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", NULL
);
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 device
->has_mmap_offset
=
481 anv_gem_get_param(fd
, I915_PARAM_MMAP_GTT_VERSION
) >= 4;
483 /* GENs prior to 8 do not support EU/Subslice info */
484 if (device
->info
.gen
>= 8) {
485 device
->subslice_total
= anv_gem_get_param(fd
, I915_PARAM_SUBSLICE_TOTAL
);
486 device
->eu_total
= anv_gem_get_param(fd
, I915_PARAM_EU_TOTAL
);
488 /* Without this information, we cannot get the right Braswell
489 * brandstrings, and we have to use conservative numbers for GPGPU on
490 * many platforms, but otherwise, things will just work.
492 if (device
->subslice_total
< 1 || device
->eu_total
< 1) {
493 intel_logw("Kernel 4.1 required to properly query GPU properties");
495 } else if (device
->info
.gen
== 7) {
496 device
->subslice_total
= 1 << (device
->info
.gt
- 1);
499 if (device
->info
.is_cherryview
&&
500 device
->subslice_total
> 0 && device
->eu_total
> 0) {
501 /* Logical CS threads = EUs per subslice * num threads per EU */
502 uint32_t max_cs_threads
=
503 device
->eu_total
/ device
->subslice_total
* device
->info
.num_thread_per_eu
;
505 /* Fuse configurations may give more threads than expected, never less. */
506 if (max_cs_threads
> device
->info
.max_cs_threads
)
507 device
->info
.max_cs_threads
= max_cs_threads
;
510 device
->compiler
= brw_compiler_create(NULL
, &device
->info
);
511 if (device
->compiler
== NULL
) {
512 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
515 device
->compiler
->shader_debug_log
= compiler_debug_log
;
516 device
->compiler
->shader_perf_log
= compiler_perf_log
;
517 device
->compiler
->supports_pull_constants
= false;
518 device
->compiler
->constant_buffer_0_is_relative
=
519 device
->info
.gen
< 8 || !device
->has_context_isolation
;
520 device
->compiler
->supports_shader_constants
= true;
521 device
->compiler
->compact_params
= false;
523 /* Broadwell PRM says:
525 * "Before Gen8, there was a historical configuration control field to
526 * swizzle address bit[6] for in X/Y tiling modes. This was set in three
527 * different places: TILECTL[1:0], ARB_MODE[5:4], and
528 * DISP_ARB_CTL[14:13].
530 * For Gen8 and subsequent generations, the swizzle fields are all
531 * reserved, and the CPU's memory controller performs all address
532 * swizzling modifications."
535 device
->info
.gen
< 8 && anv_gem_get_bit6_swizzle(fd
, I915_TILING_X
);
537 isl_device_init(&device
->isl_dev
, &device
->info
, swizzled
);
539 result
= anv_physical_device_init_uuids(device
);
540 if (result
!= VK_SUCCESS
)
543 anv_physical_device_init_disk_cache(device
);
545 if (instance
->enabled_extensions
.KHR_display
) {
546 master_fd
= open(primary_path
, O_RDWR
| O_CLOEXEC
);
547 if (master_fd
>= 0) {
548 /* prod the device with a GETPARAM call which will fail if
549 * we don't have permission to even render on this device
551 if (anv_gem_get_param(master_fd
, I915_PARAM_CHIPSET_ID
) == 0) {
557 device
->master_fd
= master_fd
;
559 result
= anv_init_wsi(device
);
560 if (result
!= VK_SUCCESS
)
561 goto fail_disk_cache
;
563 device
->perf
= anv_get_perf(&device
->info
, fd
);
565 anv_physical_device_get_supported_extensions(device
,
566 &device
->supported_extensions
);
569 device
->local_fd
= fd
;
571 *device_out
= device
;
576 anv_physical_device_free_disk_cache(device
);
578 ralloc_free(device
->compiler
);
580 vk_free(&instance
->alloc
, device
);
589 anv_physical_device_destroy(struct anv_physical_device
*device
)
591 anv_finish_wsi(device
);
592 anv_physical_device_free_disk_cache(device
);
593 ralloc_free(device
->compiler
);
594 ralloc_free(device
->perf
);
595 close(device
->local_fd
);
596 if (device
->master_fd
>= 0)
597 close(device
->master_fd
);
598 vk_free(&device
->instance
->alloc
, device
);
602 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
603 VkSystemAllocationScope allocationScope
)
609 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
610 size_t align
, VkSystemAllocationScope allocationScope
)
612 return realloc(pOriginal
, size
);
616 default_free_func(void *pUserData
, void *pMemory
)
621 static const VkAllocationCallbacks default_alloc
= {
623 .pfnAllocation
= default_alloc_func
,
624 .pfnReallocation
= default_realloc_func
,
625 .pfnFree
= default_free_func
,
628 VkResult
anv_EnumerateInstanceExtensionProperties(
629 const char* pLayerName
,
630 uint32_t* pPropertyCount
,
631 VkExtensionProperties
* pProperties
)
633 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
635 for (int i
= 0; i
< ANV_INSTANCE_EXTENSION_COUNT
; i
++) {
636 if (anv_instance_extensions_supported
.extensions
[i
]) {
637 vk_outarray_append(&out
, prop
) {
638 *prop
= anv_instance_extensions
[i
];
643 return vk_outarray_status(&out
);
646 VkResult
anv_CreateInstance(
647 const VkInstanceCreateInfo
* pCreateInfo
,
648 const VkAllocationCallbacks
* pAllocator
,
649 VkInstance
* pInstance
)
651 struct anv_instance
*instance
;
654 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
656 struct anv_instance_extension_table enabled_extensions
= {};
657 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
659 for (idx
= 0; idx
< ANV_INSTANCE_EXTENSION_COUNT
; idx
++) {
660 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
661 anv_instance_extensions
[idx
].extensionName
) == 0)
665 if (idx
>= ANV_INSTANCE_EXTENSION_COUNT
)
666 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
668 if (!anv_instance_extensions_supported
.extensions
[idx
])
669 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
671 enabled_extensions
.extensions
[idx
] = true;
674 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
675 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
677 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
679 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
682 instance
->alloc
= *pAllocator
;
684 instance
->alloc
= default_alloc
;
686 instance
->app_info
= (struct anv_app_info
) { .api_version
= 0 };
687 if (pCreateInfo
->pApplicationInfo
) {
688 const VkApplicationInfo
*app
= pCreateInfo
->pApplicationInfo
;
690 instance
->app_info
.app_name
=
691 vk_strdup(&instance
->alloc
, app
->pApplicationName
,
692 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
693 instance
->app_info
.app_version
= app
->applicationVersion
;
695 instance
->app_info
.engine_name
=
696 vk_strdup(&instance
->alloc
, app
->pEngineName
,
697 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
698 instance
->app_info
.engine_version
= app
->engineVersion
;
700 instance
->app_info
.api_version
= app
->apiVersion
;
703 if (instance
->app_info
.api_version
== 0)
704 instance
->app_info
.api_version
= VK_API_VERSION_1_0
;
706 instance
->enabled_extensions
= enabled_extensions
;
708 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->dispatch
.entrypoints
); i
++) {
709 /* Vulkan requires that entrypoints for extensions which have not been
710 * enabled must not be advertised.
712 if (!anv_instance_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
713 &instance
->enabled_extensions
)) {
714 instance
->dispatch
.entrypoints
[i
] = NULL
;
716 instance
->dispatch
.entrypoints
[i
] =
717 anv_instance_dispatch_table
.entrypoints
[i
];
721 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->physical_device_dispatch
.entrypoints
); i
++) {
722 /* Vulkan requires that entrypoints for extensions which have not been
723 * enabled must not be advertised.
725 if (!anv_physical_device_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
726 &instance
->enabled_extensions
)) {
727 instance
->physical_device_dispatch
.entrypoints
[i
] = NULL
;
729 instance
->physical_device_dispatch
.entrypoints
[i
] =
730 anv_physical_device_dispatch_table
.entrypoints
[i
];
734 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->device_dispatch
.entrypoints
); i
++) {
735 /* Vulkan requires that entrypoints for extensions which have not been
736 * enabled must not be advertised.
738 if (!anv_device_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
739 &instance
->enabled_extensions
, NULL
)) {
740 instance
->device_dispatch
.entrypoints
[i
] = NULL
;
742 instance
->device_dispatch
.entrypoints
[i
] =
743 anv_device_dispatch_table
.entrypoints
[i
];
747 instance
->physical_devices_enumerated
= false;
748 list_inithead(&instance
->physical_devices
);
750 result
= vk_debug_report_instance_init(&instance
->debug_report_callbacks
);
751 if (result
!= VK_SUCCESS
) {
752 vk_free2(&default_alloc
, pAllocator
, instance
);
753 return vk_error(result
);
756 instance
->pipeline_cache_enabled
=
757 env_var_as_boolean("ANV_ENABLE_PIPELINE_CACHE", true);
759 glsl_type_singleton_init_or_ref();
761 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
763 driParseOptionInfo(&instance
->available_dri_options
, anv_dri_options_xml
);
764 driParseConfigFiles(&instance
->dri_options
, &instance
->available_dri_options
,
766 instance
->app_info
.engine_name
,
767 instance
->app_info
.engine_version
);
769 *pInstance
= anv_instance_to_handle(instance
);
774 void anv_DestroyInstance(
775 VkInstance _instance
,
776 const VkAllocationCallbacks
* pAllocator
)
778 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
783 list_for_each_entry_safe(struct anv_physical_device
, pdevice
,
784 &instance
->physical_devices
, link
)
785 anv_physical_device_destroy(pdevice
);
787 vk_free(&instance
->alloc
, (char *)instance
->app_info
.app_name
);
788 vk_free(&instance
->alloc
, (char *)instance
->app_info
.engine_name
);
790 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
792 vk_debug_report_instance_destroy(&instance
->debug_report_callbacks
);
794 glsl_type_singleton_decref();
796 driDestroyOptionCache(&instance
->dri_options
);
797 driDestroyOptionInfo(&instance
->available_dri_options
);
799 vk_free(&instance
->alloc
, instance
);
803 anv_enumerate_physical_devices(struct anv_instance
*instance
)
805 if (instance
->physical_devices_enumerated
)
808 instance
->physical_devices_enumerated
= true;
810 /* TODO: Check for more devices ? */
811 drmDevicePtr devices
[8];
814 max_devices
= drmGetDevices2(0, devices
, ARRAY_SIZE(devices
));
818 VkResult result
= VK_SUCCESS
;
819 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
820 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
821 devices
[i
]->bustype
== DRM_BUS_PCI
&&
822 devices
[i
]->deviceinfo
.pci
->vendor_id
== 0x8086) {
824 struct anv_physical_device
*pdevice
;
825 result
= anv_physical_device_try_create(instance
, devices
[i
],
827 /* Incompatible DRM device, skip. */
828 if (result
== VK_ERROR_INCOMPATIBLE_DRIVER
) {
833 /* Error creating the physical device, report the error. */
834 if (result
!= VK_SUCCESS
)
837 list_addtail(&pdevice
->link
, &instance
->physical_devices
);
840 drmFreeDevices(devices
, max_devices
);
842 /* If we successfully enumerated any devices, call it success */
846 VkResult
anv_EnumeratePhysicalDevices(
847 VkInstance _instance
,
848 uint32_t* pPhysicalDeviceCount
,
849 VkPhysicalDevice
* pPhysicalDevices
)
851 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
852 VK_OUTARRAY_MAKE(out
, pPhysicalDevices
, pPhysicalDeviceCount
);
854 VkResult result
= anv_enumerate_physical_devices(instance
);
855 if (result
!= VK_SUCCESS
)
858 list_for_each_entry(struct anv_physical_device
, pdevice
,
859 &instance
->physical_devices
, link
) {
860 vk_outarray_append(&out
, i
) {
861 *i
= anv_physical_device_to_handle(pdevice
);
865 return vk_outarray_status(&out
);
868 VkResult
anv_EnumeratePhysicalDeviceGroups(
869 VkInstance _instance
,
870 uint32_t* pPhysicalDeviceGroupCount
,
871 VkPhysicalDeviceGroupProperties
* pPhysicalDeviceGroupProperties
)
873 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
874 VK_OUTARRAY_MAKE(out
, pPhysicalDeviceGroupProperties
,
875 pPhysicalDeviceGroupCount
);
877 VkResult result
= anv_enumerate_physical_devices(instance
);
878 if (result
!= VK_SUCCESS
)
881 list_for_each_entry(struct anv_physical_device
, pdevice
,
882 &instance
->physical_devices
, link
) {
883 vk_outarray_append(&out
, p
) {
884 p
->physicalDeviceCount
= 1;
885 memset(p
->physicalDevices
, 0, sizeof(p
->physicalDevices
));
886 p
->physicalDevices
[0] = anv_physical_device_to_handle(pdevice
);
887 p
->subsetAllocation
= false;
889 vk_foreach_struct(ext
, p
->pNext
)
890 anv_debug_ignored_stype(ext
->sType
);
894 return vk_outarray_status(&out
);
897 void anv_GetPhysicalDeviceFeatures(
898 VkPhysicalDevice physicalDevice
,
899 VkPhysicalDeviceFeatures
* pFeatures
)
901 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
903 *pFeatures
= (VkPhysicalDeviceFeatures
) {
904 .robustBufferAccess
= true,
905 .fullDrawIndexUint32
= true,
906 .imageCubeArray
= true,
907 .independentBlend
= true,
908 .geometryShader
= true,
909 .tessellationShader
= true,
910 .sampleRateShading
= true,
911 .dualSrcBlend
= true,
913 .multiDrawIndirect
= true,
914 .drawIndirectFirstInstance
= true,
916 .depthBiasClamp
= true,
917 .fillModeNonSolid
= true,
918 .depthBounds
= pdevice
->info
.gen
>= 12,
922 .multiViewport
= true,
923 .samplerAnisotropy
= true,
924 .textureCompressionETC2
= pdevice
->info
.gen
>= 8 ||
925 pdevice
->info
.is_baytrail
,
926 .textureCompressionASTC_LDR
= pdevice
->info
.gen
>= 9, /* FINISHME CHV */
927 .textureCompressionBC
= true,
928 .occlusionQueryPrecise
= true,
929 .pipelineStatisticsQuery
= true,
930 .fragmentStoresAndAtomics
= true,
931 .shaderTessellationAndGeometryPointSize
= true,
932 .shaderImageGatherExtended
= true,
933 .shaderStorageImageExtendedFormats
= true,
934 .shaderStorageImageMultisample
= false,
935 .shaderStorageImageReadWithoutFormat
= false,
936 .shaderStorageImageWriteWithoutFormat
= true,
937 .shaderUniformBufferArrayDynamicIndexing
= true,
938 .shaderSampledImageArrayDynamicIndexing
= true,
939 .shaderStorageBufferArrayDynamicIndexing
= true,
940 .shaderStorageImageArrayDynamicIndexing
= true,
941 .shaderClipDistance
= true,
942 .shaderCullDistance
= true,
943 .shaderFloat64
= pdevice
->info
.gen
>= 8 &&
944 pdevice
->info
.has_64bit_float
,
945 .shaderInt64
= pdevice
->info
.gen
>= 8 &&
946 pdevice
->info
.has_64bit_int
,
947 .shaderInt16
= pdevice
->info
.gen
>= 8,
948 .shaderResourceMinLod
= pdevice
->info
.gen
>= 9,
949 .variableMultisampleRate
= true,
950 .inheritedQueries
= true,
953 /* We can't do image stores in vec4 shaders */
954 pFeatures
->vertexPipelineStoresAndAtomics
=
955 pdevice
->compiler
->scalar_stage
[MESA_SHADER_VERTEX
] &&
956 pdevice
->compiler
->scalar_stage
[MESA_SHADER_GEOMETRY
];
958 struct anv_app_info
*app_info
= &pdevice
->instance
->app_info
;
960 /* The new DOOM and Wolfenstein games require depthBounds without
961 * checking for it. They seem to run fine without it so just claim it's
962 * there and accept the consequences.
