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/strtod.h"
36 #include "util/debug.h"
37 #include "util/build_id.h"
38 #include "util/disk_cache.h"
39 #include "util/mesa-sha1.h"
40 #include "util/u_string.h"
43 #include "common/gen_defines.h"
44 #include "compiler/glsl_types.h"
46 #include "genxml/gen7_pack.h"
48 /* This is probably far to big but it reflects the max size used for messages
49 * in OpenGLs KHR_debug.
51 #define MAX_DEBUG_MESSAGE_LENGTH 4096
54 compiler_debug_log(void *data
, const char *fmt
, ...)
56 char str
[MAX_DEBUG_MESSAGE_LENGTH
];
57 struct anv_device
*device
= (struct anv_device
*)data
;
59 if (list_empty(&device
->instance
->debug_report_callbacks
.callbacks
))
64 (void) vsnprintf(str
, MAX_DEBUG_MESSAGE_LENGTH
, fmt
, args
);
67 vk_debug_report(&device
->instance
->debug_report_callbacks
,
68 VK_DEBUG_REPORT_DEBUG_BIT_EXT
,
69 VK_DEBUG_REPORT_OBJECT_TYPE_UNKNOWN_EXT
,
74 compiler_perf_log(void *data
, const char *fmt
, ...)
79 if (unlikely(INTEL_DEBUG
& DEBUG_PERF
))
80 intel_logd_v(fmt
, args
);
86 anv_compute_heap_size(int fd
, uint64_t gtt_size
)
88 /* Query the total ram from the system */
92 uint64_t total_ram
= (uint64_t)info
.totalram
* (uint64_t)info
.mem_unit
;
94 /* We don't want to burn too much ram with the GPU. If the user has 4GiB
95 * or less, we use at most half. If they have more than 4GiB, we use 3/4.
97 uint64_t available_ram
;
98 if (total_ram
<= 4ull * 1024ull * 1024ull * 1024ull)
99 available_ram
= total_ram
/ 2;
101 available_ram
= total_ram
* 3 / 4;
103 /* We also want to leave some padding for things we allocate in the driver,
104 * so don't go over 3/4 of the GTT either.
106 uint64_t available_gtt
= gtt_size
* 3 / 4;
108 return MIN2(available_ram
, available_gtt
);
112 anv_physical_device_init_heaps(struct anv_physical_device
*device
, int fd
)
115 if (anv_gem_get_context_param(fd
, 0, I915_CONTEXT_PARAM_GTT_SIZE
,
117 /* If, for whatever reason, we can't actually get the GTT size from the
118 * kernel (too old?) fall back to the aperture size.
120 anv_perf_warn(NULL
, NULL
,
121 "Failed to get I915_CONTEXT_PARAM_GTT_SIZE: %m");
123 if (anv_gem_get_aperture(fd
, >t_size
) == -1) {
124 return vk_errorf(NULL
, NULL
, VK_ERROR_INITIALIZATION_FAILED
,
125 "failed to get aperture size: %m");
129 device
->supports_48bit_addresses
= (device
->info
.gen
>= 8) &&
130 gtt_size
> (4ULL << 30 /* GiB */);
132 uint64_t heap_size
= anv_compute_heap_size(fd
, gtt_size
);
134 if (heap_size
> (2ull << 30) && !device
->supports_48bit_addresses
) {
135 /* When running with an overridden PCI ID, we may get a GTT size from
136 * the kernel that is greater than 2 GiB but the execbuf check for 48bit
137 * address support can still fail. Just clamp the address space size to
138 * 2 GiB if we don't have 48-bit support.
140 intel_logw("%s:%d: The kernel reported a GTT size larger than 2 GiB but "
141 "not support for 48-bit addresses",
143 heap_size
= 2ull << 30;
146 if (heap_size
<= 3ull * (1ull << 30)) {
147 /* In this case, everything fits nicely into the 32-bit address space,
148 * so there's no need for supporting 48bit addresses on client-allocated
151 device
->memory
.heap_count
= 1;
152 device
->memory
.heaps
[0] = (struct anv_memory_heap
) {
153 .vma_start
= LOW_HEAP_MIN_ADDRESS
,
154 .vma_size
= LOW_HEAP_SIZE
,
156 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
157 .supports_48bit_addresses
= false,
160 /* Not everything will fit nicely into a 32-bit address space. In this
161 * case we need a 64-bit heap. Advertise a small 32-bit heap and a
162 * larger 48-bit heap. If we're in this case, then we have a total heap
163 * size larger than 3GiB which most likely means they have 8 GiB of
164 * video memory and so carving off 1 GiB for the 32-bit heap should be
167 const uint64_t heap_size_32bit
= 1ull << 30;
168 const uint64_t heap_size_48bit
= heap_size
- heap_size_32bit
;
170 assert(device
->supports_48bit_addresses
);
172 device
->memory
.heap_count
= 2;
173 device
->memory
.heaps
[0] = (struct anv_memory_heap
) {
174 .vma_start
= HIGH_HEAP_MIN_ADDRESS
,
175 /* Leave the last 4GiB out of the high vma range, so that no state
176 * base address + size can overflow 48 bits. For more information see
177 * the comment about Wa32bitGeneralStateOffset in anv_allocator.c
179 .vma_size
= gtt_size
- (1ull << 32) - HIGH_HEAP_MIN_ADDRESS
,
180 .size
= heap_size_48bit
,
181 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
182 .supports_48bit_addresses
= true,
184 device
->memory
.heaps
[1] = (struct anv_memory_heap
) {
185 .vma_start
= LOW_HEAP_MIN_ADDRESS
,
186 .vma_size
= LOW_HEAP_SIZE
,
187 .size
= heap_size_32bit
,
188 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
189 .supports_48bit_addresses
= false,
193 uint32_t type_count
= 0;
194 for (uint32_t heap
= 0; heap
< device
->memory
.heap_count
; heap
++) {
195 uint32_t valid_buffer_usage
= ~0;
197 /* There appears to be a hardware issue in the VF cache where it only
198 * considers the bottom 32 bits of memory addresses. If you happen to
199 * have two vertex buffers which get placed exactly 4 GiB apart and use
200 * them in back-to-back draw calls, you can get collisions. In order to
201 * solve this problem, we require vertex and index buffers be bound to
202 * memory allocated out of the 32-bit heap.
204 if (device
->memory
.heaps
[heap
].supports_48bit_addresses
) {
205 valid_buffer_usage
&= ~(VK_BUFFER_USAGE_INDEX_BUFFER_BIT
|
206 VK_BUFFER_USAGE_VERTEX_BUFFER_BIT
);
209 if (device
->info
.has_llc
) {
210 /* Big core GPUs share LLC with the CPU and thus one memory type can be
211 * both cached and coherent at the same time.
213 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
214 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
215 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
216 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
217 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
219 .valid_buffer_usage
= valid_buffer_usage
,
222 /* The spec requires that we expose a host-visible, coherent memory
223 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
224 * to give the application a choice between cached, but not coherent and
225 * coherent but uncached (WC though).
227 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
228 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
229 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
230 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
232 .valid_buffer_usage
= valid_buffer_usage
,
234 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
235 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
236 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
237 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
239 .valid_buffer_usage
= valid_buffer_usage
,
243 device
->memory
.type_count
= type_count
;
249 anv_physical_device_init_uuids(struct anv_physical_device
*device
)
251 const struct build_id_note
*note
=
252 build_id_find_nhdr_for_addr(anv_physical_device_init_uuids
);
254 return vk_errorf(device
->instance
, device
,
255 VK_ERROR_INITIALIZATION_FAILED
,
256 "Failed to find build-id");
259 unsigned build_id_len
= build_id_length(note
);
260 if (build_id_len
< 20) {
261 return vk_errorf(device
->instance
, device
,
262 VK_ERROR_INITIALIZATION_FAILED
,
263 "build-id too short. It needs to be a SHA");
266 memcpy(device
->driver_build_sha1
, build_id_data(note
), 20);
268 struct mesa_sha1 sha1_ctx
;
270 STATIC_ASSERT(VK_UUID_SIZE
<= sizeof(sha1
));
272 /* The pipeline cache UUID is used for determining when a pipeline cache is
273 * invalid. It needs both a driver build and the PCI ID of the device.
275 _mesa_sha1_init(&sha1_ctx
);
276 _mesa_sha1_update(&sha1_ctx
, build_id_data(note
), build_id_len
);
277 _mesa_sha1_update(&sha1_ctx
, &device
->chipset_id
,
278 sizeof(device
->chipset_id
));
279 _mesa_sha1_update(&sha1_ctx
, &device
->always_use_bindless
,
280 sizeof(device
->always_use_bindless
));
281 _mesa_sha1_update(&sha1_ctx
, &device
->has_a64_buffer_access
,
282 sizeof(device
->has_a64_buffer_access
));
283 _mesa_sha1_update(&sha1_ctx
, &device
->has_bindless_images
,
284 sizeof(device
->has_bindless_images
));
285 _mesa_sha1_update(&sha1_ctx
, &device
->has_bindless_samplers
,
286 sizeof(device
->has_bindless_samplers
));
287 _mesa_sha1_final(&sha1_ctx
, sha1
);
288 memcpy(device
->pipeline_cache_uuid
, sha1
, VK_UUID_SIZE
);
290 /* The driver UUID is used for determining sharability of images and memory
291 * between two Vulkan instances in separate processes. People who want to
292 * share memory need to also check the device UUID (below) so all this
293 * needs to be is the build-id.
295 memcpy(device
->driver_uuid
, build_id_data(note
), VK_UUID_SIZE
);
297 /* The device UUID uniquely identifies the given device within the machine.
298 * Since we never have more than one device, this doesn't need to be a real
299 * UUID. However, on the off-chance that someone tries to use this to
300 * cache pre-tiled images or something of the like, we use the PCI ID and
301 * some bits of ISL info to ensure that this is safe.
303 _mesa_sha1_init(&sha1_ctx
);
304 _mesa_sha1_update(&sha1_ctx
, &device
->chipset_id
,
305 sizeof(device
->chipset_id
));
306 _mesa_sha1_update(&sha1_ctx
, &device
->isl_dev
.has_bit6_swizzling
,
307 sizeof(device
->isl_dev
.has_bit6_swizzling
));
308 _mesa_sha1_final(&sha1_ctx
, sha1
);
309 memcpy(device
->device_uuid
, sha1
, VK_UUID_SIZE
);
315 anv_physical_device_init_disk_cache(struct anv_physical_device
*device
)
317 #ifdef ENABLE_SHADER_CACHE
319 MAYBE_UNUSED
int len
= snprintf(renderer
, sizeof(renderer
), "anv_%04x",
321 assert(len
== sizeof(renderer
) - 2);
324 _mesa_sha1_format(timestamp
, device
->driver_build_sha1
);
326 const uint64_t driver_flags
=
327 brw_get_compiler_config_value(device
->compiler
);
328 device
->disk_cache
= disk_cache_create(renderer
, timestamp
, driver_flags
);
330 device
->disk_cache
= NULL
;
335 anv_physical_device_free_disk_cache(struct anv_physical_device
*device
)
337 #ifdef ENABLE_SHADER_CACHE
338 if (device
->disk_cache
)
339 disk_cache_destroy(device
->disk_cache
);
341 assert(device
->disk_cache
== NULL
);
346 anv_physical_device_init(struct anv_physical_device
*device
,
347 struct anv_instance
*instance
,
348 drmDevicePtr drm_device
)
350 const char *primary_path
= drm_device
->nodes
[DRM_NODE_PRIMARY
];
351 const char *path
= drm_device
->nodes
[DRM_NODE_RENDER
];
356 brw_process_intel_debug_variable();
358 fd
= open(path
, O_RDWR
| O_CLOEXEC
);
360 return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
362 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
363 device
->instance
= instance
;
365 assert(strlen(path
) < ARRAY_SIZE(device
->path
));
366 snprintf(device
->path
, ARRAY_SIZE(device
->path
), "%s", path
);
368 device
->no_hw
= getenv("INTEL_NO_HW") != NULL
;
370 const int pci_id_override
= gen_get_pci_device_id_override();
371 if (pci_id_override
< 0) {
372 device
->chipset_id
= anv_gem_get_param(fd
, I915_PARAM_CHIPSET_ID
);
373 if (!device
->chipset_id
) {
374 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
378 device
->chipset_id
= pci_id_override
;
379 device
->no_hw
= true;
382 device
->pci_info
.domain
= drm_device
->businfo
.pci
->domain
;
383 device
->pci_info
.bus
= drm_device
->businfo
.pci
->bus
;
384 device
->pci_info
.device
= drm_device
->businfo
.pci
->dev
;
385 device
->pci_info
.function
= drm_device
->businfo
.pci
->func
;
387 device
->name
= gen_get_device_name(device
->chipset_id
);
388 if (!gen_get_device_info(device
->chipset_id
, &device
->info
)) {
389 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
393 if (device
->info
.is_haswell
) {
394 intel_logw("Haswell Vulkan support is incomplete");
395 } else if (device
->info
.gen
== 7 && !device
->info
.is_baytrail
) {
396 intel_logw("Ivy Bridge Vulkan support is incomplete");
397 } else if (device
->info
.gen
== 7 && device
->info
.is_baytrail
) {
398 intel_logw("Bay Trail Vulkan support is incomplete");
399 } else if (device
->info
.gen
>= 8 && device
->info
.gen
<= 10) {
400 /* Gen8-10 fully supported */
401 } else if (device
->info
.gen
== 11) {
402 intel_logw("Vulkan is not yet fully supported on gen11.");
404 result
= vk_errorf(device
->instance
, device
,
405 VK_ERROR_INCOMPATIBLE_DRIVER
,
406 "Vulkan not yet supported on %s", device
->name
);
410 device
->cmd_parser_version
= -1;
411 if (device
->info
.gen
== 7) {
412 device
->cmd_parser_version
=
413 anv_gem_get_param(fd
, I915_PARAM_CMD_PARSER_VERSION
);
414 if (device
->cmd_parser_version
== -1) {
415 result
= vk_errorf(device
->instance
, device
,
416 VK_ERROR_INITIALIZATION_FAILED
,
417 "failed to get command parser version");
422 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_WAIT_TIMEOUT
)) {
423 result
= vk_errorf(device
->instance
, device
,
424 VK_ERROR_INITIALIZATION_FAILED
,
425 "kernel missing gem wait");
429 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_EXECBUF2
)) {
430 result
= vk_errorf(device
->instance
, device
,
431 VK_ERROR_INITIALIZATION_FAILED
,
432 "kernel missing execbuf2");
436 if (!device
->info
.has_llc
&&
437 anv_gem_get_param(fd
, I915_PARAM_MMAP_VERSION
) < 1) {
438 result
= vk_errorf(device
->instance
, device
,
439 VK_ERROR_INITIALIZATION_FAILED
,
440 "kernel missing wc mmap");
444 result
= anv_physical_device_init_heaps(device
, fd
);
445 if (result
!= VK_SUCCESS
)
448 device
->has_exec_async
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_ASYNC
);
449 device
->has_exec_capture
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_CAPTURE
);
450 device
->has_exec_fence
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE
);
451 device
->has_syncobj
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE_ARRAY
);
452 device
->has_syncobj_wait
= device
->has_syncobj
&&
453 anv_gem_supports_syncobj_wait(fd
);
454 device
->has_context_priority
= anv_gem_has_context_priority(fd
);
456 device
->use_softpin
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_SOFTPIN
)
457 && device
->supports_48bit_addresses
;
459 device
->has_context_isolation
=
460 anv_gem_get_param(fd
, I915_PARAM_HAS_CONTEXT_ISOLATION
);
462 device
->always_use_bindless
=
463 env_var_as_boolean("ANV_ALWAYS_BINDLESS", false);
465 /* We first got the A64 messages on broadwell and we can only use them if
466 * we can pass addresses directly into the shader which requires softpin.
