anv: Enable Vulkan 1.2 support
[mesa.git] / src / intel / vulkan / anv_device.c
1 /*
2 * Copyright © 2015 Intel Corporation
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 */
23
24 #include <assert.h>
25 #include <stdbool.h>
26 #include <string.h>
27 #include <sys/mman.h>
28 #include <sys/sysinfo.h>
29 #include <unistd.h>
30 #include <fcntl.h>
31 #include <xf86drm.h>
32 #include "drm-uapi/drm_fourcc.h"
33
34 #include "anv_private.h"
35 #include "util/debug.h"
36 #include "util/build_id.h"
37 #include "util/disk_cache.h"
38 #include "util/mesa-sha1.h"
39 #include "util/os_file.h"
40 #include "util/u_atomic.h"
41 #include "util/u_string.h"
42 #include "util/xmlpool.h"
43 #include "git_sha1.h"
44 #include "vk_util.h"
45 #include "common/gen_aux_map.h"
46 #include "common/gen_defines.h"
47 #include "compiler/glsl_types.h"
48
49 #include "genxml/gen7_pack.h"
50
51 static const char anv_dri_options_xml[] =
52 DRI_CONF_BEGIN
53 DRI_CONF_SECTION_PERFORMANCE
54 DRI_CONF_VK_X11_OVERRIDE_MIN_IMAGE_COUNT(0)
55 DRI_CONF_VK_X11_STRICT_IMAGE_COUNT("false")
56 DRI_CONF_SECTION_END
57
58 DRI_CONF_SECTION_DEBUG
59 DRI_CONF_ALWAYS_FLUSH_CACHE("false")
60 DRI_CONF_VK_WSI_FORCE_BGRA8_UNORM_FIRST("false")
61 DRI_CONF_SECTION_END
62 DRI_CONF_END;
63
64 /* This is probably far to big but it reflects the max size used for messages
65 * in OpenGLs KHR_debug.
66 */
67 #define MAX_DEBUG_MESSAGE_LENGTH 4096
68
69 static void
70 compiler_debug_log(void *data, const char *fmt, ...)
71 {
72 char str[MAX_DEBUG_MESSAGE_LENGTH];
73 struct anv_device *device = (struct anv_device *)data;
74
75 if (list_is_empty(&device->instance->debug_report_callbacks.callbacks))
76 return;
77
78 va_list args;
79 va_start(args, fmt);
80 (void) vsnprintf(str, MAX_DEBUG_MESSAGE_LENGTH, fmt, args);
81 va_end(args);
82
83 vk_debug_report(&device->instance->debug_report_callbacks,
84 VK_DEBUG_REPORT_DEBUG_BIT_EXT,
85 VK_DEBUG_REPORT_OBJECT_TYPE_UNKNOWN_EXT,
86 0, 0, 0, "anv", str);
87 }
88
89 static void
90 compiler_perf_log(void *data, const char *fmt, ...)
91 {
92 va_list args;
93 va_start(args, fmt);
94
95 if (unlikely(INTEL_DEBUG & DEBUG_PERF))
96 intel_logd_v(fmt, args);
97
98 va_end(args);
99 }
100
101 static uint64_t
102 anv_compute_heap_size(int fd, uint64_t gtt_size)
103 {
104 /* Query the total ram from the system */
105 struct sysinfo info;
106 sysinfo(&info);
107
108 uint64_t total_ram = (uint64_t)info.totalram * (uint64_t)info.mem_unit;
109
110 /* We don't want to burn too much ram with the GPU. If the user has 4GiB
111 * or less, we use at most half. If they have more than 4GiB, we use 3/4.
112 */
113 uint64_t available_ram;
114 if (total_ram <= 4ull * 1024ull * 1024ull * 1024ull)
115 available_ram = total_ram / 2;
116 else
117 available_ram = total_ram * 3 / 4;
118
119 /* We also want to leave some padding for things we allocate in the driver,
120 * so don't go over 3/4 of the GTT either.
121 */
122 uint64_t available_gtt = gtt_size * 3 / 4;
123
124 return MIN2(available_ram, available_gtt);
125 }
126
127 static VkResult
128 anv_physical_device_init_heaps(struct anv_physical_device *device, int fd)
129 {
130 if (anv_gem_get_context_param(fd, 0, I915_CONTEXT_PARAM_GTT_SIZE,
131 &device->gtt_size) == -1) {
132 /* If, for whatever reason, we can't actually get the GTT size from the
133 * kernel (too old?) fall back to the aperture size.
134 */
135 anv_perf_warn(NULL, NULL,
136 "Failed to get I915_CONTEXT_PARAM_GTT_SIZE: %m");
137
138 if (anv_gem_get_aperture(fd, &device->gtt_size) == -1) {
139 return vk_errorf(NULL, NULL, VK_ERROR_INITIALIZATION_FAILED,
140 "failed to get aperture size: %m");
141 }
142 }
143
144 /* We only allow 48-bit addresses with softpin because knowing the actual
145 * address is required for the vertex cache flush workaround.
146 */
147 device->supports_48bit_addresses = (device->info.gen >= 8) &&
148 device->has_softpin &&
149 device->gtt_size > (4ULL << 30 /* GiB */);
150
151 uint64_t heap_size = anv_compute_heap_size(fd, device->gtt_size);
152
153 if (heap_size > (2ull << 30) && !device->supports_48bit_addresses) {
154 /* When running with an overridden PCI ID, we may get a GTT size from
155 * the kernel that is greater than 2 GiB but the execbuf check for 48bit
156 * address support can still fail. Just clamp the address space size to
157 * 2 GiB if we don't have 48-bit support.
158 */
159 intel_logw("%s:%d: The kernel reported a GTT size larger than 2 GiB but "
160 "not support for 48-bit addresses",
161 __FILE__, __LINE__);
162 heap_size = 2ull << 30;
163 }
164
165 device->memory.heap_count = 1;
166 device->memory.heaps[0] = (struct anv_memory_heap) {
167 .size = heap_size,
168 .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
169 };
170
171 uint32_t type_count = 0;
172 for (uint32_t heap = 0; heap < device->memory.heap_count; heap++) {
173 if (device->info.has_llc) {
174 /* Big core GPUs share LLC with the CPU and thus one memory type can be
175 * both cached and coherent at the same time.
176 */
177 device->memory.types[type_count++] = (struct anv_memory_type) {
178 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
179 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
180 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
181 VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
182 .heapIndex = heap,
183 };
184 } else {
185 /* The spec requires that we expose a host-visible, coherent memory
186 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
187 * to give the application a choice between cached, but not coherent and
188 * coherent but uncached (WC though).
189 */
190 device->memory.types[type_count++] = (struct anv_memory_type) {
191 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
192 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
193 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
194 .heapIndex = heap,
195 };
196 device->memory.types[type_count++] = (struct anv_memory_type) {
197 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
198 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
199 VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
200 .heapIndex = heap,
201 };
202 }
203 }
204 device->memory.type_count = type_count;
205
206 return VK_SUCCESS;
207 }
208
209 static VkResult
210 anv_physical_device_init_uuids(struct anv_physical_device *device)
211 {
212 const struct build_id_note *note =
213 build_id_find_nhdr_for_addr(anv_physical_device_init_uuids);
214 if (!note) {
215 return vk_errorf(device->instance, device,
216 VK_ERROR_INITIALIZATION_FAILED,
217 "Failed to find build-id");
218 }
219
220 unsigned build_id_len = build_id_length(note);
221 if (build_id_len < 20) {
222 return vk_errorf(device->instance, device,
223 VK_ERROR_INITIALIZATION_FAILED,
224 "build-id too short. It needs to be a SHA");
225 }
226
227 memcpy(device->driver_build_sha1, build_id_data(note), 20);
228
229 struct mesa_sha1 sha1_ctx;
230 uint8_t sha1[20];
231 STATIC_ASSERT(VK_UUID_SIZE <= sizeof(sha1));
232
233 /* The pipeline cache UUID is used for determining when a pipeline cache is
234 * invalid. It needs both a driver build and the PCI ID of the device.
235 */
236 _mesa_sha1_init(&sha1_ctx);
237 _mesa_sha1_update(&sha1_ctx, build_id_data(note), build_id_len);
238 _mesa_sha1_update(&sha1_ctx, &device->chipset_id,
239 sizeof(device->chipset_id));
240 _mesa_sha1_update(&sha1_ctx, &device->always_use_bindless,
241 sizeof(device->always_use_bindless));
242 _mesa_sha1_update(&sha1_ctx, &device->has_a64_buffer_access,
243 sizeof(device->has_a64_buffer_access));
244 _mesa_sha1_update(&sha1_ctx, &device->has_bindless_images,
245 sizeof(device->has_bindless_images));
246 _mesa_sha1_update(&sha1_ctx, &device->has_bindless_samplers,
247 sizeof(device->has_bindless_samplers));
248 _mesa_sha1_final(&sha1_ctx, sha1);
249 memcpy(device->pipeline_cache_uuid, sha1, VK_UUID_SIZE);
250
251 /* The driver UUID is used for determining sharability of images and memory
252 * between two Vulkan instances in separate processes. People who want to
253 * share memory need to also check the device UUID (below) so all this
254 * needs to be is the build-id.
255 */
256 memcpy(device->driver_uuid, build_id_data(note), VK_UUID_SIZE);
257
258 /* The device UUID uniquely identifies the given device within the machine.
259 * Since we never have more than one device, this doesn't need to be a real
260 * UUID. However, on the off-chance that someone tries to use this to
261 * cache pre-tiled images or something of the like, we use the PCI ID and
262 * some bits of ISL info to ensure that this is safe.
263 */
264 _mesa_sha1_init(&sha1_ctx);
265 _mesa_sha1_update(&sha1_ctx, &device->chipset_id,
266 sizeof(device->chipset_id));
267 _mesa_sha1_update(&sha1_ctx, &device->isl_dev.has_bit6_swizzling,
268 sizeof(device->isl_dev.has_bit6_swizzling));
269 _mesa_sha1_final(&sha1_ctx, sha1);
270 memcpy(device->device_uuid, sha1, VK_UUID_SIZE);
271
272 return VK_SUCCESS;
273 }
274
275 static void
276 anv_physical_device_init_disk_cache(struct anv_physical_device *device)
277 {
278 #ifdef ENABLE_SHADER_CACHE
279 char renderer[10];
280 ASSERTED int len = snprintf(renderer, sizeof(renderer), "anv_%04x",
281 device->chipset_id);
282 assert(len == sizeof(renderer) - 2);
283
284 char timestamp[41];
285 _mesa_sha1_format(timestamp, device->driver_build_sha1);
286
287 const uint64_t driver_flags =
288 brw_get_compiler_config_value(device->compiler);
289 device->disk_cache = disk_cache_create(renderer, timestamp, driver_flags);
290 #else
291 device->disk_cache = NULL;
292 #endif
293 }
294
295 static void
296 anv_physical_device_free_disk_cache(struct anv_physical_device *device)
297 {
298 #ifdef ENABLE_SHADER_CACHE
299 if (device->disk_cache)
300 disk_cache_destroy(device->disk_cache);
301 #else
302 assert(device->disk_cache == NULL);
303 #endif
304 }
305
306 static uint64_t
307 get_available_system_memory()
308 {
309 char *meminfo = os_read_file("/proc/meminfo");
310 if (!meminfo)
311 return 0;
312
313 char *str = strstr(meminfo, "MemAvailable:");
314 if (!str) {
315 free(meminfo);
316 return 0;
317 }
318
319 uint64_t kb_mem_available;
320 if (sscanf(str, "MemAvailable: %" PRIx64, &kb_mem_available) == 1) {
321 free(meminfo);
322 return kb_mem_available << 10;
323 }
324
325 free(meminfo);
326 return 0;
327 }
328
329 static VkResult
330 anv_physical_device_init(struct anv_physical_device *device,
331 struct anv_instance *instance,
332 drmDevicePtr drm_device)
333 {
334 const char *primary_path = drm_device->nodes[DRM_NODE_PRIMARY];
335 const char *path = drm_device->nodes[DRM_NODE_RENDER];
336 VkResult result;
337 int fd;
338 int master_fd = -1;
339
340 brw_process_intel_debug_variable();
341
342 fd = open(path, O_RDWR | O_CLOEXEC);
343 if (fd < 0)
344 return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
345
346 device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
347 device->instance = instance;
348
349 assert(strlen(path) < ARRAY_SIZE(device->path));
350 snprintf(device->path, ARRAY_SIZE(device->path), "%s", path);
351
352 if (!gen_get_device_info_from_fd(fd, &device->info)) {
353 result = vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
354 goto fail;
355 }
356 device->chipset_id = device->info.chipset_id;
357 device->no_hw = device->info.no_hw;
358
359 if (getenv("INTEL_NO_HW") != NULL)
360 device->no_hw = true;
361
362 device->pci_info.domain = drm_device->businfo.pci->domain;
363 device->pci_info.bus = drm_device->businfo.pci->bus;
364 device->pci_info.device = drm_device->businfo.pci->dev;
365 device->pci_info.function = drm_device->businfo.pci->func;
366
367 device->name = gen_get_device_name(device->chipset_id);
368
369 if (device->info.is_haswell) {
370 intel_logw("Haswell Vulkan support is incomplete");
371 } else if (device->info.gen == 7 && !device->info.is_baytrail) {
372 intel_logw("Ivy Bridge Vulkan support is incomplete");
373 } else if (device->info.gen == 7 && device->info.is_baytrail) {
374 intel_logw("Bay Trail Vulkan support is incomplete");
375 } else if (device->info.gen >= 8 && device->info.gen <= 11) {
376 /* Gen8-11 fully supported */
377 } else if (device->info.gen == 12) {
378 intel_logw("Vulkan is not yet fully supported on gen12");
379 } else {
380 result = vk_errorf(device->instance, device,
381 VK_ERROR_INCOMPATIBLE_DRIVER,
382 "Vulkan not yet supported on %s", device->name);
383 goto fail;
384 }
385
386 device->cmd_parser_version = -1;
387 if (device->info.gen == 7) {
388 device->cmd_parser_version =
389 anv_gem_get_param(fd, I915_PARAM_CMD_PARSER_VERSION);
390 if (device->cmd_parser_version == -1) {
391 result = vk_errorf(device->instance, device,
392 VK_ERROR_INITIALIZATION_FAILED,
393 "failed to get command parser version");
394 goto fail;
395 }
396 }
397
398 if (!anv_gem_get_param(fd, I915_PARAM_HAS_WAIT_TIMEOUT)) {
399 result = vk_errorf(device->instance, device,
400 VK_ERROR_INITIALIZATION_FAILED,
401 "kernel missing gem wait");
402 goto fail;
403 }
404
405 if (!anv_gem_get_param(fd, I915_PARAM_HAS_EXECBUF2)) {
406 result = vk_errorf(device->instance, device,
407 VK_ERROR_INITIALIZATION_FAILED,
408 "kernel missing execbuf2");
409 goto fail;
410 }
411
412 if (!device->info.has_llc &&
413 anv_gem_get_param(fd, I915_PARAM_MMAP_VERSION) < 1) {
414 result = vk_errorf(device->instance, device,
415 VK_ERROR_INITIALIZATION_FAILED,
416 "kernel missing wc mmap");
417 goto fail;
418 }
419
420 device->has_softpin = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_SOFTPIN);
421 device->has_exec_async = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_ASYNC);
422 device->has_exec_capture = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_CAPTURE);
423 device->has_exec_fence = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_FENCE);
424 device->has_syncobj = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_FENCE_ARRAY);
425 device->has_syncobj_wait = device->has_syncobj &&
426 anv_gem_supports_syncobj_wait(fd);
427 device->has_context_priority = anv_gem_has_context_priority(fd);
428
429 result = anv_physical_device_init_heaps(device, fd);
430 if (result != VK_SUCCESS)
431 goto fail;
432
433 device->use_softpin = device->has_softpin &&
434 device->supports_48bit_addresses;
435
436 device->has_context_isolation =
437 anv_gem_get_param(fd, I915_PARAM_HAS_CONTEXT_ISOLATION);
438
439 device->always_use_bindless =
440 env_var_as_boolean("ANV_ALWAYS_BINDLESS", false);
441
442 /* We first got the A64 messages on broadwell and we can only use them if
443 * we can pass addresses directly into the shader which requires softpin.
444 */
445 device->has_a64_buffer_access = device->info.gen >= 8 &&
446 device->use_softpin;
447
448 /* We first get bindless image access on Skylake and we can only really do
449 * it if we don't have any relocations so we need softpin.
450 */
451 device->has_bindless_images = device->info.gen >= 9 &&
452 device->use_softpin;
453
454 /* We've had bindless samplers since Ivy Bridge (forever in Vulkan terms)
455 * because it's just a matter of setting the sampler address in the sample
456 * message header. However, we've not bothered to wire it up for vec4 so
457 * we leave it disabled on gen7.
458 */
459 device->has_bindless_samplers = device->info.gen >= 8;
460
461 device->has_mem_available = get_available_system_memory() != 0;
462
463 device->always_flush_cache =
464 driQueryOptionb(&instance->dri_options, "always_flush_cache");
465
466 /* Starting with Gen10, the timestamp frequency of the command streamer may
467 * vary from one part to another. We can query the value from the kernel.
468 */
469 if (device->info.gen >= 10) {
470 int timestamp_frequency =
471 anv_gem_get_param(fd, I915_PARAM_CS_TIMESTAMP_FREQUENCY);
472
473 if (timestamp_frequency < 0)
474 intel_logw("Kernel 4.16-rc1+ required to properly query CS timestamp frequency");
475 else
476 device->info.timestamp_frequency = timestamp_frequency;
477 }
478
479 /* GENs prior to 8 do not support EU/Subslice info */
480 if (device->info.gen >= 8) {
481 device->subslice_total = anv_gem_get_param(fd, I915_PARAM_SUBSLICE_TOTAL);
482 device->eu_total = anv_gem_get_param(fd, I915_PARAM_EU_TOTAL);
483
484 /* Without this information, we cannot get the right Braswell
485 * brandstrings, and we have to use conservative numbers for GPGPU on
486 * many platforms, but otherwise, things will just work.
487 */
488 if (device->subslice_total < 1 || device->eu_total < 1) {
489 intel_logw("Kernel 4.1 required to properly query GPU properties");
490 }
491 } else if (device->info.gen == 7) {
492 device->subslice_total = 1 << (device->info.gt - 1);
493 }
494
495 if (device->info.is_cherryview &&
496 device->subslice_total > 0 && device->eu_total > 0) {
497 /* Logical CS threads = EUs per subslice * num threads per EU */
498 uint32_t max_cs_threads =
499 device->eu_total / device->subslice_total * device->info.num_thread_per_eu;
500
501 /* Fuse configurations may give more threads than expected, never less. */
502 if (max_cs_threads > device->info.max_cs_threads)
503 device->info.max_cs_threads = max_cs_threads;
504 }
505
506 device->compiler = brw_compiler_create(NULL, &device->info);
507 if (device->compiler == NULL) {
508 result = vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
509 goto fail;
510 }
511 device->compiler->shader_debug_log = compiler_debug_log;
512 device->compiler->shader_perf_log = compiler_perf_log;
513 device->compiler->supports_pull_constants = false;
514 device->compiler->constant_buffer_0_is_relative =
515 device->info.gen < 8 || !device->has_context_isolation;
516 device->compiler->supports_shader_constants = true;
517 device->compiler->compact_params = false;
518
519 /* Broadwell PRM says:
520 *
521 * "Before Gen8, there was a historical configuration control field to
522 * swizzle address bit[6] for in X/Y tiling modes. This was set in three
523 * different places: TILECTL[1:0], ARB_MODE[5:4], and
524 * DISP_ARB_CTL[14:13].
525 *
526 * For Gen8 and subsequent generations, the swizzle fields are all
527 * reserved, and the CPU's memory controller performs all address
528 * swizzling modifications."
