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 struct anv_instance_extension_table enabled_extensions
= {};
518 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
520 for (idx
= 0; idx
< ANV_INSTANCE_EXTENSION_COUNT
; idx
++) {
521 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
522 anv_instance_extensions
[idx
].extensionName
) == 0)
526 if (idx
>= ANV_INSTANCE_EXTENSION_COUNT
)
527 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
529 if (!anv_instance_extensions_supported
.extensions
[idx
])
530 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
532 enabled_extensions
.extensions
[idx
] = true;
535 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
536 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
538 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
540 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
543 instance
->alloc
= *pAllocator
;
545 instance
->alloc
= default_alloc
;
547 if (pCreateInfo
->pApplicationInfo
&&
548 pCreateInfo
->pApplicationInfo
->apiVersion
!= 0) {
549 instance
->apiVersion
= pCreateInfo
->pApplicationInfo
->apiVersion
;
551 anv_EnumerateInstanceVersion(&instance
->apiVersion
);
554 instance
->enabled_extensions
= enabled_extensions
;
556 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->dispatch
.entrypoints
); i
++) {
557 /* Vulkan requires that entrypoints for extensions which have not been
558 * enabled must not be advertised.
560 if (!anv_entrypoint_is_enabled(i
, instance
->apiVersion
,
561 &instance
->enabled_extensions
, NULL
)) {
562 instance
->dispatch
.entrypoints
[i
] = NULL
;
563 } else if (anv_dispatch_table
.entrypoints
[i
] != NULL
) {
564 instance
->dispatch
.entrypoints
[i
] = anv_dispatch_table
.entrypoints
[i
];
566 instance
->dispatch
.entrypoints
[i
] =
567 anv_tramp_dispatch_table
.entrypoints
[i
];
571 instance
->physicalDeviceCount
= -1;
573 result
= vk_debug_report_instance_init(&instance
->debug_report_callbacks
);
574 if (result
!= VK_SUCCESS
) {
575 vk_free2(&default_alloc
, pAllocator
, instance
);
576 return vk_error(result
);
581 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
583 *pInstance
= anv_instance_to_handle(instance
);
588 void anv_DestroyInstance(
589 VkInstance _instance
,
590 const VkAllocationCallbacks
* pAllocator
)
592 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
597 if (instance
->physicalDeviceCount
> 0) {
598 /* We support at most one physical device. */
599 assert(instance
->physicalDeviceCount
== 1);
600 anv_physical_device_finish(&instance
->physicalDevice
);
603 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
605 vk_debug_report_instance_destroy(&instance
->debug_report_callbacks
);
609 vk_free(&instance
->alloc
, instance
);
613 anv_enumerate_devices(struct anv_instance
*instance
)
615 /* TODO: Check for more devices ? */
616 drmDevicePtr devices
[8];
617 VkResult result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
620 instance
->physicalDeviceCount
= 0;
622 max_devices
= drmGetDevices2(0, devices
, ARRAY_SIZE(devices
));
624 return VK_ERROR_INCOMPATIBLE_DRIVER
;
626 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
627 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
628 devices
[i
]->bustype
== DRM_BUS_PCI
&&
629 devices
[i
]->deviceinfo
.pci
->vendor_id
== 0x8086) {
631 result
= anv_physical_device_init(&instance
->physicalDevice
,
633 devices
[i
]->nodes
[DRM_NODE_RENDER
]);
634 if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
638 drmFreeDevices(devices
, max_devices
);
640 if (result
== VK_SUCCESS
)
641 instance
->physicalDeviceCount
= 1;
647 anv_instance_ensure_physical_device(struct anv_instance
*instance
)
649 if (instance
->physicalDeviceCount
< 0) {
650 VkResult result
= anv_enumerate_devices(instance
);
651 if (result
!= VK_SUCCESS
&&
652 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
659 VkResult
anv_EnumeratePhysicalDevices(
660 VkInstance _instance
,
661 uint32_t* pPhysicalDeviceCount
,
662 VkPhysicalDevice
* pPhysicalDevices
)
664 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
665 VK_OUTARRAY_MAKE(out
, pPhysicalDevices
, pPhysicalDeviceCount
);
667 VkResult result
= anv_instance_ensure_physical_device(instance
);
668 if (result
!= VK_SUCCESS
)
671 if (instance
->physicalDeviceCount
== 0)
674 assert(instance
->physicalDeviceCount
== 1);
675 vk_outarray_append(&out
, i
) {
676 *i
= anv_physical_device_to_handle(&instance
->physicalDevice
);
679 return vk_outarray_status(&out
);
682 VkResult
anv_EnumeratePhysicalDeviceGroups(
683 VkInstance _instance
,
684 uint32_t* pPhysicalDeviceGroupCount
,
685 VkPhysicalDeviceGroupProperties
* pPhysicalDeviceGroupProperties
)
687 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
688 VK_OUTARRAY_MAKE(out
, pPhysicalDeviceGroupProperties
,
689 pPhysicalDeviceGroupCount
);
691 VkResult result
= anv_instance_ensure_physical_device(instance
);
692 if (result
!= VK_SUCCESS
)
695 if (instance
->physicalDeviceCount
== 0)
698 assert(instance
->physicalDeviceCount
== 1);
700 vk_outarray_append(&out
, p
) {
701 p
->physicalDeviceCount
= 1;
702 memset(p
->physicalDevices
, 0, sizeof(p
->physicalDevices
));
703 p
->physicalDevices
[0] =
704 anv_physical_device_to_handle(&instance
->physicalDevice
);
705 p
->subsetAllocation
= VK_FALSE
;
707 vk_foreach_struct(ext
, p
->pNext
)
708 anv_debug_ignored_stype(ext
->sType
);
711 return vk_outarray_status(&out
);
714 void anv_GetPhysicalDeviceFeatures(
715 VkPhysicalDevice physicalDevice
,
716 VkPhysicalDeviceFeatures
* pFeatures
)
718 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
720 *pFeatures
= (VkPhysicalDeviceFeatures
) {
721 .robustBufferAccess
= true,
722 .fullDrawIndexUint32
= true,
723 .imageCubeArray
= true,
724 .independentBlend
= true,
725 .geometryShader
= true,
726 .tessellationShader
= true,
727 .sampleRateShading
= true,
728 .dualSrcBlend
= true,
730 .multiDrawIndirect
= true,
731 .drawIndirectFirstInstance
= true,
733 .depthBiasClamp
= true,
734 .fillModeNonSolid
= true,
735 .depthBounds
= false,
739 .multiViewport
= true,
740 .samplerAnisotropy
= true,
741 .textureCompressionETC2
= pdevice
->info
.gen
>= 8 ||
742 pdevice
->info
.is_baytrail
,
743 .textureCompressionASTC_LDR
= pdevice
->info
.gen
>= 9, /* FINISHME CHV */
744 .textureCompressionBC
= true,
745 .occlusionQueryPrecise
= true,
746 .pipelineStatisticsQuery
= true,
747 .fragmentStoresAndAtomics
= true,
748 .shaderTessellationAndGeometryPointSize
= true,
749 .shaderImageGatherExtended
= true,
750 .shaderStorageImageExtendedFormats
= true,
751 .shaderStorageImageMultisample
= false,
752 .shaderStorageImageReadWithoutFormat
= false,
753 .shaderStorageImageWriteWithoutFormat
= true,
754 .shaderUniformBufferArrayDynamicIndexing
= true,
755 .shaderSampledImageArrayDynamicIndexing
= true,
756 .shaderStorageBufferArrayDynamicIndexing
= true,
757 .shaderStorageImageArrayDynamicIndexing
= true,
758 .shaderClipDistance
= true,
759 .shaderCullDistance
= true,
760 .shaderFloat64
= pdevice
->info
.gen
>= 8,
761 .shaderInt64
= pdevice
->info
.gen
>= 8,
762 .shaderInt16
= false,
763 .shaderResourceMinLod
= false,
764 .variableMultisampleRate
= false,
765 .inheritedQueries
= true,
768 /* We can't do image stores in vec4 shaders */
769 pFeatures
->vertexPipelineStoresAndAtomics
=
770 pdevice
->compiler
->scalar_stage
[MESA_SHADER_VERTEX
] &&
771 pdevice
->compiler
->scalar_stage
[MESA_SHADER_GEOMETRY
];
774 void anv_GetPhysicalDeviceFeatures2(
775 VkPhysicalDevice physicalDevice
,
776 VkPhysicalDeviceFeatures2
* pFeatures
)
778 anv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
780 vk_foreach_struct(ext
, pFeatures
->pNext
) {
781 switch (ext
->sType
) {
782 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES
: {
783 VkPhysicalDeviceProtectedMemoryFeatures
*features
= (void *)ext
;
784 features
->protectedMemory
= VK_FALSE
;
788 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES
: {
789 VkPhysicalDeviceMultiviewFeatures
*features
=
790 (VkPhysicalDeviceMultiviewFeatures
*)ext
;
791 features
->multiview
= true;
792 features
->multiviewGeometryShader
= true;
793 features
->multiviewTessellationShader
= true;
797 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTER_FEATURES
: {
798 VkPhysicalDeviceVariablePointerFeatures
*features
= (void *)ext
;
799 features
->variablePointersStorageBuffer
= true;
800 features
->variablePointers
= true;
804 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES
: {
805 VkPhysicalDeviceSamplerYcbcrConversionFeatures
*features
=
806 (VkPhysicalDeviceSamplerYcbcrConversionFeatures
*) ext
;
807 features
->samplerYcbcrConversion
= true;
811 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETER_FEATURES
: {
812 VkPhysicalDeviceShaderDrawParameterFeatures
*features
= (void *)ext
;
813 features
->shaderDrawParameters
= true;
817 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES_KHR
: {
818 VkPhysicalDevice16BitStorageFeaturesKHR
*features
=
819 (VkPhysicalDevice16BitStorageFeaturesKHR
*)ext
;
820 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
822 features
->storageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
823 features
->uniformAndStorageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
824 features
->storagePushConstant16
= pdevice
->info
.gen
>= 8;
825 features
->storageInputOutput16
= false;
830 anv_debug_ignored_stype(ext
->sType
);
836 void anv_GetPhysicalDeviceProperties(
837 VkPhysicalDevice physicalDevice
,
838 VkPhysicalDeviceProperties
* pProperties
)
840 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
841 const struct gen_device_info
*devinfo
= &pdevice
->info
;
843 /* See assertions made when programming the buffer surface state. */
844 const uint32_t max_raw_buffer_sz
= devinfo
->gen
>= 7 ?
