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>
33 #include "anv_private.h"
34 #include "util/strtod.h"
35 #include "util/debug.h"
36 #include "util/build_id.h"
37 #include "util/mesa-sha1.h"
38 #include "util/vk_util.h"
40 #include "genxml/gen7_pack.h"
43 compiler_debug_log(void *data
, const char *fmt
, ...)
47 compiler_perf_log(void *data
, const char *fmt
, ...)
52 if (unlikely(INTEL_DEBUG
& DEBUG_PERF
))
53 vfprintf(stderr
, fmt
, args
);
59 anv_compute_heap_size(int fd
, uint64_t *heap_size
)
62 if (anv_gem_get_context_param(fd
, 0, I915_CONTEXT_PARAM_GTT_SIZE
,
64 /* If, for whatever reason, we can't actually get the GTT size from the
65 * kernel (too old?) fall back to the aperture size.
67 anv_perf_warn("Failed to get I915_CONTEXT_PARAM_GTT_SIZE: %m");
69 if (anv_gem_get_aperture(fd
, >t_size
) == -1) {
70 return vk_errorf(VK_ERROR_INITIALIZATION_FAILED
,
71 "failed to get aperture size: %m");
75 /* Query the total ram from the system */
79 uint64_t total_ram
= (uint64_t)info
.totalram
* (uint64_t)info
.mem_unit
;
81 /* We don't want to burn too much ram with the GPU. If the user has 4GiB
82 * or less, we use at most half. If they have more than 4GiB, we use 3/4.
84 uint64_t available_ram
;
85 if (total_ram
<= 4ull * 1024ull * 1024ull * 1024ull)
86 available_ram
= total_ram
/ 2;
88 available_ram
= total_ram
* 3 / 4;
90 /* We also want to leave some padding for things we allocate in the driver,
91 * so don't go over 3/4 of the GTT either.
93 uint64_t available_gtt
= gtt_size
* 3 / 4;
95 *heap_size
= MIN2(available_ram
, available_gtt
);
101 anv_physical_device_init_uuids(struct anv_physical_device
*device
)
103 const struct build_id_note
*note
= build_id_find_nhdr("libvulkan_intel.so");
105 return vk_errorf(VK_ERROR_INITIALIZATION_FAILED
,
106 "Failed to find build-id");
109 unsigned build_id_len
= build_id_length(note
);
110 if (build_id_len
< 20) {
111 return vk_errorf(VK_ERROR_INITIALIZATION_FAILED
,
112 "build-id too short. It needs to be a SHA");
115 struct mesa_sha1 sha1_ctx
;
117 STATIC_ASSERT(VK_UUID_SIZE
<= sizeof(sha1
));
119 /* The pipeline cache UUID is used for determining when a pipeline cache is
120 * invalid. It needs both a driver build and the PCI ID of the device.
122 _mesa_sha1_init(&sha1_ctx
);
123 _mesa_sha1_update(&sha1_ctx
, build_id_data(note
), build_id_len
);
124 _mesa_sha1_update(&sha1_ctx
, &device
->chipset_id
,
125 sizeof(device
->chipset_id
));
126 _mesa_sha1_final(&sha1_ctx
, sha1
);
127 memcpy(device
->pipeline_cache_uuid
, sha1
, VK_UUID_SIZE
);
129 /* The driver UUID is used for determining sharability of images and memory
130 * between two Vulkan instances in separate processes. People who want to
131 * share memory need to also check the device UUID (below) so all this
132 * needs to be is the build-id.
134 memcpy(device
->driver_uuid
, build_id_data(note
), VK_UUID_SIZE
);
136 /* The device UUID uniquely identifies the given device within the machine.
137 * Since we never have more than one device, this doesn't need to be a real
138 * UUID. However, on the off-chance that someone tries to use this to
139 * cache pre-tiled images or something of the like, we use the PCI ID and
140 * some bits of ISL info to ensure that this is safe.
142 _mesa_sha1_init(&sha1_ctx
);
143 _mesa_sha1_update(&sha1_ctx
, &device
->chipset_id
,
144 sizeof(device
->chipset_id
));
145 _mesa_sha1_update(&sha1_ctx
, &device
->isl_dev
.has_bit6_swizzling
,
146 sizeof(device
->isl_dev
.has_bit6_swizzling
));
147 _mesa_sha1_final(&sha1_ctx
, sha1
);
148 memcpy(device
->device_uuid
, sha1
, VK_UUID_SIZE
);
154 anv_physical_device_init(struct anv_physical_device
*device
,
155 struct anv_instance
*instance
,
161 fd
= open(path
, O_RDWR
| O_CLOEXEC
);
163 return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
165 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
166 device
->instance
= instance
;
168 assert(strlen(path
) < ARRAY_SIZE(device
->path
));
169 strncpy(device
->path
, path
, ARRAY_SIZE(device
->path
));
171 device
->chipset_id
= anv_gem_get_param(fd
, I915_PARAM_CHIPSET_ID
);
172 if (!device
->chipset_id
) {
173 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
177 device
->name
= gen_get_device_name(device
->chipset_id
);
178 if (!gen_get_device_info(device
->chipset_id
, &device
->info
)) {
179 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
183 if (device
->info
.is_haswell
) {
184 fprintf(stderr
, "WARNING: Haswell Vulkan support is incomplete\n");
185 } else if (device
->info
.gen
== 7 && !device
->info
.is_baytrail
) {
186 fprintf(stderr
, "WARNING: Ivy Bridge Vulkan support is incomplete\n");
187 } else if (device
->info
.gen
== 7 && device
->info
.is_baytrail
) {
188 fprintf(stderr
, "WARNING: Bay Trail Vulkan support is incomplete\n");
189 } else if (device
->info
.gen
>= 8) {
190 /* Broadwell, Cherryview, Skylake, Broxton, Kabylake is as fully
191 * supported as anything */
193 result
= vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER
,
194 "Vulkan not yet supported on %s", device
->name
);
198 device
->cmd_parser_version
= -1;
199 if (device
->info
.gen
== 7) {
200 device
->cmd_parser_version
=
201 anv_gem_get_param(fd
, I915_PARAM_CMD_PARSER_VERSION
);
202 if (device
->cmd_parser_version
== -1) {
203 result
= vk_errorf(VK_ERROR_INITIALIZATION_FAILED
,
204 "failed to get command parser version");
209 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_WAIT_TIMEOUT
)) {
210 result
= vk_errorf(VK_ERROR_INITIALIZATION_FAILED
,
211 "kernel missing gem wait");
215 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_EXECBUF2
)) {
216 result
= vk_errorf(VK_ERROR_INITIALIZATION_FAILED
,
217 "kernel missing execbuf2");
221 if (!device
->info
.has_llc
&&
222 anv_gem_get_param(fd
, I915_PARAM_MMAP_VERSION
) < 1) {
223 result
= vk_errorf(VK_ERROR_INITIALIZATION_FAILED
,
224 "kernel missing wc mmap");
228 device
->supports_48bit_addresses
= anv_gem_supports_48b_addresses(fd
);
230 result
= anv_compute_heap_size(fd
, &device
->heap_size
);
231 if (result
!= VK_SUCCESS
)
234 device
->has_exec_async
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_ASYNC
);
236 bool swizzled
= anv_gem_get_bit6_swizzle(fd
, I915_TILING_X
);
238 /* GENs prior to 8 do not support EU/Subslice info */
239 if (device
->info
.gen
>= 8) {
240 device
->subslice_total
= anv_gem_get_param(fd
, I915_PARAM_SUBSLICE_TOTAL
);
241 device
->eu_total
= anv_gem_get_param(fd
, I915_PARAM_EU_TOTAL
);
243 /* Without this information, we cannot get the right Braswell
244 * brandstrings, and we have to use conservative numbers for GPGPU on
245 * many platforms, but otherwise, things will just work.