468 device
->has_a64_buffer_access
= device
->info
.gen
>= 8 &&
471 /* We first get bindless image access on Skylake and we can only really do
472 * it if we don't have any relocations so we need softpin.
474 device
->has_bindless_images
= device
->info
.gen
>= 9 &&
477 /* We've had bindless samplers since Ivy Bridge (forever in Vulkan terms)
478 * because it's just a matter of setting the sampler address in the sample
479 * message header. However, we've not bothered to wire it up for vec4 so
480 * we leave it disabled on gen7.
482 device
->has_bindless_samplers
= device
->info
.gen
>= 8;
484 /* Starting with Gen10, the timestamp frequency of the command streamer may
485 * vary from one part to another. We can query the value from the kernel.
487 if (device
->info
.gen
>= 10) {
488 int timestamp_frequency
=
489 anv_gem_get_param(fd
, I915_PARAM_CS_TIMESTAMP_FREQUENCY
);
491 if (timestamp_frequency
< 0)
492 intel_logw("Kernel 4.16-rc1+ required to properly query CS timestamp frequency");
494 device
->info
.timestamp_frequency
= timestamp_frequency
;
497 /* GENs prior to 8 do not support EU/Subslice info */
498 if (device
->info
.gen
>= 8) {
499 device
->subslice_total
= anv_gem_get_param(fd
, I915_PARAM_SUBSLICE_TOTAL
);
500 device
->eu_total
= anv_gem_get_param(fd
, I915_PARAM_EU_TOTAL
);
502 /* Without this information, we cannot get the right Braswell
503 * brandstrings, and we have to use conservative numbers for GPGPU on
504 * many platforms, but otherwise, things will just work.
506 if (device
->subslice_total
< 1 || device
->eu_total
< 1) {
507 intel_logw("Kernel 4.1 required to properly query GPU properties");
509 } else if (device
->info
.gen
== 7) {
510 device
->subslice_total
= 1 << (device
->info
.gt
- 1);
513 if (device
->info
.is_cherryview
&&
514 device
->subslice_total
> 0 && device
->eu_total
> 0) {
515 /* Logical CS threads = EUs per subslice * num threads per EU */
516 uint32_t max_cs_threads
=
517 device
->eu_total
/ device
->subslice_total
* device
->info
.num_thread_per_eu
;
519 /* Fuse configurations may give more threads than expected, never less. */
520 if (max_cs_threads
> device
->info
.max_cs_threads
)
521 device
->info
.max_cs_threads
= max_cs_threads
;
524 device
->compiler
= brw_compiler_create(NULL
, &device
->info
);
525 if (device
->compiler
== NULL
) {
526 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
529 device
->compiler
->shader_debug_log
= compiler_debug_log
;
530 device
->compiler
->shader_perf_log
= compiler_perf_log
;
531 device
->compiler
->supports_pull_constants
= false;
532 device
->compiler
->constant_buffer_0_is_relative
=
533 device
->info
.gen
< 8 || !device
->has_context_isolation
;
534 device
->compiler
->supports_shader_constants
= true;
536 /* Broadwell PRM says:
538 * "Before Gen8, there was a historical configuration control field to
539 * swizzle address bit[6] for in X/Y tiling modes. This was set in three
540 * different places: TILECTL[1:0], ARB_MODE[5:4], and
541 * DISP_ARB_CTL[14:13].
543 * For Gen8 and subsequent generations, the swizzle fields are all
544 * reserved, and the CPU's memory controller performs all address
545 * swizzling modifications."
548 device
->info
.gen
< 8 && anv_gem_get_bit6_swizzle(fd
, I915_TILING_X
);
550 isl_device_init(&device
->isl_dev
, &device
->info
, swizzled
);
552 result
= anv_physical_device_init_uuids(device
);
553 if (result
!= VK_SUCCESS
)
556 anv_physical_device_init_disk_cache(device
);
558 if (instance
->enabled_extensions
.KHR_display
) {
559 master_fd
= open(primary_path
, O_RDWR
| O_CLOEXEC
);
560 if (master_fd
>= 0) {
561 /* prod the device with a GETPARAM call which will fail if
562 * we don't have permission to even render on this device
564 if (anv_gem_get_param(master_fd
, I915_PARAM_CHIPSET_ID
) == 0) {
570 device
->master_fd
= master_fd
;
572 result
= anv_init_wsi(device
);
573 if (result
!= VK_SUCCESS
) {
574 ralloc_free(device
->compiler
);
575 anv_physical_device_free_disk_cache(device
);
579 anv_physical_device_get_supported_extensions(device
,
580 &device
->supported_extensions
);
583 device
->local_fd
= fd
;
595 anv_physical_device_finish(struct anv_physical_device
*device
)
597 anv_finish_wsi(device
);
598 anv_physical_device_free_disk_cache(device
);
599 ralloc_free(device
->compiler
);
600 close(device
->local_fd
);
601 if (device
->master_fd
>= 0)
602 close(device
->master_fd
);
606 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
607 VkSystemAllocationScope allocationScope
)
613 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
614 size_t align
, VkSystemAllocationScope allocationScope
)
616 return realloc(pOriginal
, size
);
620 default_free_func(void *pUserData
, void *pMemory
)
625 static const VkAllocationCallbacks default_alloc
= {
627 .pfnAllocation
= default_alloc_func
,
628 .pfnReallocation
= default_realloc_func
,
629 .pfnFree
= default_free_func
,
632 VkResult
anv_EnumerateInstanceExtensionProperties(
633 const char* pLayerName
,
634 uint32_t* pPropertyCount
,
635 VkExtensionProperties
* pProperties
)
637 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
639 for (int i
= 0; i
< ANV_INSTANCE_EXTENSION_COUNT
; i
++) {
640 if (anv_instance_extensions_supported
.extensions
[i
]) {
641 vk_outarray_append(&out
, prop
) {
642 *prop
= anv_instance_extensions
[i
];
647 return vk_outarray_status(&out
);
650 VkResult
anv_CreateInstance(
651 const VkInstanceCreateInfo
* pCreateInfo
,
652 const VkAllocationCallbacks
* pAllocator
,
653 VkInstance
* pInstance
)
655 struct anv_instance
*instance
;
658 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
660 struct anv_instance_extension_table enabled_extensions
= {};
661 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
663 for (idx
= 0; idx
< ANV_INSTANCE_EXTENSION_COUNT
; idx
++) {
664 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
665 anv_instance_extensions
[idx
].extensionName
) == 0)
669 if (idx
>= ANV_INSTANCE_EXTENSION_COUNT
)
670 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
672 if (!anv_instance_extensions_supported
.extensions
[idx
])
673 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
675 enabled_extensions
.extensions
[idx
] = true;
678 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
679 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
681 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
683 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
686 instance
->alloc
= *pAllocator
;
688 instance
->alloc
= default_alloc
;
690 instance
->app_info
= (struct anv_app_info
) { .api_version
= 0 };
691 if (pCreateInfo
->pApplicationInfo
) {
692 const VkApplicationInfo
*app
= pCreateInfo
->pApplicationInfo
;
694 instance
->app_info
.app_name
=
695 vk_strdup(&instance
->alloc
, app
->pApplicationName
,
696 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
697 instance
->app_info
.app_version
= app
->applicationVersion
;
699 instance
->app_info
.engine_name
=
700 vk_strdup(&instance
->alloc
, app
->pEngineName
,
701 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
702 instance
->app_info
.engine_version
= app
->engineVersion
;
704 instance
->app_info
.api_version
= app
->apiVersion
;
707 if (instance
->app_info
.api_version
== 0)
708 instance
->app_info
.api_version
= VK_API_VERSION_1_0
;
710 instance
->enabled_extensions
= enabled_extensions
;
712 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->dispatch
.entrypoints
); i
++) {
713 /* Vulkan requires that entrypoints for extensions which have not been
714 * enabled must not be advertised.
716 if (!anv_instance_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
717 &instance
->enabled_extensions
)) {
718 instance
->dispatch
.entrypoints
[i
] = NULL
;
720 instance
->dispatch
.entrypoints
[i
] =
721 anv_instance_dispatch_table
.entrypoints
[i
];
725 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->device_dispatch
.entrypoints
); i
++) {
726 /* Vulkan requires that entrypoints for extensions which have not been
727 * enabled must not be advertised.
729 if (!anv_device_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
730 &instance
->enabled_extensions
, NULL
)) {
731 instance
->device_dispatch
.entrypoints
[i
] = NULL
;
733 instance
->device_dispatch
.entrypoints
[i
] =
734 anv_device_dispatch_table
.entrypoints
[i
];
738 instance
->physicalDeviceCount
= -1;
740 result
= vk_debug_report_instance_init(&instance
->debug_report_callbacks
);
741 if (result
!= VK_SUCCESS
) {
742 vk_free2(&default_alloc
, pAllocator
, instance
);
743 return vk_error(result
);
746 instance
->pipeline_cache_enabled
=
747 env_var_as_boolean("ANV_ENABLE_PIPELINE_CACHE", true);
750 glsl_type_singleton_init_or_ref();
752 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
754 *pInstance
= anv_instance_to_handle(instance
);
759 void anv_DestroyInstance(
760 VkInstance _instance
,
761 const VkAllocationCallbacks
* pAllocator
)
763 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
768 if (instance
->physicalDeviceCount
> 0) {
769 /* We support at most one physical device. */
770 assert(instance
->physicalDeviceCount
== 1);
771 anv_physical_device_finish(&instance
->physicalDevice
);
774 vk_free(&instance
->alloc
, (char *)instance
->app_info
.app_name
);
775 vk_free(&instance
->alloc
, (char *)instance
->app_info
.engine_name
);
777 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
779 vk_debug_report_instance_destroy(&instance
->debug_report_callbacks
);
781 glsl_type_singleton_decref();
784 vk_free(&instance
->alloc
, instance
);
788 anv_enumerate_devices(struct anv_instance
*instance
)
790 /* TODO: Check for more devices ? */
791 drmDevicePtr devices
[8];
792 VkResult result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
795 instance
->physicalDeviceCount
= 0;
797 max_devices
= drmGetDevices2(0, devices
, ARRAY_SIZE(devices
));
799 return VK_ERROR_INCOMPATIBLE_DRIVER
;
801 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
802 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
803 devices
[i
]->bustype
== DRM_BUS_PCI
&&
804 devices
[i
]->deviceinfo
.pci
->vendor_id
== 0x8086) {
806 result
= anv_physical_device_init(&instance
->physicalDevice
,
807 instance
, devices
[i
]);
808 if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
812 drmFreeDevices(devices
, max_devices
);
814 if (result
== VK_SUCCESS
)
815 instance
->physicalDeviceCount
= 1;
821 anv_instance_ensure_physical_device(struct anv_instance
*instance
)
823 if (instance
->physicalDeviceCount
< 0) {
824 VkResult result
= anv_enumerate_devices(instance
);
825 if (result
!= VK_SUCCESS
&&
826 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
833 VkResult
anv_EnumeratePhysicalDevices(
834 VkInstance _instance
,
835 uint32_t* pPhysicalDeviceCount
,
836 VkPhysicalDevice
* pPhysicalDevices
)
838 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
839 VK_OUTARRAY_MAKE(out
, pPhysicalDevices
, pPhysicalDeviceCount
);
841 VkResult result
= anv_instance_ensure_physical_device(instance
);
842 if (result
!= VK_SUCCESS
)
845 if (instance
->physicalDeviceCount
== 0)
848 assert(instance
->physicalDeviceCount
== 1);
849 vk_outarray_append(&out
, i
) {
850 *i
= anv_physical_device_to_handle(&instance
->physicalDevice
);
853 return vk_outarray_status(&out
);
856 VkResult
anv_EnumeratePhysicalDeviceGroups(
857 VkInstance _instance
,
858 uint32_t* pPhysicalDeviceGroupCount
,
859 VkPhysicalDeviceGroupProperties
* pPhysicalDeviceGroupProperties
)
861 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
862 VK_OUTARRAY_MAKE(out
, pPhysicalDeviceGroupProperties
,
863 pPhysicalDeviceGroupCount
);
865 VkResult result
= anv_instance_ensure_physical_device(instance
);
866 if (result
!= VK_SUCCESS
)
869 if (instance
->physicalDeviceCount
== 0)
872 assert(instance
->physicalDeviceCount
== 1);
874 vk_outarray_append(&out
, p
) {
875 p
->physicalDeviceCount
= 1;
876 memset(p
->physicalDevices
, 0, sizeof(p
->physicalDevices
));
877 p
->physicalDevices
[0] =
878 anv_physical_device_to_handle(&instance
->physicalDevice
);
879 p
->subsetAllocation
= false;
881 vk_foreach_struct(ext
, p
->pNext
)
882 anv_debug_ignored_stype(ext
->sType
);
885 return vk_outarray_status(&out
);
888 void anv_GetPhysicalDeviceFeatures(
889 VkPhysicalDevice physicalDevice
,
890 VkPhysicalDeviceFeatures
* pFeatures
)
892 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
894 *pFeatures
= (VkPhysicalDeviceFeatures
) {
895 .robustBufferAccess
= true,
896 .fullDrawIndexUint32
= true,
897 .imageCubeArray
= true,
898 .independentBlend
= true,
899 .geometryShader
= true,
900 .tessellationShader
= true,
901 .sampleRateShading
= true,
902 .dualSrcBlend
= true,
904 .multiDrawIndirect
= true,
905 .drawIndirectFirstInstance
= true,
907 .depthBiasClamp
= true,
908 .fillModeNonSolid
= true,
909 .depthBounds
= false,
913 .multiViewport
= true,
914 .samplerAnisotropy
= true,
915 .textureCompressionETC2
= pdevice
->info
.gen
>= 8 ||
916 pdevice
->info
.is_baytrail
,
917 .textureCompressionASTC_LDR
= pdevice
->info
.gen
>= 9, /* FINISHME CHV */
918 .textureCompressionBC
= true,
919 .occlusionQueryPrecise
= true,
920 .pipelineStatisticsQuery
= true,
921 .fragmentStoresAndAtomics
= true,
922 .shaderTessellationAndGeometryPointSize
= true,
923 .shaderImageGatherExtended
= true,
924 .shaderStorageImageExtendedFormats
= true,
925 .shaderStorageImageMultisample
= false,
926 .shaderStorageImageReadWithoutFormat
= false,
927 .shaderStorageImageWriteWithoutFormat
= true,
928 .shaderUniformBufferArrayDynamicIndexing
= true,
929 .shaderSampledImageArrayDynamicIndexing
= true,
930 .shaderStorageBufferArrayDynamicIndexing
= true,
931 .shaderStorageImageArrayDynamicIndexing
= true,
932 .shaderClipDistance
= true,
933 .shaderCullDistance
= true,
934 .shaderFloat64
= pdevice
->info
.gen
>= 8 &&
935 pdevice
->info
.has_64bit_types
,
936 .shaderInt64
= pdevice
->info
.gen
>= 8 &&
937 pdevice
->info
.has_64bit_types
,
938 .shaderInt16
= pdevice
->info
.gen
>= 8,
939 .shaderResourceMinLod
= pdevice
->info
.gen
>= 9,
940 .variableMultisampleRate
= true,
941 .inheritedQueries
= true,
944 /* We can't do image stores in vec4 shaders */
945 pFeatures
->vertexPipelineStoresAndAtomics
=
946 pdevice
->compiler
->scalar_stage
[MESA_SHADER_VERTEX
] &&
947 pdevice
->compiler
->scalar_stage
[MESA_SHADER_GEOMETRY
];
949 struct anv_app_info
*app_info
= &pdevice
->instance
->app_info
;
951 /* The new DOOM and Wolfenstein games require depthBounds without
952 * checking for it. They seem to run fine without it so just claim it's
953 * there and accept the consequences.