845 (1ul << 30) : (1ul << 27);
847 const uint32_t max_samplers
= (devinfo
->gen
>= 8 || devinfo
->is_haswell
) ?
850 VkSampleCountFlags sample_counts
=
851 isl_device_get_sample_counts(&pdevice
->isl_dev
);
853 VkPhysicalDeviceLimits limits
= {
854 .maxImageDimension1D
= (1 << 14),
855 .maxImageDimension2D
= (1 << 14),
856 .maxImageDimension3D
= (1 << 11),
857 .maxImageDimensionCube
= (1 << 14),
858 .maxImageArrayLayers
= (1 << 11),
859 .maxTexelBufferElements
= 128 * 1024 * 1024,
860 .maxUniformBufferRange
= (1ul << 27),
861 .maxStorageBufferRange
= max_raw_buffer_sz
,
862 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
863 .maxMemoryAllocationCount
= UINT32_MAX
,
864 .maxSamplerAllocationCount
= 64 * 1024,
865 .bufferImageGranularity
= 64, /* A cache line */
866 .sparseAddressSpaceSize
= 0,
867 .maxBoundDescriptorSets
= MAX_SETS
,
868 .maxPerStageDescriptorSamplers
= max_samplers
,
869 .maxPerStageDescriptorUniformBuffers
= 64,
870 .maxPerStageDescriptorStorageBuffers
= 64,
871 .maxPerStageDescriptorSampledImages
= max_samplers
,
872 .maxPerStageDescriptorStorageImages
= 64,
873 .maxPerStageDescriptorInputAttachments
= 64,
874 .maxPerStageResources
= 250,
875 .maxDescriptorSetSamplers
= 6 * max_samplers
, /* number of stages * maxPerStageDescriptorSamplers */
876 .maxDescriptorSetUniformBuffers
= 6 * 64, /* number of stages * maxPerStageDescriptorUniformBuffers */
877 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
878 .maxDescriptorSetStorageBuffers
= 6 * 64, /* number of stages * maxPerStageDescriptorStorageBuffers */
879 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
880 .maxDescriptorSetSampledImages
= 6 * max_samplers
, /* number of stages * maxPerStageDescriptorSampledImages */
881 .maxDescriptorSetStorageImages
= 6 * 64, /* number of stages * maxPerStageDescriptorStorageImages */
882 .maxDescriptorSetInputAttachments
= 256,
883 .maxVertexInputAttributes
= MAX_VBS
,
884 .maxVertexInputBindings
= MAX_VBS
,
885 .maxVertexInputAttributeOffset
= 2047,
886 .maxVertexInputBindingStride
= 2048,
887 .maxVertexOutputComponents
= 128,
888 .maxTessellationGenerationLevel
= 64,
889 .maxTessellationPatchSize
= 32,
890 .maxTessellationControlPerVertexInputComponents
= 128,
891 .maxTessellationControlPerVertexOutputComponents
= 128,
892 .maxTessellationControlPerPatchOutputComponents
= 128,
893 .maxTessellationControlTotalOutputComponents
= 2048,
894 .maxTessellationEvaluationInputComponents
= 128,
895 .maxTessellationEvaluationOutputComponents
= 128,
896 .maxGeometryShaderInvocations
= 32,
897 .maxGeometryInputComponents
= 64,
898 .maxGeometryOutputComponents
= 128,
899 .maxGeometryOutputVertices
= 256,
900 .maxGeometryTotalOutputComponents
= 1024,
901 .maxFragmentInputComponents
= 128,
902 .maxFragmentOutputAttachments
= 8,
903 .maxFragmentDualSrcAttachments
= 1,
904 .maxFragmentCombinedOutputResources
= 8,
905 .maxComputeSharedMemorySize
= 32768,
906 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
907 .maxComputeWorkGroupInvocations
= 16 * devinfo
->max_cs_threads
,
908 .maxComputeWorkGroupSize
= {
909 16 * devinfo
->max_cs_threads
,
910 16 * devinfo
->max_cs_threads
,
911 16 * devinfo
->max_cs_threads
,
913 .subPixelPrecisionBits
= 4 /* FIXME */,
914 .subTexelPrecisionBits
= 4 /* FIXME */,
915 .mipmapPrecisionBits
= 4 /* FIXME */,
916 .maxDrawIndexedIndexValue
= UINT32_MAX
,
917 .maxDrawIndirectCount
= UINT32_MAX
,
918 .maxSamplerLodBias
= 16,
919 .maxSamplerAnisotropy
= 16,
920 .maxViewports
= MAX_VIEWPORTS
,
921 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
922 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
923 .viewportSubPixelBits
= 13, /* We take a float? */
924 .minMemoryMapAlignment
= 4096, /* A page */
925 .minTexelBufferOffsetAlignment
= 1,
926 /* We need 16 for UBO block reads to work and 32 for push UBOs */
927 .minUniformBufferOffsetAlignment
= 32,
928 .minStorageBufferOffsetAlignment
= 4,
929 .minTexelOffset
= -8,
931 .minTexelGatherOffset
= -32,
932 .maxTexelGatherOffset
= 31,
933 .minInterpolationOffset
= -0.5,
934 .maxInterpolationOffset
= 0.4375,
935 .subPixelInterpolationOffsetBits
= 4,
936 .maxFramebufferWidth
= (1 << 14),
937 .maxFramebufferHeight
= (1 << 14),
938 .maxFramebufferLayers
= (1 << 11),
939 .framebufferColorSampleCounts
= sample_counts
,
940 .framebufferDepthSampleCounts
= sample_counts
,
941 .framebufferStencilSampleCounts
= sample_counts
,
942 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
943 .maxColorAttachments
= MAX_RTS
,
944 .sampledImageColorSampleCounts
= sample_counts
,
945 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
946 .sampledImageDepthSampleCounts
= sample_counts
,
947 .sampledImageStencilSampleCounts
= sample_counts
,
948 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
949 .maxSampleMaskWords
= 1,
950 .timestampComputeAndGraphics
= false,
951 .timestampPeriod
= 1000000000.0 / devinfo
->timestamp_frequency
,
952 .maxClipDistances
= 8,
953 .maxCullDistances
= 8,
954 .maxCombinedClipAndCullDistances
= 8,
955 .discreteQueuePriorities
= 1,
956 .pointSizeRange
= { 0.125, 255.875 },
957 .lineWidthRange
= { 0.0, 7.9921875 },
958 .pointSizeGranularity
= (1.0 / 8.0),
959 .lineWidthGranularity
= (1.0 / 128.0),
960 .strictLines
= false, /* FINISHME */
961 .standardSampleLocations
= true,
962 .optimalBufferCopyOffsetAlignment
= 128,
963 .optimalBufferCopyRowPitchAlignment
= 128,
964 .nonCoherentAtomSize
= 64,
967 *pProperties
= (VkPhysicalDeviceProperties
) {
968 .apiVersion
= anv_physical_device_api_version(pdevice
),
969 .driverVersion
= vk_get_driver_version(),
971 .deviceID
= pdevice
->chipset_id
,
972 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
974 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
977 snprintf(pProperties
->deviceName
, sizeof(pProperties
->deviceName
),
978 "%s", pdevice
->name
);
979 memcpy(pProperties
->pipelineCacheUUID
,
980 pdevice
->pipeline_cache_uuid
, VK_UUID_SIZE
);
983 void anv_GetPhysicalDeviceProperties2(
984 VkPhysicalDevice physicalDevice
,
985 VkPhysicalDeviceProperties2
* pProperties
)
987 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
989 anv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
991 vk_foreach_struct(ext
, pProperties
->pNext
) {
992 switch (ext
->sType
) {
993 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
994 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
995 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
997 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
1001 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES
: {
1002 VkPhysicalDeviceIDProperties
*id_props
=
1003 (VkPhysicalDeviceIDProperties
*)ext
;
1004 memcpy(id_props
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
1005 memcpy(id_props
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
1006 /* The LUID is for Windows. */
1007 id_props
->deviceLUIDValid
= false;
1011 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES
: {
1012 VkPhysicalDeviceMaintenance3Properties
*props
=
1013 (VkPhysicalDeviceMaintenance3Properties
*)ext
;
1014 /* This value doesn't matter for us today as our per-stage
1015 * descriptors are the real limit.