529 */
530 bool swizzled =
531 device->info.gen < 8 && anv_gem_get_bit6_swizzle(fd, I915_TILING_X);
532
533 isl_device_init(&device->isl_dev, &device->info, swizzled);
534
535 result = anv_physical_device_init_uuids(device);
536 if (result != VK_SUCCESS)
537 goto fail;
538
539 anv_physical_device_init_disk_cache(device);
540
541 if (instance->enabled_extensions.KHR_display) {
542 master_fd = open(primary_path, O_RDWR | O_CLOEXEC);
543 if (master_fd >= 0) {
544 /* prod the device with a GETPARAM call which will fail if
545 * we don't have permission to even render on this device
546 */
547 if (anv_gem_get_param(master_fd, I915_PARAM_CHIPSET_ID) == 0) {
548 close(master_fd);
549 master_fd = -1;
550 }
551 }
552 }
553 device->master_fd = master_fd;
554
555 result = anv_init_wsi(device);
556 if (result != VK_SUCCESS) {
557 ralloc_free(device->compiler);
558 anv_physical_device_free_disk_cache(device);
559 goto fail;
560 }
561
562 device->perf = anv_get_perf(&device->info, fd);
563
564 anv_physical_device_get_supported_extensions(device,
565 &device->supported_extensions);
566
567
568 device->local_fd = fd;
569
570 return VK_SUCCESS;
571
572 fail:
573 close(fd);
574 if (master_fd != -1)
575 close(master_fd);
576 return result;
577 }
578
579 static void
580 anv_physical_device_finish(struct anv_physical_device *device)
581 {
582 anv_finish_wsi(device);
583 anv_physical_device_free_disk_cache(device);
584 ralloc_free(device->compiler);
585 ralloc_free(device->perf);
586 close(device->local_fd);
587 if (device->master_fd >= 0)
588 close(device->master_fd);
589 }
590
591 static void *
592 default_alloc_func(void *pUserData, size_t size, size_t align,
593 VkSystemAllocationScope allocationScope)
594 {
595 return malloc(size);
596 }
597
598 static void *
599 default_realloc_func(void *pUserData, void *pOriginal, size_t size,
600 size_t align, VkSystemAllocationScope allocationScope)
601 {
602 return realloc(pOriginal, size);
603 }
604
605 static void
606 default_free_func(void *pUserData, void *pMemory)
607 {
608 free(pMemory);
609 }
610
611 static const VkAllocationCallbacks default_alloc = {
612 .pUserData = NULL,
613 .pfnAllocation = default_alloc_func,
614 .pfnReallocation = default_realloc_func,
615 .pfnFree = default_free_func,
616 };
617
618 VkResult anv_EnumerateInstanceExtensionProperties(
619 const char* pLayerName,
620 uint32_t* pPropertyCount,
621 VkExtensionProperties* pProperties)
622 {
623 VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
624
625 for (int i = 0; i < ANV_INSTANCE_EXTENSION_COUNT; i++) {
626 if (anv_instance_extensions_supported.extensions[i]) {
627 vk_outarray_append(&out, prop) {
628 *prop = anv_instance_extensions[i];
629 }
630 }
631 }
632
633 return vk_outarray_status(&out);
634 }
635
636 VkResult anv_CreateInstance(
637 const VkInstanceCreateInfo* pCreateInfo,
638 const VkAllocationCallbacks* pAllocator,
639 VkInstance* pInstance)
640 {
641 struct anv_instance *instance;
642 VkResult result;
643
644 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO);
645
646 struct anv_instance_extension_table enabled_extensions = {};
647 for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
648 int idx;
649 for (idx = 0; idx < ANV_INSTANCE_EXTENSION_COUNT; idx++) {
650 if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
651 anv_instance_extensions[idx].extensionName) == 0)
652 break;
653 }
654
655 if (idx >= ANV_INSTANCE_EXTENSION_COUNT)
656 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
657
658 if (!anv_instance_extensions_supported.extensions[idx])
659 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
660
661 enabled_extensions.extensions[idx] = true;
662 }
663
664 instance = vk_alloc2(&default_alloc, pAllocator, sizeof(*instance), 8,
665 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
666 if (!instance)
667 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
668
669 instance->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
670
671 if (pAllocator)
672 instance->alloc = *pAllocator;
673 else
674 instance->alloc = default_alloc;
675
676 instance->app_info = (struct anv_app_info) { .api_version = 0 };
677 if (pCreateInfo->pApplicationInfo) {
678 const VkApplicationInfo *app = pCreateInfo->pApplicationInfo;
679
680 instance->app_info.app_name =
681 vk_strdup(&instance->alloc, app->pApplicationName,
682 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
683 instance->app_info.app_version = app->applicationVersion;
684
685 instance->app_info.engine_name =
686 vk_strdup(&instance->alloc, app->pEngineName,
687 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
688 instance->app_info.engine_version = app->engineVersion;
689
690 instance->app_info.api_version = app->apiVersion;
691 }
692
693 if (instance->app_info.api_version == 0)
694 instance->app_info.api_version = VK_API_VERSION_1_0;
695
696 instance->enabled_extensions = enabled_extensions;
697
698 for (unsigned i = 0; i < ARRAY_SIZE(instance->dispatch.entrypoints); i++) {
699 /* Vulkan requires that entrypoints for extensions which have not been
700 * enabled must not be advertised.
701 */
702 if (!anv_instance_entrypoint_is_enabled(i, instance->app_info.api_version,
703 &instance->enabled_extensions)) {
704 instance->dispatch.entrypoints[i] = NULL;
705 } else {
706 instance->dispatch.entrypoints[i] =
707 anv_instance_dispatch_table.entrypoints[i];
708 }
709 }
710
711 struct anv_physical_device *pdevice = &instance->physicalDevice;
712 for (unsigned i = 0; i < ARRAY_SIZE(pdevice->dispatch.entrypoints); i++) {
713 /* Vulkan requires that entrypoints for extensions which have not been
714 * enabled must not be advertised.
715 */
716 if (!anv_physical_device_entrypoint_is_enabled(i, instance->app_info.api_version,
717 &instance->enabled_extensions)) {
718 pdevice->dispatch.entrypoints[i] = NULL;
719 } else {
720 pdevice->dispatch.entrypoints[i] =
721 anv_physical_device_dispatch_table.entrypoints[i];
722 }
723 }
724
725 for (unsigned i = 0; i < ARRAY_SIZE(instance->device_dispatch.entrypoints); i++) {
726 /* Vulkan requires that entrypoints for extensions which have not been
727 * enabled must not be advertised.
728 */
729 if (!anv_device_entrypoint_is_enabled(i, instance->app_info.api_version,
730 &instance->enabled_extensions, NULL)) {
731 instance->device_dispatch.entrypoints[i] = NULL;
732 } else {
733 instance->device_dispatch.entrypoints[i] =
734 anv_device_dispatch_table.entrypoints[i];
735 }
736 }
737
738 instance->physicalDeviceCount = -1;
739
740 result = vk_debug_report_instance_init(&instance->debug_report_callbacks);
741 if (result != VK_SUCCESS) {
742 vk_free2(&default_alloc, pAllocator, instance);
743 return vk_error(result);
744 }
745
746 instance->pipeline_cache_enabled =
747 env_var_as_boolean("ANV_ENABLE_PIPELINE_CACHE", true);
748
749 glsl_type_singleton_init_or_ref();
750
751 VG(VALGRIND_CREATE_MEMPOOL(instance, 0, false));
752
753 driParseOptionInfo(&instance->available_dri_options, anv_dri_options_xml);
754 driParseConfigFiles(&instance->dri_options, &instance->available_dri_options,
755 0, "anv", NULL,
756 instance->app_info.engine_name,
757 instance->app_info.engine_version);
758
759 *pInstance = anv_instance_to_handle(instance);
760
761 return VK_SUCCESS;
762 }
763
764 void anv_DestroyInstance(
765 VkInstance _instance,
766 const VkAllocationCallbacks* pAllocator)
767 {
768 ANV_FROM_HANDLE(anv_instance, instance, _instance);
769
770 if (!instance)
771 return;
772
773 if (instance->physicalDeviceCount > 0) {
774 /* We support at most one physical device. */
775 assert(instance->physicalDeviceCount == 1);
776 anv_physical_device_finish(&instance->physicalDevice);
777 }
778
779 vk_free(&instance->alloc, (char *)instance->app_info.app_name);
780 vk_free(&instance->alloc, (char *)instance->app_info.engine_name);
781
782 VG(VALGRIND_DESTROY_MEMPOOL(instance));
783
784 vk_debug_report_instance_destroy(&instance->debug_report_callbacks);
785
786 glsl_type_singleton_decref();
787
788 driDestroyOptionCache(&instance->dri_options);
789 driDestroyOptionInfo(&instance->available_dri_options);
790
791 vk_free(&instance->alloc, instance);
792 }
793
794 static VkResult
795 anv_enumerate_devices(struct anv_instance *instance)
796 {
797 /* TODO: Check for more devices ? */
798 drmDevicePtr devices[8];
799 VkResult result = VK_ERROR_INCOMPATIBLE_DRIVER;
800 int max_devices;
801
802 instance->physicalDeviceCount = 0;
803
804 max_devices = drmGetDevices2(0, devices, ARRAY_SIZE(devices));
805 if (max_devices < 1)
806 return VK_ERROR_INCOMPATIBLE_DRIVER;
807
808 for (unsigned i = 0; i < (unsigned)max_devices; i++) {
809 if (devices[i]->available_nodes & 1 << DRM_NODE_RENDER &&
810 devices[i]->bustype == DRM_BUS_PCI &&
811 devices[i]->deviceinfo.pci->vendor_id == 0x8086) {
812
813 result = anv_physical_device_init(&instance->physicalDevice,
814 instance, devices[i]);
815 if (result != VK_ERROR_INCOMPATIBLE_DRIVER)
816 break;
817 }
818 }
819 drmFreeDevices(devices, max_devices);
820
821 if (result == VK_SUCCESS)
822 instance->physicalDeviceCount = 1;
823
824 return result;
825 }
826
827 static VkResult
828 anv_instance_ensure_physical_device(struct anv_instance *instance)
829 {
830 if (instance->physicalDeviceCount < 0) {
831 VkResult result = anv_enumerate_devices(instance);
832 if (result != VK_SUCCESS &&
833 result != VK_ERROR_INCOMPATIBLE_DRIVER)
834 return result;
835 }
836
837 return VK_SUCCESS;
838 }
839
840 VkResult anv_EnumeratePhysicalDevices(
841 VkInstance _instance,
842 uint32_t* pPhysicalDeviceCount,
843 VkPhysicalDevice* pPhysicalDevices)
844 {
845 ANV_FROM_HANDLE(anv_instance, instance, _instance);
846 VK_OUTARRAY_MAKE(out, pPhysicalDevices, pPhysicalDeviceCount);
847
848 VkResult result = anv_instance_ensure_physical_device(instance);
849 if (result != VK_SUCCESS)
850 return result;
851
852 if (instance->physicalDeviceCount == 0)
853 return VK_SUCCESS;
854
855 assert(instance->physicalDeviceCount == 1);
856 vk_outarray_append(&out, i) {
857 *i = anv_physical_device_to_handle(&instance->physicalDevice);
858 }
859
860 return vk_outarray_status(&out);
861 }
862
863 VkResult anv_EnumeratePhysicalDeviceGroups(
864 VkInstance _instance,
865 uint32_t* pPhysicalDeviceGroupCount,
866 VkPhysicalDeviceGroupProperties* pPhysicalDeviceGroupProperties)
867 {
868 ANV_FROM_HANDLE(anv_instance, instance, _instance);
869 VK_OUTARRAY_MAKE(out, pPhysicalDeviceGroupProperties,
870 pPhysicalDeviceGroupCount);
871
872 VkResult result = anv_instance_ensure_physical_device(instance);
873 if (result != VK_SUCCESS)
874 return result;
875
876 if (instance->physicalDeviceCount == 0)
877 return VK_SUCCESS;
878
879 assert(instance->physicalDeviceCount == 1);
880
881 vk_outarray_append(&out, p) {
882 p->physicalDeviceCount = 1;
883 memset(p->physicalDevices, 0, sizeof(p->physicalDevices));
884 p->physicalDevices[0] =
885 anv_physical_device_to_handle(&instance->physicalDevice);
886 p->subsetAllocation = false;
887
888 vk_foreach_struct(ext, p->pNext)
889 anv_debug_ignored_stype(ext->sType);
890 }
891
892 return vk_outarray_status(&out);
893 }
894
895 void anv_GetPhysicalDeviceFeatures(
896 VkPhysicalDevice physicalDevice,
897 VkPhysicalDeviceFeatures* pFeatures)
898 {
899 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
900
901 *pFeatures = (VkPhysicalDeviceFeatures) {
902 .robustBufferAccess = true,
903 .fullDrawIndexUint32 = true,
904 .imageCubeArray = true,
905 .independentBlend = true,
906 .geometryShader = true,
907 .tessellationShader = true,
908 .sampleRateShading = true,
909 .dualSrcBlend = true,
910 .logicOp = true,
911 .multiDrawIndirect = true,
912 .drawIndirectFirstInstance = true,
913 .depthClamp = true,
914 .depthBiasClamp = true,
915 .fillModeNonSolid = true,
916 .depthBounds = pdevice->info.gen >= 12,
917 .wideLines = true,
918 .largePoints = true,
919 .alphaToOne = true,
920 .multiViewport = true,
921 .samplerAnisotropy = true,
922 .textureCompressionETC2 = pdevice->info.gen >= 8 ||
923 pdevice->info.is_baytrail,
924 .textureCompressionASTC_LDR = pdevice->info.gen >= 9, /* FINISHME CHV */
925 .textureCompressionBC = true,
926 .occlusionQueryPrecise = true,
927 .pipelineStatisticsQuery = true,
928 .fragmentStoresAndAtomics = true,
929 .shaderTessellationAndGeometryPointSize = true,
930 .shaderImageGatherExtended = true,
931 .shaderStorageImageExtendedFormats = true,
932 .shaderStorageImageMultisample = false,
933 .shaderStorageImageReadWithoutFormat = false,
934 .shaderStorageImageWriteWithoutFormat = true,
935 .shaderUniformBufferArrayDynamicIndexing = true,
936 .shaderSampledImageArrayDynamicIndexing = true,
937 .shaderStorageBufferArrayDynamicIndexing = true,
938 .shaderStorageImageArrayDynamicIndexing = true,
939 .shaderClipDistance = true,
940 .shaderCullDistance = true,
941 .shaderFloat64 = pdevice->info.gen >= 8 &&
942 pdevice->info.has_64bit_types,
943 .shaderInt64 = pdevice->info.gen >= 8 &&
944 pdevice->info.has_64bit_types,
945 .shaderInt16 = pdevice->info.gen >= 8,
946 .shaderResourceMinLod = pdevice->info.gen >= 9,
947 .variableMultisampleRate = true,
948 .inheritedQueries = true,
949 };
950
951 /* We can't do image stores in vec4 shaders */
952 pFeatures->vertexPipelineStoresAndAtomics =
953 pdevice->compiler->scalar_stage[MESA_SHADER_VERTEX] &&
954 pdevice->compiler->scalar_stage[MESA_SHADER_GEOMETRY];
955
956 struct anv_app_info *app_info = &pdevice->instance->app_info;
957
958 /* The new DOOM and Wolfenstein games require depthBounds without
959 * checking for it. They seem to run fine without it so just claim it's
960 * there and accept the consequences.
961 */
962 if (app_info->engine_name && strcmp(app_info->engine_name, "idTech") == 0)
963 pFeatures->depthBounds = true;
964 }
965
966 static void
967 anv_get_physical_device_features_1_1(struct anv_physical_device *pdevice,
968 VkPhysicalDeviceVulkan11Features *f)
969 {
970 assert(f->sType == VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES);
971
972 f->storageBuffer16BitAccess = pdevice->info.gen >= 8;
973 f->uniformAndStorageBuffer16BitAccess = pdevice->info.gen >= 8;
974 f->storagePushConstant16 = pdevice->info.gen >= 8;
975 f->storageInputOutput16 = false;
976 f->multiview = true;
977 f->multiviewGeometryShader = true;
978 f->multiviewTessellationShader = true;
979 f->variablePointersStorageBuffer = true;
980 f->variablePointers = true;
981 f->protectedMemory = false;
982 f->samplerYcbcrConversion = true;
983 f->shaderDrawParameters = true;
984 }
985
986 static void
987 anv_get_physical_device_features_1_2(struct anv_physical_device *pdevice,
988 VkPhysicalDeviceVulkan12Features *f)
989 {
990 assert(f->sType == VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES);
991
992 f->samplerMirrorClampToEdge = true;
993 f->drawIndirectCount = true;
994 f->storageBuffer8BitAccess = pdevice->info.gen >= 8;
995 f->uniformAndStorageBuffer8BitAccess = pdevice->info.gen >= 8;
996 f->storagePushConstant8 = pdevice->info.gen >= 8;
997 f->shaderBufferInt64Atomics = pdevice->info.gen >= 9 &&
998 pdevice->use_softpin;
999 f->shaderSharedInt64Atomics = false;
1000 f->shaderFloat16 = pdevice->info.gen >= 8;
1001 f->shaderInt8 = pdevice->info.gen >= 8;
1002
1003 bool descIndexing = pdevice->has_a64_buffer_access &&
1004 pdevice->has_bindless_images;
1005 f->descriptorIndexing = descIndexing;
1006 f->shaderInputAttachmentArrayDynamicIndexing = false;
1007 f->shaderUniformTexelBufferArrayDynamicIndexing = descIndexing;
1008 f->shaderStorageTexelBufferArrayDynamicIndexing = descIndexing;
1009 f->shaderUniformBufferArrayNonUniformIndexing = false;
1010 f->shaderSampledImageArrayNonUniformIndexing = descIndexing;
1011 f->shaderStorageBufferArrayNonUniformIndexing = descIndexing;
1012 f->shaderStorageImageArrayNonUniformIndexing = descIndexing;
1013 f->shaderInputAttachmentArrayNonUniformIndexing = false;
1014 f->shaderUniformTexelBufferArrayNonUniformIndexing = descIndexing;
1015 f->shaderStorageTexelBufferArrayNonUniformIndexing = descIndexing;
1016 f->descriptorBindingUniformBufferUpdateAfterBind = false;
1017 f->descriptorBindingSampledImageUpdateAfterBind = descIndexing;
1018 f->descriptorBindingStorageImageUpdateAfterBind = descIndexing;
1019 f->descriptorBindingStorageBufferUpdateAfterBind = descIndexing;
1020 f->descriptorBindingUniformTexelBufferUpdateAfterBind = descIndexing;
1021 f->descriptorBindingStorageTexelBufferUpdateAfterBind = descIndexing;
1022 f->descriptorBindingUpdateUnusedWhilePending = descIndexing;
1023 f->descriptorBindingPartiallyBound = descIndexing;
1024 f->descriptorBindingVariableDescriptorCount = false;
1025 f->runtimeDescriptorArray = descIndexing;
1026
1027 f->samplerFilterMinmax = pdevice->info.gen >= 9;
1028 f->scalarBlockLayout = true;
1029 f->imagelessFramebuffer = true;
1030 f->uniformBufferStandardLayout = true;
1031 f->shaderSubgroupExtendedTypes = true;
1032 f->separateDepthStencilLayouts = true;
1033 f->hostQueryReset = true;
1034 f->timelineSemaphore = true;
1035 f->bufferDeviceAddress = pdevice->has_a64_buffer_access;
1036 f->bufferDeviceAddressCaptureReplay = pdevice->has_a64_buffer_access;
1037 f->bufferDeviceAddressMultiDevice = false;
1038 f->vulkanMemoryModel = true;
1039 f->vulkanMemoryModelDeviceScope = true;
1040 f->vulkanMemoryModelAvailabilityVisibilityChains = true;
1041 f->shaderOutputViewportIndex = true;
1042 f->shaderOutputLayer = true;
1043 f->subgroupBroadcastDynamicId = true;
1044 }
1045
1046 void anv_GetPhysicalDeviceFeatures2(
1047 VkPhysicalDevice physicalDevice,
1048 VkPhysicalDeviceFeatures2* pFeatures)
1049 {
1050 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
1051 anv_GetPhysicalDeviceFeatures(physicalDevice, &pFeatures->features);
1052
1053 VkPhysicalDeviceVulkan11Features core_1_1 = {
1054 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES,
1055 };
1056 anv_get_physical_device_features_1_1(pdevice, &core_1_1);
1057
1058 VkPhysicalDeviceVulkan12Features core_1_2 = {
1059 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES,
1060 };
1061 anv_get_physical_device_features_1_2(pdevice, &core_1_2);
1062
1063 #define CORE_FEATURE(major, minor, feature) \
1064 features->feature = core_##major##_##minor.feature
1065
1066
1067 vk_foreach_struct(ext, pFeatures->pNext) {
1068 switch (ext->sType) {
1069 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES_KHR: {
1070 VkPhysicalDevice8BitStorageFeaturesKHR *features =
1071 (VkPhysicalDevice8BitStorageFeaturesKHR *)ext;
1072 CORE_FEATURE(1, 2, storageBuffer8BitAccess);
1073 CORE_FEATURE(1, 2, uniformAndStorageBuffer8BitAccess);
1074 CORE_FEATURE(1, 2, storagePushConstant8);
1075 break;
1076 }
1077
1078 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES: {
1079 VkPhysicalDevice16BitStorageFeatures *features =
1080 (VkPhysicalDevice16BitStorageFeatures *)ext;
1081 CORE_FEATURE(1, 1, storageBuffer16BitAccess);
1082 CORE_FEATURE(1, 1, uniformAndStorageBuffer16BitAccess);
1083 CORE_FEATURE(1, 1, storagePushConstant16);
1084 CORE_FEATURE(1, 1, storageInputOutput16);
1085 break;
1086 }
1087
1088 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_EXT: {
1089 VkPhysicalDeviceBufferDeviceAddressFeaturesEXT *features = (void *)ext;
1090 features->bufferDeviceAddress = pdevice->has_a64_buffer_access;
1091 features->bufferDeviceAddressCaptureReplay = false;
1092 features->bufferDeviceAddressMultiDevice = false;
1093 break;
1094 }
1095
1096 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_KHR: {
1097 VkPhysicalDeviceBufferDeviceAddressFeaturesKHR *features = (void *)ext;
1098 CORE_FEATURE(1, 2, bufferDeviceAddress);
1099 CORE_FEATURE(1, 2, bufferDeviceAddressCaptureReplay);
1100 CORE_FEATURE(1, 2, bufferDeviceAddressMultiDevice);
1101 break;
1102 }
1103
1104 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COMPUTE_SHADER_DERIVATIVES_FEATURES_NV: {
1105 VkPhysicalDeviceComputeShaderDerivativesFeaturesNV *features =
1106 (VkPhysicalDeviceComputeShaderDerivativesFeaturesNV *)ext;
1107 features->computeDerivativeGroupQuads = true;
1108 features->computeDerivativeGroupLinear = true;
1109 break;
1110 }
1111
1112 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT: {
1113 VkPhysicalDeviceConditionalRenderingFeaturesEXT *features =
1114 (VkPhysicalDeviceConditionalRenderingFeaturesEXT*)ext;
1115 features->conditionalRendering = pdevice->info.gen >= 8 ||
1116 pdevice->info.is_haswell;
1117 features->inheritedConditionalRendering = pdevice->info.gen >= 8 ||
1118 pdevice->info.is_haswell;
1119 break;
1120 }
1121
1122 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_CLIP_ENABLE_FEATURES_EXT: {
1123 VkPhysicalDeviceDepthClipEnableFeaturesEXT *features =
1124 (VkPhysicalDeviceDepthClipEnableFeaturesEXT *)ext;
1125 features->depthClipEnable = true;
1126 break;
1127 }
1128
1129 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR: {
1130 VkPhysicalDeviceFloat16Int8FeaturesKHR *features = (void *)ext;
1131 CORE_FEATURE(1, 2, shaderFloat16);
1132 CORE_FEATURE(1, 2, shaderInt8);
1133 break;
1134 }
1135
1136 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FRAGMENT_SHADER_INTERLOCK_FEATURES_EXT: {
1137 VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT *features =
1138 (VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT *)ext;
1139 features->fragmentShaderSampleInterlock = pdevice->info.gen >= 9;
1140 features->fragmentShaderPixelInterlock = pdevice->info.gen >= 9;
1141 features->fragmentShaderShadingRateInterlock = false;
1142 break;
1143 }
1144
1145 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_QUERY_RESET_FEATURES_EXT: {
1146 VkPhysicalDeviceHostQueryResetFeaturesEXT *features =
1147 (VkPhysicalDeviceHostQueryResetFeaturesEXT *)ext;
1148 CORE_FEATURE(1, 2, hostQueryReset);
1149 break;
1150 }
1151
1152 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT: {
1153 VkPhysicalDeviceDescriptorIndexingFeaturesEXT *features =
1154 (VkPhysicalDeviceDescriptorIndexingFeaturesEXT *)ext;
1155 CORE_FEATURE(1, 2, shaderInputAttachmentArrayDynamicIndexing);
1156 CORE_FEATURE(1, 2, shaderUniformTexelBufferArrayDynamicIndexing);
1157 CORE_FEATURE(1, 2, shaderStorageTexelBufferArrayDynamicIndexing);
1158 CORE_FEATURE(1, 2, shaderUniformBufferArrayNonUniformIndexing);
1159 CORE_FEATURE(1, 2, shaderSampledImageArrayNonUniformIndexing);
1160 CORE_FEATURE(1, 2, shaderStorageBufferArrayNonUniformIndexing);
1161 CORE_FEATURE(1, 2, shaderStorageImageArrayNonUniformIndexing);
1162 CORE_FEATURE(1, 2, shaderInputAttachmentArrayNonUniformIndexing);
1163 CORE_FEATURE(1, 2, shaderUniformTexelBufferArrayNonUniformIndexing);
1164 CORE_FEATURE(1, 2, shaderStorageTexelBufferArrayNonUniformIndexing);
1165 CORE_FEATURE(1, 2, descriptorBindingUniformBufferUpdateAfterBind);
1166 CORE_FEATURE(1, 2, descriptorBindingSampledImageUpdateAfterBind);
1167 CORE_FEATURE(1, 2, descriptorBindingStorageImageUpdateAfterBind);
1168 CORE_FEATURE(1, 2, descriptorBindingStorageBufferUpdateAfterBind);
1169 CORE_FEATURE(1, 2, descriptorBindingUniformTexelBufferUpdateAfterBind);
1170 CORE_FEATURE(1, 2, descriptorBindingStorageTexelBufferUpdateAfterBind);
1171 CORE_FEATURE(1, 2, descriptorBindingUpdateUnusedWhilePending);
1172 CORE_FEATURE(1, 2, descriptorBindingPartiallyBound);
1173 CORE_FEATURE(1, 2, descriptorBindingVariableDescriptorCount);
1174 CORE_FEATURE(1, 2, runtimeDescriptorArray);
1175 break;
1176 }
1177
1178 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INDEX_TYPE_UINT8_FEATURES_EXT: {
1179 VkPhysicalDeviceIndexTypeUint8FeaturesEXT *features =
1180 (VkPhysicalDeviceIndexTypeUint8FeaturesEXT *)ext;
1181 features->indexTypeUint8 = true;
1182 break;
1183 }
1184
1185 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_FEATURES_EXT: {
1186 VkPhysicalDeviceInlineUniformBlockFeaturesEXT *features =
1187 (VkPhysicalDeviceInlineUniformBlockFeaturesEXT *)ext;
1188 features->inlineUniformBlock = true;
1189 features->descriptorBindingInlineUniformBlockUpdateAfterBind = true;
1190 break;
1191 }
1192
1193 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_FEATURES_EXT: {
1194 VkPhysicalDeviceLineRasterizationFeaturesEXT *features =
1195 (VkPhysicalDeviceLineRasterizationFeaturesEXT *)ext;
1196 features->rectangularLines = true;
1197 features->bresenhamLines = true;
1198 /* Support for Smooth lines with MSAA was removed on gen11. From the
1199 * BSpec section "Multisample ModesState" table for "AA Line Support
1200 * Requirements":
1201 *
1202 * GEN10:BUG:######## NUM_MULTISAMPLES == 1
1203 *
1204 * Fortunately, this isn't a case most people care about.
