2 * Copyright © 2015 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
28 #include <sys/sysinfo.h>
32 #include <drm_fourcc.h>
34 #include "anv_private.h"
35 #include "util/strtod.h"
36 #include "util/debug.h"
37 #include "util/build_id.h"
38 #include "util/mesa-sha1.h"
40 #include "common/gen_defines.h"
42 #include "genxml/gen7_pack.h"
45 compiler_debug_log(void *data
, const char *fmt
, ...)
49 compiler_perf_log(void *data
, const char *fmt
, ...)
54 if (unlikely(INTEL_DEBUG
& DEBUG_PERF
))
55 intel_logd_v(fmt
, args
);
61 anv_compute_heap_size(int fd
, uint64_t *heap_size
)
64 if (anv_gem_get_context_param(fd
, 0, I915_CONTEXT_PARAM_GTT_SIZE
,
66 /* If, for whatever reason, we can't actually get the GTT size from the
67 * kernel (too old?) fall back to the aperture size.
69 anv_perf_warn(NULL
, NULL
,
70 "Failed to get I915_CONTEXT_PARAM_GTT_SIZE: %m");
72 if (anv_gem_get_aperture(fd
, >t_size
) == -1) {
73 return vk_errorf(NULL
, NULL
, VK_ERROR_INITIALIZATION_FAILED
,
74 "failed to get aperture size: %m");
78 /* Query the total ram from the system */
82 uint64_t total_ram
= (uint64_t)info
.totalram
* (uint64_t)info
.mem_unit
;
84 /* We don't want to burn too much ram with the GPU. If the user has 4GiB
85 * or less, we use at most half. If they have more than 4GiB, we use 3/4.
87 uint64_t available_ram
;
88 if (total_ram
<= 4ull * 1024ull * 1024ull * 1024ull)
89 available_ram
= total_ram
/ 2;
91 available_ram
= total_ram
* 3 / 4;
93 /* We also want to leave some padding for things we allocate in the driver,
94 * so don't go over 3/4 of the GTT either.
96 uint64_t available_gtt
= gtt_size
* 3 / 4;
98 *heap_size
= MIN2(available_ram
, available_gtt
);
104 anv_physical_device_init_heaps(struct anv_physical_device
*device
, int fd
)
106 /* The kernel query only tells us whether or not the kernel supports the
107 * EXEC_OBJECT_SUPPORTS_48B_ADDRESS flag and not whether or not the
108 * hardware has actual 48bit address support.
110 device
->supports_48bit_addresses
=
111 (device
->info
.gen
>= 8) && anv_gem_supports_48b_addresses(fd
);
114 VkResult result
= anv_compute_heap_size(fd
, &heap_size
);
115 if (result
!= VK_SUCCESS
)
118 if (heap_size
> (2ull << 30) && !device
->supports_48bit_addresses
) {
119 /* When running with an overridden PCI ID, we may get a GTT size from
120 * the kernel that is greater than 2 GiB but the execbuf check for 48bit
121 * address support can still fail. Just clamp the address space size to
122 * 2 GiB if we don't have 48-bit support.
124 intel_logw("%s:%d: The kernel reported a GTT size larger than 2 GiB but "
125 "not support for 48-bit addresses",
127 heap_size
= 2ull << 30;
130 if (heap_size
<= 3ull * (1ull << 30)) {
131 /* In this case, everything fits nicely into the 32-bit address space,
132 * so there's no need for supporting 48bit addresses on client-allocated
135 device
->memory
.heap_count
= 1;
136 device
->memory
.heaps
[0] = (struct anv_memory_heap
) {
138 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
139 .supports_48bit_addresses
= false,
142 /* Not everything will fit nicely into a 32-bit address space. In this
143 * case we need a 64-bit heap. Advertise a small 32-bit heap and a
144 * larger 48-bit heap. If we're in this case, then we have a total heap
145 * size larger than 3GiB which most likely means they have 8 GiB of
146 * video memory and so carving off 1 GiB for the 32-bit heap should be
149 const uint64_t heap_size_32bit
= 1ull << 30;
150 const uint64_t heap_size_48bit
= heap_size
- heap_size_32bit
;
152 assert(device
->supports_48bit_addresses
);
154 device
->memory
.heap_count
= 2;
155 device
->memory
.heaps
[0] = (struct anv_memory_heap
) {
156 .size
= heap_size_48bit
,
157 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
158 .supports_48bit_addresses
= true,
160 device
->memory
.heaps
[1] = (struct anv_memory_heap
) {
161 .size
= heap_size_32bit
,
162 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
163 .supports_48bit_addresses
= false,
167 uint32_t type_count
= 0;
168 for (uint32_t heap
= 0; heap
< device
->memory
.heap_count
; heap
++) {
169 uint32_t valid_buffer_usage
= ~0;
171 /* There appears to be a hardware issue in the VF cache where it only
172 * considers the bottom 32 bits of memory addresses. If you happen to
173 * have two vertex buffers which get placed exactly 4 GiB apart and use
174 * them in back-to-back draw calls, you can get collisions. In order to
175 * solve this problem, we require vertex and index buffers be bound to
176 * memory allocated out of the 32-bit heap.
178 if (device
->memory
.heaps
[heap
].supports_48bit_addresses
) {
179 valid_buffer_usage
&= ~(VK_BUFFER_USAGE_INDEX_BUFFER_BIT
|
180 VK_BUFFER_USAGE_VERTEX_BUFFER_BIT
);
183 if (device
->info
.has_llc
) {
184 /* Big core GPUs share LLC with the CPU and thus one memory type can be
185 * both cached and coherent at the same time.
187 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
188 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
189 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
190 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
191 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
193 .valid_buffer_usage
= valid_buffer_usage
,
196 /* The spec requires that we expose a host-visible, coherent memory
197 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
198 * to give the application a choice between cached, but not coherent and
199 * coherent but uncached (WC though).
201 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
202 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
203 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
204 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
206 .valid_buffer_usage
= valid_buffer_usage
,
208 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
209 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
210 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
211 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
213 .valid_buffer_usage
= valid_buffer_usage
,
217 device
->memory
.type_count
= type_count
;
223 anv_physical_device_init_uuids(struct anv_physical_device
*device
)
225 const struct build_id_note
*note
=
226 build_id_find_nhdr_for_addr(anv_physical_device_init_uuids
);
228 return vk_errorf(device
->instance
, device
,
229 VK_ERROR_INITIALIZATION_FAILED
,
230 "Failed to find build-id");
233 unsigned build_id_len
= build_id_length(note
);
234 if (build_id_len
< 20) {
235 return vk_errorf(device
->instance
, device
,
236 VK_ERROR_INITIALIZATION_FAILED
,
237 "build-id too short. It needs to be a SHA");
240 struct mesa_sha1 sha1_ctx
;
242 STATIC_ASSERT(VK_UUID_SIZE
<= sizeof(sha1
));
244 /* The pipeline cache UUID is used for determining when a pipeline cache is
245 * invalid. It needs both a driver build and the PCI ID of the device.
247 _mesa_sha1_init(&sha1_ctx
);
248 _mesa_sha1_update(&sha1_ctx
, build_id_data(note
), build_id_len
);
249 _mesa_sha1_update(&sha1_ctx
, &device
->chipset_id
,
250 sizeof(device
->chipset_id
));
251 _mesa_sha1_final(&sha1_ctx
, sha1
);
252 memcpy(device
->pipeline_cache_uuid
, sha1
, VK_UUID_SIZE
);
254 /* The driver UUID is used for determining sharability of images and memory
255 * between two Vulkan instances in separate processes. People who want to
256 * share memory need to also check the device UUID (below) so all this
257 * needs to be is the build-id.
259 memcpy(device
->driver_uuid
, build_id_data(note
), VK_UUID_SIZE
);
261 /* The device UUID uniquely identifies the given device within the machine.
262 * Since we never have more than one device, this doesn't need to be a real
263 * UUID. However, on the off-chance that someone tries to use this to
264 * cache pre-tiled images or something of the like, we use the PCI ID and
265 * some bits of ISL info to ensure that this is safe.
267 _mesa_sha1_init(&sha1_ctx
);
268 _mesa_sha1_update(&sha1_ctx
, &device
->chipset_id
,
269 sizeof(device
->chipset_id
));
270 _mesa_sha1_update(&sha1_ctx
, &device
->isl_dev
.has_bit6_swizzling
,
271 sizeof(device
->isl_dev
.has_bit6_swizzling
));
272 _mesa_sha1_final(&sha1_ctx
, sha1
);
273 memcpy(device
->device_uuid
, sha1
, VK_UUID_SIZE
);
279 anv_physical_device_init(struct anv_physical_device
*device
,
280 struct anv_instance
*instance
,
286 brw_process_intel_debug_variable();
288 fd
= open(path
, O_RDWR
| O_CLOEXEC
);
290 return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
292 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
293 device
->instance
= instance
;
295 assert(strlen(path
) < ARRAY_SIZE(device
->path
));
296 strncpy(device
->path
, path
, ARRAY_SIZE(device
->path
));
298 device
->no_hw
= getenv("INTEL_NO_HW") != NULL
;
300 const int pci_id_override
= gen_get_pci_device_id_override();
301 if (pci_id_override
< 0) {
302 device
->chipset_id
= anv_gem_get_param(fd
, I915_PARAM_CHIPSET_ID
);
303 if (!device
->chipset_id
) {
304 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
308 device
->chipset_id
= pci_id_override
;
309 device
->no_hw
= true;
312 device
->name
= gen_get_device_name(device
->chipset_id
);
313 if (!gen_get_device_info(device
->chipset_id
, &device
->info
)) {
314 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
318 if (device
->info
.is_haswell
) {
319 intel_logw("Haswell Vulkan support is incomplete");
320 } else if (device
->info
.gen
== 7 && !device
->info
.is_baytrail
) {
321 intel_logw("Ivy Bridge Vulkan support is incomplete");
322 } else if (device
->info
.gen
== 7 && device
->info
.is_baytrail
) {
323 intel_logw("Bay Trail Vulkan support is incomplete");
324 } else if (device
->info
.gen
>= 8 && device
->info
.gen
<= 10) {
325 /* Gen8-10 fully supported */
327 result
= vk_errorf(device
->instance
, device
,
328 VK_ERROR_INCOMPATIBLE_DRIVER
,
329 "Vulkan not yet supported on %s", device
->name
);
333 device
->cmd_parser_version
= -1;
334 if (device
->info
.gen
== 7) {
335 device
->cmd_parser_version
=
336 anv_gem_get_param(fd
, I915_PARAM_CMD_PARSER_VERSION
);
337 if (device
->cmd_parser_version
== -1) {
338 result
= vk_errorf(device
->instance
, device
,
339 VK_ERROR_INITIALIZATION_FAILED
,
340 "failed to get command parser version");
345 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_WAIT_TIMEOUT
)) {
346 result
= vk_errorf(device
->instance
, device
,
347 VK_ERROR_INITIALIZATION_FAILED
,
348 "kernel missing gem wait");
352 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_EXECBUF2
)) {
353 result
= vk_errorf(device
->instance
, device
,
354 VK_ERROR_INITIALIZATION_FAILED
,
355 "kernel missing execbuf2");
359 if (!device
->info
.has_llc
&&
360 anv_gem_get_param(fd
, I915_PARAM_MMAP_VERSION
) < 1) {
361 result
= vk_errorf(device
->instance
, device
,
362 VK_ERROR_INITIALIZATION_FAILED
,
363 "kernel missing wc mmap");
367 result
= anv_physical_device_init_heaps(device
, fd
);
368 if (result
!= VK_SUCCESS
)
371 device
->has_exec_async
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_ASYNC
);
372 device
->has_exec_capture
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_CAPTURE
);
373 device
->has_exec_fence
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE
);
374 device
->has_syncobj
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE_ARRAY
);
375 device
->has_syncobj_wait
= device
->has_syncobj
&&
376 anv_gem_supports_syncobj_wait(fd
);
377 device
->has_context_priority
= anv_gem_has_context_priority(fd
);
379 bool swizzled
= anv_gem_get_bit6_swizzle(fd
, I915_TILING_X
);
381 /* Starting with Gen10, the timestamp frequency of the command streamer may
382 * vary from one part to another. We can query the value from the kernel.
