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"
41 #include "genxml/gen7_pack.h"
44 compiler_debug_log(void *data
, const char *fmt
, ...)
48 compiler_perf_log(void *data
, const char *fmt
, ...)
53 if (unlikely(INTEL_DEBUG
& DEBUG_PERF
))
54 intel_logd_v(fmt
, args
);
60 anv_compute_heap_size(int fd
, uint64_t *heap_size
)
63 if (anv_gem_get_context_param(fd
, 0, I915_CONTEXT_PARAM_GTT_SIZE
,
65 /* If, for whatever reason, we can't actually get the GTT size from the
66 * kernel (too old?) fall back to the aperture size.
68 anv_perf_warn(NULL
, NULL
,
69 "Failed to get I915_CONTEXT_PARAM_GTT_SIZE: %m");
71 if (anv_gem_get_aperture(fd
, >t_size
) == -1) {
72 return vk_errorf(NULL
, NULL
, VK_ERROR_INITIALIZATION_FAILED
,
73 "failed to get aperture size: %m");
77 /* Query the total ram from the system */
81 uint64_t total_ram
= (uint64_t)info
.totalram
* (uint64_t)info
.mem_unit
;
83 /* We don't want to burn too much ram with the GPU. If the user has 4GiB
84 * or less, we use at most half. If they have more than 4GiB, we use 3/4.
86 uint64_t available_ram
;
87 if (total_ram
<= 4ull * 1024ull * 1024ull * 1024ull)
88 available_ram
= total_ram
/ 2;
90 available_ram
= total_ram
* 3 / 4;
92 /* We also want to leave some padding for things we allocate in the driver,
93 * so don't go over 3/4 of the GTT either.
95 uint64_t available_gtt
= gtt_size
* 3 / 4;
97 *heap_size
= MIN2(available_ram
, available_gtt
);
103 anv_physical_device_init_heaps(struct anv_physical_device
*device
, int fd
)
105 /* The kernel query only tells us whether or not the kernel supports the
106 * EXEC_OBJECT_SUPPORTS_48B_ADDRESS flag and not whether or not the
107 * hardware has actual 48bit address support.
109 device
->supports_48bit_addresses
=
110 (device
->info
.gen
>= 8) && anv_gem_supports_48b_addresses(fd
);
113 VkResult result
= anv_compute_heap_size(fd
, &heap_size
);
114 if (result
!= VK_SUCCESS
)
117 if (heap_size
> (2ull << 30) && !device
->supports_48bit_addresses
) {
118 /* When running with an overridden PCI ID, we may get a GTT size from
119 * the kernel that is greater than 2 GiB but the execbuf check for 48bit
120 * address support can still fail. Just clamp the address space size to
121 * 2 GiB if we don't have 48-bit support.
123 intel_logw("%s:%d: The kernel reported a GTT size larger than 2 GiB but "
124 "not support for 48-bit addresses",
126 heap_size
= 2ull << 30;
129 if (heap_size
<= 3ull * (1ull << 30)) {
130 /* In this case, everything fits nicely into the 32-bit address space,
131 * so there's no need for supporting 48bit addresses on client-allocated
134 device
->memory
.heap_count
= 1;
135 device
->memory
.heaps
[0] = (struct anv_memory_heap
) {
137 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
138 .supports_48bit_addresses
= false,
141 /* Not everything will fit nicely into a 32-bit address space. In this
142 * case we need a 64-bit heap. Advertise a small 32-bit heap and a
143 * larger 48-bit heap. If we're in this case, then we have a total heap
144 * size larger than 3GiB which most likely means they have 8 GiB of
145 * video memory and so carving off 1 GiB for the 32-bit heap should be
148 const uint64_t heap_size_32bit
= 1ull << 30;
149 const uint64_t heap_size_48bit
= heap_size
- heap_size_32bit
;
151 assert(device
->supports_48bit_addresses
);
153 device
->memory
.heap_count
= 2;
154 device
->memory
.heaps
[0] = (struct anv_memory_heap
) {
155 .size
= heap_size_48bit
,
156 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
157 .supports_48bit_addresses
= true,
159 device
->memory
.heaps
[1] = (struct anv_memory_heap
) {
160 .size
= heap_size_32bit
,
161 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
162 .supports_48bit_addresses
= false,
166 uint32_t type_count
= 0;
167 for (uint32_t heap
= 0; heap
< device
->memory
.heap_count
; heap
++) {
168 uint32_t valid_buffer_usage
= ~0;
170 /* There appears to be a hardware issue in the VF cache where it only
171 * considers the bottom 32 bits of memory addresses. If you happen to
172 * have two vertex buffers which get placed exactly 4 GiB apart and use
173 * them in back-to-back draw calls, you can get collisions. In order to
174 * solve this problem, we require vertex and index buffers be bound to
175 * memory allocated out of the 32-bit heap.
177 if (device
->memory
.heaps
[heap
].supports_48bit_addresses
) {
178 valid_buffer_usage
&= ~(VK_BUFFER_USAGE_INDEX_BUFFER_BIT
|
179 VK_BUFFER_USAGE_VERTEX_BUFFER_BIT
);
182 if (device
->info
.has_llc
) {
183 /* Big core GPUs share LLC with the CPU and thus one memory type can be
184 * both cached and coherent at the same time.
186 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
187 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
188 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
189 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
190 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
192 .valid_buffer_usage
= valid_buffer_usage
,
195 /* The spec requires that we expose a host-visible, coherent memory
196 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
197 * to give the application a choice between cached, but not coherent and
198 * coherent but uncached (WC though).
200 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
201 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
202 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
203 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
205 .valid_buffer_usage
= valid_buffer_usage
,
207 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
208 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
209 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
210 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
212 .valid_buffer_usage
= valid_buffer_usage
,
216 device
->memory
.type_count
= type_count
;
222 anv_physical_device_init_uuids(struct anv_physical_device
*device
)
224 const struct build_id_note
*note
=
225 build_id_find_nhdr_for_addr(anv_physical_device_init_uuids
);
227 return vk_errorf(device
->instance
, device
,
228 VK_ERROR_INITIALIZATION_FAILED
,
229 "Failed to find build-id");
232 unsigned build_id_len
= build_id_length(note
);
233 if (build_id_len
< 20) {
234 return vk_errorf(device
->instance
, device
,
235 VK_ERROR_INITIALIZATION_FAILED
,
236 "build-id too short. It needs to be a SHA");
239 struct mesa_sha1 sha1_ctx
;
241 STATIC_ASSERT(VK_UUID_SIZE
<= sizeof(sha1
));
243 /* The pipeline cache UUID is used for determining when a pipeline cache is
244 * invalid. It needs both a driver build and the PCI ID of the device.
246 _mesa_sha1_init(&sha1_ctx
);
247 _mesa_sha1_update(&sha1_ctx
, build_id_data(note
), build_id_len
);
248 _mesa_sha1_update(&sha1_ctx
, &device
->chipset_id
,
249 sizeof(device
->chipset_id
));
250 _mesa_sha1_final(&sha1_ctx
, sha1
);
251 memcpy(device
->pipeline_cache_uuid
, sha1
, VK_UUID_SIZE
);
253 /* The driver UUID is used for determining sharability of images and memory
254 * between two Vulkan instances in separate processes. People who want to
255 * share memory need to also check the device UUID (below) so all this
256 * needs to be is the build-id.
258 memcpy(device
->driver_uuid
, build_id_data(note
), VK_UUID_SIZE
);
260 /* The device UUID uniquely identifies the given device within the machine.
261 * Since we never have more than one device, this doesn't need to be a real
262 * UUID. However, on the off-chance that someone tries to use this to
263 * cache pre-tiled images or something of the like, we use the PCI ID and
264 * some bits of ISL info to ensure that this is safe.
