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 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->device_dispatch
.entrypoints
); i
++) {
754 /* Vulkan requires that entrypoints for extensions which have not been
755 * enabled must not be advertised.
757 if (!anv_device_entrypoint_is_enabled(i
, instance
->app_info
.api_version
,
758 &instance
->enabled_extensions
, NULL
)) {
759 instance
->device_dispatch
.entrypoints
[i
] = NULL
;
761 instance
->device_dispatch
.entrypoints
[i
] =
762 anv_device_dispatch_table
.entrypoints
[i
];
766 instance
->physicalDeviceCount
= -1;
768 result
= vk_debug_report_instance_init(&instance
->debug_report_callbacks
);
769 if (result
!= VK_SUCCESS
) {
770 vk_free2(&default_alloc
, pAllocator
, instance
);
771 return vk_error(result
);
774 instance
->pipeline_cache_enabled
=
775 env_var_as_boolean("ANV_ENABLE_PIPELINE_CACHE", true);
778 glsl_type_singleton_init_or_ref();
780 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
782 driParseOptionInfo(&instance
->available_dri_options
, anv_dri_options_xml
);
783 driParseConfigFiles(&instance
->dri_options
, &instance
->available_dri_options
,
785 instance
->app_info
.engine_name
,
786 instance
->app_info
.engine_version
);
788 *pInstance
= anv_instance_to_handle(instance
);
793 void anv_DestroyInstance(
794 VkInstance _instance
,
795 const VkAllocationCallbacks
* pAllocator
)
797 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
802 if (instance
->physicalDeviceCount
> 0) {
803 /* We support at most one physical device. */
804 assert(instance
->physicalDeviceCount
== 1);
805 anv_physical_device_finish(&instance
->physicalDevice
);
808 vk_free(&instance
->alloc
, (char *)instance
->app_info
.app_name
);
809 vk_free(&instance
->alloc
, (char *)instance
->app_info
.engine_name
);
811 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
813 vk_debug_report_instance_destroy(&instance
->debug_report_callbacks
);
815 glsl_type_singleton_decref();
818 driDestroyOptionCache(&instance
->dri_options
);
819 driDestroyOptionInfo(&instance
->available_dri_options
);
821 vk_free(&instance
->alloc
, instance
);
825 anv_enumerate_devices(struct anv_instance
*instance
)
827 /* TODO: Check for more devices ? */
828 drmDevicePtr devices
[8];
829 VkResult result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
832 instance
->physicalDeviceCount
= 0;
834 max_devices
= drmGetDevices2(0, devices
, ARRAY_SIZE(devices
));
836 return VK_ERROR_INCOMPATIBLE_DRIVER
;
838 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
839 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
840 devices
[i
]->bustype
== DRM_BUS_PCI
&&
841 devices
[i
]->deviceinfo
.pci
->vendor_id
== 0x8086) {
843 result
= anv_physical_device_init(&instance
->physicalDevice
,
844 instance
, devices
[i
]);
845 if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
849 drmFreeDevices(devices
, max_devices
);
851 if (result
== VK_SUCCESS
)
852 instance
->physicalDeviceCount
= 1;
858 anv_instance_ensure_physical_device(struct anv_instance
*instance
)
860 if (instance
->physicalDeviceCount
< 0) {
861 VkResult result
= anv_enumerate_devices(instance
);
862 if (result
!= VK_SUCCESS
&&
863 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
870 VkResult
anv_EnumeratePhysicalDevices(
871 VkInstance _instance
,
872 uint32_t* pPhysicalDeviceCount
,
873 VkPhysicalDevice
* pPhysicalDevices
)
875 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
876 VK_OUTARRAY_MAKE(out
, pPhysicalDevices
, pPhysicalDeviceCount
);
878 VkResult result
= anv_instance_ensure_physical_device(instance
);
879 if (result
!= VK_SUCCESS
)
882 if (instance
->physicalDeviceCount
== 0)
885 assert(instance
->physicalDeviceCount
== 1);
886 vk_outarray_append(&out
, i
) {
887 *i
= anv_physical_device_to_handle(&instance
->physicalDevice
);
890 return vk_outarray_status(&out
);
893 VkResult
anv_EnumeratePhysicalDeviceGroups(
894 VkInstance _instance
,
895 uint32_t* pPhysicalDeviceGroupCount
,
896 VkPhysicalDeviceGroupProperties
* pPhysicalDeviceGroupProperties
)
898 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
899 VK_OUTARRAY_MAKE(out
, pPhysicalDeviceGroupProperties
,
900 pPhysicalDeviceGroupCount
);
902 VkResult result
= anv_instance_ensure_physical_device(instance
);
903 if (result
!= VK_SUCCESS
)
906 if (instance
->physicalDeviceCount
== 0)
909 assert(instance
->physicalDeviceCount
== 1);
911 vk_outarray_append(&out
, p
) {
912 p
->physicalDeviceCount
= 1;
913 memset(p
->physicalDevices
, 0, sizeof(p
->physicalDevices
));
914 p
->physicalDevices
[0] =
915 anv_physical_device_to_handle(&instance
->physicalDevice
);
916 p
->subsetAllocation
= false;
918 vk_foreach_struct(ext
, p
->pNext
)
919 anv_debug_ignored_stype(ext
->sType
);
922 return vk_outarray_status(&out
);
925 void anv_GetPhysicalDeviceFeatures(
926 VkPhysicalDevice physicalDevice
,
927 VkPhysicalDeviceFeatures
* pFeatures
)
929 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
931 *pFeatures
= (VkPhysicalDeviceFeatures
) {
932 .robustBufferAccess
= true,
933 .fullDrawIndexUint32
= true,
934 .imageCubeArray
= true,
935 .independentBlend
= true,
936 .geometryShader
= true,
937 .tessellationShader
= true,
938 .sampleRateShading
= true,
939 .dualSrcBlend
= true,
941 .multiDrawIndirect
= true,
942 .drawIndirectFirstInstance
= true,
944 .depthBiasClamp
= true,
945 .fillModeNonSolid
= true,
946 .depthBounds
= false,
950 .multiViewport
= true,
951 .samplerAnisotropy
= true,
952 .textureCompressionETC2
= pdevice
->info
.gen
>= 8 ||
953 pdevice
->info
.is_baytrail
,
954 .textureCompressionASTC_LDR
= pdevice
->info
.gen
>= 9, /* FINISHME CHV */
955 .textureCompressionBC
= true,
956 .occlusionQueryPrecise
= true,
957 .pipelineStatisticsQuery
= true,
958 .fragmentStoresAndAtomics
= true,
959 .shaderTessellationAndGeometryPointSize
= true,
960 .shaderImageGatherExtended
= true,
961 .shaderStorageImageExtendedFormats
= true,
962 .shaderStorageImageMultisample
= false,
963 .shaderStorageImageReadWithoutFormat
= false,
964 .shaderStorageImageWriteWithoutFormat
= true,
965 .shaderUniformBufferArrayDynamicIndexing
= true,
966 .shaderSampledImageArrayDynamicIndexing
= true,
967 .shaderStorageBufferArrayDynamicIndexing
= true,
968 .shaderStorageImageArrayDynamicIndexing
= true,
969 .shaderClipDistance
= true,
970 .shaderCullDistance
= true,
971 .shaderFloat64
= pdevice
->info
.gen
>= 8 &&
972 pdevice
->info
.has_64bit_types
,
973 .shaderInt64
= pdevice
->info
.gen
>= 8 &&
974 pdevice
->info
.has_64bit_types
,
975 .shaderInt16
= pdevice
->info
.gen
>= 8,
976 .shaderResourceMinLod
= pdevice
->info
.gen
>= 9,
977 .variableMultisampleRate
= true,
978 .inheritedQueries
= true,
981 /* We can't do image stores in vec4 shaders */
982 pFeatures
->vertexPipelineStoresAndAtomics
=
983 pdevice
->compiler
->scalar_stage
[MESA_SHADER_VERTEX
] &&
984 pdevice
->compiler
->scalar_stage
[MESA_SHADER_GEOMETRY
];
986 struct anv_app_info
*app_info
= &pdevice
->instance
->app_info
;
988 /* The new DOOM and Wolfenstein games require depthBounds without
989 * checking for it. They seem to run fine without it so just claim it's
990 * there and accept the consequences.
992 if (app_info
->engine_name
&& strcmp(app_info
->engine_name
, "idTech") == 0)
993 pFeatures
->depthBounds
= true;
996 void anv_GetPhysicalDeviceFeatures2(
997 VkPhysicalDevice physicalDevice
,
998 VkPhysicalDeviceFeatures2
* pFeatures
)
1000 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1001 anv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
1003 vk_foreach_struct(ext
, pFeatures
->pNext
) {
1004 switch (ext
->sType
) {
1005 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES_KHR
: {
1006 VkPhysicalDevice8BitStorageFeaturesKHR
*features
=
1007 (VkPhysicalDevice8BitStorageFeaturesKHR
*)ext
;
1008 features
->storageBuffer8BitAccess
= pdevice
->info
.gen
>= 8;
1009 features
->uniformAndStorageBuffer8BitAccess
= pdevice
->info
.gen
>= 8;
1010 features
->storagePushConstant8
= pdevice
->info
.gen
>= 8;
1014 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES
: {
1015 VkPhysicalDevice16BitStorageFeatures
*features
=
1016 (VkPhysicalDevice16BitStorageFeatures
*)ext
;
1017 features
->storageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
1018 features
->uniformAndStorageBuffer16BitAccess
= pdevice
->info
.gen
>= 8;
1019 features
->storagePushConstant16
= pdevice
->info
.gen
>= 8;
1020 features
->storageInputOutput16
= false;
1024 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_EXT
: {
1025 VkPhysicalDeviceBufferDeviceAddressFeaturesEXT
*features
= (void *)ext
;
1026 features
->bufferDeviceAddress
= pdevice
->has_a64_buffer_access
;
1027 features
->bufferDeviceAddressCaptureReplay
= false;
1028 features
->bufferDeviceAddressMultiDevice
= false;
1032 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COMPUTE_SHADER_DERIVATIVES_FEATURES_NV
: {
1033 VkPhysicalDeviceComputeShaderDerivativesFeaturesNV
*features
=
1034 (VkPhysicalDeviceComputeShaderDerivativesFeaturesNV
*)ext
;
1035 features
->computeDerivativeGroupQuads
= true;
1036 features
->computeDerivativeGroupLinear
= true;
1040 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT
: {
1041 VkPhysicalDeviceConditionalRenderingFeaturesEXT
*features
=
1042 (VkPhysicalDeviceConditionalRenderingFeaturesEXT
*)ext
;
1043 features
->conditionalRendering
= pdevice
->info
.gen
>= 8 ||
1044 pdevice
->info
.is_haswell
;
1045 features
->inheritedConditionalRendering
= pdevice
->info
.gen
>= 8 ||
1046 pdevice
->info
.is_haswell
;
1050 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_CLIP_ENABLE_FEATURES_EXT
: {
1051 VkPhysicalDeviceDepthClipEnableFeaturesEXT
*features
=
1052 (VkPhysicalDeviceDepthClipEnableFeaturesEXT
*)ext
;
1053 features
->depthClipEnable
= true;
1057 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR
: {
1058 VkPhysicalDeviceFloat16Int8FeaturesKHR
*features
= (void *)ext
;
1059 features
->shaderFloat16
= pdevice
->info
.gen
>= 8;
1060 features
->shaderInt8
= pdevice
->info
.gen
>= 8;
1064 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FRAGMENT_SHADER_INTERLOCK_FEATURES_EXT
: {
1065 VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT
*features
=
1066 (VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT
*)ext
;
1067 features
->fragmentShaderSampleInterlock
= pdevice
->info
.gen
>= 9;
1068 features
->fragmentShaderPixelInterlock
= pdevice
->info
.gen
>= 9;
1069 features
->fragmentShaderShadingRateInterlock
= false;
1073 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_QUERY_RESET_FEATURES_EXT
: {
1074 VkPhysicalDeviceHostQueryResetFeaturesEXT
*features
=
1075 (VkPhysicalDeviceHostQueryResetFeaturesEXT
*)ext
;
1076 features
->hostQueryReset
= true;
1080 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT
: {
1081 VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*features
=
1082 (VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*)ext
;
1083 features
->shaderInputAttachmentArrayDynamicIndexing
= false;
1084 features
->shaderUniformTexelBufferArrayDynamicIndexing
= true;
1085 features
->shaderStorageTexelBufferArrayDynamicIndexing
= true;
1086 features
->shaderUniformBufferArrayNonUniformIndexing
= false;
1087 features
->shaderSampledImageArrayNonUniformIndexing
= true;
1088 features
->shaderStorageBufferArrayNonUniformIndexing
= true;
1089 features
->shaderStorageImageArrayNonUniformIndexing
= true;
1090 features
->shaderInputAttachmentArrayNonUniformIndexing
= false;
1091 features
->shaderUniformTexelBufferArrayNonUniformIndexing
= true;
1092 features
->shaderStorageTexelBufferArrayNonUniformIndexing
= true;
1093 features
->descriptorBindingUniformBufferUpdateAfterBind
= false;
1094 features
->descriptorBindingSampledImageUpdateAfterBind
= true;
1095 features
->descriptorBindingStorageImageUpdateAfterBind
= true;
1096 features
->descriptorBindingStorageBufferUpdateAfterBind
= true;
1097 features
->descriptorBindingUniformTexelBufferUpdateAfterBind
= true;
1098 features
->descriptorBindingStorageTexelBufferUpdateAfterBind
= true;
1099 features
->descriptorBindingUpdateUnusedWhilePending
= true;
1100 features
->descriptorBindingPartiallyBound
= true;
1101 features
