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 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(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 fprintf(stderr
, "WARNING: Haswell Vulkan support is incomplete\n");
298 } else if (device
->info
.gen
== 7 && !device
->info
.is_baytrail
) {
299 fprintf(stderr
, "WARNING: Ivy Bridge Vulkan support is incomplete\n");
300 } else if (device
->info
.gen
== 7 && device
->info
.is_baytrail
) {
301 fprintf(stderr
, "WARNING: Bay Trail Vulkan support is incomplete\n");
302 } else if (device
->info
.gen
>= 8 && device
->info
.gen
<= 9) {
303 /* Broadwell, Cherryview, Skylake, Broxton, Kabylake is as fully
304 * supported as anything */
306 result
= vk_errorf(device
->instance
, device
,
307 VK_ERROR_INCOMPATIBLE_DRIVER
,
308 "Vulkan not yet supported on %s", device
->name
);
312 device
->cmd_parser_version
= -1;
313 if (device
->info
.gen
== 7) {
314 device
->cmd_parser_version
=
315 anv_gem_get_param(fd
, I915_PARAM_CMD_PARSER_VERSION
);
316 if (device
->cmd_parser_version
== -1) {
317 result
= vk_errorf(device
->instance
, device
,
318 VK_ERROR_INITIALIZATION_FAILED
,
319 "failed to get command parser version");
324 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_WAIT_TIMEOUT
)) {
325 result
= vk_errorf(device
->instance
, device
,
326 VK_ERROR_INITIALIZATION_FAILED
,
327 "kernel missing gem wait");
331 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_EXECBUF2
)) {
332 result
= vk_errorf(device
->instance
, device
,
333 VK_ERROR_INITIALIZATION_FAILED
,
334 "kernel missing execbuf2");
338 if (!device
->info
.has_llc
&&
339 anv_gem_get_param(fd
, I915_PARAM_MMAP_VERSION
) < 1) {
340 result
= vk_errorf(device
->instance
, device
,
341 VK_ERROR_INITIALIZATION_FAILED
,
342 "kernel missing wc mmap");
346 result
= anv_physical_device_init_heaps(device
, fd
);
347 if (result
!= VK_SUCCESS
)
350 device
->has_exec_async
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_ASYNC
);
351 device
->has_exec_fence
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE
);
352 device
->has_syncobj
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE_ARRAY
);
353 device
->has_syncobj_wait
= device
->has_syncobj
&&
354 anv_gem_supports_syncobj_wait(fd
);
356 bool swizzled
= anv_gem_get_bit6_swizzle(fd
, I915_TILING_X
);
358 /* GENs prior to 8 do not support EU/Subslice info */
359 if (device
->info
.gen
>= 8) {
360 device
->subslice_total
= anv_gem_get_param(fd
, I915_PARAM_SUBSLICE_TOTAL
);
361 device
->eu_total
= anv_gem_get_param(fd
, I915_PARAM_EU_TOTAL
);
363 /* Without this information, we cannot get the right Braswell
364 * brandstrings, and we have to use conservative numbers for GPGPU on
365 * many platforms, but otherwise, things will just work.
367 if (device
->subslice_total
< 1 || device
->eu_total
< 1) {
368 fprintf(stderr
, "WARNING: Kernel 4.1 required to properly"
369 " query GPU properties.\n");
371 } else if (device
->info
.gen
== 7) {
372 device
->subslice_total
= 1 << (device
->info
.gt
- 1);
375 if (device
->info
.is_cherryview
&&
376 device
->subslice_total
> 0 && device
->eu_total
> 0) {
377 /* Logical CS threads = EUs per subslice * num threads per EU */
378 uint32_t max_cs_threads
=
379 device
->eu_total
/ device
->subslice_total
* device
->info
.num_thread_per_eu
;
381 /* Fuse configurations may give more threads than expected, never less. */
382 if (max_cs_threads
> device
->info
.max_cs_threads
)
383 device
->info
.max_cs_threads
= max_cs_threads
;
386 device
->compiler
= brw_compiler_create(NULL
, &device
->info
);
387 if (device
->compiler
== NULL
) {
388 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
391 device
->compiler
->shader_debug_log
= compiler_debug_log
;
392 device
->compiler
->shader_perf_log
= compiler_perf_log
;
393 device
->compiler
->supports_pull_constants
= false;
395 isl_device_init(&device
->isl_dev
, &device
->info
, swizzled
);
397 result
= anv_physical_device_init_uuids(device
);
398 if (result
!= VK_SUCCESS
)
401 result
= anv_init_wsi(device
);
402 if (result
!= VK_SUCCESS
) {
403 ralloc_free(device
->compiler
);
407 device
->local_fd
= fd
;
416 anv_physical_device_finish(struct anv_physical_device
*device
)
418 anv_finish_wsi(device
);
419 ralloc_free(device
->compiler
);
420 close(device
->local_fd
);
424 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
425 VkSystemAllocationScope allocationScope
)
431 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
432 size_t align
, VkSystemAllocationScope allocationScope
)
434 return realloc(pOriginal
, size
);
438 default_free_func(void *pUserData
, void *pMemory
)
443 static const VkAllocationCallbacks default_alloc
= {
445 .pfnAllocation
= default_alloc_func
,
446 .pfnReallocation
= default_realloc_func
,
447 .pfnFree
= default_free_func
,
450 VkResult
anv_CreateInstance(
451 const VkInstanceCreateInfo
* pCreateInfo
,
452 const VkAllocationCallbacks
* pAllocator
,
453 VkInstance
* pInstance
)
455 struct anv_instance
*instance
;
457 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
459 /* Check if user passed a debug report callback to be used during
460 * Create/Destroy of instance.