964 if (app_info
->engine_name
&& strcmp(app_info
->engine_name
, "idTech") == 0)
965 pFeatures
->depthBounds
= true;
969 anv_get_physical_device_features_1_1(struct anv_physical_device
*pdevice
,
970 VkPhysicalDeviceVulkan11Features
*f
)
972 assert(f
->sType
== VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES
);
974 f
->storageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
975 f
->uniformAndStorageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
976 f
->storagePushConstant16
= pdevice
->info
.gen
>= 8;
977 f
->storageInputOutput16
= false;
979 f
->multiviewGeometryShader
= true;
980 f
->multiviewTessellationShader
= true;
981 f
->variablePointersStorageBuffer
= true;
982 f
->variablePointers
= true;
983 f
->protectedMemory
= false;
984 f
->samplerYcbcrConversion
= true;
985 f
->shaderDrawParameters
= true;
989 anv_get_physical_device_features_1_2(struct anv_physical_device
*pdevice
,
990 VkPhysicalDeviceVulkan12Features
*f
)
992 assert(f
->sType
== VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES
);
994 f
->samplerMirrorClampToEdge
= true;
995 f
->drawIndirectCount
= true;
996 f
->storageBuffer8BitAccess
= pdevice
->info
.gen
>= 8;
997 f
->uniformAndStorageBuffer8BitAccess
= pdevice
->info
.gen
>= 8;
998 f
->storagePushConstant8
= pdevice
->info
.gen
>= 8;
999 f
->shaderBufferInt64Atomics
= pdevice
->info
.gen
>= 9 &&
1000 pdevice
->use_softpin
;
1001 f
->shaderSharedInt64Atomics
= false;
1002 f
->shaderFloat16
= pdevice
->info
.gen
>= 8;
1003 f
->shaderInt8
= pdevice
->info
.gen
>= 8;
1005 bool descIndexing
= pdevice
->has_a64_buffer_access
&&
1006 pdevice
->has_bindless_images
;
1007 f
->descriptorIndexing
= descIndexing
;
1008 f
->shaderInputAttachmentArrayDynamicIndexing
= false;
1009 f
->shaderUniformTexelBufferArrayDynamicIndexing
= descIndexing
;
1010 f
->shaderStorageTexelBufferArrayDynamicIndexing
= descIndexing
;
1011 f
->shaderUniformBufferArrayNonUniformIndexing
= false;
1012 f
->shaderSampledImageArrayNonUniformIndexing
= descIndexing
;
1013 f
->shaderStorageBufferArrayNonUniformIndexing
= descIndexing
;
1014 f
->shaderStorageImageArrayNonUniformIndexing
= descIndexing
;
1015 f
->shaderInputAttachmentArrayNonUniformIndexing
= false;
1016 f
->shaderUniformTexelBufferArrayNonUniformIndexing
= descIndexing
;
1017 f
->shaderStorageTexelBufferArrayNonUniformIndexing
= descIndexing
;
1018 f
->descriptorBindingUniformBufferUpdateAfterBind
= false;
1019 f
->descriptorBindingSampledImageUpdateAfterBind
= descIndexing
;
1020 f
->descriptorBindingStorageImageUpdateAfterBind
= descIndexing
;
1021 f
->descriptorBindingStorageBufferUpdateAfterBind
= descIndexing
;
1022 f
->descriptorBindingUniformTexelBufferUpdateAfterBind
= descIndexing
;
1023 f
->descriptorBindingStorageTexelBufferUpdateAfterBind
= descIndexing
;
1024 f
->descriptorBindingUpdateUnusedWhilePending
= descIndexing
;
1025 f
->descriptorBindingPartiallyBound
= descIndexing
;
1026 f
->descriptorBindingVariableDescriptorCount
= false;
1027 f
->runtimeDescriptorArray
= descIndexing
;
1029 f
->samplerFilterMinmax
= pdevice
->info
.gen
>= 9;
1030 f
->scalarBlockLayout
= true;
1031 f
->imagelessFramebuffer
= true;
1032 f
->uniformBufferStandardLayout
= true;
1033 f
->shaderSubgroupExtendedTypes
= true;
1034 f
->separateDepthStencilLayouts
= true;
1035 f
->hostQueryReset
= true;
1036 f
->timelineSemaphore
= true;
1037 f
->bufferDeviceAddress
= pdevice
->has_a64_buffer_access
;
1038 f
->bufferDeviceAddressCaptureReplay
= pdevice
->has_a64_buffer_access
;
1039 f
->bufferDeviceAddressMultiDevice
= false;
1040 f
->vulkanMemoryModel
= true;
1041 f
->vulkanMemoryModelDeviceScope
= true;
1042 f
->vulkanMemoryModelAvailabilityVisibilityChains
= true;
1043 f
->shaderOutputViewportIndex
= true;
1044 f
->shaderOutputLayer
= true;
1045 f
->subgroupBroadcastDynamicId
= true;
1048 void anv_GetPhysicalDeviceFeatures2(
1049 VkPhysicalDevice physicalDevice
,
1050 VkPhysicalDeviceFeatures2
* pFeatures
)
1052 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1053 anv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
1055 VkPhysicalDeviceVulkan11Features core_1_1
= {
1056 .sType
= VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES
,
1058 anv_get_physical_device_features_1_1(pdevice
, &core_1_1
);
1060 VkPhysicalDeviceVulkan12Features core_1_2
= {
1061 .sType
= VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES
,
1063 anv_get_physical_device_features_1_2(pdevice
, &core_1_2
);
1065 #define CORE_FEATURE(major, minor, feature) \
1066 features->feature = core_##major##_##minor.feature
1069 vk_foreach_struct(ext
, pFeatures
->pNext
) {
1070 switch (ext
->sType
) {
1071 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES_KHR
: {
1072 VkPhysicalDevice8BitStorageFeaturesKHR
*features
=
1073 (VkPhysicalDevice8BitStorageFeaturesKHR
*)ext
;
1074 CORE_FEATURE(1, 2, storageBuffer8BitAccess
);
1075 CORE_FEATURE(1, 2, uniformAndStorageBuffer8BitAccess
);
1076 CORE_FEATURE(1, 2, storagePushConstant8
);
1080 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES
: {
1081 VkPhysicalDevice16BitStorageFeatures
*features
=
1082 (VkPhysicalDevice16BitStorageFeatures
*)ext
;
1083 CORE_FEATURE(1, 1, storageBuffer16BitAccess
);
1084 CORE_FEATURE(1, 1, uniformAndStorageBuffer16BitAccess
);
1085 CORE_FEATURE(1, 1, storagePushConstant16
);
1086 CORE_FEATURE(1, 1, storageInputOutput16
);
1090 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_EXT
: {
1091 VkPhysicalDeviceBufferDeviceAddressFeaturesEXT
*features
= (void *)ext
;
1092 features
->bufferDeviceAddress
= pdevice
->has_a64_buffer_access
;
1093 features
->bufferDeviceAddressCaptureReplay
= false;
1094 features
->bufferDeviceAddressMultiDevice
= false;
1098 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_KHR
: {
1099 VkPhysicalDeviceBufferDeviceAddressFeaturesKHR
*features
= (void *)ext
;
1100 CORE_FEATURE(1, 2, bufferDeviceAddress
);
1101 CORE_FEATURE(1, 2, bufferDeviceAddressCaptureReplay
);
1102 CORE_FEATURE(1, 2, bufferDeviceAddressMultiDevice
);
1106 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COMPUTE_SHADER_DERIVATIVES_FEATURES_NV
: {
1107 VkPhysicalDeviceComputeShaderDerivativesFeaturesNV
*features
=
1108 (VkPhysicalDeviceComputeShaderDerivativesFeaturesNV
*)ext
;
1109 features
->computeDerivativeGroupQuads
= true;
1110 features
->computeDerivativeGroupLinear
= true;
1114 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT
: {
1115 VkPhysicalDeviceConditionalRenderingFeaturesEXT
*features
=
1116 (VkPhysicalDeviceConditionalRenderingFeaturesEXT
*)ext
;
1117 features
->conditionalRendering
= pdevice
->info
.gen
>= 8 ||
1118 pdevice
->info
.is_haswell
;
1119 features
->inheritedConditionalRendering
= pdevice
->info
.gen
>= 8 ||
1120 pdevice
->info
.is_haswell
;
1124 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_CLIP_ENABLE_FEATURES_EXT
: {
1125 VkPhysicalDeviceDepthClipEnableFeaturesEXT
*features
=
1126 (VkPhysicalDeviceDepthClipEnableFeaturesEXT
*)ext
;
1127 features
->depthClipEnable
= true;
1131 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR
: {
1132 VkPhysicalDeviceFloat16Int8FeaturesKHR
*features
= (void *)ext
;
1133 CORE_FEATURE(1, 2, shaderFloat16
);
1134 CORE_FEATURE(1, 2, shaderInt8
);
1138 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FRAGMENT_SHADER_INTERLOCK_FEATURES_EXT
: {
1139 VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT
*features
=
1140 (VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT
*)ext
;
1141 features
->fragmentShaderSampleInterlock
= pdevice
->info
.gen
>= 9;
1142 features
->fragmentShaderPixelInterlock
= pdevice
->info
.gen
>= 9;
1143 features
->fragmentShaderShadingRateInterlock
= false;
1147 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_QUERY_RESET_FEATURES_EXT
: {
1148 VkPhysicalDeviceHostQueryResetFeaturesEXT
*features
=
1149 (VkPhysicalDeviceHostQueryResetFeaturesEXT
*)ext
;
1150 CORE_FEATURE(1, 2, hostQueryReset
);
1154 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT
: {
1155 VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*features
=
1156 (VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*)ext
;
1157 CORE_FEATURE(1, 2, shaderInputAttachmentArrayDynamicIndexing
);
1158 CORE_FEATURE(1, 2, shaderUniformTexelBufferArrayDynamicIndexing
);
1159 CORE_FEATURE(1, 2, shaderStorageTexelBufferArrayDynamicIndexing
);
1160 CORE_FEATURE(1, 2, shaderUniformBufferArrayNonUniformIndexing
);
1161 CORE_FEATURE(1, 2, shaderSampledImageArrayNonUniformIndexing
);
1162 CORE_FEATURE(1, 2, shaderStorageBufferArrayNonUniformIndexing
);
1163 CORE_FEATURE(1, 2, shaderStorageImageArrayNonUniformIndexing
);
1164 CORE_FEATURE(1, 2, shaderInputAttachmentArrayNonUniformIndexing
);
1165 CORE_FEATURE(1, 2, shaderUniformTexelBufferArrayNonUniformIndexing
);
1166 CORE_FEATURE(1, 2, shaderStorageTexelBufferArrayNonUniformIndexing
);
1167 CORE_FEATURE(1, 2, descriptorBindingUniformBufferUpdateAfterBind
);
1168 CORE_FEATURE(1, 2, descriptorBindingSampledImageUpdateAfterBind
);
1169 CORE_FEATURE(1, 2, descriptorBindingStorageImageUpdateAfterBind
);
1170 CORE_FEATURE(1, 2, descriptorBindingStorageBufferUpdateAfterBind
);
1171 CORE_FEATURE(1, 2, descriptorBindingUniformTexelBufferUpdateAfterBind
);
1172 CORE_FEATURE(1, 2, descriptorBindingStorageTexelBufferUpdateAfterBind
);
1173 CORE_FEATURE(1, 2, descriptorBindingUpdateUnusedWhilePending
);
1174 CORE_FEATURE(1, 2, descriptorBindingPartiallyBound
);
1175 CORE_FEATURE(1, 2, descriptorBindingVariableDescriptorCount
);
1176 CORE_FEATURE(1, 2, runtimeDescriptorArray
);
1180 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INDEX_TYPE_UINT8_FEATURES_EXT
: {
1181 VkPhysicalDeviceIndexTypeUint8FeaturesEXT
*features
=
1182 (VkPhysicalDeviceIndexTypeUint8FeaturesEXT
*)ext
;
1183 features
->indexTypeUint8
= true;
1187 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_FEATURES_EXT
: {
1188 VkPhysicalDeviceInlineUniformBlockFeaturesEXT
*features
=
1189 (VkPhysicalDeviceInlineUniformBlockFeaturesEXT
*)ext
;
1190 features
->inlineUniformBlock
= true;
1191 features
->descriptorBindingInlineUniformBlockUpdateAfterBind
= true;
1195 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_FEATURES_EXT
: {
1196 VkPhysicalDeviceLineRasterizationFeaturesEXT
*features
=
1197 (VkPhysicalDeviceLineRasterizationFeaturesEXT
*)ext
;
1198 features
->rectangularLines
= true;
1199 features
->bresenhamLines
= true;
1200 /* Support for Smooth lines with MSAA was removed on gen11. From the
1201 * BSpec section "Multisample ModesState" table for "AA Line Support
1204 * GEN10:BUG:######## NUM_MULTISAMPLES == 1
1206 * Fortunately, this isn't a case most people care about.
1208 features
->smoothLines
= pdevice
->info
.gen
< 10;
1209 features
->stippledRectangularLines
= false;
1210 features
->stippledBresenhamLines
= true;
1211 features
->stippledSmoothLines
= false;
1215 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES
: {
1216 VkPhysicalDeviceMultiviewFeatures
*features
=
1217 (VkPhysicalDeviceMultiviewFeatures
*)ext
;
1218 CORE_FEATURE(1, 1, multiview
);
1219 CORE_FEATURE(1, 1, multiviewGeometryShader
);
1220 CORE_FEATURE(1, 1, multiviewTessellationShader
);
1224 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGELESS_FRAMEBUFFER_FEATURES_KHR
: {
1225 VkPhysicalDeviceImagelessFramebufferFeaturesKHR
*features
=
1226 (VkPhysicalDeviceImagelessFramebufferFeaturesKHR
*)ext
;
1227 CORE_FEATURE(1, 2, imagelessFramebuffer
);
1231 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR
: {
1232 VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR
*features
=
1233 (VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR
*)ext
;
1234 features
->pipelineExecutableInfo
= true;
1238 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES
: {
1239 VkPhysicalDeviceProtectedMemoryFeatures
*features
= (void *)ext
;
1240 CORE_FEATURE(1, 1, protectedMemory
);
1244 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES
: {
1245 VkPhysicalDeviceSamplerYcbcrConversionFeatures
*features
=
1246 (VkPhysicalDeviceSamplerYcbcrConversionFeatures
*) ext
;
1247 CORE_FEATURE(1, 1, samplerYcbcrConversion
);
1251 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES_EXT
: {
1252 VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*features
=
1253 (VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*)ext
;
1254 CORE_FEATURE(1, 2, scalarBlockLayout
);
1258 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SEPARATE_DEPTH_STENCIL_LAYOUTS_FEATURES_KHR
: {
1259 VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR
*features
=
1260 (VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR
*)ext
;
1261 CORE_FEATURE(1, 2, separateDepthStencilLayouts
);
1265 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_INT64_FEATURES_KHR
: {
1266 VkPhysicalDeviceShaderAtomicInt64FeaturesKHR
*features
= (void *)ext
;
1267 CORE_FEATURE(1, 2, shaderBufferInt64Atomics
);
1268 CORE_FEATURE(1, 2, shaderSharedInt64Atomics
);
1272 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DEMOTE_TO_HELPER_INVOCATION_FEATURES_EXT
: {
1273 VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT
*features
= (void *)ext
;
1274 features
->shaderDemoteToHelperInvocation
= true;
1278 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CLOCK_FEATURES_KHR
: {
1279 VkPhysicalDeviceShaderClockFeaturesKHR
*features
=
1280 (VkPhysicalDeviceShaderClockFeaturesKHR
*)ext
;
1281 features
->shaderSubgroupClock
= true;
1282 features
->shaderDeviceClock
= false;
1286 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETERS_FEATURES
: {
1287 VkPhysicalDeviceShaderDrawParametersFeatures
*features
= (void *)ext
;
1288 CORE_FEATURE(1, 1, shaderDrawParameters
);
1292 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_SUBGROUP_EXTENDED_TYPES_FEATURES_KHR
: {
1293 VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR
*features
=
1294 (VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR
*)ext
;
1295 CORE_FEATURE(1, 2, shaderSubgroupExtendedTypes
);
1299 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_FEATURES_EXT
: {
1300 VkPhysicalDeviceSubgroupSizeControlFeaturesEXT
*features
=
1301 (VkPhysicalDeviceSubgroupSizeControlFeaturesEXT
*)ext
;
1302 features
->subgroupSizeControl
= true;
1303 features
->computeFullSubgroups
= true;
1307 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_FEATURES_EXT
: {
1308 VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT
*features
=
1309 (VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT
*)ext
;
1310 features
->texelBufferAlignment
= true;
1314 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_FEATURES_KHR
: {
1315 VkPhysicalDeviceTimelineSemaphoreFeaturesKHR
*features
=
1316 (VkPhysicalDeviceTimelineSemaphoreFeaturesKHR
*) ext
;
1317 CORE_FEATURE(1, 2, timelineSemaphore
);
1321 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES
: {
1322 VkPhysicalDeviceVariablePointersFeatures
*features
= (void *)ext
;
1323 CORE_FEATURE(1, 1, variablePointersStorageBuffer
);
1324 CORE_FEATURE(1, 1, variablePointers
);
1328 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT
: {
1329 VkPhysicalDeviceTransformFeedbackFeaturesEXT
*features
=
1330 (VkPhysicalDeviceTransformFeedbackFeaturesEXT
*)ext
;
1331 features
->transformFeedback
= true;
1332 features
->geometryStreams
= true;
1336 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_UNIFORM_BUFFER_STANDARD_LAYOUT_FEATURES_KHR
: {
1337 VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR
*features
=
1338 (VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR
*)ext
;
1339 CORE_FEATURE(1, 2, uniformBufferStandardLayout
);
1343 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT
: {
1344 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*features
=
1345 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*)ext
;
1346 features
->vertexAttributeInstanceRateDivisor
= true;
1347 features
->vertexAttributeInstanceRateZeroDivisor
= true;
1351 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES
:
1352 anv_get_physical_device_features_1_1(pdevice
, (void *)ext
);
1355 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES
:
1356 anv_get_physical_device_features_1_2(pdevice
, (void *)ext
);
1359 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_MEMORY_MODEL_FEATURES_KHR
: {
1360 VkPhysicalDeviceVulkanMemoryModelFeaturesKHR
*features
= (void *)ext
;
1361 CORE_FEATURE(1, 2, vulkanMemoryModel
);
1362 CORE_FEATURE(1, 2, vulkanMemoryModelDeviceScope
);
1363 CORE_FEATURE(1, 2, vulkanMemoryModelAvailabilityVisibilityChains
);
1367 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_YCBCR_IMAGE_ARRAYS_FEATURES_EXT
: {
1368 VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*features
=
1369 (VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*)ext
;
1370 features
->ycbcrImageArrays
= true;
1375 anv_debug_ignored_stype(ext
->sType
);
1383 #define MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS 64
1385 #define MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS 64
1386 #define MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS 256
1388 void anv_GetPhysicalDeviceProperties(
1389 VkPhysicalDevice physicalDevice
,
1390 VkPhysicalDeviceProperties
* pProperties
)
1392 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1393 const struct gen_device_info
*devinfo
= &pdevice
->info
;
1395 /* See assertions made when programming the buffer surface state. */
1396 const uint32_t max_raw_buffer_sz
= devinfo
->gen
>= 7 ?
