anv: Replace ANV_BO_EXTERNAL with anv_bo::is_external
[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 DRI_CONF_END;
58
59 /* This is probably far to big but it reflects the max size used for messages
60 * in OpenGLs KHR_debug.
61 */
62 #define MAX_DEBUG_MESSAGE_LENGTH 4096
63
64 static void
65 compiler_debug_log(void *data, const char *fmt, ...)
66 {
67 char str[MAX_DEBUG_MESSAGE_LENGTH];
68 struct anv_device *device = (struct anv_device *)data;
69
70 if (list_is_empty(&device->instance->debug_report_callbacks.callbacks))
71 return;
72
73 va_list args;
74 va_start(args, fmt);
75 (void) vsnprintf(str, MAX_DEBUG_MESSAGE_LENGTH, fmt, args);
76 va_end(args);
77
78 vk_debug_report(&device->instance->debug_report_callbacks,
79 VK_DEBUG_REPORT_DEBUG_BIT_EXT,
80 VK_DEBUG_REPORT_OBJECT_TYPE_UNKNOWN_EXT,
81 0, 0, 0, "anv", str);
82 }
83
84 static void
85 compiler_perf_log(void *data, const char *fmt, ...)
86 {
87 va_list args;
88 va_start(args, fmt);
89
90 if (unlikely(INTEL_DEBUG & DEBUG_PERF))
91 intel_logd_v(fmt, args);
92
93 va_end(args);
94 }
95
96 static uint64_t
97 anv_compute_heap_size(int fd, uint64_t gtt_size)
98 {
99 /* Query the total ram from the system */
100 struct sysinfo info;
101 sysinfo(&info);
102
103 uint64_t total_ram = (uint64_t)info.totalram * (uint64_t)info.mem_unit;
104
105 /* We don't want to burn too much ram with the GPU. If the user has 4GiB
106 * or less, we use at most half. If they have more than 4GiB, we use 3/4.
107 */
108 uint64_t available_ram;
109 if (total_ram <= 4ull * 1024ull * 1024ull * 1024ull)
110 available_ram = total_ram / 2;
111 else
112 available_ram = total_ram * 3 / 4;
113
114 /* We also want to leave some padding for things we allocate in the driver,
115 * so don't go over 3/4 of the GTT either.
116 */
117 uint64_t available_gtt = gtt_size * 3 / 4;
118
119 return MIN2(available_ram, available_gtt);
120 }
121
122 static VkResult
123 anv_physical_device_init_heaps(struct anv_physical_device *device, int fd)
124 {
125 uint64_t gtt_size;
126 if (anv_gem_get_context_param(fd, 0, I915_CONTEXT_PARAM_GTT_SIZE,
127 &gtt_size) == -1) {
128 /* If, for whatever reason, we can't actually get the GTT size from the
129 * kernel (too old?) fall back to the aperture size.
130 */
131 anv_perf_warn(NULL, NULL,
132 "Failed to get I915_CONTEXT_PARAM_GTT_SIZE: %m");
133
134 if (anv_gem_get_aperture(fd, &gtt_size) == -1) {
135 return vk_errorf(NULL, NULL, VK_ERROR_INITIALIZATION_FAILED,
136 "failed to get aperture size: %m");
137 }
138 }
139
140 device->supports_48bit_addresses = (device->info.gen >= 8) &&
141 gtt_size > (4ULL << 30 /* GiB */);
142
143 uint64_t heap_size = anv_compute_heap_size(fd, gtt_size);
144
145 if (heap_size > (2ull << 30) && !device->supports_48bit_addresses) {
146 /* When running with an overridden PCI ID, we may get a GTT size from
147 * the kernel that is greater than 2 GiB but the execbuf check for 48bit
148 * address support can still fail. Just clamp the address space size to
149 * 2 GiB if we don't have 48-bit support.
150 */
151 intel_logw("%s:%d: The kernel reported a GTT size larger than 2 GiB but "
152 "not support for 48-bit addresses",
153 __FILE__, __LINE__);
154 heap_size = 2ull << 30;
155 }
156
157 if (heap_size <= 3ull * (1ull << 30)) {
158 /* In this case, everything fits nicely into the 32-bit address space,
159 * so there's no need for supporting 48bit addresses on client-allocated
160 * memory objects.
161 */
162 device->memory.heap_count = 1;
163 device->memory.heaps[0] = (struct anv_memory_heap) {
164 .vma_start = LOW_HEAP_MIN_ADDRESS,
165 .vma_size = LOW_HEAP_SIZE,
166 .size = heap_size,
167 .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
168 .supports_48bit_addresses = false,
169 };
170 } else {
171 /* Not everything will fit nicely into a 32-bit address space. In this
172 * case we need a 64-bit heap. Advertise a small 32-bit heap and a
173 * larger 48-bit heap. If we're in this case, then we have a total heap
174 * size larger than 3GiB which most likely means they have 8 GiB of
175 * video memory and so carving off 1 GiB for the 32-bit heap should be
176 * reasonable.
177 */
178 const uint64_t heap_size_32bit = 1ull << 30;
179 const uint64_t heap_size_48bit = heap_size - heap_size_32bit;
180
181 assert(device->supports_48bit_addresses);
182
183 device->memory.heap_count = 2;
184 device->memory.heaps[0] = (struct anv_memory_heap) {
185 .vma_start = HIGH_HEAP_MIN_ADDRESS,
186 /* Leave the last 4GiB out of the high vma range, so that no state
187 * base address + size can overflow 48 bits. For more information see
188 * the comment about Wa32bitGeneralStateOffset in anv_allocator.c
189 */
190 .vma_size = gtt_size - (1ull << 32) - HIGH_HEAP_MIN_ADDRESS,
191 .size = heap_size_48bit,
192 .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
193 .supports_48bit_addresses = true,
194 };
195 device->memory.heaps[1] = (struct anv_memory_heap) {
196 .vma_start = LOW_HEAP_MIN_ADDRESS,
197 .vma_size = LOW_HEAP_SIZE,
198 .size = heap_size_32bit,
199 .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
200 .supports_48bit_addresses = false,
201 };
202 }
203
204 uint32_t type_count = 0;
205 for (uint32_t heap = 0; heap < device->memory.heap_count; heap++) {
206 uint32_t valid_buffer_usage = ~0;
207
208 /* There appears to be a hardware issue in the VF cache where it only
209 * considers the bottom 32 bits of memory addresses. If you happen to
210 * have two vertex buffers which get placed exactly 4 GiB apart and use
211 * them in back-to-back draw calls, you can get collisions. In order to
212 * solve this problem, we require vertex and index buffers be bound to
213 * memory allocated out of the 32-bit heap.
214 */
215 if (device->memory.heaps[heap].supports_48bit_addresses) {
216 valid_buffer_usage &= ~(VK_BUFFER_USAGE_INDEX_BUFFER_BIT |
217 VK_BUFFER_USAGE_VERTEX_BUFFER_BIT);
218 }
219
220 if (device->info.has_llc) {
221 /* Big core GPUs share LLC with the CPU and thus one memory type can be
222 * both cached and coherent at the same time.
223 */
224 device->memory.types[type_count++] = (struct anv_memory_type) {
225 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
226 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
227 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
228 VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
229 .heapIndex = heap,
230 .valid_buffer_usage = valid_buffer_usage,
231 };
232 } else {
233 /* The spec requires that we expose a host-visible, coherent memory
234 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
235 * to give the application a choice between cached, but not coherent and
236 * coherent but uncached (WC though).
237 */
238 device->memory.types[type_count++] = (struct anv_memory_type) {
239 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
240 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
241 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
242 .heapIndex = heap,
243 .valid_buffer_usage = valid_buffer_usage,
244 };
245 device->memory.types[type_count++] = (struct anv_memory_type) {
246 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
247 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
248 VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
249 .heapIndex = heap,
250 .valid_buffer_usage = valid_buffer_usage,
251 };
252 }
253 }
254 device->memory.type_count = type_count;
255
256 return VK_SUCCESS;
257 }
258
259 static VkResult
260 anv_physical_device_init_uuids(struct anv_physical_device *device)
261 {
262 const struct build_id_note *note =
263 build_id_find_nhdr_for_addr(anv_physical_device_init_uuids);
264 if (!note) {
265 return vk_errorf(device->instance, device,
266 VK_ERROR_INITIALIZATION_FAILED,
267 "Failed to find build-id");
268 }
269
270 unsigned build_id_len = build_id_length(note);
271 if (build_id_len < 20) {
272 return vk_errorf(device->instance, device,
273 VK_ERROR_INITIALIZATION_FAILED,
274 "build-id too short. It needs to be a SHA");
275 }
276
277 memcpy(device->driver_build_sha1, build_id_data(note), 20);
278
279 struct mesa_sha1 sha1_ctx;
280 uint8_t sha1[20];
281 STATIC_ASSERT(VK_UUID_SIZE <= sizeof(sha1));
282
283 /* The pipeline cache UUID is used for determining when a pipeline cache is
284 * invalid. It needs both a driver build and the PCI ID of the device.
285 */
286 _mesa_sha1_init(&sha1_ctx);
287 _mesa_sha1_update(&sha1_ctx, build_id_data(note), build_id_len);
288 _mesa_sha1_update(&sha1_ctx, &device->chipset_id,
289 sizeof(device->chipset_id));
290 _mesa_sha1_update(&sha1_ctx, &device->always_use_bindless,
291 sizeof(device->always_use_bindless));
292 _mesa_sha1_update(&sha1_ctx, &device->has_a64_buffer_access,
293 sizeof(device->has_a64_buffer_access));
294 _mesa_sha1_update(&sha1_ctx, &device->has_bindless_images,
295 sizeof(device->has_bindless_images));
296 _mesa_sha1_update(&sha1_ctx, &device->has_bindless_samplers,
297 sizeof(device->has_bindless_samplers));
298 _mesa_sha1_final(&sha1_ctx, sha1);
299 memcpy(device->pipeline_cache_uuid, sha1, VK_UUID_SIZE);
300
301 /* The driver UUID is used for determining sharability of images and memory
302 * between two Vulkan instances in separate processes. People who want to
303 * share memory need to also check the device UUID (below) so all this
304 * needs to be is the build-id.
305 */
306 memcpy(device->driver_uuid, build_id_data(note), VK_UUID_SIZE);
307
308 /* The device UUID uniquely identifies the given device within the machine.
309 * Since we never have more than one device, this doesn't need to be a real
310 * UUID. However, on the off-chance that someone tries to use this to
311 * cache pre-tiled images or something of the like, we use the PCI ID and
312 * some bits of ISL info to ensure that this is safe.
313 */
314 _mesa_sha1_init(&sha1_ctx);
315 _mesa_sha1_update(&sha1_ctx, &device->chipset_id,
316 sizeof(device->chipset_id));
317 _mesa_sha1_update(&sha1_ctx, &device->isl_dev.has_bit6_swizzling,
318 sizeof(device->isl_dev.has_bit6_swizzling));
319 _mesa_sha1_final(&sha1_ctx, sha1);
320 memcpy(device->device_uuid, sha1, VK_UUID_SIZE);
321
322 return VK_SUCCESS;
323 }
324
325 static void
326 anv_physical_device_init_disk_cache(struct anv_physical_device *device)
327 {
328 #ifdef ENABLE_SHADER_CACHE
329 char renderer[10];
330 ASSERTED int len = snprintf(renderer, sizeof(renderer), "anv_%04x",
331 device->chipset_id);
332 assert(len == sizeof(renderer) - 2);
333
334 char timestamp[41];
335 _mesa_sha1_format(timestamp, device->driver_build_sha1);
336
337 const uint64_t driver_flags =
338 brw_get_compiler_config_value(device->compiler);
339 device->disk_cache = disk_cache_create(renderer, timestamp, driver_flags);
340 #else
341 device->disk_cache = NULL;
342 #endif
343 }
344
345 static void
346 anv_physical_device_free_disk_cache(struct anv_physical_device *device)
347 {
348 #ifdef ENABLE_SHADER_CACHE
349 if (device->disk_cache)
350 disk_cache_destroy(device->disk_cache);
351 #else
352 assert(device->disk_cache == NULL);
353 #endif
354 }
355
356 static uint64_t
357 get_available_system_memory()
358 {
359 char *meminfo = os_read_file("/proc/meminfo");
360 if (!meminfo)
361 return 0;
362
363 char *str = strstr(meminfo, "MemAvailable:");
364 if (!str) {
365 free(meminfo);
366 return 0;
367 }
368
369 uint64_t kb_mem_available;
370 if (sscanf(str, "MemAvailable: %" PRIx64, &kb_mem_available) == 1) {
371 free(meminfo);
372 return kb_mem_available << 10;
373 }
374
375 free(meminfo);
376 return 0;
377 }
378
379 static VkResult
380 anv_physical_device_init(struct anv_physical_device *device,
381 struct anv_instance *instance,
382 drmDevicePtr drm_device)
383 {
384 const char *primary_path = drm_device->nodes[DRM_NODE_PRIMARY];
385 const char *path = drm_device->nodes[DRM_NODE_RENDER];
386 VkResult result;
387 int fd;
388 int master_fd = -1;
389
390 brw_process_intel_debug_variable();
391
392 fd = open(path, O_RDWR | O_CLOEXEC);
393 if (fd < 0)
394 return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
395
396 device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
397 device->instance = instance;
398
399 assert(strlen(path) < ARRAY_SIZE(device->path));
400 snprintf(device->path, ARRAY_SIZE(device->path), "%s", path);
401
402 if (!gen_get_device_info_from_fd(fd, &device->info)) {
403 result = vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
404 goto fail;
405 }
406 device->chipset_id = device->info.chipset_id;
407 device->no_hw = device->info.no_hw;
408
409 if (getenv("INTEL_NO_HW") != NULL)
410 device->no_hw = true;
411
412 device->pci_info.domain = drm_device->businfo.pci->domain;
413 device->pci_info.bus = drm_device->businfo.pci->bus;
414 device->pci_info.device = drm_device->businfo.pci->dev;
415 device->pci_info.function = drm_device->businfo.pci->func;
416
417 device->name = gen_get_device_name(device->chipset_id);
418
419 if (device->info.is_haswell) {
420 intel_logw("Haswell Vulkan support is incomplete");
421 } else if (device->info.gen == 7 && !device->info.is_baytrail) {
422 intel_logw("Ivy Bridge Vulkan support is incomplete");
423 } else if (device->info.gen == 7 && device->info.is_baytrail) {
424 intel_logw("Bay Trail Vulkan support is incomplete");
425 } else if (device->info.gen >= 8 && device->info.gen <= 11) {
426 /* Gen8-11 fully supported */
427 } else if (device->info.gen == 12) {
428 intel_logw("Vulkan is not yet fully supported on gen12");
429 } else {
430 result = vk_errorf(device->instance, device,
431 VK_ERROR_INCOMPATIBLE_DRIVER,
432 "Vulkan not yet supported on %s", device->name);
433 goto fail;
434 }
435
436 device->cmd_parser_version = -1;
437 if (device->info.gen == 7) {
438 device->cmd_parser_version =
439 anv_gem_get_param(fd, I915_PARAM_CMD_PARSER_VERSION);
440 if (device->cmd_parser_version == -1) {
441 result = vk_errorf(device->instance, device,
442 VK_ERROR_INITIALIZATION_FAILED,
443 "failed to get command parser version");
444 goto fail;
445 }
446 }
447
448 if (!anv_gem_get_param(fd, I915_PARAM_HAS_WAIT_TIMEOUT)) {
449 result = vk_errorf(device->instance, device,
450 VK_ERROR_INITIALIZATION_FAILED,
451 "kernel missing gem wait");
452 goto fail;
453 }
454
455 if (!anv_gem_get_param(fd, I915_PARAM_HAS_EXECBUF2)) {
456 result = vk_errorf(device->instance, device,
457 VK_ERROR_INITIALIZATION_FAILED,
458 "kernel missing execbuf2");
459 goto fail;
460 }
461
462 if (!device->info.has_llc &&
463 anv_gem_get_param(fd, I915_PARAM_MMAP_VERSION) < 1) {
464 result = vk_errorf(device->instance, device,
465 VK_ERROR_INITIALIZATION_FAILED,
466 "kernel missing wc mmap");
467 goto fail;
468 }
469
470 result = anv_physical_device_init_heaps(device, fd);
471 if (result != VK_SUCCESS)
472 goto fail;
473
474 device->has_exec_async = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_ASYNC);
475 device->has_exec_capture = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_CAPTURE);
476 device->has_exec_fence = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_FENCE);
477 device->has_syncobj = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_FENCE_ARRAY);
478 device->has_syncobj_wait = device->has_syncobj &&
479 anv_gem_supports_syncobj_wait(fd);
480 device->has_context_priority = anv_gem_has_context_priority(fd);
481
482 device->use_softpin = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_SOFTPIN)
483 && device->supports_48bit_addresses;
484
485 device->has_context_isolation =
486 anv_gem_get_param(fd, I915_PARAM_HAS_CONTEXT_ISOLATION);
487
488 device->always_use_bindless =
489 env_var_as_boolean("ANV_ALWAYS_BINDLESS", false);
490
491 /* We first got the A64 messages on broadwell and we can only use them if
492 * we can pass addresses directly into the shader which requires softpin.
493 */
494 device->has_a64_buffer_access = device->info.gen >= 8 &&
495 device->use_softpin;
496
497 /* We first get bindless image access on Skylake and we can only really do
498 * it if we don't have any relocations so we need softpin.
499 */
500 device->has_bindless_images = device->info.gen >= 9 &&
501 device->use_softpin;
502
503 /* We've had bindless samplers since Ivy Bridge (forever in Vulkan terms)
504 * because it's just a matter of setting the sampler address in the sample
505 * message header. However, we've not bothered to wire it up for vec4 so
506 * we leave it disabled on gen7.
507 */
508 device->has_bindless_samplers = device->info.gen >= 8;
509
510 device->has_mem_available = get_available_system_memory() != 0;
511
512 /* Starting with Gen10, the timestamp frequency of the command streamer may
513 * vary from one part to another. We can query the value from the kernel.
514 */
515 if (device->info.gen >= 10) {
516 int timestamp_frequency =
517 anv_gem_get_param(fd, I915_PARAM_CS_TIMESTAMP_FREQUENCY);
518
519 if (timestamp_frequency < 0)
520 intel_logw("Kernel 4.16-rc1+ required to properly query CS timestamp frequency");
521 else
522 device->info.timestamp_frequency = timestamp_frequency;
523 }
524
525 /* GENs prior to 8 do not support EU/Subslice info */
526 if (device->info.gen >= 8) {
527 device->subslice_total = anv_gem_get_param(fd, I915_PARAM_SUBSLICE_TOTAL);
528 device->eu_total = anv_gem_get_param(fd, I915_PARAM_EU_TOTAL);
529
530 /* Without this information, we cannot get the right Braswell
531 * brandstrings, and we have to use conservative numbers for GPGPU on
532 * many platforms, but otherwise, things will just work.
533 */
534 if (device->subslice_total < 1 || device->eu_total < 1) {
535 intel_logw("Kernel 4.1 required to properly query GPU properties");
536 }
537 } else if (device->info.gen == 7) {
538 device->subslice_total = 1 << (device->info.gt - 1);
539 }
540
541 if (device->info.is_cherryview &&
542 device->subslice_total > 0 && device->eu_total > 0) {
543 /* Logical CS threads = EUs per subslice * num threads per EU */
544 uint32_t max_cs_threads =
545 device->eu_total / device->subslice_total * device->info.num_thread_per_eu;
546
547 /* Fuse configurations may give more threads than expected, never less. */
548 if (max_cs_threads > device->info.max_cs_threads)
549 device->info.max_cs_threads = max_cs_threads;
550 }
551
552 device->compiler = brw_compiler_create(NULL, &device->info);
553 if (device->compiler == NULL) {
554 result = vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
555 goto fail;
556 }
557 device->compiler->shader_debug_log = compiler_debug_log;
558 device->compiler->shader_perf_log = compiler_perf_log;
559 device->compiler->supports_pull_constants = false;
560 device->compiler->constant_buffer_0_is_relative =
561 device->info.gen < 8 || !device->has_context_isolation;
562 device->compiler->supports_shader_constants = true;
563
564 /* Broadwell PRM says:
565 *
566 * "Before Gen8, there was a historical configuration control field to
567 * swizzle address bit[6] for in X/Y tiling modes. This was set in three
568 * different places: TILECTL[1:0], ARB_MODE[5:4], and
569 * DISP_ARB_CTL[14:13].