462 const VkDebugReportCallbackCreateInfoEXT
*ctor_cb
=
463 vk_find_struct_const(pCreateInfo
->pNext
,
464 DEBUG_REPORT_CALLBACK_CREATE_INFO_EXT
);
466 uint32_t client_version
;
467 if (pCreateInfo
->pApplicationInfo
&&
468 pCreateInfo
->pApplicationInfo
->apiVersion
!= 0) {
469 client_version
= pCreateInfo
->pApplicationInfo
->apiVersion
;
471 client_version
= VK_MAKE_VERSION(1, 0, 0);
474 if (VK_MAKE_VERSION(1, 0, 0) > client_version
||
475 client_version
> VK_MAKE_VERSION(1, 0, 0xfff)) {
477 if (ctor_cb
&& ctor_cb
->flags
& VK_DEBUG_REPORT_ERROR_BIT_EXT
)
478 ctor_cb
->pfnCallback(VK_DEBUG_REPORT_ERROR_BIT_EXT
,
479 VK_DEBUG_REPORT_OBJECT_TYPE_INSTANCE_EXT
,
480 VK_NULL_HANDLE
, /* No handle available yet. */
484 "incompatible driver version",
487 return vk_errorf(NULL
, NULL
, VK_ERROR_INCOMPATIBLE_DRIVER
,
488 "Client requested version %d.%d.%d",
489 VK_VERSION_MAJOR(client_version
),
490 VK_VERSION_MINOR(client_version
),
491 VK_VERSION_PATCH(client_version
));
494 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
495 const char *ext_name
= pCreateInfo
->ppEnabledExtensionNames
[i
];
496 if (!anv_instance_extension_supported(ext_name
))
497 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
500 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
501 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
503 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
505 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
508 instance
->alloc
= *pAllocator
;
510 instance
->alloc
= default_alloc
;
512 instance
->apiVersion
= client_version
;
513 instance
->physicalDeviceCount
= -1;
515 if (pthread_mutex_init(&instance
->callbacks_mutex
, NULL
) != 0) {
516 vk_free2(&default_alloc
, pAllocator
, instance
);
517 return vk_error(VK_ERROR_INITIALIZATION_FAILED
);
520 list_inithead(&instance
->callbacks
);
522 /* Store report debug callback to be used during DestroyInstance. */
524 instance
->destroy_debug_cb
.flags
= ctor_cb
->flags
;
525 instance
->destroy_debug_cb
.callback
= ctor_cb
->pfnCallback
;
526 instance
->destroy_debug_cb
.data
= ctor_cb
->pUserData
;
531 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
533 *pInstance
= anv_instance_to_handle(instance
);
538 void anv_DestroyInstance(
539 VkInstance _instance
,
540 const VkAllocationCallbacks
* pAllocator
)
542 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
547 if (instance
->physicalDeviceCount
> 0) {
548 /* We support at most one physical device. */
549 assert(instance
->physicalDeviceCount
== 1);
550 anv_physical_device_finish(&instance
->physicalDevice
);
553 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
555 pthread_mutex_destroy(&instance
->callbacks_mutex
);
559 vk_free(&instance
->alloc
, instance
);
563 anv_enumerate_devices(struct anv_instance
*instance
)
565 /* TODO: Check for more devices ? */
566 drmDevicePtr devices
[8];
567 VkResult result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
570 instance
->physicalDeviceCount
= 0;
572 max_devices
= drmGetDevices2(0, devices
, ARRAY_SIZE(devices
));
574 return VK_ERROR_INCOMPATIBLE_DRIVER
;
576 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
577 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
578 devices
[i
]->bustype
== DRM_BUS_PCI
&&
579 devices
[i
]->deviceinfo
.pci
->vendor_id
== 0x8086) {
581 result
= anv_physical_device_init(&instance
->physicalDevice
,
583 devices
[i
]->nodes
[DRM_NODE_RENDER
]);
584 if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
588 drmFreeDevices(devices
, max_devices
);
590 if (result
== VK_SUCCESS
)
591 instance
->physicalDeviceCount
= 1;
597 VkResult
anv_EnumeratePhysicalDevices(
598 VkInstance _instance
,
599 uint32_t* pPhysicalDeviceCount
,
600 VkPhysicalDevice
* pPhysicalDevices
)
602 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
603 VK_OUTARRAY_MAKE(out
, pPhysicalDevices
, pPhysicalDeviceCount
);
606 if (instance
->physicalDeviceCount
< 0) {
607 result
= anv_enumerate_devices(instance
);
608 if (result
!= VK_SUCCESS
&&
609 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
613 if (instance
->physicalDeviceCount
> 0) {
614 assert(instance
->physicalDeviceCount
== 1);
615 vk_outarray_append(&out
, i
) {
616 *i
= anv_physical_device_to_handle(&instance
->physicalDevice
);
620 return vk_outarray_status(&out
);
623 void anv_GetPhysicalDeviceFeatures(
624 VkPhysicalDevice physicalDevice
,
625 VkPhysicalDeviceFeatures
* pFeatures
)
627 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
629 *pFeatures
= (VkPhysicalDeviceFeatures
) {
630 .robustBufferAccess
= true,
631 .fullDrawIndexUint32
= true,
632 .imageCubeArray
= true,
633 .independentBlend
= true,
634 .geometryShader
= true,
635 .tessellationShader
= true,
636 .sampleRateShading
= true,
637 .dualSrcBlend
= true,
639 .multiDrawIndirect
= true,
640 .drawIndirectFirstInstance
= true,
642 .depthBiasClamp
= true,
643 .fillModeNonSolid
= true,
644 .depthBounds
= false,
648 .multiViewport
= true,
649 .samplerAnisotropy
= true,
650 .textureCompressionETC2
= pdevice
->info
.gen
>= 8 ||
651 pdevice
->info
.is_baytrail
,
652 .textureCompressionASTC_LDR
= pdevice
->info
.gen
>= 9, /* FINISHME CHV */
653 .textureCompressionBC
= true,
654 .occlusionQueryPrecise
= true,
655 .pipelineStatisticsQuery
= true,
656 .fragmentStoresAndAtomics
= true,
657 .shaderTessellationAndGeometryPointSize
= true,
658 .shaderImageGatherExtended
= true,
659 .shaderStorageImageExtendedFormats
= true,
660 .shaderStorageImageMultisample
= false,
661 .shaderStorageImageReadWithoutFormat
= false,
662 .shaderStorageImageWriteWithoutFormat
= true,
663 .shaderUniformBufferArrayDynamicIndexing
= true,
664 .shaderSampledImageArrayDynamicIndexing
= true,
665 .shaderStorageBufferArrayDynamicIndexing
= true,
666 .shaderStorageImageArrayDynamicIndexing
= true,
667 .shaderClipDistance
= true,
668 .shaderCullDistance
= true,
669 .shaderFloat64
= pdevice
->info
.gen
>= 8,
670 .shaderInt64
= pdevice
->info
.gen
>= 8,
671 .shaderInt16
= false,
672 .shaderResourceMinLod
= false,
673 .variableMultisampleRate
= false,
674 .inheritedQueries
= true,
677 /* We can't do image stores in vec4 shaders */
678 pFeatures
->vertexPipelineStoresAndAtomics
=
679 pdevice
->compiler
->scalar_stage
[MESA_SHADER_VERTEX
] &&
680 pdevice
->compiler
->scalar_stage
[MESA_SHADER_GEOMETRY
];
683 void anv_GetPhysicalDeviceFeatures2KHR(
684 VkPhysicalDevice physicalDevice
,
685 VkPhysicalDeviceFeatures2KHR
* pFeatures
)
687 anv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
689 vk_foreach_struct(ext
, pFeatures
->pNext
) {
690 switch (ext
->sType
) {
691 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES_KHX
: {
692 VkPhysicalDeviceMultiviewFeaturesKHX
*features
=
693 (VkPhysicalDeviceMultiviewFeaturesKHX
*)ext
;
694 features
->multiview
= true;
695 features
->multiviewGeometryShader
= true;
696 features
->multiviewTessellationShader
= true;
700 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTER_FEATURES_KHR
: {
701 VkPhysicalDeviceVariablePointerFeaturesKHR
*features
= (void *)ext
;
702 features
->variablePointersStorageBuffer
= true;
703 features
->variablePointers
= false;
707 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES_KHR
: {
708 VkPhysicalDeviceSamplerYcbcrConversionFeaturesKHR
*features
=
709 (VkPhysicalDeviceSamplerYcbcrConversionFeaturesKHR
*) ext
;
710 features
->samplerYcbcrConversion
= true;
715 anv_debug_ignored_stype(ext
->sType
);
721 void anv_GetPhysicalDeviceProperties(
722 VkPhysicalDevice physicalDevice
,
723 VkPhysicalDeviceProperties
* pProperties
)
725 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
726 const struct gen_device_info
*devinfo
= &pdevice
->info
;
728 /* See assertions made when programming the buffer surface state. */
729 const uint32_t max_raw_buffer_sz
= devinfo
->gen
>= 7 ?
