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")
60 DRI_CONF_VK_WSI_FORCE_BGRA8_UNORM_FIRST("false")
64 /* This is probably far to big but it reflects the max size used for messages
65 * in OpenGLs KHR_debug.
67 #define MAX_DEBUG_MESSAGE_LENGTH 4096
70 compiler_debug_log(void *data
, const char *fmt
, ...)
72 char str
[MAX_DEBUG_MESSAGE_LENGTH
];
73 struct anv_device
*device
= (struct anv_device
*)data
;
74 struct anv_instance
*instance
= device
->physical
->instance
;
76 if (list_is_empty(&instance
->debug_report_callbacks
.callbacks
))
81 (void) vsnprintf(str
, MAX_DEBUG_MESSAGE_LENGTH
, fmt
, args
);
84 vk_debug_report(&instance
->debug_report_callbacks
,
85 VK_DEBUG_REPORT_DEBUG_BIT_EXT
,
86 VK_DEBUG_REPORT_OBJECT_TYPE_UNKNOWN_EXT
,
91 compiler_perf_log(void *data
, const char *fmt
, ...)
96 if (unlikely(INTEL_DEBUG
& DEBUG_PERF
))
97 intel_logd_v(fmt
, args
);
103 anv_compute_heap_size(int fd
, uint64_t gtt_size
)
105 /* Query the total ram from the system */
109 uint64_t total_ram
= (uint64_t)info
.totalram
* (uint64_t)info
.mem_unit
;
111 /* We don't want to burn too much ram with the GPU. If the user has 4GiB
112 * or less, we use at most half. If they have more than 4GiB, we use 3/4.
114 uint64_t available_ram
;
115 if (total_ram
<= 4ull * 1024ull * 1024ull * 1024ull)
116 available_ram
= total_ram
/ 2;
118 available_ram
= total_ram
* 3 / 4;
120 /* We also want to leave some padding for things we allocate in the driver,
121 * so don't go over 3/4 of the GTT either.
123 uint64_t available_gtt
= gtt_size
* 3 / 4;
125 return MIN2(available_ram
, available_gtt
);
129 anv_physical_device_init_heaps(struct anv_physical_device
*device
, int fd
)
131 if (anv_gem_get_context_param(fd
, 0, I915_CONTEXT_PARAM_GTT_SIZE
,
132 &device
->gtt_size
) == -1) {
133 /* If, for whatever reason, we can't actually get the GTT size from the
134 * kernel (too old?) fall back to the aperture size.
136 anv_perf_warn(NULL
, NULL
,
137 "Failed to get I915_CONTEXT_PARAM_GTT_SIZE: %m");
139 if (anv_gem_get_aperture(fd
, &device
->gtt_size
) == -1) {
140 return vk_errorfi(device
->instance
, NULL
,
141 VK_ERROR_INITIALIZATION_FAILED
,
142 "failed to get aperture size: %m");
146 /* We only allow 48-bit addresses with softpin because knowing the actual
147 * address is required for the vertex cache flush workaround.
149 device
->supports_48bit_addresses
= (device
->info
.gen
>= 8) &&
150 device
->has_softpin
&&
151 device
->gtt_size
> (4ULL << 30 /* GiB */);
153 uint64_t heap_size
= anv_compute_heap_size(fd
, device
->gtt_size
);
155 if (heap_size
> (2ull << 30) && !device
->supports_48bit_addresses
) {
156 /* When running with an overridden PCI ID, we may get a GTT size from
157 * the kernel that is greater than 2 GiB but the execbuf check for 48bit
158 * address support can still fail. Just clamp the address space size to
159 * 2 GiB if we don't have 48-bit support.
161 intel_logw("%s:%d: The kernel reported a GTT size larger than 2 GiB but "
162 "not support for 48-bit addresses",
164 heap_size
= 2ull << 30;
167 device
->memory
.heap_count
= 1;
168 device
->memory
.heaps
[0] = (struct anv_memory_heap
) {
170 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
173 uint32_t type_count
= 0;
174 for (uint32_t heap
= 0; heap
< device
->memory
.heap_count
; heap
++) {
175 if (device
->info
.has_llc
) {
176 /* Big core GPUs share LLC with the CPU and thus one memory type can be
177 * both cached and coherent at the same time.
179 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
180 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
181 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
182 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
183 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
187 /* The spec requires that we expose a host-visible, coherent memory
188 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
189 * to give the application a choice between cached, but not coherent and
190 * coherent but uncached (WC though).
192 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
193 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
194 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
195 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
198 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
199 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
200 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
201 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
206 device
->memory
.type_count
= type_count
;
212 anv_physical_device_init_uuids(struct anv_physical_device
*device
)
214 const struct build_id_note
*note
=
215 build_id_find_nhdr_for_addr(anv_physical_device_init_uuids
);
217 return vk_errorfi(device
->instance
, NULL
,
218 VK_ERROR_INITIALIZATION_FAILED
,
219 "Failed to find build-id");
222 unsigned build_id_len
= build_id_length(note
);
223 if (build_id_len
< 20) {
224 return vk_errorfi(device
->instance
, NULL
,
225 VK_ERROR_INITIALIZATION_FAILED
,
226 "build-id too short. It needs to be a SHA");
229 memcpy(device
->driver_build_sha1
, build_id_data(note
), 20);
231 struct mesa_sha1 sha1_ctx
;
233 STATIC_ASSERT(VK_UUID_SIZE
<= sizeof(sha1
));
235 /* The pipeline cache UUID is used for determining when a pipeline cache is
236 * invalid. It needs both a driver build and the PCI ID of the device.
238 _mesa_sha1_init(&sha1_ctx
);
239 _mesa_sha1_update(&sha1_ctx
, build_id_data(note
), build_id_len
);
240 _mesa_sha1_update(&sha1_ctx
, &device
->info
.chipset_id
,
241 sizeof(device
->info
.chipset_id
));
242 _mesa_sha1_update(&sha1_ctx
, &device
->always_use_bindless
,
243 sizeof(device
->always_use_bindless
));
244 _mesa_sha1_update(&sha1_ctx
, &device
->has_a64_buffer_access
,
245 sizeof(device
->has_a64_buffer_access
));
246 _mesa_sha1_update(&sha1_ctx
, &device
->has_bindless_images
,
247 sizeof(device
->has_bindless_images
));
248 _mesa_sha1_update(&sha1_ctx
, &device
->has_bindless_samplers
,
249 sizeof(device
->has_bindless_samplers
));
250 _mesa_sha1_final(&sha1_ctx
, sha1
);
251 memcpy(device
->pipeline_cache_uuid
, sha1
, VK_UUID_SIZE
);
253 /* The driver UUID is used for determining sharability of images and memory
254 * between two Vulkan instances in separate processes. People who want to
255 * share memory need to also check the device UUID (below) so all this
256 * needs to be is the build-id.
258 memcpy(device
->driver_uuid
, build_id_data(note
), VK_UUID_SIZE
);
260 /* The device UUID uniquely identifies the given device within the machine.
261 * Since we never have more than one device, this doesn't need to be a real
262 * UUID. However, on the off-chance that someone tries to use this to
263 * cache pre-tiled images or something of the like, we use the PCI ID and
264 * some bits of ISL info to ensure that this is safe.
266 _mesa_sha1_init(&sha1_ctx
);
267 _mesa_sha1_update(&sha1_ctx
, &device
->info
.chipset_id
,
268 sizeof(device
->info
.chipset_id
));
269 _mesa_sha1_update(&sha1_ctx
, &device
->isl_dev
.has_bit6_swizzling
,
270 sizeof(device
->isl_dev
.has_bit6_swizzling
));
271 _mesa_sha1_final(&sha1_ctx
, sha1
);
272 memcpy(device
->device_uuid
, sha1
, VK_UUID_SIZE
);
278 anv_physical_device_init_disk_cache(struct anv_physical_device
*device
)
280 #ifdef ENABLE_SHADER_CACHE
282 ASSERTED
int len
= snprintf(renderer
, sizeof(renderer
), "anv_%04x",
283 device
->info
.chipset_id
);
284 assert(len
== sizeof(renderer
) - 2);
287 _mesa_sha1_format(timestamp
, device
->driver_build_sha1
);
289 const uint64_t driver_flags
=
290 brw_get_compiler_config_value(device
->compiler
);
291 device
->disk_cache
= disk_cache_create(renderer
, timestamp
, driver_flags
);
293 device
->disk_cache
= NULL
;
298 anv_physical_device_free_disk_cache(struct anv_physical_device
*device
)
300 #ifdef ENABLE_SHADER_CACHE
301 if (device
->disk_cache
)
302 disk_cache_destroy(device
->disk_cache
);
304 assert(device
->disk_cache
== NULL
);
309 get_available_system_memory()
311 char *meminfo
= os_read_file("/proc/meminfo");
315 char *str
= strstr(meminfo
, "MemAvailable:");
321 uint64_t kb_mem_available
;
322 if (sscanf(str
, "MemAvailable: %" PRIx64
, &kb_mem_available
) == 1) {
324 return kb_mem_available
<< 10;
332 anv_physical_device_try_create(struct anv_instance
*instance
,
333 drmDevicePtr drm_device
,
334 struct anv_physical_device
**device_out
)
336 const char *primary_path
= drm_device
->nodes
[DRM_NODE_PRIMARY
];
337 const char *path
= drm_device
->nodes
[DRM_NODE_RENDER
];
342 brw_process_intel_debug_variable();
344 fd
= open(path
, O_RDWR
| O_CLOEXEC
);
346 return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
348 struct gen_device_info devinfo
;
349 if (!gen_get_device_info_from_fd(fd
, &devinfo
)) {
350 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
354 const char *device_name
= gen_get_device_name(devinfo
.chipset_id
);
356 if (devinfo
.is_haswell
) {
357 intel_logw("Haswell Vulkan support is incomplete");
358 } else if (devinfo
.gen
== 7 && !devinfo
.is_baytrail
) {
359 intel_logw("Ivy Bridge Vulkan support is incomplete");
360 } else if (devinfo
.gen
== 7 && devinfo
.is_baytrail
) {
361 intel_logw("Bay Trail Vulkan support is incomplete");
362 } else if (devinfo
.gen
>= 8 && devinfo
.gen
<= 11) {
363 /* Gen8-11 fully supported */
364 } else if (devinfo
.gen
== 12) {
365 intel_logw("Vulkan is not yet fully supported on gen12");
367 result
= vk_errorfi(instance
, NULL
, VK_ERROR_INCOMPATIBLE_DRIVER
,
368 "Vulkan not yet supported on %s", device_name
);
372 struct anv_physical_device
*device
=
373 vk_alloc(&instance
->alloc
, sizeof(*device
), 8,
374 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
375 if (device
== NULL
) {
376 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
380 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
381 device
->instance
= instance
;
383 assert(strlen(path
) < ARRAY_SIZE(device
->path
));
384 snprintf(device
->path
, ARRAY_SIZE(device
->path
), "%s", path
);
386 device
->info
= devinfo
;
387 device
->name
= device_name
;
389 device
->no_hw
= device
->info
.no_hw
;
390 if (getenv("INTEL_NO_HW") != NULL
)
391 device
->no_hw
= true;
393 device
->pci_info
.domain
= drm_device
->businfo
.pci
->domain
;
394 device
->pci_info
.bus
= drm_device
->businfo
.pci
->bus
;
395 device
->pci_info
.device
= drm_device
->businfo
.pci
->dev
;
396 device
->pci_info
.function
= drm_device
->businfo
.pci
->func
;
398 device
->cmd_parser_version
= -1;
399 if (device
->info
.gen
== 7) {
400 device
->cmd_parser_version
=
401 anv_gem_get_param(fd
, I915_PARAM_CMD_PARSER_VERSION
);
402 if (device
->cmd_parser_version
== -1) {
403 result
= vk_errorfi(device
->instance
, NULL
,
404 VK_ERROR_INITIALIZATION_FAILED
,
405 "failed to get command parser version");
410 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_WAIT_TIMEOUT
)) {
411 result
= vk_errorfi(device
->instance
, NULL
,
412 VK_ERROR_INITIALIZATION_FAILED
,
413 "kernel missing gem wait");
417 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_EXECBUF2
)) {
418 result
= vk_errorfi(device
->instance
, NULL
,
419 VK_ERROR_INITIALIZATION_FAILED
,
420 "kernel missing execbuf2");
424 if (!device
->info
.has_llc
&&
425 anv_gem_get_param(fd
, I915_PARAM_MMAP_VERSION
) < 1) {
426 result
= vk_errorfi(device
->instance
, NULL
,
427 VK_ERROR_INITIALIZATION_FAILED
,
428 "kernel missing wc mmap");
432 device
->has_softpin
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_SOFTPIN
);
433 device
->has_exec_async
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_ASYNC
);
434 device
->has_exec_capture
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_CAPTURE
);
435 device
->has_exec_fence
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE
);
436 device
->has_syncobj
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE_ARRAY
);
437 device
->has_syncobj_wait
= device
->has_syncobj
&&
438 anv_gem_supports_syncobj_wait(fd
);
439 device
->has_context_priority
= anv_gem_has_context_priority(fd
);
441 result
= anv_physical_device_init_heaps(device
, fd
);
442 if (result
!= VK_SUCCESS
)
445 device
->use_softpin
= device
->has_softpin
&&
446 device
->supports_48bit_addresses
;
448 device
->has_context_isolation
=
449 anv_gem_get_param(fd
, I915_PARAM_HAS_CONTEXT_ISOLATION
);
451 device
->always_use_bindless
=
452 env_var_as_boolean("ANV_ALWAYS_BINDLESS", false);
454 /* We first got the A64 messages on broadwell and we can only use them if
455 * we can pass addresses directly into the shader which requires softpin.
457 device
->has_a64_buffer_access
= device
->info
.gen
>= 8 &&
460 /* We first get bindless image access on Skylake and we can only really do
461 * it if we don't have any relocations so we need softpin.
463 device
->has_bindless_images
= device
->info
.gen
>= 9 &&
466 /* We've had bindless samplers since Ivy Bridge (forever in Vulkan terms)
467 * because it's just a matter of setting the sampler address in the sample
468 * message header. However, we've not bothered to wire it up for vec4 so
469 * we leave it disabled on gen7.
471 device
->has_bindless_samplers
= device
->info
.gen
>= 8;
473 device
->has_implicit_ccs
= device
->info
.has_aux_map
;
475 device
->has_mem_available
= get_available_system_memory() != 0;
477 device
->always_flush_cache
=
478 driQueryOptionb(&instance
->dri_options
, "always_flush_cache");
480 /* Starting with Gen10, the timestamp frequency of the command streamer may
481 * vary from one part to another. We can query the value from the kernel.
483 if (device
->info
.gen
>= 10) {
484 int timestamp_frequency
=
485 anv_gem_get_param(fd
, I915_PARAM_CS_TIMESTAMP_FREQUENCY
);
487 if (timestamp_frequency
< 0)
488 intel_logw("Kernel 4.16-rc1+ required to properly query CS timestamp frequency");
490 device
->info
.timestamp_frequency
= timestamp_frequency
;
493 /* GENs prior to 8 do not support EU/Subslice info */
494 if (device
->info
.gen
>= 8) {
495 device
->subslice_total
= anv_gem_get_param(fd
, I915_PARAM_SUBSLICE_TOTAL
);
496 device
->eu_total
= anv_gem_get_param(fd
, I915_PARAM_EU_TOTAL
);
498 /* Without this information, we cannot get the right Braswell
499 * brandstrings, and we have to use conservative numbers for GPGPU on
500 * many platforms, but otherwise, things will just work.
502 if (device
->subslice_total
< 1 || device
->eu_total
< 1) {
503 intel_logw("Kernel 4.1 required to properly query GPU properties");
505 } else if (device
->info
.gen
== 7) {
506 device
->subslice_total
= 1 << (device
->info
.gt
- 1);
509 if (device
->info
.is_cherryview
&&
510 device
->subslice_total
> 0 && device
->eu_total
> 0) {
511 /* Logical CS threads = EUs per subslice * num threads per EU */
512 uint32_t max_cs_threads
=
513 device
->eu_total
/ device
->subslice_total
* device
->info
.num_thread_per_eu
;
515 /* Fuse configurations may give more threads than expected, never less. */
516 if (max_cs_threads
> device
->info
.max_cs_threads
)
517 device
->info
.max_cs_threads
= max_cs_threads
;
520 device
->compiler
= brw_compiler_create(NULL
, &device
->info
);
521 if (device
->compiler
== NULL
) {
522 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
525 device
->compiler
->shader_debug_log
= compiler_debug_log
;
526 device
->compiler
->shader_perf_log
= compiler_perf_log
;
527 device
->compiler
->supports_pull_constants
= false;
528 device
->compiler
->constant_buffer_0_is_relative
=
529 device
->info
.gen
< 8 || !device
->has_context_isolation
;
530 device
->compiler
->supports_shader_constants
= true;
531 device
->compiler
->compact_params
= false;
533 /* Broadwell PRM says:
535 * "Before Gen8, there was a historical configuration control field to
536 * swizzle address bit[6] for in X/Y tiling modes. This was set in three
537 * different places: TILECTL[1:0], ARB_MODE[5:4], and
538 * DISP_ARB_CTL[14:13].
540 * For Gen8 and subsequent generations, the swizzle fields are all
541 * reserved, and the CPU's memory controller performs all address
542 * swizzling modifications."
