2 * Copyright © 2015 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
28 #include <sys/sysinfo.h>
32 #include <drm_fourcc.h>
34 #include "anv_private.h"
35 #include "util/strtod.h"
36 #include "util/debug.h"
37 #include "util/build_id.h"
38 #include "util/mesa-sha1.h"
41 #include "genxml/gen7_pack.h"
44 compiler_debug_log(void *data
, const char *fmt
, ...)
48 compiler_perf_log(void *data
, const char *fmt
, ...)
53 if (unlikely(INTEL_DEBUG
& DEBUG_PERF
))
54 intel_logd_v(fmt
, args
);
60 anv_compute_heap_size(int fd
, uint64_t *heap_size
)
63 if (anv_gem_get_context_param(fd
, 0, I915_CONTEXT_PARAM_GTT_SIZE
,
65 /* If, for whatever reason, we can't actually get the GTT size from the
66 * kernel (too old?) fall back to the aperture size.
68 anv_perf_warn(NULL
, NULL
,
69 "Failed to get I915_CONTEXT_PARAM_GTT_SIZE: %m");
71 if (anv_gem_get_aperture(fd
, >t_size
) == -1) {
72 return vk_errorf(NULL
, NULL
, VK_ERROR_INITIALIZATION_FAILED
,
73 "failed to get aperture size: %m");
77 /* Query the total ram from the system */
81 uint64_t total_ram
= (uint64_t)info
.totalram
* (uint64_t)info
.mem_unit
;
83 /* We don't want to burn too much ram with the GPU. If the user has 4GiB
84 * or less, we use at most half. If they have more than 4GiB, we use 3/4.
86 uint64_t available_ram
;
87 if (total_ram
<= 4ull * 1024ull * 1024ull * 1024ull)
88 available_ram
= total_ram
/ 2;
90 available_ram
= total_ram
* 3 / 4;
92 /* We also want to leave some padding for things we allocate in the driver,
93 * so don't go over 3/4 of the GTT either.
95 uint64_t available_gtt
= gtt_size
* 3 / 4;
97 *heap_size
= MIN2(available_ram
, available_gtt
);
103 anv_physical_device_init_heaps(struct anv_physical_device
*device
, int fd
)
105 /* The kernel query only tells us whether or not the kernel supports the
106 * EXEC_OBJECT_SUPPORTS_48B_ADDRESS flag and not whether or not the
107 * hardware has actual 48bit address support.
109 device
->supports_48bit_addresses
=
110 (device
->info
.gen
>= 8) && anv_gem_supports_48b_addresses(fd
);
113 VkResult result
= anv_compute_heap_size(fd
, &heap_size
);
114 if (result
!= VK_SUCCESS
)
117 if (heap_size
> (2ull << 30) && !device
->supports_48bit_addresses
) {
118 /* When running with an overridden PCI ID, we may get a GTT size from
119 * the kernel that is greater than 2 GiB but the execbuf check for 48bit
120 * address support can still fail. Just clamp the address space size to
121 * 2 GiB if we don't have 48-bit support.
123 intel_logw("%s:%d: The kernel reported a GTT size larger than 2 GiB but "
124 "not support for 48-bit addresses",
126 heap_size
= 2ull << 30;
129 if (heap_size
<= 3ull * (1ull << 30)) {
130 /* In this case, everything fits nicely into the 32-bit address space,
131 * so there's no need for supporting 48bit addresses on client-allocated
134 device
->memory
.heap_count
= 1;
135 device
->memory
.heaps
[0] = (struct anv_memory_heap
) {
137 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
138 .supports_48bit_addresses
= false,
141 /* Not everything will fit nicely into a 32-bit address space. In this
142 * case we need a 64-bit heap. Advertise a small 32-bit heap and a
143 * larger 48-bit heap. If we're in this case, then we have a total heap
144 * size larger than 3GiB which most likely means they have 8 GiB of
145 * video memory and so carving off 1 GiB for the 32-bit heap should be
148 const uint64_t heap_size_32bit
= 1ull << 30;
149 const uint64_t heap_size_48bit
= heap_size
- heap_size_32bit
;
151 assert(device
->supports_48bit_addresses
);
153 device
->memory
.heap_count
= 2;
154 device
->memory
.heaps
[0] = (struct anv_memory_heap
) {
155 .size
= heap_size_48bit
,
156 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
157 .supports_48bit_addresses
= true,
159 device
->memory
.heaps
[1] = (struct anv_memory_heap
) {
160 .size
= heap_size_32bit
,
161 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
162 .supports_48bit_addresses
= false,
166 uint32_t type_count
= 0;
167 for (uint32_t heap
= 0; heap
< device
->memory
.heap_count
; heap
++) {
168 uint32_t valid_buffer_usage
= ~0;
170 /* There appears to be a hardware issue in the VF cache where it only
171 * considers the bottom 32 bits of memory addresses. If you happen to
172 * have two vertex buffers which get placed exactly 4 GiB apart and use
173 * them in back-to-back draw calls, you can get collisions. In order to
174 * solve this problem, we require vertex and index buffers be bound to
175 * memory allocated out of the 32-bit heap.
177 if (device
->memory
.heaps
[heap
].supports_48bit_addresses
) {
178 valid_buffer_usage
&= ~(VK_BUFFER_USAGE_INDEX_BUFFER_BIT
|
179 VK_BUFFER_USAGE_VERTEX_BUFFER_BIT
);
182 if (device
->info
.has_llc
) {
183 /* Big core GPUs share LLC with the CPU and thus one memory type can be
184 * both cached and coherent at the same time.
186 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
187 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
188 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
189 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
190 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
192 .valid_buffer_usage
= valid_buffer_usage
,
195 /* The spec requires that we expose a host-visible, coherent memory
196 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
197 * to give the application a choice between cached, but not coherent and
198 * coherent but uncached (WC though).
200 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
201 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
202 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
203 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
205 .valid_buffer_usage
= valid_buffer_usage
,
207 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
208 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
209 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
210 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
212 .valid_buffer_usage
= valid_buffer_usage
,
216 device
->memory
.type_count
= type_count
;
222 anv_physical_device_init_uuids(struct anv_physical_device
*device
)
224 const struct build_id_note
*note
=
225 build_id_find_nhdr_for_addr(anv_physical_device_init_uuids
);
227 return vk_errorf(device
->instance
, device
,
228 VK_ERROR_INITIALIZATION_FAILED
,
229 "Failed to find build-id");
232 unsigned build_id_len
= build_id_length(note
);
233 if (build_id_len
< 20) {
234 return vk_errorf(device
->instance
, device
,
235 VK_ERROR_INITIALIZATION_FAILED
,
236 "build-id too short. It needs to be a SHA");
239 struct mesa_sha1 sha1_ctx
;
241 STATIC_ASSERT(VK_UUID_SIZE
<= sizeof(sha1
));
243 /* The pipeline cache UUID is used for determining when a pipeline cache is
244 * invalid. It needs both a driver build and the PCI ID of the device.
246 _mesa_sha1_init(&sha1_ctx
);
247 _mesa_sha1_update(&sha1_ctx
, build_id_data(note
), build_id_len
);
248 _mesa_sha1_update(&sha1_ctx
, &device
->chipset_id
,
249 sizeof(device
->chipset_id
));
250 _mesa_sha1_final(&sha1_ctx
, sha1
);
251 memcpy(device
->pipeline_cache_uuid
, sha1
, VK_UUID_SIZE
);
253 /* The driver UUID is used for determining sharability of images and memory
254 * between two Vulkan instances in separate processes. People who want to
255 * share memory need to also check the device UUID (below) so all this
256 * needs to be is the build-id.
258 memcpy(device
->driver_uuid
, build_id_data(note
), VK_UUID_SIZE
);
260 /* The device UUID uniquely identifies the given device within the machine.
261 * Since we never have more than one device, this doesn't need to be a real
262 * UUID. However, on the off-chance that someone tries to use this to
263 * cache pre-tiled images or something of the like, we use the PCI ID and
264 * some bits of ISL info to ensure that this is safe.
