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/disk_cache.h"
39 #include "util/mesa-sha1.h"
41 #include "common/gen_defines.h"
43 #include "genxml/gen7_pack.h"
46 compiler_debug_log(void *data
, const char *fmt
, ...)
50 compiler_perf_log(void *data
, const char *fmt
, ...)
55 if (unlikely(INTEL_DEBUG
& DEBUG_PERF
))
56 intel_logd_v(fmt
, args
);
62 anv_compute_heap_size(int fd
, uint64_t gtt_size
, uint64_t *heap_size
)
64 /* Query the total ram from the system */
68 uint64_t total_ram
= (uint64_t)info
.totalram
* (uint64_t)info
.mem_unit
;
70 /* We don't want to burn too much ram with the GPU. If the user has 4GiB
71 * or less, we use at most half. If they have more than 4GiB, we use 3/4.
73 uint64_t available_ram
;
74 if (total_ram
<= 4ull * 1024ull * 1024ull * 1024ull)
75 available_ram
= total_ram
/ 2;
77 available_ram
= total_ram
* 3 / 4;
79 /* We also want to leave some padding for things we allocate in the driver,
80 * so don't go over 3/4 of the GTT either.
82 uint64_t available_gtt
= gtt_size
* 3 / 4;
84 *heap_size
= MIN2(available_ram
, available_gtt
);
90 anv_physical_device_init_heaps(struct anv_physical_device
*device
, int fd
)
93 if (anv_gem_get_context_param(fd
, 0, I915_CONTEXT_PARAM_GTT_SIZE
,
95 /* If, for whatever reason, we can't actually get the GTT size from the
96 * kernel (too old?) fall back to the aperture size.
98 anv_perf_warn(NULL
, NULL
,
99 "Failed to get I915_CONTEXT_PARAM_GTT_SIZE: %m");
101 if (anv_gem_get_aperture(fd
, >t_size
) == -1) {
102 return vk_errorf(NULL
, NULL
, VK_ERROR_INITIALIZATION_FAILED
,
103 "failed to get aperture size: %m");
107 device
->supports_48bit_addresses
= (device
->info
.gen
>= 8) &&
108 gtt_size
> (4ULL << 30 /* GiB */);
110 uint64_t heap_size
= 0;
111 VkResult result
= anv_compute_heap_size(fd
, gtt_size
, &heap_size
);
112 if (result
!= VK_SUCCESS
)
115 if (heap_size
> (2ull << 30) && !device
->supports_48bit_addresses
) {
116 /* When running with an overridden PCI ID, we may get a GTT size from
117 * the kernel that is greater than 2 GiB but the execbuf check for 48bit
118 * address support can still fail. Just clamp the address space size to
119 * 2 GiB if we don't have 48-bit support.
121 intel_logw("%s:%d: The kernel reported a GTT size larger than 2 GiB but "
122 "not support for 48-bit addresses",
124 heap_size
= 2ull << 30;
127 if (heap_size
<= 3ull * (1ull << 30)) {
128 /* In this case, everything fits nicely into the 32-bit address space,
129 * so there's no need for supporting 48bit addresses on client-allocated
132 device
->memory
.heap_count
= 1;
133 device
->memory
.heaps
[0] = (struct anv_memory_heap
) {
135 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
136 .supports_48bit_addresses
= false,
139 /* Not everything will fit nicely into a 32-bit address space. In this
140 * case we need a 64-bit heap. Advertise a small 32-bit heap and a
141 * larger 48-bit heap. If we're in this case, then we have a total heap
142 * size larger than 3GiB which most likely means they have 8 GiB of
143 * video memory and so carving off 1 GiB for the 32-bit heap should be
146 const uint64_t heap_size_32bit
= 1ull << 30;
147 const uint64_t heap_size_48bit
= heap_size
- heap_size_32bit
;
149 assert(device
->supports_48bit_addresses
);
151 device
->memory
.heap_count
= 2;
152 device
->memory
.heaps
[0] = (struct anv_memory_heap
) {
153 .size
= heap_size_48bit
,
154 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
155 .supports_48bit_addresses
= true,
157 device
->memory
.heaps
[1] = (struct anv_memory_heap
) {
158 .size
= heap_size_32bit
,
159 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
160 .supports_48bit_addresses
= false,
164 uint32_t type_count
= 0;
165 for (uint32_t heap
= 0; heap
< device
->memory
.heap_count
; heap
++) {
166 uint32_t valid_buffer_usage
= ~0;
168 /* There appears to be a hardware issue in the VF cache where it only
169 * considers the bottom 32 bits of memory addresses. If you happen to
170 * have two vertex buffers which get placed exactly 4 GiB apart and use
171 * them in back-to-back draw calls, you can get collisions. In order to
172 * solve this problem, we require vertex and index buffers be bound to
173 * memory allocated out of the 32-bit heap.
175 if (device
->memory
.heaps
[heap
].supports_48bit_addresses
) {
176 valid_buffer_usage
&= ~(VK_BUFFER_USAGE_INDEX_BUFFER_BIT
|
177 VK_BUFFER_USAGE_VERTEX_BUFFER_BIT
);
180 if (device
->info
.has_llc
) {
181 /* Big core GPUs share LLC with the CPU and thus one memory type can be
182 * both cached and coherent at the same time.
184 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
185 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
186 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
187 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
188 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
190 .valid_buffer_usage
= valid_buffer_usage
,
193 /* The spec requires that we expose a host-visible, coherent memory
194 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
195 * to give the application a choice between cached, but not coherent and
196 * coherent but uncached (WC though).
198 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
199 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
200 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
201 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
203 .valid_buffer_usage
= valid_buffer_usage
,
205 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
206 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
207 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
208 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
210 .valid_buffer_usage
= valid_buffer_usage
,
214 device
->memory
.type_count
= type_count
;
220 anv_physical_device_init_uuids(struct anv_physical_device
*device
)
222 const struct build_id_note
*note
=
223 build_id_find_nhdr_for_addr(anv_physical_device_init_uuids
);
225 return vk_errorf(device
->instance
, device
,
226 VK_ERROR_INITIALIZATION_FAILED
,
227 "Failed to find build-id");
230 unsigned build_id_len
= build_id_length(note
);
231 if (build_id_len
< 20) {
232 return vk_errorf(device
->instance
, device
,
233 VK_ERROR_INITIALIZATION_FAILED
,
234 "build-id too short. It needs to be a SHA");
237 memcpy(device
->driver_build_sha1
, build_id_data(note
), 20);
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_disk_cache(struct anv_physical_device
*device
)
280 #ifdef ENABLE_SHADER_CACHE
282 MAYBE_UNUSED
int len
= snprintf(renderer
, sizeof(renderer
), "anv_%04x",
284 assert(len
== sizeof(renderer
) - 2);
287 _mesa_sha1_format(timestamp
, device
->driver_build_sha1
);
289 const uint64_t driver_flags
=
290 brw_get_compiler_config_value(device
->compiler
);
291 device
->disk_cache
= disk_cache_create(renderer
, timestamp
, driver_flags
);
293 device
->disk_cache
= NULL
;
298 anv_physical_device_free_disk_cache(struct anv_physical_device
*device
)
300 #ifdef ENABLE_SHADER_CACHE
301 if (device
->disk_cache
)
302 disk_cache_destroy(device
->disk_cache
);
304 assert(device
->disk_cache
== NULL
);
309 anv_physical_device_init(struct anv_physical_device
*device
,
310 struct anv_instance
*instance
,
311 const char *primary_path
,
318 brw_process_intel_debug_variable();
320 fd
= open(path
, O_RDWR
| O_CLOEXEC
);
322 return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
324 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
325 device
->instance
= instance
;
327 assert(strlen(path
) < ARRAY_SIZE(device
->path
));
328 snprintf(device
->path
, ARRAY_SIZE(device
->path
), "%s", path
);
330 device
->no_hw
= getenv("INTEL_NO_HW") != NULL
;
332 const int pci_id_override
= gen_get_pci_device_id_override();
333 if (pci_id_override
< 0) {
334 device
->chipset_id
= anv_gem_get_param(fd
, I915_PARAM_CHIPSET_ID
);
335 if (!device
->chipset_id
) {
336 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
340 device
->chipset_id
= pci_id_override
;
341 device
->no_hw
= true;
344 device
->name
= gen_get_device_name(device
->chipset_id
);
345 if (!gen_get_device_info(device
->chipset_id
, &device
->info
)) {
346 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
350 if (device
->info
.is_haswell
) {
351 intel_logw("Haswell Vulkan support is incomplete");
352 } else if (device
->info
.gen
== 7 && !device
->info
.is_baytrail
) {
353 intel_logw("Ivy Bridge Vulkan support is incomplete");
354 } else if (device
->info
.gen
== 7 && device
->info
.is_baytrail
) {
355 intel_logw("Bay Trail Vulkan support is incomplete");
356 } else if (device
->info
.gen
>= 8 && device
->info
.gen
<= 10) {
357 /* Gen8-10 fully supported */
358 } else if (device
->info
.gen
== 11) {
359 intel_logw("Vulkan is not yet fully supported on gen11.");
361 result
= vk_errorf(device
->instance
, device
,
362 VK_ERROR_INCOMPATIBLE_DRIVER
,
363 "Vulkan not yet supported on %s", device
->name
);
367 device
->cmd_parser_version
= -1;
368 if (device
->info
.gen
== 7) {
369 device
->cmd_parser_version
=
370 anv_gem_get_param(fd
, I915_PARAM_CMD_PARSER_VERSION
);
371 if (device
->cmd_parser_version
== -1) {
372 result
= vk_errorf(device
->instance
, device
,
373 VK_ERROR_INITIALIZATION_FAILED
,
374 "failed to get command parser version");
379 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_WAIT_TIMEOUT
)) {
380 result
= vk_errorf(device
->instance
, device
,
381 VK_ERROR_INITIALIZATION_FAILED
,
382 "kernel missing gem wait");
386 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_EXECBUF2
)) {
387 result
= vk_errorf(device
->instance
, device
,
388 VK_ERROR_INITIALIZATION_FAILED
,
389 "kernel missing execbuf2");
393 if (!device
->info
.has_llc
&&
394 anv_gem_get_param(fd
, I915_PARAM_MMAP_VERSION
) < 1) {
395 result
= vk_errorf(device
->instance
, device
,
396 VK_ERROR_INITIALIZATION_FAILED
,
397 "kernel missing wc mmap");
401 result
= anv_physical_device_init_heaps(device
, fd
);
402 if (result
!= VK_SUCCESS
)
405 device
->has_exec_async
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_ASYNC
);
406 device
->has_exec_capture
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_CAPTURE
);
407 device
->has_exec_fence
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE
);
408 device
->has_syncobj
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE_ARRAY
);
409 device
->has_syncobj_wait
= device
->has_syncobj
&&
410 anv_gem_supports_syncobj_wait(fd
);
411 device
->has_context_priority
= anv_gem_has_context_priority(fd
);
413 device
->use_softpin
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_SOFTPIN
)
414 && device
->supports_48bit_addresses
;
416 device
->has_context_isolation
=
417 anv_gem_get_param(fd
, I915_PARAM_HAS_CONTEXT_ISOLATION
);
419 bool swizzled
= anv_gem_get_bit6_swizzle(fd
, I915_TILING_X
);
421 /* Starting with Gen10, the timestamp frequency of the command streamer may
422 * vary from one part to another. We can query the value from the kernel.