247 if (device
->subslice_total
< 1 || device
->eu_total
< 1) {
248 fprintf(stderr
, "WARNING: Kernel 4.1 required to properly"
249 " query GPU properties.\n");
251 } else if (device
->info
.gen
== 7) {
252 device
->subslice_total
= 1 << (device
->info
.gt
- 1);
255 if (device
->info
.is_cherryview
&&
256 device
->subslice_total
> 0 && device
->eu_total
> 0) {
257 /* Logical CS threads = EUs per subslice * 7 threads per EU */
258 uint32_t max_cs_threads
= device
->eu_total
/ device
->subslice_total
* 7;
260 /* Fuse configurations may give more threads than expected, never less. */
261 if (max_cs_threads
> device
->info
.max_cs_threads
)
262 device
->info
.max_cs_threads
= max_cs_threads
;
265 brw_process_intel_debug_variable();
267 device
->compiler
= brw_compiler_create(NULL
, &device
->info
);
268 if (device
->compiler
== NULL
) {
269 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
272 device
->compiler
->shader_debug_log
= compiler_debug_log
;
273 device
->compiler
->shader_perf_log
= compiler_perf_log
;
275 isl_device_init(&device
->isl_dev
, &device
->info
, swizzled
);
277 result
= anv_physical_device_init_uuids(device
);
278 if (result
!= VK_SUCCESS
)
281 result
= anv_init_wsi(device
);
282 if (result
!= VK_SUCCESS
) {
283 ralloc_free(device
->compiler
);
287 device
->local_fd
= fd
;
296 anv_physical_device_finish(struct anv_physical_device
*device
)
298 anv_finish_wsi(device
);
299 ralloc_free(device
->compiler
);
300 close(device
->local_fd
);
303 static const VkExtensionProperties global_extensions
[] = {
305 .extensionName
= VK_KHR_GET_PHYSICAL_DEVICE_PROPERTIES_2_EXTENSION_NAME
,
309 .extensionName
= VK_KHR_SURFACE_EXTENSION_NAME
,
312 #ifdef VK_USE_PLATFORM_WAYLAND_KHR
314 .extensionName
= VK_KHR_WAYLAND_SURFACE_EXTENSION_NAME
,
318 #ifdef VK_USE_PLATFORM_XCB_KHR
320 .extensionName
= VK_KHR_XCB_SURFACE_EXTENSION_NAME
,
324 #ifdef VK_USE_PLATFORM_XLIB_KHR
326 .extensionName
= VK_KHR_XLIB_SURFACE_EXTENSION_NAME
,
331 .extensionName
= VK_KHX_EXTERNAL_MEMORY_CAPABILITIES_EXTENSION_NAME
,
336 static const VkExtensionProperties device_extensions
[] = {
338 .extensionName
= VK_KHR_DESCRIPTOR_UPDATE_TEMPLATE_EXTENSION_NAME
,
342 .extensionName
= VK_KHR_INCREMENTAL_PRESENT_EXTENSION_NAME
,
346 .extensionName
= VK_KHR_MAINTENANCE1_EXTENSION_NAME
,
350 .extensionName
= VK_KHR_PUSH_DESCRIPTOR_EXTENSION_NAME
,
354 .extensionName
= VK_KHR_SAMPLER_MIRROR_CLAMP_TO_EDGE_EXTENSION_NAME
,
358 .extensionName
= VK_KHR_SHADER_DRAW_PARAMETERS_EXTENSION_NAME
,
362 .extensionName
= VK_KHR_SWAPCHAIN_EXTENSION_NAME
,
366 .extensionName
= VK_KHX_EXTERNAL_MEMORY_EXTENSION_NAME
,
370 .extensionName
= VK_KHX_EXTERNAL_MEMORY_FD_EXTENSION_NAME
,
376 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
377 VkSystemAllocationScope allocationScope
)
383 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
384 size_t align
, VkSystemAllocationScope allocationScope
)
386 return realloc(pOriginal
, size
);
390 default_free_func(void *pUserData
, void *pMemory
)
395 static const VkAllocationCallbacks default_alloc
= {
397 .pfnAllocation
= default_alloc_func
,
398 .pfnReallocation
= default_realloc_func
,
399 .pfnFree
= default_free_func
,
402 VkResult
anv_CreateInstance(
403 const VkInstanceCreateInfo
* pCreateInfo
,
404 const VkAllocationCallbacks
* pAllocator
,
405 VkInstance
* pInstance
)
407 struct anv_instance
*instance
;
409 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
411 uint32_t client_version
;
412 if (pCreateInfo
->pApplicationInfo
&&
413 pCreateInfo
->pApplicationInfo
->apiVersion
!= 0) {
414 client_version
= pCreateInfo
->pApplicationInfo
->apiVersion
;
416 client_version
= VK_MAKE_VERSION(1, 0, 0);
419 if (VK_MAKE_VERSION(1, 0, 0) > client_version
||
420 client_version
> VK_MAKE_VERSION(1, 0, 0xfff)) {
421 return vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER
,
422 "Client requested version %d.%d.%d",
423 VK_VERSION_MAJOR(client_version
),
424 VK_VERSION_MINOR(client_version
),
425 VK_VERSION_PATCH(client_version
));
428 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
430 for (uint32_t j
= 0; j
< ARRAY_SIZE(global_extensions
); j
++) {
431 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
432 global_extensions
[j
].extensionName
) == 0) {
438 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
441 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
442 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
444 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
446 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
449 instance
->alloc
= *pAllocator
;
451 instance
->alloc
= default_alloc
;
453 instance
->apiVersion
= client_version
;
454 instance
->physicalDeviceCount
= -1;
458 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
460 *pInstance
= anv_instance_to_handle(instance
);
465 void anv_DestroyInstance(
466 VkInstance _instance
,
467 const VkAllocationCallbacks
* pAllocator
)
469 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
474 if (instance
->physicalDeviceCount
> 0) {
475 /* We support at most one physical device. */
476 assert(instance
->physicalDeviceCount
== 1);
477 anv_physical_device_finish(&instance
->physicalDevice
);
480 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
484 vk_free(&instance
->alloc
, instance
);
488 anv_enumerate_devices(struct anv_instance
*instance
)
490 /* TODO: Check for more devices ? */
491 drmDevicePtr devices
[8];
492 VkResult result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
495 instance
->physicalDeviceCount
= 0;
497 max_devices
= drmGetDevices2(0, devices
, sizeof(devices
));
499 return VK_ERROR_INCOMPATIBLE_DRIVER
;
501 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
502 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
503 devices
[i
]->bustype
== DRM_BUS_PCI
&&
504 devices
[i
]->deviceinfo
.pci
->vendor_id
== 0x8086) {
506 result
= anv_physical_device_init(&instance
->physicalDevice
,
508 devices
[i
]->nodes
[DRM_NODE_RENDER
]);
509 if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
514 if (result
== VK_SUCCESS
)
515 instance
->physicalDeviceCount
= 1;
521 VkResult
anv_EnumeratePhysicalDevices(