570 *
571 * For Gen8 and subsequent generations, the swizzle fields are all
572 * reserved, and the CPU's memory controller performs all address
573 * swizzling modifications."
574 */
575 bool swizzled =
576 device->info.gen < 8 && anv_gem_get_bit6_swizzle(fd, I915_TILING_X);
577
578 isl_device_init(&device->isl_dev, &device->info, swizzled);
579
580 result = anv_physical_device_init_uuids(device);
581 if (result != VK_SUCCESS)
582 goto fail;
583
584 anv_physical_device_init_disk_cache(device);
585
586 if (instance->enabled_extensions.KHR_display) {
587 master_fd = open(primary_path, O_RDWR | O_CLOEXEC);
588 if (master_fd >= 0) {
589 /* prod the device with a GETPARAM call which will fail if
590 * we don't have permission to even render on this device
591 */
592 if (anv_gem_get_param(master_fd, I915_PARAM_CHIPSET_ID) == 0) {
593 close(master_fd);
594 master_fd = -1;
595 }
596 }
597 }
598 device->master_fd = master_fd;
599
600 result = anv_init_wsi(device);
601 if (result != VK_SUCCESS) {
602 ralloc_free(device->compiler);
603 anv_physical_device_free_disk_cache(device);
604 goto fail;
605 }
606
607 device->perf = anv_get_perf(&device->info, fd);
608
609 anv_physical_device_get_supported_extensions(device,
610 &device->supported_extensions);
611
612
613 device->local_fd = fd;
614
615 return VK_SUCCESS;
616
617 fail:
618 close(fd);
619 if (master_fd != -1)
620 close(master_fd);
621 return result;
622 }
623
624 static void
625 anv_physical_device_finish(struct anv_physical_device *device)
626 {
627 anv_finish_wsi(device);
628 anv_physical_device_free_disk_cache(device);
629 ralloc_free(device->compiler);
630 ralloc_free(device->perf);
631 close(device->local_fd);
632 if (device->master_fd >= 0)
633 close(device->master_fd);
634 }
635
636 static void *
637 default_alloc_func(void *pUserData, size_t size, size_t align,
638 VkSystemAllocationScope allocationScope)
639 {
640 return malloc(size);
641 }
642
643 static void *
644 default_realloc_func(void *pUserData, void *pOriginal, size_t size,
645 size_t align, VkSystemAllocationScope allocationScope)
646 {
647 return realloc(pOriginal, size);
648 }
649
650 static void
651 default_free_func(void *pUserData, void *pMemory)
652 {
653 free(pMemory);
654 }
655
656 static const VkAllocationCallbacks default_alloc = {
657 .pUserData = NULL,
658 .pfnAllocation = default_alloc_func,
659 .pfnReallocation = default_realloc_func,
660 .pfnFree = default_free_func,
661 };
662
663 VkResult anv_EnumerateInstanceExtensionProperties(
664 const char* pLayerName,
665 uint32_t* pPropertyCount,
666 VkExtensionProperties* pProperties)
667 {
668 VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
669
670 for (int i = 0; i < ANV_INSTANCE_EXTENSION_COUNT; i++) {
671 if (anv_instance_extensions_supported.extensions[i]) {
672 vk_outarray_append(&out, prop) {
673 *prop = anv_instance_extensions[i];
674 }
675 }
676 }
677
678 return vk_outarray_status(&out);
679 }
680
681 VkResult anv_CreateInstance(
682 const VkInstanceCreateInfo* pCreateInfo,
683 const VkAllocationCallbacks* pAllocator,
684 VkInstance* pInstance)
685 {
686 struct anv_instance *instance;
687 VkResult result;
688
689 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO);
690
691 struct anv_instance_extension_table enabled_extensions = {};
692 for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
693 int idx;
694 for (idx = 0; idx < ANV_INSTANCE_EXTENSION_COUNT; idx++) {
695 if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
696 anv_instance_extensions[idx].extensionName) == 0)
697 break;
698 }
699
700 if (idx >= ANV_INSTANCE_EXTENSION_COUNT)
701 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
702
703 if (!anv_instance_extensions_supported.extensions[idx])
704 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
705
706 enabled_extensions.extensions[idx] = true;
707 }
708
709 instance = vk_alloc2(&default_alloc, pAllocator, sizeof(*instance), 8,
710 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
711 if (!instance)
712 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
713
714 instance->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
715
716 if (pAllocator)
717 instance->alloc = *pAllocator;
718 else
719 instance->alloc = default_alloc;
720
721 instance->app_info = (struct anv_app_info) { .api_version = 0 };
722 if (pCreateInfo->pApplicationInfo) {
723 const VkApplicationInfo *app = pCreateInfo->pApplicationInfo;
724
725 instance->app_info.app_name =
726 vk_strdup(&instance->alloc, app->pApplicationName,
727 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
728 instance->app_info.app_version = app->applicationVersion;
729
730 instance->app_info.engine_name =
731 vk_strdup(&instance->alloc, app->pEngineName,
732 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
733 instance->app_info.engine_version = app->engineVersion;
734
735 instance->app_info.api_version = app->apiVersion;
736 }
737
738 if (instance->app_info.api_version == 0)
739 instance->app_info.api_version = VK_API_VERSION_1_0;
740
741 instance->enabled_extensions = enabled_extensions;
742
743 for (unsigned i = 0; i < ARRAY_SIZE(instance->dispatch.entrypoints); i++) {
744 /* Vulkan requires that entrypoints for extensions which have not been
745 * enabled must not be advertised.
746 */
747 if (!anv_instance_entrypoint_is_enabled(i, instance->app_info.api_version,
748 &instance->enabled_extensions)) {
749 instance->dispatch.entrypoints[i] = NULL;
750 } else {
751 instance->dispatch.entrypoints[i] =
752 anv_instance_dispatch_table.entrypoints[i];
753 }
754 }
755
756 struct anv_physical_device *pdevice = &instance->physicalDevice;
757 for (unsigned i = 0; i < ARRAY_SIZE(pdevice->dispatch.entrypoints); i++) {
758 /* Vulkan requires that entrypoints for extensions which have not been
759 * enabled must not be advertised.
760 */
761 if (!anv_physical_device_entrypoint_is_enabled(i, instance->app_info.api_version,
762 &instance->enabled_extensions)) {
763 pdevice->dispatch.entrypoints[i] = NULL;
764 } else {
765 pdevice->dispatch.entrypoints[i] =
766 anv_physical_device_dispatch_table.entrypoints[i];
767 }
768 }
769
770 for (unsigned i = 0; i < ARRAY_SIZE(instance->device_dispatch.entrypoints); i++) {
771 /* Vulkan requires that entrypoints for extensions which have not been
772 * enabled must not be advertised.
773 */
774 if (!anv_device_entrypoint_is_enabled(i, instance->app_info.api_version,
775 &instance->enabled_extensions, NULL)) {
776 instance->device_dispatch.entrypoints[i] = NULL;
777 } else {
778 instance->device_dispatch.entrypoints[i] =
779 anv_device_dispatch_table.entrypoints[i];
780 }
781 }
782
783 instance->physicalDeviceCount = -1;
784
785 result = vk_debug_report_instance_init(&instance->debug_report_callbacks);
786 if (result != VK_SUCCESS) {
787 vk_free2(&default_alloc, pAllocator, instance);
788 return vk_error(result);
789 }
790
791 instance->pipeline_cache_enabled =
792 env_var_as_boolean("ANV_ENABLE_PIPELINE_CACHE", true);
793
794 glsl_type_singleton_init_or_ref();
795
796 VG(VALGRIND_CREATE_MEMPOOL(instance, 0, false));
797
798 driParseOptionInfo(&instance->available_dri_options, anv_dri_options_xml);
799 driParseConfigFiles(&instance->dri_options, &instance->available_dri_options,
800 0, "anv", NULL,
801 instance->app_info.engine_name,
802 instance->app_info.engine_version);
803
804 *pInstance = anv_instance_to_handle(instance);
805
806 return VK_SUCCESS;
807 }
808
809 void anv_DestroyInstance(
810 VkInstance _instance,
811 const VkAllocationCallbacks* pAllocator)
812 {
813 ANV_FROM_HANDLE(anv_instance, instance, _instance);
814
815 if (!instance)
816 return;
817
818 if (instance->physicalDeviceCount > 0) {
819 /* We support at most one physical device. */
820 assert(instance->physicalDeviceCount == 1);
821 anv_physical_device_finish(&instance->physicalDevice);
822 }
823
824 vk_free(&instance->alloc, (char *)instance->app_info.app_name);
825 vk_free(&instance->alloc, (char *)instance->app_info.engine_name);
826
827 VG(VALGRIND_DESTROY_MEMPOOL(instance));
828
829 vk_debug_report_instance_destroy(&instance->debug_report_callbacks);
830
831 glsl_type_singleton_decref();
832
833 driDestroyOptionCache(&instance->dri_options);
834 driDestroyOptionInfo(&instance->available_dri_options);
835
836 vk_free(&instance->alloc, instance);
837 }
838
839 static VkResult
840 anv_enumerate_devices(struct anv_instance *instance)
841 {
842 /* TODO: Check for more devices ? */
843 drmDevicePtr devices[8];
844 VkResult result = VK_ERROR_INCOMPATIBLE_DRIVER;
845 int max_devices;
846
847 instance->physicalDeviceCount = 0;
848
849 max_devices = drmGetDevices2(0, devices, ARRAY_SIZE(devices));
850 if (max_devices < 1)
851 return VK_ERROR_INCOMPATIBLE_DRIVER;
852
853 for (unsigned i = 0; i < (unsigned)max_devices; i++) {
854 if (devices[i]->available_nodes & 1 << DRM_NODE_RENDER &&
855 devices[i]->bustype == DRM_BUS_PCI &&
856 devices[i]->deviceinfo.pci->vendor_id == 0x8086) {
857
858 result = anv_physical_device_init(&instance->physicalDevice,
859 instance, devices[i]);
860 if (result != VK_ERROR_INCOMPATIBLE_DRIVER)
861 break;
862 }
863 }
864 drmFreeDevices(devices, max_devices);
865
866 if (result == VK_SUCCESS)
867 instance->physicalDeviceCount = 1;
868
869 return result;
870 }
871
872 static VkResult
873 anv_instance_ensure_physical_device(struct anv_instance *instance)
874 {
875 if (instance->physicalDeviceCount < 0) {
876 VkResult result = anv_enumerate_devices(instance);
877 if (result != VK_SUCCESS &&
878 result != VK_ERROR_INCOMPATIBLE_DRIVER)
879 return result;
880 }
881
882 return VK_SUCCESS;
883 }
884
885 VkResult anv_EnumeratePhysicalDevices(
886 VkInstance _instance,
887 uint32_t* pPhysicalDeviceCount,
888 VkPhysicalDevice* pPhysicalDevices)
889 {
890 ANV_FROM_HANDLE(anv_instance, instance, _instance);
891 VK_OUTARRAY_MAKE(out, pPhysicalDevices, pPhysicalDeviceCount);
892
893 VkResult result = anv_instance_ensure_physical_device(instance);
894 if (result != VK_SUCCESS)
895 return result;
896
897 if (instance->physicalDeviceCount == 0)
898 return VK_SUCCESS;
899
900 assert(instance->physicalDeviceCount == 1);
901 vk_outarray_append(&out, i) {
902 *i = anv_physical_device_to_handle(&instance->physicalDevice);
903 }
904
905 return vk_outarray_status(&out);
906 }
907
908 VkResult anv_EnumeratePhysicalDeviceGroups(
909 VkInstance _instance,
910 uint32_t* pPhysicalDeviceGroupCount,
911 VkPhysicalDeviceGroupProperties* pPhysicalDeviceGroupProperties)
912 {
913 ANV_FROM_HANDLE(anv_instance, instance, _instance);
914 VK_OUTARRAY_MAKE(out, pPhysicalDeviceGroupProperties,
915 pPhysicalDeviceGroupCount);
916
917 VkResult result = anv_instance_ensure_physical_device(instance);
918 if (result != VK_SUCCESS)
919 return result;
920
921 if (instance->physicalDeviceCount == 0)
922 return VK_SUCCESS;
923
924 assert(instance->physicalDeviceCount == 1);
925
926 vk_outarray_append(&out, p) {
927 p->physicalDeviceCount = 1;
928 memset(p->physicalDevices, 0, sizeof(p->physicalDevices));
929 p->physicalDevices[0] =
930 anv_physical_device_to_handle(&instance->physicalDevice);
931 p->subsetAllocation = false;
932
933 vk_foreach_struct(ext, p->pNext)
934 anv_debug_ignored_stype(ext->sType);
935 }
936
937 return vk_outarray_status(&out);
938 }
939
940 void anv_GetPhysicalDeviceFeatures(
941 VkPhysicalDevice physicalDevice,
942 VkPhysicalDeviceFeatures* pFeatures)
943 {
944 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
945
946 *pFeatures = (VkPhysicalDeviceFeatures) {
947 .robustBufferAccess = true,
948 .fullDrawIndexUint32 = true,
949 .imageCubeArray = true,
950 .independentBlend = true,
951 .geometryShader = true,
952 .tessellationShader = true,
953 .sampleRateShading = true,
954 .dualSrcBlend = true,
955 .logicOp = true,
956 .multiDrawIndirect = true,
957 .drawIndirectFirstInstance = true,
958 .depthClamp = true,
959 .depthBiasClamp = true,
960 .fillModeNonSolid = true,
961 .depthBounds = pdevice->info.gen >= 12,
962 .wideLines = true,
963 .largePoints = true,
964 .alphaToOne = true,
965 .multiViewport = true,
966 .samplerAnisotropy = true,
967 .textureCompressionETC2 = pdevice->info.gen >= 8 ||
968 pdevice->info.is_baytrail,
969 .textureCompressionASTC_LDR = pdevice->info.gen >= 9, /* FINISHME CHV */
970 .textureCompressionBC = true,
971 .occlusionQueryPrecise = true,
972 .pipelineStatisticsQuery = true,
973 .fragmentStoresAndAtomics = true,
974 .shaderTessellationAndGeometryPointSize = true,
975 .shaderImageGatherExtended = true,
976 .shaderStorageImageExtendedFormats = true,
977 .shaderStorageImageMultisample = false,
978 .shaderStorageImageReadWithoutFormat = false,
979 .shaderStorageImageWriteWithoutFormat = true,
980 .shaderUniformBufferArrayDynamicIndexing = true,
981 .shaderSampledImageArrayDynamicIndexing = true,
982 .shaderStorageBufferArrayDynamicIndexing = true,
983 .shaderStorageImageArrayDynamicIndexing = true,
984 .shaderClipDistance = true,
985 .shaderCullDistance = true,
986 .shaderFloat64 = pdevice->info.gen >= 8 &&
987 pdevice->info.has_64bit_types,
988 .shaderInt64 = pdevice->info.gen >= 8 &&
989 pdevice->info.has_64bit_types,
990 .shaderInt16 = pdevice->info.gen >= 8,
991 .shaderResourceMinLod = pdevice->info.gen >= 9,
992 .variableMultisampleRate = true,
993 .inheritedQueries = true,
994 };
995
996 /* We can't do image stores in vec4 shaders */
997 pFeatures->vertexPipelineStoresAndAtomics =
998 pdevice->compiler->scalar_stage[MESA_SHADER_VERTEX] &&
999 pdevice->compiler->scalar_stage[MESA_SHADER_GEOMETRY];
1000
1001 struct anv_app_info *app_info = &pdevice->instance->app_info;
1002
1003 /* The new DOOM and Wolfenstein games require depthBounds without
1004 * checking for it. They seem to run fine without it so just claim it's
1005 * there and accept the consequences.
