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/debug.h"
36 #include "util/build_id.h"
37 #include "util/disk_cache.h"
38 #include "util/mesa-sha1.h"
39 #include "util/os_file.h"
40 #include "util/u_atomic.h"
41 #include "util/u_string.h"
42 #include "util/xmlpool.h"
45 #include "common/gen_aux_map.h"
46 #include "common/gen_defines.h"
47 #include "compiler/glsl_types.h"
49 #include "genxml/gen7_pack.h"
51 static const char anv_dri_options_xml
[] =
53 DRI_CONF_SECTION_PERFORMANCE
54 DRI_CONF_VK_X11_OVERRIDE_MIN_IMAGE_COUNT(0)
55 DRI_CONF_VK_X11_STRICT_IMAGE_COUNT("false")
59 /* This is probably far to big but it reflects the max size used for messages
60 * in OpenGLs KHR_debug.
62 #define MAX_DEBUG_MESSAGE_LENGTH 4096
65 compiler_debug_log(void *data
, const char *fmt
, ...)
67 char str
[MAX_DEBUG_MESSAGE_LENGTH
];
68 struct anv_device
*device
= (struct anv_device
*)data
;
70 if (list_is_empty(&device
->instance
->debug_report_callbacks
.callbacks
))
75 (void) vsnprintf(str
, MAX_DEBUG_MESSAGE_LENGTH
, fmt
, args
);
78 vk_debug_report(&device
->instance
->debug_report_callbacks
,
79 VK_DEBUG_REPORT_DEBUG_BIT_EXT
,
80 VK_DEBUG_REPORT_OBJECT_TYPE_UNKNOWN_EXT
,
85 compiler_perf_log(void *data
, const char *fmt
, ...)
90 if (unlikely(INTEL_DEBUG
& DEBUG_PERF
))
91 intel_logd_v(fmt
, args
);
97 anv_compute_heap_size(int fd
, uint64_t gtt_size
)
99 /* Query the total ram from the system */
103 uint64_t total_ram
= (uint64_t)info
.totalram
* (uint64_t)info
.mem_unit
;
105 /* We don't want to burn too much ram with the GPU. If the user has 4GiB
106 * or less, we use at most half. If they have more than 4GiB, we use 3/4.
108 uint64_t available_ram
;
109 if (total_ram
<= 4ull * 1024ull * 1024ull * 1024ull)
110 available_ram
= total_ram
/ 2;
112 available_ram
= total_ram
* 3 / 4;
114 /* We also want to leave some padding for things we allocate in the driver,
115 * so don't go over 3/4 of the GTT either.
117 uint64_t available_gtt
= gtt_size
* 3 / 4;
119 return MIN2(available_ram
, available_gtt
);
123 anv_physical_device_init_heaps(struct anv_physical_device
*device
, int fd
)
126 if (anv_gem_get_context_param(fd
, 0, I915_CONTEXT_PARAM_GTT_SIZE
,
128 /* If, for whatever reason, we can't actually get the GTT size from the
129 * kernel (too old?) fall back to the aperture size.
131 anv_perf_warn(NULL
, NULL
,
132 "Failed to get I915_CONTEXT_PARAM_GTT_SIZE: %m");
134 if (anv_gem_get_aperture(fd
, >t_size
) == -1) {
135 return vk_errorf(NULL
, NULL
, VK_ERROR_INITIALIZATION_FAILED
,
136 "failed to get aperture size: %m");
140 device
->supports_48bit_addresses
= (device
->info
.gen
>= 8) &&
141 gtt_size
> (4ULL << 30 /* GiB */);
143 uint64_t heap_size
= anv_compute_heap_size(fd
, gtt_size
);
145 if (heap_size
> (2ull << 30) && !device
->supports_48bit_addresses
) {
146 /* When running with an overridden PCI ID, we may get a GTT size from
147 * the kernel that is greater than 2 GiB but the execbuf check for 48bit
148 * address support can still fail. Just clamp the address space size to
149 * 2 GiB if we don't have 48-bit support.
151 intel_logw("%s:%d: The kernel reported a GTT size larger than 2 GiB but "
152 "not support for 48-bit addresses",
154 heap_size
= 2ull << 30;
157 if (heap_size
<= 3ull * (1ull << 30)) {
158 /* In this case, everything fits nicely into the 32-bit address space,
159 * so there's no need for supporting 48bit addresses on client-allocated
162 device
->memory
.heap_count
= 1;
163 device
->memory
.heaps
[0] = (struct anv_memory_heap
) {
164 .vma_start
= LOW_HEAP_MIN_ADDRESS
,
165 .vma_size
= LOW_HEAP_SIZE
,
167 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
168 .supports_48bit_addresses
= false,
171 /* Not everything will fit nicely into a 32-bit address space. In this
172 * case we need a 64-bit heap. Advertise a small 32-bit heap and a
173 * larger 48-bit heap. If we're in this case, then we have a total heap
174 * size larger than 3GiB which most likely means they have 8 GiB of
175 * video memory and so carving off 1 GiB for the 32-bit heap should be
178 const uint64_t heap_size_32bit
= 1ull << 30;
179 const uint64_t heap_size_48bit
= heap_size
- heap_size_32bit
;
181 assert(device
->supports_48bit_addresses
);
183 device
->memory
.heap_count
= 2;
184 device
->memory
.heaps
[0] = (struct anv_memory_heap
) {
185 .vma_start
= HIGH_HEAP_MIN_ADDRESS
,
186 /* Leave the last 4GiB out of the high vma range, so that no state
187 * base address + size can overflow 48 bits. For more information see
188 * the comment about Wa32bitGeneralStateOffset in anv_allocator.c
190 .vma_size
= gtt_size
- (1ull << 32) - HIGH_HEAP_MIN_ADDRESS
,
191 .size
= heap_size_48bit
,
192 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
193 .supports_48bit_addresses
= true,
195 device
->memory
.heaps
[1] = (struct anv_memory_heap
) {
196 .vma_start
= LOW_HEAP_MIN_ADDRESS
,
197 .vma_size
= LOW_HEAP_SIZE
,
198 .size
= heap_size_32bit
,
199 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
200 .supports_48bit_addresses
= false,
204 uint32_t type_count
= 0;
205 for (uint32_t heap
= 0; heap
< device
->memory
.heap_count
; heap
++) {
206 uint32_t valid_buffer_usage
= ~0;
208 /* There appears to be a hardware issue in the VF cache where it only
209 * considers the bottom 32 bits of memory addresses. If you happen to
210 * have two vertex buffers which get placed exactly 4 GiB apart and use
211 * them in back-to-back draw calls, you can get collisions. In order to
212 * solve this problem, we require vertex and index buffers be bound to
213 * memory allocated out of the 32-bit heap.
215 if (device
->memory
.heaps
[heap
].supports_48bit_addresses
) {
216 valid_buffer_usage
&= ~(VK_BUFFER_USAGE_INDEX_BUFFER_BIT
|
217 VK_BUFFER_USAGE_VERTEX_BUFFER_BIT
);
220 if (device
->info
.has_llc
) {
221 /* Big core GPUs share LLC with the CPU and thus one memory type can be
222 * both cached and coherent at the same time.
224 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
225 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
226 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
227 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
228 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
230 .valid_buffer_usage
= valid_buffer_usage
,
233 /* The spec requires that we expose a host-visible, coherent memory
234 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
235 * to give the application a choice between cached, but not coherent and
236 * coherent but uncached (WC though).
238 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
239 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
240 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
241 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
243 .valid_buffer_usage
= valid_buffer_usage
,
245 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
246 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
247 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
248 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
250 .valid_buffer_usage
= valid_buffer_usage
,
254 device
->memory
.type_count
= type_count
;
260 anv_physical_device_init_uuids(struct anv_physical_device
*device
)
262 const struct build_id_note
*note
=
263 build_id_find_nhdr_for_addr(anv_physical_device_init_uuids
);
265 return vk_errorf(device
->instance
, device
,
266 VK_ERROR_INITIALIZATION_FAILED
,
267 "Failed to find build-id");
270 unsigned build_id_len
= build_id_length(note
);
271 if (build_id_len
< 20) {
272 return vk_errorf(device
->instance
, device
,
273 VK_ERROR_INITIALIZATION_FAILED
,
274 "build-id too short. It needs to be a SHA");
277 memcpy(device
->driver_build_sha1
, build_id_data(note
), 20);
279 struct mesa_sha1 sha1_ctx
;
281 STATIC_ASSERT(VK_UUID_SIZE
<= sizeof(sha1
));
283 /* The pipeline cache UUID is used for determining when a pipeline cache is
284 * invalid. It needs both a driver build and the PCI ID of the device.
286 _mesa_sha1_init(&sha1_ctx
);
287 _mesa_sha1_update(&sha1_ctx
, build_id_data(note
), build_id_len
);
288 _mesa_sha1_update(&sha1_ctx
, &device
->chipset_id
,
289 sizeof(device
->chipset_id
));
290 _mesa_sha1_update(&sha1_ctx
, &device
->always_use_bindless
,
291 sizeof(device
->always_use_bindless
));
292 _mesa_sha1_update(&sha1_ctx
, &device
->has_a64_buffer_access
,
293 sizeof(device
->has_a64_buffer_access
));
294 _mesa_sha1_update(&sha1_ctx
, &device
->has_bindless_images
,
295 sizeof(device
->has_bindless_images
));
296 _mesa_sha1_update(&sha1_ctx
, &device
->has_bindless_samplers
,
297 sizeof(device
->has_bindless_samplers
));
298 _mesa_sha1_final(&sha1_ctx
, sha1
);
299 memcpy(device
->pipeline_cache_uuid
, sha1
, VK_UUID_SIZE
);
301 /* The driver UUID is used for determining sharability of images and memory
302 * between two Vulkan instances in separate processes. People who want to
303 * share memory need to also check the device UUID (below) so all this
304 * needs to be is the build-id.
306 memcpy(device
->driver_uuid
, build_id_data(note
), VK_UUID_SIZE
);
308 /* The device UUID uniquely identifies the given device within the machine.
309 * Since we never have more than one device, this doesn't need to be a real
310 * UUID. However, on the off-chance that someone tries to use this to
311 * cache pre-tiled images or something of the like, we use the PCI ID and
312 * some bits of ISL info to ensure that this is safe.
314 _mesa_sha1_init(&sha1_ctx
);
315 _mesa_sha1_update(&sha1_ctx
, &device
->chipset_id
,
316 sizeof(device
->chipset_id
));
317 _mesa_sha1_update(&sha1_ctx
, &device
->isl_dev
.has_bit6_swizzling
,
318 sizeof(device
->isl_dev
.has_bit6_swizzling
));
319 _mesa_sha1_final(&sha1_ctx
, sha1
);
320 memcpy(device
->device_uuid
, sha1
, VK_UUID_SIZE
);
326 anv_physical_device_init_disk_cache(struct anv_physical_device
*device
)
328 #ifdef ENABLE_SHADER_CACHE
330 ASSERTED
int len
= snprintf(renderer
, sizeof(renderer
), "anv_%04x",
332 assert(len
== sizeof(renderer
) - 2);
335 _mesa_sha1_format(timestamp
, device
->driver_build_sha1
);
337 const uint64_t driver_flags
=
338 brw_get_compiler_config_value(device
->compiler
);
339 device
->disk_cache
= disk_cache_create(renderer
, timestamp
, driver_flags
);
341 device
->disk_cache
= NULL
;
346 anv_physical_device_free_disk_cache(struct anv_physical_device
*device
)
348 #ifdef ENABLE_SHADER_CACHE
349 if (device
->disk_cache
)
350 disk_cache_destroy(device
->disk_cache
);
352 assert(device
->disk_cache
== NULL
);
357 get_available_system_memory()
359 char *meminfo
= os_read_file("/proc/meminfo");
363 char *str
= strstr(meminfo
, "MemAvailable:");
369 uint64_t kb_mem_available
;
370 if (sscanf(str
, "MemAvailable: %" PRIx64
, &kb_mem_available
) == 1) {
372 return kb_mem_available
<< 10;
380 anv_physical_device_init(struct anv_physical_device
*device
,
381 struct anv_instance
*instance
,
382 drmDevicePtr drm_device
)
384 const char *primary_path
= drm_device
->nodes
[DRM_NODE_PRIMARY
];
385 const char *path
= drm_device
->nodes
[DRM_NODE_RENDER
];
390 brw_process_intel_debug_variable();
392 fd
= open(path
, O_RDWR
| O_CLOEXEC
);
394 return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
396 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
397 device
->instance
= instance
;
399 assert(strlen(path
) < ARRAY_SIZE(device
->path
));
400 snprintf(device
->path
, ARRAY_SIZE(device
->path
), "%s", path
);
402 if (!gen_get_device_info_from_fd(fd
, &device
->info
)) {
403 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
406 device
->chipset_id
= device
->info
.chipset_id
;
407 device
->no_hw
= device
->info
.no_hw
;
409 if (getenv("INTEL_NO_HW") != NULL
)
410 device
->no_hw
= true;
412 device
->pci_info
.domain
= drm_device
->businfo
.pci
->domain
;
413 device
->pci_info
.bus
= drm_device
->businfo
.pci
->bus
;
414 device
->pci_info
.device
= drm_device
->businfo
.pci
->dev
;
415 device
->pci_info
.function
= drm_device
->businfo
.pci
->func
;
417 device
->name
= gen_get_device_name(device
->chipset_id
);
419 if (device
->info
.is_haswell
) {
420 intel_logw("Haswell Vulkan support is incomplete");
421 } else if (device
->info
.gen
== 7 && !device
->info
.is_baytrail
) {
422 intel_logw("Ivy Bridge Vulkan support is incomplete");
423 } else if (device
->info
.gen
== 7 && device
->info
.is_baytrail
) {
424 intel_logw("Bay Trail Vulkan support is incomplete");
425 } else if (device
->info
.gen
>= 8 && device
->info
.gen
<= 11) {
426 /* Gen8-11 fully supported */
427 } else if (device
->info
.gen
== 12) {
428 intel_logw("Vulkan is not yet fully supported on gen12");
430 result
= vk_errorf(device
->instance
, device
,
431 VK_ERROR_INCOMPATIBLE_DRIVER
,
432 "Vulkan not yet supported on %s", device
->name
);
436 device
->cmd_parser_version
= -1;
437 if (device
->info
.gen
== 7) {
438 device
->cmd_parser_version
=
439 anv_gem_get_param(fd
, I915_PARAM_CMD_PARSER_VERSION
);
440 if (device
->cmd_parser_version
== -1) {
441 result
= vk_errorf(device
->instance
, device
,
442 VK_ERROR_INITIALIZATION_FAILED
,
443 "failed to get command parser version");
448 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_WAIT_TIMEOUT
)) {
449 result
= vk_errorf(device
->instance
, device
,
450 VK_ERROR_INITIALIZATION_FAILED
,
451 "kernel missing gem wait");
455 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_EXECBUF2
)) {
456 result
= vk_errorf(device
->instance
, device
,
457 VK_ERROR_INITIALIZATION_FAILED
,
458 "kernel missing execbuf2");
462 if (!device
->info
.has_llc
&&
463 anv_gem_get_param(fd
, I915_PARAM_MMAP_VERSION
) < 1) {
464 result
= vk_errorf(device
->instance
, device
,
465 VK_ERROR_INITIALIZATION_FAILED
,
466 "kernel missing wc mmap");
470 result
= anv_physical_device_init_heaps(device
, fd
);
471 if (result
!= VK_SUCCESS
)
474 device
->has_exec_async
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_ASYNC
);
475 device
->has_exec_capture
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_CAPTURE
);
476 device
->has_exec_fence
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE
);
477 device
->has_syncobj
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE_ARRAY
);
478 device
->has_syncobj_wait
= device
->has_syncobj
&&
479 anv_gem_supports_syncobj_wait(fd
);
480 device
->has_context_priority
= anv_gem_has_context_priority(fd
);
482 device
->use_softpin
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_SOFTPIN
)
483 && device
->supports_48bit_addresses
;
485 device
->has_context_isolation
=
486 anv_gem_get_param(fd
, I915_PARAM_HAS_CONTEXT_ISOLATION
);
488 device
->always_use_bindless
=
489 env_var_as_boolean("ANV_ALWAYS_BINDLESS", false);
491 /* We first got the A64 messages on broadwell and we can only use them if
492 * we can pass addresses directly into the shader which requires softpin.
494 device
->has_a64_buffer_access
= device
->info
.gen
>= 8 &&
497 /* We first get bindless image access on Skylake and we can only really do
498 * it if we don't have any relocations so we need softpin.
500 device
->has_bindless_images
= device
->info
.gen
>= 9 &&
503 /* We've had bindless samplers since Ivy Bridge (forever in Vulkan terms)
504 * because it's just a matter of setting the sampler address in the sample
505 * message header. However, we've not bothered to wire it up for vec4 so
506 * we leave it disabled on gen7.
508 device
->has_bindless_samplers
= device
->info
.gen
>= 8;
510 device
->has_mem_available
= get_available_system_memory() != 0;
512 /* Starting with Gen10, the timestamp frequency of the command streamer may
513 * vary from one part to another. We can query the value from the kernel.
515 if (device
->info
.gen
>= 10) {
516 int timestamp_frequency
=
517 anv_gem_get_param(fd
, I915_PARAM_CS_TIMESTAMP_FREQUENCY
);
519 if (timestamp_frequency
< 0)
520 intel_logw("Kernel 4.16-rc1+ required to properly query CS timestamp frequency");
522 device
->info
.timestamp_frequency
= timestamp_frequency
;
525 /* GENs prior to 8 do not support EU/Subslice info */
526 if (device
->info
.gen
>= 8) {
527 device
->subslice_total
= anv_gem_get_param(fd
, I915_PARAM_SUBSLICE_TOTAL
);
528 device
->eu_total
= anv_gem_get_param(fd
, I915_PARAM_EU_TOTAL
);
530 /* Without this information, we cannot get the right Braswell
531 * brandstrings, and we have to use conservative numbers for GPGPU on
532 * many platforms, but otherwise, things will just work.