955 if (app_info
->engine_name
&& strcmp(app_info
->engine_name
, "idTech") == 0)
956 pFeatures
->depthBounds
= true;
959 void anv_GetPhysicalDeviceFeatures2(
960 VkPhysicalDevice physicalDevice
,
961 VkPhysicalDeviceFeatures2
* pFeatures
)
963 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
964 anv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
966 vk_foreach_struct(ext
, pFeatures
->pNext
) {
967 switch (ext
->sType
) {
968 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES_KHR
: {
969 VkPhysicalDevice8BitStorageFeaturesKHR
*features
=
970 (VkPhysicalDevice8BitStorageFeaturesKHR
*)ext
;
971 features
->storageBuffer8BitAccess
= pdevice
->info
.gen
>= 8;
972 features
->uniformAndStorageBuffer8BitAccess
= pdevice
->info
.gen
>= 8;
973 features
->storagePushConstant8
= pdevice
->info
.gen
>= 8;
977 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES
: {
978 VkPhysicalDevice16BitStorageFeatures
*features
=
979 (VkPhysicalDevice16BitStorageFeatures
*)ext
;
980 features
->storageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
981 features
->uniformAndStorageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
982 features
->storagePushConstant16
= pdevice
->info
.gen
>= 8;
983 features
->storageInputOutput16
= false;
987 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_EXT
: {
988 VkPhysicalDeviceBufferDeviceAddressFeaturesEXT
*features
= (void *)ext
;
989 features
->bufferDeviceAddress
= pdevice
->has_a64_buffer_access
;
990 features
->bufferDeviceAddressCaptureReplay
= false;
991 features
->bufferDeviceAddressMultiDevice
= false;
995 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COMPUTE_SHADER_DERIVATIVES_FEATURES_NV
: {
996 VkPhysicalDeviceComputeShaderDerivativesFeaturesNV
*features
=
997 (VkPhysicalDeviceComputeShaderDerivativesFeaturesNV
*)ext
;
998 features
->computeDerivativeGroupQuads
= true;
999 features
->computeDerivativeGroupLinear
= true;
1003 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT
: {
1004 VkPhysicalDeviceConditionalRenderingFeaturesEXT
*features
=
1005 (VkPhysicalDeviceConditionalRenderingFeaturesEXT
*)ext
;
1006 features
->conditionalRendering
= pdevice
->info
.gen
>= 8 ||
1007 pdevice
->info
.is_haswell
;
1008 features
->inheritedConditionalRendering
= pdevice
->info
.gen
>= 8 ||
1009 pdevice
->info
.is_haswell
;
1013 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_CLIP_ENABLE_FEATURES_EXT
: {
1014 VkPhysicalDeviceDepthClipEnableFeaturesEXT
*features
=
1015 (VkPhysicalDeviceDepthClipEnableFeaturesEXT
*)ext
;
1016 features
->depthClipEnable
= true;
1020 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR
: {
1021 VkPhysicalDeviceFloat16Int8FeaturesKHR
*features
= (void *)ext
;
1022 features
->shaderFloat16
= pdevice
->info
.gen
>= 8;
1023 features
->shaderInt8
= pdevice
->info
.gen
>= 8;
1027 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_QUERY_RESET_FEATURES_EXT
: {
1028 VkPhysicalDeviceHostQueryResetFeaturesEXT
*features
=
1029 (VkPhysicalDeviceHostQueryResetFeaturesEXT
*)ext
;
1030 features
->hostQueryReset
= true;
1034 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT
: {
1035 VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*features
=
1036 (VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*)ext
;
1037 features
->shaderInputAttachmentArrayDynamicIndexing
= false;
1038 features
->shaderUniformTexelBufferArrayDynamicIndexing
= true;
1039 features
->shaderStorageTexelBufferArrayDynamicIndexing
= true;
1040 features
->shaderUniformBufferArrayNonUniformIndexing
= false;
1041 features
->shaderSampledImageArrayNonUniformIndexing
= true;
1042 features
->shaderStorageBufferArrayNonUniformIndexing
= true;
1043 features
->shaderStorageImageArrayNonUniformIndexing
= true;
1044 features
->shaderInputAttachmentArrayNonUniformIndexing
= false;
1045 features
->shaderUniformTexelBufferArrayNonUniformIndexing
= true;
1046 features
->shaderStorageTexelBufferArrayNonUniformIndexing
= true;
1047 features
->descriptorBindingUniformBufferUpdateAfterBind
= false;
1048 features
->descriptorBindingSampledImageUpdateAfterBind
= true;
1049 features
->descriptorBindingStorageImageUpdateAfterBind
= true;
1050 features
->descriptorBindingStorageBufferUpdateAfterBind
= true;
1051 features
->descriptorBindingUniformTexelBufferUpdateAfterBind
= true;
1052 features
->descriptorBindingStorageTexelBufferUpdateAfterBind
= true;
1053 features
->descriptorBindingUpdateUnusedWhilePending
= true;
1054 features
->descriptorBindingPartiallyBound
= true;
1055 features
->descriptorBindingVariableDescriptorCount
= false;
1056 features
->runtimeDescriptorArray
= true;
1060 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_FEATURES_EXT
: {
1061 VkPhysicalDeviceInlineUniformBlockFeaturesEXT
*features
=
1062 (VkPhysicalDeviceInlineUniformBlockFeaturesEXT
*)ext
;
1063 features
->inlineUniformBlock
= true;
1064 features
->descriptorBindingInlineUniformBlockUpdateAfterBind
= true;
1068 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES
: {
1069 VkPhysicalDeviceMultiviewFeatures
*features
=
1070 (VkPhysicalDeviceMultiviewFeatures
*)ext
;
1071 features
->multiview
= true;
1072 features
->multiviewGeometryShader
= true;
1073 features
->multiviewTessellationShader
= true;
1077 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES
: {
1078 VkPhysicalDeviceProtectedMemoryFeatures
*features
= (void *)ext
;
1079 features
->protectedMemory
= false;
1083 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES
: {
1084 VkPhysicalDeviceSamplerYcbcrConversionFeatures
*features
=
1085 (VkPhysicalDeviceSamplerYcbcrConversionFeatures
*) ext
;
1086 features
->samplerYcbcrConversion
= true;
1090 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES_EXT
: {
1091 VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*features
=
1092 (VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*)ext
;
1093 features
->scalarBlockLayout
= true;
1097 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_INT64_FEATURES_KHR
: {
1098 VkPhysicalDeviceShaderAtomicInt64FeaturesKHR
*features
= (void *)ext
;
1099 features
->shaderBufferInt64Atomics
=
1100 pdevice
->info
.gen
>= 9 && pdevice
->use_softpin
;
1101 features
->shaderSharedInt64Atomics
= VK_FALSE
;
1105 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETERS_FEATURES
: {
1106 VkPhysicalDeviceShaderDrawParametersFeatures
*features
= (void *)ext
;
1107 features
->shaderDrawParameters
= true;
1111 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES
: {
1112 VkPhysicalDeviceVariablePointersFeatures
*features
= (void *)ext
;
1113 features
->variablePointersStorageBuffer
= true;
1114 features
->variablePointers
= true;
1118 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT
: {
1119 VkPhysicalDeviceTransformFeedbackFeaturesEXT
*features
=
1120 (VkPhysicalDeviceTransformFeedbackFeaturesEXT
*)ext
;
1121 features
->transformFeedback
= true;
1122 features
->geometryStreams
= true;
1126 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT
: {
1127 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*features
=
1128 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*)ext
;
1129 features
->vertexAttributeInstanceRateDivisor
= true;
1130 features
->vertexAttributeInstanceRateZeroDivisor
= true;
1134 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_YCBCR_IMAGE_ARRAYS_FEATURES_EXT
: {
1135 VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*features
=
1136 (VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*)ext
;
1137 features
->ycbcrImageArrays
= true;
1142 anv_debug_ignored_stype(ext
->sType
);
1148 void anv_GetPhysicalDeviceProperties(
1149 VkPhysicalDevice physicalDevice
,
1150 VkPhysicalDeviceProperties
* pProperties
)
1152 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1153 const struct gen_device_info
*devinfo
= &pdevice
->info
;
1155 /* See assertions made when programming the buffer surface state. */
1156 const uint32_t max_raw_buffer_sz
= devinfo
->gen
>= 7 ?
1157 (1ul << 30) : (1ul << 27);
1159 const uint32_t max_ssbos
= pdevice
->has_a64_buffer_access
? UINT16_MAX
: 64;
1160 const uint32_t max_textures
=
1161 pdevice
->has_bindless_images
? UINT16_MAX
: 128;
1162 const uint32_t max_samplers
=
1163 pdevice
->has_bindless_samplers
? UINT16_MAX
:
1164 (devinfo
->gen
>= 8 || devinfo
->is_haswell
) ? 128 : 16;
1165 const uint32_t max_images
=
1166 pdevice
->has_bindless_images
? UINT16_MAX
: MAX_IMAGES
;
1168 /* The moment we have anything bindless, claim a high per-stage limit */
1169 const uint32_t max_per_stage
=
1170 pdevice
->has_a64_buffer_access
? UINT32_MAX
:
1171 MAX_BINDING_TABLE_SIZE
- MAX_RTS
;
1173 VkSampleCountFlags sample_counts
=
1174 isl_device_get_sample_counts(&pdevice
->isl_dev
);
1177 VkPhysicalDeviceLimits limits
= {
1178 .maxImageDimension1D
= (1 << 14),
1179 .maxImageDimension2D
= (1 << 14),
1180 .maxImageDimension3D
= (1 << 11),
1181 .maxImageDimensionCube
= (1 << 14),
1182 .maxImageArrayLayers
= (1 << 11),
1183 .maxTexelBufferElements
= 128 * 1024 * 1024,
1184 .maxUniformBufferRange
= (1ul << 27),
1185 .maxStorageBufferRange
= max_raw_buffer_sz
,
1186 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
1187 .maxMemoryAllocationCount
= UINT32_MAX
,
1188 .maxSamplerAllocationCount
= 64 * 1024,
1189 .bufferImageGranularity
= 64, /* A cache line */
1190 .sparseAddressSpaceSize
= 0,
1191 .maxBoundDescriptorSets
= MAX_SETS
,
1192 .maxPerStageDescriptorSamplers
= max_samplers
,
1193 .maxPerStageDescriptorUniformBuffers
= 64,
1194 .maxPerStageDescriptorStorageBuffers
= max_ssbos
,
1195 .maxPerStageDescriptorSampledImages
= max_textures
,
1196 .maxPerStageDescriptorStorageImages
= max_images
,
1197 .maxPerStageDescriptorInputAttachments
= 64,
1198 .maxPerStageResources
= max_per_stage
,
1199 .maxDescriptorSetSamplers
= 6 * max_samplers
, /* number of stages * maxPerStageDescriptorSamplers */
1200 .maxDescriptorSetUniformBuffers
= 6 * 64, /* number of stages * maxPerStageDescriptorUniformBuffers */
1201 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
1202 .maxDescriptorSetStorageBuffers
= 6 * max_ssbos
, /* number of stages * maxPerStageDescriptorStorageBuffers */
1203 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
1204 .maxDescriptorSetSampledImages
= 6 * max_textures
, /* number of stages * maxPerStageDescriptorSampledImages */
1205 .maxDescriptorSetStorageImages
= 6 * max_images
, /* number of stages * maxPerStageDescriptorStorageImages */
1206 .maxDescriptorSetInputAttachments
= 256,
1207 .maxVertexInputAttributes
= MAX_VBS
,
1208 .maxVertexInputBindings
= MAX_VBS
,
1209 .maxVertexInputAttributeOffset
= 2047,
1210 .maxVertexInputBindingStride
= 2048,
1211 .maxVertexOutputComponents
= 128,
1212 .maxTessellationGenerationLevel
= 64,
1213 .maxTessellationPatchSize
= 32,
1214 .maxTessellationControlPerVertexInputComponents
= 128,
1215 .maxTessellationControlPerVertexOutputComponents
= 128,
1216 .maxTessellationControlPerPatchOutputComponents
= 128,
1217 .maxTessellationControlTotalOutputComponents
= 2048,
1218 .maxTessellationEvaluationInputComponents
= 128,
1219 .maxTessellationEvaluationOutputComponents
= 128,
1220 .maxGeometryShaderInvocations
= 32,
1221 .maxGeometryInputComponents
= 64,
1222 .maxGeometryOutputComponents
= 128,
1223 .maxGeometryOutputVertices
= 256,
1224 .maxGeometryTotalOutputComponents
= 1024,
1225 .maxFragmentInputComponents
= 112, /* 128 components - (POS, PSIZ, CLIP_DIST0, CLIP_DIST1) */
1226 .maxFragmentOutputAttachments
= 8,
1227 .maxFragmentDualSrcAttachments
= 1,
1228 .maxFragmentCombinedOutputResources
= 8,
1229 .maxComputeSharedMemorySize
= 32768,
1230 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
1231 .maxComputeWorkGroupInvocations
= 32 * devinfo
->max_cs_threads
,
1232 .maxComputeWorkGroupSize
= {
1233 16 * devinfo
->max_cs_threads
,
1234 16 * devinfo
->max_cs_threads
,
1235 16 * devinfo
->max_cs_threads
,
1237 .subPixelPrecisionBits
= 8,
1238 .subTexelPrecisionBits
= 8,
1239 .mipmapPrecisionBits
= 8,
1240 .maxDrawIndexedIndexValue
= UINT32_MAX
,
1241 .maxDrawIndirectCount
= UINT32_MAX
,
1242 .maxSamplerLodBias
= 16,
1243 .maxSamplerAnisotropy
= 16,
1244 .maxViewports
= MAX_VIEWPORTS
,
1245 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
1246 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
1247 .viewportSubPixelBits
= 13, /* We take a float? */
1248 .minMemoryMapAlignment
= 4096, /* A page */
1249 .minTexelBufferOffsetAlignment
= 1,
1250 /* We need 16 for UBO block reads to work and 32 for push UBOs */
1251 .minUniformBufferOffsetAlignment
= 32,
1252 .minStorageBufferOffsetAlignment
= 4,
1253 .minTexelOffset
= -8,
1254 .maxTexelOffset
= 7,
1255 .minTexelGatherOffset
= -32,
1256 .maxTexelGatherOffset
= 31,
1257 .minInterpolationOffset
= -0.5,
1258 .maxInterpolationOffset
= 0.4375,
1259 .subPixelInterpolationOffsetBits
= 4,
1260 .maxFramebufferWidth
= (1 << 14),
1261 .maxFramebufferHeight
= (1 << 14),
1262 .maxFramebufferLayers
= (1 << 11),
1263 .framebufferColorSampleCounts
= sample_counts
,
1264 .framebufferDepthSampleCounts
= sample_counts
,
1265 .framebufferStencilSampleCounts
= sample_counts
,
1266 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
1267 .maxColorAttachments
= MAX_RTS
,
1268 .sampledImageColorSampleCounts
= sample_counts
,
1269 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
1270 .sampledImageDepthSampleCounts
= sample_counts
,
1271 .sampledImageStencilSampleCounts
= sample_counts
,
1272 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
1273 .maxSampleMaskWords
= 1,
1274 .timestampComputeAndGraphics
= false,
1275 .timestampPeriod
= 1000000000.0 / devinfo
->timestamp_frequency
,
1276 .maxClipDistances
= 8,
1277 .maxCullDistances
= 8,
1278 .maxCombinedClipAndCullDistances
= 8,
1279 .discreteQueuePriorities
= 2,
1280 .pointSizeRange
= { 0.125, 255.875 },
1281 .lineWidthRange
= { 0.0, 7.9921875 },
1282 .pointSizeGranularity
= (1.0 / 8.0),
1283 .lineWidthGranularity
= (1.0 / 128.0),
1284 .strictLines
= false, /* FINISHME */
1285 .standardSampleLocations
= true,
1286 .optimalBufferCopyOffsetAlignment
= 128,
1287 .optimalBufferCopyRowPitchAlignment
= 128,
1288 .nonCoherentAtomSize
= 64,
1291 *pProperties
= (VkPhysicalDeviceProperties
) {
1292 .apiVersion
= anv_physical_device_api_version(pdevice
),
1293 .driverVersion
= vk_get_driver_version(),
1295 .deviceID
= pdevice
->chipset_id
,
1296 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
1298 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
1301 snprintf(pProperties
->deviceName
, sizeof(pProperties
->deviceName
),
1302 "%s", pdevice
->name
);
1303 memcpy(pProperties
->pipelineCacheUUID
,
1304 pdevice
->pipeline_cache_uuid
, VK_UUID_SIZE
);
1307 void anv_GetPhysicalDeviceProperties2(
1308 VkPhysicalDevice physicalDevice
,
1309 VkPhysicalDeviceProperties2
* pProperties
)
1311 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1313 anv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
1315 vk_foreach_struct(ext
, pProperties
->pNext
) {
1316 switch (ext
->sType
) {
1317 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_STENCIL_RESOLVE_PROPERTIES_KHR
: {
1318 VkPhysicalDeviceDepthStencilResolvePropertiesKHR
*props
=
1319 (VkPhysicalDeviceDepthStencilResolvePropertiesKHR
*)ext
;
1321 /* We support all of the depth resolve modes */
1322 props
->supportedDepthResolveModes
=
1323 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
|
1324 VK_RESOLVE_MODE_AVERAGE_BIT_KHR
|
1325 VK_RESOLVE_MODE_MIN_BIT_KHR
|
1326 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1328 /* Average doesn't make sense for stencil so we don't support that */
1329 props
->supportedStencilResolveModes
=
1330 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
;
1331 if (pdevice
->info
.gen
>= 8) {
1332 /* The advanced stencil resolve modes currently require stencil
1333 * sampling be supported by the hardware.
