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;
563 device
->compiler
->compact_params
= true;
565 /* Broadwell PRM says:
567 * "Before Gen8, there was a historical configuration control field to
568 * swizzle address bit[6] for in X/Y tiling modes. This was set in three
569 * different places: TILECTL[1:0], ARB_MODE[5:4], and
570 * DISP_ARB_CTL[14:13].
572 * For Gen8 and subsequent generations, the swizzle fields are all
573 * reserved, and the CPU's memory controller performs all address
574 * swizzling modifications."
577 device
->info
.gen
< 8 && anv_gem_get_bit6_swizzle(fd
, I915_TILING_X
);
579 isl_device_init(&device
->isl_dev
, &device
->info
, swizzled
);
581 result
= anv_physical_device_init_uuids(device
);
582 if (result
!= VK_SUCCESS
)
585 anv_physical_device_init_disk_cache(device
);
587 if (instance
->enabled_extensions
.KHR_display
) {
588 master_fd
= open(primary_path
, O_RDWR
| O_CLOEXEC
);
589 if (master_fd
>= 0) {
590 /* prod the device with a GETPARAM call which will fail if
591 * we don't have permission to even render on this device
593 if (anv_gem_get_param(master_fd
, I915_PARAM_CHIPSET_ID
) == 0) {
599 device
->master_fd
= master_fd
;
601 result
= anv_init_wsi(device
);
602 if (result
!= VK_SUCCESS
) {
603 ralloc_free(device
->compiler
);
604 anv_physical_device_free_disk_cache(device
);
608 device
->perf
= anv_get_perf(&device
->info
, fd
);
610 anv_physical_device_get_supported_extensions(device
,
611 &device
->supported_extensions
);
614 device
->local_fd
= fd
;
626 anv_physical_device_finish(struct anv_physical_device
*device
)
628 anv_finish_wsi(device
);
629 anv_physical_device_free_disk_cache(device
);
630 ralloc_free(device
->compiler
);
631 ralloc_free(device
->perf
);
632 close(device
->local_fd
);
633 if (device
->master_fd
>= 0)
634 close(device
->master_fd
);
638 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
639 VkSystemAllocationScope allocationScope
)
645 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
646 size_t align
, VkSystemAllocationScope allocationScope
)
648 return realloc(pOriginal
, size
);
652 default_free_func(void *pUserData
, void *pMemory
)
657 static const VkAllocationCallbacks default_alloc
= {
659 .pfnAllocation
= default_alloc_func
,
660 .pfnReallocation
= default_realloc_func
,
661 .pfnFree
= default_free_func
,
664 VkResult
anv_EnumerateInstanceExtensionProperties(
665 const char* pLayerName
,
666 uint32_t* pPropertyCount
,
667 VkExtensionProperties
* pProperties
)
669 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
671 for (int i
= 0; i
< ANV_INSTANCE_EXTENSION_COUNT
; i
++) {
672 if (anv_instance_extensions_supported
.extensions
[i
]) {
673 vk_outarray_append(&out
, prop
) {
674 *prop
= anv_instance_extensions
[i
];
679 return vk_outarray_status(&out
);
682 VkResult
anv_CreateInstance(
683 const VkInstanceCreateInfo
* pCreateInfo
,
684 const VkAllocationCallbacks
* pAllocator
,
685 VkInstance
* pInstance
)
687 struct anv_instance
*instance
;
690 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
692 struct anv_instance_extension_table enabled_extensions
= {};
693 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
695 for (idx
= 0; idx
< ANV_INSTANCE_EXTENSION_COUNT
; idx
++) {
696 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
697 anv_instance_extensions
[idx
].extensionName
) == 0)
701 if (idx
>= ANV_INSTANCE_EXTENSION_COUNT
)
702 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
704 if (!anv_instance_extensions_supported
.extensions
[idx
])
705 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
707 enabled_extensions
.extensions
[idx
] = true;
710 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
711 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
713 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
715 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
718 instance
->alloc
= *pAllocator
;
720 instance
->alloc
= default_alloc
;
722 instance
->app_info
= (struct anv_app_info
) { .api_version
= 0 };
723 if (pCreateInfo
->pApplicationInfo
) {
724 const VkApplicationInfo
*app
= pCreateInfo
->pApplicationInfo
;
726 instance
->app_info
.app_name
=
727 vk_strdup(&instance
->alloc
, app
->pApplicationName
,
728 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
729 instance
->app_info
.app_version
= app
->applicationVersion
;
731 instance
->app_info
.engine_name
=
732 vk_strdup(&instance
->alloc
, app
->pEngineName
,
733 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
734 instance
->app_info
.engine_version
= app
->engineVersion
;
736 instance
->app_info
.api_version
= app
->apiVersion
;
739 if (instance
->app_info
.api_version
== 0)
740 instance
->app_info
.api_version
= VK_API_VERSION_1_0
;
742 instance
->enabled_extensions
= enabled_extensions
;
744 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->dispatch
.entrypoints
); i
++) {
745 /* Vulkan requires that entrypoints for extensions which have not been
746 * enabled must not be advertised.
748 if (!anv_instance_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
749 &instance
->enabled_extensions
)) {
750 instance
->dispatch
.entrypoints
[i
] = NULL
;
752 instance
->dispatch
.entrypoints
[i
] =
753 anv_instance_dispatch_table
.entrypoints
[i
];
757 struct anv_physical_device
*pdevice
= &instance
->physicalDevice
;
758 for (unsigned i
= 0; i
< ARRAY_SIZE(pdevice
->dispatch
.entrypoints
); i
++) {
759 /* Vulkan requires that entrypoints for extensions which have not been
760 * enabled must not be advertised.
762 if (!anv_physical_device_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
763 &instance
->enabled_extensions
)) {
764 pdevice
->dispatch
.entrypoints
[i
] = NULL
;
766 pdevice
->dispatch
.entrypoints
[i
] =
767 anv_physical_device_dispatch_table
.entrypoints
[i
];
771 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->device_dispatch
.entrypoints
); i
++) {
772 /* Vulkan requires that entrypoints for extensions which have not been
773 * enabled must not be advertised.
775 if (!anv_device_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
776 &instance
->enabled_extensions
, NULL
)) {
777 instance
->device_dispatch
.entrypoints
[i
] = NULL
;
779 instance
->device_dispatch
.entrypoints
[i
] =
780 anv_device_dispatch_table
.entrypoints
[i
];
784 instance
->physicalDeviceCount
= -1;
786 result
= vk_debug_report_instance_init(&instance
->debug_report_callbacks
);
787 if (result
!= VK_SUCCESS
) {
788 vk_free2(&default_alloc
, pAllocator
, instance
);
789 return vk_error(result
);
792 instance
->pipeline_cache_enabled
=
793 env_var_as_boolean("ANV_ENABLE_PIPELINE_CACHE", true);
795 glsl_type_singleton_init_or_ref();
797 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
799 driParseOptionInfo(&instance
->available_dri_options
, anv_dri_options_xml
);
800 driParseConfigFiles(&instance
->dri_options
, &instance
->available_dri_options
,
802 instance
->app_info
.engine_name
,
803 instance
->app_info
.engine_version
);
805 *pInstance
= anv_instance_to_handle(instance
);
810 void anv_DestroyInstance(
811 VkInstance _instance
,
812 const VkAllocationCallbacks
* pAllocator
)
814 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
819 if (instance
->physicalDeviceCount
> 0) {
820 /* We support at most one physical device. */
821 assert(instance
->physicalDeviceCount
== 1);
822 anv_physical_device_finish(&instance
->physicalDevice
);
825 vk_free(&instance
->alloc
, (char *)instance
->app_info
.app_name
);
826 vk_free(&instance
->alloc
, (char *)instance
->app_info
.engine_name
);
828 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
830 vk_debug_report_instance_destroy(&instance
->debug_report_callbacks
);
832 glsl_type_singleton_decref();
834 driDestroyOptionCache(&instance
->dri_options
);
835 driDestroyOptionInfo(&instance
->available_dri_options
);
837 vk_free(&instance
->alloc
, instance
);
841 anv_enumerate_devices(struct anv_instance
*instance
)
843 /* TODO: Check for more devices ? */
844 drmDevicePtr devices
[8];
845 VkResult result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
848 instance
->physicalDeviceCount
= 0;
850 max_devices
= drmGetDevices2(0, devices
, ARRAY_SIZE(devices
));
852 return VK_ERROR_INCOMPATIBLE_DRIVER
;
854 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
855 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
856 devices
[i
]->bustype
== DRM_BUS_PCI
&&
857 devices
[i
]->deviceinfo
.pci
->vendor_id
== 0x8086) {
859 result
= anv_physical_device_init(&instance
->physicalDevice
,
860 instance
, devices
[i
]);
861 if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
865 drmFreeDevices(devices
, max_devices
);
867 if (result
== VK_SUCCESS
)
868 instance
->physicalDeviceCount
= 1;
874 anv_instance_ensure_physical_device(struct anv_instance
*instance
)
876 if (instance
->physicalDeviceCount
< 0) {
877 VkResult result
= anv_enumerate_devices(instance
);
878 if (result
!= VK_SUCCESS
&&
879 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
886 VkResult
anv_EnumeratePhysicalDevices(
887 VkInstance _instance
,
888 uint32_t* pPhysicalDeviceCount
,
889 VkPhysicalDevice
* pPhysicalDevices
)
891 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
892 VK_OUTARRAY_MAKE(out
, pPhysicalDevices
, pPhysicalDeviceCount
);
894 VkResult result
= anv_instance_ensure_physical_device(instance
);
895 if (result
!= VK_SUCCESS
)
898 if (instance
->physicalDeviceCount
== 0)
901 assert(instance
->physicalDeviceCount
== 1);
902 vk_outarray_append(&out
, i
) {
903 *i
= anv_physical_device_to_handle(&instance
->physicalDevice
);
906 return vk_outarray_status(&out
);
909 VkResult
anv_EnumeratePhysicalDeviceGroups(
910 VkInstance _instance
,
911 uint32_t* pPhysicalDeviceGroupCount
,
912 VkPhysicalDeviceGroupProperties
* pPhysicalDeviceGroupProperties
)
914 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
915 VK_OUTARRAY_MAKE(out
, pPhysicalDeviceGroupProperties
,
916 pPhysicalDeviceGroupCount
);
918 VkResult result
= anv_instance_ensure_physical_device(instance
);
919 if (result
!= VK_SUCCESS
)
922 if (instance
->physicalDeviceCount
== 0)
925 assert(instance
->physicalDeviceCount
== 1);
927 vk_outarray_append(&out
, p
) {
928 p
->physicalDeviceCount
= 1;
929 memset(p
->physicalDevices
, 0, sizeof(p
->physicalDevices
));
930 p
->physicalDevices
[0] =
931 anv_physical_device_to_handle(&instance
->physicalDevice
);
932 p
->subsetAllocation
= false;
934 vk_foreach_struct(ext
, p
->pNext
)
935 anv_debug_ignored_stype(ext
->sType
);
938 return vk_outarray_status(&out
);
941 void anv_GetPhysicalDeviceFeatures(
942 VkPhysicalDevice physicalDevice
,
943 VkPhysicalDeviceFeatures
* pFeatures
)
945 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
947 *pFeatures
= (VkPhysicalDeviceFeatures
) {
948 .robustBufferAccess
= true,
949 .fullDrawIndexUint32
= true,
950 .imageCubeArray
= true,
951 .independentBlend
= true,
952 .geometryShader
= true,
953 .tessellationShader
= true,
954 .sampleRateShading
= true,
955 .dualSrcBlend
= true,
957 .multiDrawIndirect
= true,
958 .drawIndirectFirstInstance
= true,
960 .depthBiasClamp
= true,
961 .fillModeNonSolid
= true,
962 .depthBounds
= pdevice
->info
.gen
>= 12,
966 .multiViewport
= true,
967 .samplerAnisotropy
= true,
968 .textureCompressionETC2
= pdevice
->info
.gen
>= 8 ||
969 pdevice
->info
.is_baytrail
,
970 .textureCompressionASTC_LDR
= pdevice
->info
.gen
>= 9, /* FINISHME CHV */
971 .textureCompressionBC
= true,
972 .occlusionQueryPrecise
= true,
973 .pipelineStatisticsQuery
= true,
974 .fragmentStoresAndAtomics
= true,
975 .shaderTessellationAndGeometryPointSize
= true,
976 .shaderImageGatherExtended
= true,
977 .shaderStorageImageExtendedFormats
= true,
978 .shaderStorageImageMultisample
= false,
979 .shaderStorageImageReadWithoutFormat
= false,
980 .shaderStorageImageWriteWithoutFormat
= true,
981 .shaderUniformBufferArrayDynamicIndexing
= true,
982 .shaderSampledImageArrayDynamicIndexing
= true,
983 .shaderStorageBufferArrayDynamicIndexing
= true,
984 .shaderStorageImageArrayDynamicIndexing
= true,
985 .shaderClipDistance
= true,
986 .shaderCullDistance
= true,
987 .shaderFloat64
= pdevice
->info
.gen
>= 8 &&
988 pdevice
->info
.has_64bit_types
,
989 .shaderInt64
= pdevice
->info
.gen
>= 8 &&
990 pdevice
->info
.has_64bit_types
,
991 .shaderInt16
= pdevice
->info
.gen
>= 8,
992 .shaderResourceMinLod
= pdevice
->info
.gen
>= 9,
993 .variableMultisampleRate
= true,
994 .inheritedQueries
= true,
997 /* We can't do image stores in vec4 shaders */
998 pFeatures
->vertexPipelineStoresAndAtomics
=
999 pdevice
->compiler
->scalar_stage
[MESA_SHADER_VERTEX
] &&
1000 pdevice
->compiler
->scalar_stage
[MESA_SHADER_GEOMETRY
];
1002 struct anv_app_info
*app_info
= &pdevice
->instance
->app_info
;
1004 /* The new DOOM and Wolfenstein games require depthBounds without
1005 * checking for it. They seem to run fine without it so just claim it's
1006 * there and accept the consequences.
