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;
1113 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES
: {
1114 VkPhysicalDeviceMultiviewFeatures
*features
=
1115 (VkPhysicalDeviceMultiviewFeatures
*)ext
;
1116 features
->multiview
= true;
1117 features
->multiviewGeometryShader
= true;
1118 features
->multiviewTessellationShader
= true;
1122 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGELESS_FRAMEBUFFER_FEATURES_KHR
: {
1123 VkPhysicalDeviceImagelessFramebufferFeaturesKHR
*features
=
1124 (VkPhysicalDeviceImagelessFramebufferFeaturesKHR
*)ext
;
1125 features
->imagelessFramebuffer
= true;
1129 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES
: {
1130 VkPhysicalDeviceProtectedMemoryFeatures
*features
= (void *)ext
;
1131 features
->protectedMemory
= false;
1135 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES
: {
1136 VkPhysicalDeviceSamplerYcbcrConversionFeatures
*features
=
1137 (VkPhysicalDeviceSamplerYcbcrConversionFeatures
*) ext
;
1138 features
->samplerYcbcrConversion
= true;
1142 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES_EXT
: {
1143 VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*features
=
1144 (VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*)ext
;
1145 features
->scalarBlockLayout
= true;
1149 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_INT64_FEATURES_KHR
: {
1150 VkPhysicalDeviceShaderAtomicInt64FeaturesKHR
*features
= (void *)ext
;
1151 features
->shaderBufferInt64Atomics
=
1152 pdevice
->info
.gen
>= 9 && pdevice
->use_softpin
;
1153 features
->shaderSharedInt64Atomics
= VK_FALSE
;
1157 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DEMOTE_TO_HELPER_INVOCATION_FEATURES_EXT
: {
1158 VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT
*features
= (void *)ext
;
1159 features
->shaderDemoteToHelperInvocation
= true;
1163 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETERS_FEATURES
: {
1164 VkPhysicalDeviceShaderDrawParametersFeatures
*features
= (void *)ext
;
1165 features
->shaderDrawParameters
= true;
1169 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_FEATURES_EXT
: {
1170 VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT
*features
=
1171 (VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT
*)ext
;
1172 features
->texelBufferAlignment
= true;
1176 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES
: {
1177 VkPhysicalDeviceVariablePointersFeatures
*features
= (void *)ext
;
1178 features
->variablePointersStorageBuffer
= true;
1179 features
->variablePointers
= true;
1183 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT
: {
1184 VkPhysicalDeviceTransformFeedbackFeaturesEXT
*features
=
1185 (VkPhysicalDeviceTransformFeedbackFeaturesEXT
*)ext
;
1186 features
->transformFeedback
= true;
1187 features
->geometryStreams
= true;
1191 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_UNIFORM_BUFFER_STANDARD_LAYOUT_FEATURES_KHR
: {
1192 VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR
*features
=
1193 (VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR
*)ext
;
1194 features
->uniformBufferStandardLayout
= true;
1198 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT
: {
1199 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*features
=
1200 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*)ext
;
1201 features
->vertexAttributeInstanceRateDivisor
= true;
1202 features
->vertexAttributeInstanceRateZeroDivisor
= true;
1206 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_YCBCR_IMAGE_ARRAYS_FEATURES_EXT
: {
1207 VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*features
=
1208 (VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*)ext
;
1209 features
->ycbcrImageArrays
= true;
1214 anv_debug_ignored_stype(ext
->sType
);
1220 #define MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS 64
1222 #define MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS 64
1223 #define MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS 256
1225 void anv_GetPhysicalDeviceProperties(
1226 VkPhysicalDevice physicalDevice
,
1227 VkPhysicalDeviceProperties
* pProperties
)
1229 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1230 const struct gen_device_info
*devinfo
= &pdevice
->info
;
1232 /* See assertions made when programming the buffer surface state. */
1233 const uint32_t max_raw_buffer_sz
= devinfo
->gen
>= 7 ?
1234 (1ul << 30) : (1ul << 27);
1236 const uint32_t max_ssbos
= pdevice
->has_a64_buffer_access
? UINT16_MAX
: 64;
1237 const uint32_t max_textures
=
1238 pdevice
->has_bindless_images
? UINT16_MAX
: 128;
1239 const uint32_t max_samplers
=
1240 pdevice
->has_bindless_samplers
? UINT16_MAX
:
1241 (devinfo
->gen
>= 8 || devinfo
->is_haswell
) ? 128 : 16;
1242 const uint32_t max_images
=
1243 pdevice
->has_bindless_images
? UINT16_MAX
: MAX_IMAGES
;
1245 /* The moment we have anything bindless, claim a high per-stage limit */
1246 const uint32_t max_per_stage
=
1247 pdevice
->has_a64_buffer_access
? UINT32_MAX
:
1248 MAX_BINDING_TABLE_SIZE
- MAX_RTS
;
1250 VkSampleCountFlags sample_counts
=
1251 isl_device_get_sample_counts(&pdevice
->isl_dev
);
1254 VkPhysicalDeviceLimits limits
= {
1255 .maxImageDimension1D
= (1 << 14),
1256 .maxImageDimension2D
= (1 << 14),
1257 .maxImageDimension3D
= (1 << 11),
1258 .maxImageDimensionCube
= (1 << 14),
1259 .maxImageArrayLayers
= (1 << 11),
1260 .maxTexelBufferElements
= 128 * 1024 * 1024,
1261 .maxUniformBufferRange
= (1ul << 27),
1262 .maxStorageBufferRange
= max_raw_buffer_sz
,
1263 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
1264 .maxMemoryAllocationCount
= UINT32_MAX
,
1265 .maxSamplerAllocationCount
= 64 * 1024,
1266 .bufferImageGranularity
= 64, /* A cache line */
1267 .sparseAddressSpaceSize
= 0,
1268 .maxBoundDescriptorSets
= MAX_SETS
,
1269 .maxPerStageDescriptorSamplers
= max_samplers
,
1270 .maxPerStageDescriptorUniformBuffers
= MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
,
1271 .maxPerStageDescriptorStorageBuffers
= max_ssbos
,
1272 .maxPerStageDescriptorSampledImages
= max_textures
,
1273 .maxPerStageDescriptorStorageImages
= max_images
,
1274 .maxPerStageDescriptorInputAttachments
= MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS
,
1275 .maxPerStageResources
= max_per_stage
,
1276 .maxDescriptorSetSamplers
= 6 * max_samplers
, /* number of stages * maxPerStageDescriptorSamplers */
1277 .maxDescriptorSetUniformBuffers
= 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
, /* number of stages * maxPerStageDescriptorUniformBuffers */
1278 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
1279 .maxDescriptorSetStorageBuffers
= 6 * max_ssbos
, /* number of stages * maxPerStageDescriptorStorageBuffers */
1280 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
1281 .maxDescriptorSetSampledImages
= 6 * max_textures
, /* number of stages * maxPerStageDescriptorSampledImages */
1282 .maxDescriptorSetStorageImages
= 6 * max_images
, /* number of stages * maxPerStageDescriptorStorageImages */
1283 .maxDescriptorSetInputAttachments
= MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS
,
1284 .maxVertexInputAttributes
= MAX_VBS
,
1285 .maxVertexInputBindings
= MAX_VBS
,
1286 .maxVertexInputAttributeOffset
= 2047,
1287 .maxVertexInputBindingStride
= 2048,
1288 .maxVertexOutputComponents
= 128,
1289 .maxTessellationGenerationLevel
= 64,
1290 .maxTessellationPatchSize
= 32,
1291 .maxTessellationControlPerVertexInputComponents
= 128,
1292 .maxTessellationControlPerVertexOutputComponents
= 128,
1293 .maxTessellationControlPerPatchOutputComponents
= 128,
1294 .maxTessellationControlTotalOutputComponents
= 2048,
1295 .maxTessellationEvaluationInputComponents
= 128,
1296 .maxTessellationEvaluationOutputComponents
= 128,
1297 .maxGeometryShaderInvocations
= 32,
1298 .maxGeometryInputComponents
= 64,
1299 .maxGeometryOutputComponents
= 128,
1300 .maxGeometryOutputVertices
= 256,
1301 .maxGeometryTotalOutputComponents
= 1024,
1302 .maxFragmentInputComponents
= 116, /* 128 components - (PSIZ, CLIP_DIST0, CLIP_DIST1) */
1303 .maxFragmentOutputAttachments
= 8,
1304 .maxFragmentDualSrcAttachments
= 1,
1305 .maxFragmentCombinedOutputResources
= 8,
1306 .maxComputeSharedMemorySize
= 64 * 1024,
1307 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
1308 .maxComputeWorkGroupInvocations
= 32 * devinfo
->max_cs_threads
,
1309 .maxComputeWorkGroupSize
= {
1310 16 * devinfo
->max_cs_threads
,
1311 16 * devinfo
->max_cs_threads
,
1312 16 * devinfo
->max_cs_threads
,
1314 .subPixelPrecisionBits
= 8,
1315 .subTexelPrecisionBits
= 8,
1316 .mipmapPrecisionBits
= 8,
1317 .maxDrawIndexedIndexValue
= UINT32_MAX
,
1318 .maxDrawIndirectCount
= UINT32_MAX
,
1319 .maxSamplerLodBias
= 16,
1320 .maxSamplerAnisotropy
= 16,
1321 .maxViewports
= MAX_VIEWPORTS
,
1322 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
1323 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
1324 .viewportSubPixelBits
= 13, /* We take a float? */
1325 .minMemoryMapAlignment
= 4096, /* A page */
1326 /* The dataport requires texel alignment so we need to assume a worst
1327 * case of R32G32B32A32 which is 16 bytes.
1329 .minTexelBufferOffsetAlignment
= 16,
1330 /* We need 16 for UBO block reads to work and 32 for push UBOs */
1331 .minUniformBufferOffsetAlignment
= 32,
1332 .minStorageBufferOffsetAlignment
= 4,
1333 .minTexelOffset
= -8,
1334 .maxTexelOffset
= 7,
1335 .minTexelGatherOffset
= -32,
1336 .maxTexelGatherOffset
= 31,
1337 .minInterpolationOffset
= -0.5,
1338 .maxInterpolationOffset
= 0.4375,
1339 .subPixelInterpolationOffsetBits
= 4,
1340 .maxFramebufferWidth
= (1 << 14),
1341 .maxFramebufferHeight
= (1 << 14),
1342 .maxFramebufferLayers
= (1 << 11),
1343 .framebufferColorSampleCounts
= sample_counts
,
1344 .framebufferDepthSampleCounts
= sample_counts
,
1345 .framebufferStencilSampleCounts
= sample_counts
,
1346 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
1347 .maxColorAttachments
= MAX_RTS
,
1348 .sampledImageColorSampleCounts
= sample_counts
,
1349 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
1350 .sampledImageDepthSampleCounts
= sample_counts
,
1351 .sampledImageStencilSampleCounts
= sample_counts
,
1352 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
1353 .maxSampleMaskWords
= 1,
1354 .timestampComputeAndGraphics
= true,
1355 .timestampPeriod
= 1000000000.0 / devinfo
->timestamp_frequency
,
1356 .maxClipDistances
= 8,
1357 .maxCullDistances
= 8,
1358 .maxCombinedClipAndCullDistances
= 8,
1359 .discreteQueuePriorities
= 2,
1360 .pointSizeRange
= { 0.125, 255.875 },
1363 (devinfo
->gen
>= 9 || devinfo
->is_cherryview
) ?
