2 * Copyright © 2015 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
28 #include <sys/sysinfo.h>
32 #include "drm-uapi/drm_fourcc.h"
34 #include "anv_private.h"
35 #include "util/strtod.h"
36 #include "util/debug.h"
37 #include "util/build_id.h"
38 #include "util/disk_cache.h"
39 #include "util/mesa-sha1.h"
40 #include "util/os_file.h"
41 #include "util/u_atomic.h"
42 #include "util/u_string.h"
45 #include "common/gen_defines.h"
46 #include "compiler/glsl_types.h"
48 #include "genxml/gen7_pack.h"
50 /* This is probably far to big but it reflects the max size used for messages
51 * in OpenGLs KHR_debug.
53 #define MAX_DEBUG_MESSAGE_LENGTH 4096
56 compiler_debug_log(void *data
, const char *fmt
, ...)
58 char str
[MAX_DEBUG_MESSAGE_LENGTH
];
59 struct anv_device
*device
= (struct anv_device
*)data
;
61 if (list_empty(&device
->instance
->debug_report_callbacks
.callbacks
))
66 (void) vsnprintf(str
, MAX_DEBUG_MESSAGE_LENGTH
, fmt
, args
);
69 vk_debug_report(&device
->instance
->debug_report_callbacks
,
70 VK_DEBUG_REPORT_DEBUG_BIT_EXT
,
71 VK_DEBUG_REPORT_OBJECT_TYPE_UNKNOWN_EXT
,
76 compiler_perf_log(void *data
, const char *fmt
, ...)
81 if (unlikely(INTEL_DEBUG
& DEBUG_PERF
))
82 intel_logd_v(fmt
, args
);
88 anv_compute_heap_size(int fd
, uint64_t gtt_size
)
90 /* Query the total ram from the system */
94 uint64_t total_ram
= (uint64_t)info
.totalram
* (uint64_t)info
.mem_unit
;
96 /* We don't want to burn too much ram with the GPU. If the user has 4GiB
97 * or less, we use at most half. If they have more than 4GiB, we use 3/4.
99 uint64_t available_ram
;
100 if (total_ram
<= 4ull * 1024ull * 1024ull * 1024ull)
101 available_ram
= total_ram
/ 2;
103 available_ram
= total_ram
* 3 / 4;
105 /* We also want to leave some padding for things we allocate in the driver,
106 * so don't go over 3/4 of the GTT either.
108 uint64_t available_gtt
= gtt_size
* 3 / 4;
110 return MIN2(available_ram
, available_gtt
);
114 anv_physical_device_init_heaps(struct anv_physical_device
*device
, int fd
)
117 if (anv_gem_get_context_param(fd
, 0, I915_CONTEXT_PARAM_GTT_SIZE
,
119 /* If, for whatever reason, we can't actually get the GTT size from the
120 * kernel (too old?) fall back to the aperture size.
122 anv_perf_warn(NULL
, NULL
,
123 "Failed to get I915_CONTEXT_PARAM_GTT_SIZE: %m");
125 if (anv_gem_get_aperture(fd
, >t_size
) == -1) {
126 return vk_errorf(NULL
, NULL
, VK_ERROR_INITIALIZATION_FAILED
,
127 "failed to get aperture size: %m");
131 device
->supports_48bit_addresses
= (device
->info
.gen
>= 8) &&
132 gtt_size
> (4ULL << 30 /* GiB */);
134 uint64_t heap_size
= anv_compute_heap_size(fd
, gtt_size
);
136 if (heap_size
> (2ull << 30) && !device
->supports_48bit_addresses
) {
137 /* When running with an overridden PCI ID, we may get a GTT size from
138 * the kernel that is greater than 2 GiB but the execbuf check for 48bit
139 * address support can still fail. Just clamp the address space size to
140 * 2 GiB if we don't have 48-bit support.
142 intel_logw("%s:%d: The kernel reported a GTT size larger than 2 GiB but "
143 "not support for 48-bit addresses",
145 heap_size
= 2ull << 30;
148 if (heap_size
<= 3ull * (1ull << 30)) {
149 /* In this case, everything fits nicely into the 32-bit address space,
150 * so there's no need for supporting 48bit addresses on client-allocated
153 device
->memory
.heap_count
= 1;
154 device
->memory
.heaps
[0] = (struct anv_memory_heap
) {
155 .vma_start
= LOW_HEAP_MIN_ADDRESS
,
156 .vma_size
= LOW_HEAP_SIZE
,
158 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
159 .supports_48bit_addresses
= false,
162 /* Not everything will fit nicely into a 32-bit address space. In this
163 * case we need a 64-bit heap. Advertise a small 32-bit heap and a
164 * larger 48-bit heap. If we're in this case, then we have a total heap
165 * size larger than 3GiB which most likely means they have 8 GiB of
166 * video memory and so carving off 1 GiB for the 32-bit heap should be
169 const uint64_t heap_size_32bit
= 1ull << 30;
170 const uint64_t heap_size_48bit
= heap_size
- heap_size_32bit
;
172 assert(device
->supports_48bit_addresses
);
174 device
->memory
.heap_count
= 2;
175 device
->memory
.heaps
[0] = (struct anv_memory_heap
) {
176 .vma_start
= HIGH_HEAP_MIN_ADDRESS
,
177 /* Leave the last 4GiB out of the high vma range, so that no state
178 * base address + size can overflow 48 bits. For more information see
179 * the comment about Wa32bitGeneralStateOffset in anv_allocator.c
181 .vma_size
= gtt_size
- (1ull << 32) - HIGH_HEAP_MIN_ADDRESS
,
182 .size
= heap_size_48bit
,
183 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
184 .supports_48bit_addresses
= true,
186 device
->memory
.heaps
[1] = (struct anv_memory_heap
) {
187 .vma_start
= LOW_HEAP_MIN_ADDRESS
,
188 .vma_size
= LOW_HEAP_SIZE
,
189 .size
= heap_size_32bit
,
190 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
191 .supports_48bit_addresses
= false,
195 uint32_t type_count
= 0;
196 for (uint32_t heap
= 0; heap
< device
->memory
.heap_count
; heap
++) {
197 uint32_t valid_buffer_usage
= ~0;
199 /* There appears to be a hardware issue in the VF cache where it only
200 * considers the bottom 32 bits of memory addresses. If you happen to
201 * have two vertex buffers which get placed exactly 4 GiB apart and use
202 * them in back-to-back draw calls, you can get collisions. In order to
203 * solve this problem, we require vertex and index buffers be bound to
204 * memory allocated out of the 32-bit heap.
206 if (device
->memory
.heaps
[heap
].supports_48bit_addresses
) {
207 valid_buffer_usage
&= ~(VK_BUFFER_USAGE_INDEX_BUFFER_BIT
|
208 VK_BUFFER_USAGE_VERTEX_BUFFER_BIT
);
211 if (device
->info
.has_llc
) {
212 /* Big core GPUs share LLC with the CPU and thus one memory type can be
213 * both cached and coherent at the same time.
215 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
216 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
217 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
218 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
219 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
221 .valid_buffer_usage
= valid_buffer_usage
,
224 /* The spec requires that we expose a host-visible, coherent memory
225 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
226 * to give the application a choice between cached, but not coherent and
227 * coherent but uncached (WC though).
229 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
230 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
231 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
232 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
234 .valid_buffer_usage
= valid_buffer_usage
,
236 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
237 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
238 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
239 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
241 .valid_buffer_usage
= valid_buffer_usage
,
245 device
->memory
.type_count
= type_count
;
251 anv_physical_device_init_uuids(struct anv_physical_device
*device
)
253 const struct build_id_note
*note
=
254 build_id_find_nhdr_for_addr(anv_physical_device_init_uuids
);
256 return vk_errorf(device
->instance
, device
,
257 VK_ERROR_INITIALIZATION_FAILED
,
258 "Failed to find build-id");
261 unsigned build_id_len
= build_id_length(note
);
262 if (build_id_len
< 20) {
263 return vk_errorf(device
->instance
, device
,
264 VK_ERROR_INITIALIZATION_FAILED
,
265 "build-id too short. It needs to be a SHA");
268 memcpy(device
->driver_build_sha1
, build_id_data(note
), 20);
270 struct mesa_sha1 sha1_ctx
;
272 STATIC_ASSERT(VK_UUID_SIZE
<= sizeof(sha1
));
274 /* The pipeline cache UUID is used for determining when a pipeline cache is
275 * invalid. It needs both a driver build and the PCI ID of the device.
277 _mesa_sha1_init(&sha1_ctx
);
278 _mesa_sha1_update(&sha1_ctx
, build_id_data(note
), build_id_len
);
279 _mesa_sha1_update(&sha1_ctx
, &device
->chipset_id
,
280 sizeof(device
->chipset_id
));
281 _mesa_sha1_update(&sha1_ctx
, &device
->always_use_bindless
,
282 sizeof(device
->always_use_bindless
));
283 _mesa_sha1_update(&sha1_ctx
, &device
->has_a64_buffer_access
,
284 sizeof(device
->has_a64_buffer_access
));
285 _mesa_sha1_update(&sha1_ctx
, &device
->has_bindless_images
,
286 sizeof(device
->has_bindless_images
));
287 _mesa_sha1_update(&sha1_ctx
, &device
->has_bindless_samplers
,
288 sizeof(device
->has_bindless_samplers
));
289 _mesa_sha1_final(&sha1_ctx
, sha1
);
290 memcpy(device
->pipeline_cache_uuid
, sha1
, VK_UUID_SIZE
);
292 /* The driver UUID is used for determining sharability of images and memory
293 * between two Vulkan instances in separate processes. People who want to
294 * share memory need to also check the device UUID (below) so all this
295 * needs to be is the build-id.
297 memcpy(device
->driver_uuid
, build_id_data(note
), VK_UUID_SIZE
);
299 /* The device UUID uniquely identifies the given device within the machine.
300 * Since we never have more than one device, this doesn't need to be a real
301 * UUID. However, on the off-chance that someone tries to use this to
302 * cache pre-tiled images or something of the like, we use the PCI ID and
303 * some bits of ISL info to ensure that this is safe.
305 _mesa_sha1_init(&sha1_ctx
);
306 _mesa_sha1_update(&sha1_ctx
, &device
->chipset_id
,
307 sizeof(device
->chipset_id
));
308 _mesa_sha1_update(&sha1_ctx
, &device
->isl_dev
.has_bit6_swizzling
,
309 sizeof(device
->isl_dev
.has_bit6_swizzling
));
310 _mesa_sha1_final(&sha1_ctx
, sha1
);
311 memcpy(device
->device_uuid
, sha1
, VK_UUID_SIZE
);
317 anv_physical_device_init_disk_cache(struct anv_physical_device
*device
)
319 #ifdef ENABLE_SHADER_CACHE
321 MAYBE_UNUSED
int len
= snprintf(renderer
, sizeof(renderer
), "anv_%04x",
323 assert(len
== sizeof(renderer
) - 2);
326 _mesa_sha1_format(timestamp
, device
->driver_build_sha1
);
328 const uint64_t driver_flags
=
329 brw_get_compiler_config_value(device
->compiler
);
330 device
->disk_cache
= disk_cache_create(renderer
, timestamp
, driver_flags
);
332 device
->disk_cache
= NULL
;
337 anv_physical_device_free_disk_cache(struct anv_physical_device
*device
)
339 #ifdef ENABLE_SHADER_CACHE
340 if (device
->disk_cache
)
341 disk_cache_destroy(device
->disk_cache
);
343 assert(device
->disk_cache
== NULL
);
348 get_available_system_memory()
350 char *meminfo
= os_read_file("/proc/meminfo");
354 char *str
= strstr(meminfo
, "MemAvailable:");
360 uint64_t kb_mem_available
;
361 if (sscanf(str
, "MemAvailable: %" PRIx64
, &kb_mem_available
) == 1) {
363 return kb_mem_available
<< 10;
371 anv_physical_device_init(struct anv_physical_device
*device
,
372 struct anv_instance
*instance
,
373 drmDevicePtr drm_device
)
375 const char *primary_path
= drm_device
->nodes
[DRM_NODE_PRIMARY
];
376 const char *path
= drm_device
->nodes
[DRM_NODE_RENDER
];
381 brw_process_intel_debug_variable();
383 fd
= open(path
, O_RDWR
| O_CLOEXEC
);
385 return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
387 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
388 device
->instance
= instance
;
390 assert(strlen(path
) < ARRAY_SIZE(device
->path
));
391 snprintf(device
->path
, ARRAY_SIZE(device
->path
), "%s", path
);
393 device
->no_hw
= getenv("INTEL_NO_HW") != NULL
;
395 const int pci_id_override
= gen_get_pci_device_id_override();
396 if (pci_id_override
< 0) {
397 device
->chipset_id
= anv_gem_get_param(fd
, I915_PARAM_CHIPSET_ID
);
398 if (!device
->chipset_id
) {
399 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
403 device
->chipset_id
= pci_id_override
;
404 device
->no_hw
= true;
407 device
->pci_info
.domain
= drm_device
->businfo
.pci
->domain
;
408 device
->pci_info
.bus
= drm_device
->businfo
.pci
->bus
;
409 device
->pci_info
.device
= drm_device
->businfo
.pci
->dev
;
410 device
->pci_info
.function
= drm_device
->businfo
.pci
->func
;
412 device
->name
= gen_get_device_name(device
->chipset_id
);
413 if (!gen_get_device_info(device
->chipset_id
, &device
->info
)) {
414 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
418 if (device
->info
.is_haswell
) {
419 intel_logw("Haswell Vulkan support is incomplete");
420 } else if (device
->info
.gen
== 7 && !device
->info
.is_baytrail
) {
421 intel_logw("Ivy Bridge Vulkan support is incomplete");
422 } else if (device
->info
.gen
== 7 && device
->info
.is_baytrail
) {
423 intel_logw("Bay Trail Vulkan support is incomplete");
424 } else if (device
->info
.gen
>= 8 && device
->info
.gen
<= 11) {
425 /* Gen8-11 fully supported */
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 result
= anv_physical_device_init_heaps(device
, fd
);
468 if (result
!= VK_SUCCESS
)
471 device
->has_exec_async
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_ASYNC
);
472 device
->has_exec_capture
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_CAPTURE
);
473 device
->has_exec_fence
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE
);
474 device
->has_syncobj
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE_ARRAY
);
475 device
->has_syncobj_wait
= device
->has_syncobj
&&
476 anv_gem_supports_syncobj_wait(fd
);
477 device
->has_context_priority
= anv_gem_has_context_priority(fd
);
479 device
->use_softpin
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_SOFTPIN
)
480 && device
->supports_48bit_addresses
;
482 device
->has_context_isolation
=
483 anv_gem_get_param(fd
, I915_PARAM_HAS_CONTEXT_ISOLATION
);
485 device
->always_use_bindless
=
486 env_var_as_boolean("ANV_ALWAYS_BINDLESS", false);
488 /* We first got the A64 messages on broadwell and we can only use them if
489 * we can pass addresses directly into the shader which requires softpin.
491 device
->has_a64_buffer_access
= device
->info
.gen
>= 8 &&
494 /* We first get bindless image access on Skylake and we can only really do
495 * it if we don't have any relocations so we need softpin.
497 device
->has_bindless_images
= device
->info
.gen
>= 9 &&
500 /* We've had bindless samplers since Ivy Bridge (forever in Vulkan terms)
501 * because it's just a matter of setting the sampler address in the sample
502 * message header. However, we've not bothered to wire it up for vec4 so
503 * we leave it disabled on gen7.
