2 * Copyright © 2015 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
28 #include <sys/sysinfo.h>
32 #include "drm-uapi/drm_fourcc.h"
34 #include "anv_private.h"
35 #include "util/debug.h"
36 #include "util/build_id.h"
37 #include "util/disk_cache.h"
38 #include "util/mesa-sha1.h"
39 #include "util/os_file.h"
40 #include "util/u_atomic.h"
41 #include "util/u_string.h"
42 #include "util/xmlpool.h"
45 #include "common/gen_aux_map.h"
46 #include "common/gen_defines.h"
47 #include "compiler/glsl_types.h"
49 #include "genxml/gen7_pack.h"
51 static const char anv_dri_options_xml
[] =
53 DRI_CONF_SECTION_PERFORMANCE
54 DRI_CONF_VK_X11_OVERRIDE_MIN_IMAGE_COUNT(0)
55 DRI_CONF_VK_X11_STRICT_IMAGE_COUNT("false")
58 DRI_CONF_SECTION_DEBUG
59 DRI_CONF_ALWAYS_FLUSH_CACHE("false")
60 DRI_CONF_VK_WSI_FORCE_BGRA8_UNORM_FIRST("false")
64 /* This is probably far to big but it reflects the max size used for messages
65 * in OpenGLs KHR_debug.
67 #define MAX_DEBUG_MESSAGE_LENGTH 4096
70 compiler_debug_log(void *data
, const char *fmt
, ...)
72 char str
[MAX_DEBUG_MESSAGE_LENGTH
];
73 struct anv_device
*device
= (struct anv_device
*)data
;
74 struct anv_instance
*instance
= device
->physical
->instance
;
76 if (list_is_empty(&instance
->debug_report_callbacks
.callbacks
))
81 (void) vsnprintf(str
, MAX_DEBUG_MESSAGE_LENGTH
, fmt
, args
);
84 vk_debug_report(&instance
->debug_report_callbacks
,
85 VK_DEBUG_REPORT_DEBUG_BIT_EXT
,
86 VK_DEBUG_REPORT_OBJECT_TYPE_UNKNOWN_EXT
,
91 compiler_perf_log(void *data
, const char *fmt
, ...)
96 if (unlikely(INTEL_DEBUG
& DEBUG_PERF
))
97 intel_logd_v(fmt
, args
);
103 anv_compute_heap_size(int fd
, uint64_t gtt_size
)
105 /* Query the total ram from the system */
109 uint64_t total_ram
= (uint64_t)info
.totalram
* (uint64_t)info
.mem_unit
;
111 /* We don't want to burn too much ram with the GPU. If the user has 4GiB
112 * or less, we use at most half. If they have more than 4GiB, we use 3/4.
114 uint64_t available_ram
;
115 if (total_ram
<= 4ull * 1024ull * 1024ull * 1024ull)
116 available_ram
= total_ram
/ 2;
118 available_ram
= total_ram
* 3 / 4;
120 /* We also want to leave some padding for things we allocate in the driver,
121 * so don't go over 3/4 of the GTT either.
123 uint64_t available_gtt
= gtt_size
* 3 / 4;
125 return MIN2(available_ram
, available_gtt
);
129 anv_physical_device_init_heaps(struct anv_physical_device
*device
, int fd
)
131 if (anv_gem_get_context_param(fd
, 0, I915_CONTEXT_PARAM_GTT_SIZE
,
132 &device
->gtt_size
) == -1) {
133 /* If, for whatever reason, we can't actually get the GTT size from the
134 * kernel (too old?) fall back to the aperture size.
136 anv_perf_warn(NULL
, NULL
,
137 "Failed to get I915_CONTEXT_PARAM_GTT_SIZE: %m");
139 if (anv_gem_get_aperture(fd
, &device
->gtt_size
) == -1) {
140 return vk_errorfi(device
->instance
, NULL
,
141 VK_ERROR_INITIALIZATION_FAILED
,
142 "failed to get aperture size: %m");
146 /* We only allow 48-bit addresses with softpin because knowing the actual
147 * address is required for the vertex cache flush workaround.
149 device
->supports_48bit_addresses
= (device
->info
.gen
>= 8) &&
150 device
->has_softpin
&&
151 device
->gtt_size
> (4ULL << 30 /* GiB */);
153 uint64_t heap_size
= anv_compute_heap_size(fd
, device
->gtt_size
);
155 if (heap_size
> (2ull << 30) && !device
->supports_48bit_addresses
) {
156 /* When running with an overridden PCI ID, we may get a GTT size from
157 * the kernel that is greater than 2 GiB but the execbuf check for 48bit
158 * address support can still fail. Just clamp the address space size to
159 * 2 GiB if we don't have 48-bit support.
161 intel_logw("%s:%d: The kernel reported a GTT size larger than 2 GiB but "
162 "not support for 48-bit addresses",
164 heap_size
= 2ull << 30;
167 device
->memory
.heap_count
= 1;
168 device
->memory
.heaps
[0] = (struct anv_memory_heap
) {
170 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
173 uint32_t type_count
= 0;
174 for (uint32_t heap
= 0; heap
< device
->memory
.heap_count
; heap
++) {
175 if (device
->info
.has_llc
) {
176 /* Big core GPUs share LLC with the CPU and thus one memory type can be
177 * both cached and coherent at the same time.
179 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
180 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
181 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
182 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
183 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
187 /* The spec requires that we expose a host-visible, coherent memory
188 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
189 * to give the application a choice between cached, but not coherent and
190 * coherent but uncached (WC though).
192 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
193 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
194 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
195 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
198 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
199 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
200 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
201 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
206 device
->memory
.type_count
= type_count
;
212 anv_physical_device_init_uuids(struct anv_physical_device
*device
)
214 const struct build_id_note
*note
=
215 build_id_find_nhdr_for_addr(anv_physical_device_init_uuids
);
217 return vk_errorfi(device
->instance
, NULL
,
218 VK_ERROR_INITIALIZATION_FAILED
,
219 "Failed to find build-id");
222 unsigned build_id_len
= build_id_length(note
);
223 if (build_id_len
< 20) {
224 return vk_errorfi(device
->instance
, NULL
,
225 VK_ERROR_INITIALIZATION_FAILED
,
226 "build-id too short. It needs to be a SHA");
229 memcpy(device
->driver_build_sha1
, build_id_data(note
), 20);
231 struct mesa_sha1 sha1_ctx
;
233 STATIC_ASSERT(VK_UUID_SIZE
<= sizeof(sha1
));
235 /* The pipeline cache UUID is used for determining when a pipeline cache is
236 * invalid. It needs both a driver build and the PCI ID of the device.
238 _mesa_sha1_init(&sha1_ctx
);
239 _mesa_sha1_update(&sha1_ctx
, build_id_data(note
), build_id_len
);
240 _mesa_sha1_update(&sha1_ctx
, &device
->info
.chipset_id
,
241 sizeof(device
->info
.chipset_id
));
242 _mesa_sha1_update(&sha1_ctx
, &device
->always_use_bindless
,
243 sizeof(device
->always_use_bindless
));
244 _mesa_sha1_update(&sha1_ctx
, &device
->has_a64_buffer_access
,
245 sizeof(device
->has_a64_buffer_access
));
246 _mesa_sha1_update(&sha1_ctx
, &device
->has_bindless_images
,
247 sizeof(device
->has_bindless_images
));
248 _mesa_sha1_update(&sha1_ctx
, &device
->has_bindless_samplers
,
249 sizeof(device
->has_bindless_samplers
));
250 _mesa_sha1_final(&sha1_ctx
, sha1
);
251 memcpy(device
->pipeline_cache_uuid
, sha1
, VK_UUID_SIZE
);
253 /* The driver UUID is used for determining sharability of images and memory
254 * between two Vulkan instances in separate processes. People who want to
255 * share memory need to also check the device UUID (below) so all this
256 * needs to be is the build-id.
258 memcpy(device
->driver_uuid
, build_id_data(note
), VK_UUID_SIZE
);
260 /* The device UUID uniquely identifies the given device within the machine.
261 * Since we never have more than one device, this doesn't need to be a real
262 * UUID. However, on the off-chance that someone tries to use this to
263 * cache pre-tiled images or something of the like, we use the PCI ID and
264 * some bits of ISL info to ensure that this is safe.
266 _mesa_sha1_init(&sha1_ctx
);
267 _mesa_sha1_update(&sha1_ctx
, &device
->info
.chipset_id
,
268 sizeof(device
->info
.chipset_id
));
269 _mesa_sha1_update(&sha1_ctx
, &device
->isl_dev
.has_bit6_swizzling
,
270 sizeof(device
->isl_dev
.has_bit6_swizzling
));
271 _mesa_sha1_final(&sha1_ctx
, sha1
);
272 memcpy(device
->device_uuid
, sha1
, VK_UUID_SIZE
);
278 anv_physical_device_init_disk_cache(struct anv_physical_device
*device
)
280 #ifdef ENABLE_SHADER_CACHE
282 ASSERTED
int len
= snprintf(renderer
, sizeof(renderer
), "anv_%04x",
283 device
->info
.chipset_id
);
284 assert(len
== sizeof(renderer
) - 2);
287 _mesa_sha1_format(timestamp
, device
->driver_build_sha1
);
289 const uint64_t driver_flags
=
290 brw_get_compiler_config_value(device
->compiler
);
291 device
->disk_cache
= disk_cache_create(renderer
, timestamp
, driver_flags
);
293 device
->disk_cache
= NULL
;
298 anv_physical_device_free_disk_cache(struct anv_physical_device
*device
)
300 #ifdef ENABLE_SHADER_CACHE
301 if (device
->disk_cache
)
302 disk_cache_destroy(device
->disk_cache
);
304 assert(device
->disk_cache
== NULL
);
309 get_available_system_memory()
311 char *meminfo
= os_read_file("/proc/meminfo");
315 char *str
= strstr(meminfo
, "MemAvailable:");
321 uint64_t kb_mem_available
;
322 if (sscanf(str
, "MemAvailable: %" PRIx64
, &kb_mem_available
) == 1) {
324 return kb_mem_available
<< 10;
332 anv_physical_device_try_create(struct anv_instance
*instance
,
333 drmDevicePtr drm_device
,
334 struct anv_physical_device
**device_out
)
336 const char *primary_path
= drm_device
->nodes
[DRM_NODE_PRIMARY
];
337 const char *path
= drm_device
->nodes
[DRM_NODE_RENDER
];
342 brw_process_intel_debug_variable();
344 fd
= open(path
, O_RDWR
| O_CLOEXEC
);
346 return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
348 struct gen_device_info devinfo
;
349 if (!gen_get_device_info_from_fd(fd
, &devinfo
)) {
350 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
354 const char *device_name
= gen_get_device_name(devinfo
.chipset_id
);
356 if (devinfo
.is_haswell
) {
357 intel_logw("Haswell Vulkan support is incomplete");
358 } else if (devinfo
.gen
== 7 && !devinfo
.is_baytrail
) {
359 intel_logw("Ivy Bridge Vulkan support is incomplete");
360 } else if (devinfo
.gen
== 7 && devinfo
.is_baytrail
) {
361 intel_logw("Bay Trail Vulkan support is incomplete");
362 } else if (devinfo
.gen
>= 8 && devinfo
.gen
<= 11) {
363 /* Gen8-11 fully supported */
364 } else if (devinfo
.gen
== 12) {
365 intel_logw("Vulkan is not yet fully supported on gen12");
367 result
= vk_errorfi(instance
, NULL
, VK_ERROR_INCOMPATIBLE_DRIVER
,
368 "Vulkan not yet supported on %s", device_name
);
372 struct anv_physical_device
*device
=
373 vk_alloc(&instance
->alloc
, sizeof(*device
), 8,
374 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
375 if (device
== NULL
) {
376 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
380 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
381 device
->instance
= instance
;
383 assert(strlen(path
) < ARRAY_SIZE(device
->path
));
384 snprintf(device
->path
, ARRAY_SIZE(device
->path
), "%s", path
);
386 device
->info
= devinfo
;
387 device
->name
= device_name
;
389 device
->no_hw
= device
->info
.no_hw
;
390 if (getenv("INTEL_NO_HW") != NULL
)
391 device
->no_hw
= true;
393 device
->pci_info
.domain
= drm_device
->businfo
.pci
->domain
;
394 device
->pci_info
.bus
= drm_device
->businfo
.pci
->bus
;
395 device
->pci_info
.device
= drm_device
->businfo
.pci
->dev
;
396 device
->pci_info
.function
= drm_device
->businfo
.pci
->func
;
398 device
->cmd_parser_version
= -1;
399 if (device
->info
.gen
== 7) {
400 device
->cmd_parser_version
=
401 anv_gem_get_param(fd
, I915_PARAM_CMD_PARSER_VERSION
);
402 if (device
->cmd_parser_version
== -1) {
403 result
= vk_errorfi(device
->instance
, NULL
,
404 VK_ERROR_INITIALIZATION_FAILED
,
405 "failed to get command parser version");
410 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_WAIT_TIMEOUT
)) {
411 result
= vk_errorfi(device
->instance
, NULL
,
412 VK_ERROR_INITIALIZATION_FAILED
,
413 "kernel missing gem wait");
417 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_EXECBUF2
)) {
418 result
= vk_errorfi(device
->instance
, NULL
,
419 VK_ERROR_INITIALIZATION_FAILED
,
420 "kernel missing execbuf2");
424 if (!device
->info
.has_llc
&&
425 anv_gem_get_param(fd
, I915_PARAM_MMAP_VERSION
) < 1) {
426 result
= vk_errorfi(device
->instance
, NULL
,
427 VK_ERROR_INITIALIZATION_FAILED
,
428 "kernel missing wc mmap");
432 device
->has_softpin
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_SOFTPIN
);
433 device
->has_exec_async
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_ASYNC
);
434 device
->has_exec_capture
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_CAPTURE
);
435 device
->has_exec_fence
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE
);
436 device
->has_syncobj
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE_ARRAY
);
437 device
->has_syncobj_wait
= device
->has_syncobj
&&
438 anv_gem_supports_syncobj_wait(fd
);
439 device
->has_context_priority
= anv_gem_has_context_priority(fd
);
441 result
= anv_physical_device_init_heaps(device
, fd
);
442 if (result
!= VK_SUCCESS
)
445 device
->use_softpin
= device
->has_softpin
&&
446 device
->supports_48bit_addresses
;
448 device
->has_context_isolation
=
449 anv_gem_get_param(fd
, I915_PARAM_HAS_CONTEXT_ISOLATION
);
451 device
->always_use_bindless
=
452 env_var_as_boolean("ANV_ALWAYS_BINDLESS", false);
454 /* We first got the A64 messages on broadwell and we can only use them if
455 * we can pass addresses directly into the shader which requires softpin.
457 device
->has_a64_buffer_access
= device
->info
.gen
>= 8 &&
460 /* We first get bindless image access on Skylake and we can only really do
461 * it if we don't have any relocations so we need softpin.
463 device
->has_bindless_images
= device
->info
.gen
>= 9 &&
466 /* We've had bindless samplers since Ivy Bridge (forever in Vulkan terms)
467 * because it's just a matter of setting the sampler address in the sample
468 * message header. However, we've not bothered to wire it up for vec4 so
469 * we leave it disabled on gen7.
471 device
->has_bindless_samplers
= device
->info
.gen
>= 8;
473 device
->has_mem_available
= get_available_system_memory() != 0;
475 device
->always_flush_cache
=
476 driQueryOptionb(&instance
->dri_options
, "always_flush_cache");
478 /* Starting with Gen10, the timestamp frequency of the command streamer may
479 * vary from one part to another. We can query the value from the kernel.
481 if (device
->info
.gen
>= 10) {
482 int timestamp_frequency
=
483 anv_gem_get_param(fd
, I915_PARAM_CS_TIMESTAMP_FREQUENCY
);
485 if (timestamp_frequency
< 0)
486 intel_logw("Kernel 4.16-rc1+ required to properly query CS timestamp frequency");
488 device
->info
.timestamp_frequency
= timestamp_frequency
;
491 /* GENs prior to 8 do not support EU/Subslice info */
492 if (device
->info
.gen
>= 8) {
493 device
->subslice_total
= anv_gem_get_param(fd
, I915_PARAM_SUBSLICE_TOTAL
);
494 device
->eu_total
= anv_gem_get_param(fd
, I915_PARAM_EU_TOTAL
);
496 /* Without this information, we cannot get the right Braswell
497 * brandstrings, and we have to use conservative numbers for GPGPU on
498 * many platforms, but otherwise, things will just work.
500 if (device
->subslice_total
< 1 || device
->eu_total
< 1) {
501 intel_logw("Kernel 4.1 required to properly query GPU properties");
503 } else if (device
->info
.gen
== 7) {
504 device
->subslice_total
= 1 << (device
->info
.gt
- 1);
507 if (device
->info
.is_cherryview
&&
508 device
->subslice_total
> 0 && device
->eu_total
> 0) {
509 /* Logical CS threads = EUs per subslice * num threads per EU */
510 uint32_t max_cs_threads
=
511 device
->eu_total
/ device
->subslice_total
* device
->info
.num_thread_per_eu
;
513 /* Fuse configurations may give more threads than expected, never less. */
514 if (max_cs_threads
> device
->info
.max_cs_threads
)
515 device
->info
.max_cs_threads
= max_cs_threads
;
518 device
->compiler
= brw_compiler_create(NULL
, &device
->info
);
519 if (device
->compiler
== NULL
) {
520 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
523 device
->compiler
->shader_debug_log
= compiler_debug_log
;
524 device
->compiler
->shader_perf_log
= compiler_perf_log
;
525 device
->compiler
->supports_pull_constants
= false;
526 device
->compiler
->constant_buffer_0_is_relative
=
527 device
->info
.gen
< 8 || !device
->has_context_isolation
;
528 device
->compiler
->supports_shader_constants
= true;
529 device
->compiler
->compact_params
= false;
531 /* Broadwell PRM says:
533 * "Before Gen8, there was a historical configuration control field to
534 * swizzle address bit[6] for in X/Y tiling modes. This was set in three
535 * different places: TILECTL[1:0], ARB_MODE[5:4], and
536 * DISP_ARB_CTL[14:13].
538 * For Gen8 and subsequent generations, the swizzle fields are all
539 * reserved, and the CPU's memory controller performs all address
540 * swizzling modifications."