534 if (device
->subslice_total
< 1 || device
->eu_total
< 1) {
535 intel_logw("Kernel 4.1 required to properly query GPU properties");
537 } else if (device
->info
.gen
== 7) {
538 device
->subslice_total
= 1 << (device
->info
.gt
- 1);
541 if (device
->info
.is_cherryview
&&
542 device
->subslice_total
> 0 && device
->eu_total
> 0) {
543 /* Logical CS threads = EUs per subslice * num threads per EU */
544 uint32_t max_cs_threads
=
545 device
->eu_total
/ device
->subslice_total
* device
->info
.num_thread_per_eu
;
547 /* Fuse configurations may give more threads than expected, never less. */
548 if (max_cs_threads
> device
->info
.max_cs_threads
)
549 device
->info
.max_cs_threads
= max_cs_threads
;
552 device
->compiler
= brw_compiler_create(NULL
, &device
->info
);
553 if (device
->compiler
== NULL
) {
554 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
557 device
->compiler
->shader_debug_log
= compiler_debug_log
;
558 device
->compiler
->shader_perf_log
= compiler_perf_log
;
559 device
->compiler
->supports_pull_constants
= false;
560 device
->compiler
->constant_buffer_0_is_relative
=
561 device
->info
.gen
< 8 || !device
->has_context_isolation
;
562 device
->compiler
->supports_shader_constants
= true;
564 /* Broadwell PRM says:
566 * "Before Gen8, there was a historical configuration control field to
567 * swizzle address bit[6] for in X/Y tiling modes. This was set in three
568 * different places: TILECTL[1:0], ARB_MODE[5:4], and
569 * DISP_ARB_CTL[14:13].
571 * For Gen8 and subsequent generations, the swizzle fields are all
572 * reserved, and the CPU's memory controller performs all address
573 * swizzling modifications."
576 device
->info
.gen
< 8 && anv_gem_get_bit6_swizzle(fd
, I915_TILING_X
);
578 isl_device_init(&device
->isl_dev
, &device
->info
, swizzled
);
580 result
= anv_physical_device_init_uuids(device
);
581 if (result
!= VK_SUCCESS
)
584 anv_physical_device_init_disk_cache(device
);
586 if (instance
->enabled_extensions
.KHR_display
) {
587 master_fd
= open(primary_path
, O_RDWR
| O_CLOEXEC
);
588 if (master_fd
>= 0) {
589 /* prod the device with a GETPARAM call which will fail if
590 * we don't have permission to even render on this device
592 if (anv_gem_get_param(master_fd
, I915_PARAM_CHIPSET_ID
) == 0) {
598 device
->master_fd
= master_fd
;
600 result
= anv_init_wsi(device
);
601 if (result
!= VK_SUCCESS
) {
602 ralloc_free(device
->compiler
);
603 anv_physical_device_free_disk_cache(device
);
607 device
->perf
= anv_get_perf(&device
->info
, fd
);
609 anv_physical_device_get_supported_extensions(device
,
610 &device
->supported_extensions
);
613 device
->local_fd
= fd
;
625 anv_physical_device_finish(struct anv_physical_device
*device
)
627 anv_finish_wsi(device
);
628 anv_physical_device_free_disk_cache(device
);
629 ralloc_free(device
->compiler
);
630 ralloc_free(device
->perf
);
631 close(device
->local_fd
);
632 if (device
->master_fd
>= 0)
633 close(device
->master_fd
);
637 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
638 VkSystemAllocationScope allocationScope
)
644 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
645 size_t align
, VkSystemAllocationScope allocationScope
)
647 return realloc(pOriginal
, size
);
651 default_free_func(void *pUserData
, void *pMemory
)
656 static const VkAllocationCallbacks default_alloc
= {
658 .pfnAllocation
= default_alloc_func
,
659 .pfnReallocation
= default_realloc_func
,
660 .pfnFree
= default_free_func
,
663 VkResult
anv_EnumerateInstanceExtensionProperties(
664 const char* pLayerName
,
665 uint32_t* pPropertyCount
,
666 VkExtensionProperties
* pProperties
)
668 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
670 for (int i
= 0; i
< ANV_INSTANCE_EXTENSION_COUNT
; i
++) {
671 if (anv_instance_extensions_supported
.extensions
[i
]) {
672 vk_outarray_append(&out
, prop
) {
673 *prop
= anv_instance_extensions
[i
];
678 return vk_outarray_status(&out
);
681 VkResult
anv_CreateInstance(
682 const VkInstanceCreateInfo
* pCreateInfo
,
683 const VkAllocationCallbacks
* pAllocator
,
684 VkInstance
* pInstance
)
686 struct anv_instance
*instance
;
689 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
691 struct anv_instance_extension_table enabled_extensions
= {};
692 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
694 for (idx
= 0; idx
< ANV_INSTANCE_EXTENSION_COUNT
; idx
++) {
695 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
696 anv_instance_extensions
[idx
].extensionName
) == 0)
700 if (idx
>= ANV_INSTANCE_EXTENSION_COUNT
)
701 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
703 if (!anv_instance_extensions_supported
.extensions
[idx
])
704 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
706 enabled_extensions
.extensions
[idx
] = true;
709 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
710 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
712 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
714 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
717 instance
->alloc
= *pAllocator
;
719 instance
->alloc
= default_alloc
;
721 instance
->app_info
= (struct anv_app_info
) { .api_version
= 0 };
722 if (pCreateInfo
->pApplicationInfo
) {
723 const VkApplicationInfo
*app
= pCreateInfo
->pApplicationInfo
;
725 instance
->app_info
.app_name
=
726 vk_strdup(&instance
->alloc
, app
->pApplicationName
,
727 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
728 instance
->app_info
.app_version
= app
->applicationVersion
;
730 instance
->app_info
.engine_name
=
731 vk_strdup(&instance
->alloc
, app
->pEngineName
,
732 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
733 instance
->app_info
.engine_version
= app
->engineVersion
;
735 instance
->app_info
.api_version
= app
->apiVersion
;
738 if (instance
->app_info
.api_version
== 0)
739 instance
->app_info
.api_version
= VK_API_VERSION_1_0
;
741 instance
->enabled_extensions
= enabled_extensions
;
743 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->dispatch
.entrypoints
); i
++) {
744 /* Vulkan requires that entrypoints for extensions which have not been
745 * enabled must not be advertised.
747 if (!anv_instance_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
748 &instance
->enabled_extensions
)) {
749 instance
->dispatch
.entrypoints
[i
] = NULL
;
751 instance
->dispatch
.entrypoints
[i
] =
752 anv_instance_dispatch_table
.entrypoints
[i
];
756 struct anv_physical_device
*pdevice
= &instance
->physicalDevice
;
757 for (unsigned i
= 0; i
< ARRAY_SIZE(pdevice
->dispatch
.entrypoints
); i
++) {
758 /* Vulkan requires that entrypoints for extensions which have not been
759 * enabled must not be advertised.
761 if (!anv_physical_device_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
762 &instance
->enabled_extensions
)) {
763 pdevice
->dispatch
.entrypoints
[i
] = NULL
;
765 pdevice
->dispatch
.entrypoints
[i
] =
766 anv_physical_device_dispatch_table
.entrypoints
[i
];
770 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->device_dispatch
.entrypoints
); i
++) {
771 /* Vulkan requires that entrypoints for extensions which have not been
772 * enabled must not be advertised.
774 if (!anv_device_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
775 &instance
->enabled_extensions
, NULL
)) {
776 instance
->device_dispatch
.entrypoints
[i
] = NULL
;
778 instance
->device_dispatch
.entrypoints
[i
] =
779 anv_device_dispatch_table
.entrypoints
[i
];
783 instance
->physicalDeviceCount
= -1;
785 result
= vk_debug_report_instance_init(&instance
->debug_report_callbacks
);
786 if (result
!= VK_SUCCESS
) {
787 vk_free2(&default_alloc
, pAllocator
, instance
);
788 return vk_error(result
);
791 instance
->pipeline_cache_enabled
=
792 env_var_as_boolean("ANV_ENABLE_PIPELINE_CACHE", true);
794 glsl_type_singleton_init_or_ref();
796 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
798 driParseOptionInfo(&instance
->available_dri_options
, anv_dri_options_xml
);
799 driParseConfigFiles(&instance
->dri_options
, &instance
->available_dri_options
,
801 instance
->app_info
.engine_name
,
802 instance
->app_info
.engine_version
);
804 *pInstance
= anv_instance_to_handle(instance
);
809 void anv_DestroyInstance(
810 VkInstance _instance
,
811 const VkAllocationCallbacks
* pAllocator
)
813 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
818 if (instance
->physicalDeviceCount
> 0) {
819 /* We support at most one physical device. */
820 assert(instance
->physicalDeviceCount
== 1);
821 anv_physical_device_finish(&instance
->physicalDevice
);
824 vk_free(&instance
->alloc
, (char *)instance
->app_info
.app_name
);
825 vk_free(&instance
->alloc
, (char *)instance
->app_info
.engine_name
);
827 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
829 vk_debug_report_instance_destroy(&instance
->debug_report_callbacks
);
831 glsl_type_singleton_decref();
833 driDestroyOptionCache(&instance
->dri_options
);
834 driDestroyOptionInfo(&instance
->available_dri_options
);
836 vk_free(&instance
->alloc
, instance
);
840 anv_enumerate_devices(struct anv_instance
*instance
)
842 /* TODO: Check for more devices ? */
843 drmDevicePtr devices
[8];
844 VkResult result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
847 instance
->physicalDeviceCount
= 0;
849 max_devices
= drmGetDevices2(0, devices
, ARRAY_SIZE(devices
));
851 return VK_ERROR_INCOMPATIBLE_DRIVER
;
853 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
854 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
855 devices
[i
]->bustype
== DRM_BUS_PCI
&&
856 devices
[i
]->deviceinfo
.pci
->vendor_id
== 0x8086) {
858 result
= anv_physical_device_init(&instance
->physicalDevice
,
859 instance
, devices
[i
]);
860 if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
864 drmFreeDevices(devices
, max_devices
);
866 if (result
== VK_SUCCESS
)
867 instance
->physicalDeviceCount
= 1;
873 anv_instance_ensure_physical_device(struct anv_instance
*instance
)
875 if (instance
->physicalDeviceCount
< 0) {
876 VkResult result
= anv_enumerate_devices(instance
);
877 if (result
!= VK_SUCCESS
&&
878 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
885 VkResult
anv_EnumeratePhysicalDevices(
886 VkInstance _instance
,
887 uint32_t* pPhysicalDeviceCount
,
888 VkPhysicalDevice
* pPhysicalDevices
)
890 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
891 VK_OUTARRAY_MAKE(out
, pPhysicalDevices
, pPhysicalDeviceCount
);
893 VkResult result
= anv_instance_ensure_physical_device(instance
);
894 if (result
!= VK_SUCCESS
)
897 if (instance
->physicalDeviceCount
== 0)
900 assert(instance
->physicalDeviceCount
== 1);
901 vk_outarray_append(&out
, i
) {
902 *i
= anv_physical_device_to_handle(&instance
->physicalDevice
);
905 return vk_outarray_status(&out
);
908 VkResult
anv_EnumeratePhysicalDeviceGroups(
909 VkInstance _instance
,
910 uint32_t* pPhysicalDeviceGroupCount
,
911 VkPhysicalDeviceGroupProperties
* pPhysicalDeviceGroupProperties
)
913 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
914 VK_OUTARRAY_MAKE(out
, pPhysicalDeviceGroupProperties
,
915 pPhysicalDeviceGroupCount
);
917 VkResult result
= anv_instance_ensure_physical_device(instance
);
918 if (result
!= VK_SUCCESS
)
921 if (instance
->physicalDeviceCount
== 0)
924 assert(instance
->physicalDeviceCount
== 1);
926 vk_outarray_append(&out
, p
) {
927 p
->physicalDeviceCount
= 1;
928 memset(p
->physicalDevices
, 0, sizeof(p
->physicalDevices
));
929 p
->physicalDevices
[0] =
930 anv_physical_device_to_handle(&instance
->physicalDevice
);
931 p
->subsetAllocation
= false;
933 vk_foreach_struct(ext
, p
->pNext
)
934 anv_debug_ignored_stype(ext
->sType
);
937 return vk_outarray_status(&out
);
940 void anv_GetPhysicalDeviceFeatures(
941 VkPhysicalDevice physicalDevice
,
942 VkPhysicalDeviceFeatures
* pFeatures
)
944 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
946 *pFeatures
= (VkPhysicalDeviceFeatures
) {
947 .robustBufferAccess
= true,
948 .fullDrawIndexUint32
= true,
949 .imageCubeArray
= true,
950 .independentBlend
= true,
951 .geometryShader
= true,
952 .tessellationShader
= true,
953 .sampleRateShading
= true,
954 .dualSrcBlend
= true,
956 .multiDrawIndirect
= true,
957 .drawIndirectFirstInstance
= true,
959 .depthBiasClamp
= true,
960 .fillModeNonSolid
= true,
961 .depthBounds
= pdevice
->info
.gen
>= 12,
965 .multiViewport
= true,
966 .samplerAnisotropy
= true,
967 .textureCompressionETC2
= pdevice
->info
.gen
>= 8 ||
968 pdevice
->info
.is_baytrail
,
969 .textureCompressionASTC_LDR
= pdevice
->info
.gen
>= 9, /* FINISHME CHV */
970 .textureCompressionBC
= true,
971 .occlusionQueryPrecise
= true,
972 .pipelineStatisticsQuery
= true,
973 .fragmentStoresAndAtomics
= true,
974 .shaderTessellationAndGeometryPointSize
= true,
975 .shaderImageGatherExtended
= true,
976 .shaderStorageImageExtendedFormats
= true,
977 .shaderStorageImageMultisample
= false,
978 .shaderStorageImageReadWithoutFormat
= false,
979 .shaderStorageImageWriteWithoutFormat
= true,
980 .shaderUniformBufferArrayDynamicIndexing
= true,
981 .shaderSampledImageArrayDynamicIndexing
= true,
982 .shaderStorageBufferArrayDynamicIndexing
= true,
983 .shaderStorageImageArrayDynamicIndexing
= true,
984 .shaderClipDistance
= true,
985 .shaderCullDistance
= true,
986 .shaderFloat64
= pdevice
->info
.gen
>= 8 &&
987 pdevice
->info
.has_64bit_types
,
988 .shaderInt64
= pdevice
->info
.gen
>= 8 &&
989 pdevice
->info
.has_64bit_types
,
990 .shaderInt16
= pdevice
->info
.gen
>= 8,
991 .shaderResourceMinLod
= pdevice
->info
.gen
>= 9,
992 .variableMultisampleRate
= true,
993 .inheritedQueries
= true,
996 /* We can't do image stores in vec4 shaders */
997 pFeatures
->vertexPipelineStoresAndAtomics
=
998 pdevice
->compiler
->scalar_stage
[MESA_SHADER_VERTEX
] &&
999 pdevice
->compiler
->scalar_stage
[MESA_SHADER_GEOMETRY
];
1001 struct anv_app_info
*app_info
= &pdevice
->instance
->app_info
;
1003 /* The new DOOM and Wolfenstein games require depthBounds without
1004 * checking for it. They seem to run fine without it so just claim it's
1005 * there and accept the consequences.