1397 (1ul << 30) : (1ul << 27);
1399 const uint32_t max_ssbos
= pdevice
->has_a64_buffer_access
? UINT16_MAX
: 64;
1400 const uint32_t max_textures
=
1401 pdevice
->has_bindless_images
? UINT16_MAX
: 128;
1402 const uint32_t max_samplers
=
1403 pdevice
->has_bindless_samplers
? UINT16_MAX
:
1404 (devinfo
->gen
>= 8 || devinfo
->is_haswell
) ? 128 : 16;
1405 const uint32_t max_images
=
1406 pdevice
->has_bindless_images
? UINT16_MAX
: MAX_IMAGES
;
1408 /* If we can use bindless for everything, claim a high per-stage limit,
1409 * otherwise use the binding table size, minus the slots reserved for
1410 * render targets and one slot for the descriptor buffer. */
1411 const uint32_t max_per_stage
=
1412 pdevice
->has_bindless_images
&& pdevice
->has_a64_buffer_access
1413 ? UINT32_MAX
: MAX_BINDING_TABLE_SIZE
- MAX_RTS
- 1;
1415 /* Limit max_threads to 64 for the GPGPU_WALKER command */
1416 const uint32_t max_workgroup_size
= 32 * MIN2(64, devinfo
->max_cs_threads
);
1418 VkSampleCountFlags sample_counts
=
1419 isl_device_get_sample_counts(&pdevice
->isl_dev
);
1422 VkPhysicalDeviceLimits limits
= {
1423 .maxImageDimension1D
= (1 << 14),
1424 .maxImageDimension2D
= (1 << 14),
1425 .maxImageDimension3D
= (1 << 11),
1426 .maxImageDimensionCube
= (1 << 14),
1427 .maxImageArrayLayers
= (1 << 11),
1428 .maxTexelBufferElements
= 128 * 1024 * 1024,
1429 .maxUniformBufferRange
= (1ul << 27),
1430 .maxStorageBufferRange
= max_raw_buffer_sz
,
1431 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
1432 .maxMemoryAllocationCount
= UINT32_MAX
,
1433 .maxSamplerAllocationCount
= 64 * 1024,
1434 .bufferImageGranularity
= 64, /* A cache line */
1435 .sparseAddressSpaceSize
= 0,
1436 .maxBoundDescriptorSets
= MAX_SETS
,
1437 .maxPerStageDescriptorSamplers
= max_samplers
,
1438 .maxPerStageDescriptorUniformBuffers
= MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
,
1439 .maxPerStageDescriptorStorageBuffers
= max_ssbos
,
1440 .maxPerStageDescriptorSampledImages
= max_textures
,
1441 .maxPerStageDescriptorStorageImages
= max_images
,
1442 .maxPerStageDescriptorInputAttachments
= MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS
,
1443 .maxPerStageResources
= max_per_stage
,
1444 .maxDescriptorSetSamplers
= 6 * max_samplers
, /* number of stages * maxPerStageDescriptorSamplers */
1445 .maxDescriptorSetUniformBuffers
= 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
, /* number of stages * maxPerStageDescriptorUniformBuffers */
1446 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
1447 .maxDescriptorSetStorageBuffers
= 6 * max_ssbos
, /* number of stages * maxPerStageDescriptorStorageBuffers */
1448 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
1449 .maxDescriptorSetSampledImages
= 6 * max_textures
, /* number of stages * maxPerStageDescriptorSampledImages */
1450 .maxDescriptorSetStorageImages
= 6 * max_images
, /* number of stages * maxPerStageDescriptorStorageImages */
1451 .maxDescriptorSetInputAttachments
= MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS
,
1452 .maxVertexInputAttributes
= MAX_VBS
,
1453 .maxVertexInputBindings
= MAX_VBS
,
1454 .maxVertexInputAttributeOffset
= 2047,
1455 .maxVertexInputBindingStride
= 2048,
1456 .maxVertexOutputComponents
= 128,
1457 .maxTessellationGenerationLevel
= 64,
1458 .maxTessellationPatchSize
= 32,
1459 .maxTessellationControlPerVertexInputComponents
= 128,
1460 .maxTessellationControlPerVertexOutputComponents
= 128,
1461 .maxTessellationControlPerPatchOutputComponents
= 128,
1462 .maxTessellationControlTotalOutputComponents
= 2048,
1463 .maxTessellationEvaluationInputComponents
= 128,
1464 .maxTessellationEvaluationOutputComponents
= 128,
1465 .maxGeometryShaderInvocations
= 32,
1466 .maxGeometryInputComponents
= 64,
1467 .maxGeometryOutputComponents
= 128,
1468 .maxGeometryOutputVertices
= 256,
1469 .maxGeometryTotalOutputComponents
= 1024,
1470 .maxFragmentInputComponents
= 116, /* 128 components - (PSIZ, CLIP_DIST0, CLIP_DIST1) */
1471 .maxFragmentOutputAttachments
= 8,
1472 .maxFragmentDualSrcAttachments
= 1,
1473 .maxFragmentCombinedOutputResources
= 8,
1474 .maxComputeSharedMemorySize
= 64 * 1024,
1475 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
1476 .maxComputeWorkGroupInvocations
= max_workgroup_size
,
1477 .maxComputeWorkGroupSize
= {
1482 .subPixelPrecisionBits
= 8,
1483 .subTexelPrecisionBits
= 8,
1484 .mipmapPrecisionBits
= 8,
1485 .maxDrawIndexedIndexValue
= UINT32_MAX
,
1486 .maxDrawIndirectCount
= UINT32_MAX
,
1487 .maxSamplerLodBias
= 16,
1488 .maxSamplerAnisotropy
= 16,
1489 .maxViewports
= MAX_VIEWPORTS
,
1490 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
1491 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
1492 .viewportSubPixelBits
= 13, /* We take a float? */
1493 .minMemoryMapAlignment
= 4096, /* A page */
1494 /* The dataport requires texel alignment so we need to assume a worst
1495 * case of R32G32B32A32 which is 16 bytes.
1497 .minTexelBufferOffsetAlignment
= 16,
1498 /* We need 16 for UBO block reads to work and 32 for push UBOs */
1499 .minUniformBufferOffsetAlignment
= 32,
1500 .minStorageBufferOffsetAlignment
= 4,
1501 .minTexelOffset
= -8,
1502 .maxTexelOffset
= 7,
1503 .minTexelGatherOffset
= -32,
1504 .maxTexelGatherOffset
= 31,
1505 .minInterpolationOffset
= -0.5,
1506 .maxInterpolationOffset
= 0.4375,
1507 .subPixelInterpolationOffsetBits
= 4,
1508 .maxFramebufferWidth
= (1 << 14),
1509 .maxFramebufferHeight
= (1 << 14),
1510 .maxFramebufferLayers
= (1 << 11),
1511 .framebufferColorSampleCounts
= sample_counts
,
1512 .framebufferDepthSampleCounts
= sample_counts
,
1513 .framebufferStencilSampleCounts
= sample_counts
,
1514 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
1515 .maxColorAttachments
= MAX_RTS
,
1516 .sampledImageColorSampleCounts
= sample_counts
,
1517 .sampledImageIntegerSampleCounts
= sample_counts
,
1518 .sampledImageDepthSampleCounts
= sample_counts
,
1519 .sampledImageStencilSampleCounts
= sample_counts
,
1520 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
1521 .maxSampleMaskWords
= 1,
1522 .timestampComputeAndGraphics
= true,
1523 .timestampPeriod
= 1000000000.0 / devinfo
->timestamp_frequency
,
1524 .maxClipDistances
= 8,
1525 .maxCullDistances
= 8,
1526 .maxCombinedClipAndCullDistances
= 8,
1527 .discreteQueuePriorities
= 2,
1528 .pointSizeRange
= { 0.125, 255.875 },
1531 (devinfo
->gen
>= 9 || devinfo
->is_cherryview
) ?
1532 2047.9921875 : 7.9921875,
1534 .pointSizeGranularity
= (1.0 / 8.0),
1535 .lineWidthGranularity
= (1.0 / 128.0),
1536 .strictLines
= false,
1537 .standardSampleLocations
= true,
1538 .optimalBufferCopyOffsetAlignment
= 128,
1539 .optimalBufferCopyRowPitchAlignment
= 128,
1540 .nonCoherentAtomSize
= 64,
1543 *pProperties
= (VkPhysicalDeviceProperties
) {
1544 .apiVersion
= anv_physical_device_api_version(pdevice
),
1545 .driverVersion
= vk_get_driver_version(),
1547 .deviceID
= pdevice
->info
.chipset_id
,
1548 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
1550 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
1553 snprintf(pProperties
->deviceName
, sizeof(pProperties
->deviceName
),
1554 "%s", pdevice
->name
);
1555 memcpy(pProperties
->pipelineCacheUUID
,
1556 pdevice
->pipeline_cache_uuid
, VK_UUID_SIZE
);
1560 anv_get_physical_device_properties_1_1(struct anv_physical_device
*pdevice
,
1561 VkPhysicalDeviceVulkan11Properties
*p
)
1563 assert(p
->sType
== VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES
);
1565 memcpy(p
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
1566 memcpy(p
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
1567 memset(p
->deviceLUID
, 0, VK_LUID_SIZE
);
1568 p
->deviceNodeMask
= 0;
1569 p
->deviceLUIDValid
= false;
1571 p
->subgroupSize
= BRW_SUBGROUP_SIZE
;
1572 VkShaderStageFlags scalar_stages
= 0;
1573 for (unsigned stage
= 0; stage
< MESA_SHADER_STAGES
; stage
++) {
1574 if (pdevice
->compiler
->scalar_stage
[stage
])
1575 scalar_stages
|= mesa_to_vk_shader_stage(stage
);
1577 p
->subgroupSupportedStages
= scalar_stages
;
1578 p
->subgroupSupportedOperations
= VK_SUBGROUP_FEATURE_BASIC_BIT
|
1579 VK_SUBGROUP_FEATURE_VOTE_BIT
|
1580 VK_SUBGROUP_FEATURE_BALLOT_BIT
|
1581 VK_SUBGROUP_FEATURE_SHUFFLE_BIT
|
1582 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT
|
1583 VK_SUBGROUP_FEATURE_QUAD_BIT
;
1584 if (pdevice
->info
.gen
>= 8) {
1585 /* TODO: There's no technical reason why these can't be made to
1586 * work on gen7 but they don't at the moment so it's best to leave
1587 * the feature disabled than enabled and broken.
1589 p
->subgroupSupportedOperations
|= VK_SUBGROUP_FEATURE_ARITHMETIC_BIT
|
1590 VK_SUBGROUP_FEATURE_CLUSTERED_BIT
;
1592 p
->subgroupQuadOperationsInAllStages
= pdevice
->info
.gen
>= 8;
1594 p
->pointClippingBehavior
= VK_POINT_CLIPPING_BEHAVIOR_USER_CLIP_PLANES_ONLY
;
1595 p
->maxMultiviewViewCount
= 16;
1596 p
->maxMultiviewInstanceIndex
= UINT32_MAX
/ 16;
1597 p
->protectedNoFault
= false;
1598 /* This value doesn't matter for us today as our per-stage descriptors are
1601 p
->maxPerSetDescriptors
= 1024;
1602 p
->maxMemoryAllocationSize
= MAX_MEMORY_ALLOCATION_SIZE
;
1606 anv_get_physical_device_properties_1_2(struct anv_physical_device
*pdevice
,
1607 VkPhysicalDeviceVulkan12Properties
*p
)
1609 assert(p
->sType
== VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES
);
1611 p
->driverID
= VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA_KHR
;
1612 memset(p
->driverName
, 0, sizeof(p
->driverName
));
1613 snprintf(p
->driverName
, VK_MAX_DRIVER_NAME_SIZE_KHR
,
1614 "Intel open-source Mesa driver");
1615 memset(p
->driverInfo
, 0, sizeof(p
->driverInfo
));
1616 snprintf(p
->driverInfo
, VK_MAX_DRIVER_INFO_SIZE_KHR
,
1617 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1
);
1618 p
->conformanceVersion
= (VkConformanceVersionKHR
) {
1625 p
->denormBehaviorIndependence
=
1626 VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_ALL_KHR
;
1627 p
->roundingModeIndependence
=
1628 VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_NONE_KHR
;
1630 /* Broadwell does not support HF denorms and there are restrictions
1631 * other gens. According to Kabylake's PRM:
1633 * "math - Extended Math Function
1635 * Restriction : Half-float denorms are always retained."
1637 p
->shaderDenormFlushToZeroFloat16
= false;
1638 p
->shaderDenormPreserveFloat16
= pdevice
->info
.gen
> 8;
1639 p
->shaderRoundingModeRTEFloat16
= true;
1640 p
->shaderRoundingModeRTZFloat16
= true;
1641 p
->shaderSignedZeroInfNanPreserveFloat16
= true;
1643 p
->shaderDenormFlushToZeroFloat32
= true;
1644 p
->shaderDenormPreserveFloat32
= true;
1645 p
->shaderRoundingModeRTEFloat32
= true;
1646 p
->shaderRoundingModeRTZFloat32
= true;
1647 p
->shaderSignedZeroInfNanPreserveFloat32
= true;
1649 p
->shaderDenormFlushToZeroFloat64
= true;
1650 p
->shaderDenormPreserveFloat64
= true;
1651 p
->shaderRoundingModeRTEFloat64
= true;
1652 p
->shaderRoundingModeRTZFloat64
= true;
1653 p
->shaderSignedZeroInfNanPreserveFloat64
= true;
1655 /* It's a bit hard to exactly map our implementation to the limits
1656 * described here. The bindless surface handle in the extended
1657 * message descriptors is 20 bits and it's an index into the table of
1658 * RENDER_SURFACE_STATE structs that starts at bindless surface base
1659 * address. Given that most things consume two surface states per
1660 * view (general/sampled for textures and write-only/read-write for
1661 * images), we claim 2^19 things.
1663 * For SSBOs, we just use A64 messages so there is no real limit
1664 * there beyond the limit on the total size of a descriptor set.
1666 const unsigned max_bindless_views
= 1 << 19;
1667 p
->maxUpdateAfterBindDescriptorsInAllPools
= max_bindless_views
;
1668 p
->shaderUniformBufferArrayNonUniformIndexingNative
= false;
1669 p
->shaderSampledImageArrayNonUniformIndexingNative
= false;
1670 p
->shaderStorageBufferArrayNonUniformIndexingNative
= true;
1671 p
->shaderStorageImageArrayNonUniformIndexingNative
= false;
1672 p
->shaderInputAttachmentArrayNonUniformIndexingNative
= false;
1673 p
->robustBufferAccessUpdateAfterBind
= true;
1674 p
->quadDivergentImplicitLod
= false;
1675 p
->maxPerStageDescriptorUpdateAfterBindSamplers
= max_bindless_views
;
1676 p
->maxPerStageDescriptorUpdateAfterBindUniformBuffers
= MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
;
1677 p
->maxPerStageDescriptorUpdateAfterBindStorageBuffers
= UINT32_MAX
;
1678 p
->maxPerStageDescriptorUpdateAfterBindSampledImages
= max_bindless_views
;
1679 p
->maxPerStageDescriptorUpdateAfterBindStorageImages
= max_bindless_views
;
1680 p
->maxPerStageDescriptorUpdateAfterBindInputAttachments
= MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS
;
1681 p
->maxPerStageUpdateAfterBindResources
= UINT32_MAX
;
1682 p
->maxDescriptorSetUpdateAfterBindSamplers
= max_bindless_views
;
1683 p
->maxDescriptorSetUpdateAfterBindUniformBuffers
= 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
;
1684 p
->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2;
1685 p
->maxDescriptorSetUpdateAfterBindStorageBuffers
= UINT32_MAX
;
1686 p
->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2;
1687 p
->maxDescriptorSetUpdateAfterBindSampledImages
= max_bindless_views
;
1688 p
->maxDescriptorSetUpdateAfterBindStorageImages
= max_bindless_views
;
1689 p
->maxDescriptorSetUpdateAfterBindInputAttachments
= MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS
;
1691 /* We support all of the depth resolve modes */
1692 p
->supportedDepthResolveModes
= VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
|
1693 VK_RESOLVE_MODE_AVERAGE_BIT_KHR
|
1694 VK_RESOLVE_MODE_MIN_BIT_KHR
|
1695 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1696 /* Average doesn't make sense for stencil so we don't support that */
1697 p
->supportedStencilResolveModes
= VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
;
1698 if (pdevice
->info
.gen
>= 8) {
1699 /* The advanced stencil resolve modes currently require stencil
1700 * sampling be supported by the hardware.
