
1017 f->shaderStorageTexelBufferArrayNonUniformIndexing = descIndexing;
1018 f->descriptorBindingUniformBufferUpdateAfterBind = false;
1019 f->descriptorBindingSampledImageUpdateAfterBind = descIndexing;
1020 f->descriptorBindingStorageImageUpdateAfterBind = descIndexing;
1021 f->descriptorBindingStorageBufferUpdateAfterBind = descIndexing;
1022 f->descriptorBindingUniformTexelBufferUpdateAfterBind = descIndexing;
1023 f->descriptorBindingStorageTexelBufferUpdateAfterBind = descIndexing;
1024 f->descriptorBindingUpdateUnusedWhilePending = descIndexing;
1025 f->descriptorBindingPartiallyBound = descIndexing;
1026 f->descriptorBindingVariableDescriptorCount = false;
1027 f->runtimeDescriptorArray = descIndexing;
1028
1029 f->samplerFilterMinmax = pdevice->info.gen >= 9;
1030 f->scalarBlockLayout = true;
1031 f->imagelessFramebuffer = true;
1032 f->uniformBufferStandardLayout = true;
1033 f->shaderSubgroupExtendedTypes = true;
1034 f->separateDepthStencilLayouts = true;
1035 f->hostQueryReset = true;
1036 f->timelineSemaphore = true;
1037 f->bufferDeviceAddress = pdevice->has_a64_buffer_access;
1038 f->bufferDeviceAddressCaptureReplay = pdevice->has_a64_buffer_access;
1039 f->bufferDeviceAddressMultiDevice = false;
1040 f->vulkanMemoryModel = true;
1041 f->vulkanMemoryModelDeviceScope = true;
1042 f->vulkanMemoryModelAvailabilityVisibilityChains = true;
1043 f->shaderOutputViewportIndex = true;
1044 f->shaderOutputLayer = true;
1045 f->subgroupBroadcastDynamicId = true;
1046 }
1047
1048 void anv_GetPhysicalDeviceFeatures2(
1049 VkPhysicalDevice physicalDevice,
1050 VkPhysicalDeviceFeatures2* pFeatures)
1051 {
1052 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
1053 anv_GetPhysicalDeviceFeatures(physicalDevice, &pFeatures->features);
1054
1055 VkPhysicalDeviceVulkan11Features core_1_1 = {
1056 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES,
1057 };
1058 anv_get_physical_device_features_1_1(pdevice, &core_1_1);
1059
1060 VkPhysicalDeviceVulkan12Features core_1_2 = {
1061 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES,
1062 };
1063 anv_get_physical_device_features_1_2(pdevice, &core_1_2);
1064
1065 #define CORE_FEATURE(major, minor, feature) \
1066 features->feature = core_##major##_##minor.feature
1067
1068
1069 vk_foreach_struct(ext, pFeatures->pNext) {
1070 switch (ext->sType) {
1071 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES_KHR: {
1072 VkPhysicalDevice8BitStorageFeaturesKHR *features =
1073 (VkPhysicalDevice8BitStorageFeaturesKHR *)ext;
1074 CORE_FEATURE(1, 2, storageBuffer8BitAccess);
1075 CORE_FEATURE(1, 2, uniformAndStorageBuffer8BitAccess);
1076 CORE_FEATURE(1, 2, storagePushConstant8);
1077 break;
1078 }
1079
1080 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES: {
1081 VkPhysicalDevice16BitStorageFeatures *features =
1082 (VkPhysicalDevice16BitStorageFeatures *)ext;
1083 CORE_FEATURE(1, 1, storageBuffer16BitAccess);
1084 CORE_FEATURE(1, 1, uniformAndStorageBuffer16BitAccess);
1085 CORE_FEATURE(1, 1, storagePushConstant16);
1086 CORE_FEATURE(1, 1, storageInputOutput16);
1087 break;
1088 }
1089
1090 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_EXT: {
1091 VkPhysicalDeviceBufferDeviceAddressFeaturesEXT *features = (void *)ext;
1092 features->bufferDeviceAddress = pdevice->has_a64_buffer_access;
1093 features->bufferDeviceAddressCaptureReplay = false;
1094 features->bufferDeviceAddressMultiDevice = false;
1095 break;
1096 }
1097
1098 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_KHR: {
1099 VkPhysicalDeviceBufferDeviceAddressFeaturesKHR *features = (void *)ext;
1100 CORE_FEATURE(1, 2, bufferDeviceAddress);
1101 CORE_FEATURE(1, 2, bufferDeviceAddressCaptureReplay);
1102 CORE_FEATURE(1, 2, bufferDeviceAddressMultiDevice);
1103 break;
1104 }
1105
1106 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COMPUTE_SHADER_DERIVATIVES_FEATURES_NV: {
1107 VkPhysicalDeviceComputeShaderDerivativesFeaturesNV *features =
1108 (VkPhysicalDeviceComputeShaderDerivativesFeaturesNV *)ext;
1109 features->computeDerivativeGroupQuads = true;
1110 features->computeDerivativeGroupLinear = true;
1111 break;
1112 }
1113
1114 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT: {
1115 VkPhysicalDeviceConditionalRenderingFeaturesEXT *features =
1116 (VkPhysicalDeviceConditionalRenderingFeaturesEXT*)ext;
1117 features->conditionalRendering = pdevice->info.gen >= 8 ||
1118 pdevice->info.is_haswell;
1119 features->inheritedConditionalRendering = pdevice->info.gen >= 8 ||
1120 pdevice->info.is_haswell;
1121 break;
1122 }
1123
1124 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_CLIP_ENABLE_FEATURES_EXT: {
1125 VkPhysicalDeviceDepthClipEnableFeaturesEXT *features =
1126 (VkPhysicalDeviceDepthClipEnableFeaturesEXT *)ext;
1127 features->depthClipEnable = true;
1128 break;
1129 }
1130
1131 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR: {
1132 VkPhysicalDeviceFloat16Int8FeaturesKHR *features = (void *)ext;
1133 CORE_FEATURE(1, 2, shaderFloat16);
1134 CORE_FEATURE(1, 2, shaderInt8);
1135 break;
1136 }
1137
1138 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FRAGMENT_SHADER_INTERLOCK_FEATURES_EXT: {
1139 VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT *features =
1140 (VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT *)ext;
1141 features->fragmentShaderSampleInterlock = pdevice->info.gen >= 9;
1142 features->fragmentShaderPixelInterlock = pdevice->info.gen >= 9;
1143 features->fragmentShaderShadingRateInterlock = false;
1144 break;
1145 }
1146
1147 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_QUERY_RESET_FEATURES_EXT: {
1148 VkPhysicalDeviceHostQueryResetFeaturesEXT *features =
1149 (VkPhysicalDeviceHostQueryResetFeaturesEXT *)ext;
1150 CORE_FEATURE(1, 2, hostQueryReset);
1151 break;
1152 }
1153
1154 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT: {
1155 VkPhysicalDeviceDescriptorIndexingFeaturesEXT *features =
1156 (VkPhysicalDeviceDescriptorIndexingFeaturesEXT *)ext;
1157 CORE_FEATURE(1, 2, shaderInputAttachmentArrayDynamicIndexing);
1158 CORE_FEATURE(1, 2, shaderUniformTexelBufferArrayDynamicIndexing);
1159 CORE_FEATURE(1, 2, shaderStorageTexelBufferArrayDynamicIndexing);
1160 CORE_FEATURE(1, 2, shaderUniformBufferArrayNonUniformIndexing);
1161 CORE_FEATURE(1, 2, shaderSampledImageArrayNonUniformIndexing);
1162 CORE_FEATURE(1, 2, shaderStorageBufferArrayNonUniformIndexing);
1163 CORE_FEATURE(1, 2, shaderStorageImageArrayNonUniformIndexing);
1164 CORE_FEATURE(1, 2, shaderInputAttachmentArrayNonUniformIndexing);
1165 CORE_FEATURE(1, 2, shaderUniformTexelBufferArrayNonUniformIndexing);
1166 CORE_FEATURE(1, 2, shaderStorageTexelBufferArrayNonUniformIndexing);
1167 CORE_FEATURE(1, 2, descriptorBindingUniformBufferUpdateAfterBind);
1168 CORE_FEATURE(1, 2, descriptorBindingSampledImageUpdateAfterBind);
1169 CORE_FEATURE(1, 2, descriptorBindingStorageImageUpdateAfterBind);
1170 CORE_FEATURE(1, 2, descriptorBindingStorageBufferUpdateAfterBind);
1171 CORE_FEATURE(1, 2, descriptorBindingUniformTexelBufferUpdateAfterBind);
1172 CORE_FEATURE(1, 2, descriptorBindingStorageTexelBufferUpdateAfterBind);
1173 CORE_FEATURE(1, 2, descriptorBindingUpdateUnusedWhilePending);
1174 CORE_FEATURE(1, 2, descriptorBindingPartiallyBound);
1175 CORE_FEATURE(1, 2, descriptorBindingVariableDescriptorCount);
1176 CORE_FEATURE(1, 2, runtimeDescriptorArray);
1177 break;
1178 }
1179
1180 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INDEX_TYPE_UINT8_FEATURES_EXT: {
1181 VkPhysicalDeviceIndexTypeUint8FeaturesEXT *features =
1182 (VkPhysicalDeviceIndexTypeUint8FeaturesEXT *)ext;
1183 features->indexTypeUint8 = true;
1184 break;
1185 }
1186
1187 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_FEATURES_EXT: {
1188 VkPhysicalDeviceInlineUniformBlockFeaturesEXT *features =
1189 (VkPhysicalDeviceInlineUniformBlockFeaturesEXT *)ext;
1190 features->inlineUniformBlock = true;
1191 features->descriptorBindingInlineUniformBlockUpdateAfterBind = true;
1192 break;
1193 }
1194
1195 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_FEATURES_EXT: {
1196 VkPhysicalDeviceLineRasterizationFeaturesEXT *features =
1197 (VkPhysicalDeviceLineRasterizationFeaturesEXT *)ext;
1198 features->rectangularLines = true;
1199 features->bresenhamLines = true;
1200 /* Support for Smooth lines with MSAA was removed on gen11. From the
1201 * BSpec section "Multisample ModesState" table for "AA Line Support
1202 * Requirements":
1203 *
1204 * GEN10:BUG:######## NUM_MULTISAMPLES == 1
1205 *
1206 * Fortunately, this isn't a case most people care about.
1207 */
1208 features->smoothLines = pdevice->info.gen < 10;
1209 features->stippledRectangularLines = false;
1210 features->stippledBresenhamLines = true;
1211 features->stippledSmoothLines = false;
1212 break;
1213 }
1214
1215 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES: {
1216 VkPhysicalDeviceMultiviewFeatures *features =
1217 (VkPhysicalDeviceMultiviewFeatures *)ext;
1218 CORE_FEATURE(1, 1, multiview);
1219 CORE_FEATURE(1, 1, multiviewGeometryShader);
1220 CORE_FEATURE(1, 1, multiviewTessellationShader);
1221 break;
1222 }
1223
1224 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGELESS_FRAMEBUFFER_FEATURES_KHR: {
1225 VkPhysicalDeviceImagelessFramebufferFeaturesKHR *features =
1226 (VkPhysicalDeviceImagelessFramebufferFeaturesKHR *)ext;
1227 CORE_FEATURE(1, 2, imagelessFramebuffer);
1228 break;
1229 }
1230
1231 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR: {
1232 VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR *features =
1233 (VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR *)ext;
1234 features->pipelineExecutableInfo = true;
1235 break;
1236 }
1237
1238 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES: {
1239 VkPhysicalDeviceProtectedMemoryFeatures *features = (void *)ext;
1240 CORE_FEATURE(1, 1, protectedMemory);
1241 break;
1242 }
1243
1244 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES: {
1245 VkPhysicalDeviceSamplerYcbcrConversionFeatures *features =
1246 (VkPhysicalDeviceSamplerYcbcrConversionFeatures *) ext;
1247 CORE_FEATURE(1, 1, samplerYcbcrConversion);
1248 break;
1249 }
1250
1251 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES_EXT: {
1252 VkPhysicalDeviceScalarBlockLayoutFeaturesEXT *features =
1253 (VkPhysicalDeviceScalarBlockLayoutFeaturesEXT *)ext;
1254 CORE_FEATURE(1, 2, scalarBlockLayout);
1255 break;
1256 }
1257
1258 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SEPARATE_DEPTH_STENCIL_LAYOUTS_FEATURES_KHR: {
1259 VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR *features =
1260 (VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR *)ext;
1261 CORE_FEATURE(1, 2, separateDepthStencilLayouts);
1262 break;
1263 }
1264
1265 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_INT64_FEATURES_KHR: {
1266 VkPhysicalDeviceShaderAtomicInt64FeaturesKHR *features = (void *)ext;
1267 CORE_FEATURE(1, 2, shaderBufferInt64Atomics);
1268 CORE_FEATURE(1, 2, shaderSharedInt64Atomics);
1269 break;
1270 }
1271
1272 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DEMOTE_TO_HELPER_INVOCATION_FEATURES_EXT: {
1273 VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT *features = (void *)ext;
1274 features->shaderDemoteToHelperInvocation = true;
1275 break;
1276 }
1277
1278 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CLOCK_FEATURES_KHR: {
1279 VkPhysicalDeviceShaderClockFeaturesKHR *features =
1280 (VkPhysicalDeviceShaderClockFeaturesKHR *)ext;
1281 features->shaderSubgroupClock = true;
1282 features->shaderDeviceClock = false;
1283 break;
1284 }
1285
1286 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETERS_FEATURES: {
1287 VkPhysicalDeviceShaderDrawParametersFeatures *features = (void *)ext;
1288 CORE_FEATURE(1, 1, shaderDrawParameters);
1289 break;
1290 }
1291
1292 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_SUBGROUP_EXTENDED_TYPES_FEATURES_KHR: {
1293 VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR *features =
1294 (VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR *)ext;
1295 CORE_FEATURE(1, 2, shaderSubgroupExtendedTypes);
1296 break;
1297 }
1298
1299 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_FEATURES_EXT: {
1300 VkPhysicalDeviceSubgroupSizeControlFeaturesEXT *features =
1301 (VkPhysicalDeviceSubgroupSizeControlFeaturesEXT *)ext;
1302 features->subgroupSizeControl = true;
1303 features->computeFullSubgroups = true;
1304 break;
1305 }
1306
1307 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_FEATURES_EXT: {
1308 VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT *features =
1309 (VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT *)ext;
1310 features->texelBufferAlignment = true;
1311 break;
1312 }
1313
1314 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_FEATURES_KHR: {
1315 VkPhysicalDeviceTimelineSemaphoreFeaturesKHR *features =
1316 (VkPhysicalDeviceTimelineSemaphoreFeaturesKHR *) ext;
1317 CORE_FEATURE(1, 2, timelineSemaphore);
1318 break;
1319 }
1320
1321 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES: {
1322 VkPhysicalDeviceVariablePointersFeatures *features = (void *)ext;
1323 CORE_FEATURE(1, 1, variablePointersStorageBuffer);
1324 CORE_FEATURE(1, 1, variablePointers);
1325 break;
1326 }
1327
1328 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT: {
1329 VkPhysicalDeviceTransformFeedbackFeaturesEXT *features =
1330 (VkPhysicalDeviceTransformFeedbackFeaturesEXT *)ext;
1331 features->transformFeedback = true;
1332 features->geometryStreams = true;
1333 break;
1334 }
1335
1336 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_UNIFORM_BUFFER_STANDARD_LAYOUT_FEATURES_KHR: {
1337 VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR *features =
1338 (VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR *)ext;
1339 CORE_FEATURE(1, 2, uniformBufferStandardLayout);
1340 break;
1341 }
1342
1343 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT: {
1344 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *features =
1345 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *)ext;
1346 features->vertexAttributeInstanceRateDivisor = true;
1347 features->vertexAttributeInstanceRateZeroDivisor = true;
1348 break;
1349 }
1350
1351 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES:
1352 anv_get_physical_device_features_1_1(pdevice, (void *)ext);
1353 break;
1354
1355 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES:
1356 anv_get_physical_device_features_1_2(pdevice, (void *)ext);
1357 break;
1358
1359 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_MEMORY_MODEL_FEATURES_KHR: {
1360 VkPhysicalDeviceVulkanMemoryModelFeaturesKHR *features = (void *)ext;
1361 CORE_FEATURE(1, 2, vulkanMemoryModel);
1362 CORE_FEATURE(1, 2, vulkanMemoryModelDeviceScope);
1363 CORE_FEATURE(1, 2, vulkanMemoryModelAvailabilityVisibilityChains);
1364 break;
1365 }
1366
1367 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_YCBCR_IMAGE_ARRAYS_FEATURES_EXT: {
1368 VkPhysicalDeviceYcbcrImageArraysFeaturesEXT *features =
1369 (VkPhysicalDeviceYcbcrImageArraysFeaturesEXT *)ext;
1370 features->ycbcrImageArrays = true;
1371 break;
1372 }
1373
1374 default:
1375 anv_debug_ignored_stype(ext->sType);
1376 break;
1377 }
1378 }
1379
1380 #undef CORE_FEATURE
1381 }
1382
1383 #define MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS 64
1384
1385 #define MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS 64
1386 #define MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS 256
1387
1388 void anv_GetPhysicalDeviceProperties(
1389 VkPhysicalDevice physicalDevice,
1390 VkPhysicalDeviceProperties* pProperties)
1391 {
1392 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
1393 const struct gen_device_info *devinfo = &pdevice->info;
1394
1395 /* See assertions made when programming the buffer surface state. */
1396 const uint32_t max_raw_buffer_sz = devinfo->gen >= 7 ?