522 VkInstance _instance
,
523 uint32_t* pPhysicalDeviceCount
,
524 VkPhysicalDevice
* pPhysicalDevices
)
526 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
527 VK_OUTARRAY_MAKE(out
, pPhysicalDevices
, pPhysicalDeviceCount
);
530 if (instance
->physicalDeviceCount
< 0) {
531 result
= anv_enumerate_devices(instance
);
532 if (result
!= VK_SUCCESS
&&
533 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
537 if (instance
->physicalDeviceCount
> 0) {
538 assert(instance
->physicalDeviceCount
== 1);
539 vk_outarray_append(&out
, i
) {
540 *i
= anv_physical_device_to_handle(&instance
->physicalDevice
);
544 return vk_outarray_status(&out
);
547 void anv_GetPhysicalDeviceFeatures(
548 VkPhysicalDevice physicalDevice
,
549 VkPhysicalDeviceFeatures
* pFeatures
)
551 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
553 *pFeatures
= (VkPhysicalDeviceFeatures
) {
554 .robustBufferAccess
= true,
555 .fullDrawIndexUint32
= true,
556 .imageCubeArray
= true,
557 .independentBlend
= true,
558 .geometryShader
= true,
559 .tessellationShader
= true,
560 .sampleRateShading
= true,
561 .dualSrcBlend
= true,
563 .multiDrawIndirect
= false,
564 .drawIndirectFirstInstance
= true,
566 .depthBiasClamp
= true,
567 .fillModeNonSolid
= true,
568 .depthBounds
= false,
572 .multiViewport
= true,
573 .samplerAnisotropy
= true,
574 .textureCompressionETC2
= pdevice
->info
.gen
>= 8 ||
575 pdevice
->info
.is_baytrail
,
576 .textureCompressionASTC_LDR
= pdevice
->info
.gen
>= 9, /* FINISHME CHV */
577 .textureCompressionBC
= true,
578 .occlusionQueryPrecise
= true,
579 .pipelineStatisticsQuery
= true,
580 .fragmentStoresAndAtomics
= true,
581 .shaderTessellationAndGeometryPointSize
= true,
582 .shaderImageGatherExtended
= true,
583 .shaderStorageImageExtendedFormats
= true,
584 .shaderStorageImageMultisample
= false,
585 .shaderStorageImageReadWithoutFormat
= false,
586 .shaderStorageImageWriteWithoutFormat
= true,
587 .shaderUniformBufferArrayDynamicIndexing
= true,
588 .shaderSampledImageArrayDynamicIndexing
= true,
589 .shaderStorageBufferArrayDynamicIndexing
= true,
590 .shaderStorageImageArrayDynamicIndexing
= true,
591 .shaderClipDistance
= true,
592 .shaderCullDistance
= true,
593 .shaderFloat64
= pdevice
->info
.gen
>= 8,
594 .shaderInt64
= pdevice
->info
.gen
>= 8,
595 .shaderInt16
= false,
596 .shaderResourceMinLod
= false,
597 .variableMultisampleRate
= false,
598 .inheritedQueries
= true,
601 /* We can't do image stores in vec4 shaders */
602 pFeatures
->vertexPipelineStoresAndAtomics
=
603 pdevice
->compiler
->scalar_stage
[MESA_SHADER_VERTEX
] &&
604 pdevice
->compiler
->scalar_stage
[MESA_SHADER_GEOMETRY
];
607 void anv_GetPhysicalDeviceFeatures2KHR(
608 VkPhysicalDevice physicalDevice
,
609 VkPhysicalDeviceFeatures2KHR
* pFeatures
)
611 anv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
613 vk_foreach_struct(ext
, pFeatures
->pNext
) {
614 switch (ext
->sType
) {
615 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES_KHX
: {
616 VkPhysicalDeviceMultiviewFeaturesKHX
*features
=
617 (VkPhysicalDeviceMultiviewFeaturesKHX
*)ext
;
618 features
->multiview
= true;
619 features
->multiviewGeometryShader
= true;
620 features
->multiviewTessellationShader
= true;
625 anv_debug_ignored_stype(ext
->sType
);
631 void anv_GetPhysicalDeviceProperties(
632 VkPhysicalDevice physicalDevice
,
633 VkPhysicalDeviceProperties
* pProperties
)
635 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
636 const struct gen_device_info
*devinfo
= &pdevice
->info
;
638 /* See assertions made when programming the buffer surface state. */
639 const uint32_t max_raw_buffer_sz
= devinfo
->gen
>= 7 ?
640 (1ul << 30) : (1ul << 27);
642 VkSampleCountFlags sample_counts
=
643 isl_device_get_sample_counts(&pdevice
->isl_dev
);
645 VkPhysicalDeviceLimits limits
= {
646 .maxImageDimension1D
= (1 << 14),
647 .maxImageDimension2D
= (1 << 14),
648 .maxImageDimension3D
= (1 << 11),
649 .maxImageDimensionCube
= (1 << 14),
650 .maxImageArrayLayers
= (1 << 11),
651 .maxTexelBufferElements
= 128 * 1024 * 1024,
652 .maxUniformBufferRange
= (1ul << 27),
653 .maxStorageBufferRange
= max_raw_buffer_sz
,
654 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
655 .maxMemoryAllocationCount
= UINT32_MAX
,
656 .maxSamplerAllocationCount
= 64 * 1024,
657 .bufferImageGranularity
= 64, /* A cache line */
658 .sparseAddressSpaceSize
= 0,
659 .maxBoundDescriptorSets
= MAX_SETS
,
660 .maxPerStageDescriptorSamplers
= 64,
661 .maxPerStageDescriptorUniformBuffers
= 64,
662 .maxPerStageDescriptorStorageBuffers
= 64,
663 .maxPerStageDescriptorSampledImages
= 64,
664 .maxPerStageDescriptorStorageImages
= 64,
665 .maxPerStageDescriptorInputAttachments
= 64,
666 .maxPerStageResources
= 128,
667 .maxDescriptorSetSamplers
= 256,
668 .maxDescriptorSetUniformBuffers
= 256,
669 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
670 .maxDescriptorSetStorageBuffers
= 256,
671 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
672 .maxDescriptorSetSampledImages
= 256,
673 .maxDescriptorSetStorageImages
= 256,
674 .maxDescriptorSetInputAttachments
= 256,
675 .maxVertexInputAttributes
= MAX_VBS
,
676 .maxVertexInputBindings
= MAX_VBS
,
677 .maxVertexInputAttributeOffset
= 2047,
678 .maxVertexInputBindingStride
= 2048,
679 .maxVertexOutputComponents
= 128,
680 .maxTessellationGenerationLevel
= 64,
681 .maxTessellationPatchSize
= 32,
682 .maxTessellationControlPerVertexInputComponents
= 128,
683 .maxTessellationControlPerVertexOutputComponents
= 128,
684 .maxTessellationControlPerPatchOutputComponents
= 128,
685 .maxTessellationControlTotalOutputComponents
= 2048,
686 .maxTessellationEvaluationInputComponents
= 128,
687 .maxTessellationEvaluationOutputComponents
= 128,
688 .maxGeometryShaderInvocations
= 32,
689 .maxGeometryInputComponents
= 64,
690 .maxGeometryOutputComponents
= 128,
691 .maxGeometryOutputVertices
= 256,
692 .maxGeometryTotalOutputComponents
= 1024,
693 .maxFragmentInputComponents
= 128,
694 .maxFragmentOutputAttachments
= 8,
695 .maxFragmentDualSrcAttachments
= 1,
696 .maxFragmentCombinedOutputResources
= 8,
697 .maxComputeSharedMemorySize
= 32768,
