2 * Copyright © 2015 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
28 #include <sys/sysinfo.h>
32 #include "drm-uapi/drm_fourcc.h"
34 #include "anv_private.h"
35 #include "util/strtod.h"
36 #include "util/debug.h"
37 #include "util/build_id.h"
38 #include "util/disk_cache.h"
39 #include "util/mesa-sha1.h"
40 #include "util/os_file.h"
41 #include "util/u_atomic.h"
42 #include "util/u_string.h"
45 #include "common/gen_defines.h"
46 #include "compiler/glsl_types.h"
48 #include "genxml/gen7_pack.h"
50 /* This is probably far to big but it reflects the max size used for messages
51 * in OpenGLs KHR_debug.
53 #define MAX_DEBUG_MESSAGE_LENGTH 4096
56 compiler_debug_log(void *data
, const char *fmt
, ...)
58 char str
[MAX_DEBUG_MESSAGE_LENGTH
];
59 struct anv_device
*device
= (struct anv_device
*)data
;
61 if (list_empty(&device
->instance
->debug_report_callbacks
.callbacks
))
66 (void) vsnprintf(str
, MAX_DEBUG_MESSAGE_LENGTH
, fmt
, args
);
69 vk_debug_report(&device
->instance
->debug_report_callbacks
,
70 VK_DEBUG_REPORT_DEBUG_BIT_EXT
,
71 VK_DEBUG_REPORT_OBJECT_TYPE_UNKNOWN_EXT
,
76 compiler_perf_log(void *data
, const char *fmt
, ...)
81 if (unlikely(INTEL_DEBUG
& DEBUG_PERF
))
82 intel_logd_v(fmt
, args
);
88 anv_compute_heap_size(int fd
, uint64_t gtt_size
)
90 /* Query the total ram from the system */
94 uint64_t total_ram
= (uint64_t)info
.totalram
* (uint64_t)info
.mem_unit
;
96 /* We don't want to burn too much ram with the GPU. If the user has 4GiB
97 * or less, we use at most half. If they have more than 4GiB, we use 3/4.
99 uint64_t available_ram
;
100 if (total_ram
<= 4ull * 1024ull * 1024ull * 1024ull)
101 available_ram
= total_ram
/ 2;
103 available_ram
= total_ram
* 3 / 4;
105 /* We also want to leave some padding for things we allocate in the driver,
106 * so don't go over 3/4 of the GTT either.
108 uint64_t available_gtt
= gtt_size
* 3 / 4;
110 return MIN2(available_ram
, available_gtt
);
114 anv_physical_device_init_heaps(struct anv_physical_device
*device
, int fd
)
117 if (anv_gem_get_context_param(fd
, 0, I915_CONTEXT_PARAM_GTT_SIZE
,
119 /* If, for whatever reason, we can't actually get the GTT size from the
120 * kernel (too old?) fall back to the aperture size.
122 anv_perf_warn(NULL
, NULL
,
123 "Failed to get I915_CONTEXT_PARAM_GTT_SIZE: %m");
125 if (anv_gem_get_aperture(fd
, >t_size
) == -1) {
126 return vk_errorf(NULL
, NULL
, VK_ERROR_INITIALIZATION_FAILED
,
127 "failed to get aperture size: %m");
131 device
->supports_48bit_addresses
= (device
->info
.gen
>= 8) &&
132 gtt_size
> (4ULL << 30 /* GiB */);
134 uint64_t heap_size
= anv_compute_heap_size(fd
, gtt_size
);
136 if (heap_size
> (2ull << 30) && !device
->supports_48bit_addresses
) {
137 /* When running with an overridden PCI ID, we may get a GTT size from
138 * the kernel that is greater than 2 GiB but the execbuf check for 48bit
139 * address support can still fail. Just clamp the address space size to
140 * 2 GiB if we don't have 48-bit support.
142 intel_logw("%s:%d: The kernel reported a GTT size larger than 2 GiB but "
143 "not support for 48-bit addresses",
145 heap_size
= 2ull << 30;
148 if (heap_size
<= 3ull * (1ull << 30)) {
149 /* In this case, everything fits nicely into the 32-bit address space,
150 * so there's no need for supporting 48bit addresses on client-allocated
153 device
->memory
.heap_count
= 1;
154 device
->memory
.heaps
[0] = (struct anv_memory_heap
) {
155 .vma_start
= LOW_HEAP_MIN_ADDRESS
,
156 .vma_size
= LOW_HEAP_SIZE
,
158 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
159 .supports_48bit_addresses
= false,
162 /* Not everything will fit nicely into a 32-bit address space. In this
163 * case we need a 64-bit heap. Advertise a small 32-bit heap and a
164 * larger 48-bit heap. If we're in this case, then we have a total heap
165 * size larger than 3GiB which most likely means they have 8 GiB of
166 * video memory and so carving off 1 GiB for the 32-bit heap should be
169 const uint64_t heap_size_32bit
= 1ull << 30;
170 const uint64_t heap_size_48bit
= heap_size
- heap_size_32bit
;
172 assert(device
->supports_48bit_addresses
);
174 device
->memory
.heap_count
= 2;
175 device
->memory
.heaps
[0] = (struct anv_memory_heap
) {
176 .vma_start
= HIGH_HEAP_MIN_ADDRESS
,
177 /* Leave the last 4GiB out of the high vma range, so that no state
178 * base address + size can overflow 48 bits. For more information see
179 * the comment about Wa32bitGeneralStateOffset in anv_allocator.c
181 .vma_size
= gtt_size
- (1ull << 32) - HIGH_HEAP_MIN_ADDRESS
,
182 .size
= heap_size_48bit
,
183 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
184 .supports_48bit_addresses
= true,
186 device
->memory
.heaps
[1] = (struct anv_memory_heap
) {
187 .vma_start
= LOW_HEAP_MIN_ADDRESS
,
188 .vma_size
= LOW_HEAP_SIZE
,
189 .size
= heap_size_32bit
,
190 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
191 .supports_48bit_addresses
= false,
195 uint32_t type_count
= 0;
196 for (uint32_t heap
= 0; heap
< device
->memory
.heap_count
; heap
++) {
197 uint32_t valid_buffer_usage
= ~0;
199 /* There appears to be a hardware issue in the VF cache where it only
200 * considers the bottom 32 bits of memory addresses. If you happen to
201 * have two vertex buffers which get placed exactly 4 GiB apart and use
202 * them in back-to-back draw calls, you can get collisions. In order to
203 * solve this problem, we require vertex and index buffers be bound to
204 * memory allocated out of the 32-bit heap.
206 if (device
->memory
.heaps
[heap
].supports_48bit_addresses
) {
207 valid_buffer_usage
&= ~(VK_BUFFER_USAGE_INDEX_BUFFER_BIT
|
208 VK_BUFFER_USAGE_VERTEX_BUFFER_BIT
);
211 if (device
->info
.has_llc
) {
212 /* Big core GPUs share LLC with the CPU and thus one memory type can be
213 * both cached and coherent at the same time.
215 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
216 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
217 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
218 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
219 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
221 .valid_buffer_usage
= valid_buffer_usage
,
224 /* The spec requires that we expose a host-visible, coherent memory
225 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
226 * to give the application a choice between cached, but not coherent and
227 * coherent but uncached (WC though).
229 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
230 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
231 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
232 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
234 .valid_buffer_usage
= valid_buffer_usage
,
236 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
237 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
238 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
239 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
241 .valid_buffer_usage
= valid_buffer_usage
,
245 device
->memory
.type_count
= type_count
;
251 anv_physical_device_init_uuids(struct anv_physical_device
*device
)
253 const struct build_id_note
*note
=
254 build_id_find_nhdr_for_addr(anv_physical_device_init_uuids
);
256 return vk_errorf(device
->instance
, device
,
257 VK_ERROR_INITIALIZATION_FAILED
,
258 "Failed to find build-id");
261 unsigned build_id_len
= build_id_length(note
);
262 if (build_id_len
< 20) {
263 return vk_errorf(device
->instance
, device
,
264 VK_ERROR_INITIALIZATION_FAILED
,
265 "build-id too short. It needs to be a SHA");
268 memcpy(device
->driver_build_sha1
, build_id_data(note
), 20);
270 struct mesa_sha1 sha1_ctx
;
272 STATIC_ASSERT(VK_UUID_SIZE
<= sizeof(sha1
));
274 /* The pipeline cache UUID is used for determining when a pipeline cache is
275 * invalid. It needs both a driver build and the PCI ID of the device.
277 _mesa_sha1_init(&sha1_ctx
);
278 _mesa_sha1_update(&sha1_ctx
, build_id_data(note
), build_id_len
);
279 _mesa_sha1_update(&sha1_ctx
, &device
->chipset_id
,
280 sizeof(device
->chipset_id
));
281 _mesa_sha1_update(&sha1_ctx
, &device
->always_use_bindless
,
282 sizeof(device
->always_use_bindless
));
283 _mesa_sha1_update(&sha1_ctx
, &device
->has_a64_buffer_access
,
284 sizeof(device
->has_a64_buffer_access
));
285 _mesa_sha1_update(&sha1_ctx
, &device
->has_bindless_images
,
286 sizeof(device
->has_bindless_images
));
287 _mesa_sha1_update(&sha1_ctx
, &device
->has_bindless_samplers
,
288 sizeof(device
->has_bindless_samplers
));
289 _mesa_sha1_final(&sha1_ctx
, sha1
);
290 memcpy(device
->pipeline_cache_uuid
, sha1
, VK_UUID_SIZE
);
292 /* The driver UUID is used for determining sharability of images and memory
293 * between two Vulkan instances in separate processes. People who want to
294 * share memory need to also check the device UUID (below) so all this
295 * needs to be is the build-id.
297 memcpy(device
->driver_uuid
, build_id_data(note
), VK_UUID_SIZE
);
299 /* The device UUID uniquely identifies the given device within the machine.
300 * Since we never have more than one device, this doesn't need to be a real
301 * UUID. However, on the off-chance that someone tries to use this to
302 * cache pre-tiled images or something of the like, we use the PCI ID and
303 * some bits of ISL info to ensure that this is safe.
305 _mesa_sha1_init(&sha1_ctx
);
306 _mesa_sha1_update(&sha1_ctx
, &device
->chipset_id
,
307 sizeof(device
->chipset_id
));
308 _mesa_sha1_update(&sha1_ctx
, &device
->isl_dev
.has_bit6_swizzling
,
309 sizeof(device
->isl_dev
.has_bit6_swizzling
));
310 _mesa_sha1_final(&sha1_ctx
, sha1
);
311 memcpy(device
->device_uuid
, sha1
, VK_UUID_SIZE
);
317 anv_physical_device_init_disk_cache(struct anv_physical_device
*device
)
319 #ifdef ENABLE_SHADER_CACHE
321 ASSERTED
int len
= snprintf(renderer
, sizeof(renderer
), "anv_%04x",
323 assert(len
== sizeof(renderer
) - 2);
326 _mesa_sha1_format(timestamp
, device
->driver_build_sha1
);
328 const uint64_t driver_flags
=
329 brw_get_compiler_config_value(device
->compiler
);
330 device
->disk_cache
= disk_cache_create(renderer
, timestamp
, driver_flags
);
332 device
->disk_cache
= NULL
;
337 anv_physical_device_free_disk_cache(struct anv_physical_device
*device
)
339 #ifdef ENABLE_SHADER_CACHE
340 if (device
->disk_cache
)
341 disk_cache_destroy(device
->disk_cache
);
343 assert(device
->disk_cache
== NULL
);
348 get_available_system_memory()
350 char *meminfo
= os_read_file("/proc/meminfo");
354 char *str
= strstr(meminfo
, "MemAvailable:");
360 uint64_t kb_mem_available
;
361 if (sscanf(str
, "MemAvailable: %" PRIx64
, &kb_mem_available
) == 1) {
363 return kb_mem_available
<< 10;
371 anv_physical_device_init(struct anv_physical_device
*device
,
372 struct anv_instance
*instance
,
373 drmDevicePtr drm_device
)
375 const char *primary_path
= drm_device
->nodes
[DRM_NODE_PRIMARY
];
376 const char *path
= drm_device
->nodes
[DRM_NODE_RENDER
];
381 brw_process_intel_debug_variable();
383 fd
= open(path
, O_RDWR
| O_CLOEXEC
);
385 return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
387 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
388 device
->instance
= instance
;
390 assert(strlen(path
) < ARRAY_SIZE(device
->path
));
391 snprintf(device
->path
, ARRAY_SIZE(device
->path
), "%s", path
);
393 if (!gen_get_device_info_from_fd(fd
, &device
->info
)) {
394 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
397 device
->chipset_id
= device
->info
.chipset_id
;
398 device
->no_hw
= device
->info
.no_hw
;
400 if (getenv("INTEL_NO_HW") != NULL
)
401 device
->no_hw
= true;
403 device
->pci_info
.domain
= drm_device
->businfo
.pci
->domain
;
404 device
->pci_info
.bus
= drm_device
->businfo
.pci
->bus
;
405 device
->pci_info
.device
= drm_device
->businfo
.pci
->dev
;
406 device
->pci_info
.function
= drm_device
->businfo
.pci
->func
;
408 device
->name
= gen_get_device_name(device
->chipset_id
);
410 if (device
->info
.is_haswell
) {
411 intel_logw("Haswell Vulkan support is incomplete");
412 } else if (device
->info
.gen
== 7 && !device
->info
.is_baytrail
) {
413 intel_logw("Ivy Bridge Vulkan support is incomplete");
414 } else if (device
->info
.gen
== 7 && device
->info
.is_baytrail
) {
415 intel_logw("Bay Trail Vulkan support is incomplete");
416 } else if (device
->info
.gen
>= 8 && device
->info
.gen
<= 11) {
417 /* Gen8-11 fully supported */
419 result
= vk_errorf(device
->instance
, device
,
420 VK_ERROR_INCOMPATIBLE_DRIVER
,
421 "Vulkan not yet supported on %s", device
->name
);
425 device
->cmd_parser_version
= -1;
426 if (device
->info
.gen
== 7) {
427 device
->cmd_parser_version
=
428 anv_gem_get_param(fd
, I915_PARAM_CMD_PARSER_VERSION
);
429 if (device
->cmd_parser_version
== -1) {
430 result
= vk_errorf(device
->instance
, device
,
431 VK_ERROR_INITIALIZATION_FAILED
,
432 "failed to get command parser version");
437 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_WAIT_TIMEOUT
)) {
438 result
= vk_errorf(device
->instance
, device
,
439 VK_ERROR_INITIALIZATION_FAILED
,
440 "kernel missing gem wait");
444 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_EXECBUF2
)) {
445 result
= vk_errorf(device
->instance
, device
,
446 VK_ERROR_INITIALIZATION_FAILED
,
447 "kernel missing execbuf2");
451 if (!device
->info
.has_llc
&&
452 anv_gem_get_param(fd
, I915_PARAM_MMAP_VERSION
) < 1) {
453 result
= vk_errorf(device
->instance
, device
,
454 VK_ERROR_INITIALIZATION_FAILED
,
455 "kernel missing wc mmap");
459 result
= anv_physical_device_init_heaps(device
, fd
);
460 if (result
!= VK_SUCCESS
)
463 device
->has_exec_async
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_ASYNC
);
464 device
->has_exec_capture
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_CAPTURE
);
465 device
->has_exec_fence
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE
);
466 device
->has_syncobj
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE_ARRAY
);
467 device
->has_syncobj_wait
= device
->has_syncobj
&&
468 anv_gem_supports_syncobj_wait(fd
);
469 device
->has_context_priority
= anv_gem_has_context_priority(fd
);
471 device
->use_softpin
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_SOFTPIN
)
472 && device
->supports_48bit_addresses
;
474 device
->has_context_isolation
=
475 anv_gem_get_param(fd
, I915_PARAM_HAS_CONTEXT_ISOLATION
);
477 device
->always_use_bindless
=
478 env_var_as_boolean("ANV_ALWAYS_BINDLESS", false);
480 /* We first got the A64 messages on broadwell and we can only use them if
481 * we can pass addresses directly into the shader which requires softpin.
483 device
->has_a64_buffer_access
= device
->info
.gen
>= 8 &&
486 /* We first get bindless image access on Skylake and we can only really do
487 * it if we don't have any relocations so we need softpin.
489 device
->has_bindless_images
= device
->info
.gen
>= 9 &&
492 /* We've had bindless samplers since Ivy Bridge (forever in Vulkan terms)
493 * because it's just a matter of setting the sampler address in the sample
494 * message header. However, we've not bothered to wire it up for vec4 so
495 * we leave it disabled on gen7.