1017 props
->maxPerSetDescriptors
= 1024;
1018 props
->maxMemoryAllocationSize
= MAX_MEMORY_ALLOCATION_SIZE
;
1022 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES
: {
1023 VkPhysicalDeviceMultiviewProperties
*properties
=
1024 (VkPhysicalDeviceMultiviewProperties
*)ext
;
1025 properties
->maxMultiviewViewCount
= 16;
1026 properties
->maxMultiviewInstanceIndex
= UINT32_MAX
/ 16;
1030 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES
: {
1031 VkPhysicalDevicePointClippingProperties
*properties
=
1032 (VkPhysicalDevicePointClippingProperties
*) ext
;
1033 properties
->pointClippingBehavior
= VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES
;
1034 anv_finishme("Implement pop-free point clipping");
1038 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES
: {
1039 VkPhysicalDeviceSubgroupProperties
*properties
= (void *)ext
;
1041 properties
->subgroupSize
= BRW_SUBGROUP_SIZE
;
1043 VkShaderStageFlags scalar_stages
= 0;
1044 for (unsigned stage
= 0; stage
< MESA_SHADER_STAGES
; stage
++) {
1045 if (pdevice
->compiler
->scalar_stage
[stage
])
1046 scalar_stages
|= mesa_to_vk_shader_stage(stage
);
1048 properties
->supportedStages
= scalar_stages
;
1050 properties
->supportedOperations
= VK_SUBGROUP_FEATURE_BASIC_BIT
|
1051 VK_SUBGROUP_FEATURE_VOTE_BIT
|
1052 VK_SUBGROUP_FEATURE_ARITHMETIC_BIT
|
1053 VK_SUBGROUP_FEATURE_BALLOT_BIT
|
1054 VK_SUBGROUP_FEATURE_SHUFFLE_BIT
|
1055 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT
|
1056 VK_SUBGROUP_FEATURE_CLUSTERED_BIT
|
1057 VK_SUBGROUP_FEATURE_QUAD_BIT
;
1058 properties
->quadOperationsInAllStages
= VK_TRUE
;
1063 anv_debug_ignored_stype(ext
->sType
);
1069 /* We support exactly one queue family. */
1070 static const VkQueueFamilyProperties
1071 anv_queue_family_properties
= {
1072 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
1073 VK_QUEUE_COMPUTE_BIT
|
1074 VK_QUEUE_TRANSFER_BIT
,
1076 .timestampValidBits
= 36, /* XXX: Real value here */
1077 .minImageTransferGranularity
= { 1, 1, 1 },
1080 void anv_GetPhysicalDeviceQueueFamilyProperties(
1081 VkPhysicalDevice physicalDevice
,
1083 VkQueueFamilyProperties
* pQueueFamilyProperties
)
1085 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pCount
);
1087 vk_outarray_append(&out
, p
) {
1088 *p
= anv_queue_family_properties
;
1092 void anv_GetPhysicalDeviceQueueFamilyProperties2(
1093 VkPhysicalDevice physicalDevice
,
1094 uint32_t* pQueueFamilyPropertyCount
,
1095 VkQueueFamilyProperties2
* pQueueFamilyProperties
)
1098 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pQueueFamilyPropertyCount
);
1100 vk_outarray_append(&out
, p
) {
1101 p
->queueFamilyProperties
= anv_queue_family_properties
;
1103 vk_foreach_struct(s
, p
->pNext
) {
1104 anv_debug_ignored_stype(s
->sType
);
1109 void anv_GetPhysicalDeviceMemoryProperties(
1110 VkPhysicalDevice physicalDevice
,
1111 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
1113 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
1115 pMemoryProperties
->memoryTypeCount
= physical_device
->memory
.type_count
;
1116 for (uint32_t i
= 0; i
< physical_device
->memory
.type_count
; i
++) {
1117 pMemoryProperties
->memoryTypes
[i
] = (VkMemoryType
) {
1118 .propertyFlags
= physical_device
->memory
.types
[i
].propertyFlags
,
1119 .heapIndex
= physical_device
->memory
.types
[i
].heapIndex
,
1123 pMemoryProperties
->memoryHeapCount
= physical_device
->memory
.heap_count
;
1124 for (uint32_t i
= 0; i
< physical_device
->memory
.heap_count
; i
++) {
1125 pMemoryProperties
->memoryHeaps
[i
] = (VkMemoryHeap
) {
1126 .size
= physical_device
->memory
.heaps
[i
].size
,
1127 .flags
= physical_device
->memory
.heaps
[i
].flags
,
1132 void anv_GetPhysicalDeviceMemoryProperties2(
1133 VkPhysicalDevice physicalDevice
,
1134 VkPhysicalDeviceMemoryProperties2
* pMemoryProperties
)
1136 anv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
1137 &pMemoryProperties
->memoryProperties
);
1139 vk_foreach_struct(ext
, pMemoryProperties
->pNext
) {
1140 switch (ext
->sType
) {
1142 anv_debug_ignored_stype(ext
->sType
);
1149 anv_GetDeviceGroupPeerMemoryFeatures(
1152 uint32_t localDeviceIndex
,
1153 uint32_t remoteDeviceIndex
,
1154 VkPeerMemoryFeatureFlags
* pPeerMemoryFeatures
)
1156 assert(localDeviceIndex
== 0 && remoteDeviceIndex
== 0);
1157 *pPeerMemoryFeatures
= VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT
|
1158 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT
|
1159 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT
|
1160 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT
;
1163 PFN_vkVoidFunction
anv_GetInstanceProcAddr(
1164 VkInstance _instance
,
1167 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1169 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
1170 * when we have to return valid function pointers, NULL, or it's left
1171 * undefined. See the table for exact details.