730 (1ul << 30) : (1ul << 27);
732 const uint32_t max_samplers
= (devinfo
->gen
>= 8 || devinfo
->is_haswell
) ?
735 VkSampleCountFlags sample_counts
=
736 isl_device_get_sample_counts(&pdevice
->isl_dev
);
738 VkPhysicalDeviceLimits limits
= {
739 .maxImageDimension1D
= (1 << 14),
740 .maxImageDimension2D
= (1 << 14),
741 .maxImageDimension3D
= (1 << 11),
742 .maxImageDimensionCube
= (1 << 14),
743 .maxImageArrayLayers
= (1 << 11),
744 .maxTexelBufferElements
= 128 * 1024 * 1024,
745 .maxUniformBufferRange
= (1ul << 27),
746 .maxStorageBufferRange
= max_raw_buffer_sz
,
747 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
748 .maxMemoryAllocationCount
= UINT32_MAX
,
749 .maxSamplerAllocationCount
= 64 * 1024,
750 .bufferImageGranularity
= 64, /* A cache line */
751 .sparseAddressSpaceSize
= 0,
752 .maxBoundDescriptorSets
= MAX_SETS
,
753 .maxPerStageDescriptorSamplers
= max_samplers
,
754 .maxPerStageDescriptorUniformBuffers
= 64,
755 .maxPerStageDescriptorStorageBuffers
= 64,
756 .maxPerStageDescriptorSampledImages
= max_samplers
,
757 .maxPerStageDescriptorStorageImages
= 64,
758 .maxPerStageDescriptorInputAttachments
= 64,
759 .maxPerStageResources
= 250,
760 .maxDescriptorSetSamplers
= 256,
761 .maxDescriptorSetUniformBuffers
= 256,
762 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
763 .maxDescriptorSetStorageBuffers
= 256,
764 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
765 .maxDescriptorSetSampledImages
= 256,
766 .maxDescriptorSetStorageImages
= 256,
767 .maxDescriptorSetInputAttachments
= 256,
768 .maxVertexInputAttributes
= MAX_VBS
,
769 .maxVertexInputBindings
= MAX_VBS
,
770 .maxVertexInputAttributeOffset
= 2047,
771 .maxVertexInputBindingStride
= 2048,
772 .maxVertexOutputComponents
= 128,
773 .maxTessellationGenerationLevel
= 64,
774 .maxTessellationPatchSize
= 32,
775 .maxTessellationControlPerVertexInputComponents
= 128,
776 .maxTessellationControlPerVertexOutputComponents
= 128,
777 .maxTessellationControlPerPatchOutputComponents
= 128,
778 .maxTessellationControlTotalOutputComponents
= 2048,
779 .maxTessellationEvaluationInputComponents
= 128,
780 .maxTessellationEvaluationOutputComponents
= 128,
781 .maxGeometryShaderInvocations
= 32,
782 .maxGeometryInputComponents
= 64,
783 .maxGeometryOutputComponents
= 128,
784 .maxGeometryOutputVertices
= 256,
785 .maxGeometryTotalOutputComponents
= 1024,
786 .maxFragmentInputComponents
= 128,
787 .maxFragmentOutputAttachments
= 8,
788 .maxFragmentDualSrcAttachments
= 1,
789 .maxFragmentCombinedOutputResources
= 8,
790 .maxComputeSharedMemorySize
= 32768,
791 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
792 .maxComputeWorkGroupInvocations
= 16 * devinfo
->max_cs_threads
,
793 .maxComputeWorkGroupSize
= {
794 16 * devinfo
->max_cs_threads
,
795 16 * devinfo
->max_cs_threads
,
796 16 * devinfo
->max_cs_threads
,
798 .subPixelPrecisionBits
= 4 /* FIXME */,
799 .subTexelPrecisionBits
= 4 /* FIXME */,
800 .mipmapPrecisionBits
= 4 /* FIXME */,
801 .maxDrawIndexedIndexValue
= UINT32_MAX
,
802 .maxDrawIndirectCount
= UINT32_MAX
,
803 .maxSamplerLodBias
= 16,
804 .maxSamplerAnisotropy
= 16,
805 .maxViewports
= MAX_VIEWPORTS
,
806 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
807 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
808 .viewportSubPixelBits
= 13, /* We take a float? */
809 .minMemoryMapAlignment
= 4096, /* A page */
810 .minTexelBufferOffsetAlignment
= 1,
811 .minUniformBufferOffsetAlignment
= 16,
812 .minStorageBufferOffsetAlignment
= 4,
813 .minTexelOffset
= -8,
815 .minTexelGatherOffset
= -32,
816 .maxTexelGatherOffset
= 31,
817 .minInterpolationOffset
= -0.5,
818 .maxInterpolationOffset
= 0.4375,
819 .subPixelInterpolationOffsetBits
= 4,
820 .maxFramebufferWidth
= (1 << 14),
821 .maxFramebufferHeight
= (1 << 14),
822 .maxFramebufferLayers
= (1 << 11),
823 .framebufferColorSampleCounts
= sample_counts
,
824 .framebufferDepthSampleCounts
= sample_counts
,
825 .framebufferStencilSampleCounts
= sample_counts
,
826 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
827 .maxColorAttachments
= MAX_RTS
,
828 .sampledImageColorSampleCounts
= sample_counts
,
829 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
830 .sampledImageDepthSampleCounts
= sample_counts
,
831 .sampledImageStencilSampleCounts
= sample_counts
,
832 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
833 .maxSampleMaskWords
= 1,
834 .timestampComputeAndGraphics
= false,
835 .timestampPeriod
= 1000000000.0 / devinfo
->timestamp_frequency
,
836 .maxClipDistances
= 8,
837 .maxCullDistances
= 8,
838 .maxCombinedClipAndCullDistances
= 8,
839 .discreteQueuePriorities
= 1,
840 .pointSizeRange
= { 0.125, 255.875 },
841 .lineWidthRange
= { 0.0, 7.9921875 },
842 .pointSizeGranularity
= (1.0 / 8.0),
843 .lineWidthGranularity
= (1.0 / 128.0),
844 .strictLines
= false, /* FINISHME */
845 .standardSampleLocations
= true,
846 .optimalBufferCopyOffsetAlignment
= 128,
847 .optimalBufferCopyRowPitchAlignment
= 128,
848 .nonCoherentAtomSize
= 64,
851 *pProperties
= (VkPhysicalDeviceProperties
) {
852 .apiVersion
= anv_physical_device_api_version(pdevice
),
853 .driverVersion
= vk_get_driver_version(),
855 .deviceID
= pdevice
->chipset_id
,
856 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
858 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
861 snprintf(pProperties
->deviceName
, sizeof(pProperties
->deviceName
),
862 "%s", pdevice
->name
);
863 memcpy(pProperties
->pipelineCacheUUID
,
864 pdevice
->pipeline_cache_uuid
, VK_UUID_SIZE
);
867 void anv_GetPhysicalDeviceProperties2KHR(
868 VkPhysicalDevice physicalDevice
,
869 VkPhysicalDeviceProperties2KHR
* pProperties
)
871 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
873 anv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
875 vk_foreach_struct(ext
, pProperties
->pNext
) {
876 switch (ext
->sType
) {
877 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
878 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
879 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
881 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
885 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES_KHR
: {
886 VkPhysicalDeviceIDPropertiesKHR
*id_props
=
887 (VkPhysicalDeviceIDPropertiesKHR
*)ext
;
888 memcpy(id_props
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
889 memcpy(id_props
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
890 /* The LUID is for Windows. */
891 id_props
->deviceLUIDValid
= false;
895 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES_KHX
: {
896 VkPhysicalDeviceMultiviewPropertiesKHX
*properties
=
897 (VkPhysicalDeviceMultiviewPropertiesKHX
*)ext
;
898 properties
->maxMultiviewViewCount
= 16;
899 properties
->maxMultiviewInstanceIndex
= UINT32_MAX
/ 16;
903 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES_KHR
: {
904 VkPhysicalDevicePointClippingPropertiesKHR
*properties
=
905 (VkPhysicalDevicePointClippingPropertiesKHR
*) ext
;
906 properties
->pointClippingBehavior
= VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES_KHR
;