505 device
->has_bindless_samplers
= device
->info
.gen
>= 8;
507 device
->has_mem_available
= get_available_system_memory() != 0;
509 /* Starting with Gen10, the timestamp frequency of the command streamer may
510 * vary from one part to another. We can query the value from the kernel.
512 if (device
->info
.gen
>= 10) {
513 int timestamp_frequency
=
514 anv_gem_get_param(fd
, I915_PARAM_CS_TIMESTAMP_FREQUENCY
);
516 if (timestamp_frequency
< 0)
517 intel_logw("Kernel 4.16-rc1+ required to properly query CS timestamp frequency");
519 device
->info
.timestamp_frequency
= timestamp_frequency
;
522 /* GENs prior to 8 do not support EU/Subslice info */
523 if (device
->info
.gen
>= 8) {
524 device
->subslice_total
= anv_gem_get_param(fd
, I915_PARAM_SUBSLICE_TOTAL
);
525 device
->eu_total
= anv_gem_get_param(fd
, I915_PARAM_EU_TOTAL
);
527 /* Without this information, we cannot get the right Braswell
528 * brandstrings, and we have to use conservative numbers for GPGPU on
529 * many platforms, but otherwise, things will just work.
531 if (device
->subslice_total
< 1 || device
->eu_total
< 1) {
532 intel_logw("Kernel 4.1 required to properly query GPU properties");
534 } else if (device
->info
.gen
== 7) {
535 device
->subslice_total
= 1 << (device
->info
.gt
- 1);
538 if (device
->info
.is_cherryview
&&
539 device
->subslice_total
> 0 && device
->eu_total
> 0) {
540 /* Logical CS threads = EUs per subslice * num threads per EU */
541 uint32_t max_cs_threads
=
542 device
->eu_total
/ device
->subslice_total
* device
->info
.num_thread_per_eu
;
544 /* Fuse configurations may give more threads than expected, never less. */
545 if (max_cs_threads
> device
->info
.max_cs_threads
)
546 device
->info
.max_cs_threads
= max_cs_threads
;
549 device
->compiler
= brw_compiler_create(NULL
, &device
->info
);
550 if (device
->compiler
== NULL
) {
551 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
554 device
->compiler
->shader_debug_log
= compiler_debug_log
;
555 device
->compiler
->shader_perf_log
= compiler_perf_log
;
556 device
->compiler
->supports_pull_constants
= false;
557 device
->compiler
->constant_buffer_0_is_relative
=
558 device
->info
.gen
< 8 || !device
->has_context_isolation
;
559 device
->compiler
->supports_shader_constants
= true;
561 /* Broadwell PRM says:
563 * "Before Gen8, there was a historical configuration control field to
564 * swizzle address bit[6] for in X/Y tiling modes. This was set in three
565 * different places: TILECTL[1:0], ARB_MODE[5:4], and
566 * DISP_ARB_CTL[14:13].
568 * For Gen8 and subsequent generations, the swizzle fields are all
569 * reserved, and the CPU's memory controller performs all address
570 * swizzling modifications."
573 device
->info
.gen
< 8 && anv_gem_get_bit6_swizzle(fd
, I915_TILING_X
);
575 isl_device_init(&device
->isl_dev
, &device
->info
, swizzled
);
577 result
= anv_physical_device_init_uuids(device
);
578 if (result
!= VK_SUCCESS
)
581 anv_physical_device_init_disk_cache(device
);
583 if (instance
->enabled_extensions
.KHR_display
) {
584 master_fd
= open(primary_path
, O_RDWR
| O_CLOEXEC
);
585 if (master_fd
>= 0) {
586 /* prod the device with a GETPARAM call which will fail if
587 * we don't have permission to even render on this device
589 if (anv_gem_get_param(master_fd
, I915_PARAM_CHIPSET_ID
) == 0) {
595 device
->master_fd
= master_fd
;
597 result
= anv_init_wsi(device
);
598 if (result
!= VK_SUCCESS
) {
599 ralloc_free(device
->compiler
);
600 anv_physical_device_free_disk_cache(device
);
604 anv_physical_device_get_supported_extensions(device
,
605 &device
->supported_extensions
);
608 device
->local_fd
= fd
;
620 anv_physical_device_finish(struct anv_physical_device
*device
)
622 anv_finish_wsi(device
);
623 anv_physical_device_free_disk_cache(device
);
624 ralloc_free(device
->compiler
);
625 close(device
->local_fd
);
626 if (device
->master_fd
>= 0)
627 close(device
->master_fd
);
631 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
632 VkSystemAllocationScope allocationScope
)
638 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
639 size_t align
, VkSystemAllocationScope allocationScope
)
641 return realloc(pOriginal
, size
);
645 default_free_func(void *pUserData
, void *pMemory
)
650 static const VkAllocationCallbacks default_alloc
= {
652 .pfnAllocation
= default_alloc_func
,
653 .pfnReallocation
= default_realloc_func
,
654 .pfnFree
= default_free_func
,
657 VkResult
anv_EnumerateInstanceExtensionProperties(
658 const char* pLayerName
,
659 uint32_t* pPropertyCount
,
660 VkExtensionProperties
* pProperties
)
662 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
664 for (int i
= 0; i
< ANV_INSTANCE_EXTENSION_COUNT
; i
++) {
665 if (anv_instance_extensions_supported
.extensions
[i
]) {
666 vk_outarray_append(&out
, prop
) {
667 *prop
= anv_instance_extensions
[i
];
672 return vk_outarray_status(&out
);
675 VkResult
anv_CreateInstance(
676 const VkInstanceCreateInfo
* pCreateInfo
,
677 const VkAllocationCallbacks
* pAllocator
,
678 VkInstance
* pInstance
)
680 struct anv_instance
*instance
;
683 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
685 struct anv_instance_extension_table enabled_extensions
= {};
686 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
688 for (idx
= 0; idx
< ANV_INSTANCE_EXTENSION_COUNT
; idx
++) {
689 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
690 anv_instance_extensions
[idx
].extensionName
) == 0)
694 if (idx
>= ANV_INSTANCE_EXTENSION_COUNT
)
695 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
697 if (!anv_instance_extensions_supported
.extensions
[idx
])
698 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
700 enabled_extensions
.extensions
[idx
] = true;
703 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
704 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
706 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
708 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
711 instance
->alloc
= *pAllocator
;
713 instance
->alloc
= default_alloc
;
715 instance
->app_info
= (struct anv_app_info
) { .api_version
= 0 };
716 if (pCreateInfo
->pApplicationInfo
) {
717 const VkApplicationInfo
*app
= pCreateInfo
->pApplicationInfo
;
719 instance
->app_info
.app_name
=
720 vk_strdup(&instance
->alloc
, app
->pApplicationName
,
721 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
722 instance
->app_info
.app_version
= app
->applicationVersion
;
724 instance
->app_info
.engine_name
=
725 vk_strdup(&instance
->alloc
, app
->pEngineName
,
726 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
727 instance
->app_info
.engine_version
= app
->engineVersion
;
729 instance
->app_info
.api_version
= app
->apiVersion
;
732 if (instance
->app_info
.api_version
== 0)
733 instance
->app_info
.api_version
= VK_API_VERSION_1_0
;
735 instance
->enabled_extensions
= enabled_extensions
;
737 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->dispatch
.entrypoints
); i
++) {
738 /* Vulkan requires that entrypoints for extensions which have not been
739 * enabled must not be advertised.
741 if (!anv_instance_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
742 &instance
->enabled_extensions
)) {
743 instance
->dispatch
.entrypoints
[i
] = NULL
;
745 instance
->dispatch
.entrypoints
[i
] =
746 anv_instance_dispatch_table
.entrypoints
[i
];
750 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->device_dispatch
.entrypoints
); i
++) {
751 /* Vulkan requires that entrypoints for extensions which have not been
752 * enabled must not be advertised.
754 if (!anv_device_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
755 &instance
->enabled_extensions
, NULL
)) {
756 instance
->device_dispatch
.entrypoints
[i
] = NULL
;
758 instance
->device_dispatch
.entrypoints
[i
] =
759 anv_device_dispatch_table
.entrypoints
[i
];
763 instance
->physicalDeviceCount
= -1;
765 result
= vk_debug_report_instance_init(&instance
->debug_report_callbacks
);
766 if (result
!= VK_SUCCESS
) {
767 vk_free2(&default_alloc
, pAllocator
, instance
);
768 return vk_error(result
);
771 instance
->pipeline_cache_enabled
=
772 env_var_as_boolean("ANV_ENABLE_PIPELINE_CACHE", true);
775 glsl_type_singleton_init_or_ref();
777 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
779 *pInstance
= anv_instance_to_handle(instance
);
784 void anv_DestroyInstance(
785 VkInstance _instance
,
786 const VkAllocationCallbacks
* pAllocator
)
788 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
793 if (instance
->physicalDeviceCount
> 0) {
794 /* We support at most one physical device. */
795 assert(instance
->physicalDeviceCount
== 1);
796 anv_physical_device_finish(&instance
->physicalDevice
);
799 vk_free(&instance
->alloc
, (char *)instance
->app_info
.app_name
);
800 vk_free(&instance
->alloc
, (char *)instance
->app_info
.engine_name
);
802 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
804 vk_debug_report_instance_destroy(&instance
->debug_report_callbacks
);
806 glsl_type_singleton_decref();
809 vk_free(&instance
->alloc
, instance
);
813 anv_enumerate_devices(struct anv_instance
*instance
)
815 /* TODO: Check for more devices ? */
816 drmDevicePtr devices
[8];
817 VkResult result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
820 instance
->physicalDeviceCount
= 0;
822 max_devices
= drmGetDevices2(0, devices
, ARRAY_SIZE(devices
));
824 return VK_ERROR_INCOMPATIBLE_DRIVER
;
826 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
827 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
828 devices
[i
]->bustype
== DRM_BUS_PCI
&&
829 devices
[i
]->deviceinfo
.pci
->vendor_id
== 0x8086) {
831 result
= anv_physical_device_init(&instance
->physicalDevice
,
832 instance
, devices
[i
]);
833 if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
837 drmFreeDevices(devices
, max_devices
);
839 if (result
== VK_SUCCESS
)
840 instance
->physicalDeviceCount
= 1;
846 anv_instance_ensure_physical_device(struct anv_instance
*instance
)
848 if (instance
->physicalDeviceCount
< 0) {
849 VkResult result
= anv_enumerate_devices(instance
);
850 if (result
!= VK_SUCCESS
&&
851 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
858 VkResult
anv_EnumeratePhysicalDevices(
859 VkInstance _instance
,
860 uint32_t* pPhysicalDeviceCount
,
861 VkPhysicalDevice
* pPhysicalDevices
)
863 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
864 VK_OUTARRAY_MAKE(out
, pPhysicalDevices
, pPhysicalDeviceCount
);
866 VkResult result
= anv_instance_ensure_physical_device(instance
);
867 if (result
!= VK_SUCCESS
)
870 if (instance
->physicalDeviceCount
== 0)
873 assert(instance
->physicalDeviceCount
== 1);
874 vk_outarray_append(&out
, i
) {
875 *i
= anv_physical_device_to_handle(&instance
->physicalDevice
);
878 return vk_outarray_status(&out
);
881 VkResult
anv_EnumeratePhysicalDeviceGroups(
882 VkInstance _instance
,
883 uint32_t* pPhysicalDeviceGroupCount
,
884 VkPhysicalDeviceGroupProperties
* pPhysicalDeviceGroupProperties
)
886 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
887 VK_OUTARRAY_MAKE(out
, pPhysicalDeviceGroupProperties
,
888 pPhysicalDeviceGroupCount
);
890 VkResult result
= anv_instance_ensure_physical_device(instance
);
891 if (result
!= VK_SUCCESS
)
894 if (instance
->physicalDeviceCount
== 0)
897 assert(instance
->physicalDeviceCount
== 1);
899 vk_outarray_append(&out
, p
) {
900 p
->physicalDeviceCount
= 1;
901 memset(p
->physicalDevices
, 0, sizeof(p
->physicalDevices
));
902 p
->physicalDevices
[0] =
903 anv_physical_device_to_handle(&instance
->physicalDevice
);
904 p
->subsetAllocation
= false;
906 vk_foreach_struct(ext
, p
->pNext
)
907 anv_debug_ignored_stype(ext
->sType
);
910 return vk_outarray_status(&out
);
913 void anv_GetPhysicalDeviceFeatures(
914 VkPhysicalDevice physicalDevice
,
915 VkPhysicalDeviceFeatures
* pFeatures
)
917 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
919 *pFeatures
= (VkPhysicalDeviceFeatures
) {
920 .robustBufferAccess
= true,
921 .fullDrawIndexUint32
= true,
922 .imageCubeArray
= true,
923 .independentBlend
= true,
924 .geometryShader
= true,
925 .tessellationShader
= true,
926 .sampleRateShading
= true,
927 .dualSrcBlend
= true,
929 .multiDrawIndirect
= true,
930 .drawIndirectFirstInstance
= true,
932 .depthBiasClamp
= true,
933 .fillModeNonSolid
= true,
934 .depthBounds
= false,
938 .multiViewport
= true,
939 .samplerAnisotropy
= true,
940 .textureCompressionETC2
= pdevice
->info
.gen
>= 8 ||
941 pdevice
->info
.is_baytrail
,
942 .textureCompressionASTC_LDR
= pdevice
->info
.gen
>= 9, /* FINISHME CHV */
943 .textureCompressionBC
= true,
944 .occlusionQueryPrecise
= true,
945 .pipelineStatisticsQuery
= true,
946 .fragmentStoresAndAtomics
= true,
947 .shaderTessellationAndGeometryPointSize
= true,
948 .shaderImageGatherExtended
= true,
949 .shaderStorageImageExtendedFormats
= true,
950 .shaderStorageImageMultisample
= false,
951 .shaderStorageImageReadWithoutFormat
= false,
952 .shaderStorageImageWriteWithoutFormat
= true,
953 .shaderUniformBufferArrayDynamicIndexing
= true,
954 .shaderSampledImageArrayDynamicIndexing
= true,
955 .shaderStorageBufferArrayDynamicIndexing
= true,
956 .shaderStorageImageArrayDynamicIndexing
= true,
957 .shaderClipDistance
= true,
958 .shaderCullDistance
= true,
959 .shaderFloat64
= pdevice
->info
.gen
>= 8 &&
960 pdevice
->info
.has_64bit_types
,
961 .shaderInt64
= pdevice
->info
.gen
>= 8 &&
962 pdevice
->info
.has_64bit_types
,
963 .shaderInt16
= pdevice
->info
.gen
>= 8,
964 .shaderResourceMinLod
= pdevice
->info
.gen
>= 9,
965 .variableMultisampleRate
= true,
966 .inheritedQueries
= true,
969 /* We can't do image stores in vec4 shaders */
970 pFeatures
->vertexPipelineStoresAndAtomics
=
971 pdevice
->compiler
->scalar_stage
[MESA_SHADER_VERTEX
] &&
972 pdevice
->compiler
->scalar_stage
[MESA_SHADER_GEOMETRY
];
974 struct anv_app_info
*app_info
= &pdevice
->instance
->app_info
;
976 /* The new DOOM and Wolfenstein games require depthBounds without
977 * checking for it. They seem to run fine without it so just claim it's
978 * there and accept the consequences.
