projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
ANV: Stop advertising smoothLines support on gen10+
[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_SECTION_END
61 DRI_CONF_END;
62
63 /* This is probably far to big but it reflects the max size used for messages
64 * in OpenGLs KHR_debug.
65 */
66 #define MAX_DEBUG_MESSAGE_LENGTH 4096
67
68 static void
69 compiler_debug_log(void *data, const char *fmt, ...)
70 {
71 char str[MAX_DEBUG_MESSAGE_LENGTH];
72 struct anv_device *device = (struct anv_device *)data;
73
74 if (list_is_empty(&device->instance->debug_report_callbacks.callbacks))
75 return;
76
77 va_list args;
78 va_start(args, fmt);
79 (void) vsnprintf(str, MAX_DEBUG_MESSAGE_LENGTH, fmt, args);
80 va_end(args);
81
82 vk_debug_report(&device->instance->debug_report_callbacks,
83 VK_DEBUG_REPORT_DEBUG_BIT_EXT,
84 VK_DEBUG_REPORT_OBJECT_TYPE_UNKNOWN_EXT,
85 0, 0, 0, "anv", str);
86 }
87
88 static void
89 compiler_perf_log(void *data, const char *fmt, ...)
90 {
91 va_list args;
92 va_start(args, fmt);
93
94 if (unlikely(INTEL_DEBUG & DEBUG_PERF))
95 intel_logd_v(fmt, args);
96
97 va_end(args);
98 }
99
100 static uint64_t
101 anv_compute_heap_size(int fd, uint64_t gtt_size)
102 {
103 /* Query the total ram from the system */
104 struct sysinfo info;
105 sysinfo(&info);
106
107 uint64_t total_ram = (uint64_t)info.totalram * (uint64_t)info.mem_unit;
108
109 /* We don't want to burn too much ram with the GPU. If the user has 4GiB
110 * or less, we use at most half. If they have more than 4GiB, we use 3/4.
111 */
112 uint64_t available_ram;
113 if (total_ram <= 4ull * 1024ull * 1024ull * 1024ull)
114 available_ram = total_ram / 2;
115 else
116 available_ram = total_ram * 3 / 4;
117
118 /* We also want to leave some padding for things we allocate in the driver,
119 * so don't go over 3/4 of the GTT either.
120 */
121 uint64_t available_gtt = gtt_size * 3 / 4;
122
123 return MIN2(available_ram, available_gtt);
124 }
125
126 static VkResult
127 anv_physical_device_init_heaps(struct anv_physical_device *device, int fd)
128 {
129 if (anv_gem_get_context_param(fd, 0, I915_CONTEXT_PARAM_GTT_SIZE,
130 &device->gtt_size) == -1) {
131 /* If, for whatever reason, we can't actually get the GTT size from the
132 * kernel (too old?) fall back to the aperture size.
133 */
134 anv_perf_warn(NULL, NULL,
135 "Failed to get I915_CONTEXT_PARAM_GTT_SIZE: %m");
136
137 if (anv_gem_get_aperture(fd, &device->gtt_size) == -1) {
138 return vk_errorf(NULL, NULL, VK_ERROR_INITIALIZATION_FAILED,
139 "failed to get aperture size: %m");
140 }
141 }
142
143 /* We only allow 48-bit addresses with softpin because knowing the actual
144 * address is required for the vertex cache flush workaround.
145 */
146 device->supports_48bit_addresses = (device->info.gen >= 8) &&
147 device->has_softpin &&
148 device->gtt_size > (4ULL << 30 /* GiB */);
149
150 uint64_t heap_size = anv_compute_heap_size(fd, device->gtt_size);
151
152 if (heap_size > (2ull << 30) && !device->supports_48bit_addresses) {
153 /* When running with an overridden PCI ID, we may get a GTT size from
154 * the kernel that is greater than 2 GiB but the execbuf check for 48bit
155 * address support can still fail. Just clamp the address space size to
156 * 2 GiB if we don't have 48-bit support.
157 */
158 intel_logw("%s:%d: The kernel reported a GTT size larger than 2 GiB but "
159 "not support for 48-bit addresses",
160 __FILE__, __LINE__);
161 heap_size = 2ull << 30;
162 }
163
164 device->memory.heap_count = 1;
165 device->memory.heaps[0] = (struct anv_memory_heap) {
166 .size = heap_size,
167 .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
168 };
169
170 uint32_t type_count = 0;
171 for (uint32_t heap = 0; heap < device->memory.heap_count; heap++) {
172 if (device->info.has_llc) {
173 /* Big core GPUs share LLC with the CPU and thus one memory type can be
174 * both cached and coherent at the same time.
175 */
176 device->memory.types[type_count++] = (struct anv_memory_type) {
177 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
178 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
179 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
180 VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
181 .heapIndex = heap,
182 };
183 } else {
184 /* The spec requires that we expose a host-visible, coherent memory
185 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
186 * to give the application a choice between cached, but not coherent and
187 * coherent but uncached (WC though).
188 */
189 device->memory.types[type_count++] = (struct anv_memory_type) {
190 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
191 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
192 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
193 .heapIndex = heap,
194 };
195 device->memory.types[type_count++] = (struct anv_memory_type) {
196 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
197 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
198 VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
199 .heapIndex = heap,
200 };
201 }
202 }
203 device->memory.type_count = type_count;
204
205 return VK_SUCCESS;
206 }
207
208 static VkResult
209 anv_physical_device_init_uuids(struct anv_physical_device *device)
210 {
211 const struct build_id_note *note =
212 build_id_find_nhdr_for_addr(anv_physical_device_init_uuids);
213 if (!note) {
214 return vk_errorf(device->instance, device,
215 VK_ERROR_INITIALIZATION_FAILED,
216 "Failed to find build-id");
217 }
218
219 unsigned build_id_len = build_id_length(note);
220 if (build_id_len < 20) {
221 return vk_errorf(device->instance, device,
222 VK_ERROR_INITIALIZATION_FAILED,
223 "build-id too short. It needs to be a SHA");
224 }
225
226 memcpy(device->driver_build_sha1, build_id_data(note), 20);
227
228 struct mesa_sha1 sha1_ctx;
229 uint8_t sha1[20];
230 STATIC_ASSERT(VK_UUID_SIZE <= sizeof(sha1));
231
232 /* The pipeline cache UUID is used for determining when a pipeline cache is
233 * invalid. It needs both a driver build and the PCI ID of the device.
234 */
235 _mesa_sha1_init(&sha1_ctx);
236 _mesa_sha1_update(&sha1_ctx, build_id_data(note), build_id_len);
237 _mesa_sha1_update(&sha1_ctx, &device->chipset_id,
238 sizeof(device->chipset_id));
239 _mesa_sha1_update(&sha1_ctx, &device->always_use_bindless,
240 sizeof(device->always_use_bindless));
241 _mesa_sha1_update(&sha1_ctx, &device->has_a64_buffer_access,
242 sizeof(device->has_a64_buffer_access));
243 _mesa_sha1_update(&sha1_ctx, &device->has_bindless_images,
244 sizeof(device->has_bindless_images));
245 _mesa_sha1_update(&sha1_ctx, &device->has_bindless_samplers,
246 sizeof(device->has_bindless_samplers));
247 _mesa_sha1_final(&sha1_ctx, sha1);
248 memcpy(device->pipeline_cache_uuid, sha1, VK_UUID_SIZE);
249
250 /* The driver UUID is used for determining sharability of images and memory
251 * between two Vulkan instances in separate processes. People who want to
252 * share memory need to also check the device UUID (below) so all this
253 * needs to be is the build-id.
254 */
255 memcpy(device->driver_uuid, build_id_data(note), VK_UUID_SIZE);
256
257 /* The device UUID uniquely identifies the given device within the machine.
258 * Since we never have more than one device, this doesn't need to be a real
259 * UUID. However, on the off-chance that someone tries to use this to
260 * cache pre-tiled images or something of the like, we use the PCI ID and
261 * some bits of ISL info to ensure that this is safe.
262 */
263 _mesa_sha1_init(&sha1_ctx);
264 _mesa_sha1_update(&sha1_ctx, &device->chipset_id,
265 sizeof(device->chipset_id));
266 _mesa_sha1_update(&sha1_ctx, &device->isl_dev.has_bit6_swizzling,
267 sizeof(device->isl_dev.has_bit6_swizzling));
268 _mesa_sha1_final(&sha1_ctx, sha1);
269 memcpy(device->device_uuid, sha1, VK_UUID_SIZE);
270
271 return VK_SUCCESS;
272 }
273
274 static void
275 anv_physical_device_init_disk_cache(struct anv_physical_device *device)
276 {
277 #ifdef ENABLE_SHADER_CACHE
278 char renderer[10];
279 ASSERTED int len = snprintf(renderer, sizeof(renderer), "anv_%04x",
280 device->chipset_id);
281 assert(len == sizeof(renderer) - 2);
282
283 char timestamp[41];
284 _mesa_sha1_format(timestamp, device->driver_build_sha1);
285
286 const uint64_t driver_flags =
287 brw_get_compiler_config_value(device->compiler);
288 device->disk_cache = disk_cache_create(renderer, timestamp, driver_flags);
289 #else
290 device->disk_cache = NULL;
291 #endif
292 }
293
294 static void
295 anv_physical_device_free_disk_cache(struct anv_physical_device *device)
296 {
297 #ifdef ENABLE_SHADER_CACHE
298 if (device->disk_cache)
299 disk_cache_destroy(device->disk_cache);
300 #else
301 assert(device->disk_cache == NULL);
302 #endif
303 }
304
305 static uint64_t
306 get_available_system_memory()
307 {
308 char *meminfo = os_read_file("/proc/meminfo");
309 if (!meminfo)
310 return 0;
311
312 char *str = strstr(meminfo, "MemAvailable:");
313 if (!str) {
314 free(meminfo);
315 return 0;
316 }
317
318 uint64_t kb_mem_available;
319 if (sscanf(str, "MemAvailable: %" PRIx64, &kb_mem_available) == 1) {
320 free(meminfo);
321 return kb_mem_available << 10;
322 }
323
324 free(meminfo);
325 return 0;
326 }
327
328 static VkResult
329 anv_physical_device_init(struct anv_physical_device *device,
330 struct anv_instance *instance,
331 drmDevicePtr drm_device)
332 {
333 const char *primary_path = drm_device->nodes[DRM_NODE_PRIMARY];
334 const char *path = drm_device->nodes[DRM_NODE_RENDER];
335 VkResult result;
336 int fd;
337 int master_fd = -1;
338
339 brw_process_intel_debug_variable();
340
341 fd = open(path, O_RDWR | O_CLOEXEC);
342 if (fd < 0)
343 return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
344
345 device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
346 device->instance = instance;
347
348 assert(strlen(path) < ARRAY_SIZE(device->path));
349 snprintf(device->path, ARRAY_SIZE(device->path), "%s", path);
350
351 if (!gen_get_device_info_from_fd(fd, &device->info)) {
352 result = vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
353 goto fail;
354 }
355 device->chipset_id = device->info.chipset_id;
356 device->no_hw = device->info.no_hw;
357
358 if (getenv("INTEL_NO_HW") != NULL)
359 device->no_hw = true;
360
361 device->pci_info.domain = drm_device->businfo.pci->domain;
362 device->pci_info.bus = drm_device->businfo.pci->bus;
363 device->pci_info.device = drm_device->businfo.pci->dev;
364 device->pci_info.function = drm_device->businfo.pci->func;
365
366 device->name = gen_get_device_name(device->chipset_id);
367
368 if (device->info.is_haswell) {
369 intel_logw("Haswell Vulkan support is incomplete");
370 } else if (device->info.gen == 7 && !device->info.is_baytrail) {
371 intel_logw("Ivy Bridge Vulkan support is incomplete");
372 } else if (device->info.gen == 7 && device->info.is_baytrail) {
373 intel_logw("Bay Trail Vulkan support is incomplete");
374 } else if (device->info.gen >= 8 && device->info.gen <= 11) {
375 /* Gen8-11 fully supported */
376 } else if (device->info.gen == 12) {
377 intel_logw("Vulkan is not yet fully supported on gen12");
378 } else {
379 result = vk_errorf(device->instance, device,
380 VK_ERROR_INCOMPATIBLE_DRIVER,
381 "Vulkan not yet supported on %s", device->name);
382 goto fail;
383 }
384
385 device->cmd_parser_version = -1;
386 if (device->info.gen == 7) {
387 device->cmd_parser_version =
388 anv_gem_get_param(fd, I915_PARAM_CMD_PARSER_VERSION);
389 if (device->cmd_parser_version == -1) {
390 result = vk_errorf(device->instance, device,
391 VK_ERROR_INITIALIZATION_FAILED,
392 "failed to get command parser version");
393 goto fail;
394 }
395 }
396
397 if (!anv_gem_get_param(fd, I915_PARAM_HAS_WAIT_TIMEOUT)) {
398 result = vk_errorf(device->instance, device,
399 VK_ERROR_INITIALIZATION_FAILED,
400 "kernel missing gem wait");
401 goto fail;
402 }
403
404 if (!anv_gem_get_param(fd, I915_PARAM_HAS_EXECBUF2)) {
405 result = vk_errorf(device->instance, device,
406 VK_ERROR_INITIALIZATION_FAILED,
407 "kernel missing execbuf2");
408 goto fail;
409 }
410
411 if (!device->info.has_llc &&
412 anv_gem_get_param(fd, I915_PARAM_MMAP_VERSION) < 1) {
413 result = vk_errorf(device->instance, device,
414 VK_ERROR_INITIALIZATION_FAILED,
415 "kernel missing wc mmap");
416 goto fail;
417 }
418
419 device->has_softpin = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_SOFTPIN);
420 device->has_exec_async = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_ASYNC);
421 device->has_exec_capture = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_CAPTURE);
422 device->has_exec_fence = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_FENCE);
423 device->has_syncobj = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_FENCE_ARRAY);
424 device->has_syncobj_wait = device->has_syncobj &&
425 anv_gem_supports_syncobj_wait(fd);
426 device->has_context_priority = anv_gem_has_context_priority(fd);
427
428 result = anv_physical_device_init_heaps(device, fd);
429 if (result != VK_SUCCESS)
430 goto fail;
431
432 device->use_softpin = device->has_softpin &&
433 device->supports_48bit_addresses;
434
435 device->has_context_isolation =
436 anv_gem_get_param(fd, I915_PARAM_HAS_CONTEXT_ISOLATION);
437
438 device->always_use_bindless =
439 env_var_as_boolean("ANV_ALWAYS_BINDLESS", false);
440
441 /* We first got the A64 messages on broadwell and we can only use them if
442 * we can pass addresses directly into the shader which requires softpin.
443 */
444 device->has_a64_buffer_access = device->info.gen >= 8 &&
445 device->use_softpin;
446
447 /* We first get bindless image access on Skylake and we can only really do
448 * it if we don't have any relocations so we need softpin.
449 */
450 device->has_bindless_images = device->info.gen >= 9 &&
451 device->use_softpin;
452
453 /* We've had bindless samplers since Ivy Bridge (forever in Vulkan terms)
454 * because it's just a matter of setting the sampler address in the sample
455 * message header. However, we've not bothered to wire it up for vec4 so
456 * we leave it disabled on gen7.
457 */
458 device->has_bindless_samplers = device->info.gen >= 8;
459
460 device->has_mem_available = get_available_system_memory() != 0;
461
462 device->always_flush_cache =
463 driQueryOptionb(&instance->dri_options, "always_flush_cache");
464
465 /* Starting with Gen10, the timestamp frequency of the command streamer may
466 * vary from one part to another. We can query the value from the kernel.
467 */
468 if (device->info.gen >= 10) {
469 int timestamp_frequency =
470 anv_gem_get_param(fd, I915_PARAM_CS_TIMESTAMP_FREQUENCY);
471
472 if (timestamp_frequency < 0)
473 intel_logw("Kernel 4.16-rc1+ required to properly query CS timestamp frequency");
474 else
475 device->info.timestamp_frequency = timestamp_frequency;
476 }
477
478 /* GENs prior to 8 do not support EU/Subslice info */
479 if (device->info.gen >= 8) {
480 device->subslice_total = anv_gem_get_param(fd, I915_PARAM_SUBSLICE_TOTAL);
481 device->eu_total = anv_gem_get_param(fd, I915_PARAM_EU_TOTAL);
482
483 /* Without this information, we cannot get the right Braswell
484 * brandstrings, and we have to use conservative numbers for GPGPU on
485 * many platforms, but otherwise, things will just work.
486 */
487 if (device->subslice_total < 1 || device->eu_total < 1) {
488 intel_logw("Kernel 4.1 required to properly query GPU properties");
489 }
490 } else if (device->info.gen == 7) {
491 device->subslice_total = 1 << (device->info.gt - 1);
492 }
493
494 if (device->info.is_cherryview &&
495 device->subslice_total > 0 && device->eu_total > 0) {
496 /* Logical CS threads = EUs per subslice * num threads per EU */
497 uint32_t max_cs_threads =
498 device->eu_total / device->subslice_total * device->info.num_thread_per_eu;
499
500 /* Fuse configurations may give more threads than expected, never less. */
501 if (max_cs_threads > device->info.max_cs_threads)
502 device->info.max_cs_threads = max_cs_threads;
503 }
504
505 device->compiler = brw_compiler_create(NULL, &device->info);
506 if (device->compiler == NULL) {
507 result = vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
508 goto fail;
509 }
510 device->compiler->shader_debug_log = compiler_debug_log;
511 device->compiler->shader_perf_log = compiler_perf_log;
512 device->compiler->supports_pull_constants = false;
513 device->compiler->constant_buffer_0_is_relative =
514 device->info.gen < 8 || !device->has_context_isolation;
515 device->compiler->supports_shader_constants = true;
516 device->compiler->compact_params = false;
517
518 /* Broadwell PRM says:
519 *
520 * "Before Gen8, there was a historical configuration control field to
521 * swizzle address bit[6] for in X/Y tiling modes. This was set in three
522 * different places: TILECTL[1:0], ARB_MODE[5:4], and
523 * DISP_ARB_CTL[14:13].
524 *
525 * For Gen8 and subsequent generations, the swizzle fields are all
526 * reserved, and the CPU's memory controller performs all address
527 * swizzling modifications."
