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_ROBUSTNESS_2_FEATURES_EXT
: {
1245 VkPhysicalDeviceRobustness2FeaturesEXT
*features
= (void *)ext
;
1246 features
->robustBufferAccess2
= true;
1247 features
->robustImageAccess2
= true;
1248 features
->nullDescriptor
= true;
1252 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES
: {
1253 VkPhysicalDeviceSamplerYcbcrConversionFeatures
*features
=
1254 (VkPhysicalDeviceSamplerYcbcrConversionFeatures
*) ext
;
1255 CORE_FEATURE(1, 1, samplerYcbcrConversion
);
1259 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES_EXT
: {
1260 VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*features
=
1261 (VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*)ext
;
1262 CORE_FEATURE(1, 2, scalarBlockLayout
);
1266 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SEPARATE_DEPTH_STENCIL_LAYOUTS_FEATURES_KHR
: {
1267 VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR
*features
=
1268 (VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR
*)ext
;
1269 CORE_FEATURE(1, 2, separateDepthStencilLayouts
);
1273 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_INT64_FEATURES_KHR
: {
1274 VkPhysicalDeviceShaderAtomicInt64FeaturesKHR
*features
= (void *)ext
;
1275 CORE_FEATURE(1, 2, shaderBufferInt64Atomics
);
1276 CORE_FEATURE(1, 2, shaderSharedInt64Atomics
);
1280 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DEMOTE_TO_HELPER_INVOCATION_FEATURES_EXT
: {
1281 VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT
*features
= (void *)ext
;
1282 features
->shaderDemoteToHelperInvocation
= true;
1286 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CLOCK_FEATURES_KHR
: {
1287 VkPhysicalDeviceShaderClockFeaturesKHR
*features
=
1288 (VkPhysicalDeviceShaderClockFeaturesKHR
*)ext
;
1289 features
->shaderSubgroupClock
= true;
1290 features
->shaderDeviceClock
= false;
1294 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETERS_FEATURES
: {
1295 VkPhysicalDeviceShaderDrawParametersFeatures
*features
= (void *)ext
;
1296 CORE_FEATURE(1, 1, shaderDrawParameters
);
1300 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_SUBGROUP_EXTENDED_TYPES_FEATURES_KHR
: {
1301 VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR
*features
=
1302 (VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR
*)ext
;
1303 CORE_FEATURE(1, 2, shaderSubgroupExtendedTypes
);
1307 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_FEATURES_EXT
: {
1308 VkPhysicalDeviceSubgroupSizeControlFeaturesEXT
*features
=
1309 (VkPhysicalDeviceSubgroupSizeControlFeaturesEXT
*)ext
;
1310 features
->subgroupSizeControl
= true;
1311 features
->computeFullSubgroups
= true;
1315 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_FEATURES_EXT
: {
1316 VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT
*features
=
1317 (VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT
*)ext
;
1318 features
->texelBufferAlignment
= true;
1322 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_FEATURES_KHR
: {
1323 VkPhysicalDeviceTimelineSemaphoreFeaturesKHR
*features
=
1324 (VkPhysicalDeviceTimelineSemaphoreFeaturesKHR
*) ext
;
1325 CORE_FEATURE(1, 2, timelineSemaphore
);
1329 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES
: {
1330 VkPhysicalDeviceVariablePointersFeatures
*features
= (void *)ext
;
1331 CORE_FEATURE(1, 1, variablePointersStorageBuffer
);
1332 CORE_FEATURE(1, 1, variablePointers
);
1336 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT
: {
1337 VkPhysicalDeviceTransformFeedbackFeaturesEXT
*features
=
1338 (VkPhysicalDeviceTransformFeedbackFeaturesEXT
*)ext
;
1339 features
->transformFeedback
= true;
1340 features
->geometryStreams
= true;
1344 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_UNIFORM_BUFFER_STANDARD_LAYOUT_FEATURES_KHR
: {
1345 VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR
*features
=
1346 (VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR
*)ext
;
1347 CORE_FEATURE(1, 2, uniformBufferStandardLayout
);
1351 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT
: {
1352 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*features
=
1353 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*)ext
;
1354 features
->vertexAttributeInstanceRateDivisor
= true;
1355 features
->vertexAttributeInstanceRateZeroDivisor
= true;
1359 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES
:
1360 anv_get_physical_device_features_1_1(pdevice
, (void *)ext
);
1363 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES
:
1364 anv_get_physical_device_features_1_2(pdevice
, (void *)ext
);
1367 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_MEMORY_MODEL_FEATURES_KHR
: {
1368 VkPhysicalDeviceVulkanMemoryModelFeaturesKHR
*features
= (void *)ext
;
1369 CORE_FEATURE(1, 2, vulkanMemoryModel
);
1370 CORE_FEATURE(1, 2, vulkanMemoryModelDeviceScope
);
1371 CORE_FEATURE(1, 2, vulkanMemoryModelAvailabilityVisibilityChains
);
1375 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_YCBCR_IMAGE_ARRAYS_FEATURES_EXT
: {
1376 VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*features
=
1377 (VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*)ext
;
1378 features
->ycbcrImageArrays
= true;
1383 anv_debug_ignored_stype(ext
->sType
);
1391 #define MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS 64
1393 #define MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS 64
1394 #define MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS 256
1396 void anv_GetPhysicalDeviceProperties(
1397 VkPhysicalDevice physicalDevice
,
1398 VkPhysicalDeviceProperties
* pProperties
)
1400 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1401 const struct gen_device_info
*devinfo
= &pdevice
->info
;
1403 /* See assertions made when programming the buffer surface state. */
1404 const uint32_t max_raw_buffer_sz
= devinfo
->gen
>= 7 ?
1405 (1ul << 30) : (1ul << 27);
1407 const uint32_t max_ssbos
= pdevice
->has_a64_buffer_access
? UINT16_MAX
: 64;
1408 const uint32_t max_textures
=
1409 pdevice
->has_bindless_images
? UINT16_MAX
: 128;
1410 const uint32_t max_samplers
=
1411 pdevice
->has_bindless_samplers
? UINT16_MAX
:
1412 (devinfo
->gen
>= 8 || devinfo
->is_haswell
) ? 128 : 16;
1413 const uint32_t max_images
=
1414 pdevice
->has_bindless_images
? UINT16_MAX
: MAX_IMAGES
;
1416 /* If we can use bindless for everything, claim a high per-stage limit,
1417 * otherwise use the binding table size, minus the slots reserved for
1418 * render targets and one slot for the descriptor buffer. */
1419 const uint32_t max_per_stage
=
1420 pdevice
->has_bindless_images
&& pdevice
->has_a64_buffer_access
1421 ? UINT32_MAX
: MAX_BINDING_TABLE_SIZE
- MAX_RTS
- 1;
1423 /* Limit max_threads to 64 for the GPGPU_WALKER command */
1424 const uint32_t max_workgroup_size
= 32 * MIN2(64, devinfo
->max_cs_threads
);
1426 VkSampleCountFlags sample_counts
=
1427 isl_device_get_sample_counts(&pdevice
->isl_dev
);
1430 VkPhysicalDeviceLimits limits
= {
1431 .maxImageDimension1D
= (1 << 14),
1432 .maxImageDimension2D
= (1 << 14),
1433 .maxImageDimension3D
= (1 << 11),
1434 .maxImageDimensionCube
= (1 << 14),
1435 .maxImageArrayLayers
= (1 << 11),
1436 .maxTexelBufferElements
= 128 * 1024 * 1024,
1437 .maxUniformBufferRange
= (1ul << 27),
1438 .maxStorageBufferRange
= max_raw_buffer_sz
,
1439 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
1440 .maxMemoryAllocationCount
= UINT32_MAX
,
1441 .maxSamplerAllocationCount
= 64 * 1024,
1442 .bufferImageGranularity
= 64, /* A cache line */
1443 .sparseAddressSpaceSize
= 0,
1444 .maxBoundDescriptorSets
= MAX_SETS
,
1445 .maxPerStageDescriptorSamplers
= max_samplers
,
1446 .maxPerStageDescriptorUniformBuffers
= MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
,
1447 .maxPerStageDescriptorStorageBuffers
= max_ssbos
,
1448 .maxPerStageDescriptorSampledImages
= max_textures
,
1449 .maxPerStageDescriptorStorageImages
= max_images
,
1450 .maxPerStageDescriptorInputAttachments
= MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS
,
1451 .maxPerStageResources
= max_per_stage
,
1452 .maxDescriptorSetSamplers
= 6 * max_samplers
, /* number of stages * maxPerStageDescriptorSamplers */
1453 .maxDescriptorSetUniformBuffers
= 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
, /* number of stages * maxPerStageDescriptorUniformBuffers */
1454 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
1455 .maxDescriptorSetStorageBuffers
= 6 * max_ssbos
, /* number of stages * maxPerStageDescriptorStorageBuffers */
1456 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
1457 .maxDescriptorSetSampledImages
= 6 * max_textures
, /* number of stages * maxPerStageDescriptorSampledImages */
1458 .maxDescriptorSetStorageImages
= 6 * max_images
, /* number of stages * maxPerStageDescriptorStorageImages */
1459 .maxDescriptorSetInputAttachments
= MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS
,
1460 .maxVertexInputAttributes
= MAX_VBS
,
1461 .maxVertexInputBindings
= MAX_VBS
,
1462 .maxVertexInputAttributeOffset
= 2047,
1463 .maxVertexInputBindingStride
= 2048,
1464 .maxVertexOutputComponents
= 128,
1465 .maxTessellationGenerationLevel
= 64,
1466 .maxTessellationPatchSize
= 32,
1467 .maxTessellationControlPerVertexInputComponents
= 128,
1468 .maxTessellationControlPerVertexOutputComponents
= 128,
1469 .maxTessellationControlPerPatchOutputComponents
= 128,
1470 .maxTessellationControlTotalOutputComponents
= 2048,
1471 .maxTessellationEvaluationInputComponents
= 128,
1472 .maxTessellationEvaluationOutputComponents
= 128,
1473 .maxGeometryShaderInvocations
= 32,
1474 .maxGeometryInputComponents
= 64,
1475 .maxGeometryOutputComponents
= 128,
1476 .maxGeometryOutputVertices
= 256,
1477 .maxGeometryTotalOutputComponents
= 1024,
1478 .maxFragmentInputComponents
= 116, /* 128 components - (PSIZ, CLIP_DIST0, CLIP_DIST1) */
1479 .maxFragmentOutputAttachments
= 8,
1480 .maxFragmentDualSrcAttachments
= 1,
1481 .maxFragmentCombinedOutputResources
= 8,
1482 .maxComputeSharedMemorySize
= 64 * 1024,
1483 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
1484 .maxComputeWorkGroupInvocations
= max_workgroup_size
,
1485 .maxComputeWorkGroupSize
= {
1490 .subPixelPrecisionBits
= 8,
1491 .subTexelPrecisionBits
= 8,
1492 .mipmapPrecisionBits
= 8,
1493 .maxDrawIndexedIndexValue
= UINT32_MAX
,
1494 .maxDrawIndirectCount
= UINT32_MAX
,
1495 .maxSamplerLodBias
= 16,
1496 .maxSamplerAnisotropy
= 16,
1497 .maxViewports
= MAX_VIEWPORTS
,
1498 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
1499 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
1500 .viewportSubPixelBits
= 13, /* We take a float? */
1501 .minMemoryMapAlignment
= 4096, /* A page */
1502 /* The dataport requires texel alignment so we need to assume a worst
1503 * case of R32G32B32A32 which is 16 bytes.
1505 .minTexelBufferOffsetAlignment
= 16,
1506 /* We need 16 for UBO block reads to work and 32 for push UBOs */
1507 .minUniformBufferOffsetAlignment
= 32,
1508 .minStorageBufferOffsetAlignment
= 4,
1509 .minTexelOffset
= -8,
1510 .maxTexelOffset
= 7,
1511 .minTexelGatherOffset
= -32,
1512 .maxTexelGatherOffset
= 31,
1513 .minInterpolationOffset
= -0.5,
1514 .maxInterpolationOffset
= 0.4375,
1515 .subPixelInterpolationOffsetBits
= 4,
1516 .maxFramebufferWidth
= (1 << 14),
1517 .maxFramebufferHeight
= (1 << 14),
1518 .maxFramebufferLayers
= (1 << 11),
1519 .framebufferColorSampleCounts
= sample_counts
,
1520 .framebufferDepthSampleCounts
= sample_counts
,
1521 .framebufferStencilSampleCounts
= sample_counts
,
1522 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
1523 .maxColorAttachments
= MAX_RTS
,
1524 .sampledImageColorSampleCounts
= sample_counts
,
1525 .sampledImageIntegerSampleCounts
= sample_counts
,
1526 .sampledImageDepthSampleCounts
= sample_counts
,
1527 .sampledImageStencilSampleCounts
= sample_counts
,
1528 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
1529 .maxSampleMaskWords
= 1,
1530 .timestampComputeAndGraphics
= true,
1531 .timestampPeriod
= 1000000000.0 / devinfo
->timestamp_frequency
,
1532 .maxClipDistances
= 8,
1533 .maxCullDistances
= 8,
1534 .maxCombinedClipAndCullDistances
= 8,
1535 .discreteQueuePriorities
= 2,
1536 .pointSizeRange
= { 0.125, 255.875 },
1539 (devinfo
->gen
>= 9 || devinfo
->is_cherryview
) ?