1335 props
->supportedStencilResolveModes
|=
1336 VK_RESOLVE_MODE_MIN_BIT_KHR
|
1337 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1340 props
->independentResolveNone
= true;
1341 props
->independentResolve
= true;
1345 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES_EXT
: {
1346 VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*props
=
1347 (VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*)ext
;
1349 /* It's a bit hard to exactly map our implementation to the limits
1350 * described here. The bindless surface handle in the extended
1351 * message descriptors is 20 bits and it's an index into the table of
1352 * RENDER_SURFACE_STATE structs that starts at bindless surface base
1353 * address. Given that most things consume two surface states per
1354 * view (general/sampled for textures and write-only/read-write for
1355 * images), we claim 2^19 things.
1357 * For SSBOs, we just use A64 messages so there is no real limit
1358 * there beyond the limit on the total size of a descriptor set.
1360 const unsigned max_bindless_views
= 1 << 19;
1362 props
->maxUpdateAfterBindDescriptorsInAllPools
= max_bindless_views
;
1363 props
->shaderUniformBufferArrayNonUniformIndexingNative
= false;
1364 props
->shaderSampledImageArrayNonUniformIndexingNative
= false;
1365 props
->shaderStorageBufferArrayNonUniformIndexingNative
= true;
1366 props
->shaderStorageImageArrayNonUniformIndexingNative
= false;
1367 props
->shaderInputAttachmentArrayNonUniformIndexingNative
= false;
1368 props
->robustBufferAccessUpdateAfterBind
= true;
1369 props
->quadDivergentImplicitLod
= false;
1370 props
->maxPerStageDescriptorUpdateAfterBindSamplers
= max_bindless_views
;
1371 props
->maxPerStageDescriptorUpdateAfterBindUniformBuffers
= 0;
1372 props
->maxPerStageDescriptorUpdateAfterBindStorageBuffers
= UINT32_MAX
;
1373 props
->maxPerStageDescriptorUpdateAfterBindSampledImages
= max_bindless_views
;
1374 props
->maxPerStageDescriptorUpdateAfterBindStorageImages
= max_bindless_views
;
1375 props
->maxPerStageDescriptorUpdateAfterBindInputAttachments
= 0;
1376 props
->maxPerStageUpdateAfterBindResources
= UINT32_MAX
;
1377 props
->maxDescriptorSetUpdateAfterBindSamplers
= max_bindless_views
;
1378 props
->maxDescriptorSetUpdateAfterBindUniformBuffers
= 0;
1379 props
->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic
= 0;
1380 props
->maxDescriptorSetUpdateAfterBindStorageBuffers
= UINT32_MAX
;
1381 props
->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2;
1382 props
->maxDescriptorSetUpdateAfterBindSampledImages
= max_bindless_views
;
1383 props
->maxDescriptorSetUpdateAfterBindStorageImages
= max_bindless_views
;
1384 props
->maxDescriptorSetUpdateAfterBindInputAttachments
= 0;
1388 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR
: {
1389 VkPhysicalDeviceDriverPropertiesKHR
*driver_props
=
1390 (VkPhysicalDeviceDriverPropertiesKHR
*) ext
;
1392 driver_props
->driverID
= VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA_KHR
;
1393 util_snprintf(driver_props
->driverName
, VK_MAX_DRIVER_NAME_SIZE_KHR
,
1394 "Intel open-source Mesa driver");
1396 util_snprintf(driver_props
->driverInfo
, VK_MAX_DRIVER_INFO_SIZE_KHR
,
1397 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1
);
1399 driver_props
->conformanceVersion
= (VkConformanceVersionKHR
) {
1408 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT
: {
1409 VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*props
=
1410 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*) ext
;
1411 /* Userptr needs page aligned memory. */
1412 props
->minImportedHostPointerAlignment
= 4096;
1416 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES
: {
1417 VkPhysicalDeviceIDProperties
*id_props
=
1418 (VkPhysicalDeviceIDProperties
*)ext
;
1419 memcpy(id_props
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
1420 memcpy(id_props
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
1421 /* The LUID is for Windows. */
1422 id_props
->deviceLUIDValid
= false;
1426 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_PROPERTIES_EXT
: {
1427 VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*props
=
1428 (VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*)ext
;
1429 props
->maxInlineUniformBlockSize
= MAX_INLINE_UNIFORM_BLOCK_SIZE
;
1430 props
->maxPerStageDescriptorInlineUniformBlocks
=
1431 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1432 props
->maxPerStageDescriptorUpdateAfterBindInlineUniformBlocks
=
1433 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1434 props
->maxDescriptorSetInlineUniformBlocks
=
1435 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1436 props
->maxDescriptorSetUpdateAfterBindInlineUniformBlocks
=
1437 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1441 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES
: {
1442 VkPhysicalDeviceMaintenance3Properties
*props
=
1443 (VkPhysicalDeviceMaintenance3Properties
*)ext
;
1444 /* This value doesn't matter for us today as our per-stage
1445 * descriptors are the real limit.
1447 props
->maxPerSetDescriptors
= 1024;
1448 props
->maxMemoryAllocationSize
= MAX_MEMORY_ALLOCATION_SIZE
;
1452 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES
: {
1453 VkPhysicalDeviceMultiviewProperties
*properties
=
1454 (VkPhysicalDeviceMultiviewProperties
*)ext
;
1455 properties
->maxMultiviewViewCount
= 16;
1456 properties
->maxMultiviewInstanceIndex
= UINT32_MAX
/ 16;
1460 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT
: {
1461 VkPhysicalDevicePCIBusInfoPropertiesEXT
*properties
=
1462 (VkPhysicalDevicePCIBusInfoPropertiesEXT
*)ext
;
1463 properties
->pciDomain
= pdevice
->pci_info
.domain
;
1464 properties
->pciBus
= pdevice
->pci_info
.bus
;
1465 properties
->pciDevice
= pdevice
->pci_info
.device
;
1466 properties
->pciFunction
= pdevice
->pci_info
.function
;
1470 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES
: {
1471 VkPhysicalDevicePointClippingProperties
*properties
=
1472 (VkPhysicalDevicePointClippingProperties
*) ext
;
1473 properties
->pointClippingBehavior
= VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES
;
1474 anv_finishme("Implement pop-free point clipping");
1478 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES
: {
1479 VkPhysicalDeviceProtectedMemoryProperties
*props
=
1480 (VkPhysicalDeviceProtectedMemoryProperties
*)ext
;
1481 props
->protectedNoFault
= false;
1485 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
1486 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
1487 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
1489 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
1493 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT
: {
1494 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*properties
=
1495 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*)ext
;
1496 properties
->filterMinmaxImageComponentMapping
= pdevice
->info
.gen
>= 9;
1497 properties
->filterMinmaxSingleComponentFormats
= true;
1501 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES
: {
1502 VkPhysicalDeviceSubgroupProperties
*properties
= (void *)ext
;
1504 properties
->subgroupSize
= BRW_SUBGROUP_SIZE
;
1506 VkShaderStageFlags scalar_stages
= 0;
1507 for (unsigned stage
= 0; stage
< MESA_SHADER_STAGES
; stage
++) {
1508 if (pdevice
->compiler
->scalar_stage
[stage
])
1509 scalar_stages
|= mesa_to_vk_shader_stage(stage
);
1511 properties
->supportedStages
= scalar_stages
;
1513 properties
->supportedOperations
= VK_SUBGROUP_FEATURE_BASIC_BIT
|
1514 VK_SUBGROUP_FEATURE_VOTE_BIT
|
1515 VK_SUBGROUP_FEATURE_ARITHMETIC_BIT
|
1516 VK_SUBGROUP_FEATURE_BALLOT_BIT
|
1517 VK_SUBGROUP_FEATURE_SHUFFLE_BIT
|
1518 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT
|
1519 VK_SUBGROUP_FEATURE_CLUSTERED_BIT
|
1520 VK_SUBGROUP_FEATURE_QUAD_BIT
;
1521 properties
->quadOperationsInAllStages
= true;
1525 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT
: {
1526 VkPhysicalDeviceTransformFeedbackPropertiesEXT
*props
=
1527 (VkPhysicalDeviceTransformFeedbackPropertiesEXT
*)ext
;
1529 props
->maxTransformFeedbackStreams
= MAX_XFB_STREAMS
;
1530 props
->maxTransformFeedbackBuffers
= MAX_XFB_BUFFERS
;
1531 props
->maxTransformFeedbackBufferSize
= (1ull << 32);
1532 props
->maxTransformFeedbackStreamDataSize
= 128 * 4;
1533 props
->maxTransformFeedbackBufferDataSize
= 128 * 4;
1534 props
->maxTransformFeedbackBufferDataStride
= 2048;
1535 props
->transformFeedbackQueries
= true;
1536 props
->transformFeedbackStreamsLinesTriangles
= false;
1537 props
->transformFeedbackRasterizationStreamSelect
= false;
1538 props
->transformFeedbackDraw
= true;
1542 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT
: {
1543 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*props
=
1544 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*)ext
;
1545 /* We have to restrict this a bit for multiview */
1546 props
->maxVertexAttribDivisor
= UINT32_MAX
/ 16;
1551 anv_debug_ignored_stype(ext
->sType
);
1557 /* We support exactly one queue family. */
1558 static const VkQueueFamilyProperties
1559 anv_queue_family_properties
= {
1560 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
1561 VK_QUEUE_COMPUTE_BIT
|
1562 VK_QUEUE_TRANSFER_BIT
,
1564 .timestampValidBits
= 36, /* XXX: Real value here */
1565 .minImageTransferGranularity
= { 1, 1, 1 },
1568 void anv_GetPhysicalDeviceQueueFamilyProperties(
1569 VkPhysicalDevice physicalDevice
,
1571 VkQueueFamilyProperties
* pQueueFamilyProperties
)
1573 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pCount
);
1575 vk_outarray_append(&out
, p
) {
1576 *p
= anv_queue_family_properties
;
1580 void anv_GetPhysicalDeviceQueueFamilyProperties2(
1581 VkPhysicalDevice physicalDevice
,
1582 uint32_t* pQueueFamilyPropertyCount
,
1583 VkQueueFamilyProperties2
* pQueueFamilyProperties
)
1586 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pQueueFamilyPropertyCount
);
1588 vk_outarray_append(&out
, p
) {
1589 p
->queueFamilyProperties
= anv_queue_family_properties
;
1591 vk_foreach_struct(s
, p
->pNext
) {
1592 anv_debug_ignored_stype(s
->sType
);
1597 void anv_GetPhysicalDeviceMemoryProperties(
1598 VkPhysicalDevice physicalDevice
,
1599 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
1601 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
1603 pMemoryProperties
->memoryTypeCount
= physical_device
->memory
.type_count
;
1604 for (uint32_t i
= 0; i
< physical_device
->memory
.type_count
; i
++) {
1605 pMemoryProperties
->memoryTypes
[i
] = (VkMemoryType
) {
1606 .propertyFlags
= physical_device
->memory
.types
[i
].propertyFlags
,
1607 .heapIndex
= physical_device
->memory
.types
[i
].heapIndex
,
1611 pMemoryProperties
->memoryHeapCount
= physical_device
->memory
.heap_count
;
1612 for (uint32_t i
= 0; i
< physical_device
->memory
.heap_count
; i
++) {
1613 pMemoryProperties
->memoryHeaps
[i
] = (VkMemoryHeap
) {
1614 .size
= physical_device
->memory
.heaps
[i
].size
,
1615 .flags
= physical_device
->memory
.heaps
[i
].flags
,
1620 void anv_GetPhysicalDeviceMemoryProperties2(
1621 VkPhysicalDevice physicalDevice
,
1622 VkPhysicalDeviceMemoryProperties2
* pMemoryProperties
)
1624 anv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
1625 &pMemoryProperties
->memoryProperties
);
1627 vk_foreach_struct(ext
, pMemoryProperties
->pNext
) {
1628 switch (ext
->sType
) {
1630 anv_debug_ignored_stype(ext
->sType
);
1637 anv_GetDeviceGroupPeerMemoryFeatures(
1640 uint32_t localDeviceIndex
,
1641 uint32_t remoteDeviceIndex
,
1642 VkPeerMemoryFeatureFlags
* pPeerMemoryFeatures
)
1644 assert(localDeviceIndex
== 0 && remoteDeviceIndex
== 0);
1645 *pPeerMemoryFeatures
= VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT
|
1646 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT
|
1647 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT
|
1648 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT
;
1651 PFN_vkVoidFunction
anv_GetInstanceProcAddr(
1652 VkInstance _instance
,
1655 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1657 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
1658 * when we have to return valid function pointers, NULL, or it's left
1659 * undefined. See the table for exact details.