266 _mesa_sha1_init(&sha1_ctx
);
267 _mesa_sha1_update(&sha1_ctx
, &device
->chipset_id
,
268 sizeof(device
->chipset_id
));
269 _mesa_sha1_update(&sha1_ctx
, &device
->isl_dev
.has_bit6_swizzling
,
270 sizeof(device
->isl_dev
.has_bit6_swizzling
));
271 _mesa_sha1_final(&sha1_ctx
, sha1
);
272 memcpy(device
->device_uuid
, sha1
, VK_UUID_SIZE
);
278 anv_physical_device_init(struct anv_physical_device
*device
,
279 struct anv_instance
*instance
,
285 brw_process_intel_debug_variable();
287 fd
= open(path
, O_RDWR
| O_CLOEXEC
);
289 return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
291 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
292 device
->instance
= instance
;
294 assert(strlen(path
) < ARRAY_SIZE(device
->path
));
295 strncpy(device
->path
, path
, ARRAY_SIZE(device
->path
));
297 device
->chipset_id
= anv_gem_get_param(fd
, I915_PARAM_CHIPSET_ID
);
298 if (!device
->chipset_id
) {
299 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
303 device
->name
= gen_get_device_name(device
->chipset_id
);
304 if (!gen_get_device_info(device
->chipset_id
, &device
->info
)) {
305 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
309 if (device
->info
.is_haswell
) {
310 intel_logw("Haswell Vulkan support is incomplete");
311 } else if (device
->info
.gen
== 7 && !device
->info
.is_baytrail
) {
312 intel_logw("Ivy Bridge Vulkan support is incomplete");
313 } else if (device
->info
.gen
== 7 && device
->info
.is_baytrail
) {
314 intel_logw("Bay Trail Vulkan support is incomplete");
315 } else if (device
->info
.gen
>= 8 && device
->info
.gen
<= 9) {
316 /* Broadwell, Cherryview, Skylake, Broxton, Kabylake, Coffelake is as
317 * fully supported as anything */
318 } else if (device
->info
.gen
== 10) {
319 intel_logw("Cannonlake Vulkan support is alpha");
321 result
= vk_errorf(device
->instance
, device
,
322 VK_ERROR_INCOMPATIBLE_DRIVER
,
323 "Vulkan not yet supported on %s", device
->name
);
327 device
->cmd_parser_version
= -1;
328 if (device
->info
.gen
== 7) {
329 device
->cmd_parser_version
=
330 anv_gem_get_param(fd
, I915_PARAM_CMD_PARSER_VERSION
);
331 if (device
->cmd_parser_version
== -1) {
332 result
= vk_errorf(device
->instance
, device
,
333 VK_ERROR_INITIALIZATION_FAILED
,
334 "failed to get command parser version");
339 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_WAIT_TIMEOUT
)) {
340 result
= vk_errorf(device
->instance
, device
,
341 VK_ERROR_INITIALIZATION_FAILED
,
342 "kernel missing gem wait");
346 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_EXECBUF2
)) {
347 result
= vk_errorf(device
->instance
, device
,
348 VK_ERROR_INITIALIZATION_FAILED
,
349 "kernel missing execbuf2");
353 if (!device
->info
.has_llc
&&
354 anv_gem_get_param(fd
, I915_PARAM_MMAP_VERSION
) < 1) {
355 result
= vk_errorf(device
->instance
, device
,
356 VK_ERROR_INITIALIZATION_FAILED
,
357 "kernel missing wc mmap");
361 result
= anv_physical_device_init_heaps(device
, fd
);
362 if (result
!= VK_SUCCESS
)
365 device
->has_exec_async
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_ASYNC
);
366 device
->has_exec_capture
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_CAPTURE
);
367 device
->has_exec_fence
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE
);
368 device
->has_syncobj
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE_ARRAY
);
369 device
->has_syncobj_wait
= device
->has_syncobj
&&
370 anv_gem_supports_syncobj_wait(fd
);
372 bool swizzled
= anv_gem_get_bit6_swizzle(fd
, I915_TILING_X
);
374 /* Starting with Gen10, the timestamp frequency of the command streamer may
375 * vary from one part to another. We can query the value from the kernel.
377 if (device
->info
.gen
>= 10) {
378 int timestamp_frequency
=
379 anv_gem_get_param(fd
, I915_PARAM_CS_TIMESTAMP_FREQUENCY
);
381 if (timestamp_frequency
< 0)
382 intel_logw("Kernel 4.16-rc1+ required to properly query CS timestamp frequency");
384 device
->info
.timestamp_frequency
= timestamp_frequency
;
387 /* GENs prior to 8 do not support EU/Subslice info */
388 if (device
->info
.gen
>= 8) {
389 device
->subslice_total
= anv_gem_get_param(fd
, I915_PARAM_SUBSLICE_TOTAL
);
390 device
->eu_total
= anv_gem_get_param(fd
, I915_PARAM_EU_TOTAL
);
392 /* Without this information, we cannot get the right Braswell
393 * brandstrings, and we have to use conservative numbers for GPGPU on
394 * many platforms, but otherwise, things will just work.
396 if (device
->subslice_total
< 1 || device
->eu_total
< 1) {
397 intel_logw("Kernel 4.1 required to properly query GPU properties");
399 } else if (device
->info
.gen
== 7) {
400 device
->subslice_total
= 1 << (device
->info
.gt
- 1);
403 if (device
->info
.is_cherryview
&&
404 device
->subslice_total
> 0 && device
->eu_total
> 0) {
405 /* Logical CS threads = EUs per subslice * num threads per EU */
406 uint32_t max_cs_threads
=
407 device
->eu_total
/ device
->subslice_total
* device
->info
.num_thread_per_eu
;
409 /* Fuse configurations may give more threads than expected, never less. */
410 if (max_cs_threads
> device
->info
.max_cs_threads
)
411 device
->info
.max_cs_threads
= max_cs_threads
;
414 device
->compiler
= brw_compiler_create(NULL
, &device
->info
);
415 if (device
->compiler
== NULL
) {
416 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
419 device
->compiler
->shader_debug_log
= compiler_debug_log
;
420 device
->compiler
->shader_perf_log
= compiler_perf_log
;
421 device
->compiler
->supports_pull_constants
= false;
423 isl_device_init(&device
->isl_dev
, &device
->info
, swizzled
);
425 result
= anv_physical_device_init_uuids(device
);
426 if (result
!= VK_SUCCESS
)
429 result
= anv_init_wsi(device
);
430 if (result
!= VK_SUCCESS
) {
431 ralloc_free(device
->compiler
);
435 device
->local_fd
= fd
;
444 anv_physical_device_finish(struct anv_physical_device
*device
)
446 anv_finish_wsi(device
);
447 ralloc_free(device
->compiler
);
448 close(device
->local_fd
);
452 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
453 VkSystemAllocationScope allocationScope
)
459 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
460 size_t align
, VkSystemAllocationScope allocationScope
)
462 return realloc(pOriginal
, size
);
466 default_free_func(void *pUserData
, void *pMemory
)
471 static const VkAllocationCallbacks default_alloc
= {
473 .pfnAllocation
= default_alloc_func
,
474 .pfnReallocation
= default_realloc_func
,
475 .pfnFree
= default_free_func
,
478 VkResult
anv_CreateInstance(
479 const VkInstanceCreateInfo
* pCreateInfo
,
480 const VkAllocationCallbacks
* pAllocator
,
481 VkInstance
* pInstance
)
483 struct anv_instance
*instance
;
485 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
487 /* Check if user passed a debug report callback to be used during
488 * Create/Destroy of instance.
490 const VkDebugReportCallbackCreateInfoEXT
*ctor_cb
=
491 vk_find_struct_const(pCreateInfo
->pNext
,
492 DEBUG_REPORT_CALLBACK_CREATE_INFO_EXT
);
494 uint32_t client_version
;
495 if (pCreateInfo
->pApplicationInfo
&&
496 pCreateInfo
->pApplicationInfo
->apiVersion
!= 0) {
497 client_version
= pCreateInfo
->pApplicationInfo
->apiVersion
;
499 client_version
= VK_MAKE_VERSION(1, 0, 0);
502 if (VK_MAKE_VERSION(1, 0, 0) > client_version
||
503 client_version
> VK_MAKE_VERSION(1, 0, 0xfff)) {
505 if (ctor_cb
&& ctor_cb
->flags
& VK_DEBUG_REPORT_ERROR_BIT_EXT
)
506 ctor_cb
->pfnCallback(VK_DEBUG_REPORT_ERROR_BIT_EXT
,
507 VK_DEBUG_REPORT_OBJECT_TYPE_INSTANCE_EXT
,
508 VK_NULL_HANDLE
, /* No handle available yet. */
512 "incompatible driver version",
