2 * Copyright © 2015 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
28 #include <sys/sysinfo.h>
32 #include "drm-uapi/drm_fourcc.h"
34 #include "anv_private.h"
35 #include "util/strtod.h"
36 #include "util/debug.h"
37 #include "util/build_id.h"
38 #include "util/disk_cache.h"
39 #include "util/mesa-sha1.h"
40 #include "util/u_string.h"
43 #include "common/gen_defines.h"
44 #include "compiler/glsl_types.h"
46 #include "genxml/gen7_pack.h"
48 /* This is probably far to big but it reflects the max size used for messages
49 * in OpenGLs KHR_debug.
51 #define MAX_DEBUG_MESSAGE_LENGTH 4096
54 compiler_debug_log(void *data
, const char *fmt
, ...)
56 char str
[MAX_DEBUG_MESSAGE_LENGTH
];
57 struct anv_device
*device
= (struct anv_device
*)data
;
59 if (list_empty(&device
->instance
->debug_report_callbacks
.callbacks
))
64 (void) vsnprintf(str
, MAX_DEBUG_MESSAGE_LENGTH
, fmt
, args
);
67 vk_debug_report(&device
->instance
->debug_report_callbacks
,
68 VK_DEBUG_REPORT_DEBUG_BIT_EXT
,
69 VK_DEBUG_REPORT_OBJECT_TYPE_UNKNOWN_EXT
,
74 compiler_perf_log(void *data
, const char *fmt
, ...)
79 if (unlikely(INTEL_DEBUG
& DEBUG_PERF
))
80 intel_logd_v(fmt
, args
);
86 anv_compute_heap_size(int fd
, uint64_t gtt_size
)
88 /* Query the total ram from the system */
92 uint64_t total_ram
= (uint64_t)info
.totalram
* (uint64_t)info
.mem_unit
;
94 /* We don't want to burn too much ram with the GPU. If the user has 4GiB
95 * or less, we use at most half. If they have more than 4GiB, we use 3/4.
97 uint64_t available_ram
;
98 if (total_ram
<= 4ull * 1024ull * 1024ull * 1024ull)
99 available_ram
= total_ram
/ 2;
101 available_ram
= total_ram
* 3 / 4;
103 /* We also want to leave some padding for things we allocate in the driver,
104 * so don't go over 3/4 of the GTT either.
106 uint64_t available_gtt
= gtt_size
* 3 / 4;
108 return MIN2(available_ram
, available_gtt
);
112 anv_physical_device_init_heaps(struct anv_physical_device
*device
, int fd
)
115 if (anv_gem_get_context_param(fd
, 0, I915_CONTEXT_PARAM_GTT_SIZE
,
117 /* If, for whatever reason, we can't actually get the GTT size from the
118 * kernel (too old?) fall back to the aperture size.
120 anv_perf_warn(NULL
, NULL
,
121 "Failed to get I915_CONTEXT_PARAM_GTT_SIZE: %m");
123 if (anv_gem_get_aperture(fd
, >t_size
) == -1) {
124 return vk_errorf(NULL
, NULL
, VK_ERROR_INITIALIZATION_FAILED
,
125 "failed to get aperture size: %m");
129 device
->supports_48bit_addresses
= (device
->info
.gen
>= 8) &&
130 gtt_size
> (4ULL << 30 /* GiB */);
132 uint64_t heap_size
= anv_compute_heap_size(fd
, gtt_size
);
134 if (heap_size
> (2ull << 30) && !device
->supports_48bit_addresses
) {
135 /* When running with an overridden PCI ID, we may get a GTT size from
136 * the kernel that is greater than 2 GiB but the execbuf check for 48bit
137 * address support can still fail. Just clamp the address space size to
138 * 2 GiB if we don't have 48-bit support.
140 intel_logw("%s:%d: The kernel reported a GTT size larger than 2 GiB but "
141 "not support for 48-bit addresses",
143 heap_size
= 2ull << 30;
146 if (heap_size
<= 3ull * (1ull << 30)) {
147 /* In this case, everything fits nicely into the 32-bit address space,
148 * so there's no need for supporting 48bit addresses on client-allocated
151 device
->memory
.heap_count
= 1;
152 device
->memory
.heaps
[0] = (struct anv_memory_heap
) {
153 .vma_start
= LOW_HEAP_MIN_ADDRESS
,
154 .vma_size
= LOW_HEAP_SIZE
,
156 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
157 .supports_48bit_addresses
= false,
160 /* Not everything will fit nicely into a 32-bit address space. In this
161 * case we need a 64-bit heap. Advertise a small 32-bit heap and a
162 * larger 48-bit heap. If we're in this case, then we have a total heap
163 * size larger than 3GiB which most likely means they have 8 GiB of
164 * video memory and so carving off 1 GiB for the 32-bit heap should be
167 const uint64_t heap_size_32bit
= 1ull << 30;
168 const uint64_t heap_size_48bit
= heap_size
- heap_size_32bit
;
170 assert(device
->supports_48bit_addresses
);
172 device
->memory
.heap_count
= 2;
173 device
->memory
.heaps
[0] = (struct anv_memory_heap
) {
174 .vma_start
= HIGH_HEAP_MIN_ADDRESS
,
175 /* Leave the last 4GiB out of the high vma range, so that no state
176 * base address + size can overflow 48 bits. For more information see
177 * the comment about Wa32bitGeneralStateOffset in anv_allocator.c
179 .vma_size
= gtt_size
- (1ull << 32) - HIGH_HEAP_MIN_ADDRESS
,
180 .size
= heap_size_48bit
,
181 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
182 .supports_48bit_addresses
= true,
184 device
->memory
.heaps
[1] = (struct anv_memory_heap
) {
185 .vma_start
= LOW_HEAP_MIN_ADDRESS
,
186 .vma_size
= LOW_HEAP_SIZE
,
187 .size
= heap_size_32bit
,
188 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
189 .supports_48bit_addresses
= false,
193 uint32_t type_count
= 0;
194 for (uint32_t heap
= 0; heap
< device
->memory
.heap_count
; heap
++) {
195 uint32_t valid_buffer_usage
= ~0;
197 /* There appears to be a hardware issue in the VF cache where it only
198 * considers the bottom 32 bits of memory addresses. If you happen to
199 * have two vertex buffers which get placed exactly 4 GiB apart and use
200 * them in back-to-back draw calls, you can get collisions. In order to
201 * solve this problem, we require vertex and index buffers be bound to
202 * memory allocated out of the 32-bit heap.
204 if (device
->memory
.heaps
[heap
].supports_48bit_addresses
) {
205 valid_buffer_usage
&= ~(VK_BUFFER_USAGE_INDEX_BUFFER_BIT
|
206 VK_BUFFER_USAGE_VERTEX_BUFFER_BIT
);
209 if (device
->info
.has_llc
) {
210 /* Big core GPUs share LLC with the CPU and thus one memory type can be
211 * both cached and coherent at the same time.
213 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
214 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
215 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
216 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
217 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
219 .valid_buffer_usage
= valid_buffer_usage
,
222 /* The spec requires that we expose a host-visible, coherent memory
223 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
224 * to give the application a choice between cached, but not coherent and
225 * coherent but uncached (WC though).
227 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
228 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
229 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
230 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
232 .valid_buffer_usage
= valid_buffer_usage
,
234 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
235 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
236 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
237 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
239 .valid_buffer_usage
= valid_buffer_usage
,
243 device
->memory
.type_count
= type_count
;
249 anv_physical_device_init_uuids(struct anv_physical_device
*device
)
251 const struct build_id_note
*note
=
252 build_id_find_nhdr_for_addr(anv_physical_device_init_uuids
);
254 return vk_errorf(device
->instance
, device
,
255 VK_ERROR_INITIALIZATION_FAILED
,
256 "Failed to find build-id");
259 unsigned build_id_len
= build_id_length(note
);
260 if (build_id_len
< 20) {
261 return vk_errorf(device
->instance
, device
,
262 VK_ERROR_INITIALIZATION_FAILED
,
263 "build-id too short. It needs to be a SHA");
266 memcpy(device
->driver_build_sha1
, build_id_data(note
), 20);
268 struct mesa_sha1 sha1_ctx
;
270 STATIC_ASSERT(VK_UUID_SIZE
<= sizeof(sha1
));
272 /* The pipeline cache UUID is used for determining when a pipeline cache is
273 * invalid. It needs both a driver build and the PCI ID of the device.
275 _mesa_sha1_init(&sha1_ctx
);
276 _mesa_sha1_update(&sha1_ctx
, build_id_data(note
), build_id_len
);
277 _mesa_sha1_update(&sha1_ctx
, &device
->chipset_id
,
278 sizeof(device
->chipset_id
));
279 _mesa_sha1_final(&sha1_ctx
, sha1
);
280 memcpy(device
->pipeline_cache_uuid
, sha1
, VK_UUID_SIZE
);
282 /* The driver UUID is used for determining sharability of images and memory
283 * between two Vulkan instances in separate processes. People who want to
284 * share memory need to also check the device UUID (below) so all this
285 * needs to be is the build-id.
287 memcpy(device
->driver_uuid
, build_id_data(note
), VK_UUID_SIZE
);
289 /* The device UUID uniquely identifies the given device within the machine.
290 * Since we never have more than one device, this doesn't need to be a real
291 * UUID. However, on the off-chance that someone tries to use this to
292 * cache pre-tiled images or something of the like, we use the PCI ID and
293 * some bits of ISL info to ensure that this is safe.
295 _mesa_sha1_init(&sha1_ctx
);
296 _mesa_sha1_update(&sha1_ctx
, &device
->chipset_id
,
297 sizeof(device
->chipset_id
));
298 _mesa_sha1_update(&sha1_ctx
, &device
->isl_dev
.has_bit6_swizzling
,
299 sizeof(device
->isl_dev
.has_bit6_swizzling
));
300 _mesa_sha1_final(&sha1_ctx
, sha1
);
301 memcpy(device
->device_uuid
, sha1
, VK_UUID_SIZE
);
307 anv_physical_device_init_disk_cache(struct anv_physical_device
*device
)
309 #ifdef ENABLE_SHADER_CACHE
311 MAYBE_UNUSED
int len
= snprintf(renderer
, sizeof(renderer
), "anv_%04x",
313 assert(len
== sizeof(renderer
) - 2);
316 _mesa_sha1_format(timestamp
, device
->driver_build_sha1
);
318 const uint64_t driver_flags
=
319 brw_get_compiler_config_value(device
->compiler
);
320 device
->disk_cache
= disk_cache_create(renderer
, timestamp
, driver_flags
);
322 device
->disk_cache
= NULL
;
327 anv_physical_device_free_disk_cache(struct anv_physical_device
*device
)
329 #ifdef ENABLE_SHADER_CACHE
330 if (device
->disk_cache
)
331 disk_cache_destroy(device
->disk_cache
);
333 assert(device
->disk_cache
== NULL
);
338 anv_physical_device_init(struct anv_physical_device
*device
,
339 struct anv_instance
*instance
,
340 drmDevicePtr drm_device
)
342 const char *primary_path
= drm_device
->nodes
[DRM_NODE_PRIMARY
];
343 const char *path
= drm_device
->nodes
[DRM_NODE_RENDER
];
348 brw_process_intel_debug_variable();
350 fd
= open(path
, O_RDWR
| O_CLOEXEC
);
352 return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
354 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
355 device
->instance
= instance
;
357 assert(strlen(path
) < ARRAY_SIZE(device
->path
));
358 snprintf(device
->path
, ARRAY_SIZE(device
->path
), "%s", path
);
360 device
->no_hw
= getenv("INTEL_NO_HW") != NULL
;
362 const int pci_id_override
= gen_get_pci_device_id_override();
363 if (pci_id_override
< 0) {
364 device
->chipset_id
= anv_gem_get_param(fd
, I915_PARAM_CHIPSET_ID
);
365 if (!device
->chipset_id
) {
366 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
370 device
->chipset_id
= pci_id_override
;
371 device
->no_hw
= true;
374 device
->pci_info
.domain
= drm_device
->businfo
.pci
->domain
;
375 device
->pci_info
.bus
= drm_device
->businfo
.pci
->bus
;
376 device
->pci_info
.device
= drm_device
->businfo
.pci
->dev
;
377 device
->pci_info
.function
= drm_device
->businfo
.pci
->func
;
379 device
->name
= gen_get_device_name(device
->chipset_id
);
380 if (!gen_get_device_info(device
->chipset_id
, &device
->info
)) {
381 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
385 if (device
->info
.is_haswell
) {
386 intel_logw("Haswell Vulkan support is incomplete");
387 } else if (device
->info
.gen
== 7 && !device
->info
.is_baytrail
) {
388 intel_logw("Ivy Bridge Vulkan support is incomplete");
389 } else if (device
->info
.gen
== 7 && device
->info
.is_baytrail
) {
390 intel_logw("Bay Trail Vulkan support is incomplete");
391 } else if (device
->info
.gen
>= 8 && device
->info
.gen
<= 10) {
392 /* Gen8-10 fully supported */
393 } else if (device
->info
.gen
== 11) {
394 intel_logw("Vulkan is not yet fully supported on gen11.");
396 result
= vk_errorf(device
->instance
, device
,
397 VK_ERROR_INCOMPATIBLE_DRIVER
,
398 "Vulkan not yet supported on %s", device
->name
);
402 device
->cmd_parser_version
= -1;
403 if (device
->info
.gen
== 7) {
404 device
->cmd_parser_version
=
405 anv_gem_get_param(fd
, I915_PARAM_CMD_PARSER_VERSION
);
406 if (device
->cmd_parser_version
== -1) {
407 result
= vk_errorf(device
->instance
, device
,
408 VK_ERROR_INITIALIZATION_FAILED
,
409 "failed to get command parser version");
414 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_WAIT_TIMEOUT
)) {
415 result
= vk_errorf(device
->instance
, device
,
416 VK_ERROR_INITIALIZATION_FAILED
,
417 "kernel missing gem wait");
421 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_EXECBUF2
)) {
422 result
= vk_errorf(device
->instance
, device
,
423 VK_ERROR_INITIALIZATION_FAILED
,
424 "kernel missing execbuf2");
428 if (!device
->info
.has_llc
&&
429 anv_gem_get_param(fd
, I915_PARAM_MMAP_VERSION
) < 1) {
430 result
= vk_errorf(device
->instance
, device
,
431 VK_ERROR_INITIALIZATION_FAILED
,
432 "kernel missing wc mmap");
436 result
= anv_physical_device_init_heaps(device
, fd
);
437 if (result
!= VK_SUCCESS
)
440 device
->has_exec_async
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_ASYNC
);
441 device
->has_exec_capture
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_CAPTURE
);
442 device
->has_exec_fence
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE
);
443 device
->has_syncobj
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE_ARRAY
);
444 device
->has_syncobj_wait
= device
->has_syncobj
&&
445 anv_gem_supports_syncobj_wait(fd
);
446 device
->has_context_priority
= anv_gem_has_context_priority(fd
);
448 device
->use_softpin
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_SOFTPIN
)
449 && device
->supports_48bit_addresses
;
451 device
->has_context_isolation
=
452 anv_gem_get_param(fd
, I915_PARAM_HAS_CONTEXT_ISOLATION
);
454 /* Starting with Gen10, the timestamp frequency of the command streamer may
455 * vary from one part to another. We can query the value from the kernel.
