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 gtt_size
, uint64_t *heap_size
)
63 /* Query the total ram from the system */
67 uint64_t total_ram
= (uint64_t)info
.totalram
* (uint64_t)info
.mem_unit
;
69 /* We don't want to burn too much ram with the GPU. If the user has 4GiB
70 * or less, we use at most half. If they have more than 4GiB, we use 3/4.
72 uint64_t available_ram
;
73 if (total_ram
<= 4ull * 1024ull * 1024ull * 1024ull)
74 available_ram
= total_ram
/ 2;
76 available_ram
= total_ram
* 3 / 4;
78 /* We also want to leave some padding for things we allocate in the driver,
79 * so don't go over 3/4 of the GTT either.
81 uint64_t available_gtt
= gtt_size
* 3 / 4;
83 *heap_size
= MIN2(available_ram
, available_gtt
);
89 anv_physical_device_init_heaps(struct anv_physical_device
*device
, int fd
)
92 if (anv_gem_get_context_param(fd
, 0, I915_CONTEXT_PARAM_GTT_SIZE
,
94 /* If, for whatever reason, we can't actually get the GTT size from the
95 * kernel (too old?) fall back to the aperture size.
97 anv_perf_warn(NULL
, NULL
,
98 "Failed to get I915_CONTEXT_PARAM_GTT_SIZE: %m");
100 if (anv_gem_get_aperture(fd
, >t_size
) == -1) {
101 return vk_errorf(NULL
, NULL
, VK_ERROR_INITIALIZATION_FAILED
,
102 "failed to get aperture size: %m");
106 device
->supports_48bit_addresses
= (device
->info
.gen
>= 8) &&
107 gtt_size
> (4ULL << 30 /* GiB */);
109 uint64_t heap_size
= 0;
110 VkResult result
= anv_compute_heap_size(fd
, gtt_size
, &heap_size
);
111 if (result
!= VK_SUCCESS
)
114 if (heap_size
> (2ull << 30) && !device
->supports_48bit_addresses
) {
115 /* When running with an overridden PCI ID, we may get a GTT size from
116 * the kernel that is greater than 2 GiB but the execbuf check for 48bit
117 * address support can still fail. Just clamp the address space size to
118 * 2 GiB if we don't have 48-bit support.
120 intel_logw("%s:%d: The kernel reported a GTT size larger than 2 GiB but "
121 "not support for 48-bit addresses",
123 heap_size
= 2ull << 30;
126 if (heap_size
<= 3ull * (1ull << 30)) {
127 /* In this case, everything fits nicely into the 32-bit address space,
128 * so there's no need for supporting 48bit addresses on client-allocated
131 device
->memory
.heap_count
= 1;
132 device
->memory
.heaps
[0] = (struct anv_memory_heap
) {
134 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
135 .supports_48bit_addresses
= false,
138 /* Not everything will fit nicely into a 32-bit address space. In this
139 * case we need a 64-bit heap. Advertise a small 32-bit heap and a
140 * larger 48-bit heap. If we're in this case, then we have a total heap
141 * size larger than 3GiB which most likely means they have 8 GiB of
142 * video memory and so carving off 1 GiB for the 32-bit heap should be
145 const uint64_t heap_size_32bit
= 1ull << 30;
146 const uint64_t heap_size_48bit
= heap_size
- heap_size_32bit
;
148 assert(device
->supports_48bit_addresses
);
150 device
->memory
.heap_count
= 2;
151 device
->memory
.heaps
[0] = (struct anv_memory_heap
) {
152 .size
= heap_size_48bit
,
153 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
154 .supports_48bit_addresses
= true,
156 device
->memory
.heaps
[1] = (struct anv_memory_heap
) {
157 .size
= heap_size_32bit
,
158 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
159 .supports_48bit_addresses
= false,
163 uint32_t type_count
= 0;
164 for (uint32_t heap
= 0; heap
< device
->memory
.heap_count
; heap
++) {
165 uint32_t valid_buffer_usage
= ~0;
167 /* There appears to be a hardware issue in the VF cache where it only
168 * considers the bottom 32 bits of memory addresses. If you happen to
169 * have two vertex buffers which get placed exactly 4 GiB apart and use
170 * them in back-to-back draw calls, you can get collisions. In order to
171 * solve this problem, we require vertex and index buffers be bound to
172 * memory allocated out of the 32-bit heap.
174 if (device
->memory
.heaps
[heap
].supports_48bit_addresses
) {
175 valid_buffer_usage
&= ~(VK_BUFFER_USAGE_INDEX_BUFFER_BIT
|
176 VK_BUFFER_USAGE_VERTEX_BUFFER_BIT
);
179 if (device
->info
.has_llc
) {
180 /* Big core GPUs share LLC with the CPU and thus one memory type can be
181 * both cached and coherent at the same time.
183 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
184 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
185 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
186 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
187 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
189 .valid_buffer_usage
= valid_buffer_usage
,
192 /* The spec requires that we expose a host-visible, coherent memory
193 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
194 * to give the application a choice between cached, but not coherent and
195 * coherent but uncached (WC though).
197 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
198 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
199 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
200 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
202 .valid_buffer_usage
= valid_buffer_usage
,
204 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
205 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
206 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
207 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
209 .valid_buffer_usage
= valid_buffer_usage
,
213 device
->memory
.type_count
= type_count
;
219 anv_physical_device_init_uuids(struct anv_physical_device
*device
)
221 const struct build_id_note
*note
=
222 build_id_find_nhdr_for_addr(anv_physical_device_init_uuids
);
224 return vk_errorf(device
->instance
, device
,
225 VK_ERROR_INITIALIZATION_FAILED
,
226 "Failed to find build-id");
229 unsigned build_id_len
= build_id_length(note
);
230 if (build_id_len
< 20) {
231 return vk_errorf(device
->instance
, device
,
232 VK_ERROR_INITIALIZATION_FAILED
,
233 "build-id too short. It needs to be a SHA");
236 struct mesa_sha1 sha1_ctx
;
238 STATIC_ASSERT(VK_UUID_SIZE
<= sizeof(sha1
));
240 /* The pipeline cache UUID is used for determining when a pipeline cache is
241 * invalid. It needs both a driver build and the PCI ID of the device.
243 _mesa_sha1_init(&sha1_ctx
);
244 _mesa_sha1_update(&sha1_ctx
, build_id_data(note
), build_id_len
);
245 _mesa_sha1_update(&sha1_ctx
, &device
->chipset_id
,
246 sizeof(device
->chipset_id
));
247 _mesa_sha1_final(&sha1_ctx
, sha1
);
248 memcpy(device
->pipeline_cache_uuid
, sha1
, VK_UUID_SIZE
);
250 /* The driver UUID is used for determining sharability of images and memory
251 * between two Vulkan instances in separate processes. People who want to
252 * share memory need to also check the device UUID (below) so all this
253 * needs to be is the build-id.
255 memcpy(device
->driver_uuid
, build_id_data(note
), VK_UUID_SIZE
);
257 /* The device UUID uniquely identifies the given device within the machine.
258 * Since we never have more than one device, this doesn't need to be a real
259 * UUID. However, on the off-chance that someone tries to use this to
260 * cache pre-tiled images or something of the like, we use the PCI ID and
261 * some bits of ISL info to ensure that this is safe.
263 _mesa_sha1_init(&sha1_ctx
);
264 _mesa_sha1_update(&sha1_ctx
, &device
->chipset_id
,
265 sizeof(device
->chipset_id
));
266 _mesa_sha1_update(&sha1_ctx
, &device
->isl_dev
.has_bit6_swizzling
,
267 sizeof(device
->isl_dev
.has_bit6_swizzling
));
268 _mesa_sha1_final(&sha1_ctx
, sha1
);
269 memcpy(device
->device_uuid
, sha1
, VK_UUID_SIZE
);
275 anv_physical_device_init(struct anv_physical_device
*device
,
276 struct anv_instance
*instance
,
282 brw_process_intel_debug_variable();
284 fd
= open(path
, O_RDWR
| O_CLOEXEC
);
286 return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
288 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
289 device
->instance
= instance
;
291 assert(strlen(path
) < ARRAY_SIZE(device
->path
));
292 strncpy(device
->path
, path
, ARRAY_SIZE(device
->path
));
294 device
->no_hw
= getenv("INTEL_NO_HW") != NULL
;
296 const int pci_id_override
= gen_get_pci_device_id_override();
297 if (pci_id_override
< 0) {
298 device
->chipset_id
= anv_gem_get_param(fd
, I915_PARAM_CHIPSET_ID
);
299 if (!device
->chipset_id
) {
300 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
304 device
->chipset_id
= pci_id_override
;
305 device
->no_hw
= true;
308 device
->name
= gen_get_device_name(device
->chipset_id
);
309 if (!gen_get_device_info(device
->chipset_id
, &device
->info
)) {
310 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
314 if (device
->info
.is_haswell
) {
315 intel_logw("Haswell Vulkan support is incomplete");
316 } else if (device
->info
.gen
== 7 && !device
->info
.is_baytrail
) {
317 intel_logw("Ivy Bridge Vulkan support is incomplete");
318 } else if (device
->info
.gen
== 7 && device
->info
.is_baytrail
) {
319 intel_logw("Bay Trail Vulkan support is incomplete");
320 } else if (device
->info
.gen
>= 8 && device
->info
.gen
<= 10) {
321 /* Gen8-10 fully supported */
322 } else if (device
->info
.gen
== 11) {
323 intel_logw("Vulkan is not yet fully supported on gen11.");
325 result
= vk_errorf(device
->instance
, device
,
326 VK_ERROR_INCOMPATIBLE_DRIVER
,
327 "Vulkan not yet supported on %s", device
->name
);
331 device
->cmd_parser_version
= -1;
332 if (device
->info
.gen
== 7) {
333 device
->cmd_parser_version
=
334 anv_gem_get_param(fd
, I915_PARAM_CMD_PARSER_VERSION
);
335 if (device
->cmd_parser_version
== -1) {
336 result
= vk_errorf(device
->instance
, device
,
337 VK_ERROR_INITIALIZATION_FAILED
,
338 "failed to get command parser version");
343 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_WAIT_TIMEOUT
)) {
344 result
= vk_errorf(device
->instance
, device
,
345 VK_ERROR_INITIALIZATION_FAILED
,
346 "kernel missing gem wait");
350 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_EXECBUF2
)) {
351 result
= vk_errorf(device
->instance
, device
,
352 VK_ERROR_INITIALIZATION_FAILED
,
353 "kernel missing execbuf2");
357 if (!device
->info
.has_llc
&&
358 anv_gem_get_param(fd
, I915_PARAM_MMAP_VERSION
) < 1) {
359 result
= vk_errorf(device
->instance
, device
,
360 VK_ERROR_INITIALIZATION_FAILED
,
361 "kernel missing wc mmap");
365 result
= anv_physical_device_init_heaps(device
, fd
);
366 if (result
!= VK_SUCCESS
)
369 device
->has_exec_async
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_ASYNC
);
370 device
->has_exec_capture
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_CAPTURE
);
371 device
->has_exec_fence
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE
);
372 device
->has_syncobj
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE_ARRAY
);
373 device
->has_syncobj_wait
= device
->has_syncobj
&&
374 anv_gem_supports_syncobj_wait(fd
);
375 device
->has_context_priority
= anv_gem_has_context_priority(fd
);
377 device
->use_softpin
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_SOFTPIN
)
378 && device
->supports_48bit_addresses
;
380 bool swizzled
= anv_gem_get_bit6_swizzle(fd
, I915_TILING_X
);
382 /* Starting with Gen10, the timestamp frequency of the command streamer may
383 * vary from one part to another. We can query the value from the kernel.