1702 p
->supportedStencilResolveModes
|= VK_RESOLVE_MODE_MIN_BIT_KHR
|
1703 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1705 p
->independentResolveNone
= true;
1706 p
->independentResolve
= true;
1708 p
->filterMinmaxSingleComponentFormats
= pdevice
->info
.gen
>= 9;
1709 p
->filterMinmaxImageComponentMapping
= pdevice
->info
.gen
>= 9;
1711 p
->maxTimelineSemaphoreValueDifference
= UINT64_MAX
;
1713 p
->framebufferIntegerColorSampleCounts
=
1714 isl_device_get_sample_counts(&pdevice
->isl_dev
);
1717 void anv_GetPhysicalDeviceProperties2(
1718 VkPhysicalDevice physicalDevice
,
1719 VkPhysicalDeviceProperties2
* pProperties
)
1721 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1723 anv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
1725 VkPhysicalDeviceVulkan11Properties core_1_1
= {
1726 .sType
= VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES
,
1728 anv_get_physical_device_properties_1_1(pdevice
, &core_1_1
);
1730 VkPhysicalDeviceVulkan12Properties core_1_2
= {
1731 .sType
= VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES
,
1733 anv_get_physical_device_properties_1_2(pdevice
, &core_1_2
);
1735 #define CORE_RENAMED_PROPERTY(major, minor, ext_property, core_property) \
1736 memcpy(&properties->ext_property, &core_##major##_##minor.core_property, \
1737 sizeof(core_##major##_##minor.core_property))
1739 #define CORE_PROPERTY(major, minor, property) \
1740 CORE_RENAMED_PROPERTY(major, minor, property, property)
1742 vk_foreach_struct(ext
, pProperties
->pNext
) {
1743 switch (ext
->sType
) {
1744 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_STENCIL_RESOLVE_PROPERTIES_KHR
: {
1745 VkPhysicalDeviceDepthStencilResolvePropertiesKHR
*properties
=
1746 (VkPhysicalDeviceDepthStencilResolvePropertiesKHR
*)ext
;
1747 CORE_PROPERTY(1, 2, supportedDepthResolveModes
);
1748 CORE_PROPERTY(1, 2, supportedStencilResolveModes
);
1749 CORE_PROPERTY(1, 2, independentResolveNone
);
1750 CORE_PROPERTY(1, 2, independentResolve
);
1754 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES_EXT
: {
1755 VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*properties
=
1756 (VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*)ext
;
1757 CORE_PROPERTY(1, 2, maxUpdateAfterBindDescriptorsInAllPools
);
1758 CORE_PROPERTY(1, 2, shaderUniformBufferArrayNonUniformIndexingNative
);
1759 CORE_PROPERTY(1, 2, shaderSampledImageArrayNonUniformIndexingNative
);
1760 CORE_PROPERTY(1, 2, shaderStorageBufferArrayNonUniformIndexingNative
);
1761 CORE_PROPERTY(1, 2, shaderStorageImageArrayNonUniformIndexingNative
);
1762 CORE_PROPERTY(1, 2, shaderInputAttachmentArrayNonUniformIndexingNative
);
1763 CORE_PROPERTY(1, 2, robustBufferAccessUpdateAfterBind
);
1764 CORE_PROPERTY(1, 2, quadDivergentImplicitLod
);
1765 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindSamplers
);
1766 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindUniformBuffers
);
1767 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindStorageBuffers
);
1768 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindSampledImages
);
1769 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindStorageImages
);
1770 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindInputAttachments
);
1771 CORE_PROPERTY(1, 2, maxPerStageUpdateAfterBindResources
);
1772 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindSamplers
);
1773 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindUniformBuffers
);
1774 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindUniformBuffersDynamic
);
1775 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindStorageBuffers
);
1776 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindStorageBuffersDynamic
);
1777 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindSampledImages
);
1778 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindStorageImages
);
1779 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindInputAttachments
);
1783 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR
: {
1784 VkPhysicalDeviceDriverPropertiesKHR
*properties
=
1785 (VkPhysicalDeviceDriverPropertiesKHR
*) ext
;
1786 CORE_PROPERTY(1, 2, driverID
);
1787 CORE_PROPERTY(1, 2, driverName
);
1788 CORE_PROPERTY(1, 2, driverInfo
);
1789 CORE_PROPERTY(1, 2, conformanceVersion
);
1793 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT
: {
1794 VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*props
=
1795 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*) ext
;
1796 /* Userptr needs page aligned memory. */
1797 props
->minImportedHostPointerAlignment
= 4096;
1801 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES
: {
1802 VkPhysicalDeviceIDProperties
*properties
=
1803 (VkPhysicalDeviceIDProperties
*)ext
;
1804 CORE_PROPERTY(1, 1, deviceUUID
);
1805 CORE_PROPERTY(1, 1, driverUUID
);
1806 CORE_PROPERTY(1, 1, deviceLUID
);
1807 CORE_PROPERTY(1, 1, deviceLUIDValid
);
1811 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_PROPERTIES_EXT
: {
1812 VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*props
=
1813 (VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*)ext
;
1814 props
->maxInlineUniformBlockSize
= MAX_INLINE_UNIFORM_BLOCK_SIZE
;
1815 props
->maxPerStageDescriptorInlineUniformBlocks
=
1816 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1817 props
->maxPerStageDescriptorUpdateAfterBindInlineUniformBlocks
=
1818 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1819 props
->maxDescriptorSetInlineUniformBlocks
=
1820 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1821 props
->maxDescriptorSetUpdateAfterBindInlineUniformBlocks
=
1822 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1826 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_PROPERTIES_EXT
: {
1827 VkPhysicalDeviceLineRasterizationPropertiesEXT
*props
=
1828 (VkPhysicalDeviceLineRasterizationPropertiesEXT
*)ext
;
1829 /* In the Skylake PRM Vol. 7, subsection titled "GIQ (Diamond)
1830 * Sampling Rules - Legacy Mode", it says the following:
1832 * "Note that the device divides a pixel into a 16x16 array of
1833 * subpixels, referenced by their upper left corners."
1835 * This is the only known reference in the PRMs to the subpixel
1836 * precision of line rasterization and a "16x16 array of subpixels"
1837 * implies 4 subpixel precision bits. Empirical testing has shown
1838 * that 4 subpixel precision bits applies to all line rasterization
1841 props
->lineSubPixelPrecisionBits
= 4;
1845 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES
: {
1846 VkPhysicalDeviceMaintenance3Properties
*properties
=
1847 (VkPhysicalDeviceMaintenance3Properties
*)ext
;
1848 /* This value doesn't matter for us today as our per-stage
1849 * descriptors are the real limit.
1851 CORE_PROPERTY(1, 1, maxPerSetDescriptors
);
1852 CORE_PROPERTY(1, 1, maxMemoryAllocationSize
);
1856 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES
: {
1857 VkPhysicalDeviceMultiviewProperties
*properties
=
1858 (VkPhysicalDeviceMultiviewProperties
*)ext
;
1859 CORE_PROPERTY(1, 1, maxMultiviewViewCount
);
1860 CORE_PROPERTY(1, 1, maxMultiviewInstanceIndex
);
1864 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT
: {
1865 VkPhysicalDevicePCIBusInfoPropertiesEXT
*properties
=
1866 (VkPhysicalDevicePCIBusInfoPropertiesEXT
*)ext
;
1867 properties
->pciDomain
= pdevice
->pci_info
.domain
;
1868 properties
->pciBus
= pdevice
->pci_info
.bus
;
1869 properties
->pciDevice
= pdevice
->pci_info
.device
;
1870 properties
->pciFunction
= pdevice
->pci_info
.function
;
1874 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES
: {
1875 VkPhysicalDevicePointClippingProperties
*properties
=
1876 (VkPhysicalDevicePointClippingProperties
*) ext
;
1877 CORE_PROPERTY(1, 1, pointClippingBehavior
);
1881 #pragma GCC diagnostic push
1882 #pragma GCC diagnostic ignored "-Wswitch"
1883 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PRESENTATION_PROPERTIES_ANDROID
: {
1884 VkPhysicalDevicePresentationPropertiesANDROID
*props
=
1885 (VkPhysicalDevicePresentationPropertiesANDROID
*)ext
;
1886 props
->sharedImage
= VK_FALSE
;
1889 #pragma GCC diagnostic pop
1891 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES
: {
1892 VkPhysicalDeviceProtectedMemoryProperties
*properties
=
1893 (VkPhysicalDeviceProtectedMemoryProperties
*)ext
;
1894 CORE_PROPERTY(1, 1, protectedNoFault
);
1898 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
1899 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
1900 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
1901 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
1905 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT
: {
1906 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*properties
=
1907 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*)ext
;
1908 CORE_PROPERTY(1, 2, filterMinmaxImageComponentMapping
);
1909 CORE_PROPERTY(1, 2, filterMinmaxSingleComponentFormats
);
1913 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES
: {
1914 VkPhysicalDeviceSubgroupProperties
*properties
= (void *)ext
;
1915 CORE_PROPERTY(1, 1, subgroupSize
);
1916 CORE_RENAMED_PROPERTY(1, 1, supportedStages
,
1917 subgroupSupportedStages
);
1918 CORE_RENAMED_PROPERTY(1, 1, supportedOperations
,
1919 subgroupSupportedOperations
);
1920 CORE_RENAMED_PROPERTY(1, 1, quadOperationsInAllStages
,
1921 subgroupQuadOperationsInAllStages
);
1925 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_PROPERTIES_EXT
: {
1926 VkPhysicalDeviceSubgroupSizeControlPropertiesEXT
*props
=
1927 (VkPhysicalDeviceSubgroupSizeControlPropertiesEXT
*)ext
;
1928 STATIC_ASSERT(8 <= BRW_SUBGROUP_SIZE
&& BRW_SUBGROUP_SIZE
<= 32);
1929 props
->minSubgroupSize
= 8;
1930 props
->maxSubgroupSize
= 32;
1931 props
->maxComputeWorkgroupSubgroups
= pdevice
->info
.max_cs_threads
;
1932 props
->requiredSubgroupSizeStages
= VK_SHADER_STAGE_COMPUTE_BIT
;
1935 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT_CONTROLS_PROPERTIES_KHR
: {
1936 VkPhysicalDeviceFloatControlsPropertiesKHR
*properties
= (void *)ext
;
1937 CORE_PROPERTY(1, 2, denormBehaviorIndependence
);
1938 CORE_PROPERTY(1, 2, roundingModeIndependence
);
1939 CORE_PROPERTY(1, 2, shaderDenormFlushToZeroFloat16
);
1940 CORE_PROPERTY(1, 2, shaderDenormPreserveFloat16
);
1941 CORE_PROPERTY(1, 2, shaderRoundingModeRTEFloat16
);
1942 CORE_PROPERTY(1, 2, shaderRoundingModeRTZFloat16
);
1943 CORE_PROPERTY(1, 2, shaderSignedZeroInfNanPreserveFloat16
);
1944 CORE_PROPERTY(1, 2, shaderDenormFlushToZeroFloat32
);
1945 CORE_PROPERTY(1, 2, shaderDenormPreserveFloat32
);
1946 CORE_PROPERTY(1, 2, shaderRoundingModeRTEFloat32
);
1947 CORE_PROPERTY(1, 2, shaderRoundingModeRTZFloat32
);
1948 CORE_PROPERTY(1, 2, shaderSignedZeroInfNanPreserveFloat32
);
1949 CORE_PROPERTY(1, 2, shaderDenormFlushToZeroFloat64
);
1950 CORE_PROPERTY(1, 2, shaderDenormPreserveFloat64
);
1951 CORE_PROPERTY(1, 2, shaderRoundingModeRTEFloat64
);
1952 CORE_PROPERTY(1, 2, shaderRoundingModeRTZFloat64
);
1953 CORE_PROPERTY(1, 2, shaderSignedZeroInfNanPreserveFloat64
);
1957 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_PROPERTIES_EXT
: {
1958 VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT
*props
=
1959 (VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT
*)ext
;
1961 /* From the SKL PRM Vol. 2d, docs for RENDER_SURFACE_STATE::Surface
1964 * "For SURFTYPE_BUFFER non-rendertarget surfaces, this field
1965 * specifies the base address of the first element of the surface,
1966 * computed in software by adding the surface base address to the
1967 * byte offset of the element in the buffer. The base address must
1968 * be aligned to element size."
1970 * The typed dataport messages require that things be texel aligned.
1971 * Otherwise, we may just load/store the wrong data or, in the worst
1972 * case, there may be hangs.
1974 props
->storageTexelBufferOffsetAlignmentBytes
= 16;
1975 props
->storageTexelBufferOffsetSingleTexelAlignment
= true;
1977 /* The sampler, however, is much more forgiving and it can handle
1978 * arbitrary byte alignment for linear and buffer surfaces. It's
1979 * hard to find a good PRM citation for this but years of empirical
1980 * experience demonstrate that this is true.
1982 props
->uniformTexelBufferOffsetAlignmentBytes
= 1;
1983 props
->uniformTexelBufferOffsetSingleTexelAlignment
= false;
1987 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_PROPERTIES_KHR
: {
1988 VkPhysicalDeviceTimelineSemaphorePropertiesKHR
*properties
=
1989 (VkPhysicalDeviceTimelineSemaphorePropertiesKHR
*) ext
;
1990 CORE_PROPERTY(1, 2, maxTimelineSemaphoreValueDifference
);
1994 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT
: {
1995 VkPhysicalDeviceTransformFeedbackPropertiesEXT
*props
=
1996 (VkPhysicalDeviceTransformFeedbackPropertiesEXT
*)ext
;
1998 props
->maxTransformFeedbackStreams
= MAX_XFB_STREAMS
;
1999 props
->maxTransformFeedbackBuffers
= MAX_XFB_BUFFERS
;
2000 props
->maxTransformFeedbackBufferSize
= (1ull << 32);
2001 props
->maxTransformFeedbackStreamDataSize
= 128 * 4;
2002 props
->maxTransformFeedbackBufferDataSize
= 128 * 4;
2003 props
->maxTransformFeedbackBufferDataStride
= 2048;
2004 props
->transformFeedbackQueries
= true;
2005 props
->transformFeedbackStreamsLinesTriangles
= false;
2006 props
->transformFeedbackRasterizationStreamSelect
= false;
2007 props
->transformFeedbackDraw
= true;
2011 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT
: {
2012 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*props
=
2013 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*)ext
;
2014 /* We have to restrict this a bit for multiview */
2015 props
->maxVertexAttribDivisor
= UINT32_MAX
/ 16;
2019 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES
:
2020 anv_get_physical_device_properties_1_1(pdevice
, (void *)ext
);
2023 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES
:
2024 anv_get_physical_device_properties_1_2(pdevice
, (void *)ext
);
2028 anv_debug_ignored_stype(ext
->sType
);
2033 #undef CORE_RENAMED_PROPERTY
2034 #undef CORE_PROPERTY
2037 /* We support exactly one queue family. */
2038 static const VkQueueFamilyProperties
2039 anv_queue_family_properties
= {
2040 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
2041 VK_QUEUE_COMPUTE_BIT
|
2042 VK_QUEUE_TRANSFER_BIT
,
2044 .timestampValidBits
= 36, /* XXX: Real value here */
2045 .minImageTransferGranularity
= { 1, 1, 1 },
2048 void anv_GetPhysicalDeviceQueueFamilyProperties(
2049 VkPhysicalDevice physicalDevice
,
2051 VkQueueFamilyProperties
* pQueueFamilyProperties
)
2053 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pCount
);
2055 vk_outarray_append(&out
, p
) {
2056 *p
= anv_queue_family_properties
;
2060 void anv_GetPhysicalDeviceQueueFamilyProperties2(
2061 VkPhysicalDevice physicalDevice
,
2062 uint32_t* pQueueFamilyPropertyCount
,
2063 VkQueueFamilyProperties2
* pQueueFamilyProperties
)
2066 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pQueueFamilyPropertyCount
);
2068 vk_outarray_append(&out
, p
) {
2069 p
->queueFamilyProperties
= anv_queue_family_properties
;
2071 vk_foreach_struct(s
, p
->pNext
) {
2072 anv_debug_ignored_stype(s
->sType
);
2077 void anv_GetPhysicalDeviceMemoryProperties(
2078 VkPhysicalDevice physicalDevice
,
2079 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
2081 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
2083 pMemoryProperties
->memoryTypeCount
= physical_device
->memory
.type_count
;
2084 for (uint32_t i
= 0; i
< physical_device
->memory
.type_count
; i
++) {
2085 pMemoryProperties
->memoryTypes
[i
] = (VkMemoryType
) {
2086 .propertyFlags
= physical_device
->memory
.types
[i
].propertyFlags
,
2087 .heapIndex
= physical_device
->memory
.types
[i
].heapIndex
,
2091 pMemoryProperties
->memoryHeapCount
= physical_device
->memory
.heap_count
;
2092 for (uint32_t i
= 0; i
< physical_device
->memory
.heap_count
; i
++) {
2093 pMemoryProperties
->memoryHeaps
[i
] = (VkMemoryHeap
) {
2094 .size
= physical_device
->memory
.heaps
[i
].size
,
2095 .flags
= physical_device
->memory
.heaps
[i
].flags
,
2101 anv_get_memory_budget(VkPhysicalDevice physicalDevice
,
2102 VkPhysicalDeviceMemoryBudgetPropertiesEXT
*memoryBudget
)
2104 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
2105 uint64_t sys_available
= get_available_system_memory();
2106 assert(sys_available
> 0);
2108 VkDeviceSize total_heaps_size
= 0;
2109 for (size_t i
= 0; i
< device
->memory
.heap_count
; i
++)
2110 total_heaps_size
+= device
->memory
.heaps
[i
].size
;
2112 for (size_t i
= 0; i
< device
->memory
.heap_count
; i
++) {
2113 VkDeviceSize heap_size
= device
->memory
.heaps
[i
].size
;
2114 VkDeviceSize heap_used
= device
->memory
.heaps
[i
].used
;
2115 VkDeviceSize heap_budget
;
2117 double heap_proportion
= (double) heap_size
/ total_heaps_size
;
2118 VkDeviceSize sys_available_prop
= sys_available
* heap_proportion
;
2121 * Let's not incite the app to starve the system: report at most 90% of
2122 * available system memory.