497 device
->has_bindless_samplers
= device
->info
.gen
>= 8;
499 device
->has_mem_available
= get_available_system_memory() != 0;
501 /* Starting with Gen10, the timestamp frequency of the command streamer may
502 * vary from one part to another. We can query the value from the kernel.
504 if (device
->info
.gen
>= 10) {
505 int timestamp_frequency
=
506 anv_gem_get_param(fd
, I915_PARAM_CS_TIMESTAMP_FREQUENCY
);
508 if (timestamp_frequency
< 0)
509 intel_logw("Kernel 4.16-rc1+ required to properly query CS timestamp frequency");
511 device
->info
.timestamp_frequency
= timestamp_frequency
;
514 /* GENs prior to 8 do not support EU/Subslice info */
515 if (device
->info
.gen
>= 8) {
516 device
->subslice_total
= anv_gem_get_param(fd
, I915_PARAM_SUBSLICE_TOTAL
);
517 device
->eu_total
= anv_gem_get_param(fd
, I915_PARAM_EU_TOTAL
);
519 /* Without this information, we cannot get the right Braswell
520 * brandstrings, and we have to use conservative numbers for GPGPU on
521 * many platforms, but otherwise, things will just work.
523 if (device
->subslice_total
< 1 || device
->eu_total
< 1) {
524 intel_logw("Kernel 4.1 required to properly query GPU properties");
526 } else if (device
->info
.gen
== 7) {
527 device
->subslice_total
= 1 << (device
->info
.gt
- 1);
530 if (device
->info
.is_cherryview
&&
531 device
->subslice_total
> 0 && device
->eu_total
> 0) {
532 /* Logical CS threads = EUs per subslice * num threads per EU */
533 uint32_t max_cs_threads
=
534 device
->eu_total
/ device
->subslice_total
* device
->info
.num_thread_per_eu
;
536 /* Fuse configurations may give more threads than expected, never less. */
537 if (max_cs_threads
> device
->info
.max_cs_threads
)
538 device
->info
.max_cs_threads
= max_cs_threads
;
541 device
->compiler
= brw_compiler_create(NULL
, &device
->info
);
542 if (device
->compiler
== NULL
) {
543 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
546 device
->compiler
->shader_debug_log
= compiler_debug_log
;
547 device
->compiler
->shader_perf_log
= compiler_perf_log
;
548 device
->compiler
->supports_pull_constants
= false;
549 device
->compiler
->constant_buffer_0_is_relative
=
550 device
->info
.gen
< 8 || !device
->has_context_isolation
;
551 device
->compiler
->supports_shader_constants
= true;
553 /* Broadwell PRM says:
555 * "Before Gen8, there was a historical configuration control field to
556 * swizzle address bit[6] for in X/Y tiling modes. This was set in three
557 * different places: TILECTL[1:0], ARB_MODE[5:4], and
558 * DISP_ARB_CTL[14:13].
560 * For Gen8 and subsequent generations, the swizzle fields are all
561 * reserved, and the CPU's memory controller performs all address
562 * swizzling modifications."
565 device
->info
.gen
< 8 && anv_gem_get_bit6_swizzle(fd
, I915_TILING_X
);
567 isl_device_init(&device
->isl_dev
, &device
->info
, swizzled
);
569 result
= anv_physical_device_init_uuids(device
);
570 if (result
!= VK_SUCCESS
)
573 anv_physical_device_init_disk_cache(device
);
575 if (instance
->enabled_extensions
.KHR_display
) {
576 master_fd
= open(primary_path
, O_RDWR
| O_CLOEXEC
);
577 if (master_fd
>= 0) {
578 /* prod the device with a GETPARAM call which will fail if
579 * we don't have permission to even render on this device
581 if (anv_gem_get_param(master_fd
, I915_PARAM_CHIPSET_ID
) == 0) {
587 device
->master_fd
= master_fd
;
589 result
= anv_init_wsi(device
);
590 if (result
!= VK_SUCCESS
) {
591 ralloc_free(device
->compiler
);
592 anv_physical_device_free_disk_cache(device
);
596 anv_physical_device_get_supported_extensions(device
,
597 &device
->supported_extensions
);
600 device
->local_fd
= fd
;
612 anv_physical_device_finish(struct anv_physical_device
*device
)
614 anv_finish_wsi(device
);
615 anv_physical_device_free_disk_cache(device
);
616 ralloc_free(device
->compiler
);
617 close(device
->local_fd
);
618 if (device
->master_fd
>= 0)
619 close(device
->master_fd
);
623 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
624 VkSystemAllocationScope allocationScope
)
630 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
631 size_t align
, VkSystemAllocationScope allocationScope
)
633 return realloc(pOriginal
, size
);
637 default_free_func(void *pUserData
, void *pMemory
)
642 static const VkAllocationCallbacks default_alloc
= {
644 .pfnAllocation
= default_alloc_func
,
645 .pfnReallocation
= default_realloc_func
,
646 .pfnFree
= default_free_func
,
649 VkResult
anv_EnumerateInstanceExtensionProperties(
650 const char* pLayerName
,
651 uint32_t* pPropertyCount
,
652 VkExtensionProperties
* pProperties
)
654 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
656 for (int i
= 0; i
< ANV_INSTANCE_EXTENSION_COUNT
; i
++) {
657 if (anv_instance_extensions_supported
.extensions
[i
]) {
658 vk_outarray_append(&out
, prop
) {
659 *prop
= anv_instance_extensions
[i
];
664 return vk_outarray_status(&out
);
667 VkResult
anv_CreateInstance(
668 const VkInstanceCreateInfo
* pCreateInfo
,
669 const VkAllocationCallbacks
* pAllocator
,
670 VkInstance
* pInstance
)
672 struct anv_instance
*instance
;
675 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
677 struct anv_instance_extension_table enabled_extensions
= {};
678 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
680 for (idx
= 0; idx
< ANV_INSTANCE_EXTENSION_COUNT
; idx
++) {
681 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
682 anv_instance_extensions
[idx
].extensionName
) == 0)
686 if (idx
>= ANV_INSTANCE_EXTENSION_COUNT
)
687 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
689 if (!anv_instance_extensions_supported
.extensions
[idx
])
690 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
692 enabled_extensions
.extensions
[idx
] = true;
695 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
696 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
698 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
700 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
703 instance
->alloc
= *pAllocator
;
705 instance
->alloc
= default_alloc
;
707 instance
->app_info
= (struct anv_app_info
) { .api_version
= 0 };
708 if (pCreateInfo
->pApplicationInfo
) {
709 const VkApplicationInfo
*app
= pCreateInfo
->pApplicationInfo
;
711 instance
->app_info
.app_name
=
712 vk_strdup(&instance
->alloc
, app
->pApplicationName
,
713 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
714 instance
->app_info
.app_version
= app
->applicationVersion
;
716 instance
->app_info
.engine_name
=
717 vk_strdup(&instance
->alloc
, app
->pEngineName
,
718 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
719 instance
->app_info
.engine_version
= app
->engineVersion
;
721 instance
->app_info
.api_version
= app
->apiVersion
;
724 if (instance
->app_info
.api_version
== 0)
725 instance
->app_info
.api_version
= VK_API_VERSION_1_0
;
727 instance
->enabled_extensions
= enabled_extensions
;
729 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->dispatch
.entrypoints
); i
++) {
730 /* Vulkan requires that entrypoints for extensions which have not been
731 * enabled must not be advertised.
733 if (!anv_instance_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
734 &instance
->enabled_extensions
)) {
735 instance
->dispatch
.entrypoints
[i
] = NULL
;
737 instance
->dispatch
.entrypoints
[i
] =
738 anv_instance_dispatch_table
.entrypoints
[i
];
742 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->device_dispatch
.entrypoints
); i
++) {
743 /* Vulkan requires that entrypoints for extensions which have not been
744 * enabled must not be advertised.
746 if (!anv_device_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
747 &instance
->enabled_extensions
, NULL
)) {
748 instance
->device_dispatch
.entrypoints
[i
] = NULL
;
750 instance
->device_dispatch
.entrypoints
[i
] =
751 anv_device_dispatch_table
.entrypoints
[i
];
755 instance
->physicalDeviceCount
= -1;
757 result
= vk_debug_report_instance_init(&instance
->debug_report_callbacks
);
758 if (result
!= VK_SUCCESS
) {
759 vk_free2(&default_alloc
, pAllocator
, instance
);
760 return vk_error(result
);
763 instance
->pipeline_cache_enabled
=
764 env_var_as_boolean("ANV_ENABLE_PIPELINE_CACHE", true);
767 glsl_type_singleton_init_or_ref();
769 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
771 *pInstance
= anv_instance_to_handle(instance
);
776 void anv_DestroyInstance(
777 VkInstance _instance
,
778 const VkAllocationCallbacks
* pAllocator
)
780 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
785 if (instance
->physicalDeviceCount
> 0) {
786 /* We support at most one physical device. */
787 assert(instance
->physicalDeviceCount
== 1);
788 anv_physical_device_finish(&instance
->physicalDevice
);
791 vk_free(&instance
->alloc
, (char *)instance
->app_info
.app_name
);
792 vk_free(&instance
->alloc
, (char *)instance
->app_info
.engine_name
);
794 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
796 vk_debug_report_instance_destroy(&instance
->debug_report_callbacks
);
798 glsl_type_singleton_decref();
801 vk_free(&instance
->alloc
, instance
);
805 anv_enumerate_devices(struct anv_instance
*instance
)
807 /* TODO: Check for more devices ? */
808 drmDevicePtr devices
[8];
809 VkResult result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
812 instance
->physicalDeviceCount
= 0;
814 max_devices
= drmGetDevices2(0, devices
, ARRAY_SIZE(devices
));
816 return VK_ERROR_INCOMPATIBLE_DRIVER
;
818 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
819 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
820 devices
[i
]->bustype
== DRM_BUS_PCI
&&
821 devices
[i
]->deviceinfo
.pci
->vendor_id
== 0x8086) {
823 result
= anv_physical_device_init(&instance
->physicalDevice
,
824 instance
, devices
[i
]);
825 if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
829 drmFreeDevices(devices
, max_devices
);
831 if (result
== VK_SUCCESS
)
832 instance
->physicalDeviceCount
= 1;
838 anv_instance_ensure_physical_device(struct anv_instance
*instance
)
840 if (instance
->physicalDeviceCount
< 0) {
841 VkResult result
= anv_enumerate_devices(instance
);
842 if (result
!= VK_SUCCESS
&&
843 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
850 VkResult
anv_EnumeratePhysicalDevices(
851 VkInstance _instance
,
852 uint32_t* pPhysicalDeviceCount
,
853 VkPhysicalDevice
* pPhysicalDevices
)
855 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
856 VK_OUTARRAY_MAKE(out
, pPhysicalDevices
, pPhysicalDeviceCount
);
858 VkResult result
= anv_instance_ensure_physical_device(instance
);
859 if (result
!= VK_SUCCESS
)
862 if (instance
->physicalDeviceCount
== 0)
865 assert(instance
->physicalDeviceCount
== 1);
866 vk_outarray_append(&out
, i
) {
867 *i
= anv_physical_device_to_handle(&instance
->physicalDevice
);
870 return vk_outarray_status(&out
);
873 VkResult
anv_EnumeratePhysicalDeviceGroups(
874 VkInstance _instance
,
875 uint32_t* pPhysicalDeviceGroupCount
,
876 VkPhysicalDeviceGroupProperties
* pPhysicalDeviceGroupProperties
)
878 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
879 VK_OUTARRAY_MAKE(out
, pPhysicalDeviceGroupProperties
,
880 pPhysicalDeviceGroupCount
);
882 VkResult result
= anv_instance_ensure_physical_device(instance
);
883 if (result
!= VK_SUCCESS
)
886 if (instance
->physicalDeviceCount
== 0)
889 assert(instance
->physicalDeviceCount
== 1);
891 vk_outarray_append(&out
, p
) {
892 p
->physicalDeviceCount
= 1;
893 memset(p
->physicalDevices
, 0, sizeof(p
->physicalDevices
));
894 p
->physicalDevices
[0] =
895 anv_physical_device_to_handle(&instance
->physicalDevice
);
896 p
->subsetAllocation
= false;
898 vk_foreach_struct(ext
, p
->pNext
)
899 anv_debug_ignored_stype(ext
->sType
);
902 return vk_outarray_status(&out
);
905 void anv_GetPhysicalDeviceFeatures(
906 VkPhysicalDevice physicalDevice
,
907 VkPhysicalDeviceFeatures
* pFeatures
)
909 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
911 *pFeatures
= (VkPhysicalDeviceFeatures
) {
912 .robustBufferAccess
= true,
913 .fullDrawIndexUint32
= true,
914 .imageCubeArray
= true,
915 .independentBlend
= true,
916 .geometryShader
= true,
917 .tessellationShader
= true,
918 .sampleRateShading
= true,
919 .dualSrcBlend
= true,
921 .multiDrawIndirect
= true,
922 .drawIndirectFirstInstance
= true,
924 .depthBiasClamp
= true,
925 .fillModeNonSolid
= true,
926 .depthBounds
= false,
930 .multiViewport
= true,
931 .samplerAnisotropy
= true,
932 .textureCompressionETC2
= pdevice
->info
.gen
>= 8 ||
933 pdevice
->info
.is_baytrail
,
934 .textureCompressionASTC_LDR
= pdevice
->info
.gen
>= 9, /* FINISHME CHV */
935 .textureCompressionBC
= true,
936 .occlusionQueryPrecise
= true,
937 .pipelineStatisticsQuery
= true,
938 .fragmentStoresAndAtomics
= true,
939 .shaderTessellationAndGeometryPointSize
= true,
940 .shaderImageGatherExtended
= true,
941 .shaderStorageImageExtendedFormats
= true,
942 .shaderStorageImageMultisample
= false,
943 .shaderStorageImageReadWithoutFormat
= false,
944 .shaderStorageImageWriteWithoutFormat
= true,
945 .shaderUniformBufferArrayDynamicIndexing
= true,
946 .shaderSampledImageArrayDynamicIndexing
= true,
947 .shaderStorageBufferArrayDynamicIndexing
= true,
948 .shaderStorageImageArrayDynamicIndexing
= true,
949 .shaderClipDistance
= true,
950 .shaderCullDistance
= true,
951 .shaderFloat64
= pdevice
->info
.gen
>= 8 &&
952 pdevice
->info
.has_64bit_types
,
953 .shaderInt64
= pdevice
->info
.gen
>= 8 &&
954 pdevice
->info
.has_64bit_types
,
955 .shaderInt16
= pdevice
->info
.gen
>= 8,
956 .shaderResourceMinLod
= pdevice
->info
.gen
>= 9,
957 .variableMultisampleRate
= true,
958 .inheritedQueries
= true,
961 /* We can't do image stores in vec4 shaders */
962 pFeatures
->vertexPipelineStoresAndAtomics
=
963 pdevice
->compiler
->scalar_stage
[MESA_SHADER_VERTEX
] &&
964 pdevice
->compiler
->scalar_stage
[MESA_SHADER_GEOMETRY
];
966 struct anv_app_info
*app_info
= &pdevice
->instance
->app_info
;
968 /* The new DOOM and Wolfenstein games require depthBounds without
969 * checking for it. They seem to run fine without it so just claim it's
970 * there and accept the consequences.