424 if (device
->info
.gen
>= 10) {
425 int timestamp_frequency
=
426 anv_gem_get_param(fd
, I915_PARAM_CS_TIMESTAMP_FREQUENCY
);
428 if (timestamp_frequency
< 0)
429 intel_logw("Kernel 4.16-rc1+ required to properly query CS timestamp frequency");
431 device
->info
.timestamp_frequency
= timestamp_frequency
;
434 /* GENs prior to 8 do not support EU/Subslice info */
435 if (device
->info
.gen
>= 8) {
436 device
->subslice_total
= anv_gem_get_param(fd
, I915_PARAM_SUBSLICE_TOTAL
);
437 device
->eu_total
= anv_gem_get_param(fd
, I915_PARAM_EU_TOTAL
);
439 /* Without this information, we cannot get the right Braswell
440 * brandstrings, and we have to use conservative numbers for GPGPU on
441 * many platforms, but otherwise, things will just work.
443 if (device
->subslice_total
< 1 || device
->eu_total
< 1) {
444 intel_logw("Kernel 4.1 required to properly query GPU properties");
446 } else if (device
->info
.gen
== 7) {
447 device
->subslice_total
= 1 << (device
->info
.gt
- 1);
450 if (device
->info
.is_cherryview
&&
451 device
->subslice_total
> 0 && device
->eu_total
> 0) {
452 /* Logical CS threads = EUs per subslice * num threads per EU */
453 uint32_t max_cs_threads
=
454 device
->eu_total
/ device
->subslice_total
* device
->info
.num_thread_per_eu
;
456 /* Fuse configurations may give more threads than expected, never less. */
457 if (max_cs_threads
> device
->info
.max_cs_threads
)
458 device
->info
.max_cs_threads
= max_cs_threads
;
461 device
->compiler
= brw_compiler_create(NULL
, &device
->info
);
462 if (device
->compiler
== NULL
) {
463 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
466 device
->compiler
->shader_debug_log
= compiler_debug_log
;
467 device
->compiler
->shader_perf_log
= compiler_perf_log
;
468 device
->compiler
->supports_pull_constants
= false;
469 device
->compiler
->constant_buffer_0_is_relative
=
470 device
->info
.gen
< 8 || !device
->has_context_isolation
;
471 device
->compiler
->supports_shader_constants
= true;
473 isl_device_init(&device
->isl_dev
, &device
->info
, swizzled
);
475 result
= anv_physical_device_init_uuids(device
);
476 if (result
!= VK_SUCCESS
)
479 anv_physical_device_init_disk_cache(device
);
481 if (instance
->enabled_extensions
.KHR_display
) {
482 master_fd
= open(primary_path
, O_RDWR
| O_CLOEXEC
);
483 if (master_fd
>= 0) {
484 /* prod the device with a GETPARAM call which will fail if
485 * we don't have permission to even render on this device
487 if (anv_gem_get_param(master_fd
, I915_PARAM_CHIPSET_ID
) == 0) {
493 device
->master_fd
= master_fd
;
495 result
= anv_init_wsi(device
);
496 if (result
!= VK_SUCCESS
) {
497 ralloc_free(device
->compiler
);
498 anv_physical_device_free_disk_cache(device
);
502 anv_physical_device_get_supported_extensions(device
,
503 &device
->supported_extensions
);
506 device
->local_fd
= fd
;
518 anv_physical_device_finish(struct anv_physical_device
*device
)
520 anv_finish_wsi(device
);
521 anv_physical_device_free_disk_cache(device
);
522 ralloc_free(device
->compiler
);
523 close(device
->local_fd
);
524 if (device
->master_fd
>= 0)
525 close(device
->master_fd
);
529 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
530 VkSystemAllocationScope allocationScope
)
536 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
537 size_t align
, VkSystemAllocationScope allocationScope
)
539 return realloc(pOriginal
, size
);
543 default_free_func(void *pUserData
, void *pMemory
)
548 static const VkAllocationCallbacks default_alloc
= {
550 .pfnAllocation
= default_alloc_func
,
551 .pfnReallocation
= default_realloc_func
,
552 .pfnFree
= default_free_func
,
555 VkResult
anv_EnumerateInstanceExtensionProperties(
556 const char* pLayerName
,
557 uint32_t* pPropertyCount
,
558 VkExtensionProperties
* pProperties
)
560 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
562 for (int i
= 0; i
< ANV_INSTANCE_EXTENSION_COUNT
; i
++) {
563 if (anv_instance_extensions_supported
.extensions
[i
]) {
564 vk_outarray_append(&out
, prop
) {
565 *prop
= anv_instance_extensions
[i
];
570 return vk_outarray_status(&out
);
573 VkResult
anv_CreateInstance(
574 const VkInstanceCreateInfo
* pCreateInfo
,
575 const VkAllocationCallbacks
* pAllocator
,
576 VkInstance
* pInstance
)
578 struct anv_instance
*instance
;
581 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
583 struct anv_instance_extension_table enabled_extensions
= {};
584 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
586 for (idx
= 0; idx
< ANV_INSTANCE_EXTENSION_COUNT
; idx
++) {
587 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
588 anv_instance_extensions
[idx
].extensionName
) == 0)
592 if (idx
>= ANV_INSTANCE_EXTENSION_COUNT
)
593 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
595 if (!anv_instance_extensions_supported
.extensions
[idx
])
596 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
598 enabled_extensions
.extensions
[idx
] = true;
601 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
602 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
604 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
606 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
609 instance
->alloc
= *pAllocator
;
611 instance
->alloc
= default_alloc
;
613 instance
->app_info
= (struct anv_app_info
) { .api_version
= 0 };
614 if (pCreateInfo
->pApplicationInfo
) {
615 const VkApplicationInfo
*app
= pCreateInfo
->pApplicationInfo
;
617 instance
->app_info
.app_name
=
618 vk_strdup(&instance
->alloc
, app
->pApplicationName
,
619 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
620 instance
->app_info
.app_version
= app
->applicationVersion
;
622 instance
->app_info
.engine_name
=
623 vk_strdup(&instance
->alloc
, app
->pEngineName
,
624 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
625 instance
->app_info
.engine_version
= app
->engineVersion
;
627 instance
->app_info
.api_version
= app
->apiVersion
;
630 if (instance
->app_info
.api_version
== 0)
631 anv_EnumerateInstanceVersion(&instance
->app_info
.api_version
);
633 instance
->enabled_extensions
= enabled_extensions
;
635 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->dispatch
.entrypoints
); i
++) {
636 /* Vulkan requires that entrypoints for extensions which have not been
637 * enabled must not be advertised.
639 if (!anv_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
640 &instance
->enabled_extensions
, NULL
)) {
641 instance
->dispatch
.entrypoints
[i
] = NULL
;
642 } else if (anv_dispatch_table
.entrypoints
[i
] != NULL
) {
643 instance
->dispatch
.entrypoints
[i
] = anv_dispatch_table
.entrypoints
[i
];
645 instance
->dispatch
.entrypoints
[i
] =
646 anv_tramp_dispatch_table
.entrypoints
[i
];
650 instance
->physicalDeviceCount
= -1;
652 result
= vk_debug_report_instance_init(&instance
->debug_report_callbacks
);
653 if (result
!= VK_SUCCESS
) {
654 vk_free2(&default_alloc
, pAllocator
, instance
);
655 return vk_error(result
);
658 instance
->pipeline_cache_enabled
=
659 env_var_as_boolean("ANV_ENABLE_PIPELINE_CACHE", true);
663 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
665 *pInstance
= anv_instance_to_handle(instance
);
670 void anv_DestroyInstance(
671 VkInstance _instance
,
672 const VkAllocationCallbacks
* pAllocator
)
674 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
679 if (instance
->physicalDeviceCount
> 0) {
680 /* We support at most one physical device. */
681 assert(instance
->physicalDeviceCount
== 1);
682 anv_physical_device_finish(&instance
->physicalDevice
);
685 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
687 vk_debug_report_instance_destroy(&instance
->debug_report_callbacks
);
691 vk_free(&instance
->alloc
, instance
);
695 anv_enumerate_devices(struct anv_instance
*instance
)
697 /* TODO: Check for more devices ? */
698 drmDevicePtr devices
[8];
699 VkResult result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
702 instance
->physicalDeviceCount
= 0;
704 max_devices
= drmGetDevices2(0, devices
, ARRAY_SIZE(devices
));
706 return VK_ERROR_INCOMPATIBLE_DRIVER
;
708 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
709 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
710 devices
[i
]->bustype
== DRM_BUS_PCI
&&
711 devices
[i
]->deviceinfo
.pci
->vendor_id
== 0x8086) {
713 result
= anv_physical_device_init(&instance
->physicalDevice
,
715 devices
[i
]->nodes
[DRM_NODE_PRIMARY
],
716 devices
[i
]->nodes
[DRM_NODE_RENDER
]);
717 if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
721 drmFreeDevices(devices
, max_devices
);
723 if (result
== VK_SUCCESS
)
724 instance
->physicalDeviceCount
= 1;
730 anv_instance_ensure_physical_device(struct anv_instance
*instance
)
732 if (instance
->physicalDeviceCount
< 0) {
733 VkResult result
= anv_enumerate_devices(instance
);
734 if (result
!= VK_SUCCESS
&&
735 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
742 VkResult
anv_EnumeratePhysicalDevices(
743 VkInstance _instance
,
744 uint32_t* pPhysicalDeviceCount
,
745 VkPhysicalDevice
* pPhysicalDevices
)
747 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
748 VK_OUTARRAY_MAKE(out
, pPhysicalDevices
, pPhysicalDeviceCount
);
750 VkResult result
= anv_instance_ensure_physical_device(instance
);
751 if (result
!= VK_SUCCESS
)
754 if (instance
->physicalDeviceCount
== 0)
757 assert(instance
->physicalDeviceCount
== 1);
758 vk_outarray_append(&out
, i
) {
759 *i
= anv_physical_device_to_handle(&instance
->physicalDevice
);
762 return vk_outarray_status(&out
);
765 VkResult
anv_EnumeratePhysicalDeviceGroups(
766 VkInstance _instance
,
767 uint32_t* pPhysicalDeviceGroupCount
,
768 VkPhysicalDeviceGroupProperties
* pPhysicalDeviceGroupProperties
)
770 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
771 VK_OUTARRAY_MAKE(out
, pPhysicalDeviceGroupProperties
,
772 pPhysicalDeviceGroupCount
);
774 VkResult result
= anv_instance_ensure_physical_device(instance
);
775 if (result
!= VK_SUCCESS
)
778 if (instance
->physicalDeviceCount
== 0)
781 assert(instance
->physicalDeviceCount
== 1);
783 vk_outarray_append(&out
, p
) {
784 p
->physicalDeviceCount
= 1;
785 memset(p
->physicalDevices
, 0, sizeof(p
->physicalDevices
));
786 p
->physicalDevices
[0] =
787 anv_physical_device_to_handle(&instance
->physicalDevice
);
788 p
->subsetAllocation
= VK_FALSE
;
790 vk_foreach_struct(ext
, p
->pNext
)
791 anv_debug_ignored_stype(ext
->sType
);
794 return vk_outarray_status(&out
);
797 void anv_GetPhysicalDeviceFeatures(
798 VkPhysicalDevice physicalDevice
,
799 VkPhysicalDeviceFeatures
* pFeatures
)
801 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
803 *pFeatures
= (VkPhysicalDeviceFeatures
) {
804 .robustBufferAccess
= true,
805 .fullDrawIndexUint32
= true,
806 .imageCubeArray
= true,
807 .independentBlend
= true,
808 .geometryShader
= true,
809 .tessellationShader
= true,
810 .sampleRateShading
= true,
811 .dualSrcBlend
= true,
813 .multiDrawIndirect
= true,
814 .drawIndirectFirstInstance
= true,
816 .depthBiasClamp
= true,
817 .fillModeNonSolid
= true,
818 .depthBounds
= false,
822 .multiViewport
= true,
823 .samplerAnisotropy
= true,
824 .textureCompressionETC2
= pdevice
->info
.gen
>= 8 ||
825 pdevice
->info
.is_baytrail
,
826 .textureCompressionASTC_LDR
= pdevice
->info
.gen
>= 9, /* FINISHME CHV */
827 .textureCompressionBC
= true,
828 .occlusionQueryPrecise
= true,
829 .pipelineStatisticsQuery
= true,
830 .fragmentStoresAndAtomics
= true,
831 .shaderTessellationAndGeometryPointSize
= true,
832 .shaderImageGatherExtended
= true,
833 .shaderStorageImageExtendedFormats
= true,
834 .shaderStorageImageMultisample
= false,
835 .shaderStorageImageReadWithoutFormat
= false,
836 .shaderStorageImageWriteWithoutFormat
= true,
837 .shaderUniformBufferArrayDynamicIndexing
= true,
838 .shaderSampledImageArrayDynamicIndexing
= true,
839 .shaderStorageBufferArrayDynamicIndexing
= true,
840 .shaderStorageImageArrayDynamicIndexing
= true,
841 .shaderClipDistance
= true,
842 .shaderCullDistance
= true,
843 .shaderFloat64
= pdevice
->info
.gen
>= 8 &&
844 pdevice
->info
.has_64bit_types
,
845 .shaderInt64
= pdevice
->info
.gen
>= 8 &&
846 pdevice
->info
.has_64bit_types
,
847 .shaderInt16
= pdevice
->info
.gen
>= 8,
848 .shaderResourceMinLod
= false,
849 .variableMultisampleRate
= true,
850 .inheritedQueries
= true,
853 /* We can't do image stores in vec4 shaders */
854 pFeatures
->vertexPipelineStoresAndAtomics
=
855 pdevice
->compiler
->scalar_stage
[MESA_SHADER_VERTEX
] &&
856 pdevice
->compiler
->scalar_stage
[MESA_SHADER_GEOMETRY
];
858 struct anv_app_info
*app_info
= &pdevice
->instance
->app_info
;
860 /* The new DOOM and Wolfenstein games require depthBounds without
861 * checking for it. They seem to run fine without it so just claim it's
862 * there and accept the consequences.