907 anv_finishme("Implement pop-free point clipping");
912 anv_debug_ignored_stype(ext
->sType
);
918 /* We support exactly one queue family. */
919 static const VkQueueFamilyProperties
920 anv_queue_family_properties
= {
921 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
922 VK_QUEUE_COMPUTE_BIT
|
923 VK_QUEUE_TRANSFER_BIT
,
925 .timestampValidBits
= 36, /* XXX: Real value here */
926 .minImageTransferGranularity
= { 1, 1, 1 },
929 void anv_GetPhysicalDeviceQueueFamilyProperties(
930 VkPhysicalDevice physicalDevice
,
932 VkQueueFamilyProperties
* pQueueFamilyProperties
)
934 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pCount
);
936 vk_outarray_append(&out
, p
) {
937 *p
= anv_queue_family_properties
;
941 void anv_GetPhysicalDeviceQueueFamilyProperties2KHR(
942 VkPhysicalDevice physicalDevice
,
943 uint32_t* pQueueFamilyPropertyCount
,
944 VkQueueFamilyProperties2KHR
* pQueueFamilyProperties
)
947 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pQueueFamilyPropertyCount
);
949 vk_outarray_append(&out
, p
) {
950 p
->queueFamilyProperties
= anv_queue_family_properties
;
952 vk_foreach_struct(s
, p
->pNext
) {
953 anv_debug_ignored_stype(s
->sType
);
958 void anv_GetPhysicalDeviceMemoryProperties(
959 VkPhysicalDevice physicalDevice
,
960 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
962 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
964 pMemoryProperties
->memoryTypeCount
= physical_device
->memory
.type_count
;
965 for (uint32_t i
= 0; i
< physical_device
->memory
.type_count
; i
++) {
966 pMemoryProperties
->memoryTypes
[i
] = (VkMemoryType
) {
967 .propertyFlags
= physical_device
->memory
.types
[i
].propertyFlags
,
968 .heapIndex
= physical_device
->memory
.types
[i
].heapIndex
,
972 pMemoryProperties
->memoryHeapCount
= physical_device
->memory
.heap_count
;
973 for (uint32_t i
= 0; i
< physical_device
->memory
.heap_count
; i
++) {
974 pMemoryProperties
->memoryHeaps
[i
] = (VkMemoryHeap
) {
975 .size
= physical_device
->memory
.heaps
[i
].size
,
976 .flags
= physical_device
->memory
.heaps
[i
].flags
,
981 void anv_GetPhysicalDeviceMemoryProperties2KHR(
982 VkPhysicalDevice physicalDevice
,
983 VkPhysicalDeviceMemoryProperties2KHR
* pMemoryProperties
)
985 anv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
986 &pMemoryProperties
->memoryProperties
);
988 vk_foreach_struct(ext
, pMemoryProperties
->pNext
) {
989 switch (ext
->sType
) {
991 anv_debug_ignored_stype(ext
->sType
);
997 PFN_vkVoidFunction
anv_GetInstanceProcAddr(
1001 return anv_lookup_entrypoint(NULL
, pName
);
1004 /* With version 1+ of the loader interface the ICD should expose
1005 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
1008 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1009 VkInstance instance
,
1013 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1014 VkInstance instance
,
1017 return anv_GetInstanceProcAddr(instance
, pName
);
1020 PFN_vkVoidFunction
anv_GetDeviceProcAddr(
1024 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1025 return anv_lookup_entrypoint(&device
->info
, pName
);
1029 anv_queue_init(struct anv_device
*device
, struct anv_queue
*queue
)
1031 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1032 queue
->device
= device
;
1033 queue
->pool
= &device
->surface_state_pool
;
1037 anv_queue_finish(struct anv_queue
*queue
)
1041 static struct anv_state
1042 anv_state_pool_emit_data(struct anv_state_pool
*pool
, size_t size
, size_t align
, const void *p
)
1044 struct anv_state state
;
1046 state
= anv_state_pool_alloc(pool
, size
, align
);
1047 memcpy(state
.map
, p
, size
);
1049 anv_state_flush(pool
->block_pool
.device
, state
);
1054 struct gen8_border_color
{
1059 /* Pad out to 64 bytes */
1064 anv_device_init_border_colors(struct anv_device
*device
)
1066 static const struct gen8_border_color border_colors
[] = {
1067 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
1068 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
1069 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
1070 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
1071 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
1072 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
1075 device
->border_colors
= anv_state_pool_emit_data(&device
->dynamic_state_pool
,
1076 sizeof(border_colors
), 64,
1081 anv_device_init_trivial_batch(struct anv_device
*device
)
1083 anv_bo_init_new(&device
->trivial_batch_bo
, device
, 4096);
1085 if (device
->instance
->physicalDevice
.has_exec_async
)
1086 device
->trivial_batch_bo
.flags
|= EXEC_OBJECT_ASYNC
;
1088 void *map
= anv_gem_mmap(device
, device
->trivial_batch_bo
.gem_handle
,
1091 struct anv_batch batch
= {
1097 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
1098 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
1100 if (!device
->info
.has_llc
)
1101 gen_clflush_range(map
, batch
.next
- map
);
1103 anv_gem_munmap(map
, device
->trivial_batch_bo
.size
);
1106 VkResult
anv_CreateDevice(
1107 VkPhysicalDevice physicalDevice
,
1108 const VkDeviceCreateInfo
* pCreateInfo
,
1109 const VkAllocationCallbacks
* pAllocator
,
1112 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
1114 struct anv_device
*device
;
1116 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO
);
1118 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
1119 const char *ext_name
= pCreateInfo
->ppEnabledExtensionNames
[i
];
1120 if (!anv_physical_device_extension_supported(physical_device
, ext_name
))
1121 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
1124 /* Check enabled features */
1125 if (pCreateInfo
->pEnabledFeatures
) {
1126 VkPhysicalDeviceFeatures supported_features
;
1127 anv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
1128 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
1129 VkBool32
*enabled_feature
= (VkBool32
*)pCreateInfo
->pEnabledFeatures
;
1130 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
1131 for (uint32_t i
= 0; i
< num_features
; i
++) {
1132 if (enabled_feature
[i
] && !supported_feature
[i
])
1133 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
1137 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
1139 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
1141 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1143 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1144 device
->instance
= physical_device
->instance
;
1145 device
->chipset_id
= physical_device
->chipset_id
;
1146 device
->lost
= false;
1149 device
->alloc
= *pAllocator
;
1151 device
->alloc
= physical_device
->instance
->alloc
;
1153 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
1154 device
->fd
= open(physical_device
->path
, O_RDWR
| O_CLOEXEC
);
1155 if (device
->fd
== -1) {
1156 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1160 device
->context_id
= anv_gem_create_context(device
);
1161 if (device
->context_id
== -1) {
1162 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1166 device
->info
= physical_device
->info
;
1167 device
->isl_dev
= physical_device
->isl_dev
;
