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
<= 10) {
425 /* Gen8-10 fully supported */
426 } else if (device
->info
.gen
== 11) {
427 intel_logw("Vulkan is not yet fully supported on gen11.");
429 result
= vk_errorf(device
->instance
, device
,
430 VK_ERROR_INCOMPATIBLE_DRIVER
,
431 "Vulkan not yet supported on %s", device
->name
);
435 device
->cmd_parser_version
= -1;
436 if (device
->info
.gen
== 7) {
437 device
->cmd_parser_version
=
438 anv_gem_get_param(fd
, I915_PARAM_CMD_PARSER_VERSION
);
439 if (device
->cmd_parser_version
== -1) {
440 result
= vk_errorf(device
->instance
, device
,
441 VK_ERROR_INITIALIZATION_FAILED
,
442 "failed to get command parser version");
447 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_WAIT_TIMEOUT
)) {
448 result
= vk_errorf(device
->instance
, device
,
449 VK_ERROR_INITIALIZATION_FAILED
,
450 "kernel missing gem wait");
454 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_EXECBUF2
)) {
455 result
= vk_errorf(device
->instance
, device
,
456 VK_ERROR_INITIALIZATION_FAILED
,
457 "kernel missing execbuf2");
461 if (!device
->info
.has_llc
&&
462 anv_gem_get_param(fd
, I915_PARAM_MMAP_VERSION
) < 1) {
463 result
= vk_errorf(device
->instance
, device
,
464 VK_ERROR_INITIALIZATION_FAILED
,
465 "kernel missing wc mmap");
469 result
= anv_physical_device_init_heaps(device
, fd
);
470 if (result
!= VK_SUCCESS
)
473 device
->has_exec_async
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_ASYNC
);
474 device
->has_exec_capture
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_CAPTURE
);
475 device
->has_exec_fence
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE
);
476 device
->has_syncobj
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE_ARRAY
);
477 device
->has_syncobj_wait
= device
->has_syncobj
&&
478 anv_gem_supports_syncobj_wait(fd
);
479 device
->has_context_priority
= anv_gem_has_context_priority(fd
);
481 device
->use_softpin
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_SOFTPIN
)
482 && device
->supports_48bit_addresses
;
484 device
->has_context_isolation
=
485 anv_gem_get_param(fd
, I915_PARAM_HAS_CONTEXT_ISOLATION
);
487 device
->always_use_bindless
=
488 env_var_as_boolean("ANV_ALWAYS_BINDLESS", false);
490 /* We first got the A64 messages on broadwell and we can only use them if
491 * we can pass addresses directly into the shader which requires softpin.
493 device
->has_a64_buffer_access
= device
->info
.gen
>= 8 &&
496 /* We first get bindless image access on Skylake and we can only really do
497 * it if we don't have any relocations so we need softpin.
499 device
->has_bindless_images
= device
->info
.gen
>= 9 &&
502 /* We've had bindless samplers since Ivy Bridge (forever in Vulkan terms)
503 * because it's just a matter of setting the sampler address in the sample
504 * message header. However, we've not bothered to wire it up for vec4 so
505 * we leave it disabled on gen7.
507 device
->has_bindless_samplers
= device
->info
.gen
>= 8;
509 device
->has_mem_available
= get_available_system_memory() != 0;
511 /* Starting with Gen10, the timestamp frequency of the command streamer may
512 * vary from one part to another. We can query the value from the kernel.
514 if (device
->info
.gen
>= 10) {
515 int timestamp_frequency
=
516 anv_gem_get_param(fd
, I915_PARAM_CS_TIMESTAMP_FREQUENCY
);
518 if (timestamp_frequency
< 0)
519 intel_logw("Kernel 4.16-rc1+ required to properly query CS timestamp frequency");
521 device
->info
.timestamp_frequency
= timestamp_frequency
;
524 /* GENs prior to 8 do not support EU/Subslice info */
525 if (device
->info
.gen
>= 8) {
526 device
->subslice_total
= anv_gem_get_param(fd
, I915_PARAM_SUBSLICE_TOTAL
);
527 device
->eu_total
= anv_gem_get_param(fd
, I915_PARAM_EU_TOTAL
);
529 /* Without this information, we cannot get the right Braswell
530 * brandstrings, and we have to use conservative numbers for GPGPU on
531 * many platforms, but otherwise, things will just work.
533 if (device
->subslice_total
< 1 || device
->eu_total
< 1) {
534 intel_logw("Kernel 4.1 required to properly query GPU properties");
536 } else if (device
->info
.gen
== 7) {
537 device
->subslice_total
= 1 << (device
->info
.gt
- 1);
540 if (device
->info
.is_cherryview
&&
541 device
->subslice_total
> 0 && device
->eu_total
> 0) {
542 /* Logical CS threads = EUs per subslice * num threads per EU */
543 uint32_t max_cs_threads
=
544 device
->eu_total
/ device
->subslice_total
* device
->info
.num_thread_per_eu
;
546 /* Fuse configurations may give more threads than expected, never less. */
547 if (max_cs_threads
> device
->info
.max_cs_threads
)
548 device
->info
.max_cs_threads
= max_cs_threads
;
551 device
->compiler
= brw_compiler_create(NULL
, &device
->info
);
552 if (device
->compiler
== NULL
) {
553 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
556 device
->compiler
->shader_debug_log
= compiler_debug_log
;
557 device
->compiler
->shader_perf_log
= compiler_perf_log
;
558 device
->compiler
->supports_pull_constants
= false;
559 device
->compiler
->constant_buffer_0_is_relative
=
560 device
->info
.gen
< 8 || !device
->has_context_isolation
;
561 device
->compiler
->supports_shader_constants
= true;
563 /* Broadwell PRM says:
565 * "Before Gen8, there was a historical configuration control field to
566 * swizzle address bit[6] for in X/Y tiling modes. This was set in three
567 * different places: TILECTL[1:0], ARB_MODE[5:4], and
568 * DISP_ARB_CTL[14:13].
570 * For Gen8 and subsequent generations, the swizzle fields are all
571 * reserved, and the CPU's memory controller performs all address
572 * swizzling modifications."
575 device
->info
.gen
< 8 && anv_gem_get_bit6_swizzle(fd
, I915_TILING_X
);
577 isl_device_init(&device
->isl_dev
, &device
->info
, swizzled
);
579 result
= anv_physical_device_init_uuids(device
);
580 if (result
!= VK_SUCCESS
)
583 anv_physical_device_init_disk_cache(device
);
585 if (instance
->enabled_extensions
.KHR_display
) {
586 master_fd
= open(primary_path
, O_RDWR
| O_CLOEXEC
);
587 if (master_fd
>= 0) {
588 /* prod the device with a GETPARAM call which will fail if
589 * we don't have permission to even render on this device
591 if (anv_gem_get_param(master_fd
, I915_PARAM_CHIPSET_ID
) == 0) {
597 device
->master_fd
= master_fd
;
599 result
= anv_init_wsi(device
);
600 if (result
!= VK_SUCCESS
) {
601 ralloc_free(device
->compiler
);
602 anv_physical_device_free_disk_cache(device
);
606 anv_physical_device_get_supported_extensions(device
,
607 &device
->supported_extensions
);
610 device
->local_fd
= fd
;
622 anv_physical_device_finish(struct anv_physical_device
*device
)
624 anv_finish_wsi(device
);
625 anv_physical_device_free_disk_cache(device
);
626 ralloc_free(device
->compiler
);
627 close(device
->local_fd
);
628 if (device
->master_fd
>= 0)
629 close(device
->master_fd
);
633 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
634 VkSystemAllocationScope allocationScope
)
640 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
641 size_t align
, VkSystemAllocationScope allocationScope
)
643 return realloc(pOriginal
, size
);
647 default_free_func(void *pUserData
, void *pMemory
)
652 static const VkAllocationCallbacks default_alloc
= {
654 .pfnAllocation
= default_alloc_func
,
655 .pfnReallocation
= default_realloc_func
,
656 .pfnFree
= default_free_func
,
659 VkResult
anv_EnumerateInstanceExtensionProperties(
660 const char* pLayerName
,
661 uint32_t* pPropertyCount
,
662 VkExtensionProperties
* pProperties
)
664 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
666 for (int i
= 0; i
< ANV_INSTANCE_EXTENSION_COUNT
; i
++) {
667 if (anv_instance_extensions_supported
.extensions
[i
]) {
668 vk_outarray_append(&out
, prop
) {
669 *prop
= anv_instance_extensions
[i
];
674 return vk_outarray_status(&out
);
677 VkResult
anv_CreateInstance(
678 const VkInstanceCreateInfo
* pCreateInfo
,
679 const VkAllocationCallbacks
* pAllocator
,
680 VkInstance
* pInstance
)
682 struct anv_instance
*instance
;
685 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
687 struct anv_instance_extension_table enabled_extensions
= {};
688 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
690 for (idx
= 0; idx
< ANV_INSTANCE_EXTENSION_COUNT
; idx
++) {
691 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
692 anv_instance_extensions
[idx
].extensionName
) == 0)
696 if (idx
>= ANV_INSTANCE_EXTENSION_COUNT
)
697 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
699 if (!anv_instance_extensions_supported
.extensions
[idx
])
700 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
702 enabled_extensions
.extensions
[idx
] = true;
705 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
706 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
708 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
710 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
713 instance
->alloc
= *pAllocator
;
715 instance
->alloc
= default_alloc
;
717 instance
->app_info
= (struct anv_app_info
) { .api_version
= 0 };
718 if (pCreateInfo
->pApplicationInfo
) {
719 const VkApplicationInfo
*app
= pCreateInfo
->pApplicationInfo
;
721 instance
->app_info
.app_name
=
722 vk_strdup(&instance
->alloc
, app
->pApplicationName
,
723 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
724 instance
->app_info
.app_version
= app
->applicationVersion
;
726 instance
->app_info
.engine_name
=
727 vk_strdup(&instance
->alloc
, app
->pEngineName
,
728 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
729 instance
->app_info
.engine_version
= app
->engineVersion
;
731 instance
->app_info
.api_version
= app
->apiVersion
;
734 if (instance
->app_info
.api_version
== 0)
735 instance
->app_info
.api_version
= VK_API_VERSION_1_0
;
737 instance
->enabled_extensions
= enabled_extensions
;
739 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->dispatch
.entrypoints
); i
++) {
740 /* Vulkan requires that entrypoints for extensions which have not been
741 * enabled must not be advertised.
743 if (!anv_instance_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
744 &instance
->enabled_extensions
)) {
745 instance
->dispatch
.entrypoints
[i
] = NULL
;
747 instance
->dispatch
.entrypoints
[i
] =
748 anv_instance_dispatch_table
.entrypoints
[i
];
752 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->device_dispatch
.entrypoints
); i
++) {
753 /* Vulkan requires that entrypoints for extensions which have not been
754 * enabled must not be advertised.
