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/debug.h"
36 #include "util/build_id.h"
37 #include "util/disk_cache.h"
38 #include "util/mesa-sha1.h"
39 #include "util/os_file.h"
40 #include "util/u_atomic.h"
41 #include "util/u_string.h"
42 #include "util/xmlpool.h"
45 #include "common/gen_aux_map.h"
46 #include "common/gen_defines.h"
47 #include "compiler/glsl_types.h"
49 #include "genxml/gen7_pack.h"
51 static const char anv_dri_options_xml
[] =
53 DRI_CONF_SECTION_PERFORMANCE
54 DRI_CONF_VK_X11_OVERRIDE_MIN_IMAGE_COUNT(0)
55 DRI_CONF_VK_X11_STRICT_IMAGE_COUNT("false")
58 DRI_CONF_SECTION_DEBUG
59 DRI_CONF_ALWAYS_FLUSH_CACHE("false")
63 /* This is probably far to big but it reflects the max size used for messages
64 * in OpenGLs KHR_debug.
66 #define MAX_DEBUG_MESSAGE_LENGTH 4096
69 compiler_debug_log(void *data
, const char *fmt
, ...)
71 char str
[MAX_DEBUG_MESSAGE_LENGTH
];
72 struct anv_device
*device
= (struct anv_device
*)data
;
74 if (list_is_empty(&device
->instance
->debug_report_callbacks
.callbacks
))
79 (void) vsnprintf(str
, MAX_DEBUG_MESSAGE_LENGTH
, fmt
, args
);
82 vk_debug_report(&device
->instance
->debug_report_callbacks
,
83 VK_DEBUG_REPORT_DEBUG_BIT_EXT
,
84 VK_DEBUG_REPORT_OBJECT_TYPE_UNKNOWN_EXT
,
89 compiler_perf_log(void *data
, const char *fmt
, ...)
94 if (unlikely(INTEL_DEBUG
& DEBUG_PERF
))
95 intel_logd_v(fmt
, args
);
101 anv_compute_heap_size(int fd
, uint64_t gtt_size
)
103 /* Query the total ram from the system */
107 uint64_t total_ram
= (uint64_t)info
.totalram
* (uint64_t)info
.mem_unit
;
109 /* We don't want to burn too much ram with the GPU. If the user has 4GiB
110 * or less, we use at most half. If they have more than 4GiB, we use 3/4.
112 uint64_t available_ram
;
113 if (total_ram
<= 4ull * 1024ull * 1024ull * 1024ull)
114 available_ram
= total_ram
/ 2;
116 available_ram
= total_ram
* 3 / 4;
118 /* We also want to leave some padding for things we allocate in the driver,
119 * so don't go over 3/4 of the GTT either.
121 uint64_t available_gtt
= gtt_size
* 3 / 4;
123 return MIN2(available_ram
, available_gtt
);
127 anv_physical_device_init_heaps(struct anv_physical_device
*device
, int fd
)
129 if (anv_gem_get_context_param(fd
, 0, I915_CONTEXT_PARAM_GTT_SIZE
,
130 &device
->gtt_size
) == -1) {
131 /* If, for whatever reason, we can't actually get the GTT size from the
132 * kernel (too old?) fall back to the aperture size.
134 anv_perf_warn(NULL
, NULL
,
135 "Failed to get I915_CONTEXT_PARAM_GTT_SIZE: %m");
137 if (anv_gem_get_aperture(fd
, &device
->gtt_size
) == -1) {
138 return vk_errorf(NULL
, NULL
, VK_ERROR_INITIALIZATION_FAILED
,
139 "failed to get aperture size: %m");
143 /* We only allow 48-bit addresses with softpin because knowing the actual
144 * address is required for the vertex cache flush workaround.
146 device
->supports_48bit_addresses
= (device
->info
.gen
>= 8) &&
147 device
->has_softpin
&&
148 device
->gtt_size
> (4ULL << 30 /* GiB */);
150 uint64_t heap_size
= anv_compute_heap_size(fd
, device
->gtt_size
);
152 if (heap_size
> (2ull << 30) && !device
->supports_48bit_addresses
) {
153 /* When running with an overridden PCI ID, we may get a GTT size from
154 * the kernel that is greater than 2 GiB but the execbuf check for 48bit
155 * address support can still fail. Just clamp the address space size to
156 * 2 GiB if we don't have 48-bit support.
158 intel_logw("%s:%d: The kernel reported a GTT size larger than 2 GiB but "
159 "not support for 48-bit addresses",
161 heap_size
= 2ull << 30;
164 device
->memory
.heap_count
= 1;
165 device
->memory
.heaps
[0] = (struct anv_memory_heap
) {
167 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
170 uint32_t type_count
= 0;
171 for (uint32_t heap
= 0; heap
< device
->memory
.heap_count
; heap
++) {
172 if (device
->info
.has_llc
) {
173 /* Big core GPUs share LLC with the CPU and thus one memory type can be
174 * both cached and coherent at the same time.
176 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
177 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
178 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
179 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
180 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
184 /* The spec requires that we expose a host-visible, coherent memory
185 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
186 * to give the application a choice between cached, but not coherent and
187 * coherent but uncached (WC though).
189 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
190 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
191 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
192 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
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_CACHED_BIT
,
203 device
->memory
.type_count
= type_count
;
209 anv_physical_device_init_uuids(struct anv_physical_device
*device
)
211 const struct build_id_note
*note
=
212 build_id_find_nhdr_for_addr(anv_physical_device_init_uuids
);
214 return vk_errorf(device
->instance
, device
,
215 VK_ERROR_INITIALIZATION_FAILED
,
216 "Failed to find build-id");
219 unsigned build_id_len
= build_id_length(note
);
220 if (build_id_len
< 20) {
221 return vk_errorf(device
->instance
, device
,
222 VK_ERROR_INITIALIZATION_FAILED
,
223 "build-id too short. It needs to be a SHA");
226 memcpy(device
->driver_build_sha1
, build_id_data(note
), 20);
228 struct mesa_sha1 sha1_ctx
;
230 STATIC_ASSERT(VK_UUID_SIZE
<= sizeof(sha1
));
232 /* The pipeline cache UUID is used for determining when a pipeline cache is
233 * invalid. It needs both a driver build and the PCI ID of the device.
235 _mesa_sha1_init(&sha1_ctx
);
236 _mesa_sha1_update(&sha1_ctx
, build_id_data(note
), build_id_len
);
237 _mesa_sha1_update(&sha1_ctx
, &device
->chipset_id
,
238 sizeof(device
->chipset_id
));
239 _mesa_sha1_update(&sha1_ctx
, &device
->always_use_bindless
,
240 sizeof(device
->always_use_bindless
));
241 _mesa_sha1_update(&sha1_ctx
, &device
->has_a64_buffer_access
,
242 sizeof(device
->has_a64_buffer_access
));
243 _mesa_sha1_update(&sha1_ctx
, &device
->has_bindless_images
,
244 sizeof(device
->has_bindless_images
));
245 _mesa_sha1_update(&sha1_ctx
, &device
->has_bindless_samplers
,
246 sizeof(device
->has_bindless_samplers
));
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 ASSERTED
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 get_available_system_memory()
308 char *meminfo
= os_read_file("/proc/meminfo");
312 char *str
= strstr(meminfo
, "MemAvailable:");
318 uint64_t kb_mem_available
;
319 if (sscanf(str
, "MemAvailable: %" PRIx64
, &kb_mem_available
) == 1) {
321 return kb_mem_available
<< 10;
329 anv_physical_device_init(struct anv_physical_device
*device
,
330 struct anv_instance
*instance
,
331 drmDevicePtr drm_device
)
333 const char *primary_path
= drm_device
->nodes
[DRM_NODE_PRIMARY
];
334 const char *path
= drm_device
->nodes
[DRM_NODE_RENDER
];
339 brw_process_intel_debug_variable();
341 fd
= open(path
, O_RDWR
| O_CLOEXEC
);
343 return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
345 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
346 device
->instance
= instance
;
348 assert(strlen(path
) < ARRAY_SIZE(device
->path
));
349 snprintf(device
->path
, ARRAY_SIZE(device
->path
), "%s", path
);
351 if (!gen_get_device_info_from_fd(fd
, &device
->info
)) {
352 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
355 device
->chipset_id
= device
->info
.chipset_id
;
356 device
->no_hw
= device
->info
.no_hw
;
358 if (getenv("INTEL_NO_HW") != NULL
)
359 device
->no_hw
= true;
361 device
->pci_info
.domain
= drm_device
->businfo
.pci
->domain
;
362 device
->pci_info
.bus
= drm_device
->businfo
.pci
->bus
;
363 device
->pci_info
.device
= drm_device
->businfo
.pci
->dev
;
364 device
->pci_info
.function
= drm_device
->businfo
.pci
->func
;
366 device
->name
= gen_get_device_name(device
->chipset_id
);
368 if (device
->info
.is_haswell
) {
369 intel_logw("Haswell Vulkan support is incomplete");
370 } else if (device
->info
.gen
== 7 && !device
->info
.is_baytrail
) {
371 intel_logw("Ivy Bridge Vulkan support is incomplete");
372 } else if (device
->info
.gen
== 7 && device
->info
.is_baytrail
) {
373 intel_logw("Bay Trail Vulkan support is incomplete");
374 } else if (device
->info
.gen
>= 8 && device
->info
.gen
<= 11) {
375 /* Gen8-11 fully supported */
376 } else if (device
->info
.gen
== 12) {
377 intel_logw("Vulkan is not yet fully supported on gen12");
379 result
= vk_errorf(device
->instance
, device
,
380 VK_ERROR_INCOMPATIBLE_DRIVER
,
381 "Vulkan not yet supported on %s", device
->name
);
385 device
->cmd_parser_version
= -1;
386 if (device
->info
.gen
== 7) {
387 device
->cmd_parser_version
=
388 anv_gem_get_param(fd
, I915_PARAM_CMD_PARSER_VERSION
);
389 if (device
->cmd_parser_version
== -1) {
390 result
= vk_errorf(device
->instance
, device
,
391 VK_ERROR_INITIALIZATION_FAILED
,
392 "failed to get command parser version");
397 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_WAIT_TIMEOUT
)) {
398 result
= vk_errorf(device
->instance
, device
,
399 VK_ERROR_INITIALIZATION_FAILED
,
400 "kernel missing gem wait");
404 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_EXECBUF2
)) {
405 result
= vk_errorf(device
->instance
, device
,
406 VK_ERROR_INITIALIZATION_FAILED
,
407 "kernel missing execbuf2");
411 if (!device
->info
.has_llc
&&
412 anv_gem_get_param(fd
, I915_PARAM_MMAP_VERSION
) < 1) {
413 result
= vk_errorf(device
->instance
, device
,
414 VK_ERROR_INITIALIZATION_FAILED
,
415 "kernel missing wc mmap");
419 device
->has_softpin
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_SOFTPIN
);
420 device
->has_exec_async
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_ASYNC
);
421 device
->has_exec_capture
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_CAPTURE
);
422 device
->has_exec_fence
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE
);
423 device
->has_syncobj
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE_ARRAY
);
424 device
->has_syncobj_wait
= device
->has_syncobj
&&
425 anv_gem_supports_syncobj_wait(fd
);
426 device
->has_context_priority
= anv_gem_has_context_priority(fd
);
428 result
= anv_physical_device_init_heaps(device
, fd
);
429 if (result
!= VK_SUCCESS
)
432 device
->use_softpin
= device
->has_softpin
&&
433 device
->supports_48bit_addresses
;
435 device
->has_context_isolation
=
436 anv_gem_get_param(fd
, I915_PARAM_HAS_CONTEXT_ISOLATION
);
438 device
->always_use_bindless
=
439 env_var_as_boolean("ANV_ALWAYS_BINDLESS", false);
441 /* We first got the A64 messages on broadwell and we can only use them if
442 * we can pass addresses directly into the shader which requires softpin.
444 device
->has_a64_buffer_access
= device
->info
.gen
>= 8 &&
447 /* We first get bindless image access on Skylake and we can only really do
448 * it if we don't have any relocations so we need softpin.
450 device
->has_bindless_images
= device
->info
.gen
>= 9 &&
453 /* We've had bindless samplers since Ivy Bridge (forever in Vulkan terms)
454 * because it's just a matter of setting the sampler address in the sample
455 * message header. However, we've not bothered to wire it up for vec4 so
456 * we leave it disabled on gen7.
458 device
->has_bindless_samplers
= device
->info
.gen
>= 8;
460 device
->has_mem_available
= get_available_system_memory() != 0;
462 device
->always_flush_cache
=
463 driQueryOptionb(&instance
->dri_options
, "always_flush_cache");
465 /* Starting with Gen10, the timestamp frequency of the command streamer may
466 * vary from one part to another. We can query the value from the kernel.
468 if (device
->info
.gen
>= 10) {
469 int timestamp_frequency
=
470 anv_gem_get_param(fd
, I915_PARAM_CS_TIMESTAMP_FREQUENCY
);
472 if (timestamp_frequency
< 0)
473 intel_logw("Kernel 4.16-rc1+ required to properly query CS timestamp frequency");
475 device
->info
.timestamp_frequency
= timestamp_frequency
;
478 /* GENs prior to 8 do not support EU/Subslice info */
479 if (device
->info
.gen
>= 8) {
480 device
->subslice_total
= anv_gem_get_param(fd
, I915_PARAM_SUBSLICE_TOTAL
);
481 device
->eu_total
= anv_gem_get_param(fd
, I915_PARAM_EU_TOTAL
);
483 /* Without this information, we cannot get the right Braswell
484 * brandstrings, and we have to use conservative numbers for GPGPU on
485 * many platforms, but otherwise, things will just work.
487 if (device
->subslice_total
< 1 || device
->eu_total
< 1) {
488 intel_logw("Kernel 4.1 required to properly query GPU properties");
490 } else if (device
->info
.gen
== 7) {
491 device
->subslice_total
= 1 << (device
->info
.gt
- 1);
494 if (device
->info
.is_cherryview
&&
495 device
->subslice_total
> 0 && device
->eu_total
> 0) {
496 /* Logical CS threads = EUs per subslice * num threads per EU */
497 uint32_t max_cs_threads
=
498 device
->eu_total
/ device
->subslice_total
* device
->info
.num_thread_per_eu
;
500 /* Fuse configurations may give more threads than expected, never less. */
501 if (max_cs_threads
> device
->info
.max_cs_threads
)
502 device
->info
.max_cs_threads
= max_cs_threads
;
505 device
->compiler
= brw_compiler_create(NULL
, &device
->info
);
506 if (device
->compiler
== NULL
) {
507 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
510 device
->compiler
->shader_debug_log
= compiler_debug_log
;
511 device
->compiler
->shader_perf_log
= compiler_perf_log
;
512 device
->compiler
->supports_pull_constants
= false;
513 device
->compiler
->constant_buffer_0_is_relative
=
514 device
->info
.gen
< 8 || !device
->has_context_isolation
;
515 device
->compiler
->supports_shader_constants
= true;
516 device
->compiler
->compact_params
= false;
518 /* Broadwell PRM says:
520 * "Before Gen8, there was a historical configuration control field to
521 * swizzle address bit[6] for in X/Y tiling modes. This was set in three
522 * different places: TILECTL[1:0], ARB_MODE[5:4], and
523 * DISP_ARB_CTL[14:13].
525 * For Gen8 and subsequent generations, the swizzle fields are all
526 * reserved, and the CPU's memory controller performs all address
527 * swizzling modifications."
