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 device
->local_fd
= fd
;
445 anv_physical_device_finish(struct anv_physical_device
*device
)
447 anv_finish_wsi(device
);
448 ralloc_free(device
->compiler
);
449 close(device
->local_fd
);
453 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
454 VkSystemAllocationScope allocationScope
)
460 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
461 size_t align
, VkSystemAllocationScope allocationScope
)
463 return realloc(pOriginal
, size
);
467 default_free_func(void *pUserData
, void *pMemory
)
472 static const VkAllocationCallbacks default_alloc
= {
474 .pfnAllocation
= default_alloc_func
,
475 .pfnReallocation
= default_realloc_func
,
476 .pfnFree
= default_free_func
,
479 VkResult
anv_EnumerateInstanceExtensionProperties(
480 const char* pLayerName
,
481 uint32_t* pPropertyCount
,
482 VkExtensionProperties
* pProperties
)
484 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
486 for (int i
= 0; i
< ANV_INSTANCE_EXTENSION_COUNT
; i
++) {
487 if (anv_instance_extensions_supported
.extensions
[i
]) {
488 vk_outarray_append(&out
, prop
) {
489 *prop
= anv_instance_extensions
[i
];
494 return vk_outarray_status(&out
);
497 VkResult
anv_CreateInstance(
498 const VkInstanceCreateInfo
* pCreateInfo
,
499 const VkAllocationCallbacks
* pAllocator
,
500 VkInstance
* pInstance
)
502 struct anv_instance
*instance
;
505 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
507 /* Check if user passed a debug report callback to be used during
508 * Create/Destroy of instance.
510 const VkDebugReportCallbackCreateInfoEXT
*ctor_cb
=
511 vk_find_struct_const(pCreateInfo
->pNext
,
512 DEBUG_REPORT_CALLBACK_CREATE_INFO_EXT
);
514 uint32_t client_version
;
515 if (pCreateInfo
->pApplicationInfo
&&
516 pCreateInfo
->pApplicationInfo
->apiVersion
!= 0) {
517 client_version
= pCreateInfo
->pApplicationInfo
->apiVersion
;
519 client_version
= VK_MAKE_VERSION(1, 0, 0);
522 if (VK_MAKE_VERSION(1, 0, 0) > client_version
||
523 client_version
> VK_MAKE_VERSION(1, 0, 0xfff)) {
525 if (ctor_cb
&& ctor_cb
->flags
& VK_DEBUG_REPORT_ERROR_BIT_EXT
)
526 ctor_cb
->pfnCallback(VK_DEBUG_REPORT_ERROR_BIT_EXT
,
527 VK_DEBUG_REPORT_OBJECT_TYPE_INSTANCE_EXT
,
528 VK_NULL_HANDLE
, /* No handle available yet. */
532 "incompatible driver version",
535 return vk_errorf(NULL
, NULL
, VK_ERROR_INCOMPATIBLE_DRIVER
,
536 "Client requested version %d.%d.%d",
537 VK_VERSION_MAJOR(client_version
),
538 VK_VERSION_MINOR(client_version
),
539 VK_VERSION_PATCH(client_version
));
542 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
544 for (idx
= 0; idx
< ANV_INSTANCE_EXTENSION_COUNT
; idx
++) {
545 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
546 anv_instance_extensions
[idx
].extensionName
) == 0)
550 if (idx
>= ANV_INSTANCE_EXTENSION_COUNT
)
551 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
553 if (!anv_instance_extensions_supported
.extensions
[idx
])
554 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
557 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
558 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
560 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
562 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
565 instance
->alloc
= *pAllocator
;
567 instance
->alloc
= default_alloc
;
569 instance
->apiVersion
= client_version
;
570 instance
->physicalDeviceCount
= -1;
572 result
= vk_debug_report_instance_init(&instance
->debug_report_callbacks
);
573 if (result
!= VK_SUCCESS
) {
574 vk_free2(&default_alloc
, pAllocator
, instance
);
575 return vk_error(result
);
580 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
582 *pInstance
= anv_instance_to_handle(instance
);
587 void anv_DestroyInstance(
588 VkInstance _instance
,
589 const VkAllocationCallbacks
* pAllocator
)
591 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
596 if (instance
->physicalDeviceCount
> 0) {
597 /* We support at most one physical device. */
598 assert(instance
->physicalDeviceCount
== 1);
599 anv_physical_device_finish(&instance
->physicalDevice
);
602 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
604 vk_debug_report_instance_destroy(&instance
->debug_report_callbacks
);
608 vk_free(&instance
->alloc
, instance
);
612 anv_enumerate_devices(struct anv_instance
*instance
)
614 /* TODO: Check for more devices ? */
615 drmDevicePtr devices
[8];
616 VkResult result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
619 instance
->physicalDeviceCount
= 0;
621 max_devices
= drmGetDevices2(0, devices
, ARRAY_SIZE(devices
));
623 return VK_ERROR_INCOMPATIBLE_DRIVER
;
625 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
626 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
627 devices
[i
]->bustype
== DRM_BUS_PCI
&&
628 devices
[i
]->deviceinfo
.pci
->vendor_id
== 0x8086) {
630 result
= anv_physical_device_init(&instance
->physicalDevice
,
632 devices
[i
]->nodes
[DRM_NODE_RENDER
]);
633 if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
637 drmFreeDevices(devices
, max_devices
);
639 if (result
== VK_SUCCESS
)
640 instance
->physicalDeviceCount
= 1;
646 VkResult
anv_EnumeratePhysicalDevices(
647 VkInstance _instance
,
648 uint32_t* pPhysicalDeviceCount
,
649 VkPhysicalDevice
* pPhysicalDevices
)
651 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
652 VK_OUTARRAY_MAKE(out
, pPhysicalDevices
, pPhysicalDeviceCount
);
655 if (instance
->physicalDeviceCount
< 0) {
656 result
= anv_enumerate_devices(instance
);
657 if (result
!= VK_SUCCESS
&&
658 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
662 if (instance
->physicalDeviceCount
> 0) {
663 assert(instance
->physicalDeviceCount
== 1);
664 vk_outarray_append(&out
, i
) {
665 *i
= anv_physical_device_to_handle(&instance
->physicalDevice
);
669 return vk_outarray_status(&out
);
672 void anv_GetPhysicalDeviceFeatures(
673 VkPhysicalDevice physicalDevice
,
674 VkPhysicalDeviceFeatures
* pFeatures
)
676 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
678 *pFeatures
= (VkPhysicalDeviceFeatures
) {
679 .robustBufferAccess
= true,
680 .fullDrawIndexUint32
= true,
681 .imageCubeArray
= true,
682 .independentBlend
= true,
683 .geometryShader
= true,
684 .tessellationShader
= true,
685 .sampleRateShading
= true,
686 .dualSrcBlend
= true,
688 .multiDrawIndirect
= true,
689 .drawIndirectFirstInstance
= true,
691 .depthBiasClamp
= true,
692 .fillModeNonSolid
= true,
693 .depthBounds
= false,
697 .multiViewport
= true,
698 .samplerAnisotropy
= true,
699 .textureCompressionETC2
= pdevice
->info
.gen
>= 8 ||
700 pdevice
->info
.is_baytrail
,
701 .textureCompressionASTC_LDR
= pdevice
->info
.gen
>= 9, /* FINISHME CHV */
702 .textureCompressionBC
= true,
703 .occlusionQueryPrecise
= true,
704 .pipelineStatisticsQuery
= true,
705 .fragmentStoresAndAtomics
= true,
706 .shaderTessellationAndGeometryPointSize
= true,
707 .shaderImageGatherExtended
= true,
708 .shaderStorageImageExtendedFormats
= true,
709 .shaderStorageImageMultisample
= false,
710 .shaderStorageImageReadWithoutFormat
= false,
711 .shaderStorageImageWriteWithoutFormat
= true,
712 .shaderUniformBufferArrayDynamicIndexing
= true,
713 .shaderSampledImageArrayDynamicIndexing
= true,
714 .shaderStorageBufferArrayDynamicIndexing
= true,
715 .shaderStorageImageArrayDynamicIndexing
= true,
716 .shaderClipDistance
= true,
717 .shaderCullDistance
= true,
718 .shaderFloat64
= pdevice
->info
.gen