1540 2047.9921875 : 7.9921875,
1542 .pointSizeGranularity
= (1.0 / 8.0),
1543 .lineWidthGranularity
= (1.0 / 128.0),
1544 .strictLines
= false,
1545 .standardSampleLocations
= true,
1546 .optimalBufferCopyOffsetAlignment
= 128,
1547 .optimalBufferCopyRowPitchAlignment
= 128,
1548 .nonCoherentAtomSize
= 64,
1551 *pProperties
= (VkPhysicalDeviceProperties
) {
1552 .apiVersion
= anv_physical_device_api_version(pdevice
),
1553 .driverVersion
= vk_get_driver_version(),
1555 .deviceID
= pdevice
->info
.chipset_id
,
1556 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
1558 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
1561 snprintf(pProperties
->deviceName
, sizeof(pProperties
->deviceName
),
1562 "%s", pdevice
->name
);
1563 memcpy(pProperties
->pipelineCacheUUID
,
1564 pdevice
->pipeline_cache_uuid
, VK_UUID_SIZE
);
1568 anv_get_physical_device_properties_1_1(struct anv_physical_device
*pdevice
,
1569 VkPhysicalDeviceVulkan11Properties
*p
)
1571 assert(p
->sType
== VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES
);
1573 memcpy(p
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
1574 memcpy(p
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
1575 memset(p
->deviceLUID
, 0, VK_LUID_SIZE
);
1576 p
->deviceNodeMask
= 0;
1577 p
->deviceLUIDValid
= false;
1579 p
->subgroupSize
= BRW_SUBGROUP_SIZE
;
1580 VkShaderStageFlags scalar_stages
= 0;
1581 for (unsigned stage
= 0; stage
< MESA_SHADER_STAGES
; stage
++) {
1582 if (pdevice
->compiler
->scalar_stage
[stage
])
1583 scalar_stages
|= mesa_to_vk_shader_stage(stage
);
1585 p
->subgroupSupportedStages
= scalar_stages
;
1586 p
->subgroupSupportedOperations
= VK_SUBGROUP_FEATURE_BASIC_BIT
|
1587 VK_SUBGROUP_FEATURE_VOTE_BIT
|
1588 VK_SUBGROUP_FEATURE_BALLOT_BIT
|
1589 VK_SUBGROUP_FEATURE_SHUFFLE_BIT
|
1590 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT
|
1591 VK_SUBGROUP_FEATURE_QUAD_BIT
;
1592 if (pdevice
->info
.gen
>= 8) {
1593 /* TODO: There's no technical reason why these can't be made to
1594 * work on gen7 but they don't at the moment so it's best to leave
1595 * the feature disabled than enabled and broken.
1597 p
->subgroupSupportedOperations
|= VK_SUBGROUP_FEATURE_ARITHMETIC_BIT
|
1598 VK_SUBGROUP_FEATURE_CLUSTERED_BIT
;
1600 p
->subgroupQuadOperationsInAllStages
= pdevice
->info
.gen
>= 8;
1602 p
->pointClippingBehavior
= VK_POINT_CLIPPING_BEHAVIOR_USER_CLIP_PLANES_ONLY
;
1603 p
->maxMultiviewViewCount
= 16;
1604 p
->maxMultiviewInstanceIndex
= UINT32_MAX
/ 16;
1605 p
->protectedNoFault
= false;
1606 /* This value doesn't matter for us today as our per-stage descriptors are
1609 p
->maxPerSetDescriptors
= 1024;
1610 p
->maxMemoryAllocationSize
= MAX_MEMORY_ALLOCATION_SIZE
;
1614 anv_get_physical_device_properties_1_2(struct anv_physical_device
*pdevice
,
1615 VkPhysicalDeviceVulkan12Properties
*p
)
1617 assert(p
->sType
== VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES
);
1619 p
->driverID
= VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA_KHR
;
1620 memset(p
->driverName
, 0, sizeof(p
->driverName
));
1621 snprintf(p
->driverName
, VK_MAX_DRIVER_NAME_SIZE_KHR
,
1622 "Intel open-source Mesa driver");
1623 memset(p
->driverInfo
, 0, sizeof(p
->driverInfo
));
1624 snprintf(p
->driverInfo
, VK_MAX_DRIVER_INFO_SIZE_KHR
,
1625 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1
);
1626 p
->conformanceVersion
= (VkConformanceVersionKHR
) {
1633 p
->denormBehaviorIndependence
=
1634 VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_ALL_KHR
;
1635 p
->roundingModeIndependence
=
1636 VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_NONE_KHR
;
1638 /* Broadwell does not support HF denorms and there are restrictions
1639 * other gens. According to Kabylake's PRM:
1641 * "math - Extended Math Function
1643 * Restriction : Half-float denorms are always retained."
1645 p
->shaderDenormFlushToZeroFloat16
= false;
1646 p
->shaderDenormPreserveFloat16
= pdevice
->info
.gen
> 8;
1647 p
->shaderRoundingModeRTEFloat16
= true;
1648 p
->shaderRoundingModeRTZFloat16
= true;
1649 p
->shaderSignedZeroInfNanPreserveFloat16
= true;
1651 p
->shaderDenormFlushToZeroFloat32
= true;
1652 p
->shaderDenormPreserveFloat32
= true;
1653 p
->shaderRoundingModeRTEFloat32
= true;
1654 p
->shaderRoundingModeRTZFloat32
= true;
1655 p
->shaderSignedZeroInfNanPreserveFloat32
= true;
1657 p
->shaderDenormFlushToZeroFloat64
= true;
1658 p
->shaderDenormPreserveFloat64
= true;
1659 p
->shaderRoundingModeRTEFloat64
= true;
1660 p
->shaderRoundingModeRTZFloat64
= true;
1661 p
->shaderSignedZeroInfNanPreserveFloat64
= true;
1663 /* It's a bit hard to exactly map our implementation to the limits
1664 * described here. The bindless surface handle in the extended
1665 * message descriptors is 20 bits and it's an index into the table of
1666 * RENDER_SURFACE_STATE structs that starts at bindless surface base
1667 * address. Given that most things consume two surface states per
1668 * view (general/sampled for textures and write-only/read-write for
1669 * images), we claim 2^19 things.
1671 * For SSBOs, we just use A64 messages so there is no real limit
1672 * there beyond the limit on the total size of a descriptor set.
1674 const unsigned max_bindless_views
= 1 << 19;
1675 p
->maxUpdateAfterBindDescriptorsInAllPools
= max_bindless_views
;
1676 p
->shaderUniformBufferArrayNonUniformIndexingNative
= false;
1677 p
->shaderSampledImageArrayNonUniformIndexingNative
= false;
1678 p
->shaderStorageBufferArrayNonUniformIndexingNative
= true;
1679 p
->shaderStorageImageArrayNonUniformIndexingNative
= false;
1680 p
->shaderInputAttachmentArrayNonUniformIndexingNative
= false;
1681 p
->robustBufferAccessUpdateAfterBind
= true;
1682 p
->quadDivergentImplicitLod
= false;
1683 p
->maxPerStageDescriptorUpdateAfterBindSamplers
= max_bindless_views
;
1684 p
->maxPerStageDescriptorUpdateAfterBindUniformBuffers
= MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
;
1685 p
->maxPerStageDescriptorUpdateAfterBindStorageBuffers
= UINT32_MAX
;
1686 p
->maxPerStageDescriptorUpdateAfterBindSampledImages
= max_bindless_views
;
1687 p
->maxPerStageDescriptorUpdateAfterBindStorageImages
= max_bindless_views
;
1688 p
->maxPerStageDescriptorUpdateAfterBindInputAttachments
= MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS
;
1689 p
->maxPerStageUpdateAfterBindResources
= UINT32_MAX
;
1690 p
->maxDescriptorSetUpdateAfterBindSamplers
= max_bindless_views
;
1691 p
->maxDescriptorSetUpdateAfterBindUniformBuffers
= 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
;
1692 p
->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2;
1693 p
->maxDescriptorSetUpdateAfterBindStorageBuffers
= UINT32_MAX
;
1694 p
->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2;
1695 p
->maxDescriptorSetUpdateAfterBindSampledImages
= max_bindless_views
;
1696 p
->maxDescriptorSetUpdateAfterBindStorageImages
= max_bindless_views
;
1697 p
->maxDescriptorSetUpdateAfterBindInputAttachments
= MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS
;
1699 /* We support all of the depth resolve modes */
1700 p
->supportedDepthResolveModes
= VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
|
1701 VK_RESOLVE_MODE_AVERAGE_BIT_KHR
|
1702 VK_RESOLVE_MODE_MIN_BIT_KHR
|
1703 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1704 /* Average doesn't make sense for stencil so we don't support that */
1705 p
->supportedStencilResolveModes
= VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
;
1706 if (pdevice
->info
.gen
>= 8) {
1707 /* The advanced stencil resolve modes currently require stencil
1708 * sampling be supported by the hardware.
1710 p
->supportedStencilResolveModes
|= VK_RESOLVE_MODE_MIN_BIT_KHR
|
1711 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1713 p
->independentResolveNone
= true;
1714 p
->independentResolve
= true;
1716 p
->filterMinmaxSingleComponentFormats
= pdevice
->info
.gen
>= 9;
1717 p
->filterMinmaxImageComponentMapping
= pdevice
->info
.gen
>= 9;
1719 p
->maxTimelineSemaphoreValueDifference
= UINT64_MAX
;
1721 p
->framebufferIntegerColorSampleCounts
=
1722 isl_device_get_sample_counts(&pdevice
->isl_dev
);
1725 void anv_GetPhysicalDeviceProperties2(
1726 VkPhysicalDevice physicalDevice
,
1727 VkPhysicalDeviceProperties2
* pProperties
)
1729 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1731 anv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
1733 VkPhysicalDeviceVulkan11Properties core_1_1
= {
1734 .sType
= VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES
,
1736 anv_get_physical_device_properties_1_1(pdevice
, &core_1_1
);
1738 VkPhysicalDeviceVulkan12Properties core_1_2
= {
1739 .sType
= VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES
,
1741 anv_get_physical_device_properties_1_2(pdevice
, &core_1_2
);
1743 #define CORE_RENAMED_PROPERTY(major, minor, ext_property, core_property) \
1744 memcpy(&properties->ext_property, &core_##major##_##minor.core_property, \
1745 sizeof(core_##major##_##minor.core_property))
1747 #define CORE_PROPERTY(major, minor, property) \
1748 CORE_RENAMED_PROPERTY(major, minor, property, property)
1750 vk_foreach_struct(ext
, pProperties
->pNext
) {
1751 switch (ext
->sType
) {
1752 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_STENCIL_RESOLVE_PROPERTIES_KHR
: {
1753 VkPhysicalDeviceDepthStencilResolvePropertiesKHR
*properties
=
1754 (VkPhysicalDeviceDepthStencilResolvePropertiesKHR
*)ext
;
1755 CORE_PROPERTY(1, 2, supportedDepthResolveModes
);
1756 CORE_PROPERTY(1, 2, supportedStencilResolveModes
);
1757 CORE_PROPERTY(1, 2, independentResolveNone
);
1758 CORE_PROPERTY(1, 2, independentResolve
);
1762 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES_EXT
: {
1763 VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*properties
=
1764 (VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*)ext
;
1765 CORE_PROPERTY(1, 2, maxUpdateAfterBindDescriptorsInAllPools
);
1766 CORE_PROPERTY(1, 2, shaderUniformBufferArrayNonUniformIndexingNative
);
1767 CORE_PROPERTY(1, 2, shaderSampledImageArrayNonUniformIndexingNative
);
1768 CORE_PROPERTY(1, 2, shaderStorageBufferArrayNonUniformIndexingNative
);
1769 CORE_PROPERTY(1, 2, shaderStorageImageArrayNonUniformIndexingNative
);
1770 CORE_PROPERTY(1, 2, shaderInputAttachmentArrayNonUniformIndexingNative
);
1771 CORE_PROPERTY(1, 2, robustBufferAccessUpdateAfterBind
);
1772 CORE_PROPERTY(1, 2, quadDivergentImplicitLod
);
1773 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindSamplers
);
1774 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindUniformBuffers
);
1775 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindStorageBuffers
);
1776 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindSampledImages
);
1777 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindStorageImages
);
1778 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindInputAttachments
);
1779 CORE_PROPERTY(1, 2, maxPerStageUpdateAfterBindResources
);
1780 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindSamplers
);
1781 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindUniformBuffers
);
1782 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindUniformBuffersDynamic
);
1783 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindStorageBuffers
);
1784 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindStorageBuffersDynamic
);
1785 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindSampledImages
);
1786 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindStorageImages
);
1787 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindInputAttachments
);
1791 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR
: {
1792 VkPhysicalDeviceDriverPropertiesKHR
*properties
=
1793 (VkPhysicalDeviceDriverPropertiesKHR
*) ext
;
1794 CORE_PROPERTY(1, 2, driverID
);
1795 CORE_PROPERTY(1, 2, driverName
);
1796 CORE_PROPERTY(1, 2, driverInfo
);
1797 CORE_PROPERTY(1, 2, conformanceVersion
);
1801 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT
: {
1802 VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*props
=
1803 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*) ext
;
1804 /* Userptr needs page aligned memory. */
1805 props
->minImportedHostPointerAlignment
= 4096;
1809 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES
: {
1810 VkPhysicalDeviceIDProperties
*properties
=
1811 (VkPhysicalDeviceIDProperties
*)ext
;
1812 CORE_PROPERTY(1, 1, deviceUUID
);
1813 CORE_PROPERTY(1, 1, driverUUID
);
1814 CORE_PROPERTY(1, 1, deviceLUID
);
1815 CORE_PROPERTY(1, 1, deviceLUIDValid
);
1819 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_PROPERTIES_EXT
: {
1820 VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*props
=
1821 (VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*)ext
;
1822 props
->maxInlineUniformBlockSize
= MAX_INLINE_UNIFORM_BLOCK_SIZE
;
1823 props
->maxPerStageDescriptorInlineUniformBlocks
=
1824 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1825 props
->maxPerStageDescriptorUpdateAfterBindInlineUniformBlocks
=
1826 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1827 props
->maxDescriptorSetInlineUniformBlocks
=
1828 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1829 props
->maxDescriptorSetUpdateAfterBindInlineUniformBlocks
=
1830 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1834 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_PROPERTIES_EXT
: {
1835 VkPhysicalDeviceLineRasterizationPropertiesEXT
*props
=
1836 (VkPhysicalDeviceLineRasterizationPropertiesEXT
*)ext
;
1837 /* In the Skylake PRM Vol. 7, subsection titled "GIQ (Diamond)
1838 * Sampling Rules - Legacy Mode", it says the following:
1840 * "Note that the device divides a pixel into a 16x16 array of
1841 * subpixels, referenced by their upper left corners."