530 device
->info
.gen
< 8 && anv_gem_get_bit6_swizzle(fd
, I915_TILING_X
);
532 isl_device_init(&device
->isl_dev
, &device
->info
, swizzled
);
534 result
= anv_physical_device_init_uuids(device
);
535 if (result
!= VK_SUCCESS
)
538 anv_physical_device_init_disk_cache(device
);
540 if (instance
->enabled_extensions
.KHR_display
) {
541 master_fd
= open(primary_path
, O_RDWR
| O_CLOEXEC
);
542 if (master_fd
>= 0) {
543 /* prod the device with a GETPARAM call which will fail if
544 * we don't have permission to even render on this device
546 if (anv_gem_get_param(master_fd
, I915_PARAM_CHIPSET_ID
) == 0) {
552 device
->master_fd
= master_fd
;
554 result
= anv_init_wsi(device
);
555 if (result
!= VK_SUCCESS
) {
556 ralloc_free(device
->compiler
);
557 anv_physical_device_free_disk_cache(device
);
561 device
->perf
= anv_get_perf(&device
->info
, fd
);
563 anv_physical_device_get_supported_extensions(device
,
564 &device
->supported_extensions
);
567 device
->local_fd
= fd
;
579 anv_physical_device_finish(struct anv_physical_device
*device
)
581 anv_finish_wsi(device
);
582 anv_physical_device_free_disk_cache(device
);
583 ralloc_free(device
->compiler
);
584 ralloc_free(device
->perf
);
585 close(device
->local_fd
);
586 if (device
->master_fd
>= 0)
587 close(device
->master_fd
);
591 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
592 VkSystemAllocationScope allocationScope
)
598 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
599 size_t align
, VkSystemAllocationScope allocationScope
)
601 return realloc(pOriginal
, size
);
605 default_free_func(void *pUserData
, void *pMemory
)
610 static const VkAllocationCallbacks default_alloc
= {
612 .pfnAllocation
= default_alloc_func
,
613 .pfnReallocation
= default_realloc_func
,
614 .pfnFree
= default_free_func
,
617 VkResult
anv_EnumerateInstanceExtensionProperties(
618 const char* pLayerName
,
619 uint32_t* pPropertyCount
,
620 VkExtensionProperties
* pProperties
)
622 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
624 for (int i
= 0; i
< ANV_INSTANCE_EXTENSION_COUNT
; i
++) {
625 if (anv_instance_extensions_supported
.extensions
[i
]) {
626 vk_outarray_append(&out
, prop
) {
627 *prop
= anv_instance_extensions
[i
];
632 return vk_outarray_status(&out
);
635 VkResult
anv_CreateInstance(
636 const VkInstanceCreateInfo
* pCreateInfo
,
637 const VkAllocationCallbacks
* pAllocator
,
638 VkInstance
* pInstance
)
640 struct anv_instance
*instance
;
643 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
645 struct anv_instance_extension_table enabled_extensions
= {};
646 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
648 for (idx
= 0; idx
< ANV_INSTANCE_EXTENSION_COUNT
; idx
++) {
649 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
650 anv_instance_extensions
[idx
].extensionName
) == 0)
654 if (idx
>= ANV_INSTANCE_EXTENSION_COUNT
)
655 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
657 if (!anv_instance_extensions_supported
.extensions
[idx
])
658 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
660 enabled_extensions
.extensions
[idx
] = true;
663 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
664 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
666 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
668 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
671 instance
->alloc
= *pAllocator
;
673 instance
->alloc
= default_alloc
;
675 instance
->app_info
= (struct anv_app_info
) { .api_version
= 0 };
676 if (pCreateInfo
->pApplicationInfo
) {
677 const VkApplicationInfo
*app
= pCreateInfo
->pApplicationInfo
;
679 instance
->app_info
.app_name
=
680 vk_strdup(&instance
->alloc
, app
->pApplicationName
,
681 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
682 instance
->app_info
.app_version
= app
->applicationVersion
;
684 instance
->app_info
.engine_name
=
685 vk_strdup(&instance
->alloc
, app
->pEngineName
,
686 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
687 instance
->app_info
.engine_version
= app
->engineVersion
;
689 instance
->app_info
.api_version
= app
->apiVersion
;
692 if (instance
->app_info
.api_version
== 0)
693 instance
->app_info
.api_version
= VK_API_VERSION_1_0
;
695 instance
->enabled_extensions
= enabled_extensions
;
697 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->dispatch
.entrypoints
); i
++) {
698 /* Vulkan requires that entrypoints for extensions which have not been
699 * enabled must not be advertised.
701 if (!anv_instance_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
702 &instance
->enabled_extensions
)) {
703 instance
->dispatch
.entrypoints
[i
] = NULL
;
705 instance
->dispatch
.entrypoints
[i
] =
706 anv_instance_dispatch_table
.entrypoints
[i
];
710 struct anv_physical_device
*pdevice
= &instance
->physicalDevice
;
711 for (unsigned i
= 0; i
< ARRAY_SIZE(pdevice
->dispatch
.entrypoints
); i
++) {
712 /* Vulkan requires that entrypoints for extensions which have not been
713 * enabled must not be advertised.
715 if (!anv_physical_device_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
716 &instance
->enabled_extensions
)) {
717 pdevice
->dispatch
.entrypoints
[i
] = NULL
;
719 pdevice
->dispatch
.entrypoints
[i
] =
720 anv_physical_device_dispatch_table
.entrypoints
[i
];
724 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->device_dispatch
.entrypoints
); i
++) {
725 /* Vulkan requires that entrypoints for extensions which have not been
726 * enabled must not be advertised.
728 if (!anv_device_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
729 &instance
->enabled_extensions
, NULL
)) {
730 instance
->device_dispatch
.entrypoints
[i
] = NULL
;
732 instance
->device_dispatch
.entrypoints
[i
] =
733 anv_device_dispatch_table
.entrypoints
[i
];
737 instance
->physicalDeviceCount
= -1;
739 result
= vk_debug_report_instance_init(&instance
->debug_report_callbacks
);
740 if (result
!= VK_SUCCESS
) {
741 vk_free2(&default_alloc
, pAllocator
, instance
);
742 return vk_error(result
);
745 instance
->pipeline_cache_enabled
=
746 env_var_as_boolean("ANV_ENABLE_PIPELINE_CACHE", true);
748 glsl_type_singleton_init_or_ref();
750 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
752 driParseOptionInfo(&instance
->available_dri_options
, anv_dri_options_xml
);
753 driParseConfigFiles(&instance
->dri_options
, &instance
->available_dri_options
,
755 instance
->app_info
.engine_name
,
756 instance
->app_info
.engine_version
);
758 *pInstance
= anv_instance_to_handle(instance
);
763 void anv_DestroyInstance(
764 VkInstance _instance
,
765 const VkAllocationCallbacks
* pAllocator
)
767 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
772 if (instance
->physicalDeviceCount
> 0) {
773 /* We support at most one physical device. */
774 assert(instance
->physicalDeviceCount
== 1);
775 anv_physical_device_finish(&instance
->physicalDevice
);
778 vk_free(&instance
->alloc
, (char *)instance
->app_info
.app_name
);
779 vk_free(&instance
->alloc
, (char *)instance
->app_info
.engine_name
);
781 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
783 vk_debug_report_instance_destroy(&instance
->debug_report_callbacks
);
785 glsl_type_singleton_decref();
787 driDestroyOptionCache(&instance
->dri_options
);
788 driDestroyOptionInfo(&instance
->available_dri_options
);
790 vk_free(&instance
->alloc
, instance
);
794 anv_enumerate_devices(struct anv_instance
*instance
)
796 /* TODO: Check for more devices ? */
797 drmDevicePtr devices
[8];
798 VkResult result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
801 instance
->physicalDeviceCount
= 0;
803 max_devices
= drmGetDevices2(0, devices
, ARRAY_SIZE(devices
));
805 return VK_ERROR_INCOMPATIBLE_DRIVER
;
807 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
808 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
809 devices
[i
]->bustype
== DRM_BUS_PCI
&&
810 devices
[i
]->deviceinfo
.pci
->vendor_id
== 0x8086) {
812 result
= anv_physical_device_init(&instance
->physicalDevice
,
813 instance
, devices
[i
]);
814 if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
818 drmFreeDevices(devices
, max_devices
);
820 if (result
== VK_SUCCESS
)
821 instance
->physicalDeviceCount
= 1;
827 anv_instance_ensure_physical_device(struct anv_instance
*instance
)
829 if (instance
->physicalDeviceCount
< 0) {
830 VkResult result
= anv_enumerate_devices(instance
);
831 if (result
!= VK_SUCCESS
&&
832 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
839 VkResult
anv_EnumeratePhysicalDevices(
840 VkInstance _instance
,
841 uint32_t* pPhysicalDeviceCount
,
842 VkPhysicalDevice
* pPhysicalDevices
)
844 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
845 VK_OUTARRAY_MAKE(out
, pPhysicalDevices
, pPhysicalDeviceCount
);
847 VkResult result
= anv_instance_ensure_physical_device(instance
);
848 if (result
!= VK_SUCCESS
)
851 if (instance
->physicalDeviceCount
== 0)
854 assert(instance
->physicalDeviceCount
== 1);
855 vk_outarray_append(&out
, i
) {
856 *i
= anv_physical_device_to_handle(&instance
->physicalDevice
);
859 return vk_outarray_status(&out
);
862 VkResult
anv_EnumeratePhysicalDeviceGroups(
863 VkInstance _instance
,
864 uint32_t* pPhysicalDeviceGroupCount
,
865 VkPhysicalDeviceGroupProperties
* pPhysicalDeviceGroupProperties
)
867 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
868 VK_OUTARRAY_MAKE(out
, pPhysicalDeviceGroupProperties
,
869 pPhysicalDeviceGroupCount
);
871 VkResult result
= anv_instance_ensure_physical_device(instance
);
872 if (result
!= VK_SUCCESS
)
875 if (instance
->physicalDeviceCount
== 0)
878 assert(instance
->physicalDeviceCount
== 1);
880 vk_outarray_append(&out
, p
) {
881 p
->physicalDeviceCount
= 1;
882 memset(p
->physicalDevices
, 0, sizeof(p
->physicalDevices
));
883 p
->physicalDevices
[0] =
884 anv_physical_device_to_handle(&instance
->physicalDevice
);
885 p
->subsetAllocation
= false;
887 vk_foreach_struct(ext
, p
->pNext
)
888 anv_debug_ignored_stype(ext
->sType
);
891 return vk_outarray_status(&out
);
894 void anv_GetPhysicalDeviceFeatures(
895 VkPhysicalDevice physicalDevice
,
896 VkPhysicalDeviceFeatures
* pFeatures
)
898 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
900 *pFeatures
= (VkPhysicalDeviceFeatures
) {
901 .robustBufferAccess
= true,
902 .fullDrawIndexUint32
= true,
903 .imageCubeArray
= true,
904 .independentBlend
= true,
905 .geometryShader
= true,
906 .tessellationShader
= true,
907 .sampleRateShading
= true,
908 .dualSrcBlend
= true,
910 .multiDrawIndirect
= true,
911 .drawIndirectFirstInstance
= true,
913 .depthBiasClamp
= true,
914 .fillModeNonSolid
= true,
915 .depthBounds
= pdevice
->info
.gen
>= 12,
919 .multiViewport
= true,
920 .samplerAnisotropy
= true,
921 .textureCompressionETC2
= pdevice
->info
.gen
>= 8 ||
922 pdevice
->info
.is_baytrail
,
923 .textureCompressionASTC_LDR
= pdevice
->info
.gen
>= 9, /* FINISHME CHV */
924 .textureCompressionBC
= true,
925 .occlusionQueryPrecise
= true,
926 .pipelineStatisticsQuery
= true,
927 .fragmentStoresAndAtomics
= true,
928 .shaderTessellationAndGeometryPointSize
= true,
929 .shaderImageGatherExtended
= true,
930 .shaderStorageImageExtendedFormats
= true,
931 .shaderStorageImageMultisample
= false,
932 .shaderStorageImageReadWithoutFormat
= false,
933 .shaderStorageImageWriteWithoutFormat
= true,
934 .shaderUniformBufferArrayDynamicIndexing
= true,
935 .shaderSampledImageArrayDynamicIndexing
= true,
936 .shaderStorageBufferArrayDynamicIndexing
= true,
937 .shaderStorageImageArrayDynamicIndexing
= true,
938 .shaderClipDistance
= true,
939 .shaderCullDistance
= true,
940 .shaderFloat64
= pdevice
->info
.gen
>= 8 &&
941 pdevice
->info
.has_64bit_types
,
942 .shaderInt64
= pdevice
->info
.gen
>= 8 &&
943 pdevice
->info
.has_64bit_types
,
944 .shaderInt16
= pdevice
->info
.gen
>= 8,
945 .shaderResourceMinLod
= pdevice
->info
.gen
>= 9,
946 .variableMultisampleRate
= true,
947 .inheritedQueries
= true,
950 /* We can't do image stores in vec4 shaders */
951 pFeatures
->vertexPipelineStoresAndAtomics
=
952 pdevice
->compiler
->scalar_stage
[MESA_SHADER_VERTEX
] &&
953 pdevice
->compiler
->scalar_stage
[MESA_SHADER_GEOMETRY
];
955 struct anv_app_info
*app_info
= &pdevice
->instance
->app_info
;
957 /* The new DOOM and Wolfenstein games require depthBounds without
958 * checking for it. They seem to run fine without it so just claim it's
959 * there and accept the consequences.
961 if (app_info
->engine_name
&& strcmp(app_info
->engine_name
, "idTech") == 0)
962 pFeatures
->depthBounds
= true;
965 void anv_GetPhysicalDeviceFeatures2(
966 VkPhysicalDevice physicalDevice
,
967 VkPhysicalDeviceFeatures2
* pFeatures
)
969 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
970 anv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
972 vk_foreach_struct(ext
, pFeatures
->pNext
) {
973 switch (ext
->sType
) {
974 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES_KHR
: {
975 VkPhysicalDevice8BitStorageFeaturesKHR
*features
=
976 (VkPhysicalDevice8BitStorageFeaturesKHR
*)ext
;
977 features
->storageBuffer8BitAccess
= pdevice
->info
.gen
>= 8;
978 features
->uniformAndStorageBuffer8BitAccess
= pdevice
->info
.gen
>= 8;
979 features
->storagePushConstant8
= pdevice
->info
.gen
>= 8;
983 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES
: {
984 VkPhysicalDevice16BitStorageFeatures
*features
=
985 (VkPhysicalDevice16BitStorageFeatures
*)ext
;
986 features
->storageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
987 features
->uniformAndStorageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
988 features
->storagePushConstant16
= pdevice
->info
.gen
>= 8;
989 features
->storageInputOutput16
= false;
993 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_EXT
: {
994 VkPhysicalDeviceBufferDeviceAddressFeaturesEXT
*features
= (void *)ext
;
995 features
->bufferDeviceAddress
= pdevice
->has_a64_buffer_access
;
996 features
->bufferDeviceAddressCaptureReplay
= false;
997 features
->bufferDeviceAddressMultiDevice
= false;
1001 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COMPUTE_SHADER_DERIVATIVES_FEATURES_NV
: {
1002 VkPhysicalDeviceComputeShaderDerivativesFeaturesNV
*features
=
1003 (VkPhysicalDeviceComputeShaderDerivativesFeaturesNV
*)ext
;
1004 features
->computeDerivativeGroupQuads
= true;
1005 features
->computeDerivativeGroupLinear
= true;
1009 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT
: {
1010 VkPhysicalDeviceConditionalRenderingFeaturesEXT
*features
=
1011 (VkPhysicalDeviceConditionalRenderingFeaturesEXT
*)ext
;
1012 features
->conditionalRendering
= pdevice
->info
.gen
>= 8 ||
1013 pdevice
->info
.is_haswell
;
1014 features
->inheritedConditionalRendering
= pdevice
->info
.gen
>= 8 ||
1015 pdevice
->info
.is_haswell
;
1019 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_CLIP_ENABLE_FEATURES_EXT
: {
1020 VkPhysicalDeviceDepthClipEnableFeaturesEXT
*features
=
1021 (VkPhysicalDeviceDepthClipEnableFeaturesEXT
*)ext
;
1022 features
->depthClipEnable
= true;
1026 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR
: {
1027 VkPhysicalDeviceFloat16Int8FeaturesKHR
*features
= (void *)ext
;
1028 features
->shaderFloat16
= pdevice
->info
.gen
>= 8;
1029 features
->shaderInt8
= pdevice
->info
.gen
>= 8;
1033 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FRAGMENT_SHADER_INTERLOCK_FEATURES_EXT
: {
1034 VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT
*features
=
1035 (VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT
*)ext
;
1036 features
->fragmentShaderSampleInterlock
= pdevice
->info
.gen
>= 9;
1037 features
->fragmentShaderPixelInterlock
= pdevice
->info
.gen
>= 9;
1038 features
->fragmentShaderShadingRateInterlock
= false;
1042 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_QUERY_RESET_FEATURES_EXT
: {
1043 VkPhysicalDeviceHostQueryResetFeaturesEXT
*features
=
1044 (VkPhysicalDeviceHostQueryResetFeaturesEXT
*)ext
;
1045 features
->hostQueryReset
= true;
1049 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT
: {
1050 VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*features
=
1051 (VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*)ext
;
1052 features
->shaderInputAttachmentArrayDynamicIndexing
= false;
1053 features
->shaderUniformTexelBufferArrayDynamicIndexing
= true;
1054 features
->shaderStorageTexelBufferArrayDynamicIndexing
= true;
1055 features
->shaderUniformBufferArrayNonUniformIndexing
= false;
1056 features
->shaderSampledImageArrayNonUniformIndexing
= true;
1057 features
->shaderStorageBufferArrayNonUniformIndexing
= true;
1058 features
->shaderStorageImageArrayNonUniformIndexing