1007 if (app_info
->engine_name
&& strcmp(app_info
->engine_name
, "idTech") == 0)
1008 pFeatures
->depthBounds
= true;
1011 void anv_GetPhysicalDeviceFeatures2(
1012 VkPhysicalDevice physicalDevice
,
1013 VkPhysicalDeviceFeatures2
* pFeatures
)
1015 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1016 anv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
1018 vk_foreach_struct(ext
, pFeatures
->pNext
) {
1019 switch (ext
->sType
) {
1020 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES_KHR
: {
1021 VkPhysicalDevice8BitStorageFeaturesKHR
*features
=
1022 (VkPhysicalDevice8BitStorageFeaturesKHR
*)ext
;
1023 features
->storageBuffer8BitAccess
= pdevice
->info
.gen
>= 8;
1024 features
->uniformAndStorageBuffer8BitAccess
= pdevice
->info
.gen
>= 8;
1025 features
->storagePushConstant8
= pdevice
->info
.gen
>= 8;
1029 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES
: {
1030 VkPhysicalDevice16BitStorageFeatures
*features
=
1031 (VkPhysicalDevice16BitStorageFeatures
*)ext
;
1032 features
->storageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
1033 features
->uniformAndStorageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
1034 features
->storagePushConstant16
= pdevice
->info
.gen
>= 8;
1035 features
->storageInputOutput16
= false;
1039 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_EXT
: {
1040 VkPhysicalDeviceBufferDeviceAddressFeaturesEXT
*features
= (void *)ext
;
1041 features
->bufferDeviceAddress
= pdevice
->has_a64_buffer_access
;
1042 features
->bufferDeviceAddressCaptureReplay
= false;
1043 features
->bufferDeviceAddressMultiDevice
= false;
1047 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COMPUTE_SHADER_DERIVATIVES_FEATURES_NV
: {
1048 VkPhysicalDeviceComputeShaderDerivativesFeaturesNV
*features
=
1049 (VkPhysicalDeviceComputeShaderDerivativesFeaturesNV
*)ext
;
1050 features
->computeDerivativeGroupQuads
= true;
1051 features
->computeDerivativeGroupLinear
= true;
1055 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT
: {
1056 VkPhysicalDeviceConditionalRenderingFeaturesEXT
*features
=
1057 (VkPhysicalDeviceConditionalRenderingFeaturesEXT
*)ext
;
1058 features
->conditionalRendering
= pdevice
->info
.gen
>= 8 ||
1059 pdevice
->info
.is_haswell
;
1060 features
->inheritedConditionalRendering
= pdevice
->info
.gen
>= 8 ||
1061 pdevice
->info
.is_haswell
;
1065 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_CLIP_ENABLE_FEATURES_EXT
: {
1066 VkPhysicalDeviceDepthClipEnableFeaturesEXT
*features
=
1067 (VkPhysicalDeviceDepthClipEnableFeaturesEXT
*)ext
;
1068 features
->depthClipEnable
= true;
1072 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR
: {
1073 VkPhysicalDeviceFloat16Int8FeaturesKHR
*features
= (void *)ext
;
1074 features
->shaderFloat16
= pdevice
->info
.gen
>= 8;
1075 features
->shaderInt8
= pdevice
->info
.gen
>= 8;
1079 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FRAGMENT_SHADER_INTERLOCK_FEATURES_EXT
: {
1080 VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT
*features
=
1081 (VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT
*)ext
;
1082 features
->fragmentShaderSampleInterlock
= pdevice
->info
.gen
>= 9;
1083 features
->fragmentShaderPixelInterlock
= pdevice
->info
.gen
>= 9;
1084 features
->fragmentShaderShadingRateInterlock
= false;
1088 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_QUERY_RESET_FEATURES_EXT
: {
1089 VkPhysicalDeviceHostQueryResetFeaturesEXT
*features
=
1090 (VkPhysicalDeviceHostQueryResetFeaturesEXT
*)ext
;
1091 features
->hostQueryReset
= true;
1095 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT
: {
1096 VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*features
=
1097 (VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*)ext
;
1098 features
->shaderInputAttachmentArrayDynamicIndexing
= false;
1099 features
->shaderUniformTexelBufferArrayDynamicIndexing
= true;
1100 features
->shaderStorageTexelBufferArrayDynamicIndexing
= true;
1101 features
->shaderUniformBufferArrayNonUniformIndexing
= false;
1102 features
->shaderSampledImageArrayNonUniformIndexing
= true;
1103 features
->shaderStorageBufferArrayNonUniformIndexing
= true;
1104 features
->shaderStorageImageArrayNonUniformIndexing
= true;
1105 features
->shaderInputAttachmentArrayNonUniformIndexing
= false;
1106 features
->shaderUniformTexelBufferArrayNonUniformIndexing
= true;
1107 features
->shaderStorageTexelBufferArrayNonUniformIndexing
= true;
1108 features
->descriptorBindingUniformBufferUpdateAfterBind
= false;
1109 features
->descriptorBindingSampledImageUpdateAfterBind
= true;
1110 features
->descriptorBindingStorageImageUpdateAfterBind
= true;
1111 features
->descriptorBindingStorageBufferUpdateAfterBind
= true;
1112 features
->descriptorBindingUniformTexelBufferUpdateAfterBind
= true;
1113 features
->descriptorBindingStorageTexelBufferUpdateAfterBind
= true;
1114 features
->descriptorBindingUpdateUnusedWhilePending
= true;
1115 features
->descriptorBindingPartiallyBound
= true;
1116 features
->descriptorBindingVariableDescriptorCount
= false;
1117 features
->runtimeDescriptorArray
= true;
1121 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INDEX_TYPE_UINT8_FEATURES_EXT
: {
1122 VkPhysicalDeviceIndexTypeUint8FeaturesEXT
*features
=
1123 (VkPhysicalDeviceIndexTypeUint8FeaturesEXT
*)ext
;
1124 features
->indexTypeUint8
= true;
1128 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_FEATURES_EXT
: {
1129 VkPhysicalDeviceInlineUniformBlockFeaturesEXT
*features
=
1130 (VkPhysicalDeviceInlineUniformBlockFeaturesEXT
*)ext
;
1131 features
->inlineUniformBlock
= true;
1132 features
->descriptorBindingInlineUniformBlockUpdateAfterBind
= true;
1136 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_FEATURES_EXT
: {
1137 VkPhysicalDeviceLineRasterizationFeaturesEXT
*features
=
1138 (VkPhysicalDeviceLineRasterizationFeaturesEXT
*)ext
;
1139 features
->rectangularLines
= true;
1140 features
->bresenhamLines
= true;
1141 features
->smoothLines
= true;
1142 features
->stippledRectangularLines
= false;
1143 features
->stippledBresenhamLines
= true;
1144 features
->stippledSmoothLines
= false;
1148 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES
: {
1149 VkPhysicalDeviceMultiviewFeatures
*features
=
1150 (VkPhysicalDeviceMultiviewFeatures
*)ext
;
1151 features
->multiview
= true;
1152 features
->multiviewGeometryShader
= true;
1153 features
->multiviewTessellationShader
= true;
1157 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGELESS_FRAMEBUFFER_FEATURES_KHR
: {
1158 VkPhysicalDeviceImagelessFramebufferFeaturesKHR
*features
=
1159 (VkPhysicalDeviceImagelessFramebufferFeaturesKHR
*)ext
;
1160 features
->imagelessFramebuffer
= true;
1164 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR
: {
1165 VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR
*features
=
1166 (VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR
*)ext
;
1167 features
->pipelineExecutableInfo
= true;
1171 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES
: {
1172 VkPhysicalDeviceProtectedMemoryFeatures
*features
= (void *)ext
;
1173 features
->protectedMemory
= false;
1177 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES
: {
1178 VkPhysicalDeviceSamplerYcbcrConversionFeatures
*features
=
1179 (VkPhysicalDeviceSamplerYcbcrConversionFeatures
*) ext
;
1180 features
->samplerYcbcrConversion
= true;
1184 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES_EXT
: {
1185 VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*features
=
1186 (VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*)ext
;
1187 features
->scalarBlockLayout
= true;
1191 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SEPARATE_DEPTH_STENCIL_LAYOUTS_FEATURES_KHR
: {
1192 VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR
*features
=
1193 (VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR
*)ext
;
1194 features
->separateDepthStencilLayouts
= true;
1198 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_INT64_FEATURES_KHR
: {
1199 VkPhysicalDeviceShaderAtomicInt64FeaturesKHR
*features
= (void *)ext
;
1200 features
->shaderBufferInt64Atomics
=
1201 pdevice
->info
.gen
>= 9 && pdevice
->use_softpin
;
1202 features
->shaderSharedInt64Atomics
= VK_FALSE
;
1206 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DEMOTE_TO_HELPER_INVOCATION_FEATURES_EXT
: {
1207 VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT
*features
= (void *)ext
;
1208 features
->shaderDemoteToHelperInvocation
= true;
1212 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CLOCK_FEATURES_KHR
: {
1213 VkPhysicalDeviceShaderClockFeaturesKHR
*features
=
1214 (VkPhysicalDeviceShaderClockFeaturesKHR
*)ext
;
1215 features
->shaderSubgroupClock
= true;
1216 features
->shaderDeviceClock
= false;
1220 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETERS_FEATURES
: {
1221 VkPhysicalDeviceShaderDrawParametersFeatures
*features
= (void *)ext
;
1222 features
->shaderDrawParameters
= true;
1226 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_SUBGROUP_EXTENDED_TYPES_FEATURES_KHR
: {
1227 VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR
*features
=
1228 (VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR
*)ext
;
1229 features
->shaderSubgroupExtendedTypes
= true;
1233 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_FEATURES_EXT
: {
1234 VkPhysicalDeviceSubgroupSizeControlFeaturesEXT
*features
=
1235 (VkPhysicalDeviceSubgroupSizeControlFeaturesEXT
*)ext
;
1236 features
->subgroupSizeControl
= true;
1237 features
->computeFullSubgroups
= true;
1241 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_FEATURES_EXT
: {
1242 VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT
*features
=
1243 (VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT
*)ext
;
1244 features
->texelBufferAlignment
= true;
1248 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_FEATURES_KHR
: {
1249 VkPhysicalDeviceTimelineSemaphoreFeaturesKHR
*features
=
1250 (VkPhysicalDeviceTimelineSemaphoreFeaturesKHR
*) ext
;
1251 features
->timelineSemaphore
= true;
1255 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES
: {
1256 VkPhysicalDeviceVariablePointersFeatures
*features
= (void *)ext
;
1257 features
->variablePointersStorageBuffer
= true;
1258 features
->variablePointers
= true;
1262 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT
: {
1263 VkPhysicalDeviceTransformFeedbackFeaturesEXT
*features
=
1264 (VkPhysicalDeviceTransformFeedbackFeaturesEXT
*)ext
;
1265 features
->transformFeedback
= true;
1266 features
->geometryStreams
= true;
1270 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_UNIFORM_BUFFER_STANDARD_LAYOUT_FEATURES_KHR
: {
1271 VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR
*features
=
1272 (VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR
*)ext
;
1273 features
->uniformBufferStandardLayout
= true;
1277 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT
: {
1278 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*features
=
1279 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*)ext
;
1280 features
->vertexAttributeInstanceRateDivisor
= true;
1281 features
->vertexAttributeInstanceRateZeroDivisor
= true;
1285 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_MEMORY_MODEL_FEATURES_KHR
: {
1286 VkPhysicalDeviceVulkanMemoryModelFeaturesKHR
*features
= (void *)ext
;
1287 features
->vulkanMemoryModel
= true;
1288 features
->vulkanMemoryModelDeviceScope
= true;
1289 features
->vulkanMemoryModelAvailabilityVisibilityChains
= true;
1293 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_YCBCR_IMAGE_ARRAYS_FEATURES_EXT
: {
1294 VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*features
=
1295 (VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*)ext
;
1296 features
->ycbcrImageArrays
= true;
1301 anv_debug_ignored_stype(ext
->sType
);
1307 #define MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS 64
1309 #define MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS 64
1310 #define MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS 256
1312 void anv_GetPhysicalDeviceProperties(
1313 VkPhysicalDevice physicalDevice
,
1314 VkPhysicalDeviceProperties
* pProperties
)
1316 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1317 const struct gen_device_info
*devinfo
= &pdevice
->info
;
1319 /* See assertions made when programming the buffer surface state. */
1320 const uint32_t max_raw_buffer_sz
= devinfo
->gen
>= 7 ?
1321 (1ul << 30) : (1ul << 27);
1323 const uint32_t max_ssbos
= pdevice
->has_a64_buffer_access
? UINT16_MAX
: 64;
1324 const uint32_t max_textures
=
1325 pdevice
->has_bindless_images
? UINT16_MAX
: 128;
1326 const uint32_t max_samplers
=
1327 pdevice
->has_bindless_samplers
? UINT16_MAX
:
1328 (devinfo
->gen
>= 8 || devinfo
->is_haswell
) ? 128 : 16;
1329 const uint32_t max_images
=
1330 pdevice
->has_bindless_images
? UINT16_MAX
: MAX_IMAGES
;
1332 /* If we can use bindless for everything, claim a high per-stage limit,
1333 * otherwise use the binding table size, minus the slots reserved for
1334 * render targets and one slot for the descriptor buffer. */
1335 const uint32_t max_per_stage
=
1336 pdevice
->has_bindless_images
&& pdevice
->has_a64_buffer_access
1337 ? UINT32_MAX
: MAX_BINDING_TABLE_SIZE
- MAX_RTS
- 1;
1339 const uint32_t max_workgroup_size
= 32 * devinfo
->max_cs_threads
;
1341 VkSampleCountFlags sample_counts
=
1342 isl_device_get_sample_counts(&pdevice
->isl_dev
);
1345 VkPhysicalDeviceLimits limits
= {
1346 .maxImageDimension1D
= (1 << 14),
1347 .maxImageDimension2D
= (1 << 14),
1348 .maxImageDimension3D
= (1 << 11),
1349 .maxImageDimensionCube
= (1 << 14),
1350 .maxImageArrayLayers
= (1 << 11),
1351 .maxTexelBufferElements
= 128 * 1024 * 1024,
1352 .maxUniformBufferRange
= (1ul << 27),
1353 .maxStorageBufferRange
= max_raw_buffer_sz
,
1354 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
1355 .maxMemoryAllocationCount
= UINT32_MAX
,
1356 .maxSamplerAllocationCount
= 64 * 1024,
1357 .bufferImageGranularity
= 64, /* A cache line */
1358 .sparseAddressSpaceSize
= 0,
1359 .maxBoundDescriptorSets
= MAX_SETS
,
1360 .maxPerStageDescriptorSamplers
= max_samplers
,
1361 .maxPerStageDescriptorUniformBuffers
= MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
,
1362 .maxPerStageDescriptorStorageBuffers
= max_ssbos
,
1363 .maxPerStageDescriptorSampledImages
= max_textures
,
1364 .maxPerStageDescriptorStorageImages
= max_images
,
1365 .maxPerStageDescriptorInputAttachments
= MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS
,
1366 .maxPerStageResources
= max_per_stage
,
1367 .maxDescriptorSetSamplers
= 6 * max_samplers
, /* number of stages * maxPerStageDescriptorSamplers */
1368 .maxDescriptorSetUniformBuffers
= 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
, /* number of stages * maxPerStageDescriptorUniformBuffers */
1369 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
1370 .maxDescriptorSetStorageBuffers
= 6 * max_ssbos
, /* number of stages * maxPerStageDescriptorStorageBuffers */
1371 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
1372 .maxDescriptorSetSampledImages
= 6 * max_textures
, /* number of stages * maxPerStageDescriptorSampledImages */
1373 .maxDescriptorSetStorageImages
= 6 * max_images
, /* number of stages * maxPerStageDescriptorStorageImages */
1374 .maxDescriptorSetInputAttachments
= MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS
,
1375 .maxVertexInputAttributes
= MAX_VBS
,
1376 .maxVertexInputBindings
= MAX_VBS
,
1377 .maxVertexInputAttributeOffset
= 2047,
1378 .maxVertexInputBindingStride
= 2048,
1379 .maxVertexOutputComponents
= 128,
1380 .maxTessellationGenerationLevel
= 64,
1381 .maxTessellationPatchSize
= 32,
1382 .maxTessellationControlPerVertexInputComponents
= 128,
1383 .maxTessellationControlPerVertexOutputComponents
= 128,
1384 .maxTessellationControlPerPatchOutputComponents
= 128,
1385 .maxTessellationControlTotalOutputComponents
= 2048,
1386 .maxTessellationEvaluationInputComponents
= 128,
1387 .maxTessellationEvaluationOutputComponents
= 128,
1388 .maxGeometryShaderInvocations
= 32,
1389 .maxGeometryInputComponents
= 64,
1390 .maxGeometryOutputComponents
= 128,
1391 .maxGeometryOutputVertices
= 256,
1392 .maxGeometryTotalOutputComponents
= 1024,
1393 .maxFragmentInputComponents
= 116, /* 128 components - (PSIZ, CLIP_DIST0, CLIP_DIST1) */
1394 .maxFragmentOutputAttachments
= 8,
1395 .maxFragmentDualSrcAttachments
= 1,
1396 .maxFragmentCombinedOutputResources
= 8,
1397 .maxComputeSharedMemorySize
= 64 * 1024,
1398 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
1399 .maxComputeWorkGroupInvocations
= max_workgroup_size
,
1400 .maxComputeWorkGroupSize
= {
1405 .subPixelPrecisionBits
= 8,
1406 .subTexelPrecisionBits
= 8,
1407 .mipmapPrecisionBits
= 8,
1408 .maxDrawIndexedIndexValue
= UINT32_MAX
,
1409 .maxDrawIndirectCount
= UINT32_MAX
,
1410 .maxSamplerLodBias
= 16,
1411 .maxSamplerAnisotropy
= 16,
1412 .maxViewports
= MAX_VIEWPORTS
,
1413 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
1414 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
1415 .viewportSubPixelBits
= 13, /* We take a float? */
1416 .minMemoryMapAlignment
= 4096, /* A page */
1417 /* The dataport requires texel alignment so we need to assume a worst
1418 * case of R32G32B32A32 which is 16 bytes.
1420 .minTexelBufferOffsetAlignment
= 16,
1421 /* We need 16 for UBO block reads to work and 32 for push UBOs */
1422 .minUniformBufferOffsetAlignment
= 32,
1423 .minStorageBufferOffsetAlignment
= 4,
1424 .minTexelOffset
= -8,
1425 .maxTexelOffset
= 7,
1426 .minTexelGatherOffset
= -32,
1427 .maxTexelGatherOffset
= 31,
1428 .minInterpolationOffset
= -0.5,
1429 .maxInterpolationOffset
= 0.4375,
1430 .subPixelInterpolationOffsetBits
= 4,
1431 .maxFramebufferWidth
= (1 << 14),
1432 .maxFramebufferHeight
= (1 << 14),
1433 .maxFramebufferLayers
= (1 << 11),
1434 .framebufferColorSampleCounts
= sample_counts
,
1435 .framebufferDepthSampleCounts
= sample_counts
,
1436 .framebufferStencilSampleCounts
= sample_counts
,
1437 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
1438 .maxColorAttachments
= MAX_RTS
,
1439 .sampledImageColorSampleCounts
= sample_counts
,
1440 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
1441 .sampledImageDepthSampleCounts
= sample_counts
,
1442 .sampledImageStencilSampleCounts
= sample_counts
,
1443 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
1444 .maxSampleMaskWords
= 1,
1445 .timestampComputeAndGraphics
= true,
1446 .timestampPeriod
= 1000000000.0 / devinfo
->timestamp_frequency
,
1447 .maxClipDistances
= 8,
1448 .maxCullDistances
= 8,
1449 .maxCombinedClipAndCullDistances
= 8,
1450 .discreteQueuePriorities
= 2,
1451 .pointSizeRange
= { 0.125, 255.875 },
1454 (devinfo
->gen
>= 9 || devinfo
->is_cherryview
) ?