515 return vk_errorf(NULL
, NULL
, VK_ERROR_INCOMPATIBLE_DRIVER
,
516 "Client requested version %d.%d.%d",
517 VK_VERSION_MAJOR(client_version
),
518 VK_VERSION_MINOR(client_version
),
519 VK_VERSION_PATCH(client_version
));
522 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
523 const char *ext_name
= pCreateInfo
->ppEnabledExtensionNames
[i
];
524 if (!anv_instance_extension_supported(ext_name
))
525 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
528 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
529 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
531 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
533 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
536 instance
->alloc
= *pAllocator
;
538 instance
->alloc
= default_alloc
;
540 instance
->apiVersion
= client_version
;
541 instance
->physicalDeviceCount
= -1;
543 if (pthread_mutex_init(&instance
->callbacks_mutex
, NULL
) != 0) {
544 vk_free2(&default_alloc
, pAllocator
, instance
);
545 return vk_error(VK_ERROR_INITIALIZATION_FAILED
);
548 list_inithead(&instance
->callbacks
);
550 /* Store report debug callback to be used during DestroyInstance. */
552 instance
->destroy_debug_cb
.flags
= ctor_cb
->flags
;
553 instance
->destroy_debug_cb
.callback
= ctor_cb
->pfnCallback
;
554 instance
->destroy_debug_cb
.data
= ctor_cb
->pUserData
;
559 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
561 *pInstance
= anv_instance_to_handle(instance
);
566 void anv_DestroyInstance(
567 VkInstance _instance
,
568 const VkAllocationCallbacks
* pAllocator
)
570 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
575 if (instance
->physicalDeviceCount
> 0) {
576 /* We support at most one physical device. */
577 assert(instance
->physicalDeviceCount
== 1);
578 anv_physical_device_finish(&instance
->physicalDevice
);
581 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
583 pthread_mutex_destroy(&instance
->callbacks_mutex
);
587 vk_free(&instance
->alloc
, instance
);
591 anv_enumerate_devices(struct anv_instance
*instance
)
593 /* TODO: Check for more devices ? */
594 drmDevicePtr devices
[8];
595 VkResult result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
598 instance
->physicalDeviceCount
= 0;
600 max_devices
= drmGetDevices2(0, devices
, ARRAY_SIZE(devices
));
602 return VK_ERROR_INCOMPATIBLE_DRIVER
;
604 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
605 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
606 devices
[i
]->bustype
== DRM_BUS_PCI
&&
607 devices
[i
]->deviceinfo
.pci
->vendor_id
== 0x8086) {
609 result
= anv_physical_device_init(&instance
->physicalDevice
,
611 devices
[i
]->nodes
[DRM_NODE_RENDER
]);
612 if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
616 drmFreeDevices(devices
, max_devices
);
618 if (result
== VK_SUCCESS
)
619 instance
->physicalDeviceCount
= 1;
625 VkResult
anv_EnumeratePhysicalDevices(
626 VkInstance _instance
,
627 uint32_t* pPhysicalDeviceCount
,
628 VkPhysicalDevice
* pPhysicalDevices
)
630 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
631 VK_OUTARRAY_MAKE(out
, pPhysicalDevices
, pPhysicalDeviceCount
);
634 if (instance
->physicalDeviceCount
< 0) {
635 result
= anv_enumerate_devices(instance
);
636 if (result
!= VK_SUCCESS
&&
637 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
641 if (instance
->physicalDeviceCount
> 0) {
642 assert(instance
->physicalDeviceCount
== 1);
643 vk_outarray_append(&out
, i
) {
644 *i
= anv_physical_device_to_handle(&instance
->physicalDevice
);
648 return vk_outarray_status(&out
);
651 void anv_GetPhysicalDeviceFeatures(
652 VkPhysicalDevice physicalDevice
,
653 VkPhysicalDeviceFeatures
* pFeatures
)
655 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
657 *pFeatures
= (VkPhysicalDeviceFeatures
) {
658 .robustBufferAccess
= true,
659 .fullDrawIndexUint32
= true,
660 .imageCubeArray
= true,
661 .independentBlend
= true,
662 .geometryShader
= true,
663 .tessellationShader
= true,
664 .sampleRateShading
= true,
665 .dualSrcBlend
= true,
667 .multiDrawIndirect
= true,
668 .drawIndirectFirstInstance
= true,
670 .depthBiasClamp
= true,
671 .fillModeNonSolid
= true,
672 .depthBounds
= false,
676 .multiViewport
= true,
677 .samplerAnisotropy
= true,
678 .textureCompressionETC2
= pdevice
->info
.gen
>= 8 ||
679 pdevice
->info
.is_baytrail
,
680 .textureCompressionASTC_LDR
= pdevice
->info
.gen
>= 9, /* FINISHME CHV */
681 .textureCompressionBC
= true,
682 .occlusionQueryPrecise
= true,
683 .pipelineStatisticsQuery
= true,
684 .fragmentStoresAndAtomics
= true,
685 .shaderTessellationAndGeometryPointSize
= true,
686 .shaderImageGatherExtended
= true,
687 .shaderStorageImageExtendedFormats
= true,
688 .shaderStorageImageMultisample
= false,
689 .shaderStorageImageReadWithoutFormat
= false,
690 .shaderStorageImageWriteWithoutFormat
= true,
691 .shaderUniformBufferArrayDynamicIndexing
= true,
692 .shaderSampledImageArrayDynamicIndexing
= true,
693 .shaderStorageBufferArrayDynamicIndexing
= true,
694 .shaderStorageImageArrayDynamicIndexing
= true,
695 .shaderClipDistance
= true,
696 .shaderCullDistance
= true,
697 .shaderFloat64
= pdevice
->info
.gen
>= 8,
698 .shaderInt64
= pdevice
->info
.gen
>= 8,
699 .shaderInt16
= false,
700 .shaderResourceMinLod
= false,
701 .variableMultisampleRate
= false,
702 .inheritedQueries
= true,
705 /* We can't do image stores in vec4 shaders */
706 pFeatures
->vertexPipelineStoresAndAtomics
=
707 pdevice
->compiler
->scalar_stage
[MESA_SHADER_VERTEX
] &&
708 pdevice
->compiler
->scalar_stage
[MESA_SHADER_GEOMETRY
];
711 void anv_GetPhysicalDeviceFeatures2KHR(
712 VkPhysicalDevice physicalDevice
,
713 VkPhysicalDeviceFeatures2KHR
* pFeatures
)
715 anv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
717 vk_foreach_struct(ext
, pFeatures
->pNext
) {
718 switch (ext
->sType
) {
719 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES_KHX
: {
720 VkPhysicalDeviceMultiviewFeaturesKHX
*features
=
721 (VkPhysicalDeviceMultiviewFeaturesKHX
*)ext
;
722 features
->multiview
= true;
723 features
->multiviewGeometryShader
= true;
724 features
->multiviewTessellationShader
= true;
728 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTER_FEATURES_KHR
: {
729 VkPhysicalDeviceVariablePointerFeaturesKHR
*features
= (void *)ext
;
730 features
->variablePointersStorageBuffer
= true;
731 features
->variablePointers
= true;
735 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES_KHR
: {
736 VkPhysicalDeviceSamplerYcbcrConversionFeaturesKHR
*features
=
737 (VkPhysicalDeviceSamplerYcbcrConversionFeaturesKHR
*) ext
;
738 features
->samplerYcbcrConversion
= true;
742 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES_KHR
: {
743 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
745 VkPhysicalDevice16BitStorageFeaturesKHR
*features
=
746 (VkPhysicalDevice16BitStorageFeaturesKHR
*)ext
;
748 features
->storageBuffer16BitAccess
= false;
749 features
->uniformAndStorageBuffer16BitAccess
= false;
750 features
->storagePushConstant16
= false;
751 features
->storageInputOutput16
= false;
756 anv_debug_ignored_stype(ext
->sType
);
762 void anv_GetPhysicalDeviceProperties(
763 VkPhysicalDevice physicalDevice
,
764 VkPhysicalDeviceProperties
* pProperties
)
766 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
767 const struct gen_device_info
*devinfo
= &pdevice
->info
;
769 /* See assertions made when programming the buffer surface state. */
770 const uint32_t max_raw_buffer_sz
= devinfo
->gen
>= 7 ?
771 (1ul << 30) : (1ul << 27);
773 const uint32_t max_samplers
= (devinfo
->gen
>= 8 || devinfo
->is_haswell
) ?