>= 8,
719 .shaderInt64
= pdevice
->info
.gen
>= 8,
720 .shaderInt16
= false,
721 .shaderResourceMinLod
= false,
722 .variableMultisampleRate
= false,
723 .inheritedQueries
= true,
726 /* We can't do image stores in vec4 shaders */
727 pFeatures
->vertexPipelineStoresAndAtomics
=
728 pdevice
->compiler
->scalar_stage
[MESA_SHADER_VERTEX
] &&
729 pdevice
->compiler
->scalar_stage
[MESA_SHADER_GEOMETRY
];
732 void anv_GetPhysicalDeviceFeatures2KHR(
733 VkPhysicalDevice physicalDevice
,
734 VkPhysicalDeviceFeatures2KHR
* pFeatures
)
736 anv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
738 vk_foreach_struct(ext
, pFeatures
->pNext
) {
739 switch (ext
->sType
) {
740 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES_KHX
: {
741 VkPhysicalDeviceMultiviewFeaturesKHX
*features
=
742 (VkPhysicalDeviceMultiviewFeaturesKHX
*)ext
;
743 features
->multiview
= true;
744 features
->multiviewGeometryShader
= true;
745 features
->multiviewTessellationShader
= true;
749 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTER_FEATURES_KHR
: {
750 VkPhysicalDeviceVariablePointerFeaturesKHR
*features
= (void *)ext
;
751 features
->variablePointersStorageBuffer
= true;
752 features
->variablePointers
= true;
756 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES_KHR
: {
757 VkPhysicalDeviceSamplerYcbcrConversionFeaturesKHR
*features
=
758 (VkPhysicalDeviceSamplerYcbcrConversionFeaturesKHR
*) ext
;
759 features
->samplerYcbcrConversion
= true;
763 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES_KHR
: {
764 VkPhysicalDevice16BitStorageFeaturesKHR
*features
=
765 (VkPhysicalDevice16BitStorageFeaturesKHR
*)ext
;
767 features
->storageBuffer16BitAccess
= false;
768 features
->uniformAndStorageBuffer16BitAccess
= false;
769 features
->storagePushConstant16
= false;
770 features
->storageInputOutput16
= false;
775 anv_debug_ignored_stype(ext
->sType
);
781 void anv_GetPhysicalDeviceProperties(
782 VkPhysicalDevice physicalDevice
,
783 VkPhysicalDeviceProperties
* pProperties
)
785 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
786 const struct gen_device_info
*devinfo
= &pdevice
->info
;
788 /* See assertions made when programming the buffer surface state. */
789 const uint32_t max_raw_buffer_sz
= devinfo
->gen
>= 7 ?
790 (1ul << 30) : (1ul << 27);
792 const uint32_t max_samplers
= (devinfo
->gen
>= 8 || devinfo
->is_haswell
) ?
795 VkSampleCountFlags sample_counts
=
796 isl_device_get_sample_counts(&pdevice
->isl_dev
);
798 VkPhysicalDeviceLimits limits
= {
799 .maxImageDimension1D
= (1 << 14),
800 .maxImageDimension2D
= (1 << 14),
801 .maxImageDimension3D
= (1 << 11),
802 .maxImageDimensionCube
= (1 << 14),
803 .maxImageArrayLayers
= (1 << 11),
804 .maxTexelBufferElements
= 128 * 1024 * 1024,
805 .maxUniformBufferRange
= (1ul << 27),
806 .maxStorageBufferRange
= max_raw_buffer_sz
,
807 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
808 .maxMemoryAllocationCount
= UINT32_MAX
,
809 .maxSamplerAllocationCount
= 64 * 1024,
810 .bufferImageGranularity
= 64, /* A cache line */
811 .sparseAddressSpaceSize
= 0,
812 .maxBoundDescriptorSets
= MAX_SETS
,
813 .maxPerStageDescriptorSamplers
= max_samplers
,
814 .maxPerStageDescriptorUniformBuffers
= 64,
815 .maxPerStageDescriptorStorageBuffers
= 64,
816 .maxPerStageDescriptorSampledImages
= max_samplers
,
817 .maxPerStageDescriptorStorageImages
= 64,
818 .maxPerStageDescriptorInputAttachments
= 64,
819 .maxPerStageResources
= 250,
820 .maxDescriptorSetSamplers
= 6 * max_samplers
, /* number of stages * maxPerStageDescriptorSamplers */
821 .maxDescriptorSetUniformBuffers
= 6 * 64, /* number of stages * maxPerStageDescriptorUniformBuffers */
822 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
823 .maxDescriptorSetStorageBuffers
= 6 * 64, /* number of stages * maxPerStageDescriptorStorageBuffers */
824 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
825 .maxDescriptorSetSampledImages
= 6 * max_samplers
, /* number of stages * maxPerStageDescriptorSampledImages */
826 .maxDescriptorSetStorageImages
= 6 * 64, /* number of stages * maxPerStageDescriptorStorageImages */
827 .maxDescriptorSetInputAttachments
= 256,
828 .maxVertexInputAttributes
= MAX_VBS
,
829 .maxVertexInputBindings
= MAX_VBS
,
830 .maxVertexInputAttributeOffset
= 2047,
831 .maxVertexInputBindingStride
= 2048,
832 .maxVertexOutputComponents
= 128,
833 .maxTessellationGenerationLevel
= 64,
834 .maxTessellationPatchSize
= 32,
835 .maxTessellationControlPerVertexInputComponents
= 128,
836 .maxTessellationControlPerVertexOutputComponents
= 128,
837 .maxTessellationControlPerPatchOutputComponents
= 128,
838 .maxTessellationControlTotalOutputComponents
= 2048,
839 .maxTessellationEvaluationInputComponents
= 128,
840 .maxTessellationEvaluationOutputComponents
= 128,
841 .maxGeometryShaderInvocations
= 32,
842 .maxGeometryInputComponents
= 64,
843 .maxGeometryOutputComponents
= 128,
844 .maxGeometryOutputVertices
= 256,
845 .maxGeometryTotalOutputComponents
= 1024,
846 .maxFragmentInputComponents
= 128,
847 .maxFragmentOutputAttachments
= 8,
848 .maxFragmentDualSrcAttachments
= 1,
849 .maxFragmentCombinedOutputResources
= 8,
850 .maxComputeSharedMemorySize
= 32768,
851 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
852 .maxComputeWorkGroupInvocations
= 16 * devinfo
->max_cs_threads
,
853 .maxComputeWorkGroupSize
= {
854 16 * devinfo
->max_cs_threads
,
855 16 * devinfo
->max_cs_threads
,
856 16 * devinfo
->max_cs_threads
,
858 .subPixelPrecisionBits
= 4 /* FIXME */,
859 .subTexelPrecisionBits
= 4 /* FIXME */,
860 .mipmapPrecisionBits
= 4 /* FIXME */,
861 .maxDrawIndexedIndexValue
= UINT32_MAX
,
862 .maxDrawIndirectCount
= UINT32_MAX
,
863 .maxSamplerLodBias
= 16,
864 .maxSamplerAnisotropy
= 16,
865 .maxViewports
= MAX_VIEWPORTS
,
866 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
867 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
868 .viewportSubPixelBits
= 13, /* We take a float? */
869 .minMemoryMapAlignment
= 4096, /* A page */
870 .minTexelBufferOffsetAlignment
= 1,
871 /* We need 16 for UBO block reads to work and 32 for push UBOs */
872 .minUniformBufferOffsetAlignment
= 32,
873 .minStorageBufferOffsetAlignment
= 4,
874 .minTexelOffset
= -8,
876 .minTexelGatherOffset
= -32,
877 .maxTexelGatherOffset
= 31,
878 .minInterpolationOffset
= -0.5,
879 .maxInterpolationOffset
= 0.4375,
880 .subPixelInterpolationOffsetBits
= 4,
881 .maxFramebufferWidth
= (1 << 14),
882 .maxFramebufferHeight
= (1 << 14),
883 .maxFramebufferLayers
= (1 << 11),
884 .framebufferColorSampleCounts
= sample_counts
,
885 .framebufferDepthSampleCounts
= sample_counts
,
886 .framebufferStencilSampleCounts
= sample_counts
,
887 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
888 .maxColorAttachments
= MAX_RTS
,
889 .sampledImageColorSampleCounts
= sample_counts
,
890 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
891 .sampledImageDepthSampleCounts
= sample_counts
,
892 .sampledImageStencilSampleCounts
= sample_counts
,
893 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
894 .maxSampleMaskWords
= 1,
895 .timestampComputeAndGraphics
= false,
896 .timestampPeriod
= 1000000000.0 / devinfo
->timestamp_frequency
,
897 .maxClipDistances
= 8,
898 .maxCullDistances
= 8,
899 .maxCombinedClipAndCullDistances
= 8,
900 .discreteQueuePriorities
= 1,
901 .pointSizeRange
= { 0.125, 255.875 },
902 .lineWidthRange
= { 0.0, 7.9921875 },
903 .pointSizeGranularity
