2 * Copyright © 2015 Intel Corporation
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:
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
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
28 #include <sys/sysinfo.h>
32 #include "drm-uapi/drm_fourcc.h"
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"
45 #include "common/gen_aux_map.h"
46 #include "common/gen_defines.h"
47 #include "compiler/glsl_types.h"
49 #include "genxml/gen7_pack.h"
51 static const char anv_dri_options_xml
[] =
53 DRI_CONF_SECTION_PERFORMANCE
54 DRI_CONF_VK_X11_OVERRIDE_MIN_IMAGE_COUNT(0)
55 DRI_CONF_VK_X11_STRICT_IMAGE_COUNT("false")
58 DRI_CONF_SECTION_DEBUG
59 DRI_CONF_ALWAYS_FLUSH_CACHE("false")
63 /* This is probably far to big but it reflects the max size used for messages
64 * in OpenGLs KHR_debug.
66 #define MAX_DEBUG_MESSAGE_LENGTH 4096
69 compiler_debug_log(void *data
, const char *fmt
, ...)
71 char str
[MAX_DEBUG_MESSAGE_LENGTH
];
72 struct anv_device
*device
= (struct anv_device
*)data
;
74 if (list_is_empty(&device
->instance
->debug_report_callbacks
.callbacks
))
79 (void) vsnprintf(str
, MAX_DEBUG_MESSAGE_LENGTH
, fmt
, args
);
82 vk_debug_report(&device
->instance
->debug_report_callbacks
,
83 VK_DEBUG_REPORT_DEBUG_BIT_EXT
,
84 VK_DEBUG_REPORT_OBJECT_TYPE_UNKNOWN_EXT
,
89 compiler_perf_log(void *data
, const char *fmt
, ...)
94 if (unlikely(INTEL_DEBUG
& DEBUG_PERF
))
95 intel_logd_v(fmt
, args
);
101 anv_compute_heap_size(int fd
, uint64_t gtt_size
)
103 /* Query the total ram from the system */
107 uint64_t total_ram
= (uint64_t)info
.totalram
* (uint64_t)info
.mem_unit
;
109 /* We don't want to burn too much ram with the GPU. If the user has 4GiB
110 * or less, we use at most half. If they have more than 4GiB, we use 3/4.
112 uint64_t available_ram
;
113 if (total_ram
<= 4ull * 1024ull * 1024ull * 1024ull)
114 available_ram
= total_ram
/ 2;
116 available_ram
= total_ram
* 3 / 4;
118 /* We also want to leave some padding for things we allocate in the driver,
119 * so don't go over 3/4 of the GTT either.
121 uint64_t available_gtt
= gtt_size
* 3 / 4;
123 return MIN2(available_ram
, available_gtt
);
127 anv_physical_device_init_heaps(struct anv_physical_device
*device
, int fd
)
129 if (anv_gem_get_context_param(fd
, 0, I915_CONTEXT_PARAM_GTT_SIZE
,
130 &device
->gtt_size
) == -1) {
131 /* If, for whatever reason, we can't actually get the GTT size from the
132 * kernel (too old?) fall back to the aperture size.
134 anv_perf_warn(NULL
, NULL
,
135 "Failed to get I915_CONTEXT_PARAM_GTT_SIZE: %m");
137 if (anv_gem_get_aperture(fd
, &device
->gtt_size
) == -1) {
138 return vk_errorf(NULL
, NULL
, VK_ERROR_INITIALIZATION_FAILED
,
139 "failed to get aperture size: %m");
143 /* We only allow 48-bit addresses with softpin because knowing the actual
144 * address is required for the vertex cache flush workaround.
146 device
->supports_48bit_addresses
= (device
->info
.gen
>= 8) &&
147 device
->has_softpin
&&
148 device
->gtt_size
> (4ULL << 30 /* GiB */);
150 uint64_t heap_size
= anv_compute_heap_size(fd
, device
->gtt_size
);
152 if (heap_size
> (2ull << 30) && !device
->supports_48bit_addresses
) {
153 /* When running with an overridden PCI ID, we may get a GTT size from
154 * the kernel that is greater than 2 GiB but the execbuf check for 48bit
155 * address support can still fail. Just clamp the address space size to
156 * 2 GiB if we don't have 48-bit support.
158 intel_logw("%s:%d: The kernel reported a GTT size larger than 2 GiB but "
159 "not support for 48-bit addresses",
161 heap_size
= 2ull << 30;
164 device
->memory
.heap_count
= 1;
165 device
->memory
.heaps
[0] = (struct anv_memory_heap
) {
167 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
170 uint32_t type_count
= 0;
171 for (uint32_t heap
= 0; heap
< device
->memory
.heap_count
; heap
++) {
172 if (device
->info
.has_llc
) {
173 /* Big core GPUs share LLC with the CPU and thus one memory type can be
174 * both cached and coherent at the same time.
176 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
177 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
178 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
179 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
180 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
184 /* The spec requires that we expose a host-visible, coherent memory
185 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
186 * to give the application a choice between cached, but not coherent and
187 * coherent but uncached (WC though).
189 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
190 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
191 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
192 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
195 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
196 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
197 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
198 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
203 device
->memory
.type_count
= type_count
;
209 anv_physical_device_init_uuids(struct anv_physical_device
*device
)
211 const struct build_id_note
*note
=
212 build_id_find_nhdr_for_addr(anv_physical_device_init_uuids
);
214 return vk_errorf(device
->instance
, device
,
215 VK_ERROR_INITIALIZATION_FAILED
,
216 "Failed to find build-id");
219 unsigned build_id_len
= build_id_length(note
);
220 if (build_id_len
< 20) {
221 return vk_errorf(device
->instance
, device
,
222 VK_ERROR_INITIALIZATION_FAILED
,
223 "build-id too short. It needs to be a SHA");
226 memcpy(device
->driver_build_sha1
, build_id_data(note
), 20);
228 struct mesa_sha1 sha1_ctx
;
230 STATIC_ASSERT(VK_UUID_SIZE
<= sizeof(sha1
));
232 /* The pipeline cache UUID is used for determining when a pipeline cache is
233 * invalid. It needs both a driver build and the PCI ID of the device.
235 _mesa_sha1_init(&sha1_ctx
);
236 _mesa_sha1_update(&sha1_ctx
, build_id_data(note
), build_id_len
);
237 _mesa_sha1_update(&sha1_ctx
, &device
->chipset_id
,
238 sizeof(device
->chipset_id
));
239 _mesa_sha1_update(&sha1_ctx
, &device
->always_use_bindless
,
240 sizeof(device
->always_use_bindless
));
241 _mesa_sha1_update(&sha1_ctx
, &device
->has_a64_buffer_access
,
242 sizeof(device
->has_a64_buffer_access
));
243 _mesa_sha1_update(&sha1_ctx
, &device
->has_bindless_images
,
244 sizeof(device
->has_bindless_images
));
245 _mesa_sha1_update(&sha1_ctx
, &device
->has_bindless_samplers
,
246 sizeof(device
->has_bindless_samplers
));
247 _mesa_sha1_final(&sha1_ctx
, sha1
);
248 memcpy(device
->pipeline_cache_uuid
, sha1
, VK_UUID_SIZE
);
250 /* The driver UUID is used for determining sharability of images and memory
251 * between two Vulkan instances in separate processes. People who want to
252 * share memory need to also check the device UUID (below) so all this
253 * needs to be is the build-id.
255 memcpy(device
->driver_uuid
, build_id_data(note
), VK_UUID_SIZE
);
257 /* The device UUID uniquely identifies the given device within the machine.
258 * Since we never have more than one device, this doesn't need to be a real
259 * UUID. However, on the off-chance that someone tries to use this to
260 * cache pre-tiled images or something of the like, we use the PCI ID and
261 * some bits of ISL info to ensure that this is safe.
263 _mesa_sha1_init(&sha1_ctx
);
264 _mesa_sha1_update(&sha1_ctx
, &device
->chipset_id
,
265 sizeof(device
->chipset_id
));
266 _mesa_sha1_update(&sha1_ctx
, &device
->isl_dev
.has_bit6_swizzling
,
267 sizeof(device
->isl_dev
.has_bit6_swizzling
));
268 _mesa_sha1_final(&sha1_ctx
, sha1
);
269 memcpy(device
->device_uuid
, sha1
, VK_UUID_SIZE
);
275 anv_physical_device_init_disk_cache(struct anv_physical_device
*device
)
277 #ifdef ENABLE_SHADER_CACHE
279 ASSERTED
int len
= snprintf(renderer
, sizeof(renderer
), "anv_%04x",
281 assert(len
== sizeof(renderer
) - 2);
284 _mesa_sha1_format(timestamp
, device
->driver_build_sha1
);
286 const uint64_t driver_flags
=
287 brw_get_compiler_config_value(device
->compiler
);
288 device
->disk_cache
= disk_cache_create(renderer
, timestamp
, driver_flags
);
290 device
->disk_cache
= NULL
;
295 anv_physical_device_free_disk_cache(struct anv_physical_device
*device
)
297 #ifdef ENABLE_SHADER_CACHE
298 if (device
->disk_cache
)
299 disk_cache_destroy(device
->disk_cache
);
301 assert(device
->disk_cache
== NULL
);
306 get_available_system_memory()
308 char *meminfo
= os_read_file("/proc/meminfo");
312 char *str
= strstr(meminfo
, "MemAvailable:");
318 uint64_t kb_mem_available
;
319 if (sscanf(str
, "MemAvailable: %" PRIx64
, &kb_mem_available
) == 1) {
321 return kb_mem_available
<< 10;
329 anv_physical_device_init(struct anv_physical_device
*device
,
330 struct anv_instance
*instance
,
331 drmDevicePtr drm_device
)
333 const char *primary_path
= drm_device
->nodes
[DRM_NODE_PRIMARY
];
334 const char *path
= drm_device
->nodes
[DRM_NODE_RENDER
];
339 brw_process_intel_debug_variable();
341 fd
= open(path
, O_RDWR
| O_CLOEXEC
);
343 return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
345 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
346 device
->instance
= instance
;
348 assert(strlen(path
) < ARRAY_SIZE(device
->path
));
349 snprintf(device
->path
, ARRAY_SIZE(device
->path
), "%s", path
);
351 if (!gen_get_device_info_from_fd(fd
, &device
->info
)) {
352 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
355 device
->chipset_id
= device
->info
.chipset_id
;
356 device
->no_hw
= device
->info
.no_hw
;
358 if (getenv("INTEL_NO_HW") != NULL
)
359 device
->no_hw
= true;
361 device
->pci_info
.domain
= drm_device
->businfo
.pci
->domain
;
362 device
->pci_info
.bus
= drm_device
->businfo
.pci
->bus
;
363 device
->pci_info
.device
= drm_device
->businfo
.pci
->dev
;
364 device
->pci_info
.function
= drm_device
->businfo
.pci
->func
;
366 device
->name
= gen_get_device_name(device
->chipset_id
);
368 if (device
->info
.is_haswell
) {
369 intel_logw("Haswell Vulkan support is incomplete");
370 } else if (device
->info
.gen
== 7 && !device
->info
.is_baytrail
) {
371 intel_logw("Ivy Bridge Vulkan support is incomplete");
372 } else if (device
->info
.gen
== 7 && device
->info
.is_baytrail
) {
373 intel_logw("Bay Trail Vulkan support is incomplete");
374 } else if (device
->info
.gen
>= 8 && device
->info
.gen
<= 11) {
375 /* Gen8-11 fully supported */
376 } else if (device
->info
.gen
== 12) {
377 intel_logw("Vulkan is not yet fully supported on gen12");
379 result
= vk_errorf(device
->instance
, device
,
380 VK_ERROR_INCOMPATIBLE_DRIVER
,
381 "Vulkan not yet supported on %s", device
->name
);
385 device
->cmd_parser_version
= -1;
386 if (device
->info
.gen
== 7) {
387 device
->cmd_parser_version
=
388 anv_gem_get_param(fd
, I915_PARAM_CMD_PARSER_VERSION
);
389 if (device
->cmd_parser_version
== -1) {
390 result
= vk_errorf(device
->instance
, device
,
391 VK_ERROR_INITIALIZATION_FAILED
,
392 "failed to get command parser version");
397 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_WAIT_TIMEOUT
)) {
398 result
= vk_errorf(device
->instance
, device
,
399 VK_ERROR_INITIALIZATION_FAILED
,
400 "kernel missing gem wait");
404 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_EXECBUF2
)) {
405 result
= vk_errorf(device
->instance
, device
,
406 VK_ERROR_INITIALIZATION_FAILED
,
407 "kernel missing execbuf2");
411 if (!device
->info
.has_llc
&&
412 anv_gem_get_param(fd
, I915_PARAM_MMAP_VERSION
) < 1) {
413 result
= vk_errorf(device
->instance
, device
,
414 VK_ERROR_INITIALIZATION_FAILED
,
415 "kernel missing wc mmap");
419 device
->has_softpin
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_SOFTPIN
);
420 device
->has_exec_async
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_ASYNC
);
421 device
->has_exec_capture
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_CAPTURE
);
422 device
->has_exec_fence
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE
);
423 device
->has_syncobj
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE_ARRAY
);
424 device
->has_syncobj_wait
= device
->has_syncobj
&&
425 anv_gem_supports_syncobj_wait(fd
);
426 device
->has_context_priority
= anv_gem_has_context_priority(fd
);
428 result
= anv_physical_device_init_heaps(device
, fd
);
429 if (result
!= VK_SUCCESS
)
432 device
->use_softpin
= device
->has_softpin
&&
433 device
->supports_48bit_addresses
;
435 device
->has_context_isolation
=
436 anv_gem_get_param(fd
, I915_PARAM_HAS_CONTEXT_ISOLATION
);
438 device
->always_use_bindless
=
439 env_var_as_boolean("ANV_ALWAYS_BINDLESS", false);
441 /* We first got the A64 messages on broadwell and we can only use them if
442 * we can pass addresses directly into the shader which requires softpin.
444 device
->has_a64_buffer_access
= device
->info
.gen
>= 8 &&
447 /* We first get bindless image access on Skylake and we can only really do
448 * it if we don't have any relocations so we need softpin.
450 device
->has_bindless_images
= device
->info
.gen
>= 9 &&
453 /* We've had bindless samplers since Ivy Bridge (forever in Vulkan terms)
454 * because it's just a matter of setting the sampler address in the sample
455 * message header. However, we've not bothered to wire it up for vec4 so
456 * we leave it disabled on gen7.
458 device
->has_bindless_samplers
= device
->info
.gen
>= 8;
460 device
->has_mem_available
= get_available_system_memory() != 0;
462 device
->always_flush_cache
=
463 driQueryOptionb(&instance
->dri_options
, "always_flush_cache");
465 /* Starting with Gen10, the timestamp frequency of the command streamer may
466 * vary from one part to another. We can query the value from the kernel.
468 if (device
->info
.gen
>= 10) {
469 int timestamp_frequency
=
470 anv_gem_get_param(fd
, I915_PARAM_CS_TIMESTAMP_FREQUENCY
);
472 if (timestamp_frequency
< 0)
473 intel_logw("Kernel 4.16-rc1+ required to properly query CS timestamp frequency");
475 device
->info
.timestamp_frequency
= timestamp_frequency
;
478 /* GENs prior to 8 do not support EU/Subslice info */
479 if (device
->info
.gen
>= 8) {
480 device
->subslice_total
= anv_gem_get_param(fd
, I915_PARAM_SUBSLICE_TOTAL
);
481 device
->eu_total
= anv_gem_get_param(fd
, I915_PARAM_EU_TOTAL
);
483 /* Without this information, we cannot get the right Braswell
484 * brandstrings, and we have to use conservative numbers for GPGPU on
485 * many platforms, but otherwise, things will just work.
487 if (device
->subslice_total
< 1 || device
->eu_total
< 1) {
488 intel_logw("Kernel 4.1 required to properly query GPU properties");
490 } else if (device
->info
.gen
== 7) {
491 device
->subslice_total
= 1 << (device
->info
.gt
- 1);
494 if (device
->info
.is_cherryview
&&
495 device
->subslice_total
> 0 && device
->eu_total
> 0) {
496 /* Logical CS threads = EUs per subslice * num threads per EU */
497 uint32_t max_cs_threads
=
498 device
->eu_total
/ device
->subslice_total
* device
->info
.num_thread_per_eu
;
500 /* Fuse configurations may give more threads than expected, never less. */
501 if (max_cs_threads
> device
->info
.max_cs_threads
)
502 device
->info
.max_cs_threads
= max_cs_threads
;
505 device
->compiler
= brw_compiler_create(NULL
, &device
->info
);
506 if (device
->compiler
== NULL
) {
507 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
510 device
->compiler
->shader_debug_log
= compiler_debug_log
;
511 device
->compiler
->shader_perf_log
= compiler_perf_log
;
512 device
->compiler
->supports_pull_constants
= false;
513 device
->compiler
->constant_buffer_0_is_relative
=
514 device
->info
.gen
< 8 || !device
->has_context_isolation
;
515 device
->compiler
->supports_shader_constants
= true;
516 device
->compiler
->compact_params
= false;
518 /* Broadwell PRM says:
520 * "Before Gen8, there was a historical configuration control field to
521 * swizzle address bit[6] for in X/Y tiling modes. This was set in three
522 * different places: TILECTL[1:0], ARB_MODE[5:4], and
523 * DISP_ARB_CTL[14:13].
525 * For Gen8 and subsequent generations, the swizzle fields are all
526 * reserved, and the CPU's memory controller performs all address
527 * swizzling modifications."