543 device
->info
.gen
< 8 && anv_gem_get_bit6_swizzle(fd
, I915_TILING_X
);
545 isl_device_init(&device
->isl_dev
, &device
->info
, swizzled
);
547 result
= anv_physical_device_init_uuids(device
);
548 if (result
!= VK_SUCCESS
)
551 anv_physical_device_init_disk_cache(device
);
553 if (instance
->enabled_extensions
.KHR_display
) {
554 master_fd
= open(primary_path
, O_RDWR
| O_CLOEXEC
);
555 if (master_fd
>= 0) {
556 /* prod the device with a GETPARAM call which will fail if
557 * we don't have permission to even render on this device
559 if (anv_gem_get_param(master_fd
, I915_PARAM_CHIPSET_ID
) == 0) {
565 device
->master_fd
= master_fd
;
567 result
= anv_init_wsi(device
);
568 if (result
!= VK_SUCCESS
)
569 goto fail_disk_cache
;
571 device
->perf
= anv_get_perf(&device
->info
, fd
);
573 anv_physical_device_get_supported_extensions(device
,
574 &device
->supported_extensions
);
577 device
->local_fd
= fd
;
579 *device_out
= device
;
584 anv_physical_device_free_disk_cache(device
);
586 ralloc_free(device
->compiler
);
588 vk_free(&instance
->alloc
, device
);
597 anv_physical_device_destroy(struct anv_physical_device
*device
)
599 anv_finish_wsi(device
);
600 anv_physical_device_free_disk_cache(device
);
601 ralloc_free(device
->compiler
);
602 ralloc_free(device
->perf
);
603 close(device
->local_fd
);
604 if (device
->master_fd
>= 0)
605 close(device
->master_fd
);
606 vk_free(&device
->instance
->alloc
, device
);
610 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
611 VkSystemAllocationScope allocationScope
)
617 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
618 size_t align
, VkSystemAllocationScope allocationScope
)
620 return realloc(pOriginal
, size
);
624 default_free_func(void *pUserData
, void *pMemory
)
629 static const VkAllocationCallbacks default_alloc
= {
631 .pfnAllocation
= default_alloc_func
,
632 .pfnReallocation
= default_realloc_func
,
633 .pfnFree
= default_free_func
,
636 VkResult
anv_EnumerateInstanceExtensionProperties(
637 const char* pLayerName
,
638 uint32_t* pPropertyCount
,
639 VkExtensionProperties
* pProperties
)
641 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
643 for (int i
= 0; i
< ANV_INSTANCE_EXTENSION_COUNT
; i
++) {
644 if (anv_instance_extensions_supported
.extensions
[i
]) {
645 vk_outarray_append(&out
, prop
) {
646 *prop
= anv_instance_extensions
[i
];
651 return vk_outarray_status(&out
);
654 VkResult
anv_CreateInstance(
655 const VkInstanceCreateInfo
* pCreateInfo
,
656 const VkAllocationCallbacks
* pAllocator
,
657 VkInstance
* pInstance
)
659 struct anv_instance
*instance
;
662 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
664 struct anv_instance_extension_table enabled_extensions
= {};
665 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
667 for (idx
= 0; idx
< ANV_INSTANCE_EXTENSION_COUNT
; idx
++) {
668 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
669 anv_instance_extensions
[idx
].extensionName
) == 0)
673 if (idx
>= ANV_INSTANCE_EXTENSION_COUNT
)
674 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
676 if (!anv_instance_extensions_supported
.extensions
[idx
])
677 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
679 enabled_extensions
.extensions
[idx
] = true;
682 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
683 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
685 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
687 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
690 instance
->alloc
= *pAllocator
;
692 instance
->alloc
= default_alloc
;
694 instance
->app_info
= (struct anv_app_info
) { .api_version
= 0 };
695 if (pCreateInfo
->pApplicationInfo
) {
696 const VkApplicationInfo
*app
= pCreateInfo
->pApplicationInfo
;
698 instance
->app_info
.app_name
=
699 vk_strdup(&instance
->alloc
, app
->pApplicationName
,
700 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
701 instance
->app_info
.app_version
= app
->applicationVersion
;
703 instance
->app_info
.engine_name
=
704 vk_strdup(&instance
->alloc
, app
->pEngineName
,
705 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
706 instance
->app_info
.engine_version
= app
->engineVersion
;
708 instance
->app_info
.api_version
= app
->apiVersion
;
711 if (instance
->app_info
.api_version
== 0)
712 instance
->app_info
.api_version
= VK_API_VERSION_1_0
;
714 instance
->enabled_extensions
= enabled_extensions
;
716 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->dispatch
.entrypoints
); i
++) {
717 /* Vulkan requires that entrypoints for extensions which have not been
718 * enabled must not be advertised.
720 if (!anv_instance_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
721 &instance
->enabled_extensions
)) {
722 instance
->dispatch
.entrypoints
[i
] = NULL
;
724 instance
->dispatch
.entrypoints
[i
] =
725 anv_instance_dispatch_table
.entrypoints
[i
];
729 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->physical_device_dispatch
.entrypoints
); i
++) {
730 /* Vulkan requires that entrypoints for extensions which have not been
731 * enabled must not be advertised.
733 if (!anv_physical_device_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
734 &instance
->enabled_extensions
)) {
735 instance
->physical_device_dispatch
.entrypoints
[i
] = NULL
;
737 instance
->physical_device_dispatch
.entrypoints
[i
] =
738 anv_physical_device_dispatch_table
.entrypoints
[i
];
742 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->device_dispatch
.entrypoints
); i
++) {
743 /* Vulkan requires that entrypoints for extensions which have not been
744 * enabled must not be advertised.
746 if (!anv_device_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
747 &instance
->enabled_extensions
, NULL
)) {
748 instance
->device_dispatch
.entrypoints
[i
] = NULL
;
750 instance
->device_dispatch
.entrypoints
[i
] =
751 anv_device_dispatch_table
.entrypoints
[i
];
755 instance
->physical_devices_enumerated
= false;
756 list_inithead(&instance
->physical_devices
);
758 result
= vk_debug_report_instance_init(&instance
->debug_report_callbacks
);
759 if (result
!= VK_SUCCESS
) {
760 vk_free2(&default_alloc
, pAllocator
, instance
);
761 return vk_error(result
);
764 instance
->pipeline_cache_enabled
=
765 env_var_as_boolean("ANV_ENABLE_PIPELINE_CACHE", true);
767 glsl_type_singleton_init_or_ref();
769 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
771 driParseOptionInfo(&instance
->available_dri_options
, anv_dri_options_xml
);
772 driParseConfigFiles(&instance
->dri_options
, &instance
->available_dri_options
,
774 instance
->app_info
.engine_name
,
775 instance
->app_info
.engine_version
);
777 *pInstance
= anv_instance_to_handle(instance
);
782 void anv_DestroyInstance(
783 VkInstance _instance
,
784 const VkAllocationCallbacks
* pAllocator
)
786 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
791 list_for_each_entry_safe(struct anv_physical_device
, pdevice
,
792 &instance
->physical_devices
, link
)
793 anv_physical_device_destroy(pdevice
);
795 vk_free(&instance
->alloc
, (char *)instance
->app_info
.app_name
);
796 vk_free(&instance
->alloc
, (char *)instance
->app_info
.engine_name
);
798 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
800 vk_debug_report_instance_destroy(&instance
->debug_report_callbacks
);
802 glsl_type_singleton_decref();
804 driDestroyOptionCache(&instance
->dri_options
);
805 driDestroyOptionInfo(&instance
->available_dri_options
);
807 vk_free(&instance
->alloc
, instance
);
811 anv_enumerate_physical_devices(struct anv_instance
*instance
)
813 if (instance
->physical_devices_enumerated
)
816 instance
->physical_devices_enumerated
= true;
818 /* TODO: Check for more devices ? */
819 drmDevicePtr devices
[8];
822 max_devices
= drmGetDevices2(0, devices
, ARRAY_SIZE(devices
));
826 VkResult result
= VK_SUCCESS
;
827 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
828 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
829 devices
[i
]->bustype
== DRM_BUS_PCI
&&
830 devices
[i
]->deviceinfo
.pci
->vendor_id
== 0x8086) {
832 struct anv_physical_device
*pdevice
;
833 result
= anv_physical_device_try_create(instance
, devices
[i
],
835 /* Incompatible DRM device, skip. */
836 if (result
== VK_ERROR_INCOMPATIBLE_DRIVER
) {
841 /* Error creating the physical device, report the error. */
842 if (result
!= VK_SUCCESS
)
845 list_addtail(&pdevice
->link
, &instance
->physical_devices
);
848 drmFreeDevices(devices
, max_devices
);
850 /* If we successfully enumerated any devices, call it success */
854 VkResult
anv_EnumeratePhysicalDevices(
855 VkInstance _instance
,
856 uint32_t* pPhysicalDeviceCount
,
857 VkPhysicalDevice
* pPhysicalDevices
)
859 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
860 VK_OUTARRAY_MAKE(out
, pPhysicalDevices
, pPhysicalDeviceCount
);
862 VkResult result
= anv_enumerate_physical_devices(instance
);
863 if (result
!= VK_SUCCESS
)
866 list_for_each_entry(struct anv_physical_device
, pdevice
,
867 &instance
->physical_devices
, link
) {
868 vk_outarray_append(&out
, i
) {
869 *i
= anv_physical_device_to_handle(pdevice
);
873 return vk_outarray_status(&out
);
876 VkResult
anv_EnumeratePhysicalDeviceGroups(
877 VkInstance _instance
,
878 uint32_t* pPhysicalDeviceGroupCount
,
879 VkPhysicalDeviceGroupProperties
* pPhysicalDeviceGroupProperties
)
881 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
882 VK_OUTARRAY_MAKE(out
, pPhysicalDeviceGroupProperties
,
883 pPhysicalDeviceGroupCount
);
885 VkResult result
= anv_enumerate_physical_devices(instance
);
886 if (result
!= VK_SUCCESS
)
889 list_for_each_entry(struct anv_physical_device
, pdevice
,
890 &instance
->physical_devices
, link
) {
891 vk_outarray_append(&out
, p
) {
892 p
->physicalDeviceCount
= 1;
893 memset(p
->physicalDevices
, 0, sizeof(p
->physicalDevices
));
894 p
->physicalDevices
[0] = anv_physical_device_to_handle(pdevice
);
895 p
->subsetAllocation
= false;
897 vk_foreach_struct(ext
, p
->pNext
)
898 anv_debug_ignored_stype(ext
->sType
);
902 return vk_outarray_status(&out
);
905 void anv_GetPhysicalDeviceFeatures(
906 VkPhysicalDevice physicalDevice
,
907 VkPhysicalDeviceFeatures
* pFeatures
)
909 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
911 *pFeatures
= (VkPhysicalDeviceFeatures
) {
912 .robustBufferAccess
= true,
913 .fullDrawIndexUint32
= true,
914 .imageCubeArray
= true,
915 .independentBlend
= true,
916 .geometryShader
= true,
917 .tessellationShader
= true,
918 .sampleRateShading
= true,
919 .dualSrcBlend
= true,
921 .multiDrawIndirect
= true,
922 .drawIndirectFirstInstance
= true,
924 .depthBiasClamp
= true,
925 .fillModeNonSolid
= true,
926 .depthBounds
= pdevice
->info
.gen
>= 12,
930 .multiViewport
= true,
931 .samplerAnisotropy
= true,
932 .textureCompressionETC2
= pdevice
->info
.gen
>= 8 ||
933 pdevice
->info
.is_baytrail
,
934 .textureCompressionASTC_LDR
= pdevice
->info
.gen
>= 9, /* FINISHME CHV */
935 .textureCompressionBC
= true,
936 .occlusionQueryPrecise
= true,
937 .pipelineStatisticsQuery
= true,
938 .fragmentStoresAndAtomics
= true,
939 .shaderTessellationAndGeometryPointSize
= true,
940 .shaderImageGatherExtended
= true,
941 .shaderStorageImageExtendedFormats
= true,
942 .shaderStorageImageMultisample
= false,
943 .shaderStorageImageReadWithoutFormat
= false,
944 .shaderStorageImageWriteWithoutFormat
= true,
945 .shaderUniformBufferArrayDynamicIndexing
= true,
946 .shaderSampledImageArrayDynamicIndexing
= true,
947 .shaderStorageBufferArrayDynamicIndexing
= true,
948 .shaderStorageImageArrayDynamicIndexing
= true,
949 .shaderClipDistance
= true,
950 .shaderCullDistance
= true,
951 .shaderFloat64
= pdevice
->info
.gen
>= 8 &&
952 pdevice
->info
.has_64bit_float
,
953 .shaderInt64
= pdevice
->info
.gen
>= 8 &&
954 pdevice
->info
.has_64bit_int
,
955 .shaderInt16
= pdevice
->info
.gen
>= 8,
956 .shaderResourceMinLod
= pdevice
->info
.gen
>= 9,
957 .variableMultisampleRate
= true,
958 .inheritedQueries
= true,
961 /* We can't do image stores in vec4 shaders */
962 pFeatures
->vertexPipelineStoresAndAtomics
=
963 pdevice
->compiler
->scalar_stage
[MESA_SHADER_VERTEX
] &&
964 pdevice
->compiler
->scalar_stage
[MESA_SHADER_GEOMETRY
];
966 struct anv_app_info
*app_info
= &pdevice
->instance
->app_info
;
968 /* The new DOOM and Wolfenstein games require depthBounds without
969 * checking for it. They seem to run fine without it so just claim it's
970 * there and accept the consequences.
972 if (app_info
->engine_name
&& strcmp(app_info
->engine_name
, "idTech") == 0)
973 pFeatures
->depthBounds
= true;
977 anv_get_physical_device_features_1_1(struct anv_physical_device
*pdevice
,
978 VkPhysicalDeviceVulkan11Features
*f
)
980 assert(f
->sType
== VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES
);
982 f
->storageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
983 f
->uniformAndStorageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
984 f
->storagePushConstant16
= pdevice
->info
.gen
>= 8;
985 f
->storageInputOutput16
= false;
987 f
->multiviewGeometryShader
= true;
988 f
->multiviewTessellationShader
= true;
989 f
->variablePointersStorageBuffer
= true;
990 f
->variablePointers
= true;
991 f
->protectedMemory
= false;
992 f
->samplerYcbcrConversion
= true;
993 f
->shaderDrawParameters
= true;
997 anv_get_physical_device_features_1_2(struct anv_physical_device
*pdevice
,
998 VkPhysicalDeviceVulkan12Features
*f
)
1000 assert(f
->sType
== VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES
);
1002 f
->samplerMirrorClampToEdge
= true;
1003 f
->drawIndirectCount
= true;
1004 f
->storageBuffer8BitAccess
= pdevice
->info
.gen
>= 8;
1005 f
->uniformAndStorageBuffer8BitAccess
= pdevice
->info
.gen
>= 8;
1006 f
->storagePushConstant8
= pdevice
->info
.gen
>= 8;
1007 f
->shaderBufferInt64Atomics
= pdevice
->info
.gen
>= 9 &&
1008 pdevice
->use_softpin
;
1009 f
->shaderSharedInt64Atomics
= false;
1010 f
->shaderFloat16
= pdevice
->info
.gen
>= 8;
1011 f
->shaderInt8
= pdevice
->info
.gen
>= 8;
1013 bool descIndexing
= pdevice
->has_a64_buffer_access
&&
1014 pdevice
->has_bindless_images
;
1015 f
->descriptorIndexing
= descIndexing
;
1016 f
->shaderInputAttachmentArrayDynamicIndexing
= false;
1017 f
->shaderUniformTexelBufferArrayDynamicIndexing
= descIndexing
;
1018 f
->shaderStorageTexelBufferArrayDynamicIndexing
= descIndexing
;
1019 f
->shaderUniformBufferArrayNonUniformIndexing
= false;
1020 f
->shaderSampledImageArrayNonUniformIndexing
= descIndexing
;
1021 f
->shaderStorageBufferArrayNonUniformIndexing
= descIndexing
;
1022 f
->shaderStorageImageArrayNonUniformIndexing
= descIndexing
;
1023 f
->shaderInputAttachmentArrayNonUniformIndexing
= false;
1024 f
->shaderUniformTexelBufferArrayNonUniformIndexing
= descIndexing
;
1025 f
->shaderStorageTexelBufferArrayNonUniformIndexing
= descIndexing
;
1026 f
->descriptorBindingUniformBufferUpdateAfterBind
= false;
1027 f
->descriptorBindingSampledImageUpdateAfterBind
= descIndexing
;
1028 f
->descriptorBindingStorageImageUpdateAfterBind
= descIndexing
;
1029 f
->descriptorBindingStorageBufferUpdateAfterBind
= descIndexing
;
1030 f
->descriptorBindingUniformTexelBufferUpdateAfterBind
= descIndexing
;
1031 f
->descriptorBindingStorageTexelBufferUpdateAfterBind
= descIndexing
;
1032 f
->descriptorBindingUpdateUnusedWhilePending
= descIndexing
;
1033 f
->descriptorBindingPartiallyBound
= descIndexing
;
1034 f
->descriptorBindingVariableDescriptorCount
= false;
1035 f
->runtimeDescriptorArray
= descIndexing
;
1037 f
->samplerFilterMinmax
= pdevice
->info
.gen
>= 9;
1038 f
->scalarBlockLayout
= true;
1039 f
->imagelessFramebuffer
= true;
1040 f
->uniformBufferStandardLayout
= true;
1041 f
->shaderSubgroupExtendedTypes
= true;
1042 f
->separateDepthStencilLayouts
= true;
1043 f
->hostQueryReset
= true;
1044 f
->timelineSemaphore
= true;
1045 f
->bufferDeviceAddress
= pdevice
->has_a64_buffer_access
;
1046 f
->bufferDeviceAddressCaptureReplay
= pdevice
->has_a64_buffer_access
;
1047 f
->bufferDeviceAddressMultiDevice
= false;
1048 f
->vulkanMemoryModel
= true;
1049 f
->vulkanMemoryModelDeviceScope
= true;
1050 f
->vulkanMemoryModelAvailabilityVisibilityChains
= true;
1051 f
->shaderOutputViewportIndex
= true;
1052 f
->shaderOutputLayer
= true;
1053 f
->subgroupBroadcastDynamicId
= true;
1056 void anv_GetPhysicalDeviceFeatures2(
1057 VkPhysicalDevice physicalDevice
,
1058 VkPhysicalDeviceFeatures2
* pFeatures
)
1060 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1061 anv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
1063 VkPhysicalDeviceVulkan11Features core_1_1
= {
1064 .sType
= VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES
,
1066 anv_get_physical_device_features_1_1(pdevice
, &core_1_1
);
1068 VkPhysicalDeviceVulkan12Features core_1_2
= {
1069 .sType
= VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES
,
1071 anv_get_physical_device_features_1_2(pdevice
, &core_1_2
);
1073 #define CORE_FEATURE(major, minor, feature) \
1074 features->feature = core_##major##_##minor.feature
1077 vk_foreach_struct(ext
, pFeatures
->pNext
) {
1078 switch (ext
->sType
) {
1079 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES_KHR
: {
1080 VkPhysicalDevice8BitStorageFeaturesKHR
*features
=
1081 (VkPhysicalDevice8BitStorageFeaturesKHR
*)ext
;
1082 CORE_FEATURE(1, 2, storageBuffer8BitAccess
);
1083 CORE_FEATURE(1, 2, uniformAndStorageBuffer8BitAccess
);
1084 CORE_FEATURE(1, 2, storagePushConstant8
);
1088 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES
: {
1089 VkPhysicalDevice16BitStorageFeatures
*features
=
1090 (VkPhysicalDevice16BitStorageFeatures
*)ext
;
1091 CORE_FEATURE(1, 1, storageBuffer16BitAccess
);
1092 CORE_FEATURE(1, 1, uniformAndStorageBuffer16BitAccess
);
1093 CORE_FEATURE(1, 1, storagePushConstant16
);
1094 CORE_FEATURE(1, 1, storageInputOutput16
);
1098 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_EXT
: {
1099 VkPhysicalDeviceBufferDeviceAddressFeaturesEXT
*features
= (void *)ext
;
1100 features
->bufferDeviceAddress
= pdevice
->has_a64_buffer_access
;
1101 features
->bufferDeviceAddressCaptureReplay
= false;
1102 features
->bufferDeviceAddressMultiDevice
= false;
1106 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_KHR
: {
1107 VkPhysicalDeviceBufferDeviceAddressFeaturesKHR
*features
= (void *)ext
;
1108 CORE_FEATURE(1, 2, bufferDeviceAddress
);
1109 CORE_FEATURE(1, 2, bufferDeviceAddressCaptureReplay
);
1110 CORE_FEATURE(1, 2, bufferDeviceAddressMultiDevice
);
1114 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COMPUTE_SHADER_DERIVATIVES_FEATURES_NV
: {
1115 VkPhysicalDeviceComputeShaderDerivativesFeaturesNV
*features
=
1116 (VkPhysicalDeviceComputeShaderDerivativesFeaturesNV
*)ext
;
1117 features
->computeDerivativeGroupQuads
= true;
1118 features
->computeDerivativeGroupLinear
= true;
1122 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT
: {
1123 VkPhysicalDeviceConditionalRenderingFeaturesEXT
*features
=
1124 (VkPhysicalDeviceConditionalRenderingFeaturesEXT
*)ext
;
1125 features
->conditionalRendering
= pdevice
->info
.gen
>= 8 ||
1126 pdevice
->info
.is_haswell
;
1127 features
->inheritedConditionalRendering
= pdevice
->info
.gen
>= 8 ||
1128 pdevice
->info
.is_haswell
;
1132 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_CLIP_ENABLE_FEATURES_EXT
: {
1133 VkPhysicalDeviceDepthClipEnableFeaturesEXT
*features
=
1134 (VkPhysicalDeviceDepthClipEnableFeaturesEXT
*)ext
;
1135 features
->depthClipEnable
= true;
1139 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR
: {
1140 VkPhysicalDeviceFloat16Int8FeaturesKHR
*features
= (void *)ext
;
1141 CORE_FEATURE(1, 2, shaderFloat16
);
1142 CORE_FEATURE(1, 2, shaderInt8
);
1146 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FRAGMENT_SHADER_INTERLOCK_FEATURES_EXT
: {
1147 VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT
*features
=
1148 (VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT
*)ext
;
1149 features
->fragmentShaderSampleInterlock
= pdevice
->info
.gen
>= 9;
1150 features
->fragmentShaderPixelInterlock
= pdevice
->info
.gen
>= 9;
1151 features
->fragmentShaderShadingRateInterlock
= false;
1155 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_QUERY_RESET_FEATURES_EXT
: {
1156 VkPhysicalDeviceHostQueryResetFeaturesEXT
*features
=
1157 (VkPhysicalDeviceHostQueryResetFeaturesEXT
*)ext
;
1158 CORE_FEATURE(1, 2, hostQueryReset
);
1162 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT
: {
1163 VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*features
=
1164 (VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*)ext
;
1165 CORE_FEATURE(1, 2, shaderInputAttachmentArrayDynamicIndexing
);
1166 CORE_FEATURE(1, 2, shaderUniformTexelBufferArrayDynamicIndexing
);
1167 CORE_FEATURE(1, 2, shaderStorageTexelBufferArrayDynamicIndexing
);
1168 CORE_FEATURE(1, 2, shaderUniformBufferArrayNonUniformIndexing
);
1169 CORE_FEATURE(1, 2, shaderSampledImageArrayNonUniformIndexing
);
1170 CORE_FEATURE(1, 2, shaderStorageBufferArrayNonUniformIndexing
);
1171 CORE_FEATURE(1, 2, shaderStorageImageArrayNonUniformIndexing
);
1172 CORE_FEATURE(1, 2, shaderInputAttachmentArrayNonUniformIndexing
);
1173 CORE_FEATURE(1, 2, shaderUniformTexelBufferArrayNonUniformIndexing
);
1174 CORE_FEATURE(1, 2, shaderStorageTexelBufferArrayNonUniformIndexing
);
1175 CORE_FEATURE(1, 2, descriptorBindingUniformBufferUpdateAfterBind
);
1176 CORE_FEATURE(1, 2, descriptorBindingSampledImageUpdateAfterBind
);
1177 CORE_FEATURE(1, 2, descriptorBindingStorageImageUpdateAfterBind
);
1178 CORE_FEATURE(1, 2, descriptorBindingStorageBufferUpdateAfterBind
);
1179 CORE_FEATURE(1, 2, descriptorBindingUniformTexelBufferUpdateAfterBind
);
1180 CORE_FEATURE(1, 2, descriptorBindingStorageTexelBufferUpdateAfterBind
);
1181 CORE_FEATURE(1, 2, descriptorBindingUpdateUnusedWhilePending
);
1182 CORE_FEATURE(1, 2, descriptorBindingPartiallyBound
);
1183 CORE_FEATURE(1, 2, descriptorBindingVariableDescriptorCount
);
1184 CORE_FEATURE(1, 2, runtimeDescriptorArray
);
1188 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INDEX_TYPE_UINT8_FEATURES_EXT
: {
1189 VkPhysicalDeviceIndexTypeUint8FeaturesEXT
*features
=
1190 (VkPhysicalDeviceIndexTypeUint8FeaturesEXT
*)ext
;
1191 features
->indexTypeUint8
= true;
1195 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_FEATURES_EXT
: {
1196 VkPhysicalDeviceInlineUniformBlockFeaturesEXT
*features
=
1197 (VkPhysicalDeviceInlineUniformBlockFeaturesEXT
*)ext
;
1198 features
->inlineUniformBlock
= true;
1199 features
->descriptorBindingInlineUniformBlockUpdateAfterBind
= true;
1203 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_FEATURES_EXT
: {
1204 VkPhysicalDeviceLineRasterizationFeaturesEXT
*features
=
1205 (VkPhysicalDeviceLineRasterizationFeaturesEXT
*)ext
;
1206 features
->rectangularLines
= true;
1207 features
->bresenhamLines
= true;
1208 /* Support for Smooth lines with MSAA was removed on gen11. From the
1209 * BSpec section "Multisample ModesState" table for "AA Line Support
1212 * GEN10:BUG:######## NUM_MULTISAMPLES == 1
1214 * Fortunately, this isn't a case most people care about.
