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"
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 intel_logd_v(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(NULL
, NULL
,
68 "Failed to get I915_CONTEXT_PARAM_GTT_SIZE: %m");
70 if (anv_gem_get_aperture(fd
, >t_size
) == -1) {
71 return vk_errorf(NULL
, NULL
, VK_ERROR_INITIALIZATION_FAILED
,
72 "failed to get aperture size: %m");
76 /* Query the total ram from the system */
80 uint64_t total_ram
= (uint64_t)info
.totalram
* (uint64_t)info
.mem_unit
;
82 /* We don't want to burn too much ram with the GPU. If the user has 4GiB
83 * or less, we use at most half. If they have more than 4GiB, we use 3/4.
85 uint64_t available_ram
;
86 if (total_ram
<= 4ull * 1024ull * 1024ull * 1024ull)
87 available_ram
= total_ram
/ 2;
89 available_ram
= total_ram
* 3 / 4;
91 /* We also want to leave some padding for things we allocate in the driver,
92 * so don't go over 3/4 of the GTT either.
94 uint64_t available_gtt
= gtt_size
* 3 / 4;
96 *heap_size
= MIN2(available_ram
, available_gtt
);
102 anv_physical_device_init_heaps(struct anv_physical_device
*device
, int fd
)
104 /* The kernel query only tells us whether or not the kernel supports the
105 * EXEC_OBJECT_SUPPORTS_48B_ADDRESS flag and not whether or not the
106 * hardware has actual 48bit address support.
108 device
->supports_48bit_addresses
=
109 (device
->info
.gen
>= 8) && anv_gem_supports_48b_addresses(fd
);
112 VkResult result
= anv_compute_heap_size(fd
, &heap_size
);
113 if (result
!= VK_SUCCESS
)
116 if (heap_size
<= 3ull * (1ull << 30)) {
117 /* In this case, everything fits nicely into the 32-bit address space,
118 * so there's no need for supporting 48bit addresses on client-allocated
121 device
->memory
.heap_count
= 1;
122 device
->memory
.heaps
[0] = (struct anv_memory_heap
) {
124 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
125 .supports_48bit_addresses
= false,
128 /* Not everything will fit nicely into a 32-bit address space. In this
129 * case we need a 64-bit heap. Advertise a small 32-bit heap and a
130 * larger 48-bit heap. If we're in this case, then we have a total heap
131 * size larger than 3GiB which most likely means they have 8 GiB of
132 * video memory and so carving off 1 GiB for the 32-bit heap should be
135 const uint64_t heap_size_32bit
= 1ull << 30;
136 const uint64_t heap_size_48bit
= heap_size
- heap_size_32bit
;
138 assert(device
->supports_48bit_addresses
);
140 device
->memory
.heap_count
= 2;
141 device
->memory
.heaps
[0] = (struct anv_memory_heap
) {
142 .size
= heap_size_48bit
,
143 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
144 .supports_48bit_addresses
= true,
146 device
->memory
.heaps
[1] = (struct anv_memory_heap
) {
147 .size
= heap_size_32bit
,
148 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
149 .supports_48bit_addresses
= false,
153 uint32_t type_count
= 0;
154 for (uint32_t heap
= 0; heap
< device
->memory
.heap_count
; heap
++) {
155 uint32_t valid_buffer_usage
= ~0;
157 /* There appears to be a hardware issue in the VF cache where it only
158 * considers the bottom 32 bits of memory addresses. If you happen to
159 * have two vertex buffers which get placed exactly 4 GiB apart and use
160 * them in back-to-back draw calls, you can get collisions. In order to
161 * solve this problem, we require vertex and index buffers be bound to
162 * memory allocated out of the 32-bit heap.
164 if (device
->memory
.heaps
[heap
].supports_48bit_addresses
) {
165 valid_buffer_usage
&= ~(VK_BUFFER_USAGE_INDEX_BUFFER_BIT
|
166 VK_BUFFER_USAGE_VERTEX_BUFFER_BIT
);
169 if (device
->info
.has_llc
) {
170 /* Big core GPUs share LLC with the CPU and thus one memory type can be
171 * both cached and coherent at the same time.
173 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
174 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
175 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
176 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
177 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
179 .valid_buffer_usage
= valid_buffer_usage
,
182 /* The spec requires that we expose a host-visible, coherent memory
183 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
184 * to give the application a choice between cached, but not coherent and
185 * coherent but uncached (WC though).
187 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
188 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
189 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
190 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
192 .valid_buffer_usage
= valid_buffer_usage
,
194 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
195 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
196 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
197 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
199 .valid_buffer_usage
= valid_buffer_usage
,
203 device
->memory
.type_count
= type_count
;
209 anv_physical_device_init_uuids(struct anv_physical_device
*device
)
211 const struct build_id_note
*note
=
212 build_id_find_nhdr_for_addr(anv_physical_device_init_uuids
);
214 return vk_errorf(device
->instance
, device
,
215 VK_ERROR_INITIALIZATION_FAILED
,
216 "Failed to find build-id");
219 unsigned build_id_len
= build_id_length(note
);
220 if (build_id_len
< 20) {
221 return vk_errorf(device
->instance
, device
,
222 VK_ERROR_INITIALIZATION_FAILED
,
223 "build-id too short. It needs to be a SHA");
226 struct mesa_sha1 sha1_ctx
;
228 STATIC_ASSERT(VK_UUID_SIZE
<= sizeof(sha1
));
230 /* The pipeline cache UUID is used for determining when a pipeline cache is
231 * invalid. It needs both a driver build and the PCI ID of the device.
233 _mesa_sha1_init(&sha1_ctx
);
234 _mesa_sha1_update(&sha1_ctx
, build_id_data(note
), build_id_len
);
235 _mesa_sha1_update(&sha1_ctx
, &device
->chipset_id
,
236 sizeof(device
->chipset_id
));
237 _mesa_sha1_final(&sha1_ctx
, sha1
);
238 memcpy(device
->pipeline_cache_uuid
, sha1
, VK_UUID_SIZE
);
240 /* The driver UUID is used for determining sharability of images and memory
241 * between two Vulkan instances in separate processes. People who want to
242 * share memory need to also check the device UUID (below) so all this
243 * needs to be is the build-id.
245 memcpy(device
->driver_uuid
, build_id_data(note
), VK_UUID_SIZE
);
247 /* The device UUID uniquely identifies the given device within the machine.
248 * Since we never have more than one device, this doesn't need to be a real
249 * UUID. However, on the off-chance that someone tries to use this to
250 * cache pre-tiled images or something of the like, we use the PCI ID and
251 * some bits of ISL info to ensure that this is safe.
