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
;
75 if (list_is_empty(&device
->instance
->debug_report_callbacks
.callbacks
))
80 (void) vsnprintf(str
, MAX_DEBUG_MESSAGE_LENGTH
, fmt
, args
);
83 vk_debug_report(&device
->instance
->debug_report_callbacks
,
84 VK_DEBUG_REPORT_DEBUG_BIT_EXT
,
85 VK_DEBUG_REPORT_OBJECT_TYPE_UNKNOWN_EXT
,
90 compiler_perf_log(void *data
, const char *fmt
, ...)
95 if (unlikely(INTEL_DEBUG
& DEBUG_PERF
))
96 intel_logd_v(fmt
, args
);
102 anv_compute_heap_size(int fd
, uint64_t gtt_size
)
104 /* Query the total ram from the system */
108 uint64_t total_ram
= (uint64_t)info
.totalram
* (uint64_t)info
.mem_unit
;
110 /* We don't want to burn too much ram with the GPU. If the user has 4GiB
111 * or less, we use at most half. If they have more than 4GiB, we use 3/4.
113 uint64_t available_ram
;
114 if (total_ram
<= 4ull * 1024ull * 1024ull * 1024ull)
115 available_ram
= total_ram
/ 2;
117 available_ram
= total_ram
* 3 / 4;
119 /* We also want to leave some padding for things we allocate in the driver,
120 * so don't go over 3/4 of the GTT either.
122 uint64_t available_gtt
= gtt_size
* 3 / 4;
124 return MIN2(available_ram
, available_gtt
);
128 anv_physical_device_init_heaps(struct anv_physical_device
*device
, int fd
)
130 if (anv_gem_get_context_param(fd
, 0, I915_CONTEXT_PARAM_GTT_SIZE
,
131 &device
->gtt_size
) == -1) {
132 /* If, for whatever reason, we can't actually get the GTT size from the
133 * kernel (too old?) fall back to the aperture size.
135 anv_perf_warn(NULL
, NULL
,
136 "Failed to get I915_CONTEXT_PARAM_GTT_SIZE: %m");
138 if (anv_gem_get_aperture(fd
, &device
->gtt_size
) == -1) {
139 return vk_errorf(NULL
, NULL
, VK_ERROR_INITIALIZATION_FAILED
,
140 "failed to get aperture size: %m");
144 /* We only allow 48-bit addresses with softpin because knowing the actual
145 * address is required for the vertex cache flush workaround.
147 device
->supports_48bit_addresses
= (device
->info
.gen
>= 8) &&
148 device
->has_softpin
&&
149 device
->gtt_size
> (4ULL << 30 /* GiB */);
151 uint64_t heap_size
= anv_compute_heap_size(fd
, device
->gtt_size
);
153 if (heap_size
> (2ull << 30) && !device
->supports_48bit_addresses
) {
154 /* When running with an overridden PCI ID, we may get a GTT size from
155 * the kernel that is greater than 2 GiB but the execbuf check for 48bit
156 * address support can still fail. Just clamp the address space size to
157 * 2 GiB if we don't have 48-bit support.
159 intel_logw("%s:%d: The kernel reported a GTT size larger than 2 GiB but "
160 "not support for 48-bit addresses",
162 heap_size
= 2ull << 30;
165 device
->memory
.heap_count
= 1;
166 device
->memory
.heaps
[0] = (struct anv_memory_heap
) {
168 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
171 uint32_t type_count
= 0;
172 for (uint32_t heap
= 0; heap
< device
->memory
.heap_count
; heap
++) {
173 if (device
->info
.has_llc
) {
174 /* Big core GPUs share LLC with the CPU and thus one memory type can be
175 * both cached and coherent at the same time.
177 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
178 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
179 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
180 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
181 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
185 /* The spec requires that we expose a host-visible, coherent memory
186 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
187 * to give the application a choice between cached, but not coherent and
188 * coherent but uncached (WC though).
190 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
191 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
192 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
193 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
196 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
197 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
198 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
199 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
204 device
->memory
.type_count
= type_count
;
210 anv_physical_device_init_uuids(struct anv_physical_device
*device
)
212 const struct build_id_note
*note
=
213 build_id_find_nhdr_for_addr(anv_physical_device_init_uuids
);
215 return vk_errorf(device
->instance
, device
,
216 VK_ERROR_INITIALIZATION_FAILED
,
217 "Failed to find build-id");
220 unsigned build_id_len
= build_id_length(note
);
221 if (build_id_len
< 20) {
222 return vk_errorf(device
->instance
, device
,
223 VK_ERROR_INITIALIZATION_FAILED
,
224 "build-id too short. It needs to be a SHA");
227 memcpy(device
->driver_build_sha1
, build_id_data(note
), 20);
229 struct mesa_sha1 sha1_ctx
;
231 STATIC_ASSERT(VK_UUID_SIZE
<= sizeof(sha1
));
233 /* The pipeline cache UUID is used for determining when a pipeline cache is
234 * invalid. It needs both a driver build and the PCI ID of the device.
236 _mesa_sha1_init(&sha1_ctx
);
237 _mesa_sha1_update(&sha1_ctx
, build_id_data(note
), build_id_len
);
238 _mesa_sha1_update(&sha1_ctx
, &device
->chipset_id
,
239 sizeof(device
->chipset_id
));
240 _mesa_sha1_update(&sha1_ctx
, &device
->always_use_bindless
,
241 sizeof(device
->always_use_bindless
));
242 _mesa_sha1_update(&sha1_ctx
, &device
->has_a64_buffer_access
,
243 sizeof(device
->has_a64_buffer_access
));
244 _mesa_sha1_update(&sha1_ctx
, &device
->has_bindless_images
,
245 sizeof(device
->has_bindless_images
));
246 _mesa_sha1_update(&sha1_ctx
, &device
->has_bindless_samplers
,
247 sizeof(device
->has_bindless_samplers
));
248 _mesa_sha1_final(&sha1_ctx
, sha1
);
249 memcpy(device
->pipeline_cache_uuid
, sha1
, VK_UUID_SIZE
);
251 /* The driver UUID is used for determining sharability of images and memory
252 * between two Vulkan instances in separate processes. People who want to
253 * share memory need to also check the device UUID (below) so all this
254 * needs to be is the build-id.
256 memcpy(device
->driver_uuid
, build_id_data(note
), VK_UUID_SIZE
);
258 /* The device UUID uniquely identifies the given device within the machine.
259 * Since we never have more than one device, this doesn't need to be a real
260 * UUID. However, on the off-chance that someone tries to use this to
261 * cache pre-tiled images or something of the like, we use the PCI ID and
262 * some bits of ISL info to ensure that this is safe.
264 _mesa_sha1_init(&sha1_ctx
);
265 _mesa_sha1_update(&sha1_ctx
, &device
->chipset_id
,
266 sizeof(device
->chipset_id
));
267 _mesa_sha1_update(&sha1_ctx
, &device
->isl_dev
.has_bit6_swizzling
,
268 sizeof(device
->isl_dev
.has_bit6_swizzling
));
269 _mesa_sha1_final(&sha1_ctx
, sha1
);
270 memcpy(device
->device_uuid
, sha1
, VK_UUID_SIZE
);
276 anv_physical_device_init_disk_cache(struct anv_physical_device
*device
)
278 #ifdef ENABLE_SHADER_CACHE
280 ASSERTED
int len
= snprintf(renderer
, sizeof(renderer
), "anv_%04x",
282 assert(len
== sizeof(renderer
) - 2);
285 _mesa_sha1_format(timestamp
, device
->driver_build_sha1
);
287 const uint64_t driver_flags
=
288 brw_get_compiler_config_value(device
->compiler
);
289 device
->disk_cache
= disk_cache_create(renderer
, timestamp
, driver_flags
);
291 device
->disk_cache
= NULL
;
296 anv_physical_device_free_disk_cache(struct anv_physical_device
*device
)
298 #ifdef ENABLE_SHADER_CACHE
299 if (device
->disk_cache
)
300 disk_cache_destroy(device
->disk_cache
);
302 assert(device
->disk_cache
== NULL
);
307 get_available_system_memory()
309 char *meminfo
= os_read_file("/proc/meminfo");
313 char *str
= strstr(meminfo
, "MemAvailable:");
319 uint64_t kb_mem_available
;
320 if (sscanf(str
, "MemAvailable: %" PRIx64
, &kb_mem_available
) == 1) {
322 return kb_mem_available
<< 10;
330 anv_physical_device_init(struct anv_physical_device
*device
,
331 struct anv_instance
*instance
,
332 drmDevicePtr drm_device
)
334 const char *primary_path
= drm_device
->nodes
[DRM_NODE_PRIMARY
];
335 const char *path
= drm_device
->nodes
[DRM_NODE_RENDER
];
340 brw_process_intel_debug_variable();
342 fd
= open(path
, O_RDWR
| O_CLOEXEC
);
344 return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
346 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
347 device
->instance
= instance
;
349 assert(strlen(path
) < ARRAY_SIZE(device
->path
));
350 snprintf(device
->path
, ARRAY_SIZE(device
->path
), "%s", path
);
352 if (!gen_get_device_info_from_fd(fd
, &device
->info
)) {
353 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
356 device
->chipset_id
= device
->info
.chipset_id
;
357 device
->no_hw
= device
->info
.no_hw
;
359 if (getenv("INTEL_NO_HW") != NULL
)
360 device
->no_hw
= true;
362 device
->pci_info
.domain
= drm_device
->businfo
.pci
->domain
;
363 device
->pci_info
.bus
= drm_device
->businfo
.pci
->bus
;
364 device
->pci_info
.device
= drm_device
->businfo
.pci
->dev
;
365 device
->pci_info
.function
= drm_device
->businfo
.pci
->func
;
367 device
->name
= gen_get_device_name(device
->chipset_id
);
369 if (device
->info
.is_haswell
) {
370 intel_logw("Haswell Vulkan support is incomplete");
371 } else if (device
->info
.gen
== 7 && !device
->info
.is_baytrail
) {
372 intel_logw("Ivy Bridge Vulkan support is incomplete");
373 } else if (device
->info
.gen
== 7 && device
->info
.is_baytrail
) {
374 intel_logw("Bay Trail Vulkan support is incomplete");
375 } else if (device
->info
.gen
>= 8 && device
->info
.gen
<= 11) {
376 /* Gen8-11 fully supported */
377 } else if (device
->info
.gen
== 12) {
378 intel_logw("Vulkan is not yet fully supported on gen12");
380 result
= vk_errorf(device
->instance
, device
,
381 VK_ERROR_INCOMPATIBLE_DRIVER
,
382 "Vulkan not yet supported on %s", device
->name
);
386 device
->cmd_parser_version
= -1;
387 if (device
->info
.gen
== 7) {
388 device
->cmd_parser_version
=
389 anv_gem_get_param(fd
, I915_PARAM_CMD_PARSER_VERSION
);
390 if (device
->cmd_parser_version
== -1) {
391 result
= vk_errorf(device
->instance
, device
,
392 VK_ERROR_INITIALIZATION_FAILED
,
393 "failed to get command parser version");
398 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_WAIT_TIMEOUT
)) {
399 result
= vk_errorf(device
->instance
, device
,
400 VK_ERROR_INITIALIZATION_FAILED
,
401 "kernel missing gem wait");
405 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_EXECBUF2
)) {
406 result
= vk_errorf(device
->instance
, device
,
407 VK_ERROR_INITIALIZATION_FAILED
,
408 "kernel missing execbuf2");
412 if (!device
->info
.has_llc
&&
413 anv_gem_get_param(fd
, I915_PARAM_MMAP_VERSION
) < 1) {
414 result
= vk_errorf(device
->instance
, device
,
415 VK_ERROR_INITIALIZATION_FAILED
,
416 "kernel missing wc mmap");
420 device
->has_softpin
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_SOFTPIN
);
421 device
->has_exec_async
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_ASYNC
);
422 device
->has_exec_capture
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_CAPTURE
);
423 device
->has_exec_fence
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE
);
424 device
->has_syncobj
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE_ARRAY
);
425 device
->has_syncobj_wait
= device
->has_syncobj
&&
426 anv_gem_supports_syncobj_wait(fd
);
427 device
->has_context_priority
= anv_gem_has_context_priority(fd
);
429 result
= anv_physical_device_init_heaps(device
, fd
);
430 if (result
!= VK_SUCCESS
)
433 device
->use_softpin
= device
->has_softpin
&&
434 device
->supports_48bit_addresses
;
436 device
->has_context_isolation
=
437 anv_gem_get_param(fd
, I915_PARAM_HAS_CONTEXT_ISOLATION
);
439 device
->always_use_bindless
=
440 env_var_as_boolean("ANV_ALWAYS_BINDLESS", false);
442 /* We first got the A64 messages on broadwell and we can only use them if
443 * we can pass addresses directly into the shader which requires softpin.
445 device
->has_a64_buffer_access
= device
->info
.gen
>= 8 &&
448 /* We first get bindless image access on Skylake and we can only really do
449 * it if we don't have any relocations so we need softpin.
451 device
->has_bindless_images
= device
->info
.gen
>= 9 &&
454 /* We've had bindless samplers since Ivy Bridge (forever in Vulkan terms)
455 * because it's just a matter of setting the sampler address in the sample
456 * message header. However, we've not bothered to wire it up for vec4 so
457 * we leave it disabled on gen7.
459 device
->has_bindless_samplers
= device
->info
.gen
>= 8;
461 device
->has_mem_available
= get_available_system_memory() != 0;
463 device
->always_flush_cache
=
464 driQueryOptionb(&instance
->dri_options
, "always_flush_cache");
466 /* Starting with Gen10, the timestamp frequency of the command streamer may
467 * vary from one part to another. We can query the value from the kernel.
469 if (device
->info
.gen
>= 10) {
470 int timestamp_frequency
=
471 anv_gem_get_param(fd
, I915_PARAM_CS_TIMESTAMP_FREQUENCY
);
473 if (timestamp_frequency
< 0)
474 intel_logw("Kernel 4.16-rc1+ required to properly query CS timestamp frequency");
476 device
->info
.timestamp_frequency
= timestamp_frequency
;
479 /* GENs prior to 8 do not support EU/Subslice info */
480 if (device
->info
.gen
>= 8) {
481 device
->subslice_total
= anv_gem_get_param(fd
, I915_PARAM_SUBSLICE_TOTAL
);
482 device
->eu_total
= anv_gem_get_param(fd
, I915_PARAM_EU_TOTAL
);
484 /* Without this information, we cannot get the right Braswell
485 * brandstrings, and we have to use conservative numbers for GPGPU on
486 * many platforms, but otherwise, things will just work.
488 if (device
->subslice_total
< 1 || device
->eu_total
< 1) {
489 intel_logw("Kernel 4.1 required to properly query GPU properties");
491 } else if (device
->info
.gen
== 7) {
492 device
->subslice_total
= 1 << (device
->info
.gt
- 1);
495 if (device
->info
.is_cherryview
&&
496 device
->subslice_total
> 0 && device
->eu_total
> 0) {
497 /* Logical CS threads = EUs per subslice * num threads per EU */
498 uint32_t max_cs_threads
=
499 device
->eu_total
/ device
->subslice_total
* device
->info
.num_thread_per_eu
;
501 /* Fuse configurations may give more threads than expected, never less. */
502 if (max_cs_threads
> device
->info
.max_cs_threads
)
503 device
->info
.max_cs_threads
= max_cs_threads
;
506 device
->compiler
= brw_compiler_create(NULL
, &device
->info
);
507 if (device
->compiler
== NULL
) {
508 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
511 device
->compiler
->shader_debug_log
= compiler_debug_log
;
512 device
->compiler
->shader_perf_log
= compiler_perf_log
;
513 device
->compiler
->supports_pull_constants
= false;
514 device
->compiler
->constant_buffer_0_is_relative
=
515 device
->info
.gen
< 8 || !device
->has_context_isolation
;
516 device
->compiler
->supports_shader_constants
= true;
517 device
->compiler
->compact_params
= false;
519 /* Broadwell PRM says:
521 * "Before Gen8, there was a historical configuration control field to
522 * swizzle address bit[6] for in X/Y tiling modes. This was set in three
523 * different places: TILECTL[1:0], ARB_MODE[5:4], and
524 * DISP_ARB_CTL[14:13].
526 * For Gen8 and subsequent generations, the swizzle fields are all
527 * reserved, and the CPU's memory controller performs all address
528 * swizzling modifications."
531 device
->info
.gen
< 8 && anv_gem_get_bit6_swizzle(fd
, I915_TILING_X
);
533 isl_device_init(&device
->isl_dev
, &device
->info
, swizzled
);
535 result
= anv_physical_device_init_uuids(device
);
536 if (result
!= VK_SUCCESS
)
539 anv_physical_device_init_disk_cache(device
);
541 if (instance
->enabled_extensions
.KHR_display
) {
542 master_fd
= open(primary_path
, O_RDWR
| O_CLOEXEC
);
543 if (master_fd
>= 0) {
544 /* prod the device with a GETPARAM call which will fail if
545 * we don't have permission to even render on this device
547 if (anv_gem_get_param(master_fd
, I915_PARAM_CHIPSET_ID
) == 0) {
553 device
->master_fd
= master_fd
;
555 result
= anv_init_wsi(device
);
556 if (result
!= VK_SUCCESS
) {
557 ralloc_free(device
->compiler
);
558 anv_physical_device_free_disk_cache(device
);
562 device
->perf
= anv_get_perf(&device
->info
, fd
);
564 anv_physical_device_get_supported_extensions(device
,
565 &device
->supported_extensions
);
568 device
->local_fd
= fd
;
580 anv_physical_device_finish(struct anv_physical_device
*device
)
582 anv_finish_wsi(device
);
583 anv_physical_device_free_disk_cache(device
);
584 ralloc_free(device
->compiler
);
585 ralloc_free(device
->perf
);
586 close(device
->local_fd
);
587 if (device
->master_fd
>= 0)
588 close(device
->master_fd
);
592 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
593 VkSystemAllocationScope allocationScope
)
599 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
600 size_t align
, VkSystemAllocationScope allocationScope
)
602 return realloc(pOriginal
, size
);
606 default_free_func(void *pUserData
, void *pMemory
)
611 static const VkAllocationCallbacks default_alloc
= {
613 .pfnAllocation
= default_alloc_func
,
614 .pfnReallocation
= default_realloc_func
,
615 .pfnFree
= default_free_func
,
618 VkResult
anv_EnumerateInstanceExtensionProperties(
619 const char* pLayerName
,
620 uint32_t* pPropertyCount
,
621 VkExtensionProperties
* pProperties
)
623 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
625 for (int i
= 0; i
< ANV_INSTANCE_EXTENSION_COUNT
; i
++) {
626 if (anv_instance_extensions_supported
.extensions
[i
]) {
627 vk_outarray_append(&out
, prop
) {
628 *prop
= anv_instance_extensions
[i
];
633 return vk_outarray_status(&out
);
636 VkResult
anv_CreateInstance(
637 const VkInstanceCreateInfo
* pCreateInfo
,
638 const VkAllocationCallbacks
* pAllocator
,
639 VkInstance
* pInstance
)
641 struct anv_instance
*instance
;
644 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
646 struct anv_instance_extension_table enabled_extensions
= {};
647 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
649 for (idx
= 0; idx
< ANV_INSTANCE_EXTENSION_COUNT
; idx
++) {
650 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
651 anv_instance_extensions
[idx
].extensionName
) == 0)
655 if (idx
>= ANV_INSTANCE_EXTENSION_COUNT
)
656 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
658 if (!anv_instance_extensions_supported
.extensions
[idx
])
659 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
661 enabled_extensions
.extensions
[idx
] = true;
664 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
665 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
667 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
669 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
672 instance
->alloc
= *pAllocator
;
674 instance
->alloc
= default_alloc
;
676 instance
->app_info
= (struct anv_app_info
) { .api_version
= 0 };
677 if (pCreateInfo
->pApplicationInfo
) {
678 const VkApplicationInfo
*app
= pCreateInfo
->pApplicationInfo
;
680 instance
->app_info
.app_name
=
681 vk_strdup(&instance
->alloc
, app
->pApplicationName
,
682 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
683 instance
->app_info
.app_version
= app
->applicationVersion
;
685 instance
->app_info
.engine_name
=
686 vk_strdup(&instance
->alloc
, app
->pEngineName
,
687 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
688 instance
->app_info
.engine_version
= app
->engineVersion
;
690 instance
->app_info
.api_version
= app
->apiVersion
;
693 if (instance
->app_info
.api_version
== 0)
694 instance
->app_info
.api_version
= VK_API_VERSION_1_0
;
696 instance
->enabled_extensions
= enabled_extensions
;
698 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->dispatch
.entrypoints
); i
++) {
699 /* Vulkan requires that entrypoints for extensions which have not been
700 * enabled must not be advertised.