530 device
->info
.gen
< 8 && anv_gem_get_bit6_swizzle(fd
, I915_TILING_X
);
532 isl_device_init(&device
->isl_dev
, &device
->info
, swizzled
);
534 result
= anv_physical_device_init_uuids(device
);
535 if (result
!= VK_SUCCESS
)
538 anv_physical_device_init_disk_cache(device
);
540 if (instance
->enabled_extensions
.KHR_display
) {
541 master_fd
= open(primary_path
, O_RDWR
| O_CLOEXEC
);
542 if (master_fd
>= 0) {
543 /* prod the device with a GETPARAM call which will fail if
544 * we don't have permission to even render on this device
546 if (anv_gem_get_param(master_fd
, I915_PARAM_CHIPSET_ID
) == 0) {
552 device
->master_fd
= master_fd
;
554 result
= anv_init_wsi(device
);
555 if (result
!= VK_SUCCESS
) {
556 ralloc_free(device
->compiler
);
557 anv_physical_device_free_disk_cache(device
);
561 device
->perf
= anv_get_perf(&device
->info
, fd
);
563 anv_physical_device_get_supported_extensions(device
,
564 &device
->supported_extensions
);
567 device
->local_fd
= fd
;
579 anv_physical_device_finish(struct anv_physical_device
*device
)
581 anv_finish_wsi(device
);
582 anv_physical_device_free_disk_cache(device
);
583 ralloc_free(device
->compiler
);
584 ralloc_free(device
->perf
);
585 close(device
->local_fd
);
586 if (device
->master_fd
>= 0)
587 close(device
->master_fd
);
591 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
592 VkSystemAllocationScope allocationScope
)
598 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
599 size_t align
, VkSystemAllocationScope allocationScope
)
601 return realloc(pOriginal
, size
);
605 default_free_func(void *pUserData
, void *pMemory
)
610 static const VkAllocationCallbacks default_alloc
= {
612 .pfnAllocation
= default_alloc_func
,
613 .pfnReallocation
= default_realloc_func
,
614 .pfnFree
= default_free_func
,
617 VkResult
anv_EnumerateInstanceExtensionProperties(
618 const char* pLayerName
,
619 uint32_t* pPropertyCount
,
620 VkExtensionProperties
* pProperties
)
622 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
624 for (int i
= 0; i
< ANV_INSTANCE_EXTENSION_COUNT
; i
++) {
625 if (anv_instance_extensions_supported
.extensions
[i
]) {
626 vk_outarray_append(&out
, prop
) {
627 *prop
= anv_instance_extensions
[i
];
632 return vk_outarray_status(&out
);
635 VkResult
anv_CreateInstance(
636 const VkInstanceCreateInfo
* pCreateInfo
,
637 const VkAllocationCallbacks
* pAllocator
,
638 VkInstance
* pInstance
)
640 struct anv_instance
*instance
;
643 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
645 struct anv_instance_extension_table enabled_extensions
= {};
646 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
648 for (idx
= 0; idx
< ANV_INSTANCE_EXTENSION_COUNT
; idx
++) {
649 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
650 anv_instance_extensions
[idx
].extensionName
) == 0)
654 if (idx
>= ANV_INSTANCE_EXTENSION_COUNT
)
655 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
657 if (!anv_instance_extensions_supported
.extensions
[idx
])
658 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
660 enabled_extensions
.extensions
[idx
] = true;
663 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
664 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
666 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
668 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
671 instance
->alloc
= *pAllocator
;
673 instance
->alloc
= default_alloc
;
675 instance
->app_info
= (struct anv_app_info
) { .api_version
= 0 };
676 if (pCreateInfo
->pApplicationInfo
) {
677 const VkApplicationInfo
*app
= pCreateInfo
->pApplicationInfo
;
679 instance
->app_info
.app_name
=
680 vk_strdup(&instance
->alloc
, app
->pApplicationName
,
681 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
682 instance
->app_info
.app_version
= app
->applicationVersion
;
684 instance
->app_info
.engine_name
=
685 vk_strdup(&instance
->alloc
, app
->pEngineName
,
686 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
687 instance
->app_info
.engine_version
= app
->engineVersion
;
689 instance
->app_info
.api_version
= app
->apiVersion
;
692 if (instance
->app_info
.api_version
== 0)
693 instance
->app_info
.api_version
= VK_API_VERSION_1_0
;
695 instance
->enabled_extensions
= enabled_extensions
;
697 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->dispatch
.entrypoints
); i
++) {
698 /* Vulkan requires that entrypoints for extensions which have not been
699 * enabled must not be advertised.
701 if (!anv_instance_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
702 &instance
->enabled_extensions
)) {
703 instance
->dispatch
.entrypoints
[i
] = NULL
;
705 instance
->dispatch
.entrypoints
[i
] =
706 anv_instance_dispatch_table
.entrypoints
[i
];
710 struct anv_physical_device
*pdevice
= &instance
->physicalDevice
;
711 for (unsigned i
= 0; i
< ARRAY_SIZE(pdevice
->dispatch
.entrypoints
); i
++) {
712 /* Vulkan requires that entrypoints for extensions which have not been
713 * enabled must not be advertised.
715 if (!anv_physical_device_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
716 &instance
->enabled_extensions
)) {
717 pdevice
->dispatch
.entrypoints
[i
] = NULL
;
719 pdevice
->dispatch
.entrypoints
[i
] =
720 anv_physical_device_dispatch_table
.entrypoints
[i
];
724 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->device_dispatch
.entrypoints
); i
++) {
725 /* Vulkan requires that entrypoints for extensions which have not been
726 * enabled must not be advertised.
728 if (!anv_device_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
729 &instance
->enabled_extensions
, NULL
)) {
730 instance
->device_dispatch
.entrypoints
[i
] = NULL
;
732 instance
->device_dispatch
.entrypoints
[i
] =
733 anv_device_dispatch_table
.entrypoints
[i
];
737 instance
->physicalDeviceCount
= -1;
739 result
= vk_debug_report_instance_init(&instance
->debug_report_callbacks
);
740 if (result
!= VK_SUCCESS
) {
741 vk_free2(&default_alloc
, pAllocator
, instance
);
742 return vk_error(result
);
745 instance
->pipeline_cache_enabled
=
746 env_var_as_boolean("ANV_ENABLE_PIPELINE_CACHE", true);
748 glsl_type_singleton_init_or_ref();
750 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
752 driParseOptionInfo(&instance
->available_dri_options
, anv_dri_options_xml
);
753 driParseConfigFiles(&instance
->dri_options
, &instance
->available_dri_options
,
755 instance
->app_info
.engine_name
,
756 instance
->app_info
.engine_version
);
758 *pInstance
= anv_instance_to_handle(instance
);
763 void anv_DestroyInstance(
764 VkInstance _instance
,
765 const VkAllocationCallbacks
* pAllocator
)
767 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
772 if (instance
->physicalDeviceCount
> 0) {
773 /* We support at most one physical device. */
774 assert(instance
->physicalDeviceCount
== 1);
775 anv_physical_device_finish(&instance
->physicalDevice
);
778 vk_free(&instance
->alloc
, (char *)instance
->app_info
.app_name
);
779 vk_free(&instance
->alloc
, (char *)instance
->app_info
.engine_name
);
781 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
783 vk_debug_report_instance_destroy(&instance
->debug_report_callbacks
);
785 glsl_type_singleton_decref();
787 driDestroyOptionCache(&instance
->dri_options
);
788 driDestroyOptionInfo(&instance
->available_dri_options
);
790 vk_free(&instance
->alloc
, instance
);
794 anv_enumerate_devices(struct anv_instance
*instance
)
796 /* TODO: Check for more devices ? */
797 drmDevicePtr devices
[8];
798 VkResult result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
801 instance
->physicalDeviceCount
= 0;
803 max_devices
= drmGetDevices2(0, devices
, ARRAY_SIZE(devices
));
805 return VK_ERROR_INCOMPATIBLE_DRIVER
;
807 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
808 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
809 devices
[i
]->bustype
== DRM_BUS_PCI
&&
810 devices
[i
]->deviceinfo
.pci
->vendor_id
== 0x8086) {
812 result
= anv_physical_device_init(&instance
->physicalDevice
,
813 instance
, devices
[i
]);
814 if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
818 drmFreeDevices(devices
, max_devices
);
820 if (result
== VK_SUCCESS
)
821 instance
->physicalDeviceCount
= 1;
827 anv_instance_ensure_physical_device(struct anv_instance
*instance
)
829 if (instance
->physicalDeviceCount
< 0) {
830 VkResult result
= anv_enumerate_devices(instance
);
831 if (result
!= VK_SUCCESS
&&
832 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
839 VkResult
anv_EnumeratePhysicalDevices(
840 VkInstance _instance
,
841 uint32_t* pPhysicalDeviceCount
,
842 VkPhysicalDevice
* pPhysicalDevices
)
844 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
845 VK_OUTARRAY_MAKE(out
, pPhysicalDevices
, pPhysicalDeviceCount
);
847 VkResult result
= anv_instance_ensure_physical_device(instance
);
848 if (result
!= VK_SUCCESS
)
851 if (instance
->physicalDeviceCount
== 0)
854 assert(instance
->physicalDeviceCount
== 1);
855 vk_outarray_append(&out
, i
) {
856 *i
= anv_physical_device_to_handle(&instance
->physicalDevice
);
859 return vk_outarray_status(&out
);
862 VkResult
anv_EnumeratePhysicalDeviceGroups(
863 VkInstance _instance
,
864 uint32_t* pPhysicalDeviceGroupCount
,
865 VkPhysicalDeviceGroupProperties
* pPhysicalDeviceGroupProperties
)
867 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
868 VK_OUTARRAY_MAKE(out
, pPhysicalDeviceGroupProperties
,
869 pPhysicalDeviceGroupCount
);
871 VkResult result
= anv_instance_ensure_physical_device(instance
);
872 if (result
!= VK_SUCCESS
)
875 if (instance
->physicalDeviceCount
== 0)
878 assert(instance
->physicalDeviceCount
== 1);
880 vk_outarray_append(&out
, p
) {
881 p
->physicalDeviceCount
= 1;
882 memset(p
->physicalDevices
, 0, sizeof(p
->physicalDevices
));
883 p
->physicalDevices
[0] =
884 anv_physical_device_to_handle(&instance
->physicalDevice
);
885 p
->subsetAllocation
= false;
887 vk_foreach_struct(ext
, p
->pNext
)
888 anv_debug_ignored_stype(ext
->sType
);
891 return vk_outarray_status(&out
);
894 void anv_GetPhysicalDeviceFeatures(
895 VkPhysicalDevice physicalDevice
,
896 VkPhysicalDeviceFeatures
* pFeatures
)
898 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
900 *pFeatures
= (VkPhysicalDeviceFeatures
) {
901 .robustBufferAccess
= true,
902 .fullDrawIndexUint32
= true,
903 .imageCubeArray
= true,
904 .independentBlend
= true,
905 .geometryShader
= true,
906 .tessellationShader
= true,
907 .sampleRateShading
= true,
908 .dualSrcBlend
= true,
910 .multiDrawIndirect
= true,
911 .drawIndirectFirstInstance
= true,
913 .depthBiasClamp
= true,
914 .fillModeNonSolid
= true,
915 .depthBounds
= pdevice
->info
.gen
>= 12,
919 .multiViewport
= true,
920 .samplerAnisotropy
= true,
921 .textureCompressionETC2
= pdevice
->info
.gen
>= 8 ||
922 pdevice
->info
.is_baytrail
,
923 .textureCompressionASTC_LDR
= pdevice
->info
.gen
>= 9, /* FINISHME CHV */
924 .textureCompressionBC
= true,
925 .occlusionQueryPrecise
= true,
926 .pipelineStatisticsQuery
= true,
927 .fragmentStoresAndAtomics
= true,
928 .shaderTessellationAndGeometryPointSize
= true,
929 .shaderImageGatherExtended
= true,
930 .shaderStorageImageExtendedFormats
= true,
931 .shaderStorageImageMultisample
= false,
932 .shaderStorageImageReadWithoutFormat
= false,
933 .shaderStorageImageWriteWithoutFormat
= true,
934 .shaderUniformBufferArrayDynamicIndexing
= true,
935 .shaderSampledImageArrayDynamicIndexing
= true,
936 .shaderStorageBufferArrayDynamicIndexing
= true,
937 .shaderStorageImageArrayDynamicIndexing
= true,
938 .shaderClipDistance
= true,
939 .shaderCullDistance
= true,
940 .shaderFloat64
= pdevice
->info
.gen
>= 8 &&
941 pdevice
->info
.has_64bit_types
,
942 .shaderInt64
= pdevice
->info
.gen
>= 8 &&
943 pdevice
->info
.has_64bit_types
,
944 .shaderInt16
= pdevice
->info
.gen
>= 8,
945 .shaderResourceMinLod
= pdevice
->info
.gen
>= 9,
946 .variableMultisampleRate
= true,
947 .inheritedQueries
= true,
950 /* We can't do image stores in vec4 shaders */
951 pFeatures
->vertexPipelineStoresAndAtomics
=
952 pdevice
->compiler
->scalar_stage
[MESA_SHADER_VERTEX
] &&
953 pdevice
->compiler
->scalar_stage
[MESA_SHADER_GEOMETRY
];
955 struct anv_app_info
*app_info
= &pdevice
->instance
->app_info
;
957 /* The new DOOM and Wolfenstein games require depthBounds without
958 * checking for it. They seem to run fine without it so just claim it's
959 * there and accept the consequences.
961 if (app_info
->engine_name
&& strcmp(app_info
->engine_name
, "idTech") == 0)
962 pFeatures
->depthBounds
= true;
965 void anv_GetPhysicalDeviceFeatures2(
966 VkPhysicalDevice physicalDevice
,
967 VkPhysicalDeviceFeatures2
* pFeatures
)
969 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
970 anv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
972 vk_foreach_struct(ext
, pFeatures
->pNext
) {
973 switch (ext
->sType
) {
974 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES_KHR
: {
975 VkPhysicalDevice8BitStorageFeaturesKHR
*features
=
976 (VkPhysicalDevice8BitStorageFeaturesKHR
*)ext
;
977 features
->storageBuffer8BitAccess
= pdevice
->info
.gen
>= 8;
978 features
->uniformAndStorageBuffer8BitAccess
= pdevice
->info
.gen
>= 8;
979 features
->storagePushConstant8
= pdevice
->info
.gen
>= 8;
983 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES
: {
984 VkPhysicalDevice16BitStorageFeatures
*features
=
985 (VkPhysicalDevice16BitStorageFeatures
*)ext
;
986 features
->storageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
987 features
->uniformAndStorageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
988 features
->storagePushConstant16
= pdevice
->info
.gen
>= 8;
989 features
->storageInputOutput16
= false;
993 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_EXT
: {
994 VkPhysicalDeviceBufferDeviceAddressFeaturesEXT
*features
= (void *)ext
;
995 features
->bufferDeviceAddress
= pdevice
->has_a64_buffer_access
;
996 features
->bufferDeviceAddressCaptureReplay
= false;
997 features
->bufferDeviceAddressMultiDevice
= false;
1001 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_KHR
: {
1002 VkPhysicalDeviceBufferDeviceAddressFeaturesKHR
*features
= (void *)ext
;
1003 features
->bufferDeviceAddress
= pdevice
->has_a64_buffer_access
;
1004 features
->bufferDeviceAddressCaptureReplay
=
1005 pdevice
->has_a64_buffer_access
;
1006 features
->bufferDeviceAddressMultiDevice
= false;
1010 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COMPUTE_SHADER_DERIVATIVES_FEATURES_NV
: {
1011 VkPhysicalDeviceComputeShaderDerivativesFeaturesNV
*features
=
1012 (VkPhysicalDeviceComputeShaderDerivativesFeaturesNV
*)ext
;
1013 features
->computeDerivativeGroupQuads
= true;
1014 features
->computeDerivativeGroupLinear
= true;
1018 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT
: {
1019 VkPhysicalDeviceConditionalRenderingFeaturesEXT
*features
=
1020 (VkPhysicalDeviceConditionalRenderingFeaturesEXT
*)ext
;
1021 features
->conditionalRendering
= pdevice
->info
.gen
>= 8 ||
1022 pdevice
->info
.is_haswell
;
1023 features
->inheritedConditionalRendering
= pdevice
->info
.gen
>= 8 ||
1024 pdevice
->info
.is_haswell
;
1028 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_CLIP_ENABLE_FEATURES_EXT
: {
1029 VkPhysicalDeviceDepthClipEnableFeaturesEXT
*features
=
1030 (VkPhysicalDeviceDepthClipEnableFeaturesEXT
*)ext
;
1031 features
->depthClipEnable
= true;
1035 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR
: {
1036 VkPhysicalDeviceFloat16Int8FeaturesKHR
*features
= (void *)ext
;
1037 features
->shaderFloat16
= pdevice
->info
.gen
>= 8;
1038 features
->shaderInt8
= pdevice
->info
.gen
>= 8;
1042 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FRAGMENT_SHADER_INTERLOCK_FEATURES_EXT
: {
1043 VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT
*features
=
1044 (VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT
*)ext
;
1045 features
->fragmentShaderSampleInterlock
= pdevice
->info
.gen
>= 9;
1046 features
->fragmentShaderPixelInterlock
= pdevice
->info
.gen
>= 9;
1047 features
->fragmentShaderShadingRateInterlock
= false;
1051 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_QUERY_RESET_FEATURES_EXT
: {
1052 VkPhysicalDeviceHostQueryResetFeaturesEXT
*features
=
1053 (VkPhysicalDeviceHostQueryResetFeaturesEXT
*)ext
;
1054 features
->hostQueryReset
= true;
1058 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT
: {
1059 VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*features
=
1060 (VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*)ext
;
1061 bool descIndexing
= pdevice
->has_a64_buffer_access
&&
1062 pdevice
->has_bindless_images
;
1063 features
->shaderInputAttachmentArrayDynamicIndexing
= false;
1064 features
->shaderUniformTexelBufferArrayDynamicIndexing
= descIndexing
;
1065 features
->shaderStorageTexelBufferArrayDynamicIndexing
= descIndexing
;
1066 features
->shaderUniformBufferArrayNonUniformIndexing
= false;
1067 features
->shaderSampledImageArrayNonUniformIndexing
= descIndexing
;
1068 features
->shaderStorageBufferArrayNonUniformIndexing
= descIndexing
;
1069 features
->shaderStorageImageArrayNonUniformIndexing
= descIndexing
;
1070 features
->shaderInputAttachmentArrayNonUniformIndexing
= false;
1071 features
->shaderUniformTexelBufferArrayNonUniformIndexing
= descIndexing
;
1072 features
->shaderStorageTexelBufferArrayNonUniformIndexing
= descIndexing
;
1073 features
->descriptorBindingUniformBufferUpdateAfterBind
= false;
1074 features
->descriptorBindingSampledImageUpdateAfterBind
= descIndexing
;
1075 features
->descriptorBindingStorageImageUpdateAfterBind
= descIndexing
;
1076 features
->descriptorBindingStorageBufferUpdateAfterBind
= descIndexing
;
1077 features
->descriptorBindingUniformTexelBufferUpdateAfterBind
= descIndexing
;
1078 features
->descriptorBindingStorageTexelBufferUpdateAfterBind
= descIndexing
;
1079 features
->descriptorBindingUpdateUnusedWhilePending
= descIndexing
;
1080 features
->descriptorBindingPartiallyBound
= descIndexing
;
1081 features
->descriptorBindingVariableDescriptorCount
= false;
1082 features
->runtimeDescriptorArray
= descIndexing
;
1086 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INDEX_TYPE_UINT8_FEATURES_EXT
: {
1087 VkPhysicalDeviceIndexTypeUint8FeaturesEXT
*features
=
1088 (VkPhysicalDeviceIndexTypeUint8FeaturesEXT
*)ext
;
1089 features
->indexTypeUint8
= true;
1093 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_FEATURES_EXT
: {
1094 VkPhysicalDeviceInlineUniformBlockFeaturesEXT
*features
=
1095 (VkPhysicalDeviceInlineUniformBlockFeaturesEXT
*)ext
;
1096 features
->inlineUniformBlock
= true;
1097 features
->descriptorBindingInlineUniformBlockUpdateAfterBind
= true;
1101 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_FEATURES_EXT
: {
1102 VkPhysicalDeviceLineRasterizationFeaturesEXT
*features
=
1103 (VkPhysicalDeviceLineRasterizationFeaturesEXT
*)ext
;
1104 features
->rectangularLines
= true;
1105 features
->bresenhamLines
= true;
1106 /* Support for Smooth lines with MSAA was removed on gen11. From the
1107 * BSpec section "Multisample ModesState" table for "AA Line Support
1110 * GEN10:BUG:######## NUM_MULTISAMPLES == 1
1112 * Fortunately, this isn't a case most people care about.
