2 * Copyright © 2016 Red Hat.
3 * Copyright © 2016 Bas Nieuwenhuizen
5 * based in part on anv driver which is:
6 * Copyright © 2015 Intel Corporation
8 * Permission is hereby granted, free of charge, to any person obtaining a
9 * copy of this software and associated documentation files (the "Software"),
10 * to deal in the Software without restriction, including without limitation
11 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
12 * and/or sell copies of the Software, and to permit persons to whom the
13 * Software is furnished to do so, subject to the following conditions:
15 * The above copyright notice and this permission notice (including the next
16 * paragraph) shall be included in all copies or substantial portions of the
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
22 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
23 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
24 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
31 #include <linux/audit.h>
32 #include <linux/bpf.h>
33 #include <linux/filter.h>
34 #include <linux/seccomp.h>
35 #include <linux/unistd.h>
36 #include <stdatomic.h>
41 #include <sys/prctl.h>
46 #include "radv_debug.h"
47 #include "radv_private.h"
48 #include "radv_shader.h"
50 #include "util/disk_cache.h"
54 #include "drm-uapi/amdgpu_drm.h"
55 #include "winsys/amdgpu/radv_amdgpu_winsys_public.h"
56 #include "winsys/null/radv_null_winsys_public.h"
57 #include "ac_llvm_util.h"
58 #include "vk_format.h"
61 #include "util/build_id.h"
62 #include "util/debug.h"
63 #include "util/mesa-sha1.h"
64 #include "util/timespec.h"
65 #include "util/u_atomic.h"
66 #include "compiler/glsl_types.h"
67 #include "util/driconf.h"
69 static struct radv_timeline_point
*
70 radv_timeline_find_point_at_least_locked(struct radv_device
*device
,
71 struct radv_timeline
*timeline
,
74 static struct radv_timeline_point
*
75 radv_timeline_add_point_locked(struct radv_device
*device
,
76 struct radv_timeline
*timeline
,
80 radv_timeline_trigger_waiters_locked(struct radv_timeline
*timeline
,
81 struct list_head
*processing_list
);
84 void radv_destroy_semaphore_part(struct radv_device
*device
,
85 struct radv_semaphore_part
*part
);
88 radv_create_pthread_cond(pthread_cond_t
*cond
);
90 uint64_t radv_get_current_time(void)
93 clock_gettime(CLOCK_MONOTONIC
, &tv
);
94 return tv
.tv_nsec
+ tv
.tv_sec
*1000000000ull;
97 static uint64_t radv_get_absolute_timeout(uint64_t timeout
)
99 uint64_t current_time
= radv_get_current_time();
101 timeout
= MIN2(UINT64_MAX
- current_time
, timeout
);
103 return current_time
+ timeout
;
107 radv_device_get_cache_uuid(enum radeon_family family
, void *uuid
)
109 struct mesa_sha1 ctx
;
110 unsigned char sha1
[20];
111 unsigned ptr_size
= sizeof(void*);
113 memset(uuid
, 0, VK_UUID_SIZE
);
114 _mesa_sha1_init(&ctx
);
116 if (!disk_cache_get_function_identifier(radv_device_get_cache_uuid
, &ctx
) ||
117 !disk_cache_get_function_identifier(LLVMInitializeAMDGPUTargetInfo
, &ctx
))
120 _mesa_sha1_update(&ctx
, &family
, sizeof(family
));
121 _mesa_sha1_update(&ctx
, &ptr_size
, sizeof(ptr_size
));
122 _mesa_sha1_final(&ctx
, sha1
);
124 memcpy(uuid
, sha1
, VK_UUID_SIZE
);
129 radv_get_driver_uuid(void *uuid
)
131 ac_compute_driver_uuid(uuid
, VK_UUID_SIZE
);
135 radv_get_device_uuid(struct radeon_info
*info
, void *uuid
)
137 ac_compute_device_uuid(info
, uuid
, VK_UUID_SIZE
);
141 radv_get_visible_vram_size(struct radv_physical_device
*device
)
143 return MIN2(device
->rad_info
.vram_size
, device
->rad_info
.vram_vis_size
);
147 radv_get_vram_size(struct radv_physical_device
*device
)
149 return device
->rad_info
.vram_size
- radv_get_visible_vram_size(device
);
153 radv_physical_device_init_mem_types(struct radv_physical_device
*device
)
155 uint64_t visible_vram_size
= radv_get_visible_vram_size(device
);
156 uint64_t vram_size
= radv_get_vram_size(device
);
157 int vram_index
= -1, visible_vram_index
= -1, gart_index
= -1;
158 device
->memory_properties
.memoryHeapCount
= 0;
160 vram_index
= device
->memory_properties
.memoryHeapCount
++;
161 device
->memory_properties
.memoryHeaps
[vram_index
] = (VkMemoryHeap
) {
163 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
167 if (device
->rad_info
.gart_size
> 0) {
168 gart_index
= device
->memory_properties
.memoryHeapCount
++;
169 device
->memory_properties
.memoryHeaps
[gart_index
] = (VkMemoryHeap
) {
170 .size
= device
->rad_info
.gart_size
,
175 if (visible_vram_size
) {
176 visible_vram_index
= device
->memory_properties
.memoryHeapCount
++;
177 device
->memory_properties
.memoryHeaps
[visible_vram_index
] = (VkMemoryHeap
) {
178 .size
= visible_vram_size
,
179 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
183 unsigned type_count
= 0;
185 if (vram_index
>= 0 || visible_vram_index
>= 0) {
186 device
->memory_domains
[type_count
] = RADEON_DOMAIN_VRAM
;
187 device
->memory_flags
[type_count
] = RADEON_FLAG_NO_CPU_ACCESS
;
188 device
->memory_properties
.memoryTypes
[type_count
++] = (VkMemoryType
) {
189 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
,
190 .heapIndex
= vram_index
>= 0 ? vram_index
: visible_vram_index
,
194 if (gart_index
>= 0) {
195 device
->memory_domains
[type_count
] = RADEON_DOMAIN_GTT
;
196 device
->memory_flags
[type_count
] = RADEON_FLAG_GTT_WC
| RADEON_FLAG_CPU_ACCESS
;
197 device
->memory_properties
.memoryTypes
[type_count
++] = (VkMemoryType
) {
198 .propertyFlags
= VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
199 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
200 .heapIndex
= gart_index
,
203 if (visible_vram_index
>= 0) {
204 device
->memory_domains
[type_count
] = RADEON_DOMAIN_VRAM
;
205 device
->memory_flags
[type_count
] = RADEON_FLAG_CPU_ACCESS
;
206 device
->memory_properties
.memoryTypes
[type_count
++] = (VkMemoryType
) {
207 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
208 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
209 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
210 .heapIndex
= visible_vram_index
,
214 if (gart_index
>= 0) {
215 device
->memory_domains
[type_count
] = RADEON_DOMAIN_GTT
;
216 device
->memory_flags
[type_count
] = RADEON_FLAG_CPU_ACCESS
;
217 device
->memory_properties
.memoryTypes
[type_count
++] = (VkMemoryType
) {
218 .propertyFlags
= VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
219 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
220 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
221 .heapIndex
= gart_index
,
224 device
->memory_properties
.memoryTypeCount
= type_count
;
226 if (device
->rad_info
.has_l2_uncached
) {
227 for (int i
= 0; i
< device
->memory_properties
.memoryTypeCount
; i
++) {
228 VkMemoryType mem_type
= device
->memory_properties
.memoryTypes
[i
];
230 if ((mem_type
.propertyFlags
& (VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
231 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
)) ||
232 mem_type
.propertyFlags
== VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
) {
234 VkMemoryPropertyFlags property_flags
= mem_type
.propertyFlags
|
235 VK_MEMORY_PROPERTY_DEVICE_COHERENT_BIT_AMD
|
236 VK_MEMORY_PROPERTY_DEVICE_UNCACHED_BIT_AMD
;
238 device
->memory_domains
[type_count
] = device
->memory_domains
[i
];
239 device
->memory_flags
[type_count
] = device
->memory_flags
[i
] | RADEON_FLAG_VA_UNCACHED
;
240 device
->memory_properties
.memoryTypes
[type_count
++] = (VkMemoryType
) {
241 .propertyFlags
= property_flags
,
242 .heapIndex
= mem_type
.heapIndex
,
246 device
->memory_properties
.memoryTypeCount
= type_count
;
251 radv_get_compiler_string(struct radv_physical_device
*pdevice
)
253 if (!pdevice
->use_llvm
) {
254 /* Some games like SotTR apply shader workarounds if the LLVM
255 * version is too old or if the LLVM version string is
256 * missing. This gives 2-5% performance with SotTR and ACO.
258 if (driQueryOptionb(&pdevice
->instance
->dri_options
,
259 "radv_report_llvm9_version_string")) {
260 return "ACO/LLVM 9.0.1";
266 return "LLVM " MESA_LLVM_VERSION_STRING
;
270 radv_physical_device_try_create(struct radv_instance
*instance
,
271 drmDevicePtr drm_device
,
272 struct radv_physical_device
**device_out
)
279 const char *path
= drm_device
->nodes
[DRM_NODE_RENDER
];
280 drmVersionPtr version
;
282 fd
= open(path
, O_RDWR
| O_CLOEXEC
);
284 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
285 radv_logi("Could not open device '%s'", path
);
287 return vk_error(instance
, VK_ERROR_INCOMPATIBLE_DRIVER
);
290 version
= drmGetVersion(fd
);
294 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
295 radv_logi("Could not get the kernel driver version for device '%s'", path
);
297 return vk_errorf(instance
, VK_ERROR_INCOMPATIBLE_DRIVER
,
298 "failed to get version %s: %m", path
);
301 if (strcmp(version
->name
, "amdgpu")) {
302 drmFreeVersion(version
);
305 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
306 radv_logi("Device '%s' is not using the amdgpu kernel driver.", path
);
308 return VK_ERROR_INCOMPATIBLE_DRIVER
;
310 drmFreeVersion(version
);
312 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
313 radv_logi("Found compatible device '%s'.", path
);
316 struct radv_physical_device
*device
=
317 vk_zalloc2(&instance
->alloc
, NULL
, sizeof(*device
), 8,
318 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
320 result
= vk_error(instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
324 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
325 device
->instance
= instance
;
328 device
->ws
= radv_amdgpu_winsys_create(fd
, instance
->debug_flags
,
329 instance
->perftest_flags
);
331 device
->ws
= radv_null_winsys_create();
335 result
= vk_errorf(instance
, VK_ERROR_INITIALIZATION_FAILED
,
336 "failed to initialize winsys");
340 if (drm_device
&& instance
->enabled_extensions
.KHR_display
) {
341 master_fd
= open(drm_device
->nodes
[DRM_NODE_PRIMARY
], O_RDWR
| O_CLOEXEC
);
342 if (master_fd
>= 0) {
343 uint32_t accel_working
= 0;
344 struct drm_amdgpu_info request
= {
345 .return_pointer
= (uintptr_t)&accel_working
,
346 .return_size
= sizeof(accel_working
),
347 .query
= AMDGPU_INFO_ACCEL_WORKING
350 if (drmCommandWrite(master_fd
, DRM_AMDGPU_INFO
, &request
, sizeof (struct drm_amdgpu_info
)) < 0 || !accel_working
) {
357 device
->master_fd
= master_fd
;
358 device
->local_fd
= fd
;
359 device
->ws
->query_info(device
->ws
, &device
->rad_info
);
361 device
->use_llvm
= instance
->debug_flags
& RADV_DEBUG_LLVM
;
363 snprintf(device
->name
, sizeof(device
->name
),
365 device
->rad_info
.name
, radv_get_compiler_string(device
));
367 if (radv_device_get_cache_uuid(device
->rad_info
.family
, device
->cache_uuid
)) {
368 result
= vk_errorf(instance
, VK_ERROR_INITIALIZATION_FAILED
,
369 "cannot generate UUID");
373 /* These flags affect shader compilation. */
374 uint64_t shader_env_flags
= (device
->use_llvm
? 0 : 0x2);
376 /* The gpu id is already embedded in the uuid so we just pass "radv"
377 * when creating the cache.
379 char buf
[VK_UUID_SIZE
* 2 + 1];
380 disk_cache_format_hex_id(buf
, device
->cache_uuid
, VK_UUID_SIZE
* 2);
381 device
->disk_cache
= disk_cache_create(device
->name
, buf
, shader_env_flags
);
383 if (device
->rad_info
.chip_class
< GFX8
|| !device
->use_llvm
)
384 fprintf(stderr
, "WARNING: radv is not a conformant vulkan implementation, testing use only.\n");
386 radv_get_driver_uuid(&device
->driver_uuid
);
387 radv_get_device_uuid(&device
->rad_info
, &device
->device_uuid
);
389 device
->out_of_order_rast_allowed
= device
->rad_info
.has_out_of_order_rast
&&
390 !(device
->instance
->debug_flags
& RADV_DEBUG_NO_OUT_OF_ORDER
);
392 device
->dcc_msaa_allowed
=
393 (device
->instance
->perftest_flags
& RADV_PERFTEST_DCC_MSAA
);
395 device
->use_ngg
= device
->rad_info
.chip_class
>= GFX10
&&
396 device
->rad_info
.family
!= CHIP_NAVI14
&&
397 !(device
->instance
->debug_flags
& RADV_DEBUG_NO_NGG
);
399 /* TODO: Implement NGG GS with ACO. */
400 device
->use_ngg_gs
= device
->use_ngg
&& device
->use_llvm
;
401 device
->use_ngg_streamout
= false;
403 /* Determine the number of threads per wave for all stages. */
404 device
->cs_wave_size
= 64;
405 device
->ps_wave_size
= 64;
406 device
->ge_wave_size
= 64;
408 if (device
->rad_info
.chip_class
>= GFX10
) {
409 if (device
->instance
->perftest_flags
& RADV_PERFTEST_CS_WAVE_32
)
410 device
->cs_wave_size
= 32;
412 /* For pixel shaders, wave64 is recommanded. */
413 if (device
->instance
->perftest_flags
& RADV_PERFTEST_PS_WAVE_32
)
414 device
->ps_wave_size
= 32;
416 if (device
->instance
->perftest_flags
& RADV_PERFTEST_GE_WAVE_32
)
417 device
->ge_wave_size
= 32;
420 radv_physical_device_init_mem_types(device
);
422 radv_physical_device_get_supported_extensions(device
,
423 &device
->supported_extensions
);
426 device
->bus_info
= *drm_device
->businfo
.pci
;
428 if ((device
->instance
->debug_flags
& RADV_DEBUG_INFO
))
429 ac_print_gpu_info(&device
->rad_info
);
431 /* The WSI is structured as a layer on top of the driver, so this has
432 * to be the last part of initialization (at least until we get other
435 result
= radv_init_wsi(device
);
436 if (result
!= VK_SUCCESS
) {
437 vk_error(instance
, result
);
438 goto fail_disk_cache
;
441 *device_out
= device
;
446 disk_cache_destroy(device
->disk_cache
);
448 device
->ws
->destroy(device
->ws
);
450 vk_free(&instance
->alloc
, device
);
460 radv_physical_device_destroy(struct radv_physical_device
*device
)
462 radv_finish_wsi(device
);
463 device
->ws
->destroy(device
->ws
);
464 disk_cache_destroy(device
->disk_cache
);
465 close(device
->local_fd
);
466 if (device
->master_fd
!= -1)
467 close(device
->master_fd
);
468 vk_free(&device
->instance
->alloc
, device
);
472 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
473 VkSystemAllocationScope allocationScope
)
479 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
480 size_t align
, VkSystemAllocationScope allocationScope
)
482 return realloc(pOriginal
, size
);
486 default_free_func(void *pUserData
, void *pMemory
)
491 static const VkAllocationCallbacks default_alloc
= {
493 .pfnAllocation
= default_alloc_func
,
494 .pfnReallocation
= default_realloc_func
,
495 .pfnFree
= default_free_func
,
498 static const struct debug_control radv_debug_options
[] = {
499 {"nofastclears", RADV_DEBUG_NO_FAST_CLEARS
},
500 {"nodcc", RADV_DEBUG_NO_DCC
},
501 {"shaders", RADV_DEBUG_DUMP_SHADERS
},
502 {"nocache", RADV_DEBUG_NO_CACHE
},
503 {"shaderstats", RADV_DEBUG_DUMP_SHADER_STATS
},
504 {"nohiz", RADV_DEBUG_NO_HIZ
},
505 {"nocompute", RADV_DEBUG_NO_COMPUTE_QUEUE
},
506 {"allbos", RADV_DEBUG_ALL_BOS
},
507 {"noibs", RADV_DEBUG_NO_IBS
},
508 {"spirv", RADV_DEBUG_DUMP_SPIRV
},
509 {"vmfaults", RADV_DEBUG_VM_FAULTS
},
510 {"zerovram", RADV_DEBUG_ZERO_VRAM
},
511 {"syncshaders", RADV_DEBUG_SYNC_SHADERS
},
512 {"preoptir", RADV_DEBUG_PREOPTIR
},
513 {"nodynamicbounds", RADV_DEBUG_NO_DYNAMIC_BOUNDS
},
514 {"nooutoforder", RADV_DEBUG_NO_OUT_OF_ORDER
},
515 {"info", RADV_DEBUG_INFO
},
516 {"errors", RADV_DEBUG_ERRORS
},
517 {"startup", RADV_DEBUG_STARTUP
},
518 {"checkir", RADV_DEBUG_CHECKIR
},
519 {"nothreadllvm", RADV_DEBUG_NOTHREADLLVM
},
520 {"nobinning", RADV_DEBUG_NOBINNING
},
521 {"nongg", RADV_DEBUG_NO_NGG
},
522 {"allentrypoints", RADV_DEBUG_ALL_ENTRYPOINTS
},
523 {"metashaders", RADV_DEBUG_DUMP_META_SHADERS
},
524 {"nomemorycache", RADV_DEBUG_NO_MEMORY_CACHE
},
525 {"llvm", RADV_DEBUG_LLVM
},
530 radv_get_debug_option_name(int id
)
532 assert(id
< ARRAY_SIZE(radv_debug_options
) - 1);
533 return radv_debug_options
[id
].string
;
536 static const struct debug_control radv_perftest_options
[] = {
537 {"localbos", RADV_PERFTEST_LOCAL_BOS
},
538 {"dccmsaa", RADV_PERFTEST_DCC_MSAA
},
539 {"bolist", RADV_PERFTEST_BO_LIST
},
540 {"tccompatcmask", RADV_PERFTEST_TC_COMPAT_CMASK
},
541 {"cswave32", RADV_PERFTEST_CS_WAVE_32
},
542 {"pswave32", RADV_PERFTEST_PS_WAVE_32
},
543 {"gewave32", RADV_PERFTEST_GE_WAVE_32
},
544 {"dfsm", RADV_PERFTEST_DFSM
},
549 radv_get_perftest_option_name(int id
)
551 assert(id
< ARRAY_SIZE(radv_perftest_options
) - 1);
552 return radv_perftest_options
[id
].string
;
556 radv_handle_per_app_options(struct radv_instance
*instance
,
557 const VkApplicationInfo
*info
)
559 const char *name
= info
? info
->pApplicationName
: NULL
;
560 const char *engine_name
= info
? info
->pEngineName
: NULL
;
563 if (!strcmp(name
, "DOOM_VFR")) {
564 /* Work around a Doom VFR game bug */
565 instance
->debug_flags
|= RADV_DEBUG_NO_DYNAMIC_BOUNDS
;
566 } else if (!strcmp(name
, "Fledge")) {
568 * Zero VRAM for "The Surge 2"
570 * This avoid a hang when when rendering any level. Likely
571 * uninitialized data in an indirect draw.
573 instance
->debug_flags
|= RADV_DEBUG_ZERO_VRAM
;
574 } else if (!strcmp(name
, "No Man's Sky")) {
575 /* Work around a NMS game bug */
576 instance
->debug_flags
|= RADV_DEBUG_DISCARD_TO_DEMOTE
;
577 } else if (!strcmp(name
, "DOOMEternal")) {
578 /* Zero VRAM for Doom Eternal to fix rendering issues. */
579 instance
->debug_flags
|= RADV_DEBUG_ZERO_VRAM
;
580 } else if (!strcmp(name
, "Red Dead Redemption 2")) {
581 /* Work around a RDR2 game bug */
582 instance
->debug_flags
|= RADV_DEBUG_DISCARD_TO_DEMOTE
;
587 if (!strcmp(engine_name
, "vkd3d")) {
588 /* Zero VRAM for all VKD3D (DX12->VK) games to fix
591 instance
->debug_flags
|= RADV_DEBUG_ZERO_VRAM
;
592 } else if (!strcmp(engine_name
, "Quantic Dream Engine")) {
593 /* Fix various artifacts in Detroit: Become Human */
594 instance
->debug_flags
|= RADV_DEBUG_ZERO_VRAM
|
595 RADV_DEBUG_DISCARD_TO_DEMOTE
;
599 instance
->enable_mrt_output_nan_fixup
=
600 driQueryOptionb(&instance
->dri_options
,
601 "radv_enable_mrt_output_nan_fixup");
603 if (driQueryOptionb(&instance
->dri_options
, "radv_no_dynamic_bounds"))
604 instance
->debug_flags
|= RADV_DEBUG_NO_DYNAMIC_BOUNDS
;
607 static const char radv_dri_options_xml
[] =
609 DRI_CONF_SECTION_PERFORMANCE
610 DRI_CONF_ADAPTIVE_SYNC("true")
611 DRI_CONF_VK_X11_OVERRIDE_MIN_IMAGE_COUNT(0)
612 DRI_CONF_VK_X11_STRICT_IMAGE_COUNT("false")
613 DRI_CONF_VK_X11_ENSURE_MIN_IMAGE_COUNT("false")
614 DRI_CONF_RADV_REPORT_LLVM9_VERSION_STRING("false")
615 DRI_CONF_RADV_ENABLE_MRT_OUTPUT_NAN_FIXUP("false")
616 DRI_CONF_RADV_NO_DYNAMIC_BOUNDS("false")
619 DRI_CONF_SECTION_DEBUG
620 DRI_CONF_VK_WSI_FORCE_BGRA8_UNORM_FIRST("false")
624 static void radv_init_dri_options(struct radv_instance
*instance
)
626 driParseOptionInfo(&instance
->available_dri_options
, radv_dri_options_xml
);
627 driParseConfigFiles(&instance
->dri_options
,
628 &instance
->available_dri_options
,
630 instance
->engineName
,
631 instance
->engineVersion
);
634 VkResult
radv_CreateInstance(
635 const VkInstanceCreateInfo
* pCreateInfo
,
636 const VkAllocationCallbacks
* pAllocator
,
637 VkInstance
* pInstance
)
639 struct radv_instance
*instance
;
642 instance
= vk_zalloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
643 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
645 return vk_error(NULL
, VK_ERROR_OUT_OF_HOST_MEMORY
);
647 vk_object_base_init(NULL
, &instance
->base
, VK_OBJECT_TYPE_INSTANCE
);
650 instance
->alloc
= *pAllocator
;
652 instance
->alloc
= default_alloc
;
654 if (pCreateInfo
->pApplicationInfo
) {
655 const VkApplicationInfo
*app
= pCreateInfo
->pApplicationInfo
;
657 instance
->engineName
=
658 vk_strdup(&instance
->alloc
, app
->pEngineName
,
659 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
660 instance
->engineVersion
= app
->engineVersion
;
661 instance
->apiVersion
= app
->apiVersion
;
664 if (instance
->apiVersion
== 0)
665 instance
->apiVersion
= VK_API_VERSION_1_0
;
667 instance
->debug_flags
= parse_debug_string(getenv("RADV_DEBUG"),
670 instance
->perftest_flags
= parse_debug_string(getenv("RADV_PERFTEST"),
671 radv_perftest_options
);
673 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
674 radv_logi("Created an instance");
676 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
678 for (idx
= 0; idx
< RADV_INSTANCE_EXTENSION_COUNT
; idx
++) {
679 if (!strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
680 radv_instance_extensions
[idx
].extensionName
))
684 if (idx
>= RADV_INSTANCE_EXTENSION_COUNT
||
685 !radv_instance_extensions_supported
.extensions
[idx
]) {
686 vk_object_base_finish(&instance
->base
);
687 vk_free2(&default_alloc
, pAllocator
, instance
);
688 return vk_error(instance
, VK_ERROR_EXTENSION_NOT_PRESENT
);
691 instance
->enabled_extensions
.extensions
[idx
] = true;
694 bool unchecked
= instance
->debug_flags
& RADV_DEBUG_ALL_ENTRYPOINTS
;
696 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->dispatch
.entrypoints
); i
++) {
697 /* Vulkan requires that entrypoints for extensions which have
698 * not been enabled must not be advertised.
701 !radv_instance_entrypoint_is_enabled(i
, instance
->apiVersion
,
702 &instance
->enabled_extensions
)) {
703 instance
->dispatch
.entrypoints
[i
] = NULL
;
705 instance
->dispatch
.entrypoints
[i
] =
706 radv_instance_dispatch_table
.entrypoints
[i
];
710 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->physical_device_dispatch
.entrypoints
); i
++) {
711 /* Vulkan requires that entrypoints for extensions which have
712 * not been enabled must not be advertised.
715 !radv_physical_device_entrypoint_is_enabled(i
, instance
->apiVersion
,
716 &instance
->enabled_extensions
)) {
717 instance
->physical_device_dispatch
.entrypoints
[i
] = NULL
;
719 instance
->physical_device_dispatch
.entrypoints
[i
] =
720 radv_physical_device_dispatch_table
.entrypoints
[i
];
724 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->device_dispatch
.entrypoints
); i
++) {
725 /* Vulkan requires that entrypoints for extensions which have
726 * not been enabled must not be advertised.
729 !radv_device_entrypoint_is_enabled(i
, instance
->apiVersion
,
730 &instance
->enabled_extensions
, NULL
)) {
731 instance
->device_dispatch
.entrypoints
[i
] = NULL
;
733 instance
->device_dispatch
.entrypoints
[i
] =
734 radv_device_dispatch_table
.entrypoints
[i
];
738 instance
->physical_devices_enumerated
= false;
739 list_inithead(&instance
->physical_devices
);
741 result
= vk_debug_report_instance_init(&instance
->debug_report_callbacks
);
742 if (result
!= VK_SUCCESS
) {
743 vk_object_base_finish(&instance
->base
);
744 vk_free2(&default_alloc
, pAllocator
, instance
);
745 return vk_error(instance
, result
);
748 glsl_type_singleton_init_or_ref();
750 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
752 radv_init_dri_options(instance
);
753 radv_handle_per_app_options(instance
, pCreateInfo
->pApplicationInfo
);
755 *pInstance
= radv_instance_to_handle(instance
);
760 void radv_DestroyInstance(
761 VkInstance _instance
,
762 const VkAllocationCallbacks
* pAllocator
)
764 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
769 list_for_each_entry_safe(struct radv_physical_device
, pdevice
,
770 &instance
->physical_devices
, link
) {
771 radv_physical_device_destroy(pdevice
);
774 vk_free(&instance
->alloc
, instance
->engineName
);
776 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
778 glsl_type_singleton_decref();
780 driDestroyOptionCache(&instance
->dri_options
);
781 driDestroyOptionInfo(&instance
->available_dri_options
);
783 vk_debug_report_instance_destroy(&instance
->debug_report_callbacks
);
785 vk_object_base_finish(&instance
->base
);
786 vk_free(&instance
->alloc
, instance
);
790 radv_enumerate_physical_devices(struct radv_instance
*instance
)
792 if (instance
->physical_devices_enumerated
)
795 instance
->physical_devices_enumerated
= true;
797 /* TODO: Check for more devices ? */
798 drmDevicePtr devices
[8];
799 VkResult result
= VK_SUCCESS
;
802 if (getenv("RADV_FORCE_FAMILY")) {
803 /* When RADV_FORCE_FAMILY is set, the driver creates a nul
804 * device that allows to test the compiler without having an
807 struct radv_physical_device
*pdevice
;
809 result
= radv_physical_device_try_create(instance
, NULL
, &pdevice
);
810 if (result
!= VK_SUCCESS
)
813 list_addtail(&pdevice
->link
, &instance
->physical_devices
);
817 max_devices
= drmGetDevices2(0, devices
, ARRAY_SIZE(devices
));
819 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
820 radv_logi("Found %d drm nodes", max_devices
);
823 return vk_error(instance
, VK_SUCCESS
);
825 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
826 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
827 devices
[i
]->bustype
== DRM_BUS_PCI
&&
828 devices
[i
]->deviceinfo
.pci
->vendor_id
== ATI_VENDOR_ID
) {
830 struct radv_physical_device
*pdevice
;
831 result
= radv_physical_device_try_create(instance
, devices
[i
],
833 /* Incompatible DRM device, skip. */
834 if (result
== VK_ERROR_INCOMPATIBLE_DRIVER
) {
839 /* Error creating the physical device, report the error. */
840 if (result
!= VK_SUCCESS
)
843 list_addtail(&pdevice
->link
, &instance
->physical_devices
);
846 drmFreeDevices(devices
, max_devices
);
848 /* If we successfully enumerated any devices, call it success */
852 VkResult
radv_EnumeratePhysicalDevices(
853 VkInstance _instance
,
854 uint32_t* pPhysicalDeviceCount
,
855 VkPhysicalDevice
* pPhysicalDevices
)
857 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
858 VK_OUTARRAY_MAKE(out
, pPhysicalDevices
, pPhysicalDeviceCount
);
860 VkResult result
= radv_enumerate_physical_devices(instance
);
861 if (result
!= VK_SUCCESS
)
864 list_for_each_entry(struct radv_physical_device
, pdevice
,
865 &instance
->physical_devices
, link
) {
866 vk_outarray_append(&out
, i
) {
867 *i
= radv_physical_device_to_handle(pdevice
);
871 return vk_outarray_status(&out
);
874 VkResult
radv_EnumeratePhysicalDeviceGroups(
875 VkInstance _instance
,
876 uint32_t* pPhysicalDeviceGroupCount
,
877 VkPhysicalDeviceGroupProperties
* pPhysicalDeviceGroupProperties
)
879 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
880 VK_OUTARRAY_MAKE(out
, pPhysicalDeviceGroupProperties
,
881 pPhysicalDeviceGroupCount
);
883 VkResult result
= radv_enumerate_physical_devices(instance
);
884 if (result
!= VK_SUCCESS
)
887 list_for_each_entry(struct radv_physical_device
, pdevice
,
888 &instance
->physical_devices
, link
) {
889 vk_outarray_append(&out
, p
) {
890 p
->physicalDeviceCount
= 1;
891 memset(p
->physicalDevices
, 0, sizeof(p
->physicalDevices
));
892 p
->physicalDevices
[0] = radv_physical_device_to_handle(pdevice
);
893 p
->subsetAllocation
= false;
897 return vk_outarray_status(&out
);
900 void radv_GetPhysicalDeviceFeatures(
901 VkPhysicalDevice physicalDevice
,
902 VkPhysicalDeviceFeatures
* pFeatures
)
904 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
905 memset(pFeatures
, 0, sizeof(*pFeatures
));
907 *pFeatures
= (VkPhysicalDeviceFeatures
) {
908 .robustBufferAccess
= true,
909 .fullDrawIndexUint32
= true,
910 .imageCubeArray
= true,
911 .independentBlend
= true,
912 .geometryShader
= true,
913 .tessellationShader
= true,
914 .sampleRateShading
= true,
915 .dualSrcBlend
= true,
917 .multiDrawIndirect
= true,
918 .drawIndirectFirstInstance
= true,
920 .depthBiasClamp
= true,
921 .fillModeNonSolid
= true,
926 .multiViewport
= true,
927 .samplerAnisotropy
= true,
928 .textureCompressionETC2
= radv_device_supports_etc(pdevice
),
929 .textureCompressionASTC_LDR
= false,
930 .textureCompressionBC
= true,
931 .occlusionQueryPrecise
= true,
932 .pipelineStatisticsQuery
= true,
933 .vertexPipelineStoresAndAtomics
= true,
934 .fragmentStoresAndAtomics
= true,
935 .shaderTessellationAndGeometryPointSize
= true,
936 .shaderImageGatherExtended
= true,
937 .shaderStorageImageExtendedFormats
= true,
938 .shaderStorageImageMultisample
= true,
939 .shaderUniformBufferArrayDynamicIndexing
= true,
940 .shaderSampledImageArrayDynamicIndexing
= true,
941 .shaderStorageBufferArrayDynamicIndexing
= true,
942 .shaderStorageImageArrayDynamicIndexing
= true,
943 .shaderStorageImageReadWithoutFormat
= true,
944 .shaderStorageImageWriteWithoutFormat
= true,
945 .shaderClipDistance
= true,
946 .shaderCullDistance
= true,
947 .shaderFloat64
= true,
950 .sparseBinding
= true,
951 .variableMultisampleRate
= true,
952 .shaderResourceMinLod
= true,
953 .inheritedQueries
= true,
958 radv_get_physical_device_features_1_1(struct radv_physical_device
*pdevice
,
959 VkPhysicalDeviceVulkan11Features
*f
)
961 assert(f
->sType
== VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES
);
963 f
->storageBuffer16BitAccess
= true;
964 f
->uniformAndStorageBuffer16BitAccess
= true;
965 f
->storagePushConstant16
= true;
966 f
->storageInputOutput16
= pdevice
->rad_info
.has_packed_math_16bit
&& (LLVM_VERSION_MAJOR
>= 9 || !pdevice
->use_llvm
);
968 f
->multiviewGeometryShader
= true;
969 f
->multiviewTessellationShader
= true;
970 f
->variablePointersStorageBuffer
= true;
971 f
->variablePointers
= true;
972 f
->protectedMemory
= false;
973 f
->samplerYcbcrConversion
= true;
974 f
->shaderDrawParameters
= true;
978 radv_get_physical_device_features_1_2(struct radv_physical_device
*pdevice
,
979 VkPhysicalDeviceVulkan12Features
*f
)
981 assert(f
->sType
== VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES
);
983 f
->samplerMirrorClampToEdge
= true;
984 f
->drawIndirectCount
= true;
985 f
->storageBuffer8BitAccess
= true;
986 f
->uniformAndStorageBuffer8BitAccess
= true;
987 f
->storagePushConstant8
= true;
988 f
->shaderBufferInt64Atomics
= LLVM_VERSION_MAJOR
>= 9 || !pdevice
->use_llvm
;
989 f
->shaderSharedInt64Atomics
= LLVM_VERSION_MAJOR
>= 9 || !pdevice
->use_llvm
;
990 f
->shaderFloat16
= pdevice
->rad_info
.has_packed_math_16bit
;
991 f
->shaderInt8
= true;
993 f
->descriptorIndexing
= true;
994 f
->shaderInputAttachmentArrayDynamicIndexing
= true;
995 f
->shaderUniformTexelBufferArrayDynamicIndexing
= true;
996 f
->shaderStorageTexelBufferArrayDynamicIndexing
= true;
997 f
->shaderUniformBufferArrayNonUniformIndexing
= true;
998 f
->shaderSampledImageArrayNonUniformIndexing
= true;
999 f
->shaderStorageBufferArrayNonUniformIndexing
= true;
1000 f
->shaderStorageImageArrayNonUniformIndexing
= true;
1001 f
->shaderInputAttachmentArrayNonUniformIndexing
= true;
1002 f
->shaderUniformTexelBufferArrayNonUniformIndexing
= true;
1003 f
->shaderStorageTexelBufferArrayNonUniformIndexing
= true;
1004 f
->descriptorBindingUniformBufferUpdateAfterBind
= true;
1005 f
->descriptorBindingSampledImageUpdateAfterBind
= true;
1006 f
->descriptorBindingStorageImageUpdateAfterBind
= true;
1007 f
->descriptorBindingStorageBufferUpdateAfterBind
= true;
1008 f
->descriptorBindingUniformTexelBufferUpdateAfterBind
= true;
1009 f
->descriptorBindingStorageTexelBufferUpdateAfterBind
= true;
1010 f
->descriptorBindingUpdateUnusedWhilePending
= true;
1011 f
->descriptorBindingPartiallyBound
= true;
1012 f
->descriptorBindingVariableDescriptorCount
= true;
1013 f
->runtimeDescriptorArray
= true;
1015 f
->samplerFilterMinmax
= true;
1016 f
->scalarBlockLayout
= pdevice
->rad_info
.chip_class
>= GFX7
;
1017 f
->imagelessFramebuffer
= true;
1018 f
->uniformBufferStandardLayout
= true;
1019 f
->shaderSubgroupExtendedTypes
= true;
1020 f
->separateDepthStencilLayouts
= true;
1021 f
->hostQueryReset
= true;
1022 f
->timelineSemaphore
= pdevice
->rad_info
.has_syncobj_wait_for_submit
;
1023 f
->bufferDeviceAddress
= true;
1024 f
->bufferDeviceAddressCaptureReplay
= false;
1025 f
->bufferDeviceAddressMultiDevice
= false;
1026 f
->vulkanMemoryModel
= false;
1027 f
->vulkanMemoryModelDeviceScope
= false;
1028 f
->vulkanMemoryModelAvailabilityVisibilityChains
= false;
1029 f
->shaderOutputViewportIndex
= true;
1030 f
->shaderOutputLayer
= true;
1031 f
->subgroupBroadcastDynamicId
= true;
1034 void radv_GetPhysicalDeviceFeatures2(
1035 VkPhysicalDevice physicalDevice
,
1036 VkPhysicalDeviceFeatures2
*pFeatures
)
1038 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
1039 radv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
1041 VkPhysicalDeviceVulkan11Features core_1_1
= {
1042 .sType
= VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES
,
1044 radv_get_physical_device_features_1_1(pdevice
, &core_1_1
);
1046 VkPhysicalDeviceVulkan12Features core_1_2
= {
1047 .sType
= VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES
,
1049 radv_get_physical_device_features_1_2(pdevice
, &core_1_2
);
1051 #define CORE_FEATURE(major, minor, feature) \
1052 features->feature = core_##major##_##minor.feature
1054 vk_foreach_struct(ext
, pFeatures
->pNext
) {
1055 switch (ext
->sType
) {
1056 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES
: {
1057 VkPhysicalDeviceVariablePointersFeatures
*features
= (void *)ext
;
1058 CORE_FEATURE(1, 1, variablePointersStorageBuffer
);
1059 CORE_FEATURE(1, 1, variablePointers
);
1062 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES
: {
1063 VkPhysicalDeviceMultiviewFeatures
*features
= (VkPhysicalDeviceMultiviewFeatures
*)ext
;
1064 CORE_FEATURE(1, 1, multiview
);
1065 CORE_FEATURE(1, 1, multiviewGeometryShader
);
1066 CORE_FEATURE(1, 1, multiviewTessellationShader
);
1069 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETERS_FEATURES
: {
1070 VkPhysicalDeviceShaderDrawParametersFeatures
*features
=
