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-uapi/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"
40 #include "util/u_string.h"
43 #include "common/gen_defines.h"
45 #include "genxml/gen7_pack.h"
48 compiler_debug_log(void *data
, const char *fmt
, ...)
52 compiler_perf_log(void *data
, const char *fmt
, ...)
57 if (unlikely(INTEL_DEBUG
& DEBUG_PERF
))
58 intel_logd_v(fmt
, args
);
64 anv_compute_heap_size(int fd
, uint64_t gtt_size
)
66 /* Query the total ram from the system */
70 uint64_t total_ram
= (uint64_t)info
.totalram
* (uint64_t)info
.mem_unit
;
72 /* We don't want to burn too much ram with the GPU. If the user has 4GiB
73 * or less, we use at most half. If they have more than 4GiB, we use 3/4.
75 uint64_t available_ram
;
76 if (total_ram
<= 4ull * 1024ull * 1024ull * 1024ull)
77 available_ram
= total_ram
/ 2;
79 available_ram
= total_ram
* 3 / 4;
81 /* We also want to leave some padding for things we allocate in the driver,
82 * so don't go over 3/4 of the GTT either.
84 uint64_t available_gtt
= gtt_size
* 3 / 4;
86 return 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
= anv_compute_heap_size(fd
, gtt_size
);
112 if (heap_size
> (2ull << 30) && !device
->supports_48bit_addresses
) {
113 /* When running with an overridden PCI ID, we may get a GTT size from
114 * the kernel that is greater than 2 GiB but the execbuf check for 48bit
115 * address support can still fail. Just clamp the address space size to
116 * 2 GiB if we don't have 48-bit support.
118 intel_logw("%s:%d: The kernel reported a GTT size larger than 2 GiB but "
119 "not support for 48-bit addresses",
121 heap_size
= 2ull << 30;
124 if (heap_size
<= 3ull * (1ull << 30)) {
125 /* In this case, everything fits nicely into the 32-bit address space,
126 * so there's no need for supporting 48bit addresses on client-allocated
129 device
->memory
.heap_count
= 1;
130 device
->memory
.heaps
[0] = (struct anv_memory_heap
) {
132 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
133 .supports_48bit_addresses
= false,
136 /* Not everything will fit nicely into a 32-bit address space. In this
137 * case we need a 64-bit heap. Advertise a small 32-bit heap and a
138 * larger 48-bit heap. If we're in this case, then we have a total heap
139 * size larger than 3GiB which most likely means they have 8 GiB of
140 * video memory and so carving off 1 GiB for the 32-bit heap should be
143 const uint64_t heap_size_32bit
= 1ull << 30;
144 const uint64_t heap_size_48bit
= heap_size
- heap_size_32bit
;
146 assert(device
->supports_48bit_addresses
);
148 device
->memory
.heap_count
= 2;
149 device
->memory
.heaps
[0] = (struct anv_memory_heap
) {
150 .size
= heap_size_48bit
,
151 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
152 .supports_48bit_addresses
= true,
154 device
->memory
.heaps
[1] = (struct anv_memory_heap
) {
155 .size
= heap_size_32bit
,
156 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
157 .supports_48bit_addresses
= false,
161 uint32_t type_count
= 0;
162 for (uint32_t heap
= 0; heap
< device
->memory
.heap_count
; heap
++) {
163 uint32_t valid_buffer_usage
= ~0;
165 /* There appears to be a hardware issue in the VF cache where it only
166 * considers the bottom 32 bits of memory addresses. If you happen to
167 * have two vertex buffers which get placed exactly 4 GiB apart and use
168 * them in back-to-back draw calls, you can get collisions. In order to
169 * solve this problem, we require vertex and index buffers be bound to
170 * memory allocated out of the 32-bit heap.
172 if (device
->memory
.heaps
[heap
].supports_48bit_addresses
) {
173 valid_buffer_usage
&= ~(VK_BUFFER_USAGE_INDEX_BUFFER_BIT
|
174 VK_BUFFER_USAGE_VERTEX_BUFFER_BIT
);
177 if (device
->info
.has_llc
) {
178 /* Big core GPUs share LLC with the CPU and thus one memory type can be
179 * both cached and coherent at the same time.
181 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
182 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
183 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
184 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
185 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
187 .valid_buffer_usage
= valid_buffer_usage
,
190 /* The spec requires that we expose a host-visible, coherent memory
191 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
192 * to give the application a choice between cached, but not coherent and
193 * coherent but uncached (WC though).
195 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
196 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
197 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
198 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
200 .valid_buffer_usage
= valid_buffer_usage
,
202 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
203 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
204 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
205 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
207 .valid_buffer_usage
= valid_buffer_usage
,
211 device
->memory
.type_count
= type_count
;
217 anv_physical_device_init_uuids(struct anv_physical_device
*device
)
219 const struct build_id_note
*note
=
220 build_id_find_nhdr_for_addr(anv_physical_device_init_uuids
);
222 return vk_errorf(device
->instance
, device
,
223 VK_ERROR_INITIALIZATION_FAILED
,
224 "Failed to find build-id");
227 unsigned build_id_len
= build_id_length(note
);
228 if (build_id_len
< 20) {
229 return vk_errorf(device
->instance
, device
,
230 VK_ERROR_INITIALIZATION_FAILED
,
231 "build-id too short. It needs to be a SHA");
234 memcpy(device
->driver_build_sha1
, build_id_data(note
), 20);
236 struct mesa_sha1 sha1_ctx
;
238 STATIC_ASSERT(VK_UUID_SIZE
<= sizeof(sha1
));
240 /* The pipeline cache UUID is used for determining when a pipeline cache is
241 * invalid. It needs both a driver build and the PCI ID of the device.
243 _mesa_sha1_init(&sha1_ctx
);
244 _mesa_sha1_update(&sha1_ctx
, build_id_data(note
), build_id_len
);
245 _mesa_sha1_update(&sha1_ctx
, &device
->chipset_id
,
246 sizeof(device
->chipset_id
));
247 _mesa_sha1_final(&sha1_ctx
, sha1
);
248 memcpy(device
->pipeline_cache_uuid
, sha1
, VK_UUID_SIZE
);
250 /* The driver UUID is used for determining sharability of images and memory
251 * between two Vulkan instances in separate processes. People who want to
252 * share memory need to also check the device UUID (below) so all this
253 * needs to be is the build-id.
255 memcpy(device
->driver_uuid
, build_id_data(note
), VK_UUID_SIZE
);
257 /* The device UUID uniquely identifies the given device within the machine.
258 * Since we never have more than one device, this doesn't need to be a real
259 * UUID. However, on the off-chance that someone tries to use this to
260 * cache pre-tiled images or something of the like, we use the PCI ID and
261 * some bits of ISL info to ensure that this is safe.
263 _mesa_sha1_init(&sha1_ctx
);
264 _mesa_sha1_update(&sha1_ctx
, &device
->chipset_id
,
265 sizeof(device
->chipset_id
));
266 _mesa_sha1_update(&sha1_ctx
, &device
->isl_dev
.has_bit6_swizzling
,
267 sizeof(device
->isl_dev
.has_bit6_swizzling
));
268 _mesa_sha1_final(&sha1_ctx
, sha1
);
269 memcpy(device
->device_uuid
, sha1
, VK_UUID_SIZE
);
275 anv_physical_device_init_disk_cache(struct anv_physical_device
*device
)
277 #ifdef ENABLE_SHADER_CACHE
279 MAYBE_UNUSED
int len
= snprintf(renderer
, sizeof(renderer
), "anv_%04x",
281 assert(len
== sizeof(renderer
) - 2);
284 _mesa_sha1_format(timestamp
, device
->driver_build_sha1
);
286 const uint64_t driver_flags
=
287 brw_get_compiler_config_value(device
->compiler
);
288 device
->disk_cache
= disk_cache_create(renderer
, timestamp
, driver_flags
);
290 device
->disk_cache
= NULL
;
295 anv_physical_device_free_disk_cache(struct anv_physical_device
*device
)
297 #ifdef ENABLE_SHADER_CACHE
298 if (device
->disk_cache
)
299 disk_cache_destroy(device
->disk_cache
);
301 assert(device
->disk_cache
== NULL
);
306 anv_physical_device_init(struct anv_physical_device
*device
,
307 struct anv_instance
*instance
,
308 drmDevicePtr drm_device
)
310 const char *primary_path
= drm_device
->nodes
[DRM_NODE_PRIMARY
];
311 const char *path
= drm_device
->nodes
[DRM_NODE_RENDER
];
316 brw_process_intel_debug_variable();
318 fd
= open(path
, O_RDWR
| O_CLOEXEC
);
320 return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
322 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
323 device
->instance
= instance
;
325 assert(strlen(path
) < ARRAY_SIZE(device
->path
));
326 snprintf(device
->path
, ARRAY_SIZE(device
->path
), "%s", path
);
328 device
->no_hw
= getenv("INTEL_NO_HW") != NULL
;
330 const int pci_id_override
= gen_get_pci_device_id_override();
331 if (pci_id_override
< 0) {
332 device
->chipset_id
= anv_gem_get_param(fd
, I915_PARAM_CHIPSET_ID
);
333 if (!device
->chipset_id
) {
334 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
338 device
->chipset_id
= pci_id_override
;
339 device
->no_hw
= true;
342 device
->pci_info
.domain
= drm_device
->businfo
.pci
->domain
;
343 device
->pci_info
.bus
= drm_device
->businfo
.pci
->bus
;
344 device
->pci_info
.device
= drm_device
->businfo
.pci
->dev
;
345 device
->pci_info
.function
= drm_device
->businfo
.pci
->func
;
347 device
->name
= gen_get_device_name(device
->chipset_id
);
348 if (!gen_get_device_info(device
->chipset_id
, &device
->info
)) {
349 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
353 if (device
->info
.is_haswell
) {
354 intel_logw("Haswell Vulkan support is incomplete");
355 } else if (device
->info
.gen
== 7 && !device
->info
.is_baytrail
) {
356 intel_logw("Ivy Bridge Vulkan support is incomplete");
357 } else if (device
->info
.gen
== 7 && device
->info
.is_baytrail
) {
358 intel_logw("Bay Trail Vulkan support is incomplete");
359 } else if (device
->info
.gen
>= 8 && device
->info
.gen
<= 10) {
360 /* Gen8-10 fully supported */
361 } else if (device
->info
.gen
== 11) {
362 intel_logw("Vulkan is not yet fully supported on gen11.");
364 result
= vk_errorf(device
->instance
, device
,
365 VK_ERROR_INCOMPATIBLE_DRIVER
,
366 "Vulkan not yet supported on %s", device
->name
);
370 device
->cmd_parser_version
= -1;
371 if (device
->info
.gen
== 7) {
372 device
->cmd_parser_version
=
373 anv_gem_get_param(fd
, I915_PARAM_CMD_PARSER_VERSION
);
374 if (device
->cmd_parser_version
== -1) {
375 result
= vk_errorf(device
->instance
, device
,
376 VK_ERROR_INITIALIZATION_FAILED
,
377 "failed to get command parser version");
382 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_WAIT_TIMEOUT
)) {
383 result
= vk_errorf(device
->instance
, device
,
384 VK_ERROR_INITIALIZATION_FAILED
,
385 "kernel missing gem wait");
389 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_EXECBUF2
)) {
390 result
= vk_errorf(device
->instance
, device
,
391 VK_ERROR_INITIALIZATION_FAILED
,
392 "kernel missing execbuf2");
396 if (!device
->info
.has_llc
&&
397 anv_gem_get_param(fd
, I915_PARAM_MMAP_VERSION
) < 1) {
398 result
= vk_errorf(device
->instance
, device
,
399 VK_ERROR_INITIALIZATION_FAILED
,
400 "kernel missing wc mmap");
404 result
= anv_physical_device_init_heaps(device
, fd
);
405 if (result
!= VK_SUCCESS
)
408 device
->has_exec_async
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_ASYNC
);
409 device
->has_exec_capture
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_CAPTURE
);
410 device
->has_exec_fence
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE
);
411 device
->has_syncobj
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE_ARRAY
);
412 device
->has_syncobj_wait
= device
->has_syncobj
&&
413 anv_gem_supports_syncobj_wait(fd
);
414 device
->has_context_priority
= anv_gem_has_context_priority(fd
);
416 device
->use_softpin
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_SOFTPIN
)
417 && device
->supports_48bit_addresses
;
419 device
->has_context_isolation
=
420 anv_gem_get_param(fd
, I915_PARAM_HAS_CONTEXT_ISOLATION
);
422 /* Starting with Gen10, the timestamp frequency of the command streamer may
423 * vary from one part to another. We can query the value from the kernel.
425 if (device
->info
.gen
>= 10) {
426 int timestamp_frequency
=
427 anv_gem_get_param(fd
, I915_PARAM_CS_TIMESTAMP_FREQUENCY
);
429 if (timestamp_frequency
< 0)
430 intel_logw("Kernel 4.16-rc1+ required to properly query CS timestamp frequency");
432 device
->info
.timestamp_frequency
= timestamp_frequency
;
435 /* GENs prior to 8 do not support EU/Subslice info */
436 if (device
->info
.gen
>= 8) {
437 device
->subslice_total
= anv_gem_get_param(fd
, I915_PARAM_SUBSLICE_TOTAL
);
438 device
->eu_total
= anv_gem_get_param(fd
, I915_PARAM_EU_TOTAL
);
440 /* Without this information, we cannot get the right Braswell
441 * brandstrings, and we have to use conservative numbers for GPGPU on
442 * many platforms, but otherwise, things will just work.