1216 features
->smoothLines
= pdevice
->info
.gen
< 10;
1217 features
->stippledRectangularLines
= false;
1218 features
->stippledBresenhamLines
= true;
1219 features
->stippledSmoothLines
= false;
1223 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES
: {
1224 VkPhysicalDeviceMultiviewFeatures
*features
=
1225 (VkPhysicalDeviceMultiviewFeatures
*)ext
;
1226 CORE_FEATURE(1, 1, multiview
);
1227 CORE_FEATURE(1, 1, multiviewGeometryShader
);
1228 CORE_FEATURE(1, 1, multiviewTessellationShader
);
1232 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGELESS_FRAMEBUFFER_FEATURES_KHR
: {
1233 VkPhysicalDeviceImagelessFramebufferFeaturesKHR
*features
=
1234 (VkPhysicalDeviceImagelessFramebufferFeaturesKHR
*)ext
;
1235 CORE_FEATURE(1, 2, imagelessFramebuffer
);
1239 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR
: {
1240 VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR
*features
=
1241 (VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR
*)ext
;
1242 features
->pipelineExecutableInfo
= true;
1246 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES
: {
1247 VkPhysicalDeviceProtectedMemoryFeatures
*features
= (void *)ext
;
1248 CORE_FEATURE(1, 1, protectedMemory
);
1252 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES
: {
1253 VkPhysicalDeviceSamplerYcbcrConversionFeatures
*features
=
1254 (VkPhysicalDeviceSamplerYcbcrConversionFeatures
*) ext
;
1255 CORE_FEATURE(1, 1, samplerYcbcrConversion
);
1259 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES_EXT
: {
1260 VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*features
=
1261 (VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*)ext
;
1262 CORE_FEATURE(1, 2, scalarBlockLayout
);
1266 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SEPARATE_DEPTH_STENCIL_LAYOUTS_FEATURES_KHR
: {
1267 VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR
*features
=
1268 (VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR
*)ext
;
1269 CORE_FEATURE(1, 2, separateDepthStencilLayouts
);
1273 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_INT64_FEATURES_KHR
: {
1274 VkPhysicalDeviceShaderAtomicInt64FeaturesKHR
*features
= (void *)ext
;
1275 CORE_FEATURE(1, 2, shaderBufferInt64Atomics
);
1276 CORE_FEATURE(1, 2, shaderSharedInt64Atomics
);
1280 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DEMOTE_TO_HELPER_INVOCATION_FEATURES_EXT
: {
1281 VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT
*features
= (void *)ext
;
1282 features
->shaderDemoteToHelperInvocation
= true;
1286 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CLOCK_FEATURES_KHR
: {
1287 VkPhysicalDeviceShaderClockFeaturesKHR
*features
=
1288 (VkPhysicalDeviceShaderClockFeaturesKHR
*)ext
;
1289 features
->shaderSubgroupClock
= true;
1290 features
->shaderDeviceClock
= false;
1294 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETERS_FEATURES
: {
1295 VkPhysicalDeviceShaderDrawParametersFeatures
*features
= (void *)ext
;
1296 CORE_FEATURE(1, 1, shaderDrawParameters
);
1300 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_SUBGROUP_EXTENDED_TYPES_FEATURES_KHR
: {
1301 VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR
*features
=
1302 (VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR
*)ext
;
1303 CORE_FEATURE(1, 2, shaderSubgroupExtendedTypes
);
1307 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_FEATURES_EXT
: {
1308 VkPhysicalDeviceSubgroupSizeControlFeaturesEXT
*features
=
1309 (VkPhysicalDeviceSubgroupSizeControlFeaturesEXT
*)ext
;
1310 features
->subgroupSizeControl
= true;
1311 features
->computeFullSubgroups
= true;
1315 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_FEATURES_EXT
: {
1316 VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT
*features
=
1317 (VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT
*)ext
;
1318 features
->texelBufferAlignment
= true;
1322 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_FEATURES_KHR
: {
1323 VkPhysicalDeviceTimelineSemaphoreFeaturesKHR
*features
=
1324 (VkPhysicalDeviceTimelineSemaphoreFeaturesKHR
*) ext
;
1325 CORE_FEATURE(1, 2, timelineSemaphore
);
1329 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES
: {
1330 VkPhysicalDeviceVariablePointersFeatures
*features
= (void *)ext
;
1331 CORE_FEATURE(1, 1, variablePointersStorageBuffer
);
1332 CORE_FEATURE(1, 1, variablePointers
);
1336 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT
: {
1337 VkPhysicalDeviceTransformFeedbackFeaturesEXT
*features
=
1338 (VkPhysicalDeviceTransformFeedbackFeaturesEXT
*)ext
;
1339 features
->transformFeedback
= true;
1340 features
->geometryStreams
= true;
1344 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_UNIFORM_BUFFER_STANDARD_LAYOUT_FEATURES_KHR
: {
1345 VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR
*features
=
1346 (VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR
*)ext
;
1347 CORE_FEATURE(1, 2, uniformBufferStandardLayout
);
1351 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT
: {
1352 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*features
=
1353 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*)ext
;
1354 features
->vertexAttributeInstanceRateDivisor
= true;
1355 features
->vertexAttributeInstanceRateZeroDivisor
= true;
1359 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES
:
1360 anv_get_physical_device_features_1_1(pdevice
, (void *)ext
);
1363 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES
:
1364 anv_get_physical_device_features_1_2(pdevice
, (void *)ext
);
1367 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_MEMORY_MODEL_FEATURES_KHR
: {
1368 VkPhysicalDeviceVulkanMemoryModelFeaturesKHR
*features
= (void *)ext
;
1369 CORE_FEATURE(1, 2, vulkanMemoryModel
);
1370 CORE_FEATURE(1, 2, vulkanMemoryModelDeviceScope
);
1371 CORE_FEATURE(1, 2, vulkanMemoryModelAvailabilityVisibilityChains
);
1375 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_YCBCR_IMAGE_ARRAYS_FEATURES_EXT
: {
1376 VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*features
=
1377 (VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*)ext
;
1378 features
->ycbcrImageArrays
= true;
1383 anv_debug_ignored_stype(ext
->sType
);
1391 #define MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS 64
1393 #define MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS 64
1394 #define MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS 256
1396 void anv_GetPhysicalDeviceProperties(
1397 VkPhysicalDevice physicalDevice
,
1398 VkPhysicalDeviceProperties
* pProperties
)
1400 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1401 const struct gen_device_info
*devinfo
= &pdevice
->info
;
1403 /* See assertions made when programming the buffer surface state. */
1404 const uint32_t max_raw_buffer_sz
= devinfo
->gen
>= 7 ?
1405 (1ul << 30) : (1ul << 27);
1407 const uint32_t max_ssbos
= pdevice
->has_a64_buffer_access
? UINT16_MAX
: 64;
1408 const uint32_t max_textures
=
1409 pdevice
->has_bindless_images
? UINT16_MAX
: 128;
1410 const uint32_t max_samplers
=
1411 pdevice
->has_bindless_samplers
? UINT16_MAX
:
1412 (devinfo
->gen
>= 8 || devinfo
->is_haswell
) ? 128 : 16;
1413 const uint32_t max_images
=
1414 pdevice
->has_bindless_images
? UINT16_MAX
: MAX_IMAGES
;
1416 /* If we can use bindless for everything, claim a high per-stage limit,
1417 * otherwise use the binding table size, minus the slots reserved for
1418 * render targets and one slot for the descriptor buffer. */
1419 const uint32_t max_per_stage
=
1420 pdevice
->has_bindless_images
&& pdevice
->has_a64_buffer_access
1421 ? UINT32_MAX
: MAX_BINDING_TABLE_SIZE
- MAX_RTS
- 1;
1423 const uint32_t max_workgroup_size
= 32 * devinfo
->max_cs_threads
;
1425 VkSampleCountFlags sample_counts
=
1426 isl_device_get_sample_counts(&pdevice
->isl_dev
);
1429 VkPhysicalDeviceLimits limits
= {
1430 .maxImageDimension1D
= (1 << 14),
1431 .maxImageDimension2D
= (1 << 14),
1432 .maxImageDimension3D
= (1 << 11),
1433 .maxImageDimensionCube
= (1 << 14),
1434 .maxImageArrayLayers
= (1 << 11),
1435 .maxTexelBufferElements
= 128 * 1024 * 1024,
1436 .maxUniformBufferRange
= (1ul << 27),
1437 .maxStorageBufferRange
= max_raw_buffer_sz
,
1438 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
1439 .maxMemoryAllocationCount
= UINT32_MAX
,
1440 .maxSamplerAllocationCount
= 64 * 1024,
1441 .bufferImageGranularity
= 64, /* A cache line */
1442 .sparseAddressSpaceSize
= 0,
1443 .maxBoundDescriptorSets
= MAX_SETS
,
1444 .maxPerStageDescriptorSamplers
= max_samplers
,
1445 .maxPerStageDescriptorUniformBuffers
= MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
,
1446 .maxPerStageDescriptorStorageBuffers
= max_ssbos
,
1447 .maxPerStageDescriptorSampledImages
= max_textures
,
1448 .maxPerStageDescriptorStorageImages
= max_images
,
1449 .maxPerStageDescriptorInputAttachments
= MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS
,
1450 .maxPerStageResources
= max_per_stage
,
1451 .maxDescriptorSetSamplers
= 6 * max_samplers
, /* number of stages * maxPerStageDescriptorSamplers */
1452 .maxDescriptorSetUniformBuffers
= 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
, /* number of stages * maxPerStageDescriptorUniformBuffers */
1453 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
1454 .maxDescriptorSetStorageBuffers
= 6 * max_ssbos
, /* number of stages * maxPerStageDescriptorStorageBuffers */
1455 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
1456 .maxDescriptorSetSampledImages
= 6 * max_textures
, /* number of stages * maxPerStageDescriptorSampledImages */
1457 .maxDescriptorSetStorageImages
= 6 * max_images
, /* number of stages * maxPerStageDescriptorStorageImages */
1458 .maxDescriptorSetInputAttachments
= MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS
,
1459 .maxVertexInputAttributes
= MAX_VBS
,
1460 .maxVertexInputBindings
= MAX_VBS
,
1461 .maxVertexInputAttributeOffset
= 2047,
1462 .maxVertexInputBindingStride
= 2048,
1463 .maxVertexOutputComponents
= 128,
1464 .maxTessellationGenerationLevel
= 64,
1465 .maxTessellationPatchSize
= 32,
1466 .maxTessellationControlPerVertexInputComponents
= 128,
1467 .maxTessellationControlPerVertexOutputComponents
= 128,
1468 .maxTessellationControlPerPatchOutputComponents
= 128,
1469 .maxTessellationControlTotalOutputComponents
= 2048,
1470 .maxTessellationEvaluationInputComponents
= 128,
1471 .maxTessellationEvaluationOutputComponents
= 128,
1472 .maxGeometryShaderInvocations
= 32,
1473 .maxGeometryInputComponents
= 64,
1474 .maxGeometryOutputComponents
= 128,
1475 .maxGeometryOutputVertices
= 256,
1476 .maxGeometryTotalOutputComponents
= 1024,
1477 .maxFragmentInputComponents
= 116, /* 128 components - (PSIZ, CLIP_DIST0, CLIP_DIST1) */
1478 .maxFragmentOutputAttachments
= 8,
1479 .maxFragmentDualSrcAttachments
= 1,
1480 .maxFragmentCombinedOutputResources
= 8,
1481 .maxComputeSharedMemorySize
= 64 * 1024,
1482 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
1483 .maxComputeWorkGroupInvocations
= max_workgroup_size
,
1484 .maxComputeWorkGroupSize
= {
1489 .subPixelPrecisionBits
= 8,
1490 .subTexelPrecisionBits
= 8,
1491 .mipmapPrecisionBits
= 8,
1492 .maxDrawIndexedIndexValue
= UINT32_MAX
,
1493 .maxDrawIndirectCount
= UINT32_MAX
,
1494 .maxSamplerLodBias
= 16,
1495 .maxSamplerAnisotropy
= 16,
1496 .maxViewports
= MAX_VIEWPORTS
,
1497 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
1498 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
1499 .viewportSubPixelBits
= 13, /* We take a float? */
1500 .minMemoryMapAlignment
= 4096, /* A page */
1501 /* The dataport requires texel alignment so we need to assume a worst
1502 * case of R32G32B32A32 which is 16 bytes.
1504 .minTexelBufferOffsetAlignment
= 16,
1505 /* We need 16 for UBO block reads to work and 32 for push UBOs */
1506 .minUniformBufferOffsetAlignment
= 32,
1507 .minStorageBufferOffsetAlignment
= 4,
1508 .minTexelOffset
= -8,
1509 .maxTexelOffset
= 7,
1510 .minTexelGatherOffset
= -32,
1511 .maxTexelGatherOffset
= 31,
1512 .minInterpolationOffset
= -0.5,
1513 .maxInterpolationOffset
= 0.4375,
1514 .subPixelInterpolationOffsetBits
= 4,
1515 .maxFramebufferWidth
= (1 << 14),
1516 .maxFramebufferHeight
= (1 << 14),
1517 .maxFramebufferLayers
= (1 << 11),
1518 .framebufferColorSampleCounts
= sample_counts
,
1519 .framebufferDepthSampleCounts
= sample_counts
,
1520 .framebufferStencilSampleCounts
= sample_counts
,
1521 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
1522 .maxColorAttachments
= MAX_RTS
,
1523 .sampledImageColorSampleCounts
= sample_counts
,
1524 .sampledImageIntegerSampleCounts
= sample_counts
,
1525 .sampledImageDepthSampleCounts
= sample_counts
,
1526 .sampledImageStencilSampleCounts
= sample_counts
,
1527 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
1528 .maxSampleMaskWords
= 1,
1529 .timestampComputeAndGraphics
= true,
1530 .timestampPeriod
= 1000000000.0 / devinfo
->timestamp_frequency
,
1531 .maxClipDistances
= 8,
1532 .maxCullDistances
= 8,
1533 .maxCombinedClipAndCullDistances
= 8,
1534 .discreteQueuePriorities
= 2,
1535 .pointSizeRange
= { 0.125, 255.875 },
1538 (devinfo
->gen
>= 9 || devinfo
->is_cherryview
) ?
