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
);
113 uint64_t heap_size
= 0;
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 */
326 } else if (device
->info
.gen
== 11) {
327 intel_logw("Vulkan is not yet fully supported on gen11.");
329 result
= vk_errorf(device
->instance
, device
,
330 VK_ERROR_INCOMPATIBLE_DRIVER
,
331 "Vulkan not yet supported on %s", device
->name
);
335 device
->cmd_parser_version
= -1;
336 if (device
->info
.gen
== 7) {
337 device
->cmd_parser_version
=
338 anv_gem_get_param(fd
, I915_PARAM_CMD_PARSER_VERSION
);
339 if (device
->cmd_parser_version
== -1) {
340 result
= vk_errorf(device
->instance
, device
,
341 VK_ERROR_INITIALIZATION_FAILED
,
342 "failed to get command parser version");
347 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_WAIT_TIMEOUT
)) {
348 result
= vk_errorf(device
->instance
, device
,
349 VK_ERROR_INITIALIZATION_FAILED
,
350 "kernel missing gem wait");
354 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_EXECBUF2
)) {
355 result
= vk_errorf(device
->instance
, device
,
356 VK_ERROR_INITIALIZATION_FAILED
,
357 "kernel missing execbuf2");
361 if (!device
->info
.has_llc
&&
362 anv_gem_get_param(fd
, I915_PARAM_MMAP_VERSION
) < 1) {
363 result
= vk_errorf(device
->instance
, device
,
364 VK_ERROR_INITIALIZATION_FAILED
,
365 "kernel missing wc mmap");
369 result
= anv_physical_device_init_heaps(device
, fd
);
370 if (result
!= VK_SUCCESS
)
373 device
->has_exec_async
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_ASYNC
);
374 device
->has_exec_capture
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_CAPTURE
);
375 device
->has_exec_fence
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE
);
376 device
->has_syncobj
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE_ARRAY
);
377 device
->has_syncobj_wait
= device
->has_syncobj
&&
378 anv_gem_supports_syncobj_wait(fd
);
379 device
->has_context_priority
= anv_gem_has_context_priority(fd
);
381 bool swizzled
= anv_gem_get_bit6_swizzle(fd
, I915_TILING_X
);
383 /* Starting with Gen10, the timestamp frequency of the command streamer may
384 * vary from one part to another. We can query the value from the kernel.
386 if (device
->info
.gen
>= 10) {
387 int timestamp_frequency
=
388 anv_gem_get_param(fd
, I915_PARAM_CS_TIMESTAMP_FREQUENCY
);
390 if (timestamp_frequency
< 0)
391 intel_logw("Kernel 4.16-rc1+ required to properly query CS timestamp frequency");
393 device
->info
.timestamp_frequency
= timestamp_frequency
;
396 /* GENs prior to 8 do not support EU/Subslice info */
397 if (device
->info
.gen
>= 8) {
398 device
->subslice_total
= anv_gem_get_param(fd
, I915_PARAM_SUBSLICE_TOTAL
);
399 device
->eu_total
= anv_gem_get_param(fd
, I915_PARAM_EU_TOTAL
);
401 /* Without this information, we cannot get the right Braswell
402 * brandstrings, and we have to use conservative numbers for GPGPU on
403 * many platforms, but otherwise, things will just work.
405 if (device
->subslice_total
< 1 || device
->eu_total
< 1) {
406 intel_logw("Kernel 4.1 required to properly query GPU properties");
408 } else if (device
->info
.gen
== 7) {
409 device
->subslice_total
= 1 << (device
->info
.gt
- 1);
412 if (device
->info
.is_cherryview
&&
413 device
->subslice_total
> 0 && device
->eu_total
> 0) {
414 /* Logical CS threads = EUs per subslice * num threads per EU */
415 uint32_t max_cs_threads
=
416 device
->eu_total
/ device
->subslice_total
* device
->info
.num_thread_per_eu
;
418 /* Fuse configurations may give more threads than expected, never less. */
419 if (max_cs_threads
> device
->info
.max_cs_threads
)
420 device
->info
.max_cs_threads
= max_cs_threads
;
423 device
->compiler
= brw_compiler_create(NULL
, &device
->info
);
424 if (device
->compiler
== NULL
) {
425 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
428 device
->compiler
->shader_debug_log
= compiler_debug_log
;
429 device
->compiler
->shader_perf_log
= compiler_perf_log
;
430 device
->compiler
->supports_pull_constants
= false;
431 device
->compiler
->constant_buffer_0_is_relative
= true;
433 isl_device_init(&device
->isl_dev
, &device
->info
, swizzled
);
435 result
= anv_physical_device_init_uuids(device
);
436 if (result
!= VK_SUCCESS
)
439 result
= anv_init_wsi(device
);
440 if (result
!= VK_SUCCESS
) {
441 ralloc_free(device
->compiler
);
445 anv_physical_device_get_supported_extensions(device
,
446 &device
->supported_extensions
);
448 device
->local_fd
= fd
;
457 anv_physical_device_finish(struct anv_physical_device
*device
)
459 anv_finish_wsi(device
);
460 ralloc_free(device
->compiler
);
461 close(device
->local_fd
);
465 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
466 VkSystemAllocationScope allocationScope
)
472 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
473 size_t align
, VkSystemAllocationScope allocationScope
)
475 return realloc(pOriginal
, size
);
479 default_free_func(void *pUserData
, void *pMemory
)
484 static const VkAllocationCallbacks default_alloc
= {
486 .pfnAllocation
= default_alloc_func
,
487 .pfnReallocation
= default_realloc_func
,
488 .pfnFree
= default_free_func
,
491 VkResult
anv_EnumerateInstanceExtensionProperties(
492 const char* pLayerName
,
493 uint32_t* pPropertyCount
,
494 VkExtensionProperties
* pProperties
)
496 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
498 for (int i
= 0; i
< ANV_INSTANCE_EXTENSION_COUNT
; i
++) {
499 if (anv_instance_extensions_supported
.extensions
[i
]) {
500 vk_outarray_append(&out
, prop
) {
501 *prop
= anv_instance_extensions
[i
];
506 return vk_outarray_status(&out
);
509 VkResult
anv_CreateInstance(
510 const VkInstanceCreateInfo
* pCreateInfo
,
511 const VkAllocationCallbacks
* pAllocator
,
512 VkInstance
* pInstance
)
514 struct anv_instance
*instance
;
517 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
519 struct anv_instance_extension_table enabled_extensions
= {};
520 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
522 for (idx
= 0; idx
< ANV_INSTANCE_EXTENSION_COUNT
; idx
++) {
523 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
524 anv_instance_extensions
[idx
].extensionName
) == 0)
528 if (idx
>= ANV_INSTANCE_EXTENSION_COUNT
)
529 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
531 if (!anv_instance_extensions_supported
.extensions
[idx
])
532 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
534 enabled_extensions
.extensions
[idx
] = true;
537 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
538 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
540 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
542 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
545 instance
->alloc
= *pAllocator
;
547 instance
->alloc
= default_alloc
;
549 if (pCreateInfo
->pApplicationInfo
&&
550 pCreateInfo
->pApplicationInfo
->apiVersion
!= 0) {
551 instance
->apiVersion
= pCreateInfo
->pApplicationInfo
->apiVersion
;
553 anv_EnumerateInstanceVersion(&instance
->apiVersion
);
556 instance
->enabled_extensions
= enabled_extensions
;
558 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->dispatch
.entrypoints
); i
++) {
559 /* Vulkan requires that entrypoints for extensions which have not been
560 * enabled must not be advertised.