980 if (app_info
->engine_name
&& strcmp(app_info
->engine_name
, "idTech") == 0)
981 pFeatures
->depthBounds
= true;
984 void anv_GetPhysicalDeviceFeatures2(
985 VkPhysicalDevice physicalDevice
,
986 VkPhysicalDeviceFeatures2
* pFeatures
)
988 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
989 anv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
991 vk_foreach_struct(ext
, pFeatures
->pNext
) {
992 switch (ext
->sType
) {
993 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES_KHR
: {
994 VkPhysicalDevice8BitStorageFeaturesKHR
*features
=
995 (VkPhysicalDevice8BitStorageFeaturesKHR
*)ext
;
996 features
->storageBuffer8BitAccess
= pdevice
->info
.gen
>= 8;
997 features
->uniformAndStorageBuffer8BitAccess
= pdevice
->info
.gen
>= 8;
998 features
->storagePushConstant8
= pdevice
->info
.gen
>= 8;
1002 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES
: {
1003 VkPhysicalDevice16BitStorageFeatures
*features
=
1004 (VkPhysicalDevice16BitStorageFeatures
*)ext
;
1005 features
->storageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
1006 features
->uniformAndStorageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
1007 features
->storagePushConstant16
= pdevice
->info
.gen
>= 8;
1008 features
->storageInputOutput16
= false;
1012 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_EXT
: {
1013 VkPhysicalDeviceBufferDeviceAddressFeaturesEXT
*features
= (void *)ext
;
1014 features
->bufferDeviceAddress
= pdevice
->has_a64_buffer_access
;
1015 features
->bufferDeviceAddressCaptureReplay
= false;
1016 features
->bufferDeviceAddressMultiDevice
= false;
1020 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COMPUTE_SHADER_DERIVATIVES_FEATURES_NV
: {
1021 VkPhysicalDeviceComputeShaderDerivativesFeaturesNV
*features
=
1022 (VkPhysicalDeviceComputeShaderDerivativesFeaturesNV
*)ext
;
1023 features
->computeDerivativeGroupQuads
= true;
1024 features
->computeDerivativeGroupLinear
= true;
1028 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT
: {
1029 VkPhysicalDeviceConditionalRenderingFeaturesEXT
*features
=
1030 (VkPhysicalDeviceConditionalRenderingFeaturesEXT
*)ext
;
1031 features
->conditionalRendering
= pdevice
->info
.gen
>= 8 ||
1032 pdevice
->info
.is_haswell
;
1033 features
->inheritedConditionalRendering
= pdevice
->info
.gen
>= 8 ||
1034 pdevice
->info
.is_haswell
;
1038 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_CLIP_ENABLE_FEATURES_EXT
: {
1039 VkPhysicalDeviceDepthClipEnableFeaturesEXT
*features
=
1040 (VkPhysicalDeviceDepthClipEnableFeaturesEXT
*)ext
;
1041 features
->depthClipEnable
= true;
1045 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR
: {
1046 VkPhysicalDeviceFloat16Int8FeaturesKHR
*features
= (void *)ext
;
1047 features
->shaderFloat16
= pdevice
->info
.gen
>= 8;
1048 features
->shaderInt8
= pdevice
->info
.gen
>= 8;
1052 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FRAGMENT_SHADER_INTERLOCK_FEATURES_EXT
: {
1053 VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT
*features
=
1054 (VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT
*)ext
;
1055 features
->fragmentShaderSampleInterlock
= pdevice
->info
.gen
>= 9;
1056 features
->fragmentShaderPixelInterlock
= pdevice
->info
.gen
>= 9;
1057 features
->fragmentShaderShadingRateInterlock
= false;
1061 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_QUERY_RESET_FEATURES_EXT
: {
1062 VkPhysicalDeviceHostQueryResetFeaturesEXT
*features
=
1063 (VkPhysicalDeviceHostQueryResetFeaturesEXT
*)ext
;
1064 features
->hostQueryReset
= true;
1068 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT
: {
1069 VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*features
=
1070 (VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*)ext
;
1071 features
->shaderInputAttachmentArrayDynamicIndexing
= false;
1072 features
->shaderUniformTexelBufferArrayDynamicIndexing
= true;
1073 features
->shaderStorageTexelBufferArrayDynamicIndexing
= true;
1074 features
->shaderUniformBufferArrayNonUniformIndexing
= false;
1075 features
->shaderSampledImageArrayNonUniformIndexing
= true;
1076 features
->shaderStorageBufferArrayNonUniformIndexing
= true;
1077 features
->shaderStorageImageArrayNonUniformIndexing
= true;
1078 features
->shaderInputAttachmentArrayNonUniformIndexing
= false;
1079 features
->shaderUniformTexelBufferArrayNonUniformIndexing
= true;
1080 features
->shaderStorageTexelBufferArrayNonUniformIndexing
= true;
1081 features
->descriptorBindingUniformBufferUpdateAfterBind
= false;
1082 features
->descriptorBindingSampledImageUpdateAfterBind
= true;
1083 features
->descriptorBindingStorageImageUpdateAfterBind
= true;
1084 features
->descriptorBindingStorageBufferUpdateAfterBind
= true;
1085 features
->descriptorBindingUniformTexelBufferUpdateAfterBind
= true;
1086 features
->descriptorBindingStorageTexelBufferUpdateAfterBind
= true;
1087 features
->descriptorBindingUpdateUnusedWhilePending
= true;
1088 features
->descriptorBindingPartiallyBound
= true;
1089 features
->descriptorBindingVariableDescriptorCount
= false;
1090 features
->runtimeDescriptorArray
= true;
1094 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_FEATURES_EXT
: {
1095 VkPhysicalDeviceInlineUniformBlockFeaturesEXT
*features
=
1096 (VkPhysicalDeviceInlineUniformBlockFeaturesEXT
*)ext
;
1097 features
->inlineUniformBlock
= true;
1098 features
->descriptorBindingInlineUniformBlockUpdateAfterBind
= true;
1102 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES
: {
1103 VkPhysicalDeviceMultiviewFeatures
*features
=
1104 (VkPhysicalDeviceMultiviewFeatures
*)ext
;
1105 features
->multiview
= true;
1106 features
->multiviewGeometryShader
= true;
1107 features
->multiviewTessellationShader
= true;
1111 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES
: {
1112 VkPhysicalDeviceProtectedMemoryFeatures
*features
= (void *)ext
;
1113 features
->protectedMemory
= false;
1117 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES
: {
1118 VkPhysicalDeviceSamplerYcbcrConversionFeatures
*features
=
1119 (VkPhysicalDeviceSamplerYcbcrConversionFeatures
*) ext
;
1120 features
->samplerYcbcrConversion
= true;
1124 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES_EXT
: {
1125 VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*features
=
1126 (VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*)ext
;
1127 features
->scalarBlockLayout
= true;
1131 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_INT64_FEATURES_KHR
: {
1132 VkPhysicalDeviceShaderAtomicInt64FeaturesKHR
*features
= (void *)ext
;
1133 features
->shaderBufferInt64Atomics
=
1134 pdevice
->info
.gen
>= 9 && pdevice
->use_softpin
;
1135 features
->shaderSharedInt64Atomics
= VK_FALSE
;
1139 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETERS_FEATURES
: {
1140 VkPhysicalDeviceShaderDrawParametersFeatures
*features
= (void *)ext
;
1141 features
->shaderDrawParameters
= true;
1145 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES
: {
1146 VkPhysicalDeviceVariablePointersFeatures
*features
= (void *)ext
;
1147 features
->variablePointersStorageBuffer
= true;
1148 features
->variablePointers
= true;
1152 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT
: {
1153 VkPhysicalDeviceTransformFeedbackFeaturesEXT
*features
=
1154 (VkPhysicalDeviceTransformFeedbackFeaturesEXT
*)ext
;
1155 features
->transformFeedback
= true;
1156 features
->geometryStreams
= true;
1160 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_UNIFORM_BUFFER_STANDARD_LAYOUT_FEATURES_KHR
: {
1161 VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR
*features
=
1162 (VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR
*)ext
;
1163 features
->uniformBufferStandardLayout
= true;
1167 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT
: {
1168 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*features
=
1169 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*)ext
;
1170 features
->vertexAttributeInstanceRateDivisor
= true;
1171 features
->vertexAttributeInstanceRateZeroDivisor
= true;
1175 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_YCBCR_IMAGE_ARRAYS_FEATURES_EXT
: {
1176 VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*features
=
1177 (VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*)ext
;
1178 features
->ycbcrImageArrays
= true;
1183 anv_debug_ignored_stype(ext
->sType
);
1189 #define MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS 64
1191 #define MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS 64
1192 #define MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS 256
1194 void anv_GetPhysicalDeviceProperties(
1195 VkPhysicalDevice physicalDevice
,
1196 VkPhysicalDeviceProperties
* pProperties
)
1198 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1199 const struct gen_device_info
*devinfo
= &pdevice
->info
;
1201 /* See assertions made when programming the buffer surface state. */
1202 const uint32_t max_raw_buffer_sz
= devinfo
->gen
>= 7 ?
1203 (1ul << 30) : (1ul << 27);
1205 const uint32_t max_ssbos
= pdevice
->has_a64_buffer_access
? UINT16_MAX
: 64;
1206 const uint32_t max_textures
=
1207 pdevice
->has_bindless_images
? UINT16_MAX
: 128;
1208 const uint32_t max_samplers
=
1209 pdevice
->has_bindless_samplers
? UINT16_MAX
:
1210 (devinfo
->gen
>= 8 || devinfo
->is_haswell
) ? 128 : 16;
1211 const uint32_t max_images
=
1212 pdevice
->has_bindless_images
? UINT16_MAX
: MAX_IMAGES
;
1214 /* The moment we have anything bindless, claim a high per-stage limit */
1215 const uint32_t max_per_stage
=
1216 pdevice
->has_a64_buffer_access
? UINT32_MAX
:
1217 MAX_BINDING_TABLE_SIZE
- MAX_RTS
;
1219 VkSampleCountFlags sample_counts
=
1220 isl_device_get_sample_counts(&pdevice
->isl_dev
);
1223 VkPhysicalDeviceLimits limits
= {
1224 .maxImageDimension1D
= (1 << 14),
1225 .maxImageDimension2D
= (1 << 14),
1226 .maxImageDimension3D
= (1 << 11),
1227 .maxImageDimensionCube
= (1 << 14),
1228 .maxImageArrayLayers
= (1 << 11),
1229 .maxTexelBufferElements
= 128 * 1024 * 1024,
1230 .maxUniformBufferRange
= (1ul << 27),
1231 .maxStorageBufferRange
= max_raw_buffer_sz
,
1232 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
1233 .maxMemoryAllocationCount
= UINT32_MAX
,
1234 .maxSamplerAllocationCount
= 64 * 1024,
1235 .bufferImageGranularity
= 64, /* A cache line */
1236 .sparseAddressSpaceSize
= 0,
1237 .maxBoundDescriptorSets
= MAX_SETS
,
1238 .maxPerStageDescriptorSamplers
= max_samplers
,
1239 .maxPerStageDescriptorUniformBuffers
= MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
,
1240 .maxPerStageDescriptorStorageBuffers
= max_ssbos
,
1241 .maxPerStageDescriptorSampledImages
= max_textures
,
1242 .maxPerStageDescriptorStorageImages
= max_images
,
1243 .maxPerStageDescriptorInputAttachments
= MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS
,
1244 .maxPerStageResources
= max_per_stage
,
1245 .maxDescriptorSetSamplers
= 6 * max_samplers
, /* number of stages * maxPerStageDescriptorSamplers */
1246 .maxDescriptorSetUniformBuffers
= 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
, /* number of stages * maxPerStageDescriptorUniformBuffers */
1247 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
1248 .maxDescriptorSetStorageBuffers
= 6 * max_ssbos
, /* number of stages * maxPerStageDescriptorStorageBuffers */
1249 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
1250 .maxDescriptorSetSampledImages
= 6 * max_textures
, /* number of stages * maxPerStageDescriptorSampledImages */
1251 .maxDescriptorSetStorageImages
= 6 * max_images
, /* number of stages * maxPerStageDescriptorStorageImages */
1252 .maxDescriptorSetInputAttachments
= MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS
,
1253 .maxVertexInputAttributes
= MAX_VBS
,
1254 .maxVertexInputBindings
= MAX_VBS
,
1255 .maxVertexInputAttributeOffset
= 2047,
1256 .maxVertexInputBindingStride
= 2048,
1257 .maxVertexOutputComponents
= 128,
1258 .maxTessellationGenerationLevel
= 64,
1259 .maxTessellationPatchSize
= 32,
1260 .maxTessellationControlPerVertexInputComponents
= 128,
1261 .maxTessellationControlPerVertexOutputComponents
= 128,
1262 .maxTessellationControlPerPatchOutputComponents
= 128,
1263 .maxTessellationControlTotalOutputComponents
= 2048,
1264 .maxTessellationEvaluationInputComponents
= 128,
1265 .maxTessellationEvaluationOutputComponents
= 128,
1266 .maxGeometryShaderInvocations
= 32,
1267 .maxGeometryInputComponents
= 64,
1268 .maxGeometryOutputComponents
= 128,
1269 .maxGeometryOutputVertices
= 256,
1270 .maxGeometryTotalOutputComponents
= 1024,
1271 .maxFragmentInputComponents
= 116, /* 128 components - (PSIZ, CLIP_DIST0, CLIP_DIST1) */
1272 .maxFragmentOutputAttachments
= 8,
1273 .maxFragmentDualSrcAttachments
= 1,
1274 .maxFragmentCombinedOutputResources
= 8,
1275 .maxComputeSharedMemorySize
= 32768,
1276 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
1277 .maxComputeWorkGroupInvocations
= 32 * devinfo
->max_cs_threads
,
1278 .maxComputeWorkGroupSize
= {
1279 16 * devinfo
->max_cs_threads
,
1280 16 * devinfo
->max_cs_threads
,
1281 16 * devinfo
->max_cs_threads
,
1283 .subPixelPrecisionBits
= 8,
1284 .subTexelPrecisionBits
= 8,
1285 .mipmapPrecisionBits
= 8,
1286 .maxDrawIndexedIndexValue
= UINT32_MAX
,
1287 .maxDrawIndirectCount
= UINT32_MAX
,
1288 .maxSamplerLodBias
= 16,
1289 .maxSamplerAnisotropy
= 16,
1290 .maxViewports
= MAX_VIEWPORTS
,
1291 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
1292 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
1293 .viewportSubPixelBits
= 13, /* We take a float? */
1294 .minMemoryMapAlignment
= 4096, /* A page */
1295 .minTexelBufferOffsetAlignment
= 1,
1296 /* We need 16 for UBO block reads to work and 32 for push UBOs */
1297 .minUniformBufferOffsetAlignment
= 32,
1298 .minStorageBufferOffsetAlignment
= 4,
1299 .minTexelOffset
= -8,
1300 .maxTexelOffset
= 7,
1301 .minTexelGatherOffset
= -32,
1302 .maxTexelGatherOffset
= 31,
1303 .minInterpolationOffset
= -0.5,
1304 .maxInterpolationOffset
= 0.4375,
1305 .subPixelInterpolationOffsetBits
= 4,
1306 .maxFramebufferWidth
= (1 << 14),
1307 .maxFramebufferHeight
= (1 << 14),
1308 .maxFramebufferLayers
= (1 << 11),
1309 .framebufferColorSampleCounts
= sample_counts
,
1310 .framebufferDepthSampleCounts
= sample_counts
,
1311 .framebufferStencilSampleCounts
= sample_counts
,
1312 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
1313 .maxColorAttachments
= MAX_RTS
,
1314 .sampledImageColorSampleCounts
= sample_counts
,
1315 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
1316 .sampledImageDepthSampleCounts
= sample_counts
,
1317 .sampledImageStencilSampleCounts
= sample_counts
,
1318 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
1319 .maxSampleMaskWords
= 1,
1320 .timestampComputeAndGraphics
= false,
1321 .timestampPeriod
= 1000000000.0 / devinfo
->timestamp_frequency
,
1322 .maxClipDistances
= 8,
1323 .maxCullDistances
= 8,
1324 .maxCombinedClipAndCullDistances
= 8,
1325 .discreteQueuePriorities
= 2,
1326 .pointSizeRange
= { 0.125, 255.875 },
1327 .lineWidthRange
= { 0.0, 7.9921875 },
1328 .pointSizeGranularity
= (1.0 / 8.0),
1329 .lineWidthGranularity
= (1.0 / 128.0),
1330 .strictLines
= false, /* FINISHME */
1331 .standardSampleLocations
= true,
1332 .optimalBufferCopyOffsetAlignment
= 128,
1333 .optimalBufferCopyRowPitchAlignment
= 128,
1334 .nonCoherentAtomSize
= 64,
1337 *pProperties
= (VkPhysicalDeviceProperties
) {
1338 .apiVersion
= anv_physical_device_api_version(pdevice
),
1339 .driverVersion
= vk_get_driver_version(),
1341 .deviceID
= pdevice
->chipset_id
,
1342 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
1344 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
1347 snprintf(pProperties
->deviceName
, sizeof(pProperties
->deviceName
),
1348 "%s", pdevice
->name
);
1349 memcpy(pProperties
->pipelineCacheUUID
,
1350 pdevice
->pipeline_cache_uuid
, VK_UUID_SIZE
);
1353 void anv_GetPhysicalDeviceProperties2(
1354 VkPhysicalDevice physicalDevice
,
1355 VkPhysicalDeviceProperties2
* pProperties
)
1357 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1359 anv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
1361 vk_foreach_struct(ext
, pProperties
->pNext
) {
1362 switch (ext
->sType
) {
1363 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_STENCIL_RESOLVE_PROPERTIES_KHR
: {
1364 VkPhysicalDeviceDepthStencilResolvePropertiesKHR
*props
=
1365 (VkPhysicalDeviceDepthStencilResolvePropertiesKHR
*)ext
;
1367 /* We support all of the depth resolve modes */
1368 props
->supportedDepthResolveModes
=
1369 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
|
1370 VK_RESOLVE_MODE_AVERAGE_BIT_KHR
|
1371 VK_RESOLVE_MODE_MIN_BIT_KHR
|
1372 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1374 /* Average doesn't make sense for stencil so we don't support that */
1375 props
->supportedStencilResolveModes
=
1376 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
;
1377 if (pdevice
->info
.gen
>= 8) {
1378 /* The advanced stencil resolve modes currently require stencil
1379 * sampling be supported by the hardware.