1539 2047.9921875 : 7.9921875,
1541 .pointSizeGranularity
= (1.0 / 8.0),
1542 .lineWidthGranularity
= (1.0 / 128.0),
1543 .strictLines
= false,
1544 .standardSampleLocations
= true,
1545 .optimalBufferCopyOffsetAlignment
= 128,
1546 .optimalBufferCopyRowPitchAlignment
= 128,
1547 .nonCoherentAtomSize
= 64,
1550 *pProperties
= (VkPhysicalDeviceProperties
) {
1551 .apiVersion
= anv_physical_device_api_version(pdevice
),
1552 .driverVersion
= vk_get_driver_version(),
1554 .deviceID
= pdevice
->info
.chipset_id
,
1555 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
1557 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
1560 snprintf(pProperties
->deviceName
, sizeof(pProperties
->deviceName
),
1561 "%s", pdevice
->name
);
1562 memcpy(pProperties
->pipelineCacheUUID
,
1563 pdevice
->pipeline_cache_uuid
, VK_UUID_SIZE
);
1567 anv_get_physical_device_properties_1_1(struct anv_physical_device
*pdevice
,
1568 VkPhysicalDeviceVulkan11Properties
*p
)
1570 assert(p
->sType
== VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES
);
1572 memcpy(p
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
1573 memcpy(p
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
1574 memset(p
->deviceLUID
, 0, VK_LUID_SIZE
);
1575 p
->deviceNodeMask
= 0;
1576 p
->deviceLUIDValid
= false;
1578 p
->subgroupSize
= BRW_SUBGROUP_SIZE
;
1579 VkShaderStageFlags scalar_stages
= 0;
1580 for (unsigned stage
= 0; stage
< MESA_SHADER_STAGES
; stage
++) {
1581 if (pdevice
->compiler
->scalar_stage
[stage
])
1582 scalar_stages
|= mesa_to_vk_shader_stage(stage
);
1584 p
->subgroupSupportedStages
= scalar_stages
;
1585 p
->subgroupSupportedOperations
= VK_SUBGROUP_FEATURE_BASIC_BIT
|
1586 VK_SUBGROUP_FEATURE_VOTE_BIT
|
1587 VK_SUBGROUP_FEATURE_BALLOT_BIT
|
1588 VK_SUBGROUP_FEATURE_SHUFFLE_BIT
|
1589 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT
|
1590 VK_SUBGROUP_FEATURE_QUAD_BIT
;
1591 if (pdevice
->info
.gen
>= 8) {
1592 /* TODO: There's no technical reason why these can't be made to
1593 * work on gen7 but they don't at the moment so it's best to leave
1594 * the feature disabled than enabled and broken.
1596 p
->subgroupSupportedOperations
|= VK_SUBGROUP_FEATURE_ARITHMETIC_BIT
|
1597 VK_SUBGROUP_FEATURE_CLUSTERED_BIT
;
1599 p
->subgroupQuadOperationsInAllStages
= pdevice
->info
.gen
>= 8;
1601 p
->pointClippingBehavior
= VK_POINT_CLIPPING_BEHAVIOR_USER_CLIP_PLANES_ONLY
;
1602 p
->maxMultiviewViewCount
= 16;
1603 p
->maxMultiviewInstanceIndex
= UINT32_MAX
/ 16;
1604 p
->protectedNoFault
= false;
1605 /* This value doesn't matter for us today as our per-stage descriptors are
1608 p
->maxPerSetDescriptors
= 1024;
1609 p
->maxMemoryAllocationSize
= MAX_MEMORY_ALLOCATION_SIZE
;
1613 anv_get_physical_device_properties_1_2(struct anv_physical_device
*pdevice
,
1614 VkPhysicalDeviceVulkan12Properties
*p
)
1616 assert(p
->sType
== VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES
);
1618 p
->driverID
= VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA_KHR
;
1619 memset(p
->driverName
, 0, sizeof(p
->driverName
));
1620 snprintf(p
->driverName
, VK_MAX_DRIVER_NAME_SIZE_KHR
,
1621 "Intel open-source Mesa driver");
1622 memset(p
->driverInfo
, 0, sizeof(p
->driverInfo
));
1623 snprintf(p
->driverInfo
, VK_MAX_DRIVER_INFO_SIZE_KHR
,
1624 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1
);
1625 p
->conformanceVersion
= (VkConformanceVersionKHR
) {
1632 p
->denormBehaviorIndependence
=
1633 VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_ALL_KHR
;
1634 p
->roundingModeIndependence
=
1635 VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_NONE_KHR
;
1637 /* Broadwell does not support HF denorms and there are restrictions
1638 * other gens. According to Kabylake's PRM:
1640 * "math - Extended Math Function
1642 * Restriction : Half-float denorms are always retained."
1644 p
->shaderDenormFlushToZeroFloat16
= false;
1645 p
->shaderDenormPreserveFloat16
= pdevice
->info
.gen
> 8;
1646 p
->shaderRoundingModeRTEFloat16
= true;
1647 p
->shaderRoundingModeRTZFloat16
= true;
1648 p
->shaderSignedZeroInfNanPreserveFloat16
= true;
1650 p
->shaderDenormFlushToZeroFloat32
= true;
1651 p
->shaderDenormPreserveFloat32
= true;
1652 p
->shaderRoundingModeRTEFloat32
= true;
1653 p
->shaderRoundingModeRTZFloat32
= true;
1654 p
->shaderSignedZeroInfNanPreserveFloat32
= true;
1656 p
->shaderDenormFlushToZeroFloat64
= true;
1657 p
->shaderDenormPreserveFloat64
= true;
1658 p
->shaderRoundingModeRTEFloat64
= true;
1659 p
->shaderRoundingModeRTZFloat64
= true;
1660 p
->shaderSignedZeroInfNanPreserveFloat64
= true;
1662 /* It's a bit hard to exactly map our implementation to the limits
1663 * described here. The bindless surface handle in the extended
1664 * message descriptors is 20 bits and it's an index into the table of
1665 * RENDER_SURFACE_STATE structs that starts at bindless surface base
1666 * address. Given that most things consume two surface states per
1667 * view (general/sampled for textures and write-only/read-write for
1668 * images), we claim 2^19 things.
1670 * For SSBOs, we just use A64 messages so there is no real limit
1671 * there beyond the limit on the total size of a descriptor set.
1673 const unsigned max_bindless_views
= 1 << 19;
1674 p
->maxUpdateAfterBindDescriptorsInAllPools
= max_bindless_views
;
1675 p
->shaderUniformBufferArrayNonUniformIndexingNative
= false;
1676 p
->shaderSampledImageArrayNonUniformIndexingNative
= false;
1677 p
->shaderStorageBufferArrayNonUniformIndexingNative
= true;
1678 p
->shaderStorageImageArrayNonUniformIndexingNative
= false;
1679 p
->shaderInputAttachmentArrayNonUniformIndexingNative
= false;
1680 p
->robustBufferAccessUpdateAfterBind
= true;
1681 p
->quadDivergentImplicitLod
= false;
1682 p
->maxPerStageDescriptorUpdateAfterBindSamplers
= max_bindless_views
;
1683 p
->maxPerStageDescriptorUpdateAfterBindUniformBuffers
= MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
;
1684 p
->maxPerStageDescriptorUpdateAfterBindStorageBuffers
= UINT32_MAX
;
1685 p
->maxPerStageDescriptorUpdateAfterBindSampledImages
= max_bindless_views
;
1686 p
->maxPerStageDescriptorUpdateAfterBindStorageImages
= max_bindless_views
;
1687 p
->maxPerStageDescriptorUpdateAfterBindInputAttachments
= MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS
;
1688 p
->maxPerStageUpdateAfterBindResources
= UINT32_MAX
;
1689 p
->maxDescriptorSetUpdateAfterBindSamplers
= max_bindless_views
;
1690 p
->maxDescriptorSetUpdateAfterBindUniformBuffers
= 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
;
1691 p
->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2;
1692 p
->maxDescriptorSetUpdateAfterBindStorageBuffers
= UINT32_MAX
;
1693 p
->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2;
1694 p
->maxDescriptorSetUpdateAfterBindSampledImages
= max_bindless_views
;
1695 p
->maxDescriptorSetUpdateAfterBindStorageImages
= max_bindless_views
;
1696 p
->maxDescriptorSetUpdateAfterBindInputAttachments
= MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS
;
1698 /* We support all of the depth resolve modes */
1699 p
->supportedDepthResolveModes
= VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
|
1700 VK_RESOLVE_MODE_AVERAGE_BIT_KHR
|
1701 VK_RESOLVE_MODE_MIN_BIT_KHR
|
1702 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1703 /* Average doesn't make sense for stencil so we don't support that */
1704 p
->supportedStencilResolveModes
= VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
;
1705 if (pdevice
->info
.gen
>= 8) {
1706 /* The advanced stencil resolve modes currently require stencil
1707 * sampling be supported by the hardware.
1709 p
->supportedStencilResolveModes
|= VK_RESOLVE_MODE_MIN_BIT_KHR
|
1710 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1712 p
->independentResolveNone
= true;
1713 p
->independentResolve
= true;
1715 p
->filterMinmaxSingleComponentFormats
= pdevice
->info
.gen
>= 9;
1716 p
->filterMinmaxImageComponentMapping
= pdevice
->info
.gen
>= 9;
1718 p
->maxTimelineSemaphoreValueDifference
= UINT64_MAX
;
1720 p
->framebufferIntegerColorSampleCounts
=
1721 isl_device_get_sample_counts(&pdevice
->isl_dev
);
1724 void anv_GetPhysicalDeviceProperties2(
1725 VkPhysicalDevice physicalDevice
,
1726 VkPhysicalDeviceProperties2
* pProperties
)
1728 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1730 anv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
1732 VkPhysicalDeviceVulkan11Properties core_1_1
= {
1733 .sType
= VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES
,
1735 anv_get_physical_device_properties_1_1(pdevice
, &core_1_1
);
1737 VkPhysicalDeviceVulkan12Properties core_1_2
= {
1738 .sType
= VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES
,
1740 anv_get_physical_device_properties_1_2(pdevice
, &core_1_2
);
1742 #define CORE_RENAMED_PROPERTY(major, minor, ext_property, core_property) \
1743 memcpy(&properties->ext_property, &core_##major##_##minor.core_property, \
1744 sizeof(core_##major##_##minor.core_property))
1746 #define CORE_PROPERTY(major, minor, property) \
1747 CORE_RENAMED_PROPERTY(major, minor, property, property)
1749 vk_foreach_struct(ext
, pProperties
->pNext
) {
1750 switch (ext
->sType
) {
1751 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_STENCIL_RESOLVE_PROPERTIES_KHR
: {
1752 VkPhysicalDeviceDepthStencilResolvePropertiesKHR
*properties
=
1753 (VkPhysicalDeviceDepthStencilResolvePropertiesKHR
*)ext
;
1754 CORE_PROPERTY(1, 2, supportedDepthResolveModes
);
1755 CORE_PROPERTY(1, 2, supportedStencilResolveModes
);
1756 CORE_PROPERTY(1, 2, independentResolveNone
);
1757 CORE_PROPERTY(1, 2, independentResolve
);
1761 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES_EXT
: {
1762 VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*properties
=
1763 (VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*)ext
;
1764 CORE_PROPERTY(1, 2, maxUpdateAfterBindDescriptorsInAllPools
);
1765 CORE_PROPERTY(1, 2, shaderUniformBufferArrayNonUniformIndexingNative
);
1766 CORE_PROPERTY(1, 2, shaderSampledImageArrayNonUniformIndexingNative
);
1767 CORE_PROPERTY(1, 2, shaderStorageBufferArrayNonUniformIndexingNative
);
1768 CORE_PROPERTY(1, 2, shaderStorageImageArrayNonUniformIndexingNative
);
1769 CORE_PROPERTY(1, 2, shaderInputAttachmentArrayNonUniformIndexingNative
);
1770 CORE_PROPERTY(1, 2, robustBufferAccessUpdateAfterBind
);
1771 CORE_PROPERTY(1, 2, quadDivergentImplicitLod
);
1772 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindSamplers
);
1773 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindUniformBuffers
);
1774 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindStorageBuffers
);
1775 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindSampledImages
);
1776 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindStorageImages
);
1777 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindInputAttachments
);
1778 CORE_PROPERTY(1, 2, maxPerStageUpdateAfterBindResources
);
1779 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindSamplers
);
1780 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindUniformBuffers
);
1781 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindUniformBuffersDynamic
);
1782 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindStorageBuffers
);
1783 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindStorageBuffersDynamic
);
1784 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindSampledImages
);
1785 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindStorageImages
);
1786 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindInputAttachments
);
1790 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR
: {
1791 VkPhysicalDeviceDriverPropertiesKHR
*properties
=
1792 (VkPhysicalDeviceDriverPropertiesKHR
*) ext
;
1793 CORE_PROPERTY(1, 2, driverID
);
1794 CORE_PROPERTY(1, 2, driverName
);
1795 CORE_PROPERTY(1, 2, driverInfo
);
1796 CORE_PROPERTY(1, 2, conformanceVersion
);
1800 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT
: {
1801 VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*props
=
1802 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*) ext
;
1803 /* Userptr needs page aligned memory. */
1804 props
->minImportedHostPointerAlignment
= 4096;
1808 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES
: {
1809 VkPhysicalDeviceIDProperties
*properties
=
1810 (VkPhysicalDeviceIDProperties
*)ext
;
1811 CORE_PROPERTY(1, 1, deviceUUID
);
1812 CORE_PROPERTY(1, 1, driverUUID
);
1813 CORE_PROPERTY(1, 1, deviceLUID
);
1814 CORE_PROPERTY(1, 1, deviceLUIDValid
);
1818 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_PROPERTIES_EXT
: {
1819 VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*props
=
1820 (VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*)ext
;
1821 props
->maxInlineUniformBlockSize
= MAX_INLINE_UNIFORM_BLOCK_SIZE
;
1822 props
->maxPerStageDescriptorInlineUniformBlocks
=
1823 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1824 props
->maxPerStageDescriptorUpdateAfterBindInlineUniformBlocks
=
1825 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1826 props
->maxDescriptorSetInlineUniformBlocks
=
1827 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1828 props
->maxDescriptorSetUpdateAfterBindInlineUniformBlocks
=
1829 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1833 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_PROPERTIES_EXT
: {
1834 VkPhysicalDeviceLineRasterizationPropertiesEXT
*props
=
1835 (VkPhysicalDeviceLineRasterizationPropertiesEXT
*)ext
;
1836 /* In the Skylake PRM Vol. 7, subsection titled "GIQ (Diamond)
1837 * Sampling Rules - Legacy Mode", it says the following:
1839 * "Note that the device divides a pixel into a 16x16 array of
1840 * subpixels, referenced by their upper left corners."
1842 * This is the only known reference in the PRMs to the subpixel
1843 * precision of line rasterization and a "16x16 array of subpixels"
1844 * implies 4 subpixel precision bits. Empirical testing has shown
1845 * that 4 subpixel precision bits applies to all line rasterization
1848 props
->lineSubPixelPrecisionBits
= 4;
1852 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES
: {
1853 VkPhysicalDeviceMaintenance3Properties
*properties
=
1854 (VkPhysicalDeviceMaintenance3Properties
*)ext
;
1855 /* This value doesn't matter for us today as our per-stage
1856 * descriptors are the real limit.
1858 CORE_PROPERTY(1, 1, maxPerSetDescriptors
);
1859 CORE_PROPERTY(1, 1, maxMemoryAllocationSize
);
1863 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES
: {
1864 VkPhysicalDeviceMultiviewProperties
*properties
=
1865 (VkPhysicalDeviceMultiviewProperties
*)ext
;
1866 CORE_PROPERTY(1, 1, maxMultiviewViewCount
);
1867 CORE_PROPERTY(1, 1, maxMultiviewInstanceIndex
);
1871 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT
: {
1872 VkPhysicalDevicePCIBusInfoPropertiesEXT
*properties
=
1873 (VkPhysicalDevicePCIBusInfoPropertiesEXT
*)ext
;
1874 properties
->pciDomain
= pdevice
->pci_info
.domain
;
1875 properties
->pciBus
= pdevice
->pci_info
.bus
;
1876 properties
->pciDevice
= pdevice
->pci_info
.device
;
1877 properties
->pciFunction
= pdevice
->pci_info
.function
;
1881 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES
: {
1882 VkPhysicalDevicePointClippingProperties
*properties
=
1883 (VkPhysicalDevicePointClippingProperties
*) ext
;
1884 CORE_PROPERTY(1, 1, pointClippingBehavior
);
1888 #pragma GCC diagnostic push
1889 #pragma GCC diagnostic ignored "-Wswitch"
1890 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PRESENTATION_PROPERTIES_ANDROID
: {
1891 VkPhysicalDevicePresentationPropertiesANDROID
*props
=
1892 (VkPhysicalDevicePresentationPropertiesANDROID
*)ext
;
1893 props
->sharedImage
= VK_FALSE
;
1896 #pragma GCC diagnostic pop
1898 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES
: {
1899 VkPhysicalDeviceProtectedMemoryProperties
*properties
=
1900 (VkPhysicalDeviceProtectedMemoryProperties
*)ext
;
1901 CORE_PROPERTY(1, 1, protectedNoFault
);
1905 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
1906 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
1907 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
1908 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
1912 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT
: {
1913 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*properties
=
1914 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*)ext
;
1915 CORE_PROPERTY(1, 2, filterMinmaxImageComponentMapping
);
1916 CORE_PROPERTY(1, 2, filterMinmaxSingleComponentFormats
);
1920 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES
: {
1921 VkPhysicalDeviceSubgroupProperties
*properties
= (void *)ext
;
1922 CORE_PROPERTY(1, 1, subgroupSize
);
1923 CORE_RENAMED_PROPERTY(1, 1, supportedStages
,
1924 subgroupSupportedStages
);
1925 CORE_RENAMED_PROPERTY(1, 1, supportedOperations
,
1926 subgroupSupportedOperations
);
1927 CORE_RENAMED_PROPERTY(1, 1, quadOperationsInAllStages
,
1928 subgroupQuadOperationsInAllStages
);
1932 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_PROPERTIES_EXT
: {
1933 VkPhysicalDeviceSubgroupSizeControlPropertiesEXT
*props
=
1934 (VkPhysicalDeviceSubgroupSizeControlPropertiesEXT
*)ext
;
1935 STATIC_ASSERT(8 <= BRW_SUBGROUP_SIZE
&& BRW_SUBGROUP_SIZE
<= 32);
1936 props
->minSubgroupSize
= 8;
1937 props
->maxSubgroupSize
= 32;
1938 props
->maxComputeWorkgroupSubgroups
= pdevice
->info
.max_cs_threads
;
1939 props
->requiredSubgroupSizeStages
= VK_SHADER_STAGE_COMPUTE_BIT
;
1942 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT_CONTROLS_PROPERTIES_KHR
: {
1943 VkPhysicalDeviceFloatControlsPropertiesKHR
*properties
= (void *)ext
;
1944 CORE_PROPERTY(1, 2, denormBehaviorIndependence
);
1945 CORE_PROPERTY(1, 2, roundingModeIndependence
);
1946 CORE_PROPERTY(1, 2, shaderDenormFlushToZeroFloat16
);
1947 CORE_PROPERTY(1, 2, shaderDenormPreserveFloat16
);
1948 CORE_PROPERTY(1, 2, shaderRoundingModeRTEFloat16
);
1949 CORE_PROPERTY(1, 2, shaderRoundingModeRTZFloat16
);
1950 CORE_PROPERTY(1, 2, shaderSignedZeroInfNanPreserveFloat16
);
1951 CORE_PROPERTY(1, 2, shaderDenormFlushToZeroFloat32
);
1952 CORE_PROPERTY(1, 2, shaderDenormPreserveFloat32
);
1953 CORE_PROPERTY(1, 2, shaderRoundingModeRTEFloat32
);
1954 CORE_PROPERTY(1, 2, shaderRoundingModeRTZFloat32
);
1955 CORE_PROPERTY(1, 2, shaderSignedZeroInfNanPreserveFloat32
);
1956 CORE_PROPERTY(1, 2, shaderDenormFlushToZeroFloat64
);
1957 CORE_PROPERTY(1, 2, shaderDenormPreserveFloat64
);
1958 CORE_PROPERTY(1, 2, shaderRoundingModeRTEFloat64
);
1959 CORE_PROPERTY(1, 2, shaderRoundingModeRTZFloat64
);
1960 CORE_PROPERTY(1, 2, shaderSignedZeroInfNanPreserveFloat64
);
1964 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_PROPERTIES_EXT
: {
1965 VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT
*props
=
1966 (VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT
*)ext
;
1968 /* From the SKL PRM Vol. 2d, docs for RENDER_SURFACE_STATE::Surface
1971 * "For SURFTYPE_BUFFER non-rendertarget surfaces, this field
1972 * specifies the base address of the first element of the surface,
1973 * computed in software by adding the surface base address to the
1974 * byte offset of the element in the buffer. The base address must
1975 * be aligned to element size."