1071 (VkPhysicalDeviceShaderDrawParametersFeatures
*)ext
;
1072 CORE_FEATURE(1, 1, shaderDrawParameters
);
1075 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES
: {
1076 VkPhysicalDeviceProtectedMemoryFeatures
*features
=
1077 (VkPhysicalDeviceProtectedMemoryFeatures
*)ext
;
1078 CORE_FEATURE(1, 1, protectedMemory
);
1081 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES
: {
1082 VkPhysicalDevice16BitStorageFeatures
*features
=
1083 (VkPhysicalDevice16BitStorageFeatures
*)ext
;
1084 CORE_FEATURE(1, 1, storageBuffer16BitAccess
);
1085 CORE_FEATURE(1, 1, uniformAndStorageBuffer16BitAccess
);
1086 CORE_FEATURE(1, 1, storagePushConstant16
);
1087 CORE_FEATURE(1, 1, storageInputOutput16
);
1090 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES
: {
1091 VkPhysicalDeviceSamplerYcbcrConversionFeatures
*features
=
1092 (VkPhysicalDeviceSamplerYcbcrConversionFeatures
*)ext
;
1093 CORE_FEATURE(1, 1, samplerYcbcrConversion
);
1096 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES
: {
1097 VkPhysicalDeviceDescriptorIndexingFeatures
*features
=
1098 (VkPhysicalDeviceDescriptorIndexingFeatures
*)ext
;
1099 CORE_FEATURE(1, 2, shaderInputAttachmentArrayDynamicIndexing
);
1100 CORE_FEATURE(1, 2, shaderUniformTexelBufferArrayDynamicIndexing
);
1101 CORE_FEATURE(1, 2, shaderStorageTexelBufferArrayDynamicIndexing
);
1102 CORE_FEATURE(1, 2, shaderUniformBufferArrayNonUniformIndexing
);
1103 CORE_FEATURE(1, 2, shaderSampledImageArrayNonUniformIndexing
);
1104 CORE_FEATURE(1, 2, shaderStorageBufferArrayNonUniformIndexing
);
1105 CORE_FEATURE(1, 2, shaderStorageImageArrayNonUniformIndexing
);
1106 CORE_FEATURE(1, 2, shaderInputAttachmentArrayNonUniformIndexing
);
1107 CORE_FEATURE(1, 2, shaderUniformTexelBufferArrayNonUniformIndexing
);
1108 CORE_FEATURE(1, 2, shaderStorageTexelBufferArrayNonUniformIndexing
);
1109 CORE_FEATURE(1, 2, descriptorBindingUniformBufferUpdateAfterBind
);
1110 CORE_FEATURE(1, 2, descriptorBindingSampledImageUpdateAfterBind
);
1111 CORE_FEATURE(1, 2, descriptorBindingStorageImageUpdateAfterBind
);
1112 CORE_FEATURE(1, 2, descriptorBindingStorageBufferUpdateAfterBind
);
1113 CORE_FEATURE(1, 2, descriptorBindingUniformTexelBufferUpdateAfterBind
);
1114 CORE_FEATURE(1, 2, descriptorBindingStorageTexelBufferUpdateAfterBind
);
1115 CORE_FEATURE(1, 2, descriptorBindingUpdateUnusedWhilePending
);
1116 CORE_FEATURE(1, 2, descriptorBindingPartiallyBound
);
1117 CORE_FEATURE(1, 2, descriptorBindingVariableDescriptorCount
);
1118 CORE_FEATURE(1, 2, runtimeDescriptorArray
);
1121 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT
: {
1122 VkPhysicalDeviceConditionalRenderingFeaturesEXT
*features
=
1123 (VkPhysicalDeviceConditionalRenderingFeaturesEXT
*)ext
;
1124 features
->conditionalRendering
= true;
1125 features
->inheritedConditionalRendering
= false;
1128 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT
: {
1129 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*features
=
1130 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*)ext
;
1131 features
->vertexAttributeInstanceRateDivisor
= true;
1132 features
->vertexAttributeInstanceRateZeroDivisor
= true;
1135 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT
: {
1136 VkPhysicalDeviceTransformFeedbackFeaturesEXT
*features
=
1137 (VkPhysicalDeviceTransformFeedbackFeaturesEXT
*)ext
;
1138 features
->transformFeedback
= true;
1139 features
->geometryStreams
= !pdevice
->use_ngg_streamout
;
1142 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES
: {
1143 VkPhysicalDeviceScalarBlockLayoutFeatures
*features
=
1144 (VkPhysicalDeviceScalarBlockLayoutFeatures
*)ext
;
1145 CORE_FEATURE(1, 2, scalarBlockLayout
);
1148 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_PRIORITY_FEATURES_EXT
: {
1149 VkPhysicalDeviceMemoryPriorityFeaturesEXT
*features
=
1150 (VkPhysicalDeviceMemoryPriorityFeaturesEXT
*)ext
;
1151 features
->memoryPriority
= true;
1154 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_EXT
: {
1155 VkPhysicalDeviceBufferDeviceAddressFeaturesEXT
*features
=
1156 (VkPhysicalDeviceBufferDeviceAddressFeaturesEXT
*)ext
;
1157 features
->bufferDeviceAddress
= true;
1158 features
->bufferDeviceAddressCaptureReplay
= false;
1159 features
->bufferDeviceAddressMultiDevice
= false;
1162 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES
: {
1163 VkPhysicalDeviceBufferDeviceAddressFeatures
*features
=
1164 (VkPhysicalDeviceBufferDeviceAddressFeatures
*)ext
;
1165 CORE_FEATURE(1, 2, bufferDeviceAddress
);
1166 CORE_FEATURE(1, 2, bufferDeviceAddressCaptureReplay
);
1167 CORE_FEATURE(1, 2, bufferDeviceAddressMultiDevice
);
1170 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_CLIP_ENABLE_FEATURES_EXT
: {
1171 VkPhysicalDeviceDepthClipEnableFeaturesEXT
*features
=
1172 (VkPhysicalDeviceDepthClipEnableFeaturesEXT
*)ext
;
1173 features
->depthClipEnable
= true;
1176 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_QUERY_RESET_FEATURES
: {
1177 VkPhysicalDeviceHostQueryResetFeatures
*features
=
1178 (VkPhysicalDeviceHostQueryResetFeatures
*)ext
;
1179 CORE_FEATURE(1, 2, hostQueryReset
);
1182 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES
: {
1183 VkPhysicalDevice8BitStorageFeatures
*features
=
1184 (VkPhysicalDevice8BitStorageFeatures
*)ext
;
1185 CORE_FEATURE(1, 2, storageBuffer8BitAccess
);
1186 CORE_FEATURE(1, 2, uniformAndStorageBuffer8BitAccess
);
1187 CORE_FEATURE(1, 2, storagePushConstant8
);
1190 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_FLOAT16_INT8_FEATURES
: {
1191 VkPhysicalDeviceShaderFloat16Int8Features
*features
=
1192 (VkPhysicalDeviceShaderFloat16Int8Features
*)ext
;
1193 CORE_FEATURE(1, 2, shaderFloat16
);
1194 CORE_FEATURE(1, 2, shaderInt8
);
1197 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_INT64_FEATURES
: {
1198 VkPhysicalDeviceShaderAtomicInt64Features
*features
=
1199 (VkPhysicalDeviceShaderAtomicInt64Features
*)ext
;
1200 CORE_FEATURE(1, 2, shaderBufferInt64Atomics
);
1201 CORE_FEATURE(1, 2, shaderSharedInt64Atomics
);
1204 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DEMOTE_TO_HELPER_INVOCATION_FEATURES_EXT
: {
1205 VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT
*features
=
1206 (VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT
*)ext
;
1207 features
->shaderDemoteToHelperInvocation
= LLVM_VERSION_MAJOR
>= 9 || !pdevice
->use_llvm
;
1210 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_FEATURES_EXT
: {
1211 VkPhysicalDeviceInlineUniformBlockFeaturesEXT
*features
=
1212 (VkPhysicalDeviceInlineUniformBlockFeaturesEXT
*)ext
;
1214 features
->inlineUniformBlock
= true;
1215 features
->descriptorBindingInlineUniformBlockUpdateAfterBind
= true;
1218 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COMPUTE_SHADER_DERIVATIVES_FEATURES_NV
: {
1219 VkPhysicalDeviceComputeShaderDerivativesFeaturesNV
*features
=
1220 (VkPhysicalDeviceComputeShaderDerivativesFeaturesNV
*)ext
;
1221 features
->computeDerivativeGroupQuads
= false;
1222 features
->computeDerivativeGroupLinear
= true;
1225 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_YCBCR_IMAGE_ARRAYS_FEATURES_EXT
: {
1226 VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*features
=
1227 (VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*)ext
;
1228 features
->ycbcrImageArrays
= true;
1231 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_UNIFORM_BUFFER_STANDARD_LAYOUT_FEATURES
: {
1232 VkPhysicalDeviceUniformBufferStandardLayoutFeatures
*features
=
1233 (VkPhysicalDeviceUniformBufferStandardLayoutFeatures
*)ext
;
1234 CORE_FEATURE(1, 2, uniformBufferStandardLayout
);
1237 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INDEX_TYPE_UINT8_FEATURES_EXT
: {
1238 VkPhysicalDeviceIndexTypeUint8FeaturesEXT
*features
=
1239 (VkPhysicalDeviceIndexTypeUint8FeaturesEXT
*)ext
;
1240 features
->indexTypeUint8
= pdevice
->rad_info
.chip_class
>= GFX8
;
1243 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGELESS_FRAMEBUFFER_FEATURES
: {
1244 VkPhysicalDeviceImagelessFramebufferFeatures
*features
=
1245 (VkPhysicalDeviceImagelessFramebufferFeatures
*)ext
;
1246 CORE_FEATURE(1, 2, imagelessFramebuffer
);
1249 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR
: {
1250 VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR
*features
=
1251 (VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR
*)ext
;
1252 features
->pipelineExecutableInfo
= true;
1255 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CLOCK_FEATURES_KHR
: {
1256 VkPhysicalDeviceShaderClockFeaturesKHR
*features
=
1257 (VkPhysicalDeviceShaderClockFeaturesKHR
*)ext
;
1258 features
->shaderSubgroupClock
= true;
1259 features
->shaderDeviceClock
= pdevice
->rad_info
.chip_class
>= GFX8
;
1262 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_FEATURES_EXT
: {
1263 VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT
*features
=
1264 (VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT
*)ext
;
1265 features
->texelBufferAlignment
= true;
1268 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_FEATURES
: {
1269 VkPhysicalDeviceTimelineSemaphoreFeatures
*features
=
1270 (VkPhysicalDeviceTimelineSemaphoreFeatures
*) ext
;
1271 CORE_FEATURE(1, 2, timelineSemaphore
);
1274 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_FEATURES_EXT
: {
1275 VkPhysicalDeviceSubgroupSizeControlFeaturesEXT
*features
=
1276 (VkPhysicalDeviceSubgroupSizeControlFeaturesEXT
*)ext
;
1277 features
->subgroupSizeControl
= true;
1278 features
->computeFullSubgroups
= true;
1281 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COHERENT_MEMORY_FEATURES_AMD
: {
1282 VkPhysicalDeviceCoherentMemoryFeaturesAMD
*features
=
1283 (VkPhysicalDeviceCoherentMemoryFeaturesAMD
*)ext
;
1284 features
->deviceCoherentMemory
= pdevice
->rad_info
.has_l2_uncached
;
1287 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_SUBGROUP_EXTENDED_TYPES_FEATURES
: {
1288 VkPhysicalDeviceShaderSubgroupExtendedTypesFeatures
*features
=
1289 (VkPhysicalDeviceShaderSubgroupExtendedTypesFeatures
*)ext
;
1290 CORE_FEATURE(1, 2, shaderSubgroupExtendedTypes
);
1293 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SEPARATE_DEPTH_STENCIL_LAYOUTS_FEATURES_KHR
: {
1294 VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR
*features
=
1295 (VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR
*)ext
;
1296 CORE_FEATURE(1, 2, separateDepthStencilLayouts
);
1299 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES
: {
1300 radv_get_physical_device_features_1_1(pdevice
, (void *)ext
);
1303 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES
: {
1304 radv_get_physical_device_features_1_2(pdevice
, (void *)ext
);
1307 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_FEATURES_EXT
: {
1308 VkPhysicalDeviceLineRasterizationFeaturesEXT
*features
=
1309 (VkPhysicalDeviceLineRasterizationFeaturesEXT
*)ext
;
1310 features
->rectangularLines
= false;
1311 features
->bresenhamLines
= true;
1312 features
->smoothLines
= false;
1313 features
->stippledRectangularLines
= false;
1314 features
->stippledBresenhamLines
= true;
1315 features
->stippledSmoothLines
= false;
1318 case VK_STRUCTURE_TYPE_DEVICE_MEMORY_OVERALLOCATION_CREATE_INFO_AMD
: {
1319 VkDeviceMemoryOverallocationCreateInfoAMD
*features
=
1320 (VkDeviceMemoryOverallocationCreateInfoAMD
*)ext
;
1321 features
->overallocationBehavior
= true;
1324 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ROBUSTNESS_2_FEATURES_EXT
: {
1325 VkPhysicalDeviceRobustness2FeaturesEXT
*features
=
1326 (VkPhysicalDeviceRobustness2FeaturesEXT
*)ext
;
1327 features
->robustBufferAccess2
= true;
1328 features
->robustImageAccess2
= true;
1329 features
->nullDescriptor
= true;
1332 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CUSTOM_BORDER_COLOR_FEATURES_EXT
: {
1333 VkPhysicalDeviceCustomBorderColorFeaturesEXT
*features
=
1334 (VkPhysicalDeviceCustomBorderColorFeaturesEXT
*)ext
;
1335 features
->customBorderColors
= true;
1336 features
->customBorderColorWithoutFormat
= true;
1339 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PRIVATE_DATA_FEATURES_EXT
: {
1340 VkPhysicalDevicePrivateDataFeaturesEXT
*features
=
1341 (VkPhysicalDevicePrivateDataFeaturesEXT
*)ext
;
1342 features
->privateData
= true;
1345 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_CREATION_CACHE_CONTROL_FEATURES_EXT
: {
1346 VkPhysicalDevicePipelineCreationCacheControlFeaturesEXT
*features
=
1347 (VkPhysicalDevicePipelineCreationCacheControlFeaturesEXT
*)ext
;
1348 features
-> pipelineCreationCacheControl
= true;
1351 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTENDED_DYNAMIC_STATE_FEATURES_EXT
: {
1352 VkPhysicalDeviceExtendedDynamicStateFeaturesEXT
*features
=
1353 (VkPhysicalDeviceExtendedDynamicStateFeaturesEXT
*) ext
;
1354 features
->extendedDynamicState
= true;
1357 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGE_ROBUSTNESS_FEATURES_EXT
: {
1358 VkPhysicalDeviceImageRobustnessFeaturesEXT
*features
=
1359 (VkPhysicalDeviceImageRobustnessFeaturesEXT
*)ext
;
1360 features
->robustImageAccess
= true;
1363 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_FLOAT_FEATURES_EXT
: {
1364 VkPhysicalDeviceShaderAtomicFloatFeaturesEXT
*features
=
1365 (VkPhysicalDeviceShaderAtomicFloatFeaturesEXT
*)ext
;
1366 features
->shaderBufferFloat32Atomics
= true;
1367 features
->shaderBufferFloat32AtomicAdd
= false;
1368 features
->shaderBufferFloat64Atomics
= true;
1369 features
->shaderBufferFloat64AtomicAdd
= false;
1370 features
->shaderSharedFloat32Atomics
= true;
1371 features
->shaderSharedFloat32AtomicAdd
= pdevice
->rad_info
.chip_class
>= GFX8
&&
1372 (!pdevice
->use_llvm
|| LLVM_VERSION_MAJOR
>= 10);
1373 features
->shaderSharedFloat64Atomics
= true;
1374 features
->shaderSharedFloat64AtomicAdd
= false;
1375 features
->shaderImageFloat32Atomics
= true;
1376 features
->shaderImageFloat32AtomicAdd
= false;
1377 features
->sparseImageFloat32Atomics
= false;
1378 features
->sparseImageFloat32AtomicAdd
= false;
1381 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_4444_FORMATS_FEATURES_EXT
: {
1382 VkPhysicalDevice4444FormatsFeaturesEXT
*features
=
1383 (VkPhysicalDevice4444FormatsFeaturesEXT
*)ext
;
1384 features
->formatA4R4G4B4
= true;
1385 features
->formatA4B4G4R4
= true;
1396 radv_max_descriptor_set_size()
1398 /* make sure that the entire descriptor set is addressable with a signed
1399 * 32-bit int. So the sum of all limits scaled by descriptor size has to
1400 * be at most 2 GiB. the combined image & samples object count as one of
1401 * both. This limit is for the pipeline layout, not for the set layout, but
1402 * there is no set limit, so we just set a pipeline limit. I don't think
1403 * any app is going to hit this soon. */
1404 return ((1ull << 31) - 16 * MAX_DYNAMIC_BUFFERS
1405 - MAX_INLINE_UNIFORM_BLOCK_SIZE
* MAX_INLINE_UNIFORM_BLOCK_COUNT
) /
1406 (32 /* uniform buffer, 32 due to potential space wasted on alignment */ +
1407 32 /* storage buffer, 32 due to potential space wasted on alignment */ +
1408 32 /* sampler, largest when combined with image */ +
1409 64 /* sampled image */ +
1410 64 /* storage image */);
1413 void radv_GetPhysicalDeviceProperties(
1414 VkPhysicalDevice physicalDevice
,
1415 VkPhysicalDeviceProperties
* pProperties
)
1417 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
1418 VkSampleCountFlags sample_counts
= 0xf;
1420 size_t max_descriptor_set_size
= radv_max_descriptor_set_size();
1422 VkPhysicalDeviceLimits limits
= {
1423 .maxImageDimension1D
= (1 << 14),
1424 .maxImageDimension2D
= (1 << 14),
1425 .maxImageDimension3D
= (1 << 11),
1426 .maxImageDimensionCube
= (1 << 14),
1427 .maxImageArrayLayers
= (1 << 11),
1428 .maxTexelBufferElements
= UINT32_MAX
,
1429 .maxUniformBufferRange
= UINT32_MAX
,
1430 .maxStorageBufferRange
= UINT32_MAX
,
1431 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
1432 .maxMemoryAllocationCount
= UINT32_MAX
,
1433 .maxSamplerAllocationCount
= 64 * 1024,
1434 .bufferImageGranularity
= 64, /* A cache line */
1435 .sparseAddressSpaceSize
= RADV_MAX_MEMORY_ALLOCATION_SIZE
, /* buffer max size */
1436 .maxBoundDescriptorSets
= MAX_SETS
,
1437 .maxPerStageDescriptorSamplers
= max_descriptor_set_size
,
1438 .maxPerStageDescriptorUniformBuffers
= max_descriptor_set_size
,
1439 .maxPerStageDescriptorStorageBuffers
= max_descriptor_set_size
,
1440 .maxPerStageDescriptorSampledImages
= max_descriptor_set_size
,
1441 .maxPerStageDescriptorStorageImages
= max_descriptor_set_size
,
1442 .maxPerStageDescriptorInputAttachments
= max_descriptor_set_size
,
1443 .maxPerStageResources
= max_descriptor_set_size
,
1444 .maxDescriptorSetSamplers
= max_descriptor_set_size
,
1445 .maxDescriptorSetUniformBuffers
= max_descriptor_set_size
,
1446 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_UNIFORM_BUFFERS
,
1447 .maxDescriptorSetStorageBuffers
= max_descriptor_set_size
,
1448 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_STORAGE_BUFFERS
,
1449 .maxDescriptorSetSampledImages
= max_descriptor_set_size
,
1450 .maxDescriptorSetStorageImages
= max_descriptor_set_size
,
1451 .maxDescriptorSetInputAttachments
= max_descriptor_set_size
,
1452 .maxVertexInputAttributes
= MAX_VERTEX_ATTRIBS
,
1453 .maxVertexInputBindings
= MAX_VBS
,
1454 .maxVertexInputAttributeOffset
= 2047,
1455 .maxVertexInputBindingStride
= 2048,
1456 .maxVertexOutputComponents
= 128,
1457 .maxTessellationGenerationLevel
= 64,
1458 .maxTessellationPatchSize
= 32,
1459 .maxTessellationControlPerVertexInputComponents
= 128,
1460 .maxTessellationControlPerVertexOutputComponents
= 128,
1461 .maxTessellationControlPerPatchOutputComponents
= 120,
1462 .maxTessellationControlTotalOutputComponents
= 4096,
1463 .maxTessellationEvaluationInputComponents
= 128,
1464 .maxTessellationEvaluationOutputComponents
= 128,
1465 .maxGeometryShaderInvocations
= 127,
1466 .maxGeometryInputComponents
= 64,
1467 .maxGeometryOutputComponents
= 128,
1468 .maxGeometryOutputVertices
= 256,
1469 .maxGeometryTotalOutputComponents
= 1024,
1470 .maxFragmentInputComponents
= 128,
1471 .maxFragmentOutputAttachments
= 8,
1472 .maxFragmentDualSrcAttachments
= 1,
1473 .maxFragmentCombinedOutputResources
= 8,
1474 .maxComputeSharedMemorySize
= 32768,
1475 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
1476 .maxComputeWorkGroupInvocations
= 1024,
1477 .maxComputeWorkGroupSize
= {
1482 .subPixelPrecisionBits
= 8,
1483 .subTexelPrecisionBits
= 8,
1484 .mipmapPrecisionBits
= 8,
1485 .maxDrawIndexedIndexValue
= UINT32_MAX
,
1486 .maxDrawIndirectCount
= UINT32_MAX
,
1487 .maxSamplerLodBias
= 16,
1488 .maxSamplerAnisotropy
= 16,
1489 .maxViewports
= MAX_VIEWPORTS
,
1490 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
1491 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
1492 .viewportSubPixelBits
= 8,
1493 .minMemoryMapAlignment
= 4096, /* A page */
1494 .minTexelBufferOffsetAlignment
= 4,
1495 .minUniformBufferOffsetAlignment
= 4,
1496 .minStorageBufferOffsetAlignment
= 4,
1497 .minTexelOffset
= -32,
1498 .maxTexelOffset
= 31,
1499 .minTexelGatherOffset
= -32,
1500 .maxTexelGatherOffset
= 31,
1501 .minInterpolationOffset
= -2,
1502 .maxInterpolationOffset
= 2,
1503 .subPixelInterpolationOffsetBits
= 8,
1504 .maxFramebufferWidth
= (1 << 14),
1505 .maxFramebufferHeight
= (1 << 14),
1506 .maxFramebufferLayers
= (1 << 10),
1507 .framebufferColorSampleCounts
= sample_counts
,
1508 .framebufferDepthSampleCounts
= sample_counts
,
1509 .framebufferStencilSampleCounts
= sample_counts
,
1510 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
1511 .maxColorAttachments
= MAX_RTS
,
1512 .sampledImageColorSampleCounts
= sample_counts
,
1513 .sampledImageIntegerSampleCounts
= sample_counts
,
1514 .sampledImageDepthSampleCounts
= sample_counts
,
1515 .sampledImageStencilSampleCounts
= sample_counts
,
1516 .storageImageSampleCounts
= sample_counts
,
1517 .maxSampleMaskWords
= 1,
1518 .timestampComputeAndGraphics
= true,
1519 .timestampPeriod
= 1000000.0 / pdevice
->rad_info
.clock_crystal_freq
,
1520 .maxClipDistances
= 8,
1521 .maxCullDistances
= 8,
1522 .maxCombinedClipAndCullDistances
= 8,
1523 .discreteQueuePriorities
= 2,
1524 .pointSizeRange
= { 0.0, 8191.875 },
1525 .lineWidthRange
= { 0.0, 8191.875 },
1526 .pointSizeGranularity
= (1.0 / 8.0),
1527 .lineWidthGranularity
= (1.0 / 8.0),
1528 .strictLines
= false, /* FINISHME */
1529 .standardSampleLocations
= true,
1530 .optimalBufferCopyOffsetAlignment
= 128,
1531 .optimalBufferCopyRowPitchAlignment
= 128,
1532 .nonCoherentAtomSize
= 64,
1535 *pProperties
= (VkPhysicalDeviceProperties
) {
1536 .apiVersion
= radv_physical_device_api_version(pdevice
),
1537 .driverVersion
= vk_get_driver_version(),
1538 .vendorID
= ATI_VENDOR_ID
,
1539 .deviceID
= pdevice
->rad_info
.pci_id
,
1540 .deviceType
= pdevice
->rad_info
.has_dedicated_vram
? VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU
: VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
1542 .sparseProperties
= {0},
1545 strcpy(pProperties
->deviceName
, pdevice
->name
);
1546 memcpy(pProperties
->pipelineCacheUUID
, pdevice
->cache_uuid
, VK_UUID_SIZE
);
1550 radv_get_physical_device_properties_1_1(struct radv_physical_device
*pdevice
,
1551 VkPhysicalDeviceVulkan11Properties
*p
)
1553 assert(p
->sType
== VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES
);
1555 memcpy(p
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
1556 memcpy(p
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
1557 memset(p
->deviceLUID
, 0, VK_LUID_SIZE
);
1558 /* The LUID is for Windows. */
1559 p
->deviceLUIDValid
= false;
1560 p
->deviceNodeMask
= 0;
1562 p
->subgroupSize
= RADV_SUBGROUP_SIZE
;
1563 p
->subgroupSupportedStages
= VK_SHADER_STAGE_ALL_GRAPHICS
|
1564 VK_SHADER_STAGE_COMPUTE_BIT
;
1565 p
->subgroupSupportedOperations
= VK_SUBGROUP_FEATURE_BASIC_BIT
|
1566 VK_SUBGROUP_FEATURE_VOTE_BIT
|
1567 VK_SUBGROUP_FEATURE_ARITHMETIC_BIT
|
1568 VK_SUBGROUP_FEATURE_BALLOT_BIT
|
1569 VK_SUBGROUP_FEATURE_CLUSTERED_BIT
|
1570 VK_SUBGROUP_FEATURE_QUAD_BIT
|
1571 VK_SUBGROUP_FEATURE_SHUFFLE_BIT
|
1572 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT
;
1573 p
->subgroupQuadOperationsInAllStages
= true;
1575 p
->pointClippingBehavior
= VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES
;
1576 p
->maxMultiviewViewCount
= MAX_VIEWS
;
1577 p
->maxMultiviewInstanceIndex
= INT_MAX
;
1578 p
->protectedNoFault
= false;
1579 p
->maxPerSetDescriptors
= RADV_MAX_PER_SET_DESCRIPTORS
;
1580 p
->maxMemoryAllocationSize
= RADV_MAX_MEMORY_ALLOCATION_SIZE
;
1584 radv_get_physical_device_properties_1_2(struct radv_physical_device
*pdevice
,
1585 VkPhysicalDeviceVulkan12Properties
*p
)
1587 assert(p
->sType
== VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES
);
1589 p
->driverID
= VK_DRIVER_ID_MESA_RADV
;
1590 snprintf(p
->driverName
, VK_MAX_DRIVER_NAME_SIZE
, "radv");
1591 snprintf(p
->driverInfo
, VK_MAX_DRIVER_INFO_SIZE
,
1592 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1
" (%s)",
1593 radv_get_compiler_string(pdevice
));
1594 p
->conformanceVersion
= (VkConformanceVersion
) {
1601 /* On AMD hardware, denormals and rounding modes for fp16/fp64 are
1602 * controlled by the same config register.
1604 if (pdevice
->rad_info
.has_packed_math_16bit
) {
1605 p
->denormBehaviorIndependence
= VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_32_BIT_ONLY_KHR
;
1606 p
->roundingModeIndependence
= VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_32_BIT_ONLY_KHR
;
1608 p
->denormBehaviorIndependence
= VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_ALL_KHR
;
1609 p
->roundingModeIndependence
= VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_ALL_KHR
;
1612 /* With LLVM, do not allow both preserving and flushing denorms because
1613 * different shaders in the same pipeline can have different settings and
1614 * this won't work for merged shaders. To make it work, this requires LLVM
1615 * support for changing the register. The same logic applies for the
1616 * rounding modes because they are configured with the same config
1619 p
->shaderDenormFlushToZeroFloat32
= true;
1620 p
->shaderDenormPreserveFloat32
= !pdevice
->use_llvm
;
1621 p
->shaderRoundingModeRTEFloat32
= true;
1622 p
->shaderRoundingModeRTZFloat32
= !pdevice
->use_llvm
;
1623 p
->shaderSignedZeroInfNanPreserveFloat32
= true;
1625 p
->shaderDenormFlushToZeroFloat16
= pdevice
->rad_info
.has_packed_math_16bit
&& !pdevice
->use_llvm
;
1626 p
->shaderDenormPreserveFloat16
= pdevice
->rad_info
.has_packed_math_16bit
;
1627 p
->shaderRoundingModeRTEFloat16
= pdevice
->rad_info
.has_packed_math_16bit
;
1628 p
->shaderRoundingModeRTZFloat16
= pdevice
->rad_info
.has_packed_math_16bit
&& !pdevice
->use_llvm
;
1629 p
->shaderSignedZeroInfNanPreserveFloat16
= pdevice
->rad_info
.has_packed_math_16bit
;
1631 p
->shaderDenormFlushToZeroFloat64
= pdevice
->rad_info
.chip_class
>= GFX8
&& !pdevice
->use_llvm
;
1632 p
->shaderDenormPreserveFloat64
= pdevice
->rad_info
.chip_class
>= GFX8
;
1633 p
->shaderRoundingModeRTEFloat64
= pdevice
->rad_info
.chip_class
>= GFX8
;
1634 p
->shaderRoundingModeRTZFloat64
= pdevice
->rad_info
.chip_class
>= GFX8
&& !pdevice
->use_llvm
;
1635 p
->shaderSignedZeroInfNanPreserveFloat64
= pdevice
->rad_info
.chip_class
>= GFX8
;
1637 p
->maxUpdateAfterBindDescriptorsInAllPools
= UINT32_MAX
/ 64;
1638 p
->shaderUniformBufferArrayNonUniformIndexingNative
= false;
1639 p
->shaderSampledImageArrayNonUniformIndexingNative
= false;
1640 p
->shaderStorageBufferArrayNonUniformIndexingNative
= false;
1641 p
->shaderStorageImageArrayNonUniformIndexingNative
= false;
1642 p
->shaderInputAttachmentArrayNonUniformIndexingNative
= false;
1643 p
->robustBufferAccessUpdateAfterBind
= false;
1644 p
->quadDivergentImplicitLod
= false;
1646 size_t max_descriptor_set_size
= ((1ull << 31) - 16 * MAX_DYNAMIC_BUFFERS
-
1647 MAX_INLINE_UNIFORM_BLOCK_SIZE
* MAX_INLINE_UNIFORM_BLOCK_COUNT
) /
1648 (32 /* uniform buffer, 32 due to potential space wasted on alignment */ +
1649 32 /* storage buffer, 32 due to potential space wasted on alignment */ +
1650 32 /* sampler, largest when combined with image */ +
1651 64 /* sampled image */ +
1652 64 /* storage image */);
1653 p
->maxPerStageDescriptorUpdateAfterBindSamplers
= max_descriptor_set_size
;
1654 p
->maxPerStageDescriptorUpdateAfterBindUniformBuffers
= max_descriptor_set_size
;
1655 p
->maxPerStageDescriptorUpdateAfterBindStorageBuffers
= max_descriptor_set_size
;
1656 p
->maxPerStageDescriptorUpdateAfterBindSampledImages
= max_descriptor_set_size
;
1657 p
->maxPerStageDescriptorUpdateAfterBindStorageImages
= max_descriptor_set_size
;
1658 p
->maxPerStageDescriptorUpdateAfterBindInputAttachments
= max_descriptor_set_size
;
1659 p
->maxPerStageUpdateAfterBindResources
= max_descriptor_set_size
;
1660 p
->maxDescriptorSetUpdateAfterBindSamplers
= max_descriptor_set_size
;
1661 p
->maxDescriptorSetUpdateAfterBindUniformBuffers
= max_descriptor_set_size
;
1662 p
->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic
= MAX_DYNAMIC_UNIFORM_BUFFERS
;
1663 p
->maxDescriptorSetUpdateAfterBindStorageBuffers
= max_descriptor_set_size
;
1664 p
->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic
= MAX_DYNAMIC_STORAGE_BUFFERS
;
1665 p
->maxDescriptorSetUpdateAfterBindSampledImages
= max_descriptor_set_size
;
1666 p
->maxDescriptorSetUpdateAfterBindStorageImages
= max_descriptor_set_size
;
1667 p
->maxDescriptorSetUpdateAfterBindInputAttachments
= max_descriptor_set_size
;
1669 /* We support all of the depth resolve modes */
1670 p
->supportedDepthResolveModes
= VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
|
1671 VK_RESOLVE_MODE_AVERAGE_BIT_KHR
|
1672 VK_RESOLVE_MODE_MIN_BIT_KHR
|
1673 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1675 /* Average doesn't make sense for stencil so we don't support that */
1676 p
->supportedStencilResolveModes
= VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
|
1677 VK_RESOLVE_MODE_MIN_BIT_KHR
|
1678 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1680 p
->independentResolveNone
= true;
1681 p
->independentResolve
= true;
1683 /* GFX6-8 only support single channel min/max filter. */
1684 p
->filterMinmaxImageComponentMapping
= pdevice
->rad_info
.chip_class
>= GFX9
;
1685 p
->filterMinmaxSingleComponentFormats
= true;
1687 p
->maxTimelineSemaphoreValueDifference
= UINT64_MAX
;
1689 p
->framebufferIntegerColorSampleCounts
= VK_SAMPLE_COUNT_1_BIT
;
1692 void radv_GetPhysicalDeviceProperties2(
1693 VkPhysicalDevice physicalDevice
,
1694 VkPhysicalDeviceProperties2
*pProperties
)
1696 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
1697 radv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
1699 VkPhysicalDeviceVulkan11Properties core_1_1
= {
1700 .sType
= VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES
,
1702 radv_get_physical_device_properties_1_1(pdevice
, &core_1_1
);
1704 VkPhysicalDeviceVulkan12Properties core_1_2
= {
1705 .sType
= VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES
,
1707 radv_get_physical_device_properties_1_2(pdevice
, &core_1_2
);
1709 #define CORE_RENAMED_PROPERTY(major, minor, ext_property, core_property) \
1710 memcpy(&properties->ext_property, &core_##major##_##minor.core_property, \
1711 sizeof(core_##major##_##minor.core_property))
1713 #define CORE_PROPERTY(major, minor, property) \
1714 CORE_RENAMED_PROPERTY(major, minor, property, property)
1716 vk_foreach_struct(ext
, pProperties
->pNext
) {
1717 switch (ext
->sType
) {
1718 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
1719 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
1720 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
1721 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
1724 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES
: {
1725 VkPhysicalDeviceIDProperties
*properties
= (VkPhysicalDeviceIDProperties
*)ext
;
1726 CORE_PROPERTY(1, 1, deviceUUID
);
1727 CORE_PROPERTY(1, 1, driverUUID
);
1728 CORE_PROPERTY(1, 1, deviceLUID
);
1729 CORE_PROPERTY(1, 1, deviceLUIDValid
);
1732 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES
: {
1733 VkPhysicalDeviceMultiviewProperties
*properties
= (VkPhysicalDeviceMultiviewProperties
*)ext
;
1734 CORE_PROPERTY(1, 1, maxMultiviewViewCount
);
1735 CORE_PROPERTY(1, 1, maxMultiviewInstanceIndex
);
1738 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES
: {
1739 VkPhysicalDevicePointClippingProperties
*properties
=
1740 (VkPhysicalDevicePointClippingProperties
*)ext
;
1741 CORE_PROPERTY(1, 1, pointClippingBehavior
);
1744 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DISCARD_RECTANGLE_PROPERTIES_EXT
: {
1745 VkPhysicalDeviceDiscardRectanglePropertiesEXT
*properties
=
1746 (VkPhysicalDeviceDiscardRectanglePropertiesEXT
*)ext
;
1747 properties
->maxDiscardRectangles
= MAX_DISCARD_RECTANGLES
;
1750 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT
: {
1751 VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*properties
=
1752 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*) ext
;
1753 properties
->minImportedHostPointerAlignment
= 4096;
1756 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES
: {
1757 VkPhysicalDeviceSubgroupProperties
*properties
=
1758 (VkPhysicalDeviceSubgroupProperties
*)ext
;
1759 CORE_PROPERTY(1, 1, subgroupSize
);
1760 CORE_RENAMED_PROPERTY(1, 1, supportedStages
,
1761 subgroupSupportedStages
);
1762 CORE_RENAMED_PROPERTY(1, 1, supportedOperations
,
1763 subgroupSupportedOperations
);
1764 CORE_RENAMED_PROPERTY(1, 1, quadOperationsInAllStages
,
1765 subgroupQuadOperationsInAllStages
);
1768 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES
: {
1769 VkPhysicalDeviceMaintenance3Properties
*properties
=
1770 (VkPhysicalDeviceMaintenance3Properties
*)ext
;
1771 CORE_PROPERTY(1, 1, maxPerSetDescriptors
);
1772 CORE_PROPERTY(1, 1, maxMemoryAllocationSize
);
1775 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES
: {
1776 VkPhysicalDeviceSamplerFilterMinmaxProperties
*properties
=
1777 (VkPhysicalDeviceSamplerFilterMinmaxProperties
*)ext
;
1778 CORE_PROPERTY(1, 2, filterMinmaxImageComponentMapping
);
1779 CORE_PROPERTY(1, 2, filterMinmaxSingleComponentFormats
);
1782 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CORE_PROPERTIES_AMD
: {
1783 VkPhysicalDeviceShaderCorePropertiesAMD
*properties
=
1784 (VkPhysicalDeviceShaderCorePropertiesAMD
*)ext
;
1786 /* Shader engines. */
1787 properties
->shaderEngineCount
=
1788 pdevice
->rad_info
.max_se
;
1789 properties
->shaderArraysPerEngineCount
=
1790 pdevice
->rad_info
.max_sh_per_se
;
1791 properties
->computeUnitsPerShaderArray
=
1792 pdevice
->rad_info
.min_good_cu_per_sa
;
1793 properties
->simdPerComputeUnit
=
1794 pdevice
->rad_info
.num_simd_per_compute_unit
;
1795 properties
->wavefrontsPerSimd
=
1796 pdevice
->rad_info
.max_wave64_per_simd
;
1797 properties
->wavefrontSize
= 64;
1800 properties
->sgprsPerSimd
=
1801 pdevice
->rad_info
.num_physical_sgprs_per_simd
;
1802 properties
->minSgprAllocation
=
1803 pdevice
->rad_info
.min_sgpr_alloc
;
1804 properties
->maxSgprAllocation
=
1805 pdevice
->rad_info
.max_sgpr_alloc
;
1806 properties
->sgprAllocationGranularity
=
1807 pdevice
->rad_info