1169 /* On Broadwell and later, we can use batch chaining to more efficiently
1170 * implement growing command buffers. Prior to Haswell, the kernel
1171 * command parser gets in the way and we have to fall back to growing
1174 device
->can_chain_batches
= device
->info
.gen
>= 8;
1176 device
->robust_buffer_access
= pCreateInfo
->pEnabledFeatures
&&
1177 pCreateInfo
->pEnabledFeatures
->robustBufferAccess
;
1179 if (pthread_mutex_init(&device
->mutex
, NULL
) != 0) {
1180 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1181 goto fail_context_id
;
1184 pthread_condattr_t condattr
;
1185 if (pthread_condattr_init(&condattr
) != 0) {
1186 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1189 if (pthread_condattr_setclock(&condattr
, CLOCK_MONOTONIC
) != 0) {
1190 pthread_condattr_destroy(&condattr
);
1191 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1194 if (pthread_cond_init(&device
->queue_submit
, NULL
) != 0) {
1195 pthread_condattr_destroy(&condattr
);
1196 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1199 pthread_condattr_destroy(&condattr
);
1201 anv_bo_pool_init(&device
->batch_bo_pool
, device
);
1203 result
= anv_bo_cache_init(&device
->bo_cache
);
1204 if (result
!= VK_SUCCESS
)
1205 goto fail_batch_bo_pool
;
1207 result
= anv_state_pool_init(&device
->dynamic_state_pool
, device
, 16384);
1208 if (result
!= VK_SUCCESS
)
1211 result
= anv_state_pool_init(&device
->instruction_state_pool
, device
, 16384);
1212 if (result
!= VK_SUCCESS
)
1213 goto fail_dynamic_state_pool
;
1215 result
= anv_state_pool_init(&device
->surface_state_pool
, device
, 4096);
1216 if (result
!= VK_SUCCESS
)
1217 goto fail_instruction_state_pool
;
1219 result
= anv_bo_init_new(&device
->workaround_bo
, device
, 1024);
1220 if (result
!= VK_SUCCESS
)
1221 goto fail_surface_state_pool
;
1223 anv_device_init_trivial_batch(device
);
1225 anv_scratch_pool_init(device
, &device
->scratch_pool
);
1227 anv_queue_init(device
, &device
->queue
);
1229 switch (device
->info
.gen
) {
1231 if (!device
->info
.is_haswell
)
1232 result
= gen7_init_device_state(device
);
1234 result
= gen75_init_device_state(device
);
1237 result
= gen8_init_device_state(device
);
1240 result
= gen9_init_device_state(device
);
1243 result
= gen10_init_device_state(device
);
1246 /* Shouldn't get here as we don't create physical devices for any other
1248 unreachable("unhandled gen");
1250 if (result
!= VK_SUCCESS
)
1251 goto fail_workaround_bo
;
1253 anv_device_init_blorp(device
);
1255 anv_device_init_border_colors(device
);
1257 *pDevice
= anv_device_to_handle(device
);
1262 anv_queue_finish(&device
->queue
);
1263 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
1264 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
1265 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
1266 fail_surface_state_pool
:
1267 anv_state_pool_finish(&device
->surface_state_pool
);
1268 fail_instruction_state_pool
:
1269 anv_state_pool_finish(&device
->instruction_state_pool
);
1270 fail_dynamic_state_pool
:
1271 anv_state_pool_finish(&device
->dynamic_state_pool
);
1273 anv_bo_cache_finish(&device
->bo_cache
);
1275 anv_bo_pool_finish(&device
->batch_bo_pool
);
1276 pthread_cond_destroy(&device
->queue_submit
);
1278 pthread_mutex_destroy(&device
->mutex
);
1280 anv_gem_destroy_context(device
, device
->context_id
);
1284 vk_free(&device
->alloc
, device
);
1289 void anv_DestroyDevice(
1291 const VkAllocationCallbacks
* pAllocator
)
1293 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1298 anv_device_finish_blorp(device
);
1300 anv_queue_finish(&device
->queue
);
1302 #ifdef HAVE_VALGRIND
1303 /* We only need to free these to prevent valgrind errors. The backing
1304 * BO will go away in a couple of lines so we don't actually leak.
1306 anv_state_pool_free(&device
->dynamic_state_pool
, device
->border_colors
);
1309 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
1311 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
1312 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
1314 anv_gem_close(device
, device
->trivial_batch_bo
.gem_handle
);
1316 anv_state_pool_finish(&device
->surface_state_pool
);
1317 anv_state_pool_finish(&device
->instruction_state_pool
);
1318 anv_state_pool_finish(&device
->dynamic_state_pool
);
1320 anv_bo_cache_finish(&device
->bo_cache
);
1322 anv_bo_pool_finish(&device
->batch_bo_pool
);
1324 pthread_cond_destroy(&device
->queue_submit
);
1325 pthread_mutex_destroy(&device
->mutex
);
1327 anv_gem_destroy_context(device
, device
->context_id
);
1331 vk_free(&device
->alloc
, device
);
1334 VkResult
anv_EnumerateInstanceLayerProperties(
1335 uint32_t* pPropertyCount
,
1336 VkLayerProperties
* pProperties
)
1338 if (pProperties
== NULL
) {
1339 *pPropertyCount
= 0;
1343 /* None supported at this time */
1344 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1347 VkResult
anv_EnumerateDeviceLayerProperties(
1348 VkPhysicalDevice physicalDevice
,
1349 uint32_t* pPropertyCount
,
1350 VkLayerProperties
* pProperties
)
1352 if (pProperties
== NULL
) {
1353 *pPropertyCount
= 0;
1357 /* None supported at this time */
1358 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1361 void anv_GetDeviceQueue(
1363 uint32_t queueNodeIndex
,
1364 uint32_t queueIndex
,
1367 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1369 assert(queueIndex
== 0);
1371 *pQueue
= anv_queue_to_handle(&device
->queue
);
1375 anv_device_query_status(struct anv_device
*device
)
1377 /* This isn't likely as most of the callers of this function already check
1378 * for it. However, it doesn't hurt to check and it potentially lets us
1381 if (unlikely(device
->lost
))
1382 return VK_ERROR_DEVICE_LOST
;
1384 uint32_t active
, pending
;
1385 int ret
= anv_gem_gpu_get_reset_stats(device
, &active
, &pending
);
1387 /* We don't know the real error. */
1388 device
->lost
= true;
1389 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1390 "get_reset_stats failed: %m");
1394 device
->lost
= true;
1395 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1396 "GPU hung on one of our command buffers");
1397 } else if (pending
) {
1398 device
->lost
= true;
1399 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1400 "GPU hung with commands in-flight");
1407 anv_device_bo_busy(struct anv_device
*device
, struct anv_bo
*bo
)
1409 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
1410 * Other usages of the BO (such as on different hardware) will not be
1411 * flagged as "busy" by this ioctl. Use with care.
1413 int ret
= anv_gem_busy(device
, bo
->gem_handle
);
1415 return VK_NOT_READY
;
1416 } else if (ret
== -1) {
1417 /* We don't know the real error. */
1418 device
->lost
= true;
1419 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1420 "gem wait failed: %m");
1423 /* Query for device status after the busy call. If the BO we're checking
1424 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
1425 * client because it clearly doesn't have valid data. Yes, this most
1426 * likely means an ioctl, but we just did an ioctl to query the busy status
1427 * so it's no great loss.