1977 * The typed dataport messages require that things be texel aligned.
1978 * Otherwise, we may just load/store the wrong data or, in the worst
1979 * case, there may be hangs.
1981 props
->storageTexelBufferOffsetAlignmentBytes
= 16;
1982 props
->storageTexelBufferOffsetSingleTexelAlignment
= true;
1984 /* The sampler, however, is much more forgiving and it can handle
1985 * arbitrary byte alignment for linear and buffer surfaces. It's
1986 * hard to find a good PRM citation for this but years of empirical
1987 * experience demonstrate that this is true.
1989 props
->uniformTexelBufferOffsetAlignmentBytes
= 1;
1990 props
->uniformTexelBufferOffsetSingleTexelAlignment
= false;
1994 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_PROPERTIES_KHR
: {
1995 VkPhysicalDeviceTimelineSemaphorePropertiesKHR
*properties
=
1996 (VkPhysicalDeviceTimelineSemaphorePropertiesKHR
*) ext
;
1997 CORE_PROPERTY(1, 2, maxTimelineSemaphoreValueDifference
);
2001 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT
: {
2002 VkPhysicalDeviceTransformFeedbackPropertiesEXT
*props
=
2003 (VkPhysicalDeviceTransformFeedbackPropertiesEXT
*)ext
;
2005 props
->maxTransformFeedbackStreams
= MAX_XFB_STREAMS
;
2006 props
->maxTransformFeedbackBuffers
= MAX_XFB_BUFFERS
;
2007 props
->maxTransformFeedbackBufferSize
= (1ull << 32);
2008 props
->maxTransformFeedbackStreamDataSize
= 128 * 4;
2009 props
->maxTransformFeedbackBufferDataSize
= 128 * 4;
2010 props
->maxTransformFeedbackBufferDataStride
= 2048;
2011 props
->transformFeedbackQueries
= true;
2012 props
->transformFeedbackStreamsLinesTriangles
= false;
2013 props
->transformFeedbackRasterizationStreamSelect
= false;
2014 props
->transformFeedbackDraw
= true;
2018 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT
: {
2019 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*props
=
2020 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*)ext
;
2021 /* We have to restrict this a bit for multiview */
2022 props
->maxVertexAttribDivisor
= UINT32_MAX
/ 16;
2026 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES
:
2027 anv_get_physical_device_properties_1_1(pdevice
, (void *)ext
);
2030 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES
:
2031 anv_get_physical_device_properties_1_2(pdevice
, (void *)ext
);
2035 anv_debug_ignored_stype(ext
->sType
);
2040 #undef CORE_RENAMED_PROPERTY
2041 #undef CORE_PROPERTY
2044 /* We support exactly one queue family. */
2045 static const VkQueueFamilyProperties
2046 anv_queue_family_properties
= {
2047 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
2048 VK_QUEUE_COMPUTE_BIT
|
2049 VK_QUEUE_TRANSFER_BIT
,
2051 .timestampValidBits
= 36, /* XXX: Real value here */
2052 .minImageTransferGranularity
= { 1, 1, 1 },
2055 void anv_GetPhysicalDeviceQueueFamilyProperties(
2056 VkPhysicalDevice physicalDevice
,
2058 VkQueueFamilyProperties
* pQueueFamilyProperties
)
2060 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pCount
);
2062 vk_outarray_append(&out
, p
) {
2063 *p
= anv_queue_family_properties
;
2067 void anv_GetPhysicalDeviceQueueFamilyProperties2(
2068 VkPhysicalDevice physicalDevice
,
2069 uint32_t* pQueueFamilyPropertyCount
,
2070 VkQueueFamilyProperties2
* pQueueFamilyProperties
)
2073 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pQueueFamilyPropertyCount
);
2075 vk_outarray_append(&out
, p
) {
2076 p
->queueFamilyProperties
= anv_queue_family_properties
;
2078 vk_foreach_struct(s
, p
->pNext
) {
2079 anv_debug_ignored_stype(s
->sType
);
2084 void anv_GetPhysicalDeviceMemoryProperties(
2085 VkPhysicalDevice physicalDevice
,
2086 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
2088 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
2090 pMemoryProperties
->memoryTypeCount
= physical_device
->memory
.type_count
;
2091 for (uint32_t i
= 0; i
< physical_device
->memory
.type_count
; i
++) {
2092 pMemoryProperties
->memoryTypes
[i
] = (VkMemoryType
) {
2093 .propertyFlags
= physical_device
->memory
.types
[i
].propertyFlags
,
2094 .heapIndex
= physical_device
->memory
.types
[i
].heapIndex
,
2098 pMemoryProperties
->memoryHeapCount
= physical_device
->memory
.heap_count
;
2099 for (uint32_t i
= 0; i
< physical_device
->memory
.heap_count
; i
++) {
2100 pMemoryProperties
->memoryHeaps
[i
] = (VkMemoryHeap
) {
2101 .size
= physical_device
->memory
.heaps
[i
].size
,
2102 .flags
= physical_device
->memory
.heaps
[i
].flags
,
2108 anv_get_memory_budget(VkPhysicalDevice physicalDevice
,
2109 VkPhysicalDeviceMemoryBudgetPropertiesEXT
*memoryBudget
)
2111 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
2112 uint64_t sys_available
= get_available_system_memory();
2113 assert(sys_available
> 0);
2115 VkDeviceSize total_heaps_size
= 0;
2116 for (size_t i
= 0; i
< device
->memory
.heap_count
; i
++)
2117 total_heaps_size
+= device
->memory
.heaps
[i
].size
;
2119 for (size_t i
= 0; i
< device
->memory
.heap_count
; i
++) {
2120 VkDeviceSize heap_size
= device
->memory
.heaps
[i
].size
;
2121 VkDeviceSize heap_used
= device
->memory
.heaps
[i
].used
;
2122 VkDeviceSize heap_budget
;
2124 double heap_proportion
= (double) heap_size
/ total_heaps_size
;
2125 VkDeviceSize sys_available_prop
= sys_available
* heap_proportion
;
2128 * Let's not incite the app to starve the system: report at most 90% of
2129 * available system memory.
2131 uint64_t heap_available
= sys_available_prop
* 9 / 10;
2132 heap_budget
= MIN2(heap_size
, heap_used
+ heap_available
);
2135 * Round down to the nearest MB
2137 heap_budget
&= ~((1ull << 20) - 1);
2140 * The heapBudget value must be non-zero for array elements less than
2141 * VkPhysicalDeviceMemoryProperties::memoryHeapCount. The heapBudget
2142 * value must be less than or equal to VkMemoryHeap::size for each heap.
2144 assert(0 < heap_budget
&& heap_budget
<= heap_size
);
2146 memoryBudget
->heapUsage
[i
] = heap_used
;
2147 memoryBudget
->heapBudget
[i
] = heap_budget
;
2150 /* The heapBudget and heapUsage values must be zero for array elements
2151 * greater than or equal to VkPhysicalDeviceMemoryProperties::memoryHeapCount
2153 for (uint32_t i
= device
->memory
.heap_count
; i
< VK_MAX_MEMORY_HEAPS
; i
++) {
2154 memoryBudget
->heapBudget
[i
] = 0;
2155 memoryBudget
->heapUsage
[i
] = 0;
2159 void anv_GetPhysicalDeviceMemoryProperties2(
2160 VkPhysicalDevice physicalDevice
,
2161 VkPhysicalDeviceMemoryProperties2
* pMemoryProperties
)
2163 anv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
2164 &pMemoryProperties
->memoryProperties
);
2166 vk_foreach_struct(ext
, pMemoryProperties
->pNext
) {
2167 switch (ext
->sType
) {
2168 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT
:
2169 anv_get_memory_budget(physicalDevice
, (void*)ext
);
2172 anv_debug_ignored_stype(ext
->sType
);
2179 anv_GetDeviceGroupPeerMemoryFeatures(
2182 uint32_t localDeviceIndex
,
2183 uint32_t remoteDeviceIndex
,
2184 VkPeerMemoryFeatureFlags
* pPeerMemoryFeatures
)
2186 assert(localDeviceIndex
== 0 && remoteDeviceIndex
== 0);
2187 *pPeerMemoryFeatures
= VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT
|
2188 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT
|
2189 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT
|
2190 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT
;
2193 PFN_vkVoidFunction
anv_GetInstanceProcAddr(
2194 VkInstance _instance
,
2197 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2199 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
2200 * when we have to return valid function pointers, NULL, or it's left
2201 * undefined. See the table for exact details.
2206 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
2207 if (strcmp(pName, "vk" #entrypoint) == 0) \
2208 return (PFN_vkVoidFunction)anv_##entrypoint
2210 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties
);
2211 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties
);
2212 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceVersion
);
2213 LOOKUP_ANV_ENTRYPOINT(CreateInstance
);
2215 #undef LOOKUP_ANV_ENTRYPOINT
2217 if (instance
== NULL
)
2220 int idx
= anv_get_instance_entrypoint_index(pName
);
2222 return instance
->dispatch
.entrypoints
[idx
];
2224 idx
= anv_get_physical_device_entrypoint_index(pName
);
2226 return instance
->physical_device_dispatch
.entrypoints
[idx
];
2228 idx
= anv_get_device_entrypoint_index(pName
);
2230 return instance
->device_dispatch
.entrypoints
[idx
];
2235 /* With version 1+ of the loader interface the ICD should expose
2236 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
2239 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
2240 VkInstance instance
,
2244 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
2245 VkInstance instance
,
2248 return anv_GetInstanceProcAddr(instance
, pName
);
2251 PFN_vkVoidFunction
anv_GetDeviceProcAddr(
2255 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2257 if (!device
|| !pName
)
2260 int idx
= anv_get_device_entrypoint_index(pName
);
2264 return device
->dispatch
.entrypoints
[idx
];
2267 /* With version 4+ of the loader interface the ICD should expose
2268 * vk_icdGetPhysicalDeviceProcAddr()
2271 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetPhysicalDeviceProcAddr(
2272 VkInstance _instance
,
2275 PFN_vkVoidFunction
vk_icdGetPhysicalDeviceProcAddr(
2276 VkInstance _instance
,
2279 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2281 if (!pName
|| !instance
)
2284 int idx
= anv_get_physical_device_entrypoint_index(pName
);
2288 return instance
->physical_device_dispatch
.entrypoints
[idx
];
2293 anv_CreateDebugReportCallbackEXT(VkInstance _instance
,
2294 const VkDebugReportCallbackCreateInfoEXT
* pCreateInfo
,
2295 const VkAllocationCallbacks
* pAllocator
,
2296 VkDebugReportCallbackEXT
* pCallback
)
2298 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2299 return vk_create_debug_report_callback(&instance
->debug_report_callbacks
,
2300 pCreateInfo
, pAllocator
, &instance
->alloc
,
2305 anv_DestroyDebugReportCallbackEXT(VkInstance _instance
,
2306 VkDebugReportCallbackEXT _callback
,
2307 const VkAllocationCallbacks
* pAllocator
)
2309 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2310 vk_destroy_debug_report_callback(&instance
->debug_report_callbacks
,
2311 _callback
, pAllocator
, &instance
->alloc
);
2315 anv_DebugReportMessageEXT(VkInstance _instance
,
2316 VkDebugReportFlagsEXT flags
,
2317 VkDebugReportObjectTypeEXT objectType
,
2320 int32_t messageCode
,
2321 const char* pLayerPrefix
,
2322 const char* pMessage
)
2324 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2325 vk_debug_report(&instance
->debug_report_callbacks
, flags
, objectType
,
2326 object
, location
, messageCode
, pLayerPrefix
, pMessage
);
2329 static struct anv_state
2330 anv_state_pool_emit_data(struct anv_state_pool
*pool
, size_t size
, size_t align
, const void *p
)
2332 struct anv_state state
;
2334 state
= anv_state_pool_alloc(pool
, size
, align
);
2335 memcpy(state
.map
, p
, size
);
2340 /* Haswell border color is a bit of a disaster. Float and unorm formats use a
2341 * straightforward 32-bit float color in the first 64 bytes. Instead of using
2342 * a nice float/integer union like Gen8+, Haswell specifies the integer border
2343 * color as a separate entry /after/ the float color. The layout of this entry
2344 * also depends on the format's bpp (with extra hacks for RG32), and overlaps.
2346 * Since we don't know the format/bpp, we can't make any of the border colors
2347 * containing '1' work for all formats, as it would be in the wrong place for
2348 * some of them. We opt to make 32-bit integers work as this seems like the
2349 * most common option. Fortunately, transparent black works regardless, as
2350 * all zeroes is the same in every bit-size.
2352 struct hsw_border_color
{
2356 uint32_t _pad1
[108];
2359 struct gen8_border_color
{
2364 /* Pad out to 64 bytes */
2369 anv_device_init_border_colors(struct anv_device
*device
)
2371 if (device
->info
.is_haswell
) {
2372 static const struct hsw_border_color border_colors
[] = {
2373 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
2374 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
2375 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
2376 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
2377 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
2378 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
2381 device
->border_colors
=
2382 anv_state_pool_emit_data(&device
->dynamic_state_pool
,
2383 sizeof(border_colors
), 512, border_colors
);
2385 static const struct gen8_border_color border_colors
[] = {
2386 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
2387 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
2388 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
2389 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
2390 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
2391 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
2394 device
->border_colors
=
2395 anv_state_pool_emit_data(&device
->dynamic_state_pool
,
2396 sizeof(border_colors
), 64, border_colors
);
2401 anv_device_init_trivial_batch(struct anv_device
*device
)
2403 VkResult result
= anv_device_alloc_bo(device
, 4096,
2404 ANV_BO_ALLOC_MAPPED
,
2405 0 /* explicit_address */,
2406 &device
->trivial_batch_bo
);
2407 if (result
!= VK_SUCCESS
)
2410 struct anv_batch batch
= {
2411 .start
= device
->trivial_batch_bo
->map
,
2412 .next
= device
->trivial_batch_bo
->map
,
2413 .end
= device
->trivial_batch_bo
->map
+ 4096,
2416 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
2417 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
2419 if (!device
->info
.has_llc
)
2420 gen_clflush_range(batch
.start
, batch
.next
- batch
.start
);
2425 VkResult
anv_EnumerateDeviceExtensionProperties(
2426 VkPhysicalDevice physicalDevice
,
2427 const char* pLayerName
,
2428 uint32_t* pPropertyCount
,
2429 VkExtensionProperties
* pProperties
)
2431 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
2432 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
2434 for (int i
= 0; i
< ANV_DEVICE_EXTENSION_COUNT
; i
++) {
2435 if (device
->supported_extensions
.extensions
[i
]) {
2436 vk_outarray_append(&out
, prop
) {
2437 *prop
= anv_device_extensions
[i
];
2442 return vk_outarray_status(&out
);
2446 anv_device_init_dispatch(struct anv_device
*device
)
2448 const struct anv_instance
*instance
= device
->physical
->instance
;
2450 const struct anv_device_dispatch_table
*genX_table
;
2451 switch (device
->info
.gen
) {
2453 genX_table
= &gen12_device_dispatch_table
;
2456 genX_table
= &gen11_device_dispatch_table
;
2459 genX_table
= &gen10_device_dispatch_table
;
2462 genX_table
= &gen9_device_dispatch_table
;
2465 genX_table
= &gen8_device_dispatch_table
;
2468 if (device
->info
.is_haswell
)
2469 genX_table
= &gen75_device_dispatch_table
;
2471 genX_table
= &gen7_device_dispatch_table
;
2474 unreachable("unsupported gen\n");
2477 for (unsigned i
= 0; i
< ARRAY_SIZE(device
->dispatch
.entrypoints
); i
++) {
2478 /* Vulkan requires that entrypoints for extensions which have not been
2479 * enabled must not be advertised.