253 _mesa_sha1_init(&sha1_ctx
);
254 _mesa_sha1_update(&sha1_ctx
, &device
->chipset_id
,
255 sizeof(device
->chipset_id
));
256 _mesa_sha1_update(&sha1_ctx
, &device
->isl_dev
.has_bit6_swizzling
,
257 sizeof(device
->isl_dev
.has_bit6_swizzling
));
258 _mesa_sha1_final(&sha1_ctx
, sha1
);
259 memcpy(device
->device_uuid
, sha1
, VK_UUID_SIZE
);
265 anv_physical_device_init(struct anv_physical_device
*device
,
266 struct anv_instance
*instance
,
272 brw_process_intel_debug_variable();
274 fd
= open(path
, O_RDWR
| O_CLOEXEC
);
276 return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
278 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
279 device
->instance
= instance
;
281 assert(strlen(path
) < ARRAY_SIZE(device
->path
));
282 strncpy(device
->path
, path
, ARRAY_SIZE(device
->path
));
284 device
->chipset_id
= anv_gem_get_param(fd
, I915_PARAM_CHIPSET_ID
);
285 if (!device
->chipset_id
) {
286 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
290 device
->name
= gen_get_device_name(device
->chipset_id
);
291 if (!gen_get_device_info(device
->chipset_id
, &device
->info
)) {
292 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
296 if (device
->info
.is_haswell
) {
297 intel_logw("Haswell Vulkan support is incomplete");
298 } else if (device
->info
.gen
== 7 && !device
->info
.is_baytrail
) {
299 intel_logw("Ivy Bridge Vulkan support is incomplete");
300 } else if (device
->info
.gen
== 7 && device
->info
.is_baytrail
) {
301 intel_logw("Bay Trail Vulkan support is incomplete");
302 } else if (device
->info
.gen
>= 8 && device
->info
.gen
<= 9) {
303 /* Broadwell, Cherryview, Skylake, Broxton, Kabylake, Coffelake is as
304 * fully supported as anything */
305 } else if (device
->info
.gen
== 10) {
306 intel_logw("Cannonlake Vulkan support is alpha");
308 result
= vk_errorf(device
->instance
, device
,
309 VK_ERROR_INCOMPATIBLE_DRIVER
,
310 "Vulkan not yet supported on %s", device
->name
);
314 device
->cmd_parser_version
= -1;
315 if (device
->info
.gen
== 7) {
316 device
->cmd_parser_version
=
317 anv_gem_get_param(fd
, I915_PARAM_CMD_PARSER_VERSION
);
318 if (device
->cmd_parser_version
== -1) {
319 result
= vk_errorf(device
->instance
, device
,
320 VK_ERROR_INITIALIZATION_FAILED
,
321 "failed to get command parser version");
326 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_WAIT_TIMEOUT
)) {
327 result
= vk_errorf(device
->instance
, device
,
328 VK_ERROR_INITIALIZATION_FAILED
,
329 "kernel missing gem wait");
333 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_EXECBUF2
)) {
334 result
= vk_errorf(device
->instance
, device
,
335 VK_ERROR_INITIALIZATION_FAILED
,
336 "kernel missing execbuf2");
340 if (!device
->info
.has_llc
&&
341 anv_gem_get_param(fd
, I915_PARAM_MMAP_VERSION
) < 1) {
342 result
= vk_errorf(device
->instance
, device
,
343 VK_ERROR_INITIALIZATION_FAILED
,
344 "kernel missing wc mmap");
348 result
= anv_physical_device_init_heaps(device
, fd
);
349 if (result
!= VK_SUCCESS
)
352 device
->has_exec_async
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_ASYNC
);
353 device
->has_exec_fence
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE
);
354 device
->has_syncobj
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE_ARRAY
);
355 device
->has_syncobj_wait
= device
->has_syncobj
&&
356 anv_gem_supports_syncobj_wait(fd
);
358 bool swizzled
= anv_gem_get_bit6_swizzle(fd
, I915_TILING_X
);
360 /* GENs prior to 8 do not support EU/Subslice info */
361 if (device
->info
.gen
>= 8) {
362 device
->subslice_total
= anv_gem_get_param(fd
, I915_PARAM_SUBSLICE_TOTAL
);
363 device
->eu_total
= anv_gem_get_param(fd
, I915_PARAM_EU_TOTAL
);
365 /* Without this information, we cannot get the right Braswell
366 * brandstrings, and we have to use conservative numbers for GPGPU on
367 * many platforms, but otherwise, things will just work.
369 if (device
->subslice_total
< 1 || device
->eu_total
< 1) {
370 intel_logw("Kernel 4.1 required to properly query GPU properties");
372 } else if (device
->info
.gen
== 7) {
373 device
->subslice_total
= 1 << (device
->info
.gt
- 1);
376 if (device
->info
.is_cherryview
&&
377 device
->subslice_total
> 0 && device
->eu_total
> 0) {
378 /* Logical CS threads = EUs per subslice * num threads per EU */
379 uint32_t max_cs_threads
=
380 device
->eu_total
/ device
->subslice_total
* device
->info
.num_thread_per_eu
;
382 /* Fuse configurations may give more threads than expected, never less. */
383 if (max_cs_threads
> device
->info
.max_cs_threads
)
384 device
->info
.max_cs_threads
= max_cs_threads
;
387 device
->compiler
= brw_compiler_create(NULL
, &device
->info
);
388 if (device
->compiler
== NULL
) {
389 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
392 device
->compiler
->shader_debug_log
= compiler_debug_log
;
393 device
->compiler
->shader_perf_log
= compiler_perf_log
;
394 device
->compiler
->supports_pull_constants
= false;
396 isl_device_init(&device
->isl_dev
, &device
->info
, swizzled
);
398 result
= anv_physical_device_init_uuids(device
);
399 if (result
!= VK_SUCCESS
)
402 result
= anv_init_wsi(device
);
403 if (result
!= VK_SUCCESS
) {
404 ralloc_free(device
->compiler
);
408 device
->local_fd
= fd
;
417 anv_physical_device_finish(struct anv_physical_device
*device
)
419 anv_finish_wsi(device
);
420 ralloc_free(device
->compiler
);
421 close(device
->local_fd
);
425 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
426 VkSystemAllocationScope allocationScope
)
432 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
433 size_t align
, VkSystemAllocationScope allocationScope
)
435 return realloc(pOriginal
, size
);
439 default_free_func(void *pUserData
, void *pMemory
)
444 static const VkAllocationCallbacks default_alloc
= {
446 .pfnAllocation
= default_alloc_func
,
447 .pfnReallocation
= default_realloc_func
,
448 .pfnFree
= default_free_func
,
451 VkResult
anv_CreateInstance(
452 const VkInstanceCreateInfo
* pCreateInfo
,
453 const VkAllocationCallbacks
* pAllocator
,
454 VkInstance
* pInstance
)
456 struct anv_instance
*instance
;
458 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
460 /* Check if user passed a debug report callback to be used during
461 * Create/Destroy of instance.
463 const VkDebugReportCallbackCreateInfoEXT
*ctor_cb
=
464 vk_find_struct_const(pCreateInfo
->pNext
,
465 DEBUG_REPORT_CALLBACK_CREATE_INFO_EXT
);
467 uint32_t client_version
;
468 if (pCreateInfo
->pApplicationInfo
&&
469 pCreateInfo
->pApplicationInfo
->apiVersion
!= 0) {
470 client_version
= pCreateInfo
->pApplicationInfo
->apiVersion
;
472 client_version
= VK_MAKE_VERSION(1, 0, 0);
475 if (VK_MAKE_VERSION(1, 0, 0) > client_version
||
476 client_version
> VK_MAKE_VERSION(1, 0, 0xfff)) {
478 if (ctor_cb
&& ctor_cb
->flags
& VK_DEBUG_REPORT_ERROR_BIT_EXT
)
479 ctor_cb
->pfnCallback(VK_DEBUG_REPORT_ERROR_BIT_EXT
,
480 VK_DEBUG_REPORT_OBJECT_TYPE_INSTANCE_EXT
,
481 VK_NULL_HANDLE
, /* No handle available yet. */
485 "incompatible driver version",
488 return vk_errorf(NULL
, NULL
, VK_ERROR_INCOMPATIBLE_DRIVER
,
489 "Client requested version %d.%d.%d",
490 VK_VERSION_MAJOR(client_version
),
491 VK_VERSION_MINOR(client_version
),
492 VK_VERSION_PATCH(client_version
));
495 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
496 const char *ext_name
= pCreateInfo
->ppEnabledExtensionNames
[i
];
497 if (!anv_instance_extension_supported(ext_name
))
498 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
501 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
502 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
504 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
506 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
509 instance
->alloc
= *pAllocator
;
511 instance
->alloc
= default_alloc
;
513 instance
->apiVersion
= client_version
;
514 instance
->physicalDeviceCount
= -1;
516 if (pthread_mutex_init(&instance
->callbacks_mutex
, NULL
) != 0) {
517 vk_free2(&default_alloc
, pAllocator
, instance