528 */
529 bool swizzled =
530 device->info.gen < 8 && anv_gem_get_bit6_swizzle(fd, I915_TILING_X);
531
532 isl_device_init(&device->isl_dev, &device->info, swizzled);
533
534 result = anv_physical_device_init_uuids(device);
535 if (result != VK_SUCCESS)
536 goto fail;
537
538 anv_physical_device_init_disk_cache(device);
539
540 if (instance->enabled_extensions.KHR_display) {
541 master_fd = open(primary_path, O_RDWR | O_CLOEXEC);
542 if (master_fd >= 0) {
543 /* prod the device with a GETPARAM call which will fail if
544 * we don't have permission to even render on this device
545 */
546 if (anv_gem_get_param(master_fd, I915_PARAM_CHIPSET_ID) == 0) {
547 close(master_fd);
548 master_fd = -1;
549 }
550 }
551 }
552 device->master_fd = master_fd;
553
554 result = anv_init_wsi(device);
555 if (result != VK_SUCCESS) {
556 ralloc_free(device->compiler);
557 anv_physical_device_free_disk_cache(device);
558 goto fail;
559 }
560
561 device->perf = anv_get_perf(&device->info, fd);
562
563 anv_physical_device_get_supported_extensions(device,
564 &device->supported_extensions);
565
566
567 device->local_fd = fd;
568
569 return VK_SUCCESS;
570
571 fail:
572 close(fd);
573 if (master_fd != -1)
574 close(master_fd);
575 return result;
576 }
577
578 static void
579 anv_physical_device_finish(struct anv_physical_device *device)
580 {
581 anv_finish_wsi(device);
582 anv_physical_device_free_disk_cache(device);
583 ralloc_free(device->compiler);
584 ralloc_free(device->perf);
585 close(device->local_fd);
586 if (device->master_fd >= 0)
587 close(device->master_fd);
588 }
589
590 static void *
591 default_alloc_func(void *pUserData, size_t size, size_t align,
592 VkSystemAllocationScope allocationScope)
593 {
594 return malloc(size);
595 }
596
597 static void *
598 default_realloc_func(void *pUserData, void *pOriginal, size_t size,
599 size_t align, VkSystemAllocationScope allocationScope)
600 {
601 return realloc(pOriginal, size);
602 }
603
604 static void
605 default_free_func(void *pUserData, void *pMemory)
606 {
607 free(pMemory);
608 }
609
610 static const VkAllocationCallbacks default_alloc = {
611 .pUserData = NULL,
612 .pfnAllocation = default_alloc_func,
613 .pfnReallocation = default_realloc_func,
614 .pfnFree = default_free_func,
615 };
616
617 VkResult anv_EnumerateInstanceExtensionProperties(
618 const char* pLayerName,
619 uint32_t* pPropertyCount,
620 VkExtensionProperties* pProperties)
621 {
622 VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
623
624 for (int i = 0; i < ANV_INSTANCE_EXTENSION_COUNT; i++) {
625 if (anv_instance_extensions_supported.extensions[i]) {
626 vk_outarray_append(&out, prop) {
627 *prop = anv_instance_extensions[i];
628 }
629 }
630 }
631
632 return vk_outarray_status(&out);
633 }
634
635 VkResult anv_CreateInstance(
636 const VkInstanceCreateInfo* pCreateInfo,
637 const VkAllocationCallbacks* pAllocator,
638 VkInstance* pInstance)
639 {
640 struct anv_instance *instance;
641 VkResult result;
642
643 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO);
644
645 struct anv_instance_extension_table enabled_extensions = {};
646 for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
647 int idx;
648 for (idx = 0; idx < ANV_INSTANCE_EXTENSION_COUNT; idx++) {
649 if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
650 anv_instance_extensions[idx].extensionName) == 0)
651 break;
652 }
653
654 if (idx >= ANV_INSTANCE_EXTENSION_COUNT)
655 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
656
657 if (!anv_instance_extensions_supported.extensions[idx])
658 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
659
660 enabled_extensions.extensions[idx] = true;
661 }
662
663 instance = vk_alloc2(&default_alloc, pAllocator, sizeof(*instance), 8,
664 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
665 if (!instance)
666 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
667
668 instance->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
669
670 if (pAllocator)
671 instance->alloc = *pAllocator;
672 else
673 instance->alloc = default_alloc;
674
675 instance->app_info = (struct anv_app_info) { .api_version = 0 };
676 if (pCreateInfo->pApplicationInfo) {
677 const VkApplicationInfo *app = pCreateInfo->pApplicationInfo;
678
679 instance->app_info.app_name =
680 vk_strdup(&instance->alloc, app->pApplicationName,
681 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
682 instance->app_info.app_version = app->applicationVersion;
683
684 instance->app_info.engine_name =
685 vk_strdup(&instance->alloc, app->pEngineName,
686 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
687 instance->app_info.engine_version = app->engineVersion;
688
689 instance->app_info.api_version = app->apiVersion;
690 }
691
692 if (instance->app_info.api_version == 0)
693 instance->app_info.api_version = VK_API_VERSION_1_0;
694
695 instance->enabled_extensions = enabled_extensions;
696
697 for (unsigned i = 0; i < ARRAY_SIZE(instance->dispatch.entrypoints); i++) {
698 /* Vulkan requires that entrypoints for extensions which have not been
699 * enabled must not be advertised.
700 */
701 if (!anv_instance_entrypoint_is_enabled(i, instance->app_info.api_version,
702 &instance->enabled_extensions)) {
703 instance->dispatch.entrypoints[i] = NULL;
704 } else {
705 instance->dispatch.entrypoints[i] =
706 anv_instance_dispatch_table.entrypoints[i];
707 }
708 }
709
710 struct anv_physical_device *pdevice = &instance->physicalDevice;
711 for (unsigned i = 0; i < ARRAY_SIZE(pdevice->dispatch.entrypoints); i++) {
712 /* Vulkan requires that entrypoints for extensions which have not been
713 * enabled must not be advertised.
714 */
715 if (!anv_physical_device_entrypoint_is_enabled(i, instance->app_info.api_version,
716 &instance->enabled_extensions)) {
717 pdevice->dispatch.entrypoints[i] = NULL;
718 } else {
719 pdevice->dispatch.entrypoints[i] =
720 anv_physical_device_dispatch_table.entrypoints[i];
721 }
722 }
723
724 for (unsigned i = 0; i < ARRAY_SIZE(instance->device_dispatch.entrypoints); i++) {
725 /* Vulkan requires that entrypoints for extensions which have not been
726 * enabled must not be advertised.
727 */
728 if (!anv_device_entrypoint_is_enabled(i, instance->app_info.api_version,
729 &instance->enabled_extensions, NULL)) {
730 instance->device_dispatch.entrypoints[i] = NULL;
731 } else {
732 instance->device_dispatch.entrypoints[i] =
733 anv_device_dispatch_table.entrypoints[i];
734 }
735 }
736
737 instance->physicalDeviceCount = -1;
738
739 result = vk_debug_report_instance_init(&instance->debug_report_callbacks);
740 if (result != VK_SUCCESS) {
741 vk_free2(&default_alloc, pAllocator, instance);
742 return vk_error(result);
743 }
744
745 instance->pipeline_cache_enabled =
746 env_var_as_boolean("ANV_ENABLE_PIPELINE_CACHE", true);
747
748 glsl_type_singleton_init_or_ref();
749
750 VG(VALGRIND_CREATE_MEMPOOL(instance, 0, false));
751
752 driParseOptionInfo(&instance->available_dri_options, anv_dri_options_xml);
753 driParseConfigFiles(&instance->dri_options, &instance->available_dri_options,
754 0, "anv", NULL,
755 instance->app_info.engine_name,
756 instance->app_info.engine_version);
757
758 *pInstance = anv_instance_to_handle(instance);
759
760 return VK_SUCCESS;
761 }
762
763 void anv_DestroyInstance(
764 VkInstance _instance,
765 const VkAllocationCallbacks* pAllocator)
766 {
767 ANV_FROM_HANDLE(anv_instance, instance, _instance);
768
769 if (!instance)
770 return;
771
772 if (instance->physicalDeviceCount > 0) {
773 /* We support at most one physical device. */
774 assert(instance->physicalDeviceCount == 1);
775 anv_physical_device_finish(&instance->physicalDevice);
776 }
777
778 vk_free(&instance->alloc, (char *)instance->app_info.app_name);
779 vk_free(&instance->alloc, (char *)instance->app_info.engine_name);
780
781 VG(VALGRIND_DESTROY_MEMPOOL(instance));
782
783 vk_debug_report_instance_destroy(&instance->debug_report_callbacks);
784
785 glsl_type_singleton_decref();
786
787 driDestroyOptionCache(&instance->dri_options);
788 driDestroyOptionInfo(&instance->available_dri_options);
789
790 vk_free(&instance->alloc, instance);
791 }
792
793 static VkResult
794 anv_enumerate_devices(struct anv_instance *instance)
795 {
796 /* TODO: Check for more devices ? */
797 drmDevicePtr devices[8];
798 VkResult result = VK_ERROR_INCOMPATIBLE_DRIVER;
799 int max_devices;
800
801 instance->physicalDeviceCount = 0;
802
803 max_devices = drmGetDevices2(0, devices, ARRAY_SIZE(devices));
804 if (max_devices < 1)
805 return VK_ERROR_INCOMPATIBLE_DRIVER;
806
807 for (unsigned i = 0; i < (unsigned)max_devices; i++) {
808 if (devices[i]->available_nodes & 1 << DRM_NODE_RENDER &&
809 devices[i]->bustype == DRM_BUS_PCI &&
810 devices[i]->deviceinfo.pci->vendor_id == 0x8086) {
811
812 result = anv_physical_device_init(&instance->physicalDevice,
813 instance, devices[i]);
814 if (result != VK_ERROR_INCOMPATIBLE_DRIVER)
815 break;
816 }
817 }
818 drmFreeDevices(devices, max_devices);
819
820 if (result == VK_SUCCESS)
821 instance->physicalDeviceCount = 1;
822
823 return result;
824 }
825
826 static VkResult
827 anv_instance_ensure_physical_device(struct anv_instance *instance)
828 {
829 if (instance->physicalDeviceCount < 0) {
830 VkResult result = anv_enumerate_devices(instance);
831 if (result != VK_SUCCESS &&
832 result != VK_ERROR_INCOMPATIBLE_DRIVER)
833 return result;
834 }
835
836 return VK_SUCCESS;
837 }
838
839 VkResult anv_EnumeratePhysicalDevices(
840 VkInstance _instance,
841 uint32_t* pPhysicalDeviceCount,
842 VkPhysicalDevice* pPhysicalDevices)
843 {
844 ANV_FROM_HANDLE(anv_instance, instance, _instance);
845 VK_OUTARRAY_MAKE(out, pPhysicalDevices, pPhysicalDeviceCount);
846
847 VkResult result = anv_instance_ensure_physical_device(instance);
848 if (result != VK_SUCCESS)
849 return result;
850
851 if (instance->physicalDeviceCount == 0)
852 return VK_SUCCESS;
853
854 assert(instance->physicalDeviceCount == 1);
855 vk_outarray_append(&out, i) {
856 *i = anv_physical_device_to_handle(&instance->physicalDevice);
857 }
858
859 return vk_outarray_status(&out);
860 }
861
862 VkResult anv_EnumeratePhysicalDeviceGroups(
863 VkInstance _instance,
864 uint32_t* pPhysicalDeviceGroupCount,
865 VkPhysicalDeviceGroupProperties* pPhysicalDeviceGroupProperties)
866 {
867 ANV_FROM_HANDLE(anv_instance, instance, _instance);
868 VK_OUTARRAY_MAKE(out, pPhysicalDeviceGroupProperties,
869 pPhysicalDeviceGroupCount);
870
871 VkResult result = anv_instance_ensure_physical_device(instance);
872 if (result != VK_SUCCESS)
873 return result;
874
875 if (instance->physicalDeviceCount == 0)
876 return VK_SUCCESS;
877
878 assert(instance->physicalDeviceCount == 1);
879
880 vk_outarray_append(&out, p) {
881 p->physicalDeviceCount = 1;
882 memset(p->physicalDevices, 0, sizeof(p->physicalDevices));
883 p->physicalDevices[0] =
884 anv_physical_device_to_handle(&instance->physicalDevice);
885 p->subsetAllocation = false;
886
887 vk_foreach_struct(ext, p->pNext)
888 anv_debug_ignored_stype(ext->sType);
889 }
890
891 return vk_outarray_status(&out);
892 }
893
894 void anv_GetPhysicalDeviceFeatures(
895 VkPhysicalDevice physicalDevice,
896 VkPhysicalDeviceFeatures* pFeatures)
897 {
898 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
899
900 *pFeatures = (VkPhysicalDeviceFeatures) {
901 .robustBufferAccess = true,
902 .fullDrawIndexUint32 = true,
903 .imageCubeArray = true,
904 .independentBlend = true,
905 .geometryShader = true,
906 .tessellationShader = true,
907 .sampleRateShading = true,
908 .dualSrcBlend = true,
909 .logicOp = true,
910 .multiDrawIndirect = true,
911 .drawIndirectFirstInstance = true,
912 .depthClamp = true,
913 .depthBiasClamp = true,
914 .fillModeNonSolid = true,
915 .depthBounds = pdevice->info.gen >= 12,
916 .wideLines = true,
917 .largePoints = true,
918 .alphaToOne = true,
919 .multiViewport = true,
920 .samplerAnisotropy = true,
921 .textureCompressionETC2 = pdevice->info.gen >= 8 ||
922 pdevice->info.is_baytrail,
923 .textureCompressionASTC_LDR = pdevice->info.gen >= 9, /* FINISHME CHV */
924 .textureCompressionBC = true,
925 .occlusionQueryPrecise = true,
926 .pipelineStatisticsQuery = true,
927 .fragmentStoresAndAtomics = true,
928 .shaderTessellationAndGeometryPointSize = true,
929 .shaderImageGatherExtended = true,
930 .shaderStorageImageExtendedFormats = true,
931 .shaderStorageImageMultisample = false,
932 .shaderStorageImageReadWithoutFormat = false,
933 .shaderStorageImageWriteWithoutFormat = true,
934 .shaderUniformBufferArrayDynamicIndexing = true,
935 .shaderSampledImageArrayDynamicIndexing = true,
936 .shaderStorageBufferArrayDynamicIndexing = true,
937 .shaderStorageImageArrayDynamicIndexing = true,
938 .shaderClipDistance = true,
939 .shaderCullDistance = true,
940 .shaderFloat64 = pdevice->info.gen >= 8 &&
941 pdevice->info.has_64bit_types,
942 .shaderInt64 = pdevice->info.gen >= 8 &&
943 pdevice->info.has_64bit_types,
944 .shaderInt16 = pdevice->info.gen >= 8,
945 .shaderResourceMinLod = pdevice->info.gen >= 9,
946 .variableMultisampleRate = true,
947 .inheritedQueries = true,
948 };
949
950 /* We can't do image stores in vec4 shaders */
951 pFeatures->vertexPipelineStoresAndAtomics =
952 pdevice->compiler->scalar_stage[MESA_SHADER_VERTEX] &&
953 pdevice->compiler->scalar_stage[MESA_SHADER_GEOMETRY];
954
955 struct anv_app_info *app_info = &pdevice->instance->app_info;
956
957 /* The new DOOM and Wolfenstein games require depthBounds without
958 * checking for it. They seem to run fine without it so just claim it's
959 * there and accept the consequences.
960 */
961 if (app_info->engine_name && strcmp(app_info->engine_name, "idTech") == 0)
962 pFeatures->depthBounds = true;
963 }
964
965 void anv_GetPhysicalDeviceFeatures2(
966 VkPhysicalDevice physicalDevice,
967 VkPhysicalDeviceFeatures2* pFeatures)
968 {
969 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
970 anv_GetPhysicalDeviceFeatures(physicalDevice, &pFeatures->features);
971
972 vk_foreach_struct(ext, pFeatures->pNext) {
973 switch (ext->sType) {
974 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES_KHR: {
975 VkPhysicalDevice8BitStorageFeaturesKHR *features =
976 (VkPhysicalDevice8BitStorageFeaturesKHR *)ext;
977 features->storageBuffer8BitAccess = pdevice->info.gen >= 8;
978 features->uniformAndStorageBuffer8BitAccess = pdevice->info.gen >= 8;
979 features->storagePushConstant8 = pdevice->info.gen >= 8;
980 break;
981 }
982
983 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES: {
984 VkPhysicalDevice16BitStorageFeatures *features =
985 (VkPhysicalDevice16BitStorageFeatures *)ext;
986 features->storageBuffer16BitAccess = pdevice->info.gen >= 8;
987 features->uniformAndStorageBuffer16BitAccess = pdevice->info.gen >= 8;
988 features->storagePushConstant16 = pdevice->info.gen >= 8;
989 features->storageInputOutput16 = false;
990 break;
991 }
992
993 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_EXT: {
994 VkPhysicalDeviceBufferDeviceAddressFeaturesEXT *features = (void *)ext;
995 features->bufferDeviceAddress = pdevice->has_a64_buffer_access;
996 features->bufferDeviceAddressCaptureReplay = false;
997 features->bufferDeviceAddressMultiDevice = false;
998 break;
999 }
1000
1001 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_KHR: {
1002 VkPhysicalDeviceBufferDeviceAddressFeaturesKHR *features = (void *)ext;
1003 features->bufferDeviceAddress = pdevice->has_a64_buffer_access;
1004 features->bufferDeviceAddressCaptureReplay =
1005 pdevice->has_a64_buffer_access;
1006 features->bufferDeviceAddressMultiDevice = false;
1007 break;
1008 }
1009
1010 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COMPUTE_SHADER_DERIVATIVES_FEATURES_NV: {
1011 VkPhysicalDeviceComputeShaderDerivativesFeaturesNV *features =
1012 (VkPhysicalDeviceComputeShaderDerivativesFeaturesNV *)ext;
1013 features->computeDerivativeGroupQuads = true;
1014 features->computeDerivativeGroupLinear = true;
1015 break;
1016 }
1017
1018 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT: {
1019 VkPhysicalDeviceConditionalRenderingFeaturesEXT *features =
1020 (VkPhysicalDeviceConditionalRenderingFeaturesEXT*)ext;
1021 features->conditionalRendering = pdevice->info.gen >= 8 ||
1022 pdevice->info.is_haswell;
1023 features->inheritedConditionalRendering = pdevice->info.gen >= 8 ||
1024 pdevice->info.is_haswell;
1025 break;
1026 }
1027
1028 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_CLIP_ENABLE_FEATURES_EXT: {
1029 VkPhysicalDeviceDepthClipEnableFeaturesEXT *features =
1030 (VkPhysicalDeviceDepthClipEnableFeaturesEXT *)ext;
1031 features->depthClipEnable = true;
1032 break;
1033 }
1034
1035 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR: {
1036 VkPhysicalDeviceFloat16Int8FeaturesKHR *features = (void *)ext;
1037 features->shaderFloat16 = pdevice->info.gen >= 8;
1038 features->shaderInt8 = pdevice->info.gen >= 8;
1039 break;
1040 }
1041
1042 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FRAGMENT_SHADER_INTERLOCK_FEATURES_EXT: {
1043 VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT *features =
1044 (VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT *)ext;
1045 features->fragmentShaderSampleInterlock = pdevice->info.gen >= 9;
1046 features->fragmentShaderPixelInterlock = pdevice->info.gen >= 9;
1047 features->fragmentShaderShadingRateInterlock = false;
1048 break;
1049 }
1050
1051 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_QUERY_RESET_FEATURES_EXT: {
1052 VkPhysicalDeviceHostQueryResetFeaturesEXT *features =
1053 (VkPhysicalDeviceHostQueryResetFeaturesEXT *)ext;
1054 features->hostQueryReset = true;
1055 break;
1056 }
1057
1058 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT: {
1059 VkPhysicalDeviceDescriptorIndexingFeaturesEXT *features =
1060 (VkPhysicalDeviceDescriptorIndexingFeaturesEXT *)ext;
1061 features->shaderInputAttachmentArrayDynamicIndexing = false;
1062 features->shaderUniformTexelBufferArrayDynamicIndexing = true;
1063 features->shaderStorageTexelBufferArrayDynamicIndexing = true;
1064 features->shaderUniformBufferArrayNonUniformIndexing = false;
1065 features->shaderSampledImageArrayNonUniformIndexing = true;
1066 features->shaderStorageBufferArrayNonUniformIndexing = true;
1067 features->shaderStorageImageArrayNonUniformIndexing = true;
1068 features->shaderInputAttachmentArrayNonUniformIndexing = false;
1069 features->shaderUniformTexelBufferArrayNonUniformIndexing = true;
1070 features->shaderStorageTexelBufferArrayNonUniformIndexing = true;
1071 features->descriptorBindingUniformBufferUpdateAfterBind = false;
1072 features->descriptorBindingSampledImageUpdateAfterBind = true;
1073 features->descriptorBindingStorageImageUpdateAfterBind = true;
1074 features->descriptorBindingStorageBufferUpdateAfterBind = true;
1075 features->descriptorBindingUniformTexelBufferUpdateAfterBind = true;
1076 features->descriptorBindingStorageTexelBufferUpdateAfterBind = true;
1077 features->descriptorBindingUpdateUnusedWhilePending = true;
1078 features->descriptorBindingPartiallyBound = true;
1079 features->descriptorBindingVariableDescriptorCount = false;
1080 features->runtimeDescriptorArray = true;
1081 break;
1082 }
1083
1084 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INDEX_TYPE_UINT8_FEATURES_EXT: {
1085 VkPhysicalDeviceIndexTypeUint8FeaturesEXT *features =
1086 (VkPhysicalDeviceIndexTypeUint8FeaturesEXT *)ext;
1087 features->indexTypeUint8 = true;
1088 break;
1089 }
1090
1091 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_FEATURES_EXT: {
1092 VkPhysicalDeviceInlineUniformBlockFeaturesEXT *features =
1093 (VkPhysicalDeviceInlineUniformBlockFeaturesEXT *)ext;
1094 features->inlineUniformBlock = true;
1095 features->descriptorBindingInlineUniformBlockUpdateAfterBind = true;
1096 break;
1097 }
1098
1099 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_FEATURES_EXT: {
1100 VkPhysicalDeviceLineRasterizationFeaturesEXT *features =
1101 (VkPhysicalDeviceLineRasterizationFeaturesEXT *)ext;
1102 features->rectangularLines = true;
1103 features->bresenhamLines = true;
1104 /* Support for Smooth lines with MSAA was removed on gen11. From the
1105 * BSpec section "Multisample ModesState" table for "AA Line Support
1106 * Requirements":
1107 *
1108 * GEN10:BUG:######## NUM_MULTISAMPLES == 1
1109 *
1110 * Fortunately, this isn't a case most people care about.