1176 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
1177 if (strcmp(pName, "vk" #entrypoint) == 0) \
1178 return (PFN_vkVoidFunction)anv_##entrypoint
1180 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties
);
1181 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties
);
1182 LOOKUP_ANV_ENTRYPOINT(CreateInstance
);
1184 #undef LOOKUP_ANV_ENTRYPOINT
1186 if (instance
== NULL
)
1189 int idx
= anv_get_entrypoint_index(pName
);
1193 return instance
->dispatch
.entrypoints
[idx
];
1196 /* With version 1+ of the loader interface the ICD should expose
1197 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
1200 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1201 VkInstance instance
,
1205 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1206 VkInstance instance
,
1209 return anv_GetInstanceProcAddr(instance
, pName
);
1212 PFN_vkVoidFunction
anv_GetDeviceProcAddr(
1216 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1218 if (!device
|| !pName
)
1221 int idx
= anv_get_entrypoint_index(pName
);
1225 return device
->dispatch
.entrypoints
[idx
];
1229 anv_CreateDebugReportCallbackEXT(VkInstance _instance
,
1230 const VkDebugReportCallbackCreateInfoEXT
* pCreateInfo
,
1231 const VkAllocationCallbacks
* pAllocator
,
1232 VkDebugReportCallbackEXT
* pCallback
)
1234 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1235 return vk_create_debug_report_callback(&instance
->debug_report_callbacks
,
1236 pCreateInfo
, pAllocator
, &instance
->alloc
,
1241 anv_DestroyDebugReportCallbackEXT(VkInstance _instance
,
1242 VkDebugReportCallbackEXT _callback
,
1243 const VkAllocationCallbacks
* pAllocator
)
1245 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1246 vk_destroy_debug_report_callback(&instance
->debug_report_callbacks
,
1247 _callback
, pAllocator
, &instance
->alloc
);
1251 anv_DebugReportMessageEXT(VkInstance _instance
,
1252 VkDebugReportFlagsEXT flags
,
1253 VkDebugReportObjectTypeEXT objectType
,
1256 int32_t messageCode
,
1257 const char* pLayerPrefix
,
1258 const char* pMessage
)
1260 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1261 vk_debug_report(&instance
->debug_report_callbacks
, flags
, objectType
,
1262 object
, location
, messageCode
, pLayerPrefix
, pMessage
);
1266 anv_queue_init(struct anv_device
*device
, struct anv_queue
*queue
)
1268 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1269 queue
->device
= device
;
1270 queue
->pool
= &device
->surface_state_pool
;
1275 anv_queue_finish(struct anv_queue
*queue
)
1279 static struct anv_state
1280 anv_state_pool_emit_data(struct anv_state_pool
*pool
, size_t size
, size_t align
, const void *p
)
1282 struct anv_state state
;
1284 state
= anv_state_pool_alloc(pool
, size
, align
);
1285 memcpy(state
.map
, p
, size
);
1287 anv_state_flush(pool
->block_pool
.device
, state
);
1292 struct gen8_border_color
{
1297 /* Pad out to 64 bytes */
1302 anv_device_init_border_colors(struct anv_device
*device
)
1304 static const struct gen8_border_color border_colors
[] = {
1305 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
1306 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
1307 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
1308 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
1309 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
1310 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
1313 device
->border_colors
= anv_state_pool_emit_data(&device
->dynamic_state_pool
,
1314 sizeof(border_colors
), 64,
1319 anv_device_init_trivial_batch(struct anv_device
*device
)
1321 anv_bo_init_new(&device
->trivial_batch_bo
, device
, 4096);
1323 if (device
->instance
->physicalDevice
.has_exec_async
)
1324 device
->trivial_batch_bo
.flags
|= EXEC_OBJECT_ASYNC
;
1326 void *map
= anv_gem_mmap(device
, device
->trivial_batch_bo
.gem_handle
,
1329 struct anv_batch batch
= {
1335 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
1336 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
1338 if (!device
->info
.has_llc
)
1339 gen_clflush_range(map
, batch
.next
- map
);
1341 anv_gem_munmap(map
, device
->trivial_batch_bo
.size
);
1344 VkResult
anv_EnumerateDeviceExtensionProperties(
1345 VkPhysicalDevice physicalDevice
,
1346 const char* pLayerName
,
1347 uint32_t* pPropertyCount
,
1348 VkExtensionProperties
* pProperties
)
1350 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
1351 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
1354 for (int i
= 0; i
< ANV_DEVICE_EXTENSION_COUNT
; i
++) {
1355 if (device
->supported_extensions
.extensions
[i
]) {
1356 vk_outarray_append(&out
, prop
) {
1357 *prop
= anv_device_extensions
[i
];
1362 return vk_outarray_status(&out
);
1366 anv_device_init_dispatch(struct anv_device
*device
)
1368 const struct anv_dispatch_table
*genX_table
;
1369 switch (device
->info
.gen
) {
1371 genX_table
= &gen10_dispatch_table
;
1374 genX_table
= &gen9_dispatch_table
;
1377 genX_table
= &gen8_dispatch_table
;
1380 if (device
->info
.is_haswell
)
1381 genX_table
= &gen75_dispatch_table
;
1383 genX_table
= &gen7_dispatch_table
;
1386 unreachable("unsupported gen\n");
1389 for (unsigned i
= 0; i
< ARRAY_SIZE(device
->dispatch
.entrypoints
); i
++) {
1390 /* Vulkan requires that entrypoints for extensions which have not been
1391 * enabled must not be advertised.
1393 if (!anv_entrypoint_is_enabled(i
, device
->instance
->apiVersion
,
1394 &device
->instance
->enabled_extensions
,
1395 &device
->enabled_extensions
)) {
1396 device
->dispatch
.entrypoints
[i
] = NULL
;
1397 } else if (genX_table
->entrypoints
[i
]) {
1398 device
->dispatch
.entrypoints
[i
] = genX_table
->entrypoints
[i
];
1400 device
->dispatch
.entrypoints
[i
] = anv_dispatch_table
.entrypoints
[i
];
1406 vk_priority_to_gen(int priority
)
1409 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT
:
1410 return GEN_CONTEXT_LOW_PRIORITY
;
1411 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
:
1412 return GEN_CONTEXT_MEDIUM_PRIORITY
;
1413 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT
:
1414 return GEN_CONTEXT_HIGH_PRIORITY
;
1415 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT
:
1416 return GEN_CONTEXT_REALTIME_PRIORITY
;
1418 unreachable("Invalid priority");
1422 VkResult
anv_CreateDevice(
1423 VkPhysicalDevice physicalDevice
,
1424 const VkDeviceCreateInfo
* pCreateInfo
,
1425 const VkAllocationCallbacks
* pAllocator
,
1428 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
1430 struct anv_device
*device
;
1432 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO
);
1434 struct anv_device_extension_table enabled_extensions
= { };
1435 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
1437 for (idx
= 0; idx
< ANV_DEVICE_EXTENSION_COUNT
; idx
++) {
1438 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
1439 anv_device_extensions
[idx
].extensionName
) == 0)
1443 if (idx
>= ANV_DEVICE_EXTENSION_COUNT
)
1444 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
1446 if (!physical_device
->supported_extensions
.extensions
[idx
])
1447 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
1449 enabled_extensions
.extensions
[idx
] = true;
1452 /* Check enabled features */
1453 if (pCreateInfo
->pEnabledFeatures
) {
1454 VkPhysicalDeviceFeatures supported_features
;
1455 anv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
1456 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
1457 VkBool32
*enabled_feature
= (VkBool32
*)pCreateInfo
->pEnabledFeatures
;
1458 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
1459 for (uint32_t i
= 0; i
< num_features
; i
++) {
1460 if (enabled_feature
[i
] && !supported_feature
[i
])
1461 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
1465 /* Check requested queues and fail if we are requested to create any
1466 * queues with flags we don't support.