1429 return anv_device_query_status(device
);
1433 anv_device_wait(struct anv_device
*device
, struct anv_bo
*bo
,
1436 int ret
= anv_gem_wait(device
, bo
->gem_handle
, &timeout
);
1437 if (ret
== -1 && errno
== ETIME
) {
1439 } else if (ret
== -1) {
1440 /* We don't know the real error. */
1441 device
->lost
= true;
1442 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1443 "gem wait failed: %m");
1446 /* Query for device status after the wait. If the BO we're waiting on got
1447 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
1448 * because it clearly doesn't have valid data. Yes, this most likely means
1449 * an ioctl, but we just did an ioctl to wait so it's no great loss.
1451 return anv_device_query_status(device
);
1454 VkResult
anv_DeviceWaitIdle(
1457 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1458 if (unlikely(device
->lost
))
1459 return VK_ERROR_DEVICE_LOST
;
1461 struct anv_batch batch
;
1464 batch
.start
= batch
.next
= cmds
;
1465 batch
.end
= (void *) cmds
+ sizeof(cmds
);
1467 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
1468 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
1470 return anv_device_submit_simple_batch(device
, &batch
);
1474 anv_bo_init_new(struct anv_bo
*bo
, struct anv_device
*device
, uint64_t size
)
1476 uint32_t gem_handle
= anv_gem_create(device
, size
);
1478 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
1480 anv_bo_init(bo
, gem_handle
, size
);
1485 VkResult
anv_AllocateMemory(
1487 const VkMemoryAllocateInfo
* pAllocateInfo
,
1488 const VkAllocationCallbacks
* pAllocator
,
1489 VkDeviceMemory
* pMem
)
1491 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1492 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
1493 struct anv_device_memory
*mem
;
1494 VkResult result
= VK_SUCCESS
;
1496 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
1498 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
1499 assert(pAllocateInfo
->allocationSize
> 0);
1501 /* The kernel relocation API has a limitation of a 32-bit delta value
1502 * applied to the address before it is written which, in spite of it being
1503 * unsigned, is treated as signed . Because of the way that this maps to
1504 * the Vulkan API, we cannot handle an offset into a buffer that does not
1505 * fit into a signed 32 bits. The only mechanism we have for dealing with
1506 * this at the moment is to limit all VkDeviceMemory objects to a maximum
1507 * of 2GB each. The Vulkan spec allows us to do this:
1509 * "Some platforms may have a limit on the maximum size of a single
1510 * allocation. For example, certain systems may fail to create
1511 * allocations with a size greater than or equal to 4GB. Such a limit is
1512 * implementation-dependent, and if such a failure occurs then the error
1513 * VK_ERROR_OUT_OF_DEVICE_MEMORY should be returned."
1515 * We don't use vk_error here because it's not an error so much as an
1516 * indication to the application that the allocation is too large.
1518 if (pAllocateInfo
->allocationSize
> (1ull << 31))
1519 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
1521 /* FINISHME: Fail if allocation request exceeds heap size. */
1523 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
1524 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1526 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1528 assert(pAllocateInfo
->memoryTypeIndex
< pdevice
->memory
.type_count
);
1529 mem
->type
= &pdevice
->memory
.types
[pAllocateInfo
->memoryTypeIndex
];
1533 const VkImportMemoryFdInfoKHR
*fd_info
=
1534 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_FD_INFO_KHR
);
1536 /* The Vulkan spec permits handleType to be 0, in which case the struct is
1539 if (fd_info
&& fd_info
->handleType
) {
1540 /* At the moment, we only support the OPAQUE_FD memory type which is
1541 * just a GEM buffer.
1543 assert(fd_info
->handleType
==
1544 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
);
1546 result
= anv_bo_cache_import(device
, &device
->bo_cache
,
1547 fd_info
->fd
, pAllocateInfo
->allocationSize
,
1549 if (result
!= VK_SUCCESS
)
1552 result
= anv_bo_cache_alloc(device
, &device
->bo_cache
,
1553 pAllocateInfo
->allocationSize
,
1555 if (result
!= VK_SUCCESS
)
1559 assert(mem
->type
->heapIndex
< pdevice
->memory
.heap_count
);
1560 if (pdevice
->memory
.heaps
[mem
->type
->heapIndex
].supports_48bit_addresses
)
1561 mem
->bo
->flags
|= EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
1563 if (pdevice
->has_exec_async
)
1564 mem
->bo
->flags
|= EXEC_OBJECT_ASYNC
;
1566 *pMem
= anv_device_memory_to_handle(mem
);
1571 vk_free2(&device
->alloc
, pAllocator
, mem
);
1576 VkResult
anv_GetMemoryFdKHR(
1578 const VkMemoryGetFdInfoKHR
* pGetFdInfo
,
1581 ANV_FROM_HANDLE(anv_device
, dev
, device_h
);
1582 ANV_FROM_HANDLE(anv_device_memory
, mem
, pGetFdInfo
->memory
);
1584 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
1586 /* We support only one handle type. */
1587 assert(pGetFdInfo
->handleType
==
1588 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
);
1590 return anv_bo_cache_export(dev
, &dev
->bo_cache
, mem
->bo
, pFd
);
1593 VkResult
anv_GetMemoryFdPropertiesKHR(
1595 VkExternalMemoryHandleTypeFlagBitsKHR handleType
,
1597 VkMemoryFdPropertiesKHR
* pMemoryFdProperties
)
1599 /* The valid usage section for this function says:
1601 * "handleType must not be one of the handle types defined as opaque."
1603 * Since we only handle opaque handles for now, there are no FD properties.