.sgpr_alloc_granularity
;
1810 properties
->vgprsPerSimd
=
1811 pdevice
->rad_info
.num_physical_wave64_vgprs_per_simd
;
1812 properties
->minVgprAllocation
=
1813 pdevice
->rad_info
.min_wave64_vgpr_alloc
;
1814 properties
->maxVgprAllocation
=
1815 pdevice
->rad_info
.max_vgpr_alloc
;
1816 properties
->vgprAllocationGranularity
=
1817 pdevice
->rad_info
.wave64_vgpr_alloc_granularity
;
1820 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CORE_PROPERTIES_2_AMD
: {
1821 VkPhysicalDeviceShaderCoreProperties2AMD
*properties
=
1822 (VkPhysicalDeviceShaderCoreProperties2AMD
*)ext
;
1824 properties
->shaderCoreFeatures
= 0;
1825 properties
->activeComputeUnitCount
=
1826 pdevice
->rad_info
.num_good_compute_units
;
1829 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT
: {
1830 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*properties
=
1831 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*)ext
;
1832 properties
->maxVertexAttribDivisor
= UINT32_MAX
;
1835 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES
: {
1836 VkPhysicalDeviceDescriptorIndexingProperties
*properties
=
1837 (VkPhysicalDeviceDescriptorIndexingProperties
*)ext
;
1838 CORE_PROPERTY(1, 2, maxUpdateAfterBindDescriptorsInAllPools
);
1839 CORE_PROPERTY(1, 2, shaderUniformBufferArrayNonUniformIndexingNative
);
1840 CORE_PROPERTY(1, 2, shaderSampledImageArrayNonUniformIndexingNative
);
1841 CORE_PROPERTY(1, 2, shaderStorageBufferArrayNonUniformIndexingNative
);
1842 CORE_PROPERTY(1, 2, shaderStorageImageArrayNonUniformIndexingNative
);
1843 CORE_PROPERTY(1, 2, shaderInputAttachmentArrayNonUniformIndexingNative
);
1844 CORE_PROPERTY(1, 2, robustBufferAccessUpdateAfterBind
);
1845 CORE_PROPERTY(1, 2, quadDivergentImplicitLod
);
1846 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindSamplers
);
1847 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindUniformBuffers
);
1848 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindStorageBuffers
);
1849 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindSampledImages
);
1850 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindStorageImages
);
1851 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindInputAttachments
);
1852 CORE_PROPERTY(1, 2, maxPerStageUpdateAfterBindResources
);
1853 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindSamplers
);
1854 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindUniformBuffers
);
1855 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindUniformBuffersDynamic
);
1856 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindStorageBuffers
);
1857 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindStorageBuffersDynamic
);
1858 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindSampledImages
);
1859 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindStorageImages
);
1860 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindInputAttachments
);
1863 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES
: {
1864 VkPhysicalDeviceProtectedMemoryProperties
*properties
=
1865 (VkPhysicalDeviceProtectedMemoryProperties
*)ext
;
1866 CORE_PROPERTY(1, 1, protectedNoFault
);
1869 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONSERVATIVE_RASTERIZATION_PROPERTIES_EXT
: {
1870 VkPhysicalDeviceConservativeRasterizationPropertiesEXT
*properties
=
1871 (VkPhysicalDeviceConservativeRasterizationPropertiesEXT
*)ext
;
1872 properties
->primitiveOverestimationSize
= 0;
1873 properties
->maxExtraPrimitiveOverestimationSize
= 0;
1874 properties
->extraPrimitiveOverestimationSizeGranularity
= 0;
1875 properties
->primitiveUnderestimation
= false;
1876 properties
->conservativePointAndLineRasterization
= false;
1877 properties
->degenerateTrianglesRasterized
= false;
1878 properties
->degenerateLinesRasterized
= false;
1879 properties
->fullyCoveredFragmentShaderInputVariable
= false;
1880 properties
->conservativeRasterizationPostDepthCoverage
= false;
1883 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT
: {
1884 VkPhysicalDevicePCIBusInfoPropertiesEXT
*properties
=
1885 (VkPhysicalDevicePCIBusInfoPropertiesEXT
*)ext
;
1886 properties
->pciDomain
= pdevice
->bus_info
.domain
;
1887 properties
->pciBus
= pdevice
->bus_info
.bus
;
1888 properties
->pciDevice
= pdevice
->bus_info
.dev
;
1889 properties
->pciFunction
= pdevice
->bus_info
.func
;
1892 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES
: {
1893 VkPhysicalDeviceDriverProperties
*properties
=
1894 (VkPhysicalDeviceDriverProperties
*) ext
;
1895 CORE_PROPERTY(1, 2, driverID
);
1896 CORE_PROPERTY(1, 2, driverName
);
1897 CORE_PROPERTY(1, 2, driverInfo
);
1898 CORE_PROPERTY(1, 2, conformanceVersion
);
1901 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT
: {
1902 VkPhysicalDeviceTransformFeedbackPropertiesEXT
*properties
=
1903 (VkPhysicalDeviceTransformFeedbackPropertiesEXT
*)ext
;
1904 properties
->maxTransformFeedbackStreams
= MAX_SO_STREAMS
;
1905 properties
->maxTransformFeedbackBuffers
= MAX_SO_BUFFERS
;
1906 properties
->maxTransformFeedbackBufferSize
= UINT32_MAX
;
1907 properties
->maxTransformFeedbackStreamDataSize
= 512;
1908 properties
->maxTransformFeedbackBufferDataSize
= UINT32_MAX
;
1909 properties
->maxTransformFeedbackBufferDataStride
= 512;
1910 properties
->transformFeedbackQueries
= !pdevice
->use_ngg_streamout
;
1911 properties
->transformFeedbackStreamsLinesTriangles
= !pdevice
->use_ngg_streamout
;
1912 properties
->transformFeedbackRasterizationStreamSelect
= false;
1913 properties
->transformFeedbackDraw
= true;
1916 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_PROPERTIES_EXT
: {
1917 VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*props
=
1918 (VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*)ext
;
1920 props
->maxInlineUniformBlockSize
= MAX_INLINE_UNIFORM_BLOCK_SIZE
;
1921 props
->maxPerStageDescriptorInlineUniformBlocks
= MAX_INLINE_UNIFORM_BLOCK_SIZE
* MAX_SETS
;
1922 props
->maxPerStageDescriptorUpdateAfterBindInlineUniformBlocks
= MAX_INLINE_UNIFORM_BLOCK_SIZE
* MAX_SETS
;
1923 props
->maxDescriptorSetInlineUniformBlocks
= MAX_INLINE_UNIFORM_BLOCK_COUNT
;
1924 props
->maxDescriptorSetUpdateAfterBindInlineUniformBlocks
= MAX_INLINE_UNIFORM_BLOCK_COUNT
;
1927 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLE_LOCATIONS_PROPERTIES_EXT
: {
1928 VkPhysicalDeviceSampleLocationsPropertiesEXT
*properties
=
1929 (VkPhysicalDeviceSampleLocationsPropertiesEXT
*)ext
;
1930 properties
->sampleLocationSampleCounts
= VK_SAMPLE_COUNT_2_BIT
|
1931 VK_SAMPLE_COUNT_4_BIT
|
1932 VK_SAMPLE_COUNT_8_BIT
;
1933 properties
->maxSampleLocationGridSize
= (VkExtent2D
){ 2 , 2 };
1934 properties
->sampleLocationCoordinateRange
[0] = 0.0f
;
1935 properties
->sampleLocationCoordinateRange
[1] = 0.9375f
;
1936 properties
->sampleLocationSubPixelBits
= 4;
1937 properties
->variableSampleLocations
= false;
1940 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_STENCIL_RESOLVE_PROPERTIES
: {
1941 VkPhysicalDeviceDepthStencilResolveProperties
*properties
=
1942 (VkPhysicalDeviceDepthStencilResolveProperties
*)ext
;
1943 CORE_PROPERTY(1, 2, supportedDepthResolveModes
);
1944 CORE_PROPERTY(1, 2, supportedStencilResolveModes
);
1945 CORE_PROPERTY(1, 2, independentResolveNone
);
1946 CORE_PROPERTY(1, 2, independentResolve
);
1949 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_PROPERTIES_EXT
: {
1950 VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT
*properties
=
1951 (VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT
*)ext
;
1952 properties
->storageTexelBufferOffsetAlignmentBytes
= 4;
1953 properties
->storageTexelBufferOffsetSingleTexelAlignment
= true;
1954 properties
->uniformTexelBufferOffsetAlignmentBytes
= 4;
1955 properties
->uniformTexelBufferOffsetSingleTexelAlignment
= true;
1958 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT_CONTROLS_PROPERTIES
: {
1959 VkPhysicalDeviceFloatControlsProperties
*properties
=
1960 (VkPhysicalDeviceFloatControlsProperties
*)ext
;
1961 CORE_PROPERTY(1, 2, denormBehaviorIndependence
);
1962 CORE_PROPERTY(1, 2, roundingModeIndependence
);
1963 CORE_PROPERTY(1, 2, shaderDenormFlushToZeroFloat16
);
1964 CORE_PROPERTY(1, 2, shaderDenormPreserveFloat16
);
1965 CORE_PROPERTY(1, 2, shaderRoundingModeRTEFloat16
);
1966 CORE_PROPERTY(1, 2, shaderRoundingModeRTZFloat16
);
1967 CORE_PROPERTY(1, 2, shaderSignedZeroInfNanPreserveFloat16
);
1968 CORE_PROPERTY(1, 2, shaderDenormFlushToZeroFloat32
);
1969 CORE_PROPERTY(1, 2, shaderDenormPreserveFloat32
);
1970 CORE_PROPERTY(1, 2, shaderRoundingModeRTEFloat32
);
1971 CORE_PROPERTY(1, 2, shaderRoundingModeRTZFloat32
);
1972 CORE_PROPERTY(1, 2, shaderSignedZeroInfNanPreserveFloat32
);
1973 CORE_PROPERTY(1, 2, shaderDenormFlushToZeroFloat64
);
1974 CORE_PROPERTY(1, 2, shaderDenormPreserveFloat64
);
1975 CORE_PROPERTY(1, 2, shaderRoundingModeRTEFloat64
);
1976 CORE_PROPERTY(1, 2, shaderRoundingModeRTZFloat64
);
1977 CORE_PROPERTY(1, 2, shaderSignedZeroInfNanPreserveFloat64
);
1980 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_PROPERTIES
: {
1981 VkPhysicalDeviceTimelineSemaphoreProperties
*properties
=
1982 (VkPhysicalDeviceTimelineSemaphoreProperties
*) ext
;
1983 CORE_PROPERTY(1, 2, maxTimelineSemaphoreValueDifference
);
1986 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_PROPERTIES_EXT
: {
1987 VkPhysicalDeviceSubgroupSizeControlPropertiesEXT
*props
=
1988 (VkPhysicalDeviceSubgroupSizeControlPropertiesEXT
*)ext
;
1989 props
->minSubgroupSize
= 64;
1990 props
->maxSubgroupSize
= 64;
1991 props
->maxComputeWorkgroupSubgroups
= UINT32_MAX
;
1992 props
->requiredSubgroupSizeStages
= 0;
1994 if (pdevice
->rad_info
.chip_class
>= GFX10
) {
1995 /* Only GFX10+ supports wave32. */
1996 props
->minSubgroupSize
= 32;
1997 props
->requiredSubgroupSizeStages
= VK_SHADER_STAGE_COMPUTE_BIT
;
2001 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES
:
2002 radv_get_physical_device_properties_1_1(pdevice
, (void *)ext
);
2004 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES
:
2005 radv_get_physical_device_properties_1_2(pdevice
, (void *)ext
);
2007 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_PROPERTIES_EXT
: {
2008 VkPhysicalDeviceLineRasterizationPropertiesEXT
*props
=
2009 (VkPhysicalDeviceLineRasterizationPropertiesEXT
*)ext
;
2010 props
->lineSubPixelPrecisionBits
= 4;
2013 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ROBUSTNESS_2_PROPERTIES_EXT
: {
2014 VkPhysicalDeviceRobustness2PropertiesEXT
*properties
=
2015 (VkPhysicalDeviceRobustness2PropertiesEXT
*)ext
;
2016 properties
->robustStorageBufferAccessSizeAlignment
= 4;
2017 properties
->robustUniformBufferAccessSizeAlignment
= 4;
2020 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CUSTOM_BORDER_COLOR_PROPERTIES_EXT
: {
2021 VkPhysicalDeviceCustomBorderColorPropertiesEXT
*props
=
2022 (VkPhysicalDeviceCustomBorderColorPropertiesEXT
*)ext
;
2023 props
->maxCustomBorderColorSamplers
= RADV_BORDER_COLOR_COUNT
;
2032 static void radv_get_physical_device_queue_family_properties(
2033 struct radv_physical_device
* pdevice
,
2035 VkQueueFamilyProperties
** pQueueFamilyProperties
)
2037 int num_queue_families
= 1;
2039 if (pdevice
->rad_info
.num_rings
[RING_COMPUTE
] > 0 &&
2040 !(pdevice
->instance
->debug_flags
& RADV_DEBUG_NO_COMPUTE_QUEUE
))
2041 num_queue_families
++;
2043 if (pQueueFamilyProperties
== NULL
) {
2044 *pCount
= num_queue_families
;
2053 *pQueueFamilyProperties
[idx
] = (VkQueueFamilyProperties
) {
2054 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
2055 VK_QUEUE_COMPUTE_BIT
|
2056 VK_QUEUE_TRANSFER_BIT
|
2057 VK_QUEUE_SPARSE_BINDING_BIT
,
2059 .timestampValidBits
= 64,
2060 .minImageTransferGranularity
= (VkExtent3D
) { 1, 1, 1 },
2065 if (pdevice
->rad_info
.num_rings
[RING_COMPUTE
] > 0 &&
2066 !(pdevice
->instance
->debug_flags
& RADV_DEBUG_NO_COMPUTE_QUEUE
)) {
2067 if (*pCount
> idx
) {
2068 *pQueueFamilyProperties
[idx
] = (VkQueueFamilyProperties
) {
2069 .queueFlags
= VK_QUEUE_COMPUTE_BIT
|
2070 VK_QUEUE_TRANSFER_BIT
|
2071 VK_QUEUE_SPARSE_BINDING_BIT
,
2072 .queueCount
= pdevice
->rad_info
.num_rings
[RING_COMPUTE
],
2073 .timestampValidBits
= 64,
2074 .minImageTransferGranularity
= (VkExtent3D
) { 1, 1, 1 },
2082 void radv_GetPhysicalDeviceQueueFamilyProperties(
2083 VkPhysicalDevice physicalDevice
,
2085 VkQueueFamilyProperties
* pQueueFamilyProperties
)
2087 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
2088 if (!pQueueFamilyProperties
) {
2089 radv_get_physical_device_queue_family_properties(pdevice
, pCount
, NULL
);
2092 VkQueueFamilyProperties
*properties
[] = {
2093 pQueueFamilyProperties
+ 0,
2094 pQueueFamilyProperties
+ 1,
2095 pQueueFamilyProperties
+ 2,
2097 radv_get_physical_device_queue_family_properties(pdevice
, pCount
, properties
);
2098 assert(*pCount
<= 3);
2101 void radv_GetPhysicalDeviceQueueFamilyProperties2(
2102 VkPhysicalDevice physicalDevice
,
2104 VkQueueFamilyProperties2
*pQueueFamilyProperties
)
2106 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
2107 if (!pQueueFamilyProperties
) {
2108 radv_get_physical_device_queue_family_properties(pdevice
, pCount
, NULL
);
2111 VkQueueFamilyProperties
*properties
[] = {
2112 &pQueueFamilyProperties
[0].queueFamilyProperties
,
2113 &pQueueFamilyProperties
[1].queueFamilyProperties
,
2114 &pQueueFamilyProperties
[2].queueFamilyProperties
,
2116 radv_get_physical_device_queue_family_properties(pdevice
, pCount
, properties
);
2117 assert(*pCount
<= 3);
2120 void radv_GetPhysicalDeviceMemoryProperties(
2121 VkPhysicalDevice physicalDevice
,
2122 VkPhysicalDeviceMemoryProperties
*pMemoryProperties
)
2124 RADV_FROM_HANDLE(radv_physical_device
, physical_device
, physicalDevice
);
2126 *pMemoryProperties
= physical_device
->memory_properties
;
2130 radv_get_memory_budget_properties(VkPhysicalDevice physicalDevice
,
2131 VkPhysicalDeviceMemoryBudgetPropertiesEXT
*memoryBudget
)
2133 RADV_FROM_HANDLE(radv_physical_device
, device
, physicalDevice
);
2134 VkPhysicalDeviceMemoryProperties
*memory_properties
= &device
->memory_properties
;
2135 uint64_t visible_vram_size
= radv_get_visible_vram_size(device
);
2136 uint64_t vram_size
= radv_get_vram_size(device
);
2137 uint64_t gtt_size
= device
->rad_info
.gart_size
;
2138 uint64_t heap_budget
, heap_usage
;
2140 /* For all memory heaps, the computation of budget is as follow:
2141 * heap_budget = heap_size - global_heap_usage + app_heap_usage
2143 * The Vulkan spec 1.1.97 says that the budget should include any
2144 * currently allocated device memory.
2146 * Note that the application heap usages are not really accurate (eg.
2147 * in presence of shared buffers).
2149 for (int i
= 0; i
< device
->memory_properties
.memoryTypeCount
; i
++) {
2150 uint32_t heap_index
= device
->memory_properties
.memoryTypes
[i
].heapIndex
;
2152 if ((device
->memory_domains
[i
] & RADEON_DOMAIN_VRAM
) && (device
->memory_flags
[i
] & RADEON_FLAG_NO_CPU_ACCESS
)) {
2153 heap_usage
= device
->ws
->query_value(device
->ws
,
2154 RADEON_ALLOCATED_VRAM
);
2156 heap_budget
= vram_size
-
2157 device
->ws
->query_value(device
->ws
, RADEON_VRAM_USAGE
) +
2160 memoryBudget
->heapBudget
[heap_index
] = heap_budget
;
2161 memoryBudget
->heapUsage
[heap_index
] = heap_usage
;
2162 } else if (device
->memory_domains
[i
] & RADEON_DOMAIN_VRAM
) {
2163 heap_usage
= device
->ws
->query_value(device
->ws
,
2164 RADEON_ALLOCATED_VRAM_VIS
);
2166 heap_budget
= visible_vram_size
-
2167 device
->ws
->query_value(device
->ws
, RADEON_VRAM_VIS_USAGE
) +
2170 memoryBudget
->heapBudget
[heap_index
] = heap_budget
;
2171 memoryBudget
->heapUsage
[heap_index
] = heap_usage
;
2173 assert(device
->memory_domains
[i
] & RADEON_DOMAIN_GTT
);
2175 heap_usage
= device
->ws
->query_value(device
->ws
,
2176 RADEON_ALLOCATED_GTT
);
2178 heap_budget
= gtt_size
-
2179 device
->ws
->query_value(device
->ws
, RADEON_GTT_USAGE
) +
2182 memoryBudget
->heapBudget
[heap_index
] = heap_budget
;
2183 memoryBudget
->heapUsage
[heap_index
] = heap_usage
;
2187 /* The heapBudget and heapUsage values must be zero for array elements
2188 * greater than or equal to
2189 * VkPhysicalDeviceMemoryProperties::memoryHeapCount.
2191 for (uint32_t i
= memory_properties
->memoryHeapCount
; i
< VK_MAX_MEMORY_HEAPS
; i
++) {
2192 memoryBudget
->heapBudget
[i
] = 0;
2193 memoryBudget
->heapUsage
[i
] = 0;
2197 void radv_GetPhysicalDeviceMemoryProperties2(
2198 VkPhysicalDevice physicalDevice
,
2199 VkPhysicalDeviceMemoryProperties2
*pMemoryProperties
)
2201 radv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
2202 &pMemoryProperties
->memoryProperties
);
2204 VkPhysicalDeviceMemoryBudgetPropertiesEXT
*memory_budget
=
2205 vk_find_struct(pMemoryProperties
->pNext
,
2206 PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT
);
2208 radv_get_memory_budget_properties(physicalDevice
, memory_budget
);
2211 VkResult
radv_GetMemoryHostPointerPropertiesEXT(
2213 VkExternalMemoryHandleTypeFlagBits handleType
,
2214 const void *pHostPointer
,
2215 VkMemoryHostPointerPropertiesEXT
*pMemoryHostPointerProperties
)
2217 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2221 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
: {
2222 const struct radv_physical_device
*physical_device
= device
->physical_device
;
2223 uint32_t memoryTypeBits
= 0;
2224 for (int i
= 0; i
< physical_device
->memory_properties
.memoryTypeCount
; i
++) {
2225 if (physical_device
->memory_domains
[i
] == RADEON_DOMAIN_GTT
&&
2226 !(physical_device
->memory_flags
[i
] & RADEON_FLAG_GTT_WC
)) {
2227 memoryTypeBits
= (1 << i
);
2231 pMemoryHostPointerProperties
->memoryTypeBits
= memoryTypeBits
;
2235 return VK_ERROR_INVALID_EXTERNAL_HANDLE
;
2239 static enum radeon_ctx_priority
2240 radv_get_queue_global_priority(const VkDeviceQueueGlobalPriorityCreateInfoEXT
*pObj
)
2242 /* Default to MEDIUM when a specific global priority isn't requested */
2244 return RADEON_CTX_PRIORITY_MEDIUM
;
2246 switch(pObj
->globalPriority
) {
2247 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT
:
2248 return RADEON_CTX_PRIORITY_REALTIME
;
2249 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT
:
2250 return RADEON_CTX_PRIORITY_HIGH
;
2251 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
:
2252 return RADEON_CTX_PRIORITY_MEDIUM
;
2253 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT
:
2254 return RADEON_CTX_PRIORITY_LOW
;
2256 unreachable("Illegal global priority value");
2257 return RADEON_CTX_PRIORITY_INVALID
;
2262 radv_queue_init(struct radv_device
*device
, struct radv_queue
*queue
,
2263 uint32_t queue_family_index
, int idx
,
2264 VkDeviceQueueCreateFlags flags
,
2265 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*global_priority
)
2267 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
2268 queue
->device
= device
;
2269 queue
->queue_family_index
= queue_family_index
;
2270 queue
->queue_idx
= idx
;
2271 queue
->priority
= radv_get_queue_global_priority(global_priority
);
2272 queue
->flags
= flags
;
2273 queue
->hw_ctx
= NULL
;
2275 VkResult result
= device
->ws
->ctx_create(device
->ws
, queue
->priority
, &queue
->hw_ctx
);
2276 if (result
!= VK_SUCCESS
)
2277 return vk_error(device
->instance
, result
);
2279 list_inithead(&queue
->pending_submissions
);
2280 pthread_mutex_init(&queue
->pending_mutex
, NULL
);
2282 pthread_mutex_init(&queue
->thread_mutex
, NULL
);
2283 queue
->thread_submission
= NULL
;
2284 queue
->thread_running
= queue
->thread_exit
= false;
2285 result
= radv_create_pthread_cond(&queue
->thread_cond
);
2286 if (result
!= VK_SUCCESS
)
2287 return vk_error(device
->instance
, result
);
2293 radv_queue_finish(struct radv_queue
*queue
)
2295 if (queue
->thread_running
) {
2296 p_atomic_set(&queue
->thread_exit
, true);
2297 pthread_cond_broadcast(&queue
->thread_cond
);
2298 pthread_join(queue
->submission_thread
, NULL
);
2300 pthread_cond_destroy(&queue
->thread_cond
);
2301 pthread_mutex_destroy(&queue
->pending_mutex
);
2302 pthread_mutex_destroy(&queue
->thread_mutex
);
2305 queue
->device
->ws
->ctx_destroy(queue
->hw_ctx
);
2307 if (queue
->initial_full_flush_preamble_cs
)
2308 queue
->device
->ws
->cs_destroy(queue
->initial_full_flush_preamble_cs
);
2309 if (queue
->initial_preamble_cs
)
2310 queue
->device
->ws
->cs_destroy(queue
->initial_preamble_cs
);
2311 if (queue
->continue_preamble_cs
)
2312 queue
->device
->ws
->cs_destroy(queue
->continue_preamble_cs
);
2313 if (queue
->descriptor_bo
)
2314 queue
->device
->ws
->buffer_destroy(queue
->descriptor_bo
);
2315 if (queue
->scratch_bo
)
2316 queue
->device
->ws
->buffer_destroy(queue
->scratch_bo
);
2317 if (queue
->esgs_ring_bo
)
2318 queue
->device
->ws
->buffer_destroy(queue
->esgs_ring_bo
);
2319 if (queue
->gsvs_ring_bo
)
2320 queue
->device
->ws
->buffer_destroy(queue
->gsvs_ring_bo
);
2321 if (queue
->tess_rings_bo
)
2322 queue
->device
->ws
->buffer_destroy(queue
->tess_rings_bo
);
2324 queue
->device
->ws
->buffer_destroy(queue
->gds_bo
);
2325 if (queue
->gds_oa_bo
)
2326 queue
->device
->ws
->buffer_destroy(queue
->gds_oa_bo
);
2327 if (queue
->compute_scratch_bo
)
2328 queue
->device
->ws
->buffer_destroy(queue
->compute_scratch_bo
);
2332 radv_bo_list_init(struct radv_bo_list
*bo_list
)
2334 pthread_mutex_init(&bo_list
->mutex
, NULL
);
2335 bo_list
->list
.count
= bo_list
->capacity
= 0;
2336 bo_list
->list
.bos
= NULL
;
2340 radv_bo_list_finish(struct radv_bo_list
*bo_list
)
2342 free(bo_list
->list
.bos
);
2343 pthread_mutex_destroy(&bo_list
->mutex
);
2346 VkResult
radv_bo_list_add(struct radv_device
*device
,
2347 struct radeon_winsys_bo
*bo
)
2349 struct radv_bo_list
*bo_list
= &device
->bo_list
;
2354 if (unlikely(!device
->use_global_bo_list
))
2357 pthread_mutex_lock(&bo_list
->mutex
);
2358 if (bo_list
->list
.count
== bo_list
->capacity
) {
2359 unsigned capacity
= MAX2(4, bo_list
->capacity
* 2);
2360 void *data
= realloc(bo_list
->list
.bos
, capacity
* sizeof(struct radeon_winsys_bo
*));
2363 pthread_mutex_unlock(&bo_list
->mutex
);
2364 return VK_ERROR_OUT_OF_HOST_MEMORY
;
2367 bo_list
->list
.bos
= (struct radeon_winsys_bo
**)data
;
2368 bo_list
->capacity
= capacity
;
2371 bo_list
->list
.bos
[bo_list
->list
.count
++] = bo
;
2372 pthread_mutex_unlock(&bo_list
->mutex
);
2376 void radv_bo_list_remove(struct radv_device
*device
,
2377 struct radeon_winsys_bo
*bo
)
2379 struct radv_bo_list
*bo_list
= &device
->bo_list
;
2384 if (unlikely(!device
->use_global_bo_list
))
2387 pthread_mutex_lock(&bo_list
->mutex
);
2388 /* Loop the list backwards so we find the most recently added
2390 for(unsigned i
= bo_list
->list
.count
; i
-- > 0;) {
2391 if (bo_list
->list
.bos
[i
] == bo
) {
2392 bo_list
->list
.bos
[i
] = bo_list
->list
.bos
[bo_list
->list
.count
- 1];
2393 --bo_list
->list
.count
;
2397 pthread_mutex_unlock(&bo_list
->mutex
);
2401 radv_device_init_gs_info(struct radv_device
*device
)
2403 device
->gs_table_depth
= ac_get_gs_table_depth(device
->physical_device
->rad_info
.chip_class
,
2404 device
->physical_device
->rad_info
.family
);
2407 static int radv_get_device_extension_index(const char *name
)
2409 for (unsigned i
= 0; i
< RADV_DEVICE_EXTENSION_COUNT
; ++i
) {
2410 if (strcmp(name
, radv_device_extensions
[i
].extensionName
) == 0)
2417 radv_get_int_debug_option(const char *name
, int default_value
)
2424 result
= default_value
;
2428 result
= strtol(str
, &endptr
, 0);
2429 if (str
== endptr
) {
2430 /* No digits founs. */
2431 result
= default_value
;
2439 radv_device_init_dispatch(struct radv_device
*device
)
2441 const struct radv_instance
*instance
= device
->physical_device
->instance
;
2442 const struct radv_device_dispatch_table
*dispatch_table_layer
= NULL
;
2443 bool unchecked
= instance
->debug_flags
& RADV_DEBUG_ALL_ENTRYPOINTS
;
2444 int radv_thread_trace
= radv_get_int_debug_option("RADV_THREAD_TRACE", -1);
2446 if (radv_thread_trace
>= 0) {
2447 /* Use device entrypoints from the SQTT layer if enabled. */
2448 dispatch_table_layer
= &sqtt_device_dispatch_table
;
2451 for (unsigned i
= 0; i
< ARRAY_SIZE(device
->dispatch
.entrypoints
); i
++) {
2452 /* Vulkan requires that entrypoints for extensions which have not been
2453 * enabled must not be advertised.
2456 !radv_device_entrypoint_is_enabled(i
, instance
->apiVersion
,
2457 &instance
->enabled_extensions
,
2458 &device
->enabled_extensions
)) {
2459 device
->dispatch
.entrypoints
[i
] = NULL
;
2460 } else if (dispatch_table_layer
&&
2461 dispatch_table_layer
->entrypoints
[i
]) {
2462 device
->dispatch
.entrypoints
[i
] =
2463 dispatch_table_layer
->entrypoints
[i
];
2465 device
->dispatch
.entrypoints
[i
] =
2466 radv_device_dispatch_table
.entrypoints
[i
];
2472 radv_create_pthread_cond(pthread_cond_t
*cond
)
2474 pthread_condattr_t condattr
;
2475 if (pthread_condattr_init(&condattr
)) {
2476 return VK_ERROR_INITIALIZATION_FAILED
;
2479 if (pthread_condattr_setclock(&condattr
, CLOCK_MONOTONIC
)) {
2480 pthread_condattr_destroy(&condattr
);
2481 return VK_ERROR_INITIALIZATION_FAILED
;
2483 if (pthread_cond_init(cond
, &condattr
)) {
2484 pthread_condattr_destroy(&condattr
);
2485 return VK_ERROR_INITIALIZATION_FAILED
;
2487 pthread_condattr_destroy(&condattr
);
2492 check_physical_device_features(VkPhysicalDevice physicalDevice
,
2493 const VkPhysicalDeviceFeatures
*features
)
2495 RADV_FROM_HANDLE(radv_physical_device
, physical_device
, physicalDevice
);
2496 VkPhysicalDeviceFeatures supported_features
;
2497 radv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
2498 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
2499 VkBool32
*enabled_feature
= (VkBool32
*)features
;
2500 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
2501 for (uint32_t i
= 0; i
< num_features
; i
++) {
2502 if (enabled_feature
[i
] && !supported_feature
[i
])
2503 return vk_error(physical_device
->instance
, VK_ERROR_FEATURE_NOT_PRESENT
);
2509 static VkResult
radv_device_init_border_color(struct radv_device
*device
)
2511 device
->border_color_data
.bo
=
2512 device
->ws
->buffer_create(device
->ws
,
2513 RADV_BORDER_COLOR_BUFFER_SIZE
,
2516 RADEON_FLAG_CPU_ACCESS
|
2517 RADEON_FLAG_READ_ONLY
|
2518 RADEON_FLAG_NO_INTERPROCESS_SHARING
,
2519 RADV_BO_PRIORITY_SHADER
);
2521 if (device
->border_color_data
.bo
== NULL
)
2522 return vk_error(device
->physical_device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
2524 device
->border_color_data
.colors_gpu_ptr
=
2525 device
->ws
->buffer_map(device
->border_color_data
.bo
);
2526 if (!device
->border_color_data
.colors_gpu_ptr
)
2527 return vk_error(device
->physical_device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
2528 pthread_mutex_init(&device
->border_color_data
.mutex
, NULL
);
2533 static void radv_device_finish_border_color(struct radv_device
*device
)
2535 if (device
->border_color_data
.bo
) {
2536 device
->ws
->buffer_destroy(device
->border_color_data
.bo
);
2538 pthread_mutex_destroy(&device
->border_color_data
.mutex
);
2542 VkResult
radv_CreateDevice(
2543 VkPhysicalDevice physicalDevice
,
2544 const VkDeviceCreateInfo
* pCreateInfo
,
2545 const VkAllocationCallbacks
* pAllocator
,
2548 RADV_FROM_HANDLE(radv_physical_device
, physical_device
, physicalDevice
);
2550 struct radv_device
*device
;
2552 bool keep_shader_info
= false;
2553 bool robust_buffer_access
= false;
2554 bool overallocation_disallowed
= false;
2555 bool custom_border_colors
= false;
2557 /* Check enabled features */
2558 if (pCreateInfo
->pEnabledFeatures
) {
2559 result
= check_physical_device_features(physicalDevice
,
2560 pCreateInfo
->pEnabledFeatures
);
2561 if (result
!= VK_SUCCESS
)
2564 if (pCreateInfo
->pEnabledFeatures
->robustBufferAccess
)
2565 robust_buffer_access
= true;
2568 vk_foreach_struct_const(ext
, pCreateInfo
->pNext
) {
2569 switch (ext
->sType
) {
2570 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2
: {
2571 const VkPhysicalDeviceFeatures2
*features
= (const void *)ext
;
2572 result
= check_physical_device_features(physicalDevice
,
2573 &features
->features
);
2574 if (result
!= VK_SUCCESS
)
2577 if (features
->features
.robustBufferAccess
)
2578 robust_buffer_access
= true;
2581 case VK_STRUCTURE_TYPE_DEVICE_MEMORY_OVERALLOCATION_CREATE_INFO_AMD
: {
2582 const VkDeviceMemoryOverallocationCreateInfoAMD
*overallocation
= (const void *)ext
;
2583 if (overallocation
->overallocationBehavior
== VK_MEMORY_OVERALLOCATION_BEHAVIOR_DISALLOWED_AMD
)
2584 overallocation_disallowed
= true;
2587 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CUSTOM_BORDER_COLOR_FEATURES_EXT
: {
2588 const VkPhysicalDeviceCustomBorderColorFeaturesEXT
*border_color_features
= (const void *)ext
;
2589 custom_border_colors
= border_color_features
->customBorderColors
;
2597 device
= vk_zalloc2(&physical_device
->instance
->alloc
, pAllocator
,
2599 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
2601 return vk_error(physical_device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
2603 vk_device_init(&device
->vk
, pCreateInfo
,
2604 &physical_device
->instance
->alloc
, pAllocator
);
2606 device
->instance
= physical_device
->instance
;
2607 device
->physical_device
= physical_device
;
2609 device
->ws
= physical_device
->ws
;
2611 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
2612 const char *ext_name
= pCreateInfo
->ppEnabledExtensionNames
[i
];
2613 int index
= radv_get_device_extension_index(ext_name
);
2614 if (index
< 0 || !physical_device
->supported_extensions
.extensions
[index
]) {
2615 vk_free(&device
->vk
.alloc
, device
);
2616 return vk_error(physical_device
->instance
, VK_ERROR_EXTENSION_NOT_PRESENT
);
2619 device
->enabled_extensions
.extensions
[index
] = true;
2622 radv_device_init_dispatch(device
);
2624 keep_shader_info
= device
->enabled_extensions
.AMD_shader_info
;
2626 /* With update after bind we can't attach bo's to the command buffer
2627 * from the descriptor set anymore, so we have to use a global BO list.
2629 device
->use_global_bo_list
=
2630 (device
->instance
->perftest_flags
& RADV_PERFTEST_BO_LIST
) ||
2631 device
->enabled_extensions
.EXT_descriptor_indexing
||
2632 device
->enabled_extensions
.EXT_buffer_device_address
||
2633 device
->enabled_extensions
.KHR_buffer_device_address
;
2635 device
->robust_buffer_access
= robust_buffer_access
;
2637 mtx_init(&device
->shader_slab_mutex
, mtx_plain
);
2638 list_inithead(&device
->shader_slabs
);
2640 device
->overallocation_disallowed
= overallocation_disallowed
;
2641 mtx_init(&device
->overallocation_mutex
, mtx_plain
);
2643 radv_bo_list_init(&device
->bo_list
);
2645 for (unsigned i
= 0; i
< pCreateInfo
->queueCreateInfoCount
; i
++) {
2646 const VkDeviceQueueCreateInfo
*queue_create
= &pCreateInfo
->pQueueCreateInfos
[i
];
2647 uint32_t qfi
= queue_create
->queueFamilyIndex
;
2648 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*global_priority
=
2649 vk_find_struct_const(queue_create
->pNext
, DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT
);
2651 assert(!global_priority
|| device
->physical_device
->rad_info
.has_ctx_priority
);
2653 device
->queues
[qfi
] = vk_alloc(&device
->vk
.alloc
,
2654 queue_create
->queueCount
* sizeof(struct radv_queue
), 8, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
2655 if (!device
->queues
[qfi
]) {
2656 result
= VK_ERROR_OUT_OF_HOST_MEMORY
;
2660 memset(device
->queues
[qfi
], 0, queue_create
->queueCount
* sizeof(struct radv_queue
));
2662 device
->queue_count
[qfi
] = queue_create
->queueCount
;
2664 for (unsigned q
= 0; q
< queue_create
->queueCount
; q
++) {
2665 result
= radv_queue_init(device
, &device
->queues
[qfi
][q
],
2666 qfi
, q
, queue_create
->flags
,
2668 if (result
!= VK_SUCCESS
)
2673 device
->pbb_allowed
= device
->physical_device
->rad_info
.chip_class
>= GFX9
&&
2674 !(device
->instance
->debug_flags
& RADV_DEBUG_NOBINNING
);
2676 /* Disable DFSM by default. As of 2019-09-15 Talos on Low is still 3% slower on Raven. */
2677 device
->dfsm_allowed
= device
->pbb_allowed
&&
2678 (device
->instance
->perftest_flags
& RADV_PERFTEST_DFSM
);
2680 device
->always_use_syncobj
= device
->physical_device
->rad_info
.has_syncobj_wait_for_submit
;
2682 /* The maximum number of scratch waves. Scratch space isn't divided
2683 * evenly between CUs. The number is only a function of the number of CUs.
2684 * We can decrease the constant to decrease the scratch buffer size.
2686 * sctx->scratch_waves must be >= the maximum possible size of
2687 * 1 threadgroup, so that the hw doesn't hang from being unable
2690 * The recommended value is 4 per CU at most. Higher numbers don't
2691 * bring much benefit, but they still occupy chip resources (think
2692 * async compute). I've seen ~2% performance difference between 4 and 32.
2694 uint32_t max_threads_per_block
= 2048;
2695 device
->scratch_waves
= MAX2(32 * physical_device
->rad_info
.num_good_compute_units
,
2696 max_threads_per_block
/ 64);
2698 device
->dispatch_initiator
= S_00B800_COMPUTE_SHADER_EN(1);
2700 if (device
->physical_device
->rad_info
.chip_class
>= GFX7
) {
2701 /* If the KMD allows it (there is a KMD hw register for it),
2702 * allow launching waves out-of-order.