972 if (app_info
->engine_name
&& strcmp(app_info
->engine_name
, "idTech") == 0)
973 pFeatures
->depthBounds
= true;
976 void anv_GetPhysicalDeviceFeatures2(
977 VkPhysicalDevice physicalDevice
,
978 VkPhysicalDeviceFeatures2
* pFeatures
)
980 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
981 anv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
983 vk_foreach_struct(ext
, pFeatures
->pNext
) {
984 switch (ext
->sType
) {
985 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES_KHR
: {
986 VkPhysicalDevice8BitStorageFeaturesKHR
*features
=
987 (VkPhysicalDevice8BitStorageFeaturesKHR
*)ext
;
988 features
->storageBuffer8BitAccess
= pdevice
->info
.gen
>= 8;
989 features
->uniformAndStorageBuffer8BitAccess
= pdevice
->info
.gen
>= 8;
990 features
->storagePushConstant8
= pdevice
->info
.gen
>= 8;
994 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES
: {
995 VkPhysicalDevice16BitStorageFeatures
*features
=
996 (VkPhysicalDevice16BitStorageFeatures
*)ext
;
997 features
->storageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
998 features
->uniformAndStorageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
999 features
->storagePushConstant16
= pdevice
->info
.gen
>= 8;
1000 features
->storageInputOutput16
= false;
1004 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_EXT
: {
1005 VkPhysicalDeviceBufferDeviceAddressFeaturesEXT
*features
= (void *)ext
;
1006 features
->bufferDeviceAddress
= pdevice
->has_a64_buffer_access
;
1007 features
->bufferDeviceAddressCaptureReplay
= false;
1008 features
->bufferDeviceAddressMultiDevice
= false;
1012 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COMPUTE_SHADER_DERIVATIVES_FEATURES_NV
: {
1013 VkPhysicalDeviceComputeShaderDerivativesFeaturesNV
*features
=
1014 (VkPhysicalDeviceComputeShaderDerivativesFeaturesNV
*)ext
;
1015 features
->computeDerivativeGroupQuads
= true;
1016 features
->computeDerivativeGroupLinear
= true;
1020 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT
: {
1021 VkPhysicalDeviceConditionalRenderingFeaturesEXT
*features
=
1022 (VkPhysicalDeviceConditionalRenderingFeaturesEXT
*)ext
;
1023 features
->conditionalRendering
= pdevice
->info
.gen
>= 8 ||
1024 pdevice
->info
.is_haswell
;
1025 features
->inheritedConditionalRendering
= pdevice
->info
.gen
>= 8 ||
1026 pdevice
->info
.is_haswell
;
1030 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_CLIP_ENABLE_FEATURES_EXT
: {
1031 VkPhysicalDeviceDepthClipEnableFeaturesEXT
*features
=
1032 (VkPhysicalDeviceDepthClipEnableFeaturesEXT
*)ext
;
1033 features
->depthClipEnable
= true;
1037 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR
: {
1038 VkPhysicalDeviceFloat16Int8FeaturesKHR
*features
= (void *)ext
;
1039 features
->shaderFloat16
= pdevice
->info
.gen
>= 8;
1040 features
->shaderInt8
= pdevice
->info
.gen
>= 8;
1044 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FRAGMENT_SHADER_INTERLOCK_FEATURES_EXT
: {
1045 VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT
*features
=
1046 (VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT
*)ext
;
1047 features
->fragmentShaderSampleInterlock
= pdevice
->info
.gen
>= 9;
1048 features
->fragmentShaderPixelInterlock
= pdevice
->info
.gen
>= 9;
1049 features
->fragmentShaderShadingRateInterlock
= false;
1053 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_QUERY_RESET_FEATURES_EXT
: {
1054 VkPhysicalDeviceHostQueryResetFeaturesEXT
*features
=
1055 (VkPhysicalDeviceHostQueryResetFeaturesEXT
*)ext
;
1056 features
->hostQueryReset
= true;
1060 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT
: {
1061 VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*features
=
1062 (VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*)ext
;
1063 features
->shaderInputAttachmentArrayDynamicIndexing
= false;
1064 features
->shaderUniformTexelBufferArrayDynamicIndexing
= true;
1065 features
->shaderStorageTexelBufferArrayDynamicIndexing
= true;
1066 features
->shaderUniformBufferArrayNonUniformIndexing
= false;
1067 features
->shaderSampledImageArrayNonUniformIndexing
= true;
1068 features
->shaderStorageBufferArrayNonUniformIndexing
= true;
1069 features
->shaderStorageImageArrayNonUniformIndexing
= true;
1070 features
->shaderInputAttachmentArrayNonUniformIndexing
= false;
1071 features
->shaderUniformTexelBufferArrayNonUniformIndexing
= true;
1072 features
->shaderStorageTexelBufferArrayNonUniformIndexing
= true;
1073 features
->descriptorBindingUniformBufferUpdateAfterBind
= false;
1074 features
->descriptorBindingSampledImageUpdateAfterBind
= true;
1075 features
->descriptorBindingStorageImageUpdateAfterBind
= true;
1076 features
->descriptorBindingStorageBufferUpdateAfterBind
= true;
1077 features
->descriptorBindingUniformTexelBufferUpdateAfterBind
= true;
1078 features
->descriptorBindingStorageTexelBufferUpdateAfterBind
= true;
1079 features
->descriptorBindingUpdateUnusedWhilePending
= true;
1080 features
->descriptorBindingPartiallyBound
= true;
1081 features
->descriptorBindingVariableDescriptorCount
= false;
1082 features
->runtimeDescriptorArray
= true;
1086 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INDEX_TYPE_UINT8_FEATURES_EXT
: {
1087 VkPhysicalDeviceIndexTypeUint8FeaturesEXT
*features
=
1088 (VkPhysicalDeviceIndexTypeUint8FeaturesEXT
*)ext
;
1089 features
->indexTypeUint8
= true;
1093 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_FEATURES_EXT
: {
1094 VkPhysicalDeviceInlineUniformBlockFeaturesEXT
*features
=
1095 (VkPhysicalDeviceInlineUniformBlockFeaturesEXT
*)ext
;
1096 features
->inlineUniformBlock
= true;
1097 features
->descriptorBindingInlineUniformBlockUpdateAfterBind
= true;
1101 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_FEATURES_EXT
: {
1102 VkPhysicalDeviceLineRasterizationFeaturesEXT
*features
=
1103 (VkPhysicalDeviceLineRasterizationFeaturesEXT
*)ext
;
1104 features
->rectangularLines
= true;
1105 features
->bresenhamLines
= true;
1106 features
->smoothLines
= true;
1107 features
->stippledRectangularLines
= false;
1108 features
->stippledBresenhamLines
= true;
1109 features
->stippledSmoothLines
= false;
1112 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES
: {
1113 VkPhysicalDeviceMultiviewFeatures
*features
=
1114 (VkPhysicalDeviceMultiviewFeatures
*)ext
;
1115 features
->multiview
= true;
1116 features
->multiviewGeometryShader
= true;
1117 features
->multiviewTessellationShader
= true;
1121 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGELESS_FRAMEBUFFER_FEATURES_KHR
: {
1122 VkPhysicalDeviceImagelessFramebufferFeaturesKHR
*features
=
1123 (VkPhysicalDeviceImagelessFramebufferFeaturesKHR
*)ext
;
1124 features
->imagelessFramebuffer
= true;
1128 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES
: {
1129 VkPhysicalDeviceProtectedMemoryFeatures
*features
= (void *)ext
;
1130 features
->protectedMemory
= false;
1134 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES
: {
1135 VkPhysicalDeviceSamplerYcbcrConversionFeatures
*features
=
1136 (VkPhysicalDeviceSamplerYcbcrConversionFeatures
*) ext
;
1137 features
->samplerYcbcrConversion
= true;
1141 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES_EXT
: {
1142 VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*features
=
1143 (VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*)ext
;
1144 features
->scalarBlockLayout
= true;
1148 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_INT64_FEATURES_KHR
: {
1149 VkPhysicalDeviceShaderAtomicInt64FeaturesKHR
*features
= (void *)ext
;
1150 features
->shaderBufferInt64Atomics
=
1151 pdevice
->info
.gen
>= 9 && pdevice
->use_softpin
;
1152 features
->shaderSharedInt64Atomics
= VK_FALSE
;
1156 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DEMOTE_TO_HELPER_INVOCATION_FEATURES_EXT
: {
1157 VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT
*features
= (void *)ext
;
1158 features
->shaderDemoteToHelperInvocation
= true;
1162 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETERS_FEATURES
: {
1163 VkPhysicalDeviceShaderDrawParametersFeatures
*features
= (void *)ext
;
1164 features
->shaderDrawParameters
= true;
1168 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_FEATURES_EXT
: {
1169 VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT
*features
=
1170 (VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT
*)ext
;
1171 features
->texelBufferAlignment
= true;
1175 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES
: {
1176 VkPhysicalDeviceVariablePointersFeatures
*features
= (void *)ext
;
1177 features
->variablePointersStorageBuffer
= true;
1178 features
->variablePointers
= true;
1182 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT
: {
1183 VkPhysicalDeviceTransformFeedbackFeaturesEXT
*features
=
1184 (VkPhysicalDeviceTransformFeedbackFeaturesEXT
*)ext
;
1185 features
->transformFeedback
= true;
1186 features
->geometryStreams
= true;
1190 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_UNIFORM_BUFFER_STANDARD_LAYOUT_FEATURES_KHR
: {
1191 VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR
*features
=
1192 (VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR
*)ext
;
1193 features
->uniformBufferStandardLayout
= true;
1197 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT
: {
1198 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*features
=
1199 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*)ext
;
1200 features
->vertexAttributeInstanceRateDivisor
= true;
1201 features
->vertexAttributeInstanceRateZeroDivisor
= true;
1205 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_YCBCR_IMAGE_ARRAYS_FEATURES_EXT
: {
1206 VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*features
=
1207 (VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*)ext
;
1208 features
->ycbcrImageArrays
= true;
1213 anv_debug_ignored_stype(ext
->sType
);
1219 #define MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS 64
1221 #define MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS 64
1222 #define MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS 256
1224 void anv_GetPhysicalDeviceProperties(
1225 VkPhysicalDevice physicalDevice
,
1226 VkPhysicalDeviceProperties
* pProperties
)
1228 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1229 const struct gen_device_info
*devinfo
= &pdevice
->info
;
1231 /* See assertions made when programming the buffer surface state. */
1232 const uint32_t max_raw_buffer_sz
= devinfo
->gen
>= 7 ?
1233 (1ul << 30) : (1ul << 27);
1235 const uint32_t max_ssbos
= pdevice
->has_a64_buffer_access
? UINT16_MAX
: 64;
1236 const uint32_t max_textures
=
1237 pdevice
->has_bindless_images
? UINT16_MAX
: 128;
1238 const uint32_t max_samplers
=
1239 pdevice
->has_bindless_samplers
? UINT16_MAX
:
1240 (devinfo
->gen
>= 8 || devinfo
->is_haswell
) ? 128 : 16;
1241 const uint32_t max_images
=
1242 pdevice
->has_bindless_images
? UINT16_MAX
: MAX_IMAGES
;
1244 /* The moment we have anything bindless, claim a high per-stage limit */
1245 const uint32_t max_per_stage
=
1246 pdevice
->has_a64_buffer_access
? UINT32_MAX
:
1247 MAX_BINDING_TABLE_SIZE
- MAX_RTS
;
1249 VkSampleCountFlags sample_counts
=
1250 isl_device_get_sample_counts(&pdevice
->isl_dev
);
1253 VkPhysicalDeviceLimits limits
= {
1254 .maxImageDimension1D
= (1 << 14),
1255 .maxImageDimension2D
= (1 << 14),
1256 .maxImageDimension3D
= (1 << 11),
1257 .maxImageDimensionCube
= (1 << 14),
1258 .maxImageArrayLayers
= (1 << 11),
1259 .maxTexelBufferElements
= 128 * 1024 * 1024,
1260 .maxUniformBufferRange
= (1ul << 27),
1261 .maxStorageBufferRange
= max_raw_buffer_sz
,
1262 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
1263 .maxMemoryAllocationCount
= UINT32_MAX
,
1264 .maxSamplerAllocationCount
= 64 * 1024,
1265 .bufferImageGranularity
= 64, /* A cache line */
1266 .sparseAddressSpaceSize
= 0,
1267 .maxBoundDescriptorSets
= MAX_SETS
,
1268 .maxPerStageDescriptorSamplers
= max_samplers
,
1269 .maxPerStageDescriptorUniformBuffers
= MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
,
1270 .maxPerStageDescriptorStorageBuffers
= max_ssbos
,
1271 .maxPerStageDescriptorSampledImages
= max_textures
,
1272 .maxPerStageDescriptorStorageImages
= max_images
,
1273 .maxPerStageDescriptorInputAttachments
= MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS
,
1274 .maxPerStageResources
= max_per_stage
,
1275 .maxDescriptorSetSamplers
= 6 * max_samplers
, /* number of stages * maxPerStageDescriptorSamplers */
1276 .maxDescriptorSetUniformBuffers
= 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
, /* number of stages * maxPerStageDescriptorUniformBuffers */
1277 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
1278 .maxDescriptorSetStorageBuffers
= 6 * max_ssbos
, /* number of stages * maxPerStageDescriptorStorageBuffers */
1279 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
1280 .maxDescriptorSetSampledImages
= 6 * max_textures
, /* number of stages * maxPerStageDescriptorSampledImages */
1281 .maxDescriptorSetStorageImages
= 6 * max_images
, /* number of stages * maxPerStageDescriptorStorageImages */
1282 .maxDescriptorSetInputAttachments
= MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS
,
1283 .maxVertexInputAttributes
= MAX_VBS
,
1284 .maxVertexInputBindings
= MAX_VBS
,
1285 .maxVertexInputAttributeOffset
= 2047,
1286 .maxVertexInputBindingStride
= 2048,
1287 .maxVertexOutputComponents
= 128,
1288 .maxTessellationGenerationLevel
= 64,
1289 .maxTessellationPatchSize
= 32,
1290 .maxTessellationControlPerVertexInputComponents
= 128,
1291 .maxTessellationControlPerVertexOutputComponents
= 128,
1292 .maxTessellationControlPerPatchOutputComponents
= 128,
1293 .maxTessellationControlTotalOutputComponents
= 2048,
1294 .maxTessellationEvaluationInputComponents
= 128,
1295 .maxTessellationEvaluationOutputComponents
= 128,
1296 .maxGeometryShaderInvocations
= 32,
1297 .maxGeometryInputComponents
= 64,
1298 .maxGeometryOutputComponents
= 128,
1299 .maxGeometryOutputVertices
= 256,
1300 .maxGeometryTotalOutputComponents
= 1024,
1301 .maxFragmentInputComponents
= 116, /* 128 components - (PSIZ, CLIP_DIST0, CLIP_DIST1) */
1302 .maxFragmentOutputAttachments
= 8,
1303 .maxFragmentDualSrcAttachments
= 1,
1304 .maxFragmentCombinedOutputResources
= 8,
1305 .maxComputeSharedMemorySize
= 64 * 1024,
1306 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
1307 .maxComputeWorkGroupInvocations
= 32 * devinfo
->max_cs_threads
,
1308 .maxComputeWorkGroupSize
= {
1309 16 * devinfo
->max_cs_threads
,
1310 16 * devinfo
->max_cs_threads
,
1311 16 * devinfo
->max_cs_threads
,
1313 .subPixelPrecisionBits
= 8,
1314 .subTexelPrecisionBits
= 8,
1315 .mipmapPrecisionBits
= 8,
1316 .maxDrawIndexedIndexValue
= UINT32_MAX
,
1317 .maxDrawIndirectCount
= UINT32_MAX
,
1318 .maxSamplerLodBias
= 16,
1319 .maxSamplerAnisotropy
= 16,
1320 .maxViewports
= MAX_VIEWPORTS
,
1321 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
1322 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
1323 .viewportSubPixelBits
= 13, /* We take a float? */
1324 .minMemoryMapAlignment
= 4096, /* A page */
1325 /* The dataport requires texel alignment so we need to assume a worst
1326 * case of R32G32B32A32 which is 16 bytes.
1328 .minTexelBufferOffsetAlignment
= 16,
1329 /* We need 16 for UBO block reads to work and 32 for push UBOs */
1330 .minUniformBufferOffsetAlignment
= 32,
1331 .minStorageBufferOffsetAlignment
= 4,
1332 .minTexelOffset
= -8,
1333 .maxTexelOffset
= 7,
1334 .minTexelGatherOffset
= -32,
1335 .maxTexelGatherOffset
= 31,
1336 .minInterpolationOffset
= -0.5,
1337 .maxInterpolationOffset
= 0.4375,
1338 .subPixelInterpolationOffsetBits
= 4,
1339 .maxFramebufferWidth
= (1 << 14),
1340 .maxFramebufferHeight
= (1 << 14),
1341 .maxFramebufferLayers
= (1 << 11),
1342 .framebufferColorSampleCounts
= sample_counts
,
1343 .framebufferDepthSampleCounts
= sample_counts
,
1344 .framebufferStencilSampleCounts
= sample_counts
,
1345 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
1346 .maxColorAttachments
= MAX_RTS
,
1347 .sampledImageColorSampleCounts
= sample_counts
,
1348 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
1349 .sampledImageDepthSampleCounts
= sample_counts
,
1350 .sampledImageStencilSampleCounts
= sample_counts
,
1351 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
1352 .maxSampleMaskWords
= 1,
1353 .timestampComputeAndGraphics
= true,
1354 .timestampPeriod
= 1000000000.0 / devinfo
->timestamp_frequency
,
1355 .maxClipDistances
= 8,
1356 .maxCullDistances
= 8,
1357 .maxCombinedClipAndCullDistances
= 8,
1358 .discreteQueuePriorities
= 2,
1359 .pointSizeRange
= { 0.125, 255.875 },
1362 (devinfo
->gen
>= 9 || devinfo
->is_cherryview
) ?