1843 * This is the only known reference in the PRMs to the subpixel
1844 * precision of line rasterization and a "16x16 array of subpixels"
1845 * implies 4 subpixel precision bits. Empirical testing has shown
1846 * that 4 subpixel precision bits applies to all line rasterization
1849 props
->lineSubPixelPrecisionBits
= 4;
1853 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES
: {
1854 VkPhysicalDeviceMaintenance3Properties
*properties
=
1855 (VkPhysicalDeviceMaintenance3Properties
*)ext
;
1856 /* This value doesn't matter for us today as our per-stage
1857 * descriptors are the real limit.
1859 CORE_PROPERTY(1, 1, maxPerSetDescriptors
);
1860 CORE_PROPERTY(1, 1, maxMemoryAllocationSize
);
1864 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES
: {
1865 VkPhysicalDeviceMultiviewProperties
*properties
=
1866 (VkPhysicalDeviceMultiviewProperties
*)ext
;
1867 CORE_PROPERTY(1, 1, maxMultiviewViewCount
);
1868 CORE_PROPERTY(1, 1, maxMultiviewInstanceIndex
);
1872 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT
: {
1873 VkPhysicalDevicePCIBusInfoPropertiesEXT
*properties
=
1874 (VkPhysicalDevicePCIBusInfoPropertiesEXT
*)ext
;
1875 properties
->pciDomain
= pdevice
->pci_info
.domain
;
1876 properties
->pciBus
= pdevice
->pci_info
.bus
;
1877 properties
->pciDevice
= pdevice
->pci_info
.device
;
1878 properties
->pciFunction
= pdevice
->pci_info
.function
;
1882 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES
: {
1883 VkPhysicalDevicePointClippingProperties
*properties
=
1884 (VkPhysicalDevicePointClippingProperties
*) ext
;
1885 CORE_PROPERTY(1, 1, pointClippingBehavior
);
1889 #pragma GCC diagnostic push
1890 #pragma GCC diagnostic ignored "-Wswitch"
1891 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PRESENTATION_PROPERTIES_ANDROID
: {
1892 VkPhysicalDevicePresentationPropertiesANDROID
*props
=
1893 (VkPhysicalDevicePresentationPropertiesANDROID
*)ext
;
1894 props
->sharedImage
= VK_FALSE
;
1897 #pragma GCC diagnostic pop
1899 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES
: {
1900 VkPhysicalDeviceProtectedMemoryProperties
*properties
=
1901 (VkPhysicalDeviceProtectedMemoryProperties
*)ext
;
1902 CORE_PROPERTY(1, 1, protectedNoFault
);
1906 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
1907 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
1908 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
1909 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
1913 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ROBUSTNESS_2_PROPERTIES_EXT
: {
1914 VkPhysicalDeviceRobustness2PropertiesEXT
*properties
= (void *)ext
;
1915 properties
->robustStorageBufferAccessSizeAlignment
=
1916 ANV_SSBO_BOUNDS_CHECK_ALIGNMENT
;
1917 properties
->robustUniformBufferAccessSizeAlignment
=
1918 ANV_UBO_BOUNDS_CHECK_ALIGNMENT
;
1922 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT
: {
1923 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*properties
=
1924 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*)ext
;
1925 CORE_PROPERTY(1, 2, filterMinmaxImageComponentMapping
);
1926 CORE_PROPERTY(1, 2, filterMinmaxSingleComponentFormats
);
1930 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES
: {
1931 VkPhysicalDeviceSubgroupProperties
*properties
= (void *)ext
;
1932 CORE_PROPERTY(1, 1, subgroupSize
);
1933 CORE_RENAMED_PROPERTY(1, 1, supportedStages
,
1934 subgroupSupportedStages
);
1935 CORE_RENAMED_PROPERTY(1, 1, supportedOperations
,
1936 subgroupSupportedOperations
);
1937 CORE_RENAMED_PROPERTY(1, 1, quadOperationsInAllStages
,
1938 subgroupQuadOperationsInAllStages
);
1942 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_PROPERTIES_EXT
: {
1943 VkPhysicalDeviceSubgroupSizeControlPropertiesEXT
*props
=
1944 (VkPhysicalDeviceSubgroupSizeControlPropertiesEXT
*)ext
;
1945 STATIC_ASSERT(8 <= BRW_SUBGROUP_SIZE
&& BRW_SUBGROUP_SIZE
<= 32);
1946 props
->minSubgroupSize
= 8;
1947 props
->maxSubgroupSize
= 32;
1948 props
->maxComputeWorkgroupSubgroups
= pdevice
->info
.max_cs_threads
;
1949 props
->requiredSubgroupSizeStages
= VK_SHADER_STAGE_COMPUTE_BIT
;
1952 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT_CONTROLS_PROPERTIES_KHR
: {
1953 VkPhysicalDeviceFloatControlsPropertiesKHR
*properties
= (void *)ext
;
1954 CORE_PROPERTY(1, 2, denormBehaviorIndependence
);
1955 CORE_PROPERTY(1, 2, roundingModeIndependence
);
1956 CORE_PROPERTY(1, 2, shaderDenormFlushToZeroFloat16
);
1957 CORE_PROPERTY(1, 2, shaderDenormPreserveFloat16
);
1958 CORE_PROPERTY(1, 2, shaderRoundingModeRTEFloat16
);
1959 CORE_PROPERTY(1, 2, shaderRoundingModeRTZFloat16
);
1960 CORE_PROPERTY(1, 2, shaderSignedZeroInfNanPreserveFloat16
);
1961 CORE_PROPERTY(1, 2, shaderDenormFlushToZeroFloat32
);
1962 CORE_PROPERTY(1, 2, shaderDenormPreserveFloat32
);
1963 CORE_PROPERTY(1, 2, shaderRoundingModeRTEFloat32
);
1964 CORE_PROPERTY(1, 2, shaderRoundingModeRTZFloat32
);
1965 CORE_PROPERTY(1, 2, shaderSignedZeroInfNanPreserveFloat32
);
1966 CORE_PROPERTY(1, 2, shaderDenormFlushToZeroFloat64
);
1967 CORE_PROPERTY(1, 2, shaderDenormPreserveFloat64
);
1968 CORE_PROPERTY(1, 2, shaderRoundingModeRTEFloat64
);
1969 CORE_PROPERTY(1, 2, shaderRoundingModeRTZFloat64
);
1970 CORE_PROPERTY(1, 2, shaderSignedZeroInfNanPreserveFloat64
);
1974 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_PROPERTIES_EXT
: {
1975 VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT
*props
=
1976 (VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT
*)ext
;
1978 /* From the SKL PRM Vol. 2d, docs for RENDER_SURFACE_STATE::Surface
1981 * "For SURFTYPE_BUFFER non-rendertarget surfaces, this field
1982 * specifies the base address of the first element of the surface,
1983 * computed in software by adding the surface base address to the
1984 * byte offset of the element in the buffer. The base address must
1985 * be aligned to element size."
1987 * The typed dataport messages require that things be texel aligned.
1988 * Otherwise, we may just load/store the wrong data or, in the worst
1989 * case, there may be hangs.
1991 props
->storageTexelBufferOffsetAlignmentBytes
= 16;
1992 props
->storageTexelBufferOffsetSingleTexelAlignment
= true;
1994 /* The sampler, however, is much more forgiving and it can handle
1995 * arbitrary byte alignment for linear and buffer surfaces. It's
1996 * hard to find a good PRM citation for this but years of empirical
1997 * experience demonstrate that this is true.
1999 props
->uniformTexelBufferOffsetAlignmentBytes
= 1;
2000 props
->uniformTexelBufferOffsetSingleTexelAlignment
= false;
2004 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_PROPERTIES_KHR
: {
2005 VkPhysicalDeviceTimelineSemaphorePropertiesKHR
*properties
=
2006 (VkPhysicalDeviceTimelineSemaphorePropertiesKHR
*) ext
;
2007 CORE_PROPERTY(1, 2, maxTimelineSemaphoreValueDifference
);
2011 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT
: {
2012 VkPhysicalDeviceTransformFeedbackPropertiesEXT
*props
=
2013 (VkPhysicalDeviceTransformFeedbackPropertiesEXT
*)ext
;
2015 props
->maxTransformFeedbackStreams
= MAX_XFB_STREAMS
;
2016 props
->maxTransformFeedbackBuffers
= MAX_XFB_BUFFERS
;
2017 props
->maxTransformFeedbackBufferSize
= (1ull << 32);
2018 props
->maxTransformFeedbackStreamDataSize
= 128 * 4;
2019 props
->maxTransformFeedbackBufferDataSize
= 128 * 4;
2020 props
->maxTransformFeedbackBufferDataStride
= 2048;
2021 props
->transformFeedbackQueries
= true;
2022 props
->transformFeedbackStreamsLinesTriangles
= false;
2023 props
->transformFeedbackRasterizationStreamSelect
= false;
2024 props
->transformFeedbackDraw
= true;
2028 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT
: {
2029 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*props
=
2030 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*)ext
;
2031 /* We have to restrict this a bit for multiview */
2032 props
->maxVertexAttribDivisor
= UINT32_MAX
/ 16;
2036 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES
:
2037 anv_get_physical_device_properties_1_1(pdevice
, (void *)ext
);
2040 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES
:
2041 anv_get_physical_device_properties_1_2(pdevice
, (void *)ext
);
2045 anv_debug_ignored_stype(ext
->sType
);
2050 #undef CORE_RENAMED_PROPERTY
2051 #undef CORE_PROPERTY
2054 /* We support exactly one queue family. */
2055 static const VkQueueFamilyProperties
2056 anv_queue_family_properties
= {
2057 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
2058 VK_QUEUE_COMPUTE_BIT
|
2059 VK_QUEUE_TRANSFER_BIT
,
2061 .timestampValidBits
= 36, /* XXX: Real value here */
2062 .minImageTransferGranularity
= { 1, 1, 1 },
2065 void anv_GetPhysicalDeviceQueueFamilyProperties(
2066 VkPhysicalDevice physicalDevice
,
2068 VkQueueFamilyProperties
* pQueueFamilyProperties
)
2070 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pCount
);
2072 vk_outarray_append(&out
, p
) {
2073 *p
= anv_queue_family_properties
;
2077 void anv_GetPhysicalDeviceQueueFamilyProperties2(
2078 VkPhysicalDevice physicalDevice
,
2079 uint32_t* pQueueFamilyPropertyCount
,
2080 VkQueueFamilyProperties2
* pQueueFamilyProperties
)
2083 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pQueueFamilyPropertyCount
);
2085 vk_outarray_append(&out
, p
) {
2086 p
->queueFamilyProperties
= anv_queue_family_properties
;
2088 vk_foreach_struct(s
, p
->pNext
) {
2089 anv_debug_ignored_stype(s
->sType
);
2094 void anv_GetPhysicalDeviceMemoryProperties(
2095 VkPhysicalDevice physicalDevice
,
2096 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
2098 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
2100 pMemoryProperties
->memoryTypeCount
= physical_device
->memory
.type_count
;
2101 for (uint32_t i
= 0; i
< physical_device
->memory
.type_count
; i
++) {
2102 pMemoryProperties
->memoryTypes
[i
] = (VkMemoryType
) {
2103 .propertyFlags
= physical_device
->memory
.types
[i
].propertyFlags
,
2104 .heapIndex
= physical_device
->memory
.types
[i
].heapIndex
,
2108 pMemoryProperties
->memoryHeapCount
= physical_device
->memory
.heap_count
;
2109 for (uint32_t i
= 0; i
< physical_device
->memory
.heap_count
; i
++) {
2110 pMemoryProperties
->memoryHeaps
[i
] = (VkMemoryHeap
) {
2111 .size
= physical_device
->memory
.heaps
[i
].size
,
2112 .flags
= physical_device
->memory
.heaps
[i
].flags
,
2118 anv_get_memory_budget(VkPhysicalDevice physicalDevice
,
2119 VkPhysicalDeviceMemoryBudgetPropertiesEXT
*memoryBudget
)
2121 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
2122 uint64_t sys_available
= get_available_system_memory();
2123 assert(sys_available
> 0);
2125 VkDeviceSize total_heaps_size
= 0;
2126 for (size_t i
= 0; i
< device
->memory
.heap_count
; i
++)
2127 total_heaps_size
+= device
->memory
.heaps
[i
].size
;
2129 for (size_t i
= 0; i
< device
->memory
.heap_count
; i
++) {
2130 VkDeviceSize heap_size
= device
->memory
.heaps
[i
].size
;
2131 VkDeviceSize heap_used
= device
->memory
.heaps
[i
].used
;
2132 VkDeviceSize heap_budget
;
2134 double heap_proportion
= (double) heap_size
/ total_heaps_size
;
2135 VkDeviceSize sys_available_prop
= sys_available
* heap_proportion
;
2138 * Let's not incite the app to starve the system: report at most 90% of
2139 * available system memory.
2141 uint64_t heap_available
= sys_available_prop
* 9 / 10;
2142 heap_budget
= MIN2(heap_size
, heap_used
+ heap_available
);
2145 * Round down to the nearest MB
2147 heap_budget
&= ~((1ull << 20) - 1);
2150 * The heapBudget value must be non-zero for array elements less than
2151 * VkPhysicalDeviceMemoryProperties::memoryHeapCount. The heapBudget
2152 * value must be less than or equal to VkMemoryHeap::size for each heap.