756 if (!anv_device_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
757 &instance
->enabled_extensions
, NULL
)) {
758 instance
->device_dispatch
.entrypoints
[i
] = NULL
;
760 instance
->device_dispatch
.entrypoints
[i
] =
761 anv_device_dispatch_table
.entrypoints
[i
];
765 instance
->physicalDeviceCount
= -1;
767 result
= vk_debug_report_instance_init(&instance
->debug_report_callbacks
);
768 if (result
!= VK_SUCCESS
) {
769 vk_free2(&default_alloc
, pAllocator
, instance
);
770 return vk_error(result
);
773 instance
->pipeline_cache_enabled
=
774 env_var_as_boolean("ANV_ENABLE_PIPELINE_CACHE", true);
777 glsl_type_singleton_init_or_ref();
779 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
781 *pInstance
= anv_instance_to_handle(instance
);
786 void anv_DestroyInstance(
787 VkInstance _instance
,
788 const VkAllocationCallbacks
* pAllocator
)
790 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
795 if (instance
->physicalDeviceCount
> 0) {
796 /* We support at most one physical device. */
797 assert(instance
->physicalDeviceCount
== 1);
798 anv_physical_device_finish(&instance
->physicalDevice
);
801 vk_free(&instance
->alloc
, (char *)instance
->app_info
.app_name
);
802 vk_free(&instance
->alloc
, (char *)instance
->app_info
.engine_name
);
804 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
806 vk_debug_report_instance_destroy(&instance
->debug_report_callbacks
);
808 glsl_type_singleton_decref();
811 vk_free(&instance
->alloc
, instance
);
815 anv_enumerate_devices(struct anv_instance
*instance
)
817 /* TODO: Check for more devices ? */
818 drmDevicePtr devices
[8];
819 VkResult result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
822 instance
->physicalDeviceCount
= 0;
824 max_devices
= drmGetDevices2(0, devices
, ARRAY_SIZE(devices
));
826 return VK_ERROR_INCOMPATIBLE_DRIVER
;
828 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
829 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
830 devices
[i
]->bustype
== DRM_BUS_PCI
&&
831 devices
[i
]->deviceinfo
.pci
->vendor_id
== 0x8086) {
833 result
= anv_physical_device_init(&instance
->physicalDevice
,
834 instance
, devices
[i
]);
835 if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
839 drmFreeDevices(devices
, max_devices
);
841 if (result
== VK_SUCCESS
)
842 instance
->physicalDeviceCount
= 1;
848 anv_instance_ensure_physical_device(struct anv_instance
*instance
)
850 if (instance
->physicalDeviceCount
< 0) {
851 VkResult result
= anv_enumerate_devices(instance
);
852 if (result
!= VK_SUCCESS
&&
853 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
860 VkResult
anv_EnumeratePhysicalDevices(
861 VkInstance _instance
,
862 uint32_t* pPhysicalDeviceCount
,
863 VkPhysicalDevice
* pPhysicalDevices
)
865 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
866 VK_OUTARRAY_MAKE(out
, pPhysicalDevices
, pPhysicalDeviceCount
);
868 VkResult result
= anv_instance_ensure_physical_device(instance
);
869 if (result
!= VK_SUCCESS
)
872 if (instance
->physicalDeviceCount
== 0)
875 assert(instance
->physicalDeviceCount
== 1);
876 vk_outarray_append(&out
, i
) {
877 *i
= anv_physical_device_to_handle(&instance
->physicalDevice
);
880 return vk_outarray_status(&out
);
883 VkResult
anv_EnumeratePhysicalDeviceGroups(
884 VkInstance _instance
,
885 uint32_t* pPhysicalDeviceGroupCount
,
886 VkPhysicalDeviceGroupProperties
* pPhysicalDeviceGroupProperties
)
888 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
889 VK_OUTARRAY_MAKE(out
, pPhysicalDeviceGroupProperties
,
890 pPhysicalDeviceGroupCount
);
892 VkResult result
= anv_instance_ensure_physical_device(instance
);
893 if (result
!= VK_SUCCESS
)
896 if (instance
->physicalDeviceCount
== 0)
899 assert(instance
->physicalDeviceCount
== 1);
901 vk_outarray_append(&out
, p
) {
902 p
->physicalDeviceCount
= 1;
903 memset(p
->physicalDevices
, 0, sizeof(p
->physicalDevices
));
904 p
->physicalDevices
[0] =
905 anv_physical_device_to_handle(&instance
->physicalDevice
);
906 p
->subsetAllocation
= false;
908 vk_foreach_struct(ext
, p
->pNext
)
909 anv_debug_ignored_stype(ext
->sType
);
912 return vk_outarray_status(&out
);
915 void anv_GetPhysicalDeviceFeatures(
916 VkPhysicalDevice physicalDevice
,
917 VkPhysicalDeviceFeatures
* pFeatures
)
919 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
921 *pFeatures
= (VkPhysicalDeviceFeatures
) {
922 .robustBufferAccess
= true,
923 .fullDrawIndexUint32
= true,
924 .imageCubeArray
= true,
925 .independentBlend
= true,
926 .geometryShader
= true,
927 .tessellationShader
= true,
928 .sampleRateShading
= true,
929 .dualSrcBlend
= true,
931 .multiDrawIndirect
= true,
932 .drawIndirectFirstInstance
= true,
934 .depthBiasClamp
= true,
935 .fillModeNonSolid
= true,
936 .depthBounds
= false,
940 .multiViewport
= true,
941 .samplerAnisotropy
= true,
942 .textureCompressionETC2
= pdevice
->info
.gen
>= 8 ||
943 pdevice
->info
.is_baytrail
,
944 .textureCompressionASTC_LDR
= pdevice
->info
.gen
>= 9, /* FINISHME CHV */
945 .textureCompressionBC
= true,
946 .occlusionQueryPrecise
= true,
947 .pipelineStatisticsQuery
= true,
948 .fragmentStoresAndAtomics
= true,
949 .shaderTessellationAndGeometryPointSize
= true,
950 .shaderImageGatherExtended
= true,
951 .shaderStorageImageExtendedFormats
= true,
952 .shaderStorageImageMultisample
= false,
953 .shaderStorageImageReadWithoutFormat
= false,
954 .shaderStorageImageWriteWithoutFormat
= true,
955 .shaderUniformBufferArrayDynamicIndexing
= true,
956 .shaderSampledImageArrayDynamicIndexing
= true,
957 .shaderStorageBufferArrayDynamicIndexing
= true,
958 .shaderStorageImageArrayDynamicIndexing
= true,
959 .shaderClipDistance
= true,
960 .shaderCullDistance
= true,
961 .shaderFloat64
= pdevice
->info
.gen
>= 8 &&
962 pdevice
->info
.has_64bit_types
,
963 .shaderInt64
= pdevice
->info
.gen
>= 8 &&
964 pdevice
->info
.has_64bit_types
,
965 .shaderInt16
= pdevice
->info
.gen
>= 8,
966 .shaderResourceMinLod
= pdevice
->info
.gen
>= 9,
967 .variableMultisampleRate
= true,
968 .inheritedQueries
= true,
971 /* We can't do image stores in vec4 shaders */
972 pFeatures
->vertexPipelineStoresAndAtomics
=
973 pdevice
->compiler
->scalar_stage
[MESA_SHADER_VERTEX
] &&
974 pdevice
->compiler
->scalar_stage
[MESA_SHADER_GEOMETRY
];
976 struct anv_app_info
*app_info
= &pdevice
->instance
->app_info
;
978 /* The new DOOM and Wolfenstein games require depthBounds without
979 * checking for it. They seem to run fine without it so just claim it's
980 * there and accept the consequences.
982 if (app_info
->engine_name
&& strcmp(app_info
->engine_name
, "idTech") == 0)
983 pFeatures
->depthBounds
= true;
986 void anv_GetPhysicalDeviceFeatures2(
987 VkPhysicalDevice physicalDevice
,
988 VkPhysicalDeviceFeatures2
* pFeatures
)
990 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
991 anv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
993 vk_foreach_struct(ext
, pFeatures
->pNext
) {
994 switch (ext
->sType
) {
995 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES_KHR
: {
996 VkPhysicalDevice8BitStorageFeaturesKHR
*features
=
997 (VkPhysicalDevice8BitStorageFeaturesKHR
*)ext
;
998 features
->storageBuffer8BitAccess
= pdevice
->info
.gen
>= 8;
999 features
->uniformAndStorageBuffer8BitAccess
= pdevice
->info
.gen
>= 8;
1000 features
->storagePushConstant8
= pdevice
->info
.gen
>= 8;
1004 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES
: {
1005 VkPhysicalDevice16BitStorageFeatures
*features
=
1006 (VkPhysicalDevice16BitStorageFeatures
*)ext
;
1007 features
->storageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
1008 features
->uniformAndStorageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
1009 features
->storagePushConstant16
= pdevice
->info
.gen
>= 8;
1010 features
->storageInputOutput16
= false;
1014 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_EXT
: {
1015 VkPhysicalDeviceBufferDeviceAddressFeaturesEXT
*features
= (void *)ext
;
1016 features
->bufferDeviceAddress
= pdevice
->has_a64_buffer_access
;
1017 features
->bufferDeviceAddressCaptureReplay
= false;
1018 features
->bufferDeviceAddressMultiDevice
= false;
1022 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COMPUTE_SHADER_DERIVATIVES_FEATURES_NV
: {
1023 VkPhysicalDeviceComputeShaderDerivativesFeaturesNV
*features
=
1024 (VkPhysicalDeviceComputeShaderDerivativesFeaturesNV
*)ext
;
1025 features
->computeDerivativeGroupQuads
= true;
1026 features
->computeDerivativeGroupLinear
= true;
1030 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT
: {
1031 VkPhysicalDeviceConditionalRenderingFeaturesEXT
*features
=
1032 (VkPhysicalDeviceConditionalRenderingFeaturesEXT
*)ext
;
1033 features
->conditionalRendering
= pdevice
->info
.gen
>= 8 ||
1034 pdevice
->info
.is_haswell
;
1035 features
->inheritedConditionalRendering
= pdevice
->info
.gen
>= 8 ||
1036 pdevice
->info
.is_haswell
;
1040 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_CLIP_ENABLE_FEATURES_EXT
: {
1041 VkPhysicalDeviceDepthClipEnableFeaturesEXT
*features
=
1042 (VkPhysicalDeviceDepthClipEnableFeaturesEXT
*)ext
;
1043 features
->depthClipEnable
= true;
1047 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR
: {
1048 VkPhysicalDeviceFloat16Int8FeaturesKHR
*features
= (void *)ext
;
1049 features
->shaderFloat16
= pdevice
->info
.gen
>= 8;
1050 features
->shaderInt8
= pdevice
->info
.gen
>= 8;
1054 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_QUERY_RESET_FEATURES_EXT
: {
1055 VkPhysicalDeviceHostQueryResetFeaturesEXT
*features
=
1056 (VkPhysicalDeviceHostQueryResetFeaturesEXT
*)ext
;
1057 features
->hostQueryReset
= true;
1061 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT
: {
1062 VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*features
=
1063 (VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*)ext
;
1064 features
->shaderInputAttachmentArrayDynamicIndexing
= false;
1065 features
->shaderUniformTexelBufferArrayDynamicIndexing
= true;
1066 features
->shaderStorageTexelBufferArrayDynamicIndexing
= true;
1067 features
->shaderUniformBufferArrayNonUniformIndexing
= false;
1068 features
->shaderSampledImageArrayNonUniformIndexing
= true;
1069 features
->shaderStorageBufferArrayNonUniformIndexing
= true;
1070 features
->shaderStorageImageArrayNonUniformIndexing
= true;
1071 features
->shaderInputAttachmentArrayNonUniformIndexing
= false;
1072 features
->shaderUniformTexelBufferArrayNonUniformIndexing
= true;
1073 features
->shaderStorageTexelBufferArrayNonUniformIndexing
= true;
1074 features
->descriptorBindingUniformBufferUpdateAfterBind
= false;
1075 features
->descriptorBindingSampledImageUpdateAfterBind
= true;
1076 features
->descriptorBindingStorageImageUpdateAfterBind
= true;
1077 features
->descriptorBindingStorageBufferUpdateAfterBind
= true;
1078 features
->descriptorBindingUniformTexelBufferUpdateAfterBind
= true;
1079 features
->descriptorBindingStorageTexelBufferUpdateAfterBind
= true;
1080 features
->descriptorBindingUpdateUnusedWhilePending
= true;
1081 features
->descriptorBindingPartiallyBound
= true;
1082 features
->descriptorBindingVariableDescriptorCount
= false;
1083 features
->runtimeDescriptorArray
= true;
1087 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_FEATURES_EXT
: {
1088 VkPhysicalDeviceInlineUniformBlockFeaturesEXT
*features
=
1089 (VkPhysicalDeviceInlineUniformBlockFeaturesEXT
*)ext
;
1090 features
->inlineUniformBlock
= true;
1091 features
->descriptorBindingInlineUniformBlockUpdateAfterBind
= true;
1095 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES
: {
1096 VkPhysicalDeviceMultiviewFeatures
*features
=
1097 (VkPhysicalDeviceMultiviewFeatures
*)ext
;
1098 features
->multiview
= true;
1099 features
->multiviewGeometryShader
= true;
1100 features
->multiviewTessellationShader
= true;
1104 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES
: {
1105 VkPhysicalDeviceProtectedMemoryFeatures
*features
= (void *)ext
;
1106 features
->protectedMemory
= false;
1110 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES
: {
1111 VkPhysicalDeviceSamplerYcbcrConversionFeatures
*features
=
1112 (VkPhysicalDeviceSamplerYcbcrConversionFeatures
*) ext
;
1113 features
->samplerYcbcrConversion
= true;
1117 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES_EXT
: {
1118 VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*features
=
1119 (VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*)ext
;
1120 features
->scalarBlockLayout
= true;
1124 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_INT64_FEATURES_KHR
: {
1125 VkPhysicalDeviceShaderAtomicInt64FeaturesKHR
*features
= (void *)ext
;
1126 features
->shaderBufferInt64Atomics
=
1127 pdevice
->info
.gen
>= 9 && pdevice
->use_softpin
;
1128 features
->shaderSharedInt64Atomics
= VK_FALSE
;
1132 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETERS_FEATURES
: {
1133 VkPhysicalDeviceShaderDrawParametersFeatures
*features
= (void *)ext
;
1134 features
->shaderDrawParameters
= true;
1138 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES
: {
1139 VkPhysicalDeviceVariablePointersFeatures
*features
= (void *)ext
;
1140 features
->variablePointersStorageBuffer
= true;
1141 features
->variablePointers
= true;
1145 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT
: {
1146 VkPhysicalDeviceTransformFeedbackFeaturesEXT
*features
=
1147 (VkPhysicalDeviceTransformFeedbackFeaturesEXT
*)ext
;
1148 features
->transformFeedback
= true;
1149 features
->geometryStreams
= true;
1153 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT
: {
1154 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*features
=
1155 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*)ext
;
1156 features
->vertexAttributeInstanceRateDivisor
= true;
1157 features
->vertexAttributeInstanceRateZeroDivisor
= true;
1161 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_YCBCR_IMAGE_ARRAYS_FEATURES_EXT
: {
1162 VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*features
=
1163 (VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*)ext
;
1164 features
->ycbcrImageArrays
= true;
1169 anv_debug_ignored_stype(ext
->sType
);
1175 void anv_GetPhysicalDeviceProperties(
1176 VkPhysicalDevice physicalDevice
,
1177 VkPhysicalDeviceProperties
* pProperties
)
1179 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1180 const struct gen_device_info
*devinfo
= &pdevice
->info
;
1182 /* See assertions made when programming the buffer surface state. */
1183 const uint32_t max_raw_buffer_sz
= devinfo
->gen
>= 7 ?