1381 props
->supportedStencilResolveModes
|=
1382 VK_RESOLVE_MODE_MIN_BIT_KHR
|
1383 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1386 props
->independentResolveNone
= true;
1387 props
->independentResolve
= true;
1391 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES_EXT
: {
1392 VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*props
=
1393 (VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*)ext
;
1395 /* It's a bit hard to exactly map our implementation to the limits
1396 * described here. The bindless surface handle in the extended
1397 * message descriptors is 20 bits and it's an index into the table of
1398 * RENDER_SURFACE_STATE structs that starts at bindless surface base
1399 * address. Given that most things consume two surface states per
1400 * view (general/sampled for textures and write-only/read-write for
1401 * images), we claim 2^19 things.
1403 * For SSBOs, we just use A64 messages so there is no real limit
1404 * there beyond the limit on the total size of a descriptor set.
1406 const unsigned max_bindless_views
= 1 << 19;
1408 props
->maxUpdateAfterBindDescriptorsInAllPools
= max_bindless_views
;
1409 props
->shaderUniformBufferArrayNonUniformIndexingNative
= false;
1410 props
->shaderSampledImageArrayNonUniformIndexingNative
= false;
1411 props
->shaderStorageBufferArrayNonUniformIndexingNative
= true;
1412 props
->shaderStorageImageArrayNonUniformIndexingNative
= false;
1413 props
->shaderInputAttachmentArrayNonUniformIndexingNative
= false;
1414 props
->robustBufferAccessUpdateAfterBind
= true;
1415 props
->quadDivergentImplicitLod
= false;
1416 props
->maxPerStageDescriptorUpdateAfterBindSamplers
= max_bindless_views
;
1417 props
->maxPerStageDescriptorUpdateAfterBindUniformBuffers
= MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
;
1418 props
->maxPerStageDescriptorUpdateAfterBindStorageBuffers
= UINT32_MAX
;
1419 props
->maxPerStageDescriptorUpdateAfterBindSampledImages
= max_bindless_views
;
1420 props
->maxPerStageDescriptorUpdateAfterBindStorageImages
= max_bindless_views
;
1421 props
->maxPerStageDescriptorUpdateAfterBindInputAttachments
= MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS
;
1422 props
->maxPerStageUpdateAfterBindResources
= UINT32_MAX
;
1423 props
->maxDescriptorSetUpdateAfterBindSamplers
= max_bindless_views
;
1424 props
->maxDescriptorSetUpdateAfterBindUniformBuffers
= 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
;
1425 props
->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2;
1426 props
->maxDescriptorSetUpdateAfterBindStorageBuffers
= UINT32_MAX
;
1427 props
->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2;
1428 props
->maxDescriptorSetUpdateAfterBindSampledImages
= max_bindless_views
;
1429 props
->maxDescriptorSetUpdateAfterBindStorageImages
= max_bindless_views
;
1430 props
->maxDescriptorSetUpdateAfterBindInputAttachments
= MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS
;
1434 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR
: {
1435 VkPhysicalDeviceDriverPropertiesKHR
*driver_props
=
1436 (VkPhysicalDeviceDriverPropertiesKHR
*) ext
;
1438 driver_props
->driverID
= VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA_KHR
;
1439 util_snprintf(driver_props
->driverName
, VK_MAX_DRIVER_NAME_SIZE_KHR
,
1440 "Intel open-source Mesa driver");
1442 util_snprintf(driver_props
->driverInfo
, VK_MAX_DRIVER_INFO_SIZE_KHR
,
1443 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1
);
1445 driver_props
->conformanceVersion
= (VkConformanceVersionKHR
) {
1454 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT
: {
1455 VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*props
=
1456 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*) ext
;
1457 /* Userptr needs page aligned memory. */
1458 props
->minImportedHostPointerAlignment
= 4096;
1462 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES
: {
1463 VkPhysicalDeviceIDProperties
*id_props
=
1464 (VkPhysicalDeviceIDProperties
*)ext
;
1465 memcpy(id_props
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
1466 memcpy(id_props
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
1467 /* The LUID is for Windows. */
1468 id_props
->deviceLUIDValid
= false;
1472 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_PROPERTIES_EXT
: {
1473 VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*props
=
1474 (VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*)ext
;
1475 props
->maxInlineUniformBlockSize
= MAX_INLINE_UNIFORM_BLOCK_SIZE
;
1476 props
->maxPerStageDescriptorInlineUniformBlocks
=
1477 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1478 props
->maxPerStageDescriptorUpdateAfterBindInlineUniformBlocks
=
1479 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1480 props
->maxDescriptorSetInlineUniformBlocks
=
1481 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1482 props
->maxDescriptorSetUpdateAfterBindInlineUniformBlocks
=
1483 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1487 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES
: {
1488 VkPhysicalDeviceMaintenance3Properties
*props
=
1489 (VkPhysicalDeviceMaintenance3Properties
*)ext
;
1490 /* This value doesn't matter for us today as our per-stage
1491 * descriptors are the real limit.
1493 props
->maxPerSetDescriptors
= 1024;
1494 props
->maxMemoryAllocationSize
= MAX_MEMORY_ALLOCATION_SIZE
;
1498 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES
: {
1499 VkPhysicalDeviceMultiviewProperties
*properties
=
1500 (VkPhysicalDeviceMultiviewProperties
*)ext
;
1501 properties
->maxMultiviewViewCount
= 16;
1502 properties
->maxMultiviewInstanceIndex
= UINT32_MAX
/ 16;
1506 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT
: {
1507 VkPhysicalDevicePCIBusInfoPropertiesEXT
*properties
=
1508 (VkPhysicalDevicePCIBusInfoPropertiesEXT
*)ext
;
1509 properties
->pciDomain
= pdevice
->pci_info
.domain
;
1510 properties
->pciBus
= pdevice
->pci_info
.bus
;
1511 properties
->pciDevice
= pdevice
->pci_info
.device
;
1512 properties
->pciFunction
= pdevice
->pci_info
.function
;
1516 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES
: {
1517 VkPhysicalDevicePointClippingProperties
*properties
=
1518 (VkPhysicalDevicePointClippingProperties
*) ext
;
1519 properties
->pointClippingBehavior
= VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES
;
1520 anv_finishme("Implement pop-free point clipping");
1524 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES
: {
1525 VkPhysicalDeviceProtectedMemoryProperties
*props
=
1526 (VkPhysicalDeviceProtectedMemoryProperties
*)ext
;
1527 props
->protectedNoFault
= false;
1531 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
1532 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
1533 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
1535 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
1539 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT
: {
1540 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*properties
=
1541 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*)ext
;
1542 properties
->filterMinmaxImageComponentMapping
= pdevice
->info
.gen
>= 9;
1543 properties
->filterMinmaxSingleComponentFormats
= true;
1547 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES
: {
1548 VkPhysicalDeviceSubgroupProperties
*properties
= (void *)ext
;
1550 properties
->subgroupSize
= BRW_SUBGROUP_SIZE
;
1552 VkShaderStageFlags scalar_stages
= 0;
1553 for (unsigned stage
= 0; stage
< MESA_SHADER_STAGES
; stage
++) {
1554 if (pdevice
->compiler
->scalar_stage
[stage
])
1555 scalar_stages
|= mesa_to_vk_shader_stage(stage
);
1557 properties
->supportedStages
= scalar_stages
;
1559 properties
->supportedOperations
= VK_SUBGROUP_FEATURE_BASIC_BIT
|
1560 VK_SUBGROUP_FEATURE_VOTE_BIT
|
1561 VK_SUBGROUP_FEATURE_ARITHMETIC_BIT
|
1562 VK_SUBGROUP_FEATURE_BALLOT_BIT
|
1563 VK_SUBGROUP_FEATURE_SHUFFLE_BIT
|
1564 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT
|
1565 VK_SUBGROUP_FEATURE_CLUSTERED_BIT
|
1566 VK_SUBGROUP_FEATURE_QUAD_BIT
;
1567 properties
->quadOperationsInAllStages
= true;
1571 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT
: {
1572 VkPhysicalDeviceTransformFeedbackPropertiesEXT
*props
=
1573 (VkPhysicalDeviceTransformFeedbackPropertiesEXT
*)ext
;
1575 props
->maxTransformFeedbackStreams
= MAX_XFB_STREAMS
;
1576 props
->maxTransformFeedbackBuffers
= MAX_XFB_BUFFERS
;
1577 props
->maxTransformFeedbackBufferSize
= (1ull << 32);
1578 props
->maxTransformFeedbackStreamDataSize
= 128 * 4;
1579 props
->maxTransformFeedbackBufferDataSize
= 128 * 4;
1580 props
->maxTransformFeedbackBufferDataStride
= 2048;
1581 props
->transformFeedbackQueries
= true;
1582 props
->transformFeedbackStreamsLinesTriangles
= false;
1583 props
->transformFeedbackRasterizationStreamSelect
= false;
1584 props
->transformFeedbackDraw
= true;
1588 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT
: {
1589 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*props
=
1590 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*)ext
;
1591 /* We have to restrict this a bit for multiview */
1592 props
->maxVertexAttribDivisor
= UINT32_MAX
/ 16;
1597 anv_debug_ignored_stype(ext
->sType
);
1603 /* We support exactly one queue family. */
1604 static const VkQueueFamilyProperties
1605 anv_queue_family_properties
= {
1606 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
1607 VK_QUEUE_COMPUTE_BIT
|
1608 VK_QUEUE_TRANSFER_BIT
,
1610 .timestampValidBits
= 36, /* XXX: Real value here */
1611 .minImageTransferGranularity
= { 1, 1, 1 },
1614 void anv_GetPhysicalDeviceQueueFamilyProperties(
1615 VkPhysicalDevice physicalDevice
,
1617 VkQueueFamilyProperties
* pQueueFamilyProperties
)
1619 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pCount
);
1621 vk_outarray_append(&out
, p
) {
1622 *p
= anv_queue_family_properties
;
1626 void anv_GetPhysicalDeviceQueueFamilyProperties2(
1627 VkPhysicalDevice physicalDevice
,
1628 uint32_t* pQueueFamilyPropertyCount
,
1629 VkQueueFamilyProperties2
* pQueueFamilyProperties
)
1632 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pQueueFamilyPropertyCount
);
1634 vk_outarray_append(&out
, p
) {
1635 p
->queueFamilyProperties
= anv_queue_family_properties
;
1637 vk_foreach_struct(s
, p
->pNext
) {
1638 anv_debug_ignored_stype(s
->sType
);
1643 void anv_GetPhysicalDeviceMemoryProperties(
1644 VkPhysicalDevice physicalDevice
,
1645 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
1647 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
1649 pMemoryProperties
->memoryTypeCount
= physical_device
->memory
.type_count
;
1650 for (uint32_t i
= 0; i
< physical_device
->memory
.type_count
; i
++) {
1651 pMemoryProperties
->memoryTypes
[i
] = (VkMemoryType
) {
1652 .propertyFlags
= physical_device
->memory
.types
[i
].propertyFlags
,
1653 .heapIndex
= physical_device
->memory
.types
[i
].heapIndex
,
1657 pMemoryProperties
->memoryHeapCount
= physical_device
->memory
.heap_count
;
1658 for (uint32_t i
= 0; i
< physical_device
->memory
.heap_count
; i
++) {
1659 pMemoryProperties
->memoryHeaps
[i
] = (VkMemoryHeap
) {
1660 .size
= physical_device
->memory
.heaps
[i
].size
,
1661 .flags
= physical_device
->memory
.heaps
[i
].flags
,
1667 anv_get_memory_budget(VkPhysicalDevice physicalDevice
,
1668 VkPhysicalDeviceMemoryBudgetPropertiesEXT
*memoryBudget
)
1670 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
1671 uint64_t sys_available
= get_available_system_memory();
1672 assert(sys_available
> 0);
1674 VkDeviceSize total_heaps_size
= 0;
1675 for (size_t i
= 0; i
< device
->memory
.heap_count
; i
++)
1676 total_heaps_size
+= device
->memory
.heaps
[i
].size
;
1678 for (size_t i
= 0; i
< device
->memory
.heap_count
; i
++) {
1679 VkDeviceSize heap_size
= device
->memory
.heaps
[i
].size
;
1680 VkDeviceSize heap_used
= device
->memory
.heaps
[i
].used
;
1681 VkDeviceSize heap_budget
;
1683 double heap_proportion
= (double) heap_size
/ total_heaps_size
;
1684 VkDeviceSize sys_available_prop
= sys_available
* heap_proportion
;
1687 * Let's not incite the app to starve the system: report at most 90% of
1688 * available system memory.
1690 uint64_t heap_available
= sys_available_prop
* 9 / 10;
1691 heap_budget
= MIN2(heap_size
, heap_used
+ heap_available
);
1694 * Round down to the nearest MB
1696 heap_budget
&= ~((1ull << 20) - 1);
1699 * The heapBudget value must be non-zero for array elements less than
1700 * VkPhysicalDeviceMemoryProperties::memoryHeapCount. The heapBudget
1701 * value must be less than or equal to VkMemoryHeap::size for each heap.