1364 2047.9921875 : 7.9921875,
1366 .pointSizeGranularity
= (1.0 / 8.0),
1367 .lineWidthGranularity
= (1.0 / 128.0),
1368 .strictLines
= false,
1369 .standardSampleLocations
= true,
1370 .optimalBufferCopyOffsetAlignment
= 128,
1371 .optimalBufferCopyRowPitchAlignment
= 128,
1372 .nonCoherentAtomSize
= 64,
1375 *pProperties
= (VkPhysicalDeviceProperties
) {
1376 .apiVersion
= anv_physical_device_api_version(pdevice
),
1377 .driverVersion
= vk_get_driver_version(),
1379 .deviceID
= pdevice
->chipset_id
,
1380 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
1382 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
1385 snprintf(pProperties
->deviceName
, sizeof(pProperties
->deviceName
),
1386 "%s", pdevice
->name
);
1387 memcpy(pProperties
->pipelineCacheUUID
,
1388 pdevice
->pipeline_cache_uuid
, VK_UUID_SIZE
);
1391 void anv_GetPhysicalDeviceProperties2(
1392 VkPhysicalDevice physicalDevice
,
1393 VkPhysicalDeviceProperties2
* pProperties
)
1395 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1397 anv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
1399 vk_foreach_struct(ext
, pProperties
->pNext
) {
1400 switch (ext
->sType
) {
1401 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_STENCIL_RESOLVE_PROPERTIES_KHR
: {
1402 VkPhysicalDeviceDepthStencilResolvePropertiesKHR
*props
=
1403 (VkPhysicalDeviceDepthStencilResolvePropertiesKHR
*)ext
;
1405 /* We support all of the depth resolve modes */
1406 props
->supportedDepthResolveModes
=
1407 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
|
1408 VK_RESOLVE_MODE_AVERAGE_BIT_KHR
|
1409 VK_RESOLVE_MODE_MIN_BIT_KHR
|
1410 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1412 /* Average doesn't make sense for stencil so we don't support that */
1413 props
->supportedStencilResolveModes
=
1414 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
;
1415 if (pdevice
->info
.gen
>= 8) {
1416 /* The advanced stencil resolve modes currently require stencil
1417 * sampling be supported by the hardware.
1419 props
->supportedStencilResolveModes
|=
1420 VK_RESOLVE_MODE_MIN_BIT_KHR
|
1421 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1424 props
->independentResolveNone
= true;
1425 props
->independentResolve
= true;
1429 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES_EXT
: {
1430 VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*props
=
1431 (VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*)ext
;
1433 /* It's a bit hard to exactly map our implementation to the limits
1434 * described here. The bindless surface handle in the extended
1435 * message descriptors is 20 bits and it's an index into the table of
1436 * RENDER_SURFACE_STATE structs that starts at bindless surface base
1437 * address. Given that most things consume two surface states per
1438 * view (general/sampled for textures and write-only/read-write for
1439 * images), we claim 2^19 things.
1441 * For SSBOs, we just use A64 messages so there is no real limit
1442 * there beyond the limit on the total size of a descriptor set.
1444 const unsigned max_bindless_views
= 1 << 19;
1446 props
->maxUpdateAfterBindDescriptorsInAllPools
= max_bindless_views
;
1447 props
->shaderUniformBufferArrayNonUniformIndexingNative
= false;
1448 props
->shaderSampledImageArrayNonUniformIndexingNative
= false;
1449 props
->shaderStorageBufferArrayNonUniformIndexingNative
= true;
1450 props
->shaderStorageImageArrayNonUniformIndexingNative
= false;
1451 props
->shaderInputAttachmentArrayNonUniformIndexingNative
= false;
1452 props
->robustBufferAccessUpdateAfterBind
= true;
1453 props
->quadDivergentImplicitLod
= false;
1454 props
->maxPerStageDescriptorUpdateAfterBindSamplers
= max_bindless_views
;
1455 props
->maxPerStageDescriptorUpdateAfterBindUniformBuffers
= MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
;
1456 props
->maxPerStageDescriptorUpdateAfterBindStorageBuffers
= UINT32_MAX
;
1457 props
->maxPerStageDescriptorUpdateAfterBindSampledImages
= max_bindless_views
;
1458 props
->maxPerStageDescriptorUpdateAfterBindStorageImages
= max_bindless_views
;
1459 props
->maxPerStageDescriptorUpdateAfterBindInputAttachments
= MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS
;
1460 props
->maxPerStageUpdateAfterBindResources
= UINT32_MAX
;
1461 props
->maxDescriptorSetUpdateAfterBindSamplers
= max_bindless_views
;
1462 props
->maxDescriptorSetUpdateAfterBindUniformBuffers
= 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
;
1463 props
->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2;
1464 props
->maxDescriptorSetUpdateAfterBindStorageBuffers
= UINT32_MAX
;
1465 props
->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2;
1466 props
->maxDescriptorSetUpdateAfterBindSampledImages
= max_bindless_views
;
1467 props
->maxDescriptorSetUpdateAfterBindStorageImages
= max_bindless_views
;
1468 props
->maxDescriptorSetUpdateAfterBindInputAttachments
= MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS
;
1472 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR
: {
1473 VkPhysicalDeviceDriverPropertiesKHR
*driver_props
=
1474 (VkPhysicalDeviceDriverPropertiesKHR
*) ext
;
1476 driver_props
->driverID
= VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA_KHR
;
1477 snprintf(driver_props
->driverName
, VK_MAX_DRIVER_NAME_SIZE_KHR
,
1478 "Intel open-source Mesa driver");
1480 snprintf(driver_props
->driverInfo
, VK_MAX_DRIVER_INFO_SIZE_KHR
,
1481 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1
);
1483 driver_props
->conformanceVersion
= (VkConformanceVersionKHR
) {
1492 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT
: {
1493 VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*props
=
1494 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*) ext
;
1495 /* Userptr needs page aligned memory. */
1496 props
->minImportedHostPointerAlignment
= 4096;
1500 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES
: {
1501 VkPhysicalDeviceIDProperties
*id_props
=
1502 (VkPhysicalDeviceIDProperties
*)ext
;
1503 memcpy(id_props
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
1504 memcpy(id_props
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
1505 /* The LUID is for Windows. */
1506 id_props
->deviceLUIDValid
= false;
1510 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_PROPERTIES_EXT
: {
1511 VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*props
=
1512 (VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*)ext
;
1513 props
->maxInlineUniformBlockSize
= MAX_INLINE_UNIFORM_BLOCK_SIZE
;
1514 props
->maxPerStageDescriptorInlineUniformBlocks
=
1515 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1516 props
->maxPerStageDescriptorUpdateAfterBindInlineUniformBlocks
=
1517 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1518 props
->maxDescriptorSetInlineUniformBlocks
=
1519 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1520 props
->maxDescriptorSetUpdateAfterBindInlineUniformBlocks
=
1521 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1525 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_PROPERTIES_EXT
: {
1526 VkPhysicalDeviceLineRasterizationPropertiesEXT
*props
=
1527 (VkPhysicalDeviceLineRasterizationPropertiesEXT
*)ext
;
1528 /* In the Skylake PRM Vol. 7, subsection titled "GIQ (Diamond)
1529 * Sampling Rules - Legacy Mode", it says the following:
1531 * "Note that the device divides a pixel into a 16x16 array of
1532 * subpixels, referenced by their upper left corners."
1534 * This is the only known reference in the PRMs to the subpixel
1535 * precision of line rasterization and a "16x16 array of subpixels"
1536 * implies 4 subpixel precision bits. Empirical testing has shown
1537 * that 4 subpixel precision bits applies to all line rasterization
1540 props
->lineSubPixelPrecisionBits
= 4;
1544 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES
: {
1545 VkPhysicalDeviceMaintenance3Properties
*props
=
1546 (VkPhysicalDeviceMaintenance3Properties
*)ext
;
1547 /* This value doesn't matter for us today as our per-stage
1548 * descriptors are the real limit.
1550 props
->maxPerSetDescriptors
= 1024;
1551 props
->maxMemoryAllocationSize
= MAX_MEMORY_ALLOCATION_SIZE
;
1555 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES
: {
1556 VkPhysicalDeviceMultiviewProperties
*properties
=
1557 (VkPhysicalDeviceMultiviewProperties
*)ext
;
1558 properties
->maxMultiviewViewCount
= 16;
1559 properties
->maxMultiviewInstanceIndex
= UINT32_MAX
/ 16;
1563 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT
: {
1564 VkPhysicalDevicePCIBusInfoPropertiesEXT
*properties
=
1565 (VkPhysicalDevicePCIBusInfoPropertiesEXT
*)ext
;
1566 properties
->pciDomain
= pdevice
->pci_info
.domain
;
1567 properties
->pciBus
= pdevice
->pci_info
.bus
;
1568 properties
->pciDevice
= pdevice
->pci_info
.device
;
1569 properties
->pciFunction
= pdevice
->pci_info
.function
;
1573 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES
: {
1574 VkPhysicalDevicePointClippingProperties
*properties
=
1575 (VkPhysicalDevicePointClippingProperties
*) ext
;
1576 properties
->pointClippingBehavior
= VK_POINT_CLIPPING_BEHAVIOR_USER_CLIP_PLANES_ONLY
;
1580 #pragma GCC diagnostic push
1581 #pragma GCC diagnostic ignored "-Wswitch"
1582 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PRESENTATION_PROPERTIES_ANDROID
: {
1583 VkPhysicalDevicePresentationPropertiesANDROID
*props
=
1584 (VkPhysicalDevicePresentationPropertiesANDROID
*)ext
;
1585 props
->sharedImage
= VK_FALSE
;
1588 #pragma GCC diagnostic pop
1590 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES
: {
1591 VkPhysicalDeviceProtectedMemoryProperties
*props
=
1592 (VkPhysicalDeviceProtectedMemoryProperties
*)ext
;
1593 props
->protectedNoFault
= false;
1597 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
1598 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
1599 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
1601 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
1605 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT
: {
1606 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*properties
=
1607 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*)ext
;
1608 properties
->filterMinmaxImageComponentMapping
= pdevice
->info
.gen
>= 9;
1609 properties
->filterMinmaxSingleComponentFormats
= true;
1613 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES
: {
1614 VkPhysicalDeviceSubgroupProperties
*properties
= (void *)ext
;
1616 properties
->subgroupSize
= BRW_SUBGROUP_SIZE
;
1618 VkShaderStageFlags scalar_stages
= 0;
1619 for (unsigned stage
= 0; stage
< MESA_SHADER_STAGES
; stage
++) {
1620 if (pdevice
->compiler
->scalar_stage
[stage
])
1621 scalar_stages
|= mesa_to_vk_shader_stage(stage
);
1623 properties
->supportedStages
= scalar_stages
;
1625 properties
->supportedOperations
= VK_SUBGROUP_FEATURE_BASIC_BIT
|
1626 VK_SUBGROUP_FEATURE_VOTE_BIT
|
1627 VK_SUBGROUP_FEATURE_BALLOT_BIT
|
1628 VK_SUBGROUP_FEATURE_SHUFFLE_BIT
|
1629 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT
|
1630 VK_SUBGROUP_FEATURE_QUAD_BIT
;
1631 if (pdevice
->info
.gen
>= 8) {
1632 /* TODO: There's no technical reason why these can't be made to
1633 * work on gen7 but they don't at the moment so it's best to leave
1634 * the feature disabled than enabled and broken.
1636 properties
->supportedOperations
|=
1637 VK_SUBGROUP_FEATURE_ARITHMETIC_BIT
|
1638 VK_SUBGROUP_FEATURE_CLUSTERED_BIT
;
1640 properties
->quadOperationsInAllStages
= pdevice
->info
.gen
>= 8;
1644 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_PROPERTIES_EXT
: {
1645 VkPhysicalDeviceSubgroupSizeControlPropertiesEXT
*props
=
1646 (VkPhysicalDeviceSubgroupSizeControlPropertiesEXT
*)ext
;
1647 STATIC_ASSERT(8 <= BRW_SUBGROUP_SIZE
&& BRW_SUBGROUP_SIZE
<= 32);
1648 props
->minSubgroupSize
= 8;
1649 props
->maxSubgroupSize
= 32;
1650 props
->maxComputeWorkgroupSubgroups
= pdevice
->info
.max_cs_threads
;
1651 props
->requiredSubgroupSizeStages
= VK_SHADER_STAGE_COMPUTE_BIT
;
1655 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_PROPERTIES_EXT
: {
1656 VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT
*props
=
1657 (VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT
*)ext
;
1659 /* From the SKL PRM Vol. 2d, docs for RENDER_SURFACE_STATE::Surface
1662 * "For SURFTYPE_BUFFER non-rendertarget surfaces, this field
1663 * specifies the base address of the first element of the surface,
1664 * computed in software by adding the surface base address to the
1665 * byte offset of the element in the buffer. The base address must
1666 * be aligned to element size."