457 if (device
->info
.gen
>= 10) {
458 int timestamp_frequency
=
459 anv_gem_get_param(fd
, I915_PARAM_CS_TIMESTAMP_FREQUENCY
);
461 if (timestamp_frequency
< 0)
462 intel_logw("Kernel 4.16-rc1+ required to properly query CS timestamp frequency");
464 device
->info
.timestamp_frequency
= timestamp_frequency
;
467 /* GENs prior to 8 do not support EU/Subslice info */
468 if (device
->info
.gen
>= 8) {
469 device
->subslice_total
= anv_gem_get_param(fd
, I915_PARAM_SUBSLICE_TOTAL
);
470 device
->eu_total
= anv_gem_get_param(fd
, I915_PARAM_EU_TOTAL
);
472 /* Without this information, we cannot get the right Braswell
473 * brandstrings, and we have to use conservative numbers for GPGPU on
474 * many platforms, but otherwise, things will just work.
476 if (device
->subslice_total
< 1 || device
->eu_total
< 1) {
477 intel_logw("Kernel 4.1 required to properly query GPU properties");
479 } else if (device
->info
.gen
== 7) {
480 device
->subslice_total
= 1 << (device
->info
.gt
- 1);
483 if (device
->info
.is_cherryview
&&
484 device
->subslice_total
> 0 && device
->eu_total
> 0) {
485 /* Logical CS threads = EUs per subslice * num threads per EU */
486 uint32_t max_cs_threads
=
487 device
->eu_total
/ device
->subslice_total
* device
->info
.num_thread_per_eu
;
489 /* Fuse configurations may give more threads than expected, never less. */
490 if (max_cs_threads
> device
->info
.max_cs_threads
)
491 device
->info
.max_cs_threads
= max_cs_threads
;
494 device
->compiler
= brw_compiler_create(NULL
, &device
->info
);
495 if (device
->compiler
== NULL
) {
496 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
499 device
->compiler
->shader_debug_log
= compiler_debug_log
;
500 device
->compiler
->shader_perf_log
= compiler_perf_log
;
501 device
->compiler
->supports_pull_constants
= false;
502 device
->compiler
->constant_buffer_0_is_relative
=
503 device
->info
.gen
< 8 || !device
->has_context_isolation
;
504 device
->compiler
->supports_shader_constants
= true;
506 /* Broadwell PRM says:
508 * "Before Gen8, there was a historical configuration control field to
509 * swizzle address bit[6] for in X/Y tiling modes. This was set in three
510 * different places: TILECTL[1:0], ARB_MODE[5:4], and
511 * DISP_ARB_CTL[14:13].
513 * For Gen8 and subsequent generations, the swizzle fields are all
514 * reserved, and the CPU's memory controller performs all address
515 * swizzling modifications."
518 device
->info
.gen
< 8 && anv_gem_get_bit6_swizzle(fd
, I915_TILING_X
);
520 isl_device_init(&device
->isl_dev
, &device
->info
, swizzled
);
522 result
= anv_physical_device_init_uuids(device
);
523 if (result
!= VK_SUCCESS
)
526 anv_physical_device_init_disk_cache(device
);
528 if (instance
->enabled_extensions
.KHR_display
) {
529 master_fd
= open(primary_path
, O_RDWR
| O_CLOEXEC
);
530 if (master_fd
>= 0) {
531 /* prod the device with a GETPARAM call which will fail if
532 * we don't have permission to even render on this device
534 if (anv_gem_get_param(master_fd
, I915_PARAM_CHIPSET_ID
) == 0) {
540 device
->master_fd
= master_fd
;
542 result
= anv_init_wsi(device
);
543 if (result
!= VK_SUCCESS
) {
544 ralloc_free(device
->compiler
);
545 anv_physical_device_free_disk_cache(device
);
549 anv_physical_device_get_supported_extensions(device
,
550 &device
->supported_extensions
);
553 device
->local_fd
= fd
;
565 anv_physical_device_finish(struct anv_physical_device
*device
)
567 anv_finish_wsi(device
);
568 anv_physical_device_free_disk_cache(device
);
569 ralloc_free(device
->compiler
);
570 close(device
->local_fd
);
571 if (device
->master_fd
>= 0)
572 close(device
->master_fd
);
576 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
577 VkSystemAllocationScope allocationScope
)
583 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
584 size_t align
, VkSystemAllocationScope allocationScope
)
586 return realloc(pOriginal
, size
);
590 default_free_func(void *pUserData
, void *pMemory
)
595 static const VkAllocationCallbacks default_alloc
= {
597 .pfnAllocation
= default_alloc_func
,
598 .pfnReallocation
= default_realloc_func
,
599 .pfnFree
= default_free_func
,
602 VkResult
anv_EnumerateInstanceExtensionProperties(
603 const char* pLayerName
,
604 uint32_t* pPropertyCount
,
605 VkExtensionProperties
* pProperties
)
607 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
609 for (int i
= 0; i
< ANV_INSTANCE_EXTENSION_COUNT
; i
++) {
610 if (anv_instance_extensions_supported
.extensions
[i
]) {
611 vk_outarray_append(&out
, prop
) {
612 *prop
= anv_instance_extensions
[i
];
617 return vk_outarray_status(&out
);
620 VkResult
anv_CreateInstance(
621 const VkInstanceCreateInfo
* pCreateInfo
,
622 const VkAllocationCallbacks
* pAllocator
,
623 VkInstance
* pInstance
)
625 struct anv_instance
*instance
;
628 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
630 struct anv_instance_extension_table enabled_extensions
= {};
631 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
633 for (idx
= 0; idx
< ANV_INSTANCE_EXTENSION_COUNT
; idx
++) {
634 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
635 anv_instance_extensions
[idx
].extensionName
) == 0)
639 if (idx
>= ANV_INSTANCE_EXTENSION_COUNT
)
640 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
642 if (!anv_instance_extensions_supported
.extensions
[idx
])
643 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
645 enabled_extensions
.extensions
[idx
] = true;
648 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
649 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
651 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
653 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
656 instance
->alloc
= *pAllocator
;
658 instance
->alloc
= default_alloc
;
660 instance
->app_info
= (struct anv_app_info
) { .api_version
= 0 };
661 if (pCreateInfo
->pApplicationInfo
) {
662 const VkApplicationInfo
*app
= pCreateInfo
->pApplicationInfo
;
664 instance
->app_info
.app_name
=
665 vk_strdup(&instance
->alloc
, app
->pApplicationName
,
666 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
667 instance
->app_info
.app_version
= app
->applicationVersion
;
669 instance
->app_info
.engine_name
=
670 vk_strdup(&instance
->alloc
, app
->pEngineName
,
671 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
672 instance
->app_info
.engine_version
= app
->engineVersion
;
674 instance
->app_info
.api_version
= app
->apiVersion
;
677 if (instance
->app_info
.api_version
== 0)
678 instance
->app_info
.api_version
= VK_API_VERSION_1_0
;
680 instance
->enabled_extensions
= enabled_extensions
;
682 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->dispatch
.entrypoints
); i
++) {
683 /* Vulkan requires that entrypoints for extensions which have not been
684 * enabled must not be advertised.
686 if (!anv_instance_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
687 &instance
->enabled_extensions
)) {
688 instance
->dispatch
.entrypoints
[i
] = NULL
;
690 instance
->dispatch
.entrypoints
[i
] =
691 anv_instance_dispatch_table
.entrypoints
[i
];
695 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->device_dispatch
.entrypoints
); i
++) {
696 /* Vulkan requires that entrypoints for extensions which have not been
697 * enabled must not be advertised.
699 if (!anv_device_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
700 &instance
->enabled_extensions
, NULL
)) {
701 instance
->device_dispatch
.entrypoints
[i
] = NULL
;
703 instance
->device_dispatch
.entrypoints
[i
] =
704 anv_device_dispatch_table
.entrypoints
[i
];
708 instance
->physicalDeviceCount
= -1;
710 result
= vk_debug_report_instance_init(&instance
->debug_report_callbacks
);
711 if (result
!= VK_SUCCESS
) {
712 vk_free2(&default_alloc
, pAllocator
, instance
);
713 return vk_error(result
);
716 instance
->pipeline_cache_enabled
=
717 env_var_as_boolean("ANV_ENABLE_PIPELINE_CACHE", true);
720 glsl_type_singleton_init_or_ref();
722 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
724 *pInstance
= anv_instance_to_handle(instance
);
729 void anv_DestroyInstance(
730 VkInstance _instance
,
731 const VkAllocationCallbacks
* pAllocator
)
733 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
738 if (instance
->physicalDeviceCount
> 0) {
739 /* We support at most one physical device. */
740 assert(instance
->physicalDeviceCount
== 1);
741 anv_physical_device_finish(&instance
->physicalDevice
);
744 vk_free(&instance
->alloc
, (char *)instance
->app_info
.app_name
);
745 vk_free(&instance
->alloc
, (char *)instance
->app_info
.engine_name
);
747 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
749 vk_debug_report_instance_destroy(&instance
->debug_report_callbacks
);
751 glsl_type_singleton_decref();
754 vk_free(&instance
->alloc
, instance
);
758 anv_enumerate_devices(struct anv_instance
*instance
)
760 /* TODO: Check for more devices ? */
761 drmDevicePtr devices
[8];
762 VkResult result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
765 instance
->physicalDeviceCount
= 0;
767 max_devices
= drmGetDevices2(0, devices
, ARRAY_SIZE(devices
));
769 return VK_ERROR_INCOMPATIBLE_DRIVER
;
771 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
772 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
773 devices
[i
]->bustype
== DRM_BUS_PCI
&&
774 devices
[i
]->deviceinfo
.pci
->vendor_id
== 0x8086) {
776 result
= anv_physical_device_init(&instance
->physicalDevice
,
777 instance
, devices
[i
]);
778 if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
782 drmFreeDevices(devices
, max_devices
);
784 if (result
== VK_SUCCESS
)
785 instance
->physicalDeviceCount
= 1;
791 anv_instance_ensure_physical_device(struct anv_instance
*instance
)
793 if (instance
->physicalDeviceCount
< 0) {
794 VkResult result
= anv_enumerate_devices(instance
);
795 if (result
!= VK_SUCCESS
&&
796 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
803 VkResult
anv_EnumeratePhysicalDevices(
804 VkInstance _instance
,
805 uint32_t* pPhysicalDeviceCount
,
806 VkPhysicalDevice
* pPhysicalDevices
)
808 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
809 VK_OUTARRAY_MAKE(out
, pPhysicalDevices
, pPhysicalDeviceCount
);
811 VkResult result
= anv_instance_ensure_physical_device(instance
);
812 if (result
!= VK_SUCCESS
)
815 if (instance
->physicalDeviceCount
== 0)
818 assert(instance
->physicalDeviceCount
== 1);
819 vk_outarray_append(&out
, i
) {
820 *i
= anv_physical_device_to_handle(&instance
->physicalDevice
);
823 return vk_outarray_status(&out
);
826 VkResult
anv_EnumeratePhysicalDeviceGroups(
827 VkInstance _instance
,
828 uint32_t* pPhysicalDeviceGroupCount
,
829 VkPhysicalDeviceGroupProperties
* pPhysicalDeviceGroupProperties
)
831 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
832 VK_OUTARRAY_MAKE(out
, pPhysicalDeviceGroupProperties
,
833 pPhysicalDeviceGroupCount
);
835 VkResult result
= anv_instance_ensure_physical_device(instance
);
836 if (result
!= VK_SUCCESS
)
839 if (instance
->physicalDeviceCount
== 0)
842 assert(instance
->physicalDeviceCount
== 1);
844 vk_outarray_append(&out
, p
) {
845 p
->physicalDeviceCount
= 1;
846 memset(p
->physicalDevices
, 0, sizeof(p
->physicalDevices
));
847 p
->physicalDevices
[0] =
848 anv_physical_device_to_handle(&instance
->physicalDevice
);
849 p
->subsetAllocation
= false;
851 vk_foreach_struct(ext
, p
->pNext
)
852 anv_debug_ignored_stype(ext
->sType
);
855 return vk_outarray_status(&out
);
858 void anv_GetPhysicalDeviceFeatures(
859 VkPhysicalDevice physicalDevice
,
860 VkPhysicalDeviceFeatures
* pFeatures
)
862 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
864 *pFeatures
= (VkPhysicalDeviceFeatures
) {
865 .robustBufferAccess
= true,
866 .fullDrawIndexUint32
= true,
867 .imageCubeArray
= true,
868 .independentBlend
= true,
869 .geometryShader
= true,
870 .tessellationShader
= true,
871 .sampleRateShading
= true,
872 .dualSrcBlend
= true,
874 .multiDrawIndirect
= true,
875 .drawIndirectFirstInstance
= true,
877 .depthBiasClamp
= true,
878 .fillModeNonSolid
= true,
879 .depthBounds
= false,
883 .multiViewport
= true,
884 .samplerAnisotropy
= true,
885 .textureCompressionETC2
= pdevice
->info
.gen
>= 8 ||
886 pdevice
->info
.is_baytrail
,
887 .textureCompressionASTC_LDR
= pdevice
->info
.gen
>= 9, /* FINISHME CHV */
888 .textureCompressionBC
= true,
889 .occlusionQueryPrecise
= true,
890 .pipelineStatisticsQuery
= true,
891 .fragmentStoresAndAtomics
= true,
892 .shaderTessellationAndGeometryPointSize
= true,
893 .shaderImageGatherExtended
= true,
894 .shaderStorageImageExtendedFormats
= true,
895 .shaderStorageImageMultisample
= false,
896 .shaderStorageImageReadWithoutFormat
= false,
897 .shaderStorageImageWriteWithoutFormat
= true,
898 .shaderUniformBufferArrayDynamicIndexing
= true,
899 .shaderSampledImageArrayDynamicIndexing
= true,
900 .shaderStorageBufferArrayDynamicIndexing
= true,
901 .shaderStorageImageArrayDynamicIndexing
= true,
902 .shaderClipDistance
= true,
903 .shaderCullDistance
= true,
904 .shaderFloat64
= pdevice
->info
.gen
>= 8 &&
905 pdevice
->info
.has_64bit_types
,
906 .shaderInt64
= pdevice
->info
.gen
>= 8 &&
907 pdevice
->info
.has_64bit_types
,
908 .shaderInt16
= pdevice
->info
.gen
>= 8,
909 .shaderResourceMinLod
= pdevice
->info
.gen
>= 9,
910 .variableMultisampleRate
= true,
911 .inheritedQueries
= true,
914 /* We can't do image stores in vec4 shaders */
915 pFeatures
->vertexPipelineStoresAndAtomics
=
916 pdevice
->compiler
->scalar_stage
[MESA_SHADER_VERTEX
] &&
917 pdevice
->compiler
->scalar_stage
[MESA_SHADER_GEOMETRY
];
919 struct anv_app_info
*app_info
= &pdevice
->instance
->app_info
;
921 /* The new DOOM and Wolfenstein games require depthBounds without
922 * checking for it. They seem to run fine without it so just claim it's
923 * there and accept the consequences.