1205 */
1206 features->smoothLines = pdevice->info.gen < 10;
1207 features->stippledRectangularLines = false;
1208 features->stippledBresenhamLines = true;
1209 features->stippledSmoothLines = false;
1210 break;
1211 }
1212
1213 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES: {
1214 VkPhysicalDeviceMultiviewFeatures *features =
1215 (VkPhysicalDeviceMultiviewFeatures *)ext;
1216 CORE_FEATURE(1, 1, multiview);
1217 CORE_FEATURE(1, 1, multiviewGeometryShader);
1218 CORE_FEATURE(1, 1, multiviewTessellationShader);
1219 break;
1220 }
1221
1222 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGELESS_FRAMEBUFFER_FEATURES_KHR: {
1223 VkPhysicalDeviceImagelessFramebufferFeaturesKHR *features =
1224 (VkPhysicalDeviceImagelessFramebufferFeaturesKHR *)ext;
1225 CORE_FEATURE(1, 2, imagelessFramebuffer);
1226 break;
1227 }
1228
1229 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR: {
1230 VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR *features =
1231 (VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR *)ext;
1232 features->pipelineExecutableInfo = true;
1233 break;
1234 }
1235
1236 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES: {
1237 VkPhysicalDeviceProtectedMemoryFeatures *features = (void *)ext;
1238 CORE_FEATURE(1, 1, protectedMemory);
1239 break;
1240 }
1241
1242 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES: {
1243 VkPhysicalDeviceSamplerYcbcrConversionFeatures *features =
1244 (VkPhysicalDeviceSamplerYcbcrConversionFeatures *) ext;
1245 CORE_FEATURE(1, 1, samplerYcbcrConversion);
1246 break;
1247 }
1248
1249 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES_EXT: {
1250 VkPhysicalDeviceScalarBlockLayoutFeaturesEXT *features =
1251 (VkPhysicalDeviceScalarBlockLayoutFeaturesEXT *)ext;
1252 CORE_FEATURE(1, 2, scalarBlockLayout);
1253 break;
1254 }
1255
1256 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SEPARATE_DEPTH_STENCIL_LAYOUTS_FEATURES_KHR: {
1257 VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR *features =
1258 (VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR *)ext;
1259 CORE_FEATURE(1, 2, separateDepthStencilLayouts);
1260 break;
1261 }
1262
1263 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_INT64_FEATURES_KHR: {
1264 VkPhysicalDeviceShaderAtomicInt64FeaturesKHR *features = (void *)ext;
1265 CORE_FEATURE(1, 2, shaderBufferInt64Atomics);
1266 CORE_FEATURE(1, 2, shaderSharedInt64Atomics);
1267 break;
1268 }
1269
1270 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DEMOTE_TO_HELPER_INVOCATION_FEATURES_EXT: {
1271 VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT *features = (void *)ext;
1272 features->shaderDemoteToHelperInvocation = true;
1273 break;
1274 }
1275
1276 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CLOCK_FEATURES_KHR: {
1277 VkPhysicalDeviceShaderClockFeaturesKHR *features =
1278 (VkPhysicalDeviceShaderClockFeaturesKHR *)ext;
1279 features->shaderSubgroupClock = true;
1280 features->shaderDeviceClock = false;
1281 break;
1282 }
1283
1284 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETERS_FEATURES: {
1285 VkPhysicalDeviceShaderDrawParametersFeatures *features = (void *)ext;
1286 CORE_FEATURE(1, 1, shaderDrawParameters);
1287 break;
1288 }
1289
1290 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_SUBGROUP_EXTENDED_TYPES_FEATURES_KHR: {
1291 VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR *features =
1292 (VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR *)ext;
1293 CORE_FEATURE(1, 2, shaderSubgroupExtendedTypes);
1294 break;
1295 }
1296
1297 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_FEATURES_EXT: {
1298 VkPhysicalDeviceSubgroupSizeControlFeaturesEXT *features =
1299 (VkPhysicalDeviceSubgroupSizeControlFeaturesEXT *)ext;
1300 features->subgroupSizeControl = true;
1301 features->computeFullSubgroups = true;
1302 break;
1303 }
1304
1305 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_FEATURES_EXT: {
1306 VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT *features =
1307 (VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT *)ext;
1308 features->texelBufferAlignment = true;
1309 break;
1310 }
1311
1312 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_FEATURES_KHR: {
1313 VkPhysicalDeviceTimelineSemaphoreFeaturesKHR *features =
1314 (VkPhysicalDeviceTimelineSemaphoreFeaturesKHR *) ext;
1315 CORE_FEATURE(1, 2, timelineSemaphore);
1316 break;
1317 }
1318
1319 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES: {
1320 VkPhysicalDeviceVariablePointersFeatures *features = (void *)ext;
1321 CORE_FEATURE(1, 1, variablePointersStorageBuffer);
1322 CORE_FEATURE(1, 1, variablePointers);
1323 break;
1324 }
1325
1326 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT: {
1327 VkPhysicalDeviceTransformFeedbackFeaturesEXT *features =
1328 (VkPhysicalDeviceTransformFeedbackFeaturesEXT *)ext;
1329 features->transformFeedback = true;
1330 features->geometryStreams = true;
1331 break;
1332 }
1333
1334 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_UNIFORM_BUFFER_STANDARD_LAYOUT_FEATURES_KHR: {
1335 VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR *features =
1336 (VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR *)ext;
1337 CORE_FEATURE(1, 2, uniformBufferStandardLayout);
1338 break;
1339 }
1340
1341 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT: {
1342 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *features =
1343 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *)ext;
1344 features->vertexAttributeInstanceRateDivisor = true;
1345 features->vertexAttributeInstanceRateZeroDivisor = true;
1346 break;
1347 }
1348
1349 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES:
1350 anv_get_physical_device_features_1_1(pdevice, (void *)ext);
1351 break;
1352
1353 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES:
1354 anv_get_physical_device_features_1_2(pdevice, (void *)ext);
1355 break;
1356
1357 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_MEMORY_MODEL_FEATURES_KHR: {
1358 VkPhysicalDeviceVulkanMemoryModelFeaturesKHR *features = (void *)ext;
1359 CORE_FEATURE(1, 2, vulkanMemoryModel);
1360 CORE_FEATURE(1, 2, vulkanMemoryModelDeviceScope);
1361 CORE_FEATURE(1, 2, vulkanMemoryModelAvailabilityVisibilityChains);
1362 break;
1363 }
1364
1365 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_YCBCR_IMAGE_ARRAYS_FEATURES_EXT: {
1366 VkPhysicalDeviceYcbcrImageArraysFeaturesEXT *features =
1367 (VkPhysicalDeviceYcbcrImageArraysFeaturesEXT *)ext;
1368 features->ycbcrImageArrays = true;
1369 break;
1370 }
1371
1372 default:
1373 anv_debug_ignored_stype(ext->sType);
1374 break;
1375 }
1376 }
1377
1378 #undef CORE_FEATURE
1379 }
1380
1381 #define MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS 64
1382
1383 #define MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS 64
1384 #define MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS 256
1385
1386 void anv_GetPhysicalDeviceProperties(
1387 VkPhysicalDevice physicalDevice,
1388 VkPhysicalDeviceProperties* pProperties)
1389 {
1390 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
1391 const struct gen_device_info *devinfo = &pdevice->info;
1392
1393 /* See assertions made when programming the buffer surface state. */
1394 const uint32_t max_raw_buffer_sz = devinfo->gen >= 7 ?
1395 (1ul << 30) : (1ul << 27);
1396
1397 const uint32_t max_ssbos = pdevice->has_a64_buffer_access ? UINT16_MAX : 64;
1398 const uint32_t max_textures =
1399 pdevice->has_bindless_images ? UINT16_MAX : 128;
1400 const uint32_t max_samplers =
1401 pdevice->has_bindless_samplers ? UINT16_MAX :
1402 (devinfo->gen >= 8 || devinfo->is_haswell) ? 128 : 16;
1403 const uint32_t max_images =
1404 pdevice->has_bindless_images ? UINT16_MAX : MAX_IMAGES;
1405
1406 /* If we can use bindless for everything, claim a high per-stage limit,
1407 * otherwise use the binding table size, minus the slots reserved for
1408 * render targets and one slot for the descriptor buffer. */
1409 const uint32_t max_per_stage =
1410 pdevice->has_bindless_images && pdevice->has_a64_buffer_access
1411 ? UINT32_MAX : MAX_BINDING_TABLE_SIZE - MAX_RTS - 1;
1412
1413 const uint32_t max_workgroup_size = 32 * devinfo->max_cs_threads;
1414
1415 VkSampleCountFlags sample_counts =
1416 isl_device_get_sample_counts(&pdevice->isl_dev);
1417
1418
1419 VkPhysicalDeviceLimits limits = {
1420 .maxImageDimension1D = (1 << 14),
1421 .maxImageDimension2D = (1 << 14),
1422 .maxImageDimension3D = (1 << 11),
1423 .maxImageDimensionCube = (1 << 14),
1424 .maxImageArrayLayers = (1 << 11),
1425 .maxTexelBufferElements = 128 * 1024 * 1024,
1426 .maxUniformBufferRange = (1ul << 27),
1427 .maxStorageBufferRange = max_raw_buffer_sz,
1428 .maxPushConstantsSize = MAX_PUSH_CONSTANTS_SIZE,
1429 .maxMemoryAllocationCount = UINT32_MAX,
1430 .maxSamplerAllocationCount = 64 * 1024,
1431 .bufferImageGranularity = 64, /* A cache line */
1432 .sparseAddressSpaceSize = 0,
1433 .maxBoundDescriptorSets = MAX_SETS,
1434 .maxPerStageDescriptorSamplers = max_samplers,
1435 .maxPerStageDescriptorUniformBuffers = MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS,
1436 .maxPerStageDescriptorStorageBuffers = max_ssbos,
1437 .maxPerStageDescriptorSampledImages = max_textures,
1438 .maxPerStageDescriptorStorageImages = max_images,
1439 .maxPerStageDescriptorInputAttachments = MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS,
1440 .maxPerStageResources = max_per_stage,
1441 .maxDescriptorSetSamplers = 6 * max_samplers, /* number of stages * maxPerStageDescriptorSamplers */
1442 .maxDescriptorSetUniformBuffers = 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS, /* number of stages * maxPerStageDescriptorUniformBuffers */
1443 .maxDescriptorSetUniformBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2,
1444 .maxDescriptorSetStorageBuffers = 6 * max_ssbos, /* number of stages * maxPerStageDescriptorStorageBuffers */
1445 .maxDescriptorSetStorageBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2,
1446 .maxDescriptorSetSampledImages = 6 * max_textures, /* number of stages * maxPerStageDescriptorSampledImages */
1447 .maxDescriptorSetStorageImages = 6 * max_images, /* number of stages * maxPerStageDescriptorStorageImages */
1448 .maxDescriptorSetInputAttachments = MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS,
1449 .maxVertexInputAttributes = MAX_VBS,
1450 .maxVertexInputBindings = MAX_VBS,
1451 .maxVertexInputAttributeOffset = 2047,
1452 .maxVertexInputBindingStride = 2048,
1453 .maxVertexOutputComponents = 128,
1454 .maxTessellationGenerationLevel = 64,
1455 .maxTessellationPatchSize = 32,
1456 .maxTessellationControlPerVertexInputComponents = 128,
1457 .maxTessellationControlPerVertexOutputComponents = 128,
1458 .maxTessellationControlPerPatchOutputComponents = 128,
1459 .maxTessellationControlTotalOutputComponents = 2048,
1460 .maxTessellationEvaluationInputComponents = 128,
1461 .maxTessellationEvaluationOutputComponents = 128,
1462 .maxGeometryShaderInvocations = 32,
1463 .maxGeometryInputComponents = 64,
1464 .maxGeometryOutputComponents = 128,
1465 .maxGeometryOutputVertices = 256,
1466 .maxGeometryTotalOutputComponents = 1024,
1467 .maxFragmentInputComponents = 116, /* 128 components - (PSIZ, CLIP_DIST0, CLIP_DIST1) */
1468 .maxFragmentOutputAttachments = 8,
1469 .maxFragmentDualSrcAttachments = 1,
1470 .maxFragmentCombinedOutputResources = 8,
1471 .maxComputeSharedMemorySize = 64 * 1024,
1472 .maxComputeWorkGroupCount = { 65535, 65535, 65535 },
1473 .maxComputeWorkGroupInvocations = max_workgroup_size,
1474 .maxComputeWorkGroupSize = {
1475 max_workgroup_size,
1476 max_workgroup_size,
1477 max_workgroup_size,
1478 },
1479 .subPixelPrecisionBits = 8,
1480 .subTexelPrecisionBits = 8,
1481 .mipmapPrecisionBits = 8,
1482 .maxDrawIndexedIndexValue = UINT32_MAX,
1483 .maxDrawIndirectCount = UINT32_MAX,
1484 .maxSamplerLodBias = 16,
1485 .maxSamplerAnisotropy = 16,
1486 .maxViewports = MAX_VIEWPORTS,
1487 .maxViewportDimensions = { (1 << 14), (1 << 14) },
1488 .viewportBoundsRange = { INT16_MIN, INT16_MAX },
1489 .viewportSubPixelBits = 13, /* We take a float? */
1490 .minMemoryMapAlignment = 4096, /* A page */
1491 /* The dataport requires texel alignment so we need to assume a worst
1492 * case of R32G32B32A32 which is 16 bytes.
1493 */
1494 .minTexelBufferOffsetAlignment = 16,
1495 /* We need 16 for UBO block reads to work and 32 for push UBOs */
1496 .minUniformBufferOffsetAlignment = 32,
1497 .minStorageBufferOffsetAlignment = 4,
1498 .minTexelOffset = -8,
1499 .maxTexelOffset = 7,
1500 .minTexelGatherOffset = -32,
1501 .maxTexelGatherOffset = 31,
1502 .minInterpolationOffset = -0.5,
1503 .maxInterpolationOffset = 0.4375,
1504 .subPixelInterpolationOffsetBits = 4,
1505 .maxFramebufferWidth = (1 << 14),
1506 .maxFramebufferHeight = (1 << 14),
1507 .maxFramebufferLayers = (1 << 11),
1508 .framebufferColorSampleCounts = sample_counts,
1509 .framebufferDepthSampleCounts = sample_counts,
1510 .framebufferStencilSampleCounts = sample_counts,
1511 .framebufferNoAttachmentsSampleCounts = sample_counts,
1512 .maxColorAttachments = MAX_RTS,
1513 .sampledImageColorSampleCounts = sample_counts,
1514 .sampledImageIntegerSampleCounts = sample_counts,
1515 .sampledImageDepthSampleCounts = sample_counts,
1516 .sampledImageStencilSampleCounts = sample_counts,
1517 .storageImageSampleCounts = VK_SAMPLE_COUNT_1_BIT,
1518 .maxSampleMaskWords = 1,
1519 .timestampComputeAndGraphics = true,
1520 .timestampPeriod = 1000000000.0 / devinfo->timestamp_frequency,
1521 .maxClipDistances = 8,
1522 .maxCullDistances = 8,
1523 .maxCombinedClipAndCullDistances = 8,
1524 .discreteQueuePriorities = 2,
1525 .pointSizeRange = { 0.125, 255.875 },
1526 .lineWidthRange = {
1527 0.0,
1528 (devinfo->gen >= 9 || devinfo->is_cherryview) ?
1529 2047.9921875 : 7.9921875,
1530 },
1531 .pointSizeGranularity = (1.0 / 8.0),
1532 .lineWidthGranularity = (1.0 / 128.0),
1533 .strictLines = false,
1534 .standardSampleLocations = true,
1535 .optimalBufferCopyOffsetAlignment = 128,
1536 .optimalBufferCopyRowPitchAlignment = 128,
1537 .nonCoherentAtomSize = 64,
1538 };
1539
1540 *pProperties = (VkPhysicalDeviceProperties) {
1541 .apiVersion = anv_physical_device_api_version(pdevice),
1542 .driverVersion = vk_get_driver_version(),
1543 .vendorID = 0x8086,
1544 .deviceID = pdevice->chipset_id,
1545 .deviceType = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
1546 .limits = limits,
1547 .sparseProperties = {0}, /* Broadwell doesn't do sparse. */
1548 };
1549
1550 snprintf(pProperties->deviceName, sizeof(pProperties->deviceName),
1551 "%s", pdevice->name);
1552 memcpy(pProperties->pipelineCacheUUID,
1553 pdevice->pipeline_cache_uuid, VK_UUID_SIZE);
1554 }
1555
1556 static void
1557 anv_get_physical_device_properties_1_1(struct anv_physical_device *pdevice,
1558 VkPhysicalDeviceVulkan11Properties *p)
1559 {
1560 assert(p->sType == VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES);
1561
1562 memcpy(p->deviceUUID, pdevice->device_uuid, VK_UUID_SIZE);
1563 memcpy(p->driverUUID, pdevice->driver_uuid, VK_UUID_SIZE);
1564 memset(p->deviceLUID, 0, VK_LUID_SIZE);
1565 p->deviceNodeMask = 0;
1566 p->deviceLUIDValid = false;
1567
1568 p->subgroupSize = BRW_SUBGROUP_SIZE;
1569 VkShaderStageFlags scalar_stages = 0;
1570 for (unsigned stage = 0; stage < MESA_SHADER_STAGES; stage++) {
1571 if (pdevice->compiler->scalar_stage[stage])
1572 scalar_stages |= mesa_to_vk_shader_stage(stage);
1573 }
1574 p->subgroupSupportedStages = scalar_stages;
1575 p->subgroupSupportedOperations = VK_SUBGROUP_FEATURE_BASIC_BIT |
1576 VK_SUBGROUP_FEATURE_VOTE_BIT |
1577 VK_SUBGROUP_FEATURE_BALLOT_BIT |
1578 VK_SUBGROUP_FEATURE_SHUFFLE_BIT |
1579 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT |
1580 VK_SUBGROUP_FEATURE_QUAD_BIT;
1581 if (pdevice->info.gen >= 8) {
1582 /* TODO: There's no technical reason why these can't be made to
1583 * work on gen7 but they don't at the moment so it's best to leave
1584 * the feature disabled than enabled and broken.
