2 * Copyright © 2015 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
28 #include <sys/sysinfo.h>
32 #include "drm-uapi/drm_fourcc.h"
34 #include "anv_private.h"
35 #include "util/debug.h"
36 #include "util/build_id.h"
37 #include "util/disk_cache.h"
38 #include "util/mesa-sha1.h"
39 #include "util/os_file.h"
40 #include "util/u_atomic.h"
41 #include "util/u_string.h"
42 #include "util/xmlpool.h"
45 #include "common/gen_aux_map.h"
46 #include "common/gen_defines.h"
47 #include "compiler/glsl_types.h"
49 #include "genxml/gen7_pack.h"
51 static const char anv_dri_options_xml
[] =
53 DRI_CONF_SECTION_PERFORMANCE
54 DRI_CONF_VK_X11_OVERRIDE_MIN_IMAGE_COUNT(0)
55 DRI_CONF_VK_X11_STRICT_IMAGE_COUNT("false")
58 DRI_CONF_SECTION_DEBUG
59 DRI_CONF_ALWAYS_FLUSH_CACHE("false")
63 /* This is probably far to big but it reflects the max size used for messages
64 * in OpenGLs KHR_debug.
66 #define MAX_DEBUG_MESSAGE_LENGTH 4096
69 compiler_debug_log(void *data
, const char *fmt
, ...)
71 char str
[MAX_DEBUG_MESSAGE_LENGTH
];
72 struct anv_device
*device
= (struct anv_device
*)data
;
74 if (list_is_empty(&device
->instance
->debug_report_callbacks
.callbacks
))
79 (void) vsnprintf(str
, MAX_DEBUG_MESSAGE_LENGTH
, fmt
, args
);
82 vk_debug_report(&device
->instance
->debug_report_callbacks
,
83 VK_DEBUG_REPORT_DEBUG_BIT_EXT
,
84 VK_DEBUG_REPORT_OBJECT_TYPE_UNKNOWN_EXT
,
89 compiler_perf_log(void *data
, const char *fmt
, ...)
94 if (unlikely(INTEL_DEBUG
& DEBUG_PERF
))
95 intel_logd_v(fmt
, args
);
101 anv_compute_heap_size(int fd
, uint64_t gtt_size
)
103 /* Query the total ram from the system */
107 uint64_t total_ram
= (uint64_t)info
.totalram
* (uint64_t)info
.mem_unit
;
109 /* We don't want to burn too much ram with the GPU. If the user has 4GiB
110 * or less, we use at most half. If they have more than 4GiB, we use 3/4.
112 uint64_t available_ram
;
113 if (total_ram
<= 4ull * 1024ull * 1024ull * 1024ull)
114 available_ram
= total_ram
/ 2;
116 available_ram
= total_ram
* 3 / 4;
118 /* We also want to leave some padding for things we allocate in the driver,
119 * so don't go over 3/4 of the GTT either.
121 uint64_t available_gtt
= gtt_size
* 3 / 4;
123 return MIN2(available_ram
, available_gtt
);
127 anv_physical_device_init_heaps(struct anv_physical_device
*device
, int fd
)
129 if (anv_gem_get_context_param(fd
, 0, I915_CONTEXT_PARAM_GTT_SIZE
,
130 &device
->gtt_size
) == -1) {
131 /* If, for whatever reason, we can't actually get the GTT size from the
132 * kernel (too old?) fall back to the aperture size.
134 anv_perf_warn(NULL
, NULL
,
135 "Failed to get I915_CONTEXT_PARAM_GTT_SIZE: %m");
137 if (anv_gem_get_aperture(fd
, &device
->gtt_size
) == -1) {
138 return vk_errorf(NULL
, NULL
, VK_ERROR_INITIALIZATION_FAILED
,
139 "failed to get aperture size: %m");
143 /* We only allow 48-bit addresses with softpin because knowing the actual
144 * address is required for the vertex cache flush workaround.
146 device
->supports_48bit_addresses
= (device
->info
.gen
>= 8) &&
147 device
->has_softpin
&&
148 device
->gtt_size
> (4ULL << 30 /* GiB */);
150 uint64_t heap_size
= anv_compute_heap_size(fd
, device
->gtt_size
);
152 if (heap_size
> (2ull << 30) && !device
->supports_48bit_addresses
) {
153 /* When running with an overridden PCI ID, we may get a GTT size from
154 * the kernel that is greater than 2 GiB but the execbuf check for 48bit
155 * address support can still fail. Just clamp the address space size to
156 * 2 GiB if we don't have 48-bit support.
158 intel_logw("%s:%d: The kernel reported a GTT size larger than 2 GiB but "
159 "not support for 48-bit addresses",
161 heap_size
= 2ull << 30;
164 if (heap_size
<= 3ull * (1ull << 30)) {
165 /* In this case, everything fits nicely into the 32-bit address space,
166 * so there's no need for supporting 48bit addresses on client-allocated
169 device
->memory
.heap_count
= 1;
170 device
->memory
.heaps
[0] = (struct anv_memory_heap
) {
172 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
173 .supports_48bit_addresses
= false,
176 /* Not everything will fit nicely into a 32-bit address space. In this
177 * case we need a 64-bit heap. Advertise a small 32-bit heap and a
178 * larger 48-bit heap. If we're in this case, then we have a total heap
179 * size larger than 3GiB which most likely means they have 8 GiB of
180 * video memory and so carving off 1 GiB for the 32-bit heap should be
183 const uint64_t heap_size_32bit
= 1ull << 30;
184 const uint64_t heap_size_48bit
= heap_size
- heap_size_32bit
;
186 assert(device
->supports_48bit_addresses
);
188 device
->memory
.heap_count
= 2;
189 device
->memory
.heaps
[0] = (struct anv_memory_heap
) {
190 .size
= heap_size_48bit
,
191 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
192 .supports_48bit_addresses
= true,
194 device
->memory
.heaps
[1] = (struct anv_memory_heap
) {
195 .size
= heap_size_32bit
,
196 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
197 .supports_48bit_addresses
= false,
201 uint32_t type_count
= 0;
202 for (uint32_t heap
= 0; heap
< device
->memory
.heap_count
; heap
++) {
203 uint32_t valid_buffer_usage
= ~0;
205 /* There appears to be a hardware issue in the VF cache where it only
206 * considers the bottom 32 bits of memory addresses. If you happen to
207 * have two vertex buffers which get placed exactly 4 GiB apart and use
208 * them in back-to-back draw calls, you can get collisions. In order to
209 * solve this problem, we require vertex and index buffers be bound to
210 * memory allocated out of the 32-bit heap.
212 if (device
->memory
.heaps
[heap
].supports_48bit_addresses
) {
213 valid_buffer_usage
&= ~(VK_BUFFER_USAGE_INDEX_BUFFER_BIT
|
214 VK_BUFFER_USAGE_VERTEX_BUFFER_BIT
);
217 if (device
->info
.has_llc
) {
218 /* Big core GPUs share LLC with the CPU and thus one memory type can be
219 * both cached and coherent at the same time.
221 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
222 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
223 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
224 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
225 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
227 .valid_buffer_usage
= valid_buffer_usage
,
230 /* The spec requires that we expose a host-visible, coherent memory
231 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
232 * to give the application a choice between cached, but not coherent and
233 * coherent but uncached (WC though).
235 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
236 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
237 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
238 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
240 .valid_buffer_usage
= valid_buffer_usage
,
242 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
243 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
244 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
245 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
247 .valid_buffer_usage
= valid_buffer_usage
,
251 device
->memory
.type_count
= type_count
;
257 anv_physical_device_init_uuids(struct anv_physical_device
*device
)
259 const struct build_id_note
*note
=
260 build_id_find_nhdr_for_addr(anv_physical_device_init_uuids
);
262 return vk_errorf(device
->instance
, device
,
263 VK_ERROR_INITIALIZATION_FAILED
,
264 "Failed to find build-id");
267 unsigned build_id_len
= build_id_length(note
);
268 if (build_id_len
< 20) {
269 return vk_errorf(device
->instance
, device
,
270 VK_ERROR_INITIALIZATION_FAILED
,
271 "build-id too short. It needs to be a SHA");
274 memcpy(device
->driver_build_sha1
, build_id_data(note
), 20);
276 struct mesa_sha1 sha1_ctx
;
278 STATIC_ASSERT(VK_UUID_SIZE
<= sizeof(sha1
));
280 /* The pipeline cache UUID is used for determining when a pipeline cache is
281 * invalid. It needs both a driver build and the PCI ID of the device.
283 _mesa_sha1_init(&sha1_ctx
);
284 _mesa_sha1_update(&sha1_ctx
, build_id_data(note
), build_id_len
);
285 _mesa_sha1_update(&sha1_ctx
, &device
->chipset_id
,
286 sizeof(device
->chipset_id
));
287 _mesa_sha1_update(&sha1_ctx
, &device
->always_use_bindless
,
288 sizeof(device
->always_use_bindless
));
289 _mesa_sha1_update(&sha1_ctx
, &device
->has_a64_buffer_access
,
290 sizeof(device
->has_a64_buffer_access
));
291 _mesa_sha1_update(&sha1_ctx
, &device
->has_bindless_images
,
292 sizeof(device
->has_bindless_images
));
293 _mesa_sha1_update(&sha1_ctx
, &device
->has_bindless_samplers
,
294 sizeof(device
->has_bindless_samplers
));
295 _mesa_sha1_final(&sha1_ctx
, sha1
);
296 memcpy(device
->pipeline_cache_uuid
, sha1
, VK_UUID_SIZE
);
298 /* The driver UUID is used for determining sharability of images and memory
299 * between two Vulkan instances in separate processes. People who want to
300 * share memory need to also check the device UUID (below) so all this
301 * needs to be is the build-id.
303 memcpy(device
->driver_uuid
, build_id_data(note
), VK_UUID_SIZE
);
305 /* The device UUID uniquely identifies the given device within the machine.
306 * Since we never have more than one device, this doesn't need to be a real
307 * UUID. However, on the off-chance that someone tries to use this to
308 * cache pre-tiled images or something of the like, we use the PCI ID and
309 * some bits of ISL info to ensure that this is safe.
311 _mesa_sha1_init(&sha1_ctx
);
312 _mesa_sha1_update(&sha1_ctx
, &device
->chipset_id
,
313 sizeof(device
->chipset_id
));
314 _mesa_sha1_update(&sha1_ctx
, &device
->isl_dev
.has_bit6_swizzling
,
315 sizeof(device
->isl_dev
.has_bit6_swizzling
));
316 _mesa_sha1_final(&sha1_ctx
, sha1
);
317 memcpy(device
->device_uuid
, sha1
, VK_UUID_SIZE
);
323 anv_physical_device_init_disk_cache(struct anv_physical_device
*device
)
325 #ifdef ENABLE_SHADER_CACHE
327 ASSERTED
int len
= snprintf(renderer
, sizeof(renderer
), "anv_%04x",
329 assert(len
== sizeof(renderer
) - 2);
332 _mesa_sha1_format(timestamp
, device
->driver_build_sha1
);
334 const uint64_t driver_flags
=
335 brw_get_compiler_config_value(device
->compiler
);
336 device
->disk_cache
= disk_cache_create(renderer
, timestamp
, driver_flags
);
338 device
->disk_cache
= NULL
;
343 anv_physical_device_free_disk_cache(struct anv_physical_device
*device
)
345 #ifdef ENABLE_SHADER_CACHE
346 if (device
->disk_cache
)
347 disk_cache_destroy(device
->disk_cache
);
349 assert(device
->disk_cache
== NULL
);
354 get_available_system_memory()
356 char *meminfo
= os_read_file("/proc/meminfo");
360 char *str
= strstr(meminfo
, "MemAvailable:");
366 uint64_t kb_mem_available
;
367 if (sscanf(str
, "MemAvailable: %" PRIx64
, &kb_mem_available
) == 1) {
369 return kb_mem_available
<< 10;
377 anv_physical_device_init(struct anv_physical_device
*device
,
378 struct anv_instance
*instance
,
379 drmDevicePtr drm_device
)
381 const char *primary_path
= drm_device
->nodes
[DRM_NODE_PRIMARY
];
382 const char *path
= drm_device
->nodes
[DRM_NODE_RENDER
];
387 brw_process_intel_debug_variable();
389 fd
= open(path
, O_RDWR
| O_CLOEXEC
);
391 return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
393 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
394 device
->instance
= instance
;
396 assert(strlen(path
) < ARRAY_SIZE(device
->path
));
397 snprintf(device
->path
, ARRAY_SIZE(device
->path
), "%s", path
);
399 if (!gen_get_device_info_from_fd(fd
, &device
->info
)) {
400 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
403 device
->chipset_id
= device
->info
.chipset_id
;
404 device
->no_hw
= device
->info
.no_hw
;
406 if (getenv("INTEL_NO_HW") != NULL
)
407 device
->no_hw
= true;
409 device
->pci_info
.domain
= drm_device
->businfo
.pci
->domain
;
410 device
->pci_info
.bus
= drm_device
->businfo
.pci
->bus
;
411 device
->pci_info
.device
= drm_device
->businfo
.pci
->dev
;
412 device
->pci_info
.function
= drm_device
->businfo
.pci
->func
;
414 device
->name
= gen_get_device_name(device
->chipset_id
);
416 if (device
->info
.is_haswell
) {
417 intel_logw("Haswell Vulkan support is incomplete");
418 } else if (device
->info
.gen
== 7 && !device
->info
.is_baytrail
) {
419 intel_logw("Ivy Bridge Vulkan support is incomplete");
420 } else if (device
->info
.gen
== 7 && device
->info
.is_baytrail
) {
421 intel_logw("Bay Trail Vulkan support is incomplete");
422 } else if (device
->info
.gen
>= 8 && device
->info
.gen
<= 11) {
423 /* Gen8-11 fully supported */
424 } else if (device
->info
.gen
== 12) {
425 intel_logw("Vulkan is not yet fully supported on gen12");
427 result
= vk_errorf(device
->instance
, device
,
428 VK_ERROR_INCOMPATIBLE_DRIVER
,
429 "Vulkan not yet supported on %s", device
->name
);
433 device
->cmd_parser_version
= -1;
434 if (device
->info
.gen
== 7) {
435 device
->cmd_parser_version
=
436 anv_gem_get_param(fd
, I915_PARAM_CMD_PARSER_VERSION
);
437 if (device
->cmd_parser_version
== -1) {
438 result
= vk_errorf(device
->instance
, device
,
439 VK_ERROR_INITIALIZATION_FAILED
,
440 "failed to get command parser version");
445 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_WAIT_TIMEOUT
)) {
446 result
= vk_errorf(device
->instance
, device
,
447 VK_ERROR_INITIALIZATION_FAILED
,
448 "kernel missing gem wait");
452 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_EXECBUF2
)) {
453 result
= vk_errorf(device
->instance
, device
,
454 VK_ERROR_INITIALIZATION_FAILED
,
455 "kernel missing execbuf2");
459 if (!device
->info
.has_llc
&&
460 anv_gem_get_param(fd
, I915_PARAM_MMAP_VERSION
) < 1) {
461 result
= vk_errorf(device
->instance
, device
,
462 VK_ERROR_INITIALIZATION_FAILED
,
463 "kernel missing wc mmap");
467 device
->has_softpin
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_SOFTPIN
);
468 device
->has_exec_async
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_ASYNC
);
469 device
->has_exec_capture
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_CAPTURE
);
470 device
->has_exec_fence
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE
);
471 device
->has_syncobj
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE_ARRAY
);
472 device
->has_syncobj_wait
= device
->has_syncobj
&&
473 anv_gem_supports_syncobj_wait(fd
);
474 device
->has_context_priority
= anv_gem_has_context_priority(fd
);
476 result
= anv_physical_device_init_heaps(device
, fd
);
477 if (result
!= VK_SUCCESS
)
480 device
->use_softpin
= device
->has_softpin
&&
481 device
->supports_48bit_addresses
;
483 device
->has_context_isolation
=
484 anv_gem_get_param(fd
, I915_PARAM_HAS_CONTEXT_ISOLATION
);
486 device
->always_use_bindless
=
487 env_var_as_boolean("ANV_ALWAYS_BINDLESS", false);
489 /* We first got the A64 messages on broadwell and we can only use them if
490 * we can pass addresses directly into the shader which requires softpin.
492 device
->has_a64_buffer_access
= device
->info
.gen
>= 8 &&
495 /* We first get bindless image access on Skylake and we can only really do
496 * it if we don't have any relocations so we need softpin.
498 device
->has_bindless_images
= device
->info
.gen
>= 9 &&
501 /* We've had bindless samplers since Ivy Bridge (forever in Vulkan terms)
502 * because it's just a matter of setting the sampler address in the sample
503 * message header. However, we've not bothered to wire it up for vec4 so
504 * we leave it disabled on gen7.
506 device
->has_bindless_samplers
= device
->info
.gen
>= 8;
508 device
->has_mem_available
= get_available_system_memory() != 0;
510 device
->always_flush_cache
=
511 driQueryOptionb(&instance
->dri_options
, "always_flush_cache");
513 /* Starting with Gen10, the timestamp frequency of the command streamer may
514 * vary from one part to another. We can query the value from the kernel.
516 if (device
->info
.gen
>= 10) {
517 int timestamp_frequency
=
518 anv_gem_get_param(fd
, I915_PARAM_CS_TIMESTAMP_FREQUENCY
);
520 if (timestamp_frequency
< 0)
521 intel_logw("Kernel 4.16-rc1+ required to properly query CS timestamp frequency");
523 device
->info
.timestamp_frequency
= timestamp_frequency
;
526 /* GENs prior to 8 do not support EU/Subslice info */
527 if (device
->info
.gen
>= 8) {
528 device
->subslice_total
= anv_gem_get_param(fd
, I915_PARAM_SUBSLICE_TOTAL
);
529 device
->eu_total
= anv_gem_get_param(fd
, I915_PARAM_EU_TOTAL
);
531 /* Without this information, we cannot get the right Braswell
532 * brandstrings, and we have to use conservative numbers for GPGPU on
533 * many platforms, but otherwise, things will just work.