1397 (1ul << 30) : (1ul << 27);
1398
1399 const uint32_t max_ssbos = pdevice->has_a64_buffer_access ? UINT16_MAX : 64;
1400 const uint32_t max_textures =
1401 pdevice->has_bindless_images ? UINT16_MAX : 128;
1402 const uint32_t max_samplers =
1403 pdevice->has_bindless_samplers ? UINT16_MAX :
1404 (devinfo->gen >= 8 || devinfo->is_haswell) ? 128 : 16;
1405 const uint32_t max_images =
1406 pdevice->has_bindless_images ? UINT16_MAX : MAX_IMAGES;
1407
1408 /* If we can use bindless for everything, claim a high per-stage limit,
1409 * otherwise use the binding table size, minus the slots reserved for
1410 * render targets and one slot for the descriptor buffer. */
1411 const uint32_t max_per_stage =
1412 pdevice->has_bindless_images && pdevice->has_a64_buffer_access
1413 ? UINT32_MAX : MAX_BINDING_TABLE_SIZE - MAX_RTS - 1;
1414
1415 const uint32_t max_workgroup_size = 32 * devinfo->max_cs_threads;
1416
1417 VkSampleCountFlags sample_counts =
1418 isl_device_get_sample_counts(&pdevice->isl_dev);
1419
1420
1421 VkPhysicalDeviceLimits limits = {
1422 .maxImageDimension1D = (1 << 14),
1423 .maxImageDimension2D = (1 << 14),
1424 .maxImageDimension3D = (1 << 11),
1425 .maxImageDimensionCube = (1 << 14),
1426 .maxImageArrayLayers = (1 << 11),
1427 .maxTexelBufferElements = 128 * 1024 * 1024,
1428 .maxUniformBufferRange = (1ul << 27),
1429 .maxStorageBufferRange = max_raw_buffer_sz,
1430 .maxPushConstantsSize = MAX_PUSH_CONSTANTS_SIZE,
1431 .maxMemoryAllocationCount = UINT32_MAX,
1432 .maxSamplerAllocationCount = 64 * 1024,
1433 .bufferImageGranularity = 64, /* A cache line */
1434 .sparseAddressSpaceSize = 0,
1435 .maxBoundDescriptorSets = MAX_SETS,
1436 .maxPerStageDescriptorSamplers = max_samplers,
1437 .maxPerStageDescriptorUniformBuffers = MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS,
1438 .maxPerStageDescriptorStorageBuffers = max_ssbos,
1439 .maxPerStageDescriptorSampledImages = max_textures,
1440 .maxPerStageDescriptorStorageImages = max_images,
1441 .maxPerStageDescriptorInputAttachments = MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS,
1442 .maxPerStageResources = max_per_stage,
1443 .maxDescriptorSetSamplers = 6 * max_samplers, /* number of stages * maxPerStageDescriptorSamplers */
1444 .maxDescriptorSetUniformBuffers = 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS, /* number of stages * maxPerStageDescriptorUniformBuffers */
1445 .maxDescriptorSetUniformBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2,
1446 .maxDescriptorSetStorageBuffers = 6 * max_ssbos, /* number of stages * maxPerStageDescriptorStorageBuffers */
1447 .maxDescriptorSetStorageBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2,
1448 .maxDescriptorSetSampledImages = 6 * max_textures, /* number of stages * maxPerStageDescriptorSampledImages */
1449 .maxDescriptorSetStorageImages = 6 * max_images, /* number of stages * maxPerStageDescriptorStorageImages */
1450 .maxDescriptorSetInputAttachments = MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS,
1451 .maxVertexInputAttributes = MAX_VBS,
1452 .maxVertexInputBindings = MAX_VBS,
1453 .maxVertexInputAttributeOffset = 2047,
1454 .maxVertexInputBindingStride = 2048,
1455 .maxVertexOutputComponents = 128,
1456 .maxTessellationGenerationLevel = 64,
1457 .maxTessellationPatchSize = 32,
1458 .maxTessellationControlPerVertexInputComponents = 128,
1459 .maxTessellationControlPerVertexOutputComponents = 128,
1460 .maxTessellationControlPerPatchOutputComponents = 128,
1461 .maxTessellationControlTotalOutputComponents = 2048,
1462 .maxTessellationEvaluationInputComponents = 128,
1463 .maxTessellationEvaluationOutputComponents = 128,
1464 .maxGeometryShaderInvocations = 32,
1465 .maxGeometryInputComponents = 64,
1466 .maxGeometryOutputComponents = 128,
1467 .maxGeometryOutputVertices = 256,
1468 .maxGeometryTotalOutputComponents = 1024,
1469 .maxFragmentInputComponents = 116, /* 128 components - (PSIZ, CLIP_DIST0, CLIP_DIST1) */
1470 .maxFragmentOutputAttachments = 8,
1471 .maxFragmentDualSrcAttachments = 1,
1472 .maxFragmentCombinedOutputResources = 8,
1473 .maxComputeSharedMemorySize = 64 * 1024,
1474 .maxComputeWorkGroupCount = { 65535, 65535, 65535 },
1475 .maxComputeWorkGroupInvocations = max_workgroup_size,
1476 .maxComputeWorkGroupSize = {
1477 max_workgroup_size,
1478 max_workgroup_size,
1479 max_workgroup_size,
1480 },
1481 .subPixelPrecisionBits = 8,
1482 .subTexelPrecisionBits = 8,
1483 .mipmapPrecisionBits = 8,
1484 .maxDrawIndexedIndexValue = UINT32_MAX,
1485 .maxDrawIndirectCount = UINT32_MAX,
1486 .maxSamplerLodBias = 16,
1487 .maxSamplerAnisotropy = 16,
1488 .maxViewports = MAX_VIEWPORTS,
1489 .maxViewportDimensions = { (1 << 14), (1 << 14) },
1490 .viewportBoundsRange = { INT16_MIN, INT16_MAX },
1491 .viewportSubPixelBits = 13, /* We take a float? */
1492 .minMemoryMapAlignment = 4096, /* A page */
1493 /* The dataport requires texel alignment so we need to assume a worst
1494 * case of R32G32B32A32 which is 16 bytes.
1495 */
1496 .minTexelBufferOffsetAlignment = 16,
1497 /* We need 16 for UBO block reads to work and 32 for push UBOs */
1498 .minUniformBufferOffsetAlignment = 32,
1499 .minStorageBufferOffsetAlignment = 4,
1500 .minTexelOffset = -8,
1501 .maxTexelOffset = 7,
1502 .minTexelGatherOffset = -32,
1503 .maxTexelGatherOffset = 31,
1504 .minInterpolationOffset = -0.5,
1505 .maxInterpolationOffset = 0.4375,
1506 .subPixelInterpolationOffsetBits = 4,
1507 .maxFramebufferWidth = (1 << 14),
1508 .maxFramebufferHeight = (1 << 14),
1509 .maxFramebufferLayers = (1 << 11),
1510 .framebufferColorSampleCounts = sample_counts,
1511 .framebufferDepthSampleCounts = sample_counts,
1512 .framebufferStencilSampleCounts = sample_counts,
1513 .framebufferNoAttachmentsSampleCounts = sample_counts,
1514 .maxColorAttachments = MAX_RTS,
1515 .sampledImageColorSampleCounts = sample_counts,
1516 .sampledImageIntegerSampleCounts = sample_counts,
1517 .sampledImageDepthSampleCounts = sample_counts,
1518 .sampledImageStencilSampleCounts = sample_counts,
1519 .storageImageSampleCounts = VK_SAMPLE_COUNT_1_BIT,
1520 .maxSampleMaskWords = 1,
1521 .timestampComputeAndGraphics = true,
1522 .timestampPeriod = 1000000000.0 / devinfo->timestamp_frequency,
1523 .maxClipDistances = 8,
1524 .maxCullDistances = 8,
1525 .maxCombinedClipAndCullDistances = 8,
1526 .discreteQueuePriorities = 2,
1527 .pointSizeRange = { 0.125, 255.875 },
1528 .lineWidthRange = {
1529 0.0,
1530 (devinfo->gen >= 9 || devinfo->is_cherryview) ?
1531 2047.9921875 : 7.9921875,
1532 },
1533 .pointSizeGranularity = (1.0 / 8.0),
1534 .lineWidthGranularity = (1.0 / 128.0),
1535 .strictLines = false,
1536 .standardSampleLocations = true,
1537 .optimalBufferCopyOffsetAlignment = 128,
1538 .optimalBufferCopyRowPitchAlignment = 128,
1539 .nonCoherentAtomSize = 64,
1540 };
1541
1542 *pProperties = (VkPhysicalDeviceProperties) {
1543 .apiVersion = anv_physical_device_api_version(pdevice),
1544 .driverVersion = vk_get_driver_version(),
1545 .vendorID = 0x8086,
1546 .deviceID = pdevice->info.chipset_id,
1547 .deviceType = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
1548 .limits = limits,
1549 .sparseProperties = {0}, /* Broadwell doesn't do sparse. */
1550 };
1551
1552 snprintf(pProperties->deviceName, sizeof(pProperties->deviceName),
1553 "%s", pdevice->name);
1554 memcpy(pProperties->pipelineCacheUUID,
1555 pdevice->pipeline_cache_uuid, VK_UUID_SIZE);
1556 }
1557
1558 static void
1559 anv_get_physical_device_properties_1_1(struct anv_physical_device *pdevice,
1560 VkPhysicalDeviceVulkan11Properties *p)
1561 {
1562 assert(p->sType == VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES);
1563
1564 memcpy(p->deviceUUID, pdevice->device_uuid, VK_UUID_SIZE);
1565 memcpy(p->driverUUID, pdevice->driver_uuid, VK_UUID_SIZE);
1566 memset(p->deviceLUID, 0, VK_LUID_SIZE);
1567 p->deviceNodeMask = 0;
1568 p->deviceLUIDValid = false;
1569
1570 p->subgroupSize = BRW_SUBGROUP_SIZE;
1571 VkShaderStageFlags scalar_stages = 0;
1572 for (unsigned stage = 0; stage < MESA_SHADER_STAGES; stage++) {
1573 if (pdevice->compiler->scalar_stage[stage])
1574 scalar_stages |= mesa_to_vk_shader_stage(stage);
1575 }
1576 p->subgroupSupportedStages = scalar_stages;
1577 p->subgroupSupportedOperations = VK_SUBGROUP_FEATURE_BASIC_BIT |
1578 VK_SUBGROUP_FEATURE_VOTE_BIT |
1579 VK_SUBGROUP_FEATURE_BALLOT_BIT |
1580 VK_SUBGROUP_FEATURE_SHUFFLE_BIT |
1581 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT |
1582 VK_SUBGROUP_FEATURE_QUAD_BIT;
1583 if (pdevice->info.gen >= 8) {
1584 /* TODO: There's no technical reason why these can't be made to
1585 * work on gen7 but they don't at the moment so it's best to leave
1586 * the feature disabled than enabled and broken.
1587 */
1588 p->subgroupSupportedOperations |= VK_SUBGROUP_FEATURE_ARITHMETIC_BIT |
1589 VK_SUBGROUP_FEATURE_CLUSTERED_BIT;
1590 }
1591 p->subgroupQuadOperationsInAllStages = pdevice->info.gen >= 8;
1592
1593 p->pointClippingBehavior = VK_POINT_CLIPPING_BEHAVIOR_USER_CLIP_PLANES_ONLY;
1594 p->maxMultiviewViewCount = 16;
1595 p->maxMultiviewInstanceIndex = UINT32_MAX / 16;
1596 p->protectedNoFault = false;
1597 /* This value doesn't matter for us today as our per-stage descriptors are
1598 * the real limit.
1599 */
1600 p->maxPerSetDescriptors = 1024;
1601 p->maxMemoryAllocationSize = MAX_MEMORY_ALLOCATION_SIZE;
1602 }
1603
1604 static void
1605 anv_get_physical_device_properties_1_2(struct anv_physical_device *pdevice,
1606 VkPhysicalDeviceVulkan12Properties *p)
1607 {
1608 assert(p->sType == VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES);
1609
1610 p->driverID = VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA_KHR;
1611 memset(p->driverName, 0, sizeof(p->driverName));
1612 snprintf(p->driverName, VK_MAX_DRIVER_NAME_SIZE_KHR,
1613 "Intel open-source Mesa driver");
1614 memset(p->driverInfo, 0, sizeof(p->driverInfo));
1615 snprintf(p->driverInfo, VK_MAX_DRIVER_INFO_SIZE_KHR,
1616 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1);
1617 p->conformanceVersion = (VkConformanceVersionKHR) {
1618 .major = 1,
1619 .minor = 2,
1620 .subminor = 0,
1621 .patch = 0,
1622 };
1623
1624 p->denormBehaviorIndependence =
1625 VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_ALL_KHR;
1626 p->roundingModeIndependence =
1627 VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_NONE_KHR;
1628
1629 /* Broadwell does not support HF denorms and there are restrictions
1630 * other gens. According to Kabylake's PRM:
1631 *
1632 * "math - Extended Math Function
1633 * [...]
1634 * Restriction : Half-float denorms are always retained."
1635 */
1636 p->shaderDenormFlushToZeroFloat16 = false;
1637 p->shaderDenormPreserveFloat16 = pdevice->info.gen > 8;
1638 p->shaderRoundingModeRTEFloat16 = true;
1639 p->shaderRoundingModeRTZFloat16 = true;
1640 p->shaderSignedZeroInfNanPreserveFloat16 = true;
1641
1642 p->shaderDenormFlushToZeroFloat32 = true;
1643 p->shaderDenormPreserveFloat32 = true;
1644 p->shaderRoundingModeRTEFloat32 = true;
1645 p->shaderRoundingModeRTZFloat32 = true;
1646 p->shaderSignedZeroInfNanPreserveFloat32 = true;
1647
1648 p->shaderDenormFlushToZeroFloat64 = true;
1649 p->shaderDenormPreserveFloat64 = true;
1650 p->shaderRoundingModeRTEFloat64 = true;
1651 p->shaderRoundingModeRTZFloat64 = true;
1652 p->shaderSignedZeroInfNanPreserveFloat64 = true;
1653
1654 /* It's a bit hard to exactly map our implementation to the limits
1655 * described here. The bindless surface handle in the extended
1656 * message descriptors is 20 bits and it's an index into the table of
1657 * RENDER_SURFACE_STATE structs that starts at bindless surface base
1658 * address. Given that most things consume two surface states per
1659 * view (general/sampled for textures and write-only/read-write for
1660 * images), we claim 2^19 things.
1661 *
1662 * For SSBOs, we just use A64 messages so there is no real limit
1663 * there beyond the limit on the total size of a descriptor set.
1664 */
1665 const unsigned max_bindless_views = 1 << 19;
1666 p->maxUpdateAfterBindDescriptorsInAllPools = max_bindless_views;
1667 p->shaderUniformBufferArrayNonUniformIndexingNative = false;
1668 p->shaderSampledImageArrayNonUniformIndexingNative = false;
1669 p->shaderStorageBufferArrayNonUniformIndexingNative = true;
1670 p->shaderStorageImageArrayNonUniformIndexingNative = false;
1671 p->shaderInputAttachmentArrayNonUniformIndexingNative = false;
1672 p->robustBufferAccessUpdateAfterBind = true;
1673 p->quadDivergentImplicitLod = false;
1674 p->maxPerStageDescriptorUpdateAfterBindSamplers = max_bindless_views;
1675 p->maxPerStageDescriptorUpdateAfterBindUniformBuffers = MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS;
1676 p->maxPerStageDescriptorUpdateAfterBindStorageBuffers = UINT32_MAX;
1677 p->maxPerStageDescriptorUpdateAfterBindSampledImages = max_bindless_views;
1678 p->maxPerStageDescriptorUpdateAfterBindStorageImages = max_bindless_views;
1679 p->maxPerStageDescriptorUpdateAfterBindInputAttachments = MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS;
1680 p->maxPerStageUpdateAfterBindResources = UINT32_MAX;
1681 p->maxDescriptorSetUpdateAfterBindSamplers = max_bindless_views;
1682 p->maxDescriptorSetUpdateAfterBindUniformBuffers = 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS;
1683 p->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2;
1684 p->maxDescriptorSetUpdateAfterBindStorageBuffers = UINT32_MAX;
1685 p->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2;
1686 p->maxDescriptorSetUpdateAfterBindSampledImages = max_bindless_views;
1687 p->maxDescriptorSetUpdateAfterBindStorageImages = max_bindless_views;
1688 p->maxDescriptorSetUpdateAfterBindInputAttachments = MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS;
1689
1690 /* We support all of the depth resolve modes */
1691 p->supportedDepthResolveModes = VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR |
1692 VK_RESOLVE_MODE_AVERAGE_BIT_KHR |
1693 VK_RESOLVE_MODE_MIN_BIT_KHR |
1694 VK_RESOLVE_MODE_MAX_BIT_KHR;
1695 /* Average doesn't make sense for stencil so we don't support that */
1696 p->supportedStencilResolveModes = VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR;
1697 if (pdevice->info.gen >= 8) {
1698 /* The advanced stencil resolve modes currently require stencil
1699 * sampling be supported by the hardware.