->descriptorBindingVariableDescriptorCount
= false;
1102 features
->runtimeDescriptorArray
= true;
1106 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INDEX_TYPE_UINT8_FEATURES_EXT
: {
1107 VkPhysicalDeviceIndexTypeUint8FeaturesEXT
*features
=
1108 (VkPhysicalDeviceIndexTypeUint8FeaturesEXT
*)ext
;
1109 features
->indexTypeUint8
= true;
1113 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_FEATURES_EXT
: {
1114 VkPhysicalDeviceInlineUniformBlockFeaturesEXT
*features
=
1115 (VkPhysicalDeviceInlineUniformBlockFeaturesEXT
*)ext
;
1116 features
->inlineUniformBlock
= true;
1117 features
->descriptorBindingInlineUniformBlockUpdateAfterBind
= true;
1121 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_FEATURES_EXT
: {
1122 VkPhysicalDeviceLineRasterizationFeaturesEXT
*features
=
1123 (VkPhysicalDeviceLineRasterizationFeaturesEXT
*)ext
;
1124 features
->rectangularLines
= true;
1125 features
->bresenhamLines
= true;
1126 features
->smoothLines
= true;
1127 features
->stippledRectangularLines
= false;
1128 features
->stippledBresenhamLines
= true;
1129 features
->stippledSmoothLines
= false;
1133 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES
: {
1134 VkPhysicalDeviceMultiviewFeatures
*features
=
1135 (VkPhysicalDeviceMultiviewFeatures
*)ext
;
1136 features
->multiview
= true;
1137 features
->multiviewGeometryShader
= true;
1138 features
->multiviewTessellationShader
= true;
1142 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGELESS_FRAMEBUFFER_FEATURES_KHR
: {
1143 VkPhysicalDeviceImagelessFramebufferFeaturesKHR
*features
=
1144 (VkPhysicalDeviceImagelessFramebufferFeaturesKHR
*)ext
;
1145 features
->imagelessFramebuffer
= true;
1149 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR
: {
1150 VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR
*features
=
1151 (VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR
*)ext
;
1152 features
->pipelineExecutableInfo
= true;
1156 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES
: {
1157 VkPhysicalDeviceProtectedMemoryFeatures
*features
= (void *)ext
;
1158 features
->protectedMemory
= false;
1162 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES
: {
1163 VkPhysicalDeviceSamplerYcbcrConversionFeatures
*features
=
1164 (VkPhysicalDeviceSamplerYcbcrConversionFeatures
*) ext
;
1165 features
->samplerYcbcrConversion
= true;
1169 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES_EXT
: {
1170 VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*features
=
1171 (VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*)ext
;
1172 features
->scalarBlockLayout
= true;
1176 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_INT64_FEATURES_KHR
: {
1177 VkPhysicalDeviceShaderAtomicInt64FeaturesKHR
*features
= (void *)ext
;
1178 features
->shaderBufferInt64Atomics
=
1179 pdevice
->info
.gen
>= 9 && pdevice
->use_softpin
;
1180 features
->shaderSharedInt64Atomics
= VK_FALSE
;
1184 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DEMOTE_TO_HELPER_INVOCATION_FEATURES_EXT
: {
1185 VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT
*features
= (void *)ext
;
1186 features
->shaderDemoteToHelperInvocation
= true;
1190 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETERS_FEATURES
: {
1191 VkPhysicalDeviceShaderDrawParametersFeatures
*features
= (void *)ext
;
1192 features
->shaderDrawParameters
= true;
1196 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_FEATURES_EXT
: {
1197 VkPhysicalDeviceSubgroupSizeControlFeaturesEXT
*features
=
1198 (VkPhysicalDeviceSubgroupSizeControlFeaturesEXT
*)ext
;
1199 features
->subgroupSizeControl
= true;
1200 features
->computeFullSubgroups
= true;
1204 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_FEATURES_EXT
: {
1205 VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT
*features
=
1206 (VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT
*)ext
;
1207 features
->texelBufferAlignment
= true;
1211 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES
: {
1212 VkPhysicalDeviceVariablePointersFeatures
*features
= (void *)ext
;
1213 features
->variablePointersStorageBuffer
= true;
1214 features
->variablePointers
= true;
1218 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT
: {
1219 VkPhysicalDeviceTransformFeedbackFeaturesEXT
*features
=
1220 (VkPhysicalDeviceTransformFeedbackFeaturesEXT
*)ext
;
1221 features
->transformFeedback
= true;
1222 features
->geometryStreams
= true;
1226 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_UNIFORM_BUFFER_STANDARD_LAYOUT_FEATURES_KHR
: {
1227 VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR
*features
=
1228 (VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR
*)ext
;
1229 features
->uniformBufferStandardLayout
= true;
1233 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT
: {
1234 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*features
=
1235 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*)ext
;
1236 features
->vertexAttributeInstanceRateDivisor
= true;
1237 features
->vertexAttributeInstanceRateZeroDivisor
= true;
1241 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_YCBCR_IMAGE_ARRAYS_FEATURES_EXT
: {
1242 VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*features
=
1243 (VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*)ext
;
1244 features
->ycbcrImageArrays
= true;
1249 anv_debug_ignored_stype(ext
->sType
);
1255 #define MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS 64
1257 #define MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS 64
1258 #define MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS 256
1260 void anv_GetPhysicalDeviceProperties(
1261 VkPhysicalDevice physicalDevice
,
1262 VkPhysicalDeviceProperties
* pProperties
)
1264 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1265 const struct gen_device_info
*devinfo
= &pdevice
->info
;
1267 /* See assertions made when programming the buffer surface state. */
1268 const uint32_t max_raw_buffer_sz
= devinfo
->gen
>= 7 ?
1269 (1ul << 30) : (1ul << 27);
1271 const uint32_t max_ssbos
= pdevice
->has_a64_buffer_access
? UINT16_MAX
: 64;
1272 const uint32_t max_textures
=
1273 pdevice
->has_bindless_images
? UINT16_MAX
: 128;
1274 const uint32_t max_samplers
=
1275 pdevice
->has_bindless_samplers
? UINT16_MAX
:
1276 (devinfo
->gen
>= 8 || devinfo
->is_haswell
) ? 128 : 16;
1277 const uint32_t max_images
=
1278 pdevice
->has_bindless_images
? UINT16_MAX
: MAX_IMAGES
;
1280 /* The moment we have anything bindless, claim a high per-stage limit */
1281 const uint32_t max_per_stage
=
1282 pdevice
->has_a64_buffer_access
? UINT32_MAX
:
1283 MAX_BINDING_TABLE_SIZE
- MAX_RTS
;
1285 const uint32_t max_workgroup_size
= 32 * devinfo
->max_cs_threads
;
1287 VkSampleCountFlags sample_counts
=
1288 isl_device_get_sample_counts(&pdevice
->isl_dev
);
1291 VkPhysicalDeviceLimits limits
= {
1292 .maxImageDimension1D
= (1 << 14),
1293 .maxImageDimension2D
= (1 << 14),
1294 .maxImageDimension3D
= (1 << 11),
1295 .maxImageDimensionCube
= (1 << 14),
1296 .maxImageArrayLayers
= (1 << 11),
1297 .maxTexelBufferElements
= 128 * 1024 * 1024,
1298 .maxUniformBufferRange
= (1ul << 27),
1299 .maxStorageBufferRange
= max_raw_buffer_sz
,
1300 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
1301 .maxMemoryAllocationCount
= UINT32_MAX
,
1302 .maxSamplerAllocationCount
= 64 * 1024,
1303 .bufferImageGranularity
= 64, /* A cache line */
1304 .sparseAddressSpaceSize
= 0,
1305 .maxBoundDescriptorSets
= MAX_SETS
,
1306 .maxPerStageDescriptorSamplers
= max_samplers
,
1307 .maxPerStageDescriptorUniformBuffers
= MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
,
1308 .maxPerStageDescriptorStorageBuffers
= max_ssbos
,
1309 .maxPerStageDescriptorSampledImages
= max_textures
,
1310 .maxPerStageDescriptorStorageImages
= max_images
,
1311 .maxPerStageDescriptorInputAttachments
= MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS
,
1312 .maxPerStageResources
= max_per_stage
,
1313 .maxDescriptorSetSamplers
= 6 * max_samplers
, /* number of stages * maxPerStageDescriptorSamplers */
1314 .maxDescriptorSetUniformBuffers
= 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
, /* number of stages * maxPerStageDescriptorUniformBuffers */
1315 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
1316 .maxDescriptorSetStorageBuffers
= 6 * max_ssbos
, /* number of stages * maxPerStageDescriptorStorageBuffers */
1317 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
1318 .maxDescriptorSetSampledImages
= 6 * max_textures
, /* number of stages * maxPerStageDescriptorSampledImages */
1319 .maxDescriptorSetStorageImages
= 6 * max_images
, /* number of stages * maxPerStageDescriptorStorageImages */
1320 .maxDescriptorSetInputAttachments
= MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS
,
1321 .maxVertexInputAttributes
= MAX_VBS
,
1322 .maxVertexInputBindings
= MAX_VBS
,
1323 .maxVertexInputAttributeOffset
= 2047,
1324 .maxVertexInputBindingStride
= 2048,
1325 .maxVertexOutputComponents
= 128,
1326 .maxTessellationGenerationLevel
= 64,
1327 .maxTessellationPatchSize
= 32,
1328 .maxTessellationControlPerVertexInputComponents
= 128,
1329 .maxTessellationControlPerVertexOutputComponents
= 128,
1330 .maxTessellationControlPerPatchOutputComponents
= 128,
1331 .maxTessellationControlTotalOutputComponents
= 2048,
1332 .maxTessellationEvaluationInputComponents
= 128,
1333 .maxTessellationEvaluationOutputComponents
= 128,
1334 .maxGeometryShaderInvocations
= 32,
1335 .maxGeometryInputComponents
= 64,
1336 .maxGeometryOutputComponents
= 128,
1337 .maxGeometryOutputVertices
= 256,
1338 .maxGeometryTotalOutputComponents
= 1024,
1339 .maxFragmentInputComponents
= 116, /* 128 components - (PSIZ, CLIP_DIST0, CLIP_DIST1) */
1340 .maxFragmentOutputAttachments
= 8,
1341 .maxFragmentDualSrcAttachments
= 1,
1342 .maxFragmentCombinedOutputResources
= 8,
1343 .maxComputeSharedMemorySize
= 64 * 1024,
1344 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
1345 .maxComputeWorkGroupInvocations
= max_workgroup_size
,
1346 .maxComputeWorkGroupSize
= {
1351 .subPixelPrecisionBits
= 8,
1352 .subTexelPrecisionBits
= 8,
1353 .mipmapPrecisionBits
= 8,
1354 .maxDrawIndexedIndexValue
= UINT32_MAX
,
1355 .maxDrawIndirectCount
= UINT32_MAX
,
1356 .maxSamplerLodBias
= 16,
1357 .maxSamplerAnisotropy
= 16,
1358 .maxViewports
= MAX_VIEWPORTS
,
1359 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
1360 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
1361 .viewportSubPixelBits
= 13, /* We take a float? */
1362 .minMemoryMapAlignment
= 4096, /* A page */
1363 /* The dataport requires texel alignment so we need to assume a worst
1364 * case of R32G32B32A32 which is 16 bytes.
1366 .minTexelBufferOffsetAlignment
= 16,
1367 /* We need 16 for UBO block reads to work and 32 for push UBOs */
1368 .minUniformBufferOffsetAlignment
= 32,
1369 .minStorageBufferOffsetAlignment
= 4,
1370 .minTexelOffset
= -8,
1371 .maxTexelOffset
= 7,
1372 .minTexelGatherOffset
= -32,
1373 .maxTexelGatherOffset
= 31,
1374 .minInterpolationOffset
= -0.5,
1375 .maxInterpolationOffset
= 0.4375,
1376 .subPixelInterpolationOffsetBits
= 4,
1377 .maxFramebufferWidth
= (1 << 14),
1378 .maxFramebufferHeight
= (1 << 14),
1379 .maxFramebufferLayers
= (1 << 11),
1380 .framebufferColorSampleCounts
= sample_counts
,
1381 .framebufferDepthSampleCounts
= sample_counts
,
1382 .framebufferStencilSampleCounts
= sample_counts
,
1383 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
1384 .maxColorAttachments
= MAX_RTS
,
1385 .sampledImageColorSampleCounts
= sample_counts
,
1386 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
1387 .sampledImageDepthSampleCounts
= sample_counts
,
1388 .sampledImageStencilSampleCounts
= sample_counts
,
1389 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
1390 .maxSampleMaskWords
= 1,
1391 .timestampComputeAndGraphics
= true,
1392 .timestampPeriod
= 1000000000.0 / devinfo
->timestamp_frequency
,
1393 .maxClipDistances
= 8,
1394 .maxCullDistances
= 8,
1395 .maxCombinedClipAndCullDistances
= 8,
1396 .discreteQueuePriorities
= 2,
1397 .pointSizeRange
= { 0.125, 255.875 },
1400 (devinfo
->gen
>= 9 || devinfo
->is_cherryview
) ?