384 if (device
->info
.gen
>= 10) {
385 int timestamp_frequency
=
386 anv_gem_get_param(fd
, I915_PARAM_CS_TIMESTAMP_FREQUENCY
);
388 if (timestamp_frequency
< 0)
389 intel_logw("Kernel 4.16-rc1+ required to properly query CS timestamp frequency");
391 device
->info
.timestamp_frequency
= timestamp_frequency
;
394 /* GENs prior to 8 do not support EU/Subslice info */
395 if (device
->info
.gen
>= 8) {
396 device
->subslice_total
= anv_gem_get_param(fd
, I915_PARAM_SUBSLICE_TOTAL
);
397 device
->eu_total
= anv_gem_get_param(fd
, I915_PARAM_EU_TOTAL
);
399 /* Without this information, we cannot get the right Braswell
400 * brandstrings, and we have to use conservative numbers for GPGPU on
401 * many platforms, but otherwise, things will just work.
403 if (device
->subslice_total
< 1 || device
->eu_total
< 1) {
404 intel_logw("Kernel 4.1 required to properly query GPU properties");
406 } else if (device
->info
.gen
== 7) {
407 device
->subslice_total
= 1 << (device
->info
.gt
- 1);
410 if (device
->info
.is_cherryview
&&
411 device
->subslice_total
> 0 && device
->eu_total
> 0) {
412 /* Logical CS threads = EUs per subslice * num threads per EU */
413 uint32_t max_cs_threads
=
414 device
->eu_total
/ device
->subslice_total
* device
->info
.num_thread_per_eu
;
416 /* Fuse configurations may give more threads than expected, never less. */
417 if (max_cs_threads
> device
->info
.max_cs_threads
)
418 device
->info
.max_cs_threads
= max_cs_threads
;
421 device
->compiler
= brw_compiler_create(NULL
, &device
->info
);
422 if (device
->compiler
== NULL
) {
423 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
426 device
->compiler
->shader_debug_log
= compiler_debug_log
;
427 device
->compiler
->shader_perf_log
= compiler_perf_log
;
428 device
->compiler
->supports_pull_constants
= false;
429 device
->compiler
->constant_buffer_0_is_relative
= true;
431 isl_device_init(&device
->isl_dev
, &device
->info
, swizzled
);
433 result
= anv_physical_device_init_uuids(device
);
434 if (result
!= VK_SUCCESS
)
437 result
= anv_init_wsi(device
);
438 if (result
!= VK_SUCCESS
) {
439 ralloc_free(device
->compiler
);
443 anv_physical_device_get_supported_extensions(device
,
444 &device
->supported_extensions
);
446 device
->local_fd
= fd
;
455 anv_physical_device_finish(struct anv_physical_device
*device
)
457 anv_finish_wsi(device
);
458 ralloc_free(device
->compiler
);
459 close(device
->local_fd
);
463 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
464 VkSystemAllocationScope allocationScope
)
470 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
471 size_t align
, VkSystemAllocationScope allocationScope
)
473 return realloc(pOriginal
, size
);
477 default_free_func(void *pUserData
, void *pMemory
)
482 static const VkAllocationCallbacks default_alloc
= {
484 .pfnAllocation
= default_alloc_func
,
485 .pfnReallocation
= default_realloc_func
,
486 .pfnFree
= default_free_func
,
489 VkResult
anv_EnumerateInstanceExtensionProperties(
490 const char* pLayerName
,
491 uint32_t* pPropertyCount
,
492 VkExtensionProperties
* pProperties
)
494 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
496 for (int i
= 0; i
< ANV_INSTANCE_EXTENSION_COUNT
; i
++) {
497 if (anv_instance_extensions_supported
.extensions
[i
]) {
498 vk_outarray_append(&out
, prop
) {
499 *prop
= anv_instance_extensions
[i
];
504 return vk_outarray_status(&out
);
507 VkResult
anv_CreateInstance(
508 const VkInstanceCreateInfo
* pCreateInfo
,
509 const VkAllocationCallbacks
* pAllocator
,
510 VkInstance
* pInstance
)
512 struct anv_instance
*instance
;
515 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
517 /* Check if user passed a debug report callback to be used during
518 * Create/Destroy of instance.
520 const VkDebugReportCallbackCreateInfoEXT
*ctor_cb
=
521 vk_find_struct_const(pCreateInfo
->pNext
,
522 DEBUG_REPORT_CALLBACK_CREATE_INFO_EXT
);
524 uint32_t client_version
;
525 if (pCreateInfo
->pApplicationInfo
&&
526 pCreateInfo
->pApplicationInfo
->apiVersion
!= 0) {
527 client_version
= pCreateInfo
->pApplicationInfo
->apiVersion
;
529 client_version
= VK_MAKE_VERSION(1, 0, 0);
532 if (VK_MAKE_VERSION(1, 0, 0) > client_version
||
533 client_version
> VK_MAKE_VERSION(1, 0, 0xfff)) {
535 if (ctor_cb
&& ctor_cb
->flags
& VK_DEBUG_REPORT_ERROR_BIT_EXT
)
536 ctor_cb
->pfnCallback(VK_DEBUG_REPORT_ERROR_BIT_EXT
,
537 VK_DEBUG_REPORT_OBJECT_TYPE_INSTANCE_EXT
,
538 VK_NULL_HANDLE
, /* No handle available yet. */
542 "incompatible driver version",
545 return vk_errorf(NULL
, NULL
, VK_ERROR_INCOMPATIBLE_DRIVER
,
546 "Client requested version %d.%d.%d",
547 VK_VERSION_MAJOR(client_version
),
548 VK_VERSION_MINOR(client_version
),
549 VK_VERSION_PATCH(client_version
));
552 struct anv_instance_extension_table enabled_extensions
= {};
553 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
555 for (idx
= 0; idx
< ANV_INSTANCE_EXTENSION_COUNT
; idx
++) {
556 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
557 anv_instance_extensions
[idx
].extensionName
) == 0)
561 if (idx
>= ANV_INSTANCE_EXTENSION_COUNT
)
562 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
564 if (!anv_instance_extensions_supported
.extensions
[idx
])
565 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
567 enabled_extensions
.extensions
[idx
] = true;
570 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
571 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
573 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
575 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
578 instance
->alloc
= *pAllocator
;
580 instance
->alloc
= default_alloc
;
582 instance
->apiVersion
= client_version
;
583 instance
->enabled_extensions
= enabled_extensions
;
585 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->dispatch
.entrypoints
); i
++) {
586 /* Vulkan requires that entrypoints for extensions which have not been
587 * enabled must not be advertised.