864 if (app_info
->engine_name
&& strcmp(app_info
->engine_name
, "idTech") == 0)
865 pFeatures
->depthBounds
= true;
868 void anv_GetPhysicalDeviceFeatures2(
869 VkPhysicalDevice physicalDevice
,
870 VkPhysicalDeviceFeatures2
* pFeatures
)
872 anv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
874 vk_foreach_struct(ext
, pFeatures
->pNext
) {
875 switch (ext
->sType
) {
876 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES
: {
877 VkPhysicalDeviceProtectedMemoryFeatures
*features
= (void *)ext
;
878 features
->protectedMemory
= VK_FALSE
;
882 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES
: {
883 VkPhysicalDeviceMultiviewFeatures
*features
=
884 (VkPhysicalDeviceMultiviewFeatures
*)ext
;
885 features
->multiview
= true;
886 features
->multiviewGeometryShader
= true;
887 features
->multiviewTessellationShader
= true;
891 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTER_FEATURES
: {
892 VkPhysicalDeviceVariablePointerFeatures
*features
= (void *)ext
;
893 features
->variablePointersStorageBuffer
= true;
894 features
->variablePointers
= true;
898 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES
: {
899 VkPhysicalDeviceSamplerYcbcrConversionFeatures
*features
=
900 (VkPhysicalDeviceSamplerYcbcrConversionFeatures
*) ext
;
901 features
->samplerYcbcrConversion
= true;
905 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETER_FEATURES
: {
906 VkPhysicalDeviceShaderDrawParameterFeatures
*features
= (void *)ext
;
907 features
->shaderDrawParameters
= true;
911 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES_KHR
: {
912 VkPhysicalDevice16BitStorageFeaturesKHR
*features
=
913 (VkPhysicalDevice16BitStorageFeaturesKHR
*)ext
;
914 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
916 features
->storageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
917 features
->uniformAndStorageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
918 features
->storagePushConstant16
= pdevice
->info
.gen
>= 8;
919 features
->storageInputOutput16
= false;
923 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES_KHR
: {
924 VkPhysicalDevice8BitStorageFeaturesKHR
*features
=
925 (VkPhysicalDevice8BitStorageFeaturesKHR
*)ext
;
926 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
928 features
->storageBuffer8BitAccess
= pdevice
->info
.gen
>= 8;
929 features
->uniformAndStorageBuffer8BitAccess
= pdevice
->info
.gen
>= 8;
930 features
->storagePushConstant8
= pdevice
->info
.gen
>= 8;
935 anv_debug_ignored_stype(ext
->sType
);
941 void anv_GetPhysicalDeviceProperties(
942 VkPhysicalDevice physicalDevice
,
943 VkPhysicalDeviceProperties
* pProperties
)
945 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
946 const struct gen_device_info
*devinfo
= &pdevice
->info
;
948 /* See assertions made when programming the buffer surface state. */
949 const uint32_t max_raw_buffer_sz
= devinfo
->gen
>= 7 ?
950 (1ul << 30) : (1ul << 27);
952 const uint32_t max_samplers
= (devinfo
->gen
>= 8 || devinfo
->is_haswell
) ?
955 VkSampleCountFlags sample_counts
=
956 isl_device_get_sample_counts(&pdevice
->isl_dev
);
958 VkPhysicalDeviceLimits limits
= {
959 .maxImageDimension1D
= (1 << 14),
960 .maxImageDimension2D
= (1 << 14),
961 .maxImageDimension3D
= (1 << 11),
962 .maxImageDimensionCube
= (1 << 14),
963 .maxImageArrayLayers
= (1 << 11),
964 .maxTexelBufferElements
= 128 * 1024 * 1024,
965 .maxUniformBufferRange
= (1ul << 27),
966 .maxStorageBufferRange
= max_raw_buffer_sz
,
967 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
968 .maxMemoryAllocationCount
= UINT32_MAX
,
969 .maxSamplerAllocationCount
= 64 * 1024,
970 .bufferImageGranularity
= 64, /* A cache line */
971 .sparseAddressSpaceSize
= 0,
972 .maxBoundDescriptorSets
= MAX_SETS
,
973 .maxPerStageDescriptorSamplers
= max_samplers
,
974 .maxPerStageDescriptorUniformBuffers
= 64,
975 .maxPerStageDescriptorStorageBuffers
= 64,
976 .maxPerStageDescriptorSampledImages
= max_samplers
,
977 .maxPerStageDescriptorStorageImages
= 64,
978 .maxPerStageDescriptorInputAttachments
= 64,
979 .maxPerStageResources
= 250,
980 .maxDescriptorSetSamplers
= 6 * max_samplers
, /* number of stages * maxPerStageDescriptorSamplers */
981 .maxDescriptorSetUniformBuffers
= 6 * 64, /* number of stages * maxPerStageDescriptorUniformBuffers */
982 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
983 .maxDescriptorSetStorageBuffers
= 6 * 64, /* number of stages * maxPerStageDescriptorStorageBuffers */
984 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
985 .maxDescriptorSetSampledImages
= 6 * max_samplers
, /* number of stages * maxPerStageDescriptorSampledImages */
986 .maxDescriptorSetStorageImages
= 6 * 64, /* number of stages * maxPerStageDescriptorStorageImages */
987 .maxDescriptorSetInputAttachments
= 256,
988 .maxVertexInputAttributes
= MAX_VBS
,
989 .maxVertexInputBindings
= MAX_VBS
,
990 .maxVertexInputAttributeOffset
= 2047,
991 .maxVertexInputBindingStride
= 2048,
992 .maxVertexOutputComponents
= 128,
993 .maxTessellationGenerationLevel
= 64,
994 .maxTessellationPatchSize
= 32,
995 .maxTessellationControlPerVertexInputComponents
= 128,
996 .maxTessellationControlPerVertexOutputComponents
= 128,
997 .maxTessellationControlPerPatchOutputComponents
= 128,
998 .maxTessellationControlTotalOutputComponents
= 2048,
999 .maxTessellationEvaluationInputComponents
= 128,
1000 .maxTessellationEvaluationOutputComponents
= 128,
1001 .maxGeometryShaderInvocations
= 32,
1002 .maxGeometryInputComponents
= 64,
1003 .maxGeometryOutputComponents
= 128,
1004 .maxGeometryOutputVertices
= 256,
1005 .maxGeometryTotalOutputComponents
= 1024,
1006 .maxFragmentInputComponents
= 112, /* 128 components - (POS, PSIZ, CLIP_DIST0, CLIP_DIST1) */
1007 .maxFragmentOutputAttachments
= 8,
1008 .maxFragmentDualSrcAttachments
= 1,
1009 .maxFragmentCombinedOutputResources
= 8,
1010 .maxComputeSharedMemorySize
= 32768,
1011 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
1012 .maxComputeWorkGroupInvocations
= 16 * devinfo
->max_cs_threads
,
1013 .maxComputeWorkGroupSize
= {
1014 16 * devinfo
->max_cs_threads
,
1015 16 * devinfo
->max_cs_threads
,
1016 16 * devinfo
->max_cs_threads
,
1018 .subPixelPrecisionBits
= 4 /* FIXME */,
1019 .subTexelPrecisionBits
= 4 /* FIXME */,
1020 .mipmapPrecisionBits
= 4 /* FIXME */,
1021 .maxDrawIndexedIndexValue
= UINT32_MAX
,
1022 .maxDrawIndirectCount
= UINT32_MAX
,
1023 .maxSamplerLodBias
= 16,
1024 .maxSamplerAnisotropy
= 16,
1025 .maxViewports
= MAX_VIEWPORTS
,
1026 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
1027 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
1028 .viewportSubPixelBits
= 13, /* We take a float? */
1029 .minMemoryMapAlignment
= 4096, /* A page */
1030 .minTexelBufferOffsetAlignment
= 1,
1031 /* We need 16 for UBO block reads to work and 32 for push UBOs */
1032 .minUniformBufferOffsetAlignment
= 32,
1033 .minStorageBufferOffsetAlignment
= 4,
1034 .minTexelOffset
= -8,
1035 .maxTexelOffset
= 7,
1036 .minTexelGatherOffset
= -32,
1037 .maxTexelGatherOffset
= 31,
1038 .minInterpolationOffset
= -0.5,
1039 .maxInterpolationOffset
= 0.4375,
1040 .subPixelInterpolationOffsetBits
= 4,
1041 .maxFramebufferWidth
= (1 << 14),
1042 .maxFramebufferHeight
= (1 << 14),
1043 .maxFramebufferLayers
= (1 << 11),
1044 .framebufferColorSampleCounts
= sample_counts
,
1045 .framebufferDepthSampleCounts
= sample_counts
,
1046 .framebufferStencilSampleCounts
= sample_counts
,
1047 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
1048 .maxColorAttachments
= MAX_RTS
,
1049 .sampledImageColorSampleCounts
= sample_counts
,
1050 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
1051 .sampledImageDepthSampleCounts
= sample_counts
,
1052 .sampledImageStencilSampleCounts
= sample_counts
,
1053 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
1054 .maxSampleMaskWords
= 1,
1055 .timestampComputeAndGraphics
= false,
1056 .timestampPeriod
= 1000000000.0 / devinfo
->timestamp_frequency
,
1057 .maxClipDistances
= 8,
1058 .maxCullDistances
= 8,
1059 .maxCombinedClipAndCullDistances
= 8,
1060 .discreteQueuePriorities
= 1,
1061 .pointSizeRange
= { 0.125, 255.875 },
1062 .lineWidthRange
= { 0.0, 7.9921875 },
1063 .pointSizeGranularity
= (1.0 / 8.0),
1064 .lineWidthGranularity
= (1.0 / 128.0),
1065 .strictLines
= false, /* FINISHME */
1066 .standardSampleLocations
= true,
1067 .optimalBufferCopyOffsetAlignment
= 128,
1068 .optimalBufferCopyRowPitchAlignment
= 128,
1069 .nonCoherentAtomSize
= 64,
1072 *pProperties
= (VkPhysicalDeviceProperties
) {
1073 .apiVersion
= anv_physical_device_api_version(pdevice
),
1074 .driverVersion
= vk_get_driver_version(),
1076 .deviceID
= pdevice
->chipset_id
,
1077 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
1079 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
1082 snprintf(pProperties
->deviceName
, sizeof(pProperties
->deviceName
),
1083 "%s", pdevice
->name
);
1084 memcpy(pProperties
->pipelineCacheUUID
,
1085 pdevice
->pipeline_cache_uuid
, VK_UUID_SIZE
);
1088 void anv_GetPhysicalDeviceProperties2(
1089 VkPhysicalDevice physicalDevice
,
1090 VkPhysicalDeviceProperties2
* pProperties
)
1092 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1094 anv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
1096 vk_foreach_struct(ext
, pProperties
->pNext
) {
1097 switch (ext
->sType
) {
1098 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
1099 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
1100 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
1102 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
1106 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES
: {
1107 VkPhysicalDeviceIDProperties
*id_props
=
1108 (VkPhysicalDeviceIDProperties
*)ext
;
1109 memcpy(id_props
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
1110 memcpy(id_props
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
1111 /* The LUID is for Windows. */
1112 id_props
->deviceLUIDValid
= false;
1116 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES
: {
1117 VkPhysicalDeviceMaintenance3Properties
*props
=
1118 (VkPhysicalDeviceMaintenance3Properties
*)ext
;
1119 /* This value doesn't matter for us today as our per-stage
1120 * descriptors are the real limit.