2704 device
->dispatch_initiator
|= S_00B800_ORDER_MODE(1);
2707 radv_device_init_gs_info(device
);
2709 device
->tess_offchip_block_dw_size
=
2710 device
->physical_device
->rad_info
.family
== CHIP_HAWAII
? 4096 : 8192;
2712 if (getenv("RADV_TRACE_FILE")) {
2713 const char *filename
= getenv("RADV_TRACE_FILE");
2715 keep_shader_info
= true;
2717 if (!radv_init_trace(device
))
2720 fprintf(stderr
, "*****************************************************************************\n");
2721 fprintf(stderr
, "* WARNING: RADV_TRACE_FILE is costly and should only be used for debugging! *\n");
2722 fprintf(stderr
, "*****************************************************************************\n");
2724 fprintf(stderr
, "Trace file will be dumped to %s\n", filename
);
2725 radv_dump_enabled_options(device
, stderr
);
2728 int radv_thread_trace
= radv_get_int_debug_option("RADV_THREAD_TRACE", -1);
2729 if (radv_thread_trace
>= 0) {
2730 fprintf(stderr
, "*************************************************\n");
2731 fprintf(stderr
, "* WARNING: Thread trace support is experimental *\n");
2732 fprintf(stderr
, "*************************************************\n");
2734 if (device
->physical_device
->rad_info
.chip_class
< GFX8
) {
2735 fprintf(stderr
, "GPU hardware not supported: refer to "
2736 "the RGP documentation for the list of "
2737 "supported GPUs!\n");
2741 /* Default buffer size set to 1MB per SE. */
2742 device
->thread_trace_buffer_size
=
2743 radv_get_int_debug_option("RADV_THREAD_TRACE_BUFFER_SIZE", 1024 * 1024);
2744 device
->thread_trace_start_frame
= radv_thread_trace
;
2746 if (!radv_thread_trace_init(device
))
2750 device
->keep_shader_info
= keep_shader_info
;
2751 result
= radv_device_init_meta(device
);
2752 if (result
!= VK_SUCCESS
)
2755 radv_device_init_msaa(device
);
2757 /* If the border color extension is enabled, let's create the buffer we need. */
2758 if (custom_border_colors
) {
2759 result
= radv_device_init_border_color(device
);
2760 if (result
!= VK_SUCCESS
)
2764 for (int family
= 0; family
< RADV_MAX_QUEUE_FAMILIES
; ++family
) {
2765 device
->empty_cs
[family
] = device
->ws
->cs_create(device
->ws
, family
);
2766 if (!device
->empty_cs
[family
])
2770 case RADV_QUEUE_GENERAL
:
2771 radeon_emit(device
->empty_cs
[family
], PKT3(PKT3_CONTEXT_CONTROL
, 1, 0));
2772 radeon_emit(device
->empty_cs
[family
], CC0_UPDATE_LOAD_ENABLES(1));
2773 radeon_emit(device
->empty_cs
[family
], CC1_UPDATE_SHADOW_ENABLES(1));
2775 case RADV_QUEUE_COMPUTE
:
2776 radeon_emit(device
->empty_cs
[family
], PKT3(PKT3_NOP
, 0, 0));
2777 radeon_emit(device
->empty_cs
[family
], 0);
2781 result
= device
->ws
->cs_finalize(device
->empty_cs
[family
]);
2782 if (result
!= VK_SUCCESS
)
2786 if (device
->physical_device
->rad_info
.chip_class
>= GFX7
)
2787 cik_create_gfx_config(device
);
2789 VkPipelineCacheCreateInfo ci
;
2790 ci
.sType
= VK_STRUCTURE_TYPE_PIPELINE_CACHE_CREATE_INFO
;
2793 ci
.pInitialData
= NULL
;
2794 ci
.initialDataSize
= 0;
2796 result
= radv_CreatePipelineCache(radv_device_to_handle(device
),
2798 if (result
!= VK_SUCCESS
)
2801 device
->mem_cache
= radv_pipeline_cache_from_handle(pc
);
2803 result
= radv_create_pthread_cond(&device
->timeline_cond
);
2804 if (result
!= VK_SUCCESS
)
2805 goto fail_mem_cache
;
2807 device
->force_aniso
=
2808 MIN2(16, radv_get_int_debug_option("RADV_TEX_ANISO", -1));
2809 if (device
->force_aniso
>= 0) {
2810 fprintf(stderr
, "radv: Forcing anisotropy filter to %ix\n",
2811 1 << util_logbase2(device
->force_aniso
));
2814 *pDevice
= radv_device_to_handle(device
);
2818 radv_DestroyPipelineCache(radv_device_to_handle(device
), pc
, NULL
);
2820 radv_device_finish_meta(device
);
2822 radv_bo_list_finish(&device
->bo_list
);
2824 radv_thread_trace_finish(device
);
2826 if (device
->trace_bo
)
2827 device
->ws
->buffer_destroy(device
->trace_bo
);
2829 if (device
->gfx_init
)
2830 device
->ws
->buffer_destroy(device
->gfx_init
);
2832 radv_device_finish_border_color(device
);
2834 for (unsigned i
= 0; i
< RADV_MAX_QUEUE_FAMILIES
; i
++) {
2835 for (unsigned q
= 0; q
< device
->queue_count
[i
]; q
++)
2836 radv_queue_finish(&device
->queues
[i
][q
]);
2837 if (device
->queue_count
[i
])
2838 vk_free(&device
->vk
.alloc
, device
->queues
[i
]);
2841 vk_free(&device
->vk
.alloc
, device
);
2845 void radv_DestroyDevice(
2847 const VkAllocationCallbacks
* pAllocator
)
2849 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2854 if (device
->trace_bo
)
2855 device
->ws
->buffer_destroy(device
->trace_bo
);
2857 if (device
->gfx_init
)
2858 device
->ws
->buffer_destroy(device
->gfx_init
);
2860 radv_device_finish_border_color(device
);
2862 for (unsigned i
= 0; i
< RADV_MAX_QUEUE_FAMILIES
; i
++) {
2863 for (unsigned q
= 0; q
< device
->queue_count
[i
]; q
++)
2864 radv_queue_finish(&device
->queues
[i
][q
]);
2865 if (device
->queue_count
[i
])
2866 vk_free(&device
->vk
.alloc
, device
->queues
[i
]);
2867 if (device
->empty_cs
[i
])
2868 device
->ws
->cs_destroy(device
->empty_cs
[i
]);
2870 radv_device_finish_meta(device
);
2872 VkPipelineCache pc
= radv_pipeline_cache_to_handle(device
->mem_cache
);
2873 radv_DestroyPipelineCache(radv_device_to_handle(device
), pc
, NULL
);
2875 radv_destroy_shader_slabs(device
);
2877 pthread_cond_destroy(&device
->timeline_cond
);
2878 radv_bo_list_finish(&device
->bo_list
);
2880 radv_thread_trace_finish(device
);
2882 vk_free(&device
->vk
.alloc
, device
);
2885 VkResult
radv_EnumerateInstanceLayerProperties(
2886 uint32_t* pPropertyCount
,
2887 VkLayerProperties
* pProperties
)
2889 if (pProperties
== NULL
) {
2890 *pPropertyCount
= 0;
2894 /* None supported at this time */
2895 return vk_error(NULL
, VK_ERROR_LAYER_NOT_PRESENT
);
2898 VkResult
radv_EnumerateDeviceLayerProperties(
2899 VkPhysicalDevice physicalDevice
,
2900 uint32_t* pPropertyCount
,
2901 VkLayerProperties
* pProperties
)
2903 if (pProperties
== NULL
) {
2904 *pPropertyCount
= 0;
2908 /* None supported at this time */
2909 return vk_error(NULL
, VK_ERROR_LAYER_NOT_PRESENT
);
2912 void radv_GetDeviceQueue2(
2914 const VkDeviceQueueInfo2
* pQueueInfo
,
2917 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2918 struct radv_queue
*queue
;
2920 queue
= &device
->queues
[pQueueInfo
->queueFamilyIndex
][pQueueInfo
->queueIndex
];
2921 if (pQueueInfo
->flags
!= queue
->flags
) {
2922 /* From the Vulkan 1.1.70 spec:
2924 * "The queue returned by vkGetDeviceQueue2 must have the same
2925 * flags value from this structure as that used at device
2926 * creation time in a VkDeviceQueueCreateInfo instance. If no
2927 * matching flags were specified at device creation time then
2928 * pQueue will return VK_NULL_HANDLE."
2930 *pQueue
= VK_NULL_HANDLE
;
2934 *pQueue
= radv_queue_to_handle(queue
);
2937 void radv_GetDeviceQueue(
2939 uint32_t queueFamilyIndex
,
2940 uint32_t queueIndex
,
2943 const VkDeviceQueueInfo2 info
= (VkDeviceQueueInfo2
) {
2944 .sType
= VK_STRUCTURE_TYPE_DEVICE_QUEUE_INFO_2
,
2945 .queueFamilyIndex
= queueFamilyIndex
,
2946 .queueIndex
= queueIndex
2949 radv_GetDeviceQueue2(_device
, &info
, pQueue
);
2953 fill_geom_tess_rings(struct radv_queue
*queue
,
2955 bool add_sample_positions
,
2956 uint32_t esgs_ring_size
,
2957 struct radeon_winsys_bo
*esgs_ring_bo
,
2958 uint32_t gsvs_ring_size
,
2959 struct radeon_winsys_bo
*gsvs_ring_bo
,
2960 uint32_t tess_factor_ring_size
,
2961 uint32_t tess_offchip_ring_offset
,
2962 uint32_t tess_offchip_ring_size
,
2963 struct radeon_winsys_bo
*tess_rings_bo
)
2965 uint32_t *desc
= &map
[4];
2968 uint64_t esgs_va
= radv_buffer_get_va(esgs_ring_bo
);
2970 /* stride 0, num records - size, add tid, swizzle, elsize4,
2973 desc
[1] = S_008F04_BASE_ADDRESS_HI(esgs_va
>> 32) |
2974 S_008F04_SWIZZLE_ENABLE(true);
2975 desc
[2] = esgs_ring_size
;
2976 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
2977 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
2978 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
2979 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
2980 S_008F0C_INDEX_STRIDE(3) |
2981 S_008F0C_ADD_TID_ENABLE(1);
2983 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
2984 desc
[3] |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT
) |
2985 S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED
) |
2986 S_008F0C_RESOURCE_LEVEL(1);
2988 desc
[3] |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
2989 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
2990 S_008F0C_ELEMENT_SIZE(1);
2993 /* GS entry for ES->GS ring */
2994 /* stride 0, num records - size, elsize0,
2997 desc
[5] = S_008F04_BASE_ADDRESS_HI(esgs_va
>> 32);
2998 desc
[6] = esgs_ring_size
;
2999 desc
[7] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
3000 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
3001 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
3002 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
);
3004 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
3005 desc
[7] |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT
) |
3006 S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED
) |
3007 S_008F0C_RESOURCE_LEVEL(1);
3009 desc
[7] |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
3010 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
);
3017 uint64_t gsvs_va
= radv_buffer_get_va(gsvs_ring_bo
);
3019 /* VS entry for GS->VS ring */
3020 /* stride 0, num records - size, elsize0,
3023 desc
[1] = S_008F04_BASE_ADDRESS_HI(gsvs_va
>> 32);
3024 desc
[2] = gsvs_ring_size
;
3025 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
3026 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
3027 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
3028 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
);
3030 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
3031 desc
[3] |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT
) |
3032 S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED
) |
3033 S_008F0C_RESOURCE_LEVEL(1);
3035 desc
[3] |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
3036 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
);
3039 /* stride gsvs_itemsize, num records 64
3040 elsize 4, index stride 16 */
3041 /* shader will patch stride and desc[2] */
3043 desc
[5] = S_008F04_BASE_ADDRESS_HI(gsvs_va
>> 32) |
3044 S_008F04_SWIZZLE_ENABLE(1);
3046 desc
[7] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
3047 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
3048 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
3049 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
3050 S_008F0C_INDEX_STRIDE(1) |
3051 S_008F0C_ADD_TID_ENABLE(true);
3053 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
3054 desc
[7] |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT
) |
3055 S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED
) |
3056 S_008F0C_RESOURCE_LEVEL(1);
3058 desc
[7] |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
3059 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
3060 S_008F0C_ELEMENT_SIZE(1);
3067 if (tess_rings_bo
) {
3068 uint64_t tess_va
= radv_buffer_get_va(tess_rings_bo
);
3069 uint64_t tess_offchip_va
= tess_va
+ tess_offchip_ring_offset
;
3072 desc
[1] = S_008F04_BASE_ADDRESS_HI(tess_va
>> 32);
3073 desc
[2] = tess_factor_ring_size
;
3074 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
3075 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
3076 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
3077 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
);
3079 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
3080 desc
[3] |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT
) |
3081 S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_RAW
) |
3082 S_008F0C_RESOURCE_LEVEL(1);
3084 desc
[3] |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
3085 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
);
3088 desc
[4] = tess_offchip_va
;
3089 desc
[5] = S_008F04_BASE_ADDRESS_HI(tess_offchip_va
>> 32);
3090 desc
[6] = tess_offchip_ring_size
;
3091 desc
[7] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
3092 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
3093 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
3094 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
);
3096 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
3097 desc
[7] |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT
) |
3098 S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_RAW
) |
3099 S_008F0C_RESOURCE_LEVEL(1);
3101 desc
[7] |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
3102 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
);
3108 if (add_sample_positions
) {
3109 /* add sample positions after all rings */
3110 memcpy(desc
, queue
->device
->sample_locations_1x
, 8);
3112 memcpy(desc
, queue
->device
->sample_locations_2x
, 16);
3114 memcpy(desc
, queue
->device
->sample_locations_4x
, 32);
3116 memcpy(desc
, queue
->device
->sample_locations_8x
, 64);
3121 radv_get_hs_offchip_param(struct radv_device
*device
, uint32_t *max_offchip_buffers_p
)
3123 bool double_offchip_buffers
= device
->physical_device
->rad_info
.chip_class
>= GFX7
&&
3124 device
->physical_device
->rad_info
.family
!= CHIP_CARRIZO
&&
3125 device
->physical_device
->rad_info
.family
!= CHIP_STONEY
;
3126 unsigned max_offchip_buffers_per_se
= double_offchip_buffers
? 128 : 64;
3127 unsigned max_offchip_buffers
;
3128 unsigned offchip_granularity
;
3129 unsigned hs_offchip_param
;
3133 * This must be one less than the maximum number due to a hw limitation.
3134 * Various hardware bugs need thGFX7
3137 * Vega10 should limit max_offchip_buffers to 508 (4 * 127).
3138 * Gfx7 should limit max_offchip_buffers to 508
3139 * Gfx6 should limit max_offchip_buffers to 126 (2 * 63)
3141 * Follow AMDVLK here.
3143 if (device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
3144 max_offchip_buffers_per_se
= 256;
3145 } else if (device
->physical_device
->rad_info
.family
== CHIP_VEGA10
||
3146 device
->physical_device
->rad_info
.chip_class
== GFX7
||
3147 device
->physical_device
->rad_info
.chip_class
== GFX6
)
3148 --max_offchip_buffers_per_se
;
3150 max_offchip_buffers
= max_offchip_buffers_per_se
*
3151 device
->physical_device
->rad_info
.max_se
;
3153 /* Hawaii has a bug with offchip buffers > 256 that can be worked
3154 * around by setting 4K granularity.
3156 if (device
->tess_offchip_block_dw_size
== 4096) {
3157 assert(device
->physical_device
->rad_info
.family
== CHIP_HAWAII
);
3158 offchip_granularity
= V_03093C_X_4K_DWORDS
;
3160 assert(device
->tess_offchip_block_dw_size
== 8192);
3161 offchip_granularity
= V_03093C_X_8K_DWORDS
;
3164 switch (device
->physical_device
->rad_info
.chip_class
) {
3166 max_offchip_buffers
= MIN2(max_offchip_buffers
, 126);
3171 max_offchip_buffers
= MIN2(max_offchip_buffers
, 508);
3179 *max_offchip_buffers_p
= max_offchip_buffers
;
3180 if (device
->physical_device
->rad_info
.chip_class
>= GFX10_3
) {
3181 hs_offchip_param
= S_03093C_OFFCHIP_BUFFERING_GFX103(max_offchip_buffers
- 1) |
3182 S_03093C_OFFCHIP_GRANULARITY_GFX103(offchip_granularity
);
3183 } else if (device
->physical_device
->rad_info
.chip_class
>= GFX7
) {
3184 if (device
->physical_device
->rad_info
.chip_class
>= GFX8
)
3185 --max_offchip_buffers
;
3187 S_03093C_OFFCHIP_BUFFERING(max_offchip_buffers
) |
3188 S_03093C_OFFCHIP_GRANULARITY(offchip_granularity
);
3191 S_0089B0_OFFCHIP_BUFFERING(max_offchip_buffers
);
3193 return hs_offchip_param
;
3197 radv_emit_gs_ring_sizes(struct radv_queue
*queue
, struct radeon_cmdbuf
*cs
,
3198 struct radeon_winsys_bo
*esgs_ring_bo
,
3199 uint32_t esgs_ring_size
,
3200 struct radeon_winsys_bo
*gsvs_ring_bo
,
3201 uint32_t gsvs_ring_size
)
3203 if (!esgs_ring_bo
&& !gsvs_ring_bo
)
3207 radv_cs_add_buffer(queue
->device
->ws
, cs
, esgs_ring_bo
);
3210 radv_cs_add_buffer(queue
->device
->ws
, cs
, gsvs_ring_bo
);
3212 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX7
) {
3213 radeon_set_uconfig_reg_seq(cs
, R_030900_VGT_ESGS_RING_SIZE
, 2);
3214 radeon_emit(cs
, esgs_ring_size
>> 8);
3215 radeon_emit(cs
, gsvs_ring_size
>> 8);
3217 radeon_set_config_reg_seq(cs
, R_0088C8_VGT_ESGS_RING_SIZE
, 2);
3218 radeon_emit(cs
, esgs_ring_size
>> 8);
3219 radeon_emit(cs
, gsvs_ring_size
>> 8);
3224 radv_emit_tess_factor_ring(struct radv_queue
*queue
, struct radeon_cmdbuf
*cs
,
3225 unsigned hs_offchip_param
, unsigned tf_ring_size
,
3226 struct radeon_winsys_bo
*tess_rings_bo
)
3233 tf_va
= radv_buffer_get_va(tess_rings_bo
);
3235 radv_cs_add_buffer(queue
->device
->ws
, cs
, tess_rings_bo
);
3237 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX7
) {
3238 radeon_set_uconfig_reg(cs
, R_030938_VGT_TF_RING_SIZE
,
3239 S_030938_SIZE(tf_ring_size
/ 4));
3240 radeon_set_uconfig_reg(cs
, R_030940_VGT_TF_MEMORY_BASE
,
3243 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
3244 radeon_set_uconfig_reg(cs
, R_030984_VGT_TF_MEMORY_BASE_HI_UMD
,
3245 S_030984_BASE_HI(tf_va
>> 40));
3246 } else if (queue
->device
->physical_device
->rad_info
.chip_class
== GFX9
) {
3247 radeon_set_uconfig_reg(cs
, R_030944_VGT_TF_MEMORY_BASE_HI
,
3248 S_030944_BASE_HI(tf_va
>> 40));
3250 radeon_set_uconfig_reg(cs
, R_03093C_VGT_HS_OFFCHIP_PARAM
,
3253 radeon_set_config_reg(cs
, R_008988_VGT_TF_RING_SIZE
,
3254 S_008988_SIZE(tf_ring_size
/ 4));
3255 radeon_set_config_reg(cs
, R_0089B8_VGT_TF_MEMORY_BASE
,
3257 radeon_set_config_reg(cs
, R_0089B0_VGT_HS_OFFCHIP_PARAM
,
3263 radv_emit_graphics_scratch(struct radv_queue
*queue
, struct radeon_cmdbuf
*cs
,
3264 uint32_t size_per_wave
, uint32_t waves
,
3265 struct radeon_winsys_bo
*scratch_bo
)
3267 if (queue
->queue_family_index
!= RADV_QUEUE_GENERAL
)
3273 radv_cs_add_buffer(queue
->device
->ws
, cs
, scratch_bo
);
3275 radeon_set_context_reg(cs
, R_0286E8_SPI_TMPRING_SIZE
,
3276 S_0286E8_WAVES(waves
) |
3277 S_0286E8_WAVESIZE(round_up_u32(size_per_wave
, 1024)));
3281 radv_emit_compute_scratch(struct radv_queue
*queue
, struct radeon_cmdbuf
*cs
,
3282 uint32_t size_per_wave
, uint32_t waves
,
3283 struct radeon_winsys_bo
*compute_scratch_bo
)
3285 uint64_t scratch_va
;
3287 if (!compute_scratch_bo
)
3290 scratch_va
= radv_buffer_get_va(compute_scratch_bo
);
3292 radv_cs_add_buffer(queue
->device
->ws
, cs
, compute_scratch_bo
);
3294 radeon_set_sh_reg_seq(cs
, R_00B900_COMPUTE_USER_DATA_0
, 2);
3295 radeon_emit(cs
, scratch_va
);
3296 radeon_emit(cs
, S_008F04_BASE_ADDRESS_HI(scratch_va
>> 32) |
3297 S_008F04_SWIZZLE_ENABLE(1));
3299 radeon_set_sh_reg(cs
, R_00B860_COMPUTE_TMPRING_SIZE
,
3300 S_00B860_WAVES(waves
) |
3301 S_00B860_WAVESIZE(round_up_u32(size_per_wave
, 1024)));
3305 radv_emit_global_shader_pointers(struct radv_queue
*queue
,
3306 struct radeon_cmdbuf
*cs
,
3307 struct radeon_winsys_bo
*descriptor_bo
)
3314 va
= radv_buffer_get_va(descriptor_bo
);
3316 radv_cs_add_buffer(queue
->device
->ws
, cs
, descriptor_bo
);
3318 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
3319 uint32_t regs
[] = {R_00B030_SPI_SHADER_USER_DATA_PS_0
,
3320 R_00B130_SPI_SHADER_USER_DATA_VS_0
,
3321 R_00B208_SPI_SHADER_USER_DATA_ADDR_LO_GS
,
3322 R_00B408_SPI_SHADER_USER_DATA_ADDR_LO_HS
};
3324 for (int i
= 0; i
< ARRAY_SIZE(regs
); ++i
) {
3325 radv_emit_shader_pointer(queue
->device
, cs
, regs
[i
],
3328 } else if (queue
->device
->physical_device
->rad_info
.chip_class
== GFX9
) {
3329 uint32_t regs
[] = {R_00B030_SPI_SHADER_USER_DATA_PS_0
,
3330 R_00B130_SPI_SHADER_USER_DATA_VS_0
,
3331 R_00B208_SPI_SHADER_USER_DATA_ADDR_LO_GS
,
3332 R_00B408_SPI_SHADER_USER_DATA_ADDR_LO_HS
};
3334 for (int i
= 0; i
< ARRAY_SIZE(regs
); ++i
) {
3335 radv_emit_shader_pointer(queue
->device
, cs
, regs
[i
],
3339 uint32_t regs
[] = {R_00B030_SPI_SHADER_USER_DATA_PS_0
,
3340 R_00B130_SPI_SHADER_USER_DATA_VS_0
,
3341 R_00B230_SPI_SHADER_USER_DATA_GS_0
,
3342 R_00B330_SPI_SHADER_USER_DATA_ES_0
,
3343 R_00B430_SPI_SHADER_USER_DATA_HS_0
,
3344 R_00B530_SPI_SHADER_USER_DATA_LS_0
};
3346 for (int i
= 0; i
< ARRAY_SIZE(regs
); ++i
) {
3347 radv_emit_shader_pointer(queue
->device
, cs
, regs
[i
],
3354 radv_init_graphics_state(struct radeon_cmdbuf
*cs
, struct radv_queue
*queue
)
3356 struct radv_device
*device
= queue
->device
;
3358 if (device
->gfx_init
) {
3359 uint64_t va
= radv_buffer_get_va(device
->gfx_init
);
3361 radeon_emit(cs
, PKT3(PKT3_INDIRECT_BUFFER_CIK
, 2, 0));
3362 radeon_emit(cs
, va
);
3363 radeon_emit(cs
, va
>> 32);
3364 radeon_emit(cs
, device
->gfx_init_size_dw
& 0xffff);
3366 radv_cs_add_buffer(device
->ws
, cs
, device
->gfx_init
);
3368 si_emit_graphics(device
, cs
);
3373 radv_init_compute_state(struct radeon_cmdbuf
*cs
, struct radv_queue
*queue
)
3375 struct radv_physical_device
*physical_device
= queue
->device
->physical_device
;
3376 si_emit_compute(physical_device
, cs
);
3380 radv_get_preamble_cs(struct radv_queue
*queue
,
3381 uint32_t scratch_size_per_wave
,
3382 uint32_t scratch_waves
,
3383 uint32_t compute_scratch_size_per_wave
,
3384 uint32_t compute_scratch_waves
,
3385 uint32_t esgs_ring_size
,
3386 uint32_t gsvs_ring_size
,
3387 bool needs_tess_rings
,
3390 bool needs_sample_positions
,
3391 struct radeon_cmdbuf
**initial_full_flush_preamble_cs
,
3392 struct radeon_cmdbuf
**initial_preamble_cs
,
3393 struct radeon_cmdbuf
**continue_preamble_cs
)
3395 struct radeon_winsys_bo
*scratch_bo
= NULL
;
3396 struct radeon_winsys_bo
*descriptor_bo
= NULL
;
3397 struct radeon_winsys_bo
*compute_scratch_bo
= NULL
;
3398 struct radeon_winsys_bo
*esgs_ring_bo
= NULL
;
3399 struct radeon_winsys_bo
*gsvs_ring_bo
= NULL
;
3400 struct radeon_winsys_bo
*tess_rings_bo
= NULL
;
3401 struct radeon_winsys_bo
*gds_bo
= NULL
;
3402 struct radeon_winsys_bo
*gds_oa_bo
= NULL
;
3403 struct radeon_cmdbuf
*dest_cs
[3] = {0};
3404 bool add_tess_rings
= false, add_gds
= false, add_gds_oa
= false, add_sample_positions
= false;
3405 unsigned tess_factor_ring_size
= 0, tess_offchip_ring_size
= 0;
3406 unsigned max_offchip_buffers
;
3407 unsigned hs_offchip_param
= 0;
3408 unsigned tess_offchip_ring_offset
;
3409 uint32_t ring_bo_flags
= RADEON_FLAG_NO_CPU_ACCESS
| RADEON_FLAG_NO_INTERPROCESS_SHARING
;
3410 if (!queue
->has_tess_rings
) {
3411 if (needs_tess_rings
)
3412 add_tess_rings
= true;
3414 if (!queue
->has_gds
) {
3418 if (!queue
->has_gds_oa
) {
3422 if (!queue
->has_sample_positions
) {
3423 if (needs_sample_positions
)
3424 add_sample_positions
= true;
3426 tess_factor_ring_size
= 32768 * queue
->device
->physical_device
->rad_info
.max_se
;
3427 hs_offchip_param
= radv_get_hs_offchip_param(queue
->device
,
3428 &max_offchip_buffers
);
3429 tess_offchip_ring_offset
= align(tess_factor_ring_size
, 64 * 1024);
3430 tess_offchip_ring_size
= max_offchip_buffers
*
3431 queue
->device
->tess_offchip_block_dw_size
* 4;
3433 scratch_size_per_wave
= MAX2(scratch_size_per_wave
, queue
->scratch_size_per_wave
);
3434 if (scratch_size_per_wave
)
3435 scratch_waves
= MIN2(scratch_waves
, UINT32_MAX
/ scratch_size_per_wave
);
3439 compute_scratch_size_per_wave
= MAX2(compute_scratch_size_per_wave
, queue
->compute_scratch_size_per_wave
);
3440 if (compute_scratch_size_per_wave
)
3441 compute_scratch_waves
= MIN2(compute_scratch_waves
, UINT32_MAX
/ compute_scratch_size_per_wave
);
3443 compute_scratch_waves
= 0;
3445 if (scratch_size_per_wave
<= queue
->scratch_size_per_wave
&&
3446 scratch_waves
<= queue
->scratch_waves
&&
3447 compute_scratch_size_per_wave
<= queue
->compute_scratch_size_per_wave
&&
3448 compute_scratch_waves
<= queue
->compute_scratch_waves
&&
3449 esgs_ring_size
<= queue
->esgs_ring_size
&&
3450 gsvs_ring_size
<= queue
->gsvs_ring_size
&&
3451 !add_tess_rings
&& !add_gds
&& !add_gds_oa
&& !add_sample_positions
&&
3452 queue
->initial_preamble_cs
) {
3453 *initial_full_flush_preamble_cs
= queue
->initial_full_flush_preamble_cs
;
3454 *initial_preamble_cs
= queue
->initial_preamble_cs
;
3455 *continue_preamble_cs
= queue
->continue_preamble_cs
;
3456 if (!scratch_size_per_wave
&& !compute_scratch_size_per_wave
&&
3457 !esgs_ring_size
&& !gsvs_ring_size
&& !needs_tess_rings
&&
3458 !needs_gds
&& !needs_gds_oa
&& !needs_sample_positions
)
3459 *continue_preamble_cs
= NULL
;
3463 uint32_t scratch_size
= scratch_size_per_wave
* scratch_waves
;
3464 uint32_t queue_scratch_size
= queue
->scratch_size_per_wave
* queue
->scratch_waves
;
3465 if (scratch_size
> queue_scratch_size
) {
3466 scratch_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
3471 RADV_BO_PRIORITY_SCRATCH
);
3475 scratch_bo
= queue
->scratch_bo
;
3477 uint32_t compute_scratch_size
= compute_scratch_size_per_wave
* compute_scratch_waves
;
3478 uint32_t compute_queue_scratch_size
= queue
->compute_scratch_size_per_wave
* queue
->compute_scratch_waves
;
3479 if (compute_scratch_size
> compute_queue_scratch_size
) {
3480 compute_scratch_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
3481 compute_scratch_size
,
3485 RADV_BO_PRIORITY_SCRATCH
);
3486 if (!compute_scratch_bo
)
3490 compute_scratch_bo
= queue
->compute_scratch_bo
;
3492 if (esgs_ring_size
> queue
->esgs_ring_size
) {
3493 esgs_ring_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
3498 RADV_BO_PRIORITY_SCRATCH
);
3502 esgs_ring_bo
= queue
->esgs_ring_bo
;
3503 esgs_ring_size
= queue
->esgs_ring_size
;
3506 if (gsvs_ring_size
> queue
->gsvs_ring_size
) {
3507 gsvs_ring_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
3512 RADV_BO_PRIORITY_SCRATCH
);
3516 gsvs_ring_bo
= queue
->gsvs_ring_bo
;
3517 gsvs_ring_size
= queue
->gsvs_ring_size
;
3520 if (add_tess_rings
) {
3521 tess_rings_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
3522 tess_offchip_ring_offset
+ tess_offchip_ring_size
,
3526 RADV_BO_PRIORITY_SCRATCH
);
3530 tess_rings_bo
= queue
->tess_rings_bo
;
3534 assert(queue
->device
->physical_device
->rad_info
.chip_class
>= GFX10
);
3536 /* 4 streamout GDS counters.
3537 * We need 256B (64 dw) of GDS, otherwise streamout hangs.
3539 gds_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
3543 RADV_BO_PRIORITY_SCRATCH
);
3547 gds_bo
= queue
->gds_bo
;
3551 assert(queue
->device
->physical_device
->rad_info
.chip_class
>= GFX10
);
3553 gds_oa_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
3557 RADV_BO_PRIORITY_SCRATCH
);
3561 gds_oa_bo
= queue
->gds_oa_bo
;
3564 if (scratch_bo
!= queue
->scratch_bo
||
3565 esgs_ring_bo
!= queue
->esgs_ring_bo
||
3566 gsvs_ring_bo
!= queue
->gsvs_ring_bo
||
3567 tess_rings_bo
!= queue
->tess_rings_bo
||
3568 add_sample_positions
) {
3570 if (gsvs_ring_bo
|| esgs_ring_bo
||
3571 tess_rings_bo
|| add_sample_positions
) {
3572 size
= 112; /* 2 dword + 2 padding + 4 dword * 6 */
3573 if (add_sample_positions
)
3574 size
+= 128; /* 64+32+16+8 = 120 bytes */
3576 else if (scratch_bo
)
3577 size
= 8; /* 2 dword */
3579 descriptor_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
3583 RADEON_FLAG_CPU_ACCESS
|
3584 RADEON_FLAG_NO_INTERPROCESS_SHARING
|
3585 RADEON_FLAG_READ_ONLY
,
3586 RADV_BO_PRIORITY_DESCRIPTOR
);
3590 descriptor_bo
= queue
->descriptor_bo
;
3592 if (descriptor_bo
!= queue
->descriptor_bo
) {
3593 uint32_t *map
= (uint32_t*)queue
->device
->ws
->buffer_map(descriptor_bo
);
3598 uint64_t scratch_va
= radv_buffer_get_va(scratch_bo
);
3599 uint32_t rsrc1
= S_008F04_BASE_ADDRESS_HI(scratch_va
>> 32) |
3600 S_008F04_SWIZZLE_ENABLE(1);
3601 map
[0] = scratch_va
;
3605 if (esgs_ring_bo
|| gsvs_ring_bo
|| tess_rings_bo
|| add_sample_positions
)
3606 fill_geom_tess_rings(queue
, map
, add_sample_positions
,
3607 esgs_ring_size
, esgs_ring_bo
,
3608 gsvs_ring_size
, gsvs_ring_bo
,
3609 tess_factor_ring_size
,
3610 tess_offchip_ring_offset
,
3611 tess_offchip_ring_size
,
3614 queue
->device
->ws
->buffer_unmap(descriptor_bo
);
3617 for(int i
= 0; i
< 3; ++i
) {
3618 struct radeon_cmdbuf
*cs
= NULL
;
3619 cs
= queue
->device
->ws
->cs_create(queue
->device
->ws
,
3620 queue
->queue_family_index
? RING_COMPUTE
: RING_GFX
);
3627 radv_cs_add_buffer(queue
->device
->ws
, cs
, scratch_bo
);
3629 /* Emit initial configuration. */
3630 switch (queue
->queue_family_index
) {
3631 case RADV_QUEUE_GENERAL
:
3632 radv_init_graphics_state(cs
, queue
);
3634 case RADV_QUEUE_COMPUTE
:
3635 radv_init_compute_state(cs
, queue
);
3637 case RADV_QUEUE_TRANSFER
:
3641 if (esgs_ring_bo
|| gsvs_ring_bo
|| tess_rings_bo
) {
3642 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
3643 radeon_emit(cs
, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
3645 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
3646 radeon_emit(cs
, EVENT_TYPE(V_028A90_VGT_FLUSH
) | EVENT_INDEX(0));
3649 radv_emit_gs_ring_sizes(queue
, cs
, esgs_ring_bo
, esgs_ring_size
,
3650 gsvs_ring_bo
, gsvs_ring_size
);
3651 radv_emit_tess_factor_ring(queue
, cs
, hs_offchip_param
,
3652 tess_factor_ring_size
, tess_rings_bo
);
3653 radv_emit_global_shader_pointers(queue
, cs
, descriptor_bo
);
3654 radv_emit_compute_scratch(queue
, cs
, compute_scratch_size_per_wave
,
3655 compute_scratch_waves
, compute_scratch_bo
);
3656 radv_emit_graphics_scratch(queue
, cs
, scratch_size_per_wave
,
3657 scratch_waves
, scratch_bo
);
3660 radv_cs_add_buffer(queue
->device
->ws
, cs
, gds_bo
);
3662 radv_cs_add_buffer(queue
->device
->ws
, cs
, gds_oa_bo
);
3664 if (queue
->device
->trace_bo
)
3665 radv_cs_add_buffer(queue
->device
->ws
, cs
, queue
->device
->trace_bo
);
3667 if (queue
->device
->border_color_data
.bo
)
3668 radv_cs_add_buffer(queue
->device
->ws
, cs
,
3669 queue
->device
->border_color_data
.bo
);
3672 si_cs_emit_cache_flush(cs
,
3673 queue
->device
->physical_device
->rad_info
.chip_class
,
3675 queue
->queue_family_index
== RING_COMPUTE
&&
3676 queue
->device
->physical_device
->rad_info
.chip_class
>= GFX7
,
3677 (queue
->queue_family_index
== RADV_QUEUE_COMPUTE
? RADV_CMD_FLAG_CS_PARTIAL_FLUSH
: (RADV_CMD_FLAG_CS_PARTIAL_FLUSH
| RADV_CMD_FLAG_PS_PARTIAL_FLUSH
)) |
3678 RADV_CMD_FLAG_INV_ICACHE
|
3679 RADV_CMD_FLAG_INV_SCACHE
|
3680 RADV_CMD_FLAG_INV_VCACHE
|
3681 RADV_CMD_FLAG_INV_L2
|
3682 RADV_CMD_FLAG_START_PIPELINE_STATS
, 0);
3683 } else if (i
== 1) {
3684 si_cs_emit_cache_flush(cs
,
3685 queue
->device
->physical_device
->rad_info
.chip_class
,
3687 queue
->queue_family_index
== RING_COMPUTE
&&
3688 queue
->device
->physical_device
->rad_info
.chip_class
>= GFX7
,
3689 RADV_CMD_FLAG_INV_ICACHE
|
3690 RADV_CMD_FLAG_INV_SCACHE
|
3691 RADV_CMD_FLAG_INV_VCACHE
|
3692 RADV_CMD_FLAG_INV_L2
|
3693 RADV_CMD_FLAG_START_PIPELINE_STATS
, 0);
3696 if (queue
->device
->ws
->cs_finalize(cs
) != VK_SUCCESS
)
3700 if (queue
->initial_full_flush_preamble_cs
)
3701 queue
->device
->ws
->cs_destroy(queue
->initial_full_flush_preamble_cs
);
3703 if (queue
->initial_preamble_cs
)
3704 queue
->device
->ws
->cs_destroy(queue
->initial_preamble_cs
);
3706 if (queue
->continue_preamble_cs
)
3707 queue
->device
->ws
->cs_destroy(queue
->continue_preamble_cs
);
3709 queue
->initial_full_flush_preamble_cs
= dest_cs
[0];
3710 queue
->initial_preamble_cs
= dest_cs
[1];
3711 queue
->continue_preamble_cs
= dest_cs
[2];
3713 if (scratch_bo
!= queue
->scratch_bo
) {
3714 if (queue
->scratch_bo
)
3715 queue
->device
->ws
->buffer_destroy(queue
->scratch_bo
);
3716 queue
->scratch_bo
= scratch_bo
;
3718 queue
->scratch_size_per_wave
= scratch_size_per_wave
;
3719 queue
->scratch_waves
= scratch_waves
;
3721 if (compute_scratch_bo
!= queue
->compute_scratch_bo
) {
3722 if (queue
->compute_scratch_bo
)
3723 queue
->device
->ws
->buffer_destroy(queue
->compute_scratch_bo
);
3724 queue
->compute_scratch_bo
= compute_scratch_bo
;
3726 queue
->compute_scratch_size_per_wave
= compute_scratch_size_per_wave
;
3727 queue
->compute_scratch_waves
= compute_scratch_waves
;
3729 if (esgs_ring_bo
!= queue
->esgs_ring_bo
) {
3730 if (queue
->esgs_ring_bo
)
3731 queue
->device
->ws
->buffer_destroy(queue
->esgs_ring_bo
);
3732 queue
->esgs_ring_bo
= esgs_ring_bo
;
3733 queue
->esgs_ring_size
= esgs_ring_size
;
3736 if (gsvs_ring_bo
!= queue
->gsvs_ring_bo
) {
3737 if (queue
->gsvs_ring_bo
)
3738 queue
->device
->ws
->buffer_destroy(queue
->gsvs_ring_bo
);
3739 queue
->gsvs_ring_bo
= gsvs_ring_bo
;
3740 queue
->gsvs_ring_size
= gsvs_ring_size
;
3743 if (tess_rings_bo
!= queue
->tess_rings_bo
) {
3744 queue
->tess_rings_bo
= tess_rings_bo
;
3745 queue
->has_tess_rings
= true;
3748 if (gds_bo
!= queue
->gds_bo
) {
3749 queue
->gds_bo
= gds_bo
;
3750 queue
->has_gds
= true;
3753 if (gds_oa_bo
!= queue
->gds_oa_bo
) {
3754 queue
->gds_oa_bo
= gds_oa_bo
;
3755 queue
->has_gds_oa
= true;
3758 if (descriptor_bo
!= queue
->descriptor_bo
) {
3759 if (queue
->descriptor_bo
)
3760 queue
->device
->ws
->buffer_destroy(queue
->descriptor_bo
);
3762 queue
->descriptor_bo
= descriptor_bo
;
3765 if (add_sample_positions
)
3766 queue
->has_sample_positions
= true;
3768 *initial_full_flush_preamble_cs
= queue
->initial_full_flush_preamble_cs
;
3769 *initial_preamble_cs
= queue
->initial_preamble_cs
;
3770 *continue_preamble_cs
= queue
->continue_preamble_cs
;
3771 if (!scratch_size
&& !compute_scratch_size
&& !esgs_ring_size
&& !gsvs_ring_size
)
3772 *continue_preamble_cs
= NULL
;
3775 for (int i
= 0; i
< ARRAY_SIZE(dest_cs
); ++i
)
3777 queue
->device
->ws
->cs_destroy(dest_cs
[i
]);
3778 if (descriptor_bo
&& descriptor_bo
!= queue
->descriptor_bo
)
3779 queue
->device
->ws
->buffer_destroy(descriptor_bo
);
3780 if (scratch_bo
&& scratch_bo
!= queue
->scratch_bo
)
3781 queue
->device
->ws
->buffer_destroy(scratch_bo
);
3782 if (compute_scratch_bo
&& compute_scratch_bo
!= queue
->compute_scratch_bo
)
3783 queue
->device
->ws
->buffer_destroy(compute_scratch_bo
);
3784 if (esgs_ring_bo
&& esgs_ring_bo
!= queue
->esgs_ring_bo
)
3785 queue
->device
->ws
->buffer_destroy(esgs_ring_bo
);
3786 if (gsvs_ring_bo
&& gsvs_ring_bo
!= queue
->gsvs_ring_bo
)
3787 queue
->device
->ws
->buffer_destroy(gsvs_ring_bo
);
3788 if (tess_rings_bo
&& tess_rings_bo
!= queue
->tess_rings_bo
)
3789 queue
->device
->ws
->buffer_destroy(tess_rings_bo
);
3790 if (gds_bo
&& gds_bo
!= queue
->gds_bo
)
3791 queue
->device
->ws
->buffer_destroy(gds_bo
);
3792 if (gds_oa_bo
&& gds_oa_bo
!= queue
->gds_oa_bo
)
3793 queue
->device
->ws
->buffer_destroy(gds_oa_bo
);
3795 return vk_error(queue
->device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
3798 static VkResult
radv_alloc_sem_counts(struct radv_device
*device
,
3799 struct radv_winsys_sem_counts
*counts
,
3801 struct radv_semaphore_part
**sems
,
3802 const uint64_t *timeline_values
,
3806 int syncobj_idx
= 0, non_reset_idx
= 0, sem_idx
= 0, timeline_idx
= 0;
3808 if (num_sems
== 0 && _fence
== VK_NULL_HANDLE
)
3811 for (uint32_t i
= 0; i
< num_sems
; i
++) {
3812 switch(sems
[i
]->kind
) {
3813 case RADV_SEMAPHORE_SYNCOBJ
:
3814 counts
->syncobj_count
++;
3815 counts
->syncobj_reset_count
++;
3817 case RADV_SEMAPHORE_WINSYS
:
3818 counts
->sem_count
++;
3820 case RADV_SEMAPHORE_NONE
:
3822 case RADV_SEMAPHORE_TIMELINE
:
3823 counts
->syncobj_count
++;
3825 case RADV_SEMAPHORE_TIMELINE_SYNCOBJ
:
3826 counts
->timeline_syncobj_count
++;
3831 if (_fence
!= VK_NULL_HANDLE
) {
3832 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
3834 struct radv_fence_part
*part
=
3835 fence
->temporary
.kind
!= RADV_FENCE_NONE
?