2154 assert(0 < heap_budget
&& heap_budget
<= heap_size
);
2156 memoryBudget
->heapUsage
[i
] = heap_used
;
2157 memoryBudget
->heapBudget
[i
] = heap_budget
;
2160 /* The heapBudget and heapUsage values must be zero for array elements
2161 * greater than or equal to VkPhysicalDeviceMemoryProperties::memoryHeapCount
2163 for (uint32_t i
= device
->memory
.heap_count
; i
< VK_MAX_MEMORY_HEAPS
; i
++) {
2164 memoryBudget
->heapBudget
[i
] = 0;
2165 memoryBudget
->heapUsage
[i
] = 0;
2169 void anv_GetPhysicalDeviceMemoryProperties2(
2170 VkPhysicalDevice physicalDevice
,
2171 VkPhysicalDeviceMemoryProperties2
* pMemoryProperties
)
2173 anv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
2174 &pMemoryProperties
->memoryProperties
);
2176 vk_foreach_struct(ext
, pMemoryProperties
->pNext
) {
2177 switch (ext
->sType
) {
2178 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT
:
2179 anv_get_memory_budget(physicalDevice
, (void*)ext
);
2182 anv_debug_ignored_stype(ext
->sType
);
2189 anv_GetDeviceGroupPeerMemoryFeatures(
2192 uint32_t localDeviceIndex
,
2193 uint32_t remoteDeviceIndex
,
2194 VkPeerMemoryFeatureFlags
* pPeerMemoryFeatures
)
2196 assert(localDeviceIndex
== 0 && remoteDeviceIndex
== 0);
2197 *pPeerMemoryFeatures
= VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT
|
2198 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT
|
2199 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT
|
2200 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT
;
2203 PFN_vkVoidFunction
anv_GetInstanceProcAddr(
2204 VkInstance _instance
,
2207 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2209 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
2210 * when we have to return valid function pointers, NULL, or it's left
2211 * undefined. See the table for exact details.
2216 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
2217 if (strcmp(pName, "vk" #entrypoint) == 0) \
2218 return (PFN_vkVoidFunction)anv_##entrypoint
2220 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties
);
2221 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties
);
2222 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceVersion
);
2223 LOOKUP_ANV_ENTRYPOINT(CreateInstance
);
2225 /* GetInstanceProcAddr() can also be called with a NULL instance.
2226 * See https://gitlab.khronos.org/vulkan/vulkan/issues/2057
2228 LOOKUP_ANV_ENTRYPOINT(GetInstanceProcAddr
);
2230 #undef LOOKUP_ANV_ENTRYPOINT
2232 if (instance
== NULL
)
2235 int idx
= anv_get_instance_entrypoint_index(pName
);
2237 return instance
->dispatch
.entrypoints
[idx
];
2239 idx
= anv_get_physical_device_entrypoint_index(pName
);
2241 return instance
->physical_device_dispatch
.entrypoints
[idx
];
2243 idx
= anv_get_device_entrypoint_index(pName
);
2245 return instance
->device_dispatch
.entrypoints
[idx
];
2250 /* With version 1+ of the loader interface the ICD should expose
2251 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
2254 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
2255 VkInstance instance
,
2259 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
2260 VkInstance instance
,
2263 return anv_GetInstanceProcAddr(instance
, pName
);
2266 PFN_vkVoidFunction
anv_GetDeviceProcAddr(
2270 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2272 if (!device
|| !pName
)
2275 int idx
= anv_get_device_entrypoint_index(pName
);
2279 return device
->dispatch
.entrypoints
[idx
];
2282 /* With version 4+ of the loader interface the ICD should expose
2283 * vk_icdGetPhysicalDeviceProcAddr()
2286 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetPhysicalDeviceProcAddr(
2287 VkInstance _instance
,
2290 PFN_vkVoidFunction
vk_icdGetPhysicalDeviceProcAddr(
2291 VkInstance _instance
,
2294 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2296 if (!pName
|| !instance
)
2299 int idx
= anv_get_physical_device_entrypoint_index(pName
);
2303 return instance
->physical_device_dispatch
.entrypoints
[idx
];
2308 anv_CreateDebugReportCallbackEXT(VkInstance _instance
,
2309 const VkDebugReportCallbackCreateInfoEXT
* pCreateInfo
,
2310 const VkAllocationCallbacks
* pAllocator
,
2311 VkDebugReportCallbackEXT
* pCallback
)
2313 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2314 return vk_create_debug_report_callback(&instance
->debug_report_callbacks
,
2315 pCreateInfo
, pAllocator
, &instance
->alloc
,
2320 anv_DestroyDebugReportCallbackEXT(VkInstance _instance
,
2321 VkDebugReportCallbackEXT _callback
,
2322 const VkAllocationCallbacks
* pAllocator
)
2324 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2325 vk_destroy_debug_report_callback(&instance
->debug_report_callbacks
,
2326 _callback
, pAllocator
, &instance
->alloc
);
2330 anv_DebugReportMessageEXT(VkInstance _instance
,
2331 VkDebugReportFlagsEXT flags
,
2332 VkDebugReportObjectTypeEXT objectType
,
2335 int32_t messageCode
,
2336 const char* pLayerPrefix
,
2337 const char* pMessage
)
2339 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2340 vk_debug_report(&instance
->debug_report_callbacks
, flags
, objectType
,
2341 object
, location
, messageCode
, pLayerPrefix
, pMessage
);
2344 static struct anv_state
2345 anv_state_pool_emit_data(struct anv_state_pool
*pool
, size_t size
, size_t align
, const void *p
)
2347 struct anv_state state
;
2349 state
= anv_state_pool_alloc(pool
, size
, align
);
2350 memcpy(state
.map
, p
, size
);
2355 /* Haswell border color is a bit of a disaster. Float and unorm formats use a
2356 * straightforward 32-bit float color in the first 64 bytes. Instead of using
2357 * a nice float/integer union like Gen8+, Haswell specifies the integer border
2358 * color as a separate entry /after/ the float color. The layout of this entry
2359 * also depends on the format's bpp (with extra hacks for RG32), and overlaps.
2361 * Since we don't know the format/bpp, we can't make any of the border colors
2362 * containing '1' work for all formats, as it would be in the wrong place for
2363 * some of them. We opt to make 32-bit integers work as this seems like the
2364 * most common option. Fortunately, transparent black works regardless, as
2365 * all zeroes is the same in every bit-size.
2367 struct hsw_border_color
{
2371 uint32_t _pad1
[108];
2374 struct gen8_border_color
{
2379 /* Pad out to 64 bytes */
2384 anv_device_init_border_colors(struct anv_device
*device
)
2386 if (device
->info
.is_haswell
) {
2387 static const struct hsw_border_color border_colors
[] = {
2388 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
2389 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
2390 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
2391 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
2392 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
2393 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
2396 device
->border_colors
=
2397 anv_state_pool_emit_data(&device
->dynamic_state_pool
,
2398 sizeof(border_colors
), 512, border_colors
);
2400 static const struct gen8_border_color border_colors
[] = {
2401 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
2402 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
2403 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
2404 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
2405 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
2406 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
2409 device
->border_colors
=
2410 anv_state_pool_emit_data(&device
->dynamic_state_pool
,
2411 sizeof(border_colors
), 64, border_colors
);
2416 anv_device_init_trivial_batch(struct anv_device
*device
)
2418 VkResult result
= anv_device_alloc_bo(device
, 4096,
2419 ANV_BO_ALLOC_MAPPED
,
2420 0 /* explicit_address */,
2421 &device
->trivial_batch_bo
);
2422 if (result
!= VK_SUCCESS
)
2425 struct anv_batch batch
= {
2426 .start
= device
->trivial_batch_bo
->map
,
2427 .next
= device
->trivial_batch_bo
->map
,
2428 .end
= device
->trivial_batch_bo
->map
+ 4096,
2431 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
2432 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
2434 if (!device
->info
.has_llc
)
2435 gen_clflush_range(batch
.start
, batch
.next
- batch
.start
);
2440 VkResult
anv_EnumerateDeviceExtensionProperties(
2441 VkPhysicalDevice physicalDevice
,
2442 const char* pLayerName
,
2443 uint32_t* pPropertyCount
,
2444 VkExtensionProperties
* pProperties
)
2446 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
2447 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
2449 for (int i
= 0; i
< ANV_DEVICE_EXTENSION_COUNT
; i
++) {
2450 if (device
->supported_extensions
.extensions
[i
]) {
2451 vk_outarray_append(&out
, prop
) {
2452 *prop
= anv_device_extensions
[i
];
2457 return vk_outarray_status(&out
);
2461 anv_device_init_dispatch(struct anv_device
*device
)
2463 const struct anv_instance
*instance
= device
->physical
->instance
;
2465 const struct anv_device_dispatch_table
*genX_table
;
2466 switch (device
->info
.gen
) {
2468 genX_table
= &gen12_device_dispatch_table
;
2471 genX_table
= &gen11_device_dispatch_table
;
2474 genX_table
= &gen10_device_dispatch_table
;
2477 genX_table
= &gen9_device_dispatch_table
;
2480 genX_table
= &gen8_device_dispatch_table
;
2483 if (device
->info
.is_haswell
)
2484 genX_table
= &gen75_device_dispatch_table
;
2486 genX_table
= &gen7_device_dispatch_table
;
2489 unreachable("unsupported gen\n");
2492 for (unsigned i
= 0; i
< ARRAY_SIZE(device
->dispatch
.entrypoints
); i
++) {
2493 /* Vulkan requires that entrypoints for extensions which have not been
2494 * enabled must not be advertised.
2496 if (!anv_device_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
2497 &instance
->enabled_extensions
,
2498 &device
->enabled_extensions
)) {
2499 device
->dispatch
.entrypoints
[i
] = NULL
;
2500 } else if (genX_table
->entrypoints
[i
]) {
2501 device
->dispatch
.entrypoints
[i
] = genX_table
->entrypoints
[i
];
2503 device
->dispatch
.entrypoints
[i
] =
2504 anv_device_dispatch_table
.entrypoints
[i
];
2510 vk_priority_to_gen(int priority
)
2513 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT
:
2514 return GEN_CONTEXT_LOW_PRIORITY
;
2515 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
:
2516 return GEN_CONTEXT_MEDIUM_PRIORITY
;
2517 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT
:
2518 return GEN_CONTEXT_HIGH_PRIORITY
;
2519 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT
:
2520 return GEN_CONTEXT_REALTIME_PRIORITY
;
2522 unreachable("Invalid priority");
2527 anv_device_init_hiz_clear_value_bo(struct anv_device
*device
)
2529 VkResult result
= anv_device_alloc_bo(device
, 4096,
2530 ANV_BO_ALLOC_MAPPED
,
2531 0 /* explicit_address */,
2532 &device
->hiz_clear_bo
);
2533 if (result
!= VK_SUCCESS
)
2536 union isl_color_value hiz_clear
= { .u32
= { 0, } };
2537 hiz_clear
.f32
[0] = ANV_HZ_FC_VAL
;
2539 memcpy(device
->hiz_clear_bo
->map
, hiz_clear
.u32
, sizeof(hiz_clear
.u32
));
2541 if (!device
->info
.has_llc
)
2542 gen_clflush_range(device
->hiz_clear_bo
->map
, sizeof(hiz_clear
.u32
));
2548 get_bo_from_pool(struct gen_batch_decode_bo
*ret
,
2549 struct anv_block_pool
*pool
,
2552 anv_block_pool_foreach_bo(bo
, pool
) {
2553 uint64_t bo_address
= gen_48b_address(bo
->offset
);
2554 if (address
>= bo_address
&& address
< (bo_address
+ bo
->size
)) {
2555 *ret
= (struct gen_batch_decode_bo
) {
2566 /* Finding a buffer for batch decoding */
2567 static struct gen_batch_decode_bo
2568 decode_get_bo(void *v_batch
, bool ppgtt
, uint64_t address
)
2570 struct anv_device
*device
= v_batch
;
2571 struct gen_batch_decode_bo ret_bo
= {};
2575 if (get_bo_from_pool(&ret_bo
, &device
->dynamic_state_pool
.block_pool
, address
))
2577 if (get_bo_from_pool(&ret_bo
, &device
->instruction_state_pool
.block_pool
, address
))
2579 if (get_bo_from_pool(&ret_bo
, &device
->binding_table_pool
.block_pool
, address
))
2581 if (get_bo_from_pool(&ret_bo
, &device
->surface_state_pool
.block_pool
, address
))
2584 if (!device
->cmd_buffer_being_decoded
)
2585 return (struct gen_batch_decode_bo
) { };
2587 struct anv_batch_bo
**bo
;
2589 u_vector_foreach(bo
, &device
->cmd_buffer_being_decoded
->seen_bbos
) {
2590 /* The decoder zeroes out the top 16 bits, so we need to as well */
2591 uint64_t bo_address
= (*bo
)->bo
->offset
& (~0ull >> 16);
2593 if (address
>= bo_address
&& address
< bo_address
+ (*bo
)->bo
->size
) {
2594 return (struct gen_batch_decode_bo
) {
2596 .size
= (*bo
)->bo
->size
,
2597 .map
= (*bo
)->bo
->map
,
2602 return (struct gen_batch_decode_bo
) { };
2605 struct gen_aux_map_buffer
{
2606 struct gen_buffer base
;
2607 struct anv_state state
;
2610 static struct gen_buffer
*
2611 gen_aux_map_buffer_alloc(void *driver_ctx
, uint32_t size
)
2613 struct gen_aux_map_buffer
*buf
= malloc(sizeof(struct gen_aux_map_buffer
));
2617 struct anv_device
*device
= (struct anv_device
*)driver_ctx
;
2618 assert(device
->physical
->supports_48bit_addresses
&&
2619 device
->physical
->use_softpin
);
2621 struct anv_state_pool
*pool
= &device
->dynamic_state_pool
;
2622 buf
->state
= anv_state_pool_alloc(pool
, size
, size
);
2624 buf
->base
.gpu
= pool
->block_pool
.bo
->offset
+ buf
->state
.offset
;
2625 buf
->base
.gpu_end
= buf
->base
.gpu
+ buf
->state
.alloc_size
;
2626 buf
->base
.map
= buf
->state
.map
;
2627 buf
->base
.driver_bo
= &buf
->state
;
2632 gen_aux_map_buffer_free(void *driver_ctx
, struct gen_buffer
*buffer
)
2634 struct gen_aux_map_buffer
*buf
= (struct gen_aux_map_buffer
*)buffer
;
2635 struct anv_device
*device
= (struct anv_device
*)driver_ctx
;
2636 struct anv_state_pool
*pool
= &device
->dynamic_state_pool
;
2637 anv_state_pool_free(pool
, buf
->state
);
2641 static struct gen_mapped_pinned_buffer_alloc aux_map_allocator
= {
2642 .alloc
= gen_aux_map_buffer_alloc
,
2643 .free
= gen_aux_map_buffer_free
,
2647 check_physical_device_features(VkPhysicalDevice physicalDevice
,
2648 const VkPhysicalDeviceFeatures
*features
)
2650 VkPhysicalDeviceFeatures supported_features
;
2651 anv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
2652 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
2653 VkBool32
*enabled_feature
= (VkBool32
*)features
;
2654 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
2655 for (uint32_t i
= 0; i
< num_features
; i
++) {
2656 if (enabled_feature
[i
] && !supported_feature
[i
])
2657 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
2663 VkResult
anv_CreateDevice(
2664 VkPhysicalDevice physicalDevice
,
2665 const VkDeviceCreateInfo
* pCreateInfo
,
2666 const VkAllocationCallbacks
* pAllocator
,
2669 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
2671 struct anv_device
*device
;
2673 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO
);
2675 struct anv_device_extension_table enabled_extensions
= { };
2676 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
2678 for (idx
= 0; idx
< ANV_DEVICE_EXTENSION_COUNT
; idx
++) {
2679 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
2680 anv_device_extensions
[idx
].extensionName
) == 0)
2684 if (idx
>= ANV_DEVICE_EXTENSION_COUNT
)
2685 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
2687 if (!physical_device
->supported_extensions
.extensions
[idx
])
2688 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
2690 enabled_extensions
.extensions
[idx
] = true;
2693 /* Check enabled features */
2694 bool robust_buffer_access
= false;
2695 if (pCreateInfo
->pEnabledFeatures
) {
2696 result
= check_physical_device_features(physicalDevice
,
2697 pCreateInfo
->pEnabledFeatures
);
2698 if (result
!= VK_SUCCESS
)
2701 if (pCreateInfo
->pEnabledFeatures
->robustBufferAccess
)
2702 robust_buffer_access
= true;
2705 vk_foreach_struct_const(ext
, pCreateInfo
->pNext
) {
2706 switch (ext
->sType
) {
2707 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2
: {
2708 const VkPhysicalDeviceFeatures2
*features
= (const void *)ext
;
2709 result
= check_physical_device_features(physicalDevice
,
2710 &features
->features
);
2711 if (result
!= VK_SUCCESS
)
2714 if (features
->features
.robustBufferAccess
)
2715 robust_buffer_access
= true;
2725 /* Check requested queues and fail if we are requested to create any
2726 * queues with flags we don't support.
