2 * Copyright © 2015 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
28 #include <sys/sysinfo.h>
32 #include "drm-uapi/drm_fourcc.h"
34 #include "anv_private.h"
35 #include "util/strtod.h"
36 #include "util/debug.h"
37 #include "util/build_id.h"
38 #include "util/disk_cache.h"
39 #include "util/mesa-sha1.h"
40 #include "util/os_file.h"
41 #include "util/u_atomic.h"
42 #include "util/u_string.h"
43 #include "util/xmlpool.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_empty(&device
->instance
->debug_report_callbacks
.callbacks
))
75 (void) vsnprintf(str
, MAX_DEBUG_MESSAGE_LENGTH
, fmt
, args
);
78 vk_debug_report(&device
->instance
->debug_report_callbacks
,
79 VK_DEBUG_REPORT_DEBUG_BIT_EXT
,
80 VK_DEBUG_REPORT_OBJECT_TYPE_UNKNOWN_EXT
,
85 compiler_perf_log(void *data
, const char *fmt
, ...)
90 if (unlikely(INTEL_DEBUG
& DEBUG_PERF
))
91 intel_logd_v(fmt
, args
);
97 anv_compute_heap_size(int fd
, uint64_t gtt_size
)
99 /* Query the total ram from the system */
103 uint64_t total_ram
= (uint64_t)info
.totalram
* (uint64_t)info
.mem_unit
;
105 /* We don't want to burn too much ram with the GPU. If the user has 4GiB
106 * or less, we use at most half. If they have more than 4GiB, we use 3/4.
108 uint64_t available_ram
;
109 if (total_ram
<= 4ull * 1024ull * 1024ull * 1024ull)
110 available_ram
= total_ram
/ 2;
112 available_ram
= total_ram
* 3 / 4;
114 /* We also want to leave some padding for things we allocate in the driver,
115 * so don't go over 3/4 of the GTT either.
117 uint64_t available_gtt
= gtt_size
* 3 / 4;
119 return MIN2(available_ram
, available_gtt
);
123 anv_physical_device_init_heaps(struct anv_physical_device
*device
, int fd
)
126 if (anv_gem_get_context_param(fd
, 0, I915_CONTEXT_PARAM_GTT_SIZE
,
128 /* If, for whatever reason, we can't actually get the GTT size from the
129 * kernel (too old?) fall back to the aperture size.
131 anv_perf_warn(NULL
, NULL
,
132 "Failed to get I915_CONTEXT_PARAM_GTT_SIZE: %m");
134 if (anv_gem_get_aperture(fd
, >t_size
) == -1) {
135 return vk_errorf(NULL
, NULL
, VK_ERROR_INITIALIZATION_FAILED
,
136 "failed to get aperture size: %m");
140 device
->supports_48bit_addresses
= (device
->info
.gen
>= 8) &&
141 gtt_size
> (4ULL << 30 /* GiB */);
143 uint64_t heap_size
= anv_compute_heap_size(fd
, gtt_size
);
145 if (heap_size
> (2ull << 30) && !device
->supports_48bit_addresses
) {
146 /* When running with an overridden PCI ID, we may get a GTT size from
147 * the kernel that is greater than 2 GiB but the execbuf check for 48bit
148 * address support can still fail. Just clamp the address space size to
149 * 2 GiB if we don't have 48-bit support.
151 intel_logw("%s:%d: The kernel reported a GTT size larger than 2 GiB but "
152 "not support for 48-bit addresses",
154 heap_size
= 2ull << 30;
157 if (heap_size
<= 3ull * (1ull << 30)) {
158 /* In this case, everything fits nicely into the 32-bit address space,
159 * so there's no need for supporting 48bit addresses on client-allocated
162 device
->memory
.heap_count
= 1;
163 device
->memory
.heaps
[0] = (struct anv_memory_heap
) {
164 .vma_start
= LOW_HEAP_MIN_ADDRESS
,
165 .vma_size
= LOW_HEAP_SIZE
,
167 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
168 .supports_48bit_addresses
= false,
171 /* Not everything will fit nicely into a 32-bit address space. In this
172 * case we need a 64-bit heap. Advertise a small 32-bit heap and a
173 * larger 48-bit heap. If we're in this case, then we have a total heap
174 * size larger than 3GiB which most likely means they have 8 GiB of
175 * video memory and so carving off 1 GiB for the 32-bit heap should be
178 const uint64_t heap_size_32bit
= 1ull << 30;
179 const uint64_t heap_size_48bit
= heap_size
- heap_size_32bit
;
181 assert(device
->supports_48bit_addresses
);
183 device
->memory
.heap_count
= 2;
184 device
->memory
.heaps
[0] = (struct anv_memory_heap
) {
185 .vma_start
= HIGH_HEAP_MIN_ADDRESS
,
186 /* Leave the last 4GiB out of the high vma range, so that no state
187 * base address + size can overflow 48 bits. For more information see
188 * the comment about Wa32bitGeneralStateOffset in anv_allocator.c
190 .vma_size
= gtt_size
- (1ull << 32) - HIGH_HEAP_MIN_ADDRESS
,
191 .size
= heap_size_48bit
,
192 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
193 .supports_48bit_addresses
= true,
195 device
->memory
.heaps
[1] = (struct anv_memory_heap
) {
196 .vma_start
= LOW_HEAP_MIN_ADDRESS
,
197 .vma_size
= LOW_HEAP_SIZE
,
198 .size
= heap_size_32bit
,
199 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
200 .supports_48bit_addresses
= false,
204 uint32_t type_count
= 0;
205 for (uint32_t heap
= 0; heap
< device
->memory
.heap_count
; heap
++) {
206 uint32_t valid_buffer_usage
= ~0;
208 /* There appears to be a hardware issue in the VF cache where it only
209 * considers the bottom 32 bits of memory addresses. If you happen to
210 * have two vertex buffers which get placed exactly 4 GiB apart and use
211 * them in back-to-back draw calls, you can get collisions. In order to
212 * solve this problem, we require vertex and index buffers be bound to
213 * memory allocated out of the 32-bit heap.
215 if (device
->memory
.heaps
[heap
].supports_48bit_addresses
) {
216 valid_buffer_usage
&= ~(VK_BUFFER_USAGE_INDEX_BUFFER_BIT
|
217 VK_BUFFER_USAGE_VERTEX_BUFFER_BIT
);
220 if (device
->info
.has_llc
) {
221 /* Big core GPUs share LLC with the CPU and thus one memory type can be
222 * both cached and coherent at the same time.
224 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
225 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
226 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
227 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
228 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
230 .valid_buffer_usage
= valid_buffer_usage
,
233 /* The spec requires that we expose a host-visible, coherent memory
234 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
235 * to give the application a choice between cached, but not coherent and
236 * coherent but uncached (WC though).
238 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
239 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
240 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
241 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
243 .valid_buffer_usage
= valid_buffer_usage
,
245 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
246 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
247 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
248 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
250 .valid_buffer_usage
= valid_buffer_usage
,
254 device
->memory
.type_count
= type_count
;
260 anv_physical_device_init_uuids(struct anv_physical_device
*device
)
262 const struct build_id_note
*note
=
263 build_id_find_nhdr_for_addr(anv_physical_device_init_uuids
);
265 return vk_errorf(device
->instance
, device
,
266 VK_ERROR_INITIALIZATION_FAILED
,
267 "Failed to find build-id");
270 unsigned build_id_len
= build_id_length(note
);
271 if (build_id_len
< 20) {
272 return vk_errorf(device
->instance
, device
,
273 VK_ERROR_INITIALIZATION_FAILED
,
274 "build-id too short. It needs to be a SHA");
277 memcpy(device
->driver_build_sha1
, build_id_data(note
), 20);
279 struct mesa_sha1 sha1_ctx
;
281 STATIC_ASSERT(VK_UUID_SIZE
<= sizeof(sha1
));
283 /* The pipeline cache UUID is used for determining when a pipeline cache is
284 * invalid. It needs both a driver build and the PCI ID of the device.
286 _mesa_sha1_init(&sha1_ctx
);
287 _mesa_sha1_update(&sha1_ctx
, build_id_data(note
), build_id_len
);
288 _mesa_sha1_update(&sha1_ctx
, &device
->chipset_id
,
289 sizeof(device
->chipset_id
));
290 _mesa_sha1_update(&sha1_ctx
, &device
->always_use_bindless
,
291 sizeof(device
->always_use_bindless
));
292 _mesa_sha1_update(&sha1_ctx
, &device
->has_a64_buffer_access
,
293 sizeof(device
->has_a64_buffer_access
));
294 _mesa_sha1_update(&sha1_ctx
, &device
->has_bindless_images
,
295 sizeof(device
->has_bindless_images
));
296 _mesa_sha1_update(&sha1_ctx
, &device
->has_bindless_samplers
,
297 sizeof(device
->has_bindless_samplers
));
298 _mesa_sha1_final(&sha1_ctx
, sha1
);
299 memcpy(device
->pipeline_cache_uuid
, sha1
, VK_UUID_SIZE
);
301 /* The driver UUID is used for determining sharability of images and memory
302 * between two Vulkan instances in separate processes. People who want to
303 * share memory need to also check the device UUID (below) so all this
304 * needs to be is the build-id.
306 memcpy(device
->driver_uuid
, build_id_data(note
), VK_UUID_SIZE
);
308 /* The device UUID uniquely identifies the given device within the machine.
309 * Since we never have more than one device, this doesn't need to be a real
310 * UUID. However, on the off-chance that someone tries to use this to
311 * cache pre-tiled images or something of the like, we use the PCI ID and
312 * some bits of ISL info to ensure that this is safe.
314 _mesa_sha1_init(&sha1_ctx
);
315 _mesa_sha1_update(&sha1_ctx
, &device
->chipset_id
,
316 sizeof(device
->chipset_id
));
317 _mesa_sha1_update(&sha1_ctx
, &device
->isl_dev
.has_bit6_swizzling
,
318 sizeof(device
->isl_dev
.has_bit6_swizzling
));
319 _mesa_sha1_final(&sha1_ctx
, sha1
);
320 memcpy(device
->device_uuid
, sha1
, VK_UUID_SIZE
);
326 anv_physical_device_init_disk_cache(struct anv_physical_device
*device
)
328 #ifdef ENABLE_SHADER_CACHE
330 ASSERTED
int len
= snprintf(renderer
, sizeof(renderer
), "anv_%04x",
332 assert(len
== sizeof(renderer
) - 2);
335 _mesa_sha1_format(timestamp
, device
->driver_build_sha1
);
337 const uint64_t driver_flags
=
338 brw_get_compiler_config_value(device
->compiler
);
339 device
->disk_cache
= disk_cache_create(renderer
, timestamp
, driver_flags
);
341 device
->disk_cache
= NULL
;
346 anv_physical_device_free_disk_cache(struct anv_physical_device
*device
)
348 #ifdef ENABLE_SHADER_CACHE
349 if (device
->disk_cache
)
350 disk_cache_destroy(device
->disk_cache
);
352 assert(device
->disk_cache
== NULL
);
357 get_available_system_memory()
359 char *meminfo
= os_read_file("/proc/meminfo");
363 char *str
= strstr(meminfo
, "MemAvailable:");
369 uint64_t kb_mem_available
;
370 if (sscanf(str
, "MemAvailable: %" PRIx64
, &kb_mem_available
) == 1) {
372 return kb_mem_available
<< 10;
380 anv_physical_device_init(struct anv_physical_device
*device
,
381 struct anv_instance
*instance
,
382 drmDevicePtr drm_device
)
384 const char *primary_path
= drm_device
->nodes
[DRM_NODE_PRIMARY
];
385 const char *path
= drm_device
->nodes
[DRM_NODE_RENDER
];
390 brw_process_intel_debug_variable();
392 fd
= open(path
, O_RDWR
| O_CLOEXEC
);
394 return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
396 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
397 device
->instance
= instance
;
399 assert(strlen(path
) < ARRAY_SIZE(device
->path
));
400 snprintf(device
->path
, ARRAY_SIZE(device
->path
), "%s", path
);
402 if (!gen_get_device_info_from_fd(fd
, &device
->info
)) {
403 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
406 device
->chipset_id
= device
->info
.chipset_id
;
407 device
->no_hw
= device
->info
.no_hw
;
409 if (getenv("INTEL_NO_HW") != NULL
)
410 device
->no_hw
= true;
412 device
->pci_info
.domain
= drm_device
->businfo
.pci
->domain
;
413 device
->pci_info
.bus
= drm_device
->businfo
.pci
->bus
;
414 device
->pci_info
.device
= drm_device
->businfo
.pci
->dev
;
415 device
->pci_info
.function
= drm_device
->businfo
.pci
->func
;
417 device
->name
= gen_get_device_name(device
->chipset_id
);
419 if (device
->info
.is_haswell
) {
420 intel_logw("Haswell Vulkan support is incomplete");
421 } else if (device
->info
.gen
== 7 && !device
->info
.is_baytrail
) {
422 intel_logw("Ivy Bridge Vulkan support is incomplete");
423 } else if (device
->info
.gen
== 7 && device
->info
.is_baytrail
) {
424 intel_logw("Bay Trail Vulkan support is incomplete");
425 } else if (device
->info
.gen
>= 8 && device
->info
.gen
<= 11) {
426 /* Gen8-11 fully supported */
427 } else if (device
->info
.gen
== 12) {
428 intel_logw("Vulkan is not yet fully supported on gen12");
430 result
= vk_errorf(device
->instance
, device
,
431 VK_ERROR_INCOMPATIBLE_DRIVER
,
432 "Vulkan not yet supported on %s", device
->name
);
436 device
->cmd_parser_version
= -1;
437 if (device
->info
.gen
== 7) {
438 device
->cmd_parser_version
=
439 anv_gem_get_param(fd
, I915_PARAM_CMD_PARSER_VERSION
);
440 if (device
->cmd_parser_version
== -1) {
441 result
= vk_errorf(device
->instance
, device
,
442 VK_ERROR_INITIALIZATION_FAILED
,
443 "failed to get command parser version");
448 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_WAIT_TIMEOUT
)) {
449 result
= vk_errorf(device
->instance
, device
,
450 VK_ERROR_INITIALIZATION_FAILED
,
451 "kernel missing gem wait");
455 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_EXECBUF2
)) {
456 result
= vk_errorf(device
->instance
, device
,
457 VK_ERROR_INITIALIZATION_FAILED
,
458 "kernel missing execbuf2");
462 if (!device
->info
.has_llc
&&
463 anv_gem_get_param(fd
, I915_PARAM_MMAP_VERSION
) < 1) {
464 result
= vk_errorf(device
->instance
, device
,
465 VK_ERROR_INITIALIZATION_FAILED
,
466 "kernel missing wc mmap");
470 result
= anv_physical_device_init_heaps(device
, fd
);
471 if (result
!= VK_SUCCESS
)
474 device
->has_exec_async
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_ASYNC
);
475 device
->has_exec_capture
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_CAPTURE
);
476 device
->has_exec_fence
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE
);
477 device
->has_syncobj
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_FENCE_ARRAY
);
478 device
->has_syncobj_wait
= device
->has_syncobj
&&
479 anv_gem_supports_syncobj_wait(fd
);
480 device
->has_context_priority
= anv_gem_has_context_priority(fd
);
482 device
->use_softpin
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_SOFTPIN
)
483 && device
->supports_48bit_addresses
;
485 device
->has_context_isolation
=
486 anv_gem_get_param(fd
, I915_PARAM_HAS_CONTEXT_ISOLATION
);
488 device
->always_use_bindless
=
489 env_var_as_boolean("ANV_ALWAYS_BINDLESS", false);
491 /* We first got the A64 messages on broadwell and we can only use them if
492 * we can pass addresses directly into the shader which requires softpin.
494 device
->has_a64_buffer_access
= device
->info
.gen
>= 8 &&
497 /* We first get bindless image access on Skylake and we can only really do
498 * it if we don't have any relocations so we need softpin.
500 device
->has_bindless_images
= device
->info
.gen
>= 9 &&
503 /* We've had bindless samplers since Ivy Bridge (forever in Vulkan terms)
504 * because it's just a matter of setting the sampler address in the sample
505 * message header. However, we've not bothered to wire it up for vec4 so
506 * we leave it disabled on gen7.
508 device
->has_bindless_samplers
= device
->info
.gen
>= 8;
510 device
->has_mem_available
= get_available_system_memory() != 0;
512 /* Starting with Gen10, the timestamp frequency of the command streamer may
513 * vary from one part to another. We can query the value from the kernel.
515 if (device
->info
.gen
>= 10) {
516 int timestamp_frequency
=
517 anv_gem_get_param(fd
, I915_PARAM_CS_TIMESTAMP_FREQUENCY
);
519 if (timestamp_frequency
< 0)
520 intel_logw("Kernel 4.16-rc1+ required to properly query CS timestamp frequency");
522 device
->info
.timestamp_frequency
= timestamp_frequency
;
525 /* GENs prior to 8 do not support EU/Subslice info */
526 if (device
->info
.gen
>= 8) {
527 device
->subslice_total
= anv_gem_get_param(fd
, I915_PARAM_SUBSLICE_TOTAL
);
528 device
->eu_total
= anv_gem_get_param(fd
, I915_PARAM_EU_TOTAL
);
530 /* Without this information, we cannot get the right Braswell
531 * brandstrings, and we have to use conservative numbers for GPGPU on
532 * many platforms, but otherwise, things will just work.