266 _mesa_sha1_init(&sha1_ctx
);
267 _mesa_sha1_update(&sha1_ctx
, &device
->chipset_id
,
268 sizeof(device
->chipset_id
));
269 _mesa_sha1_update(&sha1_ctx
, &device
->isl_dev
.has_bit6_swizzling
,
270 sizeof(device
->isl_dev
.has_bit6_swizzling
));
271 _mesa_sha1_final(&sha1_ctx
, sha1
);
272 memcpy(device
->device_uuid
, sha1
, VK_UUID_SIZE
);
278 anv_physical_device_init(struct anv_physical_device
*device
,
279 struct anv_instance
*instance
,
285 brw_process_intel_debug_variable();
287 fd
= open(path
, O_RDWR
| O_CLOEXEC
);
289 return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
291 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
292 device
->instance
= instance
;
294 assert(strlen(path
) < ARRAY_SIZE(device
->path
));
295 strncpy(device
->path
, path
, ARRAY_SIZE(device
->path
));
297 device
->chipset_id
= anv_gem_get_param(fd
, I915_PARAM_CHIPSET_ID
);
298 if (!device
->chipset_id
) {
299 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
303 device
->name
= gen_get_device_name(device
->chipset_id
);
304 if (!gen_get_device_info(device
->chipset_id
, &device
->info
)) {
305 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
309 if (device
->info
.is_haswell
) {
310 intel_logw("Haswell Vulkan support is incomplete");
311 } else if (device
->info
.gen
== 7 && !device
->info
.is_baytrail
) {
312 intel_logw("Ivy Bridge Vulkan support is incomplete");
313 } else if (device
->info
.gen
== 7 && device
->info
.is_baytrail
) {
314 intel_logw("Bay Trail Vulkan support is incomplete");
315 } else if (device
->info
.gen
>= 8 && device
->info
.gen
<= 9) {
316 /* Broadwell, Cherryview, Skylake, Broxton, Kabylake, Coffelake is as
317 * fully supported as anything */
318 } else if (device
->info
.gen
== 10) {
319 intel_logw("Cannonlake Vulkan support is alpha");
321 result
= vk_errorf(device
->instance
, device
,
322 VK_ERROR_INCOMPATIBLE_DRIVER
,
323 "Vulkan not yet supported on %s", device
->name
);
327 device
->cmd_parser_version
= -1;
328 if (device
->info
.gen
== 7) {
329 device
->cmd_parser_version
=
330 anv_gem_get_param(fd
, I915_PARAM_CMD_PARSER_VERSION
);
331 if (device
->cmd_parser_version
== -1) {
332 result
= vk_errorf(device
->instance
, device
,
333 VK_ERROR_INITIALIZATION_FAILED
,
334 "failed to get command parser version");
339 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_WAIT_TIMEOUT
)) {
340 result
= vk_errorf(device
->instance
, device
,
341 VK_ERROR_INITIALIZATION_FAILED
,
342 "kernel missing gem wait");
346 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_EXECBUF2
)) {
347 result
= vk_errorf(device
->instance
, device
,
348 VK_ERROR_INITIALIZATION_FAILED
,
349 "kernel missing execbuf2");
353 if (!device
->info
.has_llc
&&
354 anv_gem_get_param(fd
, I915_PARAM_MMAP_VERSION
) < 1) {
355 result
= vk_errorf(device
->instance
, device
,
356 VK_ERROR_INITIALIZATION_FAILED
,
357 "kernel missing wc mmap");
361 result
= anv_physical_device_init_heaps(device
, fd
);
362 if (result
!= VK_SUCCESS
)
365 device
->has_exec_async
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_ASYNC
);
366 device
->has_exec_capture
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_CAPTURE
);
367 device
->has_exec_fence
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE
);
368 device
->has_syncobj
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE_ARRAY
);
369 device
->has_syncobj_wait
= device
->has_syncobj
&&
370 anv_gem_supports_syncobj_wait(fd
);
372 bool swizzled
= anv_gem_get_bit6_swizzle(fd
, I915_TILING_X
);
374 /* Starting with Gen10, the timestamp frequency of the command streamer may
375 * vary from one part to another. We can query the value from the kernel.
377 if (device
->info
.gen
>= 10) {
378 int timestamp_frequency
=
379 anv_gem_get_param(fd
, I915_PARAM_CS_TIMESTAMP_FREQUENCY
);
381 if (timestamp_frequency
< 0)
382 intel_logw("Kernel 4.16-rc1+ required to properly query CS timestamp frequency");
384 device
->info
.timestamp_frequency
= timestamp_frequency
;
387 /* GENs prior to 8 do not support EU/Subslice info */
388 if (device
->info
.gen
>= 8) {
389 device
->subslice_total
= anv_gem_get_param(fd
, I915_PARAM_SUBSLICE_TOTAL
);
390 device
->eu_total
= anv_gem_get_param(fd
, I915_PARAM_EU_TOTAL
);
392 /* Without this information, we cannot get the right Braswell
393 * brandstrings, and we have to use conservative numbers for GPGPU on
394 * many platforms, but otherwise, things will just work.
396 if (device
->subslice_total
< 1 || device
->eu_total
< 1) {
397 intel_logw("Kernel 4.1 required to properly query GPU properties");
399 } else if (device
->info
.gen
== 7) {
400 device
->subslice_total
= 1 << (device
->info
.gt
- 1);
403 if (device
->info
.is_cherryview
&&
404 device
->subslice_total
> 0 && device
->eu_total
> 0) {
405 /* Logical CS threads = EUs per subslice * num threads per EU */
406 uint32_t max_cs_threads
=
407 device
->eu_total
/ device
->subslice_total
* device
->info
.num_thread_per_eu
;
409 /* Fuse configurations may give more threads than expected, never less. */
410 if (max_cs_threads
> device
->info
.max_cs_threads
)
411 device
->info
.max_cs_threads
= max_cs_threads
;
414 device
->compiler
= brw_compiler_create(NULL
, &device
->info
);
415 if (device
->compiler
== NULL
) {
416 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
419 device
->compiler
->shader_debug_log
= compiler_debug_log
;
420 device
->compiler
->shader_perf_log
= compiler_perf_log
;
421 device
->compiler
->supports_pull_constants
= false;
422 device
->compiler
->constant_buffer_0_is_relative
= true;
424 isl_device_init(&device
->isl_dev
, &device
->info
, swizzled
);
426 result
= anv_physical_device_init_uuids(device
);
427 if (result
!= VK_SUCCESS
)
430 result
= anv_init_wsi(device
);
431 if (result
!= VK_SUCCESS
) {
432 ralloc_free(device
->compiler
);
436 anv_physical_device_get_supported_extensions(device
,
437 &device
->supported_extensions
);
439 device
->local_fd
= fd
;
448 anv_physical_device_finish(struct anv_physical_device
*device
)
450 anv_finish_wsi(device
);
451 ralloc_free(device
->compiler
);
452 close(device
->local_fd
);
456 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
457 VkSystemAllocationScope allocationScope
)
463 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
464 size_t align
, VkSystemAllocationScope allocationScope
)
466 return realloc(pOriginal
, size
);
470 default_free_func(void *pUserData
, void *pMemory
)
475 static const VkAllocationCallbacks default_alloc
= {
477 .pfnAllocation
= default_alloc_func
,
478 .pfnReallocation
= default_realloc_func
,
479 .pfnFree
= default_free_func
,
482 VkResult
anv_EnumerateInstanceExtensionProperties(
483 const char* pLayerName
,
484 uint32_t* pPropertyCount
,
485 VkExtensionProperties
* pProperties
)
487 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
489 for (int i
= 0; i
< ANV_INSTANCE_EXTENSION_COUNT
; i
++) {
490 if (anv_instance_extensions_supported
.extensions
[i
]) {
491 vk_outarray_append(&out
, prop
) {
492 *prop
= anv_instance_extensions
[i
];
497 return vk_outarray_status(&out
);
500 VkResult
anv_CreateInstance(
501 const VkInstanceCreateInfo
* pCreateInfo
,
502 const VkAllocationCallbacks
* pAllocator
,
503 VkInstance
* pInstance
)
505 struct anv_instance
*instance
;
508 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
510 /* Check if user passed a debug report callback to be used during
511 * Create/Destroy of instance.
513 const VkDebugReportCallbackCreateInfoEXT
*ctor_cb
=
514 vk_find_struct_const(pCreateInfo
->pNext
,
515 DEBUG_REPORT_CALLBACK_CREATE_INFO_EXT
);
517 uint32_t client_version
;
518 if (pCreateInfo
->pApplicationInfo
&&
519 pCreateInfo
->pApplicationInfo
->apiVersion
!= 0) {
520 client_version
= pCreateInfo
->pApplicationInfo
->apiVersion
;
522 client_version
= VK_MAKE_VERSION(1, 0, 0);
525 if (VK_MAKE_VERSION(1, 0, 0) > client_version
||
526 client_version
> VK_MAKE_VERSION(1, 0, 0xfff)) {
528 if (ctor_cb
&& ctor_cb
->flags
& VK_DEBUG_REPORT_ERROR_BIT_EXT
)
529 ctor_cb
->pfnCallback(VK_DEBUG_REPORT_ERROR_BIT_EXT
,
530 VK_DEBUG_REPORT_OBJECT_TYPE_INSTANCE_EXT
,
531 VK_NULL_HANDLE
, /* No handle available yet. */
535 "incompatible driver version",
538 return vk_errorf(NULL
, NULL
, VK_ERROR_INCOMPATIBLE_DRIVER
,
539 "Client requested version %d.%d.%d",
540 VK_VERSION_MAJOR(client_version
),
541 VK_VERSION_MINOR(client_version
),
542 VK_VERSION_PATCH(client_version
));
545 struct anv_instance_extension_table enabled_extensions
= {};
546 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
548 for (idx
= 0; idx
< ANV_INSTANCE_EXTENSION_COUNT
; idx
++) {
549 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
550 anv_instance_extensions
[idx
].extensionName
) == 0)
554 if (idx
>= ANV_INSTANCE_EXTENSION_COUNT
)
555 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
557 if (!anv_instance_extensions_supported
.extensions
[idx
])
558 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
560 enabled_extensions
.extensions
[idx
] = true;
563 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
564 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
566 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