1363 2047.9921875 : 7.9921875,
1365 .pointSizeGranularity
= (1.0 / 8.0),
1366 .lineWidthGranularity
= (1.0 / 128.0),
1367 .strictLines
= false,
1368 .standardSampleLocations
= true,
1369 .optimalBufferCopyOffsetAlignment
= 128,
1370 .optimalBufferCopyRowPitchAlignment
= 128,
1371 .nonCoherentAtomSize
= 64,
1374 *pProperties
= (VkPhysicalDeviceProperties
) {
1375 .apiVersion
= anv_physical_device_api_version(pdevice
),
1376 .driverVersion
= vk_get_driver_version(),
1378 .deviceID
= pdevice
->chipset_id
,
1379 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
1381 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
1384 snprintf(pProperties
->deviceName
, sizeof(pProperties
->deviceName
),
1385 "%s", pdevice
->name
);
1386 memcpy(pProperties
->pipelineCacheUUID
,
1387 pdevice
->pipeline_cache_uuid
, VK_UUID_SIZE
);
1390 void anv_GetPhysicalDeviceProperties2(
1391 VkPhysicalDevice physicalDevice
,
1392 VkPhysicalDeviceProperties2
* pProperties
)
1394 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1396 anv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
1398 vk_foreach_struct(ext
, pProperties
->pNext
) {
1399 switch (ext
->sType
) {
1400 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_STENCIL_RESOLVE_PROPERTIES_KHR
: {
1401 VkPhysicalDeviceDepthStencilResolvePropertiesKHR
*props
=
1402 (VkPhysicalDeviceDepthStencilResolvePropertiesKHR
*)ext
;
1404 /* We support all of the depth resolve modes */
1405 props
->supportedDepthResolveModes
=
1406 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
|
1407 VK_RESOLVE_MODE_AVERAGE_BIT_KHR
|
1408 VK_RESOLVE_MODE_MIN_BIT_KHR
|
1409 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1411 /* Average doesn't make sense for stencil so we don't support that */
1412 props
->supportedStencilResolveModes
=
1413 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
;
1414 if (pdevice
->info
.gen
>= 8) {
1415 /* The advanced stencil resolve modes currently require stencil
1416 * sampling be supported by the hardware.
1418 props
->supportedStencilResolveModes
|=
1419 VK_RESOLVE_MODE_MIN_BIT_KHR
|
1420 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1423 props
->independentResolveNone
= true;
1424 props
->independentResolve
= true;
1428 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES_EXT
: {
1429 VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*props
=
1430 (VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*)ext
;
1432 /* It's a bit hard to exactly map our implementation to the limits
1433 * described here. The bindless surface handle in the extended
1434 * message descriptors is 20 bits and it's an index into the table of
1435 * RENDER_SURFACE_STATE structs that starts at bindless surface base
1436 * address. Given that most things consume two surface states per
1437 * view (general/sampled for textures and write-only/read-write for
1438 * images), we claim 2^19 things.
1440 * For SSBOs, we just use A64 messages so there is no real limit
1441 * there beyond the limit on the total size of a descriptor set.
1443 const unsigned max_bindless_views
= 1 << 19;
1445 props
->maxUpdateAfterBindDescriptorsInAllPools
= max_bindless_views
;
1446 props
->shaderUniformBufferArrayNonUniformIndexingNative
= false;
1447 props
->shaderSampledImageArrayNonUniformIndexingNative
= false;
1448 props
->shaderStorageBufferArrayNonUniformIndexingNative
= true;
1449 props
->shaderStorageImageArrayNonUniformIndexingNative
= false;
1450 props
->shaderInputAttachmentArrayNonUniformIndexingNative
= false;
1451 props
->robustBufferAccessUpdateAfterBind
= true;
1452 props
->quadDivergentImplicitLod
= false;
1453 props
->maxPerStageDescriptorUpdateAfterBindSamplers
= max_bindless_views
;
1454 props
->maxPerStageDescriptorUpdateAfterBindUniformBuffers
= MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
;
1455 props
->maxPerStageDescriptorUpdateAfterBindStorageBuffers
= UINT32_MAX
;
1456 props
->maxPerStageDescriptorUpdateAfterBindSampledImages
= max_bindless_views
;
1457 props
->maxPerStageDescriptorUpdateAfterBindStorageImages
= max_bindless_views
;
1458 props
->maxPerStageDescriptorUpdateAfterBindInputAttachments
= MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS
;
1459 props
->maxPerStageUpdateAfterBindResources
= UINT32_MAX
;
1460 props
->maxDescriptorSetUpdateAfterBindSamplers
= max_bindless_views
;
1461 props
->maxDescriptorSetUpdateAfterBindUniformBuffers
= 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
;
1462 props
->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2;
1463 props
->maxDescriptorSetUpdateAfterBindStorageBuffers
= UINT32_MAX
;
1464 props
->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2;
1465 props
->maxDescriptorSetUpdateAfterBindSampledImages
= max_bindless_views
;
1466 props
->maxDescriptorSetUpdateAfterBindStorageImages
= max_bindless_views
;
1467 props
->maxDescriptorSetUpdateAfterBindInputAttachments
= MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS
;
1471 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR
: {
1472 VkPhysicalDeviceDriverPropertiesKHR
*driver_props
=
1473 (VkPhysicalDeviceDriverPropertiesKHR
*) ext
;
1475 driver_props
->driverID
= VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA_KHR
;
1476 snprintf(driver_props
->driverName
, VK_MAX_DRIVER_NAME_SIZE_KHR
,
1477 "Intel open-source Mesa driver");
1479 snprintf(driver_props
->driverInfo
, VK_MAX_DRIVER_INFO_SIZE_KHR
,
1480 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1
);
1482 driver_props
->conformanceVersion
= (VkConformanceVersionKHR
) {
1491 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT
: {
1492 VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*props
=
1493 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*) ext
;
1494 /* Userptr needs page aligned memory. */
1495 props
->minImportedHostPointerAlignment
= 4096;
1499 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES
: {
1500 VkPhysicalDeviceIDProperties
*id_props
=
1501 (VkPhysicalDeviceIDProperties
*)ext
;
1502 memcpy(id_props
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
1503 memcpy(id_props
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
1504 /* The LUID is for Windows. */
1505 id_props
->deviceLUIDValid
= false;
1509 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_PROPERTIES_EXT
: {
1510 VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*props
=
1511 (VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*)ext
;
1512 props
->maxInlineUniformBlockSize
= MAX_INLINE_UNIFORM_BLOCK_SIZE
;
1513 props
->maxPerStageDescriptorInlineUniformBlocks
=
1514 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1515 props
->maxPerStageDescriptorUpdateAfterBindInlineUniformBlocks
=
1516 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1517 props
->maxDescriptorSetInlineUniformBlocks
=
1518 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1519 props
->maxDescriptorSetUpdateAfterBindInlineUniformBlocks
=
1520 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1524 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_PROPERTIES_EXT
: {
1525 VkPhysicalDeviceLineRasterizationPropertiesEXT
*props
=
1526 (VkPhysicalDeviceLineRasterizationPropertiesEXT
*)ext
;
1527 /* In the Skylake PRM Vol. 7, subsection titled "GIQ (Diamond)
1528 * Sampling Rules - Legacy Mode", it says the following:
1530 * "Note that the device divides a pixel into a 16x16 array of
1531 * subpixels, referenced by their upper left corners."
1533 * This is the only known reference in the PRMs to the subpixel
1534 * precision of line rasterization and a "16x16 array of subpixels"
1535 * implies 4 subpixel precision bits. Empirical testing has shown
1536 * that 4 subpixel precision bits applies to all line rasterization
1539 props
->lineSubPixelPrecisionBits
= 4;
1543 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES
: {
1544 VkPhysicalDeviceMaintenance3Properties
*props
=
1545 (VkPhysicalDeviceMaintenance3Properties
*)ext
;
1546 /* This value doesn't matter for us today as our per-stage
1547 * descriptors are the real limit.
1549 props
->maxPerSetDescriptors
= 1024;
1550 props
->maxMemoryAllocationSize
= MAX_MEMORY_ALLOCATION_SIZE
;
1554 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES
: {
1555 VkPhysicalDeviceMultiviewProperties
*properties
=
1556 (VkPhysicalDeviceMultiviewProperties
*)ext
;
1557 properties
->maxMultiviewViewCount
= 16;
1558 properties
->maxMultiviewInstanceIndex
= UINT32_MAX
/ 16;
1562 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT
: {
1563 VkPhysicalDevicePCIBusInfoPropertiesEXT
*properties
=
1564 (VkPhysicalDevicePCIBusInfoPropertiesEXT
*)ext
;
1565 properties
->pciDomain
= pdevice
->pci_info
.domain
;
1566 properties
->pciBus
= pdevice
->pci_info
.bus
;
1567 properties
->pciDevice
= pdevice
->pci_info
.device
;
1568 properties
->pciFunction
= pdevice
->pci_info
.function
;
1572 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES
: {
1573 VkPhysicalDevicePointClippingProperties
*properties
=
1574 (VkPhysicalDevicePointClippingProperties
*) ext
;
1575 properties
->pointClippingBehavior
= VK_POINT_CLIPPING_BEHAVIOR_USER_CLIP_PLANES_ONLY
;
1579 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES
: {
1580 VkPhysicalDeviceProtectedMemoryProperties
*props
=
1581 (VkPhysicalDeviceProtectedMemoryProperties
*)ext
;
1582 props
->protectedNoFault
= false;
1586 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
1587 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
1588 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
1590 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
1594 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT
: {
1595 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*properties
=
1596 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*)ext
;
1597 properties
->filterMinmaxImageComponentMapping
= pdevice
->info
.gen
>= 9;
1598 properties
->filterMinmaxSingleComponentFormats
= true;
1602 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES
: {
1603 VkPhysicalDeviceSubgroupProperties
*properties
= (void *)ext
;
1605 properties
->subgroupSize
= BRW_SUBGROUP_SIZE
;
1607 VkShaderStageFlags scalar_stages
= 0;
1608 for (unsigned stage
= 0; stage
< MESA_SHADER_STAGES
; stage
++) {
1609 if (pdevice
->compiler
->scalar_stage
[stage
])
1610 scalar_stages
|= mesa_to_vk_shader_stage(stage
);
1612 properties
->supportedStages
= scalar_stages
;
1614 properties
->supportedOperations
= VK_SUBGROUP_FEATURE_BASIC_BIT
|
1615 VK_SUBGROUP_FEATURE_VOTE_BIT
|
1616 VK_SUBGROUP_FEATURE_BALLOT_BIT
|
1617 VK_SUBGROUP_FEATURE_SHUFFLE_BIT
|
1618 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT
|
1619 VK_SUBGROUP_FEATURE_QUAD_BIT
;
1620 if (pdevice
->info
.gen
>= 8) {
1621 /* TODO: There's no technical reason why these can't be made to
1622 * work on gen7 but they don't at the moment so it's best to leave
1623 * the feature disabled than enabled and broken.
1625 properties
->supportedOperations
|=
1626 VK_SUBGROUP_FEATURE_ARITHMETIC_BIT
|
1627 VK_SUBGROUP_FEATURE_CLUSTERED_BIT
;
1629 properties
->quadOperationsInAllStages
= pdevice
->info
.gen
>= 8;
1633 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_PROPERTIES_EXT
: {
1634 VkPhysicalDeviceSubgroupSizeControlPropertiesEXT
*props
=
1635 (VkPhysicalDeviceSubgroupSizeControlPropertiesEXT
*)ext
;
1636 STATIC_ASSERT(8 <= BRW_SUBGROUP_SIZE
&& BRW_SUBGROUP_SIZE
<= 32);
1637 props
->minSubgroupSize
= 8;
1638 props
->maxSubgroupSize
= 32;
1639 props
->maxComputeWorkgroupSubgroups
= pdevice
->info
.max_cs_threads
;
1640 props
->requiredSubgroupSizeStages
= VK_SHADER_STAGE_COMPUTE_BIT
;
1644 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_PROPERTIES_EXT
: {
1645 VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT
*props
=
1646 (VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT
*)ext
;
1648 /* From the SKL PRM Vol. 2d, docs for RENDER_SURFACE_STATE::Surface
1651 * "For SURFTYPE_BUFFER non-rendertarget surfaces, this field
1652 * specifies the base address of the first element of the surface,
1653 * computed in software by adding the surface base address to the
1654 * byte offset of the element in the buffer. The base address must
1655 * be aligned to element size."
1657 * The typed dataport messages require that things be texel aligned.
1658 * Otherwise, we may just load/store the wrong data or, in the worst
1659 * case, there may be hangs.
1661 props
->storageTexelBufferOffsetAlignmentBytes
= 16;
1662 props
->storageTexelBufferOffsetSingleTexelAlignment
= true;
1664 /* The sampler, however, is much more forgiving and it can handle
1665 * arbitrary byte alignment for linear and buffer surfaces. It's
1666 * hard to find a good PRM citation for this but years of empirical
1667 * experience demonstrate that this is true.
1669 props
->uniformTexelBufferOffsetAlignmentBytes
= 1;
1670 props
->uniformTexelBufferOffsetSingleTexelAlignment
= false;
1674 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT
: {
1675 VkPhysicalDeviceTransformFeedbackPropertiesEXT
*props
=
1676 (VkPhysicalDeviceTransformFeedbackPropertiesEXT
*)ext
;
1678 props
->maxTransformFeedbackStreams
= MAX_XFB_STREAMS
;
1679 props
->maxTransformFeedbackBuffers
= MAX_XFB_BUFFERS
;
1680 props
->maxTransformFeedbackBufferSize
= (1ull << 32);
1681 props
->maxTransformFeedbackStreamDataSize
= 128 * 4;
1682 props
->maxTransformFeedbackBufferDataSize
= 128 * 4;
1683 props
->maxTransformFeedbackBufferDataStride
= 2048;
1684 props
->transformFeedbackQueries
= true;
1685 props
->transformFeedbackStreamsLinesTriangles
= false;
1686 props
->transformFeedbackRasterizationStreamSelect
= false;
1687 props
->transformFeedbackDraw
= true;
1691 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT
: {
1692 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*props
=
1693 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*)ext
;
1694 /* We have to restrict this a bit for multiview */
1695 props
->maxVertexAttribDivisor
= UINT32_MAX
/ 16;
1700 anv_debug_ignored_stype(ext
->sType
);
1706 /* We support exactly one queue family. */
1707 static const VkQueueFamilyProperties
1708 anv_queue_family_properties
= {
1709 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
1710 VK_QUEUE_COMPUTE_BIT
|
1711 VK_QUEUE_TRANSFER_BIT
,
1713 .timestampValidBits
= 36, /* XXX: Real value here */
1714 .minImageTransferGranularity
= { 1, 1, 1 },
1717 void anv_GetPhysicalDeviceQueueFamilyProperties(
1718 VkPhysicalDevice physicalDevice
,
1720 VkQueueFamilyProperties
* pQueueFamilyProperties
)
1722 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pCount
);
1724 vk_outarray_append(&out
, p
) {
1725 *p
= anv_queue_family_properties
;
1729 void anv_GetPhysicalDeviceQueueFamilyProperties2(
1730 VkPhysicalDevice physicalDevice
,
1731 uint32_t* pQueueFamilyPropertyCount
,
1732 VkQueueFamilyProperties2
* pQueueFamilyProperties
)
1735 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pQueueFamilyPropertyCount
);
1737 vk_outarray_append(&out
, p
) {
1738 p
->queueFamilyProperties
= anv_queue_family_properties
;
1740 vk_foreach_struct(s
, p
->pNext
) {
1741 anv_debug_ignored_stype(s
->sType
);
1746 void anv_GetPhysicalDeviceMemoryProperties(
1747 VkPhysicalDevice physicalDevice
,
1748 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
1750 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
1752 pMemoryProperties
->memoryTypeCount
= physical_device
->memory
.type_count
;
1753 for (uint32_t i
= 0; i
< physical_device
->memory
.type_count
; i
++) {
1754 pMemoryProperties
->memoryTypes
[i
] = (VkMemoryType
) {
1755 .propertyFlags
= physical_device
->memory
.types
[i
].propertyFlags
,
1756 .heapIndex
= physical_device
->memory
.types
[i
].heapIndex
,
1760 pMemoryProperties
->memoryHeapCount
= physical_device
->memory
.heap_count
;
1761 for (uint32_t i
= 0; i
< physical_device
->memory
.heap_count
; i
++) {
1762 pMemoryProperties
->memoryHeaps
[i
] = (VkMemoryHeap
) {
1763 .size
= physical_device
->memory
.heaps
[i
].size
,
1764 .flags
= physical_device
->memory
.heaps
[i
].flags
,
1770 anv_get_memory_budget(VkPhysicalDevice physicalDevice
,
1771 VkPhysicalDeviceMemoryBudgetPropertiesEXT
*memoryBudget
)
1773 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
1774 uint64_t sys_available
= get_available_system_memory();
1775 assert(sys_available
> 0);
1777 VkDeviceSize total_heaps_size
= 0;
1778 for (size_t i
= 0; i
< device
->memory
.heap_count
; i
++)
1779 total_heaps_size
+= device
->memory
.heaps
[i
].size
;
1781 for (size_t i
= 0; i
< device
->memory
.heap_count
; i
++) {
1782 VkDeviceSize heap_size
= device
->memory
.heaps
[i
].size
;
1783 VkDeviceSize heap_used
= device
->memory
.heaps
[i
].used
;
1784 VkDeviceSize heap_budget
;
1786 double heap_proportion
= (double) heap_size
/ total_heaps_size
;
1787 VkDeviceSize sys_available_prop
= sys_available
* heap_proportion
;
1790 * Let's not incite the app to starve the system: report at most 90% of
1791 * available system memory.