);
518 return vk_error(VK_ERROR_INITIALIZATION_FAILED
);
521 list_inithead(&instance
->callbacks
);
523 /* Store report debug callback to be used during DestroyInstance. */
525 instance
->destroy_debug_cb
.flags
= ctor_cb
->flags
;
526 instance
->destroy_debug_cb
.callback
= ctor_cb
->pfnCallback
;
527 instance
->destroy_debug_cb
.data
= ctor_cb
->pUserData
;
532 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
534 *pInstance
= anv_instance_to_handle(instance
);
539 void anv_DestroyInstance(
540 VkInstance _instance
,
541 const VkAllocationCallbacks
* pAllocator
)
543 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
548 if (instance
->physicalDeviceCount
> 0) {
549 /* We support at most one physical device. */
550 assert(instance
->physicalDeviceCount
== 1);
551 anv_physical_device_finish(&instance
->physicalDevice
);
554 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
556 pthread_mutex_destroy(&instance
->callbacks_mutex
);
560 vk_free(&instance
->alloc
, instance
);
564 anv_enumerate_devices(struct anv_instance
*instance
)
566 /* TODO: Check for more devices ? */
567 drmDevicePtr devices
[8];
568 VkResult result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
571 instance
->physicalDeviceCount
= 0;
573 max_devices
= drmGetDevices2(0, devices
, ARRAY_SIZE(devices
));
575 return VK_ERROR_INCOMPATIBLE_DRIVER
;
577 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
578 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
579 devices
[i
]->bustype
== DRM_BUS_PCI
&&
580 devices
[i
]->deviceinfo
.pci
->vendor_id
== 0x8086) {
582 result
= anv_physical_device_init(&instance
->physicalDevice
,
584 devices
[i
]->nodes
[DRM_NODE_RENDER
]);
585 if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
589 drmFreeDevices(devices
, max_devices
);
591 if (result
== VK_SUCCESS
)
592 instance
->physicalDeviceCount
= 1;
598 VkResult
anv_EnumeratePhysicalDevices(
599 VkInstance _instance
,
600 uint32_t* pPhysicalDeviceCount
,
601 VkPhysicalDevice
* pPhysicalDevices
)
603 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
604 VK_OUTARRAY_MAKE(out
, pPhysicalDevices
, pPhysicalDeviceCount
);
607 if (instance
->physicalDeviceCount
< 0) {
608 result
= anv_enumerate_devices(instance
);
609 if (result
!= VK_SUCCESS
&&
610 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
614 if (instance
->physicalDeviceCount
> 0) {
615 assert(instance
->physicalDeviceCount
== 1);
616 vk_outarray_append(&out
, i
) {
617 *i
= anv_physical_device_to_handle(&instance
->physicalDevice
);
621 return vk_outarray_status(&out
);
624 void anv_GetPhysicalDeviceFeatures(
625 VkPhysicalDevice physicalDevice
,
626 VkPhysicalDeviceFeatures
* pFeatures
)
628 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
630 *pFeatures
= (VkPhysicalDeviceFeatures
) {
631 .robustBufferAccess
= true,
632 .fullDrawIndexUint32
= true,
633 .imageCubeArray
= true,
634 .independentBlend
= true,
635 .geometryShader
= true,
636 .tessellationShader
= true,
637 .sampleRateShading
= true,
638 .dualSrcBlend
= true,
640 .multiDrawIndirect
= true,
641 .drawIndirectFirstInstance
= true,
643 .depthBiasClamp
= true,
644 .fillModeNonSolid
= true,
645 .depthBounds
= false,
649 .multiViewport
= true,
650 .samplerAnisotropy
= true,
651 .textureCompressionETC2
= pdevice
->info
.gen
>= 8 ||
652 pdevice
->info
.is_baytrail
,
653 .textureCompressionASTC_LDR
= pdevice
->info
.gen
>= 9, /* FINISHME CHV */
654 .textureCompressionBC
= true,
655 .occlusionQueryPrecise
= true,
656 .pipelineStatisticsQuery
= true,
657 .fragmentStoresAndAtomics
= true,
658 .shaderTessellationAndGeometryPointSize
= true,
659 .shaderImageGatherExtended
= true,
660 .shaderStorageImageExtendedFormats
= true,
661 .shaderStorageImageMultisample
= false,
662 .shaderStorageImageReadWithoutFormat
= false,
663 .shaderStorageImageWriteWithoutFormat
= true,
664 .shaderUniformBufferArrayDynamicIndexing
= true,
665 .shaderSampledImageArrayDynamicIndexing
= true,
666 .shaderStorageBufferArrayDynamicIndexing
= true,
667 .shaderStorageImageArrayDynamicIndexing
= true,
668 .shaderClipDistance
= true,
669 .shaderCullDistance
= true,
670 .shaderFloat64
= pdevice
->info
.gen
>= 8,
671 .shaderInt64
= pdevice
->info
.gen
>= 8,
672 .shaderInt16
= false,
673 .shaderResourceMinLod
= false,
674 .variableMultisampleRate
= false,
675 .inheritedQueries
= true,
678 /* We can't do image stores in vec4 shaders */
679 pFeatures
->vertexPipelineStoresAndAtomics
=
680 pdevice
->compiler
->scalar_stage
[MESA_SHADER_VERTEX
] &&
681 pdevice
->compiler
->scalar_stage
[MESA_SHADER_GEOMETRY
];
684 void anv_GetPhysicalDeviceFeatures2KHR(
685 VkPhysicalDevice physicalDevice
,
686 VkPhysicalDeviceFeatures2KHR
* pFeatures
)
688 anv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
690 vk_foreach_struct(ext
, pFeatures
->pNext
) {
691 switch (ext
->sType
) {
692 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES_KHX
: {
693 VkPhysicalDeviceMultiviewFeaturesKHX
*features
=
694 (VkPhysicalDeviceMultiviewFeaturesKHX
*)ext
;
695 features
->multiview
= true;
696 features
->multiviewGeometryShader
= true;
697 features
->multiviewTessellationShader
= true;
701 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTER_FEATURES_KHR
: {
702 VkPhysicalDeviceVariablePointerFeaturesKHR
*features
= (void *)ext
;
703 features
->variablePointersStorageBuffer
= true;
704 features
->variablePointers
= false;
708 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES_KHR
: {
709 VkPhysicalDeviceSamplerYcbcrConversionFeaturesKHR
*features
=
710 (VkPhysicalDeviceSamplerYcbcrConversionFeaturesKHR
*) ext
;
711 features
->samplerYcbcrConversion
= true;
716 anv_debug_ignored_stype(ext
->sType
);
722 void anv_GetPhysicalDeviceProperties(
723 VkPhysicalDevice physicalDevice
,
724 VkPhysicalDeviceProperties
* pProperties
)
726 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
727 const struct gen_device_info
*devinfo
= &pdevice
->info
;
729 /* See assertions made when programming the buffer surface state. */
730 const uint32_t max_raw_buffer_sz
= devinfo
->gen
>= 7 ?
731 (1ul << 30) : (1ul << 27);
733 const uint32_t max_samplers
= (devinfo
->gen
>= 8 || devinfo
->is_haswell
) ?
736 VkSampleCountFlags sample_counts
=
737 isl_device_get_sample_counts(&pdevice
->isl_dev
);
739 VkPhysicalDeviceLimits limits
= {
740 .maxImageDimension1D
= (1 << 14),
741 .maxImageDimension2D
= (1 << 14),
742 .maxImageDimension3D
= (1 << 11),
743 .maxImageDimensionCube
= (1 << 14),
744 .maxImageArrayLayers
= (1 << 11),
745 .maxTexelBufferElements
= 128 * 1024 * 1024,
746 .maxUniformBufferRange
= (1ul << 27),
747 .maxStorageBufferRange
= max_raw_buffer_sz
,
748 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
749 .maxMemoryAllocationCount
= UINT32_MAX
,
750 .maxSamplerAllocationCount
= 64 * 1024,
751 .bufferImageGranularity
= 64, /* A cache line */
752 .sparseAddressSpaceSize
= 0,
753 .maxBoundDescriptorSets
= MAX_SETS
,
754 .maxPerStageDescriptorSamplers
= max_samplers
,
755 .maxPerStageDescriptorUniformBuffers
= 64,
756 .maxPerStageDescriptorStorageBuffers
= 64,
757 .maxPerStageDescriptorSampledImages
= max_samplers
,
758 .maxPerStageDescriptorStorageImages
= 64,
759 .maxPerStageDescriptorInputAttachments
= 64,
760 .maxPerStageResources
= 250,
761 .maxDescriptorSetSamplers
= 256,
762 .maxDescriptorSetUniformBuffers
= 256,
763 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
764 .maxDescriptorSetStorageBuffers
= 256,
765 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
766 .maxDescriptorSetSampledImages
= 256,
767 .maxDescriptorSetStorageImages
= 256,
768 .maxDescriptorSetInputAttachments
= 256,
769 .maxVertexInputAttributes
= MAX_VBS
,
770 .maxVertexInputBindings
= MAX_VBS
,
771 .maxVertexInputAttributeOffset
= 2047,
772 .maxVertexInputBindingStride
= 2048,
773 .maxVertexOutputComponents
= 128,
774 .maxTessellationGenerationLevel
= 64,
775 .maxTessellationPatchSize
= 32,
776 .maxTessellationControlPerVertexInputComponents
= 128,
777 .maxTessellationControlPerVertexOutputComponents