1455 2047.9921875 : 7.9921875,
1457 .pointSizeGranularity
= (1.0 / 8.0),
1458 .lineWidthGranularity
= (1.0 / 128.0),
1459 .strictLines
= false,
1460 .standardSampleLocations
= true,
1461 .optimalBufferCopyOffsetAlignment
= 128,
1462 .optimalBufferCopyRowPitchAlignment
= 128,
1463 .nonCoherentAtomSize
= 64,
1466 *pProperties
= (VkPhysicalDeviceProperties
) {
1467 .apiVersion
= anv_physical_device_api_version(pdevice
),
1468 .driverVersion
= vk_get_driver_version(),
1470 .deviceID
= pdevice
->chipset_id
,
1471 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
1473 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
1476 snprintf(pProperties
->deviceName
, sizeof(pProperties
->deviceName
),
1477 "%s", pdevice
->name
);
1478 memcpy(pProperties
->pipelineCacheUUID
,
1479 pdevice
->pipeline_cache_uuid
, VK_UUID_SIZE
);
1482 void anv_GetPhysicalDeviceProperties2(
1483 VkPhysicalDevice physicalDevice
,
1484 VkPhysicalDeviceProperties2
* pProperties
)
1486 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1488 anv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
1490 vk_foreach_struct(ext
, pProperties
->pNext
) {
1491 switch (ext
->sType
) {
1492 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_STENCIL_RESOLVE_PROPERTIES_KHR
: {
1493 VkPhysicalDeviceDepthStencilResolvePropertiesKHR
*props
=
1494 (VkPhysicalDeviceDepthStencilResolvePropertiesKHR
*)ext
;
1496 /* We support all of the depth resolve modes */
1497 props
->supportedDepthResolveModes
=
1498 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
|
1499 VK_RESOLVE_MODE_AVERAGE_BIT_KHR
|
1500 VK_RESOLVE_MODE_MIN_BIT_KHR
|
1501 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1503 /* Average doesn't make sense for stencil so we don't support that */
1504 props
->supportedStencilResolveModes
=
1505 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
;
1506 if (pdevice
->info
.gen
>= 8) {
1507 /* The advanced stencil resolve modes currently require stencil
1508 * sampling be supported by the hardware.
1510 props
->supportedStencilResolveModes
|=
1511 VK_RESOLVE_MODE_MIN_BIT_KHR
|
1512 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1515 props
->independentResolveNone
= true;
1516 props
->independentResolve
= true;
1520 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES_EXT
: {
1521 VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*props
=
1522 (VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*)ext
;
1524 /* It's a bit hard to exactly map our implementation to the limits
1525 * described here. The bindless surface handle in the extended
1526 * message descriptors is 20 bits and it's an index into the table of
1527 * RENDER_SURFACE_STATE structs that starts at bindless surface base
1528 * address. Given that most things consume two surface states per
1529 * view (general/sampled for textures and write-only/read-write for
1530 * images), we claim 2^19 things.
1532 * For SSBOs, we just use A64 messages so there is no real limit
1533 * there beyond the limit on the total size of a descriptor set.
1535 const unsigned max_bindless_views
= 1 << 19;
1537 props
->maxUpdateAfterBindDescriptorsInAllPools
= max_bindless_views
;
1538 props
->shaderUniformBufferArrayNonUniformIndexingNative
= false;
1539 props
->shaderSampledImageArrayNonUniformIndexingNative
= false;
1540 props
->shaderStorageBufferArrayNonUniformIndexingNative
= true;
1541 props
->shaderStorageImageArrayNonUniformIndexingNative
= false;
1542 props
->shaderInputAttachmentArrayNonUniformIndexingNative
= false;
1543 props
->robustBufferAccessUpdateAfterBind
= true;
1544 props
->quadDivergentImplicitLod
= false;
1545 props
->maxPerStageDescriptorUpdateAfterBindSamplers
= max_bindless_views
;
1546 props
->maxPerStageDescriptorUpdateAfterBindUniformBuffers
= MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
;
1547 props
->maxPerStageDescriptorUpdateAfterBindStorageBuffers
= UINT32_MAX
;
1548 props
->maxPerStageDescriptorUpdateAfterBindSampledImages
= max_bindless_views
;
1549 props
->maxPerStageDescriptorUpdateAfterBindStorageImages
= max_bindless_views
;
1550 props
->maxPerStageDescriptorUpdateAfterBindInputAttachments
= MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS
;
1551 props
->maxPerStageUpdateAfterBindResources
= UINT32_MAX
;
1552 props
->maxDescriptorSetUpdateAfterBindSamplers
= max_bindless_views
;
1553 props
->maxDescriptorSetUpdateAfterBindUniformBuffers
= 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
;
1554 props
->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2;
1555 props
->maxDescriptorSetUpdateAfterBindStorageBuffers
= UINT32_MAX
;
1556 props
->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2;
1557 props
->maxDescriptorSetUpdateAfterBindSampledImages
= max_bindless_views
;
1558 props
->maxDescriptorSetUpdateAfterBindStorageImages
= max_bindless_views
;
1559 props
->maxDescriptorSetUpdateAfterBindInputAttachments
= MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS
;
1563 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR
: {
1564 VkPhysicalDeviceDriverPropertiesKHR
*driver_props
=
1565 (VkPhysicalDeviceDriverPropertiesKHR
*) ext
;
1567 driver_props
->driverID
= VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA_KHR
;
1568 snprintf(driver_props
->driverName
, VK_MAX_DRIVER_NAME_SIZE_KHR
,
1569 "Intel open-source Mesa driver");
1571 snprintf(driver_props
->driverInfo
, VK_MAX_DRIVER_INFO_SIZE_KHR
,
1572 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1
);
1574 driver_props
->conformanceVersion
= (VkConformanceVersionKHR
) {
1583 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT
: {
1584 VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*props
=
1585 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*) ext
;
1586 /* Userptr needs page aligned memory. */
1587 props
->minImportedHostPointerAlignment
= 4096;
1591 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES
: {
1592 VkPhysicalDeviceIDProperties
*id_props
=
1593 (VkPhysicalDeviceIDProperties
*)ext
;
1594 memcpy(id_props
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
1595 memcpy(id_props
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
1596 /* The LUID is for Windows. */
1597 id_props
->deviceLUIDValid
= false;
1601 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_PROPERTIES_EXT
: {
1602 VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*props
=
1603 (VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*)ext
;
1604 props
->maxInlineUniformBlockSize
= MAX_INLINE_UNIFORM_BLOCK_SIZE
;
1605 props
->maxPerStageDescriptorInlineUniformBlocks
=
1606 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1607 props
->maxPerStageDescriptorUpdateAfterBindInlineUniformBlocks
=
1608 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1609 props
->maxDescriptorSetInlineUniformBlocks
=
1610 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1611 props
->maxDescriptorSetUpdateAfterBindInlineUniformBlocks
=
1612 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1616 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_PROPERTIES_EXT
: {
1617 VkPhysicalDeviceLineRasterizationPropertiesEXT
*props
=
1618 (VkPhysicalDeviceLineRasterizationPropertiesEXT
*)ext
;
1619 /* In the Skylake PRM Vol. 7, subsection titled "GIQ (Diamond)
1620 * Sampling Rules - Legacy Mode", it says the following:
1622 * "Note that the device divides a pixel into a 16x16 array of
1623 * subpixels, referenced by their upper left corners."
1625 * This is the only known reference in the PRMs to the subpixel
1626 * precision of line rasterization and a "16x16 array of subpixels"
1627 * implies 4 subpixel precision bits. Empirical testing has shown
1628 * that 4 subpixel precision bits applies to all line rasterization
1631 props
->lineSubPixelPrecisionBits
= 4;
1635 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES
: {
1636 VkPhysicalDeviceMaintenance3Properties
*props
=
1637 (VkPhysicalDeviceMaintenance3Properties
*)ext
;
1638 /* This value doesn't matter for us today as our per-stage
1639 * descriptors are the real limit.
1641 props
->maxPerSetDescriptors
= 1024;
1642 props
->maxMemoryAllocationSize
= MAX_MEMORY_ALLOCATION_SIZE
;
1646 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES
: {
1647 VkPhysicalDeviceMultiviewProperties
*properties
=
1648 (VkPhysicalDeviceMultiviewProperties
*)ext
;
1649 properties
->maxMultiviewViewCount
= 16;
1650 properties
->maxMultiviewInstanceIndex
= UINT32_MAX
/ 16;
1654 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT
: {
1655 VkPhysicalDevicePCIBusInfoPropertiesEXT
*properties
=
1656 (VkPhysicalDevicePCIBusInfoPropertiesEXT
*)ext
;
1657 properties
->pciDomain
= pdevice
->pci_info
.domain
;
1658 properties
->pciBus
= pdevice
->pci_info
.bus
;
1659 properties
->pciDevice
= pdevice
->pci_info
.device
;
1660 properties
->pciFunction
= pdevice
->pci_info
.function
;
1664 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES
: {
1665 VkPhysicalDevicePointClippingProperties
*properties
=
1666 (VkPhysicalDevicePointClippingProperties
*) ext
;
1667 properties
->pointClippingBehavior
= VK_POINT_CLIPPING_BEHAVIOR_USER_CLIP_PLANES_ONLY
;
1671 #pragma GCC diagnostic push
1672 #pragma GCC diagnostic ignored "-Wswitch"
1673 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PRESENTATION_PROPERTIES_ANDROID
: {
1674 VkPhysicalDevicePresentationPropertiesANDROID
*props
=
1675 (VkPhysicalDevicePresentationPropertiesANDROID
*)ext
;
1676 props
->sharedImage
= VK_FALSE
;
1679 #pragma GCC diagnostic pop
1681 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES
: {
1682 VkPhysicalDeviceProtectedMemoryProperties
*props
=
1683 (VkPhysicalDeviceProtectedMemoryProperties
*)ext
;
1684 props
->protectedNoFault
= false;
1688 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
1689 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
1690 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
1692 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
1696 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT
: {
1697 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*properties
=
1698 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*)ext
;
1699 properties
->filterMinmaxImageComponentMapping
= pdevice
->info
.gen
>= 9;
1700 properties
->filterMinmaxSingleComponentFormats
= true;
1704 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES
: {
1705 VkPhysicalDeviceSubgroupProperties
*properties
= (void *)ext
;
1707 properties
->subgroupSize
= BRW_SUBGROUP_SIZE
;
1709 VkShaderStageFlags scalar_stages
= 0;
1710 for (unsigned stage
= 0; stage
< MESA_SHADER_STAGES
; stage
++) {
1711 if (pdevice
->compiler
->scalar_stage
[stage
])
1712 scalar_stages
|= mesa_to_vk_shader_stage(stage
);
1714 properties
->supportedStages
= scalar_stages
;
1716 properties
->supportedOperations
= VK_SUBGROUP_FEATURE_BASIC_BIT
|
1717 VK_SUBGROUP_FEATURE_VOTE_BIT
|
1718 VK_SUBGROUP_FEATURE_BALLOT_BIT
|
1719 VK_SUBGROUP_FEATURE_SHUFFLE_BIT
|
1720 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT
|
1721 VK_SUBGROUP_FEATURE_QUAD_BIT
;
1722 if (pdevice
->info
.gen
>= 8) {
1723 /* TODO: There's no technical reason why these can't be made to
1724 * work on gen7 but they don't at the moment so it's best to leave
1725 * the feature disabled than enabled and broken.
1727 properties
->supportedOperations
|=
1728 VK_SUBGROUP_FEATURE_ARITHMETIC_BIT
|
1729 VK_SUBGROUP_FEATURE_CLUSTERED_BIT
;
1731 properties
->quadOperationsInAllStages
= pdevice
->info
.gen
>= 8;
1735 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_PROPERTIES_EXT
: {
1736 VkPhysicalDeviceSubgroupSizeControlPropertiesEXT
*props
=
1737 (VkPhysicalDeviceSubgroupSizeControlPropertiesEXT
*)ext
;
1738 STATIC_ASSERT(8 <= BRW_SUBGROUP_SIZE
&& BRW_SUBGROUP_SIZE
<= 32);
1739 props
->minSubgroupSize
= 8;
1740 props
->maxSubgroupSize
= 32;
1741 props
->maxComputeWorkgroupSubgroups
= pdevice
->info
.max_cs_threads
;
1742 props
->requiredSubgroupSizeStages
= VK_SHADER_STAGE_COMPUTE_BIT
;
1745 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT_CONTROLS_PROPERTIES_KHR
: {
1746 VkPhysicalDeviceFloatControlsPropertiesKHR
*properties
= (void *)ext
;
1747 properties
->denormBehaviorIndependence
= VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_ALL_KHR
;
1748 properties
->roundingModeIndependence
= VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_NONE_KHR
;
1750 /* Broadwell does not support HF denorms and there are restrictions
1751 * other gens. According to Kabylake's PRM:
1753 * "math - Extended Math Function
1755 * Restriction : Half-float denorms are always retained."
1757 properties
->shaderDenormFlushToZeroFloat16
= false;
1758 properties
->shaderDenormPreserveFloat16
= pdevice
->info
.gen
> 8;
1759 properties
->shaderRoundingModeRTEFloat16
= true;
1760 properties
->shaderRoundingModeRTZFloat16
= true;
1761 properties
->shaderSignedZeroInfNanPreserveFloat16
= true;
1763 properties
->shaderDenormFlushToZeroFloat32
= true;
1764 properties
->shaderDenormPreserveFloat32
= true;
1765 properties
->shaderRoundingModeRTEFloat32
= true;
1766 properties
->shaderRoundingModeRTZFloat32
= true;
1767 properties
->shaderSignedZeroInfNanPreserveFloat32
= true;
1769 properties
->shaderDenormFlushToZeroFloat64
= true;
1770 properties
->shaderDenormPreserveFloat64
= true;
1771 properties
->shaderRoundingModeRTEFloat64
= true;
1772 properties
->shaderRoundingModeRTZFloat64
= true;
1773 properties
->shaderSignedZeroInfNanPreserveFloat64
= true;
1777 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_PROPERTIES_EXT
: {
1778 VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT
*props
=
1779 (VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT
*)ext
;
1781 /* From the SKL PRM Vol. 2d, docs for RENDER_SURFACE_STATE::Surface
1784 * "For SURFTYPE_BUFFER non-rendertarget surfaces, this field
1785 * specifies the base address of the first element of the surface,
1786 * computed in software by adding the surface base address to the
1787 * byte offset of the element in the buffer. The base address must
1788 * be aligned to element size."
1790 * The typed dataport messages require that things be texel aligned.
1791 * Otherwise, we may just load/store the wrong data or, in the worst
1792 * case, there may be hangs.
1794 props
->storageTexelBufferOffsetAlignmentBytes
= 16;
1795 props
->storageTexelBufferOffsetSingleTexelAlignment
= true;
1797 /* The sampler, however, is much more forgiving and it can handle
1798 * arbitrary byte alignment for linear and buffer surfaces. It's
1799 * hard to find a good PRM citation for this but years of empirical
1800 * experience demonstrate that this is true.