534 if (device
->subslice_total
< 1 || device
->eu_total
< 1) {
535 intel_logw("Kernel 4.1 required to properly query GPU properties");
537 } else if (device
->info
.gen
== 7) {
538 device
->subslice_total
= 1 << (device
->info
.gt
- 1);
541 if (device
->info
.is_cherryview
&&
542 device
->subslice_total
> 0 && device
->eu_total
> 0) {
543 /* Logical CS threads = EUs per subslice * num threads per EU */
544 uint32_t max_cs_threads
=
545 device
->eu_total
/ device
->subslice_total
* device
->info
.num_thread_per_eu
;
547 /* Fuse configurations may give more threads than expected, never less. */
548 if (max_cs_threads
> device
->info
.max_cs_threads
)
549 device
->info
.max_cs_threads
= max_cs_threads
;
552 device
->compiler
= brw_compiler_create(NULL
, &device
->info
);
553 if (device
->compiler
== NULL
) {
554 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
557 device
->compiler
->shader_debug_log
= compiler_debug_log
;
558 device
->compiler
->shader_perf_log
= compiler_perf_log
;
559 device
->compiler
->supports_pull_constants
= false;
560 device
->compiler
->constant_buffer_0_is_relative
=
561 device
->info
.gen
< 8 || !device
->has_context_isolation
;
562 device
->compiler
->supports_shader_constants
= true;
564 /* Broadwell PRM says:
566 * "Before Gen8, there was a historical configuration control field to
567 * swizzle address bit[6] for in X/Y tiling modes. This was set in three
568 * different places: TILECTL[1:0], ARB_MODE[5:4], and
569 * DISP_ARB_CTL[14:13].
571 * For Gen8 and subsequent generations, the swizzle fields are all
572 * reserved, and the CPU's memory controller performs all address
573 * swizzling modifications."
576 device
->info
.gen
< 8 && anv_gem_get_bit6_swizzle(fd
, I915_TILING_X
);
578 isl_device_init(&device
->isl_dev
, &device
->info
, swizzled
);
580 result
= anv_physical_device_init_uuids(device
);
581 if (result
!= VK_SUCCESS
)
584 anv_physical_device_init_disk_cache(device
);
586 if (instance
->enabled_extensions
.KHR_display
) {
587 master_fd
= open(primary_path
, O_RDWR
| O_CLOEXEC
);
588 if (master_fd
>= 0) {
589 /* prod the device with a GETPARAM call which will fail if
590 * we don't have permission to even render on this device
592 if (anv_gem_get_param(master_fd
, I915_PARAM_CHIPSET_ID
) == 0) {
598 device
->master_fd
= master_fd
;
600 result
= anv_init_wsi(device
);
601 if (result
!= VK_SUCCESS
) {
602 ralloc_free(device
->compiler
);
603 anv_physical_device_free_disk_cache(device
);
607 anv_physical_device_get_supported_extensions(device
,
608 &device
->supported_extensions
);
611 device
->local_fd
= fd
;
623 anv_physical_device_finish(struct anv_physical_device
*device
)
625 anv_finish_wsi(device
);
626 anv_physical_device_free_disk_cache(device
);
627 ralloc_free(device
->compiler
);
628 close(device
->local_fd
);
629 if (device
->master_fd
>= 0)
630 close(device
->master_fd
);
634 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
635 VkSystemAllocationScope allocationScope
)
641 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
642 size_t align
, VkSystemAllocationScope allocationScope
)
644 return realloc(pOriginal
, size
);
648 default_free_func(void *pUserData
, void *pMemory
)
653 static const VkAllocationCallbacks default_alloc
= {
655 .pfnAllocation
= default_alloc_func
,
656 .pfnReallocation
= default_realloc_func
,
657 .pfnFree
= default_free_func
,
660 VkResult
anv_EnumerateInstanceExtensionProperties(
661 const char* pLayerName
,
662 uint32_t* pPropertyCount
,
663 VkExtensionProperties
* pProperties
)
665 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
667 for (int i
= 0; i
< ANV_INSTANCE_EXTENSION_COUNT
; i
++) {
668 if (anv_instance_extensions_supported
.extensions
[i
]) {
669 vk_outarray_append(&out
, prop
) {
670 *prop
= anv_instance_extensions
[i
];
675 return vk_outarray_status(&out
);
678 VkResult
anv_CreateInstance(
679 const VkInstanceCreateInfo
* pCreateInfo
,
680 const VkAllocationCallbacks
* pAllocator
,
681 VkInstance
* pInstance
)
683 struct anv_instance
*instance
;
686 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
688 struct anv_instance_extension_table enabled_extensions
= {};
689 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
691 for (idx
= 0; idx
< ANV_INSTANCE_EXTENSION_COUNT
; idx
++) {
692 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
693 anv_instance_extensions
[idx
].extensionName
) == 0)
697 if (idx
>= ANV_INSTANCE_EXTENSION_COUNT
)
698 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
700 if (!anv_instance_extensions_supported
.extensions
[idx
])
701 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
703 enabled_extensions
.extensions
[idx
] = true;
706 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
707 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
709 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
711 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
714 instance
->alloc
= *pAllocator
;
716 instance
->alloc
= default_alloc
;
718 instance
->app_info
= (struct anv_app_info
) { .api_version
= 0 };
719 if (pCreateInfo
->pApplicationInfo
) {
720 const VkApplicationInfo
*app
= pCreateInfo
->pApplicationInfo
;
722 instance
->app_info
.app_name
=
723 vk_strdup(&instance
->alloc
, app
->pApplicationName
,
724 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
725 instance
->app_info
.app_version
= app
->applicationVersion
;
727 instance
->app_info
.engine_name
=
728 vk_strdup(&instance
->alloc
, app
->pEngineName
,
729 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
730 instance
->app_info
.engine_version
= app
->engineVersion
;
732 instance
->app_info
.api_version
= app
->apiVersion
;
735 if (instance
->app_info
.api_version
== 0)
736 instance
->app_info
.api_version
= VK_API_VERSION_1_0
;
738 instance
->enabled_extensions
= enabled_extensions
;
740 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->dispatch
.entrypoints
); i
++) {
741 /* Vulkan requires that entrypoints for extensions which have not been
742 * enabled must not be advertised.
744 if (!anv_instance_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
745 &instance
->enabled_extensions
)) {
746 instance
->dispatch
.entrypoints
[i
] = NULL
;
748 instance
->dispatch
.entrypoints
[i
] =
749 anv_instance_dispatch_table
.entrypoints
[i
];
753 struct anv_physical_device
*pdevice
= &instance
->physicalDevice
;
754 for (unsigned i
= 0; i
< ARRAY_SIZE(pdevice
->dispatch
.entrypoints
); i
++) {
755 /* Vulkan requires that entrypoints for extensions which have not been
756 * enabled must not be advertised.
758 if (!anv_physical_device_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
759 &instance
->enabled_extensions
)) {
760 pdevice
->dispatch
.entrypoints
[i
] = NULL
;
762 pdevice
->dispatch
.entrypoints
[i
] =
763 anv_physical_device_dispatch_table
.entrypoints
[i
];
767 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->device_dispatch
.entrypoints
); i
++) {
768 /* Vulkan requires that entrypoints for extensions which have not been
769 * enabled must not be advertised.
771 if (!anv_device_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
772 &instance
->enabled_extensions
, NULL
)) {
773 instance
->device_dispatch
.entrypoints
[i
] = NULL
;
775 instance
->device_dispatch
.entrypoints
[i
] =
776 anv_device_dispatch_table
.entrypoints
[i
];
780 instance
->physicalDeviceCount
= -1;
782 result
= vk_debug_report_instance_init(&instance
->debug_report_callbacks
);
783 if (result
!= VK_SUCCESS
) {
784 vk_free2(&default_alloc
, pAllocator
, instance
);
785 return vk_error(result
);
788 instance
->pipeline_cache_enabled
=
789 env_var_as_boolean("ANV_ENABLE_PIPELINE_CACHE", true);
792 glsl_type_singleton_init_or_ref();
794 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
796 driParseOptionInfo(&instance
->available_dri_options
, anv_dri_options_xml
);
797 driParseConfigFiles(&instance
->dri_options
, &instance
->available_dri_options
,
799 instance
->app_info
.engine_name
,
800 instance
->app_info
.engine_version
);
802 *pInstance
= anv_instance_to_handle(instance
);
807 void anv_DestroyInstance(
808 VkInstance _instance
,
809 const VkAllocationCallbacks
* pAllocator
)
811 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
816 if (instance
->physicalDeviceCount
> 0) {
817 /* We support at most one physical device. */
818 assert(instance
->physicalDeviceCount
== 1);
819 anv_physical_device_finish(&instance
->physicalDevice
);
822 vk_free(&instance
->alloc
, (char *)instance
->app_info
.app_name
);
823 vk_free(&instance
->alloc
, (char *)instance
->app_info
.engine_name
);
825 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
827 vk_debug_report_instance_destroy(&instance
->debug_report_callbacks
);
829 glsl_type_singleton_decref();
832 driDestroyOptionCache(&instance
->dri_options
);
833 driDestroyOptionInfo(&instance
->available_dri_options
);
835 vk_free(&instance
->alloc
, instance
);
839 anv_enumerate_devices(struct anv_instance
*instance
)
841 /* TODO: Check for more devices ? */
842 drmDevicePtr devices
[8];
843 VkResult result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
846 instance
->physicalDeviceCount
= 0;
848 max_devices
= drmGetDevices2(0, devices
, ARRAY_SIZE(devices
));
850 return VK_ERROR_INCOMPATIBLE_DRIVER
;
852 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
853 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
854 devices
[i
]->bustype
== DRM_BUS_PCI
&&
855 devices
[i
]->deviceinfo
.pci
->vendor_id
== 0x8086) {
857 result
= anv_physical_device_init(&instance
->physicalDevice
,
858 instance
, devices
[i
]);
859 if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
863 drmFreeDevices(devices
, max_devices
);
865 if (result
== VK_SUCCESS
)
866 instance
->physicalDeviceCount
= 1;
872 anv_instance_ensure_physical_device(struct anv_instance
*instance
)
874 if (instance
->physicalDeviceCount
< 0) {
875 VkResult result
= anv_enumerate_devices(instance
);
876 if (result
!= VK_SUCCESS
&&
877 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
884 VkResult
anv_EnumeratePhysicalDevices(
885 VkInstance _instance
,
886 uint32_t* pPhysicalDeviceCount
,
887 VkPhysicalDevice
* pPhysicalDevices
)
889 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
890 VK_OUTARRAY_MAKE(out
, pPhysicalDevices
, pPhysicalDeviceCount
);
892 VkResult result
= anv_instance_ensure_physical_device(instance
);
893 if (result
!= VK_SUCCESS
)
896 if (instance
->physicalDeviceCount
== 0)
899 assert(instance
->physicalDeviceCount
== 1);
900 vk_outarray_append(&out
, i
) {
901 *i
= anv_physical_device_to_handle(&instance
->physicalDevice
);
904 return vk_outarray_status(&out
);
907 VkResult
anv_EnumeratePhysicalDeviceGroups(
908 VkInstance _instance
,
909 uint32_t* pPhysicalDeviceGroupCount
,
910 VkPhysicalDeviceGroupProperties
* pPhysicalDeviceGroupProperties
)
912 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
913 VK_OUTARRAY_MAKE(out
, pPhysicalDeviceGroupProperties
,
914 pPhysicalDeviceGroupCount
);
916 VkResult result
= anv_instance_ensure_physical_device(instance
);
917 if (result
!= VK_SUCCESS
)
920 if (instance
->physicalDeviceCount
== 0)
923 assert(instance
->physicalDeviceCount
== 1);
925 vk_outarray_append(&out
, p
) {
926 p
->physicalDeviceCount
= 1;
927 memset(p
->physicalDevices
, 0, sizeof(p
->physicalDevices
));
928 p
->physicalDevices
[0] =
929 anv_physical_device_to_handle(&instance
->physicalDevice
);
930 p
->subsetAllocation
= false;
932 vk_foreach_struct(ext
, p
->pNext
)
933 anv_debug_ignored_stype(ext
->sType
);
936 return vk_outarray_status(&out
);
939 void anv_GetPhysicalDeviceFeatures(
940 VkPhysicalDevice physicalDevice
,
941 VkPhysicalDeviceFeatures
* pFeatures
)
943 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
945 *pFeatures
= (VkPhysicalDeviceFeatures
) {
946 .robustBufferAccess
= true,
947 .fullDrawIndexUint32
= true,
948 .imageCubeArray
= true,
949 .independentBlend
= true,
950 .geometryShader
= true,
951 .tessellationShader
= true,
952 .sampleRateShading
= true,
953 .dualSrcBlend
= true,
955 .multiDrawIndirect
= true,
956 .drawIndirectFirstInstance
= true,
958 .depthBiasClamp
= true,
959 .fillModeNonSolid
= true,
960 .depthBounds
= false,
964 .multiViewport
= true,
965 .samplerAnisotropy
= true,
966 .textureCompressionETC2
= pdevice
->info
.gen
>= 8 ||
967 pdevice
->info
.is_baytrail
,
968 .textureCompressionASTC_LDR
= pdevice
->info
.gen
>= 9, /* FINISHME CHV */
969 .textureCompressionBC
= true,
970 .occlusionQueryPrecise
= true,
971 .pipelineStatisticsQuery
= true,
972 .fragmentStoresAndAtomics
= true,
973 .shaderTessellationAndGeometryPointSize
= true,
974 .shaderImageGatherExtended
= true,
975 .shaderStorageImageExtendedFormats
= true,
976 .shaderStorageImageMultisample
= false,
977 .shaderStorageImageReadWithoutFormat
= false,
978 .shaderStorageImageWriteWithoutFormat
= true,
979 .shaderUniformBufferArrayDynamicIndexing
= true,
980 .shaderSampledImageArrayDynamicIndexing
= true,
981 .shaderStorageBufferArrayDynamicIndexing
= true,
982 .shaderStorageImageArrayDynamicIndexing
= true,
983 .shaderClipDistance
= true,
984 .shaderCullDistance
= true,
985 .shaderFloat64
= pdevice
->info
.gen
>= 8 &&
986 pdevice
->info
.has_64bit_types
,
987 .shaderInt64
= pdevice
->info
.gen
>= 8 &&
988 pdevice
->info
.has_64bit_types
,
989 .shaderInt16
= pdevice
->info
.gen
>= 8,
990 .shaderResourceMinLod
= pdevice
->info
.gen
>= 9,
991 .variableMultisampleRate
= true,
992 .inheritedQueries
= true,
995 /* We can't do image stores in vec4 shaders */
996 pFeatures
->vertexPipelineStoresAndAtomics
=
997 pdevice
->compiler
->scalar_stage
[MESA_SHADER_VERTEX
] &&
998 pdevice
->compiler
->scalar_stage
[MESA_SHADER_GEOMETRY
];
1000 struct anv_app_info
*app_info
= &pdevice
->instance
->app_info
;
1002 /* The new DOOM and Wolfenstein games require depthBounds without
1003 * checking for it. They seem to run fine without it so just claim it's
1004 * there and accept the consequences.