1668 * The typed dataport messages require that things be texel aligned.
1669 * Otherwise, we may just load/store the wrong data or, in the worst
1670 * case, there may be hangs.
1672 props
->storageTexelBufferOffsetAlignmentBytes
= 16;
1673 props
->storageTexelBufferOffsetSingleTexelAlignment
= true;
1675 /* The sampler, however, is much more forgiving and it can handle
1676 * arbitrary byte alignment for linear and buffer surfaces. It's
1677 * hard to find a good PRM citation for this but years of empirical
1678 * experience demonstrate that this is true.
1680 props
->uniformTexelBufferOffsetAlignmentBytes
= 1;
1681 props
->uniformTexelBufferOffsetSingleTexelAlignment
= false;
1685 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT
: {
1686 VkPhysicalDeviceTransformFeedbackPropertiesEXT
*props
=
1687 (VkPhysicalDeviceTransformFeedbackPropertiesEXT
*)ext
;
1689 props
->maxTransformFeedbackStreams
= MAX_XFB_STREAMS
;
1690 props
->maxTransformFeedbackBuffers
= MAX_XFB_BUFFERS
;
1691 props
->maxTransformFeedbackBufferSize
= (1ull << 32);
1692 props
->maxTransformFeedbackStreamDataSize
= 128 * 4;
1693 props
->maxTransformFeedbackBufferDataSize
= 128 * 4;
1694 props
->maxTransformFeedbackBufferDataStride
= 2048;
1695 props
->transformFeedbackQueries
= true;
1696 props
->transformFeedbackStreamsLinesTriangles
= false;
1697 props
->transformFeedbackRasterizationStreamSelect
= false;
1698 props
->transformFeedbackDraw
= true;
1702 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT
: {
1703 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*props
=
1704 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*)ext
;
1705 /* We have to restrict this a bit for multiview */
1706 props
->maxVertexAttribDivisor
= UINT32_MAX
/ 16;
1711 anv_debug_ignored_stype(ext
->sType
);
1717 /* We support exactly one queue family. */
1718 static const VkQueueFamilyProperties
1719 anv_queue_family_properties
= {
1720 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
1721 VK_QUEUE_COMPUTE_BIT
|
1722 VK_QUEUE_TRANSFER_BIT
,
1724 .timestampValidBits
= 36, /* XXX: Real value here */
1725 .minImageTransferGranularity
= { 1, 1, 1 },
1728 void anv_GetPhysicalDeviceQueueFamilyProperties(
1729 VkPhysicalDevice physicalDevice
,
1731 VkQueueFamilyProperties
* pQueueFamilyProperties
)
1733 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pCount
);
1735 vk_outarray_append(&out
, p
) {
1736 *p
= anv_queue_family_properties
;
1740 void anv_GetPhysicalDeviceQueueFamilyProperties2(
1741 VkPhysicalDevice physicalDevice
,
1742 uint32_t* pQueueFamilyPropertyCount
,
1743 VkQueueFamilyProperties2
* pQueueFamilyProperties
)
1746 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pQueueFamilyPropertyCount
);
1748 vk_outarray_append(&out
, p
) {
1749 p
->queueFamilyProperties
= anv_queue_family_properties
;
1751 vk_foreach_struct(s
, p
->pNext
) {
1752 anv_debug_ignored_stype(s
->sType
);
1757 void anv_GetPhysicalDeviceMemoryProperties(
1758 VkPhysicalDevice physicalDevice
,
1759 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
1761 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
1763 pMemoryProperties
->memoryTypeCount
= physical_device
->memory
.type_count
;
1764 for (uint32_t i
= 0; i
< physical_device
->memory
.type_count
; i
++) {
1765 pMemoryProperties
->memoryTypes
[i
] = (VkMemoryType
) {
1766 .propertyFlags
= physical_device
->memory
.types
[i
].propertyFlags
,
1767 .heapIndex
= physical_device
->memory
.types
[i
].heapIndex
,
1771 pMemoryProperties
->memoryHeapCount
= physical_device
->memory
.heap_count
;
1772 for (uint32_t i
= 0; i
< physical_device
->memory
.heap_count
; i
++) {
1773 pMemoryProperties
->memoryHeaps
[i
] = (VkMemoryHeap
) {
1774 .size
= physical_device
->memory
.heaps
[i
].size
,
1775 .flags
= physical_device
->memory
.heaps
[i
].flags
,
1781 anv_get_memory_budget(VkPhysicalDevice physicalDevice
,
1782 VkPhysicalDeviceMemoryBudgetPropertiesEXT
*memoryBudget
)
1784 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
1785 uint64_t sys_available
= get_available_system_memory();
1786 assert(sys_available
> 0);
1788 VkDeviceSize total_heaps_size
= 0;
1789 for (size_t i
= 0; i
< device
->memory
.heap_count
; i
++)
1790 total_heaps_size
+= device
->memory
.heaps
[i
].size
;
1792 for (size_t i
= 0; i
< device
->memory
.heap_count
; i
++) {
1793 VkDeviceSize heap_size
= device
->memory
.heaps
[i
].size
;
1794 VkDeviceSize heap_used
= device
->memory
.heaps
[i
].used
;
1795 VkDeviceSize heap_budget
;
1797 double heap_proportion
= (double) heap_size
/ total_heaps_size
;
1798 VkDeviceSize sys_available_prop
= sys_available
* heap_proportion
;
1801 * Let's not incite the app to starve the system: report at most 90% of
1802 * available system memory.
1804 uint64_t heap_available
= sys_available_prop
* 9 / 10;
1805 heap_budget
= MIN2(heap_size
, heap_used
+ heap_available
);
1808 * Round down to the nearest MB
1810 heap_budget
&= ~((1ull << 20) - 1);
1813 * The heapBudget value must be non-zero for array elements less than
1814 * VkPhysicalDeviceMemoryProperties::memoryHeapCount. The heapBudget
1815 * value must be less than or equal to VkMemoryHeap::size for each heap.
1817 assert(0 < heap_budget
&& heap_budget
<= heap_size
);
1819 memoryBudget
->heapUsage
[i
] = heap_used
;
1820 memoryBudget
->heapBudget
[i
] = heap_budget
;
1823 /* The heapBudget and heapUsage values must be zero for array elements
1824 * greater than or equal to VkPhysicalDeviceMemoryProperties::memoryHeapCount
1826 for (uint32_t i
= device
->memory
.heap_count
; i
< VK_MAX_MEMORY_HEAPS
; i
++) {
1827 memoryBudget
->heapBudget
[i
] = 0;
1828 memoryBudget
->heapUsage
[i
] = 0;
1832 void anv_GetPhysicalDeviceMemoryProperties2(
1833 VkPhysicalDevice physicalDevice
,
1834 VkPhysicalDeviceMemoryProperties2
* pMemoryProperties
)
1836 anv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
1837 &pMemoryProperties
->memoryProperties
);
1839 vk_foreach_struct(ext
, pMemoryProperties
->pNext
) {
1840 switch (ext
->sType
) {
1841 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT
:
1842 anv_get_memory_budget(physicalDevice
, (void*)ext
);
1845 anv_debug_ignored_stype(ext
->sType
);
1852 anv_GetDeviceGroupPeerMemoryFeatures(
1855 uint32_t localDeviceIndex
,
1856 uint32_t remoteDeviceIndex
,
1857 VkPeerMemoryFeatureFlags
* pPeerMemoryFeatures
)
1859 assert(localDeviceIndex
== 0 && remoteDeviceIndex
== 0);
1860 *pPeerMemoryFeatures
= VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT
|
1861 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT
|
1862 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT
|
1863 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT
;
1866 PFN_vkVoidFunction
anv_GetInstanceProcAddr(
1867 VkInstance _instance
,
1870 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1872 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
1873 * when we have to return valid function pointers, NULL, or it's left
1874 * undefined. See the table for exact details.
1879 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
1880 if (strcmp(pName, "vk" #entrypoint) == 0) \
1881 return (PFN_vkVoidFunction)anv_##entrypoint
1883 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties
);
1884 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties
);
1885 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceVersion
);
1886 LOOKUP_ANV_ENTRYPOINT(CreateInstance
);
1888 #undef LOOKUP_ANV_ENTRYPOINT
1890 if (instance
== NULL
)
1893 int idx
= anv_get_instance_entrypoint_index(pName
);
1895 return instance
->dispatch
.entrypoints
[idx
];
1897 idx
= anv_get_device_entrypoint_index(pName
);
1899 return instance
->device_dispatch
.entrypoints
[idx
];
1904 /* With version 1+ of the loader interface the ICD should expose
1905 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
1908 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1909 VkInstance instance
,
1913 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1914 VkInstance instance
,
1917 return anv_GetInstanceProcAddr(instance
, pName
);
1920 PFN_vkVoidFunction
anv_GetDeviceProcAddr(
1924 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1926 if (!device
|| !pName
)
1929 int idx
= anv_get_device_entrypoint_index(pName
);
1933 return device
->dispatch
.entrypoints
[idx
];
1937 anv_CreateDebugReportCallbackEXT(VkInstance _instance
,
1938 const VkDebugReportCallbackCreateInfoEXT
* pCreateInfo
,
1939 const VkAllocationCallbacks
* pAllocator
,
1940 VkDebugReportCallbackEXT
* pCallback
)
1942 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1943 return vk_create_debug_report_callback(&instance
->debug_report_callbacks
,
1944 pCreateInfo
, pAllocator
, &instance
->alloc
,
1949 anv_DestroyDebugReportCallbackEXT(VkInstance _instance
,
1950 VkDebugReportCallbackEXT _callback
,
1951 const VkAllocationCallbacks
* pAllocator
)
1953 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1954 vk_destroy_debug_report_callback(&instance
->debug_report_callbacks
,
1955 _callback
, pAllocator
, &instance
->alloc
);
1959 anv_DebugReportMessageEXT(VkInstance _instance
,
1960 VkDebugReportFlagsEXT flags
,
1961 VkDebugReportObjectTypeEXT objectType
,
1964 int32_t messageCode
,
1965 const char* pLayerPrefix
,
1966 const char* pMessage
)
1968 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1969 vk_debug_report(&instance
->debug_report_callbacks
, flags
, objectType
,
1970 object
, location
, messageCode
, pLayerPrefix
, pMessage
);
1974 anv_queue_init(struct anv_device
*device
, struct anv_queue
*queue
)
1976 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1977 queue
->device
= device
;
1982 anv_queue_finish(struct anv_queue
*queue
)
1986 static struct anv_state
1987 anv_state_pool_emit_data(struct anv_state_pool
*pool
, size_t size
, size_t align
, const void *p
)
1989 struct anv_state state
;
1991 state
= anv_state_pool_alloc(pool
, size
, align
);
1992 memcpy(state
.map
, p
, size
);
1997 /* Haswell border color is a bit of a disaster. Float and unorm formats use a
1998 * straightforward 32-bit float color in the first 64 bytes. Instead of using
1999 * a nice float/integer union like Gen8+, Haswell specifies the integer border
2000 * color as a separate entry /after/ the float color. The layout of this entry
2001 * also depends on the format's bpp (with extra hacks for RG32), and overlaps.