1184 (1ul << 30) : (1ul << 27);
1186 const uint32_t max_ssbos
= pdevice
->has_a64_buffer_access
? UINT16_MAX
: 64;
1187 const uint32_t max_textures
=
1188 pdevice
->has_bindless_images
? UINT16_MAX
: 128;
1189 const uint32_t max_samplers
=
1190 pdevice
->has_bindless_samplers
? UINT16_MAX
:
1191 (devinfo
->gen
>= 8 || devinfo
->is_haswell
) ? 128 : 16;
1192 const uint32_t max_images
=
1193 pdevice
->has_bindless_images
? UINT16_MAX
: MAX_IMAGES
;
1195 /* The moment we have anything bindless, claim a high per-stage limit */
1196 const uint32_t max_per_stage
=
1197 pdevice
->has_a64_buffer_access
? UINT32_MAX
:
1198 MAX_BINDING_TABLE_SIZE
- MAX_RTS
;
1200 VkSampleCountFlags sample_counts
=
1201 isl_device_get_sample_counts(&pdevice
->isl_dev
);
1204 VkPhysicalDeviceLimits limits
= {
1205 .maxImageDimension1D
= (1 << 14),
1206 .maxImageDimension2D
= (1 << 14),
1207 .maxImageDimension3D
= (1 << 11),
1208 .maxImageDimensionCube
= (1 << 14),
1209 .maxImageArrayLayers
= (1 << 11),
1210 .maxTexelBufferElements
= 128 * 1024 * 1024,
1211 .maxUniformBufferRange
= (1ul << 27),
1212 .maxStorageBufferRange
= max_raw_buffer_sz
,
1213 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
1214 .maxMemoryAllocationCount
= UINT32_MAX
,
1215 .maxSamplerAllocationCount
= 64 * 1024,
1216 .bufferImageGranularity
= 64, /* A cache line */
1217 .sparseAddressSpaceSize
= 0,
1218 .maxBoundDescriptorSets
= MAX_SETS
,
1219 .maxPerStageDescriptorSamplers
= max_samplers
,
1220 .maxPerStageDescriptorUniformBuffers
= 64,
1221 .maxPerStageDescriptorStorageBuffers
= max_ssbos
,
1222 .maxPerStageDescriptorSampledImages
= max_textures
,
1223 .maxPerStageDescriptorStorageImages
= max_images
,
1224 .maxPerStageDescriptorInputAttachments
= 64,
1225 .maxPerStageResources
= max_per_stage
,
1226 .maxDescriptorSetSamplers
= 6 * max_samplers
, /* number of stages * maxPerStageDescriptorSamplers */
1227 .maxDescriptorSetUniformBuffers
= 6 * 64, /* number of stages * maxPerStageDescriptorUniformBuffers */
1228 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
1229 .maxDescriptorSetStorageBuffers
= 6 * max_ssbos
, /* number of stages * maxPerStageDescriptorStorageBuffers */
1230 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
1231 .maxDescriptorSetSampledImages
= 6 * max_textures
, /* number of stages * maxPerStageDescriptorSampledImages */
1232 .maxDescriptorSetStorageImages
= 6 * max_images
, /* number of stages * maxPerStageDescriptorStorageImages */
1233 .maxDescriptorSetInputAttachments
= 256,
1234 .maxVertexInputAttributes
= MAX_VBS
,
1235 .maxVertexInputBindings
= MAX_VBS
,
1236 .maxVertexInputAttributeOffset
= 2047,
1237 .maxVertexInputBindingStride
= 2048,
1238 .maxVertexOutputComponents
= 128,
1239 .maxTessellationGenerationLevel
= 64,
1240 .maxTessellationPatchSize
= 32,
1241 .maxTessellationControlPerVertexInputComponents
= 128,
1242 .maxTessellationControlPerVertexOutputComponents
= 128,
1243 .maxTessellationControlPerPatchOutputComponents
= 128,
1244 .maxTessellationControlTotalOutputComponents
= 2048,
1245 .maxTessellationEvaluationInputComponents
= 128,
1246 .maxTessellationEvaluationOutputComponents
= 128,
1247 .maxGeometryShaderInvocations
= 32,
1248 .maxGeometryInputComponents
= 64,
1249 .maxGeometryOutputComponents
= 128,
1250 .maxGeometryOutputVertices
= 256,
1251 .maxGeometryTotalOutputComponents
= 1024,
1252 .maxFragmentInputComponents
= 116, /* 128 components - (PSIZ, CLIP_DIST0, CLIP_DIST1) */
1253 .maxFragmentOutputAttachments
= 8,
1254 .maxFragmentDualSrcAttachments
= 1,
1255 .maxFragmentCombinedOutputResources
= 8,
1256 .maxComputeSharedMemorySize
= 32768,
1257 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
1258 .maxComputeWorkGroupInvocations
= 32 * devinfo
->max_cs_threads
,
1259 .maxComputeWorkGroupSize
= {
1260 16 * devinfo
->max_cs_threads
,
1261 16 * devinfo
->max_cs_threads
,
1262 16 * devinfo
->max_cs_threads
,
1264 .subPixelPrecisionBits
= 8,
1265 .subTexelPrecisionBits
= 8,
1266 .mipmapPrecisionBits
= 8,
1267 .maxDrawIndexedIndexValue
= UINT32_MAX
,
1268 .maxDrawIndirectCount
= UINT32_MAX
,
1269 .maxSamplerLodBias
= 16,
1270 .maxSamplerAnisotropy
= 16,
1271 .maxViewports
= MAX_VIEWPORTS
,
1272 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
1273 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
1274 .viewportSubPixelBits
= 13, /* We take a float? */
1275 .minMemoryMapAlignment
= 4096, /* A page */
1276 .minTexelBufferOffsetAlignment
= 1,
1277 /* We need 16 for UBO block reads to work and 32 for push UBOs */
1278 .minUniformBufferOffsetAlignment
= 32,
1279 .minStorageBufferOffsetAlignment
= 4,
1280 .minTexelOffset
= -8,
1281 .maxTexelOffset
= 7,
1282 .minTexelGatherOffset
= -32,
1283 .maxTexelGatherOffset
= 31,
1284 .minInterpolationOffset
= -0.5,
1285 .maxInterpolationOffset
= 0.4375,
1286 .subPixelInterpolationOffsetBits
= 4,
1287 .maxFramebufferWidth
= (1 << 14),
1288 .maxFramebufferHeight
= (1 << 14),
1289 .maxFramebufferLayers
= (1 << 11),
1290 .framebufferColorSampleCounts
= sample_counts
,
1291 .framebufferDepthSampleCounts
= sample_counts
,
1292 .framebufferStencilSampleCounts
= sample_counts
,
1293 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
1294 .maxColorAttachments
= MAX_RTS
,
1295 .sampledImageColorSampleCounts
= sample_counts
,
1296 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
1297 .sampledImageDepthSampleCounts
= sample_counts
,
1298 .sampledImageStencilSampleCounts
= sample_counts
,
1299 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
1300 .maxSampleMaskWords
= 1,
1301 .timestampComputeAndGraphics
= false,
1302 .timestampPeriod
= 1000000000.0 / devinfo
->timestamp_frequency
,
1303 .maxClipDistances
= 8,
1304 .maxCullDistances
= 8,
1305 .maxCombinedClipAndCullDistances
= 8,
1306 .discreteQueuePriorities
= 2,
1307 .pointSizeRange
= { 0.125, 255.875 },
1308 .lineWidthRange
= { 0.0, 7.9921875 },
1309 .pointSizeGranularity
= (1.0 / 8.0),
1310 .lineWidthGranularity
= (1.0 / 128.0),
1311 .strictLines
= false, /* FINISHME */
1312 .standardSampleLocations
= true,
1313 .optimalBufferCopyOffsetAlignment
= 128,
1314 .optimalBufferCopyRowPitchAlignment
= 128,
1315 .nonCoherentAtomSize
= 64,
1318 *pProperties
= (VkPhysicalDeviceProperties
) {
1319 .apiVersion
= anv_physical_device_api_version(pdevice
),
1320 .driverVersion
= vk_get_driver_version(),
1322 .deviceID
= pdevice
->chipset_id
,
1323 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
1325 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
1328 snprintf(pProperties
->deviceName
, sizeof(pProperties
->deviceName
),
1329 "%s", pdevice
->name
);
1330 memcpy(pProperties
->pipelineCacheUUID
,
1331 pdevice
->pipeline_cache_uuid
, VK_UUID_SIZE
);
1334 void anv_GetPhysicalDeviceProperties2(
1335 VkPhysicalDevice physicalDevice
,
1336 VkPhysicalDeviceProperties2
* pProperties
)
1338 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1340 anv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
1342 vk_foreach_struct(ext
, pProperties
->pNext
) {
1343 switch (ext
->sType
) {
1344 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_STENCIL_RESOLVE_PROPERTIES_KHR
: {
1345 VkPhysicalDeviceDepthStencilResolvePropertiesKHR
*props
=
1346 (VkPhysicalDeviceDepthStencilResolvePropertiesKHR
*)ext
;
1348 /* We support all of the depth resolve modes */
1349 props
->supportedDepthResolveModes
=
1350 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
|
1351 VK_RESOLVE_MODE_AVERAGE_BIT_KHR
|
1352 VK_RESOLVE_MODE_MIN_BIT_KHR
|
1353 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1355 /* Average doesn't make sense for stencil so we don't support that */
1356 props
->supportedStencilResolveModes
=
1357 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
;
1358 if (pdevice
->info
.gen
>= 8) {
1359 /* The advanced stencil resolve modes currently require stencil
1360 * sampling be supported by the hardware.
1362 props
->supportedStencilResolveModes
|=
1363 VK_RESOLVE_MODE_MIN_BIT_KHR
|
1364 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1367 props
->independentResolveNone
= true;
1368 props
->independentResolve
= true;
1372 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES_EXT
: {
1373 VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*props
=
1374 (VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*)ext
;
1376 /* It's a bit hard to exactly map our implementation to the limits
1377 * described here. The bindless surface handle in the extended
1378 * message descriptors is 20 bits and it's an index into the table of
1379 * RENDER_SURFACE_STATE structs that starts at bindless surface base
1380 * address. Given that most things consume two surface states per
1381 * view (general/sampled for textures and write-only/read-write for
1382 * images), we claim 2^19 things.
1384 * For SSBOs, we just use A64 messages so there is no real limit
1385 * there beyond the limit on the total size of a descriptor set.
1387 const unsigned max_bindless_views
= 1 << 19;
1389 props
->maxUpdateAfterBindDescriptorsInAllPools
= max_bindless_views
;
1390 props
->shaderUniformBufferArrayNonUniformIndexingNative
= false;
1391 props
->shaderSampledImageArrayNonUniformIndexingNative
= false;
1392 props
->shaderStorageBufferArrayNonUniformIndexingNative
= true;
1393 props
->shaderStorageImageArrayNonUniformIndexingNative
= false;
1394 props
->shaderInputAttachmentArrayNonUniformIndexingNative
= false;
1395 props
->robustBufferAccessUpdateAfterBind
= true;
1396 props
->quadDivergentImplicitLod
= false;
1397 props
->maxPerStageDescriptorUpdateAfterBindSamplers
= max_bindless_views
;
1398 props
->maxPerStageDescriptorUpdateAfterBindUniformBuffers
= 0;
1399 props
->maxPerStageDescriptorUpdateAfterBindStorageBuffers
= UINT32_MAX
;
1400 props
->maxPerStageDescriptorUpdateAfterBindSampledImages
= max_bindless_views
;
1401 props
->maxPerStageDescriptorUpdateAfterBindStorageImages
= max_bindless_views
;
1402 props
->maxPerStageDescriptorUpdateAfterBindInputAttachments
= 0;
1403 props
->maxPerStageUpdateAfterBindResources
= UINT32_MAX
;
1404 props
->maxDescriptorSetUpdateAfterBindSamplers
= max_bindless_views
;
1405 props
->maxDescriptorSetUpdateAfterBindUniformBuffers
= 0;
1406 props
->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic
= 0;
1407 props
->maxDescriptorSetUpdateAfterBindStorageBuffers
= UINT32_MAX
;
1408 props
->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2;
1409 props
->maxDescriptorSetUpdateAfterBindSampledImages
= max_bindless_views
;
1410 props
->maxDescriptorSetUpdateAfterBindStorageImages
= max_bindless_views
;
1411 props
->maxDescriptorSetUpdateAfterBindInputAttachments
= 0;
1415 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR
: {
1416 VkPhysicalDeviceDriverPropertiesKHR
*driver_props
=
1417 (VkPhysicalDeviceDriverPropertiesKHR
*) ext
;
1419 driver_props
->driverID
= VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA_KHR
;
1420 util_snprintf(driver_props
->driverName
, VK_MAX_DRIVER_NAME_SIZE_KHR
,
1421 "Intel open-source Mesa driver");
1423 util_snprintf(driver_props
->driverInfo
, VK_MAX_DRIVER_INFO_SIZE_KHR
,
1424 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1
);
1426 driver_props
->conformanceVersion
= (VkConformanceVersionKHR
) {
1435 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT
: {
1436 VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*props
=
1437 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*) ext
;
1438 /* Userptr needs page aligned memory. */
1439 props
->minImportedHostPointerAlignment
= 4096;
1443 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES
: {
1444 VkPhysicalDeviceIDProperties
*id_props
=
1445 (VkPhysicalDeviceIDProperties
*)ext
;
1446 memcpy(id_props
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
1447 memcpy(id_props
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
1448 /* The LUID is for Windows. */
1449 id_props
->deviceLUIDValid
= false;
1453 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_PROPERTIES_EXT
: {
1454 VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*props
=
1455 (VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*)ext
;
1456 props
->maxInlineUniformBlockSize
= MAX_INLINE_UNIFORM_BLOCK_SIZE
;
1457 props
->maxPerStageDescriptorInlineUniformBlocks
=
1458 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1459 props
->maxPerStageDescriptorUpdateAfterBindInlineUniformBlocks
=
1460 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1461 props
->maxDescriptorSetInlineUniformBlocks
=
1462 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1463 props
->maxDescriptorSetUpdateAfterBindInlineUniformBlocks
=
1464 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1468 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES
: {
1469 VkPhysicalDeviceMaintenance3Properties
*props
=
1470 (VkPhysicalDeviceMaintenance3Properties
*)ext
;
1471 /* This value doesn't matter for us today as our per-stage
1472 * descriptors are the real limit.