1703 assert(0 < heap_budget
&& heap_budget
<= heap_size
);
1705 memoryBudget
->heapUsage
[i
] = heap_used
;
1706 memoryBudget
->heapBudget
[i
] = heap_budget
;
1709 /* The heapBudget and heapUsage values must be zero for array elements
1710 * greater than or equal to VkPhysicalDeviceMemoryProperties::memoryHeapCount
1712 for (uint32_t i
= device
->memory
.heap_count
; i
< VK_MAX_MEMORY_HEAPS
; i
++) {
1713 memoryBudget
->heapBudget
[i
] = 0;
1714 memoryBudget
->heapUsage
[i
] = 0;
1718 void anv_GetPhysicalDeviceMemoryProperties2(
1719 VkPhysicalDevice physicalDevice
,
1720 VkPhysicalDeviceMemoryProperties2
* pMemoryProperties
)
1722 anv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
1723 &pMemoryProperties
->memoryProperties
);
1725 vk_foreach_struct(ext
, pMemoryProperties
->pNext
) {
1726 switch (ext
->sType
) {
1727 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT
:
1728 anv_get_memory_budget(physicalDevice
, (void*)ext
);
1731 anv_debug_ignored_stype(ext
->sType
);
1738 anv_GetDeviceGroupPeerMemoryFeatures(
1741 uint32_t localDeviceIndex
,
1742 uint32_t remoteDeviceIndex
,
1743 VkPeerMemoryFeatureFlags
* pPeerMemoryFeatures
)
1745 assert(localDeviceIndex
== 0 && remoteDeviceIndex
== 0);
1746 *pPeerMemoryFeatures
= VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT
|
1747 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT
|
1748 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT
|
1749 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT
;
1752 PFN_vkVoidFunction
anv_GetInstanceProcAddr(
1753 VkInstance _instance
,
1756 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1758 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
1759 * when we have to return valid function pointers, NULL, or it's left
1760 * undefined. See the table for exact details.
1765 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
1766 if (strcmp(pName, "vk" #entrypoint) == 0) \
1767 return (PFN_vkVoidFunction)anv_##entrypoint
1769 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties
);
1770 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties
);
1771 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceVersion
);
1772 LOOKUP_ANV_ENTRYPOINT(CreateInstance
);
1774 #undef LOOKUP_ANV_ENTRYPOINT
1776 if (instance
== NULL
)
1779 int idx
= anv_get_instance_entrypoint_index(pName
);
1781 return instance
->dispatch
.entrypoints
[idx
];
1783 idx
= anv_get_device_entrypoint_index(pName
);
1785 return instance
->device_dispatch
.entrypoints
[idx
];
1790 /* With version 1+ of the loader interface the ICD should expose
1791 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
1794 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1795 VkInstance instance
,
1799 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1800 VkInstance instance
,
1803 return anv_GetInstanceProcAddr(instance
, pName
);
1806 PFN_vkVoidFunction
anv_GetDeviceProcAddr(
1810 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1812 if (!device
|| !pName
)
1815 int idx
= anv_get_device_entrypoint_index(pName
);
1819 return device
->dispatch
.entrypoints
[idx
];
1823 anv_CreateDebugReportCallbackEXT(VkInstance _instance
,
1824 const VkDebugReportCallbackCreateInfoEXT
* pCreateInfo
,
1825 const VkAllocationCallbacks
* pAllocator
,
1826 VkDebugReportCallbackEXT
* pCallback
)
1828 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1829 return vk_create_debug_report_callback(&instance
->debug_report_callbacks
,
1830 pCreateInfo
, pAllocator
, &instance
->alloc
,
1835 anv_DestroyDebugReportCallbackEXT(VkInstance _instance
,
1836 VkDebugReportCallbackEXT _callback
,
1837 const VkAllocationCallbacks
* pAllocator
)
1839 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1840 vk_destroy_debug_report_callback(&instance
->debug_report_callbacks
,
1841 _callback
, pAllocator
, &instance
->alloc
);
1845 anv_DebugReportMessageEXT(VkInstance _instance
,
1846 VkDebugReportFlagsEXT flags
,
1847 VkDebugReportObjectTypeEXT objectType
,
1850 int32_t messageCode
,
1851 const char* pLayerPrefix
,
1852 const char* pMessage
)
1854 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1855 vk_debug_report(&instance
->debug_report_callbacks
, flags
, objectType
,
1856 object
, location
, messageCode
, pLayerPrefix
, pMessage
);
1860 anv_queue_init(struct anv_device
*device
, struct anv_queue
*queue
)
1862 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1863 queue
->device
= device
;
1868 anv_queue_finish(struct anv_queue
*queue
)
1872 static struct anv_state
1873 anv_state_pool_emit_data(struct anv_state_pool
*pool
, size_t size
, size_t align
, const void *p
)
1875 struct anv_state state
;
1877 state
= anv_state_pool_alloc(pool
, size
, align
);
1878 memcpy(state
.map
, p
, size
);
1883 struct gen8_border_color
{
1888 /* Pad out to 64 bytes */
1893 anv_device_init_border_colors(struct anv_device
*device
)
1895 static const struct gen8_border_color border_colors
[] = {
1896 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
1897 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
1898 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
1899 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
1900 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
1901 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
1904 device
->border_colors
= anv_state_pool_emit_data(&device
->dynamic_state_pool
,
1905 sizeof(border_colors
), 64,
1910 anv_device_init_trivial_batch(struct anv_device
*device
)
1912 anv_bo_init_new(&device
->trivial_batch_bo
, device
, 4096);
1914 if (device
->instance
->physicalDevice
.has_exec_async
)
1915 device
->trivial_batch_bo
.flags
|= EXEC_OBJECT_ASYNC
;
1917 if (device
->instance
->physicalDevice
.use_softpin
)
1918 device
->trivial_batch_bo
.flags
|= EXEC_OBJECT_PINNED
;
1920 anv_vma_alloc(device
, &device
->trivial_batch_bo
);
1922 void *map
= anv_gem_mmap(device
, device
->trivial_batch_bo
.gem_handle
,
1925 struct anv_batch batch
= {
1931 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
1932 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
1934 if (!device
->info
.has_llc
)
1935 gen_clflush_range(map
, batch
.next
- map
);
1937 anv_gem_munmap(map
, device
->trivial_batch_bo
.size
);
1940 VkResult
anv_EnumerateDeviceExtensionProperties(
1941 VkPhysicalDevice physicalDevice
,
1942 const char* pLayerName
,
1943 uint32_t* pPropertyCount
,
1944 VkExtensionProperties
* pProperties
)
1946 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
1947 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
1949 for (int i
= 0; i
< ANV_DEVICE_EXTENSION_COUNT
; i
++) {
1950 if (device
->supported_extensions
.extensions
[i
]) {
1951 vk_outarray_append(&out
, prop
) {
1952 *prop
= anv_device_extensions
[i
];
1957 return vk_outarray_status(&out
);
1961 anv_device_init_dispatch(struct anv_device
*device
)
1963 const struct anv_device_dispatch_table
*genX_table
;
1964 switch (device
->info
.gen
) {
1966 genX_table
= &gen11_device_dispatch_table
;
1969 genX_table
= &gen10_device_dispatch_table
;
1972 genX_table
= &gen9_device_dispatch_table
;
1975 genX_table
= &gen8_device_dispatch_table
;
1978 if (device
->info
.is_haswell
)
1979 genX_table
= &gen75_device_dispatch_table
;
1981 genX_table
= &gen7_device_dispatch_table
;
1984 unreachable("unsupported gen\n");
1987 for (unsigned i
= 0; i
< ARRAY_SIZE(device
->dispatch
.entrypoints
); i
++) {
1988 /* Vulkan requires that entrypoints for extensions which have not been
1989 * enabled must not be advertised.
1991 if (!anv_device_entrypoint_is_enabled(i
, device
->instance
->app_info
.api_version
,
1992 &device
->instance
->enabled_extensions
,
1993 &device
->enabled_extensions
)) {
1994 device
->dispatch
.entrypoints
[i
] = NULL
;
1995 } else if (genX_table
->entrypoints
[i
]) {
1996 device
->dispatch
.entrypoints
[i
] = genX_table
->entrypoints
[i
];
1998 device
->dispatch
.entrypoints
[i
] =
1999 anv_device_dispatch_table
.entrypoints
[i
];
2005 vk_priority_to_gen(int priority
)
2008 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT
:
2009 return GEN_CONTEXT_LOW_PRIORITY
;
2010 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
:
2011 return GEN_CONTEXT_MEDIUM_PRIORITY
;
2012 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT
:
2013 return GEN_CONTEXT_HIGH_PRIORITY
;
2014 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT
:
2015 return GEN_CONTEXT_REALTIME_PRIORITY
;
2017 unreachable("Invalid priority");
2022 anv_device_init_hiz_clear_value_bo(struct anv_device
*device
)
2024 anv_bo_init_new(&device
->hiz_clear_bo
, device
, 4096);
2026 if (device
->instance
->physicalDevice
.has_exec_async
)
2027 device
->hiz_clear_bo
.flags
|= EXEC_OBJECT_ASYNC
;
2029 if (device
->instance
->physicalDevice
.use_softpin
)
2030 device
->hiz_clear_bo
.flags
|= EXEC_OBJECT_PINNED
;
2032 anv_vma_alloc(device
, &device
->hiz_clear_bo
);
2034 uint32_t *map
= anv_gem_mmap(device
, device
->hiz_clear_bo
.gem_handle
,
2037 union isl_color_value hiz_clear
= { .u32
= { 0, } };
2038 hiz_clear
.f32
[0] = ANV_HZ_FC_VAL
;
2040 memcpy(map
, hiz_clear
.u32
, sizeof(hiz_clear
.u32
));
2041 anv_gem_munmap(map
, device
->hiz_clear_bo
.size
);
2045 get_bo_from_pool(struct gen_batch_decode_bo
*ret
,
2046 struct anv_block_pool
*pool
,
2049 for (uint32_t i
= 0; i
< pool
->nbos
; i
++) {
2050 uint64_t bo_address
= pool
->bos
[i
].offset
& (~0ull >> 16);
2051 uint32_t bo_size
= pool
->bos
[i
].size
;
2052 if (address
>= bo_address
&& address
< (bo_address
+ bo_size
)) {
2053 *ret
= (struct gen_batch_decode_bo
) {
2056 .map
= pool
->bos
[i
].map
,
2064 /* Finding a buffer for batch decoding */
2065 static struct gen_batch_decode_bo
2066 decode_get_bo(void *v_batch
, bool ppgtt
, uint64_t address
)
2068 struct anv_device
*device
= v_batch
;
2069 struct gen_batch_decode_bo ret_bo
= {};
2073 if (get_bo_from_pool(&ret_bo
, &device
->dynamic_state_pool
.block_pool
, address
))
2075 if (get_bo_from_pool(&ret_bo
, &device
->instruction_state_pool
.block_pool
, address
))
2077 if (get_bo_from_pool(&ret_bo
, &device
->binding_table_pool
.block_pool
, address
))
2079 if (get_bo_from_pool(&ret_bo
, &device
->surface_state_pool
.block_pool
, address
))
2082 if (!device
->cmd_buffer_being_decoded
)
2083 return (struct gen_batch_decode_bo
) { };
2085 struct anv_batch_bo
**bo
;
2087 u_vector_foreach(bo
, &device
->cmd_buffer_being_decoded
->seen_bbos
) {
2088 /* The decoder zeroes out the top 16 bits, so we need to as well */
2089 uint64_t bo_address
= (*bo
)->bo
.offset
& (~0ull >> 16);
2091 if (address
>= bo_address
&& address
< bo_address
+ (*bo
)->bo
.size
) {
2092 return (struct gen_batch_decode_bo
) {
2094 .size
= (*bo
)->bo
.size
,
2095 .map
= (*bo
)->bo
.map
,
2100 return (struct gen_batch_decode_bo
) { };
2103 VkResult
anv_CreateDevice(
2104 VkPhysicalDevice physicalDevice
,
2105 const VkDeviceCreateInfo
* pCreateInfo
,
2106 const VkAllocationCallbacks
* pAllocator
,
2109 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
2111 struct anv_device
*device
;
2113 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO
);
2115 struct anv_device_extension_table enabled_extensions
= { };
2116 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
2118 for (idx
= 0; idx
< ANV_DEVICE_EXTENSION_COUNT
; idx
++) {
2119 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
2120 anv_device_extensions
[idx
].extensionName
) == 0)
2124 if (idx
>= ANV_DEVICE_EXTENSION_COUNT
)
2125 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
2127 if (!physical_device
->supported_extensions
.extensions
[idx
])
2128 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
2130 enabled_extensions
.extensions
[idx
] = true;
2133 /* Check enabled features */
2134 if (pCreateInfo
->pEnabledFeatures
) {
2135 VkPhysicalDeviceFeatures supported_features
;
2136 anv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
2137 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
2138 VkBool32
*enabled_feature
= (VkBool32
*)pCreateInfo
->pEnabledFeatures
;
2139 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
2140 for (uint32_t i
= 0; i
< num_features
; i
++) {
2141 if (enabled_feature
[i
] && !supported_feature
[i
])
2142 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
2146 /* Check requested queues and fail if we are requested to create any
2147 * queues with flags we don't support.