= true;
1059 features
->shaderInputAttachmentArrayNonUniformIndexing
= false;
1060 features
->shaderUniformTexelBufferArrayNonUniformIndexing
= true;
1061 features
->shaderStorageTexelBufferArrayNonUniformIndexing
= true;
1062 features
->descriptorBindingUniformBufferUpdateAfterBind
= false;
1063 features
->descriptorBindingSampledImageUpdateAfterBind
= true;
1064 features
->descriptorBindingStorageImageUpdateAfterBind
= true;
1065 features
->descriptorBindingStorageBufferUpdateAfterBind
= true;
1066 features
->descriptorBindingUniformTexelBufferUpdateAfterBind
= true;
1067 features
->descriptorBindingStorageTexelBufferUpdateAfterBind
= true;
1068 features
->descriptorBindingUpdateUnusedWhilePending
= true;
1069 features
->descriptorBindingPartiallyBound
= true;
1070 features
->descriptorBindingVariableDescriptorCount
= false;
1071 features
->runtimeDescriptorArray
= true;
1075 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INDEX_TYPE_UINT8_FEATURES_EXT
: {
1076 VkPhysicalDeviceIndexTypeUint8FeaturesEXT
*features
=
1077 (VkPhysicalDeviceIndexTypeUint8FeaturesEXT
*)ext
;
1078 features
->indexTypeUint8
= true;
1082 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_FEATURES_EXT
: {
1083 VkPhysicalDeviceInlineUniformBlockFeaturesEXT
*features
=
1084 (VkPhysicalDeviceInlineUniformBlockFeaturesEXT
*)ext
;
1085 features
->inlineUniformBlock
= true;
1086 features
->descriptorBindingInlineUniformBlockUpdateAfterBind
= true;
1090 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_FEATURES_EXT
: {
1091 VkPhysicalDeviceLineRasterizationFeaturesEXT
*features
=
1092 (VkPhysicalDeviceLineRasterizationFeaturesEXT
*)ext
;
1093 features
->rectangularLines
= true;
1094 features
->bresenhamLines
= true;
1095 features
->smoothLines
= true;
1096 features
->stippledRectangularLines
= false;
1097 features
->stippledBresenhamLines
= true;
1098 features
->stippledSmoothLines
= false;
1102 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES
: {
1103 VkPhysicalDeviceMultiviewFeatures
*features
=
1104 (VkPhysicalDeviceMultiviewFeatures
*)ext
;
1105 features
->multiview
= true;
1106 features
->multiviewGeometryShader
= true;
1107 features
->multiviewTessellationShader
= true;
1111 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGELESS_FRAMEBUFFER_FEATURES_KHR
: {
1112 VkPhysicalDeviceImagelessFramebufferFeaturesKHR
*features
=
1113 (VkPhysicalDeviceImagelessFramebufferFeaturesKHR
*)ext
;
1114 features
->imagelessFramebuffer
= true;
1118 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR
: {
1119 VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR
*features
=
1120 (VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR
*)ext
;
1121 features
->pipelineExecutableInfo
= true;
1125 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES
: {
1126 VkPhysicalDeviceProtectedMemoryFeatures
*features
= (void *)ext
;
1127 features
->protectedMemory
= false;
1131 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES
: {
1132 VkPhysicalDeviceSamplerYcbcrConversionFeatures
*features
=
1133 (VkPhysicalDeviceSamplerYcbcrConversionFeatures
*) ext
;
1134 features
->samplerYcbcrConversion
= true;
1138 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES_EXT
: {
1139 VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*features
=
1140 (VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*)ext
;
1141 features
->scalarBlockLayout
= true;
1145 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SEPARATE_DEPTH_STENCIL_LAYOUTS_FEATURES_KHR
: {
1146 VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR
*features
=
1147 (VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR
*)ext
;
1148 features
->separateDepthStencilLayouts
= true;
1152 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_INT64_FEATURES_KHR
: {
1153 VkPhysicalDeviceShaderAtomicInt64FeaturesKHR
*features
= (void *)ext
;
1154 features
->shaderBufferInt64Atomics
=
1155 pdevice
->info
.gen
>= 9 && pdevice
->use_softpin
;
1156 features
->shaderSharedInt64Atomics
= VK_FALSE
;
1160 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DEMOTE_TO_HELPER_INVOCATION_FEATURES_EXT
: {
1161 VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT
*features
= (void *)ext
;
1162 features
->shaderDemoteToHelperInvocation
= true;
1166 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CLOCK_FEATURES_KHR
: {
1167 VkPhysicalDeviceShaderClockFeaturesKHR
*features
=
1168 (VkPhysicalDeviceShaderClockFeaturesKHR
*)ext
;
1169 features
->shaderSubgroupClock
= true;
1170 features
->shaderDeviceClock
= false;
1174 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETERS_FEATURES
: {
1175 VkPhysicalDeviceShaderDrawParametersFeatures
*features
= (void *)ext
;
1176 features
->shaderDrawParameters
= true;
1180 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_SUBGROUP_EXTENDED_TYPES_FEATURES_KHR
: {
1181 VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR
*features
=
1182 (VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR
*)ext
;
1183 features
->shaderSubgroupExtendedTypes
= true;
1187 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_FEATURES_EXT
: {
1188 VkPhysicalDeviceSubgroupSizeControlFeaturesEXT
*features
=
1189 (VkPhysicalDeviceSubgroupSizeControlFeaturesEXT
*)ext
;
1190 features
->subgroupSizeControl
= true;
1191 features
->computeFullSubgroups
= true;
1195 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_FEATURES_EXT
: {
1196 VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT
*features
=
1197 (VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT
*)ext
;
1198 features
->texelBufferAlignment
= true;
1202 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_FEATURES_KHR
: {
1203 VkPhysicalDeviceTimelineSemaphoreFeaturesKHR
*features
=
1204 (VkPhysicalDeviceTimelineSemaphoreFeaturesKHR
*) ext
;
1205 features
->timelineSemaphore
= true;
1209 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES
: {
1210 VkPhysicalDeviceVariablePointersFeatures
*features
= (void *)ext
;
1211 features
->variablePointersStorageBuffer
= true;
1212 features
->variablePointers
= true;
1216 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT
: {
1217 VkPhysicalDeviceTransformFeedbackFeaturesEXT
*features
=
1218 (VkPhysicalDeviceTransformFeedbackFeaturesEXT
*)ext
;
1219 features
->transformFeedback
= true;
1220 features
->geometryStreams
= true;
1224 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_UNIFORM_BUFFER_STANDARD_LAYOUT_FEATURES_KHR
: {
1225 VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR
*features
=
1226 (VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR
*)ext
;
1227 features
->uniformBufferStandardLayout
= true;
1231 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT
: {
1232 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*features
=
1233 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*)ext
;
1234 features
->vertexAttributeInstanceRateDivisor
= true;
1235 features
->vertexAttributeInstanceRateZeroDivisor
= true;
1239 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_MEMORY_MODEL_FEATURES_KHR
: {
1240 VkPhysicalDeviceVulkanMemoryModelFeaturesKHR
*features
= (void *)ext
;
1241 features
->vulkanMemoryModel
= true;
1242 features
->vulkanMemoryModelDeviceScope
= true;
1243 features
->vulkanMemoryModelAvailabilityVisibilityChains
= true;
1247 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_YCBCR_IMAGE_ARRAYS_FEATURES_EXT
: {
1248 VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*features
=
1249 (VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*)ext
;
1250 features
->ycbcrImageArrays
= true;
1255 anv_debug_ignored_stype(ext
->sType
);
1261 #define MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS 64
1263 #define MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS 64
1264 #define MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS 256
1266 void anv_GetPhysicalDeviceProperties(
1267 VkPhysicalDevice physicalDevice
,
1268 VkPhysicalDeviceProperties
* pProperties
)
1270 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1271 const struct gen_device_info
*devinfo
= &pdevice
->info
;
1273 /* See assertions made when programming the buffer surface state. */
1274 const uint32_t max_raw_buffer_sz
= devinfo
->gen
>= 7 ?
1275 (1ul << 30) : (1ul << 27);
1277 const uint32_t max_ssbos
= pdevice
->has_a64_buffer_access
? UINT16_MAX
: 64;
1278 const uint32_t max_textures
=
1279 pdevice
->has_bindless_images
? UINT16_MAX
: 128;
1280 const uint32_t max_samplers
=
1281 pdevice
->has_bindless_samplers
? UINT16_MAX
:
1282 (devinfo
->gen
>= 8 || devinfo
->is_haswell
) ? 128 : 16;
1283 const uint32_t max_images
=
1284 pdevice
->has_bindless_images
? UINT16_MAX
: MAX_IMAGES
;
1286 /* If we can use bindless for everything, claim a high per-stage limit,
1287 * otherwise use the binding table size, minus the slots reserved for
1288 * render targets and one slot for the descriptor buffer. */
1289 const uint32_t max_per_stage
=
1290 pdevice
->has_bindless_images
&& pdevice
->has_a64_buffer_access
1291 ? UINT32_MAX
: MAX_BINDING_TABLE_SIZE
- MAX_RTS
- 1;
1293 const uint32_t max_workgroup_size
= 32 * devinfo
->max_cs_threads
;
1295 VkSampleCountFlags sample_counts
=
1296 isl_device_get_sample_counts(&pdevice
->isl_dev
);
1299 VkPhysicalDeviceLimits limits
= {
1300 .maxImageDimension1D
= (1 << 14),
1301 .maxImageDimension2D
= (1 << 14),
1302 .maxImageDimension3D
= (1 << 11),
1303 .maxImageDimensionCube
= (1 << 14),
1304 .maxImageArrayLayers
= (1 << 11),
1305 .maxTexelBufferElements
= 128 * 1024 * 1024,
1306 .maxUniformBufferRange
= (1ul << 27),
1307 .maxStorageBufferRange
= max_raw_buffer_sz
,
1308 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
1309 .maxMemoryAllocationCount
= UINT32_MAX
,
1310 .maxSamplerAllocationCount
= 64 * 1024,
1311 .bufferImageGranularity
= 64, /* A cache line */
1312 .sparseAddressSpaceSize
= 0,
1313 .maxBoundDescriptorSets
= MAX_SETS
,
1314 .maxPerStageDescriptorSamplers
= max_samplers
,
1315 .maxPerStageDescriptorUniformBuffers
= MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
,
1316 .maxPerStageDescriptorStorageBuffers
= max_ssbos
,
1317 .maxPerStageDescriptorSampledImages
= max_textures
,
1318 .maxPerStageDescriptorStorageImages
= max_images
,
1319 .maxPerStageDescriptorInputAttachments
= MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS
,
1320 .maxPerStageResources
= max_per_stage
,
1321 .maxDescriptorSetSamplers
= 6 * max_samplers
, /* number of stages * maxPerStageDescriptorSamplers */
1322 .maxDescriptorSetUniformBuffers
= 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
, /* number of stages * maxPerStageDescriptorUniformBuffers */
1323 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
1324 .maxDescriptorSetStorageBuffers
= 6 * max_ssbos
, /* number of stages * maxPerStageDescriptorStorageBuffers */
1325 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
1326 .maxDescriptorSetSampledImages
= 6 * max_textures
, /* number of stages * maxPerStageDescriptorSampledImages */
1327 .maxDescriptorSetStorageImages
= 6 * max_images
, /* number of stages * maxPerStageDescriptorStorageImages */
1328 .maxDescriptorSetInputAttachments
= MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS
,
1329 .maxVertexInputAttributes
= MAX_VBS
,
1330 .maxVertexInputBindings
= MAX_VBS
,
1331 .maxVertexInputAttributeOffset
= 2047,
1332 .maxVertexInputBindingStride
= 2048,
1333 .maxVertexOutputComponents
= 128,
1334 .maxTessellationGenerationLevel
= 64,
1335 .maxTessellationPatchSize
= 32,
1336 .maxTessellationControlPerVertexInputComponents
= 128,
1337 .maxTessellationControlPerVertexOutputComponents
= 128,
1338 .maxTessellationControlPerPatchOutputComponents
= 128,
1339 .maxTessellationControlTotalOutputComponents
= 2048,
1340 .maxTessellationEvaluationInputComponents
= 128,
1341 .maxTessellationEvaluationOutputComponents
= 128,
1342 .maxGeometryShaderInvocations
= 32,
1343 .maxGeometryInputComponents
= 64,
1344 .maxGeometryOutputComponents
= 128,
1345 .maxGeometryOutputVertices
= 256,
1346 .maxGeometryTotalOutputComponents
= 1024,
1347 .maxFragmentInputComponents
= 116, /* 128 components - (PSIZ, CLIP_DIST0, CLIP_DIST1) */
1348 .maxFragmentOutputAttachments
= 8,
1349 .maxFragmentDualSrcAttachments
= 1,
1350 .maxFragmentCombinedOutputResources
= 8,
1351 .maxComputeSharedMemorySize
= 64 * 1024,
1352 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
1353 .maxComputeWorkGroupInvocations
= max_workgroup_size
,
1354 .maxComputeWorkGroupSize
= {
1359 .subPixelPrecisionBits
= 8,
1360 .subTexelPrecisionBits
= 8,
1361 .mipmapPrecisionBits
= 8,
1362 .maxDrawIndexedIndexValue
= UINT32_MAX
,
1363 .maxDrawIndirectCount
= UINT32_MAX
,
1364 .maxSamplerLodBias
= 16,
1365 .maxSamplerAnisotropy
= 16,
1366 .maxViewports
= MAX_VIEWPORTS
,
1367 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
1368 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
1369 .viewportSubPixelBits
= 13, /* We take a float? */
1370 .minMemoryMapAlignment
= 4096, /* A page */
1371 /* The dataport requires texel alignment so we need to assume a worst
1372 * case of R32G32B32A32 which is 16 bytes.
1374 .minTexelBufferOffsetAlignment
= 16,
1375 /* We need 16 for UBO block reads to work and 32 for push UBOs */
1376 .minUniformBufferOffsetAlignment
= 32,
1377 .minStorageBufferOffsetAlignment
= 4,
1378 .minTexelOffset
= -8,
1379 .maxTexelOffset
= 7,
1380 .minTexelGatherOffset
= -32,
1381 .maxTexelGatherOffset
= 31,
1382 .minInterpolationOffset
= -0.5,
1383 .maxInterpolationOffset
= 0.4375,
1384 .subPixelInterpolationOffsetBits
= 4,
1385 .maxFramebufferWidth
= (1 << 14),
1386 .maxFramebufferHeight
= (1 << 14),
1387 .maxFramebufferLayers
= (1 << 11),
1388 .framebufferColorSampleCounts
= sample_counts
,
1389 .framebufferDepthSampleCounts
= sample_counts
,
1390 .framebufferStencilSampleCounts
= sample_counts
,
1391 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
1392 .maxColorAttachments
= MAX_RTS
,
1393 .sampledImageColorSampleCounts
= sample_counts
,
1394 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
1395 .sampledImageDepthSampleCounts
= sample_counts
,
1396 .sampledImageStencilSampleCounts
= sample_counts
,
1397 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
1398 .maxSampleMaskWords
= 1,
1399 .timestampComputeAndGraphics
= true,
1400 .timestampPeriod
= 1000000000.0 / devinfo
->timestamp_frequency
,
1401 .maxClipDistances
= 8,
1402 .maxCullDistances
= 8,
1403 .maxCombinedClipAndCullDistances
= 8,
1404 .discreteQueuePriorities
= 2,
1405 .pointSizeRange
= { 0.125, 255.875 },
1408 (devinfo
->gen
>= 9 || devinfo
->is_cherryview
) ?
1409 2047.9921875 : 7.9921875,
1411 .pointSizeGranularity
= (1.0 / 8.0),
1412 .lineWidthGranularity
= (1.0 / 128.0),
1413 .strictLines
= false,
1414 .standardSampleLocations
= true,
1415 .optimalBufferCopyOffsetAlignment
= 128,
1416 .optimalBufferCopyRowPitchAlignment
= 128,
1417 .nonCoherentAtomSize
= 64,
1420 *pProperties
= (VkPhysicalDeviceProperties
) {
1421 .apiVersion
= anv_physical_device_api_version(pdevice
),
1422 .driverVersion
= vk_get_driver_version(),
1424 .deviceID
= pdevice
->chipset_id
,
1425 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
1427 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
1430 snprintf(pProperties
->deviceName
, sizeof(pProperties
->deviceName
),
1431 "%s", pdevice
->name
);
1432 memcpy(pProperties
->pipelineCacheUUID
,
1433 pdevice
->pipeline_cache_uuid
, VK_UUID_SIZE
);
1436 void anv_GetPhysicalDeviceProperties2(
1437 VkPhysicalDevice physicalDevice
,
1438 VkPhysicalDeviceProperties2
* pProperties
)
1440 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1442 anv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
1444 vk_foreach_struct(ext
, pProperties
->pNext
) {
1445 switch (ext
->sType
) {
1446 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_STENCIL_RESOLVE_PROPERTIES_KHR
: {
1447 VkPhysicalDeviceDepthStencilResolvePropertiesKHR
*props
=
1448 (VkPhysicalDeviceDepthStencilResolvePropertiesKHR
*)ext
;
1450 /* We support all of the depth resolve modes */
1451 props
->supportedDepthResolveModes
=
1452 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
|
1453 VK_RESOLVE_MODE_AVERAGE_BIT_KHR
|
1454 VK_RESOLVE_MODE_MIN_BIT_KHR
|
1455 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1457 /* Average doesn't make sense for stencil so we don't support that */
1458 props
->supportedStencilResolveModes
=
1459 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
;
1460 if (pdevice
->info
.gen
>= 8) {
1461 /* The advanced stencil resolve modes currently require stencil
1462 * sampling be supported by the hardware.