562 if (!anv_entrypoint_is_enabled(i
, instance
->apiVersion
,
563 &instance
->enabled_extensions
, NULL
)) {
564 instance
->dispatch
.entrypoints
[i
] = NULL
;
565 } else if (anv_dispatch_table
.entrypoints
[i
] != NULL
) {
566 instance
->dispatch
.entrypoints
[i
] = anv_dispatch_table
.entrypoints
[i
];
568 instance
->dispatch
.entrypoints
[i
] =
569 anv_tramp_dispatch_table
.entrypoints
[i
];
573 instance
->physicalDeviceCount
= -1;
575 result
= vk_debug_report_instance_init(&instance
->debug_report_callbacks
);
576 if (result
!= VK_SUCCESS
) {
577 vk_free2(&default_alloc
, pAllocator
, instance
);
578 return vk_error(result
);
583 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
585 *pInstance
= anv_instance_to_handle(instance
);
590 void anv_DestroyInstance(
591 VkInstance _instance
,
592 const VkAllocationCallbacks
* pAllocator
)
594 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
599 if (instance
->physicalDeviceCount
> 0) {
600 /* We support at most one physical device. */
601 assert(instance
->physicalDeviceCount
== 1);
602 anv_physical_device_finish(&instance
->physicalDevice
);
605 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
607 vk_debug_report_instance_destroy(&instance
->debug_report_callbacks
);
611 vk_free(&instance
->alloc
, instance
);
615 anv_enumerate_devices(struct anv_instance
*instance
)
617 /* TODO: Check for more devices ? */
618 drmDevicePtr devices
[8];
619 VkResult result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
622 instance
->physicalDeviceCount
= 0;
624 max_devices
= drmGetDevices2(0, devices
, ARRAY_SIZE(devices
));
626 return VK_ERROR_INCOMPATIBLE_DRIVER
;
628 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
629 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
630 devices
[i
]->bustype
== DRM_BUS_PCI
&&
631 devices
[i
]->deviceinfo
.pci
->vendor_id
== 0x8086) {
633 result
= anv_physical_device_init(&instance
->physicalDevice
,
635 devices
[i
]->nodes
[DRM_NODE_RENDER
]);
636 if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
640 drmFreeDevices(devices
, max_devices
);
642 if (result
== VK_SUCCESS
)
643 instance
->physicalDeviceCount
= 1;
649 anv_instance_ensure_physical_device(struct anv_instance
*instance
)
651 if (instance
->physicalDeviceCount
< 0) {
652 VkResult result
= anv_enumerate_devices(instance
);
653 if (result
!= VK_SUCCESS
&&
654 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
661 VkResult
anv_EnumeratePhysicalDevices(
662 VkInstance _instance
,
663 uint32_t* pPhysicalDeviceCount
,
664 VkPhysicalDevice
* pPhysicalDevices
)
666 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
667 VK_OUTARRAY_MAKE(out
, pPhysicalDevices
, pPhysicalDeviceCount
);
669 VkResult result
= anv_instance_ensure_physical_device(instance
);
670 if (result
!= VK_SUCCESS
)
673 if (instance
->physicalDeviceCount
== 0)
676 assert(instance
->physicalDeviceCount
== 1);
677 vk_outarray_append(&out
, i
) {
678 *i
= anv_physical_device_to_handle(&instance
->physicalDevice
);
681 return vk_outarray_status(&out
);
684 VkResult
anv_EnumeratePhysicalDeviceGroups(
685 VkInstance _instance
,
686 uint32_t* pPhysicalDeviceGroupCount
,
687 VkPhysicalDeviceGroupProperties
* pPhysicalDeviceGroupProperties
)
689 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
690 VK_OUTARRAY_MAKE(out
, pPhysicalDeviceGroupProperties
,
691 pPhysicalDeviceGroupCount
);
693 VkResult result
= anv_instance_ensure_physical_device(instance
);
694 if (result
!= VK_SUCCESS
)
697 if (instance
->physicalDeviceCount
== 0)
700 assert(instance
->physicalDeviceCount
== 1);
702 vk_outarray_append(&out
, p
) {
703 p
->physicalDeviceCount
= 1;
704 memset(p
->physicalDevices
, 0, sizeof(p
->physicalDevices
));
705 p
->physicalDevices
[0] =
706 anv_physical_device_to_handle(&instance
->physicalDevice
);
707 p
->subsetAllocation
= VK_FALSE
;
709 vk_foreach_struct(ext
, p
->pNext
)
710 anv_debug_ignored_stype(ext
->sType
);
713 return vk_outarray_status(&out
);
716 void anv_GetPhysicalDeviceFeatures(
717 VkPhysicalDevice physicalDevice
,
718 VkPhysicalDeviceFeatures
* pFeatures
)
720 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
722 *pFeatures
= (VkPhysicalDeviceFeatures
) {
723 .robustBufferAccess
= true,
724 .fullDrawIndexUint32
= true,
725 .imageCubeArray
= true,
726 .independentBlend
= true,
727 .geometryShader
= true,
728 .tessellationShader
= true,
729 .sampleRateShading
= true,
730 .dualSrcBlend
= true,
732 .multiDrawIndirect
= true,
733 .drawIndirectFirstInstance
= true,
735 .depthBiasClamp
= true,
736 .fillModeNonSolid
= true,
737 .depthBounds
= false,
741 .multiViewport
= true,
742 .samplerAnisotropy
= true,
743 .textureCompressionETC2
= pdevice
->info
.gen
>= 8 ||
744 pdevice
->info
.is_baytrail
,
745 .textureCompressionASTC_LDR
= pdevice
->info
.gen
>= 9, /* FINISHME CHV */
746 .textureCompressionBC
= true,
747 .occlusionQueryPrecise
= true,
748 .pipelineStatisticsQuery
= true,
749 .fragmentStoresAndAtomics
= true,
750 .shaderTessellationAndGeometryPointSize
= true,
751 .shaderImageGatherExtended
= true,
752 .shaderStorageImageExtendedFormats
= true,
753 .shaderStorageImageMultisample
= false,
754 .shaderStorageImageReadWithoutFormat
= false,
755 .shaderStorageImageWriteWithoutFormat
= true,
756 .shaderUniformBufferArrayDynamicIndexing
= true,
757 .shaderSampledImageArrayDynamicIndexing
= true,
758 .shaderStorageBufferArrayDynamicIndexing
= true,
759 .shaderStorageImageArrayDynamicIndexing
= true,
760 .shaderClipDistance
= true,
761 .shaderCullDistance
= true,
762 .shaderFloat64
= pdevice
->info
.gen
>= 8,
763 .shaderInt64
= pdevice
->info
.gen
>= 8,
764 .shaderInt16
= false,
765 .shaderResourceMinLod
= false,
766 .variableMultisampleRate
= false,
767 .inheritedQueries
= true,
770 /* We can't do image stores in vec4 shaders */
771 pFeatures
->vertexPipelineStoresAndAtomics
=
772 pdevice
->compiler
->scalar_stage
[MESA_SHADER_VERTEX
] &&
773 pdevice
->compiler
->scalar_stage
[MESA_SHADER_GEOMETRY
];
776 void anv_GetPhysicalDeviceFeatures2(
777 VkPhysicalDevice physicalDevice
,
778 VkPhysicalDeviceFeatures2
* pFeatures
)
780 anv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
782 vk_foreach_struct(ext
, pFeatures
->pNext
) {
783 switch (ext
->sType
) {
784 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES
: {
785 VkPhysicalDeviceProtectedMemoryFeatures
*features
= (void *)ext
;
786 features
->protectedMemory
= VK_FALSE
;
790 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES
: {
791 VkPhysicalDeviceMultiviewFeatures
*features
=
792 (VkPhysicalDeviceMultiviewFeatures
*)ext
;
793 features
->multiview
= true;
794 features
->multiviewGeometryShader
= true;
795 features
->multiviewTessellationShader
= true;
799 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTER_FEATURES
: {
800 VkPhysicalDeviceVariablePointerFeatures
*features
= (void *)ext
;
801 features
->variablePointersStorageBuffer
= true;
802 features
->variablePointers
= true;
806 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES
: {
807 VkPhysicalDeviceSamplerYcbcrConversionFeatures
*features
=
808 (VkPhysicalDeviceSamplerYcbcrConversionFeatures
*) ext
;
809 features
->samplerYcbcrConversion
= true;
813 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETER_FEATURES
: {
814 VkPhysicalDeviceShaderDrawParameterFeatures
*features
= (void *)ext
;
815 features
->shaderDrawParameters
= true;
819 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES_KHR
: {
820 VkPhysicalDevice16BitStorageFeaturesKHR
*features
=
821 (VkPhysicalDevice16BitStorageFeaturesKHR
*)ext
;
822 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
824 features
->storageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
825 features
->uniformAndStorageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
826 features
->storagePushConstant16
= pdevice
->info
.gen
>= 8;
827 features
->storageInputOutput16
= false;
832 anv_debug_ignored_stype(ext
->sType
);
838 void anv_GetPhysicalDeviceProperties(
839 VkPhysicalDevice physicalDevice
,
840 VkPhysicalDeviceProperties
* pProperties
)
842 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
843 const struct gen_device_info
*devinfo
= &pdevice
->info
;
845 /* See assertions made when programming the buffer surface state. */
846 const uint32_t max_raw_buffer_sz
= devinfo
->gen
>= 7 ?
847 (1ul << 30) : (1ul << 27);
849 const uint32_t max_samplers
= (devinfo
->gen
>= 8 || devinfo
->is_haswell
) ?