1006 */
1007 if (app_info->engine_name && strcmp(app_info->engine_name, "idTech") == 0)
1008 pFeatures->depthBounds = true;
1009 }
1010
1011 void anv_GetPhysicalDeviceFeatures2(
1012 VkPhysicalDevice physicalDevice,
1013 VkPhysicalDeviceFeatures2* pFeatures)
1014 {
1015 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
1016 anv_GetPhysicalDeviceFeatures(physicalDevice, &pFeatures->features);
1017
1018 vk_foreach_struct(ext, pFeatures->pNext) {
1019 switch (ext->sType) {
1020 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES_KHR: {
1021 VkPhysicalDevice8BitStorageFeaturesKHR *features =
1022 (VkPhysicalDevice8BitStorageFeaturesKHR *)ext;
1023 features->storageBuffer8BitAccess = pdevice->info.gen >= 8;
1024 features->uniformAndStorageBuffer8BitAccess = pdevice->info.gen >= 8;
1025 features->storagePushConstant8 = pdevice->info.gen >= 8;
1026 break;
1027 }
1028
1029 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES: {
1030 VkPhysicalDevice16BitStorageFeatures *features =
1031 (VkPhysicalDevice16BitStorageFeatures *)ext;
1032 features->storageBuffer16BitAccess = pdevice->info.gen >= 8;
1033 features->uniformAndStorageBuffer16BitAccess = pdevice->info.gen >= 8;
1034 features->storagePushConstant16 = pdevice->info.gen >= 8;
1035 features->storageInputOutput16 = false;
1036 break;
1037 }
1038
1039 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_EXT: {
1040 VkPhysicalDeviceBufferDeviceAddressFeaturesEXT *features = (void *)ext;
1041 features->bufferDeviceAddress = pdevice->has_a64_buffer_access;
1042 features->bufferDeviceAddressCaptureReplay = false;
1043 features->bufferDeviceAddressMultiDevice = false;
1044 break;
1045 }
1046
1047 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COMPUTE_SHADER_DERIVATIVES_FEATURES_NV: {
1048 VkPhysicalDeviceComputeShaderDerivativesFeaturesNV *features =
1049 (VkPhysicalDeviceComputeShaderDerivativesFeaturesNV *)ext;
1050 features->computeDerivativeGroupQuads = true;
1051 features->computeDerivativeGroupLinear = true;
1052 break;
1053 }
1054
1055 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT: {
1056 VkPhysicalDeviceConditionalRenderingFeaturesEXT *features =
1057 (VkPhysicalDeviceConditionalRenderingFeaturesEXT*)ext;
1058 features->conditionalRendering = pdevice->info.gen >= 8 ||
1059 pdevice->info.is_haswell;
1060 features->inheritedConditionalRendering = pdevice->info.gen >= 8 ||
1061 pdevice->info.is_haswell;
1062 break;
1063 }
1064
1065 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_CLIP_ENABLE_FEATURES_EXT: {
1066 VkPhysicalDeviceDepthClipEnableFeaturesEXT *features =
1067 (VkPhysicalDeviceDepthClipEnableFeaturesEXT *)ext;
1068 features->depthClipEnable = true;
1069 break;
1070 }
1071
1072 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR: {
1073 VkPhysicalDeviceFloat16Int8FeaturesKHR *features = (void *)ext;
1074 features->shaderFloat16 = pdevice->info.gen >= 8;
1075 features->shaderInt8 = pdevice->info.gen >= 8;
1076 break;
1077 }
1078
1079 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FRAGMENT_SHADER_INTERLOCK_FEATURES_EXT: {
1080 VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT *features =
1081 (VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT *)ext;
1082 features->fragmentShaderSampleInterlock = pdevice->info.gen >= 9;
1083 features->fragmentShaderPixelInterlock = pdevice->info.gen >= 9;
1084 features->fragmentShaderShadingRateInterlock = false;
1085 break;
1086 }
1087
1088 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_QUERY_RESET_FEATURES_EXT: {
1089 VkPhysicalDeviceHostQueryResetFeaturesEXT *features =
1090 (VkPhysicalDeviceHostQueryResetFeaturesEXT *)ext;
1091 features->hostQueryReset = true;
1092 break;
1093 }
1094
1095 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT: {
1096 VkPhysicalDeviceDescriptorIndexingFeaturesEXT *features =
1097 (VkPhysicalDeviceDescriptorIndexingFeaturesEXT *)ext;
1098 features->shaderInputAttachmentArrayDynamicIndexing = false;
1099 features->shaderUniformTexelBufferArrayDynamicIndexing = true;
1100 features->shaderStorageTexelBufferArrayDynamicIndexing = true;
1101 features->shaderUniformBufferArrayNonUniformIndexing = false;
1102 features->shaderSampledImageArrayNonUniformIndexing = true;
1103 features->shaderStorageBufferArrayNonUniformIndexing = true;
1104 features->shaderStorageImageArrayNonUniformIndexing = true;
1105 features->shaderInputAttachmentArrayNonUniformIndexing = false;
1106 features->shaderUniformTexelBufferArrayNonUniformIndexing = true;
1107 features->shaderStorageTexelBufferArrayNonUniformIndexing = true;
1108 features->descriptorBindingUniformBufferUpdateAfterBind = false;
1109 features->descriptorBindingSampledImageUpdateAfterBind = true;
1110 features->descriptorBindingStorageImageUpdateAfterBind = true;
1111 features->descriptorBindingStorageBufferUpdateAfterBind = true;
1112 features->descriptorBindingUniformTexelBufferUpdateAfterBind = true;
1113 features->descriptorBindingStorageTexelBufferUpdateAfterBind = true;
1114 features->descriptorBindingUpdateUnusedWhilePending = true;
1115 features->descriptorBindingPartiallyBound = true;
1116 features->descriptorBindingVariableDescriptorCount = false;
1117 features->runtimeDescriptorArray = true;
1118 break;
1119 }
1120
1121 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INDEX_TYPE_UINT8_FEATURES_EXT: {
1122 VkPhysicalDeviceIndexTypeUint8FeaturesEXT *features =
1123 (VkPhysicalDeviceIndexTypeUint8FeaturesEXT *)ext;
1124 features->indexTypeUint8 = true;
1125 break;
1126 }
1127
1128 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_FEATURES_EXT: {
1129 VkPhysicalDeviceInlineUniformBlockFeaturesEXT *features =
1130 (VkPhysicalDeviceInlineUniformBlockFeaturesEXT *)ext;
1131 features->inlineUniformBlock = true;
1132 features->descriptorBindingInlineUniformBlockUpdateAfterBind = true;
1133 break;
1134 }
1135
1136 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_FEATURES_EXT: {
1137 VkPhysicalDeviceLineRasterizationFeaturesEXT *features =
1138 (VkPhysicalDeviceLineRasterizationFeaturesEXT *)ext;
1139 features->rectangularLines = true;
1140 features->bresenhamLines = true;
1141 features->smoothLines = true;
1142 features->stippledRectangularLines = false;
1143 features->stippledBresenhamLines = true;
1144 features->stippledSmoothLines = false;
1145 break;
1146 }
1147
1148 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES: {
1149 VkPhysicalDeviceMultiviewFeatures *features =
1150 (VkPhysicalDeviceMultiviewFeatures *)ext;
1151 features->multiview = true;
1152 features->multiviewGeometryShader = true;
1153 features->multiviewTessellationShader = true;
1154 break;
1155 }
1156
1157 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGELESS_FRAMEBUFFER_FEATURES_KHR: {
1158 VkPhysicalDeviceImagelessFramebufferFeaturesKHR *features =
1159 (VkPhysicalDeviceImagelessFramebufferFeaturesKHR *)ext;
1160 features->imagelessFramebuffer = true;
1161 break;
1162 }
1163
1164 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR: {
1165 VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR *features =
1166 (VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR *)ext;
1167 features->pipelineExecutableInfo = true;
1168 break;
1169 }
1170
1171 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES: {
1172 VkPhysicalDeviceProtectedMemoryFeatures *features = (void *)ext;
1173 features->protectedMemory = false;
1174 break;
1175 }
1176
1177 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES: {
1178 VkPhysicalDeviceSamplerYcbcrConversionFeatures *features =
1179 (VkPhysicalDeviceSamplerYcbcrConversionFeatures *) ext;
1180 features->samplerYcbcrConversion = true;
1181 break;
1182 }
1183
1184 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES_EXT: {
1185 VkPhysicalDeviceScalarBlockLayoutFeaturesEXT *features =
1186 (VkPhysicalDeviceScalarBlockLayoutFeaturesEXT *)ext;
1187 features->scalarBlockLayout = true;
1188 break;
1189 }
1190
1191 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_INT64_FEATURES_KHR: {
1192 VkPhysicalDeviceShaderAtomicInt64FeaturesKHR *features = (void *)ext;
1193 features->shaderBufferInt64Atomics =
1194 pdevice->info.gen >= 9 && pdevice->use_softpin;
1195 features->shaderSharedInt64Atomics = VK_FALSE;
1196 break;
1197 }
1198
1199 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DEMOTE_TO_HELPER_INVOCATION_FEATURES_EXT: {
1200 VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT *features = (void *)ext;
1201 features->shaderDemoteToHelperInvocation = true;
1202 break;
1203 }
1204
1205 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CLOCK_FEATURES_KHR: {
1206 VkPhysicalDeviceShaderClockFeaturesKHR *features =
1207 (VkPhysicalDeviceShaderClockFeaturesKHR *)ext;
1208 features->shaderSubgroupClock = true;
1209 features->shaderDeviceClock = false;
1210 break;
1211 }
1212
1213 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETERS_FEATURES: {
1214 VkPhysicalDeviceShaderDrawParametersFeatures *features = (void *)ext;
1215 features->shaderDrawParameters = true;
1216 break;
1217 }
1218
1219 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_SUBGROUP_EXTENDED_TYPES_FEATURES_KHR: {
1220 VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR *features =
1221 (VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR *)ext;
1222 features->shaderSubgroupExtendedTypes = true;
1223 break;
1224 }
1225
1226 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_FEATURES_EXT: {
1227 VkPhysicalDeviceSubgroupSizeControlFeaturesEXT *features =
1228 (VkPhysicalDeviceSubgroupSizeControlFeaturesEXT *)ext;
1229 features->subgroupSizeControl = true;
1230 features->computeFullSubgroups = true;
1231 break;
1232 }
1233
1234 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_FEATURES_EXT: {
1235 VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT *features =
1236 (VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT *)ext;
1237 features->texelBufferAlignment = true;
1238 break;
1239 }
1240
1241 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES: {
1242 VkPhysicalDeviceVariablePointersFeatures *features = (void *)ext;
1243 features->variablePointersStorageBuffer = true;
1244 features->variablePointers = true;
1245 break;
1246 }
1247
1248 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT: {
1249 VkPhysicalDeviceTransformFeedbackFeaturesEXT *features =
1250 (VkPhysicalDeviceTransformFeedbackFeaturesEXT *)ext;
1251 features->transformFeedback = true;
1252 features->geometryStreams = true;
1253 break;
1254 }
1255
1256 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_UNIFORM_BUFFER_STANDARD_LAYOUT_FEATURES_KHR: {
1257 VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR *features =
1258 (VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR *)ext;
1259 features->uniformBufferStandardLayout = true;
1260 break;
1261 }
1262
1263 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT: {
1264 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *features =
1265 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *)ext;
1266 features->vertexAttributeInstanceRateDivisor = true;
1267 features->vertexAttributeInstanceRateZeroDivisor = true;
1268 break;
1269 }
1270
1271 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_MEMORY_MODEL_FEATURES_KHR: {
1272 VkPhysicalDeviceVulkanMemoryModelFeaturesKHR *features = (void *)ext;
1273 features->vulkanMemoryModel = true;
1274 features->vulkanMemoryModelDeviceScope = true;
1275 features->vulkanMemoryModelAvailabilityVisibilityChains = true;
1276 break;
1277 }
1278
1279 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_YCBCR_IMAGE_ARRAYS_FEATURES_EXT: {
1280 VkPhysicalDeviceYcbcrImageArraysFeaturesEXT *features =
1281 (VkPhysicalDeviceYcbcrImageArraysFeaturesEXT *)ext;
1282 features->ycbcrImageArrays = true;
1283 break;
1284 }
1285
1286 default:
1287 anv_debug_ignored_stype(ext->sType);
1288 break;
1289 }
1290 }
1291 }
1292
1293 #define MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS 64
1294
1295 #define MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS 64
1296 #define MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS 256
1297
1298 void anv_GetPhysicalDeviceProperties(
1299 VkPhysicalDevice physicalDevice,
1300 VkPhysicalDeviceProperties* pProperties)
1301 {
1302 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
1303 const struct gen_device_info *devinfo = &pdevice->info;
1304
1305 /* See assertions made when programming the buffer surface state. */
1306 const uint32_t max_raw_buffer_sz = devinfo->gen >= 7 ?
1307 (1ul << 30) : (1ul << 27);
1308
1309 const uint32_t max_ssbos = pdevice->has_a64_buffer_access ? UINT16_MAX : 64;
1310 const uint32_t max_textures =
1311 pdevice->has_bindless_images ? UINT16_MAX : 128;
1312 const uint32_t max_samplers =
1313 pdevice->has_bindless_samplers ? UINT16_MAX :
1314 (devinfo->gen >= 8 || devinfo->is_haswell) ? 128 : 16;
1315 const uint32_t max_images =
1316 pdevice->has_bindless_images ? UINT16_MAX : MAX_IMAGES;
1317
1318 /* If we can use bindless for everything, claim a high per-stage limit,
1319 * otherwise use the binding table size, minus the slots reserved for
1320 * render targets and one slot for the descriptor buffer. */
1321 const uint32_t max_per_stage =
1322 pdevice->has_bindless_images && pdevice->has_a64_buffer_access
1323 ? UINT32_MAX : MAX_BINDING_TABLE_SIZE - MAX_RTS - 1;
1324
1325 const uint32_t max_workgroup_size = 32 * devinfo->max_cs_threads;
1326
1327 VkSampleCountFlags sample_counts =
1328 isl_device_get_sample_counts(&pdevice->isl_dev);
1329
1330
1331 VkPhysicalDeviceLimits limits = {
1332 .maxImageDimension1D = (1 << 14),
1333 .maxImageDimension2D = (1 << 14),
1334 .maxImageDimension3D = (1 << 11),
1335 .maxImageDimensionCube = (1 << 14),
1336 .maxImageArrayLayers = (1 << 11),
1337 .maxTexelBufferElements = 128 * 1024 * 1024,
1338 .maxUniformBufferRange = (1ul << 27),
1339 .maxStorageBufferRange = max_raw_buffer_sz,
1340 .maxPushConstantsSize = MAX_PUSH_CONSTANTS_SIZE,
1341 .maxMemoryAllocationCount = UINT32_MAX,
1342 .maxSamplerAllocationCount = 64 * 1024,
1343 .bufferImageGranularity = 64, /* A cache line */
1344 .sparseAddressSpaceSize = 0,
1345 .maxBoundDescriptorSets = MAX_SETS,
1346 .maxPerStageDescriptorSamplers = max_samplers,
1347 .maxPerStageDescriptorUniformBuffers = MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS,
1348 .maxPerStageDescriptorStorageBuffers = max_ssbos,
1349 .maxPerStageDescriptorSampledImages = max_textures,
1350 .maxPerStageDescriptorStorageImages = max_images,
1351 .maxPerStageDescriptorInputAttachments = MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS,
1352 .maxPerStageResources = max_per_stage,
1353 .maxDescriptorSetSamplers = 6 * max_samplers, /* number of stages * maxPerStageDescriptorSamplers */
1354 .maxDescriptorSetUniformBuffers = 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS, /* number of stages * maxPerStageDescriptorUniformBuffers */
1355 .maxDescriptorSetUniformBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2,
1356 .maxDescriptorSetStorageBuffers = 6 * max_ssbos, /* number of stages * maxPerStageDescriptorStorageBuffers */
1357 .maxDescriptorSetStorageBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2,
1358 .maxDescriptorSetSampledImages = 6 * max_textures, /* number of stages * maxPerStageDescriptorSampledImages */
1359 .maxDescriptorSetStorageImages = 6 * max_images, /* number of stages * maxPerStageDescriptorStorageImages */
1360 .maxDescriptorSetInputAttachments = MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS,
1361 .maxVertexInputAttributes = MAX_VBS,
1362 .maxVertexInputBindings = MAX_VBS,
1363 .maxVertexInputAttributeOffset = 2047,
1364 .maxVertexInputBindingStride = 2048,
1365 .maxVertexOutputComponents = 128,
1366 .maxTessellationGenerationLevel = 64,
1367 .maxTessellationPatchSize = 32,
1368 .maxTessellationControlPerVertexInputComponents = 128,
1369 .maxTessellationControlPerVertexOutputComponents = 128,
1370 .maxTessellationControlPerPatchOutputComponents = 128,
1371 .maxTessellationControlTotalOutputComponents = 2048,
1372 .maxTessellationEvaluationInputComponents = 128,
1373 .maxTessellationEvaluationOutputComponents = 128,
1374 .maxGeometryShaderInvocations = 32,
1375 .maxGeometryInputComponents = 64,
1376 .maxGeometryOutputComponents = 128,
1377 .maxGeometryOutputVertices = 256,
1378 .maxGeometryTotalOutputComponents = 1024,
1379 .maxFragmentInputComponents = 116, /* 128 components - (PSIZ, CLIP_DIST0, CLIP_DIST1) */
1380 .maxFragmentOutputAttachments = 8,
1381 .maxFragmentDualSrcAttachments = 1,
1382 .maxFragmentCombinedOutputResources = 8,
1383 .maxComputeSharedMemorySize = 64 * 1024,
1384 .maxComputeWorkGroupCount = { 65535, 65535, 65535 },
1385 .maxComputeWorkGroupInvocations = max_workgroup_size,
1386 .maxComputeWorkGroupSize = {
1387 max_workgroup_size,
1388 max_workgroup_size,
1389 max_workgroup_size,
1390 },
1391 .subPixelPrecisionBits = 8,
1392 .subTexelPrecisionBits = 8,
1393 .mipmapPrecisionBits = 8,
1394 .maxDrawIndexedIndexValue = UINT32_MAX,
1395 .maxDrawIndirectCount = UINT32_MAX,
1396 .maxSamplerLodBias = 16,
1397 .maxSamplerAnisotropy = 16,
1398 .maxViewports = MAX_VIEWPORTS,
1399 .maxViewportDimensions = { (1 << 14), (1 << 14) },
1400 .viewportBoundsRange = { INT16_MIN, INT16_MAX },
1401 .viewportSubPixelBits = 13, /* We take a float? */
1402 .minMemoryMapAlignment = 4096, /* A page */
1403 /* The dataport requires texel alignment so we need to assume a worst
1404 * case of R32G32B32A32 which is 16 bytes.
1405 */
1406 .minTexelBufferOffsetAlignment = 16,
1407 /* We need 16 for UBO block reads to work and 32 for push UBOs */
1408 .minUniformBufferOffsetAlignment = 32,
1409 .minStorageBufferOffsetAlignment = 4,
1410 .minTexelOffset = -8,
1411 .maxTexelOffset = 7,
1412 .minTexelGatherOffset = -32,
1413 .maxTexelGatherOffset = 31,
1414 .minInterpolationOffset = -0.5,
1415 .maxInterpolationOffset = 0.4375,
1416 .subPixelInterpolationOffsetBits = 4,
1417 .maxFramebufferWidth = (1 << 14),
1418 .maxFramebufferHeight = (1 << 14),
1419 .maxFramebufferLayers = (1 << 11),
1420 .framebufferColorSampleCounts = sample_counts,
1421 .framebufferDepthSampleCounts = sample_counts,
1422 .framebufferStencilSampleCounts = sample_counts,
1423 .framebufferNoAttachmentsSampleCounts = sample_counts,
1424 .maxColorAttachments = MAX_RTS,
1425 .sampledImageColorSampleCounts = sample_counts,
1426 .sampledImageIntegerSampleCounts = VK_SAMPLE_COUNT_1_BIT,
1427 .sampledImageDepthSampleCounts = sample_counts,
1428 .sampledImageStencilSampleCounts = sample_counts,
1429 .storageImageSampleCounts = VK_SAMPLE_COUNT_1_BIT,
1430 .maxSampleMaskWords = 1,
1431 .timestampComputeAndGraphics = true,
1432 .timestampPeriod = 1000000000.0 / devinfo->timestamp_frequency,
1433 .maxClipDistances = 8,
1434 .maxCullDistances = 8,
1435 .maxCombinedClipAndCullDistances = 8,
1436 .discreteQueuePriorities = 2,
1437 .pointSizeRange = { 0.125, 255.875 },
1438 .lineWidthRange = {
1439 0.0,
1440 (devinfo->gen >= 9 || devinfo->is_cherryview) ?
1441 2047.9921875 : 7.9921875,
1442 },
1443 .pointSizeGranularity = (1.0 / 8.0),
1444 .lineWidthGranularity = (1.0 / 128.0),
1445 .strictLines = false,
1446 .standardSampleLocations = true,
1447 .optimalBufferCopyOffsetAlignment = 128,
1448 .optimalBufferCopyRowPitchAlignment = 128,
1449 .nonCoherentAtomSize = 64,
1450 };
1451
1452 *pProperties = (VkPhysicalDeviceProperties) {
1453 .apiVersion = anv_physical_device_api_version(pdevice),
1454 .driverVersion = vk_get_driver_version(),
1455 .vendorID = 0x8086,
1456 .deviceID = pdevice->chipset_id,
1457 .deviceType = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
1458 .limits = limits,
1459 .sparseProperties = {0}, /* Broadwell doesn't do sparse. */
1460 };
1461
1462 snprintf(pProperties->deviceName, sizeof(pProperties->deviceName),
1463 "%s", pdevice->name);
1464 memcpy(pProperties->pipelineCacheUUID,
1465 pdevice->pipeline_cache_uuid, VK_UUID_SIZE);
1466 }
1467
1468 void anv_GetPhysicalDeviceProperties2(
1469 VkPhysicalDevice physicalDevice,
1470 VkPhysicalDeviceProperties2* pProperties)
1471 {
1472 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
1473
1474 anv_GetPhysicalDeviceProperties(physicalDevice, &pProperties->properties);
1475
1476 vk_foreach_struct(ext, pProperties->pNext) {
1477 switch (ext->sType) {
1478 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_STENCIL_RESOLVE_PROPERTIES_KHR: {
1479 VkPhysicalDeviceDepthStencilResolvePropertiesKHR *props =
1480 (VkPhysicalDeviceDepthStencilResolvePropertiesKHR *)ext;
1481
1482 /* We support all of the depth resolve modes */
1483 props->supportedDepthResolveModes =
1484 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR |
1485 VK_RESOLVE_MODE_AVERAGE_BIT_KHR |
1486 VK_RESOLVE_MODE_MIN_BIT_KHR |
1487 VK_RESOLVE_MODE_MAX_BIT_KHR;
1488
1489 /* Average doesn't make sense for stencil so we don't support that */
1490 props->supportedStencilResolveModes =
1491 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR;
1492 if (pdevice->info.gen >= 8) {
1493 /* The advanced stencil resolve modes currently require stencil
1494 * sampling be supported by the hardware.
1495 */
1496 props->supportedStencilResolveModes |=
1497 VK_RESOLVE_MODE_MIN_BIT_KHR |
1498 VK_RESOLVE_MODE_MAX_BIT_KHR;
1499 }
1500
1501 props->independentResolveNone = true;
1502 props->independentResolve = true;
1503 break;
1504 }
1505
1506 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES_EXT: {
1507 VkPhysicalDeviceDescriptorIndexingPropertiesEXT *props =
1508 (VkPhysicalDeviceDescriptorIndexingPropertiesEXT *)ext;
1509
1510 /* It's a bit hard to exactly map our implementation to the limits
1511 * described here. The bindless surface handle in the extended
1512 * message descriptors is 20 bits and it's an index into the table of
1513 * RENDER_SURFACE_STATE structs that starts at bindless surface base
1514 * address. Given that most things consume two surface states per
1515 * view (general/sampled for textures and write-only/read-write for
1516 * images), we claim 2^19 things.
1517 *
1518 * For SSBOs, we just use A64 messages so there is no real limit
1519 * there beyond the limit on the total size of a descriptor set.