1111 */
1112 features->smoothLines = pdevice->info.gen < 10;
1113 features->stippledRectangularLines = false;
1114 features->stippledBresenhamLines = true;
1115 features->stippledSmoothLines = false;
1116 break;
1117 }
1118
1119 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES: {
1120 VkPhysicalDeviceMultiviewFeatures *features =
1121 (VkPhysicalDeviceMultiviewFeatures *)ext;
1122 features->multiview = true;
1123 features->multiviewGeometryShader = true;
1124 features->multiviewTessellationShader = true;
1125 break;
1126 }
1127
1128 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGELESS_FRAMEBUFFER_FEATURES_KHR: {
1129 VkPhysicalDeviceImagelessFramebufferFeaturesKHR *features =
1130 (VkPhysicalDeviceImagelessFramebufferFeaturesKHR *)ext;
1131 features->imagelessFramebuffer = true;
1132 break;
1133 }
1134
1135 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR: {
1136 VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR *features =
1137 (VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR *)ext;
1138 features->pipelineExecutableInfo = true;
1139 break;
1140 }
1141
1142 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES: {
1143 VkPhysicalDeviceProtectedMemoryFeatures *features = (void *)ext;
1144 features->protectedMemory = false;
1145 break;
1146 }
1147
1148 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES: {
1149 VkPhysicalDeviceSamplerYcbcrConversionFeatures *features =
1150 (VkPhysicalDeviceSamplerYcbcrConversionFeatures *) ext;
1151 features->samplerYcbcrConversion = true;
1152 break;
1153 }
1154
1155 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES_EXT: {
1156 VkPhysicalDeviceScalarBlockLayoutFeaturesEXT *features =
1157 (VkPhysicalDeviceScalarBlockLayoutFeaturesEXT *)ext;
1158 features->scalarBlockLayout = true;
1159 break;
1160 }
1161
1162 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SEPARATE_DEPTH_STENCIL_LAYOUTS_FEATURES_KHR: {
1163 VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR *features =
1164 (VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR *)ext;
1165 features->separateDepthStencilLayouts = true;
1166 break;
1167 }
1168
1169 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_INT64_FEATURES_KHR: {
1170 VkPhysicalDeviceShaderAtomicInt64FeaturesKHR *features = (void *)ext;
1171 features->shaderBufferInt64Atomics =
1172 pdevice->info.gen >= 9 && pdevice->use_softpin;
1173 features->shaderSharedInt64Atomics = VK_FALSE;
1174 break;
1175 }
1176
1177 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DEMOTE_TO_HELPER_INVOCATION_FEATURES_EXT: {
1178 VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT *features = (void *)ext;
1179 features->shaderDemoteToHelperInvocation = true;
1180 break;
1181 }
1182
1183 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CLOCK_FEATURES_KHR: {
1184 VkPhysicalDeviceShaderClockFeaturesKHR *features =
1185 (VkPhysicalDeviceShaderClockFeaturesKHR *)ext;
1186 features->shaderSubgroupClock = true;
1187 features->shaderDeviceClock = false;
1188 break;
1189 }
1190
1191 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETERS_FEATURES: {
1192 VkPhysicalDeviceShaderDrawParametersFeatures *features = (void *)ext;
1193 features->shaderDrawParameters = true;
1194 break;
1195 }
1196
1197 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_SUBGROUP_EXTENDED_TYPES_FEATURES_KHR: {
1198 VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR *features =
1199 (VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR *)ext;
1200 features->shaderSubgroupExtendedTypes = true;
1201 break;
1202 }
1203
1204 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_FEATURES_EXT: {
1205 VkPhysicalDeviceSubgroupSizeControlFeaturesEXT *features =
1206 (VkPhysicalDeviceSubgroupSizeControlFeaturesEXT *)ext;
1207 features->subgroupSizeControl = true;
1208 features->computeFullSubgroups = true;
1209 break;
1210 }
1211
1212 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_FEATURES_EXT: {
1213 VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT *features =
1214 (VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT *)ext;
1215 features->texelBufferAlignment = true;
1216 break;
1217 }
1218
1219 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_FEATURES_KHR: {
1220 VkPhysicalDeviceTimelineSemaphoreFeaturesKHR *features =
1221 (VkPhysicalDeviceTimelineSemaphoreFeaturesKHR *) ext;
1222 features->timelineSemaphore = true;
1223 break;
1224 }
1225
1226 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES: {
1227 VkPhysicalDeviceVariablePointersFeatures *features = (void *)ext;
1228 features->variablePointersStorageBuffer = true;
1229 features->variablePointers = true;
1230 break;
1231 }
1232
1233 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT: {
1234 VkPhysicalDeviceTransformFeedbackFeaturesEXT *features =
1235 (VkPhysicalDeviceTransformFeedbackFeaturesEXT *)ext;
1236 features->transformFeedback = true;
1237 features->geometryStreams = true;
1238 break;
1239 }
1240
1241 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_UNIFORM_BUFFER_STANDARD_LAYOUT_FEATURES_KHR: {
1242 VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR *features =
1243 (VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR *)ext;
1244 features->uniformBufferStandardLayout = true;
1245 break;
1246 }
1247
1248 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT: {
1249 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *features =
1250 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *)ext;
1251 features->vertexAttributeInstanceRateDivisor = true;
1252 features->vertexAttributeInstanceRateZeroDivisor = true;
1253 break;
1254 }
1255
1256 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_MEMORY_MODEL_FEATURES_KHR: {
1257 VkPhysicalDeviceVulkanMemoryModelFeaturesKHR *features = (void *)ext;
1258 features->vulkanMemoryModel = true;
1259 features->vulkanMemoryModelDeviceScope = true;
1260 features->vulkanMemoryModelAvailabilityVisibilityChains = true;
1261 break;
1262 }
1263
1264 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_YCBCR_IMAGE_ARRAYS_FEATURES_EXT: {
1265 VkPhysicalDeviceYcbcrImageArraysFeaturesEXT *features =
1266 (VkPhysicalDeviceYcbcrImageArraysFeaturesEXT *)ext;
1267 features->ycbcrImageArrays = true;
1268 break;
1269 }
1270
1271 default:
1272 anv_debug_ignored_stype(ext->sType);
1273 break;
1274 }
1275 }
1276 }
1277
1278 #define MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS 64
1279
1280 #define MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS 64
1281 #define MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS 256
1282
1283 void anv_GetPhysicalDeviceProperties(
1284 VkPhysicalDevice physicalDevice,
1285 VkPhysicalDeviceProperties* pProperties)
1286 {
1287 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
1288 const struct gen_device_info *devinfo = &pdevice->info;
1289
1290 /* See assertions made when programming the buffer surface state. */
1291 const uint32_t max_raw_buffer_sz = devinfo->gen >= 7 ?
1292 (1ul << 30) : (1ul << 27);
1293
1294 const uint32_t max_ssbos = pdevice->has_a64_buffer_access ? UINT16_MAX : 64;
1295 const uint32_t max_textures =
1296 pdevice->has_bindless_images ? UINT16_MAX : 128;
1297 const uint32_t max_samplers =
1298 pdevice->has_bindless_samplers ? UINT16_MAX :
1299 (devinfo->gen >= 8 || devinfo->is_haswell) ? 128 : 16;
1300 const uint32_t max_images =
1301 pdevice->has_bindless_images ? UINT16_MAX : MAX_IMAGES;
1302
1303 /* If we can use bindless for everything, claim a high per-stage limit,
1304 * otherwise use the binding table size, minus the slots reserved for
1305 * render targets and one slot for the descriptor buffer. */
1306 const uint32_t max_per_stage =
1307 pdevice->has_bindless_images && pdevice->has_a64_buffer_access
1308 ? UINT32_MAX : MAX_BINDING_TABLE_SIZE - MAX_RTS - 1;
1309
1310 const uint32_t max_workgroup_size = 32 * devinfo->max_cs_threads;
1311
1312 VkSampleCountFlags sample_counts =
1313 isl_device_get_sample_counts(&pdevice->isl_dev);
1314
1315
1316 VkPhysicalDeviceLimits limits = {
1317 .maxImageDimension1D = (1 << 14),
1318 .maxImageDimension2D = (1 << 14),
1319 .maxImageDimension3D = (1 << 11),
1320 .maxImageDimensionCube = (1 << 14),
1321 .maxImageArrayLayers = (1 << 11),
1322 .maxTexelBufferElements = 128 * 1024 * 1024,
1323 .maxUniformBufferRange = (1ul << 27),
1324 .maxStorageBufferRange = max_raw_buffer_sz,
1325 .maxPushConstantsSize = MAX_PUSH_CONSTANTS_SIZE,
1326 .maxMemoryAllocationCount = UINT32_MAX,
1327 .maxSamplerAllocationCount = 64 * 1024,
1328 .bufferImageGranularity = 64, /* A cache line */
1329 .sparseAddressSpaceSize = 0,
1330 .maxBoundDescriptorSets = MAX_SETS,
1331 .maxPerStageDescriptorSamplers = max_samplers,
1332 .maxPerStageDescriptorUniformBuffers = MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS,
1333 .maxPerStageDescriptorStorageBuffers = max_ssbos,
1334 .maxPerStageDescriptorSampledImages = max_textures,
1335 .maxPerStageDescriptorStorageImages = max_images,
1336 .maxPerStageDescriptorInputAttachments = MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS,
1337 .maxPerStageResources = max_per_stage,
1338 .maxDescriptorSetSamplers = 6 * max_samplers, /* number of stages * maxPerStageDescriptorSamplers */
1339 .maxDescriptorSetUniformBuffers = 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS, /* number of stages * maxPerStageDescriptorUniformBuffers */
1340 .maxDescriptorSetUniformBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2,
1341 .maxDescriptorSetStorageBuffers = 6 * max_ssbos, /* number of stages * maxPerStageDescriptorStorageBuffers */
1342 .maxDescriptorSetStorageBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2,
1343 .maxDescriptorSetSampledImages = 6 * max_textures, /* number of stages * maxPerStageDescriptorSampledImages */
1344 .maxDescriptorSetStorageImages = 6 * max_images, /* number of stages * maxPerStageDescriptorStorageImages */
1345 .maxDescriptorSetInputAttachments = MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS,
1346 .maxVertexInputAttributes = MAX_VBS,
1347 .maxVertexInputBindings = MAX_VBS,
1348 .maxVertexInputAttributeOffset = 2047,
1349 .maxVertexInputBindingStride = 2048,
1350 .maxVertexOutputComponents = 128,
1351 .maxTessellationGenerationLevel = 64,
1352 .maxTessellationPatchSize = 32,
1353 .maxTessellationControlPerVertexInputComponents = 128,
1354 .maxTessellationControlPerVertexOutputComponents = 128,
1355 .maxTessellationControlPerPatchOutputComponents = 128,
1356 .maxTessellationControlTotalOutputComponents = 2048,
1357 .maxTessellationEvaluationInputComponents = 128,
1358 .maxTessellationEvaluationOutputComponents = 128,
1359 .maxGeometryShaderInvocations = 32,
1360 .maxGeometryInputComponents = 64,
1361 .maxGeometryOutputComponents = 128,
1362 .maxGeometryOutputVertices = 256,
1363 .maxGeometryTotalOutputComponents = 1024,
1364 .maxFragmentInputComponents = 116, /* 128 components - (PSIZ, CLIP_DIST0, CLIP_DIST1) */
1365 .maxFragmentOutputAttachments = 8,
1366 .maxFragmentDualSrcAttachments = 1,
1367 .maxFragmentCombinedOutputResources = 8,
1368 .maxComputeSharedMemorySize = 64 * 1024,
1369 .maxComputeWorkGroupCount = { 65535, 65535, 65535 },
1370 .maxComputeWorkGroupInvocations = max_workgroup_size,
1371 .maxComputeWorkGroupSize = {
1372 max_workgroup_size,
1373 max_workgroup_size,
1374 max_workgroup_size,
1375 },
1376 .subPixelPrecisionBits = 8,
1377 .subTexelPrecisionBits = 8,
1378 .mipmapPrecisionBits = 8,
1379 .maxDrawIndexedIndexValue = UINT32_MAX,
1380 .maxDrawIndirectCount = UINT32_MAX,
1381 .maxSamplerLodBias = 16,
1382 .maxSamplerAnisotropy = 16,
1383 .maxViewports = MAX_VIEWPORTS,
1384 .maxViewportDimensions = { (1 << 14), (1 << 14) },
1385 .viewportBoundsRange = { INT16_MIN, INT16_MAX },
1386 .viewportSubPixelBits = 13, /* We take a float? */
1387 .minMemoryMapAlignment = 4096, /* A page */
1388 /* The dataport requires texel alignment so we need to assume a worst
1389 * case of R32G32B32A32 which is 16 bytes.
1390 */
1391 .minTexelBufferOffsetAlignment = 16,
1392 /* We need 16 for UBO block reads to work and 32 for push UBOs */
1393 .minUniformBufferOffsetAlignment = 32,
1394 .minStorageBufferOffsetAlignment = 4,
1395 .minTexelOffset = -8,
1396 .maxTexelOffset = 7,
1397 .minTexelGatherOffset = -32,
1398 .maxTexelGatherOffset = 31,
1399 .minInterpolationOffset = -0.5,
1400 .maxInterpolationOffset = 0.4375,
1401 .subPixelInterpolationOffsetBits = 4,
1402 .maxFramebufferWidth = (1 << 14),
1403 .maxFramebufferHeight = (1 << 14),
1404 .maxFramebufferLayers = (1 << 11),
1405 .framebufferColorSampleCounts = sample_counts,
1406 .framebufferDepthSampleCounts = sample_counts,
1407 .framebufferStencilSampleCounts = sample_counts,
1408 .framebufferNoAttachmentsSampleCounts = sample_counts,
1409 .maxColorAttachments = MAX_RTS,
1410 .sampledImageColorSampleCounts = sample_counts,
1411 .sampledImageIntegerSampleCounts = VK_SAMPLE_COUNT_1_BIT,
1412 .sampledImageDepthSampleCounts = sample_counts,
1413 .sampledImageStencilSampleCounts = sample_counts,
1414 .storageImageSampleCounts = VK_SAMPLE_COUNT_1_BIT,
1415 .maxSampleMaskWords = 1,
1416 .timestampComputeAndGraphics = true,
1417 .timestampPeriod = 1000000000.0 / devinfo->timestamp_frequency,
1418 .maxClipDistances = 8,
1419 .maxCullDistances = 8,
1420 .maxCombinedClipAndCullDistances = 8,
1421 .discreteQueuePriorities = 2,
1422 .pointSizeRange = { 0.125, 255.875 },
1423 .lineWidthRange = {
1424 0.0,
1425 (devinfo->gen >= 9 || devinfo->is_cherryview) ?
1426 2047.9921875 : 7.9921875,
1427 },
1428 .pointSizeGranularity = (1.0 / 8.0),
1429 .lineWidthGranularity = (1.0 / 128.0),
1430 .strictLines = false,
1431 .standardSampleLocations = true,
1432 .optimalBufferCopyOffsetAlignment = 128,
1433 .optimalBufferCopyRowPitchAlignment = 128,
1434 .nonCoherentAtomSize = 64,
1435 };
1436
1437 *pProperties = (VkPhysicalDeviceProperties) {
1438 .apiVersion = anv_physical_device_api_version(pdevice),
1439 .driverVersion = vk_get_driver_version(),
1440 .vendorID = 0x8086,
1441 .deviceID = pdevice->chipset_id,
1442 .deviceType = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
1443 .limits = limits,
1444 .sparseProperties = {0}, /* Broadwell doesn't do sparse. */
1445 };
1446
1447 snprintf(pProperties->deviceName, sizeof(pProperties->deviceName),
1448 "%s", pdevice->name);
1449 memcpy(pProperties->pipelineCacheUUID,
1450 pdevice->pipeline_cache_uuid, VK_UUID_SIZE);
1451 }
1452
1453 void anv_GetPhysicalDeviceProperties2(
1454 VkPhysicalDevice physicalDevice,
1455 VkPhysicalDeviceProperties2* pProperties)
1456 {
1457 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
1458
1459 anv_GetPhysicalDeviceProperties(physicalDevice, &pProperties->properties);
1460
1461 vk_foreach_struct(ext, pProperties->pNext) {
1462 switch (ext->sType) {
1463 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_STENCIL_RESOLVE_PROPERTIES_KHR: {
1464 VkPhysicalDeviceDepthStencilResolvePropertiesKHR *props =
1465 (VkPhysicalDeviceDepthStencilResolvePropertiesKHR *)ext;
1466
1467 /* We support all of the depth resolve modes */
1468 props->supportedDepthResolveModes =
1469 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR |
1470 VK_RESOLVE_MODE_AVERAGE_BIT_KHR |
1471 VK_RESOLVE_MODE_MIN_BIT_KHR |
1472 VK_RESOLVE_MODE_MAX_BIT_KHR;
1473
1474 /* Average doesn't make sense for stencil so we don't support that */
1475 props->supportedStencilResolveModes =
1476 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR;
1477 if (pdevice->info.gen >= 8) {
1478 /* The advanced stencil resolve modes currently require stencil
1479 * sampling be supported by the hardware.
1480 */
1481 props->supportedStencilResolveModes |=
1482 VK_RESOLVE_MODE_MIN_BIT_KHR |
1483 VK_RESOLVE_MODE_MAX_BIT_KHR;
1484 }
1485
1486 props->independentResolveNone = true;
1487 props->independentResolve = true;
1488 break;
1489 }
1490
1491 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES_EXT: {
1492 VkPhysicalDeviceDescriptorIndexingPropertiesEXT *props =
1493 (VkPhysicalDeviceDescriptorIndexingPropertiesEXT *)ext;
1494
1495 /* It's a bit hard to exactly map our implementation to the limits
1496 * described here. The bindless surface handle in the extended
1497 * message descriptors is 20 bits and it's an index into the table of
1498 * RENDER_SURFACE_STATE structs that starts at bindless surface base
1499 * address. Given that most things consume two surface states per
1500 * view (general/sampled for textures and write-only/read-write for
1501 * images), we claim 2^19 things.
1502 *
1503 * For SSBOs, we just use A64 messages so there is no real limit
1504 * there beyond the limit on the total size of a descriptor set.