589 if (!anv_entrypoint_is_enabled(i
, instance
->apiVersion
,
590 &instance
->enabled_extensions
, NULL
)) {
591 instance
->dispatch
.entrypoints
[i
] = NULL
;
592 } else if (anv_dispatch_table
.entrypoints
[i
] != NULL
) {
593 instance
->dispatch
.entrypoints
[i
] = anv_dispatch_table
.entrypoints
[i
];
595 instance
->dispatch
.entrypoints
[i
] =
596 anv_tramp_dispatch_table
.entrypoints
[i
];
600 instance
->physicalDeviceCount
= -1;
602 result
= vk_debug_report_instance_init(&instance
->debug_report_callbacks
);
603 if (result
!= VK_SUCCESS
) {
604 vk_free2(&default_alloc
, pAllocator
, instance
);
605 return vk_error(result
);
610 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
612 *pInstance
= anv_instance_to_handle(instance
);
617 void anv_DestroyInstance(
618 VkInstance _instance
,
619 const VkAllocationCallbacks
* pAllocator
)
621 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
626 if (instance
->physicalDeviceCount
> 0) {
627 /* We support at most one physical device. */
628 assert(instance
->physicalDeviceCount
== 1);
629 anv_physical_device_finish(&instance
->physicalDevice
);
632 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
634 vk_debug_report_instance_destroy(&instance
->debug_report_callbacks
);
638 vk_free(&instance
->alloc
, instance
);
642 anv_enumerate_devices(struct anv_instance
*instance
)
644 /* TODO: Check for more devices ? */
645 drmDevicePtr devices
[8];
646 VkResult result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
649 instance
->physicalDeviceCount
= 0;
651 max_devices
= drmGetDevices2(0, devices
, ARRAY_SIZE(devices
));
653 return VK_ERROR_INCOMPATIBLE_DRIVER
;
655 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
656 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
657 devices
[i
]->bustype
== DRM_BUS_PCI
&&
658 devices
[i
]->deviceinfo
.pci
->vendor_id
== 0x8086) {
660 result
= anv_physical_device_init(&instance
->physicalDevice
,
662 devices
[i
]->nodes
[DRM_NODE_RENDER
]);
663 if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
667 drmFreeDevices(devices
, max_devices
);
669 if (result
== VK_SUCCESS
)
670 instance
->physicalDeviceCount
= 1;
676 anv_instance_ensure_physical_device(struct anv_instance
*instance
)
678 if (instance
->physicalDeviceCount
< 0) {
679 VkResult result
= anv_enumerate_devices(instance
);
680 if (result
!= VK_SUCCESS
&&
681 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
688 VkResult
anv_EnumeratePhysicalDevices(
689 VkInstance _instance
,
690 uint32_t* pPhysicalDeviceCount
,
691 VkPhysicalDevice
* pPhysicalDevices
)
693 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
694 VK_OUTARRAY_MAKE(out
, pPhysicalDevices
, pPhysicalDeviceCount
);
696 VkResult result
= anv_instance_ensure_physical_device(instance
);
697 if (result
!= VK_SUCCESS
)
700 if (instance
->physicalDeviceCount
== 0)
703 assert(instance
->physicalDeviceCount
== 1);
704 vk_outarray_append(&out
, i
) {
705 *i
= anv_physical_device_to_handle(&instance
->physicalDevice
);
708 return vk_outarray_status(&out
);
711 VkResult
anv_EnumeratePhysicalDeviceGroups(
712 VkInstance _instance
,
713 uint32_t* pPhysicalDeviceGroupCount
,
714 VkPhysicalDeviceGroupProperties
* pPhysicalDeviceGroupProperties
)
716 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
717 VK_OUTARRAY_MAKE(out
, pPhysicalDeviceGroupProperties
,
718 pPhysicalDeviceGroupCount
);
720 VkResult result
= anv_instance_ensure_physical_device(instance
);
721 if (result
!= VK_SUCCESS
)
724 if (instance
->physicalDeviceCount
== 0)
727 assert(instance
->physicalDeviceCount
== 1);
729 vk_outarray_append(&out
, p
) {
730 p
->physicalDeviceCount
= 1;
731 memset(p
->physicalDevices
, 0, sizeof(p
->physicalDevices
));
732 p
->physicalDevices
[0] =
733 anv_physical_device_to_handle(&instance
->physicalDevice
);
734 p
->subsetAllocation
= VK_FALSE
;
736 vk_foreach_struct(ext
, p
->pNext
)
737 anv_debug_ignored_stype(ext
->sType
);
740 return vk_outarray_status(&out
);
743 void anv_GetPhysicalDeviceFeatures(
744 VkPhysicalDevice physicalDevice
,
745 VkPhysicalDeviceFeatures
* pFeatures
)
747 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
749 *pFeatures
= (VkPhysicalDeviceFeatures
) {
750 .robustBufferAccess
= true,
751 .fullDrawIndexUint32
= true,
752 .imageCubeArray
= true,
753 .independentBlend
= true,
754 .geometryShader
= true,
755 .tessellationShader
= true,
756 .sampleRateShading
= true,
757 .dualSrcBlend
= true,
759 .multiDrawIndirect
= true,
760 .drawIndirectFirstInstance
= true,
762 .depthBiasClamp
= true,
763 .fillModeNonSolid
= true,
764 .depthBounds
= false,
768 .multiViewport
= true,
769 .samplerAnisotropy
= true,
770 .textureCompressionETC2
= pdevice
->info
.gen
>= 8 ||
771 pdevice
->info
.is_baytrail
,
772 .textureCompressionASTC_LDR
= pdevice
->info
.gen
>= 9, /* FINISHME CHV */
773 .textureCompressionBC
= true,
774 .occlusionQueryPrecise
= true,
775 .pipelineStatisticsQuery
= true,
776 .fragmentStoresAndAtomics
= true,
777 .shaderTessellationAndGeometryPointSize
= true,
778 .shaderImageGatherExtended
= true,
779 .shaderStorageImageExtendedFormats
= true,
780 .shaderStorageImageMultisample
= false,
781 .shaderStorageImageReadWithoutFormat
= false,
782 .shaderStorageImageWriteWithoutFormat
= true,
783 .shaderUniformBufferArrayDynamicIndexing
= true,
784 .shaderSampledImageArrayDynamicIndexing
= true,
785 .shaderStorageBufferArrayDynamicIndexing
= true,
786 .shaderStorageImageArrayDynamicIndexing
= true,
787 .shaderClipDistance
= true,
788 .shaderCullDistance
= true,
789 .shaderFloat64
= pdevice
->info
.gen
>= 8,
790 .shaderInt64
= pdevice
->info
.gen
>= 8,
791 .shaderInt16
= false,
792 .shaderResourceMinLod
= false,
793 .variableMultisampleRate
= false,
794 .inheritedQueries
= true,
797 /* We can't do image stores in vec4 shaders */
798 pFeatures
->vertexPipelineStoresAndAtomics
=
799 pdevice
->compiler
->scalar_stage
[MESA_SHADER_VERTEX
] &&
800 pdevice
->compiler
->scalar_stage
[MESA_SHADER_GEOMETRY
];
803 void anv_GetPhysicalDeviceFeatures2(
804 VkPhysicalDevice physicalDevice
,
805 VkPhysicalDeviceFeatures2
* pFeatures
)
807 anv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
809 vk_foreach_struct(ext
, pFeatures
->pNext
) {
810 switch (ext
->sType
) {
811 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES
: {
812 VkPhysicalDeviceProtectedMemoryFeatures
*features
= (void *)ext
;
813 features
->protectedMemory
= VK_FALSE
;
817 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES
: {
818 VkPhysicalDeviceMultiviewFeatures
*features
=
819 (VkPhysicalDeviceMultiviewFeatures
*)ext
;
820 features
->multiview
= true;
821 features
->multiviewGeometryShader
= true;
822 features
->multiviewTessellationShader
= true;
826 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTER_FEATURES
: {
827 VkPhysicalDeviceVariablePointerFeatures
*features
= (void *)ext
;
828 features
->variablePointersStorageBuffer
= true;
829 features
->variablePointers
= true;
833 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES
: {
834 VkPhysicalDeviceSamplerYcbcrConversionFeatures
*features
=
835 (VkPhysicalDeviceSamplerYcbcrConversionFeatures
*) ext
;
836 features
->samplerYcbcrConversion
= true;
840 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETER_FEATURES
: {
841 VkPhysicalDeviceShaderDrawParameterFeatures
*features
= (void *)ext
;
842 features
->shaderDrawParameters
= true;
846 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES_KHR
: {
847 VkPhysicalDevice16BitStorageFeaturesKHR
*features
=
848 (VkPhysicalDevice16BitStorageFeaturesKHR
*)ext
;
849 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
851 features
->storageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
852 features
->uniformAndStorageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
853 features
->storagePushConstant16
= pdevice
->info
.gen
>= 8;
854 features
->storageInputOutput16
= false;
859 anv_debug_ignored_stype(ext
->sType
);
865 void anv_GetPhysicalDeviceProperties(
866 VkPhysicalDevice physicalDevice
,
867 VkPhysicalDeviceProperties
* pProperties
)
869 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
870 const struct gen_device_info
*devinfo
= &pdevice
->info
;
872 /* See assertions made when programming the buffer surface state. */
873 const uint32_t max_raw_buffer_sz
= devinfo
->gen
>= 7 ?
874 (1ul << 30) : (1ul << 27);
876 const uint32_t max_samplers
= (devinfo
->gen
>= 8 || devinfo
->is_haswell
) ?