2149 assert(pCreateInfo
->queueCreateInfoCount
> 0);
2150 for (uint32_t i
= 0; i
< pCreateInfo
->queueCreateInfoCount
; i
++) {
2151 if (pCreateInfo
->pQueueCreateInfos
[i
].flags
!= 0)
2152 return vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2155 /* Check if client specified queue priority. */
2156 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*queue_priority
=
2157 vk_find_struct_const(pCreateInfo
->pQueueCreateInfos
[0].pNext
,
2158 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT
);
2160 VkQueueGlobalPriorityEXT priority
=
2161 queue_priority
? queue_priority
->globalPriority
:
2162 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
;
2164 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
2166 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
2168 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2170 const unsigned decode_flags
=
2171 GEN_BATCH_DECODE_FULL
|
2172 ((INTEL_DEBUG
& DEBUG_COLOR
) ? GEN_BATCH_DECODE_IN_COLOR
: 0) |
2173 GEN_BATCH_DECODE_OFFSETS
|
2174 GEN_BATCH_DECODE_FLOATS
;
2176 gen_batch_decode_ctx_init(&device
->decoder_ctx
,
2177 &physical_device
->info
,
2178 stderr
, decode_flags
, NULL
,
2179 decode_get_bo
, NULL
, device
);
2181 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
2182 device
->instance
= physical_device
->instance
;
2183 device
->chipset_id
= physical_device
->chipset_id
;
2184 device
->no_hw
= physical_device
->no_hw
;
2185 device
->_lost
= false;
2188 device
->alloc
= *pAllocator
;
2190 device
->alloc
= physical_device
->instance
->alloc
;
2192 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
2193 device
->fd
= open(physical_device
->path
, O_RDWR
| O_CLOEXEC
);
2194 if (device
->fd
== -1) {
2195 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2199 device
->context_id
= anv_gem_create_context(device
);
2200 if (device
->context_id
== -1) {
2201 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2205 if (physical_device
->use_softpin
) {
2206 if (pthread_mutex_init(&device
->vma_mutex
, NULL
) != 0) {
2207 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2211 /* keep the page with address zero out of the allocator */
2212 struct anv_memory_heap
*low_heap
=
2213 &physical_device
->memory
.heaps
[physical_device
->memory
.heap_count
- 1];
2214 util_vma_heap_init(&device
->vma_lo
, low_heap
->vma_start
, low_heap
->vma_size
);
2215 device
->vma_lo_available
= low_heap
->size
;
2217 struct anv_memory_heap
*high_heap
=
2218 &physical_device
->memory
.heaps
[0];
2219 util_vma_heap_init(&device
->vma_hi
, high_heap
->vma_start
, high_heap
->vma_size
);
2220 device
->vma_hi_available
= physical_device
->memory
.heap_count
== 1 ? 0 :
2224 list_inithead(&device
->memory_objects
);
2226 /* As per spec, the driver implementation may deny requests to acquire
2227 * a priority above the default priority (MEDIUM) if the caller does not
2228 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
2231 if (physical_device
->has_context_priority
) {
2232 int err
= anv_gem_set_context_param(device
->fd
, device
->context_id
,
2233 I915_CONTEXT_PARAM_PRIORITY
,
2234 vk_priority_to_gen(priority
));
2235 if (err
!= 0 && priority
> VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
) {
2236 result
= vk_error(VK_ERROR_NOT_PERMITTED_EXT
);
2241 device
->info
= physical_device
->info
;
2242 device
->isl_dev
= physical_device
->isl_dev
;
2244 /* On Broadwell and later, we can use batch chaining to more efficiently
2245 * implement growing command buffers. Prior to Haswell, the kernel
2246 * command parser gets in the way and we have to fall back to growing
2249 device
->can_chain_batches
= device
->info
.gen
>= 8;
2251 device
->robust_buffer_access
= pCreateInfo
->pEnabledFeatures
&&
2252 pCreateInfo
->pEnabledFeatures
->robustBufferAccess
;
2253 device
->enabled_extensions
= enabled_extensions
;
2255 anv_device_init_dispatch(device
);
2257 if (pthread_mutex_init(&device
->mutex
, NULL
) != 0) {
2258 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2259 goto fail_context_id
;
2262 pthread_condattr_t condattr
;
2263 if (pthread_condattr_init(&condattr
) != 0) {
2264 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2267 if (pthread_condattr_setclock(&condattr
, CLOCK_MONOTONIC
) != 0) {
2268 pthread_condattr_destroy(&condattr
);
2269 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2272 if (pthread_cond_init(&device
->queue_submit
, &condattr
) != 0) {
2273 pthread_condattr_destroy(&condattr
);
2274 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2277 pthread_condattr_destroy(&condattr
);
2280 (physical_device
->supports_48bit_addresses
? EXEC_OBJECT_SUPPORTS_48B_ADDRESS
: 0) |
2281 (physical_device
->has_exec_async
? EXEC_OBJECT_ASYNC
: 0) |
2282 (physical_device
->has_exec_capture
? EXEC_OBJECT_CAPTURE
: 0) |
2283 (physical_device
->use_softpin
? EXEC_OBJECT_PINNED
: 0);
2285 anv_bo_pool_init(&device
->batch_bo_pool
, device
, bo_flags
);
2287 result
= anv_bo_cache_init(&device
->bo_cache
);
2288 if (result
!= VK_SUCCESS
)
2289 goto fail_batch_bo_pool
;
2291 if (!physical_device
->use_softpin
)
2292 bo_flags
&= ~EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
2294 result
= anv_state_pool_init(&device
->dynamic_state_pool
, device
,
2295 DYNAMIC_STATE_POOL_MIN_ADDRESS
,
2298 if (result
!= VK_SUCCESS
)
2301 result
= anv_state_pool_init(&device
->instruction_state_pool
, device
,
2302 INSTRUCTION_STATE_POOL_MIN_ADDRESS
,
2305 if (result
!= VK_SUCCESS
)
2306 goto fail_dynamic_state_pool
;
2308 result
= anv_state_pool_init(&device
->surface_state_pool
, device
,
2309 SURFACE_STATE_POOL_MIN_ADDRESS
,
2312 if (result
!= VK_SUCCESS
)
2313 goto fail_instruction_state_pool
;
2315 if (physical_device
->use_softpin
) {
2316 result
= anv_state_pool_init(&device
->binding_table_pool
, device
,
2317 BINDING_TABLE_POOL_MIN_ADDRESS
,
2320 if (result
!= VK_SUCCESS
)
2321 goto fail_surface_state_pool
;
2324 result
= anv_bo_init_new(&device
->workaround_bo
, device
, 1024);
2325 if (result
!= VK_SUCCESS
)
2326 goto fail_binding_table_pool
;
2328 if (physical_device
->use_softpin
)
2329 device
->workaround_bo
.flags
|= EXEC_OBJECT_PINNED
;
2331 if (!anv_vma_alloc(device
, &device
->workaround_bo
))
2332 goto fail_workaround_bo
;
2334 anv_device_init_trivial_batch(device
);
2336 if (device
->info
.gen
>= 10)
2337 anv_device_init_hiz_clear_value_bo(device
);
2339 anv_scratch_pool_init(device
, &device
->scratch_pool
);
2341 anv_queue_init(device
, &device
->queue
);
2343 switch (device
->info
.gen
) {
2345 if (!device
->info
.is_haswell
)
2346 result
= gen7_init_device_state(device
);
2348 result
= gen75_init_device_state(device
);
2351 result
= gen8_init_device_state(device
);
2354 result
= gen9_init_device_state(device
);
2357 result
= gen10_init_device_state(device
);
2360 result
= gen11_init_device_state(device
);
2363 /* Shouldn't get here as we don't create physical devices for any other
2365 unreachable("unhandled gen");
2367 if (result
!= VK_SUCCESS
)
2368 goto fail_workaround_bo
;
2370 anv_pipeline_cache_init(&device
->default_pipeline_cache
, device
, true);
2372 anv_device_init_blorp(device
);
2374 anv_device_init_border_colors(device
);
2376 *pDevice
= anv_device_to_handle(device
);
2381 anv_queue_finish(&device
->queue
);
2382 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
2383 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
2384 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
2385 fail_binding_table_pool
:
2386 if (physical_device
->use_softpin
)
2387 anv_state_pool_finish(&device
->binding_table_pool
);
2388 fail_surface_state_pool
:
2389 anv_state_pool_finish(&device
->surface_state_pool
);
2390 fail_instruction_state_pool
:
2391 anv_state_pool_finish(&device
->instruction_state_pool
);
2392 fail_dynamic_state_pool
:
2393 anv_state_pool_finish(&device
->dynamic_state_pool
);
2395 anv_bo_cache_finish(&device
->bo_cache
);
2397 anv_bo_pool_finish(&device
->batch_bo_pool
);
2398 pthread_cond_destroy(&device
->queue_submit
);
2400 pthread_mutex_destroy(&device
->mutex
);
2402 anv_gem_destroy_context(device
, device
->context_id
);
2406 vk_free(&device
->alloc
, device
);
2411 void anv_DestroyDevice(
2413 const VkAllocationCallbacks
* pAllocator
)
2415 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2416 struct anv_physical_device
*physical_device
;
2421 physical_device
= &device
->instance
->physicalDevice
;
2423 anv_device_finish_blorp(device
);
2425 anv_pipeline_cache_finish(&device
->default_pipeline_cache
);
2427 anv_queue_finish(&device
->queue
);
2429 #ifdef HAVE_VALGRIND
2430 /* We only need to free these to prevent valgrind errors. The backing
2431 * BO will go away in a couple of lines so we don't actually leak.
2433 anv_state_pool_free(&device
->dynamic_state_pool
, device
->border_colors
);
2436 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
2438 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
2439 anv_vma_free(device
, &device
->workaround_bo
);
2440 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
2442 anv_vma_free(device
, &device
->trivial_batch_bo
);
2443 anv_gem_close(device
, device
->trivial_batch_bo
.gem_handle
);
2444 if (device
->info
.gen
>= 10)
2445 anv_gem_close(device
, device
->hiz_clear_bo
.gem_handle
);
2447 if (physical_device
->use_softpin
)
2448 anv_state_pool_finish(&device
->binding_table_pool
);
2449 anv_state_pool_finish(&device
->surface_state_pool
);
2450 anv_state_pool_finish(&device
->instruction_state_pool
);
2451 anv_state_pool_finish(&device
->dynamic_state_pool
);
2453 anv_bo_cache_finish(&device
->bo_cache
);
2455 anv_bo_pool_finish(&device
->batch_bo_pool
);
2457 pthread_cond_destroy(&device
->queue_submit
);
2458 pthread_mutex_destroy(&device
->mutex
);
2460 anv_gem_destroy_context(device
, device
->context_id
);
2462 gen_batch_decode_ctx_finish(&device
->decoder_ctx
);
2466 vk_free(&device
->alloc
, device
);
2469 VkResult
anv_EnumerateInstanceLayerProperties(
2470 uint32_t* pPropertyCount
,
2471 VkLayerProperties
* pProperties
)
2473 if (pProperties
== NULL
) {
2474 *pPropertyCount
= 0;
2478 /* None supported at this time */
2479 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
2482 VkResult
anv_EnumerateDeviceLayerProperties(
2483 VkPhysicalDevice physicalDevice
,
2484 uint32_t* pPropertyCount
,
2485 VkLayerProperties
* pProperties
)
2487 if (pProperties
== NULL
) {
2488 *pPropertyCount
= 0;
2492 /* None supported at this time */
2493 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
2496 void anv_GetDeviceQueue(
2498 uint32_t queueNodeIndex
,
2499 uint32_t queueIndex
,
2502 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2504 assert(queueIndex
== 0);
2506 *pQueue
= anv_queue_to_handle(&device
->queue
);
2509 void anv_GetDeviceQueue2(
2511 const VkDeviceQueueInfo2
* pQueueInfo
,
2514 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2516 assert(pQueueInfo
->queueIndex
== 0);
2518 if (pQueueInfo
->flags
== device
->queue
.flags
)
2519 *pQueue
= anv_queue_to_handle(&device
->queue
);
2525 _anv_device_set_lost(struct anv_device
*device
,
2526 const char *file
, int line
,
2527 const char *msg
, ...)
2532 device
->_lost
= true;
2535 err
= __vk_errorv(device
->instance
, device
,
2536 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT
,
2537 VK_ERROR_DEVICE_LOST
, file
, line
, msg
, ap
);
2540 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
2547 anv_device_query_status(struct anv_device
*device
)
2549 /* This isn't likely as most of the callers of this function already check
2550 * for it. However, it doesn't hurt to check and it potentially lets us
2553 if (anv_device_is_lost(device
))
2554 return VK_ERROR_DEVICE_LOST
;
2556 uint32_t active
, pending
;
2557 int ret
= anv_gem_gpu_get_reset_stats(device
, &active
, &pending
);
2559 /* We don't know the real error. */
2560 return anv_device_set_lost(device
, "get_reset_stats failed: %m");
2564 return anv_device_set_lost(device
, "GPU hung on one of our command buffers");
2565 } else if (pending
) {
2566 return anv_device_set_lost(device
, "GPU hung with commands in-flight");
2573 anv_device_bo_busy(struct anv_device
*device
, struct anv_bo
*bo
)
2575 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
2576 * Other usages of the BO (such as on different hardware) will not be
2577 * flagged as "busy" by this ioctl. Use with care.
2579 int ret
= anv_gem_busy(device
, bo
->gem_handle
);
2581 return VK_NOT_READY
;
2582 } else if (ret
== -1) {
2583 /* We don't know the real error. */
2584 return anv_device_set_lost(device
, "gem wait failed: %m");
2587 /* Query for device status after the busy call. If the BO we're checking
2588 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
2589 * client because it clearly doesn't have valid data. Yes, this most
2590 * likely means an ioctl, but we just did an ioctl to query the busy status
2591 * so it's no great loss.
2593 return anv_device_query_status(device
);
2597 anv_device_wait(struct anv_device
*device
, struct anv_bo
*bo
,
2600 int ret
= anv_gem_wait(device
, bo
->gem_handle
, &timeout
);
2601 if (ret
== -1 && errno
== ETIME
) {
2603 } else if (ret
== -1) {
2604 /* We don't know the real error. */
2605 return anv_device_set_lost(device
, "gem wait failed: %m");
2608 /* Query for device status after the wait. If the BO we're waiting on got
2609 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
2610 * because it clearly doesn't have valid data. Yes, this most likely means
2611 * an ioctl, but we just did an ioctl to wait so it's no great loss.
2613 return anv_device_query_status(device
);
2616 VkResult
anv_DeviceWaitIdle(
2619 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2620 if (anv_device_is_lost(device
))
2621 return VK_ERROR_DEVICE_LOST
;
2623 struct anv_batch batch
;
2626 batch
.start
= batch
.next
= cmds
;
2627 batch
.end
= (void *) cmds
+ sizeof(cmds
);
2629 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
2630 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
2632 return anv_device_submit_simple_batch(device
, &batch
);
2636 anv_vma_alloc(struct anv_device
*device
, struct anv_bo
*bo
)
2638 if (!(bo
->flags
& EXEC_OBJECT_PINNED
))
2641 pthread_mutex_lock(&device
->vma_mutex
);
2645 if (bo
->flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
&&
2646 device
->vma_hi_available
>= bo
->size
) {
2647 uint64_t addr
= util_vma_heap_alloc(&device
->vma_hi
, bo
->size
, 4096);
2649 bo
->offset
= gen_canonical_address(addr
);
2650 assert(addr
== gen_48b_address(bo
->offset
));
2651 device
->vma_hi_available
-= bo
->size
;
2655 if (bo
->offset
== 0 && device
->vma_lo_available
>= bo
->size
) {
2656 uint64_t addr
= util_vma_heap_alloc(&device
->vma_lo
, bo
->size
, 4096);
2658 bo
->offset
= gen_canonical_address(addr
);
2659 assert(addr
== gen_48b_address(bo
->offset
));
2660 device
->vma_lo_available
-= bo
->size
;
2664 pthread_mutex_unlock(&device
->vma_mutex
);
2666 return bo
->offset
!= 0;
2670 anv_vma_free(struct anv_device
*device
, struct anv_bo
*bo
)
2672 if (!(bo
->flags
& EXEC_OBJECT_PINNED
))
2675 const uint64_t addr_48b
= gen_48b_address(bo
->offset
);
2677 pthread_mutex_lock(&device
->vma_mutex
);
2679 if (addr_48b
>= LOW_HEAP_MIN_ADDRESS
&&
2680 addr_48b
<= LOW_HEAP_MAX_ADDRESS
) {
2681 util_vma_heap_free(&device
->vma_lo
, addr_48b
, bo
->size
);
2682 device
->vma_lo_available
+= bo
->size
;
2684 MAYBE_UNUSED
const struct anv_physical_device
*physical_device
=
2685 &device
->instance
->physicalDevice
;
2686 assert(addr_48b
>= physical_device
->memory
.heaps
[0].vma_start
&&
2687 addr_48b
< (physical_device
->memory
.heaps
[0].vma_start
+
2688 physical_device
->memory
.heaps
[0].vma_size
));
2689 util_vma_heap_free(&device
->vma_hi
, addr_48b
, bo
->size
);
2690 device
->vma_hi_available
+= bo
->size
;
2693 pthread_mutex_unlock(&device
->vma_mutex
);
2699 anv_bo_init_new(struct anv_bo
*bo
, struct anv_device
*device
, uint64_t size
)
2701 uint32_t gem_handle
= anv_gem_create(device
, size
);
2703 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
2705 anv_bo_init(bo
, gem_handle
, size
);
2710 VkResult
anv_AllocateMemory(
2712 const VkMemoryAllocateInfo
* pAllocateInfo
,
2713 const VkAllocationCallbacks
* pAllocator
,
2714 VkDeviceMemory
* pMem
)
2716 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2717 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2718 struct anv_device_memory
*mem
;
2719 VkResult result
= VK_SUCCESS
;
2721 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
2723 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
2724 assert(pAllocateInfo
->allocationSize
> 0);
2726 if (pAllocateInfo
->allocationSize
> MAX_MEMORY_ALLOCATION_SIZE
)
2727 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
2729 /* FINISHME: Fail if allocation request exceeds heap size. */
2731 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
2732 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2734 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2736 assert(pAllocateInfo
->memoryTypeIndex
< pdevice
->memory
.type_count
);
2737 mem
->type
= &pdevice
->memory
.types
[pAllocateInfo
->memoryTypeIndex
];
2741 mem
->host_ptr
= NULL
;
2743 uint64_t bo_flags
= 0;
2745 assert(mem
->type
->heapIndex
< pdevice
->memory
.heap_count
);
2746 if (pdevice
->memory
.heaps
[mem
->type
->heapIndex
].supports_48bit_addresses
)
2747 bo_flags
|= EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
2749 const struct wsi_memory_allocate_info
*wsi_info
=
2750 vk_find_struct_const(pAllocateInfo
->pNext
, WSI_MEMORY_ALLOCATE_INFO_MESA
);
2751 if (wsi_info
&& wsi_info
->implicit_sync
) {
2752 /* We need to set the WRITE flag on window system buffers so that GEM
2753 * will know we're writing to them and synchronize uses on other rings
2754 * (eg if the display server uses the blitter ring).