2124 uint64_t heap_available
= sys_available_prop
* 9 / 10;
2125 heap_budget
= MIN2(heap_size
, heap_used
+ heap_available
);
2128 * Round down to the nearest MB
2130 heap_budget
&= ~((1ull << 20) - 1);
2133 * The heapBudget value must be non-zero for array elements less than
2134 * VkPhysicalDeviceMemoryProperties::memoryHeapCount. The heapBudget
2135 * value must be less than or equal to VkMemoryHeap::size for each heap.
2137 assert(0 < heap_budget
&& heap_budget
<= heap_size
);
2139 memoryBudget
->heapUsage
[i
] = heap_used
;
2140 memoryBudget
->heapBudget
[i
] = heap_budget
;
2143 /* The heapBudget and heapUsage values must be zero for array elements
2144 * greater than or equal to VkPhysicalDeviceMemoryProperties::memoryHeapCount
2146 for (uint32_t i
= device
->memory
.heap_count
; i
< VK_MAX_MEMORY_HEAPS
; i
++) {
2147 memoryBudget
->heapBudget
[i
] = 0;
2148 memoryBudget
->heapUsage
[i
] = 0;
2152 void anv_GetPhysicalDeviceMemoryProperties2(
2153 VkPhysicalDevice physicalDevice
,
2154 VkPhysicalDeviceMemoryProperties2
* pMemoryProperties
)
2156 anv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
2157 &pMemoryProperties
->memoryProperties
);
2159 vk_foreach_struct(ext
, pMemoryProperties
->pNext
) {
2160 switch (ext
->sType
) {
2161 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT
:
2162 anv_get_memory_budget(physicalDevice
, (void*)ext
);
2165 anv_debug_ignored_stype(ext
->sType
);
2172 anv_GetDeviceGroupPeerMemoryFeatures(
2175 uint32_t localDeviceIndex
,
2176 uint32_t remoteDeviceIndex
,
2177 VkPeerMemoryFeatureFlags
* pPeerMemoryFeatures
)
2179 assert(localDeviceIndex
== 0 && remoteDeviceIndex
== 0);
2180 *pPeerMemoryFeatures
= VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT
|
2181 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT
|
2182 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT
|
2183 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT
;
2186 PFN_vkVoidFunction
anv_GetInstanceProcAddr(
2187 VkInstance _instance
,
2190 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2192 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
2193 * when we have to return valid function pointers, NULL, or it's left
2194 * undefined. See the table for exact details.
2199 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
2200 if (strcmp(pName, "vk" #entrypoint) == 0) \
2201 return (PFN_vkVoidFunction)anv_##entrypoint
2203 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties
);
2204 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties
);
2205 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceVersion
);
2206 LOOKUP_ANV_ENTRYPOINT(CreateInstance
);
2208 /* GetInstanceProcAddr() can also be called with a NULL instance.
2209 * See https://gitlab.khronos.org/vulkan/vulkan/issues/2057
2211 LOOKUP_ANV_ENTRYPOINT(GetInstanceProcAddr
);
2213 #undef LOOKUP_ANV_ENTRYPOINT
2215 if (instance
== NULL
)
2218 int idx
= anv_get_instance_entrypoint_index(pName
);
2220 return instance
->dispatch
.entrypoints
[idx
];
2222 idx
= anv_get_physical_device_entrypoint_index(pName
);
2224 return instance
->physical_device_dispatch
.entrypoints
[idx
];
2226 idx
= anv_get_device_entrypoint_index(pName
);
2228 return instance
->device_dispatch
.entrypoints
[idx
];
2233 /* With version 1+ of the loader interface the ICD should expose
2234 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
2237 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
2238 VkInstance instance
,
2242 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
2243 VkInstance instance
,
2246 return anv_GetInstanceProcAddr(instance
, pName
);
2249 PFN_vkVoidFunction
anv_GetDeviceProcAddr(
2253 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2255 if (!device
|| !pName
)
2258 int idx
= anv_get_device_entrypoint_index(pName
);
2262 return device
->dispatch
.entrypoints
[idx
];
2265 /* With version 4+ of the loader interface the ICD should expose
2266 * vk_icdGetPhysicalDeviceProcAddr()
2269 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetPhysicalDeviceProcAddr(
2270 VkInstance _instance
,
2273 PFN_vkVoidFunction
vk_icdGetPhysicalDeviceProcAddr(
2274 VkInstance _instance
,
2277 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2279 if (!pName
|| !instance
)
2282 int idx
= anv_get_physical_device_entrypoint_index(pName
);
2286 return instance
->physical_device_dispatch
.entrypoints
[idx
];
2291 anv_CreateDebugReportCallbackEXT(VkInstance _instance
,
2292 const VkDebugReportCallbackCreateInfoEXT
* pCreateInfo
,
2293 const VkAllocationCallbacks
* pAllocator
,
2294 VkDebugReportCallbackEXT
* pCallback
)
2296 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2297 return vk_create_debug_report_callback(&instance
->debug_report_callbacks
,
2298 pCreateInfo
, pAllocator
, &instance
->alloc
,
2303 anv_DestroyDebugReportCallbackEXT(VkInstance _instance
,
2304 VkDebugReportCallbackEXT _callback
,
2305 const VkAllocationCallbacks
* pAllocator
)
2307 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2308 vk_destroy_debug_report_callback(&instance
->debug_report_callbacks
,
2309 _callback
, pAllocator
, &instance
->alloc
);
2313 anv_DebugReportMessageEXT(VkInstance _instance
,
2314 VkDebugReportFlagsEXT flags
,
2315 VkDebugReportObjectTypeEXT objectType
,
2318 int32_t messageCode
,
2319 const char* pLayerPrefix
,
2320 const char* pMessage
)
2322 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2323 vk_debug_report(&instance
->debug_report_callbacks
, flags
, objectType
,
2324 object
, location
, messageCode
, pLayerPrefix
, pMessage
);
2327 static struct anv_state
2328 anv_state_pool_emit_data(struct anv_state_pool
*pool
, size_t size
, size_t align
, const void *p
)
2330 struct anv_state state
;
2332 state
= anv_state_pool_alloc(pool
, size
, align
);
2333 memcpy(state
.map
, p
, size
);
2338 /* Haswell border color is a bit of a disaster. Float and unorm formats use a
2339 * straightforward 32-bit float color in the first 64 bytes. Instead of using
2340 * a nice float/integer union like Gen8+, Haswell specifies the integer border
2341 * color as a separate entry /after/ the float color. The layout of this entry
2342 * also depends on the format's bpp (with extra hacks for RG32), and overlaps.
2344 * Since we don't know the format/bpp, we can't make any of the border colors
2345 * containing '1' work for all formats, as it would be in the wrong place for
2346 * some of them. We opt to make 32-bit integers work as this seems like the
2347 * most common option. Fortunately, transparent black works regardless, as
2348 * all zeroes is the same in every bit-size.
2350 struct hsw_border_color
{
2354 uint32_t _pad1
[108];
2357 struct gen8_border_color
{
2362 /* Pad out to 64 bytes */
2367 anv_device_init_border_colors(struct anv_device
*device
)
2369 if (device
->info
.is_haswell
) {
2370 static const struct hsw_border_color border_colors
[] = {
2371 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
2372 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
2373 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
2374 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
2375 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
2376 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
2379 device
->border_colors
=
2380 anv_state_pool_emit_data(&device
->dynamic_state_pool
,
2381 sizeof(border_colors
), 512, border_colors
);
2383 static const struct gen8_border_color border_colors
[] = {
2384 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
2385 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
2386 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
2387 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
2388 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
2389 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
2392 device
->border_colors
=
2393 anv_state_pool_emit_data(&device
->dynamic_state_pool
,
2394 sizeof(border_colors
), 64, border_colors
);
2399 anv_device_init_trivial_batch(struct anv_device
*device
)
2401 VkResult result
= anv_device_alloc_bo(device
, 4096,
2402 ANV_BO_ALLOC_MAPPED
,
2403 0 /* explicit_address */,
2404 &device
->trivial_batch_bo
);
2405 if (result
!= VK_SUCCESS
)
2408 struct anv_batch batch
= {
2409 .start
= device
->trivial_batch_bo
->map
,
2410 .next
= device
->trivial_batch_bo
->map
,
2411 .end
= device
->trivial_batch_bo
->map
+ 4096,
2414 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
2415 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
2417 if (!device
->info
.has_llc
)
2418 gen_clflush_range(batch
.start
, batch
.next
- batch
.start
);
2423 VkResult
anv_EnumerateDeviceExtensionProperties(
2424 VkPhysicalDevice physicalDevice
,
2425 const char* pLayerName
,
2426 uint32_t* pPropertyCount
,
2427 VkExtensionProperties
* pProperties
)
2429 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
2430 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
2432 for (int i
= 0; i
< ANV_DEVICE_EXTENSION_COUNT
; i
++) {
2433 if (device
->supported_extensions
.extensions
[i
]) {
2434 vk_outarray_append(&out
, prop
) {
2435 *prop
= anv_device_extensions
[i
];
2440 return vk_outarray_status(&out
);
2444 anv_device_init_dispatch(struct anv_device
*device
)
2446 const struct anv_instance
*instance
= device
->physical
->instance
;
2448 const struct anv_device_dispatch_table
*genX_table
;
2449 switch (device
->info
.gen
) {
2451 genX_table
= &gen12_device_dispatch_table
;
2454 genX_table
= &gen11_device_dispatch_table
;
2457 genX_table
= &gen10_device_dispatch_table
;
2460 genX_table
= &gen9_device_dispatch_table
;
2463 genX_table
= &gen8_device_dispatch_table
;
2466 if (device
->info
.is_haswell
)
2467 genX_table
= &gen75_device_dispatch_table
;
2469 genX_table
= &gen7_device_dispatch_table
;
2472 unreachable("unsupported gen\n");
2475 for (unsigned i
= 0; i
< ARRAY_SIZE(device
->dispatch
.entrypoints
); i
++) {
2476 /* Vulkan requires that entrypoints for extensions which have not been
2477 * enabled must not be advertised.
2479 if (!anv_device_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
2480 &instance
->enabled_extensions
,
2481 &device
->enabled_extensions
)) {
2482 device
->dispatch
.entrypoints
[i
] = NULL
;
2483 } else if (genX_table
->entrypoints
[i
]) {
2484 device
->dispatch
.entrypoints
[i
] = genX_table
->entrypoints
[i
];
2486 device
->dispatch
.entrypoints
[i
] =
2487 anv_device_dispatch_table
.entrypoints
[i
];
2493 vk_priority_to_gen(int priority
)
2496 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT
:
2497 return GEN_CONTEXT_LOW_PRIORITY
;
2498 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
:
2499 return GEN_CONTEXT_MEDIUM_PRIORITY
;
2500 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT
:
2501 return GEN_CONTEXT_HIGH_PRIORITY
;
2502 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT
:
2503 return GEN_CONTEXT_REALTIME_PRIORITY
;
2505 unreachable("Invalid priority");
2510 anv_device_init_hiz_clear_value_bo(struct anv_device
*device
)
2512 VkResult result
= anv_device_alloc_bo(device
, 4096,
2513 ANV_BO_ALLOC_MAPPED
,
2514 0 /* explicit_address */,
2515 &device
->hiz_clear_bo
);
2516 if (result
!= VK_SUCCESS
)
2519 union isl_color_value hiz_clear
= { .u32
= { 0, } };
2520 hiz_clear
.f32
[0] = ANV_HZ_FC_VAL
;
2522 memcpy(device
->hiz_clear_bo
->map
, hiz_clear
.u32
, sizeof(hiz_clear
.u32
));
2524 if (!device
->info
.has_llc
)
2525 gen_clflush_range(device
->hiz_clear_bo
->map
, sizeof(hiz_clear
.u32
));
2531 get_bo_from_pool(struct gen_batch_decode_bo
*ret
,
2532 struct anv_block_pool
*pool
,
2535 anv_block_pool_foreach_bo(bo
, pool
) {
2536 uint64_t bo_address
= gen_48b_address(bo
->offset
);
2537 if (address
>= bo_address
&& address
< (bo_address
+ bo
->size
)) {
2538 *ret
= (struct gen_batch_decode_bo
) {
2549 /* Finding a buffer for batch decoding */
2550 static struct gen_batch_decode_bo
2551 decode_get_bo(void *v_batch
, bool ppgtt
, uint64_t address
)
2553 struct anv_device
*device
= v_batch
;
2554 struct gen_batch_decode_bo ret_bo
= {};
2558 if (get_bo_from_pool(&ret_bo
, &device
->dynamic_state_pool
.block_pool
, address
))
2560 if (get_bo_from_pool(&ret_bo
, &device
->instruction_state_pool
.block_pool
, address
))
2562 if (get_bo_from_pool(&ret_bo
, &device
->binding_table_pool
.block_pool
, address
))
2564 if (get_bo_from_pool(&ret_bo
, &device
->surface_state_pool
.block_pool
, address
))
2567 if (!device
->cmd_buffer_being_decoded
)
2568 return (struct gen_batch_decode_bo
) { };
2570 struct anv_batch_bo
**bo
;
2572 u_vector_foreach(bo
, &device
->cmd_buffer_being_decoded
->seen_bbos
) {
2573 /* The decoder zeroes out the top 16 bits, so we need to as well */
2574 uint64_t bo_address
= (*bo
)->bo
->offset
& (~0ull >> 16);
2576 if (address
>= bo_address
&& address
< bo_address
+ (*bo
)->bo
->size
) {
2577 return (struct gen_batch_decode_bo
) {
2579 .size
= (*bo
)->bo
->size
,
2580 .map
= (*bo
)->bo
->map
,
2585 return (struct gen_batch_decode_bo
) { };
2588 struct gen_aux_map_buffer
{
2589 struct gen_buffer base
;
2590 struct anv_state state
;
2593 static struct gen_buffer
*
2594 gen_aux_map_buffer_alloc(void *driver_ctx
, uint32_t size
)
2596 struct gen_aux_map_buffer
*buf
= malloc(sizeof(struct gen_aux_map_buffer
));
2600 struct anv_device
*device
= (struct anv_device
*)driver_ctx
;
2601 assert(device
->physical
->supports_48bit_addresses
&&
2602 device
->physical
->use_softpin
);
2604 struct anv_state_pool
*pool
= &device
->dynamic_state_pool
;
2605 buf
->state
= anv_state_pool_alloc(pool
, size
, size
);
2607 buf
->base
.gpu
= pool
->block_pool
.bo
->offset
+ buf
->state
.offset
;
2608 buf
->base
.gpu_end
= buf
->base
.gpu
+ buf
->state
.alloc_size
;
2609 buf
->base
.map
= buf
->state
.map
;
2610 buf
->base
.driver_bo
= &buf
->state
;
2615 gen_aux_map_buffer_free(void *driver_ctx
, struct gen_buffer
*buffer
)
2617 struct gen_aux_map_buffer
*buf
= (struct gen_aux_map_buffer
*)buffer
;
2618 struct anv_device
*device
= (struct anv_device
*)driver_ctx
;
2619 struct anv_state_pool
*pool
= &device
->dynamic_state_pool
;
2620 anv_state_pool_free(pool
, buf
->state
);
2624 static struct gen_mapped_pinned_buffer_alloc aux_map_allocator
= {
2625 .alloc
= gen_aux_map_buffer_alloc
,
2626 .free
= gen_aux_map_buffer_free
,
2630 check_physical_device_features(VkPhysicalDevice physicalDevice
,
2631 const VkPhysicalDeviceFeatures
*features
)
2633 VkPhysicalDeviceFeatures supported_features
;
2634 anv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
2635 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
2636 VkBool32
*enabled_feature
= (VkBool32
*)features
;
2637 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
2638 for (uint32_t i
= 0; i
< num_features
; i
++) {
2639 if (enabled_feature
[i
] && !supported_feature
[i
])
2640 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
2646 VkResult
anv_CreateDevice(
2647 VkPhysicalDevice physicalDevice
,
2648 const VkDeviceCreateInfo
* pCreateInfo
,
2649 const VkAllocationCallbacks
* pAllocator
,
2652 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
2654 struct anv_device
*device
;
2656 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO
);
2658 struct anv_device_extension_table enabled_extensions
= { };
2659 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
2661 for (idx
= 0; idx
< ANV_DEVICE_EXTENSION_COUNT
; idx
++) {
2662 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
2663 anv_device_extensions
[idx
].extensionName
) == 0)
2667 if (idx
>= ANV_DEVICE_EXTENSION_COUNT
)
2668 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
2670 if (!physical_device
->supported_extensions
.extensions
[idx
])
2671 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
2673 enabled_extensions
.extensions
[idx
] = true;
2676 /* Check enabled features */
2677 bool robust_buffer_access
= false;
2678 if (pCreateInfo
->pEnabledFeatures
) {
2679 result
= check_physical_device_features(physicalDevice
,
2680 pCreateInfo
->pEnabledFeatures
);
2681 if (result
!= VK_SUCCESS
)
2684 if (pCreateInfo
->pEnabledFeatures
->robustBufferAccess
)
2685 robust_buffer_access
= true;
2688 vk_foreach_struct_const(ext
, pCreateInfo
->pNext
) {
2689 switch (ext
->sType
) {
2690 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2
: {
2691 const VkPhysicalDeviceFeatures2
*features
= (const void *)ext
;
2692 result
= check_physical_device_features(physicalDevice
,
2693 &features
->features
);
2694 if (result
!= VK_SUCCESS
)
2697 if (features
->features
.robustBufferAccess
)
2698 robust_buffer_access
= true;
2708 /* Check requested queues and fail if we are requested to create any
2709 * queues with flags we don't support.