2728 assert(pCreateInfo
->queueCreateInfoCount
> 0);
2729 for (uint32_t i
= 0; i
< pCreateInfo
->queueCreateInfoCount
; i
++) {
2730 if (pCreateInfo
->pQueueCreateInfos
[i
].flags
!= 0)
2731 return vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2734 /* Check if client specified queue priority. */
2735 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*queue_priority
=
2736 vk_find_struct_const(pCreateInfo
->pQueueCreateInfos
[0].pNext
,
2737 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT
);
2739 VkQueueGlobalPriorityEXT priority
=
2740 queue_priority
? queue_priority
->globalPriority
:
2741 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
;
2743 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
2745 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
2747 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2749 vk_device_init(&device
->vk
, pCreateInfo
,
2750 &physical_device
->instance
->alloc
, pAllocator
);
2752 if (INTEL_DEBUG
& DEBUG_BATCH
) {
2753 const unsigned decode_flags
=
2754 GEN_BATCH_DECODE_FULL
|
2755 ((INTEL_DEBUG
& DEBUG_COLOR
) ? GEN_BATCH_DECODE_IN_COLOR
: 0) |
2756 GEN_BATCH_DECODE_OFFSETS
|
2757 GEN_BATCH_DECODE_FLOATS
;
2759 gen_batch_decode_ctx_init(&device
->decoder_ctx
,
2760 &physical_device
->info
,
2761 stderr
, decode_flags
, NULL
,
2762 decode_get_bo
, NULL
, device
);
2765 device
->physical
= physical_device
;
2766 device
->no_hw
= physical_device
->no_hw
;
2767 device
->_lost
= false;
2769 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
2770 device
->fd
= open(physical_device
->path
, O_RDWR
| O_CLOEXEC
);
2771 if (device
->fd
== -1) {
2772 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2776 device
->context_id
= anv_gem_create_context(device
);
2777 if (device
->context_id
== -1) {
2778 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2782 result
= anv_queue_init(device
, &device
->queue
);
2783 if (result
!= VK_SUCCESS
)
2784 goto fail_context_id
;
2786 if (physical_device
->use_softpin
) {
2787 if (pthread_mutex_init(&device
->vma_mutex
, NULL
) != 0) {
2788 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2792 /* keep the page with address zero out of the allocator */
2793 util_vma_heap_init(&device
->vma_lo
,
2794 LOW_HEAP_MIN_ADDRESS
, LOW_HEAP_SIZE
);
2796 util_vma_heap_init(&device
->vma_cva
, CLIENT_VISIBLE_HEAP_MIN_ADDRESS
,
2797 CLIENT_VISIBLE_HEAP_SIZE
);
2799 /* Leave the last 4GiB out of the high vma range, so that no state
2800 * base address + size can overflow 48 bits. For more information see
2801 * the comment about Wa32bitGeneralStateOffset in anv_allocator.c
2803 util_vma_heap_init(&device
->vma_hi
, HIGH_HEAP_MIN_ADDRESS
,
2804 physical_device
->gtt_size
- (1ull << 32) -
2805 HIGH_HEAP_MIN_ADDRESS
);
2808 list_inithead(&device
->memory_objects
);
2810 /* As per spec, the driver implementation may deny requests to acquire
2811 * a priority above the default priority (MEDIUM) if the caller does not
2812 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
2815 if (physical_device
->has_context_priority
) {
2816 int err
= anv_gem_set_context_param(device
->fd
, device
->context_id
,
2817 I915_CONTEXT_PARAM_PRIORITY
,
2818 vk_priority_to_gen(priority
));
2819 if (err
!= 0 && priority
> VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
) {
2820 result
= vk_error(VK_ERROR_NOT_PERMITTED_EXT
);
2825 device
->info
= physical_device
->info
;
2826 device
->isl_dev
= physical_device
->isl_dev
;
2828 /* On Broadwell and later, we can use batch chaining to more efficiently
2829 * implement growing command buffers. Prior to Haswell, the kernel
2830 * command parser gets in the way and we have to fall back to growing
2833 device
->can_chain_batches
= device
->info
.gen
>= 8;
2835 device
->robust_buffer_access
= robust_buffer_access
;
2836 device
->enabled_extensions
= enabled_extensions
;
2838 anv_device_init_dispatch(device
);
2840 if (pthread_mutex_init(&device
->mutex
, NULL
) != 0) {
2841 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2845 pthread_condattr_t condattr
;
2846 if (pthread_condattr_init(&condattr
) != 0) {
2847 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2850 if (pthread_condattr_setclock(&condattr
, CLOCK_MONOTONIC
) != 0) {
2851 pthread_condattr_destroy(&condattr
);
2852 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2855 if (pthread_cond_init(&device
->queue_submit
, &condattr
) != 0) {
2856 pthread_condattr_destroy(&condattr
);
2857 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2860 pthread_condattr_destroy(&condattr
);
2862 result
= anv_bo_cache_init(&device
->bo_cache
);
2863 if (result
!= VK_SUCCESS
)
2864 goto fail_queue_cond
;
2866 anv_bo_pool_init(&device
->batch_bo_pool
, device
);
2868 result
= anv_state_pool_init(&device
->dynamic_state_pool
, device
,
2869 DYNAMIC_STATE_POOL_MIN_ADDRESS
, 16384);
2870 if (result
!= VK_SUCCESS
)
2871 goto fail_batch_bo_pool
;
2873 result
= anv_state_pool_init(&device
->instruction_state_pool
, device
,
2874 INSTRUCTION_STATE_POOL_MIN_ADDRESS
, 16384);
2875 if (result
!= VK_SUCCESS
)
2876 goto fail_dynamic_state_pool
;
2878 result
= anv_state_pool_init(&device
->surface_state_pool
, device
,
2879 SURFACE_STATE_POOL_MIN_ADDRESS
, 4096);
2880 if (result
!= VK_SUCCESS
)
2881 goto fail_instruction_state_pool
;
2883 if (physical_device
->use_softpin
) {
2884 result
= anv_state_pool_init(&device
->binding_table_pool
, device
,
2885 BINDING_TABLE_POOL_MIN_ADDRESS
, 4096);
2886 if (result
!= VK_SUCCESS
)
2887 goto fail_surface_state_pool
;
2890 if (device
->info
.gen
>= 12) {
2891 device
->aux_map_ctx
= gen_aux_map_init(device
, &aux_map_allocator
,
2892 &physical_device
->info
);
2893 if (!device
->aux_map_ctx
)
2894 goto fail_binding_table_pool
;
2897 result
= anv_device_alloc_bo(device
, 4096, 0 /* flags */,
2898 0 /* explicit_address */,
2899 &device
->workaround_bo
);
2900 if (result
!= VK_SUCCESS
)
2901 goto fail_surface_aux_map_pool
;
2903 result
= anv_device_init_trivial_batch(device
);
2904 if (result
!= VK_SUCCESS
)
2905 goto fail_workaround_bo
;
2907 /* Allocate a null surface state at surface state offset 0. This makes
2908 * NULL descriptor handling trivial because we can just memset structures
2909 * to zero and they have a valid descriptor.
2911 device
->null_surface_state
=
2912 anv_state_pool_alloc(&device
->surface_state_pool
,
2913 device
->isl_dev
.ss
.size
,
2914 device
->isl_dev
.ss
.align
);
2915 isl_null_fill_state(&device
->isl_dev
, device
->null_surface_state
.map
,
2916 isl_extent3d(1, 1, 1) /* This shouldn't matter */);
2917 assert(device
->null_surface_state
.offset
== 0);
2919 if (device
->info
.gen
>= 10) {
2920 result
= anv_device_init_hiz_clear_value_bo(device
);
2921 if (result
!= VK_SUCCESS
)
2922 goto fail_trivial_batch_bo
;
2925 anv_scratch_pool_init(device
, &device
->scratch_pool
);
2927 switch (device
->info
.gen
) {
2929 if (!device
->info
.is_haswell
)
2930 result
= gen7_init_device_state(device
);
2932 result
= gen75_init_device_state(device
);
2935 result
= gen8_init_device_state(device
);
2938 result
= gen9_init_device_state(device
);
2941 result
= gen10_init_device_state(device
);
2944 result
= gen11_init_device_state(device
);
2947 result
= gen12_init_device_state(device
);
2950 /* Shouldn't get here as we don't create physical devices for any other
2952 unreachable("unhandled gen");
2954 if (result
!= VK_SUCCESS
)
2955 goto fail_workaround_bo
;
2957 anv_pipeline_cache_init(&device
->default_pipeline_cache
, device
, true);
2959 anv_device_init_blorp(device
);
2961 anv_device_init_border_colors(device
);
2963 anv_device_perf_init(device
);
2965 *pDevice
= anv_device_to_handle(device
);
2970 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
2971 if (device
->info
.gen
>= 10)
2972 anv_device_release_bo(device
, device
->hiz_clear_bo
);
2973 anv_device_release_bo(device
, device
->workaround_bo
);
2974 fail_trivial_batch_bo
:
2975 anv_device_release_bo(device
, device
->trivial_batch_bo
);
2976 fail_surface_aux_map_pool
:
2977 if (device
->info
.gen
>= 12) {
2978 gen_aux_map_finish(device
->aux_map_ctx
);
2979 device
->aux_map_ctx
= NULL
;
2981 fail_binding_table_pool
:
2982 if (physical_device
->use_softpin
)
2983 anv_state_pool_finish(&device
->binding_table_pool
);
2984 fail_surface_state_pool
:
2985 anv_state_pool_finish(&device
->surface_state_pool
);
2986 fail_instruction_state_pool
:
2987 anv_state_pool_finish(&device
->instruction_state_pool
);
2988 fail_dynamic_state_pool
:
2989 anv_state_pool_finish(&device
->dynamic_state_pool
);
2991 anv_bo_pool_finish(&device
->batch_bo_pool
);
2992 anv_bo_cache_finish(&device
->bo_cache
);
2994 pthread_cond_destroy(&device
->queue_submit
);
2996 pthread_mutex_destroy(&device
->mutex
);
2998 if (physical_device
->use_softpin
) {
2999 util_vma_heap_finish(&device
->vma_hi
);
3000 util_vma_heap_finish(&device
->vma_cva
);
3001 util_vma_heap_finish(&device
->vma_lo
);
3004 anv_queue_finish(&device
->queue
);
3006 anv_gem_destroy_context(device
, device
->context_id
);
3010 vk_free(&device
->vk
.alloc
, device
);
3015 void anv_DestroyDevice(
3017 const VkAllocationCallbacks
* pAllocator
)
3019 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3024 anv_device_finish_blorp(device
);
3026 anv_pipeline_cache_finish(&device
->default_pipeline_cache
);
3028 anv_queue_finish(&device
->queue
);
3030 #ifdef HAVE_VALGRIND
3031 /* We only need to free these to prevent valgrind errors. The backing
3032 * BO will go away in a couple of lines so we don't actually leak.
