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_fourcc.h>
34 #include "anv_private.h"
35 #include "util/strtod.h"
36 #include "util/debug.h"
37 #include "util/build_id.h"
38 #include "util/mesa-sha1.h"
40 #include "common/gen_defines.h"
42 #include "genxml/gen7_pack.h"
45 compiler_debug_log(void *data
, const char *fmt
, ...)
49 compiler_perf_log(void *data
, const char *fmt
, ...)
54 if (unlikely(INTEL_DEBUG
& DEBUG_PERF
))
55 intel_logd_v(fmt
, args
);
61 anv_compute_heap_size(int fd
, uint64_t *heap_size
)
64 if (anv_gem_get_context_param(fd
, 0, I915_CONTEXT_PARAM_GTT_SIZE
,
66 /* If, for whatever reason, we can't actually get the GTT size from the
67 * kernel (too old?) fall back to the aperture size.
69 anv_perf_warn(NULL
, NULL
,
70 "Failed to get I915_CONTEXT_PARAM_GTT_SIZE: %m");
72 if (anv_gem_get_aperture(fd
, >t_size
) == -1) {
73 return vk_errorf(NULL
, NULL
, VK_ERROR_INITIALIZATION_FAILED
,
74 "failed to get aperture size: %m");
78 /* Query the total ram from the system */
82 uint64_t total_ram
= (uint64_t)info
.totalram
* (uint64_t)info
.mem_unit
;
84 /* We don't want to burn too much ram with the GPU. If the user has 4GiB
85 * or less, we use at most half. If they have more than 4GiB, we use 3/4.
87 uint64_t available_ram
;
88 if (total_ram
<= 4ull * 1024ull * 1024ull * 1024ull)
89 available_ram
= total_ram
/ 2;
91 available_ram
= total_ram
* 3 / 4;
93 /* We also want to leave some padding for things we allocate in the driver,
94 * so don't go over 3/4 of the GTT either.
96 uint64_t available_gtt
= gtt_size
* 3 / 4;
98 *heap_size
= MIN2(available_ram
, available_gtt
);
104 anv_physical_device_init_heaps(struct anv_physical_device
*device
, int fd
)
106 /* The kernel query only tells us whether or not the kernel supports the
107 * EXEC_OBJECT_SUPPORTS_48B_ADDRESS flag and not whether or not the
108 * hardware has actual 48bit address support.
110 device
->supports_48bit_addresses
=
111 (device
->info
.gen
>= 8) && anv_gem_supports_48b_addresses(fd
);
114 VkResult result
= anv_compute_heap_size(fd
, &heap_size
);
115 if (result
!= VK_SUCCESS
)
118 if (heap_size
> (2ull << 30) && !device
->supports_48bit_addresses
) {
119 /* When running with an overridden PCI ID, we may get a GTT size from
120 * the kernel that is greater than 2 GiB but the execbuf check for 48bit
121 * address support can still fail. Just clamp the address space size to
122 * 2 GiB if we don't have 48-bit support.
124 intel_logw("%s:%d: The kernel reported a GTT size larger than 2 GiB but "
125 "not support for 48-bit addresses",
127 heap_size
= 2ull << 30;
130 if (heap_size
<= 3ull * (1ull << 30)) {
131 /* In this case, everything fits nicely into the 32-bit address space,
132 * so there's no need for supporting 48bit addresses on client-allocated
135 device
->memory
.heap_count
= 1;
136 device
->memory
.heaps
[0] = (struct anv_memory_heap
) {
138 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
139 .supports_48bit_addresses
= false,
142 /* Not everything will fit nicely into a 32-bit address space. In this
143 * case we need a 64-bit heap. Advertise a small 32-bit heap and a
144 * larger 48-bit heap. If we're in this case, then we have a total heap
145 * size larger than 3GiB which most likely means they have 8 GiB of
146 * video memory and so carving off 1 GiB for the 32-bit heap should be
149 const uint64_t heap_size_32bit
= 1ull << 30;
150 const uint64_t heap_size_48bit
= heap_size
- heap_size_32bit
;
152 assert(device
->supports_48bit_addresses
);
154 device
->memory
.heap_count
= 2;
155 device
->memory
.heaps
[0] = (struct anv_memory_heap
) {
156 .size
= heap_size_48bit
,
157 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
158 .supports_48bit_addresses
= true,
160 device
->memory
.heaps
[1] = (struct anv_memory_heap
) {
161 .size
= heap_size_32bit
,
162 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
163 .supports_48bit_addresses
= false,
167 uint32_t type_count
= 0;
168 for (uint32_t heap
= 0; heap
< device
->memory
.heap_count
; heap
++) {
169 uint32_t valid_buffer_usage
= ~0;
171 /* There appears to be a hardware issue in the VF cache where it only
172 * considers the bottom 32 bits of memory addresses. If you happen to
173 * have two vertex buffers which get placed exactly 4 GiB apart and use
174 * them in back-to-back draw calls, you can get collisions. In order to
175 * solve this problem, we require vertex and index buffers be bound to
176 * memory allocated out of the 32-bit heap.
178 if (device
->memory
.heaps
[heap
].supports_48bit_addresses
) {
179 valid_buffer_usage
&= ~(VK_BUFFER_USAGE_INDEX_BUFFER_BIT
|
180 VK_BUFFER_USAGE_VERTEX_BUFFER_BIT
);
183 if (device
->info
.has_llc
) {
184 /* Big core GPUs share LLC with the CPU and thus one memory type can be
185 * both cached and coherent at the same time.
187 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
188 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
189 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
190 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
191 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
193 .valid_buffer_usage
= valid_buffer_usage
,
196 /* The spec requires that we expose a host-visible, coherent memory
197 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
198 * to give the application a choice between cached, but not coherent and
199 * coherent but uncached (WC though).
201 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
202 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
203 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
204 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
206 .valid_buffer_usage
= valid_buffer_usage
,
208 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
209 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
210 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
211 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
213 .valid_buffer_usage
= valid_buffer_usage
,
217 device
->memory
.type_count
= type_count
;
223 anv_physical_device_init_uuids(struct anv_physical_device
*device
)
225 const struct build_id_note
*note
=
226 build_id_find_nhdr_for_addr(anv_physical_device_init_uuids
);
228 return vk_errorf(device
->instance
, device
,
229 VK_ERROR_INITIALIZATION_FAILED
,
230 "Failed to find build-id");
233 unsigned build_id_len
= build_id_length(note
);
234 if (build_id_len
< 20) {
235 return vk_errorf(device
->instance
, device
,
236 VK_ERROR_INITIALIZATION_FAILED
,
237 "build-id too short. It needs to be a SHA");
240 struct mesa_sha1 sha1_ctx
;
242 STATIC_ASSERT(VK_UUID_SIZE
<= sizeof(sha1
));
244 /* The pipeline cache UUID is used for determining when a pipeline cache is
245 * invalid. It needs both a driver build and the PCI ID of the device.
247 _mesa_sha1_init(&sha1_ctx
);
248 _mesa_sha1_update(&sha1_ctx
, build_id_data(note
), build_id_len
);
249 _mesa_sha1_update(&sha1_ctx
, &device
->chipset_id
,
250 sizeof(device
->chipset_id
));
251 _mesa_sha1_final(&sha1_ctx
, sha1
);
252 memcpy(device
->pipeline_cache_uuid
, sha1
, VK_UUID_SIZE
);
254 /* The driver UUID is used for determining sharability of images and memory
255 * between two Vulkan instances in separate processes. People who want to
256 * share memory need to also check the device UUID (below) so all this
257 * needs to be is the build-id.
259 memcpy(device
->driver_uuid
, build_id_data(note
), VK_UUID_SIZE
);
261 /* The device UUID uniquely identifies the given device within the machine.
262 * Since we never have more than one device, this doesn't need to be a real
263 * UUID. However, on the off-chance that someone tries to use this to
264 * cache pre-tiled images or something of the like, we use the PCI ID and
265 * some bits of ISL info to ensure that this is safe.
267 _mesa_sha1_init(&sha1_ctx
);
268 _mesa_sha1_update(&sha1_ctx
, &device
->chipset_id
,
269 sizeof(device
->chipset_id
));
270 _mesa_sha1_update(&sha1_ctx
, &device
->isl_dev
.has_bit6_swizzling
,
271 sizeof(device
->isl_dev
.has_bit6_swizzling
));
272 _mesa_sha1_final(&sha1_ctx
, sha1
);
273 memcpy(device
->device_uuid
, sha1
, VK_UUID_SIZE
);
279 anv_physical_device_init(struct anv_physical_device
*device
,
280 struct anv_instance
*instance
,
286 brw_process_intel_debug_variable();
288 fd
= open(path
, O_RDWR
| O_CLOEXEC
);
290 return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
292 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
293 device
->instance
= instance
;
295 assert(strlen(path
) < ARRAY_SIZE(device
->path
));
296 strncpy(device
->path
, path
, ARRAY_SIZE(device
->path
));
298 device
->no_hw
= getenv("INTEL_NO_HW") != NULL
;
300 const int pci_id_override
= gen_get_pci_device_id_override();
301 if (pci_id_override
< 0) {
302 device
->chipset_id
= anv_gem_get_param(fd
, I915_PARAM_CHIPSET_ID
);
303 if (!device
->chipset_id
) {
304 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
308 device
->chipset_id
= pci_id_override
;
309 device
->no_hw
= true;
312 device
->name
= gen_get_device_name(device
->chipset_id
);
313 if (!gen_get_device_info(device
->chipset_id
, &device
->info
)) {
314 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
318 if (device
->info
.is_haswell
) {
319 intel_logw("Haswell Vulkan support is incomplete");
320 } else if (device
->info
.gen
== 7 && !device
->info
.is_baytrail
) {
321 intel_logw("Ivy Bridge Vulkan support is incomplete");
322 } else if (device
->info
.gen
== 7 && device
->info
.is_baytrail
) {
323 intel_logw("Bay Trail Vulkan support is incomplete");
324 } else if (device
->info
.gen
>= 8 && device
->info
.gen
<= 10) {
325 /* Gen8-10 fully supported */
327 result
= vk_errorf(device
->instance
, device
,
328 VK_ERROR_INCOMPATIBLE_DRIVER
,
329 "Vulkan not yet supported on %s", device
->name
);
333 device
->cmd_parser_version
= -1;
334 if (device
->info
.gen
== 7) {
335 device
->cmd_parser_version
=
336 anv_gem_get_param(fd
, I915_PARAM_CMD_PARSER_VERSION
);
337 if (device
->cmd_parser_version
== -1) {
338 result
= vk_errorf(device
->instance
, device
,
339 VK_ERROR_INITIALIZATION_FAILED
,
340 "failed to get command parser version");
345 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_WAIT_TIMEOUT
)) {
346 result
= vk_errorf(device
->instance
, device
,
347 VK_ERROR_INITIALIZATION_FAILED
,
348 "kernel missing gem wait");
352 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_EXECBUF2
)) {
353 result
= vk_errorf(device
->instance
, device
,
354 VK_ERROR_INITIALIZATION_FAILED
,
355 "kernel missing execbuf2");
359 if (!device
->info
.has_llc
&&
360 anv_gem_get_param(fd
, I915_PARAM_MMAP_VERSION
) < 1) {
361 result
= vk_errorf(device
->instance
, device
,
362 VK_ERROR_INITIALIZATION_FAILED
,
363 "kernel missing wc mmap");
367 result
= anv_physical_device_init_heaps(device
, fd
);
368 if (result
!= VK_SUCCESS
)
371 device
->has_exec_async
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_ASYNC
);
372 device
->has_exec_capture
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_CAPTURE
);
373 device
->has_exec_fence
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE
);
374 device
->has_syncobj
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE_ARRAY
);
375 device
->has_syncobj_wait
= device
->has_syncobj
&&
376 anv_gem_supports_syncobj_wait(fd
);
377 device
->has_context_priority
= anv_gem_has_context_priority(fd
);
379 bool swizzled
= anv_gem_get_bit6_swizzle(fd
, I915_TILING_X
);
381 /* Starting with Gen10, the timestamp frequency of the command streamer may
382 * vary from one part to another. We can query the value from the kernel.