852 VkSampleCountFlags sample_counts
=
853 isl_device_get_sample_counts(&pdevice
->isl_dev
);
855 VkPhysicalDeviceLimits limits
= {
856 .maxImageDimension1D
= (1 << 14),
857 .maxImageDimension2D
= (1 << 14),
858 .maxImageDimension3D
= (1 << 11),
859 .maxImageDimensionCube
= (1 << 14),
860 .maxImageArrayLayers
= (1 << 11),
861 .maxTexelBufferElements
= 128 * 1024 * 1024,
862 .maxUniformBufferRange
= (1ul << 27),
863 .maxStorageBufferRange
= max_raw_buffer_sz
,
864 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
865 .maxMemoryAllocationCount
= UINT32_MAX
,
866 .maxSamplerAllocationCount
= 64 * 1024,
867 .bufferImageGranularity
= 64, /* A cache line */
868 .sparseAddressSpaceSize
= 0,
869 .maxBoundDescriptorSets
= MAX_SETS
,
870 .maxPerStageDescriptorSamplers
= max_samplers
,
871 .maxPerStageDescriptorUniformBuffers
= 64,
872 .maxPerStageDescriptorStorageBuffers
= 64,
873 .maxPerStageDescriptorSampledImages
= max_samplers
,
874 .maxPerStageDescriptorStorageImages
= 64,
875 .maxPerStageDescriptorInputAttachments
= 64,
876 .maxPerStageResources
= 250,
877 .maxDescriptorSetSamplers
= 6 * max_samplers
, /* number of stages * maxPerStageDescriptorSamplers */
878 .maxDescriptorSetUniformBuffers
= 6 * 64, /* number of stages * maxPerStageDescriptorUniformBuffers */
879 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
880 .maxDescriptorSetStorageBuffers
= 6 * 64, /* number of stages * maxPerStageDescriptorStorageBuffers */
881 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
882 .maxDescriptorSetSampledImages
= 6 * max_samplers
, /* number of stages * maxPerStageDescriptorSampledImages */
883 .maxDescriptorSetStorageImages
= 6 * 64, /* number of stages * maxPerStageDescriptorStorageImages */
884 .maxDescriptorSetInputAttachments
= 256,
885 .maxVertexInputAttributes
= MAX_VBS
,
886 .maxVertexInputBindings
= MAX_VBS
,
887 .maxVertexInputAttributeOffset
= 2047,
888 .maxVertexInputBindingStride
= 2048,
889 .maxVertexOutputComponents
= 128,
890 .maxTessellationGenerationLevel
= 64,
891 .maxTessellationPatchSize
= 32,
892 .maxTessellationControlPerVertexInputComponents
= 128,
893 .maxTessellationControlPerVertexOutputComponents
= 128,
894 .maxTessellationControlPerPatchOutputComponents
= 128,
895 .maxTessellationControlTotalOutputComponents
= 2048,
896 .maxTessellationEvaluationInputComponents
= 128,
897 .maxTessellationEvaluationOutputComponents
= 128,
898 .maxGeometryShaderInvocations
= 32,
899 .maxGeometryInputComponents
= 64,
900 .maxGeometryOutputComponents
= 128,
901 .maxGeometryOutputVertices
= 256,
902 .maxGeometryTotalOutputComponents
= 1024,
903 .maxFragmentInputComponents
= 128,
904 .maxFragmentOutputAttachments
= 8,
905 .maxFragmentDualSrcAttachments
= 1,
906 .maxFragmentCombinedOutputResources
= 8,
907 .maxComputeSharedMemorySize
= 32768,
908 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
909 .maxComputeWorkGroupInvocations
= 16 * devinfo
->max_cs_threads
,
910 .maxComputeWorkGroupSize
= {
911 16 * devinfo
->max_cs_threads
,
912 16 * devinfo
->max_cs_threads
,
913 16 * devinfo
->max_cs_threads
,
915 .subPixelPrecisionBits
= 4 /* FIXME */,
916 .subTexelPrecisionBits
= 4 /* FIXME */,
917 .mipmapPrecisionBits
= 4 /* FIXME */,
918 .maxDrawIndexedIndexValue
= UINT32_MAX
,
919 .maxDrawIndirectCount
= UINT32_MAX
,
920 .maxSamplerLodBias
= 16,
921 .maxSamplerAnisotropy
= 16,
922 .maxViewports
= MAX_VIEWPORTS
,
923 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
924 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
925 .viewportSubPixelBits
= 13, /* We take a float? */
926 .minMemoryMapAlignment
= 4096, /* A page */
927 .minTexelBufferOffsetAlignment
= 1,
928 /* We need 16 for UBO block reads to work and 32 for push UBOs */
929 .minUniformBufferOffsetAlignment
= 32,
930 .minStorageBufferOffsetAlignment
= 4,
931 .minTexelOffset
= -8,
933 .minTexelGatherOffset
= -32,
934 .maxTexelGatherOffset
= 31,
935 .minInterpolationOffset
= -0.5,
936 .maxInterpolationOffset
= 0.4375,
937 .subPixelInterpolationOffsetBits
= 4,
938 .maxFramebufferWidth
= (1 << 14),
939 .maxFramebufferHeight
= (1 << 14),
940 .maxFramebufferLayers
= (1 << 11),
941 .framebufferColorSampleCounts
= sample_counts
,
942 .framebufferDepthSampleCounts
= sample_counts
,
943 .framebufferStencilSampleCounts
= sample_counts
,
944 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
945 .maxColorAttachments
= MAX_RTS
,
946 .sampledImageColorSampleCounts
= sample_counts
,
947 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
948 .sampledImageDepthSampleCounts
= sample_counts
,
949 .sampledImageStencilSampleCounts
= sample_counts
,
950 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
951 .maxSampleMaskWords
= 1,
952 .timestampComputeAndGraphics
= false,
953 .timestampPeriod
= 1000000000.0 / devinfo
->timestamp_frequency
,
954 .maxClipDistances
= 8,
955 .maxCullDistances
= 8,
956 .maxCombinedClipAndCullDistances
= 8,
957 .discreteQueuePriorities
= 1,
958 .pointSizeRange
= { 0.125, 255.875 },
959 .lineWidthRange
= { 0.0, 7.9921875 },
960 .pointSizeGranularity
= (1.0 / 8.0),
961 .lineWidthGranularity
= (1.0 / 128.0),
962 .strictLines
= false, /* FINISHME */
963 .standardSampleLocations
= true,
964 .optimalBufferCopyOffsetAlignment
= 128,
965 .optimalBufferCopyRowPitchAlignment
= 128,
966 .nonCoherentAtomSize
= 64,
969 *pProperties
= (VkPhysicalDeviceProperties
) {
970 .apiVersion
= anv_physical_device_api_version(pdevice
),
971 .driverVersion
= vk_get_driver_version(),
973 .deviceID
= pdevice
->chipset_id
,
974 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
976 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
979 snprintf(pProperties
->deviceName
, sizeof(pProperties
->deviceName
),
980 "%s", pdevice
->name
);
981 memcpy(pProperties
->pipelineCacheUUID
,
982 pdevice
->pipeline_cache_uuid
, VK_UUID_SIZE
);
985 void anv_GetPhysicalDeviceProperties2(
986 VkPhysicalDevice physicalDevice
,
987 VkPhysicalDeviceProperties2
* pProperties
)
989 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
991 anv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
993 vk_foreach_struct(ext
, pProperties
->pNext
) {
994 switch (ext
->sType
) {
995 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
996 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
997 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
999 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
1003 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES
: {
1004 VkPhysicalDeviceIDProperties
*id_props
=
1005 (VkPhysicalDeviceIDProperties
*)ext
;
1006 memcpy(id_props
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
1007 memcpy(id_props
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
1008 /* The LUID is for Windows. */
1009 id_props
->deviceLUIDValid
= false;
1013 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES
: {
1014 VkPhysicalDeviceMaintenance3Properties
*props
=
1015 (VkPhysicalDeviceMaintenance3Properties
*)ext
;
1016 /* This value doesn't matter for us today as our per-stage
1017 * descriptors are the real limit.
1019 props
->maxPerSetDescriptors
= 1024;
1020 props
->maxMemoryAllocationSize
= MAX_MEMORY_ALLOCATION_SIZE
;
1024 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES
: {
1025 VkPhysicalDeviceMultiviewProperties
*properties
=
1026 (VkPhysicalDeviceMultiviewProperties
*)ext
;
1027 properties
->maxMultiviewViewCount
= 16;
1028 properties
->maxMultiviewInstanceIndex
= UINT32_MAX
/ 16;
1032 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES
: {
1033 VkPhysicalDevicePointClippingProperties
*properties
=
1034 (VkPhysicalDevicePointClippingProperties
*) ext
;
1035 properties
->pointClippingBehavior
= VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES
;
1036 anv_finishme("Implement pop-free point clipping");
1040 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES
: {
1041 VkPhysicalDeviceSubgroupProperties
*properties
= (void *)ext
;
1043 properties
->subgroupSize
= BRW_SUBGROUP_SIZE
;
1045 VkShaderStageFlags scalar_stages
= 0;
1046 for (unsigned stage
= 0; stage
< MESA_SHADER_STAGES
; stage
++) {
1047 if (pdevice
->compiler
->scalar_stage
[stage
])
1048 scalar_stages
|= mesa_to_vk_shader_stage(stage
);
1050 properties
->supportedStages
= scalar_stages
;
1052 properties
->supportedOperations
= VK_SUBGROUP_FEATURE_BASIC_BIT
|
1053 VK_SUBGROUP_FEATURE_VOTE_BIT
|
1054 VK_SUBGROUP_FEATURE_ARITHMETIC_BIT
|
1055 VK_SUBGROUP_FEATURE_BALLOT_BIT
|
1056 VK_SUBGROUP_FEATURE_SHUFFLE_BIT
|
1057 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT
|
1058 VK_SUBGROUP_FEATURE_CLUSTERED_BIT
|
1059 VK_SUBGROUP_FEATURE_QUAD_BIT
;
1060 properties
->quadOperationsInAllStages