1802 props
->uniformTexelBufferOffsetAlignmentBytes
= 1;
1803 props
->uniformTexelBufferOffsetSingleTexelAlignment
= false;
1807 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_PROPERTIES_KHR
: {
1808 VkPhysicalDeviceTimelineSemaphorePropertiesKHR
*props
=
1809 (VkPhysicalDeviceTimelineSemaphorePropertiesKHR
*) ext
;
1810 props
->maxTimelineSemaphoreValueDifference
= UINT64_MAX
;
1814 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT
: {
1815 VkPhysicalDeviceTransformFeedbackPropertiesEXT
*props
=
1816 (VkPhysicalDeviceTransformFeedbackPropertiesEXT
*)ext
;
1818 props
->maxTransformFeedbackStreams
= MAX_XFB_STREAMS
;
1819 props
->maxTransformFeedbackBuffers
= MAX_XFB_BUFFERS
;
1820 props
->maxTransformFeedbackBufferSize
= (1ull << 32);
1821 props
->maxTransformFeedbackStreamDataSize
= 128 * 4;
1822 props
->maxTransformFeedbackBufferDataSize
= 128 * 4;
1823 props
->maxTransformFeedbackBufferDataStride
= 2048;
1824 props
->transformFeedbackQueries
= true;
1825 props
->transformFeedbackStreamsLinesTriangles
= false;
1826 props
->transformFeedbackRasterizationStreamSelect
= false;
1827 props
->transformFeedbackDraw
= true;
1831 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT
: {
1832 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*props
=
1833 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*)ext
;
1834 /* We have to restrict this a bit for multiview */
1835 props
->maxVertexAttribDivisor
= UINT32_MAX
/ 16;
1840 anv_debug_ignored_stype(ext
->sType
);
1846 /* We support exactly one queue family. */
1847 static const VkQueueFamilyProperties
1848 anv_queue_family_properties
= {
1849 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
1850 VK_QUEUE_COMPUTE_BIT
|
1851 VK_QUEUE_TRANSFER_BIT
,
1853 .timestampValidBits
= 36, /* XXX: Real value here */
1854 .minImageTransferGranularity
= { 1, 1, 1 },
1857 void anv_GetPhysicalDeviceQueueFamilyProperties(
1858 VkPhysicalDevice physicalDevice
,
1860 VkQueueFamilyProperties
* pQueueFamilyProperties
)
1862 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pCount
);
1864 vk_outarray_append(&out
, p
) {
1865 *p
= anv_queue_family_properties
;
1869 void anv_GetPhysicalDeviceQueueFamilyProperties2(
1870 VkPhysicalDevice physicalDevice
,
1871 uint32_t* pQueueFamilyPropertyCount
,
1872 VkQueueFamilyProperties2
* pQueueFamilyProperties
)
1875 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pQueueFamilyPropertyCount
);
1877 vk_outarray_append(&out
, p
) {
1878 p
->queueFamilyProperties
= anv_queue_family_properties
;
1880 vk_foreach_struct(s
, p
->pNext
) {
1881 anv_debug_ignored_stype(s
->sType
);
1886 void anv_GetPhysicalDeviceMemoryProperties(
1887 VkPhysicalDevice physicalDevice
,
1888 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
1890 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
1892 pMemoryProperties
->memoryTypeCount
= physical_device
->memory
.type_count
;
1893 for (uint32_t i
= 0; i
< physical_device
->memory
.type_count
; i
++) {
1894 pMemoryProperties
->memoryTypes
[i
] = (VkMemoryType
) {
1895 .propertyFlags
= physical_device
->memory
.types
[i
].propertyFlags
,
1896 .heapIndex
= physical_device
->memory
.types
[i
].heapIndex
,
1900 pMemoryProperties
->memoryHeapCount
= physical_device
->memory
.heap_count
;
1901 for (uint32_t i
= 0; i
< physical_device
->memory
.heap_count
; i
++) {
1902 pMemoryProperties
->memoryHeaps
[i
] = (VkMemoryHeap
) {
1903 .size
= physical_device
->memory
.heaps
[i
].size
,
1904 .flags
= physical_device
->memory
.heaps
[i
].flags
,
1910 anv_get_memory_budget(VkPhysicalDevice physicalDevice
,
1911 VkPhysicalDeviceMemoryBudgetPropertiesEXT
*memoryBudget
)
1913 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
1914 uint64_t sys_available
= get_available_system_memory();
1915 assert(sys_available
> 0);
1917 VkDeviceSize total_heaps_size
= 0;
1918 for (size_t i
= 0; i
< device
->memory
.heap_count
; i
++)
1919 total_heaps_size
+= device
->memory
.heaps
[i
].size
;
1921 for (size_t i
= 0; i
< device
->memory
.heap_count
; i
++) {
1922 VkDeviceSize heap_size
= device
->memory
.heaps
[i
].size
;
1923 VkDeviceSize heap_used
= device
->memory
.heaps
[i
].used
;
1924 VkDeviceSize heap_budget
;
1926 double heap_proportion
= (double) heap_size
/ total_heaps_size
;
1927 VkDeviceSize sys_available_prop
= sys_available
* heap_proportion
;
1930 * Let's not incite the app to starve the system: report at most 90% of
1931 * available system memory.
1933 uint64_t heap_available
= sys_available_prop
* 9 / 10;
1934 heap_budget
= MIN2(heap_size
, heap_used
+ heap_available
);
1937 * Round down to the nearest MB
1939 heap_budget
&= ~((1ull << 20) - 1);
1942 * The heapBudget value must be non-zero for array elements less than
1943 * VkPhysicalDeviceMemoryProperties::memoryHeapCount. The heapBudget
1944 * value must be less than or equal to VkMemoryHeap::size for each heap.
1946 assert(0 < heap_budget
&& heap_budget
<= heap_size
);
1948 memoryBudget
->heapUsage
[i
] = heap_used
;
1949 memoryBudget
->heapBudget
[i
] = heap_budget
;
1952 /* The heapBudget and heapUsage values must be zero for array elements
1953 * greater than or equal to VkPhysicalDeviceMemoryProperties::memoryHeapCount
1955 for (uint32_t i
= device
->memory
.heap_count
; i
< VK_MAX_MEMORY_HEAPS
; i
++) {
1956 memoryBudget
->heapBudget
[i
] = 0;
1957 memoryBudget
->heapUsage
[i
] = 0;
1961 void anv_GetPhysicalDeviceMemoryProperties2(
1962 VkPhysicalDevice physicalDevice
,
1963 VkPhysicalDeviceMemoryProperties2
* pMemoryProperties
)
1965 anv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
1966 &pMemoryProperties
->memoryProperties
);
1968 vk_foreach_struct(ext
, pMemoryProperties
->pNext
) {
1969 switch (ext
->sType
) {
1970 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT
:
1971 anv_get_memory_budget(physicalDevice
, (void*)ext
);
1974 anv_debug_ignored_stype(ext
->sType
);
1981 anv_GetDeviceGroupPeerMemoryFeatures(
1984 uint32_t localDeviceIndex
,
1985 uint32_t remoteDeviceIndex
,
1986 VkPeerMemoryFeatureFlags
* pPeerMemoryFeatures
)
1988 assert(localDeviceIndex
== 0 && remoteDeviceIndex
== 0);
1989 *pPeerMemoryFeatures
= VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT
|
1990 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT
|
1991 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT
|
1992 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT
;
1995 PFN_vkVoidFunction
anv_GetInstanceProcAddr(
1996 VkInstance _instance
,
1999 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2001 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
2002 * when we have to return valid function pointers, NULL, or it's left
2003 * undefined. See the table for exact details.
2008 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
2009 if (strcmp(pName, "vk" #entrypoint) == 0) \
2010 return (PFN_vkVoidFunction)anv_##entrypoint
2012 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties
);
2013 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties
);
2014 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceVersion
);
2015 LOOKUP_ANV_ENTRYPOINT(CreateInstance
);
2017 #undef LOOKUP_ANV_ENTRYPOINT
2019 if (instance
== NULL
)
2022 int idx
= anv_get_instance_entrypoint_index(pName
);
2024 return instance
->dispatch
.entrypoints
[idx
];
2026 idx
= anv_get_physical_device_entrypoint_index(pName
);
2028 return instance
->physicalDevice
.dispatch
.entrypoints
[idx
];
2030 idx
= anv_get_device_entrypoint_index(pName
);
2032 return instance
->device_dispatch
.entrypoints
[idx
];
2037 /* With version 1+ of the loader interface the ICD should expose
2038 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
2041 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
2042 VkInstance instance
,
2046 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
2047 VkInstance instance
,
2050 return anv_GetInstanceProcAddr(instance
, pName
);
2053 PFN_vkVoidFunction
anv_GetDeviceProcAddr(
2057 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2059 if (!device
|| !pName
)
2062 int idx
= anv_get_device_entrypoint_index(pName
);
2066 return device
->dispatch
.entrypoints
[idx
];
2069 /* With version 4+ of the loader interface the ICD should expose
2070 * vk_icdGetPhysicalDeviceProcAddr()
2073 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetPhysicalDeviceProcAddr(
2074 VkInstance _instance
,
2077 PFN_vkVoidFunction
vk_icdGetPhysicalDeviceProcAddr(
2078 VkInstance _instance
,
2081 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2083 if (!pName
|| !instance
)
2086 int idx
= anv_get_physical_device_entrypoint_index(pName
);
2090 return instance
->physicalDevice
.dispatch
.entrypoints
[idx
];
2095 anv_CreateDebugReportCallbackEXT(VkInstance _instance
,
2096 const VkDebugReportCallbackCreateInfoEXT
* pCreateInfo
,
2097 const VkAllocationCallbacks
* pAllocator
,
2098 VkDebugReportCallbackEXT
* pCallback
)
2100 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2101 return vk_create_debug_report_callback(&instance
->debug_report_callbacks
,
2102 pCreateInfo
, pAllocator
, &instance
->alloc
,
2107 anv_DestroyDebugReportCallbackEXT(VkInstance _instance
,
2108 VkDebugReportCallbackEXT _callback
,
2109 const VkAllocationCallbacks
* pAllocator
)
2111 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2112 vk_destroy_debug_report_callback(&instance
->debug_report_callbacks
,
2113 _callback
, pAllocator
, &instance
->alloc
);
2117 anv_DebugReportMessageEXT(VkInstance _instance
,
2118 VkDebugReportFlagsEXT flags
,
2119 VkDebugReportObjectTypeEXT objectType
,
2122 int32_t messageCode
,
2123 const char* pLayerPrefix
,
2124 const char* pMessage
)
2126 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2127 vk_debug_report(&instance
->debug_report_callbacks
, flags
, objectType
,
2128 object
, location
, messageCode
, pLayerPrefix
, pMessage
);
2131 static struct anv_state
2132 anv_state_pool_emit_data(struct anv_state_pool
*pool
, size_t size
, size_t align
, const void *p
)
2134 struct anv_state state
;
2136 state
= anv_state_pool_alloc(pool
, size
, align
);
2137 memcpy(state
.map
, p
, size
);
2142 /* Haswell border color is a bit of a disaster. Float and unorm formats use a
2143 * straightforward 32-bit float color in the first 64 bytes. Instead of using
2144 * a nice float/integer union like Gen8+, Haswell specifies the integer border
2145 * color as a separate entry /after/ the float color. The layout of this entry
2146 * also depends on the format's bpp (with extra hacks for RG32), and overlaps.
2148 * Since we don't know the format/bpp, we can't make any of the border colors
2149 * containing '1' work for all formats, as it would be in the wrong place for
2150 * some of them. We opt to make 32-bit integers work as this seems like the
2151 * most common option. Fortunately, transparent black works regardless, as
2152 * all zeroes is the same in every bit-size.
2154 struct hsw_border_color
{
2158 uint32_t _pad1
[108];
2161 struct gen8_border_color
{
2166 /* Pad out to 64 bytes */
2171 anv_device_init_border_colors(struct anv_device
*device
)
2173 if (device
->info
.is_haswell
) {
2174 static const struct hsw_border_color border_colors
[] = {
2175 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
2176 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
2177 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
2178 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
2179 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
2180 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
2183 device
->border_colors
=
2184 anv_state_pool_emit_data(&device
->dynamic_state_pool
,
2185 sizeof(border_colors
), 512, border_colors
);
2187 static const struct gen8_border_color border_colors
[] = {
2188 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
2189 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
2190 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
2191 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
2192 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
2193 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
2196 device
->border_colors
=
2197 anv_state_pool_emit_data(&device
->dynamic_state_pool
,
2198 sizeof(border_colors
), 64, border_colors
);
2203 anv_device_init_trivial_batch(struct anv_device
*device
)
2205 VkResult result
= anv_device_alloc_bo(device
, 4096,
2206 ANV_BO_ALLOC_MAPPED
,
2207 &device
->trivial_batch_bo
);
2208 if (result
!= VK_SUCCESS
)
2211 struct anv_batch batch
= {
2212 .start
= device
->trivial_batch_bo
->map
,
2213 .next
= device
->trivial_batch_bo
->map
,
2214 .end
= device
->trivial_batch_bo
->map
+ 4096,
2217 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
2218 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
2220 if (!device
->info
.has_llc
)
2221 gen_clflush_range(batch
.start
, batch
.next
- batch
.start
);
2226 VkResult
anv_EnumerateDeviceExtensionProperties(
2227 VkPhysicalDevice physicalDevice
,
2228 const char* pLayerName
,
2229 uint32_t* pPropertyCount
,
2230 VkExtensionProperties
* pProperties
)
2232 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
2233 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
2235 for (int i
= 0; i
< ANV_DEVICE_EXTENSION_COUNT
; i
++) {
2236 if (device
->supported_extensions
.extensions
[i
]) {
2237 vk_outarray_append(&out
, prop
) {
2238 *prop
= anv_device_extensions
[i
];
2243 return vk_outarray_status(&out
);
2247 anv_device_init_dispatch(struct anv_device
*device
)
2249 const struct anv_device_dispatch_table
*genX_table
;
2250 switch (device
->info
.gen
) {
2252 genX_table
= &gen12_device_dispatch_table
;
2255 genX_table
= &gen11_device_dispatch_table
;
2258 genX_table
= &gen10_device_dispatch_table
;
2261 genX_table
= &gen9_device_dispatch_table
;
2264 genX_table
= &gen8_device_dispatch_table
;
2267 if (device
->info
.is_haswell
)
2268 genX_table
= &gen75_device_dispatch_table
;
2270 genX_table
= &gen7_device_dispatch_table
;
2273 unreachable("unsupported gen\n");
2276 for (unsigned i
= 0; i
< ARRAY_SIZE(device
->dispatch
.entrypoints
); i
++) {
2277 /* Vulkan requires that entrypoints for extensions which have not been
2278 * enabled must not be advertised.
2280 if (!anv_device_entrypoint_is_enabled(i
, device
->instance
->app_info
.api_version
,
2281 &device
->instance
->enabled_extensions
,
2282 &device
->enabled_extensions
)) {
2283 device
->dispatch
.entrypoints
[i
] = NULL
;
2284 } else if (genX_table
->entrypoints
[i
]) {
2285 device
->dispatch
.entrypoints
[i
] = genX_table
->entrypoints
[i
];
2287 device
->dispatch
.entrypoints
[i
] =
2288 anv_device_dispatch_table
.entrypoints
[i
];
2294 vk_priority_to_gen(int priority
)
2297 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT
:
2298 return GEN_CONTEXT_LOW_PRIORITY
;
2299 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
:
2300 return GEN_CONTEXT_MEDIUM_PRIORITY
;
2301 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT
:
2302 return GEN_CONTEXT_HIGH_PRIORITY
;
2303 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT
:
2304 return GEN_CONTEXT_REALTIME_PRIORITY
;
2306 unreachable("Invalid priority");
2311 anv_device_init_hiz_clear_value_bo(struct anv_device
*device
)
2313 VkResult result
= anv_device_alloc_bo(device
, 4096,
2314 ANV_BO_ALLOC_MAPPED
,
2315 &device
->hiz_clear_bo
);
2316 if (result
!= VK_SUCCESS
)
2319 union isl_color_value hiz_clear
= { .u32
= { 0, } };
2320 hiz_clear
.f32
[0] = ANV_HZ_FC_VAL
;
2322 memcpy(device
->hiz_clear_bo
->map
, hiz_clear
.u32
, sizeof(hiz_clear
.u32
));
2324 if (!device
->info
.has_llc
)
2325 gen_clflush_range(device
->hiz_clear_bo
->map
, sizeof(hiz_clear
.u32
));
2331 get_bo_from_pool(struct gen_batch_decode_bo
*ret
,
2332 struct anv_block_pool
*pool
,
2335 anv_block_pool_foreach_bo(bo
, pool
) {
2336 uint64_t bo_address
= gen_48b_address(bo
->offset
);
2337 if (address
>= bo_address
&& address
< (bo_address
+ bo
->size
)) {
2338 *ret
= (struct gen_batch_decode_bo
) {
2349 /* Finding a buffer for batch decoding */
2350 static struct gen_batch_decode_bo
2351 decode_get_bo(void *v_batch
, bool ppgtt
, uint64_t address
)
2353 struct anv_device
*device
= v_batch
;
2354 struct gen_batch_decode_bo ret_bo
= {};
2358 if (get_bo_from_pool(&ret_bo
, &device
->dynamic_state_pool
.block_pool
, address
))
2360 if (get_bo_from_pool(&ret_bo
, &device
->instruction_state_pool
.block_pool
, address
))
2362 if (get_bo_from_pool(&ret_bo
, &device
->binding_table_pool
.block_pool
, address
))
2364 if (get_bo_from_pool(&ret_bo
, &device
->surface_state_pool
.block_pool
, address
))
2367 if (!device
->cmd_buffer_being_decoded
)
2368 return (struct gen_batch_decode_bo
) { };
2370 struct anv_batch_bo
**bo
;
2372 u_vector_foreach(bo
, &device
->cmd_buffer_being_decoded
->seen_bbos
) {
2373 /* The decoder zeroes out the top 16 bits, so we need to as well */
2374 uint64_t bo_address
= (*bo
)->bo
->offset
& (~0ull >> 16);
2376 if (address
>= bo_address
&& address
< bo_address
+ (*bo
)->bo
->size
) {
2377 return (struct gen_batch_decode_bo
) {
2379 .size
= (*bo
)->bo
->size
,
2380 .map
= (*bo
)->bo
->map
,
2385 return (struct gen_batch_decode_bo
) { };
2388 struct gen_aux_map_buffer
{
2389 struct gen_buffer base
;
2390 struct anv_state state
;
2393 static struct gen_buffer
*
2394 gen_aux_map_buffer_alloc(void *driver_ctx
, uint32_t size
)
2396 struct gen_aux_map_buffer
*buf
= malloc(sizeof(struct gen_aux_map_buffer
));
2400 struct anv_device
*device
= (struct anv_device
*)driver_ctx
;
2401 assert(device
->instance
->physicalDevice
.supports_48bit_addresses
&&
2402 device
->instance
->physicalDevice
.use_softpin
);
2404 struct anv_state_pool
*pool
= &device
->dynamic_state_pool
;
2405 buf
->state
= anv_state_pool_alloc(pool
, size
, size
);
2407 buf
->base
.gpu
= pool
->block_pool
.bo
->offset
+ buf
->state
.offset
;
2408 buf
->base
.gpu_end
= buf
->base
.gpu
+ buf
->state
.alloc_size
;
2409 buf
->base
.map
= buf
->state
.map
;
2410 buf
->base
.driver_bo
= &buf
->state
;
2415 gen_aux_map_buffer_free(void *driver_ctx
, struct gen_buffer
*buffer
)
2417 struct gen_aux_map_buffer
*buf
= (struct gen_aux_map_buffer
*)buffer
;
2418 struct anv_device
*device
= (struct anv_device
*)driver_ctx
;
2419 struct anv_state_pool
*pool
= &device
->dynamic_state_pool
;
2420 anv_state_pool_free(pool
, buf
->state
);
2424 static struct gen_mapped_pinned_buffer_alloc aux_map_allocator
= {
2425 .alloc
= gen_aux_map_buffer_alloc
,
2426 .free
= gen_aux_map_buffer_free
,
2429 VkResult
anv_CreateDevice(
2430 VkPhysicalDevice physicalDevice
,
2431 const VkDeviceCreateInfo
* pCreateInfo
,
2432 const VkAllocationCallbacks
* pAllocator
,
2435 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
2437 struct anv_device
*device
;
2439 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO
);
2441 struct anv_device_extension_table enabled_extensions
= { };
2442 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
2444 for (idx
= 0; idx
< ANV_DEVICE_EXTENSION_COUNT
; idx
++) {
2445 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
2446 anv_device_extensions
[idx
].extensionName
) == 0)
2450 if (idx
>= ANV_DEVICE_EXTENSION_COUNT
)
2451 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
2453 if (!physical_device
->supported_extensions
.extensions
[idx
])
2454 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
2456 enabled_extensions
.extensions
[idx
] = true;
2459 /* Check enabled features */
2460 if (pCreateInfo
->pEnabledFeatures
) {
2461 VkPhysicalDeviceFeatures supported_features
;
2462 anv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
2463 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
2464 VkBool32
*enabled_feature
= (VkBool32
*)pCreateInfo
->pEnabledFeatures
;
2465 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
2466 for (uint32_t i
= 0; i
< num_features
; i
++) {
2467 if (enabled_feature
[i
] && !supported_feature
[i
])
2468 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
2472 /* Check requested queues and fail if we are requested to create any
2473 * queues with flags we don't support.