444 if (device
->subslice_total
< 1 || device
->eu_total
< 1) {
445 intel_logw("Kernel 4.1 required to properly query GPU properties");
447 } else if (device
->info
.gen
== 7) {
448 device
->subslice_total
= 1 << (device
->info
.gt
- 1);
451 if (device
->info
.is_cherryview
&&
452 device
->subslice_total
> 0 && device
->eu_total
> 0) {
453 /* Logical CS threads = EUs per subslice * num threads per EU */
454 uint32_t max_cs_threads
=
455 device
->eu_total
/ device
->subslice_total
* device
->info
.num_thread_per_eu
;
457 /* Fuse configurations may give more threads than expected, never less. */
458 if (max_cs_threads
> device
->info
.max_cs_threads
)
459 device
->info
.max_cs_threads
= max_cs_threads
;
462 device
->compiler
= brw_compiler_create(NULL
, &device
->info
);
463 if (device
->compiler
== NULL
) {
464 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
467 device
->compiler
->shader_debug_log
= compiler_debug_log
;
468 device
->compiler
->shader_perf_log
= compiler_perf_log
;
469 device
->compiler
->supports_pull_constants
= false;
470 device
->compiler
->constant_buffer_0_is_relative
=
471 device
->info
.gen
< 8 || !device
->has_context_isolation
;
472 device
->compiler
->supports_shader_constants
= true;
474 /* Broadwell PRM says:
476 * "Before Gen8, there was a historical configuration control field to
477 * swizzle address bit[6] for in X/Y tiling modes. This was set in three
478 * different places: TILECTL[1:0], ARB_MODE[5:4], and
479 * DISP_ARB_CTL[14:13].
481 * For Gen8 and subsequent generations, the swizzle fields are all
482 * reserved, and the CPU's memory controller performs all address
483 * swizzling modifications."
486 device
->info
.gen
< 8 && anv_gem_get_bit6_swizzle(fd
, I915_TILING_X
);
488 isl_device_init(&device
->isl_dev
, &device
->info
, swizzled
);
490 result
= anv_physical_device_init_uuids(device
);
491 if (result
!= VK_SUCCESS
)
494 anv_physical_device_init_disk_cache(device
);
496 if (instance
->enabled_extensions
.KHR_display
) {
497 master_fd
= open(primary_path
, O_RDWR
| O_CLOEXEC
);
498 if (master_fd
>= 0) {
499 /* prod the device with a GETPARAM call which will fail if
500 * we don't have permission to even render on this device
502 if (anv_gem_get_param(master_fd
, I915_PARAM_CHIPSET_ID
) == 0) {
508 device
->master_fd
= master_fd
;
510 result
= anv_init_wsi(device
);
511 if (result
!= VK_SUCCESS
) {
512 ralloc_free(device
->compiler
);
513 anv_physical_device_free_disk_cache(device
);
517 anv_physical_device_get_supported_extensions(device
,
518 &device
->supported_extensions
);
521 device
->local_fd
= fd
;
533 anv_physical_device_finish(struct anv_physical_device
*device
)
535 anv_finish_wsi(device
);
536 anv_physical_device_free_disk_cache(device
);
537 ralloc_free(device
->compiler
);
538 close(device
->local_fd
);
539 if (device
->master_fd
>= 0)
540 close(device
->master_fd
);
544 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
545 VkSystemAllocationScope allocationScope
)
551 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
552 size_t align
, VkSystemAllocationScope allocationScope
)
554 return realloc(pOriginal
, size
);
558 default_free_func(void *pUserData
, void *pMemory
)
563 static const VkAllocationCallbacks default_alloc
= {
565 .pfnAllocation
= default_alloc_func
,
566 .pfnReallocation
= default_realloc_func
,
567 .pfnFree
= default_free_func
,
570 VkResult
anv_EnumerateInstanceExtensionProperties(
571 const char* pLayerName
,
572 uint32_t* pPropertyCount
,
573 VkExtensionProperties
* pProperties
)
575 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
577 for (int i
= 0; i
< ANV_INSTANCE_EXTENSION_COUNT
; i
++) {
578 if (anv_instance_extensions_supported
.extensions
[i
]) {
579 vk_outarray_append(&out
, prop
) {
580 *prop
= anv_instance_extensions
[i
];
585 return vk_outarray_status(&out
);
588 VkResult
anv_CreateInstance(
589 const VkInstanceCreateInfo
* pCreateInfo
,
590 const VkAllocationCallbacks
* pAllocator
,
591 VkInstance
* pInstance
)
593 struct anv_instance
*instance
;
596 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
598 struct anv_instance_extension_table enabled_extensions
= {};
599 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
601 for (idx
= 0; idx
< ANV_INSTANCE_EXTENSION_COUNT
; idx
++) {
602 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
603 anv_instance_extensions
[idx
].extensionName
) == 0)
607 if (idx
>= ANV_INSTANCE_EXTENSION_COUNT
)
608 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
610 if (!anv_instance_extensions_supported
.extensions
[idx
])
611 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
613 enabled_extensions
.extensions
[idx
] = true;
616 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
617 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
619 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
621 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
624 instance
->alloc
= *pAllocator
;
626 instance
->alloc
= default_alloc
;
628 instance
->app_info
= (struct anv_app_info
) { .api_version
= 0 };
629 if (pCreateInfo
->pApplicationInfo
) {
630 const VkApplicationInfo
*app
= pCreateInfo
->pApplicationInfo
;
632 instance
->app_info
.app_name
=
633 vk_strdup(&instance
->alloc
, app
->pApplicationName
,
634 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
635 instance
->app_info
.app_version
= app
->applicationVersion
;
637 instance
->app_info
.engine_name
=
638 vk_strdup(&instance
->alloc
, app
->pEngineName
,
639 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
640 instance
->app_info
.engine_version
= app
->engineVersion
;
642 instance
->app_info
.api_version
= app
->apiVersion
;
645 if (instance
->app_info
.api_version
== 0)
646 instance
->app_info
.api_version
= VK_API_VERSION_1_0
;
648 instance
->enabled_extensions
= enabled_extensions
;
650 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->dispatch
.entrypoints
); i
++) {
651 /* Vulkan requires that entrypoints for extensions which have not been
652 * enabled must not be advertised.
654 if (!anv_instance_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
655 &instance
->enabled_extensions
)) {
656 instance
->dispatch
.entrypoints
[i
] = NULL
;
658 instance
->dispatch
.entrypoints
[i
] =
659 anv_instance_dispatch_table
.entrypoints
[i
];
663 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->device_dispatch
.entrypoints
); i
++) {
664 /* Vulkan requires that entrypoints for extensions which have not been
665 * enabled must not be advertised.
667 if (!anv_device_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
668 &instance
->enabled_extensions
, NULL
)) {
669 instance
->device_dispatch
.entrypoints
[i
] = NULL
;
671 instance
->device_dispatch
.entrypoints
[i
] =
672 anv_device_dispatch_table
.entrypoints
[i
];
676 instance
->physicalDeviceCount
= -1;
678 result
= vk_debug_report_instance_init(&instance
->debug_report_callbacks
);
679 if (result
!= VK_SUCCESS
) {
680 vk_free2(&default_alloc
, pAllocator
, instance
);
681 return vk_error(result
);
684 instance
->pipeline_cache_enabled
=
685 env_var_as_boolean("ANV_ENABLE_PIPELINE_CACHE", true);
689 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
691 *pInstance
= anv_instance_to_handle(instance
);
696 void anv_DestroyInstance(
697 VkInstance _instance
,
698 const VkAllocationCallbacks
* pAllocator
)
700 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
705 if (instance
->physicalDeviceCount
> 0) {
706 /* We support at most one physical device. */
707 assert(instance
->physicalDeviceCount
== 1);
708 anv_physical_device_finish(&instance
->physicalDevice
);
711 vk_free(&instance
->alloc
, (char *)instance
->app_info
.app_name
);
712 vk_free(&instance
->alloc
, (char *)instance
->app_info
.engine_name
);
714 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
716 vk_debug_report_instance_destroy(&instance
->debug_report_callbacks
);
720 vk_free(&instance
->alloc
, instance
);
724 anv_enumerate_devices(struct anv_instance
*instance
)
726 /* TODO: Check for more devices ? */
727 drmDevicePtr devices
[8];
728 VkResult result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
731 instance
->physicalDeviceCount
= 0;
733 max_devices
= drmGetDevices2(0, devices
, ARRAY_SIZE(devices
));
735 return VK_ERROR_INCOMPATIBLE_DRIVER
;
737 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
738 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
739 devices
[i
]->bustype
== DRM_BUS_PCI
&&
740 devices
[i
]->deviceinfo
.pci
->vendor_id
== 0x8086) {
742 result
= anv_physical_device_init(&instance
->physicalDevice
,
743 instance
, devices
[i
]);
744 if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
748 drmFreeDevices(devices
, max_devices
);
750 if (result
== VK_SUCCESS
)
751 instance
->physicalDeviceCount
= 1;
757 anv_instance_ensure_physical_device(struct anv_instance
*instance
)
759 if (instance
->physicalDeviceCount
< 0) {
760 VkResult result
= anv_enumerate_devices(instance
);
761 if (result
!= VK_SUCCESS
&&
762 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
769 VkResult
anv_EnumeratePhysicalDevices(
770 VkInstance _instance
,
771 uint32_t* pPhysicalDeviceCount
,
772 VkPhysicalDevice
* pPhysicalDevices
)
774 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
775 VK_OUTARRAY_MAKE(out
, pPhysicalDevices
, pPhysicalDeviceCount
);
777 VkResult result
= anv_instance_ensure_physical_device(instance
);
778 if (result
!= VK_SUCCESS
)
781 if (instance
->physicalDeviceCount
== 0)
784 assert(instance
->physicalDeviceCount
== 1);
785 vk_outarray_append(&out
, i
) {
786 *i
= anv_physical_device_to_handle(&instance
->physicalDevice
);
789 return vk_outarray_status(&out
);
792 VkResult
anv_EnumeratePhysicalDeviceGroups(
793 VkInstance _instance
,
794 uint32_t* pPhysicalDeviceGroupCount
,
795 VkPhysicalDeviceGroupProperties
* pPhysicalDeviceGroupProperties
)
797 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
798 VK_OUTARRAY_MAKE(out
, pPhysicalDeviceGroupProperties
,
799 pPhysicalDeviceGroupCount
);
801 VkResult result
= anv_instance_ensure_physical_device(instance
);
802 if (result
!= VK_SUCCESS
)
805 if (instance
->physicalDeviceCount
== 0)
808 assert(instance
->physicalDeviceCount
== 1);
810 vk_outarray_append(&out
, p
) {
811 p
->physicalDeviceCount
= 1;
812 memset(p
->physicalDevices
, 0, sizeof(p
->physicalDevices
));
813 p
->physicalDevices
[0] =
814 anv_physical_device_to_handle(&instance
->physicalDevice
);
815 p
->subsetAllocation
= VK_FALSE
;
817 vk_foreach_struct(ext
, p
->pNext
)
818 anv_debug_ignored_stype(ext
->sType
);
821 return vk_outarray_status(&out
);
824 void anv_GetPhysicalDeviceFeatures(
825 VkPhysicalDevice physicalDevice
,
826 VkPhysicalDeviceFeatures
* pFeatures
)
828 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
830 *pFeatures
= (VkPhysicalDeviceFeatures
) {
831 .robustBufferAccess
= true,
832 .fullDrawIndexUint32
= true,
833 .imageCubeArray
= true,
834 .independentBlend
= true,
835 .geometryShader
= true,
836 .tessellationShader
= true,
837 .sampleRateShading
= true,
838 .dualSrcBlend
= true,
840 .multiDrawIndirect
= true,
841 .drawIndirectFirstInstance
= true,
843 .depthBiasClamp
= true,
844 .fillModeNonSolid
= true,
845 .depthBounds
= false,
849 .multiViewport
= true,
850 .samplerAnisotropy
= true,
851 .textureCompressionETC2
= pdevice
->info
.gen
>= 8 ||
852 pdevice
->info
.is_baytrail
,
853 .textureCompressionASTC_LDR
= pdevice
->info
.gen
>= 9, /* FINISHME CHV */
854 .textureCompressionBC
= true,
855 .occlusionQueryPrecise
= true,
856 .pipelineStatisticsQuery
= true,
857 .fragmentStoresAndAtomics
= true,
858 .shaderTessellationAndGeometryPointSize
= true,
859 .shaderImageGatherExtended
= true,
860 .shaderStorageImageExtendedFormats
= true,
861 .shaderStorageImageMultisample
= false,
862 .shaderStorageImageReadWithoutFormat
= false,
863 .shaderStorageImageWriteWithoutFormat
= true,
864 .shaderUniformBufferArrayDynamicIndexing
= true,
865 .shaderSampledImageArrayDynamicIndexing
= true,
866 .shaderStorageBufferArrayDynamicIndexing
= true,
867 .shaderStorageImageArrayDynamicIndexing
= true,
868 .shaderClipDistance
= true,
869 .shaderCullDistance
= true,
870 .shaderFloat64
= pdevice
->info
.gen
>= 8 &&
871 pdevice
->info
.has_64bit_types
,
872 .shaderInt64
= pdevice
->info
.gen
>= 8 &&
873 pdevice
->info
.has_64bit_types
,
874 .shaderInt16
= pdevice
->info
.gen
>= 8,
875 .shaderResourceMinLod
= pdevice
->info
.gen
>= 9,
876 .variableMultisampleRate
= true,
877 .inheritedQueries
= true,
880 /* We can't do image stores in vec4 shaders */
881 pFeatures
->vertexPipelineStoresAndAtomics
=
882 pdevice
->compiler
->scalar_stage
[MESA_SHADER_VERTEX
] &&
883 pdevice
->compiler
->scalar_stage
[MESA_SHADER_GEOMETRY
];
885 struct anv_app_info
*app_info
= &pdevice
->instance
->app_info
;
887 /* The new DOOM and Wolfenstein games require depthBounds without
888 * checking for it. They seem to run fine without it so just claim it's
889 * there and accept the consequences.