1793 uint64_t heap_available
= sys_available_prop
* 9 / 10;
1794 heap_budget
= MIN2(heap_size
, heap_used
+ heap_available
);
1797 * Round down to the nearest MB
1799 heap_budget
&= ~((1ull << 20) - 1);
1802 * The heapBudget value must be non-zero for array elements less than
1803 * VkPhysicalDeviceMemoryProperties::memoryHeapCount. The heapBudget
1804 * value must be less than or equal to VkMemoryHeap::size for each heap.
1806 assert(0 < heap_budget
&& heap_budget
<= heap_size
);
1808 memoryBudget
->heapUsage
[i
] = heap_used
;
1809 memoryBudget
->heapBudget
[i
] = heap_budget
;
1812 /* The heapBudget and heapUsage values must be zero for array elements
1813 * greater than or equal to VkPhysicalDeviceMemoryProperties::memoryHeapCount
1815 for (uint32_t i
= device
->memory
.heap_count
; i
< VK_MAX_MEMORY_HEAPS
; i
++) {
1816 memoryBudget
->heapBudget
[i
] = 0;
1817 memoryBudget
->heapUsage
[i
] = 0;
1821 void anv_GetPhysicalDeviceMemoryProperties2(
1822 VkPhysicalDevice physicalDevice
,
1823 VkPhysicalDeviceMemoryProperties2
* pMemoryProperties
)
1825 anv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
1826 &pMemoryProperties
->memoryProperties
);
1828 vk_foreach_struct(ext
, pMemoryProperties
->pNext
) {
1829 switch (ext
->sType
) {
1830 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT
:
1831 anv_get_memory_budget(physicalDevice
, (void*)ext
);
1834 anv_debug_ignored_stype(ext
->sType
);
1841 anv_GetDeviceGroupPeerMemoryFeatures(
1844 uint32_t localDeviceIndex
,
1845 uint32_t remoteDeviceIndex
,
1846 VkPeerMemoryFeatureFlags
* pPeerMemoryFeatures
)
1848 assert(localDeviceIndex
== 0 && remoteDeviceIndex
== 0);
1849 *pPeerMemoryFeatures
= VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT
|
1850 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT
|
1851 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT
|
1852 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT
;
1855 PFN_vkVoidFunction
anv_GetInstanceProcAddr(
1856 VkInstance _instance
,
1859 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1861 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
1862 * when we have to return valid function pointers, NULL, or it's left
1863 * undefined. See the table for exact details.
1868 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
1869 if (strcmp(pName, "vk" #entrypoint) == 0) \
1870 return (PFN_vkVoidFunction)anv_##entrypoint
1872 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties
);
1873 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties
);
1874 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceVersion
);
1875 LOOKUP_ANV_ENTRYPOINT(CreateInstance
);
1877 #undef LOOKUP_ANV_ENTRYPOINT
1879 if (instance
== NULL
)
1882 int idx
= anv_get_instance_entrypoint_index(pName
);
1884 return instance
->dispatch
.entrypoints
[idx
];
1886 idx
= anv_get_device_entrypoint_index(pName
);
1888 return instance
->device_dispatch
.entrypoints
[idx
];
1893 /* With version 1+ of the loader interface the ICD should expose
1894 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
1897 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1898 VkInstance instance
,
1902 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1903 VkInstance instance
,
1906 return anv_GetInstanceProcAddr(instance
, pName
);
1909 PFN_vkVoidFunction
anv_GetDeviceProcAddr(
1913 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1915 if (!device
|| !pName
)
1918 int idx
= anv_get_device_entrypoint_index(pName
);
1922 return device
->dispatch
.entrypoints
[idx
];
1926 anv_CreateDebugReportCallbackEXT(VkInstance _instance
,
1927 const VkDebugReportCallbackCreateInfoEXT
* pCreateInfo
,
1928 const VkAllocationCallbacks
* pAllocator
,
1929 VkDebugReportCallbackEXT
* pCallback
)
1931 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1932 return vk_create_debug_report_callback(&instance
->debug_report_callbacks
,
1933 pCreateInfo
, pAllocator
, &instance
->alloc
,
1938 anv_DestroyDebugReportCallbackEXT(VkInstance _instance
,
1939 VkDebugReportCallbackEXT _callback
,
1940 const VkAllocationCallbacks
* pAllocator
)
1942 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1943 vk_destroy_debug_report_callback(&instance
->debug_report_callbacks
,
1944 _callback
, pAllocator
, &instance
->alloc
);
1948 anv_DebugReportMessageEXT(VkInstance _instance
,
1949 VkDebugReportFlagsEXT flags
,
1950 VkDebugReportObjectTypeEXT objectType
,
1953 int32_t messageCode
,
1954 const char* pLayerPrefix
,
1955 const char* pMessage
)
1957 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1958 vk_debug_report(&instance
->debug_report_callbacks
, flags
, objectType
,
1959 object
, location
, messageCode
, pLayerPrefix
, pMessage
);
1963 anv_queue_init(struct anv_device
*device
, struct anv_queue
*queue
)
1965 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1966 queue
->device
= device
;
1971 anv_queue_finish(struct anv_queue
*queue
)
1975 static struct anv_state
1976 anv_state_pool_emit_data(struct anv_state_pool
*pool
, size_t size
, size_t align
, const void *p
)
1978 struct anv_state state
;
1980 state
= anv_state_pool_alloc(pool
, size
, align
);
1981 memcpy(state
.map
, p
, size
);
1986 /* Haswell border color is a bit of a disaster. Float and unorm formats use a
1987 * straightforward 32-bit float color in the first 64 bytes. Instead of using
1988 * a nice float/integer union like Gen8+, Haswell specifies the integer border
1989 * color as a separate entry /after/ the float color. The layout of this entry
1990 * also depends on the format's bpp (with extra hacks for RG32), and overlaps.
1992 * Since we don't know the format/bpp, we can't make any of the border colors
1993 * containing '1' work for all formats, as it would be in the wrong place for
1994 * some of them. We opt to make 32-bit integers work as this seems like the
1995 * most common option. Fortunately, transparent black works regardless, as
1996 * all zeroes is the same in every bit-size.
1998 struct hsw_border_color
{
2002 uint32_t _pad1
[108];
2005 struct gen8_border_color
{
2010 /* Pad out to 64 bytes */
2015 anv_device_init_border_colors(struct anv_device
*device
)
2017 if (device
->info
.is_haswell
) {
2018 static const struct hsw_border_color border_colors
[] = {
2019 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
2020 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
2021 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
2022 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
2023 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
2024 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
2027 device
->border_colors
=
2028 anv_state_pool_emit_data(&device
->dynamic_state_pool
,
2029 sizeof(border_colors
), 512, border_colors
);
2031 static const struct gen8_border_color border_colors
[] = {
2032 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
2033 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
2034 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
2035 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
2036 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
2037 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
2040 device
->border_colors
=
2041 anv_state_pool_emit_data(&device
->dynamic_state_pool
,
2042 sizeof(border_colors
), 64, border_colors
);
2047 anv_device_init_trivial_batch(struct anv_device
*device
)
2049 anv_bo_init_new(&device
->trivial_batch_bo
, device
, 4096);
2051 if (device
->instance
->physicalDevice
.has_exec_async
)
2052 device
->trivial_batch_bo
.flags
|= EXEC_OBJECT_ASYNC
;
2054 if (device
->instance
->physicalDevice
.use_softpin
)
2055 device
->trivial_batch_bo
.flags
|= EXEC_OBJECT_PINNED
;
2057 anv_vma_alloc(device
, &device
->trivial_batch_bo
);
2059 void *map
= anv_gem_mmap(device
, device
->trivial_batch_bo
.gem_handle
,
2062 struct anv_batch batch
= {
2068 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
2069 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
2071 if (!device
->info
.has_llc
)
2072 gen_clflush_range(map
, batch
.next
- map
);
2074 anv_gem_munmap(map
, device
->trivial_batch_bo
.size
);
2077 VkResult
anv_EnumerateDeviceExtensionProperties(
2078 VkPhysicalDevice physicalDevice
,
2079 const char* pLayerName
,
2080 uint32_t* pPropertyCount
,
2081 VkExtensionProperties
* pProperties
)
2083 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
2084 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
2086 for (int i
= 0; i
< ANV_DEVICE_EXTENSION_COUNT
; i
++) {
2087 if (device
->supported_extensions
.extensions
[i
]) {
2088 vk_outarray_append(&out
, prop
) {
2089 *prop
= anv_device_extensions
[i
];
2094 return vk_outarray_status(&out
);
2098 anv_device_init_dispatch(struct anv_device
*device
)
2100 const struct anv_device_dispatch_table
*genX_table
;
2101 switch (device
->info
.gen
) {
2103 genX_table
= &gen11_device_dispatch_table
;
2106 genX_table
= &gen10_device_dispatch_table
;
2109 genX_table
= &gen9_device_dispatch_table
;
2112 genX_table
= &gen8_device_dispatch_table
;
2115 if (device
->info
.is_haswell
)
2116 genX_table
= &gen75_device_dispatch_table
;
2118 genX_table
= &gen7_device_dispatch_table
;
2121 unreachable("unsupported gen\n");
2124 for (unsigned i
= 0; i
< ARRAY_SIZE(device
->dispatch
.entrypoints
); i
++) {
2125 /* Vulkan requires that entrypoints for extensions which have not been
2126 * enabled must not be advertised.
2128 if (!anv_device_entrypoint_is_enabled(i
, device
->instance
->app_info
.api_version
,
2129 &device
->instance
->enabled_extensions
,
2130 &device
->enabled_extensions
)) {
2131 device
->dispatch
.entrypoints
[i
] = NULL
;
2132 } else if (genX_table
->entrypoints
[i
]) {
2133 device
->dispatch
.entrypoints
[i
] = genX_table
->entrypoints
[i
];
2135 device
->dispatch
.entrypoints
[i
] =
2136 anv_device_dispatch_table
.entrypoints
[i
];
2142 vk_priority_to_gen(int priority
)
2145 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT
:
2146 return GEN_CONTEXT_LOW_PRIORITY
;
2147 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
:
2148 return GEN_CONTEXT_MEDIUM_PRIORITY
;
2149 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT
:
2150 return GEN_CONTEXT_HIGH_PRIORITY
;
2151 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT
:
2152 return GEN_CONTEXT_REALTIME_PRIORITY
;
2154 unreachable("Invalid priority");
2159 anv_device_init_hiz_clear_value_bo(struct anv_device
*device
)
2161 anv_bo_init_new(&device
->hiz_clear_bo
, device
, 4096);
2163 if (device
->instance
->physicalDevice
.has_exec_async
)
2164 device
->hiz_clear_bo
.flags
|= EXEC_OBJECT_ASYNC
;
2166 if (device
->instance
->physicalDevice
.use_softpin
)
2167 device
->hiz_clear_bo
.flags
|= EXEC_OBJECT_PINNED
;
2169 anv_vma_alloc(device
, &device
->hiz_clear_bo
);
2171 uint32_t *map
= anv_gem_mmap(device
, device
->hiz_clear_bo
.gem_handle
,
2174 union isl_color_value hiz_clear
= { .u32
= { 0, } };
2175 hiz_clear
.f32
[0] = ANV_HZ_FC_VAL
;
2177 memcpy(map
, hiz_clear
.u32
, sizeof(hiz_clear
.u32
));
2178 anv_gem_munmap(map
, device
->hiz_clear_bo
.size
);
2182 get_bo_from_pool(struct gen_batch_decode_bo
*ret
,
2183 struct anv_block_pool
*pool
,
2186 for (uint32_t i
= 0; i
< pool
->nbos
; i
++) {
2187 uint64_t bo_address
= pool
->bos
[i
].offset
& (~0ull >> 16);
2188 uint32_t bo_size
= pool
->bos
[i
].size
;
2189 if (address
>= bo_address
&& address
< (bo_address
+ bo_size
)) {
2190 *ret
= (struct gen_batch_decode_bo
) {
2193 .map
= pool
->bos
[i
].map
,
2201 /* Finding a buffer for batch decoding */
2202 static struct gen_batch_decode_bo
2203 decode_get_bo(void *v_batch
, bool ppgtt
, uint64_t address
)
2205 struct anv_device
*device
= v_batch
;
2206 struct gen_batch_decode_bo ret_bo
= {};
2210 if (get_bo_from_pool(&ret_bo
, &device
->dynamic_state_pool
.block_pool
, address
))
2212 if (get_bo_from_pool(&ret_bo
, &device
->instruction_state_pool
.block_pool
, address
))
2214 if (get_bo_from_pool(&ret_bo
, &device
->binding_table_pool
.block_pool
, address
))
2216 if (get_bo_from_pool(&ret_bo
, &device
->surface_state_pool
.block_pool
, address
))
2219 if (!device
->cmd_buffer_being_decoded
)
2220 return (struct gen_batch_decode_bo
) { };
2222 struct anv_batch_bo
**bo
;
2224 u_vector_foreach(bo
, &device
->cmd_buffer_being_decoded
->seen_bbos
) {
2225 /* The decoder zeroes out the top 16 bits, so we need to as well */
2226 uint64_t bo_address
= (*bo
)->bo
.offset
& (~0ull >> 16);
2228 if (address
>= bo_address
&& address
< bo_address
+ (*bo
)->bo
.size
) {
2229 return (struct gen_batch_decode_bo
) {
2231 .size
= (*bo
)->bo
.size
,
2232 .map
= (*bo
)->bo
.map
,
2237 return (struct gen_batch_decode_bo
) { };
2240 VkResult
anv_CreateDevice(
2241 VkPhysicalDevice physicalDevice
,
2242 const VkDeviceCreateInfo
* pCreateInfo
,
2243 const VkAllocationCallbacks
* pAllocator
,
2246 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
2248 struct anv_device
*device
;
2250 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO
);
2252 struct anv_device_extension_table enabled_extensions
= { };
2253 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
2255 for (idx
= 0; idx
< ANV_DEVICE_EXTENSION_COUNT
; idx
++) {
2256 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
2257 anv_device_extensions
[idx
].extensionName
) == 0)
2261 if (idx
>= ANV_DEVICE_EXTENSION_COUNT
)
2262 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
2264 if (!physical_device
->supported_extensions
.extensions
[idx
])
2265 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
2267 enabled_extensions
.extensions
[idx
] = true;
2270 /* Check enabled features */
2271 if (pCreateInfo
->pEnabledFeatures
) {
2272 VkPhysicalDeviceFeatures supported_features
;
2273 anv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
2274 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
2275 VkBool32
*enabled_feature
= (VkBool32
*)pCreateInfo
->pEnabledFeatures
;
2276 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
2277 for (uint32_t i
= 0; i
< num_features
; i
++) {
2278 if (enabled_feature
[i
] && !supported_feature
[i
])
2279 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
2283 /* Check requested queues and fail if we are requested to create any
2284 * queues with flags we don't support.