1585 */
1586 p->subgroupSupportedOperations |= VK_SUBGROUP_FEATURE_ARITHMETIC_BIT |
1587 VK_SUBGROUP_FEATURE_CLUSTERED_BIT;
1588 }
1589 p->subgroupQuadOperationsInAllStages = pdevice->info.gen >= 8;
1590
1591 p->pointClippingBehavior = VK_POINT_CLIPPING_BEHAVIOR_USER_CLIP_PLANES_ONLY;
1592 p->maxMultiviewViewCount = 16;
1593 p->maxMultiviewInstanceIndex = UINT32_MAX / 16;
1594 p->protectedNoFault = false;
1595 /* This value doesn't matter for us today as our per-stage descriptors are
1596 * the real limit.
1597 */
1598 p->maxPerSetDescriptors = 1024;
1599 p->maxMemoryAllocationSize = MAX_MEMORY_ALLOCATION_SIZE;
1600 }
1601
1602 static void
1603 anv_get_physical_device_properties_1_2(struct anv_physical_device *pdevice,
1604 VkPhysicalDeviceVulkan12Properties *p)
1605 {
1606 assert(p->sType == VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES);
1607
1608 p->driverID = VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA_KHR;
1609 memset(p->driverName, 0, sizeof(p->driverName));
1610 snprintf(p->driverName, VK_MAX_DRIVER_NAME_SIZE_KHR,
1611 "Intel open-source Mesa driver");
1612 memset(p->driverInfo, 0, sizeof(p->driverInfo));
1613 snprintf(p->driverInfo, VK_MAX_DRIVER_INFO_SIZE_KHR,
1614 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1);
1615 p->conformanceVersion = (VkConformanceVersionKHR) {
1616 .major = 1,
1617 .minor = 2,
1618 .subminor = 0,
1619 .patch = 0,
1620 };
1621
1622 p->denormBehaviorIndependence =
1623 VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_ALL_KHR;
1624 p->roundingModeIndependence =
1625 VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_NONE_KHR;
1626
1627 /* Broadwell does not support HF denorms and there are restrictions
1628 * other gens. According to Kabylake's PRM:
1629 *
1630 * "math - Extended Math Function
1631 * [...]
1632 * Restriction : Half-float denorms are always retained."
1633 */
1634 p->shaderDenormFlushToZeroFloat16 = false;
1635 p->shaderDenormPreserveFloat16 = pdevice->info.gen > 8;
1636 p->shaderRoundingModeRTEFloat16 = true;
1637 p->shaderRoundingModeRTZFloat16 = true;
1638 p->shaderSignedZeroInfNanPreserveFloat16 = true;
1639
1640 p->shaderDenormFlushToZeroFloat32 = true;
1641 p->shaderDenormPreserveFloat32 = true;
1642 p->shaderRoundingModeRTEFloat32 = true;
1643 p->shaderRoundingModeRTZFloat32 = true;
1644 p->shaderSignedZeroInfNanPreserveFloat32 = true;
1645
1646 p->shaderDenormFlushToZeroFloat64 = true;
1647 p->shaderDenormPreserveFloat64 = true;
1648 p->shaderRoundingModeRTEFloat64 = true;
1649 p->shaderRoundingModeRTZFloat64 = true;
1650 p->shaderSignedZeroInfNanPreserveFloat64 = true;
1651
1652 /* It's a bit hard to exactly map our implementation to the limits
1653 * described here. The bindless surface handle in the extended
1654 * message descriptors is 20 bits and it's an index into the table of
1655 * RENDER_SURFACE_STATE structs that starts at bindless surface base
1656 * address. Given that most things consume two surface states per
1657 * view (general/sampled for textures and write-only/read-write for
1658 * images), we claim 2^19 things.
1659 *
1660 * For SSBOs, we just use A64 messages so there is no real limit
1661 * there beyond the limit on the total size of a descriptor set.
1662 */
1663 const unsigned max_bindless_views = 1 << 19;
1664 p->maxUpdateAfterBindDescriptorsInAllPools = max_bindless_views;
1665 p->shaderUniformBufferArrayNonUniformIndexingNative = false;
1666 p->shaderSampledImageArrayNonUniformIndexingNative = false;
1667 p->shaderStorageBufferArrayNonUniformIndexingNative = true;
1668 p->shaderStorageImageArrayNonUniformIndexingNative = false;
1669 p->shaderInputAttachmentArrayNonUniformIndexingNative = false;
1670 p->robustBufferAccessUpdateAfterBind = true;
1671 p->quadDivergentImplicitLod = false;
1672 p->maxPerStageDescriptorUpdateAfterBindSamplers = max_bindless_views;
1673 p->maxPerStageDescriptorUpdateAfterBindUniformBuffers = MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS;
1674 p->maxPerStageDescriptorUpdateAfterBindStorageBuffers = UINT32_MAX;
1675 p->maxPerStageDescriptorUpdateAfterBindSampledImages = max_bindless_views;
1676 p->maxPerStageDescriptorUpdateAfterBindStorageImages = max_bindless_views;
1677 p->maxPerStageDescriptorUpdateAfterBindInputAttachments = MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS;
1678 p->maxPerStageUpdateAfterBindResources = UINT32_MAX;
1679 p->maxDescriptorSetUpdateAfterBindSamplers = max_bindless_views;
1680 p->maxDescriptorSetUpdateAfterBindUniformBuffers = 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS;
1681 p->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2;
1682 p->maxDescriptorSetUpdateAfterBindStorageBuffers = UINT32_MAX;
1683 p->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2;
1684 p->maxDescriptorSetUpdateAfterBindSampledImages = max_bindless_views;
1685 p->maxDescriptorSetUpdateAfterBindStorageImages = max_bindless_views;
1686 p->maxDescriptorSetUpdateAfterBindInputAttachments = MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS;
1687
1688 /* We support all of the depth resolve modes */
1689 p->supportedDepthResolveModes = VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR |
1690 VK_RESOLVE_MODE_AVERAGE_BIT_KHR |
1691 VK_RESOLVE_MODE_MIN_BIT_KHR |
1692 VK_RESOLVE_MODE_MAX_BIT_KHR;
1693 /* Average doesn't make sense for stencil so we don't support that */
1694 p->supportedStencilResolveModes = VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR;
1695 if (pdevice->info.gen >= 8) {
1696 /* The advanced stencil resolve modes currently require stencil
1697 * sampling be supported by the hardware.
1698 */
1699 p->supportedStencilResolveModes |= VK_RESOLVE_MODE_MIN_BIT_KHR |
1700 VK_RESOLVE_MODE_MAX_BIT_KHR;
1701 }
1702 p->independentResolveNone = true;
1703 p->independentResolve = true;
1704
1705 p->filterMinmaxSingleComponentFormats = pdevice->info.gen >= 9;
1706 p->filterMinmaxImageComponentMapping = pdevice->info.gen >= 9;
1707
1708 p->maxTimelineSemaphoreValueDifference = UINT64_MAX;
1709
1710 p->framebufferIntegerColorSampleCounts =
1711 isl_device_get_sample_counts(&pdevice->isl_dev);
1712 }
1713
1714 void anv_GetPhysicalDeviceProperties2(
1715 VkPhysicalDevice physicalDevice,
1716 VkPhysicalDeviceProperties2* pProperties)
1717 {
1718 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
1719
1720 anv_GetPhysicalDeviceProperties(physicalDevice, &pProperties->properties);
1721
1722 VkPhysicalDeviceVulkan11Properties core_1_1 = {
1723 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES,
1724 };
1725 anv_get_physical_device_properties_1_1(pdevice, &core_1_1);
1726
1727 VkPhysicalDeviceVulkan12Properties core_1_2 = {
1728 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES,
1729 };
1730 anv_get_physical_device_properties_1_2(pdevice, &core_1_2);
1731
1732 #define CORE_RENAMED_PROPERTY(major, minor, ext_property, core_property) \
1733 memcpy(&properties->ext_property, &core_##major##_##minor.core_property, \
1734 sizeof(core_##major##_##minor.core_property))
1735
1736 #define CORE_PROPERTY(major, minor, property) \
1737 CORE_RENAMED_PROPERTY(major, minor, property, property)
1738
1739 vk_foreach_struct(ext, pProperties->pNext) {
1740 switch (ext->sType) {
1741 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_STENCIL_RESOLVE_PROPERTIES_KHR: {
1742 VkPhysicalDeviceDepthStencilResolvePropertiesKHR *properties =
1743 (VkPhysicalDeviceDepthStencilResolvePropertiesKHR *)ext;
1744 CORE_PROPERTY(1, 2, supportedDepthResolveModes);
1745 CORE_PROPERTY(1, 2, supportedStencilResolveModes);
1746 CORE_PROPERTY(1, 2, independentResolveNone);
1747 CORE_PROPERTY(1, 2, independentResolve);
1748 break;
1749 }
1750
1751 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES_EXT: {
1752 VkPhysicalDeviceDescriptorIndexingPropertiesEXT *properties =
1753 (VkPhysicalDeviceDescriptorIndexingPropertiesEXT *)ext;
1754 CORE_PROPERTY(1, 2, maxUpdateAfterBindDescriptorsInAllPools);
1755 CORE_PROPERTY(1, 2, shaderUniformBufferArrayNonUniformIndexingNative);
1756 CORE_PROPERTY(1, 2, shaderSampledImageArrayNonUniformIndexingNative);
1757 CORE_PROPERTY(1, 2, shaderStorageBufferArrayNonUniformIndexingNative);
1758 CORE_PROPERTY(1, 2, shaderStorageImageArrayNonUniformIndexingNative);
1759 CORE_PROPERTY(1, 2, shaderInputAttachmentArrayNonUniformIndexingNative);
1760 CORE_PROPERTY(1, 2, robustBufferAccessUpdateAfterBind);
1761 CORE_PROPERTY(1, 2, quadDivergentImplicitLod);
1762 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindSamplers);
1763 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindUniformBuffers);
1764 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindStorageBuffers);
1765 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindSampledImages);
1766 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindStorageImages);
1767 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindInputAttachments);
1768 CORE_PROPERTY(1, 2, maxPerStageUpdateAfterBindResources);
1769 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindSamplers);
1770 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindUniformBuffers);
1771 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindUniformBuffersDynamic);
1772 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindStorageBuffers);
1773 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindStorageBuffersDynamic);
1774 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindSampledImages);
1775 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindStorageImages);
1776 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindInputAttachments);
1777 break;
1778 }
1779
1780 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR: {
1781 VkPhysicalDeviceDriverPropertiesKHR *properties =
1782 (VkPhysicalDeviceDriverPropertiesKHR *) ext;
1783 CORE_PROPERTY(1, 2, driverID);
1784 CORE_PROPERTY(1, 2, driverName);
1785 CORE_PROPERTY(1, 2, driverInfo);
1786 CORE_PROPERTY(1, 2, conformanceVersion);
1787 break;
1788 }
1789
1790 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT: {
1791 VkPhysicalDeviceExternalMemoryHostPropertiesEXT *props =
1792 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT *) ext;
1793 /* Userptr needs page aligned memory. */
1794 props->minImportedHostPointerAlignment = 4096;
1795 break;
1796 }
1797
1798 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES: {
1799 VkPhysicalDeviceIDProperties *properties =
1800 (VkPhysicalDeviceIDProperties *)ext;
1801 CORE_PROPERTY(1, 1, deviceUUID);
1802 CORE_PROPERTY(1, 1, driverUUID);
1803 CORE_PROPERTY(1, 1, deviceLUID);
1804 CORE_PROPERTY(1, 1, deviceLUIDValid);
1805 break;
1806 }
1807
1808 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_PROPERTIES_EXT: {
1809 VkPhysicalDeviceInlineUniformBlockPropertiesEXT *props =
1810 (VkPhysicalDeviceInlineUniformBlockPropertiesEXT *)ext;
1811 props->maxInlineUniformBlockSize = MAX_INLINE_UNIFORM_BLOCK_SIZE;
1812 props->maxPerStageDescriptorInlineUniformBlocks =
1813 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1814 props->maxPerStageDescriptorUpdateAfterBindInlineUniformBlocks =
1815 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1816 props->maxDescriptorSetInlineUniformBlocks =
1817 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1818 props->maxDescriptorSetUpdateAfterBindInlineUniformBlocks =
1819 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1820 break;
1821 }
1822
1823 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_PROPERTIES_EXT: {
1824 VkPhysicalDeviceLineRasterizationPropertiesEXT *props =
1825 (VkPhysicalDeviceLineRasterizationPropertiesEXT *)ext;
1826 /* In the Skylake PRM Vol. 7, subsection titled "GIQ (Diamond)
1827 * Sampling Rules - Legacy Mode", it says the following:
1828 *
1829 * "Note that the device divides a pixel into a 16x16 array of
1830 * subpixels, referenced by their upper left corners."
1831 *
1832 * This is the only known reference in the PRMs to the subpixel
1833 * precision of line rasterization and a "16x16 array of subpixels"
1834 * implies 4 subpixel precision bits. Empirical testing has shown
1835 * that 4 subpixel precision bits applies to all line rasterization
1836 * types.
1837 */
1838 props->lineSubPixelPrecisionBits = 4;
1839 break;
1840 }
1841
1842 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES: {
1843 VkPhysicalDeviceMaintenance3Properties *properties =
1844 (VkPhysicalDeviceMaintenance3Properties *)ext;
1845 /* This value doesn't matter for us today as our per-stage
1846 * descriptors are the real limit.
1847 */
1848 CORE_PROPERTY(1, 1, maxPerSetDescriptors);
1849 CORE_PROPERTY(1, 1, maxMemoryAllocationSize);
1850 break;
1851 }
1852
1853 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES: {
1854 VkPhysicalDeviceMultiviewProperties *properties =
1855 (VkPhysicalDeviceMultiviewProperties *)ext;
1856 CORE_PROPERTY(1, 1, maxMultiviewViewCount);
1857 CORE_PROPERTY(1, 1, maxMultiviewInstanceIndex);
1858 break;
1859 }
1860
1861 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT: {
1862 VkPhysicalDevicePCIBusInfoPropertiesEXT *properties =
1863 (VkPhysicalDevicePCIBusInfoPropertiesEXT *)ext;
1864 properties->pciDomain = pdevice->pci_info.domain;
1865 properties->pciBus = pdevice->pci_info.bus;
1866 properties->pciDevice = pdevice->pci_info.device;
1867 properties->pciFunction = pdevice->pci_info.function;
1868 break;
1869 }
1870
1871 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES: {
1872 VkPhysicalDevicePointClippingProperties *properties =
1873 (VkPhysicalDevicePointClippingProperties *) ext;
1874 CORE_PROPERTY(1, 1, pointClippingBehavior);
1875 break;
1876 }
1877
1878 #pragma GCC diagnostic push
1879 #pragma GCC diagnostic ignored "-Wswitch"
1880 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PRESENTATION_PROPERTIES_ANDROID: {
1881 VkPhysicalDevicePresentationPropertiesANDROID *props =
1882 (VkPhysicalDevicePresentationPropertiesANDROID *)ext;
1883 props->sharedImage = VK_FALSE;
1884 break;
1885 }
1886 #pragma GCC diagnostic pop
1887
1888 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES: {
1889 VkPhysicalDeviceProtectedMemoryProperties *properties =
1890 (VkPhysicalDeviceProtectedMemoryProperties *)ext;
1891 CORE_PROPERTY(1, 1, protectedNoFault);
1892 break;
1893 }
1894
1895 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR: {
1896 VkPhysicalDevicePushDescriptorPropertiesKHR *properties =
1897 (VkPhysicalDevicePushDescriptorPropertiesKHR *) ext;
1898 properties->maxPushDescriptors = MAX_PUSH_DESCRIPTORS;
1899 break;
1900 }
1901
1902 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT: {
1903 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT *properties =
1904 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT *)ext;
1905 CORE_PROPERTY(1, 2, filterMinmaxImageComponentMapping);
1906 CORE_PROPERTY(1, 2, filterMinmaxSingleComponentFormats);
1907 break;
1908 }
1909
1910 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES: {
1911 VkPhysicalDeviceSubgroupProperties *properties = (void *)ext;
1912 CORE_PROPERTY(1, 1, subgroupSize);
1913 CORE_RENAMED_PROPERTY(1, 1, supportedStages,
1914 subgroupSupportedStages);
1915 CORE_RENAMED_PROPERTY(1, 1, supportedOperations,
1916 subgroupSupportedOperations);
1917 CORE_RENAMED_PROPERTY(1, 1, quadOperationsInAllStages,
1918 subgroupQuadOperationsInAllStages);
1919 break;
1920 }
1921
1922 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_PROPERTIES_EXT: {
1923 VkPhysicalDeviceSubgroupSizeControlPropertiesEXT *props =
1924 (VkPhysicalDeviceSubgroupSizeControlPropertiesEXT *)ext;
1925 STATIC_ASSERT(8 <= BRW_SUBGROUP_SIZE && BRW_SUBGROUP_SIZE <= 32);
1926 props->minSubgroupSize = 8;
1927 props->maxSubgroupSize = 32;
1928 props->maxComputeWorkgroupSubgroups = pdevice->info.max_cs_threads;
1929 props->requiredSubgroupSizeStages = VK_SHADER_STAGE_COMPUTE_BIT;
1930 break;
1931 }
1932 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT_CONTROLS_PROPERTIES_KHR : {
1933 VkPhysicalDeviceFloatControlsPropertiesKHR *properties = (void *)ext;
1934 CORE_PROPERTY(1, 2, denormBehaviorIndependence);
1935 CORE_PROPERTY(1, 2, roundingModeIndependence);
1936 CORE_PROPERTY(1, 2, shaderDenormFlushToZeroFloat16);
1937 CORE_PROPERTY(1, 2, shaderDenormPreserveFloat16);
1938 CORE_PROPERTY(1, 2, shaderRoundingModeRTEFloat16);
1939 CORE_PROPERTY(1, 2, shaderRoundingModeRTZFloat16);
1940 CORE_PROPERTY(1, 2, shaderSignedZeroInfNanPreserveFloat16);
1941 CORE_PROPERTY(1, 2, shaderDenormFlushToZeroFloat32);
1942 CORE_PROPERTY(1, 2, shaderDenormPreserveFloat32);
1943 CORE_PROPERTY(1, 2, shaderRoundingModeRTEFloat32);
1944 CORE_PROPERTY(1, 2, shaderRoundingModeRTZFloat32);
1945 CORE_PROPERTY(1, 2, shaderSignedZeroInfNanPreserveFloat32);
1946 CORE_PROPERTY(1, 2, shaderDenormFlushToZeroFloat64);
1947 CORE_PROPERTY(1, 2, shaderDenormPreserveFloat64);
1948 CORE_PROPERTY(1, 2, shaderRoundingModeRTEFloat64);
1949 CORE_PROPERTY(1, 2, shaderRoundingModeRTZFloat64);
1950 CORE_PROPERTY(1, 2, shaderSignedZeroInfNanPreserveFloat64);
1951 break;
1952 }
1953
1954 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_PROPERTIES_EXT: {
1955 VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT *props =
1956 (VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT *)ext;
1957
1958 /* From the SKL PRM Vol. 2d, docs for RENDER_SURFACE_STATE::Surface
1959 * Base Address:
1960 *
1961 * "For SURFTYPE_BUFFER non-rendertarget surfaces, this field
1962 * specifies the base address of the first element of the surface,
1963 * computed in software by adding the surface base address to the
1964 * byte offset of the element in the buffer. The base address must
1965 * be aligned to element size."
1966 *
1967 * The typed dataport messages require that things be texel aligned.
1968 * Otherwise, we may just load/store the wrong data or, in the worst
1969 * case, there may be hangs.
1970 */
1971 props->storageTexelBufferOffsetAlignmentBytes = 16;
1972 props->storageTexelBufferOffsetSingleTexelAlignment = true;
1973
1974 /* The sampler, however, is much more forgiving and it can handle
1975 * arbitrary byte alignment for linear and buffer surfaces. It's
1976 * hard to find a good PRM citation for this but years of empirical
1977 * experience demonstrate that this is true.