535 if (device
->subslice_total
< 1 || device
->eu_total
< 1) {
536 intel_logw("Kernel 4.1 required to properly query GPU properties");
538 } else if (device
->info
.gen
== 7) {
539 device
->subslice_total
= 1 << (device
->info
.gt
- 1);
542 if (device
->info
.is_cherryview
&&
543 device
->subslice_total
> 0 && device
->eu_total
> 0) {
544 /* Logical CS threads = EUs per subslice * num threads per EU */
545 uint32_t max_cs_threads
=
546 device
->eu_total
/ device
->subslice_total
* device
->info
.num_thread_per_eu
;
548 /* Fuse configurations may give more threads than expected, never less. */
549 if (max_cs_threads
> device
->info
.max_cs_threads
)
550 device
->info
.max_cs_threads
= max_cs_threads
;
553 device
->compiler
= brw_compiler_create(NULL
, &device
->info
);
554 if (device
->compiler
== NULL
) {
555 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
558 device
->compiler
->shader_debug_log
= compiler_debug_log
;
559 device
->compiler
->shader_perf_log
= compiler_perf_log
;
560 device
->compiler
->supports_pull_constants
= false;
561 device
->compiler
->constant_buffer_0_is_relative
=
562 device
->info
.gen
< 8 || !device
->has_context_isolation
;
563 device
->compiler
->supports_shader_constants
= true;
564 device
->compiler
->compact_params
= false;
566 /* Broadwell PRM says:
568 * "Before Gen8, there was a historical configuration control field to
569 * swizzle address bit[6] for in X/Y tiling modes. This was set in three
570 * different places: TILECTL[1:0], ARB_MODE[5:4], and
571 * DISP_ARB_CTL[14:13].
573 * For Gen8 and subsequent generations, the swizzle fields are all
574 * reserved, and the CPU's memory controller performs all address
575 * swizzling modifications."
578 device
->info
.gen
< 8 && anv_gem_get_bit6_swizzle(fd
, I915_TILING_X
);
580 isl_device_init(&device
->isl_dev
, &device
->info
, swizzled
);
582 result
= anv_physical_device_init_uuids(device
);
583 if (result
!= VK_SUCCESS
)
586 anv_physical_device_init_disk_cache(device
);
588 if (instance
->enabled_extensions
.KHR_display
) {
589 master_fd
= open(primary_path
, O_RDWR
| O_CLOEXEC
);
590 if (master_fd
>= 0) {
591 /* prod the device with a GETPARAM call which will fail if
592 * we don't have permission to even render on this device
594 if (anv_gem_get_param(master_fd
, I915_PARAM_CHIPSET_ID
) == 0) {
600 device
->master_fd
= master_fd
;
602 result
= anv_init_wsi(device
);
603 if (result
!= VK_SUCCESS
) {
604 ralloc_free(device
->compiler
);
605 anv_physical_device_free_disk_cache(device
);
609 device
->perf
= anv_get_perf(&device
->info
, fd
);
611 anv_physical_device_get_supported_extensions(device
,
612 &device
->supported_extensions
);
615 device
->local_fd
= fd
;
627 anv_physical_device_finish(struct anv_physical_device
*device
)
629 anv_finish_wsi(device
);
630 anv_physical_device_free_disk_cache(device
);
631 ralloc_free(device
->compiler
);
632 ralloc_free(device
->perf
);
633 close(device
->local_fd
);
634 if (device
->master_fd
>= 0)
635 close(device
->master_fd
);
639 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
640 VkSystemAllocationScope allocationScope
)
646 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
647 size_t align
, VkSystemAllocationScope allocationScope
)
649 return realloc(pOriginal
, size
);
653 default_free_func(void *pUserData
, void *pMemory
)
658 static const VkAllocationCallbacks default_alloc
= {
660 .pfnAllocation
= default_alloc_func
,
661 .pfnReallocation
= default_realloc_func
,
662 .pfnFree
= default_free_func
,
665 VkResult
anv_EnumerateInstanceExtensionProperties(
666 const char* pLayerName
,
667 uint32_t* pPropertyCount
,
668 VkExtensionProperties
* pProperties
)
670 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
672 for (int i
= 0; i
< ANV_INSTANCE_EXTENSION_COUNT
; i
++) {
673 if (anv_instance_extensions_supported
.extensions
[i
]) {
674 vk_outarray_append(&out
, prop
) {
675 *prop
= anv_instance_extensions
[i
];
680 return vk_outarray_status(&out
);
683 VkResult
anv_CreateInstance(
684 const VkInstanceCreateInfo
* pCreateInfo
,
685 const VkAllocationCallbacks
* pAllocator
,
686 VkInstance
* pInstance
)
688 struct anv_instance
*instance
;
691 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
693 struct anv_instance_extension_table enabled_extensions
= {};
694 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
696 for (idx
= 0; idx
< ANV_INSTANCE_EXTENSION_COUNT
; idx
++) {
697 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
698 anv_instance_extensions
[idx
].extensionName
) == 0)
702 if (idx
>= ANV_INSTANCE_EXTENSION_COUNT
)
703 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
705 if (!anv_instance_extensions_supported
.extensions
[idx
])
706 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
708 enabled_extensions
.extensions
[idx
] = true;
711 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
712 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
714 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
716 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
719 instance
->alloc
= *pAllocator
;
721 instance
->alloc
= default_alloc
;
723 instance
->app_info
= (struct anv_app_info
) { .api_version
= 0 };
724 if (pCreateInfo
->pApplicationInfo
) {
725 const VkApplicationInfo
*app
= pCreateInfo
->pApplicationInfo
;
727 instance
->app_info
.app_name
=
728 vk_strdup(&instance
->alloc
, app
->pApplicationName
,
729 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
730 instance
->app_info
.app_version
= app
->applicationVersion
;
732 instance
->app_info
.engine_name
=
733 vk_strdup(&instance
->alloc
, app
->pEngineName
,
734 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
735 instance
->app_info
.engine_version
= app
->engineVersion
;
737 instance
->app_info
.api_version
= app
->apiVersion
;
740 if (instance
->app_info
.api_version
== 0)
741 instance
->app_info
.api_version
= VK_API_VERSION_1_0
;
743 instance
->enabled_extensions
= enabled_extensions
;
745 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->dispatch
.entrypoints
); i
++) {
746 /* Vulkan requires that entrypoints for extensions which have not been
747 * enabled must not be advertised.
749 if (!anv_instance_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
750 &instance
->enabled_extensions
)) {
751 instance
->dispatch
.entrypoints
[i
] = NULL
;
753 instance
->dispatch
.entrypoints
[i
] =
754 anv_instance_dispatch_table
.entrypoints
[i
];
758 struct anv_physical_device
*pdevice
= &instance
->physicalDevice
;
759 for (unsigned i
= 0; i
< ARRAY_SIZE(pdevice
->dispatch
.entrypoints
); i
++) {
760 /* Vulkan requires that entrypoints for extensions which have not been
761 * enabled must not be advertised.
763 if (!anv_physical_device_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
764 &instance
->enabled_extensions
)) {
765 pdevice
->dispatch
.entrypoints
[i
] = NULL
;
767 pdevice
->dispatch
.entrypoints
[i
] =
768 anv_physical_device_dispatch_table
.entrypoints
[i
];
772 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->device_dispatch
.entrypoints
); i
++) {
773 /* Vulkan requires that entrypoints for extensions which have not been
774 * enabled must not be advertised.
776 if (!anv_device_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
777 &instance
->enabled_extensions
, NULL
)) {
778 instance
->device_dispatch
.entrypoints
[i
] = NULL
;
780 instance
->device_dispatch
.entrypoints
[i
] =
781 anv_device_dispatch_table
.entrypoints
[i
];
785 instance
->physicalDeviceCount
= -1;
787 result
= vk_debug_report_instance_init(&instance
->debug_report_callbacks
);
788 if (result
!= VK_SUCCESS
) {
789 vk_free2(&default_alloc
, pAllocator
, instance
);
790 return vk_error(result
);
793 instance
->pipeline_cache_enabled
=
794 env_var_as_boolean("ANV_ENABLE_PIPELINE_CACHE", true);
796 glsl_type_singleton_init_or_ref();
798 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
800 driParseOptionInfo(&instance
->available_dri_options
, anv_dri_options_xml
);
801 driParseConfigFiles(&instance
->dri_options
, &instance
->available_dri_options
,
803 instance
->app_info
.engine_name
,
804 instance
->app_info
.engine_version
);
806 *pInstance
= anv_instance_to_handle(instance
);
811 void anv_DestroyInstance(
812 VkInstance _instance
,
813 const VkAllocationCallbacks
* pAllocator
)
815 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
820 if (instance
->physicalDeviceCount
> 0) {
821 /* We support at most one physical device. */
822 assert(instance
->physicalDeviceCount
== 1);
823 anv_physical_device_finish(&instance
->physicalDevice
);
826 vk_free(&instance
->alloc
, (char *)instance
->app_info
.app_name
);
827 vk_free(&instance
->alloc
, (char *)instance
->app_info
.engine_name
);
829 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
831 vk_debug_report_instance_destroy(&instance
->debug_report_callbacks
);
833 glsl_type_singleton_decref();
835 driDestroyOptionCache(&instance
->dri_options
);
836 driDestroyOptionInfo(&instance
->available_dri_options
);
838 vk_free(&instance
->alloc
, instance
);
842 anv_enumerate_devices(struct anv_instance
*instance
)
844 /* TODO: Check for more devices ? */
845 drmDevicePtr devices
[8];
846 VkResult result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
849 instance
->physicalDeviceCount
= 0;
851 max_devices
= drmGetDevices2(0, devices
, ARRAY_SIZE(devices
));
853 return VK_ERROR_INCOMPATIBLE_DRIVER
;
855 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
856 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
857 devices
[i
]->bustype
== DRM_BUS_PCI
&&
858 devices
[i
]->deviceinfo
.pci
->vendor_id
== 0x8086) {
860 result
= anv_physical_device_init(&instance
->physicalDevice
,
861 instance
, devices
[i
]);
862 if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
866 drmFreeDevices(devices
, max_devices
);
868 if (result
== VK_SUCCESS
)
869 instance
->physicalDeviceCount
= 1;
875 anv_instance_ensure_physical_device(struct anv_instance
*instance
)
877 if (instance
->physicalDeviceCount
< 0) {
878 VkResult result
= anv_enumerate_devices(instance
);
879 if (result
!= VK_SUCCESS
&&
880 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
887 VkResult
anv_EnumeratePhysicalDevices(
888 VkInstance _instance
,
889 uint32_t* pPhysicalDeviceCount
,
890 VkPhysicalDevice
* pPhysicalDevices
)
892 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
893 VK_OUTARRAY_MAKE(out
, pPhysicalDevices
, pPhysicalDeviceCount
);
895 VkResult result
= anv_instance_ensure_physical_device(instance
);
896 if (result
!= VK_SUCCESS
)
899 if (instance
->physicalDeviceCount
== 0)
902 assert(instance
->physicalDeviceCount
== 1);
903 vk_outarray_append(&out
, i
) {
904 *i
= anv_physical_device_to_handle(&instance
->physicalDevice
);
907 return vk_outarray_status(&out
);
910 VkResult
anv_EnumeratePhysicalDeviceGroups(
911 VkInstance _instance
,
912 uint32_t* pPhysicalDeviceGroupCount
,
913 VkPhysicalDeviceGroupProperties
* pPhysicalDeviceGroupProperties
)
915 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
916 VK_OUTARRAY_MAKE(out
, pPhysicalDeviceGroupProperties
,
917 pPhysicalDeviceGroupCount
);
919 VkResult result
= anv_instance_ensure_physical_device(instance
);
920 if (result
!= VK_SUCCESS
)
923 if (instance
->physicalDeviceCount
== 0)
926 assert(instance
->physicalDeviceCount
== 1);
928 vk_outarray_append(&out
, p
) {
929 p
->physicalDeviceCount
= 1;
930 memset(p
->physicalDevices
, 0, sizeof(p
->physicalDevices
));
931 p
->physicalDevices
[0] =
932 anv_physical_device_to_handle(&instance
->physicalDevice
);
933 p
->subsetAllocation
= false;
935 vk_foreach_struct(ext
, p
->pNext
)
936 anv_debug_ignored_stype(ext
->sType
);
939 return vk_outarray_status(&out
);
942 void anv_GetPhysicalDeviceFeatures(
943 VkPhysicalDevice physicalDevice
,
944 VkPhysicalDeviceFeatures
* pFeatures
)
946 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
948 *pFeatures
= (VkPhysicalDeviceFeatures
) {
949 .robustBufferAccess
= true,
950 .fullDrawIndexUint32
= true,
951 .imageCubeArray
= true,
952 .independentBlend
= true,
953 .geometryShader
= true,
954 .tessellationShader
= true,
955 .sampleRateShading
= true,
956 .dualSrcBlend
= true,
958 .multiDrawIndirect
= true,
959 .drawIndirectFirstInstance
= true,
961 .depthBiasClamp
= true,
962 .fillModeNonSolid
= true,
963 .depthBounds
= pdevice
->info
.gen
>= 12,
967 .multiViewport
= true,
968 .samplerAnisotropy
= true,
969 .textureCompressionETC2
= pdevice
->info
.gen
>= 8 ||
970 pdevice
->info
.is_baytrail
,
971 .textureCompressionASTC_LDR
= pdevice
->info
.gen
>= 9, /* FINISHME CHV */
972 .textureCompressionBC
= true,
973 .occlusionQueryPrecise
= true,
974 .pipelineStatisticsQuery
= true,
975 .fragmentStoresAndAtomics
= true,
976 .shaderTessellationAndGeometryPointSize
= true,
977 .shaderImageGatherExtended
= true,
978 .shaderStorageImageExtendedFormats
= true,
979 .shaderStorageImageMultisample
= false,
980 .shaderStorageImageReadWithoutFormat
= false,
981 .shaderStorageImageWriteWithoutFormat
= true,
982 .shaderUniformBufferArrayDynamicIndexing
= true,
983 .shaderSampledImageArrayDynamicIndexing
= true,
984 .shaderStorageBufferArrayDynamicIndexing
= true,
985 .shaderStorageImageArrayDynamicIndexing
= true,
986 .shaderClipDistance
= true,
987 .shaderCullDistance
= true,
988 .shaderFloat64
= pdevice
->info
.gen
>= 8 &&
989 pdevice
->info
.has_64bit_types
,
990 .shaderInt64
= pdevice
->info
.gen
>= 8 &&
991 pdevice
->info
.has_64bit_types
,
992 .shaderInt16
= pdevice
->info
.gen
>= 8,
993 .shaderResourceMinLod
= pdevice
->info
.gen
>= 9,
994 .variableMultisampleRate
= true,
995 .inheritedQueries
= true,
998 /* We can't do image stores in vec4 shaders */
999 pFeatures
->vertexPipelineStoresAndAtomics
=
1000 pdevice
->compiler
->scalar_stage
[MESA_SHADER_VERTEX
] &&
1001 pdevice
->compiler
->scalar_stage
[MESA_SHADER_GEOMETRY
];
1003 struct anv_app_info
*app_info
= &pdevice
->instance
->app_info
;
1005 /* The new DOOM and Wolfenstein games require depthBounds without
1006 * checking for it. They seem to run fine without it so just claim it's
1007 * there and accept the consequences.