1505 */
1506 const unsigned max_bindless_views = 1 << 19;
1507
1508 props->maxUpdateAfterBindDescriptorsInAllPools = max_bindless_views;
1509 props->shaderUniformBufferArrayNonUniformIndexingNative = false;
1510 props->shaderSampledImageArrayNonUniformIndexingNative = false;
1511 props->shaderStorageBufferArrayNonUniformIndexingNative = true;
1512 props->shaderStorageImageArrayNonUniformIndexingNative = false;
1513 props->shaderInputAttachmentArrayNonUniformIndexingNative = false;
1514 props->robustBufferAccessUpdateAfterBind = true;
1515 props->quadDivergentImplicitLod = false;
1516 props->maxPerStageDescriptorUpdateAfterBindSamplers = max_bindless_views;
1517 props->maxPerStageDescriptorUpdateAfterBindUniformBuffers = MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS;
1518 props->maxPerStageDescriptorUpdateAfterBindStorageBuffers = UINT32_MAX;
1519 props->maxPerStageDescriptorUpdateAfterBindSampledImages = max_bindless_views;
1520 props->maxPerStageDescriptorUpdateAfterBindStorageImages = max_bindless_views;
1521 props->maxPerStageDescriptorUpdateAfterBindInputAttachments = MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS;
1522 props->maxPerStageUpdateAfterBindResources = UINT32_MAX;
1523 props->maxDescriptorSetUpdateAfterBindSamplers = max_bindless_views;
1524 props->maxDescriptorSetUpdateAfterBindUniformBuffers = 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS;
1525 props->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2;
1526 props->maxDescriptorSetUpdateAfterBindStorageBuffers = UINT32_MAX;
1527 props->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2;
1528 props->maxDescriptorSetUpdateAfterBindSampledImages = max_bindless_views;
1529 props->maxDescriptorSetUpdateAfterBindStorageImages = max_bindless_views;
1530 props->maxDescriptorSetUpdateAfterBindInputAttachments = MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS;
1531 break;
1532 }
1533
1534 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR: {
1535 VkPhysicalDeviceDriverPropertiesKHR *driver_props =
1536 (VkPhysicalDeviceDriverPropertiesKHR *) ext;
1537
1538 driver_props->driverID = VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA_KHR;
1539 snprintf(driver_props->driverName, VK_MAX_DRIVER_NAME_SIZE_KHR,
1540 "Intel open-source Mesa driver");
1541
1542 snprintf(driver_props->driverInfo, VK_MAX_DRIVER_INFO_SIZE_KHR,
1543 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1);
1544
1545 driver_props->conformanceVersion = (VkConformanceVersionKHR) {
1546 .major = 1,
1547 .minor = 1,
1548 .subminor = 2,
1549 .patch = 0,
1550 };
1551 break;
1552 }
1553
1554 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT: {
1555 VkPhysicalDeviceExternalMemoryHostPropertiesEXT *props =
1556 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT *) ext;
1557 /* Userptr needs page aligned memory. */
1558 props->minImportedHostPointerAlignment = 4096;
1559 break;
1560 }
1561
1562 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES: {
1563 VkPhysicalDeviceIDProperties *id_props =
1564 (VkPhysicalDeviceIDProperties *)ext;
1565 memcpy(id_props->deviceUUID, pdevice->device_uuid, VK_UUID_SIZE);
1566 memcpy(id_props->driverUUID, pdevice->driver_uuid, VK_UUID_SIZE);
1567 /* The LUID is for Windows. */
1568 id_props->deviceLUIDValid = false;
1569 break;
1570 }
1571
1572 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_PROPERTIES_EXT: {
1573 VkPhysicalDeviceInlineUniformBlockPropertiesEXT *props =
1574 (VkPhysicalDeviceInlineUniformBlockPropertiesEXT *)ext;
1575 props->maxInlineUniformBlockSize = MAX_INLINE_UNIFORM_BLOCK_SIZE;
1576 props->maxPerStageDescriptorInlineUniformBlocks =
1577 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1578 props->maxPerStageDescriptorUpdateAfterBindInlineUniformBlocks =
1579 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1580 props->maxDescriptorSetInlineUniformBlocks =
1581 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1582 props->maxDescriptorSetUpdateAfterBindInlineUniformBlocks =
1583 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1584 break;
1585 }
1586
1587 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_PROPERTIES_EXT: {
1588 VkPhysicalDeviceLineRasterizationPropertiesEXT *props =
1589 (VkPhysicalDeviceLineRasterizationPropertiesEXT *)ext;
1590 /* In the Skylake PRM Vol. 7, subsection titled "GIQ (Diamond)
1591 * Sampling Rules - Legacy Mode", it says the following:
1592 *
1593 * "Note that the device divides a pixel into a 16x16 array of
1594 * subpixels, referenced by their upper left corners."
1595 *
1596 * This is the only known reference in the PRMs to the subpixel
1597 * precision of line rasterization and a "16x16 array of subpixels"
1598 * implies 4 subpixel precision bits. Empirical testing has shown
1599 * that 4 subpixel precision bits applies to all line rasterization
1600 * types.
1601 */
1602 props->lineSubPixelPrecisionBits = 4;
1603 break;
1604 }
1605
1606 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES: {
1607 VkPhysicalDeviceMaintenance3Properties *props =
1608 (VkPhysicalDeviceMaintenance3Properties *)ext;
1609 /* This value doesn't matter for us today as our per-stage
1610 * descriptors are the real limit.
1611 */
1612 props->maxPerSetDescriptors = 1024;
1613 props->maxMemoryAllocationSize = MAX_MEMORY_ALLOCATION_SIZE;
1614 break;
1615 }
1616
1617 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES: {
1618 VkPhysicalDeviceMultiviewProperties *properties =
1619 (VkPhysicalDeviceMultiviewProperties *)ext;
1620 properties->maxMultiviewViewCount = 16;
1621 properties->maxMultiviewInstanceIndex = UINT32_MAX / 16;
1622 break;
1623 }
1624
1625 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT: {
1626 VkPhysicalDevicePCIBusInfoPropertiesEXT *properties =
1627 (VkPhysicalDevicePCIBusInfoPropertiesEXT *)ext;
1628 properties->pciDomain = pdevice->pci_info.domain;
1629 properties->pciBus = pdevice->pci_info.bus;
1630 properties->pciDevice = pdevice->pci_info.device;
1631 properties->pciFunction = pdevice->pci_info.function;
1632 break;
1633 }
1634
1635 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES: {
1636 VkPhysicalDevicePointClippingProperties *properties =
1637 (VkPhysicalDevicePointClippingProperties *) ext;
1638 properties->pointClippingBehavior = VK_POINT_CLIPPING_BEHAVIOR_USER_CLIP_PLANES_ONLY;
1639 break;
1640 }
1641
1642 #pragma GCC diagnostic push
1643 #pragma GCC diagnostic ignored "-Wswitch"
1644 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PRESENTATION_PROPERTIES_ANDROID: {
1645 VkPhysicalDevicePresentationPropertiesANDROID *props =
1646 (VkPhysicalDevicePresentationPropertiesANDROID *)ext;
1647 props->sharedImage = VK_FALSE;
1648 break;
1649 }
1650 #pragma GCC diagnostic pop
1651
1652 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES: {
1653 VkPhysicalDeviceProtectedMemoryProperties *props =
1654 (VkPhysicalDeviceProtectedMemoryProperties *)ext;
1655 props->protectedNoFault = false;
1656 break;
1657 }
1658
1659 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR: {
1660 VkPhysicalDevicePushDescriptorPropertiesKHR *properties =
1661 (VkPhysicalDevicePushDescriptorPropertiesKHR *) ext;
1662
1663 properties->maxPushDescriptors = MAX_PUSH_DESCRIPTORS;
1664 break;
1665 }
1666
1667 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT: {
1668 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT *properties =
1669 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT *)ext;
1670 properties->filterMinmaxImageComponentMapping = pdevice->info.gen >= 9;
1671 properties->filterMinmaxSingleComponentFormats = true;
1672 break;
1673 }
1674
1675 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES: {
1676 VkPhysicalDeviceSubgroupProperties *properties = (void *)ext;
1677
1678 properties->subgroupSize = BRW_SUBGROUP_SIZE;
1679
1680 VkShaderStageFlags scalar_stages = 0;
1681 for (unsigned stage = 0; stage < MESA_SHADER_STAGES; stage++) {
1682 if (pdevice->compiler->scalar_stage[stage])
1683 scalar_stages |= mesa_to_vk_shader_stage(stage);
1684 }
1685 properties->supportedStages = scalar_stages;
1686
1687 properties->supportedOperations = VK_SUBGROUP_FEATURE_BASIC_BIT |
1688 VK_SUBGROUP_FEATURE_VOTE_BIT |
1689 VK_SUBGROUP_FEATURE_BALLOT_BIT |
1690 VK_SUBGROUP_FEATURE_SHUFFLE_BIT |
1691 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT |
1692 VK_SUBGROUP_FEATURE_QUAD_BIT;
1693 if (pdevice->info.gen >= 8) {
1694 /* TODO: There's no technical reason why these can't be made to
1695 * work on gen7 but they don't at the moment so it's best to leave
1696 * the feature disabled than enabled and broken.
1697 */
1698 properties->supportedOperations |=
1699 VK_SUBGROUP_FEATURE_ARITHMETIC_BIT |
1700 VK_SUBGROUP_FEATURE_CLUSTERED_BIT;
1701 }
1702 properties->quadOperationsInAllStages = pdevice->info.gen >= 8;
1703 break;
1704 }
1705
1706 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_PROPERTIES_EXT: {
1707 VkPhysicalDeviceSubgroupSizeControlPropertiesEXT *props =
1708 (VkPhysicalDeviceSubgroupSizeControlPropertiesEXT *)ext;
1709 STATIC_ASSERT(8 <= BRW_SUBGROUP_SIZE && BRW_SUBGROUP_SIZE <= 32);
1710 props->minSubgroupSize = 8;
1711 props->maxSubgroupSize = 32;
1712 props->maxComputeWorkgroupSubgroups = pdevice->info.max_cs_threads;
1713 props->requiredSubgroupSizeStages = VK_SHADER_STAGE_COMPUTE_BIT;
1714 break;
1715 }
1716 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT_CONTROLS_PROPERTIES_KHR : {
1717 VkPhysicalDeviceFloatControlsPropertiesKHR *properties = (void *)ext;
1718 properties->denormBehaviorIndependence = VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_ALL_KHR;
1719 properties->roundingModeIndependence = VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_NONE_KHR;
1720
1721 /* Broadwell does not support HF denorms and there are restrictions
1722 * other gens. According to Kabylake's PRM:
1723 *
1724 * "math - Extended Math Function
1725 * [...]
1726 * Restriction : Half-float denorms are always retained."
1727 */
1728 properties->shaderDenormFlushToZeroFloat16 = false;
1729 properties->shaderDenormPreserveFloat16 = pdevice->info.gen > 8;
1730 properties->shaderRoundingModeRTEFloat16 = true;
1731 properties->shaderRoundingModeRTZFloat16 = true;
1732 properties->shaderSignedZeroInfNanPreserveFloat16 = true;
1733
1734 properties->shaderDenormFlushToZeroFloat32 = true;
1735 properties->shaderDenormPreserveFloat32 = true;
1736 properties->shaderRoundingModeRTEFloat32 = true;
1737 properties->shaderRoundingModeRTZFloat32 = true;
1738 properties->shaderSignedZeroInfNanPreserveFloat32 = true;
1739
1740 properties->shaderDenormFlushToZeroFloat64 = true;
1741 properties->shaderDenormPreserveFloat64 = true;
1742 properties->shaderRoundingModeRTEFloat64 = true;
1743 properties->shaderRoundingModeRTZFloat64 = true;
1744 properties->shaderSignedZeroInfNanPreserveFloat64 = true;
1745 break;
1746 }
1747
1748 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_PROPERTIES_EXT: {
1749 VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT *props =
1750 (VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT *)ext;
1751
1752 /* From the SKL PRM Vol. 2d, docs for RENDER_SURFACE_STATE::Surface
1753 * Base Address:
1754 *
1755 * "For SURFTYPE_BUFFER non-rendertarget surfaces, this field
1756 * specifies the base address of the first element of the surface,
1757 * computed in software by adding the surface base address to the
1758 * byte offset of the element in the buffer. The base address must
1759 * be aligned to element size."
1760 *
1761 * The typed dataport messages require that things be texel aligned.
1762 * Otherwise, we may just load/store the wrong data or, in the worst
1763 * case, there may be hangs.
1764 */
1765 props->storageTexelBufferOffsetAlignmentBytes = 16;
1766 props->storageTexelBufferOffsetSingleTexelAlignment = true;
1767
1768 /* The sampler, however, is much more forgiving and it can handle
1769 * arbitrary byte alignment for linear and buffer surfaces. It's
1770 * hard to find a good PRM citation for this but years of empirical
1771 * experience demonstrate that this is true.
1772 */
1773 props->uniformTexelBufferOffsetAlignmentBytes = 1;
1774 props->uniformTexelBufferOffsetSingleTexelAlignment = false;
1775 break;
1776 }
1777
1778 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_PROPERTIES_KHR: {
1779 VkPhysicalDeviceTimelineSemaphorePropertiesKHR *props =
1780 (VkPhysicalDeviceTimelineSemaphorePropertiesKHR *) ext;
1781 props->maxTimelineSemaphoreValueDifference = UINT64_MAX;
1782 break;
1783 }
1784
1785 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT: {
1786 VkPhysicalDeviceTransformFeedbackPropertiesEXT *props =
1787 (VkPhysicalDeviceTransformFeedbackPropertiesEXT *)ext;
1788
1789 props->maxTransformFeedbackStreams = MAX_XFB_STREAMS;
1790 props->maxTransformFeedbackBuffers = MAX_XFB_BUFFERS;
1791 props->maxTransformFeedbackBufferSize = (1ull << 32);
1792 props->maxTransformFeedbackStreamDataSize = 128 * 4;
1793 props->maxTransformFeedbackBufferDataSize = 128 * 4;
1794 props->maxTransformFeedbackBufferDataStride = 2048;
1795 props->transformFeedbackQueries = true;
1796 props->transformFeedbackStreamsLinesTriangles = false;
1797 props->transformFeedbackRasterizationStreamSelect = false;
1798 props->transformFeedbackDraw = true;
1799 break;
1800 }
1801
1802 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT: {
1803 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *props =
1804 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *)ext;
1805 /* We have to restrict this a bit for multiview */
1806 props->maxVertexAttribDivisor = UINT32_MAX / 16;
1807 break;
1808 }
1809
1810 default:
1811 anv_debug_ignored_stype(ext->sType);
1812 break;
1813 }
1814 }
1815 }
1816
1817 /* We support exactly one queue family. */
1818 static const VkQueueFamilyProperties
1819 anv_queue_family_properties = {
1820 .queueFlags = VK_QUEUE_GRAPHICS_BIT |
1821 VK_QUEUE_COMPUTE_BIT |
1822 VK_QUEUE_TRANSFER_BIT,
1823 .queueCount = 1,
1824 .timestampValidBits = 36, /* XXX: Real value here */
1825 .minImageTransferGranularity = { 1, 1, 1 },
1826 };
1827
1828 void anv_GetPhysicalDeviceQueueFamilyProperties(
1829 VkPhysicalDevice physicalDevice,
1830 uint32_t* pCount,
1831 VkQueueFamilyProperties* pQueueFamilyProperties)
1832 {
1833 VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pCount);
1834
1835 vk_outarray_append(&out, p) {
1836 *p = anv_queue_family_properties;
1837 }
1838 }
1839
1840 void anv_GetPhysicalDeviceQueueFamilyProperties2(
1841 VkPhysicalDevice physicalDevice,
1842 uint32_t* pQueueFamilyPropertyCount,
1843 VkQueueFamilyProperties2* pQueueFamilyProperties)
1844 {
1845
1846 VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pQueueFamilyPropertyCount);
1847
1848 vk_outarray_append(&out, p) {
1849 p->queueFamilyProperties = anv_queue_family_properties;
1850
1851 vk_foreach_struct(s, p->pNext) {
1852 anv_debug_ignored_stype(s->sType);
1853 }
1854 }
1855 }
1856
1857 void anv_GetPhysicalDeviceMemoryProperties(
1858 VkPhysicalDevice physicalDevice,
1859 VkPhysicalDeviceMemoryProperties* pMemoryProperties)
1860 {
1861 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
1862
1863 pMemoryProperties->memoryTypeCount = physical_device->memory.type_count;
1864 for (uint32_t i = 0; i < physical_device->memory.type_count; i++) {
1865 pMemoryProperties->memoryTypes[i] = (VkMemoryType) {
1866 .propertyFlags = physical_device->memory.types[i].propertyFlags,
1867 .heapIndex = physical_device->memory.types[i].heapIndex,
1868 };
1869 }
1870
1871 pMemoryProperties->memoryHeapCount = physical_device->memory.heap_count;
1872 for (uint32_t i = 0; i < physical_device->memory.heap_count; i++) {
1873 pMemoryProperties->memoryHeaps[i] = (VkMemoryHeap) {
1874 .size = physical_device->memory.heaps[i].size,
1875 .flags = physical_device->memory.heaps[i].flags,
1876 };
1877 }
1878 }
1879
1880 static void
1881 anv_get_memory_budget(VkPhysicalDevice physicalDevice,
1882 VkPhysicalDeviceMemoryBudgetPropertiesEXT *memoryBudget)
1883 {
1884 ANV_FROM_HANDLE(anv_physical_device, device, physicalDevice);
1885 uint64_t sys_available = get_available_system_memory();
1886 assert(sys_available > 0);
1887
1888 VkDeviceSize total_heaps_size = 0;
1889 for (size_t i = 0; i < device->memory.heap_count; i++)
1890 total_heaps_size += device->memory.heaps[i].size;
1891
1892 for (size_t i = 0; i < device->memory.heap_count; i++) {
1893 VkDeviceSize heap_size = device->memory.heaps[i].size;
1894 VkDeviceSize heap_used = device->memory.heaps[i].used;
1895 VkDeviceSize heap_budget;
1896
1897 double heap_proportion = (double) heap_size / total_heaps_size;
1898 VkDeviceSize sys_available_prop = sys_available * heap_proportion;
1899
1900 /*
1901 * Let's not incite the app to starve the system: report at most 90% of
1902 * available system memory.
1903 */
1904 uint64_t heap_available = sys_available_prop * 9 / 10;
1905 heap_budget = MIN2(heap_size, heap_used + heap_available);
1906
1907 /*
1908 * Round down to the nearest MB
1909 */
1910 heap_budget &= ~((1ull << 20) - 1);
1911
1912 /*
1913 * The heapBudget value must be non-zero for array elements less than
1914 * VkPhysicalDeviceMemoryProperties::memoryHeapCount. The heapBudget
1915 * value must be less than or equal to VkMemoryHeap::size for each heap.
1916 */
1917 assert(0 < heap_budget && heap_budget <= heap_size);
1918
1919 memoryBudget->heapUsage[i] = heap_used;
1920 memoryBudget->heapBudget[i] = heap_budget;
1921 }
1922
1923 /* The heapBudget and heapUsage values must be zero for array elements
1924 * greater than or equal to VkPhysicalDeviceMemoryProperties::memoryHeapCount
1925 */
1926 for (uint32_t i = device->memory.heap_count; i < VK_MAX_MEMORY_HEAPS; i++) {
1927 memoryBudget->heapBudget[i] = 0;
1928 memoryBudget->heapUsage[i] = 0;
1929 }
1930 }
1931
1932 void anv_GetPhysicalDeviceMemoryProperties2(
1933 VkPhysicalDevice physicalDevice,
1934 VkPhysicalDeviceMemoryProperties2* pMemoryProperties)
1935 {
1936 anv_GetPhysicalDeviceMemoryProperties(physicalDevice,
1937 &pMemoryProperties->memoryProperties);
1938
1939 vk_foreach_struct(ext, pMemoryProperties->pNext) {
1940 switch (ext->sType) {
1941 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT:
1942 anv_get_memory_budget(physicalDevice, (void*)ext);
1943 break;
1944 default:
1945 anv_debug_ignored_stype(ext->sType);
1946 break;
1947 }
1948 }
1949 }
1950
1951 void
1952 anv_GetDeviceGroupPeerMemoryFeatures(
1953 VkDevice device,
1954 uint32_t heapIndex,
1955 uint32_t localDeviceIndex,
1956 uint32_t remoteDeviceIndex,
1957 VkPeerMemoryFeatureFlags* pPeerMemoryFeatures)
1958 {
1959 assert(localDeviceIndex == 0 && remoteDeviceIndex == 0);
1960 *pPeerMemoryFeatures = VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT |
1961 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT |
1962 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT |
1963 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT;
1964 }
1965
1966 PFN_vkVoidFunction anv_GetInstanceProcAddr(
1967 VkInstance _instance,
1968 const char* pName)
1969 {
1970 ANV_FROM_HANDLE(anv_instance, instance, _instance);
1971
1972 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
1973 * when we have to return valid function pointers, NULL, or it's left
1974 * undefined. See the table for exact details.