702 if (!anv_instance_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
703 &instance
->enabled_extensions
)) {
704 instance
->dispatch
.entrypoints
[i
] = NULL
;
706 instance
->dispatch
.entrypoints
[i
] =
707 anv_instance_dispatch_table
.entrypoints
[i
];
711 struct anv_physical_device
*pdevice
= &instance
->physicalDevice
;
712 for (unsigned i
= 0; i
< ARRAY_SIZE(pdevice
->dispatch
.entrypoints
); i
++) {
713 /* Vulkan requires that entrypoints for extensions which have not been
714 * enabled must not be advertised.
716 if (!anv_physical_device_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
717 &instance
->enabled_extensions
)) {
718 pdevice
->dispatch
.entrypoints
[i
] = NULL
;
720 pdevice
->dispatch
.entrypoints
[i
] =
721 anv_physical_device_dispatch_table
.entrypoints
[i
];
725 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->device_dispatch
.entrypoints
); i
++) {
726 /* Vulkan requires that entrypoints for extensions which have not been
727 * enabled must not be advertised.
729 if (!anv_device_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
730 &instance
->enabled_extensions
, NULL
)) {
731 instance
->device_dispatch
.entrypoints
[i
] = NULL
;
733 instance
->device_dispatch
.entrypoints
[i
] =
734 anv_device_dispatch_table
.entrypoints
[i
];
738 instance
->physicalDeviceCount
= -1;
740 result
= vk_debug_report_instance_init(&instance
->debug_report_callbacks
);
741 if (result
!= VK_SUCCESS
) {
742 vk_free2(&default_alloc
, pAllocator
, instance
);
743 return vk_error(result
);
746 instance
->pipeline_cache_enabled
=
747 env_var_as_boolean("ANV_ENABLE_PIPELINE_CACHE", true);
749 glsl_type_singleton_init_or_ref();
751 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
753 driParseOptionInfo(&instance
->available_dri_options
, anv_dri_options_xml
);
754 driParseConfigFiles(&instance
->dri_options
, &instance
->available_dri_options
,
756 instance
->app_info
.engine_name
,
757 instance
->app_info
.engine_version
);
759 *pInstance
= anv_instance_to_handle(instance
);
764 void anv_DestroyInstance(
765 VkInstance _instance
,
766 const VkAllocationCallbacks
* pAllocator
)
768 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
773 if (instance
->physicalDeviceCount
> 0) {
774 /* We support at most one physical device. */
775 assert(instance
->physicalDeviceCount
== 1);
776 anv_physical_device_finish(&instance
->physicalDevice
);
779 vk_free(&instance
->alloc
, (char *)instance
->app_info
.app_name
);
780 vk_free(&instance
->alloc
, (char *)instance
->app_info
.engine_name
);
782 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
784 vk_debug_report_instance_destroy(&instance
->debug_report_callbacks
);
786 glsl_type_singleton_decref();
788 driDestroyOptionCache(&instance
->dri_options
);
789 driDestroyOptionInfo(&instance
->available_dri_options
);
791 vk_free(&instance
->alloc
, instance
);
795 anv_enumerate_devices(struct anv_instance
*instance
)
797 /* TODO: Check for more devices ? */
798 drmDevicePtr devices
[8];
799 VkResult result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
802 instance
->physicalDeviceCount
= 0;
804 max_devices
= drmGetDevices2(0, devices
, ARRAY_SIZE(devices
));
806 return VK_ERROR_INCOMPATIBLE_DRIVER
;
808 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
809 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
810 devices
[i
]->bustype
== DRM_BUS_PCI
&&
811 devices
[i
]->deviceinfo
.pci
->vendor_id
== 0x8086) {
813 result
= anv_physical_device_init(&instance
->physicalDevice
,
814 instance
, devices
[i
]);
815 if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
819 drmFreeDevices(devices
, max_devices
);
821 if (result
== VK_SUCCESS
)
822 instance
->physicalDeviceCount
= 1;
828 anv_instance_ensure_physical_device(struct anv_instance
*instance
)
830 if (instance
->physicalDeviceCount
< 0) {
831 VkResult result
= anv_enumerate_devices(instance
);
832 if (result
!= VK_SUCCESS
&&
833 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
840 VkResult
anv_EnumeratePhysicalDevices(
841 VkInstance _instance
,
842 uint32_t* pPhysicalDeviceCount
,
843 VkPhysicalDevice
* pPhysicalDevices
)
845 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
846 VK_OUTARRAY_MAKE(out
, pPhysicalDevices
, pPhysicalDeviceCount
);
848 VkResult result
= anv_instance_ensure_physical_device(instance
);
849 if (result
!= VK_SUCCESS
)
852 if (instance
->physicalDeviceCount
== 0)
855 assert(instance
->physicalDeviceCount
== 1);
856 vk_outarray_append(&out
, i
) {
857 *i
= anv_physical_device_to_handle(&instance
->physicalDevice
);
860 return vk_outarray_status(&out
);
863 VkResult
anv_EnumeratePhysicalDeviceGroups(
864 VkInstance _instance
,
865 uint32_t* pPhysicalDeviceGroupCount
,
866 VkPhysicalDeviceGroupProperties
* pPhysicalDeviceGroupProperties
)
868 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
869 VK_OUTARRAY_MAKE(out
, pPhysicalDeviceGroupProperties
,
870 pPhysicalDeviceGroupCount
);
872 VkResult result
= anv_instance_ensure_physical_device(instance
);
873 if (result
!= VK_SUCCESS
)
876 if (instance
->physicalDeviceCount
== 0)
879 assert(instance
->physicalDeviceCount
== 1);
881 vk_outarray_append(&out
, p
) {
882 p
->physicalDeviceCount
= 1;
883 memset(p
->physicalDevices
, 0, sizeof(p
->physicalDevices
));
884 p
->physicalDevices
[0] =
885 anv_physical_device_to_handle(&instance
->physicalDevice
);
886 p
->subsetAllocation
= false;
888 vk_foreach_struct(ext
, p
->pNext
)
889 anv_debug_ignored_stype(ext
->sType
);
892 return vk_outarray_status(&out
);
895 void anv_GetPhysicalDeviceFeatures(
896 VkPhysicalDevice physicalDevice
,
897 VkPhysicalDeviceFeatures
* pFeatures
)
899 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
901 *pFeatures
= (VkPhysicalDeviceFeatures
) {
902 .robustBufferAccess
= true,
903 .fullDrawIndexUint32
= true,
904 .imageCubeArray
= true,
905 .independentBlend
= true,
906 .geometryShader
= true,
907 .tessellationShader
= true,
908 .sampleRateShading
= true,
909 .dualSrcBlend
= true,
911 .multiDrawIndirect
= true,
912 .drawIndirectFirstInstance
= true,
914 .depthBiasClamp
= true,
915 .fillModeNonSolid
= true,
916 .depthBounds
= pdevice
->info
.gen
>= 12,
920 .multiViewport
= true,
921 .samplerAnisotropy
= true,
922 .textureCompressionETC2
= pdevice
->info
.gen
>= 8 ||
923 pdevice
->info
.is_baytrail
,
924 .textureCompressionASTC_LDR
= pdevice
->info
.gen
>= 9, /* FINISHME CHV */
925 .textureCompressionBC
= true,
926 .occlusionQueryPrecise
= true,
927 .pipelineStatisticsQuery
= true,
928 .fragmentStoresAndAtomics
= true,
929 .shaderTessellationAndGeometryPointSize
= true,
930 .shaderImageGatherExtended
= true,
931 .shaderStorageImageExtendedFormats
= true,
932 .shaderStorageImageMultisample
= false,
933 .shaderStorageImageReadWithoutFormat
= false,
934 .shaderStorageImageWriteWithoutFormat
= true,
935 .shaderUniformBufferArrayDynamicIndexing
= true,
936 .shaderSampledImageArrayDynamicIndexing
= true,
937 .shaderStorageBufferArrayDynamicIndexing
= true,
938 .shaderStorageImageArrayDynamicIndexing
= true,
939 .shaderClipDistance
= true,
940 .shaderCullDistance
= true,
941 .shaderFloat64
= pdevice
->info
.gen
>= 8 &&
942 pdevice
->info
.has_64bit_types
,
943 .shaderInt64
= pdevice
->info
.gen
>= 8 &&
944 pdevice
->info
.has_64bit_types
,
945 .shaderInt16
= pdevice
->info
.gen
>= 8,
946 .shaderResourceMinLod
= pdevice
->info
.gen
>= 9,
947 .variableMultisampleRate
= true,
948 .inheritedQueries
= true,
951 /* We can't do image stores in vec4 shaders */
952 pFeatures
->vertexPipelineStoresAndAtomics
=
953 pdevice
->compiler
->scalar_stage
[MESA_SHADER_VERTEX
] &&
954 pdevice
->compiler
->scalar_stage
[MESA_SHADER_GEOMETRY
];
956 struct anv_app_info
*app_info
= &pdevice
->instance
->app_info
;
958 /* The new DOOM and Wolfenstein games require depthBounds without
959 * checking for it. They seem to run fine without it so just claim it's
960 * there and accept the consequences.
962 if (app_info
->engine_name
&& strcmp(app_info
->engine_name
, "idTech") == 0)
963 pFeatures
->depthBounds
= true;
967 anv_get_physical_device_features_1_1(struct anv_physical_device
*pdevice
,
968 VkPhysicalDeviceVulkan11Features
*f
)
970 assert(f
->sType
== VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES
);
972 f
->storageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
973 f
->uniformAndStorageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
974 f
->storagePushConstant16
= pdevice
->info
.gen
>= 8;
975 f
->storageInputOutput16
= false;
977 f
->multiviewGeometryShader
= true;
978 f
->multiviewTessellationShader
= true;
979 f
->variablePointersStorageBuffer
= true;
980 f
->variablePointers
= true;
981 f
->protectedMemory
= false;
982 f
->samplerYcbcrConversion
= true;
983 f
->shaderDrawParameters
= true;
987 anv_get_physical_device_features_1_2(struct anv_physical_device
*pdevice
,
988 VkPhysicalDeviceVulkan12Features
*f
)
990 assert(f
->sType
== VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES
);
992 f
->samplerMirrorClampToEdge
= true;
993 f
->drawIndirectCount
= true;
994 f
->storageBuffer8BitAccess
= pdevice
->info
.gen
>= 8;
995 f
->uniformAndStorageBuffer8BitAccess
= pdevice
->info
.gen
>= 8;
996 f
->storagePushConstant8
= pdevice
->info
.gen
>= 8;
997 f
->shaderBufferInt64Atomics
= pdevice
->info
.gen
>= 9 &&
998 pdevice
->use_softpin
;
999 f
->shaderSharedInt64Atomics
= false;
1000 f
->shaderFloat16
= pdevice
->info
.gen
>= 8;
1001 f
->shaderInt8
= pdevice
->info
.gen
>= 8;
1003 bool descIndexing
= pdevice
->has_a64_buffer_access
&&
1004 pdevice
->has_bindless_images
;
1005 f
->descriptorIndexing
= descIndexing
;
1006 f
->shaderInputAttachmentArrayDynamicIndexing
= false;
1007 f
->shaderUniformTexelBufferArrayDynamicIndexing
= descIndexing
;
1008 f
->shaderStorageTexelBufferArrayDynamicIndexing
= descIndexing
;
1009 f
->shaderUniformBufferArrayNonUniformIndexing
= false;
1010 f
->shaderSampledImageArrayNonUniformIndexing
= descIndexing
;
1011 f
->shaderStorageBufferArrayNonUniformIndexing
= descIndexing
;
1012 f
->shaderStorageImageArrayNonUniformIndexing
= descIndexing
;
1013 f
->shaderInputAttachmentArrayNonUniformIndexing
= false;
1014 f
->shaderUniformTexelBufferArrayNonUniformIndexing
= descIndexing
;
1015 f
->shaderStorageTexelBufferArrayNonUniformIndexing
= descIndexing
;
1016 f
->descriptorBindingUniformBufferUpdateAfterBind
= false;
1017 f
->descriptorBindingSampledImageUpdateAfterBind
= descIndexing
;
1018 f
->descriptorBindingStorageImageUpdateAfterBind
= descIndexing
;
1019 f
->descriptorBindingStorageBufferUpdateAfterBind
= descIndexing
;
1020 f
->descriptorBindingUniformTexelBufferUpdateAfterBind
= descIndexing
;
1021 f
->descriptorBindingStorageTexelBufferUpdateAfterBind
= descIndexing
;
1022 f
->descriptorBindingUpdateUnusedWhilePending
= descIndexing
;
1023 f
->descriptorBindingPartiallyBound
= descIndexing
;
1024 f
->descriptorBindingVariableDescriptorCount
= false;
1025 f
->runtimeDescriptorArray
= descIndexing
;
1027 f
->samplerFilterMinmax
= pdevice
->info
.gen
>= 9;
1028 f
->scalarBlockLayout
= true;
1029 f
->imagelessFramebuffer
= true;
1030 f
->uniformBufferStandardLayout
= true;
1031 f
->shaderSubgroupExtendedTypes
= true;
1032 f
->separateDepthStencilLayouts
= true;
1033 f
->hostQueryReset
= true;
1034 f
->timelineSemaphore
= true;
1035 f
->bufferDeviceAddress
= pdevice
->has_a64_buffer_access
;
1036 f
->bufferDeviceAddressCaptureReplay
= pdevice
->has_a64_buffer_access
;
1037 f
->bufferDeviceAddressMultiDevice
= false;
1038 f
->vulkanMemoryModel
= true;
1039 f
->vulkanMemoryModelDeviceScope
= true;
1040 f
->vulkanMemoryModelAvailabilityVisibilityChains
= true;
1041 f
->shaderOutputViewportIndex
= true;
1042 f
->shaderOutputLayer
= true;
1043 f
->subgroupBroadcastDynamicId
= true;
1046 void anv_GetPhysicalDeviceFeatures2(
1047 VkPhysicalDevice physicalDevice
,
1048 VkPhysicalDeviceFeatures2
* pFeatures
)
1050 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1051 anv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
1053 VkPhysicalDeviceVulkan11Features core_1_1
= {
1054 .sType
= VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES
,
1056 anv_get_physical_device_features_1_1(pdevice
, &core_1_1
);
1058 VkPhysicalDeviceVulkan12Features core_1_2
= {
1059 .sType
= VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES
,
1061 anv_get_physical_device_features_1_2(pdevice
, &core_1_2
);
1063 #define CORE_FEATURE(major, minor, feature) \
1064 features->feature = core_##major##_##minor.feature
1067 vk_foreach_struct(ext
, pFeatures
->pNext
) {
1068 switch (ext
->sType
) {
1069 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES_KHR
: {
1070 VkPhysicalDevice8BitStorageFeaturesKHR
*features
=
1071 (VkPhysicalDevice8BitStorageFeaturesKHR
*)ext
;
1072 CORE_FEATURE(1, 2, storageBuffer8BitAccess
);
1073 CORE_FEATURE(1, 2, uniformAndStorageBuffer8BitAccess
);
1074 CORE_FEATURE(1, 2, storagePushConstant8
);
1078 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES
: {
1079 VkPhysicalDevice16BitStorageFeatures
*features
=
1080 (VkPhysicalDevice16BitStorageFeatures
*)ext
;
1081 CORE_FEATURE(1, 1, storageBuffer16BitAccess
);
1082 CORE_FEATURE(1, 1, uniformAndStorageBuffer16BitAccess
);
1083 CORE_FEATURE(1, 1, storagePushConstant16
);
1084 CORE_FEATURE(1, 1, storageInputOutput16
);
1088 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_EXT
: {
1089 VkPhysicalDeviceBufferDeviceAddressFeaturesEXT
*features
= (void *)ext
;
1090 features
->bufferDeviceAddress
= pdevice
->has_a64_buffer_access
;
1091 features
->bufferDeviceAddressCaptureReplay
= false;
1092 features
->bufferDeviceAddressMultiDevice
= false;
1096 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_KHR
: {
1097 VkPhysicalDeviceBufferDeviceAddressFeaturesKHR
*features
= (void *)ext
;
1098 CORE_FEATURE(1, 2, bufferDeviceAddress
);
1099 CORE_FEATURE(1, 2, bufferDeviceAddressCaptureReplay
);
1100 CORE_FEATURE(1, 2, bufferDeviceAddressMultiDevice
);
1104 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COMPUTE_SHADER_DERIVATIVES_FEATURES_NV
: {
1105 VkPhysicalDeviceComputeShaderDerivativesFeaturesNV
*features
=
1106 (VkPhysicalDeviceComputeShaderDerivativesFeaturesNV
*)ext
;
1107 features
->computeDerivativeGroupQuads
= true;
1108 features
->computeDerivativeGroupLinear
= true;
1112 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT
: {
1113 VkPhysicalDeviceConditionalRenderingFeaturesEXT
*features
=
1114 (VkPhysicalDeviceConditionalRenderingFeaturesEXT
*)ext
;
1115 features
->conditionalRendering
= pdevice
->info
.gen
>= 8 ||
1116 pdevice
->info
.is_haswell
;
1117 features
->inheritedConditionalRendering
= pdevice
->info
.gen
>= 8 ||
1118 pdevice
->info
.is_haswell
;
1122 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_CLIP_ENABLE_FEATURES_EXT
: {
1123 VkPhysicalDeviceDepthClipEnableFeaturesEXT
*features
=
1124 (VkPhysicalDeviceDepthClipEnableFeaturesEXT
*)ext
;
1125 features
->depthClipEnable
= true;
1129 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR
: {
1130 VkPhysicalDeviceFloat16Int8FeaturesKHR
*features
= (void *)ext
;
1131 CORE_FEATURE(1, 2, shaderFloat16
);
1132 CORE_FEATURE(1, 2, shaderInt8
);
1136 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FRAGMENT_SHADER_INTERLOCK_FEATURES_EXT
: {
1137 VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT
*features
=
1138 (VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT
*)ext
;
1139 features
->fragmentShaderSampleInterlock
= pdevice
->info
.gen
>= 9;
1140 features
->fragmentShaderPixelInterlock
= pdevice
->info
.gen
>= 9;
1141 features
->fragmentShaderShadingRateInterlock
= false;
1145 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_QUERY_RESET_FEATURES_EXT
: {
1146 VkPhysicalDeviceHostQueryResetFeaturesEXT
*features
=
1147 (VkPhysicalDeviceHostQueryResetFeaturesEXT
*)ext
;
1148 CORE_FEATURE(1, 2, hostQueryReset
);
1152 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT
: {
1153 VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*features
=
1154 (VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*)ext
;
1155 CORE_FEATURE(1, 2, shaderInputAttachmentArrayDynamicIndexing
);
1156 CORE_FEATURE(1, 2, shaderUniformTexelBufferArrayDynamicIndexing
);
1157 CORE_FEATURE(1, 2, shaderStorageTexelBufferArrayDynamicIndexing
);
1158 CORE_FEATURE(1, 2, shaderUniformBufferArrayNonUniformIndexing
);
1159 CORE_FEATURE(1, 2, shaderSampledImageArrayNonUniformIndexing
);
1160 CORE_FEATURE(1, 2, shaderStorageBufferArrayNonUniformIndexing
);
1161 CORE_FEATURE(1, 2, shaderStorageImageArrayNonUniformIndexing
);
1162 CORE_FEATURE(1, 2, shaderInputAttachmentArrayNonUniformIndexing
);
1163 CORE_FEATURE(1, 2, shaderUniformTexelBufferArrayNonUniformIndexing
);
1164 CORE_FEATURE(1, 2, shaderStorageTexelBufferArrayNonUniformIndexing
);
1165 CORE_FEATURE(1, 2, descriptorBindingUniformBufferUpdateAfterBind
);
1166 CORE_FEATURE(1, 2, descriptorBindingSampledImageUpdateAfterBind
);
1167 CORE_FEATURE(1, 2, descriptorBindingStorageImageUpdateAfterBind
);
1168 CORE_FEATURE(1, 2, descriptorBindingStorageBufferUpdateAfterBind
);
1169 CORE_FEATURE(1, 2, descriptorBindingUniformTexelBufferUpdateAfterBind
);
1170 CORE_FEATURE(1, 2, descriptorBindingStorageTexelBufferUpdateAfterBind
);
1171 CORE_FEATURE(1, 2, descriptorBindingUpdateUnusedWhilePending
);
1172 CORE_FEATURE(1, 2, descriptorBindingPartiallyBound
);
1173 CORE_FEATURE(1, 2, descriptorBindingVariableDescriptorCount
);
1174 CORE_FEATURE(1, 2, runtimeDescriptorArray
);
1178 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INDEX_TYPE_UINT8_FEATURES_EXT
: {
1179 VkPhysicalDeviceIndexTypeUint8FeaturesEXT
*features
=
1180 (VkPhysicalDeviceIndexTypeUint8FeaturesEXT
*)ext
;
1181 features
->indexTypeUint8
= true;
1185 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_FEATURES_EXT
: {
1186 VkPhysicalDeviceInlineUniformBlockFeaturesEXT
*features
=
1187 (VkPhysicalDeviceInlineUniformBlockFeaturesEXT
*)ext
;
1188 features
->inlineUniformBlock
= true;
1189 features
->descriptorBindingInlineUniformBlockUpdateAfterBind
= true;
1193 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_FEATURES_EXT
: {
1194 VkPhysicalDeviceLineRasterizationFeaturesEXT
*features
=
1195 (VkPhysicalDeviceLineRasterizationFeaturesEXT
*)ext
;
1196 features
->rectangularLines
= true;
1197 features
->bresenhamLines
= true;
1198 /* Support for Smooth lines with MSAA was removed on gen11. From the
1199 * BSpec section "Multisample ModesState" table for "AA Line Support
1202 * GEN10:BUG:######## NUM_MULTISAMPLES == 1
1204 * Fortunately, this isn't a case most people care about.