385 if (device
->info
.gen
>= 10) {
386 int timestamp_frequency
=
387 anv_gem_get_param(fd
, I915_PARAM_CS_TIMESTAMP_FREQUENCY
);
389 if (timestamp_frequency
< 0)
390 intel_logw("Kernel 4.16-rc1+ required to properly query CS timestamp frequency");
392 device
->info
.timestamp_frequency
= timestamp_frequency
;
395 /* GENs prior to 8 do not support EU/Subslice info */
396 if (device
->info
.gen
>= 8) {
397 device
->subslice_total
= anv_gem_get_param(fd
, I915_PARAM_SUBSLICE_TOTAL
);
398 device
->eu_total
= anv_gem_get_param(fd
, I915_PARAM_EU_TOTAL
);
400 /* Without this information, we cannot get the right Braswell
401 * brandstrings, and we have to use conservative numbers for GPGPU on
402 * many platforms, but otherwise, things will just work.
404 if (device
->subslice_total
< 1 || device
->eu_total
< 1) {
405 intel_logw("Kernel 4.1 required to properly query GPU properties");
407 } else if (device
->info
.gen
== 7) {
408 device
->subslice_total
= 1 << (device
->info
.gt
- 1);
411 if (device
->info
.is_cherryview
&&
412 device
->subslice_total
> 0 && device
->eu_total
> 0) {
413 /* Logical CS threads = EUs per subslice * num threads per EU */
414 uint32_t max_cs_threads
=
415 device
->eu_total
/ device
->subslice_total
* device
->info
.num_thread_per_eu
;
417 /* Fuse configurations may give more threads than expected, never less. */
418 if (max_cs_threads
> device
->info
.max_cs_threads
)
419 device
->info
.max_cs_threads
= max_cs_threads
;
422 device
->compiler
= brw_compiler_create(NULL
, &device
->info
);
423 if (device
->compiler
== NULL
) {
424 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
427 device
->compiler
->shader_debug_log
= compiler_debug_log
;
428 device
->compiler
->shader_perf_log
= compiler_perf_log
;
429 device
->compiler
->supports_pull_constants
= false;
430 device
->compiler
->constant_buffer_0_is_relative
= true;
432 isl_device_init(&device
->isl_dev
, &device
->info
, swizzled
);
434 result
= anv_physical_device_init_uuids(device
);
435 if (result
!= VK_SUCCESS
)
438 result
= anv_init_wsi(device
);
439 if (result
!= VK_SUCCESS
) {
440 ralloc_free(device
->compiler
);
444 anv_physical_device_get_supported_extensions(device
,
445 &device
->supported_extensions
);
447 device
->local_fd
= fd
;
456 anv_physical_device_finish(struct anv_physical_device
*device
)
458 anv_finish_wsi(device
);
459 ralloc_free(device
->compiler
);
460 close(device
->local_fd
);
464 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
465 VkSystemAllocationScope allocationScope
)
471 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
472 size_t align
, VkSystemAllocationScope allocationScope
)
474 return realloc(pOriginal
, size
);
478 default_free_func(void *pUserData
, void *pMemory
)
483 static const VkAllocationCallbacks default_alloc
= {
485 .pfnAllocation
= default_alloc_func
,
486 .pfnReallocation
= default_realloc_func
,
487 .pfnFree
= default_free_func
,
490 VkResult
anv_EnumerateInstanceExtensionProperties(
491 const char* pLayerName
,
492 uint32_t* pPropertyCount
,
493 VkExtensionProperties
* pProperties
)
495 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
497 for (int i
= 0; i
< ANV_INSTANCE_EXTENSION_COUNT
; i
++) {
498 if (anv_instance_extensions_supported
.extensions
[i
]) {
499 vk_outarray_append(&out
, prop
) {
500 *prop
= anv_instance_extensions
[i
];
505 return vk_outarray_status(&out
);
508 VkResult
anv_CreateInstance(
509 const VkInstanceCreateInfo
* pCreateInfo
,
510 const VkAllocationCallbacks
* pAllocator
,
511 VkInstance
* pInstance
)
513 struct anv_instance
*instance
;
516 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
518 struct anv_instance_extension_table enabled_extensions
= {};
519 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
521 for (idx
= 0; idx
< ANV_INSTANCE_EXTENSION_COUNT
; idx
++) {
522 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
523 anv_instance_extensions
[idx
].extensionName
) == 0)
527 if (idx
>= ANV_INSTANCE_EXTENSION_COUNT
)
528 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
530 if (!anv_instance_extensions_supported
.extensions
[idx
])
531 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
533 enabled_extensions
.extensions
[idx
] = true;
536 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
537 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
539 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
541 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
544 instance
->alloc
= *pAllocator
;
546 instance
->alloc
= default_alloc
;
548 if (pCreateInfo
->pApplicationInfo
&&
549 pCreateInfo
->pApplicationInfo
->apiVersion
!= 0) {
550 instance
->apiVersion
= pCreateInfo
->pApplicationInfo
->apiVersion
;
552 anv_EnumerateInstanceVersion(&instance
->apiVersion
);
555 instance
->enabled_extensions
= enabled_extensions
;
557 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->dispatch
.entrypoints
); i
++) {
558 /* Vulkan requires that entrypoints for extensions which have not been
559 * enabled must not be advertised.
561 if (!anv_entrypoint_is_enabled(i
, instance
->apiVersion
,
562 &instance
->enabled_extensions
, NULL
)) {
563 instance
->dispatch
.entrypoints
[i
] = NULL
;
564 } else if (anv_dispatch_table
.entrypoints
[i
] != NULL
) {
565 instance
->dispatch
.entrypoints
[i
] = anv_dispatch_table
.entrypoints
[i
];
567 instance
->dispatch
.entrypoints
[i
] =
568 anv_tramp_dispatch_table
.entrypoints
[i
];
572 instance
->physicalDeviceCount
= -1;
574 result
= vk_debug_report_instance_init(&instance
->debug_report_callbacks
);
575 if (result
!= VK_SUCCESS
) {
576 vk_free2(&default_alloc
, pAllocator
, instance
);
577 return vk_error(result
);
582 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
584 *pInstance
= anv_instance_to_handle(instance
);
589 void anv_DestroyInstance(
590 VkInstance _instance
,
591 const VkAllocationCallbacks
* pAllocator
)
593 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
598 if (instance
->physicalDeviceCount
> 0) {
599 /* We support at most one physical device. */
600 assert(instance
->physicalDeviceCount
== 1);
601 anv_physical_device_finish(&instance
->physicalDevice
);
604 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
606 vk_debug_report_instance_destroy(&instance
->debug_report_callbacks
);
610 vk_free(&instance
->alloc
, instance
);
614 anv_enumerate_devices(struct anv_instance
*instance
)
616 /* TODO: Check for more devices ? */
617 drmDevicePtr devices
[8];
618 VkResult result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
621 instance
->physicalDeviceCount
= 0;
623 max_devices
= drmGetDevices2(0, devices
, ARRAY_SIZE(devices
));
625 return VK_ERROR_INCOMPATIBLE_DRIVER
;
627 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
628 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
629 devices
[i
]->bustype
== DRM_BUS_PCI
&&
630 devices
[i
]->deviceinfo
.pci
->vendor_id
== 0x8086) {
632 result
= anv_physical_device_init(&instance
->physicalDevice
,
634 devices
[i
]->nodes
[DRM_NODE_RENDER
]);
635 if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
639 drmFreeDevices(devices
, max_devices
);
641 if (result
== VK_SUCCESS
)
642 instance
->physicalDeviceCount
= 1;
648 anv_instance_ensure_physical_device(struct anv_instance
*instance
)
650 if (instance
->physicalDeviceCount
< 0) {
651 VkResult result
= anv_enumerate_devices(instance
);
652 if (result
!= VK_SUCCESS
&&
653 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
660 VkResult
anv_EnumeratePhysicalDevices(
661 VkInstance _instance
,
662 uint32_t* pPhysicalDeviceCount
,
663 VkPhysicalDevice
* pPhysicalDevices
)
665 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
666 VK_OUTARRAY_MAKE(out
, pPhysicalDevices
, pPhysicalDeviceCount
);
668 VkResult result
= anv_instance_ensure_physical_device(instance
);
669 if (result
!= VK_SUCCESS
)
672 if (instance
->physicalDeviceCount
== 0)
675 assert(instance
->physicalDeviceCount
== 1);
676 vk_outarray_append(&out
, i
) {
677 *i
= anv_physical_device_to_handle(&instance
->physicalDevice
);
680 return vk_outarray_status(&out
);
683 VkResult
anv_EnumeratePhysicalDeviceGroups(
684 VkInstance _instance
,
685 uint32_t* pPhysicalDeviceGroupCount
,
686 VkPhysicalDeviceGroupProperties
* pPhysicalDeviceGroupProperties
)
688 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
689 VK_OUTARRAY_MAKE(out
, pPhysicalDeviceGroupProperties
,
690 pPhysicalDeviceGroupCount
);
692 VkResult result
= anv_instance_ensure_physical_device(instance
);
693 if (result
!= VK_SUCCESS
)
696 if (instance
->physicalDeviceCount
== 0)
699 assert(instance
->physicalDeviceCount
== 1);
701 vk_outarray_append(&out
, p
) {
702 p
->physicalDeviceCount
= 1;
703 memset(p
->physicalDevices
, 0, sizeof(p
->physicalDevices
));
704 p
->physicalDevices
[0] =
705 anv_physical_device_to_handle(&instance
->physicalDevice
);
706 p
->subsetAllocation
= VK_FALSE
;
708 vk_foreach_struct(ext
, p
->pNext
)
709 anv_debug_ignored_stype(ext
->sType
);
712 return vk_outarray_status(&out
);
715 void anv_GetPhysicalDeviceFeatures(
716 VkPhysicalDevice physicalDevice
,
717 VkPhysicalDeviceFeatures
* pFeatures
)
719 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
721 *pFeatures
= (VkPhysicalDeviceFeatures
) {
722 .robustBufferAccess
= true,
723 .fullDrawIndexUint32
= true,
724 .imageCubeArray
= true,
725 .independentBlend
= true,
726 .geometryShader
= true,
727 .tessellationShader
= true,
728 .sampleRateShading
= true,
729 .dualSrcBlend
= true,
731 .multiDrawIndirect
= true,
732 .drawIndirectFirstInstance
= true,
734 .depthBiasClamp
= true,
735 .fillModeNonSolid
= true,
736 .depthBounds
= false,
740 .multiViewport
= true,
741 .samplerAnisotropy
= true,
742 .textureCompressionETC2
= pdevice
->info
.gen
>= 8 ||
743 pdevice
->info
.is_baytrail
,
744 .textureCompressionASTC_LDR
= pdevice
->info
.gen
>= 9, /* FINISHME CHV */
745 .textureCompressionBC
= true,
746 .occlusionQueryPrecise
= true,
747 .pipelineStatisticsQuery
= true,
748 .fragmentStoresAndAtomics
= true,
749 .shaderTessellationAndGeometryPointSize
= true,
750 .shaderImageGatherExtended
= true,
751 .shaderStorageImageExtendedFormats
= true,
752 .shaderStorageImageMultisample
= false,
753 .shaderStorageImageReadWithoutFormat
= false,
754 .shaderStorageImageWriteWithoutFormat
= true,