1520 */
1521 const unsigned max_bindless_views = 1 << 19;
1522
1523 props->maxUpdateAfterBindDescriptorsInAllPools = max_bindless_views;
1524 props->shaderUniformBufferArrayNonUniformIndexingNative = false;
1525 props->shaderSampledImageArrayNonUniformIndexingNative = false;
1526 props->shaderStorageBufferArrayNonUniformIndexingNative = true;
1527 props->shaderStorageImageArrayNonUniformIndexingNative = false;
1528 props->shaderInputAttachmentArrayNonUniformIndexingNative = false;
1529 props->robustBufferAccessUpdateAfterBind = true;
1530 props->quadDivergentImplicitLod = false;
1531 props->maxPerStageDescriptorUpdateAfterBindSamplers = max_bindless_views;
1532 props->maxPerStageDescriptorUpdateAfterBindUniformBuffers = MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS;
1533 props->maxPerStageDescriptorUpdateAfterBindStorageBuffers = UINT32_MAX;
1534 props->maxPerStageDescriptorUpdateAfterBindSampledImages = max_bindless_views;
1535 props->maxPerStageDescriptorUpdateAfterBindStorageImages = max_bindless_views;
1536 props->maxPerStageDescriptorUpdateAfterBindInputAttachments = MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS;
1537 props->maxPerStageUpdateAfterBindResources = UINT32_MAX;
1538 props->maxDescriptorSetUpdateAfterBindSamplers = max_bindless_views;
1539 props->maxDescriptorSetUpdateAfterBindUniformBuffers = 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS;
1540 props->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2;
1541 props->maxDescriptorSetUpdateAfterBindStorageBuffers = UINT32_MAX;
1542 props->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2;
1543 props->maxDescriptorSetUpdateAfterBindSampledImages = max_bindless_views;
1544 props->maxDescriptorSetUpdateAfterBindStorageImages = max_bindless_views;
1545 props->maxDescriptorSetUpdateAfterBindInputAttachments = MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS;
1546 break;
1547 }
1548
1549 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR: {
1550 VkPhysicalDeviceDriverPropertiesKHR *driver_props =
1551 (VkPhysicalDeviceDriverPropertiesKHR *) ext;
1552
1553 driver_props->driverID = VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA_KHR;
1554 snprintf(driver_props->driverName, VK_MAX_DRIVER_NAME_SIZE_KHR,
1555 "Intel open-source Mesa driver");
1556
1557 snprintf(driver_props->driverInfo, VK_MAX_DRIVER_INFO_SIZE_KHR,
1558 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1);
1559
1560 driver_props->conformanceVersion = (VkConformanceVersionKHR) {
1561 .major = 1,
1562 .minor = 1,
1563 .subminor = 2,
1564 .patch = 0,
1565 };
1566 break;
1567 }
1568
1569 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT: {
1570 VkPhysicalDeviceExternalMemoryHostPropertiesEXT *props =
1571 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT *) ext;
1572 /* Userptr needs page aligned memory. */
1573 props->minImportedHostPointerAlignment = 4096;
1574 break;
1575 }
1576
1577 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES: {
1578 VkPhysicalDeviceIDProperties *id_props =
1579 (VkPhysicalDeviceIDProperties *)ext;
1580 memcpy(id_props->deviceUUID, pdevice->device_uuid, VK_UUID_SIZE);
1581 memcpy(id_props->driverUUID, pdevice->driver_uuid, VK_UUID_SIZE);
1582 /* The LUID is for Windows. */
1583 id_props->deviceLUIDValid = false;
1584 break;
1585 }
1586
1587 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_PROPERTIES_EXT: {
1588 VkPhysicalDeviceInlineUniformBlockPropertiesEXT *props =
1589 (VkPhysicalDeviceInlineUniformBlockPropertiesEXT *)ext;
1590 props->maxInlineUniformBlockSize = MAX_INLINE_UNIFORM_BLOCK_SIZE;
1591 props->maxPerStageDescriptorInlineUniformBlocks =
1592 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1593 props->maxPerStageDescriptorUpdateAfterBindInlineUniformBlocks =
1594 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1595 props->maxDescriptorSetInlineUniformBlocks =
1596 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1597 props->maxDescriptorSetUpdateAfterBindInlineUniformBlocks =
1598 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
1599 break;
1600 }
1601
1602 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_PROPERTIES_EXT: {
1603 VkPhysicalDeviceLineRasterizationPropertiesEXT *props =
1604 (VkPhysicalDeviceLineRasterizationPropertiesEXT *)ext;
1605 /* In the Skylake PRM Vol. 7, subsection titled "GIQ (Diamond)
1606 * Sampling Rules - Legacy Mode", it says the following:
1607 *
1608 * "Note that the device divides a pixel into a 16x16 array of
1609 * subpixels, referenced by their upper left corners."
1610 *
1611 * This is the only known reference in the PRMs to the subpixel
1612 * precision of line rasterization and a "16x16 array of subpixels"
1613 * implies 4 subpixel precision bits. Empirical testing has shown
1614 * that 4 subpixel precision bits applies to all line rasterization
1615 * types.
1616 */
1617 props->lineSubPixelPrecisionBits = 4;
1618 break;
1619 }
1620
1621 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES: {
1622 VkPhysicalDeviceMaintenance3Properties *props =
1623 (VkPhysicalDeviceMaintenance3Properties *)ext;
1624 /* This value doesn't matter for us today as our per-stage
1625 * descriptors are the real limit.
1626 */
1627 props->maxPerSetDescriptors = 1024;
1628 props->maxMemoryAllocationSize = MAX_MEMORY_ALLOCATION_SIZE;
1629 break;
1630 }
1631
1632 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES: {
1633 VkPhysicalDeviceMultiviewProperties *properties =
1634 (VkPhysicalDeviceMultiviewProperties *)ext;
1635 properties->maxMultiviewViewCount = 16;
1636 properties->maxMultiviewInstanceIndex = UINT32_MAX / 16;
1637 break;
1638 }
1639
1640 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT: {
1641 VkPhysicalDevicePCIBusInfoPropertiesEXT *properties =
1642 (VkPhysicalDevicePCIBusInfoPropertiesEXT *)ext;
1643 properties->pciDomain = pdevice->pci_info.domain;
1644 properties->pciBus = pdevice->pci_info.bus;
1645 properties->pciDevice = pdevice->pci_info.device;
1646 properties->pciFunction = pdevice->pci_info.function;
1647 break;
1648 }
1649
1650 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES: {
1651 VkPhysicalDevicePointClippingProperties *properties =
1652 (VkPhysicalDevicePointClippingProperties *) ext;
1653 properties->pointClippingBehavior = VK_POINT_CLIPPING_BEHAVIOR_USER_CLIP_PLANES_ONLY;
1654 break;
1655 }
1656
1657 #pragma GCC diagnostic push
1658 #pragma GCC diagnostic ignored "-Wswitch"
1659 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PRESENTATION_PROPERTIES_ANDROID: {
1660 VkPhysicalDevicePresentationPropertiesANDROID *props =
1661 (VkPhysicalDevicePresentationPropertiesANDROID *)ext;
1662 props->sharedImage = VK_FALSE;
1663 break;
1664 }
1665 #pragma GCC diagnostic pop
1666
1667 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES: {
1668 VkPhysicalDeviceProtectedMemoryProperties *props =
1669 (VkPhysicalDeviceProtectedMemoryProperties *)ext;
1670 props->protectedNoFault = false;
1671 break;
1672 }
1673
1674 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR: {
1675 VkPhysicalDevicePushDescriptorPropertiesKHR *properties =
1676 (VkPhysicalDevicePushDescriptorPropertiesKHR *) ext;
1677
1678 properties->maxPushDescriptors = MAX_PUSH_DESCRIPTORS;
1679 break;
1680 }
1681
1682 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT: {
1683 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT *properties =
1684 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT *)ext;
1685 properties->filterMinmaxImageComponentMapping = pdevice->info.gen >= 9;
1686 properties->filterMinmaxSingleComponentFormats = true;
1687 break;
1688 }
1689
1690 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES: {
1691 VkPhysicalDeviceSubgroupProperties *properties = (void *)ext;
1692
1693 properties->subgroupSize = BRW_SUBGROUP_SIZE;
1694
1695 VkShaderStageFlags scalar_stages = 0;
1696 for (unsigned stage = 0; stage < MESA_SHADER_STAGES; stage++) {
1697 if (pdevice->compiler->scalar_stage[stage])
1698 scalar_stages |= mesa_to_vk_shader_stage(stage);
1699 }
1700 properties->supportedStages = scalar_stages;
1701
1702 properties->supportedOperations = VK_SUBGROUP_FEATURE_BASIC_BIT |
1703 VK_SUBGROUP_FEATURE_VOTE_BIT |
1704 VK_SUBGROUP_FEATURE_BALLOT_BIT |
1705 VK_SUBGROUP_FEATURE_SHUFFLE_BIT |
1706 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT |
1707 VK_SUBGROUP_FEATURE_QUAD_BIT;
1708 if (pdevice->info.gen >= 8) {
1709 /* TODO: There's no technical reason why these can't be made to
1710 * work on gen7 but they don't at the moment so it's best to leave
1711 * the feature disabled than enabled and broken.
1712 */
1713 properties->supportedOperations |=
1714 VK_SUBGROUP_FEATURE_ARITHMETIC_BIT |
1715 VK_SUBGROUP_FEATURE_CLUSTERED_BIT;
1716 }
1717 properties->quadOperationsInAllStages = pdevice->info.gen >= 8;
1718 break;
1719 }
1720
1721 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_PROPERTIES_EXT: {
1722 VkPhysicalDeviceSubgroupSizeControlPropertiesEXT *props =
1723 (VkPhysicalDeviceSubgroupSizeControlPropertiesEXT *)ext;
1724 STATIC_ASSERT(8 <= BRW_SUBGROUP_SIZE && BRW_SUBGROUP_SIZE <= 32);
1725 props->minSubgroupSize = 8;
1726 props->maxSubgroupSize = 32;
1727 props->maxComputeWorkgroupSubgroups = pdevice->info.max_cs_threads;
1728 props->requiredSubgroupSizeStages = VK_SHADER_STAGE_COMPUTE_BIT;
1729 break;
1730 }
1731 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT_CONTROLS_PROPERTIES_KHR : {
1732 VkPhysicalDeviceFloatControlsPropertiesKHR *properties = (void *)ext;
1733 properties->denormBehaviorIndependence = VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_ALL_KHR;
1734 properties->roundingModeIndependence = VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_NONE_KHR;
1735
1736 /* Broadwell does not support HF denorms and there are restrictions
1737 * other gens. According to Kabylake's PRM:
1738 *
1739 * "math - Extended Math Function
1740 * [...]
1741 * Restriction : Half-float denorms are always retained."
1742 */
1743 properties->shaderDenormFlushToZeroFloat16 = false;
1744 properties->shaderDenormPreserveFloat16 = pdevice->info.gen > 8;
1745 properties->shaderRoundingModeRTEFloat16 = true;
1746 properties->shaderRoundingModeRTZFloat16 = true;
1747 properties->shaderSignedZeroInfNanPreserveFloat16 = true;
1748
1749 properties->shaderDenormFlushToZeroFloat32 = true;
1750 properties->shaderDenormPreserveFloat32 = true;
1751 properties->shaderRoundingModeRTEFloat32 = true;
1752 properties->shaderRoundingModeRTZFloat32 = true;
1753 properties->shaderSignedZeroInfNanPreserveFloat32 = true;
1754
1755 properties->shaderDenormFlushToZeroFloat64 = true;
1756 properties->shaderDenormPreserveFloat64 = true;
1757 properties->shaderRoundingModeRTEFloat64 = true;
1758 properties->shaderRoundingModeRTZFloat64 = true;
1759 properties->shaderSignedZeroInfNanPreserveFloat64 = true;
1760 break;
1761 }
1762
1763 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_PROPERTIES_EXT: {
1764 VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT *props =
1765 (VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT *)ext;
1766
1767 /* From the SKL PRM Vol. 2d, docs for RENDER_SURFACE_STATE::Surface
1768 * Base Address:
1769 *
1770 * "For SURFTYPE_BUFFER non-rendertarget surfaces, this field
1771 * specifies the base address of the first element of the surface,
1772 * computed in software by adding the surface base address to the
1773 * byte offset of the element in the buffer. The base address must
1774 * be aligned to element size."
1775 *
1776 * The typed dataport messages require that things be texel aligned.
1777 * Otherwise, we may just load/store the wrong data or, in the worst
1778 * case, there may be hangs.
1779 */
1780 props->storageTexelBufferOffsetAlignmentBytes = 16;
1781 props->storageTexelBufferOffsetSingleTexelAlignment = true;
1782
1783 /* The sampler, however, is much more forgiving and it can handle
1784 * arbitrary byte alignment for linear and buffer surfaces. It's
1785 * hard to find a good PRM citation for this but years of empirical
1786 * experience demonstrate that this is true.
1787 */
1788 props->uniformTexelBufferOffsetAlignmentBytes = 1;
1789 props->uniformTexelBufferOffsetSingleTexelAlignment = false;
1790 break;
1791 }
1792
1793 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT: {
1794 VkPhysicalDeviceTransformFeedbackPropertiesEXT *props =
1795 (VkPhysicalDeviceTransformFeedbackPropertiesEXT *)ext;
1796
1797 props->maxTransformFeedbackStreams = MAX_XFB_STREAMS;
1798 props->maxTransformFeedbackBuffers = MAX_XFB_BUFFERS;
1799 props->maxTransformFeedbackBufferSize = (1ull << 32);
1800 props->maxTransformFeedbackStreamDataSize = 128 * 4;
1801 props->maxTransformFeedbackBufferDataSize = 128 * 4;
1802 props->maxTransformFeedbackBufferDataStride = 2048;
1803 props->transformFeedbackQueries = true;
1804 props->transformFeedbackStreamsLinesTriangles = false;
1805 props->transformFeedbackRasterizationStreamSelect = false;
1806 props->transformFeedbackDraw = true;
1807 break;
1808 }
1809
1810 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT: {
1811 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *props =
1812 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *)ext;
1813 /* We have to restrict this a bit for multiview */
1814 props->maxVertexAttribDivisor = UINT32_MAX / 16;
1815 break;
1816 }
1817
1818 default:
1819 anv_debug_ignored_stype(ext->sType);
1820 break;
1821 }
1822 }
1823 }
1824
1825 /* We support exactly one queue family. */
1826 static const VkQueueFamilyProperties
1827 anv_queue_family_properties = {
1828 .queueFlags = VK_QUEUE_GRAPHICS_BIT |
1829 VK_QUEUE_COMPUTE_BIT |
1830 VK_QUEUE_TRANSFER_BIT,
1831 .queueCount = 1,
1832 .timestampValidBits = 36, /* XXX: Real value here */
1833 .minImageTransferGranularity = { 1, 1, 1 },
1834 };
1835
1836 void anv_GetPhysicalDeviceQueueFamilyProperties(
1837 VkPhysicalDevice physicalDevice,
1838 uint32_t* pCount,
1839 VkQueueFamilyProperties* pQueueFamilyProperties)
1840 {
1841 VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pCount);
1842
1843 vk_outarray_append(&out, p) {
1844 *p = anv_queue_family_properties;
1845 }
1846 }
1847
1848 void anv_GetPhysicalDeviceQueueFamilyProperties2(
1849 VkPhysicalDevice physicalDevice,
1850 uint32_t* pQueueFamilyPropertyCount,
1851 VkQueueFamilyProperties2* pQueueFamilyProperties)
1852 {
1853
1854 VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pQueueFamilyPropertyCount);
1855
1856 vk_outarray_append(&out, p) {
1857 p->queueFamilyProperties = anv_queue_family_properties;
1858
1859 vk_foreach_struct(s, p->pNext) {
1860 anv_debug_ignored_stype(s->sType);
1861 }
1862 }
1863 }
1864
1865 void anv_GetPhysicalDeviceMemoryProperties(
1866 VkPhysicalDevice physicalDevice,
1867 VkPhysicalDeviceMemoryProperties* pMemoryProperties)
1868 {
1869 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
1870
1871 pMemoryProperties->memoryTypeCount = physical_device->memory.type_count;
1872 for (uint32_t i = 0; i < physical_device->memory.type_count; i++) {
1873 pMemoryProperties->memoryTypes[i] = (VkMemoryType) {
1874 .propertyFlags = physical_device->memory.types[i].propertyFlags,
1875 .heapIndex = physical_device->memory.types[i].heapIndex,
1876 };
1877 }
1878
1879 pMemoryProperties->memoryHeapCount = physical_device->memory.heap_count;
1880 for (uint32_t i = 0; i < physical_device->memory.heap_count; i++) {
1881 pMemoryProperties->memoryHeaps[i] = (VkMemoryHeap) {
1882 .size = physical_device->memory.heaps[i].size,
1883 .flags = physical_device->memory.heaps[i].flags,
1884 };
1885 }
1886 }
1887
1888 static void
1889 anv_get_memory_budget(VkPhysicalDevice physicalDevice,
1890 VkPhysicalDeviceMemoryBudgetPropertiesEXT *memoryBudget)
1891 {
1892 ANV_FROM_HANDLE(anv_physical_device, device, physicalDevice);
1893 uint64_t sys_available = get_available_system_memory();
1894 assert(sys_available > 0);
1895
1896 VkDeviceSize total_heaps_size = 0;
1897 for (size_t i = 0; i < device->memory.heap_count; i++)
1898 total_heaps_size += device->memory.heaps[i].size;
1899
1900 for (size_t i = 0; i < device->memory.heap_count; i++) {
1901 VkDeviceSize heap_size = device->memory.heaps[i].size;
1902 VkDeviceSize heap_used = device->memory.heaps[i].used;
1903 VkDeviceSize heap_budget;
1904
1905 double heap_proportion = (double) heap_size / total_heaps_size;
1906 VkDeviceSize sys_available_prop = sys_available * heap_proportion;
1907
1908 /*
1909 * Let's not incite the app to starve the system: report at most 90% of
1910 * available system memory.
1911 */
1912 uint64_t heap_available = sys_available_prop * 9 / 10;
1913 heap_budget = MIN2(heap_size, heap_used + heap_available);
1914
1915 /*
1916 * Round down to the nearest MB
1917 */
1918 heap_budget &= ~((1ull << 20) - 1);
1919
1920 /*
1921 * The heapBudget value must be non-zero for array elements less than
1922 * VkPhysicalDeviceMemoryProperties::memoryHeapCount. The heapBudget
1923 * value must be less than or equal to VkMemoryHeap::size for each heap.
1924 */
1925 assert(0 < heap_budget && heap_budget <= heap_size);
1926
1927 memoryBudget->heapUsage[i] = heap_used;
1928 memoryBudget->heapBudget[i] = heap_budget;
1929 }
1930
1931 /* The heapBudget and heapUsage values must be zero for array elements
1932 * greater than or equal to VkPhysicalDeviceMemoryProperties::memoryHeapCount
1933 */
1934 for (uint32_t i = device->memory.heap_count; i < VK_MAX_MEMORY_HEAPS; i++) {
1935 memoryBudget->heapBudget[i] = 0;
1936 memoryBudget->heapUsage[i] = 0;
1937 }
1938 }
1939
1940 void anv_GetPhysicalDeviceMemoryProperties2(
1941 VkPhysicalDevice physicalDevice,
1942 VkPhysicalDeviceMemoryProperties2* pMemoryProperties)
1943 {
1944 anv_GetPhysicalDeviceMemoryProperties(physicalDevice,
1945 &pMemoryProperties->memoryProperties);
1946
1947 vk_foreach_struct(ext, pMemoryProperties->pNext) {
1948 switch (ext->sType) {
1949 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT:
1950 anv_get_memory_budget(physicalDevice, (void*)ext);
1951 break;
1952 default:
1953 anv_debug_ignored_stype(ext->sType);
1954 break;
1955 }
1956 }
1957 }
1958
1959 void
1960 anv_GetDeviceGroupPeerMemoryFeatures(
1961 VkDevice device,
1962 uint32_t heapIndex,
1963 uint32_t localDeviceIndex,
1964 uint32_t remoteDeviceIndex,
1965 VkPeerMemoryFeatureFlags* pPeerMemoryFeatures)
1966 {
1967 assert(localDeviceIndex == 0 && remoteDeviceIndex == 0);
1968 *pPeerMemoryFeatures = VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT |
1969 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT |
1970 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT |
1971 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT;
1972 }
1973
1974 PFN_vkVoidFunction anv_GetInstanceProcAddr(
1975 VkInstance _instance,
1976 const char* pName)
1977 {
1978 ANV_FROM_HANDLE(anv_instance, instance, _instance);
1979
1980 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
1981 * when we have to return valid function pointers, NULL, or it's left
1982 * undefined. See the table for exact details.