925 if (app_info
->engine_name
&& strcmp(app_info
->engine_name
, "idTech") == 0)
926 pFeatures
->depthBounds
= true;
929 void anv_GetPhysicalDeviceFeatures2(
930 VkPhysicalDevice physicalDevice
,
931 VkPhysicalDeviceFeatures2
* pFeatures
)
933 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
934 anv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
936 vk_foreach_struct(ext
, pFeatures
->pNext
) {
937 switch (ext
->sType
) {
938 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES_KHR
: {
939 VkPhysicalDevice8BitStorageFeaturesKHR
*features
=
940 (VkPhysicalDevice8BitStorageFeaturesKHR
*)ext
;
941 features
->storageBuffer8BitAccess
= pdevice
->info
.gen
>= 8;
942 features
->uniformAndStorageBuffer8BitAccess
= pdevice
->info
.gen
>= 8;
943 features
->storagePushConstant8
= pdevice
->info
.gen
>= 8;
947 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES
: {
948 VkPhysicalDevice16BitStorageFeatures
*features
=
949 (VkPhysicalDevice16BitStorageFeatures
*)ext
;
950 features
->storageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
951 features
->uniformAndStorageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
952 features
->storagePushConstant16
= pdevice
->info
.gen
>= 8;
953 features
->storageInputOutput16
= false;
957 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_EXT
: {
958 VkPhysicalDeviceBufferDeviceAddressFeaturesEXT
*features
= (void *)ext
;
959 features
->bufferDeviceAddress
= pdevice
->use_softpin
&&
960 pdevice
->info
.gen
>= 8;
961 features
->bufferDeviceAddressCaptureReplay
= false;
962 features
->bufferDeviceAddressMultiDevice
= false;
966 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COMPUTE_SHADER_DERIVATIVES_FEATURES_NV
: {
967 VkPhysicalDeviceComputeShaderDerivativesFeaturesNV
*features
=
968 (VkPhysicalDeviceComputeShaderDerivativesFeaturesNV
*)ext
;
969 features
->computeDerivativeGroupQuads
= true;
970 features
->computeDerivativeGroupLinear
= true;
974 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT
: {
975 VkPhysicalDeviceConditionalRenderingFeaturesEXT
*features
=
976 (VkPhysicalDeviceConditionalRenderingFeaturesEXT
*)ext
;
977 features
->conditionalRendering
= pdevice
->info
.gen
>= 8 ||
978 pdevice
->info
.is_haswell
;
979 features
->inheritedConditionalRendering
= pdevice
->info
.gen
>= 8 ||
980 pdevice
->info
.is_haswell
;
984 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_CLIP_ENABLE_FEATURES_EXT
: {
985 VkPhysicalDeviceDepthClipEnableFeaturesEXT
*features
=
986 (VkPhysicalDeviceDepthClipEnableFeaturesEXT
*)ext
;
987 features
->depthClipEnable
= true;
991 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR
: {
992 VkPhysicalDeviceFloat16Int8FeaturesKHR
*features
= (void *)ext
;
993 features
->shaderFloat16
= pdevice
->info
.gen
>= 8;
994 features
->shaderInt8
= pdevice
->info
.gen
>= 8;
998 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_QUERY_RESET_FEATURES_EXT
: {
999 VkPhysicalDeviceHostQueryResetFeaturesEXT
*features
=
1000 (VkPhysicalDeviceHostQueryResetFeaturesEXT
*)ext
;
1001 features
->hostQueryReset
= true;
1005 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_FEATURES_EXT
: {
1006 VkPhysicalDeviceInlineUniformBlockFeaturesEXT
*features
=
1007 (VkPhysicalDeviceInlineUniformBlockFeaturesEXT
*)ext
;
1008 features
->inlineUniformBlock
= true;
1009 features
->descriptorBindingInlineUniformBlockUpdateAfterBind
= false;
1013 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES
: {
1014 VkPhysicalDeviceMultiviewFeatures
*features
=
1015 (VkPhysicalDeviceMultiviewFeatures
*)ext
;
1016 features
->multiview
= true;
1017 features
->multiviewGeometryShader
= true;
1018 features
->multiviewTessellationShader
= true;
1022 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES
: {
1023 VkPhysicalDeviceProtectedMemoryFeatures
*features
= (void *)ext
;
1024 features
->protectedMemory
= false;
1028 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES
: {
1029 VkPhysicalDeviceSamplerYcbcrConversionFeatures
*features
=
1030 (VkPhysicalDeviceSamplerYcbcrConversionFeatures
*) ext
;
1031 features
->samplerYcbcrConversion
= true;
1035 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES_EXT
: {
1036 VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*features
=
1037 (VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*)ext
;
1038 features
->scalarBlockLayout
= true;
1042 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETERS_FEATURES
: {
1043 VkPhysicalDeviceShaderDrawParametersFeatures
*features
= (void *)ext
;
1044 features
->shaderDrawParameters
= true;
1048 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES
: {
1049 VkPhysicalDeviceVariablePointersFeatures
*features
= (void *)ext
;
1050 features
->variablePointersStorageBuffer
= true;
1051 features
->variablePointers
= true;
1055 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT
: {
1056 VkPhysicalDeviceTransformFeedbackFeaturesEXT
*features
=
1057 (VkPhysicalDeviceTransformFeedbackFeaturesEXT
*)ext
;
1058 features
->transformFeedback
= true;
1059 features
->geometryStreams
= true;
1063 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT
: {
1064 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*features
=
1065 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*)ext
;
1066 features
->vertexAttributeInstanceRateDivisor
= true;
1067 features
->vertexAttributeInstanceRateZeroDivisor
= true;
1071 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_YCBCR_IMAGE_ARRAYS_FEATURES_EXT
: {
1072 VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*features
=
1073 (VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*)ext
;
1074 features
->ycbcrImageArrays
= true;
1079 anv_debug_ignored_stype(ext
->sType
);
1085 void anv_GetPhysicalDeviceProperties(
1086 VkPhysicalDevice physicalDevice
,
1087 VkPhysicalDeviceProperties
* pProperties
)
1089 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1090 const struct gen_device_info
*devinfo
= &pdevice
->info
;
1092 /* See assertions made when programming the buffer surface state. */
1093 const uint32_t max_raw_buffer_sz
= devinfo
->gen
>= 7 ?
1094 (1ul << 30) : (1ul << 27);
1096 const uint32_t max_samplers
= (devinfo
->gen
>= 8 || devinfo
->is_haswell
) ?
1099 VkSampleCountFlags sample_counts
=
1100 isl_device_get_sample_counts(&pdevice
->isl_dev
);
1103 VkPhysicalDeviceLimits limits
= {
1104 .maxImageDimension1D
= (1 << 14),
1105 .maxImageDimension2D
= (1 << 14),
1106 .maxImageDimension3D
= (1 << 11),
1107 .maxImageDimensionCube
= (1 << 14),
1108 .maxImageArrayLayers
= (1 << 11),
1109 .maxTexelBufferElements
= 128 * 1024 * 1024,
1110 .maxUniformBufferRange
= (1ul << 27),
1111 .maxStorageBufferRange
= max_raw_buffer_sz
,
1112 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
1113 .maxMemoryAllocationCount
= UINT32_MAX
,
1114 .maxSamplerAllocationCount
= 64 * 1024,
1115 .bufferImageGranularity
= 64, /* A cache line */
1116 .sparseAddressSpaceSize
= 0,
1117 .maxBoundDescriptorSets
= MAX_SETS
,
1118 .maxPerStageDescriptorSamplers
= max_samplers
,
1119 .maxPerStageDescriptorUniformBuffers
= 64,
1120 .maxPerStageDescriptorStorageBuffers
= 64,
1121 .maxPerStageDescriptorSampledImages
= max_samplers
,
1122 .maxPerStageDescriptorStorageImages
= MAX_IMAGES
,
1123 .maxPerStageDescriptorInputAttachments
= 64,
1124 .maxPerStageResources
= 250,
1125 .maxDescriptorSetSamplers
= 6 * max_samplers
, /* number of stages * maxPerStageDescriptorSamplers */
1126 .maxDescriptorSetUniformBuffers
= 6 * 64, /* number of stages * maxPerStageDescriptorUniformBuffers */
1127 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
1128 .maxDescriptorSetStorageBuffers
= 6 * 64, /* number of stages * maxPerStageDescriptorStorageBuffers */
1129 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
1130 .maxDescriptorSetSampledImages
= 6 * max_samplers
, /* number of stages * maxPerStageDescriptorSampledImages */
1131 .maxDescriptorSetStorageImages
= 6 * MAX_IMAGES
, /* number of stages * maxPerStageDescriptorStorageImages */
1132 .maxDescriptorSetInputAttachments
= 256,
1133 .maxVertexInputAttributes
= MAX_VBS
,
1134 .maxVertexInputBindings
= MAX_VBS
,
1135 .maxVertexInputAttributeOffset
= 2047,
1136 .maxVertexInputBindingStride
= 2048,
1137 .maxVertexOutputComponents
= 128,
1138 .maxTessellationGenerationLevel
= 64,
1139 .maxTessellationPatchSize
= 32,
1140 .maxTessellationControlPerVertexInputComponents
= 128,
1141 .maxTessellationControlPerVertexOutputComponents
= 128,
1142 .maxTessellationControlPerPatchOutputComponents
= 128,
1143 .maxTessellationControlTotalOutputComponents
= 2048,
1144 .maxTessellationEvaluationInputComponents
= 128,
1145 .maxTessellationEvaluationOutputComponents
= 128,
1146 .maxGeometryShaderInvocations
= 32,
1147 .maxGeometryInputComponents
= 64,
1148 .maxGeometryOutputComponents
= 128,
1149 .maxGeometryOutputVertices
= 256,
1150 .maxGeometryTotalOutputComponents
= 1024,
1151 .maxFragmentInputComponents
= 112, /* 128 components - (POS, PSIZ, CLIP_DIST0, CLIP_DIST1) */
1152 .maxFragmentOutputAttachments
= 8,
1153 .maxFragmentDualSrcAttachments
= 1,
1154 .maxFragmentCombinedOutputResources
= 8,
1155 .maxComputeSharedMemorySize
= 32768,
1156 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
1157 .maxComputeWorkGroupInvocations
= 32 * devinfo
->max_cs_threads
,
1158 .maxComputeWorkGroupSize
= {
1159 16 * devinfo
->max_cs_threads
,
1160 16 * devinfo
->max_cs_threads
,
1161 16 * devinfo
->max_cs_threads
,
1163 .subPixelPrecisionBits
= 8,
1164 .subTexelPrecisionBits
= 8,
1165 .mipmapPrecisionBits
= 8,
1166 .maxDrawIndexedIndexValue
= UINT32_MAX
,
1167 .maxDrawIndirectCount
= UINT32_MAX
,
1168 .maxSamplerLodBias
= 16,
1169 .maxSamplerAnisotropy
= 16,
1170 .maxViewports
= MAX_VIEWPORTS
,
1171 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
1172 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
1173 .viewportSubPixelBits
= 13, /* We take a float? */
1174 .minMemoryMapAlignment
= 4096, /* A page */
1175 .minTexelBufferOffsetAlignment
= 1,
1176 /* We need 16 for UBO block reads to work and 32 for push UBOs */
1177 .minUniformBufferOffsetAlignment
= 32,
1178 .minStorageBufferOffsetAlignment
= 4,
1179 .minTexelOffset
= -8,
1180 .maxTexelOffset
= 7,
1181 .minTexelGatherOffset
= -32,
1182 .maxTexelGatherOffset
= 31,
1183 .minInterpolationOffset
= -0.5,
1184 .maxInterpolationOffset
= 0.4375,
1185 .subPixelInterpolationOffsetBits
= 4,
1186 .maxFramebufferWidth
= (1 << 14),
1187 .maxFramebufferHeight
= (1 << 14),
1188 .maxFramebufferLayers
= (1 << 11),
1189 .framebufferColorSampleCounts
= sample_counts
,
1190 .framebufferDepthSampleCounts
= sample_counts
,
1191 .framebufferStencilSampleCounts
= sample_counts
,
1192 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
1193 .maxColorAttachments
= MAX_RTS
,
1194 .sampledImageColorSampleCounts
= sample_counts
,
1195 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
1196 .sampledImageDepthSampleCounts
= sample_counts
,
1197 .sampledImageStencilSampleCounts
= sample_counts
,
1198 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
1199 .maxSampleMaskWords
= 1,
1200 .timestampComputeAndGraphics
= false,
1201 .timestampPeriod
= 1000000000.0 / devinfo
->timestamp_frequency
,
1202 .maxClipDistances
= 8,
1203 .maxCullDistances
= 8,
1204 .maxCombinedClipAndCullDistances
= 8,
1205 .discreteQueuePriorities
= 2,
1206 .pointSizeRange
= { 0.125, 255.875 },
1207 .lineWidthRange
= { 0.0, 7.9921875 },
1208 .pointSizeGranularity
= (1.0 / 8.0),
1209 .lineWidthGranularity
= (1.0 / 128.0),
1210 .strictLines
= false, /* FINISHME */
1211 .standardSampleLocations
= true,
1212 .optimalBufferCopyOffsetAlignment
= 128,
1213 .optimalBufferCopyRowPitchAlignment
= 128,
1214 .nonCoherentAtomSize
= 64,
1217 *pProperties
= (VkPhysicalDeviceProperties
) {
1218 .apiVersion
= anv_physical_device_api_version(pdevice
),
1219 .driverVersion
= vk_get_driver_version(),
1221 .deviceID
= pdevice
->chipset_id
,
1222 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
1224 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
1227 snprintf(pProperties
->deviceName
, sizeof(pProperties
->deviceName
),
1228 "%s", pdevice
->name
);
1229 memcpy(pProperties
->pipelineCacheUUID
,
1230 pdevice
->pipeline_cache_uuid
, VK_UUID_SIZE
);
1233 void anv_GetPhysicalDeviceProperties2(
1234 VkPhysicalDevice physicalDevice
,
1235 VkPhysicalDeviceProperties2
* pProperties
)
1237 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1239 anv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
1241 vk_foreach_struct(ext
, pProperties
->pNext
) {
1242 switch (ext
->sType
) {
1243 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_STENCIL_RESOLVE_PROPERTIES_KHR
: {
1244 VkPhysicalDeviceDepthStencilResolvePropertiesKHR
*props
=
1245 (VkPhysicalDeviceDepthStencilResolvePropertiesKHR
*)ext
;
1247 /* We support all of the depth resolve modes */
1248 props
->supportedDepthResolveModes
=
1249 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
|
1250 VK_RESOLVE_MODE_AVERAGE_BIT_KHR
|
1251 VK_RESOLVE_MODE_MIN_BIT_KHR
|
1252 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1254 /* Average doesn't make sense for stencil so we don't support that */
1255 props
->supportedStencilResolveModes
=
1256 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
;
1257 if (pdevice
->info
.gen
>= 8) {
1258 /* The advanced stencil resolve modes currently require stencil
1259 * sampling be supported by the hardware.