568 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
571 instance
->alloc
= *pAllocator
;
573 instance
->alloc
= default_alloc
;
575 instance
->apiVersion
= client_version
;
576 instance
->enabled_extensions
= enabled_extensions
;
577 instance
->physicalDeviceCount
= -1;
579 result
= vk_debug_report_instance_init(&instance
->debug_report_callbacks
);
580 if (result
!= VK_SUCCESS
) {
581 vk_free2(&default_alloc
, pAllocator
, instance
);
582 return vk_error(result
);
587 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
589 *pInstance
= anv_instance_to_handle(instance
);
594 void anv_DestroyInstance(
595 VkInstance _instance
,
596 const VkAllocationCallbacks
* pAllocator
)
598 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
603 if (instance
->physicalDeviceCount
> 0) {
604 /* We support at most one physical device. */
605 assert(instance
->physicalDeviceCount
== 1);
606 anv_physical_device_finish(&instance
->physicalDevice
);
609 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
611 vk_debug_report_instance_destroy(&instance
->debug_report_callbacks
);
615 vk_free(&instance
->alloc
, instance
);
619 anv_enumerate_devices(struct anv_instance
*instance
)
621 /* TODO: Check for more devices ? */
622 drmDevicePtr devices
[8];
623 VkResult result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
626 instance
->physicalDeviceCount
= 0;
628 max_devices
= drmGetDevices2(0, devices
, ARRAY_SIZE(devices
));
630 return VK_ERROR_INCOMPATIBLE_DRIVER
;
632 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
633 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
634 devices
[i
]->bustype
== DRM_BUS_PCI
&&
635 devices
[i
]->deviceinfo
.pci
->vendor_id
== 0x8086) {
637 result
= anv_physical_device_init(&instance
->physicalDevice
,
639 devices
[i
]->nodes
[DRM_NODE_RENDER
]);
640 if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
644 drmFreeDevices(devices
, max_devices
);
646 if (result
== VK_SUCCESS
)
647 instance
->physicalDeviceCount
= 1;
653 VkResult
anv_EnumeratePhysicalDevices(
654 VkInstance _instance
,
655 uint32_t* pPhysicalDeviceCount
,
656 VkPhysicalDevice
* pPhysicalDevices
)
658 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
659 VK_OUTARRAY_MAKE(out
, pPhysicalDevices
, pPhysicalDeviceCount
);
662 if (instance
->physicalDeviceCount
< 0) {
663 result
= anv_enumerate_devices(instance
);
664 if (result
!= VK_SUCCESS
&&
665 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
669 if (instance
->physicalDeviceCount
> 0) {
670 assert(instance
->physicalDeviceCount
== 1);
671 vk_outarray_append(&out
, i
) {
672 *i
= anv_physical_device_to_handle(&instance
->physicalDevice
);
676 return vk_outarray_status(&out
);
679 void anv_GetPhysicalDeviceFeatures(
680 VkPhysicalDevice physicalDevice
,
681 VkPhysicalDeviceFeatures
* pFeatures
)
683 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
685 *pFeatures
= (VkPhysicalDeviceFeatures
) {
686 .robustBufferAccess
= true,
687 .fullDrawIndexUint32
= true,
688 .imageCubeArray
= true,
689 .independentBlend
= true,
690 .geometryShader
= true,
691 .tessellationShader
= true,
692 .sampleRateShading
= true,
693 .dualSrcBlend
= true,
695 .multiDrawIndirect
= true,
696 .drawIndirectFirstInstance
= true,
698 .depthBiasClamp
= true,
699 .fillModeNonSolid
= true,
700 .depthBounds
= false,
704 .multiViewport
= true,
705 .samplerAnisotropy
= true,
706 .textureCompressionETC2
= pdevice
->info
.gen
>= 8 ||
707 pdevice
->info
.is_baytrail
,
708 .textureCompressionASTC_LDR
= pdevice
->info
.gen
>= 9, /* FINISHME CHV */
709 .textureCompressionBC
= true,
710 .occlusionQueryPrecise
= true,
711 .pipelineStatisticsQuery
= true,
712 .fragmentStoresAndAtomics
= true,
713 .shaderTessellationAndGeometryPointSize
= true,
714 .shaderImageGatherExtended
= true,
715 .shaderStorageImageExtendedFormats
= true,
716 .shaderStorageImageMultisample
= false,
717 .shaderStorageImageReadWithoutFormat
= false,
718 .shaderStorageImageWriteWithoutFormat
= true,
719 .shaderUniformBufferArrayDynamicIndexing
= true,
720 .shaderSampledImageArrayDynamicIndexing
= true,
721 .shaderStorageBufferArrayDynamicIndexing
= true,
722 .shaderStorageImageArrayDynamicIndexing
= true,
723 .shaderClipDistance
= true,
724 .shaderCullDistance
= true,
725 .shaderFloat64
= pdevice
->info
.gen
>= 8,
726 .shaderInt64
= pdevice
->info
.gen
>= 8,
727 .shaderInt16
= false,
728 .shaderResourceMinLod
= false,
729 .variableMultisampleRate
= false,
730 .inheritedQueries
= true,
733 /* We can't do image stores in vec4 shaders */
734 pFeatures
->vertexPipelineStoresAndAtomics
=
735 pdevice
->compiler
->scalar_stage
[MESA_SHADER_VERTEX
] &&
736 pdevice
->compiler
->scalar_stage
[MESA_SHADER_GEOMETRY
];
739 void anv_GetPhysicalDeviceFeatures2KHR(
740 VkPhysicalDevice physicalDevice
,
741 VkPhysicalDeviceFeatures2KHR
* pFeatures
)
743 anv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
745 vk_foreach_struct(ext
, pFeatures
->pNext
) {
746 switch (ext
->sType
) {
747 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES_KHX
: {
748 VkPhysicalDeviceMultiviewFeaturesKHX
*features
=
749 (VkPhysicalDeviceMultiviewFeaturesKHX
*)ext
;
750 features
->multiview
= true;
751 features
->multiviewGeometryShader
= true;
752 features
->multiviewTessellationShader
= true;
756 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTER_FEATURES_KHR
: {
757 VkPhysicalDeviceVariablePointerFeaturesKHR
*features
= (void *)ext
;
758 features
->variablePointersStorageBuffer
= true;
759 features
->variablePointers
= true;
763 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES_KHR
: {
764 VkPhysicalDeviceSamplerYcbcrConversionFeaturesKHR
*features
=
765 (VkPhysicalDeviceSamplerYcbcrConversionFeaturesKHR
*) ext
;
766 features
->samplerYcbcrConversion
= true;
770 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES_KHR
: {
771 VkPhysicalDevice16BitStorageFeaturesKHR
*features
=
772 (VkPhysicalDevice16BitStorageFeaturesKHR
*)ext
;
774 features
->storageBuffer16BitAccess
= false;
775 features
->uniformAndStorageBuffer16BitAccess
= false;
776 features
->storagePushConstant16
= false;
777 features
->storageInputOutput16
= false;
782 anv_debug_ignored_stype(ext
->sType
);
788 void anv_GetPhysicalDeviceProperties(
789 VkPhysicalDevice physicalDevice
,
790 VkPhysicalDeviceProperties
* pProperties
)
792 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
793 const struct gen_device_info
*devinfo
= &pdevice
->info
;
795 /* See assertions made when programming the buffer surface state. */
796 const uint32_t max_raw_buffer_sz
= devinfo
->gen
>= 7 ?
797 (1ul << 30) : (1ul << 27);
799 const uint32_t max_samplers
= (devinfo
->gen
>= 8 || devinfo
->is_haswell
) ?
802 VkSampleCountFlags sample_counts
=
803 isl_device_get_sample_counts(&pdevice
->isl_dev
);
805 VkPhysicalDeviceLimits limits
= {
806 .maxImageDimension1D
= (1 << 14),
807 .maxImageDimension2D
= (1 << 14),
808 .maxImageDimension3D
= (1 << 11),
809 .maxImageDimensionCube
= (1 << 14),
810 .maxImageArrayLayers
= (1 << 11),
811 .maxTexelBufferElements
= 128 * 1024 * 1024,
812 .maxUniformBufferRange
= (1ul << 27),
813 .maxStorageBufferRange
= max_raw_buffer_sz
,
814 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
815 .maxMemoryAllocationCount
= UINT32_MAX
,
816 .maxSamplerAllocationCount
= 64 * 1024,
817 .bufferImageGranularity
= 64, /* A cache line */
818 .sparseAddressSpaceSize
= 0,
819 .maxBoundDescriptorSets
= MAX_SETS
,
820 .maxPerStageDescriptorSamplers
= max_samplers
,
821 .maxPerStageDescriptorUniformBuffers
= 64,
822 .maxPerStageDescriptorStorageBuffers
= 64,
823 .maxPerStageDescriptorSampledImages
= max_samplers
,
824 .maxPerStageDescriptorStorageImages
= 64,
825 .maxPerStageDescriptorInputAttachments
= 64,
826 .maxPerStageResources
= 250,
827 .maxDescriptorSetSamplers
= 6 * max_samplers
, /* number of stages * maxPerStageDescriptorSamplers */
828 .maxDescriptorSetUniformBuffers
= 6 * 64, /* number of stages * maxPerStageDescriptorUniformBuffers */
829 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
830 .maxDescriptorSetStorageBuffers
= 6 * 64, /* number of stages * maxPerStageDescriptorStorageBuffers */
831 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
832 .maxDescriptorSetSampledImages
= 6 * max_samplers
, /* number of stages * maxPerStageDescriptorSampledImages */
833 .maxDescriptorSetStorageImages
= 6 * 64, /* number of stages * maxPerStageDescriptorStorageImages */
834 .maxDescriptorSetInputAttachments
= 256,
835 .maxVertexInputAttributes
= MAX_VBS
,
836 .maxVertexInputBindings
= MAX_VBS
,
837 .maxVertexInputAttributeOffset
= 2047,
838 .maxVertexInputBindingStride
= 2048,
839 .maxVertexOutputComponents
= 128,
840 .maxTessellationGenerationLevel
= 64,
841 .maxTessellationPatchSize
= 32,
842 .maxTessellationControlPerVertexInputComponents
= 128,
843 .maxTessellationControlPerVertexOutputComponents
= 128,
844 .maxTessellationControlPerPatchOutputComponents
= 128,
845 .maxTessellationControlTotalOutputComponents
= 2048,
846 .maxTessellationEvaluationInputComponents
= 128,
847 .maxTessellationEvaluationOutputComponents
= 128,
848 .maxGeometryShaderInvocations
= 32,
849 .maxGeometryInputComponents
= 64,