2003 * Since we don't know the format/bpp, we can't make any of the border colors
2004 * containing '1' work for all formats, as it would be in the wrong place for
2005 * some of them. We opt to make 32-bit integers work as this seems like the
2006 * most common option. Fortunately, transparent black works regardless, as
2007 * all zeroes is the same in every bit-size.
2009 struct hsw_border_color
{
2013 uint32_t _pad1
[108];
2016 struct gen8_border_color
{
2021 /* Pad out to 64 bytes */
2026 anv_device_init_border_colors(struct anv_device
*device
)
2028 if (device
->info
.is_haswell
) {
2029 static const struct hsw_border_color border_colors
[] = {
2030 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
2031 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
2032 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
2033 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
2034 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
2035 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
2038 device
->border_colors
=
2039 anv_state_pool_emit_data(&device
->dynamic_state_pool
,
2040 sizeof(border_colors
), 512, border_colors
);
2042 static const struct gen8_border_color border_colors
[] = {
2043 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
2044 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
2045 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
2046 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
2047 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
2048 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
2051 device
->border_colors
=
2052 anv_state_pool_emit_data(&device
->dynamic_state_pool
,
2053 sizeof(border_colors
), 64, border_colors
);
2058 anv_device_init_trivial_batch(struct anv_device
*device
)
2060 anv_bo_init_new(&device
->trivial_batch_bo
, device
, 4096);
2062 if (device
->instance
->physicalDevice
.has_exec_async
)
2063 device
->trivial_batch_bo
.flags
|= EXEC_OBJECT_ASYNC
;
2065 if (device
->instance
->physicalDevice
.use_softpin
)
2066 device
->trivial_batch_bo
.flags
|= EXEC_OBJECT_PINNED
;
2068 anv_vma_alloc(device
, &device
->trivial_batch_bo
);
2070 void *map
= anv_gem_mmap(device
, device
->trivial_batch_bo
.gem_handle
,
2073 struct anv_batch batch
= {
2079 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
2080 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
2082 if (!device
->info
.has_llc
)
2083 gen_clflush_range(map
, batch
.next
- map
);
2085 anv_gem_munmap(map
, device
->trivial_batch_bo
.size
);
2088 VkResult
anv_EnumerateDeviceExtensionProperties(
2089 VkPhysicalDevice physicalDevice
,
2090 const char* pLayerName
,
2091 uint32_t* pPropertyCount
,
2092 VkExtensionProperties
* pProperties
)
2094 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
2095 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
2097 for (int i
= 0; i
< ANV_DEVICE_EXTENSION_COUNT
; i
++) {
2098 if (device
->supported_extensions
.extensions
[i
]) {
2099 vk_outarray_append(&out
, prop
) {
2100 *prop
= anv_device_extensions
[i
];
2105 return vk_outarray_status(&out
);
2109 anv_device_init_dispatch(struct anv_device
*device
)
2111 const struct anv_device_dispatch_table
*genX_table
;
2112 switch (device
->info
.gen
) {
2114 genX_table
= &gen11_device_dispatch_table
;
2117 genX_table
= &gen10_device_dispatch_table
;
2120 genX_table
= &gen9_device_dispatch_table
;
2123 genX_table
= &gen8_device_dispatch_table
;
2126 if (device
->info
.is_haswell
)
2127 genX_table
= &gen75_device_dispatch_table
;
2129 genX_table
= &gen7_device_dispatch_table
;
2132 unreachable("unsupported gen\n");
2135 for (unsigned i
= 0; i
< ARRAY_SIZE(device
->dispatch
.entrypoints
); i
++) {
2136 /* Vulkan requires that entrypoints for extensions which have not been
2137 * enabled must not be advertised.
2139 if (!anv_device_entrypoint_is_enabled(i
, device
->instance
->app_info
.api_version
,
2140 &device
->instance
->enabled_extensions
,
2141 &device
->enabled_extensions
)) {
2142 device
->dispatch
.entrypoints
[i
] = NULL
;
2143 } else if (genX_table
->entrypoints
[i
]) {
2144 device
->dispatch
.entrypoints
[i
] = genX_table
->entrypoints
[i
];
2146 device
->dispatch
.entrypoints
[i
] =
2147 anv_device_dispatch_table
.entrypoints
[i
];
2153 vk_priority_to_gen(int priority
)
2156 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT
:
2157 return GEN_CONTEXT_LOW_PRIORITY
;
2158 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
:
2159 return GEN_CONTEXT_MEDIUM_PRIORITY
;
2160 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT
:
2161 return GEN_CONTEXT_HIGH_PRIORITY
;
2162 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT
:
2163 return GEN_CONTEXT_REALTIME_PRIORITY
;
2165 unreachable("Invalid priority");
2170 anv_device_init_hiz_clear_value_bo(struct anv_device
*device
)
2172 anv_bo_init_new(&device
->hiz_clear_bo
, device
, 4096);
2174 if (device
->instance
->physicalDevice
.has_exec_async
)
2175 device
->hiz_clear_bo
.flags
|= EXEC_OBJECT_ASYNC
;
2177 if (device
->instance
->physicalDevice
.use_softpin
)
2178 device
->hiz_clear_bo
.flags
|= EXEC_OBJECT_PINNED
;
2180 anv_vma_alloc(device
, &device
->hiz_clear_bo
);
2182 uint32_t *map
= anv_gem_mmap(device
, device
->hiz_clear_bo
.gem_handle
,
2185 union isl_color_value hiz_clear
= { .u32
= { 0, } };
2186 hiz_clear
.f32
[0] = ANV_HZ_FC_VAL
;
2188 memcpy(map
, hiz_clear
.u32
, sizeof(hiz_clear
.u32
));
2189 anv_gem_munmap(map
, device
->hiz_clear_bo
.size
);
2193 get_bo_from_pool(struct gen_batch_decode_bo
*ret
,
2194 struct anv_block_pool
*pool
,
2197 for (uint32_t i
= 0; i
< pool
->nbos
; i
++) {
2198 uint64_t bo_address
= pool
->bos
[i
].offset
& (~0ull >> 16);
2199 uint32_t bo_size
= pool
->bos
[i
].size
;
2200 if (address
>= bo_address
&& address
< (bo_address
+ bo_size
)) {
2201 *ret
= (struct gen_batch_decode_bo
) {
2204 .map
= pool
->bos
[i
].map
,
2212 /* Finding a buffer for batch decoding */
2213 static struct gen_batch_decode_bo
2214 decode_get_bo(void *v_batch
, bool ppgtt
, uint64_t address
)
2216 struct anv_device
*device
= v_batch
;
2217 struct gen_batch_decode_bo ret_bo
= {};
2221 if (get_bo_from_pool(&ret_bo
, &device
->dynamic_state_pool
.block_pool
, address
))
2223 if (get_bo_from_pool(&ret_bo
, &device
->instruction_state_pool
.block_pool
, address
))
2225 if (get_bo_from_pool(&ret_bo
, &device
->binding_table_pool
.block_pool
, address
))
2227 if (get_bo_from_pool(&ret_bo
, &device
->surface_state_pool
.block_pool
, address
))
2230 if (!device
->cmd_buffer_being_decoded
)
2231 return (struct gen_batch_decode_bo
) { };
2233 struct anv_batch_bo
**bo
;
2235 u_vector_foreach(bo
, &device
->cmd_buffer_being_decoded
->seen_bbos
) {
2236 /* The decoder zeroes out the top 16 bits, so we need to as well */
2237 uint64_t bo_address
= (*bo
)->bo
.offset
& (~0ull >> 16);
2239 if (address
>= bo_address
&& address
< bo_address
+ (*bo
)->bo
.size
) {
2240 return (struct gen_batch_decode_bo
) {
2242 .size
= (*bo
)->bo
.size
,
2243 .map
= (*bo
)->bo
.map
,
2248 return (struct gen_batch_decode_bo
) { };
2251 VkResult
anv_CreateDevice(
2252 VkPhysicalDevice physicalDevice
,
2253 const VkDeviceCreateInfo
* pCreateInfo
,
2254 const VkAllocationCallbacks
* pAllocator
,
2257 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
2259 struct anv_device
*device
;
2261 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO
);
2263 struct anv_device_extension_table enabled_extensions
= { };
2264 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
2266 for (idx
= 0; idx
< ANV_DEVICE_EXTENSION_COUNT
; idx
++) {
2267 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
2268 anv_device_extensions
[idx
].extensionName
) == 0)
2272 if (idx
>= ANV_DEVICE_EXTENSION_COUNT
)
2273 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
2275 if (!physical_device
->supported_extensions
.extensions
[idx
])
2276 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
2278 enabled_extensions
.extensions
[idx
] = true;
2281 /* Check enabled features */
2282 if (pCreateInfo
->pEnabledFeatures
) {
2283 VkPhysicalDeviceFeatures supported_features
;
2284 anv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
2285 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
2286 VkBool32
*enabled_feature
= (VkBool32
*)pCreateInfo
->pEnabledFeatures
;
2287 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
2288 for (uint32_t i
= 0; i
< num_features
; i
++) {
2289 if (enabled_feature
[i
] && !supported_feature
[i
])
2290 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
2294 /* Check requested queues and fail if we are requested to create any
2295 * queues with flags we don't support.