3836 &fence
->temporary
: &fence
->permanent
;
3837 if (part
->kind
== RADV_FENCE_SYNCOBJ
)
3838 counts
->syncobj_count
++;
3841 if (counts
->syncobj_count
|| counts
->timeline_syncobj_count
) {
3842 counts
->points
= (uint64_t *)malloc(
3843 sizeof(*counts
->syncobj
) * counts
->syncobj_count
+
3844 (sizeof(*counts
->syncobj
) + sizeof(*counts
->points
)) * counts
->timeline_syncobj_count
);
3845 if (!counts
->points
)
3846 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3847 counts
->syncobj
= (uint32_t*)(counts
->points
+ counts
->timeline_syncobj_count
);
3850 if (counts
->sem_count
) {
3851 counts
->sem
= (struct radeon_winsys_sem
**)malloc(sizeof(struct radeon_winsys_sem
*) * counts
->sem_count
);
3853 free(counts
->syncobj
);
3854 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3858 non_reset_idx
= counts
->syncobj_reset_count
;
3860 for (uint32_t i
= 0; i
< num_sems
; i
++) {
3861 switch(sems
[i
]->kind
) {
3862 case RADV_SEMAPHORE_NONE
:
3863 unreachable("Empty semaphore");
3865 case RADV_SEMAPHORE_SYNCOBJ
:
3866 counts
->syncobj
[syncobj_idx
++] = sems
[i
]->syncobj
;
3868 case RADV_SEMAPHORE_WINSYS
:
3869 counts
->sem
[sem_idx
++] = sems
[i
]->ws_sem
;
3871 case RADV_SEMAPHORE_TIMELINE
: {
3872 pthread_mutex_lock(&sems
[i
]->timeline
.mutex
);
3873 struct radv_timeline_point
*point
= NULL
;
3875 point
= radv_timeline_add_point_locked(device
, &sems
[i
]->timeline
, timeline_values
[i
]);
3877 point
= radv_timeline_find_point_at_least_locked(device
, &sems
[i
]->timeline
, timeline_values
[i
]);
3880 pthread_mutex_unlock(&sems
[i
]->timeline
.mutex
);
3883 counts
->syncobj
[non_reset_idx
++] = point
->syncobj
;
3885 /* Explicitly remove the semaphore so we might not find
3886 * a point later post-submit. */
3891 case RADV_SEMAPHORE_TIMELINE_SYNCOBJ
:
3892 counts
->syncobj
[counts
->syncobj_count
+ timeline_idx
] = sems
[i
]->syncobj
;
3893 counts
->points
[timeline_idx
] = timeline_values
[i
];
3899 if (_fence
!= VK_NULL_HANDLE
) {
3900 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
3902 struct radv_fence_part
*part
=
3903 fence
->temporary
.kind
!= RADV_FENCE_NONE
?
3904 &fence
->temporary
: &fence
->permanent
;
3905 if (part
->kind
== RADV_FENCE_SYNCOBJ
)
3906 counts
->syncobj
[non_reset_idx
++] = part
->syncobj
;
3909 assert(MAX2(syncobj_idx
, non_reset_idx
) <= counts
->syncobj_count
);
3910 counts
->syncobj_count
= MAX2(syncobj_idx
, non_reset_idx
);
3916 radv_free_sem_info(struct radv_winsys_sem_info
*sem_info
)
3918 free(sem_info
->wait
.points
);
3919 free(sem_info
->wait
.sem
);
3920 free(sem_info
->signal
.points
);
3921 free(sem_info
->signal
.sem
);
3925 static void radv_free_temp_syncobjs(struct radv_device
*device
,
3927 struct radv_semaphore_part
*sems
)
3929 for (uint32_t i
= 0; i
< num_sems
; i
++) {
3930 radv_destroy_semaphore_part(device
, sems
+ i
);
3935 radv_alloc_sem_info(struct radv_device
*device
,
3936 struct radv_winsys_sem_info
*sem_info
,
3938 struct radv_semaphore_part
**wait_sems
,
3939 const uint64_t *wait_values
,
3940 int num_signal_sems
,
3941 struct radv_semaphore_part
**signal_sems
,
3942 const uint64_t *signal_values
,
3946 memset(sem_info
, 0, sizeof(*sem_info
));
3948 ret
= radv_alloc_sem_counts(device
, &sem_info
->wait
, num_wait_sems
, wait_sems
, wait_values
, VK_NULL_HANDLE
, false);
3951 ret
= radv_alloc_sem_counts(device
, &sem_info
->signal
, num_signal_sems
, signal_sems
, signal_values
, fence
, true);
3953 radv_free_sem_info(sem_info
);
3955 /* caller can override these */
3956 sem_info
->cs_emit_wait
= true;
3957 sem_info
->cs_emit_signal
= true;
3962 radv_finalize_timelines(struct radv_device
*device
,
3963 uint32_t num_wait_sems
,
3964 struct radv_semaphore_part
**wait_sems
,
3965 const uint64_t *wait_values
,
3966 uint32_t num_signal_sems
,
3967 struct radv_semaphore_part
**signal_sems
,
3968 const uint64_t *signal_values
,
3969 struct list_head
*processing_list
)
3971 for (uint32_t i
= 0; i
< num_wait_sems
; ++i
) {
3972 if (wait_sems
[i
] && wait_sems
[i
]->kind
== RADV_SEMAPHORE_TIMELINE
) {
3973 pthread_mutex_lock(&wait_sems
[i
]->timeline
.mutex
);
3974 struct radv_timeline_point
*point
=
3975 radv_timeline_find_point_at_least_locked(device
, &wait_sems
[i
]->timeline
, wait_values
[i
]);
3976 point
->wait_count
-= 2;
3977 pthread_mutex_unlock(&wait_sems
[i
]->timeline
.mutex
);
3980 for (uint32_t i
= 0; i
< num_signal_sems
; ++i
) {
3981 if (signal_sems
[i
] && signal_sems
[i
]->kind
== RADV_SEMAPHORE_TIMELINE
) {
3982 pthread_mutex_lock(&signal_sems
[i
]->timeline
.mutex
);
3983 struct radv_timeline_point
*point
=
3984 radv_timeline_find_point_at_least_locked(device
, &signal_sems
[i
]->timeline
, signal_values
[i
]);
3985 signal_sems
[i
]->timeline
.highest_submitted
=
3986 MAX2(signal_sems
[i
]->timeline
.highest_submitted
, point
->value
);
3987 point
->wait_count
-= 2;
3988 radv_timeline_trigger_waiters_locked(&signal_sems
[i
]->timeline
, processing_list
);
3989 pthread_mutex_unlock(&signal_sems
[i
]->timeline
.mutex
);
3990 } else if (signal_sems
[i
] && signal_sems
[i
]->kind
== RADV_SEMAPHORE_TIMELINE_SYNCOBJ
) {
3991 signal_sems
[i
]->timeline_syncobj
.max_point
=
3992 MAX2(signal_sems
[i
]->timeline_syncobj
.max_point
, signal_values
[i
]);
3998 radv_sparse_buffer_bind_memory(struct radv_device
*device
,
3999 const VkSparseBufferMemoryBindInfo
*bind
)
4001 RADV_FROM_HANDLE(radv_buffer
, buffer
, bind
->buffer
);
4004 for (uint32_t i
= 0; i
< bind
->bindCount
; ++i
) {
4005 struct radv_device_memory
*mem
= NULL
;
4007 if (bind
->pBinds
[i
].memory
!= VK_NULL_HANDLE
)
4008 mem
= radv_device_memory_from_handle(bind
->pBinds
[i
].memory
);
4010 result
= device
->ws
->buffer_virtual_bind(buffer
->bo
,
4011 bind
->pBinds
[i
].resourceOffset
,
4012 bind
->pBinds
[i
].size
,
4013 mem
? mem
->bo
: NULL
,
4014 bind
->pBinds
[i
].memoryOffset
);
4015 if (result
!= VK_SUCCESS
)
4023 radv_sparse_image_opaque_bind_memory(struct radv_device
*device
,
4024 const VkSparseImageOpaqueMemoryBindInfo
*bind
)
4026 RADV_FROM_HANDLE(radv_image
, image
, bind
->image
);
4029 for (uint32_t i
= 0; i
< bind
->bindCount
; ++i
) {
4030 struct radv_device_memory
*mem
= NULL
;
4032 if (bind
->pBinds
[i
].memory
!= VK_NULL_HANDLE
)
4033 mem
= radv_device_memory_from_handle(bind
->pBinds
[i
].memory
);
4035 result
= device
->ws
->buffer_virtual_bind(image
->bo
,
4036 bind
->pBinds
[i
].resourceOffset
,
4037 bind
->pBinds
[i
].size
,
4038 mem
? mem
->bo
: NULL
,
4039 bind
->pBinds
[i
].memoryOffset
);
4040 if (result
!= VK_SUCCESS
)
4048 radv_get_preambles(struct radv_queue
*queue
,
4049 const VkCommandBuffer
*cmd_buffers
,
4050 uint32_t cmd_buffer_count
,
4051 struct radeon_cmdbuf
**initial_full_flush_preamble_cs
,
4052 struct radeon_cmdbuf
**initial_preamble_cs
,
4053 struct radeon_cmdbuf
**continue_preamble_cs
)
4055 uint32_t scratch_size_per_wave
= 0, waves_wanted
= 0;
4056 uint32_t compute_scratch_size_per_wave
= 0, compute_waves_wanted
= 0;
4057 uint32_t esgs_ring_size
= 0, gsvs_ring_size
= 0;
4058 bool tess_rings_needed
= false;
4059 bool gds_needed
= false;
4060 bool gds_oa_needed
= false;
4061 bool sample_positions_needed
= false;
4063 for (uint32_t j
= 0; j
< cmd_buffer_count
; j
++) {
4064 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
,
4067 scratch_size_per_wave
= MAX2(scratch_size_per_wave
, cmd_buffer
->scratch_size_per_wave_needed
);
4068 waves_wanted
= MAX2(waves_wanted
, cmd_buffer
->scratch_waves_wanted
);
4069 compute_scratch_size_per_wave
= MAX2(compute_scratch_size_per_wave
,
4070 cmd_buffer
->compute_scratch_size_per_wave_needed
);
4071 compute_waves_wanted
= MAX2(compute_waves_wanted
,
4072 cmd_buffer
->compute_scratch_waves_wanted
);
4073 esgs_ring_size
= MAX2(esgs_ring_size
, cmd_buffer
->esgs_ring_size_needed
);
4074 gsvs_ring_size
= MAX2(gsvs_ring_size
, cmd_buffer
->gsvs_ring_size_needed
);
4075 tess_rings_needed
|= cmd_buffer
->tess_rings_needed
;
4076 gds_needed
|= cmd_buffer
->gds_needed
;
4077 gds_oa_needed
|= cmd_buffer
->gds_oa_needed
;
4078 sample_positions_needed
|= cmd_buffer
->sample_positions_needed
;
4081 return radv_get_preamble_cs(queue
, scratch_size_per_wave
, waves_wanted
,
4082 compute_scratch_size_per_wave
, compute_waves_wanted
,
4083 esgs_ring_size
, gsvs_ring_size
, tess_rings_needed
,
4084 gds_needed
, gds_oa_needed
, sample_positions_needed
,
4085 initial_full_flush_preamble_cs
,
4086 initial_preamble_cs
, continue_preamble_cs
);
4089 struct radv_deferred_queue_submission
{
4090 struct radv_queue
*queue
;
4091 VkCommandBuffer
*cmd_buffers
;
4092 uint32_t cmd_buffer_count
;
4094 /* Sparse bindings that happen on a queue. */
4095 VkSparseBufferMemoryBindInfo
*buffer_binds
;
4096 uint32_t buffer_bind_count
;
4097 VkSparseImageOpaqueMemoryBindInfo
*image_opaque_binds
;
4098 uint32_t image_opaque_bind_count
;
4101 VkShaderStageFlags wait_dst_stage_mask
;
4102 struct radv_semaphore_part
**wait_semaphores
;
4103 uint32_t wait_semaphore_count
;
4104 struct radv_semaphore_part
**signal_semaphores
;
4105 uint32_t signal_semaphore_count
;
4108 uint64_t *wait_values
;
4109 uint64_t *signal_values
;
4111 struct radv_semaphore_part
*temporary_semaphore_parts
;
4112 uint32_t temporary_semaphore_part_count
;
4114 struct list_head queue_pending_list
;
4115 uint32_t submission_wait_count
;
4116 struct radv_timeline_waiter
*wait_nodes
;
4118 struct list_head processing_list
;
4121 struct radv_queue_submission
{
4122 const VkCommandBuffer
*cmd_buffers
;
4123 uint32_t cmd_buffer_count
;
4125 /* Sparse bindings that happen on a queue. */
4126 const VkSparseBufferMemoryBindInfo
*buffer_binds
;
4127 uint32_t buffer_bind_count
;
4128 const VkSparseImageOpaqueMemoryBindInfo
*image_opaque_binds
;
4129 uint32_t image_opaque_bind_count
;
4132 VkPipelineStageFlags wait_dst_stage_mask
;
4133 const VkSemaphore
*wait_semaphores
;
4134 uint32_t wait_semaphore_count
;
4135 const VkSemaphore
*signal_semaphores
;
4136 uint32_t signal_semaphore_count
;
4139 const uint64_t *wait_values
;
4140 uint32_t wait_value_count
;
4141 const uint64_t *signal_values
;
4142 uint32_t signal_value_count
;
4146 radv_queue_trigger_submission(struct radv_deferred_queue_submission
*submission
,
4148 struct list_head
*processing_list
);
4151 radv_create_deferred_submission(struct radv_queue
*queue
,
4152 const struct radv_queue_submission
*submission
,
4153 struct radv_deferred_queue_submission
**out
)
4155 struct radv_deferred_queue_submission
*deferred
= NULL
;
4156 size_t size
= sizeof(struct radv_deferred_queue_submission
);
4158 uint32_t temporary_count
= 0;
4159 for (uint32_t i
= 0; i
< submission
->wait_semaphore_count
; ++i
) {
4160 RADV_FROM_HANDLE(radv_semaphore
, semaphore
, submission
->wait_semaphores
[i
]);
4161 if (semaphore
->temporary
.kind
!= RADV_SEMAPHORE_NONE
)
4165 size
+= submission
->cmd_buffer_count
* sizeof(VkCommandBuffer
);
4166 size
+= submission
->buffer_bind_count
* sizeof(VkSparseBufferMemoryBindInfo
);
4167 size
+= submission
->image_opaque_bind_count
* sizeof(VkSparseImageOpaqueMemoryBindInfo
);
4168 size
+= submission
->wait_semaphore_count
* sizeof(struct radv_semaphore_part
*);
4169 size
+= temporary_count
* sizeof(struct radv_semaphore_part
);
4170 size
+= submission
->signal_semaphore_count
* sizeof(struct radv_semaphore_part
*);
4171 size
+= submission
->wait_value_count
* sizeof(uint64_t);
4172 size
+= submission
->signal_value_count
* sizeof(uint64_t);
4173 size
+= submission
->wait_semaphore_count
* sizeof(struct radv_timeline_waiter
);
4175 deferred
= calloc(1, size
);
4177 return VK_ERROR_OUT_OF_HOST_MEMORY
;
4179 deferred
->queue
= queue
;
4181 deferred
->cmd_buffers
= (void*)(deferred
+ 1);
4182 deferred
->cmd_buffer_count
= submission
->cmd_buffer_count
;
4183 memcpy(deferred
->cmd_buffers
, submission
->cmd_buffers
,
4184 submission
->cmd_buffer_count
* sizeof(*deferred
->cmd_buffers
));
4186 deferred
->buffer_binds
= (void*)(deferred
->cmd_buffers
+ submission
->cmd_buffer_count
);
4187 deferred
->buffer_bind_count
= submission
->buffer_bind_count
;
4188 memcpy(deferred
->buffer_binds
, submission
->buffer_binds
,
4189 submission
->buffer_bind_count
* sizeof(*deferred
->buffer_binds
));
4191 deferred
->image_opaque_binds
= (void*)(deferred
->buffer_binds
+ submission
->buffer_bind_count
);
4192 deferred
->image_opaque_bind_count
= submission
->image_opaque_bind_count
;
4193 memcpy(deferred
->image_opaque_binds
, submission
->image_opaque_binds
,
4194 submission
->image_opaque_bind_count
* sizeof(*deferred
->image_opaque_binds
));
4196 deferred
->flush_caches
= submission
->flush_caches
;
4197 deferred
->wait_dst_stage_mask
= submission
->wait_dst_stage_mask
;
4199 deferred
->wait_semaphores
= (void*)(deferred
->image_opaque_binds
+ deferred
->image_opaque_bind_count
);
4200 deferred
->wait_semaphore_count
= submission
->wait_semaphore_count
;
4202 deferred
->signal_semaphores
= (void*)(deferred
->wait_semaphores
+ deferred
->wait_semaphore_count
);
4203 deferred
->signal_semaphore_count
= submission
->signal_semaphore_count
;
4205 deferred
->fence
= submission
->fence
;
4207 deferred
->temporary_semaphore_parts
= (void*)(deferred
->signal_semaphores
+ deferred
->signal_semaphore_count
);
4208 deferred
->temporary_semaphore_part_count
= temporary_count
;
4210 uint32_t temporary_idx
= 0;
4211 for (uint32_t i
= 0; i
< submission
->wait_semaphore_count
; ++i
) {
4212 RADV_FROM_HANDLE(radv_semaphore
, semaphore
, submission
->wait_semaphores
[i
]);
4213 if (semaphore
->temporary
.kind
!= RADV_SEMAPHORE_NONE
) {
4214 deferred
->wait_semaphores
[i
] = &deferred
->temporary_semaphore_parts
[temporary_idx
];
4215 deferred
->temporary_semaphore_parts
[temporary_idx
] = semaphore
->temporary
;
4216 semaphore
->temporary
.kind
= RADV_SEMAPHORE_NONE
;
4219 deferred
->wait_semaphores
[i
] = &semaphore
->permanent
;
4222 for (uint32_t i
= 0; i
< submission
->signal_semaphore_count
; ++i
) {
4223 RADV_FROM_HANDLE(radv_semaphore
, semaphore
, submission
->signal_semaphores
[i
]);
4224 if (semaphore
->temporary
.kind
!= RADV_SEMAPHORE_NONE
) {
4225 deferred
->signal_semaphores
[i
] = &semaphore
->temporary
;
4227 deferred
->signal_semaphores
[i
] = &semaphore
->permanent
;
4231 deferred
->wait_values
= (void*)(deferred
->temporary_semaphore_parts
+ temporary_count
);
4232 memcpy(deferred
->wait_values
, submission
->wait_values
, submission
->wait_value_count
* sizeof(uint64_t));
4233 deferred
->signal_values
= deferred
->wait_values
+ submission
->wait_value_count
;
4234 memcpy(deferred
->signal_values
, submission
->signal_values
, submission
->signal_value_count
* sizeof(uint64_t));
4236 deferred
->wait_nodes
= (void*)(deferred
->signal_values
+ submission
->signal_value_count
);
4237 /* This is worst-case. radv_queue_enqueue_submission will fill in further, but this
4238 * ensure the submission is not accidentally triggered early when adding wait timelines. */
4239 deferred
->submission_wait_count
= 1 + submission
->wait_semaphore_count
;
4246 radv_queue_enqueue_submission(struct radv_deferred_queue_submission
*submission
,
4247 struct list_head
*processing_list
)
4249 uint32_t wait_cnt
= 0;
4250 struct radv_timeline_waiter
*waiter
= submission
->wait_nodes
;
4251 for (uint32_t i
= 0; i
< submission
->wait_semaphore_count
; ++i
) {
4252 if (submission
->wait_semaphores
[i
]->kind
== RADV_SEMAPHORE_TIMELINE
) {
4253 pthread_mutex_lock(&submission
->wait_semaphores
[i
]->timeline
.mutex
);
4254 if (submission
->wait_semaphores
[i
]->timeline
.highest_submitted
< submission
->wait_values
[i
]) {
4256 waiter
->value
= submission
->wait_values
[i
];
4257 waiter
->submission
= submission
;
4258 list_addtail(&waiter
->list
, &submission
->wait_semaphores
[i
]->timeline
.waiters
);
4261 pthread_mutex_unlock(&submission
->wait_semaphores
[i
]->timeline
.mutex
);
4265 pthread_mutex_lock(&submission
->queue
->pending_mutex
);
4267 bool is_first
= list_is_empty(&submission
->queue
->pending_submissions
);
4268 list_addtail(&submission
->queue_pending_list
, &submission
->queue
->pending_submissions
);
4270 pthread_mutex_unlock(&submission
->queue
->pending_mutex
);
4272 /* If there is already a submission in the queue, that will decrement the counter by 1 when
4273 * submitted, but if the queue was empty, we decrement ourselves as there is no previous
4275 uint32_t decrement
= submission
->wait_semaphore_count
- wait_cnt
+ (is_first
? 1 : 0);
4276 return radv_queue_trigger_submission(submission
, decrement
, processing_list
);
4280 radv_queue_submission_update_queue(struct radv_deferred_queue_submission
*submission
,
4281 struct list_head
*processing_list
)
4283 pthread_mutex_lock(&submission
->queue
->pending_mutex
);
4284 list_del(&submission
->queue_pending_list
);
4286 /* trigger the next submission in the queue. */
4287 if (!list_is_empty(&submission
->queue
->pending_submissions
)) {
4288 struct radv_deferred_queue_submission
*next_submission
=
4289 list_first_entry(&submission
->queue
->pending_submissions
,
4290 struct radv_deferred_queue_submission
,
4291 queue_pending_list
);
4292 radv_queue_trigger_submission(next_submission
, 1, processing_list
);
4294 pthread_mutex_unlock(&submission
->queue
->pending_mutex
);
4296 pthread_cond_broadcast(&submission
->queue
->device
->timeline_cond
);
4300 radv_queue_submit_deferred(struct radv_deferred_queue_submission
*submission
,
4301 struct list_head
*processing_list
)
4303 RADV_FROM_HANDLE(radv_fence
, fence
, submission
->fence
);
4304 struct radv_queue
*queue
= submission
->queue
;
4305 struct radeon_winsys_ctx
*ctx
= queue
->hw_ctx
;
4306 uint32_t max_cs_submission
= queue
->device
->trace_bo
? 1 : RADV_MAX_IBS_PER_SUBMIT
;
4307 struct radeon_winsys_fence
*base_fence
= NULL
;
4308 bool do_flush
= submission
->flush_caches
|| submission
->wait_dst_stage_mask
;
4309 bool can_patch
= true;
4311 struct radv_winsys_sem_info sem_info
;
4313 struct radeon_cmdbuf
*initial_preamble_cs
= NULL
;
4314 struct radeon_cmdbuf
*initial_flush_preamble_cs
= NULL
;
4315 struct radeon_cmdbuf
*continue_preamble_cs
= NULL
;
4318 /* Under most circumstances, out fences won't be temporary.
4319 * However, the spec does allow it for opaque_fd.
4321 * From the Vulkan 1.0.53 spec:
4323 * "If the import is temporary, the implementation must
4324 * restore the semaphore to its prior permanent state after
4325 * submitting the next semaphore wait operation."
4327 struct radv_fence_part
*part
=
4328 fence
->temporary
.kind
!= RADV_FENCE_NONE
?
4329 &fence
->temporary
: &fence
->permanent
;
4330 if (part
->kind
== RADV_FENCE_WINSYS
)
4331 base_fence
= part
->fence
;
4334 result
= radv_get_preambles(queue
, submission
->cmd_buffers
,
4335 submission
->cmd_buffer_count
,
4336 &initial_preamble_cs
,
4337 &initial_flush_preamble_cs
,
4338 &continue_preamble_cs
);
4339 if (result
!= VK_SUCCESS
)
4342 result
= radv_alloc_sem_info(queue
->device
,
4344 submission
->wait_semaphore_count
,
4345 submission
->wait_semaphores
,
4346 submission
->wait_values
,
4347 submission
->signal_semaphore_count
,
4348 submission
->signal_semaphores
,
4349 submission
->signal_values
,
4351 if (result
!= VK_SUCCESS
)
4354 for (uint32_t i
= 0; i
< submission
->buffer_bind_count
; ++i
) {
4355 result
= radv_sparse_buffer_bind_memory(queue
->device
,
4356 submission
->buffer_binds
+ i
);
4357 if (result
!= VK_SUCCESS
)
4361 for (uint32_t i
= 0; i
< submission
->image_opaque_bind_count
; ++i
) {
4362 result
= radv_sparse_image_opaque_bind_memory(queue
->device
,
4363 submission
->image_opaque_binds
+ i
);
4364 if (result
!= VK_SUCCESS
)
4368 if (!submission
->cmd_buffer_count
) {
4369 result
= queue
->device
->ws
->cs_submit(ctx
, queue
->queue_idx
,
4370 &queue
->device
->empty_cs
[queue
->queue_family_index
],
4374 if (result
!= VK_SUCCESS
)
4377 struct radeon_cmdbuf
**cs_array
= malloc(sizeof(struct radeon_cmdbuf
*) *
4378 (submission
->cmd_buffer_count
));
4380 for (uint32_t j
= 0; j
< submission
->cmd_buffer_count
; j
++) {
4381 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, submission
->cmd_buffers
[j
]);
4382 assert(cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_PRIMARY
);
4384 cs_array
[j
] = cmd_buffer
->cs
;
4385 if ((cmd_buffer
->usage_flags
& VK_COMMAND_BUFFER_USAGE_SIMULTANEOUS_USE_BIT
))
4388 cmd_buffer
->status
= RADV_CMD_BUFFER_STATUS_PENDING
;
4391 for (uint32_t j
= 0; j
< submission
->cmd_buffer_count
; j
+= advance
) {
4392 struct radeon_cmdbuf
*initial_preamble
= (do_flush
&& !j
) ? initial_flush_preamble_cs
: initial_preamble_cs
;
4393 const struct radv_winsys_bo_list
*bo_list
= NULL
;
4395 advance
= MIN2(max_cs_submission
,
4396 submission
->cmd_buffer_count
- j
);
4398 if (queue
->device
->trace_bo
)
4399 *queue
->device
->trace_id_ptr
= 0;
4401 sem_info
.cs_emit_wait
= j
== 0;
4402 sem_info
.cs_emit_signal
= j
+ advance
== submission
->cmd_buffer_count
;
4404 if (unlikely(queue
->device
->use_global_bo_list
)) {
4405 pthread_mutex_lock(&queue
->device
->bo_list
.mutex
);
4406 bo_list
= &queue
->device
->bo_list
.list
;
4409 result
= queue
->device
->ws
->cs_submit(ctx
, queue
->queue_idx
, cs_array
+ j
,
4410 advance
, initial_preamble
, continue_preamble_cs
,
4412 can_patch
, base_fence
);
4414 if (unlikely(queue
->device
->use_global_bo_list
))
4415 pthread_mutex_unlock(&queue
->device
->bo_list
.mutex
);
4417 if (result
!= VK_SUCCESS
)
4420 if (queue
->device
->trace_bo
) {
4421 radv_check_gpu_hangs(queue
, cs_array
[j
]);
4428 radv_free_temp_syncobjs(queue
->device
,
4429 submission
->temporary_semaphore_part_count
,
4430 submission
->temporary_semaphore_parts
);
4431 radv_finalize_timelines(queue
->device
,
4432 submission
->wait_semaphore_count
,
4433 submission
->wait_semaphores
,
4434 submission
->wait_values
,
4435 submission
->signal_semaphore_count
,
4436 submission
->signal_semaphores
,
4437 submission
->signal_values
,
4439 /* Has to happen after timeline finalization to make sure the
4440 * condition variable is only triggered when timelines and queue have
4442 radv_queue_submission_update_queue(submission
, processing_list
);
4443 radv_free_sem_info(&sem_info
);
4448 if (result
!= VK_SUCCESS
&& result
!= VK_ERROR_DEVICE_LOST
) {
4449 /* When something bad happened during the submission, such as
4450 * an out of memory issue, it might be hard to recover from
4451 * this inconsistent state. To avoid this sort of problem, we
4452 * assume that we are in a really bad situation and return
4453 * VK_ERROR_DEVICE_LOST to ensure the clients do not attempt
4454 * to submit the same job again to this device.
4456 result
= VK_ERROR_DEVICE_LOST
;
4459 radv_free_temp_syncobjs(queue
->device
,
4460 submission
->temporary_semaphore_part_count
,
4461 submission
->temporary_semaphore_parts
);
4467 radv_process_submissions(struct list_head
*processing_list
)
4469 while(!list_is_empty(processing_list
)) {
4470 struct radv_deferred_queue_submission
*submission
=
4471 list_first_entry(processing_list
, struct radv_deferred_queue_submission
, processing_list
);
4472 list_del(&submission
->processing_list
);
4474 VkResult result
= radv_queue_submit_deferred(submission
, processing_list
);
4475 if (result
!= VK_SUCCESS
)
4482 wait_for_submission_timelines_available(struct radv_deferred_queue_submission
*submission
,
4485 struct radv_device
*device
= submission
->queue
->device
;
4486 uint32_t syncobj_count
= 0;
4487 uint32_t syncobj_idx
= 0;
4489 for (uint32_t i
= 0; i
< submission
->wait_semaphore_count
; ++i
) {
4490 if (submission
->wait_semaphores
[i
]->kind
!= RADV_SEMAPHORE_TIMELINE_SYNCOBJ
)
4493 if (submission
->wait_semaphores
[i
]->timeline_syncobj
.max_point
>= submission
->wait_values
[i
])
4501 uint64_t *points
= malloc((sizeof(uint64_t) + sizeof(uint32_t)) * syncobj_count
);
4503 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
4505 uint32_t *syncobj
= (uint32_t*)(points
+ syncobj_count
);
4507 for (uint32_t i
= 0; i
< submission
->wait_semaphore_count
; ++i
) {
4508 if (submission
->wait_semaphores
[i
]->kind
!= RADV_SEMAPHORE_TIMELINE_SYNCOBJ
)
4511 if (submission
->wait_semaphores
[i
]->timeline_syncobj
.max_point
>= submission
->wait_values
[i
])
4514 syncobj
[syncobj_idx
] = submission
->wait_semaphores
[i
]->syncobj
;
4515 points
[syncobj_idx
] = submission
->wait_values
[i
];
4518 bool success
= device
->ws
->wait_timeline_syncobj(device
->ws
, syncobj
, points
, syncobj_idx
, true, true, timeout
);
4521 return success
? VK_SUCCESS
: VK_TIMEOUT
;
4524 static void* radv_queue_submission_thread_run(void *q
)
4526 struct radv_queue
*queue
= q
;
4528 pthread_mutex_lock(&queue
->thread_mutex
);
4529 while (!p_atomic_read(&queue
->thread_exit
)) {
4530 struct radv_deferred_queue_submission
*submission
= queue
->thread_submission
;
4531 struct list_head processing_list
;
4532 VkResult result
= VK_SUCCESS
;
4534 pthread_cond_wait(&queue
->thread_cond
, &queue
->thread_mutex
);
4537 pthread_mutex_unlock(&queue
->thread_mutex
);
4539 /* Wait at most 5 seconds so we have a chance to notice shutdown when
4540 * a semaphore never gets signaled. If it takes longer we just retry
4541 * the wait next iteration. */
4542 result
= wait_for_submission_timelines_available(submission
,
4543 radv_get_absolute_timeout(5000000000));
4544 if (result
!= VK_SUCCESS
) {
4545 pthread_mutex_lock(&queue
->thread_mutex
);
4549 /* The lock isn't held but nobody will add one until we finish
4550 * the current submission. */
4551 p_atomic_set(&queue
->thread_submission
, NULL
);
4553 list_inithead(&processing_list
);
4554 list_addtail(&submission
->processing_list
, &processing_list
);
4555 result
= radv_process_submissions(&processing_list
);
4557 pthread_mutex_lock(&queue
->thread_mutex
);
4559 pthread_mutex_unlock(&queue
->thread_mutex
);
4564 radv_queue_trigger_submission(struct radv_deferred_queue_submission
*submission
,
4566 struct list_head
*processing_list
)
4568 struct radv_queue
*queue
= submission
->queue
;
4570 if (p_atomic_add_return(&submission
->submission_wait_count
, -decrement
))
4573 if (wait_for_submission_timelines_available(submission
, radv_get_absolute_timeout(0)) == VK_SUCCESS
) {
4574 list_addtail(&submission
->processing_list
, processing_list
);
4578 pthread_mutex_lock(&queue
->thread_mutex
);
4580 /* A submission can only be ready for the thread if it doesn't have
4581 * any predecessors in the same queue, so there can only be one such
4582 * submission at a time. */
4583 assert(queue
->thread_submission
== NULL
);
4585 /* Only start the thread on demand to save resources for the many games
4586 * which only use binary semaphores. */
4587 if (!queue
->thread_running
) {
4588 ret
= pthread_create(&queue
->submission_thread
, NULL
,
4589 radv_queue_submission_thread_run
, queue
);
4591 pthread_mutex_unlock(&queue
->thread_mutex
);
4592 return vk_errorf(queue
->device
->instance
,
4593 VK_ERROR_DEVICE_LOST
,
4594 "Failed to start submission thread");
4596 queue
->thread_running
= true;
4599 queue
->thread_submission
= submission
;
4600 pthread_mutex_unlock(&queue
->thread_mutex
);
4602 pthread_cond_signal(&queue
->thread_cond
);
4606 static VkResult
radv_queue_submit(struct radv_queue
*queue
,
4607 const struct radv_queue_submission
*submission
)
4609 struct radv_deferred_queue_submission
*deferred
= NULL
;
4611 VkResult result
= radv_create_deferred_submission(queue
, submission
, &deferred
);
4612 if (result
!= VK_SUCCESS
)
4615 struct list_head processing_list
;
4616 list_inithead(&processing_list
);
4618 result
= radv_queue_enqueue_submission(deferred
, &processing_list
);
4619 if (result
!= VK_SUCCESS
) {
4620 /* If anything is in the list we leak. */
4621 assert(list_is_empty(&processing_list
));
4624 return radv_process_submissions(&processing_list
);
4628 radv_queue_internal_submit(struct radv_queue
*queue
, struct radeon_cmdbuf
*cs
)
4630 struct radeon_winsys_ctx
*ctx
= queue
->hw_ctx
;
4631 struct radv_winsys_sem_info sem_info
;
4634 result
= radv_alloc_sem_info(queue
->device
, &sem_info
, 0, NULL
, 0, 0,
4635 0, NULL
, VK_NULL_HANDLE
);
4636 if (result
!= VK_SUCCESS
)
4639 result
= queue
->device
->ws
->cs_submit(ctx
, queue
->queue_idx
, &cs
, 1,
4640 NULL
, NULL
, &sem_info
, NULL
,
4642 radv_free_sem_info(&sem_info
);
4643 if (result
!= VK_SUCCESS
)
4650 /* Signals fence as soon as all the work currently put on queue is done. */
4651 static VkResult
radv_signal_fence(struct radv_queue
*queue
,
4654 return radv_queue_submit(queue
, &(struct radv_queue_submission
) {
4659 static bool radv_submit_has_effects(const VkSubmitInfo
*info
)
4661 return info
->commandBufferCount
||
4662 info
->waitSemaphoreCount
||
4663 info
->signalSemaphoreCount
;
4666 VkResult
radv_QueueSubmit(
4668 uint32_t submitCount
,
4669 const VkSubmitInfo
* pSubmits
,
4672 RADV_FROM_HANDLE(radv_queue
, queue
, _queue
);
4674 uint32_t fence_idx
= 0;
4675 bool flushed_caches
= false;
4677 if (fence
!= VK_NULL_HANDLE
) {
4678 for (uint32_t i
= 0; i
< submitCount
; ++i
)
4679 if (radv_submit_has_effects(pSubmits
+ i
))
4682 fence_idx
= UINT32_MAX
;
4684 for (uint32_t i
= 0; i
< submitCount
; i
++) {
4685 if (!radv_submit_has_effects(pSubmits
+ i
) && fence_idx
!= i
)
4688 VkPipelineStageFlags wait_dst_stage_mask
= 0;
4689 for (unsigned j
= 0; j
< pSubmits
[i
].waitSemaphoreCount
; ++j
) {
4690 wait_dst_stage_mask
|= pSubmits
[i
].pWaitDstStageMask
[j
];
4693 const VkTimelineSemaphoreSubmitInfo
*timeline_info
=
4694 vk_find_struct_const(pSubmits
[i
].pNext
, TIMELINE_SEMAPHORE_SUBMIT_INFO
);
4696 result
= radv_queue_submit(queue
, &(struct radv_queue_submission
) {
4697 .cmd_buffers
= pSubmits
[i
].pCommandBuffers
,
4698 .cmd_buffer_count
= pSubmits
[i
].commandBufferCount
,
4699 .wait_dst_stage_mask
= wait_dst_stage_mask
,
4700 .flush_caches
= !flushed_caches
,
4701 .wait_semaphores
= pSubmits
[i
].pWaitSemaphores
,
4702 .wait_semaphore_count
= pSubmits
[i
].waitSemaphoreCount
,
4703 .signal_semaphores
= pSubmits
[i
].pSignalSemaphores
,
4704 .signal_semaphore_count
= pSubmits
[i
].signalSemaphoreCount
,
4705 .fence
= i
== fence_idx
? fence
: VK_NULL_HANDLE
,
4706 .wait_values
= timeline_info
? timeline_info
->pWaitSemaphoreValues
: NULL
,
4707 .wait_value_count
= timeline_info
&& timeline_info
->pWaitSemaphoreValues
? timeline_info
->waitSemaphoreValueCount
: 0,
4708 .signal_values
= timeline_info
? timeline_info
->pSignalSemaphoreValues
: NULL
,
4709 .signal_value_count
= timeline_info
&& timeline_info
->pSignalSemaphoreValues
? timeline_info
->signalSemaphoreValueCount
: 0,
4711 if (result
!= VK_SUCCESS
)
4714 flushed_caches
= true;
4717 if (fence
!= VK_NULL_HANDLE
&& !submitCount
) {
4718 result
= radv_signal_fence(queue
, fence
);
4719 if (result
!= VK_SUCCESS
)
4726 VkResult
radv_QueueWaitIdle(
4729 RADV_FROM_HANDLE(radv_queue
, queue
, _queue
);
4731 pthread_mutex_lock(&queue
->pending_mutex
);
4732 while (!list_is_empty(&queue
->pending_submissions
)) {
4733 pthread_cond_wait(&queue
->device
->timeline_cond
, &queue
->pending_mutex
);
4735 pthread_mutex_unlock(&queue
->pending_mutex
);
4737 if (!queue
->device
->ws
->ctx_wait_idle(queue
->hw_ctx
,
4738 radv_queue_family_to_ring(queue
->queue_family_index
),
4740 return VK_ERROR_DEVICE_LOST
;
4745 VkResult
radv_DeviceWaitIdle(
4748 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4750 for (unsigned i
= 0; i
< RADV_MAX_QUEUE_FAMILIES
; i
++) {
4751 for (unsigned q
= 0; q
< device
->queue_count
[i
]; q
++) {
4753 radv_QueueWaitIdle(radv_queue_to_handle(&device
->queues
[i
][q
]));
4755 if (result
!= VK_SUCCESS
)
4762 VkResult
radv_EnumerateInstanceExtensionProperties(
4763 const char* pLayerName
,
4764 uint32_t* pPropertyCount
,
4765 VkExtensionProperties
* pProperties
)
4767 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
4769 for (int i
= 0; i
< RADV_INSTANCE_EXTENSION_COUNT
; i
++) {
4770 if (radv_instance_extensions_supported
.extensions
[i
]) {
4771 vk_outarray_append(&out
, prop
) {
4772 *prop
= radv_instance_extensions
[i
];
4777 return vk_outarray_status(&out
);
4780 VkResult
radv_EnumerateDeviceExtensionProperties(
4781 VkPhysicalDevice physicalDevice
,
4782 const char* pLayerName
,
4783 uint32_t* pPropertyCount
,
4784 VkExtensionProperties
* pProperties
)
4786 RADV_FROM_HANDLE(radv_physical_device
, device
, physicalDevice
);
4787 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
4789 for (int i
= 0; i
< RADV_DEVICE_EXTENSION_COUNT
; i
++) {
4790 if (device
->supported_extensions
.extensions
[i
]) {
4791 vk_outarray_append(&out
, prop
) {
4792 *prop
= radv_device_extensions
[i
];
4797 return vk_outarray_status(&out
);
4800 PFN_vkVoidFunction
radv_GetInstanceProcAddr(
4801 VkInstance _instance
,
4804 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
4806 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
4807 * when we have to return valid function pointers, NULL, or it's left
4808 * undefined. See the table for exact details.