= VK_TRUE
;
1065 anv_debug_ignored_stype(ext
->sType
);
1071 /* We support exactly one queue family. */
1072 static const VkQueueFamilyProperties
1073 anv_queue_family_properties
= {
1074 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
1075 VK_QUEUE_COMPUTE_BIT
|
1076 VK_QUEUE_TRANSFER_BIT
,
1078 .timestampValidBits
= 36, /* XXX: Real value here */
1079 .minImageTransferGranularity
= { 1, 1, 1 },
1082 void anv_GetPhysicalDeviceQueueFamilyProperties(
1083 VkPhysicalDevice physicalDevice
,
1085 VkQueueFamilyProperties
* pQueueFamilyProperties
)
1087 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pCount
);
1089 vk_outarray_append(&out
, p
) {
1090 *p
= anv_queue_family_properties
;
1094 void anv_GetPhysicalDeviceQueueFamilyProperties2(
1095 VkPhysicalDevice physicalDevice
,
1096 uint32_t* pQueueFamilyPropertyCount
,
1097 VkQueueFamilyProperties2
* pQueueFamilyProperties
)
1100 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pQueueFamilyPropertyCount
);
1102 vk_outarray_append(&out
, p
) {
1103 p
->queueFamilyProperties
= anv_queue_family_properties
;
1105 vk_foreach_struct(s
, p
->pNext
) {
1106 anv_debug_ignored_stype(s
->sType
);
1111 void anv_GetPhysicalDeviceMemoryProperties(
1112 VkPhysicalDevice physicalDevice
,
1113 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
1115 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
1117 pMemoryProperties
->memoryTypeCount
= physical_device
->memory
.type_count
;
1118 for (uint32_t i
= 0; i
< physical_device
->memory
.type_count
; i
++) {
1119 pMemoryProperties
->memoryTypes
[i
] = (VkMemoryType
) {
1120 .propertyFlags
= physical_device
->memory
.types
[i
].propertyFlags
,
1121 .heapIndex
= physical_device
->memory
.types
[i
].heapIndex
,
1125 pMemoryProperties
->memoryHeapCount
= physical_device
->memory
.heap_count
;
1126 for (uint32_t i
= 0; i
< physical_device
->memory
.heap_count
; i
++) {
1127 pMemoryProperties
->memoryHeaps
[i
] = (VkMemoryHeap
) {
1128 .size
= physical_device
->memory
.heaps
[i
].size
,
1129 .flags
= physical_device
->memory
.heaps
[i
].flags
,
1134 void anv_GetPhysicalDeviceMemoryProperties2(
1135 VkPhysicalDevice physicalDevice
,
1136 VkPhysicalDeviceMemoryProperties2
* pMemoryProperties
)
1138 anv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
1139 &pMemoryProperties
->memoryProperties
);
1141 vk_foreach_struct(ext
, pMemoryProperties
->pNext
) {
1142 switch (ext
->sType
) {
1144 anv_debug_ignored_stype(ext
->sType
);
1151 anv_GetDeviceGroupPeerMemoryFeatures(
1154 uint32_t localDeviceIndex
,
1155 uint32_t remoteDeviceIndex
,
1156 VkPeerMemoryFeatureFlags
* pPeerMemoryFeatures
)
1158 assert(localDeviceIndex
== 0 && remoteDeviceIndex
== 0);
1159 *pPeerMemoryFeatures
= VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT
|
1160 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT
|
1161 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT
|
1162 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT
;
1165 PFN_vkVoidFunction
anv_GetInstanceProcAddr(
1166 VkInstance _instance
,
1169 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1171 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
1172 * when we have to return valid function pointers, NULL, or it's left
1173 * undefined. See the table for exact details.
1178 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
1179 if (strcmp(pName, "vk" #entrypoint) == 0) \
1180 return (PFN_vkVoidFunction)anv_##entrypoint
1182 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties
);
1183 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties
);
1184 LOOKUP_ANV_ENTRYPOINT(CreateInstance
);
1186 #undef LOOKUP_ANV_ENTRYPOINT
1188 if (instance
== NULL
)
1191 int idx
= anv_get_entrypoint_index(pName
);
1195 return instance
->dispatch
.entrypoints
[idx
];
1198 /* With version 1+ of the loader interface the ICD should expose
1199 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
1202 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1203 VkInstance instance
,
1207 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1208 VkInstance instance
,
1211 return anv_GetInstanceProcAddr(instance
, pName
);
1214 PFN_vkVoidFunction
anv_GetDeviceProcAddr(
1218 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1220 if (!device
|| !pName
)
1223 int idx
= anv_get_entrypoint_index(pName
);
1227 return device
->dispatch
.entrypoints
[idx
];
1231 anv_CreateDebugReportCallbackEXT(VkInstance _instance
,
1232 const VkDebugReportCallbackCreateInfoEXT
* pCreateInfo
,
1233 const VkAllocationCallbacks
* pAllocator
,
1234 VkDebugReportCallbackEXT
* pCallback
)
1236 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1237 return vk_create_debug_report_callback(&instance
->debug_report_callbacks
,
1238 pCreateInfo
, pAllocator
, &instance
->alloc
,
1243 anv_DestroyDebugReportCallbackEXT(VkInstance _instance
,
1244 VkDebugReportCallbackEXT _callback
,
1245 const VkAllocationCallbacks
* pAllocator
)
1247 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1248 vk_destroy_debug_report_callback(&instance
->debug_report_callbacks
,
1249 _callback
, pAllocator
, &instance
->alloc
);
1253 anv_DebugReportMessageEXT(VkInstance _instance
,
1254 VkDebugReportFlagsEXT flags
,
1255 VkDebugReportObjectTypeEXT objectType
,
1258 int32_t messageCode
,
1259 const char* pLayerPrefix
,
1260 const char* pMessage
)
1262 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1263 vk_debug_report(&instance
->debug_report_callbacks
, flags
, objectType
,
1264 object
, location
, messageCode
, pLayerPrefix
, pMessage
);
1268 anv_queue_init(struct anv_device
*device
, struct anv_queue
*queue
)
1270 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1271 queue
->device
= device
;
1272 queue
->pool
= &device
->surface_state_pool
;
1277 anv_queue_finish(struct anv_queue
*queue
)
1281 static struct anv_state
1282 anv_state_pool_emit_data(struct anv_state_pool
*pool
, size_t size
, size_t align
, const void *p
)
1284 struct anv_state state
;
1286 state
= anv_state_pool_alloc(pool
, size
, align
);
1287 memcpy(state
.map
, p
, size
);
1289 anv_state_flush(pool
->block_pool
.device
, state
);
1294 struct gen8_border_color
{
1299 /* Pad out to 64 bytes */
1304 anv_device_init_border_colors(struct anv_device
*device
)
1306 static const struct gen8_border_color border_colors
[] = {
1307 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
1308 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
1309 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
1310 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
1311 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
1312 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
1315 device
->border_colors
= anv_state_pool_emit_data(&device
->dynamic_state_pool
,
1316 sizeof(border_colors
), 64,
1321 anv_device_init_trivial_batch(struct anv_device
*device
)
1323 anv_bo_init_new(&device
->trivial_batch_bo
, device
, 4096);
1325 if (device
->instance
->physicalDevice
.has_exec_async
)
1326 device
->trivial_batch_bo
.flags
|= EXEC_OBJECT_ASYNC
;
1328 void *map
= anv_gem_mmap(device
, device
->trivial_batch_bo
.gem_handle
,
1331 struct anv_batch batch
= {
1337 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
1338 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
1340 if (!device
->info
.has_llc
)
1341 gen_clflush_range(map
, batch
.next
- map
);
1343 anv_gem_munmap(map
, device
->trivial_batch_bo
.size
);
1346 VkResult
anv_EnumerateDeviceExtensionProperties(
1347 VkPhysicalDevice physicalDevice
,
1348 const char* pLayerName
,
1349 uint32_t* pPropertyCount
,
1350 VkExtensionProperties
* pProperties
)
1352 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
1353 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
1356 for (int i
= 0; i
< ANV_DEVICE_EXTENSION_COUNT
; i
++) {
1357 if (device
->supported_extensions
.extensions
[i
]) {
1358 vk_outarray_append(&out
, prop
) {
1359 *prop
= anv_device_extensions
[i
];
1364 return vk_outarray_status(&out
);
1368 anv_device_init_dispatch(struct anv_device
*device
)
1370 const struct anv_dispatch_table
*genX_table
;
1371 switch (device
->info
.gen
) {
1373 genX_table
= &gen11_dispatch_table
;
1376 genX_table
= &gen10_dispatch_table
;
1379 genX_table
= &gen9_dispatch_table
;
1382 genX_table
= &gen8_dispatch_table
;
1385 if (device
->info
.is_haswell
)
1386 genX_table
= &gen75_dispatch_table
;
1388 genX_table
= &gen7_dispatch_table
;
1391 unreachable("unsupported gen\n");
1394 for (unsigned i
= 0; i
< ARRAY_SIZE(device
->dispatch
.entrypoints
); i
++) {
1395 /* Vulkan requires that entrypoints for extensions which have not been
1396 * enabled must not be advertised.