1464 props
->supportedStencilResolveModes
|=
1465 VK_RESOLVE_MODE_MIN_BIT_KHR
|
1466 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1469 props
->independentResolveNone
= true;
1470 props
->independentResolve
= true;
1474 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES_EXT
: {
1475 VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*props
=
1476 (VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*)ext
;
1478 /* It's a bit hard to exactly map our implementation to the limits
1479 * described here. The bindless surface handle in the extended
1480 * message descriptors is 20 bits and it's an index into the table of
1481 * RENDER_SURFACE_STATE structs that starts at bindless surface base
1482 * address. Given that most things consume two surface states per
1483 * view (general/sampled for textures and write-only/read-write for
1484 * images), we claim 2^19 things.
1486 * For SSBOs, we just use A64 messages so there is no real limit
1487 * there beyond the limit on the total size of a descriptor set.
1489 const unsigned max_bindless_views
= 1 << 19;
1491 props
->maxUpdateAfterBindDescriptorsInAllPools
= max_bindless_views
;
1492 props
->shaderUniformBufferArrayNonUniformIndexingNative
= false;
1493 props
->shaderSampledImageArrayNonUniformIndexingNative
= false;
1494 props
->shaderStorageBufferArrayNonUniformIndexingNative
= true;
1495 props
->shaderStorageImageArrayNonUniformIndexingNative
= false;
1496 props
->shaderInputAttachmentArrayNonUniformIndexingNative
= false;
1497 props
->robustBufferAccessUpdateAfterBind
= true;
1498 props
->quadDivergentImplicitLod
= false;
1499 props
->maxPerStageDescriptorUpdateAfterBindSamplers
= max_bindless_views
;
1500 props
->maxPerStageDescriptorUpdateAfterBindUniformBuffers
= MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
;
1501 props
->maxPerStageDescriptorUpdateAfterBindStorageBuffers
= UINT32_MAX
;
1502 props
->maxPerStageDescriptorUpdateAfterBindSampledImages
= max_bindless_views
;
1503 props
->maxPerStageDescriptorUpdateAfterBindStorageImages
= max_bindless_views
;
1504 props
->maxPerStageDescriptorUpdateAfterBindInputAttachments
= MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS
;
1505 props
->maxPerStageUpdateAfterBindResources
= UINT32_MAX
;
1506 props
->maxDescriptorSetUpdateAfterBindSamplers
= max_bindless_views
;
1507 props
->maxDescriptorSetUpdateAfterBindUniformBuffers
= 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
;
1508 props
->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2;
1509 props
->maxDescriptorSetUpdateAfterBindStorageBuffers
= UINT32_MAX
;
1510 props
->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2;
1511 props
->maxDescriptorSetUpdateAfterBindSampledImages
= max_bindless_views
;
1512 props
->maxDescriptorSetUpdateAfterBindStorageImages
= max_bindless_views
;
1513 props
->maxDescriptorSetUpdateAfterBindInputAttachments
= MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS
;
1517 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR
: {
1518 VkPhysicalDeviceDriverPropertiesKHR
*driver_props
=
1519 (VkPhysicalDeviceDriverPropertiesKHR
*) ext
;
1521 driver_props
->driverID
= VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA_KHR
;
1522 snprintf(driver_props
->driverName
, VK_MAX_DRIVER_NAME_SIZE_KHR
,
1523 "Intel open-source Mesa driver");
1525 snprintf(driver_props
->driverInfo
, VK_MAX_DRIVER_INFO_SIZE_KHR
,
1526 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1
);
1528 driver_props
->conformanceVersion
= (VkConformanceVersionKHR
) {
1537 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT
: {
1538 VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*props
=
1539 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*) ext
;
1540 /* Userptr needs page aligned memory. */
1541 props
->minImportedHostPointerAlignment
= 4096;
1545 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES
: {
1546 VkPhysicalDeviceIDProperties
*id_props
=
1547 (VkPhysicalDeviceIDProperties
*)ext
;
1548 memcpy(id_props
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
1549 memcpy(id_props
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
1550 /* The LUID is for Windows. */
1551 id_props
->deviceLUIDValid
= false;
1555 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_PROPERTIES_EXT
: {
1556 VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*props
=
1557 (VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*)ext
;
1558 props
->maxInlineUniformBlockSize
= MAX_INLINE_UNIFORM_BLOCK_SIZE
;
1559 props
->maxPerStageDescriptorInlineUniformBlocks
=
1560 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1561 props
->maxPerStageDescriptorUpdateAfterBindInlineUniformBlocks
=
1562 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1563 props
->maxDescriptorSetInlineUniformBlocks
=
1564 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1565 props
->maxDescriptorSetUpdateAfterBindInlineUniformBlocks
=
1566 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1570 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_PROPERTIES_EXT
: {
1571 VkPhysicalDeviceLineRasterizationPropertiesEXT
*props
=
1572 (VkPhysicalDeviceLineRasterizationPropertiesEXT
*)ext
;
1573 /* In the Skylake PRM Vol. 7, subsection titled "GIQ (Diamond)
1574 * Sampling Rules - Legacy Mode", it says the following:
1576 * "Note that the device divides a pixel into a 16x16 array of
1577 * subpixels, referenced by their upper left corners."
1579 * This is the only known reference in the PRMs to the subpixel
1580 * precision of line rasterization and a "16x16 array of subpixels"
1581 * implies 4 subpixel precision bits. Empirical testing has shown
1582 * that 4 subpixel precision bits applies to all line rasterization
1585 props
->lineSubPixelPrecisionBits
= 4;
1589 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES
: {
1590 VkPhysicalDeviceMaintenance3Properties
*props
=
1591 (VkPhysicalDeviceMaintenance3Properties
*)ext
;
1592 /* This value doesn't matter for us today as our per-stage
1593 * descriptors are the real limit.
1595 props
->maxPerSetDescriptors
= 1024;
1596 props
->maxMemoryAllocationSize
= MAX_MEMORY_ALLOCATION_SIZE
;
1600 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES
: {
1601 VkPhysicalDeviceMultiviewProperties
*properties
=
1602 (VkPhysicalDeviceMultiviewProperties
*)ext
;
1603 properties
->maxMultiviewViewCount
= 16;
1604 properties
->maxMultiviewInstanceIndex
= UINT32_MAX
/ 16;
1608 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT
: {
1609 VkPhysicalDevicePCIBusInfoPropertiesEXT
*properties
=
1610 (VkPhysicalDevicePCIBusInfoPropertiesEXT
*)ext
;
1611 properties
->pciDomain
= pdevice
->pci_info
.domain
;
1612 properties
->pciBus
= pdevice
->pci_info
.bus
;
1613 properties
->pciDevice
= pdevice
->pci_info
.device
;
1614 properties
->pciFunction
= pdevice
->pci_info
.function
;
1618 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES
: {
1619 VkPhysicalDevicePointClippingProperties
*properties
=
1620 (VkPhysicalDevicePointClippingProperties
*) ext
;
1621 properties
->pointClippingBehavior
= VK_POINT_CLIPPING_BEHAVIOR_USER_CLIP_PLANES_ONLY
;
1625 #pragma GCC diagnostic push
1626 #pragma GCC diagnostic ignored "-Wswitch"
1627 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PRESENTATION_PROPERTIES_ANDROID
: {
1628 VkPhysicalDevicePresentationPropertiesANDROID
*props
=
1629 (VkPhysicalDevicePresentationPropertiesANDROID
*)ext
;
1630 props
->sharedImage
= VK_FALSE
;
1633 #pragma GCC diagnostic pop
1635 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES
: {
1636 VkPhysicalDeviceProtectedMemoryProperties
*props
=
1637 (VkPhysicalDeviceProtectedMemoryProperties
*)ext
;
1638 props
->protectedNoFault
= false;
1642 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
1643 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
1644 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
1646 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
1650 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT
: {
1651 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*properties
=
1652 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*)ext
;
1653 properties
->filterMinmaxImageComponentMapping
= pdevice
->info
.gen
>= 9;
1654 properties
->filterMinmaxSingleComponentFormats
= true;
1658 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES
: {
1659 VkPhysicalDeviceSubgroupProperties
*properties
= (void *)ext
;
1661 properties
->subgroupSize
= BRW_SUBGROUP_SIZE
;
1663 VkShaderStageFlags scalar_stages
= 0;
1664 for (unsigned stage
= 0; stage
< MESA_SHADER_STAGES
; stage
++) {
1665 if (pdevice
->compiler
->scalar_stage
[stage
])
1666 scalar_stages
|= mesa_to_vk_shader_stage(stage
);
1668 properties
->supportedStages
= scalar_stages
;
1670 properties
->supportedOperations
= VK_SUBGROUP_FEATURE_BASIC_BIT
|
1671 VK_SUBGROUP_FEATURE_VOTE_BIT
|
1672 VK_SUBGROUP_FEATURE_BALLOT_BIT
|
1673 VK_SUBGROUP_FEATURE_SHUFFLE_BIT
|
1674 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT
|
1675 VK_SUBGROUP_FEATURE_QUAD_BIT
;
1676 if (pdevice
->info
.gen
>= 8) {
1677 /* TODO: There's no technical reason why these can't be made to
1678 * work on gen7 but they don't at the moment so it's best to leave
1679 * the feature disabled than enabled and broken.
1681 properties
->supportedOperations
|=
1682 VK_SUBGROUP_FEATURE_ARITHMETIC_BIT
|
1683 VK_SUBGROUP_FEATURE_CLUSTERED_BIT
;
1685 properties
->quadOperationsInAllStages
= pdevice
->info
.gen
>= 8;
1689 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_PROPERTIES_EXT
: {
1690 VkPhysicalDeviceSubgroupSizeControlPropertiesEXT
*props
=
1691 (VkPhysicalDeviceSubgroupSizeControlPropertiesEXT
*)ext
;
1692 STATIC_ASSERT(8 <= BRW_SUBGROUP_SIZE
&& BRW_SUBGROUP_SIZE
<= 32);
1693 props
->minSubgroupSize
= 8;
1694 props
->maxSubgroupSize
= 32;
1695 props
->maxComputeWorkgroupSubgroups
= pdevice
->info
.max_cs_threads
;
1696 props
->requiredSubgroupSizeStages
= VK_SHADER_STAGE_COMPUTE_BIT
;
1699 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT_CONTROLS_PROPERTIES_KHR
: {
1700 VkPhysicalDeviceFloatControlsPropertiesKHR
*properties
= (void *)ext
;
1701 properties
->denormBehaviorIndependence
= VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_ALL_KHR
;
1702 properties
->roundingModeIndependence
= VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_NONE_KHR
;
1704 /* Broadwell does not support HF denorms and there are restrictions
1705 * other gens. According to Kabylake's PRM:
1707 * "math - Extended Math Function
1709 * Restriction : Half-float denorms are always retained."
1711 properties
->shaderDenormFlushToZeroFloat16
= false;
1712 properties
->shaderDenormPreserveFloat16
= pdevice
->info
.gen
> 8;
1713 properties
->shaderRoundingModeRTEFloat16
= true;
1714 properties
->shaderRoundingModeRTZFloat16
= true;
1715 properties
->shaderSignedZeroInfNanPreserveFloat16
= true;
1717 properties
->shaderDenormFlushToZeroFloat32
= true;
1718 properties
->shaderDenormPreserveFloat32
= true;
1719 properties
->shaderRoundingModeRTEFloat32
= true;
1720 properties
->shaderRoundingModeRTZFloat32
= true;
1721 properties
->shaderSignedZeroInfNanPreserveFloat32
= true;
1723 properties
->shaderDenormFlushToZeroFloat64
= true;
1724 properties
->shaderDenormPreserveFloat64
= true;
1725 properties
->shaderRoundingModeRTEFloat64
= true;
1726 properties
->shaderRoundingModeRTZFloat64
= true;
1727 properties
->shaderSignedZeroInfNanPreserveFloat64
= true;
1731 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_PROPERTIES_EXT
: {
1732 VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT
*props
=
1733 (VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT
*)ext
;
1735 /* From the SKL PRM Vol. 2d, docs for RENDER_SURFACE_STATE::Surface
1738 * "For SURFTYPE_BUFFER non-rendertarget surfaces, this field
1739 * specifies the base address of the first element of the surface,
1740 * computed in software by adding the surface base address to the
1741 * byte offset of the element in the buffer. The base address must
1742 * be aligned to element size."
1744 * The typed dataport messages require that things be texel aligned.
1745 * Otherwise, we may just load/store the wrong data or, in the worst
1746 * case, there may be hangs.
1748 props
->storageTexelBufferOffsetAlignmentBytes
= 16;
1749 props
->storageTexelBufferOffsetSingleTexelAlignment
= true;
1751 /* The sampler, however, is much more forgiving and it can handle
1752 * arbitrary byte alignment for linear and buffer surfaces. It's
1753 * hard to find a good PRM citation for this but years of empirical
1754 * experience demonstrate that this is true.
1756 props
->uniformTexelBufferOffsetAlignmentBytes
= 1;
1757 props
->uniformTexelBufferOffsetSingleTexelAlignment
= false;
1761 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_PROPERTIES_KHR
: {
1762 VkPhysicalDeviceTimelineSemaphorePropertiesKHR
*props
=
1763 (VkPhysicalDeviceTimelineSemaphorePropertiesKHR
*) ext
;
1764 props
->maxTimelineSemaphoreValueDifference
= UINT64_MAX
;
1768 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT
: {
1769 VkPhysicalDeviceTransformFeedbackPropertiesEXT
*props
=
1770 (VkPhysicalDeviceTransformFeedbackPropertiesEXT
*)ext
;
1772 props
->maxTransformFeedbackStreams
= MAX_XFB_STREAMS
;
1773 props
->maxTransformFeedbackBuffers
= MAX_XFB_BUFFERS
;
1774 props
->maxTransformFeedbackBufferSize
= (1ull << 32);
1775 props
->maxTransformFeedbackStreamDataSize
= 128 * 4;
1776 props
->maxTransformFeedbackBufferDataSize
= 128 * 4;
1777 props
->maxTransformFeedbackBufferDataStride
= 2048;
1778 props
->transformFeedbackQueries
= true;
1779 props
->transformFeedbackStreamsLinesTriangles
= false;
1780 props
->transformFeedbackRasterizationStreamSelect
= false;
1781 props
->transformFeedbackDraw
= true;
1785 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT
: {
1786 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*props
=
1787 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*)ext
;
1788 /* We have to restrict this a bit for multiview */
1789 props
->maxVertexAttribDivisor
= UINT32_MAX
/ 16;
1794 anv_debug_ignored_stype(ext
->sType
);
1800 /* We support exactly one queue family. */
1801 static const VkQueueFamilyProperties
1802 anv_queue_family_properties
= {
1803 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
1804 VK_QUEUE_COMPUTE_BIT
|
1805 VK_QUEUE_TRANSFER_BIT
,
1807 .timestampValidBits
= 36, /* XXX: Real value here */
1808 .minImageTransferGranularity
= { 1, 1, 1 },
1811 void anv_GetPhysicalDeviceQueueFamilyProperties(
1812 VkPhysicalDevice physicalDevice
,
1814 VkQueueFamilyProperties
* pQueueFamilyProperties
)
1816 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pCount
);
1818 vk_outarray_append(&out
, p
) {
1819 *p
= anv_queue_family_properties
;
1823 void anv_GetPhysicalDeviceQueueFamilyProperties2(
1824 VkPhysicalDevice physicalDevice
,
1825 uint32_t* pQueueFamilyPropertyCount
,
1826 VkQueueFamilyProperties2
* pQueueFamilyProperties
)
1829 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pQueueFamilyPropertyCount
);
1831 vk_outarray_append(&out
, p
) {
1832 p
->queueFamilyProperties
= anv_queue_family_properties
;
1834 vk_foreach_struct(s
, p
->pNext
) {
1835 anv_debug_ignored_stype(s
->sType
);
1840 void anv_GetPhysicalDeviceMemoryProperties(
1841 VkPhysicalDevice physicalDevice
,
1842 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
1844 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
1846 pMemoryProperties
->memoryTypeCount
= physical_device
->memory
.type_count
;
1847 for (uint32_t i
= 0; i
< physical_device
->memory
.type_count
; i
++) {
1848 pMemoryProperties
->memoryTypes
[i
] = (VkMemoryType
) {
1849 .propertyFlags
= physical_device
->memory
.types
[i
].propertyFlags
,
1850 .heapIndex
= physical_device
->memory
.types
[i
].heapIndex
,
1854 pMemoryProperties
->memoryHeapCount
= physical_device
->memory
.heap_count
;
1855 for (uint32_t i
= 0; i
< physical_device
->memory
.heap_count
; i
++) {
1856 pMemoryProperties
->memoryHeaps
[i
] = (VkMemoryHeap
) {
1857 .size
= physical_device
->memory
.heaps
[i
].size
,
1858 .flags
= physical_device
->memory
.heaps
[i
].flags
,
1864 anv_get_memory_budget(VkPhysicalDevice physicalDevice
,
1865 VkPhysicalDeviceMemoryBudgetPropertiesEXT
*memoryBudget
)
1867 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
1868 uint64_t sys_available
= get_available_system_memory();
1869 assert(sys_available
> 0);
1871 VkDeviceSize total_heaps_size
= 0;
1872 for (size_t i
= 0; i
< device
->memory
.heap_count
; i
++)
1873 total_heaps_size
+= device
->memory
.heaps
[i
].size
;
1875 for (size_t i
= 0; i
< device
->memory
.heap_count
; i
++) {
1876 VkDeviceSize heap_size
= device
->memory
.heaps
[i
].size
;
1877 VkDeviceSize heap_used
= device
->memory
.heaps
[i
].used
;
1878 VkDeviceSize heap_budget
;
1880 double heap_proportion
= (double) heap_size
/ total_heaps_size
;
1881 VkDeviceSize sys_available_prop
= sys_available
* heap_proportion
;
1884 * Let's not incite the app to starve the system: report at most 90% of
1885 * available system memory.