879 VkSampleCountFlags sample_counts
=
880 isl_device_get_sample_counts(&pdevice
->isl_dev
);
882 VkPhysicalDeviceLimits limits
= {
883 .maxImageDimension1D
= (1 << 14),
884 .maxImageDimension2D
= (1 << 14),
885 .maxImageDimension3D
= (1 << 11),
886 .maxImageDimensionCube
= (1 << 14),
887 .maxImageArrayLayers
= (1 << 11),
888 .maxTexelBufferElements
= 128 * 1024 * 1024,
889 .maxUniformBufferRange
= (1ul << 27),
890 .maxStorageBufferRange
= max_raw_buffer_sz
,
891 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
892 .maxMemoryAllocationCount
= UINT32_MAX
,
893 .maxSamplerAllocationCount
= 64 * 1024,
894 .bufferImageGranularity
= 64, /* A cache line */
895 .sparseAddressSpaceSize
= 0,
896 .maxBoundDescriptorSets
= MAX_SETS
,
897 .maxPerStageDescriptorSamplers
= max_samplers
,
898 .maxPerStageDescriptorUniformBuffers
= 64,
899 .maxPerStageDescriptorStorageBuffers
= 64,
900 .maxPerStageDescriptorSampledImages
= max_samplers
,
901 .maxPerStageDescriptorStorageImages
= 64,
902 .maxPerStageDescriptorInputAttachments
= 64,
903 .maxPerStageResources
= 250,
904 .maxDescriptorSetSamplers
= 6 * max_samplers
, /* number of stages * maxPerStageDescriptorSamplers */
905 .maxDescriptorSetUniformBuffers
= 6 * 64, /* number of stages * maxPerStageDescriptorUniformBuffers */
906 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
907 .maxDescriptorSetStorageBuffers
= 6 * 64, /* number of stages * maxPerStageDescriptorStorageBuffers */
908 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
909 .maxDescriptorSetSampledImages
= 6 * max_samplers
, /* number of stages * maxPerStageDescriptorSampledImages */
910 .maxDescriptorSetStorageImages
= 6 * 64, /* number of stages * maxPerStageDescriptorStorageImages */
911 .maxDescriptorSetInputAttachments
= 256,
912 .maxVertexInputAttributes
= MAX_VBS
,
913 .maxVertexInputBindings
= MAX_VBS
,
914 .maxVertexInputAttributeOffset
= 2047,
915 .maxVertexInputBindingStride
= 2048,
916 .maxVertexOutputComponents
= 128,
917 .maxTessellationGenerationLevel
= 64,
918 .maxTessellationPatchSize
= 32,
919 .maxTessellationControlPerVertexInputComponents
= 128,
920 .maxTessellationControlPerVertexOutputComponents
= 128,
921 .maxTessellationControlPerPatchOutputComponents
= 128,
922 .maxTessellationControlTotalOutputComponents
= 2048,
923 .maxTessellationEvaluationInputComponents
= 128,
924 .maxTessellationEvaluationOutputComponents
= 128,
925 .maxGeometryShaderInvocations
= 32,
926 .maxGeometryInputComponents
= 64,
927 .maxGeometryOutputComponents
= 128,
928 .maxGeometryOutputVertices
= 256,
929 .maxGeometryTotalOutputComponents
= 1024,
930 .maxFragmentInputComponents
= 128,
931 .maxFragmentOutputAttachments
= 8,
932 .maxFragmentDualSrcAttachments
= 1,
933 .maxFragmentCombinedOutputResources
= 8,
934 .maxComputeSharedMemorySize
= 32768,
935 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
936 .maxComputeWorkGroupInvocations
= 16 * devinfo
->max_cs_threads
,
937 .maxComputeWorkGroupSize
= {
938 16 * devinfo
->max_cs_threads
,
939 16 * devinfo
->max_cs_threads
,
940 16 * devinfo
->max_cs_threads
,
942 .subPixelPrecisionBits
= 4 /* FIXME */,
943 .subTexelPrecisionBits
= 4 /* FIXME */,
944 .mipmapPrecisionBits
= 4 /* FIXME */,
945 .maxDrawIndexedIndexValue
= UINT32_MAX
,
946 .maxDrawIndirectCount
= UINT32_MAX
,
947 .maxSamplerLodBias
= 16,
948 .maxSamplerAnisotropy
= 16,
949 .maxViewports
= MAX_VIEWPORTS
,
950 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
951 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
952 .viewportSubPixelBits
= 13, /* We take a float? */
953 .minMemoryMapAlignment
= 4096, /* A page */
954 .minTexelBufferOffsetAlignment
= 1,
955 /* We need 16 for UBO block reads to work and 32 for push UBOs */
956 .minUniformBufferOffsetAlignment
= 32,
957 .minStorageBufferOffsetAlignment
= 4,
958 .minTexelOffset
= -8,
960 .minTexelGatherOffset
= -32,
961 .maxTexelGatherOffset
= 31,
962 .minInterpolationOffset
= -0.5,
963 .maxInterpolationOffset
= 0.4375,
964 .subPixelInterpolationOffsetBits
= 4,
965 .maxFramebufferWidth
= (1 << 14),
966 .maxFramebufferHeight
= (1 << 14),
967 .maxFramebufferLayers
= (1 << 11),
968 .framebufferColorSampleCounts
= sample_counts
,
969 .framebufferDepthSampleCounts
= sample_counts
,
970 .framebufferStencilSampleCounts
= sample_counts
,
971 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
972 .maxColorAttachments
= MAX_RTS
,
973 .sampledImageColorSampleCounts
= sample_counts
,
974 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
975 .sampledImageDepthSampleCounts
= sample_counts
,
976 .sampledImageStencilSampleCounts
= sample_counts
,
977 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
978 .maxSampleMaskWords
= 1,
979 .timestampComputeAndGraphics
= false,
980 .timestampPeriod
= 1000000000.0 / devinfo
->timestamp_frequency
,
981 .maxClipDistances
= 8,
982 .maxCullDistances
= 8,
983 .maxCombinedClipAndCullDistances
= 8,
984 .discreteQueuePriorities
= 1,
985 .pointSizeRange
= { 0.125, 255.875 },
986 .lineWidthRange
= { 0.0, 7.9921875 },
987 .pointSizeGranularity
= (1.0 / 8.0),
988 .lineWidthGranularity
= (1.0 / 128.0),
989 .strictLines
= false, /* FINISHME */
990 .standardSampleLocations
= true,
991 .optimalBufferCopyOffsetAlignment
= 128,
992 .optimalBufferCopyRowPitchAlignment
= 128,
993 .nonCoherentAtomSize
= 64,
996 *pProperties
= (VkPhysicalDeviceProperties
) {
997 .apiVersion
= anv_physical_device_api_version(pdevice
),
998 .driverVersion
= vk_get_driver_version(),
1000 .deviceID
= pdevice
->chipset_id
,
1001 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
1003 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
1006 snprintf(pProperties
->deviceName
, sizeof(pProperties
->deviceName
),
1007 "%s", pdevice
->name
);
1008 memcpy(pProperties
->pipelineCacheUUID
,
1009 pdevice
->pipeline_cache_uuid
, VK_UUID_SIZE
);
1012 void anv_GetPhysicalDeviceProperties2(
1013 VkPhysicalDevice physicalDevice
,
1014 VkPhysicalDeviceProperties2
* pProperties
)
1016 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1018 anv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
1020 vk_foreach_struct(ext
, pProperties
->pNext
) {
1021 switch (ext
->sType
) {
1022 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
1023 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
1024 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
1026 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
1030 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES
: {
1031 VkPhysicalDeviceIDProperties
*id_props
=
1032 (VkPhysicalDeviceIDProperties
*)ext
;
1033 memcpy(id_props
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
1034 memcpy(id_props
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
1035 /* The LUID is for Windows. */
1036 id_props
->deviceLUIDValid
= false;
1040 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES
: {
1041 VkPhysicalDeviceMaintenance3Properties
*props
=
1042 (VkPhysicalDeviceMaintenance3Properties
*)ext
;
1043 /* This value doesn't matter for us today as our per-stage
1044 * descriptors are the real limit.
1046 props
->maxPerSetDescriptors
= 1024;
1047 props
->maxMemoryAllocationSize
= MAX_MEMORY_ALLOCATION_SIZE
;
1051 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES
: {
1052 VkPhysicalDeviceMultiviewProperties
*properties
=
1053 (VkPhysicalDeviceMultiviewProperties
*)ext
;
1054 properties
->maxMultiviewViewCount
= 16;
1055 properties
->maxMultiviewInstanceIndex
= UINT32_MAX
/ 16;
1059 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES
: {
1060 VkPhysicalDevicePointClippingProperties
*properties
=
1061 (VkPhysicalDevicePointClippingProperties
*) ext
;
1062 properties
->pointClippingBehavior
= VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES
;
1063 anv_finishme("Implement pop-free point clipping");
1068 anv_debug_ignored_stype(ext
->sType
);
1074 /* We support exactly one queue family. */
1075 static const VkQueueFamilyProperties
1076 anv_queue_family_properties
= {
1077 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
1078 VK_QUEUE_COMPUTE_BIT
|
1079 VK_QUEUE_TRANSFER_BIT
,
1081 .timestampValidBits
= 36, /* XXX: Real value here */
1082 .minImageTransferGranularity
= { 1, 1, 1 },
1085 void anv_GetPhysicalDeviceQueueFamilyProperties(
1086 VkPhysicalDevice physicalDevice
,
1088 VkQueueFamilyProperties
* pQueueFamilyProperties
)
1090 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pCount
);
1092 vk_outarray_append(&out
, p
) {
1093 *p
= anv_queue_family_properties
;
1097 void anv_GetPhysicalDeviceQueueFamilyProperties2(
1098 VkPhysicalDevice physicalDevice
,
1099 uint32_t* pQueueFamilyPropertyCount
,
1100 VkQueueFamilyProperties2
* pQueueFamilyProperties
)
1103 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pQueueFamilyPropertyCount
);
1105 vk_outarray_append(&out
, p
) {
1106 p
->queueFamilyProperties
= anv_queue_family_properties
;
1108 vk_foreach_struct(s
, p
->pNext
) {
1109 anv_debug_ignored_stype(s
->sType
);
1114 void anv_GetPhysicalDeviceMemoryProperties(
1115 VkPhysicalDevice physicalDevice
,
1116 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
1118 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
1120 pMemoryProperties
->memoryTypeCount
= physical_device
->memory
.type_count
;
1121 for (uint32_t i
= 0; i
< physical_device
->memory
.type_count
; i
++) {
1122 pMemoryProperties
->memoryTypes
[i
] = (VkMemoryType
) {
1123 .propertyFlags
= physical_device
->memory
.types
[i
].propertyFlags
,
1124 .heapIndex
= physical_device
->memory
.types
[i
].heapIndex
,
1128 pMemoryProperties
->memoryHeapCount
= physical_device
->memory
.heap_count
;
1129 for (uint32_t i
= 0; i
< physical_device
->memory
.heap_count
; i
++) {
1130 pMemoryProperties
->memoryHeaps
[i
] = (VkMemoryHeap
) {
1131 .size
= physical_device
->memory
.heaps
[i
].size
,
1132 .flags
= physical_device
->memory
.heaps
[i
].flags
,
1137 void anv_GetPhysicalDeviceMemoryProperties2(
1138 VkPhysicalDevice physicalDevice
,
1139 VkPhysicalDeviceMemoryProperties2
* pMemoryProperties
)
1141 anv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
1142 &pMemoryProperties
->memoryProperties
);
1144 vk_foreach_struct(ext
, pMemoryProperties
->pNext
) {
1145 switch (ext
->sType
) {
1147 anv_debug_ignored_stype(ext
->sType
);
1154 anv_GetDeviceGroupPeerMemoryFeatures(
1157 uint32_t localDeviceIndex
,
1158 uint32_t remoteDeviceIndex
,
1159 VkPeerMemoryFeatureFlags
* pPeerMemoryFeatures
)
1161 assert(localDeviceIndex
== 0 && remoteDeviceIndex
== 0);
1162 *pPeerMemoryFeatures
= VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT
|
1163 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT
|
1164 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT
|
1165 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT
;
1168 PFN_vkVoidFunction
anv_GetInstanceProcAddr(
1169 VkInstance _instance
,
1172 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1174 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
1175 * when we have to return valid function pointers, NULL, or it's left
1176 * undefined. See the table for exact details.