= (1.0 / 8.0),
904 .lineWidthGranularity
= (1.0 / 128.0),
905 .strictLines
= false, /* FINISHME */
906 .standardSampleLocations
= true,
907 .optimalBufferCopyOffsetAlignment
= 128,
908 .optimalBufferCopyRowPitchAlignment
= 128,
909 .nonCoherentAtomSize
= 64,
912 *pProperties
= (VkPhysicalDeviceProperties
) {
913 .apiVersion
= anv_physical_device_api_version(pdevice
),
914 .driverVersion
= vk_get_driver_version(),
916 .deviceID
= pdevice
->chipset_id
,
917 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
919 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
922 snprintf(pProperties
->deviceName
, sizeof(pProperties
->deviceName
),
923 "%s", pdevice
->name
);
924 memcpy(pProperties
->pipelineCacheUUID
,
925 pdevice
->pipeline_cache_uuid
, VK_UUID_SIZE
);
928 void anv_GetPhysicalDeviceProperties2KHR(
929 VkPhysicalDevice physicalDevice
,
930 VkPhysicalDeviceProperties2KHR
* pProperties
)
932 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
934 anv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
936 vk_foreach_struct(ext
, pProperties
->pNext
) {
937 switch (ext
->sType
) {
938 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
939 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
940 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
942 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
946 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES_KHR
: {
947 VkPhysicalDeviceIDPropertiesKHR
*id_props
=
948 (VkPhysicalDeviceIDPropertiesKHR
*)ext
;
949 memcpy(id_props
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
950 memcpy(id_props
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
951 /* The LUID is for Windows. */
952 id_props
->deviceLUIDValid
= false;
956 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES_KHX
: {
957 VkPhysicalDeviceMultiviewPropertiesKHX
*properties
=
958 (VkPhysicalDeviceMultiviewPropertiesKHX
*)ext
;
959 properties
->maxMultiviewViewCount
= 16;
960 properties
->maxMultiviewInstanceIndex
= UINT32_MAX
/ 16;
964 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES_KHR
: {
965 VkPhysicalDevicePointClippingPropertiesKHR
*properties
=
966 (VkPhysicalDevicePointClippingPropertiesKHR
*) ext
;
967 properties
->pointClippingBehavior
= VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES_KHR
;
968 anv_finishme("Implement pop-free point clipping");
973 anv_debug_ignored_stype(ext
->sType
);
979 /* We support exactly one queue family. */
980 static const VkQueueFamilyProperties
981 anv_queue_family_properties
= {
982 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
983 VK_QUEUE_COMPUTE_BIT
|
984 VK_QUEUE_TRANSFER_BIT
,
986 .timestampValidBits
= 36, /* XXX: Real value here */
987 .minImageTransferGranularity
= { 1, 1, 1 },
990 void anv_GetPhysicalDeviceQueueFamilyProperties(
991 VkPhysicalDevice physicalDevice
,
993 VkQueueFamilyProperties
* pQueueFamilyProperties
)
995 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pCount
);
997 vk_outarray_append(&out
, p
) {
998 *p
= anv_queue_family_properties
;
1002 void anv_GetPhysicalDeviceQueueFamilyProperties2KHR(
1003 VkPhysicalDevice physicalDevice
,
1004 uint32_t* pQueueFamilyPropertyCount
,
1005 VkQueueFamilyProperties2KHR
* pQueueFamilyProperties
)
1008 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pQueueFamilyPropertyCount
);
1010 vk_outarray_append(&out
, p
) {
1011 p
->queueFamilyProperties
= anv_queue_family_properties
;
1013 vk_foreach_struct(s
, p
->pNext
) {
1014 anv_debug_ignored_stype(s
->sType
);
1019 void anv_GetPhysicalDeviceMemoryProperties(
1020 VkPhysicalDevice physicalDevice
,
1021 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
1023 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
1025 pMemoryProperties
->memoryTypeCount
= physical_device
->memory
.type_count
;
1026 for (uint32_t i
= 0; i
< physical_device
->memory
.type_count
; i
++) {
1027 pMemoryProperties
->memoryTypes
[i
] = (VkMemoryType
) {
1028 .propertyFlags
= physical_device
->memory
.types
[i
].propertyFlags
,
1029 .heapIndex
= physical_device
->memory
.types
[i
].heapIndex
,
1033 pMemoryProperties
->memoryHeapCount
= physical_device
->memory
.heap_count
;
1034 for (uint32_t i
= 0; i
< physical_device
->memory
.heap_count
; i
++) {
1035 pMemoryProperties
->memoryHeaps
[i
] = (VkMemoryHeap
) {
1036 .size
= physical_device
->memory
.heaps
[i
].size
,
1037 .flags
= physical_device
->memory
.heaps
[i
].flags
,
1042 void anv_GetPhysicalDeviceMemoryProperties2KHR(
1043 VkPhysicalDevice physicalDevice
,
1044 VkPhysicalDeviceMemoryProperties2KHR
* pMemoryProperties
)
1046 anv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
1047 &pMemoryProperties
->memoryProperties
);
1049 vk_foreach_struct(ext
, pMemoryProperties
->pNext
) {
1050 switch (ext
->sType
) {
1052 anv_debug_ignored_stype(ext
->sType
);
1058 PFN_vkVoidFunction
anv_GetInstanceProcAddr(
1059 VkInstance instance
,
1062 return anv_lookup_entrypoint(NULL
, pName
);
1065 /* With version 1+ of the loader interface the ICD should expose
1066 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
1069 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1070 VkInstance instance
,
1074 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1075 VkInstance instance
,
1078 return anv_GetInstanceProcAddr(instance
, pName
);
1081 PFN_vkVoidFunction
anv_GetDeviceProcAddr(
1085 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1086 return anv_lookup_entrypoint(&device
->info
, pName
);
1090 anv_CreateDebugReportCallbackEXT(VkInstance _instance
,
1091 const VkDebugReportCallbackCreateInfoEXT
* pCreateInfo
,
1092 const VkAllocationCallbacks
* pAllocator
,
1093 VkDebugReportCallbackEXT
* pCallback
)
1095 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1096 return vk_create_debug_report_callback(&instance
->debug_report_callbacks
,
1097 pCreateInfo
, pAllocator
, &instance
->alloc
,
1102 anv_DestroyDebugReportCallbackEXT(VkInstance _instance
,
1103 VkDebugReportCallbackEXT _callback
,
1104 const VkAllocationCallbacks
* pAllocator
)
1106 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1107 vk_destroy_debug_report_callback(&instance
->debug_report_callbacks
,
1108 _callback
, pAllocator
, &instance
->alloc
);
1112 anv_DebugReportMessageEXT(VkInstance _instance
,
1113 VkDebugReportFlagsEXT flags
,
1114 VkDebugReportObjectTypeEXT objectType
,
1117 int32_t messageCode
,
1118 const char* pLayerPrefix
,
1119 const char* pMessage
)
1121 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1122 vk_debug_report(&instance
->debug_report_callbacks
, flags
, objectType
,
1123 object
, location
, messageCode
, pLayerPrefix
, pMessage
);
1127 anv_queue_init(struct anv_device
*device
, struct anv_queue
*queue
)
1129 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1130 queue
->device
= device
;
1131 queue
->pool
= &device
->surface_state_pool
;
1135 anv_queue_finish(struct anv_queue
*queue
)
1139 static struct anv_state
1140 anv_state_pool_emit_data(struct anv_state_pool
*pool
, size_t size
, size_t align
, const void *p
)
1142 struct anv_state state
;
1144 state
= anv_state_pool_alloc(pool
, size
, align
);
1145 memcpy(state
.map
, p
, size
);
1147 anv_state_flush(pool
->block_pool
.device
, state
);
1152 struct gen8_border_color
{
1157 /* Pad out to 64 bytes */
1162 anv_device_init_border_colors(struct anv_device
*device
)
1164 static const struct gen8_border_color border_colors
[] = {