1122 props
->maxPerSetDescriptors
= 1024;
1123 props
->maxMemoryAllocationSize
= MAX_MEMORY_ALLOCATION_SIZE
;
1127 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES
: {
1128 VkPhysicalDeviceMultiviewProperties
*properties
=
1129 (VkPhysicalDeviceMultiviewProperties
*)ext
;
1130 properties
->maxMultiviewViewCount
= 16;
1131 properties
->maxMultiviewInstanceIndex
= UINT32_MAX
/ 16;
1135 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES
: {
1136 VkPhysicalDevicePointClippingProperties
*properties
=
1137 (VkPhysicalDevicePointClippingProperties
*) ext
;
1138 properties
->pointClippingBehavior
= VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES
;
1139 anv_finishme("Implement pop-free point clipping");
1143 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT
: {
1144 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*properties
=
1145 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*)ext
;
1146 properties
->filterMinmaxImageComponentMapping
= pdevice
->info
.gen
>= 9;
1147 properties
->filterMinmaxSingleComponentFormats
= true;
1151 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES
: {
1152 VkPhysicalDeviceSubgroupProperties
*properties
= (void *)ext
;
1154 properties
->subgroupSize
= BRW_SUBGROUP_SIZE
;
1156 VkShaderStageFlags scalar_stages
= 0;
1157 for (unsigned stage
= 0; stage
< MESA_SHADER_STAGES
; stage
++) {
1158 if (pdevice
->compiler
->scalar_stage
[stage
])
1159 scalar_stages
|= mesa_to_vk_shader_stage(stage
);
1161 properties
->supportedStages
= scalar_stages
;
1163 properties
->supportedOperations
= VK_SUBGROUP_FEATURE_BASIC_BIT
|
1164 VK_SUBGROUP_FEATURE_VOTE_BIT
|
1165 VK_SUBGROUP_FEATURE_ARITHMETIC_BIT
|
1166 VK_SUBGROUP_FEATURE_BALLOT_BIT
|
1167 VK_SUBGROUP_FEATURE_SHUFFLE_BIT
|
1168 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT
|
1169 VK_SUBGROUP_FEATURE_CLUSTERED_BIT
|
1170 VK_SUBGROUP_FEATURE_QUAD_BIT
;
1171 properties
->quadOperationsInAllStages
= VK_TRUE
;
1175 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT
: {
1176 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*props
=
1177 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*)ext
;
1178 /* We have to restrict this a bit for multiview */
1179 props
->maxVertexAttribDivisor
= UINT32_MAX
/ 16;
1183 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES
: {
1184 VkPhysicalDeviceProtectedMemoryProperties
*props
=
1185 (VkPhysicalDeviceProtectedMemoryProperties
*)ext
;
1186 props
->protectedNoFault
= false;
1191 anv_debug_ignored_stype(ext
->sType
);
1197 /* We support exactly one queue family. */
1198 static const VkQueueFamilyProperties
1199 anv_queue_family_properties
= {
1200 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
1201 VK_QUEUE_COMPUTE_BIT
|
1202 VK_QUEUE_TRANSFER_BIT
,
1204 .timestampValidBits
= 36, /* XXX: Real value here */
1205 .minImageTransferGranularity
= { 1, 1, 1 },
1208 void anv_GetPhysicalDeviceQueueFamilyProperties(
1209 VkPhysicalDevice physicalDevice
,
1211 VkQueueFamilyProperties
* pQueueFamilyProperties
)
1213 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pCount
);
1215 vk_outarray_append(&out
, p
) {
1216 *p
= anv_queue_family_properties
;
1220 void anv_GetPhysicalDeviceQueueFamilyProperties2(
1221 VkPhysicalDevice physicalDevice
,
1222 uint32_t* pQueueFamilyPropertyCount
,
1223 VkQueueFamilyProperties2
* pQueueFamilyProperties
)
1226 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pQueueFamilyPropertyCount
);
1228 vk_outarray_append(&out
, p
) {
1229 p
->queueFamilyProperties
= anv_queue_family_properties
;
1231 vk_foreach_struct(s
, p
->pNext
) {
1232 anv_debug_ignored_stype(s
->sType
);
1237 void anv_GetPhysicalDeviceMemoryProperties(
1238 VkPhysicalDevice physicalDevice
,
1239 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
1241 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
1243 pMemoryProperties
->memoryTypeCount
= physical_device
->memory
.type_count
;
1244 for (uint32_t i
= 0; i
< physical_device
->memory
.type_count
; i
++) {
1245 pMemoryProperties
->memoryTypes
[i
] = (VkMemoryType
) {
1246 .propertyFlags
= physical_device
->memory
.types
[i
].propertyFlags
,
1247 .heapIndex
= physical_device
->memory
.types
[i
].heapIndex
,
1251 pMemoryProperties
->memoryHeapCount
= physical_device
->memory
.heap_count
;
1252 for (uint32_t i
= 0; i
< physical_device
->memory
.heap_count
; i
++) {
1253 pMemoryProperties
->memoryHeaps
[i
] = (VkMemoryHeap
) {
1254 .size
= physical_device
->memory
.heaps
[i
].size
,
1255 .flags
= physical_device
->memory
.heaps
[i
].flags
,
1260 void anv_GetPhysicalDeviceMemoryProperties2(
1261 VkPhysicalDevice physicalDevice
,
1262 VkPhysicalDeviceMemoryProperties2
* pMemoryProperties
)
1264 anv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
1265 &pMemoryProperties
->memoryProperties
);
1267 vk_foreach_struct(ext
, pMemoryProperties
->pNext
) {
1268 switch (ext
->sType
) {
1270 anv_debug_ignored_stype(ext
->sType
);
1277 anv_GetDeviceGroupPeerMemoryFeatures(
1280 uint32_t localDeviceIndex
,
1281 uint32_t remoteDeviceIndex
,
1282 VkPeerMemoryFeatureFlags
* pPeerMemoryFeatures
)
1284 assert(localDeviceIndex
== 0 && remoteDeviceIndex
== 0);
1285 *pPeerMemoryFeatures
= VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT
|
1286 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT
|
1287 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT
|
1288 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT
;
1291 PFN_vkVoidFunction
anv_GetInstanceProcAddr(
1292 VkInstance _instance
,
1295 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1297 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
1298 * when we have to return valid function pointers, NULL, or it's left
1299 * undefined. See the table for exact details.