1398 if (!anv_entrypoint_is_enabled(i
, device
->instance
->apiVersion
,
1399 &device
->instance
->enabled_extensions
,
1400 &device
->enabled_extensions
)) {
1401 device
->dispatch
.entrypoints
[i
] = NULL
;
1402 } else if (genX_table
->entrypoints
[i
]) {
1403 device
->dispatch
.entrypoints
[i
] = genX_table
->entrypoints
[i
];
1405 device
->dispatch
.entrypoints
[i
] = anv_dispatch_table
.entrypoints
[i
];
1411 vk_priority_to_gen(int priority
)
1414 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT
:
1415 return GEN_CONTEXT_LOW_PRIORITY
;
1416 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
:
1417 return GEN_CONTEXT_MEDIUM_PRIORITY
;
1418 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT
:
1419 return GEN_CONTEXT_HIGH_PRIORITY
;
1420 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT
:
1421 return GEN_CONTEXT_REALTIME_PRIORITY
;
1423 unreachable("Invalid priority");
1428 anv_device_init_hiz_clear_batch(struct anv_device
*device
)
1430 anv_bo_init_new(&device
->hiz_clear_bo
, device
, 4096);
1431 uint32_t *map
= anv_gem_mmap(device
, device
->hiz_clear_bo
.gem_handle
,
1434 union isl_color_value hiz_clear
= { .u32
= { 0, } };
1435 hiz_clear
.f32
[0] = ANV_HZ_FC_VAL
;
1437 memcpy(map
, hiz_clear
.u32
, sizeof(hiz_clear
.u32
));
1438 anv_gem_munmap(map
, device
->hiz_clear_bo
.size
);
1441 VkResult
anv_CreateDevice(
1442 VkPhysicalDevice physicalDevice
,
1443 const VkDeviceCreateInfo
* pCreateInfo
,
1444 const VkAllocationCallbacks
* pAllocator
,
1447 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
1449 struct anv_device
*device
;
1451 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO
);
1453 struct anv_device_extension_table enabled_extensions
= { };
1454 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
1456 for (idx
= 0; idx
< ANV_DEVICE_EXTENSION_COUNT
; idx
++) {
1457 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
1458 anv_device_extensions
[idx
].extensionName
) == 0)
1462 if (idx
>= ANV_DEVICE_EXTENSION_COUNT
)
1463 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
1465 if (!physical_device
->supported_extensions
.extensions
[idx
])
1466 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
1468 enabled_extensions
.extensions
[idx
] = true;
1471 /* Check enabled features */
1472 if (pCreateInfo
->pEnabledFeatures
) {
1473 VkPhysicalDeviceFeatures supported_features
;
1474 anv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
1475 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
1476 VkBool32
*enabled_feature
= (VkBool32
*)pCreateInfo
->pEnabledFeatures
;
1477 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
1478 for (uint32_t i
= 0; i
< num_features
; i
++) {
1479 if (enabled_feature
[i
] && !supported_feature
[i
])
1480 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
1484 /* Check requested queues and fail if we are requested to create any
1485 * queues with flags we don't support.
1487 assert(pCreateInfo
->queueCreateInfoCount
> 0);
1488 for (uint32_t i
= 0; i
< pCreateInfo
->queueCreateInfoCount
; i
++) {
1489 if (pCreateInfo
->pQueueCreateInfos
[i
].flags
!= 0)
1490 return vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1493 /* Check if client specified queue priority. */
1494 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*queue_priority
=
1495 vk_find_struct_const(pCreateInfo
->pQueueCreateInfos
[0].pNext
,
1496 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT
);
1498 VkQueueGlobalPriorityEXT priority
=
1499 queue_priority
? queue_priority
->globalPriority
:
1500 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
;
1502 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
1504 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
1506 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1508 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1509 device
->instance
= physical_device
->instance
;
1510 device
->chipset_id
= physical_device
->chipset_id
;
1511 device
->no_hw
= physical_device
->no_hw
;
1512 device
->lost
= false;
1515 device
->alloc
= *pAllocator
;
1517 device
->alloc
= physical_device
->instance
->alloc
;
1519 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
1520 device
->fd
= open(physical_device
->path
, O_RDWR
| O_CLOEXEC
);
1521 if (device
->fd
== -1) {
1522 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1526 device
->context_id
= anv_gem_create_context(device
);
1527 if (device
->context_id
== -1) {
1528 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1532 /* As per spec, the driver implementation may deny requests to acquire
1533 * a priority above the default priority (MEDIUM) if the caller does not
1534 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
1537 if (physical_device
->has_context_priority
) {
1538 int err
= anv_gem_set_context_param(device
->fd
, device
->context_id
,
1539 I915_CONTEXT_PARAM_PRIORITY
,
1540 vk_priority_to_gen(priority
));
1541 if (err
!= 0 && priority
> VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
) {
1542 result
= vk_error(VK_ERROR_NOT_PERMITTED_EXT
);
1547 device
->info
= physical_device
->info
;
1548 device
->isl_dev
= physical_device
->isl_dev
;
1550 /* On Broadwell and later, we can use batch chaining to more efficiently
1551 * implement growing command buffers. Prior to Haswell, the kernel
1552 * command parser gets in the way and we have to fall back to growing
1555 device
->can_chain_batches
= device
->info
.gen
>= 8;
1557 device
->robust_buffer_access
= pCreateInfo
->pEnabledFeatures
&&
1558 pCreateInfo
->pEnabledFeatures
->robustBufferAccess
;
1559 device
->enabled_extensions
= enabled_extensions
;
1561 anv_device_init_dispatch(device
);
1563 if (pthread_mutex_init(&device
->mutex
, NULL
) != 0) {
1564 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1565 goto fail_context_id
;
1568 pthread_condattr_t condattr
;
1569 if (pthread_condattr_init(&condattr
) != 0) {
1570 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1573 if (pthread_condattr_setclock(&condattr
, CLOCK_MONOTONIC
) != 0) {
1574 pthread_condattr_destroy(&condattr
);
1575 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1578 if (pthread_cond_init(&device
->queue_submit
, NULL
) != 0) {
1579 pthread_condattr_destroy(&condattr
);
1580 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1583 pthread_condattr_destroy(&condattr
);
1586 (physical_device
->supports_48bit_addresses
? EXEC_OBJECT_SUPPORTS_48B_ADDRESS
: 0) |
1587 (physical_device
->has_exec_async
? EXEC_OBJECT_ASYNC
: 0) |
1588 (physical_device
->has_exec_capture
? EXEC_OBJECT_CAPTURE
: 0);
1590 anv_bo_pool_init(&device
->batch_bo_pool
, device
, bo_flags
);
1592 result
= anv_bo_cache_init(&device
->bo_cache
);
1593 if (result
!= VK_SUCCESS
)
1594 goto fail_batch_bo_pool
;
1596 /* For the state pools we explicitly disable 48bit. */
1597 bo_flags
= (physical_device
->has_exec_async
? EXEC_OBJECT_ASYNC
: 0) |
1598 (physical_device
->has_exec_capture
? EXEC_OBJECT_CAPTURE
: 0);
1600 result
= anv_state_pool_init(&device
->dynamic_state_pool
, device
, 16384,
1602 if (result
!= VK_SUCCESS
)
1605 result
= anv_state_pool_init(&device
->instruction_state_pool
, device
, 16384,
1607 if (result
!= VK_SUCCESS
)
1608 goto fail_dynamic_state_pool
;
1610 result
= anv_state_pool_init(&device
->surface_state_pool
, device
, 4096,
1612 if (result
!= VK_SUCCESS
)
1613 goto fail_instruction_state_pool
;
1615 result
= anv_bo_init_new(&device
->workaround_bo
, device
, 1024);
1616 if (result
!= VK_SUCCESS
)
1617 goto fail_surface_state_pool
;
1619 anv_device_init_trivial_batch(device
);
1621 if (device
->info
.gen
>= 10)
1622 anv_device_init_hiz_clear_batch(device
);
1624 anv_scratch_pool_init(device
, &device
->scratch_pool
);
1626 anv_queue_init(device
, &device
->queue
);
1628 switch (device
->info
.gen
) {
1630 if (!device
->info
.is_haswell
)
1631 result
= gen7_init_device_state(device
);
1633 result
= gen75_init_device_state(device
);
1636 result
= gen8_init_device_state(device
);
1639 result
= gen9_init_device_state(device
);
1642 result
= gen10_init_device_state(device
);
1645 result
= gen11_init_device_state(device
);
1648 /* Shouldn't get here as we don't create physical devices for any other
1650 unreachable("unhandled gen");
1652 if (result
!= VK_SUCCESS
)
1653 goto fail_workaround_bo
;
1655 anv_device_init_blorp(device
);
1657 anv_device_init_border_colors(device
);
1659 *pDevice
= anv_device_to_handle(device
);
1664 anv_queue_finish(&device
->queue
);
1665 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
1666 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
1667 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
1668 fail_surface_state_pool
:
1669 anv_state_pool_finish(&device
->surface_state_pool
);
1670 fail_instruction_state_pool
:
1671 anv_state_pool_finish(&device
->instruction_state_pool
);
1672 fail_dynamic_state_pool
:
1673 anv_state_pool_finish(&device
->dynamic_state_pool
);
1675 anv_bo_cache_finish(&device
->bo_cache
);
1677 anv_bo_pool_finish(&device
->batch_bo_pool
);
1678 pthread_cond_destroy(&device
->queue_submit
);
1680 pthread_mutex_destroy(&device
->mutex
);
1682 anv_gem_destroy_context(device
, device
->context_id
);
1686 vk_free(&device
->alloc
, device
);
1691 void anv_DestroyDevice(
1693 const VkAllocationCallbacks
* pAllocator
)
1695 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1700 anv_device_finish_blorp(device
);
1702 anv_queue_finish(&device
->queue
);
1704 #ifdef HAVE_VALGRIND
1705 /* We only need to free these to prevent valgrind errors. The backing
1706 * BO will go away in a couple of lines so we don't actually leak.