3034 anv_state_pool_free(&device
->dynamic_state_pool
, device
->border_colors
);
3035 anv_state_pool_free(&device
->dynamic_state_pool
, device
->slice_hash
);
3038 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
3040 anv_device_release_bo(device
, device
->workaround_bo
);
3041 anv_device_release_bo(device
, device
->trivial_batch_bo
);
3042 if (device
->info
.gen
>= 10)
3043 anv_device_release_bo(device
, device
->hiz_clear_bo
);
3045 if (device
->info
.gen
>= 12) {
3046 gen_aux_map_finish(device
->aux_map_ctx
);
3047 device
->aux_map_ctx
= NULL
;
3050 if (device
->physical
->use_softpin
)
3051 anv_state_pool_finish(&device
->binding_table_pool
);
3052 anv_state_pool_finish(&device
->surface_state_pool
);
3053 anv_state_pool_finish(&device
->instruction_state_pool
);
3054 anv_state_pool_finish(&device
->dynamic_state_pool
);
3056 anv_bo_pool_finish(&device
->batch_bo_pool
);
3058 anv_bo_cache_finish(&device
->bo_cache
);
3060 if (device
->physical
->use_softpin
) {
3061 util_vma_heap_finish(&device
->vma_hi
);
3062 util_vma_heap_finish(&device
->vma_cva
);
3063 util_vma_heap_finish(&device
->vma_lo
);
3066 pthread_cond_destroy(&device
->queue_submit
);
3067 pthread_mutex_destroy(&device
->mutex
);
3069 anv_gem_destroy_context(device
, device
->context_id
);
3071 if (INTEL_DEBUG
& DEBUG_BATCH
)
3072 gen_batch_decode_ctx_finish(&device
->decoder_ctx
);
3076 vk_device_finish(&device
->vk
);
3077 vk_free(&device
->vk
.alloc
, device
);
3080 VkResult
anv_EnumerateInstanceLayerProperties(
3081 uint32_t* pPropertyCount
,
3082 VkLayerProperties
* pProperties
)
3084 if (pProperties
== NULL
) {
3085 *pPropertyCount
= 0;
3089 /* None supported at this time */
3090 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
3093 VkResult
anv_EnumerateDeviceLayerProperties(
3094 VkPhysicalDevice physicalDevice
,
3095 uint32_t* pPropertyCount
,
3096 VkLayerProperties
* pProperties
)
3098 if (pProperties
== NULL
) {
3099 *pPropertyCount
= 0;
3103 /* None supported at this time */
3104 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
3107 void anv_GetDeviceQueue(
3109 uint32_t queueNodeIndex
,
3110 uint32_t queueIndex
,
3113 const VkDeviceQueueInfo2 info
= {
3114 .sType
= VK_STRUCTURE_TYPE_DEVICE_QUEUE_INFO_2
,
3117 .queueFamilyIndex
= queueNodeIndex
,
3118 .queueIndex
= queueIndex
,
3121 anv_GetDeviceQueue2(_device
, &info
, pQueue
);
3124 void anv_GetDeviceQueue2(
3126 const VkDeviceQueueInfo2
* pQueueInfo
,
3129 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3131 assert(pQueueInfo
->queueIndex
== 0);
3133 if (pQueueInfo
->flags
== device
->queue
.flags
)
3134 *pQueue
= anv_queue_to_handle(&device
->queue
);
3140 _anv_device_set_lost(struct anv_device
*device
,
3141 const char *file
, int line
,
3142 const char *msg
, ...)
3147 p_atomic_inc(&device
->_lost
);
3150 err
= __vk_errorv(device
->physical
->instance
, device
,
3151 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT
,
3152 VK_ERROR_DEVICE_LOST
, file
, line
, msg
, ap
);
3155 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
3162 _anv_queue_set_lost(struct anv_queue
*queue
,
3163 const char *file
, int line
,
3164 const char *msg
, ...)
3169 p_atomic_inc(&queue
->device
->_lost
);
3172 err
= __vk_errorv(queue
->device
->physical
->instance
, queue
->device
,
3173 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT
,
3174 VK_ERROR_DEVICE_LOST
, file
, line
, msg
, ap
);
3177 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
3184 anv_device_query_status(struct anv_device
*device
)
3186 /* This isn't likely as most of the callers of this function already check
3187 * for it. However, it doesn't hurt to check and it potentially lets us
3190 if (anv_device_is_lost(device
))
3191 return VK_ERROR_DEVICE_LOST
;
3193 uint32_t active
, pending
;
3194 int ret
= anv_gem_gpu_get_reset_stats(device
, &active
, &pending
);
3196 /* We don't know the real error. */
3197 return anv_device_set_lost(device
, "get_reset_stats failed: %m");
3201 return anv_device_set_lost(device
, "GPU hung on one of our command buffers");
3202 } else if (pending
) {
3203 return anv_device_set_lost(device
, "GPU hung with commands in-flight");
3210 anv_device_bo_busy(struct anv_device
*device
, struct anv_bo
*bo
)
3212 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
3213 * Other usages of the BO (such as on different hardware) will not be
3214 * flagged as "busy" by this ioctl. Use with care.
3216 int ret
= anv_gem_busy(device
, bo
->gem_handle
);
3218 return VK_NOT_READY
;
3219 } else if (ret
== -1) {
3220 /* We don't know the real error. */
3221 return anv_device_set_lost(device
, "gem wait failed: %m");
3224 /* Query for device status after the busy call. If the BO we're checking
3225 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
3226 * client because it clearly doesn't have valid data. Yes, this most
3227 * likely means an ioctl, but we just did an ioctl to query the busy status
3228 * so it's no great loss.
3230 return anv_device_query_status(device
);
3234 anv_device_wait(struct anv_device
*device
, struct anv_bo
*bo
,
3237 int ret
= anv_gem_wait(device
, bo
->gem_handle
, &timeout
);
3238 if (ret
== -1 && errno
== ETIME
) {
3240 } else if (ret
== -1) {
3241 /* We don't know the real error. */
3242 return anv_device_set_lost(device
, "gem wait failed: %m");
3245 /* Query for device status after the wait. If the BO we're waiting on got
3246 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
3247 * because it clearly doesn't have valid data. Yes, this most likely means
3248 * an ioctl, but we just did an ioctl to wait so it's no great loss.
3250 return anv_device_query_status(device
);
3253 VkResult
anv_DeviceWaitIdle(
3256 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3258 if (anv_device_is_lost(device
))
3259 return VK_ERROR_DEVICE_LOST
;
3261 return anv_queue_submit_simple_batch(&device
->queue
, NULL
);
3265 anv_vma_alloc(struct anv_device
*device
,
3266 uint64_t size
, uint64_t align
,
3267 enum anv_bo_alloc_flags alloc_flags
,
3268 uint64_t client_address
)
3270 pthread_mutex_lock(&device
->vma_mutex
);
3274 if (alloc_flags
& ANV_BO_ALLOC_CLIENT_VISIBLE_ADDRESS
) {
3275 if (client_address
) {
3276 if (util_vma_heap_alloc_addr(&device
->vma_cva
,
3277 client_address
, size
)) {
3278 addr
= client_address
;
3281 addr
= util_vma_heap_alloc(&device
->vma_cva
, size
, align
);
3283 /* We don't want to fall back to other heaps */
3287 assert(client_address
== 0);
3289 if (!(alloc_flags
& ANV_BO_ALLOC_32BIT_ADDRESS
))
3290 addr
= util_vma_heap_alloc(&device
->vma_hi
, size
, align
);
3293 addr
= util_vma_heap_alloc(&device
->vma_lo
, size
, align
);
3296 pthread_mutex_unlock(&device
->vma_mutex
);
3298 assert(addr
== gen_48b_address(addr
));
3299 return gen_canonical_address(addr
);
3303 anv_vma_free(struct anv_device
*device
,
3304 uint64_t address
, uint64_t size
)
3306 const uint64_t addr_48b
= gen_48b_address(address
);
3308 pthread_mutex_lock(&device
->vma_mutex
);
3310 if (addr_48b
>= LOW_HEAP_MIN_ADDRESS
&&
3311 addr_48b
<= LOW_HEAP_MAX_ADDRESS
) {
3312 util_vma_heap_free(&device
->vma_lo
, addr_48b
, size
);
3313 } else if (addr_48b
>= CLIENT_VISIBLE_HEAP_MIN_ADDRESS
&&
3314 addr_48b
<= CLIENT_VISIBLE_HEAP_MAX_ADDRESS
) {
3315 util_vma_heap_free(&device
->vma_cva
, addr_48b
, size
);
3317 assert(addr_48b
>= HIGH_HEAP_MIN_ADDRESS
);
3318 util_vma_heap_free(&device
->vma_hi
, addr_48b
, size
);
3321 pthread_mutex_unlock(&device
->vma_mutex
);
3324 VkResult
anv_AllocateMemory(
3326 const VkMemoryAllocateInfo
* pAllocateInfo
,
3327 const VkAllocationCallbacks
* pAllocator
,
3328 VkDeviceMemory
* pMem
)
3330 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3331 struct anv_physical_device
*pdevice
= device
->physical
;
3332 struct anv_device_memory
*mem
;
3333 VkResult result
= VK_SUCCESS
;
3335 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
3337 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
3338 assert(pAllocateInfo
->allocationSize
> 0);
3340 VkDeviceSize aligned_alloc_size
=
3341 align_u64(pAllocateInfo
->allocationSize
, 4096);
3343 if (aligned_alloc_size
> MAX_MEMORY_ALLOCATION_SIZE
)
3344 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
3346 assert(pAllocateInfo
->memoryTypeIndex
< pdevice
->memory
.type_count
);
3347 struct anv_memory_type
*mem_type
=
3348 &pdevice
->memory
.types
[pAllocateInfo
->memoryTypeIndex
];
3349 assert(mem_type
->heapIndex
< pdevice
->memory
.heap_count
);
3350 struct anv_memory_heap
*mem_heap
=
3351 &pdevice
->memory
.heaps
[mem_type
->heapIndex
];
3353 uint64_t mem_heap_used
= p_atomic_read(&mem_heap
->used
);
3354 if (mem_heap_used
+ aligned_alloc_size
> mem_heap
->size
)
3355 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
3357 mem
= vk_alloc2(&device
->vk
.alloc
, pAllocator
, sizeof(*mem
), 8,
3358 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3360 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3362 assert(pAllocateInfo
->memoryTypeIndex
< pdevice
->memory
.type_count
);
3363 mem
->type
= mem_type
;
3367 mem
->host_ptr
= NULL
;
3369 enum anv_bo_alloc_flags alloc_flags
= 0;
3371 const VkExportMemoryAllocateInfo
*export_info
= NULL
;
3372 const VkImportAndroidHardwareBufferInfoANDROID
*ahw_import_info
= NULL
;
3373 const VkImportMemoryFdInfoKHR
*fd_info
= NULL
;
3374 const VkImportMemoryHostPointerInfoEXT
*host_ptr_info
= NULL
;
3375 const VkMemoryDedicatedAllocateInfo
*dedicated_info
= NULL
;
3376 VkMemoryAllocateFlags vk_flags
= 0;
3377 uint64_t client_address
= 0;
3379 vk_foreach_struct_const(ext
, pAllocateInfo
->pNext
) {
3380 switch (ext
->sType
) {
3381 case VK_STRUCTURE_TYPE_EXPORT_MEMORY_ALLOCATE_INFO
:
3382 export_info
= (void *)ext
;
3385 case VK_STRUCTURE_TYPE_IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID
:
3386 ahw_import_info
= (void *)ext
;
3389 case VK_STRUCTURE_TYPE_IMPORT_MEMORY_FD_INFO_KHR
:
3390 fd_info
= (void *)ext
;
3393 case VK_STRUCTURE_TYPE_IMPORT_MEMORY_HOST_POINTER_INFO_EXT
:
3394 host_ptr_info
= (void *)ext
;
3397 case VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_FLAGS_INFO
: {
3398 const VkMemoryAllocateFlagsInfo
*flags_info
= (void *)ext
;
3399 vk_flags
= flags_info
->flags
;
3403 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_ALLOCATE_INFO
:
3404 dedicated_info
= (void *)ext
;
3407 case VK_STRUCTURE_TYPE_MEMORY_OPAQUE_CAPTURE_ADDRESS_ALLOCATE_INFO_KHR
: {
3408 const VkMemoryOpaqueCaptureAddressAllocateInfoKHR
*addr_info
=
3409 (const VkMemoryOpaqueCaptureAddressAllocateInfoKHR
*)ext
;
3410 client_address
= addr_info
->opaqueCaptureAddress
;
3415 anv_debug_ignored_stype(ext
->sType
);
3420 /* By default, we want all VkDeviceMemory objects to support CCS */
3421 if (device
->physical
->has_implicit_ccs
)
3422 alloc_flags
|= ANV_BO_ALLOC_IMPLICIT_CCS
;
3424 if (vk_flags
& VK_MEMORY_ALLOCATE_DEVICE_ADDRESS_BIT_KHR
)
3425 alloc_flags
|= ANV_BO_ALLOC_CLIENT_VISIBLE_ADDRESS
;
3427 if ((export_info
&& export_info
->handleTypes
) ||
3428 (fd_info
&& fd_info
->handleType
) ||
3429 (host_ptr_info
&& host_ptr_info
->handleType
)) {
3430 /* Anything imported or exported is EXTERNAL */
3431 alloc_flags
|= ANV_BO_ALLOC_EXTERNAL
;
3433 /* We can't have implicit CCS on external memory with an AUX-table.
3434 * Doing so would require us to sync the aux tables across processes
3435 * which is impractical.
3437 if (device
->info
.has_aux_map
)
3438 alloc_flags
&= ~ANV_BO_ALLOC_IMPLICIT_CCS
;
3441 /* Check if we need to support Android HW buffer export. If so,
3442 * create AHardwareBuffer and import memory from it.
3444 bool android_export
= false;
3445 if (export_info
&& export_info
->handleTypes
&
3446 VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID
)
3447 android_export
= true;
3449 if (ahw_import_info
) {
3450 result
= anv_import_ahw_memory(_device
, mem
, ahw_import_info
);
3451 if (result
!= VK_SUCCESS
)
3455 } else if (android_export
) {
3456 result
= anv_create_ahw_memory(_device
, mem
, pAllocateInfo
);
3457 if (result
!= VK_SUCCESS
)
3460 const VkImportAndroidHardwareBufferInfoANDROID import_info
= {
3463 result
= anv_import_ahw_memory(_device
, mem
, &import_info
);
3464 if (result
!= VK_SUCCESS
)
3470 /* The Vulkan spec permits handleType to be 0, in which case the struct is
3473 if (fd_info
&& fd_info
->handleType
) {
3474 /* At the moment, we support only the below handle types. */
3475 assert(fd_info
->handleType
==
3476 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
3477 fd_info
->handleType
==
3478 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
3480 result
= anv_device_import_bo(device
, fd_info
->fd
, alloc_flags
,
3481 client_address
, &mem
->bo
);
3482 if (result
!= VK_SUCCESS
)
3485 /* For security purposes, we reject importing the bo if it's smaller
3486 * than the requested allocation size. This prevents a malicious client
3487 * from passing a buffer to a trusted client, lying about the size, and
3488 * telling the trusted client to try and texture from an image that goes
3489 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
3490 * in the trusted client. The trusted client can protect itself against
3491 * this sort of attack but only if it can trust the buffer size.