1206 features
->smoothLines
= pdevice
->info
.gen
< 10;
1207 features
->stippledRectangularLines
= false;
1208 features
->stippledBresenhamLines
= true;
1209 features
->stippledSmoothLines
= false;
1213 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES
: {
1214 VkPhysicalDeviceMultiviewFeatures
*features
=
1215 (VkPhysicalDeviceMultiviewFeatures
*)ext
;
1216 CORE_FEATURE(1, 1, multiview
);
1217 CORE_FEATURE(1, 1, multiviewGeometryShader
);
1218 CORE_FEATURE(1, 1, multiviewTessellationShader
);
1222 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGELESS_FRAMEBUFFER_FEATURES_KHR
: {
1223 VkPhysicalDeviceImagelessFramebufferFeaturesKHR
*features
=
1224 (VkPhysicalDeviceImagelessFramebufferFeaturesKHR
*)ext
;
1225 CORE_FEATURE(1, 2, imagelessFramebuffer
);
1229 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR
: {
1230 VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR
*features
=
1231 (VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR
*)ext
;
1232 features
->pipelineExecutableInfo
= true;
1236 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES
: {
1237 VkPhysicalDeviceProtectedMemoryFeatures
*features
= (void *)ext
;
1238 CORE_FEATURE(1, 1, protectedMemory
);
1242 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES
: {
1243 VkPhysicalDeviceSamplerYcbcrConversionFeatures
*features
=
1244 (VkPhysicalDeviceSamplerYcbcrConversionFeatures
*) ext
;
1245 CORE_FEATURE(1, 1, samplerYcbcrConversion
);
1249 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES_EXT
: {
1250 VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*features
=
1251 (VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*)ext
;
1252 CORE_FEATURE(1, 2, scalarBlockLayout
);
1256 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SEPARATE_DEPTH_STENCIL_LAYOUTS_FEATURES_KHR
: {
1257 VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR
*features
=
1258 (VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR
*)ext
;
1259 CORE_FEATURE(1, 2, separateDepthStencilLayouts
);
1263 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_INT64_FEATURES_KHR
: {
1264 VkPhysicalDeviceShaderAtomicInt64FeaturesKHR
*features
= (void *)ext
;
1265 CORE_FEATURE(1, 2, shaderBufferInt64Atomics
);
1266 CORE_FEATURE(1, 2, shaderSharedInt64Atomics
);
1270 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DEMOTE_TO_HELPER_INVOCATION_FEATURES_EXT
: {
1271 VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT
*features
= (void *)ext
;
1272 features
->shaderDemoteToHelperInvocation
= true;
1276 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CLOCK_FEATURES_KHR
: {
1277 VkPhysicalDeviceShaderClockFeaturesKHR
*features
=
1278 (VkPhysicalDeviceShaderClockFeaturesKHR
*)ext
;
1279 features
->shaderSubgroupClock
= true;
1280 features
->shaderDeviceClock
= false;
1284 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETERS_FEATURES
: {
1285 VkPhysicalDeviceShaderDrawParametersFeatures
*features
= (void *)ext
;
1286 CORE_FEATURE(1, 1, shaderDrawParameters
);
1290 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_SUBGROUP_EXTENDED_TYPES_FEATURES_KHR
: {
1291 VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR
*features
=
1292 (VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR
*)ext
;
1293 CORE_FEATURE(1, 2, shaderSubgroupExtendedTypes
);
1297 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_FEATURES_EXT
: {
1298 VkPhysicalDeviceSubgroupSizeControlFeaturesEXT
*features
=
1299 (VkPhysicalDeviceSubgroupSizeControlFeaturesEXT
*)ext
;
1300 features
->subgroupSizeControl
= true;
1301 features
->computeFullSubgroups
= true;
1305 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_FEATURES_EXT
: {
1306 VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT
*features
=
1307 (VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT
*)ext
;
1308 features
->texelBufferAlignment
= true;
1312 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_FEATURES_KHR
: {
1313 VkPhysicalDeviceTimelineSemaphoreFeaturesKHR
*features
=
1314 (VkPhysicalDeviceTimelineSemaphoreFeaturesKHR
*) ext
;
1315 CORE_FEATURE(1, 2, timelineSemaphore
);
1319 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES
: {
1320 VkPhysicalDeviceVariablePointersFeatures
*features
= (void *)ext
;
1321 CORE_FEATURE(1, 1, variablePointersStorageBuffer
);
1322 CORE_FEATURE(1, 1, variablePointers
);
1326 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT
: {
1327 VkPhysicalDeviceTransformFeedbackFeaturesEXT
*features
=
1328 (VkPhysicalDeviceTransformFeedbackFeaturesEXT
*)ext
;
1329 features
->transformFeedback
= true;
1330 features
->geometryStreams
= true;
1334 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_UNIFORM_BUFFER_STANDARD_LAYOUT_FEATURES_KHR
: {
1335 VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR
*features
=
1336 (VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR
*)ext
;
1337 CORE_FEATURE(1, 2, uniformBufferStandardLayout
);
1341 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT
: {
1342 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*features
=
1343 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*)ext
;
1344 features
->vertexAttributeInstanceRateDivisor
= true;
1345 features
->vertexAttributeInstanceRateZeroDivisor
= true;
1349 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES
:
1350 anv_get_physical_device_features_1_1(pdevice
, (void *)ext
);
1353 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES
:
1354 anv_get_physical_device_features_1_2(pdevice
, (void *)ext
);
1357 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_MEMORY_MODEL_FEATURES_KHR
: {
1358 VkPhysicalDeviceVulkanMemoryModelFeaturesKHR
*features
= (void *)ext
;
1359 CORE_FEATURE(1, 2, vulkanMemoryModel
);
1360 CORE_FEATURE(1, 2, vulkanMemoryModelDeviceScope
);
1361 CORE_FEATURE(1, 2, vulkanMemoryModelAvailabilityVisibilityChains
);
1365 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_YCBCR_IMAGE_ARRAYS_FEATURES_EXT
: {
1366 VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*features
=
1367 (VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*)ext
;
1368 features
->ycbcrImageArrays
= true;
1373 anv_debug_ignored_stype(ext
->sType
);
1381 #define MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS 64
1383 #define MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS 64
1384 #define MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS 256
1386 void anv_GetPhysicalDeviceProperties(
1387 VkPhysicalDevice physicalDevice
,
1388 VkPhysicalDeviceProperties
* pProperties
)
1390 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1391 const struct gen_device_info
*devinfo
= &pdevice
->info
;
1393 /* See assertions made when programming the buffer surface state. */
1394 const uint32_t max_raw_buffer_sz
= devinfo
->gen
>= 7 ?
1395 (1ul << 30) : (1ul << 27);
1397 const uint32_t max_ssbos
= pdevice
->has_a64_buffer_access
? UINT16_MAX
: 64;
1398 const uint32_t max_textures
=
1399 pdevice
->has_bindless_images
? UINT16_MAX
: 128;
1400 const uint32_t max_samplers
=
1401 pdevice
->has_bindless_samplers
? UINT16_MAX
:
1402 (devinfo
->gen
>= 8 || devinfo
->is_haswell
) ? 128 : 16;
1403 const uint32_t max_images
=
1404 pdevice
->has_bindless_images
? UINT16_MAX
: MAX_IMAGES
;
1406 /* If we can use bindless for everything, claim a high per-stage limit,
1407 * otherwise use the binding table size, minus the slots reserved for
1408 * render targets and one slot for the descriptor buffer. */
1409 const uint32_t max_per_stage
=
1410 pdevice
->has_bindless_images
&& pdevice
->has_a64_buffer_access
1411 ? UINT32_MAX
: MAX_BINDING_TABLE_SIZE
- MAX_RTS
- 1;
1413 const uint32_t max_workgroup_size
= 32 * devinfo
->max_cs_threads
;
1415 VkSampleCountFlags sample_counts
=
1416 isl_device_get_sample_counts(&pdevice
->isl_dev
);
1419 VkPhysicalDeviceLimits limits
= {
1420 .maxImageDimension1D
= (1 << 14),
1421 .maxImageDimension2D
= (1 << 14),
1422 .maxImageDimension3D
= (1 << 11),
1423 .maxImageDimensionCube
= (1 << 14),
1424 .maxImageArrayLayers
= (1 << 11),
1425 .maxTexelBufferElements
= 128 * 1024 * 1024,
1426 .maxUniformBufferRange
= (1ul << 27),
1427 .maxStorageBufferRange
= max_raw_buffer_sz
,
1428 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
1429 .maxMemoryAllocationCount
= UINT32_MAX
,
1430 .maxSamplerAllocationCount
= 64 * 1024,
1431 .bufferImageGranularity
= 64, /* A cache line */
1432 .sparseAddressSpaceSize
= 0,
1433 .maxBoundDescriptorSets
= MAX_SETS
,
1434 .maxPerStageDescriptorSamplers
= max_samplers
,
1435 .maxPerStageDescriptorUniformBuffers
= MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
,
1436 .maxPerStageDescriptorStorageBuffers
= max_ssbos
,
1437 .maxPerStageDescriptorSampledImages
= max_textures
,
1438 .maxPerStageDescriptorStorageImages
= max_images
,
1439 .maxPerStageDescriptorInputAttachments
= MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS
,
1440 .maxPerStageResources
= max_per_stage
,
1441 .maxDescriptorSetSamplers
= 6 * max_samplers
, /* number of stages * maxPerStageDescriptorSamplers */
1442 .maxDescriptorSetUniformBuffers
= 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
, /* number of stages * maxPerStageDescriptorUniformBuffers */
1443 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
1444 .maxDescriptorSetStorageBuffers
= 6 * max_ssbos
, /* number of stages * maxPerStageDescriptorStorageBuffers */
1445 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
1446 .maxDescriptorSetSampledImages
= 6 * max_textures
, /* number of stages * maxPerStageDescriptorSampledImages */
1447 .maxDescriptorSetStorageImages
= 6 * max_images
, /* number of stages * maxPerStageDescriptorStorageImages */
1448 .maxDescriptorSetInputAttachments
= MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS
,
1449 .maxVertexInputAttributes
= MAX_VBS
,
1450 .maxVertexInputBindings
= MAX_VBS
,
1451 .maxVertexInputAttributeOffset
= 2047,
1452 .maxVertexInputBindingStride
= 2048,
1453 .maxVertexOutputComponents
= 128,
1454 .maxTessellationGenerationLevel
= 64,
1455 .maxTessellationPatchSize
= 32,
1456 .maxTessellationControlPerVertexInputComponents
= 128,
1457 .maxTessellationControlPerVertexOutputComponents
= 128,
1458 .maxTessellationControlPerPatchOutputComponents
= 128,
1459 .maxTessellationControlTotalOutputComponents
= 2048,
1460 .maxTessellationEvaluationInputComponents
= 128,
1461 .maxTessellationEvaluationOutputComponents
= 128,
1462 .maxGeometryShaderInvocations
= 32,
1463 .maxGeometryInputComponents
= 64,
1464 .maxGeometryOutputComponents
= 128,
1465 .maxGeometryOutputVertices
= 256,
1466 .maxGeometryTotalOutputComponents
= 1024,
1467 .maxFragmentInputComponents
= 116, /* 128 components - (PSIZ, CLIP_DIST0, CLIP_DIST1) */
1468 .maxFragmentOutputAttachments
= 8,
1469 .maxFragmentDualSrcAttachments
= 1,
1470 .maxFragmentCombinedOutputResources
= 8,
1471 .maxComputeSharedMemorySize
= 64 * 1024,
1472 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
1473 .maxComputeWorkGroupInvocations
= max_workgroup_size
,
1474 .maxComputeWorkGroupSize
= {
1479 .subPixelPrecisionBits
= 8,
1480 .subTexelPrecisionBits
= 8,
1481 .mipmapPrecisionBits
= 8,
1482 .maxDrawIndexedIndexValue
= UINT32_MAX
,
1483 .maxDrawIndirectCount
= UINT32_MAX
,
1484 .maxSamplerLodBias
= 16,
1485 .maxSamplerAnisotropy
= 16,
1486 .maxViewports
= MAX_VIEWPORTS
,
1487 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
1488 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
1489 .viewportSubPixelBits
= 13, /* We take a float? */
1490 .minMemoryMapAlignment
= 4096, /* A page */
1491 /* The dataport requires texel alignment so we need to assume a worst
1492 * case of R32G32B32A32 which is 16 bytes.
1494 .minTexelBufferOffsetAlignment
= 16,
1495 /* We need 16 for UBO block reads to work and 32 for push UBOs */
1496 .minUniformBufferOffsetAlignment
= 32,
1497 .minStorageBufferOffsetAlignment
= 4,
1498 .minTexelOffset
= -8,
1499 .maxTexelOffset
= 7,
1500 .minTexelGatherOffset
= -32,
1501 .maxTexelGatherOffset
= 31,
1502 .minInterpolationOffset
= -0.5,
1503 .maxInterpolationOffset
= 0.4375,
1504 .subPixelInterpolationOffsetBits
= 4,
1505 .maxFramebufferWidth
= (1 << 14),
1506 .maxFramebufferHeight
= (1 << 14),
1507 .maxFramebufferLayers
= (1 << 11),
1508 .framebufferColorSampleCounts
= sample_counts
,
1509 .framebufferDepthSampleCounts
= sample_counts
,
1510 .framebufferStencilSampleCounts
= sample_counts
,
1511 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
1512 .maxColorAttachments
= MAX_RTS
,
1513 .sampledImageColorSampleCounts
= sample_counts
,
1514 .sampledImageIntegerSampleCounts
= sample_counts
,
1515 .sampledImageDepthSampleCounts
= sample_counts
,
1516 .sampledImageStencilSampleCounts
= sample_counts
,
1517 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
1518 .maxSampleMaskWords
= 1,
1519 .timestampComputeAndGraphics
= true,
1520 .timestampPeriod
= 1000000000.0 / devinfo
->timestamp_frequency
,
1521 .maxClipDistances
= 8,
1522 .maxCullDistances
= 8,
1523 .maxCombinedClipAndCullDistances
= 8,
1524 .discreteQueuePriorities
= 2,
1525 .pointSizeRange
= { 0.125, 255.875 },
1528 (devinfo
->gen
>= 9 || devinfo
->is_cherryview
) ?