1978 */
1979 props->uniformTexelBufferOffsetAlignmentBytes = 1;
1980 props->uniformTexelBufferOffsetSingleTexelAlignment = false;
1981 break;
1982 }
1983
1984 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_PROPERTIES_KHR: {
1985 VkPhysicalDeviceTimelineSemaphorePropertiesKHR *properties =
1986 (VkPhysicalDeviceTimelineSemaphorePropertiesKHR *) ext;
1987 CORE_PROPERTY(1, 2, maxTimelineSemaphoreValueDifference);
1988 break;
1989 }
1990
1991 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT: {
1992 VkPhysicalDeviceTransformFeedbackPropertiesEXT *props =
1993 (VkPhysicalDeviceTransformFeedbackPropertiesEXT *)ext;
1994
1995 props->maxTransformFeedbackStreams = MAX_XFB_STREAMS;
1996 props->maxTransformFeedbackBuffers = MAX_XFB_BUFFERS;
1997 props->maxTransformFeedbackBufferSize = (1ull << 32);
1998 props->maxTransformFeedbackStreamDataSize = 128 * 4;
1999 props->maxTransformFeedbackBufferDataSize = 128 * 4;
2000 props->maxTransformFeedbackBufferDataStride = 2048;
2001 props->transformFeedbackQueries = true;
2002 props->transformFeedbackStreamsLinesTriangles = false;
2003 props->transformFeedbackRasterizationStreamSelect = false;
2004 props->transformFeedbackDraw = true;
2005 break;
2006 }
2007
2008 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT: {
2009 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *props =
2010 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *)ext;
2011 /* We have to restrict this a bit for multiview */
2012 props->maxVertexAttribDivisor = UINT32_MAX / 16;
2013 break;
2014 }
2015
2016 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES:
2017 anv_get_physical_device_properties_1_1(pdevice, (void *)ext);
2018 break;
2019
2020 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES:
2021 anv_get_physical_device_properties_1_2(pdevice, (void *)ext);
2022 break;
2023
2024 default:
2025 anv_debug_ignored_stype(ext->sType);
2026 break;
2027 }
2028 }
2029
2030 #undef CORE_RENAMED_PROPERTY
2031 #undef CORE_PROPERTY
2032 }
2033
2034 /* We support exactly one queue family. */
2035 static const VkQueueFamilyProperties
2036 anv_queue_family_properties = {
2037 .queueFlags = VK_QUEUE_GRAPHICS_BIT |
2038 VK_QUEUE_COMPUTE_BIT |
2039 VK_QUEUE_TRANSFER_BIT,
2040 .queueCount = 1,
2041 .timestampValidBits = 36, /* XXX: Real value here */
2042 .minImageTransferGranularity = { 1, 1, 1 },
2043 };
2044
2045 void anv_GetPhysicalDeviceQueueFamilyProperties(
2046 VkPhysicalDevice physicalDevice,
2047 uint32_t* pCount,
2048 VkQueueFamilyProperties* pQueueFamilyProperties)
2049 {
2050 VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pCount);
2051
2052 vk_outarray_append(&out, p) {
2053 *p = anv_queue_family_properties;
2054 }
2055 }
2056
2057 void anv_GetPhysicalDeviceQueueFamilyProperties2(
2058 VkPhysicalDevice physicalDevice,
2059 uint32_t* pQueueFamilyPropertyCount,
2060 VkQueueFamilyProperties2* pQueueFamilyProperties)
2061 {
2062
2063 VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pQueueFamilyPropertyCount);
2064
2065 vk_outarray_append(&out, p) {
2066 p->queueFamilyProperties = anv_queue_family_properties;
2067
2068 vk_foreach_struct(s, p->pNext) {
2069 anv_debug_ignored_stype(s->sType);
2070 }
2071 }
2072 }
2073
2074 void anv_GetPhysicalDeviceMemoryProperties(
2075 VkPhysicalDevice physicalDevice,
2076 VkPhysicalDeviceMemoryProperties* pMemoryProperties)
2077 {
2078 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
2079
2080 pMemoryProperties->memoryTypeCount = physical_device->memory.type_count;
2081 for (uint32_t i = 0; i < physical_device->memory.type_count; i++) {
2082 pMemoryProperties->memoryTypes[i] = (VkMemoryType) {
2083 .propertyFlags = physical_device->memory.types[i].propertyFlags,
2084 .heapIndex = physical_device->memory.types[i].heapIndex,
2085 };
2086 }
2087
2088 pMemoryProperties->memoryHeapCount = physical_device->memory.heap_count;
2089 for (uint32_t i = 0; i < physical_device->memory.heap_count; i++) {
2090 pMemoryProperties->memoryHeaps[i] = (VkMemoryHeap) {
2091 .size = physical_device->memory.heaps[i].size,
2092 .flags = physical_device->memory.heaps[i].flags,
2093 };
2094 }
2095 }
2096
2097 static void
2098 anv_get_memory_budget(VkPhysicalDevice physicalDevice,
2099 VkPhysicalDeviceMemoryBudgetPropertiesEXT *memoryBudget)
2100 {
2101 ANV_FROM_HANDLE(anv_physical_device, device, physicalDevice);
2102 uint64_t sys_available = get_available_system_memory();
2103 assert(sys_available > 0);
2104
2105 VkDeviceSize total_heaps_size = 0;
2106 for (size_t i = 0; i < device->memory.heap_count; i++)
2107 total_heaps_size += device->memory.heaps[i].size;
2108
2109 for (size_t i = 0; i < device->memory.heap_count; i++) {
2110 VkDeviceSize heap_size = device->memory.heaps[i].size;
2111 VkDeviceSize heap_used = device->memory.heaps[i].used;
2112 VkDeviceSize heap_budget;
2113
2114 double heap_proportion = (double) heap_size / total_heaps_size;
2115 VkDeviceSize sys_available_prop = sys_available * heap_proportion;
2116
2117 /*
2118 * Let's not incite the app to starve the system: report at most 90% of
2119 * available system memory.
2120 */
2121 uint64_t heap_available = sys_available_prop * 9 / 10;
2122 heap_budget = MIN2(heap_size, heap_used + heap_available);
2123
2124 /*
2125 * Round down to the nearest MB
2126 */
2127 heap_budget &= ~((1ull << 20) - 1);
2128
2129 /*
2130 * The heapBudget value must be non-zero for array elements less than
2131 * VkPhysicalDeviceMemoryProperties::memoryHeapCount. The heapBudget
2132 * value must be less than or equal to VkMemoryHeap::size for each heap.
2133 */
2134 assert(0 < heap_budget && heap_budget <= heap_size);
2135
2136 memoryBudget->heapUsage[i] = heap_used;
2137 memoryBudget->heapBudget[i] = heap_budget;
2138 }
2139
2140 /* The heapBudget and heapUsage values must be zero for array elements
2141 * greater than or equal to VkPhysicalDeviceMemoryProperties::memoryHeapCount
2142 */
2143 for (uint32_t i = device->memory.heap_count; i < VK_MAX_MEMORY_HEAPS; i++) {
2144 memoryBudget->heapBudget[i] = 0;
2145 memoryBudget->heapUsage[i] = 0;
2146 }
2147 }
2148
2149 void anv_GetPhysicalDeviceMemoryProperties2(
2150 VkPhysicalDevice physicalDevice,
2151 VkPhysicalDeviceMemoryProperties2* pMemoryProperties)
2152 {
2153 anv_GetPhysicalDeviceMemoryProperties(physicalDevice,
2154 &pMemoryProperties->memoryProperties);
2155
2156 vk_foreach_struct(ext, pMemoryProperties->pNext) {
2157 switch (ext->sType) {
2158 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT:
2159 anv_get_memory_budget(physicalDevice, (void*)ext);
2160 break;
2161 default:
2162 anv_debug_ignored_stype(ext->sType);
2163 break;
2164 }
2165 }
2166 }
2167
2168 void
2169 anv_GetDeviceGroupPeerMemoryFeatures(
2170 VkDevice device,
2171 uint32_t heapIndex,
2172 uint32_t localDeviceIndex,
2173 uint32_t remoteDeviceIndex,
2174 VkPeerMemoryFeatureFlags* pPeerMemoryFeatures)
2175 {
2176 assert(localDeviceIndex == 0 && remoteDeviceIndex == 0);
2177 *pPeerMemoryFeatures = VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT |
2178 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT |
2179 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT |
2180 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT;
2181 }
2182
2183 PFN_vkVoidFunction anv_GetInstanceProcAddr(
2184 VkInstance _instance,
2185 const char* pName)
2186 {
2187 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2188
2189 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
2190 * when we have to return valid function pointers, NULL, or it's left
2191 * undefined. See the table for exact details.
2192 */
2193 if (pName == NULL)
2194 return NULL;
2195
2196 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
2197 if (strcmp(pName, "vk" #entrypoint) == 0) \
2198 return (PFN_vkVoidFunction)anv_##entrypoint
2199
2200 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties);
2201 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties);
2202 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceVersion);
2203 LOOKUP_ANV_ENTRYPOINT(CreateInstance);
2204
2205 #undef LOOKUP_ANV_ENTRYPOINT
2206
2207 if (instance == NULL)
2208 return NULL;
2209
2210 int idx = anv_get_instance_entrypoint_index(pName);
2211 if (idx >= 0)
2212 return instance->dispatch.entrypoints[idx];
2213
2214 idx = anv_get_physical_device_entrypoint_index(pName);
2215 if (idx >= 0)
2216 return instance->physicalDevice.dispatch.entrypoints[idx];
2217
2218 idx = anv_get_device_entrypoint_index(pName);
2219 if (idx >= 0)
2220 return instance->device_dispatch.entrypoints[idx];
2221
2222 return NULL;
2223 }
2224
2225 /* With version 1+ of the loader interface the ICD should expose
2226 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
2227 */
2228 PUBLIC
2229 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
2230 VkInstance instance,
2231 const char* pName);
2232
2233 PUBLIC
2234 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
2235 VkInstance instance,
2236 const char* pName)
2237 {
2238 return anv_GetInstanceProcAddr(instance, pName);
2239 }
2240
2241 PFN_vkVoidFunction anv_GetDeviceProcAddr(
2242 VkDevice _device,
2243 const char* pName)
2244 {
2245 ANV_FROM_HANDLE(anv_device, device, _device);
2246
2247 if (!device || !pName)
2248 return NULL;
2249
2250 int idx = anv_get_device_entrypoint_index(pName);
2251 if (idx < 0)
2252 return NULL;
2253
2254 return device->dispatch.entrypoints[idx];
2255 }
2256
2257 /* With version 4+ of the loader interface the ICD should expose
2258 * vk_icdGetPhysicalDeviceProcAddr()
2259 */
2260 PUBLIC
2261 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetPhysicalDeviceProcAddr(
2262 VkInstance _instance,
2263 const char* pName);
2264
2265 PFN_vkVoidFunction vk_icdGetPhysicalDeviceProcAddr(
2266 VkInstance _instance,
2267 const char* pName)
2268 {
2269 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2270
2271 if (!pName || !instance)
2272 return NULL;
2273
2274 int idx = anv_get_physical_device_entrypoint_index(pName);
2275 if (idx < 0)
2276 return NULL;
2277
2278 return instance->physicalDevice.dispatch.entrypoints[idx];
2279 }
2280
2281
2282 VkResult
2283 anv_CreateDebugReportCallbackEXT(VkInstance _instance,
2284 const VkDebugReportCallbackCreateInfoEXT* pCreateInfo,
2285 const VkAllocationCallbacks* pAllocator,
2286 VkDebugReportCallbackEXT* pCallback)
2287 {
2288 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2289 return vk_create_debug_report_callback(&instance->debug_report_callbacks,
2290 pCreateInfo, pAllocator, &instance->alloc,
2291 pCallback);
2292 }
2293
2294 void
2295 anv_DestroyDebugReportCallbackEXT(VkInstance _instance,
2296 VkDebugReportCallbackEXT _callback,
2297 const VkAllocationCallbacks* pAllocator)
2298 {
2299 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2300 vk_destroy_debug_report_callback(&instance->debug_report_callbacks,
2301 _callback, pAllocator, &instance->alloc);
2302 }
2303
2304 void
2305 anv_DebugReportMessageEXT(VkInstance _instance,
2306 VkDebugReportFlagsEXT flags,
2307 VkDebugReportObjectTypeEXT objectType,
2308 uint64_t object,
2309 size_t location,
2310 int32_t messageCode,
2311 const char* pLayerPrefix,
2312 const char* pMessage)
2313 {
2314 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2315 vk_debug_report(&instance->debug_report_callbacks, flags, objectType,
2316 object, location, messageCode, pLayerPrefix, pMessage);
2317 }
2318
2319 static struct anv_state
2320 anv_state_pool_emit_data(struct anv_state_pool *pool, size_t size, size_t align, const void *p)
2321 {
2322 struct anv_state state;
2323
2324 state = anv_state_pool_alloc(pool, size, align);
2325 memcpy(state.map, p, size);
2326
2327 return state;
2328 }
2329
2330 /* Haswell border color is a bit of a disaster. Float and unorm formats use a
2331 * straightforward 32-bit float color in the first 64 bytes. Instead of using
2332 * a nice float/integer union like Gen8+, Haswell specifies the integer border
2333 * color as a separate entry /after/ the float color. The layout of this entry
2334 * also depends on the format's bpp (with extra hacks for RG32), and overlaps.
2335 *
2336 * Since we don't know the format/bpp, we can't make any of the border colors
2337 * containing '1' work for all formats, as it would be in the wrong place for
2338 * some of them. We opt to make 32-bit integers work as this seems like the
2339 * most common option. Fortunately, transparent black works regardless, as
2340 * all zeroes is the same in every bit-size.
2341 */
2342 struct hsw_border_color {
2343 float float32[4];
2344 uint32_t _pad0[12];
2345 uint32_t uint32[4];
2346 uint32_t _pad1[108];
2347 };
2348
2349 struct gen8_border_color {
2350 union {
2351 float float32[4];
2352 uint32_t uint32[4];
2353 };
2354 /* Pad out to 64 bytes */
2355 uint32_t _pad[12];
2356 };
2357
2358 static void
2359 anv_device_init_border_colors(struct anv_device *device)
2360 {
2361 if (device->info.is_haswell) {
2362 static const struct hsw_border_color border_colors[] = {
2363 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
2364 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
2365 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
2366 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = { .uint32 = { 0, 0, 0, 0 } },
2367 [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = { .uint32 = { 0, 0, 0, 1 } },
2368 [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = { .uint32 = { 1, 1, 1, 1 } },
2369 };
2370
2371 device->border_colors =
2372 anv_state_pool_emit_data(&device->dynamic_state_pool,
2373 sizeof(border_colors), 512, border_colors);
2374 } else {
2375 static const struct gen8_border_color border_colors[] = {
2376 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
2377 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
2378 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
2379 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = { .uint32 = { 0, 0, 0, 0 } },
2380 [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = { .uint32 = { 0, 0, 0, 1 } },
2381 [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = { .uint32 = { 1, 1, 1, 1 } },
2382 };
2383
2384 device->border_colors =
2385 anv_state_pool_emit_data(&device->dynamic_state_pool,
2386 sizeof(border_colors), 64, border_colors);
2387 }
2388 }
2389
2390 static VkResult
2391 anv_device_init_trivial_batch(struct anv_device *device)
2392 {
2393 VkResult result = anv_device_alloc_bo(device, 4096,
2394 ANV_BO_ALLOC_MAPPED,
2395 0 /* explicit_address */,
2396 &device->trivial_batch_bo);
2397 if (result != VK_SUCCESS)
2398 return result;
2399
2400 struct anv_batch batch = {
2401 .start = device->trivial_batch_bo->map,
2402 .next = device->trivial_batch_bo->map,
2403 .end = device->trivial_batch_bo->map + 4096,
2404 };
2405
2406 anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
2407 anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
2408
2409 if (!device->info.has_llc)
2410 gen_clflush_range(batch.start, batch.next - batch.start);
2411
2412 return VK_SUCCESS;
2413 }
2414
2415 VkResult anv_EnumerateDeviceExtensionProperties(
2416 VkPhysicalDevice physicalDevice,
2417 const char* pLayerName,
2418 uint32_t* pPropertyCount,
2419 VkExtensionProperties* pProperties)
2420 {
2421 ANV_FROM_HANDLE(anv_physical_device, device, physicalDevice);
2422 VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
2423
2424 for (int i = 0; i < ANV_DEVICE_EXTENSION_COUNT; i++) {
2425 if (device->supported_extensions.extensions[i]) {
2426 vk_outarray_append(&out, prop) {
2427 *prop = anv_device_extensions[i];
2428 }
2429 }
2430 }
2431
2432 return vk_outarray_status(&out);
2433 }
2434
2435 static void
2436 anv_device_init_dispatch(struct anv_device *device)
2437 {
2438 const struct anv_device_dispatch_table *genX_table;
2439 switch (device->info.gen) {
2440 case 12:
2441 genX_table = &gen12_device_dispatch_table;
2442 break;
2443 case 11:
2444 genX_table = &gen11_device_dispatch_table;
2445 break;
2446 case 10:
2447 genX_table = &gen10_device_dispatch_table;
2448 break;
2449 case 9:
2450 genX_table = &gen9_device_dispatch_table;
2451 break;
2452 case 8:
2453 genX_table = &gen8_device_dispatch_table;
2454 break;
2455 case 7:
2456 if (device->info.is_haswell)
2457 genX_table = &gen75_device_dispatch_table;
2458 else
2459 genX_table = &gen7_device_dispatch_table;
2460 break;
2461 default:
2462 unreachable("unsupported gen\n");
2463 }
2464
2465 for (unsigned i = 0; i < ARRAY_SIZE(device->dispatch.entrypoints); i++) {
2466 /* Vulkan requires that entrypoints for extensions which have not been
2467 * enabled must not be advertised.
2468 */
2469 if (!anv_device_entrypoint_is_enabled(i, device->instance->app_info.api_version,
2470 &device->instance->enabled_extensions,
2471 &device->enabled_extensions)) {
2472 device->dispatch.entrypoints[i] = NULL;
2473 } else if (genX_table->entrypoints[i]) {
2474 device->dispatch.entrypoints[i] = genX_table->entrypoints[i];
2475 } else {
2476 device->dispatch.entrypoints[i] =
2477 anv_device_dispatch_table.entrypoints[i];
2478 }
2479 }
2480 }
2481
2482 static int
2483 vk_priority_to_gen(int priority)
2484 {
2485 switch (priority) {
2486 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT:
2487 return GEN_CONTEXT_LOW_PRIORITY;
2488 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT:
2489 return GEN_CONTEXT_MEDIUM_PRIORITY;
2490 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT:
2491 return GEN_CONTEXT_HIGH_PRIORITY;
2492 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT:
2493 return GEN_CONTEXT_REALTIME_PRIORITY;
2494 default:
2495 unreachable("Invalid priority");
2496 }
2497 }
2498
2499 static VkResult
2500 anv_device_init_hiz_clear_value_bo(struct anv_device *device)
2501 {
2502 VkResult result = anv_device_alloc_bo(device, 4096,
2503 ANV_BO_ALLOC_MAPPED,
2504 0 /* explicit_address */,
2505 &device->hiz_clear_bo);
2506 if (result != VK_SUCCESS)
2507 return result;
2508
2509 union isl_color_value hiz_clear = { .u32 = { 0, } };
2510 hiz_clear.f32[0] = ANV_HZ_FC_VAL;
2511
2512 memcpy(device->hiz_clear_bo->map, hiz_clear.u32, sizeof(hiz_clear.u32));
2513
2514 if (!device->info.has_llc)
2515 gen_clflush_range(device->hiz_clear_bo->map, sizeof(hiz_clear.u32));
2516
2517 return VK_SUCCESS;
2518 }
2519
2520 static bool
2521 get_bo_from_pool(struct gen_batch_decode_bo *ret,
2522 struct anv_block_pool *pool,
2523 uint64_t address)
2524 {
2525 anv_block_pool_foreach_bo(bo, pool) {
2526 uint64_t bo_address = gen_48b_address(bo->offset);
2527 if (address >= bo_address && address < (bo_address + bo->size)) {
2528 *ret = (struct gen_batch_decode_bo) {
2529 .addr = bo_address,
2530 .size = bo->size,
2531 .map = bo->map,
2532 };
2533 return true;
2534 }
2535 }
2536 return false;
2537 }
2538
2539 /* Finding a buffer for batch decoding */
2540 static struct gen_batch_decode_bo
2541 decode_get_bo(void *v_batch, bool ppgtt, uint64_t address)
2542 {
2543 struct anv_device *device = v_batch;
2544 struct gen_batch_decode_bo ret_bo = {};
2545
2546 assert(ppgtt);
2547
2548 if (get_bo_from_pool(&ret_bo, &device->dynamic_state_pool.block_pool, address))
2549 return ret_bo;
2550 if (get_bo_from_pool(&ret_bo, &device->instruction_state_pool.block_pool, address))
2551 return ret_bo;
2552 if (get_bo_from_pool(&ret_bo, &device->binding_table_pool.block_pool, address))
2553 return ret_bo;
2554 if (get_bo_from_pool(&ret_bo, &device->surface_state_pool.block_pool, address))
2555 return ret_bo;
2556
2557 if (!device->cmd_buffer_being_decoded)
2558 return (struct gen_batch_decode_bo) { };
2559
2560 struct anv_batch_bo **bo;
2561
2562 u_vector_foreach(bo, &device->cmd_buffer_being_decoded->seen_bbos) {
2563 /* The decoder zeroes out the top 16 bits, so we need to as well */
2564 uint64_t bo_address = (*bo)->bo->offset & (~0ull >> 16);
2565
2566 if (address >= bo_address && address < bo_address + (*bo)->bo->size) {
2567 return (struct gen_batch_decode_bo) {
2568 .addr = bo_address,
2569 .size = (*bo)->bo->size,
2570 .map = (*bo)->bo->map,
2571 };
2572 }
2573 }
2574
2575 return (struct gen_batch_decode_bo) { };
2576 }
2577
2578 struct gen_aux_map_buffer {
2579 struct gen_buffer base;
2580 struct anv_state state;
2581 };
2582
2583 static struct gen_buffer *
2584 gen_aux_map_buffer_alloc(void *driver_ctx, uint32_t size)
2585 {
2586 struct gen_aux_map_buffer *buf = malloc(sizeof(struct gen_aux_map_buffer));
2587 if (!buf)
2588 return NULL;
2589
2590 struct anv_device *device = (struct anv_device*)driver_ctx;
2591 assert(device->instance->physicalDevice.supports_48bit_addresses &&
2592 device->instance->physicalDevice.use_softpin);
2593
2594 struct anv_state_pool *pool = &device->dynamic_state_pool;
2595 buf->state = anv_state_pool_alloc(pool, size, size);
2596
2597 buf->base.gpu = pool->block_pool.bo->offset + buf->state.offset;
2598 buf->base.gpu_end = buf->base.gpu + buf->state.alloc_size;
2599 buf->base.map = buf->state.map;
2600 buf->base.driver_bo = &buf->state;
2601 return &buf->base;
2602 }
2603
2604 static void
2605 gen_aux_map_buffer_free(void *driver_ctx, struct gen_buffer *buffer)
2606 {
2607 struct gen_aux_map_buffer *buf = (struct gen_aux_map_buffer*)buffer;
2608 struct anv_device *device = (struct anv_device*)driver_ctx;
2609 struct anv_state_pool *pool = &device->dynamic_state_pool;
2610 anv_state_pool_free(pool, buf->state);
2611 free(buf);
2612 }
2613
2614 static struct gen_mapped_pinned_buffer_alloc aux_map_allocator = {
2615 .alloc = gen_aux_map_buffer_alloc,
2616 .free = gen_aux_map_buffer_free,
2617 };
2618
2619 VkResult anv_CreateDevice(
2620 VkPhysicalDevice physicalDevice,
2621 const VkDeviceCreateInfo* pCreateInfo,
2622 const VkAllocationCallbacks* pAllocator,
2623 VkDevice* pDevice)
2624 {
2625 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
2626 VkResult result;
2627 struct anv_device *device;
2628
2629 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO);
2630
2631 struct anv_device_extension_table enabled_extensions = { };
2632 for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
2633 int idx;
2634 for (idx = 0; idx < ANV_DEVICE_EXTENSION_COUNT; idx++) {
2635 if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
2636 anv_device_extensions[idx].extensionName) == 0)
2637 break;
2638 }
2639
2640 if (idx >= ANV_DEVICE_EXTENSION_COUNT)
2641 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
2642
2643 if (!physical_device->supported_extensions.extensions[idx])
2644 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
2645
2646 enabled_extensions.extensions[idx] = true;
2647 }
2648
2649 /* Check enabled features */
2650 if (pCreateInfo->pEnabledFeatures) {
2651 VkPhysicalDeviceFeatures supported_features;
2652 anv_GetPhysicalDeviceFeatures(physicalDevice, &supported_features);
2653 VkBool32 *supported_feature = (VkBool32 *)&supported_features;
2654 VkBool32 *enabled_feature = (VkBool32 *)pCreateInfo->pEnabledFeatures;
2655 unsigned num_features = sizeof(VkPhysicalDeviceFeatures) / sizeof(VkBool32);
2656 for (uint32_t i = 0; i < num_features; i++) {
2657 if (enabled_feature[i] && !supported_feature[i])
2658 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
2659 }
2660 }
2661
2662 /* Check requested queues and fail if we are requested to create any
2663 * queues with flags we don't support.