1708 anv_state_pool_free(&device
->dynamic_state_pool
, device
->border_colors
);
1711 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
1713 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
1714 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
1716 anv_gem_close(device
, device
->trivial_batch_bo
.gem_handle
);
1717 if (device
->info
.gen
>= 10)
1718 anv_gem_close(device
, device
->hiz_clear_bo
.gem_handle
);
1720 anv_state_pool_finish(&device
->surface_state_pool
);
1721 anv_state_pool_finish(&device
->instruction_state_pool
);
1722 anv_state_pool_finish(&device
->dynamic_state_pool
);
1724 anv_bo_cache_finish(&device
->bo_cache
);
1726 anv_bo_pool_finish(&device
->batch_bo_pool
);
1728 pthread_cond_destroy(&device
->queue_submit
);
1729 pthread_mutex_destroy(&device
->mutex
);
1731 anv_gem_destroy_context(device
, device
->context_id
);
1735 vk_free(&device
->alloc
, device
);
1738 VkResult
anv_EnumerateInstanceLayerProperties(
1739 uint32_t* pPropertyCount
,
1740 VkLayerProperties
* pProperties
)
1742 if (pProperties
== NULL
) {
1743 *pPropertyCount
= 0;
1747 /* None supported at this time */
1748 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1751 VkResult
anv_EnumerateDeviceLayerProperties(
1752 VkPhysicalDevice physicalDevice
,
1753 uint32_t* pPropertyCount
,
1754 VkLayerProperties
* pProperties
)
1756 if (pProperties
== NULL
) {
1757 *pPropertyCount
= 0;
1761 /* None supported at this time */
1762 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1765 void anv_GetDeviceQueue(
1767 uint32_t queueNodeIndex
,
1768 uint32_t queueIndex
,
1771 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1773 assert(queueIndex
== 0);
1775 *pQueue
= anv_queue_to_handle(&device
->queue
);
1778 void anv_GetDeviceQueue2(
1780 const VkDeviceQueueInfo2
* pQueueInfo
,
1783 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1785 assert(pQueueInfo
->queueIndex
== 0);
1787 if (pQueueInfo
->flags
== device
->queue
.flags
)
1788 *pQueue
= anv_queue_to_handle(&device
->queue
);
1794 anv_device_query_status(struct anv_device
*device
)
1796 /* This isn't likely as most of the callers of this function already check
1797 * for it. However, it doesn't hurt to check and it potentially lets us
1800 if (unlikely(device
->lost
))
1801 return VK_ERROR_DEVICE_LOST
;
1803 uint32_t active
, pending
;
1804 int ret
= anv_gem_gpu_get_reset_stats(device
, &active
, &pending
);
1806 /* We don't know the real error. */
1807 device
->lost
= true;
1808 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1809 "get_reset_stats failed: %m");
1813 device
->lost
= true;
1814 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1815 "GPU hung on one of our command buffers");
1816 } else if (pending
) {
1817 device
->lost
= true;
1818 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1819 "GPU hung with commands in-flight");
1826 anv_device_bo_busy(struct anv_device
*device
, struct anv_bo
*bo
)
1828 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
1829 * Other usages of the BO (such as on different hardware) will not be
1830 * flagged as "busy" by this ioctl. Use with care.
1832 int ret
= anv_gem_busy(device
, bo
->gem_handle
);
1834 return VK_NOT_READY
;
1835 } else if (ret
== -1) {
1836 /* We don't know the real error. */
1837 device
->lost
= true;
1838 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1839 "gem wait failed: %m");
1842 /* Query for device status after the busy call. If the BO we're checking
1843 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
1844 * client because it clearly doesn't have valid data. Yes, this most
1845 * likely means an ioctl, but we just did an ioctl to query the busy status
1846 * so it's no great loss.
1848 return anv_device_query_status(device
);
1852 anv_device_wait(struct anv_device
*device
, struct anv_bo
*bo
,
1855 int ret
= anv_gem_wait(device
, bo
->gem_handle
, &timeout
);
1856 if (ret
== -1 && errno
== ETIME
) {
1858 } else if (ret
== -1) {
1859 /* We don't know the real error. */
1860 device
->lost
= true;
1861 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1862 "gem wait failed: %m");
1865 /* Query for device status after the wait. If the BO we're waiting on got
1866 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
1867 * because it clearly doesn't have valid data. Yes, this most likely means
1868 * an ioctl, but we just did an ioctl to wait so it's no great loss.
1870 return anv_device_query_status(device
);
1873 VkResult
anv_DeviceWaitIdle(
1876 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1877 if (unlikely(device
->lost
))
1878 return VK_ERROR_DEVICE_LOST
;
1880 struct anv_batch batch
;
1883 batch
.start
= batch
.next
= cmds
;
1884 batch
.end
= (void *) cmds
+ sizeof(cmds
);
1886 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
1887 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
1889 return anv_device_submit_simple_batch(device
, &batch
);
1893 anv_bo_init_new(struct anv_bo
*bo
, struct anv_device
*device
, uint64_t size
)
1895 uint32_t gem_handle
= anv_gem_create(device
, size
);
1897 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
1899 anv_bo_init(bo
, gem_handle
, size
);
1904 VkResult
anv_AllocateMemory(
1906 const VkMemoryAllocateInfo
* pAllocateInfo
,
1907 const VkAllocationCallbacks
* pAllocator
,
1908 VkDeviceMemory
* pMem
)
1910 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1911 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
1912 struct anv_device_memory
*mem
;
1913 VkResult result
= VK_SUCCESS
;
1915 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
1917 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
1918 assert(pAllocateInfo
->allocationSize
> 0);
1920 if (pAllocateInfo
->allocationSize
> MAX_MEMORY_ALLOCATION_SIZE
)
1921 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
1923 /* FINISHME: Fail if allocation request exceeds heap size. */
1925 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
1926 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1928 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1930 assert(pAllocateInfo
->memoryTypeIndex
< pdevice
->memory
.type_count
);
1931 mem
->type
= &pdevice
->memory
.types
[pAllocateInfo
->memoryTypeIndex
];
1935 const VkImportMemoryFdInfoKHR
*fd_info
=
1936 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_FD_INFO_KHR
);
1938 /* The Vulkan spec permits handleType to be 0, in which case the struct is
1941 if (fd_info
&& fd_info
->handleType
) {
1942 /* At the moment, we support only the below handle types. */
1943 assert(fd_info
->handleType
==
1944 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
1945 fd_info
->handleType
==
1946 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
1948 result
= anv_bo_cache_import(device
, &device
->bo_cache
,
1949 fd_info
->fd
, &mem
->bo
);
1950 if (result
!= VK_SUCCESS
)
1953 VkDeviceSize aligned_alloc_size
=
1954 align_u64(pAllocateInfo
->allocationSize
, 4096);
1956 /* For security purposes, we reject importing the bo if it's smaller
1957 * than the requested allocation size. This prevents a malicious client
1958 * from passing a buffer to a trusted client, lying about the size, and
1959 * telling the trusted client to try and texture from an image that goes
1960 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
1961 * in the trusted client. The trusted client can protect itself against
1962 * this sort of attack but only if it can trust the buffer size.
1964 if (mem
->bo
->size
< aligned_alloc_size
) {
1965 result
= vk_errorf(device
->instance
, device
,
1966 VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
,
1967 "aligned allocationSize too large for "
1968 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR: "
1969 "%"PRIu64
"B > %"PRIu64
"B",
1970 aligned_alloc_size
, mem
->bo
->size
);
1971 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
1975 /* From the Vulkan spec:
1977 * "Importing memory from a file descriptor transfers ownership of
1978 * the file descriptor from the application to the Vulkan
1979 * implementation. The application must not perform any operations on
1980 * the file descriptor after a successful import."
1982 * If the import fails, we leave the file descriptor open.
1986 result
= anv_bo_cache_alloc(device
, &device
->bo_cache
,
1987 pAllocateInfo
->allocationSize
,
1989 if (result
!= VK_SUCCESS
)
1992 const VkMemoryDedicatedAllocateInfoKHR
*dedicated_info
=
1993 vk_find_struct_const(pAllocateInfo
->pNext
, MEMORY_DEDICATED_ALLOCATE_INFO_KHR
);
1994 if (dedicated_info
&& dedicated_info
->image
!= VK_NULL_HANDLE
) {
1995 ANV_FROM_HANDLE(anv_image
, image
, dedicated_info
->image
);
1997 /* Some legacy (non-modifiers) consumers need the tiling to be set on
1998 * the BO. In this case, we have a dedicated allocation.
2000 if (image
->needs_set_tiling
) {
2001 const uint32_t i915_tiling
=
2002 isl_tiling_to_i915_tiling(image
->planes
[0].surface
.isl
.tiling
);
2003 int ret
= anv_gem_set_tiling(device
, mem
->bo
->gem_handle
,
2004 image
->planes
[0].surface
.isl
.row_pitch
,
2007 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
2008 return vk_errorf(device
->instance
, NULL
,
2009 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
2010 "failed to set BO tiling: %m");
2016 assert(mem
->type
->heapIndex
< pdevice
->memory
.heap_count
);
2017 if (pdevice
->memory
.heaps
[mem
->type
->heapIndex
].supports_48bit_addresses
)
2018 mem
->bo
->flags
|= EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
2020 const struct wsi_memory_allocate_info
*wsi_info
=
2021 vk_find_struct_const(pAllocateInfo
->pNext
, WSI_MEMORY_ALLOCATE_INFO_MESA
);
2022 if (wsi_info
&& wsi_info
->implicit_sync
) {
2023 /* We need to set the WRITE flag on window system buffers so that GEM
2024 * will know we're writing to them and synchronize uses on other rings
2025 * (eg if the display server uses the blitter ring).