384 if (device
->info
.gen
>= 10) {
385 int timestamp_frequency
=
386 anv_gem_get_param(fd
, I915_PARAM_CS_TIMESTAMP_FREQUENCY
);
388 if (timestamp_frequency
< 0)
389 intel_logw("Kernel 4.16-rc1+ required to properly query CS timestamp frequency");
391 device
->info
.timestamp_frequency
= timestamp_frequency
;
394 /* GENs prior to 8 do not support EU/Subslice info */
395 if (device
->info
.gen
>= 8) {
396 device
->subslice_total
= anv_gem_get_param(fd
, I915_PARAM_SUBSLICE_TOTAL
);
397 device
->eu_total
= anv_gem_get_param(fd
, I915_PARAM_EU_TOTAL
);
399 /* Without this information, we cannot get the right Braswell
400 * brandstrings, and we have to use conservative numbers for GPGPU on
401 * many platforms, but otherwise, things will just work.
403 if (device
->subslice_total
< 1 || device
->eu_total
< 1) {
404 intel_logw("Kernel 4.1 required to properly query GPU properties");
406 } else if (device
->info
.gen
== 7) {
407 device
->subslice_total
= 1 << (device
->info
.gt
- 1);
410 if (device
->info
.is_cherryview
&&
411 device
->subslice_total
> 0 && device
->eu_total
> 0) {
412 /* Logical CS threads = EUs per subslice * num threads per EU */
413 uint32_t max_cs_threads
=
414 device
->eu_total
/ device
->subslice_total
* device
->info
.num_thread_per_eu
;
416 /* Fuse configurations may give more threads than expected, never less. */
417 if (max_cs_threads
> device
->info
.max_cs_threads
)
418 device
->info
.max_cs_threads
= max_cs_threads
;
421 device
->compiler
= brw_compiler_create(NULL
, &device
->info
);
422 if (device
->compiler
== NULL
) {
423 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
426 device
->compiler
->shader_debug_log
= compiler_debug_log
;
427 device
->compiler
->shader_perf_log
= compiler_perf_log
;
428 device
->compiler
->supports_pull_constants
= false;
429 device
->compiler
->constant_buffer_0_is_relative
= true;
431 isl_device_init(&device
->isl_dev
, &device
->info
, swizzled
);
433 result
= anv_physical_device_init_uuids(device
);
434 if (result
!= VK_SUCCESS
)
437 result
= anv_init_wsi(device
);
438 if (result
!= VK_SUCCESS
) {
439 ralloc_free(device
->compiler
);
443 anv_physical_device_get_supported_extensions(device
,
444 &device
->supported_extensions
);
446 device
->local_fd
= fd
;
455 anv_physical_device_finish(struct anv_physical_device
*device
)
457 anv_finish_wsi(device
);
458 ralloc_free(device
->compiler
);
459 close(device
->local_fd
);
463 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
464 VkSystemAllocationScope allocationScope
)
470 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
471 size_t align
, VkSystemAllocationScope allocationScope
)
473 return realloc(pOriginal
, size
);
477 default_free_func(void *pUserData
, void *pMemory
)
482 static const VkAllocationCallbacks default_alloc
= {
484 .pfnAllocation
= default_alloc_func
,
485 .pfnReallocation
= default_realloc_func
,
486 .pfnFree
= default_free_func
,
489 VkResult
anv_EnumerateInstanceExtensionProperties(
490 const char* pLayerName
,
491 uint32_t* pPropertyCount
,
492 VkExtensionProperties
* pProperties
)
494 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
496 for (int i
= 0; i
< ANV_INSTANCE_EXTENSION_COUNT
; i
++) {
497 if (anv_instance_extensions_supported
.extensions
[i
]) {
498 vk_outarray_append(&out
, prop
) {
499 *prop
= anv_instance_extensions
[i
];
504 return vk_outarray_status(&out
);
507 VkResult
anv_CreateInstance(
508 const VkInstanceCreateInfo
* pCreateInfo
,
509 const VkAllocationCallbacks
* pAllocator
,
510 VkInstance
* pInstance
)
512 struct anv_instance
*instance
;
515 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
517 /* Check if user passed a debug report callback to be used during
518 * Create/Destroy of instance.
520 const VkDebugReportCallbackCreateInfoEXT
*ctor_cb
=
521 vk_find_struct_const(pCreateInfo
->pNext
,
522 DEBUG_REPORT_CALLBACK_CREATE_INFO_EXT
);
524 struct anv_instance_extension_table enabled_extensions
= {};
525 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
527 for (idx
= 0; idx
< ANV_INSTANCE_EXTENSION_COUNT
; idx
++) {
528 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
529 anv_instance_extensions
[idx
].extensionName
) == 0)
533 if (idx
>= ANV_INSTANCE_EXTENSION_COUNT
)
534 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
536 if (!anv_instance_extensions_supported
.extensions
[idx
])
537 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
539 enabled_extensions
.extensions
[idx
] = true;
542 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
543 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
545 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
547 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
550 instance
->alloc
= *pAllocator
;
552 instance
->alloc
= default_alloc
;
554 if (pCreateInfo
->pApplicationInfo
&&
555 pCreateInfo
->pApplicationInfo
->apiVersion
!= 0) {
556 instance
->apiVersion
= pCreateInfo
->pApplicationInfo
->apiVersion
;
558 anv_EnumerateInstanceVersion(&instance
->apiVersion
);
561 instance
->enabled_extensions
= enabled_extensions
;
563 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->dispatch
.entrypoints
); i
++) {
564 /* Vulkan requires that entrypoints for extensions which have not been
565 * enabled must not be advertised.
567 if (!anv_entrypoint_is_enabled(i
, instance
->apiVersion
,
568 &instance
->enabled_extensions
, NULL
)) {
569 instance
->dispatch
.entrypoints
[i
] = NULL
;
570 } else if (anv_dispatch_table
.entrypoints
[i
] != NULL
) {
571 instance
->dispatch
.entrypoints
[i
] = anv_dispatch_table
.entrypoints
[i
];
573 instance
->dispatch
.entrypoints
[i
] =
574 anv_tramp_dispatch_table
.entrypoints
[i
];
578 instance
->physicalDeviceCount
= -1;
580 result
= vk_debug_report_instance_init(&instance
->debug_report_callbacks
);
581 if (result
!= VK_SUCCESS
) {
582 vk_free2(&default_alloc
, pAllocator
, instance
);
583 return vk_error(result
);
588 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
590 *pInstance
= anv_instance_to_handle(instance
);
595 void anv_DestroyInstance(
596 VkInstance _instance
,
597 const VkAllocationCallbacks
* pAllocator
)
599 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
604 if (instance
->physicalDeviceCount
> 0) {
605 /* We support at most one physical device. */
606 assert(instance
->physicalDeviceCount
== 1);
607 anv_physical_device_finish(&instance
->physicalDevice
);
610 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
612 vk_debug_report_instance_destroy(&instance
->debug_report_callbacks
);
616 vk_free(&instance
->alloc
, instance
);
620 anv_enumerate_devices(struct anv_instance
*instance
)
622 /* TODO: Check for more devices ? */
623 drmDevicePtr devices
[8];
624 VkResult result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
627 instance
->physicalDeviceCount
= 0;
629 max_devices
= drmGetDevices2(0, devices
, ARRAY_SIZE(devices
));
631 return VK_ERROR_INCOMPATIBLE_DRIVER
;
633 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
634 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
635 devices
[i
]->bustype
== DRM_BUS_PCI
&&
636 devices
[i
]->deviceinfo
.pci
->vendor_id
== 0x8086) {
638 result
= anv_physical_device_init(&instance
->physicalDevice
,
640 devices
[i
]->nodes
[DRM_NODE_RENDER
]);
641 if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
645 drmFreeDevices(devices
, max_devices
);
647 if (result
== VK_SUCCESS
)
648 instance
->physicalDeviceCount
= 1;
654 anv_instance_ensure_physical_device(struct anv_instance
*instance
)
656 if (instance
->physicalDeviceCount
< 0) {
657 VkResult result
= anv_enumerate_devices(instance
);
658 if (result
!= VK_SUCCESS
&&
659 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
666 VkResult
anv_EnumeratePhysicalDevices(
667 VkInstance _instance
,
668 uint32_t* pPhysicalDeviceCount
,
669 VkPhysicalDevice
* pPhysicalDevices
)
671 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
672 VK_OUTARRAY_MAKE(out
, pPhysicalDevices
, pPhysicalDeviceCount
);
674 VkResult result
= anv_instance_ensure_physical_device(instance
);
675 if (result
!= VK_SUCCESS
)
678 if (instance
->physicalDeviceCount
== 0)
681 assert(instance
->physicalDeviceCount
== 1);
682 vk_outarray_append(&out
, i
) {
683 *i
= anv_physical_device_to_handle(&instance
->physicalDevice
);
686 return vk_outarray_status(&out
);
689 VkResult
anv_EnumeratePhysicalDeviceGroups(
690 VkInstance _instance
,
691 uint32_t* pPhysicalDeviceGroupCount
,
692 VkPhysicalDeviceGroupProperties
* pPhysicalDeviceGroupProperties
)
694 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