850 .maxGeometryOutputComponents
= 128,
851 .maxGeometryOutputVertices
= 256,
852 .maxGeometryTotalOutputComponents
= 1024,
853 .maxFragmentInputComponents
= 128,
854 .maxFragmentOutputAttachments
= 8,
855 .maxFragmentDualSrcAttachments
= 1,
856 .maxFragmentCombinedOutputResources
= 8,
857 .maxComputeSharedMemorySize
= 32768,
858 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
859 .maxComputeWorkGroupInvocations
= 16 * devinfo
->max_cs_threads
,
860 .maxComputeWorkGroupSize
= {
861 16 * devinfo
->max_cs_threads
,
862 16 * devinfo
->max_cs_threads
,
863 16 * devinfo
->max_cs_threads
,
865 .subPixelPrecisionBits
= 4 /* FIXME */,
866 .subTexelPrecisionBits
= 4 /* FIXME */,
867 .mipmapPrecisionBits
= 4 /* FIXME */,
868 .maxDrawIndexedIndexValue
= UINT32_MAX
,
869 .maxDrawIndirectCount
= UINT32_MAX
,
870 .maxSamplerLodBias
= 16,
871 .maxSamplerAnisotropy
= 16,
872 .maxViewports
= MAX_VIEWPORTS
,
873 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
874 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
875 .viewportSubPixelBits
= 13, /* We take a float? */
876 .minMemoryMapAlignment
= 4096, /* A page */
877 .minTexelBufferOffsetAlignment
= 1,
878 /* We need 16 for UBO block reads to work and 32 for push UBOs */
879 .minUniformBufferOffsetAlignment
= 32,
880 .minStorageBufferOffsetAlignment
= 4,
881 .minTexelOffset
= -8,
883 .minTexelGatherOffset
= -32,
884 .maxTexelGatherOffset
= 31,
885 .minInterpolationOffset
= -0.5,
886 .maxInterpolationOffset
= 0.4375,
887 .subPixelInterpolationOffsetBits
= 4,
888 .maxFramebufferWidth
= (1 << 14),
889 .maxFramebufferHeight
= (1 << 14),
890 .maxFramebufferLayers
= (1 << 11),
891 .framebufferColorSampleCounts
= sample_counts
,
892 .framebufferDepthSampleCounts
= sample_counts
,
893 .framebufferStencilSampleCounts
= sample_counts
,
894 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
895 .maxColorAttachments
= MAX_RTS
,
896 .sampledImageColorSampleCounts
= sample_counts
,
897 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
898 .sampledImageDepthSampleCounts
= sample_counts
,
899 .sampledImageStencilSampleCounts
= sample_counts
,
900 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
901 .maxSampleMaskWords
= 1,
902 .timestampComputeAndGraphics
= false,
903 .timestampPeriod
= 1000000000.0 / devinfo
->timestamp_frequency
,
904 .maxClipDistances
= 8,
905 .maxCullDistances
= 8,
906 .maxCombinedClipAndCullDistances
= 8,
907 .discreteQueuePriorities
= 1,
908 .pointSizeRange
= { 0.125, 255.875 },
909 .lineWidthRange
= { 0.0, 7.9921875 },
910 .pointSizeGranularity
= (1.0 / 8.0),
911 .lineWidthGranularity
= (1.0 / 128.0),
912 .strictLines
= false, /* FINISHME */
913 .standardSampleLocations
= true,
914 .optimalBufferCopyOffsetAlignment
= 128,
915 .optimalBufferCopyRowPitchAlignment
= 128,
916 .nonCoherentAtomSize
= 64,
919 *pProperties
= (VkPhysicalDeviceProperties
) {
920 .apiVersion
= anv_physical_device_api_version(pdevice
),
921 .driverVersion
= vk_get_driver_version(),
923 .deviceID
= pdevice
->chipset_id
,
924 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
926 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
929 snprintf(pProperties
->deviceName
, sizeof(pProperties
->deviceName
),
930 "%s", pdevice
->name
);
931 memcpy(pProperties
->pipelineCacheUUID
,
932 pdevice
->pipeline_cache_uuid
, VK_UUID_SIZE
);
935 void anv_GetPhysicalDeviceProperties2KHR(
936 VkPhysicalDevice physicalDevice
,
937 VkPhysicalDeviceProperties2KHR
* pProperties
)
939 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
941 anv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
943 vk_foreach_struct(ext
, pProperties
->pNext
) {
944 switch (ext
->sType
) {
945 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
946 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
947 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
949 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
953 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES_KHR
: {
954 VkPhysicalDeviceIDPropertiesKHR
*id_props
=
955 (VkPhysicalDeviceIDPropertiesKHR
*)ext
;
956 memcpy(id_props
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
957 memcpy(id_props
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
958 /* The LUID is for Windows. */
959 id_props
->deviceLUIDValid
= false;
963 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES_KHX
: {
964 VkPhysicalDeviceMultiviewPropertiesKHX
*properties
=
965 (VkPhysicalDeviceMultiviewPropertiesKHX
*)ext
;
966 properties
->maxMultiviewViewCount
= 16;
967 properties
->maxMultiviewInstanceIndex
= UINT32_MAX
/ 16;
971 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES_KHR
: {
972 VkPhysicalDevicePointClippingPropertiesKHR
*properties
=
973 (VkPhysicalDevicePointClippingPropertiesKHR
*) ext
;
974 properties
->pointClippingBehavior
= VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES_KHR
;
975 anv_finishme("Implement pop-free point clipping");
980 anv_debug_ignored_stype(ext
->sType
);
986 /* We support exactly one queue family. */
987 static const VkQueueFamilyProperties
988 anv_queue_family_properties
= {
989 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
990 VK_QUEUE_COMPUTE_BIT
|
991 VK_QUEUE_TRANSFER_BIT
,
993 .timestampValidBits
= 36, /* XXX: Real value here */
994 .minImageTransferGranularity
= { 1, 1, 1 },
997 void anv_GetPhysicalDeviceQueueFamilyProperties(
998 VkPhysicalDevice physicalDevice
,
1000 VkQueueFamilyProperties
* pQueueFamilyProperties
)
1002 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pCount
);
1004 vk_outarray_append(&out
, p
) {
1005 *p
= anv_queue_family_properties
;
1009 void anv_GetPhysicalDeviceQueueFamilyProperties2KHR(
1010 VkPhysicalDevice physicalDevice
,
1011 uint32_t* pQueueFamilyPropertyCount
,
1012 VkQueueFamilyProperties2KHR
* pQueueFamilyProperties
)
1015 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pQueueFamilyPropertyCount
);
1017 vk_outarray_append(&out
, p
) {
1018 p
->queueFamilyProperties
= anv_queue_family_properties
;
1020 vk_foreach_struct(s
, p
->pNext
) {
1021 anv_debug_ignored_stype(s
->sType
);
1026 void anv_GetPhysicalDeviceMemoryProperties(
1027 VkPhysicalDevice physicalDevice
,
1028 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
1030 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
1032 pMemoryProperties
->memoryTypeCount
= physical_device
->memory
.type_count
;
1033 for (uint32_t i
= 0; i
< physical_device
->memory
.type_count
; i
++) {
1034 pMemoryProperties
->memoryTypes
[i
] = (VkMemoryType
) {
1035 .propertyFlags
= physical_device
->memory
.types
[i
].propertyFlags
,
1036 .heapIndex
= physical_device
->memory
.types
[i
].heapIndex
,
1040 pMemoryProperties
->memoryHeapCount
= physical_device
->memory
.heap_count
;
1041 for (uint32_t i
= 0; i
< physical_device
->memory
.heap_count
; i
++) {
1042 pMemoryProperties
->memoryHeaps
[i
] = (VkMemoryHeap
) {
1043 .size
= physical_device
->memory
.heaps
[i
].size
,
1044 .flags
= physical_device
->memory
.heaps
[i
].flags
,
1049 void anv_GetPhysicalDeviceMemoryProperties2KHR(
1050 VkPhysicalDevice physicalDevice
,
1051 VkPhysicalDeviceMemoryProperties2KHR
* pMemoryProperties
)
1053 anv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
1054 &pMemoryProperties
->memoryProperties
);
1056 vk_foreach_struct(ext
, pMemoryProperties
->pNext
) {
1057 switch (ext
->sType
) {
1059 anv_debug_ignored_stype(ext
->sType
);
1065 PFN_vkVoidFunction
anv_GetInstanceProcAddr(
1066 VkInstance instance
,
1069 return anv_lookup_entrypoint(NULL
, pName
);
1072 /* With version 1+ of the loader interface the ICD should expose
1073 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
1076 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1077 VkInstance instance
,
1081 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1082 VkInstance instance
,
1085 return anv_GetInstanceProcAddr(instance
, pName
);
1088 PFN_vkVoidFunction
anv_GetDeviceProcAddr(
1092 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1093 return anv_lookup_entrypoint(&device
->info
, pName
);
1097 anv_CreateDebugReportCallbackEXT(VkInstance _instance
,
1098 const VkDebugReportCallbackCreateInfoEXT
* pCreateInfo
,
1099 const VkAllocationCallbacks
* pAllocator
,
1100 VkDebugReportCallbackEXT
* pCallback
)
1102 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1103 return vk_create_debug_report_callback(&instance
->debug_report_callbacks
,
1104 pCreateInfo
, pAllocator
, &instance
->alloc
,
1109 anv_DestroyDebugReportCallbackEXT(VkInstance _instance
,
1110 VkDebugReportCallbackEXT _callback
,
1111 const VkAllocationCallbacks
* pAllocator
)
1113 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1114 vk_destroy_debug_report_callback(&instance
->debug_report_callbacks
,
1115 _callback
, pAllocator
, &instance
->alloc
);
1119 anv_DebugReportMessageEXT(VkInstance _instance
,
1120 VkDebugReportFlagsEXT flags
,
1121 VkDebugReportObjectTypeEXT objectType
,
1124 int32_t messageCode
,
1125 const char* pLayerPrefix
,
1126 const char* pMessage
)
1128 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1129 vk_debug_report(&instance
->debug_report_callbacks
, flags
, objectType
,
1130 object
, location