1008 if (app_info
->engine_name
&& strcmp(app_info
->engine_name
, "idTech") == 0)
1009 pFeatures
->depthBounds
= true;
1012 void anv_GetPhysicalDeviceFeatures2(
1013 VkPhysicalDevice physicalDevice
,
1014 VkPhysicalDeviceFeatures2
* pFeatures
)
1016 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1017 anv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
1019 vk_foreach_struct(ext
, pFeatures
->pNext
) {
1020 switch (ext
->sType
) {
1021 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES_KHR
: {
1022 VkPhysicalDevice8BitStorageFeaturesKHR
*features
=
1023 (VkPhysicalDevice8BitStorageFeaturesKHR
*)ext
;
1024 features
->storageBuffer8BitAccess
= pdevice
->info
.gen
>= 8;
1025 features
->uniformAndStorageBuffer8BitAccess
= pdevice
->info
.gen
>= 8;
1026 features
->storagePushConstant8
= pdevice
->info
.gen
>= 8;
1030 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES
: {
1031 VkPhysicalDevice16BitStorageFeatures
*features
=
1032 (VkPhysicalDevice16BitStorageFeatures
*)ext
;
1033 features
->storageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
1034 features
->uniformAndStorageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
1035 features
->storagePushConstant16
= pdevice
->info
.gen
>= 8;
1036 features
->storageInputOutput16
= false;
1040 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_EXT
: {
1041 VkPhysicalDeviceBufferDeviceAddressFeaturesEXT
*features
= (void *)ext
;
1042 features
->bufferDeviceAddress
= pdevice
->has_a64_buffer_access
;
1043 features
->bufferDeviceAddressCaptureReplay
= false;
1044 features
->bufferDeviceAddressMultiDevice
= false;
1048 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COMPUTE_SHADER_DERIVATIVES_FEATURES_NV
: {
1049 VkPhysicalDeviceComputeShaderDerivativesFeaturesNV
*features
=
1050 (VkPhysicalDeviceComputeShaderDerivativesFeaturesNV
*)ext
;
1051 features
->computeDerivativeGroupQuads
= true;
1052 features
->computeDerivativeGroupLinear
= true;
1056 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT
: {
1057 VkPhysicalDeviceConditionalRenderingFeaturesEXT
*features
=
1058 (VkPhysicalDeviceConditionalRenderingFeaturesEXT
*)ext
;
1059 features
->conditionalRendering
= pdevice
->info
.gen
>= 8 ||
1060 pdevice
->info
.is_haswell
;
1061 features
->inheritedConditionalRendering
= pdevice
->info
.gen
>= 8 ||
1062 pdevice
->info
.is_haswell
;
1066 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_CLIP_ENABLE_FEATURES_EXT
: {
1067 VkPhysicalDeviceDepthClipEnableFeaturesEXT
*features
=
1068 (VkPhysicalDeviceDepthClipEnableFeaturesEXT
*)ext
;
1069 features
->depthClipEnable
= true;
1073 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR
: {
1074 VkPhysicalDeviceFloat16Int8FeaturesKHR
*features
= (void *)ext
;
1075 features
->shaderFloat16
= pdevice
->info
.gen
>= 8;
1076 features
->shaderInt8
= pdevice
->info
.gen
>= 8;
1080 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FRAGMENT_SHADER_INTERLOCK_FEATURES_EXT
: {
1081 VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT
*features
=
1082 (VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT
*)ext
;
1083 features
->fragmentShaderSampleInterlock
= pdevice
->info
.gen
>= 9;
1084 features
->fragmentShaderPixelInterlock
= pdevice
->info
.gen
>= 9;
1085 features
->fragmentShaderShadingRateInterlock
= false;
1089 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_QUERY_RESET_FEATURES_EXT
: {
1090 VkPhysicalDeviceHostQueryResetFeaturesEXT
*features
=
1091 (VkPhysicalDeviceHostQueryResetFeaturesEXT
*)ext
;
1092 features
->hostQueryReset
= true;
1096 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT
: {
1097 VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*features
=
1098 (VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*)ext
;
1099 features
->shaderInputAttachmentArrayDynamicIndexing
= false;
1100 features
->shaderUniformTexelBufferArrayDynamicIndexing
= true;
1101 features
->shaderStorageTexelBufferArrayDynamicIndexing
= true;
1102 features
->shaderUniformBufferArrayNonUniformIndexing
= false;
1103 features
->shaderSampledImageArrayNonUniformIndexing
= true;
1104 features
->shaderStorageBufferArrayNonUniformIndexing
= true;
1105 features
->shaderStorageImageArrayNonUniformIndexing
= true;
1106 features
->shaderInputAttachmentArrayNonUniformIndexing
= false;
1107 features
->shaderUniformTexelBufferArrayNonUniformIndexing
= true;
1108 features
->shaderStorageTexelBufferArrayNonUniformIndexing
= true;
1109 features
->descriptorBindingUniformBufferUpdateAfterBind
= false;
1110 features
->descriptorBindingSampledImageUpdateAfterBind
= true;
1111 features
->descriptorBindingStorageImageUpdateAfterBind
= true;
1112 features
->descriptorBindingStorageBufferUpdateAfterBind
= true;
1113 features
->descriptorBindingUniformTexelBufferUpdateAfterBind
= true;
1114 features
->descriptorBindingStorageTexelBufferUpdateAfterBind
= true;
1115 features
->descriptorBindingUpdateUnusedWhilePending
= true;
1116 features
->descriptorBindingPartiallyBound
= true;
1117 features
->descriptorBindingVariableDescriptorCount
= false;
1118 features
->runtimeDescriptorArray
= true;
1122 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INDEX_TYPE_UINT8_FEATURES_EXT
: {
1123 VkPhysicalDeviceIndexTypeUint8FeaturesEXT
*features
=
1124 (VkPhysicalDeviceIndexTypeUint8FeaturesEXT
*)ext
;
1125 features
->indexTypeUint8
= true;
1129 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_FEATURES_EXT
: {
1130 VkPhysicalDeviceInlineUniformBlockFeaturesEXT
*features
=
1131 (VkPhysicalDeviceInlineUniformBlockFeaturesEXT
*)ext
;
1132 features
->inlineUniformBlock
= true;
1133 features
->descriptorBindingInlineUniformBlockUpdateAfterBind
= true;
1137 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_FEATURES_EXT
: {
1138 VkPhysicalDeviceLineRasterizationFeaturesEXT
*features
=
1139 (VkPhysicalDeviceLineRasterizationFeaturesEXT
*)ext
;
1140 features
->rectangularLines
= true;
1141 features
->bresenhamLines
= true;
1142 features
->smoothLines
= true;
1143 features
->stippledRectangularLines
= false;
1144 features
->stippledBresenhamLines
= true;
1145 features
->stippledSmoothLines
= false;
1149 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES
: {
1150 VkPhysicalDeviceMultiviewFeatures
*features
=
1151 (VkPhysicalDeviceMultiviewFeatures
*)ext
;
1152 features
->multiview
= true;
1153 features
->multiviewGeometryShader
= true;
1154 features
->multiviewTessellationShader
= true;
1158 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGELESS_FRAMEBUFFER_FEATURES_KHR
: {
1159 VkPhysicalDeviceImagelessFramebufferFeaturesKHR
*features
=
1160 (VkPhysicalDeviceImagelessFramebufferFeaturesKHR
*)ext
;
1161 features
->imagelessFramebuffer
= true;
1165 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR
: {
1166 VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR
*features
=
1167 (VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR
*)ext
;
1168 features
->pipelineExecutableInfo
= true;
1172 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES
: {
1173 VkPhysicalDeviceProtectedMemoryFeatures
*features
= (void *)ext
;
1174 features
->protectedMemory
= false;
1178 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES
: {
1179 VkPhysicalDeviceSamplerYcbcrConversionFeatures
*features
=
1180 (VkPhysicalDeviceSamplerYcbcrConversionFeatures
*) ext
;
1181 features
->samplerYcbcrConversion
= true;
1185 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES_EXT
: {
1186 VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*features
=
1187 (VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*)ext
;
1188 features
->scalarBlockLayout
= true;
1192 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SEPARATE_DEPTH_STENCIL_LAYOUTS_FEATURES_KHR
: {
1193 VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR
*features
=
1194 (VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR
*)ext
;
1195 features
->separateDepthStencilLayouts
= true;
1199 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_INT64_FEATURES_KHR
: {
1200 VkPhysicalDeviceShaderAtomicInt64FeaturesKHR
*features
= (void *)ext
;
1201 features
->shaderBufferInt64Atomics
=
1202 pdevice
->info
.gen
>= 9 && pdevice
->use_softpin
;
1203 features
->shaderSharedInt64Atomics
= VK_FALSE
;
1207 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DEMOTE_TO_HELPER_INVOCATION_FEATURES_EXT
: {
1208 VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT
*features
= (void *)ext
;
1209 features
->shaderDemoteToHelperInvocation
= true;
1213 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CLOCK_FEATURES_KHR
: {
1214 VkPhysicalDeviceShaderClockFeaturesKHR
*features
=
1215 (VkPhysicalDeviceShaderClockFeaturesKHR
*)ext
;
1216 features
->shaderSubgroupClock
= true;
1217 features
->shaderDeviceClock
= false;
1221 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETERS_FEATURES
: {
1222 VkPhysicalDeviceShaderDrawParametersFeatures
*features
= (void *)ext
;
1223 features
->shaderDrawParameters
= true;
1227 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_SUBGROUP_EXTENDED_TYPES_FEATURES_KHR
: {
1228 VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR
*features
=
1229 (VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR
*)ext
;
1230 features
->shaderSubgroupExtendedTypes
= true;
1234 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_FEATURES_EXT
: {
1235 VkPhysicalDeviceSubgroupSizeControlFeaturesEXT
*features
=
1236 (VkPhysicalDeviceSubgroupSizeControlFeaturesEXT
*)ext
;
1237 features
->subgroupSizeControl
= true;
1238 features
->computeFullSubgroups
= true;
1242 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_FEATURES_EXT
: {
1243 VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT
*features
=
1244 (VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT
*)ext
;
1245 features
->texelBufferAlignment
= true;
1249 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_FEATURES_KHR
: {
1250 VkPhysicalDeviceTimelineSemaphoreFeaturesKHR
*features
=
1251 (VkPhysicalDeviceTimelineSemaphoreFeaturesKHR
*) ext
;
1252 features
->timelineSemaphore
= true;
1256 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES
: {
1257 VkPhysicalDeviceVariablePointersFeatures
*features
= (void *)ext
;
1258 features
->variablePointersStorageBuffer
= true;
1259 features
->variablePointers
= true;
1263 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT
: {
1264 VkPhysicalDeviceTransformFeedbackFeaturesEXT
*features
=
1265 (VkPhysicalDeviceTransformFeedbackFeaturesEXT
*)ext
;
1266 features
->transformFeedback
= true;
1267 features
->geometryStreams
= true;
1271 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_UNIFORM_BUFFER_STANDARD_LAYOUT_FEATURES_KHR
: {
1272 VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR
*features
=
1273 (VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR
*)ext
;
1274 features
->uniformBufferStandardLayout
= true;
1278 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT
: {
1279 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*features
=
1280 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*)ext
;
1281 features
->vertexAttributeInstanceRateDivisor
= true;
1282 features
->vertexAttributeInstanceRateZeroDivisor
= true;
1286 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_MEMORY_MODEL_FEATURES_KHR
: {
1287 VkPhysicalDeviceVulkanMemoryModelFeaturesKHR
*features
= (void *)ext
;
1288 features
->vulkanMemoryModel
= true;
1289 features
->vulkanMemoryModelDeviceScope
= true;
1290 features
->vulkanMemoryModelAvailabilityVisibilityChains
= true;
1294 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_YCBCR_IMAGE_ARRAYS_FEATURES_EXT
: {
1295 VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*features
=
1296 (VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*)ext
;
1297 features
->ycbcrImageArrays
= true;
1302 anv_debug_ignored_stype(ext
->sType
);
1308 #define MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS 64
1310 #define MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS 64
1311 #define MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS 256
1313 void anv_GetPhysicalDeviceProperties(
1314 VkPhysicalDevice physicalDevice
,
1315 VkPhysicalDeviceProperties
* pProperties
)
1317 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1318 const struct gen_device_info
*devinfo
= &pdevice
->info
;
1320 /* See assertions made when programming the buffer surface state. */
1321 const uint32_t max_raw_buffer_sz
= devinfo
->gen
>= 7 ?
1322 (1ul << 30) : (1ul << 27);
1324 const uint32_t max_ssbos
= pdevice
->has_a64_buffer_access
? UINT16_MAX
: 64;
1325 const uint32_t max_textures
=
1326 pdevice
->has_bindless_images
? UINT16_MAX
: 128;
1327 const uint32_t max_samplers
=
1328 pdevice
->has_bindless_samplers
? UINT16_MAX
:
1329 (devinfo
->gen
>= 8 || devinfo
->is_haswell
) ? 128 : 16;
1330 const uint32_t max_images
=
1331 pdevice
->has_bindless_images
? UINT16_MAX
: MAX_IMAGES
;
1333 /* If we can use bindless for everything, claim a high per-stage limit,
1334 * otherwise use the binding table size, minus the slots reserved for
1335 * render targets and one slot for the descriptor buffer. */
1336 const uint32_t max_per_stage
=
1337 pdevice
->has_bindless_images
&& pdevice
->has_a64_buffer_access
1338 ? UINT32_MAX
: MAX_BINDING_TABLE_SIZE
- MAX_RTS
- 1;
1340 const uint32_t max_workgroup_size
= 32 * devinfo
->max_cs_threads
;
1342 VkSampleCountFlags sample_counts
=
1343 isl_device_get_sample_counts(&pdevice
->isl_dev
);
1346 VkPhysicalDeviceLimits limits
= {
1347 .maxImageDimension1D
= (1 << 14),
1348 .maxImageDimension2D
= (1 << 14),
1349 .maxImageDimension3D
= (1 << 11),
1350 .maxImageDimensionCube
= (1 << 14),
1351 .maxImageArrayLayers
= (1 << 11),
1352 .maxTexelBufferElements
= 128 * 1024 * 1024,
1353 .maxUniformBufferRange
= (1ul << 27),
1354 .maxStorageBufferRange
= max_raw_buffer_sz
,
1355 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
1356 .maxMemoryAllocationCount
= UINT32_MAX
,
1357 .maxSamplerAllocationCount
= 64 * 1024,
1358 .bufferImageGranularity
= 64, /* A cache line */
1359 .sparseAddressSpaceSize
= 0,
1360 .maxBoundDescriptorSets
= MAX_SETS
,
1361 .maxPerStageDescriptorSamplers
= max_samplers
,
1362 .maxPerStageDescriptorUniformBuffers
= MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
,
1363 .maxPerStageDescriptorStorageBuffers
= max_ssbos
,
1364 .maxPerStageDescriptorSampledImages
= max_textures
,
1365 .maxPerStageDescriptorStorageImages
= max_images
,
1366 .maxPerStageDescriptorInputAttachments
= MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS
,
1367 .maxPerStageResources
= max_per_stage
,
1368 .maxDescriptorSetSamplers
= 6 * max_samplers
, /* number of stages * maxPerStageDescriptorSamplers */
1369 .maxDescriptorSetUniformBuffers
= 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
, /* number of stages * maxPerStageDescriptorUniformBuffers */
1370 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
1371 .maxDescriptorSetStorageBuffers
= 6 * max_ssbos
, /* number of stages * maxPerStageDescriptorStorageBuffers */
1372 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
1373 .maxDescriptorSetSampledImages
= 6 * max_textures
, /* number of stages * maxPerStageDescriptorSampledImages */
1374 .maxDescriptorSetStorageImages
= 6 * max_images
, /* number of stages * maxPerStageDescriptorStorageImages */
1375 .maxDescriptorSetInputAttachments
= MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS
,
1376 .maxVertexInputAttributes
= MAX_VBS
,
1377 .maxVertexInputBindings
= MAX_VBS
,
1378 .maxVertexInputAttributeOffset
= 2047,
1379 .maxVertexInputBindingStride
= 2048,
1380 .maxVertexOutputComponents
= 128,
1381 .maxTessellationGenerationLevel
= 64,
1382 .maxTessellationPatchSize
= 32,
1383 .maxTessellationControlPerVertexInputComponents
= 128,
1384 .maxTessellationControlPerVertexOutputComponents
= 128,
1385 .maxTessellationControlPerPatchOutputComponents
= 128,
1386 .maxTessellationControlTotalOutputComponents
= 2048,
1387 .maxTessellationEvaluationInputComponents
= 128,
1388 .maxTessellationEvaluationOutputComponents
= 128,
1389 .maxGeometryShaderInvocations
= 32,
1390 .maxGeometryInputComponents
= 64,
1391 .maxGeometryOutputComponents
= 128,
1392 .maxGeometryOutputVertices
= 256,
1393 .maxGeometryTotalOutputComponents
= 1024,
1394 .maxFragmentInputComponents
= 116, /* 128 components - (PSIZ, CLIP_DIST0, CLIP_DIST1) */
1395 .maxFragmentOutputAttachments
= 8,
1396 .maxFragmentDualSrcAttachments
= 1,
1397 .maxFragmentCombinedOutputResources
= 8,
1398 .maxComputeSharedMemorySize
= 64 * 1024,
1399 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
1400 .maxComputeWorkGroupInvocations
= max_workgroup_size
,
1401 .maxComputeWorkGroupSize
= {
1406 .subPixelPrecisionBits
= 8,
1407 .subTexelPrecisionBits
= 8,
1408 .mipmapPrecisionBits
= 8,
1409 .maxDrawIndexedIndexValue
= UINT32_MAX
,
1410 .maxDrawIndirectCount
= UINT32_MAX
,
1411 .maxSamplerLodBias
= 16,
1412 .maxSamplerAnisotropy
= 16,
1413 .maxViewports
= MAX_VIEWPORTS
,
1414 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
1415 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
1416 .viewportSubPixelBits
= 13, /* We take a float? */
1417 .minMemoryMapAlignment
= 4096, /* A page */
1418 /* The dataport requires texel alignment so we need to assume a worst
1419 * case of R32G32B32A32 which is 16 bytes.