1261 props
->supportedStencilResolveModes
|=
1262 VK_RESOLVE_MODE_MIN_BIT_KHR
|
1263 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1266 props
->independentResolveNone
= true;
1267 props
->independentResolve
= true;
1271 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR
: {
1272 VkPhysicalDeviceDriverPropertiesKHR
*driver_props
=
1273 (VkPhysicalDeviceDriverPropertiesKHR
*) ext
;
1275 driver_props
->driverID
= VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA_KHR
;
1276 util_snprintf(driver_props
->driverName
, VK_MAX_DRIVER_NAME_SIZE_KHR
,
1277 "Intel open-source Mesa driver");
1279 util_snprintf(driver_props
->driverInfo
, VK_MAX_DRIVER_INFO_SIZE_KHR
,
1280 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1
);
1282 driver_props
->conformanceVersion
= (VkConformanceVersionKHR
) {
1291 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT
: {
1292 VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*props
=
1293 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*) ext
;
1294 /* Userptr needs page aligned memory. */
1295 props
->minImportedHostPointerAlignment
= 4096;
1299 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES
: {
1300 VkPhysicalDeviceIDProperties
*id_props
=
1301 (VkPhysicalDeviceIDProperties
*)ext
;
1302 memcpy(id_props
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
1303 memcpy(id_props
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
1304 /* The LUID is for Windows. */
1305 id_props
->deviceLUIDValid
= false;
1309 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_PROPERTIES_EXT
: {
1310 VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*props
=
1311 (VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*)ext
;
1312 props
->maxInlineUniformBlockSize
= MAX_INLINE_UNIFORM_BLOCK_SIZE
;
1313 props
->maxPerStageDescriptorInlineUniformBlocks
=
1314 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1315 props
->maxPerStageDescriptorUpdateAfterBindInlineUniformBlocks
=
1316 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1317 props
->maxDescriptorSetInlineUniformBlocks
=
1318 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1319 props
->maxDescriptorSetUpdateAfterBindInlineUniformBlocks
=
1320 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1324 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES
: {
1325 VkPhysicalDeviceMaintenance3Properties
*props
=
1326 (VkPhysicalDeviceMaintenance3Properties
*)ext
;
1327 /* This value doesn't matter for us today as our per-stage
1328 * descriptors are the real limit.
1330 props
->maxPerSetDescriptors
= 1024;
1331 props
->maxMemoryAllocationSize
= MAX_MEMORY_ALLOCATION_SIZE
;
1335 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES
: {
1336 VkPhysicalDeviceMultiviewProperties
*properties
=
1337 (VkPhysicalDeviceMultiviewProperties
*)ext
;
1338 properties
->maxMultiviewViewCount
= 16;
1339 properties
->maxMultiviewInstanceIndex
= UINT32_MAX
/ 16;
1343 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT
: {
1344 VkPhysicalDevicePCIBusInfoPropertiesEXT
*properties
=
1345 (VkPhysicalDevicePCIBusInfoPropertiesEXT
*)ext
;
1346 properties
->pciDomain
= pdevice
->pci_info
.domain
;
1347 properties
->pciBus
= pdevice
->pci_info
.bus
;
1348 properties
->pciDevice
= pdevice
->pci_info
.device
;
1349 properties
->pciFunction
= pdevice
->pci_info
.function
;
1353 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES
: {
1354 VkPhysicalDevicePointClippingProperties
*properties
=
1355 (VkPhysicalDevicePointClippingProperties
*) ext
;
1356 properties
->pointClippingBehavior
= VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES
;
1357 anv_finishme("Implement pop-free point clipping");
1361 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES
: {
1362 VkPhysicalDeviceProtectedMemoryProperties
*props
=
1363 (VkPhysicalDeviceProtectedMemoryProperties
*)ext
;
1364 props
->protectedNoFault
= false;
1368 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
1369 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
1370 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
1372 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
1376 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT
: {
1377 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*properties
=
1378 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*)ext
;
1379 properties
->filterMinmaxImageComponentMapping
= pdevice
->info
.gen
>= 9;
1380 properties
->filterMinmaxSingleComponentFormats
= true;
1384 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES
: {
1385 VkPhysicalDeviceSubgroupProperties
*properties
= (void *)ext
;
1387 properties
->subgroupSize
= BRW_SUBGROUP_SIZE
;
1389 VkShaderStageFlags scalar_stages
= 0;
1390 for (unsigned stage
= 0; stage
< MESA_SHADER_STAGES
; stage
++) {
1391 if (pdevice
->compiler
->scalar_stage
[stage
])
1392 scalar_stages
|= mesa_to_vk_shader_stage(stage
);
1394 properties
->supportedStages
= scalar_stages
;
1396 properties
->supportedOperations
= VK_SUBGROUP_FEATURE_BASIC_BIT
|
1397 VK_SUBGROUP_FEATURE_VOTE_BIT
|
1398 VK_SUBGROUP_FEATURE_ARITHMETIC_BIT
|
1399 VK_SUBGROUP_FEATURE_BALLOT_BIT
|
1400 VK_SUBGROUP_FEATURE_SHUFFLE_BIT
|
1401 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT
|
1402 VK_SUBGROUP_FEATURE_CLUSTERED_BIT
|
1403 VK_SUBGROUP_FEATURE_QUAD_BIT
;
1404 properties
->quadOperationsInAllStages
= true;
1408 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT
: {
1409 VkPhysicalDeviceTransformFeedbackPropertiesEXT
*props
=
1410 (VkPhysicalDeviceTransformFeedbackPropertiesEXT
*)ext
;
1412 props
->maxTransformFeedbackStreams
= MAX_XFB_STREAMS
;
1413 props
->maxTransformFeedbackBuffers
= MAX_XFB_BUFFERS
;
1414 props
->maxTransformFeedbackBufferSize
= (1ull << 32);
1415 props
->maxTransformFeedbackStreamDataSize
= 128 * 4;
1416 props
->maxTransformFeedbackBufferDataSize
= 128 * 4;
1417 props
->maxTransformFeedbackBufferDataStride
= 2048;
1418 props
->transformFeedbackQueries
= true;
1419 props
->transformFeedbackStreamsLinesTriangles
= false;
1420 props
->transformFeedbackRasterizationStreamSelect
= false;
1421 props
->transformFeedbackDraw
= true;
1425 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT
: {
1426 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*props
=
1427 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*)ext
;
1428 /* We have to restrict this a bit for multiview */
1429 props
->maxVertexAttribDivisor
= UINT32_MAX
/ 16;
1434 anv_debug_ignored_stype(ext
->sType
);
1440 /* We support exactly one queue family. */
1441 static const VkQueueFamilyProperties
1442 anv_queue_family_properties
= {
1443 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
1444 VK_QUEUE_COMPUTE_BIT
|
1445 VK_QUEUE_TRANSFER_BIT
,
1447 .timestampValidBits
= 36, /* XXX: Real value here */
1448 .minImageTransferGranularity
= { 1, 1, 1 },
1451 void anv_GetPhysicalDeviceQueueFamilyProperties(
1452 VkPhysicalDevice physicalDevice
,
1454 VkQueueFamilyProperties
* pQueueFamilyProperties
)
1456 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pCount
);
1458 vk_outarray_append(&out
, p
) {
1459 *p
= anv_queue_family_properties
;
1463 void anv_GetPhysicalDeviceQueueFamilyProperties2(
1464 VkPhysicalDevice physicalDevice
,
1465 uint32_t* pQueueFamilyPropertyCount
,
1466 VkQueueFamilyProperties2
* pQueueFamilyProperties
)
1469 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pQueueFamilyPropertyCount
);
1471 vk_outarray_append(&out
, p
) {
1472 p
->queueFamilyProperties
= anv_queue_family_properties
;
1474 vk_foreach_struct(s
, p
->pNext
) {
1475 anv_debug_ignored_stype(s
->sType
);
1480 void anv_GetPhysicalDeviceMemoryProperties(
1481 VkPhysicalDevice physicalDevice
,
1482 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
1484 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
1486 pMemoryProperties
->memoryTypeCount
= physical_device
->memory
.type_count
;
1487 for (uint32_t i
= 0; i
< physical_device
->memory
.type_count
; i
++) {
1488 pMemoryProperties
->memoryTypes
[i
] = (VkMemoryType
) {
1489 .propertyFlags
= physical_device
->memory
.types
[i
].propertyFlags
,
1490 .heapIndex
= physical_device
->memory
.types
[i
].heapIndex
,
1494 pMemoryProperties
->memoryHeapCount
= physical_device
->memory
.heap_count
;
1495 for (uint32_t i
= 0; i
< physical_device
->memory
.heap_count
; i
++) {
1496 pMemoryProperties
->memoryHeaps
[i
] = (VkMemoryHeap
) {
1497 .size
= physical_device
->memory
.heaps
[i
].size
,
1498 .flags
= physical_device
->memory
.heaps
[i
].flags
,
1503 void anv_GetPhysicalDeviceMemoryProperties2(
1504 VkPhysicalDevice physicalDevice
,
1505 VkPhysicalDeviceMemoryProperties2
* pMemoryProperties
)
1507 anv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
1508 &pMemoryProperties
->memoryProperties
);
1510 vk_foreach_struct(ext
, pMemoryProperties
->pNext
) {
1511 switch (ext
->sType
) {
1513 anv_debug_ignored_stype(ext
->sType
);
1520 anv_GetDeviceGroupPeerMemoryFeatures(
1523 uint32_t localDeviceIndex
,
1524 uint32_t remoteDeviceIndex
,
1525 VkPeerMemoryFeatureFlags
* pPeerMemoryFeatures
)
1527 assert(localDeviceIndex
== 0 && remoteDeviceIndex
== 0);
1528 *pPeerMemoryFeatures
= VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT
|
1529 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT
|
1530 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT
|
1531 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT
;
1534 PFN_vkVoidFunction
anv_GetInstanceProcAddr(
1535 VkInstance _instance
,
1538 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1540 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
1541 * when we have to return valid function pointers, NULL, or it's left
1542 * undefined. See the table for exact details.