2711 assert(pCreateInfo
->queueCreateInfoCount
> 0);
2712 for (uint32_t i
= 0; i
< pCreateInfo
->queueCreateInfoCount
; i
++) {
2713 if (pCreateInfo
->pQueueCreateInfos
[i
].flags
!= 0)
2714 return vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2717 /* Check if client specified queue priority. */
2718 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*queue_priority
=
2719 vk_find_struct_const(pCreateInfo
->pQueueCreateInfos
[0].pNext
,
2720 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT
);
2722 VkQueueGlobalPriorityEXT priority
=
2723 queue_priority
? queue_priority
->globalPriority
:
2724 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
;
2726 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
2728 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
2730 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2732 if (INTEL_DEBUG
& DEBUG_BATCH
) {
2733 const unsigned decode_flags
=
2734 GEN_BATCH_DECODE_FULL
|
2735 ((INTEL_DEBUG
& DEBUG_COLOR
) ? GEN_BATCH_DECODE_IN_COLOR
: 0) |
2736 GEN_BATCH_DECODE_OFFSETS
|
2737 GEN_BATCH_DECODE_FLOATS
;
2739 gen_batch_decode_ctx_init(&device
->decoder_ctx
,
2740 &physical_device
->info
,
2741 stderr
, decode_flags
, NULL
,
2742 decode_get_bo
, NULL
, device
);
2745 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
2746 device
->physical
= physical_device
;
2747 device
->no_hw
= physical_device
->no_hw
;
2748 device
->_lost
= false;
2751 device
->alloc
= *pAllocator
;
2753 device
->alloc
= physical_device
->instance
->alloc
;
2755 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
2756 device
->fd
= open(physical_device
->path
, O_RDWR
| O_CLOEXEC
);
2757 if (device
->fd
== -1) {
2758 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2762 device
->context_id
= anv_gem_create_context(device
);
2763 if (device
->context_id
== -1) {
2764 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2768 result
= anv_queue_init(device
, &device
->queue
);
2769 if (result
!= VK_SUCCESS
)
2770 goto fail_context_id
;
2772 if (physical_device
->use_softpin
) {
2773 if (pthread_mutex_init(&device
->vma_mutex
, NULL
) != 0) {
2774 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2778 /* keep the page with address zero out of the allocator */
2779 util_vma_heap_init(&device
->vma_lo
,
2780 LOW_HEAP_MIN_ADDRESS
, LOW_HEAP_SIZE
);
2782 util_vma_heap_init(&device
->vma_cva
, CLIENT_VISIBLE_HEAP_MIN_ADDRESS
,
2783 CLIENT_VISIBLE_HEAP_SIZE
);
2785 /* Leave the last 4GiB out of the high vma range, so that no state
2786 * base address + size can overflow 48 bits. For more information see
2787 * the comment about Wa32bitGeneralStateOffset in anv_allocator.c
2789 util_vma_heap_init(&device
->vma_hi
, HIGH_HEAP_MIN_ADDRESS
,
2790 physical_device
->gtt_size
- (1ull << 32) -
2791 HIGH_HEAP_MIN_ADDRESS
);
2794 list_inithead(&device
->memory_objects
);
2796 /* As per spec, the driver implementation may deny requests to acquire
2797 * a priority above the default priority (MEDIUM) if the caller does not
2798 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
2801 if (physical_device
->has_context_priority
) {
2802 int err
= anv_gem_set_context_param(device
->fd
, device
->context_id
,
2803 I915_CONTEXT_PARAM_PRIORITY
,
2804 vk_priority_to_gen(priority
));
2805 if (err
!= 0 && priority
> VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
) {
2806 result
= vk_error(VK_ERROR_NOT_PERMITTED_EXT
);
2811 device
->info
= physical_device
->info
;
2812 device
->isl_dev
= physical_device
->isl_dev
;
2814 /* On Broadwell and later, we can use batch chaining to more efficiently
2815 * implement growing command buffers. Prior to Haswell, the kernel
2816 * command parser gets in the way and we have to fall back to growing
2819 device
->can_chain_batches
= device
->info
.gen
>= 8;
2821 device
->robust_buffer_access
= robust_buffer_access
;
2822 device
->enabled_extensions
= enabled_extensions
;
2824 anv_device_init_dispatch(device
);
2826 if (pthread_mutex_init(&device
->mutex
, NULL
) != 0) {
2827 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2831 pthread_condattr_t condattr
;
2832 if (pthread_condattr_init(&condattr
) != 0) {
2833 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2836 if (pthread_condattr_setclock(&condattr
, CLOCK_MONOTONIC
) != 0) {
2837 pthread_condattr_destroy(&condattr
);
2838 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2841 if (pthread_cond_init(&device
->queue_submit
, &condattr
) != 0) {
2842 pthread_condattr_destroy(&condattr
);
2843 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2846 pthread_condattr_destroy(&condattr
);
2848 result
= anv_bo_cache_init(&device
->bo_cache
);
2849 if (result
!= VK_SUCCESS
)
2850 goto fail_queue_cond
;
2852 anv_bo_pool_init(&device
->batch_bo_pool
, device
);
2854 result
= anv_state_pool_init(&device
->dynamic_state_pool
, device
,
2855 DYNAMIC_STATE_POOL_MIN_ADDRESS
, 16384);
2856 if (result
!= VK_SUCCESS
)
2857 goto fail_batch_bo_pool
;
2859 result
= anv_state_pool_init(&device
->instruction_state_pool
, device
,
2860 INSTRUCTION_STATE_POOL_MIN_ADDRESS
, 16384);
2861 if (result
!= VK_SUCCESS
)
2862 goto fail_dynamic_state_pool
;
2864 result
= anv_state_pool_init(&device
->surface_state_pool
, device
,
2865 SURFACE_STATE_POOL_MIN_ADDRESS
, 4096);
2866 if (result
!= VK_SUCCESS
)
2867 goto fail_instruction_state_pool
;
2869 if (physical_device
->use_softpin
) {
2870 result
= anv_state_pool_init(&device
->binding_table_pool
, device
,
2871 BINDING_TABLE_POOL_MIN_ADDRESS
, 4096);
2872 if (result
!= VK_SUCCESS
)
2873 goto fail_surface_state_pool
;
2876 if (device
->info
.gen
>= 12) {
2877 device
->aux_map_ctx
= gen_aux_map_init(device
, &aux_map_allocator
,
2878 &physical_device
->info
);
2879 if (!device
->aux_map_ctx
)
2880 goto fail_binding_table_pool
;
2883 result
= anv_device_alloc_bo(device
, 4096, 0 /* flags */,
2884 0 /* explicit_address */,
2885 &device
->workaround_bo
);
2886 if (result
!= VK_SUCCESS
)
2887 goto fail_surface_aux_map_pool
;
2889 result
= anv_device_init_trivial_batch(device
);
2890 if (result
!= VK_SUCCESS
)
2891 goto fail_workaround_bo
;
2893 if (device
->info
.gen
>= 10) {
2894 result
= anv_device_init_hiz_clear_value_bo(device
);
2895 if (result
!= VK_SUCCESS
)
2896 goto fail_trivial_batch_bo
;
2899 anv_scratch_pool_init(device
, &device
->scratch_pool
);
2901 switch (device
->info
.gen
) {
2903 if (!device
->info
.is_haswell
)
2904 result
= gen7_init_device_state(device
);
2906 result
= gen75_init_device_state(device
);
2909 result
= gen8_init_device_state(device
);
2912 result
= gen9_init_device_state(device
);
2915 result
= gen10_init_device_state(device
);
2918 result
= gen11_init_device_state(device
);
2921 result
= gen12_init_device_state(device
);
2924 /* Shouldn't get here as we don't create physical devices for any other
2926 unreachable("unhandled gen");
2928 if (result
!= VK_SUCCESS
)
2929 goto fail_workaround_bo
;
2931 anv_pipeline_cache_init(&device
->default_pipeline_cache
, device
, true);
2933 anv_device_init_blorp(device
);
2935 anv_device_init_border_colors(device
);
2937 anv_device_perf_init(device
);
2939 *pDevice
= anv_device_to_handle(device
);
2944 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
2945 if (device
->info
.gen
>= 10)
2946 anv_device_release_bo(device
, device
->hiz_clear_bo
);
2947 anv_device_release_bo(device
, device
->workaround_bo
);
2948 fail_trivial_batch_bo
:
2949 anv_device_release_bo(device
, device
->trivial_batch_bo
);
2950 fail_surface_aux_map_pool
:
2951 if (device
->info
.gen
>= 12) {
2952 gen_aux_map_finish(device
->aux_map_ctx
);
2953 device
->aux_map_ctx
= NULL
;
2955 fail_binding_table_pool
:
2956 if (physical_device
->use_softpin
)
2957 anv_state_pool_finish(&device
->binding_table_pool
);
2958 fail_surface_state_pool
:
2959 anv_state_pool_finish(&device
->surface_state_pool
);
2960 fail_instruction_state_pool
:
2961 anv_state_pool_finish(&device
->instruction_state_pool
);
2962 fail_dynamic_state_pool
:
2963 anv_state_pool_finish(&device
->dynamic_state_pool
);
2965 anv_bo_pool_finish(&device
->batch_bo_pool
);
2966 anv_bo_cache_finish(&device
->bo_cache
);
2968 pthread_cond_destroy(&device
->queue_submit
);
2970 pthread_mutex_destroy(&device
->mutex
);
2972 if (physical_device
->use_softpin
) {
2973 util_vma_heap_finish(&device
->vma_hi
);
2974 util_vma_heap_finish(&device
->vma_cva
);
2975 util_vma_heap_finish(&device
->vma_lo
);
2978 anv_queue_finish(&device
->queue
);
2980 anv_gem_destroy_context(device
, device
->context_id
);
2984 vk_free(&device
->alloc
, device
);
2989 void anv_DestroyDevice(
2991 const VkAllocationCallbacks
* pAllocator
)
2993 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2998 anv_device_finish_blorp(device
);
3000 anv_pipeline_cache_finish(&device
->default_pipeline_cache
);
3002 anv_queue_finish(&device
->queue
);
3004 #ifdef HAVE_VALGRIND
3005 /* We only need to free these to prevent valgrind errors. The backing
3006 * BO will go away in a couple of lines so we don't actually leak.
3008 anv_state_pool_free(&device
->dynamic_state_pool
, device
->border_colors
);
3009 anv_state_pool_free(&device
->dynamic_state_pool
, device
->slice_hash
);
3012 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
3014 anv_device_release_bo(device
, device
->workaround_bo
);
3015 anv_device_release_bo(device
, device
->trivial_batch_bo
);
3016 if (device
->info
.gen
>= 10)
3017 anv_device_release_bo(device
, device
->hiz_clear_bo
);
3019 if (device
->info
.gen
>= 12) {
3020 gen_aux_map_finish(device
->aux_map_ctx
);
3021 device
->aux_map_ctx
= NULL
;
3024 if (device
->physical
->use_softpin
)
3025 anv_state_pool_finish(&device
->binding_table_pool
);
3026 anv_state_pool_finish(&device
->surface_state_pool
);
3027 anv_state_pool_finish(&device
->instruction_state_pool
);
3028 anv_state_pool_finish(&device
->dynamic_state_pool
);
3030 anv_bo_pool_finish(&device
->batch_bo_pool
);
3032 anv_bo_cache_finish(&device
->bo_cache
);
3034 if (device
->physical
->use_softpin
) {
3035 util_vma_heap_finish(&device
->vma_hi
);
3036 util_vma_heap_finish(&device
->vma_cva
);
3037 util_vma_heap_finish(&device
->vma_lo
);
3040 pthread_cond_destroy(&device
->queue_submit
);
3041 pthread_mutex_destroy(&device
->mutex
);
3043 anv_gem_destroy_context(device
, device
->context_id
);
3045 if (INTEL_DEBUG
& DEBUG_BATCH
)
3046 gen_batch_decode_ctx_finish(&device
->decoder_ctx
);
3050 vk_free(&device
->alloc
, device
);
3053 VkResult
anv_EnumerateInstanceLayerProperties(
3054 uint32_t* pPropertyCount
,
3055 VkLayerProperties
* pProperties
)
3057 if (pProperties
== NULL
) {
3058 *pPropertyCount
= 0;
3062 /* None supported at this time */
3063 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
3066 VkResult
anv_EnumerateDeviceLayerProperties(
3067 VkPhysicalDevice physicalDevice
,
3068 uint32_t* pPropertyCount
,
3069 VkLayerProperties
* pProperties
)
3071 if (pProperties
== NULL
) {
3072 *pPropertyCount
= 0;
3076 /* None supported at this time */
3077 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
3080 void anv_GetDeviceQueue(
3082 uint32_t queueNodeIndex
,
3083 uint32_t queueIndex
,
3086 const VkDeviceQueueInfo2 info
= {
3087 .sType
= VK_STRUCTURE_TYPE_DEVICE_QUEUE_INFO_2
,
3090 .queueFamilyIndex
= queueNodeIndex
,
3091 .queueIndex
= queueIndex
,
3094 anv_GetDeviceQueue2(_device
, &info
, pQueue
);
3097 void anv_GetDeviceQueue2(
3099 const VkDeviceQueueInfo2
* pQueueInfo
,
3102 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3104 assert(pQueueInfo
->queueIndex
== 0);
3106 if (pQueueInfo
->flags
== device
->queue
.flags
)
3107 *pQueue
= anv_queue_to_handle(&device
->queue
);
3113 _anv_device_set_lost(struct anv_device
*device
,
3114 const char *file
, int line
,
3115 const char *msg
, ...)
3120 p_atomic_inc(&device
->_lost
);
3123 err
= __vk_errorv(device
->physical
->instance
, device
,
3124 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT
,
3125 VK_ERROR_DEVICE_LOST
, file
, line
, msg
, ap
);
3128 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
3135 _anv_queue_set_lost(struct anv_queue
*queue
,
3136 const char *file
, int line
,
3137 const char *msg
, ...)