1165 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
1166 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
1167 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
1168 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
1169 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
1170 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
1173 device
->border_colors
= anv_state_pool_emit_data(&device
->dynamic_state_pool
,
1174 sizeof(border_colors
), 64,
1179 anv_device_init_trivial_batch(struct anv_device
*device
)
1181 anv_bo_init_new(&device
->trivial_batch_bo
, device
, 4096);
1183 if (device
->instance
->physicalDevice
.has_exec_async
)
1184 device
->trivial_batch_bo
.flags
|= EXEC_OBJECT_ASYNC
;
1186 void *map
= anv_gem_mmap(device
, device
->trivial_batch_bo
.gem_handle
,
1189 struct anv_batch batch
= {
1195 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
1196 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
1198 if (!device
->info
.has_llc
)
1199 gen_clflush_range(map
, batch
.next
- map
);
1201 anv_gem_munmap(map
, device
->trivial_batch_bo
.size
);
1204 VkResult
anv_CreateDevice(
1205 VkPhysicalDevice physicalDevice
,
1206 const VkDeviceCreateInfo
* pCreateInfo
,
1207 const VkAllocationCallbacks
* pAllocator
,
1210 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
1212 struct anv_device
*device
;
1214 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO
);
1216 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
1217 const char *ext_name
= pCreateInfo
->ppEnabledExtensionNames
[i
];
1218 if (!anv_physical_device_extension_supported(physical_device
, ext_name
))
1219 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
1222 /* Check enabled features */
1223 if (pCreateInfo
->pEnabledFeatures
) {
1224 VkPhysicalDeviceFeatures supported_features
;
1225 anv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
1226 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
1227 VkBool32
*enabled_feature
= (VkBool32
*)pCreateInfo
->pEnabledFeatures
;
1228 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
1229 for (uint32_t i
= 0; i
< num_features
; i
++) {
1230 if (enabled_feature
[i
] && !supported_feature
[i
])
1231 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
1235 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
1237 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
1239 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1241 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1242 device
->instance
= physical_device
->instance
;
1243 device
->chipset_id
= physical_device
->chipset_id
;
1244 device
->lost
= false;
1247 device
->alloc
= *pAllocator
;
1249 device
->alloc
= physical_device
->instance
->alloc
;
1251 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
1252 device
->fd
= open(physical_device
->path
, O_RDWR
| O_CLOEXEC
);
1253 if (device
->fd
== -1) {
1254 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1258 device
->context_id
= anv_gem_create_context(device
);
1259 if (device
->context_id
== -1) {
1260 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1264 device
->info
= physical_device
->info
;
1265 device
->isl_dev
= physical_device
->isl_dev
;
1267 /* On Broadwell and later, we can use batch chaining to more efficiently
1268 * implement growing command buffers. Prior to Haswell, the kernel
1269 * command parser gets in the way and we have to fall back to growing
1272 device
->can_chain_batches
= device
->info
.gen
>= 8;
1274 device
->robust_buffer_access
= pCreateInfo
->pEnabledFeatures
&&
1275 pCreateInfo
->pEnabledFeatures
->robustBufferAccess
;
1277 if (pthread_mutex_init(&device
->mutex
, NULL
) != 0) {
1278 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1279 goto fail_context_id
;
1282 pthread_condattr_t condattr
;
1283 if (pthread_condattr_init(&condattr
) != 0) {
1284 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1287 if (pthread_condattr_setclock(&condattr
, CLOCK_MONOTONIC
) != 0) {
1288 pthread_condattr_destroy(&condattr
);
1289 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1292 if (pthread_cond_init(&device
->queue_submit
, NULL
) != 0) {
1293 pthread_condattr_destroy(&condattr
);
1294 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1297 pthread_condattr_destroy(&condattr
);
1300 (physical_device
->supports_48bit_addresses
? EXEC_OBJECT_SUPPORTS_48B_ADDRESS
: 0) |
1301 (physical_device
->has_exec_async
? EXEC_OBJECT_ASYNC
: 0) |
1302 (physical_device
->has_exec_capture
? EXEC_OBJECT_CAPTURE
: 0);
1304 anv_bo_pool_init(&device
->batch_bo_pool
, device
, bo_flags
);
1306 result
= anv_bo_cache_init(&device
->bo_cache
);
1307 if (result
!= VK_SUCCESS
)
1308 goto fail_batch_bo_pool
;
1310 /* For the state pools we explicitly disable 48bit. */
1311 bo_flags
= (physical_device
->has_exec_async
? EXEC_OBJECT_ASYNC
: 0) |
1312 (physical_device
->has_exec_capture
? EXEC_OBJECT_CAPTURE
: 0);
1314 result
= anv_state_pool_init(&device
->dynamic_state_pool
, device
, 16384,
1316 if (result
!= VK_SUCCESS
)
1319 result
= anv_state_pool_init(&device
->instruction_state_pool
, device
, 16384,
1321 if (result
!= VK_SUCCESS
)
1322 goto fail_dynamic_state_pool
;
1324 result
= anv_state_pool_init(&device
->surface_state_pool
, device
, 4096,
1326 if (result
!= VK_SUCCESS
)
1327 goto fail_instruction_state_pool
;
1329 result
= anv_bo_init_new(&device
->workaround_bo
, device
, 1024);
1330 if (result
!= VK_SUCCESS
)
1331 goto fail_surface_state_pool
;
1333 anv_device_init_trivial_batch(device
);
1335 anv_scratch_pool_init(device
, &device
->scratch_pool
);
1337 anv_queue_init(device
, &device
->queue
);
1339 switch (device
->info
.gen
) {
1341 if (!device
->info
.is_haswell
)
1342 result
= gen7_init_device_state(device
);
1344 result
= gen75_init_device_state(device
);
1347 result
= gen8_init_device_state(device
);
1350 result
= gen9_init_device_state(device
);
1353 result
= gen10_init_device_state(device
);
1356 /* Shouldn't get here as we don't create physical devices for any other
1358 unreachable("unhandled gen");
1360 if (result
!= VK_SUCCESS
)
1361 goto fail_workaround_bo
;
1363 anv_device_init_blorp(device
);
1365 anv_device_init_border_colors(device
);
1367 *pDevice
= anv_device_to_handle(device
);
1372 anv_queue_finish(&device
->queue
);
1373 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
1374 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
1375 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
1376 fail_surface_state_pool
:
1377 anv_state_pool_finish(&device
->surface_state_pool
);
1378 fail_instruction_state_pool
:
1379 anv_state_pool_finish(&device
->instruction_state_pool
);
1380 fail_dynamic_state_pool
:
1381 anv_state_pool_finish(&device
->dynamic_state_pool
);
1383 anv_bo_cache_finish(&device
->bo_cache
);
1385 anv_bo_pool_finish(&device
->batch_bo_pool
);
1386 pthread_cond_destroy(&device
->queue_submit
);
1388 pthread_mutex_destroy(&device
->mutex
);
1390 anv_gem_destroy_context(device
, device
->context_id
);
1394 vk_free(&device
->alloc
, device
);
1399 void anv_DestroyDevice(
1401 const VkAllocationCallbacks
* pAllocator
)
1403 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1408 anv_device_finish_blorp(device
);
1410 anv_queue_finish(&device
->queue
);
1412 #ifdef HAVE_VALGRIND
1413 /* We only need to free these to prevent valgrind errors. The backing
1414 * BO will go away in a couple of lines so we don't actually leak.