776 VkSampleCountFlags sample_counts
=
777 isl_device_get_sample_counts(&pdevice
->isl_dev
);
779 VkPhysicalDeviceLimits limits
= {
780 .maxImageDimension1D
= (1 << 14),
781 .maxImageDimension2D
= (1 << 14),
782 .maxImageDimension3D
= (1 << 11),
783 .maxImageDimensionCube
= (1 << 14),
784 .maxImageArrayLayers
= (1 << 11),
785 .maxTexelBufferElements
= 128 * 1024 * 1024,
786 .maxUniformBufferRange
= (1ul << 27),
787 .maxStorageBufferRange
= max_raw_buffer_sz
,
788 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
789 .maxMemoryAllocationCount
= UINT32_MAX
,
790 .maxSamplerAllocationCount
= 64 * 1024,
791 .bufferImageGranularity
= 64, /* A cache line */
792 .sparseAddressSpaceSize
= 0,
793 .maxBoundDescriptorSets
= MAX_SETS
,
794 .maxPerStageDescriptorSamplers
= max_samplers
,
795 .maxPerStageDescriptorUniformBuffers
= 64,
796 .maxPerStageDescriptorStorageBuffers
= 64,
797 .maxPerStageDescriptorSampledImages
= max_samplers
,
798 .maxPerStageDescriptorStorageImages
= 64,
799 .maxPerStageDescriptorInputAttachments
= 64,
800 .maxPerStageResources
= 250,
801 .maxDescriptorSetSamplers
= 256,
802 .maxDescriptorSetUniformBuffers
= 256,
803 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
804 .maxDescriptorSetStorageBuffers
= 256,
805 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
806 .maxDescriptorSetSampledImages
= 256,
807 .maxDescriptorSetStorageImages
= 256,
808 .maxDescriptorSetInputAttachments
= 256,
809 .maxVertexInputAttributes
= MAX_VBS
,
810 .maxVertexInputBindings
= MAX_VBS
,
811 .maxVertexInputAttributeOffset
= 2047,
812 .maxVertexInputBindingStride
= 2048,
813 .maxVertexOutputComponents
= 128,
814 .maxTessellationGenerationLevel
= 64,
815 .maxTessellationPatchSize
= 32,
816 .maxTessellationControlPerVertexInputComponents
= 128,
817 .maxTessellationControlPerVertexOutputComponents
= 128,
818 .maxTessellationControlPerPatchOutputComponents
= 128,
819 .maxTessellationControlTotalOutputComponents
= 2048,
820 .maxTessellationEvaluationInputComponents
= 128,
821 .maxTessellationEvaluationOutputComponents
= 128,
822 .maxGeometryShaderInvocations
= 32,
823 .maxGeometryInputComponents
= 64,
824 .maxGeometryOutputComponents
= 128,
825 .maxGeometryOutputVertices
= 256,
826 .maxGeometryTotalOutputComponents
= 1024,
827 .maxFragmentInputComponents
= 128,
828 .maxFragmentOutputAttachments
= 8,
829 .maxFragmentDualSrcAttachments
= 1,
830 .maxFragmentCombinedOutputResources
= 8,
831 .maxComputeSharedMemorySize
= 32768,
832 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
833 .maxComputeWorkGroupInvocations
= 16 * devinfo
->max_cs_threads
,
834 .maxComputeWorkGroupSize
= {
835 16 * devinfo
->max_cs_threads
,
836 16 * devinfo
->max_cs_threads
,
837 16 * devinfo
->max_cs_threads
,
839 .subPixelPrecisionBits
= 4 /* FIXME */,
840 .subTexelPrecisionBits
= 4 /* FIXME */,
841 .mipmapPrecisionBits
= 4 /* FIXME */,
842 .maxDrawIndexedIndexValue
= UINT32_MAX
,
843 .maxDrawIndirectCount
= UINT32_MAX
,
844 .maxSamplerLodBias
= 16,
845 .maxSamplerAnisotropy
= 16,
846 .maxViewports
= MAX_VIEWPORTS
,
847 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
848 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
849 .viewportSubPixelBits
= 13, /* We take a float? */
850 .minMemoryMapAlignment
= 4096, /* A page */
851 .minTexelBufferOffsetAlignment
= 1,
852 /* We need 16 for UBO block reads to work and 32 for push UBOs */
853 .minUniformBufferOffsetAlignment
= 32,
854 .minStorageBufferOffsetAlignment
= 4,
855 .minTexelOffset
= -8,
857 .minTexelGatherOffset
= -32,
858 .maxTexelGatherOffset
= 31,
859 .minInterpolationOffset
= -0.5,
860 .maxInterpolationOffset
= 0.4375,
861 .subPixelInterpolationOffsetBits
= 4,
862 .maxFramebufferWidth
= (1 << 14),
863 .maxFramebufferHeight
= (1 << 14),
864 .maxFramebufferLayers
= (1 << 11),
865 .framebufferColorSampleCounts
= sample_counts
,
866 .framebufferDepthSampleCounts
= sample_counts
,
867 .framebufferStencilSampleCounts
= sample_counts
,
868 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
869 .maxColorAttachments
= MAX_RTS
,
870 .sampledImageColorSampleCounts
= sample_counts
,
871 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
872 .sampledImageDepthSampleCounts
= sample_counts
,
873 .sampledImageStencilSampleCounts
= sample_counts
,
874 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
875 .maxSampleMaskWords
= 1,
876 .timestampComputeAndGraphics
= false,
877 .timestampPeriod
= 1000000000.0 / devinfo
->timestamp_frequency
,
878 .maxClipDistances
= 8,
879 .maxCullDistances
= 8,
880 .maxCombinedClipAndCullDistances
= 8,
881 .discreteQueuePriorities
= 1,
882 .pointSizeRange
= { 0.125, 255.875 },
883 .lineWidthRange
= { 0.0, 7.9921875 },
884 .pointSizeGranularity
= (1.0 / 8.0),
885 .lineWidthGranularity
= (1.0 / 128.0),
886 .strictLines
= false, /* FINISHME */
887 .standardSampleLocations
= true,
888 .optimalBufferCopyOffsetAlignment
= 128,
889 .optimalBufferCopyRowPitchAlignment
= 128,
890 .nonCoherentAtomSize
= 64,
893 *pProperties
= (VkPhysicalDeviceProperties
) {
894 .apiVersion
= anv_physical_device_api_version(pdevice
),
895 .driverVersion
= vk_get_driver_version(),
897 .deviceID
= pdevice
->chipset_id
,
898 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
900 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
903 snprintf(pProperties
->deviceName
, sizeof(pProperties
->deviceName
),
904 "%s", pdevice
->name
);
905 memcpy(pProperties
->pipelineCacheUUID
,
906 pdevice
->pipeline_cache_uuid
, VK_UUID_SIZE
);
909 void anv_GetPhysicalDeviceProperties2KHR(
910 VkPhysicalDevice physicalDevice
,
911 VkPhysicalDeviceProperties2KHR
* pProperties
)
913 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
915 anv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
917 vk_foreach_struct(ext
, pProperties
->pNext
) {
918 switch (ext
->sType
) {
919 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
920 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
921 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
923 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
927 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES_KHR
: {
928 VkPhysicalDeviceIDPropertiesKHR
*id_props
=
929 (VkPhysicalDeviceIDPropertiesKHR
*)ext
;
930 memcpy(id_props
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
931 memcpy(id_props
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
932 /* The LUID is for Windows. */
933 id_props
->deviceLUIDValid
= false;
937 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES_KHX
: {
938 VkPhysicalDeviceMultiviewPropertiesKHX
*properties
=
939 (VkPhysicalDeviceMultiviewPropertiesKHX
*)ext
;
940 properties
->maxMultiviewViewCount
= 16;
941 properties
->maxMultiviewInstanceIndex
= UINT32_MAX
/ 16;
945 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES_KHR
: {
946 VkPhysicalDevicePointClippingPropertiesKHR
*properties
=
947 (VkPhysicalDevicePointClippingPropertiesKHR
*) ext
;
948 properties
->pointClippingBehavior
= VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES_KHR
;
949 anv_finishme("Implement pop-free point clipping");
954 anv_debug_ignored_stype(ext
->sType
);
960 /* We support exactly one queue family. */
961 static const VkQueueFamilyProperties
962 anv_queue_family_properties
= {
963 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
964 VK_QUEUE_COMPUTE_BIT
|
965 VK_QUEUE_TRANSFER_BIT
,
967 .timestampValidBits
= 36, /* XXX: Real value here */
968 .minImageTransferGranularity
= { 1, 1, 1 },
971 void anv_GetPhysicalDeviceQueueFamilyProperties(
972 VkPhysicalDevice physicalDevice
,
974 VkQueueFamilyProperties
* pQueueFamilyProperties
)
976 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pCount
);
978 vk_outarray_append(&out
, p
) {
979 *p
= anv_queue_family_properties
;
983 void anv_GetPhysicalDeviceQueueFamilyProperties2KHR(
984 VkPhysicalDevice physicalDevice
,
985 uint32_t* pQueueFamilyPropertyCount
,
986 VkQueueFamilyProperties2KHR
* pQueueFamilyProperties
)
989 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pQueueFamilyPropertyCount
);
991 vk_outarray_append(&out
, p
) {
992 p
->queueFamilyProperties
= anv_queue_family_properties
;
994 vk_foreach_struct(s
, p
->pNext
) {
995 anv_debug_ignored_stype(s
->sType
);
1000 void anv_GetPhysicalDeviceMemoryProperties(
1001 VkPhysicalDevice physicalDevice
,
1002 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
1004 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
1006 pMemoryProperties
->memoryTypeCount
= physical_device
->memory
.type_count
;
1007 for (uint32_t i
= 0; i
< physical_device
->memory
.type_count
; i
++) {
1008 pMemoryProperties
->memoryTypes
[i
] = (VkMemoryType
) {
1009 .propertyFlags
= physical_device
->memory
.types
[i
].propertyFlags
,
1010 .heapIndex
= physical_device
->memory
.types
[i
].heapIndex
,
1014 pMemoryProperties
->memoryHeapCount
= physical_device
->memory
.heap_count
;
1015 for (uint32_t i
= 0; i
< physical_device
->memory
.heap_count
; i
++) {
1016 pMemoryProperties
->memoryHeaps
[i
] = (VkMemoryHeap
) {
1017 .size
= physical_device
->memory
.heaps
[i
].size
,
1018 .flags
= physical_device
->memory
.heaps
[i
].flags
,
1023 void anv_GetPhysicalDeviceMemoryProperties2KHR(
1024 VkPhysicalDevice physicalDevice
,
1025 VkPhysicalDeviceMemoryProperties2KHR
* pMemoryProperties
)
1027 anv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
1028 &pMemoryProperties
->memoryProperties
);
1030 vk_foreach_struct(ext
, pMemoryProperties
->pNext
) {
1031 switch (ext
->sType
) {
1033 anv_debug_ignored_stype(ext
->sType
);
1039 PFN_vkVoidFunction
anv_GetInstanceProcAddr(
1040 VkInstance instance
,
1043 return anv_lookup_entrypoint(NULL
, pName
);
1046 /* With version 1+ of the loader interface the ICD should expose
1047 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
1050 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1051 VkInstance instance
,
1055 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1056 VkInstance instance
,
1059 return anv_GetInstanceProcAddr(instance
, pName
);
1062 PFN_vkVoidFunction
anv_GetDeviceProcAddr(
1066 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1067 return anv_lookup_entrypoint(&device
->info
, pName
);
1071 anv_queue_init(struct anv_device
*device
, struct anv_queue
*queue
)
1073 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1074 queue
->device
= device
;
1075 queue
->pool
= &device
->surface_state_pool
;
1079 anv_queue_finish(struct anv_queue
*queue
)
1083 static struct anv_state
1084 anv_state_pool_emit_data(struct anv_state_pool
*pool
, size_t size
, size_t align