, messageCode
, pLayerPrefix
, pMessage
);
1134 anv_queue_init(struct anv_device
*device
, struct anv_queue
*queue
)
1136 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1137 queue
->device
= device
;
1138 queue
->pool
= &device
->surface_state_pool
;
1142 anv_queue_finish(struct anv_queue
*queue
)
1146 static struct anv_state
1147 anv_state_pool_emit_data(struct anv_state_pool
*pool
, size_t size
, size_t align
, const void *p
)
1149 struct anv_state state
;
1151 state
= anv_state_pool_alloc(pool
, size
, align
);
1152 memcpy(state
.map
, p
, size
);
1154 anv_state_flush(pool
->block_pool
.device
, state
);
1159 struct gen8_border_color
{
1164 /* Pad out to 64 bytes */
1169 anv_device_init_border_colors(struct anv_device
*device
)
1171 static const struct gen8_border_color border_colors
[] = {
1172 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
1173 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
1174 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
1175 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
1176 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
1177 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
1180 device
->border_colors
= anv_state_pool_emit_data(&device
->dynamic_state_pool
,
1181 sizeof(border_colors
), 64,
1186 anv_device_init_trivial_batch(struct anv_device
*device
)
1188 anv_bo_init_new(&device
->trivial_batch_bo
, device
, 4096);
1190 if (device
->instance
->physicalDevice
.has_exec_async
)
1191 device
->trivial_batch_bo
.flags
|= EXEC_OBJECT_ASYNC
;
1193 void *map
= anv_gem_mmap(device
, device
->trivial_batch_bo
.gem_handle
,
1196 struct anv_batch batch
= {
1202 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
1203 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
1205 if (!device
->info
.has_llc
)
1206 gen_clflush_range(map
, batch
.next
- map
);
1208 anv_gem_munmap(map
, device
->trivial_batch_bo
.size
);
1211 VkResult
anv_EnumerateDeviceExtensionProperties(
1212 VkPhysicalDevice physicalDevice
,
1213 const char* pLayerName
,
1214 uint32_t* pPropertyCount
,
1215 VkExtensionProperties
* pProperties
)
1217 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
1218 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
1221 for (int i
= 0; i
< ANV_DEVICE_EXTENSION_COUNT
; i
++) {
1222 if (device
->supported_extensions
.extensions
[i
]) {
1223 vk_outarray_append(&out
, prop
) {
1224 *prop
= anv_device_extensions
[i
];
1229 return vk_outarray_status(&out
);
1232 VkResult
anv_CreateDevice(
1233 VkPhysicalDevice physicalDevice
,
1234 const VkDeviceCreateInfo
* pCreateInfo
,
1235 const VkAllocationCallbacks
* pAllocator
,
1238 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
1240 struct anv_device
*device
;
1242 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO
);
1244 struct anv_device_extension_table enabled_extensions
;
1245 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
1247 for (idx
= 0; idx
< ANV_DEVICE_EXTENSION_COUNT
; idx
++) {
1248 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
1249 anv_device_extensions
[idx
].extensionName
) == 0)
1253 if (idx
>= ANV_DEVICE_EXTENSION_COUNT
)
1254 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
1256 if (!physical_device
->supported_extensions
.extensions
[idx
])
1257 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
1259 enabled_extensions
.extensions
[idx
] = true;
1262 /* Check enabled features */
1263 if (pCreateInfo
->pEnabledFeatures
) {
1264 VkPhysicalDeviceFeatures supported_features
;
1265 anv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
1266 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
1267 VkBool32
*enabled_feature
= (VkBool32
*)pCreateInfo
->pEnabledFeatures
;
1268 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
1269 for (uint32_t i
= 0; i
< num_features
; i
++) {
1270 if (enabled_feature
[i
] && !supported_feature
[i
])
1271 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
1275 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
1277 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
1279 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1281 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1282 device
->instance
= physical_device
->instance
;
1283 device
->chipset_id
= physical_device
->chipset_id
;
1284 device
->lost
= false;
1287 device
->alloc
= *pAllocator
;
1289 device
->alloc
= physical_device
->instance
->alloc
;
1291 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
1292 device
->fd
= open(physical_device
->path
, O_RDWR
| O_CLOEXEC
);
1293 if (device
->fd
== -1) {
1294 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1298 device
->context_id
= anv_gem_create_context(device
);
1299 if (device
->context_id
== -1) {
1300 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1304 device
->info
= physical_device
->info
;
1305 device
->isl_dev
= physical_device
->isl_dev
;
1307 /* On Broadwell and later, we can use batch chaining to more efficiently
1308 * implement growing command buffers. Prior to Haswell, the kernel
1309 * command parser gets in the way and we have to fall back to growing
1312 device
->can_chain_batches
= device
->info
.gen
>= 8;
1314 device
->robust_buffer_access
= pCreateInfo
->pEnabledFeatures
&&
1315 pCreateInfo
->pEnabledFeatures
->robustBufferAccess
;
1316 device
->enabled_extensions
= enabled_extensions
;
1318 if (pthread_mutex_init(&device
->mutex
, NULL
) != 0) {
1319 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1320 goto fail_context_id
;
1323 pthread_condattr_t condattr
;
1324 if (pthread_condattr_init(&condattr
) != 0) {
1325 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1328 if (pthread_condattr_setclock(&condattr
, CLOCK_MONOTONIC
) != 0) {
1329 pthread_condattr_destroy(&condattr
);
1330 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1333 if (pthread_cond_init(&device
->queue_submit
, NULL
) != 0) {
1334 pthread_condattr_destroy(&condattr
);
1335 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1338 pthread_condattr_destroy(&condattr
);
1341 (physical_device
->supports_48bit_addresses
? EXEC_OBJECT_SUPPORTS_48B_ADDRESS
: 0) |
1342 (physical_device
->has_exec_async
? EXEC_OBJECT_ASYNC
: 0) |
1343 (physical_device
->has_exec_capture
? EXEC_OBJECT_CAPTURE
: 0);
1345 anv_bo_pool_init(&device
->batch_bo_pool
, device
, bo_flags
);
1347 result
= anv_bo_cache_init(&device
->bo_cache
);
1348 if (result
!= VK_SUCCESS
)
1349 goto fail_batch_bo_pool
;
1351 /* For the state pools we explicitly disable 48bit. */
1352 bo_flags
= (physical_device
->has_exec_async
? EXEC_OBJECT_ASYNC
: 0) |
1353 (physical_device
->has_exec_capture
? EXEC_OBJECT_CAPTURE
: 0);
1355 result
= anv_state_pool_init(&device
->dynamic_state_pool
, device
, 16384,
1357 if (result
!= VK_SUCCESS
)
1360 result
= anv_state_pool_init(&device
->instruction_state_pool
, device
, 16384,
1362 if (result
!= VK_SUCCESS
)
1363 goto fail_dynamic_state_pool
;
1365 result
= anv_state_pool_init(&device
->surface_state_pool
, device
, 4096,
1367 if (result
!= VK_SUCCESS
)
1368 goto fail_instruction_state_pool
;
1370 result
= anv_bo_init_new(&device
->workaround_bo
, device
, 1024);
1371 if (result
!= VK_SUCCESS
)
1372 goto fail_surface_state_pool
;
1374 anv_device_init_trivial_batch(device
);
1376 anv_scratch_pool_init(device
, &device
->scratch_pool
);
1378 anv_queue_init(device
, &device
->queue
);
1380 switch (device
->info
.gen
) {
1382 if (!device
->info
.is_haswell
)
1383 result
= gen7_init_device_state(device
);
1385 result
= gen75_init_device_state(device
);
1388 result
= gen8_init_device_state(device
);
1391 result
= gen9_init_device_state(device
);
1394 result
= gen10_init_device_state(device
);
1397 /* Shouldn't get here as we don't create physical devices for any other
1399 unreachable("unhandled gen");
1401 if (result
!= VK_SUCCESS
)
1402 goto fail_workaround_bo
;
1404 anv_device_init_blorp(device
);
1406 anv_device_init_border_colors(device
);
1408 *pDevice
= anv_device_to_handle(device
);
1413 anv_queue_finish(&device
->queue
);
1414 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
1415 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
1416 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
1417 fail_surface_state_pool
:
1418 anv_state_pool_finish(&device
->surface_state_pool
);
1419 fail_instruction_state_pool
:
1420 anv_state_pool_finish(&device
->instruction_state_pool
);
1421 fail_dynamic_state_pool
:
1422 anv_state_pool_finish(&device
->dynamic_state_pool
);
1424 anv_bo_cache_finish(&device
->bo_cache
);
1426 anv_bo_pool_finish(&device
->batch_bo_pool
);
1427 pthread_cond_destroy(&device
->queue_submit
);
1429 pthread_mutex_destroy(&device
->mutex
);
1431 anv_gem_destroy_context(device
, device
->context_id
);
1435 vk_free(&device
->alloc
, device
);
1440 void anv_DestroyDevice(
1442 const VkAllocationCallbacks
* pAllocator
)
1444 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1449 anv_device_finish_blorp(device
);
1451 anv_queue_finish(&device
->queue
);
1453 #ifdef HAVE_VALGRIND
1454 /* We only need to free these to prevent valgrind errors. The backing
1455 * BO will go away in a couple of lines so we don't actually leak.