1009 if (app_info
->engine_name
&& strcmp(app_info
->engine_name
, "idTech") == 0)
1010 pFeatures
->depthBounds
= true;
1013 void anv_GetPhysicalDeviceFeatures2(
1014 VkPhysicalDevice physicalDevice
,
1015 VkPhysicalDeviceFeatures2
* pFeatures
)
1017 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1018 anv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
1020 vk_foreach_struct(ext
, pFeatures
->pNext
) {
1021 switch (ext
->sType
) {
1022 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES_KHR
: {
1023 VkPhysicalDevice8BitStorageFeaturesKHR
*features
=
1024 (VkPhysicalDevice8BitStorageFeaturesKHR
*)ext
;
1025 features
->storageBuffer8BitAccess
= pdevice
->info
.gen
>= 8;
1026 features
->uniformAndStorageBuffer8BitAccess
= pdevice
->info
.gen
>= 8;
1027 features
->storagePushConstant8
= pdevice
->info
.gen
>= 8;
1031 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES
: {
1032 VkPhysicalDevice16BitStorageFeatures
*features
=
1033 (VkPhysicalDevice16BitStorageFeatures
*)ext
;
1034 features
->storageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
1035 features
->uniformAndStorageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
1036 features
->storagePushConstant16
= pdevice
->info
.gen
>= 8;
1037 features
->storageInputOutput16
= false;
1041 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_EXT
: {
1042 VkPhysicalDeviceBufferDeviceAddressFeaturesEXT
*features
= (void *)ext
;
1043 features
->bufferDeviceAddress
= pdevice
->has_a64_buffer_access
;
1044 features
->bufferDeviceAddressCaptureReplay
= false;
1045 features
->bufferDeviceAddressMultiDevice
= false;
1049 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COMPUTE_SHADER_DERIVATIVES_FEATURES_NV
: {
1050 VkPhysicalDeviceComputeShaderDerivativesFeaturesNV
*features
=
1051 (VkPhysicalDeviceComputeShaderDerivativesFeaturesNV
*)ext
;
1052 features
->computeDerivativeGroupQuads
= true;
1053 features
->computeDerivativeGroupLinear
= true;
1057 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT
: {
1058 VkPhysicalDeviceConditionalRenderingFeaturesEXT
*features
=
1059 (VkPhysicalDeviceConditionalRenderingFeaturesEXT
*)ext
;
1060 features
->conditionalRendering
= pdevice
->info
.gen
>= 8 ||
1061 pdevice
->info
.is_haswell
;
1062 features
->inheritedConditionalRendering
= pdevice
->info
.gen
>= 8 ||
1063 pdevice
->info
.is_haswell
;
1067 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_CLIP_ENABLE_FEATURES_EXT
: {
1068 VkPhysicalDeviceDepthClipEnableFeaturesEXT
*features
=
1069 (VkPhysicalDeviceDepthClipEnableFeaturesEXT
*)ext
;
1070 features
->depthClipEnable
= true;
1074 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR
: {
1075 VkPhysicalDeviceFloat16Int8FeaturesKHR
*features
= (void *)ext
;
1076 features
->shaderFloat16
= pdevice
->info
.gen
>= 8;
1077 features
->shaderInt8
= pdevice
->info
.gen
>= 8;
1081 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FRAGMENT_SHADER_INTERLOCK_FEATURES_EXT
: {
1082 VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT
*features
=
1083 (VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT
*)ext
;
1084 features
->fragmentShaderSampleInterlock
= pdevice
->info
.gen
>= 9;
1085 features
->fragmentShaderPixelInterlock
= pdevice
->info
.gen
>= 9;
1086 features
->fragmentShaderShadingRateInterlock
= false;
1090 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_QUERY_RESET_FEATURES_EXT
: {
1091 VkPhysicalDeviceHostQueryResetFeaturesEXT
*features
=
1092 (VkPhysicalDeviceHostQueryResetFeaturesEXT
*)ext
;
1093 features
->hostQueryReset
= true;
1097 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT
: {
1098 VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*features
=
1099 (VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*)ext
;
1100 features
->shaderInputAttachmentArrayDynamicIndexing
= false;
1101 features
->shaderUniformTexelBufferArrayDynamicIndexing
= true;
1102 features
->shaderStorageTexelBufferArrayDynamicIndexing
= true;
1103 features
->shaderUniformBufferArrayNonUniformIndexing
= false;
1104 features
->shaderSampledImageArrayNonUniformIndexing
= true;
1105 features
->shaderStorageBufferArrayNonUniformIndexing
= true;
1106 features
->shaderStorageImageArrayNonUniformIndexing
= true;
1107 features
->shaderInputAttachmentArrayNonUniformIndexing
= false;
1108 features
->shaderUniformTexelBufferArrayNonUniformIndexing
= true;
1109 features
->shaderStorageTexelBufferArrayNonUniformIndexing
= true;
1110 features
->descriptorBindingUniformBufferUpdateAfterBind
= false;
1111 features
->descriptorBindingSampledImageUpdateAfterBind
= true;
1112 features
->descriptorBindingStorageImageUpdateAfterBind
= true;
1113 features
->descriptorBindingStorageBufferUpdateAfterBind
= true;
1114 features
->descriptorBindingUniformTexelBufferUpdateAfterBind
= true;
1115 features
->descriptorBindingStorageTexelBufferUpdateAfterBind
= true;
1116 features
->descriptorBindingUpdateUnusedWhilePending
= true;
1117 features
->descriptorBindingPartiallyBound
= true;
1118 features
->descriptorBindingVariableDescriptorCount
= false;
1119 features
->runtimeDescriptorArray
= true;
1123 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INDEX_TYPE_UINT8_FEATURES_EXT
: {
1124 VkPhysicalDeviceIndexTypeUint8FeaturesEXT
*features
=
1125 (VkPhysicalDeviceIndexTypeUint8FeaturesEXT
*)ext
;
1126 features
->indexTypeUint8
= true;
1130 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_FEATURES_EXT
: {
1131 VkPhysicalDeviceInlineUniformBlockFeaturesEXT
*features
=
1132 (VkPhysicalDeviceInlineUniformBlockFeaturesEXT
*)ext
;
1133 features
->inlineUniformBlock
= true;
1134 features
->descriptorBindingInlineUniformBlockUpdateAfterBind
= true;
1138 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_FEATURES_EXT
: {
1139 VkPhysicalDeviceLineRasterizationFeaturesEXT
*features
=
1140 (VkPhysicalDeviceLineRasterizationFeaturesEXT
*)ext
;
1141 features
->rectangularLines
= true;
1142 features
->bresenhamLines
= true;
1143 features
->smoothLines
= true;
1144 features
->stippledRectangularLines
= false;
1145 features
->stippledBresenhamLines
= true;
1146 features
->stippledSmoothLines
= false;
1150 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES
: {
1151 VkPhysicalDeviceMultiviewFeatures
*features
=
1152 (VkPhysicalDeviceMultiviewFeatures
*)ext
;
1153 features
->multiview
= true;
1154 features
->multiviewGeometryShader
= true;
1155 features
->multiviewTessellationShader
= true;
1159 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGELESS_FRAMEBUFFER_FEATURES_KHR
: {
1160 VkPhysicalDeviceImagelessFramebufferFeaturesKHR
*features
=
1161 (VkPhysicalDeviceImagelessFramebufferFeaturesKHR
*)ext
;
1162 features
->imagelessFramebuffer
= true;
1166 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR
: {
1167 VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR
*features
=
1168 (VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR
*)ext
;
1169 features
->pipelineExecutableInfo
= true;
1173 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES
: {
1174 VkPhysicalDeviceProtectedMemoryFeatures
*features
= (void *)ext
;
1175 features
->protectedMemory
= false;
1179 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES
: {
1180 VkPhysicalDeviceSamplerYcbcrConversionFeatures
*features
=
1181 (VkPhysicalDeviceSamplerYcbcrConversionFeatures
*) ext
;
1182 features
->samplerYcbcrConversion
= true;
1186 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES_EXT
: {
1187 VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*features
=
1188 (VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*)ext
;
1189 features
->scalarBlockLayout
= true;
1193 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SEPARATE_DEPTH_STENCIL_LAYOUTS_FEATURES_KHR
: {
1194 VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR
*features
=
1195 (VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR
*)ext
;
1196 features
->separateDepthStencilLayouts
= true;
1200 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_INT64_FEATURES_KHR
: {
1201 VkPhysicalDeviceShaderAtomicInt64FeaturesKHR
*features
= (void *)ext
;
1202 features
->shaderBufferInt64Atomics
=
1203 pdevice
->info
.gen
>= 9 && pdevice
->use_softpin
;
1204 features
->shaderSharedInt64Atomics
= VK_FALSE
;
1208 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DEMOTE_TO_HELPER_INVOCATION_FEATURES_EXT
: {
1209 VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT
*features
= (void *)ext
;
1210 features
->shaderDemoteToHelperInvocation
= true;
1214 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CLOCK_FEATURES_KHR
: {
1215 VkPhysicalDeviceShaderClockFeaturesKHR
*features
=
1216 (VkPhysicalDeviceShaderClockFeaturesKHR
*)ext
;
1217 features
->shaderSubgroupClock
= true;
1218 features
->shaderDeviceClock
= false;
1222 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETERS_FEATURES
: {
1223 VkPhysicalDeviceShaderDrawParametersFeatures
*features
= (void *)ext
;
1224 features
->shaderDrawParameters
= true;
1228 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_SUBGROUP_EXTENDED_TYPES_FEATURES_KHR
: {
1229 VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR
*features
=
1230 (VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR
*)ext
;
1231 features
->shaderSubgroupExtendedTypes
= true;
1235 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_FEATURES_EXT
: {
1236 VkPhysicalDeviceSubgroupSizeControlFeaturesEXT
*features
=
1237 (VkPhysicalDeviceSubgroupSizeControlFeaturesEXT
*)ext
;
1238 features
->subgroupSizeControl
= true;
1239 features
->computeFullSubgroups
= true;
1243 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_FEATURES_EXT
: {
1244 VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT
*features
=
1245 (VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT
*)ext
;
1246 features
->texelBufferAlignment
= true;
1250 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_FEATURES_KHR
: {
1251 VkPhysicalDeviceTimelineSemaphoreFeaturesKHR
*features
=
1252 (VkPhysicalDeviceTimelineSemaphoreFeaturesKHR
*) ext
;
1253 features
->timelineSemaphore
= true;
1257 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES
: {
1258 VkPhysicalDeviceVariablePointersFeatures
*features
= (void *)ext
;
1259 features
->variablePointersStorageBuffer
= true;
1260 features
->variablePointers
= true;
1264 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT
: {
1265 VkPhysicalDeviceTransformFeedbackFeaturesEXT
*features
=
1266 (VkPhysicalDeviceTransformFeedbackFeaturesEXT
*)ext
;
1267 features
->transformFeedback
= true;
1268 features
->geometryStreams
= true;
1272 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_UNIFORM_BUFFER_STANDARD_LAYOUT_FEATURES_KHR
: {
1273 VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR
*features
=
1274 (VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR
*)ext
;
1275 features
->uniformBufferStandardLayout
= true;
1279 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT
: {
1280 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*features
=
1281 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*)ext
;
1282 features
->vertexAttributeInstanceRateDivisor
= true;
1283 features
->vertexAttributeInstanceRateZeroDivisor
= true;
1287 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_MEMORY_MODEL_FEATURES_KHR
: {
1288 VkPhysicalDeviceVulkanMemoryModelFeaturesKHR
*features
= (void *)ext
;
1289 features
->vulkanMemoryModel
= true;
1290 features
->vulkanMemoryModelDeviceScope
= true;
1291 features
->vulkanMemoryModelAvailabilityVisibilityChains
= true;
1295 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_YCBCR_IMAGE_ARRAYS_FEATURES_EXT
: {
1296 VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*features
=
1297 (VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*)ext
;
1298 features
->ycbcrImageArrays
= true;
1303 anv_debug_ignored_stype(ext
->sType
);
1309 #define MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS 64
1311 #define MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS 64
1312 #define MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS 256
1314 void anv_GetPhysicalDeviceProperties(
1315 VkPhysicalDevice physicalDevice
,
1316 VkPhysicalDeviceProperties
* pProperties
)
1318 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1319 const struct gen_device_info
*devinfo
= &pdevice
->info
;
1321 /* See assertions made when programming the buffer surface state. */
1322 const uint32_t max_raw_buffer_sz
= devinfo
->gen
>= 7 ?
1323 (1ul << 30) : (1ul << 27);
1325 const uint32_t max_ssbos
= pdevice
->has_a64_buffer_access
? UINT16_MAX
: 64;
1326 const uint32_t max_textures
=
1327 pdevice
->has_bindless_images
? UINT16_MAX
: 128;
1328 const uint32_t max_samplers
=
1329 pdevice
->has_bindless_samplers
? UINT16_MAX
:
1330 (devinfo
->gen
>= 8 || devinfo
->is_haswell
) ? 128 : 16;
1331 const uint32_t max_images
=
1332 pdevice
->has_bindless_images
? UINT16_MAX
: MAX_IMAGES
;
1334 /* If we can use bindless for everything, claim a high per-stage limit,
1335 * otherwise use the binding table size, minus the slots reserved for
1336 * render targets and one slot for the descriptor buffer. */
1337 const uint32_t max_per_stage
=
1338 pdevice
->has_bindless_images
&& pdevice
->has_a64_buffer_access
1339 ? UINT32_MAX
: MAX_BINDING_TABLE_SIZE
- MAX_RTS
- 1;
1341 const uint32_t max_workgroup_size
= 32 * devinfo
->max_cs_threads
;
1343 VkSampleCountFlags sample_counts
=
1344 isl_device_get_sample_counts(&pdevice
->isl_dev
);
1347 VkPhysicalDeviceLimits limits
= {
1348 .maxImageDimension1D
= (1 << 14),
1349 .maxImageDimension2D
= (1 << 14),
1350 .maxImageDimension3D
= (1 << 11),
1351 .maxImageDimensionCube
= (1 << 14),
1352 .maxImageArrayLayers
= (1 << 11),
1353 .maxTexelBufferElements
= 128 * 1024 * 1024,
1354 .maxUniformBufferRange
= (1ul << 27),
1355 .maxStorageBufferRange
= max_raw_buffer_sz
,
1356 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
1357 .maxMemoryAllocationCount
= UINT32_MAX
,
1358 .maxSamplerAllocationCount
= 64 * 1024,
1359 .bufferImageGranularity
= 64, /* A cache line */
1360 .sparseAddressSpaceSize
= 0,
1361 .maxBoundDescriptorSets
= MAX_SETS
,
1362 .maxPerStageDescriptorSamplers
= max_samplers
,
1363 .maxPerStageDescriptorUniformBuffers
= MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
,
1364 .maxPerStageDescriptorStorageBuffers
= max_ssbos
,
1365 .maxPerStageDescriptorSampledImages
= max_textures
,
1366 .maxPerStageDescriptorStorageImages
= max_images
,
1367 .maxPerStageDescriptorInputAttachments
= MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS
,
1368 .maxPerStageResources
= max_per_stage
,
1369 .maxDescriptorSetSamplers
= 6 * max_samplers
, /* number of stages * maxPerStageDescriptorSamplers */
1370 .maxDescriptorSetUniformBuffers
= 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
, /* number of stages * maxPerStageDescriptorUniformBuffers */
1371 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
1372 .maxDescriptorSetStorageBuffers
= 6 * max_ssbos
, /* number of stages * maxPerStageDescriptorStorageBuffers */
1373 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
1374 .maxDescriptorSetSampledImages
= 6 * max_textures
, /* number of stages * maxPerStageDescriptorSampledImages */
1375 .maxDescriptorSetStorageImages
= 6 * max_images
, /* number of stages * maxPerStageDescriptorStorageImages */
1376 .maxDescriptorSetInputAttachments
= MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS
,
1377 .maxVertexInputAttributes
= MAX_VBS
,
1378 .maxVertexInputBindings
= MAX_VBS
,
1379 .maxVertexInputAttributeOffset
= 2047,
1380 .maxVertexInputBindingStride
= 2048,
1381 .maxVertexOutputComponents
= 128,
1382 .maxTessellationGenerationLevel
= 64,
1383 .maxTessellationPatchSize
= 32,
1384 .maxTessellationControlPerVertexInputComponents
= 128,
1385 .maxTessellationControlPerVertexOutputComponents
= 128,
1386 .maxTessellationControlPerPatchOutputComponents
= 128,
1387 .maxTessellationControlTotalOutputComponents
= 2048,
1388 .maxTessellationEvaluationInputComponents
= 128,
1389 .maxTessellationEvaluationOutputComponents
= 128,
1390 .maxGeometryShaderInvocations
= 32,
1391 .maxGeometryInputComponents
= 64,
1392 .maxGeometryOutputComponents
= 128,
1393 .maxGeometryOutputVertices
= 256,
1394 .maxGeometryTotalOutputComponents
= 1024,
1395 .maxFragmentInputComponents
= 116, /* 128 components - (PSIZ, CLIP_DIST0, CLIP_DIST1) */
1396 .maxFragmentOutputAttachments
= 8,
1397 .maxFragmentDualSrcAttachments
= 1,
1398 .maxFragmentCombinedOutputResources
= 8,
1399 .maxComputeSharedMemorySize
= 64 * 1024,
1400 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
1401 .maxComputeWorkGroupInvocations
= max_workgroup_size
,
1402 .maxComputeWorkGroupSize
= {
1407 .subPixelPrecisionBits
= 8,
1408 .subTexelPrecisionBits
= 8,
1409 .mipmapPrecisionBits
= 8,
1410 .maxDrawIndexedIndexValue
= UINT32_MAX
,
1411 .maxDrawIndirectCount
= UINT32_MAX
,
1412 .maxSamplerLodBias
= 16,
1413 .maxSamplerAnisotropy
= 16,
1414 .maxViewports
= MAX_VIEWPORTS
,
1415 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
1416 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
1417 .viewportSubPixelBits
= 13, /* We take a float? */
1418 .minMemoryMapAlignment
= 4096, /* A page */
1419 /* The dataport requires texel alignment so we need to assume a worst
1420 * case of R32G32B32A32 which is 16 bytes.
1422 .minTexelBufferOffsetAlignment
= 16,
1423 /* We need 16 for UBO block reads to work and 32 for push UBOs */
1424 .minUniformBufferOffsetAlignment
= 32,
1425 .minStorageBufferOffsetAlignment
= 4,
1426 .minTexelOffset
= -8,
1427 .maxTexelOffset
= 7,
1428 .minTexelGatherOffset
= -32,
1429 .maxTexelGatherOffset
= 31,
1430 .minInterpolationOffset
= -0.5,
1431 .maxInterpolationOffset
= 0.4375,
1432 .subPixelInterpolationOffsetBits
= 4,
1433 .maxFramebufferWidth
= (1 << 14),
1434 .maxFramebufferHeight
= (1 << 14),
1435 .maxFramebufferLayers
= (1 << 11),
1436 .framebufferColorSampleCounts
= sample_counts
,
1437 .framebufferDepthSampleCounts
= sample_counts
,
1438 .framebufferStencilSampleCounts
= sample_counts
,
1439 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
1440 .maxColorAttachments
= MAX_RTS
,
1441 .sampledImageColorSampleCounts
= sample_counts
,
1442 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
1443 .sampledImageDepthSampleCounts
= sample_counts
,
1444 .sampledImageStencilSampleCounts
= sample_counts
,
1445 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
1446 .maxSampleMaskWords
= 1,
1447 .timestampComputeAndGraphics
= true,
1448 .timestampPeriod
= 1000000000.0 / devinfo
->timestamp_frequency
,
1449 .maxClipDistances
= 8,
1450 .maxCullDistances
= 8,
1451 .maxCombinedClipAndCullDistances
= 8,
1452 .discreteQueuePriorities
= 2,
1453 .pointSizeRange
= { 0.125, 255.875 },
1456 (devinfo
->gen
>= 9 || devinfo
->is_cherryview
) ?