1416 anv_state_pool_free(&device
->dynamic_state_pool
, device
->border_colors
);
1419 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
1421 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
1422 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
1424 anv_gem_close(device
, device
->trivial_batch_bo
.gem_handle
);
1426 anv_state_pool_finish(&device
->surface_state_pool
);
1427 anv_state_pool_finish(&device
->instruction_state_pool
);
1428 anv_state_pool_finish(&device
->dynamic_state_pool
);
1430 anv_bo_cache_finish(&device
->bo_cache
);
1432 anv_bo_pool_finish(&device
->batch_bo_pool
);
1434 pthread_cond_destroy(&device
->queue_submit
);
1435 pthread_mutex_destroy(&device
->mutex
);
1437 anv_gem_destroy_context(device
, device
->context_id
);
1441 vk_free(&device
->alloc
, device
);
1444 VkResult
anv_EnumerateInstanceLayerProperties(
1445 uint32_t* pPropertyCount
,
1446 VkLayerProperties
* pProperties
)
1448 if (pProperties
== NULL
) {
1449 *pPropertyCount
= 0;
1453 /* None supported at this time */
1454 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1457 VkResult
anv_EnumerateDeviceLayerProperties(
1458 VkPhysicalDevice physicalDevice
,
1459 uint32_t* pPropertyCount
,
1460 VkLayerProperties
* pProperties
)
1462 if (pProperties
== NULL
) {
1463 *pPropertyCount
= 0;
1467 /* None supported at this time */
1468 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1471 void anv_GetDeviceQueue(
1473 uint32_t queueNodeIndex
,
1474 uint32_t queueIndex
,
1477 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1479 assert(queueIndex
== 0);
1481 *pQueue
= anv_queue_to_handle(&device
->queue
);
1485 anv_device_query_status(struct anv_device
*device
)
1487 /* This isn't likely as most of the callers of this function already check
1488 * for it. However, it doesn't hurt to check and it potentially lets us
1491 if (unlikely(device
->lost
))
1492 return VK_ERROR_DEVICE_LOST
;
1494 uint32_t active
, pending
;
1495 int ret
= anv_gem_gpu_get_reset_stats(device
, &active
, &pending
);
1497 /* We don't know the real error. */
1498 device
->lost
= true;
1499 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1500 "get_reset_stats failed: %m");
1504 device
->lost
= true;
1505 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1506 "GPU hung on one of our command buffers");
1507 } else if (pending
) {
1508 device
->lost
= true;
1509 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1510 "GPU hung with commands in-flight");
1517 anv_device_bo_busy(struct anv_device
*device
, struct anv_bo
*bo
)
1519 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
1520 * Other usages of the BO (such as on different hardware) will not be
1521 * flagged as "busy" by this ioctl. Use with care.
1523 int ret
= anv_gem_busy(device
, bo
->gem_handle
);
1525 return VK_NOT_READY
;
1526 } else if (ret
== -1) {
1527 /* We don't know the real error. */
1528 device
->lost
= true;
1529 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1530 "gem wait failed: %m");
1533 /* Query for device status after the busy call. If the BO we're checking
1534 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
1535 * client because it clearly doesn't have valid data. Yes, this most
1536 * likely means an ioctl, but we just did an ioctl to query the busy status
1537 * so it's no great loss.
1539 return anv_device_query_status(device
);
1543 anv_device_wait(struct anv_device
*device
, struct anv_bo
*bo
,
1546 int ret
= anv_gem_wait(device
, bo
->gem_handle
, &timeout
);
1547 if (ret
== -1 && errno
== ETIME
) {
1549 } else if (ret
== -1) {
1550 /* We don't know the real error. */
1551 device
->lost
= true;
1552 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1553 "gem wait failed: %m");
1556 /* Query for device status after the wait. If the BO we're waiting on got
1557 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
1558 * because it clearly doesn't have valid data. Yes, this most likely means
1559 * an ioctl, but we just did an ioctl to wait so it's no great loss.
1561 return anv_device_query_status(device
);
1564 VkResult
anv_DeviceWaitIdle(
1567 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1568 if (unlikely(device
->lost
))
1569 return VK_ERROR_DEVICE_LOST
;
1571 struct anv_batch batch
;
1574 batch
.start
= batch
.next
= cmds
;
1575 batch
.end
= (void *) cmds
+ sizeof(cmds
);
1577 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
1578 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
1580 return anv_device_submit_simple_batch(device
, &batch
);
1584 anv_bo_init_new(struct anv_bo
*bo
, struct anv_device
*device
, uint64_t size
)
1586 uint32_t gem_handle
= anv_gem_create(device
, size
);
1588 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
1590 anv_bo_init(bo
, gem_handle
, size
);
1595 VkResult
anv_AllocateMemory(
1597 const VkMemoryAllocateInfo
* pAllocateInfo
,
1598 const VkAllocationCallbacks
* pAllocator
,
1599 VkDeviceMemory
* pMem
)
1601 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1602 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
1603 struct anv_device_memory
*mem
;
1604 VkResult result
= VK_SUCCESS
;
1606 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
1608 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
1609 assert(pAllocateInfo
->allocationSize
> 0);
1611 /* The kernel relocation API has a limitation of a 32-bit delta value
1612 * applied to the address before it is written which, in spite of it being
1613 * unsigned, is treated as signed . Because of the way that this maps to
1614 * the Vulkan API, we cannot handle an offset into a buffer that does not
1615 * fit into a signed 32 bits. The only mechanism we have for dealing with
1616 * this at the moment is to limit all VkDeviceMemory objects to a maximum
1617 * of 2GB each. The Vulkan spec allows us to do this:
1619 * "Some platforms may have a limit on the maximum size of a single
1620 * allocation. For example, certain systems may fail to create
1621 * allocations with a size greater than or equal to 4GB. Such a limit is
1622 * implementation-dependent, and if such a failure occurs then the error
1623 * VK_ERROR_OUT_OF_DEVICE_MEMORY should be returned."
1625 * We don't use vk_error here because it's not an error so much as an
1626 * indication to the application that the allocation is too large.