= 128,
778 .maxTessellationControlPerPatchOutputComponents
= 128,
779 .maxTessellationControlTotalOutputComponents
= 2048,
780 .maxTessellationEvaluationInputComponents
= 128,
781 .maxTessellationEvaluationOutputComponents
= 128,
782 .maxGeometryShaderInvocations
= 32,
783 .maxGeometryInputComponents
= 64,
784 .maxGeometryOutputComponents
= 128,
785 .maxGeometryOutputVertices
= 256,
786 .maxGeometryTotalOutputComponents
= 1024,
787 .maxFragmentInputComponents
= 128,
788 .maxFragmentOutputAttachments
= 8,
789 .maxFragmentDualSrcAttachments
= 1,
790 .maxFragmentCombinedOutputResources
= 8,
791 .maxComputeSharedMemorySize
= 32768,
792 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
793 .maxComputeWorkGroupInvocations
= 16 * devinfo
->max_cs_threads
,
794 .maxComputeWorkGroupSize
= {
795 16 * devinfo
->max_cs_threads
,
796 16 * devinfo
->max_cs_threads
,
797 16 * devinfo
->max_cs_threads
,
799 .subPixelPrecisionBits
= 4 /* FIXME */,
800 .subTexelPrecisionBits
= 4 /* FIXME */,
801 .mipmapPrecisionBits
= 4 /* FIXME */,
802 .maxDrawIndexedIndexValue
= UINT32_MAX
,
803 .maxDrawIndirectCount
= UINT32_MAX
,
804 .maxSamplerLodBias
= 16,
805 .maxSamplerAnisotropy
= 16,
806 .maxViewports
= MAX_VIEWPORTS
,
807 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
808 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
809 .viewportSubPixelBits
= 13, /* We take a float? */
810 .minMemoryMapAlignment
= 4096, /* A page */
811 .minTexelBufferOffsetAlignment
= 1,
812 .minUniformBufferOffsetAlignment
= 16,
813 .minStorageBufferOffsetAlignment
= 4,
814 .minTexelOffset
= -8,
816 .minTexelGatherOffset
= -32,
817 .maxTexelGatherOffset
= 31,
818 .minInterpolationOffset
= -0.5,
819 .maxInterpolationOffset
= 0.4375,
820 .subPixelInterpolationOffsetBits
= 4,
821 .maxFramebufferWidth
= (1 << 14),
822 .maxFramebufferHeight
= (1 << 14),
823 .maxFramebufferLayers
= (1 << 11),
824 .framebufferColorSampleCounts
= sample_counts
,
825 .framebufferDepthSampleCounts
= sample_counts
,
826 .framebufferStencilSampleCounts
= sample_counts
,
827 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
828 .maxColorAttachments
= MAX_RTS
,
829 .sampledImageColorSampleCounts
= sample_counts
,
830 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
831 .sampledImageDepthSampleCounts
= sample_counts
,
832 .sampledImageStencilSampleCounts
= sample_counts
,
833 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
834 .maxSampleMaskWords
= 1,
835 .timestampComputeAndGraphics
= false,
836 .timestampPeriod
= 1000000000.0 / devinfo
->timestamp_frequency
,
837 .maxClipDistances
= 8,
838 .maxCullDistances
= 8,
839 .maxCombinedClipAndCullDistances
= 8,
840 .discreteQueuePriorities
= 1,
841 .pointSizeRange
= { 0.125, 255.875 },
842 .lineWidthRange
= { 0.0, 7.9921875 },
843 .pointSizeGranularity
= (1.0 / 8.0),
844 .lineWidthGranularity
= (1.0 / 128.0),
845 .strictLines
= false, /* FINISHME */
846 .standardSampleLocations
= true,
847 .optimalBufferCopyOffsetAlignment
= 128,
848 .optimalBufferCopyRowPitchAlignment
= 128,
849 .nonCoherentAtomSize
= 64,
852 *pProperties
= (VkPhysicalDeviceProperties
) {
853 .apiVersion
= anv_physical_device_api_version(pdevice
),
854 .driverVersion
= vk_get_driver_version(),
856 .deviceID
= pdevice
->chipset_id
,
857 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
859 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
862 snprintf(pProperties
->deviceName
, sizeof(pProperties
->deviceName
),
863 "%s", pdevice
->name
);
864 memcpy(pProperties
->pipelineCacheUUID
,
865 pdevice
->pipeline_cache_uuid
, VK_UUID_SIZE
);
868 void anv_GetPhysicalDeviceProperties2KHR(
869 VkPhysicalDevice physicalDevice
,
870 VkPhysicalDeviceProperties2KHR
* pProperties
)
872 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
874 anv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
876 vk_foreach_struct(ext
, pProperties
->pNext
) {
877 switch (ext
->sType
) {
878 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
879 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
880 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
882 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
886 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES_KHR
: {
887 VkPhysicalDeviceIDPropertiesKHR
*id_props
=
888 (VkPhysicalDeviceIDPropertiesKHR
*)ext
;
889 memcpy(id_props
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
890 memcpy(id_props
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
891 /* The LUID is for Windows. */
892 id_props
->deviceLUIDValid
= false;
896 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES_KHX
: {
897 VkPhysicalDeviceMultiviewPropertiesKHX
*properties
=
898 (VkPhysicalDeviceMultiviewPropertiesKHX
*)ext
;
899 properties
->maxMultiviewViewCount
= 16;
900 properties
->maxMultiviewInstanceIndex
= UINT32_MAX
/ 16;
904 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES_KHR
: {
905 VkPhysicalDevicePointClippingPropertiesKHR
*properties
=
906 (VkPhysicalDevicePointClippingPropertiesKHR
*) ext
;
907 properties
->pointClippingBehavior
= VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES_KHR
;
908 anv_finishme("Implement pop-free point clipping");
913 anv_debug_ignored_stype(ext
->sType
);
919 /* We support exactly one queue family. */
920 static const VkQueueFamilyProperties
921 anv_queue_family_properties
= {
922 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
923 VK_QUEUE_COMPUTE_BIT
|
924 VK_QUEUE_TRANSFER_BIT
,
926 .timestampValidBits
= 36, /* XXX: Real value here */
927 .minImageTransferGranularity
= { 1, 1, 1 },
930 void anv_GetPhysicalDeviceQueueFamilyProperties(
931 VkPhysicalDevice physicalDevice
,
933 VkQueueFamilyProperties
* pQueueFamilyProperties
)
935 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pCount
);
937 vk_outarray_append(&out
, p
) {
938 *p
= anv_queue_family_properties
;
942 void anv_GetPhysicalDeviceQueueFamilyProperties2KHR(
943 VkPhysicalDevice physicalDevice
,
944 uint32_t* pQueueFamilyPropertyCount
,
945 VkQueueFamilyProperties2KHR
* pQueueFamilyProperties
)
948 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pQueueFamilyPropertyCount
);
950 vk_outarray_append(&out
, p
) {
951 p
->queueFamilyProperties
= anv_queue_family_properties
;
953 vk_foreach_struct(s
, p
->pNext
) {
954 anv_debug_ignored_stype(s
->sType
);
959 void anv_GetPhysicalDeviceMemoryProperties(
960 VkPhysicalDevice physicalDevice
,
961 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
963 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
965 pMemoryProperties
->memoryTypeCount
= physical_device
->memory
.type_count
;
966 for (uint32_t i
= 0; i
< physical_device
->memory
.type_count
; i
++) {
967 pMemoryProperties
->memoryTypes
[i
] = (VkMemoryType
) {
968 .propertyFlags
= physical_device
->memory
.types
[i
].propertyFlags
,
969 .heapIndex
= physical_device
->memory
.types
[i
].heapIndex
,
973 pMemoryProperties
->memoryHeapCount
= physical_device
->memory
.heap_count
;
974 for (uint32_t i
= 0; i
< physical_device
->memory
.heap_count
; i
++) {
975 pMemoryProperties
->memoryHeaps
[i
] = (VkMemoryHeap
) {
976 .size
= physical_device
->memory
.heaps
[i
].size
,
977 .flags
= physical_device
->memory
.heaps
[i
].flags
,
982 void anv_GetPhysicalDeviceMemoryProperties2KHR(
983 VkPhysicalDevice physicalDevice
,
984 VkPhysicalDeviceMemoryProperties2KHR
* pMemoryProperties
)
986 anv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
987 &pMemoryProperties
->memoryProperties
);
989 vk_foreach_struct(ext
, pMemoryProperties
->pNext
) {
990 switch (ext
->sType
) {
992 anv_debug_ignored_stype(ext
->sType
);
998 PFN_vkVoidFunction
anv_GetInstanceProcAddr(
1002 return anv_lookup_entrypoint(NULL
, pName
);
1005 /* With version 1+ of the loader interface the ICD should expose
1006 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
1009 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1010 VkInstance instance
,
1014 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1015 VkInstance instance