755 .shaderUniformBufferArrayDynamicIndexing
= true,
756 .shaderSampledImageArrayDynamicIndexing
= true,
757 .shaderStorageBufferArrayDynamicIndexing
= true,
758 .shaderStorageImageArrayDynamicIndexing
= true,
759 .shaderClipDistance
= true,
760 .shaderCullDistance
= true,
761 .shaderFloat64
= pdevice
->info
.gen
>= 8 &&
762 pdevice
->info
.has_64bit_types
,
763 .shaderInt64
= pdevice
->info
.gen
>= 8 &&
764 pdevice
->info
.has_64bit_types
,
765 .shaderInt16
= pdevice
->info
.gen
>= 8,
766 .shaderResourceMinLod
= false,
767 .variableMultisampleRate
= true,
768 .inheritedQueries
= true,
771 /* We can't do image stores in vec4 shaders */
772 pFeatures
->vertexPipelineStoresAndAtomics
=
773 pdevice
->compiler
->scalar_stage
[MESA_SHADER_VERTEX
] &&
774 pdevice
->compiler
->scalar_stage
[MESA_SHADER_GEOMETRY
];
777 void anv_GetPhysicalDeviceFeatures2(
778 VkPhysicalDevice physicalDevice
,
779 VkPhysicalDeviceFeatures2
* pFeatures
)
781 anv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
783 vk_foreach_struct(ext
, pFeatures
->pNext
) {
784 switch (ext
->sType
) {
785 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES
: {
786 VkPhysicalDeviceProtectedMemoryFeatures
*features
= (void *)ext
;
787 features
->protectedMemory
= VK_FALSE
;
791 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES
: {
792 VkPhysicalDeviceMultiviewFeatures
*features
=
793 (VkPhysicalDeviceMultiviewFeatures
*)ext
;
794 features
->multiview
= true;
795 features
->multiviewGeometryShader
= true;
796 features
->multiviewTessellationShader
= true;
800 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTER_FEATURES
: {
801 VkPhysicalDeviceVariablePointerFeatures
*features
= (void *)ext
;
802 features
->variablePointersStorageBuffer
= true;
803 features
->variablePointers
= true;
807 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES
: {
808 VkPhysicalDeviceSamplerYcbcrConversionFeatures
*features
=
809 (VkPhysicalDeviceSamplerYcbcrConversionFeatures
*) ext
;
810 features
->samplerYcbcrConversion
= true;
814 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETER_FEATURES
: {
815 VkPhysicalDeviceShaderDrawParameterFeatures
*features
= (void *)ext
;
816 features
->shaderDrawParameters
= true;
820 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES_KHR
: {
821 VkPhysicalDevice16BitStorageFeaturesKHR
*features
=
822 (VkPhysicalDevice16BitStorageFeaturesKHR
*)ext
;
823 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
825 features
->storageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
826 features
->uniformAndStorageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
827 features
->storagePushConstant16
= pdevice
->info
.gen
>= 8;
828 features
->storageInputOutput16
= false;
833 anv_debug_ignored_stype(ext
->sType
);
839 void anv_GetPhysicalDeviceProperties(
840 VkPhysicalDevice physicalDevice
,
841 VkPhysicalDeviceProperties
* pProperties
)
843 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
844 const struct gen_device_info
*devinfo
= &pdevice
->info
;
846 /* See assertions made when programming the buffer surface state. */
847 const uint32_t max_raw_buffer_sz
= devinfo
->gen
>= 7 ?
848 (1ul << 30) : (1ul << 27);
850 const uint32_t max_samplers
= (devinfo
->gen
>= 8 || devinfo
->is_haswell
) ?
853 VkSampleCountFlags sample_counts
=
854 isl_device_get_sample_counts(&pdevice
->isl_dev
);
856 VkPhysicalDeviceLimits limits
= {
857 .maxImageDimension1D
= (1 << 14),
858 .maxImageDimension2D
= (1 << 14),
859 .maxImageDimension3D
= (1 << 11),
860 .maxImageDimensionCube
= (1 << 14),
861 .maxImageArrayLayers
= (1 << 11),
862 .maxTexelBufferElements
= 128 * 1024 * 1024,
863 .maxUniformBufferRange
= (1ul << 27),
864 .maxStorageBufferRange
= max_raw_buffer_sz
,
865 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
866 .maxMemoryAllocationCount
= UINT32_MAX
,
867 .maxSamplerAllocationCount
= 64 * 1024,
868 .bufferImageGranularity
= 64, /* A cache line */
869 .sparseAddressSpaceSize
= 0,
870 .maxBoundDescriptorSets
= MAX_SETS
,
871 .maxPerStageDescriptorSamplers
= max_samplers
,
872 .maxPerStageDescriptorUniformBuffers
= 64,
873 .maxPerStageDescriptorStorageBuffers
= 64,
874 .maxPerStageDescriptorSampledImages
= max_samplers
,
875 .maxPerStageDescriptorStorageImages
= 64,
876 .maxPerStageDescriptorInputAttachments
= 64,
877 .maxPerStageResources
= 250,
878 .maxDescriptorSetSamplers
= 6 * max_samplers
, /* number of stages * maxPerStageDescriptorSamplers */
879 .maxDescriptorSetUniformBuffers
= 6 * 64, /* number of stages * maxPerStageDescriptorUniformBuffers */
880 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
881 .maxDescriptorSetStorageBuffers
= 6 * 64, /* number of stages * maxPerStageDescriptorStorageBuffers */
882 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
883 .maxDescriptorSetSampledImages
= 6 * max_samplers
, /* number of stages * maxPerStageDescriptorSampledImages */
884 .maxDescriptorSetStorageImages
= 6 * 64, /* number of stages * maxPerStageDescriptorStorageImages */
885 .maxDescriptorSetInputAttachments
= 256,
886 .maxVertexInputAttributes
= MAX_VBS
,
887 .maxVertexInputBindings
= MAX_VBS
,
888 .maxVertexInputAttributeOffset
= 2047,
889 .maxVertexInputBindingStride
= 2048,
890 .maxVertexOutputComponents
= 128,
891 .maxTessellationGenerationLevel
= 64,
892 .maxTessellationPatchSize
= 32,
893 .maxTessellationControlPerVertexInputComponents
= 128,
894 .maxTessellationControlPerVertexOutputComponents
= 128,
895 .maxTessellationControlPerPatchOutputComponents
= 128,
896 .maxTessellationControlTotalOutputComponents
= 2048,
897 .maxTessellationEvaluationInputComponents
= 128,
898 .maxTessellationEvaluationOutputComponents
= 128,
899 .maxGeometryShaderInvocations
= 32,
900 .maxGeometryInputComponents
= 64,
901 .maxGeometryOutputComponents
= 128,
902 .maxGeometryOutputVertices
= 256,
903 .maxGeometryTotalOutputComponents
= 1024,
904 .maxFragmentInputComponents
= 128,
905 .maxFragmentOutputAttachments
= 8,
906 .maxFragmentDualSrcAttachments
= 1,
907 .maxFragmentCombinedOutputResources
= 8,
908 .maxComputeSharedMemorySize
= 32768,
909 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
910 .maxComputeWorkGroupInvocations
= 16 * devinfo
->max_cs_threads
,
911 .maxComputeWorkGroupSize
= {
912 16 * devinfo
->max_cs_threads
,
913 16 * devinfo
->max_cs_threads
,
914 16 * devinfo
->max_cs_threads
,
916 .subPixelPrecisionBits
= 4 /* FIXME */,
917 .subTexelPrecisionBits
= 4 /* FIXME */,
918 .mipmapPrecisionBits
= 4 /* FIXME */,
919 .maxDrawIndexedIndexValue
= UINT32_MAX
,
920 .maxDrawIndirectCount
= UINT32_MAX
,
921 .maxSamplerLodBias
= 16,
922 .maxSamplerAnisotropy
= 16,
923 .maxViewports
= MAX_VIEWPORTS
,
924 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
925 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
926 .viewportSubPixelBits
= 13, /* We take a float? */
927 .minMemoryMapAlignment
= 4096, /* A page */
928 .minTexelBufferOffsetAlignment
= 1,
929 /* We need 16 for UBO block reads to work and 32 for push UBOs */
930 .minUniformBufferOffsetAlignment
= 32,
931 .minStorageBufferOffsetAlignment
= 4,
932 .minTexelOffset
= -8,
934 .minTexelGatherOffset
= -32,
935 .maxTexelGatherOffset
= 31,
936 .minInterpolationOffset
= -0.5,
937 .maxInterpolationOffset
= 0.4375,
938 .subPixelInterpolationOffsetBits
= 4,
939 .maxFramebufferWidth
= (1 << 14),
940 .maxFramebufferHeight
= (1 << 14),
941 .maxFramebufferLayers
= (1 << 11),
942 .framebufferColorSampleCounts
= sample_counts
,
943 .framebufferDepthSampleCounts
= sample_counts
,
944 .framebufferStencilSampleCounts
= sample_counts
,
945 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
946 .maxColorAttachments
= MAX_RTS
,
947 .sampledImageColorSampleCounts
= sample_counts
,
948 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
949 .sampledImageDepthSampleCounts
= sample_counts
,
950 .sampledImageStencilSampleCounts
= sample_counts
,
951 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
952 .maxSampleMaskWords
= 1,
953 .timestampComputeAndGraphics
= false,
954 .timestampPeriod
= 1000000000.0 / devinfo
->timestamp_frequency
,
955 .maxClipDistances
= 8,
956 .maxCullDistances
= 8,
957 .maxCombinedClipAndCullDistances
= 8,
958 .discreteQueuePriorities
= 1,
959 .pointSizeRange
= { 0.125, 255.875 },
960 .lineWidthRange
= { 0.0, 7.9921875 },
961 .pointSizeGranularity
= (1.0 / 8.0),
962 .lineWidthGranularity
= (1.0 / 128.0),
963 .strictLines
= false, /* FINISHME */
964 .standardSampleLocations
= true,
965 .optimalBufferCopyOffsetAlignment
= 128,
966 .optimalBufferCopyRowPitchAlignment
= 128,
967 .nonCoherentAtomSize
= 64,
970 *pProperties
= (VkPhysicalDeviceProperties
) {
971 .apiVersion
= anv_physical_device_api_version(pdevice
),
972 .driverVersion
= vk_get_driver_version(),
974 .deviceID
= pdevice
->chipset_id
,
975 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
977 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
980 snprintf(pProperties
->deviceName
, sizeof(pProperties
->deviceName
),
981 "%s", pdevice
->name
);
982 memcpy(pProperties
->pipelineCacheUUID
,
983 pdevice
->pipeline_cache_uuid
, VK_UUID_SIZE
);
986 void anv_GetPhysicalDeviceProperties2(
987 VkPhysicalDevice physicalDevice
,
988 VkPhysicalDeviceProperties2
* pProperties
)
990 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
992 anv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
994 vk_foreach_struct(ext
, pProperties
->pNext
) {
995 switch (ext
->sType
) {
996 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
997 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
998 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
1000 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
1004 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES
: {
1005 VkPhysicalDeviceIDProperties
*id_props
=
1006 (VkPhysicalDeviceIDProperties
*)ext
;
1007 memcpy(id_props
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
1008 memcpy(id_props
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
1009 /* The LUID is for Windows. */
1010 id_props
->deviceLUIDValid
= false;
1014 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES
: {
1015 VkPhysicalDeviceMaintenance3Properties
*props
=
1016 (VkPhysicalDeviceMaintenance3Properties
*)ext
;
1017 /* This value doesn't matter for us today as our per-stage
1018 * descriptors are the real limit.