1529 2047.9921875 : 7.9921875,
1531 .pointSizeGranularity
= (1.0 / 8.0),
1532 .lineWidthGranularity
= (1.0 / 128.0),
1533 .strictLines
= false,
1534 .standardSampleLocations
= true,
1535 .optimalBufferCopyOffsetAlignment
= 128,
1536 .optimalBufferCopyRowPitchAlignment
= 128,
1537 .nonCoherentAtomSize
= 64,
1540 *pProperties
= (VkPhysicalDeviceProperties
) {
1541 .apiVersion
= anv_physical_device_api_version(pdevice
),
1542 .driverVersion
= vk_get_driver_version(),
1544 .deviceID
= pdevice
->chipset_id
,
1545 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
1547 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
1550 snprintf(pProperties
->deviceName
, sizeof(pProperties
->deviceName
),
1551 "%s", pdevice
->name
);
1552 memcpy(pProperties
->pipelineCacheUUID
,
1553 pdevice
->pipeline_cache_uuid
, VK_UUID_SIZE
);
1557 anv_get_physical_device_properties_1_1(struct anv_physical_device
*pdevice
,
1558 VkPhysicalDeviceVulkan11Properties
*p
)
1560 assert(p
->sType
== VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES
);
1562 memcpy(p
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
1563 memcpy(p
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
1564 memset(p
->deviceLUID
, 0, VK_LUID_SIZE
);
1565 p
->deviceNodeMask
= 0;
1566 p
->deviceLUIDValid
= false;
1568 p
->subgroupSize
= BRW_SUBGROUP_SIZE
;
1569 VkShaderStageFlags scalar_stages
= 0;
1570 for (unsigned stage
= 0; stage
< MESA_SHADER_STAGES
; stage
++) {
1571 if (pdevice
->compiler
->scalar_stage
[stage
])
1572 scalar_stages
|= mesa_to_vk_shader_stage(stage
);
1574 p
->subgroupSupportedStages
= scalar_stages
;
1575 p
->subgroupSupportedOperations
= VK_SUBGROUP_FEATURE_BASIC_BIT
|
1576 VK_SUBGROUP_FEATURE_VOTE_BIT
|
1577 VK_SUBGROUP_FEATURE_BALLOT_BIT
|
1578 VK_SUBGROUP_FEATURE_SHUFFLE_BIT
|
1579 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT
|
1580 VK_SUBGROUP_FEATURE_QUAD_BIT
;
1581 if (pdevice
->info
.gen
>= 8) {
1582 /* TODO: There's no technical reason why these can't be made to
1583 * work on gen7 but they don't at the moment so it's best to leave
1584 * the feature disabled than enabled and broken.
1586 p
->subgroupSupportedOperations
|= VK_SUBGROUP_FEATURE_ARITHMETIC_BIT
|
1587 VK_SUBGROUP_FEATURE_CLUSTERED_BIT
;
1589 p
->subgroupQuadOperationsInAllStages
= pdevice
->info
.gen
>= 8;
1591 p
->pointClippingBehavior
= VK_POINT_CLIPPING_BEHAVIOR_USER_CLIP_PLANES_ONLY
;
1592 p
->maxMultiviewViewCount
= 16;
1593 p
->maxMultiviewInstanceIndex
= UINT32_MAX
/ 16;
1594 p
->protectedNoFault
= false;
1595 /* This value doesn't matter for us today as our per-stage descriptors are
1598 p
->maxPerSetDescriptors
= 1024;
1599 p
->maxMemoryAllocationSize
= MAX_MEMORY_ALLOCATION_SIZE
;
1603 anv_get_physical_device_properties_1_2(struct anv_physical_device
*pdevice
,
1604 VkPhysicalDeviceVulkan12Properties
*p
)
1606 assert(p
->sType
== VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES
);
1608 p
->driverID
= VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA_KHR
;
1609 memset(p
->driverName
, 0, sizeof(p
->driverName
));
1610 snprintf(p
->driverName
, VK_MAX_DRIVER_NAME_SIZE_KHR
,
1611 "Intel open-source Mesa driver");
1612 memset(p
->driverInfo
, 0, sizeof(p
->driverInfo
));
1613 snprintf(p
->driverInfo
, VK_MAX_DRIVER_INFO_SIZE_KHR
,
1614 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1
);
1615 p
->conformanceVersion
= (VkConformanceVersionKHR
) {
1622 p
->denormBehaviorIndependence
=
1623 VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_ALL_KHR
;
1624 p
->roundingModeIndependence
=
1625 VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_NONE_KHR
;
1627 /* Broadwell does not support HF denorms and there are restrictions
1628 * other gens. According to Kabylake's PRM:
1630 * "math - Extended Math Function
1632 * Restriction : Half-float denorms are always retained."
1634 p
->shaderDenormFlushToZeroFloat16
= false;
1635 p
->shaderDenormPreserveFloat16
= pdevice
->info
.gen
> 8;
1636 p
->shaderRoundingModeRTEFloat16
= true;
1637 p
->shaderRoundingModeRTZFloat16
= true;
1638 p
->shaderSignedZeroInfNanPreserveFloat16
= true;
1640 p
->shaderDenormFlushToZeroFloat32
= true;
1641 p
->shaderDenormPreserveFloat32
= true;
1642 p
->shaderRoundingModeRTEFloat32
= true;
1643 p
->shaderRoundingModeRTZFloat32
= true;
1644 p
->shaderSignedZeroInfNanPreserveFloat32
= true;
1646 p
->shaderDenormFlushToZeroFloat64
= true;
1647 p
->shaderDenormPreserveFloat64
= true;
1648 p
->shaderRoundingModeRTEFloat64
= true;
1649 p
->shaderRoundingModeRTZFloat64
= true;
1650 p
->shaderSignedZeroInfNanPreserveFloat64
= true;
1652 /* It's a bit hard to exactly map our implementation to the limits
1653 * described here. The bindless surface handle in the extended
1654 * message descriptors is 20 bits and it's an index into the table of
1655 * RENDER_SURFACE_STATE structs that starts at bindless surface base
1656 * address. Given that most things consume two surface states per
1657 * view (general/sampled for textures and write-only/read-write for
1658 * images), we claim 2^19 things.
1660 * For SSBOs, we just use A64 messages so there is no real limit
1661 * there beyond the limit on the total size of a descriptor set.
1663 const unsigned max_bindless_views
= 1 << 19;
1664 p
->maxUpdateAfterBindDescriptorsInAllPools
= max_bindless_views
;
1665 p
->shaderUniformBufferArrayNonUniformIndexingNative
= false;
1666 p
->shaderSampledImageArrayNonUniformIndexingNative
= false;
1667 p
->shaderStorageBufferArrayNonUniformIndexingNative
= true;
1668 p
->shaderStorageImageArrayNonUniformIndexingNative
= false;
1669 p
->shaderInputAttachmentArrayNonUniformIndexingNative
= false;
1670 p
->robustBufferAccessUpdateAfterBind
= true;
1671 p
->quadDivergentImplicitLod
= false;
1672 p
->maxPerStageDescriptorUpdateAfterBindSamplers
= max_bindless_views
;
1673 p
->maxPerStageDescriptorUpdateAfterBindUniformBuffers
= MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
;
1674 p
->maxPerStageDescriptorUpdateAfterBindStorageBuffers
= UINT32_MAX
;
1675 p
->maxPerStageDescriptorUpdateAfterBindSampledImages
= max_bindless_views
;
1676 p
->maxPerStageDescriptorUpdateAfterBindStorageImages
= max_bindless_views
;
1677 p
->maxPerStageDescriptorUpdateAfterBindInputAttachments
= MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS
;
1678 p
->maxPerStageUpdateAfterBindResources
= UINT32_MAX
;
1679 p
->maxDescriptorSetUpdateAfterBindSamplers
= max_bindless_views
;
1680 p
->maxDescriptorSetUpdateAfterBindUniformBuffers
= 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
;
1681 p
->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2;
1682 p
->maxDescriptorSetUpdateAfterBindStorageBuffers
= UINT32_MAX
;
1683 p
->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2;
1684 p
->maxDescriptorSetUpdateAfterBindSampledImages
= max_bindless_views
;
1685 p
->maxDescriptorSetUpdateAfterBindStorageImages
= max_bindless_views
;
1686 p
->maxDescriptorSetUpdateAfterBindInputAttachments
= MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS
;
1688 /* We support all of the depth resolve modes */
1689 p
->supportedDepthResolveModes
= VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
|
1690 VK_RESOLVE_MODE_AVERAGE_BIT_KHR
|
1691 VK_RESOLVE_MODE_MIN_BIT_KHR
|
1692 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1693 /* Average doesn't make sense for stencil so we don't support that */
1694 p
->supportedStencilResolveModes
= VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
;
1695 if (pdevice
->info
.gen
>= 8) {
1696 /* The advanced stencil resolve modes currently require stencil
1697 * sampling be supported by the hardware.
1699 p
->supportedStencilResolveModes
|= VK_RESOLVE_MODE_MIN_BIT_KHR
|
1700 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1702 p
->independentResolveNone
= true;
1703 p
->independentResolve
= true;
1705 p
->filterMinmaxSingleComponentFormats
= pdevice
->info
.gen
>= 9;
1706 p
->filterMinmaxImageComponentMapping
= pdevice
->info
.gen
>= 9;
1708 p
->maxTimelineSemaphoreValueDifference
= UINT64_MAX
;
1710 p
->framebufferIntegerColorSampleCounts
=
1711 isl_device_get_sample_counts(&pdevice
->isl_dev
);
1714 void anv_GetPhysicalDeviceProperties2(
1715 VkPhysicalDevice physicalDevice
,
1716 VkPhysicalDeviceProperties2
* pProperties
)
1718 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1720 anv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
1722 VkPhysicalDeviceVulkan11Properties core_1_1
= {
1723 .sType
= VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES
,
1725 anv_get_physical_device_properties_1_1(pdevice
, &core_1_1
);
1727 VkPhysicalDeviceVulkan12Properties core_1_2
= {
1728 .sType
= VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES
,
1730 anv_get_physical_device_properties_1_2(pdevice
, &core_1_2
);
1732 #define CORE_RENAMED_PROPERTY(major, minor, ext_property, core_property) \
1733 memcpy(&properties->ext_property, &core_##major##_##minor.core_property, \
1734 sizeof(core_##major##_##minor.core_property))
1736 #define CORE_PROPERTY(major, minor, property) \
1737 CORE_RENAMED_PROPERTY(major, minor, property, property)
1739 vk_foreach_struct(ext
, pProperties
->pNext
) {
1740 switch (ext
->sType
) {
1741 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_STENCIL_RESOLVE_PROPERTIES_KHR
: {
1742 VkPhysicalDeviceDepthStencilResolvePropertiesKHR
*properties
=
1743 (VkPhysicalDeviceDepthStencilResolvePropertiesKHR
*)ext
;
1744 CORE_PROPERTY(1, 2, supportedDepthResolveModes
);
1745 CORE_PROPERTY(1, 2, supportedStencilResolveModes
);
1746 CORE_PROPERTY(1, 2, independentResolveNone
);
1747 CORE_PROPERTY(1, 2, independentResolve
);
1751 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES_EXT
: {
1752 VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*properties
=
1753 (VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*)ext
;
1754 CORE_PROPERTY(1, 2, maxUpdateAfterBindDescriptorsInAllPools
);
1755 CORE_PROPERTY(1, 2, shaderUniformBufferArrayNonUniformIndexingNative
);
1756 CORE_PROPERTY(1, 2, shaderSampledImageArrayNonUniformIndexingNative
);
1757 CORE_PROPERTY(1, 2, shaderStorageBufferArrayNonUniformIndexingNative
);
1758 CORE_PROPERTY(1, 2, shaderStorageImageArrayNonUniformIndexingNative
);
1759 CORE_PROPERTY(1, 2, shaderInputAttachmentArrayNonUniformIndexingNative
);
1760 CORE_PROPERTY(1, 2, robustBufferAccessUpdateAfterBind
);
1761 CORE_PROPERTY(1, 2, quadDivergentImplicitLod
);
1762 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindSamplers
);
1763 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindUniformBuffers
);
1764 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindStorageBuffers
);
1765 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindSampledImages
);
1766 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindStorageImages
);
1767 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindInputAttachments
);
1768 CORE_PROPERTY(1, 2, maxPerStageUpdateAfterBindResources
);
1769 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindSamplers
);
1770 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindUniformBuffers
);
1771 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindUniformBuffersDynamic
);
1772 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindStorageBuffers
);
1773 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindStorageBuffersDynamic
);
1774 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindSampledImages
);
1775 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindStorageImages
);
1776 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindInputAttachments
);
1780 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR
: {
1781 VkPhysicalDeviceDriverPropertiesKHR
*properties
=
1782 (VkPhysicalDeviceDriverPropertiesKHR
*) ext
;
1783 CORE_PROPERTY(1, 2, driverID
);
1784 CORE_PROPERTY(1, 2, driverName
);
1785 CORE_PROPERTY(1, 2, driverInfo
);
1786 CORE_PROPERTY(1, 2, conformanceVersion
);
1790 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT
: {
1791 VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*props
=
1792 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*) ext
;
1793 /* Userptr needs page aligned memory. */
1794 props
->minImportedHostPointerAlignment
= 4096;
1798 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES
: {
1799 VkPhysicalDeviceIDProperties
*properties
=
1800 (VkPhysicalDeviceIDProperties
*)ext
;
1801 CORE_PROPERTY(1, 1, deviceUUID
);
1802 CORE_PROPERTY(1, 1, driverUUID
);
1803 CORE_PROPERTY(1, 1, deviceLUID
);
1804 CORE_PROPERTY(1, 1, deviceLUIDValid
);
1808 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_PROPERTIES_EXT
: {
1809 VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*props
=
1810 (VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*)ext
;
1811 props
->maxInlineUniformBlockSize
= MAX_INLINE_UNIFORM_BLOCK_SIZE
;
1812 props
->maxPerStageDescriptorInlineUniformBlocks
=
1813 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1814 props
->maxPerStageDescriptorUpdateAfterBindInlineUniformBlocks
=
1815 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1816 props
->maxDescriptorSetInlineUniformBlocks
=
1817 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1818 props
->maxDescriptorSetUpdateAfterBindInlineUniformBlocks
=
1819 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1823 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_PROPERTIES_EXT
: {
1824 VkPhysicalDeviceLineRasterizationPropertiesEXT
*props
=
1825 (VkPhysicalDeviceLineRasterizationPropertiesEXT
*)ext
;
1826 /* In the Skylake PRM Vol. 7, subsection titled "GIQ (Diamond)
1827 * Sampling Rules - Legacy Mode", it says the following:
1829 * "Note that the device divides a pixel into a 16x16 array of
1830 * subpixels, referenced by their upper left corners."
1832 * This is the only known reference in the PRMs to the subpixel
1833 * precision of line rasterization and a "16x16 array of subpixels"
1834 * implies 4 subpixel precision bits. Empirical testing has shown
1835 * that 4 subpixel precision bits applies to all line rasterization
1838 props
->lineSubPixelPrecisionBits
= 4;
1842 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES
: {
1843 VkPhysicalDeviceMaintenance3Properties
*properties
=
1844 (VkPhysicalDeviceMaintenance3Properties
*)ext
;
1845 /* This value doesn't matter for us today as our per-stage
1846 * descriptors are the real limit.
1848 CORE_PROPERTY(1, 1, maxPerSetDescriptors
);
1849 CORE_PROPERTY(1, 1, maxMemoryAllocationSize
);
1853 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES
: {
1854 VkPhysicalDeviceMultiviewProperties
*properties
=
1855 (VkPhysicalDeviceMultiviewProperties
*)ext
;
1856 CORE_PROPERTY(1, 1, maxMultiviewViewCount
);
1857 CORE_PROPERTY(1, 1, maxMultiviewInstanceIndex
);
1861 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT
: {
1862 VkPhysicalDevicePCIBusInfoPropertiesEXT
*properties
=
1863 (VkPhysicalDevicePCIBusInfoPropertiesEXT
*)ext
;
1864 properties
->pciDomain
= pdevice
->pci_info
.domain
;
1865 properties
->pciBus
= pdevice
->pci_info
.bus
;
1866 properties
->pciDevice
= pdevice
->pci_info
.device
;
1867 properties
->pciFunction
= pdevice
->pci_info
.function
;
1871 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES
: {
1872 VkPhysicalDevicePointClippingProperties
*properties
=
1873 (VkPhysicalDevicePointClippingProperties
*) ext
;
1874 CORE_PROPERTY(1, 1, pointClippingBehavior
);
1878 #pragma GCC diagnostic push
1879 #pragma GCC diagnostic ignored "-Wswitch"
1880 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PRESENTATION_PROPERTIES_ANDROID
: {
1881 VkPhysicalDevicePresentationPropertiesANDROID
*props
=
1882 (VkPhysicalDevicePresentationPropertiesANDROID
*)ext
;
1883 props
->sharedImage
= VK_FALSE
;
1886 #pragma GCC diagnostic pop
1888 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES
: {
1889 VkPhysicalDeviceProtectedMemoryProperties
*properties
=
1890 (VkPhysicalDeviceProtectedMemoryProperties
*)ext
;
1891 CORE_PROPERTY(1, 1, protectedNoFault
);
1895 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
1896 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
1897 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
1898 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
1902 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT
: {
1903 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*properties
=
1904 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*)ext
;
1905 CORE_PROPERTY(1, 2, filterMinmaxImageComponentMapping
);
1906 CORE_PROPERTY(1, 2, filterMinmaxSingleComponentFormats
);
1910 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES
: {
1911 VkPhysicalDeviceSubgroupProperties
*properties
= (void *)ext
;
1912 CORE_PROPERTY(1, 1, subgroupSize
);
1913 CORE_RENAMED_PROPERTY(1, 1, supportedStages
,
1914 subgroupSupportedStages
);
1915 CORE_RENAMED_PROPERTY(1, 1, supportedOperations
,
1916 subgroupSupportedOperations
);
1917 CORE_RENAMED_PROPERTY(1, 1, quadOperationsInAllStages
,
1918 subgroupQuadOperationsInAllStages
);
1922 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_PROPERTIES_EXT
: {
1923 VkPhysicalDeviceSubgroupSizeControlPropertiesEXT
*props
=
1924 (VkPhysicalDeviceSubgroupSizeControlPropertiesEXT
*)ext
;
1925 STATIC_ASSERT(8 <= BRW_SUBGROUP_SIZE
&& BRW_SUBGROUP_SIZE
<= 32);
1926 props
->minSubgroupSize
= 8;
1927 props
->maxSubgroupSize
= 32;
1928 props
->maxComputeWorkgroupSubgroups
= pdevice
->info
.max_cs_threads
;
1929 props
->requiredSubgroupSizeStages
= VK_SHADER_STAGE_COMPUTE_BIT
;
1932 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT_CONTROLS_PROPERTIES_KHR
: {
1933 VkPhysicalDeviceFloatControlsPropertiesKHR
*properties
= (void *)ext
;
1934 CORE_PROPERTY(1, 2, denormBehaviorIndependence
);
1935 CORE_PROPERTY(1, 2, roundingModeIndependence
);
1936 CORE_PROPERTY(1, 2, shaderDenormFlushToZeroFloat16
);
1937 CORE_PROPERTY(1, 2, shaderDenormPreserveFloat16
);
1938 CORE_PROPERTY(1, 2, shaderRoundingModeRTEFloat16
);
1939 CORE_PROPERTY(1, 2, shaderRoundingModeRTZFloat16
);
1940 CORE_PROPERTY(1, 2, shaderSignedZeroInfNanPreserveFloat16
);
1941 CORE_PROPERTY(1, 2, shaderDenormFlushToZeroFloat32
);
1942 CORE_PROPERTY(1, 2, shaderDenormPreserveFloat32
);
1943 CORE_PROPERTY(1, 2, shaderRoundingModeRTEFloat32
);
1944 CORE_PROPERTY(1, 2, shaderRoundingModeRTZFloat32
);
1945 CORE_PROPERTY(1, 2, shaderSignedZeroInfNanPreserveFloat32
);
1946 CORE_PROPERTY(1, 2, shaderDenormFlushToZeroFloat64
);
1947 CORE_PROPERTY(1, 2, shaderDenormPreserveFloat64
);
1948 CORE_PROPERTY(1, 2, shaderRoundingModeRTEFloat64
);
1949 CORE_PROPERTY(1, 2, shaderRoundingModeRTZFloat64
);
1950 CORE_PROPERTY(1, 2, shaderSignedZeroInfNanPreserveFloat64
);
1954 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_PROPERTIES_EXT
: {
1955 VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT
*props
=
1956 (VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT
*)ext
;
1958 /* From the SKL PRM Vol. 2d, docs for RENDER_SURFACE_STATE::Surface
1961 * "For SURFTYPE_BUFFER non-rendertarget surfaces, this field
1962 * specifies the base address of the first element of the surface,
1963 * computed in software by adding the surface base address to the
1964 * byte offset of the element in the buffer. The base address must
1965 * be aligned to element size."