1468 assert(pCreateInfo
->queueCreateInfoCount
> 0);
1469 for (uint32_t i
= 0; i
< pCreateInfo
->queueCreateInfoCount
; i
++) {
1470 if (pCreateInfo
->pQueueCreateInfos
[i
].flags
!= 0)
1471 return vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1474 /* Check if client specified queue priority. */
1475 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*queue_priority
=
1476 vk_find_struct_const(pCreateInfo
->pQueueCreateInfos
[0].pNext
,
1477 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT
);
1479 VkQueueGlobalPriorityEXT priority
=
1480 queue_priority
? queue_priority
->globalPriority
:
1481 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
;
1483 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
1485 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
1487 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1489 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1490 device
->instance
= physical_device
->instance
;
1491 device
->chipset_id
= physical_device
->chipset_id
;
1492 device
->no_hw
= physical_device
->no_hw
;
1493 device
->lost
= false;
1496 device
->alloc
= *pAllocator
;
1498 device
->alloc
= physical_device
->instance
->alloc
;
1500 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
1501 device
->fd
= open(physical_device
->path
, O_RDWR
| O_CLOEXEC
);
1502 if (device
->fd
== -1) {
1503 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1507 device
->context_id
= anv_gem_create_context(device
);
1508 if (device
->context_id
== -1) {
1509 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1513 /* As per spec, the driver implementation may deny requests to acquire
1514 * a priority above the default priority (MEDIUM) if the caller does not
1515 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
1518 if (physical_device
->has_context_priority
) {
1519 int err
= anv_gem_set_context_param(device
->fd
, device
->context_id
,
1520 I915_CONTEXT_PARAM_PRIORITY
,
1521 vk_priority_to_gen(priority
));
1522 if (err
!= 0 && priority
> VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
) {
1523 result
= vk_error(VK_ERROR_NOT_PERMITTED_EXT
);
1528 device
->info
= physical_device
->info
;
1529 device
->isl_dev
= physical_device
->isl_dev
;
1531 /* On Broadwell and later, we can use batch chaining to more efficiently
1532 * implement growing command buffers. Prior to Haswell, the kernel
1533 * command parser gets in the way and we have to fall back to growing
1536 device
->can_chain_batches
= device
->info
.gen
>= 8;
1538 device
->robust_buffer_access
= pCreateInfo
->pEnabledFeatures
&&
1539 pCreateInfo
->pEnabledFeatures
->robustBufferAccess
;
1540 device
->enabled_extensions
= enabled_extensions
;
1542 anv_device_init_dispatch(device
);
1544 if (pthread_mutex_init(&device
->mutex
, NULL
) != 0) {
1545 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1546 goto fail_context_id
;
1549 pthread_condattr_t condattr
;
1550 if (pthread_condattr_init(&condattr
) != 0) {
1551 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1554 if (pthread_condattr_setclock(&condattr
, CLOCK_MONOTONIC
) != 0) {
1555 pthread_condattr_destroy(&condattr
);
1556 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1559 if (pthread_cond_init(&device
->queue_submit
, NULL
) != 0) {
1560 pthread_condattr_destroy(&condattr
);
1561 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1564 pthread_condattr_destroy(&condattr
);
1567 (physical_device
->supports_48bit_addresses
? EXEC_OBJECT_SUPPORTS_48B_ADDRESS
: 0) |
1568 (physical_device
->has_exec_async
? EXEC_OBJECT_ASYNC
: 0) |
1569 (physical_device
->has_exec_capture
? EXEC_OBJECT_CAPTURE
: 0);
1571 anv_bo_pool_init(&device
->batch_bo_pool
, device
, bo_flags
);
1573 result
= anv_bo_cache_init(&device
->bo_cache
);
1574 if (result
!= VK_SUCCESS
)
1575 goto fail_batch_bo_pool
;
1577 /* For the state pools we explicitly disable 48bit. */
1578 bo_flags
= (physical_device
->has_exec_async
? EXEC_OBJECT_ASYNC
: 0) |
1579 (physical_device
->has_exec_capture
? EXEC_OBJECT_CAPTURE
: 0);
1581 result
= anv_state_pool_init(&device
->dynamic_state_pool
, device
, 16384,
1583 if (result
!= VK_SUCCESS
)
1586 result
= anv_state_pool_init(&device
->instruction_state_pool
, device
, 16384,
1588 if (result
!= VK_SUCCESS
)
1589 goto fail_dynamic_state_pool
;
1591 result
= anv_state_pool_init(&device
->surface_state_pool
, device
, 4096,
1593 if (result
!= VK_SUCCESS
)
1594 goto fail_instruction_state_pool
;
1596 result
= anv_bo_init_new(&device
->workaround_bo
, device
, 1024);
1597 if (result
!= VK_SUCCESS
)
1598 goto fail_surface_state_pool
;
1600 anv_device_init_trivial_batch(device
);
1602 anv_scratch_pool_init(device
, &device
->scratch_pool
);
1604 anv_queue_init(device
, &device
->queue
);
1606 switch (device
->info
.gen
) {
1608 if (!device
->info
.is_haswell
)
1609 result
= gen7_init_device_state(device
);
1611 result
= gen75_init_device_state(device
);
1614 result
= gen8_init_device_state(device
);
1617 result
= gen9_init_device_state(device
);
1620 result
= gen10_init_device_state(device
);
1623 result
= gen11_init_device_state(device
);
1626 /* Shouldn't get here as we don't create physical devices for any other
1628 unreachable("unhandled gen");
1630 if (result
!= VK_SUCCESS
)
1631 goto fail_workaround_bo
;
1633 anv_device_init_blorp(device
);
1635 anv_device_init_border_colors(device
);
1637 *pDevice
= anv_device_to_handle(device
);
1642 anv_queue_finish(&device
->queue
);
1643 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
1644 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
1645 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
1646 fail_surface_state_pool
:
1647 anv_state_pool_finish(&device
->surface_state_pool
);
1648 fail_instruction_state_pool
:
1649 anv_state_pool_finish(&device
->instruction_state_pool
);
1650 fail_dynamic_state_pool
:
1651 anv_state_pool_finish(&device
->dynamic_state_pool
);
1653 anv_bo_cache_finish(&device
->bo_cache
);
1655 anv_bo_pool_finish(&device
->batch_bo_pool
);
1656 pthread_cond_destroy(&device
->queue_submit
);
1658 pthread_mutex_destroy(&device
->mutex
);
1660 anv_gem_destroy_context(device
, device
->context_id
);
1664 vk_free(&device
->alloc
, device
);
1669 void anv_DestroyDevice(
1671 const VkAllocationCallbacks
* pAllocator
)
1673 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1678 anv_device_finish_blorp(device
);
1680 anv_queue_finish(&device
->queue
);
1682 #ifdef HAVE_VALGRIND
1683 /* We only need to free these to prevent valgrind errors. The backing
1684 * BO will go away in a couple of lines so we don't actually leak.
1686 anv_state_pool_free(&device
->dynamic_state_pool
, device
->border_colors
);
1689 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
1691 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
1692 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
1694 anv_gem_close(device
, device
->trivial_batch_bo
.gem_handle
);
1696 anv_state_pool_finish(&device
->surface_state_pool
);
1697 anv_state_pool_finish(&device
->instruction_state_pool
);
1698 anv_state_pool_finish(&device
->dynamic_state_pool
);
1700 anv_bo_cache_finish(&device
->bo_cache
);
1702 anv_bo_pool_finish(&device
->batch_bo_pool
);
1704 pthread_cond_destroy(&device
->queue_submit
);
1705 pthread_mutex_destroy(&device
->mutex
);
1707 anv_gem_destroy_context(device
, device
->context_id
);
1711 vk_free(&device
->alloc
, device
);
1714 VkResult
anv_EnumerateInstanceLayerProperties(
1715 uint32_t* pPropertyCount
,
1716 VkLayerProperties
* pProperties
)
1718 if (pProperties
== NULL
) {
1719 *pPropertyCount
= 0;
1723 /* None supported at this time */
1724 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1727 VkResult
anv_EnumerateDeviceLayerProperties(
1728 VkPhysicalDevice physicalDevice
,
1729 uint32_t* pPropertyCount
,
1730 VkLayerProperties
* pProperties
)
1732 if (pProperties
== NULL
) {
1733 *pPropertyCount
= 0;
1737 /* None supported at this time */
1738 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1741 void anv_GetDeviceQueue(
1743 uint32_t queueNodeIndex
,
1744 uint32_t queueIndex
,
1747 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1749 assert(queueIndex
== 0);
1751 *pQueue
= anv_queue_to_handle(&device
->queue
);
1754 void anv_GetDeviceQueue2(
1756 const VkDeviceQueueInfo2
* pQueueInfo
,
1759 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1761 assert(pQueueInfo
->queueIndex
== 0);
1763 if (pQueueInfo
->flags
== device
->queue
.flags
)
1764 *pQueue
= anv_queue_to_handle(&device
->queue
);
1770 anv_device_query_status(struct anv_device
*device
)
1772 /* This isn't likely as most of the callers of this function already check
1773 * for it. However, it doesn't hurt to check and it potentially lets us
1776 if (unlikely(device
->lost
))
1777 return VK_ERROR_DEVICE_LOST
;
1779 uint32_t active
, pending
;
1780 int ret
= anv_gem_gpu_get_reset_stats(device
, &active
, &pending
);
1782 /* We don't know the real error. */
1783 device
->lost
= true;
1784 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1785 "get_reset_stats failed: %m");
1789 device
->lost
= true;
1790 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1791 "GPU hung on one of our command buffers");
1792 } else if (pending
) {
1793 device
->lost
= true;
1794 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1795 "GPU hung with commands in-flight");
1802 anv_device_bo_busy(struct anv_device
*device
, struct anv_bo
*bo
)
1804 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
1805 * Other usages of the BO (such as on different hardware) will not be
1806 * flagged as "busy" by this ioctl. Use with care.