1020 props
->maxPerSetDescriptors
= 1024;
1021 props
->maxMemoryAllocationSize
= MAX_MEMORY_ALLOCATION_SIZE
;
1025 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES
: {
1026 VkPhysicalDeviceMultiviewProperties
*properties
=
1027 (VkPhysicalDeviceMultiviewProperties
*)ext
;
1028 properties
->maxMultiviewViewCount
= 16;
1029 properties
->maxMultiviewInstanceIndex
= UINT32_MAX
/ 16;
1033 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES
: {
1034 VkPhysicalDevicePointClippingProperties
*properties
=
1035 (VkPhysicalDevicePointClippingProperties
*) ext
;
1036 properties
->pointClippingBehavior
= VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES
;
1037 anv_finishme("Implement pop-free point clipping");
1041 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES
: {
1042 VkPhysicalDeviceSubgroupProperties
*properties
= (void *)ext
;
1044 properties
->subgroupSize
= BRW_SUBGROUP_SIZE
;
1046 VkShaderStageFlags scalar_stages
= 0;
1047 for (unsigned stage
= 0; stage
< MESA_SHADER_STAGES
; stage
++) {
1048 if (pdevice
->compiler
->scalar_stage
[stage
])
1049 scalar_stages
|= mesa_to_vk_shader_stage(stage
);
1051 properties
->supportedStages
= scalar_stages
;
1053 properties
->supportedOperations
= VK_SUBGROUP_FEATURE_BASIC_BIT
|
1054 VK_SUBGROUP_FEATURE_VOTE_BIT
|
1055 VK_SUBGROUP_FEATURE_ARITHMETIC_BIT
|
1056 VK_SUBGROUP_FEATURE_BALLOT_BIT
|
1057 VK_SUBGROUP_FEATURE_SHUFFLE_BIT
|
1058 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT
|
1059 VK_SUBGROUP_FEATURE_CLUSTERED_BIT
|
1060 VK_SUBGROUP_FEATURE_QUAD_BIT
;
1061 properties
->quadOperationsInAllStages
= VK_TRUE
;
1066 anv_debug_ignored_stype(ext
->sType
);
1072 /* We support exactly one queue family. */
1073 static const VkQueueFamilyProperties
1074 anv_queue_family_properties
= {
1075 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
1076 VK_QUEUE_COMPUTE_BIT
|
1077 VK_QUEUE_TRANSFER_BIT
,
1079 .timestampValidBits
= 36, /* XXX: Real value here */
1080 .minImageTransferGranularity
= { 1, 1, 1 },
1083 void anv_GetPhysicalDeviceQueueFamilyProperties(
1084 VkPhysicalDevice physicalDevice
,
1086 VkQueueFamilyProperties
* pQueueFamilyProperties
)
1088 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pCount
);
1090 vk_outarray_append(&out
, p
) {
1091 *p
= anv_queue_family_properties
;
1095 void anv_GetPhysicalDeviceQueueFamilyProperties2(
1096 VkPhysicalDevice physicalDevice
,
1097 uint32_t* pQueueFamilyPropertyCount
,
1098 VkQueueFamilyProperties2
* pQueueFamilyProperties
)
1101 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pQueueFamilyPropertyCount
);
1103 vk_outarray_append(&out
, p
) {
1104 p
->queueFamilyProperties
= anv_queue_family_properties
;
1106 vk_foreach_struct(s
, p
->pNext
) {
1107 anv_debug_ignored_stype(s
->sType
);
1112 void anv_GetPhysicalDeviceMemoryProperties(
1113 VkPhysicalDevice physicalDevice
,
1114 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
1116 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
1118 pMemoryProperties
->memoryTypeCount
= physical_device
->memory
.type_count
;
1119 for (uint32_t i
= 0; i
< physical_device
->memory
.type_count
; i
++) {
1120 pMemoryProperties
->memoryTypes
[i
] = (VkMemoryType
) {
1121 .propertyFlags
= physical_device
->memory
.types
[i
].propertyFlags
,
1122 .heapIndex
= physical_device
->memory
.types
[i
].heapIndex
,
1126 pMemoryProperties
->memoryHeapCount
= physical_device
->memory
.heap_count
;
1127 for (uint32_t i
= 0; i
< physical_device
->memory
.heap_count
; i
++) {
1128 pMemoryProperties
->memoryHeaps
[i
] = (VkMemoryHeap
) {
1129 .size
= physical_device
->memory
.heaps
[i
].size
,
1130 .flags
= physical_device
->memory
.heaps
[i
].flags
,
1135 void anv_GetPhysicalDeviceMemoryProperties2(
1136 VkPhysicalDevice physicalDevice
,
1137 VkPhysicalDeviceMemoryProperties2
* pMemoryProperties
)
1139 anv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
1140 &pMemoryProperties
->memoryProperties
);
1142 vk_foreach_struct(ext
, pMemoryProperties
->pNext
) {
1143 switch (ext
->sType
) {
1145 anv_debug_ignored_stype(ext
->sType
);
1152 anv_GetDeviceGroupPeerMemoryFeatures(
1155 uint32_t localDeviceIndex
,
1156 uint32_t remoteDeviceIndex
,
1157 VkPeerMemoryFeatureFlags
* pPeerMemoryFeatures
)
1159 assert(localDeviceIndex
== 0 && remoteDeviceIndex
== 0);
1160 *pPeerMemoryFeatures
= VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT
|
1161 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT
|
1162 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT
|
1163 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT
;
1166 PFN_vkVoidFunction
anv_GetInstanceProcAddr(
1167 VkInstance _instance
,
1170 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1172 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
1173 * when we have to return valid function pointers, NULL, or it's left
1174 * undefined. See the table for exact details.
1179 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
1180 if (strcmp(pName, "vk" #entrypoint) == 0) \
1181 return (PFN_vkVoidFunction)anv_##entrypoint
1183 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties
);
1184 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties
);
1185 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceVersion
);
1186 LOOKUP_ANV_ENTRYPOINT(CreateInstance
);
1188 #undef LOOKUP_ANV_ENTRYPOINT
1190 if (instance
== NULL
)
1193 int idx
= anv_get_entrypoint_index(pName
);
1197 return instance
->dispatch
.entrypoints
[idx
];
1200 /* With version 1+ of the loader interface the ICD should expose
1201 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
1204 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1205 VkInstance instance
,
1209 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1210 VkInstance instance
,
1213 return anv_GetInstanceProcAddr(instance
, pName
);
1216 PFN_vkVoidFunction
anv_GetDeviceProcAddr(
1220 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1222 if (!device
|| !pName
)
1225 int idx
= anv_get_entrypoint_index(pName
);
1229 return device
->dispatch
.entrypoints
[idx
];
1233 anv_CreateDebugReportCallbackEXT(VkInstance _instance
,
1234 const VkDebugReportCallbackCreateInfoEXT
* pCreateInfo
,
1235 const VkAllocationCallbacks
* pAllocator
,
1236 VkDebugReportCallbackEXT
* pCallback
)
1238 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1239 return vk_create_debug_report_callback(&instance
->debug_report_callbacks
,
1240 pCreateInfo
, pAllocator
, &instance
->alloc
,
1245 anv_DestroyDebugReportCallbackEXT(VkInstance _instance
,
1246 VkDebugReportCallbackEXT _callback
,
1247 const VkAllocationCallbacks
* pAllocator
)
1249 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1250 vk_destroy_debug_report_callback(&instance
->debug_report_callbacks
,
1251 _callback
, pAllocator
, &instance
->alloc
);
1255 anv_DebugReportMessageEXT(VkInstance _instance
,
1256 VkDebugReportFlagsEXT flags
,
1257 VkDebugReportObjectTypeEXT objectType
,
1260 int32_t messageCode
,
1261 const char* pLayerPrefix
,
1262 const char* pMessage
)
1264 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1265 vk_debug_report(&instance
->debug_report_callbacks
, flags
, objectType
,
1266 object
, location
, messageCode
, pLayerPrefix
, pMessage
);
1270 anv_queue_init(struct anv_device
*device
, struct anv_queue
*queue
)
1272 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1273 queue
->device
= device
;
1278 anv_queue_finish(struct anv_queue
*queue
)
1282 static struct anv_state
1283 anv_state_pool_emit_data(struct anv_state_pool
*pool
, size_t size
, size_t align
, const void *p
)
1285 struct anv_state state
;
1287 state
= anv_state_pool_alloc(pool
, size
, align
);
1288 memcpy(state
.map
, p
, size
);
1290 anv_state_flush(pool
->block_pool
.device
, state
);
1295 struct gen8_border_color
{
1300 /* Pad out to 64 bytes */
1305 anv_device_init_border_colors(struct anv_device
*device
)
1307 static const struct gen8_border_color border_colors
[] = {
1308 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
1309 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
1310 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
1311 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
1312 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
1313 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
1316 device
->border_colors
= anv_state_pool_emit_data(&device
->dynamic_state_pool
,
1317 sizeof(border_colors
), 64,
1322 anv_device_init_trivial_batch(struct anv_device
*device
)
1324 anv_bo_init_new(&device
->trivial_batch_bo
, device
, 4096);
1326 if (device
->instance
->physicalDevice
.has_exec_async
)
1327 device
->trivial_batch_bo
.flags
|= EXEC_OBJECT_ASYNC
;
1329 void *map
= anv_gem_mmap(device
, device
->trivial_batch_bo
.gem_handle
,
1332 struct anv_batch batch
= {
1338 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
1339 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
1341 if (!device
->info
.has_llc
)
1342 gen_clflush_range(map
, batch
.next
- map
);
1344 anv_gem_munmap(map
, device
->trivial_batch_bo
.size
);
1347 VkResult
anv_EnumerateDeviceExtensionProperties(
1348 VkPhysicalDevice physicalDevice
,
1349 const char* pLayerName
,
1350 uint32_t* pPropertyCount
,
1351 VkExtensionProperties
* pProperties
)
1353 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
1354 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
1357 for (int i
= 0; i
< ANV_DEVICE_EXTENSION_COUNT
; i
++) {
1358 if (device
->supported_extensions
.extensions
[i
]) {
1359 vk_outarray_append(&out
, prop
) {
1360 *prop
= anv_device_extensions
[i
];
1365 return vk_outarray_status(&out
);
1369 anv_device_init_dispatch(struct anv_device
*device
)
1371 const struct anv_dispatch_table
*genX_table
;
1372 switch (device
->info
.gen
) {
1374 genX_table
= &gen11_dispatch_table
;
1377 genX_table
= &gen10_dispatch_table
;
1380 genX_table
= &gen9_dispatch_table
;
1383 genX_table
= &gen8_dispatch_table
;
1386 if (device
->info
.is_haswell
)
1387 genX_table
= &gen75_dispatch_table
;
1389 genX_table
= &gen7_dispatch_table
;
1392 unreachable("unsupported gen\n");
1395 for (unsigned i
= 0; i
< ARRAY_SIZE(device
->dispatch
.entrypoints
); i
++) {
1396 /* Vulkan requires that entrypoints for extensions which have not been
1397 * enabled must not be advertised.