1547 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
1548 if (strcmp(pName, "vk" #entrypoint) == 0) \
1549 return (PFN_vkVoidFunction)anv_##entrypoint
1551 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties
);
1552 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties
);
1553 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceVersion
);
1554 LOOKUP_ANV_ENTRYPOINT(CreateInstance
);
1556 #undef LOOKUP_ANV_ENTRYPOINT
1558 if (instance
== NULL
)
1561 int idx
= anv_get_instance_entrypoint_index(pName
);
1563 return instance
->dispatch
.entrypoints
[idx
];
1565 idx
= anv_get_device_entrypoint_index(pName
);
1567 return instance
->device_dispatch
.entrypoints
[idx
];
1572 /* With version 1+ of the loader interface the ICD should expose
1573 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
1576 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1577 VkInstance instance
,
1581 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1582 VkInstance instance
,
1585 return anv_GetInstanceProcAddr(instance
, pName
);
1588 PFN_vkVoidFunction
anv_GetDeviceProcAddr(
1592 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1594 if (!device
|| !pName
)
1597 int idx
= anv_get_device_entrypoint_index(pName
);
1601 return device
->dispatch
.entrypoints
[idx
];
1605 anv_CreateDebugReportCallbackEXT(VkInstance _instance
,
1606 const VkDebugReportCallbackCreateInfoEXT
* pCreateInfo
,
1607 const VkAllocationCallbacks
* pAllocator
,
1608 VkDebugReportCallbackEXT
* pCallback
)
1610 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1611 return vk_create_debug_report_callback(&instance
->debug_report_callbacks
,
1612 pCreateInfo
, pAllocator
, &instance
->alloc
,
1617 anv_DestroyDebugReportCallbackEXT(VkInstance _instance
,
1618 VkDebugReportCallbackEXT _callback
,
1619 const VkAllocationCallbacks
* pAllocator
)
1621 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1622 vk_destroy_debug_report_callback(&instance
->debug_report_callbacks
,
1623 _callback
, pAllocator
, &instance
->alloc
);
1627 anv_DebugReportMessageEXT(VkInstance _instance
,
1628 VkDebugReportFlagsEXT flags
,
1629 VkDebugReportObjectTypeEXT objectType
,
1632 int32_t messageCode
,
1633 const char* pLayerPrefix
,
1634 const char* pMessage
)
1636 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1637 vk_debug_report(&instance
->debug_report_callbacks
, flags
, objectType
,
1638 object
, location
, messageCode
, pLayerPrefix
, pMessage
);
1642 anv_queue_init(struct anv_device
*device
, struct anv_queue
*queue
)
1644 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1645 queue
->device
= device
;
1650 anv_queue_finish(struct anv_queue
*queue
)
1654 static struct anv_state
1655 anv_state_pool_emit_data(struct anv_state_pool
*pool
, size_t size
, size_t align
, const void *p
)
1657 struct anv_state state
;
1659 state
= anv_state_pool_alloc(pool
, size
, align
);
1660 memcpy(state
.map
, p
, size
);
1665 struct gen8_border_color
{
1670 /* Pad out to 64 bytes */
1675 anv_device_init_border_colors(struct anv_device
*device
)
1677 static const struct gen8_border_color border_colors
[] = {
1678 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
1679 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
1680 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
1681 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
1682 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
1683 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
1686 device
->border_colors
= anv_state_pool_emit_data(&device
->dynamic_state_pool
,
1687 sizeof(border_colors
), 64,
1692 anv_device_init_trivial_batch(struct anv_device
*device
)
1694 anv_bo_init_new(&device
->trivial_batch_bo
, device
, 4096);
1696 if (device
->instance
->physicalDevice
.has_exec_async
)
1697 device
->trivial_batch_bo
.flags
|= EXEC_OBJECT_ASYNC
;
1699 if (device
->instance
->physicalDevice
.use_softpin
)
1700 device
->trivial_batch_bo
.flags
|= EXEC_OBJECT_PINNED
;
1702 anv_vma_alloc(device
, &device
->trivial_batch_bo
);
1704 void *map
= anv_gem_mmap(device
, device
->trivial_batch_bo
.gem_handle
,
1707 struct anv_batch batch
= {
1713 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
1714 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
1716 if (!device
->info
.has_llc
)
1717 gen_clflush_range(map
, batch
.next
- map
);
1719 anv_gem_munmap(map
, device
->trivial_batch_bo
.size
);
1722 VkResult
anv_EnumerateDeviceExtensionProperties(
1723 VkPhysicalDevice physicalDevice
,
1724 const char* pLayerName
,
1725 uint32_t* pPropertyCount
,
1726 VkExtensionProperties
* pProperties
)
1728 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
1729 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
1731 for (int i
= 0; i
< ANV_DEVICE_EXTENSION_COUNT
; i
++) {
1732 if (device
->supported_extensions
.extensions
[i
]) {
1733 vk_outarray_append(&out
, prop
) {
1734 *prop
= anv_device_extensions
[i
];
1739 return vk_outarray_status(&out
);
1743 anv_device_init_dispatch(struct anv_device
*device
)
1745 const struct anv_device_dispatch_table
*genX_table
;
1746 switch (device
->info
.gen
) {
1748 genX_table
= &gen11_device_dispatch_table
;
1751 genX_table
= &gen10_device_dispatch_table
;
1754 genX_table
= &gen9_device_dispatch_table
;
1757 genX_table
= &gen8_device_dispatch_table
;
1760 if (device
->info
.is_haswell
)
1761 genX_table
= &gen75_device_dispatch_table
;
1763 genX_table
= &gen7_device_dispatch_table
;
1766 unreachable("unsupported gen\n");
1769 for (unsigned i
= 0; i
< ARRAY_SIZE(device
->dispatch
.entrypoints
); i
++) {
1770 /* Vulkan requires that entrypoints for extensions which have not been
1771 * enabled must not be advertised.
1773 if (!anv_device_entrypoint_is_enabled(i
, device
->instance
->app_info
.api_version
,
1774 &device
->instance
->enabled_extensions
,
1775 &device
->enabled_extensions
)) {
1776 device
->dispatch
.entrypoints
[i
] = NULL
;
1777 } else if (genX_table
->entrypoints
[i
]) {
1778 device
->dispatch
.entrypoints
[i
] = genX_table
->entrypoints
[i
];
1780 device
->dispatch
.entrypoints
[i
] =
1781 anv_device_dispatch_table
.entrypoints
[i
];
1787 vk_priority_to_gen(int priority
)
1790 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT
:
1791 return GEN_CONTEXT_LOW_PRIORITY
;
1792 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
:
1793 return GEN_CONTEXT_MEDIUM_PRIORITY
;
1794 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT
:
1795 return GEN_CONTEXT_HIGH_PRIORITY
;
1796 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT
:
1797 return GEN_CONTEXT_REALTIME_PRIORITY
;
1799 unreachable("Invalid priority");
1804 anv_device_init_hiz_clear_value_bo(struct anv_device
*device
)
1806 anv_bo_init_new(&device
->hiz_clear_bo
, device
, 4096);
1808 if (device
->instance
->physicalDevice
.has_exec_async
)
1809 device
->hiz_clear_bo
.flags
|= EXEC_OBJECT_ASYNC
;
1811 if (device
->instance
->physicalDevice
.use_softpin
)
1812 device
->hiz_clear_bo
.flags
|= EXEC_OBJECT_PINNED
;
1814 anv_vma_alloc(device
, &device
->hiz_clear_bo
);
1816 uint32_t *map
= anv_gem_mmap(device
, device
->hiz_clear_bo
.gem_handle
,
1819 union isl_color_value hiz_clear
= { .u32
= { 0, } };
1820 hiz_clear
.f32
[0] = ANV_HZ_FC_VAL
;
1822 memcpy(map
, hiz_clear
.u32
, sizeof(hiz_clear
.u32
));
1823 anv_gem_munmap(map
, device
->hiz_clear_bo
.size
);
1827 get_bo_from_pool(struct gen_batch_decode_bo
*ret
,
1828 struct anv_block_pool
*pool
,
1831 for (uint32_t i
= 0; i
< pool
->nbos
; i
++) {
1832 uint64_t bo_address
= pool
->bos
[i
].offset
& (~0ull >> 16);
1833 uint32_t bo_size
= pool
->bos
[i
].size
;
1834 if (address
>= bo_address
&& address
< (bo_address
+ bo_size
)) {
1835 *ret
= (struct gen_batch_decode_bo
) {
1838 .map
= pool
->bos
[i
].map
,
1846 /* Finding a buffer for batch decoding */
1847 static struct gen_batch_decode_bo
1848 decode_get_bo(void *v_batch
, bool ppgtt
, uint64_t address
)
1850 struct anv_device
*device
= v_batch
;
1851 struct gen_batch_decode_bo ret_bo
= {};
1855 if (get_bo_from_pool(&ret_bo
, &device
->dynamic_state_pool
.block_pool
, address
))
1857 if (get_bo_from_pool(&ret_bo
, &device
->instruction_state_pool
.block_pool
, address
))
1859 if (get_bo_from_pool(&ret_bo
, &device
->binding_table_pool
.block_pool
, address
))
1861 if (get_bo_from_pool(&ret_bo
, &device
->surface_state_pool
.block_pool
, address
))
1864 if (!device
->cmd_buffer_being_decoded
)
1865 return (struct gen_batch_decode_bo
) { };
1867 struct anv_batch_bo
**bo
;
1869 u_vector_foreach(bo
, &device
->cmd_buffer_being_decoded
->seen_bbos
) {
1870 /* The decoder zeroes out the top 16 bits, so we need to as well */
1871 uint64_t bo_address
= (*bo
)->bo
.offset
& (~0ull >> 16);
1873 if (address
>= bo_address
&& address
< bo_address
+ (*bo
)->bo
.size
) {
1874 return (struct gen_batch_decode_bo
) {
1876 .size
= (*bo
)->bo
.size
,
1877 .map
= (*bo
)->bo
.map
,
1882 return (struct gen_batch_decode_bo
) { };
1885 VkResult
anv_CreateDevice(
1886 VkPhysicalDevice physicalDevice
,
1887 const VkDeviceCreateInfo
* pCreateInfo
,
1888 const VkAllocationCallbacks
* pAllocator
,
1891 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
1893 struct anv_device
*device
;
1895 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO
);
1897 struct anv_device_extension_table enabled_extensions
= { };
1898 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
1900 for (idx
= 0; idx
< ANV_DEVICE_EXTENSION_COUNT
; idx
++) {
1901 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
1902 anv_device_extensions
[idx
].extensionName
) == 0)
1906 if (idx
>= ANV_DEVICE_EXTENSION_COUNT
)
1907 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
1909 if (!physical_device
->supported_extensions
.extensions
[idx
])
1910 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
1912 enabled_extensions
.extensions
[idx
] = true;
1915 /* Check enabled features */
1916 if (pCreateInfo
->pEnabledFeatures
) {
1917 VkPhysicalDeviceFeatures supported_features
;
1918 anv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
1919 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
1920 VkBool32
*enabled_feature
= (VkBool32
*)pCreateInfo
->pEnabledFeatures
;
1921 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
1922 for (uint32_t i
= 0; i
< num_features
; i
++) {
1923 if (enabled_feature
[i
] && !supported_feature
[i
])
1924 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
1928 /* Check requested queues and fail if we are requested to create any
1929 * queues with flags we don't support.
1931 assert(pCreateInfo
->queueCreateInfoCount
> 0);
1932 for (uint32_t i
= 0; i
< pCreateInfo
->queueCreateInfoCount
; i
++) {
1933 if (pCreateInfo
->pQueueCreateInfos
[i
].flags
!= 0)
1934 return vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1937 /* Check if client specified queue priority. */
1938 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*queue_priority
=
1939 vk_find_struct_const(pCreateInfo
->pQueueCreateInfos
[0].pNext
,
1940 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT
);
1942 VkQueueGlobalPriorityEXT priority
=
1943 queue_priority
? queue_priority
->globalPriority
:
1944 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
;
1946 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
1948 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
1950 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1952 const unsigned decode_flags
=
1953 GEN_BATCH_DECODE_FULL
|
1954 ((INTEL_DEBUG
& DEBUG_COLOR
) ? GEN_BATCH_DECODE_IN_COLOR
: 0) |
1955 GEN_BATCH_DECODE_OFFSETS
|
1956 GEN_BATCH_DECODE_FLOATS
;
1958 gen_batch_decode_ctx_init(&device
->decoder_ctx
,
1959 &physical_device
->info
,
1960 stderr
, decode_flags
, NULL
,
1961 decode_get_bo
, NULL
, device
);
1963 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1964 device
->instance
= physical_device
->instance
;
1965 device
->chipset_id
= physical_device
->chipset_id
;
1966 device
->no_hw
= physical_device
->no_hw
;
1967 device
->_lost
= false;
1970 device
->alloc
= *pAllocator
;
1972 device
->alloc
= physical_device
->instance
->alloc
;
1974 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
1975 device
->fd
= open(physical_device
->path
, O_RDWR
| O_CLOEXEC
);
1976 if (device
->fd
== -1) {
1977 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1981 device
->context_id
= anv_gem_create_context(device
);
1982 if (device
->context_id
== -1) {
1983 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1987 if (physical_device
->use_softpin
) {
1988 if (pthread_mutex_init(&device
->vma_mutex
, NULL
) != 0) {
1989 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1993 /* keep the page with address zero out of the allocator */
1994 struct anv_memory_heap
*low_heap
=
1995 &physical_device
->memory
.heaps
[physical_device
->memory
.heap_count
- 1];
1996 util_vma_heap_init(&device
->vma_lo
, low_heap
->vma_start
, low_heap
->vma_size
);
1997 device
->vma_lo_available
= low_heap
->size
;
1999 struct anv_memory_heap
*high_heap
=
2000 &physical_device
->memory
.heaps
[0];
2001 util_vma_heap_init(&device
->vma_hi
, high_heap
->vma_start
, high_heap
->vma_size
);
2002 device
->vma_hi_available
= physical_device
->memory
.heap_count
== 1 ? 0 :
2006 list_inithead(&device
->memory_objects
);
2008 /* As per spec, the driver implementation may deny requests to acquire
2009 * a priority above the default priority (MEDIUM) if the caller does not
2010 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
2013 if (physical_device
->has_context_priority
) {
2014 int err
= anv_gem_set_context_param(device
->fd
, device
->context_id
,
2015 I915_CONTEXT_PARAM_PRIORITY
,
2016 vk_priority_to_gen(priority
));
2017 if (err
!= 0 && priority
> VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
) {
2018 result
= vk_error(VK_ERROR_NOT_PERMITTED_EXT
);
2023 device
->info
= physical_device
->info
;
2024 device
->isl_dev
= physical_device
->isl_dev
;
2026 /* On Broadwell and later, we can use batch chaining to more efficiently
2027 * implement growing command buffers. Prior to Haswell, the kernel
2028 * command parser gets in the way and we have to fall back to growing
2031 device
->can_chain_batches
= device
->info
.gen
>= 8;
2033 device
->robust_buffer_access
= pCreateInfo
->pEnabledFeatures
&&
2034 pCreateInfo
->pEnabledFeatures
->robustBufferAccess
;
2035 device
->enabled_extensions
= enabled_extensions
;
2037 anv_device_init_dispatch(device
);
2039 if (pthread_mutex_init(&device
->mutex
, NULL
) != 0) {
2040 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2041 goto fail_context_id
;
2044 pthread_condattr_t condattr
;
2045 if (pthread_condattr_init(&condattr
) != 0) {
2046 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2049 if (pthread_condattr_setclock(&condattr
, CLOCK_MONOTONIC
) != 0) {
2050 pthread_condattr_destroy(&condattr
);
2051 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2054 if (pthread_cond_init(&device
->queue_submit
, &condattr
) != 0) {
2055 pthread_condattr_destroy(&condattr
);
2056 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2059 pthread_condattr_destroy(&condattr
);
2062 (physical_device
->supports_48bit_addresses
? EXEC_OBJECT_SUPPORTS_48B_ADDRESS
: 0) |
2063 (physical_device
->has_exec_async
? EXEC_OBJECT_ASYNC
: 0) |
2064 (physical_device
->has_exec_capture
? EXEC_OBJECT_CAPTURE
: 0) |
2065 (physical_device
->use_softpin
? EXEC_OBJECT_PINNED
: 0);
2067 anv_bo_pool_init(&device
->batch_bo_pool
, device
, bo_flags
);
2069 result
= anv_bo_cache_init(&device
->bo_cache
);
2070 if (result
!= VK_SUCCESS
)
2071 goto fail_batch_bo_pool
;
2073 if (!physical_device
->use_softpin
)
2074 bo_flags
&= ~EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
2076 result
= anv_state_pool_init(&device
->dynamic_state_pool
, device
,
2077 DYNAMIC_STATE_POOL_MIN_ADDRESS
,
2080 if (result
!= VK_SUCCESS
)
2083 result
= anv_state_pool_init(&device
->instruction_state_pool
, device
,
2084 INSTRUCTION_STATE_POOL_MIN_ADDRESS
,
2087 if (result
!= VK_SUCCESS
)
2088 goto fail_dynamic_state_pool
;
2090 result
= anv_state_pool_init(&device
->surface_state_pool
, device
,
2091 SURFACE_STATE_POOL_MIN_ADDRESS
,
2094 if (result
!= VK_SUCCESS
)
2095 goto fail_instruction_state_pool
;
2097 if (physical_device
->use_softpin
) {
2098 result
= anv_state_pool_init(&device
->binding_table_pool
, device
,
2099 BINDING_TABLE_POOL_MIN_ADDRESS
,
2102 if (result
!= VK_SUCCESS
)
2103 goto fail_surface_state_pool
;
2106 result
= anv_bo_init_new(&device
->workaround_bo
, device
, 1024);
2107 if (result
!= VK_SUCCESS
)
2108 goto fail_binding_table_pool
;
2110 if (physical_device
->use_softpin
)
2111 device
->workaround_bo
.flags
|= EXEC_OBJECT_PINNED
;
2113 if (!anv_vma_alloc(device
, &device
->workaround_bo
))
2114 goto fail_workaround_bo
;
2116 anv_device_init_trivial_batch(device
);
2118 if (device
->info
.gen
>= 10)
2119 anv_device_init_hiz_clear_value_bo(device
);
2121 anv_scratch_pool_init(device
, &device
->scratch_pool
);
2123 anv_queue_init(device
, &device
->queue
);
2125 switch (device
->info
.gen
) {
2127 if (!device
->info
.is_haswell
)
2128 result
= gen7_init_device_state(device
);
2130 result
= gen75_init_device_state(device
);
2133 result
= gen8_init_device_state(device
);
2136 result
= gen9_init_device_state(device
);
2139 result
= gen10_init_device_state(device
);
2142 result
= gen11_init_device_state(device
);
2145 /* Shouldn't get here as we don't create physical devices for any other
2147 unreachable("unhandled gen");
2149 if (result
!= VK_SUCCESS
)
2150 goto fail_workaround_bo
;
2152 anv_pipeline_cache_init(&device
->default_pipeline_cache
, device
, true);
2154 anv_device_init_blorp(device
);
2156 anv_device_init_border_colors(device
);
2158 *pDevice
= anv_device_to_handle(device
);
2163 anv_queue_finish(&device
->queue
);
2164 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
2165 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
2166 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
2167 fail_binding_table_pool
:
2168 if (physical_device
->use_softpin
)
2169 anv_state_pool_finish(&device
->binding_table_pool
);
2170 fail_surface_state_pool
:
2171 anv_state_pool_finish(&device
->surface_state_pool
);
2172 fail_instruction_state_pool
:
2173 anv_state_pool_finish(&device
->instruction_state_pool
);
2174 fail_dynamic_state_pool
:
2175 anv_state_pool_finish(&device
->dynamic_state_pool
);
2177 anv_bo_cache_finish(&device
->bo_cache
);
2179 anv_bo_pool_finish(&device
->batch_bo_pool
);
2180 pthread_cond_destroy(&device
->queue_submit
);
2182 pthread_mutex_destroy(&device
->mutex
);
2184 anv_gem_destroy_context(device
, device
->context_id
);
2188 vk_free(&device
->alloc
, device
);
2193 void anv_DestroyDevice(
2195 const VkAllocationCallbacks
* pAllocator
)
2197 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2198 struct anv_physical_device
*physical_device
;
2203 physical_device
= &device
->instance
->physicalDevice
;
2205 anv_device_finish_blorp(device
);
2207 anv_pipeline_cache_finish(&device
->default_pipeline_cache
);
2209 anv_queue_finish(&device
->queue
);
2211 #ifdef HAVE_VALGRIND
2212 /* We only need to free these to prevent valgrind errors. The backing
2213 * BO will go away in a couple of lines so we don't actually leak.