1457 anv_state_pool_free(&device
->dynamic_state_pool
, device
->border_colors
);
1460 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
1462 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
1463 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
1465 anv_gem_close(device
, device
->trivial_batch_bo
.gem_handle
);
1467 anv_state_pool_finish(&device
->surface_state_pool
);
1468 anv_state_pool_finish(&device
->instruction_state_pool
);
1469 anv_state_pool_finish(&device
->dynamic_state_pool
);
1471 anv_bo_cache_finish(&device
->bo_cache
);
1473 anv_bo_pool_finish(&device
->batch_bo_pool
);
1475 pthread_cond_destroy(&device
->queue_submit
);
1476 pthread_mutex_destroy(&device
->mutex
);
1478 anv_gem_destroy_context(device
, device
->context_id
);
1482 vk_free(&device
->alloc
, device
);
1485 VkResult
anv_EnumerateInstanceLayerProperties(
1486 uint32_t* pPropertyCount
,
1487 VkLayerProperties
* pProperties
)
1489 if (pProperties
== NULL
) {
1490 *pPropertyCount
= 0;
1494 /* None supported at this time */
1495 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1498 VkResult
anv_EnumerateDeviceLayerProperties(
1499 VkPhysicalDevice physicalDevice
,
1500 uint32_t* pPropertyCount
,
1501 VkLayerProperties
* pProperties
)
1503 if (pProperties
== NULL
) {
1504 *pPropertyCount
= 0;
1508 /* None supported at this time */
1509 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1512 void anv_GetDeviceQueue(
1514 uint32_t queueNodeIndex
,
1515 uint32_t queueIndex
,
1518 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1520 assert(queueIndex
== 0);
1522 *pQueue
= anv_queue_to_handle(&device
->queue
);
1526 anv_device_query_status(struct anv_device
*device
)
1528 /* This isn't likely as most of the callers of this function already check
1529 * for it. However, it doesn't hurt to check and it potentially lets us
1532 if (unlikely(device
->lost
))
1533 return VK_ERROR_DEVICE_LOST
;
1535 uint32_t active
, pending
;
1536 int ret
= anv_gem_gpu_get_reset_stats(device
, &active
, &pending
);
1538 /* We don't know the real error. */
1539 device
->lost
= true;
1540 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1541 "get_reset_stats failed: %m");
1545 device
->lost
= true;
1546 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1547 "GPU hung on one of our command buffers");
1548 } else if (pending
) {
1549 device
->lost
= true;
1550 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1551 "GPU hung with commands in-flight");
1558 anv_device_bo_busy(struct anv_device
*device
, struct anv_bo
*bo
)
1560 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
1561 * Other usages of the BO (such as on different hardware) will not be
1562 * flagged as "busy" by this ioctl. Use with care.
1564 int ret
= anv_gem_busy(device
, bo
->gem_handle
);
1566 return VK_NOT_READY
;
1567 } else if (ret
== -1) {
1568 /* We don't know the real error. */
1569 device
->lost
= true;
1570 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1571 "gem wait failed: %m");
1574 /* Query for device status after the busy call. If the BO we're checking
1575 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
1576 * client because it clearly doesn't have valid data. Yes, this most
1577 * likely means an ioctl, but we just did an ioctl to query the busy status
1578 * so it's no great loss.
1580 return anv_device_query_status(device
);
1584 anv_device_wait(struct anv_device
*device
, struct anv_bo
*bo
,
1587 int ret
= anv_gem_wait(device
, bo
->gem_handle
, &timeout
);
1588 if (ret
== -1 && errno
== ETIME
) {
1590 } else if (ret
== -1) {
1591 /* We don't know the real error. */
1592 device
->lost
= true;
1593 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1594 "gem wait failed: %m");
1597 /* Query for device status after the wait. If the BO we're waiting on got
1598 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
1599 * because it clearly doesn't have valid data. Yes, this most likely means
1600 * an ioctl, but we just did an ioctl to wait so it's no great loss.
1602 return anv_device_query_status(device
);
1605 VkResult
anv_DeviceWaitIdle(
1608 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1609 if (unlikely(device
->lost
))
1610 return VK_ERROR_DEVICE_LOST
;
1612 struct anv_batch batch
;
1615 batch
.start
= batch
.next
= cmds
;
1616 batch
.end
= (void *) cmds
+ sizeof(cmds
);
1618 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
1619 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
1621 return anv_device_submit_simple_batch(device
, &batch
);
1625 anv_bo_init_new(struct anv_bo
*bo
, struct anv_device
*device
, uint64_t size
)
1627 uint32_t gem_handle
= anv_gem_create(device
, size
);
1629 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
1631 anv_bo_init(bo
, gem_handle
, size
);
1636 VkResult
anv_AllocateMemory(
1638 const VkMemoryAllocateInfo
* pAllocateInfo
,
1639 const VkAllocationCallbacks
* pAllocator
,
1640 VkDeviceMemory
* pMem
)
1642 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1643 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
1644 struct anv_device_memory
*mem
;
1645 VkResult result
= VK_SUCCESS
;
1647 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
1649 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
1650 assert(pAllocateInfo
->allocationSize
> 0);
1652 /* The kernel relocation API has a limitation of a 32-bit delta value
1653 * applied to the address before it is written which, in spite of it being
1654 * unsigned, is treated as signed . Because of the way that this maps to
1655 * the Vulkan API, we cannot handle an offset into a buffer that does not
1656 * fit into a signed 32 bits. The only mechanism we have for dealing with
1657 * this at the moment is to limit all VkDeviceMemory objects to a maximum
1658 * of 2GB each. The Vulkan spec allows us to do this:
1660 * "Some platforms may have a limit on the maximum size of a single
1661 * allocation. For example, certain systems may fail to create
1662 * allocations with a size greater than or equal to 4GB. Such a limit is
1663 * implementation-dependent, and if such a failure occurs then the error
1664 * VK_ERROR_OUT_OF_DEVICE_MEMORY should be returned."
1666 * We don't use vk_error here because it's not an error so much as an
1667 * indication to the application that the allocation is too large.
1669 if (pAllocateInfo
->allocationSize
> (1ull << 31))
1670 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
1672 /* FINISHME: Fail if allocation request exceeds heap size. */
1674 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
1675 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1677 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1679 assert(pAllocateInfo
->memoryTypeIndex
< pdevice
->memory
.type_count
);
1680 mem
->type
= &pdevice
->memory
.types
[pAllocateInfo
->memoryTypeIndex
];
1684 const VkImportMemoryFdInfoKHR
*fd_info
=
1685 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_FD_INFO_KHR
);
1687 /* The Vulkan spec permits handleType to be 0, in which case the struct is
1690 if (fd_info
&& fd_info
->handleType
) {
1691 /* At the moment, we support only the below handle types. */
1692 assert(fd_info
->handleType
==
1693 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
||
1694 fd_info
->handleType
==
1695 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
1697 result
= anv_bo_cache_import(device
, &device
->bo_cache
,
1698 fd_info
->fd
, &mem
->bo
);
1699 if (result
!= VK_SUCCESS
)
1702 VkDeviceSize aligned_alloc_size
=
1703 align_u64(pAllocateInfo
->allocationSize
, 4096);
1705 /* For security purposes, we reject importing the bo if it's smaller
1706 * than the requested allocation size. This prevents a malicious client
1707 * from passing a buffer to a trusted client, lying about the size, and
1708 * telling the trusted client to try and texture from an image that goes
1709 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
1710 * in the trusted client. The trusted client can protect itself against
1711 * this sort of attack but only if it can trust the buffer size.
1713 if (mem
->bo
->size
< aligned_alloc_size
) {
1714 result
= vk_errorf(device
->instance
, device
,
1715 VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
,
1716 "aligned allocationSize too large for "
1717 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR: "
1718 "%"PRIu64
"B > %"PRIu64
"B",
1719 aligned_alloc_size
, mem
->bo
->size
);
1720 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
1724 /* From the Vulkan spec:
1726 * "Importing memory from a file descriptor transfers ownership of
1727 * the file descriptor from the application to the Vulkan
1728 * implementation. The application must not perform any operations on
1729 * the file descriptor after a successful import."
1731 * If the import fails, we leave the file descriptor open.
1735 result
= anv_bo_cache_alloc(device
, &device
->bo_cache
,
1736 pAllocateInfo
->allocationSize
,
1738 if (result
!= VK_SUCCESS
)
1741 const VkMemoryDedicatedAllocateInfoKHR
*dedicated_info
=
1742 vk_find_struct_const(pAllocateInfo
->pNext
, MEMORY_DEDICATED_ALLOCATE_INFO_KHR
);
1743 if (dedicated_info
&& dedicated_info
->image
!= VK_NULL_HANDLE
) {
1744 ANV_FROM_HANDLE(anv_image
, image
, dedicated_info
->image
);
1746 /* For images using modifiers, we require a dedicated allocation
1747 * and we set the BO tiling to match the tiling of the underlying
1748 * modifier. This is a bit unfortunate as this is completely
1749 * pointless for Vulkan. However, GL needs to be able to map things
1750 * so it needs the tiling to be set. The only way to do this in a
1751 * non-racy way is to set the tiling in the creator of the BO so that
1754 * One of these days, once the GL driver learns to not map things
1755 * through the GTT in random places, we can drop this and start
1756 * allowing multiple modified images in the same BO.