1700 */
1701 p->supportedStencilResolveModes |= VK_RESOLVE_MODE_MIN_BIT_KHR |
1702 VK_RESOLVE_MODE_MAX_BIT_KHR;
1703 }
1704 p->independentResolveNone = true;
1705 p->independentResolve = true;
1706
1707 p->filterMinmaxSingleComponentFormats = pdevice->info.gen >= 9;
1708 p->filterMinmaxImageComponentMapping = pdevice->info.gen >= 9;
1709
1710 p->maxTimelineSemaphoreValueDifference = UINT64_MAX;
1711
1712 p->framebufferIntegerColorSampleCounts =
1713 isl_device_get_sample_counts(&pdevice->isl_dev);
1714 }
1715
1716 void anv_GetPhysicalDeviceProperties2(
1717 VkPhysicalDevice physicalDevice,
1718 VkPhysicalDeviceProperties2* pProperties)
1719 {
1720 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
1721
1722 anv_GetPhysicalDeviceProperties(physicalDevice, &pProperties->properties);
1723
1724 VkPhysicalDeviceVulkan11Properties core_1_1 = {
1725 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES,
1726 };
1727 anv_get_physical_device_properties_1_1(pdevice, &core_1_1);
1728
1729 VkPhysicalDeviceVulkan12Properties core_1_2 = {
1730 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES,
1731 };
1732 anv_get_physical_device_properties_1_2(pdevice, &core_1_2);
1733
1734 #define CORE_RENAMED_PROPERTY(major, minor, ext_property, core_property) \
1735 memcpy(&properties->ext_property, &core_##major##_##minor.core_property, \
1736 sizeof(core_##major##_##minor.core_property))
1737
1738 #define CORE_PROPERTY(major, minor, property) \
1739 CORE_RENAMED_PROPERTY(major, minor, property, property)
1740
1741 vk_foreach_struct(ext, pProperties->pNext) {
1742 switch (ext->sType) {
1743 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_STENCIL_RESOLVE_PROPERTIES_KHR: {
1744 VkPhysicalDeviceDepthStencilResolvePropertiesKHR *properties =
1745 (VkPhysicalDeviceDepthStencilResolvePropertiesKHR *)ext;
1746 CORE_PROPERTY(1, 2, supportedDepthResolveModes);
1747 CORE_PROPERTY(1, 2, supportedStencilResolveModes);
1748 CORE_PROPERTY(1, 2, independentResolveNone);
1749 CORE_PROPERTY(1, 2, independentResolve);
1750 break;
1751 }
1752
1753 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES_EXT: {
1754 VkPhysicalDeviceDescriptorIndexingPropertiesEXT *properties =
1755 (VkPhysicalDeviceDescriptorIndexingPropertiesEXT *)ext;
1756 CORE_PROPERTY(1, 2, maxUpdateAfterBindDescriptorsInAllPools);
1757 CORE_PROPERTY(1, 2, shaderUniformBufferArrayNonUniformIndexingNative);
1758 CORE_PROPERTY(1, 2, shaderSampledImageArrayNonUniformIndexingNative);
1759 CORE_PROPERTY(1, 2, shaderStorageBufferArrayNonUniformIndexingNative);
1760 CORE_PROPERTY(1, 2, shaderStorageImageArrayNonUniformIndexingNative);
1761 CORE_PROPERTY(1, 2, shaderInputAttachmentArrayNonUniformIndexingNative);
1762 CORE_PROPERTY(1, 2, robustBufferAccessUpdateAfterBind);
1763 CORE_PROPERTY(1, 2, quadDivergentImplicitLod);
1764 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindSamplers);
1765 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindUniformBuffers);
1766 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindStorageBuffers);
1767 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindSampledImages);
1768 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindStorageImages);
1769 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindInputAttachments);
1770 CORE_PROPERTY(1, 2, maxPerStageUpdateAfterBindResources);
1771 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindSamplers);
1772 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindUniformBuffers);
1773 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindUniformBuffersDynamic);
1774 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindStorageBuffers);
1775 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindStorageBuffersDynamic);
1776 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindSampledImages);
1777 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindStorageImages);
1778 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindInputAttachments);
1779 break;
1780 }
1781
1782 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR: {
1783 VkPhysicalDeviceDriverPropertiesKHR *properties =
1784 (VkPhysicalDeviceDriverPropertiesKHR *) ext;
1785 CORE_PROPERTY(1, 2, driverID);
1786 CORE_PROPERTY(1, 2, driverName);
1787 CORE_PROPERTY(1, 2, driverInfo);
1788 CORE_PROPERTY(1, 2, conformanceVersion);
1789 break;
1790 }
1791
1792 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT: {
1793 VkPhysicalDeviceExternalMemoryHostPropertiesEXT *props =
1794 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT *) ext;
1795 /* Userptr needs page aligned memory. */
1796 props->minImportedHostPointerAlignment = 4096;
1797 break;
1798 }
1799
1800 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES: {
1801 VkPhysicalDeviceIDProperties *properties =
1802 (VkPhysicalDeviceIDProperties *)ext;
1803 CORE_PROPERTY(1, 1, deviceUUID);
1804 CORE_PROPERTY(1, 1, driverUUID);
1805 CORE_PROPERTY(1, 1, deviceLUID);
1806 CORE_PROPERTY(1, 1, deviceLUIDValid);
1807 break;
1808 }
1809
1810 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_PROPERTIES_EXT: {
1811 VkPhysicalDeviceInlineUniformBlockPropertiesEXT *props =
1812 (VkPhysicalDeviceInlineUniformBlockPropertiesEXT *)ext;
1813 props->maxInlineUniformBlockSize = MAX_INLINE_UNIFORM_BLOCK_SIZE;
1814 props->maxPerStageDescriptorInlineUniformBlocks =
1815 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1816 props->maxPerStageDescriptorUpdateAfterBindInlineUniformBlocks =
1817 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1818 props->maxDescriptorSetInlineUniformBlocks =
1819 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1820 props->maxDescriptorSetUpdateAfterBindInlineUniformBlocks =
1821 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1822 break;
1823 }
1824
1825 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_PROPERTIES_EXT: {
1826 VkPhysicalDeviceLineRasterizationPropertiesEXT *props =
1827 (VkPhysicalDeviceLineRasterizationPropertiesEXT *)ext;
1828 /* In the Skylake PRM Vol. 7, subsection titled "GIQ (Diamond)
1829 * Sampling Rules - Legacy Mode", it says the following:
1830 *
1831 * "Note that the device divides a pixel into a 16x16 array of
1832 * subpixels, referenced by their upper left corners."
1833 *
1834 * This is the only known reference in the PRMs to the subpixel
1835 * precision of line rasterization and a "16x16 array of subpixels"
1836 * implies 4 subpixel precision bits. Empirical testing has shown
1837 * that 4 subpixel precision bits applies to all line rasterization
1838 * types.
1839 */
1840 props->lineSubPixelPrecisionBits = 4;
1841 break;
1842 }
1843
1844 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES: {
1845 VkPhysicalDeviceMaintenance3Properties *properties =
1846 (VkPhysicalDeviceMaintenance3Properties *)ext;
1847 /* This value doesn't matter for us today as our per-stage
1848 * descriptors are the real limit.
1849 */
1850 CORE_PROPERTY(1, 1, maxPerSetDescriptors);
1851 CORE_PROPERTY(1, 1, maxMemoryAllocationSize);
1852 break;
1853 }
1854
1855 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES: {
1856 VkPhysicalDeviceMultiviewProperties *properties =
1857 (VkPhysicalDeviceMultiviewProperties *)ext;
1858 CORE_PROPERTY(1, 1, maxMultiviewViewCount);
1859 CORE_PROPERTY(1, 1, maxMultiviewInstanceIndex);
1860 break;
1861 }
1862
1863 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT: {
1864 VkPhysicalDevicePCIBusInfoPropertiesEXT *properties =
1865 (VkPhysicalDevicePCIBusInfoPropertiesEXT *)ext;
1866 properties->pciDomain = pdevice->pci_info.domain;
1867 properties->pciBus = pdevice->pci_info.bus;
1868 properties->pciDevice = pdevice->pci_info.device;
1869 properties->pciFunction = pdevice->pci_info.function;
1870 break;
1871 }
1872
1873 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES: {
1874 VkPhysicalDevicePointClippingProperties *properties =
1875 (VkPhysicalDevicePointClippingProperties *) ext;
1876 CORE_PROPERTY(1, 1, pointClippingBehavior);
1877 break;
1878 }
1879
1880 #pragma GCC diagnostic push
1881 #pragma GCC diagnostic ignored "-Wswitch"
1882 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PRESENTATION_PROPERTIES_ANDROID: {
1883 VkPhysicalDevicePresentationPropertiesANDROID *props =
1884 (VkPhysicalDevicePresentationPropertiesANDROID *)ext;
1885 props->sharedImage = VK_FALSE;
1886 break;
1887 }
1888 #pragma GCC diagnostic pop
1889
1890 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES: {
1891 VkPhysicalDeviceProtectedMemoryProperties *properties =
1892 (VkPhysicalDeviceProtectedMemoryProperties *)ext;
1893 CORE_PROPERTY(1, 1, protectedNoFault);
1894 break;
1895 }
1896
1897 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR: {
1898 VkPhysicalDevicePushDescriptorPropertiesKHR *properties =
1899 (VkPhysicalDevicePushDescriptorPropertiesKHR *) ext;
1900 properties->maxPushDescriptors = MAX_PUSH_DESCRIPTORS;
1901 break;
1902 }
1903
1904 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT: {
1905 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT *properties =
1906 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT *)ext;
1907 CORE_PROPERTY(1, 2, filterMinmaxImageComponentMapping);
1908 CORE_PROPERTY(1, 2, filterMinmaxSingleComponentFormats);
1909 break;
1910 }
1911
1912 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES: {
1913 VkPhysicalDeviceSubgroupProperties *properties = (void *)ext;
1914 CORE_PROPERTY(1, 1, subgroupSize);
1915 CORE_RENAMED_PROPERTY(1, 1, supportedStages,
1916 subgroupSupportedStages);
1917 CORE_RENAMED_PROPERTY(1, 1, supportedOperations,
1918 subgroupSupportedOperations);
1919 CORE_RENAMED_PROPERTY(1, 1, quadOperationsInAllStages,
1920 subgroupQuadOperationsInAllStages);
1921 break;
1922 }
1923
1924 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_PROPERTIES_EXT: {
1925 VkPhysicalDeviceSubgroupSizeControlPropertiesEXT *props =
1926 (VkPhysicalDeviceSubgroupSizeControlPropertiesEXT *)ext;
1927 STATIC_ASSERT(8 <= BRW_SUBGROUP_SIZE && BRW_SUBGROUP_SIZE <= 32);
1928 props->minSubgroupSize = 8;
1929 props->maxSubgroupSize = 32;
1930 props->maxComputeWorkgroupSubgroups = pdevice->info.max_cs_threads;
1931 props->requiredSubgroupSizeStages = VK_SHADER_STAGE_COMPUTE_BIT;
1932 break;
1933 }
1934 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT_CONTROLS_PROPERTIES_KHR : {
1935 VkPhysicalDeviceFloatControlsPropertiesKHR *properties = (void *)ext;
1936 CORE_PROPERTY(1, 2, denormBehaviorIndependence);
1937 CORE_PROPERTY(1, 2, roundingModeIndependence);
1938 CORE_PROPERTY(1, 2, shaderDenormFlushToZeroFloat16);
1939 CORE_PROPERTY(1, 2, shaderDenormPreserveFloat16);
1940 CORE_PROPERTY(1, 2, shaderRoundingModeRTEFloat16);
1941 CORE_PROPERTY(1, 2, shaderRoundingModeRTZFloat16);
1942 CORE_PROPERTY(1, 2, shaderSignedZeroInfNanPreserveFloat16);
1943 CORE_PROPERTY(1, 2, shaderDenormFlushToZeroFloat32);
1944 CORE_PROPERTY(1, 2, shaderDenormPreserveFloat32);
1945 CORE_PROPERTY(1, 2, shaderRoundingModeRTEFloat32);
1946 CORE_PROPERTY(1, 2, shaderRoundingModeRTZFloat32);
1947 CORE_PROPERTY(1, 2, shaderSignedZeroInfNanPreserveFloat32);
1948 CORE_PROPERTY(1, 2, shaderDenormFlushToZeroFloat64);
1949 CORE_PROPERTY(1, 2, shaderDenormPreserveFloat64);
1950 CORE_PROPERTY(1, 2, shaderRoundingModeRTEFloat64);
1951 CORE_PROPERTY(1, 2, shaderRoundingModeRTZFloat64);
1952 CORE_PROPERTY(1, 2, shaderSignedZeroInfNanPreserveFloat64);
1953 break;
1954 }
1955
1956 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_PROPERTIES_EXT: {
1957 VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT *props =
1958 (VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT *)ext;
1959
1960 /* From the SKL PRM Vol. 2d, docs for RENDER_SURFACE_STATE::Surface
1961 * Base Address:
1962 *
1963 * "For SURFTYPE_BUFFER non-rendertarget surfaces, this field
1964 * specifies the base address of the first element of the surface,
1965 * computed in software by adding the surface base address to the
1966 * byte offset of the element in the buffer. The base address must
1967 * be aligned to element size."
1968 *
1969 * The typed dataport messages require that things be texel aligned.
1970 * Otherwise, we may just load/store the wrong data or, in the worst
1971 * case, there may be hangs.
1972 */
1973 props->storageTexelBufferOffsetAlignmentBytes = 16;
1974 props->storageTexelBufferOffsetSingleTexelAlignment = true;
1975
1976 /* The sampler, however, is much more forgiving and it can handle
1977 * arbitrary byte alignment for linear and buffer surfaces. It's
1978 * hard to find a good PRM citation for this but years of empirical
1979 * experience demonstrate that this is true.
1980 */
1981 props->uniformTexelBufferOffsetAlignmentBytes = 1;
1982 props->uniformTexelBufferOffsetSingleTexelAlignment = false;
1983 break;
1984 }
1985
1986 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_PROPERTIES_KHR: {
1987 VkPhysicalDeviceTimelineSemaphorePropertiesKHR *properties =
1988 (VkPhysicalDeviceTimelineSemaphorePropertiesKHR *) ext;
1989 CORE_PROPERTY(1, 2, maxTimelineSemaphoreValueDifference);
1990 break;
1991 }
1992
1993 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT: {
1994 VkPhysicalDeviceTransformFeedbackPropertiesEXT *props =
1995 (VkPhysicalDeviceTransformFeedbackPropertiesEXT *)ext;
1996
1997 props->maxTransformFeedbackStreams = MAX_XFB_STREAMS;
1998 props->maxTransformFeedbackBuffers = MAX_XFB_BUFFERS;
1999 props->maxTransformFeedbackBufferSize = (1ull << 32);
2000 props->maxTransformFeedbackStreamDataSize = 128 * 4;
2001 props->maxTransformFeedbackBufferDataSize = 128 * 4;
2002 props->maxTransformFeedbackBufferDataStride = 2048;
2003 props->transformFeedbackQueries = true;
2004 props->transformFeedbackStreamsLinesTriangles = false;
2005 props->transformFeedbackRasterizationStreamSelect = false;
2006 props->transformFeedbackDraw = true;
2007 break;
2008 }
2009
2010 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT: {
2011 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *props =
2012 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *)ext;
2013 /* We have to restrict this a bit for multiview */
2014 props->maxVertexAttribDivisor = UINT32_MAX / 16;
2015 break;
2016 }
2017
2018 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES:
2019 anv_get_physical_device_properties_1_1(pdevice, (void *)ext);
2020 break;
2021
2022 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES:
2023 anv_get_physical_device_properties_1_2(pdevice, (void *)ext);
2024 break;
2025
2026 default:
2027 anv_debug_ignored_stype(ext->sType);
2028 break;
2029 }
2030 }
2031
2032 #undef CORE_RENAMED_PROPERTY
2033 #undef CORE_PROPERTY
2034 }
2035
2036 /* We support exactly one queue family. */
2037 static const VkQueueFamilyProperties
2038 anv_queue_family_properties = {
2039 .queueFlags = VK_QUEUE_GRAPHICS_BIT |
2040 VK_QUEUE_COMPUTE_BIT |
2041 VK_QUEUE_TRANSFER_BIT,
2042 .queueCount = 1,
2043 .timestampValidBits = 36, /* XXX: Real value here */
2044 .minImageTransferGranularity = { 1, 1, 1 },
2045 };
2046
2047 void anv_GetPhysicalDeviceQueueFamilyProperties(
2048 VkPhysicalDevice physicalDevice,
2049 uint32_t* pCount,
2050 VkQueueFamilyProperties* pQueueFamilyProperties)
2051 {
2052 VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pCount);
2053
2054 vk_outarray_append(&out, p) {
2055 *p = anv_queue_family_properties;
2056 }
2057 }
2058
2059 void anv_GetPhysicalDeviceQueueFamilyProperties2(
2060 VkPhysicalDevice physicalDevice,
2061 uint32_t* pQueueFamilyPropertyCount,
2062 VkQueueFamilyProperties2* pQueueFamilyProperties)
2063 {
2064
2065 VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pQueueFamilyPropertyCount);
2066
2067 vk_outarray_append(&out, p) {
2068 p->queueFamilyProperties = anv_queue_family_properties;
2069
2070 vk_foreach_struct(s, p->pNext) {
2071 anv_debug_ignored_stype(s->sType);
2072 }
2073 }
2074 }
2075
2076 void anv_GetPhysicalDeviceMemoryProperties(
2077 VkPhysicalDevice physicalDevice,
2078 VkPhysicalDeviceMemoryProperties* pMemoryProperties)
2079 {
2080 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
2081
2082 pMemoryProperties->memoryTypeCount = physical_device->memory.type_count;
2083 for (uint32_t i = 0; i < physical_device->memory.type_count; i++) {
2084 pMemoryProperties->memoryTypes[i] = (VkMemoryType) {
2085 .propertyFlags = physical_device->memory.types[i].propertyFlags,
2086 .heapIndex = physical_device->memory.types[i].heapIndex,
2087 };
2088 }
2089
2090 pMemoryProperties->memoryHeapCount = physical_device->memory.heap_count;
2091 for (uint32_t i = 0; i < physical_device->memory.heap_count; i++) {
2092 pMemoryProperties->memoryHeaps[i] = (VkMemoryHeap) {
2093 .size = physical_device->memory.heaps[i].size,
2094 .flags = physical_device->memory.heaps[i].flags,
2095 };
2096 }
2097 }
2098
2099 static void
2100 anv_get_memory_budget(VkPhysicalDevice physicalDevice,
2101 VkPhysicalDeviceMemoryBudgetPropertiesEXT *memoryBudget)
2102 {
2103 ANV_FROM_HANDLE(anv_physical_device, device, physicalDevice);
2104 uint64_t sys_available = get_available_system_memory();
2105 assert(sys_available > 0);
2106
2107 VkDeviceSize total_heaps_size = 0;
2108 for (size_t i = 0; i < device->memory.heap_count; i++)
2109 total_heaps_size += device->memory.heaps[i].size;
2110
2111 for (size_t i = 0; i < device->memory.heap_count; i++) {
2112 VkDeviceSize heap_size = device->memory.heaps[i].size;
2113 VkDeviceSize heap_used = device->memory.heaps[i].used;
2114 VkDeviceSize heap_budget;
2115
2116 double heap_proportion = (double) heap_size / total_heaps_size;
2117 VkDeviceSize sys_available_prop = sys_available * heap_proportion;
2118
2119 /*
2120 * Let's not incite the app to starve the system: report at most 90% of
2121 * available system memory.