1421 .minTexelBufferOffsetAlignment
= 16,
1422 /* We need 16 for UBO block reads to work and 32 for push UBOs */
1423 .minUniformBufferOffsetAlignment
= 32,
1424 .minStorageBufferOffsetAlignment
= 4,
1425 .minTexelOffset
= -8,
1426 .maxTexelOffset
= 7,
1427 .minTexelGatherOffset
= -32,
1428 .maxTexelGatherOffset
= 31,
1429 .minInterpolationOffset
= -0.5,
1430 .maxInterpolationOffset
= 0.4375,
1431 .subPixelInterpolationOffsetBits
= 4,
1432 .maxFramebufferWidth
= (1 << 14),
1433 .maxFramebufferHeight
= (1 << 14),
1434 .maxFramebufferLayers
= (1 << 11),
1435 .framebufferColorSampleCounts
= sample_counts
,
1436 .framebufferDepthSampleCounts
= sample_counts
,
1437 .framebufferStencilSampleCounts
= sample_counts
,
1438 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
1439 .maxColorAttachments
= MAX_RTS
,
1440 .sampledImageColorSampleCounts
= sample_counts
,
1441 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
1442 .sampledImageDepthSampleCounts
= sample_counts
,
1443 .sampledImageStencilSampleCounts
= sample_counts
,
1444 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
1445 .maxSampleMaskWords
= 1,
1446 .timestampComputeAndGraphics
= true,
1447 .timestampPeriod
= 1000000000.0 / devinfo
->timestamp_frequency
,
1448 .maxClipDistances
= 8,
1449 .maxCullDistances
= 8,
1450 .maxCombinedClipAndCullDistances
= 8,
1451 .discreteQueuePriorities
= 2,
1452 .pointSizeRange
= { 0.125, 255.875 },
1455 (devinfo
->gen
>= 9 || devinfo
->is_cherryview
) ?
1456 2047.9921875 : 7.9921875,
1458 .pointSizeGranularity
= (1.0 / 8.0),
1459 .lineWidthGranularity
= (1.0 / 128.0),
1460 .strictLines
= false,
1461 .standardSampleLocations
= true,
1462 .optimalBufferCopyOffsetAlignment
= 128,
1463 .optimalBufferCopyRowPitchAlignment
= 128,
1464 .nonCoherentAtomSize
= 64,
1467 *pProperties
= (VkPhysicalDeviceProperties
) {
1468 .apiVersion
= anv_physical_device_api_version(pdevice
),
1469 .driverVersion
= vk_get_driver_version(),
1471 .deviceID
= pdevice
->chipset_id
,
1472 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
1474 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
1477 snprintf(pProperties
->deviceName
, sizeof(pProperties
->deviceName
),
1478 "%s", pdevice
->name
);
1479 memcpy(pProperties
->pipelineCacheUUID
,
1480 pdevice
->pipeline_cache_uuid
, VK_UUID_SIZE
);
1483 void anv_GetPhysicalDeviceProperties2(
1484 VkPhysicalDevice physicalDevice
,
1485 VkPhysicalDeviceProperties2
* pProperties
)
1487 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1489 anv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
1491 vk_foreach_struct(ext
, pProperties
->pNext
) {
1492 switch (ext
->sType
) {
1493 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_STENCIL_RESOLVE_PROPERTIES_KHR
: {
1494 VkPhysicalDeviceDepthStencilResolvePropertiesKHR
*props
=
1495 (VkPhysicalDeviceDepthStencilResolvePropertiesKHR
*)ext
;
1497 /* We support all of the depth resolve modes */
1498 props
->supportedDepthResolveModes
=
1499 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
|
1500 VK_RESOLVE_MODE_AVERAGE_BIT_KHR
|
1501 VK_RESOLVE_MODE_MIN_BIT_KHR
|
1502 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1504 /* Average doesn't make sense for stencil so we don't support that */
1505 props
->supportedStencilResolveModes
=
1506 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
;
1507 if (pdevice
->info
.gen
>= 8) {
1508 /* The advanced stencil resolve modes currently require stencil
1509 * sampling be supported by the hardware.
1511 props
->supportedStencilResolveModes
|=
1512 VK_RESOLVE_MODE_MIN_BIT_KHR
|
1513 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1516 props
->independentResolveNone
= true;
1517 props
->independentResolve
= true;
1521 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES_EXT
: {
1522 VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*props
=
1523 (VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*)ext
;
1525 /* It's a bit hard to exactly map our implementation to the limits
1526 * described here. The bindless surface handle in the extended
1527 * message descriptors is 20 bits and it's an index into the table of
1528 * RENDER_SURFACE_STATE structs that starts at bindless surface base
1529 * address. Given that most things consume two surface states per
1530 * view (general/sampled for textures and write-only/read-write for
1531 * images), we claim 2^19 things.
1533 * For SSBOs, we just use A64 messages so there is no real limit
1534 * there beyond the limit on the total size of a descriptor set.
1536 const unsigned max_bindless_views
= 1 << 19;
1538 props
->maxUpdateAfterBindDescriptorsInAllPools
= max_bindless_views
;
1539 props
->shaderUniformBufferArrayNonUniformIndexingNative
= false;
1540 props
->shaderSampledImageArrayNonUniformIndexingNative
= false;
1541 props
->shaderStorageBufferArrayNonUniformIndexingNative
= true;
1542 props
->shaderStorageImageArrayNonUniformIndexingNative
= false;
1543 props
->shaderInputAttachmentArrayNonUniformIndexingNative
= false;
1544 props
->robustBufferAccessUpdateAfterBind
= true;
1545 props
->quadDivergentImplicitLod
= false;
1546 props
->maxPerStageDescriptorUpdateAfterBindSamplers
= max_bindless_views
;
1547 props
->maxPerStageDescriptorUpdateAfterBindUniformBuffers
= MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
;
1548 props
->maxPerStageDescriptorUpdateAfterBindStorageBuffers
= UINT32_MAX
;
1549 props
->maxPerStageDescriptorUpdateAfterBindSampledImages
= max_bindless_views
;
1550 props
->maxPerStageDescriptorUpdateAfterBindStorageImages
= max_bindless_views
;
1551 props
->maxPerStageDescriptorUpdateAfterBindInputAttachments
= MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS
;
1552 props
->maxPerStageUpdateAfterBindResources
= UINT32_MAX
;
1553 props
->maxDescriptorSetUpdateAfterBindSamplers
= max_bindless_views
;
1554 props
->maxDescriptorSetUpdateAfterBindUniformBuffers
= 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
;
1555 props
->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2;
1556 props
->maxDescriptorSetUpdateAfterBindStorageBuffers
= UINT32_MAX
;
1557 props
->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2;
1558 props
->maxDescriptorSetUpdateAfterBindSampledImages
= max_bindless_views
;
1559 props
->maxDescriptorSetUpdateAfterBindStorageImages
= max_bindless_views
;
1560 props
->maxDescriptorSetUpdateAfterBindInputAttachments
= MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS
;
1564 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR
: {
1565 VkPhysicalDeviceDriverPropertiesKHR
*driver_props
=
1566 (VkPhysicalDeviceDriverPropertiesKHR
*) ext
;
1568 driver_props
->driverID
= VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA_KHR
;
1569 snprintf(driver_props
->driverName
, VK_MAX_DRIVER_NAME_SIZE_KHR
,
1570 "Intel open-source Mesa driver");
1572 snprintf(driver_props
->driverInfo
, VK_MAX_DRIVER_INFO_SIZE_KHR
,
1573 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1
);
1575 driver_props
->conformanceVersion
= (VkConformanceVersionKHR
) {
1584 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT
: {
1585 VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*props
=
1586 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*) ext
;
1587 /* Userptr needs page aligned memory. */
1588 props
->minImportedHostPointerAlignment
= 4096;
1592 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES
: {
1593 VkPhysicalDeviceIDProperties
*id_props
=
1594 (VkPhysicalDeviceIDProperties
*)ext
;
1595 memcpy(id_props
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
1596 memcpy(id_props
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
1597 /* The LUID is for Windows. */
1598 id_props
->deviceLUIDValid
= false;
1602 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_PROPERTIES_EXT
: {
1603 VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*props
=
1604 (VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*)ext
;
1605 props
->maxInlineUniformBlockSize
= MAX_INLINE_UNIFORM_BLOCK_SIZE
;
1606 props
->maxPerStageDescriptorInlineUniformBlocks
=
1607 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1608 props
->maxPerStageDescriptorUpdateAfterBindInlineUniformBlocks
=
1609 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1610 props
->maxDescriptorSetInlineUniformBlocks
=
1611 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1612 props
->maxDescriptorSetUpdateAfterBindInlineUniformBlocks
=
1613 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1617 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_PROPERTIES_EXT
: {
1618 VkPhysicalDeviceLineRasterizationPropertiesEXT
*props
=
1619 (VkPhysicalDeviceLineRasterizationPropertiesEXT
*)ext
;
1620 /* In the Skylake PRM Vol. 7, subsection titled "GIQ (Diamond)
1621 * Sampling Rules - Legacy Mode", it says the following:
1623 * "Note that the device divides a pixel into a 16x16 array of
1624 * subpixels, referenced by their upper left corners."
1626 * This is the only known reference in the PRMs to the subpixel
1627 * precision of line rasterization and a "16x16 array of subpixels"
1628 * implies 4 subpixel precision bits. Empirical testing has shown
1629 * that 4 subpixel precision bits applies to all line rasterization
1632 props
->lineSubPixelPrecisionBits
= 4;
1636 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES
: {
1637 VkPhysicalDeviceMaintenance3Properties
*props
=
1638 (VkPhysicalDeviceMaintenance3Properties
*)ext
;
1639 /* This value doesn't matter for us today as our per-stage
1640 * descriptors are the real limit.
1642 props
->maxPerSetDescriptors
= 1024;
1643 props
->maxMemoryAllocationSize
= MAX_MEMORY_ALLOCATION_SIZE
;
1647 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES
: {
1648 VkPhysicalDeviceMultiviewProperties
*properties
=
1649 (VkPhysicalDeviceMultiviewProperties
*)ext
;
1650 properties
->maxMultiviewViewCount
= 16;
1651 properties
->maxMultiviewInstanceIndex
= UINT32_MAX
/ 16;
1655 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT
: {
1656 VkPhysicalDevicePCIBusInfoPropertiesEXT
*properties
=
1657 (VkPhysicalDevicePCIBusInfoPropertiesEXT
*)ext
;
1658 properties
->pciDomain
= pdevice
->pci_info
.domain
;
1659 properties
->pciBus
= pdevice
->pci_info
.bus
;
1660 properties
->pciDevice
= pdevice
->pci_info
.device
;
1661 properties
->pciFunction
= pdevice
->pci_info
.function
;
1665 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES
: {
1666 VkPhysicalDevicePointClippingProperties
*properties
=
1667 (VkPhysicalDevicePointClippingProperties
*) ext
;
1668 properties
->pointClippingBehavior
= VK_POINT_CLIPPING_BEHAVIOR_USER_CLIP_PLANES_ONLY
;
1672 #pragma GCC diagnostic push
1673 #pragma GCC diagnostic ignored "-Wswitch"
1674 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PRESENTATION_PROPERTIES_ANDROID
: {
1675 VkPhysicalDevicePresentationPropertiesANDROID
*props
=
1676 (VkPhysicalDevicePresentationPropertiesANDROID
*)ext
;
1677 props
->sharedImage
= VK_FALSE
;
1680 #pragma GCC diagnostic pop
1682 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES
: {
1683 VkPhysicalDeviceProtectedMemoryProperties
*props
=
1684 (VkPhysicalDeviceProtectedMemoryProperties
*)ext
;
1685 props
->protectedNoFault
= false;
1689 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
1690 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
1691 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
1693 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
1697 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT
: {
1698 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*properties
=
1699 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*)ext
;
1700 properties
->filterMinmaxImageComponentMapping
= pdevice
->info
.gen
>= 9;
1701 properties
->filterMinmaxSingleComponentFormats
= true;
1705 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES
: {
1706 VkPhysicalDeviceSubgroupProperties
*properties
= (void *)ext
;
1708 properties
->subgroupSize
= BRW_SUBGROUP_SIZE
;
1710 VkShaderStageFlags scalar_stages
= 0;
1711 for (unsigned stage
= 0; stage
< MESA_SHADER_STAGES
; stage
++) {
1712 if (pdevice
->compiler
->scalar_stage
[stage
])
1713 scalar_stages
|= mesa_to_vk_shader_stage(stage
);
1715 properties
->supportedStages
= scalar_stages
;
1717 properties
->supportedOperations
= VK_SUBGROUP_FEATURE_BASIC_BIT
|
1718 VK_SUBGROUP_FEATURE_VOTE_BIT
|
1719 VK_SUBGROUP_FEATURE_BALLOT_BIT
|
1720 VK_SUBGROUP_FEATURE_SHUFFLE_BIT
|
1721 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT
|
1722 VK_SUBGROUP_FEATURE_QUAD_BIT
;
1723 if (pdevice
->info
.gen
>= 8) {
1724 /* TODO: There's no technical reason why these can't be made to
1725 * work on gen7 but they don't at the moment so it's best to leave
1726 * the feature disabled than enabled and broken.