1975 */
1976 if (pName == NULL)
1977 return NULL;
1978
1979 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
1980 if (strcmp(pName, "vk" #entrypoint) == 0) \
1981 return (PFN_vkVoidFunction)anv_##entrypoint
1982
1983 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties);
1984 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties);
1985 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceVersion);
1986 LOOKUP_ANV_ENTRYPOINT(CreateInstance);
1987
1988 #undef LOOKUP_ANV_ENTRYPOINT
1989
1990 if (instance == NULL)
1991 return NULL;
1992
1993 int idx = anv_get_instance_entrypoint_index(pName);
1994 if (idx >= 0)
1995 return instance->dispatch.entrypoints[idx];
1996
1997 idx = anv_get_physical_device_entrypoint_index(pName);
1998 if (idx >= 0)
1999 return instance->physicalDevice.dispatch.entrypoints[idx];
2000
2001 idx = anv_get_device_entrypoint_index(pName);
2002 if (idx >= 0)
2003 return instance->device_dispatch.entrypoints[idx];
2004
2005 return NULL;
2006 }
2007
2008 /* With version 1+ of the loader interface the ICD should expose
2009 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
2010 */
2011 PUBLIC
2012 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
2013 VkInstance instance,
2014 const char* pName);
2015
2016 PUBLIC
2017 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
2018 VkInstance instance,
2019 const char* pName)
2020 {
2021 return anv_GetInstanceProcAddr(instance, pName);
2022 }
2023
2024 PFN_vkVoidFunction anv_GetDeviceProcAddr(
2025 VkDevice _device,
2026 const char* pName)
2027 {
2028 ANV_FROM_HANDLE(anv_device, device, _device);
2029
2030 if (!device || !pName)
2031 return NULL;
2032
2033 int idx = anv_get_device_entrypoint_index(pName);
2034 if (idx < 0)
2035 return NULL;
2036
2037 return device->dispatch.entrypoints[idx];
2038 }
2039
2040 /* With version 4+ of the loader interface the ICD should expose
2041 * vk_icdGetPhysicalDeviceProcAddr()
2042 */
2043 PUBLIC
2044 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetPhysicalDeviceProcAddr(
2045 VkInstance _instance,
2046 const char* pName);
2047
2048 PFN_vkVoidFunction vk_icdGetPhysicalDeviceProcAddr(
2049 VkInstance _instance,
2050 const char* pName)
2051 {
2052 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2053
2054 if (!pName || !instance)
2055 return NULL;
2056
2057 int idx = anv_get_physical_device_entrypoint_index(pName);
2058 if (idx < 0)
2059 return NULL;
2060
2061 return instance->physicalDevice.dispatch.entrypoints[idx];
2062 }
2063
2064
2065 VkResult
2066 anv_CreateDebugReportCallbackEXT(VkInstance _instance,
2067 const VkDebugReportCallbackCreateInfoEXT* pCreateInfo,
2068 const VkAllocationCallbacks* pAllocator,
2069 VkDebugReportCallbackEXT* pCallback)
2070 {
2071 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2072 return vk_create_debug_report_callback(&instance->debug_report_callbacks,
2073 pCreateInfo, pAllocator, &instance->alloc,
2074 pCallback);
2075 }
2076
2077 void
2078 anv_DestroyDebugReportCallbackEXT(VkInstance _instance,
2079 VkDebugReportCallbackEXT _callback,
2080 const VkAllocationCallbacks* pAllocator)
2081 {
2082 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2083 vk_destroy_debug_report_callback(&instance->debug_report_callbacks,
2084 _callback, pAllocator, &instance->alloc);
2085 }
2086
2087 void
2088 anv_DebugReportMessageEXT(VkInstance _instance,
2089 VkDebugReportFlagsEXT flags,
2090 VkDebugReportObjectTypeEXT objectType,
2091 uint64_t object,
2092 size_t location,
2093 int32_t messageCode,
2094 const char* pLayerPrefix,
2095 const char* pMessage)
2096 {
2097 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2098 vk_debug_report(&instance->debug_report_callbacks, flags, objectType,
2099 object, location, messageCode, pLayerPrefix, pMessage);
2100 }
2101
2102 static struct anv_state
2103 anv_state_pool_emit_data(struct anv_state_pool *pool, size_t size, size_t align, const void *p)
2104 {
2105 struct anv_state state;
2106
2107 state = anv_state_pool_alloc(pool, size, align);
2108 memcpy(state.map, p, size);
2109
2110 return state;
2111 }
2112
2113 /* Haswell border color is a bit of a disaster. Float and unorm formats use a
2114 * straightforward 32-bit float color in the first 64 bytes. Instead of using
2115 * a nice float/integer union like Gen8+, Haswell specifies the integer border
2116 * color as a separate entry /after/ the float color. The layout of this entry
2117 * also depends on the format's bpp (with extra hacks for RG32), and overlaps.
2118 *
2119 * Since we don't know the format/bpp, we can't make any of the border colors
2120 * containing '1' work for all formats, as it would be in the wrong place for
2121 * some of them. We opt to make 32-bit integers work as this seems like the
2122 * most common option. Fortunately, transparent black works regardless, as
2123 * all zeroes is the same in every bit-size.
2124 */
2125 struct hsw_border_color {
2126 float float32[4];
2127 uint32_t _pad0[12];
2128 uint32_t uint32[4];
2129 uint32_t _pad1[108];
2130 };
2131
2132 struct gen8_border_color {
2133 union {
2134 float float32[4];
2135 uint32_t uint32[4];
2136 };
2137 /* Pad out to 64 bytes */
2138 uint32_t _pad[12];
2139 };
2140
2141 static void
2142 anv_device_init_border_colors(struct anv_device *device)
2143 {
2144 if (device->info.is_haswell) {
2145 static const struct hsw_border_color border_colors[] = {
2146 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
2147 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
2148 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
2149 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = { .uint32 = { 0, 0, 0, 0 } },
2150 [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = { .uint32 = { 0, 0, 0, 1 } },
2151 [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = { .uint32 = { 1, 1, 1, 1 } },
2152 };
2153
2154 device->border_colors =
2155 anv_state_pool_emit_data(&device->dynamic_state_pool,
2156 sizeof(border_colors), 512, border_colors);
2157 } else {
2158 static const struct gen8_border_color border_colors[] = {
2159 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
2160 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
2161 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
2162 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = { .uint32 = { 0, 0, 0, 0 } },
2163 [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = { .uint32 = { 0, 0, 0, 1 } },
2164 [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = { .uint32 = { 1, 1, 1, 1 } },
2165 };
2166
2167 device->border_colors =
2168 anv_state_pool_emit_data(&device->dynamic_state_pool,
2169 sizeof(border_colors), 64, border_colors);
2170 }
2171 }
2172
2173 static VkResult
2174 anv_device_init_trivial_batch(struct anv_device *device)
2175 {
2176 VkResult result = anv_device_alloc_bo(device, 4096,
2177 ANV_BO_ALLOC_MAPPED,
2178 0 /* explicit_address */,
2179 &device->trivial_batch_bo);
2180 if (result != VK_SUCCESS)
2181 return result;
2182
2183 struct anv_batch batch = {
2184 .start = device->trivial_batch_bo->map,
2185 .next = device->trivial_batch_bo->map,
2186 .end = device->trivial_batch_bo->map + 4096,
2187 };
2188
2189 anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
2190 anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
2191
2192 if (!device->info.has_llc)
2193 gen_clflush_range(batch.start, batch.next - batch.start);
2194
2195 return VK_SUCCESS;
2196 }
2197
2198 VkResult anv_EnumerateDeviceExtensionProperties(
2199 VkPhysicalDevice physicalDevice,
2200 const char* pLayerName,
2201 uint32_t* pPropertyCount,
2202 VkExtensionProperties* pProperties)
2203 {
2204 ANV_FROM_HANDLE(anv_physical_device, device, physicalDevice);
2205 VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
2206
2207 for (int i = 0; i < ANV_DEVICE_EXTENSION_COUNT; i++) {
2208 if (device->supported_extensions.extensions[i]) {
2209 vk_outarray_append(&out, prop) {
2210 *prop = anv_device_extensions[i];
2211 }
2212 }
2213 }
2214
2215 return vk_outarray_status(&out);
2216 }
2217
2218 static void
2219 anv_device_init_dispatch(struct anv_device *device)
2220 {
2221 const struct anv_device_dispatch_table *genX_table;
2222 switch (device->info.gen) {
2223 case 12:
2224 genX_table = &gen12_device_dispatch_table;
2225 break;
2226 case 11:
2227 genX_table = &gen11_device_dispatch_table;
2228 break;
2229 case 10:
2230 genX_table = &gen10_device_dispatch_table;
2231 break;
2232 case 9:
2233 genX_table = &gen9_device_dispatch_table;
2234 break;
2235 case 8:
2236 genX_table = &gen8_device_dispatch_table;
2237 break;
2238 case 7:
2239 if (device->info.is_haswell)
2240 genX_table = &gen75_device_dispatch_table;
2241 else
2242 genX_table = &gen7_device_dispatch_table;
2243 break;
2244 default:
2245 unreachable("unsupported gen\n");
2246 }
2247
2248 for (unsigned i = 0; i < ARRAY_SIZE(device->dispatch.entrypoints); i++) {
2249 /* Vulkan requires that entrypoints for extensions which have not been
2250 * enabled must not be advertised.
2251 */
2252 if (!anv_device_entrypoint_is_enabled(i, device->instance->app_info.api_version,
2253 &device->instance->enabled_extensions,
2254 &device->enabled_extensions)) {
2255 device->dispatch.entrypoints[i] = NULL;
2256 } else if (genX_table->entrypoints[i]) {
2257 device->dispatch.entrypoints[i] = genX_table->entrypoints[i];
2258 } else {
2259 device->dispatch.entrypoints[i] =
2260 anv_device_dispatch_table.entrypoints[i];
2261 }
2262 }
2263 }
2264
2265 static int
2266 vk_priority_to_gen(int priority)
2267 {
2268 switch (priority) {
2269 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT:
2270 return GEN_CONTEXT_LOW_PRIORITY;
2271 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT:
2272 return GEN_CONTEXT_MEDIUM_PRIORITY;
2273 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT:
2274 return GEN_CONTEXT_HIGH_PRIORITY;
2275 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT:
2276 return GEN_CONTEXT_REALTIME_PRIORITY;
2277 default:
2278 unreachable("Invalid priority");
2279 }
2280 }
2281
2282 static VkResult
2283 anv_device_init_hiz_clear_value_bo(struct anv_device *device)
2284 {
2285 VkResult result = anv_device_alloc_bo(device, 4096,
2286 ANV_BO_ALLOC_MAPPED,
2287 0 /* explicit_address */,
2288 &device->hiz_clear_bo);
2289 if (result != VK_SUCCESS)
2290 return result;
2291
2292 union isl_color_value hiz_clear = { .u32 = { 0, } };
2293 hiz_clear.f32[0] = ANV_HZ_FC_VAL;
2294
2295 memcpy(device->hiz_clear_bo->map, hiz_clear.u32, sizeof(hiz_clear.u32));
2296
2297 if (!device->info.has_llc)
2298 gen_clflush_range(device->hiz_clear_bo->map, sizeof(hiz_clear.u32));
2299
2300 return VK_SUCCESS;
2301 }
2302
2303 static bool
2304 get_bo_from_pool(struct gen_batch_decode_bo *ret,
2305 struct anv_block_pool *pool,
2306 uint64_t address)
2307 {
2308 anv_block_pool_foreach_bo(bo, pool) {
2309 uint64_t bo_address = gen_48b_address(bo->offset);
2310 if (address >= bo_address && address < (bo_address + bo->size)) {
2311 *ret = (struct gen_batch_decode_bo) {
2312 .addr = bo_address,
2313 .size = bo->size,
2314 .map = bo->map,
2315 };
2316 return true;
2317 }
2318 }
2319 return false;
2320 }
2321
2322 /* Finding a buffer for batch decoding */
2323 static struct gen_batch_decode_bo
2324 decode_get_bo(void *v_batch, bool ppgtt, uint64_t address)
2325 {
2326 struct anv_device *device = v_batch;
2327 struct gen_batch_decode_bo ret_bo = {};
2328
2329 assert(ppgtt);
2330
2331 if (get_bo_from_pool(&ret_bo, &device->dynamic_state_pool.block_pool, address))
2332 return ret_bo;
2333 if (get_bo_from_pool(&ret_bo, &device->instruction_state_pool.block_pool, address))
2334 return ret_bo;
2335 if (get_bo_from_pool(&ret_bo, &device->binding_table_pool.block_pool, address))
2336 return ret_bo;
2337 if (get_bo_from_pool(&ret_bo, &device->surface_state_pool.block_pool, address))
2338 return ret_bo;
2339
2340 if (!device->cmd_buffer_being_decoded)
2341 return (struct gen_batch_decode_bo) { };
2342
2343 struct anv_batch_bo **bo;
2344
2345 u_vector_foreach(bo, &device->cmd_buffer_being_decoded->seen_bbos) {
2346 /* The decoder zeroes out the top 16 bits, so we need to as well */
2347 uint64_t bo_address = (*bo)->bo->offset & (~0ull >> 16);
2348
2349 if (address >= bo_address && address < bo_address + (*bo)->bo->size) {
2350 return (struct gen_batch_decode_bo) {
2351 .addr = bo_address,
2352 .size = (*bo)->bo->size,
2353 .map = (*bo)->bo->map,
2354 };
2355 }
2356 }
2357
2358 return (struct gen_batch_decode_bo) { };
2359 }
2360
2361 struct gen_aux_map_buffer {
2362 struct gen_buffer base;
2363 struct anv_state state;
2364 };
2365
2366 static struct gen_buffer *
2367 gen_aux_map_buffer_alloc(void *driver_ctx, uint32_t size)
2368 {
2369 struct gen_aux_map_buffer *buf = malloc(sizeof(struct gen_aux_map_buffer));
2370 if (!buf)
2371 return NULL;
2372
2373 struct anv_device *device = (struct anv_device*)driver_ctx;
2374 assert(device->instance->physicalDevice.supports_48bit_addresses &&
2375 device->instance->physicalDevice.use_softpin);
2376
2377 struct anv_state_pool *pool = &device->dynamic_state_pool;
2378 buf->state = anv_state_pool_alloc(pool, size, size);
2379
2380 buf->base.gpu = pool->block_pool.bo->offset + buf->state.offset;
2381 buf->base.gpu_end = buf->base.gpu + buf->state.alloc_size;
2382 buf->base.map = buf->state.map;
2383 buf->base.driver_bo = &buf->state;
2384 return &buf->base;
2385 }
2386
2387 static void
2388 gen_aux_map_buffer_free(void *driver_ctx, struct gen_buffer *buffer)
2389 {
2390 struct gen_aux_map_buffer *buf = (struct gen_aux_map_buffer*)buffer;
2391 struct anv_device *device = (struct anv_device*)driver_ctx;
2392 struct anv_state_pool *pool = &device->dynamic_state_pool;
2393 anv_state_pool_free(pool, buf->state);
2394 free(buf);
2395 }
2396
2397 static struct gen_mapped_pinned_buffer_alloc aux_map_allocator = {
2398 .alloc = gen_aux_map_buffer_alloc,
2399 .free = gen_aux_map_buffer_free,
2400 };
2401
2402 VkResult anv_CreateDevice(
2403 VkPhysicalDevice physicalDevice,
2404 const VkDeviceCreateInfo* pCreateInfo,
2405 const VkAllocationCallbacks* pAllocator,
2406 VkDevice* pDevice)
2407 {
2408 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
2409 VkResult result;
2410 struct anv_device *device;
2411
2412 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO);
2413
2414 struct anv_device_extension_table enabled_extensions = { };
2415 for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
2416 int idx;
2417 for (idx = 0; idx < ANV_DEVICE_EXTENSION_COUNT; idx++) {
2418 if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
2419 anv_device_extensions[idx].extensionName) == 0)
2420 break;
2421 }
2422
2423 if (idx >= ANV_DEVICE_EXTENSION_COUNT)
2424 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
2425
2426 if (!physical_device->supported_extensions.extensions[idx])
2427 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
2428
2429 enabled_extensions.extensions[idx] = true;
2430 }
2431
2432 /* Check enabled features */
2433 if (pCreateInfo->pEnabledFeatures) {
2434 VkPhysicalDeviceFeatures supported_features;
2435 anv_GetPhysicalDeviceFeatures(physicalDevice, &supported_features);
2436 VkBool32 *supported_feature = (VkBool32 *)&supported_features;
2437 VkBool32 *enabled_feature = (VkBool32 *)pCreateInfo->pEnabledFeatures;
2438 unsigned num_features = sizeof(VkPhysicalDeviceFeatures) / sizeof(VkBool32);
2439 for (uint32_t i = 0; i < num_features; i++) {
2440 if (enabled_feature[i] && !supported_feature[i])
2441 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
2442 }
2443 }
2444
2445 /* Check requested queues and fail if we are requested to create any
2446 * queues with flags we don't support.
2447 */
2448 assert(pCreateInfo->queueCreateInfoCount > 0);
2449 for (uint32_t i = 0; i < pCreateInfo->queueCreateInfoCount; i++) {
2450 if (pCreateInfo->pQueueCreateInfos[i].flags != 0)
2451 return vk_error(VK_ERROR_INITIALIZATION_FAILED);
2452 }
2453
2454 /* Check if client specified queue priority. */
2455 const VkDeviceQueueGlobalPriorityCreateInfoEXT *queue_priority =
2456 vk_find_struct_const(pCreateInfo->pQueueCreateInfos[0].pNext,
2457 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT);
2458
2459 VkQueueGlobalPriorityEXT priority =
2460 queue_priority ? queue_priority->globalPriority :
2461 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT;
2462
2463 device = vk_alloc2(&physical_device->instance->alloc, pAllocator,
2464 sizeof(*device), 8,
2465 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
2466 if (!device)
2467 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2468
2469 if (INTEL_DEBUG & DEBUG_BATCH) {
2470 const unsigned decode_flags =
2471 GEN_BATCH_DECODE_FULL |
2472 ((INTEL_DEBUG & DEBUG_COLOR) ? GEN_BATCH_DECODE_IN_COLOR : 0) |
2473 GEN_BATCH_DECODE_OFFSETS |
2474 GEN_BATCH_DECODE_FLOATS;
2475
2476 gen_batch_decode_ctx_init(&device->decoder_ctx,
2477 &physical_device->info,
2478 stderr, decode_flags, NULL,
2479 decode_get_bo, NULL, device);
2480 }
2481
2482 device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
2483 device->instance = physical_device->instance;
2484 device->chipset_id = physical_device->chipset_id;
2485 device->no_hw = physical_device->no_hw;
2486 device->_lost = false;
2487
2488 if (pAllocator)
2489 device->alloc = *pAllocator;
2490 else
2491 device->alloc = physical_device->instance->alloc;
2492
2493 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
2494 device->fd = open(physical_device->path, O_RDWR | O_CLOEXEC);
2495 if (device->fd == -1) {
2496 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2497 goto fail_device;
2498 }
2499
2500 device->context_id = anv_gem_create_context(device);
2501 if (device->context_id == -1) {
2502 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2503 goto fail_fd;
2504 }
2505
2506 result = anv_queue_init(device, &device->queue);
2507 if (result != VK_SUCCESS)
2508 goto fail_context_id;
2509
2510 if (physical_device->use_softpin) {
2511 if (pthread_mutex_init(&device->vma_mutex, NULL) != 0) {
2512 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2513 goto fail_queue;
2514 }
2515
2516 /* keep the page with address zero out of the allocator */
2517 util_vma_heap_init(&device->vma_lo,
2518 LOW_HEAP_MIN_ADDRESS, LOW_HEAP_SIZE);
2519
2520 util_vma_heap_init(&device->vma_cva, CLIENT_VISIBLE_HEAP_MIN_ADDRESS,
2521 CLIENT_VISIBLE_HEAP_SIZE);
2522
2523 /* Leave the last 4GiB out of the high vma range, so that no state
2524 * base address + size can overflow 48 bits. For more information see
2525 * the comment about Wa32bitGeneralStateOffset in anv_allocator.c
2526 */
2527 util_vma_heap_init(&device->vma_hi, HIGH_HEAP_MIN_ADDRESS,
2528 physical_device->gtt_size - (1ull << 32) -
2529 HIGH_HEAP_MIN_ADDRESS);
2530 }
2531
2532 list_inithead(&device->memory_objects);
2533
2534 /* As per spec, the driver implementation may deny requests to acquire
2535 * a priority above the default priority (MEDIUM) if the caller does not
2536 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
2537 * is returned.
2538 */
2539 if (physical_device->has_context_priority) {
2540 int err = anv_gem_set_context_param(device->fd, device->context_id,
2541 I915_CONTEXT_PARAM_PRIORITY,
2542 vk_priority_to_gen(priority));
2543 if (err != 0 && priority > VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT) {
2544 result = vk_error(VK_ERROR_NOT_PERMITTED_EXT);
2545 goto fail_vmas;
2546 }
2547 }
2548
2549 device->info = physical_device->info;
2550 device->isl_dev = physical_device->isl_dev;
2551
2552 /* On Broadwell and later, we can use batch chaining to more efficiently
2553 * implement growing command buffers. Prior to Haswell, the kernel
2554 * command parser gets in the way and we have to fall back to growing
2555 * the batch.