1474 props
->maxPerSetDescriptors
= 1024;
1475 props
->maxMemoryAllocationSize
= MAX_MEMORY_ALLOCATION_SIZE
;
1479 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES
: {
1480 VkPhysicalDeviceMultiviewProperties
*properties
=
1481 (VkPhysicalDeviceMultiviewProperties
*)ext
;
1482 properties
->maxMultiviewViewCount
= 16;
1483 properties
->maxMultiviewInstanceIndex
= UINT32_MAX
/ 16;
1487 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT
: {
1488 VkPhysicalDevicePCIBusInfoPropertiesEXT
*properties
=
1489 (VkPhysicalDevicePCIBusInfoPropertiesEXT
*)ext
;
1490 properties
->pciDomain
= pdevice
->pci_info
.domain
;
1491 properties
->pciBus
= pdevice
->pci_info
.bus
;
1492 properties
->pciDevice
= pdevice
->pci_info
.device
;
1493 properties
->pciFunction
= pdevice
->pci_info
.function
;
1497 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES
: {
1498 VkPhysicalDevicePointClippingProperties
*properties
=
1499 (VkPhysicalDevicePointClippingProperties
*) ext
;
1500 properties
->pointClippingBehavior
= VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES
;
1501 anv_finishme("Implement pop-free point clipping");
1505 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES
: {
1506 VkPhysicalDeviceProtectedMemoryProperties
*props
=
1507 (VkPhysicalDeviceProtectedMemoryProperties
*)ext
;
1508 props
->protectedNoFault
= false;
1512 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
1513 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
1514 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
1516 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
1520 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT
: {
1521 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*properties
=
1522 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*)ext
;
1523 properties
->filterMinmaxImageComponentMapping
= pdevice
->info
.gen
>= 9;
1524 properties
->filterMinmaxSingleComponentFormats
= true;
1528 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES
: {
1529 VkPhysicalDeviceSubgroupProperties
*properties
= (void *)ext
;
1531 properties
->subgroupSize
= BRW_SUBGROUP_SIZE
;
1533 VkShaderStageFlags scalar_stages
= 0;
1534 for (unsigned stage
= 0; stage
< MESA_SHADER_STAGES
; stage
++) {
1535 if (pdevice
->compiler
->scalar_stage
[stage
])
1536 scalar_stages
|= mesa_to_vk_shader_stage(stage
);
1538 properties
->supportedStages
= scalar_stages
;
1540 properties
->supportedOperations
= VK_SUBGROUP_FEATURE_BASIC_BIT
|
1541 VK_SUBGROUP_FEATURE_VOTE_BIT
|
1542 VK_SUBGROUP_FEATURE_ARITHMETIC_BIT
|
1543 VK_SUBGROUP_FEATURE_BALLOT_BIT
|
1544 VK_SUBGROUP_FEATURE_SHUFFLE_BIT
|
1545 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT
|
1546 VK_SUBGROUP_FEATURE_CLUSTERED_BIT
|
1547 VK_SUBGROUP_FEATURE_QUAD_BIT
;
1548 properties
->quadOperationsInAllStages
= true;
1552 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT
: {
1553 VkPhysicalDeviceTransformFeedbackPropertiesEXT
*props
=
1554 (VkPhysicalDeviceTransformFeedbackPropertiesEXT
*)ext
;
1556 props
->maxTransformFeedbackStreams
= MAX_XFB_STREAMS
;
1557 props
->maxTransformFeedbackBuffers
= MAX_XFB_BUFFERS
;
1558 props
->maxTransformFeedbackBufferSize
= (1ull << 32);
1559 props
->maxTransformFeedbackStreamDataSize
= 128 * 4;
1560 props
->maxTransformFeedbackBufferDataSize
= 128 * 4;
1561 props
->maxTransformFeedbackBufferDataStride
= 2048;
1562 props
->transformFeedbackQueries
= true;
1563 props
->transformFeedbackStreamsLinesTriangles
= false;
1564 props
->transformFeedbackRasterizationStreamSelect
= false;
1565 props
->transformFeedbackDraw
= true;
1569 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT
: {
1570 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*props
=
1571 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*)ext
;
1572 /* We have to restrict this a bit for multiview */
1573 props
->maxVertexAttribDivisor
= UINT32_MAX
/ 16;
1578 anv_debug_ignored_stype(ext
->sType
);
1584 /* We support exactly one queue family. */
1585 static const VkQueueFamilyProperties
1586 anv_queue_family_properties
= {
1587 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
1588 VK_QUEUE_COMPUTE_BIT
|
1589 VK_QUEUE_TRANSFER_BIT
,
1591 .timestampValidBits
= 36, /* XXX: Real value here */
1592 .minImageTransferGranularity
= { 1, 1, 1 },
1595 void anv_GetPhysicalDeviceQueueFamilyProperties(
1596 VkPhysicalDevice physicalDevice
,
1598 VkQueueFamilyProperties
* pQueueFamilyProperties
)
1600 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pCount
);
1602 vk_outarray_append(&out
, p
) {
1603 *p
= anv_queue_family_properties
;
1607 void anv_GetPhysicalDeviceQueueFamilyProperties2(
1608 VkPhysicalDevice physicalDevice
,
1609 uint32_t* pQueueFamilyPropertyCount
,
1610 VkQueueFamilyProperties2
* pQueueFamilyProperties
)
1613 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pQueueFamilyPropertyCount
);
1615 vk_outarray_append(&out
, p
) {
1616 p
->queueFamilyProperties
= anv_queue_family_properties
;
1618 vk_foreach_struct(s
, p
->pNext
) {
1619 anv_debug_ignored_stype(s
->sType
);
1624 void anv_GetPhysicalDeviceMemoryProperties(
1625 VkPhysicalDevice physicalDevice
,
1626 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
1628 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
1630 pMemoryProperties
->memoryTypeCount
= physical_device
->memory
.type_count
;
1631 for (uint32_t i
= 0; i
< physical_device
->memory
.type_count
; i
++) {
1632 pMemoryProperties
->memoryTypes
[i
] = (VkMemoryType
) {
1633 .propertyFlags
= physical_device
->memory
.types
[i
].propertyFlags
,
1634 .heapIndex
= physical_device
->memory
.types
[i
].heapIndex
,
1638 pMemoryProperties
->memoryHeapCount
= physical_device
->memory
.heap_count
;
1639 for (uint32_t i
= 0; i
< physical_device
->memory
.heap_count
; i
++) {
1640 pMemoryProperties
->memoryHeaps
[i
] = (VkMemoryHeap
) {
1641 .size
= physical_device
->memory
.heaps
[i
].size
,
1642 .flags
= physical_device
->memory
.heaps
[i
].flags
,
1648 anv_get_memory_budget(VkPhysicalDevice physicalDevice
,
1649 VkPhysicalDeviceMemoryBudgetPropertiesEXT
*memoryBudget
)
1651 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
1652 uint64_t sys_available
= get_available_system_memory();
1653 assert(sys_available
> 0);
1655 VkDeviceSize total_heaps_size
= 0;
1656 for (size_t i
= 0; i
< device
->memory
.heap_count
; i
++)
1657 total_heaps_size
+= device
->memory
.heaps
[i
].size
;
1659 for (size_t i
= 0; i
< device
->memory
.heap_count
; i
++) {
1660 VkDeviceSize heap_size
= device
->memory
.heaps
[i
].size
;
1661 VkDeviceSize heap_used
= device
->memory
.heaps
[i
].used
;
1662 VkDeviceSize heap_budget
;
1664 double heap_proportion
= (double) heap_size
/ total_heaps_size
;
1665 VkDeviceSize sys_available_prop
= sys_available
* heap_proportion
;
1668 * Let's not incite the app to starve the system: report at most 90% of
1669 * available system memory.
1671 uint64_t heap_available
= sys_available_prop
* 9 / 10;
1672 heap_budget
= MIN2(heap_size
, heap_used
+ heap_available
);
1675 * Round down to the nearest MB
1677 heap_budget
&= ~((1ull << 20) - 1);
1680 * The heapBudget value must be non-zero for array elements less than
1681 * VkPhysicalDeviceMemoryProperties::memoryHeapCount. The heapBudget
1682 * value must be less than or equal to VkMemoryHeap::size for each heap.
1684 assert(0 < heap_budget
&& heap_budget
<= heap_size
);
1686 memoryBudget
->heapUsage
[i
] = heap_used
;
1687 memoryBudget
->heapBudget
[i
] = heap_budget
;
1690 /* The heapBudget and heapUsage values must be zero for array elements
1691 * greater than or equal to VkPhysicalDeviceMemoryProperties::memoryHeapCount
1693 for (uint32_t i
= device
->memory
.heap_count
; i
< VK_MAX_MEMORY_HEAPS
; i
++) {
1694 memoryBudget
->heapBudget
[i
] = 0;
1695 memoryBudget
->heapUsage
[i
] = 0;
1699 void anv_GetPhysicalDeviceMemoryProperties2(
1700 VkPhysicalDevice physicalDevice
,
1701 VkPhysicalDeviceMemoryProperties2
* pMemoryProperties
)
1703 anv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
1704 &pMemoryProperties
->memoryProperties
);
1706 vk_foreach_struct(ext
, pMemoryProperties
->pNext
) {
1707 switch (ext
->sType
) {
1708 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT
:
1709 anv_get_memory_budget(physicalDevice
, (void*)ext
);
1712 anv_debug_ignored_stype(ext
->sType
);
1719 anv_GetDeviceGroupPeerMemoryFeatures(
1722 uint32_t localDeviceIndex
,
1723 uint32_t remoteDeviceIndex
,
1724 VkPeerMemoryFeatureFlags
* pPeerMemoryFeatures
)
1726 assert(localDeviceIndex
== 0 && remoteDeviceIndex
== 0);
1727 *pPeerMemoryFeatures
= VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT
|
1728 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT
|
1729 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT
|
1730 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT
;
1733 PFN_vkVoidFunction
anv_GetInstanceProcAddr(
1734 VkInstance _instance
,
1737 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1739 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
1740 * when we have to return valid function pointers, NULL, or it's left
1741 * undefined. See the table for exact details.
1746 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
1747 if (strcmp(pName, "vk" #entrypoint) == 0) \
1748 return (PFN_vkVoidFunction)anv_##entrypoint
1750 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties
);
1751 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties
);
1752 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceVersion
);
1753 LOOKUP_ANV_ENTRYPOINT(CreateInstance
);
1755 #undef LOOKUP_ANV_ENTRYPOINT
1757 if (instance
== NULL
)
1760 int idx
= anv_get_instance_entrypoint_index(pName
);
1762 return instance
->dispatch
.entrypoints
[idx
];
1764 idx
= anv_get_device_entrypoint_index(pName
);
1766 return instance
->device_dispatch
.entrypoints
[idx
];
1771 /* With version 1+ of the loader interface the ICD should expose
1772 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
1775 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1776 VkInstance instance
,
1780 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1781 VkInstance instance
,
1784 return anv_GetInstanceProcAddr(instance
, pName
);
1787 PFN_vkVoidFunction
anv_GetDeviceProcAddr(
1791 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1793 if (!device
|| !pName
)
1796 int idx
= anv_get_device_entrypoint_index(pName
);
1800 return device
->dispatch
.entrypoints
[idx
];
1804 anv_CreateDebugReportCallbackEXT(VkInstance _instance
,
1805 const VkDebugReportCallbackCreateInfoEXT
* pCreateInfo
,
1806 const VkAllocationCallbacks
* pAllocator
,
1807 VkDebugReportCallbackEXT
* pCallback
)
1809 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1810 return vk_create_debug_report_callback(&instance
->debug_report_callbacks
,
1811 pCreateInfo
, pAllocator
, &instance
->alloc
,
1816 anv_DestroyDebugReportCallbackEXT(VkInstance _instance
,
1817 VkDebugReportCallbackEXT _callback
,
1818 const VkAllocationCallbacks
* pAllocator
)
1820 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1821 vk_destroy_debug_report_callback(&instance
->debug_report_callbacks
,
1822 _callback
, pAllocator
, &instance
->alloc
);
1826 anv_DebugReportMessageEXT(VkInstance _instance
,
1827 VkDebugReportFlagsEXT flags
,
1828 VkDebugReportObjectTypeEXT objectType
,
1831 int32_t messageCode
,
1832 const char* pLayerPrefix
,
1833 const char* pMessage
)
1835 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1836 vk_debug_report(&instance
->debug_report_callbacks
, flags
, objectType
,
1837 object
, location
, messageCode
, pLayerPrefix
, pMessage
);
1841 anv_queue_init(struct anv_device
*device
, struct anv_queue
*queue
)
1843 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1844 queue
->device
= device
;
1849 anv_queue_finish(struct anv_queue
*queue
)
1853 static struct anv_state
1854 anv_state_pool_emit_data(struct anv_state_pool
*pool
, size_t size
, size_t align
, const void *p
)
1856 struct anv_state state
;
1858 state
= anv_state_pool_alloc(pool
, size
, align
);
1859 memcpy(state
.map
, p
, size
);
1864 struct gen8_border_color
{
1869 /* Pad out to 64 bytes */
1874 anv_device_init_border_colors(struct anv_device
*device
)
1876 static const struct gen8_border_color border_colors
[] = {
1877 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
1878 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
1879 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
1880 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
1881 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
1882 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
1885 device
->border_colors
= anv_state_pool_emit_data(&device
->dynamic_state_pool
,
1886 sizeof(border_colors
), 64,
1891 anv_device_init_trivial_batch(struct anv_device
*device
)
1893 anv_bo_init_new(&device
->trivial_batch_bo
, device
, 4096);
1895 if (device
->instance
->physicalDevice
.has_exec_async
)
1896 device
->trivial_batch_bo
.flags
|= EXEC_OBJECT_ASYNC
;
1898 if (device
->instance
->physicalDevice
.use_softpin
)
1899 device
->trivial_batch_bo
.flags
|= EXEC_OBJECT_PINNED
;
1901 anv_vma_alloc(device
, &device
->trivial_batch_bo
);
1903 void *map
= anv_gem_mmap(device
, device
->trivial_batch_bo
.gem_handle
,
1906 struct anv_batch batch
= {
1912 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
1913 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
1915 if (!device
->info
.has_llc
)
1916 gen_clflush_range(map
, batch
.next
- map
);
1918 anv_gem_munmap(map
, device
->trivial_batch_bo
.size
);
1921 VkResult
anv_EnumerateDeviceExtensionProperties(
1922 VkPhysicalDevice physicalDevice
,
1923 const char* pLayerName
,
1924 uint32_t* pPropertyCount
,
1925 VkExtensionProperties
* pProperties
)
1927 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
1928 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
1930 for (int i
= 0; i
< ANV_DEVICE_EXTENSION_COUNT
; i
++) {
1931 if (device
->supported_extensions
.extensions
[i
]) {
1932 vk_outarray_append(&out
, prop
) {
1933 *prop
= anv_device_extensions
[i
];
1938 return vk_outarray_status(&out
);
1942 anv_device_init_dispatch(struct anv_device
*device
)
1944 const struct anv_device_dispatch_table
*genX_table
;
1945 switch (device
->info
.gen
) {
1947 genX_table
= &gen11_device_dispatch_table
;
1950 genX_table
= &gen10_device_dispatch_table
;
1953 genX_table
= &gen9_device_dispatch_table
;
1956 genX_table
= &gen8_device_dispatch_table
;
1959 if (device
->info
.is_haswell
)
1960 genX_table
= &gen75_device_dispatch_table
;
1962 genX_table
= &gen7_device_dispatch_table
;
1965 unreachable("unsupported gen\n");
1968 for (unsigned i
= 0; i
< ARRAY_SIZE(device
->dispatch
.entrypoints
); i
++) {
1969 /* Vulkan requires that entrypoints for extensions which have not been
1970 * enabled must not be advertised.