1728 properties
->supportedOperations
|=
1729 VK_SUBGROUP_FEATURE_ARITHMETIC_BIT
|
1730 VK_SUBGROUP_FEATURE_CLUSTERED_BIT
;
1732 properties
->quadOperationsInAllStages
= pdevice
->info
.gen
>= 8;
1736 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_PROPERTIES_EXT
: {
1737 VkPhysicalDeviceSubgroupSizeControlPropertiesEXT
*props
=
1738 (VkPhysicalDeviceSubgroupSizeControlPropertiesEXT
*)ext
;
1739 STATIC_ASSERT(8 <= BRW_SUBGROUP_SIZE
&& BRW_SUBGROUP_SIZE
<= 32);
1740 props
->minSubgroupSize
= 8;
1741 props
->maxSubgroupSize
= 32;
1742 props
->maxComputeWorkgroupSubgroups
= pdevice
->info
.max_cs_threads
;
1743 props
->requiredSubgroupSizeStages
= VK_SHADER_STAGE_COMPUTE_BIT
;
1746 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT_CONTROLS_PROPERTIES_KHR
: {
1747 VkPhysicalDeviceFloatControlsPropertiesKHR
*properties
= (void *)ext
;
1748 properties
->denormBehaviorIndependence
= VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_ALL_KHR
;
1749 properties
->roundingModeIndependence
= VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_NONE_KHR
;
1751 /* Broadwell does not support HF denorms and there are restrictions
1752 * other gens. According to Kabylake's PRM:
1754 * "math - Extended Math Function
1756 * Restriction : Half-float denorms are always retained."
1758 properties
->shaderDenormFlushToZeroFloat16
= false;
1759 properties
->shaderDenormPreserveFloat16
= pdevice
->info
.gen
> 8;
1760 properties
->shaderRoundingModeRTEFloat16
= true;
1761 properties
->shaderRoundingModeRTZFloat16
= true;
1762 properties
->shaderSignedZeroInfNanPreserveFloat16
= true;
1764 properties
->shaderDenormFlushToZeroFloat32
= true;
1765 properties
->shaderDenormPreserveFloat32
= true;
1766 properties
->shaderRoundingModeRTEFloat32
= true;
1767 properties
->shaderRoundingModeRTZFloat32
= true;
1768 properties
->shaderSignedZeroInfNanPreserveFloat32
= true;
1770 properties
->shaderDenormFlushToZeroFloat64
= true;
1771 properties
->shaderDenormPreserveFloat64
= true;
1772 properties
->shaderRoundingModeRTEFloat64
= true;
1773 properties
->shaderRoundingModeRTZFloat64
= true;
1774 properties
->shaderSignedZeroInfNanPreserveFloat64
= true;
1778 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_PROPERTIES_EXT
: {
1779 VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT
*props
=
1780 (VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT
*)ext
;
1782 /* From the SKL PRM Vol. 2d, docs for RENDER_SURFACE_STATE::Surface
1785 * "For SURFTYPE_BUFFER non-rendertarget surfaces, this field
1786 * specifies the base address of the first element of the surface,
1787 * computed in software by adding the surface base address to the
1788 * byte offset of the element in the buffer. The base address must
1789 * be aligned to element size."
1791 * The typed dataport messages require that things be texel aligned.
1792 * Otherwise, we may just load/store the wrong data or, in the worst
1793 * case, there may be hangs.
1795 props
->storageTexelBufferOffsetAlignmentBytes
= 16;
1796 props
->storageTexelBufferOffsetSingleTexelAlignment
= true;
1798 /* The sampler, however, is much more forgiving and it can handle
1799 * arbitrary byte alignment for linear and buffer surfaces. It's
1800 * hard to find a good PRM citation for this but years of empirical
1801 * experience demonstrate that this is true.
1803 props
->uniformTexelBufferOffsetAlignmentBytes
= 1;
1804 props
->uniformTexelBufferOffsetSingleTexelAlignment
= false;
1808 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_PROPERTIES_KHR
: {
1809 VkPhysicalDeviceTimelineSemaphorePropertiesKHR
*props
=
1810 (VkPhysicalDeviceTimelineSemaphorePropertiesKHR
*) ext
;
1811 props
->maxTimelineSemaphoreValueDifference
= UINT64_MAX
;
1815 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT
: {
1816 VkPhysicalDeviceTransformFeedbackPropertiesEXT
*props
=
1817 (VkPhysicalDeviceTransformFeedbackPropertiesEXT
*)ext
;
1819 props
->maxTransformFeedbackStreams
= MAX_XFB_STREAMS
;
1820 props
->maxTransformFeedbackBuffers
= MAX_XFB_BUFFERS
;
1821 props
->maxTransformFeedbackBufferSize
= (1ull << 32);
1822 props
->maxTransformFeedbackStreamDataSize
= 128 * 4;
1823 props
->maxTransformFeedbackBufferDataSize
= 128 * 4;
1824 props
->maxTransformFeedbackBufferDataStride
= 2048;
1825 props
->transformFeedbackQueries
= true;
1826 props
->transformFeedbackStreamsLinesTriangles
= false;
1827 props
->transformFeedbackRasterizationStreamSelect
= false;
1828 props
->transformFeedbackDraw
= true;
1832 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT
: {
1833 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*props
=
1834 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*)ext
;
1835 /* We have to restrict this a bit for multiview */
1836 props
->maxVertexAttribDivisor
= UINT32_MAX
/ 16;
1841 anv_debug_ignored_stype(ext
->sType
);
1847 /* We support exactly one queue family. */
1848 static const VkQueueFamilyProperties
1849 anv_queue_family_properties
= {
1850 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
1851 VK_QUEUE_COMPUTE_BIT
|
1852 VK_QUEUE_TRANSFER_BIT
,
1854 .timestampValidBits
= 36, /* XXX: Real value here */
1855 .minImageTransferGranularity
= { 1, 1, 1 },
1858 void anv_GetPhysicalDeviceQueueFamilyProperties(
1859 VkPhysicalDevice physicalDevice
,
1861 VkQueueFamilyProperties
* pQueueFamilyProperties
)
1863 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pCount
);
1865 vk_outarray_append(&out
, p
) {
1866 *p
= anv_queue_family_properties
;
1870 void anv_GetPhysicalDeviceQueueFamilyProperties2(
1871 VkPhysicalDevice physicalDevice
,
1872 uint32_t* pQueueFamilyPropertyCount
,
1873 VkQueueFamilyProperties2
* pQueueFamilyProperties
)
1876 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pQueueFamilyPropertyCount
);
1878 vk_outarray_append(&out
, p
) {
1879 p
->queueFamilyProperties
= anv_queue_family_properties
;
1881 vk_foreach_struct(s
, p
->pNext
) {
1882 anv_debug_ignored_stype(s
->sType
);
1887 void anv_GetPhysicalDeviceMemoryProperties(
1888 VkPhysicalDevice physicalDevice
,
1889 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
1891 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
1893 pMemoryProperties
->memoryTypeCount
= physical_device
->memory
.type_count
;
1894 for (uint32_t i
= 0; i
< physical_device
->memory
.type_count
; i
++) {
1895 pMemoryProperties
->memoryTypes
[i
] = (VkMemoryType
) {
1896 .propertyFlags
= physical_device
->memory
.types
[i
].propertyFlags
,
1897 .heapIndex
= physical_device
->memory
.types
[i
].heapIndex
,
1901 pMemoryProperties
->memoryHeapCount
= physical_device
->memory
.heap_count
;
1902 for (uint32_t i
= 0; i
< physical_device
->memory
.heap_count
; i
++) {
1903 pMemoryProperties
->memoryHeaps
[i
] = (VkMemoryHeap
) {
1904 .size
= physical_device
->memory
.heaps
[i
].size
,
1905 .flags
= physical_device
->memory
.heaps
[i
].flags
,
1911 anv_get_memory_budget(VkPhysicalDevice physicalDevice
,
1912 VkPhysicalDeviceMemoryBudgetPropertiesEXT
*memoryBudget
)
1914 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
1915 uint64_t sys_available
= get_available_system_memory();
1916 assert(sys_available
> 0);
1918 VkDeviceSize total_heaps_size
= 0;
1919 for (size_t i
= 0; i
< device
->memory
.heap_count
; i
++)
1920 total_heaps_size
+= device
->memory
.heaps
[i
].size
;
1922 for (size_t i
= 0; i
< device
->memory
.heap_count
; i
++) {
1923 VkDeviceSize heap_size
= device
->memory
.heaps
[i
].size
;
1924 VkDeviceSize heap_used
= device
->memory
.heaps
[i
].used
;
1925 VkDeviceSize heap_budget
;
1927 double heap_proportion
= (double) heap_size
/ total_heaps_size
;
1928 VkDeviceSize sys_available_prop
= sys_available
* heap_proportion
;
1931 * Let's not incite the app to starve the system: report at most 90% of
1932 * available system memory.
1934 uint64_t heap_available
= sys_available_prop
* 9 / 10;
1935 heap_budget
= MIN2(heap_size
, heap_used
+ heap_available
);
1938 * Round down to the nearest MB
1940 heap_budget
&= ~((1ull << 20) - 1);
1943 * The heapBudget value must be non-zero for array elements less than
1944 * VkPhysicalDeviceMemoryProperties::memoryHeapCount. The heapBudget
1945 * value must be less than or equal to VkMemoryHeap::size for each heap.
1947 assert(0 < heap_budget
&& heap_budget
<= heap_size
);
1949 memoryBudget
->heapUsage
[i
] = heap_used
;
1950 memoryBudget
->heapBudget
[i
] = heap_budget
;
1953 /* The heapBudget and heapUsage values must be zero for array elements
1954 * greater than or equal to VkPhysicalDeviceMemoryProperties::memoryHeapCount
1956 for (uint32_t i
= device
->memory
.heap_count
; i
< VK_MAX_MEMORY_HEAPS
; i
++) {
1957 memoryBudget
->heapBudget
[i
] = 0;
1958 memoryBudget
->heapUsage
[i
] = 0;
1962 void anv_GetPhysicalDeviceMemoryProperties2(
1963 VkPhysicalDevice physicalDevice
,
1964 VkPhysicalDeviceMemoryProperties2
* pMemoryProperties
)
1966 anv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
1967 &pMemoryProperties
->memoryProperties
);
1969 vk_foreach_struct(ext
, pMemoryProperties
->pNext
) {
1970 switch (ext
->sType
) {
1971 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT
:
1972 anv_get_memory_budget(physicalDevice
, (void*)ext
);
1975 anv_debug_ignored_stype(ext
->sType
);
1982 anv_GetDeviceGroupPeerMemoryFeatures(
1985 uint32_t localDeviceIndex
,
1986 uint32_t remoteDeviceIndex
,
1987 VkPeerMemoryFeatureFlags
* pPeerMemoryFeatures
)
1989 assert(localDeviceIndex
== 0 && remoteDeviceIndex
== 0);
1990 *pPeerMemoryFeatures
= VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT
|
1991 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT
|
1992 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT
|
1993 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT
;
1996 PFN_vkVoidFunction
anv_GetInstanceProcAddr(
1997 VkInstance _instance
,
2000 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2002 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
2003 * when we have to return valid function pointers, NULL, or it's left
2004 * undefined. See the table for exact details.
2009 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
2010 if (strcmp(pName, "vk" #entrypoint) == 0) \
2011 return (PFN_vkVoidFunction)anv_##entrypoint
2013 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties
);
2014 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties
);
2015 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceVersion
);
2016 LOOKUP_ANV_ENTRYPOINT(CreateInstance
);
2018 #undef LOOKUP_ANV_ENTRYPOINT
2020 if (instance
== NULL
)
2023 int idx
= anv_get_instance_entrypoint_index(pName
);
2025 return instance
->dispatch
.entrypoints
[idx
];
2027 idx
= anv_get_physical_device_entrypoint_index(pName
);
2029 return instance
->physicalDevice
.dispatch
.entrypoints
[idx
];
2031 idx
= anv_get_device_entrypoint_index(pName
);
2033 return instance
->device_dispatch
.entrypoints
[idx
];
2038 /* With version 1+ of the loader interface the ICD should expose
2039 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
2042 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
2043 VkInstance instance
,
2047 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
2048 VkInstance instance
,
2051 return anv_GetInstanceProcAddr(instance
, pName
);
2054 PFN_vkVoidFunction
anv_GetDeviceProcAddr(
2058 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2060 if (!device
|| !pName
)
2063 int idx
= anv_get_device_entrypoint_index(pName
);
2067 return device
->dispatch
.entrypoints
[idx
];
2070 /* With version 4+ of the loader interface the ICD should expose
2071 * vk_icdGetPhysicalDeviceProcAddr()
2074 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetPhysicalDeviceProcAddr(
2075 VkInstance _instance
,
2078 PFN_vkVoidFunction
vk_icdGetPhysicalDeviceProcAddr(
2079 VkInstance _instance
,
2082 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2084 if (!pName
|| !instance
)
2087 int idx
= anv_get_physical_device_entrypoint_index(pName
);
2091 return instance
->physicalDevice
.dispatch
.entrypoints
[idx
];
2096 anv_CreateDebugReportCallbackEXT(VkInstance _instance
,
2097 const VkDebugReportCallbackCreateInfoEXT
* pCreateInfo
,
2098 const VkAllocationCallbacks
* pAllocator
,
2099 VkDebugReportCallbackEXT
* pCallback
)
2101 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2102 return vk_create_debug_report_callback(&instance
->debug_report_callbacks
,
2103 pCreateInfo
, pAllocator
, &instance
->alloc
,
2108 anv_DestroyDebugReportCallbackEXT(VkInstance _instance
,
2109 VkDebugReportCallbackEXT _callback
,
2110 const VkAllocationCallbacks
* pAllocator
)
2112 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2113 vk_destroy_debug_report_callback(&instance
->debug_report_callbacks
,
2114 _callback
, pAllocator
, &instance
->alloc
);
2118 anv_DebugReportMessageEXT(VkInstance _instance
,
2119 VkDebugReportFlagsEXT flags
,
2120 VkDebugReportObjectTypeEXT objectType
,
2123 int32_t messageCode
,
2124 const char* pLayerPrefix
,
2125 const char* pMessage
)
2127 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2128 vk_debug_report(&instance
->debug_report_callbacks
, flags
, objectType
,
2129 object
, location
, messageCode
, pLayerPrefix
, pMessage
);
2132 static struct anv_state
2133 anv_state_pool_emit_data(struct anv_state_pool
*pool
, size_t size
, size_t align
, const void *p
)
2135 struct anv_state state
;
2137 state
= anv_state_pool_alloc(pool
, size
, align
);
2138 memcpy(state
.map
, p
, size
);
2143 /* Haswell border color is a bit of a disaster. Float and unorm formats use a
2144 * straightforward 32-bit float color in the first 64 bytes. Instead of using
2145 * a nice float/integer union like Gen8+, Haswell specifies the integer border
2146 * color as a separate entry /after/ the float color. The layout of this entry
2147 * also depends on the format's bpp (with extra hacks for RG32), and overlaps.
2149 * Since we don't know the format/bpp, we can't make any of the border colors
2150 * containing '1' work for all formats, as it would be in the wrong place for
2151 * some of them. We opt to make 32-bit integers work as this seems like the
2152 * most common option. Fortunately, transparent black works regardless, as
2153 * all zeroes is the same in every bit-size.