545 device
->info
.gen
< 8 && anv_gem_get_bit6_swizzle(fd
, I915_TILING_X
);
547 isl_device_init(&device
->isl_dev
, &device
->info
, swizzled
);
549 result
= anv_physical_device_init_uuids(device
);
550 if (result
!= VK_SUCCESS
)
553 anv_physical_device_init_disk_cache(device
);
555 if (instance
->enabled_extensions
.KHR_display
) {
556 master_fd
= open(primary_path
, O_RDWR
| O_CLOEXEC
);
557 if (master_fd
>= 0) {
558 /* prod the device with a GETPARAM call which will fail if
559 * we don't have permission to even render on this device
561 if (anv_gem_get_param(master_fd
, I915_PARAM_CHIPSET_ID
) == 0) {
567 device
->master_fd
= master_fd
;
569 result
= anv_init_wsi(device
);
570 if (result
!= VK_SUCCESS
)
571 goto fail_disk_cache
;
573 device
->perf
= anv_get_perf(&device
->info
, fd
);
575 anv_physical_device_get_supported_extensions(device
,
576 &device
->supported_extensions
);
579 device
->local_fd
= fd
;
581 *device_out
= device
;
586 anv_physical_device_free_disk_cache(device
);
588 ralloc_free(device
->compiler
);
590 vk_free(&instance
->alloc
, device
);
599 anv_physical_device_destroy(struct anv_physical_device
*device
)
601 anv_finish_wsi(device
);
602 anv_physical_device_free_disk_cache(device
);
603 ralloc_free(device
->compiler
);
604 ralloc_free(device
->perf
);
605 close(device
->local_fd
);
606 if (device
->master_fd
>= 0)
607 close(device
->master_fd
);
608 vk_free(&device
->instance
->alloc
, device
);
612 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
613 VkSystemAllocationScope allocationScope
)
619 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
620 size_t align
, VkSystemAllocationScope allocationScope
)
622 return realloc(pOriginal
, size
);
626 default_free_func(void *pUserData
, void *pMemory
)
631 static const VkAllocationCallbacks default_alloc
= {
633 .pfnAllocation
= default_alloc_func
,
634 .pfnReallocation
= default_realloc_func
,
635 .pfnFree
= default_free_func
,
638 VkResult
anv_EnumerateInstanceExtensionProperties(
639 const char* pLayerName
,
640 uint32_t* pPropertyCount
,
641 VkExtensionProperties
* pProperties
)
643 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
645 for (int i
= 0; i
< ANV_INSTANCE_EXTENSION_COUNT
; i
++) {
646 if (anv_instance_extensions_supported
.extensions
[i
]) {
647 vk_outarray_append(&out
, prop
) {
648 *prop
= anv_instance_extensions
[i
];
653 return vk_outarray_status(&out
);
656 VkResult
anv_CreateInstance(
657 const VkInstanceCreateInfo
* pCreateInfo
,
658 const VkAllocationCallbacks
* pAllocator
,
659 VkInstance
* pInstance
)
661 struct anv_instance
*instance
;
664 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
666 struct anv_instance_extension_table enabled_extensions
= {};
667 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
669 for (idx
= 0; idx
< ANV_INSTANCE_EXTENSION_COUNT
; idx
++) {
670 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
671 anv_instance_extensions
[idx
].extensionName
) == 0)
675 if (idx
>= ANV_INSTANCE_EXTENSION_COUNT
)
676 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
678 if (!anv_instance_extensions_supported
.extensions
[idx
])
679 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
681 enabled_extensions
.extensions
[idx
] = true;
684 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
685 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
687 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
689 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
692 instance
->alloc
= *pAllocator
;
694 instance
->alloc
= default_alloc
;
696 instance
->app_info
= (struct anv_app_info
) { .api_version
= 0 };
697 if (pCreateInfo
->pApplicationInfo
) {
698 const VkApplicationInfo
*app
= pCreateInfo
->pApplicationInfo
;
700 instance
->app_info
.app_name
=
701 vk_strdup(&instance
->alloc
, app
->pApplicationName
,
702 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
703 instance
->app_info
.app_version
= app
->applicationVersion
;
705 instance
->app_info
.engine_name
=
706 vk_strdup(&instance
->alloc
, app
->pEngineName
,
707 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
708 instance
->app_info
.engine_version
= app
->engineVersion
;
710 instance
->app_info
.api_version
= app
->apiVersion
;
713 if (instance
->app_info
.api_version
== 0)
714 instance
->app_info
.api_version
= VK_API_VERSION_1_0
;
716 instance
->enabled_extensions
= enabled_extensions
;
718 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->dispatch
.entrypoints
); i
++) {
719 /* Vulkan requires that entrypoints for extensions which have not been
720 * enabled must not be advertised.
722 if (!anv_instance_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
723 &instance
->enabled_extensions
)) {
724 instance
->dispatch
.entrypoints
[i
] = NULL
;
726 instance
->dispatch
.entrypoints
[i
] =
727 anv_instance_dispatch_table
.entrypoints
[i
];
731 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->physical_device_dispatch
.entrypoints
); i
++) {
732 /* Vulkan requires that entrypoints for extensions which have not been
733 * enabled must not be advertised.
735 if (!anv_physical_device_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
736 &instance
->enabled_extensions
)) {
737 instance
->physical_device_dispatch
.entrypoints
[i
] = NULL
;
739 instance
->physical_device_dispatch
.entrypoints
[i
] =
740 anv_physical_device_dispatch_table
.entrypoints
[i
];
744 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->device_dispatch
.entrypoints
); i
++) {
745 /* Vulkan requires that entrypoints for extensions which have not been
746 * enabled must not be advertised.
748 if (!anv_device_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
749 &instance
->enabled_extensions
, NULL
)) {
750 instance
->device_dispatch
.entrypoints
[i
] = NULL
;
752 instance
->device_dispatch
.entrypoints
[i
] =
753 anv_device_dispatch_table
.entrypoints
[i
];
757 instance
->physical_devices_enumerated
= false;
758 list_inithead(&instance
->physical_devices
);
760 result
= vk_debug_report_instance_init(&instance
->debug_report_callbacks
);
761 if (result
!= VK_SUCCESS
) {
762 vk_free2(&default_alloc
, pAllocator
, instance
);
763 return vk_error(result
);
766 instance
->pipeline_cache_enabled
=
767 env_var_as_boolean("ANV_ENABLE_PIPELINE_CACHE", true);
769 glsl_type_singleton_init_or_ref();
771 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
773 driParseOptionInfo(&instance
->available_dri_options
, anv_dri_options_xml
);
774 driParseConfigFiles(&instance
->dri_options
, &instance
->available_dri_options
,
776 instance
->app_info
.engine_name
,
777 instance
->app_info
.engine_version
);
779 *pInstance
= anv_instance_to_handle(instance
);
784 void anv_DestroyInstance(
785 VkInstance _instance
,
786 const VkAllocationCallbacks
* pAllocator
)
788 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
793 list_for_each_entry_safe(struct anv_physical_device
, pdevice
,
794 &instance
->physical_devices
, link
)
795 anv_physical_device_destroy(pdevice
);
797 vk_free(&instance
->alloc
, (char *)instance
->app_info
.app_name
);
798 vk_free(&instance
->alloc
, (char *)instance
->app_info
.engine_name
);
800 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
802 vk_debug_report_instance_destroy(&instance
->debug_report_callbacks
);
804 glsl_type_singleton_decref();
806 driDestroyOptionCache(&instance
->dri_options
);
807 driDestroyOptionInfo(&instance
->available_dri_options
);
809 vk_free(&instance
->alloc
, instance
);
813 anv_enumerate_physical_devices(struct anv_instance
*instance
)
815 if (instance
->physical_devices_enumerated
)
818 instance
->physical_devices_enumerated
= true;
820 /* TODO: Check for more devices ? */
821 drmDevicePtr devices
[8];
824 max_devices
= drmGetDevices2(0, devices
, ARRAY_SIZE(devices
));
828 VkResult result
= VK_SUCCESS
;
829 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
830 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
831 devices
[i
]->bustype
== DRM_BUS_PCI
&&
832 devices
[i
]->deviceinfo
.pci
->vendor_id
== 0x8086) {
834 struct anv_physical_device
*pdevice
;
835 result
= anv_physical_device_try_create(instance
, devices
[i
],
837 /* Incompatible DRM device, skip. */
838 if (result
== VK_ERROR_INCOMPATIBLE_DRIVER
) {
843 /* Error creating the physical device, report the error. */
844 if (result
!= VK_SUCCESS
)
847 list_addtail(&pdevice
->link
, &instance
->physical_devices
);
850 drmFreeDevices(devices
, max_devices
);
852 /* If we successfully enumerated any devices, call it success */
856 VkResult
anv_EnumeratePhysicalDevices(
857 VkInstance _instance
,
858 uint32_t* pPhysicalDeviceCount
,
859 VkPhysicalDevice
* pPhysicalDevices
)
861 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
862 VK_OUTARRAY_MAKE(out
, pPhysicalDevices
, pPhysicalDeviceCount
);
864 VkResult result
= anv_enumerate_physical_devices(instance
);
865 if (result
!= VK_SUCCESS
)
868 list_for_each_entry(struct anv_physical_device
, pdevice
,
869 &instance
->physical_devices
, link
) {
870 vk_outarray_append(&out
, i
) {
871 *i
= anv_physical_device_to_handle(pdevice
);
875 return vk_outarray_status(&out
);
878 VkResult
anv_EnumeratePhysicalDeviceGroups(
879 VkInstance _instance
,
880 uint32_t* pPhysicalDeviceGroupCount
,
881 VkPhysicalDeviceGroupProperties
* pPhysicalDeviceGroupProperties
)
883 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
884 VK_OUTARRAY_MAKE(out
, pPhysicalDeviceGroupProperties
,
885 pPhysicalDeviceGroupCount
);
887 VkResult result
= anv_enumerate_physical_devices(instance
);
888 if (result
!= VK_SUCCESS
)
891 list_for_each_entry(struct anv_physical_device
, pdevice
,
892 &instance
->physical_devices
, link
) {
893 vk_outarray_append(&out
, p
) {
894 p
->physicalDeviceCount
= 1;
895 memset(p
->physicalDevices
, 0, sizeof(p
->physicalDevices
));
896 p
->physicalDevices
[0] = anv_physical_device_to_handle(pdevice
);
897 p
->subsetAllocation
= false;
899 vk_foreach_struct(ext
, p
->pNext
)
900 anv_debug_ignored_stype(ext
->sType
);
904 return vk_outarray_status(&out
);
907 void anv_GetPhysicalDeviceFeatures(
908 VkPhysicalDevice physicalDevice
,
909 VkPhysicalDeviceFeatures
* pFeatures
)
911 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
913 *pFeatures
= (VkPhysicalDeviceFeatures
) {
914 .robustBufferAccess
= true,
915 .fullDrawIndexUint32
= true,
916 .imageCubeArray
= true,
917 .independentBlend
= true,
918 .geometryShader
= true,
919 .tessellationShader
= true,
920 .sampleRateShading
= true,
921 .dualSrcBlend
= true,
923 .multiDrawIndirect
= true,
924 .drawIndirectFirstInstance
= true,
926 .depthBiasClamp
= true,
927 .fillModeNonSolid
= true,
928 .depthBounds
= pdevice
->info
.gen
>= 12,
932 .multiViewport
= true,
933 .samplerAnisotropy
= true,
934 .textureCompressionETC2
= pdevice
->info
.gen
>= 8 ||
935 pdevice
->info
.is_baytrail
,
936 .textureCompressionASTC_LDR
= pdevice
->info
.gen
>= 9, /* FINISHME CHV */
937 .textureCompressionBC
= true,
938 .occlusionQueryPrecise
= true,
939 .pipelineStatisticsQuery
= true,
940 .fragmentStoresAndAtomics
= true,
941 .shaderTessellationAndGeometryPointSize
= true,
942 .shaderImageGatherExtended
= true,
943 .shaderStorageImageExtendedFormats
= true,
944 .shaderStorageImageMultisample
= false,
945 .shaderStorageImageReadWithoutFormat
= false,
946 .shaderStorageImageWriteWithoutFormat
= true,
947 .shaderUniformBufferArrayDynamicIndexing
= true,
948 .shaderSampledImageArrayDynamicIndexing
= true,
949 .shaderStorageBufferArrayDynamicIndexing
= true,
950 .shaderStorageImageArrayDynamicIndexing
= true,
951 .shaderClipDistance
= true,
952 .shaderCullDistance
= true,
953 .shaderFloat64
= pdevice
->info
.gen
>= 8 &&
954 pdevice
->info
.has_64bit_float
,
955 .shaderInt64
= pdevice
->info
.gen
>= 8 &&
956 pdevice
->info
.has_64bit_int
,
957 .shaderInt16
= pdevice
->info
.gen
>= 8,
958 .shaderResourceMinLod
= pdevice
->info
.gen
>= 9,
959 .variableMultisampleRate
= true,
960 .inheritedQueries
= true,
963 /* We can't do image stores in vec4 shaders */
964 pFeatures
->vertexPipelineStoresAndAtomics
=
965 pdevice
->compiler
->scalar_stage
[MESA_SHADER_VERTEX
] &&
966 pdevice
->compiler
->scalar_stage
[MESA_SHADER_GEOMETRY
];
968 struct anv_app_info
*app_info
= &pdevice
->instance
->app_info
;
970 /* The new DOOM and Wolfenstein games require depthBounds without
971 * checking for it. They seem to run fine without it so just claim it's
972 * there and accept the consequences.
974 if (app_info
->engine_name
&& strcmp(app_info
->engine_name
, "idTech") == 0)
975 pFeatures
->depthBounds
= true;
979 anv_get_physical_device_features_1_1(struct anv_physical_device
*pdevice
,
980 VkPhysicalDeviceVulkan11Features
*f
)
982 assert(f
->sType
== VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES
);
984 f
->storageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
985 f
->uniformAndStorageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
986 f
->storagePushConstant16
= pdevice
->info
.gen
>= 8;
987 f
->storageInputOutput16
= false;
989 f
->multiviewGeometryShader
= true;
990 f
->multiviewTessellationShader
= true;
991 f
->variablePointersStorageBuffer
= true;
992 f
->variablePointers
= true;
993 f
->protectedMemory
= false;
994 f
->samplerYcbcrConversion
= true;
995 f
->shaderDrawParameters
= true;
999 anv_get_physical_device_features_1_2(struct anv_physical_device
*pdevice
,
1000 VkPhysicalDeviceVulkan12Features
*f
)
1002 assert(f
->sType
== VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES
);
1004 f
->samplerMirrorClampToEdge
= true;
1005 f
->drawIndirectCount
= true;
1006 f
->storageBuffer8BitAccess
= pdevice
->info
.gen
>= 8;
1007 f
->uniformAndStorageBuffer8BitAccess
= pdevice
->info
.gen
>= 8;
1008 f
->storagePushConstant8
= pdevice
->info
.gen
>= 8;
1009 f
->shaderBufferInt64Atomics
= pdevice
->info
.gen
>= 9 &&
1010 pdevice
->use_softpin
;
1011 f
->shaderSharedInt64Atomics
= false;
1012 f
->shaderFloat16
= pdevice
->info
.gen
>= 8;
1013 f
->shaderInt8
= pdevice
->info
.gen
>= 8;
1015 bool descIndexing
= pdevice
->has_a64_buffer_access
&&
1016 pdevice
->has_bindless_images
;
1017 f
->descriptorIndexing
= descIndexing
;
1018 f
->shaderInputAttachmentArrayDynamicIndexing
= false;
1019 f
->shaderUniformTexelBufferArrayDynamicIndexing
= descIndexing
;
1020 f
->shaderStorageTexelBufferArrayDynamicIndexing
= descIndexing
;
1021 f
->shaderUniformBufferArrayNonUniformIndexing
= false;
1022 f
->shaderSampledImageArrayNonUniformIndexing
= descIndexing
;
1023 f
->shaderStorageBufferArrayNonUniformIndexing
= descIndexing
;
1024 f
->shaderStorageImageArrayNonUniformIndexing
= descIndexing
;
1025 f
->shaderInputAttachmentArrayNonUniformIndexing
= false;
1026 f
->shaderUniformTexelBufferArrayNonUniformIndexing
= descIndexing
;
1027 f
->shaderStorageTexelBufferArrayNonUniformIndexing
= descIndexing
;
1028 f
->descriptorBindingUniformBufferUpdateAfterBind
= false;
1029 f
->descriptorBindingSampledImageUpdateAfterBind
= descIndexing
;
1030 f
->descriptorBindingStorageImageUpdateAfterBind
= descIndexing
;
1031 f
->descriptorBindingStorageBufferUpdateAfterBind
= descIndexing
;
1032 f
->descriptorBindingUniformTexelBufferUpdateAfterBind
= descIndexing
;
1033 f
->descriptorBindingStorageTexelBufferUpdateAfterBind
= descIndexing
;
1034 f
->descriptorBindingUpdateUnusedWhilePending
= descIndexing
;
1035 f
->descriptorBindingPartiallyBound
= descIndexing
;
1036 f
->descriptorBindingVariableDescriptorCount
= false;
1037 f
->runtimeDescriptorArray
= descIndexing
;
1039 f
->samplerFilterMinmax
= pdevice
->info
.gen
>= 9;
1040 f
->scalarBlockLayout
= true;
1041 f
->imagelessFramebuffer
= true;
1042 f
->uniformBufferStandardLayout
= true;
1043 f
->shaderSubgroupExtendedTypes
= true;
1044 f
->separateDepthStencilLayouts
= true;
1045 f
->hostQueryReset
= true;
1046 f
->timelineSemaphore
= true;
1047 f
->bufferDeviceAddress
= pdevice
->has_a64_buffer_access
;
1048 f
->bufferDeviceAddressCaptureReplay
= pdevice
->has_a64_buffer_access
;
1049 f
->bufferDeviceAddressMultiDevice
= false;
1050 f
->vulkanMemoryModel
= true;
1051 f
->vulkanMemoryModelDeviceScope
= true;
1052 f
->vulkanMemoryModelAvailabilityVisibilityChains
= true;
1053 f
->shaderOutputViewportIndex
= true;
1054 f
->shaderOutputLayer
= true;
1055 f
->subgroupBroadcastDynamicId
= true;
1058 void anv_GetPhysicalDeviceFeatures2(
1059 VkPhysicalDevice physicalDevice
,
1060 VkPhysicalDeviceFeatures2
* pFeatures
)
1062 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1063 anv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
1065 VkPhysicalDeviceVulkan11Features core_1_1
= {
1066 .sType
= VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES
,
1068 anv_get_physical_device_features_1_1(pdevice
, &core_1_1
);
1070 VkPhysicalDeviceVulkan12Features core_1_2
= {
1071 .sType
= VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES
,
1073 anv_get_physical_device_features_1_2(pdevice
, &core_1_2
);
1075 #define CORE_FEATURE(major, minor, feature) \
1076 features->feature = core_##major##_##minor.feature
1079 vk_foreach_struct(ext
, pFeatures
->pNext
) {
1080 switch (ext
->sType
) {
1081 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES_KHR
: {
1082 VkPhysicalDevice8BitStorageFeaturesKHR
*features
=
1083 (VkPhysicalDevice8BitStorageFeaturesKHR
*)ext
;
1084 CORE_FEATURE(1, 2, storageBuffer8BitAccess
);
1085 CORE_FEATURE(1, 2, uniformAndStorageBuffer8BitAccess
);
1086 CORE_FEATURE(1, 2, storagePushConstant8
);
1090 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES
: {
1091 VkPhysicalDevice16BitStorageFeatures
*features
=
1092 (VkPhysicalDevice16BitStorageFeatures
*)ext
;
1093 CORE_FEATURE(1, 1, storageBuffer16BitAccess
);
1094 CORE_FEATURE(1, 1, uniformAndStorageBuffer16BitAccess
);
1095 CORE_FEATURE(1, 1, storagePushConstant16
);
1096 CORE_FEATURE(1, 1, storageInputOutput16
);
1100 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_EXT
: {
1101 VkPhysicalDeviceBufferDeviceAddressFeaturesEXT
*features
= (void *)ext
;
1102 features
->bufferDeviceAddress
= pdevice
->has_a64_buffer_access
;
1103 features
->bufferDeviceAddressCaptureReplay
= false;
1104 features
->bufferDeviceAddressMultiDevice
= false;
1108 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_KHR
: {
1109 VkPhysicalDeviceBufferDeviceAddressFeaturesKHR
*features
= (void *)ext
;
1110 CORE_FEATURE(1, 2, bufferDeviceAddress
);
1111 CORE_FEATURE(1, 2, bufferDeviceAddressCaptureReplay
);
1112 CORE_FEATURE(1, 2, bufferDeviceAddressMultiDevice
);
1116 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COMPUTE_SHADER_DERIVATIVES_FEATURES_NV
: {
1117 VkPhysicalDeviceComputeShaderDerivativesFeaturesNV
*features
=
1118 (VkPhysicalDeviceComputeShaderDerivativesFeaturesNV
*)ext
;
1119 features
->computeDerivativeGroupQuads
= true;
1120 features
->computeDerivativeGroupLinear
= true;
1124 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT
: {
1125 VkPhysicalDeviceConditionalRenderingFeaturesEXT
*features
=
1126 (VkPhysicalDeviceConditionalRenderingFeaturesEXT
*)ext
;
1127 features
->conditionalRendering
= pdevice
->info
.gen
>= 8 ||
1128 pdevice
->info
.is_haswell
;
1129 features
->inheritedConditionalRendering
= pdevice
->info
.gen
>= 8 ||
1130 pdevice
->info
.is_haswell
;
1134 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_CLIP_ENABLE_FEATURES_EXT
: {
1135 VkPhysicalDeviceDepthClipEnableFeaturesEXT
*features
=
1136 (VkPhysicalDeviceDepthClipEnableFeaturesEXT
*)ext
;
1137 features
->depthClipEnable
= true;
1141 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR
: {
1142 VkPhysicalDeviceFloat16Int8FeaturesKHR
*features
= (void *)ext
;
1143 CORE_FEATURE(1, 2, shaderFloat16
);
1144 CORE_FEATURE(1, 2, shaderInt8
);
1148 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FRAGMENT_SHADER_INTERLOCK_FEATURES_EXT
: {
1149 VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT
*features
=
1150 (VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT
*)ext
;
1151 features
->fragmentShaderSampleInterlock
= pdevice
->info
.gen
>= 9;
1152 features
->fragmentShaderPixelInterlock
= pdevice
->info
.gen
>= 9;
1153 features
->fragmentShaderShadingRateInterlock
= false;
1157 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_QUERY_RESET_FEATURES_EXT
: {
1158 VkPhysicalDeviceHostQueryResetFeaturesEXT
*features
=
1159 (VkPhysicalDeviceHostQueryResetFeaturesEXT
*)ext
;
1160 CORE_FEATURE(1, 2, hostQueryReset
);
1164 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT
: {
1165 VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*features
=
1166 (VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*)ext
;
1167 CORE_FEATURE(1, 2, shaderInputAttachmentArrayDynamicIndexing
);
1168 CORE_FEATURE(1, 2, shaderUniformTexelBufferArrayDynamicIndexing
);
1169 CORE_FEATURE(1, 2, shaderStorageTexelBufferArrayDynamicIndexing
);
1170 CORE_FEATURE(1, 2, shaderUniformBufferArrayNonUniformIndexing
);
1171 CORE_FEATURE(1, 2, shaderSampledImageArrayNonUniformIndexing
);
1172 CORE_FEATURE(1, 2, shaderStorageBufferArrayNonUniformIndexing
);
1173 CORE_FEATURE(1, 2, shaderStorageImageArrayNonUniformIndexing
);
1174 CORE_FEATURE(1, 2, shaderInputAttachmentArrayNonUniformIndexing
);
1175 CORE_FEATURE(1, 2, shaderUniformTexelBufferArrayNonUniformIndexing
);
1176 CORE_FEATURE(1, 2, shaderStorageTexelBufferArrayNonUniformIndexing
);
1177 CORE_FEATURE(1, 2, descriptorBindingUniformBufferUpdateAfterBind
);
1178 CORE_FEATURE(1, 2, descriptorBindingSampledImageUpdateAfterBind
);
1179 CORE_FEATURE(1, 2, descriptorBindingStorageImageUpdateAfterBind
);
1180 CORE_FEATURE(1, 2, descriptorBindingStorageBufferUpdateAfterBind
);
1181 CORE_FEATURE(1, 2, descriptorBindingUniformTexelBufferUpdateAfterBind
);
1182 CORE_FEATURE(1, 2, descriptorBindingStorageTexelBufferUpdateAfterBind
);
1183 CORE_FEATURE(1, 2, descriptorBindingUpdateUnusedWhilePending
);
1184 CORE_FEATURE(1, 2, descriptorBindingPartiallyBound
);
1185 CORE_FEATURE(1, 2, descriptorBindingVariableDescriptorCount
);
1186 CORE_FEATURE(1, 2, runtimeDescriptorArray
);
1190 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INDEX_TYPE_UINT8_FEATURES_EXT
: {
1191 VkPhysicalDeviceIndexTypeUint8FeaturesEXT
*features
=
1192 (VkPhysicalDeviceIndexTypeUint8FeaturesEXT
*)ext
;
1193 features
->indexTypeUint8
= true;
1197 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_FEATURES_EXT
: {
1198 VkPhysicalDeviceInlineUniformBlockFeaturesEXT
*features
=
1199 (VkPhysicalDeviceInlineUniformBlockFeaturesEXT
*)ext
;
1200 features
->inlineUniformBlock
= true;
1201 features
->descriptorBindingInlineUniformBlockUpdateAfterBind
= true;
1205 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_FEATURES_EXT
: {
1206 VkPhysicalDeviceLineRasterizationFeaturesEXT
*features
=
1207 (VkPhysicalDeviceLineRasterizationFeaturesEXT
*)ext
;
1208 features
->rectangularLines
= true;
1209 features
->bresenhamLines
= true;
1210 /* Support for Smooth lines with MSAA was removed on gen11. From the
1211 * BSpec section "Multisample ModesState" table for "AA Line Support
1214 * GEN10:BUG:######## NUM_MULTISAMPLES == 1
1216 * Fortunately, this isn't a case most people care about.