1304 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
1305 if (strcmp(pName, "vk" #entrypoint) == 0) \
1306 return (PFN_vkVoidFunction)anv_##entrypoint
1308 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties
);
1309 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties
);
1310 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceVersion
);
1311 LOOKUP_ANV_ENTRYPOINT(CreateInstance
);
1313 #undef LOOKUP_ANV_ENTRYPOINT
1315 if (instance
== NULL
)
1318 int idx
= anv_get_entrypoint_index(pName
);
1322 return instance
->dispatch
.entrypoints
[idx
];
1325 /* With version 1+ of the loader interface the ICD should expose
1326 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
1329 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1330 VkInstance instance
,
1334 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1335 VkInstance instance
,
1338 return anv_GetInstanceProcAddr(instance
, pName
);
1341 PFN_vkVoidFunction
anv_GetDeviceProcAddr(
1345 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1347 if (!device
|| !pName
)
1350 int idx
= anv_get_entrypoint_index(pName
);
1354 return device
->dispatch
.entrypoints
[idx
];
1358 anv_CreateDebugReportCallbackEXT(VkInstance _instance
,
1359 const VkDebugReportCallbackCreateInfoEXT
* pCreateInfo
,
1360 const VkAllocationCallbacks
* pAllocator
,
1361 VkDebugReportCallbackEXT
* pCallback
)
1363 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1364 return vk_create_debug_report_callback(&instance
->debug_report_callbacks
,
1365 pCreateInfo
, pAllocator
, &instance
->alloc
,
1370 anv_DestroyDebugReportCallbackEXT(VkInstance _instance
,
1371 VkDebugReportCallbackEXT _callback
,
1372 const VkAllocationCallbacks
* pAllocator
)
1374 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1375 vk_destroy_debug_report_callback(&instance
->debug_report_callbacks
,
1376 _callback
, pAllocator
, &instance
->alloc
);
1380 anv_DebugReportMessageEXT(VkInstance _instance
,
1381 VkDebugReportFlagsEXT flags
,
1382 VkDebugReportObjectTypeEXT objectType
,
1385 int32_t messageCode
,
1386 const char* pLayerPrefix
,
1387 const char* pMessage
)
1389 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1390 vk_debug_report(&instance
->debug_report_callbacks
, flags
, objectType
,
1391 object
, location
, messageCode
, pLayerPrefix
, pMessage
);
1395 anv_queue_init(struct anv_device
*device
, struct anv_queue
*queue
)
1397 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1398 queue
->device
= device
;
1403 anv_queue_finish(struct anv_queue
*queue
)
1407 static struct anv_state
1408 anv_state_pool_emit_data(struct anv_state_pool
*pool
, size_t size
, size_t align
, const void *p
)
1410 struct anv_state state
;
1412 state
= anv_state_pool_alloc(pool
, size
, align
);
1413 memcpy(state
.map
, p
, size
);
1415 anv_state_flush(pool
->block_pool
.device
, state
);
1420 struct gen8_border_color
{
1425 /* Pad out to 64 bytes */
1430 anv_device_init_border_colors(struct anv_device
*device
)
1432 static const struct gen8_border_color border_colors
[] = {
1433 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
1434 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
1435 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
1436 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
1437 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
1438 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
1441 device
->border_colors
= anv_state_pool_emit_data(&device
->dynamic_state_pool
,
1442 sizeof(border_colors
), 64,
1447 anv_device_init_trivial_batch(struct anv_device
*device
)
1449 anv_bo_init_new(&device
->trivial_batch_bo
, device
, 4096);
1451 if (device
->instance
->physicalDevice
.has_exec_async
)
1452 device
->trivial_batch_bo
.flags
|= EXEC_OBJECT_ASYNC
;
1454 if (device
->instance
->physicalDevice
.use_softpin
)
1455 device
->trivial_batch_bo
.flags
|= EXEC_OBJECT_PINNED
;
1457 anv_vma_alloc(device
, &device
->trivial_batch_bo
);
1459 void *map
= anv_gem_mmap(device
, device
->trivial_batch_bo
.gem_handle
,
1462 struct anv_batch batch
= {
1468 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
1469 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
1471 if (!device
->info
.has_llc
)
1472 gen_clflush_range(map
, batch
.next
- map
);
1474 anv_gem_munmap(map
, device
->trivial_batch_bo
.size
);
1477 VkResult
anv_EnumerateDeviceExtensionProperties(
1478 VkPhysicalDevice physicalDevice
,
1479 const char* pLayerName
,
1480 uint32_t* pPropertyCount
,
1481 VkExtensionProperties
* pProperties
)
1483 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
1484 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
1486 for (int i
= 0; i
< ANV_DEVICE_EXTENSION_COUNT
; i
++) {
1487 if (device
->supported_extensions
.extensions
[i
]) {
1488 vk_outarray_append(&out
, prop
) {
1489 *prop
= anv_device_extensions
[i
];
1494 return vk_outarray_status(&out
);
1498 anv_device_init_dispatch(struct anv_device
*device
)
1500 const struct anv_dispatch_table
*genX_table
;
1501 switch (device
->info
.gen
) {
1503 genX_table
= &gen11_dispatch_table
;
1506 genX_table
= &gen10_dispatch_table
;
1509 genX_table
= &gen9_dispatch_table
;
1512 genX_table
= &gen8_dispatch_table
;
1515 if (device
->info
.is_haswell
)
1516 genX_table
= &gen75_dispatch_table
;
1518 genX_table
= &gen7_dispatch_table
;
1521 unreachable("unsupported gen\n");
1524 for (unsigned i
= 0; i
< ARRAY_SIZE(device
->dispatch
.entrypoints
); i
++) {
1525 /* Vulkan requires that entrypoints for extensions which have not been
1526 * enabled must not be advertised.
1528 if (!anv_entrypoint_is_enabled(i
, device
->instance
->app_info
.api_version
,
1529 &device
->instance
->enabled_extensions
,
1530 &device
->enabled_extensions
)) {
1531 device
->dispatch
.entrypoints
[i
] = NULL
;
1532 } else if (genX_table
->entrypoints
[i
]) {
1533 device
->dispatch
.entrypoints
[i
] = genX_table
->entrypoints
[i
];
1535 device
->dispatch
.entrypoints
[i
] = anv_dispatch_table
.entrypoints
[i
];
1541 vk_priority_to_gen(int priority
)
1544 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT
:
1545 return GEN_CONTEXT_LOW_PRIORITY
;
1546 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
:
1547 return GEN_CONTEXT_MEDIUM_PRIORITY
;
1548 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT
:
1549 return GEN_CONTEXT_HIGH_PRIORITY
;
1550 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT
:
1551 return GEN_CONTEXT_REALTIME_PRIORITY
;
1553 unreachable("Invalid priority");
1558 anv_device_init_hiz_clear_batch(struct anv_device
*device
)
1560 anv_bo_init_new(&device
->hiz_clear_bo
, device
, 4096);
1561 uint32_t *map
= anv_gem_mmap(device
, device
->hiz_clear_bo
.gem_handle
,
1564 union isl_color_value hiz_clear
= { .u32
= { 0, } };
1565 hiz_clear
.f32
[0] = ANV_HZ_FC_VAL
;
1567 memcpy(map
, hiz_clear
.u32
, sizeof(hiz_clear
.u32
));
1568 anv_gem_munmap(map
, device
->hiz_clear_bo
.size
);
1571 VkResult
anv_CreateDevice(
1572 VkPhysicalDevice physicalDevice
,
1573 const VkDeviceCreateInfo
* pCreateInfo
,
1574 const VkAllocationCallbacks
* pAllocator
,
1577 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
1579 struct anv_device
*device
;
1581 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO
);
1583 struct anv_device_extension_table enabled_extensions
= { };
1584 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
1586 for (idx
= 0; idx
< ANV_DEVICE_EXTENSION_COUNT
; idx
++) {
1587 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
1588 anv_device_extensions
[idx
].extensionName
) == 0)
1592 if (idx
>= ANV_DEVICE_EXTENSION_COUNT
)
1593 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
1595 if (!physical_device
->supported_extensions
.extensions
[idx
])
1596 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
1598 enabled_extensions
.extensions
[idx
] = true;
1601 /* Check enabled features */
1602 if (pCreateInfo
->pEnabledFeatures
) {
1603 VkPhysicalDeviceFeatures supported_features
;
1604 anv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
1605 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
1606 VkBool32
*enabled_feature
= (VkBool32
*)pCreateInfo
->pEnabledFeatures
;
1607 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
1608 for (uint32_t i
= 0; i
< num_features
; i
++) {
1609 if (enabled_feature
[i
] && !supported_feature
[i
])
1610 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
1614 /* Check requested queues and fail if we are requested to create any
1615 * queues with flags we don't support.
1617 assert(pCreateInfo
->queueCreateInfoCount
> 0);
1618 for (uint32_t i
= 0; i
< pCreateInfo
->queueCreateInfoCount
; i
++) {
1619 if (pCreateInfo
->pQueueCreateInfos
[i
].flags
!= 0)
1620 return vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1623 /* Check if client specified queue priority. */
1624 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*queue_priority
=
1625 vk_find_struct_const(pCreateInfo
->pQueueCreateInfos
[0].pNext
,
1626 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT
);
1628 VkQueueGlobalPriorityEXT priority
=
1629 queue_priority
? queue_priority
->globalPriority
:
1630 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
;
1632 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
1634 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
1636 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1638 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1639 device
->instance
= physical_device
->instance
;
1640 device
->chipset_id
= physical_device
->chipset_id
;
1641 device
->no_hw
= physical_device
->no_hw
;
1642 device
->lost
= false;
1645 device
->alloc
= *pAllocator
;
1647 device
->alloc
= physical_device
->instance
->alloc
;
1649 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
1650 device
->fd
= open(physical_device
->path
, O_RDWR
| O_CLOEXEC
);
1651 if (device
->fd
== -1) {
1652 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1656 device
->context_id
= anv_gem_create_context(device
);
1657 if (device
->context_id
== -1) {
1658 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1662 if (physical_device
->use_softpin
) {
1663 if (pthread_mutex_init(&device
->vma_mutex
, NULL
) != 0) {
1664 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1668 /* keep the page with address zero out of the allocator */
1669 util_vma_heap_init(&device
->vma_lo
, LOW_HEAP_MIN_ADDRESS
, LOW_HEAP_SIZE
);
1670 device
->vma_lo_available
=
1671 physical_device
->memory
.heaps
[physical_device
->memory
.heap_count
- 1].size
;
1673 /* Leave the last 4GiB out of the high vma range, so that no state base
1674 * address + size can overflow 48 bits. For more information see the
1675 * comment about Wa32bitGeneralStateOffset in anv_allocator.c
1677 util_vma_heap_init(&device
->vma_hi
, HIGH_HEAP_MIN_ADDRESS
,
1679 device
->vma_hi_available
= physical_device
->memory
.heap_count
== 1 ? 0 :
1680 physical_device
->memory
.heaps
[0].size
;
1683 /* As per spec, the driver implementation may deny requests to acquire
1684 * a priority above the default priority (MEDIUM) if the caller does not
1685 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
1688 if (physical_device
->has_context_priority
) {
1689 int err
= anv_gem_set_context_param(device
->fd
, device
->context_id
,
1690 I915_CONTEXT_PARAM_PRIORITY
,
1691 vk_priority_to_gen(priority
));
1692 if (err
!= 0 && priority
> VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
) {
1693 result
= vk_error(VK_ERROR_NOT_PERMITTED_EXT
);
1698 device
->info
= physical_device
->info
;
1699 device
->isl_dev
= physical_device
->isl_dev
;
1701 /* On Broadwell and later, we can use batch chaining to more efficiently
1702 * implement growing command buffers. Prior to Haswell, the kernel
1703 * command parser gets in the way and we have to fall back to growing
1706 device
->can_chain_batches
= device
->info
.gen
>= 8;
1708 device
->robust_buffer_access
= pCreateInfo
->pEnabledFeatures
&&
1709 pCreateInfo
->pEnabledFeatures
->robustBufferAccess
;
1710 device
->enabled_extensions
= enabled_extensions
;
1712 anv_device_init_dispatch(device
);
1714 if (pthread_mutex_init(&device
->mutex
, NULL
) != 0) {
1715 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1716 goto fail_context_id
;
1719 pthread_condattr_t condattr
;
1720 if (pthread_condattr_init(&condattr
) != 0) {
1721 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1724 if (pthread_condattr_setclock(&condattr
, CLOCK_MONOTONIC
) != 0) {
1725 pthread_condattr_destroy(&condattr
);
1726 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1729 if (pthread_cond_init(&device
->queue_submit
, NULL
) != 0) {
1730 pthread_condattr_destroy(&condattr
);
1731 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1734 pthread_condattr_destroy(&condattr
);
1737 (physical_device