1983 */
1984 if (pName == NULL)
1985 return NULL;
1986
1987 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
1988 if (strcmp(pName, "vk" #entrypoint) == 0) \
1989 return (PFN_vkVoidFunction)anv_##entrypoint
1990
1991 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties);
1992 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties);
1993 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceVersion);
1994 LOOKUP_ANV_ENTRYPOINT(CreateInstance);
1995
1996 #undef LOOKUP_ANV_ENTRYPOINT
1997
1998 if (instance == NULL)
1999 return NULL;
2000
2001 int idx = anv_get_instance_entrypoint_index(pName);
2002 if (idx >= 0)
2003 return instance->dispatch.entrypoints[idx];
2004
2005 idx = anv_get_physical_device_entrypoint_index(pName);
2006 if (idx >= 0)
2007 return instance->physicalDevice.dispatch.entrypoints[idx];
2008
2009 idx = anv_get_device_entrypoint_index(pName);
2010 if (idx >= 0)
2011 return instance->device_dispatch.entrypoints[idx];
2012
2013 return NULL;
2014 }
2015
2016 /* With version 1+ of the loader interface the ICD should expose
2017 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
2018 */
2019 PUBLIC
2020 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
2021 VkInstance instance,
2022 const char* pName);
2023
2024 PUBLIC
2025 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
2026 VkInstance instance,
2027 const char* pName)
2028 {
2029 return anv_GetInstanceProcAddr(instance, pName);
2030 }
2031
2032 PFN_vkVoidFunction anv_GetDeviceProcAddr(
2033 VkDevice _device,
2034 const char* pName)
2035 {
2036 ANV_FROM_HANDLE(anv_device, device, _device);
2037
2038 if (!device || !pName)
2039 return NULL;
2040
2041 int idx = anv_get_device_entrypoint_index(pName);
2042 if (idx < 0)
2043 return NULL;
2044
2045 return device->dispatch.entrypoints[idx];
2046 }
2047
2048 /* With version 4+ of the loader interface the ICD should expose
2049 * vk_icdGetPhysicalDeviceProcAddr()
2050 */
2051 PUBLIC
2052 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetPhysicalDeviceProcAddr(
2053 VkInstance _instance,
2054 const char* pName);
2055
2056 PFN_vkVoidFunction vk_icdGetPhysicalDeviceProcAddr(
2057 VkInstance _instance,
2058 const char* pName)
2059 {
2060 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2061
2062 if (!pName || !instance)
2063 return NULL;
2064
2065 int idx = anv_get_physical_device_entrypoint_index(pName);
2066 if (idx < 0)
2067 return NULL;
2068
2069 return instance->physicalDevice.dispatch.entrypoints[idx];
2070 }
2071
2072
2073 VkResult
2074 anv_CreateDebugReportCallbackEXT(VkInstance _instance,
2075 const VkDebugReportCallbackCreateInfoEXT* pCreateInfo,
2076 const VkAllocationCallbacks* pAllocator,
2077 VkDebugReportCallbackEXT* pCallback)
2078 {
2079 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2080 return vk_create_debug_report_callback(&instance->debug_report_callbacks,
2081 pCreateInfo, pAllocator, &instance->alloc,
2082 pCallback);
2083 }
2084
2085 void
2086 anv_DestroyDebugReportCallbackEXT(VkInstance _instance,
2087 VkDebugReportCallbackEXT _callback,
2088 const VkAllocationCallbacks* pAllocator)
2089 {
2090 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2091 vk_destroy_debug_report_callback(&instance->debug_report_callbacks,
2092 _callback, pAllocator, &instance->alloc);
2093 }
2094
2095 void
2096 anv_DebugReportMessageEXT(VkInstance _instance,
2097 VkDebugReportFlagsEXT flags,
2098 VkDebugReportObjectTypeEXT objectType,
2099 uint64_t object,
2100 size_t location,
2101 int32_t messageCode,
2102 const char* pLayerPrefix,
2103 const char* pMessage)
2104 {
2105 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2106 vk_debug_report(&instance->debug_report_callbacks, flags, objectType,
2107 object, location, messageCode, pLayerPrefix, pMessage);
2108 }
2109
2110 static void
2111 anv_queue_init(struct anv_device *device, struct anv_queue *queue)
2112 {
2113 queue->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
2114 queue->device = device;
2115 queue->flags = 0;
2116 }
2117
2118 static void
2119 anv_queue_finish(struct anv_queue *queue)
2120 {
2121 }
2122
2123 static struct anv_state
2124 anv_state_pool_emit_data(struct anv_state_pool *pool, size_t size, size_t align, const void *p)
2125 {
2126 struct anv_state state;
2127
2128 state = anv_state_pool_alloc(pool, size, align);
2129 memcpy(state.map, p, size);
2130
2131 return state;
2132 }
2133
2134 /* Haswell border color is a bit of a disaster. Float and unorm formats use a
2135 * straightforward 32-bit float color in the first 64 bytes. Instead of using
2136 * a nice float/integer union like Gen8+, Haswell specifies the integer border
2137 * color as a separate entry /after/ the float color. The layout of this entry
2138 * also depends on the format's bpp (with extra hacks for RG32), and overlaps.
2139 *
2140 * Since we don't know the format/bpp, we can't make any of the border colors
2141 * containing '1' work for all formats, as it would be in the wrong place for
2142 * some of them. We opt to make 32-bit integers work as this seems like the
2143 * most common option. Fortunately, transparent black works regardless, as
2144 * all zeroes is the same in every bit-size.
2145 */
2146 struct hsw_border_color {
2147 float float32[4];
2148 uint32_t _pad0[12];
2149 uint32_t uint32[4];
2150 uint32_t _pad1[108];
2151 };
2152
2153 struct gen8_border_color {
2154 union {
2155 float float32[4];
2156 uint32_t uint32[4];
2157 };
2158 /* Pad out to 64 bytes */
2159 uint32_t _pad[12];
2160 };
2161
2162 static void
2163 anv_device_init_border_colors(struct anv_device *device)
2164 {
2165 if (device->info.is_haswell) {
2166 static const struct hsw_border_color border_colors[] = {
2167 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
2168 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
2169 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
2170 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = { .uint32 = { 0, 0, 0, 0 } },
2171 [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = { .uint32 = { 0, 0, 0, 1 } },
2172 [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = { .uint32 = { 1, 1, 1, 1 } },
2173 };
2174
2175 device->border_colors =
2176 anv_state_pool_emit_data(&device->dynamic_state_pool,
2177 sizeof(border_colors), 512, border_colors);
2178 } else {
2179 static const struct gen8_border_color border_colors[] = {
2180 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
2181 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
2182 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
2183 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = { .uint32 = { 0, 0, 0, 0 } },
2184 [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = { .uint32 = { 0, 0, 0, 1 } },
2185 [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = { .uint32 = { 1, 1, 1, 1 } },
2186 };
2187
2188 device->border_colors =
2189 anv_state_pool_emit_data(&device->dynamic_state_pool,
2190 sizeof(border_colors), 64, border_colors);
2191 }
2192 }
2193
2194 static void
2195 anv_device_init_trivial_batch(struct anv_device *device)
2196 {
2197 anv_bo_init_new(&device->trivial_batch_bo, device, 4096);
2198
2199 if (device->instance->physicalDevice.has_exec_async)
2200 device->trivial_batch_bo.flags |= EXEC_OBJECT_ASYNC;
2201
2202 if (device->instance->physicalDevice.use_softpin)
2203 device->trivial_batch_bo.flags |= EXEC_OBJECT_PINNED;
2204
2205 anv_vma_alloc(device, &device->trivial_batch_bo);
2206
2207 void *map = anv_gem_mmap(device, device->trivial_batch_bo.gem_handle,
2208 0, 4096, 0);
2209
2210 struct anv_batch batch = {
2211 .start = map,
2212 .next = map,
2213 .end = map + 4096,
2214 };
2215
2216 anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
2217 anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
2218
2219 if (!device->info.has_llc)
2220 gen_clflush_range(map, batch.next - map);
2221
2222 anv_gem_munmap(map, device->trivial_batch_bo.size);
2223 }
2224
2225 VkResult anv_EnumerateDeviceExtensionProperties(
2226 VkPhysicalDevice physicalDevice,
2227 const char* pLayerName,
2228 uint32_t* pPropertyCount,
2229 VkExtensionProperties* pProperties)
2230 {
2231 ANV_FROM_HANDLE(anv_physical_device, device, physicalDevice);
2232 VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
2233
2234 for (int i = 0; i < ANV_DEVICE_EXTENSION_COUNT; i++) {
2235 if (device->supported_extensions.extensions[i]) {
2236 vk_outarray_append(&out, prop) {
2237 *prop = anv_device_extensions[i];
2238 }
2239 }
2240 }
2241
2242 return vk_outarray_status(&out);
2243 }
2244
2245 static void
2246 anv_device_init_dispatch(struct anv_device *device)
2247 {
2248 const struct anv_device_dispatch_table *genX_table;
2249 switch (device->info.gen) {
2250 case 12:
2251 genX_table = &gen12_device_dispatch_table;
2252 break;
2253 case 11:
2254 genX_table = &gen11_device_dispatch_table;
2255 break;
2256 case 10:
2257 genX_table = &gen10_device_dispatch_table;
2258 break;
2259 case 9:
2260 genX_table = &gen9_device_dispatch_table;
2261 break;
2262 case 8:
2263 genX_table = &gen8_device_dispatch_table;
2264 break;
2265 case 7:
2266 if (device->info.is_haswell)
2267 genX_table = &gen75_device_dispatch_table;
2268 else
2269 genX_table = &gen7_device_dispatch_table;
2270 break;
2271 default:
2272 unreachable("unsupported gen\n");
2273 }
2274
2275 for (unsigned i = 0; i < ARRAY_SIZE(device->dispatch.entrypoints); i++) {
2276 /* Vulkan requires that entrypoints for extensions which have not been
2277 * enabled must not be advertised.
2278 */
2279 if (!anv_device_entrypoint_is_enabled(i, device->instance->app_info.api_version,
2280 &device->instance->enabled_extensions,
2281 &device->enabled_extensions)) {
2282 device->dispatch.entrypoints[i] = NULL;
2283 } else if (genX_table->entrypoints[i]) {
2284 device->dispatch.entrypoints[i] = genX_table->entrypoints[i];
2285 } else {
2286 device->dispatch.entrypoints[i] =
2287 anv_device_dispatch_table.entrypoints[i];
2288 }
2289 }
2290 }
2291
2292 static int
2293 vk_priority_to_gen(int priority)
2294 {
2295 switch (priority) {
2296 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT:
2297 return GEN_CONTEXT_LOW_PRIORITY;
2298 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT:
2299 return GEN_CONTEXT_MEDIUM_PRIORITY;
2300 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT:
2301 return GEN_CONTEXT_HIGH_PRIORITY;
2302 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT:
2303 return GEN_CONTEXT_REALTIME_PRIORITY;
2304 default:
2305 unreachable("Invalid priority");
2306 }
2307 }
2308
2309 static void
2310 anv_device_init_hiz_clear_value_bo(struct anv_device *device)
2311 {
2312 anv_bo_init_new(&device->hiz_clear_bo, device, 4096);
2313
2314 if (device->instance->physicalDevice.has_exec_async)
2315 device->hiz_clear_bo.flags |= EXEC_OBJECT_ASYNC;
2316
2317 if (device->instance->physicalDevice.use_softpin)
2318 device->hiz_clear_bo.flags |= EXEC_OBJECT_PINNED;
2319
2320 anv_vma_alloc(device, &device->hiz_clear_bo);
2321
2322 uint32_t *map = anv_gem_mmap(device, device->hiz_clear_bo.gem_handle,
2323 0, 4096, 0);
2324
2325 union isl_color_value hiz_clear = { .u32 = { 0, } };
2326 hiz_clear.f32[0] = ANV_HZ_FC_VAL;
2327
2328 memcpy(map, hiz_clear.u32, sizeof(hiz_clear.u32));
2329 anv_gem_munmap(map, device->hiz_clear_bo.size);
2330 }
2331
2332 static bool
2333 get_bo_from_pool(struct gen_batch_decode_bo *ret,
2334 struct anv_block_pool *pool,
2335 uint64_t address)
2336 {
2337 for (uint32_t i = 0; i < pool->nbos; i++) {
2338 uint64_t bo_address = pool->bos[i].offset & (~0ull >> 16);
2339 uint32_t bo_size = pool->bos[i].size;
2340 if (address >= bo_address && address < (bo_address + bo_size)) {
2341 *ret = (struct gen_batch_decode_bo) {
2342 .addr = bo_address,
2343 .size = bo_size,
2344 .map = pool->bos[i].map,
2345 };
2346 return true;
2347 }
2348 }
2349 return false;
2350 }
2351
2352 /* Finding a buffer for batch decoding */
2353 static struct gen_batch_decode_bo
2354 decode_get_bo(void *v_batch, bool ppgtt, uint64_t address)
2355 {
2356 struct anv_device *device = v_batch;
2357 struct gen_batch_decode_bo ret_bo = {};
2358
2359 assert(ppgtt);
2360
2361 if (get_bo_from_pool(&ret_bo, &device->dynamic_state_pool.block_pool, address))
2362 return ret_bo;
2363 if (get_bo_from_pool(&ret_bo, &device->instruction_state_pool.block_pool, address))
2364 return ret_bo;
2365 if (get_bo_from_pool(&ret_bo, &device->binding_table_pool.block_pool, address))
2366 return ret_bo;
2367 if (get_bo_from_pool(&ret_bo, &device->surface_state_pool.block_pool, address))
2368 return ret_bo;
2369
2370 if (!device->cmd_buffer_being_decoded)
2371 return (struct gen_batch_decode_bo) { };
2372
2373 struct anv_batch_bo **bo;
2374
2375 u_vector_foreach(bo, &device->cmd_buffer_being_decoded->seen_bbos) {
2376 /* The decoder zeroes out the top 16 bits, so we need to as well */
2377 uint64_t bo_address = (*bo)->bo.offset & (~0ull >> 16);
2378
2379 if (address >= bo_address && address < bo_address + (*bo)->bo.size) {
2380 return (struct gen_batch_decode_bo) {
2381 .addr = bo_address,
2382 .size = (*bo)->bo.size,
2383 .map = (*bo)->bo.map,
2384 };
2385 }
2386 }
2387
2388 return (struct gen_batch_decode_bo) { };
2389 }
2390
2391 struct gen_aux_map_buffer {
2392 struct gen_buffer base;
2393 struct anv_state state;
2394 };
2395
2396 static struct gen_buffer *
2397 gen_aux_map_buffer_alloc(void *driver_ctx, uint32_t size)
2398 {
2399 struct gen_aux_map_buffer *buf = malloc(sizeof(struct gen_aux_map_buffer));
2400 if (!buf)
2401 return NULL;
2402
2403 struct anv_device *device = (struct anv_device*)driver_ctx;
2404 assert(device->instance->physicalDevice.supports_48bit_addresses &&
2405 device->instance->physicalDevice.use_softpin);
2406
2407 struct anv_state_pool *pool = &device->dynamic_state_pool;
2408 buf->state = anv_state_pool_alloc(pool, size, size);
2409
2410 buf->base.gpu = pool->block_pool.bo->offset + buf->state.offset;
2411 buf->base.gpu_end = buf->base.gpu + buf->state.alloc_size;
2412 buf->base.map = buf->state.map;
2413 buf->base.driver_bo = &buf->state;
2414 return &buf->base;
2415 }
2416
2417 static void
2418 gen_aux_map_buffer_free(void *driver_ctx, struct gen_buffer *buffer)
2419 {
2420 struct gen_aux_map_buffer *buf = (struct gen_aux_map_buffer*)buffer;
2421 struct anv_device *device = (struct anv_device*)driver_ctx;
2422 struct anv_state_pool *pool = &device->dynamic_state_pool;
2423 anv_state_pool_free(pool, buf->state);
2424 free(buf);
2425 }
2426
2427 static struct gen_mapped_pinned_buffer_alloc aux_map_allocator = {
2428 .alloc = gen_aux_map_buffer_alloc,
2429 .free = gen_aux_map_buffer_free,
2430 };
2431
2432 VkResult anv_CreateDevice(
2433 VkPhysicalDevice physicalDevice,
2434 const VkDeviceCreateInfo* pCreateInfo,
2435 const VkAllocationCallbacks* pAllocator,
2436 VkDevice* pDevice)
2437 {
2438 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
2439 VkResult result;
2440 struct anv_device *device;
2441
2442 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO);
2443
2444 struct anv_device_extension_table enabled_extensions = { };
2445 for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
2446 int idx;
2447 for (idx = 0; idx < ANV_DEVICE_EXTENSION_COUNT; idx++) {
2448 if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
2449 anv_device_extensions[idx].extensionName) == 0)
2450 break;
2451 }
2452
2453 if (idx >= ANV_DEVICE_EXTENSION_COUNT)
2454 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
2455
2456 if (!physical_device->supported_extensions.extensions[idx])
2457 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
2458
2459 enabled_extensions.extensions[idx] = true;
2460 }
2461
2462 /* Check enabled features */
2463 if (pCreateInfo->pEnabledFeatures) {
2464 VkPhysicalDeviceFeatures supported_features;
2465 anv_GetPhysicalDeviceFeatures(physicalDevice, &supported_features);
2466 VkBool32 *supported_feature = (VkBool32 *)&supported_features;
2467 VkBool32 *enabled_feature = (VkBool32 *)pCreateInfo->pEnabledFeatures;
2468 unsigned num_features = sizeof(VkPhysicalDeviceFeatures) / sizeof(VkBool32);
2469 for (uint32_t i = 0; i < num_features; i++) {
2470 if (enabled_feature[i] && !supported_feature[i])
2471 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
2472 }
2473 }
2474
2475 /* Check requested queues and fail if we are requested to create any
2476 * queues with flags we don't support.
2477 */
2478 assert(pCreateInfo->queueCreateInfoCount > 0);
2479 for (uint32_t i = 0; i < pCreateInfo->queueCreateInfoCount; i++) {
2480 if (pCreateInfo->pQueueCreateInfos[i].flags != 0)
2481 return vk_error(VK_ERROR_INITIALIZATION_FAILED);
2482 }
2483
2484 /* Check if client specified queue priority. */
2485 const VkDeviceQueueGlobalPriorityCreateInfoEXT *queue_priority =
2486 vk_find_struct_const(pCreateInfo->pQueueCreateInfos[0].pNext,
2487 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT);
2488
2489 VkQueueGlobalPriorityEXT priority =
2490 queue_priority ? queue_priority->globalPriority :
2491 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT;
2492
2493 device = vk_alloc2(&physical_device->instance->alloc, pAllocator,
2494 sizeof(*device), 8,
2495 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
2496 if (!device)
2497 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2498
2499 if (INTEL_DEBUG & DEBUG_BATCH) {
2500 const unsigned decode_flags =
2501 GEN_BATCH_DECODE_FULL |
2502 ((INTEL_DEBUG & DEBUG_COLOR) ? GEN_BATCH_DECODE_IN_COLOR : 0) |
2503 GEN_BATCH_DECODE_OFFSETS |
2504 GEN_BATCH_DECODE_FLOATS;
2505
2506 gen_batch_decode_ctx_init(&device->decoder_ctx,
2507 &physical_device->info,
2508 stderr, decode_flags, NULL,
2509 decode_get_bo, NULL, device);
2510 }
2511
2512 device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
2513 device->instance = physical_device->instance;
2514 device->chipset_id = physical_device->chipset_id;
2515 device->no_hw = physical_device->no_hw;
2516 device->_lost = false;
2517
2518 if (pAllocator)
2519 device->alloc = *pAllocator;
2520 else
2521 device->alloc = physical_device->instance->alloc;
2522
2523 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
2524 device->fd = open(physical_device->path, O_RDWR | O_CLOEXEC);
2525 if (device->fd == -1) {
2526 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2527 goto fail_device;
2528 }
2529
2530 device->context_id = anv_gem_create_context(device);
2531 if (device->context_id == -1) {
2532 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2533 goto fail_fd;
2534 }
2535
2536 if (physical_device->use_softpin) {
2537 if (pthread_mutex_init(&device->vma_mutex, NULL) != 0) {
2538 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2539 goto fail_context_id;
2540 }
2541
2542 /* keep the page with address zero out of the allocator */
2543 struct anv_memory_heap *low_heap =
2544 &physical_device->memory.heaps[physical_device->memory.heap_count - 1];
2545 util_vma_heap_init(&device->vma_lo, low_heap->vma_start, low_heap->vma_size);
2546 device->vma_lo_available = low_heap->size;
2547
2548 struct anv_memory_heap *high_heap =
2549 &physical_device->memory.heaps[0];
2550 util_vma_heap_init(&device->vma_hi, high_heap->vma_start, high_heap->vma_size);
2551 device->vma_hi_available = physical_device->memory.heap_count == 1 ? 0 :
2552 high_heap->size;
2553 }
2554
2555 list_inithead(&device->memory_objects);
2556
2557 /* As per spec, the driver implementation may deny requests to acquire
2558 * a priority above the default priority (MEDIUM) if the caller does not
2559 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
2560 * is returned.