891 if (app_info
->engine_name
&& strcmp(app_info
->engine_name
, "idTech") == 0)
892 pFeatures
->depthBounds
= true;
895 void anv_GetPhysicalDeviceFeatures2(
896 VkPhysicalDevice physicalDevice
,
897 VkPhysicalDeviceFeatures2
* pFeatures
)
899 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
900 anv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
902 vk_foreach_struct(ext
, pFeatures
->pNext
) {
903 switch (ext
->sType
) {
904 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES_KHR
: {
905 VkPhysicalDevice8BitStorageFeaturesKHR
*features
=
906 (VkPhysicalDevice8BitStorageFeaturesKHR
*)ext
;
907 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
909 features
->storageBuffer8BitAccess
= pdevice
->info
.gen
>= 8;
910 features
->uniformAndStorageBuffer8BitAccess
= pdevice
->info
.gen
>= 8;
911 features
->storagePushConstant8
= pdevice
->info
.gen
>= 8;
915 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES
: {
916 VkPhysicalDevice16BitStorageFeatures
*features
=
917 (VkPhysicalDevice16BitStorageFeatures
*)ext
;
918 features
->storageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
919 features
->uniformAndStorageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
920 features
->storagePushConstant16
= pdevice
->info
.gen
>= 8;
921 features
->storageInputOutput16
= false;
925 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_ADDRESS_FEATURES_EXT
: {
926 VkPhysicalDeviceBufferAddressFeaturesEXT
*features
= (void *)ext
;
927 features
->bufferDeviceAddress
= pdevice
->use_softpin
&&
928 pdevice
->info
.gen
>= 8;
929 features
->bufferDeviceAddressCaptureReplay
= false;
930 features
->bufferDeviceAddressMultiDevice
= false;
934 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT
: {
935 VkPhysicalDeviceConditionalRenderingFeaturesEXT
*features
=
936 (VkPhysicalDeviceConditionalRenderingFeaturesEXT
*)ext
;
937 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
939 features
->conditionalRendering
= pdevice
->info
.gen
>= 8 ||
940 pdevice
->info
.is_haswell
;
941 features
->inheritedConditionalRendering
= pdevice
->info
.gen
>= 8 ||
942 pdevice
->info
.is_haswell
;
946 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_CLIP_ENABLE_FEATURES_EXT
: {
947 VkPhysicalDeviceDepthClipEnableFeaturesEXT
*features
=
948 (VkPhysicalDeviceDepthClipEnableFeaturesEXT
*)ext
;
949 features
->depthClipEnable
= VK_TRUE
;
953 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES
: {
954 VkPhysicalDeviceMultiviewFeatures
*features
=
955 (VkPhysicalDeviceMultiviewFeatures
*)ext
;
956 features
->multiview
= true;
957 features
->multiviewGeometryShader
= true;
958 features
->multiviewTessellationShader
= true;
962 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES
: {
963 VkPhysicalDeviceProtectedMemoryFeatures
*features
= (void *)ext
;
964 features
->protectedMemory
= VK_FALSE
;
968 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES
: {
969 VkPhysicalDeviceSamplerYcbcrConversionFeatures
*features
=
970 (VkPhysicalDeviceSamplerYcbcrConversionFeatures
*) ext
;
971 features
->samplerYcbcrConversion
= true;
975 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES_EXT
: {
976 VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*features
=
977 (VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*)ext
;
978 features
->scalarBlockLayout
= true;
982 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETER_FEATURES
: {
983 VkPhysicalDeviceShaderDrawParameterFeatures
*features
= (void *)ext
;
984 features
->shaderDrawParameters
= true;
988 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTER_FEATURES
: {
989 VkPhysicalDeviceVariablePointerFeatures
*features
= (void *)ext
;
990 features
->variablePointersStorageBuffer
= true;
991 features
->variablePointers
= true;
995 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT
: {
996 VkPhysicalDeviceTransformFeedbackFeaturesEXT
*features
=
997 (VkPhysicalDeviceTransformFeedbackFeaturesEXT
*)ext
;
998 features
->transformFeedback
= VK_TRUE
;
999 features
->geometryStreams
= VK_TRUE
;
1003 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT
: {
1004 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*features
=
1005 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*)ext
;
1006 features
->vertexAttributeInstanceRateDivisor
= VK_TRUE
;
1007 features
->vertexAttributeInstanceRateZeroDivisor
= VK_TRUE
;
1012 anv_debug_ignored_stype(ext
->sType
);
1018 void anv_GetPhysicalDeviceProperties(
1019 VkPhysicalDevice physicalDevice
,
1020 VkPhysicalDeviceProperties
* pProperties
)
1022 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1023 const struct gen_device_info
*devinfo
= &pdevice
->info
;
1025 /* See assertions made when programming the buffer surface state. */
1026 const uint32_t max_raw_buffer_sz
= devinfo
->gen
>= 7 ?
1027 (1ul << 30) : (1ul << 27);
1029 const uint32_t max_samplers
= (devinfo
->gen
>= 8 || devinfo
->is_haswell
) ?
1032 const uint32_t max_images
= devinfo
->gen
< 9 ? MAX_GEN8_IMAGES
: MAX_IMAGES
;
1034 VkSampleCountFlags sample_counts
=
1035 isl_device_get_sample_counts(&pdevice
->isl_dev
);
1038 VkPhysicalDeviceLimits limits
= {
1039 .maxImageDimension1D
= (1 << 14),
1040 .maxImageDimension2D
= (1 << 14),
1041 .maxImageDimension3D
= (1 << 11),
1042 .maxImageDimensionCube
= (1 << 14),
1043 .maxImageArrayLayers
= (1 << 11),
1044 .maxTexelBufferElements
= 128 * 1024 * 1024,
1045 .maxUniformBufferRange
= (1ul << 27),
1046 .maxStorageBufferRange
= max_raw_buffer_sz
,
1047 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
1048 .maxMemoryAllocationCount
= UINT32_MAX
,
1049 .maxSamplerAllocationCount
= 64 * 1024,
1050 .bufferImageGranularity
= 64, /* A cache line */
1051 .sparseAddressSpaceSize
= 0,
1052 .maxBoundDescriptorSets
= MAX_SETS
,
1053 .maxPerStageDescriptorSamplers
= max_samplers
,
1054 .maxPerStageDescriptorUniformBuffers
= 64,
1055 .maxPerStageDescriptorStorageBuffers
= 64,
1056 .maxPerStageDescriptorSampledImages
= max_samplers
,
1057 .maxPerStageDescriptorStorageImages
= max_images
,
1058 .maxPerStageDescriptorInputAttachments
= 64,
1059 .maxPerStageResources
= 250,
1060 .maxDescriptorSetSamplers
= 6 * max_samplers
, /* number of stages * maxPerStageDescriptorSamplers */
1061 .maxDescriptorSetUniformBuffers
= 6 * 64, /* number of stages * maxPerStageDescriptorUniformBuffers */
1062 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
1063 .maxDescriptorSetStorageBuffers
= 6 * 64, /* number of stages * maxPerStageDescriptorStorageBuffers */
1064 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
1065 .maxDescriptorSetSampledImages
= 6 * max_samplers
, /* number of stages * maxPerStageDescriptorSampledImages */
1066 .maxDescriptorSetStorageImages
= 6 * max_images
, /* number of stages * maxPerStageDescriptorStorageImages */
1067 .maxDescriptorSetInputAttachments
= 256,
1068 .maxVertexInputAttributes
= MAX_VBS
,
1069 .maxVertexInputBindings
= MAX_VBS
,
1070 .maxVertexInputAttributeOffset
= 2047,
1071 .maxVertexInputBindingStride
= 2048,
1072 .maxVertexOutputComponents
= 128,
1073 .maxTessellationGenerationLevel
= 64,
1074 .maxTessellationPatchSize
= 32,
1075 .maxTessellationControlPerVertexInputComponents
= 128,
1076 .maxTessellationControlPerVertexOutputComponents
= 128,
1077 .maxTessellationControlPerPatchOutputComponents
= 128,
1078 .maxTessellationControlTotalOutputComponents
= 2048,
1079 .maxTessellationEvaluationInputComponents
= 128,
1080 .maxTessellationEvaluationOutputComponents
= 128,
1081 .maxGeometryShaderInvocations
= 32,
1082 .maxGeometryInputComponents
= 64,
1083 .maxGeometryOutputComponents
= 128,
1084 .maxGeometryOutputVertices
= 256,
1085 .maxGeometryTotalOutputComponents
= 1024,
1086 .maxFragmentInputComponents
= 112, /* 128 components - (POS, PSIZ, CLIP_DIST0, CLIP_DIST1) */
1087 .maxFragmentOutputAttachments
= 8,
1088 .maxFragmentDualSrcAttachments
= 1,
1089 .maxFragmentCombinedOutputResources
= 8,
1090 .maxComputeSharedMemorySize
= 32768,
1091 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
1092 .maxComputeWorkGroupInvocations
= 16 * devinfo
->max_cs_threads
,
1093 .maxComputeWorkGroupSize
= {
1094 16 * devinfo
->max_cs_threads
,
1095 16 * devinfo
->max_cs_threads
,
1096 16 * devinfo
->max_cs_threads
,
1098 .subPixelPrecisionBits
= 8,
1099 .subTexelPrecisionBits
= 4 /* FIXME */,
1100 .mipmapPrecisionBits
= 4 /* FIXME */,
1101 .maxDrawIndexedIndexValue
= UINT32_MAX
,
1102 .maxDrawIndirectCount
= UINT32_MAX
,
1103 .maxSamplerLodBias
= 16,
1104 .maxSamplerAnisotropy
= 16,
1105 .maxViewports
= MAX_VIEWPORTS
,
1106 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
1107 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
1108 .viewportSubPixelBits
= 13, /* We take a float? */
1109 .minMemoryMapAlignment
= 4096, /* A page */
1110 .minTexelBufferOffsetAlignment
= 1,
1111 /* We need 16 for UBO block reads to work and 32 for push UBOs */
1112 .minUniformBufferOffsetAlignment
= 32,
1113 .minStorageBufferOffsetAlignment
= 4,
1114 .minTexelOffset
= -8,
1115 .maxTexelOffset
= 7,
1116 .minTexelGatherOffset
= -32,
1117 .maxTexelGatherOffset
= 31,
1118 .minInterpolationOffset
= -0.5,
1119 .maxInterpolationOffset
= 0.4375,
1120 .subPixelInterpolationOffsetBits
= 4,
1121 .maxFramebufferWidth
= (1 << 14),
1122 .maxFramebufferHeight
= (1 << 14),
1123 .maxFramebufferLayers
= (1 << 11),
1124 .framebufferColorSampleCounts
= sample_counts
,
1125 .framebufferDepthSampleCounts
= sample_counts
,
1126 .framebufferStencilSampleCounts
= sample_counts
,
1127 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
1128 .maxColorAttachments
= MAX_RTS
,
1129 .sampledImageColorSampleCounts
= sample_counts
,
1130 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
1131 .sampledImageDepthSampleCounts
= sample_counts
,
1132 .sampledImageStencilSampleCounts
= sample_counts
,
1133 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
1134 .maxSampleMaskWords
= 1,
1135 .timestampComputeAndGraphics
= false,
1136 .timestampPeriod
= 1000000000.0 / devinfo
->timestamp_frequency
,
1137 .maxClipDistances
= 8,
1138 .maxCullDistances
= 8,
1139 .maxCombinedClipAndCullDistances
= 8,
1140 .discreteQueuePriorities
= 2,
1141 .pointSizeRange
= { 0.125, 255.875 },
1142 .lineWidthRange
= { 0.0, 7.9921875 },
1143 .pointSizeGranularity
= (1.0 / 8.0),
1144 .lineWidthGranularity
= (1.0 / 128.0),
1145 .strictLines
= false, /* FINISHME */
1146 .standardSampleLocations
= true,
1147 .optimalBufferCopyOffsetAlignment
= 128,
1148 .optimalBufferCopyRowPitchAlignment
= 128,
1149 .nonCoherentAtomSize
= 64,
1152 *pProperties
= (VkPhysicalDeviceProperties
) {
1153 .apiVersion
= anv_physical_device_api_version(pdevice
),
1154 .driverVersion
= vk_get_driver_version(),
1156 .deviceID
= pdevice
->chipset_id
,
1157 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
1159 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
1162 snprintf(pProperties
->deviceName
, sizeof(pProperties
->deviceName
),
1163 "%s", pdevice
->name
);
1164 memcpy(pProperties
->pipelineCacheUUID
,
1165 pdevice
->pipeline_cache_uuid
, VK_UUID_SIZE
);
1168 void anv_GetPhysicalDeviceProperties2(
1169 VkPhysicalDevice physicalDevice
,
1170 VkPhysicalDeviceProperties2
* pProperties
)
1172 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1174 anv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
1176 vk_foreach_struct(ext
, pProperties
->pNext
) {
1177 switch (ext
->sType
) {
1178 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_STENCIL_RESOLVE_PROPERTIES_KHR
: {
1179 VkPhysicalDeviceDepthStencilResolvePropertiesKHR
*props
=
1180 (VkPhysicalDeviceDepthStencilResolvePropertiesKHR
*)ext
;
1182 /* We support all of the depth resolve modes */
1183 props
->supportedDepthResolveModes
=
1184 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
|
1185 VK_RESOLVE_MODE_AVERAGE_BIT_KHR
|
1186 VK_RESOLVE_MODE_MIN_BIT_KHR
|
1187 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1189 /* Average doesn't make sense for stencil so we don't support that */
1190 props
->supportedStencilResolveModes
=
1191 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
;
1192 if (pdevice
->info
.gen
>= 8) {
1193 /* The advanced stencil resolve modes currently require stencil
1194 * sampling be supported by the hardware.