1628 if (pAllocateInfo
->allocationSize
> (1ull << 31))
1629 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
1631 /* FINISHME: Fail if allocation request exceeds heap size. */
1633 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
1634 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1636 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1638 assert(pAllocateInfo
->memoryTypeIndex
< pdevice
->memory
.type_count
);
1639 mem
->type
= &pdevice
->memory
.types
[pAllocateInfo
->memoryTypeIndex
];
1643 const VkImportMemoryFdInfoKHR
*fd_info
=
1644 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_FD_INFO_KHR
);
1646 /* The Vulkan spec permits handleType to be 0, in which case the struct is
1649 if (fd_info
&& fd_info
->handleType
) {
1650 /* At the moment, we support only the below handle types. */
1651 assert(fd_info
->handleType
==
1652 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
||
1653 fd_info
->handleType
==
1654 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
1656 result
= anv_bo_cache_import(device
, &device
->bo_cache
,
1657 fd_info
->fd
, &mem
->bo
);
1658 if (result
!= VK_SUCCESS
)
1661 VkDeviceSize aligned_alloc_size
=
1662 align_u64(pAllocateInfo
->allocationSize
, 4096);
1664 /* For security purposes, we reject importing the bo if it's smaller
1665 * than the requested allocation size. This prevents a malicious client
1666 * from passing a buffer to a trusted client, lying about the size, and
1667 * telling the trusted client to try and texture from an image that goes
1668 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
1669 * in the trusted client. The trusted client can protect itself against
1670 * this sort of attack but only if it can trust the buffer size.
1672 if (mem
->bo
->size
< aligned_alloc_size
) {
1673 result
= vk_errorf(device
->instance
, device
,
1674 VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
,
1675 "aligned allocationSize too large for "
1676 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR: "
1677 "%"PRIu64
"B > %"PRIu64
"B",
1678 aligned_alloc_size
, mem
->bo
->size
);
1679 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
1683 /* From the Vulkan spec:
1685 * "Importing memory from a file descriptor transfers ownership of
1686 * the file descriptor from the application to the Vulkan
1687 * implementation. The application must not perform any operations on
1688 * the file descriptor after a successful import."
1690 * If the import fails, we leave the file descriptor open.
1694 result
= anv_bo_cache_alloc(device
, &device
->bo_cache
,
1695 pAllocateInfo
->allocationSize
,
1697 if (result
!= VK_SUCCESS
)
1700 const VkMemoryDedicatedAllocateInfoKHR
*dedicated_info
=
1701 vk_find_struct_const(pAllocateInfo
->pNext
, MEMORY_DEDICATED_ALLOCATE_INFO_KHR
);
1702 if (dedicated_info
&& dedicated_info
->image
!= VK_NULL_HANDLE
) {
1703 ANV_FROM_HANDLE(anv_image
, image
, dedicated_info
->image
);
1705 /* For images using modifiers, we require a dedicated allocation
1706 * and we set the BO tiling to match the tiling of the underlying
1707 * modifier. This is a bit unfortunate as this is completely
1708 * pointless for Vulkan. However, GL needs to be able to map things
1709 * so it needs the tiling to be set. The only way to do this in a
1710 * non-racy way is to set the tiling in the creator of the BO so that
1713 * One of these days, once the GL driver learns to not map things
1714 * through the GTT in random places, we can drop this and start
1715 * allowing multiple modified images in the same BO.
1717 if (image
->drm_format_mod
!= DRM_FORMAT_MOD_INVALID
) {
1718 assert(isl_drm_modifier_get_info(image
->drm_format_mod
)->tiling
==
1719 image
->planes
[0].surface
.isl
.tiling
);
1720 const uint32_t i915_tiling
=
1721 isl_tiling_to_i915_tiling(image
->planes
[0].surface
.isl
.tiling
);
1722 int ret
= anv_gem_set_tiling(device
, mem
->bo
->gem_handle
,
1723 image
->planes
[0].surface
.isl
.row_pitch
,
1726 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
1727 return vk_errorf(device
->instance
, NULL
,
1728 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
1729 "failed to set BO tiling: %m");
1735 assert(mem
->type
->heapIndex
< pdevice
->memory
.heap_count
);
1736 if (pdevice
->memory
.heaps
[mem
->type
->heapIndex
].supports_48bit_addresses
)
1737 mem
->bo
->flags
|= EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
1739 const struct wsi_memory_allocate_info
*wsi_info
=
1740 vk_find_struct_const(pAllocateInfo
->pNext
, WSI_MEMORY_ALLOCATE_INFO_MESA
);
1741 if (wsi_info
&& wsi_info
->implicit_sync
) {
1742 /* We need to set the WRITE flag on window system buffers so that GEM
1743 * will know we're writing to them and synchronize uses on other rings
1744 * (eg if the display server uses the blitter ring).
1746 mem
->bo
->flags
|= EXEC_OBJECT_WRITE
;
1747 } else if (pdevice
->has_exec_async
) {
1748 mem
->bo
->flags
|= EXEC_OBJECT_ASYNC
;
1751 *pMem
= anv_device_memory_to_handle(mem
);
1756 vk_free2(&device
->alloc
, pAllocator
, mem
);
1761 VkResult
anv_GetMemoryFdKHR(
1763 const VkMemoryGetFdInfoKHR
* pGetFdInfo
,
1766 ANV_FROM_HANDLE(anv_device
, dev
, device_h
);
1767 ANV_FROM_HANDLE(anv_device_memory
, mem
, pGetFdInfo
->memory
);
1769 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
1771 assert(pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
||
1772 pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
1774 return anv_bo_cache_export(dev
, &dev
->bo_cache
, mem
->bo
, pFd
);
1777 VkResult
anv_GetMemoryFdPropertiesKHR(
1779 VkExternalMemoryHandleTypeFlagBitsKHR handleType
,
1781 VkMemoryFdPropertiesKHR
* pMemoryFdProperties
)
1783 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1784 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
1786 switch (handleType
) {
1787 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
:
1788 /* dma-buf can be imported as any memory type */
1789 pMemoryFdProperties
->memoryTypeBits
=
1790 (1 << pdevice
->memory
.type_count
) - 1;
1794 /* The valid usage section for this function says:
1796 * "handleType must not be one of the handle types defined as
1799 * So opaque handle types fall into the default "unsupported" case.