1605 return VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
;
1608 void anv_FreeMemory(
1610 VkDeviceMemory _mem
,
1611 const VkAllocationCallbacks
* pAllocator
)
1613 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1614 ANV_FROM_HANDLE(anv_device_memory
, mem
, _mem
);
1620 anv_UnmapMemory(_device
, _mem
);
1622 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
1624 vk_free2(&device
->alloc
, pAllocator
, mem
);
1627 VkResult
anv_MapMemory(
1629 VkDeviceMemory _memory
,
1630 VkDeviceSize offset
,
1632 VkMemoryMapFlags flags
,
1635 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1636 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
1643 if (size
== VK_WHOLE_SIZE
)
1644 size
= mem
->bo
->size
- offset
;
1646 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
1648 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
1649 * assert(size != 0);
1650 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
1651 * equal to the size of the memory minus offset
1654 assert(offset
+ size
<= mem
->bo
->size
);
1656 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
1657 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
1658 * at a time is valid. We could just mmap up front and return an offset
1659 * pointer here, but that may exhaust virtual memory on 32 bit
1662 uint32_t gem_flags
= 0;
1664 if (!device
->info
.has_llc
&&
1665 (mem
->type
->propertyFlags
& VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
))
1666 gem_flags
|= I915_MMAP_WC
;
1668 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
1669 uint64_t map_offset
= offset
& ~4095ull;
1670 assert(offset
>= map_offset
);
1671 uint64_t map_size
= (offset
+ size
) - map_offset
;
1673 /* Let's map whole pages */
1674 map_size
= align_u64(map_size
, 4096);
1676 void *map
= anv_gem_mmap(device
, mem
->bo
->gem_handle
,
1677 map_offset
, map_size
, gem_flags
);
1678 if (map
== MAP_FAILED
)
1679 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
1682 mem
->map_size
= map_size
;
1684 *ppData
= mem
->map
+ (offset
- map_offset
);
1689 void anv_UnmapMemory(
1691 VkDeviceMemory _memory
)
1693 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
1698 anv_gem_munmap(mem
->map
, mem
->map_size
);
1705 clflush_mapped_ranges(struct anv_device
*device
,
1707 const VkMappedMemoryRange
*ranges
)
1709 for (uint32_t i
= 0; i
< count
; i
++) {
1710 ANV_FROM_HANDLE(anv_device_memory
, mem
, ranges
[i
].memory
);
1711 if (ranges
[i
].offset
>= mem
->map_size
)
1714 gen_clflush_range(mem
->map
+ ranges
[i
].offset
,
1715 MIN2(ranges
[i
].size
, mem
->map_size
- ranges
[i
].offset
));
1719 VkResult
anv_FlushMappedMemoryRanges(
1721 uint32_t memoryRangeCount
,
1722 const VkMappedMemoryRange
* pMemoryRanges
)
1724 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1726 if (device
->info
.has_llc
)
1729 /* Make sure the writes we're flushing have landed. */
1730 __builtin_ia32_mfence();
1732 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
1737 VkResult
anv_InvalidateMappedMemoryRanges(
1739 uint32_t memoryRangeCount
,
1740 const VkMappedMemoryRange
* pMemoryRanges
)
1742 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1744 if (device
->info
.has_llc
)
1747 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
1749 /* Make sure no reads get moved up above the invalidate. */
1750 __builtin_ia32_mfence();
1755 void anv_GetBufferMemoryRequirements(
1758 VkMemoryRequirements
* pMemoryRequirements
)
1760 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
1761 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1762 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
1764 /* The Vulkan spec (git aaed022) says:
1766 * memoryTypeBits is a bitfield and contains one bit set for every
1767 * supported memory type for the resource. The bit `1<<i` is set if and
1768 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1769 * structure for the physical device is supported.
1771 uint32_t memory_types
= 0;
1772 for (uint32_t i
= 0; i
< pdevice
->memory
.type_count
; i
++) {
1773 uint32_t valid_usage
= pdevice
->memory
.types
[i
].valid_buffer_usage
;
1774 if ((valid_usage
& buffer
->usage
) == buffer
->usage
)
1775 memory_types
|= (1u << i
);
1778 pMemoryRequirements
->size
= buffer
->size
;
1779 pMemoryRequirements
->alignment
= 16;
1780 pMemoryRequirements
->memoryTypeBits
= memory_types
;
1783 void anv_GetBufferMemoryRequirements2KHR(
1785 const VkBufferMemoryRequirementsInfo2KHR
* pInfo
,
1786 VkMemoryRequirements2KHR
* pMemoryRequirements
)
1788 anv_GetBufferMemoryRequirements(_device
, pInfo
->buffer
,
1789 &pMemoryRequirements
->memoryRequirements
);
1791 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
1792 switch (ext
->sType
) {
1793 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS_KHR
: {
1794 VkMemoryDedicatedRequirementsKHR
*requirements
= (void *)ext
;
1795 requirements
->prefersDedicatedAllocation
= VK_FALSE
;
1796 requirements
->requiresDedicatedAllocation
= VK_FALSE
;
1801 anv_debug_ignored_stype(ext
->sType
);
1807 void anv_GetImageMemoryRequirements(
1810 VkMemoryRequirements
* pMemoryRequirements
)
1812 ANV_FROM_HANDLE(anv_image
, image
, _image
);
1813 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1814 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
1816 /* The Vulkan spec (git aaed022) says:
1818 * memoryTypeBits is a bitfield and contains one bit set for every
1819 * supported memory type for the resource. The bit `1<<i` is set if and
1820 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1821 * structure for the physical device is supported.
1823 * All types are currently supported for images.
1825 uint32_t memory_types
= (1ull << pdevice
->memory
.type_count
) - 1;
1827 pMemoryRequirements
->size
= image
->size
;
1828 pMemoryRequirements
->alignment
= image
->alignment
;
1829 pMemoryRequirements
->memoryTypeBits
= memory_types
;
1832 void anv_GetImageMemoryRequirements2KHR(
1834 const VkImageMemoryRequirementsInfo2KHR
* pInfo
,
1835 VkMemoryRequirements2KHR
* pMemoryRequirements
)
1837 anv_GetImageMemoryRequirements(_device
, pInfo
->image
,
1838 &pMemoryRequirements
->memoryRequirements
);
1840 vk_foreach_struct_const(ext
, pInfo
->pNext
) {
1841 switch (ext
->sType
) {
1842 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO_KHR
: {
1843 ANV_FROM_HANDLE(anv_image
, image
, pInfo
->image
);
1844 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1845 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
1846 const VkImagePlaneMemoryRequirementsInfoKHR
*plane_reqs
=
1847 (const VkImagePlaneMemoryRequirementsInfoKHR
*) ext
;
1848 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
,
1849 plane_reqs
->planeAspect
);
1851 assert(image
->planes
[plane
].offset
== 0);
1853 /* The Vulkan spec (git aaed022) says:
1855 * memoryTypeBits is a bitfield and contains one bit set for every
1856 * supported memory type for the resource. The bit `1<<i` is set
1857 * if and only if the memory type `i` in the
1858 * VkPhysicalDeviceMemoryProperties structure for the physical
1859 * device is supported.
1861 * All types are currently supported for images.