1196 props
->supportedStencilResolveModes
|=
1197 VK_RESOLVE_MODE_MIN_BIT_KHR
|
1198 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1201 props
->independentResolveNone
= VK_TRUE
;
1202 props
->independentResolve
= VK_TRUE
;
1206 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR
: {
1207 VkPhysicalDeviceDriverPropertiesKHR
*driver_props
=
1208 (VkPhysicalDeviceDriverPropertiesKHR
*) ext
;
1210 driver_props
->driverID
= VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA_KHR
;
1211 util_snprintf(driver_props
->driverName
, VK_MAX_DRIVER_NAME_SIZE_KHR
,
1212 "Intel open-source Mesa driver");
1214 util_snprintf(driver_props
->driverInfo
, VK_MAX_DRIVER_INFO_SIZE_KHR
,
1215 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1
);
1217 driver_props
->conformanceVersion
= (VkConformanceVersionKHR
) {
1226 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES
: {
1227 VkPhysicalDeviceIDProperties
*id_props
=
1228 (VkPhysicalDeviceIDProperties
*)ext
;
1229 memcpy(id_props
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
1230 memcpy(id_props
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
1231 /* The LUID is for Windows. */
1232 id_props
->deviceLUIDValid
= false;
1236 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES
: {
1237 VkPhysicalDeviceMaintenance3Properties
*props
=
1238 (VkPhysicalDeviceMaintenance3Properties
*)ext
;
1239 /* This value doesn't matter for us today as our per-stage
1240 * descriptors are the real limit.
1242 props
->maxPerSetDescriptors
= 1024;
1243 props
->maxMemoryAllocationSize
= MAX_MEMORY_ALLOCATION_SIZE
;
1247 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES
: {
1248 VkPhysicalDeviceMultiviewProperties
*properties
=
1249 (VkPhysicalDeviceMultiviewProperties
*)ext
;
1250 properties
->maxMultiviewViewCount
= 16;
1251 properties
->maxMultiviewInstanceIndex
= UINT32_MAX
/ 16;
1255 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT
: {
1256 VkPhysicalDevicePCIBusInfoPropertiesEXT
*properties
=
1257 (VkPhysicalDevicePCIBusInfoPropertiesEXT
*)ext
;
1258 properties
->pciDomain
= pdevice
->pci_info
.domain
;
1259 properties
->pciBus
= pdevice
->pci_info
.bus
;
1260 properties
->pciDevice
= pdevice
->pci_info
.device
;
1261 properties
->pciFunction
= pdevice
->pci_info
.function
;
1265 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES
: {
1266 VkPhysicalDevicePointClippingProperties
*properties
=
1267 (VkPhysicalDevicePointClippingProperties
*) ext
;
1268 properties
->pointClippingBehavior
= VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES
;
1269 anv_finishme("Implement pop-free point clipping");
1273 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES
: {
1274 VkPhysicalDeviceProtectedMemoryProperties
*props
=
1275 (VkPhysicalDeviceProtectedMemoryProperties
*)ext
;
1276 props
->protectedNoFault
= false;
1280 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
1281 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
1282 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
1284 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
1288 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT
: {
1289 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*properties
=
1290 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*)ext
;
1291 properties
->filterMinmaxImageComponentMapping
= pdevice
->info
.gen
>= 9;
1292 properties
->filterMinmaxSingleComponentFormats
= true;
1296 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES
: {
1297 VkPhysicalDeviceSubgroupProperties
*properties
= (void *)ext
;
1299 properties
->subgroupSize
= BRW_SUBGROUP_SIZE
;
1301 VkShaderStageFlags scalar_stages
= 0;
1302 for (unsigned stage
= 0; stage
< MESA_SHADER_STAGES
; stage
++) {
1303 if (pdevice
->compiler
->scalar_stage
[stage
])
1304 scalar_stages
|= mesa_to_vk_shader_stage(stage
);
1306 properties
->supportedStages
= scalar_stages
;
1308 properties
->supportedOperations
= VK_SUBGROUP_FEATURE_BASIC_BIT
|
1309 VK_SUBGROUP_FEATURE_VOTE_BIT
|
1310 VK_SUBGROUP_FEATURE_ARITHMETIC_BIT
|
1311 VK_SUBGROUP_FEATURE_BALLOT_BIT
|
1312 VK_SUBGROUP_FEATURE_SHUFFLE_BIT
|
1313 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT
|
1314 VK_SUBGROUP_FEATURE_CLUSTERED_BIT
|
1315 VK_SUBGROUP_FEATURE_QUAD_BIT
;
1316 properties
->quadOperationsInAllStages
= VK_TRUE
;
1320 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT
: {
1321 VkPhysicalDeviceTransformFeedbackPropertiesEXT
*props
=
1322 (VkPhysicalDeviceTransformFeedbackPropertiesEXT
*)ext
;
1324 props
->maxTransformFeedbackStreams
= MAX_XFB_STREAMS
;
1325 props
->maxTransformFeedbackBuffers
= MAX_XFB_BUFFERS
;
1326 props
->maxTransformFeedbackBufferSize
= (1ull << 32);
1327 props
->maxTransformFeedbackStreamDataSize
= 128 * 4;
1328 props
->maxTransformFeedbackBufferDataSize
= 128 * 4;
1329 props
->maxTransformFeedbackBufferDataStride
= 2048;
1330 props
->transformFeedbackQueries
= VK_TRUE
;
1331 props
->transformFeedbackStreamsLinesTriangles
= VK_FALSE
;
1332 props
->transformFeedbackRasterizationStreamSelect
= VK_FALSE
;
1333 props
->transformFeedbackDraw
= VK_TRUE
;
1337 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT
: {
1338 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*props
=
1339 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*)ext
;
1340 /* We have to restrict this a bit for multiview */
1341 props
->maxVertexAttribDivisor
= UINT32_MAX
/ 16;
1346 anv_debug_ignored_stype(ext
->sType
);
1352 /* We support exactly one queue family. */
1353 static const VkQueueFamilyProperties
1354 anv_queue_family_properties
= {
1355 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
1356 VK_QUEUE_COMPUTE_BIT
|
1357 VK_QUEUE_TRANSFER_BIT
,
1359 .timestampValidBits
= 36, /* XXX: Real value here */
1360 .minImageTransferGranularity
= { 1, 1, 1 },
1363 void anv_GetPhysicalDeviceQueueFamilyProperties(
1364 VkPhysicalDevice physicalDevice
,
1366 VkQueueFamilyProperties
* pQueueFamilyProperties
)
1368 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pCount
);
1370 vk_outarray_append(&out
, p
) {
1371 *p
= anv_queue_family_properties
;
1375 void anv_GetPhysicalDeviceQueueFamilyProperties2(
1376 VkPhysicalDevice physicalDevice
,
1377 uint32_t* pQueueFamilyPropertyCount
,
1378 VkQueueFamilyProperties2
* pQueueFamilyProperties
)
1381 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pQueueFamilyPropertyCount
);
1383 vk_outarray_append(&out
, p
) {
1384 p
->queueFamilyProperties
= anv_queue_family_properties
;
1386 vk_foreach_struct(s
, p
->pNext
) {
1387 anv_debug_ignored_stype(s
->sType
);
1392 void anv_GetPhysicalDeviceMemoryProperties(
1393 VkPhysicalDevice physicalDevice
,
1394 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
1396 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
1398 pMemoryProperties
->memoryTypeCount
= physical_device
->memory
.type_count
;
1399 for (uint32_t i
= 0; i
< physical_device
->memory
.type_count
; i
++) {
1400 pMemoryProperties
->memoryTypes
[i
] = (VkMemoryType
) {
1401 .propertyFlags
= physical_device
->memory
.types
[i
].propertyFlags
,
1402 .heapIndex
= physical_device
->memory
.types
[i
].heapIndex
,
1406 pMemoryProperties
->memoryHeapCount
= physical_device
->memory
.heap_count
;
1407 for (uint32_t i
= 0; i
< physical_device
->memory
.heap_count
; i
++) {
1408 pMemoryProperties
->memoryHeaps
[i
] = (VkMemoryHeap
) {
1409 .size
= physical_device
->memory
.heaps
[i
].size
,
1410 .flags
= physical_device
->memory
.heaps
[i
].flags
,
1415 void anv_GetPhysicalDeviceMemoryProperties2(
1416 VkPhysicalDevice physicalDevice
,
1417 VkPhysicalDeviceMemoryProperties2
* pMemoryProperties
)
1419 anv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
1420 &pMemoryProperties
->memoryProperties
);
1422 vk_foreach_struct(ext
, pMemoryProperties
->pNext
) {
1423 switch (ext
->sType
) {
1425 anv_debug_ignored_stype(ext
->sType
);
1432 anv_GetDeviceGroupPeerMemoryFeatures(
1435 uint32_t localDeviceIndex
,
1436 uint32_t remoteDeviceIndex
,
1437 VkPeerMemoryFeatureFlags
* pPeerMemoryFeatures
)
1439 assert(localDeviceIndex
== 0 && remoteDeviceIndex
== 0);
1440 *pPeerMemoryFeatures
= VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT
|
1441 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT
|
1442 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT
|
1443 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT
;
1446 PFN_vkVoidFunction
anv_GetInstanceProcAddr(
1447 VkInstance _instance
,
1450 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1452 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
1453 * when we have to return valid function pointers, NULL, or it's left
1454 * undefined. See the table for exact details.
1459 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
1460 if (strcmp(pName, "vk" #entrypoint) == 0) \
1461 return (PFN_vkVoidFunction)anv_##entrypoint
1463 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties
);
1464 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties
);
1465 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceVersion
);
1466 LOOKUP_ANV_ENTRYPOINT(CreateInstance
);
1468 #undef LOOKUP_ANV_ENTRYPOINT
1470 if (instance
== NULL
)
1473 int idx
= anv_get_instance_entrypoint_index(pName
);
1475 return instance
->dispatch
.entrypoints
[idx
];
1477 idx
= anv_get_device_entrypoint_index(pName
);
1479 return instance
->device_dispatch
.entrypoints
[idx
];
1484 /* With version 1+ of the loader interface the ICD should expose
1485 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
1488 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1489 VkInstance instance
,
1493 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1494 VkInstance instance
,
1497 return anv_GetInstanceProcAddr(instance
, pName
);
1500 PFN_vkVoidFunction
anv_GetDeviceProcAddr(
1504 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1506 if (!device
|| !pName
)
1509 int idx
= anv_get_device_entrypoint_index(pName
);
1513 return device
->dispatch
.entrypoints
[idx
];
1517 anv_CreateDebugReportCallbackEXT(VkInstance _instance
,
1518 const VkDebugReportCallbackCreateInfoEXT
* pCreateInfo
,
1519 const VkAllocationCallbacks
* pAllocator
,
1520 VkDebugReportCallbackEXT
* pCallback
)
1522 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1523 return vk_create_debug_report_callback(&instance
->debug_report_callbacks
,
1524 pCreateInfo
, pAllocator
, &instance
->alloc
,
1529 anv_DestroyDebugReportCallbackEXT(VkInstance _instance
,
1530 VkDebugReportCallbackEXT _callback
,
1531 const VkAllocationCallbacks
* pAllocator
)
1533 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1534 vk_destroy_debug_report_callback(&instance
->debug_report_callbacks
,
1535 _callback
, pAllocator
, &instance
->alloc
);
1539 anv_DebugReportMessageEXT(VkInstance _instance
,
1540 VkDebugReportFlagsEXT flags
,
1541 VkDebugReportObjectTypeEXT objectType
,
1544 int32_t messageCode
,
1545 const char* pLayerPrefix
,
1546 const char* pMessage
)
1548 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1549 vk_debug_report(&instance
->debug_report_callbacks
, flags
, objectType
,
1550 object
, location
, messageCode
, pLayerPrefix
, pMessage
);
1554 anv_queue_init(struct anv_device
*device
, struct anv_queue
*queue
)
1556 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1557 queue
->device
= device
;
1562 anv_queue_finish(struct anv_queue
*queue
)
1566 static struct anv_state
1567 anv_state_pool_emit_data(struct anv_state_pool
*pool
, size_t size
, size_t align
, const void *p
)
1569 struct anv_state state
;
1571 state
= anv_state_pool_alloc(pool
, size
, align
);
1572 memcpy(state
.map
, p
, size
);
1577 struct gen8_border_color
{
1582 /* Pad out to 64 bytes */
1587 anv_device_init_border_colors(struct anv_device
*device
)
1589 static const struct gen8_border_color border_colors
[] = {
1590 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
1591 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
1592 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
1593 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
1594 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
1595 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
1598 device
->border_colors
= anv_state_pool_emit_data(&device
->dynamic_state_pool
,
1599 sizeof(border_colors
), 64,
1604 anv_device_init_trivial_batch(struct anv_device
*device
)
1606 anv_bo_init_new(&device
->trivial_batch_bo
, device
, 4096);
1608 if (device
->instance
->physicalDevice
.has_exec_async
)
1609 device
->trivial_batch_bo
.flags
|= EXEC_OBJECT_ASYNC
;
1611 if (device
->instance
->physicalDevice
.use_softpin
)
1612 device
->trivial_batch_bo
.flags
|= EXEC_OBJECT_PINNED
;
1614 anv_vma_alloc(device
, &device
->trivial_batch_bo
);
1616 void *map
= anv_gem_mmap(device
, device
->trivial_batch_bo
.gem_handle
,
1619 struct anv_batch batch
= {
1625 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
1626 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
1628 if (!device
->info
.has_llc
)
1629 gen_clflush_range(map
, batch
.next
- map
);
1631 anv_gem_munmap(map
, device
->trivial_batch_bo
.size
);
1634 VkResult
anv_EnumerateDeviceExtensionProperties(
1635 VkPhysicalDevice physicalDevice
,
1636 const char* pLayerName
,
1637 uint32_t* pPropertyCount
,
1638 VkExtensionProperties
* pProperties
)
1640 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
1641 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
1643 for (int i
= 0; i
< ANV_DEVICE_EXTENSION_COUNT
; i
++) {
1644 if (device
->supported_extensions
.extensions
[i
]) {
1645 vk_outarray_append(&out
, prop
) {
1646 *prop
= anv_device_extensions
[i
];
1651 return vk_outarray_status(&out
);
1655 anv_device_init_dispatch(struct anv_device
*device
)
1657 const struct anv_device_dispatch_table
*genX_table
;
1658 switch (device
->info
.gen
) {
1660 genX_table
= &gen11_device_dispatch_table
;
1663 genX_table
= &gen10_device_dispatch_table
;
1666 genX_table
= &gen9_device_dispatch_table
;
1669 genX_table
= &gen8_device_dispatch_table
;
1672 if (device
->info
.is_haswell
)
1673 genX_table
= &gen75_device_dispatch_table
;
1675 genX_table
= &gen7_device_dispatch_table
;
1678 unreachable("unsupported gen\n");
1681 for (unsigned i
= 0; i
< ARRAY_SIZE(device
->dispatch
.entrypoints
); i
++) {
1682 /* Vulkan requires that entrypoints for extensions which have not been
1683 * enabled must not be advertised.