1972 if (!anv_device_entrypoint_is_enabled(i
, device
->instance
->app_info
.api_version
,
1973 &device
->instance
->enabled_extensions
,
1974 &device
->enabled_extensions
)) {
1975 device
->dispatch
.entrypoints
[i
] = NULL
;
1976 } else if (genX_table
->entrypoints
[i
]) {
1977 device
->dispatch
.entrypoints
[i
] = genX_table
->entrypoints
[i
];
1979 device
->dispatch
.entrypoints
[i
] =
1980 anv_device_dispatch_table
.entrypoints
[i
];
1986 vk_priority_to_gen(int priority
)
1989 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT
:
1990 return GEN_CONTEXT_LOW_PRIORITY
;
1991 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
:
1992 return GEN_CONTEXT_MEDIUM_PRIORITY
;
1993 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT
:
1994 return GEN_CONTEXT_HIGH_PRIORITY
;
1995 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT
:
1996 return GEN_CONTEXT_REALTIME_PRIORITY
;
1998 unreachable("Invalid priority");
2003 anv_device_init_hiz_clear_value_bo(struct anv_device
*device
)
2005 anv_bo_init_new(&device
->hiz_clear_bo
, device
, 4096);
2007 if (device
->instance
->physicalDevice
.has_exec_async
)
2008 device
->hiz_clear_bo
.flags
|= EXEC_OBJECT_ASYNC
;
2010 if (device
->instance
->physicalDevice
.use_softpin
)
2011 device
->hiz_clear_bo
.flags
|= EXEC_OBJECT_PINNED
;
2013 anv_vma_alloc(device
, &device
->hiz_clear_bo
);
2015 uint32_t *map
= anv_gem_mmap(device
, device
->hiz_clear_bo
.gem_handle
,
2018 union isl_color_value hiz_clear
= { .u32
= { 0, } };
2019 hiz_clear
.f32
[0] = ANV_HZ_FC_VAL
;
2021 memcpy(map
, hiz_clear
.u32
, sizeof(hiz_clear
.u32
));
2022 anv_gem_munmap(map
, device
->hiz_clear_bo
.size
);
2026 get_bo_from_pool(struct gen_batch_decode_bo
*ret
,
2027 struct anv_block_pool
*pool
,
2030 for (uint32_t i
= 0; i
< pool
->nbos
; i
++) {
2031 uint64_t bo_address
= pool
->bos
[i
].offset
& (~0ull >> 16);
2032 uint32_t bo_size
= pool
->bos
[i
].size
;
2033 if (address
>= bo_address
&& address
< (bo_address
+ bo_size
)) {
2034 *ret
= (struct gen_batch_decode_bo
) {
2037 .map
= pool
->bos
[i
].map
,
2045 /* Finding a buffer for batch decoding */
2046 static struct gen_batch_decode_bo
2047 decode_get_bo(void *v_batch
, bool ppgtt
, uint64_t address
)
2049 struct anv_device
*device
= v_batch
;
2050 struct gen_batch_decode_bo ret_bo
= {};
2054 if (get_bo_from_pool(&ret_bo
, &device
->dynamic_state_pool
.block_pool
, address
))
2056 if (get_bo_from_pool(&ret_bo
, &device
->instruction_state_pool
.block_pool
, address
))
2058 if (get_bo_from_pool(&ret_bo
, &device
->binding_table_pool
.block_pool
, address
))
2060 if (get_bo_from_pool(&ret_bo
, &device
->surface_state_pool
.block_pool
, address
))
2063 if (!device
->cmd_buffer_being_decoded
)
2064 return (struct gen_batch_decode_bo
) { };
2066 struct anv_batch_bo
**bo
;
2068 u_vector_foreach(bo
, &device
->cmd_buffer_being_decoded
->seen_bbos
) {
2069 /* The decoder zeroes out the top 16 bits, so we need to as well */
2070 uint64_t bo_address
= (*bo
)->bo
.offset
& (~0ull >> 16);
2072 if (address
>= bo_address
&& address
< bo_address
+ (*bo
)->bo
.size
) {
2073 return (struct gen_batch_decode_bo
) {
2075 .size
= (*bo
)->bo
.size
,
2076 .map
= (*bo
)->bo
.map
,
2081 return (struct gen_batch_decode_bo
) { };
2084 VkResult
anv_CreateDevice(
2085 VkPhysicalDevice physicalDevice
,
2086 const VkDeviceCreateInfo
* pCreateInfo
,
2087 const VkAllocationCallbacks
* pAllocator
,
2090 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
2092 struct anv_device
*device
;
2094 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO
);
2096 struct anv_device_extension_table enabled_extensions
= { };
2097 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
2099 for (idx
= 0; idx
< ANV_DEVICE_EXTENSION_COUNT
; idx
++) {
2100 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
2101 anv_device_extensions
[idx
].extensionName
) == 0)
2105 if (idx
>= ANV_DEVICE_EXTENSION_COUNT
)
2106 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
2108 if (!physical_device
->supported_extensions
.extensions
[idx
])
2109 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
2111 enabled_extensions
.extensions
[idx
] = true;
2114 /* Check enabled features */
2115 if (pCreateInfo
->pEnabledFeatures
) {
2116 VkPhysicalDeviceFeatures supported_features
;
2117 anv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
2118 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
2119 VkBool32
*enabled_feature
= (VkBool32
*)pCreateInfo
->pEnabledFeatures
;
2120 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
2121 for (uint32_t i
= 0; i
< num_features
; i
++) {
2122 if (enabled_feature
[i
] && !supported_feature
[i
])
2123 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
2127 /* Check requested queues and fail if we are requested to create any
2128 * queues with flags we don't support.
2130 assert(pCreateInfo
->queueCreateInfoCount
> 0);
2131 for (uint32_t i
= 0; i
< pCreateInfo
->queueCreateInfoCount
; i
++) {
2132 if (pCreateInfo
->pQueueCreateInfos
[i
].flags
!= 0)
2133 return vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2136 /* Check if client specified queue priority. */
2137 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*queue_priority
=
2138 vk_find_struct_const(pCreateInfo
->pQueueCreateInfos
[0].pNext
,
2139 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT
);
2141 VkQueueGlobalPriorityEXT priority
=
2142 queue_priority
? queue_priority
->globalPriority
:
2143 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
;
2145 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
2147 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
2149 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2151 const unsigned decode_flags
=
2152 GEN_BATCH_DECODE_FULL
|
2153 ((INTEL_DEBUG
& DEBUG_COLOR
) ? GEN_BATCH_DECODE_IN_COLOR
: 0) |
2154 GEN_BATCH_DECODE_OFFSETS
|
2155 GEN_BATCH_DECODE_FLOATS
;
2157 gen_batch_decode_ctx_init(&device
->decoder_ctx
,
2158 &physical_device
->info
,
2159 stderr
, decode_flags
, NULL
,
2160 decode_get_bo
, NULL
, device
);
2162 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
2163 device
->instance
= physical_device
->instance
;
2164 device
->chipset_id
= physical_device
->chipset_id
;
2165 device
->no_hw
= physical_device
->no_hw
;
2166 device
->_lost
= false;
2169 device
->alloc
= *pAllocator
;
2171 device
->alloc
= physical_device
->instance
->alloc
;
2173 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
2174 device
->fd
= open(physical_device
->path
, O_RDWR
| O_CLOEXEC
);
2175 if (device
->fd
== -1) {
2176 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2180 device
->context_id
= anv_gem_create_context(device
);
2181 if (device
->context_id
== -1) {
2182 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2186 if (physical_device
->use_softpin
) {
2187 if (pthread_mutex_init(&device
->vma_mutex
, NULL
) != 0) {
2188 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2192 /* keep the page with address zero out of the allocator */
2193 struct anv_memory_heap
*low_heap
=
2194 &physical_device
->memory
.heaps
[physical_device
->memory
.heap_count
- 1];
2195 util_vma_heap_init(&device
->vma_lo
, low_heap
->vma_start
, low_heap
->vma_size
);
2196 device
->vma_lo_available
= low_heap
->size
;
2198 struct anv_memory_heap
*high_heap
=
2199 &physical_device
->memory
.heaps
[0];
2200 util_vma_heap_init(&device
->vma_hi
, high_heap
->vma_start
, high_heap
->vma_size
);
2201 device
->vma_hi_available
= physical_device
->memory
.heap_count
== 1 ? 0 :
2205 list_inithead(&device
->memory_objects
);
2207 /* As per spec, the driver implementation may deny requests to acquire
2208 * a priority above the default priority (MEDIUM) if the caller does not
2209 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
2212 if (physical_device
->has_context_priority
) {
2213 int err
= anv_gem_set_context_param(device
->fd
, device
->context_id
,
2214 I915_CONTEXT_PARAM_PRIORITY
,
2215 vk_priority_to_gen(priority
));
2216 if (err
!= 0 && priority
> VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
) {
2217 result
= vk_error(VK_ERROR_NOT_PERMITTED_EXT
);
2222 device
->info
= physical_device
->info
;
2223 device
->isl_dev
= physical_device
->isl_dev
;
2225 /* On Broadwell and later, we can use batch chaining to more efficiently
2226 * implement growing command buffers. Prior to Haswell, the kernel
2227 * command parser gets in the way and we have to fall back to growing
2230 device
->can_chain_batches
= device
->info
.gen
>= 8;
2232 device
->robust_buffer_access
= pCreateInfo
->pEnabledFeatures
&&
2233 pCreateInfo
->pEnabledFeatures
->robustBufferAccess
;
2234 device
->enabled_extensions
= enabled_extensions
;
2236 anv_device_init_dispatch(device
);
2238 if (pthread_mutex_init(&device
->mutex
, NULL
) != 0) {
2239 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2240 goto fail_context_id
;
2243 pthread_condattr_t condattr
;
2244 if (pthread_condattr_init(&condattr
) != 0) {
2245 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2248 if (pthread_condattr_setclock(&condattr
, CLOCK_MONOTONIC
) != 0) {
2249 pthread_condattr_destroy(&condattr
);
2250 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2253 if (pthread_cond_init(&device
->queue_submit
, &condattr
) != 0) {
2254 pthread_condattr_destroy(&condattr
);
2255 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2258 pthread_condattr_destroy(&condattr
);
2261 (physical_device
->supports_48bit_addresses
? EXEC_OBJECT_SUPPORTS_48B_ADDRESS
: 0) |
2262 (physical_device
->has_exec_async
? EXEC_OBJECT_ASYNC
: 0) |
2263 (physical_device
->has_exec_capture
? EXEC_OBJECT_CAPTURE
: 0) |
2264 (physical_device
->use_softpin
? EXEC_OBJECT_PINNED
: 0);
2266 anv_bo_pool_init(&device
->batch_bo_pool
, device
, bo_flags
);
2268 result
= anv_bo_cache_init(&device
->bo_cache
);
2269 if (result
!= VK_SUCCESS
)
2270 goto fail_batch_bo_pool
;
2272 if (!physical_device
->use_softpin
)
2273 bo_flags
&= ~EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
2275 result
= anv_state_pool_init(&device
->dynamic_state_pool
, device
,
2276 DYNAMIC_STATE_POOL_MIN_ADDRESS
,
2279 if (result
!= VK_SUCCESS
)
2282 result
= anv_state_pool_init(&device
->instruction_state_pool
, device
,
2283 INSTRUCTION_STATE_POOL_MIN_ADDRESS
,
2286 if (result
!= VK_SUCCESS
)
2287 goto fail_dynamic_state_pool
;
2289 result
= anv_state_pool_init(&device
->surface_state_pool
, device
,
2290 SURFACE_STATE_POOL_MIN_ADDRESS
,
2293 if (result
!= VK_SUCCESS
)
2294 goto fail_instruction_state_pool
;
2296 if (physical_device
->use_softpin
) {
2297 result
= anv_state_pool_init(&device
->binding_table_pool
, device
,
2298 BINDING_TABLE_POOL_MIN_ADDRESS
,
2301 if (result
!= VK_SUCCESS
)
2302 goto fail_surface_state_pool
;
2305 result
= anv_bo_init_new(&device
->workaround_bo
, device
, 1024);
2306 if (result
!= VK_SUCCESS
)
2307 goto fail_binding_table_pool
;
2309 if (physical_device
->use_softpin
)
2310 device
->workaround_bo
.flags
|= EXEC_OBJECT_PINNED
;
2312 if (!anv_vma_alloc(device
, &device
->workaround_bo
))
2313 goto fail_workaround_bo
;
2315 anv_device_init_trivial_batch(device
);
2317 if (device
->info
.gen
>= 10)
2318 anv_device_init_hiz_clear_value_bo(device
);
2320 anv_scratch_pool_init(device
, &device
->scratch_pool
);
2322 anv_queue_init(device
, &device
->queue
);
2324 switch (device
->info
.gen
) {
2326 if (!device
->info
.is_haswell
)
2327 result
= gen7_init_device_state(device
);
2329 result
= gen75_init_device_state(device
);
2332 result
= gen8_init_device_state(device
);
2335 result
= gen9_init_device_state(device
);
2338 result
= gen10_init_device_state(device
);
2341 result
= gen11_init_device_state(device
);
2344 /* Shouldn't get here as we don't create physical devices for any other
2346 unreachable("unhandled gen");
2348 if (result
!= VK_SUCCESS
)
2349 goto fail_workaround_bo
;
2351 anv_pipeline_cache_init(&device
->default_pipeline_cache
, device
, true);
2353 anv_device_init_blorp(device
);
2355 anv_device_init_border_colors(device
);
2357 *pDevice
= anv_device_to_handle(device
);
2362 anv_queue_finish(&device
->queue
);
2363 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
2364 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
2365 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
2366 fail_binding_table_pool
:
2367 if (physical_device
->use_softpin
)
2368 anv_state_pool_finish(&device
->binding_table_pool
);
2369 fail_surface_state_pool
:
2370 anv_state_pool_finish(&device
->surface_state_pool
);
2371 fail_instruction_state_pool
:
2372 anv_state_pool_finish(&device
->instruction_state_pool
);
2373 fail_dynamic_state_pool
:
2374 anv_state_pool_finish(&device
->dynamic_state_pool
);
2376 anv_bo_cache_finish(&device
->bo_cache
);
2378 anv_bo_pool_finish(&device
->batch_bo_pool
);
2379 pthread_cond_destroy(&device
->queue_submit
);
2381 pthread_mutex_destroy(&device
->mutex
);
2383 anv_gem_destroy_context(device
, device
->context_id
);
2387 vk_free(&device
->alloc
, device
);
2392 void anv_DestroyDevice(
2394 const VkAllocationCallbacks
* pAllocator
)
2396 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2397 struct anv_physical_device
*physical_device
;
2402 physical_device
= &device
->instance
->physicalDevice
;
2404 anv_device_finish_blorp(device
);
2406 anv_pipeline_cache_finish(&device
->default_pipeline_cache
);
2408 anv_queue_finish(&device
->queue
);
2410 #ifdef HAVE_VALGRIND
2411 /* We only need to free these to prevent valgrind errors. The backing
2412 * BO will go away in a couple of lines so we don't actually leak.