2122 */
2123 uint64_t heap_available = sys_available_prop * 9 / 10;
2124 heap_budget = MIN2(heap_size, heap_used + heap_available);
2125
2126 /*
2127 * Round down to the nearest MB
2128 */
2129 heap_budget &= ~((1ull << 20) - 1);
2130
2131 /*
2132 * The heapBudget value must be non-zero for array elements less than
2133 * VkPhysicalDeviceMemoryProperties::memoryHeapCount. The heapBudget
2134 * value must be less than or equal to VkMemoryHeap::size for each heap.
2135 */
2136 assert(0 < heap_budget && heap_budget <= heap_size);
2137
2138 memoryBudget->heapUsage[i] = heap_used;
2139 memoryBudget->heapBudget[i] = heap_budget;
2140 }
2141
2142 /* The heapBudget and heapUsage values must be zero for array elements
2143 * greater than or equal to VkPhysicalDeviceMemoryProperties::memoryHeapCount
2144 */
2145 for (uint32_t i = device->memory.heap_count; i < VK_MAX_MEMORY_HEAPS; i++) {
2146 memoryBudget->heapBudget[i] = 0;
2147 memoryBudget->heapUsage[i] = 0;
2148 }
2149 }
2150
2151 void anv_GetPhysicalDeviceMemoryProperties2(
2152 VkPhysicalDevice physicalDevice,
2153 VkPhysicalDeviceMemoryProperties2* pMemoryProperties)
2154 {
2155 anv_GetPhysicalDeviceMemoryProperties(physicalDevice,
2156 &pMemoryProperties->memoryProperties);
2157
2158 vk_foreach_struct(ext, pMemoryProperties->pNext) {
2159 switch (ext->sType) {
2160 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT:
2161 anv_get_memory_budget(physicalDevice, (void*)ext);
2162 break;
2163 default:
2164 anv_debug_ignored_stype(ext->sType);
2165 break;
2166 }
2167 }
2168 }
2169
2170 void
2171 anv_GetDeviceGroupPeerMemoryFeatures(
2172 VkDevice device,
2173 uint32_t heapIndex,
2174 uint32_t localDeviceIndex,
2175 uint32_t remoteDeviceIndex,
2176 VkPeerMemoryFeatureFlags* pPeerMemoryFeatures)
2177 {
2178 assert(localDeviceIndex == 0 && remoteDeviceIndex == 0);
2179 *pPeerMemoryFeatures = VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT |
2180 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT |
2181 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT |
2182 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT;
2183 }
2184
2185 PFN_vkVoidFunction anv_GetInstanceProcAddr(
2186 VkInstance _instance,
2187 const char* pName)
2188 {
2189 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2190
2191 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
2192 * when we have to return valid function pointers, NULL, or it's left
2193 * undefined. See the table for exact details.
2194 */
2195 if (pName == NULL)
2196 return NULL;
2197
2198 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
2199 if (strcmp(pName, "vk" #entrypoint) == 0) \
2200 return (PFN_vkVoidFunction)anv_##entrypoint
2201
2202 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties);
2203 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties);
2204 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceVersion);
2205 LOOKUP_ANV_ENTRYPOINT(CreateInstance);
2206
2207 #undef LOOKUP_ANV_ENTRYPOINT
2208
2209 if (instance == NULL)
2210 return NULL;
2211
2212 int idx = anv_get_instance_entrypoint_index(pName);
2213 if (idx >= 0)
2214 return instance->dispatch.entrypoints[idx];
2215
2216 idx = anv_get_physical_device_entrypoint_index(pName);
2217 if (idx >= 0)
2218 return instance->physical_device_dispatch.entrypoints[idx];
2219
2220 idx = anv_get_device_entrypoint_index(pName);
2221 if (idx >= 0)
2222 return instance->device_dispatch.entrypoints[idx];
2223
2224 return NULL;
2225 }
2226
2227 /* With version 1+ of the loader interface the ICD should expose
2228 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
2229 */
2230 PUBLIC
2231 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
2232 VkInstance instance,
2233 const char* pName);
2234
2235 PUBLIC
2236 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
2237 VkInstance instance,
2238 const char* pName)
2239 {
2240 return anv_GetInstanceProcAddr(instance, pName);
2241 }
2242
2243 PFN_vkVoidFunction anv_GetDeviceProcAddr(
2244 VkDevice _device,
2245 const char* pName)
2246 {
2247 ANV_FROM_HANDLE(anv_device, device, _device);
2248
2249 if (!device || !pName)
2250 return NULL;
2251
2252 int idx = anv_get_device_entrypoint_index(pName);
2253 if (idx < 0)
2254 return NULL;
2255
2256 return device->dispatch.entrypoints[idx];
2257 }
2258
2259 /* With version 4+ of the loader interface the ICD should expose
2260 * vk_icdGetPhysicalDeviceProcAddr()
2261 */
2262 PUBLIC
2263 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetPhysicalDeviceProcAddr(
2264 VkInstance _instance,
2265 const char* pName);
2266
2267 PFN_vkVoidFunction vk_icdGetPhysicalDeviceProcAddr(
2268 VkInstance _instance,
2269 const char* pName)
2270 {
2271 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2272
2273 if (!pName || !instance)
2274 return NULL;
2275
2276 int idx = anv_get_physical_device_entrypoint_index(pName);
2277 if (idx < 0)
2278 return NULL;
2279
2280 return instance->physical_device_dispatch.entrypoints[idx];
2281 }
2282
2283
2284 VkResult
2285 anv_CreateDebugReportCallbackEXT(VkInstance _instance,
2286 const VkDebugReportCallbackCreateInfoEXT* pCreateInfo,
2287 const VkAllocationCallbacks* pAllocator,
2288 VkDebugReportCallbackEXT* pCallback)
2289 {
2290 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2291 return vk_create_debug_report_callback(&instance->debug_report_callbacks,
2292 pCreateInfo, pAllocator, &instance->alloc,
2293 pCallback);
2294 }
2295
2296 void
2297 anv_DestroyDebugReportCallbackEXT(VkInstance _instance,
2298 VkDebugReportCallbackEXT _callback,
2299 const VkAllocationCallbacks* pAllocator)
2300 {
2301 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2302 vk_destroy_debug_report_callback(&instance->debug_report_callbacks,
2303 _callback, pAllocator, &instance->alloc);
2304 }
2305
2306 void
2307 anv_DebugReportMessageEXT(VkInstance _instance,
2308 VkDebugReportFlagsEXT flags,
2309 VkDebugReportObjectTypeEXT objectType,
2310 uint64_t object,
2311 size_t location,
2312 int32_t messageCode,
2313 const char* pLayerPrefix,
2314 const char* pMessage)
2315 {
2316 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2317 vk_debug_report(&instance->debug_report_callbacks, flags, objectType,
2318 object, location, messageCode, pLayerPrefix, pMessage);
2319 }
2320
2321 static struct anv_state
2322 anv_state_pool_emit_data(struct anv_state_pool *pool, size_t size, size_t align, const void *p)
2323 {
2324 struct anv_state state;
2325
2326 state = anv_state_pool_alloc(pool, size, align);
2327 memcpy(state.map, p, size);
2328
2329 return state;
2330 }
2331
2332 /* Haswell border color is a bit of a disaster. Float and unorm formats use a
2333 * straightforward 32-bit float color in the first 64 bytes. Instead of using
2334 * a nice float/integer union like Gen8+, Haswell specifies the integer border
2335 * color as a separate entry /after/ the float color. The layout of this entry
2336 * also depends on the format's bpp (with extra hacks for RG32), and overlaps.
2337 *
2338 * Since we don't know the format/bpp, we can't make any of the border colors
2339 * containing '1' work for all formats, as it would be in the wrong place for
2340 * some of them. We opt to make 32-bit integers work as this seems like the
2341 * most common option. Fortunately, transparent black works regardless, as
2342 * all zeroes is the same in every bit-size.
2343 */
2344 struct hsw_border_color {
2345 float float32[4];
2346 uint32_t _pad0[12];
2347 uint32_t uint32[4];
2348 uint32_t _pad1[108];
2349 };
2350
2351 struct gen8_border_color {
2352 union {
2353 float float32[4];
2354 uint32_t uint32[4];
2355 };
2356 /* Pad out to 64 bytes */
2357 uint32_t _pad[12];
2358 };
2359
2360 static void
2361 anv_device_init_border_colors(struct anv_device *device)
2362 {
2363 if (device->info.is_haswell) {
2364 static const struct hsw_border_color border_colors[] = {
2365 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
2366 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
2367 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
2368 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = { .uint32 = { 0, 0, 0, 0 } },
2369 [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = { .uint32 = { 0, 0, 0, 1 } },
2370 [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = { .uint32 = { 1, 1, 1, 1 } },
2371 };
2372
2373 device->border_colors =
2374 anv_state_pool_emit_data(&device->dynamic_state_pool,
2375 sizeof(border_colors), 512, border_colors);
2376 } else {
2377 static const struct gen8_border_color border_colors[] = {
2378 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
2379 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
2380 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
2381 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = { .uint32 = { 0, 0, 0, 0 } },
2382 [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = { .uint32 = { 0, 0, 0, 1 } },
2383 [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = { .uint32 = { 1, 1, 1, 1 } },
2384 };
2385
2386 device->border_colors =
2387 anv_state_pool_emit_data(&device->dynamic_state_pool,
2388 sizeof(border_colors), 64, border_colors);
2389 }
2390 }
2391
2392 static VkResult
2393 anv_device_init_trivial_batch(struct anv_device *device)
2394 {
2395 VkResult result = anv_device_alloc_bo(device, 4096,
2396 ANV_BO_ALLOC_MAPPED,
2397 0 /* explicit_address */,
2398 &device->trivial_batch_bo);
2399 if (result != VK_SUCCESS)
2400 return result;
2401
2402 struct anv_batch batch = {
2403 .start = device->trivial_batch_bo->map,
2404 .next = device->trivial_batch_bo->map,
2405 .end = device->trivial_batch_bo->map + 4096,
2406 };
2407
2408 anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
2409 anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
2410
2411 if (!device->info.has_llc)
2412 gen_clflush_range(batch.start, batch.next - batch.start);
2413
2414 return VK_SUCCESS;
2415 }
2416
2417 VkResult anv_EnumerateDeviceExtensionProperties(
2418 VkPhysicalDevice physicalDevice,
2419 const char* pLayerName,
2420 uint32_t* pPropertyCount,
2421 VkExtensionProperties* pProperties)
2422 {
2423 ANV_FROM_HANDLE(anv_physical_device, device, physicalDevice);
2424 VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
2425
2426 for (int i = 0; i < ANV_DEVICE_EXTENSION_COUNT; i++) {
2427 if (device->supported_extensions.extensions[i]) {
2428 vk_outarray_append(&out, prop) {
2429 *prop = anv_device_extensions[i];
2430 }
2431 }
2432 }
2433
2434 return vk_outarray_status(&out);
2435 }
2436
2437 static void
2438 anv_device_init_dispatch(struct anv_device *device)
2439 {
2440 const struct anv_instance *instance = device->physical->instance;
2441
2442 const struct anv_device_dispatch_table *genX_table;
2443 switch (device->info.gen) {
2444 case 12:
2445 genX_table = &gen12_device_dispatch_table;
2446 break;
2447 case 11:
2448 genX_table = &gen11_device_dispatch_table;
2449 break;
2450 case 10:
2451 genX_table = &gen10_device_dispatch_table;
2452 break;
2453 case 9:
2454 genX_table = &gen9_device_dispatch_table;
2455 break;
2456 case 8:
2457 genX_table = &gen8_device_dispatch_table;
2458 break;
2459 case 7:
2460 if (device->info.is_haswell)
2461 genX_table = &gen75_device_dispatch_table;
2462 else
2463 genX_table = &gen7_device_dispatch_table;
2464 break;
2465 default:
2466 unreachable("unsupported gen\n");
2467 }
2468
2469 for (unsigned i = 0; i < ARRAY_SIZE(device->dispatch.entrypoints); i++) {
2470 /* Vulkan requires that entrypoints for extensions which have not been
2471 * enabled must not be advertised.