1218 features
->smoothLines
= pdevice
->info
.gen
< 10;
1219 features
->stippledRectangularLines
= false;
1220 features
->stippledBresenhamLines
= true;
1221 features
->stippledSmoothLines
= false;
1225 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES
: {
1226 VkPhysicalDeviceMultiviewFeatures
*features
=
1227 (VkPhysicalDeviceMultiviewFeatures
*)ext
;
1228 CORE_FEATURE(1, 1, multiview
);
1229 CORE_FEATURE(1, 1, multiviewGeometryShader
);
1230 CORE_FEATURE(1, 1, multiviewTessellationShader
);
1234 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGELESS_FRAMEBUFFER_FEATURES_KHR
: {
1235 VkPhysicalDeviceImagelessFramebufferFeaturesKHR
*features
=
1236 (VkPhysicalDeviceImagelessFramebufferFeaturesKHR
*)ext
;
1237 CORE_FEATURE(1, 2, imagelessFramebuffer
);
1241 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR
: {
1242 VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR
*features
=
1243 (VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR
*)ext
;
1244 features
->pipelineExecutableInfo
= true;
1248 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES
: {
1249 VkPhysicalDeviceProtectedMemoryFeatures
*features
= (void *)ext
;
1250 CORE_FEATURE(1, 1, protectedMemory
);
1254 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES
: {
1255 VkPhysicalDeviceSamplerYcbcrConversionFeatures
*features
=
1256 (VkPhysicalDeviceSamplerYcbcrConversionFeatures
*) ext
;
1257 CORE_FEATURE(1, 1, samplerYcbcrConversion
);
1261 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES_EXT
: {
1262 VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*features
=
1263 (VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*)ext
;
1264 CORE_FEATURE(1, 2, scalarBlockLayout
);
1268 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SEPARATE_DEPTH_STENCIL_LAYOUTS_FEATURES_KHR
: {
1269 VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR
*features
=
1270 (VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR
*)ext
;
1271 CORE_FEATURE(1, 2, separateDepthStencilLayouts
);
1275 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_INT64_FEATURES_KHR
: {
1276 VkPhysicalDeviceShaderAtomicInt64FeaturesKHR
*features
= (void *)ext
;
1277 CORE_FEATURE(1, 2, shaderBufferInt64Atomics
);
1278 CORE_FEATURE(1, 2, shaderSharedInt64Atomics
);
1282 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DEMOTE_TO_HELPER_INVOCATION_FEATURES_EXT
: {
1283 VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT
*features
= (void *)ext
;
1284 features
->shaderDemoteToHelperInvocation
= true;
1288 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CLOCK_FEATURES_KHR
: {
1289 VkPhysicalDeviceShaderClockFeaturesKHR
*features
=
1290 (VkPhysicalDeviceShaderClockFeaturesKHR
*)ext
;
1291 features
->shaderSubgroupClock
= true;
1292 features
->shaderDeviceClock
= false;
1296 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETERS_FEATURES
: {
1297 VkPhysicalDeviceShaderDrawParametersFeatures
*features
= (void *)ext
;
1298 CORE_FEATURE(1, 1, shaderDrawParameters
);
1302 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_SUBGROUP_EXTENDED_TYPES_FEATURES_KHR
: {
1303 VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR
*features
=
1304 (VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR
*)ext
;
1305 CORE_FEATURE(1, 2, shaderSubgroupExtendedTypes
);
1309 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_FEATURES_EXT
: {
1310 VkPhysicalDeviceSubgroupSizeControlFeaturesEXT
*features
=
1311 (VkPhysicalDeviceSubgroupSizeControlFeaturesEXT
*)ext
;
1312 features
->subgroupSizeControl
= true;
1313 features
->computeFullSubgroups
= true;
1317 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_FEATURES_EXT
: {
1318 VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT
*features
=
1319 (VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT
*)ext
;
1320 features
->texelBufferAlignment
= true;
1324 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_FEATURES_KHR
: {
1325 VkPhysicalDeviceTimelineSemaphoreFeaturesKHR
*features
=
1326 (VkPhysicalDeviceTimelineSemaphoreFeaturesKHR
*) ext
;
1327 CORE_FEATURE(1, 2, timelineSemaphore
);
1331 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES
: {
1332 VkPhysicalDeviceVariablePointersFeatures
*features
= (void *)ext
;
1333 CORE_FEATURE(1, 1, variablePointersStorageBuffer
);
1334 CORE_FEATURE(1, 1, variablePointers
);
1338 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT
: {
1339 VkPhysicalDeviceTransformFeedbackFeaturesEXT
*features
=
1340 (VkPhysicalDeviceTransformFeedbackFeaturesEXT
*)ext
;
1341 features
->transformFeedback
= true;
1342 features
->geometryStreams
= true;
1346 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_UNIFORM_BUFFER_STANDARD_LAYOUT_FEATURES_KHR
: {
1347 VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR
*features
=
1348 (VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR
*)ext
;
1349 CORE_FEATURE(1, 2, uniformBufferStandardLayout
);
1353 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT
: {
1354 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*features
=
1355 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*)ext
;
1356 features
->vertexAttributeInstanceRateDivisor
= true;
1357 features
->vertexAttributeInstanceRateZeroDivisor
= true;
1361 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES
:
1362 anv_get_physical_device_features_1_1(pdevice
, (void *)ext
);
1365 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES
:
1366 anv_get_physical_device_features_1_2(pdevice
, (void *)ext
);
1369 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_MEMORY_MODEL_FEATURES_KHR
: {
1370 VkPhysicalDeviceVulkanMemoryModelFeaturesKHR
*features
= (void *)ext
;
1371 CORE_FEATURE(1, 2, vulkanMemoryModel
);
1372 CORE_FEATURE(1, 2, vulkanMemoryModelDeviceScope
);
1373 CORE_FEATURE(1, 2, vulkanMemoryModelAvailabilityVisibilityChains
);
1377 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_YCBCR_IMAGE_ARRAYS_FEATURES_EXT
: {
1378 VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*features
=
1379 (VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*)ext
;
1380 features
->ycbcrImageArrays
= true;
1385 anv_debug_ignored_stype(ext
->sType
);
1393 #define MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS 64
1395 #define MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS 64
1396 #define MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS 256
1398 void anv_GetPhysicalDeviceProperties(
1399 VkPhysicalDevice physicalDevice
,
1400 VkPhysicalDeviceProperties
* pProperties
)
1402 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1403 const struct gen_device_info
*devinfo
= &pdevice
->info
;
1405 /* See assertions made when programming the buffer surface state. */
1406 const uint32_t max_raw_buffer_sz
= devinfo
->gen
>= 7 ?
1407 (1ul << 30) : (1ul << 27);
1409 const uint32_t max_ssbos
= pdevice
->has_a64_buffer_access
? UINT16_MAX
: 64;
1410 const uint32_t max_textures
=
1411 pdevice
->has_bindless_images
? UINT16_MAX
: 128;
1412 const uint32_t max_samplers
=
1413 pdevice
->has_bindless_samplers
? UINT16_MAX
:
1414 (devinfo
->gen
>= 8 || devinfo
->is_haswell
) ? 128 : 16;
1415 const uint32_t max_images
=
1416 pdevice
->has_bindless_images
? UINT16_MAX
: MAX_IMAGES
;
1418 /* If we can use bindless for everything, claim a high per-stage limit,
1419 * otherwise use the binding table size, minus the slots reserved for
1420 * render targets and one slot for the descriptor buffer. */
1421 const uint32_t max_per_stage
=
1422 pdevice
->has_bindless_images
&& pdevice
->has_a64_buffer_access
1423 ? UINT32_MAX
: MAX_BINDING_TABLE_SIZE
- MAX_RTS
- 1;
1425 const uint32_t max_workgroup_size
= 32 * devinfo
->max_cs_threads
;
1427 VkSampleCountFlags sample_counts
=
1428 isl_device_get_sample_counts(&pdevice
->isl_dev
);
1431 VkPhysicalDeviceLimits limits
= {
1432 .maxImageDimension1D
= (1 << 14),
1433 .maxImageDimension2D
= (1 << 14),
1434 .maxImageDimension3D
= (1 << 11),
1435 .maxImageDimensionCube
= (1 << 14),
1436 .maxImageArrayLayers
= (1 << 11),
1437 .maxTexelBufferElements
= 128 * 1024 * 1024,
1438 .maxUniformBufferRange
= (1ul << 27),
1439 .maxStorageBufferRange
= max_raw_buffer_sz
,
1440 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
1441 .maxMemoryAllocationCount
= UINT32_MAX
,
1442 .maxSamplerAllocationCount
= 64 * 1024,
1443 .bufferImageGranularity
= 64, /* A cache line */
1444 .sparseAddressSpaceSize
= 0,
1445 .maxBoundDescriptorSets
= MAX_SETS
,
1446 .maxPerStageDescriptorSamplers
= max_samplers
,
1447 .maxPerStageDescriptorUniformBuffers
= MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
,
1448 .maxPerStageDescriptorStorageBuffers
= max_ssbos
,
1449 .maxPerStageDescriptorSampledImages
= max_textures
,
1450 .maxPerStageDescriptorStorageImages
= max_images
,
1451 .maxPerStageDescriptorInputAttachments
= MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS
,
1452 .maxPerStageResources
= max_per_stage
,
1453 .maxDescriptorSetSamplers
= 6 * max_samplers
, /* number of stages * maxPerStageDescriptorSamplers */
1454 .maxDescriptorSetUniformBuffers
= 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
, /* number of stages * maxPerStageDescriptorUniformBuffers */
1455 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
1456 .maxDescriptorSetStorageBuffers
= 6 * max_ssbos
, /* number of stages * maxPerStageDescriptorStorageBuffers */
1457 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
1458 .maxDescriptorSetSampledImages
= 6 * max_textures
, /* number of stages * maxPerStageDescriptorSampledImages */
1459 .maxDescriptorSetStorageImages
= 6 * max_images
, /* number of stages * maxPerStageDescriptorStorageImages */
1460 .maxDescriptorSetInputAttachments
= MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS
,
1461 .maxVertexInputAttributes
= MAX_VBS
,
1462 .maxVertexInputBindings
= MAX_VBS
,
1463 .maxVertexInputAttributeOffset
= 2047,
1464 .maxVertexInputBindingStride
= 2048,
1465 .maxVertexOutputComponents
= 128,
1466 .maxTessellationGenerationLevel
= 64,
1467 .maxTessellationPatchSize
= 32,
1468 .maxTessellationControlPerVertexInputComponents
= 128,
1469 .maxTessellationControlPerVertexOutputComponents
= 128,
1470 .maxTessellationControlPerPatchOutputComponents
= 128,
1471 .maxTessellationControlTotalOutputComponents
= 2048,
1472 .maxTessellationEvaluationInputComponents
= 128,
1473 .maxTessellationEvaluationOutputComponents
= 128,
1474 .maxGeometryShaderInvocations
= 32,
1475 .maxGeometryInputComponents
= 64,
1476 .maxGeometryOutputComponents
= 128,
1477 .maxGeometryOutputVertices
= 256,
1478 .maxGeometryTotalOutputComponents
= 1024,
1479 .maxFragmentInputComponents
= 116, /* 128 components - (PSIZ, CLIP_DIST0, CLIP_DIST1) */
1480 .maxFragmentOutputAttachments
= 8,
1481 .maxFragmentDualSrcAttachments
= 1,
1482 .maxFragmentCombinedOutputResources
= 8,
1483 .maxComputeSharedMemorySize
= 64 * 1024,
1484 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
1485 .maxComputeWorkGroupInvocations
= max_workgroup_size
,
1486 .maxComputeWorkGroupSize
= {
1491 .subPixelPrecisionBits
= 8,
1492 .subTexelPrecisionBits
= 8,
1493 .mipmapPrecisionBits
= 8,
1494 .maxDrawIndexedIndexValue
= UINT32_MAX
,
1495 .maxDrawIndirectCount
= UINT32_MAX
,
1496 .maxSamplerLodBias
= 16,
1497 .maxSamplerAnisotropy
= 16,
1498 .maxViewports
= MAX_VIEWPORTS
,
1499 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
1500 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
1501 .viewportSubPixelBits
= 13, /* We take a float? */
1502 .minMemoryMapAlignment
= 4096, /* A page */
1503 /* The dataport requires texel alignment so we need to assume a worst
1504 * case of R32G32B32A32 which is 16 bytes.
1506 .minTexelBufferOffsetAlignment
= 16,
1507 /* We need 16 for UBO block reads to work and 32 for push UBOs */
1508 .minUniformBufferOffsetAlignment
= 32,
1509 .minStorageBufferOffsetAlignment
= 4,
1510 .minTexelOffset
= -8,
1511 .maxTexelOffset
= 7,
1512 .minTexelGatherOffset
= -32,
1513 .maxTexelGatherOffset
= 31,
1514 .minInterpolationOffset
= -0.5,
1515 .maxInterpolationOffset
= 0.4375,
1516 .subPixelInterpolationOffsetBits
= 4,
1517 .maxFramebufferWidth
= (1 << 14),
1518 .maxFramebufferHeight
= (1 << 14),
1519 .maxFramebufferLayers
= (1 << 11),
1520 .framebufferColorSampleCounts
= sample_counts
,
1521 .framebufferDepthSampleCounts
= sample_counts
,
1522 .framebufferStencilSampleCounts
= sample_counts
,
1523 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
1524 .maxColorAttachments
= MAX_RTS
,
1525 .sampledImageColorSampleCounts
= sample_counts
,
1526 .sampledImageIntegerSampleCounts
= sample_counts
,
1527 .sampledImageDepthSampleCounts
= sample_counts
,
1528 .sampledImageStencilSampleCounts
= sample_counts
,
1529 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
1530 .maxSampleMaskWords
= 1,
1531 .timestampComputeAndGraphics
= true,
1532 .timestampPeriod
= 1000000000.0 / devinfo
->timestamp_frequency
,
1533 .maxClipDistances
= 8,
1534 .maxCullDistances
= 8,
1535 .maxCombinedClipAndCullDistances
= 8,
1536 .discreteQueuePriorities
= 2,
1537 .pointSizeRange
= { 0.125, 255.875 },
1540 (devinfo
->gen
>= 9 || devinfo
->is_cherryview
) ?