, const void *p
)
1086 struct anv_state state
;
1088 state
= anv_state_pool_alloc(pool
, size
, align
);
1089 memcpy(state
.map
, p
, size
);
1091 anv_state_flush(pool
->block_pool
.device
, state
);
1096 struct gen8_border_color
{
1101 /* Pad out to 64 bytes */
1106 anv_device_init_border_colors(struct anv_device
*device
)
1108 static const struct gen8_border_color border_colors
[] = {
1109 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
1110 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
1111 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
1112 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
1113 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
1114 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
1117 device
->border_colors
= anv_state_pool_emit_data(&device
->dynamic_state_pool
,
1118 sizeof(border_colors
), 64,
1123 anv_device_init_trivial_batch(struct anv_device
*device
)
1125 anv_bo_init_new(&device
->trivial_batch_bo
, device
, 4096);
1127 if (device
->instance
->physicalDevice
.has_exec_async
)
1128 device
->trivial_batch_bo
.flags
|= EXEC_OBJECT_ASYNC
;
1130 void *map
= anv_gem_mmap(device
, device
->trivial_batch_bo
.gem_handle
,
1133 struct anv_batch batch
= {
1139 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
1140 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
1142 if (!device
->info
.has_llc
)
1143 gen_clflush_range(map
, batch
.next
- map
);
1145 anv_gem_munmap(map
, device
->trivial_batch_bo
.size
);
1148 VkResult
anv_CreateDevice(
1149 VkPhysicalDevice physicalDevice
,
1150 const VkDeviceCreateInfo
* pCreateInfo
,
1151 const VkAllocationCallbacks
* pAllocator
,
1154 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
1156 struct anv_device
*device
;
1158 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO
);
1160 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
1161 const char *ext_name
= pCreateInfo
->ppEnabledExtensionNames
[i
];
1162 if (!anv_physical_device_extension_supported(physical_device
, ext_name
))
1163 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
1166 /* Check enabled features */
1167 if (pCreateInfo
->pEnabledFeatures
) {
1168 VkPhysicalDeviceFeatures supported_features
;
1169 anv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
1170 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
1171 VkBool32
*enabled_feature
= (VkBool32
*)pCreateInfo
->pEnabledFeatures
;
1172 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
1173 for (uint32_t i
= 0; i
< num_features
; i
++) {
1174 if (enabled_feature
[i
] && !supported_feature
[i
])
1175 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
1179 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
1181 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
1183 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1185 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1186 device
->instance
= physical_device
->instance
;
1187 device
->chipset_id
= physical_device
->chipset_id
;
1188 device
->lost
= false;
1191 device
->alloc
= *pAllocator
;
1193 device
->alloc
= physical_device
->instance
->alloc
;
1195 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
1196 device
->fd
= open(physical_device
->path
, O_RDWR
| O_CLOEXEC
);
1197 if (device
->fd
== -1) {
1198 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1202 device
->context_id
= anv_gem_create_context(device
);
1203 if (device
->context_id
== -1) {
1204 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1208 device
->info
= physical_device
->info
;
1209 device
->isl_dev
= physical_device
->isl_dev
;
1211 /* On Broadwell and later, we can use batch chaining to more efficiently
1212 * implement growing command buffers. Prior to Haswell, the kernel
1213 * command parser gets in the way and we have to fall back to growing
1216 device
->can_chain_batches
= device
->info
.gen
>= 8;
1218 device
->robust_buffer_access
= pCreateInfo
->pEnabledFeatures
&&
1219 pCreateInfo
->pEnabledFeatures
->robustBufferAccess
;
1221 if (pthread_mutex_init(&device
->mutex
, NULL
) != 0) {
1222 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1223 goto fail_context_id
;
1226 pthread_condattr_t condattr
;
1227 if (pthread_condattr_init(&condattr
) != 0) {
1228 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1231 if (pthread_condattr_setclock(&condattr
, CLOCK_MONOTONIC
) != 0) {
1232 pthread_condattr_destroy(&condattr
);
1233 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1236 if (pthread_cond_init(&device
->queue_submit
, NULL
) != 0) {
1237 pthread_condattr_destroy(&condattr
);
1238 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1241 pthread_condattr_destroy(&condattr
);
1244 (physical_device
->supports_48bit_addresses
? EXEC_OBJECT_SUPPORTS_48B_ADDRESS
: 0) |
1245 (physical_device
->has_exec_async
? EXEC_OBJECT_ASYNC
: 0) |
1246 (physical_device
->has_exec_capture
? EXEC_OBJECT_CAPTURE
: 0);
1248 anv_bo_pool_init(&device
->batch_bo_pool
, device
, bo_flags
);
1250 result
= anv_bo_cache_init(&device
->bo_cache
);
1251 if (result
!= VK_SUCCESS
)
1252 goto fail_batch_bo_pool
;
1254 /* For the state pools we explicitly disable 48bit. */
1255 bo_flags
= physical_device
->has_exec_async
? EXEC_OBJECT_ASYNC
: 0;
1257 result
= anv_state_pool_init(&device
->dynamic_state_pool
, device
, 16384,
1259 if (result
!= VK_SUCCESS
)
1262 result
= anv_state_pool_init(&device
->instruction_state_pool
, device
, 16384,
1264 (physical_device
->has_exec_capture
? EXEC_OBJECT_CAPTURE
: 0));
1265 if (result
!= VK_SUCCESS
)
1266 goto fail_dynamic_state_pool
;
1268 result
= anv_state_pool_init(&device
->surface_state_pool
, device
, 4096,
1270 if (result
!= VK_SUCCESS
)
1271 goto fail_instruction_state_pool
;
1273 result
= anv_bo_init_new(&device
->workaround_bo
, device
, 1024);
1274 if (result
!= VK_SUCCESS
)
1275 goto fail_surface_state_pool
;
1277 anv_device_init_trivial_batch(device
);
1279 anv_scratch_pool_init(device
, &device
->scratch_pool
);
1281 anv_queue_init(device
, &device
->queue
);
1283 switch (device
->info
.gen
) {
1285 if (!device
->info
.is_haswell
)
1286 result
= gen7_init_device_state(device
);
1288 result
= gen75_init_device_state(device
);
1291 result
= gen8_init_device_state(device
);
1294 result
= gen9_init_device_state(device
);
1297 result
= gen10_init_device_state(device
);
1300 /* Shouldn't get here as we don't create physical devices for any other
1302 unreachable("unhandled gen");
1304 if (result
!= VK_SUCCESS
)
1305 goto fail_workaround_bo
;
1307 anv_device_init_blorp(device
);
1309 anv_device_init_border_colors(device
);
1311 *pDevice
= anv_device_to_handle(device
);
1316 anv_queue_finish(&device
->queue
);
1317 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
1318 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
1319 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
1320 fail_surface_state_pool
:
1321 anv_state_pool_finish(&device
->surface_state_pool
);
1322 fail_instruction_state_pool
:
1323 anv_state_pool_finish(&device
->instruction_state_pool
);
1324 fail_dynamic_state_pool
:
1325 anv_state_pool_finish(&device
->dynamic_state_pool
);
1327 anv_bo_cache_finish(&device
->bo_cache
);
1329 anv_bo_pool_finish(&device
->batch_bo_pool
);
1330 pthread_cond_destroy(&device
->queue_submit
);
1332 pthread_mutex_destroy(&device
->mutex
);
1334 anv_gem_destroy_context(device
, device
->context_id
);
1338 vk_free(&device
->alloc
, device
);
1343 void anv_DestroyDevice(
1345 const VkAllocationCallbacks
* pAllocator
)
1347 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1352 anv_device_finish_blorp(device
);
1354 anv_queue_finish(&device
->queue
);
1356 #ifdef HAVE_VALGRIND
1357 /* We only need to free these to prevent valgrind errors. The backing
1358 * BO will go away in a couple of lines so we don't actually leak.
1360 anv_state_pool_free(&device
->dynamic_state_pool
, device
->border_colors
);
1363 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
1365 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
1366 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
1368 anv_gem_close(device
, device
->trivial_batch_bo
.gem_handle
);
1370 anv_state_pool_finish(&device
->surface_state_pool
);
1371 anv_state_pool_finish(&device
->instruction_state_pool
);
1372 anv_state_pool_finish(&device
->dynamic_state_pool
);
1374 anv_bo_cache_finish(&device
->bo_cache
);
1376 anv_bo_pool_finish(&device
->batch_bo_pool
);
1378 pthread_cond_destroy(&device
->queue_submit
);
1379 pthread_mutex_destroy(&device
->mutex
);
1381 anv_gem_destroy_context(device
, device
->context_id
);
1385 vk_free(&device
->alloc
, device
);
1388 VkResult
anv_EnumerateInstanceLayerProperties(
1389 uint32_t* pPropertyCount
,
1390 VkLayerProperties
* pProperties
)
1392 if (pProperties
== NULL
) {
1393 *pPropertyCount
= 0;
1397 /* None supported at this time */
1398 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1401 VkResult
anv_EnumerateDeviceLayerProperties(
1402 VkPhysicalDevice physicalDevice
,
1403 uint32_t* pPropertyCount
,
1404 VkLayerProperties
* pProperties
)
1406 if (pProperties
== NULL
) {
1407 *pPropertyCount
= 0;
1411 /* None supported at this time */
1412 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1415 void anv_GetDeviceQueue(
1417 uint32_t queueNodeIndex
,
1418 uint32_t queueIndex
,
1421 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1423 assert(queueIndex
== 0);
1425 *pQueue
= anv_queue_to_handle(&device
->queue
);
1429 anv_device_query_status(struct anv_device
*device
)
1431 /* This isn't likely as most of the callers of this function already check
1432 * for it. However, it doesn't hurt to check and it potentially lets us
1435 if (unlikely(device
->lost
))
1436 return VK_ERROR_DEVICE_LOST
;
1438 uint32_t active
, pending
;
1439 int ret
= anv_gem_gpu_get_reset_stats(device
, &active
, &pending
);
1441 /* We don't know the real error. */
1442 device
->lost
= true;
1443 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1444 "get_reset_stats failed: %m");
1448 device
->lost
= true;
1449 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1450 "GPU hung on one of our command buffers");
1451 } else if (pending
) {
1452 device
->lost
= true;
1453 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1454 "GPU hung with commands in-flight");
1461 anv_device_bo_busy(struct anv_device
*device
, struct anv_bo
*bo
)
1463 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
1464 * Other usages of the BO (such as on different hardware) will not be
1465 * flagged as "busy" by this ioctl. Use with care.