3493 if (mem
->bo
->size
< aligned_alloc_size
) {
3494 result
= vk_errorf(device
, device
, VK_ERROR_INVALID_EXTERNAL_HANDLE
,
3495 "aligned allocationSize too large for "
3496 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT: "
3497 "%"PRIu64
"B > %"PRIu64
"B",
3498 aligned_alloc_size
, mem
->bo
->size
);
3499 anv_device_release_bo(device
, mem
->bo
);
3503 /* From the Vulkan spec:
3505 * "Importing memory from a file descriptor transfers ownership of
3506 * the file descriptor from the application to the Vulkan
3507 * implementation. The application must not perform any operations on
3508 * the file descriptor after a successful import."
3510 * If the import fails, we leave the file descriptor open.
3516 if (host_ptr_info
&& host_ptr_info
->handleType
) {
3517 if (host_ptr_info
->handleType
==
3518 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_MAPPED_FOREIGN_MEMORY_BIT_EXT
) {
3519 result
= vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
3523 assert(host_ptr_info
->handleType
==
3524 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
);
3526 result
= anv_device_import_bo_from_host_ptr(device
,
3527 host_ptr_info
->pHostPointer
,
3528 pAllocateInfo
->allocationSize
,
3532 if (result
!= VK_SUCCESS
)
3535 mem
->host_ptr
= host_ptr_info
->pHostPointer
;
3539 /* Regular allocate (not importing memory). */
3541 result
= anv_device_alloc_bo(device
, pAllocateInfo
->allocationSize
,
3542 alloc_flags
, client_address
, &mem
->bo
);
3543 if (result
!= VK_SUCCESS
)
3546 if (dedicated_info
&& dedicated_info
->image
!= VK_NULL_HANDLE
) {
3547 ANV_FROM_HANDLE(anv_image
, image
, dedicated_info
->image
);
3549 /* Some legacy (non-modifiers) consumers need the tiling to be set on
3550 * the BO. In this case, we have a dedicated allocation.
3552 if (image
->needs_set_tiling
) {
3553 const uint32_t i915_tiling
=
3554 isl_tiling_to_i915_tiling(image
->planes
[0].surface
.isl
.tiling
);
3555 int ret
= anv_gem_set_tiling(device
, mem
->bo
->gem_handle
,
3556 image
->planes
[0].surface
.isl
.row_pitch_B
,
3559 anv_device_release_bo(device
, mem
->bo
);
3560 result
= vk_errorf(device
, device
, VK_ERROR_OUT_OF_DEVICE_MEMORY
,
3561 "failed to set BO tiling: %m");
3568 mem_heap_used
= p_atomic_add_return(&mem_heap
->used
, mem
->bo
->size
);
3569 if (mem_heap_used
> mem_heap
->size
) {
3570 p_atomic_add(&mem_heap
->used
, -mem
->bo
->size
);
3571 anv_device_release_bo(device
, mem
->bo
);
3572 result
= vk_errorf(device
, device
, VK_ERROR_OUT_OF_DEVICE_MEMORY
,
3573 "Out of heap memory");
3577 pthread_mutex_lock(&device
->mutex
);
3578 list_addtail(&mem
->link
, &device
->memory_objects
);
3579 pthread_mutex_unlock(&device
->mutex
);
3581 *pMem
= anv_device_memory_to_handle(mem
);
3586 vk_free2(&device
->vk
.alloc
, pAllocator
, mem
);
3591 VkResult
anv_GetMemoryFdKHR(
3593 const VkMemoryGetFdInfoKHR
* pGetFdInfo
,
3596 ANV_FROM_HANDLE(anv_device
, dev
, device_h
);
3597 ANV_FROM_HANDLE(anv_device_memory
, mem
, pGetFdInfo
->memory
);
3599 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
3601 assert(pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
3602 pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
3604 return anv_device_export_bo(dev
, mem
->bo
, pFd
);
3607 VkResult
anv_GetMemoryFdPropertiesKHR(
3609 VkExternalMemoryHandleTypeFlagBits handleType
,
3611 VkMemoryFdPropertiesKHR
* pMemoryFdProperties
)
3613 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3615 switch (handleType
) {
3616 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
:
3617 /* dma-buf can be imported as any memory type */
3618 pMemoryFdProperties
->memoryTypeBits
=
3619 (1 << device
->physical
->memory
.type_count
) - 1;
3623 /* The valid usage section for this function says:
3625 * "handleType must not be one of the handle types defined as
3628 * So opaque handle types fall into the default "unsupported" case.
3630 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
3634 VkResult
anv_GetMemoryHostPointerPropertiesEXT(
3636 VkExternalMemoryHandleTypeFlagBits handleType
,
3637 const void* pHostPointer
,
3638 VkMemoryHostPointerPropertiesEXT
* pMemoryHostPointerProperties
)
3640 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3642 assert(pMemoryHostPointerProperties
->sType
==
3643 VK_STRUCTURE_TYPE_MEMORY_HOST_POINTER_PROPERTIES_EXT
);
3645 switch (handleType
) {
3646 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
:
3647 /* Host memory can be imported as any memory type. */
3648 pMemoryHostPointerProperties
->memoryTypeBits
=
3649 (1ull << device
->physical
->memory
.type_count
) - 1;
3654 return VK_ERROR_INVALID_EXTERNAL_HANDLE
;
3658 void anv_FreeMemory(
3660 VkDeviceMemory _mem
,
3661 const VkAllocationCallbacks
* pAllocator
)
3663 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3664 ANV_FROM_HANDLE(anv_device_memory
, mem
, _mem
);
3669 pthread_mutex_lock(&device
->mutex
);
3670 list_del(&mem
->link
);
3671 pthread_mutex_unlock(&device
->mutex
);
3674 anv_UnmapMemory(_device
, _mem
);
3676 p_atomic_add(&device
->physical
->memory
.heaps
[mem
->type
->heapIndex
].used
,
3679 anv_device_release_bo(device
, mem
->bo
);
3681 #if defined(ANDROID) && ANDROID_API_LEVEL >= 26
3683 AHardwareBuffer_release(mem
->ahw
);
3686 vk_free2(&device
->vk
.alloc
, pAllocator
, mem
);
3689 VkResult
anv_MapMemory(
3691 VkDeviceMemory _memory
,
3692 VkDeviceSize offset
,
3694 VkMemoryMapFlags flags
,
3697 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3698 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
3705 if (mem
->host_ptr
) {
3706 *ppData
= mem
->host_ptr
+ offset
;
3710 if (size
== VK_WHOLE_SIZE
)
3711 size
= mem
->bo
->size
- offset
;
3713 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
3715 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
3716 * assert(size != 0);
3717 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
3718 * equal to the size of the memory minus offset
3721 assert(offset
+ size
<= mem
->bo
->size
);
3723 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
3724 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
3725 * at a time is valid. We could just mmap up front and return an offset
3726 * pointer here, but that may exhaust virtual memory on 32 bit
3729 uint32_t gem_flags
= 0;
3731 if (!device
->info
.has_llc
&&
3732 (mem
->type
->propertyFlags
& VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
))
3733 gem_flags
|= I915_MMAP_WC
;
3735 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
3736 uint64_t map_offset
;
3737 if (!device
->physical
->has_mmap_offset
)
3738 map_offset
= offset
& ~4095ull;
3741 assert(offset
>= map_offset
);
3742 uint64_t map_size
= (offset
+ size
) - map_offset
;
3744 /* Let's map whole pages */
3745 map_size
= align_u64(map_size
, 4096);
3747 void *map
= anv_gem_mmap(device
, mem
->bo
->gem_handle
,
3748 map_offset
, map_size
, gem_flags
);
3749 if (map
== MAP_FAILED
)
3750 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
3753 mem
->map_size
= map_size
;
3755 *ppData
= mem
->map
+ (offset
- map_offset
);
3760 void anv_UnmapMemory(
3762 VkDeviceMemory _memory
)
3764 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3765 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
3767 if (mem
== NULL
|| mem
->host_ptr
)
3770 anv_gem_munmap(device
, mem
->map
, mem
->map_size
);
3777 clflush_mapped_ranges(struct anv_device
*device
,
3779 const VkMappedMemoryRange
*ranges
)
3781 for (uint32_t i
= 0; i
< count
; i
++) {
3782 ANV_FROM_HANDLE(anv_device_memory
, mem
, ranges
[i
].memory
);
3783 if (ranges
[i
].offset
>= mem
->map_size
)
3786 gen_clflush_range(mem
->map
+ ranges
[i
].offset
,
3787 MIN2(ranges
[i
].size
, mem
->map_size
- ranges
[i
].offset
));
3791 VkResult
anv_FlushMappedMemoryRanges(
3793 uint32_t memoryRangeCount
,
3794 const VkMappedMemoryRange
* pMemoryRanges
)
3796 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3798 if (device
->info
.has_llc
)
3801 /* Make sure the writes we're flushing have landed. */
3802 __builtin_ia32_mfence();
3804 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
3809 VkResult
anv_InvalidateMappedMemoryRanges(
3811 uint32_t memoryRangeCount
,
3812 const VkMappedMemoryRange
* pMemoryRanges
)
3814 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3816 if (device
->info
.has_llc
)
3819 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
3821 /* Make sure no reads get moved up above the invalidate. */
3822 __builtin_ia32_mfence();
3827 void anv_GetBufferMemoryRequirements(
3830 VkMemoryRequirements
* pMemoryRequirements
)
3832 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
3833 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3835 /* The Vulkan spec (git aaed022) says:
3837 * memoryTypeBits is a bitfield and contains one bit set for every
3838 * supported memory type for the resource. The bit `1<<i` is set if and
3839 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3840 * structure for the physical device is supported.
3842 uint32_t memory_types
= (1ull << device
->physical
->memory
.type_count
) - 1;
3844 /* Base alignment requirement of a cache line */
3845 uint32_t alignment
= 16;
3847 /* We need an alignment of 32 for pushing UBOs */
3848 if (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
)
3849 alignment
= MAX2(alignment
, 32);
3851 pMemoryRequirements
->size
= buffer
->size
;
3852 pMemoryRequirements
->alignment
= alignment
;
3854 /* Storage and Uniform buffers should have their size aligned to
3855 * 32-bits to avoid boundary checks when last DWord is not complete.
3856 * This would ensure that not internal padding would be needed for
3859 if (device
->robust_buffer_access
&&
3860 (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
||
3861 buffer
->usage
& VK_BUFFER_USAGE_STORAGE_BUFFER_BIT
))
3862 pMemoryRequirements
->size
= align_u64(buffer
->size
, 4);
3864 pMemoryRequirements
->memoryTypeBits
= memory_types
;
3867 void anv_GetBufferMemoryRequirements2(
3869 const VkBufferMemoryRequirementsInfo2
* pInfo
,
3870 VkMemoryRequirements2
* pMemoryRequirements
)
3872 anv_GetBufferMemoryRequirements(_device
, pInfo
->buffer
,
3873 &pMemoryRequirements
->memoryRequirements
);
3875 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3876 switch (ext
->sType
) {
3877 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
3878 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
3879 requirements
->prefersDedicatedAllocation
= false;
3880 requirements
->requiresDedicatedAllocation
= false;
3885 anv_debug_ignored_stype(ext
->sType
);
3891 void anv_GetImageMemoryRequirements(
3894 VkMemoryRequirements
* pMemoryRequirements
)
3896 ANV_FROM_HANDLE(anv_image
, image
, _image
);
3897 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3899 /* The Vulkan spec (git aaed022) says:
3901 * memoryTypeBits is a bitfield and contains one bit set for every
3902 * supported memory type for the resource. The bit `1<<i` is set if and
3903 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3904 * structure for the physical device is supported.
3906 * All types are currently supported for images.
3908 uint32_t memory_types
= (1ull << device
->physical
->memory
.type_count
) - 1;
3910 pMemoryRequirements
->size
= image
->size
;
3911 pMemoryRequirements
->alignment
= image
->alignment
;
3912 pMemoryRequirements
->memoryTypeBits
= memory_types
;
3915 void anv_GetImageMemoryRequirements2(
3917 const VkImageMemoryRequirementsInfo2
* pInfo
,
3918 VkMemoryRequirements2
* pMemoryRequirements
)
3920 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3921 ANV_FROM_HANDLE(anv_image
, image
, pInfo
->image
);
3923 anv_GetImageMemoryRequirements(_device
, pInfo
->image
,
3924 &pMemoryRequirements
->memoryRequirements
);
3926 vk_foreach_struct_const(ext
, pInfo
->pNext
) {
3927 switch (ext
->sType
) {
3928 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO
: {
3929 const VkImagePlaneMemoryRequirementsInfo
*plane_reqs
=
3930 (const VkImagePlaneMemoryRequirementsInfo
*) ext
;
3931 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
,
3932 plane_reqs
->planeAspect
);
3934 assert(image
->planes
[plane
].offset
== 0);
3936 /* The Vulkan spec (git aaed022) says:
3938 * memoryTypeBits is a bitfield and contains one bit set for every
3939 * supported memory type for the resource. The bit `1<<i` is set
3940 * if and only if the memory type `i` in the
3941 * VkPhysicalDeviceMemoryProperties structure for the physical
3942 * device is supported.
3944 * All types are currently supported for images.
3946 pMemoryRequirements
->memoryRequirements
.memoryTypeBits
=
3947 (1ull << device
->physical
->memory
.type_count
) - 1;
3949 pMemoryRequirements
->memoryRequirements
.size
= image
->planes
[plane
].size
;
3950 pMemoryRequirements
->memoryRequirements
.alignment
=
3951 image
->planes
[plane
].alignment
;
3956 anv_debug_ignored_stype(ext
->sType
);
3961 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3962 switch (ext
->sType
) {
3963 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
3964 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
3965 if (image
->needs_set_tiling
|| image
->external_format
) {
3966 /* If we need to set the tiling for external consumers, we need a
3967 * dedicated allocation.
3969 * See also anv_AllocateMemory.