2286 assert(pCreateInfo
->queueCreateInfoCount
> 0);
2287 for (uint32_t i
= 0; i
< pCreateInfo
->queueCreateInfoCount
; i
++) {
2288 if (pCreateInfo
->pQueueCreateInfos
[i
].flags
!= 0)
2289 return vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2292 /* Check if client specified queue priority. */
2293 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*queue_priority
=
2294 vk_find_struct_const(pCreateInfo
->pQueueCreateInfos
[0].pNext
,
2295 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT
);
2297 VkQueueGlobalPriorityEXT priority
=
2298 queue_priority
? queue_priority
->globalPriority
:
2299 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
;
2301 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
2303 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
2305 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2307 if (INTEL_DEBUG
& DEBUG_BATCH
) {
2308 const unsigned decode_flags
=
2309 GEN_BATCH_DECODE_FULL
|
2310 ((INTEL_DEBUG
& DEBUG_COLOR
) ? GEN_BATCH_DECODE_IN_COLOR
: 0) |
2311 GEN_BATCH_DECODE_OFFSETS
|
2312 GEN_BATCH_DECODE_FLOATS
;
2314 gen_batch_decode_ctx_init(&device
->decoder_ctx
,
2315 &physical_device
->info
,
2316 stderr
, decode_flags
, NULL
,
2317 decode_get_bo
, NULL
, device
);
2320 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
2321 device
->instance
= physical_device
->instance
;
2322 device
->chipset_id
= physical_device
->chipset_id
;
2323 device
->no_hw
= physical_device
->no_hw
;
2324 device
->_lost
= false;
2327 device
->alloc
= *pAllocator
;
2329 device
->alloc
= physical_device
->instance
->alloc
;
2331 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
2332 device
->fd
= open(physical_device
->path
, O_RDWR
| O_CLOEXEC
);
2333 if (device
->fd
== -1) {
2334 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2338 device
->context_id
= anv_gem_create_context(device
);
2339 if (device
->context_id
== -1) {
2340 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2344 if (physical_device
->use_softpin
) {
2345 if (pthread_mutex_init(&device
->vma_mutex
, NULL
) != 0) {
2346 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2350 /* keep the page with address zero out of the allocator */
2351 struct anv_memory_heap
*low_heap
=
2352 &physical_device
->memory
.heaps
[physical_device
->memory
.heap_count
- 1];
2353 util_vma_heap_init(&device
->vma_lo
, low_heap
->vma_start
, low_heap
->vma_size
);
2354 device
->vma_lo_available
= low_heap
->size
;
2356 struct anv_memory_heap
*high_heap
=
2357 &physical_device
->memory
.heaps
[0];
2358 util_vma_heap_init(&device
->vma_hi
, high_heap
->vma_start
, high_heap
->vma_size
);
2359 device
->vma_hi_available
= physical_device
->memory
.heap_count
== 1 ? 0 :
2363 list_inithead(&device
->memory_objects
);
2365 /* As per spec, the driver implementation may deny requests to acquire
2366 * a priority above the default priority (MEDIUM) if the caller does not
2367 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
2370 if (physical_device
->has_context_priority
) {
2371 int err
= anv_gem_set_context_param(device
->fd
, device
->context_id
,
2372 I915_CONTEXT_PARAM_PRIORITY
,
2373 vk_priority_to_gen(priority
));
2374 if (err
!= 0 && priority
> VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
) {
2375 result
= vk_error(VK_ERROR_NOT_PERMITTED_EXT
);
2380 device
->info
= physical_device
->info
;
2381 device
->isl_dev
= physical_device
->isl_dev
;
2383 /* On Broadwell and later, we can use batch chaining to more efficiently
2384 * implement growing command buffers. Prior to Haswell, the kernel
2385 * command parser gets in the way and we have to fall back to growing
2388 device
->can_chain_batches
= device
->info
.gen
>= 8;
2390 device
->robust_buffer_access
= pCreateInfo
->pEnabledFeatures
&&
2391 pCreateInfo
->pEnabledFeatures
->robustBufferAccess
;
2392 device
->enabled_extensions
= enabled_extensions
;
2394 anv_device_init_dispatch(device
);
2396 if (pthread_mutex_init(&device
->mutex
, NULL
) != 0) {
2397 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2398 goto fail_context_id
;
2401 pthread_condattr_t condattr
;
2402 if (pthread_condattr_init(&condattr
) != 0) {
2403 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2406 if (pthread_condattr_setclock(&condattr
, CLOCK_MONOTONIC
) != 0) {
2407 pthread_condattr_destroy(&condattr
);
2408 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2411 if (pthread_cond_init(&device
->queue_submit
, &condattr
) != 0) {
2412 pthread_condattr_destroy(&condattr
);
2413 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2416 pthread_condattr_destroy(&condattr
);
2419 (physical_device
->supports_48bit_addresses
? EXEC_OBJECT_SUPPORTS_48B_ADDRESS
: 0) |
2420 (physical_device
->has_exec_async
? EXEC_OBJECT_ASYNC
: 0) |
2421 (physical_device
->has_exec_capture
? EXEC_OBJECT_CAPTURE
: 0) |
2422 (physical_device
->use_softpin
? EXEC_OBJECT_PINNED
: 0);
2424 anv_bo_pool_init(&device
->batch_bo_pool
, device
, bo_flags
);
2426 result
= anv_bo_cache_init(&device
->bo_cache
);
2427 if (result
!= VK_SUCCESS
)
2428 goto fail_batch_bo_pool
;
2430 if (!physical_device
->use_softpin
)
2431 bo_flags
&= ~EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
2433 result
= anv_state_pool_init(&device
->dynamic_state_pool
, device
,
2434 DYNAMIC_STATE_POOL_MIN_ADDRESS
,
2437 if (result
!= VK_SUCCESS
)
2440 result
= anv_state_pool_init(&device
->instruction_state_pool
, device
,
2441 INSTRUCTION_STATE_POOL_MIN_ADDRESS
,
2444 if (result
!= VK_SUCCESS
)
2445 goto fail_dynamic_state_pool
;
2447 result
= anv_state_pool_init(&device
->surface_state_pool
, device
,
2448 SURFACE_STATE_POOL_MIN_ADDRESS
,
2451 if (result
!= VK_SUCCESS
)
2452 goto fail_instruction_state_pool
;
2454 if (physical_device
->use_softpin
) {
2455 result
= anv_state_pool_init(&device
->binding_table_pool
, device
,
2456 BINDING_TABLE_POOL_MIN_ADDRESS
,
2459 if (result
!= VK_SUCCESS
)
2460 goto fail_surface_state_pool
;
2463 result
= anv_bo_init_new(&device
->workaround_bo
, device
, 4096);
2464 if (result
!= VK_SUCCESS
)
2465 goto fail_binding_table_pool
;
2467 if (physical_device
->use_softpin
)
2468 device
->workaround_bo
.flags
|= EXEC_OBJECT_PINNED
;
2470 if (!anv_vma_alloc(device
, &device
->workaround_bo
))
2471 goto fail_workaround_bo
;
2473 anv_device_init_trivial_batch(device
);
2475 if (device
->info
.gen
>= 10)
2476 anv_device_init_hiz_clear_value_bo(device
);
2478 anv_scratch_pool_init(device
, &device
->scratch_pool
);
2480 anv_queue_init(device
, &device
->queue
);
2482 switch (device
->info
.gen
) {
2484 if (!device
->info
.is_haswell
)
2485 result
= gen7_init_device_state(device
);
2487 result
= gen75_init_device_state(device
);
2490 result
= gen8_init_device_state(device
);
2493 result
= gen9_init_device_state(device
);
2496 result
= gen10_init_device_state(device
);
2499 result
= gen11_init_device_state(device
);
2502 /* Shouldn't get here as we don't create physical devices for any other
2504 unreachable("unhandled gen");
2506 if (result
!= VK_SUCCESS
)
2507 goto fail_workaround_bo
;
2509 anv_pipeline_cache_init(&device
->default_pipeline_cache
, device
, true);
2511 anv_device_init_blorp(device
);
2513 anv_device_init_border_colors(device
);
2515 *pDevice
= anv_device_to_handle(device
);
2520 anv_queue_finish(&device
->queue
);
2521 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
2522 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
2523 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
2524 fail_binding_table_pool
:
2525 if (physical_device
->use_softpin
)
2526 anv_state_pool_finish(&device
->binding_table_pool
);
2527 fail_surface_state_pool
:
2528 anv_state_pool_finish(&device
->surface_state_pool
);
2529 fail_instruction_state_pool
:
2530 anv_state_pool_finish(&device
->instruction_state_pool
);
2531 fail_dynamic_state_pool
:
2532 anv_state_pool_finish(&device
->dynamic_state_pool
);
2534 anv_bo_cache_finish(&device
->bo_cache
);
2536 anv_bo_pool_finish(&device
->batch_bo_pool
);
2537 pthread_cond_destroy(&device
->queue_submit
);
2539 pthread_mutex_destroy(&device
->mutex
);
2541 anv_gem_destroy_context(device
, device
->context_id
);
2545 vk_free(&device
->alloc
, device
);
2550 void anv_DestroyDevice(
2552 const VkAllocationCallbacks
* pAllocator
)
2554 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2555 struct anv_physical_device
*physical_device
;
2560 physical_device
= &device
->instance
->physicalDevice
;
2562 anv_device_finish_blorp(device
);
2564 anv_pipeline_cache_finish(&device
->default_pipeline_cache
);
2566 anv_queue_finish(&device
->queue
);
2568 #ifdef HAVE_VALGRIND
2569 /* We only need to free these to prevent valgrind errors. The backing
2570 * BO will go away in a couple of lines so we don't actually leak.
2572 anv_state_pool_free(&device
->dynamic_state_pool
, device
->border_colors
);
2575 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
2577 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
2578 anv_vma_free(device
, &device
->workaround_bo
);
2579 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
2581 anv_vma_free(device
, &device
->trivial_batch_bo
);
2582 anv_gem_close(device
, device
->trivial_batch_bo
.gem_handle
);
2583 if (device
->info
.gen
>= 10)
2584 anv_gem_close(device
, device
->hiz_clear_bo
.gem_handle
);
2586 if (physical_device
->use_softpin
)
2587 anv_state_pool_finish(&device
->binding_table_pool
);
2588 anv_state_pool_finish(&device
->surface_state_pool
);
2589 anv_state_pool_finish(&device
->instruction_state_pool
);
2590 anv_state_pool_finish(&device
->dynamic_state_pool
);
2592 anv_bo_cache_finish(&device
->bo_cache
);
2594 anv_bo_pool_finish(&device
->batch_bo_pool
);
2596 pthread_cond_destroy(&device
->queue_submit
);
2597 pthread_mutex_destroy(&device
->mutex
);
2599 anv_gem_destroy_context(device
, device
->context_id
);
2601 if (INTEL_DEBUG
& DEBUG_BATCH
)
2602 gen_batch_decode_ctx_finish(&device
->decoder_ctx
);
2606 vk_free(&device
->alloc
, device
);
2609 VkResult
anv_EnumerateInstanceLayerProperties(
2610 uint32_t* pPropertyCount
,
2611 VkLayerProperties
* pProperties
)
2613 if (pProperties
== NULL
) {
2614 *pPropertyCount
= 0;
2618 /* None supported at this time */
2619 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
2622 VkResult
anv_EnumerateDeviceLayerProperties(
2623 VkPhysicalDevice physicalDevice
,
2624 uint32_t* pPropertyCount
,
2625 VkLayerProperties
* pProperties
)
2627 if (pProperties
== NULL
) {
2628 *pPropertyCount
= 0;
2632 /* None supported at this time */
2633 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
2636 void anv_GetDeviceQueue(
2638 uint32_t queueNodeIndex
,
2639 uint32_t queueIndex
,
2642 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2644 assert(queueIndex
== 0);
2646 *pQueue
= anv_queue_to_handle(&device
->queue
);
2649 void anv_GetDeviceQueue2(
2651 const VkDeviceQueueInfo2
* pQueueInfo
,
2654 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2656 assert(pQueueInfo
->queueIndex
== 0);
2658 if (pQueueInfo
->flags
== device
->queue
.flags
)
2659 *pQueue
= anv_queue_to_handle(&device
->queue
);
2665 _anv_device_set_lost(struct anv_device
*device
,
2666 const char *file
, int line
,
2667 const char *msg
, ...)
2672 device
->_lost
= true;
2675 err
= __vk_errorv(device
->instance
, device
,
2676 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT
,
2677 VK_ERROR_DEVICE_LOST
, file
, line
, msg
, ap
);
2680 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
2687 anv_device_query_status(struct anv_device
*device
)
2689 /* This isn't likely as most of the callers of this function already check
2690 * for it. However, it doesn't hurt to check and it potentially lets us
2693 if (anv_device_is_lost(device
))
2694 return VK_ERROR_DEVICE_LOST
;
2696 uint32_t active
, pending
;
2697 int ret
= anv_gem_gpu_get_reset_stats(device
, &active
, &pending
);
2699 /* We don't know the real error. */
2700 return anv_device_set_lost(device
, "get_reset_stats failed: %m");
2704 return anv_device_set_lost(device
, "GPU hung on one of our command buffers");
2705 } else if (pending
) {
2706 return anv_device_set_lost(device
, "GPU hung with commands in-flight");
2713 anv_device_bo_busy(struct anv_device
*device
, struct anv_bo
*bo
)
2715 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
2716 * Other usages of the BO (such as on different hardware) will not be
2717 * flagged as "busy" by this ioctl. Use with care.
2719 int ret
= anv_gem_busy(device
, bo
->gem_handle
);
2721 return VK_NOT_READY
;
2722 } else if (ret
== -1) {
2723 /* We don't know the real error. */
2724 return anv_device_set_lost(device
, "gem wait failed: %m");
2727 /* Query for device status after the busy call. If the BO we're checking
2728 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
2729 * client because it clearly doesn't have valid data. Yes, this most
2730 * likely means an ioctl, but we just did an ioctl to query the busy status
2731 * so it's no great loss.
2733 return anv_device_query_status(device
);
2737 anv_device_wait(struct anv_device
*device
, struct anv_bo
*bo
,
2740 int ret
= anv_gem_wait(device
, bo
->gem_handle
, &timeout
);
2741 if (ret
== -1 && errno
== ETIME
) {
2743 } else if (ret
== -1) {
2744 /* We don't know the real error. */
2745 return anv_device_set_lost(device
, "gem wait failed: %m");
2748 /* Query for device status after the wait. If the BO we're waiting on got
2749 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
2750 * because it clearly doesn't have valid data. Yes, this most likely means
2751 * an ioctl, but we just did an ioctl to wait so it's no great loss.