1664 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
1665 if (strcmp(pName, "vk" #entrypoint) == 0) \
1666 return (PFN_vkVoidFunction)anv_##entrypoint
1668 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties
);
1669 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties
);
1670 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceVersion
);
1671 LOOKUP_ANV_ENTRYPOINT(CreateInstance
);
1673 #undef LOOKUP_ANV_ENTRYPOINT
1675 if (instance
== NULL
)
1678 int idx
= anv_get_instance_entrypoint_index(pName
);
1680 return instance
->dispatch
.entrypoints
[idx
];
1682 idx
= anv_get_device_entrypoint_index(pName
);
1684 return instance
->device_dispatch
.entrypoints
[idx
];
1689 /* With version 1+ of the loader interface the ICD should expose
1690 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
1693 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1694 VkInstance instance
,
1698 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1699 VkInstance instance
,
1702 return anv_GetInstanceProcAddr(instance
, pName
);
1705 PFN_vkVoidFunction
anv_GetDeviceProcAddr(
1709 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1711 if (!device
|| !pName
)
1714 int idx
= anv_get_device_entrypoint_index(pName
);
1718 return device
->dispatch
.entrypoints
[idx
];
1722 anv_CreateDebugReportCallbackEXT(VkInstance _instance
,
1723 const VkDebugReportCallbackCreateInfoEXT
* pCreateInfo
,
1724 const VkAllocationCallbacks
* pAllocator
,
1725 VkDebugReportCallbackEXT
* pCallback
)
1727 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1728 return vk_create_debug_report_callback(&instance
->debug_report_callbacks
,
1729 pCreateInfo
, pAllocator
, &instance
->alloc
,
1734 anv_DestroyDebugReportCallbackEXT(VkInstance _instance
,
1735 VkDebugReportCallbackEXT _callback
,
1736 const VkAllocationCallbacks
* pAllocator
)
1738 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1739 vk_destroy_debug_report_callback(&instance
->debug_report_callbacks
,
1740 _callback
, pAllocator
, &instance
->alloc
);
1744 anv_DebugReportMessageEXT(VkInstance _instance
,
1745 VkDebugReportFlagsEXT flags
,
1746 VkDebugReportObjectTypeEXT objectType
,
1749 int32_t messageCode
,
1750 const char* pLayerPrefix
,
1751 const char* pMessage
)
1753 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1754 vk_debug_report(&instance
->debug_report_callbacks
, flags
, objectType
,
1755 object
, location
, messageCode
, pLayerPrefix
, pMessage
);
1759 anv_queue_init(struct anv_device
*device
, struct anv_queue
*queue
)
1761 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1762 queue
->device
= device
;
1767 anv_queue_finish(struct anv_queue
*queue
)
1771 static struct anv_state
1772 anv_state_pool_emit_data(struct anv_state_pool
*pool
, size_t size
, size_t align
, const void *p
)
1774 struct anv_state state
;
1776 state
= anv_state_pool_alloc(pool
, size
, align
);
1777 memcpy(state
.map
, p
, size
);
1782 struct gen8_border_color
{
1787 /* Pad out to 64 bytes */
1792 anv_device_init_border_colors(struct anv_device
*device
)
1794 static const struct gen8_border_color border_colors
[] = {
1795 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
1796 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
1797 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
1798 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
1799 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
1800 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
1803 device
->border_colors
= anv_state_pool_emit_data(&device
->dynamic_state_pool
,
1804 sizeof(border_colors
), 64,
1809 anv_device_init_trivial_batch(struct anv_device
*device
)
1811 anv_bo_init_new(&device
->trivial_batch_bo
, device
, 4096);
1813 if (device
->instance
->physicalDevice
.has_exec_async
)
1814 device
->trivial_batch_bo
.flags
|= EXEC_OBJECT_ASYNC
;
1816 if (device
->instance
->physicalDevice
.use_softpin
)
1817 device
->trivial_batch_bo
.flags
|= EXEC_OBJECT_PINNED
;
1819 anv_vma_alloc(device
, &device
->trivial_batch_bo
);
1821 void *map
= anv_gem_mmap(device
, device
->trivial_batch_bo
.gem_handle
,
1824 struct anv_batch batch
= {
1830 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
1831 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
1833 if (!device
->info
.has_llc
)
1834 gen_clflush_range(map
, batch
.next
- map
);
1836 anv_gem_munmap(map
, device
->trivial_batch_bo
.size
);
1839 VkResult
anv_EnumerateDeviceExtensionProperties(
1840 VkPhysicalDevice physicalDevice
,
1841 const char* pLayerName
,
1842 uint32_t* pPropertyCount
,
1843 VkExtensionProperties
* pProperties
)
1845 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
1846 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
1848 for (int i
= 0; i
< ANV_DEVICE_EXTENSION_COUNT
; i
++) {
1849 if (device
->supported_extensions
.extensions
[i
]) {
1850 vk_outarray_append(&out
, prop
) {
1851 *prop
= anv_device_extensions
[i
];
1856 return vk_outarray_status(&out
);
1860 anv_device_init_dispatch(struct anv_device
*device
)
1862 const struct anv_device_dispatch_table
*genX_table
;
1863 switch (device
->info
.gen
) {
1865 genX_table
= &gen11_device_dispatch_table
;
1868 genX_table
= &gen10_device_dispatch_table
;
1871 genX_table
= &gen9_device_dispatch_table
;
1874 genX_table
= &gen8_device_dispatch_table
;
1877 if (device
->info
.is_haswell
)
1878 genX_table
= &gen75_device_dispatch_table
;
1880 genX_table
= &gen7_device_dispatch_table
;
1883 unreachable("unsupported gen\n");
1886 for (unsigned i
= 0; i
< ARRAY_SIZE(device
->dispatch
.entrypoints
); i
++) {
1887 /* Vulkan requires that entrypoints for extensions which have not been
1888 * enabled must not be advertised.
1890 if (!anv_device_entrypoint_is_enabled(i
, device
->instance
->app_info
.api_version
,
1891 &device
->instance
->enabled_extensions
,
1892 &device
->enabled_extensions
)) {
1893 device
->dispatch
.entrypoints
[i
] = NULL
;
1894 } else if (genX_table
->entrypoints
[i
]) {
1895 device
->dispatch
.entrypoints
[i
] = genX_table
->entrypoints
[i
];
1897 device
->dispatch
.entrypoints
[i
] =
1898 anv_device_dispatch_table
.entrypoints
[i
];
1904 vk_priority_to_gen(int priority
)
1907 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT
:
1908 return GEN_CONTEXT_LOW_PRIORITY
;
1909 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
:
1910 return GEN_CONTEXT_MEDIUM_PRIORITY
;
1911 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT
:
1912 return GEN_CONTEXT_HIGH_PRIORITY
;
1913 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT
:
1914 return GEN_CONTEXT_REALTIME_PRIORITY
;
1916 unreachable("Invalid priority");
1921 anv_device_init_hiz_clear_value_bo(struct anv_device
*device
)
1923 anv_bo_init_new(&device
->hiz_clear_bo
, device
, 4096);
1925 if (device
->instance
->physicalDevice
.has_exec_async
)
1926 device
->hiz_clear_bo
.flags
|= EXEC_OBJECT_ASYNC
;
1928 if (device
->instance
->physicalDevice
.use_softpin
)
1929 device
->hiz_clear_bo
.flags
|= EXEC_OBJECT_PINNED
;
1931 anv_vma_alloc(device
, &device
->hiz_clear_bo
);
1933 uint32_t *map
= anv_gem_mmap(device
, device
->hiz_clear_bo
.gem_handle
,
1936 union isl_color_value hiz_clear
= { .u32
= { 0, } };
1937 hiz_clear
.f32
[0] = ANV_HZ_FC_VAL
;
1939 memcpy(map
, hiz_clear
.u32
, sizeof(hiz_clear
.u32
));
1940 anv_gem_munmap(map
, device
->hiz_clear_bo
.size
);
1944 get_bo_from_pool(struct gen_batch_decode_bo
*ret
,
1945 struct anv_block_pool
*pool
,
1948 for (uint32_t i
= 0; i
< pool
->nbos
; i
++) {
1949 uint64_t bo_address
= pool
->bos
[i
].offset
& (~0ull >> 16);
1950 uint32_t bo_size
= pool
->bos
[i
].size
;
1951 if (address
>= bo_address
&& address
< (bo_address
+ bo_size
)) {
1952 *ret
= (struct gen_batch_decode_bo
) {
1955 .map
= pool
->bos
[i
].map
,
1963 /* Finding a buffer for batch decoding */
1964 static struct gen_batch_decode_bo
1965 decode_get_bo(void *v_batch
, bool ppgtt
, uint64_t address
)
1967 struct anv_device
*device
= v_batch
;
1968 struct gen_batch_decode_bo ret_bo
= {};
1972 if (get_bo_from_pool(&ret_bo
, &device
->dynamic_state_pool
.block_pool
, address
))
1974 if (get_bo_from_pool(&ret_bo
, &device
->instruction_state_pool
.block_pool
, address
))
1976 if (get_bo_from_pool(&ret_bo
, &device
->binding_table_pool
.block_pool
, address
))
1978 if (get_bo_from_pool(&ret_bo
, &device
->surface_state_pool
.block_pool
, address
))
1981 if (!device
->cmd_buffer_being_decoded
)
1982 return (struct gen_batch_decode_bo
) { };
1984 struct anv_batch_bo
**bo
;
1986 u_vector_foreach(bo
, &device
->cmd_buffer_being_decoded
->seen_bbos
) {
1987 /* The decoder zeroes out the top 16 bits, so we need to as well */
1988 uint64_t bo_address
= (*bo
)->bo
.offset
& (~0ull >> 16);
1990 if (address
>= bo_address
&& address
< bo_address
+ (*bo
)->bo
.size
) {
1991 return (struct gen_batch_decode_bo
) {
1993 .size
= (*bo
)->bo
.size
,
1994 .map
= (*bo
)->bo
.map
,
1999 return (struct gen_batch_decode_bo
) { };
2002 VkResult
anv_CreateDevice(
2003 VkPhysicalDevice physicalDevice
,
2004 const VkDeviceCreateInfo
* pCreateInfo
,
2005 const VkAllocationCallbacks
* pAllocator
,
2008 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
2010 struct anv_device
*device
;
2012 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO
);
2014 struct anv_device_extension_table enabled_extensions
= { };
2015 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
2017 for (idx
= 0; idx
< ANV_DEVICE_EXTENSION_COUNT
; idx
++) {
2018 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
2019 anv_device_extensions
[idx
].extensionName
) == 0)
2023 if (idx
>= ANV_DEVICE_EXTENSION_COUNT
)
2024 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
2026 if (!physical_device
->supported_extensions
.extensions
[idx
])
2027 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
2029 enabled_extensions
.extensions
[idx
] = true;
2032 /* Check enabled features */
2033 if (pCreateInfo
->pEnabledFeatures
) {
2034 VkPhysicalDeviceFeatures supported_features
;
2035 anv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
2036 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
2037 VkBool32
*enabled_feature
= (VkBool32
*)pCreateInfo
->pEnabledFeatures
;
2038 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
2039 for (uint32_t i
= 0; i
< num_features
; i
++) {
2040 if (enabled_feature
[i
] && !supported_feature
[i
])
2041 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
2045 /* Check requested queues and fail if we are requested to create any
2046 * queues with flags we don't support.