2215 anv_state_pool_free(&device
->dynamic_state_pool
, device
->border_colors
);
2218 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
2220 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
2221 anv_vma_free(device
, &device
->workaround_bo
);
2222 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
2224 anv_vma_free(device
, &device
->trivial_batch_bo
);
2225 anv_gem_close(device
, device
->trivial_batch_bo
.gem_handle
);
2226 if (device
->info
.gen
>= 10)
2227 anv_gem_close(device
, device
->hiz_clear_bo
.gem_handle
);
2229 if (physical_device
->use_softpin
)
2230 anv_state_pool_finish(&device
->binding_table_pool
);
2231 anv_state_pool_finish(&device
->surface_state_pool
);
2232 anv_state_pool_finish(&device
->instruction_state_pool
);
2233 anv_state_pool_finish(&device
->dynamic_state_pool
);
2235 anv_bo_cache_finish(&device
->bo_cache
);
2237 anv_bo_pool_finish(&device
->batch_bo_pool
);
2239 pthread_cond_destroy(&device
->queue_submit
);
2240 pthread_mutex_destroy(&device
->mutex
);
2242 anv_gem_destroy_context(device
, device
->context_id
);
2244 gen_batch_decode_ctx_finish(&device
->decoder_ctx
);
2248 vk_free(&device
->alloc
, device
);
2251 VkResult
anv_EnumerateInstanceLayerProperties(
2252 uint32_t* pPropertyCount
,
2253 VkLayerProperties
* pProperties
)
2255 if (pProperties
== NULL
) {
2256 *pPropertyCount
= 0;
2260 /* None supported at this time */
2261 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
2264 VkResult
anv_EnumerateDeviceLayerProperties(
2265 VkPhysicalDevice physicalDevice
,
2266 uint32_t* pPropertyCount
,
2267 VkLayerProperties
* pProperties
)
2269 if (pProperties
== NULL
) {
2270 *pPropertyCount
= 0;
2274 /* None supported at this time */
2275 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
2278 void anv_GetDeviceQueue(
2280 uint32_t queueNodeIndex
,
2281 uint32_t queueIndex
,
2284 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2286 assert(queueIndex
== 0);
2288 *pQueue
= anv_queue_to_handle(&device
->queue
);
2291 void anv_GetDeviceQueue2(
2293 const VkDeviceQueueInfo2
* pQueueInfo
,
2296 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2298 assert(pQueueInfo
->queueIndex
== 0);
2300 if (pQueueInfo
->flags
== device
->queue
.flags
)
2301 *pQueue
= anv_queue_to_handle(&device
->queue
);
2307 _anv_device_set_lost(struct anv_device
*device
,
2308 const char *file
, int line
,
2309 const char *msg
, ...)
2314 device
->_lost
= true;
2317 err
= __vk_errorv(device
->instance
, device
,
2318 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT
,
2319 VK_ERROR_DEVICE_LOST
, file
, line
, msg
, ap
);
2322 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
2329 anv_device_query_status(struct anv_device
*device
)
2331 /* This isn't likely as most of the callers of this function already check
2332 * for it. However, it doesn't hurt to check and it potentially lets us
2335 if (anv_device_is_lost(device
))
2336 return VK_ERROR_DEVICE_LOST
;
2338 uint32_t active
, pending
;
2339 int ret
= anv_gem_gpu_get_reset_stats(device
, &active
, &pending
);
2341 /* We don't know the real error. */
2342 return anv_device_set_lost(device
, "get_reset_stats failed: %m");
2346 return anv_device_set_lost(device
, "GPU hung on one of our command buffers");
2347 } else if (pending
) {
2348 return anv_device_set_lost(device
, "GPU hung with commands in-flight");
2355 anv_device_bo_busy(struct anv_device
*device
, struct anv_bo
*bo
)
2357 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
2358 * Other usages of the BO (such as on different hardware) will not be
2359 * flagged as "busy" by this ioctl. Use with care.
2361 int ret
= anv_gem_busy(device
, bo
->gem_handle
);
2363 return VK_NOT_READY
;
2364 } else if (ret
== -1) {
2365 /* We don't know the real error. */
2366 return anv_device_set_lost(device
, "gem wait failed: %m");
2369 /* Query for device status after the busy call. If the BO we're checking
2370 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
2371 * client because it clearly doesn't have valid data. Yes, this most
2372 * likely means an ioctl, but we just did an ioctl to query the busy status
2373 * so it's no great loss.
2375 return anv_device_query_status(device
);
2379 anv_device_wait(struct anv_device
*device
, struct anv_bo
*bo
,
2382 int ret
= anv_gem_wait(device
, bo
->gem_handle
, &timeout
);
2383 if (ret
== -1 && errno
== ETIME
) {
2385 } else if (ret
== -1) {
2386 /* We don't know the real error. */
2387 return anv_device_set_lost(device
, "gem wait failed: %m");
2390 /* Query for device status after the wait. If the BO we're waiting on got
2391 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
2392 * because it clearly doesn't have valid data. Yes, this most likely means
2393 * an ioctl, but we just did an ioctl to wait so it's no great loss.
2395 return anv_device_query_status(device
);
2398 VkResult
anv_DeviceWaitIdle(
2401 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2402 if (anv_device_is_lost(device
))
2403 return VK_ERROR_DEVICE_LOST
;
2405 struct anv_batch batch
;
2408 batch
.start
= batch
.next
= cmds
;
2409 batch
.end
= (void *) cmds
+ sizeof(cmds
);
2411 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
2412 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
2414 return anv_device_submit_simple_batch(device
, &batch
);
2418 anv_vma_alloc(struct anv_device
*device
, struct anv_bo
*bo
)
2420 if (!(bo
->flags
& EXEC_OBJECT_PINNED
))
2423 pthread_mutex_lock(&device
->vma_mutex
);
2427 if (bo
->flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
&&
2428 device
->vma_hi_available
>= bo
->size
) {
2429 uint64_t addr
= util_vma_heap_alloc(&device
->vma_hi
, bo
->size
, 4096);
2431 bo
->offset
= gen_canonical_address(addr
);
2432 assert(addr
== gen_48b_address(bo
->offset
));
2433 device
->vma_hi_available
-= bo
->size
;
2437 if (bo
->offset
== 0 && device
->vma_lo_available
>= bo
->size
) {
2438 uint64_t addr
= util_vma_heap_alloc(&device
->vma_lo
, bo
->size
, 4096);
2440 bo
->offset
= gen_canonical_address(addr
);
2441 assert(addr
== gen_48b_address(bo
->offset
));
2442 device
->vma_lo_available
-= bo
->size
;
2446 pthread_mutex_unlock(&device
->vma_mutex
);
2448 return bo
->offset
!= 0;
2452 anv_vma_free(struct anv_device
*device
, struct anv_bo
*bo
)
2454 if (!(bo
->flags
& EXEC_OBJECT_PINNED
))
2457 const uint64_t addr_48b
= gen_48b_address(bo
->offset
);
2459 pthread_mutex_lock(&device
->vma_mutex
);
2461 if (addr_48b
>= LOW_HEAP_MIN_ADDRESS
&&
2462 addr_48b
<= LOW_HEAP_MAX_ADDRESS
) {
2463 util_vma_heap_free(&device
->vma_lo
, addr_48b
, bo
->size
);
2464 device
->vma_lo_available
+= bo
->size
;
2466 MAYBE_UNUSED
const struct anv_physical_device
*physical_device
=
2467 &device
->instance
->physicalDevice
;
2468 assert(addr_48b
>= physical_device
->memory
.heaps
[0].vma_start
&&
2469 addr_48b
< (physical_device
->memory
.heaps
[0].vma_start
+
2470 physical_device
->memory
.heaps
[0].vma_size
));
2471 util_vma_heap_free(&device
->vma_hi
, addr_48b
, bo
->size
);
2472 device
->vma_hi_available
+= bo
->size
;
2475 pthread_mutex_unlock(&device
->vma_mutex
);
2481 anv_bo_init_new(struct anv_bo
*bo
, struct anv_device
*device
, uint64_t size
)
2483 uint32_t gem_handle
= anv_gem_create(device
, size
);
2485 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
2487 anv_bo_init(bo
, gem_handle
, size
);
2492 VkResult
anv_AllocateMemory(
2494 const VkMemoryAllocateInfo
* pAllocateInfo
,
2495 const VkAllocationCallbacks
* pAllocator
,
2496 VkDeviceMemory
* pMem
)
2498 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2499 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2500 struct anv_device_memory
*mem
;
2501 VkResult result
= VK_SUCCESS
;
2503 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
2505 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
2506 assert(pAllocateInfo
->allocationSize
> 0);
2508 if (pAllocateInfo
->allocationSize
> MAX_MEMORY_ALLOCATION_SIZE
)
2509 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
2511 /* FINISHME: Fail if allocation request exceeds heap size. */
2513 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
2514 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2516 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2518 assert(pAllocateInfo
->memoryTypeIndex
< pdevice
->memory
.type_count
);
2519 mem
->type
= &pdevice
->memory
.types
[pAllocateInfo
->memoryTypeIndex
];
2523 mem
->host_ptr
= NULL
;
2525 uint64_t bo_flags
= 0;
2527 assert(mem
->type
->heapIndex
< pdevice
->memory
.heap_count
);
2528 if (pdevice
->memory
.heaps
[mem
->type
->heapIndex
].supports_48bit_addresses
)
2529 bo_flags
|= EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
2531 const struct wsi_memory_allocate_info
*wsi_info
=
2532 vk_find_struct_const(pAllocateInfo
->pNext
, WSI_MEMORY_ALLOCATE_INFO_MESA
);
2533 if (wsi_info
&& wsi_info
->implicit_sync
) {
2534 /* We need to set the WRITE flag on window system buffers so that GEM
2535 * will know we're writing to them and synchronize uses on other rings
2536 * (eg if the display server uses the blitter ring).
2538 bo_flags
|= EXEC_OBJECT_WRITE
;
2539 } else if (pdevice
->has_exec_async
) {
2540 bo_flags
|= EXEC_OBJECT_ASYNC
;
2543 if (pdevice
->use_softpin
)
2544 bo_flags
|= EXEC_OBJECT_PINNED
;
2546 const VkExportMemoryAllocateInfo
*export_info
=
2547 vk_find_struct_const(pAllocateInfo
->pNext
, EXPORT_MEMORY_ALLOCATE_INFO
);
2549 /* Check if we need to support Android HW buffer export. If so,
2550 * create AHardwareBuffer and import memory from it.