1181 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
1182 if (strcmp(pName, "vk" #entrypoint) == 0) \
1183 return (PFN_vkVoidFunction)anv_##entrypoint
1185 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties
);
1186 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties
);
1187 LOOKUP_ANV_ENTRYPOINT(CreateInstance
);
1189 #undef LOOKUP_ANV_ENTRYPOINT
1191 if (instance
== NULL
)
1194 int idx
= anv_get_entrypoint_index(pName
);
1198 return instance
->dispatch
.entrypoints
[idx
];
1201 /* With version 1+ of the loader interface the ICD should expose
1202 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
1205 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1206 VkInstance instance
,
1210 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1211 VkInstance instance
,
1214 return anv_GetInstanceProcAddr(instance
, pName
);
1217 PFN_vkVoidFunction
anv_GetDeviceProcAddr(
1221 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1223 if (!device
|| !pName
)
1226 int idx
= anv_get_entrypoint_index(pName
);
1230 return device
->dispatch
.entrypoints
[idx
];
1234 anv_CreateDebugReportCallbackEXT(VkInstance _instance
,
1235 const VkDebugReportCallbackCreateInfoEXT
* pCreateInfo
,
1236 const VkAllocationCallbacks
* pAllocator
,
1237 VkDebugReportCallbackEXT
* pCallback
)
1239 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1240 return vk_create_debug_report_callback(&instance
->debug_report_callbacks
,
1241 pCreateInfo
, pAllocator
, &instance
->alloc
,
1246 anv_DestroyDebugReportCallbackEXT(VkInstance _instance
,
1247 VkDebugReportCallbackEXT _callback
,
1248 const VkAllocationCallbacks
* pAllocator
)
1250 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1251 vk_destroy_debug_report_callback(&instance
->debug_report_callbacks
,
1252 _callback
, pAllocator
, &instance
->alloc
);
1256 anv_DebugReportMessageEXT(VkInstance _instance
,
1257 VkDebugReportFlagsEXT flags
,
1258 VkDebugReportObjectTypeEXT objectType
,
1261 int32_t messageCode
,
1262 const char* pLayerPrefix
,
1263 const char* pMessage
)
1265 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1266 vk_debug_report(&instance
->debug_report_callbacks
, flags
, objectType
,
1267 object
, location
, messageCode
, pLayerPrefix
, pMessage
);
1271 anv_queue_init(struct anv_device
*device
, struct anv_queue
*queue
)
1273 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1274 queue
->device
= device
;
1275 queue
->pool
= &device
->surface_state_pool
;
1280 anv_queue_finish(struct anv_queue
*queue
)
1284 static struct anv_state
1285 anv_state_pool_emit_data(struct anv_state_pool
*pool
, size_t size
, size_t align
, const void *p
)
1287 struct anv_state state
;
1289 state
= anv_state_pool_alloc(pool
, size
, align
);
1290 memcpy(state
.map
, p
, size
);
1292 anv_state_flush(pool
->block_pool
.device
, state
);
1297 struct gen8_border_color
{
1302 /* Pad out to 64 bytes */
1307 anv_device_init_border_colors(struct anv_device
*device
)
1309 static const struct gen8_border_color border_colors
[] = {
1310 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
1311 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
1312 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
1313 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
1314 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
1315 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
1318 device
->border_colors
= anv_state_pool_emit_data(&device
->dynamic_state_pool
,
1319 sizeof(border_colors
), 64,
1324 anv_device_init_trivial_batch(struct anv_device
*device
)
1326 anv_bo_init_new(&device
->trivial_batch_bo
, device
, 4096);
1328 if (device
->instance
->physicalDevice
.has_exec_async
)
1329 device
->trivial_batch_bo
.flags
|= EXEC_OBJECT_ASYNC
;
1331 void *map
= anv_gem_mmap(device
, device
->trivial_batch_bo
.gem_handle
,
1334 struct anv_batch batch
= {
1340 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
1341 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
1343 if (!device
->info
.has_llc
)
1344 gen_clflush_range(map
, batch
.next
- map
);
1346 anv_gem_munmap(map
, device
->trivial_batch_bo
.size
);
1349 VkResult
anv_EnumerateDeviceExtensionProperties(
1350 VkPhysicalDevice physicalDevice
,
1351 const char* pLayerName
,
1352 uint32_t* pPropertyCount
,
1353 VkExtensionProperties
* pProperties
)
1355 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
1356 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
1359 for (int i
= 0; i
< ANV_DEVICE_EXTENSION_COUNT
; i
++) {
1360 if (device
->supported_extensions
.extensions
[i
]) {
1361 vk_outarray_append(&out
, prop
) {
1362 *prop
= anv_device_extensions
[i
];
1367 return vk_outarray_status(&out
);
1371 anv_device_init_dispatch(struct anv_device
*device
)
1373 const struct anv_dispatch_table
*genX_table
;
1374 switch (device
->info
.gen
) {
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");
1427 VkResult
anv_CreateDevice(
1428 VkPhysicalDevice physicalDevice
,
1429 const VkDeviceCreateInfo
* pCreateInfo
,
1430 const VkAllocationCallbacks
* pAllocator
,
1433 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
1435 struct anv_device
*device
;
1437 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO
);
1439 struct anv_device_extension_table enabled_extensions
= { };
1440 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
1442 for (idx
= 0; idx
< ANV_DEVICE_EXTENSION_COUNT
; idx
++) {
1443 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
1444 anv_device_extensions
[idx
].extensionName
) == 0)
1448 if (idx
>= ANV_DEVICE_EXTENSION_COUNT
)
1449 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
1451 if (!physical_device
->supported_extensions
.extensions
[idx
])
1452 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
1454 enabled_extensions
.extensions
[idx
] = true;
1457 /* Check enabled features */
1458 if (pCreateInfo
->pEnabledFeatures
) {
1459 VkPhysicalDeviceFeatures supported_features
;
1460 anv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
1461 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
1462 VkBool32
*enabled_feature
= (VkBool32
*)pCreateInfo
->pEnabledFeatures
;
1463 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
1464 for (uint32_t i
= 0; i
< num_features
; i
++) {
1465 if (enabled_feature
[i
] && !supported_feature
[i
])
1466 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
1470 /* Check if client specified queue priority. */
1471 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*queue_priority
=
1472 vk_find_struct_const(pCreateInfo
->pQueueCreateInfos
[0].pNext
,
1473 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT
);
1475 VkQueueGlobalPriorityEXT priority
=
1476 queue_priority
? queue_priority
->globalPriority
:
1477 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
;
1479 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
1481 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
1483 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1485 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1486 device
->instance
= physical_device
->instance
;
1487 device
->chipset_id
= physical_device
->chipset_id
;
1488 device
->no_hw
= physical_device
->no_hw
;
1489 device
->lost
= false;
1492 device
->alloc
= *pAllocator
;
1494 device
->alloc
= physical_device
->instance
->alloc
;
1496 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
1497 device
->fd
= open(physical_device
->path
, O_RDWR
| O_CLOEXEC
);
1498 if (device
->fd
== -1) {
1499 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1503 device
->context_id
= anv_gem_create_context(device
);
1504 if (device
->context_id
== -1) {
1505 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1509 /* As per spec, the driver implementation may deny requests to acquire
1510 * a priority above the default priority (MEDIUM) if the caller does not
1511 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
1514 if (physical_device
->has_context_priority
) {
1515 int err
= anv_gem_set_context_param(device
->fd
, device
->context_id
,
1516 I915_CONTEXT_PARAM_PRIORITY
,
1517 vk_priority_to_gen(priority
));
1518 if (err
!= 0 && priority
> VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
) {
1519 result
= vk_error(VK_ERROR_NOT_PERMITTED_EXT
);
1524 device
->info
= physical_device
->info
;
1525 device
->isl_dev
= physical_device
->isl_dev
;
1527 /* On Broadwell and later, we can use batch chaining to more efficiently
1528 * implement growing command buffers. Prior to Haswell, the kernel
1529 * command parser gets in the way and we have to fall back to growing
1532 device
->can_chain_batches
= device
->info
.gen
>= 8;
1534 device
->robust_buffer_access
= pCreateInfo
->pEnabledFeatures
&&
1535 pCreateInfo
->pEnabledFeatures
->robustBufferAccess
;
1536 device
->enabled_extensions
= enabled_extensions
;
1538 anv_device_init_dispatch(device
);
1540 if (pthread_mutex_init(&device
->mutex
, NULL
) != 0) {
1541 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1542 goto fail_context_id
;
1545 pthread_condattr_t condattr
;
1546 if (pthread_condattr_init(&condattr
) != 0) {
1547 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1550 if (pthread_condattr_setclock(&condattr
, CLOCK_MONOTONIC
) != 0) {
1551 pthread_condattr_destroy(&condattr
);
1552 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1555 if (pthread_cond_init(&device
->queue_submit
, NULL
) != 0) {
1556 pthread_condattr_destroy(&condattr
);
1557 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1560 pthread_condattr_destroy(&condattr
);
1563 (physical_device
->supports_48bit_addresses
? EXEC_OBJECT_SUPPORTS_48B_ADDRESS
: 0) |
1564 (physical_device
->has_exec_async
? EXEC_OBJECT_ASYNC
: 0) |
1565 (physical_device
->has_exec_capture
? EXEC_OBJECT_CAPTURE
: 0);
1567 anv_bo_pool_init(&device
->batch_bo_pool
, device
, bo_flags
);
1569 result
= anv_bo_cache_init(&device
->bo_cache
);
1570 if (result
!= VK_SUCCESS
)
1571 goto fail_batch_bo_pool
;
1573 /* For the state pools we explicitly disable 48bit. */
1574 bo_flags
= (physical_device
->has_exec_async
? EXEC_OBJECT_ASYNC
: 0) |
1575 (physical_device
->has_exec_capture
? EXEC_OBJECT_CAPTURE
: 0);
1577 result
= anv_state_pool_init(&device
->dynamic_state_pool
, device
, 16384,
1579 if (result
!= VK_SUCCESS
)
1582 result
= anv_state_pool_init(&device
->instruction_state_pool
, device
, 16384,
1584 if (result
!= VK_SUCCESS
)
1585 goto fail_dynamic_state_pool
;
1587 result
= anv_state_pool_init(&device
->surface_state_pool
, device
, 4096,
1589 if (result
!= VK_SUCCESS
)
1590 goto fail_instruction_state_pool
;
1592 result
= anv_bo_init_new(&device
->workaround_bo
, device
, 1024);
1593 if (result
!= VK_SUCCESS
)
1594 goto fail_surface_state_pool
;
1596 anv_device_init_trivial_batch(device
);
1598 anv_scratch_pool_init(device
, &device
->scratch_pool
);
1600 anv_queue_init(device
, &device
->queue
);
1602 switch (device
->info
.gen
) {
1604 if (!device
->info
.is_haswell
)
1605 result
= gen7_init_device_state(device
);
1607 result
= gen75_init_device_state(device
);
1610 result
= gen8_init_device_state(device
);
1613 result
= gen9_init_device_state(device
);
1616 result
= gen10_init_device_state(device
);
1619 result
= gen11_init_device_state(device
);
1622 /* Shouldn't get here as we don't create physical devices for any other
1624 unreachable("unhandled gen");
1626 if (result
!= VK_SUCCESS
)
1627 goto fail_workaround_bo
;
1629 anv_device_init_blorp(device
);
1631 anv_device_init_border_colors(device
);
1633 *pDevice
= anv_device_to_handle(device
);
1638 anv_queue_finish(&device
->queue
);
1639 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
1640 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
1641 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
1642 fail_surface_state_pool
:
1643 anv_state_pool_finish(&device
->surface_state_pool
);
1644 fail_instruction_state_pool
:
1645 anv_state_pool_finish(&device
->instruction_state_pool
);
1646 fail_dynamic_state_pool
:
1647 anv_state_pool_finish(&device
->dynamic_state_pool
);
1649 anv_bo_cache_finish(&device
->bo_cache
);
1651 anv_bo_pool_finish(&device
->batch_bo_pool
);
1652 pthread_cond_destroy(&device
->queue_submit
);
1654 pthread_mutex_destroy(&device
->mutex
);
1656 anv_gem_destroy_context(device
, device
->context_id
);
1660 vk_free(&device
->alloc
, device
);
1665 void anv_DestroyDevice(
1667 const VkAllocationCallbacks
* pAllocator
)
1669 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1674 anv_device_finish_blorp(device
);
1676 anv_queue_finish(&device
->queue
);
1678 #ifdef HAVE_VALGRIND
1679 /* We only need to free these to prevent valgrind errors. The backing
1680 * BO will go away in a couple of lines so we don't actually leak.