,
1018 return anv_GetInstanceProcAddr(instance
, pName
);
1021 PFN_vkVoidFunction
anv_GetDeviceProcAddr(
1025 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1026 return anv_lookup_entrypoint(&device
->info
, pName
);
1030 anv_queue_init(struct anv_device
*device
, struct anv_queue
*queue
)
1032 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1033 queue
->device
= device
;
1034 queue
->pool
= &device
->surface_state_pool
;
1038 anv_queue_finish(struct anv_queue
*queue
)
1042 static struct anv_state
1043 anv_state_pool_emit_data(struct anv_state_pool
*pool
, size_t size
, size_t align
, const void *p
)
1045 struct anv_state state
;
1047 state
= anv_state_pool_alloc(pool
, size
, align
);
1048 memcpy(state
.map
, p
, size
);
1050 anv_state_flush(pool
->block_pool
.device
, state
);
1055 struct gen8_border_color
{
1060 /* Pad out to 64 bytes */
1065 anv_device_init_border_colors(struct anv_device
*device
)
1067 static const struct gen8_border_color border_colors
[] = {
1068 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
1069 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
1070 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
1071 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
1072 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
1073 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
1076 device
->border_colors
= anv_state_pool_emit_data(&device
->dynamic_state_pool
,
1077 sizeof(border_colors
), 64,
1082 anv_device_init_trivial_batch(struct anv_device
*device
)
1084 anv_bo_init_new(&device
->trivial_batch_bo
, device
, 4096);
1086 if (device
->instance
->physicalDevice
.has_exec_async
)
1087 device
->trivial_batch_bo
.flags
|= EXEC_OBJECT_ASYNC
;
1089 void *map
= anv_gem_mmap(device
, device
->trivial_batch_bo
.gem_handle
,
1092 struct anv_batch batch
= {
1098 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
1099 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
1101 if (!device
->info
.has_llc
)
1102 gen_clflush_range(map
, batch
.next
- map
);
1104 anv_gem_munmap(map
, device
->trivial_batch_bo
.size
);
1107 VkResult
anv_CreateDevice(
1108 VkPhysicalDevice physicalDevice
,
1109 const VkDeviceCreateInfo
* pCreateInfo
,
1110 const VkAllocationCallbacks
* pAllocator
,
1113 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
1115 struct anv_device
*device
;
1117 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO
);
1119 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
1120 const char *ext_name
= pCreateInfo
->ppEnabledExtensionNames
[i
];
1121 if (!anv_physical_device_extension_supported(physical_device
, ext_name
))
1122 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
1125 /* Check enabled features */
1126 if (pCreateInfo
->pEnabledFeatures
) {
1127 VkPhysicalDeviceFeatures supported_features
;
1128 anv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
1129 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
1130 VkBool32
*enabled_feature
= (VkBool32
*)pCreateInfo
->pEnabledFeatures
;
1131 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
1132 for (uint32_t i
= 0; i
< num_features
; i
++) {
1133 if (enabled_feature
[i
] && !supported_feature
[i
])
1134 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
1138 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
1140 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
1142 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1144 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1145 device
->instance
= physical_device
->instance
;
1146 device
->chipset_id
= physical_device
->chipset_id
;
1147 device
->lost
= false;
1150 device
->alloc
= *pAllocator
;
1152 device
->alloc
= physical_device
->instance
->alloc
;
1154 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
1155 device
->fd
= open(physical_device
->path
, O_RDWR
| O_CLOEXEC
);
1156 if (device
->fd
== -1) {
1157 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1161 device
->context_id
= anv_gem_create_context(device
);
1162 if (device
->context_id
== -1) {
1163 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1167 device
->info
= physical_device
->info
;
1168 device
->isl_dev
= physical_device
->isl_dev
;
1170 /* On Broadwell and later, we can use batch chaining to more efficiently
1171 * implement growing command buffers. Prior to Haswell, the kernel
1172 * command parser gets in the way and we have to fall back to growing
1175 device
->can_chain_batches
= device
->info
.gen
>= 8;
1177 device
->robust_buffer_access
= pCreateInfo
->pEnabledFeatures
&&
1178 pCreateInfo
->pEnabledFeatures
->robustBufferAccess
;
1180 if (pthread_mutex_init(&device
->mutex
, NULL
) != 0) {
1181 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1182 goto fail_context_id
;
1185 pthread_condattr_t condattr
;
1186 if (pthread_condattr_init(&condattr
) != 0) {
1187 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1190 if (pthread_condattr_setclock(&condattr
, CLOCK_MONOTONIC
) != 0) {
1191 pthread_condattr_destroy(&condattr
);
1192 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1195 if (pthread_cond_init(&device
->queue_submit
, NULL
) != 0) {
1196 pthread_condattr_destroy(&condattr
);
1197 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1200 pthread_condattr_destroy(&condattr
);
1202 anv_bo_pool_init(&device
->batch_bo_pool
, device
);
1204 result
= anv_bo_cache_init(&device
->bo_cache
);
1205 if (result
!= VK_SUCCESS
)
1206 goto fail_batch_bo_pool
;
1208 result
= anv_state_pool_init(&device
->dynamic_state_pool
, device
, 16384);
1209 if (result
!= VK_SUCCESS
)
1212 result
= anv_state_pool_init(&device
->instruction_state_pool
, device
, 16384);
1213 if (result
!= VK_SUCCESS
)
1214 goto fail_dynamic_state_pool
;
1216 result
= anv_state_pool_init(&device
->surface_state_pool
, device
, 4096);
1217 if (result
!= VK_SUCCESS
)
1218 goto fail_instruction_state_pool
;
1220 result
= anv_bo_init_new(&device
->workaround_bo
, device
, 1024);
1221 if (result
!= VK_SUCCESS
)
1222 goto fail_surface_state_pool
;
1224 anv_device_init_trivial_batch(device
);
1226 anv_scratch_pool_init(device
, &device
->scratch_pool
);
1228 anv_queue_init(device
, &device
->queue
);
1230 switch (device
->info
.gen
) {
1232 if (!device
->info
.is_haswell
)
1233 result
= gen7_init_device_state(device
);
1235 result
= gen75_init_device_state(device
);
1238 result
= gen8_init_device_state(device
);
1241 result
= gen9_init_device_state(device
);
1244 result
= gen10_init_device_state(device
);
1247 /* Shouldn't get here as we don't create physical devices for any other
1249 unreachable("unhandled gen");
1251 if (result
!= VK_SUCCESS
)
1252 goto fail_workaround_bo
;
1254 anv_device_init_blorp(device
);
1256 anv_device_init_border_colors(device
);
1258 *pDevice
= anv_device_to_handle(device
);
1263 anv_queue_finish(&device
->queue
);
1264 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
1265 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
1266 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
1267 fail_surface_state_pool
:
1268 anv_state_pool_finish(&device
->surface_state_pool
);
1269 fail_instruction_state_pool
:
1270 anv_state_pool_finish(&device
->instruction_state_pool
);
1271 fail_dynamic_state_pool
:
1272 anv_state_pool_finish(&device
->dynamic_state_pool
);
1274 anv_bo_cache_finish(&device
->bo_cache
);
1276 anv_bo_pool_finish(&device
->batch_bo_pool
);
1277 pthread_cond_destroy(&device
->queue_submit
);
1279 pthread_mutex_destroy(&device
->mutex
);
1281 anv_gem_destroy_context(device
, device
->context_id
);
1285 vk_free(&device
->alloc
, device
);
1290 void anv_DestroyDevice(
1292 const VkAllocationCallbacks
* pAllocator
)
1294 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1299 anv_device_finish_blorp(device
);
1301 anv_queue_finish(&device
->queue
);
1303 #ifdef HAVE_VALGRIND
1304 /* We only need to free these to prevent valgrind errors. The backing
1305 * BO will go away in a couple of lines so we don't actually leak.