2414 anv_state_pool_free(&device
->dynamic_state_pool
, device
->border_colors
);
2417 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
2419 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
2420 anv_vma_free(device
, &device
->workaround_bo
);
2421 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
2423 anv_vma_free(device
, &device
->trivial_batch_bo
);
2424 anv_gem_close(device
, device
->trivial_batch_bo
.gem_handle
);
2425 if (device
->info
.gen
>= 10)
2426 anv_gem_close(device
, device
->hiz_clear_bo
.gem_handle
);
2428 if (physical_device
->use_softpin
)
2429 anv_state_pool_finish(&device
->binding_table_pool
);
2430 anv_state_pool_finish(&device
->surface_state_pool
);
2431 anv_state_pool_finish(&device
->instruction_state_pool
);
2432 anv_state_pool_finish(&device
->dynamic_state_pool
);
2434 anv_bo_cache_finish(&device
->bo_cache
);
2436 anv_bo_pool_finish(&device
->batch_bo_pool
);
2438 pthread_cond_destroy(&device
->queue_submit
);
2439 pthread_mutex_destroy(&device
->mutex
);
2441 anv_gem_destroy_context(device
, device
->context_id
);
2443 gen_batch_decode_ctx_finish(&device
->decoder_ctx
);
2447 vk_free(&device
->alloc
, device
);
2450 VkResult
anv_EnumerateInstanceLayerProperties(
2451 uint32_t* pPropertyCount
,
2452 VkLayerProperties
* pProperties
)
2454 if (pProperties
== NULL
) {
2455 *pPropertyCount
= 0;
2459 /* None supported at this time */
2460 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
2463 VkResult
anv_EnumerateDeviceLayerProperties(
2464 VkPhysicalDevice physicalDevice
,
2465 uint32_t* pPropertyCount
,
2466 VkLayerProperties
* pProperties
)
2468 if (pProperties
== NULL
) {
2469 *pPropertyCount
= 0;
2473 /* None supported at this time */
2474 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
2477 void anv_GetDeviceQueue(
2479 uint32_t queueNodeIndex
,
2480 uint32_t queueIndex
,
2483 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2485 assert(queueIndex
== 0);
2487 *pQueue
= anv_queue_to_handle(&device
->queue
);
2490 void anv_GetDeviceQueue2(
2492 const VkDeviceQueueInfo2
* pQueueInfo
,
2495 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2497 assert(pQueueInfo
->queueIndex
== 0);
2499 if (pQueueInfo
->flags
== device
->queue
.flags
)
2500 *pQueue
= anv_queue_to_handle(&device
->queue
);
2506 _anv_device_set_lost(struct anv_device
*device
,
2507 const char *file
, int line
,
2508 const char *msg
, ...)
2513 device
->_lost
= true;
2516 err
= __vk_errorv(device
->instance
, device
,
2517 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT
,
2518 VK_ERROR_DEVICE_LOST
, file
, line
, msg
, ap
);
2521 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
2528 anv_device_query_status(struct anv_device
*device
)
2530 /* This isn't likely as most of the callers of this function already check
2531 * for it. However, it doesn't hurt to check and it potentially lets us
2534 if (anv_device_is_lost(device
))
2535 return VK_ERROR_DEVICE_LOST
;
2537 uint32_t active
, pending
;
2538 int ret
= anv_gem_gpu_get_reset_stats(device
, &active
, &pending
);
2540 /* We don't know the real error. */
2541 return anv_device_set_lost(device
, "get_reset_stats failed: %m");
2545 return anv_device_set_lost(device
, "GPU hung on one of our command buffers");
2546 } else if (pending
) {
2547 return anv_device_set_lost(device
, "GPU hung with commands in-flight");
2554 anv_device_bo_busy(struct anv_device
*device
, struct anv_bo
*bo
)
2556 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
2557 * Other usages of the BO (such as on different hardware) will not be
2558 * flagged as "busy" by this ioctl. Use with care.
2560 int ret
= anv_gem_busy(device
, bo
->gem_handle
);
2562 return VK_NOT_READY
;
2563 } else if (ret
== -1) {
2564 /* We don't know the real error. */
2565 return anv_device_set_lost(device
, "gem wait failed: %m");
2568 /* Query for device status after the busy call. If the BO we're checking
2569 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
2570 * client because it clearly doesn't have valid data. Yes, this most
2571 * likely means an ioctl, but we just did an ioctl to query the busy status
2572 * so it's no great loss.
2574 return anv_device_query_status(device
);
2578 anv_device_wait(struct anv_device
*device
, struct anv_bo
*bo
,
2581 int ret
= anv_gem_wait(device
, bo
->gem_handle
, &timeout
);
2582 if (ret
== -1 && errno
== ETIME
) {
2584 } else if (ret
== -1) {
2585 /* We don't know the real error. */
2586 return anv_device_set_lost(device
, "gem wait failed: %m");
2589 /* Query for device status after the wait. If the BO we're waiting on got
2590 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
2591 * because it clearly doesn't have valid data. Yes, this most likely means
2592 * an ioctl, but we just did an ioctl to wait so it's no great loss.
2594 return anv_device_query_status(device
);
2597 VkResult
anv_DeviceWaitIdle(
2600 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2601 if (anv_device_is_lost(device
))
2602 return VK_ERROR_DEVICE_LOST
;
2604 struct anv_batch batch
;
2607 batch
.start
= batch
.next
= cmds
;
2608 batch
.end
= (void *) cmds
+ sizeof(cmds
);
2610 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
2611 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
2613 return anv_device_submit_simple_batch(device
, &batch
);
2617 anv_vma_alloc(struct anv_device
*device
, struct anv_bo
*bo
)
2619 if (!(bo
->flags
& EXEC_OBJECT_PINNED
))
2622 pthread_mutex_lock(&device
->vma_mutex
);
2626 if (bo
->flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
&&
2627 device
->vma_hi_available
>= bo
->size
) {
2628 uint64_t addr
= util_vma_heap_alloc(&device
->vma_hi
, bo
->size
, 4096);
2630 bo
->offset
= gen_canonical_address(addr
);
2631 assert(addr
== gen_48b_address(bo
->offset
));
2632 device
->vma_hi_available
-= bo
->size
;
2636 if (bo
->offset
== 0 && device
->vma_lo_available
>= bo
->size
) {
2637 uint64_t addr
= util_vma_heap_alloc(&device
->vma_lo
, bo
->size
, 4096);
2639 bo
->offset
= gen_canonical_address(addr
);
2640 assert(addr
== gen_48b_address(bo
->offset
));
2641 device
->vma_lo_available
-= bo
->size
;
2645 pthread_mutex_unlock(&device
->vma_mutex
);
2647 return bo
->offset
!= 0;
2651 anv_vma_free(struct anv_device
*device
, struct anv_bo
*bo
)
2653 if (!(bo
->flags
& EXEC_OBJECT_PINNED
))
2656 const uint64_t addr_48b
= gen_48b_address(bo
->offset
);
2658 pthread_mutex_lock(&device
->vma_mutex
);
2660 if (addr_48b
>= LOW_HEAP_MIN_ADDRESS
&&
2661 addr_48b
<= LOW_HEAP_MAX_ADDRESS
) {
2662 util_vma_heap_free(&device
->vma_lo
, addr_48b
, bo
->size
);
2663 device
->vma_lo_available
+= bo
->size
;
2665 MAYBE_UNUSED
const struct anv_physical_device
*physical_device
=
2666 &device
->instance
->physicalDevice
;
2667 assert(addr_48b
>= physical_device
->memory
.heaps
[0].vma_start
&&
2668 addr_48b
< (physical_device
->memory
.heaps
[0].vma_start
+
2669 physical_device
->memory
.heaps
[0].vma_size
));
2670 util_vma_heap_free(&device
->vma_hi
, addr_48b
, bo
->size
);
2671 device
->vma_hi_available
+= bo
->size
;
2674 pthread_mutex_unlock(&device
->vma_mutex
);
2680 anv_bo_init_new(struct anv_bo
*bo
, struct anv_device
*device
, uint64_t size
)
2682 uint32_t gem_handle
= anv_gem_create(device
, size
);
2684 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
2686 anv_bo_init(bo
, gem_handle
, size
);
2691 VkResult
anv_AllocateMemory(
2693 const VkMemoryAllocateInfo
* pAllocateInfo
,
2694 const VkAllocationCallbacks
* pAllocator
,
2695 VkDeviceMemory
* pMem
)
2697 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2698 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2699 struct anv_device_memory
*mem
;
2700 VkResult result
= VK_SUCCESS
;
2702 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
2704 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
2705 assert(pAllocateInfo
->allocationSize
> 0);
2707 if (pAllocateInfo
->allocationSize
> MAX_MEMORY_ALLOCATION_SIZE
)
2708 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
2710 /* FINISHME: Fail if allocation request exceeds heap size. */
2712 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
2713 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2715 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2717 assert(pAllocateInfo
->memoryTypeIndex
< pdevice
->memory
.type_count
);
2718 mem
->type
= &pdevice
->memory
.types
[pAllocateInfo
->memoryTypeIndex
];
2722 mem
->host_ptr
= NULL
;
2724 uint64_t bo_flags
= 0;
2726 assert(mem
->type
->heapIndex
< pdevice
->memory
.heap_count
);
2727 if (pdevice
->memory
.heaps
[mem
->type
->heapIndex
].supports_48bit_addresses
)
2728 bo_flags
|= EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
2730 const struct wsi_memory_allocate_info
*wsi_info
=
2731 vk_find_struct_const(pAllocateInfo
->pNext
, WSI_MEMORY_ALLOCATE_INFO_MESA
);
2732 if (wsi_info
&& wsi_info
->implicit_sync
) {
2733 /* We need to set the WRITE flag on window system buffers so that GEM
2734 * will know we're writing to them and synchronize uses on other rings
2735 * (eg if the display server uses the blitter ring).
2737 bo_flags
|= EXEC_OBJECT_WRITE
;
2738 } else if (pdevice
->has_exec_async
) {
2739 bo_flags
|= EXEC_OBJECT_ASYNC
;
2742 if (pdevice
->use_softpin
)
2743 bo_flags
|= EXEC_OBJECT_PINNED
;
2745 const VkExportMemoryAllocateInfo
*export_info
=
2746 vk_find_struct_const(pAllocateInfo
->pNext
, EXPORT_MEMORY_ALLOCATE_INFO
);
2748 /* Check if we need to support Android HW buffer export. If so,
2749 * create AHardwareBuffer and import memory from it.
2751 bool android_export
= false;
2752 if (export_info
&& export_info
->handleTypes
&
2753 VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID
)
2754 android_export
= true;
2756 /* Android memory import. */
2757 const struct VkImportAndroidHardwareBufferInfoANDROID
*ahw_import_info
=
2758 vk_find_struct_const(pAllocateInfo
->pNext
,
2759 IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID
);
2761 if (ahw_import_info
) {
2762 result
= anv_import_ahw_memory(_device
, mem
, ahw_import_info
);
2763 if (result
!= VK_SUCCESS
)
2767 } else if (android_export
) {
2768 result
= anv_create_ahw_memory(_device
, mem
, pAllocateInfo
);
2769 if (result
!= VK_SUCCESS
)
2772 const struct VkImportAndroidHardwareBufferInfoANDROID import_info
= {
2775 result
= anv_import_ahw_memory(_device
, mem
, &import_info
);
2776 if (result
!= VK_SUCCESS
)
2782 const VkImportMemoryFdInfoKHR
*fd_info
=
2783 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_FD_INFO_KHR
);
2785 /* The Vulkan spec permits handleType to be 0, in which case the struct is
2788 if (fd_info
&& fd_info
->handleType
) {
2789 /* At the moment, we support only the below handle types. */
2790 assert(fd_info
->handleType
==
2791 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
2792 fd_info
->handleType
==
2793 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
2795 result
= anv_bo_cache_import(device
, &device
->bo_cache
, fd_info
->fd
,
2796 bo_flags
| ANV_BO_EXTERNAL
, &mem
->bo
);
2797 if (result
!= VK_SUCCESS
)
2800 VkDeviceSize aligned_alloc_size
=
2801 align_u64(pAllocateInfo
->allocationSize
, 4096);
2803 /* For security purposes, we reject importing the bo if it's smaller
2804 * than the requested allocation size. This prevents a malicious client
2805 * from passing a buffer to a trusted client, lying about the size, and
2806 * telling the trusted client to try and texture from an image that goes
2807 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
2808 * in the trusted client. The trusted client can protect itself against
2809 * this sort of attack but only if it can trust the buffer size.