2664 */
2665 assert(pCreateInfo->queueCreateInfoCount > 0);
2666 for (uint32_t i = 0; i < pCreateInfo->queueCreateInfoCount; i++) {
2667 if (pCreateInfo->pQueueCreateInfos[i].flags != 0)
2668 return vk_error(VK_ERROR_INITIALIZATION_FAILED);
2669 }
2670
2671 /* Check if client specified queue priority. */
2672 const VkDeviceQueueGlobalPriorityCreateInfoEXT *queue_priority =
2673 vk_find_struct_const(pCreateInfo->pQueueCreateInfos[0].pNext,
2674 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT);
2675
2676 VkQueueGlobalPriorityEXT priority =
2677 queue_priority ? queue_priority->globalPriority :
2678 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT;
2679
2680 device = vk_alloc2(&physical_device->instance->alloc, pAllocator,
2681 sizeof(*device), 8,
2682 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
2683 if (!device)
2684 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2685
2686 if (INTEL_DEBUG & DEBUG_BATCH) {
2687 const unsigned decode_flags =
2688 GEN_BATCH_DECODE_FULL |
2689 ((INTEL_DEBUG & DEBUG_COLOR) ? GEN_BATCH_DECODE_IN_COLOR : 0) |
2690 GEN_BATCH_DECODE_OFFSETS |
2691 GEN_BATCH_DECODE_FLOATS;
2692
2693 gen_batch_decode_ctx_init(&device->decoder_ctx,
2694 &physical_device->info,
2695 stderr, decode_flags, NULL,
2696 decode_get_bo, NULL, device);
2697 }
2698
2699 device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
2700 device->instance = physical_device->instance;
2701 device->chipset_id = physical_device->chipset_id;
2702 device->no_hw = physical_device->no_hw;
2703 device->_lost = false;
2704
2705 if (pAllocator)
2706 device->alloc = *pAllocator;
2707 else
2708 device->alloc = physical_device->instance->alloc;
2709
2710 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
2711 device->fd = open(physical_device->path, O_RDWR | O_CLOEXEC);
2712 if (device->fd == -1) {
2713 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2714 goto fail_device;
2715 }
2716
2717 device->context_id = anv_gem_create_context(device);
2718 if (device->context_id == -1) {
2719 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2720 goto fail_fd;
2721 }
2722
2723 result = anv_queue_init(device, &device->queue);
2724 if (result != VK_SUCCESS)
2725 goto fail_context_id;
2726
2727 if (physical_device->use_softpin) {
2728 if (pthread_mutex_init(&device->vma_mutex, NULL) != 0) {
2729 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2730 goto fail_queue;
2731 }
2732
2733 /* keep the page with address zero out of the allocator */
2734 util_vma_heap_init(&device->vma_lo,
2735 LOW_HEAP_MIN_ADDRESS, LOW_HEAP_SIZE);
2736
2737 util_vma_heap_init(&device->vma_cva, CLIENT_VISIBLE_HEAP_MIN_ADDRESS,
2738 CLIENT_VISIBLE_HEAP_SIZE);
2739
2740 /* Leave the last 4GiB out of the high vma range, so that no state
2741 * base address + size can overflow 48 bits. For more information see
2742 * the comment about Wa32bitGeneralStateOffset in anv_allocator.c
2743 */
2744 util_vma_heap_init(&device->vma_hi, HIGH_HEAP_MIN_ADDRESS,
2745 physical_device->gtt_size - (1ull << 32) -
2746 HIGH_HEAP_MIN_ADDRESS);
2747 }
2748
2749 list_inithead(&device->memory_objects);
2750
2751 /* As per spec, the driver implementation may deny requests to acquire
2752 * a priority above the default priority (MEDIUM) if the caller does not
2753 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
2754 * is returned.
2755 */
2756 if (physical_device->has_context_priority) {
2757 int err = anv_gem_set_context_param(device->fd, device->context_id,
2758 I915_CONTEXT_PARAM_PRIORITY,
2759 vk_priority_to_gen(priority));
2760 if (err != 0 && priority > VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT) {
2761 result = vk_error(VK_ERROR_NOT_PERMITTED_EXT);
2762 goto fail_vmas;
2763 }
2764 }
2765
2766 device->info = physical_device->info;
2767 device->isl_dev = physical_device->isl_dev;
2768
2769 /* On Broadwell and later, we can use batch chaining to more efficiently
2770 * implement growing command buffers. Prior to Haswell, the kernel
2771 * command parser gets in the way and we have to fall back to growing
2772 * the batch.
2773 */
2774 device->can_chain_batches = device->info.gen >= 8;
2775
2776 device->robust_buffer_access = pCreateInfo->pEnabledFeatures &&
2777 pCreateInfo->pEnabledFeatures->robustBufferAccess;
2778 device->enabled_extensions = enabled_extensions;
2779
2780 anv_device_init_dispatch(device);
2781
2782 if (pthread_mutex_init(&device->mutex, NULL) != 0) {
2783 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2784 goto fail_queue;
2785 }
2786
2787 pthread_condattr_t condattr;
2788 if (pthread_condattr_init(&condattr) != 0) {
2789 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2790 goto fail_mutex;
2791 }
2792 if (pthread_condattr_setclock(&condattr, CLOCK_MONOTONIC) != 0) {
2793 pthread_condattr_destroy(&condattr);
2794 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2795 goto fail_mutex;
2796 }
2797 if (pthread_cond_init(&device->queue_submit, &condattr) != 0) {
2798 pthread_condattr_destroy(&condattr);
2799 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2800 goto fail_mutex;
2801 }
2802 pthread_condattr_destroy(&condattr);
2803
2804 result = anv_bo_cache_init(&device->bo_cache);
2805 if (result != VK_SUCCESS)
2806 goto fail_queue_cond;
2807
2808 anv_bo_pool_init(&device->batch_bo_pool, device);
2809
2810 result = anv_state_pool_init(&device->dynamic_state_pool, device,
2811 DYNAMIC_STATE_POOL_MIN_ADDRESS, 16384);
2812 if (result != VK_SUCCESS)
2813 goto fail_batch_bo_pool;
2814
2815 result = anv_state_pool_init(&device->instruction_state_pool, device,
2816 INSTRUCTION_STATE_POOL_MIN_ADDRESS, 16384);
2817 if (result != VK_SUCCESS)
2818 goto fail_dynamic_state_pool;
2819
2820 result = anv_state_pool_init(&device->surface_state_pool, device,
2821 SURFACE_STATE_POOL_MIN_ADDRESS, 4096);
2822 if (result != VK_SUCCESS)
2823 goto fail_instruction_state_pool;
2824
2825 if (physical_device->use_softpin) {
2826 result = anv_state_pool_init(&device->binding_table_pool, device,
2827 BINDING_TABLE_POOL_MIN_ADDRESS, 4096);
2828 if (result != VK_SUCCESS)
2829 goto fail_surface_state_pool;
2830 }
2831
2832 if (device->info.gen >= 12) {
2833 device->aux_map_ctx = gen_aux_map_init(device, &aux_map_allocator,
2834 &physical_device->info);
2835 if (!device->aux_map_ctx)
2836 goto fail_binding_table_pool;
2837 }
2838
2839 result = anv_device_alloc_bo(device, 4096, 0 /* flags */,
2840 0 /* explicit_address */,
2841 &device->workaround_bo);
2842 if (result != VK_SUCCESS)
2843 goto fail_surface_aux_map_pool;
2844
2845 result = anv_device_init_trivial_batch(device);
2846 if (result != VK_SUCCESS)
2847 goto fail_workaround_bo;
2848
2849 if (device->info.gen >= 10) {
2850 result = anv_device_init_hiz_clear_value_bo(device);
2851 if (result != VK_SUCCESS)
2852 goto fail_trivial_batch_bo;
2853 }
2854
2855 anv_scratch_pool_init(device, &device->scratch_pool);
2856
2857 switch (device->info.gen) {
2858 case 7:
2859 if (!device->info.is_haswell)
2860 result = gen7_init_device_state(device);
2861 else
2862 result = gen75_init_device_state(device);
2863 break;
2864 case 8:
2865 result = gen8_init_device_state(device);
2866 break;
2867 case 9:
2868 result = gen9_init_device_state(device);
2869 break;
2870 case 10:
2871 result = gen10_init_device_state(device);
2872 break;
2873 case 11:
2874 result = gen11_init_device_state(device);
2875 break;
2876 case 12:
2877 result = gen12_init_device_state(device);
2878 break;
2879 default:
2880 /* Shouldn't get here as we don't create physical devices for any other
2881 * gens. */
2882 unreachable("unhandled gen");
2883 }
2884 if (result != VK_SUCCESS)
2885 goto fail_workaround_bo;
2886
2887 anv_pipeline_cache_init(&device->default_pipeline_cache, device, true);
2888
2889 anv_device_init_blorp(device);
2890
2891 anv_device_init_border_colors(device);
2892
2893 anv_device_perf_init(device);
2894
2895 *pDevice = anv_device_to_handle(device);
2896
2897 return VK_SUCCESS;
2898
2899 fail_workaround_bo:
2900 anv_scratch_pool_finish(device, &device->scratch_pool);
2901 if (device->info.gen >= 10)
2902 anv_device_release_bo(device, device->hiz_clear_bo);
2903 anv_device_release_bo(device, device->workaround_bo);
2904 fail_trivial_batch_bo:
2905 anv_device_release_bo(device, device->trivial_batch_bo);
2906 fail_surface_aux_map_pool:
2907 if (device->info.gen >= 12) {
2908 gen_aux_map_finish(device->aux_map_ctx);
2909 device->aux_map_ctx = NULL;
2910 }
2911 fail_binding_table_pool:
2912 if (physical_device->use_softpin)
2913 anv_state_pool_finish(&device->binding_table_pool);
2914 fail_surface_state_pool:
2915 anv_state_pool_finish(&device->surface_state_pool);
2916 fail_instruction_state_pool:
2917 anv_state_pool_finish(&device->instruction_state_pool);
2918 fail_dynamic_state_pool:
2919 anv_state_pool_finish(&device->dynamic_state_pool);
2920 fail_batch_bo_pool:
2921 anv_bo_pool_finish(&device->batch_bo_pool);
2922 anv_bo_cache_finish(&device->bo_cache);
2923 fail_queue_cond:
2924 pthread_cond_destroy(&device->queue_submit);
2925 fail_mutex:
2926 pthread_mutex_destroy(&device->mutex);
2927 fail_vmas:
2928 if (physical_device->use_softpin) {
2929 util_vma_heap_finish(&device->vma_hi);
2930 util_vma_heap_finish(&device->vma_cva);
2931 util_vma_heap_finish(&device->vma_lo);
2932 }
2933 fail_queue:
2934 anv_queue_finish(&device->queue);
2935 fail_context_id:
2936 anv_gem_destroy_context(device, device->context_id);
2937 fail_fd:
2938 close(device->fd);
2939 fail_device:
2940 vk_free(&device->alloc, device);
2941
2942 return result;
2943 }
2944
2945 void anv_DestroyDevice(
2946 VkDevice _device,
2947 const VkAllocationCallbacks* pAllocator)
2948 {
2949 ANV_FROM_HANDLE(anv_device, device, _device);
2950 struct anv_physical_device *physical_device;
2951
2952 if (!device)
2953 return;
2954
2955 physical_device = &device->instance->physicalDevice;
2956
2957 anv_device_finish_blorp(device);
2958
2959 anv_pipeline_cache_finish(&device->default_pipeline_cache);
2960
2961 anv_queue_finish(&device->queue);
2962
2963 #ifdef HAVE_VALGRIND
2964 /* We only need to free these to prevent valgrind errors. The backing
2965 * BO will go away in a couple of lines so we don't actually leak.
2966 */
2967 anv_state_pool_free(&device->dynamic_state_pool, device->border_colors);
2968 anv_state_pool_free(&device->dynamic_state_pool, device->slice_hash);
2969 #endif
2970
2971 anv_scratch_pool_finish(device, &device->scratch_pool);
2972
2973 anv_device_release_bo(device, device->workaround_bo);
2974 anv_device_release_bo(device, device->trivial_batch_bo);
2975 if (device->info.gen >= 10)
2976 anv_device_release_bo(device, device->hiz_clear_bo);
2977
2978 if (device->info.gen >= 12) {
2979 gen_aux_map_finish(device->aux_map_ctx);
2980 device->aux_map_ctx = NULL;
2981 }
2982
2983 if (physical_device->use_softpin)
2984 anv_state_pool_finish(&device->binding_table_pool);
2985 anv_state_pool_finish(&device->surface_state_pool);
2986 anv_state_pool_finish(&device->instruction_state_pool);
2987 anv_state_pool_finish(&device->dynamic_state_pool);
2988
2989 anv_bo_pool_finish(&device->batch_bo_pool);
2990
2991 anv_bo_cache_finish(&device->bo_cache);
2992
2993 if (physical_device->use_softpin) {
2994 util_vma_heap_finish(&device->vma_hi);
2995 util_vma_heap_finish(&device->vma_cva);
2996 util_vma_heap_finish(&device->vma_lo);
2997 }
2998
2999 pthread_cond_destroy(&device->queue_submit);
3000 pthread_mutex_destroy(&device->mutex);
3001
3002 anv_gem_destroy_context(device, device->context_id);
3003
3004 if (INTEL_DEBUG & DEBUG_BATCH)
3005 gen_batch_decode_ctx_finish(&device->decoder_ctx);
3006
3007 close(device->fd);
3008
3009 vk_free(&device->alloc, device);
3010 }
3011
3012 VkResult anv_EnumerateInstanceLayerProperties(
3013 uint32_t* pPropertyCount,
3014 VkLayerProperties* pProperties)
3015 {
3016 if (pProperties == NULL) {
3017 *pPropertyCount = 0;
3018 return VK_SUCCESS;
3019 }
3020
3021 /* None supported at this time */
3022 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
3023 }
3024
3025 VkResult anv_EnumerateDeviceLayerProperties(
3026 VkPhysicalDevice physicalDevice,
3027 uint32_t* pPropertyCount,
3028 VkLayerProperties* pProperties)
3029 {
3030 if (pProperties == NULL) {
3031 *pPropertyCount = 0;
3032 return VK_SUCCESS;
3033 }
3034
3035 /* None supported at this time */
3036 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
3037 }
3038
3039 void anv_GetDeviceQueue(
3040 VkDevice _device,
3041 uint32_t queueNodeIndex,
3042 uint32_t queueIndex,
3043 VkQueue* pQueue)
3044 {
3045 const VkDeviceQueueInfo2 info = {
3046 .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_INFO_2,
3047 .pNext = NULL,
3048 .flags = 0,
3049 .queueFamilyIndex = queueNodeIndex,
3050 .queueIndex = queueIndex,
3051 };
3052
3053 anv_GetDeviceQueue2(_device, &info, pQueue);
3054 }
3055
3056 void anv_GetDeviceQueue2(
3057 VkDevice _device,
3058 const VkDeviceQueueInfo2* pQueueInfo,
3059 VkQueue* pQueue)
3060 {
3061 ANV_FROM_HANDLE(anv_device, device, _device);
3062
3063 assert(pQueueInfo->queueIndex == 0);
3064
3065 if (pQueueInfo->flags == device->queue.flags)
3066 *pQueue = anv_queue_to_handle(&device->queue);
3067 else
3068 *pQueue = NULL;
3069 }
3070
3071 VkResult
3072 _anv_device_set_lost(struct anv_device *device,
3073 const char *file, int line,
3074 const char *msg, ...)
3075 {
3076 VkResult err;
3077 va_list ap;
3078
3079 p_atomic_inc(&device->_lost);
3080
3081 va_start(ap, msg);
3082 err = __vk_errorv(device->instance, device,
3083 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT,
3084 VK_ERROR_DEVICE_LOST, file, line, msg, ap);
3085 va_end(ap);
3086
3087 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
3088 abort();
3089
3090 return err;
3091 }
3092
3093 VkResult
3094 _anv_queue_set_lost(struct anv_queue *queue,
3095 const char *file, int line,
3096 const char *msg, ...)
3097 {
3098 VkResult err;
3099 va_list ap;
3100
3101 p_atomic_inc(&queue->device->_lost);
3102
3103 va_start(ap, msg);
3104 err = __vk_errorv(queue->device->instance, queue->device,
3105 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT,
3106 VK_ERROR_DEVICE_LOST, file, line, msg, ap);
3107 va_end(ap);
3108
3109 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
3110 abort();
3111
3112 return err;
3113 }
3114
3115 VkResult
3116 anv_device_query_status(struct anv_device *device)
3117 {
3118 /* This isn't likely as most of the callers of this function already check
3119 * for it. However, it doesn't hurt to check and it potentially lets us
3120 * avoid an ioctl.
3121 */
3122 if (anv_device_is_lost(device))
3123 return VK_ERROR_DEVICE_LOST;
3124
3125 uint32_t active, pending;
3126 int ret = anv_gem_gpu_get_reset_stats(device, &active, &pending);
3127 if (ret == -1) {
3128 /* We don't know the real error. */
3129 return anv_device_set_lost(device, "get_reset_stats failed: %m");
3130 }
3131
3132 if (active) {
3133 return anv_device_set_lost(device, "GPU hung on one of our command buffers");
3134 } else if (pending) {
3135 return anv_device_set_lost(device, "GPU hung with commands in-flight");
3136 }
3137
3138 return VK_SUCCESS;
3139 }
3140
3141 VkResult
3142 anv_device_bo_busy(struct anv_device *device, struct anv_bo *bo)
3143 {
3144 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
3145 * Other usages of the BO (such as on different hardware) will not be
3146 * flagged as "busy" by this ioctl. Use with care.
3147 */
3148 int ret = anv_gem_busy(device, bo->gem_handle);
3149 if (ret == 1) {
3150 return VK_NOT_READY;
3151 } else if (ret == -1) {
3152 /* We don't know the real error. */
3153 return anv_device_set_lost(device, "gem wait failed: %m");
3154 }
3155
3156 /* Query for device status after the busy call. If the BO we're checking
3157 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
3158 * client because it clearly doesn't have valid data. Yes, this most
3159 * likely means an ioctl, but we just did an ioctl to query the busy status
3160 * so it's no great loss.
3161 */
3162 return anv_device_query_status(device);
3163 }
3164
3165 VkResult
3166 anv_device_wait(struct anv_device *device, struct anv_bo *bo,
3167 int64_t timeout)
3168 {
3169 int ret = anv_gem_wait(device, bo->gem_handle, &timeout);
3170 if (ret == -1 && errno == ETIME) {
3171 return VK_TIMEOUT;
3172 } else if (ret == -1) {
3173 /* We don't know the real error. */
3174 return anv_device_set_lost(device, "gem wait failed: %m");
3175 }
3176
3177 /* Query for device status after the wait. If the BO we're waiting on got
3178 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
3179 * because it clearly doesn't have valid data. Yes, this most likely means
3180 * an ioctl, but we just did an ioctl to wait so it's no great loss.
3181 */
3182 return anv_device_query_status(device);
3183 }
3184
3185 VkResult anv_DeviceWaitIdle(
3186 VkDevice _device)
3187 {
3188 ANV_FROM_HANDLE(anv_device, device, _device);
3189
3190 if (anv_device_is_lost(device))
3191 return VK_ERROR_DEVICE_LOST;
3192
3193 return anv_queue_submit_simple_batch(&device->queue, NULL);
3194 }
3195
3196 bool
3197 anv_vma_alloc(struct anv_device *device, struct anv_bo *bo,
3198 uint64_t client_address)
3199 {
3200 const struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3201 const struct gen_device_info *devinfo = &pdevice->info;
3202 /* Gen12 CCS surface addresses need to be 64K aligned. We have no way of
3203 * telling what this allocation is for so pick the largest alignment.