1967 * The typed dataport messages require that things be texel aligned.
1968 * Otherwise, we may just load/store the wrong data or, in the worst
1969 * case, there may be hangs.
1971 props
->storageTexelBufferOffsetAlignmentBytes
= 16;
1972 props
->storageTexelBufferOffsetSingleTexelAlignment
= true;
1974 /* The sampler, however, is much more forgiving and it can handle
1975 * arbitrary byte alignment for linear and buffer surfaces. It's
1976 * hard to find a good PRM citation for this but years of empirical
1977 * experience demonstrate that this is true.
1979 props
->uniformTexelBufferOffsetAlignmentBytes
= 1;
1980 props
->uniformTexelBufferOffsetSingleTexelAlignment
= false;
1984 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_PROPERTIES_KHR
: {
1985 VkPhysicalDeviceTimelineSemaphorePropertiesKHR
*properties
=
1986 (VkPhysicalDeviceTimelineSemaphorePropertiesKHR
*) ext
;
1987 CORE_PROPERTY(1, 2, maxTimelineSemaphoreValueDifference
);
1991 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT
: {
1992 VkPhysicalDeviceTransformFeedbackPropertiesEXT
*props
=
1993 (VkPhysicalDeviceTransformFeedbackPropertiesEXT
*)ext
;
1995 props
->maxTransformFeedbackStreams
= MAX_XFB_STREAMS
;
1996 props
->maxTransformFeedbackBuffers
= MAX_XFB_BUFFERS
;
1997 props
->maxTransformFeedbackBufferSize
= (1ull << 32);
1998 props
->maxTransformFeedbackStreamDataSize
= 128 * 4;
1999 props
->maxTransformFeedbackBufferDataSize
= 128 * 4;
2000 props
->maxTransformFeedbackBufferDataStride
= 2048;
2001 props
->transformFeedbackQueries
= true;
2002 props
->transformFeedbackStreamsLinesTriangles
= false;
2003 props
->transformFeedbackRasterizationStreamSelect
= false;
2004 props
->transformFeedbackDraw
= true;
2008 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT
: {
2009 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*props
=
2010 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*)ext
;
2011 /* We have to restrict this a bit for multiview */
2012 props
->maxVertexAttribDivisor
= UINT32_MAX
/ 16;
2016 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES
:
2017 anv_get_physical_device_properties_1_1(pdevice
, (void *)ext
);
2020 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES
:
2021 anv_get_physical_device_properties_1_2(pdevice
, (void *)ext
);
2025 anv_debug_ignored_stype(ext
->sType
);
2030 #undef CORE_RENAMED_PROPERTY
2031 #undef CORE_PROPERTY
2034 /* We support exactly one queue family. */
2035 static const VkQueueFamilyProperties
2036 anv_queue_family_properties
= {
2037 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
2038 VK_QUEUE_COMPUTE_BIT
|
2039 VK_QUEUE_TRANSFER_BIT
,
2041 .timestampValidBits
= 36, /* XXX: Real value here */
2042 .minImageTransferGranularity
= { 1, 1, 1 },
2045 void anv_GetPhysicalDeviceQueueFamilyProperties(
2046 VkPhysicalDevice physicalDevice
,
2048 VkQueueFamilyProperties
* pQueueFamilyProperties
)
2050 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pCount
);
2052 vk_outarray_append(&out
, p
) {
2053 *p
= anv_queue_family_properties
;
2057 void anv_GetPhysicalDeviceQueueFamilyProperties2(
2058 VkPhysicalDevice physicalDevice
,
2059 uint32_t* pQueueFamilyPropertyCount
,
2060 VkQueueFamilyProperties2
* pQueueFamilyProperties
)
2063 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pQueueFamilyPropertyCount
);
2065 vk_outarray_append(&out
, p
) {
2066 p
->queueFamilyProperties
= anv_queue_family_properties
;
2068 vk_foreach_struct(s
, p
->pNext
) {
2069 anv_debug_ignored_stype(s
->sType
);
2074 void anv_GetPhysicalDeviceMemoryProperties(
2075 VkPhysicalDevice physicalDevice
,
2076 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
2078 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
2080 pMemoryProperties
->memoryTypeCount
= physical_device
->memory
.type_count
;
2081 for (uint32_t i
= 0; i
< physical_device
->memory
.type_count
; i
++) {
2082 pMemoryProperties
->memoryTypes
[i
] = (VkMemoryType
) {
2083 .propertyFlags
= physical_device
->memory
.types
[i
].propertyFlags
,
2084 .heapIndex
= physical_device
->memory
.types
[i
].heapIndex
,
2088 pMemoryProperties
->memoryHeapCount
= physical_device
->memory
.heap_count
;
2089 for (uint32_t i
= 0; i
< physical_device
->memory
.heap_count
; i
++) {
2090 pMemoryProperties
->memoryHeaps
[i
] = (VkMemoryHeap
) {
2091 .size
= physical_device
->memory
.heaps
[i
].size
,
2092 .flags
= physical_device
->memory
.heaps
[i
].flags
,
2098 anv_get_memory_budget(VkPhysicalDevice physicalDevice
,
2099 VkPhysicalDeviceMemoryBudgetPropertiesEXT
*memoryBudget
)
2101 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
2102 uint64_t sys_available
= get_available_system_memory();
2103 assert(sys_available
> 0);
2105 VkDeviceSize total_heaps_size
= 0;
2106 for (size_t i
= 0; i
< device
->memory
.heap_count
; i
++)
2107 total_heaps_size
+= device
->memory
.heaps
[i
].size
;
2109 for (size_t i
= 0; i
< device
->memory
.heap_count
; i
++) {
2110 VkDeviceSize heap_size
= device
->memory
.heaps
[i
].size
;
2111 VkDeviceSize heap_used
= device
->memory
.heaps
[i
].used
;
2112 VkDeviceSize heap_budget
;
2114 double heap_proportion
= (double) heap_size
/ total_heaps_size
;
2115 VkDeviceSize sys_available_prop
= sys_available
* heap_proportion
;
2118 * Let's not incite the app to starve the system: report at most 90% of
2119 * available system memory.
2121 uint64_t heap_available
= sys_available_prop
* 9 / 10;
2122 heap_budget
= MIN2(heap_size
, heap_used
+ heap_available
);
2125 * Round down to the nearest MB
2127 heap_budget
&= ~((1ull << 20) - 1);
2130 * The heapBudget value must be non-zero for array elements less than
2131 * VkPhysicalDeviceMemoryProperties::memoryHeapCount. The heapBudget
2132 * value must be less than or equal to VkMemoryHeap::size for each heap.
2134 assert(0 < heap_budget
&& heap_budget
<= heap_size
);
2136 memoryBudget
->heapUsage
[i
] = heap_used
;
2137 memoryBudget
->heapBudget
[i
] = heap_budget
;
2140 /* The heapBudget and heapUsage values must be zero for array elements
2141 * greater than or equal to VkPhysicalDeviceMemoryProperties::memoryHeapCount
2143 for (uint32_t i
= device
->memory
.heap_count
; i
< VK_MAX_MEMORY_HEAPS
; i
++) {
2144 memoryBudget
->heapBudget
[i
] = 0;
2145 memoryBudget
->heapUsage
[i
] = 0;
2149 void anv_GetPhysicalDeviceMemoryProperties2(
2150 VkPhysicalDevice physicalDevice
,
2151 VkPhysicalDeviceMemoryProperties2
* pMemoryProperties
)
2153 anv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
2154 &pMemoryProperties
->memoryProperties
);
2156 vk_foreach_struct(ext
, pMemoryProperties
->pNext
) {
2157 switch (ext
->sType
) {
2158 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT
:
2159 anv_get_memory_budget(physicalDevice
, (void*)ext
);
2162 anv_debug_ignored_stype(ext
->sType
);
2169 anv_GetDeviceGroupPeerMemoryFeatures(
2172 uint32_t localDeviceIndex
,
2173 uint32_t remoteDeviceIndex
,
2174 VkPeerMemoryFeatureFlags
* pPeerMemoryFeatures
)
2176 assert(localDeviceIndex
== 0 && remoteDeviceIndex
== 0);
2177 *pPeerMemoryFeatures
= VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT
|
2178 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT
|
2179 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT
|
2180 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT
;
2183 PFN_vkVoidFunction
anv_GetInstanceProcAddr(
2184 VkInstance _instance
,
2187 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2189 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
2190 * when we have to return valid function pointers, NULL, or it's left
2191 * undefined. See the table for exact details.
2196 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
2197 if (strcmp(pName, "vk" #entrypoint) == 0) \
2198 return (PFN_vkVoidFunction)anv_##entrypoint
2200 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties
);
2201 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties
);
2202 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceVersion
);
2203 LOOKUP_ANV_ENTRYPOINT(CreateInstance
);
2205 #undef LOOKUP_ANV_ENTRYPOINT
2207 if (instance
== NULL
)
2210 int idx
= anv_get_instance_entrypoint_index(pName
);
2212 return instance
->dispatch
.entrypoints
[idx
];
2214 idx
= anv_get_physical_device_entrypoint_index(pName
);
2216 return instance
->physicalDevice
.dispatch
.entrypoints
[idx
];
2218 idx
= anv_get_device_entrypoint_index(pName
);
2220 return instance
->device_dispatch
.entrypoints
[idx
];
2225 /* With version 1+ of the loader interface the ICD should expose
2226 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
2229 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
2230 VkInstance instance
,
2234 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
2235 VkInstance instance
,
2238 return anv_GetInstanceProcAddr(instance
, pName
);
2241 PFN_vkVoidFunction
anv_GetDeviceProcAddr(
2245 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2247 if (!device
|| !pName
)
2250 int idx
= anv_get_device_entrypoint_index(pName
);
2254 return device
->dispatch
.entrypoints
[idx
];
2257 /* With version 4+ of the loader interface the ICD should expose
2258 * vk_icdGetPhysicalDeviceProcAddr()
2261 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetPhysicalDeviceProcAddr(
2262 VkInstance _instance
,
2265 PFN_vkVoidFunction
vk_icdGetPhysicalDeviceProcAddr(
2266 VkInstance _instance
,
2269 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2271 if (!pName
|| !instance
)
2274 int idx
= anv_get_physical_device_entrypoint_index(pName
);
2278 return instance
->physicalDevice
.dispatch
.entrypoints
[idx
];
2283 anv_CreateDebugReportCallbackEXT(VkInstance _instance
,
2284 const VkDebugReportCallbackCreateInfoEXT
* pCreateInfo
,
2285 const VkAllocationCallbacks
* pAllocator
,
2286 VkDebugReportCallbackEXT
* pCallback
)
2288 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2289 return vk_create_debug_report_callback(&instance
->debug_report_callbacks
,
2290 pCreateInfo
, pAllocator
, &instance
->alloc
,
2295 anv_DestroyDebugReportCallbackEXT(VkInstance _instance
,
2296 VkDebugReportCallbackEXT _callback
,
2297 const VkAllocationCallbacks
* pAllocator
)
2299 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2300 vk_destroy_debug_report_callback(&instance
->debug_report_callbacks
,
2301 _callback
, pAllocator
, &instance
->alloc
);
2305 anv_DebugReportMessageEXT(VkInstance _instance
,
2306 VkDebugReportFlagsEXT flags
,
2307 VkDebugReportObjectTypeEXT objectType
,
2310 int32_t messageCode
,
2311 const char* pLayerPrefix
,
2312 const char* pMessage
)
2314 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2315 vk_debug_report(&instance
->debug_report_callbacks
, flags
, objectType
,
2316 object
, location
, messageCode
, pLayerPrefix
, pMessage
);
2319 static struct anv_state
2320 anv_state_pool_emit_data(struct anv_state_pool
*pool
, size_t size
, size_t align
, const void *p
)
2322 struct anv_state state
;
2324 state
= anv_state_pool_alloc(pool
, size
, align
);
2325 memcpy(state
.map
, p
, size
);
2330 /* Haswell border color is a bit of a disaster. Float and unorm formats use a
2331 * straightforward 32-bit float color in the first 64 bytes. Instead of using
2332 * a nice float/integer union like Gen8+, Haswell specifies the integer border
2333 * color as a separate entry /after/ the float color. The layout of this entry
2334 * also depends on the format's bpp (with extra hacks for RG32), and overlaps.
2336 * Since we don't know the format/bpp, we can't make any of the border colors
2337 * containing '1' work for all formats, as it would be in the wrong place for
2338 * some of them. We opt to make 32-bit integers work as this seems like the
2339 * most common option. Fortunately, transparent black works regardless, as
2340 * all zeroes is the same in every bit-size.
2342 struct hsw_border_color
{
2346 uint32_t _pad1
[108];
2349 struct gen8_border_color
{
2354 /* Pad out to 64 bytes */
2359 anv_device_init_border_colors(struct anv_device
*device
)
2361 if (device
->info
.is_haswell
) {
2362 static const struct hsw_border_color border_colors
[] = {
2363 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
2364 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
2365 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
2366 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
2367 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
2368 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
2371 device
->border_colors
=
2372 anv_state_pool_emit_data(&device
->dynamic_state_pool
,
2373 sizeof(border_colors
), 512, border_colors
);
2375 static const struct gen8_border_color border_colors
[] = {
2376 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
2377 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
2378 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
2379 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
2380 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
2381 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
2384 device
->border_colors
=
2385 anv_state_pool_emit_data(&device
->dynamic_state_pool
,
2386 sizeof(border_colors
), 64, border_colors
);
2391 anv_device_init_trivial_batch(struct anv_device
*device
)
2393 VkResult result
= anv_device_alloc_bo(device
, 4096,
2394 ANV_BO_ALLOC_MAPPED
,
2395 0 /* explicit_address */,
2396 &device
->trivial_batch_bo
);
2397 if (result
!= VK_SUCCESS
)
2400 struct anv_batch batch
= {
2401 .start
= device
->trivial_batch_bo
->map
,
2402 .next
= device
->trivial_batch_bo
->map
,
2403 .end
= device
->trivial_batch_bo
->map
+ 4096,
2406 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
2407 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
2409 if (!device
->info
.has_llc
)
2410 gen_clflush_range(batch
.start
, batch
.next
- batch
.start
);
2415 VkResult
anv_EnumerateDeviceExtensionProperties(
2416 VkPhysicalDevice physicalDevice
,
2417 const char* pLayerName
,
2418 uint32_t* pPropertyCount
,
2419 VkExtensionProperties
* pProperties
)
2421 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
2422 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
2424 for (int i
= 0; i
< ANV_DEVICE_EXTENSION_COUNT
; i
++) {
2425 if (device
->supported_extensions
.extensions
[i
]) {
2426 vk_outarray_append(&out
, prop
) {
2427 *prop
= anv_device_extensions
[i
];
2432 return vk_outarray_status(&out
);
2436 anv_device_init_dispatch(struct anv_device
*device
)
2438 const struct anv_device_dispatch_table
*genX_table
;
2439 switch (device
->info
.gen
) {
2441 genX_table
= &gen12_device_dispatch_table
;
2444 genX_table
= &gen11_device_dispatch_table
;
2447 genX_table
= &gen10_device_dispatch_table
;
2450 genX_table
= &gen9_device_dispatch_table
;
2453 genX_table
= &gen8_device_dispatch_table
;
2456 if (device
->info
.is_haswell
)
2457 genX_table
= &gen75_device_dispatch_table
;
2459 genX_table
= &gen7_device_dispatch_table
;
2462 unreachable("unsupported gen\n");
2465 for (unsigned i
= 0; i
< ARRAY_SIZE(device
->dispatch
.entrypoints
); i
++) {
2466 /* Vulkan requires that entrypoints for extensions which have not been
2467 * enabled must not be advertised.