695 VK_OUTARRAY_MAKE(out
, pPhysicalDeviceGroupProperties
,
696 pPhysicalDeviceGroupCount
);
698 VkResult result
= anv_instance_ensure_physical_device(instance
);
699 if (result
!= VK_SUCCESS
)
702 if (instance
->physicalDeviceCount
== 0)
705 assert(instance
->physicalDeviceCount
== 1);
707 vk_outarray_append(&out
, p
) {
708 p
->physicalDeviceCount
= 1;
709 memset(p
->physicalDevices
, 0, sizeof(p
->physicalDevices
));
710 p
->physicalDevices
[0] =
711 anv_physical_device_to_handle(&instance
->physicalDevice
);
712 p
->subsetAllocation
= VK_FALSE
;
714 vk_foreach_struct(ext
, p
->pNext
)
715 anv_debug_ignored_stype(ext
->sType
);
718 return vk_outarray_status(&out
);
721 void anv_GetPhysicalDeviceFeatures(
722 VkPhysicalDevice physicalDevice
,
723 VkPhysicalDeviceFeatures
* pFeatures
)
725 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
727 *pFeatures
= (VkPhysicalDeviceFeatures
) {
728 .robustBufferAccess
= true,
729 .fullDrawIndexUint32
= true,
730 .imageCubeArray
= true,
731 .independentBlend
= true,
732 .geometryShader
= true,
733 .tessellationShader
= true,
734 .sampleRateShading
= true,
735 .dualSrcBlend
= true,
737 .multiDrawIndirect
= true,
738 .drawIndirectFirstInstance
= true,
740 .depthBiasClamp
= true,
741 .fillModeNonSolid
= true,
742 .depthBounds
= false,
746 .multiViewport
= true,
747 .samplerAnisotropy
= true,
748 .textureCompressionETC2
= pdevice
->info
.gen
>= 8 ||
749 pdevice
->info
.is_baytrail
,
750 .textureCompressionASTC_LDR
= pdevice
->info
.gen
>= 9, /* FINISHME CHV */
751 .textureCompressionBC
= true,
752 .occlusionQueryPrecise
= true,
753 .pipelineStatisticsQuery
= true,
754 .fragmentStoresAndAtomics
= true,
755 .shaderTessellationAndGeometryPointSize
= true,
756 .shaderImageGatherExtended
= true,
757 .shaderStorageImageExtendedFormats
= true,
758 .shaderStorageImageMultisample
= false,
759 .shaderStorageImageReadWithoutFormat
= false,
760 .shaderStorageImageWriteWithoutFormat
= true,
761 .shaderUniformBufferArrayDynamicIndexing
= true,
762 .shaderSampledImageArrayDynamicIndexing
= true,
763 .shaderStorageBufferArrayDynamicIndexing
= true,
764 .shaderStorageImageArrayDynamicIndexing
= true,
765 .shaderClipDistance
= true,
766 .shaderCullDistance
= true,
767 .shaderFloat64
= pdevice
->info
.gen
>= 8,
768 .shaderInt64
= pdevice
->info
.gen
>= 8,
769 .shaderInt16
= false,
770 .shaderResourceMinLod
= false,
771 .variableMultisampleRate
= false,
772 .inheritedQueries
= true,
775 /* We can't do image stores in vec4 shaders */
776 pFeatures
->vertexPipelineStoresAndAtomics
=
777 pdevice
->compiler
->scalar_stage
[MESA_SHADER_VERTEX
] &&
778 pdevice
->compiler
->scalar_stage
[MESA_SHADER_GEOMETRY
];
781 void anv_GetPhysicalDeviceFeatures2(
782 VkPhysicalDevice physicalDevice
,
783 VkPhysicalDeviceFeatures2
* pFeatures
)
785 anv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
787 vk_foreach_struct(ext
, pFeatures
->pNext
) {
788 switch (ext
->sType
) {
789 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES
: {
790 VkPhysicalDeviceProtectedMemoryFeatures
*features
= (void *)ext
;
791 features
->protectedMemory
= VK_FALSE
;
795 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES
: {
796 VkPhysicalDeviceMultiviewFeatures
*features
=
797 (VkPhysicalDeviceMultiviewFeatures
*)ext
;
798 features
->multiview
= true;
799 features
->multiviewGeometryShader
= true;
800 features
->multiviewTessellationShader
= true;
804 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTER_FEATURES
: {
805 VkPhysicalDeviceVariablePointerFeatures
*features
= (void *)ext
;
806 features
->variablePointersStorageBuffer
= true;
807 features
->variablePointers
= true;
811 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES
: {
812 VkPhysicalDeviceSamplerYcbcrConversionFeatures
*features
=
813 (VkPhysicalDeviceSamplerYcbcrConversionFeatures
*) ext
;
814 features
->samplerYcbcrConversion
= true;
818 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETER_FEATURES
: {
819 VkPhysicalDeviceShaderDrawParameterFeatures
*features
= (void *)ext
;
820 features
->shaderDrawParameters
= true;
824 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES_KHR
: {
825 VkPhysicalDevice16BitStorageFeaturesKHR
*features
=
826 (VkPhysicalDevice16BitStorageFeaturesKHR
*)ext
;
827 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
829 features
->storageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
830 features
->uniformAndStorageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
831 features
->storagePushConstant16
= pdevice
->info
.gen
>= 8;
832 features
->storageInputOutput16
= false;
837 anv_debug_ignored_stype(ext
->sType
);
843 void anv_GetPhysicalDeviceProperties(
844 VkPhysicalDevice physicalDevice
,
845 VkPhysicalDeviceProperties
* pProperties
)
847 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
848 const struct gen_device_info
*devinfo
= &pdevice
->info
;
850 /* See assertions made when programming the buffer surface state. */
851 const uint32_t max_raw_buffer_sz
= devinfo
->gen
>= 7 ?
852 (1ul << 30) : (1ul << 27);
854 const uint32_t max_samplers
= (devinfo
->gen
>= 8 || devinfo
->is_haswell
) ?
857 VkSampleCountFlags sample_counts
=
858 isl_device_get_sample_counts(&pdevice
->isl_dev
);
860 VkPhysicalDeviceLimits limits
= {
861 .maxImageDimension1D
= (1 << 14),
862 .maxImageDimension2D
= (1 << 14),
863 .maxImageDimension3D
= (1 << 11),
864 .maxImageDimensionCube
= (1 << 14),
865 .maxImageArrayLayers
= (1 << 11),
866 .maxTexelBufferElements
= 128 * 1024 * 1024,
867 .maxUniformBufferRange
= (1ul << 27),
868 .maxStorageBufferRange
= max_raw_buffer_sz
,
869 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
870 .maxMemoryAllocationCount
= UINT32_MAX
,
871 .maxSamplerAllocationCount
= 64 * 1024,
872 .bufferImageGranularity
= 64, /* A cache line */
873 .sparseAddressSpaceSize
= 0,
874 .maxBoundDescriptorSets
= MAX_SETS
,
875 .maxPerStageDescriptorSamplers
= max_samplers
,
876 .maxPerStageDescriptorUniformBuffers
= 64,
877 .maxPerStageDescriptorStorageBuffers
= 64,
878 .maxPerStageDescriptorSampledImages
= max_samplers
,
879 .maxPerStageDescriptorStorageImages
= 64,
880 .maxPerStageDescriptorInputAttachments
= 64,
881 .maxPerStageResources
= 250,
882 .maxDescriptorSetSamplers
= 6 * max_samplers
, /* number of stages * maxPerStageDescriptorSamplers */
883 .maxDescriptorSetUniformBuffers
= 6 * 64, /* number of stages * maxPerStageDescriptorUniformBuffers */
884 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
885 .maxDescriptorSetStorageBuffers
= 6 * 64, /* number of stages * maxPerStageDescriptorStorageBuffers */
886 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
887 .maxDescriptorSetSampledImages
= 6 * max_samplers
, /* number of stages * maxPerStageDescriptorSampledImages */
888 .maxDescriptorSetStorageImages
= 6 * 64, /* number of stages * maxPerStageDescriptorStorageImages */
889 .maxDescriptorSetInputAttachments
= 256,
890 .maxVertexInputAttributes
= MAX_VBS
,
891 .maxVertexInputBindings
= MAX_VBS
,
892 .maxVertexInputAttributeOffset
= 2047,
893 .maxVertexInputBindingStride
= 2048,
894 .maxVertexOutputComponents
= 128,
895 .maxTessellationGenerationLevel
= 64,
896 .maxTessellationPatchSize
= 32,
897 .maxTessellationControlPerVertexInputComponents
= 128,
898 .maxTessellationControlPerVertexOutputComponents
= 128,
899 .maxTessellationControlPerPatchOutputComponents
= 128,
900 .maxTessellationControlTotalOutputComponents
= 2048,
901 .maxTessellationEvaluationInputComponents
= 128,
902 .maxTessellationEvaluationOutputComponents
= 128,
903 .maxGeometryShaderInvocations
= 32,
904 .maxGeometryInputComponents
= 64,
905 .maxGeometryOutputComponents
= 128,
906 .maxGeometryOutputVertices
= 256,
907 .maxGeometryTotalOutputComponents
= 1024,
908 .maxFragmentInputComponents
= 128,
909 .maxFragmentOutputAttachments
= 8,
910 .maxFragmentDualSrcAttachments
= 1,
911 .maxFragmentCombinedOutputResources
= 8,
912 .maxComputeSharedMemorySize
= 32768,
913 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
914 .maxComputeWorkGroupInvocations
= 16 * devinfo
->max_cs_threads
,
915 .maxComputeWorkGroupSize
= {
916 16 * devinfo
->max_cs_threads
,
917 16 * devinfo
->max_cs_threads
,
918 16 * devinfo
->max_cs_threads
,
920 .subPixelPrecisionBits
= 4 /* FIXME */,
921 .subTexelPrecisionBits
= 4 /* FIXME */,
922 .mipmapPrecisionBits
= 4 /* FIXME */,
923 .maxDrawIndexedIndexValue
= UINT32_MAX
,
924 .maxDrawIndirectCount
= UINT32_MAX
,
925 .maxSamplerLodBias
= 16,
926 .maxSamplerAnisotropy
= 16,
927 .maxViewports
= MAX_VIEWPORTS