1887 uint64_t heap_available
= sys_available_prop
* 9 / 10;
1888 heap_budget
= MIN2(heap_size
, heap_used
+ heap_available
);
1891 * Round down to the nearest MB
1893 heap_budget
&= ~((1ull << 20) - 1);
1896 * The heapBudget value must be non-zero for array elements less than
1897 * VkPhysicalDeviceMemoryProperties::memoryHeapCount. The heapBudget
1898 * value must be less than or equal to VkMemoryHeap::size for each heap.
1900 assert(0 < heap_budget
&& heap_budget
<= heap_size
);
1902 memoryBudget
->heapUsage
[i
] = heap_used
;
1903 memoryBudget
->heapBudget
[i
] = heap_budget
;
1906 /* The heapBudget and heapUsage values must be zero for array elements
1907 * greater than or equal to VkPhysicalDeviceMemoryProperties::memoryHeapCount
1909 for (uint32_t i
= device
->memory
.heap_count
; i
< VK_MAX_MEMORY_HEAPS
; i
++) {
1910 memoryBudget
->heapBudget
[i
] = 0;
1911 memoryBudget
->heapUsage
[i
] = 0;
1915 void anv_GetPhysicalDeviceMemoryProperties2(
1916 VkPhysicalDevice physicalDevice
,
1917 VkPhysicalDeviceMemoryProperties2
* pMemoryProperties
)
1919 anv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
1920 &pMemoryProperties
->memoryProperties
);
1922 vk_foreach_struct(ext
, pMemoryProperties
->pNext
) {
1923 switch (ext
->sType
) {
1924 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT
:
1925 anv_get_memory_budget(physicalDevice
, (void*)ext
);
1928 anv_debug_ignored_stype(ext
->sType
);
1935 anv_GetDeviceGroupPeerMemoryFeatures(
1938 uint32_t localDeviceIndex
,
1939 uint32_t remoteDeviceIndex
,
1940 VkPeerMemoryFeatureFlags
* pPeerMemoryFeatures
)
1942 assert(localDeviceIndex
== 0 && remoteDeviceIndex
== 0);
1943 *pPeerMemoryFeatures
= VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT
|
1944 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT
|
1945 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT
|
1946 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT
;
1949 PFN_vkVoidFunction
anv_GetInstanceProcAddr(
1950 VkInstance _instance
,
1953 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1955 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
1956 * when we have to return valid function pointers, NULL, or it's left
1957 * undefined. See the table for exact details.
1962 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
1963 if (strcmp(pName, "vk" #entrypoint) == 0) \
1964 return (PFN_vkVoidFunction)anv_##entrypoint
1966 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties
);
1967 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties
);
1968 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceVersion
);
1969 LOOKUP_ANV_ENTRYPOINT(CreateInstance
);
1971 #undef LOOKUP_ANV_ENTRYPOINT
1973 if (instance
== NULL
)
1976 int idx
= anv_get_instance_entrypoint_index(pName
);
1978 return instance
->dispatch
.entrypoints
[idx
];
1980 idx
= anv_get_physical_device_entrypoint_index(pName
);
1982 return instance
->physicalDevice
.dispatch
.entrypoints
[idx
];
1984 idx
= anv_get_device_entrypoint_index(pName
);
1986 return instance
->device_dispatch
.entrypoints
[idx
];
1991 /* With version 1+ of the loader interface the ICD should expose
1992 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
1995 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1996 VkInstance instance
,
2000 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
2001 VkInstance instance
,
2004 return anv_GetInstanceProcAddr(instance
, pName
);
2007 PFN_vkVoidFunction
anv_GetDeviceProcAddr(
2011 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2013 if (!device
|| !pName
)
2016 int idx
= anv_get_device_entrypoint_index(pName
);
2020 return device
->dispatch
.entrypoints
[idx
];
2023 /* With version 4+ of the loader interface the ICD should expose
2024 * vk_icdGetPhysicalDeviceProcAddr()
2027 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetPhysicalDeviceProcAddr(
2028 VkInstance _instance
,
2031 PFN_vkVoidFunction
vk_icdGetPhysicalDeviceProcAddr(
2032 VkInstance _instance
,
2035 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2037 if (!pName
|| !instance
)
2040 int idx
= anv_get_physical_device_entrypoint_index(pName
);
2044 return instance
->physicalDevice
.dispatch
.entrypoints
[idx
];
2049 anv_CreateDebugReportCallbackEXT(VkInstance _instance
,
2050 const VkDebugReportCallbackCreateInfoEXT
* pCreateInfo
,
2051 const VkAllocationCallbacks
* pAllocator
,
2052 VkDebugReportCallbackEXT
* pCallback
)
2054 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2055 return vk_create_debug_report_callback(&instance
->debug_report_callbacks
,
2056 pCreateInfo
, pAllocator
, &instance
->alloc
,
2061 anv_DestroyDebugReportCallbackEXT(VkInstance _instance
,
2062 VkDebugReportCallbackEXT _callback
,
2063 const VkAllocationCallbacks
* pAllocator
)
2065 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2066 vk_destroy_debug_report_callback(&instance
->debug_report_callbacks
,
2067 _callback
, pAllocator
, &instance
->alloc
);
2071 anv_DebugReportMessageEXT(VkInstance _instance
,
2072 VkDebugReportFlagsEXT flags
,
2073 VkDebugReportObjectTypeEXT objectType
,
2076 int32_t messageCode
,
2077 const char* pLayerPrefix
,
2078 const char* pMessage
)
2080 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2081 vk_debug_report(&instance
->debug_report_callbacks
, flags
, objectType
,
2082 object
, location
, messageCode
, pLayerPrefix
, pMessage
);
2085 static struct anv_state
2086 anv_state_pool_emit_data(struct anv_state_pool
*pool
, size_t size
, size_t align
, const void *p
)
2088 struct anv_state state
;
2090 state
= anv_state_pool_alloc(pool
, size
, align
);
2091 memcpy(state
.map
, p
, size
);
2096 /* Haswell border color is a bit of a disaster. Float and unorm formats use a
2097 * straightforward 32-bit float color in the first 64 bytes. Instead of using
2098 * a nice float/integer union like Gen8+, Haswell specifies the integer border
2099 * color as a separate entry /after/ the float color. The layout of this entry
2100 * also depends on the format's bpp (with extra hacks for RG32), and overlaps.
2102 * Since we don't know the format/bpp, we can't make any of the border colors
2103 * containing '1' work for all formats, as it would be in the wrong place for
2104 * some of them. We opt to make 32-bit integers work as this seems like the
2105 * most common option. Fortunately, transparent black works regardless, as
2106 * all zeroes is the same in every bit-size.
2108 struct hsw_border_color
{
2112 uint32_t _pad1
[108];
2115 struct gen8_border_color
{
2120 /* Pad out to 64 bytes */
2125 anv_device_init_border_colors(struct anv_device
*device
)
2127 if (device
->info
.is_haswell
) {
2128 static const struct hsw_border_color border_colors
[] = {
2129 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
2130 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
2131 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
2132 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
2133 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
2134 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
2137 device
->border_colors
=
2138 anv_state_pool_emit_data(&device
->dynamic_state_pool
,
2139 sizeof(border_colors
), 512, border_colors
);
2141 static const struct gen8_border_color border_colors
[] = {
2142 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
2143 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
2144 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
2145 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
2146 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
2147 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
2150 device
->border_colors
=
2151 anv_state_pool_emit_data(&device
->dynamic_state_pool
,
2152 sizeof(border_colors
), 64, border_colors
);
2157 anv_device_init_trivial_batch(struct anv_device
*device
)
2159 VkResult result
= anv_device_alloc_bo(device
, 4096,
2160 ANV_BO_ALLOC_MAPPED
,
2161 &device
->trivial_batch_bo
);
2162 if (result
!= VK_SUCCESS
)
2165 struct anv_batch batch
= {
2166 .start
= device
->trivial_batch_bo
->map
,
2167 .next
= device
->trivial_batch_bo
->map
,
2168 .end
= device
->trivial_batch_bo
->map
+ 4096,
2171 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
2172 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
2174 if (!device
->info
.has_llc
)
2175 gen_clflush_range(batch
.start
, batch
.next
- batch
.start
);
2180 VkResult
anv_EnumerateDeviceExtensionProperties(
2181 VkPhysicalDevice physicalDevice
,
2182 const char* pLayerName
,
2183 uint32_t* pPropertyCount
,
2184 VkExtensionProperties
* pProperties
)
2186 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
2187 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
2189 for (int i
= 0; i
< ANV_DEVICE_EXTENSION_COUNT
; i
++) {
2190 if (device
->supported_extensions
.extensions
[i
]) {
2191 vk_outarray_append(&out
, prop
) {
2192 *prop
= anv_device_extensions
[i
];
2197 return vk_outarray_status(&out
);
2201 anv_device_init_dispatch(struct anv_device
*device
)
2203 const struct anv_device_dispatch_table
*genX_table
;
2204 switch (device
->info
.gen
) {
2206 genX_table
= &gen12_device_dispatch_table
;
2209 genX_table
= &gen11_device_dispatch_table
;
2212 genX_table
= &gen10_device_dispatch_table
;
2215 genX_table
= &gen9_device_dispatch_table
;
2218 genX_table
= &gen8_device_dispatch_table
;
2221 if (device
->info
.is_haswell
)
2222 genX_table
= &gen75_device_dispatch_table
;
2224 genX_table
= &gen7_device_dispatch_table
;
2227 unreachable("unsupported gen\n");
2230 for (unsigned i
= 0; i
< ARRAY_SIZE(device
->dispatch
.entrypoints
); i
++) {
2231 /* Vulkan requires that entrypoints for extensions which have not been
2232 * enabled must not be advertised.
2234 if (!anv_device_entrypoint_is_enabled(i
, device
->instance
->app_info
.api_version
,
2235 &device
->instance
->enabled_extensions
,
2236 &device
->enabled_extensions
)) {
2237 device
->dispatch
.entrypoints
[i
] = NULL
;
2238 } else if (genX_table
->entrypoints
[i
]) {
2239 device
->dispatch
.entrypoints
[i
] = genX_table
->entrypoints
[i
];
2241 device
->dispatch
.entrypoints
[i
] =
2242 anv_device_dispatch_table
.entrypoints
[i
];
2248 vk_priority_to_gen(int priority
)
2251 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT
:
2252 return GEN_CONTEXT_LOW_PRIORITY
;
2253 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
:
2254 return GEN_CONTEXT_MEDIUM_PRIORITY
;
2255 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT
:
2256 return GEN_CONTEXT_HIGH_PRIORITY
;
2257 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT
:
2258 return GEN_CONTEXT_REALTIME_PRIORITY
;
2260 unreachable("Invalid priority");
2265 anv_device_init_hiz_clear_value_bo(struct anv_device
*device
)
2267 VkResult result
= anv_device_alloc_bo(device
, 4096,
2268 ANV_BO_ALLOC_MAPPED
,
2269 &device
->hiz_clear_bo
);
2270 if (result
!= VK_SUCCESS
)
2273 union isl_color_value hiz_clear
= { .u32
= { 0, } };
2274 hiz_clear
.f32
[0] = ANV_HZ_FC_VAL
;
2276 memcpy(device
->hiz_clear_bo
->map
, hiz_clear
.u32
, sizeof(hiz_clear
.u32
));
2278 if (!device
->info
.has_llc
)
2279 gen_clflush_range(device
->hiz_clear_bo
->map
, sizeof(hiz_clear
.u32
));
2285 get_bo_from_pool(struct gen_batch_decode_bo
*ret
,
2286 struct anv_block_pool
*pool
,
2289 anv_block_pool_foreach_bo(bo
, pool
) {
2290 uint64_t bo_address
= gen_48b_address(bo
->offset
);
2291 if (address
>= bo_address
&& address
< (bo_address
+ bo
->size
)) {
2292 *ret
= (struct gen_batch_decode_bo
) {
2303 /* Finding a buffer for batch decoding */
2304 static struct gen_batch_decode_bo
2305 decode_get_bo(void *v_batch
, bool ppgtt
, uint64_t address
)
2307 struct anv_device
*device
= v_batch
;
2308 struct gen_batch_decode_bo ret_bo
= {};
2312 if (get_bo_from_pool(&ret_bo
, &device
->dynamic_state_pool
.block_pool
, address
))
2314 if (get_bo_from_pool(&ret_bo
, &device
->instruction_state_pool
.block_pool
, address
))
2316 if (get_bo_from_pool(&ret_bo
, &device
->binding_table_pool
.block_pool
, address
))
2318 if (get_bo_from_pool(&ret_bo
, &device
->surface_state_pool
.block_pool
, address
))
2321 if (!device
->cmd_buffer_being_decoded
)
2322 return (struct gen_batch_decode_bo
) { };
2324 struct anv_batch_bo
**bo
;
2326 u_vector_foreach(bo
, &device
->cmd_buffer_being_decoded
->seen_bbos
) {
2327 /* The decoder zeroes out the top 16 bits, so we need to as well */
2328 uint64_t bo_address
= (*bo
)->bo
->offset
& (~0ull >> 16);
2330 if (address
>= bo_address
&& address
< bo_address
+ (*bo
)->bo
->size
) {
2331 return (struct gen_batch_decode_bo
) {
2333 .size
= (*bo
)->bo
->size
,
2334 .map
= (*bo
)->bo
->map
,
2339 return (struct gen_batch_decode_bo
) { };
2342 struct gen_aux_map_buffer
{
2343 struct gen_buffer base
;
2344 struct anv_state state
;
2347 static struct gen_buffer
*
2348 gen_aux_map_buffer_alloc(void *driver_ctx
, uint32_t size
)
2350 struct gen_aux_map_buffer
*buf
= malloc(sizeof(struct gen_aux_map_buffer
));
2354 struct anv_device
*device
= (struct anv_device
*)driver_ctx
;
2355 assert(device
->instance
->physicalDevice
.supports_48bit_addresses
&&
2356 device
->instance
->physicalDevice
.use_softpin
);
2358 struct anv_state_pool
*pool
= &device
->dynamic_state_pool
;
2359 buf
->state
= anv_state_pool_alloc(pool
, size
, size
);
2361 buf
->base
.gpu
= pool
->block_pool
.bo
->offset
+ buf
->state
.offset
;
2362 buf
->base
.gpu_end
= buf
->base
.gpu
+ buf
->state
.alloc_size
;
2363 buf
->base
.map
= buf
->state
.map
;
2364 buf
->base
.driver_bo
= &buf
->state
;
2369 gen_aux_map_buffer_free(void *driver_ctx
, struct gen_buffer
*buffer
)
2371 struct gen_aux_map_buffer
*buf
= (struct gen_aux_map_buffer
*)buffer
;
2372 struct anv_device
*device
= (struct anv_device
*)driver_ctx
;
2373 struct anv_state_pool
*pool
= &device
->dynamic_state_pool
;
2374 anv_state_pool_free(pool
, buf
->state
);
2378 static struct gen_mapped_pinned_buffer_alloc aux_map_allocator
= {
2379 .alloc
= gen_aux_map_buffer_alloc
,
2380 .free
= gen_aux_map_buffer_free
,
2383 VkResult
anv_CreateDevice(
2384 VkPhysicalDevice physicalDevice
,
2385 const VkDeviceCreateInfo
* pCreateInfo
,
2386 const VkAllocationCallbacks
* pAllocator
,
2389 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
2391 struct anv_device
*device
;
2393 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO
);
2395 struct anv_device_extension_table enabled_extensions
= { };
2396 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
2398 for (idx
= 0; idx
< ANV_DEVICE_EXTENSION_COUNT
; idx
++) {
2399 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
2400 anv_device_extensions
[idx
].extensionName
) == 0)
2404 if (idx
>= ANV_DEVICE_EXTENSION_COUNT
)
2405 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
2407 if (!physical_device
->supported_extensions
.extensions
[idx
])
2408 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
2410 enabled_extensions
.extensions
[idx
] = true;
2413 /* Check enabled features */
2414 if (pCreateInfo
->pEnabledFeatures
) {
2415 VkPhysicalDeviceFeatures supported_features
;
2416 anv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
2417 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
2418 VkBool32
*enabled_feature
= (VkBool32
*)pCreateInfo
->pEnabledFeatures
;
2419 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
2420 for (uint32_t i
= 0; i
< num_features
; i
++) {
2421 if (enabled_feature
[i
] && !supported_feature
[i
])
2422 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
2426 /* Check requested queues and fail if we are requested to create any
2427 * queues with flags we don't support.