1457 2047.9921875 : 7.9921875,
1459 .pointSizeGranularity
= (1.0 / 8.0),
1460 .lineWidthGranularity
= (1.0 / 128.0),
1461 .strictLines
= false,
1462 .standardSampleLocations
= true,
1463 .optimalBufferCopyOffsetAlignment
= 128,
1464 .optimalBufferCopyRowPitchAlignment
= 128,
1465 .nonCoherentAtomSize
= 64,
1468 *pProperties
= (VkPhysicalDeviceProperties
) {
1469 .apiVersion
= anv_physical_device_api_version(pdevice
),
1470 .driverVersion
= vk_get_driver_version(),
1472 .deviceID
= pdevice
->chipset_id
,
1473 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
1475 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
1478 snprintf(pProperties
->deviceName
, sizeof(pProperties
->deviceName
),
1479 "%s", pdevice
->name
);
1480 memcpy(pProperties
->pipelineCacheUUID
,
1481 pdevice
->pipeline_cache_uuid
, VK_UUID_SIZE
);
1484 void anv_GetPhysicalDeviceProperties2(
1485 VkPhysicalDevice physicalDevice
,
1486 VkPhysicalDeviceProperties2
* pProperties
)
1488 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1490 anv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
1492 vk_foreach_struct(ext
, pProperties
->pNext
) {
1493 switch (ext
->sType
) {
1494 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_STENCIL_RESOLVE_PROPERTIES_KHR
: {
1495 VkPhysicalDeviceDepthStencilResolvePropertiesKHR
*props
=
1496 (VkPhysicalDeviceDepthStencilResolvePropertiesKHR
*)ext
;
1498 /* We support all of the depth resolve modes */
1499 props
->supportedDepthResolveModes
=
1500 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
|
1501 VK_RESOLVE_MODE_AVERAGE_BIT_KHR
|
1502 VK_RESOLVE_MODE_MIN_BIT_KHR
|
1503 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1505 /* Average doesn't make sense for stencil so we don't support that */
1506 props
->supportedStencilResolveModes
=
1507 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
;
1508 if (pdevice
->info
.gen
>= 8) {
1509 /* The advanced stencil resolve modes currently require stencil
1510 * sampling be supported by the hardware.
1512 props
->supportedStencilResolveModes
|=
1513 VK_RESOLVE_MODE_MIN_BIT_KHR
|
1514 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1517 props
->independentResolveNone
= true;
1518 props
->independentResolve
= true;
1522 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES_EXT
: {
1523 VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*props
=
1524 (VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*)ext
;
1526 /* It's a bit hard to exactly map our implementation to the limits
1527 * described here. The bindless surface handle in the extended
1528 * message descriptors is 20 bits and it's an index into the table of
1529 * RENDER_SURFACE_STATE structs that starts at bindless surface base
1530 * address. Given that most things consume two surface states per
1531 * view (general/sampled for textures and write-only/read-write for
1532 * images), we claim 2^19 things.
1534 * For SSBOs, we just use A64 messages so there is no real limit
1535 * there beyond the limit on the total size of a descriptor set.
1537 const unsigned max_bindless_views
= 1 << 19;
1539 props
->maxUpdateAfterBindDescriptorsInAllPools
= max_bindless_views
;
1540 props
->shaderUniformBufferArrayNonUniformIndexingNative
= false;
1541 props
->shaderSampledImageArrayNonUniformIndexingNative
= false;
1542 props
->shaderStorageBufferArrayNonUniformIndexingNative
= true;
1543 props
->shaderStorageImageArrayNonUniformIndexingNative
= false;
1544 props
->shaderInputAttachmentArrayNonUniformIndexingNative
= false;
1545 props
->robustBufferAccessUpdateAfterBind
= true;
1546 props
->quadDivergentImplicitLod
= false;
1547 props
->maxPerStageDescriptorUpdateAfterBindSamplers
= max_bindless_views
;
1548 props
->maxPerStageDescriptorUpdateAfterBindUniformBuffers
= MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
;
1549 props
->maxPerStageDescriptorUpdateAfterBindStorageBuffers
= UINT32_MAX
;
1550 props
->maxPerStageDescriptorUpdateAfterBindSampledImages
= max_bindless_views
;
1551 props
->maxPerStageDescriptorUpdateAfterBindStorageImages
= max_bindless_views
;
1552 props
->maxPerStageDescriptorUpdateAfterBindInputAttachments
= MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS
;
1553 props
->maxPerStageUpdateAfterBindResources
= UINT32_MAX
;
1554 props
->maxDescriptorSetUpdateAfterBindSamplers
= max_bindless_views
;
1555 props
->maxDescriptorSetUpdateAfterBindUniformBuffers
= 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
;
1556 props
->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2;
1557 props
->maxDescriptorSetUpdateAfterBindStorageBuffers
= UINT32_MAX
;
1558 props
->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2;
1559 props
->maxDescriptorSetUpdateAfterBindSampledImages
= max_bindless_views
;
1560 props
->maxDescriptorSetUpdateAfterBindStorageImages
= max_bindless_views
;
1561 props
->maxDescriptorSetUpdateAfterBindInputAttachments
= MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS
;
1565 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR
: {
1566 VkPhysicalDeviceDriverPropertiesKHR
*driver_props
=
1567 (VkPhysicalDeviceDriverPropertiesKHR
*) ext
;
1569 driver_props
->driverID
= VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA_KHR
;
1570 snprintf(driver_props
->driverName
, VK_MAX_DRIVER_NAME_SIZE_KHR
,
1571 "Intel open-source Mesa driver");
1573 snprintf(driver_props
->driverInfo
, VK_MAX_DRIVER_INFO_SIZE_KHR
,
1574 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1
);
1576 driver_props
->conformanceVersion
= (VkConformanceVersionKHR
) {
1585 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT
: {
1586 VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*props
=
1587 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*) ext
;
1588 /* Userptr needs page aligned memory. */
1589 props
->minImportedHostPointerAlignment
= 4096;
1593 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES
: {
1594 VkPhysicalDeviceIDProperties
*id_props
=
1595 (VkPhysicalDeviceIDProperties
*)ext
;
1596 memcpy(id_props
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
1597 memcpy(id_props
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
1598 /* The LUID is for Windows. */
1599 id_props
->deviceLUIDValid
= false;
1603 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_PROPERTIES_EXT
: {
1604 VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*props
=
1605 (VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*)ext
;
1606 props
->maxInlineUniformBlockSize
= MAX_INLINE_UNIFORM_BLOCK_SIZE
;
1607 props
->maxPerStageDescriptorInlineUniformBlocks
=
1608 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1609 props
->maxPerStageDescriptorUpdateAfterBindInlineUniformBlocks
=
1610 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1611 props
->maxDescriptorSetInlineUniformBlocks
=
1612 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1613 props
->maxDescriptorSetUpdateAfterBindInlineUniformBlocks
=
1614 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1618 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_PROPERTIES_EXT
: {
1619 VkPhysicalDeviceLineRasterizationPropertiesEXT
*props
=
1620 (VkPhysicalDeviceLineRasterizationPropertiesEXT
*)ext
;
1621 /* In the Skylake PRM Vol. 7, subsection titled "GIQ (Diamond)
1622 * Sampling Rules - Legacy Mode", it says the following:
1624 * "Note that the device divides a pixel into a 16x16 array of
1625 * subpixels, referenced by their upper left corners."
1627 * This is the only known reference in the PRMs to the subpixel
1628 * precision of line rasterization and a "16x16 array of subpixels"
1629 * implies 4 subpixel precision bits. Empirical testing has shown
1630 * that 4 subpixel precision bits applies to all line rasterization
1633 props
->lineSubPixelPrecisionBits
= 4;
1637 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES
: {
1638 VkPhysicalDeviceMaintenance3Properties
*props
=
1639 (VkPhysicalDeviceMaintenance3Properties
*)ext
;
1640 /* This value doesn't matter for us today as our per-stage
1641 * descriptors are the real limit.
1643 props
->maxPerSetDescriptors
= 1024;
1644 props
->maxMemoryAllocationSize
= MAX_MEMORY_ALLOCATION_SIZE
;
1648 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES
: {
1649 VkPhysicalDeviceMultiviewProperties
*properties
=
1650 (VkPhysicalDeviceMultiviewProperties
*)ext
;
1651 properties
->maxMultiviewViewCount
= 16;
1652 properties
->maxMultiviewInstanceIndex
= UINT32_MAX
/ 16;
1656 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT
: {
1657 VkPhysicalDevicePCIBusInfoPropertiesEXT
*properties
=
1658 (VkPhysicalDevicePCIBusInfoPropertiesEXT
*)ext
;
1659 properties
->pciDomain
= pdevice
->pci_info
.domain
;
1660 properties
->pciBus
= pdevice
->pci_info
.bus
;
1661 properties
->pciDevice
= pdevice
->pci_info
.device
;
1662 properties
->pciFunction
= pdevice
->pci_info
.function
;
1666 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES
: {
1667 VkPhysicalDevicePointClippingProperties
*properties
=
1668 (VkPhysicalDevicePointClippingProperties
*) ext
;
1669 properties
->pointClippingBehavior
= VK_POINT_CLIPPING_BEHAVIOR_USER_CLIP_PLANES_ONLY
;
1673 #pragma GCC diagnostic push
1674 #pragma GCC diagnostic ignored "-Wswitch"
1675 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PRESENTATION_PROPERTIES_ANDROID
: {
1676 VkPhysicalDevicePresentationPropertiesANDROID
*props
=
1677 (VkPhysicalDevicePresentationPropertiesANDROID
*)ext
;
1678 props
->sharedImage
= VK_FALSE
;
1681 #pragma GCC diagnostic pop
1683 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES
: {
1684 VkPhysicalDeviceProtectedMemoryProperties
*props
=
1685 (VkPhysicalDeviceProtectedMemoryProperties
*)ext
;
1686 props
->protectedNoFault
= false;
1690 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
1691 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
1692 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
1694 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
1698 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT
: {
1699 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*properties
=
1700 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*)ext
;
1701 properties
->filterMinmaxImageComponentMapping
= pdevice
->info
.gen
>= 9;
1702 properties
->filterMinmaxSingleComponentFormats
= true;
1706 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES
: {
1707 VkPhysicalDeviceSubgroupProperties
*properties
= (void *)ext
;
1709 properties
->subgroupSize
= BRW_SUBGROUP_SIZE
;
1711 VkShaderStageFlags scalar_stages
= 0;
1712 for (unsigned stage
= 0; stage
< MESA_SHADER_STAGES
; stage
++) {
1713 if (pdevice
->compiler
->scalar_stage
[stage
])
1714 scalar_stages
|= mesa_to_vk_shader_stage(stage
);
1716 properties
->supportedStages
= scalar_stages
;
1718 properties
->supportedOperations
= VK_SUBGROUP_FEATURE_BASIC_BIT
|
1719 VK_SUBGROUP_FEATURE_VOTE_BIT
|
1720 VK_SUBGROUP_FEATURE_BALLOT_BIT
|
1721 VK_SUBGROUP_FEATURE_SHUFFLE_BIT
|
1722 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT
|
1723 VK_SUBGROUP_FEATURE_QUAD_BIT
;
1724 if (pdevice
->info
.gen
>= 8) {
1725 /* TODO: There's no technical reason why these can't be made to
1726 * work on gen7 but they don't at the moment so it's best to leave
1727 * the feature disabled than enabled and broken.
1729 properties
->supportedOperations
|=
1730 VK_SUBGROUP_FEATURE_ARITHMETIC_BIT
|
1731 VK_SUBGROUP_FEATURE_CLUSTERED_BIT
;
1733 properties
->quadOperationsInAllStages
= pdevice
->info
.gen
>= 8;
1737 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_PROPERTIES_EXT
: {
1738 VkPhysicalDeviceSubgroupSizeControlPropertiesEXT
*props
=
1739 (VkPhysicalDeviceSubgroupSizeControlPropertiesEXT
*)ext
;
1740 STATIC_ASSERT(8 <= BRW_SUBGROUP_SIZE
&& BRW_SUBGROUP_SIZE
<= 32);
1741 props
->minSubgroupSize
= 8;
1742 props
->maxSubgroupSize
= 32;
1743 props
->maxComputeWorkgroupSubgroups
= pdevice
->info
.max_cs_threads
;
1744 props
->requiredSubgroupSizeStages
= VK_SHADER_STAGE_COMPUTE_BIT
;
1747 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT_CONTROLS_PROPERTIES_KHR
: {
1748 VkPhysicalDeviceFloatControlsPropertiesKHR
*properties
= (void *)ext
;
1749 properties
->denormBehaviorIndependence
= VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_ALL_KHR
;
1750 properties
->roundingModeIndependence
= VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_NONE_KHR
;
1752 /* Broadwell does not support HF denorms and there are restrictions
1753 * other gens. According to Kabylake's PRM:
1755 * "math - Extended Math Function
1757 * Restriction : Half-float denorms are always retained."
1759 properties
->shaderDenormFlushToZeroFloat16
= false;
1760 properties
->shaderDenormPreserveFloat16
= pdevice
->info
.gen
> 8;
1761 properties
->shaderRoundingModeRTEFloat16
= true;
1762 properties
->shaderRoundingModeRTZFloat16
= true;
1763 properties
->shaderSignedZeroInfNanPreserveFloat16
= true;
1765 properties
->shaderDenormFlushToZeroFloat32
= true;
1766 properties
->shaderDenormPreserveFloat32
= true;
1767 properties
->shaderRoundingModeRTEFloat32
= true;
1768 properties
->shaderRoundingModeRTZFloat32
= true;
1769 properties
->shaderSignedZeroInfNanPreserveFloat32
= true;
1771 properties
->shaderDenormFlushToZeroFloat64
= true;
1772 properties
->shaderDenormPreserveFloat64
= true;
1773 properties
->shaderRoundingModeRTEFloat64
= true;
1774 properties
->shaderRoundingModeRTZFloat64
= true;
1775 properties
->shaderSignedZeroInfNanPreserveFloat64
= true;
1779 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_PROPERTIES_EXT
: {
1780 VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT
*props
=
1781 (VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT
*)ext
;
1783 /* From the SKL PRM Vol. 2d, docs for RENDER_SURFACE_STATE::Surface
1786 * "For SURFTYPE_BUFFER non-rendertarget surfaces, this field
1787 * specifies the base address of the first element of the surface,
1788 * computed in software by adding the surface base address to the
1789 * byte offset of the element in the buffer. The base address must
1790 * be aligned to element size."
1792 * The typed dataport messages require that things be texel aligned.
1793 * Otherwise, we may just load/store the wrong data or, in the worst
1794 * case, there may be hangs.
1796 props
->storageTexelBufferOffsetAlignmentBytes
= 16;
1797 props
->storageTexelBufferOffsetSingleTexelAlignment
= true;
1799 /* The sampler, however, is much more forgiving and it can handle
1800 * arbitrary byte alignment for linear and buffer surfaces. It's
1801 * hard to find a good PRM citation for this but years of empirical
1802 * experience demonstrate that this is true.