2556 */
2557 device->can_chain_batches = device->info.gen >= 8;
2558
2559 device->robust_buffer_access = pCreateInfo->pEnabledFeatures &&
2560 pCreateInfo->pEnabledFeatures->robustBufferAccess;
2561 device->enabled_extensions = enabled_extensions;
2562
2563 anv_device_init_dispatch(device);
2564
2565 if (pthread_mutex_init(&device->mutex, NULL) != 0) {
2566 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2567 goto fail_queue;
2568 }
2569
2570 pthread_condattr_t condattr;
2571 if (pthread_condattr_init(&condattr) != 0) {
2572 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2573 goto fail_mutex;
2574 }
2575 if (pthread_condattr_setclock(&condattr, CLOCK_MONOTONIC) != 0) {
2576 pthread_condattr_destroy(&condattr);
2577 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2578 goto fail_mutex;
2579 }
2580 if (pthread_cond_init(&device->queue_submit, &condattr) != 0) {
2581 pthread_condattr_destroy(&condattr);
2582 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2583 goto fail_mutex;
2584 }
2585 pthread_condattr_destroy(&condattr);
2586
2587 result = anv_bo_cache_init(&device->bo_cache);
2588 if (result != VK_SUCCESS)
2589 goto fail_queue_cond;
2590
2591 anv_bo_pool_init(&device->batch_bo_pool, device);
2592
2593 result = anv_state_pool_init(&device->dynamic_state_pool, device,
2594 DYNAMIC_STATE_POOL_MIN_ADDRESS, 16384);
2595 if (result != VK_SUCCESS)
2596 goto fail_batch_bo_pool;
2597
2598 result = anv_state_pool_init(&device->instruction_state_pool, device,
2599 INSTRUCTION_STATE_POOL_MIN_ADDRESS, 16384);
2600 if (result != VK_SUCCESS)
2601 goto fail_dynamic_state_pool;
2602
2603 result = anv_state_pool_init(&device->surface_state_pool, device,
2604 SURFACE_STATE_POOL_MIN_ADDRESS, 4096);
2605 if (result != VK_SUCCESS)
2606 goto fail_instruction_state_pool;
2607
2608 if (physical_device->use_softpin) {
2609 result = anv_state_pool_init(&device->binding_table_pool, device,
2610 BINDING_TABLE_POOL_MIN_ADDRESS, 4096);
2611 if (result != VK_SUCCESS)
2612 goto fail_surface_state_pool;
2613 }
2614
2615 if (device->info.gen >= 12) {
2616 device->aux_map_ctx = gen_aux_map_init(device, &aux_map_allocator,
2617 &physical_device->info);
2618 if (!device->aux_map_ctx)
2619 goto fail_binding_table_pool;
2620 }
2621
2622 result = anv_device_alloc_bo(device, 4096, 0 /* flags */,
2623 0 /* explicit_address */,
2624 &device->workaround_bo);
2625 if (result != VK_SUCCESS)
2626 goto fail_surface_aux_map_pool;
2627
2628 result = anv_device_init_trivial_batch(device);
2629 if (result != VK_SUCCESS)
2630 goto fail_workaround_bo;
2631
2632 if (device->info.gen >= 10) {
2633 result = anv_device_init_hiz_clear_value_bo(device);
2634 if (result != VK_SUCCESS)
2635 goto fail_trivial_batch_bo;
2636 }
2637
2638 anv_scratch_pool_init(device, &device->scratch_pool);
2639
2640 switch (device->info.gen) {
2641 case 7:
2642 if (!device->info.is_haswell)
2643 result = gen7_init_device_state(device);
2644 else
2645 result = gen75_init_device_state(device);
2646 break;
2647 case 8:
2648 result = gen8_init_device_state(device);
2649 break;
2650 case 9:
2651 result = gen9_init_device_state(device);
2652 break;
2653 case 10:
2654 result = gen10_init_device_state(device);
2655 break;
2656 case 11:
2657 result = gen11_init_device_state(device);
2658 break;
2659 case 12:
2660 result = gen12_init_device_state(device);
2661 break;
2662 default:
2663 /* Shouldn't get here as we don't create physical devices for any other
2664 * gens. */
2665 unreachable("unhandled gen");
2666 }
2667 if (result != VK_SUCCESS)
2668 goto fail_workaround_bo;
2669
2670 anv_pipeline_cache_init(&device->default_pipeline_cache, device, true);
2671
2672 anv_device_init_blorp(device);
2673
2674 anv_device_init_border_colors(device);
2675
2676 anv_device_perf_init(device);
2677
2678 *pDevice = anv_device_to_handle(device);
2679
2680 return VK_SUCCESS;
2681
2682 fail_workaround_bo:
2683 anv_scratch_pool_finish(device, &device->scratch_pool);
2684 if (device->info.gen >= 10)
2685 anv_device_release_bo(device, device->hiz_clear_bo);
2686 anv_device_release_bo(device, device->workaround_bo);
2687 fail_trivial_batch_bo:
2688 anv_device_release_bo(device, device->trivial_batch_bo);
2689 fail_surface_aux_map_pool:
2690 if (device->info.gen >= 12) {
2691 gen_aux_map_finish(device->aux_map_ctx);
2692 device->aux_map_ctx = NULL;
2693 }
2694 fail_binding_table_pool:
2695 if (physical_device->use_softpin)
2696 anv_state_pool_finish(&device->binding_table_pool);
2697 fail_surface_state_pool:
2698 anv_state_pool_finish(&device->surface_state_pool);
2699 fail_instruction_state_pool:
2700 anv_state_pool_finish(&device->instruction_state_pool);
2701 fail_dynamic_state_pool:
2702 anv_state_pool_finish(&device->dynamic_state_pool);
2703 fail_batch_bo_pool:
2704 anv_bo_pool_finish(&device->batch_bo_pool);
2705 anv_bo_cache_finish(&device->bo_cache);
2706 fail_queue_cond:
2707 pthread_cond_destroy(&device->queue_submit);
2708 fail_mutex:
2709 pthread_mutex_destroy(&device->mutex);
2710 fail_vmas:
2711 if (physical_device->use_softpin) {
2712 util_vma_heap_finish(&device->vma_hi);
2713 util_vma_heap_finish(&device->vma_cva);
2714 util_vma_heap_finish(&device->vma_lo);
2715 }
2716 fail_queue:
2717 anv_queue_finish(&device->queue);
2718 fail_context_id:
2719 anv_gem_destroy_context(device, device->context_id);
2720 fail_fd:
2721 close(device->fd);
2722 fail_device:
2723 vk_free(&device->alloc, device);
2724
2725 return result;
2726 }
2727
2728 void anv_DestroyDevice(
2729 VkDevice _device,
2730 const VkAllocationCallbacks* pAllocator)
2731 {
2732 ANV_FROM_HANDLE(anv_device, device, _device);
2733 struct anv_physical_device *physical_device;
2734
2735 if (!device)
2736 return;
2737
2738 physical_device = &device->instance->physicalDevice;
2739
2740 anv_device_finish_blorp(device);
2741
2742 anv_pipeline_cache_finish(&device->default_pipeline_cache);
2743
2744 anv_queue_finish(&device->queue);
2745
2746 #ifdef HAVE_VALGRIND
2747 /* We only need to free these to prevent valgrind errors. The backing
2748 * BO will go away in a couple of lines so we don't actually leak.
2749 */
2750 anv_state_pool_free(&device->dynamic_state_pool, device->border_colors);
2751 anv_state_pool_free(&device->dynamic_state_pool, device->slice_hash);
2752 #endif
2753
2754 anv_scratch_pool_finish(device, &device->scratch_pool);
2755
2756 anv_device_release_bo(device, device->workaround_bo);
2757 anv_device_release_bo(device, device->trivial_batch_bo);
2758 if (device->info.gen >= 10)
2759 anv_device_release_bo(device, device->hiz_clear_bo);
2760
2761 if (device->info.gen >= 12) {
2762 gen_aux_map_finish(device->aux_map_ctx);
2763 device->aux_map_ctx = NULL;
2764 }
2765
2766 if (physical_device->use_softpin)
2767 anv_state_pool_finish(&device->binding_table_pool);
2768 anv_state_pool_finish(&device->surface_state_pool);
2769 anv_state_pool_finish(&device->instruction_state_pool);
2770 anv_state_pool_finish(&device->dynamic_state_pool);
2771
2772 anv_bo_pool_finish(&device->batch_bo_pool);
2773
2774 anv_bo_cache_finish(&device->bo_cache);
2775
2776 if (physical_device->use_softpin) {
2777 util_vma_heap_finish(&device->vma_hi);
2778 util_vma_heap_finish(&device->vma_cva);
2779 util_vma_heap_finish(&device->vma_lo);
2780 }
2781
2782 pthread_cond_destroy(&device->queue_submit);
2783 pthread_mutex_destroy(&device->mutex);
2784
2785 anv_gem_destroy_context(device, device->context_id);
2786
2787 if (INTEL_DEBUG & DEBUG_BATCH)
2788 gen_batch_decode_ctx_finish(&device->decoder_ctx);
2789
2790 close(device->fd);
2791
2792 vk_free(&device->alloc, device);
2793 }
2794
2795 VkResult anv_EnumerateInstanceLayerProperties(
2796 uint32_t* pPropertyCount,
2797 VkLayerProperties* pProperties)
2798 {
2799 if (pProperties == NULL) {
2800 *pPropertyCount = 0;
2801 return VK_SUCCESS;
2802 }
2803
2804 /* None supported at this time */
2805 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
2806 }
2807
2808 VkResult anv_EnumerateDeviceLayerProperties(
2809 VkPhysicalDevice physicalDevice,
2810 uint32_t* pPropertyCount,
2811 VkLayerProperties* pProperties)
2812 {
2813 if (pProperties == NULL) {
2814 *pPropertyCount = 0;
2815 return VK_SUCCESS;
2816 }
2817
2818 /* None supported at this time */
2819 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
2820 }
2821
2822 void anv_GetDeviceQueue(
2823 VkDevice _device,
2824 uint32_t queueNodeIndex,
2825 uint32_t queueIndex,
2826 VkQueue* pQueue)
2827 {
2828 const VkDeviceQueueInfo2 info = {
2829 .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_INFO_2,
2830 .pNext = NULL,
2831 .flags = 0,
2832 .queueFamilyIndex = queueNodeIndex,
2833 .queueIndex = queueIndex,
2834 };
2835
2836 anv_GetDeviceQueue2(_device, &info, pQueue);
2837 }
2838
2839 void anv_GetDeviceQueue2(
2840 VkDevice _device,
2841 const VkDeviceQueueInfo2* pQueueInfo,
2842 VkQueue* pQueue)
2843 {
2844 ANV_FROM_HANDLE(anv_device, device, _device);
2845
2846 assert(pQueueInfo->queueIndex == 0);
2847
2848 if (pQueueInfo->flags == device->queue.flags)
2849 *pQueue = anv_queue_to_handle(&device->queue);
2850 else
2851 *pQueue = NULL;
2852 }
2853
2854 VkResult
2855 _anv_device_set_lost(struct anv_device *device,
2856 const char *file, int line,
2857 const char *msg, ...)
2858 {
2859 VkResult err;
2860 va_list ap;
2861
2862 p_atomic_inc(&device->_lost);
2863
2864 va_start(ap, msg);
2865 err = __vk_errorv(device->instance, device,
2866 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT,
2867 VK_ERROR_DEVICE_LOST, file, line, msg, ap);
2868 va_end(ap);
2869
2870 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
2871 abort();
2872
2873 return err;
2874 }
2875
2876 VkResult
2877 _anv_queue_set_lost(struct anv_queue *queue,
2878 const char *file, int line,
2879 const char *msg, ...)
2880 {
2881 VkResult err;
2882 va_list ap;
2883
2884 p_atomic_inc(&queue->device->_lost);
2885
2886 va_start(ap, msg);
2887 err = __vk_errorv(queue->device->instance, queue->device,
2888 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT,
2889 VK_ERROR_DEVICE_LOST, file, line, msg, ap);
2890 va_end(ap);
2891
2892 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
2893 abort();
2894
2895 return err;
2896 }
2897
2898 VkResult
2899 anv_device_query_status(struct anv_device *device)
2900 {
2901 /* This isn't likely as most of the callers of this function already check
2902 * for it. However, it doesn't hurt to check and it potentially lets us
2903 * avoid an ioctl.
2904 */
2905 if (anv_device_is_lost(device))
2906 return VK_ERROR_DEVICE_LOST;
2907
2908 uint32_t active, pending;
2909 int ret = anv_gem_gpu_get_reset_stats(device, &active, &pending);
2910 if (ret == -1) {
2911 /* We don't know the real error. */
2912 return anv_device_set_lost(device, "get_reset_stats failed: %m");
2913 }
2914
2915 if (active) {
2916 return anv_device_set_lost(device, "GPU hung on one of our command buffers");
2917 } else if (pending) {
2918 return anv_device_set_lost(device, "GPU hung with commands in-flight");
2919 }
2920
2921 return VK_SUCCESS;
2922 }
2923
2924 VkResult
2925 anv_device_bo_busy(struct anv_device *device, struct anv_bo *bo)
2926 {
2927 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
2928 * Other usages of the BO (such as on different hardware) will not be
2929 * flagged as "busy" by this ioctl. Use with care.
2930 */
2931 int ret = anv_gem_busy(device, bo->gem_handle);
2932 if (ret == 1) {
2933 return VK_NOT_READY;
2934 } else if (ret == -1) {
2935 /* We don't know the real error. */
2936 return anv_device_set_lost(device, "gem wait failed: %m");
2937 }
2938
2939 /* Query for device status after the busy call. If the BO we're checking
2940 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
2941 * client because it clearly doesn't have valid data. Yes, this most
2942 * likely means an ioctl, but we just did an ioctl to query the busy status
2943 * so it's no great loss.
2944 */
2945 return anv_device_query_status(device);
2946 }
2947
2948 VkResult
2949 anv_device_wait(struct anv_device *device, struct anv_bo *bo,
2950 int64_t timeout)
2951 {
2952 int ret = anv_gem_wait(device, bo->gem_handle, &timeout);
2953 if (ret == -1 && errno == ETIME) {
2954 return VK_TIMEOUT;
2955 } else if (ret == -1) {
2956 /* We don't know the real error. */
2957 return anv_device_set_lost(device, "gem wait failed: %m");
2958 }
2959
2960 /* Query for device status after the wait. If the BO we're waiting on got
2961 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
2962 * because it clearly doesn't have valid data. Yes, this most likely means
2963 * an ioctl, but we just did an ioctl to wait so it's no great loss.
2964 */
2965 return anv_device_query_status(device);
2966 }
2967
2968 VkResult anv_DeviceWaitIdle(
2969 VkDevice _device)
2970 {
2971 ANV_FROM_HANDLE(anv_device, device, _device);
2972
2973 if (anv_device_is_lost(device))
2974 return VK_ERROR_DEVICE_LOST;
2975
2976 return anv_queue_submit_simple_batch(&device->queue, NULL);
2977 }
2978
2979 bool
2980 anv_vma_alloc(struct anv_device *device, struct anv_bo *bo,
2981 uint64_t client_address)
2982 {
2983 const struct anv_physical_device *pdevice = &device->instance->physicalDevice;
2984 const struct gen_device_info *devinfo = &pdevice->info;
2985 /* Gen12 CCS surface addresses need to be 64K aligned. We have no way of
2986 * telling what this allocation is for so pick the largest alignment.
2987 */
2988 const uint32_t vma_alignment =
2989 devinfo->gen >= 12 ? (64 * 1024) : (4 * 1024);
2990
2991 if (!(bo->flags & EXEC_OBJECT_PINNED)) {
2992 assert(!(bo->has_client_visible_address));
2993 return true;
2994 }
2995
2996 pthread_mutex_lock(&device->vma_mutex);
2997
2998 bo->offset = 0;
2999
3000 if (bo->has_client_visible_address) {
3001 assert(bo->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS);
3002 if (client_address) {
3003 if (util_vma_heap_alloc_addr(&device->vma_cva,
3004 client_address, bo->size)) {
3005 bo->offset = gen_canonical_address(client_address);
3006 }
3007 } else {
3008 uint64_t addr =
3009 util_vma_heap_alloc(&device->vma_cva, bo->size, vma_alignment);
3010 if (addr) {
3011 bo->offset = gen_canonical_address(addr);
3012 assert(addr == gen_48b_address(bo->offset));
3013 }
3014 }
3015 /* We don't want to fall back to other heaps */
3016 goto done;
3017 }
3018
3019 assert(client_address == 0);
3020
3021 if (bo->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) {
3022 uint64_t addr =
3023 util_vma_heap_alloc(&device->vma_hi, bo->size, vma_alignment);
3024 if (addr) {
3025 bo->offset = gen_canonical_address(addr);
3026 assert(addr == gen_48b_address(bo->offset));
3027 }
3028 }
3029
3030 if (bo->offset == 0) {
3031 uint64_t addr =
3032 util_vma_heap_alloc(&device->vma_lo, bo->size, vma_alignment);
3033 if (addr) {
3034 bo->offset = gen_canonical_address(addr);
3035 assert(addr == gen_48b_address(bo->offset));
3036 }
3037 }
3038
3039 done:
3040 pthread_mutex_unlock(&device->vma_mutex);
3041
3042 return bo->offset != 0;
3043 }
3044
3045 void
3046 anv_vma_free(struct anv_device *device, struct anv_bo *bo)
3047 {
3048 if (!(bo->flags & EXEC_OBJECT_PINNED))
3049 return;
3050
3051 const uint64_t addr_48b = gen_48b_address(bo->offset);
3052
3053 pthread_mutex_lock(&device->vma_mutex);
3054
3055 if (addr_48b >= LOW_HEAP_MIN_ADDRESS &&
3056 addr_48b <= LOW_HEAP_MAX_ADDRESS) {
3057 util_vma_heap_free(&device->vma_lo, addr_48b, bo->size);
3058 } else if (addr_48b >= CLIENT_VISIBLE_HEAP_MIN_ADDRESS &&
3059 addr_48b <= CLIENT_VISIBLE_HEAP_MAX_ADDRESS) {
3060 util_vma_heap_free(&device->vma_cva, addr_48b, bo->size);
3061 } else {
3062 assert(addr_48b >= HIGH_HEAP_MIN_ADDRESS);
3063 util_vma_heap_free(&device->vma_hi, addr_48b, bo->size);
3064 }
3065
3066 pthread_mutex_unlock(&device->vma_mutex);
3067
3068 bo->offset = 0;
3069 }
3070
3071 VkResult anv_AllocateMemory(
3072 VkDevice _device,
3073 const VkMemoryAllocateInfo* pAllocateInfo,
3074 const VkAllocationCallbacks* pAllocator,
3075 VkDeviceMemory* pMem)
3076 {
3077 ANV_FROM_HANDLE(anv_device, device, _device);
3078 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3079 struct anv_device_memory *mem;
3080 VkResult result = VK_SUCCESS;
3081
3082 assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
3083
3084 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
3085 assert(pAllocateInfo->allocationSize > 0);
3086
3087 VkDeviceSize aligned_alloc_size =
3088 align_u64(pAllocateInfo->allocationSize, 4096);
3089
3090 if (aligned_alloc_size > MAX_MEMORY_ALLOCATION_SIZE)
3091 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
3092
3093 assert(pAllocateInfo->memoryTypeIndex < pdevice->memory.type_count);
3094 struct anv_memory_type *mem_type =
3095 &pdevice->memory.types[pAllocateInfo->memoryTypeIndex];
3096 assert(mem_type->heapIndex < pdevice->memory.heap_count);
3097 struct anv_memory_heap *mem_heap =
3098 &pdevice->memory.heaps[mem_type->heapIndex];
3099
3100 uint64_t mem_heap_used = p_atomic_read(&mem_heap->used);
3101 if (mem_heap_used + aligned_alloc_size > mem_heap->size)
3102 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
3103
3104 mem = vk_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
3105 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
3106 if (mem == NULL)
3107 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
3108
3109 assert(pAllocateInfo->memoryTypeIndex < pdevice->memory.type_count);
3110 mem->type = mem_type;
3111 mem->map = NULL;
3112 mem->map_size = 0;
3113 mem->ahw = NULL;
3114 mem->host_ptr = NULL;
3115
3116 enum anv_bo_alloc_flags alloc_flags = 0;
3117
3118 const VkExportMemoryAllocateInfo *export_info = NULL;
3119 const VkImportAndroidHardwareBufferInfoANDROID *ahw_import_info = NULL;
3120 const VkImportMemoryFdInfoKHR *fd_info = NULL;
3121 const VkImportMemoryHostPointerInfoEXT *host_ptr_info = NULL;
3122 const VkMemoryDedicatedAllocateInfo *dedicated_info = NULL;
3123 VkMemoryAllocateFlags vk_flags = 0;
3124 uint64_t client_address = 0;
3125
3126 vk_foreach_struct_const(ext, pAllocateInfo->pNext) {
3127 switch (ext->sType) {
3128 case VK_STRUCTURE_TYPE_EXPORT_MEMORY_ALLOCATE_INFO:
3129 export_info = (void *)ext;
3130 break;
3131
3132 case VK_STRUCTURE_TYPE_IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID:
3133 ahw_import_info = (void *)ext;
3134 break;
3135
3136 case VK_STRUCTURE_TYPE_IMPORT_MEMORY_FD_INFO_KHR:
3137 fd_info = (void *)ext;
3138 break;
3139
3140 case VK_STRUCTURE_TYPE_IMPORT_MEMORY_HOST_POINTER_INFO_EXT:
3141 host_ptr_info = (void *)ext;
3142 break;
3143
3144 case VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_FLAGS_INFO: {
3145 const VkMemoryAllocateFlagsInfo *flags_info = (void *)ext;
3146 vk_flags = flags_info->flags;
3147 break;
3148 }
3149
3150 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_ALLOCATE_INFO:
3151 dedicated_info = (void *)ext;
3152 break;
3153
3154 case VK_STRUCTURE_TYPE_MEMORY_OPAQUE_CAPTURE_ADDRESS_ALLOCATE_INFO_KHR: {
3155 const VkMemoryOpaqueCaptureAddressAllocateInfoKHR *addr_info =
3156 (const VkMemoryOpaqueCaptureAddressAllocateInfoKHR *)ext;
3157 client_address = addr_info->opaqueCaptureAddress;
3158 break;
3159 }
3160
3161 default:
3162 anv_debug_ignored_stype(ext->sType);
3163 break;
3164 }
3165 }
3166
3167 if (vk_flags & VK_MEMORY_ALLOCATE_DEVICE_ADDRESS_BIT_KHR)
3168 alloc_flags |= ANV_BO_ALLOC_CLIENT_VISIBLE_ADDRESS;
3169
3170 /* Check if we need to support Android HW buffer export. If so,
3171 * create AHardwareBuffer and import memory from it.