1808 int ret
= anv_gem_busy(device
, bo
->gem_handle
);
1810 return VK_NOT_READY
;
1811 } else if (ret
== -1) {
1812 /* We don't know the real error. */
1813 device
->lost
= true;
1814 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1815 "gem wait failed: %m");
1818 /* Query for device status after the busy call. If the BO we're checking
1819 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
1820 * client because it clearly doesn't have valid data. Yes, this most
1821 * likely means an ioctl, but we just did an ioctl to query the busy status
1822 * so it's no great loss.
1824 return anv_device_query_status(device
);
1828 anv_device_wait(struct anv_device
*device
, struct anv_bo
*bo
,
1831 int ret
= anv_gem_wait(device
, bo
->gem_handle
, &timeout
);
1832 if (ret
== -1 && errno
== ETIME
) {
1834 } else if (ret
== -1) {
1835 /* We don't know the real error. */
1836 device
->lost
= true;
1837 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1838 "gem wait failed: %m");
1841 /* Query for device status after the wait. If the BO we're waiting on got
1842 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
1843 * because it clearly doesn't have valid data. Yes, this most likely means
1844 * an ioctl, but we just did an ioctl to wait so it's no great loss.
1846 return anv_device_query_status(device
);
1849 VkResult
anv_DeviceWaitIdle(
1852 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1853 if (unlikely(device
->lost
))
1854 return VK_ERROR_DEVICE_LOST
;
1856 struct anv_batch batch
;
1859 batch
.start
= batch
.next
= cmds
;
1860 batch
.end
= (void *) cmds
+ sizeof(cmds
);
1862 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
1863 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
1865 return anv_device_submit_simple_batch(device
, &batch
);
1869 anv_bo_init_new(struct anv_bo
*bo
, struct anv_device
*device
, uint64_t size
)
1871 uint32_t gem_handle
= anv_gem_create(device
, size
);
1873 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
1875 anv_bo_init(bo
, gem_handle
, size
);
1880 VkResult
anv_AllocateMemory(
1882 const VkMemoryAllocateInfo
* pAllocateInfo
,
1883 const VkAllocationCallbacks
* pAllocator
,
1884 VkDeviceMemory
* pMem
)
1886 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1887 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
1888 struct anv_device_memory
*mem
;
1889 VkResult result
= VK_SUCCESS
;
1891 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
1893 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
1894 assert(pAllocateInfo
->allocationSize
> 0);
1896 if (pAllocateInfo
->allocationSize
> MAX_MEMORY_ALLOCATION_SIZE
)
1897 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
1899 /* FINISHME: Fail if allocation request exceeds heap size. */
1901 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
1902 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1904 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1906 assert(pAllocateInfo
->memoryTypeIndex
< pdevice
->memory
.type_count
);
1907 mem
->type
= &pdevice
->memory
.types
[pAllocateInfo
->memoryTypeIndex
];
1911 const VkImportMemoryFdInfoKHR
*fd_info
=
1912 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_FD_INFO_KHR
);
1914 /* The Vulkan spec permits handleType to be 0, in which case the struct is
1917 if (fd_info
&& fd_info
->handleType
) {
1918 /* At the moment, we support only the below handle types. */
1919 assert(fd_info
->handleType
==
1920 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
1921 fd_info
->handleType
==
1922 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
1924 result
= anv_bo_cache_import(device
, &device
->bo_cache
,
1925 fd_info
->fd
, &mem
->bo
);
1926 if (result
!= VK_SUCCESS
)
1929 VkDeviceSize aligned_alloc_size
=
1930 align_u64(pAllocateInfo
->allocationSize
, 4096);
1932 /* For security purposes, we reject importing the bo if it's smaller
1933 * than the requested allocation size. This prevents a malicious client
1934 * from passing a buffer to a trusted client, lying about the size, and
1935 * telling the trusted client to try and texture from an image that goes
1936 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
1937 * in the trusted client. The trusted client can protect itself against
1938 * this sort of attack but only if it can trust the buffer size.
1940 if (mem
->bo
->size
< aligned_alloc_size
) {
1941 result
= vk_errorf(device
->instance
, device
,
1942 VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
,
1943 "aligned allocationSize too large for "
1944 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR: "
1945 "%"PRIu64
"B > %"PRIu64
"B",
1946 aligned_alloc_size
, mem
->bo
->size
);
1947 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
1951 /* From the Vulkan spec:
1953 * "Importing memory from a file descriptor transfers ownership of
1954 * the file descriptor from the application to the Vulkan
1955 * implementation. The application must not perform any operations on
1956 * the file descriptor after a successful import."
1958 * If the import fails, we leave the file descriptor open.
1962 result
= anv_bo_cache_alloc(device
, &device
->bo_cache
,
1963 pAllocateInfo
->allocationSize
,
1965 if (result
!= VK_SUCCESS
)
1968 const VkMemoryDedicatedAllocateInfoKHR
*dedicated_info
=
1969 vk_find_struct_const(pAllocateInfo
->pNext
, MEMORY_DEDICATED_ALLOCATE_INFO_KHR
);
1970 if (dedicated_info
&& dedicated_info
->image
!= VK_NULL_HANDLE
) {
1971 ANV_FROM_HANDLE(anv_image
, image
, dedicated_info
->image
);
1973 /* Some legacy (non-modifiers) consumers need the tiling to be set on
1974 * the BO. In this case, we have a dedicated allocation.
1976 if (image
->needs_set_tiling
) {
1977 const uint32_t i915_tiling
=
1978 isl_tiling_to_i915_tiling(image
->planes
[0].surface
.isl
.tiling
);
1979 int ret
= anv_gem_set_tiling(device
, mem
->bo
->gem_handle
,
1980 image
->planes
[0].surface
.isl
.row_pitch
,
1983 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
1984 return vk_errorf(device
->instance
, NULL
,
1985 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
1986 "failed to set BO tiling: %m");
1992 assert(mem
->type
->heapIndex
< pdevice
->memory
.heap_count
);
1993 if (pdevice
->memory
.heaps
[mem
->type
->heapIndex
].supports_48bit_addresses
)
1994 mem
->bo
->flags
|= EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
1996 const struct wsi_memory_allocate_info
*wsi_info
=
1997 vk_find_struct_const(pAllocateInfo
->pNext
, WSI_MEMORY_ALLOCATE_INFO_MESA
);
1998 if (wsi_info
&& wsi_info
->implicit_sync
) {
1999 /* We need to set the WRITE flag on window system buffers so that GEM
2000 * will know we're writing to them and synchronize uses on other rings
2001 * (eg if the display server uses the blitter ring).
2003 mem
->bo
->flags
|= EXEC_OBJECT_WRITE
;
2004 } else if (pdevice
->has_exec_async
) {
2005 mem
->bo
->flags
|= EXEC_OBJECT_ASYNC
;
2008 *pMem
= anv_device_memory_to_handle(mem
);
2013 vk_free2(&device
->alloc
, pAllocator
, mem
);
2018 VkResult
anv_GetMemoryFdKHR(
2020 const VkMemoryGetFdInfoKHR
* pGetFdInfo
,
2023 ANV_FROM_HANDLE(anv_device
, dev
, device_h
);
2024 ANV_FROM_HANDLE(anv_device_memory
, mem
, pGetFdInfo
->memory
);
2026 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
2028 assert(pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
2029 pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
2031 return anv_bo_cache_export(dev
, &dev
->bo_cache
, mem
->bo
, pFd
);
2034 VkResult
anv_GetMemoryFdPropertiesKHR(
2036 VkExternalMemoryHandleTypeFlagBitsKHR handleType
,
2038 VkMemoryFdPropertiesKHR
* pMemoryFdProperties
)
2040 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2041 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2043 switch (handleType
) {
2044 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
:
2045 /* dma-buf can be imported as any memory type */
2046 pMemoryFdProperties
->memoryTypeBits
=
2047 (1 << pdevice
->memory
.type_count
) - 1;
2051 /* The valid usage section for this function says:
2053 * "handleType must not be one of the handle types defined as
2056 * So opaque handle types fall into the default "unsupported" case.