1682 anv_state_pool_free(&device
->dynamic_state_pool
, device
->border_colors
);
1685 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
1687 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
1688 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
1690 anv_gem_close(device
, device
->trivial_batch_bo
.gem_handle
);
1692 anv_state_pool_finish(&device
->surface_state_pool
);
1693 anv_state_pool_finish(&device
->instruction_state_pool
);
1694 anv_state_pool_finish(&device
->dynamic_state_pool
);
1696 anv_bo_cache_finish(&device
->bo_cache
);
1698 anv_bo_pool_finish(&device
->batch_bo_pool
);
1700 pthread_cond_destroy(&device
->queue_submit
);
1701 pthread_mutex_destroy(&device
->mutex
);
1703 anv_gem_destroy_context(device
, device
->context_id
);
1707 vk_free(&device
->alloc
, device
);
1710 VkResult
anv_EnumerateInstanceLayerProperties(
1711 uint32_t* pPropertyCount
,
1712 VkLayerProperties
* pProperties
)
1714 if (pProperties
== NULL
) {
1715 *pPropertyCount
= 0;
1719 /* None supported at this time */
1720 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1723 VkResult
anv_EnumerateDeviceLayerProperties(
1724 VkPhysicalDevice physicalDevice
,
1725 uint32_t* pPropertyCount
,
1726 VkLayerProperties
* pProperties
)
1728 if (pProperties
== NULL
) {
1729 *pPropertyCount
= 0;
1733 /* None supported at this time */
1734 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1737 void anv_GetDeviceQueue(
1739 uint32_t queueNodeIndex
,
1740 uint32_t queueIndex
,
1743 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1745 assert(queueIndex
== 0);
1747 *pQueue
= anv_queue_to_handle(&device
->queue
);
1750 void anv_GetDeviceQueue2(
1752 const VkDeviceQueueInfo2
* pQueueInfo
,
1755 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1757 assert(pQueueInfo
->queueIndex
== 0);
1759 if (pQueueInfo
->flags
== device
->queue
.flags
)
1760 *pQueue
= anv_queue_to_handle(&device
->queue
);
1766 anv_device_query_status(struct anv_device
*device
)
1768 /* This isn't likely as most of the callers of this function already check
1769 * for it. However, it doesn't hurt to check and it potentially lets us
1772 if (unlikely(device
->lost
))
1773 return VK_ERROR_DEVICE_LOST
;
1775 uint32_t active
, pending
;
1776 int ret
= anv_gem_gpu_get_reset_stats(device
, &active
, &pending
);
1778 /* We don't know the real error. */
1779 device
->lost
= true;
1780 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1781 "get_reset_stats failed: %m");
1785 device
->lost
= true;
1786 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1787 "GPU hung on one of our command buffers");
1788 } else if (pending
) {
1789 device
->lost
= true;
1790 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1791 "GPU hung with commands in-flight");
1798 anv_device_bo_busy(struct anv_device
*device
, struct anv_bo
*bo
)
1800 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
1801 * Other usages of the BO (such as on different hardware) will not be
1802 * flagged as "busy" by this ioctl. Use with care.
1804 int ret
= anv_gem_busy(device
, bo
->gem_handle
);
1806 return VK_NOT_READY
;
1807 } else if (ret
== -1) {
1808 /* We don't know the real error. */
1809 device
->lost
= true;
1810 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1811 "gem wait failed: %m");
1814 /* Query for device status after the busy call. If the BO we're checking
1815 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
1816 * client because it clearly doesn't have valid data. Yes, this most
1817 * likely means an ioctl, but we just did an ioctl to query the busy status
1818 * so it's no great loss.
1820 return anv_device_query_status(device
);
1824 anv_device_wait(struct anv_device
*device
, struct anv_bo
*bo
,
1827 int ret
= anv_gem_wait(device
, bo
->gem_handle
, &timeout
);
1828 if (ret
== -1 && errno
== ETIME
) {
1830 } else if (ret
== -1) {
1831 /* We don't know the real error. */
1832 device
->lost
= true;
1833 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1834 "gem wait failed: %m");
1837 /* Query for device status after the wait. If the BO we're waiting on got
1838 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
1839 * because it clearly doesn't have valid data. Yes, this most likely means
1840 * an ioctl, but we just did an ioctl to wait so it's no great loss.
1842 return anv_device_query_status(device
);
1845 VkResult
anv_DeviceWaitIdle(
1848 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1849 if (unlikely(device
->lost
))
1850 return VK_ERROR_DEVICE_LOST
;
1852 struct anv_batch batch
;
1855 batch
.start
= batch
.next
= cmds
;
1856 batch
.end
= (void *) cmds
+ sizeof(cmds
);
1858 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
1859 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
1861 return anv_device_submit_simple_batch(device
, &batch
);
1865 anv_bo_init_new(struct anv_bo
*bo
, struct anv_device
*device
, uint64_t size
)
1867 uint32_t gem_handle
= anv_gem_create(device
, size
);
1869 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
1871 anv_bo_init(bo
, gem_handle
, size
);
1876 VkResult
anv_AllocateMemory(
1878 const VkMemoryAllocateInfo
* pAllocateInfo
,
1879 const VkAllocationCallbacks
* pAllocator
,
1880 VkDeviceMemory
* pMem
)
1882 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1883 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
1884 struct anv_device_memory
*mem
;
1885 VkResult result
= VK_SUCCESS
;
1887 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
1889 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
1890 assert(pAllocateInfo
->allocationSize
> 0);
1892 if (pAllocateInfo
->allocationSize
> MAX_MEMORY_ALLOCATION_SIZE
)
1893 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
1895 /* FINISHME: Fail if allocation request exceeds heap size. */
1897 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
1898 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1900 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1902 assert(pAllocateInfo
->memoryTypeIndex
< pdevice
->memory
.type_count
);
1903 mem
->type
= &pdevice
->memory
.types
[pAllocateInfo
->memoryTypeIndex
];
1907 const VkImportMemoryFdInfoKHR
*fd_info
=
1908 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_FD_INFO_KHR
);
1910 /* The Vulkan spec permits handleType to be 0, in which case the struct is
1913 if (fd_info
&& fd_info
->handleType
) {
1914 /* At the moment, we support only the below handle types. */
1915 assert(fd_info
->handleType
==
1916 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
1917 fd_info
->handleType
==
1918 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
1920 result
= anv_bo_cache_import(device
, &device
->bo_cache
,
1921 fd_info
->fd
, &mem
->bo
);
1922 if (result
!= VK_SUCCESS
)
1925 VkDeviceSize aligned_alloc_size
=
1926 align_u64(pAllocateInfo
->allocationSize
, 4096);
1928 /* For security purposes, we reject importing the bo if it's smaller
1929 * than the requested allocation size. This prevents a malicious client
1930 * from passing a buffer to a trusted client, lying about the size, and
1931 * telling the trusted client to try and texture from an image that goes
1932 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
1933 * in the trusted client. The trusted client can protect itself against
1934 * this sort of attack but only if it can trust the buffer size.
1936 if (mem
->bo
->size
< aligned_alloc_size
) {
1937 result
= vk_errorf(device
->instance
, device
,
1938 VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
,
1939 "aligned allocationSize too large for "
1940 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR: "
1941 "%"PRIu64
"B > %"PRIu64
"B",
1942 aligned_alloc_size
, mem
->bo
->size
);
1943 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
1947 /* From the Vulkan spec:
1949 * "Importing memory from a file descriptor transfers ownership of
1950 * the file descriptor from the application to the Vulkan
1951 * implementation. The application must not perform any operations on
1952 * the file descriptor after a successful import."
1954 * If the import fails, we leave the file descriptor open.