2155 struct hsw_border_color
{
2159 uint32_t _pad1
[108];
2162 struct gen8_border_color
{
2167 /* Pad out to 64 bytes */
2172 anv_device_init_border_colors(struct anv_device
*device
)
2174 if (device
->info
.is_haswell
) {
2175 static const struct hsw_border_color border_colors
[] = {
2176 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
2177 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
2178 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
2179 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
2180 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
2181 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
2184 device
->border_colors
=
2185 anv_state_pool_emit_data(&device
->dynamic_state_pool
,
2186 sizeof(border_colors
), 512, border_colors
);
2188 static const struct gen8_border_color border_colors
[] = {
2189 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
2190 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
2191 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
2192 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
2193 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
2194 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
2197 device
->border_colors
=
2198 anv_state_pool_emit_data(&device
->dynamic_state_pool
,
2199 sizeof(border_colors
), 64, border_colors
);
2204 anv_device_init_trivial_batch(struct anv_device
*device
)
2206 VkResult result
= anv_device_alloc_bo(device
, 4096,
2207 ANV_BO_ALLOC_MAPPED
,
2208 &device
->trivial_batch_bo
);
2209 if (result
!= VK_SUCCESS
)
2212 struct anv_batch batch
= {
2213 .start
= device
->trivial_batch_bo
->map
,
2214 .next
= device
->trivial_batch_bo
->map
,
2215 .end
= device
->trivial_batch_bo
->map
+ 4096,
2218 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
2219 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
2221 if (!device
->info
.has_llc
)
2222 gen_clflush_range(batch
.start
, batch
.next
- batch
.start
);
2227 VkResult
anv_EnumerateDeviceExtensionProperties(
2228 VkPhysicalDevice physicalDevice
,
2229 const char* pLayerName
,
2230 uint32_t* pPropertyCount
,
2231 VkExtensionProperties
* pProperties
)
2233 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
2234 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
2236 for (int i
= 0; i
< ANV_DEVICE_EXTENSION_COUNT
; i
++) {
2237 if (device
->supported_extensions
.extensions
[i
]) {
2238 vk_outarray_append(&out
, prop
) {
2239 *prop
= anv_device_extensions
[i
];
2244 return vk_outarray_status(&out
);
2248 anv_device_init_dispatch(struct anv_device
*device
)
2250 const struct anv_device_dispatch_table
*genX_table
;
2251 switch (device
->info
.gen
) {
2253 genX_table
= &gen12_device_dispatch_table
;
2256 genX_table
= &gen11_device_dispatch_table
;
2259 genX_table
= &gen10_device_dispatch_table
;
2262 genX_table
= &gen9_device_dispatch_table
;
2265 genX_table
= &gen8_device_dispatch_table
;
2268 if (device
->info
.is_haswell
)
2269 genX_table
= &gen75_device_dispatch_table
;
2271 genX_table
= &gen7_device_dispatch_table
;
2274 unreachable("unsupported gen\n");
2277 for (unsigned i
= 0; i
< ARRAY_SIZE(device
->dispatch
.entrypoints
); i
++) {
2278 /* Vulkan requires that entrypoints for extensions which have not been
2279 * enabled must not be advertised.
2281 if (!anv_device_entrypoint_is_enabled(i
, device
->instance
->app_info
.api_version
,
2282 &device
->instance
->enabled_extensions
,
2283 &device
->enabled_extensions
)) {
2284 device
->dispatch
.entrypoints
[i
] = NULL
;
2285 } else if (genX_table
->entrypoints
[i
]) {
2286 device
->dispatch
.entrypoints
[i
] = genX_table
->entrypoints
[i
];
2288 device
->dispatch
.entrypoints
[i
] =
2289 anv_device_dispatch_table
.entrypoints
[i
];
2295 vk_priority_to_gen(int priority
)
2298 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT
:
2299 return GEN_CONTEXT_LOW_PRIORITY
;
2300 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
:
2301 return GEN_CONTEXT_MEDIUM_PRIORITY
;
2302 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT
:
2303 return GEN_CONTEXT_HIGH_PRIORITY
;
2304 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT
:
2305 return GEN_CONTEXT_REALTIME_PRIORITY
;
2307 unreachable("Invalid priority");
2312 anv_device_init_hiz_clear_value_bo(struct anv_device
*device
)
2314 VkResult result
= anv_device_alloc_bo(device
, 4096,
2315 ANV_BO_ALLOC_MAPPED
,
2316 &device
->hiz_clear_bo
);
2317 if (result
!= VK_SUCCESS
)
2320 union isl_color_value hiz_clear
= { .u32
= { 0, } };
2321 hiz_clear
.f32
[0] = ANV_HZ_FC_VAL
;
2323 memcpy(device
->hiz_clear_bo
->map
, hiz_clear
.u32
, sizeof(hiz_clear
.u32
));
2325 if (!device
->info
.has_llc
)
2326 gen_clflush_range(device
->hiz_clear_bo
->map
, sizeof(hiz_clear
.u32
));
2332 get_bo_from_pool(struct gen_batch_decode_bo
*ret
,
2333 struct anv_block_pool
*pool
,
2336 anv_block_pool_foreach_bo(bo
, pool
) {
2337 uint64_t bo_address
= gen_48b_address(bo
->offset
);
2338 if (address
>= bo_address
&& address
< (bo_address
+ bo
->size
)) {
2339 *ret
= (struct gen_batch_decode_bo
) {
2350 /* Finding a buffer for batch decoding */
2351 static struct gen_batch_decode_bo
2352 decode_get_bo(void *v_batch
, bool ppgtt
, uint64_t address
)
2354 struct anv_device
*device
= v_batch
;
2355 struct gen_batch_decode_bo ret_bo
= {};
2359 if (get_bo_from_pool(&ret_bo
, &device
->dynamic_state_pool
.block_pool
, address
))
2361 if (get_bo_from_pool(&ret_bo
, &device
->instruction_state_pool
.block_pool
, address
))
2363 if (get_bo_from_pool(&ret_bo
, &device
->binding_table_pool
.block_pool
, address
))
2365 if (get_bo_from_pool(&ret_bo
, &device
->surface_state_pool
.block_pool
, address
))
2368 if (!device
->cmd_buffer_being_decoded
)
2369 return (struct gen_batch_decode_bo
) { };
2371 struct anv_batch_bo
**bo
;
2373 u_vector_foreach(bo
, &device
->cmd_buffer_being_decoded
->seen_bbos
) {
2374 /* The decoder zeroes out the top 16 bits, so we need to as well */
2375 uint64_t bo_address
= (*bo
)->bo
->offset
& (~0ull >> 16);
2377 if (address
>= bo_address
&& address
< bo_address
+ (*bo
)->bo
->size
) {
2378 return (struct gen_batch_decode_bo
) {
2380 .size
= (*bo
)->bo
->size
,
2381 .map
= (*bo
)->bo
->map
,
2386 return (struct gen_batch_decode_bo
) { };
2389 struct gen_aux_map_buffer
{
2390 struct gen_buffer base
;
2391 struct anv_state state
;
2394 static struct gen_buffer
*
2395 gen_aux_map_buffer_alloc(void *driver_ctx
, uint32_t size
)
2397 struct gen_aux_map_buffer
*buf
= malloc(sizeof(struct gen_aux_map_buffer
));
2401 struct anv_device
*device
= (struct anv_device
*)driver_ctx
;
2402 assert(device
->instance
->physicalDevice
.supports_48bit_addresses
&&
2403 device
->instance
->physicalDevice
.use_softpin
);
2405 struct anv_state_pool
*pool
= &device
->dynamic_state_pool
;
2406 buf
->state
= anv_state_pool_alloc(pool
, size
, size
);
2408 buf
->base
.gpu
= pool
->block_pool
.bo
->offset
+ buf
->state
.offset
;
2409 buf
->base
.gpu_end
= buf
->base
.gpu
+ buf
->state
.alloc_size
;
2410 buf
->base
.map
= buf
->state
.map
;
2411 buf
->base
.driver_bo
= &buf
->state
;
2416 gen_aux_map_buffer_free(void *driver_ctx
, struct gen_buffer
*buffer
)
2418 struct gen_aux_map_buffer
*buf
= (struct gen_aux_map_buffer
*)buffer
;
2419 struct anv_device
*device
= (struct anv_device
*)driver_ctx
;
2420 struct anv_state_pool
*pool
= &device
->dynamic_state_pool
;
2421 anv_state_pool_free(pool
, buf
->state
);
2425 static struct gen_mapped_pinned_buffer_alloc aux_map_allocator
= {
2426 .alloc
= gen_aux_map_buffer_alloc
,
2427 .free
= gen_aux_map_buffer_free
,
2430 VkResult
anv_CreateDevice(
2431 VkPhysicalDevice physicalDevice
,
2432 const VkDeviceCreateInfo
* pCreateInfo
,
2433 const VkAllocationCallbacks
* pAllocator
,
2436 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
2438 struct anv_device
*device
;
2440 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO
);
2442 struct anv_device_extension_table enabled_extensions
= { };
2443 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
2445 for (idx
= 0; idx
< ANV_DEVICE_EXTENSION_COUNT
; idx
++) {
2446 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
2447 anv_device_extensions
[idx
].extensionName
) == 0)
2451 if (idx
>= ANV_DEVICE_EXTENSION_COUNT
)
2452 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
2454 if (!physical_device
->supported_extensions
.extensions
[idx
])
2455 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
2457 enabled_extensions
.extensions
[idx
] = true;
2460 /* Check enabled features */
2461 if (pCreateInfo
->pEnabledFeatures
) {
2462 VkPhysicalDeviceFeatures supported_features
;
2463 anv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
2464 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
2465 VkBool32
*enabled_feature
= (VkBool32
*)pCreateInfo
->pEnabledFeatures
;
2466 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
2467 for (uint32_t i
= 0; i
< num_features
; i
++) {
2468 if (enabled_feature
[i
] && !supported_feature
[i
])
2469 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
2473 /* Check requested queues and fail if we are requested to create any
2474 * queues with flags we don't support.
2476 assert(pCreateInfo
->queueCreateInfoCount
> 0);
2477 for (uint32_t i
= 0; i
< pCreateInfo
->queueCreateInfoCount
; i
++) {
2478 if (pCreateInfo
->pQueueCreateInfos
[i
].flags
!= 0)
2479 return vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2482 /* Check if client specified queue priority. */
2483 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*queue_priority
=
2484 vk_find_struct_const(pCreateInfo
->pQueueCreateInfos
[0].pNext
,
2485 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT
);
2487 VkQueueGlobalPriorityEXT priority
=
2488 queue_priority
? queue_priority
->globalPriority
:
2489 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
;
2491 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
2493 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
2495 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2497 if (INTEL_DEBUG
& DEBUG_BATCH
) {
2498 const unsigned decode_flags
=
2499 GEN_BATCH_DECODE_FULL
|
2500 ((INTEL_DEBUG
& DEBUG_COLOR
) ? GEN_BATCH_DECODE_IN_COLOR
: 0) |
2501 GEN_BATCH_DECODE_OFFSETS
|
2502 GEN_BATCH_DECODE_FLOATS
;
2504 gen_batch_decode_ctx_init(&device
->decoder_ctx
,
2505 &physical_device
->info
,
2506 stderr
, decode_flags
, NULL
,
2507 decode_get_bo
, NULL
, device
);
2510 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
2511 device
->instance
= physical_device
->instance
;
2512 device
->chipset_id
= physical_device
->chipset_id
;
2513 device
->no_hw
= physical_device
->no_hw
;
2514 device
->_lost
= false;
2517 device
->alloc
= *pAllocator
;
2519 device
->alloc
= physical_device
->instance
->alloc
;
2521 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
2522 device
->fd
= open(physical_device
->path
, O_RDWR
| O_CLOEXEC
);
2523 if (device
->fd
== -1) {
2524 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2528 device
->context_id
= anv_gem_create_context(device
);
2529 if (device
->context_id
== -1) {
2530 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2534 result
= anv_queue_init(device
, &device
->queue
);
2535 if (result
!= VK_SUCCESS
)
2536 goto fail_context_id
;
2538 if (physical_device
->use_softpin
) {
2539 if (pthread_mutex_init(&device
->vma_mutex
, NULL
) != 0) {
2540 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2544 /* keep the page with address zero out of the allocator */
2545 struct anv_memory_heap
*low_heap
=
2546 &physical_device
->memory
.heaps
[physical_device
->memory
.heap_count
- 1];
2547 util_vma_heap_init(&device
->vma_lo
, low_heap
->vma_start
, low_heap
->vma_size
);
2548 device
->vma_lo_available
= low_heap
->size
;
2550 struct anv_memory_heap
*high_heap
=
2551 &physical_device
->memory
.heaps
[0];
2552 util_vma_heap_init(&device
->vma_hi
, high_heap
->vma_start
, high_heap
->vma_size
);
2553 device
->vma_hi_available
= physical_device
->memory
.heap_count
== 1 ? 0 :
2557 list_inithead(&device
->memory_objects
);
2559 /* As per spec, the driver implementation may deny requests to acquire
2560 * a priority above the default priority (MEDIUM) if the caller does not
2561 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
2564 if (physical_device
->has_context_priority
) {
2565 int err
= anv_gem_set_context_param(device
->fd
, device
->context_id
,
2566 I915_CONTEXT_PARAM_PRIORITY
,
2567 vk_priority_to_gen(priority
));
2568 if (err
!= 0 && priority
> VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
) {
2569 result
= vk_error(VK_ERROR_NOT_PERMITTED_EXT
);
2574 device
->info
= physical_device
->info
;
2575 device
->isl_dev
= physical_device
->isl_dev
;
2577 /* On Broadwell and later, we can use batch chaining to more efficiently
2578 * implement growing command buffers. Prior to Haswell, the kernel
2579 * command parser gets in the way and we have to fall back to growing
2582 device
->can_chain_batches
= device
->info
.gen
>= 8;
2584 device
->robust_buffer_access
= pCreateInfo
->pEnabledFeatures
&&
2585 pCreateInfo
->pEnabledFeatures
->robustBufferAccess
;
2586 device
->enabled_extensions
= enabled_extensions
;
2588 anv_device_init_dispatch(device
);
2590 if (pthread_mutex_init(&device
->mutex
, NULL
) != 0) {
2591 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2595 pthread_condattr_t condattr
;
2596 if (pthread_condattr_init(&condattr
) != 0) {
2597 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2600 if (pthread_condattr_setclock(&condattr
, CLOCK_MONOTONIC
) != 0) {
2601 pthread_condattr_destroy(&condattr
);
2602 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2605 if (pthread_cond_init(&device
->queue_submit
, &condattr
) != 0) {
2606 pthread_condattr_destroy(&condattr
);
2607 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2610 pthread_condattr_destroy(&condattr
);
2613 (physical_device
->supports_48bit_addresses
? EXEC_OBJECT_SUPPORTS_48B_ADDRESS