,
928 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
929 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
930 .viewportSubPixelBits
= 13, /* We take a float? */
931 .minMemoryMapAlignment
= 4096, /* A page */
932 .minTexelBufferOffsetAlignment
= 1,
933 /* We need 16 for UBO block reads to work and 32 for push UBOs */
934 .minUniformBufferOffsetAlignment
= 32,
935 .minStorageBufferOffsetAlignment
= 4,
936 .minTexelOffset
= -8,
938 .minTexelGatherOffset
= -32,
939 .maxTexelGatherOffset
= 31,
940 .minInterpolationOffset
= -0.5,
941 .maxInterpolationOffset
= 0.4375,
942 .subPixelInterpolationOffsetBits
= 4,
943 .maxFramebufferWidth
= (1 << 14),
944 .maxFramebufferHeight
= (1 << 14),
945 .maxFramebufferLayers
= (1 << 11),
946 .framebufferColorSampleCounts
= sample_counts
,
947 .framebufferDepthSampleCounts
= sample_counts
,
948 .framebufferStencilSampleCounts
= sample_counts
,
949 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
950 .maxColorAttachments
= MAX_RTS
,
951 .sampledImageColorSampleCounts
= sample_counts
,
952 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
953 .sampledImageDepthSampleCounts
= sample_counts
,
954 .sampledImageStencilSampleCounts
= sample_counts
,
955 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
956 .maxSampleMaskWords
= 1,
957 .timestampComputeAndGraphics
= false,
958 .timestampPeriod
= 1000000000.0 / devinfo
->timestamp_frequency
,
959 .maxClipDistances
= 8,
960 .maxCullDistances
= 8,
961 .maxCombinedClipAndCullDistances
= 8,
962 .discreteQueuePriorities
= 1,
963 .pointSizeRange
= { 0.125, 255.875 },
964 .lineWidthRange
= { 0.0, 7.9921875 },
965 .pointSizeGranularity
= (1.0 / 8.0),
966 .lineWidthGranularity
= (1.0 / 128.0),
967 .strictLines
= false, /* FINISHME */
968 .standardSampleLocations
= true,
969 .optimalBufferCopyOffsetAlignment
= 128,
970 .optimalBufferCopyRowPitchAlignment
= 128,
971 .nonCoherentAtomSize
= 64,
974 *pProperties
= (VkPhysicalDeviceProperties
) {
975 .apiVersion
= anv_physical_device_api_version(pdevice
),
976 .driverVersion
= vk_get_driver_version(),
978 .deviceID
= pdevice
->chipset_id
,
979 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
981 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
984 snprintf(pProperties
->deviceName
, sizeof(pProperties
->deviceName
),
985 "%s", pdevice
->name
);
986 memcpy(pProperties
->pipelineCacheUUID
,
987 pdevice
->pipeline_cache_uuid
, VK_UUID_SIZE
);
990 void anv_GetPhysicalDeviceProperties2(
991 VkPhysicalDevice physicalDevice
,
992 VkPhysicalDeviceProperties2
* pProperties
)
994 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
996 anv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
998 vk_foreach_struct(ext
, pProperties
->pNext
) {
999 switch (ext
->sType
) {
1000 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
1001 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
1002 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
1004 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
1008 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES
: {
1009 VkPhysicalDeviceIDProperties
*id_props
=
1010 (VkPhysicalDeviceIDProperties
*)ext
;
1011 memcpy(id_props
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
1012 memcpy(id_props
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
1013 /* The LUID is for Windows. */
1014 id_props
->deviceLUIDValid
= false;
1018 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES
: {
1019 VkPhysicalDeviceMaintenance3Properties
*props
=
1020 (VkPhysicalDeviceMaintenance3Properties
*)ext
;
1021 /* This value doesn't matter for us today as our per-stage
1022 * descriptors are the real limit.
1024 props
->maxPerSetDescriptors
= 1024;
1025 props
->maxMemoryAllocationSize
= MAX_MEMORY_ALLOCATION_SIZE
;
1029 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES
: {
1030 VkPhysicalDeviceMultiviewProperties
*properties
=
1031 (VkPhysicalDeviceMultiviewProperties
*)ext
;
1032 properties
->maxMultiviewViewCount
= 16;
1033 properties
->maxMultiviewInstanceIndex
= UINT32_MAX
/ 16;
1037 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES
: {
1038 VkPhysicalDevicePointClippingProperties
*properties
=
1039 (VkPhysicalDevicePointClippingProperties
*) ext
;
1040 properties
->pointClippingBehavior
= VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES
;
1041 anv_finishme("Implement pop-free point clipping");
1045 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES
: {
1046 VkPhysicalDeviceSubgroupProperties
*properties
= (void *)ext
;
1048 properties
->subgroupSize
= BRW_SUBGROUP_SIZE
;
1050 VkShaderStageFlags scalar_stages
= 0;
1051 for (unsigned stage
= 0; stage
< MESA_SHADER_STAGES
; stage
++) {
1052 if (pdevice
->compiler
->scalar_stage
[stage
])
1053 scalar_stages
|= mesa_to_vk_shader_stage(stage
);
1055 properties
->supportedStages
= scalar_stages
;
1057 properties
->supportedOperations
= VK_SUBGROUP_FEATURE_BASIC_BIT
|
1058 VK_SUBGROUP_FEATURE_VOTE_BIT
|
1059 VK_SUBGROUP_FEATURE_ARITHMETIC_BIT
|
1060 VK_SUBGROUP_FEATURE_BALLOT_BIT
|
1061 VK_SUBGROUP_FEATURE_SHUFFLE_BIT
|
1062 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT
|
1063 VK_SUBGROUP_FEATURE_CLUSTERED_BIT
|
1064 VK_SUBGROUP_FEATURE_QUAD_BIT
;
1065 properties
->quadOperationsInAllStages
= VK_TRUE
;
1070 anv_debug_ignored_stype(ext
->sType
);
1076 /* We support exactly one queue family. */
1077 static const VkQueueFamilyProperties
1078 anv_queue_family_properties
= {
1079 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
1080 VK_QUEUE_COMPUTE_BIT
|
1081 VK_QUEUE_TRANSFER_BIT
,
1083 .timestampValidBits
= 36, /* XXX: Real value here */
1084 .minImageTransferGranularity
= { 1, 1, 1 },
1087 void anv_GetPhysicalDeviceQueueFamilyProperties(
1088 VkPhysicalDevice physicalDevice
,
1090 VkQueueFamilyProperties
* pQueueFamilyProperties
)
1092 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pCount
);
1094 vk_outarray_append(&out
, p
) {
1095 *p
= anv_queue_family_properties
;
1099 void anv_GetPhysicalDeviceQueueFamilyProperties2(
1100 VkPhysicalDevice physicalDevice
,
1101 uint32_t* pQueueFamilyPropertyCount
,
1102 VkQueueFamilyProperties2
* pQueueFamilyProperties
)
1105 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pQueueFamilyPropertyCount
);
1107 vk_outarray_append(&out
, p
) {
1108 p
->queueFamilyProperties
= anv_queue_family_properties
;
1110 vk_foreach_struct(s
, p
->pNext
) {
1111 anv_debug_ignored_stype(s
->sType
);
1116 void anv_GetPhysicalDeviceMemoryProperties(
1117 VkPhysicalDevice physicalDevice
,
1118 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
1120 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
1122 pMemoryProperties
->memoryTypeCount
= physical_device
->memory
.type_count
;
1123 for (uint32_t i
= 0; i
< physical_device
->memory
.type_count
; i
++) {
1124 pMemoryProperties
->memoryTypes
[i
] = (VkMemoryType
) {
1125 .propertyFlags
= physical_device
->memory
.types
[i
].propertyFlags
,
1126 .heapIndex
= physical_device
->memory
.types
[i
].heapIndex
,
1130 pMemoryProperties
->memoryHeapCount
= physical_device
->memory
.heap_count
;
1131 for (uint32_t i
= 0; i
< physical_device
->memory
.heap_count
; i
++) {
1132 pMemoryProperties
->memoryHeaps
[i
] = (VkMemoryHeap
) {
1133 .size
= physical_device
->memory
.heaps
[i
].size
,
1134 .flags
= physical_device
->memory
.heaps
[i
].flags
,
1139 void anv_GetPhysicalDeviceMemoryProperties2(
1140 VkPhysicalDevice physicalDevice
,
1141 VkPhysicalDeviceMemoryProperties2
* pMemoryProperties
)
1143 anv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
1144 &pMemoryProperties
->memoryProperties
);
1146 vk_foreach_struct(ext
, pMemoryProperties
->pNext
) {
1147 switch (ext
->sType
) {
1149 anv_debug_ignored_stype(ext
->sType
);
1156 anv_GetDeviceGroupPeerMemoryFeatures(
1159 uint32_t localDeviceIndex
,
1160 uint32_t remoteDeviceIndex
,
1161 VkPeerMemoryFeatureFlags
* pPeerMemoryFeatures
)
1163 assert(localDeviceIndex
== 0 && remoteDeviceIndex
== 0);
1164 *pPeerMemoryFeatures
= VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT
|
1165 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT
|
1166 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT
|
1167 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT
;
1170 PFN_vkVoidFunction
anv_GetInstanceProcAddr(
1171 VkInstance _instance
,
1174 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1176 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
1177 * when we have to return valid function pointers, NULL, or it's left
1178 * undefined. See the table for exact details.