1685 if (!anv_device_entrypoint_is_enabled(i
, device
->instance
->app_info
.api_version
,
1686 &device
->instance
->enabled_extensions
,
1687 &device
->enabled_extensions
)) {
1688 device
->dispatch
.entrypoints
[i
] = NULL
;
1689 } else if (genX_table
->entrypoints
[i
]) {
1690 device
->dispatch
.entrypoints
[i
] = genX_table
->entrypoints
[i
];
1692 device
->dispatch
.entrypoints
[i
] =
1693 anv_device_dispatch_table
.entrypoints
[i
];
1699 vk_priority_to_gen(int priority
)
1702 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT
:
1703 return GEN_CONTEXT_LOW_PRIORITY
;
1704 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
:
1705 return GEN_CONTEXT_MEDIUM_PRIORITY
;
1706 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT
:
1707 return GEN_CONTEXT_HIGH_PRIORITY
;
1708 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT
:
1709 return GEN_CONTEXT_REALTIME_PRIORITY
;
1711 unreachable("Invalid priority");
1716 anv_device_init_hiz_clear_value_bo(struct anv_device
*device
)
1718 anv_bo_init_new(&device
->hiz_clear_bo
, device
, 4096);
1720 if (device
->instance
->physicalDevice
.has_exec_async
)
1721 device
->hiz_clear_bo
.flags
|= EXEC_OBJECT_ASYNC
;
1723 if (device
->instance
->physicalDevice
.use_softpin
)
1724 device
->hiz_clear_bo
.flags
|= EXEC_OBJECT_PINNED
;
1726 anv_vma_alloc(device
, &device
->hiz_clear_bo
);
1728 uint32_t *map
= anv_gem_mmap(device
, device
->hiz_clear_bo
.gem_handle
,
1731 union isl_color_value hiz_clear
= { .u32
= { 0, } };
1732 hiz_clear
.f32
[0] = ANV_HZ_FC_VAL
;
1734 memcpy(map
, hiz_clear
.u32
, sizeof(hiz_clear
.u32
));
1735 anv_gem_munmap(map
, device
->hiz_clear_bo
.size
);
1739 get_bo_from_pool(struct gen_batch_decode_bo
*ret
,
1740 struct anv_block_pool
*pool
,
1743 for (uint32_t i
= 0; i
< pool
->nbos
; i
++) {
1744 uint64_t bo_address
= pool
->bos
[i
].offset
& (~0ull >> 16);
1745 uint32_t bo_size
= pool
->bos
[i
].size
;
1746 if (address
>= bo_address
&& address
< (bo_address
+ bo_size
)) {
1747 *ret
= (struct gen_batch_decode_bo
) {
1750 .map
= pool
->bos
[i
].map
,
1758 /* Finding a buffer for batch decoding */
1759 static struct gen_batch_decode_bo
1760 decode_get_bo(void *v_batch
, uint64_t address
)
1762 struct anv_device
*device
= v_batch
;
1763 struct gen_batch_decode_bo ret_bo
= {};
1765 if (get_bo_from_pool(&ret_bo
, &device
->dynamic_state_pool
.block_pool
, address
))
1767 if (get_bo_from_pool(&ret_bo
, &device
->instruction_state_pool
.block_pool
, address
))
1769 if (get_bo_from_pool(&ret_bo
, &device
->binding_table_pool
.block_pool
, address
))
1771 if (get_bo_from_pool(&ret_bo
, &device
->surface_state_pool
.block_pool
, address
))
1774 if (!device
->cmd_buffer_being_decoded
)
1775 return (struct gen_batch_decode_bo
) { };
1777 struct anv_batch_bo
**bo
;
1779 u_vector_foreach(bo
, &device
->cmd_buffer_being_decoded
->seen_bbos
) {
1780 /* The decoder zeroes out the top 16 bits, so we need to as well */
1781 uint64_t bo_address
= (*bo
)->bo
.offset
& (~0ull >> 16);
1783 if (address
>= bo_address
&& address
< bo_address
+ (*bo
)->bo
.size
) {
1784 return (struct gen_batch_decode_bo
) {
1786 .size
= (*bo
)->bo
.size
,
1787 .map
= (*bo
)->bo
.map
,
1792 return (struct gen_batch_decode_bo
) { };
1795 VkResult
anv_CreateDevice(
1796 VkPhysicalDevice physicalDevice
,
1797 const VkDeviceCreateInfo
* pCreateInfo
,
1798 const VkAllocationCallbacks
* pAllocator
,
1801 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
1803 struct anv_device
*device
;
1805 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO
);
1807 struct anv_device_extension_table enabled_extensions
= { };
1808 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
1810 for (idx
= 0; idx
< ANV_DEVICE_EXTENSION_COUNT
; idx
++) {
1811 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
1812 anv_device_extensions
[idx
].extensionName
) == 0)
1816 if (idx
>= ANV_DEVICE_EXTENSION_COUNT
)
1817 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
1819 if (!physical_device
->supported_extensions
.extensions
[idx
])
1820 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
1822 enabled_extensions
.extensions
[idx
] = true;
1825 /* Check enabled features */
1826 if (pCreateInfo
->pEnabledFeatures
) {
1827 VkPhysicalDeviceFeatures supported_features
;
1828 anv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
1829 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
1830 VkBool32
*enabled_feature
= (VkBool32
*)pCreateInfo
->pEnabledFeatures
;
1831 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
1832 for (uint32_t i
= 0; i
< num_features
; i
++) {
1833 if (enabled_feature
[i
] && !supported_feature
[i
])
1834 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
1838 /* Check requested queues and fail if we are requested to create any
1839 * queues with flags we don't support.
1841 assert(pCreateInfo
->queueCreateInfoCount
> 0);
1842 for (uint32_t i
= 0; i
< pCreateInfo
->queueCreateInfoCount
; i
++) {
1843 if (pCreateInfo
->pQueueCreateInfos
[i
].flags
!= 0)
1844 return vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1847 /* Check if client specified queue priority. */
1848 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*queue_priority
=
1849 vk_find_struct_const(pCreateInfo
->pQueueCreateInfos
[0].pNext
,
1850 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT
);
1852 VkQueueGlobalPriorityEXT priority
=
1853 queue_priority
? queue_priority
->globalPriority
:
1854 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
;
1856 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
1858 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
1860 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1862 const unsigned decode_flags
=
1863 GEN_BATCH_DECODE_FULL
|
1864 ((INTEL_DEBUG
& DEBUG_COLOR
) ? GEN_BATCH_DECODE_IN_COLOR
: 0) |
1865 GEN_BATCH_DECODE_OFFSETS
|
1866 GEN_BATCH_DECODE_FLOATS
;
1868 gen_batch_decode_ctx_init(&device
->decoder_ctx
,
1869 &physical_device
->info
,
1870 stderr
, decode_flags
, NULL
,
1871 decode_get_bo
, NULL
, device
);
1873 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1874 device
->instance
= physical_device
->instance
;
1875 device
->chipset_id
= physical_device
->chipset_id
;
1876 device
->no_hw
= physical_device
->no_hw
;
1877 device
->_lost
= false;
1880 device
->alloc
= *pAllocator
;
1882 device
->alloc
= physical_device
->instance
->alloc
;
1884 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
1885 device
->fd
= open(physical_device
->path
, O_RDWR
| O_CLOEXEC
);
1886 if (device
->fd
== -1) {
1887 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1891 device
->context_id
= anv_gem_create_context(device
);
1892 if (device
->context_id
== -1) {
1893 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1897 if (physical_device
->use_softpin
) {
1898 if (pthread_mutex_init(&device
->vma_mutex
, NULL
) != 0) {
1899 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1903 /* keep the page with address zero out of the allocator */
1904 util_vma_heap_init(&device
->vma_lo
, LOW_HEAP_MIN_ADDRESS
, LOW_HEAP_SIZE
);
1905 device
->vma_lo_available
=
1906 physical_device
->memory
.heaps
[physical_device
->memory
.heap_count
- 1].size
;
1908 /* Leave the last 4GiB out of the high vma range, so that no state base
1909 * address + size can overflow 48 bits. For more information see the
1910 * comment about Wa32bitGeneralStateOffset in anv_allocator.c
1912 util_vma_heap_init(&device
->vma_hi
, HIGH_HEAP_MIN_ADDRESS
,
1914 device
->vma_hi_available
= physical_device
->memory
.heap_count
== 1 ? 0 :
1915 physical_device
->memory
.heaps
[0].size
;
1918 /* As per spec, the driver implementation may deny requests to acquire
1919 * a priority above the default priority (MEDIUM) if the caller does not
1920 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
1923 if (physical_device
->has_context_priority
) {
1924 int err
= anv_gem_set_context_param(device
->fd
, device
->context_id
,
1925 I915_CONTEXT_PARAM_PRIORITY
,
1926 vk_priority_to_gen(priority
));
1927 if (err
!= 0 && priority
> VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
) {
1928 result
= vk_error(VK_ERROR_NOT_PERMITTED_EXT
);
1933 device
->info
= physical_device
->info
;
1934 device
->isl_dev
= physical_device
->isl_dev
;
1936 /* On Broadwell and later, we can use batch chaining to more efficiently
1937 * implement growing command buffers. Prior to Haswell, the kernel
1938 * command parser gets in the way and we have to fall back to growing
1941 device
->can_chain_batches
= device
->info
.gen
>= 8;
1943 device
->robust_buffer_access
= pCreateInfo
->pEnabledFeatures
&&
1944 pCreateInfo
->pEnabledFeatures
->robustBufferAccess
;
1945 device
->enabled_extensions
= enabled_extensions
;
1947 anv_device_init_dispatch(device
);
1949 if (pthread_mutex_init(&device
->mutex
, NULL
) != 0) {
1950 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1951 goto fail_context_id
;
1954 pthread_condattr_t condattr
;
1955 if (pthread_condattr_init(&condattr
) != 0) {
1956 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1959 if (pthread_condattr_setclock(&condattr
, CLOCK_MONOTONIC
) != 0) {
1960 pthread_condattr_destroy(&condattr
);
1961 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1964 if (pthread_cond_init(&device
->queue_submit
, NULL
) != 0) {
1965 pthread_condattr_destroy(&condattr
);
1966 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1969 pthread_condattr_destroy(&condattr
);
1972 (physical_device
->supports_48bit_addresses
? EXEC_OBJECT_SUPPORTS_48B_ADDRESS
: 0) |
1973 (physical_device
->has_exec_async
? EXEC_OBJECT_ASYNC
: 0) |
1974 (physical_device
->has_exec_capture
? EXEC_OBJECT_CAPTURE
: 0) |
1975 (physical_device
->use_softpin
? EXEC_OBJECT_PINNED
: 0);
1977 anv_bo_pool_init(&device
->batch_bo_pool
, device
, bo_flags
);
1979 result
= anv_bo_cache_init(&device
->bo_cache
);
1980 if (result
!= VK_SUCCESS
)
1981 goto fail_batch_bo_pool
;
1983 if (!physical_device
->use_softpin
)
1984 bo_flags
&= ~EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
1986 result
= anv_state_pool_init(&device
->dynamic_state_pool
, device
,
1987 DYNAMIC_STATE_POOL_MIN_ADDRESS
,
1990 if (result
!= VK_SUCCESS
)
1993 result
= anv_state_pool_init(&device
->instruction_state_pool
, device
,
1994 INSTRUCTION_STATE_POOL_MIN_ADDRESS
,
1997 if (result
!= VK_SUCCESS
)
1998 goto fail_dynamic_state_pool
;
2000 result
= anv_state_pool_init(&device
->surface_state_pool
, device
,
2001 SURFACE_STATE_POOL_MIN_ADDRESS
,
2004 if (result
!= VK_SUCCESS
)
2005 goto fail_instruction_state_pool
;
2007 if (physical_device
->use_softpin
) {
2008 result
= anv_state_pool_init(&device
->binding_table_pool
, device
,
2009 BINDING_TABLE_POOL_MIN_ADDRESS
,
2012 if (result
!= VK_SUCCESS
)
2013 goto fail_surface_state_pool
;
2016 result
= anv_bo_init_new(&device
->workaround_bo
, device
, 1024);
2017 if (result
!= VK_SUCCESS
)
2018 goto fail_binding_table_pool
;
2020 if (physical_device
->use_softpin
)
2021 device
->workaround_bo
.flags
|= EXEC_OBJECT_PINNED
;
2023 if (!anv_vma_alloc(device
, &device
->workaround_bo
))
2024 goto fail_workaround_bo
;
2026 anv_device_init_trivial_batch(device
);
2028 if (device
->info
.gen
>= 10)
2029 anv_device_init_hiz_clear_value_bo(device
);
2031 if (physical_device
->use_softpin
)
2032 device
->pinned_buffers
= _mesa_pointer_set_create(NULL
);
2034 anv_scratch_pool_init(device
, &device
->scratch_pool
);
2036 anv_queue_init(device
, &device
->queue
);
2038 switch (device
->info
.gen
) {
2040 if (!device
->info
.is_haswell
)
2041 result
= gen7_init_device_state(device
);
2043 result
= gen75_init_device_state(device
);
2046 result
= gen8_init_device_state(device
);
2049 result
= gen9_init_device_state(device
);
2052 result
= gen10_init_device_state(device
);
2055 result
= gen11_init_device_state(device
);
2058 /* Shouldn't get here as we don't create physical devices for any other
2060 unreachable("unhandled gen");
2062 if (result
!= VK_SUCCESS
)
2063 goto fail_workaround_bo
;
2065 anv_pipeline_cache_init(&device
->default_pipeline_cache
, device
, true);
2067 anv_device_init_blorp(device
);
2069 anv_device_init_border_colors(device
);
2071 *pDevice
= anv_device_to_handle(device
);
2076 anv_queue_finish(&device
->queue
);
2077 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
2078 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
2079 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
2080 fail_binding_table_pool
:
2081 if (physical_device
->use_softpin
)
2082 anv_state_pool_finish(&device
->binding_table_pool
);
2083 fail_surface_state_pool
:
2084 anv_state_pool_finish(&device
->surface_state_pool
);
2085 fail_instruction_state_pool
:
2086 anv_state_pool_finish(&device
->instruction_state_pool
);
2087 fail_dynamic_state_pool
:
2088 anv_state_pool_finish(&device
->dynamic_state_pool
);
2090 anv_bo_cache_finish(&device
->bo_cache
);
2092 anv_bo_pool_finish(&device
->batch_bo_pool
);
2093 pthread_cond_destroy(&device
->queue_submit
);
2095 pthread_mutex_destroy(&device
->mutex
);
2097 anv_gem_destroy_context(device
, device
->context_id
);
2101 vk_free(&device
->alloc
, device
);
2106 void anv_DestroyDevice(
2108 const VkAllocationCallbacks
* pAllocator
)
2110 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2111 struct anv_physical_device
*physical_device
;
2116 physical_device
= &device
->instance
->physicalDevice
;
2118 anv_device_finish_blorp(device
);
2120 anv_pipeline_cache_finish(&device
->default_pipeline_cache
);
2122 anv_queue_finish(&device
->queue
);
2124 if (physical_device
->use_softpin
)
2125 _mesa_set_destroy(device
->pinned_buffers
, NULL
);
2127 #ifdef HAVE_VALGRIND
2128 /* We only need to free these to prevent valgrind errors. The backing
2129 * BO will go away in a couple of lines so we don't actually leak.