: 0) |
2614 (physical_device
->has_exec_async
? EXEC_OBJECT_ASYNC
: 0) |
2615 (physical_device
->has_exec_capture
? EXEC_OBJECT_CAPTURE
: 0) |
2616 (physical_device
->use_softpin
? EXEC_OBJECT_PINNED
: 0);
2618 result
= anv_bo_cache_init(&device
->bo_cache
);
2619 if (result
!= VK_SUCCESS
)
2620 goto fail_queue_cond
;
2622 anv_bo_pool_init(&device
->batch_bo_pool
, device
, bo_flags
);
2624 result
= anv_state_pool_init(&device
->dynamic_state_pool
, device
,
2625 DYNAMIC_STATE_POOL_MIN_ADDRESS
, 16384);
2626 if (result
!= VK_SUCCESS
)
2627 goto fail_batch_bo_pool
;
2629 result
= anv_state_pool_init(&device
->instruction_state_pool
, device
,
2630 INSTRUCTION_STATE_POOL_MIN_ADDRESS
, 16384);
2631 if (result
!= VK_SUCCESS
)
2632 goto fail_dynamic_state_pool
;
2634 result
= anv_state_pool_init(&device
->surface_state_pool
, device
,
2635 SURFACE_STATE_POOL_MIN_ADDRESS
, 4096);
2636 if (result
!= VK_SUCCESS
)
2637 goto fail_instruction_state_pool
;
2639 if (physical_device
->use_softpin
) {
2640 result
= anv_state_pool_init(&device
->binding_table_pool
, device
,
2641 BINDING_TABLE_POOL_MIN_ADDRESS
, 4096);
2642 if (result
!= VK_SUCCESS
)
2643 goto fail_surface_state_pool
;
2646 if (device
->info
.gen
>= 12) {
2647 device
->aux_map_ctx
= gen_aux_map_init(device
, &aux_map_allocator
,
2648 &physical_device
->info
);
2649 if (!device
->aux_map_ctx
)
2650 goto fail_binding_table_pool
;
2653 result
= anv_device_alloc_bo(device
, 4096, 0, &device
->workaround_bo
);
2654 if (result
!= VK_SUCCESS
)
2655 goto fail_surface_aux_map_pool
;
2657 result
= anv_device_init_trivial_batch(device
);
2658 if (result
!= VK_SUCCESS
)
2659 goto fail_workaround_bo
;
2661 if (device
->info
.gen
>= 10) {
2662 result
= anv_device_init_hiz_clear_value_bo(device
);
2663 if (result
!= VK_SUCCESS
)
2664 goto fail_trivial_batch_bo
;
2667 anv_scratch_pool_init(device
, &device
->scratch_pool
);
2669 switch (device
->info
.gen
) {
2671 if (!device
->info
.is_haswell
)
2672 result
= gen7_init_device_state(device
);
2674 result
= gen75_init_device_state(device
);
2677 result
= gen8_init_device_state(device
);
2680 result
= gen9_init_device_state(device
);
2683 result
= gen10_init_device_state(device
);
2686 result
= gen11_init_device_state(device
);
2689 result
= gen12_init_device_state(device
);
2692 /* Shouldn't get here as we don't create physical devices for any other
2694 unreachable("unhandled gen");
2696 if (result
!= VK_SUCCESS
)
2697 goto fail_workaround_bo
;
2699 anv_pipeline_cache_init(&device
->default_pipeline_cache
, device
, true);
2701 anv_device_init_blorp(device
);
2703 anv_device_init_border_colors(device
);
2705 anv_device_perf_init(device
);
2707 *pDevice
= anv_device_to_handle(device
);
2712 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
2713 if (device
->info
.gen
>= 10)
2714 anv_device_release_bo(device
, device
->hiz_clear_bo
);
2715 anv_device_release_bo(device
, device
->workaround_bo
);
2716 fail_trivial_batch_bo
:
2717 anv_device_release_bo(device
, device
->trivial_batch_bo
);
2718 fail_surface_aux_map_pool
:
2719 if (device
->info
.gen
>= 12) {
2720 gen_aux_map_finish(device
->aux_map_ctx
);
2721 device
->aux_map_ctx
= NULL
;
2723 fail_binding_table_pool
:
2724 if (physical_device
->use_softpin
)
2725 anv_state_pool_finish(&device
->binding_table_pool
);
2726 fail_surface_state_pool
:
2727 anv_state_pool_finish(&device
->surface_state_pool
);
2728 fail_instruction_state_pool
:
2729 anv_state_pool_finish(&device
->instruction_state_pool
);
2730 fail_dynamic_state_pool
:
2731 anv_state_pool_finish(&device
->dynamic_state_pool
);
2733 anv_bo_pool_finish(&device
->batch_bo_pool
);
2734 anv_bo_cache_finish(&device
->bo_cache
);
2736 pthread_cond_destroy(&device
->queue_submit
);
2738 pthread_mutex_destroy(&device
->mutex
);
2740 if (physical_device
->use_softpin
) {
2741 util_vma_heap_finish(&device
->vma_hi
);
2742 util_vma_heap_finish(&device
->vma_lo
);
2745 anv_queue_finish(&device
->queue
);
2747 anv_gem_destroy_context(device
, device
->context_id
);
2751 vk_free(&device
->alloc
, device
);
2756 void anv_DestroyDevice(
2758 const VkAllocationCallbacks
* pAllocator
)
2760 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2761 struct anv_physical_device
*physical_device
;
2766 physical_device
= &device
->instance
->physicalDevice
;
2768 anv_device_finish_blorp(device
);
2770 anv_pipeline_cache_finish(&device
->default_pipeline_cache
);
2772 anv_queue_finish(&device
->queue
);
2774 #ifdef HAVE_VALGRIND
2775 /* We only need to free these to prevent valgrind errors. The backing
2776 * BO will go away in a couple of lines so we don't actually leak.
2778 anv_state_pool_free(&device
->dynamic_state_pool
, device
->border_colors
);
2779 anv_state_pool_free(&device
->dynamic_state_pool
, device
->slice_hash
);
2782 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
2784 anv_device_release_bo(device
, device
->workaround_bo
);
2785 anv_device_release_bo(device
, device
->trivial_batch_bo
);
2786 if (device
->info
.gen
>= 10)
2787 anv_device_release_bo(device
, device
->hiz_clear_bo
);
2789 if (device
->info
.gen
>= 12) {
2790 gen_aux_map_finish(device
->aux_map_ctx
);
2791 device
->aux_map_ctx
= NULL
;
2794 if (physical_device
->use_softpin
)
2795 anv_state_pool_finish(&device
->binding_table_pool
);
2796 anv_state_pool_finish(&device
->surface_state_pool
);
2797 anv_state_pool_finish(&device
->instruction_state_pool
);
2798 anv_state_pool_finish(&device
->dynamic_state_pool
);
2800 anv_bo_pool_finish(&device
->batch_bo_pool
);
2802 anv_bo_cache_finish(&device
->bo_cache
);
2804 if (physical_device
->use_softpin
) {
2805 util_vma_heap_finish(&device
->vma_hi
);
2806 util_vma_heap_finish(&device
->vma_lo
);
2809 pthread_cond_destroy(&device
->queue_submit
);
2810 pthread_mutex_destroy(&device
->mutex
);
2812 anv_gem_destroy_context(device
, device
->context_id
);
2814 if (INTEL_DEBUG
& DEBUG_BATCH
)
2815 gen_batch_decode_ctx_finish(&device
->decoder_ctx
);
2819 vk_free(&device
->alloc
, device
);
2822 VkResult
anv_EnumerateInstanceLayerProperties(
2823 uint32_t* pPropertyCount
,
2824 VkLayerProperties
* pProperties
)
2826 if (pProperties
== NULL
) {
2827 *pPropertyCount
= 0;
2831 /* None supported at this time */
2832 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
2835 VkResult
anv_EnumerateDeviceLayerProperties(
2836 VkPhysicalDevice physicalDevice
,
2837 uint32_t* pPropertyCount
,
2838 VkLayerProperties
* pProperties
)
2840 if (pProperties
== NULL
) {
2841 *pPropertyCount
= 0;
2845 /* None supported at this time */
2846 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
2849 void anv_GetDeviceQueue(
2851 uint32_t queueNodeIndex
,
2852 uint32_t queueIndex
,
2855 const VkDeviceQueueInfo2 info
= {
2856 .sType
= VK_STRUCTURE_TYPE_DEVICE_QUEUE_INFO_2
,
2859 .queueFamilyIndex
= queueNodeIndex
,
2860 .queueIndex
= queueIndex
,
2863 anv_GetDeviceQueue2(_device
, &info
, pQueue
);
2866 void anv_GetDeviceQueue2(
2868 const VkDeviceQueueInfo2
* pQueueInfo
,
2871 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2873 assert(pQueueInfo
->queueIndex
== 0);
2875 if (pQueueInfo
->flags
== device
->queue
.flags
)
2876 *pQueue
= anv_queue_to_handle(&device
->queue
);
2882 _anv_device_set_lost(struct anv_device
*device
,
2883 const char *file
, int line
,
2884 const char *msg
, ...)
2889 p_atomic_inc(&device
->_lost
);
2892 err
= __vk_errorv(device
->instance
, device
,
2893 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT
,
2894 VK_ERROR_DEVICE_LOST
, file
, line
, msg
, ap
);
2897 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
2904 _anv_queue_set_lost(struct anv_queue
*queue
,
2905 const char *file
, int line
,
2906 const char *msg
, ...)
2911 p_atomic_inc(&queue
->device
->_lost
);
2914 err
= __vk_errorv(queue
->device
->instance
, queue
->device
,
2915 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT
,
2916 VK_ERROR_DEVICE_LOST
, file
, line
, msg
, ap
);
2919 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
2926 anv_device_query_status(struct anv_device
*device
)
2928 /* This isn't likely as most of the callers of this function already check
2929 * for it. However, it doesn't hurt to check and it potentially lets us
2932 if (anv_device_is_lost(device
))
2933 return VK_ERROR_DEVICE_LOST
;
2935 uint32_t active
, pending
;
2936 int ret
= anv_gem_gpu_get_reset_stats(device
, &active
, &pending
);
2938 /* We don't know the real error. */
2939 return anv_device_set_lost(device
, "get_reset_stats failed: %m");
2943 return anv_device_set_lost(device
, "GPU hung on one of our command buffers");
2944 } else if (pending
) {
2945 return anv_device_set_lost(device
, "GPU hung with commands in-flight");
2952 anv_device_bo_busy(struct anv_device
*device
, struct anv_bo
*bo
)
2954 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
2955 * Other usages of the BO (such as on different hardware) will not be
2956 * flagged as "busy" by this ioctl. Use with care.
2958 int ret
= anv_gem_busy(device
, bo
->gem_handle
);
2960 return VK_NOT_READY
;
2961 } else if (ret
== -1) {
2962 /* We don't know the real error. */
2963 return anv_device_set_lost(device
, "gem wait failed: %m");
2966 /* Query for device status after the busy call. If the BO we're checking
2967 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
2968 * client because it clearly doesn't have valid data. Yes, this most
2969 * likely means an ioctl, but we just did an ioctl to query the busy status
2970 * so it's no great loss.
2972 return anv_device_query_status(device
);
2976 anv_device_wait(struct anv_device
*device
, struct anv_bo
*bo
,
2979 int ret
= anv_gem_wait(device
, bo
->gem_handle
, &timeout
);
2980 if (ret
== -1 && errno
== ETIME
) {
2982 } else if (ret
== -1) {
2983 /* We don't know the real error. */
2984 return anv_device_set_lost(device
, "gem wait failed: %m");
2987 /* Query for device status after the wait. If the BO we're waiting on got
2988 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
2989 * because it clearly doesn't have valid data. Yes, this most likely means
2990 * an ioctl, but we just did an ioctl to wait so it's no great loss.
2992 return anv_device_query_status(device
);
2995 VkResult
anv_DeviceWaitIdle(
2998 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3000 if (anv_device_is_lost(device
))
3001 return VK_ERROR_DEVICE_LOST
;
3003 return anv_queue_submit_simple_batch(&device
->queue
, NULL
);
3007 anv_vma_alloc(struct anv_device
*device
, struct anv_bo
*bo
)
3009 if (!(bo
->flags
& EXEC_OBJECT_PINNED
))
3012 pthread_mutex_lock(&device
->vma_mutex
);
3016 if (bo
->flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
&&
3017 device
->vma_hi_available
>= bo
->size
) {
3018 uint64_t addr
= util_vma_heap_alloc(&device
->vma_hi
, bo
->size
, 4096);
3020 bo
->offset
= gen_canonical_address(addr
);
3021 assert(addr
== gen_48b_address(bo
->offset
));
3022 device
->vma_hi_available
-= bo
->size
;
3026 if (bo
->offset
== 0 && device
->vma_lo_available
>= bo
->size
) {
3027 uint64_t addr
= util_vma_heap_alloc(&device
->vma_lo
, bo
->size
, 4096);
3029 bo
->offset
= gen_canonical_address(addr
);
3030 assert(addr
== gen_48b_address(bo
->offset
));
3031 device
->vma_lo_available
-= bo
->size
;
3035 pthread_mutex_unlock(&device
->vma_mutex
);
3037 return bo
->offset
!= 0;
3041 anv_vma_free(struct anv_device
*device
, struct anv_bo
*bo
)
3043 if (!(bo
->flags
& EXEC_OBJECT_PINNED
))
3046 const uint64_t addr_48b
= gen_48b_address(bo
->offset
);
3048 pthread_mutex_lock(&device
->vma_mutex
);
3050 if (addr_48b
>= LOW_HEAP_MIN_ADDRESS
&&
3051 addr_48b
<= LOW_HEAP_MAX_ADDRESS
) {
3052 util_vma_heap_free(&device
->vma_lo
, addr_48b
, bo
->size
);
3053 device
->vma_lo_available
+= bo
->size
;
3055 ASSERTED
const struct anv_physical_device
*physical_device
=
3056 &device
->instance
->physicalDevice
;
3057 assert(addr_48b
>= physical_device
->memory
.heaps
[0].vma_start
&&
3058 addr_48b
< (physical_device
->memory
.heaps
[0].vma_start
+
3059 physical_device
->memory
.heaps
[0].vma_size
));
3060 util_vma_heap_free(&device
->vma_hi
, addr_48b
, bo
->size
);
3061 device
->vma_hi_available
+= bo
->size
;
3064 pthread_mutex_unlock(&device
->vma_mutex
);
3069 VkResult
anv_AllocateMemory(
3071 const VkMemoryAllocateInfo
* pAllocateInfo
,
3072 const VkAllocationCallbacks
* pAllocator
,
3073 VkDeviceMemory
* pMem
)
3075 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3076 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3077 struct anv_device_memory
*mem
;
3078 VkResult result
= VK_SUCCESS
;
3080 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
3082 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
3083 assert(pAllocateInfo
->allocationSize
> 0);
3085 if (pAllocateInfo
->allocationSize
> MAX_MEMORY_ALLOCATION_SIZE
)
3086 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
3088 /* FINISHME: Fail if allocation request exceeds heap size. */
3090 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
3091 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3093 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3095 assert(pAllocateInfo
->memoryTypeIndex
< pdevice
->memory
.type_count
);
3096 mem
->type
= &pdevice
->memory
.types
[pAllocateInfo
->memoryTypeIndex
];
3100 mem
->host_ptr
= NULL
;
3102 enum anv_bo_alloc_flags alloc_flags
= 0;
3104 assert(mem
->type
->heapIndex
< pdevice
->memory
.heap_count
);
3105 if (!pdevice
->memory
.heaps
[mem
->type
->heapIndex
].supports_48bit_addresses
)
3106 alloc_flags
|= ANV_BO_ALLOC_32BIT_ADDRESS
;
3108 const struct wsi_memory_allocate_info
*wsi_info
=
3109 vk_find_struct_const(pAllocateInfo
->pNext
, WSI_MEMORY_ALLOCATE_INFO_MESA
);
3110 if (wsi_info
&& wsi_info
->implicit_sync
) {
3111 /* We need to set the WRITE flag on window system buffers so that GEM
3112 * will know we're writing to them and synchronize uses on other rings
3113 * (eg if the display server uses the blitter ring).