1183 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
1184 if (strcmp(pName, "vk" #entrypoint) == 0) \
1185 return (PFN_vkVoidFunction)anv_##entrypoint
1187 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties
);
1188 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties
);
1189 LOOKUP_ANV_ENTRYPOINT(CreateInstance
);
1191 #undef LOOKUP_ANV_ENTRYPOINT
1193 if (instance
== NULL
)
1196 int idx
= anv_get_entrypoint_index(pName
);
1200 return instance
->dispatch
.entrypoints
[idx
];
1203 /* With version 1+ of the loader interface the ICD should expose
1204 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
1207 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1208 VkInstance instance
,
1212 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1213 VkInstance instance
,
1216 return anv_GetInstanceProcAddr(instance
, pName
);
1219 PFN_vkVoidFunction
anv_GetDeviceProcAddr(
1223 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1225 if (!device
|| !pName
)
1228 int idx
= anv_get_entrypoint_index(pName
);
1232 return device
->dispatch
.entrypoints
[idx
];
1236 anv_CreateDebugReportCallbackEXT(VkInstance _instance
,
1237 const VkDebugReportCallbackCreateInfoEXT
* pCreateInfo
,
1238 const VkAllocationCallbacks
* pAllocator
,
1239 VkDebugReportCallbackEXT
* pCallback
)
1241 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1242 return vk_create_debug_report_callback(&instance
->debug_report_callbacks
,
1243 pCreateInfo
, pAllocator
, &instance
->alloc
,
1248 anv_DestroyDebugReportCallbackEXT(VkInstance _instance
,
1249 VkDebugReportCallbackEXT _callback
,
1250 const VkAllocationCallbacks
* pAllocator
)
1252 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1253 vk_destroy_debug_report_callback(&instance
->debug_report_callbacks
,
1254 _callback
, pAllocator
, &instance
->alloc
);
1258 anv_DebugReportMessageEXT(VkInstance _instance
,
1259 VkDebugReportFlagsEXT flags
,
1260 VkDebugReportObjectTypeEXT objectType
,
1263 int32_t messageCode
,
1264 const char* pLayerPrefix
,
1265 const char* pMessage
)
1267 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1268 vk_debug_report(&instance
->debug_report_callbacks
, flags
, objectType
,
1269 object
, location
, messageCode
, pLayerPrefix
, pMessage
);
1273 anv_queue_init(struct anv_device
*device
, struct anv_queue
*queue
)
1275 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1276 queue
->device
= device
;
1277 queue
->pool
= &device
->surface_state_pool
;
1282 anv_queue_finish(struct anv_queue
*queue
)
1286 static struct anv_state
1287 anv_state_pool_emit_data(struct anv_state_pool
*pool
, size_t size
, size_t align
, const void *p
)
1289 struct anv_state state
;
1291 state
= anv_state_pool_alloc(pool
, size
, align
);
1292 memcpy(state
.map
, p
, size
);
1294 anv_state_flush(pool
->block_pool
.device
, state
);
1299 struct gen8_border_color
{
1304 /* Pad out to 64 bytes */
1309 anv_device_init_border_colors(struct anv_device
*device
)
1311 static const struct gen8_border_color border_colors
[] = {
1312 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
1313 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
1314 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
1315 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
1316 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
1317 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
1320 device
->border_colors
= anv_state_pool_emit_data(&device
->dynamic_state_pool
,
1321 sizeof(border_colors
), 64,
1326 anv_device_init_trivial_batch(struct anv_device
*device
)
1328 anv_bo_init_new(&device
->trivial_batch_bo
, device
, 4096);
1330 if (device
->instance
->physicalDevice
.has_exec_async
)
1331 device
->trivial_batch_bo
.flags
|= EXEC_OBJECT_ASYNC
;
1333 void *map
= anv_gem_mmap(device
, device
->trivial_batch_bo
.gem_handle
,
1336 struct anv_batch batch
= {
1342 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
1343 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
1345 if (!device
->info
.has_llc
)
1346 gen_clflush_range(map
, batch
.next
- map
);
1348 anv_gem_munmap(map
, device
->trivial_batch_bo
.size
);
1351 VkResult
anv_EnumerateDeviceExtensionProperties(
1352 VkPhysicalDevice physicalDevice
,
1353 const char* pLayerName
,
1354 uint32_t* pPropertyCount
,
1355 VkExtensionProperties
* pProperties
)
1357 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
1358 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
1361 for (int i
= 0; i
< ANV_DEVICE_EXTENSION_COUNT
; i
++) {
1362 if (device
->supported_extensions
.extensions
[i
]) {
1363 vk_outarray_append(&out
, prop
) {
1364 *prop
= anv_device_extensions
[i
];
1369 return vk_outarray_status(&out
);
1373 anv_device_init_dispatch(struct anv_device
*device
)
1375 const struct anv_dispatch_table
*genX_table
;
1376 switch (device
->info
.gen
) {
1378 genX_table
= &gen10_dispatch_table
;
1381 genX_table
= &gen9_dispatch_table
;
1384 genX_table
= &gen8_dispatch_table
;
1387 if (device
->info
.is_haswell
)
1388 genX_table
= &gen75_dispatch_table
;
1390 genX_table
= &gen7_dispatch_table
;
1393 unreachable("unsupported gen\n");
1396 for (unsigned i
= 0; i
< ARRAY_SIZE(device
->dispatch
.entrypoints
); i
++) {
1397 /* Vulkan requires that entrypoints for extensions which have not been
1398 * enabled must not be advertised.
1400 if (!anv_entrypoint_is_enabled(i
, device
->instance
->apiVersion
,
1401 &device
->instance
->enabled_extensions
,
1402 &device
->enabled_extensions
)) {
1403 device
->dispatch
.entrypoints
[i
] = NULL
;
1404 } else if (genX_table
->entrypoints
[i
]) {
1405 device
->dispatch
.entrypoints
[i
] = genX_table
->entrypoints
[i
];
1407 device
->dispatch
.entrypoints
[i
] = anv_dispatch_table
.entrypoints
[i
];
1413 vk_priority_to_gen(int priority
)
1416 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT
:
1417 return GEN_CONTEXT_LOW_PRIORITY
;
1418 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
:
1419 return GEN_CONTEXT_MEDIUM_PRIORITY
;
1420 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT
:
1421 return GEN_CONTEXT_HIGH_PRIORITY
;
1422 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT
:
1423 return GEN_CONTEXT_REALTIME_PRIORITY
;
1425 unreachable("Invalid priority");
1429 VkResult
anv_CreateDevice(
1430 VkPhysicalDevice physicalDevice
,
1431 const VkDeviceCreateInfo
* pCreateInfo
,
1432 const VkAllocationCallbacks
* pAllocator
,
1435 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
1437 struct anv_device
*device
;
1439 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO
);
1441 struct anv_device_extension_table enabled_extensions
= { };
1442 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
1444 for (idx
= 0; idx
< ANV_DEVICE_EXTENSION_COUNT
; idx
++) {
1445 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
1446 anv_device_extensions
[idx
].extensionName
) == 0)
1450 if (idx
>= ANV_DEVICE_EXTENSION_COUNT
)
1451 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
1453 if (!physical_device
->supported_extensions
.extensions
[idx
])
1454 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
1456 enabled_extensions
.extensions
[idx
] = true;
1459 /* Check enabled features */
1460 if (pCreateInfo
->pEnabledFeatures
) {
1461 VkPhysicalDeviceFeatures supported_features
;
1462 anv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
1463 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
1464 VkBool32
*enabled_feature
= (VkBool32
*)pCreateInfo
->pEnabledFeatures
;
1465 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
1466 for (uint32_t i
= 0; i
< num_features
; i
++) {
1467 if (enabled_feature
[i
] && !supported_feature
[i
])
1468 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
1472 /* Check requested queues and fail if we are requested to create any
1473 * queues with flags we don't support.
1475 assert(pCreateInfo
->queueCreateInfoCount
> 0);
1476 for (uint32_t i
= 0; i
< pCreateInfo
->queueCreateInfoCount
; i
++) {
1477 if (pCreateInfo
->pQueueCreateInfos
[i
].flags
!= 0)
1478 return vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1481 /* Check if client specified queue priority. */
1482 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*queue_priority
=
1483 vk_find_struct_const(pCreateInfo
->pQueueCreateInfos
[0].pNext
,
1484 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT
);
1486 VkQueueGlobalPriorityEXT priority
=
1487 queue_priority
? queue_priority
->globalPriority
:
1488 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
;
1490 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
1492 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
1494 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1496 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1497 device
->instance
= physical_device
->instance
;
1498 device
->chipset_id
= physical_device
->chipset_id
;
1499 device
->no_hw
= physical_device
->no_hw
;
1500 device
->lost
= false;
1503 device
->alloc
= *pAllocator
;
1505 device
->alloc
= physical_device
->instance
->alloc
;
1507 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
1508 device
->fd
= open(physical_device
->path
, O_RDWR
| O_CLOEXEC
);
1509 if (device
->fd
== -1) {
1510 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1514 device
->context_id
= anv_gem_create_context(device
);
1515 if (device
->context_id
== -1) {
1516 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1520 /* As per spec, the driver implementation may deny requests to acquire
1521 * a priority above the default priority (MEDIUM) if the caller does not
1522 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
1525 if (physical_device
->has_context_priority
) {
1526 int err
= anv_gem_set_context_param(device
->fd
, device
->context_id
,
1527 I915_CONTEXT_PARAM_PRIORITY
,
1528 vk_priority_to_gen(priority
));
1529 if (err
!= 0 && priority
> VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
) {
1530 result
= vk_error(VK_ERROR_NOT_PERMITTED_EXT
);
1535 device
->info
= physical_device
->info
;
1536 device
->isl_dev
= physical_device
->isl_dev
;
1538 /* On Broadwell and later, we can use batch chaining to more efficiently
1539 * implement growing command buffers. Prior to Haswell, the kernel
1540 * command parser gets in the way and we have to fall back to growing
1543 device
->can_chain_batches
= device
->info
.gen
>= 8;
1545 device
->robust_buffer_access
= pCreateInfo
->pEnabledFeatures
&&
1546 pCreateInfo
->pEnabledFeatures
->robustBufferAccess
;
1547 device
->enabled_extensions
= enabled_extensions
;
1549 anv_device_init_dispatch(device
);
1551 if (pthread_mutex_init(&device
->mutex
, NULL
) != 0) {
1552 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1553 goto fail_context_id
;
1556 pthread_condattr_t condattr
;
1557 if (pthread_condattr_init(&condattr
) != 0) {
1558 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1561 if (pthread_condattr_setclock(&condattr