2475 assert(pCreateInfo
->queueCreateInfoCount
> 0);
2476 for (uint32_t i
= 0; i
< pCreateInfo
->queueCreateInfoCount
; i
++) {
2477 if (pCreateInfo
->pQueueCreateInfos
[i
].flags
!= 0)
2478 return vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2481 /* Check if client specified queue priority. */
2482 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*queue_priority
=
2483 vk_find_struct_const(pCreateInfo
->pQueueCreateInfos
[0].pNext
,
2484 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT
);
2486 VkQueueGlobalPriorityEXT priority
=
2487 queue_priority
? queue_priority
->globalPriority
:
2488 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
;
2490 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
2492 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
2494 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2496 if (INTEL_DEBUG
& DEBUG_BATCH
) {
2497 const unsigned decode_flags
=
2498 GEN_BATCH_DECODE_FULL
|
2499 ((INTEL_DEBUG
& DEBUG_COLOR
) ? GEN_BATCH_DECODE_IN_COLOR
: 0) |
2500 GEN_BATCH_DECODE_OFFSETS
|
2501 GEN_BATCH_DECODE_FLOATS
;
2503 gen_batch_decode_ctx_init(&device
->decoder_ctx
,
2504 &physical_device
->info
,
2505 stderr
, decode_flags
, NULL
,
2506 decode_get_bo
, NULL
, device
);
2509 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
2510 device
->instance
= physical_device
->instance
;
2511 device
->chipset_id
= physical_device
->chipset_id
;
2512 device
->no_hw
= physical_device
->no_hw
;
2513 device
->_lost
= false;
2516 device
->alloc
= *pAllocator
;
2518 device
->alloc
= physical_device
->instance
->alloc
;
2520 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
2521 device
->fd
= open(physical_device
->path
, O_RDWR
| O_CLOEXEC
);
2522 if (device
->fd
== -1) {
2523 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2527 device
->context_id
= anv_gem_create_context(device
);
2528 if (device
->context_id
== -1) {
2529 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2533 result
= anv_queue_init(device
, &device
->queue
);
2534 if (result
!= VK_SUCCESS
)
2535 goto fail_context_id
;
2537 if (physical_device
->use_softpin
) {
2538 if (pthread_mutex_init(&device
->vma_mutex
, NULL
) != 0) {
2539 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2543 /* keep the page with address zero out of the allocator */
2544 struct anv_memory_heap
*low_heap
=
2545 &physical_device
->memory
.heaps
[physical_device
->memory
.heap_count
- 1];
2546 util_vma_heap_init(&device
->vma_lo
, low_heap
->vma_start
, low_heap
->vma_size
);
2547 device
->vma_lo_available
= low_heap
->size
;
2549 struct anv_memory_heap
*high_heap
=
2550 &physical_device
->memory
.heaps
[0];
2551 util_vma_heap_init(&device
->vma_hi
, high_heap
->vma_start
, high_heap
->vma_size
);
2552 device
->vma_hi_available
= physical_device
->memory
.heap_count
== 1 ? 0 :
2556 list_inithead(&device
->memory_objects
);
2558 /* As per spec, the driver implementation may deny requests to acquire
2559 * a priority above the default priority (MEDIUM) if the caller does not
2560 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
2563 if (physical_device
->has_context_priority
) {
2564 int err
= anv_gem_set_context_param(device
->fd
, device
->context_id
,
2565 I915_CONTEXT_PARAM_PRIORITY
,
2566 vk_priority_to_gen(priority
));
2567 if (err
!= 0 && priority
> VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
) {
2568 result
= vk_error(VK_ERROR_NOT_PERMITTED_EXT
);
2573 device
->info
= physical_device
->info
;
2574 device
->isl_dev
= physical_device
->isl_dev
;
2576 /* On Broadwell and later, we can use batch chaining to more efficiently
2577 * implement growing command buffers. Prior to Haswell, the kernel
2578 * command parser gets in the way and we have to fall back to growing
2581 device
->can_chain_batches
= device
->info
.gen
>= 8;
2583 device
->robust_buffer_access
= pCreateInfo
->pEnabledFeatures
&&
2584 pCreateInfo
->pEnabledFeatures
->robustBufferAccess
;
2585 device
->enabled_extensions
= enabled_extensions
;
2587 anv_device_init_dispatch(device
);
2589 if (pthread_mutex_init(&device
->mutex
, NULL
) != 0) {
2590 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2594 pthread_condattr_t condattr
;
2595 if (pthread_condattr_init(&condattr
) != 0) {
2596 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2599 if (pthread_condattr_setclock(&condattr
, CLOCK_MONOTONIC
) != 0) {
2600 pthread_condattr_destroy(&condattr
);
2601 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2604 if (pthread_cond_init(&device
->queue_submit
, &condattr
) != 0) {
2605 pthread_condattr_destroy(&condattr
);
2606 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2609 pthread_condattr_destroy(&condattr
);
2612 (physical_device
->supports_48bit_addresses
? EXEC_OBJECT_SUPPORTS_48B_ADDRESS
: 0) |
2613 (physical_device
->has_exec_async
? EXEC_OBJECT_ASYNC
: 0) |
2614 (physical_device
->has_exec_capture
? EXEC_OBJECT_CAPTURE
: 0) |
2615 (physical_device
->use_softpin
? EXEC_OBJECT_PINNED
: 0);
2617 result
= anv_bo_cache_init(&device
->bo_cache
);
2618 if (result
!= VK_SUCCESS
)
2619 goto fail_queue_cond
;
2621 anv_bo_pool_init(&device
->batch_bo_pool
, device
, bo_flags
);
2623 result
= anv_state_pool_init(&device
->dynamic_state_pool
, device
,
2624 DYNAMIC_STATE_POOL_MIN_ADDRESS
, 16384);
2625 if (result
!= VK_SUCCESS
)
2626 goto fail_batch_bo_pool
;
2628 result
= anv_state_pool_init(&device
->instruction_state_pool
, device
,
2629 INSTRUCTION_STATE_POOL_MIN_ADDRESS
, 16384);
2630 if (result
!= VK_SUCCESS
)
2631 goto fail_dynamic_state_pool
;
2633 result
= anv_state_pool_init(&device
->surface_state_pool
, device
,
2634 SURFACE_STATE_POOL_MIN_ADDRESS
, 4096);
2635 if (result
!= VK_SUCCESS
)
2636 goto fail_instruction_state_pool
;
2638 if (physical_device
->use_softpin
) {
2639 result
= anv_state_pool_init(&device
->binding_table_pool
, device
,
2640 BINDING_TABLE_POOL_MIN_ADDRESS
, 4096);
2641 if (result
!= VK_SUCCESS
)
2642 goto fail_surface_state_pool
;
2645 if (device
->info
.gen
>= 12) {
2646 device
->aux_map_ctx
= gen_aux_map_init(device
, &aux_map_allocator
,
2647 &physical_device
->info
);
2648 if (!device
->aux_map_ctx
)
2649 goto fail_binding_table_pool
;
2652 result
= anv_device_alloc_bo(device
, 4096, 0, &device
->workaround_bo
);
2653 if (result
!= VK_SUCCESS
)
2654 goto fail_surface_aux_map_pool
;
2656 result
= anv_device_init_trivial_batch(device
);
2657 if (result
!= VK_SUCCESS
)
2658 goto fail_workaround_bo
;
2660 if (device
->info
.gen
>= 10) {
2661 result
= anv_device_init_hiz_clear_value_bo(device
);
2662 if (result
!= VK_SUCCESS
)
2663 goto fail_trivial_batch_bo
;
2666 anv_scratch_pool_init(device
, &device
->scratch_pool
);
2668 switch (device
->info
.gen
) {
2670 if (!device
->info
.is_haswell
)
2671 result
= gen7_init_device_state(device
);
2673 result
= gen75_init_device_state(device
);
2676 result
= gen8_init_device_state(device
);
2679 result
= gen9_init_device_state(device
);
2682 result
= gen10_init_device_state(device
);
2685 result
= gen11_init_device_state(device
);
2688 result
= gen12_init_device_state(device
);
2691 /* Shouldn't get here as we don't create physical devices for any other
2693 unreachable("unhandled gen");
2695 if (result
!= VK_SUCCESS
)
2696 goto fail_workaround_bo
;
2698 anv_pipeline_cache_init(&device
->default_pipeline_cache
, device
, true);
2700 anv_device_init_blorp(device
);
2702 anv_device_init_border_colors(device
);
2704 anv_device_perf_init(device
);
2706 *pDevice
= anv_device_to_handle(device
);
2711 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
2712 if (device
->info
.gen
>= 10)
2713 anv_device_release_bo(device
, device
->hiz_clear_bo
);
2714 anv_device_release_bo(device
, device
->workaround_bo
);
2715 fail_trivial_batch_bo
:
2716 anv_device_release_bo(device
, device
->trivial_batch_bo
);
2717 fail_surface_aux_map_pool
:
2718 if (device
->info
.gen
>= 12) {
2719 gen_aux_map_finish(device
->aux_map_ctx
);
2720 device
->aux_map_ctx
= NULL
;
2722 fail_binding_table_pool
:
2723 if (physical_device
->use_softpin
)
2724 anv_state_pool_finish(&device
->binding_table_pool
);
2725 fail_surface_state_pool
:
2726 anv_state_pool_finish(&device
->surface_state_pool
);
2727 fail_instruction_state_pool
:
2728 anv_state_pool_finish(&device
->instruction_state_pool
);
2729 fail_dynamic_state_pool
:
2730 anv_state_pool_finish(&device
->dynamic_state_pool
);
2732 anv_bo_pool_finish(&device
->batch_bo_pool
);
2733 anv_bo_cache_finish(&device
->bo_cache
);
2735 pthread_cond_destroy(&device
->queue_submit
);
2737 pthread_mutex_destroy(&device
->mutex
);
2739 if (physical_device
->use_softpin
) {
2740 util_vma_heap_finish(&device
->vma_hi
);
2741 util_vma_heap_finish(&device
->vma_lo
);
2744 anv_queue_finish(&device
->queue
);
2746 anv_gem_destroy_context(device
, device
->context_id
);
2750 vk_free(&device
->alloc
, device
);
2755 void anv_DestroyDevice(
2757 const VkAllocationCallbacks
* pAllocator
)
2759 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2760 struct anv_physical_device
*physical_device
;
2765 physical_device
= &device
->instance
->physicalDevice
;
2767 anv_device_finish_blorp(device
);
2769 anv_pipeline_cache_finish(&device
->default_pipeline_cache
);
2771 anv_queue_finish(&device
->queue
);
2773 #ifdef HAVE_VALGRIND
2774 /* We only need to free these to prevent valgrind errors. The backing
2775 * BO will go away in a couple of lines so we don't actually leak.
2777 anv_state_pool_free(&device
->dynamic_state_pool
, device
->border_colors
);
2778 anv_state_pool_free(&device
->dynamic_state_pool
, device
->slice_hash
);
2781 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
2783 anv_device_release_bo(device
, device
->workaround_bo
);
2784 anv_device_release_bo(device
, device
->trivial_batch_bo
);
2785 if (device
->info
.gen
>= 10)
2786 anv_device_release_bo(device
, device
->hiz_clear_bo
);
2788 if (device
->info
.gen
>= 12) {
2789 gen_aux_map_finish(device
->aux_map_ctx
);
2790 device
->aux_map_ctx
= NULL
;
2793 if (physical_device
->use_softpin
)
2794 anv_state_pool_finish(&device
->binding_table_pool
);
2795 anv_state_pool_finish(&device
->surface_state_pool
);
2796 anv_state_pool_finish(&device
->instruction_state_pool
);
2797 anv_state_pool_finish(&device
->dynamic_state_pool
);
2799 anv_bo_pool_finish(&device
->batch_bo_pool
);
2801 anv_bo_cache_finish(&device
->bo_cache
);
2803 if (physical_device
->use_softpin
) {
2804 util_vma_heap_finish(&device
->vma_hi
);
2805 util_vma_heap_finish(&device
->vma_lo
);
2808 pthread_cond_destroy(&device
->queue_submit
);
2809 pthread_mutex_destroy(&device
->mutex
);
2811 anv_gem_destroy_context(device
, device
->context_id
);
2813 if (INTEL_DEBUG
& DEBUG_BATCH
)
2814 gen_batch_decode_ctx_finish(&device
->decoder_ctx
);
2818 vk_free(&device
->alloc
, device
);
2821 VkResult
anv_EnumerateInstanceLayerProperties(
2822 uint32_t* pPropertyCount
,
2823 VkLayerProperties
* pProperties
)
2825 if (pProperties
== NULL
) {
2826 *pPropertyCount
= 0;
2830 /* None supported at this time */
2831 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
2834 VkResult
anv_EnumerateDeviceLayerProperties(
2835 VkPhysicalDevice physicalDevice
,
2836 uint32_t* pPropertyCount
,
2837 VkLayerProperties
* pProperties
)
2839 if (pProperties
== NULL
) {
2840 *pPropertyCount
= 0;
2844 /* None supported at this time */
2845 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
2848 void anv_GetDeviceQueue(
2850 uint32_t queueNodeIndex
,
2851 uint32_t queueIndex
,
2854 const VkDeviceQueueInfo2 info
= {
2855 .sType
= VK_STRUCTURE_TYPE_DEVICE_QUEUE_INFO_2
,
2858 .queueFamilyIndex
= queueNodeIndex
,
2859 .queueIndex
= queueIndex
,
2862 anv_GetDeviceQueue2(_device
, &info
, pQueue
);
2865 void anv_GetDeviceQueue2(
2867 const VkDeviceQueueInfo2
* pQueueInfo
,
2870 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2872 assert(pQueueInfo
->queueIndex
== 0);
2874 if (pQueueInfo
->flags
== device
->queue
.flags
)
2875 *pQueue
= anv_queue_to_handle(&device
->queue
);
2881 _anv_device_set_lost(struct anv_device
*device
,
2882 const char *file
, int line
,
2883 const char *msg
, ...)
2888 p_atomic_inc(&device
->_lost
);
2891 err
= __vk_errorv(device
->instance
, device
,
2892 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT
,
2893 VK_ERROR_DEVICE_LOST
, file
, line
, msg
, ap
);
2896 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
2903 _anv_queue_set_lost(struct anv_queue
*queue
,
2904 const char *file
, int line
,
2905 const char *msg
, ...)
2910 p_atomic_inc(&queue
->device
->_lost
);
2913 err
= __vk_errorv(queue
->device
->instance
, queue
->device
,
2914 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT
,
2915 VK_ERROR_DEVICE_LOST
, file
, line
, msg
, ap
);
2918 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
2925 anv_device_query_status(struct anv_device
*device
)
2927 /* This isn't likely as most of the callers of this function already check
2928 * for it. However, it doesn't hurt to check and it potentially lets us
2931 if (anv_device_is_lost(device
))
2932 return VK_ERROR_DEVICE_LOST
;
2934 uint32_t active
, pending
;
2935 int ret
= anv_gem_gpu_get_reset_stats(device
, &active
, &pending
);
2937 /* We don't know the real error. */
2938 return anv_device_set_lost(device
, "get_reset_stats failed: %m");
2942 return anv_device_set_lost(device
, "GPU hung on one of our command buffers");
2943 } else if (pending
) {
2944 return anv_device_set_lost(device
, "GPU hung with commands in-flight");
2951 anv_device_bo_busy(struct anv_device
*device
, struct anv_bo
*bo
)
2953 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
2954 * Other usages of the BO (such as on different hardware) will not be
2955 * flagged as "busy" by this ioctl. Use with care.
2957 int ret
= anv_gem_busy(device
, bo
->gem_handle
);
2959 return VK_NOT_READY
;
2960 } else if (ret
== -1) {
2961 /* We don't know the real error. */
2962 return anv_device_set_lost(device
, "gem wait failed: %m");
2965 /* Query for device status after the busy call. If the BO we're checking
2966 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
2967 * client because it clearly doesn't have valid data. Yes, this most
2968 * likely means an ioctl, but we just did an ioctl to query the busy status
2969 * so it's no great loss.