2297 assert(pCreateInfo
->queueCreateInfoCount
> 0);
2298 for (uint32_t i
= 0; i
< pCreateInfo
->queueCreateInfoCount
; i
++) {
2299 if (pCreateInfo
->pQueueCreateInfos
[i
].flags
!= 0)
2300 return vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2303 /* Check if client specified queue priority. */
2304 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*queue_priority
=
2305 vk_find_struct_const(pCreateInfo
->pQueueCreateInfos
[0].pNext
,
2306 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT
);
2308 VkQueueGlobalPriorityEXT priority
=
2309 queue_priority
? queue_priority
->globalPriority
:
2310 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
;
2312 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
2314 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
2316 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2318 if (INTEL_DEBUG
& DEBUG_BATCH
) {
2319 const unsigned decode_flags
=
2320 GEN_BATCH_DECODE_FULL
|
2321 ((INTEL_DEBUG
& DEBUG_COLOR
) ? GEN_BATCH_DECODE_IN_COLOR
: 0) |
2322 GEN_BATCH_DECODE_OFFSETS
|
2323 GEN_BATCH_DECODE_FLOATS
;
2325 gen_batch_decode_ctx_init(&device
->decoder_ctx
,
2326 &physical_device
->info
,
2327 stderr
, decode_flags
, NULL
,
2328 decode_get_bo
, NULL
, device
);
2331 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
2332 device
->instance
= physical_device
->instance
;
2333 device
->chipset_id
= physical_device
->chipset_id
;
2334 device
->no_hw
= physical_device
->no_hw
;
2335 device
->_lost
= false;
2338 device
->alloc
= *pAllocator
;
2340 device
->alloc
= physical_device
->instance
->alloc
;
2342 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
2343 device
->fd
= open(physical_device
->path
, O_RDWR
| O_CLOEXEC
);
2344 if (device
->fd
== -1) {
2345 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2349 device
->context_id
= anv_gem_create_context(device
);
2350 if (device
->context_id
== -1) {
2351 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2355 if (physical_device
->use_softpin
) {
2356 if (pthread_mutex_init(&device
->vma_mutex
, NULL
) != 0) {
2357 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2361 /* keep the page with address zero out of the allocator */
2362 struct anv_memory_heap
*low_heap
=
2363 &physical_device
->memory
.heaps
[physical_device
->memory
.heap_count
- 1];
2364 util_vma_heap_init(&device
->vma_lo
, low_heap
->vma_start
, low_heap
->vma_size
);
2365 device
->vma_lo_available
= low_heap
->size
;
2367 struct anv_memory_heap
*high_heap
=
2368 &physical_device
->memory
.heaps
[0];
2369 util_vma_heap_init(&device
->vma_hi
, high_heap
->vma_start
, high_heap
->vma_size
);
2370 device
->vma_hi_available
= physical_device
->memory
.heap_count
== 1 ? 0 :
2374 list_inithead(&device
->memory_objects
);
2376 /* As per spec, the driver implementation may deny requests to acquire
2377 * a priority above the default priority (MEDIUM) if the caller does not
2378 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
2381 if (physical_device
->has_context_priority
) {
2382 int err
= anv_gem_set_context_param(device
->fd
, device
->context_id
,
2383 I915_CONTEXT_PARAM_PRIORITY
,
2384 vk_priority_to_gen(priority
));
2385 if (err
!= 0 && priority
> VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
) {
2386 result
= vk_error(VK_ERROR_NOT_PERMITTED_EXT
);
2391 device
->info
= physical_device
->info
;
2392 device
->isl_dev
= physical_device
->isl_dev
;
2394 /* On Broadwell and later, we can use batch chaining to more efficiently
2395 * implement growing command buffers. Prior to Haswell, the kernel
2396 * command parser gets in the way and we have to fall back to growing
2399 device
->can_chain_batches
= device
->info
.gen
>= 8;
2401 device
->robust_buffer_access
= pCreateInfo
->pEnabledFeatures
&&
2402 pCreateInfo
->pEnabledFeatures
->robustBufferAccess
;
2403 device
->enabled_extensions
= enabled_extensions
;
2405 anv_device_init_dispatch(device
);
2407 if (pthread_mutex_init(&device
->mutex
, NULL
) != 0) {
2408 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2409 goto fail_context_id
;
2412 pthread_condattr_t condattr
;
2413 if (pthread_condattr_init(&condattr
) != 0) {
2414 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2417 if (pthread_condattr_setclock(&condattr
, CLOCK_MONOTONIC
) != 0) {
2418 pthread_condattr_destroy(&condattr
);
2419 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2422 if (pthread_cond_init(&device
->queue_submit
, &condattr
) != 0) {
2423 pthread_condattr_destroy(&condattr
);
2424 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2427 pthread_condattr_destroy(&condattr
);
2430 (physical_device
->supports_48bit_addresses
? EXEC_OBJECT_SUPPORTS_48B_ADDRESS
: 0) |
2431 (physical_device
->has_exec_async
? EXEC_OBJECT_ASYNC
: 0) |
2432 (physical_device
->has_exec_capture
? EXEC_OBJECT_CAPTURE
: 0) |
2433 (physical_device
->use_softpin
? EXEC_OBJECT_PINNED
: 0);
2435 anv_bo_pool_init(&device
->batch_bo_pool
, device
, bo_flags
);
2437 result
= anv_bo_cache_init(&device
->bo_cache
);
2438 if (result
!= VK_SUCCESS
)
2439 goto fail_batch_bo_pool
;
2441 if (!physical_device
->use_softpin
)
2442 bo_flags
&= ~EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
2444 result
= anv_state_pool_init(&device
->dynamic_state_pool
, device
,
2445 DYNAMIC_STATE_POOL_MIN_ADDRESS
,
2448 if (result
!= VK_SUCCESS
)
2451 result
= anv_state_pool_init(&device
->instruction_state_pool
, device
,
2452 INSTRUCTION_STATE_POOL_MIN_ADDRESS
,
2455 if (result
!= VK_SUCCESS
)
2456 goto fail_dynamic_state_pool
;
2458 result
= anv_state_pool_init(&device
->surface_state_pool
, device
,
2459 SURFACE_STATE_POOL_MIN_ADDRESS
,
2462 if (result
!= VK_SUCCESS
)
2463 goto fail_instruction_state_pool
;
2465 if (physical_device
->use_softpin
) {
2466 result
= anv_state_pool_init(&device
->binding_table_pool
, device
,
2467 BINDING_TABLE_POOL_MIN_ADDRESS
,
2470 if (result
!= VK_SUCCESS
)
2471 goto fail_surface_state_pool
;
2474 result
= anv_bo_init_new(&device
->workaround_bo
, device
, 4096);
2475 if (result
!= VK_SUCCESS
)
2476 goto fail_binding_table_pool
;
2478 if (physical_device
->use_softpin
)
2479 device
->workaround_bo
.flags
|= EXEC_OBJECT_PINNED
;
2481 if (!anv_vma_alloc(device
, &device
->workaround_bo
))
2482 goto fail_workaround_bo
;
2484 anv_device_init_trivial_batch(device
);
2486 if (device
->info
.gen
>= 10)
2487 anv_device_init_hiz_clear_value_bo(device
);
2489 anv_scratch_pool_init(device
, &device
->scratch_pool
);
2491 anv_queue_init(device
, &device
->queue
);
2493 switch (device
->info
.gen
) {
2495 if (!device
->info
.is_haswell
)
2496 result
= gen7_init_device_state(device
);
2498 result
= gen75_init_device_state(device
);
2501 result
= gen8_init_device_state(device
);
2504 result
= gen9_init_device_state(device
);
2507 result
= gen10_init_device_state(device
);
2510 result
= gen11_init_device_state(device
);
2513 /* Shouldn't get here as we don't create physical devices for any other
2515 unreachable("unhandled gen");
2517 if (result
!= VK_SUCCESS
)
2518 goto fail_workaround_bo
;
2520 anv_pipeline_cache_init(&device
->default_pipeline_cache
, device
, true);
2522 anv_device_init_blorp(device
);
2524 anv_device_init_border_colors(device
);
2526 *pDevice
= anv_device_to_handle(device
);
2531 anv_queue_finish(&device
->queue
);
2532 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
2533 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
2534 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
2535 fail_binding_table_pool
:
2536 if (physical_device
->use_softpin
)
2537 anv_state_pool_finish(&device
->binding_table_pool
);
2538 fail_surface_state_pool
:
2539 anv_state_pool_finish(&device
->surface_state_pool
);
2540 fail_instruction_state_pool
:
2541 anv_state_pool_finish(&device
->instruction_state_pool
);
2542 fail_dynamic_state_pool
:
2543 anv_state_pool_finish(&device
->dynamic_state_pool
);
2545 anv_bo_cache_finish(&device
->bo_cache
);
2547 anv_bo_pool_finish(&device
->batch_bo_pool
);
2548 pthread_cond_destroy(&device
->queue_submit
);
2550 pthread_mutex_destroy(&device
->mutex
);
2552 anv_gem_destroy_context(device
, device
->context_id
);
2556 vk_free(&device
->alloc
, device
);
2561 void anv_DestroyDevice(
2563 const VkAllocationCallbacks
* pAllocator
)
2565 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2566 struct anv_physical_device
*physical_device
;
2571 physical_device
= &device
->instance
->physicalDevice
;
2573 anv_device_finish_blorp(device
);
2575 anv_pipeline_cache_finish(&device
->default_pipeline_cache
);
2577 anv_queue_finish(&device
->queue
);
2579 #ifdef HAVE_VALGRIND
2580 /* We only need to free these to prevent valgrind errors. The backing
2581 * BO will go away in a couple of lines so we don't actually leak.
2583 anv_state_pool_free(&device
->dynamic_state_pool
, device
->border_colors
);
2586 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
2588 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
2589 anv_vma_free(device
, &device
->workaround_bo
);
2590 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
2592 anv_vma_free(device
, &device
->trivial_batch_bo
);
2593 anv_gem_close(device
, device
->trivial_batch_bo
.gem_handle
);
2594 if (device
->info
.gen
>= 10)
2595 anv_gem_close(device
, device
->hiz_clear_bo
.gem_handle
);
2597 if (physical_device
->use_softpin
)
2598 anv_state_pool_finish(&device
->binding_table_pool
);
2599 anv_state_pool_finish(&device
->surface_state_pool
);
2600 anv_state_pool_finish(&device
->instruction_state_pool
);
2601 anv_state_pool_finish(&device
->dynamic_state_pool
);
2603 anv_bo_cache_finish(&device
->bo_cache
);
2605 anv_bo_pool_finish(&device
->batch_bo_pool
);
2607 pthread_cond_destroy(&device
->queue_submit
);
2608 pthread_mutex_destroy(&device
->mutex
);
2610 anv_gem_destroy_context(device
, device
->context_id
);
2612 if (INTEL_DEBUG
& DEBUG_BATCH
)
2613 gen_batch_decode_ctx_finish(&device
->decoder_ctx
);
2617 vk_free(&device
->alloc
, device
);
2620 VkResult
anv_EnumerateInstanceLayerProperties(
2621 uint32_t* pPropertyCount
,
2622 VkLayerProperties
* pProperties
)
2624 if (pProperties
== NULL
) {
2625 *pPropertyCount
= 0;
2629 /* None supported at this time */
2630 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
2633 VkResult
anv_EnumerateDeviceLayerProperties(
2634 VkPhysicalDevice physicalDevice
,
2635 uint32_t* pPropertyCount
,
2636 VkLayerProperties
* pProperties
)
2638 if (pProperties
== NULL
) {
2639 *pPropertyCount
= 0;
2643 /* None supported at this time */
2644 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
2647 void anv_GetDeviceQueue(
2649 uint32_t queueNodeIndex
,
2650 uint32_t queueIndex
,
2653 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2655 assert(queueIndex
== 0);
2657 *pQueue
= anv_queue_to_handle(&device
->queue
);
2660 void anv_GetDeviceQueue2(
2662 const VkDeviceQueueInfo2
* pQueueInfo
,
2665 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2667 assert(pQueueInfo
->queueIndex
== 0);
2669 if (pQueueInfo
->flags
== device
->queue
.flags
)
2670 *pQueue
= anv_queue_to_handle(&device
->queue
);
2676 _anv_device_set_lost(struct anv_device
*device
,
2677 const char *file
, int line
,
2678 const char *msg
, ...)
2683 device
->_lost
= true;
2686 err
= __vk_errorv(device
->instance
, device
,
2687 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT
,
2688 VK_ERROR_DEVICE_LOST
, file
, line
, msg
, ap
);
2691 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
2698 anv_device_query_status(struct anv_device
*device
)
2700 /* This isn't likely as most of the callers of this function already check
2701 * for it. However, it doesn't hurt to check and it potentially lets us
2704 if (anv_device_is_lost(device
))
2705 return VK_ERROR_DEVICE_LOST
;
2707 uint32_t active
, pending
;
2708 int ret
= anv_gem_gpu_get_reset_stats(device
, &active
, &pending
);
2710 /* We don't know the real error. */
2711 return anv_device_set_lost(device
, "get_reset_stats failed: %m");
2715 return anv_device_set_lost(device
, "GPU hung on one of our command buffers");
2716 } else if (pending
) {
2717 return anv_device_set_lost(device
, "GPU hung with commands in-flight");
2724 anv_device_bo_busy(struct anv_device
*device
, struct anv_bo
*bo
)
2726 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
2727 * Other usages of the BO (such as on different hardware) will not be
2728 * flagged as "busy" by this ioctl. Use with care.
2730 int ret
= anv_gem_busy(device
, bo
->gem_handle
);
2732 return VK_NOT_READY
;
2733 } else if (ret
== -1) {
2734 /* We don't know the real error. */
2735 return anv_device_set_lost(device
, "gem wait failed: %m");
2738 /* Query for device status after the busy call. If the BO we're checking
2739 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
2740 * client because it clearly doesn't have valid data. Yes, this most
2741 * likely means an ioctl, but we just did an ioctl to query the busy status
2742 * so it's no great loss.
2744 return anv_device_query_status(device
);
2748 anv_device_wait(struct anv_device
*device
, struct anv_bo
*bo
,
2751 int ret
= anv_gem_wait(device
, bo
->gem_handle
, &timeout
);
2752 if (ret
== -1 && errno
== ETIME
) {
2754 } else if (ret
== -1) {
2755 /* We don't know the real error. */
2756 return anv_device_set_lost(device
, "gem wait failed: %m");
2759 /* Query for device status after the wait. If the BO we're waiting on got
2760 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
2761 * because it clearly doesn't have valid data. Yes, this most likely means
2762 * an ioctl, but we just did an ioctl to wait so it's no great loss.