1758 if (image
->drm_format_mod
!= DRM_FORMAT_MOD_INVALID
) {
1759 assert(isl_drm_modifier_get_info(image
->drm_format_mod
)->tiling
==
1760 image
->planes
[0].surface
.isl
.tiling
);
1761 const uint32_t i915_tiling
=
1762 isl_tiling_to_i915_tiling(image
->planes
[0].surface
.isl
.tiling
);
1763 int ret
= anv_gem_set_tiling(device
, mem
->bo
->gem_handle
,
1764 image
->planes
[0].surface
.isl
.row_pitch
,
1767 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
1768 return vk_errorf(device
->instance
, NULL
,
1769 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
1770 "failed to set BO tiling: %m");
1776 assert(mem
->type
->heapIndex
< pdevice
->memory
.heap_count
);
1777 if (pdevice
->memory
.heaps
[mem
->type
->heapIndex
].supports_48bit_addresses
)
1778 mem
->bo
->flags
|= EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
1780 const struct wsi_memory_allocate_info
*wsi_info
=
1781 vk_find_struct_const(pAllocateInfo
->pNext
, WSI_MEMORY_ALLOCATE_INFO_MESA
);
1782 if (wsi_info
&& wsi_info
->implicit_sync
) {
1783 /* We need to set the WRITE flag on window system buffers so that GEM
1784 * will know we're writing to them and synchronize uses on other rings
1785 * (eg if the display server uses the blitter ring).
1787 mem
->bo
->flags
|= EXEC_OBJECT_WRITE
;
1788 } else if (pdevice
->has_exec_async
) {
1789 mem
->bo
->flags
|= EXEC_OBJECT_ASYNC
;
1792 *pMem
= anv_device_memory_to_handle(mem
);
1797 vk_free2(&device
->alloc
, pAllocator
, mem
);
1802 VkResult
anv_GetMemoryFdKHR(
1804 const VkMemoryGetFdInfoKHR
* pGetFdInfo
,
1807 ANV_FROM_HANDLE(anv_device
, dev
, device_h
);
1808 ANV_FROM_HANDLE(anv_device_memory
, mem
, pGetFdInfo
->memory
);
1810 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
1812 assert(pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
||
1813 pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
1815 return anv_bo_cache_export(dev
, &dev
->bo_cache
, mem
->bo
, pFd
);
1818 VkResult
anv_GetMemoryFdPropertiesKHR(
1820 VkExternalMemoryHandleTypeFlagBitsKHR handleType
,
1822 VkMemoryFdPropertiesKHR
* pMemoryFdProperties
)
1824 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1825 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
1827 switch (handleType
) {
1828 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
:
1829 /* dma-buf can be imported as any memory type */
1830 pMemoryFdProperties
->memoryTypeBits
=
1831 (1 << pdevice
->memory
.type_count
) - 1;
1835 /* The valid usage section for this function says:
1837 * "handleType must not be one of the handle types defined as
1840 * So opaque handle types fall into the default "unsupported" case.
1842 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
1846 void anv_FreeMemory(
1848 VkDeviceMemory _mem
,
1849 const VkAllocationCallbacks
* pAllocator
)
1851 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1852 ANV_FROM_HANDLE(anv_device_memory
, mem
, _mem
);
1858 anv_UnmapMemory(_device
, _mem
);
1860 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
1862 vk_free2(&device
->alloc
, pAllocator
, mem
);
1865 VkResult
anv_MapMemory(
1867 VkDeviceMemory _memory
,
1868 VkDeviceSize offset
,
1870 VkMemoryMapFlags flags
,
1873 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1874 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
1881 if (size
== VK_WHOLE_SIZE
)
1882 size
= mem
->bo
->size
- offset
;
1884 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
1886 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
1887 * assert(size != 0);
1888 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
1889 * equal to the size of the memory minus offset
1892 assert(offset
+ size
<= mem
->bo
->size
);
1894 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
1895 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
1896 * at a time is valid. We could just mmap up front and return an offset
1897 * pointer here, but that may exhaust virtual memory on 32 bit
1900 uint32_t gem_flags
= 0;
1902 if (!device
->info
.has_llc
&&
1903 (mem
->type
->propertyFlags
& VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
))
1904 gem_flags
|= I915_MMAP_WC
;
1906 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
1907 uint64_t map_offset
= offset
& ~4095ull;
1908 assert(offset
>= map_offset
);
1909 uint64_t map_size
= (offset
+ size
) - map_offset
;
1911 /* Let's map whole pages */
1912 map_size
= align_u64(map_size
, 4096);
1914 void *map
= anv_gem_mmap(device
, mem
->bo
->gem_handle
,
1915 map_offset
, map_size
, gem_flags
);
1916 if (map
== MAP_FAILED
)
1917 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
1920 mem
->map_size
= map_size
;
1922 *ppData
= mem
->map
+ (offset
- map_offset
);
1927 void anv_UnmapMemory(
1929 VkDeviceMemory _memory
)
1931 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
1936 anv_gem_munmap(mem
->map
, mem
->map_size
);
1943 clflush_mapped_ranges(struct anv_device
*device
,
1945 const VkMappedMemoryRange
*ranges
)
1947 for (uint32_t i
= 0; i
< count
; i
++) {
1948 ANV_FROM_HANDLE(anv_device_memory
, mem
, ranges
[i
].memory
);
1949 if (ranges
[i
].offset
>= mem
->map_size
)
1952 gen_clflush_range(mem
->map
+ ranges
[i
].offset
,
1953 MIN2(ranges
[i
].size
, mem
->map_size
- ranges
[i
].offset
));
1957 VkResult
anv_FlushMappedMemoryRanges(
1959 uint32_t memoryRangeCount
,
1960 const VkMappedMemoryRange
* pMemoryRanges
)
1962 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1964 if (device
->info
.has_llc
)
1967 /* Make sure the writes we're flushing have landed. */
1968 __builtin_ia32_mfence();
1970 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
1975 VkResult
anv_InvalidateMappedMemoryRanges(
1977 uint32_t memoryRangeCount
,
1978 const VkMappedMemoryRange
* pMemoryRanges
)
1980 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1982 if (device
->info
.has_llc
)
1985 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
1987 /* Make sure no reads get moved up above the invalidate. */
1988 __builtin_ia32_mfence();
1993 void anv_GetBufferMemoryRequirements(
1996 VkMemoryRequirements
* pMemoryRequirements
)
1998 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
1999 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2000 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2002 /* The Vulkan spec (git aaed022) says:
2004 * memoryTypeBits is a bitfield and contains one bit set for every
2005 * supported memory type for the resource. The bit `1<<i` is set if and
2006 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
2007 * structure for the physical device is supported.
2009 uint32_t memory_types
= 0;
2010 for (uint32_t i
= 0; i
< pdevice
->memory
.type_count
; i
++) {
2011 uint32_t valid_usage
= pdevice
->memory
.types
[i
].valid_buffer_usage
;
2012 if ((valid_usage
& buffer
->usage
) == buffer
->usage
)
2013 memory_types
|= (1u << i
);
2016 /* Base alignment requirement of a cache line */
2017 uint32_t alignment
= 16;
2019 /* We need an alignment of 32 for pushing UBOs */
2020 if (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
)
2021 alignment
= MAX2(alignment
, 32);
2023 pMemoryRequirements
->size
= buffer
->size
;
2024 pMemoryRequirements
->alignment
= alignment
;
2025 pMemoryRequirements
->memoryTypeBits
= memory_types
;
2028 void anv_GetBufferMemoryRequirements2KHR(
2030 const VkBufferMemoryRequirementsInfo2KHR
* pInfo
,
2031 VkMemoryRequirements2KHR
* pMemoryRequirements
)
2033 anv_GetBufferMemoryRequirements(_device
, pInfo
->buffer
,
2034 &pMemoryRequirements
->memoryRequirements
);
2036 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
2037 switch (ext
->sType
) {
2038 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS_KHR
: {
2039 VkMemoryDedicatedRequirementsKHR
*requirements
= (void *)ext
;
2040 requirements
->prefersDedicatedAllocation
= VK_FALSE
;
2041 requirements
->requiresDedicatedAllocation
= VK_FALSE
;
2046 anv_debug_ignored_stype(ext
->sType
);
2052 void anv_GetImageMemoryRequirements(
2055 VkMemoryRequirements
* pMemoryRequirements
)
2057 ANV_FROM_HANDLE(anv_image
, image
, _image
);
2058 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2059 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2061 /* The Vulkan spec (git aaed022) says:
2063 * memoryTypeBits is a bitfield and contains one bit set for every
2064 * supported memory type for the resource. The bit `1<<i` is set if and
2065 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
2066 * structure for the physical device is supported.
2068 * All types are currently supported for images.
2070 uint32_t memory_types
= (1ull << pdevice
->memory
.type_count
) - 1;
2072 pMemoryRequirements
->size
= image
->size
;
2073 pMemoryRequirements
->alignment
= image
->alignment
;
2074 pMemoryRequirements
->memoryTypeBits
= memory_types
;
2077 void anv_GetImageMemoryRequirements2KHR(
2079 const VkImageMemoryRequirementsInfo2KHR
* pInfo
,
2080 VkMemoryRequirements2KHR
* pMemoryRequirements
)
2082 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2083 ANV_FROM_HANDLE(anv_image
, image
, pInfo
->image
);
2085 anv_GetImageMemoryRequirements(_device
, pInfo
->image
,
2086 &pMemoryRequirements
->memoryRequirements
);
2088 vk_foreach_struct_const(ext
, pInfo
->pNext
) {
2089 switch (ext
->sType
) {
2090 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO_KHR
: {
2091 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2092 const VkImagePlaneMemoryRequirementsInfoKHR
*plane_reqs
=
2093 (const VkImagePlaneMemoryRequirementsInfoKHR
*) ext
;
2094 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
,
2095 plane_reqs
->planeAspect
);
2097 assert(image
->planes
[plane
].offset
== 0);
2099 /* The Vulkan spec (git aaed022) says:
2101 * memoryTypeBits is a bitfield and contains one bit set for every
2102 * supported memory type for the resource. The bit `1<<i` is set
2103 * if and only if the memory type `i` in the
2104 * VkPhysicalDeviceMemoryProperties structure for the physical
2105 * device is supported.