2753 return anv_device_query_status(device
);
2756 VkResult
anv_DeviceWaitIdle(
2759 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2760 if (anv_device_is_lost(device
))
2761 return VK_ERROR_DEVICE_LOST
;
2763 struct anv_batch batch
;
2766 batch
.start
= batch
.next
= cmds
;
2767 batch
.end
= (void *) cmds
+ sizeof(cmds
);
2769 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
2770 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
2772 return anv_device_submit_simple_batch(device
, &batch
);
2776 anv_vma_alloc(struct anv_device
*device
, struct anv_bo
*bo
)
2778 if (!(bo
->flags
& EXEC_OBJECT_PINNED
))
2781 pthread_mutex_lock(&device
->vma_mutex
);
2785 if (bo
->flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
&&
2786 device
->vma_hi_available
>= bo
->size
) {
2787 uint64_t addr
= util_vma_heap_alloc(&device
->vma_hi
, bo
->size
, 4096);
2789 bo
->offset
= gen_canonical_address(addr
);
2790 assert(addr
== gen_48b_address(bo
->offset
));
2791 device
->vma_hi_available
-= bo
->size
;
2795 if (bo
->offset
== 0 && device
->vma_lo_available
>= bo
->size
) {
2796 uint64_t addr
= util_vma_heap_alloc(&device
->vma_lo
, bo
->size
, 4096);
2798 bo
->offset
= gen_canonical_address(addr
);
2799 assert(addr
== gen_48b_address(bo
->offset
));
2800 device
->vma_lo_available
-= bo
->size
;
2804 pthread_mutex_unlock(&device
->vma_mutex
);
2806 return bo
->offset
!= 0;
2810 anv_vma_free(struct anv_device
*device
, struct anv_bo
*bo
)
2812 if (!(bo
->flags
& EXEC_OBJECT_PINNED
))
2815 const uint64_t addr_48b
= gen_48b_address(bo
->offset
);
2817 pthread_mutex_lock(&device
->vma_mutex
);
2819 if (addr_48b
>= LOW_HEAP_MIN_ADDRESS
&&
2820 addr_48b
<= LOW_HEAP_MAX_ADDRESS
) {
2821 util_vma_heap_free(&device
->vma_lo
, addr_48b
, bo
->size
);
2822 device
->vma_lo_available
+= bo
->size
;
2824 ASSERTED
const struct anv_physical_device
*physical_device
=
2825 &device
->instance
->physicalDevice
;
2826 assert(addr_48b
>= physical_device
->memory
.heaps
[0].vma_start
&&
2827 addr_48b
< (physical_device
->memory
.heaps
[0].vma_start
+
2828 physical_device
->memory
.heaps
[0].vma_size
));
2829 util_vma_heap_free(&device
->vma_hi
, addr_48b
, bo
->size
);
2830 device
->vma_hi_available
+= bo
->size
;
2833 pthread_mutex_unlock(&device
->vma_mutex
);
2839 anv_bo_init_new(struct anv_bo
*bo
, struct anv_device
*device
, uint64_t size
)
2841 uint32_t gem_handle
= anv_gem_create(device
, size
);
2843 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
2845 anv_bo_init(bo
, gem_handle
, size
);
2850 VkResult
anv_AllocateMemory(
2852 const VkMemoryAllocateInfo
* pAllocateInfo
,
2853 const VkAllocationCallbacks
* pAllocator
,
2854 VkDeviceMemory
* pMem
)
2856 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2857 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2858 struct anv_device_memory
*mem
;
2859 VkResult result
= VK_SUCCESS
;
2861 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
2863 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
2864 assert(pAllocateInfo
->allocationSize
> 0);
2866 if (pAllocateInfo
->allocationSize
> MAX_MEMORY_ALLOCATION_SIZE
)
2867 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
2869 /* FINISHME: Fail if allocation request exceeds heap size. */
2871 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
2872 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2874 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2876 assert(pAllocateInfo
->memoryTypeIndex
< pdevice
->memory
.type_count
);
2877 mem
->type
= &pdevice
->memory
.types
[pAllocateInfo
->memoryTypeIndex
];
2881 mem
->host_ptr
= NULL
;
2883 uint64_t bo_flags
= 0;
2885 assert(mem
->type
->heapIndex
< pdevice
->memory
.heap_count
);
2886 if (pdevice
->memory
.heaps
[mem
->type
->heapIndex
].supports_48bit_addresses
)
2887 bo_flags
|= EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
2889 const struct wsi_memory_allocate_info
*wsi_info
=
2890 vk_find_struct_const(pAllocateInfo
->pNext
, WSI_MEMORY_ALLOCATE_INFO_MESA
);
2891 if (wsi_info
&& wsi_info
->implicit_sync
) {
2892 /* We need to set the WRITE flag on window system buffers so that GEM
2893 * will know we're writing to them and synchronize uses on other rings
2894 * (eg if the display server uses the blitter ring).
2896 bo_flags
|= EXEC_OBJECT_WRITE
;
2897 } else if (pdevice
->has_exec_async
) {
2898 bo_flags
|= EXEC_OBJECT_ASYNC
;
2901 if (pdevice
->use_softpin
)
2902 bo_flags
|= EXEC_OBJECT_PINNED
;
2904 const VkExportMemoryAllocateInfo
*export_info
=
2905 vk_find_struct_const(pAllocateInfo
->pNext
, EXPORT_MEMORY_ALLOCATE_INFO
);
2907 /* Check if we need to support Android HW buffer export. If so,
2908 * create AHardwareBuffer and import memory from it.
2910 bool android_export
= false;
2911 if (export_info
&& export_info
->handleTypes
&
2912 VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID
)
2913 android_export
= true;
2915 /* Android memory import. */
2916 const struct VkImportAndroidHardwareBufferInfoANDROID
*ahw_import_info
=
2917 vk_find_struct_const(pAllocateInfo
->pNext
,
2918 IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID
);
2920 if (ahw_import_info
) {
2921 result
= anv_import_ahw_memory(_device
, mem
, ahw_import_info
);
2922 if (result
!= VK_SUCCESS
)
2926 } else if (android_export
) {
2927 result
= anv_create_ahw_memory(_device
, mem
, pAllocateInfo
);
2928 if (result
!= VK_SUCCESS
)
2931 const struct VkImportAndroidHardwareBufferInfoANDROID import_info
= {
2934 result
= anv_import_ahw_memory(_device
, mem
, &import_info
);
2935 if (result
!= VK_SUCCESS
)
2941 const VkImportMemoryFdInfoKHR
*fd_info
=
2942 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_FD_INFO_KHR
);
2944 /* The Vulkan spec permits handleType to be 0, in which case the struct is
2947 if (fd_info
&& fd_info
->handleType
) {
2948 /* At the moment, we support only the below handle types. */
2949 assert(fd_info
->handleType
==
2950 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
2951 fd_info
->handleType
==
2952 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
2954 result
= anv_bo_cache_import(device
, &device
->bo_cache
, fd_info
->fd
,
2955 bo_flags
| ANV_BO_EXTERNAL
, &mem
->bo
);
2956 if (result
!= VK_SUCCESS
)
2959 VkDeviceSize aligned_alloc_size
=
2960 align_u64(pAllocateInfo
->allocationSize
, 4096);
2962 /* For security purposes, we reject importing the bo if it's smaller
2963 * than the requested allocation size. This prevents a malicious client
2964 * from passing a buffer to a trusted client, lying about the size, and
2965 * telling the trusted client to try and texture from an image that goes
2966 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
2967 * in the trusted client. The trusted client can protect itself against
2968 * this sort of attack but only if it can trust the buffer size.
2970 if (mem
->bo
->size
< aligned_alloc_size
) {
2971 result
= vk_errorf(device
->instance
, device
,
2972 VK_ERROR_INVALID_EXTERNAL_HANDLE
,
2973 "aligned allocationSize too large for "
2974 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT: "
2975 "%"PRIu64
"B > %"PRIu64
"B",
2976 aligned_alloc_size
, mem
->bo
->size
);
2977 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
2981 /* From the Vulkan spec:
2983 * "Importing memory from a file descriptor transfers ownership of
2984 * the file descriptor from the application to the Vulkan
2985 * implementation. The application must not perform any operations on
2986 * the file descriptor after a successful import."
2988 * If the import fails, we leave the file descriptor open.
2994 const VkImportMemoryHostPointerInfoEXT
*host_ptr_info
=
2995 vk_find_struct_const(pAllocateInfo
->pNext
,
2996 IMPORT_MEMORY_HOST_POINTER_INFO_EXT
);
2997 if (host_ptr_info
&& host_ptr_info
->handleType
) {
2998 if (host_ptr_info
->handleType
==
2999 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_MAPPED_FOREIGN_MEMORY_BIT_EXT
) {
3000 result
= vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
3004 assert(host_ptr_info
->handleType
==
3005 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
);
3007 result
= anv_bo_cache_import_host_ptr(
3008 device
, &device
->bo_cache
, host_ptr_info
->pHostPointer
,
3009 pAllocateInfo
->allocationSize
, bo_flags
, &mem
->bo
);
3011 if (result
!= VK_SUCCESS
)
3014 mem
->host_ptr
= host_ptr_info
->pHostPointer
;
3018 /* Regular allocate (not importing memory). */
3020 if (export_info
&& export_info
->handleTypes
)
3021 bo_flags
|= ANV_BO_EXTERNAL
;
3023 result
= anv_bo_cache_alloc(device
, &device
->bo_cache
,
3024 pAllocateInfo
->allocationSize
, bo_flags
,
3026 if (result
!= VK_SUCCESS
)
3029 const VkMemoryDedicatedAllocateInfo
*dedicated_info
=
3030 vk_find_struct_const(pAllocateInfo
->pNext
, MEMORY_DEDICATED_ALLOCATE_INFO
);
3031 if (dedicated_info
&& dedicated_info
->image
!= VK_NULL_HANDLE
) {
3032 ANV_FROM_HANDLE(anv_image
, image
, dedicated_info
->image
);
3034 /* Some legacy (non-modifiers) consumers need the tiling to be set on
3035 * the BO. In this case, we have a dedicated allocation.
3037 if (image
->needs_set_tiling
) {
3038 const uint32_t i915_tiling
=
3039 isl_tiling_to_i915_tiling(image
->planes
[0].surface
.isl
.tiling
);
3040 int ret
= anv_gem_set_tiling(device
, mem
->bo
->gem_handle
,
3041 image
->planes
[0].surface
.isl
.row_pitch_B
,
3044 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
3045 return vk_errorf(device
->instance
, NULL
,
3046 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
3047 "failed to set BO tiling: %m");
3053 pthread_mutex_lock(&device
->mutex
);
3054 list_addtail(&mem
->link
, &device
->memory_objects
);
3055 pthread_mutex_unlock(&device
->mutex
);
3057 *pMem
= anv_device_memory_to_handle(mem
);
3059 p_atomic_add(&pdevice
->memory
.heaps
[mem
->type
->heapIndex
].used
,
3065 vk_free2(&device
->alloc
, pAllocator
, mem
);
3070 VkResult
anv_GetMemoryFdKHR(
3072 const VkMemoryGetFdInfoKHR
* pGetFdInfo
,
3075 ANV_FROM_HANDLE(anv_device
, dev
, device_h
);
3076 ANV_FROM_HANDLE(anv_device_memory
, mem
, pGetFdInfo
->memory
);
3078 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
3080 assert(pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
3081 pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
3083 return anv_bo_cache_export(dev
, &dev
->bo_cache
, mem
->bo
, pFd
);
3086 VkResult
anv_GetMemoryFdPropertiesKHR(
3088 VkExternalMemoryHandleTypeFlagBits handleType
,
3090 VkMemoryFdPropertiesKHR
* pMemoryFdProperties
)
3092 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3093 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3095 switch (handleType
) {
3096 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
:
3097 /* dma-buf can be imported as any memory type */
3098 pMemoryFdProperties
->memoryTypeBits
=
3099 (1 << pdevice
->memory
.type_count
) - 1;
3103 /* The valid usage section for this function says:
3105 * "handleType must not be one of the handle types defined as
3108 * So opaque handle types fall into the default "unsupported" case.
3110 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
3114 VkResult
anv_GetMemoryHostPointerPropertiesEXT(
3116 VkExternalMemoryHandleTypeFlagBits handleType
,
3117 const void* pHostPointer
,
3118 VkMemoryHostPointerPropertiesEXT
* pMemoryHostPointerProperties
)
3120 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3122 assert(pMemoryHostPointerProperties
->sType
==
3123 VK_STRUCTURE_TYPE_MEMORY_HOST_POINTER_PROPERTIES_EXT
);
3125 switch (handleType
) {
3126 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
: {
3127 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3129 /* Host memory can be imported as any memory type. */
3130 pMemoryHostPointerProperties
->memoryTypeBits
=
3131 (1ull << pdevice
->memory
.type_count
) - 1;
3136 return VK_ERROR_INVALID_EXTERNAL_HANDLE
;
3140 void anv_FreeMemory(
3142 VkDeviceMemory _mem
,
3143 const VkAllocationCallbacks
* pAllocator
)
3145 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3146 ANV_FROM_HANDLE(anv_device_memory
, mem
, _mem
);
3147 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3152 pthread_mutex_lock(&device
->mutex
);
3153 list_del(&mem
->link
);
3154 pthread_mutex_unlock(&device
->mutex
);
3157 anv_UnmapMemory(_device
, _mem
);
3159 p_atomic_add(&pdevice
->memory
.heaps
[mem
->type
->heapIndex
].used
,
3162 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
3164 #if defined(ANDROID) && ANDROID_API_LEVEL >= 26
3166 AHardwareBuffer_release(mem
->ahw
);
3169 vk_free2(&device
->alloc
, pAllocator
, mem
);
3172 VkResult
anv_MapMemory(
3174 VkDeviceMemory _memory
,
3175 VkDeviceSize offset
,
3177 VkMemoryMapFlags flags
,
3180 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3181 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
3188 if (mem
->host_ptr
) {
3189 *ppData
= mem
->host_ptr
+ offset
;
3193 if (size
== VK_WHOLE_SIZE
)
3194 size
= mem
->bo
->size
- offset
;
3196 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
3198 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
3199 * assert(size != 0);
3200 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
3201 * equal to the size of the memory minus offset
3204 assert(offset
+ size
<= mem
->bo
->size
);
3206 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
3207 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
3208 * at a time is valid. We could just mmap up front and return an offset
3209 * pointer here, but that may exhaust virtual memory on 32 bit
3212 uint32_t gem_flags
= 0;
3214 if (!device
->info
.has_llc
&&
3215 (mem
->type
->propertyFlags
& VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
))
3216 gem_flags
|= I915_MMAP_WC
;
3218 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
3219 uint64_t map_offset
= offset
& ~4095ull;
3220 assert(offset
>= map_offset
);
3221 uint64_t map_size
= (offset
+ size
) - map_offset
;
3223 /* Let's map whole pages */
3224 map_size
= align_u64(map_size
, 4096);
3226 void *map
= anv_gem_mmap(device
, mem
->bo
->gem_handle
,
3227 map_offset
, map_size
, gem_flags
);
3228 if (map
== MAP_FAILED
)
3229 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
3232 mem
->map_size
= map_size
;
3234 *ppData
= mem
->map
+ (offset
- map_offset
);
3239 void anv_UnmapMemory(
3241 VkDeviceMemory _memory
)
3243 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
3245 if (mem
== NULL
|| mem
->host_ptr
)
3248 anv_gem_munmap(mem
->map
, mem
->map_size
);
3255 clflush_mapped_ranges(struct anv_device
*device
,
3257 const VkMappedMemoryRange
*ranges
)
3259 for (uint32_t i
= 0; i
< count
; i
++) {
3260 ANV_FROM_HANDLE(anv_device_memory
, mem
, ranges
[i
].memory
);
3261 if (ranges
[i
].offset
>= mem
->map_size
)
3264 gen_clflush_range(mem
->map
+ ranges
[i
].offset
,
3265 MIN2(ranges
[i
].size
, mem
->map_size
- ranges
[i
].offset
));
3269 VkResult
anv_FlushMappedMemoryRanges(
3271 uint32_t memoryRangeCount
,
3272 const VkMappedMemoryRange
* pMemoryRanges
)
3274 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3276 if (device
->info
.has_llc
)
3279 /* Make sure the writes we're flushing have landed. */
3280 __builtin_ia32_mfence();
3282 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
3287 VkResult
anv_InvalidateMappedMemoryRanges(
3289 uint32_t memoryRangeCount
,
3290 const VkMappedMemoryRange
* pMemoryRanges
)
3292 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3294 if (device
->info
.has_llc
)
3297 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
3299 /* Make sure no reads get moved up above the invalidate. */
3300 __builtin_ia32_mfence();
3305 void anv_GetBufferMemoryRequirements(
3308 VkMemoryRequirements
* pMemoryRequirements
)
3310 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
3311 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3312 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3314 /* The Vulkan spec (git aaed022) says:
3316 * memoryTypeBits is a bitfield and contains one bit set for every
3317 * supported memory type for the resource. The bit `1<<i` is set if and
3318 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3319 * structure for the physical device is supported.
3321 uint32_t memory_types
= 0;
3322 for (uint32_t i
= 0; i
< pdevice
->memory
.type_count
; i
++) {
3323 uint32_t valid_usage
= pdevice
->memory
.types
[i
].valid_buffer_usage
;
3324 if ((valid_usage
& buffer
->usage
) == buffer
->usage
)
3325 memory_types
|= (1u << i
);
3328 /* Base alignment requirement of a cache line */
3329 uint32_t alignment
= 16;
3331 /* We need an alignment of 32 for pushing UBOs */
3332 if (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
)
3333 alignment
= MAX2(alignment
, 32);
3335 pMemoryRequirements
->size
= buffer
->size
;
3336 pMemoryRequirements
->alignment
= alignment
;
3338 /* Storage and Uniform buffers should have their size aligned to
3339 * 32-bits to avoid boundary checks when last DWord is not complete.
3340 * This would ensure that not internal padding would be needed for
3343 if (device
->robust_buffer_access
&&
3344 (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
||
3345 buffer
->usage
& VK_BUFFER_USAGE_STORAGE_BUFFER_BIT
))
3346 pMemoryRequirements
->size
= align_u64(buffer
->size
, 4);
3348 pMemoryRequirements
->memoryTypeBits
= memory_types
;
3351 void anv_GetBufferMemoryRequirements2(
3353 const VkBufferMemoryRequirementsInfo2
* pInfo
,
3354 VkMemoryRequirements2
* pMemoryRequirements
)
3356 anv_GetBufferMemoryRequirements(_device
, pInfo
->buffer
,
3357 &pMemoryRequirements
->memoryRequirements
);
3359 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3360 switch (ext
->sType
) {
3361 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
3362 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
3363 requirements
->prefersDedicatedAllocation
= false;
3364 requirements
->requiresDedicatedAllocation
= false;
3369 anv_debug_ignored_stype(ext
->sType
);
3375 void anv_GetImageMemoryRequirements(
3378 VkMemoryRequirements
* pMemoryRequirements
)
3380 ANV_FROM_HANDLE(anv_image
, image
, _image
);
3381 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3382 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3384 /* The Vulkan spec (git aaed022) says:
3386 * memoryTypeBits is a bitfield and contains one bit set for every
3387 * supported memory type for the resource. The bit `1<<i` is set if and
3388 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3389 * structure for the physical device is supported.