2058 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
2062 void anv_FreeMemory(
2064 VkDeviceMemory _mem
,
2065 const VkAllocationCallbacks
* pAllocator
)
2067 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2068 ANV_FROM_HANDLE(anv_device_memory
, mem
, _mem
);
2074 anv_UnmapMemory(_device
, _mem
);
2076 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
2078 vk_free2(&device
->alloc
, pAllocator
, mem
);
2081 VkResult
anv_MapMemory(
2083 VkDeviceMemory _memory
,
2084 VkDeviceSize offset
,
2086 VkMemoryMapFlags flags
,
2089 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2090 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
2097 if (size
== VK_WHOLE_SIZE
)
2098 size
= mem
->bo
->size
- offset
;
2100 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
2102 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
2103 * assert(size != 0);
2104 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
2105 * equal to the size of the memory minus offset
2108 assert(offset
+ size
<= mem
->bo
->size
);
2110 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
2111 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
2112 * at a time is valid. We could just mmap up front and return an offset
2113 * pointer here, but that may exhaust virtual memory on 32 bit
2116 uint32_t gem_flags
= 0;
2118 if (!device
->info
.has_llc
&&
2119 (mem
->type
->propertyFlags
& VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
))
2120 gem_flags
|= I915_MMAP_WC
;
2122 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
2123 uint64_t map_offset
= offset
& ~4095ull;
2124 assert(offset
>= map_offset
);
2125 uint64_t map_size
= (offset
+ size
) - map_offset
;
2127 /* Let's map whole pages */
2128 map_size
= align_u64(map_size
, 4096);
2130 void *map
= anv_gem_mmap(device
, mem
->bo
->gem_handle
,
2131 map_offset
, map_size
, gem_flags
);
2132 if (map
== MAP_FAILED
)
2133 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
2136 mem
->map_size
= map_size
;
2138 *ppData
= mem
->map
+ (offset
- map_offset
);
2143 void anv_UnmapMemory(
2145 VkDeviceMemory _memory
)
2147 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
2152 anv_gem_munmap(mem
->map
, mem
->map_size
);
2159 clflush_mapped_ranges(struct anv_device
*device
,
2161 const VkMappedMemoryRange
*ranges
)
2163 for (uint32_t i
= 0; i
< count
; i
++) {
2164 ANV_FROM_HANDLE(anv_device_memory
, mem
, ranges
[i
].memory
);
2165 if (ranges
[i
].offset
>= mem
->map_size
)
2168 gen_clflush_range(mem
->map
+ ranges
[i
].offset
,
2169 MIN2(ranges
[i
].size
, mem
->map_size
- ranges
[i
].offset
));
2173 VkResult
anv_FlushMappedMemoryRanges(
2175 uint32_t memoryRangeCount
,
2176 const VkMappedMemoryRange
* pMemoryRanges
)
2178 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2180 if (device
->info
.has_llc
)
2183 /* Make sure the writes we're flushing have landed. */
2184 __builtin_ia32_mfence();
2186 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
2191 VkResult
anv_InvalidateMappedMemoryRanges(
2193 uint32_t memoryRangeCount
,
2194 const VkMappedMemoryRange
* pMemoryRanges
)
2196 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2198 if (device
->info
.has_llc
)
2201 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
2203 /* Make sure no reads get moved up above the invalidate. */
2204 __builtin_ia32_mfence();
2209 void anv_GetBufferMemoryRequirements(
2212 VkMemoryRequirements
* pMemoryRequirements
)
2214 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
2215 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2216 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2218 /* The Vulkan spec (git aaed022) says:
2220 * memoryTypeBits is a bitfield and contains one bit set for every
2221 * supported memory type for the resource. The bit `1<<i` is set if and
2222 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
2223 * structure for the physical device is supported.
2225 uint32_t memory_types
= 0;
2226 for (uint32_t i
= 0; i
< pdevice
->memory
.type_count
; i
++) {
2227 uint32_t valid_usage
= pdevice
->memory
.types
[i
].valid_buffer_usage
;
2228 if ((valid_usage
& buffer
->usage
) == buffer
->usage
)
2229 memory_types
|= (1u << i
);
2232 /* Base alignment requirement of a cache line */
2233 uint32_t alignment
= 16;
2235 /* We need an alignment of 32 for pushing UBOs */
2236 if (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
)
2237 alignment
= MAX2(alignment
, 32);
2239 pMemoryRequirements
->size
= buffer
->size
;
2240 pMemoryRequirements
->alignment
= alignment
;
2242 /* Storage and Uniform buffers should have their size aligned to
2243 * 32-bits to avoid boundary checks when last DWord is not complete.
2244 * This would ensure that not internal padding would be needed for
2247 if (device
->robust_buffer_access
&&
2248 (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
||
2249 buffer
->usage
& VK_BUFFER_USAGE_STORAGE_BUFFER_BIT
))
2250 pMemoryRequirements
->size
= align_u64(buffer
->size
, 4);
2252 pMemoryRequirements
->memoryTypeBits
= memory_types
;
2255 void anv_GetBufferMemoryRequirements2(
2257 const VkBufferMemoryRequirementsInfo2
* pInfo
,
2258 VkMemoryRequirements2
* pMemoryRequirements
)
2260 anv_GetBufferMemoryRequirements(_device
, pInfo
->buffer
,
2261 &pMemoryRequirements
->memoryRequirements
);
2263 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
2264 switch (ext
->sType
) {
2265 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
2266 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
2267 requirements
->prefersDedicatedAllocation
= VK_FALSE
;
2268 requirements
->requiresDedicatedAllocation
= VK_FALSE
;
2273 anv_debug_ignored_stype(ext
->sType
);
2279 void anv_GetImageMemoryRequirements(
2282 VkMemoryRequirements
* pMemoryRequirements
)
2284 ANV_FROM_HANDLE(anv_image
, image
, _image
);
2285 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2286 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2288 /* The Vulkan spec (git aaed022) says:
2290 * memoryTypeBits is a bitfield and contains one bit set for every
2291 * supported memory type for the resource. The bit `1<<i` is set if and
2292 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
2293 * structure for the physical device is supported.
2295 * All types are currently supported for images.
2297 uint32_t memory_types
= (1ull << pdevice
->memory
.type_count
) - 1;
2299 pMemoryRequirements
->size
= image
->size
;
2300 pMemoryRequirements
->alignment
= image
->alignment
;
2301 pMemoryRequirements
->memoryTypeBits
= memory_types
;
2304 void anv_GetImageMemoryRequirements2(
2306 const VkImageMemoryRequirementsInfo2
* pInfo
,
2307 VkMemoryRequirements2
* pMemoryRequirements
)
2309 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2310 ANV_FROM_HANDLE(anv_image
, image
, pInfo
->image
);
2312 anv_GetImageMemoryRequirements(_device
, pInfo
->image
,
2313 &pMemoryRequirements
->memoryRequirements
);
2315 vk_foreach_struct_const(ext
, pInfo
->pNext
) {
2316 switch (ext
->sType
) {
2317 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO
: {
2318 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2319 const VkImagePlaneMemoryRequirementsInfoKHR
*plane_reqs
=
2320 (const VkImagePlaneMemoryRequirementsInfoKHR
*) ext
;
2321 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
,
2322 plane_reqs
->planeAspect
);
2324 assert(image
->planes
[plane
].offset
== 0);
2326 /* The Vulkan spec (git aaed022) says:
2328 * memoryTypeBits is a bitfield and contains one bit set for every
2329 * supported memory type for the resource. The bit `1<<i` is set
2330 * if and only if the memory type `i` in the
2331 * VkPhysicalDeviceMemoryProperties structure for the physical
2332 * device is supported.
2334 * All types are currently supported for images.
2336 pMemoryRequirements
->memoryRequirements
.memoryTypeBits
=
2337 (1ull << pdevice
->memory
.type_count
) - 1;
2339 pMemoryRequirements
->memoryRequirements
.size
= image
->planes
[plane
].size
;
2340 pMemoryRequirements
->memoryRequirements
.alignment
=
2341 image
->planes
[plane
].alignment
;
2346 anv_debug_ignored_stype(ext
->sType
);
2351 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
2352 switch (ext
->sType
) {
2353 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
2354 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
2355 if (image
->needs_set_tiling
) {
2356 /* If we need to set the tiling for external consumers, we need a
2357 * dedicated allocation.
2359 * See also anv_AllocateMemory.