2481 if (!anv_device_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
2482 &instance
->enabled_extensions
,
2483 &device
->enabled_extensions
)) {
2484 device
->dispatch
.entrypoints
[i
] = NULL
;
2485 } else if (genX_table
->entrypoints
[i
]) {
2486 device
->dispatch
.entrypoints
[i
] = genX_table
->entrypoints
[i
];
2488 device
->dispatch
.entrypoints
[i
] =
2489 anv_device_dispatch_table
.entrypoints
[i
];
2495 vk_priority_to_gen(int priority
)
2498 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT
:
2499 return GEN_CONTEXT_LOW_PRIORITY
;
2500 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
:
2501 return GEN_CONTEXT_MEDIUM_PRIORITY
;
2502 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT
:
2503 return GEN_CONTEXT_HIGH_PRIORITY
;
2504 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT
:
2505 return GEN_CONTEXT_REALTIME_PRIORITY
;
2507 unreachable("Invalid priority");
2512 anv_device_init_hiz_clear_value_bo(struct anv_device
*device
)
2514 VkResult result
= anv_device_alloc_bo(device
, 4096,
2515 ANV_BO_ALLOC_MAPPED
,
2516 0 /* explicit_address */,
2517 &device
->hiz_clear_bo
);
2518 if (result
!= VK_SUCCESS
)
2521 union isl_color_value hiz_clear
= { .u32
= { 0, } };
2522 hiz_clear
.f32
[0] = ANV_HZ_FC_VAL
;
2524 memcpy(device
->hiz_clear_bo
->map
, hiz_clear
.u32
, sizeof(hiz_clear
.u32
));
2526 if (!device
->info
.has_llc
)
2527 gen_clflush_range(device
->hiz_clear_bo
->map
, sizeof(hiz_clear
.u32
));
2533 get_bo_from_pool(struct gen_batch_decode_bo
*ret
,
2534 struct anv_block_pool
*pool
,
2537 anv_block_pool_foreach_bo(bo
, pool
) {
2538 uint64_t bo_address
= gen_48b_address(bo
->offset
);
2539 if (address
>= bo_address
&& address
< (bo_address
+ bo
->size
)) {
2540 *ret
= (struct gen_batch_decode_bo
) {
2551 /* Finding a buffer for batch decoding */
2552 static struct gen_batch_decode_bo
2553 decode_get_bo(void *v_batch
, bool ppgtt
, uint64_t address
)
2555 struct anv_device
*device
= v_batch
;
2556 struct gen_batch_decode_bo ret_bo
= {};
2560 if (get_bo_from_pool(&ret_bo
, &device
->dynamic_state_pool
.block_pool
, address
))
2562 if (get_bo_from_pool(&ret_bo
, &device
->instruction_state_pool
.block_pool
, address
))
2564 if (get_bo_from_pool(&ret_bo
, &device
->binding_table_pool
.block_pool
, address
))
2566 if (get_bo_from_pool(&ret_bo
, &device
->surface_state_pool
.block_pool
, address
))
2569 if (!device
->cmd_buffer_being_decoded
)
2570 return (struct gen_batch_decode_bo
) { };
2572 struct anv_batch_bo
**bo
;
2574 u_vector_foreach(bo
, &device
->cmd_buffer_being_decoded
->seen_bbos
) {
2575 /* The decoder zeroes out the top 16 bits, so we need to as well */
2576 uint64_t bo_address
= (*bo
)->bo
->offset
& (~0ull >> 16);
2578 if (address
>= bo_address
&& address
< bo_address
+ (*bo
)->bo
->size
) {
2579 return (struct gen_batch_decode_bo
) {
2581 .size
= (*bo
)->bo
->size
,
2582 .map
= (*bo
)->bo
->map
,
2587 return (struct gen_batch_decode_bo
) { };
2590 struct gen_aux_map_buffer
{
2591 struct gen_buffer base
;
2592 struct anv_state state
;
2595 static struct gen_buffer
*
2596 gen_aux_map_buffer_alloc(void *driver_ctx
, uint32_t size
)
2598 struct gen_aux_map_buffer
*buf
= malloc(sizeof(struct gen_aux_map_buffer
));
2602 struct anv_device
*device
= (struct anv_device
*)driver_ctx
;
2603 assert(device
->physical
->supports_48bit_addresses
&&
2604 device
->physical
->use_softpin
);
2606 struct anv_state_pool
*pool
= &device
->dynamic_state_pool
;
2607 buf
->state
= anv_state_pool_alloc(pool
, size
, size
);
2609 buf
->base
.gpu
= pool
->block_pool
.bo
->offset
+ buf
->state
.offset
;
2610 buf
->base
.gpu_end
= buf
->base
.gpu
+ buf
->state
.alloc_size
;
2611 buf
->base
.map
= buf
->state
.map
;
2612 buf
->base
.driver_bo
= &buf
->state
;
2617 gen_aux_map_buffer_free(void *driver_ctx
, struct gen_buffer
*buffer
)
2619 struct gen_aux_map_buffer
*buf
= (struct gen_aux_map_buffer
*)buffer
;
2620 struct anv_device
*device
= (struct anv_device
*)driver_ctx
;
2621 struct anv_state_pool
*pool
= &device
->dynamic_state_pool
;
2622 anv_state_pool_free(pool
, buf
->state
);
2626 static struct gen_mapped_pinned_buffer_alloc aux_map_allocator
= {
2627 .alloc
= gen_aux_map_buffer_alloc
,
2628 .free
= gen_aux_map_buffer_free
,
2631 VkResult
anv_CreateDevice(
2632 VkPhysicalDevice physicalDevice
,
2633 const VkDeviceCreateInfo
* pCreateInfo
,
2634 const VkAllocationCallbacks
* pAllocator
,
2637 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
2639 struct anv_device
*device
;
2641 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO
);
2643 struct anv_device_extension_table enabled_extensions
= { };
2644 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
2646 for (idx
= 0; idx
< ANV_DEVICE_EXTENSION_COUNT
; idx
++) {
2647 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
2648 anv_device_extensions
[idx
].extensionName
) == 0)
2652 if (idx
>= ANV_DEVICE_EXTENSION_COUNT
)
2653 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
2655 if (!physical_device
->supported_extensions
.extensions
[idx
])
2656 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
2658 enabled_extensions
.extensions
[idx
] = true;
2661 /* Check enabled features */
2662 if (pCreateInfo
->pEnabledFeatures
) {
2663 VkPhysicalDeviceFeatures supported_features
;
2664 anv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
2665 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
2666 VkBool32
*enabled_feature
= (VkBool32
*)pCreateInfo
->pEnabledFeatures
;
2667 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
2668 for (uint32_t i
= 0; i
< num_features
; i
++) {
2669 if (enabled_feature
[i
] && !supported_feature
[i
])
2670 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
2674 /* Check requested queues and fail if we are requested to create any
2675 * queues with flags we don't support.
2677 assert(pCreateInfo
->queueCreateInfoCount
> 0);
2678 for (uint32_t i
= 0; i
< pCreateInfo
->queueCreateInfoCount
; i
++) {
2679 if (pCreateInfo
->pQueueCreateInfos
[i
].flags
!= 0)
2680 return vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2683 /* Check if client specified queue priority. */
2684 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*queue_priority
=
2685 vk_find_struct_const(pCreateInfo
->pQueueCreateInfos
[0].pNext
,
2686 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT
);
2688 VkQueueGlobalPriorityEXT priority
=
2689 queue_priority
? queue_priority
->globalPriority
:
2690 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
;
2692 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
2694 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
2696 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2698 if (INTEL_DEBUG
& DEBUG_BATCH
) {
2699 const unsigned decode_flags
=
2700 GEN_BATCH_DECODE_FULL
|
2701 ((INTEL_DEBUG
& DEBUG_COLOR
) ? GEN_BATCH_DECODE_IN_COLOR
: 0) |
2702 GEN_BATCH_DECODE_OFFSETS
|
2703 GEN_BATCH_DECODE_FLOATS
;
2705 gen_batch_decode_ctx_init(&device
->decoder_ctx
,
2706 &physical_device
->info
,
2707 stderr
, decode_flags
, NULL
,
2708 decode_get_bo
, NULL
, device
);
2711 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
2712 device
->physical
= physical_device
;
2713 device
->no_hw
= physical_device
->no_hw
;
2714 device
->_lost
= false;
2717 device
->alloc
= *pAllocator
;
2719 device
->alloc
= physical_device
->instance
->alloc
;
2721 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
2722 device
->fd
= open(physical_device
->path
, O_RDWR
| O_CLOEXEC
);
2723 if (device
->fd
== -1) {
2724 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2728 device
->context_id
= anv_gem_create_context(device
);
2729 if (device
->context_id
== -1) {
2730 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2734 result
= anv_queue_init(device
, &device
->queue
);
2735 if (result
!= VK_SUCCESS
)
2736 goto fail_context_id
;
2738 if (physical_device
->use_softpin
) {
2739 if (pthread_mutex_init(&device
->vma_mutex
, NULL
) != 0) {
2740 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2744 /* keep the page with address zero out of the allocator */
2745 util_vma_heap_init(&device
->vma_lo
,
2746 LOW_HEAP_MIN_ADDRESS
, LOW_HEAP_SIZE
);
2748 util_vma_heap_init(&device
->vma_cva
, CLIENT_VISIBLE_HEAP_MIN_ADDRESS
,
2749 CLIENT_VISIBLE_HEAP_SIZE
);
2751 /* Leave the last 4GiB out of the high vma range, so that no state
2752 * base address + size can overflow 48 bits. For more information see
2753 * the comment about Wa32bitGeneralStateOffset in anv_allocator.c
2755 util_vma_heap_init(&device
->vma_hi
, HIGH_HEAP_MIN_ADDRESS
,
2756 physical_device
->gtt_size
- (1ull << 32) -
2757 HIGH_HEAP_MIN_ADDRESS
);
2760 list_inithead(&device
->memory_objects
);
2762 /* As per spec, the driver implementation may deny requests to acquire
2763 * a priority above the default priority (MEDIUM) if the caller does not
2764 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
2767 if (physical_device
->has_context_priority
) {
2768 int err
= anv_gem_set_context_param(device
->fd
, device
->context_id
,
2769 I915_CONTEXT_PARAM_PRIORITY
,
2770 vk_priority_to_gen(priority
));
2771 if (err
!= 0 && priority
> VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
) {
2772 result
= vk_error(VK_ERROR_NOT_PERMITTED_EXT
);
2777 device
->info
= physical_device
->info
;
2778 device
->isl_dev
= physical_device
->isl_dev
;
2780 /* On Broadwell and later, we can use batch chaining to more efficiently
2781 * implement growing command buffers. Prior to Haswell, the kernel
2782 * command parser gets in the way and we have to fall back to growing
2785 device
->can_chain_batches
= device
->info
.gen
>= 8;
2787 device
->robust_buffer_access
= pCreateInfo
->pEnabledFeatures
&&
2788 pCreateInfo
->pEnabledFeatures
->robustBufferAccess
;
2789 device
->enabled_extensions
= enabled_extensions
;
2791 anv_device_init_dispatch(device
);
2793 if (pthread_mutex_init(&device
->mutex
, NULL
) != 0) {
2794 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2798 pthread_condattr_t condattr
;
2799 if (pthread_condattr_init(&condattr
) != 0) {
2800 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2803 if (pthread_condattr_setclock(&condattr
, CLOCK_MONOTONIC
) != 0) {
2804 pthread_condattr_destroy(&condattr
);
2805 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2808 if (pthread_cond_init(&device
->queue_submit
, &condattr
) != 0) {
2809 pthread_condattr_destroy(&condattr
);
2810 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2813 pthread_condattr_destroy(&condattr
);
2815 result
= anv_bo_cache_init(&device
->bo_cache
);
2816 if (result
!= VK_SUCCESS
)
2817 goto fail_queue_cond
;
2819 anv_bo_pool_init(&device
->batch_bo_pool
, device
);
2821 result
= anv_state_pool_init(&device
->dynamic_state_pool
, device
,
2822 DYNAMIC_STATE_POOL_MIN_ADDRESS
, 16384);
2823 if (result
!= VK_SUCCESS
)
2824 goto fail_batch_bo_pool
;
2826 result
= anv_state_pool_init(&device
->instruction_state_pool
, device
,
2827 INSTRUCTION_STATE_POOL_MIN_ADDRESS
, 16384);
2828 if (result
!= VK_SUCCESS
)
2829 goto fail_dynamic_state_pool
;
2831 result
= anv_state_pool_init(&device
->surface_state_pool
, device
,
2832 SURFACE_STATE_POOL_MIN_ADDRESS
, 4096);
2833 if (result
!= VK_SUCCESS
)
2834 goto fail_instruction_state_pool
;
2836 if (physical_device
->use_softpin
) {
2837 result
= anv_state_pool_init(&device
->binding_table_pool
, device
,
2838 BINDING_TABLE_POOL_MIN_ADDRESS
, 4096);
2839 if (result
!= VK_SUCCESS
)
2840 goto fail_surface_state_pool
;
2843 if (device
->info
.gen
>= 12) {
2844 device
->aux_map_ctx
= gen_aux_map_init(device
, &aux_map_allocator
,
2845 &physical_device
->info
);
2846 if (!device
->aux_map_ctx
)
2847 goto fail_binding_table_pool
;
2850 result
= anv_device_alloc_bo(device
, 4096, 0 /* flags */,
2851 0 /* explicit_address */,
2852 &device
->workaround_bo
);
2853 if (result
!= VK_SUCCESS
)
2854 goto fail_surface_aux_map_pool
;
2856 result
= anv_device_init_trivial_batch(device
);
2857 if (result
!= VK_SUCCESS
)
2858 goto fail_workaround_bo
;
2860 if (device
->info
.gen
>= 10) {
2861 result
= anv_device_init_hiz_clear_value_bo(device
);
2862 if (result
!= VK_SUCCESS
)
2863 goto fail_trivial_batch_bo
;
2866 anv_scratch_pool_init(device
, &device
->scratch_pool
);
2868 switch (device
->info
.gen
) {
2870 if (!device
->info
.is_haswell
)
2871 result
= gen7_init_device_state(device
);
2873 result
= gen75_init_device_state(device
);
2876 result
= gen8_init_device_state(device
);
2879 result
= gen9_init_device_state(device
);
2882 result
= gen10_init_device_state(device
);
2885 result
= gen11_init_device_state(device
);
2888 result
= gen12_init_device_state(device
);
2891 /* Shouldn't get here as we don't create physical devices for any other
2893 unreachable("unhandled gen");
2895 if (result
!= VK_SUCCESS
)
2896 goto fail_workaround_bo
;
2898 anv_pipeline_cache_init(&device
->default_pipeline_cache
, device
, true);
2900 anv_device_init_blorp(device
);
2902 anv_device_init_border_colors(device
);
2904 anv_device_perf_init(device
);
2906 *pDevice
= anv_device_to_handle(device
);
2911 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
2912 if (device
->info
.gen
>= 10)
2913 anv_device_release_bo(device
, device
->hiz_clear_bo
);
2914 anv_device_release_bo(device
, device
->workaround_bo
);
2915 fail_trivial_batch_bo
:
2916 anv_device_release_bo(device
, device
->trivial_batch_bo
);
2917 fail_surface_aux_map_pool
:
2918 if (device
->info
.gen
>= 12) {
2919 gen_aux_map_finish(device
->aux_map_ctx
);
2920 device
->aux_map_ctx
= NULL
;
2922 fail_binding_table_pool
:
2923 if (physical_device
->use_softpin
)
2924 anv_state_pool_finish(&device
->binding_table_pool
);
2925 fail_surface_state_pool
:
2926 anv_state_pool_finish(&device
->surface_state_pool
);
2927 fail_instruction_state_pool
:
2928 anv_state_pool_finish(&device
->instruction_state_pool
);
2929 fail_dynamic_state_pool
:
2930 anv_state_pool_finish(&device
->dynamic_state_pool
);
2932 anv_bo_pool_finish(&device
->batch_bo_pool
);
2933 anv_bo_cache_finish(&device
->bo_cache
);
2935 pthread_cond_destroy(&device
->queue_submit
);
2937 pthread_mutex_destroy(&device
->mutex
);
2939 if (physical_device
->use_softpin
) {
2940 util_vma_heap_finish(&device
->vma_hi
);
2941 util_vma_heap_finish(&device
->vma_cva
);
2942 util_vma_heap_finish(&device
->vma_lo
);
2945 anv_queue_finish(&device
->queue
);
2947 anv_gem_destroy_context(device
, device
->context_id
);
2951 vk_free(&device
->alloc
, device
);
2956 void anv_DestroyDevice(
2958 const VkAllocationCallbacks
* pAllocator
)
2960 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2965 anv_device_finish_blorp(device
);
2967 anv_pipeline_cache_finish(&device
->default_pipeline_cache
);
2969 anv_queue_finish(&device
->queue
);
2971 #ifdef HAVE_VALGRIND
2972 /* We only need to free these to prevent valgrind errors. The backing
2973 * BO will go away in a couple of lines so we don't actually leak.
2975 anv_state_pool_free(&device
->dynamic_state_pool
, device
->border_colors
);
2976 anv_state_pool_free(&device
->dynamic_state_pool
, device
->slice_hash
);
2979 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
2981 anv_device_release_bo(device
, device
->workaround_bo
);
2982 anv_device_release_bo(device
, device
->trivial_batch_bo
);
2983 if (device
->info
.gen
>= 10)
2984 anv_device_release_bo(device
, device
->hiz_clear_bo
);
2986 if (device
->info
.gen
>= 12) {
2987 gen_aux_map_finish(device
->aux_map_ctx
);
2988 device
->aux_map_ctx
= NULL
;
2991 if (device
->physical
->use_softpin
)
2992 anv_state_pool_finish(&device
->binding_table_pool
);
2993 anv_state_pool_finish(&device
->surface_state_pool
);
2994 anv_state_pool_finish(&device
->instruction_state_pool
);
2995 anv_state_pool_finish(&device
->dynamic_state_pool
);
2997 anv_bo_pool_finish(&device
->batch_bo_pool
);
2999 anv_bo_cache_finish(&device
->bo_cache
);
3001 if (device
->physical
->use_softpin
) {
3002 util_vma_heap_finish(&device
->vma_hi
);
3003 util_vma_heap_finish(&device
->vma_cva
);
3004 util_vma_heap_finish(&device
->vma_lo
);
3007 pthread_cond_destroy(&device
->queue_submit
);
3008 pthread_mutex_destroy(&device
->mutex
);
3010 anv_gem_destroy_context(device
, device
->context_id
);
3012 if (INTEL_DEBUG
& DEBUG_BATCH
)
3013 gen_batch_decode_ctx_finish(&device
->decoder_ctx
);
3017 vk_free(&device
->alloc
, device
);
3020 VkResult
anv_EnumerateInstanceLayerProperties(
3021 uint32_t* pPropertyCount
,
3022 VkLayerProperties
* pProperties
)
3024 if (pProperties
== NULL
) {
3025 *pPropertyCount
= 0;
3029 /* None supported at this time */
3030 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
3033 VkResult
anv_EnumerateDeviceLayerProperties(
3034 VkPhysicalDevice physicalDevice
,
3035 uint32_t* pPropertyCount
,
3036 VkLayerProperties
* pProperties
)
3038 if (pProperties
== NULL
) {
3039 *pPropertyCount
= 0;
3043 /* None supported at this time */
3044 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
3047 void anv_GetDeviceQueue(
3049 uint32_t queueNodeIndex
,
3050 uint32_t queueIndex
,
3053 const VkDeviceQueueInfo2 info
= {
3054 .sType
= VK_STRUCTURE_TYPE_DEVICE_QUEUE_INFO_2
,
3057 .queueFamilyIndex
= queueNodeIndex
,
3058 .queueIndex
= queueIndex
,
3061 anv_GetDeviceQueue2(_device
, &info
, pQueue
);
3064 void anv_GetDeviceQueue2(
3066 const VkDeviceQueueInfo2
* pQueueInfo
,
3069 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3071 assert(pQueueInfo
->queueIndex
== 0);
3073 if (pQueueInfo
->flags
== device
->queue
.flags
)
3074 *pQueue
= anv_queue_to_handle(&device
->queue
);
3080 _anv_device_set_lost(struct anv_device
*device
,
3081 const char *file
, int line
,
3082 const char *msg
, ...)
3087 p_atomic_inc(&device
->_lost
);
3090 err
= __vk_errorv(device
->physical
->instance
, device
,
3091 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT
,
3092 VK_ERROR_DEVICE_LOST
, file
, line
, msg
, ap
);
3095 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
3102 _anv_queue_set_lost(struct anv_queue
*queue
,
3103 const char *file
, int line
,
3104 const char *msg
, ...)
3109 p_atomic_inc(&queue
->device
->_lost
);
3112 err
= __vk_errorv(queue
->device
->physical
->instance
, queue
->device
,
3113 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT
,
3114 VK_ERROR_DEVICE_LOST
, file
, line
, msg
, ap
);
3117 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
3124 anv_device_query_status(struct anv_device
*device
)
3126 /* This isn't likely as most of the callers of this function already check
3127 * for it. However, it doesn't hurt to check and it potentially lets us
3130 if (anv_device_is_lost(device
))
3131 return VK_ERROR_DEVICE_LOST
;
3133 uint32_t active
, pending
;
3134 int ret
= anv_gem_gpu_get_reset_stats(device
, &active
, &pending
);
3136 /* We don't know the real error. */
3137 return anv_device_set_lost(device
, "get_reset_stats failed: %m");
3141 return anv_device_set_lost(device
, "GPU hung on one of our command buffers");
3142 } else if (pending
) {
3143 return anv_device_set_lost(device
, "GPU hung with commands in-flight");
3150 anv_device_bo_busy(struct anv_device
*device
, struct anv_bo
*bo
)
3152 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
3153 * Other usages of the BO (such as on different hardware) will not be
3154 * flagged as "busy" by this ioctl. Use with care.