->supports_48bit_addresses
? EXEC_OBJECT_SUPPORTS_48B_ADDRESS
: 0) |
1738 (physical_device
->has_exec_async
? EXEC_OBJECT_ASYNC
: 0) |
1739 (physical_device
->has_exec_capture
? EXEC_OBJECT_CAPTURE
: 0) |
1740 (physical_device
->use_softpin
? EXEC_OBJECT_PINNED
: 0);
1742 anv_bo_pool_init(&device
->batch_bo_pool
, device
, bo_flags
);
1744 result
= anv_bo_cache_init(&device
->bo_cache
);
1745 if (result
!= VK_SUCCESS
)
1746 goto fail_batch_bo_pool
;
1748 if (!physical_device
->use_softpin
)
1749 bo_flags
&= ~EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
1751 result
= anv_state_pool_init(&device
->dynamic_state_pool
, device
,
1752 DYNAMIC_STATE_POOL_MIN_ADDRESS
,
1755 if (result
!= VK_SUCCESS
)
1758 result
= anv_state_pool_init(&device
->instruction_state_pool
, device
,
1759 INSTRUCTION_STATE_POOL_MIN_ADDRESS
,
1762 if (result
!= VK_SUCCESS
)
1763 goto fail_dynamic_state_pool
;
1765 result
= anv_state_pool_init(&device
->surface_state_pool
, device
,
1766 SURFACE_STATE_POOL_MIN_ADDRESS
,
1769 if (result
!= VK_SUCCESS
)
1770 goto fail_instruction_state_pool
;
1772 if (physical_device
->use_softpin
) {
1773 result
= anv_state_pool_init(&device
->binding_table_pool
, device
,
1774 BINDING_TABLE_POOL_MIN_ADDRESS
,
1777 if (result
!= VK_SUCCESS
)
1778 goto fail_surface_state_pool
;
1781 result
= anv_bo_init_new(&device
->workaround_bo
, device
, 1024);
1782 if (result
!= VK_SUCCESS
)
1783 goto fail_binding_table_pool
;
1785 if (physical_device
->use_softpin
)
1786 device
->workaround_bo
.flags
|= EXEC_OBJECT_PINNED
;
1788 if (!anv_vma_alloc(device
, &device
->workaround_bo
))
1789 goto fail_workaround_bo
;
1791 anv_device_init_trivial_batch(device
);
1793 if (device
->info
.gen
>= 10)
1794 anv_device_init_hiz_clear_batch(device
);
1796 anv_scratch_pool_init(device
, &device
->scratch_pool
);
1798 anv_queue_init(device
, &device
->queue
);
1800 switch (device
->info
.gen
) {
1802 if (!device
->info
.is_haswell
)
1803 result
= gen7_init_device_state(device
);
1805 result
= gen75_init_device_state(device
);
1808 result
= gen8_init_device_state(device
);
1811 result
= gen9_init_device_state(device
);
1814 result
= gen10_init_device_state(device
);
1817 result
= gen11_init_device_state(device
);
1820 /* Shouldn't get here as we don't create physical devices for any other
1822 unreachable("unhandled gen");
1824 if (result
!= VK_SUCCESS
)
1825 goto fail_workaround_bo
;
1827 anv_pipeline_cache_init(&device
->default_pipeline_cache
, device
, true);
1829 anv_device_init_blorp(device
);
1831 anv_device_init_border_colors(device
);
1833 *pDevice
= anv_device_to_handle(device
);
1838 anv_queue_finish(&device
->queue
);
1839 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
1840 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
1841 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
1842 fail_binding_table_pool
:
1843 if (physical_device
->use_softpin
)
1844 anv_state_pool_finish(&device
->binding_table_pool
);
1845 fail_surface_state_pool
:
1846 anv_state_pool_finish(&device
->surface_state_pool
);
1847 fail_instruction_state_pool
:
1848 anv_state_pool_finish(&device
->instruction_state_pool
);
1849 fail_dynamic_state_pool
:
1850 anv_state_pool_finish(&device
->dynamic_state_pool
);
1852 anv_bo_cache_finish(&device
->bo_cache
);
1854 anv_bo_pool_finish(&device
->batch_bo_pool
);
1855 pthread_cond_destroy(&device
->queue_submit
);
1857 pthread_mutex_destroy(&device
->mutex
);
1859 anv_gem_destroy_context(device
, device
->context_id
);
1863 vk_free(&device
->alloc
, device
);
1868 void anv_DestroyDevice(
1870 const VkAllocationCallbacks
* pAllocator
)
1872 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1873 struct anv_physical_device
*physical_device
;
1878 physical_device
= &device
->instance
->physicalDevice
;
1880 anv_device_finish_blorp(device
);
1882 anv_pipeline_cache_finish(&device
->default_pipeline_cache
);
1884 anv_queue_finish(&device
->queue
);
1886 #ifdef HAVE_VALGRIND
1887 /* We only need to free these to prevent valgrind errors. The backing
1888 * BO will go away in a couple of lines so we don't actually leak.
1890 anv_state_pool_free(&device
->dynamic_state_pool
, device
->border_colors
);
1893 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
1895 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
1896 anv_vma_free(device
, &device
->workaround_bo
);
1897 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
1899 anv_vma_free(device
, &device
->trivial_batch_bo
);
1900 anv_gem_close(device
, device
->trivial_batch_bo
.gem_handle
);
1901 if (device
->info
.gen
>= 10)
1902 anv_gem_close(device
, device
->hiz_clear_bo
.gem_handle
);
1904 if (physical_device
->use_softpin
)
1905 anv_state_pool_finish(&device
->binding_table_pool
);
1906 anv_state_pool_finish(&device
->surface_state_pool
);
1907 anv_state_pool_finish(&device
->instruction_state_pool
);
1908 anv_state_pool_finish(&device
->dynamic_state_pool
);
1910 anv_bo_cache_finish(&device
->bo_cache
);
1912 anv_bo_pool_finish(&device
->batch_bo_pool
);
1914 pthread_cond_destroy(&device
->queue_submit
);
1915 pthread_mutex_destroy(&device
->mutex
);
1917 anv_gem_destroy_context(device
, device
->context_id
);
1921 vk_free(&device
->alloc
, device
);
1924 VkResult
anv_EnumerateInstanceLayerProperties(
1925 uint32_t* pPropertyCount
,
1926 VkLayerProperties
* pProperties
)
1928 if (pProperties
== NULL
) {
1929 *pPropertyCount
= 0;
1933 /* None supported at this time */
1934 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1937 VkResult
anv_EnumerateDeviceLayerProperties(
1938 VkPhysicalDevice physicalDevice
,
1939 uint32_t* pPropertyCount
,
1940 VkLayerProperties
* pProperties
)
1942 if (pProperties
== NULL
) {
1943 *pPropertyCount
= 0;
1947 /* None supported at this time */
1948 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1951 void anv_GetDeviceQueue(
1953 uint32_t queueNodeIndex
,
1954 uint32_t queueIndex
,
1957 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1959 assert(queueIndex
== 0);
1961 *pQueue
= anv_queue_to_handle(&device
->queue
);
1964 void anv_GetDeviceQueue2(
1966 const VkDeviceQueueInfo2
* pQueueInfo
,
1969 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1971 assert(pQueueInfo
->queueIndex
== 0);
1973 if (pQueueInfo
->flags
== device
->queue
.flags
)
1974 *pQueue
= anv_queue_to_handle(&device
->queue
);
1980 anv_device_query_status(struct anv_device
*device
)
1982 /* This isn't likely as most of the callers of this function already check
1983 * for it. However, it doesn't hurt to check and it potentially lets us
1986 if (unlikely(device
->lost
))
1987 return VK_ERROR_DEVICE_LOST
;
1989 uint32_t active
, pending
;
1990 int ret
= anv_gem_gpu_get_reset_stats(device
, &active
, &pending
);
1992 /* We don't know the real error. */
1993 device
->lost
= true;
1994 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
1995 "get_reset_stats failed: %m");
1999 device
->lost
= true;
2000 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
2001 "GPU hung on one of our command buffers");
2002 } else if (pending
) {
2003 device
->lost
= true;
2004 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
2005 "GPU hung with commands in-flight");
2012 anv_device_bo_busy(struct anv_device
*device
, struct anv_bo
*bo
)
2014 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
2015 * Other usages of the BO (such as on different hardware) will not be
2016 * flagged as "busy" by this ioctl. Use with care.
2018 int ret
= anv_gem_busy(device
, bo
->gem_handle
);
2020 return VK_NOT_READY
;
2021 } else if (ret
== -1) {
2022 /* We don't know the real error. */
2023 device
->lost
= true;
2024 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
2025 "gem wait failed: %m");
2028 /* Query for device status after the busy call. If the BO we're checking
2029 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
2030 * client because it clearly doesn't have valid data. Yes, this most
2031 * likely means an ioctl, but we just did an ioctl to query the busy status
2032 * so it's no great loss.
2034 return anv_device_query_status(device
);
2038 anv_device_wait(struct anv_device
*device
, struct anv_bo
*bo
,
2041 int ret
= anv_gem_wait(device
, bo
->gem_handle
, &timeout
);
2042 if (ret
== -1 && errno
== ETIME
) {
2044 } else if (ret
== -1) {
2045 /* We don't know the real error. */
2046 device
->lost
= true;
2047 return vk_errorf(device
->instance
, device
, VK_ERROR_DEVICE_LOST
,
2048 "gem wait failed: %m");
2051 /* Query for device status after the wait. If the BO we're waiting on got
2052 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
2053 * because it clearly doesn't have valid data. Yes, this most likely means
2054 * an ioctl, but we just did an ioctl to wait so it's no great loss.
2056 return anv_device_query_status(device
);
2059 VkResult
anv_DeviceWaitIdle(
2062 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2063 if (unlikely(device
->lost
))
2064 return VK_ERROR_DEVICE_LOST
;
2066 struct anv_batch batch
;
2069 batch
.start
= batch
.next
= cmds
;
2070 batch
.end
= (void *) cmds
+ sizeof(cmds
);
2072 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
2073 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
2075 return anv_device_submit_simple_batch(device
, &batch
);
2079 anv_vma_alloc(struct anv_device
*device
, struct anv_bo
*bo
)
2081 if (!(bo
->flags
& EXEC_OBJECT_PINNED
))
2084 pthread_mutex_lock(&device
->vma_mutex
);
2088 if (bo
->flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
&&
2089 device
->vma_hi_available
>= bo
->size
) {
2090 uint64_t addr
= util_vma_heap_alloc(&device
->vma_hi
, bo
->size
, 4096);
2092 bo
->offset
= gen_canonical_address(addr
);
2093 assert(addr
== gen_48b_address(bo
->offset
));
2094 device
->vma_hi_available
-= bo
->size
;
2098 if (bo
->offset
== 0 && device
->vma_lo_available
>= bo
->size
) {
2099 uint64_t addr
= util_vma_heap_alloc(&device
->vma_lo
, bo
->size
, 4096);
2101 bo
->offset
= gen_canonical_address(addr
);
2102 assert(addr
== gen_48b_address(bo
->offset
));
2103 device
->vma_lo_available
-= bo
->size
;
2107 pthread_mutex_unlock(&device
->vma_mutex
);
2109 return bo
->offset
!= 0;
2113 anv_vma_free(struct anv_device
*device
, struct anv_bo
*bo
)
2115 if (!(bo
->flags
& EXEC_OBJECT_PINNED
))
2118 const uint64_t addr_48b
= gen_48b_address(bo
->offset
);
2120 pthread_mutex_lock(&device
->vma_mutex
);
2122 if (addr_48b
>= LOW_HEAP_MIN_ADDRESS
&&
2123 addr_48b
<= LOW_HEAP_MAX_ADDRESS
) {
2124 util_vma_heap_free(&device
->vma_lo
, addr_48b
, bo
->size
);
2125 device
->vma_lo_available
+= bo
->size
;
2127 assert(addr_48b
>= HIGH_HEAP_MIN_ADDRESS
&&
2128 addr_48b
<= HIGH_HEAP_MAX_ADDRESS
);
2129 util_vma_heap_free(&device
->vma_hi
, addr_48b
, bo
->size
);
2130 device
->vma_hi_available
+= bo
->size
;
2133 pthread_mutex_unlock(&device
->vma_mutex
);
2139 anv_bo_init_new(struct anv_bo
*bo
, struct anv_device
*device
, uint64_t size
)
2141 uint32_t gem_handle
= anv_gem_create(device
, size
);
2143 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
2145 anv_bo_init(bo
, gem_handle
, size
);
2150 VkResult
anv_AllocateMemory(
2152 const VkMemoryAllocateInfo
* pAllocateInfo
,
2153 const VkAllocationCallbacks
* pAllocator
,
2154 VkDeviceMemory
* pMem
)
2156 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2157 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2158 struct anv_device_memory
*mem
;
2159 VkResult result
= VK_SUCCESS
;
2161 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
2163 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
2164 assert(pAllocateInfo
->allocationSize
> 0);
2166 if (pAllocateInfo
->allocationSize
> MAX_MEMORY_ALLOCATION_SIZE
)
2167 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
2169 /* FINISHME: Fail if allocation request exceeds heap size. */
2171 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
2172 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2174 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2176 assert(pAllocateInfo
->memoryTypeIndex
< pdevice
->memory
.type_count
);
2177 mem
->type
= &pdevice
->memory
.types
[pAllocateInfo
->memoryTypeIndex
];
2181 uint64_t bo_flags
= 0;
2183 assert(mem
->type
->heapIndex
< pdevice
->memory
.heap_count
);
2184 if (pdevice
->memory
.heaps
[mem
->type
->heapIndex
].supports_48bit_addresses
)
2185 bo_flags
|= EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
2187 const struct wsi_memory_allocate_info
*wsi_info
=
2188 vk_find_struct_const(pAllocateInfo
->pNext
, WSI_MEMORY_ALLOCATE_INFO_MESA
);
2189 if (wsi_info
&& wsi_info
->implicit_sync
) {
2190 /* We need to set the WRITE flag on window system buffers so that GEM
2191 * will know we're writing to them and synchronize uses on other rings
2192 * (eg if the display server uses the blitter ring).