, CLOCK_MONOTONIC
) != 0) {
1562 pthread_condattr_destroy(&condattr
);
1563 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1566 if (pthread_cond_init(&device
->queue_submit
, NULL
) != 0) {
1567 pthread_condattr_destroy(&condattr
);
1568 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1571 pthread_condattr_destroy(&condattr
);
1574 (physical_device
->supports_48bit_addresses
? EXEC_OBJECT_SUPPORTS_48B_ADDRESS
: 0) |
1575 (physical_device
->has_exec_async
? EXEC_OBJECT_ASYNC
: 0) |
1576 (physical_device
->has_exec_capture
? EXEC_OBJECT_CAPTURE
: 0);
1578 anv_bo_pool_init(&device
->batch_bo_pool
, device
, bo_flags
);
1580 result
= anv_bo_cache_init(&device
->bo_cache
);
1581 if (result
!= VK_SUCCESS
)
1582 goto fail_batch_bo_pool
;
1584 /* For the state pools we explicitly disable 48bit. */
1585 bo_flags
= (physical_device
->has_exec_async
? EXEC_OBJECT_ASYNC
: 0) |
1586 (physical_device
->has_exec_capture
? EXEC_OBJECT_CAPTURE
: 0);
1588 result
= anv_state_pool_init(&device
->dynamic_state_pool
, device
, 16384,
1590 if (result
!= VK_SUCCESS
)
1593 result
= anv_state_pool_init(&device
->instruction_state_pool
, device
, 16384,
1595 if (result
!= VK_SUCCESS
)
1596 goto fail_dynamic_state_pool
;
1598 result
= anv_state_pool_init(&device
->surface_state_pool
, device
, 4096,
1600 if (result
!= VK_SUCCESS
)
1601 goto fail_instruction_state_pool
;
1603 result
= anv_bo_init_new(&device
->workaround_bo
, device
, 1024);
1604 if (result
!= VK_SUCCESS
)
1605 goto fail_surface_state_pool
;
1607 anv_device_init_trivial_batch(device
);
1609 anv_scratch_pool_init(device
, &device
->scratch_pool
);
1611 anv_queue_init(device
, &device
->queue
);
1613 switch (device
->info
.gen
) {
1615 if (!device
->info
.is_haswell
)
1616 result
= gen7_init_device_state(device
);
1618 result
= gen75_init_device_state(device
);
1621 result
= gen8_init_device_state(device
);
1624 result
= gen9_init_device_state(device
);
1627 result
= gen10_init_device_state(device
);
1630 result
= gen11_init_device_state(device
);
1633 /* Shouldn't get here as we don't create physical devices for any other
1635 unreachable("unhandled gen");
1637 if (result
!= VK_SUCCESS
)
1638 goto fail_workaround_bo
;
1640 anv_device_init_blorp(device
);
1642 anv_device_init_border_colors(device
);
1644 *pDevice
= anv_device_to_handle(device
);
1649 anv_queue_finish(&device
->queue
);
1650 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
1651 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
1652 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
1653 fail_surface_state_pool
:
1654 anv_state_pool_finish(&device
->surface_state_pool
);
1655 fail_instruction_state_pool
:
1656 anv_state_pool_finish(&device
->instruction_state_pool
);
1657 fail_dynamic_state_pool
:
1658 anv_state_pool_finish(&device
->dynamic_state_pool
);
1660 anv_bo_cache_finish(&device
->bo_cache
);
1662 anv_bo_pool_finish(&device
->batch_bo_pool
);
1663 pthread_cond_destroy(&device
->queue_submit
);
1665 pthread_mutex_destroy(&device
->mutex
);
1667 anv_gem_destroy_context(device
, device
->context_id
);
1671 vk_free(&device
->alloc
, device
);
1676 void anv_DestroyDevice(
1678 const VkAllocationCallbacks
* pAllocator
)
1680 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1685 anv_device_finish_blorp(device
);
1687 anv_queue_finish(&device
->queue
);
1689 #ifdef HAVE_VALGRIND
1690 /* We only need to free these to prevent valgrind errors. The backing
1691 * BO will go away in a couple of lines so we don't actually leak.
1693 anv_state_pool_free(&device
->dynamic_state_pool
, device
->border_colors
);
1696 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
1698 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
1699 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
1701 anv_gem_close(device
, device
->trivial_batch_bo
.gem_handle
);
1703 anv_state_pool_finish(&device
->surface_state_pool
);
1704 anv_state_pool_finish(&device
->instruction_state_pool
);
1705 anv_state_pool_finish(&device
->dynamic_state_pool
);
1707 anv_bo_cache_finish(&device
->bo_cache
);
1709 anv_bo_pool_finish(&device
->batch_bo_pool
);
1711 pthread_cond_destroy(&device
->queue_submit
);
1712 pthread_mutex_destroy(&device
->mutex
);
1714 anv_gem_destroy_context(device
, device
->context_id
);
1718 vk_free(&device
->alloc
, device
);
1721 VkResult
anv_EnumerateInstanceLayerProperties(
1722 uint32_t* pPropertyCount
,
1723 VkLayerProperties
* pProperties
)
1725 if (pProperties
== NULL
) {
1726 *pPropertyCount
= 0;
1730 /* None supported at this time */
1731 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1734 VkResult
anv_EnumerateDeviceLayerProperties(
1735 VkPhysicalDevice physicalDevice
,
1736 uint32_t* pPropertyCount
,
1737 VkLayerProperties
* pProperties
)
1739 if (pProperties
== NULL
) {
1740 *pPropertyCount
= 0;
1744 /* None supported at this time */
1745 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1748 void anv_GetDeviceQueue(
1750 uint32_t queueNodeIndex
,
1751 uint32_t queueIndex
,
1754 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1756 assert(queueIndex
== 0);
1758 *pQueue
= anv_queue_to_handle(&device
->queue
);
1761 void anv_GetDeviceQueue2(
1763 const VkDeviceQueueInfo2
* pQueueInfo
,
1766 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1768 assert(pQueueInfo
->queueIndex
== 0);
1770 if (pQueueInfo
->flags
== device
->queue
.flags
)
1771 *pQueue
= anv_queue_to_handle(&device
->queue
);
1777 anv_device_query_status(struct anv_device
*device
)
1779 /* This isn't likely as most of the callers of this function already check
1780 * for it. However, it doesn't hurt to check and it potentially lets us
1783 if (unlikely(device
->lost
))
1784 return VK_ERROR_DEVICE_LOST
;
1786 uint32_t active
, pending
;
1787 int ret
= anv_gem_gpu_get_reset_stats(device
, &active
, &pending
);
1789 /* We don't know the real error. */
1790 device
->lost
= true;
1791 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1792 "get_reset_stats failed: %m");
1796 device
->lost
= true;
1797 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1798 "GPU hung on one of our command buffers");
1799 } else if (pending
) {
1800 device
->lost
= true;
1801 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1802 "GPU hung with commands in-flight");
1809 anv_device_bo_busy(struct anv_device
*device
, struct anv_bo
*bo
)
1811 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
1812 * Other usages of the BO (such as on different hardware) will not be
1813 * flagged as "busy" by this ioctl. Use with care.
1815 int ret
= anv_gem_busy(device
, bo
->gem_handle
);
1817 return VK_NOT_READY
;
1818 } else if (ret
== -1) {
1819 /* We don't know the real error. */
1820 device
->lost
= true;
1821 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1822 "gem wait failed: %m");
1825 /* Query for device status after the busy call. If the BO we're checking
1826 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
1827 * client because it clearly doesn't have valid data. Yes, this most
1828 * likely means an ioctl, but we just did an ioctl to query the busy status
1829 * so it's no great loss.
1831 return anv_device_query_status(device
);
1835 anv_device_wait(struct anv_device
*device
, struct anv_bo
*bo
,
1838 int ret
= anv_gem_wait(device
, bo
->gem_handle
, &timeout
);
1839 if (ret
== -1 && errno
== ETIME
) {
1841 } else if (ret
== -1) {
1842 /* We don't know the real error. */
1843 device
->lost
= true;
1844 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1845 "gem wait failed: %m");
1848 /* Query for device status after the wait. If the BO we're waiting on got
1849 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
1850 * because it clearly doesn't have valid data. Yes, this most likely means
1851 * an ioctl, but we just did an ioctl to wait so it's no great loss.
1853 return anv_device_query_status(device
);
1856 VkResult
anv_DeviceWaitIdle(
1859 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1860 if (unlikely(device
->lost
))
1861 return VK_ERROR_DEVICE_LOST
;
1863 struct anv_batch batch
;
1866 batch
.start
= batch
.next
= cmds
;
1867 batch
.end
= (void *) cmds
+ sizeof(cmds
);
1869 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
1870 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
1872 return anv_device_submit_simple_batch(device
, &batch
);
1876 anv_bo_init_new(struct anv_bo
*bo
, struct anv_device
*device
, uint64_t size
)
1878 uint32_t gem_handle
= anv_gem_create(device
, size
);
1880 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
1882 anv_bo_init(bo
, gem_handle
, size
);
1887 VkResult
anv_AllocateMemory(
1889 const VkMemoryAllocateInfo
* pAllocateInfo
,
1890 const VkAllocationCallbacks
* pAllocator
,
1891 VkDeviceMemory
* pMem
)
1893 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1894 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
1895 struct anv_device_memory
*mem
;
1896 VkResult result
= VK_SUCCESS
;
1898 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
1900 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
1901 assert(pAllocateInfo
->allocationSize
> 0);
1903 if (pAllocateInfo
->allocationSize
> MAX_MEMORY_ALLOCATION_SIZE
)
1904 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
1906 /* FINISHME: Fail if allocation request exceeds heap size. */
1908 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
1909 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1911 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1913 assert(pAllocateInfo
->memoryTypeIndex
< pdevice
->memory
.type_count
);
1914 mem
->type
= &pdevice
->memory
.types
[pAllocateInfo
->memoryTypeIndex
];
1918 const VkImportMemoryFdInfoKHR
*fd_info
=
1919 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_FD_INFO_KHR
);
1921 /* The Vulkan spec permits handleType to be 0, in which case the struct is
1924 if (fd_info
&& fd_info
->handleType
) {
1925 /* At the moment, we support only the below handle types. */
1926 assert(fd_info
->handleType
==
1927 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
1928 fd_info
->handleType
==
1929 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
1931 result
= anv_bo_cache_import(device
, &device
->bo_cache
,
1932 fd_info
->fd
, &mem
->bo
);
1933 if (result
!= VK_SUCCESS
)
1936 VkDeviceSize aligned_alloc_size
=
1937 align_u64(pAllocateInfo
->allocationSize
, 4096);
1939 /* For security purposes, we reject importing the bo if it's smaller
1940 * than the requested allocation size. This prevents a malicious client
1941 * from passing a buffer to a trusted client, lying about the size, and
1942 * telling the trusted client to try and texture from an image that goes
1943 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
1944 * in the trusted client. The trusted client can protect itself against
1945 * this sort of attack but only if it can trust the buffer size.