2811 if (mem
->bo
->size
< aligned_alloc_size
) {
2812 result
= vk_errorf(device
->instance
, device
,
2813 VK_ERROR_INVALID_EXTERNAL_HANDLE
,
2814 "aligned allocationSize too large for "
2815 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT: "
2816 "%"PRIu64
"B > %"PRIu64
"B",
2817 aligned_alloc_size
, mem
->bo
->size
);
2818 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
2822 /* From the Vulkan spec:
2824 * "Importing memory from a file descriptor transfers ownership of
2825 * the file descriptor from the application to the Vulkan
2826 * implementation. The application must not perform any operations on
2827 * the file descriptor after a successful import."
2829 * If the import fails, we leave the file descriptor open.
2835 const VkImportMemoryHostPointerInfoEXT
*host_ptr_info
=
2836 vk_find_struct_const(pAllocateInfo
->pNext
,
2837 IMPORT_MEMORY_HOST_POINTER_INFO_EXT
);
2838 if (host_ptr_info
&& host_ptr_info
->handleType
) {
2839 if (host_ptr_info
->handleType
==
2840 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_MAPPED_FOREIGN_MEMORY_BIT_EXT
) {
2841 result
= vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
2845 assert(host_ptr_info
->handleType
==
2846 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
);
2848 result
= anv_bo_cache_import_host_ptr(
2849 device
, &device
->bo_cache
, host_ptr_info
->pHostPointer
,
2850 pAllocateInfo
->allocationSize
, bo_flags
, &mem
->bo
);
2852 if (result
!= VK_SUCCESS
)
2855 mem
->host_ptr
= host_ptr_info
->pHostPointer
;
2859 /* Regular allocate (not importing memory). */
2861 if (export_info
&& export_info
->handleTypes
)
2862 bo_flags
|= ANV_BO_EXTERNAL
;
2864 result
= anv_bo_cache_alloc(device
, &device
->bo_cache
,
2865 pAllocateInfo
->allocationSize
, bo_flags
,
2867 if (result
!= VK_SUCCESS
)
2870 const VkMemoryDedicatedAllocateInfo
*dedicated_info
=
2871 vk_find_struct_const(pAllocateInfo
->pNext
, MEMORY_DEDICATED_ALLOCATE_INFO
);
2872 if (dedicated_info
&& dedicated_info
->image
!= VK_NULL_HANDLE
) {
2873 ANV_FROM_HANDLE(anv_image
, image
, dedicated_info
->image
);
2875 /* Some legacy (non-modifiers) consumers need the tiling to be set on
2876 * the BO. In this case, we have a dedicated allocation.
2878 if (image
->needs_set_tiling
) {
2879 const uint32_t i915_tiling
=
2880 isl_tiling_to_i915_tiling(image
->planes
[0].surface
.isl
.tiling
);
2881 int ret
= anv_gem_set_tiling(device
, mem
->bo
->gem_handle
,
2882 image
->planes
[0].surface
.isl
.row_pitch_B
,
2885 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
2886 return vk_errorf(device
->instance
, NULL
,
2887 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
2888 "failed to set BO tiling: %m");
2894 pthread_mutex_lock(&device
->mutex
);
2895 list_addtail(&mem
->link
, &device
->memory_objects
);
2896 pthread_mutex_unlock(&device
->mutex
);
2898 *pMem
= anv_device_memory_to_handle(mem
);
2900 p_atomic_add(&pdevice
->memory
.heaps
[mem
->type
->heapIndex
].used
,
2906 vk_free2(&device
->alloc
, pAllocator
, mem
);
2911 VkResult
anv_GetMemoryFdKHR(
2913 const VkMemoryGetFdInfoKHR
* pGetFdInfo
,
2916 ANV_FROM_HANDLE(anv_device
, dev
, device_h
);
2917 ANV_FROM_HANDLE(anv_device_memory
, mem
, pGetFdInfo
->memory
);
2919 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
2921 assert(pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
2922 pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
2924 return anv_bo_cache_export(dev
, &dev
->bo_cache
, mem
->bo
, pFd
);
2927 VkResult
anv_GetMemoryFdPropertiesKHR(
2929 VkExternalMemoryHandleTypeFlagBits handleType
,
2931 VkMemoryFdPropertiesKHR
* pMemoryFdProperties
)
2933 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2934 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2936 switch (handleType
) {
2937 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
:
2938 /* dma-buf can be imported as any memory type */
2939 pMemoryFdProperties
->memoryTypeBits
=
2940 (1 << pdevice
->memory
.type_count
) - 1;
2944 /* The valid usage section for this function says:
2946 * "handleType must not be one of the handle types defined as
2949 * So opaque handle types fall into the default "unsupported" case.
2951 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
2955 VkResult
anv_GetMemoryHostPointerPropertiesEXT(
2957 VkExternalMemoryHandleTypeFlagBits handleType
,
2958 const void* pHostPointer
,
2959 VkMemoryHostPointerPropertiesEXT
* pMemoryHostPointerProperties
)
2961 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2963 assert(pMemoryHostPointerProperties
->sType
==
2964 VK_STRUCTURE_TYPE_MEMORY_HOST_POINTER_PROPERTIES_EXT
);
2966 switch (handleType
) {
2967 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
: {
2968 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2970 /* Host memory can be imported as any memory type. */
2971 pMemoryHostPointerProperties
->memoryTypeBits
=
2972 (1ull << pdevice
->memory
.type_count
) - 1;
2977 return VK_ERROR_INVALID_EXTERNAL_HANDLE
;
2981 void anv_FreeMemory(
2983 VkDeviceMemory _mem
,
2984 const VkAllocationCallbacks
* pAllocator
)
2986 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2987 ANV_FROM_HANDLE(anv_device_memory
, mem
, _mem
);
2988 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2993 pthread_mutex_lock(&device
->mutex
);
2994 list_del(&mem
->link
);
2995 pthread_mutex_unlock(&device
->mutex
);
2998 anv_UnmapMemory(_device
, _mem
);
3000 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
3002 #if defined(ANDROID) && ANDROID_API_LEVEL >= 26
3004 AHardwareBuffer_release(mem
->ahw
);
3007 p_atomic_add(&pdevice
->memory
.heaps
[mem
->type
->heapIndex
].used
,
3010 vk_free2(&device
->alloc
, pAllocator
, mem
);
3013 VkResult
anv_MapMemory(
3015 VkDeviceMemory _memory
,
3016 VkDeviceSize offset
,
3018 VkMemoryMapFlags flags
,
3021 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3022 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
3029 if (mem
->host_ptr
) {
3030 *ppData
= mem
->host_ptr
+ offset
;
3034 if (size
== VK_WHOLE_SIZE
)
3035 size
= mem
->bo
->size
- offset
;
3037 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
3039 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
3040 * assert(size != 0);
3041 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
3042 * equal to the size of the memory minus offset
3045 assert(offset
+ size
<= mem
->bo
->size
);
3047 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
3048 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
3049 * at a time is valid. We could just mmap up front and return an offset
3050 * pointer here, but that may exhaust virtual memory on 32 bit
3053 uint32_t gem_flags
= 0;
3055 if (!device
->info
.has_llc
&&
3056 (mem
->type
->propertyFlags
& VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
))
3057 gem_flags
|= I915_MMAP_WC
;
3059 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
3060 uint64_t map_offset
= offset
& ~4095ull;
3061 assert(offset
>= map_offset
);
3062 uint64_t map_size
= (offset
+ size
) - map_offset
;
3064 /* Let's map whole pages */
3065 map_size
= align_u64(map_size
, 4096);
3067 void *map
= anv_gem_mmap(device
, mem
->bo
->gem_handle
,
3068 map_offset
, map_size
, gem_flags
);
3069 if (map
== MAP_FAILED
)
3070 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
3073 mem
->map_size
= map_size
;
3075 *ppData
= mem
->map
+ (offset
- map_offset
);
3080 void anv_UnmapMemory(
3082 VkDeviceMemory _memory
)
3084 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
3086 if (mem
== NULL
|| mem
->host_ptr
)
3089 anv_gem_munmap(mem
->map
, mem
->map_size
);
3096 clflush_mapped_ranges(struct anv_device
*device
,
3098 const VkMappedMemoryRange
*ranges
)
3100 for (uint32_t i
= 0; i
< count
; i
++) {
3101 ANV_FROM_HANDLE(anv_device_memory
, mem
, ranges
[i
].memory
);
3102 if (ranges
[i
].offset
>= mem
->map_size
)
3105 gen_clflush_range(mem
->map
+ ranges
[i
].offset
,
3106 MIN2(ranges
[i
].size
, mem
->map_size
- ranges
[i
].offset
));
3110 VkResult
anv_FlushMappedMemoryRanges(
3112 uint32_t memoryRangeCount
,
3113 const VkMappedMemoryRange
* pMemoryRanges
)
3115 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3117 if (device
->info
.has_llc
)
3120 /* Make sure the writes we're flushing have landed. */
3121 __builtin_ia32_mfence();
3123 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
3128 VkResult
anv_InvalidateMappedMemoryRanges(
3130 uint32_t memoryRangeCount
,
3131 const VkMappedMemoryRange
* pMemoryRanges
)
3133 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3135 if (device
->info
.has_llc
)
3138 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
3140 /* Make sure no reads get moved up above the invalidate. */
3141 __builtin_ia32_mfence();
3146 void anv_GetBufferMemoryRequirements(
3149 VkMemoryRequirements
* pMemoryRequirements
)
3151 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
3152 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3153 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3155 /* The Vulkan spec (git aaed022) says:
3157 * memoryTypeBits is a bitfield and contains one bit set for every
3158 * supported memory type for the resource. The bit `1<<i` is set if and
3159 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3160 * structure for the physical device is supported.
3162 uint32_t memory_types
= 0;
3163 for (uint32_t i
= 0; i
< pdevice
->memory
.type_count
; i
++) {
3164 uint32_t valid_usage
= pdevice
->memory
.types
[i
].valid_buffer_usage
;
3165 if ((valid_usage
& buffer
->usage
) == buffer
->usage
)
3166 memory_types
|= (1u << i
);
3169 /* Base alignment requirement of a cache line */
3170 uint32_t alignment
= 16;
3172 /* We need an alignment of 32 for pushing UBOs */
3173 if (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
)
3174 alignment
= MAX2(alignment
, 32);
3176 pMemoryRequirements
->size
= buffer
->size
;
3177 pMemoryRequirements
->alignment
= alignment
;
3179 /* Storage and Uniform buffers should have their size aligned to
3180 * 32-bits to avoid boundary checks when last DWord is not complete.
3181 * This would ensure that not internal padding would be needed for
3184 if (device
->robust_buffer_access
&&
3185 (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
||
3186 buffer
->usage
& VK_BUFFER_USAGE_STORAGE_BUFFER_BIT
))
3187 pMemoryRequirements
->size
= align_u64(buffer
->size
, 4);
3189 pMemoryRequirements
->memoryTypeBits
= memory_types
;
3192 void anv_GetBufferMemoryRequirements2(
3194 const VkBufferMemoryRequirementsInfo2
* pInfo
,
3195 VkMemoryRequirements2
* pMemoryRequirements
)
3197 anv_GetBufferMemoryRequirements(_device
, pInfo
->buffer
,
3198 &pMemoryRequirements
->memoryRequirements
);
3200 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3201 switch (ext
->sType
) {
3202 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
3203 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
3204 requirements
->prefersDedicatedAllocation
= false;
3205 requirements
->requiresDedicatedAllocation
= false;
3210 anv_debug_ignored_stype(ext
->sType
);
3216 void anv_GetImageMemoryRequirements(
3219 VkMemoryRequirements
* pMemoryRequirements
)
3221 ANV_FROM_HANDLE(anv_image
, image
, _image
);
3222 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3223 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3225 /* The Vulkan spec (git aaed022) says:
3227 * memoryTypeBits is a bitfield and contains one bit set for every
3228 * supported memory type for the resource. The bit `1<<i` is set if and
3229 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3230 * structure for the physical device is supported.
3232 * All types are currently supported for images.
3234 uint32_t memory_types
= (1ull << pdevice
->memory
.type_count
) - 1;
3236 /* We must have image allocated or imported at this point. According to the
3237 * specification, external images must have been bound to memory before
3238 * calling GetImageMemoryRequirements.
3240 assert(image
->size
> 0);
3242 pMemoryRequirements
->size
= image
->size
;
3243 pMemoryRequirements
->alignment
= image
->alignment
;
3244 pMemoryRequirements
->memoryTypeBits
= memory_types
;
3247 void anv_GetImageMemoryRequirements2(
3249 const VkImageMemoryRequirementsInfo2
* pInfo
,
3250 VkMemoryRequirements2
* pMemoryRequirements
)
3252 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3253 ANV_FROM_HANDLE(anv_image
, image
, pInfo
->image
);
3255 anv_GetImageMemoryRequirements(_device
, pInfo
->image
,
3256 &pMemoryRequirements
->memoryRequirements
);
3258 vk_foreach_struct_const(ext
, pInfo
->pNext
) {
3259 switch (ext
->sType
) {
3260 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO
: {
3261 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3262 const VkImagePlaneMemoryRequirementsInfo
*plane_reqs
=
3263 (const VkImagePlaneMemoryRequirementsInfo
*) ext
;
3264 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
,
3265 plane_reqs
->planeAspect
);
3267 assert(image
->planes
[plane
].offset
== 0);
3269 /* The Vulkan spec (git aaed022) says:
3271 * memoryTypeBits is a bitfield and contains one bit set for every
3272 * supported memory type for the resource. The bit `1<<i` is set
3273 * if and only if the memory type `i` in the
3274 * VkPhysicalDeviceMemoryProperties structure for the physical
3275 * device is supported.