4813 #define LOOKUP_RADV_ENTRYPOINT(entrypoint) \
4814 if (strcmp(pName, "vk" #entrypoint) == 0) \
4815 return (PFN_vkVoidFunction)radv_##entrypoint
4817 LOOKUP_RADV_ENTRYPOINT(EnumerateInstanceExtensionProperties
);
4818 LOOKUP_RADV_ENTRYPOINT(EnumerateInstanceLayerProperties
);
4819 LOOKUP_RADV_ENTRYPOINT(EnumerateInstanceVersion
);
4820 LOOKUP_RADV_ENTRYPOINT(CreateInstance
);
4822 /* GetInstanceProcAddr() can also be called with a NULL instance.
4823 * See https://gitlab.khronos.org/vulkan/vulkan/issues/2057
4825 LOOKUP_RADV_ENTRYPOINT(GetInstanceProcAddr
);
4827 #undef LOOKUP_RADV_ENTRYPOINT
4829 if (instance
== NULL
)
4832 int idx
= radv_get_instance_entrypoint_index(pName
);
4834 return instance
->dispatch
.entrypoints
[idx
];
4836 idx
= radv_get_physical_device_entrypoint_index(pName
);
4838 return instance
->physical_device_dispatch
.entrypoints
[idx
];
4840 idx
= radv_get_device_entrypoint_index(pName
);
4842 return instance
->device_dispatch
.entrypoints
[idx
];
4847 /* The loader wants us to expose a second GetInstanceProcAddr function
4848 * to work around certain LD_PRELOAD issues seen in apps.
4851 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
4852 VkInstance instance
,
4856 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
4857 VkInstance instance
,
4860 return radv_GetInstanceProcAddr(instance
, pName
);
4864 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetPhysicalDeviceProcAddr(
4865 VkInstance _instance
,
4869 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetPhysicalDeviceProcAddr(
4870 VkInstance _instance
,
4873 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
4875 if (!pName
|| !instance
)
4878 int idx
= radv_get_physical_device_entrypoint_index(pName
);
4882 return instance
->physical_device_dispatch
.entrypoints
[idx
];
4885 PFN_vkVoidFunction
radv_GetDeviceProcAddr(
4889 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4891 if (!device
|| !pName
)
4894 int idx
= radv_get_device_entrypoint_index(pName
);
4898 return device
->dispatch
.entrypoints
[idx
];
4901 bool radv_get_memory_fd(struct radv_device
*device
,
4902 struct radv_device_memory
*memory
,
4905 struct radeon_bo_metadata metadata
;
4907 if (memory
->image
) {
4908 if (memory
->image
->tiling
!= VK_IMAGE_TILING_LINEAR
)
4909 radv_init_metadata(device
, memory
->image
, &metadata
);
4910 device
->ws
->buffer_set_metadata(memory
->bo
, &metadata
);
4913 return device
->ws
->buffer_get_fd(device
->ws
, memory
->bo
,
4919 radv_free_memory(struct radv_device
*device
,
4920 const VkAllocationCallbacks
* pAllocator
,
4921 struct radv_device_memory
*mem
)
4926 #if RADV_SUPPORT_ANDROID_HARDWARE_BUFFER
4927 if (mem
->android_hardware_buffer
)
4928 AHardwareBuffer_release(mem
->android_hardware_buffer
);
4932 if (device
->overallocation_disallowed
) {
4933 mtx_lock(&device
->overallocation_mutex
);
4934 device
->allocated_memory_size
[mem
->heap_index
] -= mem
->alloc_size
;
4935 mtx_unlock(&device
->overallocation_mutex
);
4938 radv_bo_list_remove(device
, mem
->bo
);
4939 device
->ws
->buffer_destroy(mem
->bo
);
4943 vk_object_base_finish(&mem
->base
);
4944 vk_free2(&device
->vk
.alloc
, pAllocator
, mem
);
4947 static VkResult
radv_alloc_memory(struct radv_device
*device
,
4948 const VkMemoryAllocateInfo
* pAllocateInfo
,
4949 const VkAllocationCallbacks
* pAllocator
,
4950 VkDeviceMemory
* pMem
)
4952 struct radv_device_memory
*mem
;
4954 enum radeon_bo_domain domain
;
4957 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
4959 const VkImportMemoryFdInfoKHR
*import_info
=
4960 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_FD_INFO_KHR
);
4961 const VkMemoryDedicatedAllocateInfo
*dedicate_info
=
4962 vk_find_struct_const(pAllocateInfo
->pNext
, MEMORY_DEDICATED_ALLOCATE_INFO
);
4963 const VkExportMemoryAllocateInfo
*export_info
=
4964 vk_find_struct_const(pAllocateInfo
->pNext
, EXPORT_MEMORY_ALLOCATE_INFO
);
4965 const struct VkImportAndroidHardwareBufferInfoANDROID
*ahb_import_info
=
4966 vk_find_struct_const(pAllocateInfo
->pNext
,
4967 IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID
);
4968 const VkImportMemoryHostPointerInfoEXT
*host_ptr_info
=
4969 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_HOST_POINTER_INFO_EXT
);
4971 const struct wsi_memory_allocate_info
*wsi_info
=
4972 vk_find_struct_const(pAllocateInfo
->pNext
, WSI_MEMORY_ALLOCATE_INFO_MESA
);
4974 if (pAllocateInfo
->allocationSize
== 0 && !ahb_import_info
&&
4975 !(export_info
&& (export_info
->handleTypes
& VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID
))) {
4976 /* Apparently, this is allowed */
4977 *pMem
= VK_NULL_HANDLE
;
4981 mem
= vk_zalloc2(&device
->vk
.alloc
, pAllocator
, sizeof(*mem
), 8,
4982 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
4984 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
4986 vk_object_base_init(&device
->vk
, &mem
->base
,
4987 VK_OBJECT_TYPE_DEVICE_MEMORY
);
4989 if (wsi_info
&& wsi_info
->implicit_sync
)
4990 flags
|= RADEON_FLAG_IMPLICIT_SYNC
;
4992 if (dedicate_info
) {
4993 mem
->image
= radv_image_from_handle(dedicate_info
->image
);
4994 mem
->buffer
= radv_buffer_from_handle(dedicate_info
->buffer
);
5000 float priority_float
= 0.5;
5001 const struct VkMemoryPriorityAllocateInfoEXT
*priority_ext
=
5002 vk_find_struct_const(pAllocateInfo
->pNext
,
5003 MEMORY_PRIORITY_ALLOCATE_INFO_EXT
);
5005 priority_float
= priority_ext
->priority
;
5007 unsigned priority
= MIN2(RADV_BO_PRIORITY_APPLICATION_MAX
- 1,
5008 (int)(priority_float
* RADV_BO_PRIORITY_APPLICATION_MAX
));
5010 mem
->user_ptr
= NULL
;
5013 #if RADV_SUPPORT_ANDROID_HARDWARE_BUFFER
5014 mem
->android_hardware_buffer
= NULL
;
5017 if (ahb_import_info
) {
5018 result
= radv_import_ahb_memory(device
, mem
, priority
, ahb_import_info
);
5019 if (result
!= VK_SUCCESS
)
5021 } else if(export_info
&& (export_info
->handleTypes
& VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID
)) {
5022 result
= radv_create_ahb_memory(device
, mem
, priority
, pAllocateInfo
);
5023 if (result
!= VK_SUCCESS
)
5025 } else if (import_info
) {
5026 assert(import_info
->handleType
==
5027 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
5028 import_info
->handleType
==
5029 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
5030 mem
->bo
= device
->ws
->buffer_from_fd(device
->ws
, import_info
->fd
,
5033 result
= VK_ERROR_INVALID_EXTERNAL_HANDLE
;
5036 close(import_info
->fd
);
5038 } else if (host_ptr_info
) {
5039 assert(host_ptr_info
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
);
5040 mem
->bo
= device
->ws
->buffer_from_ptr(device
->ws
, host_ptr_info
->pHostPointer
,
5041 pAllocateInfo
->allocationSize
,
5044 result
= VK_ERROR_INVALID_EXTERNAL_HANDLE
;
5047 mem
->user_ptr
= host_ptr_info
->pHostPointer
;
5050 uint64_t alloc_size
= align_u64(pAllocateInfo
->allocationSize
, 4096);
5051 uint32_t heap_index
;
5053 heap_index
= device
->physical_device
->memory_properties
.memoryTypes
[pAllocateInfo
->memoryTypeIndex
].heapIndex
;
5054 domain
= device
->physical_device
->memory_domains
[pAllocateInfo
->memoryTypeIndex
];
5055 flags
|= device
->physical_device
->memory_flags
[pAllocateInfo
->memoryTypeIndex
];
5057 if (!dedicate_info
&& !import_info
&& (!export_info
|| !export_info
->handleTypes
)) {
5058 flags
|= RADEON_FLAG_NO_INTERPROCESS_SHARING
;
5059 if (device
->use_global_bo_list
) {
5060 flags
|= RADEON_FLAG_PREFER_LOCAL_BO
;
5064 if (device
->overallocation_disallowed
) {
5065 uint64_t total_size
=
5066 device
->physical_device
->memory_properties
.memoryHeaps
[heap_index
].size
;
5068 mtx_lock(&device
->overallocation_mutex
);
5069 if (device
->allocated_memory_size
[heap_index
] + alloc_size
> total_size
) {
5070 mtx_unlock(&device
->overallocation_mutex
);
5071 result
= VK_ERROR_OUT_OF_DEVICE_MEMORY
;
5074 device
->allocated_memory_size
[heap_index
] += alloc_size
;
5075 mtx_unlock(&device
->overallocation_mutex
);
5078 mem
->bo
= device
->ws
->buffer_create(device
->ws
, alloc_size
, device
->physical_device
->rad_info
.max_alignment
,
5079 domain
, flags
, priority
);
5082 if (device
->overallocation_disallowed
) {
5083 mtx_lock(&device
->overallocation_mutex
);
5084 device
->allocated_memory_size
[heap_index
] -= alloc_size
;
5085 mtx_unlock(&device
->overallocation_mutex
);
5087 result
= VK_ERROR_OUT_OF_DEVICE_MEMORY
;
5091 mem
->heap_index
= heap_index
;
5092 mem
->alloc_size
= alloc_size
;
5096 result
= radv_bo_list_add(device
, mem
->bo
);
5097 if (result
!= VK_SUCCESS
)
5101 *pMem
= radv_device_memory_to_handle(mem
);
5106 radv_free_memory(device
, pAllocator
,mem
);
5111 VkResult
radv_AllocateMemory(
5113 const VkMemoryAllocateInfo
* pAllocateInfo
,
5114 const VkAllocationCallbacks
* pAllocator
,
5115 VkDeviceMemory
* pMem
)
5117 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5118 return radv_alloc_memory(device
, pAllocateInfo
, pAllocator
, pMem
);
5121 void radv_FreeMemory(
5123 VkDeviceMemory _mem
,
5124 const VkAllocationCallbacks
* pAllocator
)
5126 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5127 RADV_FROM_HANDLE(radv_device_memory
, mem
, _mem
);
5129 radv_free_memory(device
, pAllocator
, mem
);
5132 VkResult
radv_MapMemory(
5134 VkDeviceMemory _memory
,
5135 VkDeviceSize offset
,
5137 VkMemoryMapFlags flags
,
5140 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5141 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
5149 *ppData
= mem
->user_ptr
;
5151 *ppData
= device
->ws
->buffer_map(mem
->bo
);
5158 return vk_error(device
->instance
, VK_ERROR_MEMORY_MAP_FAILED
);
5161 void radv_UnmapMemory(
5163 VkDeviceMemory _memory
)
5165 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5166 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
5171 if (mem
->user_ptr
== NULL
)
5172 device
->ws
->buffer_unmap(mem
->bo
);
5175 VkResult
radv_FlushMappedMemoryRanges(
5177 uint32_t memoryRangeCount
,
5178 const VkMappedMemoryRange
* pMemoryRanges
)
5183 VkResult
radv_InvalidateMappedMemoryRanges(
5185 uint32_t memoryRangeCount
,
5186 const VkMappedMemoryRange
* pMemoryRanges
)
5191 void radv_GetBufferMemoryRequirements(
5194 VkMemoryRequirements
* pMemoryRequirements
)
5196 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5197 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
5199 pMemoryRequirements
->memoryTypeBits
= (1u << device
->physical_device
->memory_properties
.memoryTypeCount
) - 1;
5201 if (buffer
->flags
& VK_BUFFER_CREATE_SPARSE_BINDING_BIT
)
5202 pMemoryRequirements
->alignment
= 4096;
5204 pMemoryRequirements
->alignment
= 16;
5206 pMemoryRequirements
->size
= align64(buffer
->size
, pMemoryRequirements
->alignment
);
5209 void radv_GetBufferMemoryRequirements2(
5211 const VkBufferMemoryRequirementsInfo2
*pInfo
,
5212 VkMemoryRequirements2
*pMemoryRequirements
)
5214 radv_GetBufferMemoryRequirements(device
, pInfo
->buffer
,
5215 &pMemoryRequirements
->memoryRequirements
);
5216 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
5217 switch (ext
->sType
) {
5218 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
5219 VkMemoryDedicatedRequirements
*req
=
5220 (VkMemoryDedicatedRequirements
*) ext
;
5221 req
->requiresDedicatedAllocation
= false;
5222 req
->prefersDedicatedAllocation
= req
->requiresDedicatedAllocation
;
5231 void radv_GetImageMemoryRequirements(
5234 VkMemoryRequirements
* pMemoryRequirements
)
5236 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5237 RADV_FROM_HANDLE(radv_image
, image
, _image
);
5239 pMemoryRequirements
->memoryTypeBits
= (1u << device
->physical_device
->memory_properties
.memoryTypeCount
) - 1;
5241 pMemoryRequirements
->size
= image
->size
;
5242 pMemoryRequirements
->alignment
= image
->alignment
;
5245 void radv_GetImageMemoryRequirements2(
5247 const VkImageMemoryRequirementsInfo2
*pInfo
,
5248 VkMemoryRequirements2
*pMemoryRequirements
)
5250 radv_GetImageMemoryRequirements(device
, pInfo
->image
,
5251 &pMemoryRequirements
->memoryRequirements
);
5253 RADV_FROM_HANDLE(radv_image
, image
, pInfo
->image
);
5255 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
5256 switch (ext
->sType
) {
5257 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
5258 VkMemoryDedicatedRequirements
*req
=
5259 (VkMemoryDedicatedRequirements
*) ext
;
5260 req
->requiresDedicatedAllocation
= image
->shareable
&&
5261 image
->tiling
!= VK_IMAGE_TILING_LINEAR
;
5262 req
->prefersDedicatedAllocation
= req
->requiresDedicatedAllocation
;
5271 void radv_GetImageSparseMemoryRequirements(
5274 uint32_t* pSparseMemoryRequirementCount
,
5275 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
5280 void radv_GetImageSparseMemoryRequirements2(
5282 const VkImageSparseMemoryRequirementsInfo2
*pInfo
,
5283 uint32_t* pSparseMemoryRequirementCount
,
5284 VkSparseImageMemoryRequirements2
*pSparseMemoryRequirements
)
5289 void radv_GetDeviceMemoryCommitment(
5291 VkDeviceMemory memory
,
5292 VkDeviceSize
* pCommittedMemoryInBytes
)
5294 *pCommittedMemoryInBytes
= 0;
5297 VkResult
radv_BindBufferMemory2(VkDevice device
,
5298 uint32_t bindInfoCount
,
5299 const VkBindBufferMemoryInfo
*pBindInfos
)
5301 for (uint32_t i
= 0; i
< bindInfoCount
; ++i
) {
5302 RADV_FROM_HANDLE(radv_device_memory
, mem
, pBindInfos
[i
].memory
);
5303 RADV_FROM_HANDLE(radv_buffer
, buffer
, pBindInfos
[i
].buffer
);
5306 buffer
->bo
= mem
->bo
;
5307 buffer
->offset
= pBindInfos
[i
].memoryOffset
;
5315 VkResult
radv_BindBufferMemory(
5318 VkDeviceMemory memory
,
5319 VkDeviceSize memoryOffset
)
5321 const VkBindBufferMemoryInfo info
= {
5322 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
,
5325 .memoryOffset
= memoryOffset
5328 return radv_BindBufferMemory2(device
, 1, &info
);
5331 VkResult
radv_BindImageMemory2(VkDevice device
,
5332 uint32_t bindInfoCount
,
5333 const VkBindImageMemoryInfo
*pBindInfos
)
5335 for (uint32_t i
= 0; i
< bindInfoCount
; ++i
) {
5336 RADV_FROM_HANDLE(radv_device_memory
, mem
, pBindInfos
[i
].memory
);
5337 RADV_FROM_HANDLE(radv_image
, image
, pBindInfos
[i
].image
);
5340 image
->bo
= mem
->bo
;
5341 image
->offset
= pBindInfos
[i
].memoryOffset
;
5351 VkResult
radv_BindImageMemory(
5354 VkDeviceMemory memory
,
5355 VkDeviceSize memoryOffset
)
5357 const VkBindImageMemoryInfo info
= {
5358 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
,
5361 .memoryOffset
= memoryOffset
5364 return radv_BindImageMemory2(device
, 1, &info
);
5367 static bool radv_sparse_bind_has_effects(const VkBindSparseInfo
*info
)
5369 return info
->bufferBindCount
||
5370 info
->imageOpaqueBindCount
||
5371 info
->imageBindCount
||
5372 info
->waitSemaphoreCount
||
5373 info
->signalSemaphoreCount
;
5376 VkResult
radv_QueueBindSparse(
5378 uint32_t bindInfoCount
,
5379 const VkBindSparseInfo
* pBindInfo
,
5382 RADV_FROM_HANDLE(radv_queue
, queue
, _queue
);
5384 uint32_t fence_idx
= 0;
5386 if (fence
!= VK_NULL_HANDLE
) {
5387 for (uint32_t i
= 0; i
< bindInfoCount
; ++i
)
5388 if (radv_sparse_bind_has_effects(pBindInfo
+ i
))
5391 fence_idx
= UINT32_MAX
;
5393 for (uint32_t i
= 0; i
< bindInfoCount
; ++i
) {
5394 if (i
!= fence_idx
&& !radv_sparse_bind_has_effects(pBindInfo
+ i
))
5397 const VkTimelineSemaphoreSubmitInfo
*timeline_info
=
5398 vk_find_struct_const(pBindInfo
[i
].pNext
, TIMELINE_SEMAPHORE_SUBMIT_INFO
);
5400 VkResult result
= radv_queue_submit(queue
, &(struct radv_queue_submission
) {
5401 .buffer_binds
= pBindInfo
[i
].pBufferBinds
,
5402 .buffer_bind_count
= pBindInfo
[i
].bufferBindCount
,
5403 .image_opaque_binds
= pBindInfo
[i
].pImageOpaqueBinds
,
5404 .image_opaque_bind_count
= pBindInfo
[i
].imageOpaqueBindCount
,
5405 .wait_semaphores
= pBindInfo
[i
].pWaitSemaphores
,
5406 .wait_semaphore_count
= pBindInfo
[i
].waitSemaphoreCount
,
5407 .signal_semaphores
= pBindInfo
[i
].pSignalSemaphores
,
5408 .signal_semaphore_count
= pBindInfo
[i
].signalSemaphoreCount
,
5409 .fence
= i
== fence_idx
? fence
: VK_NULL_HANDLE
,
5410 .wait_values
= timeline_info
? timeline_info
->pWaitSemaphoreValues
: NULL
,
5411 .wait_value_count
= timeline_info
&& timeline_info
->pWaitSemaphoreValues
? timeline_info
->waitSemaphoreValueCount
: 0,
5412 .signal_values
= timeline_info
? timeline_info
->pSignalSemaphoreValues
: NULL
,
5413 .signal_value_count
= timeline_info
&& timeline_info
->pSignalSemaphoreValues
? timeline_info
->signalSemaphoreValueCount
: 0,
5416 if (result
!= VK_SUCCESS
)
5420 if (fence
!= VK_NULL_HANDLE
&& !bindInfoCount
) {
5421 result
= radv_signal_fence(queue
, fence
);
5422 if (result
!= VK_SUCCESS
)
5430 radv_destroy_fence_part(struct radv_device
*device
,
5431 struct radv_fence_part
*part
)
5433 switch (part
->kind
) {
5434 case RADV_FENCE_NONE
:
5436 case RADV_FENCE_WINSYS
:
5437 device
->ws
->destroy_fence(part
->fence
);
5439 case RADV_FENCE_SYNCOBJ
:
5440 device
->ws
->destroy_syncobj(device
->ws
, part
->syncobj
);
5442 case RADV_FENCE_WSI
:
5443 part
->fence_wsi
->destroy(part
->fence_wsi
);
5446 unreachable("Invalid fence type");
5449 part
->kind
= RADV_FENCE_NONE
;
5453 radv_destroy_fence(struct radv_device
*device
,
5454 const VkAllocationCallbacks
*pAllocator
,
5455 struct radv_fence
*fence
)
5457 radv_destroy_fence_part(device
, &fence
->temporary
);
5458 radv_destroy_fence_part(device
, &fence
->permanent
);
5460 vk_object_base_finish(&fence
->base
);
5461 vk_free2(&device
->vk
.alloc
, pAllocator
, fence
);
5464 VkResult
radv_CreateFence(
5466 const VkFenceCreateInfo
* pCreateInfo
,
5467 const VkAllocationCallbacks
* pAllocator
,
5470 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5471 const VkExportFenceCreateInfo
*export
=
5472 vk_find_struct_const(pCreateInfo
->pNext
, EXPORT_FENCE_CREATE_INFO
);
5473 VkExternalFenceHandleTypeFlags handleTypes
=
5474 export
? export
->handleTypes
: 0;
5475 struct radv_fence
*fence
;
5477 fence
= vk_zalloc2(&device
->vk
.alloc
, pAllocator
, sizeof(*fence
), 8,
5478 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
5480 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
5482 vk_object_base_init(&device
->vk
, &fence
->base
, VK_OBJECT_TYPE_FENCE
);
5484 if (device
->always_use_syncobj
|| handleTypes
) {
5485 fence
->permanent
.kind
= RADV_FENCE_SYNCOBJ
;
5487 bool create_signaled
= false;
5488 if (pCreateInfo
->flags
& VK_FENCE_CREATE_SIGNALED_BIT
)
5489 create_signaled
= true;
5491 int ret
= device
->ws
->create_syncobj(device
->ws
, create_signaled
,
5492 &fence
->permanent
.syncobj
);
5494 radv_destroy_fence(device
, pAllocator
, fence
);
5495 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
5498 fence
->permanent
.kind
= RADV_FENCE_WINSYS
;
5500 fence
->permanent
.fence
= device
->ws
->create_fence();
5501 if (!fence
->permanent
.fence
) {
5502 vk_free2(&device
->vk
.alloc
, pAllocator
, fence
);
5503 radv_destroy_fence(device
, pAllocator
, fence
);
5504 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
5506 if (pCreateInfo
->flags
& VK_FENCE_CREATE_SIGNALED_BIT
)
5507 device
->ws
->signal_fence(fence
->permanent
.fence
);
5510 *pFence
= radv_fence_to_handle(fence
);
5516 void radv_DestroyFence(
5519 const VkAllocationCallbacks
* pAllocator
)
5521 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5522 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
5527 radv_destroy_fence(device
, pAllocator
, fence
);
5530 static bool radv_all_fences_plain_and_submitted(struct radv_device
*device
,
5531 uint32_t fenceCount
, const VkFence
*pFences
)
5533 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
5534 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
5536 struct radv_fence_part
*part
=
5537 fence
->temporary
.kind
!= RADV_FENCE_NONE
?
5538 &fence
->temporary
: &fence
->permanent
;
5539 if (part
->kind
!= RADV_FENCE_WINSYS
||
5540 !device
->ws
->is_fence_waitable(part
->fence
))
5546 static bool radv_all_fences_syncobj(uint32_t fenceCount
, const VkFence
*pFences
)
5548 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
5549 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
5551 struct radv_fence_part
*part
=
5552 fence
->temporary
.kind
!= RADV_FENCE_NONE
?
5553 &fence
->temporary
: &fence
->permanent
;
5554 if (part
->kind
!= RADV_FENCE_SYNCOBJ
)
5560 VkResult
radv_WaitForFences(
5562 uint32_t fenceCount
,
5563 const VkFence
* pFences
,
5567 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5568 timeout
= radv_get_absolute_timeout(timeout
);
5570 if (device
->always_use_syncobj
&&
5571 radv_all_fences_syncobj(fenceCount
, pFences
))
5573 uint32_t *handles
= malloc(sizeof(uint32_t) * fenceCount
);
5575 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
5577 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
5578 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
5580 struct radv_fence_part
*part
=
5581 fence
->temporary
.kind
!= RADV_FENCE_NONE
?
5582 &fence
->temporary
: &fence
->permanent
;
5584 assert(part
->kind
== RADV_FENCE_SYNCOBJ
);
5585 handles
[i
] = part
->syncobj
;
5588 bool success
= device
->ws
->wait_syncobj(device
->ws
, handles
, fenceCount
, waitAll
, timeout
);
5591 return success
? VK_SUCCESS
: VK_TIMEOUT
;
5594 if (!waitAll
&& fenceCount
> 1) {
5595 /* Not doing this by default for waitAll, due to needing to allocate twice. */
5596 if (device
->physical_device
->rad_info
.drm_minor
>= 10 && radv_all_fences_plain_and_submitted(device
, fenceCount
, pFences
)) {
5597 uint32_t wait_count
= 0;
5598 struct radeon_winsys_fence
**fences
= malloc(sizeof(struct radeon_winsys_fence
*) * fenceCount
);
5600 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
5602 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
5603 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
5605 struct radv_fence_part
*part
=
5606 fence
->temporary
.kind
!= RADV_FENCE_NONE
?
5607 &fence
->temporary
: &fence
->permanent
;
5608 assert(part
->kind
== RADV_FENCE_WINSYS
);
5610 if (device
->ws
->fence_wait(device
->ws
, part
->fence
, false, 0)) {
5615 fences
[wait_count
++] = part
->fence
;
5618 bool success
= device
->ws
->fences_wait(device
->ws
, fences
, wait_count
,
5619 waitAll
, timeout
- radv_get_current_time());
5622 return success
? VK_SUCCESS
: VK_TIMEOUT
;
5625 while(radv_get_current_time() <= timeout
) {
5626 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
5627 if (radv_GetFenceStatus(_device
, pFences
[i
]) == VK_SUCCESS
)
5634 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
5635 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
5636 bool expired
= false;
5638 struct radv_fence_part
*part
=
5639 fence
->temporary
.kind
!= RADV_FENCE_NONE
?
5640 &fence
->temporary
: &fence
->permanent
;
5642 switch (part
->kind
) {
5643 case RADV_FENCE_NONE
:
5645 case RADV_FENCE_WINSYS
:
5646 if (!device
->ws
->is_fence_waitable(part
->fence
)) {
5647 while (!device
->ws
->is_fence_waitable(part
->fence
) &&
5648 radv_get_current_time() <= timeout
)
5652 expired
= device
->ws
->fence_wait(device
->ws
,
5658 case RADV_FENCE_SYNCOBJ
:
5659 if (!device
->ws
->wait_syncobj(device
->ws
,
5660 &part
->syncobj
, 1, true,
5664 case RADV_FENCE_WSI
: {
5665 VkResult result
= part
->fence_wsi
->wait(part
->fence_wsi
, timeout
);
5666 if (result
!= VK_SUCCESS
)
5671 unreachable("Invalid fence type");
5678 VkResult
radv_ResetFences(VkDevice _device
,
5679 uint32_t fenceCount
,
5680 const VkFence
*pFences
)
5682 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5684 for (unsigned i
= 0; i
< fenceCount
; ++i
) {
5685 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
5687 /* From the Vulkan 1.0.53 spec:
5689 * "If any member of pFences currently has its payload
5690 * imported with temporary permanence, that fence’s prior
5691 * permanent payload is irst restored. The remaining
5692 * operations described therefore operate on the restored
5695 if (fence
->temporary
.kind
!= RADV_FENCE_NONE
)
5696 radv_destroy_fence_part(device
, &fence
->temporary
);
5698 struct radv_fence_part
*part
= &fence
->permanent
;
5700 switch (part
->kind
) {
5701 case RADV_FENCE_WSI
:
5702 device
->ws
->reset_fence(part
->fence
);
5704 case RADV_FENCE_SYNCOBJ
:
5705 device
->ws
->reset_syncobj(device
->ws
, part
->syncobj
);
5708 unreachable("Invalid fence type");
5715 VkResult
radv_GetFenceStatus(VkDevice _device
, VkFence _fence
)
5717 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5718 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
5720 struct radv_fence_part
*part
=
5721 fence
->temporary
.kind
!= RADV_FENCE_NONE
?