2048 assert(pCreateInfo
->queueCreateInfoCount
> 0);
2049 for (uint32_t i
= 0; i
< pCreateInfo
->queueCreateInfoCount
; i
++) {
2050 if (pCreateInfo
->pQueueCreateInfos
[i
].flags
!= 0)
2051 return vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2054 /* Check if client specified queue priority. */
2055 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*queue_priority
=
2056 vk_find_struct_const(pCreateInfo
->pQueueCreateInfos
[0].pNext
,
2057 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT
);
2059 VkQueueGlobalPriorityEXT priority
=
2060 queue_priority
? queue_priority
->globalPriority
:
2061 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
;
2063 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
2065 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
2067 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2069 const unsigned decode_flags
=
2070 GEN_BATCH_DECODE_FULL
|
2071 ((INTEL_DEBUG
& DEBUG_COLOR
) ? GEN_BATCH_DECODE_IN_COLOR
: 0) |
2072 GEN_BATCH_DECODE_OFFSETS
|
2073 GEN_BATCH_DECODE_FLOATS
;
2075 gen_batch_decode_ctx_init(&device
->decoder_ctx
,
2076 &physical_device
->info
,
2077 stderr
, decode_flags
, NULL
,
2078 decode_get_bo
, NULL
, device
);
2080 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
2081 device
->instance
= physical_device
->instance
;
2082 device
->chipset_id
= physical_device
->chipset_id
;
2083 device
->no_hw
= physical_device
->no_hw
;
2084 device
->_lost
= false;
2087 device
->alloc
= *pAllocator
;
2089 device
->alloc
= physical_device
->instance
->alloc
;
2091 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
2092 device
->fd
= open(physical_device
->path
, O_RDWR
| O_CLOEXEC
);
2093 if (device
->fd
== -1) {
2094 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2098 device
->context_id
= anv_gem_create_context(device
);
2099 if (device
->context_id
== -1) {
2100 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2104 if (physical_device
->use_softpin
) {
2105 if (pthread_mutex_init(&device
->vma_mutex
, NULL
) != 0) {
2106 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2110 /* keep the page with address zero out of the allocator */
2111 struct anv_memory_heap
*low_heap
=
2112 &physical_device
->memory
.heaps
[physical_device
->memory
.heap_count
- 1];
2113 util_vma_heap_init(&device
->vma_lo
, low_heap
->vma_start
, low_heap
->vma_size
);
2114 device
->vma_lo_available
= low_heap
->size
;
2116 struct anv_memory_heap
*high_heap
=
2117 &physical_device
->memory
.heaps
[0];
2118 util_vma_heap_init(&device
->vma_hi
, high_heap
->vma_start
, high_heap
->vma_size
);
2119 device
->vma_hi_available
= physical_device
->memory
.heap_count
== 1 ? 0 :
2123 list_inithead(&device
->memory_objects
);
2125 /* As per spec, the driver implementation may deny requests to acquire
2126 * a priority above the default priority (MEDIUM) if the caller does not
2127 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
2130 if (physical_device
->has_context_priority
) {
2131 int err
= anv_gem_set_context_param(device
->fd
, device
->context_id
,
2132 I915_CONTEXT_PARAM_PRIORITY
,
2133 vk_priority_to_gen(priority
));
2134 if (err
!= 0 && priority
> VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
) {
2135 result
= vk_error(VK_ERROR_NOT_PERMITTED_EXT
);
2140 device
->info
= physical_device
->info
;
2141 device
->isl_dev
= physical_device
->isl_dev
;
2143 /* On Broadwell and later, we can use batch chaining to more efficiently
2144 * implement growing command buffers. Prior to Haswell, the kernel
2145 * command parser gets in the way and we have to fall back to growing
2148 device
->can_chain_batches
= device
->info
.gen
>= 8;
2150 device
->robust_buffer_access
= pCreateInfo
->pEnabledFeatures
&&
2151 pCreateInfo
->pEnabledFeatures
->robustBufferAccess
;
2152 device
->enabled_extensions
= enabled_extensions
;
2154 anv_device_init_dispatch(device
);
2156 if (pthread_mutex_init(&device
->mutex
, NULL
) != 0) {
2157 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2158 goto fail_context_id
;
2161 pthread_condattr_t condattr
;
2162 if (pthread_condattr_init(&condattr
) != 0) {
2163 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2166 if (pthread_condattr_setclock(&condattr
, CLOCK_MONOTONIC
) != 0) {
2167 pthread_condattr_destroy(&condattr
);
2168 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2171 if (pthread_cond_init(&device
->queue_submit
, &condattr
) != 0) {
2172 pthread_condattr_destroy(&condattr
);
2173 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2176 pthread_condattr_destroy(&condattr
);
2179 (physical_device
->supports_48bit_addresses
? EXEC_OBJECT_SUPPORTS_48B_ADDRESS
: 0) |
2180 (physical_device
->has_exec_async
? EXEC_OBJECT_ASYNC
: 0) |
2181 (physical_device
->has_exec_capture
? EXEC_OBJECT_CAPTURE
: 0) |
2182 (physical_device
->use_softpin
? EXEC_OBJECT_PINNED
: 0);
2184 anv_bo_pool_init(&device
->batch_bo_pool
, device
, bo_flags
);
2186 result
= anv_bo_cache_init(&device
->bo_cache
);
2187 if (result
!= VK_SUCCESS
)
2188 goto fail_batch_bo_pool
;
2190 if (!physical_device
->use_softpin
)
2191 bo_flags
&= ~EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
2193 result
= anv_state_pool_init(&device
->dynamic_state_pool
, device
,
2194 DYNAMIC_STATE_POOL_MIN_ADDRESS
,
2197 if (result
!= VK_SUCCESS
)
2200 result
= anv_state_pool_init(&device
->instruction_state_pool
, device
,
2201 INSTRUCTION_STATE_POOL_MIN_ADDRESS
,
2204 if (result
!= VK_SUCCESS
)
2205 goto fail_dynamic_state_pool
;
2207 result
= anv_state_pool_init(&device
->surface_state_pool
, device
,
2208 SURFACE_STATE_POOL_MIN_ADDRESS
,
2211 if (result
!= VK_SUCCESS
)
2212 goto fail_instruction_state_pool
;
2214 if (physical_device
->use_softpin
) {
2215 result
= anv_state_pool_init(&device
->binding_table_pool
, device
,
2216 BINDING_TABLE_POOL_MIN_ADDRESS
,
2219 if (result
!= VK_SUCCESS
)
2220 goto fail_surface_state_pool
;
2223 result
= anv_bo_init_new(&device
->workaround_bo
, device
, 1024);
2224 if (result
!= VK_SUCCESS
)
2225 goto fail_binding_table_pool
;
2227 if (physical_device
->use_softpin
)
2228 device
->workaround_bo
.flags
|= EXEC_OBJECT_PINNED
;
2230 if (!anv_vma_alloc(device
, &device
->workaround_bo
))
2231 goto fail_workaround_bo
;
2233 anv_device_init_trivial_batch(device
);
2235 if (device
->info
.gen
>= 10)
2236 anv_device_init_hiz_clear_value_bo(device
);
2238 anv_scratch_pool_init(device
, &device
->scratch_pool
);
2240 anv_queue_init(device
, &device
->queue
);
2242 switch (device
->info
.gen
) {
2244 if (!device
->info
.is_haswell
)
2245 result
= gen7_init_device_state(device
);
2247 result
= gen75_init_device_state(device
);
2250 result
= gen8_init_device_state(device
);
2253 result
= gen9_init_device_state(device
);
2256 result
= gen10_init_device_state(device
);
2259 result
= gen11_init_device_state(device
);
2262 /* Shouldn't get here as we don't create physical devices for any other
2264 unreachable("unhandled gen");
2266 if (result
!= VK_SUCCESS
)
2267 goto fail_workaround_bo
;
2269 anv_pipeline_cache_init(&device
->default_pipeline_cache
, device
, true);
2271 anv_device_init_blorp(device
);
2273 anv_device_init_border_colors(device
);
2275 *pDevice
= anv_device_to_handle(device
);
2280 anv_queue_finish(&device
->queue
);
2281 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
2282 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
2283 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
2284 fail_binding_table_pool
:
2285 if (physical_device
->use_softpin
)
2286 anv_state_pool_finish(&device
->binding_table_pool
);
2287 fail_surface_state_pool
:
2288 anv_state_pool_finish(&device
->surface_state_pool
);
2289 fail_instruction_state_pool
:
2290 anv_state_pool_finish(&device
->instruction_state_pool
);
2291 fail_dynamic_state_pool
:
2292 anv_state_pool_finish(&device
->dynamic_state_pool
);
2294 anv_bo_cache_finish(&device
->bo_cache
);
2296 anv_bo_pool_finish(&device
->batch_bo_pool
);
2297 pthread_cond_destroy(&device
->queue_submit
);
2299 pthread_mutex_destroy(&device
->mutex
);
2301 anv_gem_destroy_context(device
, device
->context_id
);
2305 vk_free(&device
->alloc
, device
);
2310 void anv_DestroyDevice(
2312 const VkAllocationCallbacks
* pAllocator
)
2314 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2315 struct anv_physical_device
*physical_device
;
2320 physical_device
= &device
->instance
->physicalDevice
;
2322 anv_device_finish_blorp(device
);
2324 anv_pipeline_cache_finish(&device
->default_pipeline_cache
);
2326 anv_queue_finish(&device
->queue
);
2328 #ifdef HAVE_VALGRIND
2329 /* We only need to free these to prevent valgrind errors. The backing
2330 * BO will go away in a couple of lines so we don't actually leak.
2332 anv_state_pool_free(&device
->dynamic_state_pool
, device
->border_colors
);
2335 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
2337 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
2338 anv_vma_free(device
, &device
->workaround_bo
);
2339 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
2341 anv_vma_free(device
, &device
->trivial_batch_bo
);
2342 anv_gem_close(device
, device
->trivial_batch_bo
.gem_handle
);
2343 if (device
->info
.gen
>= 10)
2344 anv_gem_close(device
, device
->hiz_clear_bo
.gem_handle
);
2346 if (physical_device
->use_softpin
)
2347 anv_state_pool_finish(&device
->binding_table_pool
);
2348 anv_state_pool_finish(&device
->surface_state_pool
);
2349 anv_state_pool_finish(&device
->instruction_state_pool
);
2350 anv_state_pool_finish(&device
->dynamic_state_pool
);
2352 anv_bo_cache_finish(&device
->bo_cache
);
2354 anv_bo_pool_finish(&device
->batch_bo_pool
);
2356 pthread_cond_destroy(&device
->queue_submit
);
2357 pthread_mutex_destroy(&device
->mutex
);
2359 anv_gem_destroy_context(device
, device
->context_id
);
2361 gen_batch_decode_ctx_finish(&device
->decoder_ctx
);
2365 vk_free(&device
->alloc
, device
);
2368 VkResult
anv_EnumerateInstanceLayerProperties(
2369 uint32_t* pPropertyCount
,
2370 VkLayerProperties
* pProperties
)
2372 if (pProperties
== NULL
) {
2373 *pPropertyCount
= 0;
2377 /* None supported at this time */
2378 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
2381 VkResult
anv_EnumerateDeviceLayerProperties(
2382 VkPhysicalDevice physicalDevice
,
2383 uint32_t* pPropertyCount
,
2384 VkLayerProperties
* pProperties
)
2386 if (pProperties
== NULL
) {
2387 *pPropertyCount
= 0;
2391 /* None supported at this time */
2392 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
2395 void anv_GetDeviceQueue(
2397 uint32_t queueNodeIndex
,
2398 uint32_t queueIndex
,
2401 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2403 assert(queueIndex
== 0);
2405 *pQueue
= anv_queue_to_handle(&device
->queue
);
2408 void anv_GetDeviceQueue2(
2410 const VkDeviceQueueInfo2
* pQueueInfo
,
2413 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2415 assert(pQueueInfo
->queueIndex
== 0);
2417 if (pQueueInfo
->flags
== device
->queue
.flags
)
2418 *pQueue
= anv_queue_to_handle(&device
->queue
);
2424 _anv_device_set_lost(struct anv_device
*device
,
2425 const char *file
, int line
,
2426 const char *msg
, ...)
2431 device
->_lost
= true;
2434 err
= __vk_errorv(device
->instance
, device
,
2435 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT
,
2436 VK_ERROR_DEVICE_LOST
, file
, line
, msg
, ap
);
2439 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
2446 anv_device_query_status(struct anv_device
*device
)
2448 /* This isn't likely as most of the callers of this function already check
2449 * for it. However, it doesn't hurt to check and it potentially lets us
2452 if (anv_device_is_lost(device
))
2453 return VK_ERROR_DEVICE_LOST
;
2455 uint32_t active
, pending
;
2456 int ret
= anv_gem_gpu_get_reset_stats(device
, &active
, &pending
);
2458 /* We don't know the real error. */
2459 return anv_device_set_lost(device
, "get_reset_stats failed: %m");
2463 return anv_device_set_lost(device
, "GPU hung on one of our command buffers");
2464 } else if (pending
) {
2465 return anv_device_set_lost(device
, "GPU hung with commands in-flight");
2472 anv_device_bo_busy(struct anv_device
*device
, struct anv_bo
*bo
)
2474 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
2475 * Other usages of the BO (such as on different hardware) will not be
2476 * flagged as "busy" by this ioctl. Use with care.
2478 int ret
= anv_gem_busy(device
, bo
->gem_handle
);
2480 return VK_NOT_READY
;
2481 } else if (ret
== -1) {
2482 /* We don't know the real error. */
2483 return anv_device_set_lost(device
, "gem wait failed: %m");
2486 /* Query for device status after the busy call. If the BO we're checking
2487 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
2488 * client because it clearly doesn't have valid data. Yes, this most
2489 * likely means an ioctl, but we just did an ioctl to query the busy status
2490 * so it's no great loss.
2492 return anv_device_query_status(device
);
2496 anv_device_wait(struct anv_device
*device
, struct anv_bo
*bo
,
2499 int ret
= anv_gem_wait(device
, bo
->gem_handle
, &timeout
);
2500 if (ret
== -1 && errno
== ETIME
) {
2502 } else if (ret
== -1) {
2503 /* We don't know the real error. */
2504 return anv_device_set_lost(device
, "gem wait failed: %m");
2507 /* Query for device status after the wait. If the BO we're waiting on got
2508 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
2509 * because it clearly doesn't have valid data. Yes, this most likely means
2510 * an ioctl, but we just did an ioctl to wait so it's no great loss.
2512 return anv_device_query_status(device
);
2515 VkResult
anv_DeviceWaitIdle(
2518 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2519 if (anv_device_is_lost(device
))
2520 return VK_ERROR_DEVICE_LOST
;
2522 struct anv_batch batch
;
2525 batch
.start
= batch
.next
= cmds
;
2526 batch
.end
= (void *) cmds
+ sizeof(cmds
);
2528 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
2529 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
2531 return anv_device_submit_simple_batch(device
, &batch
);
2535 anv_vma_alloc(struct anv_device
*device
, struct anv_bo
*bo
)
2537 if (!(bo
->flags
& EXEC_OBJECT_PINNED
))
2540 pthread_mutex_lock(&device
->vma_mutex
);
2544 if (bo
->flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
&&
2545 device
->vma_hi_available
>= bo
->size
) {
2546 uint64_t addr
= util_vma_heap_alloc(&device
->vma_hi
, bo
->size
, 4096);
2548 bo
->offset
= gen_canonical_address(addr
);
2549 assert(addr
== gen_48b_address(bo
->offset
));
2550 device
->vma_hi_available
-= bo
->size
;
2554 if (bo
->offset
== 0 && device
->vma_lo_available
>= bo
->size
) {
2555 uint64_t addr
= util_vma_heap_alloc(&device
->vma_lo
, bo
->size
, 4096);
2557 bo
->offset
= gen_canonical_address(addr
);
2558 assert(addr
== gen_48b_address(bo
->offset
));
2559 device
->vma_lo_available
-= bo
->size
;
2563 pthread_mutex_unlock(&device
->vma_mutex
);
2565 return bo
->offset
!= 0;
2569 anv_vma_free(struct anv_device
*device
, struct anv_bo
*bo
)
2571 if (!(bo
->flags
& EXEC_OBJECT_PINNED
))
2574 const uint64_t addr_48b
= gen_48b_address(bo
->offset
);
2576 pthread_mutex_lock(&device
->vma_mutex
);
2578 if (addr_48b
>= LOW_HEAP_MIN_ADDRESS
&&
2579 addr_48b
<= LOW_HEAP_MAX_ADDRESS
) {
2580 util_vma_heap_free(&device
->vma_lo
, addr_48b
, bo
->size
);
2581 device
->vma_lo_available
+= bo
->size
;
2583 MAYBE_UNUSED
const struct anv_physical_device
*physical_device
=
2584 &device
->instance
->physicalDevice
;
2585 assert(addr_48b
>= physical_device
->memory
.heaps
[0].vma_start
&&
2586 addr_48b
< (physical_device
->memory
.heaps
[0].vma_start
+
2587 physical_device
->memory
.heaps
[0].vma_size
));
2588 util_vma_heap_free(&device
->vma_hi
, addr_48b
, bo
->size
);
2589 device
->vma_hi_available
+= bo
->size
;
2592 pthread_mutex_unlock(&device
->vma_mutex
);
2598 anv_bo_init_new(struct anv_bo
*bo
, struct anv_device
*device
, uint64_t size
)
2600 uint32_t gem_handle
= anv_gem_create(device
, size
);
2602 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
2604 anv_bo_init(bo
, gem_handle
, size
);
2609 VkResult
anv_AllocateMemory(
2611 const VkMemoryAllocateInfo
* pAllocateInfo
,
2612 const VkAllocationCallbacks
* pAllocator
,
2613 VkDeviceMemory
* pMem
)
2615 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2616 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2617 struct anv_device_memory
*mem
;
2618 VkResult result
= VK_SUCCESS
;
2620 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
2622 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
2623 assert(pAllocateInfo
->allocationSize
> 0);
2625 if (pAllocateInfo
->allocationSize
> MAX_MEMORY_ALLOCATION_SIZE
)
2626 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
2628 /* FINISHME: Fail if allocation request exceeds heap size. */
2630 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
2631 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2633 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2635 assert(pAllocateInfo
->memoryTypeIndex
< pdevice
->memory
.type_count
);
2636 mem
->type
= &pdevice
->memory
.types
[pAllocateInfo
->memoryTypeIndex
];
2640 mem
->host_ptr
= NULL
;
2642 uint64_t bo_flags
= 0;
2644 assert(mem
->type
->heapIndex
< pdevice
->memory
.heap_count
);
2645 if (pdevice
->memory
.heaps
[mem
->type
->heapIndex
].supports_48bit_addresses
)
2646 bo_flags
|= EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
2648 const struct wsi_memory_allocate_info
*wsi_info
=
2649 vk_find_struct_const(pAllocateInfo
->pNext
, WSI_MEMORY_ALLOCATE_INFO_MESA
);
2650 if (wsi_info
&& wsi_info
->implicit_sync
) {
2651 /* We need to set the WRITE flag on window system buffers so that GEM
2652 * will know we're writing to them and synchronize uses on other rings
2653 * (eg if the display server uses the blitter ring).