3172 */
3173 bool android_export = false;
3174 if (export_info && export_info->handleTypes &
3175 VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID)
3176 android_export = true;
3177
3178 if (ahw_import_info) {
3179 result = anv_import_ahw_memory(_device, mem, ahw_import_info);
3180 if (result != VK_SUCCESS)
3181 goto fail;
3182
3183 goto success;
3184 } else if (android_export) {
3185 result = anv_create_ahw_memory(_device, mem, pAllocateInfo);
3186 if (result != VK_SUCCESS)
3187 goto fail;
3188
3189 const struct VkImportAndroidHardwareBufferInfoANDROID import_info = {
3190 .buffer = mem->ahw,
3191 };
3192 result = anv_import_ahw_memory(_device, mem, &import_info);
3193 if (result != VK_SUCCESS)
3194 goto fail;
3195
3196 goto success;
3197 }
3198
3199 /* The Vulkan spec permits handleType to be 0, in which case the struct is
3200 * ignored.
3201 */
3202 if (fd_info && fd_info->handleType) {
3203 /* At the moment, we support only the below handle types. */
3204 assert(fd_info->handleType ==
3205 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
3206 fd_info->handleType ==
3207 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
3208
3209 result = anv_device_import_bo(device, fd_info->fd, alloc_flags,
3210 client_address, &mem->bo);
3211 if (result != VK_SUCCESS)
3212 goto fail;
3213
3214 VkDeviceSize aligned_alloc_size =
3215 align_u64(pAllocateInfo->allocationSize, 4096);
3216
3217 /* For security purposes, we reject importing the bo if it's smaller
3218 * than the requested allocation size. This prevents a malicious client
3219 * from passing a buffer to a trusted client, lying about the size, and
3220 * telling the trusted client to try and texture from an image that goes
3221 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
3222 * in the trusted client. The trusted client can protect itself against
3223 * this sort of attack but only if it can trust the buffer size.
3224 */
3225 if (mem->bo->size < aligned_alloc_size) {
3226 result = vk_errorf(device->instance, device,
3227 VK_ERROR_INVALID_EXTERNAL_HANDLE,
3228 "aligned allocationSize too large for "
3229 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT: "
3230 "%"PRIu64"B > %"PRIu64"B",
3231 aligned_alloc_size, mem->bo->size);
3232 anv_device_release_bo(device, mem->bo);
3233 goto fail;
3234 }
3235
3236 /* From the Vulkan spec:
3237 *
3238 * "Importing memory from a file descriptor transfers ownership of
3239 * the file descriptor from the application to the Vulkan
3240 * implementation. The application must not perform any operations on
3241 * the file descriptor after a successful import."
3242 *
3243 * If the import fails, we leave the file descriptor open.
3244 */
3245 close(fd_info->fd);
3246 goto success;
3247 }
3248
3249 if (host_ptr_info && host_ptr_info->handleType) {
3250 if (host_ptr_info->handleType ==
3251 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_MAPPED_FOREIGN_MEMORY_BIT_EXT) {
3252 result = vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE);
3253 goto fail;
3254 }
3255
3256 assert(host_ptr_info->handleType ==
3257 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT);
3258
3259 result = anv_device_import_bo_from_host_ptr(device,
3260 host_ptr_info->pHostPointer,
3261 pAllocateInfo->allocationSize,
3262 alloc_flags,
3263 client_address,
3264 &mem->bo);
3265 if (result != VK_SUCCESS)
3266 goto fail;
3267
3268 mem->host_ptr = host_ptr_info->pHostPointer;
3269 goto success;
3270 }
3271
3272 /* Regular allocate (not importing memory). */
3273
3274 if (export_info && export_info->handleTypes)
3275 alloc_flags |= ANV_BO_ALLOC_EXTERNAL;
3276
3277 result = anv_device_alloc_bo(device, pAllocateInfo->allocationSize,
3278 alloc_flags, client_address, &mem->bo);
3279 if (result != VK_SUCCESS)
3280 goto fail;
3281
3282 if (dedicated_info && dedicated_info->image != VK_NULL_HANDLE) {
3283 ANV_FROM_HANDLE(anv_image, image, dedicated_info->image);
3284
3285 /* Some legacy (non-modifiers) consumers need the tiling to be set on
3286 * the BO. In this case, we have a dedicated allocation.
3287 */
3288 if (image->needs_set_tiling) {
3289 const uint32_t i915_tiling =
3290 isl_tiling_to_i915_tiling(image->planes[0].surface.isl.tiling);
3291 int ret = anv_gem_set_tiling(device, mem->bo->gem_handle,
3292 image->planes[0].surface.isl.row_pitch_B,
3293 i915_tiling);
3294 if (ret) {
3295 anv_device_release_bo(device, mem->bo);
3296 result = vk_errorf(device->instance, NULL,
3297 VK_ERROR_OUT_OF_DEVICE_MEMORY,
3298 "failed to set BO tiling: %m");
3299 goto fail;
3300 }
3301 }
3302 }
3303
3304 success:
3305 mem_heap_used = p_atomic_add_return(&mem_heap->used, mem->bo->size);
3306 if (mem_heap_used > mem_heap->size) {
3307 p_atomic_add(&mem_heap->used, -mem->bo->size);
3308 anv_device_release_bo(device, mem->bo);
3309 result = vk_errorf(device->instance, NULL,
3310 VK_ERROR_OUT_OF_DEVICE_MEMORY,
3311 "Out of heap memory");
3312 goto fail;
3313 }
3314
3315 pthread_mutex_lock(&device->mutex);
3316 list_addtail(&mem->link, &device->memory_objects);
3317 pthread_mutex_unlock(&device->mutex);
3318
3319 *pMem = anv_device_memory_to_handle(mem);
3320
3321 return VK_SUCCESS;
3322
3323 fail:
3324 vk_free2(&device->alloc, pAllocator, mem);
3325
3326 return result;
3327 }
3328
3329 VkResult anv_GetMemoryFdKHR(
3330 VkDevice device_h,
3331 const VkMemoryGetFdInfoKHR* pGetFdInfo,
3332 int* pFd)
3333 {
3334 ANV_FROM_HANDLE(anv_device, dev, device_h);
3335 ANV_FROM_HANDLE(anv_device_memory, mem, pGetFdInfo->memory);
3336
3337 assert(pGetFdInfo->sType == VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR);
3338
3339 assert(pGetFdInfo->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
3340 pGetFdInfo->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
3341
3342 return anv_device_export_bo(dev, mem->bo, pFd);
3343 }
3344
3345 VkResult anv_GetMemoryFdPropertiesKHR(
3346 VkDevice _device,
3347 VkExternalMemoryHandleTypeFlagBits handleType,
3348 int fd,
3349 VkMemoryFdPropertiesKHR* pMemoryFdProperties)
3350 {
3351 ANV_FROM_HANDLE(anv_device, device, _device);
3352 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3353
3354 switch (handleType) {
3355 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT:
3356 /* dma-buf can be imported as any memory type */
3357 pMemoryFdProperties->memoryTypeBits =
3358 (1 << pdevice->memory.type_count) - 1;
3359 return VK_SUCCESS;
3360
3361 default:
3362 /* The valid usage section for this function says:
3363 *
3364 * "handleType must not be one of the handle types defined as
3365 * opaque."
3366 *
3367 * So opaque handle types fall into the default "unsupported" case.
3368 */
3369 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE);
3370 }
3371 }
3372
3373 VkResult anv_GetMemoryHostPointerPropertiesEXT(
3374 VkDevice _device,
3375 VkExternalMemoryHandleTypeFlagBits handleType,
3376 const void* pHostPointer,
3377 VkMemoryHostPointerPropertiesEXT* pMemoryHostPointerProperties)
3378 {
3379 ANV_FROM_HANDLE(anv_device, device, _device);
3380
3381 assert(pMemoryHostPointerProperties->sType ==
3382 VK_STRUCTURE_TYPE_MEMORY_HOST_POINTER_PROPERTIES_EXT);
3383
3384 switch (handleType) {
3385 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT: {
3386 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3387
3388 /* Host memory can be imported as any memory type. */
3389 pMemoryHostPointerProperties->memoryTypeBits =
3390 (1ull << pdevice->memory.type_count) - 1;
3391
3392 return VK_SUCCESS;
3393 }
3394 default:
3395 return VK_ERROR_INVALID_EXTERNAL_HANDLE;
3396 }
3397 }
3398
3399 void anv_FreeMemory(
3400 VkDevice _device,
3401 VkDeviceMemory _mem,
3402 const VkAllocationCallbacks* pAllocator)
3403 {
3404 ANV_FROM_HANDLE(anv_device, device, _device);
3405 ANV_FROM_HANDLE(anv_device_memory, mem, _mem);
3406 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3407
3408 if (mem == NULL)
3409 return;
3410
3411 pthread_mutex_lock(&device->mutex);
3412 list_del(&mem->link);
3413 pthread_mutex_unlock(&device->mutex);
3414
3415 if (mem->map)
3416 anv_UnmapMemory(_device, _mem);
3417
3418 p_atomic_add(&pdevice->memory.heaps[mem->type->heapIndex].used,
3419 -mem->bo->size);
3420
3421 anv_device_release_bo(device, mem->bo);
3422
3423 #if defined(ANDROID) && ANDROID_API_LEVEL >= 26
3424 if (mem->ahw)
3425 AHardwareBuffer_release(mem->ahw);
3426 #endif
3427
3428 vk_free2(&device->alloc, pAllocator, mem);
3429 }
3430
3431 VkResult anv_MapMemory(
3432 VkDevice _device,
3433 VkDeviceMemory _memory,
3434 VkDeviceSize offset,
3435 VkDeviceSize size,
3436 VkMemoryMapFlags flags,
3437 void** ppData)
3438 {
3439 ANV_FROM_HANDLE(anv_device, device, _device);
3440 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
3441
3442 if (mem == NULL) {
3443 *ppData = NULL;
3444 return VK_SUCCESS;
3445 }
3446
3447 if (mem->host_ptr) {
3448 *ppData = mem->host_ptr + offset;
3449 return VK_SUCCESS;
3450 }
3451
3452 if (size == VK_WHOLE_SIZE)
3453 size = mem->bo->size - offset;
3454
3455 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
3456 *
3457 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
3458 * assert(size != 0);
3459 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
3460 * equal to the size of the memory minus offset
3461 */
3462 assert(size > 0);
3463 assert(offset + size <= mem->bo->size);
3464
3465 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
3466 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
3467 * at a time is valid. We could just mmap up front and return an offset
3468 * pointer here, but that may exhaust virtual memory on 32 bit
3469 * userspace. */
3470
3471 uint32_t gem_flags = 0;
3472
3473 if (!device->info.has_llc &&
3474 (mem->type->propertyFlags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT))
3475 gem_flags |= I915_MMAP_WC;
3476
3477 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
3478 uint64_t map_offset = offset & ~4095ull;
3479 assert(offset >= map_offset);
3480 uint64_t map_size = (offset + size) - map_offset;
3481
3482 /* Let's map whole pages */
3483 map_size = align_u64(map_size, 4096);
3484
3485 void *map = anv_gem_mmap(device, mem->bo->gem_handle,
3486 map_offset, map_size, gem_flags);
3487 if (map == MAP_FAILED)
3488 return vk_error(VK_ERROR_MEMORY_MAP_FAILED);
3489
3490 mem->map = map;
3491 mem->map_size = map_size;
3492
3493 *ppData = mem->map + (offset - map_offset);
3494
3495 return VK_SUCCESS;
3496 }
3497
3498 void anv_UnmapMemory(
3499 VkDevice _device,
3500 VkDeviceMemory _memory)
3501 {
3502 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
3503
3504 if (mem == NULL || mem->host_ptr)
3505 return;
3506
3507 anv_gem_munmap(mem->map, mem->map_size);
3508
3509 mem->map = NULL;
3510 mem->map_size = 0;
3511 }
3512
3513 static void
3514 clflush_mapped_ranges(struct anv_device *device,
3515 uint32_t count,
3516 const VkMappedMemoryRange *ranges)
3517 {
3518 for (uint32_t i = 0; i < count; i++) {
3519 ANV_FROM_HANDLE(anv_device_memory, mem, ranges[i].memory);
3520 if (ranges[i].offset >= mem->map_size)
3521 continue;
3522
3523 gen_clflush_range(mem->map + ranges[i].offset,
3524 MIN2(ranges[i].size, mem->map_size - ranges[i].offset));
3525 }
3526 }
3527
3528 VkResult anv_FlushMappedMemoryRanges(
3529 VkDevice _device,
3530 uint32_t memoryRangeCount,
3531 const VkMappedMemoryRange* pMemoryRanges)
3532 {
3533 ANV_FROM_HANDLE(anv_device, device, _device);
3534
3535 if (device->info.has_llc)
3536 return VK_SUCCESS;
3537
3538 /* Make sure the writes we're flushing have landed. */
3539 __builtin_ia32_mfence();
3540
3541 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
3542
3543 return VK_SUCCESS;
3544 }
3545
3546 VkResult anv_InvalidateMappedMemoryRanges(
3547 VkDevice _device,
3548 uint32_t memoryRangeCount,
3549 const VkMappedMemoryRange* pMemoryRanges)
3550 {
3551 ANV_FROM_HANDLE(anv_device, device, _device);
3552
3553 if (device->info.has_llc)
3554 return VK_SUCCESS;
3555
3556 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
3557
3558 /* Make sure no reads get moved up above the invalidate. */
3559 __builtin_ia32_mfence();
3560
3561 return VK_SUCCESS;
3562 }
3563
3564 void anv_GetBufferMemoryRequirements(
3565 VkDevice _device,
3566 VkBuffer _buffer,
3567 VkMemoryRequirements* pMemoryRequirements)
3568 {
3569 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
3570 ANV_FROM_HANDLE(anv_device, device, _device);
3571 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3572
3573 /* The Vulkan spec (git aaed022) says:
3574 *
3575 * memoryTypeBits is a bitfield and contains one bit set for every
3576 * supported memory type for the resource. The bit `1<<i` is set if and
3577 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3578 * structure for the physical device is supported.
3579 */
3580 uint32_t memory_types = (1ull << pdevice->memory.type_count) - 1;
3581
3582 /* Base alignment requirement of a cache line */
3583 uint32_t alignment = 16;
3584
3585 /* We need an alignment of 32 for pushing UBOs */
3586 if (buffer->usage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT)
3587 alignment = MAX2(alignment, 32);
3588
3589 pMemoryRequirements->size = buffer->size;
3590 pMemoryRequirements->alignment = alignment;
3591
3592 /* Storage and Uniform buffers should have their size aligned to
3593 * 32-bits to avoid boundary checks when last DWord is not complete.
3594 * This would ensure that not internal padding would be needed for
3595 * 16-bit types.
3596 */
3597 if (device->robust_buffer_access &&
3598 (buffer->usage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT ||
3599 buffer->usage & VK_BUFFER_USAGE_STORAGE_BUFFER_BIT))
3600 pMemoryRequirements->size = align_u64(buffer->size, 4);
3601
3602 pMemoryRequirements->memoryTypeBits = memory_types;
3603 }
3604
3605 void anv_GetBufferMemoryRequirements2(
3606 VkDevice _device,
3607 const VkBufferMemoryRequirementsInfo2* pInfo,
3608 VkMemoryRequirements2* pMemoryRequirements)
3609 {
3610 anv_GetBufferMemoryRequirements(_device, pInfo->buffer,
3611 &pMemoryRequirements->memoryRequirements);
3612
3613 vk_foreach_struct(ext, pMemoryRequirements->pNext) {
3614 switch (ext->sType) {
3615 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
3616 VkMemoryDedicatedRequirements *requirements = (void *)ext;
3617 requirements->prefersDedicatedAllocation = false;
3618 requirements->requiresDedicatedAllocation = false;
3619 break;
3620 }
3621
3622 default:
3623 anv_debug_ignored_stype(ext->sType);
3624 break;
3625 }
3626 }
3627 }
3628
3629 void anv_GetImageMemoryRequirements(
3630 VkDevice _device,
3631 VkImage _image,
3632 VkMemoryRequirements* pMemoryRequirements)
3633 {
3634 ANV_FROM_HANDLE(anv_image, image, _image);
3635 ANV_FROM_HANDLE(anv_device, device, _device);
3636 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3637
3638 /* The Vulkan spec (git aaed022) says:
3639 *
3640 * memoryTypeBits is a bitfield and contains one bit set for every
3641 * supported memory type for the resource. The bit `1<<i` is set if and
3642 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3643 * structure for the physical device is supported.
3644 *
3645 * All types are currently supported for images.
3646 */
3647 uint32_t memory_types = (1ull << pdevice->memory.type_count) - 1;
3648
3649 /* We must have image allocated or imported at this point. According to the
3650 * specification, external images must have been bound to memory before
3651 * calling GetImageMemoryRequirements.
3652 */
3653 assert(image->size > 0);
3654
3655 pMemoryRequirements->size = image->size;
3656 pMemoryRequirements->alignment = image->alignment;
3657 pMemoryRequirements->memoryTypeBits = memory_types;
3658 }
3659
3660 void anv_GetImageMemoryRequirements2(
3661 VkDevice _device,
3662 const VkImageMemoryRequirementsInfo2* pInfo,
3663 VkMemoryRequirements2* pMemoryRequirements)
3664 {
3665 ANV_FROM_HANDLE(anv_device, device, _device);
3666 ANV_FROM_HANDLE(anv_image, image, pInfo->image);
3667
3668 anv_GetImageMemoryRequirements(_device, pInfo->image,
3669 &pMemoryRequirements->memoryRequirements);
3670
3671 vk_foreach_struct_const(ext, pInfo->pNext) {
3672 switch (ext->sType) {
3673 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO: {
3674 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3675 const VkImagePlaneMemoryRequirementsInfo *plane_reqs =
3676 (const VkImagePlaneMemoryRequirementsInfo *) ext;
3677 uint32_t plane = anv_image_aspect_to_plane(image->aspects,
3678 plane_reqs->planeAspect);
3679
3680 assert(image->planes[plane].offset == 0);
3681
3682 /* The Vulkan spec (git aaed022) says:
3683 *
3684 * memoryTypeBits is a bitfield and contains one bit set for every
3685 * supported memory type for the resource. The bit `1<<i` is set
3686 * if and only if the memory type `i` in the
3687 * VkPhysicalDeviceMemoryProperties structure for the physical
3688 * device is supported.