1801 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
1805 void anv_FreeMemory(
1807 VkDeviceMemory _mem
,
1808 const VkAllocationCallbacks
* pAllocator
)
1810 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1811 ANV_FROM_HANDLE(anv_device_memory
, mem
, _mem
);
1817 anv_UnmapMemory(_device
, _mem
);
1819 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
1821 vk_free2(&device
->alloc
, pAllocator
, mem
);
1824 VkResult
anv_MapMemory(
1826 VkDeviceMemory _memory
,
1827 VkDeviceSize offset
,
1829 VkMemoryMapFlags flags
,
1832 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1833 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
1840 if (size
== VK_WHOLE_SIZE
)
1841 size
= mem
->bo
->size
- offset
;
1843 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
1845 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
1846 * assert(size != 0);
1847 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
1848 * equal to the size of the memory minus offset
1851 assert(offset
+ size
<= mem
->bo
->size
);
1853 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
1854 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
1855 * at a time is valid. We could just mmap up front and return an offset
1856 * pointer here, but that may exhaust virtual memory on 32 bit
1859 uint32_t gem_flags
= 0;
1861 if (!device
->info
.has_llc
&&
1862 (mem
->type
->propertyFlags
& VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
))
1863 gem_flags
|= I915_MMAP_WC
;
1865 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
1866 uint64_t map_offset
= offset
& ~4095ull;
1867 assert(offset
>= map_offset
);
1868 uint64_t map_size
= (offset
+ size
) - map_offset
;
1870 /* Let's map whole pages */
1871 map_size
= align_u64(map_size
, 4096);
1873 void *map
= anv_gem_mmap(device
, mem
->bo
->gem_handle
,
1874 map_offset
, map_size
, gem_flags
);
1875 if (map
== MAP_FAILED
)
1876 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
1879 mem
->map_size
= map_size
;
1881 *ppData
= mem
->map
+ (offset
- map_offset
);
1886 void anv_UnmapMemory(
1888 VkDeviceMemory _memory
)
1890 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
1895 anv_gem_munmap(mem
->map
, mem
->map_size
);
1902 clflush_mapped_ranges(struct anv_device
*device
,
1904 const VkMappedMemoryRange
*ranges
)
1906 for (uint32_t i
= 0; i
< count
; i
++) {
1907 ANV_FROM_HANDLE(anv_device_memory
, mem
, ranges
[i
].memory
);
1908 if (ranges
[i
].offset
>= mem
->map_size
)
1911 gen_clflush_range(mem
->map
+ ranges
[i
].offset
,
1912 MIN2(ranges
[i
].size
, mem
->map_size
- ranges
[i
].offset
));
1916 VkResult
anv_FlushMappedMemoryRanges(
1918 uint32_t memoryRangeCount
,
1919 const VkMappedMemoryRange
* pMemoryRanges
)
1921 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1923 if (device
->info
.has_llc
)
1926 /* Make sure the writes we're flushing have landed. */
1927 __builtin_ia32_mfence();
1929 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
1934 VkResult
anv_InvalidateMappedMemoryRanges(
1936 uint32_t memoryRangeCount
,
1937 const VkMappedMemoryRange
* pMemoryRanges
)
1939 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1941 if (device
->info
.has_llc
)
1944 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
1946 /* Make sure no reads get moved up above the invalidate. */
1947 __builtin_ia32_mfence();
1952 void anv_GetBufferMemoryRequirements(
1955 VkMemoryRequirements
* pMemoryRequirements
)
1957 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
1958 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1959 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
1961 /* The Vulkan spec (git aaed022) says:
1963 * memoryTypeBits is a bitfield and contains one bit set for every
1964 * supported memory type for the resource. The bit `1<<i` is set if and
1965 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1966 * structure for the physical device is supported.
1968 uint32_t memory_types
= 0;
1969 for (uint32_t i
= 0; i
< pdevice
->memory
.type_count
; i
++) {
1970 uint32_t valid_usage
= pdevice
->memory
.types
[i
].valid_buffer_usage
;
1971 if ((valid_usage
& buffer
->usage
) == buffer
->usage
)
1972 memory_types
|= (1u << i
);
1975 /* Base alignment requirement of a cache line */
1976 uint32_t alignment
= 16;
1978 /* We need an alignment of 32 for pushing UBOs */
1979 if (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
)
1980 alignment
= MAX2(alignment
, 32);
1982 pMemoryRequirements
->size
= buffer
->size
;
1983 pMemoryRequirements
->alignment
= alignment
;
1984 pMemoryRequirements
->memoryTypeBits
= memory_types
;
1987 void anv_GetBufferMemoryRequirements2KHR(
1989 const VkBufferMemoryRequirementsInfo2KHR
* pInfo
,
1990 VkMemoryRequirements2KHR
* pMemoryRequirements
)
1992 anv_GetBufferMemoryRequirements(_device
, pInfo
->buffer
,
1993 &pMemoryRequirements
->memoryRequirements
);
1995 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
1996 switch (ext
->sType
) {
1997 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS_KHR
: {
1998 VkMemoryDedicatedRequirementsKHR
*requirements
= (void *)ext
;
1999 requirements
->prefersDedicatedAllocation
= VK_FALSE
;
2000 requirements
->requiresDedicatedAllocation
= VK_FALSE
;
2005 anv_debug_ignored_stype(ext
->sType
);
2011 void anv_GetImageMemoryRequirements(
2014 VkMemoryRequirements
* pMemoryRequirements
)
2016 ANV_FROM_HANDLE(anv_image
, image
, _image
);
2017 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2018 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2020 /* The Vulkan spec (git aaed022) says:
2022 * memoryTypeBits is a bitfield and contains one bit set for every
2023 * supported memory type for the resource. The bit `1<<i` is set if and
2024 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
2025 * structure for the physical device is supported.
2027 * All types are currently supported for images.
2029 uint32_t memory_types
= (1ull << pdevice
->memory
.type_count
) - 1;
2031 pMemoryRequirements
->size
= image
->size
;
2032 pMemoryRequirements
->alignment
= image
->alignment
;
2033 pMemoryRequirements
->memoryTypeBits
= memory_types
;
2036 void anv_GetImageMemoryRequirements2KHR(
2038 const VkImageMemoryRequirementsInfo2KHR
* pInfo
,
2039 VkMemoryRequirements2KHR
* pMemoryRequirements
)
2041 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2042 ANV_FROM_HANDLE(anv_image
, image
, pInfo
->image
);
2044 anv_GetImageMemoryRequirements(_device
, pInfo
->image
,
2045 &pMemoryRequirements
->memoryRequirements
);
2047 vk_foreach_struct_const(ext
, pInfo
->pNext
) {
2048 switch (ext
->sType
) {
2049 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO_KHR
: {
2050 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2051 const VkImagePlaneMemoryRequirementsInfoKHR
*plane_reqs
=
2052 (const VkImagePlaneMemoryRequirementsInfoKHR
*) ext
;
2053 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
,
2054 plane_reqs
->planeAspect
);
2056 assert(image
->planes
[plane
].offset
== 0);
2058 /* The Vulkan spec (git aaed022) says:
2060 * memoryTypeBits is a bitfield and contains one bit set for every
2061 * supported memory type for the resource. The bit `1<<i` is set
2062 * if and only if the memory type `i` in the
2063 * VkPhysicalDeviceMemoryProperties structure for the physical
2064 * device is supported.
2066 * All types are currently supported for images.
2068 pMemoryRequirements
->memoryRequirements
.memoryTypeBits
=
2069 (1ull << pdevice
->memory
.type_count
) - 1;
2071 pMemoryRequirements
->memoryRequirements
.size
= image
->planes
[plane
].size
;
2072 pMemoryRequirements
->memoryRequirements
.alignment
=
2073 image
->planes
[plane
].alignment
;
2078 anv_debug_ignored_stype(ext
->sType
);
2083 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
2084 switch (ext
->sType
) {
2085 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS_KHR
: {
2086 VkMemoryDedicatedRequirementsKHR
*requirements
= (void *)ext
;
2087 if (image
->drm_format_mod
!= DRM_FORMAT_MOD_INVALID
) {
2088 /* Require a dedicated allocation for images with modifiers.
2090 * See also anv_AllocateMemory.