1401 2047.9921875 : 7.9921875,
1403 .pointSizeGranularity
= (1.0 / 8.0),
1404 .lineWidthGranularity
= (1.0 / 128.0),
1405 .strictLines
= false,
1406 .standardSampleLocations
= true,
1407 .optimalBufferCopyOffsetAlignment
= 128,
1408 .optimalBufferCopyRowPitchAlignment
= 128,
1409 .nonCoherentAtomSize
= 64,
1412 *pProperties
= (VkPhysicalDeviceProperties
) {
1413 .apiVersion
= anv_physical_device_api_version(pdevice
),
1414 .driverVersion
= vk_get_driver_version(),
1416 .deviceID
= pdevice
->chipset_id
,
1417 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
1419 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
1422 snprintf(pProperties
->deviceName
, sizeof(pProperties
->deviceName
),
1423 "%s", pdevice
->name
);
1424 memcpy(pProperties
->pipelineCacheUUID
,
1425 pdevice
->pipeline_cache_uuid
, VK_UUID_SIZE
);
1428 void anv_GetPhysicalDeviceProperties2(
1429 VkPhysicalDevice physicalDevice
,
1430 VkPhysicalDeviceProperties2
* pProperties
)
1432 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
1434 anv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
1436 vk_foreach_struct(ext
, pProperties
->pNext
) {
1437 switch (ext
->sType
) {
1438 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_STENCIL_RESOLVE_PROPERTIES_KHR
: {
1439 VkPhysicalDeviceDepthStencilResolvePropertiesKHR
*props
=
1440 (VkPhysicalDeviceDepthStencilResolvePropertiesKHR
*)ext
;
1442 /* We support all of the depth resolve modes */
1443 props
->supportedDepthResolveModes
=
1444 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
|
1445 VK_RESOLVE_MODE_AVERAGE_BIT_KHR
|
1446 VK_RESOLVE_MODE_MIN_BIT_KHR
|
1447 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1449 /* Average doesn't make sense for stencil so we don't support that */
1450 props
->supportedStencilResolveModes
=
1451 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
;
1452 if (pdevice
->info
.gen
>= 8) {
1453 /* The advanced stencil resolve modes currently require stencil
1454 * sampling be supported by the hardware.
1456 props
->supportedStencilResolveModes
|=
1457 VK_RESOLVE_MODE_MIN_BIT_KHR
|
1458 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1461 props
->independentResolveNone
= true;
1462 props
->independentResolve
= true;
1466 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES_EXT
: {
1467 VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*props
=
1468 (VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*)ext
;
1470 /* It's a bit hard to exactly map our implementation to the limits
1471 * described here. The bindless surface handle in the extended
1472 * message descriptors is 20 bits and it's an index into the table of
1473 * RENDER_SURFACE_STATE structs that starts at bindless surface base
1474 * address. Given that most things consume two surface states per
1475 * view (general/sampled for textures and write-only/read-write for
1476 * images), we claim 2^19 things.
1478 * For SSBOs, we just use A64 messages so there is no real limit
1479 * there beyond the limit on the total size of a descriptor set.
1481 const unsigned max_bindless_views
= 1 << 19;
1483 props
->maxUpdateAfterBindDescriptorsInAllPools
= max_bindless_views
;
1484 props
->shaderUniformBufferArrayNonUniformIndexingNative
= false;
1485 props
->shaderSampledImageArrayNonUniformIndexingNative
= false;
1486 props
->shaderStorageBufferArrayNonUniformIndexingNative
= true;
1487 props
->shaderStorageImageArrayNonUniformIndexingNative
= false;
1488 props
->shaderInputAttachmentArrayNonUniformIndexingNative
= false;
1489 props
->robustBufferAccessUpdateAfterBind
= true;
1490 props
->quadDivergentImplicitLod
= false;
1491 props
->maxPerStageDescriptorUpdateAfterBindSamplers
= max_bindless_views
;
1492 props
->maxPerStageDescriptorUpdateAfterBindUniformBuffers
= MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
;
1493 props
->maxPerStageDescriptorUpdateAfterBindStorageBuffers
= UINT32_MAX
;
1494 props
->maxPerStageDescriptorUpdateAfterBindSampledImages
= max_bindless_views
;
1495 props
->maxPerStageDescriptorUpdateAfterBindStorageImages
= max_bindless_views
;
1496 props
->maxPerStageDescriptorUpdateAfterBindInputAttachments
= MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS
;
1497 props
->maxPerStageUpdateAfterBindResources
= UINT32_MAX
;
1498 props
->maxDescriptorSetUpdateAfterBindSamplers
= max_bindless_views
;
1499 props
->maxDescriptorSetUpdateAfterBindUniformBuffers
= 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS
;
1500 props
->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2;
1501 props
->maxDescriptorSetUpdateAfterBindStorageBuffers
= UINT32_MAX
;
1502 props
->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2;
1503 props
->maxDescriptorSetUpdateAfterBindSampledImages
= max_bindless_views
;
1504 props
->maxDescriptorSetUpdateAfterBindStorageImages
= max_bindless_views
;
1505 props
->maxDescriptorSetUpdateAfterBindInputAttachments
= MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS
;
1509 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR
: {
1510 VkPhysicalDeviceDriverPropertiesKHR
*driver_props
=
1511 (VkPhysicalDeviceDriverPropertiesKHR
*) ext
;
1513 driver_props
->driverID
= VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA_KHR
;
1514 snprintf(driver_props
->driverName
, VK_MAX_DRIVER_NAME_SIZE_KHR
,
1515 "Intel open-source Mesa driver");
1517 snprintf(driver_props
->driverInfo
, VK_MAX_DRIVER_INFO_SIZE_KHR
,
1518 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1
);
1520 driver_props
->conformanceVersion
= (VkConformanceVersionKHR
) {
1529 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT
: {
1530 VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*props
=
1531 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*) ext
;
1532 /* Userptr needs page aligned memory. */
1533 props
->minImportedHostPointerAlignment
= 4096;
1537 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES
: {
1538 VkPhysicalDeviceIDProperties
*id_props
=
1539 (VkPhysicalDeviceIDProperties
*)ext
;
1540 memcpy(id_props
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
1541 memcpy(id_props
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
1542 /* The LUID is for Windows. */
1543 id_props
->deviceLUIDValid
= false;
1547 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_PROPERTIES_EXT
: {
1548 VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*props
=
1549 (VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*)ext
;
1550 props
->maxInlineUniformBlockSize
= MAX_INLINE_UNIFORM_BLOCK_SIZE
;
1551 props
->maxPerStageDescriptorInlineUniformBlocks
=
1552 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1553 props
->maxPerStageDescriptorUpdateAfterBindInlineUniformBlocks
=
1554 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1555 props
->maxDescriptorSetInlineUniformBlocks
=
1556 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1557 props
->maxDescriptorSetUpdateAfterBindInlineUniformBlocks
=
1558 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS
;
1562 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_PROPERTIES_EXT
: {
1563 VkPhysicalDeviceLineRasterizationPropertiesEXT
*props
=
1564 (VkPhysicalDeviceLineRasterizationPropertiesEXT
*)ext
;
1565 /* In the Skylake PRM Vol. 7, subsection titled "GIQ (Diamond)
1566 * Sampling Rules - Legacy Mode", it says the following:
1568 * "Note that the device divides a pixel into a 16x16 array of
1569 * subpixels, referenced by their upper left corners."
1571 * This is the only known reference in the PRMs to the subpixel
1572 * precision of line rasterization and a "16x16 array of subpixels"
1573 * implies 4 subpixel precision bits. Empirical testing has shown
1574 * that 4 subpixel precision bits applies to all line rasterization
1577 props
->lineSubPixelPrecisionBits
= 4;
1581 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES
: {
1582 VkPhysicalDeviceMaintenance3Properties
*props
=
1583 (VkPhysicalDeviceMaintenance3Properties
*)ext
;
1584 /* This value doesn't matter for us today as our per-stage
1585 * descriptors are the real limit.
1587 props
->maxPerSetDescriptors
= 1024;
1588 props
->maxMemoryAllocationSize
= MAX_MEMORY_ALLOCATION_SIZE
;
1592 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES
: {
1593 VkPhysicalDeviceMultiviewProperties
*properties
=
1594 (VkPhysicalDeviceMultiviewProperties
*)ext
;
1595 properties
->maxMultiviewViewCount
= 16;
1596 properties
->maxMultiviewInstanceIndex
= UINT32_MAX
/ 16;
1600 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT
: {
1601 VkPhysicalDevicePCIBusInfoPropertiesEXT
*properties
=
1602 (VkPhysicalDevicePCIBusInfoPropertiesEXT
*)ext
;
1603 properties
->pciDomain
= pdevice
->pci_info
.domain
;
1604 properties
->pciBus
= pdevice
->pci_info
.bus
;
1605 properties
->pciDevice
= pdevice
->pci_info
.device
;
1606 properties
->pciFunction
= pdevice
->pci_info
.function
;
1610 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES
: {
1611 VkPhysicalDevicePointClippingProperties
*properties
=
1612 (VkPhysicalDevicePointClippingProperties
*) ext
;
1613 properties
->pointClippingBehavior
= VK_POINT_CLIPPING_BEHAVIOR_USER_CLIP_PLANES_ONLY
;
1617 #pragma GCC diagnostic push
1618 #pragma GCC diagnostic ignored "-Wswitch"
1619 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PRESENTATION_PROPERTIES_ANDROID
: {
1620 VkPhysicalDevicePresentationPropertiesANDROID
*props
=
1621 (VkPhysicalDevicePresentationPropertiesANDROID
*)ext
;
1622 props
->sharedImage
= VK_FALSE
;
1625 #pragma GCC diagnostic pop
1627 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES
: {
1628 VkPhysicalDeviceProtectedMemoryProperties
*props
=
1629 (VkPhysicalDeviceProtectedMemoryProperties
*)ext
;
1630 props
->protectedNoFault
= false;
1634 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
1635 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
1636 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
1638 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
1642 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT
: {
1643 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*properties
=
1644 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*)ext
;
1645 properties
->filterMinmaxImageComponentMapping
= pdevice
->info
.gen
>= 9;
1646 properties
->filterMinmaxSingleComponentFormats
= true;
1650 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES
: {
1651 VkPhysicalDeviceSubgroupProperties
*properties
= (void *)ext
;
1653 properties
->subgroupSize
= BRW_SUBGROUP_SIZE
;
1655 VkShaderStageFlags scalar_stages
= 0;
1656 for (unsigned stage
= 0; stage
< MESA_SHADER_STAGES
; stage
++) {
1657 if (pdevice
->compiler
->scalar_stage
[stage
])
1658 scalar_stages
|= mesa_to_vk_shader_stage(stage
);
1660 properties
->supportedStages
= scalar_stages
;
1662 properties
->supportedOperations
= VK_SUBGROUP_FEATURE_BASIC_BIT
|
1663 VK_SUBGROUP_FEATURE_VOTE_BIT
|
1664 VK_SUBGROUP_FEATURE_BALLOT_BIT
|
1665 VK_SUBGROUP_FEATURE_SHUFFLE_BIT
|
1666 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT
|
1667 VK_SUBGROUP_FEATURE_QUAD_BIT
;
1668 if (pdevice
->info
.gen
>= 8) {
1669 /* TODO: There's no technical reason why these can't be made to
1670 * work on gen7 but they don't at the moment so it's best to leave
1671 * the feature disabled than enabled and broken.
1673 properties
->supportedOperations
|=
1674 VK_SUBGROUP_FEATURE_ARITHMETIC_BIT
|
1675 VK_SUBGROUP_FEATURE_CLUSTERED_BIT
;
1677 properties
->quadOperationsInAllStages
= pdevice
->info
.gen
>= 8;
1681 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_PROPERTIES_EXT
: {
1682 VkPhysicalDeviceSubgroupSizeControlPropertiesEXT
*props
=
1683 (VkPhysicalDeviceSubgroupSizeControlPropertiesEXT
*)ext
;
1684 STATIC_ASSERT(8 <= BRW_SUBGROUP_SIZE
&& BRW_SUBGROUP_SIZE
<= 32);
1685 props
->minSubgroupSize
= 8;
1686 props
->maxSubgroupSize
= 32;
1687 props
->maxComputeWorkgroupSubgroups
= pdevice
->info
.max_cs_threads
;
1688 props
->requiredSubgroupSizeStages
= VK_SHADER_STAGE_COMPUTE_BIT
;
1692 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_PROPERTIES_EXT
: {
1693 VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT
*props
=
1694 (VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT
*)ext
;
1696 /* From the SKL PRM Vol. 2d, docs for RENDER_SURFACE_STATE::Surface
1699 * "For SURFTYPE_BUFFER non-rendertarget surfaces, this field
1700 * specifies the base address of the first element of the surface,
1701 * computed in software by adding the surface base address to the
1702 * byte offset of the element in the buffer. The base address must
1703 * be aligned to element size."
1705 * The typed dataport messages require that things be texel aligned.
1706 * Otherwise, we may just load/store the wrong data or, in the worst
1707 * case, there may be hangs.
1709 props
->storageTexelBufferOffsetAlignmentBytes
= 16;
1710 props
->storageTexelBufferOffsetSingleTexelAlignment
= true;
1712 /* The sampler, however, is much more forgiving and it can handle
1713 * arbitrary byte alignment for linear and buffer surfaces. It's
1714 * hard to find a good PRM citation for this but years of empirical
1715 * experience demonstrate that this is true.