1114 features
->smoothLines
= pdevice
->info
.gen
< 10;
1115 features
->stippledRectangularLines
= false;
1116 features
->stippledBresenhamLines
= true;
1117 features
->stippledSmoothLines
= false;
1121 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES
: {
1122 VkPhysicalDeviceMultiviewFeatures
*features
=
1123 (VkPhysicalDeviceMultiviewFeatures
*)ext
;
1124 features
->multiview
= true;
1125 features
->multiviewGeometryShader
= true;
1126 features
->multiviewTessellationShader
= true;
1130 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGELESS_FRAMEBUFFER_FEATURES_KHR
: {
1131 VkPhysicalDeviceImagelessFramebufferFeaturesKHR
*features
=
1132 (VkPhysicalDeviceImagelessFramebufferFeaturesKHR
*)ext
;
1133 features
->imagelessFramebuffer
= true;
1137 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR
: {
1138 VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR
*features
=
1139 (VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR
*)ext
;
1140 features
->pipelineExecutableInfo
= true;
1144 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES
: {
1145 VkPhysicalDeviceProtectedMemoryFeatures
*features
= (void *)ext
;
1146 features
->protectedMemory
= false;
1150 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES
: {
1151 VkPhysicalDeviceSamplerYcbcrConversionFeatures
*features
=
1152 (VkPhysicalDeviceSamplerYcbcrConversionFeatures
*) ext
;
1153 features
->samplerYcbcrConversion
= true;
1157 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES_EXT
: {
1158 VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*features
=
1159 (VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*)ext
;
1160 features
->scalarBlockLayout
= true;
1164 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SEPARATE_DEPTH_STENCIL_LAYOUTS_FEATURES_KHR
: {
1165 VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR
*features
=
1166 (VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR
*)ext
;
1167 features
->separateDepthStencilLayouts
= true;
1171 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_INT64_FEATURES_KHR
: {
1172 VkPhysicalDeviceShaderAtomicInt64FeaturesKHR
*features
= (void *)ext
;
1173 features
->shaderBufferInt64Atomics
=
1174 pdevice
->info
.gen
>= 9 && pdevice
->use_softpin
;
1175 features
->shaderSharedInt64Atomics
= VK_FALSE
;
1179 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DEMOTE_TO_HELPER_INVOCATION_FEATURES_EXT
: {
1180 VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT
*features
= (void *)ext
;
1181 features
->shaderDemoteToHelperInvocation
= true;
1185 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CLOCK_FEATURES_KHR
: {
1186 VkPhysicalDeviceShaderClockFeaturesKHR
*features
=
1187 (VkPhysicalDeviceShaderClockFeaturesKHR
*)ext
;
1188 features
->shaderSubgroupClock
= true;
1189 features
->shaderDeviceClock
= false;
1193 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETERS_FEATURES
: {
1194 VkPhysicalDeviceShaderDrawParametersFeatures
*features
= (void *)ext
;
1195 features
->shaderDrawParameters
= true;
1199 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_SUBGROUP_EXTENDED_TYPES_FEATURES_KHR
: {
1200 VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR
*features
=
1201 (VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR
*)ext
;
1202 features
->shaderSubgroupExtendedTypes
= true;
1206 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_FEATURES_EXT
: {
1207 VkPhysicalDeviceSubgroupSizeControlFeaturesEXT
*features
=
1208 (VkPhysicalDeviceSubgroupSizeControlFeaturesEXT
*)ext
;
1209 features
->subgroupSizeControl
= true;
1210 features
->computeFullSubgroups
= true;
1214 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_FEATURES_EXT
: {
1215 VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT
*features
=
1216 (VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT
*)ext
;
1217 features
->texelBufferAlignment
= true;
1221 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_FEATURES_KHR
: {
1222 VkPhysicalDeviceTimelineSemaphoreFeaturesKHR
*features
=
1223 (VkPhysicalDeviceTimelineSemaphoreFeaturesKHR
*) ext
;
1224 features
->timelineSemaphore
= true;
1228 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES
: {
1229 VkPhysicalDeviceVariablePointersFeatures
*features
= (void *)ext
;
1230 features
->variablePointersStorageBuffer
= true;
1231 features
->variablePointers
= true;
1235 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT
: {
1236 VkPhysicalDeviceTransformFeedbackFeaturesEXT
*features
=
1237 (VkPhysicalDeviceTransformFeedbackFeaturesEXT
*)ext
;
1238 features
->transformFeedback
= true;
1239 features
->geometryStreams
= true;
1243 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_UNIFORM_BUFFER_STANDARD_LAYOUT_FEATURES_KHR
: {
1244 VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR
*features
=
1245 (VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR
*)ext
;
1246 features
->uniformBufferStandardLayout
= true;
1250 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT
: {
1251 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*features
=
1252 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*)ext
;
1253 features
->vertexAttributeInstanceRateDivisor
= true;
1254 features
->vertexAttributeInstanceRateZeroDivisor
= true;
1258 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_MEMORY_MODEL_FEATURES_KHR
: {
1259 VkPhysicalDeviceVulkanMemoryModelFeaturesKHR
*features
= (void *)ext
;
1260 features
->vulkanMemoryModel
= true;
1261 features
->vulkanMemoryModelDeviceScope
= true;
1262 features
->vulkanMemoryModelAvailabilityVisibilityChains
= true;
1266 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_YCBCR_IMAGE_ARRAYS_FEATURES_EXT
: {
1267 VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*features
=
1268 (VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*)ext
;
1269 features
->ycbcrImageArrays
= true;
1274 anv_debug_ignored_stype(ext
->sType
);
1280 #define MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS 64
1282 #define MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS 64
1283 #define MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS 256
1285 void anv_GetPhysicalDeviceProperties(
1286 VkPhysicalDevice physicalDevice
,
1287 VkPhysicalDeviceProperties
* pProperties
)
1289 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1290 const struct gen_device_info
*devinfo
= &pdevice
->info
;
1292 /* See assertions made when programming the buffer surface state. */
1293 const uint32_t max_raw_buffer_sz
= devinfo
->gen
>= 7 ?
1294 (1ul << 30) : (1ul << 27);
1296 const uint32_t max_ssbos
= pdevice
->has_a64_buffer_access
? UINT16_MAX
: 64;
1297 const uint32_t max_textures
=
1298 pdevice
->has_bindless_images
? UINT16_MAX
: 128;
1299 const uint32_t max_samplers
=
1300 pdevice
->has_bindless_samplers
? UINT16_MAX
:
1301 (devinfo
->gen
>= 8 || devinfo
->is_haswell
) ? 128 : 16;
1302 const uint32_t max_images
=
1303 pdevice
->has_bindless_images
? UINT16_MAX
: MAX_IMAGES
;
1305 /* If we can use bindless for everything, claim a high per-stage limit,
1306 * otherwise use the binding table size, minus the slots reserved for
1307 * render targets and one slot for the descriptor buffer. */
1308 const uint32_t max_per_stage
=
1309 pdevice
->has_bindless_images
&& pdevice
->has_a64_buffer_access
1310 ? UINT32_MAX
: MAX_BINDING_TABLE_SIZE
- MAX_RTS
- 1;
1312 const uint32_t max_workgroup_size
= 32 * devinfo
->max_cs_threads
;
1314 VkSampleCountFlags sample_counts
=
1315 isl_device_get_sample_counts(&pdevice
->isl_dev
);
1318 VkPhysicalDeviceLimits limits
= {
1319 .maxImageDimension1D
= (1 << 14),
1320 .maxImageDimension2D
= (1 << 14),
1321 .maxImageDimension3D
= (1 << 11),
1322 .maxImageDimensionCube
= (1 << 14),
1323 .maxImageArrayLayers
= (1 << 11),
1324 .maxTexelBufferElements
= 128 * 1024 * 1024,
1325 .maxUniformBufferRange
= (1ul << 27),
1326 .maxStorageBufferRange
= max_raw_buffer_sz
,
1327 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
1328 .maxMemoryAllocationCount
= UINT32_MAX
,
1329 .maxSamplerAllocationCount
= 64 * 1024,
1330 .bufferImageGranularity
= 64, /* A cache line */
1331 .sparseAddressSpaceSize
= 0,
1332 .maxBoundDescriptorSets
= MAX_SETS
,
1333 .maxPerStageDescriptorSamplers
= max_samplers
,
1334 .maxPerStageDescriptorUniformBuffers
= MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
,
1335 .maxPerStageDescriptorStorageBuffers
= max_ssbos
,
1336 .maxPerStageDescriptorSampledImages
= max_textures
,
1337 .maxPerStageDescriptorStorageImages
= max_images
,
1338 .maxPerStageDescriptorInputAttachments
= MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS
,
1339 .maxPerStageResources
= max_per_stage
,
1340 .maxDescriptorSetSamplers
= 6 * max_samplers
, /* number of stages * maxPerStageDescriptorSamplers */
1341 .maxDescriptorSetUniformBuffers
= 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
, /* number of stages * maxPerStageDescriptorUniformBuffers */
1342 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
1343 .maxDescriptorSetStorageBuffers
= 6 * max_ssbos
, /* number of stages * maxPerStageDescriptorStorageBuffers */
1344 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
1345 .maxDescriptorSetSampledImages
= 6 * max_textures
, /* number of stages * maxPerStageDescriptorSampledImages */
1346 .maxDescriptorSetStorageImages
= 6 * max_images
, /* number of stages * maxPerStageDescriptorStorageImages */
1347 .maxDescriptorSetInputAttachments
= MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS
,
1348 .maxVertexInputAttributes
= MAX_VBS
,
1349 .maxVertexInputBindings
= MAX_VBS
,
1350 .maxVertexInputAttributeOffset
= 2047,
1351 .maxVertexInputBindingStride
= 2048,
1352 .maxVertexOutputComponents
= 128,
1353 .maxTessellationGenerationLevel
= 64,
1354 .maxTessellationPatchSize
= 32,
1355 .maxTessellationControlPerVertexInputComponents
= 128,
1356 .maxTessellationControlPerVertexOutputComponents
= 128,
1357 .maxTessellationControlPerPatchOutputComponents
= 128,
1358 .maxTessellationControlTotalOutputComponents
= 2048,
1359 .maxTessellationEvaluationInputComponents
= 128,
1360 .maxTessellationEvaluationOutputComponents
= 128,
1361 .maxGeometryShaderInvocations
= 32,
1362 .maxGeometryInputComponents
= 64,
1363 .maxGeometryOutputComponents
= 128,
1364 .maxGeometryOutputVertices
= 256,
1365 .maxGeometryTotalOutputComponents
= 1024,
1366 .maxFragmentInputComponents
= 116, /* 128 components - (PSIZ, CLIP_DIST0, CLIP_DIST1) */
1367 .maxFragmentOutputAttachments
= 8,
1368 .maxFragmentDualSrcAttachments
= 1,
1369 .maxFragmentCombinedOutputResources
= 8,
1370 .maxComputeSharedMemorySize
= 64 * 1024,
1371 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
1372 .maxComputeWorkGroupInvocations
= max_workgroup_size
,
1373 .maxComputeWorkGroupSize
= {
1378 .subPixelPrecisionBits
= 8,
1379 .subTexelPrecisionBits
= 8,
1380 .mipmapPrecisionBits
= 8,
1381 .maxDrawIndexedIndexValue
= UINT32_MAX
,
1382 .maxDrawIndirectCount
= UINT32_MAX
,
1383 .maxSamplerLodBias
= 16,
1384 .maxSamplerAnisotropy
= 16,
1385 .maxViewports
= MAX_VIEWPORTS
,
1386 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
1387 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
1388 .viewportSubPixelBits
= 13, /* We take a float? */
1389 .minMemoryMapAlignment
= 4096, /* A page */
1390 /* The dataport requires texel alignment so we need to assume a worst
1391 * case of R32G32B32A32 which is 16 bytes.
1393 .minTexelBufferOffsetAlignment
= 16,
1394 /* We need 16 for UBO block reads to work and 32 for push UBOs */
1395 .minUniformBufferOffsetAlignment
= 32,
1396 .minStorageBufferOffsetAlignment
= 4,
1397 .minTexelOffset
= -8,
1398 .maxTexelOffset
= 7,
1399 .minTexelGatherOffset
= -32,
1400 .maxTexelGatherOffset
= 31,
1401 .minInterpolationOffset
= -0.5,
1402 .maxInterpolationOffset
= 0.4375,
1403 .subPixelInterpolationOffsetBits
= 4,
1404 .maxFramebufferWidth
= (1 << 14),
1405 .maxFramebufferHeight
= (1 << 14),
1406 .maxFramebufferLayers
= (1 << 11),
1407 .framebufferColorSampleCounts
= sample_counts
,
1408 .framebufferDepthSampleCounts
= sample_counts
,
1409 .framebufferStencilSampleCounts
= sample_counts
,
1410 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
1411 .maxColorAttachments
= MAX_RTS
,
1412 .sampledImageColorSampleCounts
= sample_counts
,
1413 .sampledImageIntegerSampleCounts
= sample_counts
,
1414 .sampledImageDepthSampleCounts
= sample_counts
,
1415 .sampledImageStencilSampleCounts
= sample_counts
,
1416 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
1417 .maxSampleMaskWords
= 1,
1418 .timestampComputeAndGraphics
= true,
1419 .timestampPeriod
= 1000000000.0 / devinfo
->timestamp_frequency
,
1420 .maxClipDistances
= 8,
1421 .maxCullDistances
= 8,
1422 .maxCombinedClipAndCullDistances
= 8,
1423 .discreteQueuePriorities
= 2,
1424 .pointSizeRange
= { 0.125, 255.875 },
1427 (devinfo
->gen
>= 9 || devinfo
->is_cherryview
) ?
1428 2047.9921875 : 7.9921875,
1430 .pointSizeGranularity
= (1.0 / 8.0),
1431 .lineWidthGranularity
= (1.0 / 128.0),
1432 .strictLines
= false,
1433 .standardSampleLocations
= true,
1434 .optimalBufferCopyOffsetAlignment
= 128,
1435 .optimalBufferCopyRowPitchAlignment
= 128,
1436 .nonCoherentAtomSize
= 64,
1439 *pProperties
= (VkPhysicalDeviceProperties
) {
1440 .apiVersion
= anv_physical_device_api_version(pdevice
),
1441 .driverVersion
= vk_get_driver_version(),
1443 .deviceID
= pdevice
->chipset_id
,
1444 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
1446 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
1449 snprintf(pProperties
->deviceName
, sizeof(pProperties
->deviceName
),
1450 "%s", pdevice
->name
);
1451 memcpy(pProperties
->pipelineCacheUUID
,
1452 pdevice
->pipeline_cache_uuid
, VK_UUID_SIZE
);
1455 void anv_GetPhysicalDeviceProperties2(
1456 VkPhysicalDevice physicalDevice
,
1457 VkPhysicalDeviceProperties2
* pProperties
)
1459 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1461 anv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
1463 vk_foreach_struct(ext
, pProperties
->pNext
) {
1464 switch (ext
->sType
) {
1465 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_STENCIL_RESOLVE_PROPERTIES_KHR
: {
1466 VkPhysicalDeviceDepthStencilResolvePropertiesKHR
*props
=
1467 (VkPhysicalDeviceDepthStencilResolvePropertiesKHR
*)ext
;
1469 /* We support all of the depth resolve modes */
1470 props
->supportedDepthResolveModes
=
1471 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
|
1472 VK_RESOLVE_MODE_AVERAGE_BIT_KHR
|
1473 VK_RESOLVE_MODE_MIN_BIT_KHR
|
1474 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1476 /* Average doesn't make sense for stencil so we don't support that */
1477 props
->supportedStencilResolveModes
=
1478 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
;
1479 if (pdevice
->info
.gen
>= 8) {
1480 /* The advanced stencil resolve modes currently require stencil
1481 * sampling be supported by the hardware.