1541 2047.9921875 : 7.9921875,
1543 .pointSizeGranularity
= (1.0 / 8.0),
1544 .lineWidthGranularity
= (1.0 / 128.0),
1545 .strictLines
= false,
1546 .standardSampleLocations
= true,
1547 .optimalBufferCopyOffsetAlignment
= 128,
1548 .optimalBufferCopyRowPitchAlignment
= 128,
1549 .nonCoherentAtomSize
= 64,
1552 *pProperties
= (VkPhysicalDeviceProperties
) {
1553 .apiVersion
= anv_physical_device_api_version(pdevice
),
1554 .driverVersion
= vk_get_driver_version(),
1556 .deviceID
= pdevice
->info
.chipset_id
,
1557 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
1559 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
1562 snprintf(pProperties
->deviceName
, sizeof(pProperties
->deviceName
),
1563 "%s", pdevice
->name
);
1564 memcpy(pProperties
->pipelineCacheUUID
,
1565 pdevice
->pipeline_cache_uuid
, VK_UUID_SIZE
);
1569 anv_get_physical_device_properties_1_1(struct anv_physical_device
*pdevice
,
1570 VkPhysicalDeviceVulkan11Properties
*p
)
1572 assert(p
->sType
== VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES
);
1574 memcpy(p
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
1575 memcpy(p
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
1576 memset(p
->deviceLUID
, 0, VK_LUID_SIZE
);
1577 p
->deviceNodeMask
= 0;
1578 p
->deviceLUIDValid
= false;
1580 p
->subgroupSize
= BRW_SUBGROUP_SIZE
;
1581 VkShaderStageFlags scalar_stages
= 0;
1582 for (unsigned stage
= 0; stage
< MESA_SHADER_STAGES
; stage
++) {
1583 if (pdevice
->compiler
->scalar_stage
[stage
])
1584 scalar_stages
|= mesa_to_vk_shader_stage(stage
);
1586 p
->subgroupSupportedStages
= scalar_stages
;
1587 p
->subgroupSupportedOperations
= VK_SUBGROUP_FEATURE_BASIC_BIT
|
1588 VK_SUBGROUP_FEATURE_VOTE_BIT
|
1589 VK_SUBGROUP_FEATURE_BALLOT_BIT
|
1590 VK_SUBGROUP_FEATURE_SHUFFLE_BIT
|
1591 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT
|
1592 VK_SUBGROUP_FEATURE_QUAD_BIT
;
1593 if (pdevice
->info
.gen
>= 8) {
1594 /* TODO: There's no technical reason why these can't be made to
1595 * work on gen7 but they don't at the moment so it's best to leave
1596 * the feature disabled than enabled and broken.
1598 p
->subgroupSupportedOperations
|= VK_SUBGROUP_FEATURE_ARITHMETIC_BIT
|
1599 VK_SUBGROUP_FEATURE_CLUSTERED_BIT
;
1601 p
->subgroupQuadOperationsInAllStages
= pdevice
->info
.gen
>= 8;
1603 p
->pointClippingBehavior
= VK_POINT_CLIPPING_BEHAVIOR_USER_CLIP_PLANES_ONLY
;
1604 p
->maxMultiviewViewCount
= 16;
1605 p
->maxMultiviewInstanceIndex
= UINT32_MAX
/ 16;
1606 p
->protectedNoFault
= false;
1607 /* This value doesn't matter for us today as our per-stage descriptors are
1610 p
->maxPerSetDescriptors
= 1024;
1611 p
->maxMemoryAllocationSize
= MAX_MEMORY_ALLOCATION_SIZE
;
1615 anv_get_physical_device_properties_1_2(struct anv_physical_device
*pdevice
,
1616 VkPhysicalDeviceVulkan12Properties
*p
)
1618 assert(p
->sType
== VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES
);
1620 p
->driverID
= VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA_KHR
;
1621 memset(p
->driverName
, 0, sizeof(p
->driverName
));
1622 snprintf(p
->driverName
, VK_MAX_DRIVER_NAME_SIZE_KHR
,
1623 "Intel open-source Mesa driver");
1624 memset(p
->driverInfo
, 0, sizeof(p
->driverInfo
));
1625 snprintf(p
->driverInfo
, VK_MAX_DRIVER_INFO_SIZE_KHR
,
1626 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1
);
1627 p
->conformanceVersion
= (VkConformanceVersionKHR
) {
1634 p
->denormBehaviorIndependence
=
1635 VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_ALL_KHR
;
1636 p
->roundingModeIndependence
=
1637 VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_NONE_KHR
;
1639 /* Broadwell does not support HF denorms and there are restrictions
1640 * other gens. According to Kabylake's PRM:
1642 * "math - Extended Math Function
1644 * Restriction : Half-float denorms are always retained."
1646 p
->shaderDenormFlushToZeroFloat16
= false;
1647 p
->shaderDenormPreserveFloat16
= pdevice
->info
.gen
> 8;
1648 p
->shaderRoundingModeRTEFloat16
= true;
1649 p
->shaderRoundingModeRTZFloat16
= true;
1650 p
->shaderSignedZeroInfNanPreserveFloat16
= true;
1652 p
->shaderDenormFlushToZeroFloat32
= true;
1653 p
->shaderDenormPreserveFloat32
= true;
1654 p
->shaderRoundingModeRTEFloat32
= true;
1655 p
->shaderRoundingModeRTZFloat32
= true;
1656 p
->shaderSignedZeroInfNanPreserveFloat32
= true;
1658 p
->shaderDenormFlushToZeroFloat64
= true;
1659 p
->shaderDenormPreserveFloat64
= true;
1660 p
->shaderRoundingModeRTEFloat64
= true;
1661 p
->shaderRoundingModeRTZFloat64
= true;
1662 p
->shaderSignedZeroInfNanPreserveFloat64
= true;
1664 /* It's a bit hard to exactly map our implementation to the limits
1665 * described here. The bindless surface handle in the extended
1666 * message descriptors is 20 bits and it's an index into the table of
1667 * RENDER_SURFACE_STATE structs that starts at bindless surface base
1668 * address. Given that most things consume two surface states per
1669 * view (general/sampled for textures and write-only/read-write for
1670 * images), we claim 2^19 things.
1672 * For SSBOs, we just use A64 messages so there is no real limit
1673 * there beyond the limit on the total size of a descriptor set.
1675 const unsigned max_bindless_views
= 1 << 19;
1676 p
->maxUpdateAfterBindDescriptorsInAllPools
= max_bindless_views
;
1677 p
->shaderUniformBufferArrayNonUniformIndexingNative
= false;
1678 p
->shaderSampledImageArrayNonUniformIndexingNative
= false;
1679 p
->shaderStorageBufferArrayNonUniformIndexingNative
= true;
1680 p
->shaderStorageImageArrayNonUniformIndexingNative
= false;
1681 p
->shaderInputAttachmentArrayNonUniformIndexingNative
= false;
1682 p
->robustBufferAccessUpdateAfterBind
= true;
1683 p
->quadDivergentImplicitLod
= false;
1684 p
->maxPerStageDescriptorUpdateAfterBindSamplers
= max_bindless_views
;
1685 p
->maxPerStageDescriptorUpdateAfterBindUniformBuffers
= MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
;
1686 p
->maxPerStageDescriptorUpdateAfterBindStorageBuffers
= UINT32_MAX
;
1687 p
->maxPerStageDescriptorUpdateAfterBindSampledImages
= max_bindless_views
;
1688 p
->maxPerStageDescriptorUpdateAfterBindStorageImages
= max_bindless_views
;
1689 p
->maxPerStageDescriptorUpdateAfterBindInputAttachments
= MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS
;
1690 p
->maxPerStageUpdateAfterBindResources
= UINT32_MAX
;
1691 p
->maxDescriptorSetUpdateAfterBindSamplers
= max_bindless_views
;
1692 p
->maxDescriptorSetUpdateAfterBindUniformBuffers
= 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
;
1693 p
->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2;
1694 p
->maxDescriptorSetUpdateAfterBindStorageBuffers
= UINT32_MAX
;
1695 p
->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2;
1696 p
->maxDescriptorSetUpdateAfterBindSampledImages
= max_bindless_views
;
1697 p
->maxDescriptorSetUpdateAfterBindStorageImages
= max_bindless_views
;
1698 p
->maxDescriptorSetUpdateAfterBindInputAttachments
= MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS
;
1700 /* We support all of the depth resolve modes */
1701 p
->supportedDepthResolveModes
= VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
|
1702 VK_RESOLVE_MODE_AVERAGE_BIT_KHR
|
1703 VK_RESOLVE_MODE_MIN_BIT_KHR
|
1704 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1705 /* Average doesn't make sense for stencil so we don't support that */
1706 p
->supportedStencilResolveModes
= VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
;
1707 if (pdevice
->info
.gen
>= 8) {
1708 /* The advanced stencil resolve modes currently require stencil
1709 * sampling be supported by the hardware.
1711 p
->supportedStencilResolveModes
|= VK_RESOLVE_MODE_MIN_BIT_KHR
|
1712 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1714 p
->independentResolveNone
= true;
1715 p
->independentResolve
= true;
1717 p
->filterMinmaxSingleComponentFormats
= pdevice
->info
.gen
>= 9;
1718 p
->filterMinmaxImageComponentMapping
= pdevice
->info
.gen
>= 9;
1720 p
->maxTimelineSemaphoreValueDifference
= UINT64_MAX
;
1722 p
->framebufferIntegerColorSampleCounts
=
1723 isl_device_get_sample_counts(&pdevice
->isl_dev
);
1726 void anv_GetPhysicalDeviceProperties2(
1727 VkPhysicalDevice physicalDevice
,
1728 VkPhysicalDeviceProperties2
* pProperties
)
1730 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1732 anv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
1734 VkPhysicalDeviceVulkan11Properties core_1_1
= {
1735 .sType
= VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES
,
1737 anv_get_physical_device_properties_1_1(pdevice
, &core_1_1
);
1739 VkPhysicalDeviceVulkan12Properties core_1_2
= {
1740 .sType
= VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES
,
1742 anv_get_physical_device_properties_1_2(pdevice
, &core_1_2
);
1744 #define CORE_RENAMED_PROPERTY(major, minor, ext_property, core_property) \
1745 memcpy(&properties->ext_property, &core_##major##_##minor.core_property, \
1746 sizeof(core_##major##_##minor.core_property))
1748 #define CORE_PROPERTY(major, minor, property) \
1749 CORE_RENAMED_PROPERTY(major, minor, property, property)
1751 vk_foreach_struct(ext
, pProperties
->pNext
) {
1752 switch (ext
->sType
) {
1753 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_STENCIL_RESOLVE_PROPERTIES_KHR
: {
1754 VkPhysicalDeviceDepthStencilResolvePropertiesKHR
*properties
=
1755 (VkPhysicalDeviceDepthStencilResolvePropertiesKHR
*)ext
;
1756 CORE_PROPERTY(1, 2, supportedDepthResolveModes
);
1757 CORE_PROPERTY(1, 2, supportedStencilResolveModes
);
1758 CORE_PROPERTY(1, 2, independentResolveNone
);
1759 CORE_PROPERTY(1, 2, independentResolve
);
1763 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES_EXT
: {
1764 VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*properties
=
1765 (VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*)ext
;
1766 CORE_PROPERTY(1, 2, maxUpdateAfterBindDescriptorsInAllPools
);
1767 CORE_PROPERTY(1, 2, shaderUniformBufferArrayNonUniformIndexingNative
);
1768 CORE_PROPERTY(1, 2, shaderSampledImageArrayNonUniformIndexingNative
);
1769 CORE_PROPERTY(1, 2, shaderStorageBufferArrayNonUniformIndexingNative
);
1770 CORE_PROPERTY(1, 2, shaderStorageImageArrayNonUniformIndexingNative
);
1771 CORE_PROPERTY(1, 2, shaderInputAttachmentArrayNonUniformIndexingNative
);
1772 CORE_PROPERTY(1, 2, robustBufferAccessUpdateAfterBind
);
1773 CORE_PROPERTY(1, 2, quadDivergentImplicitLod
);
1774 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindSamplers
);
1775 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindUniformBuffers
);
1776 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindStorageBuffers
);
1777 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindSampledImages
);
1778 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindStorageImages
);
1779 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindInputAttachments
);
1780 CORE_PROPERTY(1, 2, maxPerStageUpdateAfterBindResources
);
1781 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindSamplers
);
1782 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindUniformBuffers
);
1783 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindUniformBuffersDynamic
);
1784 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindStorageBuffers
);
1785 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindStorageBuffersDynamic
);
1786 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindSampledImages
);
1787 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindStorageImages
);
1788 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindInputAttachments
);
1792 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR
: {
1793 VkPhysicalDeviceDriverPropertiesKHR
*properties
=
1794 (VkPhysicalDeviceDriverPropertiesKHR
*) ext
;
1795 CORE_PROPERTY(1, 2, driverID
);
1796 CORE_PROPERTY(1, 2, driverName
);
1797 CORE_PROPERTY(1, 2, driverInfo
);
1798 CORE_PROPERTY(1, 2, conformanceVersion
);
1802 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT
: {
1803 VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*props
=
1804 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*) ext
;
1805 /* Userptr needs page aligned memory. */
1806 props
->minImportedHostPointerAlignment
= 4096;
1810 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES
: {
1811 VkPhysicalDeviceIDProperties
*properties
=
1812 (VkPhysicalDeviceIDProperties
*)ext
;
1813 CORE_PROPERTY(1, 1, deviceUUID
);
1814 CORE_PROPERTY(1, 1, driverUUID
);
1815 CORE_PROPERTY(1, 1, deviceLUID
);
1816 CORE_PROPERTY(1, 1, deviceLUIDValid
);
1820 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_PROPERTIES_EXT
: {
1821 VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*props
=
1822 (VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*)ext
;
1823 props
->maxInlineUniformBlockSize
= MAX_INLINE_UNIFORM_BLOCK_SIZE
;
1824 props
->maxPerStageDescriptorInlineUniformBlocks
=
1825 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1826 props
->maxPerStageDescriptorUpdateAfterBindInlineUniformBlocks
=
1827 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1828 props
->maxDescriptorSetInlineUniformBlocks
=
1829 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1830 props
->maxDescriptorSetUpdateAfterBindInlineUniformBlocks
=
1831 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1835 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_PROPERTIES_EXT
: {
1836 VkPhysicalDeviceLineRasterizationPropertiesEXT
*props
=
1837 (VkPhysicalDeviceLineRasterizationPropertiesEXT
*)ext
;
1838 /* In the Skylake PRM Vol. 7, subsection titled "GIQ (Diamond)
1839 * Sampling Rules - Legacy Mode", it says the following:
1841 * "Note that the device divides a pixel into a 16x16 array of
1842 * subpixels, referenced by their upper left corners."
1844 * This is the only known reference in the PRMs to the subpixel
1845 * precision of line rasterization and a "16x16 array of subpixels"
1846 * implies 4 subpixel precision bits. Empirical testing has shown
1847 * that 4 subpixel precision bits applies to all line rasterization
1850 props
->lineSubPixelPrecisionBits
= 4;
1854 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES
: {
1855 VkPhysicalDeviceMaintenance3Properties
*properties
=
1856 (VkPhysicalDeviceMaintenance3Properties
*)ext
;
1857 /* This value doesn't matter for us today as our per-stage
1858 * descriptors are the real limit.
1860 CORE_PROPERTY(1, 1, maxPerSetDescriptors
);
1861 CORE_PROPERTY(1, 1, maxMemoryAllocationSize
);
1865 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES
: {
1866 VkPhysicalDeviceMultiviewProperties
*properties
=
1867 (VkPhysicalDeviceMultiviewProperties
*)ext
;
1868 CORE_PROPERTY(1, 1, maxMultiviewViewCount
);
1869 CORE_PROPERTY(1, 1, maxMultiviewInstanceIndex
);
1873 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT
: {
1874 VkPhysicalDevicePCIBusInfoPropertiesEXT
*properties
=
1875 (VkPhysicalDevicePCIBusInfoPropertiesEXT
*)ext
;
1876 properties
->pciDomain
= pdevice
->pci_info
.domain
;
1877 properties
->pciBus
= pdevice
->pci_info
.bus
;
1878 properties
->pciDevice
= pdevice
->pci_info
.device
;
1879 properties
->pciFunction
= pdevice
->pci_info
.function
;
1883 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES
: {
1884 VkPhysicalDevicePointClippingProperties
*properties
=
1885 (VkPhysicalDevicePointClippingProperties
*) ext
;
1886 CORE_PROPERTY(1, 1, pointClippingBehavior
);
1890 #pragma GCC diagnostic push
1891 #pragma GCC diagnostic ignored "-Wswitch"
1892 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PRESENTATION_PROPERTIES_ANDROID
: {
1893 VkPhysicalDevicePresentationPropertiesANDROID
*props
=
1894 (VkPhysicalDevicePresentationPropertiesANDROID
*)ext
;
1895 props
->sharedImage
= VK_FALSE
;
1898 #pragma GCC diagnostic pop
1900 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES
: {
1901 VkPhysicalDeviceProtectedMemoryProperties
*properties
=
1902 (VkPhysicalDeviceProtectedMemoryProperties
*)ext
;
1903 CORE_PROPERTY(1, 1, protectedNoFault
);
1907 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
1908 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
1909 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
1910 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
1914 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT
: {
1915 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*properties
=
1916 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*)ext
;
1917 CORE_PROPERTY(1, 2, filterMinmaxImageComponentMapping
);
1918 CORE_PROPERTY(1, 2, filterMinmaxSingleComponentFormats
);
1922 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES
: {
1923 VkPhysicalDeviceSubgroupProperties
*properties
= (void *)ext
;
1924 CORE_PROPERTY(1, 1, subgroupSize
);
1925 CORE_RENAMED_PROPERTY(1, 1, supportedStages
,
1926 subgroupSupportedStages
);
1927 CORE_RENAMED_PROPERTY(1, 1, supportedOperations
,
1928 subgroupSupportedOperations
);
1929 CORE_RENAMED_PROPERTY(1, 1, quadOperationsInAllStages
,
1930 subgroupQuadOperationsInAllStages
);
1934 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_PROPERTIES_EXT
: {
1935 VkPhysicalDeviceSubgroupSizeControlPropertiesEXT
*props
=
1936 (VkPhysicalDeviceSubgroupSizeControlPropertiesEXT
*)ext
;
1937 STATIC_ASSERT(8 <= BRW_SUBGROUP_SIZE
&& BRW_SUBGROUP_SIZE
<= 32);
1938 props
->minSubgroupSize
= 8;
1939 props
->maxSubgroupSize
= 32;
1940 props
->maxComputeWorkgroupSubgroups
= pdevice
->info
.max_cs_threads
;
1941 props
->requiredSubgroupSizeStages
= VK_SHADER_STAGE_COMPUTE_BIT
;
1944 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT_CONTROLS_PROPERTIES_KHR
: {
1945 VkPhysicalDeviceFloatControlsPropertiesKHR
*properties
= (void *)ext
;
1946 CORE_PROPERTY(1, 2, denormBehaviorIndependence
);
1947 CORE_PROPERTY(1, 2, roundingModeIndependence
);
1948 CORE_PROPERTY(1, 2, shaderDenormFlushToZeroFloat16
);
1949 CORE_PROPERTY(1, 2, shaderDenormPreserveFloat16
);
1950 CORE_PROPERTY(1, 2, shaderRoundingModeRTEFloat16
);
1951 CORE_PROPERTY(1, 2, shaderRoundingModeRTZFloat16
);
1952 CORE_PROPERTY(1, 2, shaderSignedZeroInfNanPreserveFloat16
);
1953 CORE_PROPERTY(1, 2, shaderDenormFlushToZeroFloat32
);
1954 CORE_PROPERTY(1, 2, shaderDenormPreserveFloat32
);
1955 CORE_PROPERTY(1, 2, shaderRoundingModeRTEFloat32
);
1956 CORE_PROPERTY(1, 2, shaderRoundingModeRTZFloat32
);
1957 CORE_PROPERTY(1, 2, shaderSignedZeroInfNanPreserveFloat32
);
1958 CORE_PROPERTY(1, 2, shaderDenormFlushToZeroFloat64
);
1959 CORE_PROPERTY(1, 2, shaderDenormPreserveFloat64
);
1960 CORE_PROPERTY(1, 2, shaderRoundingModeRTEFloat64
);
1961 CORE_PROPERTY(1, 2, shaderRoundingModeRTZFloat64
);
1962 CORE_PROPERTY(1, 2, shaderSignedZeroInfNanPreserveFloat64
);
1966 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_PROPERTIES_EXT
: {
1967 VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT
*props
=
1968 (VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT
*)ext
;
1970 /* From the SKL PRM Vol. 2d, docs for RENDER_SURFACE_STATE::Surface
1973 * "For SURFTYPE_BUFFER non-rendertarget surfaces, this field
1974 * specifies the base address of the first element of the surface,
1975 * computed in software by adding the surface base address to the
1976 * byte offset of the element in the buffer. The base address must
1977 * be aligned to element size."