1717 props
->uniformTexelBufferOffsetAlignmentBytes
= 1;
1718 props
->uniformTexelBufferOffsetSingleTexelAlignment
= false;
1722 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT
: {
1723 VkPhysicalDeviceTransformFeedbackPropertiesEXT
*props
=
1724 (VkPhysicalDeviceTransformFeedbackPropertiesEXT
*)ext
;
1726 props
->maxTransformFeedbackStreams
= MAX_XFB_STREAMS
;
1727 props
->maxTransformFeedbackBuffers
= MAX_XFB_BUFFERS
;
1728 props
->maxTransformFeedbackBufferSize
= (1ull << 32);
1729 props
->maxTransformFeedbackStreamDataSize
= 128 * 4;
1730 props
->maxTransformFeedbackBufferDataSize
= 128 * 4;
1731 props
->maxTransformFeedbackBufferDataStride
= 2048;
1732 props
->transformFeedbackQueries
= true;
1733 props
->transformFeedbackStreamsLinesTriangles
= false;
1734 props
->transformFeedbackRasterizationStreamSelect
= false;
1735 props
->transformFeedbackDraw
= true;
1739 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT
: {
1740 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*props
=
1741 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*)ext
;
1742 /* We have to restrict this a bit for multiview */
1743 props
->maxVertexAttribDivisor
= UINT32_MAX
/ 16;
1748 anv_debug_ignored_stype(ext
->sType
);
1754 /* We support exactly one queue family. */
1755 static const VkQueueFamilyProperties
1756 anv_queue_family_properties
= {
1757 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
1758 VK_QUEUE_COMPUTE_BIT
|
1759 VK_QUEUE_TRANSFER_BIT
,
1761 .timestampValidBits
= 36, /* XXX: Real value here */
1762 .minImageTransferGranularity
= { 1, 1, 1 },
1765 void anv_GetPhysicalDeviceQueueFamilyProperties(
1766 VkPhysicalDevice physicalDevice
,
1768 VkQueueFamilyProperties
* pQueueFamilyProperties
)
1770 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pCount
);
1772 vk_outarray_append(&out
, p
) {
1773 *p
= anv_queue_family_properties
;
1777 void anv_GetPhysicalDeviceQueueFamilyProperties2(
1778 VkPhysicalDevice physicalDevice
,
1779 uint32_t* pQueueFamilyPropertyCount
,
1780 VkQueueFamilyProperties2
* pQueueFamilyProperties
)
1783 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pQueueFamilyPropertyCount
);
1785 vk_outarray_append(&out
, p
) {
1786 p
->queueFamilyProperties
= anv_queue_family_properties
;
1788 vk_foreach_struct(s
, p
->pNext
) {
1789 anv_debug_ignored_stype(s
->sType
);
1794 void anv_GetPhysicalDeviceMemoryProperties(
1795 VkPhysicalDevice physicalDevice
,
1796 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
1798 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
1800 pMemoryProperties
->memoryTypeCount
= physical_device
->memory
.type_count
;
1801 for (uint32_t i
= 0; i
< physical_device
->memory
.type_count
; i
++) {
1802 pMemoryProperties
->memoryTypes
[i
] = (VkMemoryType
) {
1803 .propertyFlags
= physical_device
->memory
.types
[i
].propertyFlags
,
1804 .heapIndex
= physical_device
->memory
.types
[i
].heapIndex
,
1808 pMemoryProperties
->memoryHeapCount
= physical_device
->memory
.heap_count
;
1809 for (uint32_t i
= 0; i
< physical_device
->memory
.heap_count
; i
++) {
1810 pMemoryProperties
->memoryHeaps
[i
] = (VkMemoryHeap
) {
1811 .size
= physical_device
->memory
.heaps
[i
].size
,
1812 .flags
= physical_device
->memory
.heaps
[i
].flags
,
1818 anv_get_memory_budget(VkPhysicalDevice physicalDevice
,
1819 VkPhysicalDeviceMemoryBudgetPropertiesEXT
*memoryBudget
)
1821 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
1822 uint64_t sys_available
= get_available_system_memory();
1823 assert(sys_available
> 0);
1825 VkDeviceSize total_heaps_size
= 0;
1826 for (size_t i
= 0; i
< device
->memory
.heap_count
; i
++)
1827 total_heaps_size
+= device
->memory
.heaps
[i
].size
;
1829 for (size_t i
= 0; i
< device
->memory
.heap_count
; i
++) {
1830 VkDeviceSize heap_size
= device
->memory
.heaps
[i
].size
;
1831 VkDeviceSize heap_used
= device
->memory
.heaps
[i
].used
;
1832 VkDeviceSize heap_budget
;
1834 double heap_proportion
= (double) heap_size
/ total_heaps_size
;
1835 VkDeviceSize sys_available_prop
= sys_available
* heap_proportion
;
1838 * Let's not incite the app to starve the system: report at most 90% of
1839 * available system memory.
1841 uint64_t heap_available
= sys_available_prop
* 9 / 10;
1842 heap_budget
= MIN2(heap_size
, heap_used
+ heap_available
);
1845 * Round down to the nearest MB
1847 heap_budget
&= ~((1ull << 20) - 1);
1850 * The heapBudget value must be non-zero for array elements less than
1851 * VkPhysicalDeviceMemoryProperties::memoryHeapCount. The heapBudget
1852 * value must be less than or equal to VkMemoryHeap::size for each heap.
1854 assert(0 < heap_budget
&& heap_budget
<= heap_size
);
1856 memoryBudget
->heapUsage
[i
] = heap_used
;
1857 memoryBudget
->heapBudget
[i
] = heap_budget
;
1860 /* The heapBudget and heapUsage values must be zero for array elements
1861 * greater than or equal to VkPhysicalDeviceMemoryProperties::memoryHeapCount
1863 for (uint32_t i
= device
->memory
.heap_count
; i
< VK_MAX_MEMORY_HEAPS
; i
++) {
1864 memoryBudget
->heapBudget
[i
] = 0;
1865 memoryBudget
->heapUsage
[i
] = 0;
1869 void anv_GetPhysicalDeviceMemoryProperties2(
1870 VkPhysicalDevice physicalDevice
,
1871 VkPhysicalDeviceMemoryProperties2
* pMemoryProperties
)
1873 anv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
1874 &pMemoryProperties
->memoryProperties
);
1876 vk_foreach_struct(ext
, pMemoryProperties
->pNext
) {
1877 switch (ext
->sType
) {
1878 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT
:
1879 anv_get_memory_budget(physicalDevice
, (void*)ext
);
1882 anv_debug_ignored_stype(ext
->sType
);
1889 anv_GetDeviceGroupPeerMemoryFeatures(
1892 uint32_t localDeviceIndex
,
1893 uint32_t remoteDeviceIndex
,
1894 VkPeerMemoryFeatureFlags
* pPeerMemoryFeatures
)
1896 assert(localDeviceIndex
== 0 && remoteDeviceIndex
== 0);
1897 *pPeerMemoryFeatures
= VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT
|
1898 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT
|
1899 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT
|
1900 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT
;
1903 PFN_vkVoidFunction
anv_GetInstanceProcAddr(
1904 VkInstance _instance
,
1907 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1909 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
1910 * when we have to return valid function pointers, NULL, or it's left
1911 * undefined. See the table for exact details.
1916 #define LOOKUP_ANV_ENTRYPOINT(entrypoint) \
1917 if (strcmp(pName, "vk" #entrypoint) == 0) \
1918 return (PFN_vkVoidFunction)anv_##entrypoint
1920 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceExtensionProperties
);
1921 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceLayerProperties
);
1922 LOOKUP_ANV_ENTRYPOINT(EnumerateInstanceVersion
);
1923 LOOKUP_ANV_ENTRYPOINT(CreateInstance
);
1925 #undef LOOKUP_ANV_ENTRYPOINT
1927 if (instance
== NULL
)
1930 int idx
= anv_get_instance_entrypoint_index(pName
);
1932 return instance
->dispatch
.entrypoints
[idx
];
1934 idx
= anv_get_device_entrypoint_index(pName
);
1936 return instance
->device_dispatch
.entrypoints
[idx
];
1941 /* With version 1+ of the loader interface the ICD should expose
1942 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
1945 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1946 VkInstance instance
,
1950 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1951 VkInstance instance
,
1954 return anv_GetInstanceProcAddr(instance
, pName
);
1957 PFN_vkVoidFunction
anv_GetDeviceProcAddr(
1961 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1963 if (!device
|| !pName
)
1966 int idx
= anv_get_device_entrypoint_index(pName
);
1970 return device
->dispatch
.entrypoints
[idx
];
1974 anv_CreateDebugReportCallbackEXT(VkInstance _instance
,
1975 const VkDebugReportCallbackCreateInfoEXT
* pCreateInfo
,
1976 const VkAllocationCallbacks
* pAllocator
,
1977 VkDebugReportCallbackEXT
* pCallback
)
1979 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1980 return vk_create_debug_report_callback(&instance
->debug_report_callbacks
,
1981 pCreateInfo
, pAllocator
, &instance
->alloc
,
1986 anv_DestroyDebugReportCallbackEXT(VkInstance _instance
,
1987 VkDebugReportCallbackEXT _callback
,
1988 const VkAllocationCallbacks
* pAllocator
)
1990 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
1991 vk_destroy_debug_report_callback(&instance
->debug_report_callbacks
,
1992 _callback
, pAllocator
, &instance
->alloc
);
1996 anv_DebugReportMessageEXT(VkInstance _instance
,
1997 VkDebugReportFlagsEXT flags
,
1998 VkDebugReportObjectTypeEXT objectType
,
2001 int32_t messageCode
,
2002 const char* pLayerPrefix
,
2003 const char* pMessage
)
2005 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
2006 vk_debug_report(&instance
->debug_report_callbacks
, flags
, objectType
,
2007 object
, location
, messageCode
, pLayerPrefix
, pMessage
);
2011 anv_queue_init(struct anv_device
*device
, struct anv_queue
*queue
)
2013 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
2014 queue
->device
= device
;
2019 anv_queue_finish(struct anv_queue
*queue
)
2023 static struct anv_state
2024 anv_state_pool_emit_data(struct anv_state_pool
*pool
, size_t size
, size_t align
, const void *p
)
2026 struct anv_state state
;
2028 state
= anv_state_pool_alloc(pool
, size
, align
);
2029 memcpy(state
.map
, p
, size
);
2034 /* Haswell border color is a bit of a disaster. Float and unorm formats use a
2035 * straightforward 32-bit float color in the first 64 bytes. Instead of using
2036 * a nice float/integer union like Gen8+, Haswell specifies the integer border
2037 * color as a separate entry /after/ the float color. The layout of this entry
2038 * also depends on the format's bpp (with extra hacks for RG32), and overlaps.
2040 * Since we don't know the format/bpp, we can't make any of the border colors
2041 * containing '1' work for all formats, as it would be in the wrong place for
2042 * some of them. We opt to make 32-bit integers work as this seems like the
2043 * most common option. Fortunately, transparent black works regardless, as
2044 * all zeroes is the same in every bit-size.
2046 struct hsw_border_color
{
2050 uint32_t _pad1
[108];
2053 struct gen8_border_color
{
2058 /* Pad out to 64 bytes */
2063 anv_device_init_border_colors(struct anv_device
*device
)
2065 if (device
->info
.is_haswell
) {
2066 static const struct hsw_border_color border_colors
[] = {
2067 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
2068 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
2069 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
2070 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
2071 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
2072 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
2075 device
->border_colors
=
2076 anv_state_pool_emit_data(&device
->dynamic_state_pool
,
2077 sizeof(border_colors
), 512, border_colors
);
2079 static const struct gen8_border_color border_colors
[] = {
2080 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
2081 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
2082 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
2083 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
2084 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
2085 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
2088 device
->border_colors
=
2089 anv_state_pool_emit_data(&device
->dynamic_state_pool
,
2090 sizeof(border_colors
), 64, border_colors
);
2095 anv_device_init_trivial_batch(struct anv_device
*device
)
2097 anv_bo_init_new(&device
->trivial_batch_bo
, device
, 4096);
2099 if (device
->instance
->physicalDevice
.has_exec_async
)
2100 device
->trivial_batch_bo
.flags
|= EXEC_OBJECT_ASYNC
;
2102 if (device
->instance
->physicalDevice
.use_softpin
)
2103 device
->trivial_batch_bo
.flags
|= EXEC_OBJECT_PINNED
;
2105 anv_vma_alloc(device
, &device
->trivial_batch_bo
);
2107 void *map
= anv_gem_mmap(device
, device
->trivial_batch_bo
.gem_handle
,
2110 struct anv_batch batch
= {
2116 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
2117 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
2119 if (!device
->info
.has_llc
)
2120 gen_clflush_range(map
, batch
.next
- map
);
2122 anv_gem_munmap(map
, device
->trivial_batch_bo
.size
);
2125 VkResult
anv_EnumerateDeviceExtensionProperties(
2126 VkPhysicalDevice physicalDevice
,
2127 const char* pLayerName
,
2128 uint32_t* pPropertyCount
,
2129 VkExtensionProperties
* pProperties
)
2131 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
2132 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
2134 for (int i
= 0; i
< ANV_DEVICE_EXTENSION_COUNT
; i
++) {
2135 if (device
->supported_extensions
.extensions
[i
]) {
2136 vk_outarray_append(&out
, prop
) {
2137 *prop
= anv_device_extensions
[i
];
2142 return vk_outarray_status(&out
);
2146 anv_device_init_dispatch(struct anv_device
*device
)
2148 const struct anv_device_dispatch_table
*genX_table
;
2149 switch (device
->info
.gen
) {
2151 genX_table
= &gen12_device_dispatch_table
;
2154 genX_table
= &gen11_device_dispatch_table
;
2157 genX_table
= &gen10_device_dispatch_table
;
2160 genX_table
= &gen9_device_dispatch_table
;
2163 genX_table
= &gen8_device_dispatch_table
;
2166 if (device
->info
.is_haswell
)
2167 genX_table
= &gen75_device_dispatch_table
;
2169 genX_table
= &gen7_device_dispatch_table
;
2172 unreachable("unsupported gen\n");
2175 for (unsigned i
= 0; i
< ARRAY_SIZE(device
->dispatch
.entrypoints
); i
++) {
2176 /* Vulkan requires that entrypoints for extensions which have not been
2177 * enabled must not be advertised.