2469 if (!anv_device_entrypoint_is_enabled(i
, device
->instance
->app_info
.api_version
,
2470 &device
->instance
->enabled_extensions
,
2471 &device
->enabled_extensions
)) {
2472 device
->dispatch
.entrypoints
[i
] = NULL
;
2473 } else if (genX_table
->entrypoints
[i
]) {
2474 device
->dispatch
.entrypoints
[i
] = genX_table
->entrypoints
[i
];
2476 device
->dispatch
.entrypoints
[i
] =
2477 anv_device_dispatch_table
.entrypoints
[i
];
2483 vk_priority_to_gen(int priority
)
2486 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT
:
2487 return GEN_CONTEXT_LOW_PRIORITY
;
2488 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
:
2489 return GEN_CONTEXT_MEDIUM_PRIORITY
;
2490 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT
:
2491 return GEN_CONTEXT_HIGH_PRIORITY
;
2492 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT
:
2493 return GEN_CONTEXT_REALTIME_PRIORITY
;
2495 unreachable("Invalid priority");
2500 anv_device_init_hiz_clear_value_bo(struct anv_device
*device
)
2502 VkResult result
= anv_device_alloc_bo(device
, 4096,
2503 ANV_BO_ALLOC_MAPPED
,
2504 0 /* explicit_address */,
2505 &device
->hiz_clear_bo
);
2506 if (result
!= VK_SUCCESS
)
2509 union isl_color_value hiz_clear
= { .u32
= { 0, } };
2510 hiz_clear
.f32
[0] = ANV_HZ_FC_VAL
;
2512 memcpy(device
->hiz_clear_bo
->map
, hiz_clear
.u32
, sizeof(hiz_clear
.u32
));
2514 if (!device
->info
.has_llc
)
2515 gen_clflush_range(device
->hiz_clear_bo
->map
, sizeof(hiz_clear
.u32
));
2521 get_bo_from_pool(struct gen_batch_decode_bo
*ret
,
2522 struct anv_block_pool
*pool
,
2525 anv_block_pool_foreach_bo(bo
, pool
) {
2526 uint64_t bo_address
= gen_48b_address(bo
->offset
);
2527 if (address
>= bo_address
&& address
< (bo_address
+ bo
->size
)) {
2528 *ret
= (struct gen_batch_decode_bo
) {
2539 /* Finding a buffer for batch decoding */
2540 static struct gen_batch_decode_bo
2541 decode_get_bo(void *v_batch
, bool ppgtt
, uint64_t address
)
2543 struct anv_device
*device
= v_batch
;
2544 struct gen_batch_decode_bo ret_bo
= {};
2548 if (get_bo_from_pool(&ret_bo
, &device
->dynamic_state_pool
.block_pool
, address
))
2550 if (get_bo_from_pool(&ret_bo
, &device
->instruction_state_pool
.block_pool
, address
))
2552 if (get_bo_from_pool(&ret_bo
, &device
->binding_table_pool
.block_pool
, address
))
2554 if (get_bo_from_pool(&ret_bo
, &device
->surface_state_pool
.block_pool
, address
))
2557 if (!device
->cmd_buffer_being_decoded
)
2558 return (struct gen_batch_decode_bo
) { };
2560 struct anv_batch_bo
**bo
;
2562 u_vector_foreach(bo
, &device
->cmd_buffer_being_decoded
->seen_bbos
) {
2563 /* The decoder zeroes out the top 16 bits, so we need to as well */
2564 uint64_t bo_address
= (*bo
)->bo
->offset
& (~0ull >> 16);
2566 if (address
>= bo_address
&& address
< bo_address
+ (*bo
)->bo
->size
) {
2567 return (struct gen_batch_decode_bo
) {
2569 .size
= (*bo
)->bo
->size
,
2570 .map
= (*bo
)->bo
->map
,
2575 return (struct gen_batch_decode_bo
) { };
2578 struct gen_aux_map_buffer
{
2579 struct gen_buffer base
;
2580 struct anv_state state
;
2583 static struct gen_buffer
*
2584 gen_aux_map_buffer_alloc(void *driver_ctx
, uint32_t size
)
2586 struct gen_aux_map_buffer
*buf
= malloc(sizeof(struct gen_aux_map_buffer
));
2590 struct anv_device
*device
= (struct anv_device
*)driver_ctx
;
2591 assert(device
->physical
->supports_48bit_addresses
&&
2592 device
->physical
->use_softpin
);
2594 struct anv_state_pool
*pool
= &device
->dynamic_state_pool
;
2595 buf
->state
= anv_state_pool_alloc(pool
, size
, size
);
2597 buf
->base
.gpu
= pool
->block_pool
.bo
->offset
+ buf
->state
.offset
;
2598 buf
->base
.gpu_end
= buf
->base
.gpu
+ buf
->state
.alloc_size
;
2599 buf
->base
.map
= buf
->state
.map
;
2600 buf
->base
.driver_bo
= &buf
->state
;
2605 gen_aux_map_buffer_free(void *driver_ctx
, struct gen_buffer
*buffer
)
2607 struct gen_aux_map_buffer
*buf
= (struct gen_aux_map_buffer
*)buffer
;
2608 struct anv_device
*device
= (struct anv_device
*)driver_ctx
;
2609 struct anv_state_pool
*pool
= &device
->dynamic_state_pool
;
2610 anv_state_pool_free(pool
, buf
->state
);
2614 static struct gen_mapped_pinned_buffer_alloc aux_map_allocator
= {
2615 .alloc
= gen_aux_map_buffer_alloc
,
2616 .free
= gen_aux_map_buffer_free
,
2619 VkResult
anv_CreateDevice(
2620 VkPhysicalDevice physicalDevice
,
2621 const VkDeviceCreateInfo
* pCreateInfo
,
2622 const VkAllocationCallbacks
* pAllocator
,
2625 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
2627 struct anv_device
*device
;
2629 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO
);
2631 struct anv_device_extension_table enabled_extensions
= { };
2632 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
2634 for (idx
= 0; idx
< ANV_DEVICE_EXTENSION_COUNT
; idx
++) {
2635 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
2636 anv_device_extensions
[idx
].extensionName
) == 0)
2640 if (idx
>= ANV_DEVICE_EXTENSION_COUNT
)
2641 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
2643 if (!physical_device
->supported_extensions
.extensions
[idx
])
2644 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
2646 enabled_extensions
.extensions
[idx
] = true;
2649 /* Check enabled features */
2650 if (pCreateInfo
->pEnabledFeatures
) {
2651 VkPhysicalDeviceFeatures supported_features
;
2652 anv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
2653 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
2654 VkBool32
*enabled_feature
= (VkBool32
*)pCreateInfo
->pEnabledFeatures
;
2655 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
2656 for (uint32_t i
= 0; i
< num_features
; i
++) {
2657 if (enabled_feature
[i
] && !supported_feature
[i
])
2658 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
2662 /* Check requested queues and fail if we are requested to create any
2663 * queues with flags we don't support.
2665 assert(pCreateInfo
->queueCreateInfoCount
> 0);
2666 for (uint32_t i
= 0; i
< pCreateInfo
->queueCreateInfoCount
; i
++) {
2667 if (pCreateInfo
->pQueueCreateInfos
[i
].flags
!= 0)
2668 return vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2671 /* Check if client specified queue priority. */
2672 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*queue_priority
=
2673 vk_find_struct_const(pCreateInfo
->pQueueCreateInfos
[0].pNext
,
2674 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT
);
2676 VkQueueGlobalPriorityEXT priority
=
2677 queue_priority
? queue_priority
->globalPriority
:
2678 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
;
2680 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
2682 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
2684 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2686 if (INTEL_DEBUG
& DEBUG_BATCH
) {
2687 const unsigned decode_flags
=
2688 GEN_BATCH_DECODE_FULL
|
2689 ((INTEL_DEBUG
& DEBUG_COLOR
) ? GEN_BATCH_DECODE_IN_COLOR
: 0) |
2690 GEN_BATCH_DECODE_OFFSETS
|
2691 GEN_BATCH_DECODE_FLOATS
;
2693 gen_batch_decode_ctx_init(&device
->decoder_ctx
,
2694 &physical_device
->info
,
2695 stderr
, decode_flags
, NULL
,
2696 decode_get_bo
, NULL
, device
);
2699 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
2700 device
->instance
= physical_device
->instance
;
2701 device
->physical
= physical_device
;
2702 device
->chipset_id
= physical_device
->chipset_id
;
2703 device
->no_hw
= physical_device
->no_hw
;
2704 device
->_lost
= false;
2707 device
->alloc
= *pAllocator
;
2709 device
->alloc
= physical_device
->instance
->alloc
;
2711 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
2712 device
->fd
= open(physical_device
->path
, O_RDWR
| O_CLOEXEC
);
2713 if (device
->fd
== -1) {
2714 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2718 device
->context_id
= anv_gem_create_context(device
);
2719 if (device
->context_id
== -1) {
2720 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2724 result
= anv_queue_init(device
, &device
->queue
);
2725 if (result
!= VK_SUCCESS
)
2726 goto fail_context_id
;
2728 if (physical_device
->use_softpin
) {
2729 if (pthread_mutex_init(&device
->vma_mutex
, NULL
) != 0) {
2730 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2734 /* keep the page with address zero out of the allocator */
2735 util_vma_heap_init(&device
->vma_lo
,
2736 LOW_HEAP_MIN_ADDRESS
, LOW_HEAP_SIZE
);
2738 util_vma_heap_init(&device
->vma_cva
, CLIENT_VISIBLE_HEAP_MIN_ADDRESS
,
2739 CLIENT_VISIBLE_HEAP_SIZE
);
2741 /* Leave the last 4GiB out of the high vma range, so that no state
2742 * base address + size can overflow 48 bits. For more information see
2743 * the comment about Wa32bitGeneralStateOffset in anv_allocator.c
2745 util_vma_heap_init(&device
->vma_hi
, HIGH_HEAP_MIN_ADDRESS
,
2746 physical_device
->gtt_size
- (1ull << 32) -
2747 HIGH_HEAP_MIN_ADDRESS
);
2750 list_inithead(&device
->memory_objects
);
2752 /* As per spec, the driver implementation may deny requests to acquire
2753 * a priority above the default priority (MEDIUM) if the caller does not
2754 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
2757 if (physical_device
->has_context_priority
) {
2758 int err
= anv_gem_set_context_param(device
->fd
, device
->context_id
,
2759 I915_CONTEXT_PARAM_PRIORITY
,
2760 vk_priority_to_gen(priority
));
2761 if (err
!= 0 && priority
> VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
) {
2762 result
= vk_error(VK_ERROR_NOT_PERMITTED_EXT
);
2767 device
->info
= physical_device
->info
;
2768 device
->isl_dev
= physical_device
->isl_dev
;
2770 /* On Broadwell and later, we can use batch chaining to more efficiently
2771 * implement growing command buffers. Prior to Haswell, the kernel
2772 * command parser gets in the way and we have to fall back to growing
2775 device
->can_chain_batches
= device
->info
.gen
>= 8;
2777 device
->robust_buffer_access
= pCreateInfo
->pEnabledFeatures
&&
2778 pCreateInfo
->pEnabledFeatures
->robustBufferAccess
;
2779 device
->enabled_extensions
= enabled_extensions
;
2781 anv_device_init_dispatch(device
);
2783 if (pthread_mutex_init(&device
->mutex
, NULL
) != 0) {
2784 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2788 pthread_condattr_t condattr
;
2789 if (pthread_condattr_init(&condattr
) != 0) {
2790 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2793 if (pthread_condattr_setclock(&condattr
, CLOCK_MONOTONIC
) != 0) {
2794 pthread_condattr_destroy(&condattr
);
2795 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2798 if (pthread_cond_init(&device
->queue_submit
, &condattr
) != 0) {
2799 pthread_condattr_destroy(&condattr
);
2800 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2803 pthread_condattr_destroy(&condattr
);
2805 result
= anv_bo_cache_init(&device
->bo_cache
);
2806 if (result
!= VK_SUCCESS
)
2807 goto fail_queue_cond
;
2809 anv_bo_pool_init(&device
->batch_bo_pool
, device
);
2811 result
= anv_state_pool_init(&device
->dynamic_state_pool
, device
,
2812 DYNAMIC_STATE_POOL_MIN_ADDRESS
, 16384);
2813 if (result
!= VK_SUCCESS
)
2814 goto fail_batch_bo_pool
;
2816 result
= anv_state_pool_init(&device
->instruction_state_pool
, device
,
2817 INSTRUCTION_STATE_POOL_MIN_ADDRESS
, 16384);
2818 if (result
!= VK_SUCCESS
)
2819 goto fail_dynamic_state_pool
;
2821 result
= anv_state_pool_init(&device
->surface_state_pool
, device
,
2822 SURFACE_STATE_POOL_MIN_ADDRESS
, 4096);
2823 if (result
!= VK_SUCCESS
)
2824 goto fail_instruction_state_pool
;
2826 if (physical_device
->use_softpin
) {
2827 result
= anv_state_pool_init(&device
->binding_table_pool
, device
,
2828 BINDING_TABLE_POOL_MIN_ADDRESS
, 4096);
2829 if (result
!= VK_SUCCESS
)
2830 goto fail_surface_state_pool
;
2833 if (device
->info
.gen
>= 12) {
2834 device
->aux_map_ctx
= gen_aux_map_init(device
, &aux_map_allocator
,
2835 &physical_device
->info
);
2836 if (!device
->aux_map_ctx
)
2837 goto fail_binding_table_pool
;
2840 result
= anv_device_alloc_bo(device
, 4096, 0 /* flags */,
2841 0 /* explicit_address */,
2842 &device
->workaround_bo
);
2843 if (result
!= VK_SUCCESS
)
2844 goto fail_surface_aux_map_pool
;
2846 result
= anv_device_init_trivial_batch(device
);
2847 if (result
!= VK_SUCCESS
)
2848 goto fail_workaround_bo
;
2850 if (device
->info
.gen
>= 10) {
2851 result
= anv_device_init_hiz_clear_value_bo(device
);
2852 if (result
!= VK_SUCCESS
)
2853 goto fail_trivial_batch_bo
;
2856 anv_scratch_pool_init(device
, &device
->scratch_pool
);
2858 switch (device
->info
.gen
) {
2860 if (!device
->info
.is_haswell
)
2861 result
= gen7_init_device_state(device
);
2863 result
= gen75_init_device_state(device
);
2866 result
= gen8_init_device_state(device
);
2869 result
= gen9_init_device_state(device
);
2872 result
= gen10_init_device_state(device
);
2875 result
= gen11_init_device_state(device
);
2878 result
= gen12_init_device_state(device
);
2881 /* Shouldn't get here as we don't create physical devices for any other
2883 unreachable("unhandled gen");
2885 if (result
!= VK_SUCCESS
)
2886 goto fail_workaround_bo
;
2888 anv_pipeline_cache_init(&device
->default_pipeline_cache
, device
, true);
2890 anv_device_init_blorp(device
);
2892 anv_device_init_border_colors(device
);
2894 anv_device_perf_init(device
);
2896 *pDevice
= anv_device_to_handle(device
);
2901 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
2902 if (device
->info
.gen
>= 10)
2903 anv_device_release_bo(device
, device
->hiz_clear_bo
);
2904 anv_device_release_bo(device
, device
->workaround_bo
);
2905 fail_trivial_batch_bo
:
2906 anv_device_release_bo(device
, device
->trivial_batch_bo
);
2907 fail_surface_aux_map_pool
:
2908 if (device
->info
.gen
>= 12) {
2909 gen_aux_map_finish(device
->aux_map_ctx
);
2910 device
->aux_map_ctx
= NULL
;
2912 fail_binding_table_pool
:
2913 if (physical_device
->use_softpin
)
2914 anv_state_pool_finish(&device
->binding_table_pool
);
2915 fail_surface_state_pool
:
2916 anv_state_pool_finish(&device
->surface_state_pool
);
2917 fail_instruction_state_pool
:
2918 anv_state_pool_finish(&device
->instruction_state_pool
);
2919 fail_dynamic_state_pool
:
2920 anv_state_pool_finish(&device
->dynamic_state_pool
);
2922 anv_bo_pool_finish(&device
->batch_bo_pool
);
2923 anv_bo_cache_finish(&device
->bo_cache
);
2925 pthread_cond_destroy(&device
->queue_submit
);
2927 pthread_mutex_destroy(&device
->mutex
);
2929 if (physical_device
->use_softpin
) {
2930 util_vma_heap_finish(&device
->vma_hi
);
2931 util_vma_heap_finish(&device
->vma_cva
);
2932 util_vma_heap_finish(&device
->vma_lo
);
2935 anv_queue_finish(&device
->queue
);
2937 anv_gem_destroy_context(device
, device
->context_id
);
2941 vk_free(&device
->alloc
, device
);
2946 void anv_DestroyDevice(
2948 const VkAllocationCallbacks
* pAllocator
)
2950 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2955 anv_device_finish_blorp(device
);
2957 anv_pipeline_cache_finish(&device
->default_pipeline_cache
);
2959 anv_queue_finish(&device
->queue
);
2961 #ifdef HAVE_VALGRIND
2962 /* We only need to free these to prevent valgrind errors. The backing
2963 * BO will go away in a couple of lines so we don't actually leak.
2965 anv_state_pool_free(&device
->dynamic_state_pool
, device
->border_colors
);
2966 anv_state_pool_free(&device
->dynamic_state_pool
, device
->slice_hash
);
2969 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
2971 anv_device_release_bo(device
, device
->workaround_bo
);
2972 anv_device_release_bo(device
, device
->trivial_batch_bo
);
2973 if (device
->info
.gen
>= 10)
2974 anv_device_release_bo(device
, device
->hiz_clear_bo
);
2976 if (device
->info
.gen
>= 12) {
2977 gen_aux_map_finish(device
->aux_map_ctx
);
2978 device
->aux_map_ctx
= NULL
;
2981 if (device
->physical
->use_softpin
)
2982 anv_state_pool_finish(&device
->binding_table_pool
);
2983 anv_state_pool_finish(&device
->surface_state_pool
);
2984 anv_state_pool_finish(&device
->instruction_state_pool
);
2985 anv_state_pool_finish(&device
->dynamic_state_pool
);
2987 anv_bo_pool_finish(&device
->batch_bo_pool
);
2989 anv_bo_cache_finish(&device
->bo_cache
);
2991 if (device
->physical
->use_softpin
) {
2992 util_vma_heap_finish(&device
->vma_hi
);
2993 util_vma_heap_finish(&device
->vma_cva
);
2994 util_vma_heap_finish(&device
->vma_lo
);
2997 pthread_cond_destroy(&device
->queue_submit
);
2998 pthread_mutex_destroy(&device
->mutex
);
3000 anv_gem_destroy_context(device
, device
->context_id
);
3002 if (INTEL_DEBUG
& DEBUG_BATCH
)
3003 gen_batch_decode_ctx_finish(&device
->decoder_ctx
);
3007 vk_free(&device
->alloc
, device
);
3010 VkResult
anv_EnumerateInstanceLayerProperties(
3011 uint32_t* pPropertyCount
,
3012 VkLayerProperties
* pProperties
)
3014 if (pProperties
== NULL
) {
3015 *pPropertyCount
= 0;
3019 /* None supported at this time */
3020 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
3023 VkResult
anv_EnumerateDeviceLayerProperties(
3024 VkPhysicalDevice physicalDevice
,
3025 uint32_t* pPropertyCount
,
3026 VkLayerProperties
* pProperties
)
3028 if (pProperties
== NULL
) {
3029 *pPropertyCount
= 0;
3033 /* None supported at this time */
3034 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
3037 void anv_GetDeviceQueue(
3039 uint32_t queueNodeIndex
,
3040 uint32_t queueIndex
,
3043 const VkDeviceQueueInfo2 info
= {
3044 .sType
= VK_STRUCTURE_TYPE_DEVICE_QUEUE_INFO_2
,
3047 .queueFamilyIndex
= queueNodeIndex
,
3048 .queueIndex
= queueIndex
,
3051 anv_GetDeviceQueue2(_device
, &info
, pQueue
);
3054 void anv_GetDeviceQueue2(
3056 const VkDeviceQueueInfo2
* pQueueInfo
,
3059 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3061 assert(pQueueInfo
->queueIndex
== 0);
3063 if (pQueueInfo
->flags
== device
->queue
.flags
)
3064 *pQueue
= anv_queue_to_handle(&device
->queue
);
3070 _anv_device_set_lost(struct anv_device
*device
,
3071 const char *file
, int line
,
3072 const char *msg
, ...)
3077 p_atomic_inc(&device
->_lost
);
3080 err
= __vk_errorv(device
->instance
, device
,
3081 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT
,
3082 VK_ERROR_DEVICE_LOST
, file
, line
, msg
, ap
);
3085 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
3092 _anv_queue_set_lost(struct anv_queue
*queue
,
3093 const char *file
, int line
,
3094 const char *msg
, ...)
3099 p_atomic_inc(&queue
->device
->_lost
);
3102 err
= __vk_errorv(queue
->device
->instance
, queue
->device
,
3103 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT
,
3104 VK_ERROR_DEVICE_LOST
, file
, line
, msg
, ap
);
3107 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
3114 anv_device_query_status(struct anv_device
*device
)
3116 /* This isn't likely as most of the callers of this function already check
3117 * for it. However, it doesn't hurt to check and it potentially lets us
3120 if (anv_device_is_lost(device
))
3121 return VK_ERROR_DEVICE_LOST
;
3123 uint32_t active
, pending
;
3124 int ret
= anv_gem_gpu_get_reset_stats(device
, &active
, &pending
);
3126 /* We don't know the real error. */
3127 return anv_device_set_lost(device
, "get_reset_stats failed: %m");
3131 return anv_device_set_lost(device
, "GPU hung on one of our command buffers");
3132 } else if (pending
) {
3133 return anv_device_set_lost(device
, "GPU hung with commands in-flight");
3140 anv_device_bo_busy(struct anv_device
*device
, struct anv_bo
*bo
)
3142 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
3143 * Other usages of the BO (such as on different hardware) will not be
3144 * flagged as "busy" by this ioctl. Use with care.
3146 int ret
= anv_gem_busy(device
, bo
->gem_handle
);
3148 return VK_NOT_READY
;
3149 } else if (ret
== -1) {
3150 /* We don't know the real error. */
3151 return anv_device_set_lost(device
, "gem wait failed: %m");
3154 /* Query for device status after the busy call. If the BO we're checking
3155 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
3156 * client because it clearly doesn't have valid data. Yes, this most
3157 * likely means an ioctl, but we just did an ioctl to query the busy status
3158 * so it's no great loss.