2971 return anv_device_query_status(device
);
2975 anv_device_wait(struct anv_device
*device
, struct anv_bo
*bo
,
2978 int ret
= anv_gem_wait(device
, bo
->gem_handle
, &timeout
);
2979 if (ret
== -1 && errno
== ETIME
) {
2981 } else if (ret
== -1) {
2982 /* We don't know the real error. */
2983 return anv_device_set_lost(device
, "gem wait failed: %m");
2986 /* Query for device status after the wait. If the BO we're waiting on got
2987 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
2988 * because it clearly doesn't have valid data. Yes, this most likely means
2989 * an ioctl, but we just did an ioctl to wait so it's no great loss.
2991 return anv_device_query_status(device
);
2994 VkResult
anv_DeviceWaitIdle(
2997 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2999 if (anv_device_is_lost(device
))
3000 return VK_ERROR_DEVICE_LOST
;
3002 return anv_queue_submit_simple_batch(&device
->queue
, NULL
);
3006 anv_vma_alloc(struct anv_device
*device
, struct anv_bo
*bo
)
3008 if (!(bo
->flags
& EXEC_OBJECT_PINNED
))
3011 pthread_mutex_lock(&device
->vma_mutex
);
3015 if (bo
->flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
&&
3016 device
->vma_hi_available
>= bo
->size
) {
3017 uint64_t addr
= util_vma_heap_alloc(&device
->vma_hi
, bo
->size
, 4096);
3019 bo
->offset
= gen_canonical_address(addr
);
3020 assert(addr
== gen_48b_address(bo
->offset
));
3021 device
->vma_hi_available
-= bo
->size
;
3025 if (bo
->offset
== 0 && device
->vma_lo_available
>= bo
->size
) {
3026 uint64_t addr
= util_vma_heap_alloc(&device
->vma_lo
, bo
->size
, 4096);
3028 bo
->offset
= gen_canonical_address(addr
);
3029 assert(addr
== gen_48b_address(bo
->offset
));
3030 device
->vma_lo_available
-= bo
->size
;
3034 pthread_mutex_unlock(&device
->vma_mutex
);
3036 return bo
->offset
!= 0;
3040 anv_vma_free(struct anv_device
*device
, struct anv_bo
*bo
)
3042 if (!(bo
->flags
& EXEC_OBJECT_PINNED
))
3045 const uint64_t addr_48b
= gen_48b_address(bo
->offset
);
3047 pthread_mutex_lock(&device
->vma_mutex
);
3049 if (addr_48b
>= LOW_HEAP_MIN_ADDRESS
&&
3050 addr_48b
<= LOW_HEAP_MAX_ADDRESS
) {
3051 util_vma_heap_free(&device
->vma_lo
, addr_48b
, bo
->size
);
3052 device
->vma_lo_available
+= bo
->size
;
3054 ASSERTED
const struct anv_physical_device
*physical_device
=
3055 &device
->instance
->physicalDevice
;
3056 assert(addr_48b
>= physical_device
->memory
.heaps
[0].vma_start
&&
3057 addr_48b
< (physical_device
->memory
.heaps
[0].vma_start
+
3058 physical_device
->memory
.heaps
[0].vma_size
));
3059 util_vma_heap_free(&device
->vma_hi
, addr_48b
, bo
->size
);
3060 device
->vma_hi_available
+= bo
->size
;
3063 pthread_mutex_unlock(&device
->vma_mutex
);
3068 VkResult
anv_AllocateMemory(
3070 const VkMemoryAllocateInfo
* pAllocateInfo
,
3071 const VkAllocationCallbacks
* pAllocator
,
3072 VkDeviceMemory
* pMem
)
3074 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3075 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3076 struct anv_device_memory
*mem
;
3077 VkResult result
= VK_SUCCESS
;
3079 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
3081 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
3082 assert(pAllocateInfo
->allocationSize
> 0);
3084 if (pAllocateInfo
->allocationSize
> MAX_MEMORY_ALLOCATION_SIZE
)
3085 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
3087 /* FINISHME: Fail if allocation request exceeds heap size. */
3089 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
3090 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3092 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3094 assert(pAllocateInfo
->memoryTypeIndex
< pdevice
->memory
.type_count
);
3095 mem
->type
= &pdevice
->memory
.types
[pAllocateInfo
->memoryTypeIndex
];
3099 mem
->host_ptr
= NULL
;
3101 enum anv_bo_alloc_flags alloc_flags
= 0;
3103 assert(mem
->type
->heapIndex
< pdevice
->memory
.heap_count
);
3104 if (!pdevice
->memory
.heaps
[mem
->type
->heapIndex
].supports_48bit_addresses
)
3105 alloc_flags
|= ANV_BO_ALLOC_32BIT_ADDRESS
;
3107 const struct wsi_memory_allocate_info
*wsi_info
=
3108 vk_find_struct_const(pAllocateInfo
->pNext
, WSI_MEMORY_ALLOCATE_INFO_MESA
);
3109 if (wsi_info
&& wsi_info
->implicit_sync
) {
3110 /* We need to set the WRITE flag on window system buffers so that GEM
3111 * will know we're writing to them and synchronize uses on other rings
3112 * (eg if the display server uses the blitter ring).
3114 alloc_flags
|= ANV_BO_ALLOC_IMPLICIT_SYNC
|
3115 ANV_BO_ALLOC_IMPLICIT_WRITE
;
3118 const VkExportMemoryAllocateInfo
*export_info
=
3119 vk_find_struct_const(pAllocateInfo
->pNext
, EXPORT_MEMORY_ALLOCATE_INFO
);
3121 /* Check if we need to support Android HW buffer export. If so,
3122 * create AHardwareBuffer and import memory from it.
3124 bool android_export
= false;
3125 if (export_info
&& export_info
->handleTypes
&
3126 VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID
)
3127 android_export
= true;
3129 /* Android memory import. */
3130 const struct VkImportAndroidHardwareBufferInfoANDROID
*ahw_import_info
=
3131 vk_find_struct_const(pAllocateInfo
->pNext
,
3132 IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID
);
3134 if (ahw_import_info
) {
3135 result
= anv_import_ahw_memory(_device
, mem
, ahw_import_info
);
3136 if (result
!= VK_SUCCESS
)
3140 } else if (android_export
) {
3141 result
= anv_create_ahw_memory(_device
, mem
, pAllocateInfo
);
3142 if (result
!= VK_SUCCESS
)
3145 const struct VkImportAndroidHardwareBufferInfoANDROID import_info
= {
3148 result
= anv_import_ahw_memory(_device
, mem
, &import_info
);
3149 if (result
!= VK_SUCCESS
)
3155 const VkImportMemoryFdInfoKHR
*fd_info
=
3156 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_FD_INFO_KHR
);
3158 /* The Vulkan spec permits handleType to be 0, in which case the struct is
3161 if (fd_info
&& fd_info
->handleType
) {
3162 /* At the moment, we support only the below handle types. */
3163 assert(fd_info
->handleType
==
3164 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
3165 fd_info
->handleType
==
3166 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
3168 result
= anv_device_import_bo(device
, fd_info
->fd
, alloc_flags
,
3170 if (result
!= VK_SUCCESS
)
3173 VkDeviceSize aligned_alloc_size
=
3174 align_u64(pAllocateInfo
->allocationSize
, 4096);
3176 /* For security purposes, we reject importing the bo if it's smaller
3177 * than the requested allocation size. This prevents a malicious client
3178 * from passing a buffer to a trusted client, lying about the size, and
3179 * telling the trusted client to try and texture from an image that goes
3180 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
3181 * in the trusted client. The trusted client can protect itself against
3182 * this sort of attack but only if it can trust the buffer size.
3184 if (mem
->bo
->size
< aligned_alloc_size
) {
3185 result
= vk_errorf(device
->instance
, device
,
3186 VK_ERROR_INVALID_EXTERNAL_HANDLE
,
3187 "aligned allocationSize too large for "
3188 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT: "
3189 "%"PRIu64
"B > %"PRIu64
"B",
3190 aligned_alloc_size
, mem
->bo
->size
);
3191 anv_device_release_bo(device
, mem
->bo
);
3195 /* From the Vulkan spec:
3197 * "Importing memory from a file descriptor transfers ownership of
3198 * the file descriptor from the application to the Vulkan
3199 * implementation. The application must not perform any operations on
3200 * the file descriptor after a successful import."
3202 * If the import fails, we leave the file descriptor open.
3208 const VkImportMemoryHostPointerInfoEXT
*host_ptr_info
=
3209 vk_find_struct_const(pAllocateInfo
->pNext
,
3210 IMPORT_MEMORY_HOST_POINTER_INFO_EXT
);
3211 if (host_ptr_info
&& host_ptr_info
->handleType
) {
3212 if (host_ptr_info
->handleType
==
3213 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_MAPPED_FOREIGN_MEMORY_BIT_EXT
) {
3214 result
= vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
3218 assert(host_ptr_info
->handleType
==
3219 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
);
3221 result
= anv_device_import_bo_from_host_ptr(device
,
3222 host_ptr_info
->pHostPointer
,
3223 pAllocateInfo
->allocationSize
,
3227 if (result
!= VK_SUCCESS
)
3230 mem
->host_ptr
= host_ptr_info
->pHostPointer
;
3234 /* Regular allocate (not importing memory). */
3236 if (export_info
&& export_info
->handleTypes
)
3237 alloc_flags
|= ANV_BO_ALLOC_EXTERNAL
;
3239 result
= anv_device_alloc_bo(device
, pAllocateInfo
->allocationSize
,
3240 alloc_flags
, &mem
->bo
);
3241 if (result
!= VK_SUCCESS
)
3244 const VkMemoryDedicatedAllocateInfo
*dedicated_info
=
3245 vk_find_struct_const(pAllocateInfo
->pNext
, MEMORY_DEDICATED_ALLOCATE_INFO
);
3246 if (dedicated_info
&& dedicated_info
->image
!= VK_NULL_HANDLE
) {
3247 ANV_FROM_HANDLE(anv_image
, image
, dedicated_info
->image
);
3249 /* Some legacy (non-modifiers) consumers need the tiling to be set on
3250 * the BO. In this case, we have a dedicated allocation.
3252 if (image
->needs_set_tiling
) {
3253 const uint32_t i915_tiling
=
3254 isl_tiling_to_i915_tiling(image
->planes
[0].surface
.isl
.tiling
);
3255 int ret
= anv_gem_set_tiling(device
, mem
->bo
->gem_handle
,
3256 image
->planes
[0].surface
.isl
.row_pitch_B
,
3259 anv_device_release_bo(device
, mem
->bo
);
3260 return vk_errorf(device
->instance
, NULL
,
3261 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
3262 "failed to set BO tiling: %m");
3268 pthread_mutex_lock(&device
->mutex
);
3269 list_addtail(&mem
->link
, &device
->memory_objects
);
3270 pthread_mutex_unlock(&device
->mutex
);
3272 *pMem
= anv_device_memory_to_handle(mem
);
3274 p_atomic_add(&pdevice
->memory
.heaps
[mem
->type
->heapIndex
].used
,
3280 vk_free2(&device
->alloc
, pAllocator
, mem
);
3285 VkResult
anv_GetMemoryFdKHR(
3287 const VkMemoryGetFdInfoKHR
* pGetFdInfo
,
3290 ANV_FROM_HANDLE(anv_device
, dev
, device_h
);
3291 ANV_FROM_HANDLE(anv_device_memory
, mem
, pGetFdInfo
->memory
);
3293 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
3295 assert(pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
3296 pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
3298 return anv_device_export_bo(dev
, mem
->bo
, pFd
);
3301 VkResult
anv_GetMemoryFdPropertiesKHR(
3303 VkExternalMemoryHandleTypeFlagBits handleType
,
3305 VkMemoryFdPropertiesKHR
* pMemoryFdProperties
)
3307 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3308 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3310 switch (handleType
) {
3311 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
:
3312 /* dma-buf can be imported as any memory type */
3313 pMemoryFdProperties
->memoryTypeBits
=
3314 (1 << pdevice
->memory
.type_count
) - 1;
3318 /* The valid usage section for this function says:
3320 * "handleType must not be one of the handle types defined as
3323 * So opaque handle types fall into the default "unsupported" case.
3325 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
3329 VkResult
anv_GetMemoryHostPointerPropertiesEXT(
3331 VkExternalMemoryHandleTypeFlagBits handleType
,
3332 const void* pHostPointer
,
3333 VkMemoryHostPointerPropertiesEXT
* pMemoryHostPointerProperties
)
3335 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3337 assert(pMemoryHostPointerProperties
->sType
==
3338 VK_STRUCTURE_TYPE_MEMORY_HOST_POINTER_PROPERTIES_EXT
);
3340 switch (handleType
) {
3341 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
: {
3342 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3344 /* Host memory can be imported as any memory type. */
3345 pMemoryHostPointerProperties
->memoryTypeBits
=
3346 (1ull << pdevice
->memory
.type_count
) - 1;
3351 return VK_ERROR_INVALID_EXTERNAL_HANDLE
;
3355 void anv_FreeMemory(
3357 VkDeviceMemory _mem
,
3358 const VkAllocationCallbacks
* pAllocator
)
3360 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3361 ANV_FROM_HANDLE(anv_device_memory
, mem
, _mem
);
3362 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3367 pthread_mutex_lock(&device
->mutex
);
3368 list_del(&mem
->link
);
3369 pthread_mutex_unlock(&device
->mutex
);
3372 anv_UnmapMemory(_device
, _mem
);
3374 p_atomic_add(&pdevice
->memory
.heaps
[mem
->type
->heapIndex
].used
,
3377 anv_device_release_bo(device
, mem
->bo
);
3379 #if defined(ANDROID) && ANDROID_API_LEVEL >= 26
3381 AHardwareBuffer_release(mem
->ahw
);
3384 vk_free2(&device
->alloc
, pAllocator
, mem
);
3387 VkResult
anv_MapMemory(
3389 VkDeviceMemory _memory
,
3390 VkDeviceSize offset
,
3392 VkMemoryMapFlags flags
,
3395 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3396 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
3403 if (mem
->host_ptr
) {
3404 *ppData
= mem
->host_ptr
+ offset
;
3408 if (size
== VK_WHOLE_SIZE
)
3409 size
= mem
->bo
->size
- offset
;
3411 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
3413 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
3414 * assert(size != 0);
3415 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
3416 * equal to the size of the memory minus offset
3419 assert(offset
+ size
<= mem
->bo
->size
);
3421 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
3422 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
3423 * at a time is valid. We could just mmap up front and return an offset
3424 * pointer here, but that may exhaust virtual memory on 32 bit
3427 uint32_t gem_flags
= 0;
3429 if (!device
->info
.has_llc
&&
3430 (mem
->type
->propertyFlags
& VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
))
3431 gem_flags
|= I915_MMAP_WC
;
3433 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
3434 uint64_t map_offset
= offset
& ~4095ull;
3435 assert(offset
>= map_offset
);
3436 uint64_t map_size
= (offset
+ size
) - map_offset
;
3438 /* Let's map whole pages */
3439 map_size
= align_u64(map_size
, 4096);
3441 void *map
= anv_gem_mmap(device
, mem
->bo
->gem_handle
,
3442 map_offset
, map_size
, gem_flags
);
3443 if (map
== MAP_FAILED
)
3444 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
3447 mem
->map_size
= map_size
;
3449 *ppData
= mem
->map
+ (offset
- map_offset
);
3454 void anv_UnmapMemory(
3456 VkDeviceMemory _memory
)
3458 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
3460 if (mem
== NULL
|| mem
->host_ptr
)
3463 anv_gem_munmap(mem
->map
, mem
->map_size
);
3470 clflush_mapped_ranges(struct anv_device
*device
,
3472 const VkMappedMemoryRange
*ranges
)
3474 for (uint32_t i
= 0; i
< count
; i
++) {
3475 ANV_FROM_HANDLE(anv_device_memory
, mem
, ranges
[i
].memory
);
3476 if (ranges
[i
].offset
>= mem
->map_size
)
3479 gen_clflush_range(mem
->map
+ ranges
[i
].offset
,
3480 MIN2(ranges
[i
].size
, mem
->map_size
- ranges
[i
].offset
));
3484 VkResult
anv_FlushMappedMemoryRanges(
3486 uint32_t memoryRangeCount
,
3487 const VkMappedMemoryRange
* pMemoryRanges
)
3489 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3491 if (device
->info
.has_llc
)
3494 /* Make sure the writes we're flushing have landed. */
3495 __builtin_ia32_mfence();
3497 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
3502 VkResult
anv_InvalidateMappedMemoryRanges(
3504 uint32_t memoryRangeCount
,
3505 const VkMappedMemoryRange
* pMemoryRanges
)
3507 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3509 if (device
->info
.has_llc
)
3512 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
3514 /* Make sure no reads get moved up above the invalidate. */
3515 __builtin_ia32_mfence();
3520 void anv_GetBufferMemoryRequirements(
3523 VkMemoryRequirements
* pMemoryRequirements
)
3525 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
3526 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3527 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3529 /* The Vulkan spec (git aaed022) says:
3531 * memoryTypeBits is a bitfield and contains one bit set for every
3532 * supported memory type for the resource. The bit `1<<i` is set if and
3533 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3534 * structure for the physical device is supported.
3536 uint32_t memory_types
= 0;
3537 for (uint32_t i
= 0; i
< pdevice
->memory
.type_count
; i
++) {
3538 uint32_t valid_usage
= pdevice
->memory
.types
[i
].valid_buffer_usage
;
3539 if ((valid_usage
& buffer
->usage
) == buffer
->usage
)
3540 memory_types
|= (1u << i
);
3543 /* Base alignment requirement of a cache line */
3544 uint32_t alignment
= 16;
3546 /* We need an alignment of 32 for pushing UBOs */
3547 if (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
)
3548 alignment
= MAX2(alignment
, 32);
3550 pMemoryRequirements
->size
= buffer
->size
;
3551 pMemoryRequirements
->alignment
= alignment
;
3553 /* Storage and Uniform buffers should have their size aligned to
3554 * 32-bits to avoid boundary checks when last DWord is not complete.