1467 int ret
= anv_gem_busy(device
, bo
->gem_handle
);
1469 return VK_NOT_READY
;
1470 } else if (ret
== -1) {
1471 /* We don't know the real error. */
1472 device
->lost
= true;
1473 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1474 "gem wait failed: %m");
1477 /* Query for device status after the busy call. If the BO we're checking
1478 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
1479 * client because it clearly doesn't have valid data. Yes, this most
1480 * likely means an ioctl, but we just did an ioctl to query the busy status
1481 * so it's no great loss.
1483 return anv_device_query_status(device
);
1487 anv_device_wait(struct anv_device
*device
, struct anv_bo
*bo
,
1490 int ret
= anv_gem_wait(device
, bo
->gem_handle
, &timeout
);
1491 if (ret
== -1 && errno
== ETIME
) {
1493 } else if (ret
== -1) {
1494 /* We don't know the real error. */
1495 device
->lost
= true;
1496 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1497 "gem wait failed: %m");
1500 /* Query for device status after the wait. If the BO we're waiting on got
1501 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
1502 * because it clearly doesn't have valid data. Yes, this most likely means
1503 * an ioctl, but we just did an ioctl to wait so it's no great loss.
1505 return anv_device_query_status(device
);
1508 VkResult
anv_DeviceWaitIdle(
1511 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1512 if (unlikely(device
->lost
))
1513 return VK_ERROR_DEVICE_LOST
;
1515 struct anv_batch batch
;
1518 batch
.start
= batch
.next
= cmds
;
1519 batch
.end
= (void *) cmds
+ sizeof(cmds
);
1521 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
1522 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
1524 return anv_device_submit_simple_batch(device
, &batch
);
1528 anv_bo_init_new(struct anv_bo
*bo
, struct anv_device
*device
, uint64_t size
)
1530 uint32_t gem_handle
= anv_gem_create(device
, size
);
1532 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
1534 anv_bo_init(bo
, gem_handle
, size
);
1539 VkResult
anv_AllocateMemory(
1541 const VkMemoryAllocateInfo
* pAllocateInfo
,
1542 const VkAllocationCallbacks
* pAllocator
,
1543 VkDeviceMemory
* pMem
)
1545 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1546 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
1547 struct anv_device_memory
*mem
;
1548 VkResult result
= VK_SUCCESS
;
1550 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
1552 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
1553 assert(pAllocateInfo
->allocationSize
> 0);
1555 /* The kernel relocation API has a limitation of a 32-bit delta value
1556 * applied to the address before it is written which, in spite of it being
1557 * unsigned, is treated as signed . Because of the way that this maps to
1558 * the Vulkan API, we cannot handle an offset into a buffer that does not
1559 * fit into a signed 32 bits. The only mechanism we have for dealing with
1560 * this at the moment is to limit all VkDeviceMemory objects to a maximum
1561 * of 2GB each. The Vulkan spec allows us to do this:
1563 * "Some platforms may have a limit on the maximum size of a single
1564 * allocation. For example, certain systems may fail to create
1565 * allocations with a size greater than or equal to 4GB. Such a limit is
1566 * implementation-dependent, and if such a failure occurs then the error
1567 * VK_ERROR_OUT_OF_DEVICE_MEMORY should be returned."
1569 * We don't use vk_error here because it's not an error so much as an
1570 * indication to the application that the allocation is too large.
1572 if (pAllocateInfo
->allocationSize
> (1ull << 31))
1573 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
1575 /* FINISHME: Fail if allocation request exceeds heap size. */
1577 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
1578 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1580 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1582 assert(pAllocateInfo
->memoryTypeIndex
< pdevice
->memory
.type_count
);
1583 mem
->type
= &pdevice
->memory
.types
[pAllocateInfo
->memoryTypeIndex
];
1587 const VkImportMemoryFdInfoKHR
*fd_info
=
1588 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_FD_INFO_KHR
);
1590 /* The Vulkan spec permits handleType to be 0, in which case the struct is
1593 if (fd_info
&& fd_info
->handleType
) {
1594 /* At the moment, we support only the below handle types. */
1595 assert(fd_info
->handleType
==
1596 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
||
1597 fd_info
->handleType
==
1598 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
1600 result
= anv_bo_cache_import(device
, &device
->bo_cache
,
1601 fd_info
->fd
, &mem
->bo
);
1602 if (result
!= VK_SUCCESS
)
1605 VkDeviceSize aligned_alloc_size
=
1606 align_u64(pAllocateInfo
->allocationSize
, 4096);
1608 /* For security purposes, we reject importing the bo if it's smaller
1609 * than the requested allocation size. This prevents a malicious client
1610 * from passing a buffer to a trusted client, lying about the size, and
1611 * telling the trusted client to try and texture from an image that goes
1612 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
1613 * in the trusted client. The trusted client can protect itself against
1614 * this sort of attack but only if it can trust the buffer size.
1616 if (mem
->bo
->size
< aligned_alloc_size
) {
1617 result
= vk_errorf(device
->instance
, device
,
1618 VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
,
1619 "aligned allocationSize too large for "
1620 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR: "
1621 "%"PRIu64
"B > %"PRIu64
"B",
1622 aligned_alloc_size
, mem
->bo
->size
);
1623 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
1627 /* From the Vulkan spec:
1629 * "Importing memory from a file descriptor transfers ownership of
1630 * the file descriptor from the application to the Vulkan
1631 * implementation. The application must not perform any operations on
1632 * the file descriptor after a successful import."
1634 * If the import fails, we leave the file descriptor open.
1638 result
= anv_bo_cache_alloc(device
, &device
->bo_cache
,
1639 pAllocateInfo
->allocationSize
,
1641 if (result
!= VK_SUCCESS
)
1644 const VkMemoryDedicatedAllocateInfoKHR
*dedicated_info
=
1645 vk_find_struct_const(pAllocateInfo
->pNext
, MEMORY_DEDICATED_ALLOCATE_INFO_KHR
);
1646 if (dedicated_info
&& dedicated_info
->image
!= VK_NULL_HANDLE
) {
1647 ANV_FROM_HANDLE(anv_image
, image
, dedicated_info
->image
);
1649 /* For images using modifiers, we require a dedicated allocation
1650 * and we set the BO tiling to match the tiling of the underlying
1651 * modifier. This is a bit unfortunate as this is completely
1652 * pointless for Vulkan. However, GL needs to be able to map things
1653 * so it needs the tiling to be set. The only way to do this in a
1654 * non-racy way is to set the tiling in the creator of the BO so that
1657 * One of these days, once the GL driver learns to not map things
1658 * through the GTT in random places, we can drop this and start
1659 * allowing multiple modified images in the same BO.
1661 if (image
->drm_format_mod
!= DRM_FORMAT_MOD_INVALID
) {
1662 assert(isl_drm_modifier_get_info(image
->drm_format_mod
)->tiling
==
1663 image
->planes
[0].surface
.isl
.tiling
);
1664 const uint32_t i915_tiling
=
1665 isl_tiling_to_i915_tiling(image
->planes
[0].surface
.isl
.tiling
);
1666 int ret
= anv_gem_set_tiling(device
, mem
->bo
->gem_handle
,
1667 image
->planes
[0].surface
.isl
.row_pitch
,
1670 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
1671 return vk_errorf(device
->instance
, NULL
,
1672 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
1673 "failed to set BO tiling: %m");
1679 assert(mem
->type
->heapIndex
< pdevice
->memory
.heap_count
);
1680 if (pdevice
->memory
.heaps
[mem
->type
->heapIndex
].supports_48bit_addresses
)
1681 mem
->bo
->flags
|= EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
1683 const struct wsi_memory_allocate_info
*wsi_info
=
1684 vk_find_struct_const(pAllocateInfo
->pNext
, WSI_MEMORY_ALLOCATE_INFO_MESA
);
1685 if (wsi_info
&& wsi_info
->implicit_sync
) {
1686 /* We need to set the WRITE flag on window system buffers so that GEM
1687 * will know we're writing to them and synchronize uses on other rings
1688 * (eg if the display server uses the blitter ring).