1006 if (app_info
->engine_name
&& strcmp(app_info
->engine_name
, "idTech") == 0)
1007 pFeatures
->depthBounds
= true;
1010 void anv_GetPhysicalDeviceFeatures2(
1011 VkPhysicalDevice physicalDevice
,
1012 VkPhysicalDeviceFeatures2
* pFeatures
)
1014 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1015 anv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
1017 vk_foreach_struct(ext
, pFeatures
->pNext
) {
1018 switch (ext
->sType
) {
1019 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES_KHR
: {
1020 VkPhysicalDevice8BitStorageFeaturesKHR
*features
=
1021 (VkPhysicalDevice8BitStorageFeaturesKHR
*)ext
;
1022 features
->storageBuffer8BitAccess
= pdevice
->info
.gen
>= 8;
1023 features
->uniformAndStorageBuffer8BitAccess
= pdevice
->info
.gen
>= 8;
1024 features
->storagePushConstant8
= pdevice
->info
.gen
>= 8;
1028 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES
: {
1029 VkPhysicalDevice16BitStorageFeatures
*features
=
1030 (VkPhysicalDevice16BitStorageFeatures
*)ext
;
1031 features
->storageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
1032 features
->uniformAndStorageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
1033 features
->storagePushConstant16
= pdevice
->info
.gen
>= 8;
1034 features
->storageInputOutput16
= false;
1038 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_EXT
: {
1039 VkPhysicalDeviceBufferDeviceAddressFeaturesEXT
*features
= (void *)ext
;
1040 features
->bufferDeviceAddress
= pdevice
->has_a64_buffer_access
;
1041 features
->bufferDeviceAddressCaptureReplay
= false;
1042 features
->bufferDeviceAddressMultiDevice
= false;
1046 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COMPUTE_SHADER_DERIVATIVES_FEATURES_NV
: {
1047 VkPhysicalDeviceComputeShaderDerivativesFeaturesNV
*features
=
1048 (VkPhysicalDeviceComputeShaderDerivativesFeaturesNV
*)ext
;
1049 features
->computeDerivativeGroupQuads
= true;
1050 features
->computeDerivativeGroupLinear
= true;
1054 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT
: {
1055 VkPhysicalDeviceConditionalRenderingFeaturesEXT
*features
=
1056 (VkPhysicalDeviceConditionalRenderingFeaturesEXT
*)ext
;
1057 features
->conditionalRendering
= pdevice
->info
.gen
>= 8 ||
1058 pdevice
->info
.is_haswell
;
1059 features
->inheritedConditionalRendering
= pdevice
->info
.gen
>= 8 ||
1060 pdevice
->info
.is_haswell
;
1064 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_CLIP_ENABLE_FEATURES_EXT
: {
1065 VkPhysicalDeviceDepthClipEnableFeaturesEXT
*features
=
1066 (VkPhysicalDeviceDepthClipEnableFeaturesEXT
*)ext
;
1067 features
->depthClipEnable
= true;
1071 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR
: {
1072 VkPhysicalDeviceFloat16Int8FeaturesKHR
*features
= (void *)ext
;
1073 features
->shaderFloat16
= pdevice
->info
.gen
>= 8;
1074 features
->shaderInt8
= pdevice
->info
.gen
>= 8;
1078 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FRAGMENT_SHADER_INTERLOCK_FEATURES_EXT
: {
1079 VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT
*features
=
1080 (VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT
*)ext
;
1081 features
->fragmentShaderSampleInterlock
= pdevice
->info
.gen
>= 9;
1082 features
->fragmentShaderPixelInterlock
= pdevice
->info
.gen
>= 9;
1083 features
->fragmentShaderShadingRateInterlock
= false;
1087 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_QUERY_RESET_FEATURES_EXT
: {
1088 VkPhysicalDeviceHostQueryResetFeaturesEXT
*features
=
1089 (VkPhysicalDeviceHostQueryResetFeaturesEXT
*)ext
;
1090 features
->hostQueryReset
= true;
1094 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT
: {
1095 VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*features
=
1096 (VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*)ext
;
1097 features
->shaderInputAttachmentArrayDynamicIndexing
= false;
1098 features
->shaderUniformTexelBufferArrayDynamicIndexing
= true;
1099 features
->shaderStorageTexelBufferArrayDynamicIndexing
= true;
1100 features
->shaderUniformBufferArrayNonUniformIndexing
= false;
1101 features
->shaderSampledImageArrayNonUniformIndexing
= true;
1102 features
->shaderStorageBufferArrayNonUniformIndexing
= true;
1103 features
->shaderStorageImageArrayNonUniformIndexing
= true;
1104 features
->shaderInputAttachmentArrayNonUniformIndexing
= false;
1105 features
->shaderUniformTexelBufferArrayNonUniformIndexing
= true;
1106 features
->shaderStorageTexelBufferArrayNonUniformIndexing
= true;
1107 features
->descriptorBindingUniformBufferUpdateAfterBind
= false;
1108 features
->descriptorBindingSampledImageUpdateAfterBind
= true;
1109 features
->descriptorBindingStorageImageUpdateAfterBind
= true;
1110 features
->descriptorBindingStorageBufferUpdateAfterBind
= true;
1111 features
->descriptorBindingUniformTexelBufferUpdateAfterBind
= true;
1112 features
->descriptorBindingStorageTexelBufferUpdateAfterBind
= true;
1113 features
->descriptorBindingUpdateUnusedWhilePending
= true;
1114 features
->descriptorBindingPartiallyBound
= true;
1115 features
->descriptorBindingVariableDescriptorCount
= false;
1116 features
->runtimeDescriptorArray
= true;
1120 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INDEX_TYPE_UINT8_FEATURES_EXT
: {
1121 VkPhysicalDeviceIndexTypeUint8FeaturesEXT
*features
=
1122 (VkPhysicalDeviceIndexTypeUint8FeaturesEXT
*)ext
;
1123 features
->indexTypeUint8
= true;
1127 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_FEATURES_EXT
: {
1128 VkPhysicalDeviceInlineUniformBlockFeaturesEXT
*features
=
1129 (VkPhysicalDeviceInlineUniformBlockFeaturesEXT
*)ext
;
1130 features
->inlineUniformBlock
= true;
1131 features
->descriptorBindingInlineUniformBlockUpdateAfterBind
= true;
1135 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_FEATURES_EXT
: {
1136 VkPhysicalDeviceLineRasterizationFeaturesEXT
*features
=
1137 (VkPhysicalDeviceLineRasterizationFeaturesEXT
*)ext
;
1138 features
->rectangularLines
= true;
1139 features
->bresenhamLines
= true;
1140 features
->smoothLines
= true;
1141 features
->stippledRectangularLines
= false;
1142 features
->stippledBresenhamLines
= true;
1143 features
->stippledSmoothLines
= false;
1147 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES
: {
1148 VkPhysicalDeviceMultiviewFeatures
*features
=
1149 (VkPhysicalDeviceMultiviewFeatures
*)ext
;
1150 features
->multiview
= true;
1151 features
->multiviewGeometryShader
= true;
1152 features
->multiviewTessellationShader
= true;
1156 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGELESS_FRAMEBUFFER_FEATURES_KHR
: {
1157 VkPhysicalDeviceImagelessFramebufferFeaturesKHR
*features
=
1158 (VkPhysicalDeviceImagelessFramebufferFeaturesKHR
*)ext
;
1159 features
->imagelessFramebuffer
= true;
1163 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR
: {
1164 VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR
*features
=
1165 (VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR
*)ext
;
1166 features
->pipelineExecutableInfo
= true;
1170 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES
: {
1171 VkPhysicalDeviceProtectedMemoryFeatures
*features
= (void *)ext
;
1172 features
->protectedMemory
= false;
1176 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES
: {
1177 VkPhysicalDeviceSamplerYcbcrConversionFeatures
*features
=
1178 (VkPhysicalDeviceSamplerYcbcrConversionFeatures
*) ext
;
1179 features
->samplerYcbcrConversion
= true;
1183 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES_EXT
: {
1184 VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*features
=
1185 (VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*)ext
;
1186 features
->scalarBlockLayout
= true;
1190 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_INT64_FEATURES_KHR
: {
1191 VkPhysicalDeviceShaderAtomicInt64FeaturesKHR
*features
= (void *)ext
;
1192 features
->shaderBufferInt64Atomics
=
1193 pdevice
->info
.gen
>= 9 && pdevice
->use_softpin
;
1194 features
->shaderSharedInt64Atomics
= VK_FALSE
;
1198 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DEMOTE_TO_HELPER_INVOCATION_FEATURES_EXT
: {
1199 VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT
*features
= (void *)ext
;
1200 features
->shaderDemoteToHelperInvocation
= true;
1204 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETERS_FEATURES
: {
1205 VkPhysicalDeviceShaderDrawParametersFeatures
*features
= (void *)ext
;
1206 features
->shaderDrawParameters
= true;
1210 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_SUBGROUP_EXTENDED_TYPES_FEATURES_KHR
: {
1211 VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR
*features
=
1212 (VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR
*)ext
;
1213 features
->shaderSubgroupExtendedTypes
= true;
1217 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_FEATURES_EXT
: {
1218 VkPhysicalDeviceSubgroupSizeControlFeaturesEXT
*features
=
1219 (VkPhysicalDeviceSubgroupSizeControlFeaturesEXT
*)ext
;
1220 features
->subgroupSizeControl
= true;
1221 features
->computeFullSubgroups
= true;
1225 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_FEATURES_EXT
: {
1226 VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT
*features
=
1227 (VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT
*)ext
;
1228 features
->texelBufferAlignment
= true;
1232 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES
: {
1233 VkPhysicalDeviceVariablePointersFeatures
*features
= (void *)ext
;
1234 features
->variablePointersStorageBuffer
= true;
1235 features
->variablePointers
= true;
1239 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT
: {
1240 VkPhysicalDeviceTransformFeedbackFeaturesEXT
*features
=
1241 (VkPhysicalDeviceTransformFeedbackFeaturesEXT
*)ext
;
1242 features
->transformFeedback
= true;
1243 features
->geometryStreams
= true;
1247 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_UNIFORM_BUFFER_STANDARD_LAYOUT_FEATURES_KHR
: {
1248 VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR
*features
=
1249 (VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR
*)ext
;
1250 features
->uniformBufferStandardLayout
= true;
1254 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT
: {
1255 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*features
=
1256 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*)ext
;
1257 features
->vertexAttributeInstanceRateDivisor
= true;
1258 features
->vertexAttributeInstanceRateZeroDivisor
= true;
1262 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_YCBCR_IMAGE_ARRAYS_FEATURES_EXT
: {
1263 VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*features
=
1264 (VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*)ext
;
1265 features
->ycbcrImageArrays
= true;
1270 anv_debug_ignored_stype(ext
->sType
);
1276 #define MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS 64
1278 #define MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS 64
1279 #define MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS 256
1281 void anv_GetPhysicalDeviceProperties(
1282 VkPhysicalDevice physicalDevice
,
1283 VkPhysicalDeviceProperties
* pProperties
)
1285 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1286 const struct gen_device_info
*devinfo
= &pdevice
->info
;
1288 /* See assertions made when programming the buffer surface state. */
1289 const uint32_t max_raw_buffer_sz
= devinfo
->gen
>= 7 ?
1290 (1ul << 30) : (1ul << 27);
1292 const uint32_t max_ssbos
= pdevice
->has_a64_buffer_access
? UINT16_MAX
: 64;
1293 const uint32_t max_textures
=
1294 pdevice
->has_bindless_images
? UINT16_MAX
: 128;
1295 const uint32_t max_samplers
=
1296 pdevice
->has_bindless_samplers
? UINT16_MAX
:
1297 (devinfo
->gen
>= 8 || devinfo
->is_haswell
) ? 128 : 16;
1298 const uint32_t max_images
=
1299 pdevice
->has_bindless_images
? UINT16_MAX
: MAX_IMAGES
;
1301 /* If we can use bindless for everything, claim a high per-stage limit,
1302 * otherwise use the binding table size, minus the slots reserved for
1303 * render targets and one slot for the descriptor buffer. */
1304 const uint32_t max_per_stage
=
1305 pdevice
->has_bindless_images
&& pdevice
->has_a64_buffer_access
1306 ? UINT32_MAX
: MAX_BINDING_TABLE_SIZE
- MAX_RTS
- 1;
1308 const uint32_t max_workgroup_size
= 32 * devinfo
->max_cs_threads
;
1310 VkSampleCountFlags sample_counts
=
1311 isl_device_get_sample_counts(&pdevice
->isl_dev
);
1314 VkPhysicalDeviceLimits limits
= {
1315 .maxImageDimension1D
= (1 << 14),
1316 .maxImageDimension2D
= (1 << 14),
1317 .maxImageDimension3D
= (1 << 11),
1318 .maxImageDimensionCube
= (1 << 14),
1319 .maxImageArrayLayers
= (1 << 11),
1320 .maxTexelBufferElements
= 128 * 1024 * 1024,
1321 .maxUniformBufferRange
= (1ul << 27),
1322 .maxStorageBufferRange
= max_raw_buffer_sz
,
1323 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
1324 .maxMemoryAllocationCount
= UINT32_MAX
,
1325 .maxSamplerAllocationCount
= 64 * 1024,
1326 .bufferImageGranularity
= 64, /* A cache line */
1327 .sparseAddressSpaceSize
= 0,
1328 .maxBoundDescriptorSets
= MAX_SETS
,
1329 .maxPerStageDescriptorSamplers
= max_samplers
,
1330 .maxPerStageDescriptorUniformBuffers
= MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
,
1331 .maxPerStageDescriptorStorageBuffers
= max_ssbos
,
1332 .maxPerStageDescriptorSampledImages
= max_textures
,
1333 .maxPerStageDescriptorStorageImages
= max_images
,
1334 .maxPerStageDescriptorInputAttachments
= MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS
,
1335 .maxPerStageResources
= max_per_stage
,
1336 .maxDescriptorSetSamplers
= 6 * max_samplers
, /* number of stages * maxPerStageDescriptorSamplers */
1337 .maxDescriptorSetUniformBuffers
= 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
, /* number of stages * maxPerStageDescriptorUniformBuffers */
1338 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
1339 .maxDescriptorSetStorageBuffers
= 6 * max_ssbos
, /* number of stages * maxPerStageDescriptorStorageBuffers */
1340 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
1341 .maxDescriptorSetSampledImages
= 6 * max_textures
, /* number of stages * maxPerStageDescriptorSampledImages */
1342 .maxDescriptorSetStorageImages
= 6 * max_images
, /* number of stages * maxPerStageDescriptorStorageImages */
1343 .maxDescriptorSetInputAttachments
= MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS
,
1344 .maxVertexInputAttributes
= MAX_VBS
,
1345 .maxVertexInputBindings
= MAX_VBS
,
1346 .maxVertexInputAttributeOffset
= 2047,
1347 .maxVertexInputBindingStride
= 2048,
1348 .maxVertexOutputComponents
= 128,
1349 .maxTessellationGenerationLevel
= 64,
1350 .maxTessellationPatchSize
= 32,
1351 .maxTessellationControlPerVertexInputComponents
= 128,
1352 .maxTessellationControlPerVertexOutputComponents
= 128,
1353 .maxTessellationControlPerPatchOutputComponents
= 128,
1354 .maxTessellationControlTotalOutputComponents
= 2048,
1355 .maxTessellationEvaluationInputComponents
= 128,
1356 .maxTessellationEvaluationOutputComponents
= 128,
1357 .maxGeometryShaderInvocations
= 32,
1358 .maxGeometryInputComponents
= 64,
1359 .maxGeometryOutputComponents
= 128,
1360 .maxGeometryOutputVertices
= 256,
1361 .maxGeometryTotalOutputComponents
= 1024,
1362 .maxFragmentInputComponents
= 116, /* 128 components - (PSIZ, CLIP_DIST0, CLIP_DIST1) */
1363 .maxFragmentOutputAttachments
= 8,
1364 .maxFragmentDualSrcAttachments
= 1,
1365 .maxFragmentCombinedOutputResources
= 8,
1366 .maxComputeSharedMemorySize
= 64 * 1024,
1367 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
1368 .maxComputeWorkGroupInvocations
= max_workgroup_size
,
1369 .maxComputeWorkGroupSize
= {
1374 .subPixelPrecisionBits
= 8,
1375 .subTexelPrecisionBits
= 8,
1376 .mipmapPrecisionBits
= 8,
1377 .maxDrawIndexedIndexValue
= UINT32_MAX
,
1378 .maxDrawIndirectCount
= UINT32_MAX
,
1379 .maxSamplerLodBias
= 16,
1380 .maxSamplerAnisotropy
= 16,
1381 .maxViewports
= MAX_VIEWPORTS
,
1382 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
1383 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
1384 .viewportSubPixelBits
= 13, /* We take a float? */
1385 .minMemoryMapAlignment
= 4096, /* A page */
1386 /* The dataport requires texel alignment so we need to assume a worst
1387 * case of R32G32B32A32 which is 16 bytes.
1389 .minTexelBufferOffsetAlignment
= 16,
1390 /* We need 16 for UBO block reads to work and 32 for push UBOs */
1391 .minUniformBufferOffsetAlignment
= 32,
1392 .minStorageBufferOffsetAlignment
= 4,
1393 .minTexelOffset
= -8,
1394 .maxTexelOffset
= 7,
1395 .minTexelGatherOffset
= -32,
1396 .maxTexelGatherOffset
= 31,
1397 .minInterpolationOffset
= -0.5,
1398 .maxInterpolationOffset
= 0.4375,
1399 .subPixelInterpolationOffsetBits
= 4,
1400 .maxFramebufferWidth
= (1 << 14),
1401 .maxFramebufferHeight
= (1 << 14),
1402 .maxFramebufferLayers
= (1 << 11),
1403 .framebufferColorSampleCounts
= sample_counts
,
1404 .framebufferDepthSampleCounts
= sample_counts
,
1405 .framebufferStencilSampleCounts
= sample_counts
,
1406 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
1407 .maxColorAttachments
= MAX_RTS
,
1408 .sampledImageColorSampleCounts
= sample_counts
,
1409 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
1410 .sampledImageDepthSampleCounts
= sample_counts
,
1411 .sampledImageStencilSampleCounts
= sample_counts
,
1412 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
1413 .maxSampleMaskWords
= 1,
1414 .timestampComputeAndGraphics
= true,
1415 .timestampPeriod
= 1000000000.0 / devinfo
->timestamp_frequency
,
1416 .maxClipDistances
= 8,
1417 .maxCullDistances
= 8,
1418 .maxCombinedClipAndCullDistances
= 8,
1419 .discreteQueuePriorities
= 2,
1420 .pointSizeRange
= { 0.125, 255.875 },
1423 (devinfo
->gen
>= 9 || devinfo
->is_cherryview
) ?