2179 if (!anv_device_entrypoint_is_enabled(i
, device
->instance
->app_info
.api_version
,
2180 &device
->instance
->enabled_extensions
,
2181 &device
->enabled_extensions
)) {
2182 device
->dispatch
.entrypoints
[i
] = NULL
;
2183 } else if (genX_table
->entrypoints
[i
]) {
2184 device
->dispatch
.entrypoints
[i
] = genX_table
->entrypoints
[i
];
2186 device
->dispatch
.entrypoints
[i
] =
2187 anv_device_dispatch_table
.entrypoints
[i
];
2193 vk_priority_to_gen(int priority
)
2196 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT
:
2197 return GEN_CONTEXT_LOW_PRIORITY
;
2198 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
:
2199 return GEN_CONTEXT_MEDIUM_PRIORITY
;
2200 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT
:
2201 return GEN_CONTEXT_HIGH_PRIORITY
;
2202 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT
:
2203 return GEN_CONTEXT_REALTIME_PRIORITY
;
2205 unreachable("Invalid priority");
2210 anv_device_init_hiz_clear_value_bo(struct anv_device
*device
)
2212 anv_bo_init_new(&device
->hiz_clear_bo
, device
, 4096);
2214 if (device
->instance
->physicalDevice
.has_exec_async
)
2215 device
->hiz_clear_bo
.flags
|= EXEC_OBJECT_ASYNC
;
2217 if (device
->instance
->physicalDevice
.use_softpin
)
2218 device
->hiz_clear_bo
.flags
|= EXEC_OBJECT_PINNED
;
2220 anv_vma_alloc(device
, &device
->hiz_clear_bo
);
2222 uint32_t *map
= anv_gem_mmap(device
, device
->hiz_clear_bo
.gem_handle
,
2225 union isl_color_value hiz_clear
= { .u32
= { 0, } };
2226 hiz_clear
.f32
[0] = ANV_HZ_FC_VAL
;
2228 memcpy(map
, hiz_clear
.u32
, sizeof(hiz_clear
.u32
));
2229 anv_gem_munmap(map
, device
->hiz_clear_bo
.size
);
2233 get_bo_from_pool(struct gen_batch_decode_bo
*ret
,
2234 struct anv_block_pool
*pool
,
2237 for (uint32_t i
= 0; i
< pool
->nbos
; i
++) {
2238 uint64_t bo_address
= pool
->bos
[i
].offset
& (~0ull >> 16);
2239 uint32_t bo_size
= pool
->bos
[i
].size
;
2240 if (address
>= bo_address
&& address
< (bo_address
+ bo_size
)) {
2241 *ret
= (struct gen_batch_decode_bo
) {
2244 .map
= pool
->bos
[i
].map
,
2252 /* Finding a buffer for batch decoding */
2253 static struct gen_batch_decode_bo
2254 decode_get_bo(void *v_batch
, bool ppgtt
, uint64_t address
)
2256 struct anv_device
*device
= v_batch
;
2257 struct gen_batch_decode_bo ret_bo
= {};
2261 if (get_bo_from_pool(&ret_bo
, &device
->dynamic_state_pool
.block_pool
, address
))
2263 if (get_bo_from_pool(&ret_bo
, &device
->instruction_state_pool
.block_pool
, address
))
2265 if (get_bo_from_pool(&ret_bo
, &device
->binding_table_pool
.block_pool
, address
))
2267 if (get_bo_from_pool(&ret_bo
, &device
->surface_state_pool
.block_pool
, address
))
2270 if (!device
->cmd_buffer_being_decoded
)
2271 return (struct gen_batch_decode_bo
) { };
2273 struct anv_batch_bo
**bo
;
2275 u_vector_foreach(bo
, &device
->cmd_buffer_being_decoded
->seen_bbos
) {
2276 /* The decoder zeroes out the top 16 bits, so we need to as well */
2277 uint64_t bo_address
= (*bo
)->bo
.offset
& (~0ull >> 16);
2279 if (address
>= bo_address
&& address
< bo_address
+ (*bo
)->bo
.size
) {
2280 return (struct gen_batch_decode_bo
) {
2282 .size
= (*bo
)->bo
.size
,
2283 .map
= (*bo
)->bo
.map
,
2288 return (struct gen_batch_decode_bo
) { };
2291 VkResult
anv_CreateDevice(
2292 VkPhysicalDevice physicalDevice
,
2293 const VkDeviceCreateInfo
* pCreateInfo
,
2294 const VkAllocationCallbacks
* pAllocator
,
2297 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
2299 struct anv_device
*device
;
2301 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO
);
2303 struct anv_device_extension_table enabled_extensions
= { };
2304 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
2306 for (idx
= 0; idx
< ANV_DEVICE_EXTENSION_COUNT
; idx
++) {
2307 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
2308 anv_device_extensions
[idx
].extensionName
) == 0)
2312 if (idx
>= ANV_DEVICE_EXTENSION_COUNT
)
2313 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
2315 if (!physical_device
->supported_extensions
.extensions
[idx
])
2316 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
2318 enabled_extensions
.extensions
[idx
] = true;
2321 /* Check enabled features */
2322 if (pCreateInfo
->pEnabledFeatures
) {
2323 VkPhysicalDeviceFeatures supported_features
;
2324 anv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
2325 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
2326 VkBool32
*enabled_feature
= (VkBool32
*)pCreateInfo
->pEnabledFeatures
;
2327 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
2328 for (uint32_t i
= 0; i
< num_features
; i
++) {
2329 if (enabled_feature
[i
] && !supported_feature
[i
])
2330 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
2334 /* Check requested queues and fail if we are requested to create any
2335 * queues with flags we don't support.
2337 assert(pCreateInfo
->queueCreateInfoCount
> 0);
2338 for (uint32_t i
= 0; i
< pCreateInfo
->queueCreateInfoCount
; i
++) {
2339 if (pCreateInfo
->pQueueCreateInfos
[i
].flags
!= 0)
2340 return vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2343 /* Check if client specified queue priority. */
2344 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*queue_priority
=
2345 vk_find_struct_const(pCreateInfo
->pQueueCreateInfos
[0].pNext
,
2346 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT
);
2348 VkQueueGlobalPriorityEXT priority
=
2349 queue_priority
? queue_priority
->globalPriority
:
2350 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
;
2352 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
2354 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
2356 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2358 if (INTEL_DEBUG
& DEBUG_BATCH
) {
2359 const unsigned decode_flags
=
2360 GEN_BATCH_DECODE_FULL
|
2361 ((INTEL_DEBUG
& DEBUG_COLOR
) ? GEN_BATCH_DECODE_IN_COLOR
: 0) |
2362 GEN_BATCH_DECODE_OFFSETS
|
2363 GEN_BATCH_DECODE_FLOATS
;
2365 gen_batch_decode_ctx_init(&device
->decoder_ctx
,
2366 &physical_device
->info
,
2367 stderr
, decode_flags
, NULL
,
2368 decode_get_bo
, NULL
, device
);
2371 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
2372 device
->instance
= physical_device
->instance
;
2373 device
->chipset_id
= physical_device
->chipset_id
;
2374 device
->no_hw
= physical_device
->no_hw
;
2375 device
->_lost
= false;
2378 device
->alloc
= *pAllocator
;
2380 device
->alloc
= physical_device
->instance
->alloc
;
2382 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
2383 device
->fd
= open(physical_device
->path
, O_RDWR
| O_CLOEXEC
);
2384 if (device
->fd
== -1) {
2385 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2389 device
->context_id
= anv_gem_create_context(device
);
2390 if (device
->context_id
== -1) {
2391 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2395 if (physical_device
->use_softpin
) {
2396 if (pthread_mutex_init(&device
->vma_mutex
, NULL
) != 0) {
2397 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2401 /* keep the page with address zero out of the allocator */
2402 struct anv_memory_heap
*low_heap
=
2403 &physical_device
->memory
.heaps
[physical_device
->memory
.heap_count
- 1];
2404 util_vma_heap_init(&device
->vma_lo
, low_heap
->vma_start
, low_heap
->vma_size
);
2405 device
->vma_lo_available
= low_heap
->size
;
2407 struct anv_memory_heap
*high_heap
=
2408 &physical_device
->memory
.heaps
[0];
2409 util_vma_heap_init(&device
->vma_hi
, high_heap
->vma_start
, high_heap
->vma_size
);
2410 device
->vma_hi_available
= physical_device
->memory
.heap_count
== 1 ? 0 :
2414 list_inithead(&device
->memory_objects
);
2416 /* As per spec, the driver implementation may deny requests to acquire
2417 * a priority above the default priority (MEDIUM) if the caller does not
2418 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
2421 if (physical_device
->has_context_priority
) {
2422 int err
= anv_gem_set_context_param(device
->fd
, device
->context_id
,
2423 I915_CONTEXT_PARAM_PRIORITY
,
2424 vk_priority_to_gen(priority
));
2425 if (err
!= 0 && priority
> VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
) {
2426 result
= vk_error(VK_ERROR_NOT_PERMITTED_EXT
);
2431 device
->info
= physical_device
->info
;
2432 device
->isl_dev
= physical_device
->isl_dev
;
2434 /* On Broadwell and later, we can use batch chaining to more efficiently
2435 * implement growing command buffers. Prior to Haswell, the kernel
2436 * command parser gets in the way and we have to fall back to growing
2439 device
->can_chain_batches
= device
->info
.gen
>= 8;
2441 device
->robust_buffer_access
= pCreateInfo
->pEnabledFeatures
&&
2442 pCreateInfo
->pEnabledFeatures
->robustBufferAccess
;
2443 device
->enabled_extensions
= enabled_extensions
;
2445 anv_device_init_dispatch(device
);
2447 if (pthread_mutex_init(&device
->mutex
, NULL
) != 0) {
2448 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2449 goto fail_context_id
;
2452 pthread_condattr_t condattr
;
2453 if (pthread_condattr_init(&condattr
) != 0) {
2454 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2457 if (pthread_condattr_setclock(&condattr
, CLOCK_MONOTONIC
) != 0) {
2458 pthread_condattr_destroy(&condattr
);
2459 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2462 if (pthread_cond_init(&device
->queue_submit
, &condattr
) != 0) {
2463 pthread_condattr_destroy(&condattr
);
2464 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
2467 pthread_condattr_destroy(&condattr
);
2470 (physical_device
->supports_48bit_addresses
? EXEC_OBJECT_SUPPORTS_48B_ADDRESS
: 0) |
2471 (physical_device
->has_exec_async
? EXEC_OBJECT_ASYNC
: 0) |
2472 (physical_device
->has_exec_capture
? EXEC_OBJECT_CAPTURE
: 0) |
2473 (physical_device
->use_softpin
? EXEC_OBJECT_PINNED
: 0);
2475 anv_bo_pool_init(&device
->batch_bo_pool
, device
, bo_flags
);
2477 result
= anv_bo_cache_init(&device
->bo_cache
);
2478 if (result
!= VK_SUCCESS
)
2479 goto fail_batch_bo_pool
;
2481 if (!physical_device
->use_softpin
)
2482 bo_flags
&= ~EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
2484 result
= anv_state_pool_init(&device
->dynamic_state_pool
, device
,
2485 DYNAMIC_STATE_POOL_MIN_ADDRESS
,
2488 if (result
!= VK_SUCCESS
)
2491 result
= anv_state_pool_init(&device
->instruction_state_pool
, device
,
2492 INSTRUCTION_STATE_POOL_MIN_ADDRESS
,
2495 if (result
!= VK_SUCCESS
)
2496 goto fail_dynamic_state_pool
;
2498 result
= anv_state_pool_init(&device
->surface_state_pool
, device
,
2499 SURFACE_STATE_POOL_MIN_ADDRESS
,
2502 if (result
!= VK_SUCCESS
)
2503 goto fail_instruction_state_pool
;
2505 if (physical_device
->use_softpin
) {
2506 result
= anv_state_pool_init(&device
->binding_table_pool
, device
,
2507 BINDING_TABLE_POOL_MIN_ADDRESS
,
2510 if (result
!= VK_SUCCESS
)
2511 goto fail_surface_state_pool
;
2514 result
= anv_bo_init_new(&device
->workaround_bo
, device
, 4096);
2515 if (result
!= VK_SUCCESS
)
2516 goto fail_binding_table_pool
;
2518 if (physical_device
->use_softpin
)
2519 device
->workaround_bo
.flags
|= EXEC_OBJECT_PINNED
;
2521 if (!anv_vma_alloc(device
, &device
->workaround_bo
))
2522 goto fail_workaround_bo
;
2524 anv_device_init_trivial_batch(device
);
2526 if (device
->info
.gen
>= 10)
2527 anv_device_init_hiz_clear_value_bo(device
);
2529 anv_scratch_pool_init(device
, &device
->scratch_pool
);
2531 anv_queue_init(device
, &device
->queue
);
2533 switch (device
->info
.gen
) {
2535 if (!device
->info
.is_haswell
)
2536 result
= gen7_init_device_state(device
);
2538 result
= gen75_init_device_state(device
);
2541 result
= gen8_init_device_state(device
);
2544 result
= gen9_init_device_state(device
);
2547 result
= gen10_init_device_state(device
);
2550 result
= gen11_init_device_state(device
);
2553 result
= gen12_init_device_state(device
);
2556 /* Shouldn't get here as we don't create physical devices for any other
2558 unreachable("unhandled gen");
2560 if (result
!= VK_SUCCESS
)
2561 goto fail_workaround_bo
;
2563 anv_pipeline_cache_init(&device
->default_pipeline_cache
, device
, true);
2565 anv_device_init_blorp(device
);
2567 anv_device_init_border_colors(device
);
2569 *pDevice
= anv_device_to_handle(device
);
2574 anv_queue_finish(&device
->queue
);
2575 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
2576 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
2577 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
2578 fail_binding_table_pool
:
2579 if (physical_device
->use_softpin
)
2580 anv_state_pool_finish(&device
->binding_table_pool
);
2581 fail_surface_state_pool
:
2582 anv_state_pool_finish(&device
->surface_state_pool
);
2583 fail_instruction_state_pool
:
2584 anv_state_pool_finish(&device
->instruction_state_pool
);
2585 fail_dynamic_state_pool
:
2586 anv_state_pool_finish(&device
->dynamic_state_pool
);
2588 anv_bo_cache_finish(&device
->bo_cache
);
2590 anv_bo_pool_finish(&device
->batch_bo_pool
);
2591 pthread_cond_destroy(&device
->queue_submit
);
2593 pthread_mutex_destroy(&device
->mutex
);
2595 anv_gem_destroy_context(device
, device
->context_id
);
2599 vk_free(&device
->alloc
, device
);
2604 void anv_DestroyDevice(
2606 const VkAllocationCallbacks
* pAllocator
)
2608 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2609 struct anv_physical_device
*physical_device
;
2614 physical_device
= &device
->instance
->physicalDevice
;
2616 anv_device_finish_blorp(device
);
2618 anv_pipeline_cache_finish(&device
->default_pipeline_cache
);
2620 anv_queue_finish(&device
->queue
);
2622 #ifdef HAVE_VALGRIND
2623 /* We only need to free these to prevent valgrind errors. The backing
2624 * BO will go away in a couple of lines so we don't actually leak.