1979 * The typed dataport messages require that things be texel aligned.
1980 * Otherwise, we may just load/store the wrong data or, in the worst
1981 * case, there may be hangs.
1983 props
->storageTexelBufferOffsetAlignmentBytes
= 16;
1984 props
->storageTexelBufferOffsetSingleTexelAlignment
= true;
1986 /* The sampler, however, is much more forgiving and it can handle
1987 * arbitrary byte alignment for linear and buffer surfaces. It's
1988 * hard to find a good PRM citation for this but years of empirical
1989 * experience demonstrate that this is true.
1991 props
->uniformTexelBufferOffsetAlignmentBytes
= 1;
1992 props
->uniformTexelBufferOffsetSingleTexelAlignment
= false;
1996 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_PROPERTIES_KHR
: {
1997 VkPhysicalDeviceTimelineSemaphorePropertiesKHR
*properties
=
1998 (VkPhysicalDeviceTimelineSemaphorePropertiesKHR
*) ext
;
1999 CORE_PROPERTY(1, 2, maxTimelineSemaphoreValueDifference
);
2003 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT
: {
2004 VkPhysicalDeviceTransformFeedbackPropertiesEXT
*props
=
2005 (VkPhysicalDeviceTransformFeedbackPropertiesEXT
*)ext
;
2007 props
->maxTransformFeedbackStreams
= MAX_XFB_STREAMS
;
2008 props
->maxTransformFeedbackBuffers
= MAX_XFB_BUFFERS
;
2009 props
->maxTransformFeedbackBufferSize
= (1ull << 32);
2010 props
->maxTransformFeedbackStreamDataSize
= 128 * 4;
2011 props
->maxTransformFeedbackBufferDataSize
= 128 * 4;
2012 props
->maxTransformFeedbackBufferDataStride
= 2048;
2013 props
->transformFeedbackQueries
= true;
2014 props
->transformFeedbackStreamsLinesTriangles
= false;
2015 props
->transformFeedbackRasterizationStreamSelect
= false;
2016 props
->transformFeedbackDraw
= true;
2020 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT
: {
2021 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*props
=
2022 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*)ext
;
2023 /* We have to restrict this a bit for multiview */
2024 props
->maxVertexAttribDivisor
= UINT32_MAX
/ 16;
2028 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES
:
2029 anv_get_physical_device_properties_1_1(pdevice
, (void *)ext
);
2032 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES
:
2033 anv_get_physical_device_properties_1_2(pdevice
, (void *)ext
);
2037 anv_debug_ignored_stype(ext
->sType
);
2042 #undef CORE_RENAMED_PROPERTY
2043 #undef CORE_PROPERTY
2046 /* We support exactly one queue family. */
2047 static const VkQueueFamilyProperties
2048 anv_queue_family_properties
= {
2049 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
2050 VK_QUEUE_COMPUTE_BIT
|
2051 VK_QUEUE_TRANSFER_BIT
,
2053 .timestampValidBits
= 36, /* XXX: Real value here */
2054 .minImageTransferGranularity
= { 1, 1, 1 },
2057 void anv_GetPhysicalDeviceQueueFamilyProperties(
2058 VkPhysicalDevice physicalDevice
,
2060 VkQueueFamilyProperties
* pQueueFamilyProperties
)
2062 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pCount
);
2064 vk_outarray_append(&out
, p
) {
2065 *p
= anv_queue_family_properties
;
2069 void anv_GetPhysicalDeviceQueueFamilyProperties2(
2070 VkPhysicalDevice physicalDevice
,
2071 uint32_t* pQueueFamilyPropertyCount
,
2072 VkQueueFamilyProperties2
* pQueueFamilyProperties
)
2075 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pQueueFamilyPropertyCount
);
2077 vk_outarray_append(&out
, p
) {
2078 p
->queueFamilyProperties
= anv_queue_family_properties
;
2080 vk_foreach_struct(s
, p
->pNext
) {
2081 anv_debug_ignored_stype(s
->sType
);
2086 void anv_GetPhysicalDeviceMemoryProperties(
2087 VkPhysicalDevice physicalDevice
,
2088 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
2090 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
2092 pMemoryProperties
->memoryTypeCount
= physical_device
->memory
.type_count
;
2093 for (uint32_t i
= 0; i
< physical_device
->memory
.type_count
; i
++) {
2094 pMemoryProperties
->memoryTypes
[i
] = (VkMemoryType
) {
2095 .propertyFlags
= physical_device
->memory
.types
[i
].propertyFlags
,
2096 .heapIndex
= physical_device
->memory
.types
[i
].heapIndex
,
2100 pMemoryProperties
->memoryHeapCount
= physical_device
->memory
.heap_count
;
2101 for (uint32_t i
= 0; i
< physical_device
->memory
.heap_count
; i
++) {
2102 pMemoryProperties
->memoryHeaps
[i
] = (VkMemoryHeap
) {
2103 .size
= physical_device
->memory
.heaps
[i
].size
,
2104 .flags
= physical_device
->memory
.heaps
[i
].flags
,
2110 anv_get_memory_budget(VkPhysicalDevice physicalDevice
,
2111 VkPhysicalDeviceMemoryBudgetPropertiesEXT
*memoryBudget
)
2113 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
2114 uint64_t sys_available
= get_available_system_memory();
2115 assert(sys_available
> 0);
2117 VkDeviceSize total_heaps_size
= 0;
2118 for (size_t i
= 0; i
< device
->memory
.heap_count
; i
++)
2119 total_heaps_size
+= device
->memory
.heaps
[i
].size
;
2121 for (size_t i
= 0; i
< device
->memory
.heap_count
; i
++) {
2122 VkDeviceSize heap_size
= device
->memory
.heaps
[i
].size
;
2123 VkDeviceSize heap_used
= device
->memory
.heaps
[i
].used
;
2124 VkDeviceSize heap_budget
;
2126 double heap_proportion
= (double) heap_size
/ total_heaps_size
;
2127 VkDeviceSize sys_available_prop
= sys_available
* heap_proportion
;
2130 * Let's not incite the app to starve the system: report at most 90% of
2131 * available system memory.
2133 uint64_t heap_available
= sys_available_prop
* 9 / 10;
2134 heap_budget
= MIN2(heap_size
, heap_used
+ heap_available
);
2137 * Round down to the nearest MB
2139 heap_budget
&= ~((1ull << 20) - 1);
2142 * The heapBudget value must be non-zero for array elements less than
2143 * VkPhysicalDeviceMemoryProperties::memoryHeapCount. The heapBudget
2144 * value must be less than or equal to VkMemoryHeap::size for each heap.
2146 assert(0 < heap_budget
&& heap_budget
<= heap_size
);
2148 memoryBudget
->heapUsage
[i
] = heap_used
;
2149 memoryBudget
->heapBudget
[i
] = heap_budget
;
2152 /* The heapBudget and heapUsage values must be zero for array elements
2153 * greater than or equal to VkPhysicalDeviceMemoryProperties::memoryHeapCount
2155 for (uint32_t i
= device
->memory
.heap_count
; i
< VK_MAX_MEMORY_HEAPS
; i
++) {
2156 memoryBudget
->heapBudget
[i
] = 0;
2157 memoryBudget
->heapUsage
[i
] = 0;
2161 void anv_GetPhysicalDeviceMemoryProperties2(
2162 VkPhysicalDevice physicalDevice
,
2163 VkPhysicalDeviceMemoryProperties2
* pMemoryProperties
)
2165 anv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
2166 &pMemoryProperties
->memoryProperties
);
2168 vk_foreach_struct(ext
, pMemoryProperties
->pNext
) {
2169 switch (ext
->sType
) {
2170 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT
:
2171 anv_get_memory_budget(physicalDevice
, (void*)ext
);
2174 anv_debug_ignored_stype(ext
->sType
);
2181 anv_GetDeviceGroupPeerMemoryFeatures(
2184 uint32_t localDeviceIndex
,
2185 uint32_t remoteDeviceIndex
,
2186 VkPeerMemoryFeatureFlags
* pPeerMemoryFeatures
)
2188 assert(localDeviceIndex
== 0 && remoteDeviceIndex
== 0);
2189 *pPeerMemoryFeatures
= VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT
|
2190 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT
|
2191 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT
|
2192 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT
;
2195 PFN_vkVoidFunction
anv_GetInstanceProcAddr(
2196 VkInstance _instance
,
2199 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2201 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
2202 * when we have to return valid function pointers, NULL, or it's left
2203 * undefined. See the table for exact details.
2208 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
2209 if (strcmp(pName, "vk" #entrypoint) == 0) \
2210 return (PFN_vkVoidFunction)anv_##entrypoint
2212 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties
);
2213 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties
);
2214 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceVersion
);
2215 LOOKUP_ANV_ENTRYPOINT(CreateInstance
);
2217 #undef LOOKUP_ANV_ENTRYPOINT
2219 if (instance
== NULL
)
2222 int idx
= anv_get_instance_entrypoint_index(pName
);
2224 return instance
->dispatch
.entrypoints
[idx
];
2226 idx
= anv_get_physical_device_entrypoint_index(pName
);
2228 return instance
->physical_device_dispatch
.entrypoints
[idx
];
2230 idx
= anv_get_device_entrypoint_index(pName
);
2232 return instance
->device_dispatch
.entrypoints
[idx
];
2237 /* With version 1+ of the loader interface the ICD should expose
2238 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
2241 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
2242 VkInstance instance
,
2246 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
2247 VkInstance instance
,
2250 return anv_GetInstanceProcAddr(instance
, pName
);
2253 PFN_vkVoidFunction
anv_GetDeviceProcAddr(
2257 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2259 if (!device
|| !pName
)
2262 int idx
= anv_get_device_entrypoint_index(pName
);
2266 return device
->dispatch
.entrypoints
[idx
];
2269 /* With version 4+ of the loader interface the ICD should expose
2270 * vk_icdGetPhysicalDeviceProcAddr()
2273 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetPhysicalDeviceProcAddr(
2274 VkInstance _instance
,
2277 PFN_vkVoidFunction
vk_icdGetPhysicalDeviceProcAddr(
2278 VkInstance _instance
,
2281 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2283 if (!pName
|| !instance
)
2286 int idx
= anv_get_physical_device_entrypoint_index(pName
);
2290 return instance
->physical_device_dispatch
.entrypoints
[idx
];
2295 anv_CreateDebugReportCallbackEXT(VkInstance _instance
,
2296 const VkDebugReportCallbackCreateInfoEXT
* pCreateInfo
,
2297 const VkAllocationCallbacks
* pAllocator
,
2298 VkDebugReportCallbackEXT
* pCallback
)
2300 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2301 return vk_create_debug_report_callback(&instance
->debug_report_callbacks
,
2302 pCreateInfo
, pAllocator
, &instance
->alloc
,
2307 anv_DestroyDebugReportCallbackEXT(VkInstance _instance
,
2308 VkDebugReportCallbackEXT _callback
,
2309 const VkAllocationCallbacks
* pAllocator
)
2311 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2312 vk_destroy_debug_report_callback(&instance
->debug_report_callbacks
,
2313 _callback
, pAllocator
, &instance
->alloc
);
2317 anv_DebugReportMessageEXT(VkInstance _instance
,
2318 VkDebugReportFlagsEXT flags
,
2319 VkDebugReportObjectTypeEXT objectType
,
2322 int32_t messageCode
,
2323 const char* pLayerPrefix
,
2324 const char* pMessage
)
2326 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2327 vk_debug_report(&instance
->debug_report_callbacks
, flags
, objectType
,
2328 object
, location
, messageCode
, pLayerPrefix
, pMessage
);
2331 static struct anv_state
2332 anv_state_pool_emit_data(struct anv_state_pool
*pool
, size_t size
, size_t align
, const void *p
)
2334 struct anv_state state
;
2336 state
= anv_state_pool_alloc(pool
, size
, align
);
2337 memcpy(state
.map
, p
, size
);
2342 /* Haswell border color is a bit of a disaster. Float and unorm formats use a
2343 * straightforward 32-bit float color in the first 64 bytes. Instead of using
2344 * a nice float/integer union like Gen8+, Haswell specifies the integer border
2345 * color as a separate entry /after/ the float color. The layout of this entry
2346 * also depends on the format's bpp (with extra hacks for RG32), and overlaps.
2348 * Since we don't know the format/bpp, we can't make any of the border colors
2349 * containing '1' work for all formats, as it would be in the wrong place for
2350 * some of them. We opt to make 32-bit integers work as this seems like the
2351 * most common option. Fortunately, transparent black works regardless, as
2352 * all zeroes is the same in every bit-size.
2354 struct hsw_border_color
{
2358 uint32_t _pad1
[108];
2361 struct gen8_border_color
{
2366 /* Pad out to 64 bytes */
2371 anv_device_init_border_colors(struct anv_device
*device
)
2373 if (device
->info
.is_haswell
) {
2374 static const struct hsw_border_color border_colors
[] = {
2375 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
2376 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
2377 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
2378 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
2379 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
2380 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
2383 device
->border_colors
=
2384 anv_state_pool_emit_data(&device
->dynamic_state_pool
,
2385 sizeof(border_colors
), 512, border_colors
);
2387 static const struct gen8_border_color border_colors
[] = {
2388 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
2389 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
2390 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
2391 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
2392 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
2393 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
2396 device
->border_colors
=
2397 anv_state_pool_emit_data(&device
->dynamic_state_pool
,
2398 sizeof(border_colors
), 64, border_colors
);
2403 anv_device_init_trivial_batch(struct anv_device
*device
)
2405 VkResult result
= anv_device_alloc_bo(device
, 4096,
2406 ANV_BO_ALLOC_MAPPED
,
2407 0 /* explicit_address */,
2408 &device
->trivial_batch_bo
);
2409 if (result
!= VK_SUCCESS
)
2412 struct anv_batch batch
= {
2413 .start
= device
->trivial_batch_bo
->map
,
2414 .next
= device
->trivial_batch_bo
->map
,
2415 .end
= device
->trivial_batch_bo
->map
+ 4096,
2418 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
2419 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
2421 if (!device
->info
.has_llc
)
2422 gen_clflush_range(batch
.start
, batch
.next
- batch
.start
);
2427 VkResult
anv_EnumerateDeviceExtensionProperties(
2428 VkPhysicalDevice physicalDevice
,
2429 const char* pLayerName
,
2430 uint32_t* pPropertyCount
,
2431 VkExtensionProperties
* pProperties
)
2433 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
2434 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
2436 for (int i
= 0; i
< ANV_DEVICE_EXTENSION_COUNT
; i
++) {
2437 if (device
->supported_extensions
.extensions
[i
]) {
2438 vk_outarray_append(&out
, prop
) {
2439 *prop
= anv_device_extensions
[i
];
2444 return vk_outarray_status(&out
);
2448 anv_device_init_dispatch(struct anv_device
*device
)
2450 const struct anv_instance
*instance
= device
->physical
->instance
;
2452 const struct anv_device_dispatch_table
*genX_table
;
2453 switch (device
->info
.gen
) {
2455 genX_table
= &gen12_device_dispatch_table
;
2458 genX_table
= &gen11_device_dispatch_table
;
2461 genX_table
= &gen10_device_dispatch_table
;
2464 genX_table
= &gen9_device_dispatch_table
;
2467 genX_table
= &gen8_device_dispatch_table
;
2470 if (device
->info
.is_haswell
)
2471 genX_table
= &gen75_device_dispatch_table
;
2473 genX_table
= &gen7_device_dispatch_table
;
2476 unreachable("unsupported gen\n");
2479 for (unsigned i
= 0; i
< ARRAY_SIZE(device
->dispatch
.entrypoints
); i
++) {
2480 /* Vulkan requires that entrypoints for extensions which have not been
2481 * enabled must not be advertised.