698 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
699 .maxComputeWorkGroupInvocations
= 16 * devinfo
->max_cs_threads
,
700 .maxComputeWorkGroupSize
= {
701 16 * devinfo
->max_cs_threads
,
702 16 * devinfo
->max_cs_threads
,
703 16 * devinfo
->max_cs_threads
,
705 .subPixelPrecisionBits
= 4 /* FIXME */,
706 .subTexelPrecisionBits
= 4 /* FIXME */,
707 .mipmapPrecisionBits
= 4 /* FIXME */,
708 .maxDrawIndexedIndexValue
= UINT32_MAX
,
709 .maxDrawIndirectCount
= UINT32_MAX
,
710 .maxSamplerLodBias
= 16,
711 .maxSamplerAnisotropy
= 16,
712 .maxViewports
= MAX_VIEWPORTS
,
713 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
714 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
715 .viewportSubPixelBits
= 13, /* We take a float? */
716 .minMemoryMapAlignment
= 4096, /* A page */
717 .minTexelBufferOffsetAlignment
= 1,
718 .minUniformBufferOffsetAlignment
= 16,
719 .minStorageBufferOffsetAlignment
= 4,
720 .minTexelOffset
= -8,
722 .minTexelGatherOffset
= -32,
723 .maxTexelGatherOffset
= 31,
724 .minInterpolationOffset
= -0.5,
725 .maxInterpolationOffset
= 0.4375,
726 .subPixelInterpolationOffsetBits
= 4,
727 .maxFramebufferWidth
= (1 << 14),
728 .maxFramebufferHeight
= (1 << 14),
729 .maxFramebufferLayers
= (1 << 11),
730 .framebufferColorSampleCounts
= sample_counts
,
731 .framebufferDepthSampleCounts
= sample_counts
,
732 .framebufferStencilSampleCounts
= sample_counts
,
733 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
734 .maxColorAttachments
= MAX_RTS
,
735 .sampledImageColorSampleCounts
= sample_counts
,
736 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
737 .sampledImageDepthSampleCounts
= sample_counts
,
738 .sampledImageStencilSampleCounts
= sample_counts
,
739 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
740 .maxSampleMaskWords
= 1,
741 .timestampComputeAndGraphics
= false,
742 .timestampPeriod
= devinfo
->timebase_scale
,
743 .maxClipDistances
= 8,
744 .maxCullDistances
= 8,
745 .maxCombinedClipAndCullDistances
= 8,
746 .discreteQueuePriorities
= 1,
747 .pointSizeRange
= { 0.125, 255.875 },
748 .lineWidthRange
= { 0.0, 7.9921875 },
749 .pointSizeGranularity
= (1.0 / 8.0),
750 .lineWidthGranularity
= (1.0 / 128.0),
751 .strictLines
= false, /* FINISHME */
752 .standardSampleLocations
= true,
753 .optimalBufferCopyOffsetAlignment
= 128,
754 .optimalBufferCopyRowPitchAlignment
= 128,
755 .nonCoherentAtomSize
= 64,
758 *pProperties
= (VkPhysicalDeviceProperties
) {
759 .apiVersion
= VK_MAKE_VERSION(1, 0, 42),
762 .deviceID
= pdevice
->chipset_id
,
763 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
765 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
768 strcpy(pProperties
->deviceName
, pdevice
->name
);
769 memcpy(pProperties
->pipelineCacheUUID
,
770 pdevice
->pipeline_cache_uuid
, VK_UUID_SIZE
);
773 void anv_GetPhysicalDeviceProperties2KHR(
774 VkPhysicalDevice physicalDevice
,
775 VkPhysicalDeviceProperties2KHR
* pProperties
)
777 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
779 anv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
781 vk_foreach_struct(ext
, pProperties
->pNext
) {
782 switch (ext
->sType
) {
783 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
784 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
785 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
787 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
791 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES_KHX
: {
792 VkPhysicalDeviceIDPropertiesKHX
*id_props
=
793 (VkPhysicalDeviceIDPropertiesKHX
*)ext
;
794 memcpy(id_props
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
795 memcpy(id_props
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
796 /* The LUID is for Windows. */
797 id_props
->deviceLUIDValid
= false;
801 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES_KHX
: {
802 VkPhysicalDeviceMultiviewPropertiesKHX
*properties
=
803 (VkPhysicalDeviceMultiviewPropertiesKHX
*)ext
;
804 properties
->maxMultiviewViewCount
= 16;
805 properties
->maxMultiviewInstanceIndex
= UINT32_MAX
/ 16;
810 anv_debug_ignored_stype(ext
->sType
);
816 /* We support exactly one queue family. */
817 static const VkQueueFamilyProperties
818 anv_queue_family_properties
= {
819 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
820 VK_QUEUE_COMPUTE_BIT
|
821 VK_QUEUE_TRANSFER_BIT
,
823 .timestampValidBits
= 36, /* XXX: Real value here */
824 .minImageTransferGranularity
= { 1, 1, 1 },
827 void anv_GetPhysicalDeviceQueueFamilyProperties(
828 VkPhysicalDevice physicalDevice
,
830 VkQueueFamilyProperties
* pQueueFamilyProperties
)
832 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pCount
);
834 vk_outarray_append(&out
, p
) {
835 *p
= anv_queue_family_properties
;
839 void anv_GetPhysicalDeviceQueueFamilyProperties2KHR(
840 VkPhysicalDevice physicalDevice
,
841 uint32_t* pQueueFamilyPropertyCount
,
842 VkQueueFamilyProperties2KHR
* pQueueFamilyProperties
)
845 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pQueueFamilyPropertyCount
);
847 vk_outarray_append(&out
, p
) {
848 p
->queueFamilyProperties
= anv_queue_family_properties
;
850 vk_foreach_struct(s
, p
->pNext
) {
851 anv_debug_ignored_stype(s
->sType
);
856 void anv_GetPhysicalDeviceMemoryProperties(
857 VkPhysicalDevice physicalDevice
,
858 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
860 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
862 if (physical_device
->info
.has_llc
) {
863 /* Big core GPUs share LLC with the CPU and thus one memory type can be
864 * both cached and coherent at the same time.
866 pMemoryProperties
->memoryTypeCount
= 1;
867 pMemoryProperties
->memoryTypes
[0] = (VkMemoryType
) {
868 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
869 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
870 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
871 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
875 /* The spec requires that we expose a host-visible, coherent memory
876 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
877 * to give the application a choice between cached, but not coherent and
878 * coherent but uncached (WC though).