3391 * All types are currently supported for images.
3393 uint32_t memory_types
= (1ull << pdevice
->memory
.type_count
) - 1;
3395 /* We must have image allocated or imported at this point. According to the
3396 * specification, external images must have been bound to memory before
3397 * calling GetImageMemoryRequirements.
3399 assert(image
->size
> 0);
3401 pMemoryRequirements
->size
= image
->size
;
3402 pMemoryRequirements
->alignment
= image
->alignment
;
3403 pMemoryRequirements
->memoryTypeBits
= memory_types
;
3406 void anv_GetImageMemoryRequirements2(
3408 const VkImageMemoryRequirementsInfo2
* pInfo
,
3409 VkMemoryRequirements2
* pMemoryRequirements
)
3411 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3412 ANV_FROM_HANDLE(anv_image
, image
, pInfo
->image
);
3414 anv_GetImageMemoryRequirements(_device
, pInfo
->image
,
3415 &pMemoryRequirements
->memoryRequirements
);
3417 vk_foreach_struct_const(ext
, pInfo
->pNext
) {
3418 switch (ext
->sType
) {
3419 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO
: {
3420 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3421 const VkImagePlaneMemoryRequirementsInfo
*plane_reqs
=
3422 (const VkImagePlaneMemoryRequirementsInfo
*) ext
;
3423 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
,
3424 plane_reqs
->planeAspect
);
3426 assert(image
->planes
[plane
].offset
== 0);
3428 /* The Vulkan spec (git aaed022) says:
3430 * memoryTypeBits is a bitfield and contains one bit set for every
3431 * supported memory type for the resource. The bit `1<<i` is set
3432 * if and only if the memory type `i` in the
3433 * VkPhysicalDeviceMemoryProperties structure for the physical
3434 * device is supported.
3436 * All types are currently supported for images.
3438 pMemoryRequirements
->memoryRequirements
.memoryTypeBits
=
3439 (1ull << pdevice
->memory
.type_count
) - 1;
3441 /* We must have image allocated or imported at this point. According to the
3442 * specification, external images must have been bound to memory before
3443 * calling GetImageMemoryRequirements.
3445 assert(image
->planes
[plane
].size
> 0);
3447 pMemoryRequirements
->memoryRequirements
.size
= image
->planes
[plane
].size
;
3448 pMemoryRequirements
->memoryRequirements
.alignment
=
3449 image
->planes
[plane
].alignment
;
3454 anv_debug_ignored_stype(ext
->sType
);
3459 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3460 switch (ext
->sType
) {
3461 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
3462 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
3463 if (image
->needs_set_tiling
|| image
->external_format
) {
3464 /* If we need to set the tiling for external consumers, we need a
3465 * dedicated allocation.
3467 * See also anv_AllocateMemory.
3469 requirements
->prefersDedicatedAllocation
= true;
3470 requirements
->requiresDedicatedAllocation
= true;
3472 requirements
->prefersDedicatedAllocation
= false;
3473 requirements
->requiresDedicatedAllocation
= false;
3479 anv_debug_ignored_stype(ext
->sType
);
3485 void anv_GetImageSparseMemoryRequirements(
3488 uint32_t* pSparseMemoryRequirementCount
,
3489 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
3491 *pSparseMemoryRequirementCount
= 0;
3494 void anv_GetImageSparseMemoryRequirements2(
3496 const VkImageSparseMemoryRequirementsInfo2
* pInfo
,
3497 uint32_t* pSparseMemoryRequirementCount
,
3498 VkSparseImageMemoryRequirements2
* pSparseMemoryRequirements
)
3500 *pSparseMemoryRequirementCount
= 0;
3503 void anv_GetDeviceMemoryCommitment(
3505 VkDeviceMemory memory
,
3506 VkDeviceSize
* pCommittedMemoryInBytes
)
3508 *pCommittedMemoryInBytes
= 0;
3512 anv_bind_buffer_memory(const VkBindBufferMemoryInfo
*pBindInfo
)
3514 ANV_FROM_HANDLE(anv_device_memory
, mem
, pBindInfo
->memory
);
3515 ANV_FROM_HANDLE(anv_buffer
, buffer
, pBindInfo
->buffer
);
3517 assert(pBindInfo
->sType
== VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
);
3520 assert((buffer
->usage
& mem
->type
->valid_buffer_usage
) == buffer
->usage
);
3521 buffer
->address
= (struct anv_address
) {
3523 .offset
= pBindInfo
->memoryOffset
,
3526 buffer
->address
= ANV_NULL_ADDRESS
;
3530 VkResult
anv_BindBufferMemory(
3533 VkDeviceMemory memory
,
3534 VkDeviceSize memoryOffset
)
3536 anv_bind_buffer_memory(
3537 &(VkBindBufferMemoryInfo
) {
3538 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
,
3541 .memoryOffset
= memoryOffset
,
3547 VkResult
anv_BindBufferMemory2(
3549 uint32_t bindInfoCount
,
3550 const VkBindBufferMemoryInfo
* pBindInfos
)
3552 for (uint32_t i
= 0; i
< bindInfoCount
; i
++)
3553 anv_bind_buffer_memory(&pBindInfos
[i
]);
3558 VkResult
anv_QueueBindSparse(
3560 uint32_t bindInfoCount
,
3561 const VkBindSparseInfo
* pBindInfo
,
3564 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
3565 if (anv_device_is_lost(queue
->device
))
3566 return VK_ERROR_DEVICE_LOST
;
3568 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
3573 VkResult
anv_CreateEvent(
3575 const VkEventCreateInfo
* pCreateInfo
,
3576 const VkAllocationCallbacks
* pAllocator
,
3579 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3580 struct anv_state state
;
3581 struct anv_event
*event
;
3583 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_EVENT_CREATE_INFO
);
3585 state
= anv_state_pool_alloc(&device
->dynamic_state_pool
,
3588 event
->state
= state
;
3589 event
->semaphore
= VK_EVENT_RESET
;
3591 if (!device
->info
.has_llc
) {
3592 /* Make sure the writes we're flushing have landed. */
3593 __builtin_ia32_mfence();
3594 __builtin_ia32_clflush(event
);
3597 *pEvent
= anv_event_to_handle(event
);
3602 void anv_DestroyEvent(
3605 const VkAllocationCallbacks
* pAllocator
)
3607 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3608 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3613 anv_state_pool_free(&device
->dynamic_state_pool
, event
->state
);
3616 VkResult
anv_GetEventStatus(
3620 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3621 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3623 if (anv_device_is_lost(device
))
3624 return VK_ERROR_DEVICE_LOST
;
3626 if (!device
->info
.has_llc
) {
3627 /* Invalidate read cache before reading event written by GPU. */
3628 __builtin_ia32_clflush(event
);
3629 __builtin_ia32_mfence();
3633 return event
->semaphore
;
3636 VkResult
anv_SetEvent(
3640 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3641 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3643 event
->semaphore
= VK_EVENT_SET
;
3645 if (!device
->info
.has_llc
) {
3646 /* Make sure the writes we're flushing have landed. */
3647 __builtin_ia32_mfence();
3648 __builtin_ia32_clflush(event
);
3654 VkResult
anv_ResetEvent(
3658 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3659 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3661 event
->semaphore
= VK_EVENT_RESET
;
3663 if (!device
->info
.has_llc
) {
3664 /* Make sure the writes we're flushing have landed. */
3665 __builtin_ia32_mfence();
3666 __builtin_ia32_clflush(event
);
3674 VkResult
anv_CreateBuffer(
3676 const VkBufferCreateInfo
* pCreateInfo
,
3677 const VkAllocationCallbacks
* pAllocator
,
3680 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3681 struct anv_buffer
*buffer
;
3683 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
3685 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
3686 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3688 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3690 buffer
->size
= pCreateInfo
->size
;
3691 buffer
->usage
= pCreateInfo
->usage
;
3692 buffer
->address
= ANV_NULL_ADDRESS
;
3694 *pBuffer
= anv_buffer_to_handle(buffer
);
3699 void anv_DestroyBuffer(
3702 const VkAllocationCallbacks
* pAllocator
)
3704 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3705 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
3710 vk_free2(&device
->alloc
, pAllocator
, buffer
);
3713 VkDeviceAddress
anv_GetBufferDeviceAddressEXT(
3715 const VkBufferDeviceAddressInfoEXT
* pInfo
)
3717 ANV_FROM_HANDLE(anv_buffer
, buffer
, pInfo
->buffer
);
3719 assert(buffer
->address
.bo
->flags
& EXEC_OBJECT_PINNED
);
3721 return anv_address_physical(buffer
->address
);
3725 anv_fill_buffer_surface_state(struct anv_device
*device
, struct anv_state state
,
3726 enum isl_format format
,
3727 struct anv_address address
,
3728 uint32_t range
, uint32_t stride
)
3730 isl_buffer_fill_state(&device
->isl_dev
, state
.map
,
3731 .address
= anv_address_physical(address
),
3732 .mocs
= device
->default_mocs
,
3735 .swizzle
= ISL_SWIZZLE_IDENTITY
,
3736 .stride_B
= stride
);
3739 void anv_DestroySampler(
3742 const VkAllocationCallbacks
* pAllocator
)
3744 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3745 ANV_FROM_HANDLE(anv_sampler
, sampler
, _sampler
);
3750 if (sampler
->bindless_state
.map
) {
3751 anv_state_pool_free(&device
->dynamic_state_pool
,
3752 sampler
->bindless_state
);
3755 vk_free2(&device
->alloc
, pAllocator
, sampler
);
3758 VkResult
anv_CreateFramebuffer(
3760 const VkFramebufferCreateInfo
* pCreateInfo
,
3761 const VkAllocationCallbacks
* pAllocator
,
3762 VkFramebuffer
* pFramebuffer
)
3764 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3765 struct anv_framebuffer
*framebuffer
;
3767 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
3769 size_t size
= sizeof(*framebuffer
);
3771 /* VK_KHR_imageless_framebuffer extension says:
3773 * If flags includes VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR,
3774 * parameter pAttachments is ignored.
3776 if (!(pCreateInfo
->flags
& VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR
)) {
3777 size
+= sizeof(struct anv_image_view
*) * pCreateInfo
->attachmentCount
;
3778 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
3779 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3780 if (framebuffer
== NULL
)
3781 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3783 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
3784 ANV_FROM_HANDLE(anv_image_view
, iview
, pCreateInfo
->pAttachments
[i
]);
3785 framebuffer
->attachments
[i
] = iview
;
3787 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
3789 assert(device
->enabled_extensions
.KHR_imageless_framebuffer
);
3790 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
3791 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3792 if (framebuffer
== NULL
)
3793 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3795 framebuffer
->attachment_count
= 0;
3798 framebuffer
->width
= pCreateInfo
->width
;
3799 framebuffer
->height
= pCreateInfo
->height
;
3800 framebuffer
->layers
= pCreateInfo
->layers
;
3802 *pFramebuffer
= anv_framebuffer_to_handle(framebuffer
);
3807 void anv_DestroyFramebuffer(
3810 const VkAllocationCallbacks
* pAllocator
)
3812 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3813 ANV_FROM_HANDLE(anv_framebuffer
, fb
, _fb
);
3818 vk_free2(&device
->alloc
, pAllocator
, fb
);
3821 static const VkTimeDomainEXT anv_time_domains
[] = {
3822 VK_TIME_DOMAIN_DEVICE_EXT
,
3823 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
,
3824 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
,
3827 VkResult
anv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
3828 VkPhysicalDevice physicalDevice
,
3829 uint32_t *pTimeDomainCount
,
3830 VkTimeDomainEXT
*pTimeDomains
)
3833 VK_OUTARRAY_MAKE(out
, pTimeDomains
, pTimeDomainCount
);
3835 for (d
= 0; d
< ARRAY_SIZE(anv_time_domains
); d
++) {
3836 vk_outarray_append(&out
, i
) {
3837 *i
= anv_time_domains
[d
];
3841 return vk_outarray_status(&out
);
3845 anv_clock_gettime(clockid_t clock_id
)
3847 struct timespec current
;
3850 ret
= clock_gettime(clock_id
, ¤t
);
3851 if (ret
< 0 && clock_id
== CLOCK_MONOTONIC_RAW
)
3852 ret
= clock_gettime(CLOCK_MONOTONIC
, ¤t
);
3856 return (uint64_t) current
.tv_sec
* 1000000000ULL + current
.tv_nsec
;
3859 #define TIMESTAMP 0x2358
3861 VkResult
anv_GetCalibratedTimestampsEXT(
3863 uint32_t timestampCount
,
3864 const VkCalibratedTimestampInfoEXT
*pTimestampInfos
,
3865 uint64_t *pTimestamps
,
3866 uint64_t *pMaxDeviation
)
3868 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3869 uint64_t timestamp_frequency
= device
->info
.timestamp_frequency
;
3872 uint64_t begin
, end
;
3873 uint64_t max_clock_period
= 0;
3875 begin
= anv_clock_gettime(CLOCK_MONOTONIC_RAW
);
3877 for (d
= 0; d
< timestampCount
; d
++) {
3878 switch (pTimestampInfos
[d
].timeDomain
) {
3879 case VK_TIME_DOMAIN_DEVICE_EXT
:
3880 ret
= anv_gem_reg_read(device
, TIMESTAMP
| 1,
3884 return anv_device_set_lost(device
, "Failed to read the TIMESTAMP "
3887 uint64_t device_period
= DIV_ROUND_UP(1000000000, timestamp_frequency
);
3888 max_clock_period
= MAX2(max_clock_period
, device_period
);
3890 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
:
3891 pTimestamps
[d
] = anv_clock_gettime(CLOCK_MONOTONIC
);
3892 max_clock_period
= MAX2(max_clock_period
, 1);
3895 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
:
3896 pTimestamps
[d
] = begin
;
3904 end
= anv_clock_gettime(CLOCK_MONOTONIC_RAW
);
3907 * The maximum deviation is the sum of the interval over which we
3908 * perform the sampling and the maximum period of any sampled
3909 * clock. That's because the maximum skew between any two sampled
3910 * clock edges is when the sampled clock with the largest period is
3911 * sampled at the end of that period but right at the beginning of the
3912 * sampling interval and some other clock is sampled right at the
3913 * begining of its sampling period and right at the end of the
3914 * sampling interval. Let's assume the GPU has the longest clock
3915 * period and that the application is sampling GPU and monotonic:
3918 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
3919 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
3923 * GPU -----_____-----_____-----_____-----_____
3926 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
3927 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
3929 * Interval <----------------->
3930 * Deviation <-------------------------->
3934 * m = read(monotonic) 2
3937 * We round the sample interval up by one tick to cover sampling error
3938 * in the interval clock
3941 uint64_t sample_interval
= end
- begin
+ 1;
3943 *pMaxDeviation
= sample_interval
+ max_clock_period
;
3948 /* vk_icd.h does not declare this function, so we declare it here to
3949 * suppress Wmissing-prototypes.
3951 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
3952 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
);
3954 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
3955 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
)
3957 /* For the full details on loader interface versioning, see
3958 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
3959 * What follows is a condensed summary, to help you navigate the large and
3960 * confusing official doc.
3962 * - Loader interface v0 is incompatible with later versions. We don't
3965 * - In loader interface v1:
3966 * - The first ICD entrypoint called by the loader is
3967 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
3969 * - The ICD must statically expose no other Vulkan symbol unless it is
3970 * linked with -Bsymbolic.
3971 * - Each dispatchable Vulkan handle created by the ICD must be
3972 * a pointer to a struct whose first member is VK_LOADER_DATA. The
3973 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
3974 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
3975 * vkDestroySurfaceKHR(). The ICD must be capable of working with
3976 * such loader-managed surfaces.
3978 * - Loader interface v2 differs from v1 in:
3979 * - The first ICD entrypoint called by the loader is
3980 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
3981 * statically expose this entrypoint.
3983 * - Loader interface v3 differs from v2 in:
3984 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
3985 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
3986 * because the loader no longer does so.
3988 *pSupportedVersion
= MIN2(*pSupportedVersion
, 3u);