1307 anv_state_pool_free(&device
->dynamic_state_pool
, device
->border_colors
);
1310 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
1312 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
1313 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
1315 anv_gem_close(device
, device
->trivial_batch_bo
.gem_handle
);
1317 anv_state_pool_finish(&device
->surface_state_pool
);
1318 anv_state_pool_finish(&device
->instruction_state_pool
);
1319 anv_state_pool_finish(&device
->dynamic_state_pool
);
1321 anv_bo_cache_finish(&device
->bo_cache
);
1323 anv_bo_pool_finish(&device
->batch_bo_pool
);
1325 pthread_cond_destroy(&device
->queue_submit
);
1326 pthread_mutex_destroy(&device
->mutex
);
1328 anv_gem_destroy_context(device
, device
->context_id
);
1332 vk_free(&device
->alloc
, device
);
1335 VkResult
anv_EnumerateInstanceLayerProperties(
1336 uint32_t* pPropertyCount
,
1337 VkLayerProperties
* pProperties
)
1339 if (pProperties
== NULL
) {
1340 *pPropertyCount
= 0;
1344 /* None supported at this time */
1345 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1348 VkResult
anv_EnumerateDeviceLayerProperties(
1349 VkPhysicalDevice physicalDevice
,
1350 uint32_t* pPropertyCount
,
1351 VkLayerProperties
* pProperties
)
1353 if (pProperties
== NULL
) {
1354 *pPropertyCount
= 0;
1358 /* None supported at this time */
1359 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1362 void anv_GetDeviceQueue(
1364 uint32_t queueNodeIndex
,
1365 uint32_t queueIndex
,
1368 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1370 assert(queueIndex
== 0);
1372 *pQueue
= anv_queue_to_handle(&device
->queue
);
1376 anv_device_query_status(struct anv_device
*device
)
1378 /* This isn't likely as most of the callers of this function already check
1379 * for it. However, it doesn't hurt to check and it potentially lets us
1382 if (unlikely(device
->lost
))
1383 return VK_ERROR_DEVICE_LOST
;
1385 uint32_t active
, pending
;
1386 int ret
= anv_gem_gpu_get_reset_stats(device
, &active
, &pending
);
1388 /* We don't know the real error. */
1389 device
->lost
= true;
1390 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1391 "get_reset_stats failed: %m");
1395 device
->lost
= true;
1396 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1397 "GPU hung on one of our command buffers");
1398 } else if (pending
) {
1399 device
->lost
= true;
1400 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1401 "GPU hung with commands in-flight");
1408 anv_device_bo_busy(struct anv_device
*device
, struct anv_bo
*bo
)
1410 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
1411 * Other usages of the BO (such as on different hardware) will not be
1412 * flagged as "busy" by this ioctl. Use with care.
1414 int ret
= anv_gem_busy(device
, bo
->gem_handle
);
1416 return VK_NOT_READY
;
1417 } else if (ret
== -1) {
1418 /* We don't know the real error. */
1419 device
->lost
= true;
1420 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1421 "gem wait failed: %m");
1424 /* Query for device status after the busy call. If the BO we're checking
1425 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
1426 * client because it clearly doesn't have valid data. Yes, this most
1427 * likely means an ioctl, but we just did an ioctl to query the busy status
1428 * so it's no great loss.
1430 return anv_device_query_status(device
);
1434 anv_device_wait(struct anv_device
*device
, struct anv_bo
*bo
,
1437 int ret
= anv_gem_wait(device
, bo
->gem_handle
, &timeout
);
1438 if (ret
== -1 && errno
== ETIME
) {
1440 } else if (ret
== -1) {
1441 /* We don't know the real error. */
1442 device
->lost
= true;
1443 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1444 "gem wait failed: %m");
1447 /* Query for device status after the wait. If the BO we're waiting on got
1448 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
1449 * because it clearly doesn't have valid data. Yes, this most likely means
1450 * an ioctl, but we just did an ioctl to wait so it's no great loss.
1452 return anv_device_query_status(device
);
1455 VkResult
anv_DeviceWaitIdle(
1458 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1459 if (unlikely(device
->lost
))
1460 return VK_ERROR_DEVICE_LOST
;
1462 struct anv_batch batch
;
1465 batch
.start
= batch
.next
= cmds
;
1466 batch
.end
= (void *) cmds
+ sizeof(cmds
);
1468 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
1469 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
1471 return anv_device_submit_simple_batch(device
, &batch
);
1475 anv_bo_init_new(struct anv_bo
*bo
, struct anv_device
*device
, uint64_t size
)
1477 uint32_t gem_handle
= anv_gem_create(device
, size
);
1479 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
1481 anv_bo_init(bo
, gem_handle
, size
);
1486 VkResult
anv_AllocateMemory(
1488 const VkMemoryAllocateInfo
* pAllocateInfo
,
1489 const VkAllocationCallbacks
* pAllocator
,
1490 VkDeviceMemory
* pMem
)
1492 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1493 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
1494 struct anv_device_memory
*mem
;
1495 VkResult result
= VK_SUCCESS
;
1497 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
1499 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
1500 assert(pAllocateInfo
->allocationSize
> 0);
1502 /* The kernel relocation API has a limitation of a 32-bit delta value
1503 * applied to the address before it is written which, in spite of it being
1504 * unsigned, is treated as signed . Because of the way that this maps to
1505 * the Vulkan API, we cannot handle an offset into a buffer that does not
1506 * fit into a signed 32 bits. The only mechanism we have for dealing with
1507 * this at the moment is to limit all VkDeviceMemory objects to a maximum
1508 * of 2GB each. The Vulkan spec allows us to do this:
1510 * "Some platforms may have a limit on the maximum size of a single
1511 * allocation. For example, certain systems may fail to create
1512 * allocations with a size greater than or equal to 4GB. Such a limit is
1513 * implementation-dependent, and if such a failure occurs then the error
1514 * VK_ERROR_OUT_OF_DEVICE_MEMORY should be returned."
1516 * We don't use vk_error here because it's not an error so much as an
1517 * indication to the application that the allocation is too large.
1519 if (pAllocateInfo
->allocationSize
> (1ull << 31))
1520 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
1522 /* FINISHME: Fail if allocation request exceeds heap size. */
1524 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
1525 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1527 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1529 assert(pAllocateInfo
->memoryTypeIndex
< pdevice
->memory
.type_count
);
1530 mem
->type
= &pdevice
->memory
.types
[pAllocateInfo
->memoryTypeIndex
];
1534 const VkImportMemoryFdInfoKHR
*fd_info
=
1535 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_FD_INFO_KHR
);
1537 /* The Vulkan spec permits handleType to be 0, in which case the struct is
1540 if (fd_info
&& fd_info
->handleType
) {
1541 /* At the moment, we only support the OPAQUE_FD memory type which is
1542 * just a GEM buffer.
1544 assert(fd_info
->handleType
==
1545 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
);
1547 result
= anv_bo_cache_import(device
, &device
->bo_cache
,
1548 fd_info
->fd
, &mem
->bo
);
1549 if (result
!= VK_SUCCESS
)
1552 VkDeviceSize aligned_alloc_size
=
1553 align_u64(pAllocateInfo
->allocationSize
, 4096);
1555 /* For security purposes, we reject importing the bo if it's smaller
1556 * than the requested allocation size. This prevents a malicious client
1557 * from passing a buffer to a trusted client, lying about the size, and
1558 * telling the trusted client to try and texture from an image that goes
1559 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
1560 * in the trusted client. The trusted client can protect itself against
1561 * this sort of attack but only if it can trust the buffer size.
1563 if (mem
->bo
->size
< aligned_alloc_size
) {
1564 result
= vk_errorf(device
->instance
, device
,
1565 VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
,
1566 "aligned allocationSize too large for "
1567 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR: "
1568 "%"PRIu64
"B > %"PRIu64
"B",
1569 aligned_alloc_size
, mem
->bo
->size
);
1570 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
1574 /* From the Vulkan spec:
1576 * "Importing memory from a file descriptor transfers ownership of
1577 * the file descriptor from the application to the Vulkan
1578 * implementation. The application must not perform any operations on
1579 * the file descriptor after a successful import."
1581 * If the import fails, we leave the file descriptor open.
1585 result
= anv_bo_cache_alloc(device
, &device
->bo_cache
,
1586 pAllocateInfo
->allocationSize
,
1588 if (result
!= VK_SUCCESS
)
1592 assert(mem
->type
->heapIndex
< pdevice
->memory
.heap_count
);
1593 if (pdevice
->memory
.heaps
[mem
->type
->heapIndex
].supports_48bit_addresses
)
1594 mem
->bo
->flags
|= EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
1596 if (pdevice
->has_exec_async
)
1597 mem
->bo
->flags
|= EXEC_OBJECT_ASYNC
;
1599 *pMem
= anv_device_memory_to_handle(mem
);
1604 vk_free2(&device
->alloc
, pAllocator
, mem
);
1609 VkResult
anv_GetMemoryFdKHR(
1611 const VkMemoryGetFdInfoKHR
* pGetFdInfo
,
1614 ANV_FROM_HANDLE(anv_device
, dev
, device_h
);
1615 ANV_FROM_HANDLE(anv_device_memory
, mem
, pGetFdInfo
->memory
);
1617 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
1619 /* We support only one handle type. */
1620 assert(pGetFdInfo
->handleType
==
1621 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
);
1623 return anv_bo_cache_export(dev
, &dev
->bo_cache
, mem
->bo
, pFd
);
1626 VkResult
anv_GetMemoryFdPropertiesKHR(
1628 VkExternalMemoryHandleTypeFlagBitsKHR handleType
,
1630 VkMemoryFdPropertiesKHR
* pMemoryFdProperties
)
1632 /* The valid usage section for this function says:
1634 * "handleType must not be one of the handle types defined as opaque."