2561 */
2562 if (physical_device->has_context_priority) {
2563 int err = anv_gem_set_context_param(device->fd, device->context_id,
2564 I915_CONTEXT_PARAM_PRIORITY,
2565 vk_priority_to_gen(priority));
2566 if (err != 0 && priority > VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT) {
2567 result = vk_error(VK_ERROR_NOT_PERMITTED_EXT);
2568 goto fail_vmas;
2569 }
2570 }
2571
2572 device->info = physical_device->info;
2573 device->isl_dev = physical_device->isl_dev;
2574
2575 /* On Broadwell and later, we can use batch chaining to more efficiently
2576 * implement growing command buffers. Prior to Haswell, the kernel
2577 * command parser gets in the way and we have to fall back to growing
2578 * the batch.
2579 */
2580 device->can_chain_batches = device->info.gen >= 8;
2581
2582 device->robust_buffer_access = pCreateInfo->pEnabledFeatures &&
2583 pCreateInfo->pEnabledFeatures->robustBufferAccess;
2584 device->enabled_extensions = enabled_extensions;
2585
2586 anv_device_init_dispatch(device);
2587
2588 if (pthread_mutex_init(&device->mutex, NULL) != 0) {
2589 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2590 goto fail_context_id;
2591 }
2592
2593 pthread_condattr_t condattr;
2594 if (pthread_condattr_init(&condattr) != 0) {
2595 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2596 goto fail_mutex;
2597 }
2598 if (pthread_condattr_setclock(&condattr, CLOCK_MONOTONIC) != 0) {
2599 pthread_condattr_destroy(&condattr);
2600 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2601 goto fail_mutex;
2602 }
2603 if (pthread_cond_init(&device->queue_submit, &condattr) != 0) {
2604 pthread_condattr_destroy(&condattr);
2605 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
2606 goto fail_mutex;
2607 }
2608 pthread_condattr_destroy(&condattr);
2609
2610 uint64_t bo_flags =
2611 (physical_device->supports_48bit_addresses ? EXEC_OBJECT_SUPPORTS_48B_ADDRESS : 0) |
2612 (physical_device->has_exec_async ? EXEC_OBJECT_ASYNC : 0) |
2613 (physical_device->has_exec_capture ? EXEC_OBJECT_CAPTURE : 0) |
2614 (physical_device->use_softpin ? EXEC_OBJECT_PINNED : 0);
2615
2616 anv_bo_pool_init(&device->batch_bo_pool, device, bo_flags);
2617
2618 result = anv_bo_cache_init(&device->bo_cache);
2619 if (result != VK_SUCCESS)
2620 goto fail_batch_bo_pool;
2621
2622 if (!physical_device->use_softpin)
2623 bo_flags &= ~EXEC_OBJECT_SUPPORTS_48B_ADDRESS;
2624
2625 result = anv_state_pool_init(&device->dynamic_state_pool, device,
2626 DYNAMIC_STATE_POOL_MIN_ADDRESS,
2627 16384,
2628 bo_flags);
2629 if (result != VK_SUCCESS)
2630 goto fail_bo_cache;
2631
2632 result = anv_state_pool_init(&device->instruction_state_pool, device,
2633 INSTRUCTION_STATE_POOL_MIN_ADDRESS,
2634 16384,
2635 bo_flags);
2636 if (result != VK_SUCCESS)
2637 goto fail_dynamic_state_pool;
2638
2639 result = anv_state_pool_init(&device->surface_state_pool, device,
2640 SURFACE_STATE_POOL_MIN_ADDRESS,
2641 4096,
2642 bo_flags);
2643 if (result != VK_SUCCESS)
2644 goto fail_instruction_state_pool;
2645
2646 if (physical_device->use_softpin) {
2647 result = anv_state_pool_init(&device->binding_table_pool, device,
2648 BINDING_TABLE_POOL_MIN_ADDRESS,
2649 4096,
2650 bo_flags);
2651 if (result != VK_SUCCESS)
2652 goto fail_surface_state_pool;
2653 }
2654
2655 if (device->info.gen >= 12) {
2656 device->aux_map_ctx = gen_aux_map_init(device, &aux_map_allocator,
2657 &physical_device->info);
2658 if (!device->aux_map_ctx)
2659 goto fail_binding_table_pool;
2660 }
2661
2662 result = anv_bo_init_new(&device->workaround_bo, device, 4096);
2663 if (result != VK_SUCCESS)
2664 goto fail_surface_aux_map_pool;
2665
2666 if (physical_device->use_softpin)
2667 device->workaround_bo.flags |= EXEC_OBJECT_PINNED;
2668
2669 if (!anv_vma_alloc(device, &device->workaround_bo))
2670 goto fail_workaround_bo;
2671
2672 anv_device_init_trivial_batch(device);
2673
2674 if (device->info.gen >= 10)
2675 anv_device_init_hiz_clear_value_bo(device);
2676
2677 anv_scratch_pool_init(device, &device->scratch_pool);
2678
2679 anv_queue_init(device, &device->queue);
2680
2681 switch (device->info.gen) {
2682 case 7:
2683 if (!device->info.is_haswell)
2684 result = gen7_init_device_state(device);
2685 else
2686 result = gen75_init_device_state(device);
2687 break;
2688 case 8:
2689 result = gen8_init_device_state(device);
2690 break;
2691 case 9:
2692 result = gen9_init_device_state(device);
2693 break;
2694 case 10:
2695 result = gen10_init_device_state(device);
2696 break;
2697 case 11:
2698 result = gen11_init_device_state(device);
2699 break;
2700 case 12:
2701 result = gen12_init_device_state(device);
2702 break;
2703 default:
2704 /* Shouldn't get here as we don't create physical devices for any other
2705 * gens. */
2706 unreachable("unhandled gen");
2707 }
2708 if (result != VK_SUCCESS)
2709 goto fail_workaround_bo;
2710
2711 anv_pipeline_cache_init(&device->default_pipeline_cache, device, true);
2712
2713 anv_device_init_blorp(device);
2714
2715 anv_device_init_border_colors(device);
2716
2717 anv_device_perf_init(device);
2718
2719 *pDevice = anv_device_to_handle(device);
2720
2721 return VK_SUCCESS;
2722
2723 fail_workaround_bo:
2724 anv_queue_finish(&device->queue);
2725 anv_scratch_pool_finish(device, &device->scratch_pool);
2726 anv_gem_munmap(device->workaround_bo.map, device->workaround_bo.size);
2727 anv_gem_close(device, device->workaround_bo.gem_handle);
2728 fail_surface_aux_map_pool:
2729 if (device->info.gen >= 12) {
2730 gen_aux_map_finish(device->aux_map_ctx);
2731 device->aux_map_ctx = NULL;
2732 }
2733 fail_binding_table_pool:
2734 if (physical_device->use_softpin)
2735 anv_state_pool_finish(&device->binding_table_pool);
2736 fail_surface_state_pool:
2737 anv_state_pool_finish(&device->surface_state_pool);
2738 fail_instruction_state_pool:
2739 anv_state_pool_finish(&device->instruction_state_pool);
2740 fail_dynamic_state_pool:
2741 anv_state_pool_finish(&device->dynamic_state_pool);
2742 fail_bo_cache:
2743 anv_bo_cache_finish(&device->bo_cache);
2744 fail_batch_bo_pool:
2745 anv_bo_pool_finish(&device->batch_bo_pool);
2746 pthread_cond_destroy(&device->queue_submit);
2747 fail_mutex:
2748 pthread_mutex_destroy(&device->mutex);
2749 fail_vmas:
2750 if (physical_device->use_softpin) {
2751 util_vma_heap_finish(&device->vma_hi);
2752 util_vma_heap_finish(&device->vma_lo);
2753 }
2754 fail_context_id:
2755 anv_gem_destroy_context(device, device->context_id);
2756 fail_fd:
2757 close(device->fd);
2758 fail_device:
2759 vk_free(&device->alloc, device);
2760
2761 return result;
2762 }
2763
2764 void anv_DestroyDevice(
2765 VkDevice _device,
2766 const VkAllocationCallbacks* pAllocator)
2767 {
2768 ANV_FROM_HANDLE(anv_device, device, _device);
2769 struct anv_physical_device *physical_device;
2770
2771 if (!device)
2772 return;
2773
2774 physical_device = &device->instance->physicalDevice;
2775
2776 anv_device_finish_blorp(device);
2777
2778 anv_pipeline_cache_finish(&device->default_pipeline_cache);
2779
2780 anv_queue_finish(&device->queue);
2781
2782 #ifdef HAVE_VALGRIND
2783 /* We only need to free these to prevent valgrind errors. The backing
2784 * BO will go away in a couple of lines so we don't actually leak.
2785 */
2786 anv_state_pool_free(&device->dynamic_state_pool, device->border_colors);
2787 anv_state_pool_free(&device->dynamic_state_pool, device->slice_hash);
2788 #endif
2789
2790 anv_scratch_pool_finish(device, &device->scratch_pool);
2791
2792 anv_gem_munmap(device->workaround_bo.map, device->workaround_bo.size);
2793 anv_vma_free(device, &device->workaround_bo);
2794 anv_gem_close(device, device->workaround_bo.gem_handle);
2795
2796 anv_vma_free(device, &device->trivial_batch_bo);
2797 anv_gem_close(device, device->trivial_batch_bo.gem_handle);
2798 if (device->info.gen >= 10)
2799 anv_gem_close(device, device->hiz_clear_bo.gem_handle);
2800
2801 if (device->info.gen >= 12) {
2802 gen_aux_map_finish(device->aux_map_ctx);
2803 device->aux_map_ctx = NULL;
2804 }
2805
2806 if (physical_device->use_softpin)
2807 anv_state_pool_finish(&device->binding_table_pool);
2808 anv_state_pool_finish(&device->surface_state_pool);
2809 anv_state_pool_finish(&device->instruction_state_pool);
2810 anv_state_pool_finish(&device->dynamic_state_pool);
2811
2812 anv_bo_cache_finish(&device->bo_cache);
2813
2814 anv_bo_pool_finish(&device->batch_bo_pool);
2815
2816 if (physical_device->use_softpin) {
2817 util_vma_heap_finish(&device->vma_hi);
2818 util_vma_heap_finish(&device->vma_lo);
2819 }
2820
2821 pthread_cond_destroy(&device->queue_submit);
2822 pthread_mutex_destroy(&device->mutex);
2823
2824 anv_gem_destroy_context(device, device->context_id);
2825
2826 if (INTEL_DEBUG & DEBUG_BATCH)
2827 gen_batch_decode_ctx_finish(&device->decoder_ctx);
2828
2829 close(device->fd);
2830
2831 vk_free(&device->alloc, device);
2832 }
2833
2834 VkResult anv_EnumerateInstanceLayerProperties(
2835 uint32_t* pPropertyCount,
2836 VkLayerProperties* pProperties)
2837 {
2838 if (pProperties == NULL) {
2839 *pPropertyCount = 0;
2840 return VK_SUCCESS;
2841 }
2842
2843 /* None supported at this time */
2844 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
2845 }
2846
2847 VkResult anv_EnumerateDeviceLayerProperties(
2848 VkPhysicalDevice physicalDevice,
2849 uint32_t* pPropertyCount,
2850 VkLayerProperties* pProperties)
2851 {
2852 if (pProperties == NULL) {
2853 *pPropertyCount = 0;
2854 return VK_SUCCESS;
2855 }
2856
2857 /* None supported at this time */
2858 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
2859 }
2860
2861 void anv_GetDeviceQueue(
2862 VkDevice _device,
2863 uint32_t queueNodeIndex,
2864 uint32_t queueIndex,
2865 VkQueue* pQueue)
2866 {
2867 ANV_FROM_HANDLE(anv_device, device, _device);
2868
2869 assert(queueIndex == 0);
2870
2871 *pQueue = anv_queue_to_handle(&device->queue);
2872 }
2873
2874 void anv_GetDeviceQueue2(
2875 VkDevice _device,
2876 const VkDeviceQueueInfo2* pQueueInfo,
2877 VkQueue* pQueue)
2878 {
2879 ANV_FROM_HANDLE(anv_device, device, _device);
2880
2881 assert(pQueueInfo->queueIndex == 0);
2882
2883 if (pQueueInfo->flags == device->queue.flags)
2884 *pQueue = anv_queue_to_handle(&device->queue);
2885 else
2886 *pQueue = NULL;
2887 }
2888
2889 VkResult
2890 _anv_device_set_lost(struct anv_device *device,
2891 const char *file, int line,
2892 const char *msg, ...)
2893 {
2894 VkResult err;
2895 va_list ap;
2896
2897 device->_lost = true;
2898
2899 va_start(ap, msg);
2900 err = __vk_errorv(device->instance, device,
2901 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT,
2902 VK_ERROR_DEVICE_LOST, file, line, msg, ap);
2903 va_end(ap);
2904
2905 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
2906 abort();
2907
2908 return err;
2909 }
2910
2911 VkResult
2912 anv_device_query_status(struct anv_device *device)
2913 {
2914 /* This isn't likely as most of the callers of this function already check
2915 * for it. However, it doesn't hurt to check and it potentially lets us
2916 * avoid an ioctl.
2917 */
2918 if (anv_device_is_lost(device))
2919 return VK_ERROR_DEVICE_LOST;
2920
2921 uint32_t active, pending;
2922 int ret = anv_gem_gpu_get_reset_stats(device, &active, &pending);
2923 if (ret == -1) {
2924 /* We don't know the real error. */
2925 return anv_device_set_lost(device, "get_reset_stats failed: %m");
2926 }
2927
2928 if (active) {
2929 return anv_device_set_lost(device, "GPU hung on one of our command buffers");
2930 } else if (pending) {
2931 return anv_device_set_lost(device, "GPU hung with commands in-flight");
2932 }
2933
2934 return VK_SUCCESS;
2935 }
2936
2937 VkResult
2938 anv_device_bo_busy(struct anv_device *device, struct anv_bo *bo)
2939 {
2940 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
2941 * Other usages of the BO (such as on different hardware) will not be
2942 * flagged as "busy" by this ioctl. Use with care.
2943 */
2944 int ret = anv_gem_busy(device, bo->gem_handle);
2945 if (ret == 1) {
2946 return VK_NOT_READY;
2947 } else if (ret == -1) {
2948 /* We don't know the real error. */
2949 return anv_device_set_lost(device, "gem wait failed: %m");
2950 }
2951
2952 /* Query for device status after the busy call. If the BO we're checking
2953 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
2954 * client because it clearly doesn't have valid data. Yes, this most
2955 * likely means an ioctl, but we just did an ioctl to query the busy status
2956 * so it's no great loss.
2957 */
2958 return anv_device_query_status(device);
2959 }
2960
2961 VkResult
2962 anv_device_wait(struct anv_device *device, struct anv_bo *bo,
2963 int64_t timeout)
2964 {
2965 int ret = anv_gem_wait(device, bo->gem_handle, &timeout);
2966 if (ret == -1 && errno == ETIME) {
2967 return VK_TIMEOUT;
2968 } else if (ret == -1) {
2969 /* We don't know the real error. */
2970 return anv_device_set_lost(device, "gem wait failed: %m");
2971 }
2972
2973 /* Query for device status after the wait. If the BO we're waiting on got
2974 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
2975 * because it clearly doesn't have valid data. Yes, this most likely means
2976 * an ioctl, but we just did an ioctl to wait so it's no great loss.
2977 */
2978 return anv_device_query_status(device);
2979 }
2980
2981 VkResult anv_DeviceWaitIdle(
2982 VkDevice _device)
2983 {
2984 ANV_FROM_HANDLE(anv_device, device, _device);
2985 if (anv_device_is_lost(device))
2986 return VK_ERROR_DEVICE_LOST;
2987
2988 struct anv_batch batch;
2989
2990 uint32_t cmds[8];
2991 batch.start = batch.next = cmds;
2992 batch.end = (void *) cmds + sizeof(cmds);
2993
2994 anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
2995 anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
2996
2997 return anv_device_submit_simple_batch(device, &batch);
2998 }
2999
3000 bool
3001 anv_vma_alloc(struct anv_device *device, struct anv_bo *bo)
3002 {
3003 if (!(bo->flags & EXEC_OBJECT_PINNED))
3004 return true;
3005
3006 pthread_mutex_lock(&device->vma_mutex);
3007
3008 bo->offset = 0;
3009
3010 if (bo->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS &&
3011 device->vma_hi_available >= bo->size) {
3012 uint64_t addr = util_vma_heap_alloc(&device->vma_hi, bo->size, 4096);
3013 if (addr) {
3014 bo->offset = gen_canonical_address(addr);
3015 assert(addr == gen_48b_address(bo->offset));
3016 device->vma_hi_available -= bo->size;
3017 }
3018 }
3019
3020 if (bo->offset == 0 && device->vma_lo_available >= bo->size) {
3021 uint64_t addr = util_vma_heap_alloc(&device->vma_lo, bo->size, 4096);
3022 if (addr) {
3023 bo->offset = gen_canonical_address(addr);
3024 assert(addr == gen_48b_address(bo->offset));
3025 device->vma_lo_available -= bo->size;
3026 }
3027 }
3028
3029 pthread_mutex_unlock(&device->vma_mutex);
3030
3031 return bo->offset != 0;
3032 }
3033
3034 void
3035 anv_vma_free(struct anv_device *device, struct anv_bo *bo)
3036 {
3037 if (!(bo->flags & EXEC_OBJECT_PINNED))
3038 return;
3039
3040 const uint64_t addr_48b = gen_48b_address(bo->offset);
3041
3042 pthread_mutex_lock(&device->vma_mutex);
3043
3044 if (addr_48b >= LOW_HEAP_MIN_ADDRESS &&
3045 addr_48b <= LOW_HEAP_MAX_ADDRESS) {
3046 util_vma_heap_free(&device->vma_lo, addr_48b, bo->size);
3047 device->vma_lo_available += bo->size;
3048 } else {
3049 ASSERTED const struct anv_physical_device *physical_device =
3050 &device->instance->physicalDevice;
3051 assert(addr_48b >= physical_device->memory.heaps[0].vma_start &&
3052 addr_48b < (physical_device->memory.heaps[0].vma_start +
3053 physical_device->memory.heaps[0].vma_size));
3054 util_vma_heap_free(&device->vma_hi, addr_48b, bo->size);
3055 device->vma_hi_available += bo->size;
3056 }
3057
3058 pthread_mutex_unlock(&device->vma_mutex);
3059
3060 bo->offset = 0;
3061 }
3062
3063 VkResult
3064 anv_bo_init_new(struct anv_bo *bo, struct anv_device *device, uint64_t size)
3065 {
3066 uint32_t gem_handle = anv_gem_create(device, size);
3067 if (!gem_handle)
3068 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
3069
3070 anv_bo_init(bo, gem_handle, size);
3071
3072 return VK_SUCCESS;
3073 }
3074
3075 VkResult anv_AllocateMemory(
3076 VkDevice _device,
3077 const VkMemoryAllocateInfo* pAllocateInfo,
3078 const VkAllocationCallbacks* pAllocator,
3079 VkDeviceMemory* pMem)
3080 {
3081 ANV_FROM_HANDLE(anv_device, device, _device);
3082 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3083 struct anv_device_memory *mem;
3084 VkResult result = VK_SUCCESS;
3085
3086 assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
3087
3088 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
3089 assert(pAllocateInfo->allocationSize > 0);
3090
3091 if (pAllocateInfo->allocationSize > MAX_MEMORY_ALLOCATION_SIZE)
3092 return VK_ERROR_OUT_OF_DEVICE_MEMORY;
3093
3094 /* FINISHME: Fail if allocation request exceeds heap size. */
3095
3096 mem = vk_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
3097 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
3098 if (mem == NULL)
3099 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
3100
3101 assert(pAllocateInfo->memoryTypeIndex < pdevice->memory.type_count);
3102 mem->type = &pdevice->memory.types[pAllocateInfo->memoryTypeIndex];
3103 mem->map = NULL;
3104 mem->map_size = 0;
3105 mem->ahw = NULL;
3106 mem->host_ptr = NULL;
3107
3108 uint64_t bo_flags = 0;
3109
3110 assert(mem->type->heapIndex < pdevice->memory.heap_count);
3111 if (pdevice->memory.heaps[mem->type->heapIndex].supports_48bit_addresses)
3112 bo_flags |= EXEC_OBJECT_SUPPORTS_48B_ADDRESS;
3113
3114 const struct wsi_memory_allocate_info *wsi_info =
3115 vk_find_struct_const(pAllocateInfo->pNext, WSI_MEMORY_ALLOCATE_INFO_MESA);
3116 if (wsi_info && wsi_info->implicit_sync) {
3117 /* We need to set the WRITE flag on window system buffers so that GEM
3118 * will know we're writing to them and synchronize uses on other rings
3119 * (eg if the display server uses the blitter ring).