1424 2047.9921875 : 7.9921875,
1426 .pointSizeGranularity
= (1.0 / 8.0),
1427 .lineWidthGranularity
= (1.0 / 128.0),
1428 .strictLines
= false,
1429 .standardSampleLocations
= true,
1430 .optimalBufferCopyOffsetAlignment
= 128,
1431 .optimalBufferCopyRowPitchAlignment
= 128,
1432 .nonCoherentAtomSize
= 64,
1435 *pProperties
= (VkPhysicalDeviceProperties
) {
1436 .apiVersion
= anv_physical_device_api_version(pdevice
),
1437 .driverVersion
= vk_get_driver_version(),
1439 .deviceID
= pdevice
->chipset_id
,
1440 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
1442 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
1445 snprintf(pProperties
->deviceName
, sizeof(pProperties
->deviceName
),
1446 "%s", pdevice
->name
);
1447 memcpy(pProperties
->pipelineCacheUUID
,
1448 pdevice
->pipeline_cache_uuid
, VK_UUID_SIZE
);
1451 void anv_GetPhysicalDeviceProperties2(
1452 VkPhysicalDevice physicalDevice
,
1453 VkPhysicalDeviceProperties2
* pProperties
)
1455 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1457 anv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
1459 vk_foreach_struct(ext
, pProperties
->pNext
) {
1460 switch (ext
->sType
) {
1461 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_STENCIL_RESOLVE_PROPERTIES_KHR
: {
1462 VkPhysicalDeviceDepthStencilResolvePropertiesKHR
*props
=
1463 (VkPhysicalDeviceDepthStencilResolvePropertiesKHR
*)ext
;
1465 /* We support all of the depth resolve modes */
1466 props
->supportedDepthResolveModes
=
1467 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
|
1468 VK_RESOLVE_MODE_AVERAGE_BIT_KHR
|
1469 VK_RESOLVE_MODE_MIN_BIT_KHR
|
1470 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1472 /* Average doesn't make sense for stencil so we don't support that */
1473 props
->supportedStencilResolveModes
=
1474 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
;
1475 if (pdevice
->info
.gen
>= 8) {
1476 /* The advanced stencil resolve modes currently require stencil
1477 * sampling be supported by the hardware.
1479 props
->supportedStencilResolveModes
|=
1480 VK_RESOLVE_MODE_MIN_BIT_KHR
|
1481 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1484 props
->independentResolveNone
= true;
1485 props
->independentResolve
= true;
1489 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES_EXT
: {
1490 VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*props
=
1491 (VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*)ext
;
1493 /* It's a bit hard to exactly map our implementation to the limits
1494 * described here. The bindless surface handle in the extended
1495 * message descriptors is 20 bits and it's an index into the table of
1496 * RENDER_SURFACE_STATE structs that starts at bindless surface base
1497 * address. Given that most things consume two surface states per
1498 * view (general/sampled for textures and write-only/read-write for
1499 * images), we claim 2^19 things.
1501 * For SSBOs, we just use A64 messages so there is no real limit
1502 * there beyond the limit on the total size of a descriptor set.
1504 const unsigned max_bindless_views
= 1 << 19;
1506 props
->maxUpdateAfterBindDescriptorsInAllPools
= max_bindless_views
;
1507 props
->shaderUniformBufferArrayNonUniformIndexingNative
= false;
1508 props
->shaderSampledImageArrayNonUniformIndexingNative
= false;
1509 props
->shaderStorageBufferArrayNonUniformIndexingNative
= true;
1510 props
->shaderStorageImageArrayNonUniformIndexingNative
= false;
1511 props
->shaderInputAttachmentArrayNonUniformIndexingNative
= false;
1512 props
->robustBufferAccessUpdateAfterBind
= true;
1513 props
->quadDivergentImplicitLod
= false;
1514 props
->maxPerStageDescriptorUpdateAfterBindSamplers
= max_bindless_views
;
1515 props
->maxPerStageDescriptorUpdateAfterBindUniformBuffers
= MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
;
1516 props
->maxPerStageDescriptorUpdateAfterBindStorageBuffers
= UINT32_MAX
;
1517 props
->maxPerStageDescriptorUpdateAfterBindSampledImages
= max_bindless_views
;
1518 props
->maxPerStageDescriptorUpdateAfterBindStorageImages
= max_bindless_views
;
1519 props
->maxPerStageDescriptorUpdateAfterBindInputAttachments
= MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS
;
1520 props
->maxPerStageUpdateAfterBindResources
= UINT32_MAX
;
1521 props
->maxDescriptorSetUpdateAfterBindSamplers
= max_bindless_views
;
1522 props
->maxDescriptorSetUpdateAfterBindUniformBuffers
= 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
;
1523 props
->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2;
1524 props
->maxDescriptorSetUpdateAfterBindStorageBuffers
= UINT32_MAX
;
1525 props
->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2;
1526 props
->maxDescriptorSetUpdateAfterBindSampledImages
= max_bindless_views
;
1527 props
->maxDescriptorSetUpdateAfterBindStorageImages
= max_bindless_views
;
1528 props
->maxDescriptorSetUpdateAfterBindInputAttachments
= MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS
;
1532 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR
: {
1533 VkPhysicalDeviceDriverPropertiesKHR
*driver_props
=
1534 (VkPhysicalDeviceDriverPropertiesKHR
*) ext
;
1536 driver_props
->driverID
= VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA_KHR
;
1537 snprintf(driver_props
->driverName
, VK_MAX_DRIVER_NAME_SIZE_KHR
,
1538 "Intel open-source Mesa driver");
1540 snprintf(driver_props
->driverInfo
, VK_MAX_DRIVER_INFO_SIZE_KHR
,
1541 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1
);
1543 driver_props
->conformanceVersion
= (VkConformanceVersionKHR
) {
1552 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT
: {
1553 VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*props
=
1554 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*) ext
;
1555 /* Userptr needs page aligned memory. */
1556 props
->minImportedHostPointerAlignment
= 4096;
1560 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES
: {
1561 VkPhysicalDeviceIDProperties
*id_props
=
1562 (VkPhysicalDeviceIDProperties
*)ext
;
1563 memcpy(id_props
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
1564 memcpy(id_props
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
1565 /* The LUID is for Windows. */
1566 id_props
->deviceLUIDValid
= false;
1570 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_PROPERTIES_EXT
: {
1571 VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*props
=
1572 (VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*)ext
;
1573 props
->maxInlineUniformBlockSize
= MAX_INLINE_UNIFORM_BLOCK_SIZE
;
1574 props
->maxPerStageDescriptorInlineUniformBlocks
=
1575 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1576 props
->maxPerStageDescriptorUpdateAfterBindInlineUniformBlocks
=
1577 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1578 props
->maxDescriptorSetInlineUniformBlocks
=
1579 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1580 props
->maxDescriptorSetUpdateAfterBindInlineUniformBlocks
=
1581 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1585 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_PROPERTIES_EXT
: {
1586 VkPhysicalDeviceLineRasterizationPropertiesEXT
*props
=
1587 (VkPhysicalDeviceLineRasterizationPropertiesEXT
*)ext
;
1588 /* In the Skylake PRM Vol. 7, subsection titled "GIQ (Diamond)
1589 * Sampling Rules - Legacy Mode", it says the following:
1591 * "Note that the device divides a pixel into a 16x16 array of
1592 * subpixels, referenced by their upper left corners."
1594 * This is the only known reference in the PRMs to the subpixel
1595 * precision of line rasterization and a "16x16 array of subpixels"
1596 * implies 4 subpixel precision bits. Empirical testing has shown
1597 * that 4 subpixel precision bits applies to all line rasterization
1600 props
->lineSubPixelPrecisionBits
= 4;
1604 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES
: {
1605 VkPhysicalDeviceMaintenance3Properties
*props
=
1606 (VkPhysicalDeviceMaintenance3Properties
*)ext
;
1607 /* This value doesn't matter for us today as our per-stage
1608 * descriptors are the real limit.
1610 props
->maxPerSetDescriptors
= 1024;
1611 props
->maxMemoryAllocationSize
= MAX_MEMORY_ALLOCATION_SIZE
;
1615 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES
: {
1616 VkPhysicalDeviceMultiviewProperties
*properties
=
1617 (VkPhysicalDeviceMultiviewProperties
*)ext
;
1618 properties
->maxMultiviewViewCount
= 16;
1619 properties
->maxMultiviewInstanceIndex
= UINT32_MAX
/ 16;
1623 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT
: {
1624 VkPhysicalDevicePCIBusInfoPropertiesEXT
*properties
=
1625 (VkPhysicalDevicePCIBusInfoPropertiesEXT
*)ext
;
1626 properties
->pciDomain
= pdevice
->pci_info
.domain
;
1627 properties
->pciBus
= pdevice
->pci_info
.bus
;
1628 properties
->pciDevice
= pdevice
->pci_info
.device
;
1629 properties
->pciFunction
= pdevice
->pci_info
.function
;
1633 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES
: {
1634 VkPhysicalDevicePointClippingProperties
*properties
=
1635 (VkPhysicalDevicePointClippingProperties
*) ext
;
1636 properties
->pointClippingBehavior
= VK_POINT_CLIPPING_BEHAVIOR_USER_CLIP_PLANES_ONLY
;
1640 #pragma GCC diagnostic push
1641 #pragma GCC diagnostic ignored "-Wswitch"
1642 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PRESENTATION_PROPERTIES_ANDROID
: {
1643 VkPhysicalDevicePresentationPropertiesANDROID
*props
=
1644 (VkPhysicalDevicePresentationPropertiesANDROID
*)ext
;
1645 props
->sharedImage
= VK_FALSE
;
1648 #pragma GCC diagnostic pop
1650 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES
: {
1651 VkPhysicalDeviceProtectedMemoryProperties
*props
=
1652 (VkPhysicalDeviceProtectedMemoryProperties
*)ext
;
1653 props
->protectedNoFault
= false;
1657 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
1658 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
1659 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
1661 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
1665 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT
: {
1666 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*properties
=
1667 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*)ext
;
1668 properties
->filterMinmaxImageComponentMapping
= pdevice
->info
.gen
>= 9;
1669 properties
->filterMinmaxSingleComponentFormats
= true;
1673 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES
: {
1674 VkPhysicalDeviceSubgroupProperties
*properties
= (void *)ext
;
1676 properties
->subgroupSize
= BRW_SUBGROUP_SIZE
;
1678 VkShaderStageFlags scalar_stages
= 0;
1679 for (unsigned stage
= 0; stage
< MESA_SHADER_STAGES
; stage
++) {
1680 if (pdevice
->compiler
->scalar_stage
[stage
])
1681 scalar_stages
|= mesa_to_vk_shader_stage(stage
);
1683 properties
->supportedStages
= scalar_stages
;
1685 properties
->supportedOperations
= VK_SUBGROUP_FEATURE_BASIC_BIT
|
1686 VK_SUBGROUP_FEATURE_VOTE_BIT
|
1687 VK_SUBGROUP_FEATURE_BALLOT_BIT
|
1688 VK_SUBGROUP_FEATURE_SHUFFLE_BIT
|
1689 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT
|
1690 VK_SUBGROUP_FEATURE_QUAD_BIT
;
1691 if (pdevice
->info
.gen
>= 8) {
1692 /* TODO: There's no technical reason why these can't be made to
1693 * work on gen7 but they don't at the moment so it's best to leave
1694 * the feature disabled than enabled and broken.
1696 properties
->supportedOperations
|=
1697 VK_SUBGROUP_FEATURE_ARITHMETIC_BIT
|
1698 VK_SUBGROUP_FEATURE_CLUSTERED_BIT
;
1700 properties
->quadOperationsInAllStages
= pdevice
->info
.gen
>= 8;
1704 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_PROPERTIES_EXT
: {
1705 VkPhysicalDeviceSubgroupSizeControlPropertiesEXT
*props
=
1706 (VkPhysicalDeviceSubgroupSizeControlPropertiesEXT
*)ext
;
1707 STATIC_ASSERT(8 <= BRW_SUBGROUP_SIZE
&& BRW_SUBGROUP_SIZE
<= 32);
1708 props
->minSubgroupSize
= 8;
1709 props
->maxSubgroupSize
= 32;
1710 props
->maxComputeWorkgroupSubgroups
= pdevice
->info
.max_cs_threads
;
1711 props
->requiredSubgroupSizeStages
= VK_SHADER_STAGE_COMPUTE_BIT
;
1714 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT_CONTROLS_PROPERTIES_KHR
: {
1715 VkPhysicalDeviceFloatControlsPropertiesKHR
*properties
= (void *)ext
;
1716 properties
->denormBehaviorIndependence
= VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_ALL_KHR
;
1717 properties
->roundingModeIndependence
= VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_NONE_KHR
;
1719 /* Broadwell does not support HF denorms and there are restrictions
1720 * other gens. According to Kabylake's PRM:
1722 * "math - Extended Math Function
1724 * Restriction : Half-float denorms are always retained."
1726 properties
->shaderDenormFlushToZeroFloat16
= false;
1727 properties
->shaderDenormPreserveFloat16
= pdevice
->info
.gen
> 8;
1728 properties
->shaderRoundingModeRTEFloat16
= true;
1729 properties
->shaderRoundingModeRTZFloat16
= true;
1730 properties
->shaderSignedZeroInfNanPreserveFloat16
= true;
1732 properties
->shaderDenormFlushToZeroFloat32
= true;
1733 properties
->shaderDenormPreserveFloat32
= true;
1734 properties
->shaderRoundingModeRTEFloat32
= true;
1735 properties
->shaderRoundingModeRTZFloat32
= true;
1736 properties
->shaderSignedZeroInfNanPreserveFloat32
= true;
1738 properties
->shaderDenormFlushToZeroFloat64
= true;
1739 properties
->shaderDenormPreserveFloat64
= true;
1740 properties
->shaderRoundingModeRTEFloat64
= true;
1741 properties
->shaderRoundingModeRTZFloat64
= true;
1742 properties
->shaderSignedZeroInfNanPreserveFloat64
= true;
1746 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_PROPERTIES_EXT
: {
1747 VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT
*props
=
1748 (VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT
*)ext
;
1750 /* From the SKL PRM Vol. 2d, docs for RENDER_SURFACE_STATE::Surface
1753 * "For SURFTYPE_BUFFER non-rendertarget surfaces, this field
1754 * specifies the base address of the first element of the surface,
1755 * computed in software by adding the surface base address to the
1756 * byte offset of the element in the buffer. The base address must
1757 * be aligned to element size."
1759 * The typed dataport messages require that things be texel aligned.
1760 * Otherwise, we may just load/store the wrong data or, in the worst
1761 * case, there may be hangs.
1763 props
->storageTexelBufferOffsetAlignmentBytes
= 16;
1764 props
->storageTexelBufferOffsetSingleTexelAlignment
= true;
1766 /* The sampler, however, is much more forgiving and it can handle
1767 * arbitrary byte alignment for linear and buffer surfaces. It's
1768 * hard to find a good PRM citation for this but years of empirical
1769 * experience demonstrate that this is true.