3971 requirements
->prefersDedicatedAllocation
= true;
3972 requirements
->requiresDedicatedAllocation
= true;
3974 requirements
->prefersDedicatedAllocation
= false;
3975 requirements
->requiresDedicatedAllocation
= false;
3981 anv_debug_ignored_stype(ext
->sType
);
3987 void anv_GetImageSparseMemoryRequirements(
3990 uint32_t* pSparseMemoryRequirementCount
,
3991 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
3993 *pSparseMemoryRequirementCount
= 0;
3996 void anv_GetImageSparseMemoryRequirements2(
3998 const VkImageSparseMemoryRequirementsInfo2
* pInfo
,
3999 uint32_t* pSparseMemoryRequirementCount
,
4000 VkSparseImageMemoryRequirements2
* pSparseMemoryRequirements
)
4002 *pSparseMemoryRequirementCount
= 0;
4005 void anv_GetDeviceMemoryCommitment(
4007 VkDeviceMemory memory
,
4008 VkDeviceSize
* pCommittedMemoryInBytes
)
4010 *pCommittedMemoryInBytes
= 0;
4014 anv_bind_buffer_memory(const VkBindBufferMemoryInfo
*pBindInfo
)
4016 ANV_FROM_HANDLE(anv_device_memory
, mem
, pBindInfo
->memory
);
4017 ANV_FROM_HANDLE(anv_buffer
, buffer
, pBindInfo
->buffer
);
4019 assert(pBindInfo
->sType
== VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
);
4022 buffer
->address
= (struct anv_address
) {
4024 .offset
= pBindInfo
->memoryOffset
,
4027 buffer
->address
= ANV_NULL_ADDRESS
;
4031 VkResult
anv_BindBufferMemory(
4034 VkDeviceMemory memory
,
4035 VkDeviceSize memoryOffset
)
4037 anv_bind_buffer_memory(
4038 &(VkBindBufferMemoryInfo
) {
4039 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
,
4042 .memoryOffset
= memoryOffset
,
4048 VkResult
anv_BindBufferMemory2(
4050 uint32_t bindInfoCount
,
4051 const VkBindBufferMemoryInfo
* pBindInfos
)
4053 for (uint32_t i
= 0; i
< bindInfoCount
; i
++)
4054 anv_bind_buffer_memory(&pBindInfos
[i
]);
4059 VkResult
anv_QueueBindSparse(
4061 uint32_t bindInfoCount
,
4062 const VkBindSparseInfo
* pBindInfo
,
4065 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
4066 if (anv_device_is_lost(queue
->device
))
4067 return VK_ERROR_DEVICE_LOST
;
4069 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
4074 VkResult
anv_CreateEvent(
4076 const VkEventCreateInfo
* pCreateInfo
,
4077 const VkAllocationCallbacks
* pAllocator
,
4080 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4081 struct anv_state state
;
4082 struct anv_event
*event
;
4084 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_EVENT_CREATE_INFO
);
4086 state
= anv_state_pool_alloc(&device
->dynamic_state_pool
,
4089 event
->state
= state
;
4090 event
->semaphore
= VK_EVENT_RESET
;
4092 *pEvent
= anv_event_to_handle(event
);
4097 void anv_DestroyEvent(
4100 const VkAllocationCallbacks
* pAllocator
)
4102 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4103 ANV_FROM_HANDLE(anv_event
, event
, _event
);
4108 anv_state_pool_free(&device
->dynamic_state_pool
, event
->state
);
4111 VkResult
anv_GetEventStatus(
4115 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4116 ANV_FROM_HANDLE(anv_event
, event
, _event
);
4118 if (anv_device_is_lost(device
))
4119 return VK_ERROR_DEVICE_LOST
;
4121 return event
->semaphore
;
4124 VkResult
anv_SetEvent(
4128 ANV_FROM_HANDLE(anv_event
, event
, _event
);
4130 event
->semaphore
= VK_EVENT_SET
;
4135 VkResult
anv_ResetEvent(
4139 ANV_FROM_HANDLE(anv_event
, event
, _event
);
4141 event
->semaphore
= VK_EVENT_RESET
;
4148 VkResult
anv_CreateBuffer(
4150 const VkBufferCreateInfo
* pCreateInfo
,
4151 const VkAllocationCallbacks
* pAllocator
,
4154 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4155 struct anv_buffer
*buffer
;
4157 /* Don't allow creating buffers bigger than our address space. The real
4158 * issue here is that we may align up the buffer size and we don't want
4159 * doing so to cause roll-over. However, no one has any business
4160 * allocating a buffer larger than our GTT size.
4162 if (pCreateInfo
->size
> device
->physical
->gtt_size
)
4163 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
4165 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
4167 buffer
= vk_alloc2(&device
->vk
.alloc
, pAllocator
, sizeof(*buffer
), 8,
4168 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
4170 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
4172 buffer
->size
= pCreateInfo
->size
;
4173 buffer
->usage
= pCreateInfo
->usage
;
4174 buffer
->address
= ANV_NULL_ADDRESS
;
4176 *pBuffer
= anv_buffer_to_handle(buffer
);
4181 void anv_DestroyBuffer(
4184 const VkAllocationCallbacks
* pAllocator
)
4186 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4187 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
4192 vk_free2(&device
->vk
.alloc
, pAllocator
, buffer
);
4195 VkDeviceAddress
anv_GetBufferDeviceAddress(
4197 const VkBufferDeviceAddressInfoKHR
* pInfo
)
4199 ANV_FROM_HANDLE(anv_buffer
, buffer
, pInfo
->buffer
);
4201 assert(!anv_address_is_null(buffer
->address
));
4202 assert(buffer
->address
.bo
->flags
& EXEC_OBJECT_PINNED
);
4204 return anv_address_physical(buffer
->address
);
4207 uint64_t anv_GetBufferOpaqueCaptureAddress(
4209 const VkBufferDeviceAddressInfoKHR
* pInfo
)
4214 uint64_t anv_GetDeviceMemoryOpaqueCaptureAddress(
4216 const VkDeviceMemoryOpaqueCaptureAddressInfoKHR
* pInfo
)
4218 ANV_FROM_HANDLE(anv_device_memory
, memory
, pInfo
->memory
);
4220 assert(memory
->bo
->flags
& EXEC_OBJECT_PINNED
);
4221 assert(memory
->bo
->has_client_visible_address
);
4223 return gen_48b_address(memory
->bo
->offset
);
4227 anv_fill_buffer_surface_state(struct anv_device
*device
, struct anv_state state
,
4228 enum isl_format format
,
4229 struct anv_address address
,
4230 uint32_t range
, uint32_t stride
)
4232 isl_buffer_fill_state(&device
->isl_dev
, state
.map
,
4233 .address
= anv_address_physical(address
),
4234 .mocs
= device
->isl_dev
.mocs
.internal
,
4237 .swizzle
= ISL_SWIZZLE_IDENTITY
,
4238 .stride_B
= stride
);
4241 void anv_DestroySampler(
4244 const VkAllocationCallbacks
* pAllocator
)
4246 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4247 ANV_FROM_HANDLE(anv_sampler
, sampler
, _sampler
);
4252 if (sampler
->bindless_state
.map
) {
4253 anv_state_pool_free(&device
->dynamic_state_pool
,
4254 sampler
->bindless_state
);
4257 vk_free2(&device
->vk
.alloc
, pAllocator
, sampler
);
4260 VkResult
anv_CreateFramebuffer(
4262 const VkFramebufferCreateInfo
* pCreateInfo
,
4263 const VkAllocationCallbacks
* pAllocator
,
4264 VkFramebuffer
* pFramebuffer
)
4266 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4267 struct anv_framebuffer
*framebuffer
;
4269 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
4271 size_t size
= sizeof(*framebuffer
);
4273 /* VK_KHR_imageless_framebuffer extension says:
4275 * If flags includes VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR,
4276 * parameter pAttachments is ignored.
4278 if (!(pCreateInfo
->flags
& VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR
)) {
4279 size
+= sizeof(struct anv_image_view
*) * pCreateInfo
->attachmentCount
;
4280 framebuffer
= vk_alloc2(&device
->vk
.alloc
, pAllocator
, size
, 8,
4281 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
4282 if (framebuffer
== NULL
)
4283 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
4285 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
4286 ANV_FROM_HANDLE(anv_image_view
, iview
, pCreateInfo
->pAttachments
[i
]);
4287 framebuffer
->attachments
[i
] = iview
;
4289 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
4291 framebuffer
= vk_alloc2(&device
->vk
.alloc
, pAllocator
, size
, 8,
4292 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
4293 if (framebuffer
== NULL
)
4294 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
4296 framebuffer
->attachment_count
= 0;
4299 framebuffer
->width
= pCreateInfo
->width
;
4300 framebuffer
->height
= pCreateInfo
->height
;
4301 framebuffer
->layers
= pCreateInfo
->layers
;
4303 *pFramebuffer
= anv_framebuffer_to_handle(framebuffer
);
4308 void anv_DestroyFramebuffer(
4311 const VkAllocationCallbacks
* pAllocator
)
4313 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4314 ANV_FROM_HANDLE(anv_framebuffer
, fb
, _fb
);
4319 vk_free2(&device
->vk
.alloc
, pAllocator
, fb
);
4322 static const VkTimeDomainEXT anv_time_domains
[] = {
4323 VK_TIME_DOMAIN_DEVICE_EXT
,
4324 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
,
4325 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
,
4328 VkResult
anv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
4329 VkPhysicalDevice physicalDevice
,
4330 uint32_t *pTimeDomainCount
,
4331 VkTimeDomainEXT
*pTimeDomains
)
4334 VK_OUTARRAY_MAKE(out
, pTimeDomains
, pTimeDomainCount
);
4336 for (d
= 0; d
< ARRAY_SIZE(anv_time_domains
); d
++) {
4337 vk_outarray_append(&out
, i
) {
4338 *i
= anv_time_domains
[d
];
4342 return vk_outarray_status(&out
);
4346 anv_clock_gettime(clockid_t clock_id
)
4348 struct timespec current
;
4351 ret
= clock_gettime(clock_id
, ¤t
);
4352 if (ret
< 0 && clock_id
== CLOCK_MONOTONIC_RAW
)
4353 ret
= clock_gettime(CLOCK_MONOTONIC
, ¤t
);
4357 return (uint64_t) current
.tv_sec
* 1000000000ULL + current
.tv_nsec
;
4360 #define TIMESTAMP 0x2358
4362 VkResult
anv_GetCalibratedTimestampsEXT(
4364 uint32_t timestampCount
,
4365 const VkCalibratedTimestampInfoEXT
*pTimestampInfos
,
4366 uint64_t *pTimestamps
,
4367 uint64_t *pMaxDeviation
)
4369 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4370 uint64_t timestamp_frequency
= device
->info
.timestamp_frequency
;
4373 uint64_t begin
, end
;
4374 uint64_t max_clock_period
= 0;
4376 begin
= anv_clock_gettime(CLOCK_MONOTONIC_RAW
);
4378 for (d
= 0; d
< timestampCount
; d
++) {
4379 switch (pTimestampInfos
[d
].timeDomain
) {
4380 case VK_TIME_DOMAIN_DEVICE_EXT
:
4381 ret
= anv_gem_reg_read(device
, TIMESTAMP
| 1,
4385 return anv_device_set_lost(device
, "Failed to read the TIMESTAMP "
4388 uint64_t device_period
= DIV_ROUND_UP(1000000000, timestamp_frequency
);
4389 max_clock_period
= MAX2(max_clock_period
, device_period
);
4391 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
:
4392 pTimestamps
[d
] = anv_clock_gettime(CLOCK_MONOTONIC
);
4393 max_clock_period
= MAX2(max_clock_period
, 1);
4396 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
:
4397 pTimestamps
[d
] = begin
;
4405 end
= anv_clock_gettime(CLOCK_MONOTONIC_RAW
);
4408 * The maximum deviation is the sum of the interval over which we
4409 * perform the sampling and the maximum period of any sampled
4410 * clock. That's because the maximum skew between any two sampled
4411 * clock edges is when the sampled clock with the largest period is
4412 * sampled at the end of that period but right at the beginning of the
4413 * sampling interval and some other clock is sampled right at the
4414 * begining of its sampling period and right at the end of the
4415 * sampling interval. Let's assume the GPU has the longest clock
4416 * period and that the application is sampling GPU and monotonic:
4419 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
4420 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4424 * GPU -----_____-----_____-----_____-----_____
4427 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
4428 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4430 * Interval <----------------->
4431 * Deviation <-------------------------->
4435 * m = read(monotonic) 2
4438 * We round the sample interval up by one tick to cover sampling error
4439 * in the interval clock
4442 uint64_t sample_interval
= end
- begin
+ 1;
4444 *pMaxDeviation
= sample_interval
+ max_clock_period
;
4449 /* vk_icd.h does not declare this function, so we declare it here to
4450 * suppress Wmissing-prototypes.
4452 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4453 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
);
4455 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4456 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
)
4458 /* For the full details on loader interface versioning, see
4459 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
4460 * What follows is a condensed summary, to help you navigate the large and
4461 * confusing official doc.
4463 * - Loader interface v0 is incompatible with later versions. We don't
4466 * - In loader interface v1:
4467 * - The first ICD entrypoint called by the loader is
4468 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
4470 * - The ICD must statically expose no other Vulkan symbol unless it is
4471 * linked with -Bsymbolic.
4472 * - Each dispatchable Vulkan handle created by the ICD must be
4473 * a pointer to a struct whose first member is VK_LOADER_DATA. The
4474 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
4475 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
4476 * vkDestroySurfaceKHR(). The ICD must be capable of working with
4477 * such loader-managed surfaces.
4479 * - Loader interface v2 differs from v1 in:
4480 * - The first ICD entrypoint called by the loader is
4481 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
4482 * statically expose this entrypoint.
4484 * - Loader interface v3 differs from v2 in:
4485 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
4486 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
4487 * because the loader no longer does so.
4489 * - Loader interface v4 differs from v3 in:
4490 * - The ICD must implement vk_icdGetPhysicalDeviceProcAddr().
4492 *pSupportedVersion
= MIN2(*pSupportedVersion
, 4u);