2194 bo_flags
|= EXEC_OBJECT_WRITE
;
2195 } else if (pdevice
->has_exec_async
) {
2196 bo_flags
|= EXEC_OBJECT_ASYNC
;
2199 if (pdevice
->use_softpin
)
2200 bo_flags
|= EXEC_OBJECT_PINNED
;
2202 const VkImportMemoryFdInfoKHR
*fd_info
=
2203 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_FD_INFO_KHR
);
2205 /* The Vulkan spec permits handleType to be 0, in which case the struct is
2208 if (fd_info
&& fd_info
->handleType
) {
2209 /* At the moment, we support only the below handle types. */
2210 assert(fd_info
->handleType
==
2211 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
2212 fd_info
->handleType
==
2213 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
2215 result
= anv_bo_cache_import(device
, &device
->bo_cache
,
2216 fd_info
->fd
, bo_flags
, &mem
->bo
);
2217 if (result
!= VK_SUCCESS
)
2220 VkDeviceSize aligned_alloc_size
=
2221 align_u64(pAllocateInfo
->allocationSize
, 4096);
2223 /* For security purposes, we reject importing the bo if it's smaller
2224 * than the requested allocation size. This prevents a malicious client
2225 * from passing a buffer to a trusted client, lying about the size, and
2226 * telling the trusted client to try and texture from an image that goes
2227 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
2228 * in the trusted client. The trusted client can protect itself against
2229 * this sort of attack but only if it can trust the buffer size.
2231 if (mem
->bo
->size
< aligned_alloc_size
) {
2232 result
= vk_errorf(device
->instance
, device
,
2233 VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
,
2234 "aligned allocationSize too large for "
2235 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR: "
2236 "%"PRIu64
"B > %"PRIu64
"B",
2237 aligned_alloc_size
, mem
->bo
->size
);
2238 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
2242 /* From the Vulkan spec:
2244 * "Importing memory from a file descriptor transfers ownership of
2245 * the file descriptor from the application to the Vulkan
2246 * implementation. The application must not perform any operations on
2247 * the file descriptor after a successful import."
2249 * If the import fails, we leave the file descriptor open.
2253 result
= anv_bo_cache_alloc(device
, &device
->bo_cache
,
2254 pAllocateInfo
->allocationSize
, bo_flags
,
2256 if (result
!= VK_SUCCESS
)
2259 const VkMemoryDedicatedAllocateInfoKHR
*dedicated_info
=
2260 vk_find_struct_const(pAllocateInfo
->pNext
, MEMORY_DEDICATED_ALLOCATE_INFO_KHR
);
2261 if (dedicated_info
&& dedicated_info
->image
!= VK_NULL_HANDLE
) {
2262 ANV_FROM_HANDLE(anv_image
, image
, dedicated_info
->image
);
2264 /* Some legacy (non-modifiers) consumers need the tiling to be set on
2265 * the BO. In this case, we have a dedicated allocation.
2267 if (image
->needs_set_tiling
) {
2268 const uint32_t i915_tiling
=
2269 isl_tiling_to_i915_tiling(image
->planes
[0].surface
.isl
.tiling
);
2270 int ret
= anv_gem_set_tiling(device
, mem
->bo
->gem_handle
,
2271 image
->planes
[0].surface
.isl
.row_pitch
,
2274 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
2275 return vk_errorf(device
->instance
, NULL
,
2276 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
2277 "failed to set BO tiling: %m");
2283 *pMem
= anv_device_memory_to_handle(mem
);
2288 vk_free2(&device
->alloc
, pAllocator
, mem
);
2293 VkResult
anv_GetMemoryFdKHR(
2295 const VkMemoryGetFdInfoKHR
* pGetFdInfo
,
2298 ANV_FROM_HANDLE(anv_device
, dev
, device_h
);
2299 ANV_FROM_HANDLE(anv_device_memory
, mem
, pGetFdInfo
->memory
);
2301 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
2303 assert(pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
2304 pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
2306 return anv_bo_cache_export(dev
, &dev
->bo_cache
, mem
->bo
, pFd
);
2309 VkResult
anv_GetMemoryFdPropertiesKHR(
2311 VkExternalMemoryHandleTypeFlagBitsKHR handleType
,
2313 VkMemoryFdPropertiesKHR
* pMemoryFdProperties
)
2315 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2316 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2318 switch (handleType
) {
2319 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
:
2320 /* dma-buf can be imported as any memory type */
2321 pMemoryFdProperties
->memoryTypeBits
=
2322 (1 << pdevice
->memory
.type_count
) - 1;
2326 /* The valid usage section for this function says:
2328 * "handleType must not be one of the handle types defined as
2331 * So opaque handle types fall into the default "unsupported" case.
2333 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
2337 void anv_FreeMemory(
2339 VkDeviceMemory _mem
,
2340 const VkAllocationCallbacks
* pAllocator
)
2342 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2343 ANV_FROM_HANDLE(anv_device_memory
, mem
, _mem
);
2349 anv_UnmapMemory(_device
, _mem
);
2351 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
2353 vk_free2(&device
->alloc
, pAllocator
, mem
);
2356 VkResult
anv_MapMemory(
2358 VkDeviceMemory _memory
,
2359 VkDeviceSize offset
,
2361 VkMemoryMapFlags flags
,
2364 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2365 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
2372 if (size
== VK_WHOLE_SIZE
)
2373 size
= mem
->bo
->size
- offset
;
2375 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
2377 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
2378 * assert(size != 0);
2379 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
2380 * equal to the size of the memory minus offset
2383 assert(offset
+ size
<= mem
->bo
->size
);
2385 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
2386 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
2387 * at a time is valid. We could just mmap up front and return an offset
2388 * pointer here, but that may exhaust virtual memory on 32 bit
2391 uint32_t gem_flags
= 0;
2393 if (!device
->info
.has_llc
&&
2394 (mem
->type
->propertyFlags
& VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
))
2395 gem_flags
|= I915_MMAP_WC
;
2397 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
2398 uint64_t map_offset
= offset
& ~4095ull;
2399 assert(offset
>= map_offset
);
2400 uint64_t map_size
= (offset
+ size
) - map_offset
;
2402 /* Let's map whole pages */
2403 map_size
= align_u64(map_size
, 4096);
2405 void *map
= anv_gem_mmap(device
, mem
->bo
->gem_handle
,
2406 map_offset
, map_size
, gem_flags
);
2407 if (map
== MAP_FAILED
)
2408 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
2411 mem
->map_size
= map_size
;
2413 *ppData
= mem
->map
+ (offset
- map_offset
);
2418 void anv_UnmapMemory(
2420 VkDeviceMemory _memory
)
2422 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
2427 anv_gem_munmap(mem
->map
, mem
->map_size
);
2434 clflush_mapped_ranges(struct anv_device
*device
,
2436 const VkMappedMemoryRange
*ranges
)
2438 for (uint32_t i
= 0; i
< count
; i
++) {
2439 ANV_FROM_HANDLE(anv_device_memory
, mem
, ranges
[i
].memory
);
2440 if (ranges
[i
].offset
>= mem
->map_size
)
2443 gen_clflush_range(mem
->map
+ ranges
[i
].offset
,
2444 MIN2(ranges
[i
].size
, mem
->map_size
- ranges
[i
].offset
));
2448 VkResult
anv_FlushMappedMemoryRanges(
2450 uint32_t memoryRangeCount
,
2451 const VkMappedMemoryRange
* pMemoryRanges
)
2453 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2455 if (device
->info
.has_llc
)
2458 /* Make sure the writes we're flushing have landed. */
2459 __builtin_ia32_mfence();
2461 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
2466 VkResult
anv_InvalidateMappedMemoryRanges(
2468 uint32_t memoryRangeCount
,
2469 const VkMappedMemoryRange
* pMemoryRanges
)
2471 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2473 if (device
->info
.has_llc
)
2476 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
2478 /* Make sure no reads get moved up above the invalidate. */
2479 __builtin_ia32_mfence();
2484 void anv_GetBufferMemoryRequirements(
2487 VkMemoryRequirements
* pMemoryRequirements
)
2489 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
2490 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2491 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2493 /* The Vulkan spec (git aaed022) says:
2495 * memoryTypeBits is a bitfield and contains one bit set for every
2496 * supported memory type for the resource. The bit `1<<i` is set if and
2497 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
2498 * structure for the physical device is supported.
2500 uint32_t memory_types
= 0;
2501 for (uint32_t i
= 0; i
< pdevice
->memory
.type_count
; i
++) {
2502 uint32_t valid_usage
= pdevice
->memory
.types
[i
].valid_buffer_usage
;
2503 if ((valid_usage
& buffer
->usage
) == buffer
->usage
)
2504 memory_types
|= (1u << i
);
2507 /* Base alignment requirement of a cache line */
2508 uint32_t alignment
= 16;
2510 /* We need an alignment of 32 for pushing UBOs */
2511 if (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
)
2512 alignment
= MAX2(alignment
, 32);
2514 pMemoryRequirements
->size
= buffer
->size
;
2515 pMemoryRequirements
->alignment
= alignment
;
2517 /* Storage and Uniform buffers should have their size aligned to
2518 * 32-bits to avoid boundary checks when last DWord is not complete.
2519 * This would ensure that not internal padding would be needed for
2522 if (device
->robust_buffer_access
&&
2523 (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
||
2524 buffer
->usage
& VK_BUFFER_USAGE_STORAGE_BUFFER_BIT
))
2525 pMemoryRequirements
->size
= align_u64(buffer
->size
, 4);
2527 pMemoryRequirements
->memoryTypeBits
= memory_types
;
2530 void anv_GetBufferMemoryRequirements2(
2532 const VkBufferMemoryRequirementsInfo2
* pInfo
,
2533 VkMemoryRequirements2
* pMemoryRequirements
)
2535 anv_GetBufferMemoryRequirements(_device
, pInfo
->buffer
,
2536 &pMemoryRequirements
->memoryRequirements
);
2538 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
2539 switch (ext
->sType
) {
2540 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
2541 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
2542 requirements
->prefersDedicatedAllocation
= VK_FALSE
;
2543 requirements
->requiresDedicatedAllocation
= VK_FALSE
;
2548 anv_debug_ignored_stype(ext
->sType
);
2554 void anv_GetImageMemoryRequirements(
2557 VkMemoryRequirements
* pMemoryRequirements
)
2559 ANV_FROM_HANDLE(anv_image
, image
, _image
);
2560 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2561 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2563 /* The Vulkan spec (git aaed022) says:
2565 * memoryTypeBits is a bitfield and contains one bit set for every
2566 * supported memory type for the resource. The bit `1<<i` is set if and
2567 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
2568 * structure for the physical device is supported.