3115 alloc_flags
|= ANV_BO_ALLOC_IMPLICIT_SYNC
|
3116 ANV_BO_ALLOC_IMPLICIT_WRITE
;
3119 const VkExportMemoryAllocateInfo
*export_info
=
3120 vk_find_struct_const(pAllocateInfo
->pNext
, EXPORT_MEMORY_ALLOCATE_INFO
);
3122 /* Check if we need to support Android HW buffer export. If so,
3123 * create AHardwareBuffer and import memory from it.
3125 bool android_export
= false;
3126 if (export_info
&& export_info
->handleTypes
&
3127 VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID
)
3128 android_export
= true;
3130 /* Android memory import. */
3131 const struct VkImportAndroidHardwareBufferInfoANDROID
*ahw_import_info
=
3132 vk_find_struct_const(pAllocateInfo
->pNext
,
3133 IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID
);
3135 if (ahw_import_info
) {
3136 result
= anv_import_ahw_memory(_device
, mem
, ahw_import_info
);
3137 if (result
!= VK_SUCCESS
)
3141 } else if (android_export
) {
3142 result
= anv_create_ahw_memory(_device
, mem
, pAllocateInfo
);
3143 if (result
!= VK_SUCCESS
)
3146 const struct VkImportAndroidHardwareBufferInfoANDROID import_info
= {
3149 result
= anv_import_ahw_memory(_device
, mem
, &import_info
);
3150 if (result
!= VK_SUCCESS
)
3156 const VkImportMemoryFdInfoKHR
*fd_info
=
3157 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_FD_INFO_KHR
);
3159 /* The Vulkan spec permits handleType to be 0, in which case the struct is
3162 if (fd_info
&& fd_info
->handleType
) {
3163 /* At the moment, we support only the below handle types. */
3164 assert(fd_info
->handleType
==
3165 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
3166 fd_info
->handleType
==
3167 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
3169 result
= anv_device_import_bo(device
, fd_info
->fd
, alloc_flags
,
3171 if (result
!= VK_SUCCESS
)
3174 VkDeviceSize aligned_alloc_size
=
3175 align_u64(pAllocateInfo
->allocationSize
, 4096);
3177 /* For security purposes, we reject importing the bo if it's smaller
3178 * than the requested allocation size. This prevents a malicious client
3179 * from passing a buffer to a trusted client, lying about the size, and
3180 * telling the trusted client to try and texture from an image that goes
3181 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
3182 * in the trusted client. The trusted client can protect itself against
3183 * this sort of attack but only if it can trust the buffer size.
3185 if (mem
->bo
->size
< aligned_alloc_size
) {
3186 result
= vk_errorf(device
->instance
, device
,
3187 VK_ERROR_INVALID_EXTERNAL_HANDLE
,
3188 "aligned allocationSize too large for "
3189 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT: "
3190 "%"PRIu64
"B > %"PRIu64
"B",
3191 aligned_alloc_size
, mem
->bo
->size
);
3192 anv_device_release_bo(device
, mem
->bo
);
3196 /* From the Vulkan spec:
3198 * "Importing memory from a file descriptor transfers ownership of
3199 * the file descriptor from the application to the Vulkan
3200 * implementation. The application must not perform any operations on
3201 * the file descriptor after a successful import."
3203 * If the import fails, we leave the file descriptor open.
3209 const VkImportMemoryHostPointerInfoEXT
*host_ptr_info
=
3210 vk_find_struct_const(pAllocateInfo
->pNext
,
3211 IMPORT_MEMORY_HOST_POINTER_INFO_EXT
);
3212 if (host_ptr_info
&& host_ptr_info
->handleType
) {
3213 if (host_ptr_info
->handleType
==
3214 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_MAPPED_FOREIGN_MEMORY_BIT_EXT
) {
3215 result
= vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
3219 assert(host_ptr_info
->handleType
==
3220 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
);
3222 result
= anv_device_import_bo_from_host_ptr(device
,
3223 host_ptr_info
->pHostPointer
,
3224 pAllocateInfo
->allocationSize
,
3228 if (result
!= VK_SUCCESS
)
3231 mem
->host_ptr
= host_ptr_info
->pHostPointer
;
3235 /* Regular allocate (not importing memory). */
3237 if (export_info
&& export_info
->handleTypes
)
3238 alloc_flags
|= ANV_BO_ALLOC_EXTERNAL
;
3240 result
= anv_device_alloc_bo(device
, pAllocateInfo
->allocationSize
,
3241 alloc_flags
, &mem
->bo
);
3242 if (result
!= VK_SUCCESS
)
3245 const VkMemoryDedicatedAllocateInfo
*dedicated_info
=
3246 vk_find_struct_const(pAllocateInfo
->pNext
, MEMORY_DEDICATED_ALLOCATE_INFO
);
3247 if (dedicated_info
&& dedicated_info
->image
!= VK_NULL_HANDLE
) {
3248 ANV_FROM_HANDLE(anv_image
, image
, dedicated_info
->image
);
3250 /* Some legacy (non-modifiers) consumers need the tiling to be set on
3251 * the BO. In this case, we have a dedicated allocation.
3253 if (image
->needs_set_tiling
) {
3254 const uint32_t i915_tiling
=
3255 isl_tiling_to_i915_tiling(image
->planes
[0].surface
.isl
.tiling
);
3256 int ret
= anv_gem_set_tiling(device
, mem
->bo
->gem_handle
,
3257 image
->planes
[0].surface
.isl
.row_pitch_B
,
3260 anv_device_release_bo(device
, mem
->bo
);
3261 return vk_errorf(device
->instance
, NULL
,
3262 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
3263 "failed to set BO tiling: %m");
3269 pthread_mutex_lock(&device
->mutex
);
3270 list_addtail(&mem
->link
, &device
->memory_objects
);
3271 pthread_mutex_unlock(&device
->mutex
);
3273 *pMem
= anv_device_memory_to_handle(mem
);
3275 p_atomic_add(&pdevice
->memory
.heaps
[mem
->type
->heapIndex
].used
,
3281 vk_free2(&device
->alloc
, pAllocator
, mem
);
3286 VkResult
anv_GetMemoryFdKHR(
3288 const VkMemoryGetFdInfoKHR
* pGetFdInfo
,
3291 ANV_FROM_HANDLE(anv_device
, dev
, device_h
);
3292 ANV_FROM_HANDLE(anv_device_memory
, mem
, pGetFdInfo
->memory
);
3294 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
3296 assert(pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
3297 pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
3299 return anv_device_export_bo(dev
, mem
->bo
, pFd
);
3302 VkResult
anv_GetMemoryFdPropertiesKHR(
3304 VkExternalMemoryHandleTypeFlagBits handleType
,
3306 VkMemoryFdPropertiesKHR
* pMemoryFdProperties
)
3308 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3309 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3311 switch (handleType
) {
3312 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
:
3313 /* dma-buf can be imported as any memory type */
3314 pMemoryFdProperties
->memoryTypeBits
=
3315 (1 << pdevice
->memory
.type_count
) - 1;
3319 /* The valid usage section for this function says:
3321 * "handleType must not be one of the handle types defined as
3324 * So opaque handle types fall into the default "unsupported" case.
3326 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
3330 VkResult
anv_GetMemoryHostPointerPropertiesEXT(
3332 VkExternalMemoryHandleTypeFlagBits handleType
,
3333 const void* pHostPointer
,
3334 VkMemoryHostPointerPropertiesEXT
* pMemoryHostPointerProperties
)
3336 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3338 assert(pMemoryHostPointerProperties
->sType
==
3339 VK_STRUCTURE_TYPE_MEMORY_HOST_POINTER_PROPERTIES_EXT
);
3341 switch (handleType
) {
3342 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
: {
3343 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3345 /* Host memory can be imported as any memory type. */
3346 pMemoryHostPointerProperties
->memoryTypeBits
=
3347 (1ull << pdevice
->memory
.type_count
) - 1;
3352 return VK_ERROR_INVALID_EXTERNAL_HANDLE
;
3356 void anv_FreeMemory(
3358 VkDeviceMemory _mem
,
3359 const VkAllocationCallbacks
* pAllocator
)
3361 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3362 ANV_FROM_HANDLE(anv_device_memory
, mem
, _mem
);
3363 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3368 pthread_mutex_lock(&device
->mutex
);
3369 list_del(&mem
->link
);
3370 pthread_mutex_unlock(&device
->mutex
);
3373 anv_UnmapMemory(_device
, _mem
);
3375 p_atomic_add(&pdevice
->memory
.heaps
[mem
->type
->heapIndex
].used
,
3378 anv_device_release_bo(device
, mem
->bo
);
3380 #if defined(ANDROID) && ANDROID_API_LEVEL >= 26
3382 AHardwareBuffer_release(mem
->ahw
);
3385 vk_free2(&device
->alloc
, pAllocator
, mem
);
3388 VkResult
anv_MapMemory(
3390 VkDeviceMemory _memory
,
3391 VkDeviceSize offset
,
3393 VkMemoryMapFlags flags
,
3396 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3397 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
3404 if (mem
->host_ptr
) {
3405 *ppData
= mem
->host_ptr
+ offset
;
3409 if (size
== VK_WHOLE_SIZE
)
3410 size
= mem
->bo
->size
- offset
;
3412 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
3414 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
3415 * assert(size != 0);
3416 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
3417 * equal to the size of the memory minus offset
3420 assert(offset
+ size
<= mem
->bo
->size
);
3422 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
3423 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
3424 * at a time is valid. We could just mmap up front and return an offset
3425 * pointer here, but that may exhaust virtual memory on 32 bit
3428 uint32_t gem_flags
= 0;
3430 if (!device
->info
.has_llc
&&
3431 (mem
->type
->propertyFlags
& VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
))
3432 gem_flags
|= I915_MMAP_WC
;
3434 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
3435 uint64_t map_offset
= offset
& ~4095ull;
3436 assert(offset
>= map_offset
);
3437 uint64_t map_size
= (offset
+ size
) - map_offset
;
3439 /* Let's map whole pages */
3440 map_size
= align_u64(map_size
, 4096);
3442 void *map
= anv_gem_mmap(device
, mem
->bo
->gem_handle
,
3443 map_offset
, map_size
, gem_flags
);
3444 if (map
== MAP_FAILED
)
3445 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
3448 mem
->map_size
= map_size
;
3450 *ppData
= mem
->map
+ (offset
- map_offset
);
3455 void anv_UnmapMemory(
3457 VkDeviceMemory _memory
)
3459 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
3461 if (mem
== NULL
|| mem
->host_ptr
)
3464 anv_gem_munmap(mem
->map
, mem
->map_size
);
3471 clflush_mapped_ranges(struct anv_device
*device
,
3473 const VkMappedMemoryRange
*ranges
)
3475 for (uint32_t i
= 0; i
< count
; i
++) {
3476 ANV_FROM_HANDLE(anv_device_memory
, mem
, ranges
[i
].memory
);
3477 if (ranges
[i
].offset
>= mem
->map_size
)
3480 gen_clflush_range(mem
->map
+ ranges
[i
].offset
,
3481 MIN2(ranges
[i
].size
, mem
->map_size
- ranges
[i
].offset
));
3485 VkResult
anv_FlushMappedMemoryRanges(
3487 uint32_t memoryRangeCount
,
3488 const VkMappedMemoryRange
* pMemoryRanges
)
3490 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3492 if (device
->info
.has_llc
)
3495 /* Make sure the writes we're flushing have landed. */
3496 __builtin_ia32_mfence();
3498 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
3503 VkResult
anv_InvalidateMappedMemoryRanges(
3505 uint32_t memoryRangeCount
,
3506 const VkMappedMemoryRange
* pMemoryRanges
)
3508 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3510 if (device
->info
.has_llc
)
3513 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
3515 /* Make sure no reads get moved up above the invalidate. */
3516 __builtin_ia32_mfence();
3521 void anv_GetBufferMemoryRequirements(
3524 VkMemoryRequirements
* pMemoryRequirements
)
3526 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
3527 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3528 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3530 /* The Vulkan spec (git aaed022) says:
3532 * memoryTypeBits is a bitfield and contains one bit set for every
3533 * supported memory type for the resource. The bit `1<<i` is set if and
3534 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3535 * structure for the physical device is supported.
3537 uint32_t memory_types
= 0;
3538 for (uint32_t i
= 0; i
< pdevice
->memory
.type_count
; i
++) {
3539 uint32_t valid_usage
= pdevice
->memory
.types
[i
].valid_buffer_usage
;
3540 if ((valid_usage
& buffer
->usage
) == buffer
->usage
)
3541 memory_types
|= (1u << i
);
3544 /* Base alignment requirement of a cache line */
3545 uint32_t alignment
= 16;
3547 /* We need an alignment of 32 for pushing UBOs */
3548 if (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
)
3549 alignment
= MAX2(alignment
, 32);
3551 pMemoryRequirements
->size
= buffer
->size
;
3552 pMemoryRequirements
->alignment
= alignment
;
3554 /* Storage and Uniform buffers should have their size aligned to
3555 * 32-bits to avoid boundary checks when last DWord is not complete.
3556 * This would ensure that not internal padding would be needed for
3559 if (device
->robust_buffer_access
&&
3560 (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
||
3561 buffer
->usage
& VK_BUFFER_USAGE_STORAGE_BUFFER_BIT
))
3562 pMemoryRequirements
->size
= align_u64(buffer
->size
, 4);
3564 pMemoryRequirements
->memoryTypeBits
= memory_types
;
3567 void anv_GetBufferMemoryRequirements2(
3569 const VkBufferMemoryRequirementsInfo2
* pInfo
,
3570 VkMemoryRequirements2
* pMemoryRequirements
)
3572 anv_GetBufferMemoryRequirements(_device
, pInfo
->buffer
,
3573 &pMemoryRequirements
->memoryRequirements
);
3575 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3576 switch (ext
->sType
) {
3577 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
3578 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
3579 requirements
->prefersDedicatedAllocation
= false;
3580 requirements
->requiresDedicatedAllocation
= false;
3585 anv_debug_ignored_stype(ext
->sType
);
3591 void anv_GetImageMemoryRequirements(
3594 VkMemoryRequirements
* pMemoryRequirements
)
3596 ANV_FROM_HANDLE(anv_image
, image
, _image
);
3597 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3598 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3600 /* The Vulkan spec (git aaed022) says:
3602 * memoryTypeBits is a bitfield and contains one bit set for every
3603 * supported memory type for the resource. The bit `1<<i` is set if and
3604 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3605 * structure for the physical device is supported.
3607 * All types are currently supported for images.
3609 uint32_t memory_types
= (1ull << pdevice
->memory
.type_count
) - 1;
3611 /* We must have image allocated or imported at this point. According to the
3612 * specification, external images must have been bound to memory before
3613 * calling GetImageMemoryRequirements.