2472 */
2473 if (!anv_device_entrypoint_is_enabled(i, instance->app_info.api_version,
2474 &instance->enabled_extensions,
2475 &device->enabled_extensions)) {
2476 device->dispatch.entrypoints[i] = NULL;
2477 } else if (genX_table->entrypoints[i]) {
2478 device->dispatch.entrypoints[i] = genX_table->entrypoints[i];
2479 } else {
2480 device->dispatch.entrypoints[i] =
2481 anv_device_dispatch_table.entrypoints[i];
2482 }
2483 }
2484 }
2485
2486 static int
2487 vk_priority_to_gen(int priority)
2488 {
2489 switch (priority) {
2490 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT:
2491 return GEN_CONTEXT_LOW_PRIORITY;
2492 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT:
2493 return GEN_CONTEXT_MEDIUM_PRIORITY;
2494 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT:
2495 return GEN_CONTEXT_HIGH_PRIORITY;
2496 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT:
2497 return GEN_CONTEXT_REALTIME_PRIORITY;
2498 default:
2499 unreachable("Invalid priority");
2500 }
2501 }
2502
2503 static VkResult
2504 anv_device_init_hiz_clear_value_bo(struct anv_device *device)
2505 {
2506 VkResult result = anv_device_alloc_bo(device, 4096,
2507 ANV_BO_ALLOC_MAPPED,
2508 0 /* explicit_address */,
2509 &device->hiz_clear_bo);
2510 if (result != VK_SUCCESS)
2511 return result;
2512
2513 union isl_color_value hiz_clear = { .u32 = { 0, } };
2514 hiz_clear.f32[0] = ANV_HZ_FC_VAL;
2515
2516 memcpy(device->hiz_clear_bo->map, hiz_clear.u32, sizeof(hiz_clear.u32));
2517
2518 if (!device->info.has_llc)
2519 gen_clflush_range(device->hiz_clear_bo->map, sizeof(hiz_clear.u32));
2520
2521 return VK_SUCCESS;
2522 }
2523
2524 static bool
2525 get_bo_from_pool(struct gen_batch_decode_bo *ret,
2526 struct anv_block_pool *pool,
2527 uint64_t address)
2528 {
2529 anv_block_pool_foreach_bo(bo, pool) {
2530 uint64_t bo_address = gen_48b_address(bo->offset);
2531 if (address >= bo_address && address < (bo_address + bo->size)) {
2532 *ret = (struct gen_batch_decode_bo) {
2533 .addr = bo_address,
2534 .size = bo->size,
2535 .map = bo->map,
2536 };
2537 return true;
2538 }
2539 }
2540 return false;
2541 }
2542
2543 /* Finding a buffer for batch decoding */
2544 static struct gen_batch_decode_bo
2545 decode_get_bo(void *v_batch, bool ppgtt, uint64_t address)
2546 {
2547 struct anv_device *device = v_batch;
2548 struct gen_batch_decode_bo ret_bo = {};
2549
2550 assert(ppgtt);
2551
2552 if (get_bo_from_pool(&ret_bo, &device->dynamic_state_pool.block_pool, address))
2553 return ret_bo;
2554 if (get_bo_from_pool(&ret_bo, &device->instruction_state_pool.block_pool, address))
2555 return ret_bo;
2556 if (get_bo_from_pool(&ret_bo, &device->binding_table_pool.block_pool, address))
2557 return ret_bo;
2558 if (get_bo_from_pool(&ret_bo, &device->surface_state_pool.block_pool, address))
2559 return ret_bo;
2560
2561 if (!device->cmd_buffer_being_decoded)
2562 return (struct gen_batch_decode_bo) { };
2563
2564 struct anv_batch_bo **bo;
2565
2566 u_vector_foreach(bo, &device->cmd_buffer_being_decoded->seen_bbos) {
2567 /* The decoder zeroes out the top 16 bits, so we need to as well */
2568 uint64_t bo_address = (*bo)->bo->offset & (~0ull >> 16);
2569
2570 if (address >= bo_address && address < bo_address + (*bo)->bo->size) {
2571 return (struct gen_batch_decode_bo) {
2572 .addr = bo_address,
2573 .size = (*bo)->bo->size,
2574 .map = (*bo)->bo->map,
2575 };
2576 }
2577 }
2578
2579 return (struct gen_batch_decode_bo) { };
2580 }
2581
2582 struct gen_aux_map_buffer {
2583 struct gen_buffer base;
2584 struct anv_state state;
2585 };
2586
2587 static struct gen_buffer *
2588 gen_aux_map_buffer_alloc(void *driver_ctx, uint32_t size)
2589 {
2590 struct gen_aux_map_buffer *buf = malloc(sizeof(struct gen_aux_map_buffer));
2591 if (!buf)
2592 return NULL;
2593
2594 struct anv_device *device = (struct anv_device*)driver_ctx;
2595 assert(device->physical->supports_48bit_addresses &&
2596 device->physical->use_softpin);
2597
2598 struct anv_state_pool *pool = &device->dynamic_state_pool;
2599 buf->state = anv_state_pool_alloc(pool, size, size);
2600
2601 buf->base.gpu = pool->block_pool.bo->offset + buf->state.offset;
2602 buf->base.gpu_end = buf->base.gpu + buf->state.alloc_size;
2603 buf->base.map = buf->state.map;
2604 buf->base.driver_bo = &buf->state;
2605 return &buf->base;
2606 }
2607
2608 static void
2609 gen_aux_map_buffer_free(void *driver_ctx, struct gen_buffer *buffer)
2610 {
2611 struct gen_aux_map_buffer *buf = (struct gen_aux_map_buffer*)buffer;
2612 struct anv_device *device = (struct anv_device*)driver_ctx;
2613 struct anv_state_pool *pool = &device->dynamic_state_pool;
2614 anv_state_pool_free(pool, buf->state);
2615 free(buf);
2616 }
2617
2618 static struct gen_mapped_pinned_buffer_alloc aux_map_allocator = {
2619 .alloc = gen_aux_map_buffer_alloc,
2620 .free = gen_aux_map_buffer_free,
2621 };
2622
2623 static VkResult
2624 check_physical_device_features(VkPhysicalDevice physicalDevice,
2625 const VkPhysicalDeviceFeatures *features)
2626 {
2627 VkPhysicalDeviceFeatures supported_features;
2628 anv_GetPhysicalDeviceFeatures(physicalDevice, &supported_features);
2629 VkBool32 *supported_feature = (VkBool32 *)&supported_features;
2630 VkBool32 *enabled_feature = (VkBool32 *)features;
2631 unsigned num_features = sizeof(VkPhysicalDeviceFeatures) / sizeof(VkBool32);
2632 for (uint32_t i = 0; i < num_features; i++) {
2633 if (enabled_feature[i] && !supported_feature[i])
2634 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
2635 }
2636
2637 return VK_SUCCESS;
2638 }
2639
2640 VkResult anv_CreateDevice(
2641 VkPhysicalDevice physicalDevice,
2642 const VkDeviceCreateInfo* pCreateInfo,
2643 const VkAllocationCallbacks* pAllocator,
2644 VkDevice* pDevice)
2645 {
2646 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
2647 VkResult result;
2648 struct anv_device *device;
2649
2650 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO);
2651
2652 struct anv_device_extension_table enabled_extensions = { };
2653 for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
2654 int idx;
2655 for (idx = 0; idx < ANV_DEVICE_EXTENSION_COUNT; idx++) {
2656 if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
2657 anv_device_extensions[idx].extensionName) == 0)
2658 break;
2659 }
2660
2661 if (idx >= ANV_DEVICE_EXTENSION_COUNT)
2662 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
2663
2664 if (!physical_device->supported_extensions.extensions[idx])
2665 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
2666
2667 enabled_extensions.extensions[idx] = true;
2668 }
2669
2670 /* Check enabled features */
2671 bool robust_buffer_access = false;
2672 if (pCreateInfo->pEnabledFeatures) {
2673 result = check_physical_device_features(physicalDevice,
2674 pCreateInfo->pEnabledFeatures);
2675 if (result != VK_SUCCESS)
2676 return result;
2677
2678 if (pCreateInfo->pEnabledFeatures->robustBufferAccess)
2679 robust_buffer_access = true;
2680 }
2681
2682 vk_foreach_struct_const(ext, pCreateInfo->pNext) {
2683 switch (ext->sType) {
2684 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2: {
2685 const VkPhysicalDeviceFeatures2 *features = (const void *)ext;
2686 result = check_physical_device_features(physicalDevice,
2687 &features->features);
2688 if (result != VK_SUCCESS)
2689 return result;
2690
2691 if (features->features.robustBufferAccess)
2692 robust_buffer_access = true;
2693 break;
2694 }
2695
2696 default:
2697 /* Don't warn */
2698 break;
2699 }
2700 }
2701
2702 /* Check requested queues and fail if we are requested to create any
2703 * queues with flags we don't support.
2704 */
2705 assert(pCreateInfo->queueCreateInfoCount > 0);
2706 for (uint32_t i = 0; i < pCreateInfo->queueCreateInfoCount; i++) {
2707 if (pCreateInfo->pQueueCreateInfos[i].flags != 0)
2708 return vk_error(VK_ERROR_INITIALIZATION_FAILED);
2709 }
2710
2711 /* Check if client specified queue priority. */
2712 const VkDeviceQueueGlobalPriorityCreateInfoEXT *queue_priority =
2713 vk_find_struct_const(pCreateInfo->pQueueCreateInfos[0].pNext,
2714 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT);
2715
2716 VkQueueGlobalPriorityEXT priority =
2717 queue_priority ? queue_priority->globalPriority :
2718 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT;
2719
2720 device = vk_alloc2(&physical_device->instance->alloc, pAllocator,
2721 sizeof(*device), 8,
2722 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
2723 if (!device)
2724 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2725
2726 if (INTEL_DEBUG & DEBUG_BATCH) {
2727 const unsigned decode_flags =
2728 GEN_BATCH_DECODE_FULL |
2729 ((INTEL_DEBUG & DEBUG_COLOR) ? GEN_BATCH_DECODE_IN_COLOR : 0) |
2730 GEN_BATCH_DECODE_OFFSETS |
2731 GEN_BATCH_DECODE_FLOATS;
2732
2733 gen_batch_decode_ctx_init(&device->decoder_ctx,
2734 &physical_device->info,
2735 stderr, decode_flags, NULL,
2736 decode_get_bo, NULL, device);
2737 }
2738
2739 device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
2740 device->physical = physical_device;
2741 device->no_hw = physical_device->no_hw;
2742 device->_lost = false;
2743
2744 if (pAllocator)
2745 device->alloc = *pAllocator;
2746 else
2747 device->alloc = physical_device->instance->alloc;
2748
2749 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
2750 device->fd = open(physical_device->path, O_RDWR | O_CLOEXEC);
2751 if (device->fd == -1) {
2752 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2753 goto fail_device;
2754 }
2755
2756 device->context_id = anv_gem_create_context(device);
2757 if (device->context_id == -1) {
2758 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2759 goto fail_fd;
2760 }
2761
2762 result = anv_queue_init(device, &device->queue);
2763 if (result != VK_SUCCESS)
2764 goto fail_context_id;
2765
2766 if (physical_device->use_softpin) {
2767 if (pthread_mutex_init(&device->vma_mutex, NULL) != 0) {
2768 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2769 goto fail_queue;
2770 }
2771
2772 /* keep the page with address zero out of the allocator */
2773 util_vma_heap_init(&device->vma_lo,
2774 LOW_HEAP_MIN_ADDRESS, LOW_HEAP_SIZE);
2775
2776 util_vma_heap_init(&device->vma_cva, CLIENT_VISIBLE_HEAP_MIN_ADDRESS,
2777 CLIENT_VISIBLE_HEAP_SIZE);
2778
2779 /* Leave the last 4GiB out of the high vma range, so that no state
2780 * base address + size can overflow 48 bits. For more information see
2781 * the comment about Wa32bitGeneralStateOffset in anv_allocator.c
2782 */
2783 util_vma_heap_init(&device->vma_hi, HIGH_HEAP_MIN_ADDRESS,
2784 physical_device->gtt_size - (1ull << 32) -
2785 HIGH_HEAP_MIN_ADDRESS);
2786 }
2787
2788 list_inithead(&device->memory_objects);
2789
2790 /* As per spec, the driver implementation may deny requests to acquire
2791 * a priority above the default priority (MEDIUM) if the caller does not
2792 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
2793 * is returned.
2794 */
2795 if (physical_device->has_context_priority) {
2796 int err = anv_gem_set_context_param(device->fd, device->context_id,
2797 I915_CONTEXT_PARAM_PRIORITY,
2798 vk_priority_to_gen(priority));
2799 if (err != 0 && priority > VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT) {
2800 result = vk_error(VK_ERROR_NOT_PERMITTED_EXT);
2801 goto fail_vmas;
2802 }
2803 }
2804
2805 device->info = physical_device->info;
2806 device->isl_dev = physical_device->isl_dev;
2807
2808 /* On Broadwell and later, we can use batch chaining to more efficiently
2809 * implement growing command buffers. Prior to Haswell, the kernel
2810 * command parser gets in the way and we have to fall back to growing
2811 * the batch.
2812 */
2813 device->can_chain_batches = device->info.gen >= 8;
2814
2815 device->robust_buffer_access = robust_buffer_access;
2816 device->enabled_extensions = enabled_extensions;
2817
2818 anv_device_init_dispatch(device);
2819
2820 if (pthread_mutex_init(&device->mutex, NULL) != 0) {
2821 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2822 goto fail_queue;
2823 }
2824
2825 pthread_condattr_t condattr;
2826 if (pthread_condattr_init(&condattr) != 0) {
2827 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2828 goto fail_mutex;
2829 }
2830 if (pthread_condattr_setclock(&condattr, CLOCK_MONOTONIC) != 0) {
2831 pthread_condattr_destroy(&condattr);
2832 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2833 goto fail_mutex;
2834 }
2835 if (pthread_cond_init(&device->queue_submit, &condattr) != 0) {
2836 pthread_condattr_destroy(&condattr);
2837 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2838 goto fail_mutex;
2839 }
2840 pthread_condattr_destroy(&condattr);
2841
2842 result = anv_bo_cache_init(&device->bo_cache);
2843 if (result != VK_SUCCESS)
2844 goto fail_queue_cond;
2845
2846 anv_bo_pool_init(&device->batch_bo_pool, device);
2847
2848 result = anv_state_pool_init(&device->dynamic_state_pool, device,
2849 DYNAMIC_STATE_POOL_MIN_ADDRESS, 16384);
2850 if (result != VK_SUCCESS)
2851 goto fail_batch_bo_pool;
2852
2853 result = anv_state_pool_init(&device->instruction_state_pool, device,
2854 INSTRUCTION_STATE_POOL_MIN_ADDRESS, 16384);
2855 if (result != VK_SUCCESS)
2856 goto fail_dynamic_state_pool;
2857
2858 result = anv_state_pool_init(&device->surface_state_pool, device,
2859 SURFACE_STATE_POOL_MIN_ADDRESS, 4096);
2860 if (result != VK_SUCCESS)
2861 goto fail_instruction_state_pool;
2862
2863 if (physical_device->use_softpin) {
2864 result = anv_state_pool_init(&device->binding_table_pool, device,
2865 BINDING_TABLE_POOL_MIN_ADDRESS, 4096);
2866 if (result != VK_SUCCESS)
2867 goto fail_surface_state_pool;
2868 }
2869
2870 if (device->info.gen >= 12) {
2871 device->aux_map_ctx = gen_aux_map_init(device, &aux_map_allocator,
2872 &physical_device->info);
2873 if (!device->aux_map_ctx)
2874 goto fail_binding_table_pool;
2875 }
2876
2877 result = anv_device_alloc_bo(device, 4096, 0 /* flags */,
2878 0 /* explicit_address */,
2879 &device->workaround_bo);
2880 if (result != VK_SUCCESS)
2881 goto fail_surface_aux_map_pool;
2882
2883 result = anv_device_init_trivial_batch(device);
2884 if (result != VK_SUCCESS)
2885 goto fail_workaround_bo;
2886
2887 if (device->info.gen >= 10) {
2888 result = anv_device_init_hiz_clear_value_bo(device);
2889 if (result != VK_SUCCESS)
2890 goto fail_trivial_batch_bo;
2891 }
2892
2893 anv_scratch_pool_init(device, &device->scratch_pool);
2894
2895 switch (device->info.gen) {
2896 case 7:
2897 if (!device->info.is_haswell)
2898 result = gen7_init_device_state(device);
2899 else
2900 result = gen75_init_device_state(device);
2901 break;
2902 case 8:
2903 result = gen8_init_device_state(device);
2904 break;
2905 case 9:
2906 result = gen9_init_device_state(device);
2907 break;
2908 case 10:
2909 result = gen10_init_device_state(device);
2910 break;
2911 case 11:
2912 result = gen11_init_device_state(device);
2913 break;
2914 case 12:
2915 result = gen12_init_device_state(device);
2916 break;
2917 default:
2918 /* Shouldn't get here as we don't create physical devices for any other
2919 * gens. */
2920 unreachable("unhandled gen");
2921 }
2922 if (result != VK_SUCCESS)
2923 goto fail_workaround_bo;
2924
2925 anv_pipeline_cache_init(&device->default_pipeline_cache, device, true);
2926
2927 anv_device_init_blorp(device);
2928
2929 anv_device_init_border_colors(device);
2930
2931 anv_device_perf_init(device);
2932
2933 *pDevice = anv_device_to_handle(device);
2934
2935 return VK_SUCCESS;
2936
2937 fail_workaround_bo:
2938 anv_scratch_pool_finish(device, &device->scratch_pool);
2939 if (device->info.gen >= 10)
2940 anv_device_release_bo(device, device->hiz_clear_bo);
2941 anv_device_release_bo(device, device->workaround_bo);
2942 fail_trivial_batch_bo:
2943 anv_device_release_bo(device, device->trivial_batch_bo);
2944 fail_surface_aux_map_pool:
2945 if (device->info.gen >= 12) {
2946 gen_aux_map_finish(device->aux_map_ctx);
2947 device->aux_map_ctx = NULL;
2948 }
2949 fail_binding_table_pool:
2950 if (physical_device->use_softpin)
2951 anv_state_pool_finish(&device->binding_table_pool);
2952 fail_surface_state_pool:
2953 anv_state_pool_finish(&device->surface_state_pool);
2954 fail_instruction_state_pool:
2955 anv_state_pool_finish(&device->instruction_state_pool);
2956 fail_dynamic_state_pool:
2957 anv_state_pool_finish(&device->dynamic_state_pool);
2958 fail_batch_bo_pool:
2959 anv_bo_pool_finish(&device->batch_bo_pool);
2960 anv_bo_cache_finish(&device->bo_cache);
2961 fail_queue_cond:
2962 pthread_cond_destroy(&device->queue_submit);
2963 fail_mutex:
2964 pthread_mutex_destroy(&device->mutex);
2965 fail_vmas:
2966 if (physical_device->use_softpin) {
2967 util_vma_heap_finish(&device->vma_hi);
2968 util_vma_heap_finish(&device->vma_cva);
2969 util_vma_heap_finish(&device->vma_lo);
2970 }
2971 fail_queue:
2972 anv_queue_finish(&device->queue);
2973 fail_context_id:
2974 anv_gem_destroy_context(device, device->context_id);
2975 fail_fd:
2976 close(device->fd);
2977 fail_device:
2978 vk_free(&device->alloc, device);
2979
2980 return result;
2981 }
2982
2983 void anv_DestroyDevice(
2984 VkDevice _device,
2985 const VkAllocationCallbacks* pAllocator)
2986 {
2987 ANV_FROM_HANDLE(anv_device, device, _device);
2988
2989 if (!device)
2990 return;
2991
2992 anv_device_finish_blorp(device);
2993
2994 anv_pipeline_cache_finish(&device->default_pipeline_cache);
2995
2996 anv_queue_finish(&device->queue);
2997
2998 #ifdef HAVE_VALGRIND
2999 /* We only need to free these to prevent valgrind errors. The backing
3000 * BO will go away in a couple of lines so we don't actually leak.