1690 mem
->bo
->flags
|= EXEC_OBJECT_WRITE
;
1691 } else if (pdevice
->has_exec_async
) {
1692 mem
->bo
->flags
|= EXEC_OBJECT_ASYNC
;
1695 *pMem
= anv_device_memory_to_handle(mem
);
1700 vk_free2(&device
->alloc
, pAllocator
, mem
);
1705 VkResult
anv_GetMemoryFdKHR(
1707 const VkMemoryGetFdInfoKHR
* pGetFdInfo
,
1710 ANV_FROM_HANDLE(anv_device
, dev
, device_h
);
1711 ANV_FROM_HANDLE(anv_device_memory
, mem
, pGetFdInfo
->memory
);
1713 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
1715 assert(pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
||
1716 pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
1718 return anv_bo_cache_export(dev
, &dev
->bo_cache
, mem
->bo
, pFd
);
1721 VkResult
anv_GetMemoryFdPropertiesKHR(
1723 VkExternalMemoryHandleTypeFlagBitsKHR handleType
,
1725 VkMemoryFdPropertiesKHR
* pMemoryFdProperties
)
1727 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1728 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
1730 switch (handleType
) {
1731 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
:
1732 /* dma-buf can be imported as any memory type */
1733 pMemoryFdProperties
->memoryTypeBits
=
1734 (1 << pdevice
->memory
.type_count
) - 1;
1738 /* The valid usage section for this function says:
1740 * "handleType must not be one of the handle types defined as
1743 * So opaque handle types fall into the default "unsupported" case.
1745 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
1749 void anv_FreeMemory(
1751 VkDeviceMemory _mem
,
1752 const VkAllocationCallbacks
* pAllocator
)
1754 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1755 ANV_FROM_HANDLE(anv_device_memory
, mem
, _mem
);
1761 anv_UnmapMemory(_device
, _mem
);
1763 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
1765 vk_free2(&device
->alloc
, pAllocator
, mem
);
1768 VkResult
anv_MapMemory(
1770 VkDeviceMemory _memory
,
1771 VkDeviceSize offset
,
1773 VkMemoryMapFlags flags
,
1776 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1777 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
1784 if (size
== VK_WHOLE_SIZE
)
1785 size
= mem
->bo
->size
- offset
;
1787 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
1789 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
1790 * assert(size != 0);
1791 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
1792 * equal to the size of the memory minus offset
1795 assert(offset
+ size
<= mem
->bo
->size
);
1797 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
1798 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
1799 * at a time is valid. We could just mmap up front and return an offset
1800 * pointer here, but that may exhaust virtual memory on 32 bit
1803 uint32_t gem_flags
= 0;
1805 if (!device
->info
.has_llc
&&
1806 (mem
->type
->propertyFlags
& VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
))
1807 gem_flags
|= I915_MMAP_WC
;
1809 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
1810 uint64_t map_offset
= offset
& ~4095ull;
1811 assert(offset
>= map_offset
);
1812 uint64_t map_size
= (offset
+ size
) - map_offset
;
1814 /* Let's map whole pages */
1815 map_size
= align_u64(map_size
, 4096);
1817 void *map
= anv_gem_mmap(device
, mem
->bo
->gem_handle
,
1818 map_offset
, map_size
, gem_flags
);
1819 if (map
== MAP_FAILED
)
1820 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
1823 mem
->map_size
= map_size
;
1825 *ppData
= mem
->map
+ (offset
- map_offset
);
1830 void anv_UnmapMemory(
1832 VkDeviceMemory _memory
)
1834 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
1839 anv_gem_munmap(mem
->map
, mem
->map_size
);
1846 clflush_mapped_ranges(struct anv_device
*device
,
1848 const VkMappedMemoryRange
*ranges
)
1850 for (uint32_t i
= 0; i
< count
; i
++) {
1851 ANV_FROM_HANDLE(anv_device_memory
, mem
, ranges
[i
].memory
);
1852 if (ranges
[i
].offset
>= mem
->map_size
)
1855 gen_clflush_range(mem
->map
+ ranges
[i
].offset
,
1856 MIN2(ranges
[i
].size
, mem
->map_size
- ranges
[i
].offset
));
1860 VkResult
anv_FlushMappedMemoryRanges(
1862 uint32_t memoryRangeCount
,
1863 const VkMappedMemoryRange
* pMemoryRanges
)
1865 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1867 if (device
->info
.has_llc
)
1870 /* Make sure the writes we're flushing have landed. */
1871 __builtin_ia32_mfence();
1873 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
1878 VkResult
anv_InvalidateMappedMemoryRanges(
1880 uint32_t memoryRangeCount
,
1881 const VkMappedMemoryRange
* pMemoryRanges
)
1883 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1885 if (device
->info
.has_llc
)
1888 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
1890 /* Make sure no reads get moved up above the invalidate. */
1891 __builtin_ia32_mfence();
1896 void anv_GetBufferMemoryRequirements(
1899 VkMemoryRequirements
* pMemoryRequirements
)
1901 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
1902 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1903 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
1905 /* The Vulkan spec (git aaed022) says:
1907 * memoryTypeBits is a bitfield and contains one bit set for every
1908 * supported memory type for the resource. The bit `1<<i` is set if and
1909 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1910 * structure for the physical device is supported.
1912 uint32_t memory_types
= 0;
1913 for (uint32_t i
= 0; i
< pdevice
->memory
.type_count
; i
++) {
1914 uint32_t valid_usage
= pdevice
->memory
.types
[i
].valid_buffer_usage
;
1915 if ((valid_usage
& buffer
->usage
) == buffer
->usage
)
1916 memory_types
|= (1u << i
);
1919 /* Base alignment requirement of a cache line */
1920 uint32_t alignment
= 16;
1922 /* We need an alignment of 32 for pushing UBOs */
1923 if (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
)
1924 alignment
= MAX2(alignment
, 32);
1926 pMemoryRequirements
->size
= buffer
->size
;
1927 pMemoryRequirements
->alignment
= alignment
;
1928 pMemoryRequirements
->memoryTypeBits
= memory_types
;
1931 void anv_GetBufferMemoryRequirements2KHR(
1933 const VkBufferMemoryRequirementsInfo2KHR
* pInfo
,
1934 VkMemoryRequirements2KHR
* pMemoryRequirements
)
1936 anv_GetBufferMemoryRequirements(_device
, pInfo
->buffer
,
1937 &pMemoryRequirements
->memoryRequirements
);
1939 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
1940 switch (ext
->sType
) {
1941 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS_KHR
: {
1942 VkMemoryDedicatedRequirementsKHR
*requirements
= (void *)ext
;
1943 requirements
->prefersDedicatedAllocation
= VK_FALSE
;
1944 requirements
->requiresDedicatedAllocation
= VK_FALSE
;
1949 anv_debug_ignored_stype(ext
->sType
);
1955 void anv_GetImageMemoryRequirements(
1958 VkMemoryRequirements
* pMemoryRequirements
)
1960 ANV_FROM_HANDLE(anv_image
, image
, _image
);
1961 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1962 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
1964 /* The Vulkan spec (git aaed022) says:
1966 * memoryTypeBits is a bitfield and contains one bit set for every
1967 * supported memory type for the resource. The bit `1<<i` is set if and
1968 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1969 * structure for the physical device is supported.
1971 * All types are currently supported for images.
1973 uint32_t memory_types
= (1ull << pdevice
->memory
.type_count
) - 1;
1975 pMemoryRequirements
->size
= image
->size
;
1976 pMemoryRequirements
->alignment
= image
->alignment
;
1977 pMemoryRequirements
->memoryTypeBits
= memory_types
;
1980 void anv_GetImageMemoryRequirements2KHR(
1982 const VkImageMemoryRequirementsInfo2KHR
* pInfo
,
1983 VkMemoryRequirements2KHR
* pMemoryRequirements
)
1985 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1986 ANV_FROM_HANDLE(anv_image
, image
, pInfo
->image
);
1988 anv_GetImageMemoryRequirements(_device
, pInfo
->image
,
1989 &pMemoryRequirements
->memoryRequirements
);
1991 vk_foreach_struct_const(ext
, pInfo
->pNext
) {
1992 switch (ext
->sType
) {
1993 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO_KHR
: {
1994 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
1995 const VkImagePlaneMemoryRequirementsInfoKHR
*plane_reqs
=
1996 (const VkImagePlaneMemoryRequirementsInfoKHR
*) ext
;
1997 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
,
1998 plane_reqs
->planeAspect
);
2000 assert(image
->planes
[plane
].offset
== 0);
2002 /* The Vulkan spec (git aaed022) says:
2004 * memoryTypeBits is a bitfield and contains one bit set for every
2005 * supported memory type for the resource. The bit `1<<i` is set
2006 * if and only if the memory type `i` in the
2007 * VkPhysicalDeviceMemoryProperties structure for the physical
2008 * device is supported.
2010 * All types are currently supported for images.