1483 props
->supportedStencilResolveModes
|=
1484 VK_RESOLVE_MODE_MIN_BIT_KHR
|
1485 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1488 props
->independentResolveNone
= true;
1489 props
->independentResolve
= true;
1493 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES_EXT
: {
1494 VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*props
=
1495 (VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*)ext
;
1497 /* It's a bit hard to exactly map our implementation to the limits
1498 * described here. The bindless surface handle in the extended
1499 * message descriptors is 20 bits and it's an index into the table of
1500 * RENDER_SURFACE_STATE structs that starts at bindless surface base
1501 * address. Given that most things consume two surface states per
1502 * view (general/sampled for textures and write-only/read-write for
1503 * images), we claim 2^19 things.
1505 * For SSBOs, we just use A64 messages so there is no real limit
1506 * there beyond the limit on the total size of a descriptor set.
1508 const unsigned max_bindless_views
= 1 << 19;
1510 props
->maxUpdateAfterBindDescriptorsInAllPools
= max_bindless_views
;
1511 props
->shaderUniformBufferArrayNonUniformIndexingNative
= false;
1512 props
->shaderSampledImageArrayNonUniformIndexingNative
= false;
1513 props
->shaderStorageBufferArrayNonUniformIndexingNative
= true;
1514 props
->shaderStorageImageArrayNonUniformIndexingNative
= false;
1515 props
->shaderInputAttachmentArrayNonUniformIndexingNative
= false;
1516 props
->robustBufferAccessUpdateAfterBind
= true;
1517 props
->quadDivergentImplicitLod
= false;
1518 props
->maxPerStageDescriptorUpdateAfterBindSamplers
= max_bindless_views
;
1519 props
->maxPerStageDescriptorUpdateAfterBindUniformBuffers
= MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
;
1520 props
->maxPerStageDescriptorUpdateAfterBindStorageBuffers
= UINT32_MAX
;
1521 props
->maxPerStageDescriptorUpdateAfterBindSampledImages
= max_bindless_views
;
1522 props
->maxPerStageDescriptorUpdateAfterBindStorageImages
= max_bindless_views
;
1523 props
->maxPerStageDescriptorUpdateAfterBindInputAttachments
= MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS
;
1524 props
->maxPerStageUpdateAfterBindResources
= UINT32_MAX
;
1525 props
->maxDescriptorSetUpdateAfterBindSamplers
= max_bindless_views
;
1526 props
->maxDescriptorSetUpdateAfterBindUniformBuffers
= 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
;
1527 props
->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2;
1528 props
->maxDescriptorSetUpdateAfterBindStorageBuffers
= UINT32_MAX
;
1529 props
->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2;
1530 props
->maxDescriptorSetUpdateAfterBindSampledImages
= max_bindless_views
;
1531 props
->maxDescriptorSetUpdateAfterBindStorageImages
= max_bindless_views
;
1532 props
->maxDescriptorSetUpdateAfterBindInputAttachments
= MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS
;
1536 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR
: {
1537 VkPhysicalDeviceDriverPropertiesKHR
*driver_props
=
1538 (VkPhysicalDeviceDriverPropertiesKHR
*) ext
;
1540 driver_props
->driverID
= VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA_KHR
;
1541 snprintf(driver_props
->driverName
, VK_MAX_DRIVER_NAME_SIZE_KHR
,
1542 "Intel open-source Mesa driver");
1544 snprintf(driver_props
->driverInfo
, VK_MAX_DRIVER_INFO_SIZE_KHR
,
1545 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1
);
1547 driver_props
->conformanceVersion
= (VkConformanceVersionKHR
) {
1556 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT
: {
1557 VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*props
=
1558 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*) ext
;
1559 /* Userptr needs page aligned memory. */
1560 props
->minImportedHostPointerAlignment
= 4096;
1564 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES
: {
1565 VkPhysicalDeviceIDProperties
*id_props
=
1566 (VkPhysicalDeviceIDProperties
*)ext
;
1567 memcpy(id_props
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
1568 memcpy(id_props
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
1569 /* The LUID is for Windows. */
1570 id_props
->deviceLUIDValid
= false;
1574 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_PROPERTIES_EXT
: {
1575 VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*props
=
1576 (VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*)ext
;
1577 props
->maxInlineUniformBlockSize
= MAX_INLINE_UNIFORM_BLOCK_SIZE
;
1578 props
->maxPerStageDescriptorInlineUniformBlocks
=
1579 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1580 props
->maxPerStageDescriptorUpdateAfterBindInlineUniformBlocks
=
1581 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1582 props
->maxDescriptorSetInlineUniformBlocks
=
1583 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1584 props
->maxDescriptorSetUpdateAfterBindInlineUniformBlocks
=
1585 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1589 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_PROPERTIES_EXT
: {
1590 VkPhysicalDeviceLineRasterizationPropertiesEXT
*props
=
1591 (VkPhysicalDeviceLineRasterizationPropertiesEXT
*)ext
;
1592 /* In the Skylake PRM Vol. 7, subsection titled "GIQ (Diamond)
1593 * Sampling Rules - Legacy Mode", it says the following:
1595 * "Note that the device divides a pixel into a 16x16 array of
1596 * subpixels, referenced by their upper left corners."
1598 * This is the only known reference in the PRMs to the subpixel
1599 * precision of line rasterization and a "16x16 array of subpixels"
1600 * implies 4 subpixel precision bits. Empirical testing has shown
1601 * that 4 subpixel precision bits applies to all line rasterization
1604 props
->lineSubPixelPrecisionBits
= 4;
1608 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES
: {
1609 VkPhysicalDeviceMaintenance3Properties
*props
=
1610 (VkPhysicalDeviceMaintenance3Properties
*)ext
;
1611 /* This value doesn't matter for us today as our per-stage
1612 * descriptors are the real limit.
1614 props
->maxPerSetDescriptors
= 1024;
1615 props
->maxMemoryAllocationSize
= MAX_MEMORY_ALLOCATION_SIZE
;
1619 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES
: {
1620 VkPhysicalDeviceMultiviewProperties
*properties
=
1621 (VkPhysicalDeviceMultiviewProperties
*)ext
;
1622 properties
->maxMultiviewViewCount
= 16;
1623 properties
->maxMultiviewInstanceIndex
= UINT32_MAX
/ 16;
1627 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT
: {
1628 VkPhysicalDevicePCIBusInfoPropertiesEXT
*properties
=
1629 (VkPhysicalDevicePCIBusInfoPropertiesEXT
*)ext
;
1630 properties
->pciDomain
= pdevice
->pci_info
.domain
;
1631 properties
->pciBus
= pdevice
->pci_info
.bus
;
1632 properties
->pciDevice
= pdevice
->pci_info
.device
;
1633 properties
->pciFunction
= pdevice
->pci_info
.function
;
1637 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES
: {
1638 VkPhysicalDevicePointClippingProperties
*properties
=
1639 (VkPhysicalDevicePointClippingProperties
*) ext
;
1640 properties
->pointClippingBehavior
= VK_POINT_CLIPPING_BEHAVIOR_USER_CLIP_PLANES_ONLY
;
1644 #pragma GCC diagnostic push
1645 #pragma GCC diagnostic ignored "-Wswitch"
1646 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PRESENTATION_PROPERTIES_ANDROID
: {
1647 VkPhysicalDevicePresentationPropertiesANDROID
*props
=
1648 (VkPhysicalDevicePresentationPropertiesANDROID
*)ext
;
1649 props
->sharedImage
= VK_FALSE
;
1652 #pragma GCC diagnostic pop
1654 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES
: {
1655 VkPhysicalDeviceProtectedMemoryProperties
*props
=
1656 (VkPhysicalDeviceProtectedMemoryProperties
*)ext
;
1657 props
->protectedNoFault
= false;
1661 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
1662 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
1663 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
1665 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
1669 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT
: {
1670 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*properties
=
1671 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*)ext
;
1672 properties
->filterMinmaxImageComponentMapping
= pdevice
->info
.gen
>= 9;
1673 properties
->filterMinmaxSingleComponentFormats
= pdevice
->info
.gen
>= 9;
1677 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES
: {
1678 VkPhysicalDeviceSubgroupProperties
*properties
= (void *)ext
;
1680 properties
->subgroupSize
= BRW_SUBGROUP_SIZE
;
1682 VkShaderStageFlags scalar_stages
= 0;
1683 for (unsigned stage
= 0; stage
< MESA_SHADER_STAGES
; stage
++) {
1684 if (pdevice
->compiler
->scalar_stage
[stage
])
1685 scalar_stages
|= mesa_to_vk_shader_stage(stage
);
1687 properties
->supportedStages
= scalar_stages
;
1689 properties
->supportedOperations
= VK_SUBGROUP_FEATURE_BASIC_BIT
|
1690 VK_SUBGROUP_FEATURE_VOTE_BIT
|
1691 VK_SUBGROUP_FEATURE_BALLOT_BIT
|
1692 VK_SUBGROUP_FEATURE_SHUFFLE_BIT
|
1693 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT
|
1694 VK_SUBGROUP_FEATURE_QUAD_BIT
;
1695 if (pdevice
->info
.gen
>= 8) {
1696 /* TODO: There's no technical reason why these can't be made to
1697 * work on gen7 but they don't at the moment so it's best to leave
1698 * the feature disabled than enabled and broken.
1700 properties
->supportedOperations
|=
1701 VK_SUBGROUP_FEATURE_ARITHMETIC_BIT
|
1702 VK_SUBGROUP_FEATURE_CLUSTERED_BIT
;
1704 properties
->quadOperationsInAllStages
= pdevice
->info
.gen
>= 8;
1708 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_PROPERTIES_EXT
: {
1709 VkPhysicalDeviceSubgroupSizeControlPropertiesEXT
*props
=
1710 (VkPhysicalDeviceSubgroupSizeControlPropertiesEXT
*)ext
;
1711 STATIC_ASSERT(8 <= BRW_SUBGROUP_SIZE
&& BRW_SUBGROUP_SIZE
<= 32);
1712 props
->minSubgroupSize
= 8;
1713 props
->maxSubgroupSize
= 32;
1714 props
->maxComputeWorkgroupSubgroups
= pdevice
->info
.max_cs_threads
;
1715 props
->requiredSubgroupSizeStages
= VK_SHADER_STAGE_COMPUTE_BIT
;
1718 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT_CONTROLS_PROPERTIES_KHR
: {
1719 VkPhysicalDeviceFloatControlsPropertiesKHR
*properties
= (void *)ext
;
1720 properties
->denormBehaviorIndependence
= VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_ALL_KHR
;
1721 properties
->roundingModeIndependence
= VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_NONE_KHR
;
1723 /* Broadwell does not support HF denorms and there are restrictions
1724 * other gens. According to Kabylake's PRM:
1726 * "math - Extended Math Function
1728 * Restriction : Half-float denorms are always retained."
1730 properties
->shaderDenormFlushToZeroFloat16
= false;
1731 properties
->shaderDenormPreserveFloat16
= pdevice
->info
.gen
> 8;
1732 properties
->shaderRoundingModeRTEFloat16
= true;
1733 properties
->shaderRoundingModeRTZFloat16
= true;
1734 properties
->shaderSignedZeroInfNanPreserveFloat16
= true;
1736 properties
->shaderDenormFlushToZeroFloat32
= true;
1737 properties
->shaderDenormPreserveFloat32
= true;
1738 properties
->shaderRoundingModeRTEFloat32
= true;
1739 properties
->shaderRoundingModeRTZFloat32
= true;
1740 properties
->shaderSignedZeroInfNanPreserveFloat32
= true;
1742 properties
->shaderDenormFlushToZeroFloat64
= true;
1743 properties
->shaderDenormPreserveFloat64
= true;
1744 properties
->shaderRoundingModeRTEFloat64
= true;
1745 properties
->shaderRoundingModeRTZFloat64
= true;
1746 properties
->shaderSignedZeroInfNanPreserveFloat64
= true;
1750 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_PROPERTIES_EXT
: {
1751 VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT
*props
=
1752 (VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT
*)ext
;
1754 /* From the SKL PRM Vol. 2d, docs for RENDER_SURFACE_STATE::Surface
1757 * "For SURFTYPE_BUFFER non-rendertarget surfaces, this field
1758 * specifies the base address of the first element of the surface,
1759 * computed in software by adding the surface base address to the
1760 * byte offset of the element in the buffer. The base address must
1761 * be aligned to element size."
1763 * The typed dataport messages require that things be texel aligned.
1764 * Otherwise, we may just load/store the wrong data or, in the worst
1765 * case, there may be hangs.
1767 props
->storageTexelBufferOffsetAlignmentBytes
= 16;
1768 props
->storageTexelBufferOffsetSingleTexelAlignment
= true;
1770 /* The sampler, however, is much more forgiving and it can handle
1771 * arbitrary byte alignment for linear and buffer surfaces. It's
1772 * hard to find a good PRM citation for this but years of empirical
1773 * experience demonstrate that this is true.
1775 props
->uniformTexelBufferOffsetAlignmentBytes
= 1;
1776 props
->uniformTexelBufferOffsetSingleTexelAlignment
= false;
1780 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_PROPERTIES_KHR
: {
1781 VkPhysicalDeviceTimelineSemaphorePropertiesKHR
*props
=
1782 (VkPhysicalDeviceTimelineSemaphorePropertiesKHR
*) ext
;
1783 props
->maxTimelineSemaphoreValueDifference
= UINT64_MAX
;
1787 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT
: {
1788 VkPhysicalDeviceTransformFeedbackPropertiesEXT
*props
=
1789 (VkPhysicalDeviceTransformFeedbackPropertiesEXT
*)ext
;
1791 props
->maxTransformFeedbackStreams
= MAX_XFB_STREAMS
;
1792 props
->maxTransformFeedbackBuffers
= MAX_XFB_BUFFERS
;
1793 props
->maxTransformFeedbackBufferSize
= (1ull << 32);
1794 props
->maxTransformFeedbackStreamDataSize
= 128 * 4;
1795 props
->maxTransformFeedbackBufferDataSize
= 128 * 4;
1796 props
->maxTransformFeedbackBufferDataStride
= 2048;
1797 props
->transformFeedbackQueries
= true;
1798 props
->transformFeedbackStreamsLinesTriangles
= false;
1799 props
->transformFeedbackRasterizationStreamSelect
= false;
1800 props
->transformFeedbackDraw
= true;
1804 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT
: {
1805 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*props
=
1806 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*)ext
;
1807 /* We have to restrict this a bit for multiview */
1808 props
->maxVertexAttribDivisor
= UINT32_MAX
/ 16;
1813 anv_debug_ignored_stype(ext
->sType
);
1819 /* We support exactly one queue family. */
1820 static const VkQueueFamilyProperties
1821 anv_queue_family_properties
= {
1822 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
1823 VK_QUEUE_COMPUTE_BIT
|
1824 VK_QUEUE_TRANSFER_BIT
,
1826 .timestampValidBits
= 36, /* XXX: Real value here */
1827 .minImageTransferGranularity
= { 1, 1, 1 },
1830 void anv_GetPhysicalDeviceQueueFamilyProperties(
1831 VkPhysicalDevice physicalDevice
,
1833 VkQueueFamilyProperties
* pQueueFamilyProperties
)
1835 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pCount
);
1837 vk_outarray_append(&out
, p
) {
1838 *p
= anv_queue_family_properties
;
1842 void anv_GetPhysicalDeviceQueueFamilyProperties2(
1843 VkPhysicalDevice physicalDevice
,
1844 uint32_t* pQueueFamilyPropertyCount
,
1845 VkQueueFamilyProperties2
* pQueueFamilyProperties
)
1848 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pQueueFamilyPropertyCount
);
1850 vk_outarray_append(&out
, p
) {
1851 p
->queueFamilyProperties
= anv_queue_family_properties
;
1853 vk_foreach_struct(s
, p
->pNext
) {
1854 anv_debug_ignored_stype(s
->sType
);
1859 void anv_GetPhysicalDeviceMemoryProperties(
1860 VkPhysicalDevice physicalDevice
,
1861 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
1863 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
1865 pMemoryProperties
->memoryTypeCount
= physical_device
->memory
.type_count
;
1866 for (uint32_t i
= 0; i
< physical_device
->memory
.type_count
; i
++) {
1867 pMemoryProperties
->memoryTypes
[i
] = (VkMemoryType
) {
1868 .propertyFlags
= physical_device
->memory
.types
[i
].propertyFlags
,
1869 .heapIndex
= physical_device
->memory
.types
[i
].heapIndex
,
1873 pMemoryProperties
->memoryHeapCount
= physical_device
->memory
.heap_count
;
1874 for (uint32_t i
= 0; i
< physical_device
->memory
.heap_count
; i
++) {
1875 pMemoryProperties
->memoryHeaps
[i
] = (VkMemoryHeap
) {
1876 .size
= physical_device
->memory
.heaps
[i
].size
,
1877 .flags
= physical_device
->memory
.heaps
[i
].flags
,
1883 anv_get_memory_budget(VkPhysicalDevice physicalDevice
,
1884 VkPhysicalDeviceMemoryBudgetPropertiesEXT
*memoryBudget
)
1886 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
1887 uint64_t sys_available
= get_available_system_memory();
1888 assert(sys_available
> 0);
1890 VkDeviceSize total_heaps_size
= 0;
1891 for (size_t i
= 0; i
< device
->memory
.heap_count
; i
++)
1892 total_heaps_size
+= device
->memory
.heaps
[i
].size
;
1894 for (size_t i
= 0; i
< device
->memory
.heap_count
; i
++) {
1895 VkDeviceSize heap_size
= device
->memory
.heaps
[i
].size
;
1896 VkDeviceSize heap_used
= device
->memory
.heaps
[i
].used
;
1897 VkDeviceSize heap_budget
;
1899 double heap_proportion
= (double) heap_size
/ total_heaps_size
;
1900 VkDeviceSize sys_available_prop
= sys_available
* heap_proportion
;
1903 * Let's not incite the app to starve the system: report at most 90% of
1904 * available system memory.
