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>
40 #include <sys/prctl.h>
45 #include "radv_debug.h"
46 #include "radv_private.h"
47 #include "radv_shader.h"
49 #include "util/disk_cache.h"
53 #include "drm-uapi/amdgpu_drm.h"
54 #include "winsys/amdgpu/radv_amdgpu_winsys_public.h"
55 #include "winsys/null/radv_null_winsys_public.h"
56 #include "ac_llvm_util.h"
57 #include "vk_format.h"
60 #include "util/build_id.h"
61 #include "util/debug.h"
62 #include "util/mesa-sha1.h"
63 #include "util/timespec.h"
64 #include "util/u_atomic.h"
65 #include "compiler/glsl_types.h"
66 #include "util/xmlpool.h"
68 static struct radv_timeline_point
*
69 radv_timeline_find_point_at_least_locked(struct radv_device
*device
,
70 struct radv_timeline
*timeline
,
73 static struct radv_timeline_point
*
74 radv_timeline_add_point_locked(struct radv_device
*device
,
75 struct radv_timeline
*timeline
,
79 radv_timeline_trigger_waiters_locked(struct radv_timeline
*timeline
,
80 struct list_head
*processing_list
);
83 void radv_destroy_semaphore_part(struct radv_device
*device
,
84 struct radv_semaphore_part
*part
);
87 radv_device_get_cache_uuid(enum radeon_family family
, void *uuid
)
90 unsigned char sha1
[20];
91 unsigned ptr_size
= sizeof(void*);
93 memset(uuid
, 0, VK_UUID_SIZE
);
94 _mesa_sha1_init(&ctx
);
96 if (!disk_cache_get_function_identifier(radv_device_get_cache_uuid
, &ctx
) ||
97 !disk_cache_get_function_identifier(LLVMInitializeAMDGPUTargetInfo
, &ctx
))
100 _mesa_sha1_update(&ctx
, &family
, sizeof(family
));
101 _mesa_sha1_update(&ctx
, &ptr_size
, sizeof(ptr_size
));
102 _mesa_sha1_final(&ctx
, sha1
);
104 memcpy(uuid
, sha1
, VK_UUID_SIZE
);
109 radv_get_driver_uuid(void *uuid
)
111 ac_compute_driver_uuid(uuid
, VK_UUID_SIZE
);
115 radv_get_device_uuid(struct radeon_info
*info
, void *uuid
)
117 ac_compute_device_uuid(info
, uuid
, VK_UUID_SIZE
);
121 radv_get_visible_vram_size(struct radv_physical_device
*device
)
123 return MIN2(device
->rad_info
.vram_size
, device
->rad_info
.vram_vis_size
);
127 radv_get_vram_size(struct radv_physical_device
*device
)
129 return device
->rad_info
.vram_size
- radv_get_visible_vram_size(device
);
133 radv_physical_device_init_mem_types(struct radv_physical_device
*device
)
135 uint64_t visible_vram_size
= radv_get_visible_vram_size(device
);
136 uint64_t vram_size
= radv_get_vram_size(device
);
137 int vram_index
= -1, visible_vram_index
= -1, gart_index
= -1;
138 device
->memory_properties
.memoryHeapCount
= 0;
140 vram_index
= device
->memory_properties
.memoryHeapCount
++;
141 device
->memory_properties
.memoryHeaps
[vram_index
] = (VkMemoryHeap
) {
143 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
147 if (device
->rad_info
.gart_size
> 0) {
148 gart_index
= device
->memory_properties
.memoryHeapCount
++;
149 device
->memory_properties
.memoryHeaps
[gart_index
] = (VkMemoryHeap
) {
150 .size
= device
->rad_info
.gart_size
,
155 if (visible_vram_size
) {
156 visible_vram_index
= device
->memory_properties
.memoryHeapCount
++;
157 device
->memory_properties
.memoryHeaps
[visible_vram_index
] = (VkMemoryHeap
) {
158 .size
= visible_vram_size
,
159 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
163 unsigned type_count
= 0;
165 if (device
->rad_info
.has_dedicated_vram
) {
166 if (vram_index
>= 0) {
167 device
->memory_domains
[type_count
] = RADEON_DOMAIN_VRAM
;
168 device
->memory_flags
[type_count
] = RADEON_FLAG_NO_CPU_ACCESS
;
169 device
->memory_properties
.memoryTypes
[type_count
++] = (VkMemoryType
) {
170 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
,
171 .heapIndex
= vram_index
,