2655 bo_flags
|= EXEC_OBJECT_WRITE
;
2656 } else if (pdevice
->has_exec_async
) {
2657 bo_flags
|= EXEC_OBJECT_ASYNC
;
2660 if (pdevice
->use_softpin
)
2661 bo_flags
|= EXEC_OBJECT_PINNED
;
2663 const VkExportMemoryAllocateInfo
*export_info
=
2664 vk_find_struct_const(pAllocateInfo
->pNext
, EXPORT_MEMORY_ALLOCATE_INFO
);
2666 /* Check if we need to support Android HW buffer export. If so,
2667 * create AHardwareBuffer and import memory from it.
2669 bool android_export
= false;
2670 if (export_info
&& export_info
->handleTypes
&
2671 VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID
)
2672 android_export
= true;
2674 /* Android memory import. */
2675 const struct VkImportAndroidHardwareBufferInfoANDROID
*ahw_import_info
=
2676 vk_find_struct_const(pAllocateInfo
->pNext
,
2677 IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID
);
2679 if (ahw_import_info
) {
2680 result
= anv_import_ahw_memory(_device
, mem
, ahw_import_info
);
2681 if (result
!= VK_SUCCESS
)
2685 } else if (android_export
) {
2686 result
= anv_create_ahw_memory(_device
, mem
, pAllocateInfo
);
2687 if (result
!= VK_SUCCESS
)
2690 const struct VkImportAndroidHardwareBufferInfoANDROID import_info
= {
2693 result
= anv_import_ahw_memory(_device
, mem
, &import_info
);
2694 if (result
!= VK_SUCCESS
)
2700 const VkImportMemoryFdInfoKHR
*fd_info
=
2701 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_FD_INFO_KHR
);
2703 /* The Vulkan spec permits handleType to be 0, in which case the struct is
2706 if (fd_info
&& fd_info
->handleType
) {
2707 /* At the moment, we support only the below handle types. */
2708 assert(fd_info
->handleType
==
2709 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
2710 fd_info
->handleType
==
2711 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
2713 result
= anv_bo_cache_import(device
, &device
->bo_cache
, fd_info
->fd
,
2714 bo_flags
| ANV_BO_EXTERNAL
, &mem
->bo
);
2715 if (result
!= VK_SUCCESS
)
2718 VkDeviceSize aligned_alloc_size
=
2719 align_u64(pAllocateInfo
->allocationSize
, 4096);
2721 /* For security purposes, we reject importing the bo if it's smaller
2722 * than the requested allocation size. This prevents a malicious client
2723 * from passing a buffer to a trusted client, lying about the size, and
2724 * telling the trusted client to try and texture from an image that goes
2725 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
2726 * in the trusted client. The trusted client can protect itself against
2727 * this sort of attack but only if it can trust the buffer size.
2729 if (mem
->bo
->size
< aligned_alloc_size
) {
2730 result
= vk_errorf(device
->instance
, device
,
2731 VK_ERROR_INVALID_EXTERNAL_HANDLE
,
2732 "aligned allocationSize too large for "
2733 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT: "
2734 "%"PRIu64
"B > %"PRIu64
"B",
2735 aligned_alloc_size
, mem
->bo
->size
);
2736 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
2740 /* From the Vulkan spec:
2742 * "Importing memory from a file descriptor transfers ownership of
2743 * the file descriptor from the application to the Vulkan
2744 * implementation. The application must not perform any operations on
2745 * the file descriptor after a successful import."
2747 * If the import fails, we leave the file descriptor open.
2753 const VkImportMemoryHostPointerInfoEXT
*host_ptr_info
=
2754 vk_find_struct_const(pAllocateInfo
->pNext
,
2755 IMPORT_MEMORY_HOST_POINTER_INFO_EXT
);
2756 if (host_ptr_info
&& host_ptr_info
->handleType
) {
2757 if (host_ptr_info
->handleType
==
2758 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_MAPPED_FOREIGN_MEMORY_BIT_EXT
) {
2759 result
= vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
2763 assert(host_ptr_info
->handleType
==
2764 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
);
2766 result
= anv_bo_cache_import_host_ptr(
2767 device
, &device
->bo_cache
, host_ptr_info
->pHostPointer
,
2768 pAllocateInfo
->allocationSize
, bo_flags
, &mem
->bo
);
2770 if (result
!= VK_SUCCESS
)
2773 mem
->host_ptr
= host_ptr_info
->pHostPointer
;
2777 /* Regular allocate (not importing memory). */
2779 if (export_info
&& export_info
->handleTypes
)
2780 bo_flags
|= ANV_BO_EXTERNAL
;
2782 result
= anv_bo_cache_alloc(device
, &device
->bo_cache
,
2783 pAllocateInfo
->allocationSize
, bo_flags
,
2785 if (result
!= VK_SUCCESS
)
2788 const VkMemoryDedicatedAllocateInfo
*dedicated_info
=
2789 vk_find_struct_const(pAllocateInfo
->pNext
, MEMORY_DEDICATED_ALLOCATE_INFO
);
2790 if (dedicated_info
&& dedicated_info
->image
!= VK_NULL_HANDLE
) {
2791 ANV_FROM_HANDLE(anv_image
, image
, dedicated_info
->image
);
2793 /* Some legacy (non-modifiers) consumers need the tiling to be set on
2794 * the BO. In this case, we have a dedicated allocation.
2796 if (image
->needs_set_tiling
) {
2797 const uint32_t i915_tiling
=
2798 isl_tiling_to_i915_tiling(image
->planes
[0].surface
.isl
.tiling
);
2799 int ret
= anv_gem_set_tiling(device
, mem
->bo
->gem_handle
,
2800 image
->planes
[0].surface
.isl
.row_pitch_B
,
2803 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
2804 return vk_errorf(device
->instance
, NULL
,
2805 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
2806 "failed to set BO tiling: %m");
2812 pthread_mutex_lock(&device
->mutex
);
2813 list_addtail(&mem
->link
, &device
->memory_objects
);
2814 pthread_mutex_unlock(&device
->mutex
);
2816 *pMem
= anv_device_memory_to_handle(mem
);
2821 vk_free2(&device
->alloc
, pAllocator
, mem
);
2826 VkResult
anv_GetMemoryFdKHR(
2828 const VkMemoryGetFdInfoKHR
* pGetFdInfo
,
2831 ANV_FROM_HANDLE(anv_device
, dev
, device_h
);
2832 ANV_FROM_HANDLE(anv_device_memory
, mem
, pGetFdInfo
->memory
);
2834 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
2836 assert(pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
2837 pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
2839 return anv_bo_cache_export(dev
, &dev
->bo_cache
, mem
->bo
, pFd
);
2842 VkResult
anv_GetMemoryFdPropertiesKHR(
2844 VkExternalMemoryHandleTypeFlagBits handleType
,
2846 VkMemoryFdPropertiesKHR
* pMemoryFdProperties
)
2848 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2849 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2851 switch (handleType
) {
2852 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
:
2853 /* dma-buf can be imported as any memory type */
2854 pMemoryFdProperties
->memoryTypeBits
=
2855 (1 << pdevice
->memory
.type_count
) - 1;
2859 /* The valid usage section for this function says:
2861 * "handleType must not be one of the handle types defined as
2864 * So opaque handle types fall into the default "unsupported" case.
2866 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
2870 VkResult
anv_GetMemoryHostPointerPropertiesEXT(
2872 VkExternalMemoryHandleTypeFlagBits handleType
,
2873 const void* pHostPointer
,
2874 VkMemoryHostPointerPropertiesEXT
* pMemoryHostPointerProperties
)
2876 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2878 assert(pMemoryHostPointerProperties
->sType
==
2879 VK_STRUCTURE_TYPE_MEMORY_HOST_POINTER_PROPERTIES_EXT
);
2881 switch (handleType
) {
2882 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
: {
2883 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2885 /* Host memory can be imported as any memory type. */
2886 pMemoryHostPointerProperties
->memoryTypeBits
=
2887 (1ull << pdevice
->memory
.type_count
) - 1;
2892 return VK_ERROR_INVALID_EXTERNAL_HANDLE
;
2896 void anv_FreeMemory(
2898 VkDeviceMemory _mem
,
2899 const VkAllocationCallbacks
* pAllocator
)
2901 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2902 ANV_FROM_HANDLE(anv_device_memory
, mem
, _mem
);
2907 pthread_mutex_lock(&device
->mutex
);
2908 list_del(&mem
->link
);
2909 pthread_mutex_unlock(&device
->mutex
);
2912 anv_UnmapMemory(_device
, _mem
);
2914 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
2916 #if defined(ANDROID) && ANDROID_API_LEVEL >= 26
2918 AHardwareBuffer_release(mem
->ahw
);
2921 vk_free2(&device
->alloc
, pAllocator
, mem
);
2924 VkResult
anv_MapMemory(
2926 VkDeviceMemory _memory
,
2927 VkDeviceSize offset
,
2929 VkMemoryMapFlags flags
,
2932 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2933 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
2940 if (mem
->host_ptr
) {
2941 *ppData
= mem
->host_ptr
+ offset
;
2945 if (size
== VK_WHOLE_SIZE
)
2946 size
= mem
->bo
->size
- offset
;
2948 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
2950 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
2951 * assert(size != 0);
2952 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
2953 * equal to the size of the memory minus offset
2956 assert(offset
+ size
<= mem
->bo
->size
);
2958 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
2959 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
2960 * at a time is valid. We could just mmap up front and return an offset
2961 * pointer here, but that may exhaust virtual memory on 32 bit
2964 uint32_t gem_flags
= 0;
2966 if (!device
->info
.has_llc
&&
2967 (mem
->type
->propertyFlags
& VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
))
2968 gem_flags
|= I915_MMAP_WC
;
2970 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
2971 uint64_t map_offset
= offset
& ~4095ull;
2972 assert(offset
>= map_offset
);
2973 uint64_t map_size
= (offset
+ size
) - map_offset
;
2975 /* Let's map whole pages */
2976 map_size
= align_u64(map_size
, 4096);
2978 void *map
= anv_gem_mmap(device
, mem
->bo
->gem_handle
,
2979 map_offset
, map_size
, gem_flags
);
2980 if (map
== MAP_FAILED
)
2981 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
2984 mem
->map_size
= map_size
;
2986 *ppData
= mem
->map
+ (offset
- map_offset
);
2991 void anv_UnmapMemory(
2993 VkDeviceMemory _memory
)
2995 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
2997 if (mem
== NULL
|| mem
->host_ptr
)
3000 anv_gem_munmap(mem
->map
, mem
->map_size
);
3007 clflush_mapped_ranges(struct anv_device
*device
,
3009 const VkMappedMemoryRange
*ranges
)
3011 for (uint32_t i
= 0; i
< count
; i
++) {
3012 ANV_FROM_HANDLE(anv_device_memory
, mem
, ranges
[i
].memory
);
3013 if (ranges
[i
].offset
>= mem
->map_size
)
3016 gen_clflush_range(mem
->map
+ ranges
[i
].offset
,
3017 MIN2(ranges
[i
].size
, mem
->map_size
- ranges
[i
].offset
));
3021 VkResult
anv_FlushMappedMemoryRanges(
3023 uint32_t memoryRangeCount
,
3024 const VkMappedMemoryRange
* pMemoryRanges
)
3026 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3028 if (device
->info
.has_llc
)
3031 /* Make sure the writes we're flushing have landed. */
3032 __builtin_ia32_mfence();
3034 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
3039 VkResult
anv_InvalidateMappedMemoryRanges(
3041 uint32_t memoryRangeCount
,
3042 const VkMappedMemoryRange
* pMemoryRanges
)
3044 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3046 if (device
->info
.has_llc
)
3049 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
3051 /* Make sure no reads get moved up above the invalidate. */
3052 __builtin_ia32_mfence();
3057 void anv_GetBufferMemoryRequirements(
3060 VkMemoryRequirements
* pMemoryRequirements
)
3062 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
3063 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3064 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3066 /* The Vulkan spec (git aaed022) says:
3068 * memoryTypeBits is a bitfield and contains one bit set for every
3069 * supported memory type for the resource. The bit `1<<i` is set if and
3070 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3071 * structure for the physical device is supported.
3073 uint32_t memory_types
= 0;
3074 for (uint32_t i
= 0; i
< pdevice
->memory
.type_count
; i
++) {
3075 uint32_t valid_usage
= pdevice
->memory
.types
[i
].valid_buffer_usage
;
3076 if ((valid_usage
& buffer
->usage
) == buffer
->usage
)
3077 memory_types
|= (1u << i
);
3080 /* Base alignment requirement of a cache line */
3081 uint32_t alignment
= 16;
3083 /* We need an alignment of 32 for pushing UBOs */
3084 if (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
)
3085 alignment
= MAX2(alignment
, 32);
3087 pMemoryRequirements
->size
= buffer
->size
;
3088 pMemoryRequirements
->alignment
= alignment
;
3090 /* Storage and Uniform buffers should have their size aligned to
3091 * 32-bits to avoid boundary checks when last DWord is not complete.
3092 * This would ensure that not internal padding would be needed for
3095 if (device
->robust_buffer_access
&&
3096 (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
||
3097 buffer
->usage
& VK_BUFFER_USAGE_STORAGE_BUFFER_BIT
))
3098 pMemoryRequirements
->size
= align_u64(buffer
->size
, 4);
3100 pMemoryRequirements
->memoryTypeBits
= memory_types
;
3103 void anv_GetBufferMemoryRequirements2(
3105 const VkBufferMemoryRequirementsInfo2
* pInfo
,
3106 VkMemoryRequirements2
* pMemoryRequirements
)
3108 anv_GetBufferMemoryRequirements(_device
, pInfo
->buffer
,
3109 &pMemoryRequirements
->memoryRequirements
);
3111 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3112 switch (ext
->sType
) {
3113 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
3114 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
3115 requirements
->prefersDedicatedAllocation
= false;
3116 requirements
->requiresDedicatedAllocation
= false;
3121 anv_debug_ignored_stype(ext
->sType
);
3127 void anv_GetImageMemoryRequirements(
3130 VkMemoryRequirements
* pMemoryRequirements
)
3132 ANV_FROM_HANDLE(anv_image
, image
, _image
);
3133 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3134 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3136 /* The Vulkan spec (git aaed022) says:
3138 * memoryTypeBits is a bitfield and contains one bit set for every
3139 * supported memory type for the resource. The bit `1<<i` is set if and
3140 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3141 * structure for the physical device is supported.
3143 * All types are currently supported for images.
3145 uint32_t memory_types
= (1ull << pdevice
->memory
.type_count
) - 1;
3147 /* We must have image allocated or imported at this point. According to the
3148 * specification, external images must have been bound to memory before
3149 * calling GetImageMemoryRequirements.