1906 uint64_t heap_available
= sys_available_prop
* 9 / 10;
1907 heap_budget
= MIN2(heap_size
, heap_used
+ heap_available
);
1910 * Round down to the nearest MB
1912 heap_budget
&= ~((1ull << 20) - 1);
1915 * The heapBudget value must be non-zero for array elements less than
1916 * VkPhysicalDeviceMemoryProperties::memoryHeapCount. The heapBudget
1917 * value must be less than or equal to VkMemoryHeap::size for each heap.
1919 assert(0 < heap_budget
&& heap_budget
<= heap_size
);
1921 memoryBudget
->heapUsage
[i
] = heap_used
;
1922 memoryBudget
->heapBudget
[i
] = heap_budget
;
1925 /* The heapBudget and heapUsage values must be zero for array elements
1926 * greater than or equal to VkPhysicalDeviceMemoryProperties::memoryHeapCount
1928 for (uint32_t i
= device
->memory
.heap_count
; i
< VK_MAX_MEMORY_HEAPS
; i
++) {
1929 memoryBudget
->heapBudget
[i
] = 0;
1930 memoryBudget
->heapUsage
[i
] = 0;
1934 void anv_GetPhysicalDeviceMemoryProperties2(
1935 VkPhysicalDevice physicalDevice
,
1936 VkPhysicalDeviceMemoryProperties2
* pMemoryProperties
)
1938 anv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
1939 &pMemoryProperties
->memoryProperties
);
1941 vk_foreach_struct(ext
, pMemoryProperties
->pNext
) {
1942 switch (ext
->sType
) {
1943 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT
:
1944 anv_get_memory_budget(physicalDevice
, (void*)ext
);
1947 anv_debug_ignored_stype(ext
->sType
);
1954 anv_GetDeviceGroupPeerMemoryFeatures(
1957 uint32_t localDeviceIndex
,
1958 uint32_t remoteDeviceIndex
,
1959 VkPeerMemoryFeatureFlags
* pPeerMemoryFeatures
)
1961 assert(localDeviceIndex
== 0 && remoteDeviceIndex
== 0);
1962 *pPeerMemoryFeatures
= VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT
|
1963 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT
|
1964 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT
|
1965 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT
;
1968 PFN_vkVoidFunction
anv_GetInstanceProcAddr(
1969 VkInstance _instance
,
1972 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1974 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
1975 * when we have to return valid function pointers, NULL, or it's left
1976 * undefined. See the table for exact details.
1981 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
1982 if (strcmp(pName, "vk" #entrypoint) == 0) \
1983 return (PFN_vkVoidFunction)anv_##entrypoint
1985 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties
);
1986 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties
);
1987 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceVersion
);
1988 LOOKUP_ANV_ENTRYPOINT(CreateInstance
);
1990 #undef LOOKUP_ANV_ENTRYPOINT
1992 if (instance
== NULL
)
1995 int idx
= anv_get_instance_entrypoint_index(pName
);
1997 return instance
->dispatch
.entrypoints
[idx
];
1999 idx
= anv_get_physical_device_entrypoint_index(pName
);
2001 return instance
->physicalDevice
.dispatch
.entrypoints
[idx
];
2003 idx
= anv_get_device_entrypoint_index(pName
);
2005 return instance
->device_dispatch
.entrypoints
[idx
];
2010 /* With version 1+ of the loader interface the ICD should expose
2011 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
2014 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
2015 VkInstance instance
,
2019 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
2020 VkInstance instance
,
2023 return anv_GetInstanceProcAddr(instance
, pName
);
2026 PFN_vkVoidFunction
anv_GetDeviceProcAddr(
2030 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2032 if (!device
|| !pName
)
2035 int idx
= anv_get_device_entrypoint_index(pName
);
2039 return device
->dispatch
.entrypoints
[idx
];
2042 /* With version 4+ of the loader interface the ICD should expose
2043 * vk_icdGetPhysicalDeviceProcAddr()
2046 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetPhysicalDeviceProcAddr(
2047 VkInstance _instance
,
2050 PFN_vkVoidFunction
vk_icdGetPhysicalDeviceProcAddr(
2051 VkInstance _instance
,
2054 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2056 if (!pName
|| !instance
)
2059 int idx
= anv_get_physical_device_entrypoint_index(pName
);
2063 return instance
->physicalDevice
.dispatch
.entrypoints
[idx
];
2068 anv_CreateDebugReportCallbackEXT(VkInstance _instance
,
2069 const VkDebugReportCallbackCreateInfoEXT
* pCreateInfo
,
2070 const VkAllocationCallbacks
* pAllocator
,
2071 VkDebugReportCallbackEXT
* pCallback
)
2073 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2074 return vk_create_debug_report_callback(&instance
->debug_report_callbacks
,
2075 pCreateInfo
, pAllocator
, &instance
->alloc
,
2080 anv_DestroyDebugReportCallbackEXT(VkInstance _instance
,
2081 VkDebugReportCallbackEXT _callback
,
2082 const VkAllocationCallbacks
* pAllocator
)
2084 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2085 vk_destroy_debug_report_callback(&instance
->debug_report_callbacks
,
2086 _callback
, pAllocator
, &instance
->alloc
);
2090 anv_DebugReportMessageEXT(VkInstance _instance
,
2091 VkDebugReportFlagsEXT flags
,
2092 VkDebugReportObjectTypeEXT objectType
,
2095 int32_t messageCode
,
2096 const char* pLayerPrefix
,
2097 const char* pMessage
)
2099 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2100 vk_debug_report(&instance
->debug_report_callbacks
, flags
, objectType
,
2101 object
, location
, messageCode
, pLayerPrefix
, pMessage
);
2104 static struct anv_state
2105 anv_state_pool_emit_data(struct anv_state_pool
*pool
, size_t size
, size_t align
, const void *p
)
2107 struct anv_state state
;
2109 state
= anv_state_pool_alloc(pool
, size
, align
);
2110 memcpy(state
.map
, p
, size
);
2115 /* Haswell border color is a bit of a disaster. Float and unorm formats use a
2116 * straightforward 32-bit float color in the first 64 bytes. Instead of using
2117 * a nice float/integer union like Gen8+, Haswell specifies the integer border
2118 * color as a separate entry /after/ the float color. The layout of this entry
2119 * also depends on the format's bpp (with extra hacks for RG32), and overlaps.
2121 * Since we don't know the format/bpp, we can't make any of the border colors
2122 * containing '1' work for all formats, as it would be in the wrong place for
2123 * some of them. We opt to make 32-bit integers work as this seems like the
2124 * most common option. Fortunately, transparent black works regardless, as
2125 * all zeroes is the same in every bit-size.
2127 struct hsw_border_color
{
2131 uint32_t _pad1
[108];
2134 struct gen8_border_color
{
2139 /* Pad out to 64 bytes */
2144 anv_device_init_border_colors(struct anv_device
*device
)
2146 if (device
->info
.is_haswell
) {
2147 static const struct hsw_border_color border_colors
[] = {
2148 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
2149 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
2150 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
2151 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
2152 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
2153 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
2156 device
->border_colors
=
2157 anv_state_pool_emit_data(&device
->dynamic_state_pool
,
2158 sizeof(border_colors
), 512, border_colors
);
2160 static const struct gen8_border_color border_colors
[] = {
2161 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
2162 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
2163 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
2164 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
2165 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
2166 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
2169 device
->border_colors
=
2170 anv_state_pool_emit_data(&device
->dynamic_state_pool
,
2171 sizeof(border_colors
), 64, border_colors
);
2176 anv_device_init_trivial_batch(struct anv_device
*device
)
2178 VkResult result
= anv_device_alloc_bo(device
, 4096,
2179 ANV_BO_ALLOC_MAPPED
,
2180 0 /* explicit_address */,
2181 &device
->trivial_batch_bo
);
2182 if (result
!= VK_SUCCESS
)
2185 struct anv_batch batch
= {
2186 .start
= device
->trivial_batch_bo
->map
,
2187 .next
= device
->trivial_batch_bo
->map
,
2188 .end
= device
->trivial_batch_bo
->map
+ 4096,
2191 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
2192 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
2194 if (!device
->info
.has_llc
)
2195 gen_clflush_range(batch
.start
, batch
.next
- batch
.start
);
2200 VkResult
anv_EnumerateDeviceExtensionProperties(
2201 VkPhysicalDevice physicalDevice
,
2202 const char* pLayerName
,
2203 uint32_t* pPropertyCount
,
2204 VkExtensionProperties
* pProperties
)
2206 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
2207 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
2209 for (int i
= 0; i
< ANV_DEVICE_EXTENSION_COUNT
; i
++) {
2210 if (device
->supported_extensions
.extensions
[i
]) {
2211 vk_outarray_append(&out
, prop
) {
2212 *prop
= anv_device_extensions
[i
];
2217 return vk_outarray_status(&out
);
2221 anv_device_init_dispatch(struct anv_device
*device
)
2223 const struct anv_device_dispatch_table
*genX_table
;
2224 switch (device
->info
.gen
) {
2226 genX_table
= &gen12_device_dispatch_table
;
2229 genX_table
= &gen11_device_dispatch_table
;
2232 genX_table
= &gen10_device_dispatch_table
;
2235 genX_table
= &gen9_device_dispatch_table
;
2238 genX_table
= &gen8_device_dispatch_table
;
2241 if (device
->info
.is_haswell
)
2242 genX_table
= &gen75_device_dispatch_table
;
2244 genX_table
= &gen7_device_dispatch_table
;
2247 unreachable("unsupported gen\n");
2250 for (unsigned i
= 0; i
< ARRAY_SIZE(device
->dispatch
.entrypoints
); i
++) {
2251 /* Vulkan requires that entrypoints for extensions which have not been
2252 * enabled must not be advertised.
2254 if (!anv_device_entrypoint_is_enabled(i
, device
->instance
->app_info
.api_version
,
2255 &device
->instance
->enabled_extensions
,
2256 &device
->enabled_extensions
)) {
2257 device
->dispatch
.entrypoints
[i
] = NULL
;
2258 } else if (genX_table
->entrypoints
[i
]) {
2259 device
->dispatch
.entrypoints
[i
] = genX_table
->entrypoints
[i
];
2261 device
->dispatch
.entrypoints
[i
] =
2262 anv_device_dispatch_table
.entrypoints
[i
];
2268 vk_priority_to_gen(int priority
)
2271 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT
:
2272 return GEN_CONTEXT_LOW_PRIORITY
;
2273 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
:
2274 return GEN_CONTEXT_MEDIUM_PRIORITY
;
2275 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT
:
2276 return GEN_CONTEXT_HIGH_PRIORITY
;
2277 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT
:
2278 return GEN_CONTEXT_REALTIME_PRIORITY
;
2280 unreachable("Invalid priority");
2285 anv_device_init_hiz_clear_value_bo(struct anv_device
*device
)
2287 VkResult result
= anv_device_alloc_bo(device
, 4096,
2288 ANV_BO_ALLOC_MAPPED
,
2289 0 /* explicit_address */,
2290 &device
->hiz_clear_bo
);
2291 if (result
!= VK_SUCCESS
)
2294 union isl_color_value hiz_clear
= { .u32
= { 0, } };
2295 hiz_clear
.f32
[0] = ANV_HZ_FC_VAL
;
2297 memcpy(device
->hiz_clear_bo
->map
, hiz_clear
.u32
, sizeof(hiz_clear
.u32
));
2299 if (!device
->info
.has_llc
)
2300 gen_clflush_range(device
->hiz_clear_bo
->map
, sizeof(hiz_clear
.u32
));
2306 get_bo_from_pool(struct gen_batch_decode_bo
*ret
,
2307 struct anv_block_pool
*pool
,
2310 anv_block_pool_foreach_bo(bo
, pool
) {
2311 uint64_t bo_address
= gen_48b_address(bo
->offset
);
2312 if (address
>= bo_address
&& address
< (bo_address
+ bo
->size
)) {
2313 *ret
= (struct gen_batch_decode_bo
) {
2324 /* Finding a buffer for batch decoding */
2325 static struct gen_batch_decode_bo
2326 decode_get_bo(void *v_batch
, bool ppgtt
, uint64_t address
)
2328 struct anv_device
*device
= v_batch
;
2329 struct gen_batch_decode_bo ret_bo
= {};
2333 if (get_bo_from_pool(&ret_bo
, &device
->dynamic_state_pool
.block_pool
, address
))
2335 if (get_bo_from_pool(&ret_bo
, &device
->instruction_state_pool
.block_pool
, address
))
2337 if (get_bo_from_pool(&ret_bo
, &device
->binding_table_pool
.block_pool
, address
))
2339 if (get_bo_from_pool(&ret_bo
, &device
->surface_state_pool
.block_pool
, address
))
2342 if (!device
->cmd_buffer_being_decoded
)
2343 return (struct gen_batch_decode_bo
) { };
2345 struct anv_batch_bo
**bo
;
2347 u_vector_foreach(bo
, &device
->cmd_buffer_being_decoded
->seen_bbos
) {
2348 /* The decoder zeroes out the top 16 bits, so we need to as well */
2349 uint64_t bo_address
= (*bo
)->bo
->offset
& (~0ull >> 16);
2351 if (address
>= bo_address
&& address
< bo_address
+ (*bo
)->bo
->size
) {
2352 return (struct gen_batch_decode_bo
) {
2354 .size
= (*bo
)->bo
->size
,
2355 .map
= (*bo
)->bo
->map
,
2360 return (struct gen_batch_decode_bo
) { };
2363 struct gen_aux_map_buffer
{
2364 struct gen_buffer base
;
2365 struct anv_state state
;
2368 static struct gen_buffer
*
2369 gen_aux_map_buffer_alloc(void *driver_ctx
, uint32_t size
)
2371 struct gen_aux_map_buffer
*buf
= malloc(sizeof(struct gen_aux_map_buffer
));
2375 struct anv_device
*device
= (struct anv_device
*)driver_ctx
;
2376 assert(device
->instance
->physicalDevice
.supports_48bit_addresses
&&
2377 device
->instance
->physicalDevice
.use_softpin
);
2379 struct anv_state_pool
*pool
= &device
->dynamic_state_pool
;
2380 buf
->state
= anv_state_pool_alloc(pool
, size
, size
);
2382 buf
->base
.gpu
= pool
->block_pool
.bo
->offset
+ buf
->state
.offset
;
2383 buf
->base
.gpu_end
= buf
->base
.gpu
+ buf
->state
.alloc_size
;
2384 buf
->base
.map
= buf
->state
.map
;
2385 buf
->base
.driver_bo
= &buf
->state
;
2390 gen_aux_map_buffer_free(void *driver_ctx
, struct gen_buffer
*buffer
)
2392 struct gen_aux_map_buffer
*buf
= (struct gen_aux_map_buffer
*)buffer
;
2393 struct anv_device
*device
= (struct anv_device
*)driver_ctx
;
2394 struct anv_state_pool
*pool
= &device
->dynamic_state_pool
;
2395 anv_state_pool_free(pool
, buf
->state
);
2399 static struct gen_mapped_pinned_buffer_alloc aux_map_allocator
= {
2400 .alloc
= gen_aux_map_buffer_alloc
,
2401 .free
= gen_aux_map_buffer_free
,
2404 VkResult
anv_CreateDevice(
2405 VkPhysicalDevice physicalDevice
,
2406 const VkDeviceCreateInfo
* pCreateInfo
,
2407 const VkAllocationCallbacks
* pAllocator
,
2410 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
2412 struct anv_device
*device
;
2414 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO
);
2416 struct anv_device_extension_table enabled_extensions
= { };
2417 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
2419 for (idx
= 0; idx
< ANV_DEVICE_EXTENSION_COUNT
; idx
++) {
2420 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
2421 anv_device_extensions
[idx
].extensionName
) == 0)
2425 if (idx
>= ANV_DEVICE_EXTENSION_COUNT
)
2426 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
2428 if (!physical_device
->supported_extensions
.extensions
[idx
])
2429 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
2431 enabled_extensions
.extensions
[idx
] = true;
2434 /* Check enabled features */
2435 if (pCreateInfo
->pEnabledFeatures
) {
2436 VkPhysicalDeviceFeatures supported_features
;
2437 anv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
2438 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
2439 VkBool32
*enabled_feature
= (VkBool32
*)pCreateInfo
->pEnabledFeatures
;
2440 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
2441 for (uint32_t i
= 0; i
< num_features
; i
++) {
2442 if (enabled_feature
[i
] && !supported_feature
[i
])
2443 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
2447 /* Check requested queues and fail if we are requested to create any
2448 * queues with flags we don't support.