175 if (visible_vram_index
>= 0) {
176 device
->memory_domains
[type_count
] = RADEON_DOMAIN_VRAM
;
177 device
->memory_flags
[type_count
] = RADEON_FLAG_NO_CPU_ACCESS
;
178 device
->memory_properties
.memoryTypes
[type_count
++] = (VkMemoryType
) {
179 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
,
180 .heapIndex
= visible_vram_index
,
185 if (gart_index
>= 0) {
186 device
->memory_domains
[type_count
] = RADEON_DOMAIN_GTT
;
187 device
->memory_flags
[type_count
] = RADEON_FLAG_GTT_WC
| RADEON_FLAG_CPU_ACCESS
;
188 device
->memory_properties
.memoryTypes
[type_count
++] = (VkMemoryType
) {
189 .propertyFlags
= VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
190 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
191 .heapIndex
= gart_index
,
194 if (visible_vram_index
>= 0) {
195 device
->memory_domains
[type_count
] = RADEON_DOMAIN_VRAM
;
196 device
->memory_flags
[type_count
] = RADEON_FLAG_CPU_ACCESS
;
197 device
->memory_properties
.memoryTypes
[type_count
++] = (VkMemoryType
) {
198 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
199 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
200 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
201 .heapIndex
= visible_vram_index
,
205 if (gart_index
>= 0) {
206 device
->memory_domains
[type_count
] = RADEON_DOMAIN_GTT
;
207 device
->memory_flags
[type_count
] = RADEON_FLAG_CPU_ACCESS
;
208 device
->memory_properties
.memoryTypes
[type_count
++] = (VkMemoryType
) {
209 .propertyFlags
= VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
210 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
211 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
212 .heapIndex
= gart_index
,
215 device
->memory_properties
.memoryTypeCount
= type_count
;
217 if (device
->rad_info
.has_l2_uncached
) {
218 for (int i
= 0; i
< device
->memory_properties
.memoryTypeCount
; i
++) {
219 VkMemoryType mem_type
= device
->memory_properties
.memoryTypes
[i
];
221 if ((mem_type
.propertyFlags
& (VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
222 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
)) ||
223 mem_type
.propertyFlags
== VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
) {
225 VkMemoryPropertyFlags property_flags
= mem_type
.propertyFlags
|
226 VK_MEMORY_PROPERTY_DEVICE_COHERENT_BIT_AMD
|
227 VK_MEMORY_PROPERTY_DEVICE_UNCACHED_BIT_AMD
;
229 device
->memory_domains
[type_count
] = device
->memory_domains
[i
];
230 device
->memory_flags
[type_count
] = device
->memory_flags
[i
] | RADEON_FLAG_VA_UNCACHED
;
231 device
->memory_properties
.memoryTypes
[type_count
++] = (VkMemoryType
) {
232 .propertyFlags
= property_flags
,
233 .heapIndex
= mem_type
.heapIndex
,
237 device
->memory_properties
.memoryTypeCount
= type_count
;
242 radv_physical_device_try_create(struct radv_instance
*instance
,
243 drmDevicePtr drm_device
,
244 struct radv_physical_device
**device_out
)
251 const char *path
= drm_device
->nodes
[DRM_NODE_RENDER
];
252 drmVersionPtr version
;
254 fd
= open(path
, O_RDWR
| O_CLOEXEC
);
256 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
257 radv_logi("Could not open device '%s'", path
);
259 return vk_error(instance
, VK_ERROR_INCOMPATIBLE_DRIVER
);
262 version
= drmGetVersion(fd
);
266 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
267 radv_logi("Could not get the kernel driver version for device '%s'", path
);
269 return vk_errorf(instance
, VK_ERROR_INCOMPATIBLE_DRIVER
,
270 "failed to get version %s: %m", path
);
273 if (strcmp(version
->name
, "amdgpu")) {
274 drmFreeVersion(version
);
277 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