1771 props
->uniformTexelBufferOffsetAlignmentBytes
= 1;
1772 props
->uniformTexelBufferOffsetSingleTexelAlignment
= false;
1776 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT
: {
1777 VkPhysicalDeviceTransformFeedbackPropertiesEXT
*props
=
1778 (VkPhysicalDeviceTransformFeedbackPropertiesEXT
*)ext
;
1780 props
->maxTransformFeedbackStreams
= MAX_XFB_STREAMS
;
1781 props
->maxTransformFeedbackBuffers
= MAX_XFB_BUFFERS
;
1782 props
->maxTransformFeedbackBufferSize
= (1ull << 32);
1783 props
->maxTransformFeedbackStreamDataSize
= 128 * 4;
1784 props
->maxTransformFeedbackBufferDataSize
= 128 * 4;
1785 props
->maxTransformFeedbackBufferDataStride
= 2048;
1786 props
->transformFeedbackQueries
= true;
1787 props
->transformFeedbackStreamsLinesTriangles
= false;
1788 props
->transformFeedbackRasterizationStreamSelect
= false;
1789 props
->transformFeedbackDraw
= true;
1793 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT
: {
1794 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*props
=
1795 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*)ext
;
1796 /* We have to restrict this a bit for multiview */
1797 props
->maxVertexAttribDivisor
= UINT32_MAX
/ 16;
1802 anv_debug_ignored_stype(ext
->sType
);
1808 /* We support exactly one queue family. */
1809 static const VkQueueFamilyProperties
1810 anv_queue_family_properties
= {
1811 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
1812 VK_QUEUE_COMPUTE_BIT
|
1813 VK_QUEUE_TRANSFER_BIT
,
1815 .timestampValidBits
= 36, /* XXX: Real value here */
1816 .minImageTransferGranularity
= { 1, 1, 1 },
1819 void anv_GetPhysicalDeviceQueueFamilyProperties(
1820 VkPhysicalDevice physicalDevice
,
1822 VkQueueFamilyProperties
* pQueueFamilyProperties
)
1824 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pCount
);
1826 vk_outarray_append(&out
, p
) {
1827 *p
= anv_queue_family_properties
;
1831 void anv_GetPhysicalDeviceQueueFamilyProperties2(
1832 VkPhysicalDevice physicalDevice
,
1833 uint32_t* pQueueFamilyPropertyCount
,
1834 VkQueueFamilyProperties2
* pQueueFamilyProperties
)
1837 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pQueueFamilyPropertyCount
);
1839 vk_outarray_append(&out
, p
) {
1840 p
->queueFamilyProperties
= anv_queue_family_properties
;
1842 vk_foreach_struct(s
, p
->pNext
) {
1843 anv_debug_ignored_stype(s
->sType
);
1848 void anv_GetPhysicalDeviceMemoryProperties(
1849 VkPhysicalDevice physicalDevice
,
1850 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
1852 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
1854 pMemoryProperties
->memoryTypeCount
= physical_device
->memory
.type_count
;
1855 for (uint32_t i
= 0; i
< physical_device
->memory
.type_count
; i
++) {
1856 pMemoryProperties
->memoryTypes
[i
] = (VkMemoryType
) {
1857 .propertyFlags
= physical_device
->memory
.types
[i
].propertyFlags
,
1858 .heapIndex
= physical_device
->memory
.types
[i
].heapIndex
,
1862 pMemoryProperties
->memoryHeapCount
= physical_device
->memory
.heap_count
;
1863 for (uint32_t i
= 0; i
< physical_device
->memory
.heap_count
; i
++) {
1864 pMemoryProperties
->memoryHeaps
[i
] = (VkMemoryHeap
) {
1865 .size
= physical_device
->memory
.heaps
[i
].size
,
1866 .flags
= physical_device
->memory
.heaps
[i
].flags
,
1872 anv_get_memory_budget(VkPhysicalDevice physicalDevice
,
1873 VkPhysicalDeviceMemoryBudgetPropertiesEXT
*memoryBudget
)
1875 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
1876 uint64_t sys_available
= get_available_system_memory();
1877 assert(sys_available
> 0);
1879 VkDeviceSize total_heaps_size
= 0;
1880 for (size_t i
= 0; i
< device
->memory
.heap_count
; i
++)
1881 total_heaps_size
+= device
->memory
.heaps
[i
].size
;
1883 for (size_t i
= 0; i
< device
->memory
.heap_count
; i
++) {
1884 VkDeviceSize heap_size
= device
->memory
.heaps
[i
].size
;
1885 VkDeviceSize heap_used
= device
->memory
.heaps
[i
].used
;
1886 VkDeviceSize heap_budget
;
1888 double heap_proportion
= (double) heap_size
/ total_heaps_size
;
1889 VkDeviceSize sys_available_prop
= sys_available
* heap_proportion
;
1892 * Let's not incite the app to starve the system: report at most 90% of
1893 * available system memory.
1895 uint64_t heap_available
= sys_available_prop
* 9 / 10;
1896 heap_budget
= MIN2(heap_size
, heap_used
+ heap_available
);
1899 * Round down to the nearest MB
1901 heap_budget
&= ~((1ull << 20) - 1);
1904 * The heapBudget value must be non-zero for array elements less than
1905 * VkPhysicalDeviceMemoryProperties::memoryHeapCount. The heapBudget
1906 * value must be less than or equal to VkMemoryHeap::size for each heap.
1908 assert(0 < heap_budget
&& heap_budget
<= heap_size
);
1910 memoryBudget
->heapUsage
[i
] = heap_used
;
1911 memoryBudget
->heapBudget
[i
] = heap_budget
;
1914 /* The heapBudget and heapUsage values must be zero for array elements
1915 * greater than or equal to VkPhysicalDeviceMemoryProperties::memoryHeapCount
1917 for (uint32_t i
= device
->memory
.heap_count
; i
< VK_MAX_MEMORY_HEAPS
; i
++) {
1918 memoryBudget
->heapBudget
[i
] = 0;
1919 memoryBudget
->heapUsage
[i
] = 0;
1923 void anv_GetPhysicalDeviceMemoryProperties2(
1924 VkPhysicalDevice physicalDevice
,
1925 VkPhysicalDeviceMemoryProperties2
* pMemoryProperties
)
1927 anv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
1928 &pMemoryProperties
->memoryProperties
);
1930 vk_foreach_struct(ext
, pMemoryProperties
->pNext
) {
1931 switch (ext
->sType
) {
1932 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT
:
1933 anv_get_memory_budget(physicalDevice
, (void*)ext
);
1936 anv_debug_ignored_stype(ext
->sType
);
1943 anv_GetDeviceGroupPeerMemoryFeatures(
1946 uint32_t localDeviceIndex
,
1947 uint32_t remoteDeviceIndex
,
1948 VkPeerMemoryFeatureFlags
* pPeerMemoryFeatures
)
1950 assert(localDeviceIndex
== 0 && remoteDeviceIndex
== 0);
1951 *pPeerMemoryFeatures
= VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT
|
1952 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT
|
1953 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT
|
1954 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT
;
1957 PFN_vkVoidFunction
anv_GetInstanceProcAddr(
1958 VkInstance _instance
,
1961 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1963 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
1964 * when we have to return valid function pointers, NULL, or it's left
1965 * undefined. See the table for exact details.
1970 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
1971 if (strcmp(pName, "vk" #entrypoint) == 0) \
1972 return (PFN_vkVoidFunction)anv_##entrypoint
1974 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties
);
1975 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties
);
1976 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceVersion
);
1977 LOOKUP_ANV_ENTRYPOINT(CreateInstance
);
1979 #undef LOOKUP_ANV_ENTRYPOINT
1981 if (instance
== NULL
)
1984 int idx
= anv_get_instance_entrypoint_index(pName
);
1986 return instance
->dispatch
.entrypoints
[idx
];
1988 idx
= anv_get_physical_device_entrypoint_index(pName
);
1990 return instance
->physicalDevice
.dispatch
.entrypoints
[idx
];
1992 idx
= anv_get_device_entrypoint_index(pName
);
1994 return instance
->device_dispatch
.entrypoints
[idx
];
1999 /* With version 1+ of the loader interface the ICD should expose
2000 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
2003 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
2004 VkInstance instance
,
2008 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
2009 VkInstance instance
,
2012 return anv_GetInstanceProcAddr(instance
, pName
);
2015 PFN_vkVoidFunction
anv_GetDeviceProcAddr(
2019 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2021 if (!device
|| !pName
)
2024 int idx
= anv_get_device_entrypoint_index(pName
);
2028 return device
->dispatch
.entrypoints
[idx
];
2031 /* With version 4+ of the loader interface the ICD should expose
2032 * vk_icdGetPhysicalDeviceProcAddr()
2035 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetPhysicalDeviceProcAddr(
2036 VkInstance _instance
,
2039 PFN_vkVoidFunction
vk_icdGetPhysicalDeviceProcAddr(
2040 VkInstance _instance
,
2043 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2045 if (!pName
|| !instance
)
2048 int idx
= anv_get_physical_device_entrypoint_index(pName
);
2052 return instance
->physicalDevice
.dispatch
.entrypoints
[idx
];
2057 anv_CreateDebugReportCallbackEXT(VkInstance _instance
,
2058 const VkDebugReportCallbackCreateInfoEXT
* pCreateInfo
,
2059 const VkAllocationCallbacks
* pAllocator
,
2060 VkDebugReportCallbackEXT
* pCallback
)
2062 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2063 return vk_create_debug_report_callback(&instance
->debug_report_callbacks
,
2064 pCreateInfo
, pAllocator
, &instance
->alloc
,
2069 anv_DestroyDebugReportCallbackEXT(VkInstance _instance
,
2070 VkDebugReportCallbackEXT _callback
,
2071 const VkAllocationCallbacks
* pAllocator
)
2073 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2074 vk_destroy_debug_report_callback(&instance
->debug_report_callbacks
,
2075 _callback
, pAllocator
, &instance
->alloc
);
2079 anv_DebugReportMessageEXT(VkInstance _instance
,
2080 VkDebugReportFlagsEXT flags
,
2081 VkDebugReportObjectTypeEXT objectType
,
2084 int32_t messageCode
,
2085 const char* pLayerPrefix
,
2086 const char* pMessage
)
2088 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2089 vk_debug_report(&instance
->debug_report_callbacks
, flags
, objectType
,
2090 object
, location
, messageCode
, pLayerPrefix
, pMessage
);
2094 anv_queue_init(struct anv_device
*device
, struct anv_queue
*queue
)
2096 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
2097 queue
->device
= device
;
2102 anv_queue_finish(struct anv_queue
*queue
)
2106 static struct anv_state
2107 anv_state_pool_emit_data(struct anv_state_pool
*pool
, size_t size
, size_t align
, const void *p
)
2109 struct anv_state state
;
2111 state
= anv_state_pool_alloc(pool
, size
, align
);
2112 memcpy(state
.map
, p
, size
);
2117 /* Haswell border color is a bit of a disaster. Float and unorm formats use a
2118 * straightforward 32-bit float color in the first 64 bytes. Instead of using
2119 * a nice float/integer union like Gen8+, Haswell specifies the integer border
2120 * color as a separate entry /after/ the float color. The layout of this entry
2121 * also depends on the format's bpp (with extra hacks for RG32), and overlaps.
2123 * Since we don't know the format/bpp, we can't make any of the border colors
2124 * containing '1' work for all formats, as it would be in the wrong place for
2125 * some of them. We opt to make 32-bit integers work as this seems like the
2126 * most common option. Fortunately, transparent black works regardless, as
2127 * all zeroes is the same in every bit-size.
2129 struct hsw_border_color
{
2133 uint32_t _pad1
[108];
2136 struct gen8_border_color
{
2141 /* Pad out to 64 bytes */
2146 anv_device_init_border_colors(struct anv_device
*device
)
2148 if (device
->info
.is_haswell
) {
2149 static const struct hsw_border_color border_colors
[] = {
2150 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
2151 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
2152 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
2153 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
2154 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
2155 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
2158 device
->border_colors
=
2159 anv_state_pool_emit_data(&device
->dynamic_state_pool
,
2160 sizeof(border_colors
), 512, border_colors
);
2162 static const struct gen8_border_color border_colors
[] = {
2163 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
2164 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
2165 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
2166 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
2167 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
2168 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
2171 device
->border_colors
=
2172 anv_state_pool_emit_data(&device
->dynamic_state_pool
,
2173 sizeof(border_colors
), 64, border_colors
);
2178 anv_device_init_trivial_batch(struct anv_device
*device
)
2180 anv_bo_init_new(&device
->trivial_batch_bo
, device
, 4096);
2182 if (device
->instance
->physicalDevice
.has_exec_async
)
2183 device
->trivial_batch_bo
.flags
|= EXEC_OBJECT_ASYNC
;
2185 if (device
->instance
->physicalDevice
.use_softpin
)
2186 device
->trivial_batch_bo
.flags
|= EXEC_OBJECT_PINNED
;
2188 anv_vma_alloc(device
, &device
->trivial_batch_bo
);
2190 void *map
= anv_gem_mmap(device
, device
->trivial_batch_bo
.gem_handle
,
2193 struct anv_batch batch
= {
2199 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
2200 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
2202 if (!device
->info
.has_llc
)
2203 gen_clflush_range(map
, batch
.next
- map
);
2205 anv_gem_munmap(map
, device
->trivial_batch_bo
.size
);
2208 VkResult
anv_EnumerateDeviceExtensionProperties(
2209 VkPhysicalDevice physicalDevice
,
2210 const char* pLayerName
,
2211 uint32_t* pPropertyCount
,
2212 VkExtensionProperties
* pProperties
)
2214 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
2215 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
2217 for (int i
= 0; i
< ANV_DEVICE_EXTENSION_COUNT
; i
++) {
2218 if (device
->supported_extensions
.extensions
[i
]) {
2219 vk_outarray_append(&out
, prop
) {
2220 *prop
= anv_device_extensions
[i
];
2225 return vk_outarray_status(&out
);
2229 anv_device_init_dispatch(struct anv_device
*device
)
2231 const struct anv_device_dispatch_table
*genX_table
;
2232 switch (device
->info
.gen
) {
2234 genX_table
= &gen12_device_dispatch_table
;
2237 genX_table
= &gen11_device_dispatch_table
;
2240 genX_table
= &gen10_device_dispatch_table
;
2243 genX_table
= &gen9_device_dispatch_table
;
2246 genX_table
= &gen8_device_dispatch_table
;
2249 if (device
->info
.is_haswell
)
2250 genX_table
= &gen75_device_dispatch_table
;
2252 genX_table
= &gen7_device_dispatch_table
;
2255 unreachable("unsupported gen\n");
2258 for (unsigned i
= 0; i
< ARRAY_SIZE(device
->dispatch
.entrypoints
); i
++) {
2259 /* Vulkan requires that entrypoints for extensions which have not been
2260 * enabled must not be advertised.
2262 if (!anv_device_entrypoint_is_enabled(i
, device
->instance
->app_info
.api_version
,
2263 &device
->instance
->enabled_extensions
,
2264 &device
->enabled_extensions
)) {
2265 device
->dispatch
.entrypoints
[i
] = NULL
;
2266 } else if (genX_table
->entrypoints
[i
]) {
2267 device
->dispatch
.entrypoints
[i
] = genX_table
->entrypoints
[i
];
2269 device
->dispatch
.entrypoints
[i
] =
2270 anv_device_dispatch_table
.entrypoints
[i
];
2276 vk_priority_to_gen(int priority
)
2279 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT
:
2280 return GEN_CONTEXT_LOW_PRIORITY
;
2281 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
:
2282 return GEN_CONTEXT_MEDIUM_PRIORITY
;
2283 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT
:
2284 return GEN_CONTEXT_HIGH_PRIORITY
;
2285 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT
:
2286 return GEN_CONTEXT_REALTIME_PRIORITY
;
2288 unreachable("Invalid priority");
2293 anv_device_init_hiz_clear_value_bo(struct anv_device
*device
)
2295 anv_bo_init_new(&device
->hiz_clear_bo
, device
, 4096);
2297 if (device
->instance
->physicalDevice
.has_exec_async
)
2298 device
->hiz_clear_bo
.flags
|= EXEC_OBJECT_ASYNC
;
2300 if (device
->instance
->physicalDevice
.use_softpin
)
2301 device
->hiz_clear_bo
.flags
|= EXEC_OBJECT_PINNED
;
2303 anv_vma_alloc(device
, &device
->hiz_clear_bo
);
2305 uint32_t *map
= anv_gem_mmap(device
, device
->hiz_clear_bo
.gem_handle
,
2308 union isl_color_value hiz_clear
= { .u32
= { 0, } };
2309 hiz_clear
.f32
[0] = ANV_HZ_FC_VAL
;
2311 memcpy(map
, hiz_clear
.u32
, sizeof(hiz_clear
.u32
));
2312 anv_gem_munmap(map
, device
->hiz_clear_bo
.size
);
2316 get_bo_from_pool(struct gen_batch_decode_bo
*ret
,
2317 struct anv_block_pool
*pool
,
2320 for (uint32_t i
= 0; i
< pool
->nbos
; i
++) {
2321 uint64_t bo_address
= pool
->bos
[i
].offset
& (~0ull >> 16);
2322 uint32_t bo_size
= pool
->bos
[i
].size
;
2323 if (address
>= bo_address
&& address
< (bo_address
+ bo_size
)) {
2324 *ret
= (struct gen_batch_decode_bo
) {
2327 .map
= pool
->bos
[i
].map
,
2335 /* Finding a buffer for batch decoding */
2336 static struct gen_batch_decode_bo
2337 decode_get_bo(void *v_batch
, bool ppgtt
, uint64_t address
)
2339 struct anv_device
*device
= v_batch
;
2340 struct gen_batch_decode_bo ret_bo
= {};
2344 if (get_bo_from_pool(&ret_bo
, &device
->dynamic_state_pool
.block_pool
, address
))
2346 if (get_bo_from_pool(&ret_bo
, &device
->instruction_state_pool
.block_pool
, address
))
2348 if (get_bo_from_pool(&ret_bo
, &device
->binding_table_pool
.block_pool
, address
))
2350 if (get_bo_from_pool(&ret_bo
, &device
->surface_state_pool
.block_pool
, address
))
2353 if (!device
->cmd_buffer_being_decoded
)
2354 return (struct gen_batch_decode_bo
) { };
2356 struct anv_batch_bo
**bo
;
2358 u_vector_foreach(bo
, &device
->cmd_buffer_being_decoded
->seen_bbos
) {
2359 /* The decoder zeroes out the top 16 bits, so we need to as well */
2360 uint64_t bo_address
= (*bo
)->bo
.offset
& (~0ull >> 16);
2362 if (address
>= bo_address
&& address
< bo_address
+ (*bo
)->bo
.size
) {
2363 return (struct gen_batch_decode_bo
) {
2365 .size
= (*bo
)->bo
.size
,
2366 .map
= (*bo
)->bo
.map
,
2371 return (struct gen_batch_decode_bo
) { };
2374 VkResult
anv_CreateDevice(
2375 VkPhysicalDevice physicalDevice
,
2376 const VkDeviceCreateInfo
* pCreateInfo
,
2377 const VkAllocationCallbacks
* pAllocator
,
2380 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
2382 struct anv_device
*device
;
2384 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO
);
2386 struct anv_device_extension_table enabled_extensions
= { };
2387 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
2389 for (idx
= 0; idx
< ANV_DEVICE_EXTENSION_COUNT
; idx
++) {
2390 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
2391 anv_device_extensions
[idx
].extensionName
) == 0)
2395 if (idx
>= ANV_DEVICE_EXTENSION_COUNT
)
2396 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
2398 if (!physical_device
->supported_extensions
.extensions
[idx
])
2399 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
2401 enabled_extensions
.extensions
[idx
] = true;
2404 /* Check enabled features */
2405 if (pCreateInfo
->pEnabledFeatures
) {
2406 VkPhysicalDeviceFeatures supported_features
;
2407 anv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
2408 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
2409 VkBool32
*enabled_feature
= (VkBool32
*)pCreateInfo
->pEnabledFeatures
;
2410 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
2411 for (uint32_t i
= 0; i
< num_features
; i
++) {
2412 if (enabled_feature
[i
] && !supported_feature
[i
])
2413 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
2417 /* Check requested queues and fail if we are requested to create any
2418 * queues with flags we don't support.