2429 assert(pCreateInfo
->queueCreateInfoCount
> 0);
2430 for (uint32_t i
= 0; i
< pCreateInfo
->queueCreateInfoCount
; i
++) {
2431 if (pCreateInfo
->pQueueCreateInfos
[i
].flags
!= 0)
2432 return vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2435 /* Check if client specified queue priority. */
2436 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*queue_priority
=
2437 vk_find_struct_const(pCreateInfo
->pQueueCreateInfos
[0].pNext
,
2438 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT
);
2440 VkQueueGlobalPriorityEXT priority
=
2441 queue_priority
? queue_priority
->globalPriority
:
2442 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
;
2444 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
2446 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
2448 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2450 if (INTEL_DEBUG
& DEBUG_BATCH
) {
2451 const unsigned decode_flags
=
2452 GEN_BATCH_DECODE_FULL
|
2453 ((INTEL_DEBUG
& DEBUG_COLOR
) ? GEN_BATCH_DECODE_IN_COLOR
: 0) |
2454 GEN_BATCH_DECODE_OFFSETS
|
2455 GEN_BATCH_DECODE_FLOATS
;
2457 gen_batch_decode_ctx_init(&device
->decoder_ctx
,
2458 &physical_device
->info
,
2459 stderr
, decode_flags
, NULL
,
2460 decode_get_bo
, NULL
, device
);
2463 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
2464 device
->instance
= physical_device
->instance
;
2465 device
->chipset_id
= physical_device
->chipset_id
;
2466 device
->no_hw
= physical_device
->no_hw
;
2467 device
->_lost
= false;
2470 device
->alloc
= *pAllocator
;
2472 device
->alloc
= physical_device
->instance
->alloc
;
2474 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
2475 device
->fd
= open(physical_device
->path
, O_RDWR
| O_CLOEXEC
);
2476 if (device
->fd
== -1) {
2477 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2481 device
->context_id
= anv_gem_create_context(device
);
2482 if (device
->context_id
== -1) {
2483 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2487 result
= anv_queue_init(device
, &device
->queue
);
2488 if (result
!= VK_SUCCESS
)
2489 goto fail_context_id
;
2491 if (physical_device
->use_softpin
) {
2492 if (pthread_mutex_init(&device
->vma_mutex
, NULL
) != 0) {
2493 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2497 /* keep the page with address zero out of the allocator */
2498 util_vma_heap_init(&device
->vma_lo
,
2499 LOW_HEAP_MIN_ADDRESS
, LOW_HEAP_SIZE
);
2501 /* Leave the last 4GiB out of the high vma range, so that no state
2502 * base address + size can overflow 48 bits. For more information see
2503 * the comment about Wa32bitGeneralStateOffset in anv_allocator.c
2505 util_vma_heap_init(&device
->vma_hi
, HIGH_HEAP_MIN_ADDRESS
,
2506 physical_device
->gtt_size
- (1ull << 32) -
2507 HIGH_HEAP_MIN_ADDRESS
);
2510 list_inithead(&device
->memory_objects
);
2512 /* As per spec, the driver implementation may deny requests to acquire
2513 * a priority above the default priority (MEDIUM) if the caller does not
2514 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
2517 if (physical_device
->has_context_priority
) {
2518 int err
= anv_gem_set_context_param(device
->fd
, device
->context_id
,
2519 I915_CONTEXT_PARAM_PRIORITY
,
2520 vk_priority_to_gen(priority
));
2521 if (err
!= 0 && priority
> VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
) {
2522 result
= vk_error(VK_ERROR_NOT_PERMITTED_EXT
);
2527 device
->info
= physical_device
->info
;
2528 device
->isl_dev
= physical_device
->isl_dev
;
2530 /* On Broadwell and later, we can use batch chaining to more efficiently
2531 * implement growing command buffers. Prior to Haswell, the kernel
2532 * command parser gets in the way and we have to fall back to growing
2535 device
->can_chain_batches
= device
->info
.gen
>= 8;
2537 device
->robust_buffer_access
= pCreateInfo
->pEnabledFeatures
&&
2538 pCreateInfo
->pEnabledFeatures
->robustBufferAccess
;
2539 device
->enabled_extensions
= enabled_extensions
;
2541 anv_device_init_dispatch(device
);
2543 if (pthread_mutex_init(&device
->mutex
, NULL
) != 0) {
2544 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2548 pthread_condattr_t condattr
;
2549 if (pthread_condattr_init(&condattr
) != 0) {
2550 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2553 if (pthread_condattr_setclock(&condattr
, CLOCK_MONOTONIC
) != 0) {
2554 pthread_condattr_destroy(&condattr
);
2555 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2558 if (pthread_cond_init(&device
->queue_submit
, &condattr
) != 0) {
2559 pthread_condattr_destroy(&condattr
);
2560 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2563 pthread_condattr_destroy(&condattr
);
2565 result
= anv_bo_cache_init(&device
->bo_cache
);
2566 if (result
!= VK_SUCCESS
)
2567 goto fail_queue_cond
;
2569 anv_bo_pool_init(&device
->batch_bo_pool
, device
);
2571 result
= anv_state_pool_init(&device
->dynamic_state_pool
, device
,
2572 DYNAMIC_STATE_POOL_MIN_ADDRESS
, 16384);
2573 if (result
!= VK_SUCCESS
)
2574 goto fail_batch_bo_pool
;
2576 result
= anv_state_pool_init(&device
->instruction_state_pool
, device
,
2577 INSTRUCTION_STATE_POOL_MIN_ADDRESS
, 16384);
2578 if (result
!= VK_SUCCESS
)
2579 goto fail_dynamic_state_pool
;
2581 result
= anv_state_pool_init(&device
->surface_state_pool
, device
,
2582 SURFACE_STATE_POOL_MIN_ADDRESS
, 4096);
2583 if (result
!= VK_SUCCESS
)
2584 goto fail_instruction_state_pool
;
2586 if (physical_device
->use_softpin
) {
2587 result
= anv_state_pool_init(&device
->binding_table_pool
, device
,
2588 BINDING_TABLE_POOL_MIN_ADDRESS
, 4096);
2589 if (result
!= VK_SUCCESS
)
2590 goto fail_surface_state_pool
;
2593 if (device
->info
.gen
>= 12) {
2594 device
->aux_map_ctx
= gen_aux_map_init(device
, &aux_map_allocator
,
2595 &physical_device
->info
);
2596 if (!device
->aux_map_ctx
)
2597 goto fail_binding_table_pool
;
2600 result
= anv_device_alloc_bo(device
, 4096, 0, &device
->workaround_bo
);
2601 if (result
!= VK_SUCCESS
)
2602 goto fail_surface_aux_map_pool
;
2604 result
= anv_device_init_trivial_batch(device
);
2605 if (result
!= VK_SUCCESS
)
2606 goto fail_workaround_bo
;
2608 if (device
->info
.gen
>= 10) {
2609 result
= anv_device_init_hiz_clear_value_bo(device
);
2610 if (result
!= VK_SUCCESS
)
2611 goto fail_trivial_batch_bo
;
2614 anv_scratch_pool_init(device
, &device
->scratch_pool
);
2616 switch (device
->info
.gen
) {
2618 if (!device
->info
.is_haswell
)
2619 result
= gen7_init_device_state(device
);
2621 result
= gen75_init_device_state(device
);
2624 result
= gen8_init_device_state(device
);
2627 result
= gen9_init_device_state(device
);
2630 result
= gen10_init_device_state(device
);
2633 result
= gen11_init_device_state(device
);
2636 result
= gen12_init_device_state(device
);
2639 /* Shouldn't get here as we don't create physical devices for any other
2641 unreachable("unhandled gen");
2643 if (result
!= VK_SUCCESS
)
2644 goto fail_workaround_bo
;
2646 anv_pipeline_cache_init(&device
->default_pipeline_cache
, device
, true);
2648 anv_device_init_blorp(device
);
2650 anv_device_init_border_colors(device
);
2652 anv_device_perf_init(device
);
2654 *pDevice
= anv_device_to_handle(device
);
2659 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
2660 if (device
->info
.gen
>= 10)
2661 anv_device_release_bo(device
, device
->hiz_clear_bo
);
2662 anv_device_release_bo(device
, device
->workaround_bo
);
2663 fail_trivial_batch_bo
:
2664 anv_device_release_bo(device
, device
->trivial_batch_bo
);
2665 fail_surface_aux_map_pool
:
2666 if (device
->info
.gen
>= 12) {
2667 gen_aux_map_finish(device
->aux_map_ctx
);
2668 device
->aux_map_ctx
= NULL
;
2670 fail_binding_table_pool
:
2671 if (physical_device
->use_softpin
)
2672 anv_state_pool_finish(&device
->binding_table_pool
);
2673 fail_surface_state_pool
:
2674 anv_state_pool_finish(&device
->surface_state_pool
);
2675 fail_instruction_state_pool
:
2676 anv_state_pool_finish(&device
->instruction_state_pool
);
2677 fail_dynamic_state_pool
:
2678 anv_state_pool_finish(&device
->dynamic_state_pool
);
2680 anv_bo_pool_finish(&device
->batch_bo_pool
);
2681 anv_bo_cache_finish(&device
->bo_cache
);
2683 pthread_cond_destroy(&device
->queue_submit
);
2685 pthread_mutex_destroy(&device
->mutex
);
2687 if (physical_device
->use_softpin
) {
2688 util_vma_heap_finish(&device
->vma_hi
);
2689 util_vma_heap_finish(&device
->vma_lo
);
2692 anv_queue_finish(&device
->queue
);
2694 anv_gem_destroy_context(device
, device
->context_id
);
2698 vk_free(&device
->alloc
, device
);
2703 void anv_DestroyDevice(
2705 const VkAllocationCallbacks
* pAllocator
)
2707 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2708 struct anv_physical_device
*physical_device
;
2713 physical_device
= &device
->instance
->physicalDevice
;
2715 anv_device_finish_blorp(device
);
2717 anv_pipeline_cache_finish(&device
->default_pipeline_cache
);
2719 anv_queue_finish(&device
->queue
);
2721 #ifdef HAVE_VALGRIND
2722 /* We only need to free these to prevent valgrind errors. The backing
2723 * BO will go away in a couple of lines so we don't actually leak.
2725 anv_state_pool_free(&device
->dynamic_state_pool
, device
->border_colors
);
2726 anv_state_pool_free(&device
->dynamic_state_pool
, device
->slice_hash
);
2729 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
2731 anv_device_release_bo(device
, device
->workaround_bo
);
2732 anv_device_release_bo(device
, device
->trivial_batch_bo
);
2733 if (device
->info
.gen
>= 10)
2734 anv_device_release_bo(device
, device
->hiz_clear_bo
);
2736 if (device
->info
.gen
>= 12) {
2737 gen_aux_map_finish(device
->aux_map_ctx
);
2738 device
->aux_map_ctx
= NULL
;
2741 if (physical_device
->use_softpin
)
2742 anv_state_pool_finish(&device
->binding_table_pool
);
2743 anv_state_pool_finish(&device
->surface_state_pool
);
2744 anv_state_pool_finish(&device
->instruction_state_pool
);
2745 anv_state_pool_finish(&device
->dynamic_state_pool
);
2747 anv_bo_pool_finish(&device
->batch_bo_pool
);
2749 anv_bo_cache_finish(&device
->bo_cache
);
2751 if (physical_device
->use_softpin
) {
2752 util_vma_heap_finish(&device
->vma_hi
);
2753 util_vma_heap_finish(&device
->vma_lo
);
2756 pthread_cond_destroy(&device
->queue_submit
);
2757 pthread_mutex_destroy(&device
->mutex
);
2759 anv_gem_destroy_context(device
, device
->context_id
);
2761 if (INTEL_DEBUG
& DEBUG_BATCH
)
2762 gen_batch_decode_ctx_finish(&device
->decoder_ctx
);
2766 vk_free(&device
->alloc
, device
);
2769 VkResult
anv_EnumerateInstanceLayerProperties(
2770 uint32_t* pPropertyCount
,
2771 VkLayerProperties
* pProperties
)
2773 if (pProperties
== NULL
) {
2774 *pPropertyCount
= 0;
2778 /* None supported at this time */
2779 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
2782 VkResult
anv_EnumerateDeviceLayerProperties(
2783 VkPhysicalDevice physicalDevice
,
2784 uint32_t* pPropertyCount
,
2785 VkLayerProperties
* pProperties
)
2787 if (pProperties
== NULL
) {
2788 *pPropertyCount
= 0;
2792 /* None supported at this time */
2793 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
2796 void anv_GetDeviceQueue(
2798 uint32_t queueNodeIndex
,
2799 uint32_t queueIndex
,
2802 const VkDeviceQueueInfo2 info
= {
2803 .sType
= VK_STRUCTURE_TYPE_DEVICE_QUEUE_INFO_2
,
2806 .queueFamilyIndex
= queueNodeIndex
,
2807 .queueIndex
= queueIndex
,
2810 anv_GetDeviceQueue2(_device
, &info
, pQueue
);
2813 void anv_GetDeviceQueue2(
2815 const VkDeviceQueueInfo2
* pQueueInfo
,
2818 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2820 assert(pQueueInfo
->queueIndex
== 0);
2822 if (pQueueInfo
->flags
== device
->queue
.flags
)
2823 *pQueue
= anv_queue_to_handle(&device
->queue
);
2829 _anv_device_set_lost(struct anv_device
*device
,
2830 const char *file
, int line
,
2831 const char *msg
, ...)
2836 p_atomic_inc(&device
->_lost
);
2839 err
= __vk_errorv(device
->instance
, device
,
2840 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT
,
2841 VK_ERROR_DEVICE_LOST
, file
, line
, msg
, ap
);
2844 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
2851 _anv_queue_set_lost(struct anv_queue
*queue
,
2852 const char *file
, int line
,
2853 const char *msg
, ...)
2858 p_atomic_inc(&queue
->device
->_lost
);
2861 err
= __vk_errorv(queue
->device
->instance
, queue
->device
,
2862 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT
,
2863 VK_ERROR_DEVICE_LOST
, file
, line
, msg
, ap
);
2866 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
2873 anv_device_query_status(struct anv_device
*device
)
2875 /* This isn't likely as most of the callers of this function already check
2876 * for it. However, it doesn't hurt to check and it potentially lets us
2879 if (anv_device_is_lost(device
))
2880 return VK_ERROR_DEVICE_LOST
;
2882 uint32_t active
, pending
;
2883 int ret
= anv_gem_gpu_get_reset_stats(device
, &active
, &pending
);
2885 /* We don't know the real error. */
2886 return anv_device_set_lost(device
, "get_reset_stats failed: %m");
2890 return anv_device_set_lost(device
, "GPU hung on one of our command buffers");
2891 } else if (pending
) {
2892 return anv_device_set_lost(device
, "GPU hung with commands in-flight");
2899 anv_device_bo_busy(struct anv_device
*device
, struct anv_bo
*bo
)
2901 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
2902 * Other usages of the BO (such as on different hardware) will not be
2903 * flagged as "busy" by this ioctl. Use with care.
2905 int ret
= anv_gem_busy(device
, bo
->gem_handle
);
2907 return VK_NOT_READY
;
2908 } else if (ret
== -1) {
2909 /* We don't know the real error. */
2910 return anv_device_set_lost(device
, "gem wait failed: %m");
2913 /* Query for device status after the busy call. If the BO we're checking
2914 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
2915 * client because it clearly doesn't have valid data. Yes, this most
2916 * likely means an ioctl, but we just did an ioctl to query the busy status
2917 * so it's no great loss.
2919 return anv_device_query_status(device
);
2923 anv_device_wait(struct anv_device
*device
, struct anv_bo
*bo
,
2926 int ret
= anv_gem_wait(device
, bo
->gem_handle
, &timeout
);
2927 if (ret
== -1 && errno
== ETIME
) {
2929 } else if (ret
== -1) {
2930 /* We don't know the real error. */
2931 return anv_device_set_lost(device
, "gem wait failed: %m");
2934 /* Query for device status after the wait. If the BO we're waiting on got
2935 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
2936 * because it clearly doesn't have valid data. Yes, this most likely means
2937 * an ioctl, but we just did an ioctl to wait so it's no great loss.