2131 anv_state_pool_free(&device
->dynamic_state_pool
, device
->border_colors
);
2134 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
2136 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
2137 anv_vma_free(device
, &device
->workaround_bo
);
2138 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
2140 anv_vma_free(device
, &device
->trivial_batch_bo
);
2141 anv_gem_close(device
, device
->trivial_batch_bo
.gem_handle
);
2142 if (device
->info
.gen
>= 10)
2143 anv_gem_close(device
, device
->hiz_clear_bo
.gem_handle
);
2145 if (physical_device
->use_softpin
)
2146 anv_state_pool_finish(&device
->binding_table_pool
);
2147 anv_state_pool_finish(&device
->surface_state_pool
);
2148 anv_state_pool_finish(&device
->instruction_state_pool
);
2149 anv_state_pool_finish(&device
->dynamic_state_pool
);
2151 anv_bo_cache_finish(&device
->bo_cache
);
2153 anv_bo_pool_finish(&device
->batch_bo_pool
);
2155 pthread_cond_destroy(&device
->queue_submit
);
2156 pthread_mutex_destroy(&device
->mutex
);
2158 anv_gem_destroy_context(device
, device
->context_id
);
2160 gen_batch_decode_ctx_finish(&device
->decoder_ctx
);
2164 vk_free(&device
->alloc
, device
);
2167 VkResult
anv_EnumerateInstanceLayerProperties(
2168 uint32_t* pPropertyCount
,
2169 VkLayerProperties
* pProperties
)
2171 if (pProperties
== NULL
) {
2172 *pPropertyCount
= 0;
2176 /* None supported at this time */
2177 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
2180 VkResult
anv_EnumerateDeviceLayerProperties(
2181 VkPhysicalDevice physicalDevice
,
2182 uint32_t* pPropertyCount
,
2183 VkLayerProperties
* pProperties
)
2185 if (pProperties
== NULL
) {
2186 *pPropertyCount
= 0;
2190 /* None supported at this time */
2191 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
2194 void anv_GetDeviceQueue(
2196 uint32_t queueNodeIndex
,
2197 uint32_t queueIndex
,
2200 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2202 assert(queueIndex
== 0);
2204 *pQueue
= anv_queue_to_handle(&device
->queue
);
2207 void anv_GetDeviceQueue2(
2209 const VkDeviceQueueInfo2
* pQueueInfo
,
2212 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2214 assert(pQueueInfo
->queueIndex
== 0);
2216 if (pQueueInfo
->flags
== device
->queue
.flags
)
2217 *pQueue
= anv_queue_to_handle(&device
->queue
);
2223 _anv_device_set_lost(struct anv_device
*device
,
2224 const char *file
, int line
,
2225 const char *msg
, ...)
2230 device
->_lost
= true;
2233 err
= __vk_errorv(device
->instance
, device
,
2234 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT
,
2235 VK_ERROR_DEVICE_LOST
, file
, line
, msg
, ap
);
2238 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
2245 anv_device_query_status(struct anv_device
*device
)
2247 /* This isn't likely as most of the callers of this function already check
2248 * for it. However, it doesn't hurt to check and it potentially lets us
2251 if (anv_device_is_lost(device
))
2252 return VK_ERROR_DEVICE_LOST
;
2254 uint32_t active
, pending
;
2255 int ret
= anv_gem_gpu_get_reset_stats(device
, &active
, &pending
);
2257 /* We don't know the real error. */
2258 return anv_device_set_lost(device
, "get_reset_stats failed: %m");
2262 return anv_device_set_lost(device
, "GPU hung on one of our command buffers");
2263 } else if (pending
) {
2264 return anv_device_set_lost(device
, "GPU hung with commands in-flight");
2271 anv_device_bo_busy(struct anv_device
*device
, struct anv_bo
*bo
)
2273 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
2274 * Other usages of the BO (such as on different hardware) will not be
2275 * flagged as "busy" by this ioctl. Use with care.
2277 int ret
= anv_gem_busy(device
, bo
->gem_handle
);
2279 return VK_NOT_READY
;
2280 } else if (ret
== -1) {
2281 /* We don't know the real error. */
2282 return anv_device_set_lost(device
, "gem wait failed: %m");
2285 /* Query for device status after the busy call. If the BO we're checking
2286 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
2287 * client because it clearly doesn't have valid data. Yes, this most
2288 * likely means an ioctl, but we just did an ioctl to query the busy status
2289 * so it's no great loss.
2291 return anv_device_query_status(device
);
2295 anv_device_wait(struct anv_device
*device
, struct anv_bo
*bo
,
2298 int ret
= anv_gem_wait(device
, bo
->gem_handle
, &timeout
);
2299 if (ret
== -1 && errno
== ETIME
) {
2301 } else if (ret
== -1) {
2302 /* We don't know the real error. */
2303 return anv_device_set_lost(device
, "gem wait failed: %m");
2306 /* Query for device status after the wait. If the BO we're waiting on got
2307 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
2308 * because it clearly doesn't have valid data. Yes, this most likely means
2309 * an ioctl, but we just did an ioctl to wait so it's no great loss.
2311 return anv_device_query_status(device
);
2314 VkResult
anv_DeviceWaitIdle(
2317 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2318 if (anv_device_is_lost(device
))
2319 return VK_ERROR_DEVICE_LOST
;
2321 struct anv_batch batch
;
2324 batch
.start
= batch
.next
= cmds
;
2325 batch
.end
= (void *) cmds
+ sizeof(cmds
);
2327 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
2328 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
2330 return anv_device_submit_simple_batch(device
, &batch
);
2334 anv_vma_alloc(struct anv_device
*device
, struct anv_bo
*bo
)
2336 if (!(bo
->flags
& EXEC_OBJECT_PINNED
))
2339 pthread_mutex_lock(&device
->vma_mutex
);
2343 if (bo
->flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
&&
2344 device
->vma_hi_available
>= bo
->size
) {
2345 uint64_t addr
= util_vma_heap_alloc(&device
->vma_hi
, bo
->size
, 4096);
2347 bo
->offset
= gen_canonical_address(addr
);
2348 assert(addr
== gen_48b_address(bo
->offset
));
2349 device
->vma_hi_available
-= bo
->size
;
2353 if (bo
->offset
== 0 && device
->vma_lo_available
>= bo
->size
) {
2354 uint64_t addr
= util_vma_heap_alloc(&device
->vma_lo
, bo
->size
, 4096);
2356 bo
->offset
= gen_canonical_address(addr
);
2357 assert(addr
== gen_48b_address(bo
->offset
));
2358 device
->vma_lo_available
-= bo
->size
;
2362 pthread_mutex_unlock(&device
->vma_mutex
);
2364 return bo
->offset
!= 0;
2368 anv_vma_free(struct anv_device
*device
, struct anv_bo
*bo
)
2370 if (!(bo
->flags
& EXEC_OBJECT_PINNED
))
2373 const uint64_t addr_48b
= gen_48b_address(bo
->offset
);
2375 pthread_mutex_lock(&device
->vma_mutex
);
2377 if (addr_48b
>= LOW_HEAP_MIN_ADDRESS
&&
2378 addr_48b
<= LOW_HEAP_MAX_ADDRESS
) {
2379 util_vma_heap_free(&device
->vma_lo
, addr_48b
, bo
->size
);
2380 device
->vma_lo_available
+= bo
->size
;
2382 assert(addr_48b
>= HIGH_HEAP_MIN_ADDRESS
&&
2383 addr_48b
<= HIGH_HEAP_MAX_ADDRESS
);
2384 util_vma_heap_free(&device
->vma_hi
, addr_48b
, bo
->size
);
2385 device
->vma_hi_available
+= bo
->size
;
2388 pthread_mutex_unlock(&device
->vma_mutex
);
2394 anv_bo_init_new(struct anv_bo
*bo
, struct anv_device
*device
, uint64_t size
)
2396 uint32_t gem_handle
= anv_gem_create(device
, size
);
2398 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
2400 anv_bo_init(bo
, gem_handle
, size
);
2405 VkResult
anv_AllocateMemory(
2407 const VkMemoryAllocateInfo
* pAllocateInfo
,
2408 const VkAllocationCallbacks
* pAllocator
,
2409 VkDeviceMemory
* pMem
)
2411 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2412 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2413 struct anv_device_memory
*mem
;
2414 VkResult result
= VK_SUCCESS
;
2416 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
2418 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
2419 assert(pAllocateInfo
->allocationSize
> 0);
2421 if (pAllocateInfo
->allocationSize
> MAX_MEMORY_ALLOCATION_SIZE
)
2422 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
2424 /* FINISHME: Fail if allocation request exceeds heap size. */
2426 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
2427 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2429 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2431 assert(pAllocateInfo
->memoryTypeIndex
< pdevice
->memory
.type_count
);
2432 mem
->type
= &pdevice
->memory
.types
[pAllocateInfo
->memoryTypeIndex
];
2437 uint64_t bo_flags
= 0;
2439 assert(mem
->type
->heapIndex
< pdevice
->memory
.heap_count
);
2440 if (pdevice
->memory
.heaps
[mem
->type
->heapIndex
].supports_48bit_addresses
)
2441 bo_flags
|= EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
2443 const struct wsi_memory_allocate_info
*wsi_info
=
2444 vk_find_struct_const(pAllocateInfo
->pNext
, WSI_MEMORY_ALLOCATE_INFO_MESA
);
2445 if (wsi_info
&& wsi_info
->implicit_sync
) {
2446 /* We need to set the WRITE flag on window system buffers so that GEM
2447 * will know we're writing to them and synchronize uses on other rings
2448 * (eg if the display server uses the blitter ring).
2450 bo_flags
|= EXEC_OBJECT_WRITE
;
2451 } else if (pdevice
->has_exec_async
) {
2452 bo_flags
|= EXEC_OBJECT_ASYNC
;
2455 if (pdevice
->use_softpin
)
2456 bo_flags
|= EXEC_OBJECT_PINNED
;
2458 const VkExportMemoryAllocateInfo
*export_info
=
2459 vk_find_struct_const(pAllocateInfo
->pNext
, EXPORT_MEMORY_ALLOCATE_INFO
);
2461 /* Check if we need to support Android HW buffer export. If so,
2462 * create AHardwareBuffer and import memory from it.