3156 int ret
= anv_gem_busy(device
, bo
->gem_handle
);
3158 return VK_NOT_READY
;
3159 } else if (ret
== -1) {
3160 /* We don't know the real error. */
3161 return anv_device_set_lost(device
, "gem wait failed: %m");
3164 /* Query for device status after the busy call. If the BO we're checking
3165 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
3166 * client because it clearly doesn't have valid data. Yes, this most
3167 * likely means an ioctl, but we just did an ioctl to query the busy status
3168 * so it's no great loss.
3170 return anv_device_query_status(device
);
3174 anv_device_wait(struct anv_device
*device
, struct anv_bo
*bo
,
3177 int ret
= anv_gem_wait(device
, bo
->gem_handle
, &timeout
);
3178 if (ret
== -1 && errno
== ETIME
) {
3180 } else if (ret
== -1) {
3181 /* We don't know the real error. */
3182 return anv_device_set_lost(device
, "gem wait failed: %m");
3185 /* Query for device status after the wait. If the BO we're waiting on got
3186 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
3187 * because it clearly doesn't have valid data. Yes, this most likely means
3188 * an ioctl, but we just did an ioctl to wait so it's no great loss.
3190 return anv_device_query_status(device
);
3193 VkResult
anv_DeviceWaitIdle(
3196 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3198 if (anv_device_is_lost(device
))
3199 return VK_ERROR_DEVICE_LOST
;
3201 return anv_queue_submit_simple_batch(&device
->queue
, NULL
);
3205 anv_vma_alloc(struct anv_device
*device
, struct anv_bo
*bo
,
3206 uint64_t client_address
)
3208 const struct gen_device_info
*devinfo
= &device
->info
;
3209 /* Gen12 CCS surface addresses need to be 64K aligned. We have no way of
3210 * telling what this allocation is for so pick the largest alignment.
3212 const uint32_t vma_alignment
=
3213 devinfo
->gen
>= 12 ? (64 * 1024) : (4 * 1024);
3215 if (!(bo
->flags
& EXEC_OBJECT_PINNED
)) {
3216 assert(!(bo
->has_client_visible_address
));
3220 pthread_mutex_lock(&device
->vma_mutex
);
3224 if (bo
->has_client_visible_address
) {
3225 assert(bo
->flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
);
3226 if (client_address
) {
3227 if (util_vma_heap_alloc_addr(&device
->vma_cva
,
3228 client_address
, bo
->size
)) {
3229 bo
->offset
= gen_canonical_address(client_address
);
3233 util_vma_heap_alloc(&device
->vma_cva
, bo
->size
, vma_alignment
);
3235 bo
->offset
= gen_canonical_address(addr
);
3236 assert(addr
== gen_48b_address(bo
->offset
));
3239 /* We don't want to fall back to other heaps */
3243 assert(client_address
== 0);
3245 if (bo
->flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
) {
3247 util_vma_heap_alloc(&device
->vma_hi
, bo
->size
, vma_alignment
);
3249 bo
->offset
= gen_canonical_address(addr
);
3250 assert(addr
== gen_48b_address(bo
->offset
));
3254 if (bo
->offset
== 0) {
3256 util_vma_heap_alloc(&device
->vma_lo
, bo
->size
, vma_alignment
);
3258 bo
->offset
= gen_canonical_address(addr
);
3259 assert(addr
== gen_48b_address(bo
->offset
));
3264 pthread_mutex_unlock(&device
->vma_mutex
);
3266 return bo
->offset
!= 0;
3270 anv_vma_free(struct anv_device
*device
, struct anv_bo
*bo
)
3272 if (!(bo
->flags
& EXEC_OBJECT_PINNED
))
3275 const uint64_t addr_48b
= gen_48b_address(bo
->offset
);
3277 pthread_mutex_lock(&device
->vma_mutex
);
3279 if (addr_48b
>= LOW_HEAP_MIN_ADDRESS
&&
3280 addr_48b
<= LOW_HEAP_MAX_ADDRESS
) {
3281 util_vma_heap_free(&device
->vma_lo
, addr_48b
, bo
->size
);
3282 } else if (addr_48b
>= CLIENT_VISIBLE_HEAP_MIN_ADDRESS
&&
3283 addr_48b
<= CLIENT_VISIBLE_HEAP_MAX_ADDRESS
) {
3284 util_vma_heap_free(&device
->vma_cva
, addr_48b
, bo
->size
);
3286 assert(addr_48b
>= HIGH_HEAP_MIN_ADDRESS
);
3287 util_vma_heap_free(&device
->vma_hi
, addr_48b
, bo
->size
);
3290 pthread_mutex_unlock(&device
->vma_mutex
);
3295 VkResult
anv_AllocateMemory(
3297 const VkMemoryAllocateInfo
* pAllocateInfo
,
3298 const VkAllocationCallbacks
* pAllocator
,
3299 VkDeviceMemory
* pMem
)
3301 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3302 struct anv_physical_device
*pdevice
= device
->physical
;
3303 struct anv_device_memory
*mem
;
3304 VkResult result
= VK_SUCCESS
;
3306 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
3308 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
3309 assert(pAllocateInfo
->allocationSize
> 0);
3311 VkDeviceSize aligned_alloc_size
=
3312 align_u64(pAllocateInfo
->allocationSize
, 4096);
3314 if (aligned_alloc_size
> MAX_MEMORY_ALLOCATION_SIZE
)
3315 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
3317 assert(pAllocateInfo
->memoryTypeIndex
< pdevice
->memory
.type_count
);
3318 struct anv_memory_type
*mem_type
=
3319 &pdevice
->memory
.types
[pAllocateInfo
->memoryTypeIndex
];
3320 assert(mem_type
->heapIndex
< pdevice
->memory
.heap_count
);
3321 struct anv_memory_heap
*mem_heap
=
3322 &pdevice
->memory
.heaps
[mem_type
->heapIndex
];
3324 uint64_t mem_heap_used
= p_atomic_read(&mem_heap
->used
);
3325 if (mem_heap_used
+ aligned_alloc_size
> mem_heap
->size
)
3326 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
3328 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
3329 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3331 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3333 assert(pAllocateInfo
->memoryTypeIndex
< pdevice
->memory
.type_count
);
3334 mem
->type
= mem_type
;
3338 mem
->host_ptr
= NULL
;
3340 enum anv_bo_alloc_flags alloc_flags
= 0;
3342 const VkExportMemoryAllocateInfo
*export_info
= NULL
;
3343 const VkImportAndroidHardwareBufferInfoANDROID
*ahw_import_info
= NULL
;
3344 const VkImportMemoryFdInfoKHR
*fd_info
= NULL
;
3345 const VkImportMemoryHostPointerInfoEXT
*host_ptr_info
= NULL
;
3346 const VkMemoryDedicatedAllocateInfo
*dedicated_info
= NULL
;
3347 VkMemoryAllocateFlags vk_flags
= 0;
3348 uint64_t client_address
= 0;
3350 vk_foreach_struct_const(ext
, pAllocateInfo
->pNext
) {
3351 switch (ext
->sType
) {
3352 case VK_STRUCTURE_TYPE_EXPORT_MEMORY_ALLOCATE_INFO
:
3353 export_info
= (void *)ext
;
3356 case VK_STRUCTURE_TYPE_IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID
:
3357 ahw_import_info
= (void *)ext
;
3360 case VK_STRUCTURE_TYPE_IMPORT_MEMORY_FD_INFO_KHR
:
3361 fd_info
= (void *)ext
;
3364 case VK_STRUCTURE_TYPE_IMPORT_MEMORY_HOST_POINTER_INFO_EXT
:
3365 host_ptr_info
= (void *)ext
;
3368 case VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_FLAGS_INFO
: {
3369 const VkMemoryAllocateFlagsInfo
*flags_info
= (void *)ext
;
3370 vk_flags
= flags_info
->flags
;
3374 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_ALLOCATE_INFO
:
3375 dedicated_info
= (void *)ext
;
3378 case VK_STRUCTURE_TYPE_MEMORY_OPAQUE_CAPTURE_ADDRESS_ALLOCATE_INFO_KHR
: {
3379 const VkMemoryOpaqueCaptureAddressAllocateInfoKHR
*addr_info
=
3380 (const VkMemoryOpaqueCaptureAddressAllocateInfoKHR
*)ext
;
3381 client_address
= addr_info
->opaqueCaptureAddress
;
3386 anv_debug_ignored_stype(ext
->sType
);
3391 if (vk_flags
& VK_MEMORY_ALLOCATE_DEVICE_ADDRESS_BIT_KHR
)
3392 alloc_flags
|= ANV_BO_ALLOC_CLIENT_VISIBLE_ADDRESS
;
3394 /* Check if we need to support Android HW buffer export. If so,
3395 * create AHardwareBuffer and import memory from it.
3397 bool android_export
= false;
3398 if (export_info
&& export_info
->handleTypes
&
3399 VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID
)
3400 android_export
= true;
3402 if (ahw_import_info
) {
3403 result
= anv_import_ahw_memory(_device
, mem
, ahw_import_info
);
3404 if (result
!= VK_SUCCESS
)
3408 } else if (android_export
) {
3409 result
= anv_create_ahw_memory(_device
, mem
, pAllocateInfo
);
3410 if (result
!= VK_SUCCESS
)
3413 const VkImportAndroidHardwareBufferInfoANDROID import_info
= {
3416 result
= anv_import_ahw_memory(_device
, mem
, &import_info
);
3417 if (result
!= VK_SUCCESS
)
3423 /* The Vulkan spec permits handleType to be 0, in which case the struct is
3426 if (fd_info
&& fd_info
->handleType
) {
3427 /* At the moment, we support only the below handle types. */
3428 assert(fd_info
->handleType
==
3429 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
3430 fd_info
->handleType
==
3431 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
3433 result
= anv_device_import_bo(device
, fd_info
->fd
, alloc_flags
,
3434 client_address
, &mem
->bo
);
3435 if (result
!= VK_SUCCESS
)
3438 VkDeviceSize aligned_alloc_size
=
3439 align_u64(pAllocateInfo
->allocationSize
, 4096);
3441 /* For security purposes, we reject importing the bo if it's smaller
3442 * than the requested allocation size. This prevents a malicious client
3443 * from passing a buffer to a trusted client, lying about the size, and
3444 * telling the trusted client to try and texture from an image that goes
3445 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
3446 * in the trusted client. The trusted client can protect itself against
3447 * this sort of attack but only if it can trust the buffer size.
3449 if (mem
->bo
->size
< aligned_alloc_size
) {
3450 result
= vk_errorf(device
, device
, VK_ERROR_INVALID_EXTERNAL_HANDLE
,
3451 "aligned allocationSize too large for "
3452 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT: "
3453 "%"PRIu64
"B > %"PRIu64
"B",
3454 aligned_alloc_size
, mem
->bo
->size
);
3455 anv_device_release_bo(device
, mem
->bo
);
3459 /* From the Vulkan spec:
3461 * "Importing memory from a file descriptor transfers ownership of
3462 * the file descriptor from the application to the Vulkan
3463 * implementation. The application must not perform any operations on
3464 * the file descriptor after a successful import."
3466 * If the import fails, we leave the file descriptor open.
3472 if (host_ptr_info
&& host_ptr_info
->handleType
) {
3473 if (host_ptr_info
->handleType
==
3474 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_MAPPED_FOREIGN_MEMORY_BIT_EXT
) {
3475 result
= vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
3479 assert(host_ptr_info
->handleType
==
3480 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
);
3482 result
= anv_device_import_bo_from_host_ptr(device
,
3483 host_ptr_info
->pHostPointer
,
3484 pAllocateInfo
->allocationSize
,
3488 if (result
!= VK_SUCCESS
)
3491 mem
->host_ptr
= host_ptr_info
->pHostPointer
;
3495 /* Regular allocate (not importing memory). */
3497 if (export_info
&& export_info
->handleTypes
)
3498 alloc_flags
|= ANV_BO_ALLOC_EXTERNAL
;
3500 result
= anv_device_alloc_bo(device
, pAllocateInfo
->allocationSize
,
3501 alloc_flags
, client_address
, &mem
->bo
);
3502 if (result
!= VK_SUCCESS
)
3505 if (dedicated_info
&& dedicated_info
->image
!= VK_NULL_HANDLE
) {
3506 ANV_FROM_HANDLE(anv_image
, image
, dedicated_info
->image
);
3508 /* Some legacy (non-modifiers) consumers need the tiling to be set on
3509 * the BO. In this case, we have a dedicated allocation.
3511 if (image
->needs_set_tiling
) {
3512 const uint32_t i915_tiling
=
3513 isl_tiling_to_i915_tiling(image
->planes
[0].surface
.isl
.tiling
);
3514 int ret
= anv_gem_set_tiling(device
, mem
->bo
->gem_handle
,
3515 image
->planes
[0].surface
.isl
.row_pitch_B
,
3518 anv_device_release_bo(device
, mem
->bo
);
3519 result
= vk_errorf(device
, device
, VK_ERROR_OUT_OF_DEVICE_MEMORY
,
3520 "failed to set BO tiling: %m");
3527 mem_heap_used
= p_atomic_add_return(&mem_heap
->used
, mem
->bo
->size
);
3528 if (mem_heap_used
> mem_heap
->size
) {
3529 p_atomic_add(&mem_heap
->used
, -mem
->bo
->size
);
3530 anv_device_release_bo(device
, mem
->bo
);
3531 result
= vk_errorf(device
, device
, VK_ERROR_OUT_OF_DEVICE_MEMORY
,
3532 "Out of heap memory");
3536 pthread_mutex_lock(&device
->mutex
);
3537 list_addtail(&mem
->link
, &device
->memory_objects
);
3538 pthread_mutex_unlock(&device
->mutex
);
3540 *pMem
= anv_device_memory_to_handle(mem
);
3545 vk_free2(&device
->alloc
, pAllocator
, mem
);
3550 VkResult
anv_GetMemoryFdKHR(
3552 const VkMemoryGetFdInfoKHR
* pGetFdInfo
,
3555 ANV_FROM_HANDLE(anv_device
, dev
, device_h
);
3556 ANV_FROM_HANDLE(anv_device_memory
, mem
, pGetFdInfo
->memory
);
3558 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
3560 assert(pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
3561 pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
3563 return anv_device_export_bo(dev
, mem
->bo
, pFd
);
3566 VkResult
anv_GetMemoryFdPropertiesKHR(
3568 VkExternalMemoryHandleTypeFlagBits handleType
,
3570 VkMemoryFdPropertiesKHR
* pMemoryFdProperties
)
3572 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3574 switch (handleType
) {
3575 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
:
3576 /* dma-buf can be imported as any memory type */
3577 pMemoryFdProperties
->memoryTypeBits
=
3578 (1 << device
->physical
->memory
.type_count
) - 1;
3582 /* The valid usage section for this function says:
3584 * "handleType must not be one of the handle types defined as
3587 * So opaque handle types fall into the default "unsupported" case.
3589 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
3593 VkResult
anv_GetMemoryHostPointerPropertiesEXT(
3595 VkExternalMemoryHandleTypeFlagBits handleType
,
3596 const void* pHostPointer
,
3597 VkMemoryHostPointerPropertiesEXT
* pMemoryHostPointerProperties
)
3599 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3601 assert(pMemoryHostPointerProperties
->sType
==
3602 VK_STRUCTURE_TYPE_MEMORY_HOST_POINTER_PROPERTIES_EXT
);
3604 switch (handleType
) {
3605 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
:
3606 /* Host memory can be imported as any memory type. */
3607 pMemoryHostPointerProperties
->memoryTypeBits
=
3608 (1ull << device
->physical
->memory
.type_count
) - 1;
3613 return VK_ERROR_INVALID_EXTERNAL_HANDLE
;
3617 void anv_FreeMemory(
3619 VkDeviceMemory _mem
,
3620 const VkAllocationCallbacks
* pAllocator
)
3622 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3623 ANV_FROM_HANDLE(anv_device_memory
, mem
, _mem
);
3628 pthread_mutex_lock(&device
->mutex
);
3629 list_del(&mem
->link
);
3630 pthread_mutex_unlock(&device
->mutex
);
3633 anv_UnmapMemory(_device
, _mem
);
3635 p_atomic_add(&device
->physical
->memory
.heaps
[mem
->type
->heapIndex
].used
,
3638 anv_device_release_bo(device
, mem
->bo
);
3640 #if defined(ANDROID) && ANDROID_API_LEVEL >= 26
3642 AHardwareBuffer_release(mem
->ahw
);
3645 vk_free2(&device
->alloc
, pAllocator
, mem
);
3648 VkResult
anv_MapMemory(
3650 VkDeviceMemory _memory
,
3651 VkDeviceSize offset
,
3653 VkMemoryMapFlags flags
,
3656 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3657 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
3664 if (mem
->host_ptr
) {
3665 *ppData
= mem
->host_ptr
+ offset
;
3669 if (size
== VK_WHOLE_SIZE
)
3670 size
= mem
->bo
->size
- offset
;
3672 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
3674 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
3675 * assert(size != 0);
3676 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
3677 * equal to the size of the memory minus offset
3680 assert(offset
+ size
<= mem
->bo
->size
);
3682 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
3683 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
3684 * at a time is valid. We could just mmap up front and return an offset
3685 * pointer here, but that may exhaust virtual memory on 32 bit
3688 uint32_t gem_flags
= 0;
3690 if (!device
->info
.has_llc
&&
3691 (mem
->type
->propertyFlags
& VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
))
3692 gem_flags
|= I915_MMAP_WC
;
3694 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
3695 uint64_t map_offset
= offset
& ~4095ull;
3696 assert(offset
>= map_offset
);
3697 uint64_t map_size
= (offset
+ size
) - map_offset
;
3699 /* Let's map whole pages */
3700 map_size
= align_u64(map_size
, 4096);
3702 void *map
= anv_gem_mmap(device
, mem
->bo
->gem_handle
,
3703 map_offset
, map_size
, gem_flags
);
3704 if (map
== MAP_FAILED
)
3705 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
3708 mem
->map_size
= map_size
;
3710 *ppData
= mem
->map
+ (offset
- map_offset
);
3715 void anv_UnmapMemory(
3717 VkDeviceMemory _memory
)
3719 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
3721 if (mem
== NULL
|| mem
->host_ptr
)
3724 anv_gem_munmap(mem
->map
, mem
->map_size
);
3731 clflush_mapped_ranges(struct anv_device
*device
,
3733 const VkMappedMemoryRange
*ranges
)
3735 for (uint32_t i
= 0; i
< count
; i
++) {
3736 ANV_FROM_HANDLE(anv_device_memory
, mem
, ranges
[i
].memory
);
3737 if (ranges
[i
].offset
>= mem
->map_size
)
3740 gen_clflush_range(mem
->map
+ ranges
[i
].offset
,
3741 MIN2(ranges
[i
].size
, mem
->map_size
- ranges
[i
].offset
));
3745 VkResult
anv_FlushMappedMemoryRanges(
3747 uint32_t memoryRangeCount
,
3748 const VkMappedMemoryRange
* pMemoryRanges
)
3750 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3752 if (device
->info
.has_llc
)
3755 /* Make sure the writes we're flushing have landed. */
3756 __builtin_ia32_mfence();
3758 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
3763 VkResult
anv_InvalidateMappedMemoryRanges(
3765 uint32_t memoryRangeCount
,
3766 const VkMappedMemoryRange
* pMemoryRanges
)
3768 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3770 if (device
->info
.has_llc
)
3773 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
3775 /* Make sure no reads get moved up above the invalidate. */
3776 __builtin_ia32_mfence();
3781 void anv_GetBufferMemoryRequirements(
3784 VkMemoryRequirements
* pMemoryRequirements
)
3786 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
3787 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3789 /* The Vulkan spec (git aaed022) says:
3791 * memoryTypeBits is a bitfield and contains one bit set for every
3792 * supported memory type for the resource. The bit `1<<i` is set if and
3793 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3794 * structure for the physical device is supported.