1399 if (!anv_entrypoint_is_enabled(i
, device
->instance
->apiVersion
,
1400 &device
->instance
->enabled_extensions
,
1401 &device
->enabled_extensions
)) {
1402 device
->dispatch
.entrypoints
[i
] = NULL
;
1403 } else if (genX_table
->entrypoints
[i
]) {
1404 device
->dispatch
.entrypoints
[i
] = genX_table
->entrypoints
[i
];
1406 device
->dispatch
.entrypoints
[i
] = anv_dispatch_table
.entrypoints
[i
];
1412 vk_priority_to_gen(int priority
)
1415 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT
:
1416 return GEN_CONTEXT_LOW_PRIORITY
;
1417 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
:
1418 return GEN_CONTEXT_MEDIUM_PRIORITY
;
1419 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT
:
1420 return GEN_CONTEXT_HIGH_PRIORITY
;
1421 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT
:
1422 return GEN_CONTEXT_REALTIME_PRIORITY
;
1424 unreachable("Invalid priority");
1429 anv_device_init_hiz_clear_batch(struct anv_device
*device
)
1431 anv_bo_init_new(&device
->hiz_clear_bo
, device
, 4096);
1432 uint32_t *map
= anv_gem_mmap(device
, device
->hiz_clear_bo
.gem_handle
,
1435 union isl_color_value hiz_clear
= { .u32
= { 0, } };
1436 hiz_clear
.f32
[0] = ANV_HZ_FC_VAL
;
1438 memcpy(map
, hiz_clear
.u32
, sizeof(hiz_clear
.u32
));
1439 anv_gem_munmap(map
, device
->hiz_clear_bo
.size
);
1442 VkResult
anv_CreateDevice(
1443 VkPhysicalDevice physicalDevice
,
1444 const VkDeviceCreateInfo
* pCreateInfo
,
1445 const VkAllocationCallbacks
* pAllocator
,
1448 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
1450 struct anv_device
*device
;
1452 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO
);
1454 struct anv_device_extension_table enabled_extensions
= { };
1455 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
1457 for (idx
= 0; idx
< ANV_DEVICE_EXTENSION_COUNT
; idx
++) {
1458 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
1459 anv_device_extensions
[idx
].extensionName
) == 0)
1463 if (idx
>= ANV_DEVICE_EXTENSION_COUNT
)
1464 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
1466 if (!physical_device
->supported_extensions
.extensions
[idx
])
1467 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
1469 enabled_extensions
.extensions
[idx
] = true;
1472 /* Check enabled features */
1473 if (pCreateInfo
->pEnabledFeatures
) {
1474 VkPhysicalDeviceFeatures supported_features
;
1475 anv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
1476 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
1477 VkBool32
*enabled_feature
= (VkBool32
*)pCreateInfo
->pEnabledFeatures
;
1478 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
1479 for (uint32_t i
= 0; i
< num_features
; i
++) {
1480 if (enabled_feature
[i
] && !supported_feature
[i
])
1481 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
1485 /* Check requested queues and fail if we are requested to create any
1486 * queues with flags we don't support.
1488 assert(pCreateInfo
->queueCreateInfoCount
> 0);
1489 for (uint32_t i
= 0; i
< pCreateInfo
->queueCreateInfoCount
; i
++) {
1490 if (pCreateInfo
->pQueueCreateInfos
[i
].flags
!= 0)
1491 return vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1494 /* Check if client specified queue priority. */
1495 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*queue_priority
=
1496 vk_find_struct_const(pCreateInfo
->pQueueCreateInfos
[0].pNext
,
1497 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT
);
1499 VkQueueGlobalPriorityEXT priority
=
1500 queue_priority
? queue_priority
->globalPriority
:
1501 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
;
1503 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
1505 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
1507 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1509 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1510 device
->instance
= physical_device
->instance
;
1511 device
->chipset_id
= physical_device
->chipset_id
;
1512 device
->no_hw
= physical_device
->no_hw
;
1513 device
->lost
= false;
1516 device
->alloc
= *pAllocator
;
1518 device
->alloc
= physical_device
->instance
->alloc
;
1520 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
1521 device
->fd
= open(physical_device
->path
, O_RDWR
| O_CLOEXEC
);
1522 if (device
->fd
== -1) {
1523 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1527 device
->context_id
= anv_gem_create_context(device
);
1528 if (device
->context_id
== -1) {
1529 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1533 if (physical_device
->use_softpin
) {
1534 if (pthread_mutex_init(&device
->vma_mutex
, NULL
) != 0) {
1535 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1539 /* keep the page with address zero out of the allocator */
1540 util_vma_heap_init(&device
->vma_lo
, LOW_HEAP_MIN_ADDRESS
, LOW_HEAP_SIZE
);
1541 device
->vma_lo_available
=
1542 physical_device
->memory
.heaps
[physical_device
->memory
.heap_count
- 1].size
;
1544 /* Leave the last 4GiB out of the high vma range, so that no state base
1545 * address + size can overflow 48 bits. For more information see the
1546 * comment about Wa32bitGeneralStateOffset in anv_allocator.c
1548 util_vma_heap_init(&device
->vma_hi
, HIGH_HEAP_MIN_ADDRESS
,
1550 device
->vma_hi_available
= physical_device
->memory
.heap_count
== 1 ? 0 :
1551 physical_device
->memory
.heaps
[0].size
;
1554 /* As per spec, the driver implementation may deny requests to acquire
1555 * a priority above the default priority (MEDIUM) if the caller does not
1556 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
1559 if (physical_device
->has_context_priority
) {
1560 int err
= anv_gem_set_context_param(device
->fd
, device
->context_id
,
1561 I915_CONTEXT_PARAM_PRIORITY
,
1562 vk_priority_to_gen(priority
));
1563 if (err
!= 0 && priority
> VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
) {
1564 result
= vk_error(VK_ERROR_NOT_PERMITTED_EXT
);
1569 device
->info
= physical_device
->info
;
1570 device
->isl_dev
= physical_device
->isl_dev
;
1572 /* On Broadwell and later, we can use batch chaining to more efficiently
1573 * implement growing command buffers. Prior to Haswell, the kernel
1574 * command parser gets in the way and we have to fall back to growing
1577 device
->can_chain_batches
= device
->info
.gen
>= 8;
1579 device
->robust_buffer_access
= pCreateInfo
->pEnabledFeatures
&&
1580 pCreateInfo
->pEnabledFeatures
->robustBufferAccess
;
1581 device
->enabled_extensions
= enabled_extensions
;
1583 anv_device_init_dispatch(device
);
1585 if (pthread_mutex_init(&device
->mutex
, NULL
) != 0) {
1586 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1587 goto fail_context_id
;
1590 pthread_condattr_t condattr
;
1591 if (pthread_condattr_init(&condattr
) != 0) {
1592 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1595 if (pthread_condattr_setclock(&condattr
, CLOCK_MONOTONIC
) != 0) {
1596 pthread_condattr_destroy(&condattr
);
1597 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1600 if (pthread_cond_init(&device
->queue_submit
, NULL
) != 0) {
1601 pthread_condattr_destroy(&condattr
);
1602 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1605 pthread_condattr_destroy(&condattr
);
1608 (physical_device
->supports_48bit_addresses
? EXEC_OBJECT_SUPPORTS_48B_ADDRESS
: 0) |
1609 (physical_device
->has_exec_async
? EXEC_OBJECT_ASYNC
: 0) |
1610 (physical_device
->has_exec_capture
? EXEC_OBJECT_CAPTURE
: 0);
1612 anv_bo_pool_init(&device
->batch_bo_pool
, device
, bo_flags
);
1614 result
= anv_bo_cache_init(&device
->bo_cache
);
1615 if (result
!= VK_SUCCESS
)
1616 goto fail_batch_bo_pool
;
1618 if (physical_device
->use_softpin
)
1619 bo_flags
|= EXEC_OBJECT_PINNED
;
1621 bo_flags
&= ~EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
1623 result
= anv_state_pool_init(&device
->dynamic_state_pool
, device
,
1624 DYNAMIC_STATE_POOL_MIN_ADDRESS
,
1627 if (result
!= VK_SUCCESS
)
1630 result
= anv_state_pool_init(&device
->instruction_state_pool
, device
,
1631 INSTRUCTION_STATE_POOL_MIN_ADDRESS
,
1634 if (result
!= VK_SUCCESS
)
1635 goto fail_dynamic_state_pool
;
1637 result
= anv_state_pool_init(&device
->surface_state_pool
, device
,
1638 SURFACE_STATE_POOL_MIN_ADDRESS
,
1641 if (result
!= VK_SUCCESS
)
1642 goto fail_instruction_state_pool
;
1644 result
= anv_bo_init_new(&device
->workaround_bo
, device
, 1024);
1645 if (result
!= VK_SUCCESS
)
1646 goto fail_surface_state_pool
;
1648 anv_device_init_trivial_batch(device
);
1650 if (device
->info
.gen
>= 10)
1651 anv_device_init_hiz_clear_batch(device
);
1653 anv_scratch_pool_init(device
, &device
->scratch_pool
);
1655 anv_queue_init(device
, &device
->queue
);
1657 switch (device
->info
.gen
) {
1659 if (!device
->info
.is_haswell
)
1660 result
= gen7_init_device_state(device
);
1662 result
= gen75_init_device_state(device
);
1665 result
= gen8_init_device_state(device
);
1668 result
= gen9_init_device_state(device
);
1671 result
= gen10_init_device_state(device
);
1674 result
= gen11_init_device_state(device
);
1677 /* Shouldn't get here as we don't create physical devices for any other
1679 unreachable("unhandled gen");
1681 if (result
!= VK_SUCCESS
)
1682 goto fail_workaround_bo
;
1684 anv_device_init_blorp(device
);
1686 anv_device_init_border_colors(device
);
1688 *pDevice
= anv_device_to_handle(device
);
1693 anv_queue_finish(&device
->queue
);
1694 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
1695 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
1696 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
1697 fail_surface_state_pool
:
1698 anv_state_pool_finish(&device
->surface_state_pool
);
1699 fail_instruction_state_pool
:
1700 anv_state_pool_finish(&device
->instruction_state_pool
);
1701 fail_dynamic_state_pool
:
1702 anv_state_pool_finish(&device
->dynamic_state_pool
);
1704 anv_bo_cache_finish(&device
->bo_cache
);
1706 anv_bo_pool_finish(&device
->batch_bo_pool
);
1707 pthread_cond_destroy(&device
->queue_submit
);
1709 pthread_mutex_destroy(&device
->mutex
);
1711 anv_gem_destroy_context(device
, device
->context_id
);
1715 vk_free(&device
->alloc
, device
);
1720 void anv_DestroyDevice(
1722 const VkAllocationCallbacks
* pAllocator
)
1724 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1729 anv_device_finish_blorp(device
);
1731 anv_queue_finish(&device
->queue
);
1733 #ifdef HAVE_VALGRIND
1734 /* We only need to free these to prevent valgrind errors. The backing
1735 * BO will go away in a couple of lines so we don't actually leak.