3204 */
3205 const uint32_t vma_alignment =
3206 devinfo->gen >= 12 ? (64 * 1024) : (4 * 1024);
3207
3208 if (!(bo->flags & EXEC_OBJECT_PINNED)) {
3209 assert(!(bo->has_client_visible_address));
3210 return true;
3211 }
3212
3213 pthread_mutex_lock(&device->vma_mutex);
3214
3215 bo->offset = 0;
3216
3217 if (bo->has_client_visible_address) {
3218 assert(bo->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS);
3219 if (client_address) {
3220 if (util_vma_heap_alloc_addr(&device->vma_cva,
3221 client_address, bo->size)) {
3222 bo->offset = gen_canonical_address(client_address);
3223 }
3224 } else {
3225 uint64_t addr =
3226 util_vma_heap_alloc(&device->vma_cva, bo->size, vma_alignment);
3227 if (addr) {
3228 bo->offset = gen_canonical_address(addr);
3229 assert(addr == gen_48b_address(bo->offset));
3230 }
3231 }
3232 /* We don't want to fall back to other heaps */
3233 goto done;
3234 }
3235
3236 assert(client_address == 0);
3237
3238 if (bo->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) {
3239 uint64_t addr =
3240 util_vma_heap_alloc(&device->vma_hi, bo->size, vma_alignment);
3241 if (addr) {
3242 bo->offset = gen_canonical_address(addr);
3243 assert(addr == gen_48b_address(bo->offset));
3244 }
3245 }
3246
3247 if (bo->offset == 0) {
3248 uint64_t addr =
3249 util_vma_heap_alloc(&device->vma_lo, bo->size, vma_alignment);
3250 if (addr) {
3251 bo->offset = gen_canonical_address(addr);
3252 assert(addr == gen_48b_address(bo->offset));
3253 }
3254 }
3255
3256 done:
3257 pthread_mutex_unlock(&device->vma_mutex);
3258
3259 return bo->offset != 0;
3260 }
3261
3262 void
3263 anv_vma_free(struct anv_device *device, struct anv_bo *bo)
3264 {
3265 if (!(bo->flags & EXEC_OBJECT_PINNED))
3266 return;
3267
3268 const uint64_t addr_48b = gen_48b_address(bo->offset);
3269
3270 pthread_mutex_lock(&device->vma_mutex);
3271
3272 if (addr_48b >= LOW_HEAP_MIN_ADDRESS &&
3273 addr_48b <= LOW_HEAP_MAX_ADDRESS) {
3274 util_vma_heap_free(&device->vma_lo, addr_48b, bo->size);
3275 } else if (addr_48b >= CLIENT_VISIBLE_HEAP_MIN_ADDRESS &&
3276 addr_48b <= CLIENT_VISIBLE_HEAP_MAX_ADDRESS) {
3277 util_vma_heap_free(&device->vma_cva, addr_48b, bo->size);
3278 } else {
3279 assert(addr_48b >= HIGH_HEAP_MIN_ADDRESS);
3280 util_vma_heap_free(&device->vma_hi, addr_48b, bo->size);
3281 }
3282
3283 pthread_mutex_unlock(&device->vma_mutex);
3284
3285 bo->offset = 0;
3286 }
3287
3288 VkResult anv_AllocateMemory(
3289 VkDevice _device,
3290 const VkMemoryAllocateInfo* pAllocateInfo,
3291 const VkAllocationCallbacks* pAllocator,
3292 VkDeviceMemory* pMem)
3293 {
3294 ANV_FROM_HANDLE(anv_device, device, _device);
3295 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3296 struct anv_device_memory *mem;
3297 VkResult result = VK_SUCCESS;
3298
3299 assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
3300
3301 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
3302 assert(pAllocateInfo->allocationSize > 0);
3303
3304 VkDeviceSize aligned_alloc_size =
3305 align_u64(pAllocateInfo->allocationSize, 4096);
3306
3307 if (aligned_alloc_size > MAX_MEMORY_ALLOCATION_SIZE)
3308 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
3309
3310 assert(pAllocateInfo->memoryTypeIndex < pdevice->memory.type_count);
3311 struct anv_memory_type *mem_type =
3312 &pdevice->memory.types[pAllocateInfo->memoryTypeIndex];
3313 assert(mem_type->heapIndex < pdevice->memory.heap_count);
3314 struct anv_memory_heap *mem_heap =
3315 &pdevice->memory.heaps[mem_type->heapIndex];
3316
3317 uint64_t mem_heap_used = p_atomic_read(&mem_heap->used);
3318 if (mem_heap_used + aligned_alloc_size > mem_heap->size)
3319 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
3320
3321 mem = vk_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
3322 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
3323 if (mem == NULL)
3324 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
3325
3326 assert(pAllocateInfo->memoryTypeIndex < pdevice->memory.type_count);
3327 mem->type = mem_type;
3328 mem->map = NULL;
3329 mem->map_size = 0;
3330 mem->ahw = NULL;
3331 mem->host_ptr = NULL;
3332
3333 enum anv_bo_alloc_flags alloc_flags = 0;
3334
3335 const VkExportMemoryAllocateInfo *export_info = NULL;
3336 const VkImportAndroidHardwareBufferInfoANDROID *ahw_import_info = NULL;
3337 const VkImportMemoryFdInfoKHR *fd_info = NULL;
3338 const VkImportMemoryHostPointerInfoEXT *host_ptr_info = NULL;
3339 const VkMemoryDedicatedAllocateInfo *dedicated_info = NULL;
3340 VkMemoryAllocateFlags vk_flags = 0;
3341 uint64_t client_address = 0;
3342
3343 vk_foreach_struct_const(ext, pAllocateInfo->pNext) {
3344 switch (ext->sType) {
3345 case VK_STRUCTURE_TYPE_EXPORT_MEMORY_ALLOCATE_INFO:
3346 export_info = (void *)ext;
3347 break;
3348
3349 case VK_STRUCTURE_TYPE_IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID:
3350 ahw_import_info = (void *)ext;
3351 break;
3352
3353 case VK_STRUCTURE_TYPE_IMPORT_MEMORY_FD_INFO_KHR:
3354 fd_info = (void *)ext;
3355 break;
3356
3357 case VK_STRUCTURE_TYPE_IMPORT_MEMORY_HOST_POINTER_INFO_EXT:
3358 host_ptr_info = (void *)ext;
3359 break;
3360
3361 case VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_FLAGS_INFO: {
3362 const VkMemoryAllocateFlagsInfo *flags_info = (void *)ext;
3363 vk_flags = flags_info->flags;
3364 break;
3365 }
3366
3367 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_ALLOCATE_INFO:
3368 dedicated_info = (void *)ext;
3369 break;
3370
3371 case VK_STRUCTURE_TYPE_MEMORY_OPAQUE_CAPTURE_ADDRESS_ALLOCATE_INFO_KHR: {
3372 const VkMemoryOpaqueCaptureAddressAllocateInfoKHR *addr_info =
3373 (const VkMemoryOpaqueCaptureAddressAllocateInfoKHR *)ext;
3374 client_address = addr_info->opaqueCaptureAddress;
3375 break;
3376 }
3377
3378 default:
3379 anv_debug_ignored_stype(ext->sType);
3380 break;
3381 }
3382 }
3383
3384 if (vk_flags & VK_MEMORY_ALLOCATE_DEVICE_ADDRESS_BIT_KHR)
3385 alloc_flags |= ANV_BO_ALLOC_CLIENT_VISIBLE_ADDRESS;
3386
3387 /* Check if we need to support Android HW buffer export. If so,
3388 * create AHardwareBuffer and import memory from it.
3389 */
3390 bool android_export = false;
3391 if (export_info && export_info->handleTypes &
3392 VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID)
3393 android_export = true;
3394
3395 if (ahw_import_info) {
3396 result = anv_import_ahw_memory(_device, mem, ahw_import_info);
3397 if (result != VK_SUCCESS)
3398 goto fail;
3399
3400 goto success;
3401 } else if (android_export) {
3402 result = anv_create_ahw_memory(_device, mem, pAllocateInfo);
3403 if (result != VK_SUCCESS)
3404 goto fail;
3405
3406 const VkImportAndroidHardwareBufferInfoANDROID import_info = {
3407 .buffer = mem->ahw,
3408 };
3409 result = anv_import_ahw_memory(_device, mem, &import_info);
3410 if (result != VK_SUCCESS)
3411 goto fail;
3412
3413 goto success;
3414 }
3415
3416 /* The Vulkan spec permits handleType to be 0, in which case the struct is
3417 * ignored.
3418 */
3419 if (fd_info && fd_info->handleType) {
3420 /* At the moment, we support only the below handle types. */
3421 assert(fd_info->handleType ==
3422 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
3423 fd_info->handleType ==
3424 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
3425
3426 result = anv_device_import_bo(device, fd_info->fd, alloc_flags,
3427 client_address, &mem->bo);
3428 if (result != VK_SUCCESS)
3429 goto fail;
3430
3431 VkDeviceSize aligned_alloc_size =
3432 align_u64(pAllocateInfo->allocationSize, 4096);
3433
3434 /* For security purposes, we reject importing the bo if it's smaller
3435 * than the requested allocation size. This prevents a malicious client
3436 * from passing a buffer to a trusted client, lying about the size, and
3437 * telling the trusted client to try and texture from an image that goes
3438 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
3439 * in the trusted client. The trusted client can protect itself against
3440 * this sort of attack but only if it can trust the buffer size.
3441 */
3442 if (mem->bo->size < aligned_alloc_size) {
3443 result = vk_errorf(device->instance, device,
3444 VK_ERROR_INVALID_EXTERNAL_HANDLE,
3445 "aligned allocationSize too large for "
3446 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT: "
3447 "%"PRIu64"B > %"PRIu64"B",
3448 aligned_alloc_size, mem->bo->size);
3449 anv_device_release_bo(device, mem->bo);
3450 goto fail;
3451 }
3452
3453 /* From the Vulkan spec:
3454 *
3455 * "Importing memory from a file descriptor transfers ownership of
3456 * the file descriptor from the application to the Vulkan
3457 * implementation. The application must not perform any operations on
3458 * the file descriptor after a successful import."
3459 *
3460 * If the import fails, we leave the file descriptor open.
3461 */
3462 close(fd_info->fd);
3463 goto success;
3464 }
3465
3466 if (host_ptr_info && host_ptr_info->handleType) {
3467 if (host_ptr_info->handleType ==
3468 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_MAPPED_FOREIGN_MEMORY_BIT_EXT) {
3469 result = vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE);
3470 goto fail;
3471 }
3472
3473 assert(host_ptr_info->handleType ==
3474 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT);
3475
3476 result = anv_device_import_bo_from_host_ptr(device,
3477 host_ptr_info->pHostPointer,
3478 pAllocateInfo->allocationSize,
3479 alloc_flags,
3480 client_address,
3481 &mem->bo);
3482 if (result != VK_SUCCESS)
3483 goto fail;
3484
3485 mem->host_ptr = host_ptr_info->pHostPointer;
3486 goto success;
3487 }
3488
3489 /* Regular allocate (not importing memory). */
3490
3491 if (export_info && export_info->handleTypes)
3492 alloc_flags |= ANV_BO_ALLOC_EXTERNAL;
3493
3494 result = anv_device_alloc_bo(device, pAllocateInfo->allocationSize,
3495 alloc_flags, client_address, &mem->bo);
3496 if (result != VK_SUCCESS)
3497 goto fail;
3498
3499 if (dedicated_info && dedicated_info->image != VK_NULL_HANDLE) {
3500 ANV_FROM_HANDLE(anv_image, image, dedicated_info->image);
3501
3502 /* Some legacy (non-modifiers) consumers need the tiling to be set on
3503 * the BO. In this case, we have a dedicated allocation.
3504 */
3505 if (image->needs_set_tiling) {
3506 const uint32_t i915_tiling =
3507 isl_tiling_to_i915_tiling(image->planes[0].surface.isl.tiling);
3508 int ret = anv_gem_set_tiling(device, mem->bo->gem_handle,
3509 image->planes[0].surface.isl.row_pitch_B,
3510 i915_tiling);
3511 if (ret) {
3512 anv_device_release_bo(device, mem->bo);
3513 result = vk_errorf(device->instance, NULL,
3514 VK_ERROR_OUT_OF_DEVICE_MEMORY,
3515 "failed to set BO tiling: %m");
3516 goto fail;
3517 }
3518 }
3519 }
3520
3521 success:
3522 mem_heap_used = p_atomic_add_return(&mem_heap->used, mem->bo->size);
3523 if (mem_heap_used > mem_heap->size) {
3524 p_atomic_add(&mem_heap->used, -mem->bo->size);
3525 anv_device_release_bo(device, mem->bo);
3526 result = vk_errorf(device->instance, NULL,
3527 VK_ERROR_OUT_OF_DEVICE_MEMORY,
3528 "Out of heap memory");
3529 goto fail;
3530 }
3531
3532 pthread_mutex_lock(&device->mutex);
3533 list_addtail(&mem->link, &device->memory_objects);
3534 pthread_mutex_unlock(&device->mutex);
3535
3536 *pMem = anv_device_memory_to_handle(mem);
3537
3538 return VK_SUCCESS;
3539
3540 fail:
3541 vk_free2(&device->alloc, pAllocator, mem);
3542
3543 return result;
3544 }
3545
3546 VkResult anv_GetMemoryFdKHR(
3547 VkDevice device_h,
3548 const VkMemoryGetFdInfoKHR* pGetFdInfo,
3549 int* pFd)
3550 {
3551 ANV_FROM_HANDLE(anv_device, dev, device_h);
3552 ANV_FROM_HANDLE(anv_device_memory, mem, pGetFdInfo->memory);
3553
3554 assert(pGetFdInfo->sType == VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR);
3555
3556 assert(pGetFdInfo->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
3557 pGetFdInfo->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
3558
3559 return anv_device_export_bo(dev, mem->bo, pFd);
3560 }
3561
3562 VkResult anv_GetMemoryFdPropertiesKHR(
3563 VkDevice _device,
3564 VkExternalMemoryHandleTypeFlagBits handleType,
3565 int fd,
3566 VkMemoryFdPropertiesKHR* pMemoryFdProperties)
3567 {
3568 ANV_FROM_HANDLE(anv_device, device, _device);
3569 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3570
3571 switch (handleType) {
3572 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT:
3573 /* dma-buf can be imported as any memory type */
3574 pMemoryFdProperties->memoryTypeBits =
3575 (1 << pdevice->memory.type_count) - 1;
3576 return VK_SUCCESS;
3577
3578 default:
3579 /* The valid usage section for this function says:
3580 *
3581 * "handleType must not be one of the handle types defined as
3582 * opaque."
3583 *
3584 * So opaque handle types fall into the default "unsupported" case.
3585 */
3586 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE);
3587 }
3588 }
3589
3590 VkResult anv_GetMemoryHostPointerPropertiesEXT(
3591 VkDevice _device,
3592 VkExternalMemoryHandleTypeFlagBits handleType,
3593 const void* pHostPointer,
3594 VkMemoryHostPointerPropertiesEXT* pMemoryHostPointerProperties)
3595 {
3596 ANV_FROM_HANDLE(anv_device, device, _device);
3597
3598 assert(pMemoryHostPointerProperties->sType ==
3599 VK_STRUCTURE_TYPE_MEMORY_HOST_POINTER_PROPERTIES_EXT);
3600
3601 switch (handleType) {
3602 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT: {
3603 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3604
3605 /* Host memory can be imported as any memory type. */
3606 pMemoryHostPointerProperties->memoryTypeBits =
3607 (1ull << pdevice->memory.type_count) - 1;
3608
3609 return VK_SUCCESS;
3610 }
3611 default:
3612 return VK_ERROR_INVALID_EXTERNAL_HANDLE;
3613 }
3614 }
3615
3616 void anv_FreeMemory(
3617 VkDevice _device,
3618 VkDeviceMemory _mem,
3619 const VkAllocationCallbacks* pAllocator)
3620 {
3621 ANV_FROM_HANDLE(anv_device, device, _device);
3622 ANV_FROM_HANDLE(anv_device_memory, mem, _mem);
3623 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3624
3625 if (mem == NULL)
3626 return;
3627
3628 pthread_mutex_lock(&device->mutex);
3629 list_del(&mem->link);
3630 pthread_mutex_unlock(&device->mutex);
3631
3632 if (mem->map)
3633 anv_UnmapMemory(_device, _mem);
3634
3635 p_atomic_add(&pdevice->memory.heaps[mem->type->heapIndex].used,
3636 -mem->bo->size);
3637
3638 anv_device_release_bo(device, mem->bo);
3639
3640 #if defined(ANDROID) && ANDROID_API_LEVEL >= 26
3641 if (mem->ahw)
3642 AHardwareBuffer_release(mem->ahw);
3643 #endif
3644
3645 vk_free2(&device->alloc, pAllocator, mem);
3646 }
3647
3648 VkResult anv_MapMemory(
3649 VkDevice _device,
3650 VkDeviceMemory _memory,
3651 VkDeviceSize offset,
3652 VkDeviceSize size,
3653 VkMemoryMapFlags flags,
3654 void** ppData)
3655 {
3656 ANV_FROM_HANDLE(anv_device, device, _device);
3657 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
3658
3659 if (mem == NULL) {
3660 *ppData = NULL;
3661 return VK_SUCCESS;
3662 }
3663
3664 if (mem->host_ptr) {
3665 *ppData = mem->host_ptr + offset;
3666 return VK_SUCCESS;
3667 }
3668
3669 if (size == VK_WHOLE_SIZE)
3670 size = mem->bo->size - offset;
3671
3672 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
3673 *
3674 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
3675 * assert(size != 0);
3676 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
3677 * equal to the size of the memory minus offset
3678 */
3679 assert(size > 0);
3680 assert(offset + size <= mem->bo->size);
3681
3682 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
3683 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
3684 * at a time is valid. We could just mmap up front and return an offset
3685 * pointer here, but that may exhaust virtual memory on 32 bit
3686 * userspace. */
3687
3688 uint32_t gem_flags = 0;
3689
3690 if (!device->info.has_llc &&
3691 (mem->type->propertyFlags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT))
3692 gem_flags |= I915_MMAP_WC;
3693
3694 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
3695 uint64_t map_offset = offset & ~4095ull;
3696 assert(offset >= map_offset);
3697 uint64_t map_size = (offset + size) - map_offset;
3698
3699 /* Let's map whole pages */
3700 map_size = align_u64(map_size, 4096);
3701
3702 void *map = anv_gem_mmap(device, mem->bo->gem_handle,
3703 map_offset, map_size, gem_flags);
3704 if (map == MAP_FAILED)
3705 return vk_error(VK_ERROR_MEMORY_MAP_FAILED);
3706
3707 mem->map = map;
3708 mem->map_size = map_size;
3709
3710 *ppData = mem->map + (offset - map_offset);
3711
3712 return VK_SUCCESS;
3713 }
3714
3715 void anv_UnmapMemory(
3716 VkDevice _device,
3717 VkDeviceMemory _memory)
3718 {
3719 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
3720
3721 if (mem == NULL || mem->host_ptr)
3722 return;
3723
3724 anv_gem_munmap(mem->map, mem->map_size);
3725
3726 mem->map = NULL;
3727 mem->map_size = 0;
3728 }
3729
3730 static void
3731 clflush_mapped_ranges(struct anv_device *device,
3732 uint32_t count,
3733 const VkMappedMemoryRange *ranges)
3734 {
3735 for (uint32_t i = 0; i < count; i++) {
3736 ANV_FROM_HANDLE(anv_device_memory, mem, ranges[i].memory);
3737 if (ranges[i].offset >= mem->map_size)
3738 continue;
3739
3740 gen_clflush_range(mem->map + ranges[i].offset,
3741 MIN2(ranges[i].size, mem->map_size - ranges[i].offset));
3742 }
3743 }
3744
3745 VkResult anv_FlushMappedMemoryRanges(
3746 VkDevice _device,
3747 uint32_t memoryRangeCount,
3748 const VkMappedMemoryRange* pMemoryRanges)
3749 {
3750 ANV_FROM_HANDLE(anv_device, device, _device);
3751
3752 if (device->info.has_llc)
3753 return VK_SUCCESS;
3754
3755 /* Make sure the writes we're flushing have landed. */
3756 __builtin_ia32_mfence();
3757
3758 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
3759
3760 return VK_SUCCESS;
3761 }
3762
3763 VkResult anv_InvalidateMappedMemoryRanges(
3764 VkDevice _device,
3765 uint32_t memoryRangeCount,
3766 const VkMappedMemoryRange* pMemoryRanges)
3767 {
3768 ANV_FROM_HANDLE(anv_device, device, _device);
3769
3770 if (device->info.has_llc)
3771 return VK_SUCCESS;
3772
3773 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
3774
3775 /* Make sure no reads get moved up above the invalidate. */
3776 __builtin_ia32_mfence();
3777
3778 return VK_SUCCESS;
3779 }
3780
3781 void anv_GetBufferMemoryRequirements(
3782 VkDevice _device,
3783 VkBuffer _buffer,
3784 VkMemoryRequirements* pMemoryRequirements)
3785 {
3786 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
3787 ANV_FROM_HANDLE(anv_device, device, _device);
3788 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3789
3790 /* The Vulkan spec (git aaed022) says:
3791 *
3792 * memoryTypeBits is a bitfield and contains one bit set for every
3793 * supported memory type for the resource. The bit `1<<i` is set if and
3794 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3795 * structure for the physical device is supported.
3796 */
3797 uint32_t memory_types = (1ull << pdevice->memory.type_count) - 1;
3798
3799 /* Base alignment requirement of a cache line */
3800 uint32_t alignment = 16;
3801
3802 /* We need an alignment of 32 for pushing UBOs */
3803 if (buffer->usage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT)
3804 alignment = MAX2(alignment, 32);
3805
3806 pMemoryRequirements->size = buffer->size;
3807 pMemoryRequirements->alignment = alignment;
3808
3809 /* Storage and Uniform buffers should have their size aligned to
3810 * 32-bits to avoid boundary checks when last DWord is not complete.
3811 * This would ensure that not internal padding would be needed for
3812 * 16-bit types.
3813 */
3814 if (device->robust_buffer_access &&
3815 (buffer->usage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT ||
3816 buffer->usage & VK_BUFFER_USAGE_STORAGE_BUFFER_BIT))
3817 pMemoryRequirements->size = align_u64(buffer->size, 4);
3818
3819 pMemoryRequirements->memoryTypeBits = memory_types;
3820 }
3821
3822 void anv_GetBufferMemoryRequirements2(
3823 VkDevice _device,
3824 const VkBufferMemoryRequirementsInfo2* pInfo,
3825 VkMemoryRequirements2* pMemoryRequirements)
3826 {
3827 anv_GetBufferMemoryRequirements(_device, pInfo->buffer,
3828 &pMemoryRequirements->memoryRequirements);
3829
3830 vk_foreach_struct(ext, pMemoryRequirements->pNext) {
3831 switch (ext->sType) {
3832 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
3833 VkMemoryDedicatedRequirements *requirements = (void *)ext;
3834 requirements->prefersDedicatedAllocation = false;
3835 requirements->requiresDedicatedAllocation = false;
3836 break;
3837 }
3838
3839 default:
3840 anv_debug_ignored_stype(ext->sType);
3841 break;
3842 }
3843 }
3844 }
3845
3846 void anv_GetImageMemoryRequirements(
3847 VkDevice _device,
3848 VkImage _image,
3849 VkMemoryRequirements* pMemoryRequirements)
3850 {
3851 ANV_FROM_HANDLE(anv_image, image, _image);
3852 ANV_FROM_HANDLE(anv_device, device, _device);
3853 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3854
3855 /* The Vulkan spec (git aaed022) says:
3856 *
3857 * memoryTypeBits is a bitfield and contains one bit set for every
3858 * supported memory type for the resource. The bit `1<<i` is set if and
3859 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3860 * structure for the physical device is supported.