3796 uint32_t memory_types
= (1ull << device
->physical
->memory
.type_count
) - 1;
3798 /* Base alignment requirement of a cache line */
3799 uint32_t alignment
= 16;
3801 /* We need an alignment of 32 for pushing UBOs */
3802 if (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
)
3803 alignment
= MAX2(alignment
, 32);
3805 pMemoryRequirements
->size
= buffer
->size
;
3806 pMemoryRequirements
->alignment
= alignment
;
3808 /* Storage and Uniform buffers should have their size aligned to
3809 * 32-bits to avoid boundary checks when last DWord is not complete.
3810 * This would ensure that not internal padding would be needed for
3813 if (device
->robust_buffer_access
&&
3814 (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
||
3815 buffer
->usage
& VK_BUFFER_USAGE_STORAGE_BUFFER_BIT
))
3816 pMemoryRequirements
->size
= align_u64(buffer
->size
, 4);
3818 pMemoryRequirements
->memoryTypeBits
= memory_types
;
3821 void anv_GetBufferMemoryRequirements2(
3823 const VkBufferMemoryRequirementsInfo2
* pInfo
,
3824 VkMemoryRequirements2
* pMemoryRequirements
)
3826 anv_GetBufferMemoryRequirements(_device
, pInfo
->buffer
,
3827 &pMemoryRequirements
->memoryRequirements
);
3829 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3830 switch (ext
->sType
) {
3831 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
3832 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
3833 requirements
->prefersDedicatedAllocation
= false;
3834 requirements
->requiresDedicatedAllocation
= false;
3839 anv_debug_ignored_stype(ext
->sType
);
3845 void anv_GetImageMemoryRequirements(
3848 VkMemoryRequirements
* pMemoryRequirements
)
3850 ANV_FROM_HANDLE(anv_image
, image
, _image
);
3851 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3853 /* The Vulkan spec (git aaed022) says:
3855 * memoryTypeBits is a bitfield and contains one bit set for every
3856 * supported memory type for the resource. The bit `1<<i` is set if and
3857 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3858 * structure for the physical device is supported.
3860 * All types are currently supported for images.
3862 uint32_t memory_types
= (1ull << device
->physical
->memory
.type_count
) - 1;
3864 /* We must have image allocated or imported at this point. According to the
3865 * specification, external images must have been bound to memory before
3866 * calling GetImageMemoryRequirements.
3868 assert(image
->size
> 0);
3870 pMemoryRequirements
->size
= image
->size
;
3871 pMemoryRequirements
->alignment
= image
->alignment
;
3872 pMemoryRequirements
->memoryTypeBits
= memory_types
;
3875 void anv_GetImageMemoryRequirements2(
3877 const VkImageMemoryRequirementsInfo2
* pInfo
,
3878 VkMemoryRequirements2
* pMemoryRequirements
)
3880 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3881 ANV_FROM_HANDLE(anv_image
, image
, pInfo
->image
);
3883 anv_GetImageMemoryRequirements(_device
, pInfo
->image
,
3884 &pMemoryRequirements
->memoryRequirements
);
3886 vk_foreach_struct_const(ext
, pInfo
->pNext
) {
3887 switch (ext
->sType
) {
3888 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO
: {
3889 const VkImagePlaneMemoryRequirementsInfo
*plane_reqs
=
3890 (const VkImagePlaneMemoryRequirementsInfo
*) ext
;
3891 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
,
3892 plane_reqs
->planeAspect
);
3894 assert(image
->planes
[plane
].offset
== 0);
3896 /* The Vulkan spec (git aaed022) says:
3898 * memoryTypeBits is a bitfield and contains one bit set for every
3899 * supported memory type for the resource. The bit `1<<i` is set
3900 * if and only if the memory type `i` in the
3901 * VkPhysicalDeviceMemoryProperties structure for the physical
3902 * device is supported.
3904 * All types are currently supported for images.
3906 pMemoryRequirements
->memoryRequirements
.memoryTypeBits
=
3907 (1ull << device
->physical
->memory
.type_count
) - 1;
3909 /* We must have image allocated or imported at this point. According to the
3910 * specification, external images must have been bound to memory before
3911 * calling GetImageMemoryRequirements.
3913 assert(image
->planes
[plane
].size
> 0);
3915 pMemoryRequirements
->memoryRequirements
.size
= image
->planes
[plane
].size
;
3916 pMemoryRequirements
->memoryRequirements
.alignment
=
3917 image
->planes
[plane
].alignment
;
3922 anv_debug_ignored_stype(ext
->sType
);
3927 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3928 switch (ext
->sType
) {
3929 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
3930 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
3931 if (image
->needs_set_tiling
|| image
->external_format
) {
3932 /* If we need to set the tiling for external consumers, we need a
3933 * dedicated allocation.
3935 * See also anv_AllocateMemory.
3937 requirements
->prefersDedicatedAllocation
= true;
3938 requirements
->requiresDedicatedAllocation
= true;
3940 requirements
->prefersDedicatedAllocation
= false;
3941 requirements
->requiresDedicatedAllocation
= false;
3947 anv_debug_ignored_stype(ext
->sType
);
3953 void anv_GetImageSparseMemoryRequirements(
3956 uint32_t* pSparseMemoryRequirementCount
,
3957 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
3959 *pSparseMemoryRequirementCount
= 0;
3962 void anv_GetImageSparseMemoryRequirements2(
3964 const VkImageSparseMemoryRequirementsInfo2
* pInfo
,
3965 uint32_t* pSparseMemoryRequirementCount
,
3966 VkSparseImageMemoryRequirements2
* pSparseMemoryRequirements
)
3968 *pSparseMemoryRequirementCount
= 0;
3971 void anv_GetDeviceMemoryCommitment(
3973 VkDeviceMemory memory
,
3974 VkDeviceSize
* pCommittedMemoryInBytes
)
3976 *pCommittedMemoryInBytes
= 0;
3980 anv_bind_buffer_memory(const VkBindBufferMemoryInfo
*pBindInfo
)
3982 ANV_FROM_HANDLE(anv_device_memory
, mem
, pBindInfo
->memory
);
3983 ANV_FROM_HANDLE(anv_buffer
, buffer
, pBindInfo
->buffer
);
3985 assert(pBindInfo
->sType
== VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
);
3988 buffer
->address
= (struct anv_address
) {
3990 .offset
= pBindInfo
->memoryOffset
,
3993 buffer
->address
= ANV_NULL_ADDRESS
;
3997 VkResult
anv_BindBufferMemory(
4000 VkDeviceMemory memory
,
4001 VkDeviceSize memoryOffset
)
4003 anv_bind_buffer_memory(
4004 &(VkBindBufferMemoryInfo
) {
4005 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
,
4008 .memoryOffset
= memoryOffset
,
4014 VkResult
anv_BindBufferMemory2(
4016 uint32_t bindInfoCount
,
4017 const VkBindBufferMemoryInfo
* pBindInfos
)
4019 for (uint32_t i
= 0; i
< bindInfoCount
; i
++)
4020 anv_bind_buffer_memory(&pBindInfos
[i
]);
4025 VkResult
anv_QueueBindSparse(
4027 uint32_t bindInfoCount
,
4028 const VkBindSparseInfo
* pBindInfo
,
4031 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
4032 if (anv_device_is_lost(queue
->device
))
4033 return VK_ERROR_DEVICE_LOST
;
4035 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
4040 VkResult
anv_CreateEvent(
4042 const VkEventCreateInfo
* pCreateInfo
,
4043 const VkAllocationCallbacks
* pAllocator
,
4046 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4047 struct anv_state state
;
4048 struct anv_event
*event
;
4050 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_EVENT_CREATE_INFO
);
4052 state
= anv_state_pool_alloc(&device
->dynamic_state_pool
,
4055 event
->state
= state
;
4056 event
->semaphore
= VK_EVENT_RESET
;
4058 if (!device
->info
.has_llc
) {
4059 /* Make sure the writes we're flushing have landed. */
4060 __builtin_ia32_mfence();
4061 __builtin_ia32_clflush(event
);
4064 *pEvent
= anv_event_to_handle(event
);
4069 void anv_DestroyEvent(
4072 const VkAllocationCallbacks
* pAllocator
)
4074 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4075 ANV_FROM_HANDLE(anv_event
, event
, _event
);
4080 anv_state_pool_free(&device
->dynamic_state_pool
, event
->state
);
4083 VkResult
anv_GetEventStatus(
4087 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4088 ANV_FROM_HANDLE(anv_event
, event
, _event
);
4090 if (anv_device_is_lost(device
))
4091 return VK_ERROR_DEVICE_LOST
;
4093 if (!device
->info
.has_llc
) {
4094 /* Invalidate read cache before reading event written by GPU. */
4095 __builtin_ia32_clflush(event
);
4096 __builtin_ia32_mfence();
4100 return event
->semaphore
;
4103 VkResult
anv_SetEvent(
4107 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4108 ANV_FROM_HANDLE(anv_event
, event
, _event
);
4110 event
->semaphore
= VK_EVENT_SET
;
4112 if (!device
->info
.has_llc
) {
4113 /* Make sure the writes we're flushing have landed. */
4114 __builtin_ia32_mfence();
4115 __builtin_ia32_clflush(event
);
4121 VkResult
anv_ResetEvent(
4125 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4126 ANV_FROM_HANDLE(anv_event
, event
, _event
);
4128 event
->semaphore
= VK_EVENT_RESET
;
4130 if (!device
->info
.has_llc
) {
4131 /* Make sure the writes we're flushing have landed. */
4132 __builtin_ia32_mfence();
4133 __builtin_ia32_clflush(event
);
4141 VkResult
anv_CreateBuffer(
4143 const VkBufferCreateInfo
* pCreateInfo
,
4144 const VkAllocationCallbacks
* pAllocator
,
4147 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4148 struct anv_buffer
*buffer
;
4150 /* Don't allow creating buffers bigger than our address space. The real
4151 * issue here is that we may align up the buffer size and we don't want
4152 * doing so to cause roll-over. However, no one has any business
4153 * allocating a buffer larger than our GTT size.
4155 if (pCreateInfo
->size
> device
->physical
->gtt_size
)
4156 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
4158 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
4160 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
4161 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
4163 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
4165 buffer
->size
= pCreateInfo
->size
;
4166 buffer
->usage
= pCreateInfo
->usage
;
4167 buffer
->address
= ANV_NULL_ADDRESS
;
4169 *pBuffer
= anv_buffer_to_handle(buffer
);
4174 void anv_DestroyBuffer(
4177 const VkAllocationCallbacks
* pAllocator
)
4179 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4180 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
4185 vk_free2(&device
->alloc
, pAllocator
, buffer
);
4188 VkDeviceAddress
anv_GetBufferDeviceAddress(
4190 const VkBufferDeviceAddressInfoKHR
* pInfo
)
4192 ANV_FROM_HANDLE(anv_buffer
, buffer
, pInfo
->buffer
);
4194 assert(!anv_address_is_null(buffer
->address
));
4195 assert(buffer
->address
.bo
->flags
& EXEC_OBJECT_PINNED
);
4197 return anv_address_physical(buffer
->address
);
4200 uint64_t anv_GetBufferOpaqueCaptureAddress(
4202 const VkBufferDeviceAddressInfoKHR
* pInfo
)
4207 uint64_t anv_GetDeviceMemoryOpaqueCaptureAddress(
4209 const VkDeviceMemoryOpaqueCaptureAddressInfoKHR
* pInfo
)
4211 ANV_FROM_HANDLE(anv_device_memory
, memory
, pInfo
->memory
);
4213 assert(memory
->bo
->flags
& EXEC_OBJECT_PINNED
);
4214 assert(memory
->bo
->has_client_visible_address
);
4216 return gen_48b_address(memory
->bo
->offset
);
4220 anv_fill_buffer_surface_state(struct anv_device
*device
, struct anv_state state
,
4221 enum isl_format format
,
4222 struct anv_address address
,
4223 uint32_t range
, uint32_t stride
)
4225 isl_buffer_fill_state(&device
->isl_dev
, state
.map
,
4226 .address
= anv_address_physical(address
),
4227 .mocs
= device
->isl_dev
.mocs
.internal
,
4230 .swizzle
= ISL_SWIZZLE_IDENTITY
,
4231 .stride_B
= stride
);
4234 void anv_DestroySampler(
4237 const VkAllocationCallbacks
* pAllocator
)
4239 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4240 ANV_FROM_HANDLE(anv_sampler
, sampler
, _sampler
);
4245 if (sampler
->bindless_state
.map
) {
4246 anv_state_pool_free(&device
->dynamic_state_pool
,
4247 sampler
->bindless_state
);
4250 vk_free2(&device
->alloc
, pAllocator
, sampler
);
4253 VkResult
anv_CreateFramebuffer(
4255 const VkFramebufferCreateInfo
* pCreateInfo
,
4256 const VkAllocationCallbacks
* pAllocator
,
4257 VkFramebuffer
* pFramebuffer
)
4259 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4260 struct anv_framebuffer
*framebuffer
;
4262 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
4264 size_t size
= sizeof(*framebuffer
);
4266 /* VK_KHR_imageless_framebuffer extension says:
4268 * If flags includes VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR,
4269 * parameter pAttachments is ignored.
4271 if (!(pCreateInfo
->flags
& VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR
)) {
4272 size
+= sizeof(struct anv_image_view
*) * pCreateInfo
->attachmentCount
;
4273 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
4274 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
4275 if (framebuffer
== NULL
)
4276 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
4278 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
4279 ANV_FROM_HANDLE(anv_image_view
, iview
, pCreateInfo
->pAttachments
[i
]);
4280 framebuffer
->attachments
[i
] = iview
;
4282 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
4284 assert(device
->enabled_extensions
.KHR_imageless_framebuffer
);
4285 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
4286 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
4287 if (framebuffer
== NULL
)
4288 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
4290 framebuffer
->attachment_count
= 0;
4293 framebuffer
->width
= pCreateInfo
->width
;
4294 framebuffer
->height
= pCreateInfo
->height
;
4295 framebuffer
->layers
= pCreateInfo
->layers
;
4297 *pFramebuffer
= anv_framebuffer_to_handle(framebuffer
);
4302 void anv_DestroyFramebuffer(
4305 const VkAllocationCallbacks
* pAllocator
)
4307 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4308 ANV_FROM_HANDLE(anv_framebuffer
, fb
, _fb
);
4313 vk_free2(&device
->alloc
, pAllocator
, fb
);
4316 static const VkTimeDomainEXT anv_time_domains
[] = {
4317 VK_TIME_DOMAIN_DEVICE_EXT
,
4318 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
,
4319 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
,
4322 VkResult
anv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
4323 VkPhysicalDevice physicalDevice
,
4324 uint32_t *pTimeDomainCount
,
4325 VkTimeDomainEXT
*pTimeDomains
)
4328 VK_OUTARRAY_MAKE(out
, pTimeDomains
, pTimeDomainCount
);
4330 for (d
= 0; d
< ARRAY_SIZE(anv_time_domains
); d
++) {
4331 vk_outarray_append(&out
, i
) {
4332 *i
= anv_time_domains
[d
];
4336 return vk_outarray_status(&out
);
4340 anv_clock_gettime(clockid_t clock_id
)
4342 struct timespec current
;
4345 ret
= clock_gettime(clock_id
, ¤t
);
4346 if (ret
< 0 && clock_id
== CLOCK_MONOTONIC_RAW
)
4347 ret
= clock_gettime(CLOCK_MONOTONIC
, ¤t
);
4351 return (uint64_t) current
.tv_sec
* 1000000000ULL + current
.tv_nsec
;
4354 #define TIMESTAMP 0x2358
4356 VkResult
anv_GetCalibratedTimestampsEXT(
4358 uint32_t timestampCount
,
4359 const VkCalibratedTimestampInfoEXT
*pTimestampInfos
,
4360 uint64_t *pTimestamps
,
4361 uint64_t *pMaxDeviation
)
4363 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4364 uint64_t timestamp_frequency
= device
->info
.timestamp_frequency
;
4367 uint64_t begin
, end
;
4368 uint64_t max_clock_period
= 0;
4370 begin
= anv_clock_gettime(CLOCK_MONOTONIC_RAW
);
4372 for (d
= 0; d
< timestampCount
; d
++) {
4373 switch (pTimestampInfos
[d
].timeDomain
) {
4374 case VK_TIME_DOMAIN_DEVICE_EXT
:
4375 ret
= anv_gem_reg_read(device
, TIMESTAMP
| 1,
4379 return anv_device_set_lost(device
, "Failed to read the TIMESTAMP "
4382 uint64_t device_period
= DIV_ROUND_UP(1000000000, timestamp_frequency
);
4383 max_clock_period
= MAX2(max_clock_period
, device_period
);
4385 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
:
4386 pTimestamps
[d
] = anv_clock_gettime(CLOCK_MONOTONIC
);
4387 max_clock_period
= MAX2(max_clock_period
, 1);
4390 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
:
4391 pTimestamps
[d
] = begin
;
4399 end
= anv_clock_gettime(CLOCK_MONOTONIC_RAW
);
4402 * The maximum deviation is the sum of the interval over which we
4403 * perform the sampling and the maximum period of any sampled
4404 * clock. That's because the maximum skew between any two sampled
4405 * clock edges is when the sampled clock with the largest period is
4406 * sampled at the end of that period but right at the beginning of the
4407 * sampling interval and some other clock is sampled right at the
4408 * begining of its sampling period and right at the end of the
4409 * sampling interval. Let's assume the GPU has the longest clock
4410 * period and that the application is sampling GPU and monotonic:
4413 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
4414 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4418 * GPU -----_____-----_____-----_____-----_____
4421 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
4422 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4424 * Interval <----------------->
4425 * Deviation <-------------------------->
4429 * m = read(monotonic) 2
4432 * We round the sample interval up by one tick to cover sampling error
4433 * in the interval clock
4436 uint64_t sample_interval
= end
- begin
+ 1;
4438 *pMaxDeviation
= sample_interval
+ max_clock_period
;
4443 /* vk_icd.h does not declare this function, so we declare it here to
4444 * suppress Wmissing-prototypes.
4446 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4447 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
);
4449 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4450 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
)
4452 /* For the full details on loader interface versioning, see
4453 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
4454 * What follows is a condensed summary, to help you navigate the large and
4455 * confusing official doc.
4457 * - Loader interface v0 is incompatible with later versions. We don't
4460 * - In loader interface v1:
4461 * - The first ICD entrypoint called by the loader is
4462 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
4464 * - The ICD must statically expose no other Vulkan symbol unless it is
4465 * linked with -Bsymbolic.
4466 * - Each dispatchable Vulkan handle created by the ICD must be
4467 * a pointer to a struct whose first member is VK_LOADER_DATA. The
4468 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
4469 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
4470 * vkDestroySurfaceKHR(). The ICD must be capable of working with
4471 * such loader-managed surfaces.
4473 * - Loader interface v2 differs from v1 in:
4474 * - The first ICD entrypoint called by the loader is
4475 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
4476 * statically expose this entrypoint.
4478 * - Loader interface v3 differs from v2 in:
4479 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
4480 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
4481 * because the loader no longer does so.
4483 * - Loader interface v4 differs from v3 in:
4484 * - The ICD must implement vk_icdGetPhysicalDeviceProcAddr().
4486 *pSupportedVersion
= MIN2(*pSupportedVersion
, 4u);