2764 return anv_device_query_status(device
);
2767 VkResult
anv_DeviceWaitIdle(
2770 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2771 if (anv_device_is_lost(device
))
2772 return VK_ERROR_DEVICE_LOST
;
2774 struct anv_batch batch
;
2777 batch
.start
= batch
.next
= cmds
;
2778 batch
.end
= (void *) cmds
+ sizeof(cmds
);
2780 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
2781 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
2783 return anv_device_submit_simple_batch(device
, &batch
);
2787 anv_vma_alloc(struct anv_device
*device
, struct anv_bo
*bo
)
2789 if (!(bo
->flags
& EXEC_OBJECT_PINNED
))
2792 pthread_mutex_lock(&device
->vma_mutex
);
2796 if (bo
->flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
&&
2797 device
->vma_hi_available
>= bo
->size
) {
2798 uint64_t addr
= util_vma_heap_alloc(&device
->vma_hi
, bo
->size
, 4096);
2800 bo
->offset
= gen_canonical_address(addr
);
2801 assert(addr
== gen_48b_address(bo
->offset
));
2802 device
->vma_hi_available
-= bo
->size
;
2806 if (bo
->offset
== 0 && device
->vma_lo_available
>= bo
->size
) {
2807 uint64_t addr
= util_vma_heap_alloc(&device
->vma_lo
, bo
->size
, 4096);
2809 bo
->offset
= gen_canonical_address(addr
);
2810 assert(addr
== gen_48b_address(bo
->offset
));
2811 device
->vma_lo_available
-= bo
->size
;
2815 pthread_mutex_unlock(&device
->vma_mutex
);
2817 return bo
->offset
!= 0;
2821 anv_vma_free(struct anv_device
*device
, struct anv_bo
*bo
)
2823 if (!(bo
->flags
& EXEC_OBJECT_PINNED
))
2826 const uint64_t addr_48b
= gen_48b_address(bo
->offset
);
2828 pthread_mutex_lock(&device
->vma_mutex
);
2830 if (addr_48b
>= LOW_HEAP_MIN_ADDRESS
&&
2831 addr_48b
<= LOW_HEAP_MAX_ADDRESS
) {
2832 util_vma_heap_free(&device
->vma_lo
, addr_48b
, bo
->size
);
2833 device
->vma_lo_available
+= bo
->size
;
2835 ASSERTED
const struct anv_physical_device
*physical_device
=
2836 &device
->instance
->physicalDevice
;
2837 assert(addr_48b
>= physical_device
->memory
.heaps
[0].vma_start
&&
2838 addr_48b
< (physical_device
->memory
.heaps
[0].vma_start
+
2839 physical_device
->memory
.heaps
[0].vma_size
));
2840 util_vma_heap_free(&device
->vma_hi
, addr_48b
, bo
->size
);
2841 device
->vma_hi_available
+= bo
->size
;
2844 pthread_mutex_unlock(&device
->vma_mutex
);
2850 anv_bo_init_new(struct anv_bo
*bo
, struct anv_device
*device
, uint64_t size
)
2852 uint32_t gem_handle
= anv_gem_create(device
, size
);
2854 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
2856 anv_bo_init(bo
, gem_handle
, size
);
2861 VkResult
anv_AllocateMemory(
2863 const VkMemoryAllocateInfo
* pAllocateInfo
,
2864 const VkAllocationCallbacks
* pAllocator
,
2865 VkDeviceMemory
* pMem
)
2867 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2868 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2869 struct anv_device_memory
*mem
;
2870 VkResult result
= VK_SUCCESS
;
2872 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
2874 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
2875 assert(pAllocateInfo
->allocationSize
> 0);
2877 if (pAllocateInfo
->allocationSize
> MAX_MEMORY_ALLOCATION_SIZE
)
2878 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
2880 /* FINISHME: Fail if allocation request exceeds heap size. */
2882 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
2883 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2885 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2887 assert(pAllocateInfo
->memoryTypeIndex
< pdevice
->memory
.type_count
);
2888 mem
->type
= &pdevice
->memory
.types
[pAllocateInfo
->memoryTypeIndex
];
2892 mem
->host_ptr
= NULL
;
2894 uint64_t bo_flags
= 0;
2896 assert(mem
->type
->heapIndex
< pdevice
->memory
.heap_count
);
2897 if (pdevice
->memory
.heaps
[mem
->type
->heapIndex
].supports_48bit_addresses
)
2898 bo_flags
|= EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
2900 const struct wsi_memory_allocate_info
*wsi_info
=
2901 vk_find_struct_const(pAllocateInfo
->pNext
, WSI_MEMORY_ALLOCATE_INFO_MESA
);
2902 if (wsi_info
&& wsi_info
->implicit_sync
) {
2903 /* We need to set the WRITE flag on window system buffers so that GEM
2904 * will know we're writing to them and synchronize uses on other rings
2905 * (eg if the display server uses the blitter ring).
2907 bo_flags
|= EXEC_OBJECT_WRITE
;
2908 } else if (pdevice
->has_exec_async
) {
2909 bo_flags
|= EXEC_OBJECT_ASYNC
;
2912 if (pdevice
->use_softpin
)
2913 bo_flags
|= EXEC_OBJECT_PINNED
;
2915 const VkExportMemoryAllocateInfo
*export_info
=
2916 vk_find_struct_const(pAllocateInfo
->pNext
, EXPORT_MEMORY_ALLOCATE_INFO
);
2918 /* Check if we need to support Android HW buffer export. If so,
2919 * create AHardwareBuffer and import memory from it.
2921 bool android_export
= false;
2922 if (export_info
&& export_info
->handleTypes
&
2923 VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID
)
2924 android_export
= true;
2926 /* Android memory import. */
2927 const struct VkImportAndroidHardwareBufferInfoANDROID
*ahw_import_info
=
2928 vk_find_struct_const(pAllocateInfo
->pNext
,
2929 IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID
);
2931 if (ahw_import_info
) {
2932 result
= anv_import_ahw_memory(_device
, mem
, ahw_import_info
);
2933 if (result
!= VK_SUCCESS
)
2937 } else if (android_export
) {
2938 result
= anv_create_ahw_memory(_device
, mem
, pAllocateInfo
);
2939 if (result
!= VK_SUCCESS
)
2942 const struct VkImportAndroidHardwareBufferInfoANDROID import_info
= {
2945 result
= anv_import_ahw_memory(_device
, mem
, &import_info
);
2946 if (result
!= VK_SUCCESS
)
2952 const VkImportMemoryFdInfoKHR
*fd_info
=
2953 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_FD_INFO_KHR
);
2955 /* The Vulkan spec permits handleType to be 0, in which case the struct is
2958 if (fd_info
&& fd_info
->handleType
) {
2959 /* At the moment, we support only the below handle types. */
2960 assert(fd_info
->handleType
==
2961 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
2962 fd_info
->handleType
==
2963 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
2965 result
= anv_bo_cache_import(device
, &device
->bo_cache
, fd_info
->fd
,
2966 bo_flags
| ANV_BO_EXTERNAL
, &mem
->bo
);
2967 if (result
!= VK_SUCCESS
)
2970 VkDeviceSize aligned_alloc_size
=
2971 align_u64(pAllocateInfo
->allocationSize
, 4096);
2973 /* For security purposes, we reject importing the bo if it's smaller
2974 * than the requested allocation size. This prevents a malicious client
2975 * from passing a buffer to a trusted client, lying about the size, and
2976 * telling the trusted client to try and texture from an image that goes
2977 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
2978 * in the trusted client. The trusted client can protect itself against
2979 * this sort of attack but only if it can trust the buffer size.
2981 if (mem
->bo
->size
< aligned_alloc_size
) {
2982 result
= vk_errorf(device
->instance
, device
,
2983 VK_ERROR_INVALID_EXTERNAL_HANDLE
,
2984 "aligned allocationSize too large for "
2985 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT: "
2986 "%"PRIu64
"B > %"PRIu64
"B",
2987 aligned_alloc_size
, mem
->bo
->size
);
2988 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
2992 /* From the Vulkan spec:
2994 * "Importing memory from a file descriptor transfers ownership of
2995 * the file descriptor from the application to the Vulkan
2996 * implementation. The application must not perform any operations on
2997 * the file descriptor after a successful import."
2999 * If the import fails, we leave the file descriptor open.
3005 const VkImportMemoryHostPointerInfoEXT
*host_ptr_info
=
3006 vk_find_struct_const(pAllocateInfo
->pNext
,
3007 IMPORT_MEMORY_HOST_POINTER_INFO_EXT
);
3008 if (host_ptr_info
&& host_ptr_info
->handleType
) {
3009 if (host_ptr_info
->handleType
==
3010 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_MAPPED_FOREIGN_MEMORY_BIT_EXT
) {
3011 result
= vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
3015 assert(host_ptr_info
->handleType
==
3016 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
);
3018 result
= anv_bo_cache_import_host_ptr(
3019 device
, &device
->bo_cache
, host_ptr_info
->pHostPointer
,
3020 pAllocateInfo
->allocationSize
, bo_flags
, &mem
->bo
);
3022 if (result
!= VK_SUCCESS
)
3025 mem
->host_ptr
= host_ptr_info
->pHostPointer
;
3029 /* Regular allocate (not importing memory). */
3031 if (export_info
&& export_info
->handleTypes
)
3032 bo_flags
|= ANV_BO_EXTERNAL
;
3034 result
= anv_bo_cache_alloc(device
, &device
->bo_cache
,
3035 pAllocateInfo
->allocationSize
, bo_flags
,
3037 if (result
!= VK_SUCCESS
)
3040 const VkMemoryDedicatedAllocateInfo
*dedicated_info
=
3041 vk_find_struct_const(pAllocateInfo
->pNext
, MEMORY_DEDICATED_ALLOCATE_INFO
);
3042 if (dedicated_info
&& dedicated_info
->image
!= VK_NULL_HANDLE
) {
3043 ANV_FROM_HANDLE(anv_image
, image
, dedicated_info
->image
);
3045 /* Some legacy (non-modifiers) consumers need the tiling to be set on
3046 * the BO. In this case, we have a dedicated allocation.
3048 if (image
->needs_set_tiling
) {
3049 const uint32_t i915_tiling
=
3050 isl_tiling_to_i915_tiling(image
->planes
[0].surface
.isl
.tiling
);
3051 int ret
= anv_gem_set_tiling(device
, mem
->bo
->gem_handle
,
3052 image
->planes
[0].surface
.isl
.row_pitch_B
,
3055 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
3056 return vk_errorf(device
->instance
, NULL
,
3057 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
3058 "failed to set BO tiling: %m");
3064 pthread_mutex_lock(&device
->mutex
);
3065 list_addtail(&mem
->link
, &device
->memory_objects
);
3066 pthread_mutex_unlock(&device
->mutex
);
3068 *pMem
= anv_device_memory_to_handle(mem
);
3070 p_atomic_add(&pdevice
->memory
.heaps
[mem
->type
->heapIndex
].used
,
3076 vk_free2(&device
->alloc
, pAllocator
, mem
);
3081 VkResult
anv_GetMemoryFdKHR(
3083 const VkMemoryGetFdInfoKHR
* pGetFdInfo
,
3086 ANV_FROM_HANDLE(anv_device
, dev
, device_h
);
3087 ANV_FROM_HANDLE(anv_device_memory
, mem
, pGetFdInfo
->memory
);
3089 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
3091 assert(pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
3092 pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
3094 return anv_bo_cache_export(dev
, &dev
->bo_cache
, mem
->bo
, pFd
);
3097 VkResult
anv_GetMemoryFdPropertiesKHR(
3099 VkExternalMemoryHandleTypeFlagBits handleType
,
3101 VkMemoryFdPropertiesKHR
* pMemoryFdProperties
)
3103 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3104 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3106 switch (handleType
) {
3107 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
:
3108 /* dma-buf can be imported as any memory type */
3109 pMemoryFdProperties
->memoryTypeBits
=
3110 (1 << pdevice
->memory
.type_count
) - 1;
3114 /* The valid usage section for this function says:
3116 * "handleType must not be one of the handle types defined as
3119 * So opaque handle types fall into the default "unsupported" case.