3120 */
3121 bo_flags |= EXEC_OBJECT_WRITE;
3122 } else if (pdevice->has_exec_async) {
3123 bo_flags |= EXEC_OBJECT_ASYNC;
3124 }
3125
3126 if (pdevice->use_softpin)
3127 bo_flags |= EXEC_OBJECT_PINNED;
3128
3129 const VkExportMemoryAllocateInfo *export_info =
3130 vk_find_struct_const(pAllocateInfo->pNext, EXPORT_MEMORY_ALLOCATE_INFO);
3131
3132 /* Check if we need to support Android HW buffer export. If so,
3133 * create AHardwareBuffer and import memory from it.
3134 */
3135 bool android_export = false;
3136 if (export_info && export_info->handleTypes &
3137 VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID)
3138 android_export = true;
3139
3140 /* Android memory import. */
3141 const struct VkImportAndroidHardwareBufferInfoANDROID *ahw_import_info =
3142 vk_find_struct_const(pAllocateInfo->pNext,
3143 IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID);
3144
3145 if (ahw_import_info) {
3146 result = anv_import_ahw_memory(_device, mem, ahw_import_info);
3147 if (result != VK_SUCCESS)
3148 goto fail;
3149
3150 goto success;
3151 } else if (android_export) {
3152 result = anv_create_ahw_memory(_device, mem, pAllocateInfo);
3153 if (result != VK_SUCCESS)
3154 goto fail;
3155
3156 const struct VkImportAndroidHardwareBufferInfoANDROID import_info = {
3157 .buffer = mem->ahw,
3158 };
3159 result = anv_import_ahw_memory(_device, mem, &import_info);
3160 if (result != VK_SUCCESS)
3161 goto fail;
3162
3163 goto success;
3164 }
3165
3166 const VkImportMemoryFdInfoKHR *fd_info =
3167 vk_find_struct_const(pAllocateInfo->pNext, IMPORT_MEMORY_FD_INFO_KHR);
3168
3169 /* The Vulkan spec permits handleType to be 0, in which case the struct is
3170 * ignored.
3171 */
3172 if (fd_info && fd_info->handleType) {
3173 /* At the moment, we support only the below handle types. */
3174 assert(fd_info->handleType ==
3175 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
3176 fd_info->handleType ==
3177 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
3178
3179 result = anv_bo_cache_import(device, &device->bo_cache, fd_info->fd,
3180 bo_flags, &mem->bo);
3181 if (result != VK_SUCCESS)
3182 goto fail;
3183
3184 VkDeviceSize aligned_alloc_size =
3185 align_u64(pAllocateInfo->allocationSize, 4096);
3186
3187 /* For security purposes, we reject importing the bo if it's smaller
3188 * than the requested allocation size. This prevents a malicious client
3189 * from passing a buffer to a trusted client, lying about the size, and
3190 * telling the trusted client to try and texture from an image that goes
3191 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
3192 * in the trusted client. The trusted client can protect itself against
3193 * this sort of attack but only if it can trust the buffer size.
3194 */
3195 if (mem->bo->size < aligned_alloc_size) {
3196 result = vk_errorf(device->instance, device,
3197 VK_ERROR_INVALID_EXTERNAL_HANDLE,
3198 "aligned allocationSize too large for "
3199 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT: "
3200 "%"PRIu64"B > %"PRIu64"B",
3201 aligned_alloc_size, mem->bo->size);
3202 anv_bo_cache_release(device, &device->bo_cache, mem->bo);
3203 goto fail;
3204 }
3205
3206 /* From the Vulkan spec:
3207 *
3208 * "Importing memory from a file descriptor transfers ownership of
3209 * the file descriptor from the application to the Vulkan
3210 * implementation. The application must not perform any operations on
3211 * the file descriptor after a successful import."
3212 *
3213 * If the import fails, we leave the file descriptor open.
3214 */
3215 close(fd_info->fd);
3216 goto success;
3217 }
3218
3219 const VkImportMemoryHostPointerInfoEXT *host_ptr_info =
3220 vk_find_struct_const(pAllocateInfo->pNext,
3221 IMPORT_MEMORY_HOST_POINTER_INFO_EXT);
3222 if (host_ptr_info && host_ptr_info->handleType) {
3223 if (host_ptr_info->handleType ==
3224 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_MAPPED_FOREIGN_MEMORY_BIT_EXT) {
3225 result = vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE);
3226 goto fail;
3227 }
3228
3229 assert(host_ptr_info->handleType ==
3230 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT);
3231
3232 result = anv_bo_cache_import_host_ptr(
3233 device, &device->bo_cache, host_ptr_info->pHostPointer,
3234 pAllocateInfo->allocationSize, bo_flags, &mem->bo);
3235
3236 if (result != VK_SUCCESS)
3237 goto fail;
3238
3239 mem->host_ptr = host_ptr_info->pHostPointer;
3240 goto success;
3241 }
3242
3243 /* Regular allocate (not importing memory). */
3244
3245 bool is_external = export_info && export_info->handleTypes;
3246 result = anv_bo_cache_alloc(device, &device->bo_cache,
3247 pAllocateInfo->allocationSize,
3248 bo_flags, is_external,
3249 &mem->bo);
3250 if (result != VK_SUCCESS)
3251 goto fail;
3252
3253 const VkMemoryDedicatedAllocateInfo *dedicated_info =
3254 vk_find_struct_const(pAllocateInfo->pNext, MEMORY_DEDICATED_ALLOCATE_INFO);
3255 if (dedicated_info && dedicated_info->image != VK_NULL_HANDLE) {
3256 ANV_FROM_HANDLE(anv_image, image, dedicated_info->image);
3257
3258 /* Some legacy (non-modifiers) consumers need the tiling to be set on
3259 * the BO. In this case, we have a dedicated allocation.
3260 */
3261 if (image->needs_set_tiling) {
3262 const uint32_t i915_tiling =
3263 isl_tiling_to_i915_tiling(image->planes[0].surface.isl.tiling);
3264 int ret = anv_gem_set_tiling(device, mem->bo->gem_handle,
3265 image->planes[0].surface.isl.row_pitch_B,
3266 i915_tiling);
3267 if (ret) {
3268 anv_bo_cache_release(device, &device->bo_cache, mem->bo);
3269 return vk_errorf(device->instance, NULL,
3270 VK_ERROR_OUT_OF_DEVICE_MEMORY,
3271 "failed to set BO tiling: %m");
3272 }
3273 }
3274 }
3275
3276 success:
3277 pthread_mutex_lock(&device->mutex);
3278 list_addtail(&mem->link, &device->memory_objects);
3279 pthread_mutex_unlock(&device->mutex);
3280
3281 *pMem = anv_device_memory_to_handle(mem);
3282
3283 p_atomic_add(&pdevice->memory.heaps[mem->type->heapIndex].used,
3284 mem->bo->size);
3285
3286 return VK_SUCCESS;
3287
3288 fail:
3289 vk_free2(&device->alloc, pAllocator, mem);
3290
3291 return result;
3292 }
3293
3294 VkResult anv_GetMemoryFdKHR(
3295 VkDevice device_h,
3296 const VkMemoryGetFdInfoKHR* pGetFdInfo,
3297 int* pFd)
3298 {
3299 ANV_FROM_HANDLE(anv_device, dev, device_h);
3300 ANV_FROM_HANDLE(anv_device_memory, mem, pGetFdInfo->memory);
3301
3302 assert(pGetFdInfo->sType == VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR);
3303
3304 assert(pGetFdInfo->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
3305 pGetFdInfo->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
3306
3307 return anv_bo_cache_export(dev, &dev->bo_cache, mem->bo, pFd);
3308 }
3309
3310 VkResult anv_GetMemoryFdPropertiesKHR(
3311 VkDevice _device,
3312 VkExternalMemoryHandleTypeFlagBits handleType,
3313 int fd,
3314 VkMemoryFdPropertiesKHR* pMemoryFdProperties)
3315 {
3316 ANV_FROM_HANDLE(anv_device, device, _device);
3317 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3318
3319 switch (handleType) {
3320 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT:
3321 /* dma-buf can be imported as any memory type */
3322 pMemoryFdProperties->memoryTypeBits =
3323 (1 << pdevice->memory.type_count) - 1;
3324 return VK_SUCCESS;
3325
3326 default:
3327 /* The valid usage section for this function says:
3328 *
3329 * "handleType must not be one of the handle types defined as
3330 * opaque."
3331 *
3332 * So opaque handle types fall into the default "unsupported" case.
3333 */
3334 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE);
3335 }
3336 }
3337
3338 VkResult anv_GetMemoryHostPointerPropertiesEXT(
3339 VkDevice _device,
3340 VkExternalMemoryHandleTypeFlagBits handleType,
3341 const void* pHostPointer,
3342 VkMemoryHostPointerPropertiesEXT* pMemoryHostPointerProperties)
3343 {
3344 ANV_FROM_HANDLE(anv_device, device, _device);
3345
3346 assert(pMemoryHostPointerProperties->sType ==
3347 VK_STRUCTURE_TYPE_MEMORY_HOST_POINTER_PROPERTIES_EXT);
3348
3349 switch (handleType) {
3350 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT: {
3351 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3352
3353 /* Host memory can be imported as any memory type. */
3354 pMemoryHostPointerProperties->memoryTypeBits =
3355 (1ull << pdevice->memory.type_count) - 1;
3356
3357 return VK_SUCCESS;
3358 }
3359 default:
3360 return VK_ERROR_INVALID_EXTERNAL_HANDLE;
3361 }
3362 }
3363
3364 void anv_FreeMemory(
3365 VkDevice _device,
3366 VkDeviceMemory _mem,
3367 const VkAllocationCallbacks* pAllocator)
3368 {
3369 ANV_FROM_HANDLE(anv_device, device, _device);
3370 ANV_FROM_HANDLE(anv_device_memory, mem, _mem);
3371 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3372
3373 if (mem == NULL)
3374 return;
3375
3376 pthread_mutex_lock(&device->mutex);
3377 list_del(&mem->link);
3378 pthread_mutex_unlock(&device->mutex);
3379
3380 if (mem->map)
3381 anv_UnmapMemory(_device, _mem);
3382
3383 p_atomic_add(&pdevice->memory.heaps[mem->type->heapIndex].used,
3384 -mem->bo->size);
3385
3386 anv_bo_cache_release(device, &device->bo_cache, mem->bo);
3387
3388 #if defined(ANDROID) && ANDROID_API_LEVEL >= 26
3389 if (mem->ahw)
3390 AHardwareBuffer_release(mem->ahw);
3391 #endif
3392
3393 vk_free2(&device->alloc, pAllocator, mem);
3394 }
3395
3396 VkResult anv_MapMemory(
3397 VkDevice _device,
3398 VkDeviceMemory _memory,
3399 VkDeviceSize offset,
3400 VkDeviceSize size,
3401 VkMemoryMapFlags flags,
3402 void** ppData)
3403 {
3404 ANV_FROM_HANDLE(anv_device, device, _device);
3405 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
3406
3407 if (mem == NULL) {
3408 *ppData = NULL;
3409 return VK_SUCCESS;
3410 }
3411
3412 if (mem->host_ptr) {
3413 *ppData = mem->host_ptr + offset;
3414 return VK_SUCCESS;
3415 }
3416
3417 if (size == VK_WHOLE_SIZE)
3418 size = mem->bo->size - offset;
3419
3420 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
3421 *
3422 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
3423 * assert(size != 0);
3424 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
3425 * equal to the size of the memory minus offset
3426 */
3427 assert(size > 0);
3428 assert(offset + size <= mem->bo->size);
3429
3430 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
3431 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
3432 * at a time is valid. We could just mmap up front and return an offset
3433 * pointer here, but that may exhaust virtual memory on 32 bit
3434 * userspace. */
3435
3436 uint32_t gem_flags = 0;
3437
3438 if (!device->info.has_llc &&
3439 (mem->type->propertyFlags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT))
3440 gem_flags |= I915_MMAP_WC;
3441
3442 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
3443 uint64_t map_offset = offset & ~4095ull;
3444 assert(offset >= map_offset);
3445 uint64_t map_size = (offset + size) - map_offset;
3446
3447 /* Let's map whole pages */
3448 map_size = align_u64(map_size, 4096);
3449
3450 void *map = anv_gem_mmap(device, mem->bo->gem_handle,
3451 map_offset, map_size, gem_flags);
3452 if (map == MAP_FAILED)
3453 return vk_error(VK_ERROR_MEMORY_MAP_FAILED);
3454
3455 mem->map = map;
3456 mem->map_size = map_size;
3457
3458 *ppData = mem->map + (offset - map_offset);
3459
3460 return VK_SUCCESS;
3461 }
3462
3463 void anv_UnmapMemory(
3464 VkDevice _device,
3465 VkDeviceMemory _memory)
3466 {
3467 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
3468
3469 if (mem == NULL || mem->host_ptr)
3470 return;
3471
3472 anv_gem_munmap(mem->map, mem->map_size);
3473
3474 mem->map = NULL;
3475 mem->map_size = 0;
3476 }
3477
3478 static void
3479 clflush_mapped_ranges(struct anv_device *device,
3480 uint32_t count,
3481 const VkMappedMemoryRange *ranges)
3482 {
3483 for (uint32_t i = 0; i < count; i++) {
3484 ANV_FROM_HANDLE(anv_device_memory, mem, ranges[i].memory);
3485 if (ranges[i].offset >= mem->map_size)
3486 continue;
3487
3488 gen_clflush_range(mem->map + ranges[i].offset,
3489 MIN2(ranges[i].size, mem->map_size - ranges[i].offset));
3490 }
3491 }
3492
3493 VkResult anv_FlushMappedMemoryRanges(
3494 VkDevice _device,
3495 uint32_t memoryRangeCount,
3496 const VkMappedMemoryRange* pMemoryRanges)
3497 {
3498 ANV_FROM_HANDLE(anv_device, device, _device);
3499
3500 if (device->info.has_llc)
3501 return VK_SUCCESS;
3502
3503 /* Make sure the writes we're flushing have landed. */
3504 __builtin_ia32_mfence();
3505
3506 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
3507
3508 return VK_SUCCESS;
3509 }
3510
3511 VkResult anv_InvalidateMappedMemoryRanges(
3512 VkDevice _device,
3513 uint32_t memoryRangeCount,
3514 const VkMappedMemoryRange* pMemoryRanges)
3515 {
3516 ANV_FROM_HANDLE(anv_device, device, _device);
3517
3518 if (device->info.has_llc)
3519 return VK_SUCCESS;
3520
3521 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
3522
3523 /* Make sure no reads get moved up above the invalidate. */
3524 __builtin_ia32_mfence();
3525
3526 return VK_SUCCESS;
3527 }
3528
3529 void anv_GetBufferMemoryRequirements(
3530 VkDevice _device,
3531 VkBuffer _buffer,
3532 VkMemoryRequirements* pMemoryRequirements)
3533 {
3534 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
3535 ANV_FROM_HANDLE(anv_device, device, _device);
3536 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3537
3538 /* The Vulkan spec (git aaed022) says:
3539 *
3540 * memoryTypeBits is a bitfield and contains one bit set for every
3541 * supported memory type for the resource. The bit `1<<i` is set if and
3542 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3543 * structure for the physical device is supported.
3544 */
3545 uint32_t memory_types = 0;
3546 for (uint32_t i = 0; i < pdevice->memory.type_count; i++) {
3547 uint32_t valid_usage = pdevice->memory.types[i].valid_buffer_usage;
3548 if ((valid_usage & buffer->usage) == buffer->usage)
3549 memory_types |= (1u << i);
3550 }
3551
3552 /* Base alignment requirement of a cache line */
3553 uint32_t alignment = 16;
3554
3555 /* We need an alignment of 32 for pushing UBOs */
3556 if (buffer->usage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT)
3557 alignment = MAX2(alignment, 32);
3558
3559 pMemoryRequirements->size = buffer->size;
3560 pMemoryRequirements->alignment = alignment;
3561
3562 /* Storage and Uniform buffers should have their size aligned to
3563 * 32-bits to avoid boundary checks when last DWord is not complete.
3564 * This would ensure that not internal padding would be needed for
3565 * 16-bit types.
3566 */
3567 if (device->robust_buffer_access &&
3568 (buffer->usage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT ||
3569 buffer->usage & VK_BUFFER_USAGE_STORAGE_BUFFER_BIT))
3570 pMemoryRequirements->size = align_u64(buffer->size, 4);
3571
3572 pMemoryRequirements->memoryTypeBits = memory_types;
3573 }
3574
3575 void anv_GetBufferMemoryRequirements2(
3576 VkDevice _device,
3577 const VkBufferMemoryRequirementsInfo2* pInfo,
3578 VkMemoryRequirements2* pMemoryRequirements)
3579 {
3580 anv_GetBufferMemoryRequirements(_device, pInfo->buffer,
3581 &pMemoryRequirements->memoryRequirements);
3582
3583 vk_foreach_struct(ext, pMemoryRequirements->pNext) {
3584 switch (ext->sType) {
3585 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
3586 VkMemoryDedicatedRequirements *requirements = (void *)ext;
3587 requirements->prefersDedicatedAllocation = false;
3588 requirements->requiresDedicatedAllocation = false;
3589 break;
3590 }
3591
3592 default:
3593 anv_debug_ignored_stype(ext->sType);
3594 break;
3595 }
3596 }
3597 }
3598
3599 void anv_GetImageMemoryRequirements(
3600 VkDevice _device,
3601 VkImage _image,
3602 VkMemoryRequirements* pMemoryRequirements)
3603 {
3604 ANV_FROM_HANDLE(anv_image, image, _image);
3605 ANV_FROM_HANDLE(anv_device, device, _device);
3606 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3607
3608 /* The Vulkan spec (git aaed022) says:
3609 *
3610 * memoryTypeBits is a bitfield and contains one bit set for every
3611 * supported memory type for the resource. The bit `1<<i` is set if and
3612 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3613 * structure for the physical device is supported.
3614 *
3615 * All types are currently supported for images.
3616 */
3617 uint32_t memory_types = (1ull << pdevice->memory.type_count) - 1;
3618
3619 /* We must have image allocated or imported at this point. According to the
3620 * specification, external images must have been bound to memory before
3621 * calling GetImageMemoryRequirements.