3001 */
3002 anv_state_pool_free(&device->dynamic_state_pool, device->border_colors);
3003 anv_state_pool_free(&device->dynamic_state_pool, device->slice_hash);
3004 #endif
3005
3006 anv_scratch_pool_finish(device, &device->scratch_pool);
3007
3008 anv_device_release_bo(device, device->workaround_bo);
3009 anv_device_release_bo(device, device->trivial_batch_bo);
3010 if (device->info.gen >= 10)
3011 anv_device_release_bo(device, device->hiz_clear_bo);
3012
3013 if (device->info.gen >= 12) {
3014 gen_aux_map_finish(device->aux_map_ctx);
3015 device->aux_map_ctx = NULL;
3016 }
3017
3018 if (device->physical->use_softpin)
3019 anv_state_pool_finish(&device->binding_table_pool);
3020 anv_state_pool_finish(&device->surface_state_pool);
3021 anv_state_pool_finish(&device->instruction_state_pool);
3022 anv_state_pool_finish(&device->dynamic_state_pool);
3023
3024 anv_bo_pool_finish(&device->batch_bo_pool);
3025
3026 anv_bo_cache_finish(&device->bo_cache);
3027
3028 if (device->physical->use_softpin) {
3029 util_vma_heap_finish(&device->vma_hi);
3030 util_vma_heap_finish(&device->vma_cva);
3031 util_vma_heap_finish(&device->vma_lo);
3032 }
3033
3034 pthread_cond_destroy(&device->queue_submit);
3035 pthread_mutex_destroy(&device->mutex);
3036
3037 anv_gem_destroy_context(device, device->context_id);
3038
3039 if (INTEL_DEBUG & DEBUG_BATCH)
3040 gen_batch_decode_ctx_finish(&device->decoder_ctx);
3041
3042 close(device->fd);
3043
3044 vk_free(&device->alloc, device);
3045 }
3046
3047 VkResult anv_EnumerateInstanceLayerProperties(
3048 uint32_t* pPropertyCount,
3049 VkLayerProperties* pProperties)
3050 {
3051 if (pProperties == NULL) {
3052 *pPropertyCount = 0;
3053 return VK_SUCCESS;
3054 }
3055
3056 /* None supported at this time */
3057 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
3058 }
3059
3060 VkResult anv_EnumerateDeviceLayerProperties(
3061 VkPhysicalDevice physicalDevice,
3062 uint32_t* pPropertyCount,
3063 VkLayerProperties* pProperties)
3064 {
3065 if (pProperties == NULL) {
3066 *pPropertyCount = 0;
3067 return VK_SUCCESS;
3068 }
3069
3070 /* None supported at this time */
3071 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
3072 }
3073
3074 void anv_GetDeviceQueue(
3075 VkDevice _device,
3076 uint32_t queueNodeIndex,
3077 uint32_t queueIndex,
3078 VkQueue* pQueue)
3079 {
3080 const VkDeviceQueueInfo2 info = {
3081 .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_INFO_2,
3082 .pNext = NULL,
3083 .flags = 0,
3084 .queueFamilyIndex = queueNodeIndex,
3085 .queueIndex = queueIndex,
3086 };
3087
3088 anv_GetDeviceQueue2(_device, &info, pQueue);
3089 }
3090
3091 void anv_GetDeviceQueue2(
3092 VkDevice _device,
3093 const VkDeviceQueueInfo2* pQueueInfo,
3094 VkQueue* pQueue)
3095 {
3096 ANV_FROM_HANDLE(anv_device, device, _device);
3097
3098 assert(pQueueInfo->queueIndex == 0);
3099
3100 if (pQueueInfo->flags == device->queue.flags)
3101 *pQueue = anv_queue_to_handle(&device->queue);
3102 else
3103 *pQueue = NULL;
3104 }
3105
3106 VkResult
3107 _anv_device_set_lost(struct anv_device *device,
3108 const char *file, int line,
3109 const char *msg, ...)
3110 {
3111 VkResult err;
3112 va_list ap;
3113
3114 p_atomic_inc(&device->_lost);
3115
3116 va_start(ap, msg);
3117 err = __vk_errorv(device->physical->instance, device,
3118 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT,
3119 VK_ERROR_DEVICE_LOST, file, line, msg, ap);
3120 va_end(ap);
3121
3122 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
3123 abort();
3124
3125 return err;
3126 }
3127
3128 VkResult
3129 _anv_queue_set_lost(struct anv_queue *queue,
3130 const char *file, int line,
3131 const char *msg, ...)
3132 {
3133 VkResult err;
3134 va_list ap;
3135
3136 p_atomic_inc(&queue->device->_lost);
3137
3138 va_start(ap, msg);
3139 err = __vk_errorv(queue->device->physical->instance, queue->device,
3140 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT,
3141 VK_ERROR_DEVICE_LOST, file, line, msg, ap);
3142 va_end(ap);
3143
3144 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
3145 abort();
3146
3147 return err;
3148 }
3149
3150 VkResult
3151 anv_device_query_status(struct anv_device *device)
3152 {
3153 /* This isn't likely as most of the callers of this function already check
3154 * for it. However, it doesn't hurt to check and it potentially lets us
3155 * avoid an ioctl.
3156 */
3157 if (anv_device_is_lost(device))
3158 return VK_ERROR_DEVICE_LOST;
3159
3160 uint32_t active, pending;
3161 int ret = anv_gem_gpu_get_reset_stats(device, &active, &pending);
3162 if (ret == -1) {
3163 /* We don't know the real error. */
3164 return anv_device_set_lost(device, "get_reset_stats failed: %m");
3165 }
3166
3167 if (active) {
3168 return anv_device_set_lost(device, "GPU hung on one of our command buffers");
3169 } else if (pending) {
3170 return anv_device_set_lost(device, "GPU hung with commands in-flight");
3171 }
3172
3173 return VK_SUCCESS;
3174 }
3175
3176 VkResult
3177 anv_device_bo_busy(struct anv_device *device, struct anv_bo *bo)
3178 {
3179 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
3180 * Other usages of the BO (such as on different hardware) will not be
3181 * flagged as "busy" by this ioctl. Use with care.
3182 */
3183 int ret = anv_gem_busy(device, bo->gem_handle);
3184 if (ret == 1) {
3185 return VK_NOT_READY;
3186 } else if (ret == -1) {
3187 /* We don't know the real error. */
3188 return anv_device_set_lost(device, "gem wait failed: %m");
3189 }
3190
3191 /* Query for device status after the busy call. If the BO we're checking
3192 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
3193 * client because it clearly doesn't have valid data. Yes, this most
3194 * likely means an ioctl, but we just did an ioctl to query the busy status
3195 * so it's no great loss.
3196 */
3197 return anv_device_query_status(device);
3198 }
3199
3200 VkResult
3201 anv_device_wait(struct anv_device *device, struct anv_bo *bo,
3202 int64_t timeout)
3203 {
3204 int ret = anv_gem_wait(device, bo->gem_handle, &timeout);
3205 if (ret == -1 && errno == ETIME) {
3206 return VK_TIMEOUT;
3207 } else if (ret == -1) {
3208 /* We don't know the real error. */
3209 return anv_device_set_lost(device, "gem wait failed: %m");
3210 }
3211
3212 /* Query for device status after the wait. If the BO we're waiting on got
3213 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
3214 * because it clearly doesn't have valid data. Yes, this most likely means
3215 * an ioctl, but we just did an ioctl to wait so it's no great loss.
3216 */
3217 return anv_device_query_status(device);
3218 }
3219
3220 VkResult anv_DeviceWaitIdle(
3221 VkDevice _device)
3222 {
3223 ANV_FROM_HANDLE(anv_device, device, _device);
3224
3225 if (anv_device_is_lost(device))
3226 return VK_ERROR_DEVICE_LOST;
3227
3228 return anv_queue_submit_simple_batch(&device->queue, NULL);
3229 }
3230
3231 uint64_t
3232 anv_vma_alloc(struct anv_device *device,
3233 uint64_t size, uint64_t align,
3234 enum anv_bo_alloc_flags alloc_flags,
3235 uint64_t client_address)
3236 {
3237 pthread_mutex_lock(&device->vma_mutex);
3238
3239 uint64_t addr = 0;
3240
3241 if (alloc_flags & ANV_BO_ALLOC_CLIENT_VISIBLE_ADDRESS) {
3242 if (client_address) {
3243 if (util_vma_heap_alloc_addr(&device->vma_cva,
3244 client_address, size)) {
3245 addr = client_address;
3246 }
3247 } else {
3248 addr = util_vma_heap_alloc(&device->vma_cva, size, align);
3249 }
3250 /* We don't want to fall back to other heaps */
3251 goto done;
3252 }
3253
3254 assert(client_address == 0);
3255
3256 if (!(alloc_flags & ANV_BO_ALLOC_32BIT_ADDRESS))
3257 addr = util_vma_heap_alloc(&device->vma_hi, size, align);
3258
3259 if (addr == 0)
3260 addr = util_vma_heap_alloc(&device->vma_lo, size, align);
3261
3262 done:
3263 pthread_mutex_unlock(&device->vma_mutex);
3264
3265 assert(addr == gen_48b_address(addr));
3266 return gen_canonical_address(addr);
3267 }
3268
3269 void
3270 anv_vma_free(struct anv_device *device,
3271 uint64_t address, uint64_t size)
3272 {
3273 const uint64_t addr_48b = gen_48b_address(address);
3274
3275 pthread_mutex_lock(&device->vma_mutex);
3276
3277 if (addr_48b >= LOW_HEAP_MIN_ADDRESS &&
3278 addr_48b <= LOW_HEAP_MAX_ADDRESS) {
3279 util_vma_heap_free(&device->vma_lo, addr_48b, size);
3280 } else if (addr_48b >= CLIENT_VISIBLE_HEAP_MIN_ADDRESS &&
3281 addr_48b <= CLIENT_VISIBLE_HEAP_MAX_ADDRESS) {
3282 util_vma_heap_free(&device->vma_cva, addr_48b, size);
3283 } else {
3284 assert(addr_48b >= HIGH_HEAP_MIN_ADDRESS);
3285 util_vma_heap_free(&device->vma_hi, addr_48b, size);
3286 }
3287
3288 pthread_mutex_unlock(&device->vma_mutex);
3289 }
3290
3291 VkResult anv_AllocateMemory(
3292 VkDevice _device,
3293 const VkMemoryAllocateInfo* pAllocateInfo,
3294 const VkAllocationCallbacks* pAllocator,
3295 VkDeviceMemory* pMem)
3296 {
3297 ANV_FROM_HANDLE(anv_device, device, _device);
3298 struct anv_physical_device *pdevice = device->physical;
3299 struct anv_device_memory *mem;
3300 VkResult result = VK_SUCCESS;
3301
3302 assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
3303
3304 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
3305 assert(pAllocateInfo->allocationSize > 0);
3306
3307 VkDeviceSize aligned_alloc_size =
3308 align_u64(pAllocateInfo->allocationSize, 4096);
3309
3310 if (aligned_alloc_size > MAX_MEMORY_ALLOCATION_SIZE)
3311 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
3312
3313 assert(pAllocateInfo->memoryTypeIndex < pdevice->memory.type_count);
3314 struct anv_memory_type *mem_type =
3315 &pdevice->memory.types[pAllocateInfo->memoryTypeIndex];
3316 assert(mem_type->heapIndex < pdevice->memory.heap_count);
3317 struct anv_memory_heap *mem_heap =
3318 &pdevice->memory.heaps[mem_type->heapIndex];
3319
3320 uint64_t mem_heap_used = p_atomic_read(&mem_heap->used);
3321 if (mem_heap_used + aligned_alloc_size > mem_heap->size)
3322 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
3323
3324 mem = vk_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
3325 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
3326 if (mem == NULL)
3327 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
3328
3329 assert(pAllocateInfo->memoryTypeIndex < pdevice->memory.type_count);
3330 mem->type = mem_type;
3331 mem->map = NULL;
3332 mem->map_size = 0;
3333 mem->ahw = NULL;
3334 mem->host_ptr = NULL;
3335
3336 enum anv_bo_alloc_flags alloc_flags = 0;
3337
3338 const VkExportMemoryAllocateInfo *export_info = NULL;
3339 const VkImportAndroidHardwareBufferInfoANDROID *ahw_import_info = NULL;
3340 const VkImportMemoryFdInfoKHR *fd_info = NULL;
3341 const VkImportMemoryHostPointerInfoEXT *host_ptr_info = NULL;
3342 const VkMemoryDedicatedAllocateInfo *dedicated_info = NULL;
3343 VkMemoryAllocateFlags vk_flags = 0;
3344 uint64_t client_address = 0;
3345
3346 vk_foreach_struct_const(ext, pAllocateInfo->pNext) {
3347 switch (ext->sType) {
3348 case VK_STRUCTURE_TYPE_EXPORT_MEMORY_ALLOCATE_INFO:
3349 export_info = (void *)ext;
3350 break;
3351
3352 case VK_STRUCTURE_TYPE_IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID:
3353 ahw_import_info = (void *)ext;
3354 break;
3355
3356 case VK_STRUCTURE_TYPE_IMPORT_MEMORY_FD_INFO_KHR:
3357 fd_info = (void *)ext;
3358 break;
3359
3360 case VK_STRUCTURE_TYPE_IMPORT_MEMORY_HOST_POINTER_INFO_EXT:
3361 host_ptr_info = (void *)ext;
3362 break;
3363
3364 case VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_FLAGS_INFO: {
3365 const VkMemoryAllocateFlagsInfo *flags_info = (void *)ext;
3366 vk_flags = flags_info->flags;
3367 break;
3368 }
3369
3370 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_ALLOCATE_INFO:
3371 dedicated_info = (void *)ext;
3372 break;
3373
3374 case VK_STRUCTURE_TYPE_MEMORY_OPAQUE_CAPTURE_ADDRESS_ALLOCATE_INFO_KHR: {
3375 const VkMemoryOpaqueCaptureAddressAllocateInfoKHR *addr_info =
3376 (const VkMemoryOpaqueCaptureAddressAllocateInfoKHR *)ext;
3377 client_address = addr_info->opaqueCaptureAddress;
3378 break;
3379 }
3380
3381 default:
3382 anv_debug_ignored_stype(ext->sType);
3383 break;
3384 }
3385 }
3386
3387 /* By default, we want all VkDeviceMemory objects to support CCS */
3388 if (device->physical->has_implicit_ccs)
3389 alloc_flags |= ANV_BO_ALLOC_IMPLICIT_CCS;
3390
3391 if (vk_flags & VK_MEMORY_ALLOCATE_DEVICE_ADDRESS_BIT_KHR)
3392 alloc_flags |= ANV_BO_ALLOC_CLIENT_VISIBLE_ADDRESS;
3393
3394 if ((export_info && export_info->handleTypes) ||
3395 (fd_info && fd_info->handleType) ||
3396 (host_ptr_info && host_ptr_info->handleType)) {
3397 /* Anything imported or exported is EXTERNAL */
3398 alloc_flags |= ANV_BO_ALLOC_EXTERNAL;
3399
3400 /* We can't have implicit CCS on external memory with an AUX-table.
3401 * Doing so would require us to sync the aux tables across processes
3402 * which is impractical.
3403 */
3404 if (device->info.has_aux_map)
3405 alloc_flags &= ~ANV_BO_ALLOC_IMPLICIT_CCS;
3406 }
3407
3408 /* Check if we need to support Android HW buffer export. If so,
3409 * create AHardwareBuffer and import memory from it.
3410 */
3411 bool android_export = false;
3412 if (export_info && export_info->handleTypes &
3413 VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID)
3414 android_export = true;
3415
3416 if (ahw_import_info) {
3417 result = anv_import_ahw_memory(_device, mem, ahw_import_info);
3418 if (result != VK_SUCCESS)
3419 goto fail;
3420
3421 goto success;
3422 } else if (android_export) {
3423 result = anv_create_ahw_memory(_device, mem, pAllocateInfo);
3424 if (result != VK_SUCCESS)
3425 goto fail;
3426
3427 const VkImportAndroidHardwareBufferInfoANDROID import_info = {
3428 .buffer = mem->ahw,
3429 };
3430 result = anv_import_ahw_memory(_device, mem, &import_info);
3431 if (result != VK_SUCCESS)
3432 goto fail;
3433
3434 goto success;
3435 }
3436
3437 /* The Vulkan spec permits handleType to be 0, in which case the struct is
3438 * ignored.
3439 */
3440 if (fd_info && fd_info->handleType) {
3441 /* At the moment, we support only the below handle types. */
3442 assert(fd_info->handleType ==
3443 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
3444 fd_info->handleType ==
3445 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
3446
3447 result = anv_device_import_bo(device, fd_info->fd, alloc_flags,
3448 client_address, &mem->bo);
3449 if (result != VK_SUCCESS)
3450 goto fail;
3451
3452 /* For security purposes, we reject importing the bo if it's smaller
3453 * than the requested allocation size. This prevents a malicious client
3454 * from passing a buffer to a trusted client, lying about the size, and
3455 * telling the trusted client to try and texture from an image that goes
3456 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
3457 * in the trusted client. The trusted client can protect itself against
3458 * this sort of attack but only if it can trust the buffer size.
3459 */
3460 if (mem->bo->size < aligned_alloc_size) {
3461 result = vk_errorf(device, device, VK_ERROR_INVALID_EXTERNAL_HANDLE,
3462 "aligned allocationSize too large for "
3463 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT: "
3464 "%"PRIu64"B > %"PRIu64"B",
3465 aligned_alloc_size, mem->bo->size);
3466 anv_device_release_bo(device, mem->bo);
3467 goto fail;
3468 }
3469
3470 /* From the Vulkan spec:
3471 *
3472 * "Importing memory from a file descriptor transfers ownership of
3473 * the file descriptor from the application to the Vulkan
3474 * implementation. The application must not perform any operations on
3475 * the file descriptor after a successful import."
3476 *
3477 * If the import fails, we leave the file descriptor open.