2626 anv_state_pool_free(&device
->dynamic_state_pool
, device
->border_colors
);
2627 anv_state_pool_free(&device
->dynamic_state_pool
, device
->slice_hash
);
2630 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
2632 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
2633 anv_vma_free(device
, &device
->workaround_bo
);
2634 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
2636 anv_vma_free(device
, &device
->trivial_batch_bo
);
2637 anv_gem_close(device
, device
->trivial_batch_bo
.gem_handle
);
2638 if (device
->info
.gen
>= 10)
2639 anv_gem_close(device
, device
->hiz_clear_bo
.gem_handle
);
2641 if (physical_device
->use_softpin
)
2642 anv_state_pool_finish(&device
->binding_table_pool
);
2643 anv_state_pool_finish(&device
->surface_state_pool
);
2644 anv_state_pool_finish(&device
->instruction_state_pool
);
2645 anv_state_pool_finish(&device
->dynamic_state_pool
);
2647 anv_bo_cache_finish(&device
->bo_cache
);
2649 anv_bo_pool_finish(&device
->batch_bo_pool
);
2651 pthread_cond_destroy(&device
->queue_submit
);
2652 pthread_mutex_destroy(&device
->mutex
);
2654 anv_gem_destroy_context(device
, device
->context_id
);
2656 if (INTEL_DEBUG
& DEBUG_BATCH
)
2657 gen_batch_decode_ctx_finish(&device
->decoder_ctx
);
2661 vk_free(&device
->alloc
, device
);
2664 VkResult
anv_EnumerateInstanceLayerProperties(
2665 uint32_t* pPropertyCount
,
2666 VkLayerProperties
* pProperties
)
2668 if (pProperties
== NULL
) {
2669 *pPropertyCount
= 0;
2673 /* None supported at this time */
2674 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
2677 VkResult
anv_EnumerateDeviceLayerProperties(
2678 VkPhysicalDevice physicalDevice
,
2679 uint32_t* pPropertyCount
,
2680 VkLayerProperties
* pProperties
)
2682 if (pProperties
== NULL
) {
2683 *pPropertyCount
= 0;
2687 /* None supported at this time */
2688 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
2691 void anv_GetDeviceQueue(
2693 uint32_t queueNodeIndex
,
2694 uint32_t queueIndex
,
2697 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2699 assert(queueIndex
== 0);
2701 *pQueue
= anv_queue_to_handle(&device
->queue
);
2704 void anv_GetDeviceQueue2(
2706 const VkDeviceQueueInfo2
* pQueueInfo
,
2709 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2711 assert(pQueueInfo
->queueIndex
== 0);
2713 if (pQueueInfo
->flags
== device
->queue
.flags
)
2714 *pQueue
= anv_queue_to_handle(&device
->queue
);
2720 _anv_device_set_lost(struct anv_device
*device
,
2721 const char *file
, int line
,
2722 const char *msg
, ...)
2727 device
->_lost
= true;
2730 err
= __vk_errorv(device
->instance
, device
,
2731 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT
,
2732 VK_ERROR_DEVICE_LOST
, file
, line
, msg
, ap
);
2735 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
2742 anv_device_query_status(struct anv_device
*device
)
2744 /* This isn't likely as most of the callers of this function already check
2745 * for it. However, it doesn't hurt to check and it potentially lets us
2748 if (anv_device_is_lost(device
))
2749 return VK_ERROR_DEVICE_LOST
;
2751 uint32_t active
, pending
;
2752 int ret
= anv_gem_gpu_get_reset_stats(device
, &active
, &pending
);
2754 /* We don't know the real error. */
2755 return anv_device_set_lost(device
, "get_reset_stats failed: %m");
2759 return anv_device_set_lost(device
, "GPU hung on one of our command buffers");
2760 } else if (pending
) {
2761 return anv_device_set_lost(device
, "GPU hung with commands in-flight");
2768 anv_device_bo_busy(struct anv_device
*device
, struct anv_bo
*bo
)
2770 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
2771 * Other usages of the BO (such as on different hardware) will not be
2772 * flagged as "busy" by this ioctl. Use with care.
2774 int ret
= anv_gem_busy(device
, bo
->gem_handle
);
2776 return VK_NOT_READY
;
2777 } else if (ret
== -1) {
2778 /* We don't know the real error. */
2779 return anv_device_set_lost(device
, "gem wait failed: %m");
2782 /* Query for device status after the busy call. If the BO we're checking
2783 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
2784 * client because it clearly doesn't have valid data. Yes, this most
2785 * likely means an ioctl, but we just did an ioctl to query the busy status
2786 * so it's no great loss.
2788 return anv_device_query_status(device
);
2792 anv_device_wait(struct anv_device
*device
, struct anv_bo
*bo
,
2795 int ret
= anv_gem_wait(device
, bo
->gem_handle
, &timeout
);
2796 if (ret
== -1 && errno
== ETIME
) {
2798 } else if (ret
== -1) {
2799 /* We don't know the real error. */
2800 return anv_device_set_lost(device
, "gem wait failed: %m");
2803 /* Query for device status after the wait. If the BO we're waiting on got
2804 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
2805 * because it clearly doesn't have valid data. Yes, this most likely means
2806 * an ioctl, but we just did an ioctl to wait so it's no great loss.
2808 return anv_device_query_status(device
);
2811 VkResult
anv_DeviceWaitIdle(
2814 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2815 if (anv_device_is_lost(device
))
2816 return VK_ERROR_DEVICE_LOST
;
2818 struct anv_batch batch
;
2821 batch
.start
= batch
.next
= cmds
;
2822 batch
.end
= (void *) cmds
+ sizeof(cmds
);
2824 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
2825 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
2827 return anv_device_submit_simple_batch(device
, &batch
);
2831 anv_vma_alloc(struct anv_device
*device
, struct anv_bo
*bo
)
2833 if (!(bo
->flags
& EXEC_OBJECT_PINNED
))
2836 pthread_mutex_lock(&device
->vma_mutex
);
2840 if (bo
->flags
& EXEC_OBJECT_SUPPORTS_48B_ADDRESS
&&
2841 device
->vma_hi_available
>= bo
->size
) {
2842 uint64_t addr
= util_vma_heap_alloc(&device
->vma_hi
, bo
->size
, 4096);
2844 bo
->offset
= gen_canonical_address(addr
);
2845 assert(addr
== gen_48b_address(bo
->offset
));
2846 device
->vma_hi_available
-= bo
->size
;
2850 if (bo
->offset
== 0 && device
->vma_lo_available
>= bo
->size
) {
2851 uint64_t addr
= util_vma_heap_alloc(&device
->vma_lo
, bo
->size
, 4096);
2853 bo
->offset
= gen_canonical_address(addr
);
2854 assert(addr
== gen_48b_address(bo
->offset
));
2855 device
->vma_lo_available
-= bo
->size
;
2859 pthread_mutex_unlock(&device
->vma_mutex
);
2861 return bo
->offset
!= 0;
2865 anv_vma_free(struct anv_device
*device
, struct anv_bo
*bo
)
2867 if (!(bo
->flags
& EXEC_OBJECT_PINNED
))
2870 const uint64_t addr_48b
= gen_48b_address(bo
->offset
);
2872 pthread_mutex_lock(&device
->vma_mutex
);
2874 if (addr_48b
>= LOW_HEAP_MIN_ADDRESS
&&
2875 addr_48b
<= LOW_HEAP_MAX_ADDRESS
) {
2876 util_vma_heap_free(&device
->vma_lo
, addr_48b
, bo
->size
);
2877 device
->vma_lo_available
+= bo
->size
;
2879 ASSERTED
const struct anv_physical_device
*physical_device
=
2880 &device
->instance
->physicalDevice
;
2881 assert(addr_48b
>= physical_device
->memory
.heaps
[0].vma_start
&&
2882 addr_48b
< (physical_device
->memory
.heaps
[0].vma_start
+
2883 physical_device
->memory
.heaps
[0].vma_size
));
2884 util_vma_heap_free(&device
->vma_hi
, addr_48b
, bo
->size
);
2885 device
->vma_hi_available
+= bo
->size
;
2888 pthread_mutex_unlock(&device
->vma_mutex
);
2894 anv_bo_init_new(struct anv_bo
*bo
, struct anv_device
*device
, uint64_t size
)
2896 uint32_t gem_handle
= anv_gem_create(device
, size
);
2898 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
2900 anv_bo_init(bo
, gem_handle
, size
);
2905 VkResult
anv_AllocateMemory(
2907 const VkMemoryAllocateInfo
* pAllocateInfo
,
2908 const VkAllocationCallbacks
* pAllocator
,
2909 VkDeviceMemory
* pMem
)
2911 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2912 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
2913 struct anv_device_memory
*mem
;
2914 VkResult result
= VK_SUCCESS
;
2916 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
2918 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
2919 assert(pAllocateInfo
->allocationSize
> 0);
2921 if (pAllocateInfo
->allocationSize
> MAX_MEMORY_ALLOCATION_SIZE
)
2922 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
2924 /* FINISHME: Fail if allocation request exceeds heap size. */
2926 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
2927 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2929 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2931 assert(pAllocateInfo
->memoryTypeIndex
< pdevice
->memory
.type_count
);
2932 mem
->type
= &pdevice
->memory
.types
[pAllocateInfo
->memoryTypeIndex
];
2936 mem
->host_ptr
= NULL
;
2938 uint64_t bo_flags
= 0;
2940 assert(mem
->type
->heapIndex
< pdevice
->memory
.heap_count
);
2941 if (pdevice
->memory
.heaps
[mem
->type
->heapIndex
].supports_48bit_addresses
)
2942 bo_flags
|= EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
2944 const struct wsi_memory_allocate_info
*wsi_info
=
2945 vk_find_struct_const(pAllocateInfo
->pNext
, WSI_MEMORY_ALLOCATE_INFO_MESA
);
2946 if (wsi_info
&& wsi_info
->implicit_sync
) {
2947 /* We need to set the WRITE flag on window system buffers so that GEM
2948 * will know we're writing to them and synchronize uses on other rings
2949 * (eg if the display server uses the blitter ring).
2951 bo_flags
|= EXEC_OBJECT_WRITE
;
2952 } else if (pdevice
->has_exec_async
) {
2953 bo_flags
|= EXEC_OBJECT_ASYNC
;
2956 if (pdevice
->use_softpin
)
2957 bo_flags
|= EXEC_OBJECT_PINNED
;
2959 const VkExportMemoryAllocateInfo
*export_info
=
2960 vk_find_struct_const(pAllocateInfo
->pNext
, EXPORT_MEMORY_ALLOCATE_INFO
);
2962 /* Check if we need to support Android HW buffer export. If so,
2963 * create AHardwareBuffer and import memory from it.
2965 bool android_export
= false;
2966 if (export_info
&& export_info
->handleTypes
&
2967 VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID
)
2968 android_export
= true;
2970 /* Android memory import. */
2971 const struct VkImportAndroidHardwareBufferInfoANDROID
*ahw_import_info
=
2972 vk_find_struct_const(pAllocateInfo
->pNext
,
2973 IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID
);
2975 if (ahw_import_info
) {
2976 result
= anv_import_ahw_memory(_device
, mem
, ahw_import_info
);
2977 if (result
!= VK_SUCCESS
)
2981 } else if (android_export
) {
2982 result
= anv_create_ahw_memory(_device
, mem
, pAllocateInfo
);
2983 if (result
!= VK_SUCCESS
)
2986 const struct VkImportAndroidHardwareBufferInfoANDROID import_info
= {
2989 result
= anv_import_ahw_memory(_device
, mem
, &import_info
);
2990 if (result
!= VK_SUCCESS
)
2996 const VkImportMemoryFdInfoKHR
*fd_info
=
2997 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_FD_INFO_KHR
);
2999 /* The Vulkan spec permits handleType to be 0, in which case the struct is
3002 if (fd_info
&& fd_info
->handleType
) {
3003 /* At the moment, we support only the below handle types. */
3004 assert(fd_info
->handleType
==
3005 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
3006 fd_info
->handleType
==
3007 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
3009 result
= anv_bo_cache_import(device
, &device
->bo_cache
, fd_info
->fd
,
3010 bo_flags
| ANV_BO_EXTERNAL
, &mem
->bo
);
3011 if (result
!= VK_SUCCESS
)
3014 VkDeviceSize aligned_alloc_size
=
3015 align_u64(pAllocateInfo
->allocationSize
, 4096);
3017 /* For security purposes, we reject importing the bo if it's smaller
3018 * than the requested allocation size. This prevents a malicious client
3019 * from passing a buffer to a trusted client, lying about the size, and
3020 * telling the trusted client to try and texture from an image that goes
3021 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
3022 * in the trusted client. The trusted client can protect itself against
3023 * this sort of attack but only if it can trust the buffer size.
3025 if (mem
->bo
->size
< aligned_alloc_size
) {
3026 result
= vk_errorf(device
->instance
, device
,
3027 VK_ERROR_INVALID_EXTERNAL_HANDLE
,
3028 "aligned allocationSize too large for "
3029 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT: "
3030 "%"PRIu64
"B > %"PRIu64
"B",
3031 aligned_alloc_size
, mem
->bo
->size
);
3032 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
3036 /* From the Vulkan spec:
3038 * "Importing memory from a file descriptor transfers ownership of
3039 * the file descriptor from the application to the Vulkan
3040 * implementation. The application must not perform any operations on
3041 * the file descriptor after a successful import."
3043 * If the import fails, we leave the file descriptor open.