1804 props
->uniformTexelBufferOffsetAlignmentBytes
= 1;
1805 props
->uniformTexelBufferOffsetSingleTexelAlignment
= false;
1809 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_PROPERTIES_KHR
: {
1810 VkPhysicalDeviceTimelineSemaphorePropertiesKHR
*props
=
1811 (VkPhysicalDeviceTimelineSemaphorePropertiesKHR
*) ext
;
1812 props
->maxTimelineSemaphoreValueDifference
= UINT64_MAX
;
1816 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT
: {
1817 VkPhysicalDeviceTransformFeedbackPropertiesEXT
*props
=
1818 (VkPhysicalDeviceTransformFeedbackPropertiesEXT
*)ext
;
1820 props
->maxTransformFeedbackStreams
= MAX_XFB_STREAMS
;
1821 props
->maxTransformFeedbackBuffers
= MAX_XFB_BUFFERS
;
1822 props
->maxTransformFeedbackBufferSize
= (1ull << 32);
1823 props
->maxTransformFeedbackStreamDataSize
= 128 * 4;
1824 props
->maxTransformFeedbackBufferDataSize
= 128 * 4;
1825 props
->maxTransformFeedbackBufferDataStride
= 2048;
1826 props
->transformFeedbackQueries
= true;
1827 props
->transformFeedbackStreamsLinesTriangles
= false;
1828 props
->transformFeedbackRasterizationStreamSelect
= false;
1829 props
->transformFeedbackDraw
= true;
1833 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT
: {
1834 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*props
=
1835 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*)ext
;
1836 /* We have to restrict this a bit for multiview */
1837 props
->maxVertexAttribDivisor
= UINT32_MAX
/ 16;
1842 anv_debug_ignored_stype(ext
->sType
);
1848 /* We support exactly one queue family. */
1849 static const VkQueueFamilyProperties
1850 anv_queue_family_properties
= {
1851 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
1852 VK_QUEUE_COMPUTE_BIT
|
1853 VK_QUEUE_TRANSFER_BIT
,
1855 .timestampValidBits
= 36, /* XXX: Real value here */
1856 .minImageTransferGranularity
= { 1, 1, 1 },
1859 void anv_GetPhysicalDeviceQueueFamilyProperties(
1860 VkPhysicalDevice physicalDevice
,
1862 VkQueueFamilyProperties
* pQueueFamilyProperties
)
1864 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pCount
);
1866 vk_outarray_append(&out
, p
) {
1867 *p
= anv_queue_family_properties
;
1871 void anv_GetPhysicalDeviceQueueFamilyProperties2(
1872 VkPhysicalDevice physicalDevice
,
1873 uint32_t* pQueueFamilyPropertyCount
,
1874 VkQueueFamilyProperties2
* pQueueFamilyProperties
)
1877 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pQueueFamilyPropertyCount
);
1879 vk_outarray_append(&out
, p
) {
1880 p
->queueFamilyProperties
= anv_queue_family_properties
;
1882 vk_foreach_struct(s
, p
->pNext
) {
1883 anv_debug_ignored_stype(s
->sType
);
1888 void anv_GetPhysicalDeviceMemoryProperties(
1889 VkPhysicalDevice physicalDevice
,
1890 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
1892 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
1894 pMemoryProperties
->memoryTypeCount
= physical_device
->memory
.type_count
;
1895 for (uint32_t i
= 0; i
< physical_device
->memory
.type_count
; i
++) {
1896 pMemoryProperties
->memoryTypes
[i
] = (VkMemoryType
) {
1897 .propertyFlags
= physical_device
->memory
.types
[i
].propertyFlags
,
1898 .heapIndex
= physical_device
->memory
.types
[i
].heapIndex
,
1902 pMemoryProperties
->memoryHeapCount
= physical_device
->memory
.heap_count
;
1903 for (uint32_t i
= 0; i
< physical_device
->memory
.heap_count
; i
++) {
1904 pMemoryProperties
->memoryHeaps
[i
] = (VkMemoryHeap
) {
1905 .size
= physical_device
->memory
.heaps
[i
].size
,
1906 .flags
= physical_device
->memory
.heaps
[i
].flags
,
1912 anv_get_memory_budget(VkPhysicalDevice physicalDevice
,
1913 VkPhysicalDeviceMemoryBudgetPropertiesEXT
*memoryBudget
)
1915 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
1916 uint64_t sys_available
= get_available_system_memory();
1917 assert(sys_available
> 0);
1919 VkDeviceSize total_heaps_size
= 0;
1920 for (size_t i
= 0; i
< device
->memory
.heap_count
; i
++)
1921 total_heaps_size
+= device
->memory
.heaps
[i
].size
;
1923 for (size_t i
= 0; i
< device
->memory
.heap_count
; i
++) {
1924 VkDeviceSize heap_size
= device
->memory
.heaps
[i
].size
;
1925 VkDeviceSize heap_used
= device
->memory
.heaps
[i
].used
;
1926 VkDeviceSize heap_budget
;
1928 double heap_proportion
= (double) heap_size
/ total_heaps_size
;
1929 VkDeviceSize sys_available_prop
= sys_available
* heap_proportion
;
1932 * Let's not incite the app to starve the system: report at most 90% of
1933 * available system memory.
1935 uint64_t heap_available
= sys_available_prop
* 9 / 10;
1936 heap_budget
= MIN2(heap_size
, heap_used
+ heap_available
);
1939 * Round down to the nearest MB
1941 heap_budget
&= ~((1ull << 20) - 1);
1944 * The heapBudget value must be non-zero for array elements less than
1945 * VkPhysicalDeviceMemoryProperties::memoryHeapCount. The heapBudget
1946 * value must be less than or equal to VkMemoryHeap::size for each heap.
1948 assert(0 < heap_budget
&& heap_budget
<= heap_size
);
1950 memoryBudget
->heapUsage
[i
] = heap_used
;
1951 memoryBudget
->heapBudget
[i
] = heap_budget
;
1954 /* The heapBudget and heapUsage values must be zero for array elements
1955 * greater than or equal to VkPhysicalDeviceMemoryProperties::memoryHeapCount
1957 for (uint32_t i
= device
->memory
.heap_count
; i
< VK_MAX_MEMORY_HEAPS
; i
++) {
1958 memoryBudget
->heapBudget
[i
] = 0;
1959 memoryBudget
->heapUsage
[i
] = 0;
1963 void anv_GetPhysicalDeviceMemoryProperties2(
1964 VkPhysicalDevice physicalDevice
,
1965 VkPhysicalDeviceMemoryProperties2
* pMemoryProperties
)
1967 anv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
1968 &pMemoryProperties
->memoryProperties
);
1970 vk_foreach_struct(ext
, pMemoryProperties
->pNext
) {
1971 switch (ext
->sType
) {
1972 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT
:
1973 anv_get_memory_budget(physicalDevice
, (void*)ext
);
1976 anv_debug_ignored_stype(ext
->sType
);
1983 anv_GetDeviceGroupPeerMemoryFeatures(
1986 uint32_t localDeviceIndex
,
1987 uint32_t remoteDeviceIndex
,
1988 VkPeerMemoryFeatureFlags
* pPeerMemoryFeatures
)
1990 assert(localDeviceIndex
== 0 && remoteDeviceIndex
== 0);
1991 *pPeerMemoryFeatures
= VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT
|
1992 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT
|
1993 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT
|
1994 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT
;
1997 PFN_vkVoidFunction
anv_GetInstanceProcAddr(
1998 VkInstance _instance
,
2001 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2003 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
2004 * when we have to return valid function pointers, NULL, or it's left
2005 * undefined. See the table for exact details.
2010 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
2011 if (strcmp(pName, "vk" #entrypoint) == 0) \
2012 return (PFN_vkVoidFunction)anv_##entrypoint
2014 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties
);
2015 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties
);
2016 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceVersion
);
2017 LOOKUP_ANV_ENTRYPOINT(CreateInstance
);
2019 #undef LOOKUP_ANV_ENTRYPOINT
2021 if (instance
== NULL
)
2024 int idx
= anv_get_instance_entrypoint_index(pName
);
2026 return instance
->dispatch
.entrypoints
[idx
];
2028 idx
= anv_get_physical_device_entrypoint_index(pName
);
2030 return instance
->physicalDevice
.dispatch
.entrypoints
[idx
];
2032 idx
= anv_get_device_entrypoint_index(pName
);
2034 return instance
->device_dispatch
.entrypoints
[idx
];
2039 /* With version 1+ of the loader interface the ICD should expose
2040 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
2043 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
2044 VkInstance instance
,
2048 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
2049 VkInstance instance
,
2052 return anv_GetInstanceProcAddr(instance
, pName
);
2055 PFN_vkVoidFunction
anv_GetDeviceProcAddr(
2059 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2061 if (!device
|| !pName
)
2064 int idx
= anv_get_device_entrypoint_index(pName
);
2068 return device
->dispatch
.entrypoints
[idx
];
2071 /* With version 4+ of the loader interface the ICD should expose
2072 * vk_icdGetPhysicalDeviceProcAddr()
2075 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetPhysicalDeviceProcAddr(
2076 VkInstance _instance
,
2079 PFN_vkVoidFunction
vk_icdGetPhysicalDeviceProcAddr(
2080 VkInstance _instance
,
2083 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2085 if (!pName
|| !instance
)
2088 int idx
= anv_get_physical_device_entrypoint_index(pName
);
2092 return instance
->physicalDevice
.dispatch
.entrypoints
[idx
];
2097 anv_CreateDebugReportCallbackEXT(VkInstance _instance
,
2098 const VkDebugReportCallbackCreateInfoEXT
* pCreateInfo
,
2099 const VkAllocationCallbacks
* pAllocator
,
2100 VkDebugReportCallbackEXT
* pCallback
)
2102 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2103 return vk_create_debug_report_callback(&instance
->debug_report_callbacks
,
2104 pCreateInfo
, pAllocator
, &instance
->alloc
,
2109 anv_DestroyDebugReportCallbackEXT(VkInstance _instance
,
2110 VkDebugReportCallbackEXT _callback
,
2111 const VkAllocationCallbacks
* pAllocator
)
2113 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2114 vk_destroy_debug_report_callback(&instance
->debug_report_callbacks
,
2115 _callback
, pAllocator
, &instance
->alloc
);
2119 anv_DebugReportMessageEXT(VkInstance _instance
,
2120 VkDebugReportFlagsEXT flags
,
2121 VkDebugReportObjectTypeEXT objectType
,
2124 int32_t messageCode
,
2125 const char* pLayerPrefix
,
2126 const char* pMessage
)
2128 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2129 vk_debug_report(&instance
->debug_report_callbacks
, flags
, objectType
,
2130 object
, location
, messageCode
, pLayerPrefix
, pMessage
);
2133 static struct anv_state
2134 anv_state_pool_emit_data(struct anv_state_pool
*pool
, size_t size
, size_t align
, const void *p
)
2136 struct anv_state state
;
2138 state
= anv_state_pool_alloc(pool
, size
, align
);
2139 memcpy(state
.map
, p
, size
);
2144 /* Haswell border color is a bit of a disaster. Float and unorm formats use a
2145 * straightforward 32-bit float color in the first 64 bytes. Instead of using
2146 * a nice float/integer union like Gen8+, Haswell specifies the integer border
2147 * color as a separate entry /after/ the float color. The layout of this entry
2148 * also depends on the format's bpp (with extra hacks for RG32), and overlaps.
2150 * Since we don't know the format/bpp, we can't make any of the border colors
2151 * containing '1' work for all formats, as it would be in the wrong place for
2152 * some of them. We opt to make 32-bit integers work as this seems like the
2153 * most common option. Fortunately, transparent black works regardless, as
2154 * all zeroes is the same in every bit-size.
2156 struct hsw_border_color
{
2160 uint32_t _pad1
[108];
2163 struct gen8_border_color
{
2168 /* Pad out to 64 bytes */
2173 anv_device_init_border_colors(struct anv_device
*device
)
2175 if (device
->info
.is_haswell
) {
2176 static const struct hsw_border_color border_colors
[] = {
2177 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
2178 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
2179 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
2180 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
2181 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
2182 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
2185 device
->border_colors
=
2186 anv_state_pool_emit_data(&device
->dynamic_state_pool
,
2187 sizeof(border_colors
), 512, border_colors
);
2189 static const struct gen8_border_color border_colors
[] = {
2190 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
2191 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
2192 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
2193 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
2194 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
2195 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
2198 device
->border_colors
=
2199 anv_state_pool_emit_data(&device
->dynamic_state_pool
,
2200 sizeof(border_colors
), 64, border_colors
);
2205 anv_device_init_trivial_batch(struct anv_device
*device
)
2207 VkResult result
= anv_device_alloc_bo(device
, 4096,
2208 ANV_BO_ALLOC_MAPPED
,
2209 &device
->trivial_batch_bo
);
2210 if (result
!= VK_SUCCESS
)
2213 struct anv_batch batch
= {
2214 .start
= device
->trivial_batch_bo
->map
,
2215 .next
= device
->trivial_batch_bo
->map
,
2216 .end
= device
->trivial_batch_bo
->map
+ 4096,
2219 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
2220 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
2222 if (!device
->info
.has_llc
)
2223 gen_clflush_range(batch
.start
, batch
.next
- batch
.start
);
2228 VkResult
anv_EnumerateDeviceExtensionProperties(
2229 VkPhysicalDevice physicalDevice
,
2230 const char* pLayerName
,
2231 uint32_t* pPropertyCount
,
2232 VkExtensionProperties
* pProperties
)
2234 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
2235 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
2237 for (int i
= 0; i
< ANV_DEVICE_EXTENSION_COUNT
; i
++) {
2238 if (device
->supported_extensions
.extensions
[i
]) {
2239 vk_outarray_append(&out
, prop
) {
2240 *prop
= anv_device_extensions
[i
];
2245 return vk_outarray_status(&out
);
2249 anv_device_init_dispatch(struct anv_device
*device
)
2251 const struct anv_device_dispatch_table
*genX_table
;
2252 switch (device
->info
.gen
) {
2254 genX_table
= &gen12_device_dispatch_table
;
2257 genX_table
= &gen11_device_dispatch_table
;
2260 genX_table
= &gen10_device_dispatch_table
;
2263 genX_table
= &gen9_device_dispatch_table
;
2266 genX_table
= &gen8_device_dispatch_table
;
2269 if (device
->info
.is_haswell
)
2270 genX_table
= &gen75_device_dispatch_table
;
2272 genX_table
= &gen7_device_dispatch_table
;
2275 unreachable("unsupported gen\n");
2278 for (unsigned i
= 0; i
< ARRAY_SIZE(device
->dispatch
.entrypoints
); i
++) {
2279 /* Vulkan requires that entrypoints for extensions which have not been
2280 * enabled must not be advertised.
2282 if (!anv_device_entrypoint_is_enabled(i
, device
->instance
->app_info
.api_version
,
2283 &device
->instance
->enabled_extensions
,
2284 &device
->enabled_extensions
)) {
2285 device
->dispatch
.entrypoints
[i
] = NULL
;
2286 } else if (genX_table
->entrypoints
[i
]) {
2287 device
->dispatch
.entrypoints
[i
] = genX_table
->entrypoints
[i
];
2289 device
->dispatch
.entrypoints
[i
] =
2290 anv_device_dispatch_table
.entrypoints
[i
];
2296 vk_priority_to_gen(int priority
)
2299 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT
:
2300 return GEN_CONTEXT_LOW_PRIORITY
;
2301 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
:
2302 return GEN_CONTEXT_MEDIUM_PRIORITY
;
2303 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT
:
2304 return GEN_CONTEXT_HIGH_PRIORITY
;
2305 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT
:
2306 return GEN_CONTEXT_REALTIME_PRIORITY
;
2308 unreachable("Invalid priority");
2313 anv_device_init_hiz_clear_value_bo(struct anv_device
*device
)
2315 VkResult result
= anv_device_alloc_bo(device
, 4096,
2316 ANV_BO_ALLOC_MAPPED
,
2317 &device
->hiz_clear_bo
);
2318 if (result
!= VK_SUCCESS
)
2321 union isl_color_value hiz_clear
= { .u32
= { 0, } };
2322 hiz_clear
.f32
[0] = ANV_HZ_FC_VAL
;
2324 memcpy(device
->hiz_clear_bo
->map
, hiz_clear
.u32
, sizeof(hiz_clear
.u32
));
2326 if (!device
->info
.has_llc
)
2327 gen_clflush_range(device
->hiz_clear_bo
->map
, sizeof(hiz_clear
.u32
));
2333 get_bo_from_pool(struct gen_batch_decode_bo
*ret
,
2334 struct anv_block_pool
*pool
,
2337 anv_block_pool_foreach_bo(bo
, pool
) {
2338 uint64_t bo_address
= gen_48b_address(bo
->offset
);
2339 if (address
>= bo_address
&& address
< (bo_address
+ bo
->size
)) {
2340 *ret
= (struct gen_batch_decode_bo
) {
2351 /* Finding a buffer for batch decoding */
2352 static struct gen_batch_decode_bo
2353 decode_get_bo(void *v_batch
, bool ppgtt
, uint64_t address
)
2355 struct anv_device
*device
= v_batch
;
2356 struct gen_batch_decode_bo ret_bo
= {};
2360 if (get_bo_from_pool(&ret_bo
, &device
->dynamic_state_pool
.block_pool
, address
))
2362 if (get_bo_from_pool(&ret_bo
, &device
->instruction_state_pool
.block_pool
, address
))
2364 if (get_bo_from_pool(&ret_bo
, &device
->binding_table_pool
.block_pool
, address
))
2366 if (get_bo_from_pool(&ret_bo
, &device
->surface_state_pool
.block_pool
, address
))
2369 if (!device
->cmd_buffer_being_decoded
)
2370 return (struct gen_batch_decode_bo
) { };
2372 struct anv_batch_bo
**bo
;
2374 u_vector_foreach(bo
, &device
->cmd_buffer_being_decoded
->seen_bbos
) {
2375 /* The decoder zeroes out the top 16 bits, so we need to as well */
2376 uint64_t bo_address
= (*bo
)->bo
->offset
& (~0ull >> 16);
2378 if (address
>= bo_address
&& address
< bo_address
+ (*bo
)->bo
->size
) {
2379 return (struct gen_batch_decode_bo
) {
2381 .size
= (*bo
)->bo
->size
,
2382 .map
= (*bo
)->bo
->map
,
2387 return (struct gen_batch_decode_bo
) { };
2390 struct gen_aux_map_buffer
{
2391 struct gen_buffer base
;
2392 struct anv_state state
;
2395 static struct gen_buffer
*
2396 gen_aux_map_buffer_alloc(void *driver_ctx
, uint32_t size
)
2398 struct gen_aux_map_buffer
*buf
= malloc(sizeof(struct gen_aux_map_buffer
));
2402 struct anv_device
*device
= (struct anv_device
*)driver_ctx
;
2403 assert(device
->instance
->physicalDevice
.supports_48bit_addresses
&&
2404 device
->instance
->physicalDevice
.use_softpin
);
2406 struct anv_state_pool
*pool
= &device
->dynamic_state_pool
;
2407 buf
->state
= anv_state_pool_alloc(pool
, size
, size
);
2409 buf
->base
.gpu
= pool
->block_pool
.bo
->offset
+ buf
->state
.offset
;
2410 buf
->base
.gpu_end
= buf
->base
.gpu
+ buf
->state
.alloc_size
;
2411 buf
->base
.map
= buf
->state
.map
;
2412 buf
->base
.driver_bo
= &buf
->state
;
2417 gen_aux_map_buffer_free(void *driver_ctx
, struct gen_buffer
*buffer
)
2419 struct gen_aux_map_buffer
*buf
= (struct gen_aux_map_buffer
*)buffer
;
2420 struct anv_device
*device
= (struct anv_device
*)driver_ctx
;
2421 struct anv_state_pool
*pool
= &device
->dynamic_state_pool
;
2422 anv_state_pool_free(pool
, buf
->state
);
2426 static struct gen_mapped_pinned_buffer_alloc aux_map_allocator
= {
2427 .alloc
= gen_aux_map_buffer_alloc
,
2428 .free
= gen_aux_map_buffer_free
,
2431 VkResult
anv_CreateDevice(
2432 VkPhysicalDevice physicalDevice
,
2433 const VkDeviceCreateInfo
* pCreateInfo
,
2434 const VkAllocationCallbacks
* pAllocator
,
2437 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
2439 struct anv_device
*device
;
2441 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO
);
2443 struct anv_device_extension_table enabled_extensions
= { };
2444 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
2446 for (idx
= 0; idx
< ANV_DEVICE_EXTENSION_COUNT
; idx
++) {
2447 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
2448 anv_device_extensions
[idx
].extensionName
) == 0)
2452 if (idx
>= ANV_DEVICE_EXTENSION_COUNT
)
2453 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
2455 if (!physical_device
->supported_extensions
.extensions
[idx
])
2456 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
2458 enabled_extensions
.extensions
[idx
] = true;
2461 /* Check enabled features */
2462 if (pCreateInfo
->pEnabledFeatures
) {
2463 VkPhysicalDeviceFeatures supported_features
;
2464 anv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
2465 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
2466 VkBool32
*enabled_feature
= (VkBool32
*)pCreateInfo
->pEnabledFeatures
;
2467 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
2468 for (uint32_t i
= 0; i
< num_features
; i
++) {
2469 if (enabled_feature
[i
] && !supported_feature
[i
])
2470 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
2474 /* Check requested queues and fail if we are requested to create any
2475 * queues with flags we don't support.