1737 anv_state_pool_free(&device
->dynamic_state_pool
, device
->border_colors
);
1740 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
1742 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
1743 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
1745 anv_gem_close(device
, device
->trivial_batch_bo
.gem_handle
);
1746 if (device
->info
.gen
>= 10)
1747 anv_gem_close(device
, device
->hiz_clear_bo
.gem_handle
);
1749 anv_state_pool_finish(&device
->surface_state_pool
);
1750 anv_state_pool_finish(&device
->instruction_state_pool
);
1751 anv_state_pool_finish(&device
->dynamic_state_pool
);
1753 anv_bo_cache_finish(&device
->bo_cache
);
1755 anv_bo_pool_finish(&device
->batch_bo_pool
);
1757 pthread_cond_destroy(&device
->queue_submit
);
1758 pthread_mutex_destroy(&device
->mutex
);
1760 anv_gem_destroy_context(device
, device
->context_id
);
1764 vk_free(&device
->alloc
, device
);
1767 VkResult
anv_EnumerateInstanceLayerProperties(
1768 uint32_t* pPropertyCount
,
1769 VkLayerProperties
* pProperties
)
1771 if (pProperties
== NULL
) {
1772 *pPropertyCount
= 0;
1776 /* None supported at this time */
1777 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1780 VkResult
anv_EnumerateDeviceLayerProperties(
1781 VkPhysicalDevice physicalDevice
,
1782 uint32_t* pPropertyCount
,
1783 VkLayerProperties
* pProperties
)
1785 if (pProperties
== NULL
) {
1786 *pPropertyCount
= 0;
1790 /* None supported at this time */
1791 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1794 void anv_GetDeviceQueue(
1796 uint32_t queueNodeIndex
,
1797 uint32_t queueIndex
,
1800 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1802 assert(queueIndex
== 0);
1804 *pQueue
= anv_queue_to_handle(&device
->queue
);
1807 void anv_GetDeviceQueue2(
1809 const VkDeviceQueueInfo2
* pQueueInfo
,
1812 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1814 assert(pQueueInfo
->queueIndex
== 0);
1816 if (pQueueInfo
->flags
== device
->queue
.flags
)
1817 *pQueue
= anv_queue_to_handle(&device
->queue
);
1823 anv_device_query_status(struct anv_device
*device
)
1825 /* This isn't likely as most of the callers of this function already check
1826 * for it. However, it doesn't hurt to check and it potentially lets us
1829 if (unlikely(device
->lost
))
1830 return VK_ERROR_DEVICE_LOST
;
1832 uint32_t active
, pending
;
1833 int ret
= anv_gem_gpu_get_reset_stats(device
, &active
, &pending
);
1835 /* We don't know the real error. */
1836 device
->lost
= true;
1837 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1838 "get_reset_stats failed: %m");
1842 device
->lost
= true;
1843 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1844 "GPU hung on one of our command buffers");
1845 } else if (pending
) {
1846 device
->lost
= true;
1847 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1848 "GPU hung with commands in-flight");
1855 anv_device_bo_busy(struct anv_device
*device
, struct anv_bo
*bo
)
1857 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
1858 * Other usages of the BO (such as on different hardware) will not be
1859 * flagged as "busy" by this ioctl. Use with care.
1861 int ret
= anv_gem_busy(device
, bo
->gem_handle
);
1863 return VK_NOT_READY
;
1864 } else if (ret
== -1) {
1865 /* We don't know the real error. */
1866 device
->lost
= true;
1867 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1868 "gem wait failed: %m");
1871 /* Query for device status after the busy call. If the BO we're checking
1872 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
1873 * client because it clearly doesn't have valid data. Yes, this most
1874 * likely means an ioctl, but we just did an ioctl to query the busy status
1875 * so it's no great loss.
1877 return anv_device_query_status(device
);
1881 anv_device_wait(struct anv_device
*device
, struct anv_bo
*bo
,
1884 int ret
= anv_gem_wait(device
, bo
->gem_handle
, &timeout
);
1885 if (ret
== -1 && errno
== ETIME
) {
1887 } else if (ret
== -1) {
1888 /* We don't know the real error. */
1889 device
->lost
= true;
1890 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1891 "gem wait failed: %m");
1894 /* Query for device status after the wait. If the BO we're waiting on got
1895 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
1896 * because it clearly doesn't have valid data. Yes, this most likely means
1897 * an ioctl, but we just did an ioctl to wait so it's no great loss.
1899 return anv_device_query_status(device
);
1902 VkResult
anv_DeviceWaitIdle(
1905 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1906 if (unlikely(device
->lost
))
1907 return VK_ERROR_DEVICE_LOST
;
1909 struct anv_batch batch
;
1912 batch
.start
= batch
.next
= cmds
;
1913 batch
.end
= (void *) cmds
+ sizeof(cmds
);
1915 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
1916 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
1918 return anv_device_submit_simple_batch(device
, &batch
);
1922 anv_vma_alloc(struct anv_device
*device
, struct anv_bo
*bo
)
1924 if (!(bo
->flags
& EXEC_OBJECT_PINNED
))
1927 pthread_mutex_lock(&device
->vma_mutex
);
1931 if (bo
->flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
&&
1932 device
->vma_hi_available
>= bo
->size
) {
1933 uint64_t addr
= util_vma_heap_alloc(&device
->vma_hi
, bo
->size
, 4096);
1935 bo
->offset
= gen_canonical_address(addr
);
1936 assert(addr
== gen_48b_address(bo
->offset
));
1937 device
->vma_hi_available
-= bo
->size
;
1941 if (bo
->offset
== 0 && device
->vma_lo_available
>= bo
->size
) {
1942 uint64_t addr
= util_vma_heap_alloc(&device
->vma_lo
, bo
->size
, 4096);
1944 bo
->offset
= gen_canonical_address(addr
);
1945 assert(addr
== gen_48b_address(bo
->offset
));
1946 device
->vma_lo_available
-= bo
->size
;
1950 pthread_mutex_unlock(&device
->vma_mutex
);
1952 return bo
->offset
!= 0;
1956 anv_vma_free(struct anv_device
*device
, struct anv_bo
*bo
)
1958 if (!(bo
->flags
& EXEC_OBJECT_PINNED
))
1961 const uint64_t addr_48b
= gen_48b_address(bo
->offset
);
1963 pthread_mutex_lock(&device
->vma_mutex
);
1965 if (addr_48b
>= LOW_HEAP_MIN_ADDRESS
&&
1966 addr_48b
<= LOW_HEAP_MAX_ADDRESS
) {
1967 util_vma_heap_free(&device
->vma_lo
, addr_48b
, bo
->size
);
1968 device
->vma_lo_available
+= bo
->size
;
1970 assert(addr_48b
>= HIGH_HEAP_MIN_ADDRESS
&&
1971 addr_48b
<= HIGH_HEAP_MAX_ADDRESS
);
1972 util_vma_heap_free(&device
->vma_hi
, addr_48b
, bo
->size
);
1973 device
->vma_hi_available
+= bo
->size
;
1976 pthread_mutex_unlock(&device
->vma_mutex
);
1982 anv_bo_init_new(struct anv_bo
*bo
, struct anv_device
*device
, uint64_t size
)
1984 uint32_t gem_handle
= anv_gem_create(device
, size
);
1986 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
1988 anv_bo_init(bo
, gem_handle
, size
);
1993 VkResult
anv_AllocateMemory(
1995 const VkMemoryAllocateInfo
* pAllocateInfo
,
1996 const VkAllocationCallbacks
* pAllocator
,
1997 VkDeviceMemory
* pMem
)
1999 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2000 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2001 struct anv_device_memory
*mem
;
2002 VkResult result
= VK_SUCCESS
;
2004 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
2006 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
2007 assert(pAllocateInfo
->allocationSize
> 0);
2009 if (pAllocateInfo
->allocationSize
> MAX_MEMORY_ALLOCATION_SIZE
)
2010 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
2012 /* FINISHME: Fail if allocation request exceeds heap size. */
2014 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
2015 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2017 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2019 assert(pAllocateInfo
->memoryTypeIndex
< pdevice
->memory
.type_count
);
2020 mem
->type
= &pdevice
->memory
.types
[pAllocateInfo
->memoryTypeIndex
];
2024 const VkImportMemoryFdInfoKHR
*fd_info
=
2025 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_FD_INFO_KHR
);
2027 /* The Vulkan spec permits handleType to be 0, in which case the struct is
2030 if (fd_info
&& fd_info
->handleType
) {
2031 /* At the moment, we support only the below handle types. */
2032 assert(fd_info
->handleType
==
2033 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
2034 fd_info
->handleType
==
2035 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
2037 result
= anv_bo_cache_import(device
, &device
->bo_cache
,
2038 fd_info
->fd
, &mem
->bo
);
2039 if (result
!= VK_SUCCESS
)
2042 VkDeviceSize aligned_alloc_size
=
2043 align_u64(pAllocateInfo
->allocationSize
, 4096);
2045 /* For security purposes, we reject importing the bo if it's smaller
2046 * than the requested allocation size. This prevents a malicious client
2047 * from passing a buffer to a trusted client, lying about the size, and
2048 * telling the trusted client to try and texture from an image that goes
2049 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
2050 * in the trusted client. The trusted client can protect itself against
2051 * this sort of attack but only if it can trust the buffer size.