3049 const VkImportMemoryHostPointerInfoEXT
*host_ptr_info
=
3050 vk_find_struct_const(pAllocateInfo
->pNext
,
3051 IMPORT_MEMORY_HOST_POINTER_INFO_EXT
);
3052 if (host_ptr_info
&& host_ptr_info
->handleType
) {
3053 if (host_ptr_info
->handleType
==
3054 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_MAPPED_FOREIGN_MEMORY_BIT_EXT
) {
3055 result
= vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
3059 assert(host_ptr_info
->handleType
==
3060 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
);
3062 result
= anv_bo_cache_import_host_ptr(
3063 device
, &device
->bo_cache
, host_ptr_info
->pHostPointer
,
3064 pAllocateInfo
->allocationSize
, bo_flags
, &mem
->bo
);
3066 if (result
!= VK_SUCCESS
)
3069 mem
->host_ptr
= host_ptr_info
->pHostPointer
;
3073 /* Regular allocate (not importing memory). */
3075 if (export_info
&& export_info
->handleTypes
)
3076 bo_flags
|= ANV_BO_EXTERNAL
;
3078 result
= anv_bo_cache_alloc(device
, &device
->bo_cache
,
3079 pAllocateInfo
->allocationSize
, bo_flags
,
3081 if (result
!= VK_SUCCESS
)
3084 const VkMemoryDedicatedAllocateInfo
*dedicated_info
=
3085 vk_find_struct_const(pAllocateInfo
->pNext
, MEMORY_DEDICATED_ALLOCATE_INFO
);
3086 if (dedicated_info
&& dedicated_info
->image
!= VK_NULL_HANDLE
) {
3087 ANV_FROM_HANDLE(anv_image
, image
, dedicated_info
->image
);
3089 /* Some legacy (non-modifiers) consumers need the tiling to be set on
3090 * the BO. In this case, we have a dedicated allocation.
3092 if (image
->needs_set_tiling
) {
3093 const uint32_t i915_tiling
=
3094 isl_tiling_to_i915_tiling(image
->planes
[0].surface
.isl
.tiling
);
3095 int ret
= anv_gem_set_tiling(device
, mem
->bo
->gem_handle
,
3096 image
->planes
[0].surface
.isl
.row_pitch_B
,
3099 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
3100 return vk_errorf(device
->instance
, NULL
,
3101 VK_ERROR_OUT_OF_DEVICE_MEMORY
,
3102 "failed to set BO tiling: %m");
3108 pthread_mutex_lock(&device
->mutex
);
3109 list_addtail(&mem
->link
, &device
->memory_objects
);
3110 pthread_mutex_unlock(&device
->mutex
);
3112 *pMem
= anv_device_memory_to_handle(mem
);
3114 p_atomic_add(&pdevice
->memory
.heaps
[mem
->type
->heapIndex
].used
,
3120 vk_free2(&device
->alloc
, pAllocator
, mem
);
3125 VkResult
anv_GetMemoryFdKHR(
3127 const VkMemoryGetFdInfoKHR
* pGetFdInfo
,
3130 ANV_FROM_HANDLE(anv_device
, dev
, device_h
);
3131 ANV_FROM_HANDLE(anv_device_memory
, mem
, pGetFdInfo
->memory
);
3133 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
3135 assert(pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
3136 pGetFdInfo
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
3138 return anv_bo_cache_export(dev
, &dev
->bo_cache
, mem
->bo
, pFd
);
3141 VkResult
anv_GetMemoryFdPropertiesKHR(
3143 VkExternalMemoryHandleTypeFlagBits handleType
,
3145 VkMemoryFdPropertiesKHR
* pMemoryFdProperties
)
3147 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3148 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3150 switch (handleType
) {
3151 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
:
3152 /* dma-buf can be imported as any memory type */
3153 pMemoryFdProperties
->memoryTypeBits
=
3154 (1 << pdevice
->memory
.type_count
) - 1;
3158 /* The valid usage section for this function says:
3160 * "handleType must not be one of the handle types defined as
3163 * So opaque handle types fall into the default "unsupported" case.
3165 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE
);
3169 VkResult
anv_GetMemoryHostPointerPropertiesEXT(
3171 VkExternalMemoryHandleTypeFlagBits handleType
,
3172 const void* pHostPointer
,
3173 VkMemoryHostPointerPropertiesEXT
* pMemoryHostPointerProperties
)
3175 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3177 assert(pMemoryHostPointerProperties
->sType
==
3178 VK_STRUCTURE_TYPE_MEMORY_HOST_POINTER_PROPERTIES_EXT
);
3180 switch (handleType
) {
3181 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
: {
3182 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3184 /* Host memory can be imported as any memory type. */
3185 pMemoryHostPointerProperties
->memoryTypeBits
=
3186 (1ull << pdevice
->memory
.type_count
) - 1;
3191 return VK_ERROR_INVALID_EXTERNAL_HANDLE
;
3195 void anv_FreeMemory(
3197 VkDeviceMemory _mem
,
3198 const VkAllocationCallbacks
* pAllocator
)
3200 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3201 ANV_FROM_HANDLE(anv_device_memory
, mem
, _mem
);
3202 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3207 pthread_mutex_lock(&device
->mutex
);
3208 list_del(&mem
->link
);
3209 pthread_mutex_unlock(&device
->mutex
);
3212 anv_UnmapMemory(_device
, _mem
);
3214 p_atomic_add(&pdevice
->memory
.heaps
[mem
->type
->heapIndex
].used
,
3217 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
3219 #if defined(ANDROID) && ANDROID_API_LEVEL >= 26
3221 AHardwareBuffer_release(mem
->ahw
);
3224 vk_free2(&device
->alloc
, pAllocator
, mem
);
3227 VkResult
anv_MapMemory(
3229 VkDeviceMemory _memory
,
3230 VkDeviceSize offset
,
3232 VkMemoryMapFlags flags
,
3235 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3236 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
3243 if (mem
->host_ptr
) {
3244 *ppData
= mem
->host_ptr
+ offset
;
3248 if (size
== VK_WHOLE_SIZE
)
3249 size
= mem
->bo
->size
- offset
;
3251 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
3253 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
3254 * assert(size != 0);
3255 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
3256 * equal to the size of the memory minus offset
3259 assert(offset
+ size
<= mem
->bo
->size
);
3261 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
3262 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
3263 * at a time is valid. We could just mmap up front and return an offset
3264 * pointer here, but that may exhaust virtual memory on 32 bit
3267 uint32_t gem_flags
= 0;
3269 if (!device
->info
.has_llc
&&
3270 (mem
->type
->propertyFlags
& VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
))
3271 gem_flags
|= I915_MMAP_WC
;
3273 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
3274 uint64_t map_offset
= offset
& ~4095ull;
3275 assert(offset
>= map_offset
);
3276 uint64_t map_size
= (offset
+ size
) - map_offset
;
3278 /* Let's map whole pages */
3279 map_size
= align_u64(map_size
, 4096);
3281 void *map
= anv_gem_mmap(device
, mem
->bo
->gem_handle
,
3282 map_offset
, map_size
, gem_flags
);
3283 if (map
== MAP_FAILED
)
3284 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
3287 mem
->map_size
= map_size
;
3289 *ppData
= mem
->map
+ (offset
- map_offset
);
3294 void anv_UnmapMemory(
3296 VkDeviceMemory _memory
)
3298 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
3300 if (mem
== NULL
|| mem
->host_ptr
)
3303 anv_gem_munmap(mem
->map
, mem
->map_size
);
3310 clflush_mapped_ranges(struct anv_device
*device
,
3312 const VkMappedMemoryRange
*ranges
)
3314 for (uint32_t i
= 0; i
< count
; i
++) {
3315 ANV_FROM_HANDLE(anv_device_memory
, mem
, ranges
[i
].memory
);
3316 if (ranges
[i
].offset
>= mem
->map_size
)
3319 gen_clflush_range(mem
->map
+ ranges
[i
].offset
,
3320 MIN2(ranges
[i
].size
, mem
->map_size
- ranges
[i
].offset
));
3324 VkResult
anv_FlushMappedMemoryRanges(
3326 uint32_t memoryRangeCount
,
3327 const VkMappedMemoryRange
* pMemoryRanges
)
3329 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3331 if (device
->info
.has_llc
)
3334 /* Make sure the writes we're flushing have landed. */
3335 __builtin_ia32_mfence();
3337 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
3342 VkResult
anv_InvalidateMappedMemoryRanges(
3344 uint32_t memoryRangeCount
,
3345 const VkMappedMemoryRange
* pMemoryRanges
)
3347 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3349 if (device
->info
.has_llc
)
3352 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
3354 /* Make sure no reads get moved up above the invalidate. */
3355 __builtin_ia32_mfence();
3360 void anv_GetBufferMemoryRequirements(
3363 VkMemoryRequirements
* pMemoryRequirements
)
3365 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
3366 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3367 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3369 /* The Vulkan spec (git aaed022) says:
3371 * memoryTypeBits is a bitfield and contains one bit set for every
3372 * supported memory type for the resource. The bit `1<<i` is set if and
3373 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3374 * structure for the physical device is supported.
3376 uint32_t memory_types
= 0;
3377 for (uint32_t i
= 0; i
< pdevice
->memory
.type_count
; i
++) {
3378 uint32_t valid_usage
= pdevice
->memory
.types
[i
].valid_buffer_usage
;
3379 if ((valid_usage
& buffer
->usage
) == buffer
->usage
)
3380 memory_types
|= (1u << i
);
3383 /* Base alignment requirement of a cache line */
3384 uint32_t alignment
= 16;
3386 /* We need an alignment of 32 for pushing UBOs */
3387 if (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
)
3388 alignment
= MAX2(alignment
, 32);
3390 pMemoryRequirements
->size
= buffer
->size
;
3391 pMemoryRequirements
->alignment
= alignment
;
3393 /* Storage and Uniform buffers should have their size aligned to
3394 * 32-bits to avoid boundary checks when last DWord is not complete.
3395 * This would ensure that not internal padding would be needed for
3398 if (device
->robust_buffer_access
&&
3399 (buffer
->usage
& VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT
||
3400 buffer
->usage
& VK_BUFFER_USAGE_STORAGE_BUFFER_BIT
))
3401 pMemoryRequirements
->size
= align_u64(buffer
->size
, 4);
3403 pMemoryRequirements
->memoryTypeBits
= memory_types
;
3406 void anv_GetBufferMemoryRequirements2(
3408 const VkBufferMemoryRequirementsInfo2
* pInfo
,
3409 VkMemoryRequirements2
* pMemoryRequirements
)
3411 anv_GetBufferMemoryRequirements(_device
, pInfo
->buffer
,
3412 &pMemoryRequirements
->memoryRequirements
);
3414 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3415 switch (ext
->sType
) {
3416 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
3417 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
3418 requirements
->prefersDedicatedAllocation
= false;
3419 requirements
->requiresDedicatedAllocation
= false;
3424 anv_debug_ignored_stype(ext
->sType
);
3430 void anv_GetImageMemoryRequirements(
3433 VkMemoryRequirements
* pMemoryRequirements
)
3435 ANV_FROM_HANDLE(anv_image
, image
, _image
);
3436 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3437 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3439 /* The Vulkan spec (git aaed022) says:
3441 * memoryTypeBits is a bitfield and contains one bit set for every
3442 * supported memory type for the resource. The bit `1<<i` is set if and
3443 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
3444 * structure for the physical device is supported.
3446 * All types are currently supported for images.
3448 uint32_t memory_types
= (1ull << pdevice
->memory
.type_count
) - 1;
3450 /* We must have image allocated or imported at this point. According to the
3451 * specification, external images must have been bound to memory before
3452 * calling GetImageMemoryRequirements.
3454 assert(image
->size
> 0);
3456 pMemoryRequirements
->size
= image
->size
;
3457 pMemoryRequirements
->alignment
= image
->alignment
;
3458 pMemoryRequirements
->memoryTypeBits
= memory_types
;
3461 void anv_GetImageMemoryRequirements2(
3463 const VkImageMemoryRequirementsInfo2
* pInfo
,
3464 VkMemoryRequirements2
* pMemoryRequirements
)
3466 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3467 ANV_FROM_HANDLE(anv_image
, image
, pInfo
->image
);
3469 anv_GetImageMemoryRequirements(_device
, pInfo
->image
,
3470 &pMemoryRequirements
->memoryRequirements
);
3472 vk_foreach_struct_const(ext
, pInfo
->pNext
) {
3473 switch (ext
->sType
) {
3474 case VK_STRUCTURE_TYPE_IMAGE_PLANE_MEMORY_REQUIREMENTS_INFO
: {
3475 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
3476 const VkImagePlaneMemoryRequirementsInfo
*plane_reqs
=
3477 (const VkImagePlaneMemoryRequirementsInfo
*) ext
;
3478 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
,
3479 plane_reqs
->planeAspect
);
3481 assert(image
->planes
[plane
].offset
== 0);
3483 /* The Vulkan spec (git aaed022) says:
3485 * memoryTypeBits is a bitfield and contains one bit set for every
3486 * supported memory type for the resource. The bit `1<<i` is set
3487 * if and only if the memory type `i` in the
3488 * VkPhysicalDeviceMemoryProperties structure for the physical
3489 * device is supported.
3491 * All types are currently supported for images.
3493 pMemoryRequirements
->memoryRequirements
.memoryTypeBits
=
3494 (1ull << pdevice
->memory
.type_count
) - 1;
3496 /* We must have image allocated or imported at this point. According to the
3497 * specification, external images must have been bound to memory before
3498 * calling GetImageMemoryRequirements.