2420 assert(pCreateInfo
->queueCreateInfoCount
> 0);
2421 for (uint32_t i
= 0; i
< pCreateInfo
->queueCreateInfoCount
; i
++) {
2422 if (pCreateInfo
->pQueueCreateInfos
[i
].flags
!= 0)
2423 return vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2426 /* Check if client specified queue priority. */
2427 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*queue_priority
=
2428 vk_find_struct_const(pCreateInfo
->pQueueCreateInfos
[0].pNext
,
2429 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT
);
2431 VkQueueGlobalPriorityEXT priority
=
2432 queue_priority
? queue_priority
->globalPriority
:
2433 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
;
2435 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
2437 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
2439 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2441 if (INTEL_DEBUG
& DEBUG_BATCH
) {
2442 const unsigned decode_flags
=
2443 GEN_BATCH_DECODE_FULL
|
2444 ((INTEL_DEBUG
& DEBUG_COLOR
) ? GEN_BATCH_DECODE_IN_COLOR
: 0) |
2445 GEN_BATCH_DECODE_OFFSETS
|
2446 GEN_BATCH_DECODE_FLOATS
;
2448 gen_batch_decode_ctx_init(&device
->decoder_ctx
,
2449 &physical_device
->info
,
2450 stderr
, decode_flags
, NULL
,
2451 decode_get_bo
, NULL
, device
);
2454 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
2455 device
->instance
= physical_device
->instance
;
2456 device
->chipset_id
= physical_device
->chipset_id
;
2457 device
->no_hw
= physical_device
->no_hw
;
2458 device
->_lost
= false;
2461 device
->alloc
= *pAllocator
;
2463 device
->alloc
= physical_device
->instance
->alloc
;
2465 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
2466 device
->fd
= open(physical_device
->path
, O_RDWR
| O_CLOEXEC
);
2467 if (device
->fd
== -1) {
2468 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2472 device
->context_id
= anv_gem_create_context(device
);
2473 if (device
->context_id
== -1) {
2474 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2478 if (physical_device
->use_softpin
) {
2479 if (pthread_mutex_init(&device
->vma_mutex
, NULL
) != 0) {
2480 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2484 /* keep the page with address zero out of the allocator */
2485 struct anv_memory_heap
*low_heap
=
2486 &physical_device
->memory
.heaps
[physical_device
->memory
.heap_count
- 1];
2487 util_vma_heap_init(&device
->vma_lo
, low_heap
->vma_start
, low_heap
->vma_size
);
2488 device
->vma_lo_available
= low_heap
->size
;
2490 struct anv_memory_heap
*high_heap
=
2491 &physical_device
->memory
.heaps
[0];
2492 util_vma_heap_init(&device
->vma_hi
, high_heap
->vma_start
, high_heap
->vma_size
);
2493 device
->vma_hi_available
= physical_device
->memory
.heap_count
== 1 ? 0 :
2497 list_inithead(&device
->memory_objects
);
2499 /* As per spec, the driver implementation may deny requests to acquire
2500 * a priority above the default priority (MEDIUM) if the caller does not
2501 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
2504 if (physical_device
->has_context_priority
) {
2505 int err
= anv_gem_set_context_param(device
->fd
, device
->context_id
,
2506 I915_CONTEXT_PARAM_PRIORITY
,
2507 vk_priority_to_gen(priority
));
2508 if (err
!= 0 && priority
> VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
) {
2509 result
= vk_error(VK_ERROR_NOT_PERMITTED_EXT
);
2514 device
->info
= physical_device
->info
;
2515 device
->isl_dev
= physical_device
->isl_dev
;
2517 /* On Broadwell and later, we can use batch chaining to more efficiently
2518 * implement growing command buffers. Prior to Haswell, the kernel
2519 * command parser gets in the way and we have to fall back to growing
2522 device
->can_chain_batches
= device
->info
.gen
>= 8;
2524 device
->robust_buffer_access
= pCreateInfo
->pEnabledFeatures
&&
2525 pCreateInfo
->pEnabledFeatures
->robustBufferAccess
;
2526 device
->enabled_extensions
= enabled_extensions
;
2528 anv_device_init_dispatch(device
);
2530 if (pthread_mutex_init(&device
->mutex
, NULL
) != 0) {
2531 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2532 goto fail_context_id
;
2535 pthread_condattr_t condattr
;
2536 if (pthread_condattr_init(&condattr
) != 0) {
2537 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2540 if (pthread_condattr_setclock(&condattr
, CLOCK_MONOTONIC
) != 0) {
2541 pthread_condattr_destroy(&condattr
);
2542 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2545 if (pthread_cond_init(&device
->queue_submit
, &condattr
) != 0) {
2546 pthread_condattr_destroy(&condattr
);
2547 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2550 pthread_condattr_destroy(&condattr
);
2553 (physical_device
->supports_48bit_addresses
? EXEC_OBJECT_SUPPORTS_48B_ADDRESS
: 0) |
2554 (physical_device
->has_exec_async
? EXEC_OBJECT_ASYNC
: 0) |
2555 (physical_device
->has_exec_capture
? EXEC_OBJECT_CAPTURE
: 0) |
2556 (physical_device
->use_softpin
? EXEC_OBJECT_PINNED
: 0);
2558 anv_bo_pool_init(&device
->batch_bo_pool
, device
, bo_flags
);
2560 result
= anv_bo_cache_init(&device
->bo_cache
);
2561 if (result
!= VK_SUCCESS
)
2562 goto fail_batch_bo_pool
;
2564 if (!physical_device
->use_softpin
)
2565 bo_flags
&= ~EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
2567 result
= anv_state_pool_init(&device
->dynamic_state_pool
, device
,
2568 DYNAMIC_STATE_POOL_MIN_ADDRESS
,
2571 if (result
!= VK_SUCCESS
)
2574 result
= anv_state_pool_init(&device
->instruction_state_pool
, device
,
2575 INSTRUCTION_STATE_POOL_MIN_ADDRESS
,
2578 if (result
!= VK_SUCCESS
)
2579 goto fail_dynamic_state_pool
;
2581 result
= anv_state_pool_init(&device
->surface_state_pool
, device
,
2582 SURFACE_STATE_POOL_MIN_ADDRESS
,
2585 if (result
!= VK_SUCCESS
)
2586 goto fail_instruction_state_pool
;
2588 if (physical_device
->use_softpin
) {
2589 result
= anv_state_pool_init(&device
->binding_table_pool
, device
,
2590 BINDING_TABLE_POOL_MIN_ADDRESS
,
2593 if (result
!= VK_SUCCESS
)
2594 goto fail_surface_state_pool
;
2597 result
= anv_bo_init_new(&device
->workaround_bo
, device
, 4096);
2598 if (result
!= VK_SUCCESS
)
2599 goto fail_binding_table_pool
;
2601 if (physical_device
->use_softpin
)
2602 device
->workaround_bo
.flags
|= EXEC_OBJECT_PINNED
;
2604 if (!anv_vma_alloc(device
, &device
->workaround_bo
))
2605 goto fail_workaround_bo
;
2607 anv_device_init_trivial_batch(device
);
2609 if (device
->info
.gen
>= 10)
2610 anv_device_init_hiz_clear_value_bo(device
);
2612 anv_scratch_pool_init(device
, &device
->scratch_pool
);
2614 anv_queue_init(device
, &device
->queue
);
2616 switch (device
->info
.gen
) {
2618 if (!device
->info
.is_haswell
)
2619 result
= gen7_init_device_state(device
);
2621 result
= gen75_init_device_state(device
);
2624 result
= gen8_init_device_state(device
);
2627 result
= gen9_init_device_state(device
);
2630 result
= gen10_init_device_state(device
);
2633 result
= gen11_init_device_state(device
);
2636 result
= gen12_init_device_state(device
);
2639 /* Shouldn't get here as we don't create physical devices for any other
2641 unreachable("unhandled gen");
2643 if (result
!= VK_SUCCESS
)
2644 goto fail_workaround_bo
;
2646 anv_pipeline_cache_init(&device
->default_pipeline_cache
, device
, true);
2648 anv_device_init_blorp(device
);
2650 anv_device_init_border_colors(device
);
2652 *pDevice
= anv_device_to_handle(device
);
2657 anv_queue_finish(&device
->queue
);
2658 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
2659 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
2660 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
2661 fail_binding_table_pool
:
2662 if (physical_device
->use_softpin
)
2663 anv_state_pool_finish(&device
->binding_table_pool
);
2664 fail_surface_state_pool
:
2665 anv_state_pool_finish(&device
->surface_state_pool
);
2666 fail_instruction_state_pool
:
2667 anv_state_pool_finish(&device
->instruction_state_pool
);
2668 fail_dynamic_state_pool
:
2669 anv_state_pool_finish(&device
->dynamic_state_pool
);
2671 anv_bo_cache_finish(&device
->bo_cache
);
2673 anv_bo_pool_finish(&device
->batch_bo_pool
);
2674 pthread_cond_destroy(&device
->queue_submit
);
2676 pthread_mutex_destroy(&device
->mutex
);
2678 anv_gem_destroy_context(device
, device
->context_id
);
2682 vk_free(&device
->alloc
, device
);
2687 void anv_DestroyDevice(
2689 const VkAllocationCallbacks
* pAllocator
)
2691 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2692 struct anv_physical_device
*physical_device
;
2697 physical_device
= &device
->instance
->physicalDevice
;
2699 anv_device_finish_blorp(device
);
2701 anv_pipeline_cache_finish(&device
->default_pipeline_cache
);
2703 anv_queue_finish(&device
->queue
);
2705 #ifdef HAVE_VALGRIND
2706 /* We only need to free these to prevent valgrind errors. The backing
2707 * BO will go away in a couple of lines so we don't actually leak.
2709 anv_state_pool_free(&device
->dynamic_state_pool
, device
->border_colors
);
2710 anv_state_pool_free(&device
->dynamic_state_pool
, device
->slice_hash
);
2713 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
2715 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
2716 anv_vma_free(device
, &device
->workaround_bo
);
2717 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
2719 anv_vma_free(device
, &device
->trivial_batch_bo
);
2720 anv_gem_close(device
, device
->trivial_batch_bo
.gem_handle
);
2721 if (device
->info
.gen
>= 10)
2722 anv_gem_close(device
, device
->hiz_clear_bo
.gem_handle
);
2724 if (physical_device
->use_softpin
)
2725 anv_state_pool_finish(&device
->binding_table_pool
);
2726 anv_state_pool_finish(&device
->surface_state_pool
);
2727 anv_state_pool_finish(&device
->instruction_state_pool
);
2728 anv_state_pool_finish(&device
->dynamic_state_pool
);
2730 anv_bo_cache_finish(&device
->bo_cache
);
2732 anv_bo_pool_finish(&device
->batch_bo_pool
);
2734 pthread_cond_destroy(&device
->queue_submit
);
2735 pthread_mutex_destroy(&device
->mutex
);
2737 anv_gem_destroy_context(device
, device
->context_id
);
2739 if (INTEL_DEBUG
& DEBUG_BATCH
)
2740 gen_batch_decode_ctx_finish(&device
->decoder_ctx
);
2744 vk_free(&device
->alloc
, device
);
2747 VkResult
anv_EnumerateInstanceLayerProperties(
2748 uint32_t* pPropertyCount
,
2749 VkLayerProperties
* pProperties
)
2751 if (pProperties
== NULL
) {
2752 *pPropertyCount
= 0;
2756 /* None supported at this time */
2757 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
2760 VkResult
anv_EnumerateDeviceLayerProperties(
2761 VkPhysicalDevice physicalDevice
,
2762 uint32_t* pPropertyCount
,
2763 VkLayerProperties
* pProperties
)
2765 if (pProperties
== NULL
) {
2766 *pPropertyCount
= 0;
2770 /* None supported at this time */
2771 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
2774 void anv_GetDeviceQueue(
2776 uint32_t queueNodeIndex
,
2777 uint32_t queueIndex
,
2780 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2782 assert(queueIndex
== 0);
2784 *pQueue
= anv_queue_to_handle(&device
->queue
);
2787 void anv_GetDeviceQueue2(
2789 const VkDeviceQueueInfo2
* pQueueInfo
,
2792 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2794 assert(pQueueInfo
->queueIndex
== 0);
2796 if (pQueueInfo
->flags
== device
->queue
.flags
)
2797 *pQueue
= anv_queue_to_handle(&device
->queue
);
2803 _anv_device_set_lost(struct anv_device
*device
,
2804 const char *file
, int line
,
2805 const char *msg
, ...)
2810 device
->_lost
= true;
2813 err
= __vk_errorv(device
->instance
, device
,
2814 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT
,
2815 VK_ERROR_DEVICE_LOST
, file
, line
, msg
, ap
);
2818 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
2825 anv_device_query_status(struct anv_device
*device
)
2827 /* This isn't likely as most of the callers of this function already check
2828 * for it. However, it doesn't hurt to check and it potentially lets us
2831 if (anv_device_is_lost(device
))
2832 return VK_ERROR_DEVICE_LOST
;
2834 uint32_t active
, pending
;
2835 int ret
= anv_gem_gpu_get_reset_stats(device
, &active
, &pending
);
2837 /* We don't know the real error. */
2838 return anv_device_set_lost(device
, "get_reset_stats failed: %m");
2842 return anv_device_set_lost(device
, "GPU hung on one of our command buffers");
2843 } else if (pending
) {
2844 return anv_device_set_lost(device
, "GPU hung with commands in-flight");
2851 anv_device_bo_busy(struct anv_device
*device
, struct anv_bo
*bo
)
2853 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
2854 * Other usages of the BO (such as on different hardware) will not be
2855 * flagged as "busy" by this ioctl. Use with care.
2857 int ret
= anv_gem_busy(device
, bo
->gem_handle
);
2859 return VK_NOT_READY
;
2860 } else if (ret
== -1) {
2861 /* We don't know the real error. */
2862 return anv_device_set_lost(device
, "gem wait failed: %m");
2865 /* Query for device status after the busy call. If the BO we're checking
2866 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
2867 * client because it clearly doesn't have valid data. Yes, this most
2868 * likely means an ioctl, but we just did an ioctl to query the busy status
2869 * so it's no great loss.
2871 return anv_device_query_status(device
);
2875 anv_device_wait(struct anv_device
*device
, struct anv_bo
*bo
,
2878 int ret
= anv_gem_wait(device
, bo
->gem_handle
, &timeout
);
2879 if (ret
== -1 && errno
== ETIME
) {
2881 } else if (ret
== -1) {
2882 /* We don't know the real error. */
2883 return anv_device_set_lost(device
, "gem wait failed: %m");
2886 /* Query for device status after the wait. If the BO we're waiting on got
2887 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
2888 * because it clearly doesn't have valid data. Yes, this most likely means
2889 * an ioctl, but we just did an ioctl to wait so it's no great loss.