3151 assert(image
->size
> 0);
3153 pMemoryRequirements
->size
= image
->size
;
3154 pMemoryRequirements
->alignment
= image
->alignment
;
3155 pMemoryRequirements
->memoryTypeBits
= memory_types
;
3158 void anv_GetImageMemoryRequirements2(
3160 const VkImageMemoryRequirementsInfo2
* pInfo
,
3161 VkMemoryRequirements2
* pMemoryRequirements
)
3163 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3164 ANV_FROM_HANDLE(anv_image
, image
, pInfo
->image
);
3166 anv_GetImageMemoryRequirements(_device
, pInfo
->image
,
3167 &pMemoryRequirements
->memoryRequirements
);
3169 vk_foreach_struct_const(ext
, pInfo
->pNext
) {
3170 switch (ext
->sType
) {
3171 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO
: {
3172 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3173 const VkImagePlaneMemoryRequirementsInfo
*plane_reqs
=
3174 (const VkImagePlaneMemoryRequirementsInfo
*) ext
;
3175 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
,
3176 plane_reqs
->planeAspect
);
3178 assert(image
->planes
[plane
].offset
== 0);
3180 /* The Vulkan spec (git aaed022) says:
3182 * memoryTypeBits is a bitfield and contains one bit set for every
3183 * supported memory type for the resource. The bit `1<<i` is set
3184 * if and only if the memory type `i` in the
3185 * VkPhysicalDeviceMemoryProperties structure for the physical
3186 * device is supported.
3188 * All types are currently supported for images.
3190 pMemoryRequirements
->memoryRequirements
.memoryTypeBits
=
3191 (1ull << pdevice
->memory
.type_count
) - 1;
3193 /* We must have image allocated or imported at this point. According to the
3194 * specification, external images must have been bound to memory before
3195 * calling GetImageMemoryRequirements.
3197 assert(image
->planes
[plane
].size
> 0);
3199 pMemoryRequirements
->memoryRequirements
.size
= image
->planes
[plane
].size
;
3200 pMemoryRequirements
->memoryRequirements
.alignment
=
3201 image
->planes
[plane
].alignment
;
3206 anv_debug_ignored_stype(ext
->sType
);
3211 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3212 switch (ext
->sType
) {
3213 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
3214 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
3215 if (image
->needs_set_tiling
|| image
->external_format
) {
3216 /* If we need to set the tiling for external consumers, we need a
3217 * dedicated allocation.
3219 * See also anv_AllocateMemory.
3221 requirements
->prefersDedicatedAllocation
= true;
3222 requirements
->requiresDedicatedAllocation
= true;
3224 requirements
->prefersDedicatedAllocation
= false;
3225 requirements
->requiresDedicatedAllocation
= false;
3231 anv_debug_ignored_stype(ext
->sType
);
3237 void anv_GetImageSparseMemoryRequirements(
3240 uint32_t* pSparseMemoryRequirementCount
,
3241 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
3243 *pSparseMemoryRequirementCount
= 0;
3246 void anv_GetImageSparseMemoryRequirements2(
3248 const VkImageSparseMemoryRequirementsInfo2
* pInfo
,
3249 uint32_t* pSparseMemoryRequirementCount
,
3250 VkSparseImageMemoryRequirements2
* pSparseMemoryRequirements
)
3252 *pSparseMemoryRequirementCount
= 0;
3255 void anv_GetDeviceMemoryCommitment(
3257 VkDeviceMemory memory
,
3258 VkDeviceSize
* pCommittedMemoryInBytes
)
3260 *pCommittedMemoryInBytes
= 0;
3264 anv_bind_buffer_memory(const VkBindBufferMemoryInfo
*pBindInfo
)
3266 ANV_FROM_HANDLE(anv_device_memory
, mem
, pBindInfo
->memory
);
3267 ANV_FROM_HANDLE(anv_buffer
, buffer
, pBindInfo
->buffer
);
3269 assert(pBindInfo
->sType
== VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
);
3272 assert((buffer
->usage
& mem
->type
->valid_buffer_usage
) == buffer
->usage
);
3273 buffer
->address
= (struct anv_address
) {
3275 .offset
= pBindInfo
->memoryOffset
,
3278 buffer
->address
= ANV_NULL_ADDRESS
;
3282 VkResult
anv_BindBufferMemory(
3285 VkDeviceMemory memory
,
3286 VkDeviceSize memoryOffset
)
3288 anv_bind_buffer_memory(
3289 &(VkBindBufferMemoryInfo
) {
3290 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
,
3293 .memoryOffset
= memoryOffset
,
3299 VkResult
anv_BindBufferMemory2(
3301 uint32_t bindInfoCount
,
3302 const VkBindBufferMemoryInfo
* pBindInfos
)
3304 for (uint32_t i
= 0; i
< bindInfoCount
; i
++)
3305 anv_bind_buffer_memory(&pBindInfos
[i
]);
3310 VkResult
anv_QueueBindSparse(
3312 uint32_t bindInfoCount
,
3313 const VkBindSparseInfo
* pBindInfo
,
3316 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
3317 if (anv_device_is_lost(queue
->device
))
3318 return VK_ERROR_DEVICE_LOST
;
3320 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
3325 VkResult
anv_CreateEvent(
3327 const VkEventCreateInfo
* pCreateInfo
,
3328 const VkAllocationCallbacks
* pAllocator
,
3331 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3332 struct anv_state state
;
3333 struct anv_event
*event
;
3335 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_EVENT_CREATE_INFO
);
3337 state
= anv_state_pool_alloc(&device
->dynamic_state_pool
,
3340 event
->state
= state
;
3341 event
->semaphore
= VK_EVENT_RESET
;
3343 if (!device
->info
.has_llc
) {
3344 /* Make sure the writes we're flushing have landed. */
3345 __builtin_ia32_mfence();
3346 __builtin_ia32_clflush(event
);
3349 *pEvent
= anv_event_to_handle(event
);
3354 void anv_DestroyEvent(
3357 const VkAllocationCallbacks
* pAllocator
)
3359 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3360 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3365 anv_state_pool_free(&device
->dynamic_state_pool
, event
->state
);
3368 VkResult
anv_GetEventStatus(
3372 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3373 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3375 if (anv_device_is_lost(device
))
3376 return VK_ERROR_DEVICE_LOST
;
3378 if (!device
->info
.has_llc
) {
3379 /* Invalidate read cache before reading event written by GPU. */
3380 __builtin_ia32_clflush(event
);
3381 __builtin_ia32_mfence();
3385 return event
->semaphore
;
3388 VkResult
anv_SetEvent(
3392 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3393 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3395 event
->semaphore
= VK_EVENT_SET
;
3397 if (!device
->info
.has_llc
) {
3398 /* Make sure the writes we're flushing have landed. */
3399 __builtin_ia32_mfence();
3400 __builtin_ia32_clflush(event
);
3406 VkResult
anv_ResetEvent(
3410 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3411 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3413 event
->semaphore
= VK_EVENT_RESET
;
3415 if (!device
->info
.has_llc
) {
3416 /* Make sure the writes we're flushing have landed. */
3417 __builtin_ia32_mfence();
3418 __builtin_ia32_clflush(event
);
3426 VkResult
anv_CreateBuffer(
3428 const VkBufferCreateInfo
* pCreateInfo
,
3429 const VkAllocationCallbacks
* pAllocator
,
3432 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3433 struct anv_buffer
*buffer
;
3435 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
3437 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
3438 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3440 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3442 buffer
->size
= pCreateInfo
->size
;
3443 buffer
->usage
= pCreateInfo
->usage
;
3444 buffer
->address
= ANV_NULL_ADDRESS
;
3446 *pBuffer
= anv_buffer_to_handle(buffer
);
3451 void anv_DestroyBuffer(
3454 const VkAllocationCallbacks
* pAllocator
)
3456 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3457 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
3462 vk_free2(&device
->alloc
, pAllocator
, buffer
);
3465 VkDeviceAddress
anv_GetBufferDeviceAddressEXT(
3467 const VkBufferDeviceAddressInfoEXT
* pInfo
)
3469 ANV_FROM_HANDLE(anv_buffer
, buffer
, pInfo
->buffer
);
3471 assert(buffer
->address
.bo
->flags
& EXEC_OBJECT_PINNED
);
3473 return anv_address_physical(buffer
->address
);
3477 anv_fill_buffer_surface_state(struct anv_device
*device
, struct anv_state state
,
3478 enum isl_format format
,
3479 struct anv_address address
,
3480 uint32_t range
, uint32_t stride
)
3482 isl_buffer_fill_state(&device
->isl_dev
, state
.map
,
3483 .address
= anv_address_physical(address
),
3484 .mocs
= device
->default_mocs
,
3487 .swizzle
= ISL_SWIZZLE_IDENTITY
,
3488 .stride_B
= stride
);
3491 void anv_DestroySampler(
3494 const VkAllocationCallbacks
* pAllocator
)
3496 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3497 ANV_FROM_HANDLE(anv_sampler
, sampler
, _sampler
);
3502 if (sampler
->bindless_state
.map
) {
3503 anv_state_pool_free(&device
->dynamic_state_pool
,
3504 sampler
->bindless_state
);
3507 vk_free2(&device
->alloc
, pAllocator
, sampler
);
3510 VkResult
anv_CreateFramebuffer(
3512 const VkFramebufferCreateInfo
* pCreateInfo
,
3513 const VkAllocationCallbacks
* pAllocator
,
3514 VkFramebuffer
* pFramebuffer
)
3516 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3517 struct anv_framebuffer
*framebuffer
;
3519 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
3521 size_t size
= sizeof(*framebuffer
) +
3522 sizeof(struct anv_image_view
*) * pCreateInfo
->attachmentCount
;
3523 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
3524 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3525 if (framebuffer
== NULL
)
3526 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3528 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
3529 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
3530 VkImageView _iview
= pCreateInfo
->pAttachments
[i
];
3531 framebuffer
->attachments
[i
] = anv_image_view_from_handle(_iview
);
3534 framebuffer
->width
= pCreateInfo
->width
;
3535 framebuffer
->height
= pCreateInfo
->height
;
3536 framebuffer
->layers
= pCreateInfo
->layers
;
3538 *pFramebuffer
= anv_framebuffer_to_handle(framebuffer
);
3543 void anv_DestroyFramebuffer(
3546 const VkAllocationCallbacks
* pAllocator
)
3548 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3549 ANV_FROM_HANDLE(anv_framebuffer
, fb
, _fb
);
3554 vk_free2(&device
->alloc
, pAllocator
, fb
);
3557 static const VkTimeDomainEXT anv_time_domains
[] = {
3558 VK_TIME_DOMAIN_DEVICE_EXT
,
3559 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
,
3560 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
,
3563 VkResult
anv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
3564 VkPhysicalDevice physicalDevice
,
3565 uint32_t *pTimeDomainCount
,
3566 VkTimeDomainEXT
*pTimeDomains
)
3569 VK_OUTARRAY_MAKE(out
, pTimeDomains
, pTimeDomainCount
);
3571 for (d
= 0; d
< ARRAY_SIZE(anv_time_domains
); d
++) {
3572 vk_outarray_append(&out
, i
) {
3573 *i
= anv_time_domains
[d
];
3577 return vk_outarray_status(&out
);
3581 anv_clock_gettime(clockid_t clock_id
)
3583 struct timespec current
;
3586 ret
= clock_gettime(clock_id
, ¤t
);
3587 if (ret
< 0 && clock_id
== CLOCK_MONOTONIC_RAW
)
3588 ret
= clock_gettime(CLOCK_MONOTONIC
, ¤t
);
3592 return (uint64_t) current
.tv_sec
* 1000000000ULL + current
.tv_nsec
;
3595 #define TIMESTAMP 0x2358
3597 VkResult
anv_GetCalibratedTimestampsEXT(
3599 uint32_t timestampCount
,
3600 const VkCalibratedTimestampInfoEXT
*pTimestampInfos
,
3601 uint64_t *pTimestamps
,
3602 uint64_t *pMaxDeviation
)
3604 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3605 uint64_t timestamp_frequency
= device
->info
.timestamp_frequency
;
3608 uint64_t begin
, end
;
3609 uint64_t max_clock_period
= 0;
3611 begin
= anv_clock_gettime(CLOCK_MONOTONIC_RAW
);
3613 for (d
= 0; d
< timestampCount
; d
++) {
3614 switch (pTimestampInfos
[d
].timeDomain
) {
3615 case VK_TIME_DOMAIN_DEVICE_EXT
:
3616 ret
= anv_gem_reg_read(device
, TIMESTAMP
| 1,
3620 return anv_device_set_lost(device
, "Failed to read the TIMESTAMP "
3623 uint64_t device_period
= DIV_ROUND_UP(1000000000, timestamp_frequency
);
3624 max_clock_period
= MAX2(max_clock_period
, device_period
);
3626 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
:
3627 pTimestamps
[d
] = anv_clock_gettime(CLOCK_MONOTONIC
);
3628 max_clock_period
= MAX2(max_clock_period
, 1);
3631 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
:
3632 pTimestamps
[d
] = begin
;
3640 end
= anv_clock_gettime(CLOCK_MONOTONIC_RAW
);
3643 * The maximum deviation is the sum of the interval over which we
3644 * perform the sampling and the maximum period of any sampled
3645 * clock. That's because the maximum skew between any two sampled
3646 * clock edges is when the sampled clock with the largest period is
3647 * sampled at the end of that period but right at the beginning of the
3648 * sampling interval and some other clock is sampled right at the
3649 * begining of its sampling period and right at the end of the
3650 * sampling interval. Let's assume the GPU has the longest clock
3651 * period and that the application is sampling GPU and monotonic:
3654 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
3655 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
3659 * GPU -----_____-----_____-----_____-----_____
3662 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
3663 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
3665 * Interval <----------------->
3666 * Deviation <-------------------------->
3670 * m = read(monotonic) 2
3673 * We round the sample interval up by one tick to cover sampling error
3674 * in the interval clock
3677 uint64_t sample_interval
= end
- begin
+ 1;
3679 *pMaxDeviation
= sample_interval
+ max_clock_period
;
3684 /* vk_icd.h does not declare this function, so we declare it here to
3685 * suppress Wmissing-prototypes.
3687 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
3688 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
);
3690 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
3691 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
)
3693 /* For the full details on loader interface versioning, see
3694 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
3695 * What follows is a condensed summary, to help you navigate the large and
3696 * confusing official doc.
3698 * - Loader interface v0 is incompatible with later versions. We don't
3701 * - In loader interface v1:
3702 * - The first ICD entrypoint called by the loader is
3703 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
3705 * - The ICD must statically expose no other Vulkan symbol unless it is
3706 * linked with -Bsymbolic.
3707 * - Each dispatchable Vulkan handle created by the ICD must be
3708 * a pointer to a struct whose first member is VK_LOADER_DATA. The
3709 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
3710 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
3711 * vkDestroySurfaceKHR(). The ICD must be capable of working with
3712 * such loader-managed surfaces.
3714 * - Loader interface v2 differs from v1 in:
3715 * - The first ICD entrypoint called by the loader is
3716 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
3717 * statically expose this entrypoint.
3719 * - Loader interface v3 differs from v2 in:
3720 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
3721 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
3722 * because the loader no longer does so.
3724 *pSupportedVersion
= MIN2(*pSupportedVersion
, 3u);