2053 if (mem
->bo
->size
< aligned_alloc_size
) {
2054 result
= vk_errorf(device
->instance
, device
,
2055 VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
,
2056 "aligned allocationSize too large for "
2057 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR: "
2058 "%"PRIu64
"B > %"PRIu64
"B",
2059 aligned_alloc_size
, mem
->bo
->size
);
2060 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
2064 /* From the Vulkan spec:
2066 * "Importing memory from a file descriptor transfers ownership of
2067 * the file descriptor from the application to the Vulkan
2068 * implementation. The application must not perform any operations on
2069 * the file descriptor after a successful import."
2071 * If the import fails, we leave the file descriptor open.
2075 result
= anv_bo_cache_alloc(device
, &device
->bo_cache
,
2076 pAllocateInfo
->allocationSize
,
2078 if (result
!= VK_SUCCESS
)
2081 const VkMemoryDedicatedAllocateInfoKHR
*dedicated_info
=
2082 vk_find_struct_const(pAllocateInfo
->pNext
, MEMORY_DEDICATED_ALLOCATE_INFO_KHR
);
2083 if (dedicated_info
&& dedicated_info
->image
!= VK_NULL_HANDLE
) {
2084 ANV_FROM_HANDLE(anv_image
, image
, dedicated_info
->image
);
2086 /* Some legacy (non-modifiers) consumers need the tiling to be set on
2087 * the BO. In this case, we have a dedicated allocation.
2089 if (image
->needs_set_tiling
) {
2090 const uint32_t i915_tiling
=
2091 isl_tiling_to_i915_tiling(image
->planes
[0].surface
.isl
.tiling
);
2092 int ret
= anv_gem_set_tiling(device
, mem
->bo
->gem_handle
,
2093 image
->planes
[0].surface
.isl
.row_pitch
,
2096 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
2097 return vk_errorf(device
->instance
, NULL
,
2098 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
2099 "failed to set BO tiling: %m");
2105 assert(mem
->type
->heapIndex
< pdevice
->memory
.heap_count
);
2106 if (pdevice
->memory
.heaps
[mem
->type
->heapIndex
].supports_48bit_addresses
)
2107 mem
->bo
->flags
|= EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
2109 const struct wsi_memory_allocate_info
*wsi_info
=
2110 vk_find_struct_const(pAllocateInfo
->pNext
, WSI_MEMORY_ALLOCATE_INFO_MESA
);
2111 if (wsi_info
&& wsi_info
->implicit_sync
) {
2112 /* We need to set the WRITE flag on window system buffers so that GEM
2113 * will know we're writing to them and synchronize uses on other rings
2114 * (eg if the display server uses the blitter ring).
2116 mem
->bo
->flags
|= EXEC_OBJECT_WRITE
;
2117 } else if (pdevice
->has_exec_async
) {
2118 mem
->bo
->flags
|= EXEC_OBJECT_ASYNC
;
2121 *pMem
= anv_device_memory_to_handle(mem
);
2126 vk_free2(&device
->alloc
, pAllocator
, mem
);
2131 VkResult
anv_GetMemoryFdKHR(
2133 const VkMemoryGetFdInfoKHR
* pGetFdInfo
,
2136 ANV_FROM_HANDLE(anv_device
, dev
, device_h
);
2137 ANV_FROM_HANDLE(anv_device_memory
, mem
, pGetFdInfo
->memory
);
2139 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
2141 assert(pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
2142 pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
2144 return anv_bo_cache_export(dev
, &dev
->bo_cache
, mem
->bo
, pFd
);
2147 VkResult
anv_GetMemoryFdPropertiesKHR(
2149 VkExternalMemoryHandleTypeFlagBitsKHR handleType
,
2151 VkMemoryFdPropertiesKHR
* pMemoryFdProperties
)
2153 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2154 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2156 switch (handleType
) {
2157 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
:
2158 /* dma-buf can be imported as any memory type */
2159 pMemoryFdProperties
->memoryTypeBits
=
2160 (1 << pdevice
->memory
.type_count
) - 1;
2164 /* The valid usage section for this function says:
2166 * "handleType must not be one of the handle types defined as
2169 * So opaque handle types fall into the default "unsupported" case.
2171 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
2175 void anv_FreeMemory(
2177 VkDeviceMemory _mem
,
2178 const VkAllocationCallbacks
* pAllocator
)
2180 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2181 ANV_FROM_HANDLE(anv_device_memory
, mem
, _mem
);
2187 anv_UnmapMemory(_device
, _mem
);
2189 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
2191 vk_free2(&device
->alloc
, pAllocator
, mem
);
2194 VkResult
anv_MapMemory(
2196 VkDeviceMemory _memory
,
2197 VkDeviceSize offset
,
2199 VkMemoryMapFlags flags
,
2202 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2203 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
2210 if (size
== VK_WHOLE_SIZE
)
2211 size
= mem
->bo
->size
- offset
;
2213 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
2215 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
2216 * assert(size != 0);
2217 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
2218 * equal to the size of the memory minus offset
2221 assert(offset
+ size
<= mem
->bo
->size
);
2223 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
2224 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
2225 * at a time is valid. We could just mmap up front and return an offset
2226 * pointer here, but that may exhaust virtual memory on 32 bit
2229 uint32_t gem_flags
= 0;
2231 if (!device
->info
.has_llc
&&
2232 (mem
->type
->propertyFlags
& VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
))
2233 gem_flags
|= I915_MMAP_WC
;
2235 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
2236 uint64_t map_offset
= offset
& ~4095ull;
2237 assert(offset
>= map_offset
);
2238 uint64_t map_size
= (offset
+ size
) - map_offset
;
2240 /* Let's map whole pages */
2241 map_size
= align_u64(map_size
, 4096);
2243 void *map
= anv_gem_mmap(device
, mem
->bo
->gem_handle
,
2244 map_offset
, map_size
, gem_flags
);
2245 if (map
== MAP_FAILED
)
2246 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
2249 mem
->map_size
= map_size
;
2251 *ppData
= mem
->map
+ (offset
- map_offset
);
2256 void anv_UnmapMemory(
2258 VkDeviceMemory _memory
)
2260 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
2265 anv_gem_munmap(mem
->map
, mem
->map_size
);
2272 clflush_mapped_ranges(struct anv_device
*device
,
2274 const VkMappedMemoryRange
*ranges
)
2276 for (uint32_t i
= 0; i
< count
; i
++) {
2277 ANV_FROM_HANDLE(anv_device_memory
, mem
, ranges
[i
].memory
);
2278 if (ranges
[i
].offset
>= mem
->map_size
)
2281 gen_clflush_range(mem
->map
+ ranges
[i
].offset
,
2282 MIN2(ranges
[i
].size
, mem
->map_size
- ranges
[i
].offset
));
2286 VkResult
anv_FlushMappedMemoryRanges(
2288 uint32_t memoryRangeCount
,
2289 const VkMappedMemoryRange
* pMemoryRanges
)
2291 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2293 if (device
->info
.has_llc
)
2296 /* Make sure the writes we're flushing have landed. */
2297 __builtin_ia32_mfence();
2299 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
2304 VkResult
anv_InvalidateMappedMemoryRanges(
2306 uint32_t memoryRangeCount
,
2307 const VkMappedMemoryRange
* pMemoryRanges
)
2309 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2311 if (device
->info
.has_llc
)
2314 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
2316 /* Make sure no reads get moved up above the invalidate. */
2317 __builtin_ia32_mfence();
2322 void anv_GetBufferMemoryRequirements(
2325 VkMemoryRequirements
* pMemoryRequirements
)
2327 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
2328 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2329 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2331 /* The Vulkan spec (git aaed022) says:
2333 * memoryTypeBits is a bitfield and contains one bit set for every
2334 * supported memory type for the resource. The bit `1<<i` is set if and
2335 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
2336 * structure for the physical device is supported.
2338 uint32_t memory_types
= 0;
2339 for (uint32_t i
= 0; i
< pdevice
->memory
.type_count
; i
++) {
2340 uint32_t valid_usage
= pdevice
->memory
.types
[i
].valid_buffer_usage
;
2341 if ((valid_usage
& buffer
->usage
) == buffer
->usage
)
2342 memory_types
|= (1u << i
);
2345 /* Base alignment requirement of a cache line */
2346 uint32_t alignment
= 16;
2348 /* We need an alignment of 32 for pushing UBOs */
2349 if (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
)
2350 alignment
= MAX2(alignment
, 32);
2352 pMemoryRequirements
->size
= buffer
->size
;
2353 pMemoryRequirements
->alignment
= alignment
;
2355 /* Storage and Uniform buffers should have their size aligned to
2356 * 32-bits to avoid boundary checks when last DWord is not complete.
2357 * This would ensure that not internal padding would be needed for
2360 if (device
->robust_buffer_access
&&
2361 (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
||
2362 buffer
->usage
& VK_BUFFER_USAGE_STORAGE_BUFFER_BIT
))
2363 pMemoryRequirements
->size
= align_u64(buffer
->size
, 4);
2365 pMemoryRequirements
->memoryTypeBits
= memory_types
;
2368 void anv_GetBufferMemoryRequirements2(
2370 const VkBufferMemoryRequirementsInfo2
* pInfo
,
2371 VkMemoryRequirements2
* pMemoryRequirements
)
2373 anv_GetBufferMemoryRequirements(_device
, pInfo
->buffer
,
2374 &pMemoryRequirements
->memoryRequirements
);
2376 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
2377 switch (ext
->sType
) {
2378 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
2379 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
2380 requirements
->prefersDedicatedAllocation
= VK_FALSE
;
2381 requirements
->requiresDedicatedAllocation
= VK_FALSE
;
2386 anv_debug_ignored_stype(ext
->sType
);
2392 void anv_GetImageMemoryRequirements(
2395 VkMemoryRequirements
* pMemoryRequirements
)
2397 ANV_FROM_HANDLE(anv_image
, image
, _image
);
2398 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2399 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2401 /* The Vulkan spec (git aaed022) says:
2403 * memoryTypeBits is a bitfield and contains one bit set for every
2404 * supported memory type for the resource. The bit `1<<i` is set if and
2405 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
2406 * structure for the physical device is supported.
2408 * All types are currently supported for images.