2939 return anv_device_query_status(device
);
2942 VkResult
anv_DeviceWaitIdle(
2945 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2947 if (anv_device_is_lost(device
))
2948 return VK_ERROR_DEVICE_LOST
;
2950 return anv_queue_submit_simple_batch(&device
->queue
, NULL
);
2954 anv_vma_alloc(struct anv_device
*device
, struct anv_bo
*bo
)
2956 if (!(bo
->flags
& EXEC_OBJECT_PINNED
))
2959 pthread_mutex_lock(&device
->vma_mutex
);
2963 if (bo
->flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
) {
2964 uint64_t addr
= util_vma_heap_alloc(&device
->vma_hi
, bo
->size
, 4096);
2966 bo
->offset
= gen_canonical_address(addr
);
2967 assert(addr
== gen_48b_address(bo
->offset
));
2971 if (bo
->offset
== 0) {
2972 uint64_t addr
= util_vma_heap_alloc(&device
->vma_lo
, bo
->size
, 4096);
2974 bo
->offset
= gen_canonical_address(addr
);
2975 assert(addr
== gen_48b_address(bo
->offset
));
2979 pthread_mutex_unlock(&device
->vma_mutex
);
2981 return bo
->offset
!= 0;
2985 anv_vma_free(struct anv_device
*device
, struct anv_bo
*bo
)
2987 if (!(bo
->flags
& EXEC_OBJECT_PINNED
))
2990 const uint64_t addr_48b
= gen_48b_address(bo
->offset
);
2992 pthread_mutex_lock(&device
->vma_mutex
);
2994 if (addr_48b
>= LOW_HEAP_MIN_ADDRESS
&&
2995 addr_48b
<= LOW_HEAP_MAX_ADDRESS
) {
2996 util_vma_heap_free(&device
->vma_lo
, addr_48b
, bo
->size
);
2998 assert(addr_48b
>= HIGH_HEAP_MIN_ADDRESS
);
2999 util_vma_heap_free(&device
->vma_hi
, addr_48b
, bo
->size
);
3002 pthread_mutex_unlock(&device
->vma_mutex
);
3007 VkResult
anv_AllocateMemory(
3009 const VkMemoryAllocateInfo
* pAllocateInfo
,
3010 const VkAllocationCallbacks
* pAllocator
,
3011 VkDeviceMemory
* pMem
)
3013 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3014 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3015 struct anv_device_memory
*mem
;
3016 VkResult result
= VK_SUCCESS
;
3018 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
3020 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
3021 assert(pAllocateInfo
->allocationSize
> 0);
3023 VkDeviceSize aligned_alloc_size
=
3024 align_u64(pAllocateInfo
->allocationSize
, 4096);
3026 if (aligned_alloc_size
> MAX_MEMORY_ALLOCATION_SIZE
)
3027 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
3029 assert(pAllocateInfo
->memoryTypeIndex
< pdevice
->memory
.type_count
);
3030 struct anv_memory_type
*mem_type
=
3031 &pdevice
->memory
.types
[pAllocateInfo
->memoryTypeIndex
];
3032 assert(mem_type
->heapIndex
< pdevice
->memory
.heap_count
);
3033 struct anv_memory_heap
*mem_heap
=
3034 &pdevice
->memory
.heaps
[mem_type
->heapIndex
];
3036 uint64_t mem_heap_used
= p_atomic_read(&mem_heap
->used
);
3037 if (mem_heap_used
+ aligned_alloc_size
> mem_heap
->size
)
3038 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
3040 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
3041 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3043 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3045 assert(pAllocateInfo
->memoryTypeIndex
< pdevice
->memory
.type_count
);
3046 mem
->type
= mem_type
;
3050 mem
->host_ptr
= NULL
;
3052 enum anv_bo_alloc_flags alloc_flags
= 0;
3054 const struct wsi_memory_allocate_info
*wsi_info
=
3055 vk_find_struct_const(pAllocateInfo
->pNext
, WSI_MEMORY_ALLOCATE_INFO_MESA
);
3056 if (wsi_info
&& wsi_info
->implicit_sync
) {
3057 /* We need to set the WRITE flag on window system buffers so that GEM
3058 * will know we're writing to them and synchronize uses on other rings
3059 * (eg if the display server uses the blitter ring).
3061 alloc_flags
|= ANV_BO_ALLOC_IMPLICIT_SYNC
|
3062 ANV_BO_ALLOC_IMPLICIT_WRITE
;
3065 const VkExportMemoryAllocateInfo
*export_info
=
3066 vk_find_struct_const(pAllocateInfo
->pNext
, EXPORT_MEMORY_ALLOCATE_INFO
);
3068 /* Check if we need to support Android HW buffer export. If so,
3069 * create AHardwareBuffer and import memory from it.
3071 bool android_export
= false;
3072 if (export_info
&& export_info
->handleTypes
&
3073 VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID
)
3074 android_export
= true;
3076 /* Android memory import. */
3077 const struct VkImportAndroidHardwareBufferInfoANDROID
*ahw_import_info
=
3078 vk_find_struct_const(pAllocateInfo
->pNext
,
3079 IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID
);
3081 if (ahw_import_info
) {
3082 result
= anv_import_ahw_memory(_device
, mem
, ahw_import_info
);
3083 if (result
!= VK_SUCCESS
)
3087 } else if (android_export
) {
3088 result
= anv_create_ahw_memory(_device
, mem
, pAllocateInfo
);
3089 if (result
!= VK_SUCCESS
)
3092 const struct VkImportAndroidHardwareBufferInfoANDROID import_info
= {
3095 result
= anv_import_ahw_memory(_device
, mem
, &import_info
);
3096 if (result
!= VK_SUCCESS
)
3102 const VkImportMemoryFdInfoKHR
*fd_info
=
3103 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_FD_INFO_KHR
);
3105 /* The Vulkan spec permits handleType to be 0, in which case the struct is
3108 if (fd_info
&& fd_info
->handleType
) {
3109 /* At the moment, we support only the below handle types. */
3110 assert(fd_info
->handleType
==
3111 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
3112 fd_info
->handleType
==
3113 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
3115 result
= anv_device_import_bo(device
, fd_info
->fd
, alloc_flags
,
3117 if (result
!= VK_SUCCESS
)
3120 VkDeviceSize aligned_alloc_size
=
3121 align_u64(pAllocateInfo
->allocationSize
, 4096);
3123 /* For security purposes, we reject importing the bo if it's smaller
3124 * than the requested allocation size. This prevents a malicious client
3125 * from passing a buffer to a trusted client, lying about the size, and
3126 * telling the trusted client to try and texture from an image that goes
3127 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
3128 * in the trusted client. The trusted client can protect itself against
3129 * this sort of attack but only if it can trust the buffer size.
3131 if (mem
->bo
->size
< aligned_alloc_size
) {
3132 result
= vk_errorf(device
->instance
, device
,
3133 VK_ERROR_INVALID_EXTERNAL_HANDLE
,
3134 "aligned allocationSize too large for "
3135 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT: "
3136 "%"PRIu64
"B > %"PRIu64
"B",
3137 aligned_alloc_size
, mem
->bo
->size
);
3138 anv_device_release_bo(device
, mem
->bo
);
3142 /* From the Vulkan spec:
3144 * "Importing memory from a file descriptor transfers ownership of
3145 * the file descriptor from the application to the Vulkan
3146 * implementation. The application must not perform any operations on
3147 * the file descriptor after a successful import."
3149 * If the import fails, we leave the file descriptor open.
3155 const VkImportMemoryHostPointerInfoEXT
*host_ptr_info
=
3156 vk_find_struct_const(pAllocateInfo
->pNext
,
3157 IMPORT_MEMORY_HOST_POINTER_INFO_EXT
);
3158 if (host_ptr_info
&& host_ptr_info
->handleType
) {
3159 if (host_ptr_info
->handleType
==
3160 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_MAPPED_FOREIGN_MEMORY_BIT_EXT
) {
3161 result
= vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
3165 assert(host_ptr_info
->handleType
==
3166 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
);
3168 result
= anv_device_import_bo_from_host_ptr(device
,
3169 host_ptr_info
->pHostPointer
,
3170 pAllocateInfo
->allocationSize
,
3174 if (result
!= VK_SUCCESS
)
3177 mem
->host_ptr
= host_ptr_info
->pHostPointer
;
3181 /* Regular allocate (not importing memory). */
3183 if (export_info
&& export_info
->handleTypes
)
3184 alloc_flags
|= ANV_BO_ALLOC_EXTERNAL
;
3186 result
= anv_device_alloc_bo(device
, pAllocateInfo
->allocationSize
,
3187 alloc_flags
, &mem
->bo
);
3188 if (result
!= VK_SUCCESS
)
3191 const VkMemoryDedicatedAllocateInfo
*dedicated_info
=
3192 vk_find_struct_const(pAllocateInfo
->pNext
, MEMORY_DEDICATED_ALLOCATE_INFO
);
3193 if (dedicated_info
&& dedicated_info
->image
!= VK_NULL_HANDLE
) {
3194 ANV_FROM_HANDLE(anv_image
, image
, dedicated_info
->image
);
3196 /* Some legacy (non-modifiers) consumers need the tiling to be set on
3197 * the BO. In this case, we have a dedicated allocation.
3199 if (image
->needs_set_tiling
) {
3200 const uint32_t i915_tiling
=
3201 isl_tiling_to_i915_tiling(image
->planes
[0].surface
.isl
.tiling
);
3202 int ret
= anv_gem_set_tiling(device
, mem
->bo
->gem_handle
,
3203 image
->planes
[0].surface
.isl
.row_pitch_B
,
3206 anv_device_release_bo(device
, mem
->bo
);
3207 result
= vk_errorf(device
->instance
, NULL
,
3208 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
3209 "failed to set BO tiling: %m");
3216 mem_heap_used
= p_atomic_add_return(&mem_heap
->used
, mem
->bo
->size
);
3217 if (mem_heap_used
> mem_heap
->size
) {
3218 p_atomic_add(&mem_heap
->used
, -mem
->bo
->size
);
3219 anv_device_release_bo(device
, mem
->bo
);
3220 result
= vk_errorf(device
->instance
, NULL
,
3221 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
3222 "Out of heap memory");
3226 pthread_mutex_lock(&device
->mutex
);
3227 list_addtail(&mem
->link
, &device
->memory_objects
);
3228 pthread_mutex_unlock(&device
->mutex
);
3230 *pMem
= anv_device_memory_to_handle(mem
);
3235 vk_free2(&device
->alloc
, pAllocator
, mem
);
3240 VkResult
anv_GetMemoryFdKHR(
3242 const VkMemoryGetFdInfoKHR
* pGetFdInfo
,
3245 ANV_FROM_HANDLE(anv_device
, dev
, device_h
);
3246 ANV_FROM_HANDLE(anv_device_memory
, mem
, pGetFdInfo
->memory
);
3248 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
3250 assert(pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
3251 pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
3253 return anv_device_export_bo(dev
, mem
->bo
, pFd
);
3256 VkResult
anv_GetMemoryFdPropertiesKHR(
3258 VkExternalMemoryHandleTypeFlagBits handleType
,
3260 VkMemoryFdPropertiesKHR
* pMemoryFdProperties
)
3262 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3263 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3265 switch (handleType
) {
3266 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
:
3267 /* dma-buf can be imported as any memory type */
3268 pMemoryFdProperties
->memoryTypeBits
=
3269 (1 << pdevice
->memory
.type_count
) - 1;
3273 /* The valid usage section for this function says:
3275 * "handleType must not be one of the handle types defined as
3278 * So opaque handle types fall into the default "unsupported" case.
3280 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
3284 VkResult
anv_GetMemoryHostPointerPropertiesEXT(
3286 VkExternalMemoryHandleTypeFlagBits handleType
,
3287 const void* pHostPointer
,
3288 VkMemoryHostPointerPropertiesEXT
* pMemoryHostPointerProperties
)
3290 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3292 assert(pMemoryHostPointerProperties
->sType
==
3293 VK_STRUCTURE_TYPE_MEMORY_HOST_POINTER_PROPERTIES_EXT
);
3295 switch (handleType
) {
3296 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
: {
3297 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3299 /* Host memory can be imported as any memory type. */
3300 pMemoryHostPointerProperties
->memoryTypeBits
=
3301 (1ull << pdevice
->memory
.type_count
) - 1;
3306 return VK_ERROR_INVALID_EXTERNAL_HANDLE
;
3310 void anv_FreeMemory(
3312 VkDeviceMemory _mem
,
3313 const VkAllocationCallbacks
* pAllocator
)
3315 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3316 ANV_FROM_HANDLE(anv_device_memory
, mem
, _mem
);
3317 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3322 pthread_mutex_lock(&device
->mutex
);
3323 list_del(&mem
->link
);
3324 pthread_mutex_unlock(&device
->mutex
);
3327 anv_UnmapMemory(_device
, _mem
);
3329 p_atomic_add(&pdevice
->memory
.heaps
[mem
->type
->heapIndex
].used
,
3332 anv_device_release_bo(device
, mem
->bo
);
3334 #if defined(ANDROID) && ANDROID_API_LEVEL >= 26
3336 AHardwareBuffer_release(mem
->ahw
);
3339 vk_free2(&device
->alloc
, pAllocator
, mem
);
3342 VkResult
anv_MapMemory(
3344 VkDeviceMemory _memory
,
3345 VkDeviceSize offset
,
3347 VkMemoryMapFlags flags
,
3350 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3351 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
3358 if (mem
->host_ptr
) {
3359 *ppData
= mem
->host_ptr
+ offset
;
3363 if (size
== VK_WHOLE_SIZE
)
3364 size
= mem
->bo
->size
- offset
;
3366 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
3368 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
3369 * assert(size != 0);
3370 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
3371 * equal to the size of the memory minus offset
3374 assert(offset
+ size
<= mem
->bo
->size
);
3376 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
3377 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
3378 * at a time is valid. We could just mmap up front and return an offset
3379 * pointer here, but that may exhaust virtual memory on 32 bit
3382 uint32_t gem_flags
= 0;
3384 if (!device
->info
.has_llc
&&
3385 (mem
->type
->propertyFlags
& VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
))
3386 gem_flags
|= I915_MMAP_WC
;
3388 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
3389 uint64_t map_offset
= offset
& ~4095ull;
3390 assert(offset
>= map_offset
);
3391 uint64_t map_size
= (offset
+ size
) - map_offset
;
3393 /* Let's map whole pages */
3394 map_size
= align_u64(map_size
, 4096);
3396 void *map
= anv_gem_mmap(device
, mem
->bo
->gem_handle
,
3397 map_offset
, map_size
, gem_flags
);
3398 if (map
== MAP_FAILED
)
3399 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
3402 mem
->map_size
= map_size
;
3404 *ppData
= mem
->map
+ (offset
- map_offset
);
3409 void anv_UnmapMemory(
3411 VkDeviceMemory _memory
)
3413 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
3415 if (mem
== NULL
|| mem
->host_ptr
)
3418 anv_gem_munmap(mem
->map
, mem
->map_size
);
3425 clflush_mapped_ranges(struct anv_device
*device
,
3427 const VkMappedMemoryRange
*ranges
)
3429 for (uint32_t i
= 0; i
< count
; i
++) {
3430 ANV_FROM_HANDLE(anv_device_memory
, mem
, ranges
[i
].memory
);
3431 if (ranges
[i
].offset
>= mem
->map_size
)
3434 gen_clflush_range(mem
->map
+ ranges
[i
].offset
,
3435 MIN2(ranges
[i
].size
, mem
->map_size
- ranges
[i
].offset
));
3439 VkResult
anv_FlushMappedMemoryRanges(
3441 uint32_t memoryRangeCount
,
3442 const VkMappedMemoryRange
* pMemoryRanges
)
3444 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3446 if (device
->info
.has_llc
)
3449 /* Make sure the writes we're flushing have landed. */
3450 __builtin_ia32_mfence();
3452 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
3457 VkResult
anv_InvalidateMappedMemoryRanges(
3459 uint32_t memoryRangeCount
,
3460 const VkMappedMemoryRange
* pMemoryRanges
)
3462 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3464 if (device
->info
.has_llc
)
3467 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
3469 /* Make sure no reads get moved up above the invalidate. */
3470 __builtin_ia32_mfence();
3475 void anv_GetBufferMemoryRequirements(
3478 VkMemoryRequirements
* pMemoryRequirements
)
3480 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
3481 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3482 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3484 /* The Vulkan spec (git aaed022) says:
3486 * memoryTypeBits is a bitfield and contains one bit set for every
3487 * supported memory type for the resource. The bit `1<<i` is set if and
3488 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3489 * structure for the physical device is supported.
3491 uint32_t memory_types
= (1ull << pdevice
->memory
.type_count
) - 1;
3493 /* Base alignment requirement of a cache line */
3494 uint32_t alignment
= 16;
3496 /* We need an alignment of 32 for pushing UBOs */
3497 if (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
)
3498 alignment
= MAX2(alignment
, 32);
3500 pMemoryRequirements
->size
= buffer
->size
;
3501 pMemoryRequirements
->alignment
= alignment
;
3503 /* Storage and Uniform buffers should have their size aligned to
3504 * 32-bits to avoid boundary checks when last DWord is not complete.