3142 p_atomic_inc(&queue
->device
->_lost
);
3145 err
= __vk_errorv(queue
->device
->physical
->instance
, queue
->device
,
3146 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT
,
3147 VK_ERROR_DEVICE_LOST
, file
, line
, msg
, ap
);
3150 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
3157 anv_device_query_status(struct anv_device
*device
)
3159 /* This isn't likely as most of the callers of this function already check
3160 * for it. However, it doesn't hurt to check and it potentially lets us
3163 if (anv_device_is_lost(device
))
3164 return VK_ERROR_DEVICE_LOST
;
3166 uint32_t active
, pending
;
3167 int ret
= anv_gem_gpu_get_reset_stats(device
, &active
, &pending
);
3169 /* We don't know the real error. */
3170 return anv_device_set_lost(device
, "get_reset_stats failed: %m");
3174 return anv_device_set_lost(device
, "GPU hung on one of our command buffers");
3175 } else if (pending
) {
3176 return anv_device_set_lost(device
, "GPU hung with commands in-flight");
3183 anv_device_bo_busy(struct anv_device
*device
, struct anv_bo
*bo
)
3185 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
3186 * Other usages of the BO (such as on different hardware) will not be
3187 * flagged as "busy" by this ioctl. Use with care.
3189 int ret
= anv_gem_busy(device
, bo
->gem_handle
);
3191 return VK_NOT_READY
;
3192 } else if (ret
== -1) {
3193 /* We don't know the real error. */
3194 return anv_device_set_lost(device
, "gem wait failed: %m");
3197 /* Query for device status after the busy call. If the BO we're checking
3198 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
3199 * client because it clearly doesn't have valid data. Yes, this most
3200 * likely means an ioctl, but we just did an ioctl to query the busy status
3201 * so it's no great loss.
3203 return anv_device_query_status(device
);
3207 anv_device_wait(struct anv_device
*device
, struct anv_bo
*bo
,
3210 int ret
= anv_gem_wait(device
, bo
->gem_handle
, &timeout
);
3211 if (ret
== -1 && errno
== ETIME
) {
3213 } else if (ret
== -1) {
3214 /* We don't know the real error. */
3215 return anv_device_set_lost(device
, "gem wait failed: %m");
3218 /* Query for device status after the wait. If the BO we're waiting on got
3219 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
3220 * because it clearly doesn't have valid data. Yes, this most likely means
3221 * an ioctl, but we just did an ioctl to wait so it's no great loss.
3223 return anv_device_query_status(device
);
3226 VkResult
anv_DeviceWaitIdle(
3229 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3231 if (anv_device_is_lost(device
))
3232 return VK_ERROR_DEVICE_LOST
;
3234 return anv_queue_submit_simple_batch(&device
->queue
, NULL
);
3238 anv_vma_alloc(struct anv_device
*device
,
3239 uint64_t size
, uint64_t align
,
3240 enum anv_bo_alloc_flags alloc_flags
,
3241 uint64_t client_address
)
3243 pthread_mutex_lock(&device
->vma_mutex
);
3247 if (alloc_flags
& ANV_BO_ALLOC_CLIENT_VISIBLE_ADDRESS
) {
3248 if (client_address
) {
3249 if (util_vma_heap_alloc_addr(&device
->vma_cva
,
3250 client_address
, size
)) {
3251 addr
= client_address
;
3254 addr
= util_vma_heap_alloc(&device
->vma_cva
, size
, align
);
3256 /* We don't want to fall back to other heaps */
3260 assert(client_address
== 0);
3262 if (!(alloc_flags
& ANV_BO_ALLOC_32BIT_ADDRESS
))
3263 addr
= util_vma_heap_alloc(&device
->vma_hi
, size
, align
);
3266 addr
= util_vma_heap_alloc(&device
->vma_lo
, size
, align
);
3269 pthread_mutex_unlock(&device
->vma_mutex
);
3271 assert(addr
== gen_48b_address(addr
));
3272 return gen_canonical_address(addr
);
3276 anv_vma_free(struct anv_device
*device
,
3277 uint64_t address
, uint64_t size
)
3279 const uint64_t addr_48b
= gen_48b_address(address
);
3281 pthread_mutex_lock(&device
->vma_mutex
);
3283 if (addr_48b
>= LOW_HEAP_MIN_ADDRESS
&&
3284 addr_48b
<= LOW_HEAP_MAX_ADDRESS
) {
3285 util_vma_heap_free(&device
->vma_lo
, addr_48b
, size
);
3286 } else if (addr_48b
>= CLIENT_VISIBLE_HEAP_MIN_ADDRESS
&&
3287 addr_48b
<= CLIENT_VISIBLE_HEAP_MAX_ADDRESS
) {
3288 util_vma_heap_free(&device
->vma_cva
, addr_48b
, size
);
3290 assert(addr_48b
>= HIGH_HEAP_MIN_ADDRESS
);
3291 util_vma_heap_free(&device
->vma_hi
, addr_48b
, size
);
3294 pthread_mutex_unlock(&device
->vma_mutex
);
3297 VkResult
anv_AllocateMemory(
3299 const VkMemoryAllocateInfo
* pAllocateInfo
,
3300 const VkAllocationCallbacks
* pAllocator
,
3301 VkDeviceMemory
* pMem
)
3303 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3304 struct anv_physical_device
*pdevice
= device
->physical
;
3305 struct anv_device_memory
*mem
;
3306 VkResult result
= VK_SUCCESS
;
3308 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
3310 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
3311 assert(pAllocateInfo
->allocationSize
> 0);
3313 VkDeviceSize aligned_alloc_size
=
3314 align_u64(pAllocateInfo
->allocationSize
, 4096);
3316 if (aligned_alloc_size
> MAX_MEMORY_ALLOCATION_SIZE
)
3317 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
3319 assert(pAllocateInfo
->memoryTypeIndex
< pdevice
->memory
.type_count
);
3320 struct anv_memory_type
*mem_type
=
3321 &pdevice
->memory
.types
[pAllocateInfo
->memoryTypeIndex
];
3322 assert(mem_type
->heapIndex
< pdevice
->memory
.heap_count
);
3323 struct anv_memory_heap
*mem_heap
=
3324 &pdevice
->memory
.heaps
[mem_type
->heapIndex
];
3326 uint64_t mem_heap_used
= p_atomic_read(&mem_heap
->used
);
3327 if (mem_heap_used
+ aligned_alloc_size
> mem_heap
->size
)
3328 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
3330 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
3331 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3333 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3335 assert(pAllocateInfo
->memoryTypeIndex
< pdevice
->memory
.type_count
);
3336 mem
->type
= mem_type
;
3340 mem
->host_ptr
= NULL
;
3342 enum anv_bo_alloc_flags alloc_flags
= 0;
3344 const VkExportMemoryAllocateInfo
*export_info
= NULL
;
3345 const VkImportAndroidHardwareBufferInfoANDROID
*ahw_import_info
= NULL
;
3346 const VkImportMemoryFdInfoKHR
*fd_info
= NULL
;
3347 const VkImportMemoryHostPointerInfoEXT
*host_ptr_info
= NULL
;
3348 const VkMemoryDedicatedAllocateInfo
*dedicated_info
= NULL
;
3349 VkMemoryAllocateFlags vk_flags
= 0;
3350 uint64_t client_address
= 0;
3352 vk_foreach_struct_const(ext
, pAllocateInfo
->pNext
) {
3353 switch (ext
->sType
) {
3354 case VK_STRUCTURE_TYPE_EXPORT_MEMORY_ALLOCATE_INFO
:
3355 export_info
= (void *)ext
;
3358 case VK_STRUCTURE_TYPE_IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID
:
3359 ahw_import_info
= (void *)ext
;
3362 case VK_STRUCTURE_TYPE_IMPORT_MEMORY_FD_INFO_KHR
:
3363 fd_info
= (void *)ext
;
3366 case VK_STRUCTURE_TYPE_IMPORT_MEMORY_HOST_POINTER_INFO_EXT
:
3367 host_ptr_info
= (void *)ext
;
3370 case VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_FLAGS_INFO
: {
3371 const VkMemoryAllocateFlagsInfo
*flags_info
= (void *)ext
;
3372 vk_flags
= flags_info
->flags
;
3376 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_ALLOCATE_INFO
:
3377 dedicated_info
= (void *)ext
;
3380 case VK_STRUCTURE_TYPE_MEMORY_OPAQUE_CAPTURE_ADDRESS_ALLOCATE_INFO_KHR
: {
3381 const VkMemoryOpaqueCaptureAddressAllocateInfoKHR
*addr_info
=
3382 (const VkMemoryOpaqueCaptureAddressAllocateInfoKHR
*)ext
;
3383 client_address
= addr_info
->opaqueCaptureAddress
;
3388 anv_debug_ignored_stype(ext
->sType
);
3393 /* By default, we want all VkDeviceMemory objects to support CCS */
3394 if (device
->physical
->has_implicit_ccs
)
3395 alloc_flags
|= ANV_BO_ALLOC_IMPLICIT_CCS
;
3397 if (vk_flags
& VK_MEMORY_ALLOCATE_DEVICE_ADDRESS_BIT_KHR
)
3398 alloc_flags
|= ANV_BO_ALLOC_CLIENT_VISIBLE_ADDRESS
;
3400 if ((export_info
&& export_info
->handleTypes
) ||
3401 (fd_info
&& fd_info
->handleType
) ||
3402 (host_ptr_info
&& host_ptr_info
->handleType
)) {
3403 /* Anything imported or exported is EXTERNAL */
3404 alloc_flags
|= ANV_BO_ALLOC_EXTERNAL
;
3406 /* We can't have implicit CCS on external memory with an AUX-table.
3407 * Doing so would require us to sync the aux tables across processes
3408 * which is impractical.
3410 if (device
->info
.has_aux_map
)
3411 alloc_flags
&= ~ANV_BO_ALLOC_IMPLICIT_CCS
;
3414 /* Check if we need to support Android HW buffer export. If so,
3415 * create AHardwareBuffer and import memory from it.
3417 bool android_export
= false;
3418 if (export_info
&& export_info
->handleTypes
&
3419 VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID
)
3420 android_export
= true;
3422 if (ahw_import_info
) {
3423 result
= anv_import_ahw_memory(_device
, mem
, ahw_import_info
);
3424 if (result
!= VK_SUCCESS
)
3428 } else if (android_export
) {
3429 result
= anv_create_ahw_memory(_device
, mem
, pAllocateInfo
);
3430 if (result
!= VK_SUCCESS
)
3433 const VkImportAndroidHardwareBufferInfoANDROID import_info
= {
3436 result
= anv_import_ahw_memory(_device
, mem
, &import_info
);
3437 if (result
!= VK_SUCCESS
)
3443 /* The Vulkan spec permits handleType to be 0, in which case the struct is
3446 if (fd_info
&& fd_info
->handleType
) {
3447 /* At the moment, we support only the below handle types. */
3448 assert(fd_info
->handleType
==
3449 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
3450 fd_info
->handleType
==
3451 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
3453 result
= anv_device_import_bo(device
, fd_info
->fd
, alloc_flags
,
3454 client_address
, &mem
->bo
);
3455 if (result
!= VK_SUCCESS
)
3458 /* For security purposes, we reject importing the bo if it's smaller
3459 * than the requested allocation size. This prevents a malicious client
3460 * from passing a buffer to a trusted client, lying about the size, and
3461 * telling the trusted client to try and texture from an image that goes
3462 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
3463 * in the trusted client. The trusted client can protect itself against
3464 * this sort of attack but only if it can trust the buffer size.
3466 if (mem
->bo
->size
< aligned_alloc_size
) {
3467 result
= vk_errorf(device
, device
, VK_ERROR_INVALID_EXTERNAL_HANDLE
,
3468 "aligned allocationSize too large for "
3469 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT: "
3470 "%"PRIu64
"B > %"PRIu64
"B",
3471 aligned_alloc_size
, mem
->bo
->size
);
3472 anv_device_release_bo(device
, mem
->bo
);
3476 /* From the Vulkan spec:
3478 * "Importing memory from a file descriptor transfers ownership of
3479 * the file descriptor from the application to the Vulkan
3480 * implementation. The application must not perform any operations on
3481 * the file descriptor after a successful import."
3483 * If the import fails, we leave the file descriptor open.
3489 if (host_ptr_info
&& host_ptr_info
->handleType
) {
3490 if (host_ptr_info
->handleType
==
3491 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_MAPPED_FOREIGN_MEMORY_BIT_EXT
) {
3492 result
= vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
3496 assert(host_ptr_info
->handleType
==
3497 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
);
3499 result
= anv_device_import_bo_from_host_ptr(device
,
3500 host_ptr_info
->pHostPointer
,
3501 pAllocateInfo
->allocationSize
,
3505 if (result
!= VK_SUCCESS
)
3508 mem
->host_ptr
= host_ptr_info
->pHostPointer
;
3512 /* Regular allocate (not importing memory). */
3514 result
= anv_device_alloc_bo(device
, pAllocateInfo
->allocationSize
,
3515 alloc_flags
, client_address
, &mem
->bo
);
3516 if (result
!= VK_SUCCESS
)
3519 if (dedicated_info
&& dedicated_info
->image
!= VK_NULL_HANDLE
) {
3520 ANV_FROM_HANDLE(anv_image
, image
, dedicated_info
->image
);
3522 /* Some legacy (non-modifiers) consumers need the tiling to be set on
3523 * the BO. In this case, we have a dedicated allocation.
3525 if (image
->needs_set_tiling
) {
3526 const uint32_t i915_tiling
=
3527 isl_tiling_to_i915_tiling(image
->planes
[0].surface
.isl
.tiling
);
3528 int ret
= anv_gem_set_tiling(device
, mem
->bo
->gem_handle
,
3529 image
->planes
[0].surface
.isl
.row_pitch_B
,
3532 anv_device_release_bo(device
, mem
->bo
);
3533 result
= vk_errorf(device
, device
, VK_ERROR_OUT_OF_DEVICE_MEMORY
,
3534 "failed to set BO tiling: %m");
3541 mem_heap_used
= p_atomic_add_return(&mem_heap
->used
, mem
->bo
->size
);
3542 if (mem_heap_used
> mem_heap
->size
) {
3543 p_atomic_add(&mem_heap
->used
, -mem
->bo
->size
);
3544 anv_device_release_bo(device
, mem
->bo
);
3545 result
= vk_errorf(device
, device
, VK_ERROR_OUT_OF_DEVICE_MEMORY
,
3546 "Out of heap memory");
3550 pthread_mutex_lock(&device
->mutex
);
3551 list_addtail(&mem
->link
, &device
->memory_objects
);
3552 pthread_mutex_unlock(&device
->mutex
);
3554 *pMem
= anv_device_memory_to_handle(mem
);
3559 vk_free2(&device
->alloc
, pAllocator
, mem
);
3564 VkResult
anv_GetMemoryFdKHR(
3566 const VkMemoryGetFdInfoKHR
* pGetFdInfo
,
3569 ANV_FROM_HANDLE(anv_device
, dev
, device_h
);
3570 ANV_FROM_HANDLE(anv_device_memory
, mem
, pGetFdInfo
->memory
);
3572 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
3574 assert(pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
3575 pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
3577 return anv_device_export_bo(dev
, mem
->bo
, pFd
);
3580 VkResult
anv_GetMemoryFdPropertiesKHR(
3582 VkExternalMemoryHandleTypeFlagBits handleType
,
3584 VkMemoryFdPropertiesKHR
* pMemoryFdProperties
)
3586 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3588 switch (handleType
) {
3589 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
:
3590 /* dma-buf can be imported as any memory type */
3591 pMemoryFdProperties
->memoryTypeBits
=
3592 (1 << device
->physical
->memory
.type_count
) - 1;
3596 /* The valid usage section for this function says:
3598 * "handleType must not be one of the handle types defined as
3601 * So opaque handle types fall into the default "unsupported" case.