2570 * All types are currently supported for images.
2572 uint32_t memory_types
= (1ull << pdevice
->memory
.type_count
) - 1;
2574 pMemoryRequirements
->size
= image
->size
;
2575 pMemoryRequirements
->alignment
= image
->alignment
;
2576 pMemoryRequirements
->memoryTypeBits
= memory_types
;
2579 void anv_GetImageMemoryRequirements2(
2581 const VkImageMemoryRequirementsInfo2
* pInfo
,
2582 VkMemoryRequirements2
* pMemoryRequirements
)
2584 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2585 ANV_FROM_HANDLE(anv_image
, image
, pInfo
->image
);
2587 anv_GetImageMemoryRequirements(_device
, pInfo
->image
,
2588 &pMemoryRequirements
->memoryRequirements
);
2590 vk_foreach_struct_const(ext
, pInfo
->pNext
) {
2591 switch (ext
->sType
) {
2592 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO
: {
2593 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2594 const VkImagePlaneMemoryRequirementsInfoKHR
*plane_reqs
=
2595 (const VkImagePlaneMemoryRequirementsInfoKHR
*) ext
;
2596 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
,
2597 plane_reqs
->planeAspect
);
2599 assert(image
->planes
[plane
].offset
== 0);
2601 /* The Vulkan spec (git aaed022) says:
2603 * memoryTypeBits is a bitfield and contains one bit set for every
2604 * supported memory type for the resource. The bit `1<<i` is set
2605 * if and only if the memory type `i` in the
2606 * VkPhysicalDeviceMemoryProperties structure for the physical
2607 * device is supported.
2609 * All types are currently supported for images.
2611 pMemoryRequirements
->memoryRequirements
.memoryTypeBits
=
2612 (1ull << pdevice
->memory
.type_count
) - 1;
2614 pMemoryRequirements
->memoryRequirements
.size
= image
->planes
[plane
].size
;
2615 pMemoryRequirements
->memoryRequirements
.alignment
=
2616 image
->planes
[plane
].alignment
;
2621 anv_debug_ignored_stype(ext
->sType
);
2626 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
2627 switch (ext
->sType
) {
2628 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
2629 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
2630 if (image
->needs_set_tiling
) {
2631 /* If we need to set the tiling for external consumers, we need a
2632 * dedicated allocation.
2634 * See also anv_AllocateMemory.
2636 requirements
->prefersDedicatedAllocation
= VK_TRUE
;
2637 requirements
->requiresDedicatedAllocation
= VK_TRUE
;
2639 requirements
->prefersDedicatedAllocation
= VK_FALSE
;
2640 requirements
->requiresDedicatedAllocation
= VK_FALSE
;
2646 anv_debug_ignored_stype(ext
->sType
);
2652 void anv_GetImageSparseMemoryRequirements(
2655 uint32_t* pSparseMemoryRequirementCount
,
2656 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
2658 *pSparseMemoryRequirementCount
= 0;
2661 void anv_GetImageSparseMemoryRequirements2(
2663 const VkImageSparseMemoryRequirementsInfo2
* pInfo
,
2664 uint32_t* pSparseMemoryRequirementCount
,
2665 VkSparseImageMemoryRequirements2
* pSparseMemoryRequirements
)
2667 *pSparseMemoryRequirementCount
= 0;
2670 void anv_GetDeviceMemoryCommitment(
2672 VkDeviceMemory memory
,
2673 VkDeviceSize
* pCommittedMemoryInBytes
)
2675 *pCommittedMemoryInBytes
= 0;
2679 anv_bind_buffer_memory(const VkBindBufferMemoryInfo
*pBindInfo
)
2681 ANV_FROM_HANDLE(anv_device_memory
, mem
, pBindInfo
->memory
);
2682 ANV_FROM_HANDLE(anv_buffer
, buffer
, pBindInfo
->buffer
);
2684 assert(pBindInfo
->sType
== VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
);
2687 assert((buffer
->usage
& mem
->type
->valid_buffer_usage
) == buffer
->usage
);
2688 buffer
->address
= (struct anv_address
) {
2690 .offset
= pBindInfo
->memoryOffset
,
2693 buffer
->address
= ANV_NULL_ADDRESS
;
2697 VkResult
anv_BindBufferMemory(
2700 VkDeviceMemory memory
,
2701 VkDeviceSize memoryOffset
)
2703 anv_bind_buffer_memory(
2704 &(VkBindBufferMemoryInfo
) {
2705 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
,
2708 .memoryOffset
= memoryOffset
,
2714 VkResult
anv_BindBufferMemory2(
2716 uint32_t bindInfoCount
,
2717 const VkBindBufferMemoryInfo
* pBindInfos
)
2719 for (uint32_t i
= 0; i
< bindInfoCount
; i
++)
2720 anv_bind_buffer_memory(&pBindInfos
[i
]);
2725 VkResult
anv_QueueBindSparse(
2727 uint32_t bindInfoCount
,
2728 const VkBindSparseInfo
* pBindInfo
,
2731 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
2732 if (unlikely(queue
->device
->lost
))
2733 return VK_ERROR_DEVICE_LOST
;
2735 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
2740 VkResult
anv_CreateEvent(
2742 const VkEventCreateInfo
* pCreateInfo
,
2743 const VkAllocationCallbacks
* pAllocator
,
2746 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2747 struct anv_state state
;
2748 struct anv_event
*event
;
2750 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_EVENT_CREATE_INFO
);
2752 state
= anv_state_pool_alloc(&device
->dynamic_state_pool
,
2755 event
->state
= state
;
2756 event
->semaphore
= VK_EVENT_RESET
;
2758 if (!device
->info
.has_llc
) {
2759 /* Make sure the writes we're flushing have landed. */
2760 __builtin_ia32_mfence();
2761 __builtin_ia32_clflush(event
);
2764 *pEvent
= anv_event_to_handle(event
);
2769 void anv_DestroyEvent(
2772 const VkAllocationCallbacks
* pAllocator
)
2774 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2775 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2780 anv_state_pool_free(&device
->dynamic_state_pool
, event
->state
);
2783 VkResult
anv_GetEventStatus(
2787 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2788 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2790 if (unlikely(device
->lost
))
2791 return VK_ERROR_DEVICE_LOST
;
2793 if (!device
->info
.has_llc
) {
2794 /* Invalidate read cache before reading event written by GPU. */
2795 __builtin_ia32_clflush(event
);
2796 __builtin_ia32_mfence();
2800 return event
->semaphore
;
2803 VkResult
anv_SetEvent(
2807 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2808 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2810 event
->semaphore
= VK_EVENT_SET
;
2812 if (!device
->info
.has_llc
) {
2813 /* Make sure the writes we're flushing have landed. */
2814 __builtin_ia32_mfence();
2815 __builtin_ia32_clflush(event
);
2821 VkResult
anv_ResetEvent(
2825 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2826 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2828 event
->semaphore
= VK_EVENT_RESET
;
2830 if (!device
->info
.has_llc
) {
2831 /* Make sure the writes we're flushing have landed. */
2832 __builtin_ia32_mfence();
2833 __builtin_ia32_clflush(event
);
2841 VkResult
anv_CreateBuffer(
2843 const VkBufferCreateInfo
* pCreateInfo
,
2844 const VkAllocationCallbacks
* pAllocator
,
2847 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2848 struct anv_buffer
*buffer
;
2850 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
2852 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
2853 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2855 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2857 buffer
->size
= pCreateInfo
->size
;
2858 buffer
->usage
= pCreateInfo
->usage
;
2859 buffer
->address
= ANV_NULL_ADDRESS
;
2861 *pBuffer
= anv_buffer_to_handle(buffer
);
2866 void anv_DestroyBuffer(
2869 const VkAllocationCallbacks
* pAllocator
)
2871 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2872 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
2877 vk_free2(&device
->alloc
, pAllocator
, buffer
);
2881 anv_fill_buffer_surface_state(struct anv_device
*device
, struct anv_state state
,
2882 enum isl_format format
,
2883 struct anv_address address
,
2884 uint32_t range
, uint32_t stride
)
2886 isl_buffer_fill_state(&device
->isl_dev
, state
.map
,
2887 .address
= anv_address_physical(address
),
2888 .mocs
= device
->default_mocs
,
2893 anv_state_flush(device
, state
);
2896 void anv_DestroySampler(
2899 const VkAllocationCallbacks
* pAllocator
)
2901 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2902 ANV_FROM_HANDLE(anv_sampler
, sampler
, _sampler
);
2907 vk_free2(&device
->alloc
, pAllocator
, sampler
);
2910 VkResult
anv_CreateFramebuffer(
2912 const VkFramebufferCreateInfo
* pCreateInfo
,
2913 const VkAllocationCallbacks
* pAllocator
,
2914 VkFramebuffer
* pFramebuffer
)
2916 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2917 struct anv_framebuffer
*framebuffer
;
2919 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
2921 size_t size
= sizeof(*framebuffer
) +
2922 sizeof(struct anv_image_view
*) * pCreateInfo
->attachmentCount
;
2923 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
2924 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2925 if (framebuffer
== NULL
)
2926 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2928 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
2929 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
2930 VkImageView _iview
= pCreateInfo
->pAttachments
[i
];
2931 framebuffer
->attachments
[i
] = anv_image_view_from_handle(_iview
);
2934 framebuffer
->width
= pCreateInfo
->width
;
2935 framebuffer
->height
= pCreateInfo
->height
;
2936 framebuffer
->layers
= pCreateInfo
->layers
;
2938 *pFramebuffer
= anv_framebuffer_to_handle(framebuffer
);
2943 void anv_DestroyFramebuffer(
2946 const VkAllocationCallbacks
* pAllocator
)
2948 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2949 ANV_FROM_HANDLE(anv_framebuffer
, fb
, _fb
);
2954 vk_free2(&device
->alloc
, pAllocator
, fb
);
2957 /* vk_icd.h does not declare this function, so we declare it here to
2958 * suppress Wmissing-prototypes.
2960 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2961 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
);
2963 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2964 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
)
2966 /* For the full details on loader interface versioning, see
2967 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
2968 * What follows is a condensed summary, to help you navigate the large and
2969 * confusing official doc.
2971 * - Loader interface v0 is incompatible with later versions. We don't
2974 * - In loader interface v1:
2975 * - The first ICD entrypoint called by the loader is
2976 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
2978 * - The ICD must statically expose no other Vulkan symbol unless it is
2979 * linked with -Bsymbolic.
2980 * - Each dispatchable Vulkan handle created by the ICD must be
2981 * a pointer to a struct whose first member is VK_LOADER_DATA. The
2982 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
2983 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
2984 * vkDestroySurfaceKHR(). The ICD must be capable of working with
2985 * such loader-managed surfaces.
2987 * - Loader interface v2 differs from v1 in:
2988 * - The first ICD entrypoint called by the loader is
2989 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
2990 * statically expose this entrypoint.
2992 * - Loader interface v3 differs from v2 in:
2993 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
2994 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
2995 * because the loader no longer does so.
2997 *pSupportedVersion
= MIN2(*pSupportedVersion
, 3u);