3121 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
3125 VkResult
anv_GetMemoryHostPointerPropertiesEXT(
3127 VkExternalMemoryHandleTypeFlagBits handleType
,
3128 const void* pHostPointer
,
3129 VkMemoryHostPointerPropertiesEXT
* pMemoryHostPointerProperties
)
3131 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3133 assert(pMemoryHostPointerProperties
->sType
==
3134 VK_STRUCTURE_TYPE_MEMORY_HOST_POINTER_PROPERTIES_EXT
);
3136 switch (handleType
) {
3137 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
: {
3138 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3140 /* Host memory can be imported as any memory type. */
3141 pMemoryHostPointerProperties
->memoryTypeBits
=
3142 (1ull << pdevice
->memory
.type_count
) - 1;
3147 return VK_ERROR_INVALID_EXTERNAL_HANDLE
;
3151 void anv_FreeMemory(
3153 VkDeviceMemory _mem
,
3154 const VkAllocationCallbacks
* pAllocator
)
3156 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3157 ANV_FROM_HANDLE(anv_device_memory
, mem
, _mem
);
3158 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3163 pthread_mutex_lock(&device
->mutex
);
3164 list_del(&mem
->link
);
3165 pthread_mutex_unlock(&device
->mutex
);
3168 anv_UnmapMemory(_device
, _mem
);
3170 p_atomic_add(&pdevice
->memory
.heaps
[mem
->type
->heapIndex
].used
,
3173 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
3175 #if defined(ANDROID) && ANDROID_API_LEVEL >= 26
3177 AHardwareBuffer_release(mem
->ahw
);
3180 vk_free2(&device
->alloc
, pAllocator
, mem
);
3183 VkResult
anv_MapMemory(
3185 VkDeviceMemory _memory
,
3186 VkDeviceSize offset
,
3188 VkMemoryMapFlags flags
,
3191 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3192 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
3199 if (mem
->host_ptr
) {
3200 *ppData
= mem
->host_ptr
+ offset
;
3204 if (size
== VK_WHOLE_SIZE
)
3205 size
= mem
->bo
->size
- offset
;
3207 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
3209 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
3210 * assert(size != 0);
3211 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
3212 * equal to the size of the memory minus offset
3215 assert(offset
+ size
<= mem
->bo
->size
);
3217 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
3218 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
3219 * at a time is valid. We could just mmap up front and return an offset
3220 * pointer here, but that may exhaust virtual memory on 32 bit
3223 uint32_t gem_flags
= 0;
3225 if (!device
->info
.has_llc
&&
3226 (mem
->type
->propertyFlags
& VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
))
3227 gem_flags
|= I915_MMAP_WC
;
3229 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
3230 uint64_t map_offset
= offset
& ~4095ull;
3231 assert(offset
>= map_offset
);
3232 uint64_t map_size
= (offset
+ size
) - map_offset
;
3234 /* Let's map whole pages */
3235 map_size
= align_u64(map_size
, 4096);
3237 void *map
= anv_gem_mmap(device
, mem
->bo
->gem_handle
,
3238 map_offset
, map_size
, gem_flags
);
3239 if (map
== MAP_FAILED
)
3240 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
3243 mem
->map_size
= map_size
;
3245 *ppData
= mem
->map
+ (offset
- map_offset
);
3250 void anv_UnmapMemory(
3252 VkDeviceMemory _memory
)
3254 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
3256 if (mem
== NULL
|| mem
->host_ptr
)
3259 anv_gem_munmap(mem
->map
, mem
->map_size
);
3266 clflush_mapped_ranges(struct anv_device
*device
,
3268 const VkMappedMemoryRange
*ranges
)
3270 for (uint32_t i
= 0; i
< count
; i
++) {
3271 ANV_FROM_HANDLE(anv_device_memory
, mem
, ranges
[i
].memory
);
3272 if (ranges
[i
].offset
>= mem
->map_size
)
3275 gen_clflush_range(mem
->map
+ ranges
[i
].offset
,
3276 MIN2(ranges
[i
].size
, mem
->map_size
- ranges
[i
].offset
));
3280 VkResult
anv_FlushMappedMemoryRanges(
3282 uint32_t memoryRangeCount
,
3283 const VkMappedMemoryRange
* pMemoryRanges
)
3285 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3287 if (device
->info
.has_llc
)
3290 /* Make sure the writes we're flushing have landed. */
3291 __builtin_ia32_mfence();
3293 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
3298 VkResult
anv_InvalidateMappedMemoryRanges(
3300 uint32_t memoryRangeCount
,
3301 const VkMappedMemoryRange
* pMemoryRanges
)
3303 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3305 if (device
->info
.has_llc
)
3308 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
3310 /* Make sure no reads get moved up above the invalidate. */
3311 __builtin_ia32_mfence();
3316 void anv_GetBufferMemoryRequirements(
3319 VkMemoryRequirements
* pMemoryRequirements
)
3321 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
3322 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3323 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3325 /* The Vulkan spec (git aaed022) says:
3327 * memoryTypeBits is a bitfield and contains one bit set for every
3328 * supported memory type for the resource. The bit `1<<i` is set if and
3329 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3330 * structure for the physical device is supported.
3332 uint32_t memory_types
= 0;
3333 for (uint32_t i
= 0; i
< pdevice
->memory
.type_count
; i
++) {
3334 uint32_t valid_usage
= pdevice
->memory
.types
[i
].valid_buffer_usage
;
3335 if ((valid_usage
& buffer
->usage
) == buffer
->usage
)
3336 memory_types
|= (1u << i
);
3339 /* Base alignment requirement of a cache line */
3340 uint32_t alignment
= 16;
3342 /* We need an alignment of 32 for pushing UBOs */
3343 if (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
)
3344 alignment
= MAX2(alignment
, 32);
3346 pMemoryRequirements
->size
= buffer
->size
;
3347 pMemoryRequirements
->alignment
= alignment
;
3349 /* Storage and Uniform buffers should have their size aligned to
3350 * 32-bits to avoid boundary checks when last DWord is not complete.
3351 * This would ensure that not internal padding would be needed for
3354 if (device
->robust_buffer_access
&&
3355 (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
||
3356 buffer
->usage
& VK_BUFFER_USAGE_STORAGE_BUFFER_BIT
))
3357 pMemoryRequirements
->size
= align_u64(buffer
->size
, 4);
3359 pMemoryRequirements
->memoryTypeBits
= memory_types
;
3362 void anv_GetBufferMemoryRequirements2(
3364 const VkBufferMemoryRequirementsInfo2
* pInfo
,
3365 VkMemoryRequirements2
* pMemoryRequirements
)
3367 anv_GetBufferMemoryRequirements(_device
, pInfo
->buffer
,
3368 &pMemoryRequirements
->memoryRequirements
);
3370 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3371 switch (ext
->sType
) {
3372 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
3373 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
3374 requirements
->prefersDedicatedAllocation
= false;
3375 requirements
->requiresDedicatedAllocation
= false;
3380 anv_debug_ignored_stype(ext
->sType
);
3386 void anv_GetImageMemoryRequirements(
3389 VkMemoryRequirements
* pMemoryRequirements
)
3391 ANV_FROM_HANDLE(anv_image
, image
, _image
);
3392 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3393 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3395 /* The Vulkan spec (git aaed022) says:
3397 * memoryTypeBits is a bitfield and contains one bit set for every
3398 * supported memory type for the resource. The bit `1<<i` is set if and
3399 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3400 * structure for the physical device is supported.
3402 * All types are currently supported for images.
3404 uint32_t memory_types
= (1ull << pdevice
->memory
.type_count
) - 1;
3406 /* We must have image allocated or imported at this point. According to the
3407 * specification, external images must have been bound to memory before
3408 * calling GetImageMemoryRequirements.
3410 assert(image
->size
> 0);
3412 pMemoryRequirements
->size
= image
->size
;
3413 pMemoryRequirements
->alignment
= image
->alignment
;
3414 pMemoryRequirements
->memoryTypeBits
= memory_types
;
3417 void anv_GetImageMemoryRequirements2(
3419 const VkImageMemoryRequirementsInfo2
* pInfo
,
3420 VkMemoryRequirements2
* pMemoryRequirements
)
3422 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3423 ANV_FROM_HANDLE(anv_image
, image
, pInfo
->image
);
3425 anv_GetImageMemoryRequirements(_device
, pInfo
->image
,
3426 &pMemoryRequirements
->memoryRequirements
);
3428 vk_foreach_struct_const(ext
, pInfo
->pNext
) {
3429 switch (ext
->sType
) {
3430 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO
: {
3431 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3432 const VkImagePlaneMemoryRequirementsInfo
*plane_reqs
=
3433 (const VkImagePlaneMemoryRequirementsInfo
*) ext
;
3434 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
,
3435 plane_reqs
->planeAspect
);
3437 assert(image
->planes
[plane
].offset
== 0);
3439 /* The Vulkan spec (git aaed022) says:
3441 * memoryTypeBits is a bitfield and contains one bit set for every
3442 * supported memory type for the resource. The bit `1<<i` is set
3443 * if and only if the memory type `i` in the
3444 * VkPhysicalDeviceMemoryProperties structure for the physical
3445 * device is supported.
3447 * All types are currently supported for images.
3449 pMemoryRequirements
->memoryRequirements
.memoryTypeBits
=
3450 (1ull << pdevice
->memory
.type_count
) - 1;
3452 /* We must have image allocated or imported at this point. According to the
3453 * specification, external images must have been bound to memory before
3454 * calling GetImageMemoryRequirements.
3456 assert(image
->planes
[plane
].size
> 0);
3458 pMemoryRequirements
->memoryRequirements
.size
= image
->planes
[plane
].size
;
3459 pMemoryRequirements
->memoryRequirements
.alignment
=
3460 image
->planes
[plane
].alignment
;
3465 anv_debug_ignored_stype(ext
->sType
);
3470 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3471 switch (ext
->sType
) {
3472 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
3473 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
3474 if (image
->needs_set_tiling
|| image
->external_format
) {
3475 /* If we need to set the tiling for external consumers, we need a
3476 * dedicated allocation.
3478 * See also anv_AllocateMemory.