1958 result
= anv_bo_cache_alloc(device
, &device
->bo_cache
,
1959 pAllocateInfo
->allocationSize
,
1961 if (result
!= VK_SUCCESS
)
1964 const VkMemoryDedicatedAllocateInfoKHR
*dedicated_info
=
1965 vk_find_struct_const(pAllocateInfo
->pNext
, MEMORY_DEDICATED_ALLOCATE_INFO_KHR
);
1966 if (dedicated_info
&& dedicated_info
->image
!= VK_NULL_HANDLE
) {
1967 ANV_FROM_HANDLE(anv_image
, image
, dedicated_info
->image
);
1969 /* Some legacy (non-modifiers) consumers need the tiling to be set on
1970 * the BO. In this case, we have a dedicated allocation.
1972 if (image
->needs_set_tiling
) {
1973 const uint32_t i915_tiling
=
1974 isl_tiling_to_i915_tiling(image
->planes
[0].surface
.isl
.tiling
);
1975 int ret
= anv_gem_set_tiling(device
, mem
->bo
->gem_handle
,
1976 image
->planes
[0].surface
.isl
.row_pitch
,
1979 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
1980 return vk_errorf(device
->instance
, NULL
,
1981 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
1982 "failed to set BO tiling: %m");
1988 assert(mem
->type
->heapIndex
< pdevice
->memory
.heap_count
);
1989 if (pdevice
->memory
.heaps
[mem
->type
->heapIndex
].supports_48bit_addresses
)
1990 mem
->bo
->flags
|= EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
1992 const struct wsi_memory_allocate_info
*wsi_info
=
1993 vk_find_struct_const(pAllocateInfo
->pNext
, WSI_MEMORY_ALLOCATE_INFO_MESA
);
1994 if (wsi_info
&& wsi_info
->implicit_sync
) {
1995 /* We need to set the WRITE flag on window system buffers so that GEM
1996 * will know we're writing to them and synchronize uses on other rings
1997 * (eg if the display server uses the blitter ring).
1999 mem
->bo
->flags
|= EXEC_OBJECT_WRITE
;
2000 } else if (pdevice
->has_exec_async
) {
2001 mem
->bo
->flags
|= EXEC_OBJECT_ASYNC
;
2004 *pMem
= anv_device_memory_to_handle(mem
);
2009 vk_free2(&device
->alloc
, pAllocator
, mem
);
2014 VkResult
anv_GetMemoryFdKHR(
2016 const VkMemoryGetFdInfoKHR
* pGetFdInfo
,
2019 ANV_FROM_HANDLE(anv_device
, dev
, device_h
);
2020 ANV_FROM_HANDLE(anv_device_memory
, mem
, pGetFdInfo
->memory
);
2022 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
2024 assert(pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
2025 pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
2027 return anv_bo_cache_export(dev
, &dev
->bo_cache
, mem
->bo
, pFd
);
2030 VkResult
anv_GetMemoryFdPropertiesKHR(
2032 VkExternalMemoryHandleTypeFlagBitsKHR handleType
,
2034 VkMemoryFdPropertiesKHR
* pMemoryFdProperties
)
2036 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2037 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2039 switch (handleType
) {
2040 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
:
2041 /* dma-buf can be imported as any memory type */
2042 pMemoryFdProperties
->memoryTypeBits
=
2043 (1 << pdevice
->memory
.type_count
) - 1;
2047 /* The valid usage section for this function says:
2049 * "handleType must not be one of the handle types defined as
2052 * So opaque handle types fall into the default "unsupported" case.
2054 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
2058 void anv_FreeMemory(
2060 VkDeviceMemory _mem
,
2061 const VkAllocationCallbacks
* pAllocator
)
2063 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2064 ANV_FROM_HANDLE(anv_device_memory
, mem
, _mem
);
2070 anv_UnmapMemory(_device
, _mem
);
2072 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
2074 vk_free2(&device
->alloc
, pAllocator
, mem
);
2077 VkResult
anv_MapMemory(
2079 VkDeviceMemory _memory
,
2080 VkDeviceSize offset
,
2082 VkMemoryMapFlags flags
,
2085 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2086 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
2093 if (size
== VK_WHOLE_SIZE
)
2094 size
= mem
->bo
->size
- offset
;
2096 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
2098 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
2099 * assert(size != 0);
2100 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
2101 * equal to the size of the memory minus offset
2104 assert(offset
+ size
<= mem
->bo
->size
);
2106 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
2107 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
2108 * at a time is valid. We could just mmap up front and return an offset
2109 * pointer here, but that may exhaust virtual memory on 32 bit
2112 uint32_t gem_flags
= 0;
2114 if (!device
->info
.has_llc
&&
2115 (mem
->type
->propertyFlags
& VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
))
2116 gem_flags
|= I915_MMAP_WC
;
2118 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
2119 uint64_t map_offset
= offset
& ~4095ull;
2120 assert(offset
>= map_offset
);
2121 uint64_t map_size
= (offset
+ size
) - map_offset
;
2123 /* Let's map whole pages */
2124 map_size
= align_u64(map_size
, 4096);
2126 void *map
= anv_gem_mmap(device
, mem
->bo
->gem_handle
,
2127 map_offset
, map_size
, gem_flags
);
2128 if (map
== MAP_FAILED
)
2129 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
2132 mem
->map_size
= map_size
;
2134 *ppData
= mem
->map
+ (offset
- map_offset
);
2139 void anv_UnmapMemory(
2141 VkDeviceMemory _memory
)
2143 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
2148 anv_gem_munmap(mem
->map
, mem
->map_size
);
2155 clflush_mapped_ranges(struct anv_device
*device
,
2157 const VkMappedMemoryRange
*ranges
)
2159 for (uint32_t i
= 0; i
< count
; i
++) {
2160 ANV_FROM_HANDLE(anv_device_memory
, mem
, ranges
[i
].memory
);
2161 if (ranges
[i
].offset
>= mem
->map_size
)
2164 gen_clflush_range(mem
->map
+ ranges
[i
].offset
,
2165 MIN2(ranges
[i
].size
, mem
->map_size
- ranges
[i
].offset
));
2169 VkResult
anv_FlushMappedMemoryRanges(
2171 uint32_t memoryRangeCount
,
2172 const VkMappedMemoryRange
* pMemoryRanges
)
2174 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2176 if (device
->info
.has_llc
)
2179 /* Make sure the writes we're flushing have landed. */
2180 __builtin_ia32_mfence();
2182 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
2187 VkResult
anv_InvalidateMappedMemoryRanges(
2189 uint32_t memoryRangeCount
,
2190 const VkMappedMemoryRange
* pMemoryRanges
)
2192 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2194 if (device
->info
.has_llc
)
2197 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
2199 /* Make sure no reads get moved up above the invalidate. */
2200 __builtin_ia32_mfence();
2205 void anv_GetBufferMemoryRequirements(
2208 VkMemoryRequirements
* pMemoryRequirements
)
2210 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
2211 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2212 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2214 /* The Vulkan spec (git aaed022) says:
2216 * memoryTypeBits is a bitfield and contains one bit set for every
2217 * supported memory type for the resource. The bit `1<<i` is set if and
2218 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
2219 * structure for the physical device is supported.
2221 uint32_t memory_types
= 0;
2222 for (uint32_t i
= 0; i
< pdevice
->memory
.type_count
; i
++) {
2223 uint32_t valid_usage
= pdevice
->memory
.types
[i
].valid_buffer_usage
;
2224 if ((valid_usage
& buffer
->usage
) == buffer
->usage
)
2225 memory_types
|= (1u << i
);
2228 /* Base alignment requirement of a cache line */
2229 uint32_t alignment
= 16;
2231 /* We need an alignment of 32 for pushing UBOs */
2232 if (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
)
2233 alignment
= MAX2(alignment
, 32);
2235 pMemoryRequirements
->size
= buffer
->size
;
2236 pMemoryRequirements
->alignment
= alignment
;
2238 /* Storage and Uniform buffers should have their size aligned to
2239 * 32-bits to avoid boundary checks when last DWord is not complete.
2240 * This would ensure that not internal padding would be needed for
2243 if (device
->robust_buffer_access
&&
2244 (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
||
2245 buffer
->usage
& VK_BUFFER_USAGE_STORAGE_BUFFER_BIT
))
2246 pMemoryRequirements
->size
= align_u64(buffer
->size
, 4);
2248 pMemoryRequirements
->memoryTypeBits
= memory_types
;
2251 void anv_GetBufferMemoryRequirements2(
2253 const VkBufferMemoryRequirementsInfo2
* pInfo
,
2254 VkMemoryRequirements2
* pMemoryRequirements
)
2256 anv_GetBufferMemoryRequirements(_device
, pInfo
->buffer
,
2257 &pMemoryRequirements
->memoryRequirements
);
2259 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
2260 switch (ext
->sType
) {
2261 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
2262 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
2263 requirements
->prefersDedicatedAllocation
= VK_FALSE
;
2264 requirements
->requiresDedicatedAllocation
= VK_FALSE
;
2269 anv_debug_ignored_stype(ext
->sType
);
2275 void anv_GetImageMemoryRequirements(
2278 VkMemoryRequirements
* pMemoryRequirements
)
2280 ANV_FROM_HANDLE(anv_image
, image
, _image
);
2281 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2282 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2284 /* The Vulkan spec (git aaed022) says:
2286 * memoryTypeBits is a bitfield and contains one bit set for every
2287 * supported memory type for the resource. The bit `1<<i` is set if and
2288 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
2289 * structure for the physical device is supported.
2291 * All types are currently supported for images.