3277 * All types are currently supported for images.
3279 pMemoryRequirements
->memoryRequirements
.memoryTypeBits
=
3280 (1ull << pdevice
->memory
.type_count
) - 1;
3282 /* We must have image allocated or imported at this point. According to the
3283 * specification, external images must have been bound to memory before
3284 * calling GetImageMemoryRequirements.
3286 assert(image
->planes
[plane
].size
> 0);
3288 pMemoryRequirements
->memoryRequirements
.size
= image
->planes
[plane
].size
;
3289 pMemoryRequirements
->memoryRequirements
.alignment
=
3290 image
->planes
[plane
].alignment
;
3295 anv_debug_ignored_stype(ext
->sType
);
3300 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3301 switch (ext
->sType
) {
3302 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
3303 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
3304 if (image
->needs_set_tiling
|| image
->external_format
) {
3305 /* If we need to set the tiling for external consumers, we need a
3306 * dedicated allocation.
3308 * See also anv_AllocateMemory.
3310 requirements
->prefersDedicatedAllocation
= true;
3311 requirements
->requiresDedicatedAllocation
= true;
3313 requirements
->prefersDedicatedAllocation
= false;
3314 requirements
->requiresDedicatedAllocation
= false;
3320 anv_debug_ignored_stype(ext
->sType
);
3326 void anv_GetImageSparseMemoryRequirements(
3329 uint32_t* pSparseMemoryRequirementCount
,
3330 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
3332 *pSparseMemoryRequirementCount
= 0;
3335 void anv_GetImageSparseMemoryRequirements2(
3337 const VkImageSparseMemoryRequirementsInfo2
* pInfo
,
3338 uint32_t* pSparseMemoryRequirementCount
,
3339 VkSparseImageMemoryRequirements2
* pSparseMemoryRequirements
)
3341 *pSparseMemoryRequirementCount
= 0;
3344 void anv_GetDeviceMemoryCommitment(
3346 VkDeviceMemory memory
,
3347 VkDeviceSize
* pCommittedMemoryInBytes
)
3349 *pCommittedMemoryInBytes
= 0;
3353 anv_bind_buffer_memory(const VkBindBufferMemoryInfo
*pBindInfo
)
3355 ANV_FROM_HANDLE(anv_device_memory
, mem
, pBindInfo
->memory
);
3356 ANV_FROM_HANDLE(anv_buffer
, buffer
, pBindInfo
->buffer
);
3358 assert(pBindInfo
->sType
== VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
);
3361 assert((buffer
->usage
& mem
->type
->valid_buffer_usage
) == buffer
->usage
);
3362 buffer
->address
= (struct anv_address
) {
3364 .offset
= pBindInfo
->memoryOffset
,
3367 buffer
->address
= ANV_NULL_ADDRESS
;
3371 VkResult
anv_BindBufferMemory(
3374 VkDeviceMemory memory
,
3375 VkDeviceSize memoryOffset
)
3377 anv_bind_buffer_memory(
3378 &(VkBindBufferMemoryInfo
) {
3379 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
,
3382 .memoryOffset
= memoryOffset
,
3388 VkResult
anv_BindBufferMemory2(
3390 uint32_t bindInfoCount
,
3391 const VkBindBufferMemoryInfo
* pBindInfos
)
3393 for (uint32_t i
= 0; i
< bindInfoCount
; i
++)
3394 anv_bind_buffer_memory(&pBindInfos
[i
]);
3399 VkResult
anv_QueueBindSparse(
3401 uint32_t bindInfoCount
,
3402 const VkBindSparseInfo
* pBindInfo
,
3405 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
3406 if (anv_device_is_lost(queue
->device
))
3407 return VK_ERROR_DEVICE_LOST
;
3409 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
3414 VkResult
anv_CreateEvent(
3416 const VkEventCreateInfo
* pCreateInfo
,
3417 const VkAllocationCallbacks
* pAllocator
,
3420 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3421 struct anv_state state
;
3422 struct anv_event
*event
;
3424 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_EVENT_CREATE_INFO
);
3426 state
= anv_state_pool_alloc(&device
->dynamic_state_pool
,
3429 event
->state
= state
;
3430 event
->semaphore
= VK_EVENT_RESET
;
3432 if (!device
->info
.has_llc
) {
3433 /* Make sure the writes we're flushing have landed. */
3434 __builtin_ia32_mfence();
3435 __builtin_ia32_clflush(event
);
3438 *pEvent
= anv_event_to_handle(event
);
3443 void anv_DestroyEvent(
3446 const VkAllocationCallbacks
* pAllocator
)
3448 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3449 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3454 anv_state_pool_free(&device
->dynamic_state_pool
, event
->state
);
3457 VkResult
anv_GetEventStatus(
3461 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3462 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3464 if (anv_device_is_lost(device
))
3465 return VK_ERROR_DEVICE_LOST
;
3467 if (!device
->info
.has_llc
) {
3468 /* Invalidate read cache before reading event written by GPU. */
3469 __builtin_ia32_clflush(event
);
3470 __builtin_ia32_mfence();
3474 return event
->semaphore
;
3477 VkResult
anv_SetEvent(
3481 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3482 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3484 event
->semaphore
= VK_EVENT_SET
;
3486 if (!device
->info
.has_llc
) {
3487 /* Make sure the writes we're flushing have landed. */
3488 __builtin_ia32_mfence();
3489 __builtin_ia32_clflush(event
);
3495 VkResult
anv_ResetEvent(
3499 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3500 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3502 event
->semaphore
= VK_EVENT_RESET
;
3504 if (!device
->info
.has_llc
) {
3505 /* Make sure the writes we're flushing have landed. */
3506 __builtin_ia32_mfence();
3507 __builtin_ia32_clflush(event
);
3515 VkResult
anv_CreateBuffer(
3517 const VkBufferCreateInfo
* pCreateInfo
,
3518 const VkAllocationCallbacks
* pAllocator
,
3521 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3522 struct anv_buffer
*buffer
;
3524 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
3526 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
3527 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3529 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3531 buffer
->size
= pCreateInfo
->size
;
3532 buffer
->usage
= pCreateInfo
->usage
;
3533 buffer
->address
= ANV_NULL_ADDRESS
;
3535 *pBuffer
= anv_buffer_to_handle(buffer
);
3540 void anv_DestroyBuffer(
3543 const VkAllocationCallbacks
* pAllocator
)
3545 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3546 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
3551 vk_free2(&device
->alloc
, pAllocator
, buffer
);
3554 VkDeviceAddress
anv_GetBufferDeviceAddressEXT(
3556 const VkBufferDeviceAddressInfoEXT
* pInfo
)
3558 ANV_FROM_HANDLE(anv_buffer
, buffer
, pInfo
->buffer
);
3560 assert(buffer
->address
.bo
->flags
& EXEC_OBJECT_PINNED
);
3562 return anv_address_physical(buffer
->address
);
3566 anv_fill_buffer_surface_state(struct anv_device
*device
, struct anv_state state
,
3567 enum isl_format format
,
3568 struct anv_address address
,
3569 uint32_t range
, uint32_t stride
)
3571 isl_buffer_fill_state(&device
->isl_dev
, state
.map
,
3572 .address
= anv_address_physical(address
),
3573 .mocs
= device
->default_mocs
,
3576 .swizzle
= ISL_SWIZZLE_IDENTITY
,
3577 .stride_B
= stride
);
3580 void anv_DestroySampler(
3583 const VkAllocationCallbacks
* pAllocator
)
3585 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3586 ANV_FROM_HANDLE(anv_sampler
, sampler
, _sampler
);
3591 if (sampler
->bindless_state
.map
) {
3592 anv_state_pool_free(&device
->dynamic_state_pool
,
3593 sampler
->bindless_state
);
3596 vk_free2(&device
->alloc
, pAllocator
, sampler
);
3599 VkResult
anv_CreateFramebuffer(
3601 const VkFramebufferCreateInfo
* pCreateInfo
,
3602 const VkAllocationCallbacks
* pAllocator
,
3603 VkFramebuffer
* pFramebuffer
)
3605 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3606 struct anv_framebuffer
*framebuffer
;
3608 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
3610 size_t size
= sizeof(*framebuffer
) +
3611 sizeof(struct anv_image_view
*) * pCreateInfo
->attachmentCount
;
3612 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
3613 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3614 if (framebuffer
== NULL
)
3615 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3617 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
3618 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
3619 VkImageView _iview
= pCreateInfo
->pAttachments
[i
];
3620 framebuffer
->attachments
[i
] = anv_image_view_from_handle(_iview
);
3623 framebuffer
->width
= pCreateInfo
->width
;
3624 framebuffer
->height
= pCreateInfo
->height
;
3625 framebuffer
->layers
= pCreateInfo
->layers
;
3627 *pFramebuffer
= anv_framebuffer_to_handle(framebuffer
);
3632 void anv_DestroyFramebuffer(
3635 const VkAllocationCallbacks
* pAllocator
)
3637 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3638 ANV_FROM_HANDLE(anv_framebuffer
, fb
, _fb
);
3643 vk_free2(&device
->alloc
, pAllocator
, fb
);
3646 static const VkTimeDomainEXT anv_time_domains
[] = {
3647 VK_TIME_DOMAIN_DEVICE_EXT
,
3648 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
,
3649 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
,
3652 VkResult
anv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
3653 VkPhysicalDevice physicalDevice
,
3654 uint32_t *pTimeDomainCount
,
3655 VkTimeDomainEXT
*pTimeDomains
)
3658 VK_OUTARRAY_MAKE(out
, pTimeDomains
, pTimeDomainCount
);
3660 for (d
= 0; d
< ARRAY_SIZE(anv_time_domains
); d
++) {
3661 vk_outarray_append(&out
, i
) {
3662 *i
= anv_time_domains
[d
];
3666 return vk_outarray_status(&out
);
3670 anv_clock_gettime(clockid_t clock_id
)
3672 struct timespec current
;
3675 ret
= clock_gettime(clock_id
, ¤t
);
3676 if (ret
< 0 && clock_id
== CLOCK_MONOTONIC_RAW
)
3677 ret
= clock_gettime(CLOCK_MONOTONIC
, ¤t
);
3681 return (uint64_t) current
.tv_sec
* 1000000000ULL + current
.tv_nsec
;
3684 #define TIMESTAMP 0x2358
3686 VkResult
anv_GetCalibratedTimestampsEXT(
3688 uint32_t timestampCount
,
3689 const VkCalibratedTimestampInfoEXT
*pTimestampInfos
,
3690 uint64_t *pTimestamps
,
3691 uint64_t *pMaxDeviation
)
3693 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3694 uint64_t timestamp_frequency
= device
->info
.timestamp_frequency
;
3697 uint64_t begin
, end
;
3698 uint64_t max_clock_period
= 0;
3700 begin
= anv_clock_gettime(CLOCK_MONOTONIC_RAW
);
3702 for (d
= 0; d
< timestampCount
; d
++) {
3703 switch (pTimestampInfos
[d
].timeDomain
) {
3704 case VK_TIME_DOMAIN_DEVICE_EXT
:
3705 ret
= anv_gem_reg_read(device
, TIMESTAMP
| 1,
3709 return anv_device_set_lost(device
, "Failed to read the TIMESTAMP "
3712 uint64_t device_period
= DIV_ROUND_UP(1000000000, timestamp_frequency
);
3713 max_clock_period
= MAX2(max_clock_period
, device_period
);
3715 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
:
3716 pTimestamps
[d
] = anv_clock_gettime(CLOCK_MONOTONIC
);
3717 max_clock_period
= MAX2(max_clock_period
, 1);
3720 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
:
3721 pTimestamps
[d
] = begin
;
3729 end
= anv_clock_gettime(CLOCK_MONOTONIC_RAW
);
3732 * The maximum deviation is the sum of the interval over which we
3733 * perform the sampling and the maximum period of any sampled
3734 * clock. That's because the maximum skew between any two sampled
3735 * clock edges is when the sampled clock with the largest period is
3736 * sampled at the end of that period but right at the beginning of the
3737 * sampling interval and some other clock is sampled right at the
3738 * begining of its sampling period and right at the end of the
3739 * sampling interval. Let's assume the GPU has the longest clock
3740 * period and that the application is sampling GPU and monotonic:
3743 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
3744 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
3748 * GPU -----_____-----_____-----_____-----_____
3751 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
3752 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
3754 * Interval <----------------->
3755 * Deviation <-------------------------->
3759 * m = read(monotonic) 2
3762 * We round the sample interval up by one tick to cover sampling error
3763 * in the interval clock
3766 uint64_t sample_interval
= end
- begin
+ 1;
3768 *pMaxDeviation
= sample_interval
+ max_clock_period
;
3773 /* vk_icd.h does not declare this function, so we declare it here to
3774 * suppress Wmissing-prototypes.
3776 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
3777 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
);
3779 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
3780 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
)
3782 /* For the full details on loader interface versioning, see
3783 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
3784 * What follows is a condensed summary, to help you navigate the large and
3785 * confusing official doc.
3787 * - Loader interface v0 is incompatible with later versions. We don't
3790 * - In loader interface v1:
3791 * - The first ICD entrypoint called by the loader is
3792 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
3794 * - The ICD must statically expose no other Vulkan symbol unless it is
3795 * linked with -Bsymbolic.
3796 * - Each dispatchable Vulkan handle created by the ICD must be
3797 * a pointer to a struct whose first member is VK_LOADER_DATA. The
3798 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
3799 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
3800 * vkDestroySurfaceKHR(). The ICD must be capable of working with
3801 * such loader-managed surfaces.
3803 * - Loader interface v2 differs from v1 in:
3804 * - The first ICD entrypoint called by the loader is
3805 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
3806 * statically expose this entrypoint.
3808 * - Loader interface v3 differs from v2 in:
3809 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
3810 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
3811 * because the loader no longer does so.
3813 *pSupportedVersion
= MIN2(*pSupportedVersion
, 3u);