3480 requirements
->prefersDedicatedAllocation
= true;
3481 requirements
->requiresDedicatedAllocation
= true;
3483 requirements
->prefersDedicatedAllocation
= false;
3484 requirements
->requiresDedicatedAllocation
= false;
3490 anv_debug_ignored_stype(ext
->sType
);
3496 void anv_GetImageSparseMemoryRequirements(
3499 uint32_t* pSparseMemoryRequirementCount
,
3500 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
3502 *pSparseMemoryRequirementCount
= 0;
3505 void anv_GetImageSparseMemoryRequirements2(
3507 const VkImageSparseMemoryRequirementsInfo2
* pInfo
,
3508 uint32_t* pSparseMemoryRequirementCount
,
3509 VkSparseImageMemoryRequirements2
* pSparseMemoryRequirements
)
3511 *pSparseMemoryRequirementCount
= 0;
3514 void anv_GetDeviceMemoryCommitment(
3516 VkDeviceMemory memory
,
3517 VkDeviceSize
* pCommittedMemoryInBytes
)
3519 *pCommittedMemoryInBytes
= 0;
3523 anv_bind_buffer_memory(const VkBindBufferMemoryInfo
*pBindInfo
)
3525 ANV_FROM_HANDLE(anv_device_memory
, mem
, pBindInfo
->memory
);
3526 ANV_FROM_HANDLE(anv_buffer
, buffer
, pBindInfo
->buffer
);
3528 assert(pBindInfo
->sType
== VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
);
3531 assert((buffer
->usage
& mem
->type
->valid_buffer_usage
) == buffer
->usage
);
3532 buffer
->address
= (struct anv_address
) {
3534 .offset
= pBindInfo
->memoryOffset
,
3537 buffer
->address
= ANV_NULL_ADDRESS
;
3541 VkResult
anv_BindBufferMemory(
3544 VkDeviceMemory memory
,
3545 VkDeviceSize memoryOffset
)
3547 anv_bind_buffer_memory(
3548 &(VkBindBufferMemoryInfo
) {
3549 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
,
3552 .memoryOffset
= memoryOffset
,
3558 VkResult
anv_BindBufferMemory2(
3560 uint32_t bindInfoCount
,
3561 const VkBindBufferMemoryInfo
* pBindInfos
)
3563 for (uint32_t i
= 0; i
< bindInfoCount
; i
++)
3564 anv_bind_buffer_memory(&pBindInfos
[i
]);
3569 VkResult
anv_QueueBindSparse(
3571 uint32_t bindInfoCount
,
3572 const VkBindSparseInfo
* pBindInfo
,
3575 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
3576 if (anv_device_is_lost(queue
->device
))
3577 return VK_ERROR_DEVICE_LOST
;
3579 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
3584 VkResult
anv_CreateEvent(
3586 const VkEventCreateInfo
* pCreateInfo
,
3587 const VkAllocationCallbacks
* pAllocator
,
3590 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3591 struct anv_state state
;
3592 struct anv_event
*event
;
3594 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_EVENT_CREATE_INFO
);
3596 state
= anv_state_pool_alloc(&device
->dynamic_state_pool
,
3599 event
->state
= state
;
3600 event
->semaphore
= VK_EVENT_RESET
;
3602 if (!device
->info
.has_llc
) {
3603 /* Make sure the writes we're flushing have landed. */
3604 __builtin_ia32_mfence();
3605 __builtin_ia32_clflush(event
);
3608 *pEvent
= anv_event_to_handle(event
);
3613 void anv_DestroyEvent(
3616 const VkAllocationCallbacks
* pAllocator
)
3618 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3619 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3624 anv_state_pool_free(&device
->dynamic_state_pool
, event
->state
);
3627 VkResult
anv_GetEventStatus(
3631 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3632 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3634 if (anv_device_is_lost(device
))
3635 return VK_ERROR_DEVICE_LOST
;
3637 if (!device
->info
.has_llc
) {
3638 /* Invalidate read cache before reading event written by GPU. */
3639 __builtin_ia32_clflush(event
);
3640 __builtin_ia32_mfence();
3644 return event
->semaphore
;
3647 VkResult
anv_SetEvent(
3651 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3652 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3654 event
->semaphore
= VK_EVENT_SET
;
3656 if (!device
->info
.has_llc
) {
3657 /* Make sure the writes we're flushing have landed. */
3658 __builtin_ia32_mfence();
3659 __builtin_ia32_clflush(event
);
3665 VkResult
anv_ResetEvent(
3669 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3670 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3672 event
->semaphore
= VK_EVENT_RESET
;
3674 if (!device
->info
.has_llc
) {
3675 /* Make sure the writes we're flushing have landed. */
3676 __builtin_ia32_mfence();
3677 __builtin_ia32_clflush(event
);
3685 VkResult
anv_CreateBuffer(
3687 const VkBufferCreateInfo
* pCreateInfo
,
3688 const VkAllocationCallbacks
* pAllocator
,
3691 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3692 struct anv_buffer
*buffer
;
3694 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
3696 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
3697 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3699 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3701 buffer
->size
= pCreateInfo
->size
;
3702 buffer
->usage
= pCreateInfo
->usage
;
3703 buffer
->address
= ANV_NULL_ADDRESS
;
3705 *pBuffer
= anv_buffer_to_handle(buffer
);
3710 void anv_DestroyBuffer(
3713 const VkAllocationCallbacks
* pAllocator
)
3715 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3716 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
3721 vk_free2(&device
->alloc
, pAllocator
, buffer
);
3724 VkDeviceAddress
anv_GetBufferDeviceAddressEXT(
3726 const VkBufferDeviceAddressInfoEXT
* pInfo
)
3728 ANV_FROM_HANDLE(anv_buffer
, buffer
, pInfo
->buffer
);
3730 assert(buffer
->address
.bo
->flags
& EXEC_OBJECT_PINNED
);
3732 return anv_address_physical(buffer
->address
);
3736 anv_fill_buffer_surface_state(struct anv_device
*device
, struct anv_state state
,
3737 enum isl_format format
,
3738 struct anv_address address
,
3739 uint32_t range
, uint32_t stride
)
3741 isl_buffer_fill_state(&device
->isl_dev
, state
.map
,
3742 .address
= anv_address_physical(address
),
3743 .mocs
= device
->default_mocs
,
3746 .swizzle
= ISL_SWIZZLE_IDENTITY
,
3747 .stride_B
= stride
);
3750 void anv_DestroySampler(
3753 const VkAllocationCallbacks
* pAllocator
)
3755 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3756 ANV_FROM_HANDLE(anv_sampler
, sampler
, _sampler
);
3761 if (sampler
->bindless_state
.map
) {
3762 anv_state_pool_free(&device
->dynamic_state_pool
,
3763 sampler
->bindless_state
);
3766 vk_free2(&device
->alloc
, pAllocator
, sampler
);
3769 VkResult
anv_CreateFramebuffer(
3771 const VkFramebufferCreateInfo
* pCreateInfo
,
3772 const VkAllocationCallbacks
* pAllocator
,
3773 VkFramebuffer
* pFramebuffer
)
3775 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3776 struct anv_framebuffer
*framebuffer
;
3778 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
3780 size_t size
= sizeof(*framebuffer
);
3782 /* VK_KHR_imageless_framebuffer extension says:
3784 * If flags includes VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR,
3785 * parameter pAttachments is ignored.
3787 if (!(pCreateInfo
->flags
& VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR
)) {
3788 size
+= sizeof(struct anv_image_view
*) * pCreateInfo
->attachmentCount
;
3789 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
3790 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3791 if (framebuffer
== NULL
)
3792 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3794 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
3795 ANV_FROM_HANDLE(anv_image_view
, iview
, pCreateInfo
->pAttachments
[i
]);
3796 framebuffer
->attachments
[i
] = iview
;
3798 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
3800 assert(device
->enabled_extensions
.KHR_imageless_framebuffer
);
3801 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
3802 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3803 if (framebuffer
== NULL
)
3804 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3806 framebuffer
->attachment_count
= 0;
3809 framebuffer
->width
= pCreateInfo
->width
;
3810 framebuffer
->height
= pCreateInfo
->height
;
3811 framebuffer
->layers
= pCreateInfo
->layers
;
3813 *pFramebuffer
= anv_framebuffer_to_handle(framebuffer
);
3818 void anv_DestroyFramebuffer(
3821 const VkAllocationCallbacks
* pAllocator
)
3823 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3824 ANV_FROM_HANDLE(anv_framebuffer
, fb
, _fb
);
3829 vk_free2(&device
->alloc
, pAllocator
, fb
);
3832 static const VkTimeDomainEXT anv_time_domains
[] = {
3833 VK_TIME_DOMAIN_DEVICE_EXT
,
3834 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
,
3835 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
,
3838 VkResult
anv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
3839 VkPhysicalDevice physicalDevice
,
3840 uint32_t *pTimeDomainCount
,
3841 VkTimeDomainEXT
*pTimeDomains
)
3844 VK_OUTARRAY_MAKE(out
, pTimeDomains
, pTimeDomainCount
);
3846 for (d
= 0; d
< ARRAY_SIZE(anv_time_domains
); d
++) {
3847 vk_outarray_append(&out
, i
) {
3848 *i
= anv_time_domains
[d
];
3852 return vk_outarray_status(&out
);
3856 anv_clock_gettime(clockid_t clock_id
)
3858 struct timespec current
;
3861 ret
= clock_gettime(clock_id
, ¤t
);
3862 if (ret
< 0 && clock_id
== CLOCK_MONOTONIC_RAW
)
3863 ret
= clock_gettime(CLOCK_MONOTONIC
, ¤t
);
3867 return (uint64_t) current
.tv_sec
* 1000000000ULL + current
.tv_nsec
;
3870 #define TIMESTAMP 0x2358
3872 VkResult
anv_GetCalibratedTimestampsEXT(
3874 uint32_t timestampCount
,
3875 const VkCalibratedTimestampInfoEXT
*pTimestampInfos
,
3876 uint64_t *pTimestamps
,
3877 uint64_t *pMaxDeviation
)
3879 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3880 uint64_t timestamp_frequency
= device
->info
.timestamp_frequency
;
3883 uint64_t begin
, end
;
3884 uint64_t max_clock_period
= 0;
3886 begin
= anv_clock_gettime(CLOCK_MONOTONIC_RAW
);
3888 for (d
= 0; d
< timestampCount
; d
++) {
3889 switch (pTimestampInfos
[d
].timeDomain
) {
3890 case VK_TIME_DOMAIN_DEVICE_EXT
:
3891 ret
= anv_gem_reg_read(device
, TIMESTAMP
| 1,
3895 return anv_device_set_lost(device
, "Failed to read the TIMESTAMP "
3898 uint64_t device_period
= DIV_ROUND_UP(1000000000, timestamp_frequency
);
3899 max_clock_period
= MAX2(max_clock_period
, device_period
);
3901 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
:
3902 pTimestamps
[d
] = anv_clock_gettime(CLOCK_MONOTONIC
);
3903 max_clock_period
= MAX2(max_clock_period
, 1);
3906 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
:
3907 pTimestamps
[d
] = begin
;
3915 end
= anv_clock_gettime(CLOCK_MONOTONIC_RAW
);
3918 * The maximum deviation is the sum of the interval over which we
3919 * perform the sampling and the maximum period of any sampled
3920 * clock. That's because the maximum skew between any two sampled
3921 * clock edges is when the sampled clock with the largest period is
3922 * sampled at the end of that period but right at the beginning of the
3923 * sampling interval and some other clock is sampled right at the
3924 * begining of its sampling period and right at the end of the
3925 * sampling interval. Let's assume the GPU has the longest clock
3926 * period and that the application is sampling GPU and monotonic:
3929 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
3930 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
3934 * GPU -----_____-----_____-----_____-----_____
3937 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
3938 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
3940 * Interval <----------------->
3941 * Deviation <-------------------------->
3945 * m = read(monotonic) 2
3948 * We round the sample interval up by one tick to cover sampling error
3949 * in the interval clock
3952 uint64_t sample_interval
= end
- begin
+ 1;
3954 *pMaxDeviation
= sample_interval
+ max_clock_period
;
3959 /* vk_icd.h does not declare this function, so we declare it here to
3960 * suppress Wmissing-prototypes.
3962 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
3963 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
);
3965 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
3966 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
)
3968 /* For the full details on loader interface versioning, see
3969 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
3970 * What follows is a condensed summary, to help you navigate the large and
3971 * confusing official doc.
3973 * - Loader interface v0 is incompatible with later versions. We don't
3976 * - In loader interface v1:
3977 * - The first ICD entrypoint called by the loader is
3978 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
3980 * - The ICD must statically expose no other Vulkan symbol unless it is
3981 * linked with -Bsymbolic.
3982 * - Each dispatchable Vulkan handle created by the ICD must be
3983 * a pointer to a struct whose first member is VK_LOADER_DATA. The
3984 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
3985 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
3986 * vkDestroySurfaceKHR(). The ICD must be capable of working with
3987 * such loader-managed surfaces.
3989 * - Loader interface v2 differs from v1 in:
3990 * - The first ICD entrypoint called by the loader is
3991 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
3992 * statically expose this entrypoint.
3994 * - Loader interface v3 differs from v2 in:
3995 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
3996 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
3997 * because the loader no longer does so.
3999 *pSupportedVersion
= MIN2(*pSupportedVersion
, 3u);