2464 bool android_export
= false;
2465 if (export_info
&& export_info
->handleTypes
&
2466 VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID
)
2467 android_export
= true;
2469 /* Android memory import. */
2470 const struct VkImportAndroidHardwareBufferInfoANDROID
*ahw_import_info
=
2471 vk_find_struct_const(pAllocateInfo
->pNext
,
2472 IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID
);
2474 if (ahw_import_info
) {
2475 result
= anv_import_ahw_memory(_device
, mem
, ahw_import_info
);
2476 if (result
!= VK_SUCCESS
)
2480 } else if (android_export
) {
2481 result
= anv_create_ahw_memory(_device
, mem
, pAllocateInfo
);
2482 if (result
!= VK_SUCCESS
)
2485 const struct VkImportAndroidHardwareBufferInfoANDROID import_info
= {
2488 result
= anv_import_ahw_memory(_device
, mem
, &import_info
);
2489 if (result
!= VK_SUCCESS
)
2495 const VkImportMemoryFdInfoKHR
*fd_info
=
2496 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_FD_INFO_KHR
);
2498 /* The Vulkan spec permits handleType to be 0, in which case the struct is
2501 if (fd_info
&& fd_info
->handleType
) {
2502 /* At the moment, we support only the below handle types. */
2503 assert(fd_info
->handleType
==
2504 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
2505 fd_info
->handleType
==
2506 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
2508 result
= anv_bo_cache_import(device
, &device
->bo_cache
, fd_info
->fd
,
2509 bo_flags
| ANV_BO_EXTERNAL
, &mem
->bo
);
2510 if (result
!= VK_SUCCESS
)
2513 VkDeviceSize aligned_alloc_size
=
2514 align_u64(pAllocateInfo
->allocationSize
, 4096);
2516 /* For security purposes, we reject importing the bo if it's smaller
2517 * than the requested allocation size. This prevents a malicious client
2518 * from passing a buffer to a trusted client, lying about the size, and
2519 * telling the trusted client to try and texture from an image that goes
2520 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
2521 * in the trusted client. The trusted client can protect itself against
2522 * this sort of attack but only if it can trust the buffer size.
2524 if (mem
->bo
->size
< aligned_alloc_size
) {
2525 result
= vk_errorf(device
->instance
, device
,
2526 VK_ERROR_INVALID_EXTERNAL_HANDLE
,
2527 "aligned allocationSize too large for "
2528 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT: "
2529 "%"PRIu64
"B > %"PRIu64
"B",
2530 aligned_alloc_size
, mem
->bo
->size
);
2531 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
2535 /* From the Vulkan spec:
2537 * "Importing memory from a file descriptor transfers ownership of
2538 * the file descriptor from the application to the Vulkan
2539 * implementation. The application must not perform any operations on
2540 * the file descriptor after a successful import."
2542 * If the import fails, we leave the file descriptor open.
2548 /* Regular allocate (not importing memory). */
2550 if (export_info
&& export_info
->handleTypes
)
2551 bo_flags
|= ANV_BO_EXTERNAL
;
2553 result
= anv_bo_cache_alloc(device
, &device
->bo_cache
,
2554 pAllocateInfo
->allocationSize
, bo_flags
,
2556 if (result
!= VK_SUCCESS
)
2559 const VkMemoryDedicatedAllocateInfo
*dedicated_info
=
2560 vk_find_struct_const(pAllocateInfo
->pNext
, MEMORY_DEDICATED_ALLOCATE_INFO
);
2561 if (dedicated_info
&& dedicated_info
->image
!= VK_NULL_HANDLE
) {
2562 ANV_FROM_HANDLE(anv_image
, image
, dedicated_info
->image
);
2564 /* Some legacy (non-modifiers) consumers need the tiling to be set on
2565 * the BO. In this case, we have a dedicated allocation.
2567 if (image
->needs_set_tiling
) {
2568 const uint32_t i915_tiling
=
2569 isl_tiling_to_i915_tiling(image
->planes
[0].surface
.isl
.tiling
);
2570 int ret
= anv_gem_set_tiling(device
, mem
->bo
->gem_handle
,
2571 image
->planes
[0].surface
.isl
.row_pitch_B
,
2574 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
2575 return vk_errorf(device
->instance
, NULL
,
2576 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
2577 "failed to set BO tiling: %m");
2583 *pMem
= anv_device_memory_to_handle(mem
);
2588 vk_free2(&device
->alloc
, pAllocator
, mem
);
2593 VkResult
anv_GetMemoryFdKHR(
2595 const VkMemoryGetFdInfoKHR
* pGetFdInfo
,
2598 ANV_FROM_HANDLE(anv_device
, dev
, device_h
);
2599 ANV_FROM_HANDLE(anv_device_memory
, mem
, pGetFdInfo
->memory
);
2601 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
2603 assert(pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
2604 pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
2606 return anv_bo_cache_export(dev
, &dev
->bo_cache
, mem
->bo
, pFd
);
2609 VkResult
anv_GetMemoryFdPropertiesKHR(
2611 VkExternalMemoryHandleTypeFlagBits handleType
,
2613 VkMemoryFdPropertiesKHR
* pMemoryFdProperties
)
2615 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2616 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2618 switch (handleType
) {
2619 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
:
2620 /* dma-buf can be imported as any memory type */
2621 pMemoryFdProperties
->memoryTypeBits
=
2622 (1 << pdevice
->memory
.type_count
) - 1;
2626 /* The valid usage section for this function says:
2628 * "handleType must not be one of the handle types defined as
2631 * So opaque handle types fall into the default "unsupported" case.
2633 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
2637 void anv_FreeMemory(
2639 VkDeviceMemory _mem
,
2640 const VkAllocationCallbacks
* pAllocator
)
2642 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2643 ANV_FROM_HANDLE(anv_device_memory
, mem
, _mem
);
2649 anv_UnmapMemory(_device
, _mem
);
2651 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
2655 AHardwareBuffer_release(mem
->ahw
);
2658 vk_free2(&device
->alloc
, pAllocator
, mem
);
2661 VkResult
anv_MapMemory(
2663 VkDeviceMemory _memory
,
2664 VkDeviceSize offset
,
2666 VkMemoryMapFlags flags
,
2669 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2670 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
2677 if (size
== VK_WHOLE_SIZE
)
2678 size
= mem
->bo
->size
- offset
;
2680 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
2682 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
2683 * assert(size != 0);
2684 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
2685 * equal to the size of the memory minus offset
2688 assert(offset
+ size
<= mem
->bo
->size
);
2690 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
2691 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
2692 * at a time is valid. We could just mmap up front and return an offset
2693 * pointer here, but that may exhaust virtual memory on 32 bit
2696 uint32_t gem_flags
= 0;
2698 if (!device
->info
.has_llc
&&
2699 (mem
->type
->propertyFlags
& VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
))
2700 gem_flags
|= I915_MMAP_WC
;
2702 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
2703 uint64_t map_offset
= offset
& ~4095ull;
2704 assert(offset
>= map_offset
);
2705 uint64_t map_size
= (offset
+ size
) - map_offset
;
2707 /* Let's map whole pages */
2708 map_size
= align_u64(map_size
, 4096);
2710 void *map
= anv_gem_mmap(device
, mem
->bo
->gem_handle
,
2711 map_offset
, map_size
, gem_flags
);
2712 if (map
== MAP_FAILED
)
2713 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
2716 mem
->map_size
= map_size
;
2718 *ppData
= mem
->map
+ (offset
- map_offset
);
2723 void anv_UnmapMemory(
2725 VkDeviceMemory _memory
)
2727 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
2732 anv_gem_munmap(mem
->map
, mem
->map_size
);
2739 clflush_mapped_ranges(struct anv_device
*device
,
2741 const VkMappedMemoryRange
*ranges
)
2743 for (uint32_t i
= 0; i
< count
; i
++) {
2744 ANV_FROM_HANDLE(anv_device_memory
, mem
, ranges
[i
].memory
);
2745 if (ranges
[i
].offset
>= mem
->map_size
)
2748 gen_clflush_range(mem
->map
+ ranges
[i
].offset
,
2749 MIN2(ranges
[i
].size
, mem
->map_size
- ranges
[i
].offset
));
2753 VkResult
anv_FlushMappedMemoryRanges(
2755 uint32_t memoryRangeCount
,
2756 const VkMappedMemoryRange
* pMemoryRanges
)
2758 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2760 if (device
->info
.has_llc
)
2763 /* Make sure the writes we're flushing have landed. */
2764 __builtin_ia32_mfence();
2766 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
2771 VkResult
anv_InvalidateMappedMemoryRanges(
2773 uint32_t memoryRangeCount
,
2774 const VkMappedMemoryRange
* pMemoryRanges
)
2776 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2778 if (device
->info
.has_llc
)
2781 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
2783 /* Make sure no reads get moved up above the invalidate. */
2784 __builtin_ia32_mfence();
2789 void anv_GetBufferMemoryRequirements(
2792 VkMemoryRequirements
* pMemoryRequirements
)
2794 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
2795 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2796 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2798 /* The Vulkan spec (git aaed022) says:
2800 * memoryTypeBits is a bitfield and contains one bit set for every
2801 * supported memory type for the resource. The bit `1<<i` is set if and
2802 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
2803 * structure for the physical device is supported.
2805 uint32_t memory_types
= 0;
2806 for (uint32_t i
= 0; i
< pdevice
->memory
.type_count
; i
++) {
2807 uint32_t valid_usage
= pdevice
->memory
.types
[i
].valid_buffer_usage
;
2808 if ((valid_usage
& buffer
->usage
) == buffer
->usage
)
2809 memory_types
|= (1u << i
);
2812 /* Base alignment requirement of a cache line */
2813 uint32_t alignment
= 16;
2815 /* We need an alignment of 32 for pushing UBOs */
2816 if (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
)
2817 alignment
= MAX2(alignment
, 32);
2819 pMemoryRequirements
->size
= buffer
->size
;
2820 pMemoryRequirements
->alignment
= alignment
;
2822 /* Storage and Uniform buffers should have their size aligned to
2823 * 32-bits to avoid boundary checks when last DWord is not complete.
2824 * This would ensure that not internal padding would be needed for
2827 if (device
->robust_buffer_access
&&
2828 (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
||
2829 buffer
->usage
& VK_BUFFER_USAGE_STORAGE_BUFFER_BIT
))
2830 pMemoryRequirements
->size
= align_u64(buffer
->size
, 4);
2832 pMemoryRequirements
->memoryTypeBits
= memory_types
;
2835 void anv_GetBufferMemoryRequirements2(
2837 const VkBufferMemoryRequirementsInfo2
* pInfo
,
2838 VkMemoryRequirements2
* pMemoryRequirements
)
2840 anv_GetBufferMemoryRequirements(_device
, pInfo
->buffer
,
2841 &pMemoryRequirements
->memoryRequirements
);
2843 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
2844 switch (ext
->sType
) {
2845 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
2846 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
2847 requirements
->prefersDedicatedAllocation
= VK_FALSE
;
2848 requirements
->requiresDedicatedAllocation
= VK_FALSE
;
2853 anv_debug_ignored_stype(ext
->sType
);
2859 void anv_GetImageMemoryRequirements(
2862 VkMemoryRequirements
* pMemoryRequirements
)
2864 ANV_FROM_HANDLE(anv_image
, image
, _image
);
2865 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2866 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2868 /* The Vulkan spec (git aaed022) says:
2870 * memoryTypeBits is a bitfield and contains one bit set for every
2871 * supported memory type for the resource. The bit `1<<i` is set if and
2872 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
2873 * structure for the physical device is supported.
2875 * All types are currently supported for images.
2877 uint32_t memory_types
= (1ull << pdevice
->memory
.type_count
) - 1;
2879 /* We must have image allocated or imported at this point. According to the
2880 * specification, external images must have been bound to memory before
2881 * calling GetImageMemoryRequirements.
2883 assert(image
->size
> 0);
2885 pMemoryRequirements
->size
= image
->size
;
2886 pMemoryRequirements
->alignment
= image
->alignment
;
2887 pMemoryRequirements
->memoryTypeBits
= memory_types
;
2890 void anv_GetImageMemoryRequirements2(
2892 const VkImageMemoryRequirementsInfo2
* pInfo
,
2893 VkMemoryRequirements2
* pMemoryRequirements
)
2895 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2896 ANV_FROM_HANDLE(anv_image
, image
, pInfo
->image
);
2898 anv_GetImageMemoryRequirements(_device
, pInfo
->image
,
2899 &pMemoryRequirements
->memoryRequirements
);
2901 vk_foreach_struct_const(ext
, pInfo
->pNext
) {
2902 switch (ext
->sType
) {
2903 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO
: {
2904 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2905 const VkImagePlaneMemoryRequirementsInfo
*plane_reqs
=
2906 (const VkImagePlaneMemoryRequirementsInfo
*) ext
;
2907 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
,
2908 plane_reqs
->planeAspect
);
2910 assert(image
->planes
[plane
].offset
== 0);
2912 /* The Vulkan spec (git aaed022) says:
2914 * memoryTypeBits is a bitfield and contains one bit set for every
2915 * supported memory type for the resource. The bit `1<<i` is set
2916 * if and only if the memory type `i` in the
2917 * VkPhysicalDeviceMemoryProperties structure for the physical
2918 * device is supported.
2920 * All types are currently supported for images.
2922 pMemoryRequirements
->memoryRequirements
.memoryTypeBits
=
2923 (1ull << pdevice
->memory
.type_count
) - 1;
2925 /* We must have image allocated or imported at this point. According to the
2926 * specification, external images must have been bound to memory before
2927 * calling GetImageMemoryRequirements.