2450 assert(pCreateInfo
->queueCreateInfoCount
> 0);
2451 for (uint32_t i
= 0; i
< pCreateInfo
->queueCreateInfoCount
; i
++) {
2452 if (pCreateInfo
->pQueueCreateInfos
[i
].flags
!= 0)
2453 return vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2456 /* Check if client specified queue priority. */
2457 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*queue_priority
=
2458 vk_find_struct_const(pCreateInfo
->pQueueCreateInfos
[0].pNext
,
2459 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT
);
2461 VkQueueGlobalPriorityEXT priority
=
2462 queue_priority
? queue_priority
->globalPriority
:
2463 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
;
2465 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
2467 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
2469 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2471 if (INTEL_DEBUG
& DEBUG_BATCH
) {
2472 const unsigned decode_flags
=
2473 GEN_BATCH_DECODE_FULL
|
2474 ((INTEL_DEBUG
& DEBUG_COLOR
) ? GEN_BATCH_DECODE_IN_COLOR
: 0) |
2475 GEN_BATCH_DECODE_OFFSETS
|
2476 GEN_BATCH_DECODE_FLOATS
;
2478 gen_batch_decode_ctx_init(&device
->decoder_ctx
,
2479 &physical_device
->info
,
2480 stderr
, decode_flags
, NULL
,
2481 decode_get_bo
, NULL
, device
);
2484 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
2485 device
->instance
= physical_device
->instance
;
2486 device
->chipset_id
= physical_device
->chipset_id
;
2487 device
->no_hw
= physical_device
->no_hw
;
2488 device
->_lost
= false;
2491 device
->alloc
= *pAllocator
;
2493 device
->alloc
= physical_device
->instance
->alloc
;
2495 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
2496 device
->fd
= open(physical_device
->path
, O_RDWR
| O_CLOEXEC
);
2497 if (device
->fd
== -1) {
2498 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2502 device
->context_id
= anv_gem_create_context(device
);
2503 if (device
->context_id
== -1) {
2504 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2508 result
= anv_queue_init(device
, &device
->queue
);
2509 if (result
!= VK_SUCCESS
)
2510 goto fail_context_id
;
2512 if (physical_device
->use_softpin
) {
2513 if (pthread_mutex_init(&device
->vma_mutex
, NULL
) != 0) {
2514 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2518 /* keep the page with address zero out of the allocator */
2519 util_vma_heap_init(&device
->vma_lo
,
2520 LOW_HEAP_MIN_ADDRESS
, LOW_HEAP_SIZE
);
2522 util_vma_heap_init(&device
->vma_cva
, CLIENT_VISIBLE_HEAP_MIN_ADDRESS
,
2523 CLIENT_VISIBLE_HEAP_SIZE
);
2525 /* Leave the last 4GiB out of the high vma range, so that no state
2526 * base address + size can overflow 48 bits. For more information see
2527 * the comment about Wa32bitGeneralStateOffset in anv_allocator.c
2529 util_vma_heap_init(&device
->vma_hi
, HIGH_HEAP_MIN_ADDRESS
,
2530 physical_device
->gtt_size
- (1ull << 32) -
2531 HIGH_HEAP_MIN_ADDRESS
);
2534 list_inithead(&device
->memory_objects
);
2536 /* As per spec, the driver implementation may deny requests to acquire
2537 * a priority above the default priority (MEDIUM) if the caller does not
2538 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
2541 if (physical_device
->has_context_priority
) {
2542 int err
= anv_gem_set_context_param(device
->fd
, device
->context_id
,
2543 I915_CONTEXT_PARAM_PRIORITY
,
2544 vk_priority_to_gen(priority
));
2545 if (err
!= 0 && priority
> VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
) {
2546 result
= vk_error(VK_ERROR_NOT_PERMITTED_EXT
);
2551 device
->info
= physical_device
->info
;
2552 device
->isl_dev
= physical_device
->isl_dev
;
2554 /* On Broadwell and later, we can use batch chaining to more efficiently
2555 * implement growing command buffers. Prior to Haswell, the kernel
2556 * command parser gets in the way and we have to fall back to growing
2559 device
->can_chain_batches
= device
->info
.gen
>= 8;
2561 device
->robust_buffer_access
= pCreateInfo
->pEnabledFeatures
&&
2562 pCreateInfo
->pEnabledFeatures
->robustBufferAccess
;
2563 device
->enabled_extensions
= enabled_extensions
;
2565 anv_device_init_dispatch(device
);
2567 if (pthread_mutex_init(&device
->mutex
, NULL
) != 0) {
2568 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2572 pthread_condattr_t condattr
;
2573 if (pthread_condattr_init(&condattr
) != 0) {
2574 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2577 if (pthread_condattr_setclock(&condattr
, CLOCK_MONOTONIC
) != 0) {
2578 pthread_condattr_destroy(&condattr
);
2579 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2582 if (pthread_cond_init(&device
->queue_submit
, &condattr
) != 0) {
2583 pthread_condattr_destroy(&condattr
);
2584 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2587 pthread_condattr_destroy(&condattr
);
2589 result
= anv_bo_cache_init(&device
->bo_cache
);
2590 if (result
!= VK_SUCCESS
)
2591 goto fail_queue_cond
;
2593 anv_bo_pool_init(&device
->batch_bo_pool
, device
);
2595 result
= anv_state_pool_init(&device
->dynamic_state_pool
, device
,
2596 DYNAMIC_STATE_POOL_MIN_ADDRESS
, 16384);
2597 if (result
!= VK_SUCCESS
)
2598 goto fail_batch_bo_pool
;
2600 result
= anv_state_pool_init(&device
->instruction_state_pool
, device
,
2601 INSTRUCTION_STATE_POOL_MIN_ADDRESS
, 16384);
2602 if (result
!= VK_SUCCESS
)
2603 goto fail_dynamic_state_pool
;
2605 result
= anv_state_pool_init(&device
->surface_state_pool
, device
,
2606 SURFACE_STATE_POOL_MIN_ADDRESS
, 4096);
2607 if (result
!= VK_SUCCESS
)
2608 goto fail_instruction_state_pool
;
2610 if (physical_device
->use_softpin
) {
2611 result
= anv_state_pool_init(&device
->binding_table_pool
, device
,
2612 BINDING_TABLE_POOL_MIN_ADDRESS
, 4096);
2613 if (result
!= VK_SUCCESS
)
2614 goto fail_surface_state_pool
;
2617 if (device
->info
.gen
>= 12) {
2618 device
->aux_map_ctx
= gen_aux_map_init(device
, &aux_map_allocator
,
2619 &physical_device
->info
);
2620 if (!device
->aux_map_ctx
)
2621 goto fail_binding_table_pool
;
2624 result
= anv_device_alloc_bo(device
, 4096, 0 /* flags */,
2625 0 /* explicit_address */,
2626 &device
->workaround_bo
);
2627 if (result
!= VK_SUCCESS
)
2628 goto fail_surface_aux_map_pool
;
2630 result
= anv_device_init_trivial_batch(device
);
2631 if (result
!= VK_SUCCESS
)
2632 goto fail_workaround_bo
;
2634 if (device
->info
.gen
>= 10) {
2635 result
= anv_device_init_hiz_clear_value_bo(device
);
2636 if (result
!= VK_SUCCESS
)
2637 goto fail_trivial_batch_bo
;
2640 anv_scratch_pool_init(device
, &device
->scratch_pool
);
2642 switch (device
->info
.gen
) {
2644 if (!device
->info
.is_haswell
)
2645 result
= gen7_init_device_state(device
);
2647 result
= gen75_init_device_state(device
);
2650 result
= gen8_init_device_state(device
);
2653 result
= gen9_init_device_state(device
);
2656 result
= gen10_init_device_state(device
);
2659 result
= gen11_init_device_state(device
);
2662 result
= gen12_init_device_state(device
);
2665 /* Shouldn't get here as we don't create physical devices for any other
2667 unreachable("unhandled gen");
2669 if (result
!= VK_SUCCESS
)
2670 goto fail_workaround_bo
;
2672 anv_pipeline_cache_init(&device
->default_pipeline_cache
, device
, true);
2674 anv_device_init_blorp(device
);
2676 anv_device_init_border_colors(device
);
2678 anv_device_perf_init(device
);
2680 *pDevice
= anv_device_to_handle(device
);
2685 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
2686 if (device
->info
.gen
>= 10)
2687 anv_device_release_bo(device
, device
->hiz_clear_bo
);
2688 anv_device_release_bo(device
, device
->workaround_bo
);
2689 fail_trivial_batch_bo
:
2690 anv_device_release_bo(device
, device
->trivial_batch_bo
);
2691 fail_surface_aux_map_pool
:
2692 if (device
->info
.gen
>= 12) {
2693 gen_aux_map_finish(device
->aux_map_ctx
);
2694 device
->aux_map_ctx
= NULL
;
2696 fail_binding_table_pool
:
2697 if (physical_device
->use_softpin
)
2698 anv_state_pool_finish(&device
->binding_table_pool
);
2699 fail_surface_state_pool
:
2700 anv_state_pool_finish(&device
->surface_state_pool
);
2701 fail_instruction_state_pool
:
2702 anv_state_pool_finish(&device
->instruction_state_pool
);
2703 fail_dynamic_state_pool
:
2704 anv_state_pool_finish(&device
->dynamic_state_pool
);
2706 anv_bo_pool_finish(&device
->batch_bo_pool
);
2707 anv_bo_cache_finish(&device
->bo_cache
);
2709 pthread_cond_destroy(&device
->queue_submit
);
2711 pthread_mutex_destroy(&device
->mutex
);
2713 if (physical_device
->use_softpin
) {
2714 util_vma_heap_finish(&device
->vma_hi
);
2715 util_vma_heap_finish(&device
->vma_cva
);
2716 util_vma_heap_finish(&device
->vma_lo
);
2719 anv_queue_finish(&device
->queue
);
2721 anv_gem_destroy_context(device
, device
->context_id
);
2725 vk_free(&device
->alloc
, device
);
2730 void anv_DestroyDevice(
2732 const VkAllocationCallbacks
* pAllocator
)
2734 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2735 struct anv_physical_device
*physical_device
;
2740 physical_device
= &device
->instance
->physicalDevice
;
2742 anv_device_finish_blorp(device
);
2744 anv_pipeline_cache_finish(&device
->default_pipeline_cache
);
2746 anv_queue_finish(&device
->queue
);
2748 #ifdef HAVE_VALGRIND
2749 /* We only need to free these to prevent valgrind errors. The backing
2750 * BO will go away in a couple of lines so we don't actually leak.
2752 anv_state_pool_free(&device
->dynamic_state_pool
, device
->border_colors
);
2753 anv_state_pool_free(&device
->dynamic_state_pool
, device
->slice_hash
);
2756 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
2758 anv_device_release_bo(device
, device
->workaround_bo
);
2759 anv_device_release_bo(device
, device
->trivial_batch_bo
);
2760 if (device
->info
.gen
>= 10)
2761 anv_device_release_bo(device
, device
->hiz_clear_bo
);
2763 if (device
->info
.gen
>= 12) {
2764 gen_aux_map_finish(device
->aux_map_ctx
);
2765 device
->aux_map_ctx
= NULL
;
2768 if (physical_device
->use_softpin
)
2769 anv_state_pool_finish(&device
->binding_table_pool
);
2770 anv_state_pool_finish(&device
->surface_state_pool
);
2771 anv_state_pool_finish(&device
->instruction_state_pool
);
2772 anv_state_pool_finish(&device
->dynamic_state_pool
);
2774 anv_bo_pool_finish(&device
->batch_bo_pool
);
2776 anv_bo_cache_finish(&device
->bo_cache
);
2778 if (physical_device
->use_softpin
) {
2779 util_vma_heap_finish(&device
->vma_hi
);
2780 util_vma_heap_finish(&device
->vma_cva
);
2781 util_vma_heap_finish(&device
->vma_lo
);
2784 pthread_cond_destroy(&device
->queue_submit
);
2785 pthread_mutex_destroy(&device
->mutex
);
2787 anv_gem_destroy_context(device
, device
->context_id
);
2789 if (INTEL_DEBUG
& DEBUG_BATCH
)
2790 gen_batch_decode_ctx_finish(&device
->decoder_ctx
);
2794 vk_free(&device
->alloc
, device
);
2797 VkResult
anv_EnumerateInstanceLayerProperties(
2798 uint32_t* pPropertyCount
,
2799 VkLayerProperties
* pProperties
)
2801 if (pProperties
== NULL
) {
2802 *pPropertyCount
= 0;
2806 /* None supported at this time */
2807 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
2810 VkResult
anv_EnumerateDeviceLayerProperties(
2811 VkPhysicalDevice physicalDevice
,
2812 uint32_t* pPropertyCount
,
2813 VkLayerProperties
* pProperties
)
2815 if (pProperties
== NULL
) {
2816 *pPropertyCount
= 0;
2820 /* None supported at this time */
2821 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
2824 void anv_GetDeviceQueue(
2826 uint32_t queueNodeIndex
,
2827 uint32_t queueIndex
,
2830 const VkDeviceQueueInfo2 info
= {
2831 .sType
= VK_STRUCTURE_TYPE_DEVICE_QUEUE_INFO_2
,
2834 .queueFamilyIndex
= queueNodeIndex
,
2835 .queueIndex
= queueIndex
,
2838 anv_GetDeviceQueue2(_device
, &info
, pQueue
);
2841 void anv_GetDeviceQueue2(
2843 const VkDeviceQueueInfo2
* pQueueInfo
,
2846 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2848 assert(pQueueInfo
->queueIndex
== 0);
2850 if (pQueueInfo
->flags
== device
->queue
.flags
)
2851 *pQueue
= anv_queue_to_handle(&device
->queue
);
2857 _anv_device_set_lost(struct anv_device
*device
,
2858 const char *file
, int line
,
2859 const char *msg
, ...)
2864 p_atomic_inc(&device
->_lost
);
2867 err
= __vk_errorv(device
->instance
, device
,
2868 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT
,
2869 VK_ERROR_DEVICE_LOST
, file
, line
, msg
, ap
);
2872 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
2879 _anv_queue_set_lost(struct anv_queue
*queue
,
2880 const char *file
, int line
,
2881 const char *msg
, ...)
2886 p_atomic_inc(&queue
->device
->_lost
);
2889 err
= __vk_errorv(queue
->device
->instance
, queue
->device
,
2890 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT
,
2891 VK_ERROR_DEVICE_LOST
, file
, line
, msg
, ap
);
2894 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
2901 anv_device_query_status(struct anv_device
*device
)
2903 /* This isn't likely as most of the callers of this function already check
2904 * for it. However, it doesn't hurt to check and it potentially lets us
2907 if (anv_device_is_lost(device
))
2908 return VK_ERROR_DEVICE_LOST
;
2910 uint32_t active
, pending
;
2911 int ret
= anv_gem_gpu_get_reset_stats(device
, &active
, &pending
);
2913 /* We don't know the real error. */
2914 return anv_device_set_lost(device
, "get_reset_stats failed: %m");
2918 return anv_device_set_lost(device
, "GPU hung on one of our command buffers");
2919 } else if (pending
) {
2920 return anv_device_set_lost(device
, "GPU hung with commands in-flight");
2927 anv_device_bo_busy(struct anv_device
*device
, struct anv_bo
*bo
)
2929 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
2930 * Other usages of the BO (such as on different hardware) will not be
2931 * flagged as "busy" by this ioctl. Use with care.
2933 int ret
= anv_gem_busy(device
, bo
->gem_handle
);
2935 return VK_NOT_READY
;
2936 } else if (ret
== -1) {
2937 /* We don't know the real error. */
2938 return anv_device_set_lost(device
, "gem wait failed: %m");
2941 /* Query for device status after the busy call. If the BO we're checking
2942 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
2943 * client because it clearly doesn't have valid data. Yes, this most
2944 * likely means an ioctl, but we just did an ioctl to query the busy status
2945 * so it's no great loss.
2947 return anv_device_query_status(device
);
2951 anv_device_wait(struct anv_device
*device
, struct anv_bo
*bo
,
2954 int ret
= anv_gem_wait(device
, bo
->gem_handle
, &timeout
);
2955 if (ret
== -1 && errno
== ETIME
) {
2957 } else if (ret
== -1) {
2958 /* We don't know the real error. */
2959 return anv_device_set_lost(device
, "gem wait failed: %m");
2962 /* Query for device status after the wait. If the BO we're waiting on got
2963 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
2964 * because it clearly doesn't have valid data. Yes, this most likely means
2965 * an ioctl, but we just did an ioctl to wait so it's no great loss.
2967 return anv_device_query_status(device
);
2970 VkResult
anv_DeviceWaitIdle(
2973 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2975 if (anv_device_is_lost(device
))
2976 return VK_ERROR_DEVICE_LOST
;
2978 return anv_queue_submit_simple_batch(&device
->queue
, NULL
);
2982 anv_vma_alloc(struct anv_device
*device
, struct anv_bo
*bo
,
2983 uint64_t client_address
)
2985 const struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2986 const struct gen_device_info
*devinfo
= &pdevice
->info
;
2987 /* Gen12 CCS surface addresses need to be 64K aligned. We have no way of
2988 * telling what this allocation is for so pick the largest alignment.