1947 if (mem
->bo
->size
< aligned_alloc_size
) {
1948 result
= vk_errorf(device
->instance
, device
,
1949 VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
,
1950 "aligned allocationSize too large for "
1951 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR: "
1952 "%"PRIu64
"B > %"PRIu64
"B",
1953 aligned_alloc_size
, mem
->bo
->size
);
1954 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
1958 /* From the Vulkan spec:
1960 * "Importing memory from a file descriptor transfers ownership of
1961 * the file descriptor from the application to the Vulkan
1962 * implementation. The application must not perform any operations on
1963 * the file descriptor after a successful import."
1965 * If the import fails, we leave the file descriptor open.
1969 result
= anv_bo_cache_alloc(device
, &device
->bo_cache
,
1970 pAllocateInfo
->allocationSize
,
1972 if (result
!= VK_SUCCESS
)
1975 const VkMemoryDedicatedAllocateInfoKHR
*dedicated_info
=
1976 vk_find_struct_const(pAllocateInfo
->pNext
, MEMORY_DEDICATED_ALLOCATE_INFO_KHR
);
1977 if (dedicated_info
&& dedicated_info
->image
!= VK_NULL_HANDLE
) {
1978 ANV_FROM_HANDLE(anv_image
, image
, dedicated_info
->image
);
1980 /* Some legacy (non-modifiers) consumers need the tiling to be set on
1981 * the BO. In this case, we have a dedicated allocation.
1983 if (image
->needs_set_tiling
) {
1984 const uint32_t i915_tiling
=
1985 isl_tiling_to_i915_tiling(image
->planes
[0].surface
.isl
.tiling
);
1986 int ret
= anv_gem_set_tiling(device
, mem
->bo
->gem_handle
,
1987 image
->planes
[0].surface
.isl
.row_pitch
,
1990 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
1991 return vk_errorf(device
->instance
, NULL
,
1992 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
1993 "failed to set BO tiling: %m");
1999 assert(mem
->type
->heapIndex
< pdevice
->memory
.heap_count
);
2000 if (pdevice
->memory
.heaps
[mem
->type
->heapIndex
].supports_48bit_addresses
)
2001 mem
->bo
->flags
|= EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
2003 const struct wsi_memory_allocate_info
*wsi_info
=
2004 vk_find_struct_const(pAllocateInfo
->pNext
, WSI_MEMORY_ALLOCATE_INFO_MESA
);
2005 if (wsi_info
&& wsi_info
->implicit_sync
) {
2006 /* We need to set the WRITE flag on window system buffers so that GEM
2007 * will know we're writing to them and synchronize uses on other rings
2008 * (eg if the display server uses the blitter ring).
2010 mem
->bo
->flags
|= EXEC_OBJECT_WRITE
;
2011 } else if (pdevice
->has_exec_async
) {
2012 mem
->bo
->flags
|= EXEC_OBJECT_ASYNC
;
2015 *pMem
= anv_device_memory_to_handle(mem
);
2020 vk_free2(&device
->alloc
, pAllocator
, mem
);
2025 VkResult
anv_GetMemoryFdKHR(
2027 const VkMemoryGetFdInfoKHR
* pGetFdInfo
,
2030 ANV_FROM_HANDLE(anv_device
, dev
, device_h
);
2031 ANV_FROM_HANDLE(anv_device_memory
, mem
, pGetFdInfo
->memory
);
2033 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
2035 assert(pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
2036 pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
2038 return anv_bo_cache_export(dev
, &dev
->bo_cache
, mem
->bo
, pFd
);
2041 VkResult
anv_GetMemoryFdPropertiesKHR(
2043 VkExternalMemoryHandleTypeFlagBitsKHR handleType
,
2045 VkMemoryFdPropertiesKHR
* pMemoryFdProperties
)
2047 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2048 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2050 switch (handleType
) {
2051 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
:
2052 /* dma-buf can be imported as any memory type */
2053 pMemoryFdProperties
->memoryTypeBits
=
2054 (1 << pdevice
->memory
.type_count
) - 1;
2058 /* The valid usage section for this function says:
2060 * "handleType must not be one of the handle types defined as
2063 * So opaque handle types fall into the default "unsupported" case.
2065 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
2069 void anv_FreeMemory(
2071 VkDeviceMemory _mem
,
2072 const VkAllocationCallbacks
* pAllocator
)
2074 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2075 ANV_FROM_HANDLE(anv_device_memory
, mem
, _mem
);
2081 anv_UnmapMemory(_device
, _mem
);
2083 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
2085 vk_free2(&device
->alloc
, pAllocator
, mem
);
2088 VkResult
anv_MapMemory(
2090 VkDeviceMemory _memory
,
2091 VkDeviceSize offset
,
2093 VkMemoryMapFlags flags
,
2096 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2097 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
2104 if (size
== VK_WHOLE_SIZE
)
2105 size
= mem
->bo
->size
- offset
;
2107 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
2109 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
2110 * assert(size != 0);
2111 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
2112 * equal to the size of the memory minus offset
2115 assert(offset
+ size
<= mem
->bo
->size
);
2117 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
2118 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
2119 * at a time is valid. We could just mmap up front and return an offset
2120 * pointer here, but that may exhaust virtual memory on 32 bit
2123 uint32_t gem_flags
= 0;
2125 if (!device
->info
.has_llc
&&
2126 (mem
->type
->propertyFlags
& VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
))
2127 gem_flags
|= I915_MMAP_WC
;
2129 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
2130 uint64_t map_offset
= offset
& ~4095ull;
2131 assert(offset
>= map_offset
);
2132 uint64_t map_size
= (offset
+ size
) - map_offset
;
2134 /* Let's map whole pages */
2135 map_size
= align_u64(map_size
, 4096);
2137 void *map
= anv_gem_mmap(device
, mem
->bo
->gem_handle
,
2138 map_offset
, map_size
, gem_flags
);
2139 if (map
== MAP_FAILED
)
2140 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
2143 mem
->map_size
= map_size
;
2145 *ppData
= mem
->map
+ (offset
- map_offset
);
2150 void anv_UnmapMemory(
2152 VkDeviceMemory _memory
)
2154 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
2159 anv_gem_munmap(mem
->map
, mem
->map_size
);
2166 clflush_mapped_ranges(struct anv_device
*device
,
2168 const VkMappedMemoryRange
*ranges
)
2170 for (uint32_t i
= 0; i
< count
; i
++) {
2171 ANV_FROM_HANDLE(anv_device_memory
, mem
, ranges
[i
].memory
);
2172 if (ranges
[i
].offset
>= mem
->map_size
)
2175 gen_clflush_range(mem
->map
+ ranges
[i
].offset
,
2176 MIN2(ranges
[i
].size
, mem
->map_size
- ranges
[i
].offset
));
2180 VkResult
anv_FlushMappedMemoryRanges(
2182 uint32_t memoryRangeCount
,
2183 const VkMappedMemoryRange
* pMemoryRanges
)
2185 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2187 if (device
->info
.has_llc
)
2190 /* Make sure the writes we're flushing have landed. */
2191 __builtin_ia32_mfence();
2193 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
2198 VkResult
anv_InvalidateMappedMemoryRanges(
2200 uint32_t memoryRangeCount
,
2201 const VkMappedMemoryRange
* pMemoryRanges
)
2203 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2205 if (device
->info
.has_llc
)
2208 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
2210 /* Make sure no reads get moved up above the invalidate. */
2211 __builtin_ia32_mfence();
2216 void anv_GetBufferMemoryRequirements(
2219 VkMemoryRequirements
* pMemoryRequirements
)
2221 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
2222 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2223 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2225 /* The Vulkan spec (git aaed022) says:
2227 * memoryTypeBits is a bitfield and contains one bit set for every
2228 * supported memory type for the resource. The bit `1<<i` is set if and
2229 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
2230 * structure for the physical device is supported.
2232 uint32_t memory_types
= 0;
2233 for (uint32_t i
= 0; i
< pdevice
->memory
.type_count
; i
++) {
2234 uint32_t valid_usage
= pdevice
->memory
.types
[i
].valid_buffer_usage
;
2235 if ((valid_usage
& buffer
->usage
) == buffer
->usage
)
2236 memory_types
|= (1u << i
);
2239 /* Base alignment requirement of a cache line */
2240 uint32_t alignment
= 16;
2242 /* We need an alignment of 32 for pushing UBOs */
2243 if (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
)
2244 alignment
= MAX2(alignment
, 32);
2246 pMemoryRequirements
->size
= buffer
->size
;
2247 pMemoryRequirements
->alignment
= alignment
;
2249 /* Storage and Uniform buffers should have their size aligned to
2250 * 32-bits to avoid boundary checks when last DWord is not complete.
2251 * This would ensure that not internal padding would be needed for
2254 if (device
->robust_buffer_access
&&
2255 (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
||
2256 buffer
->usage
& VK_BUFFER_USAGE_STORAGE_BUFFER_BIT
))
2257 pMemoryRequirements
->size
= align_u64(buffer
->size
, 4);
2259 pMemoryRequirements
->memoryTypeBits
= memory_types
;
2262 void anv_GetBufferMemoryRequirements2(
2264 const VkBufferMemoryRequirementsInfo2
* pInfo
,
2265 VkMemoryRequirements2
* pMemoryRequirements
)
2267 anv_GetBufferMemoryRequirements(_device
, pInfo
->buffer
,
2268 &pMemoryRequirements
->memoryRequirements
);
2270 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
2271 switch (ext
->sType
) {
2272 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
2273 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
2274 requirements
->prefersDedicatedAllocation
= VK_FALSE
;
2275 requirements
->requiresDedicatedAllocation
= VK_FALSE
;
2280 anv_debug_ignored_stype(ext
->sType
);
2286 void anv_GetImageMemoryRequirements(
2289 VkMemoryRequirements
* pMemoryRequirements
)
2291 ANV_FROM_HANDLE(anv_image
, image
, _image
);
2292 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2293 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2295 /* The Vulkan spec (git aaed022) says:
2297 * memoryTypeBits is a bitfield and contains one bit set for every
2298 * supported memory type for the resource. The bit `1<<i` is set if and
2299 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
2300 * structure for the physical device is supported.
2302 * All types are currently supported for images.