2477 assert(pCreateInfo
->queueCreateInfoCount
> 0);
2478 for (uint32_t i
= 0; i
< pCreateInfo
->queueCreateInfoCount
; i
++) {
2479 if (pCreateInfo
->pQueueCreateInfos
[i
].flags
!= 0)
2480 return vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2483 /* Check if client specified queue priority. */
2484 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*queue_priority
=
2485 vk_find_struct_const(pCreateInfo
->pQueueCreateInfos
[0].pNext
,
2486 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT
);
2488 VkQueueGlobalPriorityEXT priority
=
2489 queue_priority
? queue_priority
->globalPriority
:
2490 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
;
2492 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
2494 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
2496 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2498 if (INTEL_DEBUG
& DEBUG_BATCH
) {
2499 const unsigned decode_flags
=
2500 GEN_BATCH_DECODE_FULL
|
2501 ((INTEL_DEBUG
& DEBUG_COLOR
) ? GEN_BATCH_DECODE_IN_COLOR
: 0) |
2502 GEN_BATCH_DECODE_OFFSETS
|
2503 GEN_BATCH_DECODE_FLOATS
;
2505 gen_batch_decode_ctx_init(&device
->decoder_ctx
,
2506 &physical_device
->info
,
2507 stderr
, decode_flags
, NULL
,
2508 decode_get_bo
, NULL
, device
);
2511 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
2512 device
->instance
= physical_device
->instance
;
2513 device
->chipset_id
= physical_device
->chipset_id
;
2514 device
->no_hw
= physical_device
->no_hw
;
2515 device
->_lost
= false;
2518 device
->alloc
= *pAllocator
;
2520 device
->alloc
= physical_device
->instance
->alloc
;
2522 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
2523 device
->fd
= open(physical_device
->path
, O_RDWR
| O_CLOEXEC
);
2524 if (device
->fd
== -1) {
2525 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2529 device
->context_id
= anv_gem_create_context(device
);
2530 if (device
->context_id
== -1) {
2531 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2535 result
= anv_queue_init(device
, &device
->queue
);
2536 if (result
!= VK_SUCCESS
)
2537 goto fail_context_id
;
2539 if (physical_device
->use_softpin
) {
2540 if (pthread_mutex_init(&device
->vma_mutex
, NULL
) != 0) {
2541 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2545 /* keep the page with address zero out of the allocator */
2546 util_vma_heap_init(&device
->vma_lo
,
2547 LOW_HEAP_MIN_ADDRESS
, LOW_HEAP_SIZE
);
2549 /* Leave the last 4GiB out of the high vma range, so that no state
2550 * base address + size can overflow 48 bits. For more information see
2551 * the comment about Wa32bitGeneralStateOffset in anv_allocator.c
2553 util_vma_heap_init(&device
->vma_hi
, HIGH_HEAP_MIN_ADDRESS
,
2554 physical_device
->gtt_size
- (1ull << 32) -
2555 HIGH_HEAP_MIN_ADDRESS
);
2558 list_inithead(&device
->memory_objects
);
2560 /* As per spec, the driver implementation may deny requests to acquire
2561 * a priority above the default priority (MEDIUM) if the caller does not
2562 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
2565 if (physical_device
->has_context_priority
) {
2566 int err
= anv_gem_set_context_param(device
->fd
, device
->context_id
,
2567 I915_CONTEXT_PARAM_PRIORITY
,
2568 vk_priority_to_gen(priority
));
2569 if (err
!= 0 && priority
> VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
) {
2570 result
= vk_error(VK_ERROR_NOT_PERMITTED_EXT
);
2575 device
->info
= physical_device
->info
;
2576 device
->isl_dev
= physical_device
->isl_dev
;
2578 /* On Broadwell and later, we can use batch chaining to more efficiently
2579 * implement growing command buffers. Prior to Haswell, the kernel
2580 * command parser gets in the way and we have to fall back to growing
2583 device
->can_chain_batches
= device
->info
.gen
>= 8;
2585 device
->robust_buffer_access
= pCreateInfo
->pEnabledFeatures
&&
2586 pCreateInfo
->pEnabledFeatures
->robustBufferAccess
;
2587 device
->enabled_extensions
= enabled_extensions
;
2589 anv_device_init_dispatch(device
);
2591 if (pthread_mutex_init(&device
->mutex
, NULL
) != 0) {
2592 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2596 pthread_condattr_t condattr
;
2597 if (pthread_condattr_init(&condattr
) != 0) {
2598 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2601 if (pthread_condattr_setclock(&condattr
, CLOCK_MONOTONIC
) != 0) {
2602 pthread_condattr_destroy(&condattr
);
2603 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2606 if (pthread_cond_init(&device
->queue_submit
, &condattr
) != 0) {
2607 pthread_condattr_destroy(&condattr
);
2608 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2611 pthread_condattr_destroy(&condattr
);
2613 result
= anv_bo_cache_init(&device
->bo_cache
);
2614 if (result
!= VK_SUCCESS
)
2615 goto fail_queue_cond
;
2617 anv_bo_pool_init(&device
->batch_bo_pool
, device
);
2619 result
= anv_state_pool_init(&device
->dynamic_state_pool
, device
,
2620 DYNAMIC_STATE_POOL_MIN_ADDRESS
, 16384);
2621 if (result
!= VK_SUCCESS
)
2622 goto fail_batch_bo_pool
;
2624 result
= anv_state_pool_init(&device
->instruction_state_pool
, device
,
2625 INSTRUCTION_STATE_POOL_MIN_ADDRESS
, 16384);
2626 if (result
!= VK_SUCCESS
)
2627 goto fail_dynamic_state_pool
;
2629 result
= anv_state_pool_init(&device
->surface_state_pool
, device
,
2630 SURFACE_STATE_POOL_MIN_ADDRESS
, 4096);
2631 if (result
!= VK_SUCCESS
)
2632 goto fail_instruction_state_pool
;
2634 if (physical_device
->use_softpin
) {
2635 result
= anv_state_pool_init(&device
->binding_table_pool
, device
,
2636 BINDING_TABLE_POOL_MIN_ADDRESS
, 4096);
2637 if (result
!= VK_SUCCESS
)
2638 goto fail_surface_state_pool
;
2641 if (device
->info
.gen
>= 12) {
2642 device
->aux_map_ctx
= gen_aux_map_init(device
, &aux_map_allocator
,
2643 &physical_device
->info
);
2644 if (!device
->aux_map_ctx
)
2645 goto fail_binding_table_pool
;
2648 result
= anv_device_alloc_bo(device
, 4096, 0, &device
->workaround_bo
);
2649 if (result
!= VK_SUCCESS
)
2650 goto fail_surface_aux_map_pool
;
2652 result
= anv_device_init_trivial_batch(device
);
2653 if (result
!= VK_SUCCESS
)
2654 goto fail_workaround_bo
;
2656 if (device
->info
.gen
>= 10) {
2657 result
= anv_device_init_hiz_clear_value_bo(device
);
2658 if (result
!= VK_SUCCESS
)
2659 goto fail_trivial_batch_bo
;
2662 anv_scratch_pool_init(device
, &device
->scratch_pool
);
2664 switch (device
->info
.gen
) {
2666 if (!device
->info
.is_haswell
)
2667 result
= gen7_init_device_state(device
);
2669 result
= gen75_init_device_state(device
);
2672 result
= gen8_init_device_state(device
);
2675 result
= gen9_init_device_state(device
);
2678 result
= gen10_init_device_state(device
);
2681 result
= gen11_init_device_state(device
);
2684 result
= gen12_init_device_state(device
);
2687 /* Shouldn't get here as we don't create physical devices for any other
2689 unreachable("unhandled gen");
2691 if (result
!= VK_SUCCESS
)
2692 goto fail_workaround_bo
;
2694 anv_pipeline_cache_init(&device
->default_pipeline_cache
, device
, true);
2696 anv_device_init_blorp(device
);
2698 anv_device_init_border_colors(device
);
2700 anv_device_perf_init(device
);
2702 *pDevice
= anv_device_to_handle(device
);
2707 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
2708 if (device
->info
.gen
>= 10)
2709 anv_device_release_bo(device
, device
->hiz_clear_bo
);
2710 anv_device_release_bo(device
, device
->workaround_bo
);
2711 fail_trivial_batch_bo
:
2712 anv_device_release_bo(device
, device
->trivial_batch_bo
);
2713 fail_surface_aux_map_pool
:
2714 if (device
->info
.gen
>= 12) {
2715 gen_aux_map_finish(device
->aux_map_ctx
);
2716 device
->aux_map_ctx
= NULL
;
2718 fail_binding_table_pool
:
2719 if (physical_device
->use_softpin
)
2720 anv_state_pool_finish(&device
->binding_table_pool
);
2721 fail_surface_state_pool
:
2722 anv_state_pool_finish(&device
->surface_state_pool
);
2723 fail_instruction_state_pool
:
2724 anv_state_pool_finish(&device
->instruction_state_pool
);
2725 fail_dynamic_state_pool
:
2726 anv_state_pool_finish(&device
->dynamic_state_pool
);
2728 anv_bo_pool_finish(&device
->batch_bo_pool
);
2729 anv_bo_cache_finish(&device
->bo_cache
);
2731 pthread_cond_destroy(&device
->queue_submit
);
2733 pthread_mutex_destroy(&device
->mutex
);
2735 if (physical_device
->use_softpin
) {
2736 util_vma_heap_finish(&device
->vma_hi
);
2737 util_vma_heap_finish(&device
->vma_lo
);
2740 anv_queue_finish(&device
->queue
);
2742 anv_gem_destroy_context(device
, device
->context_id
);
2746 vk_free(&device
->alloc
, device
);
2751 void anv_DestroyDevice(
2753 const VkAllocationCallbacks
* pAllocator
)
2755 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2756 struct anv_physical_device
*physical_device
;
2761 physical_device
= &device
->instance
->physicalDevice
;
2763 anv_device_finish_blorp(device
);
2765 anv_pipeline_cache_finish(&device
->default_pipeline_cache
);
2767 anv_queue_finish(&device
->queue
);
2769 #ifdef HAVE_VALGRIND
2770 /* We only need to free these to prevent valgrind errors. The backing
2771 * BO will go away in a couple of lines so we don't actually leak.
2773 anv_state_pool_free(&device
->dynamic_state_pool
, device
->border_colors
);
2774 anv_state_pool_free(&device
->dynamic_state_pool
, device
->slice_hash
);
2777 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
2779 anv_device_release_bo(device
, device
->workaround_bo
);
2780 anv_device_release_bo(device
, device
->trivial_batch_bo
);
2781 if (device
->info
.gen
>= 10)
2782 anv_device_release_bo(device
, device
->hiz_clear_bo
);
2784 if (device
->info
.gen
>= 12) {
2785 gen_aux_map_finish(device
->aux_map_ctx
);
2786 device
->aux_map_ctx
= NULL
;
2789 if (physical_device
->use_softpin
)
2790 anv_state_pool_finish(&device
->binding_table_pool
);
2791 anv_state_pool_finish(&device
->surface_state_pool
);
2792 anv_state_pool_finish(&device
->instruction_state_pool
);
2793 anv_state_pool_finish(&device
->dynamic_state_pool
);
2795 anv_bo_pool_finish(&device
->batch_bo_pool
);
2797 anv_bo_cache_finish(&device
->bo_cache
);
2799 if (physical_device
->use_softpin
) {
2800 util_vma_heap_finish(&device
->vma_hi
);
2801 util_vma_heap_finish(&device
->vma_lo
);
2804 pthread_cond_destroy(&device
->queue_submit
);
2805 pthread_mutex_destroy(&device
->mutex
);
2807 anv_gem_destroy_context(device
, device
->context_id
);
2809 if (INTEL_DEBUG
& DEBUG_BATCH
)
2810 gen_batch_decode_ctx_finish(&device
->decoder_ctx
);
2814 vk_free(&device
->alloc
, device
);
2817 VkResult
anv_EnumerateInstanceLayerProperties(
2818 uint32_t* pPropertyCount
,
2819 VkLayerProperties
* pProperties
)
2821 if (pProperties
== NULL
) {
2822 *pPropertyCount
= 0;
2826 /* None supported at this time */
2827 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
2830 VkResult
anv_EnumerateDeviceLayerProperties(
2831 VkPhysicalDevice physicalDevice
,
2832 uint32_t* pPropertyCount
,
2833 VkLayerProperties
* pProperties
)
2835 if (pProperties
== NULL
) {
2836 *pPropertyCount
= 0;
2840 /* None supported at this time */
2841 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
2844 void anv_GetDeviceQueue(
2846 uint32_t queueNodeIndex
,
2847 uint32_t queueIndex
,
2850 const VkDeviceQueueInfo2 info
= {
2851 .sType
= VK_STRUCTURE_TYPE_DEVICE_QUEUE_INFO_2
,
2854 .queueFamilyIndex
= queueNodeIndex
,
2855 .queueIndex
= queueIndex
,
2858 anv_GetDeviceQueue2(_device
, &info
, pQueue
);
2861 void anv_GetDeviceQueue2(
2863 const VkDeviceQueueInfo2
* pQueueInfo
,
2866 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2868 assert(pQueueInfo
->queueIndex
== 0);
2870 if (pQueueInfo
->flags
== device
->queue
.flags
)
2871 *pQueue
= anv_queue_to_handle(&device
->queue
);
2877 _anv_device_set_lost(struct anv_device
*device
,
2878 const char *file
, int line
,
2879 const char *msg
, ...)
2884 p_atomic_inc(&device
->_lost
);
2887 err
= __vk_errorv(device
->instance
, device
,
2888 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT
,
2889 VK_ERROR_DEVICE_LOST
, file
, line
, msg
, ap
);
2892 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
2899 _anv_queue_set_lost(struct anv_queue
*queue
,
2900 const char *file
, int line
,
2901 const char *msg
, ...)
2906 p_atomic_inc(&queue
->device
->_lost
);
2909 err
= __vk_errorv(queue
->device
->instance
, queue
->device
,
2910 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT
,
2911 VK_ERROR_DEVICE_LOST
, file
, line
, msg
, ap
);
2914 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
2921 anv_device_query_status(struct anv_device
*device
)
2923 /* This isn't likely as most of the callers of this function already check
2924 * for it. However, it doesn't hurt to check and it potentially lets us
2927 if (anv_device_is_lost(device
))
2928 return VK_ERROR_DEVICE_LOST
;
2930 uint32_t active
, pending
;
2931 int ret
= anv_gem_gpu_get_reset_stats(device
, &active
, &pending
);
2933 /* We don't know the real error. */
2934 return anv_device_set_lost(device
, "get_reset_stats failed: %m");
2938 return anv_device_set_lost(device
, "GPU hung on one of our command buffers");
2939 } else if (pending
) {
2940 return anv_device_set_lost(device
, "GPU hung with commands in-flight");
2947 anv_device_bo_busy(struct anv_device
*device
, struct anv_bo
*bo
)
2949 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
2950 * Other usages of the BO (such as on different hardware) will not be
2951 * flagged as "busy" by this ioctl. Use with care.
2953 int ret
= anv_gem_busy(device
, bo
->gem_handle
);
2955 return VK_NOT_READY
;
2956 } else if (ret
== -1) {
2957 /* We don't know the real error. */
2958 return anv_device_set_lost(device
, "gem wait failed: %m");
2961 /* Query for device status after the busy call. If the BO we're checking
2962 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
2963 * client because it clearly doesn't have valid data. Yes, this most
2964 * likely means an ioctl, but we just did an ioctl to query the busy status
2965 * so it's no great loss.
2967 return anv_device_query_status(device
);
2971 anv_device_wait(struct anv_device
*device
, struct anv_bo
*bo
,
2974 int ret
= anv_gem_wait(device
, bo
->gem_handle
, &timeout
);
2975 if (ret
== -1 && errno
== ETIME
) {
2977 } else if (ret
== -1) {
2978 /* We don't know the real error. */
2979 return anv_device_set_lost(device
, "gem wait failed: %m");
2982 /* Query for device status after the wait. If the BO we're waiting on got
2983 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
2984 * because it clearly doesn't have valid data. Yes, this most likely means
2985 * an ioctl, but we just did an ioctl to wait so it's no great loss.