3615 assert(image
->size
> 0);
3617 pMemoryRequirements
->size
= image
->size
;
3618 pMemoryRequirements
->alignment
= image
->alignment
;
3619 pMemoryRequirements
->memoryTypeBits
= memory_types
;
3622 void anv_GetImageMemoryRequirements2(
3624 const VkImageMemoryRequirementsInfo2
* pInfo
,
3625 VkMemoryRequirements2
* pMemoryRequirements
)
3627 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3628 ANV_FROM_HANDLE(anv_image
, image
, pInfo
->image
);
3630 anv_GetImageMemoryRequirements(_device
, pInfo
->image
,
3631 &pMemoryRequirements
->memoryRequirements
);
3633 vk_foreach_struct_const(ext
, pInfo
->pNext
) {
3634 switch (ext
->sType
) {
3635 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO
: {
3636 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3637 const VkImagePlaneMemoryRequirementsInfo
*plane_reqs
=
3638 (const VkImagePlaneMemoryRequirementsInfo
*) ext
;
3639 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
,
3640 plane_reqs
->planeAspect
);
3642 assert(image
->planes
[plane
].offset
== 0);
3644 /* The Vulkan spec (git aaed022) says:
3646 * memoryTypeBits is a bitfield and contains one bit set for every
3647 * supported memory type for the resource. The bit `1<<i` is set
3648 * if and only if the memory type `i` in the
3649 * VkPhysicalDeviceMemoryProperties structure for the physical
3650 * device is supported.
3652 * All types are currently supported for images.
3654 pMemoryRequirements
->memoryRequirements
.memoryTypeBits
=
3655 (1ull << pdevice
->memory
.type_count
) - 1;
3657 /* We must have image allocated or imported at this point. According to the
3658 * specification, external images must have been bound to memory before
3659 * calling GetImageMemoryRequirements.
3661 assert(image
->planes
[plane
].size
> 0);
3663 pMemoryRequirements
->memoryRequirements
.size
= image
->planes
[plane
].size
;
3664 pMemoryRequirements
->memoryRequirements
.alignment
=
3665 image
->planes
[plane
].alignment
;
3670 anv_debug_ignored_stype(ext
->sType
);
3675 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3676 switch (ext
->sType
) {
3677 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
3678 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
3679 if (image
->needs_set_tiling
|| image
->external_format
) {
3680 /* If we need to set the tiling for external consumers, we need a
3681 * dedicated allocation.
3683 * See also anv_AllocateMemory.
3685 requirements
->prefersDedicatedAllocation
= true;
3686 requirements
->requiresDedicatedAllocation
= true;
3688 requirements
->prefersDedicatedAllocation
= false;
3689 requirements
->requiresDedicatedAllocation
= false;
3695 anv_debug_ignored_stype(ext
->sType
);
3701 void anv_GetImageSparseMemoryRequirements(
3704 uint32_t* pSparseMemoryRequirementCount
,
3705 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
3707 *pSparseMemoryRequirementCount
= 0;
3710 void anv_GetImageSparseMemoryRequirements2(
3712 const VkImageSparseMemoryRequirementsInfo2
* pInfo
,
3713 uint32_t* pSparseMemoryRequirementCount
,
3714 VkSparseImageMemoryRequirements2
* pSparseMemoryRequirements
)
3716 *pSparseMemoryRequirementCount
= 0;
3719 void anv_GetDeviceMemoryCommitment(
3721 VkDeviceMemory memory
,
3722 VkDeviceSize
* pCommittedMemoryInBytes
)
3724 *pCommittedMemoryInBytes
= 0;
3728 anv_bind_buffer_memory(const VkBindBufferMemoryInfo
*pBindInfo
)
3730 ANV_FROM_HANDLE(anv_device_memory
, mem
, pBindInfo
->memory
);
3731 ANV_FROM_HANDLE(anv_buffer
, buffer
, pBindInfo
->buffer
);
3733 assert(pBindInfo
->sType
== VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
);
3736 assert((buffer
->usage
& mem
->type
->valid_buffer_usage
) == buffer
->usage
);
3737 buffer
->address
= (struct anv_address
) {
3739 .offset
= pBindInfo
->memoryOffset
,
3742 buffer
->address
= ANV_NULL_ADDRESS
;
3746 VkResult
anv_BindBufferMemory(
3749 VkDeviceMemory memory
,
3750 VkDeviceSize memoryOffset
)
3752 anv_bind_buffer_memory(
3753 &(VkBindBufferMemoryInfo
) {
3754 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
,
3757 .memoryOffset
= memoryOffset
,
3763 VkResult
anv_BindBufferMemory2(
3765 uint32_t bindInfoCount
,
3766 const VkBindBufferMemoryInfo
* pBindInfos
)
3768 for (uint32_t i
= 0; i
< bindInfoCount
; i
++)
3769 anv_bind_buffer_memory(&pBindInfos
[i
]);
3774 VkResult
anv_QueueBindSparse(
3776 uint32_t bindInfoCount
,
3777 const VkBindSparseInfo
* pBindInfo
,
3780 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
3781 if (anv_device_is_lost(queue
->device
))
3782 return VK_ERROR_DEVICE_LOST
;
3784 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
3789 VkResult
anv_CreateEvent(
3791 const VkEventCreateInfo
* pCreateInfo
,
3792 const VkAllocationCallbacks
* pAllocator
,
3795 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3796 struct anv_state state
;
3797 struct anv_event
*event
;
3799 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_EVENT_CREATE_INFO
);
3801 state
= anv_state_pool_alloc(&device
->dynamic_state_pool
,
3804 event
->state
= state
;
3805 event
->semaphore
= VK_EVENT_RESET
;
3807 if (!device
->info
.has_llc
) {
3808 /* Make sure the writes we're flushing have landed. */
3809 __builtin_ia32_mfence();
3810 __builtin_ia32_clflush(event
);
3813 *pEvent
= anv_event_to_handle(event
);
3818 void anv_DestroyEvent(
3821 const VkAllocationCallbacks
* pAllocator
)
3823 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3824 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3829 anv_state_pool_free(&device
->dynamic_state_pool
, event
->state
);
3832 VkResult
anv_GetEventStatus(
3836 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3837 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3839 if (anv_device_is_lost(device
))
3840 return VK_ERROR_DEVICE_LOST
;
3842 if (!device
->info
.has_llc
) {
3843 /* Invalidate read cache before reading event written by GPU. */
3844 __builtin_ia32_clflush(event
);
3845 __builtin_ia32_mfence();
3849 return event
->semaphore
;
3852 VkResult
anv_SetEvent(
3856 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3857 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3859 event
->semaphore
= VK_EVENT_SET
;
3861 if (!device
->info
.has_llc
) {
3862 /* Make sure the writes we're flushing have landed. */
3863 __builtin_ia32_mfence();
3864 __builtin_ia32_clflush(event
);
3870 VkResult
anv_ResetEvent(
3874 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3875 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3877 event
->semaphore
= VK_EVENT_RESET
;
3879 if (!device
->info
.has_llc
) {
3880 /* Make sure the writes we're flushing have landed. */
3881 __builtin_ia32_mfence();
3882 __builtin_ia32_clflush(event
);
3890 VkResult
anv_CreateBuffer(
3892 const VkBufferCreateInfo
* pCreateInfo
,
3893 const VkAllocationCallbacks
* pAllocator
,
3896 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3897 struct anv_buffer
*buffer
;
3899 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
3901 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
3902 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3904 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3906 buffer
->size
= pCreateInfo
->size
;
3907 buffer
->usage
= pCreateInfo
->usage
;
3908 buffer
->address
= ANV_NULL_ADDRESS
;
3910 *pBuffer
= anv_buffer_to_handle(buffer
);
3915 void anv_DestroyBuffer(
3918 const VkAllocationCallbacks
* pAllocator
)
3920 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3921 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
3926 vk_free2(&device
->alloc
, pAllocator
, buffer
);
3929 VkDeviceAddress
anv_GetBufferDeviceAddressEXT(
3931 const VkBufferDeviceAddressInfoEXT
* pInfo
)
3933 ANV_FROM_HANDLE(anv_buffer
, buffer
, pInfo
->buffer
);
3935 assert(buffer
->address
.bo
->flags
& EXEC_OBJECT_PINNED
);
3937 return anv_address_physical(buffer
->address
);
3941 anv_fill_buffer_surface_state(struct anv_device
*device
, struct anv_state state
,
3942 enum isl_format format
,
3943 struct anv_address address
,
3944 uint32_t range
, uint32_t stride
)
3946 isl_buffer_fill_state(&device
->isl_dev
, state
.map
,
3947 .address
= anv_address_physical(address
),
3948 .mocs
= device
->isl_dev
.mocs
.internal
,
3951 .swizzle
= ISL_SWIZZLE_IDENTITY
,
3952 .stride_B
= stride
);
3955 void anv_DestroySampler(
3958 const VkAllocationCallbacks
* pAllocator
)
3960 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3961 ANV_FROM_HANDLE(anv_sampler
, sampler
, _sampler
);
3966 if (sampler
->bindless_state
.map
) {
3967 anv_state_pool_free(&device
->dynamic_state_pool
,
3968 sampler
->bindless_state
);
3971 vk_free2(&device
->alloc
, pAllocator
, sampler
);
3974 VkResult
anv_CreateFramebuffer(
3976 const VkFramebufferCreateInfo
* pCreateInfo
,
3977 const VkAllocationCallbacks
* pAllocator
,
3978 VkFramebuffer
* pFramebuffer
)
3980 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3981 struct anv_framebuffer
*framebuffer
;
3983 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
3985 size_t size
= sizeof(*framebuffer
);
3987 /* VK_KHR_imageless_framebuffer extension says:
3989 * If flags includes VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR,
3990 * parameter pAttachments is ignored.
3992 if (!(pCreateInfo
->flags
& VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR
)) {
3993 size
+= sizeof(struct anv_image_view
*) * pCreateInfo
->attachmentCount
;
3994 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
3995 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3996 if (framebuffer
== NULL
)
3997 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3999 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
4000 ANV_FROM_HANDLE(anv_image_view
, iview
, pCreateInfo
->pAttachments
[i
]);
4001 framebuffer
->attachments
[i
] = iview
;
4003 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
4005 assert(device
->enabled_extensions
.KHR_imageless_framebuffer
);
4006 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
4007 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
4008 if (framebuffer
== NULL
)
4009 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
4011 framebuffer
->attachment_count
= 0;
4014 framebuffer
->width
= pCreateInfo
->width
;
4015 framebuffer
->height
= pCreateInfo
->height
;
4016 framebuffer
->layers
= pCreateInfo
->layers
;
4018 *pFramebuffer
= anv_framebuffer_to_handle(framebuffer
);
4023 void anv_DestroyFramebuffer(
4026 const VkAllocationCallbacks
* pAllocator
)
4028 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4029 ANV_FROM_HANDLE(anv_framebuffer
, fb
, _fb
);
4034 vk_free2(&device
->alloc
, pAllocator
, fb
);
4037 static const VkTimeDomainEXT anv_time_domains
[] = {
4038 VK_TIME_DOMAIN_DEVICE_EXT
,
4039 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
,
4040 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
,
4043 VkResult
anv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
4044 VkPhysicalDevice physicalDevice
,
4045 uint32_t *pTimeDomainCount
,
4046 VkTimeDomainEXT
*pTimeDomains
)
4049 VK_OUTARRAY_MAKE(out
, pTimeDomains
, pTimeDomainCount
);
4051 for (d
= 0; d
< ARRAY_SIZE(anv_time_domains
); d
++) {
4052 vk_outarray_append(&out
, i
) {
4053 *i
= anv_time_domains
[d
];
4057 return vk_outarray_status(&out
);
4061 anv_clock_gettime(clockid_t clock_id
)
4063 struct timespec current
;
4066 ret
= clock_gettime(clock_id
, ¤t
);
4067 if (ret
< 0 && clock_id
== CLOCK_MONOTONIC_RAW
)
4068 ret
= clock_gettime(CLOCK_MONOTONIC
, ¤t
);
4072 return (uint64_t) current
.tv_sec
* 1000000000ULL + current
.tv_nsec
;
4075 #define TIMESTAMP 0x2358
4077 VkResult
anv_GetCalibratedTimestampsEXT(
4079 uint32_t timestampCount
,
4080 const VkCalibratedTimestampInfoEXT
*pTimestampInfos
,
4081 uint64_t *pTimestamps
,
4082 uint64_t *pMaxDeviation
)
4084 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4085 uint64_t timestamp_frequency
= device
->info
.timestamp_frequency
;
4088 uint64_t begin
, end
;
4089 uint64_t max_clock_period
= 0;
4091 begin
= anv_clock_gettime(CLOCK_MONOTONIC_RAW
);
4093 for (d
= 0; d
< timestampCount
; d
++) {
4094 switch (pTimestampInfos
[d
].timeDomain
) {
4095 case VK_TIME_DOMAIN_DEVICE_EXT
:
4096 ret
= anv_gem_reg_read(device
, TIMESTAMP
| 1,
4100 return anv_device_set_lost(device
, "Failed to read the TIMESTAMP "
4103 uint64_t device_period
= DIV_ROUND_UP(1000000000, timestamp_frequency
);
4104 max_clock_period
= MAX2(max_clock_period
, device_period
);
4106 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
:
4107 pTimestamps
[d
] = anv_clock_gettime(CLOCK_MONOTONIC
);
4108 max_clock_period
= MAX2(max_clock_period
, 1);
4111 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
:
4112 pTimestamps
[d
] = begin
;
4120 end
= anv_clock_gettime(CLOCK_MONOTONIC_RAW
);
4123 * The maximum deviation is the sum of the interval over which we
4124 * perform the sampling and the maximum period of any sampled
4125 * clock. That's because the maximum skew between any two sampled
4126 * clock edges is when the sampled clock with the largest period is
4127 * sampled at the end of that period but right at the beginning of the
4128 * sampling interval and some other clock is sampled right at the
4129 * begining of its sampling period and right at the end of the
4130 * sampling interval. Let's assume the GPU has the longest clock
4131 * period and that the application is sampling GPU and monotonic:
4134 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
4135 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4139 * GPU -----_____-----_____-----_____-----_____
4142 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
4143 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4145 * Interval <----------------->
4146 * Deviation <-------------------------->
4150 * m = read(monotonic) 2
4153 * We round the sample interval up by one tick to cover sampling error
4154 * in the interval clock
4157 uint64_t sample_interval
= end
- begin
+ 1;
4159 *pMaxDeviation
= sample_interval
+ max_clock_period
;
4164 /* vk_icd.h does not declare this function, so we declare it here to
4165 * suppress Wmissing-prototypes.
4167 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4168 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
);
4170 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4171 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
)
4173 /* For the full details on loader interface versioning, see
4174 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
4175 * What follows is a condensed summary, to help you navigate the large and
4176 * confusing official doc.
4178 * - Loader interface v0 is incompatible with later versions. We don't
4181 * - In loader interface v1:
4182 * - The first ICD entrypoint called by the loader is
4183 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
4185 * - The ICD must statically expose no other Vulkan symbol unless it is
4186 * linked with -Bsymbolic.
4187 * - Each dispatchable Vulkan handle created by the ICD must be
4188 * a pointer to a struct whose first member is VK_LOADER_DATA. The
4189 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
4190 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
4191 * vkDestroySurfaceKHR(). The ICD must be capable of working with
4192 * such loader-managed surfaces.
4194 * - Loader interface v2 differs from v1 in:
4195 * - The first ICD entrypoint called by the loader is
4196 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
4197 * statically expose this entrypoint.
4199 * - Loader interface v3 differs from v2 in:
4200 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
4201 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
4202 * because the loader no longer does so.
4204 * - Loader interface v4 differs from v3 in:
4205 * - The ICD must implement vk_icdGetPhysicalDeviceProcAddr().
4207 *pSupportedVersion
= MIN2(*pSupportedVersion
, 4u);