2027 mem
->bo
->flags
|= EXEC_OBJECT_WRITE
;
2028 } else if (pdevice
->has_exec_async
) {
2029 mem
->bo
->flags
|= EXEC_OBJECT_ASYNC
;
2032 *pMem
= anv_device_memory_to_handle(mem
);
2037 vk_free2(&device
->alloc
, pAllocator
, mem
);
2042 VkResult
anv_GetMemoryFdKHR(
2044 const VkMemoryGetFdInfoKHR
* pGetFdInfo
,
2047 ANV_FROM_HANDLE(anv_device
, dev
, device_h
);
2048 ANV_FROM_HANDLE(anv_device_memory
, mem
, pGetFdInfo
->memory
);
2050 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
2052 assert(pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
2053 pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
2055 return anv_bo_cache_export(dev
, &dev
->bo_cache
, mem
->bo
, pFd
);
2058 VkResult
anv_GetMemoryFdPropertiesKHR(
2060 VkExternalMemoryHandleTypeFlagBitsKHR handleType
,
2062 VkMemoryFdPropertiesKHR
* pMemoryFdProperties
)
2064 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2065 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2067 switch (handleType
) {
2068 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
:
2069 /* dma-buf can be imported as any memory type */
2070 pMemoryFdProperties
->memoryTypeBits
=
2071 (1 << pdevice
->memory
.type_count
) - 1;
2075 /* The valid usage section for this function says:
2077 * "handleType must not be one of the handle types defined as
2080 * So opaque handle types fall into the default "unsupported" case.
2082 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
2086 void anv_FreeMemory(
2088 VkDeviceMemory _mem
,
2089 const VkAllocationCallbacks
* pAllocator
)
2091 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2092 ANV_FROM_HANDLE(anv_device_memory
, mem
, _mem
);
2098 anv_UnmapMemory(_device
, _mem
);
2100 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
2102 vk_free2(&device
->alloc
, pAllocator
, mem
);
2105 VkResult
anv_MapMemory(
2107 VkDeviceMemory _memory
,
2108 VkDeviceSize offset
,
2110 VkMemoryMapFlags flags
,
2113 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2114 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
2121 if (size
== VK_WHOLE_SIZE
)
2122 size
= mem
->bo
->size
- offset
;
2124 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
2126 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
2127 * assert(size != 0);
2128 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
2129 * equal to the size of the memory minus offset
2132 assert(offset
+ size
<= mem
->bo
->size
);
2134 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
2135 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
2136 * at a time is valid. We could just mmap up front and return an offset
2137 * pointer here, but that may exhaust virtual memory on 32 bit
2140 uint32_t gem_flags
= 0;
2142 if (!device
->info
.has_llc
&&
2143 (mem
->type
->propertyFlags
& VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
))
2144 gem_flags
|= I915_MMAP_WC
;
2146 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
2147 uint64_t map_offset
= offset
& ~4095ull;
2148 assert(offset
>= map_offset
);
2149 uint64_t map_size
= (offset
+ size
) - map_offset
;
2151 /* Let's map whole pages */
2152 map_size
= align_u64(map_size
, 4096);
2154 void *map
= anv_gem_mmap(device
, mem
->bo
->gem_handle
,
2155 map_offset
, map_size
, gem_flags
);
2156 if (map
== MAP_FAILED
)
2157 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
2160 mem
->map_size
= map_size
;
2162 *ppData
= mem
->map
+ (offset
- map_offset
);
2167 void anv_UnmapMemory(
2169 VkDeviceMemory _memory
)
2171 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
2176 anv_gem_munmap(mem
->map
, mem
->map_size
);
2183 clflush_mapped_ranges(struct anv_device
*device
,
2185 const VkMappedMemoryRange
*ranges
)
2187 for (uint32_t i
= 0; i
< count
; i
++) {
2188 ANV_FROM_HANDLE(anv_device_memory
, mem
, ranges
[i
].memory
);
2189 if (ranges
[i
].offset
>= mem
->map_size
)
2192 gen_clflush_range(mem
->map
+ ranges
[i
].offset
,
2193 MIN2(ranges
[i
].size
, mem
->map_size
- ranges
[i
].offset
));
2197 VkResult
anv_FlushMappedMemoryRanges(
2199 uint32_t memoryRangeCount
,
2200 const VkMappedMemoryRange
* pMemoryRanges
)
2202 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2204 if (device
->info
.has_llc
)
2207 /* Make sure the writes we're flushing have landed. */
2208 __builtin_ia32_mfence();
2210 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
2215 VkResult
anv_InvalidateMappedMemoryRanges(
2217 uint32_t memoryRangeCount
,
2218 const VkMappedMemoryRange
* pMemoryRanges
)
2220 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2222 if (device
->info
.has_llc
)
2225 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
2227 /* Make sure no reads get moved up above the invalidate. */
2228 __builtin_ia32_mfence();
2233 void anv_GetBufferMemoryRequirements(
2236 VkMemoryRequirements
* pMemoryRequirements
)
2238 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
2239 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2240 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2242 /* The Vulkan spec (git aaed022) says:
2244 * memoryTypeBits is a bitfield and contains one bit set for every
2245 * supported memory type for the resource. The bit `1<<i` is set if and
2246 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
2247 * structure for the physical device is supported.
2249 uint32_t memory_types
= 0;
2250 for (uint32_t i
= 0; i
< pdevice
->memory
.type_count
; i
++) {
2251 uint32_t valid_usage
= pdevice
->memory
.types
[i
].valid_buffer_usage
;
2252 if ((valid_usage
& buffer
->usage
) == buffer
->usage
)
2253 memory_types
|= (1u << i
);
2256 /* Base alignment requirement of a cache line */
2257 uint32_t alignment
= 16;
2259 /* We need an alignment of 32 for pushing UBOs */
2260 if (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
)
2261 alignment
= MAX2(alignment
, 32);
2263 pMemoryRequirements
->size
= buffer
->size
;
2264 pMemoryRequirements
->alignment
= alignment
;
2266 /* Storage and Uniform buffers should have their size aligned to
2267 * 32-bits to avoid boundary checks when last DWord is not complete.
2268 * This would ensure that not internal padding would be needed for
2271 if (device
->robust_buffer_access
&&
2272 (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
||
2273 buffer
->usage
& VK_BUFFER_USAGE_STORAGE_BUFFER_BIT
))
2274 pMemoryRequirements
->size
= align_u64(buffer
->size
, 4);
2276 pMemoryRequirements
->memoryTypeBits
= memory_types
;
2279 void anv_GetBufferMemoryRequirements2(
2281 const VkBufferMemoryRequirementsInfo2
* pInfo
,
2282 VkMemoryRequirements2
* pMemoryRequirements
)
2284 anv_GetBufferMemoryRequirements(_device
, pInfo
->buffer
,
2285 &pMemoryRequirements
->memoryRequirements
);
2287 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
2288 switch (ext
->sType
) {
2289 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
2290 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
2291 requirements
->prefersDedicatedAllocation
= VK_FALSE
;
2292 requirements
->requiresDedicatedAllocation
= VK_FALSE
;
2297 anv_debug_ignored_stype(ext
->sType
);
2303 void anv_GetImageMemoryRequirements(
2306 VkMemoryRequirements
* pMemoryRequirements
)
2308 ANV_FROM_HANDLE(anv_image
, image
, _image
);
2309 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2310 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2312 /* The Vulkan spec (git aaed022) says:
2314 * memoryTypeBits is a bitfield and contains one bit set for every
2315 * supported memory type for the resource. The bit `1<<i` is set if and
2316 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
2317 * structure for the physical device is supported.
2319 * All types are currently supported for images.
2321 uint32_t memory_types
= (1ull << pdevice
->memory
.type_count
) - 1;
2323 pMemoryRequirements
->size
= image
->size
;
2324 pMemoryRequirements
->alignment
= image
->alignment
;
2325 pMemoryRequirements
->memoryTypeBits
= memory_types
;
2328 void anv_GetImageMemoryRequirements2(
2330 const VkImageMemoryRequirementsInfo2
* pInfo
,
2331 VkMemoryRequirements2
* pMemoryRequirements
)
2333 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2334 ANV_FROM_HANDLE(anv_image
, image
, pInfo
->image
);
2336 anv_GetImageMemoryRequirements(_device
, pInfo
->image
,
2337 &pMemoryRequirements
->memoryRequirements
);
2339 vk_foreach_struct_const(ext
, pInfo
->pNext
) {
2340 switch (ext
->sType
) {
2341 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO
: {
2342 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2343 const VkImagePlaneMemoryRequirementsInfoKHR
*plane_reqs
=
2344 (const VkImagePlaneMemoryRequirementsInfoKHR
*) ext
;
2345 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
,
2346 plane_reqs
->planeAspect
);
2348 assert(image
->planes
[plane
].offset
== 0);
2350 /* The Vulkan spec (git aaed022) says:
2352 * memoryTypeBits is a bitfield and contains one bit set for every
2353 * supported memory type for the resource. The bit `1<<i` is set
2354 * if and only if the memory type `i` in the
2355 * VkPhysicalDeviceMemoryProperties structure for the physical
2356 * device is supported.
2358 * All types are currently supported for images.