1636 * Since we only handle opaque handles for now, there are no FD properties.
1638 return VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
;
1641 void anv_FreeMemory(
1643 VkDeviceMemory _mem
,
1644 const VkAllocationCallbacks
* pAllocator
)
1646 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1647 ANV_FROM_HANDLE(anv_device_memory
, mem
, _mem
);
1653 anv_UnmapMemory(_device
, _mem
);
1655 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
1657 vk_free2(&device
->alloc
, pAllocator
, mem
);
1660 VkResult
anv_MapMemory(
1662 VkDeviceMemory _memory
,
1663 VkDeviceSize offset
,
1665 VkMemoryMapFlags flags
,
1668 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1669 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
1676 if (size
== VK_WHOLE_SIZE
)
1677 size
= mem
->bo
->size
- offset
;
1679 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
1681 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
1682 * assert(size != 0);
1683 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
1684 * equal to the size of the memory minus offset
1687 assert(offset
+ size
<= mem
->bo
->size
);
1689 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
1690 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
1691 * at a time is valid. We could just mmap up front and return an offset
1692 * pointer here, but that may exhaust virtual memory on 32 bit
1695 uint32_t gem_flags
= 0;
1697 if (!device
->info
.has_llc
&&
1698 (mem
->type
->propertyFlags
& VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
))
1699 gem_flags
|= I915_MMAP_WC
;
1701 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
1702 uint64_t map_offset
= offset
& ~4095ull;
1703 assert(offset
>= map_offset
);
1704 uint64_t map_size
= (offset
+ size
) - map_offset
;
1706 /* Let's map whole pages */
1707 map_size
= align_u64(map_size
, 4096);
1709 void *map
= anv_gem_mmap(device
, mem
->bo
->gem_handle
,
1710 map_offset
, map_size
, gem_flags
);
1711 if (map
== MAP_FAILED
)
1712 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
1715 mem
->map_size
= map_size
;
1717 *ppData
= mem
->map
+ (offset
- map_offset
);
1722 void anv_UnmapMemory(
1724 VkDeviceMemory _memory
)
1726 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
1731 anv_gem_munmap(mem
->map
, mem
->map_size
);
1738 clflush_mapped_ranges(struct anv_device
*device
,
1740 const VkMappedMemoryRange
*ranges
)
1742 for (uint32_t i
= 0; i
< count
; i
++) {
1743 ANV_FROM_HANDLE(anv_device_memory
, mem
, ranges
[i
].memory
);
1744 if (ranges
[i
].offset
>= mem
->map_size
)
1747 gen_clflush_range(mem
->map
+ ranges
[i
].offset
,
1748 MIN2(ranges
[i
].size
, mem
->map_size
- ranges
[i
].offset
));
1752 VkResult
anv_FlushMappedMemoryRanges(
1754 uint32_t memoryRangeCount
,
1755 const VkMappedMemoryRange
* pMemoryRanges
)
1757 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1759 if (device
->info
.has_llc
)
1762 /* Make sure the writes we're flushing have landed. */
1763 __builtin_ia32_mfence();
1765 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
1770 VkResult
anv_InvalidateMappedMemoryRanges(
1772 uint32_t memoryRangeCount
,
1773 const VkMappedMemoryRange
* pMemoryRanges
)
1775 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1777 if (device
->info
.has_llc
)
1780 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
1782 /* Make sure no reads get moved up above the invalidate. */
1783 __builtin_ia32_mfence();
1788 void anv_GetBufferMemoryRequirements(
1791 VkMemoryRequirements
* pMemoryRequirements
)
1793 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
1794 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1795 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
1797 /* The Vulkan spec (git aaed022) says:
1799 * memoryTypeBits is a bitfield and contains one bit set for every
1800 * supported memory type for the resource. The bit `1<<i` is set if and
1801 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1802 * structure for the physical device is supported.
1804 uint32_t memory_types
= 0;
1805 for (uint32_t i
= 0; i
< pdevice
->memory
.type_count
; i
++) {
1806 uint32_t valid_usage
= pdevice
->memory
.types
[i
].valid_buffer_usage
;
1807 if ((valid_usage
& buffer
->usage
) == buffer
->usage
)
1808 memory_types
|= (1u << i
);
1811 pMemoryRequirements
->size
= buffer
->size
;
1812 pMemoryRequirements
->alignment
= 16;
1813 pMemoryRequirements
->memoryTypeBits
= memory_types
;
1816 void anv_GetBufferMemoryRequirements2KHR(
1818 const VkBufferMemoryRequirementsInfo2KHR
* pInfo
,
1819 VkMemoryRequirements2KHR
* pMemoryRequirements
)
1821 anv_GetBufferMemoryRequirements(_device
, pInfo
->buffer
,
1822 &pMemoryRequirements
->memoryRequirements
);
1824 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
1825 switch (ext
->sType
) {
1826 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS_KHR
: {
1827 VkMemoryDedicatedRequirementsKHR
*requirements
= (void *)ext
;
1828 requirements
->prefersDedicatedAllocation
= VK_FALSE
;
1829 requirements
->requiresDedicatedAllocation
= VK_FALSE
;
1834 anv_debug_ignored_stype(ext
->sType
);
1840 void anv_GetImageMemoryRequirements(
1843 VkMemoryRequirements
* pMemoryRequirements
)
1845 ANV_FROM_HANDLE(anv_image
, image
, _image
);
1846 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1847 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
1849 /* The Vulkan spec (git aaed022) says:
1851 * memoryTypeBits is a bitfield and contains one bit set for every
1852 * supported memory type for the resource. The bit `1<<i` is set if and
1853 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1854 * structure for the physical device is supported.
1856 * All types are currently supported for images.
1858 uint32_t memory_types
= (1ull << pdevice
->memory
.type_count
) - 1;
1860 pMemoryRequirements
->size
= image
->size
;
1861 pMemoryRequirements
->alignment
= image
->alignment
;
1862 pMemoryRequirements
->memoryTypeBits
= memory_types
;
1865 void anv_GetImageMemoryRequirements2KHR(
1867 const VkImageMemoryRequirementsInfo2KHR
* pInfo
,
1868 VkMemoryRequirements2KHR
* pMemoryRequirements
)
1870 anv_GetImageMemoryRequirements(_device
, pInfo
->image
,
1871 &pMemoryRequirements
->memoryRequirements
);
1873 vk_foreach_struct_const(ext
, pInfo
->pNext
) {
1874 switch (ext
->sType
) {
1875 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO_KHR
: {
1876 ANV_FROM_HANDLE(anv_image
, image
, pInfo
->image
);
1877 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1878 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
1879 const VkImagePlaneMemoryRequirementsInfoKHR
*plane_reqs
=
1880 (const VkImagePlaneMemoryRequirementsInfoKHR
*) ext
;
1881 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
,
1882 plane_reqs
->planeAspect
);
1884 assert(image
->planes
[plane
].offset
== 0);
1886 /* The Vulkan spec (git aaed022) says:
1888 * memoryTypeBits is a bitfield and contains one bit set for every
1889 * supported memory type for the resource. The bit `1<<i` is set
1890 * if and only if the memory type `i` in the
1891 * VkPhysicalDeviceMemoryProperties structure for the physical
1892 * device is supported.
1894 * All types are currently supported for images.