3555 * This would ensure that not internal padding would be needed for
3558 if (device
->robust_buffer_access
&&
3559 (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
||
3560 buffer
->usage
& VK_BUFFER_USAGE_STORAGE_BUFFER_BIT
))
3561 pMemoryRequirements
->size
= align_u64(buffer
->size
, 4);
3563 pMemoryRequirements
->memoryTypeBits
= memory_types
;
3566 void anv_GetBufferMemoryRequirements2(
3568 const VkBufferMemoryRequirementsInfo2
* pInfo
,
3569 VkMemoryRequirements2
* pMemoryRequirements
)
3571 anv_GetBufferMemoryRequirements(_device
, pInfo
->buffer
,
3572 &pMemoryRequirements
->memoryRequirements
);
3574 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3575 switch (ext
->sType
) {
3576 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
3577 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
3578 requirements
->prefersDedicatedAllocation
= false;
3579 requirements
->requiresDedicatedAllocation
= false;
3584 anv_debug_ignored_stype(ext
->sType
);
3590 void anv_GetImageMemoryRequirements(
3593 VkMemoryRequirements
* pMemoryRequirements
)
3595 ANV_FROM_HANDLE(anv_image
, image
, _image
);
3596 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3597 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3599 /* The Vulkan spec (git aaed022) says:
3601 * memoryTypeBits is a bitfield and contains one bit set for every
3602 * supported memory type for the resource. The bit `1<<i` is set if and
3603 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3604 * structure for the physical device is supported.
3606 * All types are currently supported for images.
3608 uint32_t memory_types
= (1ull << pdevice
->memory
.type_count
) - 1;
3610 /* We must have image allocated or imported at this point. According to the
3611 * specification, external images must have been bound to memory before
3612 * calling GetImageMemoryRequirements.
3614 assert(image
->size
> 0);
3616 pMemoryRequirements
->size
= image
->size
;
3617 pMemoryRequirements
->alignment
= image
->alignment
;
3618 pMemoryRequirements
->memoryTypeBits
= memory_types
;
3621 void anv_GetImageMemoryRequirements2(
3623 const VkImageMemoryRequirementsInfo2
* pInfo
,
3624 VkMemoryRequirements2
* pMemoryRequirements
)
3626 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3627 ANV_FROM_HANDLE(anv_image
, image
, pInfo
->image
);
3629 anv_GetImageMemoryRequirements(_device
, pInfo
->image
,
3630 &pMemoryRequirements
->memoryRequirements
);
3632 vk_foreach_struct_const(ext
, pInfo
->pNext
) {
3633 switch (ext
->sType
) {
3634 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO
: {
3635 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3636 const VkImagePlaneMemoryRequirementsInfo
*plane_reqs
=
3637 (const VkImagePlaneMemoryRequirementsInfo
*) ext
;
3638 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
,
3639 plane_reqs
->planeAspect
);
3641 assert(image
->planes
[plane
].offset
== 0);
3643 /* The Vulkan spec (git aaed022) says:
3645 * memoryTypeBits is a bitfield and contains one bit set for every
3646 * supported memory type for the resource. The bit `1<<i` is set
3647 * if and only if the memory type `i` in the
3648 * VkPhysicalDeviceMemoryProperties structure for the physical
3649 * device is supported.
3651 * All types are currently supported for images.
3653 pMemoryRequirements
->memoryRequirements
.memoryTypeBits
=
3654 (1ull << pdevice
->memory
.type_count
) - 1;
3656 /* We must have image allocated or imported at this point. According to the
3657 * specification, external images must have been bound to memory before
3658 * calling GetImageMemoryRequirements.
3660 assert(image
->planes
[plane
].size
> 0);
3662 pMemoryRequirements
->memoryRequirements
.size
= image
->planes
[plane
].size
;
3663 pMemoryRequirements
->memoryRequirements
.alignment
=
3664 image
->planes
[plane
].alignment
;
3669 anv_debug_ignored_stype(ext
->sType
);
3674 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3675 switch (ext
->sType
) {
3676 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
3677 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
3678 if (image
->needs_set_tiling
|| image
->external_format
) {
3679 /* If we need to set the tiling for external consumers, we need a
3680 * dedicated allocation.
3682 * See also anv_AllocateMemory.
3684 requirements
->prefersDedicatedAllocation
= true;
3685 requirements
->requiresDedicatedAllocation
= true;
3687 requirements
->prefersDedicatedAllocation
= false;
3688 requirements
->requiresDedicatedAllocation
= false;
3694 anv_debug_ignored_stype(ext
->sType
);
3700 void anv_GetImageSparseMemoryRequirements(
3703 uint32_t* pSparseMemoryRequirementCount
,
3704 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
3706 *pSparseMemoryRequirementCount
= 0;
3709 void anv_GetImageSparseMemoryRequirements2(
3711 const VkImageSparseMemoryRequirementsInfo2
* pInfo
,
3712 uint32_t* pSparseMemoryRequirementCount
,
3713 VkSparseImageMemoryRequirements2
* pSparseMemoryRequirements
)
3715 *pSparseMemoryRequirementCount
= 0;
3718 void anv_GetDeviceMemoryCommitment(
3720 VkDeviceMemory memory
,
3721 VkDeviceSize
* pCommittedMemoryInBytes
)
3723 *pCommittedMemoryInBytes
= 0;
3727 anv_bind_buffer_memory(const VkBindBufferMemoryInfo
*pBindInfo
)
3729 ANV_FROM_HANDLE(anv_device_memory
, mem
, pBindInfo
->memory
);
3730 ANV_FROM_HANDLE(anv_buffer
, buffer
, pBindInfo
->buffer
);
3732 assert(pBindInfo
->sType
== VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
);
3735 assert((buffer
->usage
& mem
->type
->valid_buffer_usage
) == buffer
->usage
);
3736 buffer
->address
= (struct anv_address
) {
3738 .offset
= pBindInfo
->memoryOffset
,
3741 buffer
->address
= ANV_NULL_ADDRESS
;
3745 VkResult
anv_BindBufferMemory(
3748 VkDeviceMemory memory
,
3749 VkDeviceSize memoryOffset
)
3751 anv_bind_buffer_memory(
3752 &(VkBindBufferMemoryInfo
) {
3753 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
,
3756 .memoryOffset
= memoryOffset
,
3762 VkResult
anv_BindBufferMemory2(
3764 uint32_t bindInfoCount
,
3765 const VkBindBufferMemoryInfo
* pBindInfos
)
3767 for (uint32_t i
= 0; i
< bindInfoCount
; i
++)
3768 anv_bind_buffer_memory(&pBindInfos
[i
]);
3773 VkResult
anv_QueueBindSparse(
3775 uint32_t bindInfoCount
,
3776 const VkBindSparseInfo
* pBindInfo
,
3779 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
3780 if (anv_device_is_lost(queue
->device
))
3781 return VK_ERROR_DEVICE_LOST
;
3783 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
3788 VkResult
anv_CreateEvent(
3790 const VkEventCreateInfo
* pCreateInfo
,
3791 const VkAllocationCallbacks
* pAllocator
,
3794 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3795 struct anv_state state
;
3796 struct anv_event
*event
;
3798 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_EVENT_CREATE_INFO
);
3800 state
= anv_state_pool_alloc(&device
->dynamic_state_pool
,
3803 event
->state
= state
;
3804 event
->semaphore
= VK_EVENT_RESET
;
3806 if (!device
->info
.has_llc
) {
3807 /* Make sure the writes we're flushing have landed. */
3808 __builtin_ia32_mfence();
3809 __builtin_ia32_clflush(event
);
3812 *pEvent
= anv_event_to_handle(event
);
3817 void anv_DestroyEvent(
3820 const VkAllocationCallbacks
* pAllocator
)
3822 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3823 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3828 anv_state_pool_free(&device
->dynamic_state_pool
, event
->state
);
3831 VkResult
anv_GetEventStatus(
3835 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3836 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3838 if (anv_device_is_lost(device
))
3839 return VK_ERROR_DEVICE_LOST
;
3841 if (!device
->info
.has_llc
) {
3842 /* Invalidate read cache before reading event written by GPU. */
3843 __builtin_ia32_clflush(event
);
3844 __builtin_ia32_mfence();
3848 return event
->semaphore
;
3851 VkResult
anv_SetEvent(
3855 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3856 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3858 event
->semaphore
= VK_EVENT_SET
;
3860 if (!device
->info
.has_llc
) {
3861 /* Make sure the writes we're flushing have landed. */
3862 __builtin_ia32_mfence();
3863 __builtin_ia32_clflush(event
);
3869 VkResult
anv_ResetEvent(
3873 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3874 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3876 event
->semaphore
= VK_EVENT_RESET
;
3878 if (!device
->info
.has_llc
) {
3879 /* Make sure the writes we're flushing have landed. */
3880 __builtin_ia32_mfence();
3881 __builtin_ia32_clflush(event
);
3889 VkResult
anv_CreateBuffer(
3891 const VkBufferCreateInfo
* pCreateInfo
,
3892 const VkAllocationCallbacks
* pAllocator
,
3895 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3896 struct anv_buffer
*buffer
;
3898 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
3900 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
3901 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3903 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3905 buffer
->size
= pCreateInfo
->size
;
3906 buffer
->usage
= pCreateInfo
->usage
;
3907 buffer
->address
= ANV_NULL_ADDRESS
;
3909 *pBuffer
= anv_buffer_to_handle(buffer
);
3914 void anv_DestroyBuffer(
3917 const VkAllocationCallbacks
* pAllocator
)
3919 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3920 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
3925 vk_free2(&device
->alloc
, pAllocator
, buffer
);
3928 VkDeviceAddress
anv_GetBufferDeviceAddressEXT(
3930 const VkBufferDeviceAddressInfoEXT
* pInfo
)
3932 ANV_FROM_HANDLE(anv_buffer
, buffer
, pInfo
->buffer
);
3934 assert(buffer
->address
.bo
->flags
& EXEC_OBJECT_PINNED
);
3936 return anv_address_physical(buffer
->address
);
3940 anv_fill_buffer_surface_state(struct anv_device
*device
, struct anv_state state
,
3941 enum isl_format format
,
3942 struct anv_address address
,
3943 uint32_t range
, uint32_t stride
)
3945 isl_buffer_fill_state(&device
->isl_dev
, state
.map
,
3946 .address
= anv_address_physical(address
),
3947 .mocs
= device
->default_mocs
,
3950 .swizzle
= ISL_SWIZZLE_IDENTITY
,
3951 .stride_B
= stride
);
3954 void anv_DestroySampler(
3957 const VkAllocationCallbacks
* pAllocator
)
3959 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3960 ANV_FROM_HANDLE(anv_sampler
, sampler
, _sampler
);
3965 if (sampler
->bindless_state
.map
) {
3966 anv_state_pool_free(&device
->dynamic_state_pool
,
3967 sampler
->bindless_state
);
3970 vk_free2(&device
->alloc
, pAllocator
, sampler
);
3973 VkResult
anv_CreateFramebuffer(
3975 const VkFramebufferCreateInfo
* pCreateInfo
,
3976 const VkAllocationCallbacks
* pAllocator
,
3977 VkFramebuffer
* pFramebuffer
)
3979 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3980 struct anv_framebuffer
*framebuffer
;
3982 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
3984 size_t size
= sizeof(*framebuffer
);
3986 /* VK_KHR_imageless_framebuffer extension says:
3988 * If flags includes VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR,
3989 * parameter pAttachments is ignored.
3991 if (!(pCreateInfo
->flags
& VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR
)) {
3992 size
+= sizeof(struct anv_image_view
*) * pCreateInfo
->attachmentCount
;
3993 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
3994 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3995 if (framebuffer
== NULL
)
3996 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3998 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
3999 ANV_FROM_HANDLE(anv_image_view
, iview
, pCreateInfo
->pAttachments
[i
]);
4000 framebuffer
->attachments
[i
] = iview
;
4002 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
4004 assert(device
->enabled_extensions
.KHR_imageless_framebuffer
);
4005 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
4006 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
4007 if (framebuffer
== NULL
)
4008 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
4010 framebuffer
->attachment_count
= 0;
4013 framebuffer
->width
= pCreateInfo
->width
;
4014 framebuffer
->height
= pCreateInfo
->height
;
4015 framebuffer
->layers
= pCreateInfo
->layers
;
4017 *pFramebuffer
= anv_framebuffer_to_handle(framebuffer
);
4022 void anv_DestroyFramebuffer(
4025 const VkAllocationCallbacks
* pAllocator
)
4027 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4028 ANV_FROM_HANDLE(anv_framebuffer
, fb
, _fb
);
4033 vk_free2(&device
->alloc
, pAllocator
, fb
);
4036 static const VkTimeDomainEXT anv_time_domains
[] = {
4037 VK_TIME_DOMAIN_DEVICE_EXT
,
4038 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
,
4039 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
,
4042 VkResult
anv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
4043 VkPhysicalDevice physicalDevice
,
4044 uint32_t *pTimeDomainCount
,
4045 VkTimeDomainEXT
*pTimeDomains
)
4048 VK_OUTARRAY_MAKE(out
, pTimeDomains
, pTimeDomainCount
);
4050 for (d
= 0; d
< ARRAY_SIZE(anv_time_domains
); d
++) {
4051 vk_outarray_append(&out
, i
) {
4052 *i
= anv_time_domains
[d
];
4056 return vk_outarray_status(&out
);
4060 anv_clock_gettime(clockid_t clock_id
)
4062 struct timespec current
;
4065 ret
= clock_gettime(clock_id
, ¤t
);
4066 if (ret
< 0 && clock_id
== CLOCK_MONOTONIC_RAW
)
4067 ret
= clock_gettime(CLOCK_MONOTONIC
, ¤t
);
4071 return (uint64_t) current
.tv_sec
* 1000000000ULL + current
.tv_nsec
;
4074 #define TIMESTAMP 0x2358
4076 VkResult
anv_GetCalibratedTimestampsEXT(
4078 uint32_t timestampCount
,
4079 const VkCalibratedTimestampInfoEXT
*pTimestampInfos
,
4080 uint64_t *pTimestamps
,
4081 uint64_t *pMaxDeviation
)
4083 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4084 uint64_t timestamp_frequency
= device
->info
.timestamp_frequency
;
4087 uint64_t begin
, end
;
4088 uint64_t max_clock_period
= 0;
4090 begin
= anv_clock_gettime(CLOCK_MONOTONIC_RAW
);
4092 for (d
= 0; d
< timestampCount
; d
++) {
4093 switch (pTimestampInfos
[d
].timeDomain
) {
4094 case VK_TIME_DOMAIN_DEVICE_EXT
:
4095 ret
= anv_gem_reg_read(device
, TIMESTAMP
| 1,
4099 return anv_device_set_lost(device
, "Failed to read the TIMESTAMP "
4102 uint64_t device_period
= DIV_ROUND_UP(1000000000, timestamp_frequency
);
4103 max_clock_period
= MAX2(max_clock_period
, device_period
);
4105 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
:
4106 pTimestamps
[d
] = anv_clock_gettime(CLOCK_MONOTONIC
);
4107 max_clock_period
= MAX2(max_clock_period
, 1);
4110 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
:
4111 pTimestamps
[d
] = begin
;
4119 end
= anv_clock_gettime(CLOCK_MONOTONIC_RAW
);
4122 * The maximum deviation is the sum of the interval over which we
4123 * perform the sampling and the maximum period of any sampled
4124 * clock. That's because the maximum skew between any two sampled
4125 * clock edges is when the sampled clock with the largest period is
4126 * sampled at the end of that period but right at the beginning of the
4127 * sampling interval and some other clock is sampled right at the
4128 * begining of its sampling period and right at the end of the
4129 * sampling interval. Let's assume the GPU has the longest clock
4130 * period and that the application is sampling GPU and monotonic:
4133 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
4134 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4138 * GPU -----_____-----_____-----_____-----_____
4141 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
4142 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4144 * Interval <----------------->
4145 * Deviation <-------------------------->
4149 * m = read(monotonic) 2
4152 * We round the sample interval up by one tick to cover sampling error
4153 * in the interval clock
4156 uint64_t sample_interval
= end
- begin
+ 1;
4158 *pMaxDeviation
= sample_interval
+ max_clock_period
;
4163 /* vk_icd.h does not declare this function, so we declare it here to
4164 * suppress Wmissing-prototypes.
4166 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4167 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
);
4169 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4170 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
)
4172 /* For the full details on loader interface versioning, see
4173 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
4174 * What follows is a condensed summary, to help you navigate the large and
4175 * confusing official doc.
4177 * - Loader interface v0 is incompatible with later versions. We don't
4180 * - In loader interface v1:
4181 * - The first ICD entrypoint called by the loader is
4182 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
4184 * - The ICD must statically expose no other Vulkan symbol unless it is
4185 * linked with -Bsymbolic.
4186 * - Each dispatchable Vulkan handle created by the ICD must be
4187 * a pointer to a struct whose first member is VK_LOADER_DATA. The
4188 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
4189 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
4190 * vkDestroySurfaceKHR(). The ICD must be capable of working with
4191 * such loader-managed surfaces.
4193 * - Loader interface v2 differs from v1 in:
4194 * - The first ICD entrypoint called by the loader is
4195 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
4196 * statically expose this entrypoint.
4198 * - Loader interface v3 differs from v2 in:
4199 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
4200 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
4201 * because the loader no longer does so.
4203 * - Loader interface v4 differs from v3 in:
4204 * - The ICD must implement vk_icdGetPhysicalDeviceProcAddr().
4206 *pSupportedVersion
= MIN2(*pSupportedVersion
, 4u);