3689 *
3690 * All types are currently supported for images.
3691 */
3692 pMemoryRequirements->memoryRequirements.memoryTypeBits =
3693 (1ull << pdevice->memory.type_count) - 1;
3694
3695 /* We must have image allocated or imported at this point. According to the
3696 * specification, external images must have been bound to memory before
3697 * calling GetImageMemoryRequirements.
3698 */
3699 assert(image->planes[plane].size > 0);
3700
3701 pMemoryRequirements->memoryRequirements.size = image->planes[plane].size;
3702 pMemoryRequirements->memoryRequirements.alignment =
3703 image->planes[plane].alignment;
3704 break;
3705 }
3706
3707 default:
3708 anv_debug_ignored_stype(ext->sType);
3709 break;
3710 }
3711 }
3712
3713 vk_foreach_struct(ext, pMemoryRequirements->pNext) {
3714 switch (ext->sType) {
3715 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
3716 VkMemoryDedicatedRequirements *requirements = (void *)ext;
3717 if (image->needs_set_tiling || image->external_format) {
3718 /* If we need to set the tiling for external consumers, we need a
3719 * dedicated allocation.
3720 *
3721 * See also anv_AllocateMemory.
3722 */
3723 requirements->prefersDedicatedAllocation = true;
3724 requirements->requiresDedicatedAllocation = true;
3725 } else {
3726 requirements->prefersDedicatedAllocation = false;
3727 requirements->requiresDedicatedAllocation = false;
3728 }
3729 break;
3730 }
3731
3732 default:
3733 anv_debug_ignored_stype(ext->sType);
3734 break;
3735 }
3736 }
3737 }
3738
3739 void anv_GetImageSparseMemoryRequirements(
3740 VkDevice device,
3741 VkImage image,
3742 uint32_t* pSparseMemoryRequirementCount,
3743 VkSparseImageMemoryRequirements* pSparseMemoryRequirements)
3744 {
3745 *pSparseMemoryRequirementCount = 0;
3746 }
3747
3748 void anv_GetImageSparseMemoryRequirements2(
3749 VkDevice device,
3750 const VkImageSparseMemoryRequirementsInfo2* pInfo,
3751 uint32_t* pSparseMemoryRequirementCount,
3752 VkSparseImageMemoryRequirements2* pSparseMemoryRequirements)
3753 {
3754 *pSparseMemoryRequirementCount = 0;
3755 }
3756
3757 void anv_GetDeviceMemoryCommitment(
3758 VkDevice device,
3759 VkDeviceMemory memory,
3760 VkDeviceSize* pCommittedMemoryInBytes)
3761 {
3762 *pCommittedMemoryInBytes = 0;
3763 }
3764
3765 static void
3766 anv_bind_buffer_memory(const VkBindBufferMemoryInfo *pBindInfo)
3767 {
3768 ANV_FROM_HANDLE(anv_device_memory, mem, pBindInfo->memory);
3769 ANV_FROM_HANDLE(anv_buffer, buffer, pBindInfo->buffer);
3770
3771 assert(pBindInfo->sType == VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO);
3772
3773 if (mem) {
3774 buffer->address = (struct anv_address) {
3775 .bo = mem->bo,
3776 .offset = pBindInfo->memoryOffset,
3777 };
3778 } else {
3779 buffer->address = ANV_NULL_ADDRESS;
3780 }
3781 }
3782
3783 VkResult anv_BindBufferMemory(
3784 VkDevice device,
3785 VkBuffer buffer,
3786 VkDeviceMemory memory,
3787 VkDeviceSize memoryOffset)
3788 {
3789 anv_bind_buffer_memory(
3790 &(VkBindBufferMemoryInfo) {
3791 .sType = VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO,
3792 .buffer = buffer,
3793 .memory = memory,
3794 .memoryOffset = memoryOffset,
3795 });
3796
3797 return VK_SUCCESS;
3798 }
3799
3800 VkResult anv_BindBufferMemory2(
3801 VkDevice device,
3802 uint32_t bindInfoCount,
3803 const VkBindBufferMemoryInfo* pBindInfos)
3804 {
3805 for (uint32_t i = 0; i < bindInfoCount; i++)
3806 anv_bind_buffer_memory(&pBindInfos[i]);
3807
3808 return VK_SUCCESS;
3809 }
3810
3811 VkResult anv_QueueBindSparse(
3812 VkQueue _queue,
3813 uint32_t bindInfoCount,
3814 const VkBindSparseInfo* pBindInfo,
3815 VkFence fence)
3816 {
3817 ANV_FROM_HANDLE(anv_queue, queue, _queue);
3818 if (anv_device_is_lost(queue->device))
3819 return VK_ERROR_DEVICE_LOST;
3820
3821 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
3822 }
3823
3824 // Event functions
3825
3826 VkResult anv_CreateEvent(
3827 VkDevice _device,
3828 const VkEventCreateInfo* pCreateInfo,
3829 const VkAllocationCallbacks* pAllocator,
3830 VkEvent* pEvent)
3831 {
3832 ANV_FROM_HANDLE(anv_device, device, _device);
3833 struct anv_state state;
3834 struct anv_event *event;
3835
3836 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_EVENT_CREATE_INFO);
3837
3838 state = anv_state_pool_alloc(&device->dynamic_state_pool,
3839 sizeof(*event), 8);
3840 event = state.map;
3841 event->state = state;
3842 event->semaphore = VK_EVENT_RESET;
3843
3844 if (!device->info.has_llc) {
3845 /* Make sure the writes we're flushing have landed. */
3846 __builtin_ia32_mfence();
3847 __builtin_ia32_clflush(event);
3848 }
3849
3850 *pEvent = anv_event_to_handle(event);
3851
3852 return VK_SUCCESS;
3853 }
3854
3855 void anv_DestroyEvent(
3856 VkDevice _device,
3857 VkEvent _event,
3858 const VkAllocationCallbacks* pAllocator)
3859 {
3860 ANV_FROM_HANDLE(anv_device, device, _device);
3861 ANV_FROM_HANDLE(anv_event, event, _event);
3862
3863 if (!event)
3864 return;
3865
3866 anv_state_pool_free(&device->dynamic_state_pool, event->state);
3867 }
3868
3869 VkResult anv_GetEventStatus(
3870 VkDevice _device,
3871 VkEvent _event)
3872 {
3873 ANV_FROM_HANDLE(anv_device, device, _device);
3874 ANV_FROM_HANDLE(anv_event, event, _event);
3875
3876 if (anv_device_is_lost(device))
3877 return VK_ERROR_DEVICE_LOST;
3878
3879 if (!device->info.has_llc) {
3880 /* Invalidate read cache before reading event written by GPU. */
3881 __builtin_ia32_clflush(event);
3882 __builtin_ia32_mfence();
3883
3884 }
3885
3886 return event->semaphore;
3887 }
3888
3889 VkResult anv_SetEvent(
3890 VkDevice _device,
3891 VkEvent _event)
3892 {
3893 ANV_FROM_HANDLE(anv_device, device, _device);
3894 ANV_FROM_HANDLE(anv_event, event, _event);
3895
3896 event->semaphore = VK_EVENT_SET;
3897
3898 if (!device->info.has_llc) {
3899 /* Make sure the writes we're flushing have landed. */
3900 __builtin_ia32_mfence();
3901 __builtin_ia32_clflush(event);
3902 }
3903
3904 return VK_SUCCESS;
3905 }
3906
3907 VkResult anv_ResetEvent(
3908 VkDevice _device,
3909 VkEvent _event)
3910 {
3911 ANV_FROM_HANDLE(anv_device, device, _device);
3912 ANV_FROM_HANDLE(anv_event, event, _event);
3913
3914 event->semaphore = VK_EVENT_RESET;
3915
3916 if (!device->info.has_llc) {
3917 /* Make sure the writes we're flushing have landed. */
3918 __builtin_ia32_mfence();
3919 __builtin_ia32_clflush(event);
3920 }
3921
3922 return VK_SUCCESS;
3923 }
3924
3925 // Buffer functions
3926
3927 VkResult anv_CreateBuffer(
3928 VkDevice _device,
3929 const VkBufferCreateInfo* pCreateInfo,
3930 const VkAllocationCallbacks* pAllocator,
3931 VkBuffer* pBuffer)
3932 {
3933 ANV_FROM_HANDLE(anv_device, device, _device);
3934 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3935 struct anv_buffer *buffer;
3936
3937 /* Don't allow creating buffers bigger than our address space. The real
3938 * issue here is that we may align up the buffer size and we don't want
3939 * doing so to cause roll-over. However, no one has any business
3940 * allocating a buffer larger than our GTT size.
3941 */
3942 if (pCreateInfo->size > pdevice->gtt_size)
3943 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
3944
3945 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
3946
3947 buffer = vk_alloc2(&device->alloc, pAllocator, sizeof(*buffer), 8,
3948 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
3949 if (buffer == NULL)
3950 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
3951
3952 buffer->size = pCreateInfo->size;
3953 buffer->usage = pCreateInfo->usage;
3954 buffer->address = ANV_NULL_ADDRESS;
3955
3956 *pBuffer = anv_buffer_to_handle(buffer);
3957
3958 return VK_SUCCESS;
3959 }
3960
3961 void anv_DestroyBuffer(
3962 VkDevice _device,
3963 VkBuffer _buffer,
3964 const VkAllocationCallbacks* pAllocator)
3965 {
3966 ANV_FROM_HANDLE(anv_device, device, _device);
3967 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
3968
3969 if (!buffer)
3970 return;
3971
3972 vk_free2(&device->alloc, pAllocator, buffer);
3973 }
3974
3975 VkDeviceAddress anv_GetBufferDeviceAddressKHR(
3976 VkDevice device,
3977 const VkBufferDeviceAddressInfoKHR* pInfo)
3978 {
3979 ANV_FROM_HANDLE(anv_buffer, buffer, pInfo->buffer);
3980
3981 assert(!anv_address_is_null(buffer->address));
3982 assert(buffer->address.bo->flags & EXEC_OBJECT_PINNED);
3983
3984 return anv_address_physical(buffer->address);
3985 }
3986
3987 uint64_t anv_GetBufferOpaqueCaptureAddressKHR(
3988 VkDevice device,
3989 const VkBufferDeviceAddressInfoKHR* pInfo)
3990 {
3991 return 0;
3992 }
3993
3994 uint64_t anv_GetDeviceMemoryOpaqueCaptureAddressKHR(
3995 VkDevice device,
3996 const VkDeviceMemoryOpaqueCaptureAddressInfoKHR* pInfo)
3997 {
3998 ANV_FROM_HANDLE(anv_device_memory, memory, pInfo->memory);
3999
4000 assert(memory->bo->flags & EXEC_OBJECT_PINNED);
4001 assert(memory->bo->has_client_visible_address);
4002
4003 return gen_48b_address(memory->bo->offset);
4004 }
4005
4006 void
4007 anv_fill_buffer_surface_state(struct anv_device *device, struct anv_state state,
4008 enum isl_format format,
4009 struct anv_address address,
4010 uint32_t range, uint32_t stride)
4011 {
4012 isl_buffer_fill_state(&device->isl_dev, state.map,
4013 .address = anv_address_physical(address),
4014 .mocs = device->isl_dev.mocs.internal,
4015 .size_B = range,
4016 .format = format,
4017 .swizzle = ISL_SWIZZLE_IDENTITY,
4018 .stride_B = stride);
4019 }
4020
4021 void anv_DestroySampler(
4022 VkDevice _device,
4023 VkSampler _sampler,
4024 const VkAllocationCallbacks* pAllocator)
4025 {
4026 ANV_FROM_HANDLE(anv_device, device, _device);
4027 ANV_FROM_HANDLE(anv_sampler, sampler, _sampler);
4028
4029 if (!sampler)
4030 return;
4031
4032 if (sampler->bindless_state.map) {
4033 anv_state_pool_free(&device->dynamic_state_pool,
4034 sampler->bindless_state);
4035 }
4036
4037 vk_free2(&device->alloc, pAllocator, sampler);
4038 }
4039
4040 VkResult anv_CreateFramebuffer(
4041 VkDevice _device,
4042 const VkFramebufferCreateInfo* pCreateInfo,
4043 const VkAllocationCallbacks* pAllocator,
4044 VkFramebuffer* pFramebuffer)
4045 {
4046 ANV_FROM_HANDLE(anv_device, device, _device);
4047 struct anv_framebuffer *framebuffer;
4048
4049 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
4050
4051 size_t size = sizeof(*framebuffer);
4052
4053 /* VK_KHR_imageless_framebuffer extension says:
4054 *
4055 * If flags includes VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR,
4056 * parameter pAttachments is ignored.
4057 */
4058 if (!(pCreateInfo->flags & VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR)) {
4059 size += sizeof(struct anv_image_view *) * pCreateInfo->attachmentCount;
4060 framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
4061 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
4062 if (framebuffer == NULL)
4063 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
4064
4065 for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
4066 ANV_FROM_HANDLE(anv_image_view, iview, pCreateInfo->pAttachments[i]);
4067 framebuffer->attachments[i] = iview;
4068 }
4069 framebuffer->attachment_count = pCreateInfo->attachmentCount;
4070 } else {
4071 assert(device->enabled_extensions.KHR_imageless_framebuffer);
4072 framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
4073 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
4074 if (framebuffer == NULL)
4075 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
4076
4077 framebuffer->attachment_count = 0;
4078 }
4079
4080 framebuffer->width = pCreateInfo->width;
4081 framebuffer->height = pCreateInfo->height;
4082 framebuffer->layers = pCreateInfo->layers;
4083
4084 *pFramebuffer = anv_framebuffer_to_handle(framebuffer);
4085
4086 return VK_SUCCESS;
4087 }
4088
4089 void anv_DestroyFramebuffer(
4090 VkDevice _device,
4091 VkFramebuffer _fb,
4092 const VkAllocationCallbacks* pAllocator)
4093 {
4094 ANV_FROM_HANDLE(anv_device, device, _device);
4095 ANV_FROM_HANDLE(anv_framebuffer, fb, _fb);
4096
4097 if (!fb)
4098 return;
4099
4100 vk_free2(&device->alloc, pAllocator, fb);
4101 }
4102
4103 static const VkTimeDomainEXT anv_time_domains[] = {
4104 VK_TIME_DOMAIN_DEVICE_EXT,
4105 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT,
4106 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT,
4107 };
4108
4109 VkResult anv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
4110 VkPhysicalDevice physicalDevice,
4111 uint32_t *pTimeDomainCount,
4112 VkTimeDomainEXT *pTimeDomains)
4113 {
4114 int d;
4115 VK_OUTARRAY_MAKE(out, pTimeDomains, pTimeDomainCount);
4116
4117 for (d = 0; d < ARRAY_SIZE(anv_time_domains); d++) {
4118 vk_outarray_append(&out, i) {
4119 *i = anv_time_domains[d];
4120 }
4121 }
4122
4123 return vk_outarray_status(&out);
4124 }
4125
4126 static uint64_t
4127 anv_clock_gettime(clockid_t clock_id)
4128 {
4129 struct timespec current;
4130 int ret;
4131
4132 ret = clock_gettime(clock_id, &current);
4133 if (ret < 0 && clock_id == CLOCK_MONOTONIC_RAW)
4134 ret = clock_gettime(CLOCK_MONOTONIC, &current);
4135 if (ret < 0)
4136 return 0;
4137
4138 return (uint64_t) current.tv_sec * 1000000000ULL + current.tv_nsec;
4139 }
4140
4141 #define TIMESTAMP 0x2358
4142
4143 VkResult anv_GetCalibratedTimestampsEXT(
4144 VkDevice _device,
4145 uint32_t timestampCount,
4146 const VkCalibratedTimestampInfoEXT *pTimestampInfos,
4147 uint64_t *pTimestamps,
4148 uint64_t *pMaxDeviation)
4149 {
4150 ANV_FROM_HANDLE(anv_device, device, _device);
4151 uint64_t timestamp_frequency = device->info.timestamp_frequency;
4152 int ret;
4153 int d;
4154 uint64_t begin, end;
4155 uint64_t max_clock_period = 0;
4156
4157 begin = anv_clock_gettime(CLOCK_MONOTONIC_RAW);
4158
4159 for (d = 0; d < timestampCount; d++) {
4160 switch (pTimestampInfos[d].timeDomain) {
4161 case VK_TIME_DOMAIN_DEVICE_EXT:
4162 ret = anv_gem_reg_read(device, TIMESTAMP | 1,
4163 &pTimestamps[d]);
4164
4165 if (ret != 0) {
4166 return anv_device_set_lost(device, "Failed to read the TIMESTAMP "
4167 "register: %m");
4168 }
4169 uint64_t device_period = DIV_ROUND_UP(1000000000, timestamp_frequency);
4170 max_clock_period = MAX2(max_clock_period, device_period);
4171 break;
4172 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT:
4173 pTimestamps[d] = anv_clock_gettime(CLOCK_MONOTONIC);
4174 max_clock_period = MAX2(max_clock_period, 1);
4175 break;
4176
4177 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT:
4178 pTimestamps[d] = begin;
4179 break;
4180 default:
4181 pTimestamps[d] = 0;
4182 break;
4183 }
4184 }
4185
4186 end = anv_clock_gettime(CLOCK_MONOTONIC_RAW);
4187
4188 /*
4189 * The maximum deviation is the sum of the interval over which we
4190 * perform the sampling and the maximum period of any sampled
4191 * clock. That's because the maximum skew between any two sampled
4192 * clock edges is when the sampled clock with the largest period is
4193 * sampled at the end of that period but right at the beginning of the
4194 * sampling interval and some other clock is sampled right at the
4195 * begining of its sampling period and right at the end of the
4196 * sampling interval. Let's assume the GPU has the longest clock
4197 * period and that the application is sampling GPU and monotonic:
4198 *
4199 * s e
4200 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
4201 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4202 *
4203 * g
4204 * 0 1 2 3
4205 * GPU -----_____-----_____-----_____-----_____
4206 *
4207 * m
4208 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
4209 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4210 *
4211 * Interval <----------------->
4212 * Deviation <-------------------------->
4213 *
4214 * s = read(raw) 2
4215 * g = read(GPU) 1
4216 * m = read(monotonic) 2
4217 * e = read(raw) b
4218 *
4219 * We round the sample interval up by one tick to cover sampling error
4220 * in the interval clock
4221 */
4222
4223 uint64_t sample_interval = end - begin + 1;
4224
4225 *pMaxDeviation = sample_interval + max_clock_period;
4226
4227 return VK_SUCCESS;
4228 }
4229
4230 /* vk_icd.h does not declare this function, so we declare it here to
4231 * suppress Wmissing-prototypes.
4232 */
4233 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4234 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion);
4235
4236 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4237 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion)
4238 {
4239 /* For the full details on loader interface versioning, see
4240 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
4241 * What follows is a condensed summary, to help you navigate the large and
4242 * confusing official doc.
4243 *
4244 * - Loader interface v0 is incompatible with later versions. We don't
4245 * support it.
4246 *
4247 * - In loader interface v1:
4248 * - The first ICD entrypoint called by the loader is
4249 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
4250 * entrypoint.
4251 * - The ICD must statically expose no other Vulkan symbol unless it is
4252 * linked with -Bsymbolic.
4253 * - Each dispatchable Vulkan handle created by the ICD must be
4254 * a pointer to a struct whose first member is VK_LOADER_DATA. The
4255 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
4256 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
4257 * vkDestroySurfaceKHR(). The ICD must be capable of working with
4258 * such loader-managed surfaces.
4259 *
4260 * - Loader interface v2 differs from v1 in:
4261 * - The first ICD entrypoint called by the loader is
4262 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
4263 * statically expose this entrypoint.
4264 *
4265 * - Loader interface v3 differs from v2 in:
4266 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
4267 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
4268 * because the loader no longer does so.
4269 *
4270 * - Loader interface v4 differs from v3 in:
4271 * - The ICD must implement vk_icdGetPhysicalDeviceProcAddr().
4272 */
4273 *pSupportedVersion = MIN2(*pSupportedVersion, 4u);
4274 return VK_SUCCESS;
4275 }