278 radv_logi("Device '%s' is not using the amdgpu kernel driver.", path
);
280 return VK_ERROR_INCOMPATIBLE_DRIVER
;
282 drmFreeVersion(version
);
284 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
285 radv_logi("Found compatible device '%s'.", path
);
288 struct radv_physical_device
*device
=
289 vk_zalloc2(&instance
->alloc
, NULL
, sizeof(*device
), 8,
290 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
292 result
= vk_error(instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
296 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
297 device
->instance
= instance
;
300 device
->ws
= radv_amdgpu_winsys_create(fd
, instance
->debug_flags
,
301 instance
->perftest_flags
);
303 device
->ws
= radv_null_winsys_create();
307 result
= vk_errorf(instance
, VK_ERROR_INITIALIZATION_FAILED
,
308 "failed to initialize winsys");
312 if (drm_device
&& instance
->enabled_extensions
.KHR_display
) {
313 master_fd
= open(drm_device
->nodes
[DRM_NODE_PRIMARY
], O_RDWR
| O_CLOEXEC
);
314 if (master_fd
>= 0) {
315 uint32_t accel_working
= 0;
316 struct drm_amdgpu_info request
= {
317 .return_pointer
= (uintptr_t)&accel_working
,
318 .return_size
= sizeof(accel_working
),
319 .query
= AMDGPU_INFO_ACCEL_WORKING
322 if (drmCommandWrite(master_fd
, DRM_AMDGPU_INFO
, &request
, sizeof (struct drm_amdgpu_info
)) < 0 || !accel_working
) {
329 device
->master_fd
= master_fd
;
330 device
->local_fd
= fd
;
331 device
->ws
->query_info(device
->ws
, &device
->rad_info
);
333 device
->use_aco
= instance
->perftest_flags
& RADV_PERFTEST_ACO
;
335 snprintf(device
->name
, sizeof(device
->name
),
337 device
->rad_info
.name
,
338 device
->use_aco
? "ACO" : "LLVM " MESA_LLVM_VERSION_STRING
);
340 if (radv_device_get_cache_uuid(device
->rad_info
.family
, device
->cache_uuid
)) {
341 result
= vk_errorf(instance
, VK_ERROR_INITIALIZATION_FAILED
,
342 "cannot generate UUID");
346 /* These flags affect shader compilation. */
347 uint64_t shader_env_flags
= (device
->use_aco
? 0x2 : 0);
349 /* The gpu id is already embedded in the uuid so we just pass "radv"
350 * when creating the cache.
352 char buf
[VK_UUID_SIZE
* 2 + 1];
353 disk_cache_format_hex_id(buf
, device
->cache_uuid
, VK_UUID_SIZE
* 2);
354 device
->disk_cache
= disk_cache_create(device
->name
, buf
, shader_env_flags
);
356 if (device
->rad_info
.chip_class
< GFX8
)
357 fprintf(stderr
, "WARNING: radv is not a conformant vulkan implementation, testing use only.\n");
359 radv_get_driver_uuid(&device
->driver_uuid
);
360 radv_get_device_uuid(&device
->rad_info
, &device
->device_uuid
);
362 device
->out_of_order_rast_allowed
= device
->rad_info
.has_out_of_order_rast
&&
363 !(device
->instance
->debug_flags
& RADV_DEBUG_NO_OUT_OF_ORDER
);
365 device
->dcc_msaa_allowed
=
366 (device
->instance
->perftest_flags
& RADV_PERFTEST_DCC_MSAA
);
368 device
->use_shader_ballot
= (device
->use_aco
&& device
->rad_info
.chip_class
>= GFX8
) ||
369 (device
->instance
->perftest_flags
& RADV_PERFTEST_SHADER_BALLOT
);
371 device
->use_ngg
= device
->rad_info
.chip_class
>= GFX10
&&
372 device
->rad_info
.family
!= CHIP_NAVI14
&&
373 !(device
->instance
->debug_flags
& RADV_DEBUG_NO_NGG
);
375 /* TODO: Implement NGG GS with ACO. */
376 device
->use_ngg_gs
= device
->use_ngg
&& !device
->use_aco
;
377 device
->use_ngg_streamout
= false;
379 /* Determine the number of threads per wave for all stages. */
380 device
->cs_wave_size
= 64;
381 device
->ps_wave_size
= 64;
382 device
->ge_wave_size
= 64;
384 if (device
->rad_info
.chip_class
>= GFX10
) {
385 if (device
->instance
->perftest_flags
& RADV_PERFTEST_CS_WAVE_32
)
386 device