2552 bool android_export
= false;
2553 if (export_info
&& export_info
->handleTypes
&
2554 VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID
)
2555 android_export
= true;
2557 /* Android memory import. */
2558 const struct VkImportAndroidHardwareBufferInfoANDROID
*ahw_import_info
=
2559 vk_find_struct_const(pAllocateInfo
->pNext
,
2560 IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID
);
2562 if (ahw_import_info
) {
2563 result
= anv_import_ahw_memory(_device
, mem
, ahw_import_info
);
2564 if (result
!= VK_SUCCESS
)
2568 } else if (android_export
) {
2569 result
= anv_create_ahw_memory(_device
, mem
, pAllocateInfo
);
2570 if (result
!= VK_SUCCESS
)
2573 const struct VkImportAndroidHardwareBufferInfoANDROID import_info
= {
2576 result
= anv_import_ahw_memory(_device
, mem
, &import_info
);
2577 if (result
!= VK_SUCCESS
)
2583 const VkImportMemoryFdInfoKHR
*fd_info
=
2584 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_FD_INFO_KHR
);
2586 /* The Vulkan spec permits handleType to be 0, in which case the struct is
2589 if (fd_info
&& fd_info
->handleType
) {
2590 /* At the moment, we support only the below handle types. */
2591 assert(fd_info
->handleType
==
2592 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
2593 fd_info
->handleType
==
2594 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
2596 result
= anv_bo_cache_import(device
, &device
->bo_cache
, fd_info
->fd
,
2597 bo_flags
| ANV_BO_EXTERNAL
, &mem
->bo
);
2598 if (result
!= VK_SUCCESS
)
2601 VkDeviceSize aligned_alloc_size
=
2602 align_u64(pAllocateInfo
->allocationSize
, 4096);
2604 /* For security purposes, we reject importing the bo if it's smaller
2605 * than the requested allocation size. This prevents a malicious client
2606 * from passing a buffer to a trusted client, lying about the size, and
2607 * telling the trusted client to try and texture from an image that goes
2608 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
2609 * in the trusted client. The trusted client can protect itself against
2610 * this sort of attack but only if it can trust the buffer size.
2612 if (mem
->bo
->size
< aligned_alloc_size
) {
2613 result
= vk_errorf(device
->instance
, device
,
2614 VK_ERROR_INVALID_EXTERNAL_HANDLE
,
2615 "aligned allocationSize too large for "
2616 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT: "
2617 "%"PRIu64
"B > %"PRIu64
"B",
2618 aligned_alloc_size
, mem
->bo
->size
);
2619 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
2623 /* From the Vulkan spec:
2625 * "Importing memory from a file descriptor transfers ownership of
2626 * the file descriptor from the application to the Vulkan
2627 * implementation. The application must not perform any operations on
2628 * the file descriptor after a successful import."
2630 * If the import fails, we leave the file descriptor open.
2636 const VkImportMemoryHostPointerInfoEXT
*host_ptr_info
=
2637 vk_find_struct_const(pAllocateInfo
->pNext
,
2638 IMPORT_MEMORY_HOST_POINTER_INFO_EXT
);
2639 if (host_ptr_info
&& host_ptr_info
->handleType
) {
2640 if (host_ptr_info
->handleType
==
2641 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_MAPPED_FOREIGN_MEMORY_BIT_EXT
) {
2642 result
= vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
2646 assert(host_ptr_info
->handleType
==
2647 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
);
2649 result
= anv_bo_cache_import_host_ptr(
2650 device
, &device
->bo_cache
, host_ptr_info
->pHostPointer
,
2651 pAllocateInfo
->allocationSize
, bo_flags
, &mem
->bo
);
2653 if (result
!= VK_SUCCESS
)
2656 mem
->host_ptr
= host_ptr_info
->pHostPointer
;
2660 /* Regular allocate (not importing memory). */
2662 if (export_info
&& export_info
->handleTypes
)
2663 bo_flags
|= ANV_BO_EXTERNAL
;
2665 result
= anv_bo_cache_alloc(device
, &device
->bo_cache
,
2666 pAllocateInfo
->allocationSize
, bo_flags
,
2668 if (result
!= VK_SUCCESS
)
2671 const VkMemoryDedicatedAllocateInfo
*dedicated_info
=
2672 vk_find_struct_const(pAllocateInfo
->pNext
, MEMORY_DEDICATED_ALLOCATE_INFO
);
2673 if (dedicated_info
&& dedicated_info
->image
!= VK_NULL_HANDLE
) {
2674 ANV_FROM_HANDLE(anv_image
, image
, dedicated_info
->image
);
2676 /* Some legacy (non-modifiers) consumers need the tiling to be set on
2677 * the BO. In this case, we have a dedicated allocation.
2679 if (image
->needs_set_tiling
) {
2680 const uint32_t i915_tiling
=
2681 isl_tiling_to_i915_tiling(image
->planes
[0].surface
.isl
.tiling
);
2682 int ret
= anv_gem_set_tiling(device
, mem
->bo
->gem_handle
,
2683 image
->planes
[0].surface
.isl
.row_pitch_B
,
2686 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
2687 return vk_errorf(device
->instance
, NULL
,
2688 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
2689 "failed to set BO tiling: %m");
2695 pthread_mutex_lock(&device
->mutex
);
2696 list_addtail(&mem
->link
, &device
->memory_objects
);
2697 pthread_mutex_unlock(&device
->mutex
);
2699 *pMem
= anv_device_memory_to_handle(mem
);
2704 vk_free2(&device
->alloc
, pAllocator
, mem
);
2709 VkResult
anv_GetMemoryFdKHR(
2711 const VkMemoryGetFdInfoKHR
* pGetFdInfo
,
2714 ANV_FROM_HANDLE(anv_device
, dev
, device_h
);
2715 ANV_FROM_HANDLE(anv_device_memory
, mem
, pGetFdInfo
->memory
);
2717 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
2719 assert(pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
2720 pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
2722 return anv_bo_cache_export(dev
, &dev
->bo_cache
, mem
->bo
, pFd
);
2725 VkResult
anv_GetMemoryFdPropertiesKHR(
2727 VkExternalMemoryHandleTypeFlagBits handleType
,
2729 VkMemoryFdPropertiesKHR
* pMemoryFdProperties
)
2731 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2732 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2734 switch (handleType
) {
2735 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
:
2736 /* dma-buf can be imported as any memory type */
2737 pMemoryFdProperties
->memoryTypeBits
=
2738 (1 << pdevice
->memory
.type_count
) - 1;
2742 /* The valid usage section for this function says:
2744 * "handleType must not be one of the handle types defined as
2747 * So opaque handle types fall into the default "unsupported" case.
2749 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
2753 VkResult
anv_GetMemoryHostPointerPropertiesEXT(
2755 VkExternalMemoryHandleTypeFlagBits handleType
,
2756 const void* pHostPointer
,
2757 VkMemoryHostPointerPropertiesEXT
* pMemoryHostPointerProperties
)
2759 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2761 assert(pMemoryHostPointerProperties
->sType
==
2762 VK_STRUCTURE_TYPE_MEMORY_HOST_POINTER_PROPERTIES_EXT
);
2764 switch (handleType
) {
2765 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
: {
2766 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2768 /* Host memory can be imported as any memory type. */
2769 pMemoryHostPointerProperties
->memoryTypeBits
=
2770 (1ull << pdevice
->memory
.type_count
) - 1;
2775 return VK_ERROR_INVALID_EXTERNAL_HANDLE
;
2779 void anv_FreeMemory(
2781 VkDeviceMemory _mem
,
2782 const VkAllocationCallbacks
* pAllocator
)
2784 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2785 ANV_FROM_HANDLE(anv_device_memory
, mem
, _mem
);
2790 pthread_mutex_lock(&device
->mutex
);
2791 list_del(&mem
->link
);
2792 pthread_mutex_unlock(&device
->mutex
);
2795 anv_UnmapMemory(_device
, _mem
);
2797 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
2799 #if defined(ANDROID) && ANDROID_API_LEVEL >= 26
2801 AHardwareBuffer_release(mem
->ahw
);
2804 vk_free2(&device
->alloc
, pAllocator
, mem
);
2807 VkResult
anv_MapMemory(
2809 VkDeviceMemory _memory
,
2810 VkDeviceSize offset
,
2812 VkMemoryMapFlags flags
,
2815 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2816 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
2823 if (mem
->host_ptr
) {
2824 *ppData
= mem
->host_ptr
+ offset
;
2828 if (size
== VK_WHOLE_SIZE
)
2829 size
= mem
->bo
->size
- offset
;
2831 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
2833 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
2834 * assert(size != 0);
2835 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
2836 * equal to the size of the memory minus offset
2839 assert(offset
+ size
<= mem
->bo
->size
);
2841 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
2842 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
2843 * at a time is valid. We could just mmap up front and return an offset
2844 * pointer here, but that may exhaust virtual memory on 32 bit
2847 uint32_t gem_flags
= 0;
2849 if (!device
->info
.has_llc
&&
2850 (mem
->type
->propertyFlags
& VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
))
2851 gem_flags
|= I915_MMAP_WC
;
2853 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
2854 uint64_t map_offset
= offset
& ~4095ull;
2855 assert(offset
>= map_offset
);
2856 uint64_t map_size
= (offset
+ size
) - map_offset
;
2858 /* Let's map whole pages */
2859 map_size
= align_u64(map_size
, 4096);
2861 void *map
= anv_gem_mmap(device
, mem
->bo
->gem_handle
,
2862 map_offset
, map_size
, gem_flags
);
2863 if (map
== MAP_FAILED
)
2864 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
2867 mem
->map_size
= map_size
;
2869 *ppData
= mem
->map
+ (offset
- map_offset
);
2874 void anv_UnmapMemory(
2876 VkDeviceMemory _memory
)
2878 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
2880 if (mem
== NULL
|| mem
->host_ptr
)
2883 anv_gem_munmap(mem
->map
, mem
->map_size
);
2890 clflush_mapped_ranges(struct anv_device
*device
,
2892 const VkMappedMemoryRange
*ranges
)
2894 for (uint32_t i
= 0; i
< count
; i
++) {
2895 ANV_FROM_HANDLE(anv_device_memory
, mem
, ranges
[i
].memory
);
2896 if (ranges
[i
].offset
>= mem
->map_size
)
2899 gen_clflush_range(mem
->map
+ ranges
[i
].offset
,
2900 MIN2(ranges
[i
].size
, mem
->map_size
- ranges
[i
].offset
));
2904 VkResult
anv_FlushMappedMemoryRanges(
2906 uint32_t memoryRangeCount
,
2907 const VkMappedMemoryRange
* pMemoryRanges
)
2909 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2911 if (device
->info
.has_llc
)
2914 /* Make sure the writes we're flushing have landed. */
2915 __builtin_ia32_mfence();
2917 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
2922 VkResult
anv_InvalidateMappedMemoryRanges(
2924 uint32_t memoryRangeCount
,
2925 const VkMappedMemoryRange
* pMemoryRanges
)
2927 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2929 if (device
->info
.has_llc
)
2932 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
2934 /* Make sure no reads get moved up above the invalidate. */
2935 __builtin_ia32_mfence();
2940 void anv_GetBufferMemoryRequirements(
2943 VkMemoryRequirements
* pMemoryRequirements
)
2945 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
2946 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2947 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2949 /* The Vulkan spec (git aaed022) says:
2951 * memoryTypeBits is a bitfield and contains one bit set for every
2952 * supported memory type for the resource. The bit `1<<i` is set if and
2953 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
2954 * structure for the physical device is supported.
2956 uint32_t memory_types
= 0;
2957 for (uint32_t i
= 0; i
< pdevice
->memory
.type_count
; i
++) {
2958 uint32_t valid_usage
= pdevice
->memory
.types
[i
].valid_buffer_usage
;
2959 if ((valid_usage
& buffer
->usage
) == buffer
->usage
)
2960 memory_types
|= (1u << i
);
2963 /* Base alignment requirement of a cache line */
2964 uint32_t alignment
= 16;
2966 /* We need an alignment of 32 for pushing UBOs */
2967 if (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
)
2968 alignment
= MAX2(alignment
, 32);
2970 pMemoryRequirements
->size
= buffer
->size
;
2971 pMemoryRequirements
->alignment
= alignment
;
2973 /* Storage and Uniform buffers should have their size aligned to
2974 * 32-bits to avoid boundary checks when last DWord is not complete.
2975 * This would ensure that not internal padding would be needed for
2978 if (device
->robust_buffer_access
&&
2979 (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
||
2980 buffer
->usage
& VK_BUFFER_USAGE_STORAGE_BUFFER_BIT
))
2981 pMemoryRequirements
->size
= align_u64(buffer
->size
, 4);
2983 pMemoryRequirements
->memoryTypeBits
= memory_types
;
2986 void anv_GetBufferMemoryRequirements2(
2988 const VkBufferMemoryRequirementsInfo2
* pInfo
,
2989 VkMemoryRequirements2
* pMemoryRequirements
)
2991 anv_GetBufferMemoryRequirements(_device
, pInfo
->buffer
,
2992 &pMemoryRequirements
->memoryRequirements
);
2994 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
2995 switch (ext
->sType
) {
2996 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
2997 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
2998 requirements
->prefersDedicatedAllocation
= false;
2999 requirements
->requiresDedicatedAllocation
= false;
3004 anv_debug_ignored_stype(ext
->sType
);
3010 void anv_GetImageMemoryRequirements(
3013 VkMemoryRequirements
* pMemoryRequirements
)
3015 ANV_FROM_HANDLE(anv_image
, image
, _image
);
3016 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3017 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3019 /* The Vulkan spec (git aaed022) says:
3021 * memoryTypeBits is a bitfield and contains one bit set for every
3022 * supported memory type for the resource. The bit `1<<i` is set if and
3023 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3024 * structure for the physical device is supported.
3026 * All types are currently supported for images.
3028 uint32_t memory_types
= (1ull << pdevice
->memory
.type_count
) - 1;
3030 /* We must have image allocated or imported at this point. According to the
3031 * specification, external images must have been bound to memory before
3032 * calling GetImageMemoryRequirements.