2987 return anv_device_query_status(device
);
2990 VkResult
anv_DeviceWaitIdle(
2993 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2995 if (anv_device_is_lost(device
))
2996 return VK_ERROR_DEVICE_LOST
;
2998 return anv_queue_submit_simple_batch(&device
->queue
, NULL
);
3002 anv_vma_alloc(struct anv_device
*device
, struct anv_bo
*bo
)
3004 if (!(bo
->flags
& EXEC_OBJECT_PINNED
))
3007 pthread_mutex_lock(&device
->vma_mutex
);
3011 if (bo
->flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
) {
3012 uint64_t addr
= util_vma_heap_alloc(&device
->vma_hi
, bo
->size
, 4096);
3014 bo
->offset
= gen_canonical_address(addr
);
3015 assert(addr
== gen_48b_address(bo
->offset
));
3019 if (bo
->offset
== 0) {
3020 uint64_t addr
= util_vma_heap_alloc(&device
->vma_lo
, bo
->size
, 4096);
3022 bo
->offset
= gen_canonical_address(addr
);
3023 assert(addr
== gen_48b_address(bo
->offset
));
3027 pthread_mutex_unlock(&device
->vma_mutex
);
3029 return bo
->offset
!= 0;
3033 anv_vma_free(struct anv_device
*device
, struct anv_bo
*bo
)
3035 if (!(bo
->flags
& EXEC_OBJECT_PINNED
))
3038 const uint64_t addr_48b
= gen_48b_address(bo
->offset
);
3040 pthread_mutex_lock(&device
->vma_mutex
);
3042 if (addr_48b
>= LOW_HEAP_MIN_ADDRESS
&&
3043 addr_48b
<= LOW_HEAP_MAX_ADDRESS
) {
3044 util_vma_heap_free(&device
->vma_lo
, addr_48b
, bo
->size
);
3046 assert(addr_48b
>= HIGH_HEAP_MIN_ADDRESS
);
3047 util_vma_heap_free(&device
->vma_hi
, addr_48b
, bo
->size
);
3050 pthread_mutex_unlock(&device
->vma_mutex
);
3055 VkResult
anv_AllocateMemory(
3057 const VkMemoryAllocateInfo
* pAllocateInfo
,
3058 const VkAllocationCallbacks
* pAllocator
,
3059 VkDeviceMemory
* pMem
)
3061 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3062 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3063 struct anv_device_memory
*mem
;
3064 VkResult result
= VK_SUCCESS
;
3066 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
3068 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
3069 assert(pAllocateInfo
->allocationSize
> 0);
3071 VkDeviceSize aligned_alloc_size
=
3072 align_u64(pAllocateInfo
->allocationSize
, 4096);
3074 if (aligned_alloc_size
> MAX_MEMORY_ALLOCATION_SIZE
)
3075 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
3077 assert(pAllocateInfo
->memoryTypeIndex
< pdevice
->memory
.type_count
);
3078 struct anv_memory_type
*mem_type
=
3079 &pdevice
->memory
.types
[pAllocateInfo
->memoryTypeIndex
];
3080 assert(mem_type
->heapIndex
< pdevice
->memory
.heap_count
);
3081 struct anv_memory_heap
*mem_heap
=
3082 &pdevice
->memory
.heaps
[mem_type
->heapIndex
];
3084 uint64_t mem_heap_used
= p_atomic_read(&mem_heap
->used
);
3085 if (mem_heap_used
+ aligned_alloc_size
> mem_heap
->size
)
3086 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
3088 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
3089 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3091 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3093 assert(pAllocateInfo
->memoryTypeIndex
< pdevice
->memory
.type_count
);
3094 mem
->type
= mem_type
;
3098 mem
->host_ptr
= NULL
;
3100 enum anv_bo_alloc_flags alloc_flags
= 0;
3102 if (!mem_heap
->supports_48bit_addresses
)
3103 alloc_flags
|= ANV_BO_ALLOC_32BIT_ADDRESS
;
3105 const struct wsi_memory_allocate_info
*wsi_info
=
3106 vk_find_struct_const(pAllocateInfo
->pNext
, WSI_MEMORY_ALLOCATE_INFO_MESA
);
3107 if (wsi_info
&& wsi_info
->implicit_sync
) {
3108 /* We need to set the WRITE flag on window system buffers so that GEM
3109 * will know we're writing to them and synchronize uses on other rings
3110 * (eg if the display server uses the blitter ring).
3112 alloc_flags
|= ANV_BO_ALLOC_IMPLICIT_SYNC
|
3113 ANV_BO_ALLOC_IMPLICIT_WRITE
;
3116 const VkExportMemoryAllocateInfo
*export_info
=
3117 vk_find_struct_const(pAllocateInfo
->pNext
, EXPORT_MEMORY_ALLOCATE_INFO
);
3119 /* Check if we need to support Android HW buffer export. If so,
3120 * create AHardwareBuffer and import memory from it.
3122 bool android_export
= false;
3123 if (export_info
&& export_info
->handleTypes
&
3124 VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID
)
3125 android_export
= true;
3127 /* Android memory import. */
3128 const struct VkImportAndroidHardwareBufferInfoANDROID
*ahw_import_info
=
3129 vk_find_struct_const(pAllocateInfo
->pNext
,
3130 IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID
);
3132 if (ahw_import_info
) {
3133 result
= anv_import_ahw_memory(_device
, mem
, ahw_import_info
);
3134 if (result
!= VK_SUCCESS
)
3138 } else if (android_export
) {
3139 result
= anv_create_ahw_memory(_device
, mem
, pAllocateInfo
);
3140 if (result
!= VK_SUCCESS
)
3143 const struct VkImportAndroidHardwareBufferInfoANDROID import_info
= {
3146 result
= anv_import_ahw_memory(_device
, mem
, &import_info
);
3147 if (result
!= VK_SUCCESS
)
3153 const VkImportMemoryFdInfoKHR
*fd_info
=
3154 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_FD_INFO_KHR
);
3156 /* The Vulkan spec permits handleType to be 0, in which case the struct is
3159 if (fd_info
&& fd_info
->handleType
) {
3160 /* At the moment, we support only the below handle types. */
3161 assert(fd_info
->handleType
==
3162 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
3163 fd_info
->handleType
==
3164 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
3166 result
= anv_device_import_bo(device
, fd_info
->fd
, alloc_flags
,
3168 if (result
!= VK_SUCCESS
)
3171 VkDeviceSize aligned_alloc_size
=
3172 align_u64(pAllocateInfo
->allocationSize
, 4096);
3174 /* For security purposes, we reject importing the bo if it's smaller
3175 * than the requested allocation size. This prevents a malicious client
3176 * from passing a buffer to a trusted client, lying about the size, and
3177 * telling the trusted client to try and texture from an image that goes
3178 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
3179 * in the trusted client. The trusted client can protect itself against
3180 * this sort of attack but only if it can trust the buffer size.
3182 if (mem
->bo
->size
< aligned_alloc_size
) {
3183 result
= vk_errorf(device
->instance
, device
,
3184 VK_ERROR_INVALID_EXTERNAL_HANDLE
,
3185 "aligned allocationSize too large for "
3186 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT: "
3187 "%"PRIu64
"B > %"PRIu64
"B",
3188 aligned_alloc_size
, mem
->bo
->size
);
3189 anv_device_release_bo(device
, mem
->bo
);
3193 /* From the Vulkan spec:
3195 * "Importing memory from a file descriptor transfers ownership of
3196 * the file descriptor from the application to the Vulkan
3197 * implementation. The application must not perform any operations on
3198 * the file descriptor after a successful import."
3200 * If the import fails, we leave the file descriptor open.
3206 const VkImportMemoryHostPointerInfoEXT
*host_ptr_info
=
3207 vk_find_struct_const(pAllocateInfo
->pNext
,
3208 IMPORT_MEMORY_HOST_POINTER_INFO_EXT
);
3209 if (host_ptr_info
&& host_ptr_info
->handleType
) {
3210 if (host_ptr_info
->handleType
==
3211 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_MAPPED_FOREIGN_MEMORY_BIT_EXT
) {
3212 result
= vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
3216 assert(host_ptr_info
->handleType
==
3217 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
);
3219 result
= anv_device_import_bo_from_host_ptr(device
,
3220 host_ptr_info
->pHostPointer
,
3221 pAllocateInfo
->allocationSize
,
3225 if (result
!= VK_SUCCESS
)
3228 mem
->host_ptr
= host_ptr_info
->pHostPointer
;
3232 /* Regular allocate (not importing memory). */
3234 if (export_info
&& export_info
->handleTypes
)
3235 alloc_flags
|= ANV_BO_ALLOC_EXTERNAL
;
3237 result
= anv_device_alloc_bo(device
, pAllocateInfo
->allocationSize
,
3238 alloc_flags
, &mem
->bo
);
3239 if (result
!= VK_SUCCESS
)
3242 const VkMemoryDedicatedAllocateInfo
*dedicated_info
=
3243 vk_find_struct_const(pAllocateInfo
->pNext
, MEMORY_DEDICATED_ALLOCATE_INFO
);
3244 if (dedicated_info
&& dedicated_info
->image
!= VK_NULL_HANDLE
) {
3245 ANV_FROM_HANDLE(anv_image
, image
, dedicated_info
->image
);
3247 /* Some legacy (non-modifiers) consumers need the tiling to be set on
3248 * the BO. In this case, we have a dedicated allocation.
3250 if (image
->needs_set_tiling
) {
3251 const uint32_t i915_tiling
=
3252 isl_tiling_to_i915_tiling(image
->planes
[0].surface
.isl
.tiling
);
3253 int ret
= anv_gem_set_tiling(device
, mem
->bo
->gem_handle
,
3254 image
->planes
[0].surface
.isl
.row_pitch_B
,
3257 anv_device_release_bo(device
, mem
->bo
);
3258 result
= vk_errorf(device
->instance
, NULL
,
3259 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
3260 "failed to set BO tiling: %m");
3267 mem_heap_used
= p_atomic_add_return(&mem_heap
->used
, mem
->bo
->size
);
3268 if (mem_heap_used
> mem_heap
->size
) {
3269 p_atomic_add(&mem_heap
->used
, -mem
->bo
->size
);
3270 anv_device_release_bo(device
, mem
->bo
);
3271 result
= vk_errorf(device
->instance
, NULL
,
3272 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
3273 "Out of heap memory");
3277 pthread_mutex_lock(&device
->mutex
);
3278 list_addtail(&mem
->link
, &device
->memory_objects
);
3279 pthread_mutex_unlock(&device
->mutex
);
3281 *pMem
= anv_device_memory_to_handle(mem
);
3286 vk_free2(&device
->alloc
, pAllocator
, mem
);
3291 VkResult
anv_GetMemoryFdKHR(
3293 const VkMemoryGetFdInfoKHR
* pGetFdInfo
,
3296 ANV_FROM_HANDLE(anv_device
, dev
, device_h
);
3297 ANV_FROM_HANDLE(anv_device_memory
, mem
, pGetFdInfo
->memory
);
3299 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
3301 assert(pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
3302 pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
3304 return anv_device_export_bo(dev
, mem
->bo
, pFd
);
3307 VkResult
anv_GetMemoryFdPropertiesKHR(
3309 VkExternalMemoryHandleTypeFlagBits handleType
,
3311 VkMemoryFdPropertiesKHR
* pMemoryFdProperties
)
3313 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3314 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3316 switch (handleType
) {
3317 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
:
3318 /* dma-buf can be imported as any memory type */
3319 pMemoryFdProperties
->memoryTypeBits
=
3320 (1 << pdevice
->memory
.type_count
) - 1;
3324 /* The valid usage section for this function says:
3326 * "handleType must not be one of the handle types defined as
3329 * So opaque handle types fall into the default "unsupported" case.
3331 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
3335 VkResult
anv_GetMemoryHostPointerPropertiesEXT(
3337 VkExternalMemoryHandleTypeFlagBits handleType
,
3338 const void* pHostPointer
,
3339 VkMemoryHostPointerPropertiesEXT
* pMemoryHostPointerProperties
)
3341 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3343 assert(pMemoryHostPointerProperties
->sType
==
3344 VK_STRUCTURE_TYPE_MEMORY_HOST_POINTER_PROPERTIES_EXT
);
3346 switch (handleType
) {
3347 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
: {
3348 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3350 /* Host memory can be imported as any memory type. */
3351 pMemoryHostPointerProperties
->memoryTypeBits
=
3352 (1ull << pdevice
->memory
.type_count
) - 1;
3357 return VK_ERROR_INVALID_EXTERNAL_HANDLE
;
3361 void anv_FreeMemory(
3363 VkDeviceMemory _mem
,
3364 const VkAllocationCallbacks
* pAllocator
)
3366 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3367 ANV_FROM_HANDLE(anv_device_memory
, mem
, _mem
);
3368 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3373 pthread_mutex_lock(&device
->mutex
);
3374 list_del(&mem
->link
);
3375 pthread_mutex_unlock(&device
->mutex
);
3378 anv_UnmapMemory(_device
, _mem
);
3380 p_atomic_add(&pdevice
->memory
.heaps
[mem
->type
->heapIndex
].used
,
3383 anv_device_release_bo(device
, mem
->bo
);
3385 #if defined(ANDROID) && ANDROID_API_LEVEL >= 26
3387 AHardwareBuffer_release(mem
->ahw
);
3390 vk_free2(&device
->alloc
, pAllocator
, mem
);
3393 VkResult
anv_MapMemory(
3395 VkDeviceMemory _memory
,
3396 VkDeviceSize offset
,
3398 VkMemoryMapFlags flags
,
3401 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3402 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
3409 if (mem
->host_ptr
) {
3410 *ppData
= mem
->host_ptr
+ offset
;
3414 if (size
== VK_WHOLE_SIZE
)
3415 size
= mem
->bo
->size
- offset
;
3417 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
3419 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
3420 * assert(size != 0);
3421 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
3422 * equal to the size of the memory minus offset
3425 assert(offset
+ size
<= mem
->bo
->size
);
3427 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
3428 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
3429 * at a time is valid. We could just mmap up front and return an offset
3430 * pointer here, but that may exhaust virtual memory on 32 bit
3433 uint32_t gem_flags
= 0;
3435 if (!device
->info
.has_llc
&&
3436 (mem
->type
->propertyFlags
& VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
))
3437 gem_flags
|= I915_MMAP_WC
;
3439 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
3440 uint64_t map_offset
= offset
& ~4095ull;
3441 assert(offset
>= map_offset
);
3442 uint64_t map_size
= (offset
+ size
) - map_offset
;
3444 /* Let's map whole pages */
3445 map_size
= align_u64(map_size
, 4096);
3447 void *map
= anv_gem_mmap(device
, mem
->bo
->gem_handle
,
3448 map_offset
, map_size
, gem_flags
);
3449 if (map
== MAP_FAILED
)
3450 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
3453 mem
->map_size
= map_size
;
3455 *ppData
= mem
->map
+ (offset
- map_offset
);
3460 void anv_UnmapMemory(
3462 VkDeviceMemory _memory
)
3464 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
3466 if (mem
== NULL
|| mem
->host_ptr
)
3469 anv_gem_munmap(mem
->map
, mem
->map_size
);
3476 clflush_mapped_ranges(struct anv_device
*device
,
3478 const VkMappedMemoryRange
*ranges
)
3480 for (uint32_t i
= 0; i
< count
; i
++) {
3481 ANV_FROM_HANDLE(anv_device_memory
, mem
, ranges
[i
].memory
);
3482 if (ranges
[i
].offset
>= mem
->map_size
)
3485 gen_clflush_range(mem
->map
+ ranges
[i
].offset
,
3486 MIN2(ranges
[i
].size
, mem
->map_size
- ranges
[i
].offset
));
3490 VkResult
anv_FlushMappedMemoryRanges(
3492 uint32_t memoryRangeCount
,
3493 const VkMappedMemoryRange
* pMemoryRanges
)
3495 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3497 if (device
->info
.has_llc
)
3500 /* Make sure the writes we're flushing have landed. */
3501 __builtin_ia32_mfence();
3503 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
3508 VkResult
anv_InvalidateMappedMemoryRanges(
3510 uint32_t memoryRangeCount
,
3511 const VkMappedMemoryRange
* pMemoryRanges
)
3513 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3515 if (device
->info
.has_llc
)
3518 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
3520 /* Make sure no reads get moved up above the invalidate. */
3521 __builtin_ia32_mfence();
3526 void anv_GetBufferMemoryRequirements(
3529 VkMemoryRequirements
* pMemoryRequirements
)
3531 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
3532 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3533 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3535 /* The Vulkan spec (git aaed022) says:
3537 * memoryTypeBits is a bitfield and contains one bit set for every
3538 * supported memory type for the resource. The bit `1<<i` is set if and
3539 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3540 * structure for the physical device is supported.
3542 uint32_t memory_types
= 0;
3543 for (uint32_t i
= 0; i
< pdevice
->memory
.type_count
; i
++) {
3544 uint32_t valid_usage
= pdevice
->memory
.types
[i
].valid_buffer_usage
;
3545 if ((valid_usage
& buffer
->usage
) == buffer
->usage
)
3546 memory_types
|= (1u << i
);
3549 /* Base alignment requirement of a cache line */
3550 uint32_t alignment
= 16;
3552 /* We need an alignment of 32 for pushing UBOs */
3553 if (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
)
3554 alignment
= MAX2(alignment
, 32);
3556 pMemoryRequirements
->size
= buffer
->size
;
3557 pMemoryRequirements
->alignment
= alignment
;
3559 /* Storage and Uniform buffers should have their size aligned to
3560 * 32-bits to avoid boundary checks when last DWord is not complete.