2756 bo_flags
|= EXEC_OBJECT_WRITE
;
2757 } else if (pdevice
->has_exec_async
) {
2758 bo_flags
|= EXEC_OBJECT_ASYNC
;
2761 if (pdevice
->use_softpin
)
2762 bo_flags
|= EXEC_OBJECT_PINNED
;
2764 const VkExportMemoryAllocateInfo
*export_info
=
2765 vk_find_struct_const(pAllocateInfo
->pNext
, EXPORT_MEMORY_ALLOCATE_INFO
);
2767 /* Check if we need to support Android HW buffer export. If so,
2768 * create AHardwareBuffer and import memory from it.
2770 bool android_export
= false;
2771 if (export_info
&& export_info
->handleTypes
&
2772 VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID
)
2773 android_export
= true;
2775 /* Android memory import. */
2776 const struct VkImportAndroidHardwareBufferInfoANDROID
*ahw_import_info
=
2777 vk_find_struct_const(pAllocateInfo
->pNext
,
2778 IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID
);
2780 if (ahw_import_info
) {
2781 result
= anv_import_ahw_memory(_device
, mem
, ahw_import_info
);
2782 if (result
!= VK_SUCCESS
)
2786 } else if (android_export
) {
2787 result
= anv_create_ahw_memory(_device
, mem
, pAllocateInfo
);
2788 if (result
!= VK_SUCCESS
)
2791 const struct VkImportAndroidHardwareBufferInfoANDROID import_info
= {
2794 result
= anv_import_ahw_memory(_device
, mem
, &import_info
);
2795 if (result
!= VK_SUCCESS
)
2801 const VkImportMemoryFdInfoKHR
*fd_info
=
2802 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_FD_INFO_KHR
);
2804 /* The Vulkan spec permits handleType to be 0, in which case the struct is
2807 if (fd_info
&& fd_info
->handleType
) {
2808 /* At the moment, we support only the below handle types. */
2809 assert(fd_info
->handleType
==
2810 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
2811 fd_info
->handleType
==
2812 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
2814 result
= anv_bo_cache_import(device
, &device
->bo_cache
, fd_info
->fd
,
2815 bo_flags
| ANV_BO_EXTERNAL
, &mem
->bo
);
2816 if (result
!= VK_SUCCESS
)
2819 VkDeviceSize aligned_alloc_size
=
2820 align_u64(pAllocateInfo
->allocationSize
, 4096);
2822 /* For security purposes, we reject importing the bo if it's smaller
2823 * than the requested allocation size. This prevents a malicious client
2824 * from passing a buffer to a trusted client, lying about the size, and
2825 * telling the trusted client to try and texture from an image that goes
2826 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
2827 * in the trusted client. The trusted client can protect itself against
2828 * this sort of attack but only if it can trust the buffer size.
2830 if (mem
->bo
->size
< aligned_alloc_size
) {
2831 result
= vk_errorf(device
->instance
, device
,
2832 VK_ERROR_INVALID_EXTERNAL_HANDLE
,
2833 "aligned allocationSize too large for "
2834 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT: "
2835 "%"PRIu64
"B > %"PRIu64
"B",
2836 aligned_alloc_size
, mem
->bo
->size
);
2837 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
2841 /* From the Vulkan spec:
2843 * "Importing memory from a file descriptor transfers ownership of
2844 * the file descriptor from the application to the Vulkan
2845 * implementation. The application must not perform any operations on
2846 * the file descriptor after a successful import."
2848 * If the import fails, we leave the file descriptor open.
2854 const VkImportMemoryHostPointerInfoEXT
*host_ptr_info
=
2855 vk_find_struct_const(pAllocateInfo
->pNext
,
2856 IMPORT_MEMORY_HOST_POINTER_INFO_EXT
);
2857 if (host_ptr_info
&& host_ptr_info
->handleType
) {
2858 if (host_ptr_info
->handleType
==
2859 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_MAPPED_FOREIGN_MEMORY_BIT_EXT
) {
2860 result
= vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
2864 assert(host_ptr_info
->handleType
==
2865 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
);
2867 result
= anv_bo_cache_import_host_ptr(
2868 device
, &device
->bo_cache
, host_ptr_info
->pHostPointer
,
2869 pAllocateInfo
->allocationSize
, bo_flags
, &mem
->bo
);
2871 if (result
!= VK_SUCCESS
)
2874 mem
->host_ptr
= host_ptr_info
->pHostPointer
;
2878 /* Regular allocate (not importing memory). */
2880 if (export_info
&& export_info
->handleTypes
)
2881 bo_flags
|= ANV_BO_EXTERNAL
;
2883 result
= anv_bo_cache_alloc(device
, &device
->bo_cache
,
2884 pAllocateInfo
->allocationSize
, bo_flags
,
2886 if (result
!= VK_SUCCESS
)
2889 const VkMemoryDedicatedAllocateInfo
*dedicated_info
=
2890 vk_find_struct_const(pAllocateInfo
->pNext
, MEMORY_DEDICATED_ALLOCATE_INFO
);
2891 if (dedicated_info
&& dedicated_info
->image
!= VK_NULL_HANDLE
) {
2892 ANV_FROM_HANDLE(anv_image
, image
, dedicated_info
->image
);
2894 /* Some legacy (non-modifiers) consumers need the tiling to be set on
2895 * the BO. In this case, we have a dedicated allocation.
2897 if (image
->needs_set_tiling
) {
2898 const uint32_t i915_tiling
=
2899 isl_tiling_to_i915_tiling(image
->planes
[0].surface
.isl
.tiling
);
2900 int ret
= anv_gem_set_tiling(device
, mem
->bo
->gem_handle
,
2901 image
->planes
[0].surface
.isl
.row_pitch_B
,
2904 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
2905 return vk_errorf(device
->instance
, NULL
,
2906 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
2907 "failed to set BO tiling: %m");
2913 pthread_mutex_lock(&device
->mutex
);
2914 list_addtail(&mem
->link
, &device
->memory_objects
);
2915 pthread_mutex_unlock(&device
->mutex
);
2917 *pMem
= anv_device_memory_to_handle(mem
);
2919 p_atomic_add(&pdevice
->memory
.heaps
[mem
->type
->heapIndex
].used
,
2925 vk_free2(&device
->alloc
, pAllocator
, mem
);
2930 VkResult
anv_GetMemoryFdKHR(
2932 const VkMemoryGetFdInfoKHR
* pGetFdInfo
,
2935 ANV_FROM_HANDLE(anv_device
, dev
, device_h
);
2936 ANV_FROM_HANDLE(anv_device_memory
, mem
, pGetFdInfo
->memory
);
2938 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
2940 assert(pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
2941 pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
2943 return anv_bo_cache_export(dev
, &dev
->bo_cache
, mem
->bo
, pFd
);
2946 VkResult
anv_GetMemoryFdPropertiesKHR(
2948 VkExternalMemoryHandleTypeFlagBits handleType
,
2950 VkMemoryFdPropertiesKHR
* pMemoryFdProperties
)
2952 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2953 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2955 switch (handleType
) {
2956 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
:
2957 /* dma-buf can be imported as any memory type */
2958 pMemoryFdProperties
->memoryTypeBits
=
2959 (1 << pdevice
->memory
.type_count
) - 1;
2963 /* The valid usage section for this function says:
2965 * "handleType must not be one of the handle types defined as
2968 * So opaque handle types fall into the default "unsupported" case.
2970 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
2974 VkResult
anv_GetMemoryHostPointerPropertiesEXT(
2976 VkExternalMemoryHandleTypeFlagBits handleType
,
2977 const void* pHostPointer
,
2978 VkMemoryHostPointerPropertiesEXT
* pMemoryHostPointerProperties
)
2980 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2982 assert(pMemoryHostPointerProperties
->sType
==
2983 VK_STRUCTURE_TYPE_MEMORY_HOST_POINTER_PROPERTIES_EXT
);
2985 switch (handleType
) {
2986 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
: {
2987 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2989 /* Host memory can be imported as any memory type. */
2990 pMemoryHostPointerProperties
->memoryTypeBits
=
2991 (1ull << pdevice
->memory
.type_count
) - 1;
2996 return VK_ERROR_INVALID_EXTERNAL_HANDLE
;
3000 void anv_FreeMemory(
3002 VkDeviceMemory _mem
,
3003 const VkAllocationCallbacks
* pAllocator
)
3005 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3006 ANV_FROM_HANDLE(anv_device_memory
, mem
, _mem
);
3007 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3012 pthread_mutex_lock(&device
->mutex
);
3013 list_del(&mem
->link
);
3014 pthread_mutex_unlock(&device
->mutex
);
3017 anv_UnmapMemory(_device
, _mem
);
3019 p_atomic_add(&pdevice
->memory
.heaps
[mem
->type
->heapIndex
].used
,
3022 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
3024 #if defined(ANDROID) && ANDROID_API_LEVEL >= 26
3026 AHardwareBuffer_release(mem
->ahw
);
3029 vk_free2(&device
->alloc
, pAllocator
, mem
);
3032 VkResult
anv_MapMemory(
3034 VkDeviceMemory _memory
,
3035 VkDeviceSize offset
,
3037 VkMemoryMapFlags flags
,
3040 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3041 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
3048 if (mem
->host_ptr
) {
3049 *ppData
= mem
->host_ptr
+ offset
;
3053 if (size
== VK_WHOLE_SIZE
)
3054 size
= mem
->bo
->size
- offset
;
3056 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
3058 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
3059 * assert(size != 0);
3060 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
3061 * equal to the size of the memory minus offset
3064 assert(offset
+ size
<= mem
->bo
->size
);
3066 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
3067 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
3068 * at a time is valid. We could just mmap up front and return an offset
3069 * pointer here, but that may exhaust virtual memory on 32 bit
3072 uint32_t gem_flags
= 0;
3074 if (!device
->info
.has_llc
&&
3075 (mem
->type
->propertyFlags
& VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
))
3076 gem_flags
|= I915_MMAP_WC
;
3078 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
3079 uint64_t map_offset
= offset
& ~4095ull;
3080 assert(offset
>= map_offset
);
3081 uint64_t map_size
= (offset
+ size
) - map_offset
;
3083 /* Let's map whole pages */
3084 map_size
= align_u64(map_size
, 4096);
3086 void *map
= anv_gem_mmap(device
, mem
->bo
->gem_handle
,
3087 map_offset
, map_size
, gem_flags
);
3088 if (map
== MAP_FAILED
)
3089 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
3092 mem
->map_size
= map_size
;
3094 *ppData
= mem
->map
+ (offset
- map_offset
);
3099 void anv_UnmapMemory(
3101 VkDeviceMemory _memory
)
3103 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
3105 if (mem
== NULL
|| mem
->host_ptr
)
3108 anv_gem_munmap(mem
->map
, mem
->map_size
);
3115 clflush_mapped_ranges(struct anv_device
*device
,
3117 const VkMappedMemoryRange
*ranges
)
3119 for (uint32_t i
= 0; i
< count
; i
++) {
3120 ANV_FROM_HANDLE(anv_device_memory
, mem
, ranges
[i
].memory
);
3121 if (ranges
[i
].offset
>= mem
->map_size
)
3124 gen_clflush_range(mem
->map
+ ranges
[i
].offset
,
3125 MIN2(ranges
[i
].size
, mem
->map_size
- ranges
[i
].offset
));
3129 VkResult
anv_FlushMappedMemoryRanges(
3131 uint32_t memoryRangeCount
,
3132 const VkMappedMemoryRange
* pMemoryRanges
)
3134 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3136 if (device
->info
.has_llc
)
3139 /* Make sure the writes we're flushing have landed. */
3140 __builtin_ia32_mfence();
3142 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
3147 VkResult
anv_InvalidateMappedMemoryRanges(
3149 uint32_t memoryRangeCount
,
3150 const VkMappedMemoryRange
* pMemoryRanges
)
3152 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3154 if (device
->info
.has_llc
)
3157 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
3159 /* Make sure no reads get moved up above the invalidate. */
3160 __builtin_ia32_mfence();
3165 void anv_GetBufferMemoryRequirements(
3168 VkMemoryRequirements
* pMemoryRequirements
)
3170 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
3171 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3172 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3174 /* The Vulkan spec (git aaed022) says:
3176 * memoryTypeBits is a bitfield and contains one bit set for every
3177 * supported memory type for the resource. The bit `1<<i` is set if and
3178 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3179 * structure for the physical device is supported.
3181 uint32_t memory_types
= 0;
3182 for (uint32_t i
= 0; i
< pdevice
->memory
.type_count
; i
++) {
3183 uint32_t valid_usage
= pdevice
->memory
.types
[i
].valid_buffer_usage
;
3184 if ((valid_usage
& buffer
->usage
) == buffer
->usage
)
3185 memory_types
|= (1u << i
);
3188 /* Base alignment requirement of a cache line */
3189 uint32_t alignment
= 16;
3191 /* We need an alignment of 32 for pushing UBOs */
3192 if (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
)
3193 alignment
= MAX2(alignment
, 32);
3195 pMemoryRequirements
->size
= buffer
->size
;
3196 pMemoryRequirements
->alignment
= alignment
;
3198 /* Storage and Uniform buffers should have their size aligned to
3199 * 32-bits to avoid boundary checks when last DWord is not complete.
3200 * This would ensure that not internal padding would be needed for
3203 if (device
->robust_buffer_access
&&
3204 (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
||
3205 buffer
->usage
& VK_BUFFER_USAGE_STORAGE_BUFFER_BIT
))
3206 pMemoryRequirements
->size
= align_u64(buffer
->size
, 4);
3208 pMemoryRequirements
->memoryTypeBits
= memory_types
;
3211 void anv_GetBufferMemoryRequirements2(
3213 const VkBufferMemoryRequirementsInfo2
* pInfo
,
3214 VkMemoryRequirements2
* pMemoryRequirements
)
3216 anv_GetBufferMemoryRequirements(_device
, pInfo
->buffer
,
3217 &pMemoryRequirements
->memoryRequirements
);
3219 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3220 switch (ext
->sType
) {
3221 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
3222 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
3223 requirements
->prefersDedicatedAllocation
= false;
3224 requirements
->requiresDedicatedAllocation
= false;
3229 anv_debug_ignored_stype(ext
->sType
);
3235 void anv_GetImageMemoryRequirements(
3238 VkMemoryRequirements
* pMemoryRequirements
)
3240 ANV_FROM_HANDLE(anv_image
, image
, _image
);
3241 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3242 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3244 /* The Vulkan spec (git aaed022) says:
3246 * memoryTypeBits is a bitfield and contains one bit set for every
3247 * supported memory type for the resource. The bit `1<<i` is set if and
3248 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3249 * structure for the physical device is supported.