2483 if (!anv_device_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
2484 &instance
->enabled_extensions
,
2485 &device
->enabled_extensions
)) {
2486 device
->dispatch
.entrypoints
[i
] = NULL
;
2487 } else if (genX_table
->entrypoints
[i
]) {
2488 device
->dispatch
.entrypoints
[i
] = genX_table
->entrypoints
[i
];
2490 device
->dispatch
.entrypoints
[i
] =
2491 anv_device_dispatch_table
.entrypoints
[i
];
2497 vk_priority_to_gen(int priority
)
2500 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT
:
2501 return GEN_CONTEXT_LOW_PRIORITY
;
2502 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
:
2503 return GEN_CONTEXT_MEDIUM_PRIORITY
;
2504 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT
:
2505 return GEN_CONTEXT_HIGH_PRIORITY
;
2506 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT
:
2507 return GEN_CONTEXT_REALTIME_PRIORITY
;
2509 unreachable("Invalid priority");
2514 anv_device_init_hiz_clear_value_bo(struct anv_device
*device
)
2516 VkResult result
= anv_device_alloc_bo(device
, 4096,
2517 ANV_BO_ALLOC_MAPPED
,
2518 0 /* explicit_address */,
2519 &device
->hiz_clear_bo
);
2520 if (result
!= VK_SUCCESS
)
2523 union isl_color_value hiz_clear
= { .u32
= { 0, } };
2524 hiz_clear
.f32
[0] = ANV_HZ_FC_VAL
;
2526 memcpy(device
->hiz_clear_bo
->map
, hiz_clear
.u32
, sizeof(hiz_clear
.u32
));
2528 if (!device
->info
.has_llc
)
2529 gen_clflush_range(device
->hiz_clear_bo
->map
, sizeof(hiz_clear
.u32
));
2535 get_bo_from_pool(struct gen_batch_decode_bo
*ret
,
2536 struct anv_block_pool
*pool
,
2539 anv_block_pool_foreach_bo(bo
, pool
) {
2540 uint64_t bo_address
= gen_48b_address(bo
->offset
);
2541 if (address
>= bo_address
&& address
< (bo_address
+ bo
->size
)) {
2542 *ret
= (struct gen_batch_decode_bo
) {
2553 /* Finding a buffer for batch decoding */
2554 static struct gen_batch_decode_bo
2555 decode_get_bo(void *v_batch
, bool ppgtt
, uint64_t address
)
2557 struct anv_device
*device
= v_batch
;
2558 struct gen_batch_decode_bo ret_bo
= {};
2562 if (get_bo_from_pool(&ret_bo
, &device
->dynamic_state_pool
.block_pool
, address
))
2564 if (get_bo_from_pool(&ret_bo
, &device
->instruction_state_pool
.block_pool
, address
))
2566 if (get_bo_from_pool(&ret_bo
, &device
->binding_table_pool
.block_pool
, address
))
2568 if (get_bo_from_pool(&ret_bo
, &device
->surface_state_pool
.block_pool
, address
))
2571 if (!device
->cmd_buffer_being_decoded
)
2572 return (struct gen_batch_decode_bo
) { };
2574 struct anv_batch_bo
**bo
;
2576 u_vector_foreach(bo
, &device
->cmd_buffer_being_decoded
->seen_bbos
) {
2577 /* The decoder zeroes out the top 16 bits, so we need to as well */
2578 uint64_t bo_address
= (*bo
)->bo
->offset
& (~0ull >> 16);
2580 if (address
>= bo_address
&& address
< bo_address
+ (*bo
)->bo
->size
) {
2581 return (struct gen_batch_decode_bo
) {
2583 .size
= (*bo
)->bo
->size
,
2584 .map
= (*bo
)->bo
->map
,
2589 return (struct gen_batch_decode_bo
) { };
2592 struct gen_aux_map_buffer
{
2593 struct gen_buffer base
;
2594 struct anv_state state
;
2597 static struct gen_buffer
*
2598 gen_aux_map_buffer_alloc(void *driver_ctx
, uint32_t size
)
2600 struct gen_aux_map_buffer
*buf
= malloc(sizeof(struct gen_aux_map_buffer
));
2604 struct anv_device
*device
= (struct anv_device
*)driver_ctx
;
2605 assert(device
->physical
->supports_48bit_addresses
&&
2606 device
->physical
->use_softpin
);
2608 struct anv_state_pool
*pool
= &device
->dynamic_state_pool
;
2609 buf
->state
= anv_state_pool_alloc(pool
, size
, size
);
2611 buf
->base
.gpu
= pool
->block_pool
.bo
->offset
+ buf
->state
.offset
;
2612 buf
->base
.gpu_end
= buf
->base
.gpu
+ buf
->state
.alloc_size
;
2613 buf
->base
.map
= buf
->state
.map
;
2614 buf
->base
.driver_bo
= &buf
->state
;
2619 gen_aux_map_buffer_free(void *driver_ctx
, struct gen_buffer
*buffer
)
2621 struct gen_aux_map_buffer
*buf
= (struct gen_aux_map_buffer
*)buffer
;
2622 struct anv_device
*device
= (struct anv_device
*)driver_ctx
;
2623 struct anv_state_pool
*pool
= &device
->dynamic_state_pool
;
2624 anv_state_pool_free(pool
, buf
->state
);
2628 static struct gen_mapped_pinned_buffer_alloc aux_map_allocator
= {
2629 .alloc
= gen_aux_map_buffer_alloc
,
2630 .free
= gen_aux_map_buffer_free
,
2634 check_physical_device_features(VkPhysicalDevice physicalDevice
,
2635 const VkPhysicalDeviceFeatures
*features
)
2637 VkPhysicalDeviceFeatures supported_features
;
2638 anv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
2639 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
2640 VkBool32
*enabled_feature
= (VkBool32
*)features
;
2641 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
2642 for (uint32_t i
= 0; i
< num_features
; i
++) {
2643 if (enabled_feature
[i
] && !supported_feature
[i
])
2644 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
2650 VkResult
anv_CreateDevice(
2651 VkPhysicalDevice physicalDevice
,
2652 const VkDeviceCreateInfo
* pCreateInfo
,
2653 const VkAllocationCallbacks
* pAllocator
,
2656 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
2658 struct anv_device
*device
;
2660 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO
);
2662 struct anv_device_extension_table enabled_extensions
= { };
2663 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
2665 for (idx
= 0; idx
< ANV_DEVICE_EXTENSION_COUNT
; idx
++) {
2666 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
2667 anv_device_extensions
[idx
].extensionName
) == 0)
2671 if (idx
>= ANV_DEVICE_EXTENSION_COUNT
)
2672 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
2674 if (!physical_device
->supported_extensions
.extensions
[idx
])
2675 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
2677 enabled_extensions
.extensions
[idx
] = true;
2680 /* Check enabled features */
2681 bool robust_buffer_access
= false;
2682 if (pCreateInfo
->pEnabledFeatures
) {
2683 result
= check_physical_device_features(physicalDevice
,
2684 pCreateInfo
->pEnabledFeatures
);
2685 if (result
!= VK_SUCCESS
)
2688 if (pCreateInfo
->pEnabledFeatures
->robustBufferAccess
)
2689 robust_buffer_access
= true;
2692 vk_foreach_struct_const(ext
, pCreateInfo
->pNext
) {
2693 switch (ext
->sType
) {
2694 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2
: {
2695 const VkPhysicalDeviceFeatures2
*features
= (const void *)ext
;
2696 result
= check_physical_device_features(physicalDevice
,
2697 &features
->features
);
2698 if (result
!= VK_SUCCESS
)
2701 if (features
->features
.robustBufferAccess
)
2702 robust_buffer_access
= true;
2712 /* Check requested queues and fail if we are requested to create any
2713 * queues with flags we don't support.
2715 assert(pCreateInfo
->queueCreateInfoCount
> 0);
2716 for (uint32_t i
= 0; i
< pCreateInfo
->queueCreateInfoCount
; i
++) {
2717 if (pCreateInfo
->pQueueCreateInfos
[i
].flags
!= 0)
2718 return vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2721 /* Check if client specified queue priority. */
2722 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*queue_priority
=
2723 vk_find_struct_const(pCreateInfo
->pQueueCreateInfos
[0].pNext
,
2724 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT
);
2726 VkQueueGlobalPriorityEXT priority
=
2727 queue_priority
? queue_priority
->globalPriority
:
2728 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
;
2730 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
2732 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
2734 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2736 if (INTEL_DEBUG
& DEBUG_BATCH
) {
2737 const unsigned decode_flags
=
2738 GEN_BATCH_DECODE_FULL
|
2739 ((INTEL_DEBUG
& DEBUG_COLOR
) ? GEN_BATCH_DECODE_IN_COLOR
: 0) |
2740 GEN_BATCH_DECODE_OFFSETS
|
2741 GEN_BATCH_DECODE_FLOATS
;
2743 gen_batch_decode_ctx_init(&device
->decoder_ctx
,
2744 &physical_device
->info
,
2745 stderr
, decode_flags
, NULL
,
2746 decode_get_bo
, NULL
, device
);
2749 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
2750 device
->physical
= physical_device
;
2751 device
->no_hw
= physical_device
->no_hw
;
2752 device
->_lost
= false;
2755 device
->alloc
= *pAllocator
;
2757 device
->alloc
= physical_device
->instance
->alloc
;
2759 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
2760 device
->fd
= open(physical_device
->path
, O_RDWR
| O_CLOEXEC
);
2761 if (device
->fd
== -1) {
2762 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2766 device
->context_id
= anv_gem_create_context(device
);
2767 if (device
->context_id
== -1) {
2768 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2772 result
= anv_queue_init(device
, &device
->queue
);
2773 if (result
!= VK_SUCCESS
)
2774 goto fail_context_id
;
2776 if (physical_device
->use_softpin
) {
2777 if (pthread_mutex_init(&device
->vma_mutex
, NULL
) != 0) {
2778 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2782 /* keep the page with address zero out of the allocator */
2783 util_vma_heap_init(&device
->vma_lo
,
2784 LOW_HEAP_MIN_ADDRESS
, LOW_HEAP_SIZE
);
2786 util_vma_heap_init(&device
->vma_cva
, CLIENT_VISIBLE_HEAP_MIN_ADDRESS
,
2787 CLIENT_VISIBLE_HEAP_SIZE
);
2789 /* Leave the last 4GiB out of the high vma range, so that no state
2790 * base address + size can overflow 48 bits. For more information see
2791 * the comment about Wa32bitGeneralStateOffset in anv_allocator.c
2793 util_vma_heap_init(&device
->vma_hi
, HIGH_HEAP_MIN_ADDRESS
,
2794 physical_device
->gtt_size
- (1ull << 32) -
2795 HIGH_HEAP_MIN_ADDRESS
);
2798 list_inithead(&device
->memory_objects
);
2800 /* As per spec, the driver implementation may deny requests to acquire
2801 * a priority above the default priority (MEDIUM) if the caller does not
2802 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
2805 if (physical_device
->has_context_priority
) {
2806 int err
= anv_gem_set_context_param(device
->fd
, device
->context_id
,
2807 I915_CONTEXT_PARAM_PRIORITY
,
2808 vk_priority_to_gen(priority
));
2809 if (err
!= 0 && priority
> VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
) {
2810 result
= vk_error(VK_ERROR_NOT_PERMITTED_EXT
);
2815 device
->info
= physical_device
->info
;
2816 device
->isl_dev
= physical_device
->isl_dev
;
2818 /* On Broadwell and later, we can use batch chaining to more efficiently
2819 * implement growing command buffers. Prior to Haswell, the kernel
2820 * command parser gets in the way and we have to fall back to growing
2823 device
->can_chain_batches
= device
->info
.gen
>= 8;
2825 device
->robust_buffer_access
= robust_buffer_access
;
2826 device
->enabled_extensions
= enabled_extensions
;
2828 anv_device_init_dispatch(device
);
2830 if (pthread_mutex_init(&device
->mutex
, NULL
) != 0) {
2831 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2835 pthread_condattr_t condattr
;
2836 if (pthread_condattr_init(&condattr
) != 0) {
2837 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2840 if (pthread_condattr_setclock(&condattr
, CLOCK_MONOTONIC
) != 0) {
2841 pthread_condattr_destroy(&condattr
);
2842 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2845 if (pthread_cond_init(&device
->queue_submit
, &condattr
) != 0) {
2846 pthread_condattr_destroy(&condattr
);
2847 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2850 pthread_condattr_destroy(&condattr
);
2852 result
= anv_bo_cache_init(&device
->bo_cache
);
2853 if (result
!= VK_SUCCESS
)
2854 goto fail_queue_cond
;
2856 anv_bo_pool_init(&device
->batch_bo_pool
, device
);
2858 result
= anv_state_pool_init(&device
->dynamic_state_pool
, device
,
2859 DYNAMIC_STATE_POOL_MIN_ADDRESS
, 16384);
2860 if (result
!= VK_SUCCESS
)
2861 goto fail_batch_bo_pool
;
2863 result
= anv_state_pool_init(&device
->instruction_state_pool
, device
,
2864 INSTRUCTION_STATE_POOL_MIN_ADDRESS
, 16384);
2865 if (result
!= VK_SUCCESS
)
2866 goto fail_dynamic_state_pool
;
2868 result
= anv_state_pool_init(&device
->surface_state_pool
, device
,
2869 SURFACE_STATE_POOL_MIN_ADDRESS
, 4096);
2870 if (result
!= VK_SUCCESS
)
2871 goto fail_instruction_state_pool
;
2873 if (physical_device
->use_softpin
) {
2874 result
= anv_state_pool_init(&device
->binding_table_pool
, device
,
2875 BINDING_TABLE_POOL_MIN_ADDRESS
, 4096);
2876 if (result
!= VK_SUCCESS
)
2877 goto fail_surface_state_pool
;
2880 if (device
->info
.gen
>= 12) {
2881 device
->aux_map_ctx
= gen_aux_map_init(device
, &aux_map_allocator
,
2882 &physical_device
->info
);
2883 if (!device
->aux_map_ctx
)
2884 goto fail_binding_table_pool
;
2887 result
= anv_device_alloc_bo(device
, 4096, 0 /* flags */,
2888 0 /* explicit_address */,
2889 &device
->workaround_bo
);
2890 if (result
!= VK_SUCCESS
)
2891 goto fail_surface_aux_map_pool
;
2893 result
= anv_device_init_trivial_batch(device
);
2894 if (result
!= VK_SUCCESS
)
2895 goto fail_workaround_bo
;
2897 if (device
->info
.gen
>= 10) {
2898 result
= anv_device_init_hiz_clear_value_bo(device
);
2899 if (result
!= VK_SUCCESS
)
2900 goto fail_trivial_batch_bo
;
2903 anv_scratch_pool_init(device
, &device
->scratch_pool
);
2905 switch (device
->info
.gen
) {
2907 if (!device
->info
.is_haswell
)
2908 result
= gen7_init_device_state(device
);
2910 result
= gen75_init_device_state(device
);
2913 result
= gen8_init_device_state(device
);
2916 result
= gen9_init_device_state(device
);
2919 result
= gen10_init_device_state(device
);
2922 result
= gen11_init_device_state(device
);
2925 result
= gen12_init_device_state(device
);
2928 /* Shouldn't get here as we don't create physical devices for any other
2930 unreachable("unhandled gen");
2932 if (result
!= VK_SUCCESS
)
2933 goto fail_workaround_bo
;
2935 anv_pipeline_cache_init(&device
->default_pipeline_cache
, device
, true);
2937 anv_device_init_blorp(device
);
2939 anv_device_init_border_colors(device
);
2941 anv_device_perf_init(device
);
2943 *pDevice
= anv_device_to_handle(device
);
2948 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
2949 if (device
->info
.gen
>= 10)
2950 anv_device_release_bo(device
, device
->hiz_clear_bo
);
2951 anv_device_release_bo(device
, device
->workaround_bo
);
2952 fail_trivial_batch_bo
:
2953 anv_device_release_bo(device
, device
->trivial_batch_bo
);
2954 fail_surface_aux_map_pool
:
2955 if (device
->info
.gen
>= 12) {
2956 gen_aux_map_finish(device
->aux_map_ctx
);
2957 device
->aux_map_ctx
= NULL
;
2959 fail_binding_table_pool
:
2960 if (physical_device
->use_softpin
)
2961 anv_state_pool_finish(&device
->binding_table_pool
);
2962 fail_surface_state_pool
:
2963 anv_state_pool_finish(&device
->surface_state_pool
);
2964 fail_instruction_state_pool
:
2965 anv_state_pool_finish(&device
->instruction_state_pool
);
2966 fail_dynamic_state_pool
:
2967 anv_state_pool_finish(&device
->dynamic_state_pool
);
2969 anv_bo_pool_finish(&device
->batch_bo_pool
);
2970 anv_bo_cache_finish(&device
->bo_cache
);
2972 pthread_cond_destroy(&device
->queue_submit
);
2974 pthread_mutex_destroy(&device
->mutex
);
2976 if (physical_device
->use_softpin
) {
2977 util_vma_heap_finish(&device
->vma_hi
);
2978 util_vma_heap_finish(&device
->vma_cva
);
2979 util_vma_heap_finish(&device
->vma_lo
);
2982 anv_queue_finish(&device
->queue
);
2984 anv_gem_destroy_context(device
, device
->context_id
);
2988 vk_free(&device
->alloc
, device
);
2993 void anv_DestroyDevice(
2995 const VkAllocationCallbacks
* pAllocator
)
2997 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3002 anv_device_finish_blorp(device
);
3004 anv_pipeline_cache_finish(&device
->default_pipeline_cache
);
3006 anv_queue_finish(&device
->queue
);
3008 #ifdef HAVE_VALGRIND
3009 /* We only need to free these to prevent valgrind errors. The backing
3010 * BO will go away in a couple of lines so we don't actually leak.
3012 anv_state_pool_free(&device
->dynamic_state_pool
, device
->border_colors
);
3013 anv_state_pool_free(&device
->dynamic_state_pool
, device
->slice_hash
);
3016 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
3018 anv_device_release_bo(device
, device
->workaround_bo
);
3019 anv_device_release_bo(device
, device
->trivial_batch_bo
);
3020 if (device
->info
.gen
>= 10)
3021 anv_device_release_bo(device
, device
->hiz_clear_bo
);
3023 if (device
->info
.gen
>= 12) {
3024 gen_aux_map_finish(device
->aux_map_ctx
);
3025 device
->aux_map_ctx
= NULL
;
3028 if (device
->physical
->use_softpin
)
3029 anv_state_pool_finish(&device
->binding_table_pool
);
3030 anv_state_pool_finish(&device
->surface_state_pool
);
3031 anv_state_pool_finish(&device
->instruction_state_pool
);
3032 anv_state_pool_finish(&device
->dynamic_state_pool
);
3034 anv_bo_pool_finish(&device
->batch_bo_pool
);
3036 anv_bo_cache_finish(&device
->bo_cache
);
3038 if (device
->physical
->use_softpin
) {
3039 util_vma_heap_finish(&device
->vma_hi
);
3040 util_vma_heap_finish(&device
->vma_cva
);
3041 util_vma_heap_finish(&device
->vma_lo
);
3044 pthread_cond_destroy(&device
->queue_submit
);
3045 pthread_mutex_destroy(&device
->mutex
);
3047 anv_gem_destroy_context(device
, device
->context_id
);
3049 if (INTEL_DEBUG
& DEBUG_BATCH
)
3050 gen_batch_decode_ctx_finish(&device
->decoder_ctx
);
3054 vk_free(&device
->alloc
, device
);
3057 VkResult
anv_EnumerateInstanceLayerProperties(
3058 uint32_t* pPropertyCount
,
3059 VkLayerProperties
* pProperties
)
3061 if (pProperties
== NULL
) {
3062 *pPropertyCount
= 0;
3066 /* None supported at this time */
3067 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
3070 VkResult
anv_EnumerateDeviceLayerProperties(
3071 VkPhysicalDevice physicalDevice
,
3072 uint32_t* pPropertyCount
,
3073 VkLayerProperties
* pProperties
)
3075 if (pProperties
== NULL
) {
3076 *pPropertyCount
= 0;
3080 /* None supported at this time */
3081 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
3084 void anv_GetDeviceQueue(
3086 uint32_t queueNodeIndex
,
3087 uint32_t queueIndex
,
3090 const VkDeviceQueueInfo2 info
= {
3091 .sType
= VK_STRUCTURE_TYPE_DEVICE_QUEUE_INFO_2
,
3094 .queueFamilyIndex
= queueNodeIndex
,
3095 .queueIndex
= queueIndex
,
3098 anv_GetDeviceQueue2(_device
, &info
, pQueue
);
3101 void anv_GetDeviceQueue2(
3103 const VkDeviceQueueInfo2
* pQueueInfo
,
3106 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3108 assert(pQueueInfo
->queueIndex
== 0);
3110 if (pQueueInfo
->flags
== device
->queue
.flags
)
3111 *pQueue
= anv_queue_to_handle(&device
->queue
);
3117 _anv_device_set_lost(struct anv_device
*device
,
3118 const char *file
, int line
,
3119 const char *msg
, ...)
3124 p_atomic_inc(&device
->_lost
);
3127 err
= __vk_errorv(device
->physical
->instance
, device
,
3128 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT
,
3129 VK_ERROR_DEVICE_LOST
, file
, line
, msg
, ap
);
3132 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
3139 _anv_queue_set_lost(struct anv_queue
*queue
,
3140 const char *file
, int line
,
3141 const char *msg
, ...)
3146 p_atomic_inc(&queue
->device
->_lost
);
3149 err
= __vk_errorv(queue
->device
->physical
->instance
, queue
->device
,
3150 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT
,
3151 VK_ERROR_DEVICE_LOST
, file
, line
, msg
, ap
);
3154 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
3161 anv_device_query_status(struct anv_device
*device
)
3163 /* This isn't likely as most of the callers of this function already check
3164 * for it. However, it doesn't hurt to check and it potentially lets us
3167 if (anv_device_is_lost(device
))
3168 return VK_ERROR_DEVICE_LOST
;
3170 uint32_t active
, pending
;
3171 int ret
= anv_gem_gpu_get_reset_stats(device
, &active
, &pending
);
3173 /* We don't know the real error. */
3174 return anv_device_set_lost(device
, "get_reset_stats failed: %m");
3178 return anv_device_set_lost(device
, "GPU hung on one of our command buffers");
3179 } else if (pending
) {
3180 return anv_device_set_lost(device
, "GPU hung with commands in-flight");
3187 anv_device_bo_busy(struct anv_device
*device
, struct anv_bo
*bo
)
3189 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
3190 * Other usages of the BO (such as on different hardware) will not be
3191 * flagged as "busy" by this ioctl. Use with care.