2304 uint32_t memory_types
= (1ull << pdevice
->memory
.type_count
) - 1;
2306 pMemoryRequirements
->size
= image
->size
;
2307 pMemoryRequirements
->alignment
= image
->alignment
;
2308 pMemoryRequirements
->memoryTypeBits
= memory_types
;
2311 void anv_GetImageMemoryRequirements2(
2313 const VkImageMemoryRequirementsInfo2
* pInfo
,
2314 VkMemoryRequirements2
* pMemoryRequirements
)
2316 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2317 ANV_FROM_HANDLE(anv_image
, image
, pInfo
->image
);
2319 anv_GetImageMemoryRequirements(_device
, pInfo
->image
,
2320 &pMemoryRequirements
->memoryRequirements
);
2322 vk_foreach_struct_const(ext
, pInfo
->pNext
) {
2323 switch (ext
->sType
) {
2324 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO
: {
2325 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2326 const VkImagePlaneMemoryRequirementsInfoKHR
*plane_reqs
=
2327 (const VkImagePlaneMemoryRequirementsInfoKHR
*) ext
;
2328 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
,
2329 plane_reqs
->planeAspect
);
2331 assert(image
->planes
[plane
].offset
== 0);
2333 /* The Vulkan spec (git aaed022) says:
2335 * memoryTypeBits is a bitfield and contains one bit set for every
2336 * supported memory type for the resource. The bit `1<<i` is set
2337 * if and only if the memory type `i` in the
2338 * VkPhysicalDeviceMemoryProperties structure for the physical
2339 * device is supported.
2341 * All types are currently supported for images.
2343 pMemoryRequirements
->memoryRequirements
.memoryTypeBits
=
2344 (1ull << pdevice
->memory
.type_count
) - 1;
2346 pMemoryRequirements
->memoryRequirements
.size
= image
->planes
[plane
].size
;
2347 pMemoryRequirements
->memoryRequirements
.alignment
=
2348 image
->planes
[plane
].alignment
;
2353 anv_debug_ignored_stype(ext
->sType
);
2358 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
2359 switch (ext
->sType
) {
2360 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
2361 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
2362 if (image
->needs_set_tiling
) {
2363 /* If we need to set the tiling for external consumers, we need a
2364 * dedicated allocation.
2366 * See also anv_AllocateMemory.
2368 requirements
->prefersDedicatedAllocation
= VK_TRUE
;
2369 requirements
->requiresDedicatedAllocation
= VK_TRUE
;
2371 requirements
->prefersDedicatedAllocation
= VK_FALSE
;
2372 requirements
->requiresDedicatedAllocation
= VK_FALSE
;
2378 anv_debug_ignored_stype(ext
->sType
);
2384 void anv_GetImageSparseMemoryRequirements(
2387 uint32_t* pSparseMemoryRequirementCount
,
2388 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
2390 *pSparseMemoryRequirementCount
= 0;
2393 void anv_GetImageSparseMemoryRequirements2(
2395 const VkImageSparseMemoryRequirementsInfo2
* pInfo
,
2396 uint32_t* pSparseMemoryRequirementCount
,
2397 VkSparseImageMemoryRequirements2
* pSparseMemoryRequirements
)
2399 *pSparseMemoryRequirementCount
= 0;
2402 void anv_GetDeviceMemoryCommitment(
2404 VkDeviceMemory memory
,
2405 VkDeviceSize
* pCommittedMemoryInBytes
)
2407 *pCommittedMemoryInBytes
= 0;
2411 anv_bind_buffer_memory(const VkBindBufferMemoryInfo
*pBindInfo
)
2413 ANV_FROM_HANDLE(anv_device_memory
, mem
, pBindInfo
->memory
);
2414 ANV_FROM_HANDLE(anv_buffer
, buffer
, pBindInfo
->buffer
);
2416 assert(pBindInfo
->sType
== VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
);
2419 assert((buffer
->usage
& mem
->type
->valid_buffer_usage
) == buffer
->usage
);
2420 buffer
->bo
= mem
->bo
;
2421 buffer
->offset
= pBindInfo
->memoryOffset
;
2428 VkResult
anv_BindBufferMemory(
2431 VkDeviceMemory memory
,
2432 VkDeviceSize memoryOffset
)
2434 anv_bind_buffer_memory(
2435 &(VkBindBufferMemoryInfo
) {
2436 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
,
2439 .memoryOffset
= memoryOffset
,
2445 VkResult
anv_BindBufferMemory2(
2447 uint32_t bindInfoCount
,
2448 const VkBindBufferMemoryInfo
* pBindInfos
)
2450 for (uint32_t i
= 0; i
< bindInfoCount
; i
++)
2451 anv_bind_buffer_memory(&pBindInfos
[i
]);
2456 VkResult
anv_QueueBindSparse(
2458 uint32_t bindInfoCount
,
2459 const VkBindSparseInfo
* pBindInfo
,
2462 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
2463 if (unlikely(queue
->device
->lost
))
2464 return VK_ERROR_DEVICE_LOST
;
2466 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
2471 VkResult
anv_CreateEvent(
2473 const VkEventCreateInfo
* pCreateInfo
,
2474 const VkAllocationCallbacks
* pAllocator
,
2477 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2478 struct anv_state state
;
2479 struct anv_event
*event
;
2481 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_EVENT_CREATE_INFO
);
2483 state
= anv_state_pool_alloc(&device
->dynamic_state_pool
,
2486 event
->state
= state
;
2487 event
->semaphore
= VK_EVENT_RESET
;
2489 if (!device
->info
.has_llc
) {
2490 /* Make sure the writes we're flushing have landed. */
2491 __builtin_ia32_mfence();
2492 __builtin_ia32_clflush(event
);
2495 *pEvent
= anv_event_to_handle(event
);
2500 void anv_DestroyEvent(
2503 const VkAllocationCallbacks
* pAllocator
)
2505 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2506 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2511 anv_state_pool_free(&device
->dynamic_state_pool
, event
->state
);
2514 VkResult
anv_GetEventStatus(
2518 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2519 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2521 if (unlikely(device
->lost
))
2522 return VK_ERROR_DEVICE_LOST
;
2524 if (!device
->info
.has_llc
) {
2525 /* Invalidate read cache before reading event written by GPU. */
2526 __builtin_ia32_clflush(event
);
2527 __builtin_ia32_mfence();
2531 return event
->semaphore
;
2534 VkResult
anv_SetEvent(
2538 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2539 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2541 event
->semaphore
= VK_EVENT_SET
;
2543 if (!device
->info
.has_llc
) {
2544 /* Make sure the writes we're flushing have landed. */
2545 __builtin_ia32_mfence();
2546 __builtin_ia32_clflush(event
);
2552 VkResult
anv_ResetEvent(
2556 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2557 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2559 event
->semaphore
= VK_EVENT_RESET
;
2561 if (!device
->info
.has_llc
) {
2562 /* Make sure the writes we're flushing have landed. */
2563 __builtin_ia32_mfence();
2564 __builtin_ia32_clflush(event
);
2572 VkResult
anv_CreateBuffer(
2574 const VkBufferCreateInfo
* pCreateInfo
,
2575 const VkAllocationCallbacks
* pAllocator
,
2578 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2579 struct anv_buffer
*buffer
;
2581 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
2583 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
2584 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2586 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2588 buffer
->size
= pCreateInfo
->size
;
2589 buffer
->usage
= pCreateInfo
->usage
;
2593 *pBuffer
= anv_buffer_to_handle(buffer
);
2598 void anv_DestroyBuffer(
2601 const VkAllocationCallbacks
* pAllocator
)
2603 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2604 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
2609 vk_free2(&device
->alloc
, pAllocator
, buffer
);
2613 anv_fill_buffer_surface_state(struct anv_device
*device
, struct anv_state state
,
2614 enum isl_format format
,
2615 uint32_t offset
, uint32_t range
, uint32_t stride
)
2617 isl_buffer_fill_state(&device
->isl_dev
, state
.map
,
2619 .mocs
= device
->default_mocs
,
2624 anv_state_flush(device
, state
);
2627 void anv_DestroySampler(
2630 const VkAllocationCallbacks
* pAllocator
)
2632 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2633 ANV_FROM_HANDLE(anv_sampler
, sampler
, _sampler
);
2638 vk_free2(&device
->alloc
, pAllocator
, sampler
);
2641 VkResult
anv_CreateFramebuffer(
2643 const VkFramebufferCreateInfo
* pCreateInfo
,
2644 const VkAllocationCallbacks
* pAllocator
,
2645 VkFramebuffer
* pFramebuffer
)
2647 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2648 struct anv_framebuffer
*framebuffer
;
2650 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
2652 size_t size
= sizeof(*framebuffer
) +
2653 sizeof(struct anv_image_view
*) * pCreateInfo
->attachmentCount
;
2654 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
2655 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2656 if (framebuffer
== NULL
)
2657 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2659 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
2660 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
2661 VkImageView _iview
= pCreateInfo
->pAttachments
[i
];
2662 framebuffer
->attachments
[i
] = anv_image_view_from_handle(_iview
);
2665 framebuffer
->width
= pCreateInfo
->width
;
2666 framebuffer
->height
= pCreateInfo
->height
;
2667 framebuffer
->layers
= pCreateInfo
->layers
;
2669 *pFramebuffer
= anv_framebuffer_to_handle(framebuffer
);
2674 void anv_DestroyFramebuffer(
2677 const VkAllocationCallbacks
* pAllocator
)
2679 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2680 ANV_FROM_HANDLE(anv_framebuffer
, fb
, _fb
);
2685 vk_free2(&device
->alloc
, pAllocator
, fb
);
2688 /* vk_icd.h does not declare this function, so we declare it here to
2689 * suppress Wmissing-prototypes.
2691 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2692 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
);
2694 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2695 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
)
2697 /* For the full details on loader interface versioning, see
2698 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
2699 * What follows is a condensed summary, to help you navigate the large and
2700 * confusing official doc.
2702 * - Loader interface v0 is incompatible with later versions. We don't
2705 * - In loader interface v1:
2706 * - The first ICD entrypoint called by the loader is
2707 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
2709 * - The ICD must statically expose no other Vulkan symbol unless it is
2710 * linked with -Bsymbolic.
2711 * - Each dispatchable Vulkan handle created by the ICD must be
2712 * a pointer to a struct whose first member is VK_LOADER_DATA. The
2713 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
2714 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
2715 * vkDestroySurfaceKHR(). The ICD must be capable of working with
2716 * such loader-managed surfaces.
2718 * - Loader interface v2 differs from v1 in:
2719 * - The first ICD entrypoint called by the loader is
2720 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
2721 * statically expose this entrypoint.
2723 * - Loader interface v3 differs from v2 in:
2724 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
2725 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
2726 * because the loader no longer does so.
2728 *pSupportedVersion
= MIN2(*pSupportedVersion
, 3u);