1863 pMemoryRequirements
->memoryRequirements
.memoryTypeBits
=
1864 (1ull << pdevice
->memory
.type_count
) - 1;
1866 pMemoryRequirements
->memoryRequirements
.size
= image
->planes
[plane
].size
;
1867 pMemoryRequirements
->memoryRequirements
.alignment
=
1868 image
->planes
[plane
].alignment
;
1873 anv_debug_ignored_stype(ext
->sType
);
1878 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
1879 switch (ext
->sType
) {
1880 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS_KHR
: {
1881 VkMemoryDedicatedRequirementsKHR
*requirements
= (void *)ext
;
1882 requirements
->prefersDedicatedAllocation
= VK_FALSE
;
1883 requirements
->requiresDedicatedAllocation
= VK_FALSE
;
1888 anv_debug_ignored_stype(ext
->sType
);
1894 void anv_GetImageSparseMemoryRequirements(
1897 uint32_t* pSparseMemoryRequirementCount
,
1898 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
1900 *pSparseMemoryRequirementCount
= 0;
1903 void anv_GetImageSparseMemoryRequirements2KHR(
1905 const VkImageSparseMemoryRequirementsInfo2KHR
* pInfo
,
1906 uint32_t* pSparseMemoryRequirementCount
,
1907 VkSparseImageMemoryRequirements2KHR
* pSparseMemoryRequirements
)
1909 *pSparseMemoryRequirementCount
= 0;
1912 void anv_GetDeviceMemoryCommitment(
1914 VkDeviceMemory memory
,
1915 VkDeviceSize
* pCommittedMemoryInBytes
)
1917 *pCommittedMemoryInBytes
= 0;
1921 anv_bind_buffer_memory(const VkBindBufferMemoryInfoKHR
*pBindInfo
)
1923 ANV_FROM_HANDLE(anv_device_memory
, mem
, pBindInfo
->memory
);
1924 ANV_FROM_HANDLE(anv_buffer
, buffer
, pBindInfo
->buffer
);
1926 assert(pBindInfo
->sType
== VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO_KHR
);
1929 assert((buffer
->usage
& mem
->type
->valid_buffer_usage
) == buffer
->usage
);
1930 buffer
->bo
= mem
->bo
;
1931 buffer
->offset
= pBindInfo
->memoryOffset
;
1938 VkResult
anv_BindBufferMemory(
1941 VkDeviceMemory memory
,
1942 VkDeviceSize memoryOffset
)
1944 anv_bind_buffer_memory(
1945 &(VkBindBufferMemoryInfoKHR
) {
1946 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO_KHR
,
1949 .memoryOffset
= memoryOffset
,
1955 VkResult
anv_BindBufferMemory2KHR(
1957 uint32_t bindInfoCount
,
1958 const VkBindBufferMemoryInfoKHR
* pBindInfos
)
1960 for (uint32_t i
= 0; i
< bindInfoCount
; i
++)
1961 anv_bind_buffer_memory(&pBindInfos
[i
]);
1966 VkResult
anv_QueueBindSparse(
1968 uint32_t bindInfoCount
,
1969 const VkBindSparseInfo
* pBindInfo
,
1972 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
1973 if (unlikely(queue
->device
->lost
))
1974 return VK_ERROR_DEVICE_LOST
;
1976 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
1981 VkResult
anv_CreateEvent(
1983 const VkEventCreateInfo
* pCreateInfo
,
1984 const VkAllocationCallbacks
* pAllocator
,
1987 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1988 struct anv_state state
;
1989 struct anv_event
*event
;
1991 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_EVENT_CREATE_INFO
);
1993 state
= anv_state_pool_alloc(&device
->dynamic_state_pool
,
1996 event
->state
= state
;
1997 event
->semaphore
= VK_EVENT_RESET
;
1999 if (!device
->info
.has_llc
) {
2000 /* Make sure the writes we're flushing have landed. */
2001 __builtin_ia32_mfence();
2002 __builtin_ia32_clflush(event
);
2005 *pEvent
= anv_event_to_handle(event
);
2010 void anv_DestroyEvent(
2013 const VkAllocationCallbacks
* pAllocator
)
2015 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2016 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2021 anv_state_pool_free(&device
->dynamic_state_pool
, event
->state
);
2024 VkResult
anv_GetEventStatus(
2028 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2029 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2031 if (unlikely(device
->lost
))
2032 return VK_ERROR_DEVICE_LOST
;
2034 if (!device
->info
.has_llc
) {
2035 /* Invalidate read cache before reading event written by GPU. */
2036 __builtin_ia32_clflush(event
);
2037 __builtin_ia32_mfence();
2041 return event
->semaphore
;
2044 VkResult
anv_SetEvent(
2048 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2049 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2051 event
->semaphore
= VK_EVENT_SET
;
2053 if (!device
->info
.has_llc
) {
2054 /* Make sure the writes we're flushing have landed. */
2055 __builtin_ia32_mfence();
2056 __builtin_ia32_clflush(event
);
2062 VkResult
anv_ResetEvent(
2066 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2067 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2069 event
->semaphore
= VK_EVENT_RESET
;
2071 if (!device
->info
.has_llc
) {
2072 /* Make sure the writes we're flushing have landed. */
2073 __builtin_ia32_mfence();
2074 __builtin_ia32_clflush(event
);
2082 VkResult
anv_CreateBuffer(
2084 const VkBufferCreateInfo
* pCreateInfo
,
2085 const VkAllocationCallbacks
* pAllocator
,
2088 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2089 struct anv_buffer
*buffer
;
2091 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
2093 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
2094 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2096 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2098 buffer
->size
= pCreateInfo
->size
;
2099 buffer
->usage
= pCreateInfo
->usage
;
2103 *pBuffer
= anv_buffer_to_handle(buffer
);
2108 void anv_DestroyBuffer(
2111 const VkAllocationCallbacks
* pAllocator
)
2113 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2114 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
2119 vk_free2(&device
->alloc
, pAllocator
, buffer
);
2123 anv_fill_buffer_surface_state(struct anv_device
*device
, struct anv_state state
,
2124 enum isl_format format
,
2125 uint32_t offset
, uint32_t range
, uint32_t stride
)
2127 isl_buffer_fill_state(&device
->isl_dev
, state
.map
,
2129 .mocs
= device
->default_mocs
,
2134 anv_state_flush(device
, state
);
2137 void anv_DestroySampler(
2140 const VkAllocationCallbacks
* pAllocator
)
2142 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2143 ANV_FROM_HANDLE(anv_sampler
, sampler
, _sampler
);
2148 vk_free2(&device
->alloc
, pAllocator
, sampler
);
2151 VkResult
anv_CreateFramebuffer(
2153 const VkFramebufferCreateInfo
* pCreateInfo
,
2154 const VkAllocationCallbacks
* pAllocator
,
2155 VkFramebuffer
* pFramebuffer
)
2157 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2158 struct anv_framebuffer
*framebuffer
;
2160 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
2162 size_t size
= sizeof(*framebuffer
) +
2163 sizeof(struct anv_image_view
*) * pCreateInfo
->attachmentCount
;
2164 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
2165 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2166 if (framebuffer
== NULL
)
2167 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2169 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
2170 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
2171 VkImageView _iview
= pCreateInfo
->pAttachments
[i
];
2172 framebuffer
->attachments
[i
] = anv_image_view_from_handle(_iview
);
2175 framebuffer
->width
= pCreateInfo
->width
;
2176 framebuffer
->height
= pCreateInfo
->height
;
2177 framebuffer
->layers
= pCreateInfo
->layers
;
2179 *pFramebuffer
= anv_framebuffer_to_handle(framebuffer
);
2184 void anv_DestroyFramebuffer(
2187 const VkAllocationCallbacks
* pAllocator
)
2189 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2190 ANV_FROM_HANDLE(anv_framebuffer
, fb
, _fb
);
2195 vk_free2(&device
->alloc
, pAllocator
, fb
);
2198 /* vk_icd.h does not declare this function, so we declare it here to
2199 * suppress Wmissing-prototypes.
2201 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2202 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
);
2204 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2205 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
)
2207 /* For the full details on loader interface versioning, see
2208 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
2209 * What follows is a condensed summary, to help you navigate the large and
2210 * confusing official doc.
2212 * - Loader interface v0 is incompatible with later versions. We don't
2215 * - In loader interface v1:
2216 * - The first ICD entrypoint called by the loader is
2217 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
2219 * - The ICD must statically expose no other Vulkan symbol unless it is
2220 * linked with -Bsymbolic.
2221 * - Each dispatchable Vulkan handle created by the ICD must be
2222 * a pointer to a struct whose first member is VK_LOADER_DATA. The
2223 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
2224 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
2225 * vkDestroySurfaceKHR(). The ICD must be capable of working with
2226 * such loader-managed surfaces.
2228 * - Loader interface v2 differs from v1 in:
2229 * - The first ICD entrypoint called by the loader is
2230 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
2231 * statically expose this entrypoint.
2233 * - Loader interface v3 differs from v2 in:
2234 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
2235 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
2236 * because the loader no longer does so.
2238 *pSupportedVersion
= MIN2(*pSupportedVersion
, 3u);