3193 int ret
= anv_gem_busy(device
, bo
->gem_handle
);
3195 return VK_NOT_READY
;
3196 } else if (ret
== -1) {
3197 /* We don't know the real error. */
3198 return anv_device_set_lost(device
, "gem wait failed: %m");
3201 /* Query for device status after the busy call. If the BO we're checking
3202 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
3203 * client because it clearly doesn't have valid data. Yes, this most
3204 * likely means an ioctl, but we just did an ioctl to query the busy status
3205 * so it's no great loss.
3207 return anv_device_query_status(device
);
3211 anv_device_wait(struct anv_device
*device
, struct anv_bo
*bo
,
3214 int ret
= anv_gem_wait(device
, bo
->gem_handle
, &timeout
);
3215 if (ret
== -1 && errno
== ETIME
) {
3217 } else if (ret
== -1) {
3218 /* We don't know the real error. */
3219 return anv_device_set_lost(device
, "gem wait failed: %m");
3222 /* Query for device status after the wait. If the BO we're waiting on got
3223 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
3224 * because it clearly doesn't have valid data. Yes, this most likely means
3225 * an ioctl, but we just did an ioctl to wait so it's no great loss.
3227 return anv_device_query_status(device
);
3230 VkResult
anv_DeviceWaitIdle(
3233 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3235 if (anv_device_is_lost(device
))
3236 return VK_ERROR_DEVICE_LOST
;
3238 return anv_queue_submit_simple_batch(&device
->queue
, NULL
);
3242 anv_vma_alloc(struct anv_device
*device
,
3243 uint64_t size
, uint64_t align
,
3244 enum anv_bo_alloc_flags alloc_flags
,
3245 uint64_t client_address
)
3247 pthread_mutex_lock(&device
->vma_mutex
);
3251 if (alloc_flags
& ANV_BO_ALLOC_CLIENT_VISIBLE_ADDRESS
) {
3252 if (client_address
) {
3253 if (util_vma_heap_alloc_addr(&device
->vma_cva
,
3254 client_address
, size
)) {
3255 addr
= client_address
;
3258 addr
= util_vma_heap_alloc(&device
->vma_cva
, size
, align
);
3260 /* We don't want to fall back to other heaps */
3264 assert(client_address
== 0);
3266 if (!(alloc_flags
& ANV_BO_ALLOC_32BIT_ADDRESS
))
3267 addr
= util_vma_heap_alloc(&device
->vma_hi
, size
, align
);
3270 addr
= util_vma_heap_alloc(&device
->vma_lo
, size
, align
);
3273 pthread_mutex_unlock(&device
->vma_mutex
);
3275 assert(addr
== gen_48b_address(addr
));
3276 return gen_canonical_address(addr
);
3280 anv_vma_free(struct anv_device
*device
,
3281 uint64_t address
, uint64_t size
)
3283 const uint64_t addr_48b
= gen_48b_address(address
);
3285 pthread_mutex_lock(&device
->vma_mutex
);
3287 if (addr_48b
>= LOW_HEAP_MIN_ADDRESS
&&
3288 addr_48b
<= LOW_HEAP_MAX_ADDRESS
) {
3289 util_vma_heap_free(&device
->vma_lo
, addr_48b
, size
);
3290 } else if (addr_48b
>= CLIENT_VISIBLE_HEAP_MIN_ADDRESS
&&
3291 addr_48b
<= CLIENT_VISIBLE_HEAP_MAX_ADDRESS
) {
3292 util_vma_heap_free(&device
->vma_cva
, addr_48b
, size
);
3294 assert(addr_48b
>= HIGH_HEAP_MIN_ADDRESS
);
3295 util_vma_heap_free(&device
->vma_hi
, addr_48b
, size
);
3298 pthread_mutex_unlock(&device
->vma_mutex
);
3301 VkResult
anv_AllocateMemory(
3303 const VkMemoryAllocateInfo
* pAllocateInfo
,
3304 const VkAllocationCallbacks
* pAllocator
,
3305 VkDeviceMemory
* pMem
)
3307 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3308 struct anv_physical_device
*pdevice
= device
->physical
;
3309 struct anv_device_memory
*mem
;
3310 VkResult result
= VK_SUCCESS
;
3312 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
3314 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
3315 assert(pAllocateInfo
->allocationSize
> 0);
3317 VkDeviceSize aligned_alloc_size
=
3318 align_u64(pAllocateInfo
->allocationSize
, 4096);
3320 if (aligned_alloc_size
> MAX_MEMORY_ALLOCATION_SIZE
)
3321 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
3323 assert(pAllocateInfo
->memoryTypeIndex
< pdevice
->memory
.type_count
);
3324 struct anv_memory_type
*mem_type
=
3325 &pdevice
->memory
.types
[pAllocateInfo
->memoryTypeIndex
];
3326 assert(mem_type
->heapIndex
< pdevice
->memory
.heap_count
);
3327 struct anv_memory_heap
*mem_heap
=
3328 &pdevice
->memory
.heaps
[mem_type
->heapIndex
];
3330 uint64_t mem_heap_used
= p_atomic_read(&mem_heap
->used
);
3331 if (mem_heap_used
+ aligned_alloc_size
> mem_heap
->size
)
3332 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
3334 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
3335 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3337 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3339 assert(pAllocateInfo
->memoryTypeIndex
< pdevice
->memory
.type_count
);
3340 mem
->type
= mem_type
;
3344 mem
->host_ptr
= NULL
;
3346 enum anv_bo_alloc_flags alloc_flags
= 0;
3348 const VkExportMemoryAllocateInfo
*export_info
= NULL
;
3349 const VkImportAndroidHardwareBufferInfoANDROID
*ahw_import_info
= NULL
;
3350 const VkImportMemoryFdInfoKHR
*fd_info
= NULL
;
3351 const VkImportMemoryHostPointerInfoEXT
*host_ptr_info
= NULL
;
3352 const VkMemoryDedicatedAllocateInfo
*dedicated_info
= NULL
;
3353 VkMemoryAllocateFlags vk_flags
= 0;
3354 uint64_t client_address
= 0;
3356 vk_foreach_struct_const(ext
, pAllocateInfo
->pNext
) {
3357 switch (ext
->sType
) {
3358 case VK_STRUCTURE_TYPE_EXPORT_MEMORY_ALLOCATE_INFO
:
3359 export_info
= (void *)ext
;
3362 case VK_STRUCTURE_TYPE_IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID
:
3363 ahw_import_info
= (void *)ext
;
3366 case VK_STRUCTURE_TYPE_IMPORT_MEMORY_FD_INFO_KHR
:
3367 fd_info
= (void *)ext
;
3370 case VK_STRUCTURE_TYPE_IMPORT_MEMORY_HOST_POINTER_INFO_EXT
:
3371 host_ptr_info
= (void *)ext
;
3374 case VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_FLAGS_INFO
: {
3375 const VkMemoryAllocateFlagsInfo
*flags_info
= (void *)ext
;
3376 vk_flags
= flags_info
->flags
;
3380 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_ALLOCATE_INFO
:
3381 dedicated_info
= (void *)ext
;
3384 case VK_STRUCTURE_TYPE_MEMORY_OPAQUE_CAPTURE_ADDRESS_ALLOCATE_INFO_KHR
: {
3385 const VkMemoryOpaqueCaptureAddressAllocateInfoKHR
*addr_info
=
3386 (const VkMemoryOpaqueCaptureAddressAllocateInfoKHR
*)ext
;
3387 client_address
= addr_info
->opaqueCaptureAddress
;
3392 anv_debug_ignored_stype(ext
->sType
);
3397 /* By default, we want all VkDeviceMemory objects to support CCS */
3398 if (device
->physical
->has_implicit_ccs
)
3399 alloc_flags
|= ANV_BO_ALLOC_IMPLICIT_CCS
;
3401 if (vk_flags
& VK_MEMORY_ALLOCATE_DEVICE_ADDRESS_BIT_KHR
)
3402 alloc_flags
|= ANV_BO_ALLOC_CLIENT_VISIBLE_ADDRESS
;
3404 if ((export_info
&& export_info
->handleTypes
) ||
3405 (fd_info
&& fd_info
->handleType
) ||
3406 (host_ptr_info
&& host_ptr_info
->handleType
)) {
3407 /* Anything imported or exported is EXTERNAL */
3408 alloc_flags
|= ANV_BO_ALLOC_EXTERNAL
;
3410 /* We can't have implicit CCS on external memory with an AUX-table.
3411 * Doing so would require us to sync the aux tables across processes
3412 * which is impractical.
3414 if (device
->info
.has_aux_map
)
3415 alloc_flags
&= ~ANV_BO_ALLOC_IMPLICIT_CCS
;
3418 /* Check if we need to support Android HW buffer export. If so,
3419 * create AHardwareBuffer and import memory from it.
3421 bool android_export
= false;
3422 if (export_info
&& export_info
->handleTypes
&
3423 VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID
)
3424 android_export
= true;
3426 if (ahw_import_info
) {
3427 result
= anv_import_ahw_memory(_device
, mem
, ahw_import_info
);
3428 if (result
!= VK_SUCCESS
)
3432 } else if (android_export
) {
3433 result
= anv_create_ahw_memory(_device
, mem
, pAllocateInfo
);
3434 if (result
!= VK_SUCCESS
)
3437 const VkImportAndroidHardwareBufferInfoANDROID import_info
= {
3440 result
= anv_import_ahw_memory(_device
, mem
, &import_info
);
3441 if (result
!= VK_SUCCESS
)
3447 /* The Vulkan spec permits handleType to be 0, in which case the struct is
3450 if (fd_info
&& fd_info
->handleType
) {
3451 /* At the moment, we support only the below handle types. */
3452 assert(fd_info
->handleType
==
3453 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
3454 fd_info
->handleType
==
3455 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
3457 result
= anv_device_import_bo(device
, fd_info
->fd
, alloc_flags
,
3458 client_address
, &mem
->bo
);
3459 if (result
!= VK_SUCCESS
)
3462 /* For security purposes, we reject importing the bo if it's smaller
3463 * than the requested allocation size. This prevents a malicious client
3464 * from passing a buffer to a trusted client, lying about the size, and
3465 * telling the trusted client to try and texture from an image that goes
3466 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
3467 * in the trusted client. The trusted client can protect itself against
3468 * this sort of attack but only if it can trust the buffer size.
3470 if (mem
->bo
->size
< aligned_alloc_size
) {
3471 result
= vk_errorf(device
, device
, VK_ERROR_INVALID_EXTERNAL_HANDLE
,
3472 "aligned allocationSize too large for "
3473 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT: "
3474 "%"PRIu64
"B > %"PRIu64
"B",
3475 aligned_alloc_size
, mem
->bo
->size
);
3476 anv_device_release_bo(device
, mem
->bo
);
3480 /* From the Vulkan spec:
3482 * "Importing memory from a file descriptor transfers ownership of
3483 * the file descriptor from the application to the Vulkan
3484 * implementation. The application must not perform any operations on
3485 * the file descriptor after a successful import."
3487 * If the import fails, we leave the file descriptor open.
3493 if (host_ptr_info
&& host_ptr_info
->handleType
) {
3494 if (host_ptr_info
->handleType
==
3495 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_MAPPED_FOREIGN_MEMORY_BIT_EXT
) {
3496 result
= vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
3500 assert(host_ptr_info
->handleType
==
3501 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
);
3503 result
= anv_device_import_bo_from_host_ptr(device
,
3504 host_ptr_info
->pHostPointer
,
3505 pAllocateInfo
->allocationSize
,
3509 if (result
!= VK_SUCCESS
)
3512 mem
->host_ptr
= host_ptr_info
->pHostPointer
;
3516 /* Regular allocate (not importing memory). */
3518 result
= anv_device_alloc_bo(device
, pAllocateInfo
->allocationSize
,
3519 alloc_flags
, client_address
, &mem
->bo
);
3520 if (result
!= VK_SUCCESS
)
3523 if (dedicated_info
&& dedicated_info
->image
!= VK_NULL_HANDLE
) {
3524 ANV_FROM_HANDLE(anv_image
, image
, dedicated_info
->image
);
3526 /* Some legacy (non-modifiers) consumers need the tiling to be set on
3527 * the BO. In this case, we have a dedicated allocation.
3529 if (image
->needs_set_tiling
) {
3530 const uint32_t i915_tiling
=
3531 isl_tiling_to_i915_tiling(image
->planes
[0].surface
.isl
.tiling
);
3532 int ret
= anv_gem_set_tiling(device
, mem
->bo
->gem_handle
,
3533 image
->planes
[0].surface
.isl
.row_pitch_B
,
3536 anv_device_release_bo(device
, mem
->bo
);
3537 result
= vk_errorf(device
, device
, VK_ERROR_OUT_OF_DEVICE_MEMORY
,
3538 "failed to set BO tiling: %m");
3545 mem_heap_used
= p_atomic_add_return(&mem_heap
->used
, mem
->bo
->size
);
3546 if (mem_heap_used
> mem_heap
->size
) {
3547 p_atomic_add(&mem_heap
->used
, -mem
->bo
->size
);
3548 anv_device_release_bo(device
, mem
->bo
);
3549 result
= vk_errorf(device
, device
, VK_ERROR_OUT_OF_DEVICE_MEMORY
,
3550 "Out of heap memory");
3554 pthread_mutex_lock(&device
->mutex
);
3555 list_addtail(&mem
->link
, &device
->memory_objects
);
3556 pthread_mutex_unlock(&device
->mutex
);
3558 *pMem
= anv_device_memory_to_handle(mem
);
3563 vk_free2(&device
->alloc
, pAllocator
, mem
);
3568 VkResult
anv_GetMemoryFdKHR(
3570 const VkMemoryGetFdInfoKHR
* pGetFdInfo
,
3573 ANV_FROM_HANDLE(anv_device
, dev
, device_h
);
3574 ANV_FROM_HANDLE(anv_device_memory
, mem
, pGetFdInfo
->memory
);
3576 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
3578 assert(pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
3579 pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
3581 return anv_device_export_bo(dev
, mem
->bo
, pFd
);
3584 VkResult
anv_GetMemoryFdPropertiesKHR(
3586 VkExternalMemoryHandleTypeFlagBits handleType
,
3588 VkMemoryFdPropertiesKHR
* pMemoryFdProperties
)
3590 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3592 switch (handleType
) {
3593 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
:
3594 /* dma-buf can be imported as any memory type */
3595 pMemoryFdProperties
->memoryTypeBits
=
3596 (1 << device
->physical
->memory
.type_count
) - 1;
3600 /* The valid usage section for this function says:
3602 * "handleType must not be one of the handle types defined as
3605 * So opaque handle types fall into the default "unsupported" case.
3607 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
3611 VkResult
anv_GetMemoryHostPointerPropertiesEXT(
3613 VkExternalMemoryHandleTypeFlagBits handleType
,
3614 const void* pHostPointer
,
3615 VkMemoryHostPointerPropertiesEXT
* pMemoryHostPointerProperties
)
3617 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3619 assert(pMemoryHostPointerProperties
->sType
==
3620 VK_STRUCTURE_TYPE_MEMORY_HOST_POINTER_PROPERTIES_EXT
);
3622 switch (handleType
) {
3623 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
:
3624 /* Host memory can be imported as any memory type. */
3625 pMemoryHostPointerProperties
->memoryTypeBits
=
3626 (1ull << device
->physical
->memory
.type_count
) - 1;
3631 return VK_ERROR_INVALID_EXTERNAL_HANDLE
;
3635 void anv_FreeMemory(
3637 VkDeviceMemory _mem
,
3638 const VkAllocationCallbacks
* pAllocator
)
3640 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3641 ANV_FROM_HANDLE(anv_device_memory
, mem
, _mem
);
3646 pthread_mutex_lock(&device
->mutex
);
3647 list_del(&mem
->link
);
3648 pthread_mutex_unlock(&device
->mutex
);
3651 anv_UnmapMemory(_device
, _mem
);
3653 p_atomic_add(&device
->physical
->memory
.heaps
[mem
->type
->heapIndex
].used
,
3656 anv_device_release_bo(device
, mem
->bo
);
3658 #if defined(ANDROID) && ANDROID_API_LEVEL >= 26
3660 AHardwareBuffer_release(mem
->ahw
);
3663 vk_free2(&device
->alloc
, pAllocator
, mem
);
3666 VkResult
anv_MapMemory(
3668 VkDeviceMemory _memory
,
3669 VkDeviceSize offset
,
3671 VkMemoryMapFlags flags
,
3674 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3675 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
3682 if (mem
->host_ptr
) {
3683 *ppData
= mem
->host_ptr
+ offset
;
3687 if (size
== VK_WHOLE_SIZE
)
3688 size
= mem
->bo
->size
- offset
;
3690 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
3692 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
3693 * assert(size != 0);
3694 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
3695 * equal to the size of the memory minus offset
3698 assert(offset
+ size
<= mem
->bo
->size
);
3700 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
3701 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
3702 * at a time is valid. We could just mmap up front and return an offset
3703 * pointer here, but that may exhaust virtual memory on 32 bit
3706 uint32_t gem_flags
= 0;
3708 if (!device
->info
.has_llc
&&
3709 (mem
->type
->propertyFlags
& VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
))
3710 gem_flags
|= I915_MMAP_WC
;
3712 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
3713 uint64_t map_offset
= offset
& ~4095ull;
3714 assert(offset
>= map_offset
);
3715 uint64_t map_size
= (offset
+ size
) - map_offset
;
3717 /* Let's map whole pages */
3718 map_size
= align_u64(map_size
, 4096);
3720 void *map
= anv_gem_mmap(device
, mem
->bo
->gem_handle
,
3721 map_offset
, map_size
, gem_flags
);
3722 if (map
== MAP_FAILED
)
3723 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
3726 mem
->map_size
= map_size
;
3728 *ppData
= mem
->map
+ (offset
- map_offset
);
3733 void anv_UnmapMemory(
3735 VkDeviceMemory _memory
)
3737 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
3739 if (mem
== NULL
|| mem
->host_ptr
)
3742 anv_gem_munmap(mem
->map
, mem
->map_size
);
3749 clflush_mapped_ranges(struct anv_device
*device
,
3751 const VkMappedMemoryRange
*ranges
)
3753 for (uint32_t i
= 0; i
< count
; i
++) {
3754 ANV_FROM_HANDLE(anv_device_memory
, mem
, ranges
[i
].memory
);
3755 if (ranges
[i
].offset
>= mem
->map_size
)
3758 gen_clflush_range(mem
->map
+ ranges
[i
].offset
,
3759 MIN2(ranges
[i
].size
, mem
->map_size
- ranges
[i
].offset
));
3763 VkResult
anv_FlushMappedMemoryRanges(
3765 uint32_t memoryRangeCount
,
3766 const VkMappedMemoryRange
* pMemoryRanges
)
3768 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3770 if (device
->info
.has_llc
)
3773 /* Make sure the writes we're flushing have landed. */
3774 __builtin_ia32_mfence();
3776 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
3781 VkResult
anv_InvalidateMappedMemoryRanges(
3783 uint32_t memoryRangeCount
,
3784 const VkMappedMemoryRange
* pMemoryRanges
)
3786 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3788 if (device
->info
.has_llc
)
3791 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
3793 /* Make sure no reads get moved up above the invalidate. */
3794 __builtin_ia32_mfence();
3799 void anv_GetBufferMemoryRequirements(
3802 VkMemoryRequirements
* pMemoryRequirements
)
3804 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
3805 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3807 /* The Vulkan spec (git aaed022) says:
3809 * memoryTypeBits is a bitfield and contains one bit set for every
3810 * supported memory type for the resource. The bit `1<<i` is set if and
3811 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3812 * structure for the physical device is supported.