3500 assert(image
->planes
[plane
].size
> 0);
3502 pMemoryRequirements
->memoryRequirements
.size
= image
->planes
[plane
].size
;
3503 pMemoryRequirements
->memoryRequirements
.alignment
=
3504 image
->planes
[plane
].alignment
;
3509 anv_debug_ignored_stype(ext
->sType
);
3514 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3515 switch (ext
->sType
) {
3516 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
3517 VkMemoryDedicatedRequirements
*requirements
= (void *)ext
;
3518 if (image
->needs_set_tiling
|| image
->external_format
) {
3519 /* If we need to set the tiling for external consumers, we need a
3520 * dedicated allocation.
3522 * See also anv_AllocateMemory.
3524 requirements
->prefersDedicatedAllocation
= true;
3525 requirements
->requiresDedicatedAllocation
= true;
3527 requirements
->prefersDedicatedAllocation
= false;
3528 requirements
->requiresDedicatedAllocation
= false;
3534 anv_debug_ignored_stype(ext
->sType
);
3540 void anv_GetImageSparseMemoryRequirements(
3543 uint32_t* pSparseMemoryRequirementCount
,
3544 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
3546 *pSparseMemoryRequirementCount
= 0;
3549 void anv_GetImageSparseMemoryRequirements2(
3551 const VkImageSparseMemoryRequirementsInfo2
* pInfo
,
3552 uint32_t* pSparseMemoryRequirementCount
,
3553 VkSparseImageMemoryRequirements2
* pSparseMemoryRequirements
)
3555 *pSparseMemoryRequirementCount
= 0;
3558 void anv_GetDeviceMemoryCommitment(
3560 VkDeviceMemory memory
,
3561 VkDeviceSize
* pCommittedMemoryInBytes
)
3563 *pCommittedMemoryInBytes
= 0;
3567 anv_bind_buffer_memory(const VkBindBufferMemoryInfo
*pBindInfo
)
3569 ANV_FROM_HANDLE(anv_device_memory
, mem
, pBindInfo
->memory
);
3570 ANV_FROM_HANDLE(anv_buffer
, buffer
, pBindInfo
->buffer
);
3572 assert(pBindInfo
->sType
== VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
);
3575 assert((buffer
->usage
& mem
->type
->valid_buffer_usage
) == buffer
->usage
);
3576 buffer
->address
= (struct anv_address
) {
3578 .offset
= pBindInfo
->memoryOffset
,
3581 buffer
->address
= ANV_NULL_ADDRESS
;
3585 VkResult
anv_BindBufferMemory(
3588 VkDeviceMemory memory
,
3589 VkDeviceSize memoryOffset
)
3591 anv_bind_buffer_memory(
3592 &(VkBindBufferMemoryInfo
) {
3593 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
,
3596 .memoryOffset
= memoryOffset
,
3602 VkResult
anv_BindBufferMemory2(
3604 uint32_t bindInfoCount
,
3605 const VkBindBufferMemoryInfo
* pBindInfos
)
3607 for (uint32_t i
= 0; i
< bindInfoCount
; i
++)
3608 anv_bind_buffer_memory(&pBindInfos
[i
]);
3613 VkResult
anv_QueueBindSparse(
3615 uint32_t bindInfoCount
,
3616 const VkBindSparseInfo
* pBindInfo
,
3619 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
3620 if (anv_device_is_lost(queue
->device
))
3621 return VK_ERROR_DEVICE_LOST
;
3623 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
3628 VkResult
anv_CreateEvent(
3630 const VkEventCreateInfo
* pCreateInfo
,
3631 const VkAllocationCallbacks
* pAllocator
,
3634 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3635 struct anv_state state
;
3636 struct anv_event
*event
;
3638 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_EVENT_CREATE_INFO
);
3640 state
= anv_state_pool_alloc(&device
->dynamic_state_pool
,
3643 event
->state
= state
;
3644 event
->semaphore
= VK_EVENT_RESET
;
3646 if (!device
->info
.has_llc
) {
3647 /* Make sure the writes we're flushing have landed. */
3648 __builtin_ia32_mfence();
3649 __builtin_ia32_clflush(event
);
3652 *pEvent
= anv_event_to_handle(event
);
3657 void anv_DestroyEvent(
3660 const VkAllocationCallbacks
* pAllocator
)
3662 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3663 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3668 anv_state_pool_free(&device
->dynamic_state_pool
, event
->state
);
3671 VkResult
anv_GetEventStatus(
3675 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3676 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3678 if (anv_device_is_lost(device
))
3679 return VK_ERROR_DEVICE_LOST
;
3681 if (!device
->info
.has_llc
) {
3682 /* Invalidate read cache before reading event written by GPU. */
3683 __builtin_ia32_clflush(event
);
3684 __builtin_ia32_mfence();
3688 return event
->semaphore
;
3691 VkResult
anv_SetEvent(
3695 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3696 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3698 event
->semaphore
= VK_EVENT_SET
;
3700 if (!device
->info
.has_llc
) {
3701 /* Make sure the writes we're flushing have landed. */
3702 __builtin_ia32_mfence();
3703 __builtin_ia32_clflush(event
);
3709 VkResult
anv_ResetEvent(
3713 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3714 ANV_FROM_HANDLE(anv_event
, event
, _event
);
3716 event
->semaphore
= VK_EVENT_RESET
;
3718 if (!device
->info
.has_llc
) {
3719 /* Make sure the writes we're flushing have landed. */
3720 __builtin_ia32_mfence();
3721 __builtin_ia32_clflush(event
);
3729 VkResult
anv_CreateBuffer(
3731 const VkBufferCreateInfo
* pCreateInfo
,
3732 const VkAllocationCallbacks
* pAllocator
,
3735 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3736 struct anv_buffer
*buffer
;
3738 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
3740 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
3741 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3743 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3745 buffer
->size
= pCreateInfo
->size
;
3746 buffer
->usage
= pCreateInfo
->usage
;
3747 buffer
->address
= ANV_NULL_ADDRESS
;
3749 *pBuffer
= anv_buffer_to_handle(buffer
);
3754 void anv_DestroyBuffer(
3757 const VkAllocationCallbacks
* pAllocator
)
3759 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3760 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
3765 vk_free2(&device
->alloc
, pAllocator
, buffer
);
3768 VkDeviceAddress
anv_GetBufferDeviceAddressEXT(
3770 const VkBufferDeviceAddressInfoEXT
* pInfo
)
3772 ANV_FROM_HANDLE(anv_buffer
, buffer
, pInfo
->buffer
);
3774 assert(buffer
->address
.bo
->flags
& EXEC_OBJECT_PINNED
);
3776 return anv_address_physical(buffer
->address
);
3780 anv_fill_buffer_surface_state(struct anv_device
*device
, struct anv_state state
,
3781 enum isl_format format
,
3782 struct anv_address address
,
3783 uint32_t range
, uint32_t stride
)
3785 isl_buffer_fill_state(&device
->isl_dev
, state
.map
,
3786 .address
= anv_address_physical(address
),
3787 .mocs
= device
->default_mocs
,
3790 .swizzle
= ISL_SWIZZLE_IDENTITY
,
3791 .stride_B
= stride
);
3794 void anv_DestroySampler(
3797 const VkAllocationCallbacks
* pAllocator
)
3799 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3800 ANV_FROM_HANDLE(anv_sampler
, sampler
, _sampler
);
3805 if (sampler
->bindless_state
.map
) {
3806 anv_state_pool_free(&device
->dynamic_state_pool
,
3807 sampler
->bindless_state
);
3810 vk_free2(&device
->alloc
, pAllocator
, sampler
);
3813 VkResult
anv_CreateFramebuffer(
3815 const VkFramebufferCreateInfo
* pCreateInfo
,
3816 const VkAllocationCallbacks
* pAllocator
,
3817 VkFramebuffer
* pFramebuffer
)
3819 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3820 struct anv_framebuffer
*framebuffer
;
3822 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
3824 size_t size
= sizeof(*framebuffer
);
3826 /* VK_KHR_imageless_framebuffer extension says:
3828 * If flags includes VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR,
3829 * parameter pAttachments is ignored.
3831 if (!(pCreateInfo
->flags
& VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR
)) {
3832 size
+= sizeof(struct anv_image_view
*) * pCreateInfo
->attachmentCount
;
3833 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
3834 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3835 if (framebuffer
== NULL
)
3836 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3838 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
3839 ANV_FROM_HANDLE(anv_image_view
, iview
, pCreateInfo
->pAttachments
[i
]);
3840 framebuffer
->attachments
[i
] = iview
;
3842 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
3844 assert(device
->enabled_extensions
.KHR_imageless_framebuffer
);
3845 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
3846 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3847 if (framebuffer
== NULL
)
3848 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3850 framebuffer
->attachment_count
= 0;
3853 framebuffer
->width
= pCreateInfo
->width
;
3854 framebuffer
->height
= pCreateInfo
->height
;
3855 framebuffer
->layers
= pCreateInfo
->layers
;
3857 *pFramebuffer
= anv_framebuffer_to_handle(framebuffer
);
3862 void anv_DestroyFramebuffer(
3865 const VkAllocationCallbacks
* pAllocator
)
3867 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3868 ANV_FROM_HANDLE(anv_framebuffer
, fb
, _fb
);
3873 vk_free2(&device
->alloc
, pAllocator
, fb
);
3876 static const VkTimeDomainEXT anv_time_domains
[] = {
3877 VK_TIME_DOMAIN_DEVICE_EXT
,
3878 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
,
3879 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
,
3882 VkResult
anv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
3883 VkPhysicalDevice physicalDevice
,
3884 uint32_t *pTimeDomainCount
,
3885 VkTimeDomainEXT
*pTimeDomains
)
3888 VK_OUTARRAY_MAKE(out
, pTimeDomains
, pTimeDomainCount
);
3890 for (d
= 0; d
< ARRAY_SIZE(anv_time_domains
); d
++) {
3891 vk_outarray_append(&out
, i
) {
3892 *i
= anv_time_domains
[d
];
3896 return vk_outarray_status(&out
);
3900 anv_clock_gettime(clockid_t clock_id
)
3902 struct timespec current
;
3905 ret
= clock_gettime(clock_id
, ¤t
);
3906 if (ret
< 0 && clock_id
== CLOCK_MONOTONIC_RAW
)
3907 ret
= clock_gettime(CLOCK_MONOTONIC
, ¤t
);
3911 return (uint64_t) current
.tv_sec
* 1000000000ULL + current
.tv_nsec
;
3914 #define TIMESTAMP 0x2358
3916 VkResult
anv_GetCalibratedTimestampsEXT(
3918 uint32_t timestampCount
,
3919 const VkCalibratedTimestampInfoEXT
*pTimestampInfos
,
3920 uint64_t *pTimestamps
,
3921 uint64_t *pMaxDeviation
)
3923 ANV_FROM_HANDLE(anv_device
, device
, _device
);
3924 uint64_t timestamp_frequency
= device
->info
.timestamp_frequency
;
3927 uint64_t begin
, end
;
3928 uint64_t max_clock_period
= 0;
3930 begin
= anv_clock_gettime(CLOCK_MONOTONIC_RAW
);
3932 for (d
= 0; d
< timestampCount
; d
++) {
3933 switch (pTimestampInfos
[d
].timeDomain
) {
3934 case VK_TIME_DOMAIN_DEVICE_EXT
:
3935 ret
= anv_gem_reg_read(device
, TIMESTAMP
| 1,
3939 return anv_device_set_lost(device
, "Failed to read the TIMESTAMP "
3942 uint64_t device_period
= DIV_ROUND_UP(1000000000, timestamp_frequency
);
3943 max_clock_period
= MAX2(max_clock_period
, device_period
);
3945 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
:
3946 pTimestamps
[d
] = anv_clock_gettime(CLOCK_MONOTONIC
);
3947 max_clock_period
= MAX2(max_clock_period
, 1);
3950 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
:
3951 pTimestamps
[d
] = begin
;
3959 end
= anv_clock_gettime(CLOCK_MONOTONIC_RAW
);
3962 * The maximum deviation is the sum of the interval over which we
3963 * perform the sampling and the maximum period of any sampled
3964 * clock. That's because the maximum skew between any two sampled
3965 * clock edges is when the sampled clock with the largest period is
3966 * sampled at the end of that period but right at the beginning of the
3967 * sampling interval and some other clock is sampled right at the
3968 * begining of its sampling period and right at the end of the
3969 * sampling interval. Let's assume the GPU has the longest clock
3970 * period and that the application is sampling GPU and monotonic:
3973 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
3974 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
3978 * GPU -----_____-----_____-----_____-----_____
3981 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
3982 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
3984 * Interval <----------------->
3985 * Deviation <-------------------------->
3989 * m = read(monotonic) 2
3992 * We round the sample interval up by one tick to cover sampling error
3993 * in the interval clock
3996 uint64_t sample_interval
= end
- begin
+ 1;
3998 *pMaxDeviation
= sample_interval
+ max_clock_period
;
4003 /* vk_icd.h does not declare this function, so we declare it here to
4004 * suppress Wmissing-prototypes.
4006 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4007 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
);
4009 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4010 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
)
4012 /* For the full details on loader interface versioning, see
4013 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
4014 * What follows is a condensed summary, to help you navigate the large and
4015 * confusing official doc.
4017 * - Loader interface v0 is incompatible with later versions. We don't
4020 * - In loader interface v1:
4021 * - The first ICD entrypoint called by the loader is
4022 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
4024 * - The ICD must statically expose no other Vulkan symbol unless it is
4025 * linked with -Bsymbolic.
4026 * - Each dispatchable Vulkan handle created by the ICD must be
4027 * a pointer to a struct whose first member is VK_LOADER_DATA. The
4028 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
4029 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
4030 * vkDestroySurfaceKHR(). The ICD must be capable of working with
4031 * such loader-managed surfaces.
4033 * - Loader interface v2 differs from v1 in:
4034 * - The first ICD entrypoint called by the loader is
4035 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
4036 * statically expose this entrypoint.
4038 * - Loader interface v3 differs from v2 in:
4039 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
4040 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
4041 * because the loader no longer does so.
4043 *pSupportedVersion
= MIN2(*pSupportedVersion
, 3u);