3603 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
3607 VkResult
anv_GetMemoryHostPointerPropertiesEXT(
3609 VkExternalMemoryHandleTypeFlagBits handleType
,
3610 const void* pHostPointer
,
3611 VkMemoryHostPointerPropertiesEXT
* pMemoryHostPointerProperties
)
3613 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3615 assert(pMemoryHostPointerProperties
->sType
==
3616 VK_STRUCTURE_TYPE_MEMORY_HOST_POINTER_PROPERTIES_EXT
);
3618 switch (handleType
) {
3619 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
:
3620 /* Host memory can be imported as any memory type. */
3621 pMemoryHostPointerProperties
->memoryTypeBits
=
3622 (1ull << device
->physical
->memory
.type_count
) - 1;
3627 return VK_ERROR_INVALID_EXTERNAL_HANDLE
;
3631 void anv_FreeMemory(
3633 VkDeviceMemory _mem
,
3634 const VkAllocationCallbacks
* pAllocator
)
3636 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3637 ANV_FROM_HANDLE(anv_device_memory
, mem
, _mem
);
3642 pthread_mutex_lock(&device
->mutex
);
3643 list_del(&mem
->link
);
3644 pthread_mutex_unlock(&device
->mutex
);
3647 anv_UnmapMemory(_device
, _mem
);
3649 p_atomic_add(&device
->physical
->memory
.heaps
[mem
->type
->heapIndex
].used
,
3652 anv_device_release_bo(device
, mem
->bo
);
3654 #if defined(ANDROID) && ANDROID_API_LEVEL >= 26
3656 AHardwareBuffer_release(mem
->ahw
);
3659 vk_free2(&device
->alloc
, pAllocator
, mem
);
3662 VkResult
anv_MapMemory(
3664 VkDeviceMemory _memory
,
3665 VkDeviceSize offset
,
3667 VkMemoryMapFlags flags
,
3670 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3671 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
3678 if (mem
->host_ptr
) {
3679 *ppData
= mem
->host_ptr
+ offset
;
3683 if (size
== VK_WHOLE_SIZE
)
3684 size
= mem
->bo
->size
- offset
;
3686 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
3688 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
3689 * assert(size != 0);
3690 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
3691 * equal to the size of the memory minus offset
3694 assert(offset
+ size
<= mem
->bo
->size
);
3696 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
3697 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
3698 * at a time is valid. We could just mmap up front and return an offset
3699 * pointer here, but that may exhaust virtual memory on 32 bit
3702 uint32_t gem_flags
= 0;
3704 if (!device
->info
.has_llc
&&
3705 (mem
->type
->propertyFlags
& VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
))
3706 gem_flags
|= I915_MMAP_WC
;
3708 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
3709 uint64_t map_offset
;
3710 if (!device
->physical
->has_mmap_offset
)
3711 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
, device
, _device
);
3738 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
3740 if (mem
== NULL
|| mem
->host_ptr
)
3743 anv_gem_munmap(device
, mem
->map
, mem
->map_size
);
3750 clflush_mapped_ranges(struct anv_device
*device
,
3752 const VkMappedMemoryRange
*ranges
)
3754 for (uint32_t i
= 0; i
< count
; i
++) {
3755 ANV_FROM_HANDLE(anv_device_memory
, mem
, ranges
[i
].memory
);
3756 if (ranges
[i
].offset
>= mem
->map_size
)
3759 gen_clflush_range(mem
->map
+ ranges
[i
].offset
,
3760 MIN2(ranges
[i
].size
, mem
->map_size
- ranges
[i
].offset
));
3764 VkResult
anv_FlushMappedMemoryRanges(
3766 uint32_t memoryRangeCount
,
3767 const VkMappedMemoryRange
* pMemoryRanges
)
3769 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3771 if (device
->info
.has_llc
)
3774 /* Make sure the writes we're flushing have landed. */
3775 __builtin_ia32_mfence();
3777 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
3782 VkResult
anv_InvalidateMappedMemoryRanges(
3784 uint32_t memoryRangeCount
,
3785 const VkMappedMemoryRange
* pMemoryRanges
)
3787 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3789 if (device
->info
.has_llc
)
3792 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
3794 /* Make sure no reads get moved up above the invalidate. */
3795 __builtin_ia32_mfence();
3800 void anv_GetBufferMemoryRequirements(
3803 VkMemoryRequirements
* pMemoryRequirements
)
3805 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
3806 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3808 /* The Vulkan spec (git aaed022) says:
3810 * memoryTypeBits is a bitfield and contains one bit set for every
3811 * supported memory type for the resource. The bit `1<<i` is set if and
3812 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3813 * structure for the physical device is supported.
3815 uint32_t memory_types
= (1ull << device
->physical
->memory
.type_count
) - 1;
3817 /* Base alignment requirement of a cache line */
3818 uint32_t alignment
= 16;
3820 /* We need an alignment of 32 for pushing UBOs */
3821 if (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
)
3822 alignment
= MAX2(alignment
, 32);
3824 pMemoryRequirements
->size
= buffer
->size
;
3825 pMemoryRequirements
->alignment
= alignment
;
3827 /* Storage and Uniform buffers should have their size aligned to
3828 * 32-bits to avoid boundary checks when last DWord is not complete.
3829 * This would ensure that not internal padding would be needed for
3832 if (device
->robust_buffer_access
&&
3833 (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
||
3834 buffer
->usage
& VK_BUFFER_USAGE_STORAGE_BUFFER_BIT
))
3835 pMemoryRequirements
->size
= align_u64(buffer
->size
, 4);
3837 pMemoryRequirements
->memoryTypeBits
= memory_types
;
3840 void anv_GetBufferMemoryRequirements2(
3842 const VkBufferMemoryRequirementsInfo2
* pInfo
,
3843 VkMemoryRequirements2
* pMemoryRequirements
)
3845 anv_GetBufferMemoryRequirements(_device
, pInfo
->buffer
,
3846 &pMemoryRequirements
->memoryRequirements
);
3848 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3849 switch (ext
->sType
) {
3850 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
3851 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
3852 requirements
->prefersDedicatedAllocation
= false;
3853 requirements
->requiresDedicatedAllocation
= false;
3858 anv_debug_ignored_stype(ext
->sType
);
3864 void anv_GetImageMemoryRequirements(
3867 VkMemoryRequirements
* pMemoryRequirements
)
3869 ANV_FROM_HANDLE(anv_image
, image
, _image
);
3870 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3872 /* The Vulkan spec (git aaed022) says:
3874 * memoryTypeBits is a bitfield and contains one bit set for every
3875 * supported memory type for the resource. The bit `1<<i` is set if and
3876 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3877 * structure for the physical device is supported.
3879 * All types are currently supported for images.
3881 uint32_t memory_types
= (1ull << device
->physical
->memory
.type_count
) - 1;
3883 /* We must have image allocated or imported at this point. According to the
3884 * specification, external images must have been bound to memory before
3885 * calling GetImageMemoryRequirements.
3887 assert(image
->size
> 0);
3889 pMemoryRequirements
->size
= image
->size
;
3890 pMemoryRequirements
->alignment
= image
->alignment
;
3891 pMemoryRequirements
->memoryTypeBits
= memory_types
;
3894 void anv_GetImageMemoryRequirements2(
3896 const VkImageMemoryRequirementsInfo2
* pInfo
,
3897 VkMemoryRequirements2
* pMemoryRequirements
)
3899 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3900 ANV_FROM_HANDLE(anv_image
, image
, pInfo
->image
);
3902 anv_GetImageMemoryRequirements(_device
, pInfo
->image
,
3903 &pMemoryRequirements
->memoryRequirements
);
3905 vk_foreach_struct_const(ext
, pInfo
->pNext
) {
3906 switch (ext
->sType
) {
3907 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO
: {
3908 const VkImagePlaneMemoryRequirementsInfo
*plane_reqs
=
3909 (const VkImagePlaneMemoryRequirementsInfo
*) ext
;
3910 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
,
3911 plane_reqs
->planeAspect
);
3913 assert(image
->planes
[plane
].offset
== 0);
3915 /* The Vulkan spec (git aaed022) says:
3917 * memoryTypeBits is a bitfield and contains one bit set for every
3918 * supported memory type for the resource. The bit `1<<i` is set
3919 * if and only if the memory type `i` in the
3920 * VkPhysicalDeviceMemoryProperties structure for the physical
3921 * device is supported.
3923 * All types are currently supported for images.
3925 pMemoryRequirements
->memoryRequirements
.memoryTypeBits
=
3926 (1ull << device
->physical
->memory
.type_count
) - 1;
3928 /* We must have image allocated or imported at this point. According to the
3929 * specification, external images must have been bound to memory before
3930 * calling GetImageMemoryRequirements.
3932 assert(image
->planes
[plane
].size
> 0);
3934 pMemoryRequirements
->memoryRequirements
.size
= image
->planes
[plane
].size
;
3935 pMemoryRequirements
->memoryRequirements
.alignment
=
3936 image
->planes
[plane
].alignment
;
3941 anv_debug_ignored_stype(ext
->sType
);
3946 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3947 switch (ext
->sType
) {
3948 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
3949 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
3950 if (image
->needs_set_tiling
|| image
->external_format
) {
3951 /* If we need to set the tiling for external consumers, we need a
3952 * dedicated allocation.
3954 * See also anv_AllocateMemory.
3956 requirements
->prefersDedicatedAllocation
= true;
3957 requirements
->requiresDedicatedAllocation
= true;
3959 requirements
->prefersDedicatedAllocation
= false;
3960 requirements
->requiresDedicatedAllocation
= false;
3966 anv_debug_ignored_stype(ext
->sType
);
3972 void anv_GetImageSparseMemoryRequirements(
3975 uint32_t* pSparseMemoryRequirementCount
,
3976 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
3978 *pSparseMemoryRequirementCount
= 0;
3981 void anv_GetImageSparseMemoryRequirements2(
3983 const VkImageSparseMemoryRequirementsInfo2
* pInfo
,
3984 uint32_t* pSparseMemoryRequirementCount
,
3985 VkSparseImageMemoryRequirements2
* pSparseMemoryRequirements
)
3987 *pSparseMemoryRequirementCount
= 0;
3990 void anv_GetDeviceMemoryCommitment(
3992 VkDeviceMemory memory
,
3993 VkDeviceSize
* pCommittedMemoryInBytes
)
3995 *pCommittedMemoryInBytes
= 0;
3999 anv_bind_buffer_memory(const VkBindBufferMemoryInfo
*pBindInfo
)
4001 ANV_FROM_HANDLE(anv_device_memory
, mem
, pBindInfo
->memory
);
4002 ANV_FROM_HANDLE(anv_buffer
, buffer
, pBindInfo
->buffer
);
4004 assert(pBindInfo
->sType
== VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
);
4007 buffer
->address
= (struct anv_address
) {
4009 .offset
= pBindInfo
->memoryOffset
,
4012 buffer
->address
= ANV_NULL_ADDRESS
;
4016 VkResult
anv_BindBufferMemory(
4019 VkDeviceMemory memory
,
4020 VkDeviceSize memoryOffset
)
4022 anv_bind_buffer_memory(
4023 &(VkBindBufferMemoryInfo
) {
4024 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
,
4027 .memoryOffset
= memoryOffset
,
4033 VkResult
anv_BindBufferMemory2(
4035 uint32_t bindInfoCount
,
4036 const VkBindBufferMemoryInfo
* pBindInfos
)
4038 for (uint32_t i
= 0; i
< bindInfoCount
; i
++)
4039 anv_bind_buffer_memory(&pBindInfos
[i
]);
4044 VkResult
anv_QueueBindSparse(
4046 uint32_t bindInfoCount
,
4047 const VkBindSparseInfo
* pBindInfo
,
4050 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
4051 if (anv_device_is_lost(queue
->device
))
4052 return VK_ERROR_DEVICE_LOST
;
4054 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
4059 VkResult
anv_CreateEvent(
4061 const VkEventCreateInfo
* pCreateInfo
,
4062 const VkAllocationCallbacks
* pAllocator
,
4065 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4066 struct anv_state state
;
4067 struct anv_event
*event
;
4069 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_EVENT_CREATE_INFO
);
4071 state
= anv_state_pool_alloc(&device
->dynamic_state_pool
,
4074 event
->state
= state
;
4075 event
->semaphore
= VK_EVENT_RESET
;
4077 if (!device
->info
.has_llc
) {
4078 /* Make sure the writes we're flushing have landed. */
4079 __builtin_ia32_mfence();
4080 __builtin_ia32_clflush(event
);
4083 *pEvent
= anv_event_to_handle(event
);
4088 void anv_DestroyEvent(
4091 const VkAllocationCallbacks
* pAllocator
)
4093 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4094 ANV_FROM_HANDLE(anv_event
, event
, _event
);
4099 anv_state_pool_free(&device
->dynamic_state_pool
, event
->state
);
4102 VkResult
anv_GetEventStatus(
4106 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4107 ANV_FROM_HANDLE(anv_event
, event
, _event
);
4109 if (anv_device_is_lost(device
))
4110 return VK_ERROR_DEVICE_LOST
;
4112 if (!device
->info
.has_llc
) {
4113 /* Invalidate read cache before reading event written by GPU. */
4114 __builtin_ia32_clflush(event
);
4115 __builtin_ia32_mfence();
4119 return event
->semaphore
;
4122 VkResult
anv_SetEvent(
4126 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4127 ANV_FROM_HANDLE(anv_event
, event
, _event
);
4129 event
->semaphore
= VK_EVENT_SET
;
4131 if (!device
->info
.has_llc
) {
4132 /* Make sure the writes we're flushing have landed. */
4133 __builtin_ia32_mfence();
4134 __builtin_ia32_clflush(event
);
4140 VkResult
anv_ResetEvent(
4144 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4145 ANV_FROM_HANDLE(anv_event
, event
, _event
);
4147 event
->semaphore
= VK_EVENT_RESET
;
4149 if (!device
->info
.has_llc
) {
4150 /* Make sure the writes we're flushing have landed. */
4151 __builtin_ia32_mfence();
4152 __builtin_ia32_clflush(event
);
4160 VkResult
anv_CreateBuffer(
4162 const VkBufferCreateInfo
* pCreateInfo
,
4163 const VkAllocationCallbacks
* pAllocator
,
4166 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4167 struct anv_buffer
*buffer
;
4169 /* Don't allow creating buffers bigger than our address space. The real
4170 * issue here is that we may align up the buffer size and we don't want
4171 * doing so to cause roll-over. However, no one has any business
4172 * allocating a buffer larger than our GTT size.
4174 if (pCreateInfo
->size
> device
->physical
->gtt_size
)
4175 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
4177 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
4179 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
4180 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
4182 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
4184 buffer
->size
= pCreateInfo
->size
;
4185 buffer
->usage
= pCreateInfo
->usage
;
4186 buffer
->address
= ANV_NULL_ADDRESS
;
4188 *pBuffer
= anv_buffer_to_handle(buffer
);
4193 void anv_DestroyBuffer(
4196 const VkAllocationCallbacks
* pAllocator
)
4198 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4199 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
4204 vk_free2(&device
->alloc
, pAllocator
, buffer
);
4207 VkDeviceAddress
anv_GetBufferDeviceAddress(
4209 const VkBufferDeviceAddressInfoKHR
* pInfo
)
4211 ANV_FROM_HANDLE(anv_buffer
, buffer
, pInfo
->buffer
);
4213 assert(!anv_address_is_null(buffer
->address
));
4214 assert(buffer
->address
.bo
->flags
& EXEC_OBJECT_PINNED
);
4216 return anv_address_physical(buffer
->address
);
4219 uint64_t anv_GetBufferOpaqueCaptureAddress(
4221 const VkBufferDeviceAddressInfoKHR
* pInfo
)
4226 uint64_t anv_GetDeviceMemoryOpaqueCaptureAddress(
4228 const VkDeviceMemoryOpaqueCaptureAddressInfoKHR
* pInfo
)
4230 ANV_FROM_HANDLE(anv_device_memory
, memory
, pInfo
->memory
);
4232 assert(memory
->bo
->flags
& EXEC_OBJECT_PINNED
);
4233 assert(memory
->bo
->has_client_visible_address
);
4235 return gen_48b_address(memory
->bo
->offset
);
4239 anv_fill_buffer_surface_state(struct anv_device
*device
, struct anv_state state
,
4240 enum isl_format format
,
4241 struct anv_address address
,
4242 uint32_t range
, uint32_t stride
)
4244 isl_buffer_fill_state(&device
->isl_dev
, state
.map
,
4245 .address
= anv_address_physical(address
),
4246 .mocs
= device
->isl_dev
.mocs
.internal
,
4249 .swizzle
= ISL_SWIZZLE_IDENTITY
,
4250 .stride_B
= stride
);
4253 void anv_DestroySampler(
4256 const VkAllocationCallbacks
* pAllocator
)
4258 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4259 ANV_FROM_HANDLE(anv_sampler
, sampler
, _sampler
);
4264 if (sampler
->bindless_state
.map
) {
4265 anv_state_pool_free(&device
->dynamic_state_pool
,
4266 sampler
->bindless_state
);
4269 vk_free2(&device
->alloc
, pAllocator
, sampler
);
4272 VkResult
anv_CreateFramebuffer(
4274 const VkFramebufferCreateInfo
* pCreateInfo
,
4275 const VkAllocationCallbacks
* pAllocator
,
4276 VkFramebuffer
* pFramebuffer
)
4278 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4279 struct anv_framebuffer
*framebuffer
;
4281 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
4283 size_t size
= sizeof(*framebuffer
);
4285 /* VK_KHR_imageless_framebuffer extension says:
4287 * If flags includes VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR,
4288 * parameter pAttachments is ignored.
4290 if (!(pCreateInfo
->flags
& VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR
)) {
4291 size
+= sizeof(struct anv_image_view
*) * pCreateInfo
->attachmentCount
;
4292 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
4293 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
4294 if (framebuffer
== NULL
)
4295 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
4297 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
4298 ANV_FROM_HANDLE(anv_image_view
, iview
, pCreateInfo
->pAttachments
[i
]);
4299 framebuffer
->attachments
[i
] = iview
;
4301 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
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);