3478 */
3479 close(fd_info->fd);
3480 goto success;
3481 }
3482
3483 if (host_ptr_info && host_ptr_info->handleType) {
3484 if (host_ptr_info->handleType ==
3485 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_MAPPED_FOREIGN_MEMORY_BIT_EXT) {
3486 result = vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE);
3487 goto fail;
3488 }
3489
3490 assert(host_ptr_info->handleType ==
3491 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT);
3492
3493 result = anv_device_import_bo_from_host_ptr(device,
3494 host_ptr_info->pHostPointer,
3495 pAllocateInfo->allocationSize,
3496 alloc_flags,
3497 client_address,
3498 &mem->bo);
3499 if (result != VK_SUCCESS)
3500 goto fail;
3501
3502 mem->host_ptr = host_ptr_info->pHostPointer;
3503 goto success;
3504 }
3505
3506 /* Regular allocate (not importing memory). */
3507
3508 result = anv_device_alloc_bo(device, pAllocateInfo->allocationSize,
3509 alloc_flags, client_address, &mem->bo);
3510 if (result != VK_SUCCESS)
3511 goto fail;
3512
3513 if (dedicated_info && dedicated_info->image != VK_NULL_HANDLE) {
3514 ANV_FROM_HANDLE(anv_image, image, dedicated_info->image);
3515
3516 /* Some legacy (non-modifiers) consumers need the tiling to be set on
3517 * the BO. In this case, we have a dedicated allocation.
3518 */
3519 if (image->needs_set_tiling) {
3520 const uint32_t i915_tiling =
3521 isl_tiling_to_i915_tiling(image->planes[0].surface.isl.tiling);
3522 int ret = anv_gem_set_tiling(device, mem->bo->gem_handle,
3523 image->planes[0].surface.isl.row_pitch_B,
3524 i915_tiling);
3525 if (ret) {
3526 anv_device_release_bo(device, mem->bo);
3527 result = vk_errorf(device, device, VK_ERROR_OUT_OF_DEVICE_MEMORY,
3528 "failed to set BO tiling: %m");
3529 goto fail;
3530 }
3531 }
3532 }
3533
3534 success:
3535 mem_heap_used = p_atomic_add_return(&mem_heap->used, mem->bo->size);
3536 if (mem_heap_used > mem_heap->size) {
3537 p_atomic_add(&mem_heap->used, -mem->bo->size);
3538 anv_device_release_bo(device, mem->bo);
3539 result = vk_errorf(device, device, VK_ERROR_OUT_OF_DEVICE_MEMORY,
3540 "Out of heap memory");
3541 goto fail;
3542 }
3543
3544 pthread_mutex_lock(&device->mutex);
3545 list_addtail(&mem->link, &device->memory_objects);
3546 pthread_mutex_unlock(&device->mutex);
3547
3548 *pMem = anv_device_memory_to_handle(mem);
3549
3550 return VK_SUCCESS;
3551
3552 fail:
3553 vk_free2(&device->alloc, pAllocator, mem);
3554
3555 return result;
3556 }
3557
3558 VkResult anv_GetMemoryFdKHR(
3559 VkDevice device_h,
3560 const VkMemoryGetFdInfoKHR* pGetFdInfo,
3561 int* pFd)
3562 {
3563 ANV_FROM_HANDLE(anv_device, dev, device_h);
3564 ANV_FROM_HANDLE(anv_device_memory, mem, pGetFdInfo->memory);
3565
3566 assert(pGetFdInfo->sType == VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR);
3567
3568 assert(pGetFdInfo->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
3569 pGetFdInfo->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
3570
3571 return anv_device_export_bo(dev, mem->bo, pFd);
3572 }
3573
3574 VkResult anv_GetMemoryFdPropertiesKHR(
3575 VkDevice _device,
3576 VkExternalMemoryHandleTypeFlagBits handleType,
3577 int fd,
3578 VkMemoryFdPropertiesKHR* pMemoryFdProperties)
3579 {
3580 ANV_FROM_HANDLE(anv_device, device, _device);
3581
3582 switch (handleType) {
3583 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT:
3584 /* dma-buf can be imported as any memory type */
3585 pMemoryFdProperties->memoryTypeBits =
3586 (1 << device->physical->memory.type_count) - 1;
3587 return VK_SUCCESS;
3588
3589 default:
3590 /* The valid usage section for this function says:
3591 *
3592 * "handleType must not be one of the handle types defined as
3593 * opaque."
3594 *
3595 * So opaque handle types fall into the default "unsupported" case.
3596 */
3597 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE);
3598 }
3599 }
3600
3601 VkResult anv_GetMemoryHostPointerPropertiesEXT(
3602 VkDevice _device,
3603 VkExternalMemoryHandleTypeFlagBits handleType,
3604 const void* pHostPointer,
3605 VkMemoryHostPointerPropertiesEXT* pMemoryHostPointerProperties)
3606 {
3607 ANV_FROM_HANDLE(anv_device, device, _device);
3608
3609 assert(pMemoryHostPointerProperties->sType ==
3610 VK_STRUCTURE_TYPE_MEMORY_HOST_POINTER_PROPERTIES_EXT);
3611
3612 switch (handleType) {
3613 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT:
3614 /* Host memory can be imported as any memory type. */
3615 pMemoryHostPointerProperties->memoryTypeBits =
3616 (1ull << device->physical->memory.type_count) - 1;
3617
3618 return VK_SUCCESS;
3619
3620 default:
3621 return VK_ERROR_INVALID_EXTERNAL_HANDLE;
3622 }
3623 }
3624
3625 void anv_FreeMemory(
3626 VkDevice _device,
3627 VkDeviceMemory _mem,
3628 const VkAllocationCallbacks* pAllocator)
3629 {
3630 ANV_FROM_HANDLE(anv_device, device, _device);
3631 ANV_FROM_HANDLE(anv_device_memory, mem, _mem);
3632
3633 if (mem == NULL)
3634 return;
3635
3636 pthread_mutex_lock(&device->mutex);
3637 list_del(&mem->link);
3638 pthread_mutex_unlock(&device->mutex);
3639
3640 if (mem->map)
3641 anv_UnmapMemory(_device, _mem);
3642
3643 p_atomic_add(&device->physical->memory.heaps[mem->type->heapIndex].used,
3644 -mem->bo->size);
3645
3646 anv_device_release_bo(device, mem->bo);
3647
3648 #if defined(ANDROID) && ANDROID_API_LEVEL >= 26
3649 if (mem->ahw)
3650 AHardwareBuffer_release(mem->ahw);
3651 #endif
3652
3653 vk_free2(&device->alloc, pAllocator, mem);
3654 }
3655
3656 VkResult anv_MapMemory(
3657 VkDevice _device,
3658 VkDeviceMemory _memory,
3659 VkDeviceSize offset,
3660 VkDeviceSize size,
3661 VkMemoryMapFlags flags,
3662 void** ppData)
3663 {
3664 ANV_FROM_HANDLE(anv_device, device, _device);
3665 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
3666
3667 if (mem == NULL) {
3668 *ppData = NULL;
3669 return VK_SUCCESS;
3670 }
3671
3672 if (mem->host_ptr) {
3673 *ppData = mem->host_ptr + offset;
3674 return VK_SUCCESS;
3675 }
3676
3677 if (size == VK_WHOLE_SIZE)
3678 size = mem->bo->size - offset;
3679
3680 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
3681 *
3682 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
3683 * assert(size != 0);
3684 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
3685 * equal to the size of the memory minus offset
3686 */
3687 assert(size > 0);
3688 assert(offset + size <= mem->bo->size);
3689
3690 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
3691 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
3692 * at a time is valid. We could just mmap up front and return an offset
3693 * pointer here, but that may exhaust virtual memory on 32 bit
3694 * userspace. */
3695
3696 uint32_t gem_flags = 0;
3697
3698 if (!device->info.has_llc &&
3699 (mem->type->propertyFlags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT))
3700 gem_flags |= I915_MMAP_WC;
3701
3702 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
3703 uint64_t map_offset;
3704 if (!device->physical->has_mmap_offset)
3705 map_offset = offset & ~4095ull;
3706 else
3707 map_offset = 0;
3708 assert(offset >= map_offset);
3709 uint64_t map_size = (offset + size) - map_offset;
3710
3711 /* Let's map whole pages */
3712 map_size = align_u64(map_size, 4096);
3713
3714 void *map = anv_gem_mmap(device, mem->bo->gem_handle,
3715 map_offset, map_size, gem_flags);
3716 if (map == MAP_FAILED)
3717 return vk_error(VK_ERROR_MEMORY_MAP_FAILED);
3718
3719 mem->map = map;
3720 mem->map_size = map_size;
3721
3722 *ppData = mem->map + (offset - map_offset);
3723
3724 return VK_SUCCESS;
3725 }
3726
3727 void anv_UnmapMemory(
3728 VkDevice _device,
3729 VkDeviceMemory _memory)
3730 {
3731 ANV_FROM_HANDLE(anv_device, device, _device);
3732 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
3733
3734 if (mem == NULL || mem->host_ptr)
3735 return;
3736
3737 anv_gem_munmap(device, mem->map, mem->map_size);
3738
3739 mem->map = NULL;
3740 mem->map_size = 0;
3741 }
3742
3743 static void
3744 clflush_mapped_ranges(struct anv_device *device,
3745 uint32_t count,
3746 const VkMappedMemoryRange *ranges)
3747 {
3748 for (uint32_t i = 0; i < count; i++) {
3749 ANV_FROM_HANDLE(anv_device_memory, mem, ranges[i].memory);
3750 if (ranges[i].offset >= mem->map_size)
3751 continue;
3752
3753 gen_clflush_range(mem->map + ranges[i].offset,
3754 MIN2(ranges[i].size, mem->map_size - ranges[i].offset));
3755 }
3756 }
3757
3758 VkResult anv_FlushMappedMemoryRanges(
3759 VkDevice _device,
3760 uint32_t memoryRangeCount,
3761 const VkMappedMemoryRange* pMemoryRanges)
3762 {
3763 ANV_FROM_HANDLE(anv_device, device, _device);
3764
3765 if (device->info.has_llc)
3766 return VK_SUCCESS;
3767
3768 /* Make sure the writes we're flushing have landed. */
3769 __builtin_ia32_mfence();
3770
3771 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
3772
3773 return VK_SUCCESS;
3774 }
3775
3776 VkResult anv_InvalidateMappedMemoryRanges(
3777 VkDevice _device,
3778 uint32_t memoryRangeCount,
3779 const VkMappedMemoryRange* pMemoryRanges)
3780 {
3781 ANV_FROM_HANDLE(anv_device, device, _device);
3782
3783 if (device->info.has_llc)
3784 return VK_SUCCESS;
3785
3786 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
3787
3788 /* Make sure no reads get moved up above the invalidate. */
3789 __builtin_ia32_mfence();
3790
3791 return VK_SUCCESS;
3792 }
3793
3794 void anv_GetBufferMemoryRequirements(
3795 VkDevice _device,
3796 VkBuffer _buffer,
3797 VkMemoryRequirements* pMemoryRequirements)
3798 {
3799 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
3800 ANV_FROM_HANDLE(anv_device, device, _device);
3801
3802 /* The Vulkan spec (git aaed022) says:
3803 *
3804 * memoryTypeBits is a bitfield and contains one bit set for every
3805 * supported memory type for the resource. The bit `1<<i` is set if and
3806 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3807 * structure for the physical device is supported.
3808 */
3809 uint32_t memory_types = (1ull << device->physical->memory.type_count) - 1;
3810
3811 /* Base alignment requirement of a cache line */
3812 uint32_t alignment = 16;
3813
3814 /* We need an alignment of 32 for pushing UBOs */
3815 if (buffer->usage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT)
3816 alignment = MAX2(alignment, 32);
3817
3818 pMemoryRequirements->size = buffer->size;
3819 pMemoryRequirements->alignment = alignment;
3820
3821 /* Storage and Uniform buffers should have their size aligned to
3822 * 32-bits to avoid boundary checks when last DWord is not complete.
3823 * This would ensure that not internal padding would be needed for
3824 * 16-bit types.
3825 */
3826 if (device->robust_buffer_access &&
3827 (buffer->usage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT ||
3828 buffer->usage & VK_BUFFER_USAGE_STORAGE_BUFFER_BIT))
3829 pMemoryRequirements->size = align_u64(buffer->size, 4);
3830
3831 pMemoryRequirements->memoryTypeBits = memory_types;
3832 }
3833
3834 void anv_GetBufferMemoryRequirements2(
3835 VkDevice _device,
3836 const VkBufferMemoryRequirementsInfo2* pInfo,
3837 VkMemoryRequirements2* pMemoryRequirements)
3838 {
3839 anv_GetBufferMemoryRequirements(_device, pInfo->buffer,
3840 &pMemoryRequirements->memoryRequirements);
3841
3842 vk_foreach_struct(ext, pMemoryRequirements->pNext) {
3843 switch (ext->sType) {
3844 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
3845 VkMemoryDedicatedRequirements *requirements = (void *)ext;
3846 requirements->prefersDedicatedAllocation = false;
3847 requirements->requiresDedicatedAllocation = false;
3848 break;
3849 }
3850
3851 default:
3852 anv_debug_ignored_stype(ext->sType);
3853 break;
3854 }
3855 }
3856 }
3857
3858 void anv_GetImageMemoryRequirements(
3859 VkDevice _device,
3860 VkImage _image,
3861 VkMemoryRequirements* pMemoryRequirements)
3862 {
3863 ANV_FROM_HANDLE(anv_image, image, _image);
3864 ANV_FROM_HANDLE(anv_device, device, _device);
3865
3866 /* The Vulkan spec (git aaed022) says:
3867 *
3868 * memoryTypeBits is a bitfield and contains one bit set for every
3869 * supported memory type for the resource. The bit `1<<i` is set if and
3870 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3871 * structure for the physical device is supported.
3872 *
3873 * All types are currently supported for images.
3874 */
3875 uint32_t memory_types = (1ull << device->physical->memory.type_count) - 1;
3876
3877 /* We must have image allocated or imported at this point. According to the
3878 * specification, external images must have been bound to memory before
3879 * calling GetImageMemoryRequirements.
3880 */
3881 assert(image->size > 0);
3882
3883 pMemoryRequirements->size = image->size;
3884 pMemoryRequirements->alignment = image->alignment;
3885 pMemoryRequirements->memoryTypeBits = memory_types;
3886 }
3887
3888 void anv_GetImageMemoryRequirements2(
3889 VkDevice _device,
3890 const VkImageMemoryRequirementsInfo2* pInfo,
3891 VkMemoryRequirements2* pMemoryRequirements)
3892 {
3893 ANV_FROM_HANDLE(anv_device, device, _device);
3894 ANV_FROM_HANDLE(anv_image, image, pInfo->image);
3895
3896 anv_GetImageMemoryRequirements(_device, pInfo->image,
3897 &pMemoryRequirements->memoryRequirements);
3898
3899 vk_foreach_struct_const(ext, pInfo->pNext) {
3900 switch (ext->sType) {
3901 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO: {
3902 const VkImagePlaneMemoryRequirementsInfo *plane_reqs =
3903 (const VkImagePlaneMemoryRequirementsInfo *) ext;
3904 uint32_t plane = anv_image_aspect_to_plane(image->aspects,
3905 plane_reqs->planeAspect);
3906
3907 assert(image->planes[plane].offset == 0);
3908
3909 /* The Vulkan spec (git aaed022) says:
3910 *
3911 * memoryTypeBits is a bitfield and contains one bit set for every
3912 * supported memory type for the resource. The bit `1<<i` is set
3913 * if and only if the memory type `i` in the
3914 * VkPhysicalDeviceMemoryProperties structure for the physical
3915 * device is supported.
3916 *
3917 * All types are currently supported for images.
3918 */
3919 pMemoryRequirements->memoryRequirements.memoryTypeBits =
3920 (1ull << device->physical->memory.type_count) - 1;
3921
3922 /* We must have image allocated or imported at this point. According to the
3923 * specification, external images must have been bound to memory before
3924 * calling GetImageMemoryRequirements.
3925 */
3926 assert(image->planes[plane].size > 0);
3927
3928 pMemoryRequirements->memoryRequirements.size = image->planes[plane].size;
3929 pMemoryRequirements->memoryRequirements.alignment =
3930 image->planes[plane].alignment;
3931 break;
3932 }
3933
3934 default:
3935 anv_debug_ignored_stype(ext->sType);
3936 break;
3937 }
3938 }
3939
3940 vk_foreach_struct(ext, pMemoryRequirements->pNext) {
3941 switch (ext->sType) {
3942 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
3943 VkMemoryDedicatedRequirements *requirements = (void *)ext;
3944 if (image->needs_set_tiling || image->external_format) {
3945 /* If we need to set the tiling for external consumers, we need a
3946 * dedicated allocation.
3947 *
3948 * See also anv_AllocateMemory.