880 pMemoryProperties
->memoryTypeCount
= 2;
881 pMemoryProperties
->memoryTypes
[0] = (VkMemoryType
) {
882 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
883 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
884 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
887 pMemoryProperties
->memoryTypes
[1] = (VkMemoryType
) {
888 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
889 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
890 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
895 pMemoryProperties
->memoryHeapCount
= 1;
896 pMemoryProperties
->memoryHeaps
[0] = (VkMemoryHeap
) {
897 .size
= physical_device
->heap_size
,
898 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
902 void anv_GetPhysicalDeviceMemoryProperties2KHR(
903 VkPhysicalDevice physicalDevice
,
904 VkPhysicalDeviceMemoryProperties2KHR
* pMemoryProperties
)
906 anv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
907 &pMemoryProperties
->memoryProperties
);
909 vk_foreach_struct(ext
, pMemoryProperties
->pNext
) {
910 switch (ext
->sType
) {
912 anv_debug_ignored_stype(ext
->sType
);
918 PFN_vkVoidFunction
anv_GetInstanceProcAddr(
922 return anv_lookup_entrypoint(NULL
, pName
);
925 /* With version 1+ of the loader interface the ICD should expose
926 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
929 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
934 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
938 return anv_GetInstanceProcAddr(instance
, pName
);
941 PFN_vkVoidFunction
anv_GetDeviceProcAddr(
945 ANV_FROM_HANDLE(anv_device
, device
, _device
);
946 return anv_lookup_entrypoint(&device
->info
, pName
);
950 anv_queue_init(struct anv_device
*device
, struct anv_queue
*queue
)
952 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
953 queue
->device
= device
;
954 queue
->pool
= &device
->surface_state_pool
;
958 anv_queue_finish(struct anv_queue
*queue
)
962 static struct anv_state
963 anv_state_pool_emit_data(struct anv_state_pool
*pool
, size_t size
, size_t align
, const void *p
)
965 struct anv_state state
;
967 state
= anv_state_pool_alloc(pool
, size
, align
);
968 memcpy(state
.map
, p
, size
);
970 anv_state_flush(pool
->block_pool
->device
, state
);
975 struct gen8_border_color
{
980 /* Pad out to 64 bytes */
985 anv_device_init_border_colors(struct anv_device
*device
)
987 static const struct gen8_border_color border_colors
[] = {
988 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
989 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
990 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
991 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
992 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
993 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
996 device
->border_colors
= anv_state_pool_emit_data(&device
->dynamic_state_pool
,
997 sizeof(border_colors
), 64,
1001 VkResult
anv_CreateDevice(
1002 VkPhysicalDevice physicalDevice
,
1003 const VkDeviceCreateInfo
* pCreateInfo
,
1004 const VkAllocationCallbacks
* pAllocator
,
1007 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
1009 struct anv_device
*device
;
1011 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO
);
1013 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
1015 for (uint32_t j
= 0; j
< ARRAY_SIZE(device_extensions
); j
++) {
1016 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
1017 device_extensions
[j
].extensionName
) == 0) {
1023 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
1026 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
1028 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
1030 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1032 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1033 device
->instance
= physical_device
->instance
;
1034 device
->chipset_id
= physical_device
->chipset_id
;
1035 device
->lost
= false;
1038 device
->alloc
= *pAllocator
;
1040 device
->alloc
= physical_device
->instance
->alloc
;
1042 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
1043 device
->fd
= open(physical_device
->path
, O_RDWR
| O_CLOEXEC
);
1044 if (device
->fd
== -1) {
1045 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1049 device
->context_id
= anv_gem_create_context(device
);
1050 if (device
->context_id
== -1) {
1051 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1055 device
->info
= physical_device
->info
;
1056 device
->isl_dev
= physical_device
->isl_dev
;
1058 /* On Broadwell and later, we can use batch chaining to more efficiently
1059 * implement growing command buffers. Prior to Haswell, the kernel
1060 * command parser gets in the way and we have to fall back to growing
1063 device
->can_chain_batches
= device
->info
.gen
>= 8;
1065 device
->robust_buffer_access
= pCreateInfo
->pEnabledFeatures
&&
1066 pCreateInfo
->pEnabledFeatures
->robustBufferAccess
;
1068 if (pthread_mutex_init(&device
->mutex
, NULL
) != 0) {
1069 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1070 goto fail_context_id
;
1073 pthread_condattr_t condattr
;
1074 if (pthread_condattr_init(&condattr
) != 0) {
1075 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1078 if (pthread_condattr_setclock(&condattr
, CLOCK_MONOTONIC
) != 0) {
1079 pthread_condattr_destroy(&condattr
);
1080 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1083 if (pthread_cond_init(&device
->queue_submit
, NULL
) != 0) {
1084 pthread_condattr_destroy(&condattr
);
1085 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1088 pthread_condattr_destroy(&condattr
);
1090 anv_bo_pool_init(&device
->batch_bo_pool
, device
);
1092 result
= anv_bo_cache_init(&device
->bo_cache
);
1093 if (result
!= VK_SUCCESS
)
1094 goto fail_batch_bo_pool
;
1096 result
= anv_block_pool_init(&device
->dynamic_state_block_pool
, device
,
1098 if (result
!= VK_SUCCESS
)
1101 anv_state_pool_init(&device
->dynamic_state_pool
,
1102 &device
->dynamic_state_block_pool
);
1104 result
= anv_block_pool_init(&device
->instruction_block_pool
, device
,
1106 if (result
!= VK_SUCCESS
)
1107 goto fail_dynamic_state_pool
;
1109 anv_state_pool_init(&device
->instruction_state_pool
,
1110 &device
->instruction_block_pool
);
1112 result
= anv_block_pool_init(&device
->surface_state_block_pool
, device
,
1114 if (result
!= VK_SUCCESS
)
1115 goto fail_instruction_state_pool
;
1117 anv_state_pool_init(&device
->surface_state_pool
,
1118 &device
->surface_state_block_pool
);
1120 result
= anv_bo_init_new(&device
->workaround_bo
, device
, 1024);
1121 if (result
!= VK_SUCCESS
)
1122 goto fail_surface_state_pool
;
1124 anv_scratch_pool_init(device
, &device
->scratch_pool
);
1126 anv_queue_init(device
, &device
->queue
);
1128 switch (device
->info
.gen
) {
1130 if (!device
->info
.is_haswell
)
1131 result
= gen7_init_device_state(device
);
1133 result
= gen75_init_device_state(device
);
1136 result
= gen8_init_device_state(device
);
1139 result
= gen9_init_device_state(device
);
1142 /* Shouldn't get here as we don't create physical devices for any other
1144 unreachable("unhandled gen");
1146 if (result
!= VK_SUCCESS
)
1147 goto fail_workaround_bo
;
1149 anv_device_init_blorp(device
);
1151 anv_device_init_border_colors(device
);
1153 *pDevice
= anv_device_to_handle(device
);
1158 anv_queue_finish(&device
->queue
);
1159 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
1160 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
1161 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
1162 fail_surface_state_pool
:
1163 anv_state_pool_finish(&device
->surface_state_pool
);
1164 anv_block_pool_finish(&device
->surface_state_block_pool
);
1165 fail_instruction_state_pool
:
1166 anv_state_pool_finish(&device
->instruction_state_pool
);
1167 anv_block_pool_finish(&device
->instruction_block_pool
);
1168 fail_dynamic_state_pool
:
1169 anv_state_pool_finish(&device
->dynamic_state_pool
);
1170 anv_block_pool_finish(&device
->dynamic_state_block_pool
);
1172 anv_bo_cache_finish(&device
->bo_cache
);
1174 anv_bo_pool_finish(&device
->batch_bo_pool
);
1175 pthread_cond_destroy(&device
->queue_submit
);
1177 pthread_mutex_destroy(&device
->mutex
);
1179 anv_gem_destroy_context(device
, device
->context_id
);
1183 vk_free(&device
->alloc
, device
);
1188 void anv_DestroyDevice(
1190 const VkAllocationCallbacks
* pAllocator
)
1192 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1197 anv_device_finish_blorp(device
);
1199 anv_queue_finish(&device
->queue
);
1201 #ifdef HAVE_VALGRIND
1202 /* We only need to free these to prevent valgrind errors. The backing
1203 * BO will go away in a couple of lines so we don't actually leak.