5722 &fence
->temporary
: &fence
->permanent
;
5724 switch (part
->kind
) {
5725 case RADV_FENCE_NONE
:
5727 case RADV_FENCE_WINSYS
:
5728 if (!device
->ws
->fence_wait(device
->ws
, part
->fence
, false, 0))
5729 return VK_NOT_READY
;
5731 case RADV_FENCE_SYNCOBJ
: {
5732 bool success
= device
->ws
->wait_syncobj(device
->ws
,
5733 &part
->syncobj
, 1, true, 0);
5735 return VK_NOT_READY
;
5738 case RADV_FENCE_WSI
: {
5739 VkResult result
= part
->fence_wsi
->wait(part
->fence_wsi
, 0);
5740 if (result
!= VK_SUCCESS
) {
5741 if (result
== VK_TIMEOUT
)
5742 return VK_NOT_READY
;
5748 unreachable("Invalid fence type");
5755 // Queue semaphore functions
5758 radv_create_timeline(struct radv_timeline
*timeline
, uint64_t value
)
5760 timeline
->highest_signaled
= value
;
5761 timeline
->highest_submitted
= value
;
5762 list_inithead(&timeline
->points
);
5763 list_inithead(&timeline
->free_points
);
5764 list_inithead(&timeline
->waiters
);
5765 pthread_mutex_init(&timeline
->mutex
, NULL
);
5769 radv_destroy_timeline(struct radv_device
*device
,
5770 struct radv_timeline
*timeline
)
5772 list_for_each_entry_safe(struct radv_timeline_point
, point
,
5773 &timeline
->free_points
, list
) {
5774 list_del(&point
->list
);
5775 device
->ws
->destroy_syncobj(device
->ws
, point
->syncobj
);
5778 list_for_each_entry_safe(struct radv_timeline_point
, point
,
5779 &timeline
->points
, list
) {
5780 list_del(&point
->list
);
5781 device
->ws
->destroy_syncobj(device
->ws
, point
->syncobj
);
5784 pthread_mutex_destroy(&timeline
->mutex
);
5788 radv_timeline_gc_locked(struct radv_device
*device
,
5789 struct radv_timeline
*timeline
)
5791 list_for_each_entry_safe(struct radv_timeline_point
, point
,
5792 &timeline
->points
, list
) {
5793 if (point
->wait_count
|| point
->value
> timeline
->highest_submitted
)
5796 if (device
->ws
->wait_syncobj(device
->ws
, &point
->syncobj
, 1, true, 0)) {
5797 timeline
->highest_signaled
= point
->value
;
5798 list_del(&point
->list
);
5799 list_add(&point
->list
, &timeline
->free_points
);
5804 static struct radv_timeline_point
*
5805 radv_timeline_find_point_at_least_locked(struct radv_device
*device
,
5806 struct radv_timeline
*timeline
,
5809 radv_timeline_gc_locked(device
, timeline
);
5811 if (p
<= timeline
->highest_signaled
)
5814 list_for_each_entry(struct radv_timeline_point
, point
,
5815 &timeline
->points
, list
) {
5816 if (point
->value
>= p
) {
5817 ++point
->wait_count
;
5824 static struct radv_timeline_point
*
5825 radv_timeline_add_point_locked(struct radv_device
*device
,
5826 struct radv_timeline
*timeline
,
5829 radv_timeline_gc_locked(device
, timeline
);
5831 struct radv_timeline_point
*ret
= NULL
;
5832 struct radv_timeline_point
*prev
= NULL
;
5835 if (p
<= timeline
->highest_signaled
)
5838 list_for_each_entry(struct radv_timeline_point
, point
,
5839 &timeline
->points
, list
) {
5840 if (point
->value
== p
) {
5844 if (point
->value
< p
)
5848 if (list_is_empty(&timeline
->free_points
)) {
5849 ret
= malloc(sizeof(struct radv_timeline_point
));
5850 r
= device
->ws
->create_syncobj(device
->ws
, false, &ret
->syncobj
);
5856 ret
= list_first_entry(&timeline
->free_points
, struct radv_timeline_point
, list
);
5857 list_del(&ret
->list
);
5859 device
->ws
->reset_syncobj(device
->ws
, ret
->syncobj
);
5863 ret
->wait_count
= 1;
5866 list_add(&ret
->list
, &prev
->list
);
5868 list_addtail(&ret
->list
, &timeline
->points
);
5875 radv_timeline_wait(struct radv_device
*device
,
5876 struct radv_timeline
*timeline
,
5878 uint64_t abs_timeout
)
5880 pthread_mutex_lock(&timeline
->mutex
);
5882 while(timeline
->highest_submitted
< value
) {
5883 struct timespec abstime
;
5884 timespec_from_nsec(&abstime
, abs_timeout
);
5886 pthread_cond_timedwait(&device
->timeline_cond
, &timeline
->mutex
, &abstime
);
5888 if (radv_get_current_time() >= abs_timeout
&& timeline
->highest_submitted
< value
) {
5889 pthread_mutex_unlock(&timeline
->mutex
);
5894 struct radv_timeline_point
*point
= radv_timeline_find_point_at_least_locked(device
, timeline
, value
);
5895 pthread_mutex_unlock(&timeline
->mutex
);
5899 bool success
= device
->ws
->wait_syncobj(device
->ws
, &point
->syncobj
, 1, true, abs_timeout
);
5901 pthread_mutex_lock(&timeline
->mutex
);
5902 point
->wait_count
--;
5903 pthread_mutex_unlock(&timeline
->mutex
);
5904 return success
? VK_SUCCESS
: VK_TIMEOUT
;
5908 radv_timeline_trigger_waiters_locked(struct radv_timeline
*timeline
,
5909 struct list_head
*processing_list
)
5911 list_for_each_entry_safe(struct radv_timeline_waiter
, waiter
,
5912 &timeline
->waiters
, list
) {
5913 if (waiter
->value
> timeline
->highest_submitted
)
5916 radv_queue_trigger_submission(waiter
->submission
, 1, processing_list
);
5917 list_del(&waiter
->list
);
5922 void radv_destroy_semaphore_part(struct radv_device
*device
,
5923 struct radv_semaphore_part
*part
)
5925 switch(part
->kind
) {
5926 case RADV_SEMAPHORE_NONE
:
5928 case RADV_SEMAPHORE_WINSYS
:
5929 device
->ws
->destroy_sem(part
->ws_sem
);
5931 case RADV_SEMAPHORE_TIMELINE
:
5932 radv_destroy_timeline(device
, &part
->timeline
);
5934 case RADV_SEMAPHORE_SYNCOBJ
:
5935 case RADV_SEMAPHORE_TIMELINE_SYNCOBJ
:
5936 device
->ws
->destroy_syncobj(device
->ws
, part
->syncobj
);
5939 part
->kind
= RADV_SEMAPHORE_NONE
;
5942 static VkSemaphoreTypeKHR
5943 radv_get_semaphore_type(const void *pNext
, uint64_t *initial_value
)
5945 const VkSemaphoreTypeCreateInfo
*type_info
=
5946 vk_find_struct_const(pNext
, SEMAPHORE_TYPE_CREATE_INFO
);
5949 return VK_SEMAPHORE_TYPE_BINARY
;
5952 *initial_value
= type_info
->initialValue
;
5953 return type_info
->semaphoreType
;
5957 radv_destroy_semaphore(struct radv_device
*device
,
5958 const VkAllocationCallbacks
*pAllocator
,
5959 struct radv_semaphore
*sem
)
5961 radv_destroy_semaphore_part(device
, &sem
->temporary
);
5962 radv_destroy_semaphore_part(device
, &sem
->permanent
);
5963 vk_object_base_finish(&sem
->base
);
5964 vk_free2(&device
->vk
.alloc
, pAllocator
, sem
);
5967 VkResult
radv_CreateSemaphore(
5969 const VkSemaphoreCreateInfo
* pCreateInfo
,
5970 const VkAllocationCallbacks
* pAllocator
,
5971 VkSemaphore
* pSemaphore
)
5973 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5974 const VkExportSemaphoreCreateInfo
*export
=
5975 vk_find_struct_const(pCreateInfo
->pNext
, EXPORT_SEMAPHORE_CREATE_INFO
);
5976 VkExternalSemaphoreHandleTypeFlags handleTypes
=
5977 export
? export
->handleTypes
: 0;
5978 uint64_t initial_value
= 0;
5979 VkSemaphoreTypeKHR type
= radv_get_semaphore_type(pCreateInfo
->pNext
, &initial_value
);
5981 struct radv_semaphore
*sem
= vk_alloc2(&device
->vk
.alloc
, pAllocator
,
5983 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
5985 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
5987 vk_object_base_init(&device
->vk
, &sem
->base
,
5988 VK_OBJECT_TYPE_SEMAPHORE
);
5990 sem
->temporary
.kind
= RADV_SEMAPHORE_NONE
;
5991 sem
->permanent
.kind
= RADV_SEMAPHORE_NONE
;
5993 if (type
== VK_SEMAPHORE_TYPE_TIMELINE
&&
5994 device
->physical_device
->rad_info
.has_timeline_syncobj
) {
5995 int ret
= device
->ws
->create_syncobj(device
->ws
, false, &sem
->permanent
.syncobj
);
5997 radv_destroy_semaphore(device
, pAllocator
, sem
);
5998 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
6000 device
->ws
->signal_syncobj(device
->ws
, sem
->permanent
.syncobj
, initial_value
);
6001 sem
->permanent
.timeline_syncobj
.max_point
= initial_value
;
6002 sem
->permanent
.kind
= RADV_SEMAPHORE_TIMELINE_SYNCOBJ
;
6003 } else if (type
== VK_SEMAPHORE_TYPE_TIMELINE
) {
6004 radv_create_timeline(&sem
->permanent
.timeline
, initial_value
);
6005 sem
->permanent
.kind
= RADV_SEMAPHORE_TIMELINE
;
6006 } else if (device
->always_use_syncobj
|| handleTypes
) {
6007 assert (device
->physical_device
->rad_info
.has_syncobj
);
6008 int ret
= device
->ws
->create_syncobj(device
->ws
, false,
6009 &sem
->permanent
.syncobj
);
6011 radv_destroy_semaphore(device
, pAllocator
, sem
);
6012 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
6014 sem
->permanent
.kind
= RADV_SEMAPHORE_SYNCOBJ
;
6016 sem
->permanent
.ws_sem
= device
->ws
->create_sem(device
->ws
);
6017 if (!sem
->permanent
.ws_sem
) {
6018 radv_destroy_semaphore(device
, pAllocator
, sem
);
6019 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
6021 sem
->permanent
.kind
= RADV_SEMAPHORE_WINSYS
;
6024 *pSemaphore
= radv_semaphore_to_handle(sem
);
6028 void radv_DestroySemaphore(
6030 VkSemaphore _semaphore
,
6031 const VkAllocationCallbacks
* pAllocator
)
6033 RADV_FROM_HANDLE(radv_device
, device
, _device
);
6034 RADV_FROM_HANDLE(radv_semaphore
, sem
, _semaphore
);
6038 radv_destroy_semaphore(device
, pAllocator
, sem
);
6042 radv_GetSemaphoreCounterValue(VkDevice _device
,
6043 VkSemaphore _semaphore
,
6046 RADV_FROM_HANDLE(radv_device
, device
, _device
);
6047 RADV_FROM_HANDLE(radv_semaphore
, semaphore
, _semaphore
);
6049 struct radv_semaphore_part
*part
=
6050 semaphore
->temporary
.kind
!= RADV_SEMAPHORE_NONE
? &semaphore
->temporary
: &semaphore
->permanent
;
6052 switch (part
->kind
) {
6053 case RADV_SEMAPHORE_TIMELINE
: {
6054 pthread_mutex_lock(&part
->timeline
.mutex
);
6055 radv_timeline_gc_locked(device
, &part
->timeline
);
6056 *pValue
= part
->timeline
.highest_signaled
;
6057 pthread_mutex_unlock(&part
->timeline
.mutex
);
6060 case RADV_SEMAPHORE_TIMELINE_SYNCOBJ
: {
6061 return device
->ws
->query_syncobj(device
->ws
, part
->syncobj
, pValue
);
6063 case RADV_SEMAPHORE_NONE
:
6064 case RADV_SEMAPHORE_SYNCOBJ
:
6065 case RADV_SEMAPHORE_WINSYS
:
6066 unreachable("Invalid semaphore type");
6068 unreachable("Unhandled semaphore type");
6073 radv_wait_timelines(struct radv_device
*device
,
6074 const VkSemaphoreWaitInfo
* pWaitInfo
,
6075 uint64_t abs_timeout
)
6077 if ((pWaitInfo
->flags
& VK_SEMAPHORE_WAIT_ANY_BIT_KHR
) && pWaitInfo
->semaphoreCount
> 1) {
6079 for(uint32_t i
= 0; i
< pWaitInfo
->semaphoreCount
; ++i
) {
6080 RADV_FROM_HANDLE(radv_semaphore
, semaphore
, pWaitInfo
->pSemaphores
[i
]);
6081 VkResult result
= radv_timeline_wait(device
, &semaphore
->permanent
.timeline
, pWaitInfo
->pValues
[i
], 0);
6083 if (result
== VK_SUCCESS
)
6086 if (radv_get_current_time() > abs_timeout
)
6091 for(uint32_t i
= 0; i
< pWaitInfo
->semaphoreCount
; ++i
) {
6092 RADV_FROM_HANDLE(radv_semaphore
, semaphore
, pWaitInfo
->pSemaphores
[i
]);
6093 VkResult result
= radv_timeline_wait(device
, &semaphore
->permanent
.timeline
, pWaitInfo
->pValues
[i
], abs_timeout
);
6095 if (result
!= VK_SUCCESS
)
6101 radv_WaitSemaphores(VkDevice _device
,
6102 const VkSemaphoreWaitInfo
* pWaitInfo
,
6105 RADV_FROM_HANDLE(radv_device
, device
, _device
);
6106 uint64_t abs_timeout
= radv_get_absolute_timeout(timeout
);
6108 if (radv_semaphore_from_handle(pWaitInfo
->pSemaphores
[0])->permanent
.kind
== RADV_SEMAPHORE_TIMELINE
)
6109 return radv_wait_timelines(device
, pWaitInfo
, abs_timeout
);
6111 if (pWaitInfo
->semaphoreCount
> UINT32_MAX
/ sizeof(uint32_t))
6112 return vk_errorf(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
, "semaphoreCount integer overflow");
6114 bool wait_all
= !(pWaitInfo
->flags
& VK_SEMAPHORE_WAIT_ANY_BIT_KHR
);
6115 uint32_t *handles
= malloc(sizeof(*handles
) * pWaitInfo
->semaphoreCount
);
6117 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
6119 for (uint32_t i
= 0; i
< pWaitInfo
->semaphoreCount
; ++i
) {
6120 RADV_FROM_HANDLE(radv_semaphore
, semaphore
, pWaitInfo
->pSemaphores
[i
]);
6121 handles
[i
] = semaphore
->permanent
.syncobj
;
6124 bool success
= device
->ws
->wait_timeline_syncobj(device
->ws
, handles
, pWaitInfo
->pValues
,
6125 pWaitInfo
->semaphoreCount
, wait_all
, false,
6128 return success
? VK_SUCCESS
: VK_TIMEOUT
;
6132 radv_SignalSemaphore(VkDevice _device
,
6133 const VkSemaphoreSignalInfo
* pSignalInfo
)
6135 RADV_FROM_HANDLE(radv_device
, device
, _device
);
6136 RADV_FROM_HANDLE(radv_semaphore
, semaphore
, pSignalInfo
->semaphore
);
6138 struct radv_semaphore_part
*part
=
6139 semaphore
->temporary
.kind
!= RADV_SEMAPHORE_NONE
? &semaphore
->temporary
: &semaphore
->permanent
;
6141 switch(part
->kind
) {
6142 case RADV_SEMAPHORE_TIMELINE
: {
6143 pthread_mutex_lock(&part
->timeline
.mutex
);
6144 radv_timeline_gc_locked(device
, &part
->timeline
);
6145 part
->timeline
.highest_submitted
= MAX2(part
->timeline
.highest_submitted
, pSignalInfo
->value
);
6146 part
->timeline
.highest_signaled
= MAX2(part
->timeline
.highest_signaled
, pSignalInfo
->value
);
6148 struct list_head processing_list
;
6149 list_inithead(&processing_list
);
6150 radv_timeline_trigger_waiters_locked(&part
->timeline
, &processing_list
);
6151 pthread_mutex_unlock(&part
->timeline
.mutex
);
6153 VkResult result
= radv_process_submissions(&processing_list
);
6155 /* This needs to happen after radv_process_submissions, so
6156 * that any submitted submissions that are now unblocked get
6157 * processed before we wake the application. This way we
6158 * ensure that any binary semaphores that are now unblocked
6159 * are usable by the application. */
6160 pthread_cond_broadcast(&device
->timeline_cond
);
6164 case RADV_SEMAPHORE_TIMELINE_SYNCOBJ
: {
6165 part
->timeline_syncobj
.max_point
= MAX2(part
->timeline_syncobj
.max_point
, pSignalInfo
->value
);
6166 device
->ws
->signal_syncobj(device
->ws
, part
->syncobj
, pSignalInfo
->value
);
6169 case RADV_SEMAPHORE_NONE
:
6170 case RADV_SEMAPHORE_SYNCOBJ
:
6171 case RADV_SEMAPHORE_WINSYS
:
6172 unreachable("Invalid semaphore type");
6177 static void radv_destroy_event(struct radv_device
*device
,
6178 const VkAllocationCallbacks
* pAllocator
,
6179 struct radv_event
*event
)
6182 device
->ws
->buffer_destroy(event
->bo
);
6184 vk_object_base_finish(&event
->base
);
6185 vk_free2(&device
->vk
.alloc
, pAllocator
, event
);
6188 VkResult
radv_CreateEvent(
6190 const VkEventCreateInfo
* pCreateInfo
,
6191 const VkAllocationCallbacks
* pAllocator
,
6194 RADV_FROM_HANDLE(radv_device
, device
, _device
);
6195 struct radv_event
*event
= vk_alloc2(&device
->vk
.alloc
, pAllocator
,
6197 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
6200 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
6202 vk_object_base_init(&device
->vk
, &event
->base
, VK_OBJECT_TYPE_EVENT
);
6204 event
->bo
= device
->ws
->buffer_create(device
->ws
, 8, 8,
6206 RADEON_FLAG_VA_UNCACHED
| RADEON_FLAG_CPU_ACCESS
| RADEON_FLAG_NO_INTERPROCESS_SHARING
,
6207 RADV_BO_PRIORITY_FENCE
);
6209 radv_destroy_event(device
, pAllocator
, event
);
6210 return vk_error(device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
6213 event
->map
= (uint64_t*)device
->ws
->buffer_map(event
->bo
);
6215 radv_destroy_event(device
, pAllocator
, event
);
6216 return vk_error(device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
6219 *pEvent
= radv_event_to_handle(event
);
6224 void radv_DestroyEvent(
6227 const VkAllocationCallbacks
* pAllocator
)
6229 RADV_FROM_HANDLE(radv_device
, device
, _device
);
6230 RADV_FROM_HANDLE(radv_event
, event
, _event
);
6235 radv_destroy_event(device
, pAllocator
, event
);
6238 VkResult
radv_GetEventStatus(
6242 RADV_FROM_HANDLE(radv_event
, event
, _event
);
6244 if (*event
->map
== 1)
6245 return VK_EVENT_SET
;
6246 return VK_EVENT_RESET
;
6249 VkResult
radv_SetEvent(
6253 RADV_FROM_HANDLE(radv_event
, event
, _event
);
6259 VkResult
radv_ResetEvent(
6263 RADV_FROM_HANDLE(radv_event
, event
, _event
);
6270 radv_destroy_buffer(struct radv_device
*device
,
6271 const VkAllocationCallbacks
*pAllocator
,
6272 struct radv_buffer
*buffer
)
6274 if ((buffer
->flags
& VK_BUFFER_CREATE_SPARSE_BINDING_BIT
) && buffer
->bo
)
6275 device
->ws
->buffer_destroy(buffer
->bo
);
6277 vk_object_base_finish(&buffer
->base
);
6278 vk_free2(&device
->vk
.alloc
, pAllocator
, buffer
);
6281 VkResult
radv_CreateBuffer(
6283 const VkBufferCreateInfo
* pCreateInfo
,
6284 const VkAllocationCallbacks
* pAllocator
,
6287 RADV_FROM_HANDLE(radv_device
, device
, _device
);
6288 struct radv_buffer
*buffer
;
6290 if (pCreateInfo
->size
> RADV_MAX_MEMORY_ALLOCATION_SIZE
)
6291 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
6293 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
6295 buffer
= vk_alloc2(&device
->vk
.alloc
, pAllocator
, sizeof(*buffer
), 8,
6296 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
6298 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
6300 vk_object_base_init(&device
->vk
, &buffer
->base
, VK_OBJECT_TYPE_BUFFER
);
6302 buffer
->size
= pCreateInfo
->size
;
6303 buffer
->usage
= pCreateInfo
->usage
;
6306 buffer
->flags
= pCreateInfo
->flags
;
6308 buffer
->shareable
= vk_find_struct_const(pCreateInfo
->pNext
,
6309 EXTERNAL_MEMORY_BUFFER_CREATE_INFO
) != NULL
;
6311 if (pCreateInfo
->flags
& VK_BUFFER_CREATE_SPARSE_BINDING_BIT
) {
6312 buffer
->bo
= device
->ws
->buffer_create(device
->ws
,
6313 align64(buffer
->size
, 4096),
6314 4096, 0, RADEON_FLAG_VIRTUAL
,
6315 RADV_BO_PRIORITY_VIRTUAL
);
6317 radv_destroy_buffer(device
, pAllocator
, buffer
);
6318 return vk_error(device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
6322 *pBuffer
= radv_buffer_to_handle(buffer
);
6327 void radv_DestroyBuffer(
6330 const VkAllocationCallbacks
* pAllocator
)
6332 RADV_FROM_HANDLE(radv_device
, device
, _device
);
6333 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
6338 radv_destroy_buffer(device
, pAllocator
, buffer
);
6341 VkDeviceAddress
radv_GetBufferDeviceAddress(
6343 const VkBufferDeviceAddressInfo
* pInfo
)
6345 RADV_FROM_HANDLE(radv_buffer
, buffer
, pInfo
->buffer
);
6346 return radv_buffer_get_va(buffer
->bo
) + buffer
->offset
;
6350 uint64_t radv_GetBufferOpaqueCaptureAddress(VkDevice device
,
6351 const VkBufferDeviceAddressInfo
* pInfo
)
6356 uint64_t radv_GetDeviceMemoryOpaqueCaptureAddress(VkDevice device
,
6357 const VkDeviceMemoryOpaqueCaptureAddressInfo
* pInfo
)
6362 static inline unsigned
6363 si_tile_mode_index(const struct radv_image_plane
*plane
, unsigned level
, bool stencil
)
6366 return plane
->surface
.u
.legacy
.stencil_tiling_index
[level
];
6368 return plane
->surface
.u
.legacy
.tiling_index
[level
];
6371 static uint32_t radv_surface_max_layer_count(struct radv_image_view
*iview
)
6373 return iview
->type
== VK_IMAGE_VIEW_TYPE_3D
? iview
->extent
.depth
: (iview
->base_layer
+ iview
->layer_count
);
6377 radv_init_dcc_control_reg(struct radv_device
*device
,
6378 struct radv_image_view
*iview
)
6380 unsigned max_uncompressed_block_size
= V_028C78_MAX_BLOCK_SIZE_256B
;
6381 unsigned min_compressed_block_size
= V_028C78_MIN_BLOCK_SIZE_32B
;
6382 unsigned max_compressed_block_size
;
6383 unsigned independent_128b_blocks
;
6384 unsigned independent_64b_blocks
;
6386 if (!radv_dcc_enabled(iview
->image
, iview
->base_mip
))
6389 if (!device
->physical_device
->rad_info
.has_dedicated_vram
) {
6390 /* amdvlk: [min-compressed-block-size] should be set to 32 for
6391 * dGPU and 64 for APU because all of our APUs to date use
6392 * DIMMs which have a request granularity size of 64B while all
6393 * other chips have a 32B request size.
6395 min_compressed_block_size
= V_028C78_MIN_BLOCK_SIZE_64B
;
6398 if (device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
6399 max_compressed_block_size
= V_028C78_MAX_BLOCK_SIZE_128B
;
6400 independent_64b_blocks
= 0;
6401 independent_128b_blocks
= 1;
6403 independent_128b_blocks
= 0;
6405 if (iview
->image
->info
.samples
> 1) {
6406 if (iview
->image
->planes
[0].surface
.bpe
== 1)
6407 max_uncompressed_block_size
= V_028C78_MAX_BLOCK_SIZE_64B
;
6408 else if (iview
->image
->planes
[0].surface
.bpe
== 2)
6409 max_uncompressed_block_size
= V_028C78_MAX_BLOCK_SIZE_128B
;
6412 if (iview
->image
->usage
& (VK_IMAGE_USAGE_SAMPLED_BIT
|
6413 VK_IMAGE_USAGE_TRANSFER_SRC_BIT
|
6414 VK_IMAGE_USAGE_INPUT_ATTACHMENT_BIT
)) {
6415 /* If this DCC image is potentially going to be used in texture
6416 * fetches, we need some special settings.
6418 independent_64b_blocks
= 1;
6419 max_compressed_block_size
= V_028C78_MAX_BLOCK_SIZE_64B
;
6421 /* MAX_UNCOMPRESSED_BLOCK_SIZE must be >=
6422 * MAX_COMPRESSED_BLOCK_SIZE. Set MAX_COMPRESSED_BLOCK_SIZE as
6423 * big as possible for better compression state.
6425 independent_64b_blocks
= 0;
6426 max_compressed_block_size
= max_uncompressed_block_size
;
6430 return S_028C78_MAX_UNCOMPRESSED_BLOCK_SIZE(max_uncompressed_block_size
) |
6431 S_028C78_MAX_COMPRESSED_BLOCK_SIZE(max_compressed_block_size
) |
6432 S_028C78_MIN_COMPRESSED_BLOCK_SIZE(min_compressed_block_size
) |
6433 S_028C78_INDEPENDENT_64B_BLOCKS(independent_64b_blocks
) |
6434 S_028C78_INDEPENDENT_128B_BLOCKS(independent_128b_blocks
);
6438 radv_initialise_color_surface(struct radv_device
*device
,
6439 struct radv_color_buffer_info
*cb
,
6440 struct radv_image_view
*iview
)
6442 const struct vk_format_description
*desc
;
6443 unsigned ntype
, format
, swap
, endian
;
6444 unsigned blend_clamp
= 0, blend_bypass
= 0;
6446 const struct radv_image_plane
*plane
= &iview
->image
->planes
[iview
->plane_id
];
6447 const struct radeon_surf
*surf
= &plane
->surface
;
6449 desc
= vk_format_description(iview
->vk_format
);
6451 memset(cb
, 0, sizeof(*cb
));
6453 /* Intensity is implemented as Red, so treat it that way. */
6454 cb
->cb_color_attrib
= S_028C74_FORCE_DST_ALPHA_1(desc
->swizzle
[3] == VK_SWIZZLE_1
);
6456 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
+ plane
->offset
;
6458 cb
->cb_color_base
= va
>> 8;
6460 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
6461 if (device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
6462 cb
->cb_color_attrib3
|= S_028EE0_COLOR_SW_MODE(surf
->u
.gfx9
.surf
.swizzle_mode
) |
6463 S_028EE0_FMASK_SW_MODE(surf
->u
.gfx9
.fmask
.swizzle_mode
) |
6464 S_028EE0_CMASK_PIPE_ALIGNED(1) |
6465 S_028EE0_DCC_PIPE_ALIGNED(surf
->u
.gfx9
.dcc
.pipe_aligned
);
6467 struct gfx9_surf_meta_flags meta
= {
6472 if (surf
->dcc_offset
)
6473 meta
= surf
->u
.gfx9
.dcc
;
6475 cb
->cb_color_attrib
|= S_028C74_COLOR_SW_MODE(surf
->u
.gfx9
.surf
.swizzle_mode
) |
6476 S_028C74_FMASK_SW_MODE(surf
->u
.gfx9
.fmask
.swizzle_mode
) |
6477 S_028C74_RB_ALIGNED(meta
.rb_aligned
) |
6478 S_028C74_PIPE_ALIGNED(meta
.pipe_aligned
);
6479 cb
->cb_mrt_epitch
= S_0287A0_EPITCH(surf
->u
.gfx9
.surf
.epitch
);
6482 cb
->cb_color_base
+= surf
->u
.gfx9
.surf_offset
>> 8;
6483 cb
->cb_color_base
|= surf
->tile_swizzle
;
6485 const struct legacy_surf_level
*level_info
= &surf
->u
.legacy
.level
[iview
->base_mip
];
6486 unsigned pitch_tile_max
, slice_tile_max
, tile_mode_index
;
6488 cb
->cb_color_base
+= level_info
->offset
>> 8;
6489 if (level_info
->mode
== RADEON_SURF_MODE_2D
)
6490 cb
->cb_color_base
|= surf
->tile_swizzle
;
6492 pitch_tile_max
= level_info
->nblk_x
/ 8 - 1;
6493 slice_tile_max
= (level_info
->nblk_x
* level_info
->nblk_y
) / 64 - 1;
6494 tile_mode_index
= si_tile_mode_index(plane
, iview
->base_mip
, false);
6496 cb
->cb_color_pitch
= S_028C64_TILE_MAX(pitch_tile_max
);
6497 cb
->cb_color_slice
= S_028C68_TILE_MAX(slice_tile_max
);
6498 cb
->cb_color_cmask_slice
= surf
->u
.legacy
.cmask_slice_tile_max
;
6500 cb
->cb_color_attrib
|= S_028C74_TILE_MODE_INDEX(tile_mode_index
);
6502 if (radv_image_has_fmask(iview
->image
)) {
6503 if (device
->physical_device
->rad_info
.chip_class
>= GFX7
)
6504 cb
->cb_color_pitch
|= S_028C64_FMASK_TILE_MAX(surf
->u
.legacy
.fmask
.pitch_in_pixels
/ 8 - 1);
6505 cb
->cb_color_attrib
|= S_028C74_FMASK_TILE_MODE_INDEX(surf
->u
.legacy
.fmask
.tiling_index
);
6506 cb
->cb_color_fmask_slice
= S_028C88_TILE_MAX(surf
->u
.legacy
.fmask
.slice_tile_max
);
6508 /* This must be set for fast clear to work without FMASK. */
6509 if (device
->physical_device
->rad_info
.chip_class
>= GFX7
)
6510 cb
->cb_color_pitch
|= S_028C64_FMASK_TILE_MAX(pitch_tile_max
);
6511 cb
->cb_color_attrib
|= S_028C74_FMASK_TILE_MODE_INDEX(tile_mode_index
);
6512 cb
->cb_color_fmask_slice
= S_028C88_TILE_MAX(slice_tile_max
);
6516 /* CMASK variables */
6517 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
;
6518 va
+= surf
->cmask_offset
;
6519 cb
->cb_color_cmask
= va
>> 8;
6521 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
;
6522 va
+= surf
->dcc_offset
;
6524 if (radv_dcc_enabled(iview
->image
, iview
->base_mip
) &&
6525 device
->physical_device
->rad_info
.chip_class
<= GFX8
)
6526 va
+= plane
->surface
.u
.legacy
.level
[iview
->base_mip
].dcc_offset
;
6528 unsigned dcc_tile_swizzle
= surf
->tile_swizzle
;
6529 dcc_tile_swizzle
&= (surf
->dcc_alignment
- 1) >> 8;
6531 cb
->cb_dcc_base
= va
>> 8;
6532 cb
->cb_dcc_base
|= dcc_tile_swizzle
;
6534 /* GFX10 field has the same base shift as the GFX6 field. */
6535 uint32_t max_slice
= radv_surface_max_layer_count(iview
) - 1;
6536 cb
->cb_color_view
= S_028C6C_SLICE_START(iview
->base_layer
) |
6537 S_028C6C_SLICE_MAX_GFX10(max_slice
);
6539 if (iview
->image
->info
.samples
> 1) {
6540 unsigned log_samples
= util_logbase2(iview
->image
->info
.samples
);
6542 cb
->cb_color_attrib
|= S_028C74_NUM_SAMPLES(log_samples
) |
6543 S_028C74_NUM_FRAGMENTS(log_samples
);
6546 if (radv_image_has_fmask(iview
->image
)) {
6547 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
+ surf
->fmask_offset
;
6548 cb
->cb_color_fmask
= va
>> 8;
6549 cb
->cb_color_fmask
|= surf
->fmask_tile_swizzle
;
6551 cb
->cb_color_fmask
= cb
->cb_color_base
;
6554 ntype
= radv_translate_color_numformat(iview
->vk_format
,
6556 vk_format_get_first_non_void_channel(iview
->vk_format
));
6557 format
= radv_translate_colorformat(iview
->vk_format
);
6558 if (format
== V_028C70_COLOR_INVALID
|| ntype
== ~0u)
6559 radv_finishme("Illegal color\n");
6560 swap
= radv_translate_colorswap(iview
->vk_format
, false);
6561 endian
= radv_colorformat_endian_swap(format
);
6563 /* blend clamp should be set for all NORM/SRGB types */
6564 if (ntype
== V_028C70_NUMBER_UNORM
||
6565 ntype
== V_028C70_NUMBER_SNORM
||
6566 ntype
== V_028C70_NUMBER_SRGB
)
6569 /* set blend bypass according to docs if SINT/UINT or
6570 8/24 COLOR variants */
6571 if (ntype
== V_028C70_NUMBER_UINT
|| ntype
== V_028C70_NUMBER_SINT
||
6572 format
== V_028C70_COLOR_8_24
|| format
== V_028C70_COLOR_24_8
||
6573 format
== V_028C70_COLOR_X24_8_32_FLOAT
) {
6578 if ((ntype
== V_028C70_NUMBER_UINT
|| ntype
== V_028C70_NUMBER_SINT
) &&
6579 (format
== V_028C70_COLOR_8
||
6580 format
== V_028C70_COLOR_8_8
||
6581 format
== V_028C70_COLOR_8_8_8_8
))
6582 ->color_is_int8
= true;
6584 cb
->cb_color_info
= S_028C70_FORMAT(format
) |
6585 S_028C70_COMP_SWAP(swap
) |
6586 S_028C70_BLEND_CLAMP(blend_clamp
) |
6587 S_028C70_BLEND_BYPASS(blend_bypass
) |
6588 S_028C70_SIMPLE_FLOAT(1) |
6589 S_028C70_ROUND_MODE(ntype
!= V_028C70_NUMBER_UNORM
&&
6590 ntype
!= V_028C70_NUMBER_SNORM
&&
6591 ntype
!= V_028C70_NUMBER_SRGB
&&
6592 format
!= V_028C70_COLOR_8_24
&&
6593 format
!= V_028C70_COLOR_24_8
) |
6594 S_028C70_NUMBER_TYPE(ntype
) |
6595 S_028C70_ENDIAN(endian
);
6596 if (radv_image_has_fmask(iview
->image
)) {
6597 cb
->cb_color_info
|= S_028C70_COMPRESSION(1);
6598 if (device
->physical_device
->rad_info
.chip_class
== GFX6
) {
6599 unsigned fmask_bankh
= util_logbase2(surf
->u
.legacy
.fmask
.bankh
);
6600 cb
->cb_color_attrib
|= S_028C74_FMASK_BANK_HEIGHT(fmask_bankh
);
6603 if (radv_image_is_tc_compat_cmask(iview
->image
)) {
6604 /* Allow the texture block to read FMASK directly
6605 * without decompressing it. This bit must be cleared
6606 * when performing FMASK_DECOMPRESS or DCC_COMPRESS,
6607 * otherwise the operation doesn't happen.
6609 cb
->cb_color_info
|= S_028C70_FMASK_COMPRESS_1FRAG_ONLY(1);
6611 /* Set CMASK into a tiling format that allows the
6612 * texture block to read it.
6614 cb
->cb_color_info
|= S_028C70_CMASK_ADDR_TYPE(2);
6618 if (radv_image_has_cmask(iview
->image
) &&
6619 !(device
->instance
->debug_flags
& RADV_DEBUG_NO_FAST_CLEARS
))
6620 cb
->cb_color_info
|= S_028C70_FAST_CLEAR(1);
6622 if (radv_dcc_enabled(iview
->image
, iview
->base_mip
))
6623 cb
->cb_color_info
|= S_028C70_DCC_ENABLE(1);
6625 cb
->cb_dcc_control
= radv_init_dcc_control_reg(device
, iview
);
6627 /* This must be set for fast clear to work without FMASK. */
6628 if (!radv_image_has_fmask(iview
->image
) &&
6629 device
->physical_device
->rad_info
.chip_class
== GFX6
) {
6630 unsigned bankh
= util_logbase2(surf
->u
.legacy
.bankh
);
6631 cb
->cb_color_attrib
|= S_028C74_FMASK_BANK_HEIGHT(bankh
);
6634 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
6635 const struct vk_format_description
*format_desc
= vk_format_description(iview
->image
->vk_format
);
6637 unsigned mip0_depth
= iview
->image
->type
== VK_IMAGE_TYPE_3D
?
6638 (iview
->extent
.depth
- 1) : (iview
->image
->info
.array_size
- 1);
6639 unsigned width
= iview
->extent
.width
/ (iview
->plane_id
? format_desc
->width_divisor
: 1);
6640 unsigned height
= iview
->extent
.height
/ (iview
->plane_id
? format_desc
->height_divisor
: 1);
6642 if (device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
6643 cb
->cb_color_view
|= S_028C6C_MIP_LEVEL_GFX10(iview
->base_mip
);
6645 cb
->cb_color_attrib3
|= S_028EE0_MIP0_DEPTH(mip0_depth
) |
6646 S_028EE0_RESOURCE_TYPE(surf
->u
.gfx9
.resource_type
) |
6647 S_028EE0_RESOURCE_LEVEL(1);
6649 cb
->cb_color_view
|= S_028C6C_MIP_LEVEL_GFX9(iview
->base_mip
);
6650 cb
->cb_color_attrib
|= S_028C74_MIP0_DEPTH(mip0_depth
) |
6651 S_028C74_RESOURCE_TYPE(surf
->u
.gfx9
.resource_type
);
6654 cb
->cb_color_attrib2
= S_028C68_MIP0_WIDTH(width
- 1) |
6655 S_028C68_MIP0_HEIGHT(height
- 1) |
6656 S_028C68_MAX_MIP(iview
->image
->info
.levels
- 1);
6661 radv_calc_decompress_on_z_planes(struct radv_device
*device
,
6662 struct radv_image_view
*iview
)
6664 unsigned max_zplanes
= 0;
6666 assert(radv_image_is_tc_compat_htile(iview
->image
));
6668 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
6669 /* Default value for 32-bit depth surfaces. */
6672 if (iview
->vk_format
== VK_FORMAT_D16_UNORM
&&
6673 iview
->image
->info
.samples
> 1)
6676 max_zplanes
= max_zplanes
+ 1;
6678 if (iview
->vk_format
== VK_FORMAT_D16_UNORM
) {
6679 /* Do not enable Z plane compression for 16-bit depth
6680 * surfaces because isn't supported on GFX8. Only
6681 * 32-bit depth surfaces are supported by the hardware.
6682 * This allows to maintain shader compatibility and to
6683 * reduce the number of depth decompressions.
6687 if (iview
->image
->info
.samples
<= 1)
6689 else if (iview
->image
->info
.samples
<= 4)
6700 radv_initialise_ds_surface(struct radv_device
*device
,
6701 struct radv_ds_buffer_info
*ds
,
6702 struct radv_image_view
*iview
)
6704 unsigned level
= iview
->base_mip
;
6705 unsigned format
, stencil_format
;
6706 uint64_t va
, s_offs
, z_offs
;
6707 bool stencil_only
= false;
6708 const struct radv_image_plane
*plane
= &iview
->image
->planes
[0];
6709 const struct radeon_surf
*surf
= &plane
->surface
;
6711 assert(vk_format_get_plane_count(iview
->image
->vk_format
) == 1);
6713 memset(ds
, 0, sizeof(*ds
));
6714 switch (iview
->image
->vk_format
) {
6715 case VK_FORMAT_D24_UNORM_S8_UINT
:
6716 case VK_FORMAT_X8_D24_UNORM_PACK32
:
6717 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-24);
6718 ds
->offset_scale
= 2.0f
;
6720 case VK_FORMAT_D16_UNORM
:
6721 case VK_FORMAT_D16_UNORM_S8_UINT
:
6722 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-16);
6723 ds
->offset_scale
= 4.0f
;
6725 case VK_FORMAT_D32_SFLOAT
:
6726 case VK_FORMAT_D32_SFLOAT_S8_UINT
:
6727 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-23) |
6728 S_028B78_POLY_OFFSET_DB_IS_FLOAT_FMT(1);
6729 ds
->offset_scale
= 1.0f
;
6731 case VK_FORMAT_S8_UINT
:
6732 stencil_only
= true;
6738 format
= radv_translate_dbformat(iview
->image
->vk_format
);
6739 stencil_format
= surf
->has_stencil
?