2360 pMemoryRequirements
->memoryRequirements
.memoryTypeBits
=
2361 (1ull << pdevice
->memory
.type_count
) - 1;
2363 pMemoryRequirements
->memoryRequirements
.size
= image
->planes
[plane
].size
;
2364 pMemoryRequirements
->memoryRequirements
.alignment
=
2365 image
->planes
[plane
].alignment
;
2370 anv_debug_ignored_stype(ext
->sType
);
2375 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
2376 switch (ext
->sType
) {
2377 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
2378 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
2379 if (image
->needs_set_tiling
) {
2380 /* If we need to set the tiling for external consumers, we need a
2381 * dedicated allocation.
2383 * See also anv_AllocateMemory.
2385 requirements
->prefersDedicatedAllocation
= VK_TRUE
;
2386 requirements
->requiresDedicatedAllocation
= VK_TRUE
;
2388 requirements
->prefersDedicatedAllocation
= VK_FALSE
;
2389 requirements
->requiresDedicatedAllocation
= VK_FALSE
;
2395 anv_debug_ignored_stype(ext
->sType
);
2401 void anv_GetImageSparseMemoryRequirements(
2404 uint32_t* pSparseMemoryRequirementCount
,
2405 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
2407 *pSparseMemoryRequirementCount
= 0;
2410 void anv_GetImageSparseMemoryRequirements2(
2412 const VkImageSparseMemoryRequirementsInfo2
* pInfo
,
2413 uint32_t* pSparseMemoryRequirementCount
,
2414 VkSparseImageMemoryRequirements2
* pSparseMemoryRequirements
)
2416 *pSparseMemoryRequirementCount
= 0;
2419 void anv_GetDeviceMemoryCommitment(
2421 VkDeviceMemory memory
,
2422 VkDeviceSize
* pCommittedMemoryInBytes
)
2424 *pCommittedMemoryInBytes
= 0;
2428 anv_bind_buffer_memory(const VkBindBufferMemoryInfo
*pBindInfo
)
2430 ANV_FROM_HANDLE(anv_device_memory
, mem
, pBindInfo
->memory
);
2431 ANV_FROM_HANDLE(anv_buffer
, buffer
, pBindInfo
->buffer
);
2433 assert(pBindInfo
->sType
== VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
);
2436 assert((buffer
->usage
& mem
->type
->valid_buffer_usage
) == buffer
->usage
);
2437 buffer
->bo
= mem
->bo
;
2438 buffer
->offset
= pBindInfo
->memoryOffset
;
2445 VkResult
anv_BindBufferMemory(
2448 VkDeviceMemory memory
,
2449 VkDeviceSize memoryOffset
)
2451 anv_bind_buffer_memory(
2452 &(VkBindBufferMemoryInfo
) {
2453 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
,
2456 .memoryOffset
= memoryOffset
,
2462 VkResult
anv_BindBufferMemory2(
2464 uint32_t bindInfoCount
,
2465 const VkBindBufferMemoryInfo
* pBindInfos
)
2467 for (uint32_t i
= 0; i
< bindInfoCount
; i
++)
2468 anv_bind_buffer_memory(&pBindInfos
[i
]);
2473 VkResult
anv_QueueBindSparse(
2475 uint32_t bindInfoCount
,
2476 const VkBindSparseInfo
* pBindInfo
,
2479 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
2480 if (unlikely(queue
->device
->lost
))
2481 return VK_ERROR_DEVICE_LOST
;
2483 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
2488 VkResult
anv_CreateEvent(
2490 const VkEventCreateInfo
* pCreateInfo
,
2491 const VkAllocationCallbacks
* pAllocator
,
2494 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2495 struct anv_state state
;
2496 struct anv_event
*event
;
2498 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_EVENT_CREATE_INFO
);
2500 state
= anv_state_pool_alloc(&device
->dynamic_state_pool
,
2503 event
->state
= state
;
2504 event
->semaphore
= VK_EVENT_RESET
;
2506 if (!device
->info
.has_llc
) {
2507 /* Make sure the writes we're flushing have landed. */
2508 __builtin_ia32_mfence();
2509 __builtin_ia32_clflush(event
);
2512 *pEvent
= anv_event_to_handle(event
);
2517 void anv_DestroyEvent(
2520 const VkAllocationCallbacks
* pAllocator
)
2522 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2523 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2528 anv_state_pool_free(&device
->dynamic_state_pool
, event
->state
);
2531 VkResult
anv_GetEventStatus(
2535 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2536 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2538 if (unlikely(device
->lost
))
2539 return VK_ERROR_DEVICE_LOST
;
2541 if (!device
->info
.has_llc
) {
2542 /* Invalidate read cache before reading event written by GPU. */
2543 __builtin_ia32_clflush(event
);
2544 __builtin_ia32_mfence();
2548 return event
->semaphore
;
2551 VkResult
anv_SetEvent(
2555 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2556 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2558 event
->semaphore
= VK_EVENT_SET
;
2560 if (!device
->info
.has_llc
) {
2561 /* Make sure the writes we're flushing have landed. */
2562 __builtin_ia32_mfence();
2563 __builtin_ia32_clflush(event
);
2569 VkResult
anv_ResetEvent(
2573 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2574 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2576 event
->semaphore
= VK_EVENT_RESET
;
2578 if (!device
->info
.has_llc
) {
2579 /* Make sure the writes we're flushing have landed. */
2580 __builtin_ia32_mfence();
2581 __builtin_ia32_clflush(event
);
2589 VkResult
anv_CreateBuffer(
2591 const VkBufferCreateInfo
* pCreateInfo
,
2592 const VkAllocationCallbacks
* pAllocator
,
2595 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2596 struct anv_buffer
*buffer
;
2598 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
2600 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
2601 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2603 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2605 buffer
->size
= pCreateInfo
->size
;
2606 buffer
->usage
= pCreateInfo
->usage
;
2610 *pBuffer
= anv_buffer_to_handle(buffer
);
2615 void anv_DestroyBuffer(
2618 const VkAllocationCallbacks
* pAllocator
)
2620 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2621 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
2626 vk_free2(&device
->alloc
, pAllocator
, buffer
);
2630 anv_fill_buffer_surface_state(struct anv_device
*device
, struct anv_state state
,
2631 enum isl_format format
,
2632 uint32_t offset
, uint32_t range
, uint32_t stride
)
2634 isl_buffer_fill_state(&device
->isl_dev
, state
.map
,
2636 .mocs
= device
->default_mocs
,
2641 anv_state_flush(device
, state
);
2644 void anv_DestroySampler(
2647 const VkAllocationCallbacks
* pAllocator
)
2649 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2650 ANV_FROM_HANDLE(anv_sampler
, sampler
, _sampler
);
2655 vk_free2(&device
->alloc
, pAllocator
, sampler
);
2658 VkResult
anv_CreateFramebuffer(
2660 const VkFramebufferCreateInfo
* pCreateInfo
,
2661 const VkAllocationCallbacks
* pAllocator
,
2662 VkFramebuffer
* pFramebuffer
)
2664 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2665 struct anv_framebuffer
*framebuffer
;
2667 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
2669 size_t size
= sizeof(*framebuffer
) +
2670 sizeof(struct anv_image_view
*) * pCreateInfo
->attachmentCount
;
2671 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
2672 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2673 if (framebuffer
== NULL
)
2674 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2676 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
2677 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
2678 VkImageView _iview
= pCreateInfo
->pAttachments
[i
];
2679 framebuffer
->attachments
[i
] = anv_image_view_from_handle(_iview
);
2682 framebuffer
->width
= pCreateInfo
->width
;
2683 framebuffer
->height
= pCreateInfo
->height
;
2684 framebuffer
->layers
= pCreateInfo
->layers
;
2686 *pFramebuffer
= anv_framebuffer_to_handle(framebuffer
);
2691 void anv_DestroyFramebuffer(
2694 const VkAllocationCallbacks
* pAllocator
)
2696 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2697 ANV_FROM_HANDLE(anv_framebuffer
, fb
, _fb
);
2702 vk_free2(&device
->alloc
, pAllocator
, fb
);
2705 /* vk_icd.h does not declare this function, so we declare it here to
2706 * suppress Wmissing-prototypes.
2708 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2709 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
);
2711 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2712 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
)
2714 /* For the full details on loader interface versioning, see
2715 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
2716 * What follows is a condensed summary, to help you navigate the large and
2717 * confusing official doc.
2719 * - Loader interface v0 is incompatible with later versions. We don't
2722 * - In loader interface v1:
2723 * - The first ICD entrypoint called by the loader is
2724 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
2726 * - The ICD must statically expose no other Vulkan symbol unless it is
2727 * linked with -Bsymbolic.
2728 * - Each dispatchable Vulkan handle created by the ICD must be
2729 * a pointer to a struct whose first member is VK_LOADER_DATA. The
2730 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
2731 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
2732 * vkDestroySurfaceKHR(). The ICD must be capable of working with
2733 * such loader-managed surfaces.
2735 * - Loader interface v2 differs from v1 in:
2736 * - The first ICD entrypoint called by the loader is
2737 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
2738 * statically expose this entrypoint.
2740 * - Loader interface v3 differs from v2 in:
2741 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
2742 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
2743 * because the loader no longer does so.
2745 *pSupportedVersion
= MIN2(*pSupportedVersion
, 3u);