1205 anv_state_pool_free(&device
->dynamic_state_pool
, device
->border_colors
);
1208 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
1210 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
1211 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
1213 anv_state_pool_finish(&device
->surface_state_pool
);
1214 anv_block_pool_finish(&device
->surface_state_block_pool
);
1215 anv_state_pool_finish(&device
->instruction_state_pool
);
1216 anv_block_pool_finish(&device
->instruction_block_pool
);
1217 anv_state_pool_finish(&device
->dynamic_state_pool
);
1218 anv_block_pool_finish(&device
->dynamic_state_block_pool
);
1220 anv_bo_cache_finish(&device
->bo_cache
);
1222 anv_bo_pool_finish(&device
->batch_bo_pool
);
1224 pthread_cond_destroy(&device
->queue_submit
);
1225 pthread_mutex_destroy(&device
->mutex
);
1227 anv_gem_destroy_context(device
, device
->context_id
);
1231 vk_free(&device
->alloc
, device
);
1234 VkResult
anv_EnumerateInstanceExtensionProperties(
1235 const char* pLayerName
,
1236 uint32_t* pPropertyCount
,
1237 VkExtensionProperties
* pProperties
)
1239 if (pProperties
== NULL
) {
1240 *pPropertyCount
= ARRAY_SIZE(global_extensions
);
1244 *pPropertyCount
= MIN2(*pPropertyCount
, ARRAY_SIZE(global_extensions
));
1245 typed_memcpy(pProperties
, global_extensions
, *pPropertyCount
);
1247 if (*pPropertyCount
< ARRAY_SIZE(global_extensions
))
1248 return VK_INCOMPLETE
;
1253 VkResult
anv_EnumerateDeviceExtensionProperties(
1254 VkPhysicalDevice physicalDevice
,
1255 const char* pLayerName
,
1256 uint32_t* pPropertyCount
,
1257 VkExtensionProperties
* pProperties
)
1259 if (pProperties
== NULL
) {
1260 *pPropertyCount
= ARRAY_SIZE(device_extensions
);
1264 *pPropertyCount
= MIN2(*pPropertyCount
, ARRAY_SIZE(device_extensions
));
1265 typed_memcpy(pProperties
, device_extensions
, *pPropertyCount
);
1267 if (*pPropertyCount
< ARRAY_SIZE(device_extensions
))
1268 return VK_INCOMPLETE
;
1273 VkResult
anv_EnumerateInstanceLayerProperties(
1274 uint32_t* pPropertyCount
,
1275 VkLayerProperties
* pProperties
)
1277 if (pProperties
== NULL
) {
1278 *pPropertyCount
= 0;
1282 /* None supported at this time */
1283 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1286 VkResult
anv_EnumerateDeviceLayerProperties(
1287 VkPhysicalDevice physicalDevice
,
1288 uint32_t* pPropertyCount
,
1289 VkLayerProperties
* pProperties
)
1291 if (pProperties
== NULL
) {
1292 *pPropertyCount
= 0;
1296 /* None supported at this time */
1297 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1300 void anv_GetDeviceQueue(
1302 uint32_t queueNodeIndex
,
1303 uint32_t queueIndex
,
1306 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1308 assert(queueIndex
== 0);
1310 *pQueue
= anv_queue_to_handle(&device
->queue
);
1314 anv_device_query_status(struct anv_device
*device
)
1316 /* This isn't likely as most of the callers of this function already check
1317 * for it. However, it doesn't hurt to check and it potentially lets us
1320 if (unlikely(device
->lost
))
1321 return VK_ERROR_DEVICE_LOST
;
1323 uint32_t active
, pending
;
1324 int ret
= anv_gem_gpu_get_reset_stats(device
, &active
, &pending
);
1326 /* We don't know the real error. */
1327 device
->lost
= true;
1328 return vk_errorf(VK_ERROR_DEVICE_LOST
, "get_reset_stats failed: %m");
1332 device
->lost
= true;
1333 return vk_errorf(VK_ERROR_DEVICE_LOST
,
1334 "GPU hung on one of our command buffers");
1335 } else if (pending
) {
1336 device
->lost
= true;
1337 return vk_errorf(VK_ERROR_DEVICE_LOST
,
1338 "GPU hung with commands in-flight");
1345 anv_device_bo_busy(struct anv_device
*device
, struct anv_bo
*bo
)
1347 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
1348 * Other usages of the BO (such as on different hardware) will not be
1349 * flagged as "busy" by this ioctl. Use with care.
1351 int ret
= anv_gem_busy(device
, bo
->gem_handle
);
1353 return VK_NOT_READY
;
1354 } else if (ret
== -1) {
1355 /* We don't know the real error. */
1356 device
->lost
= true;
1357 return vk_errorf(VK_ERROR_DEVICE_LOST
, "gem wait failed: %m");
1360 /* Query for device status after the busy call. If the BO we're checking
1361 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
1362 * client because it clearly doesn't have valid data. Yes, this most
1363 * likely means an ioctl, but we just did an ioctl to query the busy status
1364 * so it's no great loss.
1366 return anv_device_query_status(device
);
1370 anv_device_wait(struct anv_device
*device
, struct anv_bo
*bo
,
1373 int ret
= anv_gem_wait(device
, bo
->gem_handle
, &timeout
);
1374 if (ret
== -1 && errno
== ETIME
) {
1376 } else if (ret
== -1) {
1377 /* We don't know the real error. */
1378 device
->lost
= true;
1379 return vk_errorf(VK_ERROR_DEVICE_LOST
, "gem wait failed: %m");
1382 /* Query for device status after the wait. If the BO we're waiting on got
1383 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
1384 * because it clearly doesn't have valid data. Yes, this most likely means
1385 * an ioctl, but we just did an ioctl to wait so it's no great loss.
1387 return anv_device_query_status(device
);
1390 VkResult
anv_DeviceWaitIdle(
1393 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1394 if (unlikely(device
->lost
))
1395 return VK_ERROR_DEVICE_LOST
;
1397 struct anv_batch batch
;
1400 batch
.start
= batch
.next
= cmds
;
1401 batch
.end
= (void *) cmds
+ sizeof(cmds
);
1403 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
1404 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
1406 return anv_device_submit_simple_batch(device
, &batch
);
1410 anv_bo_init_new(struct anv_bo
*bo
, struct anv_device
*device
, uint64_t size
)
1412 uint32_t gem_handle
= anv_gem_create(device
, size
);
1414 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
1416 anv_bo_init(bo
, gem_handle
, size
);
1418 if (device
->instance
->physicalDevice
.supports_48bit_addresses
)
1419 bo
->flags
|= EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
1421 if (device
->instance
->physicalDevice
.has_exec_async
)
1422 bo
->flags
|= EXEC_OBJECT_ASYNC
;
1427 VkResult
anv_AllocateMemory(
1429 const VkMemoryAllocateInfo
* pAllocateInfo
,
1430 const VkAllocationCallbacks
* pAllocator
,
1431 VkDeviceMemory
* pMem
)
1433 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1434 struct anv_device_memory
*mem
;
1435 VkResult result
= VK_SUCCESS
;
1437 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
1439 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
1440 assert(pAllocateInfo
->allocationSize
> 0);
1442 /* We support exactly one memory heap. */
1443 assert(pAllocateInfo
->memoryTypeIndex
== 0 ||
1444 (!device
->info
.has_llc
&& pAllocateInfo
->memoryTypeIndex
< 2));
1446 /* The kernel relocation API has a limitation of a 32-bit delta value
1447 * applied to the address before it is written which, in spite of it being
1448 * unsigned, is treated as signed . Because of the way that this maps to
1449 * the Vulkan API, we cannot handle an offset into a buffer that does not
1450 * fit into a signed 32 bits. The only mechanism we have for dealing with
1451 * this at the moment is to limit all VkDeviceMemory objects to a maximum
1452 * of 2GB each. The Vulkan spec allows us to do this:
1454 * "Some platforms may have a limit on the maximum size of a single
1455 * allocation. For example, certain systems may fail to create
1456 * allocations with a size greater than or equal to 4GB. Such a limit is
1457 * implementation-dependent, and if such a failure occurs then the error
1458 * VK_ERROR_OUT_OF_DEVICE_MEMORY should be returned."