3160 return anv_device_query_status(device
);
3164 anv_device_wait(struct anv_device
*device
, struct anv_bo
*bo
,
3167 int ret
= anv_gem_wait(device
, bo
->gem_handle
, &timeout
);
3168 if (ret
== -1 && errno
== ETIME
) {
3170 } else if (ret
== -1) {
3171 /* We don't know the real error. */
3172 return anv_device_set_lost(device
, "gem wait failed: %m");
3175 /* Query for device status after the wait. If the BO we're waiting on got
3176 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
3177 * because it clearly doesn't have valid data. Yes, this most likely means
3178 * an ioctl, but we just did an ioctl to wait so it's no great loss.
3180 return anv_device_query_status(device
);
3183 VkResult
anv_DeviceWaitIdle(
3186 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3188 if (anv_device_is_lost(device
))
3189 return VK_ERROR_DEVICE_LOST
;
3191 return anv_queue_submit_simple_batch(&device
->queue
, NULL
);
3195 anv_vma_alloc(struct anv_device
*device
, struct anv_bo
*bo
,
3196 uint64_t client_address
)
3198 const struct gen_device_info
*devinfo
= &device
->info
;
3199 /* Gen12 CCS surface addresses need to be 64K aligned. We have no way of
3200 * telling what this allocation is for so pick the largest alignment.
3202 const uint32_t vma_alignment
=
3203 devinfo
->gen
>= 12 ? (64 * 1024) : (4 * 1024);
3205 if (!(bo
->flags
& EXEC_OBJECT_PINNED
)) {
3206 assert(!(bo
->has_client_visible_address
));
3210 pthread_mutex_lock(&device
->vma_mutex
);
3214 if (bo
->has_client_visible_address
) {
3215 assert(bo
->flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
);
3216 if (client_address
) {
3217 if (util_vma_heap_alloc_addr(&device
->vma_cva
,
3218 client_address
, bo
->size
)) {
3219 bo
->offset
= gen_canonical_address(client_address
);
3223 util_vma_heap_alloc(&device
->vma_cva
, bo
->size
, vma_alignment
);
3225 bo
->offset
= gen_canonical_address(addr
);
3226 assert(addr
== gen_48b_address(bo
->offset
));
3229 /* We don't want to fall back to other heaps */
3233 assert(client_address
== 0);
3235 if (bo
->flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
) {
3237 util_vma_heap_alloc(&device
->vma_hi
, bo
->size
, vma_alignment
);
3239 bo
->offset
= gen_canonical_address(addr
);
3240 assert(addr
== gen_48b_address(bo
->offset
));
3244 if (bo
->offset
== 0) {
3246 util_vma_heap_alloc(&device
->vma_lo
, bo
->size
, vma_alignment
);
3248 bo
->offset
= gen_canonical_address(addr
);
3249 assert(addr
== gen_48b_address(bo
->offset
));
3254 pthread_mutex_unlock(&device
->vma_mutex
);
3256 return bo
->offset
!= 0;
3260 anv_vma_free(struct anv_device
*device
, struct anv_bo
*bo
)
3262 if (!(bo
->flags
& EXEC_OBJECT_PINNED
))
3265 const uint64_t addr_48b
= gen_48b_address(bo
->offset
);
3267 pthread_mutex_lock(&device
->vma_mutex
);
3269 if (addr_48b
>= LOW_HEAP_MIN_ADDRESS
&&
3270 addr_48b
<= LOW_HEAP_MAX_ADDRESS
) {
3271 util_vma_heap_free(&device
->vma_lo
, addr_48b
, bo
->size
);
3272 } else if (addr_48b
>= CLIENT_VISIBLE_HEAP_MIN_ADDRESS
&&
3273 addr_48b
<= CLIENT_VISIBLE_HEAP_MAX_ADDRESS
) {
3274 util_vma_heap_free(&device
->vma_cva
, addr_48b
, bo
->size
);
3276 assert(addr_48b
>= HIGH_HEAP_MIN_ADDRESS
);
3277 util_vma_heap_free(&device
->vma_hi
, addr_48b
, bo
->size
);
3280 pthread_mutex_unlock(&device
->vma_mutex
);
3285 VkResult
anv_AllocateMemory(
3287 const VkMemoryAllocateInfo
* pAllocateInfo
,
3288 const VkAllocationCallbacks
* pAllocator
,
3289 VkDeviceMemory
* pMem
)
3291 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3292 struct anv_physical_device
*pdevice
= device
->physical
;
3293 struct anv_device_memory
*mem
;
3294 VkResult result
= VK_SUCCESS
;
3296 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
3298 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
3299 assert(pAllocateInfo
->allocationSize
> 0);
3301 VkDeviceSize aligned_alloc_size
=
3302 align_u64(pAllocateInfo
->allocationSize
, 4096);
3304 if (aligned_alloc_size
> MAX_MEMORY_ALLOCATION_SIZE
)
3305 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
3307 assert(pAllocateInfo
->memoryTypeIndex
< pdevice
->memory
.type_count
);
3308 struct anv_memory_type
*mem_type
=
3309 &pdevice
->memory
.types
[pAllocateInfo
->memoryTypeIndex
];
3310 assert(mem_type
->heapIndex
< pdevice
->memory
.heap_count
);
3311 struct anv_memory_heap
*mem_heap
=
3312 &pdevice
->memory
.heaps
[mem_type
->heapIndex
];
3314 uint64_t mem_heap_used
= p_atomic_read(&mem_heap
->used
);
3315 if (mem_heap_used
+ aligned_alloc_size
> mem_heap
->size
)
3316 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
3318 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
3319 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3321 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3323 assert(pAllocateInfo
->memoryTypeIndex
< pdevice
->memory
.type_count
);
3324 mem
->type
= mem_type
;
3328 mem
->host_ptr
= NULL
;
3330 enum anv_bo_alloc_flags alloc_flags
= 0;
3332 const VkExportMemoryAllocateInfo
*export_info
= NULL
;
3333 const VkImportAndroidHardwareBufferInfoANDROID
*ahw_import_info
= NULL
;
3334 const VkImportMemoryFdInfoKHR
*fd_info
= NULL
;
3335 const VkImportMemoryHostPointerInfoEXT
*host_ptr_info
= NULL
;
3336 const VkMemoryDedicatedAllocateInfo
*dedicated_info
= NULL
;
3337 VkMemoryAllocateFlags vk_flags
= 0;
3338 uint64_t client_address
= 0;
3340 vk_foreach_struct_const(ext
, pAllocateInfo
->pNext
) {
3341 switch (ext
->sType
) {
3342 case VK_STRUCTURE_TYPE_EXPORT_MEMORY_ALLOCATE_INFO
:
3343 export_info
= (void *)ext
;
3346 case VK_STRUCTURE_TYPE_IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID
:
3347 ahw_import_info
= (void *)ext
;
3350 case VK_STRUCTURE_TYPE_IMPORT_MEMORY_FD_INFO_KHR
:
3351 fd_info
= (void *)ext
;
3354 case VK_STRUCTURE_TYPE_IMPORT_MEMORY_HOST_POINTER_INFO_EXT
:
3355 host_ptr_info
= (void *)ext
;
3358 case VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_FLAGS_INFO
: {
3359 const VkMemoryAllocateFlagsInfo
*flags_info
= (void *)ext
;
3360 vk_flags
= flags_info
->flags
;
3364 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_ALLOCATE_INFO
:
3365 dedicated_info
= (void *)ext
;
3368 case VK_STRUCTURE_TYPE_MEMORY_OPAQUE_CAPTURE_ADDRESS_ALLOCATE_INFO_KHR
: {
3369 const VkMemoryOpaqueCaptureAddressAllocateInfoKHR
*addr_info
=
3370 (const VkMemoryOpaqueCaptureAddressAllocateInfoKHR
*)ext
;
3371 client_address
= addr_info
->opaqueCaptureAddress
;
3376 anv_debug_ignored_stype(ext
->sType
);
3381 if (vk_flags
& VK_MEMORY_ALLOCATE_DEVICE_ADDRESS_BIT_KHR
)
3382 alloc_flags
|= ANV_BO_ALLOC_CLIENT_VISIBLE_ADDRESS
;
3384 /* Check if we need to support Android HW buffer export. If so,
3385 * create AHardwareBuffer and import memory from it.
3387 bool android_export
= false;
3388 if (export_info
&& export_info
->handleTypes
&
3389 VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID
)
3390 android_export
= true;
3392 if (ahw_import_info
) {
3393 result
= anv_import_ahw_memory(_device
, mem
, ahw_import_info
);
3394 if (result
!= VK_SUCCESS
)
3398 } else if (android_export
) {
3399 result
= anv_create_ahw_memory(_device
, mem
, pAllocateInfo
);
3400 if (result
!= VK_SUCCESS
)
3403 const VkImportAndroidHardwareBufferInfoANDROID import_info
= {
3406 result
= anv_import_ahw_memory(_device
, mem
, &import_info
);
3407 if (result
!= VK_SUCCESS
)
3413 /* The Vulkan spec permits handleType to be 0, in which case the struct is
3416 if (fd_info
&& fd_info
->handleType
) {
3417 /* At the moment, we support only the below handle types. */
3418 assert(fd_info
->handleType
==
3419 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
3420 fd_info
->handleType
==
3421 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
3423 result
= anv_device_import_bo(device
, fd_info
->fd
, alloc_flags
,
3424 client_address
, &mem
->bo
);
3425 if (result
!= VK_SUCCESS
)
3428 VkDeviceSize aligned_alloc_size
=
3429 align_u64(pAllocateInfo
->allocationSize
, 4096);
3431 /* For security purposes, we reject importing the bo if it's smaller
3432 * than the requested allocation size. This prevents a malicious client
3433 * from passing a buffer to a trusted client, lying about the size, and
3434 * telling the trusted client to try and texture from an image that goes
3435 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
3436 * in the trusted client. The trusted client can protect itself against
3437 * this sort of attack but only if it can trust the buffer size.
3439 if (mem
->bo
->size
< aligned_alloc_size
) {
3440 result
= vk_errorf(device
->instance
, device
,
3441 VK_ERROR_INVALID_EXTERNAL_HANDLE
,
3442 "aligned allocationSize too large for "
3443 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT: "
3444 "%"PRIu64
"B > %"PRIu64
"B",
3445 aligned_alloc_size
, mem
->bo
->size
);
3446 anv_device_release_bo(device
, mem
->bo
);
3450 /* From the Vulkan spec:
3452 * "Importing memory from a file descriptor transfers ownership of
3453 * the file descriptor from the application to the Vulkan
3454 * implementation. The application must not perform any operations on
3455 * the file descriptor after a successful import."
3457 * If the import fails, we leave the file descriptor open.
3463 if (host_ptr_info
&& host_ptr_info
->handleType
) {
3464 if (host_ptr_info
->handleType
==
3465 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_MAPPED_FOREIGN_MEMORY_BIT_EXT
) {
3466 result
= vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
3470 assert(host_ptr_info
->handleType
==
3471 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
);
3473 result
= anv_device_import_bo_from_host_ptr(device
,
3474 host_ptr_info
->pHostPointer
,
3475 pAllocateInfo
->allocationSize
,
3479 if (result
!= VK_SUCCESS
)
3482 mem
->host_ptr
= host_ptr_info
->pHostPointer
;
3486 /* Regular allocate (not importing memory). */
3488 if (export_info
&& export_info
->handleTypes
)
3489 alloc_flags
|= ANV_BO_ALLOC_EXTERNAL
;
3491 result
= anv_device_alloc_bo(device
, pAllocateInfo
->allocationSize
,
3492 alloc_flags
, client_address
, &mem
->bo
);
3493 if (result
!= VK_SUCCESS
)
3496 if (dedicated_info
&& dedicated_info
->image
!= VK_NULL_HANDLE
) {
3497 ANV_FROM_HANDLE(anv_image
, image
, dedicated_info
->image
);
3499 /* Some legacy (non-modifiers) consumers need the tiling to be set on
3500 * the BO. In this case, we have a dedicated allocation.
3502 if (image
->needs_set_tiling
) {
3503 const uint32_t i915_tiling
=
3504 isl_tiling_to_i915_tiling(image
->planes
[0].surface
.isl
.tiling
);
3505 int ret
= anv_gem_set_tiling(device
, mem
->bo
->gem_handle
,
3506 image
->planes
[0].surface
.isl
.row_pitch_B
,
3509 anv_device_release_bo(device
, mem
->bo
);
3510 result
= vk_errorf(device
->instance
, NULL
,
3511 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
3512 "failed to set BO tiling: %m");
3519 mem_heap_used
= p_atomic_add_return(&mem_heap
->used
, mem
->bo
->size
);
3520 if (mem_heap_used
> mem_heap
->size
) {
3521 p_atomic_add(&mem_heap
->used
, -mem
->bo
->size
);
3522 anv_device_release_bo(device
, mem
->bo
);
3523 result
= vk_errorf(device
->instance
, NULL
,
3524 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
3525 "Out of heap memory");
3529 pthread_mutex_lock(&device
->mutex
);
3530 list_addtail(&mem
->link
, &device
->memory_objects
);
3531 pthread_mutex_unlock(&device
->mutex
);
3533 *pMem
= anv_device_memory_to_handle(mem
);
3538 vk_free2(&device
->alloc
, pAllocator
, mem
);
3543 VkResult
anv_GetMemoryFdKHR(
3545 const VkMemoryGetFdInfoKHR
* pGetFdInfo
,
3548 ANV_FROM_HANDLE(anv_device
, dev
, device_h
);
3549 ANV_FROM_HANDLE(anv_device_memory
, mem
, pGetFdInfo
->memory
);
3551 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
3553 assert(pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
3554 pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
3556 return anv_device_export_bo(dev
, mem
->bo
, pFd
);
3559 VkResult
anv_GetMemoryFdPropertiesKHR(
3561 VkExternalMemoryHandleTypeFlagBits handleType
,
3563 VkMemoryFdPropertiesKHR
* pMemoryFdProperties
)
3565 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3567 switch (handleType
) {
3568 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
:
3569 /* dma-buf can be imported as any memory type */
3570 pMemoryFdProperties
->memoryTypeBits
=
3571 (1 << device
->physical
->memory
.type_count
) - 1;
3575 /* The valid usage section for this function says:
3577 * "handleType must not be one of the handle types defined as
3580 * So opaque handle types fall into the default "unsupported" case.
3582 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
3586 VkResult
anv_GetMemoryHostPointerPropertiesEXT(
3588 VkExternalMemoryHandleTypeFlagBits handleType
,
3589 const void* pHostPointer
,
3590 VkMemoryHostPointerPropertiesEXT
* pMemoryHostPointerProperties
)
3592 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3594 assert(pMemoryHostPointerProperties
->sType
==
3595 VK_STRUCTURE_TYPE_MEMORY_HOST_POINTER_PROPERTIES_EXT
);
3597 switch (handleType
) {
3598 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
:
3599 /* Host memory can be imported as any memory type. */
3600 pMemoryHostPointerProperties
->memoryTypeBits
=
3601 (1ull << device
->physical
->memory
.type_count
) - 1;
3606 return VK_ERROR_INVALID_EXTERNAL_HANDLE
;
3610 void anv_FreeMemory(
3612 VkDeviceMemory _mem
,
3613 const VkAllocationCallbacks
* pAllocator
)
3615 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3616 ANV_FROM_HANDLE(anv_device_memory
, mem
, _mem
);
3621 pthread_mutex_lock(&device
->mutex
);
3622 list_del(&mem
->link
);
3623 pthread_mutex_unlock(&device
->mutex
);
3626 anv_UnmapMemory(_device
, _mem
);
3628 p_atomic_add(&device
->physical
->memory
.heaps
[mem
->type
->heapIndex
].used
,
3631 anv_device_release_bo(device
, mem
->bo
);
3633 #if defined(ANDROID) && ANDROID_API_LEVEL >= 26
3635 AHardwareBuffer_release(mem
->ahw
);
3638 vk_free2(&device
->alloc
, pAllocator
, mem
);
3641 VkResult
anv_MapMemory(
3643 VkDeviceMemory _memory
,
3644 VkDeviceSize offset
,
3646 VkMemoryMapFlags flags
,
3649 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3650 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
3657 if (mem
->host_ptr
) {
3658 *ppData
= mem
->host_ptr
+ offset
;
3662 if (size
== VK_WHOLE_SIZE
)
3663 size
= mem
->bo
->size
- offset
;
3665 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
3667 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
3668 * assert(size != 0);
3669 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
3670 * equal to the size of the memory minus offset
3673 assert(offset
+ size
<= mem
->bo
->size
);
3675 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
3676 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
3677 * at a time is valid. We could just mmap up front and return an offset
3678 * pointer here, but that may exhaust virtual memory on 32 bit
3681 uint32_t gem_flags
= 0;
3683 if (!device
->info
.has_llc
&&
3684 (mem
->type
->propertyFlags
& VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
))
3685 gem_flags
|= I915_MMAP_WC
;
3687 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
3688 uint64_t map_offset
= offset
& ~4095ull;
3689 assert(offset
>= map_offset
);
3690 uint64_t map_size
= (offset
+ size
) - map_offset
;
3692 /* Let's map whole pages */
3693 map_size
= align_u64(map_size
, 4096);
3695 void *map
= anv_gem_mmap(device
, mem
->bo
->gem_handle
,
3696 map_offset
, map_size
, gem_flags
);
3697 if (map
== MAP_FAILED
)
3698 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
3701 mem
->map_size
= map_size
;
3703 *ppData
= mem
->map
+ (offset
- map_offset
);
3708 void anv_UnmapMemory(
3710 VkDeviceMemory _memory
)
3712 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
3714 if (mem
== NULL
|| mem
->host_ptr
)
3717 anv_gem_munmap(mem
->map
, mem
->map_size
);
3724 clflush_mapped_ranges(struct anv_device
*device
,
3726 const VkMappedMemoryRange
*ranges
)
3728 for (uint32_t i
= 0; i
< count
; i
++) {
3729 ANV_FROM_HANDLE(anv_device_memory
, mem
, ranges
[i
].memory
);
3730 if (ranges
[i
].offset
>= mem
->map_size
)
3733 gen_clflush_range(mem
->map
+ ranges
[i
].offset
,
3734 MIN2(ranges
[i
].size
, mem
->map_size
- ranges
[i
].offset
));
3738 VkResult
anv_FlushMappedMemoryRanges(
3740 uint32_t memoryRangeCount
,
3741 const VkMappedMemoryRange
* pMemoryRanges
)
3743 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3745 if (device
->info
.has_llc
)
3748 /* Make sure the writes we're flushing have landed. */
3749 __builtin_ia32_mfence();
3751 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
3756 VkResult
anv_InvalidateMappedMemoryRanges(
3758 uint32_t memoryRangeCount
,
3759 const VkMappedMemoryRange
* pMemoryRanges
)
3761 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3763 if (device
->info
.has_llc
)
3766 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
3768 /* Make sure no reads get moved up above the invalidate. */
3769 __builtin_ia32_mfence();
3774 void anv_GetBufferMemoryRequirements(
3777 VkMemoryRequirements
* pMemoryRequirements
)
3779 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
3780 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3782 /* The Vulkan spec (git aaed022) says:
3784 * memoryTypeBits is a bitfield and contains one bit set for every
3785 * supported memory type for the resource. The bit `1<<i` is set if and
3786 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3787 * structure for the physical device is supported.