2929 assert(image
->planes
[plane
].size
> 0);
2931 pMemoryRequirements
->memoryRequirements
.size
= image
->planes
[plane
].size
;
2932 pMemoryRequirements
->memoryRequirements
.alignment
=
2933 image
->planes
[plane
].alignment
;
2938 anv_debug_ignored_stype(ext
->sType
);
2943 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
2944 switch (ext
->sType
) {
2945 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
2946 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
2947 if (image
->needs_set_tiling
|| image
->external_format
) {
2948 /* If we need to set the tiling for external consumers, we need a
2949 * dedicated allocation.
2951 * See also anv_AllocateMemory.
2953 requirements
->prefersDedicatedAllocation
= VK_TRUE
;
2954 requirements
->requiresDedicatedAllocation
= VK_TRUE
;
2956 requirements
->prefersDedicatedAllocation
= VK_FALSE
;
2957 requirements
->requiresDedicatedAllocation
= VK_FALSE
;
2963 anv_debug_ignored_stype(ext
->sType
);
2969 void anv_GetImageSparseMemoryRequirements(
2972 uint32_t* pSparseMemoryRequirementCount
,
2973 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
2975 *pSparseMemoryRequirementCount
= 0;
2978 void anv_GetImageSparseMemoryRequirements2(
2980 const VkImageSparseMemoryRequirementsInfo2
* pInfo
,
2981 uint32_t* pSparseMemoryRequirementCount
,
2982 VkSparseImageMemoryRequirements2
* pSparseMemoryRequirements
)
2984 *pSparseMemoryRequirementCount
= 0;
2987 void anv_GetDeviceMemoryCommitment(
2989 VkDeviceMemory memory
,
2990 VkDeviceSize
* pCommittedMemoryInBytes
)
2992 *pCommittedMemoryInBytes
= 0;
2996 anv_bind_buffer_memory(const VkBindBufferMemoryInfo
*pBindInfo
)
2998 ANV_FROM_HANDLE(anv_device_memory
, mem
, pBindInfo
->memory
);
2999 ANV_FROM_HANDLE(anv_buffer
, buffer
, pBindInfo
->buffer
);
3001 assert(pBindInfo
->sType
== VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
);
3004 assert((buffer
->usage
& mem
->type
->valid_buffer_usage
) == buffer
->usage
);
3005 buffer
->address
= (struct anv_address
) {
3007 .offset
= pBindInfo
->memoryOffset
,
3010 buffer
->address
= ANV_NULL_ADDRESS
;
3014 VkResult
anv_BindBufferMemory(
3017 VkDeviceMemory memory
,
3018 VkDeviceSize memoryOffset
)
3020 anv_bind_buffer_memory(
3021 &(VkBindBufferMemoryInfo
) {
3022 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
,
3025 .memoryOffset
= memoryOffset
,
3031 VkResult
anv_BindBufferMemory2(
3033 uint32_t bindInfoCount
,
3034 const VkBindBufferMemoryInfo
* pBindInfos
)
3036 for (uint32_t i
= 0; i
< bindInfoCount
; i
++)
3037 anv_bind_buffer_memory(&pBindInfos
[i
]);
3042 VkResult
anv_QueueBindSparse(
3044 uint32_t bindInfoCount
,
3045 const VkBindSparseInfo
* pBindInfo
,
3048 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
3049 if (anv_device_is_lost(queue
->device
))
3050 return VK_ERROR_DEVICE_LOST
;
3052 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
3057 VkResult
anv_CreateEvent(
3059 const VkEventCreateInfo
* pCreateInfo
,
3060 const VkAllocationCallbacks
* pAllocator
,
3063 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3064 struct anv_state state
;
3065 struct anv_event
*event
;
3067 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_EVENT_CREATE_INFO
);
3069 state
= anv_state_pool_alloc(&device
->dynamic_state_pool
,
3072 event
->state
= state
;
3073 event
->semaphore
= VK_EVENT_RESET
;
3075 if (!device
->info
.has_llc
) {
3076 /* Make sure the writes we're flushing have landed. */
3077 __builtin_ia32_mfence();
3078 __builtin_ia32_clflush(event
);
3081 *pEvent
= anv_event_to_handle(event
);
3086 void anv_DestroyEvent(
3089 const VkAllocationCallbacks
* pAllocator
)
3091 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3092 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3097 anv_state_pool_free(&device
->dynamic_state_pool
, event
->state
);
3100 VkResult
anv_GetEventStatus(
3104 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3105 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3107 if (anv_device_is_lost(device
))
3108 return VK_ERROR_DEVICE_LOST
;
3110 if (!device
->info
.has_llc
) {
3111 /* Invalidate read cache before reading event written by GPU. */
3112 __builtin_ia32_clflush(event
);
3113 __builtin_ia32_mfence();
3117 return event
->semaphore
;
3120 VkResult
anv_SetEvent(
3124 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3125 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3127 event
->semaphore
= VK_EVENT_SET
;
3129 if (!device
->info
.has_llc
) {
3130 /* Make sure the writes we're flushing have landed. */
3131 __builtin_ia32_mfence();
3132 __builtin_ia32_clflush(event
);
3138 VkResult
anv_ResetEvent(
3142 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3143 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3145 event
->semaphore
= VK_EVENT_RESET
;
3147 if (!device
->info
.has_llc
) {
3148 /* Make sure the writes we're flushing have landed. */
3149 __builtin_ia32_mfence();
3150 __builtin_ia32_clflush(event
);
3158 VkResult
anv_CreateBuffer(
3160 const VkBufferCreateInfo
* pCreateInfo
,
3161 const VkAllocationCallbacks
* pAllocator
,
3164 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3165 struct anv_buffer
*buffer
;
3167 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
3169 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
3170 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3172 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3174 buffer
->size
= pCreateInfo
->size
;
3175 buffer
->usage
= pCreateInfo
->usage
;
3176 buffer
->address
= ANV_NULL_ADDRESS
;
3178 if (buffer
->usage
& VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT_EXT
) {
3179 pthread_mutex_lock(&device
->mutex
);
3180 _mesa_set_add(device
->pinned_buffers
, buffer
);
3181 pthread_mutex_unlock(&device
->mutex
);
3184 *pBuffer
= anv_buffer_to_handle(buffer
);
3189 void anv_DestroyBuffer(
3192 const VkAllocationCallbacks
* pAllocator
)
3194 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3195 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
3200 if (buffer
->usage
& VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT_EXT
) {
3201 pthread_mutex_lock(&device
->mutex
);
3202 _mesa_set_remove_key(device
->pinned_buffers
, buffer
);
3203 pthread_mutex_unlock(&device
->mutex
);
3206 vk_free2(&device
->alloc
, pAllocator
, buffer
);
3209 VkDeviceAddress
anv_GetBufferDeviceAddressEXT(
3211 const VkBufferDeviceAddressInfoEXT
* pInfo
)
3213 ANV_FROM_HANDLE(anv_buffer
, buffer
, pInfo
->buffer
);
3215 assert(buffer
->address
.bo
->flags
& EXEC_OBJECT_PINNED
);
3217 return anv_address_physical(buffer
->address
);
3221 anv_fill_buffer_surface_state(struct anv_device
*device
, struct anv_state state
,
3222 enum isl_format format
,
3223 struct anv_address address
,
3224 uint32_t range
, uint32_t stride
)
3226 isl_buffer_fill_state(&device
->isl_dev
, state
.map
,
3227 .address
= anv_address_physical(address
),
3228 .mocs
= device
->default_mocs
,
3231 .stride_B
= stride
);
3234 void anv_DestroySampler(
3237 const VkAllocationCallbacks
* pAllocator
)
3239 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3240 ANV_FROM_HANDLE(anv_sampler
, sampler
, _sampler
);
3245 vk_free2(&device
->alloc
, pAllocator
, sampler
);
3248 VkResult
anv_CreateFramebuffer(
3250 const VkFramebufferCreateInfo
* pCreateInfo
,
3251 const VkAllocationCallbacks
* pAllocator
,
3252 VkFramebuffer
* pFramebuffer
)
3254 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3255 struct anv_framebuffer
*framebuffer
;
3257 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
3259 size_t size
= sizeof(*framebuffer
) +
3260 sizeof(struct anv_image_view
*) * pCreateInfo
->attachmentCount
;
3261 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
3262 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3263 if (framebuffer
== NULL
)
3264 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3266 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
3267 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
3268 VkImageView _iview
= pCreateInfo
->pAttachments
[i
];
3269 framebuffer
->attachments
[i
] = anv_image_view_from_handle(_iview
);
3272 framebuffer
->width
= pCreateInfo
->width
;
3273 framebuffer
->height
= pCreateInfo
->height
;
3274 framebuffer
->layers
= pCreateInfo
->layers
;
3276 *pFramebuffer
= anv_framebuffer_to_handle(framebuffer
);
3281 void anv_DestroyFramebuffer(
3284 const VkAllocationCallbacks
* pAllocator
)
3286 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3287 ANV_FROM_HANDLE(anv_framebuffer
, fb
, _fb
);
3292 vk_free2(&device
->alloc
, pAllocator
, fb
);
3295 static const VkTimeDomainEXT anv_time_domains
[] = {
3296 VK_TIME_DOMAIN_DEVICE_EXT
,
3297 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
,
3298 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
,
3301 VkResult
anv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
3302 VkPhysicalDevice physicalDevice
,
3303 uint32_t *pTimeDomainCount
,
3304 VkTimeDomainEXT
*pTimeDomains
)
3307 VK_OUTARRAY_MAKE(out
, pTimeDomains
, pTimeDomainCount
);
3309 for (d
= 0; d
< ARRAY_SIZE(anv_time_domains
); d
++) {
3310 vk_outarray_append(&out
, i
) {
3311 *i
= anv_time_domains
[d
];
3315 return vk_outarray_status(&out
);
3319 anv_clock_gettime(clockid_t clock_id
)
3321 struct timespec current
;
3324 ret
= clock_gettime(clock_id
, ¤t
);
3325 if (ret
< 0 && clock_id
== CLOCK_MONOTONIC_RAW
)
3326 ret
= clock_gettime(CLOCK_MONOTONIC
, ¤t
);
3330 return (uint64_t) current
.tv_sec
* 1000000000ULL + current
.tv_nsec
;
3333 #define TIMESTAMP 0x2358
3335 VkResult
anv_GetCalibratedTimestampsEXT(
3337 uint32_t timestampCount
,
3338 const VkCalibratedTimestampInfoEXT
*pTimestampInfos
,
3339 uint64_t *pTimestamps
,
3340 uint64_t *pMaxDeviation
)
3342 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3343 uint64_t timestamp_frequency
= device
->info
.timestamp_frequency
;
3346 uint64_t begin
, end
;
3347 uint64_t max_clock_period
= 0;
3349 begin
= anv_clock_gettime(CLOCK_MONOTONIC_RAW
);
3351 for (d
= 0; d
< timestampCount
; d
++) {
3352 switch (pTimestampInfos
[d
].timeDomain
) {
3353 case VK_TIME_DOMAIN_DEVICE_EXT
:
3354 ret
= anv_gem_reg_read(device
, TIMESTAMP
| 1,
3358 return anv_device_set_lost(device
, "Failed to read the TIMESTAMP "
3361 uint64_t device_period
= DIV_ROUND_UP(1000000000, timestamp_frequency
);
3362 max_clock_period
= MAX2(max_clock_period
, device_period
);
3364 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
:
3365 pTimestamps
[d
] = anv_clock_gettime(CLOCK_MONOTONIC
);
3366 max_clock_period
= MAX2(max_clock_period
, 1);
3369 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
:
3370 pTimestamps
[d
] = begin
;
3378 end
= anv_clock_gettime(CLOCK_MONOTONIC_RAW
);
3381 * The maximum deviation is the sum of the interval over which we
3382 * perform the sampling and the maximum period of any sampled
3383 * clock. That's because the maximum skew between any two sampled
3384 * clock edges is when the sampled clock with the largest period is
3385 * sampled at the end of that period but right at the beginning of the
3386 * sampling interval and some other clock is sampled right at the
3387 * begining of its sampling period and right at the end of the
3388 * sampling interval. Let's assume the GPU has the longest clock
3389 * period and that the application is sampling GPU and monotonic:
3392 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
3393 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
3397 * GPU -----_____-----_____-----_____-----_____
3400 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
3401 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
3403 * Interval <----------------->
3404 * Deviation <-------------------------->
3408 * m = read(monotonic) 2
3411 * We round the sample interval up by one tick to cover sampling error
3412 * in the interval clock
3415 uint64_t sample_interval
= end
- begin
+ 1;
3417 *pMaxDeviation
= sample_interval
+ max_clock_period
;
3422 /* vk_icd.h does not declare this function, so we declare it here to
3423 * suppress Wmissing-prototypes.
3425 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
3426 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
);
3428 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
3429 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
)
3431 /* For the full details on loader interface versioning, see
3432 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
3433 * What follows is a condensed summary, to help you navigate the large and
3434 * confusing official doc.
3436 * - Loader interface v0 is incompatible with later versions. We don't
3439 * - In loader interface v1:
3440 * - The first ICD entrypoint called by the loader is
3441 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
3443 * - The ICD must statically expose no other Vulkan symbol unless it is
3444 * linked with -Bsymbolic.
3445 * - Each dispatchable Vulkan handle created by the ICD must be
3446 * a pointer to a struct whose first member is VK_LOADER_DATA. The
3447 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
3448 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
3449 * vkDestroySurfaceKHR(). The ICD must be capable of working with
3450 * such loader-managed surfaces.
3452 * - Loader interface v2 differs from v1 in:
3453 * - The first ICD entrypoint called by the loader is
3454 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
3455 * statically expose this entrypoint.
3457 * - Loader interface v3 differs from v2 in:
3458 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
3459 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
3460 * because the loader no longer does so.
3462 *pSupportedVersion
= MIN2(*pSupportedVersion
, 3u);