1460 * We don't use vk_error here because it's not an error so much as an
1461 * indication to the application that the allocation is too large.
1463 if (pAllocateInfo
->allocationSize
> (1ull << 31))
1464 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
1466 /* FINISHME: Fail if allocation request exceeds heap size. */
1468 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
1469 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1471 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1473 mem
->type_index
= pAllocateInfo
->memoryTypeIndex
;
1477 const VkImportMemoryFdInfoKHX
*fd_info
=
1478 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_FD_INFO_KHX
);
1480 /* The Vulkan spec permits handleType to be 0, in which case the struct is
1483 if (fd_info
&& fd_info
->handleType
) {
1484 /* At the moment, we only support the OPAQUE_FD memory type which is
1485 * just a GEM buffer.
1487 assert(fd_info
->handleType
==
1488 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHX
);
1490 result
= anv_bo_cache_import(device
, &device
->bo_cache
,
1491 fd_info
->fd
, pAllocateInfo
->allocationSize
,
1493 if (result
!= VK_SUCCESS
)
1496 result
= anv_bo_cache_alloc(device
, &device
->bo_cache
,
1497 pAllocateInfo
->allocationSize
,
1499 if (result
!= VK_SUCCESS
)
1503 *pMem
= anv_device_memory_to_handle(mem
);
1508 vk_free2(&device
->alloc
, pAllocator
, mem
);
1513 VkResult
anv_GetMemoryFdKHX(
1515 VkDeviceMemory memory_h
,
1516 VkExternalMemoryHandleTypeFlagBitsKHX handleType
,
1519 ANV_FROM_HANDLE(anv_device
, dev
, device_h
);
1520 ANV_FROM_HANDLE(anv_device_memory
, mem
, memory_h
);
1522 /* We support only one handle type. */
1523 assert(handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHX
);
1525 return anv_bo_cache_export(dev
, &dev
->bo_cache
, mem
->bo
, pFd
);
1528 VkResult
anv_GetMemoryFdPropertiesKHX(
1530 VkExternalMemoryHandleTypeFlagBitsKHX handleType
,
1532 VkMemoryFdPropertiesKHX
* pMemoryFdProperties
)
1534 /* The valid usage section for this function says:
1536 * "handleType must not be one of the handle types defined as opaque."
1538 * Since we only handle opaque handles for now, there are no FD properties.
1540 return VK_ERROR_INVALID_EXTERNAL_HANDLE_KHX
;
1543 void anv_FreeMemory(
1545 VkDeviceMemory _mem
,
1546 const VkAllocationCallbacks
* pAllocator
)
1548 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1549 ANV_FROM_HANDLE(anv_device_memory
, mem
, _mem
);
1555 anv_UnmapMemory(_device
, _mem
);
1557 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
1559 vk_free2(&device
->alloc
, pAllocator
, mem
);
1562 VkResult
anv_MapMemory(
1564 VkDeviceMemory _memory
,
1565 VkDeviceSize offset
,
1567 VkMemoryMapFlags flags
,
1570 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1571 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
1578 if (size
== VK_WHOLE_SIZE
)
1579 size
= mem
->bo
->size
- offset
;
1581 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
1583 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
1584 * assert(size != 0);
1585 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
1586 * equal to the size of the memory minus offset
1589 assert(offset
+ size
<= mem
->bo
->size
);
1591 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
1592 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
1593 * at a time is valid. We could just mmap up front and return an offset
1594 * pointer here, but that may exhaust virtual memory on 32 bit
1597 uint32_t gem_flags
= 0;
1598 if (!device
->info
.has_llc
&& mem
->type_index
== 0)
1599 gem_flags
|= I915_MMAP_WC
;
1601 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
1602 uint64_t map_offset
= offset
& ~4095ull;
1603 assert(offset
>= map_offset
);
1604 uint64_t map_size
= (offset
+ size
) - map_offset
;
1606 /* Let's map whole pages */
1607 map_size
= align_u64(map_size
, 4096);
1609 void *map
= anv_gem_mmap(device
, mem
->bo
->gem_handle
,
1610 map_offset
, map_size
, gem_flags
);
1611 if (map
== MAP_FAILED
)
1612 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
1615 mem
->map_size
= map_size
;
1617 *ppData
= mem
->map
+ (offset
- map_offset
);
1622 void anv_UnmapMemory(
1624 VkDeviceMemory _memory
)
1626 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
1631 anv_gem_munmap(mem
->map
, mem
->map_size
);
1638 clflush_mapped_ranges(struct anv_device
*device
,
1640 const VkMappedMemoryRange
*ranges
)
1642 for (uint32_t i
= 0; i
< count
; i
++) {
1643 ANV_FROM_HANDLE(anv_device_memory
, mem
, ranges
[i
].memory
);
1644 if (ranges
[i
].offset
>= mem
->map_size
)
1647 anv_clflush_range(mem
->map
+ ranges
[i
].offset
,
1648 MIN2(ranges
[i
].size
, mem
->map_size
- ranges
[i
].offset
));
1652 VkResult
anv_FlushMappedMemoryRanges(
1654 uint32_t memoryRangeCount
,
1655 const VkMappedMemoryRange
* pMemoryRanges
)
1657 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1659 if (device
->info
.has_llc
)
1662 /* Make sure the writes we're flushing have landed. */
1663 __builtin_ia32_mfence();
1665 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
1670 VkResult
anv_InvalidateMappedMemoryRanges(
1672 uint32_t memoryRangeCount
,
1673 const VkMappedMemoryRange
* pMemoryRanges
)
1675 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1677 if (device
->info
.has_llc
)
1680 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
1682 /* Make sure no reads get moved up above the invalidate. */
1683 __builtin_ia32_mfence();
1688 void anv_GetBufferMemoryRequirements(
1691 VkMemoryRequirements
* pMemoryRequirements
)
1693 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
1694 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1696 /* The Vulkan spec (git aaed022) says:
1698 * memoryTypeBits is a bitfield and contains one bit set for every
1699 * supported memory type for the resource. The bit `1<<i` is set if and
1700 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1701 * structure for the physical device is supported.
1703 * We support exactly one memory type on LLC, two on non-LLC.
1705 pMemoryRequirements
->memoryTypeBits
= device
->info
.has_llc
? 1 : 3;
1707 pMemoryRequirements
->size
= buffer
->size
;
1708 pMemoryRequirements
->alignment
= 16;
1711 void anv_GetImageMemoryRequirements(
1714 VkMemoryRequirements
* pMemoryRequirements
)
1716 ANV_FROM_HANDLE(anv_image
, image
, _image
);
1717 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1719 /* The Vulkan spec (git aaed022) says:
1721 * memoryTypeBits is a bitfield and contains one bit set for every
1722 * supported memory type for the resource. The bit `1<<i` is set if and
1723 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1724 * structure for the physical device is supported.