3789 uint32_t memory_types
= (1ull << device
->physical
->memory
.type_count
) - 1;
3791 /* Base alignment requirement of a cache line */
3792 uint32_t alignment
= 16;
3794 /* We need an alignment of 32 for pushing UBOs */
3795 if (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
)
3796 alignment
= MAX2(alignment
, 32);
3798 pMemoryRequirements
->size
= buffer
->size
;
3799 pMemoryRequirements
->alignment
= alignment
;
3801 /* Storage and Uniform buffers should have their size aligned to
3802 * 32-bits to avoid boundary checks when last DWord is not complete.
3803 * This would ensure that not internal padding would be needed for
3806 if (device
->robust_buffer_access
&&
3807 (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
||
3808 buffer
->usage
& VK_BUFFER_USAGE_STORAGE_BUFFER_BIT
))
3809 pMemoryRequirements
->size
= align_u64(buffer
->size
, 4);
3811 pMemoryRequirements
->memoryTypeBits
= memory_types
;
3814 void anv_GetBufferMemoryRequirements2(
3816 const VkBufferMemoryRequirementsInfo2
* pInfo
,
3817 VkMemoryRequirements2
* pMemoryRequirements
)
3819 anv_GetBufferMemoryRequirements(_device
, pInfo
->buffer
,
3820 &pMemoryRequirements
->memoryRequirements
);
3822 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3823 switch (ext
->sType
) {
3824 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
3825 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
3826 requirements
->prefersDedicatedAllocation
= false;
3827 requirements
->requiresDedicatedAllocation
= false;
3832 anv_debug_ignored_stype(ext
->sType
);
3838 void anv_GetImageMemoryRequirements(
3841 VkMemoryRequirements
* pMemoryRequirements
)
3843 ANV_FROM_HANDLE(anv_image
, image
, _image
);
3844 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3846 /* The Vulkan spec (git aaed022) says:
3848 * memoryTypeBits is a bitfield and contains one bit set for every
3849 * supported memory type for the resource. The bit `1<<i` is set if and
3850 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3851 * structure for the physical device is supported.
3853 * All types are currently supported for images.
3855 uint32_t memory_types
= (1ull << device
->physical
->memory
.type_count
) - 1;
3857 /* We must have image allocated or imported at this point. According to the
3858 * specification, external images must have been bound to memory before
3859 * calling GetImageMemoryRequirements.
3861 assert(image
->size
> 0);
3863 pMemoryRequirements
->size
= image
->size
;
3864 pMemoryRequirements
->alignment
= image
->alignment
;
3865 pMemoryRequirements
->memoryTypeBits
= memory_types
;
3868 void anv_GetImageMemoryRequirements2(
3870 const VkImageMemoryRequirementsInfo2
* pInfo
,
3871 VkMemoryRequirements2
* pMemoryRequirements
)
3873 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3874 ANV_FROM_HANDLE(anv_image
, image
, pInfo
->image
);
3876 anv_GetImageMemoryRequirements(_device
, pInfo
->image
,
3877 &pMemoryRequirements
->memoryRequirements
);
3879 vk_foreach_struct_const(ext
, pInfo
->pNext
) {
3880 switch (ext
->sType
) {
3881 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO
: {
3882 const VkImagePlaneMemoryRequirementsInfo
*plane_reqs
=
3883 (const VkImagePlaneMemoryRequirementsInfo
*) ext
;
3884 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
,
3885 plane_reqs
->planeAspect
);
3887 assert(image
->planes
[plane
].offset
== 0);
3889 /* The Vulkan spec (git aaed022) says:
3891 * memoryTypeBits is a bitfield and contains one bit set for every
3892 * supported memory type for the resource. The bit `1<<i` is set
3893 * if and only if the memory type `i` in the
3894 * VkPhysicalDeviceMemoryProperties structure for the physical
3895 * device is supported.
3897 * All types are currently supported for images.
3899 pMemoryRequirements
->memoryRequirements
.memoryTypeBits
=
3900 (1ull << device
->physical
->memory
.type_count
) - 1;
3902 /* We must have image allocated or imported at this point. According to the
3903 * specification, external images must have been bound to memory before
3904 * calling GetImageMemoryRequirements.
3906 assert(image
->planes
[plane
].size
> 0);
3908 pMemoryRequirements
->memoryRequirements
.size
= image
->planes
[plane
].size
;
3909 pMemoryRequirements
->memoryRequirements
.alignment
=
3910 image
->planes
[plane
].alignment
;
3915 anv_debug_ignored_stype(ext
->sType
);
3920 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3921 switch (ext
->sType
) {
3922 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
3923 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
3924 if (image
->needs_set_tiling
|| image
->external_format
) {
3925 /* If we need to set the tiling for external consumers, we need a
3926 * dedicated allocation.
3928 * See also anv_AllocateMemory.
3930 requirements
->prefersDedicatedAllocation
= true;
3931 requirements
->requiresDedicatedAllocation
= true;
3933 requirements
->prefersDedicatedAllocation
= false;
3934 requirements
->requiresDedicatedAllocation
= false;
3940 anv_debug_ignored_stype(ext
->sType
);
3946 void anv_GetImageSparseMemoryRequirements(
3949 uint32_t* pSparseMemoryRequirementCount
,
3950 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
3952 *pSparseMemoryRequirementCount
= 0;
3955 void anv_GetImageSparseMemoryRequirements2(
3957 const VkImageSparseMemoryRequirementsInfo2
* pInfo
,
3958 uint32_t* pSparseMemoryRequirementCount
,
3959 VkSparseImageMemoryRequirements2
* pSparseMemoryRequirements
)
3961 *pSparseMemoryRequirementCount
= 0;
3964 void anv_GetDeviceMemoryCommitment(
3966 VkDeviceMemory memory
,
3967 VkDeviceSize
* pCommittedMemoryInBytes
)
3969 *pCommittedMemoryInBytes
= 0;
3973 anv_bind_buffer_memory(const VkBindBufferMemoryInfo
*pBindInfo
)
3975 ANV_FROM_HANDLE(anv_device_memory
, mem
, pBindInfo
->memory
);
3976 ANV_FROM_HANDLE(anv_buffer
, buffer
, pBindInfo
->buffer
);
3978 assert(pBindInfo
->sType
== VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
);
3981 buffer
->address
= (struct anv_address
) {
3983 .offset
= pBindInfo
->memoryOffset
,
3986 buffer
->address
= ANV_NULL_ADDRESS
;
3990 VkResult
anv_BindBufferMemory(
3993 VkDeviceMemory memory
,
3994 VkDeviceSize memoryOffset
)
3996 anv_bind_buffer_memory(
3997 &(VkBindBufferMemoryInfo
) {
3998 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
,
4001 .memoryOffset
= memoryOffset
,
4007 VkResult
anv_BindBufferMemory2(
4009 uint32_t bindInfoCount
,
4010 const VkBindBufferMemoryInfo
* pBindInfos
)
4012 for (uint32_t i
= 0; i
< bindInfoCount
; i
++)
4013 anv_bind_buffer_memory(&pBindInfos
[i
]);
4018 VkResult
anv_QueueBindSparse(
4020 uint32_t bindInfoCount
,
4021 const VkBindSparseInfo
* pBindInfo
,
4024 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
4025 if (anv_device_is_lost(queue
->device
))
4026 return VK_ERROR_DEVICE_LOST
;
4028 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
4033 VkResult
anv_CreateEvent(
4035 const VkEventCreateInfo
* pCreateInfo
,
4036 const VkAllocationCallbacks
* pAllocator
,
4039 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4040 struct anv_state state
;
4041 struct anv_event
*event
;
4043 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_EVENT_CREATE_INFO
);
4045 state
= anv_state_pool_alloc(&device
->dynamic_state_pool
,
4048 event
->state
= state
;
4049 event
->semaphore
= VK_EVENT_RESET
;
4051 if (!device
->info
.has_llc
) {
4052 /* Make sure the writes we're flushing have landed. */
4053 __builtin_ia32_mfence();
4054 __builtin_ia32_clflush(event
);
4057 *pEvent
= anv_event_to_handle(event
);
4062 void anv_DestroyEvent(
4065 const VkAllocationCallbacks
* pAllocator
)
4067 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4068 ANV_FROM_HANDLE(anv_event
, event
, _event
);
4073 anv_state_pool_free(&device
->dynamic_state_pool
, event
->state
);
4076 VkResult
anv_GetEventStatus(
4080 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4081 ANV_FROM_HANDLE(anv_event
, event
, _event
);
4083 if (anv_device_is_lost(device
))
4084 return VK_ERROR_DEVICE_LOST
;
4086 if (!device
->info
.has_llc
) {
4087 /* Invalidate read cache before reading event written by GPU. */
4088 __builtin_ia32_clflush(event
);
4089 __builtin_ia32_mfence();
4093 return event
->semaphore
;
4096 VkResult
anv_SetEvent(
4100 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4101 ANV_FROM_HANDLE(anv_event
, event
, _event
);
4103 event
->semaphore
= VK_EVENT_SET
;
4105 if (!device
->info
.has_llc
) {
4106 /* Make sure the writes we're flushing have landed. */
4107 __builtin_ia32_mfence();
4108 __builtin_ia32_clflush(event
);
4114 VkResult
anv_ResetEvent(
4118 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4119 ANV_FROM_HANDLE(anv_event
, event
, _event
);
4121 event
->semaphore
= VK_EVENT_RESET
;
4123 if (!device
->info
.has_llc
) {
4124 /* Make sure the writes we're flushing have landed. */
4125 __builtin_ia32_mfence();
4126 __builtin_ia32_clflush(event
);
4134 VkResult
anv_CreateBuffer(
4136 const VkBufferCreateInfo
* pCreateInfo
,
4137 const VkAllocationCallbacks
* pAllocator
,
4140 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4141 struct anv_buffer
*buffer
;
4143 /* Don't allow creating buffers bigger than our address space. The real
4144 * issue here is that we may align up the buffer size and we don't want
4145 * doing so to cause roll-over. However, no one has any business
4146 * allocating a buffer larger than our GTT size.
4148 if (pCreateInfo
->size
> device
->physical
->gtt_size
)
4149 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
4151 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
4153 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
4154 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
4156 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
4158 buffer
->size
= pCreateInfo
->size
;
4159 buffer
->usage
= pCreateInfo
->usage
;
4160 buffer
->address
= ANV_NULL_ADDRESS
;
4162 *pBuffer
= anv_buffer_to_handle(buffer
);
4167 void anv_DestroyBuffer(
4170 const VkAllocationCallbacks
* pAllocator
)
4172 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4173 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
4178 vk_free2(&device
->alloc
, pAllocator
, buffer
);
4181 VkDeviceAddress
anv_GetBufferDeviceAddress(
4183 const VkBufferDeviceAddressInfoKHR
* pInfo
)
4185 ANV_FROM_HANDLE(anv_buffer
, buffer
, pInfo
->buffer
);
4187 assert(!anv_address_is_null(buffer
->address
));
4188 assert(buffer
->address
.bo
->flags
& EXEC_OBJECT_PINNED
);
4190 return anv_address_physical(buffer
->address
);
4193 uint64_t anv_GetBufferOpaqueCaptureAddress(
4195 const VkBufferDeviceAddressInfoKHR
* pInfo
)
4200 uint64_t anv_GetDeviceMemoryOpaqueCaptureAddress(
4202 const VkDeviceMemoryOpaqueCaptureAddressInfoKHR
* pInfo
)
4204 ANV_FROM_HANDLE(anv_device_memory
, memory
, pInfo
->memory
);
4206 assert(memory
->bo
->flags
& EXEC_OBJECT_PINNED
);
4207 assert(memory
->bo
->has_client_visible_address
);
4209 return gen_48b_address(memory
->bo
->offset
);
4213 anv_fill_buffer_surface_state(struct anv_device
*device
, struct anv_state state
,
4214 enum isl_format format
,
4215 struct anv_address address
,
4216 uint32_t range
, uint32_t stride
)
4218 isl_buffer_fill_state(&device
->isl_dev
, state
.map
,
4219 .address
= anv_address_physical(address
),
4220 .mocs
= device
->isl_dev
.mocs
.internal
,
4223 .swizzle
= ISL_SWIZZLE_IDENTITY
,
4224 .stride_B
= stride
);
4227 void anv_DestroySampler(
4230 const VkAllocationCallbacks
* pAllocator
)
4232 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4233 ANV_FROM_HANDLE(anv_sampler
, sampler
, _sampler
);
4238 if (sampler
->bindless_state
.map
) {
4239 anv_state_pool_free(&device
->dynamic_state_pool
,
4240 sampler
->bindless_state
);
4243 vk_free2(&device
->alloc
, pAllocator
, sampler
);
4246 VkResult
anv_CreateFramebuffer(
4248 const VkFramebufferCreateInfo
* pCreateInfo
,
4249 const VkAllocationCallbacks
* pAllocator
,
4250 VkFramebuffer
* pFramebuffer
)
4252 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4253 struct anv_framebuffer
*framebuffer
;
4255 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
4257 size_t size
= sizeof(*framebuffer
);
4259 /* VK_KHR_imageless_framebuffer extension says:
4261 * If flags includes VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR,
4262 * parameter pAttachments is ignored.
4264 if (!(pCreateInfo
->flags
& VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR
)) {
4265 size
+= sizeof(struct anv_image_view
*) * pCreateInfo
->attachmentCount
;
4266 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
4267 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
4268 if (framebuffer
== NULL
)
4269 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
4271 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
4272 ANV_FROM_HANDLE(anv_image_view
, iview
, pCreateInfo
->pAttachments
[i
]);
4273 framebuffer
->attachments
[i
] = iview
;
4275 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
4277 assert(device
->enabled_extensions
.KHR_imageless_framebuffer
);
4278 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
4279 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
4280 if (framebuffer
== NULL
)
4281 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
4283 framebuffer
->attachment_count
= 0;
4286 framebuffer
->width
= pCreateInfo
->width
;
4287 framebuffer
->height
= pCreateInfo
->height
;
4288 framebuffer
->layers
= pCreateInfo
->layers
;
4290 *pFramebuffer
= anv_framebuffer_to_handle(framebuffer
);
4295 void anv_DestroyFramebuffer(
4298 const VkAllocationCallbacks
* pAllocator
)
4300 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4301 ANV_FROM_HANDLE(anv_framebuffer
, fb
, _fb
);
4306 vk_free2(&device
->alloc
, pAllocator
, fb
);
4309 static const VkTimeDomainEXT anv_time_domains
[] = {
4310 VK_TIME_DOMAIN_DEVICE_EXT
,
4311 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
,
4312 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
,
4315 VkResult
anv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
4316 VkPhysicalDevice physicalDevice
,
4317 uint32_t *pTimeDomainCount
,
4318 VkTimeDomainEXT
*pTimeDomains
)
4321 VK_OUTARRAY_MAKE(out
, pTimeDomains
, pTimeDomainCount
);
4323 for (d
= 0; d
< ARRAY_SIZE(anv_time_domains
); d
++) {
4324 vk_outarray_append(&out
, i
) {
4325 *i
= anv_time_domains
[d
];
4329 return vk_outarray_status(&out
);
4333 anv_clock_gettime(clockid_t clock_id
)
4335 struct timespec current
;
4338 ret
= clock_gettime(clock_id
, ¤t
);
4339 if (ret
< 0 && clock_id
== CLOCK_MONOTONIC_RAW
)
4340 ret
= clock_gettime(CLOCK_MONOTONIC
, ¤t
);
4344 return (uint64_t) current
.tv_sec
* 1000000000ULL + current
.tv_nsec
;
4347 #define TIMESTAMP 0x2358
4349 VkResult
anv_GetCalibratedTimestampsEXT(
4351 uint32_t timestampCount
,
4352 const VkCalibratedTimestampInfoEXT
*pTimestampInfos
,
4353 uint64_t *pTimestamps
,
4354 uint64_t *pMaxDeviation
)
4356 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4357 uint64_t timestamp_frequency
= device
->info
.timestamp_frequency
;
4360 uint64_t begin
, end
;
4361 uint64_t max_clock_period
= 0;
4363 begin
= anv_clock_gettime(CLOCK_MONOTONIC_RAW
);
4365 for (d
= 0; d
< timestampCount
; d
++) {
4366 switch (pTimestampInfos
[d
].timeDomain
) {
4367 case VK_TIME_DOMAIN_DEVICE_EXT
:
4368 ret
= anv_gem_reg_read(device
, TIMESTAMP
| 1,
4372 return anv_device_set_lost(device
, "Failed to read the TIMESTAMP "
4375 uint64_t device_period
= DIV_ROUND_UP(1000000000, timestamp_frequency
);
4376 max_clock_period
= MAX2(max_clock_period
, device_period
);
4378 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
:
4379 pTimestamps
[d
] = anv_clock_gettime(CLOCK_MONOTONIC
);
4380 max_clock_period
= MAX2(max_clock_period
, 1);
4383 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
:
4384 pTimestamps
[d
] = begin
;
4392 end
= anv_clock_gettime(CLOCK_MONOTONIC_RAW
);
4395 * The maximum deviation is the sum of the interval over which we
4396 * perform the sampling and the maximum period of any sampled
4397 * clock. That's because the maximum skew between any two sampled
4398 * clock edges is when the sampled clock with the largest period is
4399 * sampled at the end of that period but right at the beginning of the
4400 * sampling interval and some other clock is sampled right at the
4401 * begining of its sampling period and right at the end of the
4402 * sampling interval. Let's assume the GPU has the longest clock
4403 * period and that the application is sampling GPU and monotonic:
4406 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
4407 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4411 * GPU -----_____-----_____-----_____-----_____
4414 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
4415 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4417 * Interval <----------------->
4418 * Deviation <-------------------------->
4422 * m = read(monotonic) 2
4425 * We round the sample interval up by one tick to cover sampling error
4426 * in the interval clock
4429 uint64_t sample_interval
= end
- begin
+ 1;
4431 *pMaxDeviation
= sample_interval
+ max_clock_period
;
4436 /* vk_icd.h does not declare this function, so we declare it here to
4437 * suppress Wmissing-prototypes.
4439 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4440 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
);
4442 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4443 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
)
4445 /* For the full details on loader interface versioning, see
4446 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
4447 * What follows is a condensed summary, to help you navigate the large and
4448 * confusing official doc.
4450 * - Loader interface v0 is incompatible with later versions. We don't
4453 * - In loader interface v1:
4454 * - The first ICD entrypoint called by the loader is
4455 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
4457 * - The ICD must statically expose no other Vulkan symbol unless it is
4458 * linked with -Bsymbolic.
4459 * - Each dispatchable Vulkan handle created by the ICD must be
4460 * a pointer to a struct whose first member is VK_LOADER_DATA. The
4461 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
4462 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
4463 * vkDestroySurfaceKHR(). The ICD must be capable of working with
4464 * such loader-managed surfaces.
4466 * - Loader interface v2 differs from v1 in:
4467 * - The first ICD entrypoint called by the loader is
4468 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
4469 * statically expose this entrypoint.
4471 * - Loader interface v3 differs from v2 in:
4472 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
4473 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
4474 * because the loader no longer does so.
4476 * - Loader interface v4 differs from v3 in:
4477 * - The ICD must implement vk_icdGetPhysicalDeviceProcAddr().
4479 *pSupportedVersion
= MIN2(*pSupportedVersion
, 4u);