3814 uint32_t memory_types
= (1ull << device
->physical
->memory
.type_count
) - 1;
3816 /* Base alignment requirement of a cache line */
3817 uint32_t alignment
= 16;
3819 /* We need an alignment of 32 for pushing UBOs */
3820 if (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
)
3821 alignment
= MAX2(alignment
, 32);
3823 pMemoryRequirements
->size
= buffer
->size
;
3824 pMemoryRequirements
->alignment
= alignment
;
3826 /* Storage and Uniform buffers should have their size aligned to
3827 * 32-bits to avoid boundary checks when last DWord is not complete.
3828 * This would ensure that not internal padding would be needed for
3831 if (device
->robust_buffer_access
&&
3832 (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
||
3833 buffer
->usage
& VK_BUFFER_USAGE_STORAGE_BUFFER_BIT
))
3834 pMemoryRequirements
->size
= align_u64(buffer
->size
, 4);
3836 pMemoryRequirements
->memoryTypeBits
= memory_types
;
3839 void anv_GetBufferMemoryRequirements2(
3841 const VkBufferMemoryRequirementsInfo2
* pInfo
,
3842 VkMemoryRequirements2
* pMemoryRequirements
)
3844 anv_GetBufferMemoryRequirements(_device
, pInfo
->buffer
,
3845 &pMemoryRequirements
->memoryRequirements
);
3847 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3848 switch (ext
->sType
) {
3849 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
3850 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
3851 requirements
->prefersDedicatedAllocation
= false;
3852 requirements
->requiresDedicatedAllocation
= false;
3857 anv_debug_ignored_stype(ext
->sType
);
3863 void anv_GetImageMemoryRequirements(
3866 VkMemoryRequirements
* pMemoryRequirements
)
3868 ANV_FROM_HANDLE(anv_image
, image
, _image
);
3869 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3871 /* The Vulkan spec (git aaed022) says:
3873 * memoryTypeBits is a bitfield and contains one bit set for every
3874 * supported memory type for the resource. The bit `1<<i` is set if and
3875 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3876 * structure for the physical device is supported.
3878 * All types are currently supported for images.
3880 uint32_t memory_types
= (1ull << device
->physical
->memory
.type_count
) - 1;
3882 /* We must have image allocated or imported at this point. According to the
3883 * specification, external images must have been bound to memory before
3884 * calling GetImageMemoryRequirements.
3886 assert(image
->size
> 0);
3888 pMemoryRequirements
->size
= image
->size
;
3889 pMemoryRequirements
->alignment
= image
->alignment
;
3890 pMemoryRequirements
->memoryTypeBits
= memory_types
;
3893 void anv_GetImageMemoryRequirements2(
3895 const VkImageMemoryRequirementsInfo2
* pInfo
,
3896 VkMemoryRequirements2
* pMemoryRequirements
)
3898 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3899 ANV_FROM_HANDLE(anv_image
, image
, pInfo
->image
);
3901 anv_GetImageMemoryRequirements(_device
, pInfo
->image
,
3902 &pMemoryRequirements
->memoryRequirements
);
3904 vk_foreach_struct_const(ext
, pInfo
->pNext
) {
3905 switch (ext
->sType
) {
3906 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO
: {
3907 const VkImagePlaneMemoryRequirementsInfo
*plane_reqs
=
3908 (const VkImagePlaneMemoryRequirementsInfo
*) ext
;
3909 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
,
3910 plane_reqs
->planeAspect
);
3912 assert(image
->planes
[plane
].offset
== 0);
3914 /* The Vulkan spec (git aaed022) says:
3916 * memoryTypeBits is a bitfield and contains one bit set for every
3917 * supported memory type for the resource. The bit `1<<i` is set
3918 * if and only if the memory type `i` in the
3919 * VkPhysicalDeviceMemoryProperties structure for the physical
3920 * device is supported.
3922 * All types are currently supported for images.
3924 pMemoryRequirements
->memoryRequirements
.memoryTypeBits
=
3925 (1ull << device
->physical
->memory
.type_count
) - 1;
3927 /* We must have image allocated or imported at this point. According to the
3928 * specification, external images must have been bound to memory before
3929 * calling GetImageMemoryRequirements.
3931 assert(image
->planes
[plane
].size
> 0);
3933 pMemoryRequirements
->memoryRequirements
.size
= image
->planes
[plane
].size
;
3934 pMemoryRequirements
->memoryRequirements
.alignment
=
3935 image
->planes
[plane
].alignment
;
3940 anv_debug_ignored_stype(ext
->sType
);
3945 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3946 switch (ext
->sType
) {
3947 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
3948 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
3949 if (image
->needs_set_tiling
|| image
->external_format
) {
3950 /* If we need to set the tiling for external consumers, we need a
3951 * dedicated allocation.
3953 * See also anv_AllocateMemory.
3955 requirements
->prefersDedicatedAllocation
= true;
3956 requirements
->requiresDedicatedAllocation
= true;
3958 requirements
->prefersDedicatedAllocation
= false;
3959 requirements
->requiresDedicatedAllocation
= false;
3965 anv_debug_ignored_stype(ext
->sType
);
3971 void anv_GetImageSparseMemoryRequirements(
3974 uint32_t* pSparseMemoryRequirementCount
,
3975 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
3977 *pSparseMemoryRequirementCount
= 0;
3980 void anv_GetImageSparseMemoryRequirements2(
3982 const VkImageSparseMemoryRequirementsInfo2
* pInfo
,
3983 uint32_t* pSparseMemoryRequirementCount
,
3984 VkSparseImageMemoryRequirements2
* pSparseMemoryRequirements
)
3986 *pSparseMemoryRequirementCount
= 0;
3989 void anv_GetDeviceMemoryCommitment(
3991 VkDeviceMemory memory
,
3992 VkDeviceSize
* pCommittedMemoryInBytes
)
3994 *pCommittedMemoryInBytes
= 0;
3998 anv_bind_buffer_memory(const VkBindBufferMemoryInfo
*pBindInfo
)
4000 ANV_FROM_HANDLE(anv_device_memory
, mem
, pBindInfo
->memory
);
4001 ANV_FROM_HANDLE(anv_buffer
, buffer
, pBindInfo
->buffer
);
4003 assert(pBindInfo
->sType
== VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
);
4006 buffer
->address
= (struct anv_address
) {
4008 .offset
= pBindInfo
->memoryOffset
,
4011 buffer
->address
= ANV_NULL_ADDRESS
;
4015 VkResult
anv_BindBufferMemory(
4018 VkDeviceMemory memory
,
4019 VkDeviceSize memoryOffset
)
4021 anv_bind_buffer_memory(
4022 &(VkBindBufferMemoryInfo
) {
4023 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
,
4026 .memoryOffset
= memoryOffset
,
4032 VkResult
anv_BindBufferMemory2(
4034 uint32_t bindInfoCount
,
4035 const VkBindBufferMemoryInfo
* pBindInfos
)
4037 for (uint32_t i
= 0; i
< bindInfoCount
; i
++)
4038 anv_bind_buffer_memory(&pBindInfos
[i
]);
4043 VkResult
anv_QueueBindSparse(
4045 uint32_t bindInfoCount
,
4046 const VkBindSparseInfo
* pBindInfo
,
4049 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
4050 if (anv_device_is_lost(queue
->device
))
4051 return VK_ERROR_DEVICE_LOST
;
4053 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
4058 VkResult
anv_CreateEvent(
4060 const VkEventCreateInfo
* pCreateInfo
,
4061 const VkAllocationCallbacks
* pAllocator
,
4064 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4065 struct anv_state state
;
4066 struct anv_event
*event
;
4068 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_EVENT_CREATE_INFO
);
4070 state
= anv_state_pool_alloc(&device
->dynamic_state_pool
,
4073 event
->state
= state
;
4074 event
->semaphore
= VK_EVENT_RESET
;
4076 if (!device
->info
.has_llc
) {
4077 /* Make sure the writes we're flushing have landed. */
4078 __builtin_ia32_mfence();
4079 __builtin_ia32_clflush(event
);
4082 *pEvent
= anv_event_to_handle(event
);
4087 void anv_DestroyEvent(
4090 const VkAllocationCallbacks
* pAllocator
)
4092 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4093 ANV_FROM_HANDLE(anv_event
, event
, _event
);
4098 anv_state_pool_free(&device
->dynamic_state_pool
, event
->state
);
4101 VkResult
anv_GetEventStatus(
4105 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4106 ANV_FROM_HANDLE(anv_event
, event
, _event
);
4108 if (anv_device_is_lost(device
))
4109 return VK_ERROR_DEVICE_LOST
;
4111 if (!device
->info
.has_llc
) {
4112 /* Invalidate read cache before reading event written by GPU. */
4113 __builtin_ia32_clflush(event
);
4114 __builtin_ia32_mfence();
4118 return event
->semaphore
;
4121 VkResult
anv_SetEvent(
4125 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4126 ANV_FROM_HANDLE(anv_event
, event
, _event
);
4128 event
->semaphore
= VK_EVENT_SET
;
4130 if (!device
->info
.has_llc
) {
4131 /* Make sure the writes we're flushing have landed. */
4132 __builtin_ia32_mfence();
4133 __builtin_ia32_clflush(event
);
4139 VkResult
anv_ResetEvent(
4143 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4144 ANV_FROM_HANDLE(anv_event
, event
, _event
);
4146 event
->semaphore
= VK_EVENT_RESET
;
4148 if (!device
->info
.has_llc
) {
4149 /* Make sure the writes we're flushing have landed. */
4150 __builtin_ia32_mfence();
4151 __builtin_ia32_clflush(event
);
4159 VkResult
anv_CreateBuffer(
4161 const VkBufferCreateInfo
* pCreateInfo
,
4162 const VkAllocationCallbacks
* pAllocator
,
4165 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4166 struct anv_buffer
*buffer
;
4168 /* Don't allow creating buffers bigger than our address space. The real
4169 * issue here is that we may align up the buffer size and we don't want
4170 * doing so to cause roll-over. However, no one has any business
4171 * allocating a buffer larger than our GTT size.
4173 if (pCreateInfo
->size
> device
->physical
->gtt_size
)
4174 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
4176 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
4178 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
4179 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
4181 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
4183 buffer
->size
= pCreateInfo
->size
;
4184 buffer
->usage
= pCreateInfo
->usage
;
4185 buffer
->address
= ANV_NULL_ADDRESS
;
4187 *pBuffer
= anv_buffer_to_handle(buffer
);
4192 void anv_DestroyBuffer(
4195 const VkAllocationCallbacks
* pAllocator
)
4197 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4198 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
4203 vk_free2(&device
->alloc
, pAllocator
, buffer
);
4206 VkDeviceAddress
anv_GetBufferDeviceAddress(
4208 const VkBufferDeviceAddressInfoKHR
* pInfo
)
4210 ANV_FROM_HANDLE(anv_buffer
, buffer
, pInfo
->buffer
);
4212 assert(!anv_address_is_null(buffer
->address
));
4213 assert(buffer
->address
.bo
->flags
& EXEC_OBJECT_PINNED
);
4215 return anv_address_physical(buffer
->address
);
4218 uint64_t anv_GetBufferOpaqueCaptureAddress(
4220 const VkBufferDeviceAddressInfoKHR
* pInfo
)
4225 uint64_t anv_GetDeviceMemoryOpaqueCaptureAddress(
4227 const VkDeviceMemoryOpaqueCaptureAddressInfoKHR
* pInfo
)
4229 ANV_FROM_HANDLE(anv_device_memory
, memory
, pInfo
->memory
);
4231 assert(memory
->bo
->flags
& EXEC_OBJECT_PINNED
);
4232 assert(memory
->bo
->has_client_visible_address
);
4234 return gen_48b_address(memory
->bo
->offset
);
4238 anv_fill_buffer_surface_state(struct anv_device
*device
, struct anv_state state
,
4239 enum isl_format format
,
4240 struct anv_address address
,
4241 uint32_t range
, uint32_t stride
)
4243 isl_buffer_fill_state(&device
->isl_dev
, state
.map
,
4244 .address
= anv_address_physical(address
),
4245 .mocs
= device
->isl_dev
.mocs
.internal
,
4248 .swizzle
= ISL_SWIZZLE_IDENTITY
,
4249 .stride_B
= stride
);
4252 void anv_DestroySampler(
4255 const VkAllocationCallbacks
* pAllocator
)
4257 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4258 ANV_FROM_HANDLE(anv_sampler
, sampler
, _sampler
);
4263 if (sampler
->bindless_state
.map
) {
4264 anv_state_pool_free(&device
->dynamic_state_pool
,
4265 sampler
->bindless_state
);
4268 vk_free2(&device
->alloc
, pAllocator
, sampler
);
4271 VkResult
anv_CreateFramebuffer(
4273 const VkFramebufferCreateInfo
* pCreateInfo
,
4274 const VkAllocationCallbacks
* pAllocator
,
4275 VkFramebuffer
* pFramebuffer
)
4277 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4278 struct anv_framebuffer
*framebuffer
;
4280 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
4282 size_t size
= sizeof(*framebuffer
);
4284 /* VK_KHR_imageless_framebuffer extension says:
4286 * If flags includes VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR,
4287 * parameter pAttachments is ignored.
4289 if (!(pCreateInfo
->flags
& VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR
)) {
4290 size
+= sizeof(struct anv_image_view
*) * pCreateInfo
->attachmentCount
;
4291 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
4292 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
4293 if (framebuffer
== NULL
)
4294 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
4296 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
4297 ANV_FROM_HANDLE(anv_image_view
, iview
, pCreateInfo
->pAttachments
[i
]);
4298 framebuffer
->attachments
[i
] = iview
;
4300 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
4302 assert(device
->enabled_extensions
.KHR_imageless_framebuffer
);
4303 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
4304 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
4305 if (framebuffer
== NULL
)
4306 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
4308 framebuffer
->attachment_count
= 0;
4311 framebuffer
->width
= pCreateInfo
->width
;
4312 framebuffer
->height
= pCreateInfo
->height
;
4313 framebuffer
->layers
= pCreateInfo
->layers
;
4315 *pFramebuffer
= anv_framebuffer_to_handle(framebuffer
);
4320 void anv_DestroyFramebuffer(
4323 const VkAllocationCallbacks
* pAllocator
)
4325 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4326 ANV_FROM_HANDLE(anv_framebuffer
, fb
, _fb
);
4331 vk_free2(&device
->alloc
, pAllocator
, fb
);
4334 static const VkTimeDomainEXT anv_time_domains
[] = {
4335 VK_TIME_DOMAIN_DEVICE_EXT
,
4336 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
,
4337 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
,
4340 VkResult
anv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
4341 VkPhysicalDevice physicalDevice
,
4342 uint32_t *pTimeDomainCount
,
4343 VkTimeDomainEXT
*pTimeDomains
)
4346 VK_OUTARRAY_MAKE(out
, pTimeDomains
, pTimeDomainCount
);
4348 for (d
= 0; d
< ARRAY_SIZE(anv_time_domains
); d
++) {
4349 vk_outarray_append(&out
, i
) {
4350 *i
= anv_time_domains
[d
];
4354 return vk_outarray_status(&out
);
4358 anv_clock_gettime(clockid_t clock_id
)
4360 struct timespec current
;
4363 ret
= clock_gettime(clock_id
, ¤t
);
4364 if (ret
< 0 && clock_id
== CLOCK_MONOTONIC_RAW
)
4365 ret
= clock_gettime(CLOCK_MONOTONIC
, ¤t
);
4369 return (uint64_t) current
.tv_sec
* 1000000000ULL + current
.tv_nsec
;
4372 #define TIMESTAMP 0x2358
4374 VkResult
anv_GetCalibratedTimestampsEXT(
4376 uint32_t timestampCount
,
4377 const VkCalibratedTimestampInfoEXT
*pTimestampInfos
,
4378 uint64_t *pTimestamps
,
4379 uint64_t *pMaxDeviation
)
4381 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4382 uint64_t timestamp_frequency
= device
->info
.timestamp_frequency
;
4385 uint64_t begin
, end
;
4386 uint64_t max_clock_period
= 0;
4388 begin
= anv_clock_gettime(CLOCK_MONOTONIC_RAW
);
4390 for (d
= 0; d
< timestampCount
; d
++) {
4391 switch (pTimestampInfos
[d
].timeDomain
) {
4392 case VK_TIME_DOMAIN_DEVICE_EXT
:
4393 ret
= anv_gem_reg_read(device
, TIMESTAMP
| 1,
4397 return anv_device_set_lost(device
, "Failed to read the TIMESTAMP "
4400 uint64_t device_period
= DIV_ROUND_UP(1000000000, timestamp_frequency
);
4401 max_clock_period
= MAX2(max_clock_period
, device_period
);
4403 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
:
4404 pTimestamps
[d
] = anv_clock_gettime(CLOCK_MONOTONIC
);
4405 max_clock_period
= MAX2(max_clock_period
, 1);
4408 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
:
4409 pTimestamps
[d
] = begin
;
4417 end
= anv_clock_gettime(CLOCK_MONOTONIC_RAW
);
4420 * The maximum deviation is the sum of the interval over which we
4421 * perform the sampling and the maximum period of any sampled
4422 * clock. That's because the maximum skew between any two sampled
4423 * clock edges is when the sampled clock with the largest period is
4424 * sampled at the end of that period but right at the beginning of the
4425 * sampling interval and some other clock is sampled right at the
4426 * begining of its sampling period and right at the end of the
4427 * sampling interval. Let's assume the GPU has the longest clock
4428 * period and that the application is sampling GPU and monotonic:
4431 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
4432 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4436 * GPU -----_____-----_____-----_____-----_____
4439 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
4440 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4442 * Interval <----------------->
4443 * Deviation <-------------------------->
4447 * m = read(monotonic) 2
4450 * We round the sample interval up by one tick to cover sampling error
4451 * in the interval clock
4454 uint64_t sample_interval
= end
- begin
+ 1;
4456 *pMaxDeviation
= sample_interval
+ max_clock_period
;
4461 /* vk_icd.h does not declare this function, so we declare it here to
4462 * suppress Wmissing-prototypes.
4464 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4465 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
);
4467 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4468 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
)
4470 /* For the full details on loader interface versioning, see
4471 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
4472 * What follows is a condensed summary, to help you navigate the large and
4473 * confusing official doc.
4475 * - Loader interface v0 is incompatible with later versions. We don't
4478 * - In loader interface v1:
4479 * - The first ICD entrypoint called by the loader is
4480 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
4482 * - The ICD must statically expose no other Vulkan symbol unless it is
4483 * linked with -Bsymbolic.
4484 * - Each dispatchable Vulkan handle created by the ICD must be
4485 * a pointer to a struct whose first member is VK_LOADER_DATA. The
4486 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
4487 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
4488 * vkDestroySurfaceKHR(). The ICD must be capable of working with
4489 * such loader-managed surfaces.
4491 * - Loader interface v2 differs from v1 in:
4492 * - The first ICD entrypoint called by the loader is
4493 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
4494 * statically expose this entrypoint.
4496 * - Loader interface v3 differs from v2 in:
4497 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
4498 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
4499 * because the loader no longer does so.
4501 * - Loader interface v4 differs from v3 in:
4502 * - The ICD must implement vk_icdGetPhysicalDeviceProcAddr().
4504 *pSupportedVersion
= MIN2(*pSupportedVersion
, 4u);