2293 uint32_t memory_types
= (1ull << pdevice
->memory
.type_count
) - 1;
2295 pMemoryRequirements
->size
= image
->size
;
2296 pMemoryRequirements
->alignment
= image
->alignment
;
2297 pMemoryRequirements
->memoryTypeBits
= memory_types
;
2300 void anv_GetImageMemoryRequirements2(
2302 const VkImageMemoryRequirementsInfo2
* pInfo
,
2303 VkMemoryRequirements2
* pMemoryRequirements
)
2305 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2306 ANV_FROM_HANDLE(anv_image
, image
, pInfo
->image
);
2308 anv_GetImageMemoryRequirements(_device
, pInfo
->image
,
2309 &pMemoryRequirements
->memoryRequirements
);
2311 vk_foreach_struct_const(ext
, pInfo
->pNext
) {
2312 switch (ext
->sType
) {
2313 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO
: {
2314 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2315 const VkImagePlaneMemoryRequirementsInfoKHR
*plane_reqs
=
2316 (const VkImagePlaneMemoryRequirementsInfoKHR
*) ext
;
2317 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
,
2318 plane_reqs
->planeAspect
);
2320 assert(image
->planes
[plane
].offset
== 0);
2322 /* The Vulkan spec (git aaed022) says:
2324 * memoryTypeBits is a bitfield and contains one bit set for every
2325 * supported memory type for the resource. The bit `1<<i` is set
2326 * if and only if the memory type `i` in the
2327 * VkPhysicalDeviceMemoryProperties structure for the physical
2328 * device is supported.
2330 * All types are currently supported for images.
2332 pMemoryRequirements
->memoryRequirements
.memoryTypeBits
=
2333 (1ull << pdevice
->memory
.type_count
) - 1;
2335 pMemoryRequirements
->memoryRequirements
.size
= image
->planes
[plane
].size
;
2336 pMemoryRequirements
->memoryRequirements
.alignment
=
2337 image
->planes
[plane
].alignment
;
2342 anv_debug_ignored_stype(ext
->sType
);
2347 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
2348 switch (ext
->sType
) {
2349 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
2350 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
2351 if (image
->needs_set_tiling
) {
2352 /* If we need to set the tiling for external consumers, we need a
2353 * dedicated allocation.
2355 * See also anv_AllocateMemory.
2357 requirements
->prefersDedicatedAllocation
= VK_TRUE
;
2358 requirements
->requiresDedicatedAllocation
= VK_TRUE
;
2360 requirements
->prefersDedicatedAllocation
= VK_FALSE
;
2361 requirements
->requiresDedicatedAllocation
= VK_FALSE
;
2367 anv_debug_ignored_stype(ext
->sType
);
2373 void anv_GetImageSparseMemoryRequirements(
2376 uint32_t* pSparseMemoryRequirementCount
,
2377 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
2379 *pSparseMemoryRequirementCount
= 0;
2382 void anv_GetImageSparseMemoryRequirements2(
2384 const VkImageSparseMemoryRequirementsInfo2
* pInfo
,
2385 uint32_t* pSparseMemoryRequirementCount
,
2386 VkSparseImageMemoryRequirements2
* pSparseMemoryRequirements
)
2388 *pSparseMemoryRequirementCount
= 0;
2391 void anv_GetDeviceMemoryCommitment(
2393 VkDeviceMemory memory
,
2394 VkDeviceSize
* pCommittedMemoryInBytes
)
2396 *pCommittedMemoryInBytes
= 0;
2400 anv_bind_buffer_memory(const VkBindBufferMemoryInfo
*pBindInfo
)
2402 ANV_FROM_HANDLE(anv_device_memory
, mem
, pBindInfo
->memory
);
2403 ANV_FROM_HANDLE(anv_buffer
, buffer
, pBindInfo
->buffer
);
2405 assert(pBindInfo
->sType
== VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
);
2408 assert((buffer
->usage
& mem
->type
->valid_buffer_usage
) == buffer
->usage
);
2409 buffer
->bo
= mem
->bo
;
2410 buffer
->offset
= pBindInfo
->memoryOffset
;
2417 VkResult
anv_BindBufferMemory(
2420 VkDeviceMemory memory
,
2421 VkDeviceSize memoryOffset
)
2423 anv_bind_buffer_memory(
2424 &(VkBindBufferMemoryInfo
) {
2425 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
,
2428 .memoryOffset
= memoryOffset
,
2434 VkResult
anv_BindBufferMemory2(
2436 uint32_t bindInfoCount
,
2437 const VkBindBufferMemoryInfo
* pBindInfos
)
2439 for (uint32_t i
= 0; i
< bindInfoCount
; i
++)
2440 anv_bind_buffer_memory(&pBindInfos
[i
]);
2445 VkResult
anv_QueueBindSparse(
2447 uint32_t bindInfoCount
,
2448 const VkBindSparseInfo
* pBindInfo
,
2451 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
2452 if (unlikely(queue
->device
->lost
))
2453 return VK_ERROR_DEVICE_LOST
;
2455 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
2460 VkResult
anv_CreateEvent(
2462 const VkEventCreateInfo
* pCreateInfo
,
2463 const VkAllocationCallbacks
* pAllocator
,
2466 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2467 struct anv_state state
;
2468 struct anv_event
*event
;
2470 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_EVENT_CREATE_INFO
);
2472 state
= anv_state_pool_alloc(&device
->dynamic_state_pool
,
2475 event
->state
= state
;
2476 event
->semaphore
= VK_EVENT_RESET
;
2478 if (!device
->info
.has_llc
) {
2479 /* Make sure the writes we're flushing have landed. */
2480 __builtin_ia32_mfence();
2481 __builtin_ia32_clflush(event
);
2484 *pEvent
= anv_event_to_handle(event
);
2489 void anv_DestroyEvent(
2492 const VkAllocationCallbacks
* pAllocator
)
2494 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2495 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2500 anv_state_pool_free(&device
->dynamic_state_pool
, event
->state
);
2503 VkResult
anv_GetEventStatus(
2507 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2508 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2510 if (unlikely(device
->lost
))
2511 return VK_ERROR_DEVICE_LOST
;
2513 if (!device
->info
.has_llc
) {
2514 /* Invalidate read cache before reading event written by GPU. */
2515 __builtin_ia32_clflush(event
);
2516 __builtin_ia32_mfence();
2520 return event
->semaphore
;
2523 VkResult
anv_SetEvent(
2527 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2528 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2530 event
->semaphore
= VK_EVENT_SET
;
2532 if (!device
->info
.has_llc
) {
2533 /* Make sure the writes we're flushing have landed. */
2534 __builtin_ia32_mfence();
2535 __builtin_ia32_clflush(event
);
2541 VkResult
anv_ResetEvent(
2545 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2546 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2548 event
->semaphore
= VK_EVENT_RESET
;
2550 if (!device
->info
.has_llc
) {
2551 /* Make sure the writes we're flushing have landed. */
2552 __builtin_ia32_mfence();
2553 __builtin_ia32_clflush(event
);
2561 VkResult
anv_CreateBuffer(
2563 const VkBufferCreateInfo
* pCreateInfo
,
2564 const VkAllocationCallbacks
* pAllocator
,
2567 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2568 struct anv_buffer
*buffer
;
2570 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
2572 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
2573 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2575 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2577 buffer
->size
= pCreateInfo
->size
;
2578 buffer
->usage
= pCreateInfo
->usage
;
2582 *pBuffer
= anv_buffer_to_handle(buffer
);
2587 void anv_DestroyBuffer(
2590 const VkAllocationCallbacks
* pAllocator
)
2592 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2593 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
2598 vk_free2(&device
->alloc
, pAllocator
, buffer
);
2602 anv_fill_buffer_surface_state(struct anv_device
*device
, struct anv_state state
,
2603 enum isl_format format
,
2604 uint32_t offset
, uint32_t range
, uint32_t stride
)
2606 isl_buffer_fill_state(&device
->isl_dev
, state
.map
,
2608 .mocs
= device
->default_mocs
,
2613 anv_state_flush(device
, state
);
2616 void anv_DestroySampler(
2619 const VkAllocationCallbacks
* pAllocator
)
2621 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2622 ANV_FROM_HANDLE(anv_sampler
, sampler
, _sampler
);
2627 vk_free2(&device
->alloc
, pAllocator
, sampler
);
2630 VkResult
anv_CreateFramebuffer(
2632 const VkFramebufferCreateInfo
* pCreateInfo
,
2633 const VkAllocationCallbacks
* pAllocator
,
2634 VkFramebuffer
* pFramebuffer
)
2636 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2637 struct anv_framebuffer
*framebuffer
;
2639 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
2641 size_t size
= sizeof(*framebuffer
) +
2642 sizeof(struct anv_image_view
*) * pCreateInfo
->attachmentCount
;
2643 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
2644 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2645 if (framebuffer
== NULL
)
2646 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2648 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
2649 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
2650 VkImageView _iview
= pCreateInfo
->pAttachments
[i
];
2651 framebuffer
->attachments
[i
] = anv_image_view_from_handle(_iview
);
2654 framebuffer
->width
= pCreateInfo
->width
;
2655 framebuffer
->height
= pCreateInfo
->height
;
2656 framebuffer
->layers
= pCreateInfo
->layers
;
2658 *pFramebuffer
= anv_framebuffer_to_handle(framebuffer
);
2663 void anv_DestroyFramebuffer(
2666 const VkAllocationCallbacks
* pAllocator
)
2668 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2669 ANV_FROM_HANDLE(anv_framebuffer
, fb
, _fb
);
2674 vk_free2(&device
->alloc
, pAllocator
, fb
);
2677 /* vk_icd.h does not declare this function, so we declare it here to
2678 * suppress Wmissing-prototypes.
2680 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2681 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
);
2683 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2684 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
)
2686 /* For the full details on loader interface versioning, see
2687 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
2688 * What follows is a condensed summary, to help you navigate the large and
2689 * confusing official doc.
2691 * - Loader interface v0 is incompatible with later versions. We don't
2694 * - In loader interface v1:
2695 * - The first ICD entrypoint called by the loader is
2696 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
2698 * - The ICD must statically expose no other Vulkan symbol unless it is
2699 * linked with -Bsymbolic.
2700 * - Each dispatchable Vulkan handle created by the ICD must be
2701 * a pointer to a struct whose first member is VK_LOADER_DATA. The
2702 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
2703 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
2704 * vkDestroySurfaceKHR(). The ICD must be capable of working with
2705 * such loader-managed surfaces.
2707 * - Loader interface v2 differs from v1 in:
2708 * - The first ICD entrypoint called by the loader is
2709 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
2710 * statically expose this entrypoint.
2712 * - Loader interface v3 differs from v2 in:
2713 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
2714 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
2715 * because the loader no longer does so.
2717 *pSupportedVersion
= MIN2(*pSupportedVersion
, 3u);