3861 *
3862 * All types are currently supported for images.
3863 */
3864 uint32_t memory_types = (1ull << pdevice->memory.type_count) - 1;
3865
3866 /* We must have image allocated or imported at this point. According to the
3867 * specification, external images must have been bound to memory before
3868 * calling GetImageMemoryRequirements.
3869 */
3870 assert(image->size > 0);
3871
3872 pMemoryRequirements->size = image->size;
3873 pMemoryRequirements->alignment = image->alignment;
3874 pMemoryRequirements->memoryTypeBits = memory_types;
3875 }
3876
3877 void anv_GetImageMemoryRequirements2(
3878 VkDevice _device,
3879 const VkImageMemoryRequirementsInfo2* pInfo,
3880 VkMemoryRequirements2* pMemoryRequirements)
3881 {
3882 ANV_FROM_HANDLE(anv_device, device, _device);
3883 ANV_FROM_HANDLE(anv_image, image, pInfo->image);
3884
3885 anv_GetImageMemoryRequirements(_device, pInfo->image,
3886 &pMemoryRequirements->memoryRequirements);
3887
3888 vk_foreach_struct_const(ext, pInfo->pNext) {
3889 switch (ext->sType) {
3890 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO: {
3891 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3892 const VkImagePlaneMemoryRequirementsInfo *plane_reqs =
3893 (const VkImagePlaneMemoryRequirementsInfo *) ext;
3894 uint32_t plane = anv_image_aspect_to_plane(image->aspects,
3895 plane_reqs->planeAspect);
3896
3897 assert(image->planes[plane].offset == 0);
3898
3899 /* The Vulkan spec (git aaed022) says:
3900 *
3901 * memoryTypeBits is a bitfield and contains one bit set for every
3902 * supported memory type for the resource. The bit `1<<i` is set
3903 * if and only if the memory type `i` in the
3904 * VkPhysicalDeviceMemoryProperties structure for the physical
3905 * device is supported.
3906 *
3907 * All types are currently supported for images.
3908 */
3909 pMemoryRequirements->memoryRequirements.memoryTypeBits =
3910 (1ull << pdevice->memory.type_count) - 1;
3911
3912 /* We must have image allocated or imported at this point. According to the
3913 * specification, external images must have been bound to memory before
3914 * calling GetImageMemoryRequirements.
3915 */
3916 assert(image->planes[plane].size > 0);
3917
3918 pMemoryRequirements->memoryRequirements.size = image->planes[plane].size;
3919 pMemoryRequirements->memoryRequirements.alignment =
3920 image->planes[plane].alignment;
3921 break;
3922 }
3923
3924 default:
3925 anv_debug_ignored_stype(ext->sType);
3926 break;
3927 }
3928 }
3929
3930 vk_foreach_struct(ext, pMemoryRequirements->pNext) {
3931 switch (ext->sType) {
3932 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
3933 VkMemoryDedicatedRequirements *requirements = (void *)ext;
3934 if (image->needs_set_tiling || image->external_format) {
3935 /* If we need to set the tiling for external consumers, we need a
3936 * dedicated allocation.
3937 *
3938 * See also anv_AllocateMemory.
3939 */
3940 requirements->prefersDedicatedAllocation = true;
3941 requirements->requiresDedicatedAllocation = true;
3942 } else {
3943 requirements->prefersDedicatedAllocation = false;
3944 requirements->requiresDedicatedAllocation = false;
3945 }
3946 break;
3947 }
3948
3949 default:
3950 anv_debug_ignored_stype(ext->sType);
3951 break;
3952 }
3953 }
3954 }
3955
3956 void anv_GetImageSparseMemoryRequirements(
3957 VkDevice device,
3958 VkImage image,
3959 uint32_t* pSparseMemoryRequirementCount,
3960 VkSparseImageMemoryRequirements* pSparseMemoryRequirements)
3961 {
3962 *pSparseMemoryRequirementCount = 0;
3963 }
3964
3965 void anv_GetImageSparseMemoryRequirements2(
3966 VkDevice device,
3967 const VkImageSparseMemoryRequirementsInfo2* pInfo,
3968 uint32_t* pSparseMemoryRequirementCount,
3969 VkSparseImageMemoryRequirements2* pSparseMemoryRequirements)
3970 {
3971 *pSparseMemoryRequirementCount = 0;
3972 }
3973
3974 void anv_GetDeviceMemoryCommitment(
3975 VkDevice device,
3976 VkDeviceMemory memory,
3977 VkDeviceSize* pCommittedMemoryInBytes)
3978 {
3979 *pCommittedMemoryInBytes = 0;
3980 }
3981
3982 static void
3983 anv_bind_buffer_memory(const VkBindBufferMemoryInfo *pBindInfo)
3984 {
3985 ANV_FROM_HANDLE(anv_device_memory, mem, pBindInfo->memory);
3986 ANV_FROM_HANDLE(anv_buffer, buffer, pBindInfo->buffer);
3987
3988 assert(pBindInfo->sType == VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO);
3989
3990 if (mem) {
3991 buffer->address = (struct anv_address) {
3992 .bo = mem->bo,
3993 .offset = pBindInfo->memoryOffset,
3994 };
3995 } else {
3996 buffer->address = ANV_NULL_ADDRESS;
3997 }
3998 }
3999
4000 VkResult anv_BindBufferMemory(
4001 VkDevice device,
4002 VkBuffer buffer,
4003 VkDeviceMemory memory,
4004 VkDeviceSize memoryOffset)
4005 {
4006 anv_bind_buffer_memory(
4007 &(VkBindBufferMemoryInfo) {
4008 .sType = VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO,
4009 .buffer = buffer,
4010 .memory = memory,
4011 .memoryOffset = memoryOffset,
4012 });
4013
4014 return VK_SUCCESS;
4015 }
4016
4017 VkResult anv_BindBufferMemory2(
4018 VkDevice device,
4019 uint32_t bindInfoCount,
4020 const VkBindBufferMemoryInfo* pBindInfos)
4021 {
4022 for (uint32_t i = 0; i < bindInfoCount; i++)
4023 anv_bind_buffer_memory(&pBindInfos[i]);
4024
4025 return VK_SUCCESS;
4026 }
4027
4028 VkResult anv_QueueBindSparse(
4029 VkQueue _queue,
4030 uint32_t bindInfoCount,
4031 const VkBindSparseInfo* pBindInfo,
4032 VkFence fence)
4033 {
4034 ANV_FROM_HANDLE(anv_queue, queue, _queue);
4035 if (anv_device_is_lost(queue->device))
4036 return VK_ERROR_DEVICE_LOST;
4037
4038 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
4039 }
4040
4041 // Event functions
4042
4043 VkResult anv_CreateEvent(
4044 VkDevice _device,
4045 const VkEventCreateInfo* pCreateInfo,
4046 const VkAllocationCallbacks* pAllocator,
4047 VkEvent* pEvent)
4048 {
4049 ANV_FROM_HANDLE(anv_device, device, _device);
4050 struct anv_state state;
4051 struct anv_event *event;
4052
4053 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_EVENT_CREATE_INFO);
4054
4055 state = anv_state_pool_alloc(&device->dynamic_state_pool,
4056 sizeof(*event), 8);
4057 event = state.map;
4058 event->state = state;
4059 event->semaphore = VK_EVENT_RESET;
4060
4061 if (!device->info.has_llc) {
4062 /* Make sure the writes we're flushing have landed. */
4063 __builtin_ia32_mfence();
4064 __builtin_ia32_clflush(event);
4065 }
4066
4067 *pEvent = anv_event_to_handle(event);
4068
4069 return VK_SUCCESS;
4070 }
4071
4072 void anv_DestroyEvent(
4073 VkDevice _device,
4074 VkEvent _event,
4075 const VkAllocationCallbacks* pAllocator)
4076 {
4077 ANV_FROM_HANDLE(anv_device, device, _device);
4078 ANV_FROM_HANDLE(anv_event, event, _event);
4079
4080 if (!event)
4081 return;
4082
4083 anv_state_pool_free(&device->dynamic_state_pool, event->state);
4084 }
4085
4086 VkResult anv_GetEventStatus(
4087 VkDevice _device,
4088 VkEvent _event)
4089 {
4090 ANV_FROM_HANDLE(anv_device, device, _device);
4091 ANV_FROM_HANDLE(anv_event, event, _event);
4092
4093 if (anv_device_is_lost(device))
4094 return VK_ERROR_DEVICE_LOST;
4095
4096 if (!device->info.has_llc) {
4097 /* Invalidate read cache before reading event written by GPU. */
4098 __builtin_ia32_clflush(event);
4099 __builtin_ia32_mfence();
4100
4101 }
4102
4103 return event->semaphore;
4104 }
4105
4106 VkResult anv_SetEvent(
4107 VkDevice _device,
4108 VkEvent _event)
4109 {
4110 ANV_FROM_HANDLE(anv_device, device, _device);
4111 ANV_FROM_HANDLE(anv_event, event, _event);
4112
4113 event->semaphore = VK_EVENT_SET;
4114
4115 if (!device->info.has_llc) {
4116 /* Make sure the writes we're flushing have landed. */
4117 __builtin_ia32_mfence();
4118 __builtin_ia32_clflush(event);
4119 }
4120
4121 return VK_SUCCESS;
4122 }
4123
4124 VkResult anv_ResetEvent(
4125 VkDevice _device,
4126 VkEvent _event)
4127 {
4128 ANV_FROM_HANDLE(anv_device, device, _device);
4129 ANV_FROM_HANDLE(anv_event, event, _event);
4130
4131 event->semaphore = VK_EVENT_RESET;
4132
4133 if (!device->info.has_llc) {
4134 /* Make sure the writes we're flushing have landed. */
4135 __builtin_ia32_mfence();
4136 __builtin_ia32_clflush(event);
4137 }
4138
4139 return VK_SUCCESS;
4140 }
4141
4142 // Buffer functions
4143
4144 VkResult anv_CreateBuffer(
4145 VkDevice _device,
4146 const VkBufferCreateInfo* pCreateInfo,
4147 const VkAllocationCallbacks* pAllocator,
4148 VkBuffer* pBuffer)
4149 {
4150 ANV_FROM_HANDLE(anv_device, device, _device);
4151 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
4152 struct anv_buffer *buffer;
4153
4154 /* Don't allow creating buffers bigger than our address space. The real
4155 * issue here is that we may align up the buffer size and we don't want
4156 * doing so to cause roll-over. However, no one has any business
4157 * allocating a buffer larger than our GTT size.
4158 */
4159 if (pCreateInfo->size > pdevice->gtt_size)
4160 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
4161
4162 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
4163
4164 buffer = vk_alloc2(&device->alloc, pAllocator, sizeof(*buffer), 8,
4165 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
4166 if (buffer == NULL)
4167 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
4168
4169 buffer->size = pCreateInfo->size;
4170 buffer->usage = pCreateInfo->usage;
4171 buffer->address = ANV_NULL_ADDRESS;
4172
4173 *pBuffer = anv_buffer_to_handle(buffer);
4174
4175 return VK_SUCCESS;
4176 }
4177
4178 void anv_DestroyBuffer(
4179 VkDevice _device,
4180 VkBuffer _buffer,
4181 const VkAllocationCallbacks* pAllocator)
4182 {
4183 ANV_FROM_HANDLE(anv_device, device, _device);
4184 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
4185
4186 if (!buffer)
4187 return;
4188
4189 vk_free2(&device->alloc, pAllocator, buffer);
4190 }
4191
4192 VkDeviceAddress anv_GetBufferDeviceAddress(
4193 VkDevice device,
4194 const VkBufferDeviceAddressInfoKHR* pInfo)
4195 {
4196 ANV_FROM_HANDLE(anv_buffer, buffer, pInfo->buffer);
4197
4198 assert(!anv_address_is_null(buffer->address));
4199 assert(buffer->address.bo->flags & EXEC_OBJECT_PINNED);
4200
4201 return anv_address_physical(buffer->address);
4202 }
4203
4204 uint64_t anv_GetBufferOpaqueCaptureAddress(
4205 VkDevice device,
4206 const VkBufferDeviceAddressInfoKHR* pInfo)
4207 {
4208 return 0;
4209 }
4210
4211 uint64_t anv_GetDeviceMemoryOpaqueCaptureAddress(
4212 VkDevice device,
4213 const VkDeviceMemoryOpaqueCaptureAddressInfoKHR* pInfo)
4214 {
4215 ANV_FROM_HANDLE(anv_device_memory, memory, pInfo->memory);
4216
4217 assert(memory->bo->flags & EXEC_OBJECT_PINNED);
4218 assert(memory->bo->has_client_visible_address);
4219
4220 return gen_48b_address(memory->bo->offset);
4221 }
4222
4223 void
4224 anv_fill_buffer_surface_state(struct anv_device *device, struct anv_state state,
4225 enum isl_format format,
4226 struct anv_address address,
4227 uint32_t range, uint32_t stride)
4228 {
4229 isl_buffer_fill_state(&device->isl_dev, state.map,
4230 .address = anv_address_physical(address),
4231 .mocs = device->isl_dev.mocs.internal,
4232 .size_B = range,
4233 .format = format,
4234 .swizzle = ISL_SWIZZLE_IDENTITY,
4235 .stride_B = stride);
4236 }
4237
4238 void anv_DestroySampler(
4239 VkDevice _device,
4240 VkSampler _sampler,
4241 const VkAllocationCallbacks* pAllocator)
4242 {
4243 ANV_FROM_HANDLE(anv_device, device, _device);
4244 ANV_FROM_HANDLE(anv_sampler, sampler, _sampler);
4245
4246 if (!sampler)
4247 return;
4248
4249 if (sampler->bindless_state.map) {
4250 anv_state_pool_free(&device->dynamic_state_pool,
4251 sampler->bindless_state);
4252 }
4253
4254 vk_free2(&device->alloc, pAllocator, sampler);
4255 }
4256
4257 VkResult anv_CreateFramebuffer(
4258 VkDevice _device,
4259 const VkFramebufferCreateInfo* pCreateInfo,
4260 const VkAllocationCallbacks* pAllocator,
4261 VkFramebuffer* pFramebuffer)
4262 {
4263 ANV_FROM_HANDLE(anv_device, device, _device);
4264 struct anv_framebuffer *framebuffer;
4265
4266 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
4267
4268 size_t size = sizeof(*framebuffer);
4269
4270 /* VK_KHR_imageless_framebuffer extension says:
4271 *
4272 * If flags includes VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR,
4273 * parameter pAttachments is ignored.
4274 */
4275 if (!(pCreateInfo->flags & VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR)) {
4276 size += sizeof(struct anv_image_view *) * pCreateInfo->attachmentCount;
4277 framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
4278 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
4279 if (framebuffer == NULL)
4280 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
4281
4282 for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
4283 ANV_FROM_HANDLE(anv_image_view, iview, pCreateInfo->pAttachments[i]);
4284 framebuffer->attachments[i] = iview;
4285 }
4286 framebuffer->attachment_count = pCreateInfo->attachmentCount;
4287 } else {
4288 assert(device->enabled_extensions.KHR_imageless_framebuffer);
4289 framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
4290 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
4291 if (framebuffer == NULL)
4292 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
4293
4294 framebuffer->attachment_count = 0;
4295 }
4296
4297 framebuffer->width = pCreateInfo->width;
4298 framebuffer->height = pCreateInfo->height;
4299 framebuffer->layers = pCreateInfo->layers;
4300
4301 *pFramebuffer = anv_framebuffer_to_handle(framebuffer);
4302
4303 return VK_SUCCESS;
4304 }
4305
4306 void anv_DestroyFramebuffer(
4307 VkDevice _device,
4308 VkFramebuffer _fb,
4309 const VkAllocationCallbacks* pAllocator)
4310 {
4311 ANV_FROM_HANDLE(anv_device, device, _device);
4312 ANV_FROM_HANDLE(anv_framebuffer, fb, _fb);
4313
4314 if (!fb)
4315 return;
4316
4317 vk_free2(&device->alloc, pAllocator, fb);
4318 }
4319
4320 static const VkTimeDomainEXT anv_time_domains[] = {
4321 VK_TIME_DOMAIN_DEVICE_EXT,
4322 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT,
4323 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT,
4324 };
4325
4326 VkResult anv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
4327 VkPhysicalDevice physicalDevice,
4328 uint32_t *pTimeDomainCount,
4329 VkTimeDomainEXT *pTimeDomains)
4330 {
4331 int d;
4332 VK_OUTARRAY_MAKE(out, pTimeDomains, pTimeDomainCount);
4333
4334 for (d = 0; d < ARRAY_SIZE(anv_time_domains); d++) {
4335 vk_outarray_append(&out, i) {
4336 *i = anv_time_domains[d];
4337 }
4338 }
4339
4340 return vk_outarray_status(&out);
4341 }
4342
4343 static uint64_t
4344 anv_clock_gettime(clockid_t clock_id)
4345 {
4346 struct timespec current;
4347 int ret;
4348
4349 ret = clock_gettime(clock_id, &current);
4350 if (ret < 0 && clock_id == CLOCK_MONOTONIC_RAW)
4351 ret = clock_gettime(CLOCK_MONOTONIC, &current);
4352 if (ret < 0)
4353 return 0;
4354
4355 return (uint64_t) current.tv_sec * 1000000000ULL + current.tv_nsec;
4356 }
4357
4358 #define TIMESTAMP 0x2358
4359
4360 VkResult anv_GetCalibratedTimestampsEXT(
4361 VkDevice _device,
4362 uint32_t timestampCount,
4363 const VkCalibratedTimestampInfoEXT *pTimestampInfos,
4364 uint64_t *pTimestamps,
4365 uint64_t *pMaxDeviation)
4366 {
4367 ANV_FROM_HANDLE(anv_device, device, _device);
4368 uint64_t timestamp_frequency = device->info.timestamp_frequency;
4369 int ret;
4370 int d;
4371 uint64_t begin, end;
4372 uint64_t max_clock_period = 0;
4373
4374 begin = anv_clock_gettime(CLOCK_MONOTONIC_RAW);
4375
4376 for (d = 0; d < timestampCount; d++) {
4377 switch (pTimestampInfos[d].timeDomain) {
4378 case VK_TIME_DOMAIN_DEVICE_EXT:
4379 ret = anv_gem_reg_read(device, TIMESTAMP | 1,
4380 &pTimestamps[d]);
4381
4382 if (ret != 0) {
4383 return anv_device_set_lost(device, "Failed to read the TIMESTAMP "
4384 "register: %m");
4385 }
4386 uint64_t device_period = DIV_ROUND_UP(1000000000, timestamp_frequency);
4387 max_clock_period = MAX2(max_clock_period, device_period);
4388 break;
4389 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT:
4390 pTimestamps[d] = anv_clock_gettime(CLOCK_MONOTONIC);
4391 max_clock_period = MAX2(max_clock_period, 1);
4392 break;
4393
4394 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT:
4395 pTimestamps[d] = begin;
4396 break;
4397 default:
4398 pTimestamps[d] = 0;
4399 break;
4400 }
4401 }
4402
4403 end = anv_clock_gettime(CLOCK_MONOTONIC_RAW);
4404
4405 /*
4406 * The maximum deviation is the sum of the interval over which we
4407 * perform the sampling and the maximum period of any sampled
4408 * clock. That's because the maximum skew between any two sampled
4409 * clock edges is when the sampled clock with the largest period is
4410 * sampled at the end of that period but right at the beginning of the
4411 * sampling interval and some other clock is sampled right at the
4412 * begining of its sampling period and right at the end of the
4413 * sampling interval. Let's assume the GPU has the longest clock
4414 * period and that the application is sampling GPU and monotonic:
4415 *
4416 * s e
4417 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
4418 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4419 *
4420 * g
4421 * 0 1 2 3
4422 * GPU -----_____-----_____-----_____-----_____
4423 *
4424 * m
4425 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
4426 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4427 *
4428 * Interval <----------------->
4429 * Deviation <-------------------------->
4430 *
4431 * s = read(raw) 2
4432 * g = read(GPU) 1
4433 * m = read(monotonic) 2
4434 * e = read(raw) b
4435 *
4436 * We round the sample interval up by one tick to cover sampling error
4437 * in the interval clock
4438 */
4439
4440 uint64_t sample_interval = end - begin + 1;
4441
4442 *pMaxDeviation = sample_interval + max_clock_period;
4443
4444 return VK_SUCCESS;
4445 }
4446
4447 /* vk_icd.h does not declare this function, so we declare it here to
4448 * suppress Wmissing-prototypes.
4449 */
4450 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4451 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion);
4452
4453 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4454 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion)
4455 {
4456 /* For the full details on loader interface versioning, see
4457 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
4458 * What follows is a condensed summary, to help you navigate the large and
4459 * confusing official doc.
4460 *
4461 * - Loader interface v0 is incompatible with later versions. We don't
4462 * support it.
4463 *
4464 * - In loader interface v1:
4465 * - The first ICD entrypoint called by the loader is
4466 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
4467 * entrypoint.
4468 * - The ICD must statically expose no other Vulkan symbol unless it is
4469 * linked with -Bsymbolic.
4470 * - Each dispatchable Vulkan handle created by the ICD must be
4471 * a pointer to a struct whose first member is VK_LOADER_DATA. The
4472 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
4473 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
4474 * vkDestroySurfaceKHR(). The ICD must be capable of working with
4475 * such loader-managed surfaces.
4476 *
4477 * - Loader interface v2 differs from v1 in:
4478 * - The first ICD entrypoint called by the loader is
4479 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
4480 * statically expose this entrypoint.
4481 *
4482 * - Loader interface v3 differs from v2 in:
4483 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
4484 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
4485 * because the loader no longer does so.
4486 *
4487 * - Loader interface v4 differs from v3 in:
4488 * - The ICD must implement vk_icdGetPhysicalDeviceProcAddr().
4489 */
4490 *pSupportedVersion = MIN2(*pSupportedVersion, 4u);
4491 return VK_SUCCESS;
4492 }