3251 * All types are currently supported for images.
3253 uint32_t memory_types
= (1ull << pdevice
->memory
.type_count
) - 1;
3255 /* We must have image allocated or imported at this point. According to the
3256 * specification, external images must have been bound to memory before
3257 * calling GetImageMemoryRequirements.
3259 assert(image
->size
> 0);
3261 pMemoryRequirements
->size
= image
->size
;
3262 pMemoryRequirements
->alignment
= image
->alignment
;
3263 pMemoryRequirements
->memoryTypeBits
= memory_types
;
3266 void anv_GetImageMemoryRequirements2(
3268 const VkImageMemoryRequirementsInfo2
* pInfo
,
3269 VkMemoryRequirements2
* pMemoryRequirements
)
3271 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3272 ANV_FROM_HANDLE(anv_image
, image
, pInfo
->image
);
3274 anv_GetImageMemoryRequirements(_device
, pInfo
->image
,
3275 &pMemoryRequirements
->memoryRequirements
);
3277 vk_foreach_struct_const(ext
, pInfo
->pNext
) {
3278 switch (ext
->sType
) {
3279 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO
: {
3280 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3281 const VkImagePlaneMemoryRequirementsInfo
*plane_reqs
=
3282 (const VkImagePlaneMemoryRequirementsInfo
*) ext
;
3283 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
,
3284 plane_reqs
->planeAspect
);
3286 assert(image
->planes
[plane
].offset
== 0);
3288 /* The Vulkan spec (git aaed022) says:
3290 * memoryTypeBits is a bitfield and contains one bit set for every
3291 * supported memory type for the resource. The bit `1<<i` is set
3292 * if and only if the memory type `i` in the
3293 * VkPhysicalDeviceMemoryProperties structure for the physical
3294 * device is supported.
3296 * All types are currently supported for images.
3298 pMemoryRequirements
->memoryRequirements
.memoryTypeBits
=
3299 (1ull << pdevice
->memory
.type_count
) - 1;
3301 /* We must have image allocated or imported at this point. According to the
3302 * specification, external images must have been bound to memory before
3303 * calling GetImageMemoryRequirements.
3305 assert(image
->planes
[plane
].size
> 0);
3307 pMemoryRequirements
->memoryRequirements
.size
= image
->planes
[plane
].size
;
3308 pMemoryRequirements
->memoryRequirements
.alignment
=
3309 image
->planes
[plane
].alignment
;
3314 anv_debug_ignored_stype(ext
->sType
);
3319 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3320 switch (ext
->sType
) {
3321 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
3322 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
3323 if (image
->needs_set_tiling
|| image
->external_format
) {
3324 /* If we need to set the tiling for external consumers, we need a
3325 * dedicated allocation.
3327 * See also anv_AllocateMemory.
3329 requirements
->prefersDedicatedAllocation
= true;
3330 requirements
->requiresDedicatedAllocation
= true;
3332 requirements
->prefersDedicatedAllocation
= false;
3333 requirements
->requiresDedicatedAllocation
= false;
3339 anv_debug_ignored_stype(ext
->sType
);
3345 void anv_GetImageSparseMemoryRequirements(
3348 uint32_t* pSparseMemoryRequirementCount
,
3349 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
3351 *pSparseMemoryRequirementCount
= 0;
3354 void anv_GetImageSparseMemoryRequirements2(
3356 const VkImageSparseMemoryRequirementsInfo2
* pInfo
,
3357 uint32_t* pSparseMemoryRequirementCount
,
3358 VkSparseImageMemoryRequirements2
* pSparseMemoryRequirements
)
3360 *pSparseMemoryRequirementCount
= 0;
3363 void anv_GetDeviceMemoryCommitment(
3365 VkDeviceMemory memory
,
3366 VkDeviceSize
* pCommittedMemoryInBytes
)
3368 *pCommittedMemoryInBytes
= 0;
3372 anv_bind_buffer_memory(const VkBindBufferMemoryInfo
*pBindInfo
)
3374 ANV_FROM_HANDLE(anv_device_memory
, mem
, pBindInfo
->memory
);
3375 ANV_FROM_HANDLE(anv_buffer
, buffer
, pBindInfo
->buffer
);
3377 assert(pBindInfo
->sType
== VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
);
3380 assert((buffer
->usage
& mem
->type
->valid_buffer_usage
) == buffer
->usage
);
3381 buffer
->address
= (struct anv_address
) {
3383 .offset
= pBindInfo
->memoryOffset
,
3386 buffer
->address
= ANV_NULL_ADDRESS
;
3390 VkResult
anv_BindBufferMemory(
3393 VkDeviceMemory memory
,
3394 VkDeviceSize memoryOffset
)
3396 anv_bind_buffer_memory(
3397 &(VkBindBufferMemoryInfo
) {
3398 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
,
3401 .memoryOffset
= memoryOffset
,
3407 VkResult
anv_BindBufferMemory2(
3409 uint32_t bindInfoCount
,
3410 const VkBindBufferMemoryInfo
* pBindInfos
)
3412 for (uint32_t i
= 0; i
< bindInfoCount
; i
++)
3413 anv_bind_buffer_memory(&pBindInfos
[i
]);
3418 VkResult
anv_QueueBindSparse(
3420 uint32_t bindInfoCount
,
3421 const VkBindSparseInfo
* pBindInfo
,
3424 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
3425 if (anv_device_is_lost(queue
->device
))
3426 return VK_ERROR_DEVICE_LOST
;
3428 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
3433 VkResult
anv_CreateEvent(
3435 const VkEventCreateInfo
* pCreateInfo
,
3436 const VkAllocationCallbacks
* pAllocator
,
3439 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3440 struct anv_state state
;
3441 struct anv_event
*event
;
3443 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_EVENT_CREATE_INFO
);
3445 state
= anv_state_pool_alloc(&device
->dynamic_state_pool
,
3448 event
->state
= state
;
3449 event
->semaphore
= VK_EVENT_RESET
;
3451 if (!device
->info
.has_llc
) {
3452 /* Make sure the writes we're flushing have landed. */
3453 __builtin_ia32_mfence();
3454 __builtin_ia32_clflush(event
);
3457 *pEvent
= anv_event_to_handle(event
);
3462 void anv_DestroyEvent(
3465 const VkAllocationCallbacks
* pAllocator
)
3467 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3468 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3473 anv_state_pool_free(&device
->dynamic_state_pool
, event
->state
);
3476 VkResult
anv_GetEventStatus(
3480 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3481 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3483 if (anv_device_is_lost(device
))
3484 return VK_ERROR_DEVICE_LOST
;
3486 if (!device
->info
.has_llc
) {
3487 /* Invalidate read cache before reading event written by GPU. */
3488 __builtin_ia32_clflush(event
);
3489 __builtin_ia32_mfence();
3493 return event
->semaphore
;
3496 VkResult
anv_SetEvent(
3500 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3501 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3503 event
->semaphore
= VK_EVENT_SET
;
3505 if (!device
->info
.has_llc
) {
3506 /* Make sure the writes we're flushing have landed. */
3507 __builtin_ia32_mfence();
3508 __builtin_ia32_clflush(event
);
3514 VkResult
anv_ResetEvent(
3518 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3519 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3521 event
->semaphore
= VK_EVENT_RESET
;
3523 if (!device
->info
.has_llc
) {
3524 /* Make sure the writes we're flushing have landed. */
3525 __builtin_ia32_mfence();
3526 __builtin_ia32_clflush(event
);
3534 VkResult
anv_CreateBuffer(
3536 const VkBufferCreateInfo
* pCreateInfo
,
3537 const VkAllocationCallbacks
* pAllocator
,
3540 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3541 struct anv_buffer
*buffer
;
3543 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
3545 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
3546 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3548 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3550 buffer
->size
= pCreateInfo
->size
;
3551 buffer
->usage
= pCreateInfo
->usage
;
3552 buffer
->address
= ANV_NULL_ADDRESS
;
3554 *pBuffer
= anv_buffer_to_handle(buffer
);
3559 void anv_DestroyBuffer(
3562 const VkAllocationCallbacks
* pAllocator
)
3564 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3565 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
3570 vk_free2(&device
->alloc
, pAllocator
, buffer
);
3573 VkDeviceAddress
anv_GetBufferDeviceAddressEXT(
3575 const VkBufferDeviceAddressInfoEXT
* pInfo
)
3577 ANV_FROM_HANDLE(anv_buffer
, buffer
, pInfo
->buffer
);
3579 assert(buffer
->address
.bo
->flags
& EXEC_OBJECT_PINNED
);
3581 return anv_address_physical(buffer
->address
);
3585 anv_fill_buffer_surface_state(struct anv_device
*device
, struct anv_state state
,
3586 enum isl_format format
,
3587 struct anv_address address
,
3588 uint32_t range
, uint32_t stride
)
3590 isl_buffer_fill_state(&device
->isl_dev
, state
.map
,
3591 .address
= anv_address_physical(address
),
3592 .mocs
= device
->default_mocs
,
3595 .swizzle
= ISL_SWIZZLE_IDENTITY
,
3596 .stride_B
= stride
);
3599 void anv_DestroySampler(
3602 const VkAllocationCallbacks
* pAllocator
)
3604 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3605 ANV_FROM_HANDLE(anv_sampler
, sampler
, _sampler
);
3610 if (sampler
->bindless_state
.map
) {
3611 anv_state_pool_free(&device
->dynamic_state_pool
,
3612 sampler
->bindless_state
);
3615 vk_free2(&device
->alloc
, pAllocator
, sampler
);
3618 VkResult
anv_CreateFramebuffer(
3620 const VkFramebufferCreateInfo
* pCreateInfo
,
3621 const VkAllocationCallbacks
* pAllocator
,
3622 VkFramebuffer
* pFramebuffer
)
3624 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3625 struct anv_framebuffer
*framebuffer
;
3627 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
3629 size_t size
= sizeof(*framebuffer
) +
3630 sizeof(struct anv_image_view
*) * pCreateInfo
->attachmentCount
;
3631 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
3632 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3633 if (framebuffer
== NULL
)
3634 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3636 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
3637 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
3638 VkImageView _iview
= pCreateInfo
->pAttachments
[i
];
3639 framebuffer
->attachments
[i
] = anv_image_view_from_handle(_iview
);
3642 framebuffer
->width
= pCreateInfo
->width
;
3643 framebuffer
->height
= pCreateInfo
->height
;
3644 framebuffer
->layers
= pCreateInfo
->layers
;
3646 *pFramebuffer
= anv_framebuffer_to_handle(framebuffer
);
3651 void anv_DestroyFramebuffer(
3654 const VkAllocationCallbacks
* pAllocator
)
3656 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3657 ANV_FROM_HANDLE(anv_framebuffer
, fb
, _fb
);
3662 vk_free2(&device
->alloc
, pAllocator
, fb
);
3665 static const VkTimeDomainEXT anv_time_domains
[] = {
3666 VK_TIME_DOMAIN_DEVICE_EXT
,
3667 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
,
3668 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
,
3671 VkResult
anv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
3672 VkPhysicalDevice physicalDevice
,
3673 uint32_t *pTimeDomainCount
,
3674 VkTimeDomainEXT
*pTimeDomains
)
3677 VK_OUTARRAY_MAKE(out
, pTimeDomains
, pTimeDomainCount
);
3679 for (d
= 0; d
< ARRAY_SIZE(anv_time_domains
); d
++) {
3680 vk_outarray_append(&out
, i
) {
3681 *i
= anv_time_domains
[d
];
3685 return vk_outarray_status(&out
);
3689 anv_clock_gettime(clockid_t clock_id
)
3691 struct timespec current
;
3694 ret
= clock_gettime(clock_id
, ¤t
);
3695 if (ret
< 0 && clock_id
== CLOCK_MONOTONIC_RAW
)
3696 ret
= clock_gettime(CLOCK_MONOTONIC
, ¤t
);
3700 return (uint64_t) current
.tv_sec
* 1000000000ULL + current
.tv_nsec
;
3703 #define TIMESTAMP 0x2358
3705 VkResult
anv_GetCalibratedTimestampsEXT(
3707 uint32_t timestampCount
,
3708 const VkCalibratedTimestampInfoEXT
*pTimestampInfos
,
3709 uint64_t *pTimestamps
,
3710 uint64_t *pMaxDeviation
)
3712 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3713 uint64_t timestamp_frequency
= device
->info
.timestamp_frequency
;
3716 uint64_t begin
, end
;
3717 uint64_t max_clock_period
= 0;
3719 begin
= anv_clock_gettime(CLOCK_MONOTONIC_RAW
);
3721 for (d
= 0; d
< timestampCount
; d
++) {
3722 switch (pTimestampInfos
[d
].timeDomain
) {
3723 case VK_TIME_DOMAIN_DEVICE_EXT
:
3724 ret
= anv_gem_reg_read(device
, TIMESTAMP
| 1,
3728 return anv_device_set_lost(device
, "Failed to read the TIMESTAMP "
3731 uint64_t device_period
= DIV_ROUND_UP(1000000000, timestamp_frequency
);
3732 max_clock_period
= MAX2(max_clock_period
, device_period
);
3734 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
:
3735 pTimestamps
[d
] = anv_clock_gettime(CLOCK_MONOTONIC
);
3736 max_clock_period
= MAX2(max_clock_period
, 1);
3739 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
:
3740 pTimestamps
[d
] = begin
;
3748 end
= anv_clock_gettime(CLOCK_MONOTONIC_RAW
);
3751 * The maximum deviation is the sum of the interval over which we
3752 * perform the sampling and the maximum period of any sampled
3753 * clock. That's because the maximum skew between any two sampled
3754 * clock edges is when the sampled clock with the largest period is
3755 * sampled at the end of that period but right at the beginning of the
3756 * sampling interval and some other clock is sampled right at the
3757 * begining of its sampling period and right at the end of the
3758 * sampling interval. Let's assume the GPU has the longest clock
3759 * period and that the application is sampling GPU and monotonic:
3762 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
3763 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
3767 * GPU -----_____-----_____-----_____-----_____
3770 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
3771 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
3773 * Interval <----------------->
3774 * Deviation <-------------------------->
3778 * m = read(monotonic) 2
3781 * We round the sample interval up by one tick to cover sampling error
3782 * in the interval clock
3785 uint64_t sample_interval
= end
- begin
+ 1;
3787 *pMaxDeviation
= sample_interval
+ max_clock_period
;
3792 /* vk_icd.h does not declare this function, so we declare it here to
3793 * suppress Wmissing-prototypes.
3795 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
3796 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
);
3798 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
3799 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
)
3801 /* For the full details on loader interface versioning, see
3802 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
3803 * What follows is a condensed summary, to help you navigate the large and
3804 * confusing official doc.
3806 * - Loader interface v0 is incompatible with later versions. We don't
3809 * - In loader interface v1:
3810 * - The first ICD entrypoint called by the loader is
3811 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
3813 * - The ICD must statically expose no other Vulkan symbol unless it is
3814 * linked with -Bsymbolic.
3815 * - Each dispatchable Vulkan handle created by the ICD must be
3816 * a pointer to a struct whose first member is VK_LOADER_DATA. The
3817 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
3818 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
3819 * vkDestroySurfaceKHR(). The ICD must be capable of working with
3820 * such loader-managed surfaces.
3822 * - Loader interface v2 differs from v1 in:
3823 * - The first ICD entrypoint called by the loader is
3824 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
3825 * statically expose this entrypoint.
3827 * - Loader interface v3 differs from v2 in:
3828 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
3829 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
3830 * because the loader no longer does so.
3832 *pSupportedVersion
= MIN2(*pSupportedVersion
, 3u);