2891 return anv_device_query_status(device
);
2894 VkResult
anv_DeviceWaitIdle(
2897 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2898 if (anv_device_is_lost(device
))
2899 return VK_ERROR_DEVICE_LOST
;
2901 struct anv_batch batch
;
2904 batch
.start
= batch
.next
= cmds
;
2905 batch
.end
= (void *) cmds
+ sizeof(cmds
);
2907 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
2908 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
2910 return anv_device_submit_simple_batch(device
, &batch
);
2914 anv_vma_alloc(struct anv_device
*device
, struct anv_bo
*bo
)
2916 if (!(bo
->flags
& EXEC_OBJECT_PINNED
))
2919 pthread_mutex_lock(&device
->vma_mutex
);
2923 if (bo
->flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
&&
2924 device
->vma_hi_available
>= bo
->size
) {
2925 uint64_t addr
= util_vma_heap_alloc(&device
->vma_hi
, bo
->size
, 4096);
2927 bo
->offset
= gen_canonical_address(addr
);
2928 assert(addr
== gen_48b_address(bo
->offset
));
2929 device
->vma_hi_available
-= bo
->size
;
2933 if (bo
->offset
== 0 && device
->vma_lo_available
>= bo
->size
) {
2934 uint64_t addr
= util_vma_heap_alloc(&device
->vma_lo
, bo
->size
, 4096);
2936 bo
->offset
= gen_canonical_address(addr
);
2937 assert(addr
== gen_48b_address(bo
->offset
));
2938 device
->vma_lo_available
-= bo
->size
;
2942 pthread_mutex_unlock(&device
->vma_mutex
);
2944 return bo
->offset
!= 0;
2948 anv_vma_free(struct anv_device
*device
, struct anv_bo
*bo
)
2950 if (!(bo
->flags
& EXEC_OBJECT_PINNED
))
2953 const uint64_t addr_48b
= gen_48b_address(bo
->offset
);
2955 pthread_mutex_lock(&device
->vma_mutex
);
2957 if (addr_48b
>= LOW_HEAP_MIN_ADDRESS
&&
2958 addr_48b
<= LOW_HEAP_MAX_ADDRESS
) {
2959 util_vma_heap_free(&device
->vma_lo
, addr_48b
, bo
->size
);
2960 device
->vma_lo_available
+= bo
->size
;
2962 ASSERTED
const struct anv_physical_device
*physical_device
=
2963 &device
->instance
->physicalDevice
;
2964 assert(addr_48b
>= physical_device
->memory
.heaps
[0].vma_start
&&
2965 addr_48b
< (physical_device
->memory
.heaps
[0].vma_start
+
2966 physical_device
->memory
.heaps
[0].vma_size
));
2967 util_vma_heap_free(&device
->vma_hi
, addr_48b
, bo
->size
);
2968 device
->vma_hi_available
+= bo
->size
;
2971 pthread_mutex_unlock(&device
->vma_mutex
);
2977 anv_bo_init_new(struct anv_bo
*bo
, struct anv_device
*device
, uint64_t size
)
2979 uint32_t gem_handle
= anv_gem_create(device
, size
);
2981 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
2983 anv_bo_init(bo
, gem_handle
, size
);
2988 VkResult
anv_AllocateMemory(
2990 const VkMemoryAllocateInfo
* pAllocateInfo
,
2991 const VkAllocationCallbacks
* pAllocator
,
2992 VkDeviceMemory
* pMem
)
2994 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2995 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2996 struct anv_device_memory
*mem
;
2997 VkResult result
= VK_SUCCESS
;
2999 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
3001 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
3002 assert(pAllocateInfo
->allocationSize
> 0);
3004 if (pAllocateInfo
->allocationSize
> MAX_MEMORY_ALLOCATION_SIZE
)
3005 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
3007 /* FINISHME: Fail if allocation request exceeds heap size. */
3009 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
3010 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3012 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3014 assert(pAllocateInfo
->memoryTypeIndex
< pdevice
->memory
.type_count
);
3015 mem
->type
= &pdevice
->memory
.types
[pAllocateInfo
->memoryTypeIndex
];
3019 mem
->host_ptr
= NULL
;
3021 uint64_t bo_flags
= 0;
3023 assert(mem
->type
->heapIndex
< pdevice
->memory
.heap_count
);
3024 if (pdevice
->memory
.heaps
[mem
->type
->heapIndex
].supports_48bit_addresses
)
3025 bo_flags
|= EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
3027 const struct wsi_memory_allocate_info
*wsi_info
=
3028 vk_find_struct_const(pAllocateInfo
->pNext
, WSI_MEMORY_ALLOCATE_INFO_MESA
);
3029 if (wsi_info
&& wsi_info
->implicit_sync
) {
3030 /* We need to set the WRITE flag on window system buffers so that GEM
3031 * will know we're writing to them and synchronize uses on other rings
3032 * (eg if the display server uses the blitter ring).
3034 bo_flags
|= EXEC_OBJECT_WRITE
;
3035 } else if (pdevice
->has_exec_async
) {
3036 bo_flags
|= EXEC_OBJECT_ASYNC
;
3039 if (pdevice
->use_softpin
)
3040 bo_flags
|= EXEC_OBJECT_PINNED
;
3042 const VkExportMemoryAllocateInfo
*export_info
=
3043 vk_find_struct_const(pAllocateInfo
->pNext
, EXPORT_MEMORY_ALLOCATE_INFO
);
3045 /* Check if we need to support Android HW buffer export. If so,
3046 * create AHardwareBuffer and import memory from it.
3048 bool android_export
= false;
3049 if (export_info
&& export_info
->handleTypes
&
3050 VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID
)
3051 android_export
= true;
3053 /* Android memory import. */
3054 const struct VkImportAndroidHardwareBufferInfoANDROID
*ahw_import_info
=
3055 vk_find_struct_const(pAllocateInfo
->pNext
,
3056 IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID
);
3058 if (ahw_import_info
) {
3059 result
= anv_import_ahw_memory(_device
, mem
, ahw_import_info
);
3060 if (result
!= VK_SUCCESS
)
3064 } else if (android_export
) {
3065 result
= anv_create_ahw_memory(_device
, mem
, pAllocateInfo
);
3066 if (result
!= VK_SUCCESS
)
3069 const struct VkImportAndroidHardwareBufferInfoANDROID import_info
= {
3072 result
= anv_import_ahw_memory(_device
, mem
, &import_info
);
3073 if (result
!= VK_SUCCESS
)
3079 const VkImportMemoryFdInfoKHR
*fd_info
=
3080 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_FD_INFO_KHR
);
3082 /* The Vulkan spec permits handleType to be 0, in which case the struct is
3085 if (fd_info
&& fd_info
->handleType
) {
3086 /* At the moment, we support only the below handle types. */
3087 assert(fd_info
->handleType
==
3088 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
3089 fd_info
->handleType
==
3090 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
3092 result
= anv_bo_cache_import(device
, &device
->bo_cache
, fd_info
->fd
,
3093 bo_flags
| ANV_BO_EXTERNAL
, &mem
->bo
);
3094 if (result
!= VK_SUCCESS
)
3097 VkDeviceSize aligned_alloc_size
=
3098 align_u64(pAllocateInfo
->allocationSize
, 4096);
3100 /* For security purposes, we reject importing the bo if it's smaller
3101 * than the requested allocation size. This prevents a malicious client
3102 * from passing a buffer to a trusted client, lying about the size, and
3103 * telling the trusted client to try and texture from an image that goes
3104 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
3105 * in the trusted client. The trusted client can protect itself against
3106 * this sort of attack but only if it can trust the buffer size.
3108 if (mem
->bo
->size
< aligned_alloc_size
) {
3109 result
= vk_errorf(device
->instance
, device
,
3110 VK_ERROR_INVALID_EXTERNAL_HANDLE
,
3111 "aligned allocationSize too large for "
3112 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT: "
3113 "%"PRIu64
"B > %"PRIu64
"B",
3114 aligned_alloc_size
, mem
->bo
->size
);
3115 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
3119 /* From the Vulkan spec:
3121 * "Importing memory from a file descriptor transfers ownership of
3122 * the file descriptor from the application to the Vulkan
3123 * implementation. The application must not perform any operations on
3124 * the file descriptor after a successful import."
3126 * If the import fails, we leave the file descriptor open.
3132 const VkImportMemoryHostPointerInfoEXT
*host_ptr_info
=
3133 vk_find_struct_const(pAllocateInfo
->pNext
,
3134 IMPORT_MEMORY_HOST_POINTER_INFO_EXT
);
3135 if (host_ptr_info
&& host_ptr_info
->handleType
) {
3136 if (host_ptr_info
->handleType
==
3137 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_MAPPED_FOREIGN_MEMORY_BIT_EXT
) {
3138 result
= vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
3142 assert(host_ptr_info
->handleType
==
3143 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
);
3145 result
= anv_bo_cache_import_host_ptr(
3146 device
, &device
->bo_cache
, host_ptr_info
->pHostPointer
,
3147 pAllocateInfo
->allocationSize
, bo_flags
, &mem
->bo
);
3149 if (result
!= VK_SUCCESS
)
3152 mem
->host_ptr
= host_ptr_info
->pHostPointer
;
3156 /* Regular allocate (not importing memory). */
3158 if (export_info
&& export_info
->handleTypes
)
3159 bo_flags
|= ANV_BO_EXTERNAL
;
3161 result
= anv_bo_cache_alloc(device
, &device
->bo_cache
,
3162 pAllocateInfo
->allocationSize
, bo_flags
,
3164 if (result
!= VK_SUCCESS
)
3167 const VkMemoryDedicatedAllocateInfo
*dedicated_info
=
3168 vk_find_struct_const(pAllocateInfo
->pNext
, MEMORY_DEDICATED_ALLOCATE_INFO
);
3169 if (dedicated_info
&& dedicated_info
->image
!= VK_NULL_HANDLE
) {
3170 ANV_FROM_HANDLE(anv_image
, image
, dedicated_info
->image
);
3172 /* Some legacy (non-modifiers) consumers need the tiling to be set on
3173 * the BO. In this case, we have a dedicated allocation.
3175 if (image
->needs_set_tiling
) {
3176 const uint32_t i915_tiling
=
3177 isl_tiling_to_i915_tiling(image
->planes
[0].surface
.isl
.tiling
);
3178 int ret
= anv_gem_set_tiling(device
, mem
->bo
->gem_handle
,
3179 image
->planes
[0].surface
.isl
.row_pitch_B
,
3182 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
3183 return vk_errorf(device
->instance
, NULL
,
3184 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
3185 "failed to set BO tiling: %m");
3191 pthread_mutex_lock(&device
->mutex
);
3192 list_addtail(&mem
->link
, &device
->memory_objects
);
3193 pthread_mutex_unlock(&device
->mutex
);
3195 *pMem
= anv_device_memory_to_handle(mem
);
3197 p_atomic_add(&pdevice
->memory
.heaps
[mem
->type
->heapIndex
].used
,
3203 vk_free2(&device
->alloc
, pAllocator
, mem
);
3208 VkResult
anv_GetMemoryFdKHR(
3210 const VkMemoryGetFdInfoKHR
* pGetFdInfo
,
3213 ANV_FROM_HANDLE(anv_device
, dev
, device_h
);
3214 ANV_FROM_HANDLE(anv_device_memory
, mem
, pGetFdInfo
->memory
);
3216 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
3218 assert(pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
3219 pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
3221 return anv_bo_cache_export(dev
, &dev
->bo_cache
, mem
->bo
, pFd
);
3224 VkResult
anv_GetMemoryFdPropertiesKHR(
3226 VkExternalMemoryHandleTypeFlagBits handleType
,
3228 VkMemoryFdPropertiesKHR
* pMemoryFdProperties
)
3230 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3231 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3233 switch (handleType
) {
3234 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
:
3235 /* dma-buf can be imported as any memory type */
3236 pMemoryFdProperties
->memoryTypeBits
=
3237 (1 << pdevice
->memory
.type_count
) - 1;
3241 /* The valid usage section for this function says:
3243 * "handleType must not be one of the handle types defined as
3246 * So opaque handle types fall into the default "unsupported" case.
3248 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
3252 VkResult
anv_GetMemoryHostPointerPropertiesEXT(
3254 VkExternalMemoryHandleTypeFlagBits handleType
,
3255 const void* pHostPointer
,
3256 VkMemoryHostPointerPropertiesEXT
* pMemoryHostPointerProperties
)
3258 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3260 assert(pMemoryHostPointerProperties
->sType
==
3261 VK_STRUCTURE_TYPE_MEMORY_HOST_POINTER_PROPERTIES_EXT
);
3263 switch (handleType
) {
3264 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
: {
3265 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3267 /* Host memory can be imported as any memory type. */
3268 pMemoryHostPointerProperties
->memoryTypeBits
=
3269 (1ull << pdevice
->memory
.type_count
) - 1;
3274 return VK_ERROR_INVALID_EXTERNAL_HANDLE
;
3278 void anv_FreeMemory(
3280 VkDeviceMemory _mem
,
3281 const VkAllocationCallbacks
* pAllocator
)
3283 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3284 ANV_FROM_HANDLE(anv_device_memory
, mem
, _mem
);
3285 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3290 pthread_mutex_lock(&device
->mutex
);
3291 list_del(&mem
->link
);
3292 pthread_mutex_unlock(&device
->mutex
);
3295 anv_UnmapMemory(_device
, _mem
);
3297 p_atomic_add(&pdevice
->memory
.heaps
[mem
->type
->heapIndex
].used
,
3300 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
3302 #if defined(ANDROID) && ANDROID_API_LEVEL >= 26
3304 AHardwareBuffer_release(mem
->ahw
);
3307 vk_free2(&device
->alloc
, pAllocator
, mem
);
3310 VkResult
anv_MapMemory(
3312 VkDeviceMemory _memory
,
3313 VkDeviceSize offset
,
3315 VkMemoryMapFlags flags
,
3318 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3319 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
3326 if (mem
->host_ptr
) {
3327 *ppData
= mem
->host_ptr
+ offset
;
3331 if (size
== VK_WHOLE_SIZE
)
3332 size
= mem
->bo
->size
- offset
;
3334 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
3336 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
3337 * assert(size != 0);
3338 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
3339 * equal to the size of the memory minus offset
3342 assert(offset
+ size
<= mem
->bo
->size
);
3344 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
3345 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
3346 * at a time is valid. We could just mmap up front and return an offset
3347 * pointer here, but that may exhaust virtual memory on 32 bit
3350 uint32_t gem_flags
= 0;
3352 if (!device
->info
.has_llc
&&
3353 (mem
->type
->propertyFlags
& VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
))
3354 gem_flags
|= I915_MMAP_WC
;
3356 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
3357 uint64_t map_offset
= offset
& ~4095ull;
3358 assert(offset
>= map_offset
);
3359 uint64_t map_size
= (offset
+ size
) - map_offset
;
3361 /* Let's map whole pages */
3362 map_size
= align_u64(map_size
, 4096);
3364 void *map
= anv_gem_mmap(device
, mem
->bo
->gem_handle
,
3365 map_offset
, map_size
, gem_flags
);
3366 if (map
== MAP_FAILED
)
3367 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
3370 mem
->map_size
= map_size
;
3372 *ppData
= mem
->map
+ (offset
- map_offset
);
3377 void anv_UnmapMemory(
3379 VkDeviceMemory _memory
)
3381 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
3383 if (mem
== NULL
|| mem
->host_ptr
)
3386 anv_gem_munmap(mem
->map
, mem
->map_size
);
3393 clflush_mapped_ranges(struct anv_device
*device
,
3395 const VkMappedMemoryRange
*ranges
)
3397 for (uint32_t i
= 0; i
< count
; i
++) {
3398 ANV_FROM_HANDLE(anv_device_memory
, mem
, ranges
[i
].memory
);
3399 if (ranges
[i
].offset
>= mem
->map_size
)
3402 gen_clflush_range(mem
->map
+ ranges
[i
].offset
,
3403 MIN2(ranges
[i
].size
, mem
->map_size
- ranges
[i
].offset
));
3407 VkResult
anv_FlushMappedMemoryRanges(
3409 uint32_t memoryRangeCount
,
3410 const VkMappedMemoryRange
* pMemoryRanges
)
3412 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3414 if (device
->info
.has_llc
)
3417 /* Make sure the writes we're flushing have landed. */
3418 __builtin_ia32_mfence();
3420 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
3425 VkResult
anv_InvalidateMappedMemoryRanges(
3427 uint32_t memoryRangeCount
,
3428 const VkMappedMemoryRange
* pMemoryRanges
)
3430 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3432 if (device
->info
.has_llc
)
3435 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
3437 /* Make sure no reads get moved up above the invalidate. */
3438 __builtin_ia32_mfence();
3443 void anv_GetBufferMemoryRequirements(
3446 VkMemoryRequirements
* pMemoryRequirements
)
3448 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
3449 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3450 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3452 /* The Vulkan spec (git aaed022) says:
3454 * memoryTypeBits is a bitfield and contains one bit set for every
3455 * supported memory type for the resource. The bit `1<<i` is set if and
3456 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3457 * structure for the physical device is supported.
3459 uint32_t memory_types
= 0;
3460 for (uint32_t i
= 0; i
< pdevice
->memory
.type_count
; i
++) {
3461 uint32_t valid_usage
= pdevice
->memory
.types
[i
].valid_buffer_usage
;
3462 if ((valid_usage
& buffer
->usage
) == buffer
->usage
)
3463 memory_types
|= (1u << i
);
3466 /* Base alignment requirement of a cache line */
3467 uint32_t alignment
= 16;
3469 /* We need an alignment of 32 for pushing UBOs */
3470 if (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
)
3471 alignment
= MAX2(alignment
, 32);
3473 pMemoryRequirements
->size
= buffer
->size
;
3474 pMemoryRequirements
->alignment
= alignment
;
3476 /* Storage and Uniform buffers should have their size aligned to
3477 * 32-bits to avoid boundary checks when last DWord is not complete.