3034 assert(image
->size
> 0);
3036 pMemoryRequirements
->size
= image
->size
;
3037 pMemoryRequirements
->alignment
= image
->alignment
;
3038 pMemoryRequirements
->memoryTypeBits
= memory_types
;
3041 void anv_GetImageMemoryRequirements2(
3043 const VkImageMemoryRequirementsInfo2
* pInfo
,
3044 VkMemoryRequirements2
* pMemoryRequirements
)
3046 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3047 ANV_FROM_HANDLE(anv_image
, image
, pInfo
->image
);
3049 anv_GetImageMemoryRequirements(_device
, pInfo
->image
,
3050 &pMemoryRequirements
->memoryRequirements
);
3052 vk_foreach_struct_const(ext
, pInfo
->pNext
) {
3053 switch (ext
->sType
) {
3054 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO
: {
3055 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3056 const VkImagePlaneMemoryRequirementsInfo
*plane_reqs
=
3057 (const VkImagePlaneMemoryRequirementsInfo
*) ext
;
3058 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
,
3059 plane_reqs
->planeAspect
);
3061 assert(image
->planes
[plane
].offset
== 0);
3063 /* The Vulkan spec (git aaed022) says:
3065 * memoryTypeBits is a bitfield and contains one bit set for every
3066 * supported memory type for the resource. The bit `1<<i` is set
3067 * if and only if the memory type `i` in the
3068 * VkPhysicalDeviceMemoryProperties structure for the physical
3069 * device is supported.
3071 * All types are currently supported for images.
3073 pMemoryRequirements
->memoryRequirements
.memoryTypeBits
=
3074 (1ull << pdevice
->memory
.type_count
) - 1;
3076 /* We must have image allocated or imported at this point. According to the
3077 * specification, external images must have been bound to memory before
3078 * calling GetImageMemoryRequirements.
3080 assert(image
->planes
[plane
].size
> 0);
3082 pMemoryRequirements
->memoryRequirements
.size
= image
->planes
[plane
].size
;
3083 pMemoryRequirements
->memoryRequirements
.alignment
=
3084 image
->planes
[plane
].alignment
;
3089 anv_debug_ignored_stype(ext
->sType
);
3094 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3095 switch (ext
->sType
) {
3096 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
3097 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
3098 if (image
->needs_set_tiling
|| image
->external_format
) {
3099 /* If we need to set the tiling for external consumers, we need a
3100 * dedicated allocation.
3102 * See also anv_AllocateMemory.
3104 requirements
->prefersDedicatedAllocation
= true;
3105 requirements
->requiresDedicatedAllocation
= true;
3107 requirements
->prefersDedicatedAllocation
= false;
3108 requirements
->requiresDedicatedAllocation
= false;
3114 anv_debug_ignored_stype(ext
->sType
);
3120 void anv_GetImageSparseMemoryRequirements(
3123 uint32_t* pSparseMemoryRequirementCount
,
3124 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
3126 *pSparseMemoryRequirementCount
= 0;
3129 void anv_GetImageSparseMemoryRequirements2(
3131 const VkImageSparseMemoryRequirementsInfo2
* pInfo
,
3132 uint32_t* pSparseMemoryRequirementCount
,
3133 VkSparseImageMemoryRequirements2
* pSparseMemoryRequirements
)
3135 *pSparseMemoryRequirementCount
= 0;
3138 void anv_GetDeviceMemoryCommitment(
3140 VkDeviceMemory memory
,
3141 VkDeviceSize
* pCommittedMemoryInBytes
)
3143 *pCommittedMemoryInBytes
= 0;
3147 anv_bind_buffer_memory(const VkBindBufferMemoryInfo
*pBindInfo
)
3149 ANV_FROM_HANDLE(anv_device_memory
, mem
, pBindInfo
->memory
);
3150 ANV_FROM_HANDLE(anv_buffer
, buffer
, pBindInfo
->buffer
);
3152 assert(pBindInfo
->sType
== VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
);
3155 assert((buffer
->usage
& mem
->type
->valid_buffer_usage
) == buffer
->usage
);
3156 buffer
->address
= (struct anv_address
) {
3158 .offset
= pBindInfo
->memoryOffset
,
3161 buffer
->address
= ANV_NULL_ADDRESS
;
3165 VkResult
anv_BindBufferMemory(
3168 VkDeviceMemory memory
,
3169 VkDeviceSize memoryOffset
)
3171 anv_bind_buffer_memory(
3172 &(VkBindBufferMemoryInfo
) {
3173 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
,
3176 .memoryOffset
= memoryOffset
,
3182 VkResult
anv_BindBufferMemory2(
3184 uint32_t bindInfoCount
,
3185 const VkBindBufferMemoryInfo
* pBindInfos
)
3187 for (uint32_t i
= 0; i
< bindInfoCount
; i
++)
3188 anv_bind_buffer_memory(&pBindInfos
[i
]);
3193 VkResult
anv_QueueBindSparse(
3195 uint32_t bindInfoCount
,
3196 const VkBindSparseInfo
* pBindInfo
,
3199 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
3200 if (anv_device_is_lost(queue
->device
))
3201 return VK_ERROR_DEVICE_LOST
;
3203 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
3208 VkResult
anv_CreateEvent(
3210 const VkEventCreateInfo
* pCreateInfo
,
3211 const VkAllocationCallbacks
* pAllocator
,
3214 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3215 struct anv_state state
;
3216 struct anv_event
*event
;
3218 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_EVENT_CREATE_INFO
);
3220 state
= anv_state_pool_alloc(&device
->dynamic_state_pool
,
3223 event
->state
= state
;
3224 event
->semaphore
= VK_EVENT_RESET
;
3226 if (!device
->info
.has_llc
) {
3227 /* Make sure the writes we're flushing have landed. */
3228 __builtin_ia32_mfence();
3229 __builtin_ia32_clflush(event
);
3232 *pEvent
= anv_event_to_handle(event
);
3237 void anv_DestroyEvent(
3240 const VkAllocationCallbacks
* pAllocator
)
3242 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3243 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3248 anv_state_pool_free(&device
->dynamic_state_pool
, event
->state
);
3251 VkResult
anv_GetEventStatus(
3255 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3256 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3258 if (anv_device_is_lost(device
))
3259 return VK_ERROR_DEVICE_LOST
;
3261 if (!device
->info
.has_llc
) {
3262 /* Invalidate read cache before reading event written by GPU. */
3263 __builtin_ia32_clflush(event
);
3264 __builtin_ia32_mfence();
3268 return event
->semaphore
;
3271 VkResult
anv_SetEvent(
3275 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3276 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3278 event
->semaphore
= VK_EVENT_SET
;
3280 if (!device
->info
.has_llc
) {
3281 /* Make sure the writes we're flushing have landed. */
3282 __builtin_ia32_mfence();
3283 __builtin_ia32_clflush(event
);
3289 VkResult
anv_ResetEvent(
3293 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3294 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3296 event
->semaphore
= VK_EVENT_RESET
;
3298 if (!device
->info
.has_llc
) {
3299 /* Make sure the writes we're flushing have landed. */
3300 __builtin_ia32_mfence();
3301 __builtin_ia32_clflush(event
);
3309 VkResult
anv_CreateBuffer(
3311 const VkBufferCreateInfo
* pCreateInfo
,
3312 const VkAllocationCallbacks
* pAllocator
,
3315 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3316 struct anv_buffer
*buffer
;
3318 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
3320 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
3321 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3323 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3325 buffer
->size
= pCreateInfo
->size
;
3326 buffer
->usage
= pCreateInfo
->usage
;
3327 buffer
->address
= ANV_NULL_ADDRESS
;
3329 *pBuffer
= anv_buffer_to_handle(buffer
);
3334 void anv_DestroyBuffer(
3337 const VkAllocationCallbacks
* pAllocator
)
3339 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3340 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
3345 vk_free2(&device
->alloc
, pAllocator
, buffer
);
3348 VkDeviceAddress
anv_GetBufferDeviceAddressEXT(
3350 const VkBufferDeviceAddressInfoEXT
* pInfo
)
3352 ANV_FROM_HANDLE(anv_buffer
, buffer
, pInfo
->buffer
);
3354 assert(buffer
->address
.bo
->flags
& EXEC_OBJECT_PINNED
);
3356 return anv_address_physical(buffer
->address
);
3360 anv_fill_buffer_surface_state(struct anv_device
*device
, struct anv_state state
,
3361 enum isl_format format
,
3362 struct anv_address address
,
3363 uint32_t range
, uint32_t stride
)
3365 isl_buffer_fill_state(&device
->isl_dev
, state
.map
,
3366 .address
= anv_address_physical(address
),
3367 .mocs
= device
->default_mocs
,
3370 .swizzle
= ISL_SWIZZLE_IDENTITY
,
3371 .stride_B
= stride
);
3374 void anv_DestroySampler(
3377 const VkAllocationCallbacks
* pAllocator
)
3379 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3380 ANV_FROM_HANDLE(anv_sampler
, sampler
, _sampler
);
3385 vk_free2(&device
->alloc
, pAllocator
, sampler
);
3388 VkResult
anv_CreateFramebuffer(
3390 const VkFramebufferCreateInfo
* pCreateInfo
,
3391 const VkAllocationCallbacks
* pAllocator
,
3392 VkFramebuffer
* pFramebuffer
)
3394 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3395 struct anv_framebuffer
*framebuffer
;
3397 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
3399 size_t size
= sizeof(*framebuffer
) +
3400 sizeof(struct anv_image_view
*) * pCreateInfo
->attachmentCount
;
3401 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
3402 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3403 if (framebuffer
== NULL
)
3404 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3406 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
3407 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
3408 VkImageView _iview
= pCreateInfo
->pAttachments
[i
];
3409 framebuffer
->attachments
[i
] = anv_image_view_from_handle(_iview
);
3412 framebuffer
->width
= pCreateInfo
->width
;
3413 framebuffer
->height
= pCreateInfo
->height
;
3414 framebuffer
->layers
= pCreateInfo
->layers
;
3416 *pFramebuffer
= anv_framebuffer_to_handle(framebuffer
);
3421 void anv_DestroyFramebuffer(
3424 const VkAllocationCallbacks
* pAllocator
)
3426 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3427 ANV_FROM_HANDLE(anv_framebuffer
, fb
, _fb
);
3432 vk_free2(&device
->alloc
, pAllocator
, fb
);
3435 static const VkTimeDomainEXT anv_time_domains
[] = {
3436 VK_TIME_DOMAIN_DEVICE_EXT
,
3437 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
,
3438 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
,
3441 VkResult
anv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
3442 VkPhysicalDevice physicalDevice
,
3443 uint32_t *pTimeDomainCount
,
3444 VkTimeDomainEXT
*pTimeDomains
)
3447 VK_OUTARRAY_MAKE(out
, pTimeDomains
, pTimeDomainCount
);
3449 for (d
= 0; d
< ARRAY_SIZE(anv_time_domains
); d
++) {
3450 vk_outarray_append(&out
, i
) {
3451 *i
= anv_time_domains
[d
];
3455 return vk_outarray_status(&out
);
3459 anv_clock_gettime(clockid_t clock_id
)
3461 struct timespec current
;
3464 ret
= clock_gettime(clock_id
, ¤t
);
3465 if (ret
< 0 && clock_id
== CLOCK_MONOTONIC_RAW
)
3466 ret
= clock_gettime(CLOCK_MONOTONIC
, ¤t
);
3470 return (uint64_t) current
.tv_sec
* 1000000000ULL + current
.tv_nsec
;
3473 #define TIMESTAMP 0x2358
3475 VkResult
anv_GetCalibratedTimestampsEXT(
3477 uint32_t timestampCount
,
3478 const VkCalibratedTimestampInfoEXT
*pTimestampInfos
,
3479 uint64_t *pTimestamps
,
3480 uint64_t *pMaxDeviation
)
3482 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3483 uint64_t timestamp_frequency
= device
->info
.timestamp_frequency
;
3486 uint64_t begin
, end
;
3487 uint64_t max_clock_period
= 0;
3489 begin
= anv_clock_gettime(CLOCK_MONOTONIC_RAW
);
3491 for (d
= 0; d
< timestampCount
; d
++) {
3492 switch (pTimestampInfos
[d
].timeDomain
) {
3493 case VK_TIME_DOMAIN_DEVICE_EXT
:
3494 ret
= anv_gem_reg_read(device
, TIMESTAMP
| 1,
3498 return anv_device_set_lost(device
, "Failed to read the TIMESTAMP "
3501 uint64_t device_period
= DIV_ROUND_UP(1000000000, timestamp_frequency
);
3502 max_clock_period
= MAX2(max_clock_period
, device_period
);
3504 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
:
3505 pTimestamps
[d
] = anv_clock_gettime(CLOCK_MONOTONIC
);
3506 max_clock_period
= MAX2(max_clock_period
, 1);
3509 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
:
3510 pTimestamps
[d
] = begin
;
3518 end
= anv_clock_gettime(CLOCK_MONOTONIC_RAW
);
3521 * The maximum deviation is the sum of the interval over which we
3522 * perform the sampling and the maximum period of any sampled
3523 * clock. That's because the maximum skew between any two sampled
3524 * clock edges is when the sampled clock with the largest period is
3525 * sampled at the end of that period but right at the beginning of the
3526 * sampling interval and some other clock is sampled right at the
3527 * begining of its sampling period and right at the end of the
3528 * sampling interval. Let's assume the GPU has the longest clock
3529 * period and that the application is sampling GPU and monotonic:
3532 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
3533 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
3537 * GPU -----_____-----_____-----_____-----_____
3540 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
3541 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
3543 * Interval <----------------->
3544 * Deviation <-------------------------->
3548 * m = read(monotonic) 2
3551 * We round the sample interval up by one tick to cover sampling error
3552 * in the interval clock
3555 uint64_t sample_interval
= end
- begin
+ 1;
3557 *pMaxDeviation
= sample_interval
+ max_clock_period
;
3562 /* vk_icd.h does not declare this function, so we declare it here to
3563 * suppress Wmissing-prototypes.
3565 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
3566 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
);
3568 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
3569 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
)
3571 /* For the full details on loader interface versioning, see
3572 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
3573 * What follows is a condensed summary, to help you navigate the large and
3574 * confusing official doc.
3576 * - Loader interface v0 is incompatible with later versions. We don't
3579 * - In loader interface v1:
3580 * - The first ICD entrypoint called by the loader is
3581 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
3583 * - The ICD must statically expose no other Vulkan symbol unless it is
3584 * linked with -Bsymbolic.
3585 * - Each dispatchable Vulkan handle created by the ICD must be
3586 * a pointer to a struct whose first member is VK_LOADER_DATA. The
3587 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
3588 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
3589 * vkDestroySurfaceKHR(). The ICD must be capable of working with
3590 * such loader-managed surfaces.
3592 * - Loader interface v2 differs from v1 in:
3593 * - The first ICD entrypoint called by the loader is
3594 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
3595 * statically expose this entrypoint.
3597 * - Loader interface v3 differs from v2 in:
3598 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
3599 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
3600 * because the loader no longer does so.
3602 *pSupportedVersion
= MIN2(*pSupportedVersion
, 3u);