2990 const uint32_t vma_alignment
=
2991 devinfo
->gen
>= 12 ? (64 * 1024) : (4 * 1024);
2993 if (!(bo
->flags
& EXEC_OBJECT_PINNED
)) {
2994 assert(!(bo
->has_client_visible_address
));
2998 pthread_mutex_lock(&device
->vma_mutex
);
3002 if (bo
->has_client_visible_address
) {
3003 assert(bo
->flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
);
3004 if (client_address
) {
3005 if (util_vma_heap_alloc_addr(&device
->vma_cva
,
3006 client_address
, bo
->size
)) {
3007 bo
->offset
= gen_canonical_address(client_address
);
3011 util_vma_heap_alloc(&device
->vma_cva
, bo
->size
, vma_alignment
);
3013 bo
->offset
= gen_canonical_address(addr
);
3014 assert(addr
== gen_48b_address(bo
->offset
));
3017 /* We don't want to fall back to other heaps */
3021 assert(client_address
== 0);
3023 if (bo
->flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
) {
3025 util_vma_heap_alloc(&device
->vma_hi
, bo
->size
, vma_alignment
);
3027 bo
->offset
= gen_canonical_address(addr
);
3028 assert(addr
== gen_48b_address(bo
->offset
));
3032 if (bo
->offset
== 0) {
3034 util_vma_heap_alloc(&device
->vma_lo
, bo
->size
, vma_alignment
);
3036 bo
->offset
= gen_canonical_address(addr
);
3037 assert(addr
== gen_48b_address(bo
->offset
));
3042 pthread_mutex_unlock(&device
->vma_mutex
);
3044 return bo
->offset
!= 0;
3048 anv_vma_free(struct anv_device
*device
, struct anv_bo
*bo
)
3050 if (!(bo
->flags
& EXEC_OBJECT_PINNED
))
3053 const uint64_t addr_48b
= gen_48b_address(bo
->offset
);
3055 pthread_mutex_lock(&device
->vma_mutex
);
3057 if (addr_48b
>= LOW_HEAP_MIN_ADDRESS
&&
3058 addr_48b
<= LOW_HEAP_MAX_ADDRESS
) {
3059 util_vma_heap_free(&device
->vma_lo
, addr_48b
, bo
->size
);
3060 } else if (addr_48b
>= CLIENT_VISIBLE_HEAP_MIN_ADDRESS
&&
3061 addr_48b
<= CLIENT_VISIBLE_HEAP_MAX_ADDRESS
) {
3062 util_vma_heap_free(&device
->vma_cva
, addr_48b
, bo
->size
);
3064 assert(addr_48b
>= HIGH_HEAP_MIN_ADDRESS
);
3065 util_vma_heap_free(&device
->vma_hi
, addr_48b
, bo
->size
);
3068 pthread_mutex_unlock(&device
->vma_mutex
);
3073 VkResult
anv_AllocateMemory(
3075 const VkMemoryAllocateInfo
* pAllocateInfo
,
3076 const VkAllocationCallbacks
* pAllocator
,
3077 VkDeviceMemory
* pMem
)
3079 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3080 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3081 struct anv_device_memory
*mem
;
3082 VkResult result
= VK_SUCCESS
;
3084 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
3086 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
3087 assert(pAllocateInfo
->allocationSize
> 0);
3089 VkDeviceSize aligned_alloc_size
=
3090 align_u64(pAllocateInfo
->allocationSize
, 4096);
3092 if (aligned_alloc_size
> MAX_MEMORY_ALLOCATION_SIZE
)
3093 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
3095 assert(pAllocateInfo
->memoryTypeIndex
< pdevice
->memory
.type_count
);
3096 struct anv_memory_type
*mem_type
=
3097 &pdevice
->memory
.types
[pAllocateInfo
->memoryTypeIndex
];
3098 assert(mem_type
->heapIndex
< pdevice
->memory
.heap_count
);
3099 struct anv_memory_heap
*mem_heap
=
3100 &pdevice
->memory
.heaps
[mem_type
->heapIndex
];
3102 uint64_t mem_heap_used
= p_atomic_read(&mem_heap
->used
);
3103 if (mem_heap_used
+ aligned_alloc_size
> mem_heap
->size
)
3104 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
3106 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
3107 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3109 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3111 assert(pAllocateInfo
->memoryTypeIndex
< pdevice
->memory
.type_count
);
3112 mem
->type
= mem_type
;
3116 mem
->host_ptr
= NULL
;
3118 enum anv_bo_alloc_flags alloc_flags
= 0;
3120 const VkExportMemoryAllocateInfo
*export_info
= NULL
;
3121 const VkImportAndroidHardwareBufferInfoANDROID
*ahw_import_info
= NULL
;
3122 const VkImportMemoryFdInfoKHR
*fd_info
= NULL
;
3123 const VkImportMemoryHostPointerInfoEXT
*host_ptr_info
= NULL
;
3124 const VkMemoryDedicatedAllocateInfo
*dedicated_info
= NULL
;
3125 VkMemoryAllocateFlags vk_flags
= 0;
3126 uint64_t client_address
= 0;
3128 vk_foreach_struct_const(ext
, pAllocateInfo
->pNext
) {
3129 switch (ext
->sType
) {
3130 case VK_STRUCTURE_TYPE_EXPORT_MEMORY_ALLOCATE_INFO
:
3131 export_info
= (void *)ext
;
3134 case VK_STRUCTURE_TYPE_IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID
:
3135 ahw_import_info
= (void *)ext
;
3138 case VK_STRUCTURE_TYPE_IMPORT_MEMORY_FD_INFO_KHR
:
3139 fd_info
= (void *)ext
;
3142 case VK_STRUCTURE_TYPE_IMPORT_MEMORY_HOST_POINTER_INFO_EXT
:
3143 host_ptr_info
= (void *)ext
;
3146 case VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_FLAGS_INFO
: {
3147 const VkMemoryAllocateFlagsInfo
*flags_info
= (void *)ext
;
3148 vk_flags
= flags_info
->flags
;
3152 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_ALLOCATE_INFO
:
3153 dedicated_info
= (void *)ext
;
3156 case VK_STRUCTURE_TYPE_MEMORY_OPAQUE_CAPTURE_ADDRESS_ALLOCATE_INFO_KHR
: {
3157 const VkMemoryOpaqueCaptureAddressAllocateInfoKHR
*addr_info
=
3158 (const VkMemoryOpaqueCaptureAddressAllocateInfoKHR
*)ext
;
3159 client_address
= addr_info
->opaqueCaptureAddress
;
3164 anv_debug_ignored_stype(ext
->sType
);
3169 if (vk_flags
& VK_MEMORY_ALLOCATE_DEVICE_ADDRESS_BIT_KHR
)
3170 alloc_flags
|= ANV_BO_ALLOC_CLIENT_VISIBLE_ADDRESS
;
3172 /* Check if we need to support Android HW buffer export. If so,
3173 * create AHardwareBuffer and import memory from it.
3175 bool android_export
= false;
3176 if (export_info
&& export_info
->handleTypes
&
3177 VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID
)
3178 android_export
= true;
3180 if (ahw_import_info
) {
3181 result
= anv_import_ahw_memory(_device
, mem
, ahw_import_info
);
3182 if (result
!= VK_SUCCESS
)
3186 } else if (android_export
) {
3187 result
= anv_create_ahw_memory(_device
, mem
, pAllocateInfo
);
3188 if (result
!= VK_SUCCESS
)
3191 const VkImportAndroidHardwareBufferInfoANDROID import_info
= {
3194 result
= anv_import_ahw_memory(_device
, mem
, &import_info
);
3195 if (result
!= VK_SUCCESS
)
3201 /* The Vulkan spec permits handleType to be 0, in which case the struct is
3204 if (fd_info
&& fd_info
->handleType
) {
3205 /* At the moment, we support only the below handle types. */
3206 assert(fd_info
->handleType
==
3207 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
3208 fd_info
->handleType
==
3209 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
3211 result
= anv_device_import_bo(device
, fd_info
->fd
, alloc_flags
,
3212 client_address
, &mem
->bo
);
3213 if (result
!= VK_SUCCESS
)
3216 VkDeviceSize aligned_alloc_size
=
3217 align_u64(pAllocateInfo
->allocationSize
, 4096);
3219 /* For security purposes, we reject importing the bo if it's smaller
3220 * than the requested allocation size. This prevents a malicious client
3221 * from passing a buffer to a trusted client, lying about the size, and
3222 * telling the trusted client to try and texture from an image that goes
3223 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
3224 * in the trusted client. The trusted client can protect itself against
3225 * this sort of attack but only if it can trust the buffer size.
3227 if (mem
->bo
->size
< aligned_alloc_size
) {
3228 result
= vk_errorf(device
->instance
, device
,
3229 VK_ERROR_INVALID_EXTERNAL_HANDLE
,
3230 "aligned allocationSize too large for "
3231 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT: "
3232 "%"PRIu64
"B > %"PRIu64
"B",
3233 aligned_alloc_size
, mem
->bo
->size
);
3234 anv_device_release_bo(device
, mem
->bo
);
3238 /* From the Vulkan spec:
3240 * "Importing memory from a file descriptor transfers ownership of
3241 * the file descriptor from the application to the Vulkan
3242 * implementation. The application must not perform any operations on
3243 * the file descriptor after a successful import."
3245 * If the import fails, we leave the file descriptor open.
3251 if (host_ptr_info
&& host_ptr_info
->handleType
) {
3252 if (host_ptr_info
->handleType
==
3253 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_MAPPED_FOREIGN_MEMORY_BIT_EXT
) {
3254 result
= vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
3258 assert(host_ptr_info
->handleType
==
3259 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
);
3261 result
= anv_device_import_bo_from_host_ptr(device
,
3262 host_ptr_info
->pHostPointer
,
3263 pAllocateInfo
->allocationSize
,
3267 if (result
!= VK_SUCCESS
)
3270 mem
->host_ptr
= host_ptr_info
->pHostPointer
;
3274 /* Regular allocate (not importing memory). */
3276 if (export_info
&& export_info
->handleTypes
)
3277 alloc_flags
|= ANV_BO_ALLOC_EXTERNAL
;
3279 result
= anv_device_alloc_bo(device
, pAllocateInfo
->allocationSize
,
3280 alloc_flags
, client_address
, &mem
->bo
);
3281 if (result
!= VK_SUCCESS
)
3284 if (dedicated_info
&& dedicated_info
->image
!= VK_NULL_HANDLE
) {
3285 ANV_FROM_HANDLE(anv_image
, image
, dedicated_info
->image
);
3287 /* Some legacy (non-modifiers) consumers need the tiling to be set on
3288 * the BO. In this case, we have a dedicated allocation.
3290 if (image
->needs_set_tiling
) {
3291 const uint32_t i915_tiling
=
3292 isl_tiling_to_i915_tiling(image
->planes
[0].surface
.isl
.tiling
);
3293 int ret
= anv_gem_set_tiling(device
, mem
->bo
->gem_handle
,
3294 image
->planes
[0].surface
.isl
.row_pitch_B
,
3297 anv_device_release_bo(device
, mem
->bo
);
3298 result
= vk_errorf(device
->instance
, NULL
,
3299 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
3300 "failed to set BO tiling: %m");
3307 mem_heap_used
= p_atomic_add_return(&mem_heap
->used
, mem
->bo
->size
);
3308 if (mem_heap_used
> mem_heap
->size
) {
3309 p_atomic_add(&mem_heap
->used
, -mem
->bo
->size
);
3310 anv_device_release_bo(device
, mem
->bo
);
3311 result
= vk_errorf(device
->instance
, NULL
,
3312 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
3313 "Out of heap memory");
3317 pthread_mutex_lock(&device
->mutex
);
3318 list_addtail(&mem
->link
, &device
->memory_objects
);
3319 pthread_mutex_unlock(&device
->mutex
);
3321 *pMem
= anv_device_memory_to_handle(mem
);
3326 vk_free2(&device
->alloc
, pAllocator
, mem
);
3331 VkResult
anv_GetMemoryFdKHR(
3333 const VkMemoryGetFdInfoKHR
* pGetFdInfo
,
3336 ANV_FROM_HANDLE(anv_device
, dev
, device_h
);
3337 ANV_FROM_HANDLE(anv_device_memory
, mem
, pGetFdInfo
->memory
);
3339 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
3341 assert(pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
3342 pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
3344 return anv_device_export_bo(dev
, mem
->bo
, pFd
);
3347 VkResult
anv_GetMemoryFdPropertiesKHR(
3349 VkExternalMemoryHandleTypeFlagBits handleType
,
3351 VkMemoryFdPropertiesKHR
* pMemoryFdProperties
)
3353 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3354 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3356 switch (handleType
) {
3357 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
:
3358 /* dma-buf can be imported as any memory type */
3359 pMemoryFdProperties
->memoryTypeBits
=
3360 (1 << pdevice
->memory
.type_count
) - 1;
3364 /* The valid usage section for this function says:
3366 * "handleType must not be one of the handle types defined as
3369 * So opaque handle types fall into the default "unsupported" case.
3371 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
3375 VkResult
anv_GetMemoryHostPointerPropertiesEXT(
3377 VkExternalMemoryHandleTypeFlagBits handleType
,
3378 const void* pHostPointer
,
3379 VkMemoryHostPointerPropertiesEXT
* pMemoryHostPointerProperties
)
3381 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3383 assert(pMemoryHostPointerProperties
->sType
==
3384 VK_STRUCTURE_TYPE_MEMORY_HOST_POINTER_PROPERTIES_EXT
);
3386 switch (handleType
) {
3387 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
: {
3388 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3390 /* Host memory can be imported as any memory type. */
3391 pMemoryHostPointerProperties
->memoryTypeBits
=
3392 (1ull << pdevice
->memory
.type_count
) - 1;
3397 return VK_ERROR_INVALID_EXTERNAL_HANDLE
;
3401 void anv_FreeMemory(
3403 VkDeviceMemory _mem
,
3404 const VkAllocationCallbacks
* pAllocator
)
3406 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3407 ANV_FROM_HANDLE(anv_device_memory
, mem
, _mem
);
3408 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3413 pthread_mutex_lock(&device
->mutex
);
3414 list_del(&mem
->link
);
3415 pthread_mutex_unlock(&device
->mutex
);
3418 anv_UnmapMemory(_device
, _mem
);
3420 p_atomic_add(&pdevice
->memory
.heaps
[mem
->type
->heapIndex
].used
,
3423 anv_device_release_bo(device
, mem
->bo
);
3425 #if defined(ANDROID) && ANDROID_API_LEVEL >= 26
3427 AHardwareBuffer_release(mem
->ahw
);
3430 vk_free2(&device
->alloc
, pAllocator
, mem
);
3433 VkResult
anv_MapMemory(
3435 VkDeviceMemory _memory
,
3436 VkDeviceSize offset
,
3438 VkMemoryMapFlags flags
,
3441 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3442 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
3449 if (mem
->host_ptr
) {
3450 *ppData
= mem
->host_ptr
+ offset
;
3454 if (size
== VK_WHOLE_SIZE
)
3455 size
= mem
->bo
->size
- offset
;
3457 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
3459 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
3460 * assert(size != 0);
3461 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
3462 * equal to the size of the memory minus offset
3465 assert(offset
+ size
<= mem
->bo
->size
);
3467 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
3468 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
3469 * at a time is valid. We could just mmap up front and return an offset
3470 * pointer here, but that may exhaust virtual memory on 32 bit
3473 uint32_t gem_flags
= 0;
3475 if (!device
->info
.has_llc
&&
3476 (mem
->type
->propertyFlags
& VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
))
3477 gem_flags
|= I915_MMAP_WC
;
3479 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
3480 uint64_t map_offset
= offset
& ~4095ull;
3481 assert(offset
>= map_offset
);
3482 uint64_t map_size
= (offset
+ size
) - map_offset
;
3484 /* Let's map whole pages */
3485 map_size
= align_u64(map_size
, 4096);
3487 void *map
= anv_gem_mmap(device
, mem
->bo
->gem_handle
,
3488 map_offset
, map_size
, gem_flags
);
3489 if (map
== MAP_FAILED
)
3490 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
3493 mem
->map_size
= map_size
;
3495 *ppData
= mem
->map
+ (offset
- map_offset
);
3500 void anv_UnmapMemory(
3502 VkDeviceMemory _memory
)
3504 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
3506 if (mem
== NULL
|| mem
->host_ptr
)
3509 anv_gem_munmap(mem
->map
, mem
->map_size
);
3516 clflush_mapped_ranges(struct anv_device
*device
,
3518 const VkMappedMemoryRange
*ranges
)
3520 for (uint32_t i
= 0; i
< count
; i
++) {
3521 ANV_FROM_HANDLE(anv_device_memory
, mem
, ranges
[i
].memory
);
3522 if (ranges
[i
].offset
>= mem
->map_size
)
3525 gen_clflush_range(mem
->map
+ ranges
[i
].offset
,
3526 MIN2(ranges
[i
].size
, mem
->map_size
- ranges
[i
].offset
));
3530 VkResult
anv_FlushMappedMemoryRanges(
3532 uint32_t memoryRangeCount
,
3533 const VkMappedMemoryRange
* pMemoryRanges
)
3535 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3537 if (device
->info
.has_llc
)
3540 /* Make sure the writes we're flushing have landed. */
3541 __builtin_ia32_mfence();
3543 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
3548 VkResult
anv_InvalidateMappedMemoryRanges(
3550 uint32_t memoryRangeCount
,
3551 const VkMappedMemoryRange
* pMemoryRanges
)
3553 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3555 if (device
->info
.has_llc
)
3558 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
3560 /* Make sure no reads get moved up above the invalidate. */
3561 __builtin_ia32_mfence();
3566 void anv_GetBufferMemoryRequirements(
3569 VkMemoryRequirements
* pMemoryRequirements
)
3571 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
3572 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3573 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3575 /* The Vulkan spec (git aaed022) says:
3577 * memoryTypeBits is a bitfield and contains one bit set for every
3578 * supported memory type for the resource. The bit `1<<i` is set if and
3579 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3580 * structure for the physical device is supported.
3582 uint32_t memory_types
= (1ull << pdevice
->memory
.type_count
) - 1;
3584 /* Base alignment requirement of a cache line */
3585 uint32_t alignment
= 16;
3587 /* We need an alignment of 32 for pushing UBOs */
3588 if (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
)
3589 alignment
= MAX2(alignment
, 32);
3591 pMemoryRequirements
->size
= buffer
->size
;
3592 pMemoryRequirements
->alignment
= alignment
;
3594 /* Storage and Uniform buffers should have their size aligned to
3595 * 32-bits to avoid boundary checks when last DWord is not complete.