1896 pMemoryRequirements
->memoryRequirements
.memoryTypeBits
=
1897 (1ull << pdevice
->memory
.type_count
) - 1;
1899 pMemoryRequirements
->memoryRequirements
.size
= image
->planes
[plane
].size
;
1900 pMemoryRequirements
->memoryRequirements
.alignment
=
1901 image
->planes
[plane
].alignment
;
1906 anv_debug_ignored_stype(ext
->sType
);
1911 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
1912 switch (ext
->sType
) {
1913 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS_KHR
: {
1914 VkMemoryDedicatedRequirementsKHR
*requirements
= (void *)ext
;
1915 requirements
->prefersDedicatedAllocation
= VK_FALSE
;
1916 requirements
->requiresDedicatedAllocation
= VK_FALSE
;
1921 anv_debug_ignored_stype(ext
->sType
);
1927 void anv_GetImageSparseMemoryRequirements(
1930 uint32_t* pSparseMemoryRequirementCount
,
1931 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
1933 *pSparseMemoryRequirementCount
= 0;
1936 void anv_GetImageSparseMemoryRequirements2KHR(
1938 const VkImageSparseMemoryRequirementsInfo2KHR
* pInfo
,
1939 uint32_t* pSparseMemoryRequirementCount
,
1940 VkSparseImageMemoryRequirements2KHR
* pSparseMemoryRequirements
)
1942 *pSparseMemoryRequirementCount
= 0;
1945 void anv_GetDeviceMemoryCommitment(
1947 VkDeviceMemory memory
,
1948 VkDeviceSize
* pCommittedMemoryInBytes
)
1950 *pCommittedMemoryInBytes
= 0;
1954 anv_bind_buffer_memory(const VkBindBufferMemoryInfoKHR
*pBindInfo
)
1956 ANV_FROM_HANDLE(anv_device_memory
, mem
, pBindInfo
->memory
);
1957 ANV_FROM_HANDLE(anv_buffer
, buffer
, pBindInfo
->buffer
);
1959 assert(pBindInfo
->sType
== VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO_KHR
);
1962 assert((buffer
->usage
& mem
->type
->valid_buffer_usage
) == buffer
->usage
);
1963 buffer
->bo
= mem
->bo
;
1964 buffer
->offset
= pBindInfo
->memoryOffset
;
1971 VkResult
anv_BindBufferMemory(
1974 VkDeviceMemory memory
,
1975 VkDeviceSize memoryOffset
)
1977 anv_bind_buffer_memory(
1978 &(VkBindBufferMemoryInfoKHR
) {
1979 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO_KHR
,
1982 .memoryOffset
= memoryOffset
,
1988 VkResult
anv_BindBufferMemory2KHR(
1990 uint32_t bindInfoCount
,
1991 const VkBindBufferMemoryInfoKHR
* pBindInfos
)
1993 for (uint32_t i
= 0; i
< bindInfoCount
; i
++)
1994 anv_bind_buffer_memory(&pBindInfos
[i
]);
1999 VkResult
anv_QueueBindSparse(
2001 uint32_t bindInfoCount
,
2002 const VkBindSparseInfo
* pBindInfo
,
2005 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
2006 if (unlikely(queue
->device
->lost
))
2007 return VK_ERROR_DEVICE_LOST
;
2009 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
2014 VkResult
anv_CreateEvent(
2016 const VkEventCreateInfo
* pCreateInfo
,
2017 const VkAllocationCallbacks
* pAllocator
,
2020 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2021 struct anv_state state
;
2022 struct anv_event
*event
;
2024 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_EVENT_CREATE_INFO
);
2026 state
= anv_state_pool_alloc(&device
->dynamic_state_pool
,
2029 event
->state
= state
;
2030 event
->semaphore
= VK_EVENT_RESET
;
2032 if (!device
->info
.has_llc
) {
2033 /* Make sure the writes we're flushing have landed. */
2034 __builtin_ia32_mfence();
2035 __builtin_ia32_clflush(event
);
2038 *pEvent
= anv_event_to_handle(event
);
2043 void anv_DestroyEvent(
2046 const VkAllocationCallbacks
* pAllocator
)
2048 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2049 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2054 anv_state_pool_free(&device
->dynamic_state_pool
, event
->state
);
2057 VkResult
anv_GetEventStatus(
2061 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2062 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2064 if (unlikely(device
->lost
))
2065 return VK_ERROR_DEVICE_LOST
;
2067 if (!device
->info
.has_llc
) {
2068 /* Invalidate read cache before reading event written by GPU. */
2069 __builtin_ia32_clflush(event
);
2070 __builtin_ia32_mfence();
2074 return event
->semaphore
;
2077 VkResult
anv_SetEvent(
2081 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2082 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2084 event
->semaphore
= VK_EVENT_SET
;
2086 if (!device
->info
.has_llc
) {
2087 /* Make sure the writes we're flushing have landed. */
2088 __builtin_ia32_mfence();
2089 __builtin_ia32_clflush(event
);
2095 VkResult
anv_ResetEvent(
2099 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2100 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2102 event
->semaphore
= VK_EVENT_RESET
;
2104 if (!device
->info
.has_llc
) {
2105 /* Make sure the writes we're flushing have landed. */
2106 __builtin_ia32_mfence();
2107 __builtin_ia32_clflush(event
);
2115 VkResult
anv_CreateBuffer(
2117 const VkBufferCreateInfo
* pCreateInfo
,
2118 const VkAllocationCallbacks
* pAllocator
,
2121 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2122 struct anv_buffer
*buffer
;
2124 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
2126 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
2127 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2129 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2131 buffer
->size
= pCreateInfo
->size
;
2132 buffer
->usage
= pCreateInfo
->usage
;
2136 *pBuffer
= anv_buffer_to_handle(buffer
);
2141 void anv_DestroyBuffer(
2144 const VkAllocationCallbacks
* pAllocator
)
2146 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2147 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
2152 vk_free2(&device
->alloc
, pAllocator
, buffer
);
2156 anv_fill_buffer_surface_state(struct anv_device
*device
, struct anv_state state
,
2157 enum isl_format format
,
2158 uint32_t offset
, uint32_t range
, uint32_t stride
)
2160 isl_buffer_fill_state(&device
->isl_dev
, state
.map
,
2162 .mocs
= device
->default_mocs
,
2167 anv_state_flush(device
, state
);
2170 void anv_DestroySampler(
2173 const VkAllocationCallbacks
* pAllocator
)
2175 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2176 ANV_FROM_HANDLE(anv_sampler
, sampler
, _sampler
);
2181 vk_free2(&device
->alloc
, pAllocator
, sampler
);
2184 VkResult
anv_CreateFramebuffer(
2186 const VkFramebufferCreateInfo
* pCreateInfo
,
2187 const VkAllocationCallbacks
* pAllocator
,
2188 VkFramebuffer
* pFramebuffer
)
2190 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2191 struct anv_framebuffer
*framebuffer
;
2193 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
2195 size_t size
= sizeof(*framebuffer
) +
2196 sizeof(struct anv_image_view
*) * pCreateInfo
->attachmentCount
;
2197 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
2198 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2199 if (framebuffer
== NULL
)
2200 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2202 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
2203 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
2204 VkImageView _iview
= pCreateInfo
->pAttachments
[i
];
2205 framebuffer
->attachments
[i
] = anv_image_view_from_handle(_iview
);
2208 framebuffer
->width
= pCreateInfo
->width
;
2209 framebuffer
->height
= pCreateInfo
->height
;
2210 framebuffer
->layers
= pCreateInfo
->layers
;
2212 *pFramebuffer
= anv_framebuffer_to_handle(framebuffer
);
2217 void anv_DestroyFramebuffer(
2220 const VkAllocationCallbacks
* pAllocator
)
2222 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2223 ANV_FROM_HANDLE(anv_framebuffer
, fb
, _fb
);
2228 vk_free2(&device
->alloc
, pAllocator
, fb
);
2231 /* vk_icd.h does not declare this function, so we declare it here to
2232 * suppress Wmissing-prototypes.
2234 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2235 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
);
2237 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2238 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
)
2240 /* For the full details on loader interface versioning, see
2241 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
2242 * What follows is a condensed summary, to help you navigate the large and
2243 * confusing official doc.
2245 * - Loader interface v0 is incompatible with later versions. We don't
2248 * - In loader interface v1:
2249 * - The first ICD entrypoint called by the loader is
2250 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
2252 * - The ICD must statically expose no other Vulkan symbol unless it is
2253 * linked with -Bsymbolic.
2254 * - Each dispatchable Vulkan handle created by the ICD must be
2255 * a pointer to a struct whose first member is VK_LOADER_DATA. The
2256 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
2257 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
2258 * vkDestroySurfaceKHR(). The ICD must be capable of working with
2259 * such loader-managed surfaces.
2261 * - Loader interface v2 differs from v1 in:
2262 * - The first ICD entrypoint called by the loader is
2263 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
2264 * statically expose this entrypoint.
2266 * - Loader interface v3 differs from v2 in:
2267 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
2268 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
2269 * because the loader no longer does so.
2271 *pSupportedVersion
= MIN2(*pSupportedVersion
, 3u);