2092 requirements
->prefersDedicatedAllocation
= VK_TRUE
;
2093 requirements
->requiresDedicatedAllocation
= VK_TRUE
;
2095 requirements
->prefersDedicatedAllocation
= VK_FALSE
;
2096 requirements
->requiresDedicatedAllocation
= VK_FALSE
;
2102 anv_debug_ignored_stype(ext
->sType
);
2108 void anv_GetImageSparseMemoryRequirements(
2111 uint32_t* pSparseMemoryRequirementCount
,
2112 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
2114 *pSparseMemoryRequirementCount
= 0;
2117 void anv_GetImageSparseMemoryRequirements2KHR(
2119 const VkImageSparseMemoryRequirementsInfo2KHR
* pInfo
,
2120 uint32_t* pSparseMemoryRequirementCount
,
2121 VkSparseImageMemoryRequirements2KHR
* pSparseMemoryRequirements
)
2123 *pSparseMemoryRequirementCount
= 0;
2126 void anv_GetDeviceMemoryCommitment(
2128 VkDeviceMemory memory
,
2129 VkDeviceSize
* pCommittedMemoryInBytes
)
2131 *pCommittedMemoryInBytes
= 0;
2135 anv_bind_buffer_memory(const VkBindBufferMemoryInfoKHR
*pBindInfo
)
2137 ANV_FROM_HANDLE(anv_device_memory
, mem
, pBindInfo
->memory
);
2138 ANV_FROM_HANDLE(anv_buffer
, buffer
, pBindInfo
->buffer
);
2140 assert(pBindInfo
->sType
== VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO_KHR
);
2143 assert((buffer
->usage
& mem
->type
->valid_buffer_usage
) == buffer
->usage
);
2144 buffer
->bo
= mem
->bo
;
2145 buffer
->offset
= pBindInfo
->memoryOffset
;
2152 VkResult
anv_BindBufferMemory(
2155 VkDeviceMemory memory
,
2156 VkDeviceSize memoryOffset
)
2158 anv_bind_buffer_memory(
2159 &(VkBindBufferMemoryInfoKHR
) {
2160 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO_KHR
,
2163 .memoryOffset
= memoryOffset
,
2169 VkResult
anv_BindBufferMemory2KHR(
2171 uint32_t bindInfoCount
,
2172 const VkBindBufferMemoryInfoKHR
* pBindInfos
)
2174 for (uint32_t i
= 0; i
< bindInfoCount
; i
++)
2175 anv_bind_buffer_memory(&pBindInfos
[i
]);
2180 VkResult
anv_QueueBindSparse(
2182 uint32_t bindInfoCount
,
2183 const VkBindSparseInfo
* pBindInfo
,
2186 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
2187 if (unlikely(queue
->device
->lost
))
2188 return VK_ERROR_DEVICE_LOST
;
2190 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
2195 VkResult
anv_CreateEvent(
2197 const VkEventCreateInfo
* pCreateInfo
,
2198 const VkAllocationCallbacks
* pAllocator
,
2201 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2202 struct anv_state state
;
2203 struct anv_event
*event
;
2205 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_EVENT_CREATE_INFO
);
2207 state
= anv_state_pool_alloc(&device
->dynamic_state_pool
,
2210 event
->state
= state
;
2211 event
->semaphore
= VK_EVENT_RESET
;
2213 if (!device
->info
.has_llc
) {
2214 /* Make sure the writes we're flushing have landed. */
2215 __builtin_ia32_mfence();
2216 __builtin_ia32_clflush(event
);
2219 *pEvent
= anv_event_to_handle(event
);
2224 void anv_DestroyEvent(
2227 const VkAllocationCallbacks
* pAllocator
)
2229 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2230 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2235 anv_state_pool_free(&device
->dynamic_state_pool
, event
->state
);
2238 VkResult
anv_GetEventStatus(
2242 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2243 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2245 if (unlikely(device
->lost
))
2246 return VK_ERROR_DEVICE_LOST
;
2248 if (!device
->info
.has_llc
) {
2249 /* Invalidate read cache before reading event written by GPU. */
2250 __builtin_ia32_clflush(event
);
2251 __builtin_ia32_mfence();
2255 return event
->semaphore
;
2258 VkResult
anv_SetEvent(
2262 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2263 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2265 event
->semaphore
= VK_EVENT_SET
;
2267 if (!device
->info
.has_llc
) {
2268 /* Make sure the writes we're flushing have landed. */
2269 __builtin_ia32_mfence();
2270 __builtin_ia32_clflush(event
);
2276 VkResult
anv_ResetEvent(
2280 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2281 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2283 event
->semaphore
= VK_EVENT_RESET
;
2285 if (!device
->info
.has_llc
) {
2286 /* Make sure the writes we're flushing have landed. */
2287 __builtin_ia32_mfence();
2288 __builtin_ia32_clflush(event
);
2296 VkResult
anv_CreateBuffer(
2298 const VkBufferCreateInfo
* pCreateInfo
,
2299 const VkAllocationCallbacks
* pAllocator
,
2302 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2303 struct anv_buffer
*buffer
;
2305 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
2307 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
2308 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2310 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2312 buffer
->size
= pCreateInfo
->size
;
2313 buffer
->usage
= pCreateInfo
->usage
;
2317 *pBuffer
= anv_buffer_to_handle(buffer
);
2322 void anv_DestroyBuffer(
2325 const VkAllocationCallbacks
* pAllocator
)
2327 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2328 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
2333 vk_free2(&device
->alloc
, pAllocator
, buffer
);
2337 anv_fill_buffer_surface_state(struct anv_device
*device
, struct anv_state state
,
2338 enum isl_format format
,
2339 uint32_t offset
, uint32_t range
, uint32_t stride
)
2341 isl_buffer_fill_state(&device
->isl_dev
, state
.map
,
2343 .mocs
= device
->default_mocs
,
2348 anv_state_flush(device
, state
);
2351 void anv_DestroySampler(
2354 const VkAllocationCallbacks
* pAllocator
)
2356 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2357 ANV_FROM_HANDLE(anv_sampler
, sampler
, _sampler
);
2362 vk_free2(&device
->alloc
, pAllocator
, sampler
);
2365 VkResult
anv_CreateFramebuffer(
2367 const VkFramebufferCreateInfo
* pCreateInfo
,
2368 const VkAllocationCallbacks
* pAllocator
,
2369 VkFramebuffer
* pFramebuffer
)
2371 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2372 struct anv_framebuffer
*framebuffer
;
2374 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
2376 size_t size
= sizeof(*framebuffer
) +
2377 sizeof(struct anv_image_view
*) * pCreateInfo
->attachmentCount
;
2378 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
2379 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2380 if (framebuffer
== NULL
)
2381 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2383 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
2384 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
2385 VkImageView _iview
= pCreateInfo
->pAttachments
[i
];
2386 framebuffer
->attachments
[i
] = anv_image_view_from_handle(_iview
);
2389 framebuffer
->width
= pCreateInfo
->width
;
2390 framebuffer
->height
= pCreateInfo
->height
;
2391 framebuffer
->layers
= pCreateInfo
->layers
;
2393 *pFramebuffer
= anv_framebuffer_to_handle(framebuffer
);
2398 void anv_DestroyFramebuffer(
2401 const VkAllocationCallbacks
* pAllocator
)
2403 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2404 ANV_FROM_HANDLE(anv_framebuffer
, fb
, _fb
);
2409 vk_free2(&device
->alloc
, pAllocator
, fb
);
2412 /* vk_icd.h does not declare this function, so we declare it here to
2413 * suppress Wmissing-prototypes.
2415 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2416 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
);
2418 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2419 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
)
2421 /* For the full details on loader interface versioning, see
2422 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
2423 * What follows is a condensed summary, to help you navigate the large and
2424 * confusing official doc.
2426 * - Loader interface v0 is incompatible with later versions. We don't
2429 * - In loader interface v1:
2430 * - The first ICD entrypoint called by the loader is
2431 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
2433 * - The ICD must statically expose no other Vulkan symbol unless it is
2434 * linked with -Bsymbolic.
2435 * - Each dispatchable Vulkan handle created by the ICD must be
2436 * a pointer to a struct whose first member is VK_LOADER_DATA. The
2437 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
2438 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
2439 * vkDestroySurfaceKHR(). The ICD must be capable of working with
2440 * such loader-managed surfaces.
2442 * - Loader interface v2 differs from v1 in:
2443 * - The first ICD entrypoint called by the loader is
2444 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
2445 * statically expose this entrypoint.
2447 * - Loader interface v3 differs from v2 in:
2448 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
2449 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
2450 * because the loader no longer does so.
2452 *pSupportedVersion
= MIN2(*pSupportedVersion
, 3u);