2410 uint32_t memory_types
= (1ull << pdevice
->memory
.type_count
) - 1;
2412 pMemoryRequirements
->size
= image
->size
;
2413 pMemoryRequirements
->alignment
= image
->alignment
;
2414 pMemoryRequirements
->memoryTypeBits
= memory_types
;
2417 void anv_GetImageMemoryRequirements2(
2419 const VkImageMemoryRequirementsInfo2
* pInfo
,
2420 VkMemoryRequirements2
* pMemoryRequirements
)
2422 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2423 ANV_FROM_HANDLE(anv_image
, image
, pInfo
->image
);
2425 anv_GetImageMemoryRequirements(_device
, pInfo
->image
,
2426 &pMemoryRequirements
->memoryRequirements
);
2428 vk_foreach_struct_const(ext
, pInfo
->pNext
) {
2429 switch (ext
->sType
) {
2430 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO
: {
2431 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2432 const VkImagePlaneMemoryRequirementsInfoKHR
*plane_reqs
=
2433 (const VkImagePlaneMemoryRequirementsInfoKHR
*) ext
;
2434 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
,
2435 plane_reqs
->planeAspect
);
2437 assert(image
->planes
[plane
].offset
== 0);
2439 /* The Vulkan spec (git aaed022) says:
2441 * memoryTypeBits is a bitfield and contains one bit set for every
2442 * supported memory type for the resource. The bit `1<<i` is set
2443 * if and only if the memory type `i` in the
2444 * VkPhysicalDeviceMemoryProperties structure for the physical
2445 * device is supported.
2447 * All types are currently supported for images.
2449 pMemoryRequirements
->memoryRequirements
.memoryTypeBits
=
2450 (1ull << pdevice
->memory
.type_count
) - 1;
2452 pMemoryRequirements
->memoryRequirements
.size
= image
->planes
[plane
].size
;
2453 pMemoryRequirements
->memoryRequirements
.alignment
=
2454 image
->planes
[plane
].alignment
;
2459 anv_debug_ignored_stype(ext
->sType
);
2464 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
2465 switch (ext
->sType
) {
2466 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
2467 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
2468 if (image
->needs_set_tiling
) {
2469 /* If we need to set the tiling for external consumers, we need a
2470 * dedicated allocation.
2472 * See also anv_AllocateMemory.
2474 requirements
->prefersDedicatedAllocation
= VK_TRUE
;
2475 requirements
->requiresDedicatedAllocation
= VK_TRUE
;
2477 requirements
->prefersDedicatedAllocation
= VK_FALSE
;
2478 requirements
->requiresDedicatedAllocation
= VK_FALSE
;
2484 anv_debug_ignored_stype(ext
->sType
);
2490 void anv_GetImageSparseMemoryRequirements(
2493 uint32_t* pSparseMemoryRequirementCount
,
2494 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
2496 *pSparseMemoryRequirementCount
= 0;
2499 void anv_GetImageSparseMemoryRequirements2(
2501 const VkImageSparseMemoryRequirementsInfo2
* pInfo
,
2502 uint32_t* pSparseMemoryRequirementCount
,
2503 VkSparseImageMemoryRequirements2
* pSparseMemoryRequirements
)
2505 *pSparseMemoryRequirementCount
= 0;
2508 void anv_GetDeviceMemoryCommitment(
2510 VkDeviceMemory memory
,
2511 VkDeviceSize
* pCommittedMemoryInBytes
)
2513 *pCommittedMemoryInBytes
= 0;
2517 anv_bind_buffer_memory(const VkBindBufferMemoryInfo
*pBindInfo
)
2519 ANV_FROM_HANDLE(anv_device_memory
, mem
, pBindInfo
->memory
);
2520 ANV_FROM_HANDLE(anv_buffer
, buffer
, pBindInfo
->buffer
);
2522 assert(pBindInfo
->sType
== VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
);
2525 assert((buffer
->usage
& mem
->type
->valid_buffer_usage
) == buffer
->usage
);
2526 buffer
->address
= (struct anv_address
) {
2528 .offset
= pBindInfo
->memoryOffset
,
2531 buffer
->address
= ANV_NULL_ADDRESS
;
2535 VkResult
anv_BindBufferMemory(
2538 VkDeviceMemory memory
,
2539 VkDeviceSize memoryOffset
)
2541 anv_bind_buffer_memory(
2542 &(VkBindBufferMemoryInfo
) {
2543 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
,
2546 .memoryOffset
= memoryOffset
,
2552 VkResult
anv_BindBufferMemory2(
2554 uint32_t bindInfoCount
,
2555 const VkBindBufferMemoryInfo
* pBindInfos
)
2557 for (uint32_t i
= 0; i
< bindInfoCount
; i
++)
2558 anv_bind_buffer_memory(&pBindInfos
[i
]);
2563 VkResult
anv_QueueBindSparse(
2565 uint32_t bindInfoCount
,
2566 const VkBindSparseInfo
* pBindInfo
,
2569 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
2570 if (unlikely(queue
->device
->lost
))
2571 return VK_ERROR_DEVICE_LOST
;
2573 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
2578 VkResult
anv_CreateEvent(
2580 const VkEventCreateInfo
* pCreateInfo
,
2581 const VkAllocationCallbacks
* pAllocator
,
2584 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2585 struct anv_state state
;
2586 struct anv_event
*event
;
2588 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_EVENT_CREATE_INFO
);
2590 state
= anv_state_pool_alloc(&device
->dynamic_state_pool
,
2593 event
->state
= state
;
2594 event
->semaphore
= VK_EVENT_RESET
;
2596 if (!device
->info
.has_llc
) {
2597 /* Make sure the writes we're flushing have landed. */
2598 __builtin_ia32_mfence();
2599 __builtin_ia32_clflush(event
);
2602 *pEvent
= anv_event_to_handle(event
);
2607 void anv_DestroyEvent(
2610 const VkAllocationCallbacks
* pAllocator
)
2612 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2613 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2618 anv_state_pool_free(&device
->dynamic_state_pool
, event
->state
);
2621 VkResult
anv_GetEventStatus(
2625 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2626 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2628 if (unlikely(device
->lost
))
2629 return VK_ERROR_DEVICE_LOST
;
2631 if (!device
->info
.has_llc
) {
2632 /* Invalidate read cache before reading event written by GPU. */
2633 __builtin_ia32_clflush(event
);
2634 __builtin_ia32_mfence();
2638 return event
->semaphore
;
2641 VkResult
anv_SetEvent(
2645 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2646 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2648 event
->semaphore
= VK_EVENT_SET
;
2650 if (!device
->info
.has_llc
) {
2651 /* Make sure the writes we're flushing have landed. */
2652 __builtin_ia32_mfence();
2653 __builtin_ia32_clflush(event
);
2659 VkResult
anv_ResetEvent(
2663 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2664 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2666 event
->semaphore
= VK_EVENT_RESET
;
2668 if (!device
->info
.has_llc
) {
2669 /* Make sure the writes we're flushing have landed. */
2670 __builtin_ia32_mfence();
2671 __builtin_ia32_clflush(event
);
2679 VkResult
anv_CreateBuffer(
2681 const VkBufferCreateInfo
* pCreateInfo
,
2682 const VkAllocationCallbacks
* pAllocator
,
2685 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2686 struct anv_buffer
*buffer
;
2688 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
2690 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
2691 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2693 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2695 buffer
->size
= pCreateInfo
->size
;
2696 buffer
->usage
= pCreateInfo
->usage
;
2697 buffer
->address
= ANV_NULL_ADDRESS
;
2699 *pBuffer
= anv_buffer_to_handle(buffer
);
2704 void anv_DestroyBuffer(
2707 const VkAllocationCallbacks
* pAllocator
)
2709 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2710 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
2715 vk_free2(&device
->alloc
, pAllocator
, buffer
);
2719 anv_fill_buffer_surface_state(struct anv_device
*device
, struct anv_state state
,
2720 enum isl_format format
,
2721 struct anv_address address
,
2722 uint32_t range
, uint32_t stride
)
2724 isl_buffer_fill_state(&device
->isl_dev
, state
.map
,
2725 .address
= anv_address_physical(address
),
2726 .mocs
= device
->default_mocs
,
2731 anv_state_flush(device
, state
);
2734 void anv_DestroySampler(
2737 const VkAllocationCallbacks
* pAllocator
)
2739 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2740 ANV_FROM_HANDLE(anv_sampler
, sampler
, _sampler
);
2745 vk_free2(&device
->alloc
, pAllocator
, sampler
);
2748 VkResult
anv_CreateFramebuffer(
2750 const VkFramebufferCreateInfo
* pCreateInfo
,
2751 const VkAllocationCallbacks
* pAllocator
,
2752 VkFramebuffer
* pFramebuffer
)
2754 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2755 struct anv_framebuffer
*framebuffer
;
2757 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
2759 size_t size
= sizeof(*framebuffer
) +
2760 sizeof(struct anv_image_view
*) * pCreateInfo
->attachmentCount
;
2761 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
2762 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2763 if (framebuffer
== NULL
)
2764 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2766 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
2767 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
2768 VkImageView _iview
= pCreateInfo
->pAttachments
[i
];
2769 framebuffer
->attachments
[i
] = anv_image_view_from_handle(_iview
);
2772 framebuffer
->width
= pCreateInfo
->width
;
2773 framebuffer
->height
= pCreateInfo
->height
;
2774 framebuffer
->layers
= pCreateInfo
->layers
;
2776 *pFramebuffer
= anv_framebuffer_to_handle(framebuffer
);
2781 void anv_DestroyFramebuffer(
2784 const VkAllocationCallbacks
* pAllocator
)
2786 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2787 ANV_FROM_HANDLE(anv_framebuffer
, fb
, _fb
);
2792 vk_free2(&device
->alloc
, pAllocator
, fb
);
2795 /* vk_icd.h does not declare this function, so we declare it here to
2796 * suppress Wmissing-prototypes.
2798 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2799 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
);
2801 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2802 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
)
2804 /* For the full details on loader interface versioning, see
2805 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
2806 * What follows is a condensed summary, to help you navigate the large and
2807 * confusing official doc.
2809 * - Loader interface v0 is incompatible with later versions. We don't
2812 * - In loader interface v1:
2813 * - The first ICD entrypoint called by the loader is
2814 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
2816 * - The ICD must statically expose no other Vulkan symbol unless it is
2817 * linked with -Bsymbolic.
2818 * - Each dispatchable Vulkan handle created by the ICD must be
2819 * a pointer to a struct whose first member is VK_LOADER_DATA. The
2820 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
2821 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
2822 * vkDestroySurfaceKHR(). The ICD must be capable of working with
2823 * such loader-managed surfaces.
2825 * - Loader interface v2 differs from v1 in:
2826 * - The first ICD entrypoint called by the loader is
2827 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
2828 * statically expose this entrypoint.
2830 * - Loader interface v3 differs from v2 in:
2831 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
2832 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
2833 * because the loader no longer does so.
2835 *pSupportedVersion
= MIN2(*pSupportedVersion
, 3u);