3949 */
3950 requirements->prefersDedicatedAllocation = true;
3951 requirements->requiresDedicatedAllocation = true;
3952 } else {
3953 requirements->prefersDedicatedAllocation = false;
3954 requirements->requiresDedicatedAllocation = false;
3955 }
3956 break;
3957 }
3958
3959 default:
3960 anv_debug_ignored_stype(ext->sType);
3961 break;
3962 }
3963 }
3964 }
3965
3966 void anv_GetImageSparseMemoryRequirements(
3967 VkDevice device,
3968 VkImage image,
3969 uint32_t* pSparseMemoryRequirementCount,
3970 VkSparseImageMemoryRequirements* pSparseMemoryRequirements)
3971 {
3972 *pSparseMemoryRequirementCount = 0;
3973 }
3974
3975 void anv_GetImageSparseMemoryRequirements2(
3976 VkDevice device,
3977 const VkImageSparseMemoryRequirementsInfo2* pInfo,
3978 uint32_t* pSparseMemoryRequirementCount,
3979 VkSparseImageMemoryRequirements2* pSparseMemoryRequirements)
3980 {
3981 *pSparseMemoryRequirementCount = 0;
3982 }
3983
3984 void anv_GetDeviceMemoryCommitment(
3985 VkDevice device,
3986 VkDeviceMemory memory,
3987 VkDeviceSize* pCommittedMemoryInBytes)
3988 {
3989 *pCommittedMemoryInBytes = 0;
3990 }
3991
3992 static void
3993 anv_bind_buffer_memory(const VkBindBufferMemoryInfo *pBindInfo)
3994 {
3995 ANV_FROM_HANDLE(anv_device_memory, mem, pBindInfo->memory);
3996 ANV_FROM_HANDLE(anv_buffer, buffer, pBindInfo->buffer);
3997
3998 assert(pBindInfo->sType == VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO);
3999
4000 if (mem) {
4001 buffer->address = (struct anv_address) {
4002 .bo = mem->bo,
4003 .offset = pBindInfo->memoryOffset,
4004 };
4005 } else {
4006 buffer->address = ANV_NULL_ADDRESS;
4007 }
4008 }
4009
4010 VkResult anv_BindBufferMemory(
4011 VkDevice device,
4012 VkBuffer buffer,
4013 VkDeviceMemory memory,
4014 VkDeviceSize memoryOffset)
4015 {
4016 anv_bind_buffer_memory(
4017 &(VkBindBufferMemoryInfo) {
4018 .sType = VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO,
4019 .buffer = buffer,
4020 .memory = memory,
4021 .memoryOffset = memoryOffset,
4022 });
4023
4024 return VK_SUCCESS;
4025 }
4026
4027 VkResult anv_BindBufferMemory2(
4028 VkDevice device,
4029 uint32_t bindInfoCount,
4030 const VkBindBufferMemoryInfo* pBindInfos)
4031 {
4032 for (uint32_t i = 0; i < bindInfoCount; i++)
4033 anv_bind_buffer_memory(&pBindInfos[i]);
4034
4035 return VK_SUCCESS;
4036 }
4037
4038 VkResult anv_QueueBindSparse(
4039 VkQueue _queue,
4040 uint32_t bindInfoCount,
4041 const VkBindSparseInfo* pBindInfo,
4042 VkFence fence)
4043 {
4044 ANV_FROM_HANDLE(anv_queue, queue, _queue);
4045 if (anv_device_is_lost(queue->device))
4046 return VK_ERROR_DEVICE_LOST;
4047
4048 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
4049 }
4050
4051 // Event functions
4052
4053 VkResult anv_CreateEvent(
4054 VkDevice _device,
4055 const VkEventCreateInfo* pCreateInfo,
4056 const VkAllocationCallbacks* pAllocator,
4057 VkEvent* pEvent)
4058 {
4059 ANV_FROM_HANDLE(anv_device, device, _device);
4060 struct anv_state state;
4061 struct anv_event *event;
4062
4063 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_EVENT_CREATE_INFO);
4064
4065 state = anv_state_pool_alloc(&device->dynamic_state_pool,
4066 sizeof(*event), 8);
4067 event = state.map;
4068 event->state = state;
4069 event->semaphore = VK_EVENT_RESET;
4070
4071 if (!device->info.has_llc) {
4072 /* Make sure the writes we're flushing have landed. */
4073 __builtin_ia32_mfence();
4074 __builtin_ia32_clflush(event);
4075 }
4076
4077 *pEvent = anv_event_to_handle(event);
4078
4079 return VK_SUCCESS;
4080 }
4081
4082 void anv_DestroyEvent(
4083 VkDevice _device,
4084 VkEvent _event,
4085 const VkAllocationCallbacks* pAllocator)
4086 {
4087 ANV_FROM_HANDLE(anv_device, device, _device);
4088 ANV_FROM_HANDLE(anv_event, event, _event);
4089
4090 if (!event)
4091 return;
4092
4093 anv_state_pool_free(&device->dynamic_state_pool, event->state);
4094 }
4095
4096 VkResult anv_GetEventStatus(
4097 VkDevice _device,
4098 VkEvent _event)
4099 {
4100 ANV_FROM_HANDLE(anv_device, device, _device);
4101 ANV_FROM_HANDLE(anv_event, event, _event);
4102
4103 if (anv_device_is_lost(device))
4104 return VK_ERROR_DEVICE_LOST;
4105
4106 if (!device->info.has_llc) {
4107 /* Invalidate read cache before reading event written by GPU. */
4108 __builtin_ia32_clflush(event);
4109 __builtin_ia32_mfence();
4110
4111 }
4112
4113 return event->semaphore;
4114 }
4115
4116 VkResult anv_SetEvent(
4117 VkDevice _device,
4118 VkEvent _event)
4119 {
4120 ANV_FROM_HANDLE(anv_device, device, _device);
4121 ANV_FROM_HANDLE(anv_event, event, _event);
4122
4123 event->semaphore = VK_EVENT_SET;
4124
4125 if (!device->info.has_llc) {
4126 /* Make sure the writes we're flushing have landed. */
4127 __builtin_ia32_mfence();
4128 __builtin_ia32_clflush(event);
4129 }
4130
4131 return VK_SUCCESS;
4132 }
4133
4134 VkResult anv_ResetEvent(
4135 VkDevice _device,
4136 VkEvent _event)
4137 {
4138 ANV_FROM_HANDLE(anv_device, device, _device);
4139 ANV_FROM_HANDLE(anv_event, event, _event);
4140
4141 event->semaphore = VK_EVENT_RESET;
4142
4143 if (!device->info.has_llc) {
4144 /* Make sure the writes we're flushing have landed. */
4145 __builtin_ia32_mfence();
4146 __builtin_ia32_clflush(event);
4147 }
4148
4149 return VK_SUCCESS;
4150 }
4151
4152 // Buffer functions
4153
4154 VkResult anv_CreateBuffer(
4155 VkDevice _device,
4156 const VkBufferCreateInfo* pCreateInfo,
4157 const VkAllocationCallbacks* pAllocator,
4158 VkBuffer* pBuffer)
4159 {
4160 ANV_FROM_HANDLE(anv_device, device, _device);
4161 struct anv_buffer *buffer;
4162
4163 /* Don't allow creating buffers bigger than our address space. The real
4164 * issue here is that we may align up the buffer size and we don't want
4165 * doing so to cause roll-over. However, no one has any business
4166 * allocating a buffer larger than our GTT size.
4167 */
4168 if (pCreateInfo->size > device->physical->gtt_size)
4169 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
4170
4171 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
4172
4173 buffer = vk_alloc2(&device->alloc, pAllocator, sizeof(*buffer), 8,
4174 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
4175 if (buffer == NULL)
4176 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
4177
4178 buffer->size = pCreateInfo->size;
4179 buffer->usage = pCreateInfo->usage;
4180 buffer->address = ANV_NULL_ADDRESS;
4181
4182 *pBuffer = anv_buffer_to_handle(buffer);
4183
4184 return VK_SUCCESS;
4185 }
4186
4187 void anv_DestroyBuffer(
4188 VkDevice _device,
4189 VkBuffer _buffer,
4190 const VkAllocationCallbacks* pAllocator)
4191 {
4192 ANV_FROM_HANDLE(anv_device, device, _device);
4193 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
4194
4195 if (!buffer)
4196 return;
4197
4198 vk_free2(&device->alloc, pAllocator, buffer);
4199 }
4200
4201 VkDeviceAddress anv_GetBufferDeviceAddress(
4202 VkDevice device,
4203 const VkBufferDeviceAddressInfoKHR* pInfo)
4204 {
4205 ANV_FROM_HANDLE(anv_buffer, buffer, pInfo->buffer);
4206
4207 assert(!anv_address_is_null(buffer->address));
4208 assert(buffer->address.bo->flags & EXEC_OBJECT_PINNED);
4209
4210 return anv_address_physical(buffer->address);
4211 }
4212
4213 uint64_t anv_GetBufferOpaqueCaptureAddress(
4214 VkDevice device,
4215 const VkBufferDeviceAddressInfoKHR* pInfo)
4216 {
4217 return 0;
4218 }
4219
4220 uint64_t anv_GetDeviceMemoryOpaqueCaptureAddress(
4221 VkDevice device,
4222 const VkDeviceMemoryOpaqueCaptureAddressInfoKHR* pInfo)
4223 {
4224 ANV_FROM_HANDLE(anv_device_memory, memory, pInfo->memory);
4225
4226 assert(memory->bo->flags & EXEC_OBJECT_PINNED);
4227 assert(memory->bo->has_client_visible_address);
4228
4229 return gen_48b_address(memory->bo->offset);
4230 }
4231
4232 void
4233 anv_fill_buffer_surface_state(struct anv_device *device, struct anv_state state,
4234 enum isl_format format,
4235 struct anv_address address,
4236 uint32_t range, uint32_t stride)
4237 {
4238 isl_buffer_fill_state(&device->isl_dev, state.map,
4239 .address = anv_address_physical(address),
4240 .mocs = device->isl_dev.mocs.internal,
4241 .size_B = range,
4242 .format = format,
4243 .swizzle = ISL_SWIZZLE_IDENTITY,
4244 .stride_B = stride);
4245 }
4246
4247 void anv_DestroySampler(
4248 VkDevice _device,
4249 VkSampler _sampler,
4250 const VkAllocationCallbacks* pAllocator)
4251 {
4252 ANV_FROM_HANDLE(anv_device, device, _device);
4253 ANV_FROM_HANDLE(anv_sampler, sampler, _sampler);
4254
4255 if (!sampler)
4256 return;
4257
4258 if (sampler->bindless_state.map) {
4259 anv_state_pool_free(&device->dynamic_state_pool,
4260 sampler->bindless_state);
4261 }
4262
4263 vk_free2(&device->alloc, pAllocator, sampler);
4264 }
4265
4266 VkResult anv_CreateFramebuffer(
4267 VkDevice _device,
4268 const VkFramebufferCreateInfo* pCreateInfo,
4269 const VkAllocationCallbacks* pAllocator,
4270 VkFramebuffer* pFramebuffer)
4271 {
4272 ANV_FROM_HANDLE(anv_device, device, _device);
4273 struct anv_framebuffer *framebuffer;
4274
4275 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
4276
4277 size_t size = sizeof(*framebuffer);
4278
4279 /* VK_KHR_imageless_framebuffer extension says:
4280 *
4281 * If flags includes VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR,
4282 * parameter pAttachments is ignored.
4283 */
4284 if (!(pCreateInfo->flags & VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR)) {
4285 size += sizeof(struct anv_image_view *) * pCreateInfo->attachmentCount;
4286 framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
4287 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
4288 if (framebuffer == NULL)
4289 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
4290
4291 for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
4292 ANV_FROM_HANDLE(anv_image_view, iview, pCreateInfo->pAttachments[i]);
4293 framebuffer->attachments[i] = iview;
4294 }
4295 framebuffer->attachment_count = pCreateInfo->attachmentCount;
4296 } else {
4297 assert(device->enabled_extensions.KHR_imageless_framebuffer);
4298 framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
4299 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
4300 if (framebuffer == NULL)
4301 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
4302
4303 framebuffer->attachment_count = 0;
4304 }
4305
4306 framebuffer->width = pCreateInfo->width;
4307 framebuffer->height = pCreateInfo->height;
4308 framebuffer->layers = pCreateInfo->layers;
4309
4310 *pFramebuffer = anv_framebuffer_to_handle(framebuffer);
4311
4312 return VK_SUCCESS;
4313 }
4314
4315 void anv_DestroyFramebuffer(
4316 VkDevice _device,
4317 VkFramebuffer _fb,
4318 const VkAllocationCallbacks* pAllocator)
4319 {
4320 ANV_FROM_HANDLE(anv_device, device, _device);
4321 ANV_FROM_HANDLE(anv_framebuffer, fb, _fb);
4322
4323 if (!fb)
4324 return;
4325
4326 vk_free2(&device->alloc, pAllocator, fb);
4327 }
4328
4329 static const VkTimeDomainEXT anv_time_domains[] = {
4330 VK_TIME_DOMAIN_DEVICE_EXT,
4331 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT,
4332 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT,
4333 };
4334
4335 VkResult anv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
4336 VkPhysicalDevice physicalDevice,
4337 uint32_t *pTimeDomainCount,
4338 VkTimeDomainEXT *pTimeDomains)
4339 {
4340 int d;
4341 VK_OUTARRAY_MAKE(out, pTimeDomains, pTimeDomainCount);
4342
4343 for (d = 0; d < ARRAY_SIZE(anv_time_domains); d++) {
4344 vk_outarray_append(&out, i) {
4345 *i = anv_time_domains[d];
4346 }
4347 }
4348
4349 return vk_outarray_status(&out);
4350 }
4351
4352 static uint64_t
4353 anv_clock_gettime(clockid_t clock_id)
4354 {
4355 struct timespec current;
4356 int ret;
4357
4358 ret = clock_gettime(clock_id, &current);
4359 if (ret < 0 && clock_id == CLOCK_MONOTONIC_RAW)
4360 ret = clock_gettime(CLOCK_MONOTONIC, &current);
4361 if (ret < 0)
4362 return 0;
4363
4364 return (uint64_t) current.tv_sec * 1000000000ULL + current.tv_nsec;
4365 }
4366
4367 #define TIMESTAMP 0x2358
4368
4369 VkResult anv_GetCalibratedTimestampsEXT(
4370 VkDevice _device,
4371 uint32_t timestampCount,
4372 const VkCalibratedTimestampInfoEXT *pTimestampInfos,
4373 uint64_t *pTimestamps,
4374 uint64_t *pMaxDeviation)
4375 {
4376 ANV_FROM_HANDLE(anv_device, device, _device);
4377 uint64_t timestamp_frequency = device->info.timestamp_frequency;
4378 int ret;
4379 int d;
4380 uint64_t begin, end;
4381 uint64_t max_clock_period = 0;
4382
4383 begin = anv_clock_gettime(CLOCK_MONOTONIC_RAW);
4384
4385 for (d = 0; d < timestampCount; d++) {
4386 switch (pTimestampInfos[d].timeDomain) {
4387 case VK_TIME_DOMAIN_DEVICE_EXT:
4388 ret = anv_gem_reg_read(device, TIMESTAMP | 1,
4389 &pTimestamps[d]);
4390
4391 if (ret != 0) {
4392 return anv_device_set_lost(device, "Failed to read the TIMESTAMP "
4393 "register: %m");
4394 }
4395 uint64_t device_period = DIV_ROUND_UP(1000000000, timestamp_frequency);
4396 max_clock_period = MAX2(max_clock_period, device_period);
4397 break;
4398 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT:
4399 pTimestamps[d] = anv_clock_gettime(CLOCK_MONOTONIC);
4400 max_clock_period = MAX2(max_clock_period, 1);
4401 break;
4402
4403 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT:
4404 pTimestamps[d] = begin;
4405 break;
4406 default:
4407 pTimestamps[d] = 0;
4408 break;
4409 }
4410 }
4411
4412 end = anv_clock_gettime(CLOCK_MONOTONIC_RAW);
4413
4414 /*
4415 * The maximum deviation is the sum of the interval over which we
4416 * perform the sampling and the maximum period of any sampled
4417 * clock. That's because the maximum skew between any two sampled
4418 * clock edges is when the sampled clock with the largest period is
4419 * sampled at the end of that period but right at the beginning of the
4420 * sampling interval and some other clock is sampled right at the
4421 * begining of its sampling period and right at the end of the
4422 * sampling interval. Let's assume the GPU has the longest clock
4423 * period and that the application is sampling GPU and monotonic:
4424 *
4425 * s e
4426 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
4427 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4428 *
4429 * g
4430 * 0 1 2 3
4431 * GPU -----_____-----_____-----_____-----_____
4432 *
4433 * m
4434 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
4435 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4436 *
4437 * Interval <----------------->
4438 * Deviation <-------------------------->
4439 *
4440 * s = read(raw) 2
4441 * g = read(GPU) 1
4442 * m = read(monotonic) 2
4443 * e = read(raw) b
4444 *
4445 * We round the sample interval up by one tick to cover sampling error
4446 * in the interval clock
4447 */
4448
4449 uint64_t sample_interval = end - begin + 1;
4450
4451 *pMaxDeviation = sample_interval + max_clock_period;
4452
4453 return VK_SUCCESS;
4454 }
4455
4456 /* vk_icd.h does not declare this function, so we declare it here to
4457 * suppress Wmissing-prototypes.
4458 */
4459 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4460 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion);
4461
4462 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4463 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion)
4464 {
4465 /* For the full details on loader interface versioning, see
4466 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
4467 * What follows is a condensed summary, to help you navigate the large and
4468 * confusing official doc.
4469 *
4470 * - Loader interface v0 is incompatible with later versions. We don't
4471 * support it.
4472 *
4473 * - In loader interface v1:
4474 * - The first ICD entrypoint called by the loader is
4475 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
4476 * entrypoint.
4477 * - The ICD must statically expose no other Vulkan symbol unless it is
4478 * linked with -Bsymbolic.
4479 * - Each dispatchable Vulkan handle created by the ICD must be
4480 * a pointer to a struct whose first member is VK_LOADER_DATA. The
4481 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
4482 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
4483 * vkDestroySurfaceKHR(). The ICD must be capable of working with
4484 * such loader-managed surfaces.
4485 *
4486 * - Loader interface v2 differs from v1 in:
4487 * - The first ICD entrypoint called by the loader is
4488 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
4489 * statically expose this entrypoint.
4490 *
4491 * - Loader interface v3 differs from v2 in:
4492 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
4493 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
4494 * because the loader no longer does so.
4495 *
4496 * - Loader interface v4 differs from v3 in:
4497 * - The ICD must implement vk_icdGetPhysicalDeviceProcAddr().
4498 */
4499 *pSupportedVersion = MIN2(*pSupportedVersion, 4u);
4500 return VK_SUCCESS;
4501 }