3622 */
3623 assert(image->size > 0);
3624
3625 pMemoryRequirements->size = image->size;
3626 pMemoryRequirements->alignment = image->alignment;
3627 pMemoryRequirements->memoryTypeBits = memory_types;
3628 }
3629
3630 void anv_GetImageMemoryRequirements2(
3631 VkDevice _device,
3632 const VkImageMemoryRequirementsInfo2* pInfo,
3633 VkMemoryRequirements2* pMemoryRequirements)
3634 {
3635 ANV_FROM_HANDLE(anv_device, device, _device);
3636 ANV_FROM_HANDLE(anv_image, image, pInfo->image);
3637
3638 anv_GetImageMemoryRequirements(_device, pInfo->image,
3639 &pMemoryRequirements->memoryRequirements);
3640
3641 vk_foreach_struct_const(ext, pInfo->pNext) {
3642 switch (ext->sType) {
3643 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO: {
3644 struct anv_physical_device *pdevice = &device->instance->physicalDevice;
3645 const VkImagePlaneMemoryRequirementsInfo *plane_reqs =
3646 (const VkImagePlaneMemoryRequirementsInfo *) ext;
3647 uint32_t plane = anv_image_aspect_to_plane(image->aspects,
3648 plane_reqs->planeAspect);
3649
3650 assert(image->planes[plane].offset == 0);
3651
3652 /* The Vulkan spec (git aaed022) says:
3653 *
3654 * memoryTypeBits is a bitfield and contains one bit set for every
3655 * supported memory type for the resource. The bit `1<<i` is set
3656 * if and only if the memory type `i` in the
3657 * VkPhysicalDeviceMemoryProperties structure for the physical
3658 * device is supported.
3659 *
3660 * All types are currently supported for images.
3661 */
3662 pMemoryRequirements->memoryRequirements.memoryTypeBits =
3663 (1ull << pdevice->memory.type_count) - 1;
3664
3665 /* We must have image allocated or imported at this point. According to the
3666 * specification, external images must have been bound to memory before
3667 * calling GetImageMemoryRequirements.
3668 */
3669 assert(image->planes[plane].size > 0);
3670
3671 pMemoryRequirements->memoryRequirements.size = image->planes[plane].size;
3672 pMemoryRequirements->memoryRequirements.alignment =
3673 image->planes[plane].alignment;
3674 break;
3675 }
3676
3677 default:
3678 anv_debug_ignored_stype(ext->sType);
3679 break;
3680 }
3681 }
3682
3683 vk_foreach_struct(ext, pMemoryRequirements->pNext) {
3684 switch (ext->sType) {
3685 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
3686 VkMemoryDedicatedRequirements *requirements = (void *)ext;
3687 if (image->needs_set_tiling || image->external_format) {
3688 /* If we need to set the tiling for external consumers, we need a
3689 * dedicated allocation.
3690 *
3691 * See also anv_AllocateMemory.
3692 */
3693 requirements->prefersDedicatedAllocation = true;
3694 requirements->requiresDedicatedAllocation = true;
3695 } else {
3696 requirements->prefersDedicatedAllocation = false;
3697 requirements->requiresDedicatedAllocation = false;
3698 }
3699 break;
3700 }
3701
3702 default:
3703 anv_debug_ignored_stype(ext->sType);
3704 break;
3705 }
3706 }
3707 }
3708
3709 void anv_GetImageSparseMemoryRequirements(
3710 VkDevice device,
3711 VkImage image,
3712 uint32_t* pSparseMemoryRequirementCount,
3713 VkSparseImageMemoryRequirements* pSparseMemoryRequirements)
3714 {
3715 *pSparseMemoryRequirementCount = 0;
3716 }
3717
3718 void anv_GetImageSparseMemoryRequirements2(
3719 VkDevice device,
3720 const VkImageSparseMemoryRequirementsInfo2* pInfo,
3721 uint32_t* pSparseMemoryRequirementCount,
3722 VkSparseImageMemoryRequirements2* pSparseMemoryRequirements)
3723 {
3724 *pSparseMemoryRequirementCount = 0;
3725 }
3726
3727 void anv_GetDeviceMemoryCommitment(
3728 VkDevice device,
3729 VkDeviceMemory memory,
3730 VkDeviceSize* pCommittedMemoryInBytes)
3731 {
3732 *pCommittedMemoryInBytes = 0;
3733 }
3734
3735 static void
3736 anv_bind_buffer_memory(const VkBindBufferMemoryInfo *pBindInfo)
3737 {
3738 ANV_FROM_HANDLE(anv_device_memory, mem, pBindInfo->memory);
3739 ANV_FROM_HANDLE(anv_buffer, buffer, pBindInfo->buffer);
3740
3741 assert(pBindInfo->sType == VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO);
3742
3743 if (mem) {
3744 assert((buffer->usage & mem->type->valid_buffer_usage) == buffer->usage);
3745 buffer->address = (struct anv_address) {
3746 .bo = mem->bo,
3747 .offset = pBindInfo->memoryOffset,
3748 };
3749 } else {
3750 buffer->address = ANV_NULL_ADDRESS;
3751 }
3752 }
3753
3754 VkResult anv_BindBufferMemory(
3755 VkDevice device,
3756 VkBuffer buffer,
3757 VkDeviceMemory memory,
3758 VkDeviceSize memoryOffset)
3759 {
3760 anv_bind_buffer_memory(
3761 &(VkBindBufferMemoryInfo) {
3762 .sType = VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO,
3763 .buffer = buffer,
3764 .memory = memory,
3765 .memoryOffset = memoryOffset,
3766 });
3767
3768 return VK_SUCCESS;
3769 }
3770
3771 VkResult anv_BindBufferMemory2(
3772 VkDevice device,
3773 uint32_t bindInfoCount,
3774 const VkBindBufferMemoryInfo* pBindInfos)
3775 {
3776 for (uint32_t i = 0; i < bindInfoCount; i++)
3777 anv_bind_buffer_memory(&pBindInfos[i]);
3778
3779 return VK_SUCCESS;
3780 }
3781
3782 VkResult anv_QueueBindSparse(
3783 VkQueue _queue,
3784 uint32_t bindInfoCount,
3785 const VkBindSparseInfo* pBindInfo,
3786 VkFence fence)
3787 {
3788 ANV_FROM_HANDLE(anv_queue, queue, _queue);
3789 if (anv_device_is_lost(queue->device))
3790 return VK_ERROR_DEVICE_LOST;
3791
3792 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT);
3793 }
3794
3795 // Event functions
3796
3797 VkResult anv_CreateEvent(
3798 VkDevice _device,
3799 const VkEventCreateInfo* pCreateInfo,
3800 const VkAllocationCallbacks* pAllocator,
3801 VkEvent* pEvent)
3802 {
3803 ANV_FROM_HANDLE(anv_device, device, _device);
3804 struct anv_state state;
3805 struct anv_event *event;
3806
3807 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_EVENT_CREATE_INFO);
3808
3809 state = anv_state_pool_alloc(&device->dynamic_state_pool,
3810 sizeof(*event), 8);
3811 event = state.map;
3812 event->state = state;
3813 event->semaphore = VK_EVENT_RESET;
3814
3815 if (!device->info.has_llc) {
3816 /* Make sure the writes we're flushing have landed. */
3817 __builtin_ia32_mfence();
3818 __builtin_ia32_clflush(event);
3819 }
3820
3821 *pEvent = anv_event_to_handle(event);
3822
3823 return VK_SUCCESS;
3824 }
3825
3826 void anv_DestroyEvent(
3827 VkDevice _device,
3828 VkEvent _event,
3829 const VkAllocationCallbacks* pAllocator)
3830 {
3831 ANV_FROM_HANDLE(anv_device, device, _device);
3832 ANV_FROM_HANDLE(anv_event, event, _event);
3833
3834 if (!event)
3835 return;
3836
3837 anv_state_pool_free(&device->dynamic_state_pool, event->state);
3838 }
3839
3840 VkResult anv_GetEventStatus(
3841 VkDevice _device,
3842 VkEvent _event)
3843 {
3844 ANV_FROM_HANDLE(anv_device, device, _device);
3845 ANV_FROM_HANDLE(anv_event, event, _event);
3846
3847 if (anv_device_is_lost(device))
3848 return VK_ERROR_DEVICE_LOST;
3849
3850 if (!device->info.has_llc) {
3851 /* Invalidate read cache before reading event written by GPU. */
3852 __builtin_ia32_clflush(event);
3853 __builtin_ia32_mfence();
3854
3855 }
3856
3857 return event->semaphore;
3858 }
3859
3860 VkResult anv_SetEvent(
3861 VkDevice _device,
3862 VkEvent _event)
3863 {
3864 ANV_FROM_HANDLE(anv_device, device, _device);
3865 ANV_FROM_HANDLE(anv_event, event, _event);
3866
3867 event->semaphore = VK_EVENT_SET;
3868
3869 if (!device->info.has_llc) {
3870 /* Make sure the writes we're flushing have landed. */
3871 __builtin_ia32_mfence();
3872 __builtin_ia32_clflush(event);
3873 }
3874
3875 return VK_SUCCESS;
3876 }
3877
3878 VkResult anv_ResetEvent(
3879 VkDevice _device,
3880 VkEvent _event)
3881 {
3882 ANV_FROM_HANDLE(anv_device, device, _device);
3883 ANV_FROM_HANDLE(anv_event, event, _event);
3884
3885 event->semaphore = VK_EVENT_RESET;
3886
3887 if (!device->info.has_llc) {
3888 /* Make sure the writes we're flushing have landed. */
3889 __builtin_ia32_mfence();
3890 __builtin_ia32_clflush(event);
3891 }
3892
3893 return VK_SUCCESS;
3894 }
3895
3896 // Buffer functions
3897
3898 VkResult anv_CreateBuffer(
3899 VkDevice _device,
3900 const VkBufferCreateInfo* pCreateInfo,
3901 const VkAllocationCallbacks* pAllocator,
3902 VkBuffer* pBuffer)
3903 {
3904 ANV_FROM_HANDLE(anv_device, device, _device);
3905 struct anv_buffer *buffer;
3906
3907 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
3908
3909 buffer = vk_alloc2(&device->alloc, pAllocator, sizeof(*buffer), 8,
3910 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
3911 if (buffer == NULL)
3912 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
3913
3914 buffer->size = pCreateInfo->size;
3915 buffer->usage = pCreateInfo->usage;
3916 buffer->address = ANV_NULL_ADDRESS;
3917
3918 *pBuffer = anv_buffer_to_handle(buffer);
3919
3920 return VK_SUCCESS;
3921 }
3922
3923 void anv_DestroyBuffer(
3924 VkDevice _device,
3925 VkBuffer _buffer,
3926 const VkAllocationCallbacks* pAllocator)
3927 {
3928 ANV_FROM_HANDLE(anv_device, device, _device);
3929 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
3930
3931 if (!buffer)
3932 return;
3933
3934 vk_free2(&device->alloc, pAllocator, buffer);
3935 }
3936
3937 VkDeviceAddress anv_GetBufferDeviceAddressEXT(
3938 VkDevice device,
3939 const VkBufferDeviceAddressInfoEXT* pInfo)
3940 {
3941 ANV_FROM_HANDLE(anv_buffer, buffer, pInfo->buffer);
3942
3943 assert(buffer->address.bo->flags & EXEC_OBJECT_PINNED);
3944
3945 return anv_address_physical(buffer->address);
3946 }
3947
3948 void
3949 anv_fill_buffer_surface_state(struct anv_device *device, struct anv_state state,
3950 enum isl_format format,
3951 struct anv_address address,
3952 uint32_t range, uint32_t stride)
3953 {
3954 isl_buffer_fill_state(&device->isl_dev, state.map,
3955 .address = anv_address_physical(address),
3956 .mocs = device->default_mocs,
3957 .size_B = range,
3958 .format = format,
3959 .swizzle = ISL_SWIZZLE_IDENTITY,
3960 .stride_B = stride);
3961 }
3962
3963 void anv_DestroySampler(
3964 VkDevice _device,
3965 VkSampler _sampler,
3966 const VkAllocationCallbacks* pAllocator)
3967 {
3968 ANV_FROM_HANDLE(anv_device, device, _device);
3969 ANV_FROM_HANDLE(anv_sampler, sampler, _sampler);
3970
3971 if (!sampler)
3972 return;
3973
3974 if (sampler->bindless_state.map) {
3975 anv_state_pool_free(&device->dynamic_state_pool,
3976 sampler->bindless_state);
3977 }
3978
3979 vk_free2(&device->alloc, pAllocator, sampler);
3980 }
3981
3982 VkResult anv_CreateFramebuffer(
3983 VkDevice _device,
3984 const VkFramebufferCreateInfo* pCreateInfo,
3985 const VkAllocationCallbacks* pAllocator,
3986 VkFramebuffer* pFramebuffer)
3987 {
3988 ANV_FROM_HANDLE(anv_device, device, _device);
3989 struct anv_framebuffer *framebuffer;
3990
3991 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
3992
3993 size_t size = sizeof(*framebuffer);
3994
3995 /* VK_KHR_imageless_framebuffer extension says:
3996 *
3997 * If flags includes VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR,
3998 * parameter pAttachments is ignored.
3999 */
4000 if (!(pCreateInfo->flags & VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR)) {
4001 size += sizeof(struct anv_image_view *) * pCreateInfo->attachmentCount;
4002 framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
4003 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
4004 if (framebuffer == NULL)
4005 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
4006
4007 for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
4008 ANV_FROM_HANDLE(anv_image_view, iview, pCreateInfo->pAttachments[i]);
4009 framebuffer->attachments[i] = iview;
4010 }
4011 framebuffer->attachment_count = pCreateInfo->attachmentCount;
4012 } else {
4013 assert(device->enabled_extensions.KHR_imageless_framebuffer);
4014 framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
4015 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
4016 if (framebuffer == NULL)
4017 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
4018
4019 framebuffer->attachment_count = 0;
4020 }
4021
4022 framebuffer->width = pCreateInfo->width;
4023 framebuffer->height = pCreateInfo->height;
4024 framebuffer->layers = pCreateInfo->layers;
4025
4026 *pFramebuffer = anv_framebuffer_to_handle(framebuffer);
4027
4028 return VK_SUCCESS;
4029 }
4030
4031 void anv_DestroyFramebuffer(
4032 VkDevice _device,
4033 VkFramebuffer _fb,
4034 const VkAllocationCallbacks* pAllocator)
4035 {
4036 ANV_FROM_HANDLE(anv_device, device, _device);
4037 ANV_FROM_HANDLE(anv_framebuffer, fb, _fb);
4038
4039 if (!fb)
4040 return;
4041
4042 vk_free2(&device->alloc, pAllocator, fb);
4043 }
4044
4045 static const VkTimeDomainEXT anv_time_domains[] = {
4046 VK_TIME_DOMAIN_DEVICE_EXT,
4047 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT,
4048 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT,
4049 };
4050
4051 VkResult anv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
4052 VkPhysicalDevice physicalDevice,
4053 uint32_t *pTimeDomainCount,
4054 VkTimeDomainEXT *pTimeDomains)
4055 {
4056 int d;
4057 VK_OUTARRAY_MAKE(out, pTimeDomains, pTimeDomainCount);
4058
4059 for (d = 0; d < ARRAY_SIZE(anv_time_domains); d++) {
4060 vk_outarray_append(&out, i) {
4061 *i = anv_time_domains[d];
4062 }
4063 }
4064
4065 return vk_outarray_status(&out);
4066 }
4067
4068 static uint64_t
4069 anv_clock_gettime(clockid_t clock_id)
4070 {
4071 struct timespec current;
4072 int ret;
4073
4074 ret = clock_gettime(clock_id, &current);
4075 if (ret < 0 && clock_id == CLOCK_MONOTONIC_RAW)
4076 ret = clock_gettime(CLOCK_MONOTONIC, &current);
4077 if (ret < 0)
4078 return 0;
4079
4080 return (uint64_t) current.tv_sec * 1000000000ULL + current.tv_nsec;
4081 }
4082
4083 #define TIMESTAMP 0x2358
4084
4085 VkResult anv_GetCalibratedTimestampsEXT(
4086 VkDevice _device,
4087 uint32_t timestampCount,
4088 const VkCalibratedTimestampInfoEXT *pTimestampInfos,
4089 uint64_t *pTimestamps,
4090 uint64_t *pMaxDeviation)
4091 {
4092 ANV_FROM_HANDLE(anv_device, device, _device);
4093 uint64_t timestamp_frequency = device->info.timestamp_frequency;
4094 int ret;
4095 int d;
4096 uint64_t begin, end;
4097 uint64_t max_clock_period = 0;
4098
4099 begin = anv_clock_gettime(CLOCK_MONOTONIC_RAW);
4100
4101 for (d = 0; d < timestampCount; d++) {
4102 switch (pTimestampInfos[d].timeDomain) {
4103 case VK_TIME_DOMAIN_DEVICE_EXT:
4104 ret = anv_gem_reg_read(device, TIMESTAMP | 1,
4105 &pTimestamps[d]);
4106
4107 if (ret != 0) {
4108 return anv_device_set_lost(device, "Failed to read the TIMESTAMP "
4109 "register: %m");
4110 }
4111 uint64_t device_period = DIV_ROUND_UP(1000000000, timestamp_frequency);
4112 max_clock_period = MAX2(max_clock_period, device_period);
4113 break;
4114 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT:
4115 pTimestamps[d] = anv_clock_gettime(CLOCK_MONOTONIC);
4116 max_clock_period = MAX2(max_clock_period, 1);
4117 break;
4118
4119 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT:
4120 pTimestamps[d] = begin;
4121 break;
4122 default:
4123 pTimestamps[d] = 0;
4124 break;
4125 }
4126 }
4127
4128 end = anv_clock_gettime(CLOCK_MONOTONIC_RAW);
4129
4130 /*
4131 * The maximum deviation is the sum of the interval over which we
4132 * perform the sampling and the maximum period of any sampled
4133 * clock. That's because the maximum skew between any two sampled
4134 * clock edges is when the sampled clock with the largest period is
4135 * sampled at the end of that period but right at the beginning of the
4136 * sampling interval and some other clock is sampled right at the
4137 * begining of its sampling period and right at the end of the
4138 * sampling interval. Let's assume the GPU has the longest clock
4139 * period and that the application is sampling GPU and monotonic:
4140 *
4141 * s e
4142 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
4143 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4144 *
4145 * g
4146 * 0 1 2 3
4147 * GPU -----_____-----_____-----_____-----_____
4148 *
4149 * m
4150 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
4151 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4152 *
4153 * Interval <----------------->
4154 * Deviation <-------------------------->
4155 *
4156 * s = read(raw) 2
4157 * g = read(GPU) 1
4158 * m = read(monotonic) 2
4159 * e = read(raw) b
4160 *
4161 * We round the sample interval up by one tick to cover sampling error
4162 * in the interval clock
4163 */
4164
4165 uint64_t sample_interval = end - begin + 1;
4166
4167 *pMaxDeviation = sample_interval + max_clock_period;
4168
4169 return VK_SUCCESS;
4170 }
4171
4172 /* vk_icd.h does not declare this function, so we declare it here to
4173 * suppress Wmissing-prototypes.
4174 */
4175 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4176 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion);
4177
4178 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4179 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion)
4180 {
4181 /* For the full details on loader interface versioning, see
4182 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
4183 * What follows is a condensed summary, to help you navigate the large and
4184 * confusing official doc.
4185 *
4186 * - Loader interface v0 is incompatible with later versions. We don't
4187 * support it.
4188 *
4189 * - In loader interface v1:
4190 * - The first ICD entrypoint called by the loader is
4191 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
4192 * entrypoint.
4193 * - The ICD must statically expose no other Vulkan symbol unless it is
4194 * linked with -Bsymbolic.
4195 * - Each dispatchable Vulkan handle created by the ICD must be
4196 * a pointer to a struct whose first member is VK_LOADER_DATA. The
4197 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
4198 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
4199 * vkDestroySurfaceKHR(). The ICD must be capable of working with
4200 * such loader-managed surfaces.
4201 *
4202 * - Loader interface v2 differs from v1 in:
4203 * - The first ICD entrypoint called by the loader is
4204 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
4205 * statically expose this entrypoint.
4206 *
4207 * - Loader interface v3 differs from v2 in:
4208 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
4209 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
4210 * because the loader no longer does so.
4211 *
4212 * - Loader interface v4 differs from v3 in:
4213 * - The ICD must implement vk_icdGetPhysicalDeviceProcAddr().
4214 */
4215 *pSupportedVersion = MIN2(*pSupportedVersion, 4u);
4216 return VK_SUCCESS;
4217 }