1726 * We support exactly one memory type on LLC, two on non-LLC.
1728 pMemoryRequirements
->memoryTypeBits
= device
->info
.has_llc
? 1 : 3;
1730 pMemoryRequirements
->size
= image
->size
;
1731 pMemoryRequirements
->alignment
= image
->alignment
;
1734 void anv_GetImageSparseMemoryRequirements(
1737 uint32_t* pSparseMemoryRequirementCount
,
1738 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
1740 *pSparseMemoryRequirementCount
= 0;
1743 void anv_GetDeviceMemoryCommitment(
1745 VkDeviceMemory memory
,
1746 VkDeviceSize
* pCommittedMemoryInBytes
)
1748 *pCommittedMemoryInBytes
= 0;
1751 VkResult
anv_BindBufferMemory(
1754 VkDeviceMemory _memory
,
1755 VkDeviceSize memoryOffset
)
1757 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
1758 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
1761 buffer
->bo
= mem
->bo
;
1762 buffer
->offset
= memoryOffset
;
1771 VkResult
anv_QueueBindSparse(
1773 uint32_t bindInfoCount
,
1774 const VkBindSparseInfo
* pBindInfo
,
1777 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
1778 if (unlikely(queue
->device
->lost
))
1779 return VK_ERROR_DEVICE_LOST
;
1781 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
1786 VkResult
anv_CreateEvent(
1788 const VkEventCreateInfo
* pCreateInfo
,
1789 const VkAllocationCallbacks
* pAllocator
,
1792 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1793 struct anv_state state
;
1794 struct anv_event
*event
;
1796 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_EVENT_CREATE_INFO
);
1798 state
= anv_state_pool_alloc(&device
->dynamic_state_pool
,
1801 event
->state
= state
;
1802 event
->semaphore
= VK_EVENT_RESET
;
1804 if (!device
->info
.has_llc
) {
1805 /* Make sure the writes we're flushing have landed. */
1806 __builtin_ia32_mfence();
1807 __builtin_ia32_clflush(event
);
1810 *pEvent
= anv_event_to_handle(event
);
1815 void anv_DestroyEvent(
1818 const VkAllocationCallbacks
* pAllocator
)
1820 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1821 ANV_FROM_HANDLE(anv_event
, event
, _event
);
1826 anv_state_pool_free(&device
->dynamic_state_pool
, event
->state
);
1829 VkResult
anv_GetEventStatus(
1833 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1834 ANV_FROM_HANDLE(anv_event
, event
, _event
);
1836 if (unlikely(device
->lost
))
1837 return VK_ERROR_DEVICE_LOST
;
1839 if (!device
->info
.has_llc
) {
1840 /* Invalidate read cache before reading event written by GPU. */
1841 __builtin_ia32_clflush(event
);
1842 __builtin_ia32_mfence();
1846 return event
->semaphore
;
1849 VkResult
anv_SetEvent(
1853 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1854 ANV_FROM_HANDLE(anv_event
, event
, _event
);
1856 event
->semaphore
= VK_EVENT_SET
;
1858 if (!device
->info
.has_llc
) {
1859 /* Make sure the writes we're flushing have landed. */
1860 __builtin_ia32_mfence();
1861 __builtin_ia32_clflush(event
);
1867 VkResult
anv_ResetEvent(
1871 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1872 ANV_FROM_HANDLE(anv_event
, event
, _event
);
1874 event
->semaphore
= VK_EVENT_RESET
;
1876 if (!device
->info
.has_llc
) {
1877 /* Make sure the writes we're flushing have landed. */
1878 __builtin_ia32_mfence();
1879 __builtin_ia32_clflush(event
);
1887 VkResult
anv_CreateBuffer(
1889 const VkBufferCreateInfo
* pCreateInfo
,
1890 const VkAllocationCallbacks
* pAllocator
,
1893 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1894 struct anv_buffer
*buffer
;
1896 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
1898 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
1899 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1901 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1903 buffer
->size
= pCreateInfo
->size
;
1904 buffer
->usage
= pCreateInfo
->usage
;
1908 *pBuffer
= anv_buffer_to_handle(buffer
);
1913 void anv_DestroyBuffer(
1916 const VkAllocationCallbacks
* pAllocator
)
1918 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1919 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
1924 vk_free2(&device
->alloc
, pAllocator
, buffer
);
1928 anv_fill_buffer_surface_state(struct anv_device
*device
, struct anv_state state
,
1929 enum isl_format format
,
1930 uint32_t offset
, uint32_t range
, uint32_t stride
)
1932 isl_buffer_fill_state(&device
->isl_dev
, state
.map
,
1934 .mocs
= device
->default_mocs
,
1939 anv_state_flush(device
, state
);
1942 void anv_DestroySampler(
1945 const VkAllocationCallbacks
* pAllocator
)
1947 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1948 ANV_FROM_HANDLE(anv_sampler
, sampler
, _sampler
);
1953 vk_free2(&device
->alloc
, pAllocator
, sampler
);
1956 VkResult
anv_CreateFramebuffer(
1958 const VkFramebufferCreateInfo
* pCreateInfo
,
1959 const VkAllocationCallbacks
* pAllocator
,
1960 VkFramebuffer
* pFramebuffer
)
1962 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1963 struct anv_framebuffer
*framebuffer
;
1965 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
1967 size_t size
= sizeof(*framebuffer
) +
1968 sizeof(struct anv_image_view
*) * pCreateInfo
->attachmentCount
;
1969 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
1970 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1971 if (framebuffer
== NULL
)
1972 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1974 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
1975 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
1976 VkImageView _iview
= pCreateInfo
->pAttachments
[i
];
1977 framebuffer
->attachments
[i
] = anv_image_view_from_handle(_iview
);
1980 framebuffer
->width
= pCreateInfo
->width
;
1981 framebuffer
->height
= pCreateInfo
->height
;
1982 framebuffer
->layers
= pCreateInfo
->layers
;
1984 *pFramebuffer
= anv_framebuffer_to_handle(framebuffer
);
1989 void anv_DestroyFramebuffer(
1992 const VkAllocationCallbacks
* pAllocator
)
1994 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1995 ANV_FROM_HANDLE(anv_framebuffer
, fb
, _fb
);
2000 vk_free2(&device
->alloc
, pAllocator
, fb
);
2003 /* vk_icd.h does not declare this function, so we declare it here to
2004 * suppress Wmissing-prototypes.
2006 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2007 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
);
2009 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2010 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
)
2012 /* For the full details on loader interface versioning, see
2013 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
2014 * What follows is a condensed summary, to help you navigate the large and
2015 * confusing official doc.
2017 * - Loader interface v0 is incompatible with later versions. We don't
2020 * - In loader interface v1:
2021 * - The first ICD entrypoint called by the loader is
2022 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
2024 * - The ICD must statically expose no other Vulkan symbol unless it is
2025 * linked with -Bsymbolic.
2026 * - Each dispatchable Vulkan handle created by the ICD must be
2027 * a pointer to a struct whose first member is VK_LOADER_DATA. The
2028 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
2029 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
2030 * vkDestroySurfaceKHR(). The ICD must be capable of working with
2031 * such loader-managed surfaces.
2033 * - Loader interface v2 differs from v1 in:
2034 * - The first ICD entrypoint called by the loader is
2035 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
2036 * statically expose this entrypoint.
2038 * - Loader interface v3 differs from v2 in:
2039 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
2040 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
2041 * because the loader no longer does so.
2043 *pSupportedVersion
= MIN2(*pSupportedVersion
, 3u);