2107 * All types are currently supported for images.
2109 pMemoryRequirements
->memoryRequirements
.memoryTypeBits
=
2110 (1ull << pdevice
->memory
.type_count
) - 1;
2112 pMemoryRequirements
->memoryRequirements
.size
= image
->planes
[plane
].size
;
2113 pMemoryRequirements
->memoryRequirements
.alignment
=
2114 image
->planes
[plane
].alignment
;
2119 anv_debug_ignored_stype(ext
->sType
);
2124 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
2125 switch (ext
->sType
) {
2126 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS_KHR
: {
2127 VkMemoryDedicatedRequirementsKHR
*requirements
= (void *)ext
;
2128 if (image
->drm_format_mod
!= DRM_FORMAT_MOD_INVALID
) {
2129 /* Require a dedicated allocation for images with modifiers.
2131 * See also anv_AllocateMemory.
2133 requirements
->prefersDedicatedAllocation
= VK_TRUE
;
2134 requirements
->requiresDedicatedAllocation
= VK_TRUE
;
2136 requirements
->prefersDedicatedAllocation
= VK_FALSE
;
2137 requirements
->requiresDedicatedAllocation
= VK_FALSE
;
2143 anv_debug_ignored_stype(ext
->sType
);
2149 void anv_GetImageSparseMemoryRequirements(
2152 uint32_t* pSparseMemoryRequirementCount
,
2153 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
2155 *pSparseMemoryRequirementCount
= 0;
2158 void anv_GetImageSparseMemoryRequirements2KHR(
2160 const VkImageSparseMemoryRequirementsInfo2KHR
* pInfo
,
2161 uint32_t* pSparseMemoryRequirementCount
,
2162 VkSparseImageMemoryRequirements2KHR
* pSparseMemoryRequirements
)
2164 *pSparseMemoryRequirementCount
= 0;
2167 void anv_GetDeviceMemoryCommitment(
2169 VkDeviceMemory memory
,
2170 VkDeviceSize
* pCommittedMemoryInBytes
)
2172 *pCommittedMemoryInBytes
= 0;
2176 anv_bind_buffer_memory(const VkBindBufferMemoryInfoKHR
*pBindInfo
)
2178 ANV_FROM_HANDLE(anv_device_memory
, mem
, pBindInfo
->memory
);
2179 ANV_FROM_HANDLE(anv_buffer
, buffer
, pBindInfo
->buffer
);
2181 assert(pBindInfo
->sType
== VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO_KHR
);
2184 assert((buffer
->usage
& mem
->type
->valid_buffer_usage
) == buffer
->usage
);
2185 buffer
->bo
= mem
->bo
;
2186 buffer
->offset
= pBindInfo
->memoryOffset
;
2193 VkResult
anv_BindBufferMemory(
2196 VkDeviceMemory memory
,
2197 VkDeviceSize memoryOffset
)
2199 anv_bind_buffer_memory(
2200 &(VkBindBufferMemoryInfoKHR
) {
2201 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO_KHR
,
2204 .memoryOffset
= memoryOffset
,
2210 VkResult
anv_BindBufferMemory2KHR(
2212 uint32_t bindInfoCount
,
2213 const VkBindBufferMemoryInfoKHR
* pBindInfos
)
2215 for (uint32_t i
= 0; i
< bindInfoCount
; i
++)
2216 anv_bind_buffer_memory(&pBindInfos
[i
]);
2221 VkResult
anv_QueueBindSparse(
2223 uint32_t bindInfoCount
,
2224 const VkBindSparseInfo
* pBindInfo
,
2227 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
2228 if (unlikely(queue
->device
->lost
))
2229 return VK_ERROR_DEVICE_LOST
;
2231 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
2236 VkResult
anv_CreateEvent(
2238 const VkEventCreateInfo
* pCreateInfo
,
2239 const VkAllocationCallbacks
* pAllocator
,
2242 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2243 struct anv_state state
;
2244 struct anv_event
*event
;
2246 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_EVENT_CREATE_INFO
);
2248 state
= anv_state_pool_alloc(&device
->dynamic_state_pool
,
2251 event
->state
= state
;
2252 event
->semaphore
= VK_EVENT_RESET
;
2254 if (!device
->info
.has_llc
) {
2255 /* Make sure the writes we're flushing have landed. */
2256 __builtin_ia32_mfence();
2257 __builtin_ia32_clflush(event
);
2260 *pEvent
= anv_event_to_handle(event
);
2265 void anv_DestroyEvent(
2268 const VkAllocationCallbacks
* pAllocator
)
2270 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2271 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2276 anv_state_pool_free(&device
->dynamic_state_pool
, event
->state
);
2279 VkResult
anv_GetEventStatus(
2283 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2284 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2286 if (unlikely(device
->lost
))
2287 return VK_ERROR_DEVICE_LOST
;
2289 if (!device
->info
.has_llc
) {
2290 /* Invalidate read cache before reading event written by GPU. */
2291 __builtin_ia32_clflush(event
);
2292 __builtin_ia32_mfence();
2296 return event
->semaphore
;
2299 VkResult
anv_SetEvent(
2303 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2304 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2306 event
->semaphore
= VK_EVENT_SET
;
2308 if (!device
->info
.has_llc
) {
2309 /* Make sure the writes we're flushing have landed. */
2310 __builtin_ia32_mfence();
2311 __builtin_ia32_clflush(event
);
2317 VkResult
anv_ResetEvent(
2321 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2322 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2324 event
->semaphore
= VK_EVENT_RESET
;
2326 if (!device
->info
.has_llc
) {
2327 /* Make sure the writes we're flushing have landed. */
2328 __builtin_ia32_mfence();
2329 __builtin_ia32_clflush(event
);
2337 VkResult
anv_CreateBuffer(
2339 const VkBufferCreateInfo
* pCreateInfo
,
2340 const VkAllocationCallbacks
* pAllocator
,
2343 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2344 struct anv_buffer
*buffer
;
2346 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
2348 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
2349 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2351 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2353 buffer
->size
= pCreateInfo
->size
;
2354 buffer
->usage
= pCreateInfo
->usage
;
2358 *pBuffer
= anv_buffer_to_handle(buffer
);
2363 void anv_DestroyBuffer(
2366 const VkAllocationCallbacks
* pAllocator
)
2368 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2369 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
2374 vk_free2(&device
->alloc
, pAllocator
, buffer
);
2378 anv_fill_buffer_surface_state(struct anv_device
*device
, struct anv_state state
,
2379 enum isl_format format
,
2380 uint32_t offset
, uint32_t range
, uint32_t stride
)
2382 isl_buffer_fill_state(&device
->isl_dev
, state
.map
,
2384 .mocs
= device
->default_mocs
,
2389 anv_state_flush(device
, state
);
2392 void anv_DestroySampler(
2395 const VkAllocationCallbacks
* pAllocator
)
2397 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2398 ANV_FROM_HANDLE(anv_sampler
, sampler
, _sampler
);
2403 vk_free2(&device
->alloc
, pAllocator
, sampler
);
2406 VkResult
anv_CreateFramebuffer(
2408 const VkFramebufferCreateInfo
* pCreateInfo
,
2409 const VkAllocationCallbacks
* pAllocator
,
2410 VkFramebuffer
* pFramebuffer
)
2412 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2413 struct anv_framebuffer
*framebuffer
;
2415 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
2417 size_t size
= sizeof(*framebuffer
) +
2418 sizeof(struct anv_image_view
*) * pCreateInfo
->attachmentCount
;
2419 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
2420 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2421 if (framebuffer
== NULL
)
2422 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2424 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
2425 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
2426 VkImageView _iview
= pCreateInfo
->pAttachments
[i
];
2427 framebuffer
->attachments
[i
] = anv_image_view_from_handle(_iview
);
2430 framebuffer
->width
= pCreateInfo
->width
;
2431 framebuffer
->height
= pCreateInfo
->height
;
2432 framebuffer
->layers
= pCreateInfo
->layers
;
2434 *pFramebuffer
= anv_framebuffer_to_handle(framebuffer
);
2439 void anv_DestroyFramebuffer(
2442 const VkAllocationCallbacks
* pAllocator
)
2444 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2445 ANV_FROM_HANDLE(anv_framebuffer
, fb
, _fb
);
2450 vk_free2(&device
->alloc
, pAllocator
, fb
);
2453 /* vk_icd.h does not declare this function, so we declare it here to
2454 * suppress Wmissing-prototypes.
2456 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2457 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
);
2459 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2460 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
)
2462 /* For the full details on loader interface versioning, see
2463 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
2464 * What follows is a condensed summary, to help you navigate the large and
2465 * confusing official doc.
2467 * - Loader interface v0 is incompatible with later versions. We don't
2470 * - In loader interface v1:
2471 * - The first ICD entrypoint called by the loader is
2472 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
2474 * - The ICD must statically expose no other Vulkan symbol unless it is
2475 * linked with -Bsymbolic.
2476 * - Each dispatchable Vulkan handle created by the ICD must be
2477 * a pointer to a struct whose first member is VK_LOADER_DATA. The
2478 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
2479 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
2480 * vkDestroySurfaceKHR(). The ICD must be capable of working with
2481 * such loader-managed surfaces.
2483 * - Loader interface v2 differs from v1 in:
2484 * - The first ICD entrypoint called by the loader is
2485 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
2486 * statically expose this entrypoint.
2488 * - Loader interface v3 differs from v2 in:
2489 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
2490 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
2491 * because the loader no longer does so.
2493 *pSupportedVersion
= MIN2(*pSupportedVersion
, 3u);