3561 * This would ensure that not internal padding would be needed for
3564 if (device
->robust_buffer_access
&&
3565 (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
||
3566 buffer
->usage
& VK_BUFFER_USAGE_STORAGE_BUFFER_BIT
))
3567 pMemoryRequirements
->size
= align_u64(buffer
->size
, 4);
3569 pMemoryRequirements
->memoryTypeBits
= memory_types
;
3572 void anv_GetBufferMemoryRequirements2(
3574 const VkBufferMemoryRequirementsInfo2
* pInfo
,
3575 VkMemoryRequirements2
* pMemoryRequirements
)
3577 anv_GetBufferMemoryRequirements(_device
, pInfo
->buffer
,
3578 &pMemoryRequirements
->memoryRequirements
);
3580 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3581 switch (ext
->sType
) {
3582 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
3583 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
3584 requirements
->prefersDedicatedAllocation
= false;
3585 requirements
->requiresDedicatedAllocation
= false;
3590 anv_debug_ignored_stype(ext
->sType
);
3596 void anv_GetImageMemoryRequirements(
3599 VkMemoryRequirements
* pMemoryRequirements
)
3601 ANV_FROM_HANDLE(anv_image
, image
, _image
);
3602 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3603 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3605 /* The Vulkan spec (git aaed022) says:
3607 * memoryTypeBits is a bitfield and contains one bit set for every
3608 * supported memory type for the resource. The bit `1<<i` is set if and
3609 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3610 * structure for the physical device is supported.
3612 * All types are currently supported for images.
3614 uint32_t memory_types
= (1ull << pdevice
->memory
.type_count
) - 1;
3616 /* We must have image allocated or imported at this point. According to the
3617 * specification, external images must have been bound to memory before
3618 * calling GetImageMemoryRequirements.
3620 assert(image
->size
> 0);
3622 pMemoryRequirements
->size
= image
->size
;
3623 pMemoryRequirements
->alignment
= image
->alignment
;
3624 pMemoryRequirements
->memoryTypeBits
= memory_types
;
3627 void anv_GetImageMemoryRequirements2(
3629 const VkImageMemoryRequirementsInfo2
* pInfo
,
3630 VkMemoryRequirements2
* pMemoryRequirements
)
3632 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3633 ANV_FROM_HANDLE(anv_image
, image
, pInfo
->image
);
3635 anv_GetImageMemoryRequirements(_device
, pInfo
->image
,
3636 &pMemoryRequirements
->memoryRequirements
);
3638 vk_foreach_struct_const(ext
, pInfo
->pNext
) {
3639 switch (ext
->sType
) {
3640 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO
: {
3641 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3642 const VkImagePlaneMemoryRequirementsInfo
*plane_reqs
=
3643 (const VkImagePlaneMemoryRequirementsInfo
*) ext
;
3644 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
,
3645 plane_reqs
->planeAspect
);
3647 assert(image
->planes
[plane
].offset
== 0);
3649 /* The Vulkan spec (git aaed022) says:
3651 * memoryTypeBits is a bitfield and contains one bit set for every
3652 * supported memory type for the resource. The bit `1<<i` is set
3653 * if and only if the memory type `i` in the
3654 * VkPhysicalDeviceMemoryProperties structure for the physical
3655 * device is supported.
3657 * All types are currently supported for images.
3659 pMemoryRequirements
->memoryRequirements
.memoryTypeBits
=
3660 (1ull << pdevice
->memory
.type_count
) - 1;
3662 /* We must have image allocated or imported at this point. According to the
3663 * specification, external images must have been bound to memory before
3664 * calling GetImageMemoryRequirements.
3666 assert(image
->planes
[plane
].size
> 0);
3668 pMemoryRequirements
->memoryRequirements
.size
= image
->planes
[plane
].size
;
3669 pMemoryRequirements
->memoryRequirements
.alignment
=
3670 image
->planes
[plane
].alignment
;
3675 anv_debug_ignored_stype(ext
->sType
);
3680 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3681 switch (ext
->sType
) {
3682 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
3683 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
3684 if (image
->needs_set_tiling
|| image
->external_format
) {
3685 /* If we need to set the tiling for external consumers, we need a
3686 * dedicated allocation.
3688 * See also anv_AllocateMemory.
3690 requirements
->prefersDedicatedAllocation
= true;
3691 requirements
->requiresDedicatedAllocation
= true;
3693 requirements
->prefersDedicatedAllocation
= false;
3694 requirements
->requiresDedicatedAllocation
= false;
3700 anv_debug_ignored_stype(ext
->sType
);
3706 void anv_GetImageSparseMemoryRequirements(
3709 uint32_t* pSparseMemoryRequirementCount
,
3710 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
3712 *pSparseMemoryRequirementCount
= 0;
3715 void anv_GetImageSparseMemoryRequirements2(
3717 const VkImageSparseMemoryRequirementsInfo2
* pInfo
,
3718 uint32_t* pSparseMemoryRequirementCount
,
3719 VkSparseImageMemoryRequirements2
* pSparseMemoryRequirements
)
3721 *pSparseMemoryRequirementCount
= 0;
3724 void anv_GetDeviceMemoryCommitment(
3726 VkDeviceMemory memory
,
3727 VkDeviceSize
* pCommittedMemoryInBytes
)
3729 *pCommittedMemoryInBytes
= 0;
3733 anv_bind_buffer_memory(const VkBindBufferMemoryInfo
*pBindInfo
)
3735 ANV_FROM_HANDLE(anv_device_memory
, mem
, pBindInfo
->memory
);
3736 ANV_FROM_HANDLE(anv_buffer
, buffer
, pBindInfo
->buffer
);
3738 assert(pBindInfo
->sType
== VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
);
3741 assert((buffer
->usage
& mem
->type
->valid_buffer_usage
) == buffer
->usage
);
3742 buffer
->address
= (struct anv_address
) {
3744 .offset
= pBindInfo
->memoryOffset
,
3747 buffer
->address
= ANV_NULL_ADDRESS
;
3751 VkResult
anv_BindBufferMemory(
3754 VkDeviceMemory memory
,
3755 VkDeviceSize memoryOffset
)
3757 anv_bind_buffer_memory(
3758 &(VkBindBufferMemoryInfo
) {
3759 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
,
3762 .memoryOffset
= memoryOffset
,
3768 VkResult
anv_BindBufferMemory2(
3770 uint32_t bindInfoCount
,
3771 const VkBindBufferMemoryInfo
* pBindInfos
)
3773 for (uint32_t i
= 0; i
< bindInfoCount
; i
++)
3774 anv_bind_buffer_memory(&pBindInfos
[i
]);
3779 VkResult
anv_QueueBindSparse(
3781 uint32_t bindInfoCount
,
3782 const VkBindSparseInfo
* pBindInfo
,
3785 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
3786 if (anv_device_is_lost(queue
->device
))
3787 return VK_ERROR_DEVICE_LOST
;
3789 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
3794 VkResult
anv_CreateEvent(
3796 const VkEventCreateInfo
* pCreateInfo
,
3797 const VkAllocationCallbacks
* pAllocator
,
3800 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3801 struct anv_state state
;
3802 struct anv_event
*event
;
3804 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_EVENT_CREATE_INFO
);
3806 state
= anv_state_pool_alloc(&device
->dynamic_state_pool
,
3809 event
->state
= state
;
3810 event
->semaphore
= VK_EVENT_RESET
;
3812 if (!device
->info
.has_llc
) {
3813 /* Make sure the writes we're flushing have landed. */
3814 __builtin_ia32_mfence();
3815 __builtin_ia32_clflush(event
);
3818 *pEvent
= anv_event_to_handle(event
);
3823 void anv_DestroyEvent(
3826 const VkAllocationCallbacks
* pAllocator
)
3828 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3829 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3834 anv_state_pool_free(&device
->dynamic_state_pool
, event
->state
);
3837 VkResult
anv_GetEventStatus(
3841 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3842 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3844 if (anv_device_is_lost(device
))
3845 return VK_ERROR_DEVICE_LOST
;
3847 if (!device
->info
.has_llc
) {
3848 /* Invalidate read cache before reading event written by GPU. */
3849 __builtin_ia32_clflush(event
);
3850 __builtin_ia32_mfence();
3854 return event
->semaphore
;
3857 VkResult
anv_SetEvent(
3861 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3862 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3864 event
->semaphore
= VK_EVENT_SET
;
3866 if (!device
->info
.has_llc
) {
3867 /* Make sure the writes we're flushing have landed. */
3868 __builtin_ia32_mfence();
3869 __builtin_ia32_clflush(event
);
3875 VkResult
anv_ResetEvent(
3879 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3880 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3882 event
->semaphore
= VK_EVENT_RESET
;
3884 if (!device
->info
.has_llc
) {
3885 /* Make sure the writes we're flushing have landed. */
3886 __builtin_ia32_mfence();
3887 __builtin_ia32_clflush(event
);
3895 VkResult
anv_CreateBuffer(
3897 const VkBufferCreateInfo
* pCreateInfo
,
3898 const VkAllocationCallbacks
* pAllocator
,
3901 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3902 struct anv_buffer
*buffer
;
3904 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
3906 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
3907 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3909 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3911 buffer
->size
= pCreateInfo
->size
;
3912 buffer
->usage
= pCreateInfo
->usage
;
3913 buffer
->address
= ANV_NULL_ADDRESS
;
3915 *pBuffer
= anv_buffer_to_handle(buffer
);
3920 void anv_DestroyBuffer(
3923 const VkAllocationCallbacks
* pAllocator
)
3925 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3926 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
3931 vk_free2(&device
->alloc
, pAllocator
, buffer
);
3934 VkDeviceAddress
anv_GetBufferDeviceAddressEXT(
3936 const VkBufferDeviceAddressInfoEXT
* pInfo
)
3938 ANV_FROM_HANDLE(anv_buffer
, buffer
, pInfo
->buffer
);
3940 assert(buffer
->address
.bo
->flags
& EXEC_OBJECT_PINNED
);
3942 return anv_address_physical(buffer
->address
);
3946 anv_fill_buffer_surface_state(struct anv_device
*device
, struct anv_state state
,
3947 enum isl_format format
,
3948 struct anv_address address
,
3949 uint32_t range
, uint32_t stride
)
3951 isl_buffer_fill_state(&device
->isl_dev
, state
.map
,
3952 .address
= anv_address_physical(address
),
3953 .mocs
= device
->isl_dev
.mocs
.internal
,
3956 .swizzle
= ISL_SWIZZLE_IDENTITY
,
3957 .stride_B
= stride
);
3960 void anv_DestroySampler(
3963 const VkAllocationCallbacks
* pAllocator
)
3965 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3966 ANV_FROM_HANDLE(anv_sampler
, sampler
, _sampler
);
3971 if (sampler
->bindless_state
.map
) {
3972 anv_state_pool_free(&device
->dynamic_state_pool
,
3973 sampler
->bindless_state
);
3976 vk_free2(&device
->alloc
, pAllocator
, sampler
);
3979 VkResult
anv_CreateFramebuffer(
3981 const VkFramebufferCreateInfo
* pCreateInfo
,
3982 const VkAllocationCallbacks
* pAllocator
,
3983 VkFramebuffer
* pFramebuffer
)
3985 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3986 struct anv_framebuffer
*framebuffer
;
3988 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
3990 size_t size
= sizeof(*framebuffer
);
3992 /* VK_KHR_imageless_framebuffer extension says:
3994 * If flags includes VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR,
3995 * parameter pAttachments is ignored.
3997 if (!(pCreateInfo
->flags
& VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR
)) {
3998 size
+= sizeof(struct anv_image_view
*) * pCreateInfo
->attachmentCount
;
3999 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
4000 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
4001 if (framebuffer
== NULL
)
4002 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
4004 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
4005 ANV_FROM_HANDLE(anv_image_view
, iview
, pCreateInfo
->pAttachments
[i
]);
4006 framebuffer
->attachments
[i
] = iview
;
4008 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
4010 assert(device
->enabled_extensions
.KHR_imageless_framebuffer
);
4011 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
4012 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
4013 if (framebuffer
== NULL
)
4014 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
4016 framebuffer
->attachment_count
= 0;
4019 framebuffer
->width
= pCreateInfo
->width
;
4020 framebuffer
->height
= pCreateInfo
->height
;
4021 framebuffer
->layers
= pCreateInfo
->layers
;
4023 *pFramebuffer
= anv_framebuffer_to_handle(framebuffer
);
4028 void anv_DestroyFramebuffer(
4031 const VkAllocationCallbacks
* pAllocator
)
4033 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4034 ANV_FROM_HANDLE(anv_framebuffer
, fb
, _fb
);
4039 vk_free2(&device
->alloc
, pAllocator
, fb
);
4042 static const VkTimeDomainEXT anv_time_domains
[] = {
4043 VK_TIME_DOMAIN_DEVICE_EXT
,
4044 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
,
4045 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
,
4048 VkResult
anv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
4049 VkPhysicalDevice physicalDevice
,
4050 uint32_t *pTimeDomainCount
,
4051 VkTimeDomainEXT
*pTimeDomains
)
4054 VK_OUTARRAY_MAKE(out
, pTimeDomains
, pTimeDomainCount
);
4056 for (d
= 0; d
< ARRAY_SIZE(anv_time_domains
); d
++) {
4057 vk_outarray_append(&out
, i
) {
4058 *i
= anv_time_domains
[d
];
4062 return vk_outarray_status(&out
);
4066 anv_clock_gettime(clockid_t clock_id
)
4068 struct timespec current
;
4071 ret
= clock_gettime(clock_id
, ¤t
);
4072 if (ret
< 0 && clock_id
== CLOCK_MONOTONIC_RAW
)
4073 ret
= clock_gettime(CLOCK_MONOTONIC
, ¤t
);
4077 return (uint64_t) current
.tv_sec
* 1000000000ULL + current
.tv_nsec
;
4080 #define TIMESTAMP 0x2358
4082 VkResult
anv_GetCalibratedTimestampsEXT(
4084 uint32_t timestampCount
,
4085 const VkCalibratedTimestampInfoEXT
*pTimestampInfos
,
4086 uint64_t *pTimestamps
,
4087 uint64_t *pMaxDeviation
)
4089 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4090 uint64_t timestamp_frequency
= device
->info
.timestamp_frequency
;
4093 uint64_t begin
, end
;
4094 uint64_t max_clock_period
= 0;
4096 begin
= anv_clock_gettime(CLOCK_MONOTONIC_RAW
);
4098 for (d
= 0; d
< timestampCount
; d
++) {
4099 switch (pTimestampInfos
[d
].timeDomain
) {
4100 case VK_TIME_DOMAIN_DEVICE_EXT
:
4101 ret
= anv_gem_reg_read(device
, TIMESTAMP
| 1,
4105 return anv_device_set_lost(device
, "Failed to read the TIMESTAMP "
4108 uint64_t device_period
= DIV_ROUND_UP(1000000000, timestamp_frequency
);
4109 max_clock_period
= MAX2(max_clock_period
, device_period
);
4111 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
:
4112 pTimestamps
[d
] = anv_clock_gettime(CLOCK_MONOTONIC
);
4113 max_clock_period
= MAX2(max_clock_period
, 1);
4116 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
:
4117 pTimestamps
[d
] = begin
;
4125 end
= anv_clock_gettime(CLOCK_MONOTONIC_RAW
);
4128 * The maximum deviation is the sum of the interval over which we
4129 * perform the sampling and the maximum period of any sampled
4130 * clock. That's because the maximum skew between any two sampled
4131 * clock edges is when the sampled clock with the largest period is
4132 * sampled at the end of that period but right at the beginning of the
4133 * sampling interval and some other clock is sampled right at the
4134 * begining of its sampling period and right at the end of the
4135 * sampling interval. Let's assume the GPU has the longest clock
4136 * period and that the application is sampling GPU and monotonic:
4139 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
4140 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4144 * GPU -----_____-----_____-----_____-----_____
4147 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
4148 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4150 * Interval <----------------->
4151 * Deviation <-------------------------->
4155 * m = read(monotonic) 2
4158 * We round the sample interval up by one tick to cover sampling error
4159 * in the interval clock
4162 uint64_t sample_interval
= end
- begin
+ 1;
4164 *pMaxDeviation
= sample_interval
+ max_clock_period
;
4169 /* vk_icd.h does not declare this function, so we declare it here to
4170 * suppress Wmissing-prototypes.
4172 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4173 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
);
4175 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4176 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
)
4178 /* For the full details on loader interface versioning, see
4179 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
4180 * What follows is a condensed summary, to help you navigate the large and
4181 * confusing official doc.
4183 * - Loader interface v0 is incompatible with later versions. We don't
4186 * - In loader interface v1:
4187 * - The first ICD entrypoint called by the loader is
4188 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
4190 * - The ICD must statically expose no other Vulkan symbol unless it is
4191 * linked with -Bsymbolic.
4192 * - Each dispatchable Vulkan handle created by the ICD must be
4193 * a pointer to a struct whose first member is VK_LOADER_DATA. The
4194 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
4195 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
4196 * vkDestroySurfaceKHR(). The ICD must be capable of working with
4197 * such loader-managed surfaces.
4199 * - Loader interface v2 differs from v1 in:
4200 * - The first ICD entrypoint called by the loader is
4201 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
4202 * statically expose this entrypoint.
4204 * - Loader interface v3 differs from v2 in:
4205 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
4206 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
4207 * because the loader no longer does so.
4209 * - Loader interface v4 differs from v3 in:
4210 * - The ICD must implement vk_icdGetPhysicalDeviceProcAddr().
4212 *pSupportedVersion
= MIN2(*pSupportedVersion
, 4u);