3478 * This would ensure that not internal padding would be needed for
3481 if (device
->robust_buffer_access
&&
3482 (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
||
3483 buffer
->usage
& VK_BUFFER_USAGE_STORAGE_BUFFER_BIT
))
3484 pMemoryRequirements
->size
= align_u64(buffer
->size
, 4);
3486 pMemoryRequirements
->memoryTypeBits
= memory_types
;
3489 void anv_GetBufferMemoryRequirements2(
3491 const VkBufferMemoryRequirementsInfo2
* pInfo
,
3492 VkMemoryRequirements2
* pMemoryRequirements
)
3494 anv_GetBufferMemoryRequirements(_device
, pInfo
->buffer
,
3495 &pMemoryRequirements
->memoryRequirements
);
3497 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3498 switch (ext
->sType
) {
3499 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
3500 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
3501 requirements
->prefersDedicatedAllocation
= false;
3502 requirements
->requiresDedicatedAllocation
= false;
3507 anv_debug_ignored_stype(ext
->sType
);
3513 void anv_GetImageMemoryRequirements(
3516 VkMemoryRequirements
* pMemoryRequirements
)
3518 ANV_FROM_HANDLE(anv_image
, image
, _image
);
3519 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3520 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3522 /* The Vulkan spec (git aaed022) says:
3524 * memoryTypeBits is a bitfield and contains one bit set for every
3525 * supported memory type for the resource. The bit `1<<i` is set if and
3526 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3527 * structure for the physical device is supported.
3529 * All types are currently supported for images.
3531 uint32_t memory_types
= (1ull << pdevice
->memory
.type_count
) - 1;
3533 /* We must have image allocated or imported at this point. According to the
3534 * specification, external images must have been bound to memory before
3535 * calling GetImageMemoryRequirements.
3537 assert(image
->size
> 0);
3539 pMemoryRequirements
->size
= image
->size
;
3540 pMemoryRequirements
->alignment
= image
->alignment
;
3541 pMemoryRequirements
->memoryTypeBits
= memory_types
;
3544 void anv_GetImageMemoryRequirements2(
3546 const VkImageMemoryRequirementsInfo2
* pInfo
,
3547 VkMemoryRequirements2
* pMemoryRequirements
)
3549 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3550 ANV_FROM_HANDLE(anv_image
, image
, pInfo
->image
);
3552 anv_GetImageMemoryRequirements(_device
, pInfo
->image
,
3553 &pMemoryRequirements
->memoryRequirements
);
3555 vk_foreach_struct_const(ext
, pInfo
->pNext
) {
3556 switch (ext
->sType
) {
3557 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO
: {
3558 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3559 const VkImagePlaneMemoryRequirementsInfo
*plane_reqs
=
3560 (const VkImagePlaneMemoryRequirementsInfo
*) ext
;
3561 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
,
3562 plane_reqs
->planeAspect
);
3564 assert(image
->planes
[plane
].offset
== 0);
3566 /* The Vulkan spec (git aaed022) says:
3568 * memoryTypeBits is a bitfield and contains one bit set for every
3569 * supported memory type for the resource. The bit `1<<i` is set
3570 * if and only if the memory type `i` in the
3571 * VkPhysicalDeviceMemoryProperties structure for the physical
3572 * device is supported.
3574 * All types are currently supported for images.
3576 pMemoryRequirements
->memoryRequirements
.memoryTypeBits
=
3577 (1ull << pdevice
->memory
.type_count
) - 1;
3579 /* We must have image allocated or imported at this point. According to the
3580 * specification, external images must have been bound to memory before
3581 * calling GetImageMemoryRequirements.
3583 assert(image
->planes
[plane
].size
> 0);
3585 pMemoryRequirements
->memoryRequirements
.size
= image
->planes
[plane
].size
;
3586 pMemoryRequirements
->memoryRequirements
.alignment
=
3587 image
->planes
[plane
].alignment
;
3592 anv_debug_ignored_stype(ext
->sType
);
3597 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3598 switch (ext
->sType
) {
3599 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
3600 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
3601 if (image
->needs_set_tiling
|| image
->external_format
) {
3602 /* If we need to set the tiling for external consumers, we need a
3603 * dedicated allocation.
3605 * See also anv_AllocateMemory.
3607 requirements
->prefersDedicatedAllocation
= true;
3608 requirements
->requiresDedicatedAllocation
= true;
3610 requirements
->prefersDedicatedAllocation
= false;
3611 requirements
->requiresDedicatedAllocation
= false;
3617 anv_debug_ignored_stype(ext
->sType
);
3623 void anv_GetImageSparseMemoryRequirements(
3626 uint32_t* pSparseMemoryRequirementCount
,
3627 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
3629 *pSparseMemoryRequirementCount
= 0;
3632 void anv_GetImageSparseMemoryRequirements2(
3634 const VkImageSparseMemoryRequirementsInfo2
* pInfo
,
3635 uint32_t* pSparseMemoryRequirementCount
,
3636 VkSparseImageMemoryRequirements2
* pSparseMemoryRequirements
)
3638 *pSparseMemoryRequirementCount
= 0;
3641 void anv_GetDeviceMemoryCommitment(
3643 VkDeviceMemory memory
,
3644 VkDeviceSize
* pCommittedMemoryInBytes
)
3646 *pCommittedMemoryInBytes
= 0;
3650 anv_bind_buffer_memory(const VkBindBufferMemoryInfo
*pBindInfo
)
3652 ANV_FROM_HANDLE(anv_device_memory
, mem
, pBindInfo
->memory
);
3653 ANV_FROM_HANDLE(anv_buffer
, buffer
, pBindInfo
->buffer
);
3655 assert(pBindInfo
->sType
== VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
);
3658 assert((buffer
->usage
& mem
->type
->valid_buffer_usage
) == buffer
->usage
);
3659 buffer
->address
= (struct anv_address
) {
3661 .offset
= pBindInfo
->memoryOffset
,
3664 buffer
->address
= ANV_NULL_ADDRESS
;
3668 VkResult
anv_BindBufferMemory(
3671 VkDeviceMemory memory
,
3672 VkDeviceSize memoryOffset
)
3674 anv_bind_buffer_memory(
3675 &(VkBindBufferMemoryInfo
) {
3676 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
,
3679 .memoryOffset
= memoryOffset
,
3685 VkResult
anv_BindBufferMemory2(
3687 uint32_t bindInfoCount
,
3688 const VkBindBufferMemoryInfo
* pBindInfos
)
3690 for (uint32_t i
= 0; i
< bindInfoCount
; i
++)
3691 anv_bind_buffer_memory(&pBindInfos
[i
]);
3696 VkResult
anv_QueueBindSparse(
3698 uint32_t bindInfoCount
,
3699 const VkBindSparseInfo
* pBindInfo
,
3702 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
3703 if (anv_device_is_lost(queue
->device
))
3704 return VK_ERROR_DEVICE_LOST
;
3706 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
3711 VkResult
anv_CreateEvent(
3713 const VkEventCreateInfo
* pCreateInfo
,
3714 const VkAllocationCallbacks
* pAllocator
,
3717 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3718 struct anv_state state
;
3719 struct anv_event
*event
;
3721 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_EVENT_CREATE_INFO
);
3723 state
= anv_state_pool_alloc(&device
->dynamic_state_pool
,
3726 event
->state
= state
;
3727 event
->semaphore
= VK_EVENT_RESET
;
3729 if (!device
->info
.has_llc
) {
3730 /* Make sure the writes we're flushing have landed. */
3731 __builtin_ia32_mfence();
3732 __builtin_ia32_clflush(event
);
3735 *pEvent
= anv_event_to_handle(event
);
3740 void anv_DestroyEvent(
3743 const VkAllocationCallbacks
* pAllocator
)
3745 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3746 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3751 anv_state_pool_free(&device
->dynamic_state_pool
, event
->state
);
3754 VkResult
anv_GetEventStatus(
3758 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3759 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3761 if (anv_device_is_lost(device
))
3762 return VK_ERROR_DEVICE_LOST
;
3764 if (!device
->info
.has_llc
) {
3765 /* Invalidate read cache before reading event written by GPU. */
3766 __builtin_ia32_clflush(event
);
3767 __builtin_ia32_mfence();
3771 return event
->semaphore
;
3774 VkResult
anv_SetEvent(
3778 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3779 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3781 event
->semaphore
= VK_EVENT_SET
;
3783 if (!device
->info
.has_llc
) {
3784 /* Make sure the writes we're flushing have landed. */
3785 __builtin_ia32_mfence();
3786 __builtin_ia32_clflush(event
);
3792 VkResult
anv_ResetEvent(
3796 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3797 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3799 event
->semaphore
= VK_EVENT_RESET
;
3801 if (!device
->info
.has_llc
) {
3802 /* Make sure the writes we're flushing have landed. */
3803 __builtin_ia32_mfence();
3804 __builtin_ia32_clflush(event
);
3812 VkResult
anv_CreateBuffer(
3814 const VkBufferCreateInfo
* pCreateInfo
,
3815 const VkAllocationCallbacks
* pAllocator
,
3818 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3819 struct anv_buffer
*buffer
;
3821 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
3823 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
3824 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3826 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3828 buffer
->size
= pCreateInfo
->size
;
3829 buffer
->usage
= pCreateInfo
->usage
;
3830 buffer
->address
= ANV_NULL_ADDRESS
;
3832 *pBuffer
= anv_buffer_to_handle(buffer
);
3837 void anv_DestroyBuffer(
3840 const VkAllocationCallbacks
* pAllocator
)
3842 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3843 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
3848 vk_free2(&device
->alloc
, pAllocator
, buffer
);
3851 VkDeviceAddress
anv_GetBufferDeviceAddressEXT(
3853 const VkBufferDeviceAddressInfoEXT
* pInfo
)
3855 ANV_FROM_HANDLE(anv_buffer
, buffer
, pInfo
->buffer
);
3857 assert(buffer
->address
.bo
->flags
& EXEC_OBJECT_PINNED
);
3859 return anv_address_physical(buffer
->address
);
3863 anv_fill_buffer_surface_state(struct anv_device
*device
, struct anv_state state
,
3864 enum isl_format format
,
3865 struct anv_address address
,
3866 uint32_t range
, uint32_t stride
)
3868 isl_buffer_fill_state(&device
->isl_dev
, state
.map
,
3869 .address
= anv_address_physical(address
),
3870 .mocs
= device
->default_mocs
,
3873 .swizzle
= ISL_SWIZZLE_IDENTITY
,
3874 .stride_B
= stride
);
3877 void anv_DestroySampler(
3880 const VkAllocationCallbacks
* pAllocator
)
3882 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3883 ANV_FROM_HANDLE(anv_sampler
, sampler
, _sampler
);
3888 if (sampler
->bindless_state
.map
) {
3889 anv_state_pool_free(&device
->dynamic_state_pool
,
3890 sampler
->bindless_state
);
3893 vk_free2(&device
->alloc
, pAllocator
, sampler
);
3896 VkResult
anv_CreateFramebuffer(
3898 const VkFramebufferCreateInfo
* pCreateInfo
,
3899 const VkAllocationCallbacks
* pAllocator
,
3900 VkFramebuffer
* pFramebuffer
)
3902 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3903 struct anv_framebuffer
*framebuffer
;
3905 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
3907 size_t size
= sizeof(*framebuffer
);
3909 /* VK_KHR_imageless_framebuffer extension says:
3911 * If flags includes VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR,
3912 * parameter pAttachments is ignored.
3914 if (!(pCreateInfo
->flags
& VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR
)) {
3915 size
+= sizeof(struct anv_image_view
*) * pCreateInfo
->attachmentCount
;
3916 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
3917 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3918 if (framebuffer
== NULL
)
3919 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3921 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
3922 ANV_FROM_HANDLE(anv_image_view
, iview
, pCreateInfo
->pAttachments
[i
]);
3923 framebuffer
->attachments
[i
] = iview
;
3925 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
3927 assert(device
->enabled_extensions
.KHR_imageless_framebuffer
);
3928 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
3929 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3930 if (framebuffer
== NULL
)
3931 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3933 framebuffer
->attachment_count
= 0;
3936 framebuffer
->width
= pCreateInfo
->width
;
3937 framebuffer
->height
= pCreateInfo
->height
;
3938 framebuffer
->layers
= pCreateInfo
->layers
;
3940 *pFramebuffer
= anv_framebuffer_to_handle(framebuffer
);
3945 void anv_DestroyFramebuffer(
3948 const VkAllocationCallbacks
* pAllocator
)
3950 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3951 ANV_FROM_HANDLE(anv_framebuffer
, fb
, _fb
);
3956 vk_free2(&device
->alloc
, pAllocator
, fb
);
3959 static const VkTimeDomainEXT anv_time_domains
[] = {
3960 VK_TIME_DOMAIN_DEVICE_EXT
,
3961 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
,
3962 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
,
3965 VkResult
anv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
3966 VkPhysicalDevice physicalDevice
,
3967 uint32_t *pTimeDomainCount
,
3968 VkTimeDomainEXT
*pTimeDomains
)
3971 VK_OUTARRAY_MAKE(out
, pTimeDomains
, pTimeDomainCount
);
3973 for (d
= 0; d
< ARRAY_SIZE(anv_time_domains
); d
++) {
3974 vk_outarray_append(&out
, i
) {
3975 *i
= anv_time_domains
[d
];
3979 return vk_outarray_status(&out
);
3983 anv_clock_gettime(clockid_t clock_id
)
3985 struct timespec current
;
3988 ret
= clock_gettime(clock_id
, ¤t
);
3989 if (ret
< 0 && clock_id
== CLOCK_MONOTONIC_RAW
)
3990 ret
= clock_gettime(CLOCK_MONOTONIC
, ¤t
);
3994 return (uint64_t) current
.tv_sec
* 1000000000ULL + current
.tv_nsec
;
3997 #define TIMESTAMP 0x2358
3999 VkResult
anv_GetCalibratedTimestampsEXT(
4001 uint32_t timestampCount
,
4002 const VkCalibratedTimestampInfoEXT
*pTimestampInfos
,
4003 uint64_t *pTimestamps
,
4004 uint64_t *pMaxDeviation
)
4006 ANV_FROM_HANDLE(anv_device
, device
, _device
);
4007 uint64_t timestamp_frequency
= device
->info
.timestamp_frequency
;
4010 uint64_t begin
, end
;
4011 uint64_t max_clock_period
= 0;
4013 begin
= anv_clock_gettime(CLOCK_MONOTONIC_RAW
);
4015 for (d
= 0; d
< timestampCount
; d
++) {
4016 switch (pTimestampInfos
[d
].timeDomain
) {
4017 case VK_TIME_DOMAIN_DEVICE_EXT
:
4018 ret
= anv_gem_reg_read(device
, TIMESTAMP
| 1,
4022 return anv_device_set_lost(device
, "Failed to read the TIMESTAMP "
4025 uint64_t device_period
= DIV_ROUND_UP(1000000000, timestamp_frequency
);
4026 max_clock_period
= MAX2(max_clock_period
, device_period
);
4028 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
:
4029 pTimestamps
[d
] = anv_clock_gettime(CLOCK_MONOTONIC
);
4030 max_clock_period
= MAX2(max_clock_period
, 1);
4033 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
:
4034 pTimestamps
[d
] = begin
;
4042 end
= anv_clock_gettime(CLOCK_MONOTONIC_RAW
);
4045 * The maximum deviation is the sum of the interval over which we
4046 * perform the sampling and the maximum period of any sampled
4047 * clock. That's because the maximum skew between any two sampled
4048 * clock edges is when the sampled clock with the largest period is
4049 * sampled at the end of that period but right at the beginning of the
4050 * sampling interval and some other clock is sampled right at the
4051 * begining of its sampling period and right at the end of the
4052 * sampling interval. Let's assume the GPU has the longest clock
4053 * period and that the application is sampling GPU and monotonic:
4056 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
4057 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4061 * GPU -----_____-----_____-----_____-----_____
4064 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
4065 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4067 * Interval <----------------->
4068 * Deviation <-------------------------->
4072 * m = read(monotonic) 2
4075 * We round the sample interval up by one tick to cover sampling error
4076 * in the interval clock
4079 uint64_t sample_interval
= end
- begin
+ 1;
4081 *pMaxDeviation
= sample_interval
+ max_clock_period
;
4086 /* vk_icd.h does not declare this function, so we declare it here to
4087 * suppress Wmissing-prototypes.
4089 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4090 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
);
4092 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4093 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
)
4095 /* For the full details on loader interface versioning, see
4096 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
4097 * What follows is a condensed summary, to help you navigate the large and
4098 * confusing official doc.
4100 * - Loader interface v0 is incompatible with later versions. We don't
4103 * - In loader interface v1:
4104 * - The first ICD entrypoint called by the loader is
4105 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
4107 * - The ICD must statically expose no other Vulkan symbol unless it is
4108 * linked with -Bsymbolic.
4109 * - Each dispatchable Vulkan handle created by the ICD must be
4110 * a pointer to a struct whose first member is VK_LOADER_DATA. The
4111 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
4112 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
4113 * vkDestroySurfaceKHR(). The ICD must be capable of working with
4114 * such loader-managed surfaces.
4116 * - Loader interface v2 differs from v1 in:
4117 * - The first ICD entrypoint called by the loader is
4118 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
4119 * statically expose this entrypoint.
4121 * - Loader interface v3 differs from v2 in:
4122 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
4123 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
4124 * because the loader no longer does so.
4126 * - Loader interface v4 differs from v3 in:
4127 * - The ICD must implement vk_icdGetPhysicalDeviceProcAddr().
4129 *pSupportedVersion
= MIN2(*pSupportedVersion
, 4u);