6740 V_028044_STENCIL_8
: V_028044_STENCIL_INVALID
;
6742 uint32_t max_slice
= radv_surface_max_layer_count(iview
) - 1;
6743 ds
->db_depth_view
= S_028008_SLICE_START(iview
->base_layer
) |
6744 S_028008_SLICE_MAX(max_slice
);
6745 if (device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
6746 ds
->db_depth_view
|= S_028008_SLICE_START_HI(iview
->base_layer
>> 11) |
6747 S_028008_SLICE_MAX_HI(max_slice
>> 11);
6750 ds
->db_htile_data_base
= 0;
6751 ds
->db_htile_surface
= 0;
6753 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
;
6754 s_offs
= z_offs
= va
;
6756 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
6757 assert(surf
->u
.gfx9
.surf_offset
== 0);
6758 s_offs
+= surf
->u
.gfx9
.stencil_offset
;
6760 ds
->db_z_info
= S_028038_FORMAT(format
) |
6761 S_028038_NUM_SAMPLES(util_logbase2(iview
->image
->info
.samples
)) |
6762 S_028038_SW_MODE(surf
->u
.gfx9
.surf
.swizzle_mode
) |
6763 S_028038_MAXMIP(iview
->image
->info
.levels
- 1) |
6764 S_028038_ZRANGE_PRECISION(1);
6765 ds
->db_stencil_info
= S_02803C_FORMAT(stencil_format
) |
6766 S_02803C_SW_MODE(surf
->u
.gfx9
.stencil
.swizzle_mode
);
6768 if (device
->physical_device
->rad_info
.chip_class
== GFX9
) {
6769 ds
->db_z_info2
= S_028068_EPITCH(surf
->u
.gfx9
.surf
.epitch
);
6770 ds
->db_stencil_info2
= S_02806C_EPITCH(surf
->u
.gfx9
.stencil
.epitch
);
6773 ds
->db_depth_view
|= S_028008_MIPID(level
);
6774 ds
->db_depth_size
= S_02801C_X_MAX(iview
->image
->info
.width
- 1) |
6775 S_02801C_Y_MAX(iview
->image
->info
.height
- 1);
6777 if (radv_htile_enabled(iview
->image
, level
)) {
6778 ds
->db_z_info
|= S_028038_TILE_SURFACE_ENABLE(1);
6780 if (radv_image_is_tc_compat_htile(iview
->image
)) {
6781 unsigned max_zplanes
=
6782 radv_calc_decompress_on_z_planes(device
, iview
);
6784 ds
->db_z_info
|= S_028038_DECOMPRESS_ON_N_ZPLANES(max_zplanes
);
6786 if (device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
6787 ds
->db_z_info
|= S_028040_ITERATE_FLUSH(1);
6788 ds
->db_stencil_info
|= S_028044_ITERATE_FLUSH(1);
6790 ds
->db_z_info
|= S_028038_ITERATE_FLUSH(1);
6791 ds
->db_stencil_info
|= S_02803C_ITERATE_FLUSH(1);
6795 if (!surf
->has_stencil
)
6796 /* Use all of the htile_buffer for depth if there's no stencil. */
6797 ds
->db_stencil_info
|= S_02803C_TILE_STENCIL_DISABLE(1);
6798 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
+
6800 ds
->db_htile_data_base
= va
>> 8;
6801 ds
->db_htile_surface
= S_028ABC_FULL_CACHE(1) |
6802 S_028ABC_PIPE_ALIGNED(1);
6804 if (device
->physical_device
->rad_info
.chip_class
== GFX9
) {
6805 ds
->db_htile_surface
|= S_028ABC_RB_ALIGNED(1);
6809 const struct legacy_surf_level
*level_info
= &surf
->u
.legacy
.level
[level
];
6812 level_info
= &surf
->u
.legacy
.stencil_level
[level
];
6814 z_offs
+= surf
->u
.legacy
.level
[level
].offset
;
6815 s_offs
+= surf
->u
.legacy
.stencil_level
[level
].offset
;
6817 ds
->db_depth_info
= S_02803C_ADDR5_SWIZZLE_MASK(!radv_image_is_tc_compat_htile(iview
->image
));
6818 ds
->db_z_info
= S_028040_FORMAT(format
) | S_028040_ZRANGE_PRECISION(1);
6819 ds
->db_stencil_info
= S_028044_FORMAT(stencil_format
);
6821 if (iview
->image
->info
.samples
> 1)
6822 ds
->db_z_info
|= S_028040_NUM_SAMPLES(util_logbase2(iview
->image
->info
.samples
));
6824 if (device
->physical_device
->rad_info
.chip_class
>= GFX7
) {
6825 struct radeon_info
*info
= &device
->physical_device
->rad_info
;
6826 unsigned tiling_index
= surf
->u
.legacy
.tiling_index
[level
];
6827 unsigned stencil_index
= surf
->u
.legacy
.stencil_tiling_index
[level
];
6828 unsigned macro_index
= surf
->u
.legacy
.macro_tile_index
;
6829 unsigned tile_mode
= info
->si_tile_mode_array
[tiling_index
];
6830 unsigned stencil_tile_mode
= info
->si_tile_mode_array
[stencil_index
];
6831 unsigned macro_mode
= info
->cik_macrotile_mode_array
[macro_index
];
6834 tile_mode
= stencil_tile_mode
;
6836 ds
->db_depth_info
|=
6837 S_02803C_ARRAY_MODE(G_009910_ARRAY_MODE(tile_mode
)) |
6838 S_02803C_PIPE_CONFIG(G_009910_PIPE_CONFIG(tile_mode
)) |
6839 S_02803C_BANK_WIDTH(G_009990_BANK_WIDTH(macro_mode
)) |
6840 S_02803C_BANK_HEIGHT(G_009990_BANK_HEIGHT(macro_mode
)) |
6841 S_02803C_MACRO_TILE_ASPECT(G_009990_MACRO_TILE_ASPECT(macro_mode
)) |
6842 S_02803C_NUM_BANKS(G_009990_NUM_BANKS(macro_mode
));
6843 ds
->db_z_info
|= S_028040_TILE_SPLIT(G_009910_TILE_SPLIT(tile_mode
));
6844 ds
->db_stencil_info
|= S_028044_TILE_SPLIT(G_009910_TILE_SPLIT(stencil_tile_mode
));
6846 unsigned tile_mode_index
= si_tile_mode_index(&iview
->image
->planes
[0], level
, false);
6847 ds
->db_z_info
|= S_028040_TILE_MODE_INDEX(tile_mode_index
);
6848 tile_mode_index
= si_tile_mode_index(&iview
->image
->planes
[0], level
, true);
6849 ds
->db_stencil_info
|= S_028044_TILE_MODE_INDEX(tile_mode_index
);
6851 ds
->db_z_info
|= S_028040_TILE_MODE_INDEX(tile_mode_index
);
6854 ds
->db_depth_size
= S_028058_PITCH_TILE_MAX((level_info
->nblk_x
/ 8) - 1) |
6855 S_028058_HEIGHT_TILE_MAX((level_info
->nblk_y
/ 8) - 1);
6856 ds
->db_depth_slice
= S_02805C_SLICE_TILE_MAX((level_info
->nblk_x
* level_info
->nblk_y
) / 64 - 1);
6858 if (radv_htile_enabled(iview
->image
, level
)) {
6859 ds
->db_z_info
|= S_028040_TILE_SURFACE_ENABLE(1);
6861 if (!surf
->has_stencil
&&
6862 !radv_image_is_tc_compat_htile(iview
->image
))
6863 /* Use all of the htile_buffer for depth if there's no stencil. */
6864 ds
->db_stencil_info
|= S_028044_TILE_STENCIL_DISABLE(1);
6866 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
+
6868 ds
->db_htile_data_base
= va
>> 8;
6869 ds
->db_htile_surface
= S_028ABC_FULL_CACHE(1);
6871 if (radv_image_is_tc_compat_htile(iview
->image
)) {
6872 unsigned max_zplanes
=
6873 radv_calc_decompress_on_z_planes(device
, iview
);
6875 ds
->db_htile_surface
|= S_028ABC_TC_COMPATIBLE(1);
6876 ds
->db_z_info
|= S_028040_DECOMPRESS_ON_N_ZPLANES(max_zplanes
);
6881 ds
->db_z_read_base
= ds
->db_z_write_base
= z_offs
>> 8;
6882 ds
->db_stencil_read_base
= ds
->db_stencil_write_base
= s_offs
>> 8;
6885 VkResult
radv_CreateFramebuffer(
6887 const VkFramebufferCreateInfo
* pCreateInfo
,
6888 const VkAllocationCallbacks
* pAllocator
,
6889 VkFramebuffer
* pFramebuffer
)
6891 RADV_FROM_HANDLE(radv_device
, device
, _device
);
6892 struct radv_framebuffer
*framebuffer
;
6893 const VkFramebufferAttachmentsCreateInfo
*imageless_create_info
=
6894 vk_find_struct_const(pCreateInfo
->pNext
,
6895 FRAMEBUFFER_ATTACHMENTS_CREATE_INFO
);
6897 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
6899 size_t size
= sizeof(*framebuffer
);
6900 if (!imageless_create_info
)
6901 size
+= sizeof(struct radv_image_view
*) * pCreateInfo
->attachmentCount
;
6902 framebuffer
= vk_alloc2(&device
->vk
.alloc
, pAllocator
, size
, 8,
6903 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
6904 if (framebuffer
== NULL
)
6905 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
6907 vk_object_base_init(&device
->vk
, &framebuffer
->base
,
6908 VK_OBJECT_TYPE_FRAMEBUFFER
);
6910 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
6911 framebuffer
->width
= pCreateInfo
->width
;
6912 framebuffer
->height
= pCreateInfo
->height
;
6913 framebuffer
->layers
= pCreateInfo
->layers
;
6914 if (imageless_create_info
) {
6915 for (unsigned i
= 0; i
< imageless_create_info
->attachmentImageInfoCount
; ++i
) {
6916 const VkFramebufferAttachmentImageInfo
*attachment
=
6917 imageless_create_info
->pAttachmentImageInfos
+ i
;
6918 framebuffer
->width
= MIN2(framebuffer
->width
, attachment
->width
);
6919 framebuffer
->height
= MIN2(framebuffer
->height
, attachment
->height
);
6920 framebuffer
->layers
= MIN2(framebuffer
->layers
, attachment
->layerCount
);
6923 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
6924 VkImageView _iview
= pCreateInfo
->pAttachments
[i
];
6925 struct radv_image_view
*iview
= radv_image_view_from_handle(_iview
);
6926 framebuffer
->attachments
[i
] = iview
;
6927 framebuffer
->width
= MIN2(framebuffer
->width
, iview
->extent
.width
);
6928 framebuffer
->height
= MIN2(framebuffer
->height
, iview
->extent
.height
);
6929 framebuffer
->layers
= MIN2(framebuffer
->layers
, radv_surface_max_layer_count(iview
));
6933 *pFramebuffer
= radv_framebuffer_to_handle(framebuffer
);
6937 void radv_DestroyFramebuffer(
6940 const VkAllocationCallbacks
* pAllocator
)
6942 RADV_FROM_HANDLE(radv_device
, device
, _device
);
6943 RADV_FROM_HANDLE(radv_framebuffer
, fb
, _fb
);
6947 vk_object_base_finish(&fb
->base
);
6948 vk_free2(&device
->vk
.alloc
, pAllocator
, fb
);
6951 static unsigned radv_tex_wrap(VkSamplerAddressMode address_mode
)
6953 switch (address_mode
) {
6954 case VK_SAMPLER_ADDRESS_MODE_REPEAT
:
6955 return V_008F30_SQ_TEX_WRAP
;
6956 case VK_SAMPLER_ADDRESS_MODE_MIRRORED_REPEAT
:
6957 return V_008F30_SQ_TEX_MIRROR
;
6958 case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE
:
6959 return V_008F30_SQ_TEX_CLAMP_LAST_TEXEL
;
6960 case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_BORDER
:
6961 return V_008F30_SQ_TEX_CLAMP_BORDER
;
6962 case VK_SAMPLER_ADDRESS_MODE_MIRROR_CLAMP_TO_EDGE
:
6963 return V_008F30_SQ_TEX_MIRROR_ONCE_LAST_TEXEL
;
6965 unreachable("illegal tex wrap mode");
6971 radv_tex_compare(VkCompareOp op
)
6974 case VK_COMPARE_OP_NEVER
:
6975 return V_008F30_SQ_TEX_DEPTH_COMPARE_NEVER
;
6976 case VK_COMPARE_OP_LESS
:
6977 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESS
;
6978 case VK_COMPARE_OP_EQUAL
:
6979 return V_008F30_SQ_TEX_DEPTH_COMPARE_EQUAL
;
6980 case VK_COMPARE_OP_LESS_OR_EQUAL
:
6981 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESSEQUAL
;
6982 case VK_COMPARE_OP_GREATER
:
6983 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATER
;
6984 case VK_COMPARE_OP_NOT_EQUAL
:
6985 return V_008F30_SQ_TEX_DEPTH_COMPARE_NOTEQUAL
;
6986 case VK_COMPARE_OP_GREATER_OR_EQUAL
:
6987 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATEREQUAL
;
6988 case VK_COMPARE_OP_ALWAYS
:
6989 return V_008F30_SQ_TEX_DEPTH_COMPARE_ALWAYS
;
6991 unreachable("illegal compare mode");
6997 radv_tex_filter(VkFilter filter
, unsigned max_ansio
)
7000 case VK_FILTER_NEAREST
:
7001 return (max_ansio
> 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_POINT
:
7002 V_008F38_SQ_TEX_XY_FILTER_POINT
);
7003 case VK_FILTER_LINEAR
:
7004 return (max_ansio
> 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_BILINEAR
:
7005 V_008F38_SQ_TEX_XY_FILTER_BILINEAR
);
7006 case VK_FILTER_CUBIC_IMG
:
7008 fprintf(stderr
, "illegal texture filter");
7014 radv_tex_mipfilter(VkSamplerMipmapMode mode
)
7017 case VK_SAMPLER_MIPMAP_MODE_NEAREST
:
7018 return V_008F38_SQ_TEX_Z_FILTER_POINT
;
7019 case VK_SAMPLER_MIPMAP_MODE_LINEAR
:
7020 return V_008F38_SQ_TEX_Z_FILTER_LINEAR
;
7022 return V_008F38_SQ_TEX_Z_FILTER_NONE
;
7027 radv_tex_bordercolor(VkBorderColor bcolor
)
7030 case VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
:
7031 case VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
:
7032 return V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK
;
7033 case VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
:
7034 case VK_BORDER_COLOR_INT_OPAQUE_BLACK
:
7035 return V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_BLACK
;
7036 case VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
:
7037 case VK_BORDER_COLOR_INT_OPAQUE_WHITE
:
7038 return V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_WHITE
;
7039 case VK_BORDER_COLOR_FLOAT_CUSTOM_EXT
:
7040 case VK_BORDER_COLOR_INT_CUSTOM_EXT
:
7041 return V_008F3C_SQ_TEX_BORDER_COLOR_REGISTER
;
7049 radv_tex_aniso_filter(unsigned filter
)
7063 radv_tex_filter_mode(VkSamplerReductionMode mode
)
7066 case VK_SAMPLER_REDUCTION_MODE_WEIGHTED_AVERAGE_EXT
:
7067 return V_008F30_SQ_IMG_FILTER_MODE_BLEND
;
7068 case VK_SAMPLER_REDUCTION_MODE_MIN_EXT
:
7069 return V_008F30_SQ_IMG_FILTER_MODE_MIN
;
7070 case VK_SAMPLER_REDUCTION_MODE_MAX_EXT
:
7071 return V_008F30_SQ_IMG_FILTER_MODE_MAX
;
7079 radv_get_max_anisotropy(struct radv_device
*device
,
7080 const VkSamplerCreateInfo
*pCreateInfo
)
7082 if (device
->force_aniso
>= 0)
7083 return device
->force_aniso
;
7085 if (pCreateInfo
->anisotropyEnable
&&
7086 pCreateInfo
->maxAnisotropy
> 1.0f
)
7087 return (uint32_t)pCreateInfo
->maxAnisotropy
;
7092 static inline int S_FIXED(float value
, unsigned frac_bits
)
7094 return value
* (1 << frac_bits
);
7097 static uint32_t radv_register_border_color(struct radv_device
*device
,
7098 VkClearColorValue value
)
7102 pthread_mutex_lock(&device
->border_color_data
.mutex
);
7104 for (slot
= 0; slot
< RADV_BORDER_COLOR_COUNT
; slot
++) {
7105 if (!device
->border_color_data
.used
[slot
]) {
7106 /* Copy to the GPU wrt endian-ness. */
7107 util_memcpy_cpu_to_le32(&device
->border_color_data
.colors_gpu_ptr
[slot
],
7109 sizeof(VkClearColorValue
));
7111 device
->border_color_data
.used
[slot
] = true;
7116 pthread_mutex_unlock(&device
->border_color_data
.mutex
);
7121 static void radv_unregister_border_color(struct radv_device
*device
,
7124 pthread_mutex_lock(&device
->border_color_data
.mutex
);
7126 device
->border_color_data
.used
[slot
] = false;
7128 pthread_mutex_unlock(&device
->border_color_data
.mutex
);
7132 radv_init_sampler(struct radv_device
*device
,
7133 struct radv_sampler
*sampler
,
7134 const VkSamplerCreateInfo
*pCreateInfo
)
7136 uint32_t max_aniso
= radv_get_max_anisotropy(device
, pCreateInfo
);
7137 uint32_t max_aniso_ratio
= radv_tex_aniso_filter(max_aniso
);
7138 bool compat_mode
= device
->physical_device
->rad_info
.chip_class
== GFX8
||
7139 device
->physical_device
->rad_info
.chip_class
== GFX9
;
7140 unsigned filter_mode
= V_008F30_SQ_IMG_FILTER_MODE_BLEND
;
7141 unsigned depth_compare_func
= V_008F30_SQ_TEX_DEPTH_COMPARE_NEVER
;
7142 bool trunc_coord
= pCreateInfo
->minFilter
== VK_FILTER_NEAREST
&& pCreateInfo
->magFilter
== VK_FILTER_NEAREST
;
7143 bool uses_border_color
= pCreateInfo
->addressModeU
== VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_BORDER
||
7144 pCreateInfo
->addressModeV
== VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_BORDER
||
7145 pCreateInfo
->addressModeW
== VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_BORDER
;
7146 VkBorderColor border_color
= uses_border_color
? pCreateInfo
->borderColor
: VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
;
7147 uint32_t border_color_ptr
;
7149 const struct VkSamplerReductionModeCreateInfo
*sampler_reduction
=
7150 vk_find_struct_const(pCreateInfo
->pNext
,
7151 SAMPLER_REDUCTION_MODE_CREATE_INFO
);
7152 if (sampler_reduction
)
7153 filter_mode
= radv_tex_filter_mode(sampler_reduction
->reductionMode
);
7155 if (pCreateInfo
->compareEnable
)
7156 depth_compare_func
= radv_tex_compare(pCreateInfo
->compareOp
);
7158 sampler
->border_color_slot
= RADV_BORDER_COLOR_COUNT
;
7160 if (border_color
== VK_BORDER_COLOR_FLOAT_CUSTOM_EXT
|| border_color
== VK_BORDER_COLOR_INT_CUSTOM_EXT
) {
7161 const VkSamplerCustomBorderColorCreateInfoEXT
*custom_border_color
=
7162 vk_find_struct_const(pCreateInfo
->pNext
,
7163 SAMPLER_CUSTOM_BORDER_COLOR_CREATE_INFO_EXT
);
7165 assert(custom_border_color
);
7167 sampler
->border_color_slot
=
7168 radv_register_border_color(device
, custom_border_color
->customBorderColor
);
7170 /* Did we fail to find a slot? */
7171 if (sampler
->border_color_slot
== RADV_BORDER_COLOR_COUNT
) {
7172 fprintf(stderr
, "WARNING: no free border color slots, defaulting to TRANS_BLACK.\n");
7173 border_color
= VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
;
7177 /* If we don't have a custom color, set the ptr to 0 */
7178 border_color_ptr
= sampler
->border_color_slot
!= RADV_BORDER_COLOR_COUNT
7179 ? sampler
->border_color_slot
7182 sampler
->state
[0] = (S_008F30_CLAMP_X(radv_tex_wrap(pCreateInfo
->addressModeU
)) |
7183 S_008F30_CLAMP_Y(radv_tex_wrap(pCreateInfo
->addressModeV
)) |
7184 S_008F30_CLAMP_Z(radv_tex_wrap(pCreateInfo
->addressModeW
)) |
7185 S_008F30_MAX_ANISO_RATIO(max_aniso_ratio
) |
7186 S_008F30_DEPTH_COMPARE_FUNC(depth_compare_func
) |
7187 S_008F30_FORCE_UNNORMALIZED(pCreateInfo
->unnormalizedCoordinates
? 1 : 0) |
7188 S_008F30_ANISO_THRESHOLD(max_aniso_ratio
>> 1) |
7189 S_008F30_ANISO_BIAS(max_aniso_ratio
) |
7190 S_008F30_DISABLE_CUBE_WRAP(0) |
7191 S_008F30_COMPAT_MODE(compat_mode
) |
7192 S_008F30_FILTER_MODE(filter_mode
) |
7193 S_008F30_TRUNC_COORD(trunc_coord
));
7194 sampler
->state
[1] = (S_008F34_MIN_LOD(S_FIXED(CLAMP(pCreateInfo
->minLod
, 0, 15), 8)) |
7195 S_008F34_MAX_LOD(S_FIXED(CLAMP(pCreateInfo
->maxLod
, 0, 15), 8)) |
7196 S_008F34_PERF_MIP(max_aniso_ratio
? max_aniso_ratio
+ 6 : 0));
7197 sampler
->state
[2] = (S_008F38_LOD_BIAS(S_FIXED(CLAMP(pCreateInfo
->mipLodBias
, -16, 16), 8)) |
7198 S_008F38_XY_MAG_FILTER(radv_tex_filter(pCreateInfo
->magFilter
, max_aniso
)) |
7199 S_008F38_XY_MIN_FILTER(radv_tex_filter(pCreateInfo
->minFilter
, max_aniso
)) |
7200 S_008F38_MIP_FILTER(radv_tex_mipfilter(pCreateInfo
->mipmapMode
)) |
7201 S_008F38_MIP_POINT_PRECLAMP(0));
7202 sampler
->state
[3] = (S_008F3C_BORDER_COLOR_PTR(border_color_ptr
) |
7203 S_008F3C_BORDER_COLOR_TYPE(radv_tex_bordercolor(border_color
)));
7205 if (device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
7206 sampler
->state
[2] |= S_008F38_ANISO_OVERRIDE_GFX10(1);
7208 sampler
->state
[2] |=
7209 S_008F38_DISABLE_LSB_CEIL(device
->physical_device
->rad_info
.chip_class
<= GFX8
) |
7210 S_008F38_FILTER_PREC_FIX(1) |
7211 S_008F38_ANISO_OVERRIDE_GFX6(device
->physical_device
->rad_info
.chip_class
>= GFX8
);
7215 VkResult
radv_CreateSampler(
7217 const VkSamplerCreateInfo
* pCreateInfo
,
7218 const VkAllocationCallbacks
* pAllocator
,
7219 VkSampler
* pSampler
)
7221 RADV_FROM_HANDLE(radv_device
, device
, _device
);
7222 struct radv_sampler
*sampler
;
7224 const struct VkSamplerYcbcrConversionInfo
*ycbcr_conversion
=
7225 vk_find_struct_const(pCreateInfo
->pNext
,
7226 SAMPLER_YCBCR_CONVERSION_INFO
);
7228 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO
);
7230 sampler
= vk_alloc2(&device
->vk
.alloc
, pAllocator
, sizeof(*sampler
), 8,
7231 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
7233 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
7235 vk_object_base_init(&device
->vk
, &sampler
->base
,
7236 VK_OBJECT_TYPE_SAMPLER
);
7238 radv_init_sampler(device
, sampler
, pCreateInfo
);
7240 sampler
->ycbcr_sampler
= ycbcr_conversion
? radv_sampler_ycbcr_conversion_from_handle(ycbcr_conversion
->conversion
): NULL
;
7241 *pSampler
= radv_sampler_to_handle(sampler
);
7246 void radv_DestroySampler(
7249 const VkAllocationCallbacks
* pAllocator
)
7251 RADV_FROM_HANDLE(radv_device
, device
, _device
);
7252 RADV_FROM_HANDLE(radv_sampler
, sampler
, _sampler
);
7257 if (sampler
->border_color_slot
!= RADV_BORDER_COLOR_COUNT
)
7258 radv_unregister_border_color(device
, sampler
->border_color_slot
);
7260 vk_object_base_finish(&sampler
->base
);
7261 vk_free2(&device
->vk
.alloc
, pAllocator
, sampler
);
7264 /* vk_icd.h does not declare this function, so we declare it here to
7265 * suppress Wmissing-prototypes.
7267 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
7268 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t *pSupportedVersion
);
7270 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
7271 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t *pSupportedVersion
)
7273 /* For the full details on loader interface versioning, see
7274 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
7275 * What follows is a condensed summary, to help you navigate the large and
7276 * confusing official doc.
7278 * - Loader interface v0 is incompatible with later versions. We don't
7281 * - In loader interface v1:
7282 * - The first ICD entrypoint called by the loader is
7283 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
7285 * - The ICD must statically expose no other Vulkan symbol unless it is
7286 * linked with -Bsymbolic.
7287 * - Each dispatchable Vulkan handle created by the ICD must be
7288 * a pointer to a struct whose first member is VK_LOADER_DATA. The
7289 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
7290 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
7291 * vkDestroySurfaceKHR(). The ICD must be capable of working with
7292 * such loader-managed surfaces.
7294 * - Loader interface v2 differs from v1 in:
7295 * - The first ICD entrypoint called by the loader is
7296 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
7297 * statically expose this entrypoint.
7299 * - Loader interface v3 differs from v2 in:
7300 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
7301 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
7302 * because the loader no longer does so.
7304 *pSupportedVersion
= MIN2(*pSupportedVersion
, 4u);
7308 VkResult
radv_GetMemoryFdKHR(VkDevice _device
,
7309 const VkMemoryGetFdInfoKHR
*pGetFdInfo
,
7312 RADV_FROM_HANDLE(radv_device
, device
, _device
);
7313 RADV_FROM_HANDLE(radv_device_memory
, memory
, pGetFdInfo
->memory
);
7315 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
7317 /* At the moment, we support only the below handle types. */
7318 assert(pGetFdInfo
->handleType
==
7319 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
7320 pGetFdInfo
->handleType
==
7321 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
7323 bool ret
= radv_get_memory_fd(device
, memory
, pFD
);
7325 return vk_error(device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
7329 static uint32_t radv_compute_valid_memory_types_attempt(struct radv_physical_device
*dev
,
7330 enum radeon_bo_domain domains
,
7331 enum radeon_bo_flag flags
,
7332 enum radeon_bo_flag ignore_flags
)
7334 /* Don't count GTT/CPU as relevant:
7336 * - We're not fully consistent between the two.
7337 * - Sometimes VRAM gets VRAM|GTT.
7339 const enum radeon_bo_domain relevant_domains
= RADEON_DOMAIN_VRAM
|
7343 for (unsigned i
= 0; i
< dev
->memory_properties
.memoryTypeCount
; ++i
) {
7344 if ((domains
& relevant_domains
) != (dev
->memory_domains
[i
] & relevant_domains
))
7347 if ((flags
& ~ignore_flags
) != (dev
->memory_flags
[i
] & ~ignore_flags
))
7356 static uint32_t radv_compute_valid_memory_types(struct radv_physical_device
*dev
,
7357 enum radeon_bo_domain domains
,
7358 enum radeon_bo_flag flags
)
7360 enum radeon_bo_flag ignore_flags
= ~(RADEON_FLAG_NO_CPU_ACCESS
| RADEON_FLAG_GTT_WC
);
7361 uint32_t bits
= radv_compute_valid_memory_types_attempt(dev
, domains
, flags
, ignore_flags
);
7364 ignore_flags
|= RADEON_FLAG_NO_CPU_ACCESS
;
7365 bits
= radv_compute_valid_memory_types_attempt(dev
, domains
, flags
, ignore_flags
);
7370 VkResult
radv_GetMemoryFdPropertiesKHR(VkDevice _device
,
7371 VkExternalMemoryHandleTypeFlagBits handleType
,
7373 VkMemoryFdPropertiesKHR
*pMemoryFdProperties
)
7375 RADV_FROM_HANDLE(radv_device
, device
, _device
);
7377 switch (handleType
) {
7378 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
: {
7379 enum radeon_bo_domain domains
;
7380 enum radeon_bo_flag flags
;
7381 if (!device
->ws
->buffer_get_flags_from_fd(device
->ws
, fd
, &domains
, &flags
))
7382 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE
);
7384 pMemoryFdProperties
->memoryTypeBits
= radv_compute_valid_memory_types(device
->physical_device
, domains
, flags
);
7388 /* The valid usage section for this function says:
7390 * "handleType must not be one of the handle types defined as
7393 * So opaque handle types fall into the default "unsupported" case.
7395 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE
);
7399 static VkResult
radv_import_opaque_fd(struct radv_device
*device
,
7403 uint32_t syncobj_handle
= 0;
7404 int ret
= device
->ws
->import_syncobj(device
->ws
, fd
, &syncobj_handle
);
7406 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE
);
7409 device
->ws
->destroy_syncobj(device
->ws
, *syncobj
);
7411 *syncobj
= syncobj_handle
;
7417 static VkResult
radv_import_sync_fd(struct radv_device
*device
,
7421 /* If we create a syncobj we do it locally so that if we have an error, we don't
7422 * leave a syncobj in an undetermined state in the fence. */
7423 uint32_t syncobj_handle
= *syncobj
;
7424 if (!syncobj_handle
) {
7425 bool create_signaled
= fd
== -1 ? true : false;
7427 int ret
= device
->ws
->create_syncobj(device
->ws
, create_signaled
,
7430 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
7434 device
->ws
->signal_syncobj(device
->ws
, syncobj_handle
, 0);
7438 int ret
= device
->ws
->import_syncobj_from_sync_file(device
->ws
, syncobj_handle
, fd
);
7440 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE
);
7444 *syncobj
= syncobj_handle
;
7449 VkResult
radv_ImportSemaphoreFdKHR(VkDevice _device
,
7450 const VkImportSemaphoreFdInfoKHR
*pImportSemaphoreFdInfo
)
7452 RADV_FROM_HANDLE(radv_device
, device
, _device
);
7453 RADV_FROM_HANDLE(radv_semaphore
, sem
, pImportSemaphoreFdInfo
->semaphore
);
7455 struct radv_semaphore_part
*dst
= NULL
;
7456 bool timeline
= sem
->permanent
.kind
== RADV_SEMAPHORE_TIMELINE_SYNCOBJ
;
7458 if (pImportSemaphoreFdInfo
->flags
& VK_SEMAPHORE_IMPORT_TEMPORARY_BIT
) {
7460 dst
= &sem
->temporary
;
7462 dst
= &sem
->permanent
;
7465 uint32_t syncobj
= (dst
->kind
== RADV_SEMAPHORE_SYNCOBJ
||
7466 dst
->kind
== RADV_SEMAPHORE_TIMELINE_SYNCOBJ
) ? dst
->syncobj
: 0;
7468 switch(pImportSemaphoreFdInfo
->handleType
) {
7469 case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT
:
7470 result
= radv_import_opaque_fd(device
, pImportSemaphoreFdInfo
->fd
, &syncobj
);
7472 case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT
:
7474 result
= radv_import_sync_fd(device
, pImportSemaphoreFdInfo
->fd
, &syncobj
);
7477 unreachable("Unhandled semaphore handle type");
7480 if (result
== VK_SUCCESS
) {
7481 dst
->syncobj
= syncobj
;
7482 dst
->kind
= RADV_SEMAPHORE_SYNCOBJ
;
7484 dst
->kind
= RADV_SEMAPHORE_TIMELINE_SYNCOBJ
;
7485 dst
->timeline_syncobj
.max_point
= 0;
7492 VkResult
radv_GetSemaphoreFdKHR(VkDevice _device
,
7493 const VkSemaphoreGetFdInfoKHR
*pGetFdInfo
,
7496 RADV_FROM_HANDLE(radv_device
, device
, _device
);
7497 RADV_FROM_HANDLE(radv_semaphore
, sem
, pGetFdInfo
->semaphore
);
7499 uint32_t syncobj_handle
;
7501 if (sem
->temporary
.kind
!= RADV_SEMAPHORE_NONE
) {
7502 assert(sem
->temporary
.kind
== RADV_SEMAPHORE_SYNCOBJ
||
7503 sem
->temporary
.kind
== RADV_SEMAPHORE_TIMELINE_SYNCOBJ
);
7504 syncobj_handle
= sem
->temporary
.syncobj
;
7506 assert(sem
->permanent
.kind
== RADV_SEMAPHORE_SYNCOBJ
||
7507 sem
->permanent
.kind
== RADV_SEMAPHORE_TIMELINE_SYNCOBJ
);
7508 syncobj_handle
= sem
->permanent
.syncobj
;
7511 switch(pGetFdInfo
->handleType
) {
7512 case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT
:
7513 ret
= device
->ws
->export_syncobj(device
->ws
, syncobj_handle
, pFd
);
7515 return vk_error(device
->instance
, VK_ERROR_TOO_MANY_OBJECTS
);
7517 case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT
:
7518 ret
= device
->ws
->export_syncobj_to_sync_file(device
->ws
, syncobj_handle
, pFd
);
7520 return vk_error(device
->instance
, VK_ERROR_TOO_MANY_OBJECTS
);
7522 if (sem
->temporary
.kind
!= RADV_SEMAPHORE_NONE
) {
7523 radv_destroy_semaphore_part(device
, &sem
->temporary
);
7525 device
->ws
->reset_syncobj(device
->ws
, syncobj_handle
);
7529 unreachable("Unhandled semaphore handle type");
7535 void radv_GetPhysicalDeviceExternalSemaphoreProperties(
7536 VkPhysicalDevice physicalDevice
,
7537 const VkPhysicalDeviceExternalSemaphoreInfo
*pExternalSemaphoreInfo
,
7538 VkExternalSemaphoreProperties
*pExternalSemaphoreProperties
)
7540 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
7541 VkSemaphoreTypeKHR type
= radv_get_semaphore_type(pExternalSemaphoreInfo
->pNext
, NULL
);
7543 if (type
== VK_SEMAPHORE_TYPE_TIMELINE
&& pdevice
->rad_info
.has_timeline_syncobj
&&
7544 pExternalSemaphoreInfo
->handleType
== VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT
) {
7545 pExternalSemaphoreProperties
->exportFromImportedHandleTypes
= VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT
;
7546 pExternalSemaphoreProperties
->compatibleHandleTypes
= VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT
;
7547 pExternalSemaphoreProperties
->externalSemaphoreFeatures
= VK_EXTERNAL_SEMAPHORE_FEATURE_EXPORTABLE_BIT
|
7548 VK_EXTERNAL_SEMAPHORE_FEATURE_IMPORTABLE_BIT
;
7549 } else if (type
== VK_SEMAPHORE_TYPE_TIMELINE
) {
7550 pExternalSemaphoreProperties
->exportFromImportedHandleTypes
= 0;
7551 pExternalSemaphoreProperties
->compatibleHandleTypes
= 0;
7552 pExternalSemaphoreProperties
->externalSemaphoreFeatures
= 0;
7554 /* Require has_syncobj_wait_for_submit for the syncobj signal ioctl introduced at virtually the same time */
7555 } else if (pdevice
->rad_info
.has_syncobj_wait_for_submit
&&
7556 (pExternalSemaphoreInfo
->handleType
== VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT
||
7557 pExternalSemaphoreInfo
->handleType
== VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT
)) {
7558 pExternalSemaphoreProperties
->exportFromImportedHandleTypes
= VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT
| VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT
;
7559 pExternalSemaphoreProperties
->compatibleHandleTypes
= VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT
| VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT
;
7560 pExternalSemaphoreProperties
->externalSemaphoreFeatures
= VK_EXTERNAL_SEMAPHORE_FEATURE_EXPORTABLE_BIT
|
7561 VK_EXTERNAL_SEMAPHORE_FEATURE_IMPORTABLE_BIT
;
7562 } else if (pExternalSemaphoreInfo
->handleType
== VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT
) {
7563 pExternalSemaphoreProperties
->exportFromImportedHandleTypes
= VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT
;
7564 pExternalSemaphoreProperties
->compatibleHandleTypes
= VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT
;
7565 pExternalSemaphoreProperties
->externalSemaphoreFeatures
= VK_EXTERNAL_SEMAPHORE_FEATURE_EXPORTABLE_BIT
|
7566 VK_EXTERNAL_SEMAPHORE_FEATURE_IMPORTABLE_BIT
;
7568 pExternalSemaphoreProperties
->exportFromImportedHandleTypes
= 0;
7569 pExternalSemaphoreProperties
->compatibleHandleTypes
= 0;
7570 pExternalSemaphoreProperties
->externalSemaphoreFeatures
= 0;
7574 VkResult
radv_ImportFenceFdKHR(VkDevice _device
,
7575 const VkImportFenceFdInfoKHR
*pImportFenceFdInfo
)
7577 RADV_FROM_HANDLE(radv_device
, device
, _device
);
7578 RADV_FROM_HANDLE(radv_fence
, fence
, pImportFenceFdInfo
->fence
);
7579 struct radv_fence_part
*dst
= NULL
;
7582 if (pImportFenceFdInfo
->flags
& VK_FENCE_IMPORT_TEMPORARY_BIT
) {
7583 dst
= &fence
->temporary
;
7585 dst
= &fence
->permanent
;
7588 uint32_t syncobj
= dst
->kind
== RADV_FENCE_SYNCOBJ
? dst
->syncobj
: 0;
7590 switch(pImportFenceFdInfo
->handleType
) {
7591 case VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT
:
7592 result
= radv_import_opaque_fd(device
, pImportFenceFdInfo
->fd
, &syncobj
);
7594 case VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT
:
7595 result
= radv_import_sync_fd(device
, pImportFenceFdInfo
->fd
, &syncobj
);
7598 unreachable("Unhandled fence handle type");
7601 if (result
== VK_SUCCESS
) {
7602 dst
->syncobj
= syncobj
;
7603 dst
->kind
= RADV_FENCE_SYNCOBJ
;
7609 VkResult
radv_GetFenceFdKHR(VkDevice _device
,
7610 const VkFenceGetFdInfoKHR
*pGetFdInfo
,
7613 RADV_FROM_HANDLE(radv_device
, device
, _device
);
7614 RADV_FROM_HANDLE(radv_fence
, fence
, pGetFdInfo
->fence
);
7617 struct radv_fence_part
*part
=
7618 fence
->temporary
.kind
!= RADV_FENCE_NONE
?
7619 &fence
->temporary
: &fence
->permanent
;
7621 switch(pGetFdInfo
->handleType
) {
7622 case VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT
:
7623 ret
= device
->ws
->export_syncobj(device
->ws
, part
->syncobj
, pFd
);
7625 return vk_error(device
->instance
, VK_ERROR_TOO_MANY_OBJECTS
);
7627 case VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT
:
7628 ret
= device
->ws
->export_syncobj_to_sync_file(device
->ws
,
7629 part
->syncobj
, pFd
);
7631 return vk_error(device
->instance
, VK_ERROR_TOO_MANY_OBJECTS
);
7633 if (part
== &fence
->temporary
) {
7634 radv_destroy_fence_part(device
, part
);
7636 device
->ws
->reset_syncobj(device
->ws
, part
->syncobj
);
7640 unreachable("Unhandled fence handle type");
7646 void radv_GetPhysicalDeviceExternalFenceProperties(
7647 VkPhysicalDevice physicalDevice
,
7648 const VkPhysicalDeviceExternalFenceInfo
*pExternalFenceInfo
,
7649 VkExternalFenceProperties
*pExternalFenceProperties
)
7651 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
7653 if (pdevice
->rad_info
.has_syncobj_wait_for_submit
&&
7654 (pExternalFenceInfo
->handleType
== VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT
||
7655 pExternalFenceInfo
->handleType
== VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT
)) {
7656 pExternalFenceProperties
->exportFromImportedHandleTypes
= VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT
| VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT
;
7657 pExternalFenceProperties
->compatibleHandleTypes
= VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT
| VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT
;
7658 pExternalFenceProperties
->externalFenceFeatures
= VK_EXTERNAL_FENCE_FEATURE_EXPORTABLE_BIT
|
7659 VK_EXTERNAL_SEMAPHORE_FEATURE_IMPORTABLE_BIT
;
7661 pExternalFenceProperties
->exportFromImportedHandleTypes
= 0;
7662 pExternalFenceProperties
->compatibleHandleTypes
= 0;
7663 pExternalFenceProperties
->externalFenceFeatures
= 0;
7668 radv_CreateDebugReportCallbackEXT(VkInstance _instance
,
7669 const VkDebugReportCallbackCreateInfoEXT
* pCreateInfo
,
7670 const VkAllocationCallbacks
* pAllocator
,
7671 VkDebugReportCallbackEXT
* pCallback
)
7673 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
7674 return vk_create_debug_report_callback(&instance
->debug_report_callbacks
,
7675 pCreateInfo
, pAllocator
, &instance
->alloc
,
7680 radv_DestroyDebugReportCallbackEXT(VkInstance _instance
,
7681 VkDebugReportCallbackEXT _callback
,
7682 const VkAllocationCallbacks
* pAllocator
)
7684 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
7685 vk_destroy_debug_report_callback(&instance
->debug_report_callbacks
,
7686 _callback
, pAllocator
, &instance
->alloc
);
7690 radv_DebugReportMessageEXT(VkInstance _instance
,
7691 VkDebugReportFlagsEXT flags
,
7692 VkDebugReportObjectTypeEXT objectType
,
7695 int32_t messageCode
,
7696 const char* pLayerPrefix
,
7697 const char* pMessage
)
7699 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
7700 vk_debug_report(&instance
->debug_report_callbacks
, flags
, objectType
,
7701 object
, location
, messageCode
, pLayerPrefix
, pMessage
);
7705 radv_GetDeviceGroupPeerMemoryFeatures(
7708 uint32_t localDeviceIndex
,
7709 uint32_t remoteDeviceIndex
,
7710 VkPeerMemoryFeatureFlags
* pPeerMemoryFeatures
)
7712 assert(localDeviceIndex
== remoteDeviceIndex
);
7714 *pPeerMemoryFeatures
= VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT
|
7715 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT
|
7716 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT
|
7717 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT
;
7720 static const VkTimeDomainEXT radv_time_domains
[] = {
7721 VK_TIME_DOMAIN_DEVICE_EXT
,
7722 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
,
7723 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
,
7726 VkResult
radv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
7727 VkPhysicalDevice physicalDevice
,
7728 uint32_t *pTimeDomainCount
,
7729 VkTimeDomainEXT
*pTimeDomains
)
7732 VK_OUTARRAY_MAKE(out
, pTimeDomains
, pTimeDomainCount
);
7734 for (d
= 0; d
< ARRAY_SIZE(radv_time_domains
); d
++) {
7735 vk_outarray_append(&out
, i
) {
7736 *i
= radv_time_domains
[d
];
7740 return vk_outarray_status(&out
);
7744 radv_clock_gettime(clockid_t clock_id
)
7746 struct timespec current
;
7749 ret
= clock_gettime(clock_id
, ¤t
);
7750 if (ret
< 0 && clock_id
== CLOCK_MONOTONIC_RAW
)
7751 ret
= clock_gettime(CLOCK_MONOTONIC
, ¤t
);
7755 return (uint64_t) current
.tv_sec
* 1000000000ULL + current
.tv_nsec
;
7758 VkResult
radv_GetCalibratedTimestampsEXT(
7760 uint32_t timestampCount
,
7761 const VkCalibratedTimestampInfoEXT
*pTimestampInfos
,
7762 uint64_t *pTimestamps
,
7763 uint64_t *pMaxDeviation
)
7765 RADV_FROM_HANDLE(radv_device
, device
, _device
);
7766 uint32_t clock_crystal_freq
= device
->physical_device
->rad_info
.clock_crystal_freq
;
7768 uint64_t begin
, end
;
7769 uint64_t max_clock_period
= 0;
7771 begin
= radv_clock_gettime(CLOCK_MONOTONIC_RAW
);
7773 for (d
= 0; d
< timestampCount
; d
++) {
7774 switch (pTimestampInfos
[d
].timeDomain
) {
7775 case VK_TIME_DOMAIN_DEVICE_EXT
:
7776 pTimestamps
[d
] = device
->ws
->query_value(device
->ws
,
7778 uint64_t device_period
= DIV_ROUND_UP(1000000, clock_crystal_freq
);
7779 max_clock_period
= MAX2(max_clock_period
, device_period
);
7781 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
:
7782 pTimestamps
[d
] = radv_clock_gettime(CLOCK_MONOTONIC
);
7783 max_clock_period
= MAX2(max_clock_period
, 1);
7786 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
:
7787 pTimestamps
[d
] = begin
;
7795 end
= radv_clock_gettime(CLOCK_MONOTONIC_RAW
);
7798 * The maximum deviation is the sum of the interval over which we
7799 * perform the sampling and the maximum period of any sampled
7800 * clock. That's because the maximum skew between any two sampled
7801 * clock edges is when the sampled clock with the largest period is
7802 * sampled at the end of that period but right at the beginning of the
7803 * sampling interval and some other clock is sampled right at the
7804 * begining of its sampling period and right at the end of the
7805 * sampling interval. Let's assume the GPU has the longest clock
7806 * period and that the application is sampling GPU and monotonic:
7809 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
7810 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
7814 * GPU -----_____-----_____-----_____-----_____
7817 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
7818 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
7820 * Interval <----------------->
7821 * Deviation <-------------------------->
7825 * m = read(monotonic) 2
7828 * We round the sample interval up by one tick to cover sampling error
7829 * in the interval clock
7832 uint64_t sample_interval
= end
- begin
+ 1;
7834 *pMaxDeviation
= sample_interval
+ max_clock_period
;
7839 void radv_GetPhysicalDeviceMultisamplePropertiesEXT(
7840 VkPhysicalDevice physicalDevice
,
7841 VkSampleCountFlagBits samples
,
7842 VkMultisamplePropertiesEXT
* pMultisampleProperties
)
7844 if (samples
& (VK_SAMPLE_COUNT_2_BIT
|
7845 VK_SAMPLE_COUNT_4_BIT
|
7846 VK_SAMPLE_COUNT_8_BIT
)) {
7847 pMultisampleProperties
->maxSampleLocationGridSize
= (VkExtent2D
){ 2, 2 };
7849 pMultisampleProperties
->maxSampleLocationGridSize
= (VkExtent2D
){ 0, 0 };
7853 VkResult
radv_CreatePrivateDataSlotEXT(
7855 const VkPrivateDataSlotCreateInfoEXT
* pCreateInfo
,
7856 const VkAllocationCallbacks
* pAllocator
,
7857 VkPrivateDataSlotEXT
* pPrivateDataSlot
)
7859 RADV_FROM_HANDLE(radv_device
, device
, _device
);
7860 return vk_private_data_slot_create(&device
->vk
, pCreateInfo
, pAllocator
,
7864 void radv_DestroyPrivateDataSlotEXT(
7866 VkPrivateDataSlotEXT privateDataSlot
,
7867 const VkAllocationCallbacks
* pAllocator
)
7869 RADV_FROM_HANDLE(radv_device
, device
, _device
);
7870 vk_private_data_slot_destroy(&device
->vk
, privateDataSlot
, pAllocator
);
7873 VkResult
radv_SetPrivateDataEXT(
7875 VkObjectType objectType
,
7876 uint64_t objectHandle
,
7877 VkPrivateDataSlotEXT privateDataSlot
,
7880 RADV_FROM_HANDLE(radv_device
, device
, _device
);
7881 return vk_object_base_set_private_data(&device
->vk
, objectType
,
7882 objectHandle
, privateDataSlot
,
7886 void radv_GetPrivateDataEXT(
7888 VkObjectType objectType
,
7889 uint64_t objectHandle
,
7890 VkPrivateDataSlotEXT privateDataSlot
,
7893 RADV_FROM_HANDLE(radv_device
, device
, _device
);
7894 vk_object_base_get_private_data(&device
->vk
, objectType
, objectHandle
,
7895 privateDataSlot
, pData
);