2361 requirements
->prefersDedicatedAllocation
= VK_TRUE
;
2362 requirements
->requiresDedicatedAllocation
= VK_TRUE
;
2364 requirements
->prefersDedicatedAllocation
= VK_FALSE
;
2365 requirements
->requiresDedicatedAllocation
= VK_FALSE
;
2371 anv_debug_ignored_stype(ext
->sType
);
2377 void anv_GetImageSparseMemoryRequirements(
2380 uint32_t* pSparseMemoryRequirementCount
,
2381 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
2383 *pSparseMemoryRequirementCount
= 0;
2386 void anv_GetImageSparseMemoryRequirements2(
2388 const VkImageSparseMemoryRequirementsInfo2
* pInfo
,
2389 uint32_t* pSparseMemoryRequirementCount
,
2390 VkSparseImageMemoryRequirements2
* pSparseMemoryRequirements
)
2392 *pSparseMemoryRequirementCount
= 0;
2395 void anv_GetDeviceMemoryCommitment(
2397 VkDeviceMemory memory
,
2398 VkDeviceSize
* pCommittedMemoryInBytes
)
2400 *pCommittedMemoryInBytes
= 0;
2404 anv_bind_buffer_memory(const VkBindBufferMemoryInfo
*pBindInfo
)
2406 ANV_FROM_HANDLE(anv_device_memory
, mem
, pBindInfo
->memory
);
2407 ANV_FROM_HANDLE(anv_buffer
, buffer
, pBindInfo
->buffer
);
2409 assert(pBindInfo
->sType
== VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
);
2412 assert((buffer
->usage
& mem
->type
->valid_buffer_usage
) == buffer
->usage
);
2413 buffer
->bo
= mem
->bo
;
2414 buffer
->offset
= pBindInfo
->memoryOffset
;
2421 VkResult
anv_BindBufferMemory(
2424 VkDeviceMemory memory
,
2425 VkDeviceSize memoryOffset
)
2427 anv_bind_buffer_memory(
2428 &(VkBindBufferMemoryInfo
) {
2429 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
,
2432 .memoryOffset
= memoryOffset
,
2438 VkResult
anv_BindBufferMemory2(
2440 uint32_t bindInfoCount
,
2441 const VkBindBufferMemoryInfo
* pBindInfos
)
2443 for (uint32_t i
= 0; i
< bindInfoCount
; i
++)
2444 anv_bind_buffer_memory(&pBindInfos
[i
]);
2449 VkResult
anv_QueueBindSparse(
2451 uint32_t bindInfoCount
,
2452 const VkBindSparseInfo
* pBindInfo
,
2455 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
2456 if (unlikely(queue
->device
->lost
))
2457 return VK_ERROR_DEVICE_LOST
;
2459 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
2464 VkResult
anv_CreateEvent(
2466 const VkEventCreateInfo
* pCreateInfo
,
2467 const VkAllocationCallbacks
* pAllocator
,
2470 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2471 struct anv_state state
;
2472 struct anv_event
*event
;
2474 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_EVENT_CREATE_INFO
);
2476 state
= anv_state_pool_alloc(&device
->dynamic_state_pool
,
2479 event
->state
= state
;
2480 event
->semaphore
= VK_EVENT_RESET
;
2482 if (!device
->info
.has_llc
) {
2483 /* Make sure the writes we're flushing have landed. */
2484 __builtin_ia32_mfence();
2485 __builtin_ia32_clflush(event
);
2488 *pEvent
= anv_event_to_handle(event
);
2493 void anv_DestroyEvent(
2496 const VkAllocationCallbacks
* pAllocator
)
2498 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2499 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2504 anv_state_pool_free(&device
->dynamic_state_pool
, event
->state
);
2507 VkResult
anv_GetEventStatus(
2511 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2512 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2514 if (unlikely(device
->lost
))
2515 return VK_ERROR_DEVICE_LOST
;
2517 if (!device
->info
.has_llc
) {
2518 /* Invalidate read cache before reading event written by GPU. */
2519 __builtin_ia32_clflush(event
);
2520 __builtin_ia32_mfence();
2524 return event
->semaphore
;
2527 VkResult
anv_SetEvent(
2531 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2532 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2534 event
->semaphore
= VK_EVENT_SET
;
2536 if (!device
->info
.has_llc
) {
2537 /* Make sure the writes we're flushing have landed. */
2538 __builtin_ia32_mfence();
2539 __builtin_ia32_clflush(event
);
2545 VkResult
anv_ResetEvent(
2549 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2550 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2552 event
->semaphore
= VK_EVENT_RESET
;
2554 if (!device
->info
.has_llc
) {
2555 /* Make sure the writes we're flushing have landed. */
2556 __builtin_ia32_mfence();
2557 __builtin_ia32_clflush(event
);
2565 VkResult
anv_CreateBuffer(
2567 const VkBufferCreateInfo
* pCreateInfo
,
2568 const VkAllocationCallbacks
* pAllocator
,
2571 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2572 struct anv_buffer
*buffer
;
2574 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
2576 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
2577 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2579 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2581 buffer
->size
= pCreateInfo
->size
;
2582 buffer
->usage
= pCreateInfo
->usage
;
2586 *pBuffer
= anv_buffer_to_handle(buffer
);
2591 void anv_DestroyBuffer(
2594 const VkAllocationCallbacks
* pAllocator
)
2596 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2597 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
2602 vk_free2(&device
->alloc
, pAllocator
, buffer
);
2606 anv_fill_buffer_surface_state(struct anv_device
*device
, struct anv_state state
,
2607 enum isl_format format
,
2608 uint32_t offset
, uint32_t range
, uint32_t stride
)
2610 isl_buffer_fill_state(&device
->isl_dev
, state
.map
,
2612 .mocs
= device
->default_mocs
,
2617 anv_state_flush(device
, state
);
2620 void anv_DestroySampler(
2623 const VkAllocationCallbacks
* pAllocator
)
2625 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2626 ANV_FROM_HANDLE(anv_sampler
, sampler
, _sampler
);
2631 vk_free2(&device
->alloc
, pAllocator
, sampler
);
2634 VkResult
anv_CreateFramebuffer(
2636 const VkFramebufferCreateInfo
* pCreateInfo
,
2637 const VkAllocationCallbacks
* pAllocator
,
2638 VkFramebuffer
* pFramebuffer
)
2640 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2641 struct anv_framebuffer
*framebuffer
;
2643 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
2645 size_t size
= sizeof(*framebuffer
) +
2646 sizeof(struct anv_image_view
*) * pCreateInfo
->attachmentCount
;
2647 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
2648 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2649 if (framebuffer
== NULL
)
2650 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2652 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
2653 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
2654 VkImageView _iview
= pCreateInfo
->pAttachments
[i
];
2655 framebuffer
->attachments
[i
] = anv_image_view_from_handle(_iview
);
2658 framebuffer
->width
= pCreateInfo
->width
;
2659 framebuffer
->height
= pCreateInfo
->height
;
2660 framebuffer
->layers
= pCreateInfo
->layers
;
2662 *pFramebuffer
= anv_framebuffer_to_handle(framebuffer
);
2667 void anv_DestroyFramebuffer(
2670 const VkAllocationCallbacks
* pAllocator
)
2672 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2673 ANV_FROM_HANDLE(anv_framebuffer
, fb
, _fb
);
2678 vk_free2(&device
->alloc
, pAllocator
, fb
);
2681 /* vk_icd.h does not declare this function, so we declare it here to
2682 * suppress Wmissing-prototypes.
2684 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2685 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
);
2687 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2688 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
)
2690 /* For the full details on loader interface versioning, see
2691 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
2692 * What follows is a condensed summary, to help you navigate the large and
2693 * confusing official doc.
2695 * - Loader interface v0 is incompatible with later versions. We don't
2698 * - In loader interface v1:
2699 * - The first ICD entrypoint called by the loader is
2700 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
2702 * - The ICD must statically expose no other Vulkan symbol unless it is
2703 * linked with -Bsymbolic.
2704 * - Each dispatchable Vulkan handle created by the ICD must be
2705 * a pointer to a struct whose first member is VK_LOADER_DATA. The
2706 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
2707 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
2708 * vkDestroySurfaceKHR(). The ICD must be capable of working with
2709 * such loader-managed surfaces.
2711 * - Loader interface v2 differs from v1 in:
2712 * - The first ICD entrypoint called by the loader is
2713 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
2714 * statically expose this entrypoint.
2716 * - Loader interface v3 differs from v2 in:
2717 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
2718 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
2719 * because the loader no longer does so.
2721 *pSupportedVersion
= MIN2(*pSupportedVersion
, 3u);