3596 * This would ensure that not internal padding would be needed for
3599 if (device
->robust_buffer_access
&&
3600 (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
||
3601 buffer
->usage
& VK_BUFFER_USAGE_STORAGE_BUFFER_BIT
))
3602 pMemoryRequirements
->size
= align_u64(buffer
->size
, 4);
3604 pMemoryRequirements
->memoryTypeBits
= memory_types
;
3607 void anv_GetBufferMemoryRequirements2(
3609 const VkBufferMemoryRequirementsInfo2
* pInfo
,
3610 VkMemoryRequirements2
* pMemoryRequirements
)
3612 anv_GetBufferMemoryRequirements(_device
, pInfo
->buffer
,
3613 &pMemoryRequirements
->memoryRequirements
);
3615 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3616 switch (ext
->sType
) {
3617 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
3618 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
3619 requirements
->prefersDedicatedAllocation
= false;
3620 requirements
->requiresDedicatedAllocation
= false;
3625 anv_debug_ignored_stype(ext
->sType
);
3631 void anv_GetImageMemoryRequirements(
3634 VkMemoryRequirements
* pMemoryRequirements
)
3636 ANV_FROM_HANDLE(anv_image
, image
, _image
);
3637 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3638 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3640 /* The Vulkan spec (git aaed022) says:
3642 * memoryTypeBits is a bitfield and contains one bit set for every
3643 * supported memory type for the resource. The bit `1<<i` is set if and
3644 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3645 * structure for the physical device is supported.
3647 * All types are currently supported for images.
3649 uint32_t memory_types
= (1ull << pdevice
->memory
.type_count
) - 1;
3651 /* We must have image allocated or imported at this point. According to the
3652 * specification, external images must have been bound to memory before
3653 * calling GetImageMemoryRequirements.
3655 assert(image
->size
> 0);
3657 pMemoryRequirements
->size
= image
->size
;
3658 pMemoryRequirements
->alignment
= image
->alignment
;
3659 pMemoryRequirements
->memoryTypeBits
= memory_types
;
3662 void anv_GetImageMemoryRequirements2(
3664 const VkImageMemoryRequirementsInfo2
* pInfo
,
3665 VkMemoryRequirements2
* pMemoryRequirements
)
3667 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3668 ANV_FROM_HANDLE(anv_image
, image
, pInfo
->image
);
3670 anv_GetImageMemoryRequirements(_device
, pInfo
->image
,
3671 &pMemoryRequirements
->memoryRequirements
);
3673 vk_foreach_struct_const(ext
, pInfo
->pNext
) {
3674 switch (ext
->sType
) {
3675 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO
: {
3676 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3677 const VkImagePlaneMemoryRequirementsInfo
*plane_reqs
=
3678 (const VkImagePlaneMemoryRequirementsInfo
*) ext
;
3679 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
,
3680 plane_reqs
->planeAspect
);
3682 assert(image
->planes
[plane
].offset
== 0);
3684 /* The Vulkan spec (git aaed022) says:
3686 * memoryTypeBits is a bitfield and contains one bit set for every
3687 * supported memory type for the resource. The bit `1<<i` is set
3688 * if and only if the memory type `i` in the
3689 * VkPhysicalDeviceMemoryProperties structure for the physical
3690 * device is supported.
3692 * All types are currently supported for images.
3694 pMemoryRequirements
->memoryRequirements
.memoryTypeBits
=
3695 (1ull << pdevice
->memory
.type_count
) - 1;
3697 /* We must have image allocated or imported at this point. According to the
3698 * specification, external images must have been bound to memory before
3699 * calling GetImageMemoryRequirements.
3701 assert(image
->planes
[plane
].size
> 0);
3703 pMemoryRequirements
->memoryRequirements
.size
= image
->planes
[plane
].size
;
3704 pMemoryRequirements
->memoryRequirements
.alignment
=
3705 image
->planes
[plane
].alignment
;
3710 anv_debug_ignored_stype(ext
->sType
);
3715 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3716 switch (ext
->sType
) {
3717 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
3718 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
3719 if (image
->needs_set_tiling
|| image
->external_format
) {
3720 /* If we need to set the tiling for external consumers, we need a
3721 * dedicated allocation.
3723 * See also anv_AllocateMemory.
3725 requirements
->prefersDedicatedAllocation
= true;
3726 requirements
->requiresDedicatedAllocation
= true;
3728 requirements
->prefersDedicatedAllocation
= false;
3729 requirements
->requiresDedicatedAllocation
= false;
3735 anv_debug_ignored_stype(ext
->sType
);
3741 void anv_GetImageSparseMemoryRequirements(
3744 uint32_t* pSparseMemoryRequirementCount
,
3745 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
3747 *pSparseMemoryRequirementCount
= 0;
3750 void anv_GetImageSparseMemoryRequirements2(
3752 const VkImageSparseMemoryRequirementsInfo2
* pInfo
,
3753 uint32_t* pSparseMemoryRequirementCount
,
3754 VkSparseImageMemoryRequirements2
* pSparseMemoryRequirements
)
3756 *pSparseMemoryRequirementCount
= 0;
3759 void anv_GetDeviceMemoryCommitment(
3761 VkDeviceMemory memory
,
3762 VkDeviceSize
* pCommittedMemoryInBytes
)
3764 *pCommittedMemoryInBytes
= 0;
3768 anv_bind_buffer_memory(const VkBindBufferMemoryInfo
*pBindInfo
)
3770 ANV_FROM_HANDLE(anv_device_memory
, mem
, pBindInfo
->memory
);
3771 ANV_FROM_HANDLE(anv_buffer
, buffer
, pBindInfo
->buffer
);
3773 assert(pBindInfo
->sType
== VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
);
3776 buffer
->address
= (struct anv_address
) {
3778 .offset
= pBindInfo
->memoryOffset
,
3781 buffer
->address
= ANV_NULL_ADDRESS
;
3785 VkResult
anv_BindBufferMemory(
3788 VkDeviceMemory memory
,
3789 VkDeviceSize memoryOffset
)
3791 anv_bind_buffer_memory(
3792 &(VkBindBufferMemoryInfo
) {
3793 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
,
3796 .memoryOffset
= memoryOffset
,
3802 VkResult
anv_BindBufferMemory2(
3804 uint32_t bindInfoCount
,
3805 const VkBindBufferMemoryInfo
* pBindInfos
)
3807 for (uint32_t i
= 0; i
< bindInfoCount
; i
++)
3808 anv_bind_buffer_memory(&pBindInfos
[i
]);
3813 VkResult
anv_QueueBindSparse(
3815 uint32_t bindInfoCount
,
3816 const VkBindSparseInfo
* pBindInfo
,
3819 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
3820 if (anv_device_is_lost(queue
->device
))
3821 return VK_ERROR_DEVICE_LOST
;
3823 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
3828 VkResult
anv_CreateEvent(
3830 const VkEventCreateInfo
* pCreateInfo
,
3831 const VkAllocationCallbacks
* pAllocator
,
3834 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3835 struct anv_state state
;
3836 struct anv_event
*event
;
3838 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_EVENT_CREATE_INFO
);
3840 state
= anv_state_pool_alloc(&device
->dynamic_state_pool
,
3843 event
->state
= state
;
3844 event
->semaphore
= VK_EVENT_RESET
;
3846 if (!device
->info
.has_llc
) {
3847 /* Make sure the writes we're flushing have landed. */
3848 __builtin_ia32_mfence();
3849 __builtin_ia32_clflush(event
);
3852 *pEvent
= anv_event_to_handle(event
);
3857 void anv_DestroyEvent(
3860 const VkAllocationCallbacks
* pAllocator
)
3862 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3863 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3868 anv_state_pool_free(&device
->dynamic_state_pool
, event
->state
);
3871 VkResult
anv_GetEventStatus(
3875 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3876 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3878 if (anv_device_is_lost(device
))
3879 return VK_ERROR_DEVICE_LOST
;
3881 if (!device
->info
.has_llc
) {
3882 /* Invalidate read cache before reading event written by GPU. */
3883 __builtin_ia32_clflush(event
);
3884 __builtin_ia32_mfence();
3888 return event
->semaphore
;
3891 VkResult
anv_SetEvent(
3895 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3896 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3898 event
->semaphore
= VK_EVENT_SET
;
3900 if (!device
->info
.has_llc
) {
3901 /* Make sure the writes we're flushing have landed. */
3902 __builtin_ia32_mfence();
3903 __builtin_ia32_clflush(event
);
3909 VkResult
anv_ResetEvent(
3913 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3914 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3916 event
->semaphore
= VK_EVENT_RESET
;
3918 if (!device
->info
.has_llc
) {
3919 /* Make sure the writes we're flushing have landed. */
3920 __builtin_ia32_mfence();
3921 __builtin_ia32_clflush(event
);
3929 VkResult
anv_CreateBuffer(
3931 const VkBufferCreateInfo
* pCreateInfo
,
3932 const VkAllocationCallbacks
* pAllocator
,
3935 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3936 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3937 struct anv_buffer
*buffer
;
3939 /* Don't allow creating buffers bigger than our address space. The real
3940 * issue here is that we may align up the buffer size and we don't want
3941 * doing so to cause roll-over. However, no one has any business
3942 * allocating a buffer larger than our GTT size.
3944 if (pCreateInfo
->size
> pdevice
->gtt_size
)
3945 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
3947 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
3949 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
3950 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3952 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3954 buffer
->size
= pCreateInfo
->size
;
3955 buffer
->usage
= pCreateInfo
->usage
;
3956 buffer
->address
= ANV_NULL_ADDRESS
;
3958 *pBuffer
= anv_buffer_to_handle(buffer
);
3963 void anv_DestroyBuffer(
3966 const VkAllocationCallbacks
* pAllocator
)
3968 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3969 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
3974 vk_free2(&device
->alloc
, pAllocator
, buffer
);
3977 VkDeviceAddress
anv_GetBufferDeviceAddressKHR(
3979 const VkBufferDeviceAddressInfoKHR
* pInfo
)
3981 ANV_FROM_HANDLE(anv_buffer
, buffer
, pInfo
->buffer
);
3983 assert(!anv_address_is_null(buffer
->address
));
3984 assert(buffer
->address
.bo
->flags
& EXEC_OBJECT_PINNED
);
3986 return anv_address_physical(buffer
->address
);
3989 uint64_t anv_GetBufferOpaqueCaptureAddressKHR(
3991 const VkBufferDeviceAddressInfoKHR
* pInfo
)
3996 uint64_t anv_GetDeviceMemoryOpaqueCaptureAddressKHR(
3998 const VkDeviceMemoryOpaqueCaptureAddressInfoKHR
* pInfo
)
4000 ANV_FROM_HANDLE(anv_device_memory
, memory
, pInfo
->memory
);
4002 assert(memory
->bo
->flags
& EXEC_OBJECT_PINNED
);
4003 assert(memory
->bo
->has_client_visible_address
);
4005 return gen_48b_address(memory
->bo
->offset
);
4009 anv_fill_buffer_surface_state(struct anv_device
*device
, struct anv_state state
,
4010 enum isl_format format
,
4011 struct anv_address address
,
4012 uint32_t range
, uint32_t stride
)
4014 isl_buffer_fill_state(&device
->isl_dev
, state
.map
,
4015 .address
= anv_address_physical(address
),
4016 .mocs
= device
->isl_dev
.mocs
.internal
,
4019 .swizzle
= ISL_SWIZZLE_IDENTITY
,
4020 .stride_B
= stride
);
4023 void anv_DestroySampler(
4026 const VkAllocationCallbacks
* pAllocator
)
4028 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4029 ANV_FROM_HANDLE(anv_sampler
, sampler
, _sampler
);
4034 if (sampler
->bindless_state
.map
) {
4035 anv_state_pool_free(&device
->dynamic_state_pool
,
4036 sampler
->bindless_state
);
4039 vk_free2(&device
->alloc
, pAllocator
, sampler
);
4042 VkResult
anv_CreateFramebuffer(
4044 const VkFramebufferCreateInfo
* pCreateInfo
,
4045 const VkAllocationCallbacks
* pAllocator
,
4046 VkFramebuffer
* pFramebuffer
)
4048 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4049 struct anv_framebuffer
*framebuffer
;
4051 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
4053 size_t size
= sizeof(*framebuffer
);
4055 /* VK_KHR_imageless_framebuffer extension says:
4057 * If flags includes VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR,
4058 * parameter pAttachments is ignored.
4060 if (!(pCreateInfo
->flags
& VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR
)) {
4061 size
+= sizeof(struct anv_image_view
*) * pCreateInfo
->attachmentCount
;
4062 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
4063 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
4064 if (framebuffer
== NULL
)
4065 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
4067 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
4068 ANV_FROM_HANDLE(anv_image_view
, iview
, pCreateInfo
->pAttachments
[i
]);
4069 framebuffer
->attachments
[i
] = iview
;
4071 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
4073 assert(device
->enabled_extensions
.KHR_imageless_framebuffer
);
4074 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
4075 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
4076 if (framebuffer
== NULL
)
4077 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
4079 framebuffer
->attachment_count
= 0;
4082 framebuffer
->width
= pCreateInfo
->width
;
4083 framebuffer
->height
= pCreateInfo
->height
;
4084 framebuffer
->layers
= pCreateInfo
->layers
;
4086 *pFramebuffer
= anv_framebuffer_to_handle(framebuffer
);
4091 void anv_DestroyFramebuffer(
4094 const VkAllocationCallbacks
* pAllocator
)
4096 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4097 ANV_FROM_HANDLE(anv_framebuffer
, fb
, _fb
);
4102 vk_free2(&device
->alloc
, pAllocator
, fb
);
4105 static const VkTimeDomainEXT anv_time_domains
[] = {
4106 VK_TIME_DOMAIN_DEVICE_EXT
,
4107 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
,
4108 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
,
4111 VkResult
anv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
4112 VkPhysicalDevice physicalDevice
,
4113 uint32_t *pTimeDomainCount
,
4114 VkTimeDomainEXT
*pTimeDomains
)
4117 VK_OUTARRAY_MAKE(out
, pTimeDomains
, pTimeDomainCount
);
4119 for (d
= 0; d
< ARRAY_SIZE(anv_time_domains
); d
++) {
4120 vk_outarray_append(&out
, i
) {
4121 *i
= anv_time_domains
[d
];
4125 return vk_outarray_status(&out
);
4129 anv_clock_gettime(clockid_t clock_id
)
4131 struct timespec current
;
4134 ret
= clock_gettime(clock_id
, ¤t
);
4135 if (ret
< 0 && clock_id
== CLOCK_MONOTONIC_RAW
)
4136 ret
= clock_gettime(CLOCK_MONOTONIC
, ¤t
);
4140 return (uint64_t) current
.tv_sec
* 1000000000ULL + current
.tv_nsec
;
4143 #define TIMESTAMP 0x2358
4145 VkResult
anv_GetCalibratedTimestampsEXT(
4147 uint32_t timestampCount
,
4148 const VkCalibratedTimestampInfoEXT
*pTimestampInfos
,
4149 uint64_t *pTimestamps
,
4150 uint64_t *pMaxDeviation
)
4152 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4153 uint64_t timestamp_frequency
= device
->info
.timestamp_frequency
;
4156 uint64_t begin
, end
;
4157 uint64_t max_clock_period
= 0;
4159 begin
= anv_clock_gettime(CLOCK_MONOTONIC_RAW
);
4161 for (d
= 0; d
< timestampCount
; d
++) {
4162 switch (pTimestampInfos
[d
].timeDomain
) {
4163 case VK_TIME_DOMAIN_DEVICE_EXT
:
4164 ret
= anv_gem_reg_read(device
, TIMESTAMP
| 1,
4168 return anv_device_set_lost(device
, "Failed to read the TIMESTAMP "
4171 uint64_t device_period
= DIV_ROUND_UP(1000000000, timestamp_frequency
);
4172 max_clock_period
= MAX2(max_clock_period
, device_period
);
4174 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
:
4175 pTimestamps
[d
] = anv_clock_gettime(CLOCK_MONOTONIC
);
4176 max_clock_period
= MAX2(max_clock_period
, 1);
4179 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
:
4180 pTimestamps
[d
] = begin
;
4188 end
= anv_clock_gettime(CLOCK_MONOTONIC_RAW
);
4191 * The maximum deviation is the sum of the interval over which we
4192 * perform the sampling and the maximum period of any sampled
4193 * clock. That's because the maximum skew between any two sampled
4194 * clock edges is when the sampled clock with the largest period is
4195 * sampled at the end of that period but right at the beginning of the
4196 * sampling interval and some other clock is sampled right at the
4197 * begining of its sampling period and right at the end of the
4198 * sampling interval. Let's assume the GPU has the longest clock
4199 * period and that the application is sampling GPU and monotonic:
4202 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
4203 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4207 * GPU -----_____-----_____-----_____-----_____
4210 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
4211 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4213 * Interval <----------------->
4214 * Deviation <-------------------------->
4218 * m = read(monotonic) 2
4221 * We round the sample interval up by one tick to cover sampling error
4222 * in the interval clock
4225 uint64_t sample_interval
= end
- begin
+ 1;
4227 *pMaxDeviation
= sample_interval
+ max_clock_period
;
4232 /* vk_icd.h does not declare this function, so we declare it here to
4233 * suppress Wmissing-prototypes.
4235 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4236 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
);
4238 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4239 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
)
4241 /* For the full details on loader interface versioning, see
4242 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
4243 * What follows is a condensed summary, to help you navigate the large and
4244 * confusing official doc.
4246 * - Loader interface v0 is incompatible with later versions. We don't
4249 * - In loader interface v1:
4250 * - The first ICD entrypoint called by the loader is
4251 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
4253 * - The ICD must statically expose no other Vulkan symbol unless it is
4254 * linked with -Bsymbolic.
4255 * - Each dispatchable Vulkan handle created by the ICD must be
4256 * a pointer to a struct whose first member is VK_LOADER_DATA. The
4257 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
4258 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
4259 * vkDestroySurfaceKHR(). The ICD must be capable of working with
4260 * such loader-managed surfaces.
4262 * - Loader interface v2 differs from v1 in:
4263 * - The first ICD entrypoint called by the loader is
4264 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
4265 * statically expose this entrypoint.
4267 * - Loader interface v3 differs from v2 in:
4268 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
4269 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
4270 * because the loader no longer does so.
4272 * - Loader interface v4 differs from v3 in:
4273 * - The ICD must implement vk_icdGetPhysicalDeviceProcAddr().
4275 *pSupportedVersion
= MIN2(*pSupportedVersion
, 4u);