2012 pMemoryRequirements
->memoryRequirements
.memoryTypeBits
=
2013 (1ull << pdevice
->memory
.type_count
) - 1;
2015 pMemoryRequirements
->memoryRequirements
.size
= image
->planes
[plane
].size
;
2016 pMemoryRequirements
->memoryRequirements
.alignment
=
2017 image
->planes
[plane
].alignment
;
2022 anv_debug_ignored_stype(ext
->sType
);
2027 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
2028 switch (ext
->sType
) {
2029 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS_KHR
: {
2030 VkMemoryDedicatedRequirementsKHR
*requirements
= (void *)ext
;
2031 if (image
->drm_format_mod
!= DRM_FORMAT_MOD_INVALID
) {
2032 /* Require a dedicated allocation for images with modifiers.
2034 * See also anv_AllocateMemory.
2036 requirements
->prefersDedicatedAllocation
= VK_TRUE
;
2037 requirements
->requiresDedicatedAllocation
= VK_TRUE
;
2039 requirements
->prefersDedicatedAllocation
= VK_FALSE
;
2040 requirements
->requiresDedicatedAllocation
= VK_FALSE
;
2046 anv_debug_ignored_stype(ext
->sType
);
2052 void anv_GetImageSparseMemoryRequirements(
2055 uint32_t* pSparseMemoryRequirementCount
,
2056 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
2058 *pSparseMemoryRequirementCount
= 0;
2061 void anv_GetImageSparseMemoryRequirements2KHR(
2063 const VkImageSparseMemoryRequirementsInfo2KHR
* pInfo
,
2064 uint32_t* pSparseMemoryRequirementCount
,
2065 VkSparseImageMemoryRequirements2KHR
* pSparseMemoryRequirements
)
2067 *pSparseMemoryRequirementCount
= 0;
2070 void anv_GetDeviceMemoryCommitment(
2072 VkDeviceMemory memory
,
2073 VkDeviceSize
* pCommittedMemoryInBytes
)
2075 *pCommittedMemoryInBytes
= 0;
2079 anv_bind_buffer_memory(const VkBindBufferMemoryInfoKHR
*pBindInfo
)
2081 ANV_FROM_HANDLE(anv_device_memory
, mem
, pBindInfo
->memory
);
2082 ANV_FROM_HANDLE(anv_buffer
, buffer
, pBindInfo
->buffer
);
2084 assert(pBindInfo
->sType
== VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO_KHR
);
2087 assert((buffer
->usage
& mem
->type
->valid_buffer_usage
) == buffer
->usage
);
2088 buffer
->bo
= mem
->bo
;
2089 buffer
->offset
= pBindInfo
->memoryOffset
;
2096 VkResult
anv_BindBufferMemory(
2099 VkDeviceMemory memory
,
2100 VkDeviceSize memoryOffset
)
2102 anv_bind_buffer_memory(
2103 &(VkBindBufferMemoryInfoKHR
) {
2104 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO_KHR
,
2107 .memoryOffset
= memoryOffset
,
2113 VkResult
anv_BindBufferMemory2KHR(
2115 uint32_t bindInfoCount
,
2116 const VkBindBufferMemoryInfoKHR
* pBindInfos
)
2118 for (uint32_t i
= 0; i
< bindInfoCount
; i
++)
2119 anv_bind_buffer_memory(&pBindInfos
[i
]);
2124 VkResult
anv_QueueBindSparse(
2126 uint32_t bindInfoCount
,
2127 const VkBindSparseInfo
* pBindInfo
,
2130 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
2131 if (unlikely(queue
->device
->lost
))
2132 return VK_ERROR_DEVICE_LOST
;
2134 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
2139 VkResult
anv_CreateEvent(
2141 const VkEventCreateInfo
* pCreateInfo
,
2142 const VkAllocationCallbacks
* pAllocator
,
2145 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2146 struct anv_state state
;
2147 struct anv_event
*event
;
2149 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_EVENT_CREATE_INFO
);
2151 state
= anv_state_pool_alloc(&device
->dynamic_state_pool
,
2154 event
->state
= state
;
2155 event
->semaphore
= VK_EVENT_RESET
;
2157 if (!device
->info
.has_llc
) {
2158 /* Make sure the writes we're flushing have landed. */
2159 __builtin_ia32_mfence();
2160 __builtin_ia32_clflush(event
);
2163 *pEvent
= anv_event_to_handle(event
);
2168 void anv_DestroyEvent(
2171 const VkAllocationCallbacks
* pAllocator
)
2173 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2174 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2179 anv_state_pool_free(&device
->dynamic_state_pool
, event
->state
);
2182 VkResult
anv_GetEventStatus(
2186 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2187 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2189 if (unlikely(device
->lost
))
2190 return VK_ERROR_DEVICE_LOST
;
2192 if (!device
->info
.has_llc
) {
2193 /* Invalidate read cache before reading event written by GPU. */
2194 __builtin_ia32_clflush(event
);
2195 __builtin_ia32_mfence();
2199 return event
->semaphore
;
2202 VkResult
anv_SetEvent(
2206 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2207 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2209 event
->semaphore
= VK_EVENT_SET
;
2211 if (!device
->info
.has_llc
) {
2212 /* Make sure the writes we're flushing have landed. */
2213 __builtin_ia32_mfence();
2214 __builtin_ia32_clflush(event
);
2220 VkResult
anv_ResetEvent(
2224 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2225 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2227 event
->semaphore
= VK_EVENT_RESET
;
2229 if (!device
->info
.has_llc
) {
2230 /* Make sure the writes we're flushing have landed. */
2231 __builtin_ia32_mfence();
2232 __builtin_ia32_clflush(event
);
2240 VkResult
anv_CreateBuffer(
2242 const VkBufferCreateInfo
* pCreateInfo
,
2243 const VkAllocationCallbacks
* pAllocator
,
2246 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2247 struct anv_buffer
*buffer
;
2249 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
2251 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
2252 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2254 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2256 buffer
->size
= pCreateInfo
->size
;
2257 buffer
->usage
= pCreateInfo
->usage
;
2261 *pBuffer
= anv_buffer_to_handle(buffer
);
2266 void anv_DestroyBuffer(
2269 const VkAllocationCallbacks
* pAllocator
)
2271 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2272 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
2277 vk_free2(&device
->alloc
, pAllocator
, buffer
);
2281 anv_fill_buffer_surface_state(struct anv_device
*device
, struct anv_state state
,
2282 enum isl_format format
,
2283 uint32_t offset
, uint32_t range
, uint32_t stride
)
2285 isl_buffer_fill_state(&device
->isl_dev
, state
.map
,
2287 .mocs
= device
->default_mocs
,
2292 anv_state_flush(device
, state
);
2295 void anv_DestroySampler(
2298 const VkAllocationCallbacks
* pAllocator
)
2300 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2301 ANV_FROM_HANDLE(anv_sampler
, sampler
, _sampler
);
2306 vk_free2(&device
->alloc
, pAllocator
, sampler
);
2309 VkResult
anv_CreateFramebuffer(
2311 const VkFramebufferCreateInfo
* pCreateInfo
,
2312 const VkAllocationCallbacks
* pAllocator
,
2313 VkFramebuffer
* pFramebuffer
)
2315 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2316 struct anv_framebuffer
*framebuffer
;
2318 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
2320 size_t size
= sizeof(*framebuffer
) +
2321 sizeof(struct anv_image_view
*) * pCreateInfo
->attachmentCount
;
2322 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
2323 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2324 if (framebuffer
== NULL
)
2325 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2327 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
2328 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
2329 VkImageView _iview
= pCreateInfo
->pAttachments
[i
];
2330 framebuffer
->attachments
[i
] = anv_image_view_from_handle(_iview
);
2333 framebuffer
->width
= pCreateInfo
->width
;
2334 framebuffer
->height
= pCreateInfo
->height
;
2335 framebuffer
->layers
= pCreateInfo
->layers
;
2337 *pFramebuffer
= anv_framebuffer_to_handle(framebuffer
);
2342 void anv_DestroyFramebuffer(
2345 const VkAllocationCallbacks
* pAllocator
)
2347 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2348 ANV_FROM_HANDLE(anv_framebuffer
, fb
, _fb
);
2353 vk_free2(&device
->alloc
, pAllocator
, fb
);
2356 /* vk_icd.h does not declare this function, so we declare it here to
2357 * suppress Wmissing-prototypes.
2359 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2360 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
);
2362 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2363 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
)
2365 /* For the full details on loader interface versioning, see
2366 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
2367 * What follows is a condensed summary, to help you navigate the large and
2368 * confusing official doc.
2370 * - Loader interface v0 is incompatible with later versions. We don't
2373 * - In loader interface v1:
2374 * - The first ICD entrypoint called by the loader is
2375 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
2377 * - The ICD must statically expose no other Vulkan symbol unless it is
2378 * linked with -Bsymbolic.
2379 * - Each dispatchable Vulkan handle created by the ICD must be
2380 * a pointer to a struct whose first member is VK_LOADER_DATA. The
2381 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
2382 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
2383 * vkDestroySurfaceKHR(). The ICD must be capable of working with
2384 * such loader-managed surfaces.
2386 * - Loader interface v2 differs from v1 in:
2387 * - The first ICD entrypoint called by the loader is
2388 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
2389 * statically expose this entrypoint.
2391 * - Loader interface v3 differs from v2 in:
2392 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
2393 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
2394 * because the loader no longer does so.
2396 *pSupportedVersion
= MIN2(*pSupportedVersion
, 3u);