3505 * This would ensure that not internal padding would be needed for
3508 if (device
->robust_buffer_access
&&
3509 (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
||
3510 buffer
->usage
& VK_BUFFER_USAGE_STORAGE_BUFFER_BIT
))
3511 pMemoryRequirements
->size
= align_u64(buffer
->size
, 4);
3513 pMemoryRequirements
->memoryTypeBits
= memory_types
;
3516 void anv_GetBufferMemoryRequirements2(
3518 const VkBufferMemoryRequirementsInfo2
* pInfo
,
3519 VkMemoryRequirements2
* pMemoryRequirements
)
3521 anv_GetBufferMemoryRequirements(_device
, pInfo
->buffer
,
3522 &pMemoryRequirements
->memoryRequirements
);
3524 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3525 switch (ext
->sType
) {
3526 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
3527 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
3528 requirements
->prefersDedicatedAllocation
= false;
3529 requirements
->requiresDedicatedAllocation
= false;
3534 anv_debug_ignored_stype(ext
->sType
);
3540 void anv_GetImageMemoryRequirements(
3543 VkMemoryRequirements
* pMemoryRequirements
)
3545 ANV_FROM_HANDLE(anv_image
, image
, _image
);
3546 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3547 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3549 /* The Vulkan spec (git aaed022) says:
3551 * memoryTypeBits is a bitfield and contains one bit set for every
3552 * supported memory type for the resource. The bit `1<<i` is set if and
3553 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3554 * structure for the physical device is supported.
3556 * All types are currently supported for images.
3558 uint32_t memory_types
= (1ull << pdevice
->memory
.type_count
) - 1;
3560 /* We must have image allocated or imported at this point. According to the
3561 * specification, external images must have been bound to memory before
3562 * calling GetImageMemoryRequirements.
3564 assert(image
->size
> 0);
3566 pMemoryRequirements
->size
= image
->size
;
3567 pMemoryRequirements
->alignment
= image
->alignment
;
3568 pMemoryRequirements
->memoryTypeBits
= memory_types
;
3571 void anv_GetImageMemoryRequirements2(
3573 const VkImageMemoryRequirementsInfo2
* pInfo
,
3574 VkMemoryRequirements2
* pMemoryRequirements
)
3576 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3577 ANV_FROM_HANDLE(anv_image
, image
, pInfo
->image
);
3579 anv_GetImageMemoryRequirements(_device
, pInfo
->image
,
3580 &pMemoryRequirements
->memoryRequirements
);
3582 vk_foreach_struct_const(ext
, pInfo
->pNext
) {
3583 switch (ext
->sType
) {
3584 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO
: {
3585 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3586 const VkImagePlaneMemoryRequirementsInfo
*plane_reqs
=
3587 (const VkImagePlaneMemoryRequirementsInfo
*) ext
;
3588 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
,
3589 plane_reqs
->planeAspect
);
3591 assert(image
->planes
[plane
].offset
== 0);
3593 /* The Vulkan spec (git aaed022) says:
3595 * memoryTypeBits is a bitfield and contains one bit set for every
3596 * supported memory type for the resource. The bit `1<<i` is set
3597 * if and only if the memory type `i` in the
3598 * VkPhysicalDeviceMemoryProperties structure for the physical
3599 * device is supported.
3601 * All types are currently supported for images.
3603 pMemoryRequirements
->memoryRequirements
.memoryTypeBits
=
3604 (1ull << pdevice
->memory
.type_count
) - 1;
3606 /* We must have image allocated or imported at this point. According to the
3607 * specification, external images must have been bound to memory before
3608 * calling GetImageMemoryRequirements.
3610 assert(image
->planes
[plane
].size
> 0);
3612 pMemoryRequirements
->memoryRequirements
.size
= image
->planes
[plane
].size
;
3613 pMemoryRequirements
->memoryRequirements
.alignment
=
3614 image
->planes
[plane
].alignment
;
3619 anv_debug_ignored_stype(ext
->sType
);
3624 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3625 switch (ext
->sType
) {
3626 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
3627 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
3628 if (image
->needs_set_tiling
|| image
->external_format
) {
3629 /* If we need to set the tiling for external consumers, we need a
3630 * dedicated allocation.
3632 * See also anv_AllocateMemory.
3634 requirements
->prefersDedicatedAllocation
= true;
3635 requirements
->requiresDedicatedAllocation
= true;
3637 requirements
->prefersDedicatedAllocation
= false;
3638 requirements
->requiresDedicatedAllocation
= false;
3644 anv_debug_ignored_stype(ext
->sType
);
3650 void anv_GetImageSparseMemoryRequirements(
3653 uint32_t* pSparseMemoryRequirementCount
,
3654 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
3656 *pSparseMemoryRequirementCount
= 0;
3659 void anv_GetImageSparseMemoryRequirements2(
3661 const VkImageSparseMemoryRequirementsInfo2
* pInfo
,
3662 uint32_t* pSparseMemoryRequirementCount
,
3663 VkSparseImageMemoryRequirements2
* pSparseMemoryRequirements
)
3665 *pSparseMemoryRequirementCount
= 0;
3668 void anv_GetDeviceMemoryCommitment(
3670 VkDeviceMemory memory
,
3671 VkDeviceSize
* pCommittedMemoryInBytes
)
3673 *pCommittedMemoryInBytes
= 0;
3677 anv_bind_buffer_memory(const VkBindBufferMemoryInfo
*pBindInfo
)
3679 ANV_FROM_HANDLE(anv_device_memory
, mem
, pBindInfo
->memory
);
3680 ANV_FROM_HANDLE(anv_buffer
, buffer
, pBindInfo
->buffer
);
3682 assert(pBindInfo
->sType
== VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
);
3685 buffer
->address
= (struct anv_address
) {
3687 .offset
= pBindInfo
->memoryOffset
,
3690 buffer
->address
= ANV_NULL_ADDRESS
;
3694 VkResult
anv_BindBufferMemory(
3697 VkDeviceMemory memory
,
3698 VkDeviceSize memoryOffset
)
3700 anv_bind_buffer_memory(
3701 &(VkBindBufferMemoryInfo
) {
3702 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
,
3705 .memoryOffset
= memoryOffset
,
3711 VkResult
anv_BindBufferMemory2(
3713 uint32_t bindInfoCount
,
3714 const VkBindBufferMemoryInfo
* pBindInfos
)
3716 for (uint32_t i
= 0; i
< bindInfoCount
; i
++)
3717 anv_bind_buffer_memory(&pBindInfos
[i
]);
3722 VkResult
anv_QueueBindSparse(
3724 uint32_t bindInfoCount
,
3725 const VkBindSparseInfo
* pBindInfo
,
3728 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
3729 if (anv_device_is_lost(queue
->device
))
3730 return VK_ERROR_DEVICE_LOST
;
3732 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
3737 VkResult
anv_CreateEvent(
3739 const VkEventCreateInfo
* pCreateInfo
,
3740 const VkAllocationCallbacks
* pAllocator
,
3743 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3744 struct anv_state state
;
3745 struct anv_event
*event
;
3747 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_EVENT_CREATE_INFO
);
3749 state
= anv_state_pool_alloc(&device
->dynamic_state_pool
,
3752 event
->state
= state
;
3753 event
->semaphore
= VK_EVENT_RESET
;
3755 if (!device
->info
.has_llc
) {
3756 /* Make sure the writes we're flushing have landed. */
3757 __builtin_ia32_mfence();
3758 __builtin_ia32_clflush(event
);
3761 *pEvent
= anv_event_to_handle(event
);
3766 void anv_DestroyEvent(
3769 const VkAllocationCallbacks
* pAllocator
)
3771 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3772 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3777 anv_state_pool_free(&device
->dynamic_state_pool
, event
->state
);
3780 VkResult
anv_GetEventStatus(
3784 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3785 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3787 if (anv_device_is_lost(device
))
3788 return VK_ERROR_DEVICE_LOST
;
3790 if (!device
->info
.has_llc
) {
3791 /* Invalidate read cache before reading event written by GPU. */
3792 __builtin_ia32_clflush(event
);
3793 __builtin_ia32_mfence();
3797 return event
->semaphore
;
3800 VkResult
anv_SetEvent(
3804 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3805 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3807 event
->semaphore
= VK_EVENT_SET
;
3809 if (!device
->info
.has_llc
) {
3810 /* Make sure the writes we're flushing have landed. */
3811 __builtin_ia32_mfence();
3812 __builtin_ia32_clflush(event
);
3818 VkResult
anv_ResetEvent(
3822 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3823 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3825 event
->semaphore
= VK_EVENT_RESET
;
3827 if (!device
->info
.has_llc
) {
3828 /* Make sure the writes we're flushing have landed. */
3829 __builtin_ia32_mfence();
3830 __builtin_ia32_clflush(event
);
3838 VkResult
anv_CreateBuffer(
3840 const VkBufferCreateInfo
* pCreateInfo
,
3841 const VkAllocationCallbacks
* pAllocator
,
3844 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3845 struct anv_buffer
*buffer
;
3847 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
3849 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
3850 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3852 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3854 buffer
->size
= pCreateInfo
->size
;
3855 buffer
->usage
= pCreateInfo
->usage
;
3856 buffer
->address
= ANV_NULL_ADDRESS
;
3858 *pBuffer
= anv_buffer_to_handle(buffer
);
3863 void anv_DestroyBuffer(
3866 const VkAllocationCallbacks
* pAllocator
)
3868 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3869 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
3874 vk_free2(&device
->alloc
, pAllocator
, buffer
);
3877 VkDeviceAddress
anv_GetBufferDeviceAddressEXT(
3879 const VkBufferDeviceAddressInfoEXT
* pInfo
)
3881 ANV_FROM_HANDLE(anv_buffer
, buffer
, pInfo
->buffer
);
3883 assert(buffer
->address
.bo
->flags
& EXEC_OBJECT_PINNED
);
3885 return anv_address_physical(buffer
->address
);
3889 anv_fill_buffer_surface_state(struct anv_device
*device
, struct anv_state state
,
3890 enum isl_format format
,
3891 struct anv_address address
,
3892 uint32_t range
, uint32_t stride
)
3894 isl_buffer_fill_state(&device
->isl_dev
, state
.map
,
3895 .address
= anv_address_physical(address
),
3896 .mocs
= device
->isl_dev
.mocs
.internal
,
3899 .swizzle
= ISL_SWIZZLE_IDENTITY
,
3900 .stride_B
= stride
);
3903 void anv_DestroySampler(
3906 const VkAllocationCallbacks
* pAllocator
)
3908 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3909 ANV_FROM_HANDLE(anv_sampler
, sampler
, _sampler
);
3914 if (sampler
->bindless_state
.map
) {
3915 anv_state_pool_free(&device
->dynamic_state_pool
,
3916 sampler
->bindless_state
);
3919 vk_free2(&device
->alloc
, pAllocator
, sampler
);
3922 VkResult
anv_CreateFramebuffer(
3924 const VkFramebufferCreateInfo
* pCreateInfo
,
3925 const VkAllocationCallbacks
* pAllocator
,
3926 VkFramebuffer
* pFramebuffer
)
3928 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3929 struct anv_framebuffer
*framebuffer
;
3931 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
3933 size_t size
= sizeof(*framebuffer
);
3935 /* VK_KHR_imageless_framebuffer extension says:
3937 * If flags includes VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR,
3938 * parameter pAttachments is ignored.
3940 if (!(pCreateInfo
->flags
& VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR
)) {
3941 size
+= sizeof(struct anv_image_view
*) * pCreateInfo
->attachmentCount
;
3942 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
3943 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3944 if (framebuffer
== NULL
)
3945 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3947 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
3948 ANV_FROM_HANDLE(anv_image_view
, iview
, pCreateInfo
->pAttachments
[i
]);
3949 framebuffer
->attachments
[i
] = iview
;
3951 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
3953 assert(device
->enabled_extensions
.KHR_imageless_framebuffer
);
3954 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
3955 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3956 if (framebuffer
== NULL
)
3957 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3959 framebuffer
->attachment_count
= 0;
3962 framebuffer
->width
= pCreateInfo
->width
;
3963 framebuffer
->height
= pCreateInfo
->height
;
3964 framebuffer
->layers
= pCreateInfo
->layers
;
3966 *pFramebuffer
= anv_framebuffer_to_handle(framebuffer
);
3971 void anv_DestroyFramebuffer(
3974 const VkAllocationCallbacks
* pAllocator
)
3976 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3977 ANV_FROM_HANDLE(anv_framebuffer
, fb
, _fb
);
3982 vk_free2(&device
->alloc
, pAllocator
, fb
);
3985 static const VkTimeDomainEXT anv_time_domains
[] = {
3986 VK_TIME_DOMAIN_DEVICE_EXT
,
3987 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
,
3988 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
,
3991 VkResult
anv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
3992 VkPhysicalDevice physicalDevice
,
3993 uint32_t *pTimeDomainCount
,
3994 VkTimeDomainEXT
*pTimeDomains
)
3997 VK_OUTARRAY_MAKE(out
, pTimeDomains
, pTimeDomainCount
);
3999 for (d
= 0; d
< ARRAY_SIZE(anv_time_domains
); d
++) {
4000 vk_outarray_append(&out
, i
) {
4001 *i
= anv_time_domains
[d
];
4005 return vk_outarray_status(&out
);
4009 anv_clock_gettime(clockid_t clock_id
)
4011 struct timespec current
;
4014 ret
= clock_gettime(clock_id
, ¤t
);
4015 if (ret
< 0 && clock_id
== CLOCK_MONOTONIC_RAW
)
4016 ret
= clock_gettime(CLOCK_MONOTONIC
, ¤t
);
4020 return (uint64_t) current
.tv_sec
* 1000000000ULL + current
.tv_nsec
;
4023 #define TIMESTAMP 0x2358
4025 VkResult
anv_GetCalibratedTimestampsEXT(
4027 uint32_t timestampCount
,
4028 const VkCalibratedTimestampInfoEXT
*pTimestampInfos
,
4029 uint64_t *pTimestamps
,
4030 uint64_t *pMaxDeviation
)
4032 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4033 uint64_t timestamp_frequency
= device
->info
.timestamp_frequency
;
4036 uint64_t begin
, end
;
4037 uint64_t max_clock_period
= 0;
4039 begin
= anv_clock_gettime(CLOCK_MONOTONIC_RAW
);
4041 for (d
= 0; d
< timestampCount
; d
++) {
4042 switch (pTimestampInfos
[d
].timeDomain
) {
4043 case VK_TIME_DOMAIN_DEVICE_EXT
:
4044 ret
= anv_gem_reg_read(device
, TIMESTAMP
| 1,
4048 return anv_device_set_lost(device
, "Failed to read the TIMESTAMP "
4051 uint64_t device_period
= DIV_ROUND_UP(1000000000, timestamp_frequency
);
4052 max_clock_period
= MAX2(max_clock_period
, device_period
);
4054 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
:
4055 pTimestamps
[d
] = anv_clock_gettime(CLOCK_MONOTONIC
);
4056 max_clock_period
= MAX2(max_clock_period
, 1);
4059 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
:
4060 pTimestamps
[d
] = begin
;
4068 end
= anv_clock_gettime(CLOCK_MONOTONIC_RAW
);
4071 * The maximum deviation is the sum of the interval over which we
4072 * perform the sampling and the maximum period of any sampled
4073 * clock. That's because the maximum skew between any two sampled
4074 * clock edges is when the sampled clock with the largest period is
4075 * sampled at the end of that period but right at the beginning of the
4076 * sampling interval and some other clock is sampled right at the
4077 * begining of its sampling period and right at the end of the
4078 * sampling interval. Let's assume the GPU has the longest clock
4079 * period and that the application is sampling GPU and monotonic:
4082 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
4083 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4087 * GPU -----_____-----_____-----_____-----_____
4090 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
4091 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4093 * Interval <----------------->
4094 * Deviation <-------------------------->
4098 * m = read(monotonic) 2
4101 * We round the sample interval up by one tick to cover sampling error
4102 * in the interval clock
4105 uint64_t sample_interval
= end
- begin
+ 1;
4107 *pMaxDeviation
= sample_interval
+ max_clock_period
;
4112 /* vk_icd.h does not declare this function, so we declare it here to
4113 * suppress Wmissing-prototypes.
4115 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4116 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
);
4118 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4119 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
)
4121 /* For the full details on loader interface versioning, see
4122 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
4123 * What follows is a condensed summary, to help you navigate the large and
4124 * confusing official doc.
4126 * - Loader interface v0 is incompatible with later versions. We don't
4129 * - In loader interface v1:
4130 * - The first ICD entrypoint called by the loader is
4131 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
4133 * - The ICD must statically expose no other Vulkan symbol unless it is
4134 * linked with -Bsymbolic.
4135 * - Each dispatchable Vulkan handle created by the ICD must be
4136 * a pointer to a struct whose first member is VK_LOADER_DATA. The
4137 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
4138 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
4139 * vkDestroySurfaceKHR(). The ICD must be capable of working with
4140 * such loader-managed surfaces.
4142 * - Loader interface v2 differs from v1 in:
4143 * - The first ICD entrypoint called by the loader is
4144 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
4145 * statically expose this entrypoint.
4147 * - Loader interface v3 differs from v2 in:
4148 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
4149 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
4150 * because the loader no longer does so.
4152 * - Loader interface v4 differs from v3 in:
4153 * - The ICD must implement vk_icdGetPhysicalDeviceProcAddr().
4155 *pSupportedVersion
= MIN2(*pSupportedVersion
, 4u);