->cs_wave_size
= 32;
388 /* For pixel shaders, wave64 is recommanded. */
389 if (device
->instance
->perftest_flags
& RADV_PERFTEST_PS_WAVE_32
)
390 device
->ps_wave_size
= 32;
392 if (device
->instance
->perftest_flags
& RADV_PERFTEST_GE_WAVE_32
)
393 device
->ge_wave_size
= 32;
396 radv_physical_device_init_mem_types(device
);
397 radv_fill_device_extension_table(device
, &device
->supported_extensions
);
400 device
->bus_info
= *drm_device
->businfo
.pci
;
402 if ((device
->instance
->debug_flags
& RADV_DEBUG_INFO
))
403 ac_print_gpu_info(&device
->rad_info
);
405 /* The WSI is structured as a layer on top of the driver, so this has
406 * to be the last part of initialization (at least until we get other
409 result
= radv_init_wsi(device
);
410 if (result
!= VK_SUCCESS
) {
411 vk_error(instance
, result
);
412 goto fail_disk_cache
;
415 *device_out
= device
;
420 disk_cache_destroy(device
->disk_cache
);
422 device
->ws
->destroy(device
->ws
);
424 vk_free(&instance
->alloc
, device
);
433 radv_physical_device_destroy(struct radv_physical_device
*device
)
435 radv_finish_wsi(device
);
436 device
->ws
->destroy(device
->ws
);
437 disk_cache_destroy(device
->disk_cache
);
438 close(device
->local_fd
);
439 if (device
->master_fd
!= -1)
440 close(device
->master_fd
);
441 vk_free(&device
->instance
->alloc
, device
);
445 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
446 VkSystemAllocationScope allocationScope
)
452 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
453 size_t align
, VkSystemAllocationScope allocationScope
)
455 return realloc(pOriginal
, size
);
459 default_free_func(void *pUserData
, void *pMemory
)
464 static const VkAllocationCallbacks default_alloc
= {
466 .pfnAllocation
= default_alloc_func
,
467 .pfnReallocation
= default_realloc_func
,
468 .pfnFree
= default_free_func
,
471 static const struct debug_control radv_debug_options
[] = {
472 {"nofastclears", RADV_DEBUG_NO_FAST_CLEARS
},
473 {"nodcc", RADV_DEBUG_NO_DCC
},
474 {"shaders", RADV_DEBUG_DUMP_SHADERS
},
475 {"nocache", RADV_DEBUG_NO_CACHE
},
476 {"shaderstats", RADV_DEBUG_DUMP_SHADER_STATS
},
477 {"nohiz", RADV_DEBUG_NO_HIZ
},
478 {"nocompute", RADV_DEBUG_NO_COMPUTE_QUEUE
},
479 {"allbos", RADV_DEBUG_ALL_BOS
},
480 {"noibs", RADV_DEBUG_NO_IBS
},
481 {"spirv", RADV_DEBUG_DUMP_SPIRV
},
482 {"vmfaults", RADV_DEBUG_VM_FAULTS
},
483 {"zerovram", RADV_DEBUG_ZERO_VRAM
},
484 {"syncshaders", RADV_DEBUG_SYNC_SHADERS
},
485 {"preoptir", RADV_DEBUG_PREOPTIR
},
486 {"nodynamicbounds", RADV_DEBUG_NO_DYNAMIC_BOUNDS
},
487 {"nooutoforder", RADV_DEBUG_NO_OUT_OF_ORDER
},
488 {"info", RADV_DEBUG_INFO
},
489 {"errors", RADV_DEBUG_ERRORS
},
490 {"startup", RADV_DEBUG_STARTUP
},
491 {"checkir", RADV_DEBUG_CHECKIR
},
492 {"nothreadllvm", RADV_DEBUG_NOTHREADLLVM
},
493 {"nobinning", RADV_DEBUG_NOBINNING
},
494 {"noloadstoreopt", RADV_DEBUG_NO_LOAD_STORE_OPT
},
495 {"nongg", RADV_DEBUG_NO_NGG
},
496 {"noshaderballot", RADV_DEBUG_NO_SHADER_BALLOT
},
497 {"allentrypoints", RADV_DEBUG_ALL_ENTRYPOINTS
},
498 {"metashaders", RADV_DEBUG_DUMP_META_SHADERS
},
499 {"nomemorycache", RADV_DEBUG_NO_MEMORY_CACHE
},
504 radv_get_debug_option_name(int id
)
506 assert(id
< ARRAY_SIZE(radv_debug_options
) - 1);
507 return radv_debug_options
[id
].string
;
510 static const struct debug_control radv_perftest_options
[] = {
511 {"localbos", RADV_PERFTEST_LOCAL_BOS
},
512 {"dccmsaa", RADV_PERFTEST_DCC_MSAA
},
513 {"bolist", RADV_PERFTEST_BO_LIST
},
514 {"shader_ballot", RADV_PERFTEST_SHADER_BALLOT
},
515 {"tccompatcmask", RADV_PERFTEST_TC_COMPAT_CMASK
},
516 {"cswave32", RADV_PERFTEST_CS_WAVE_32
},
517 {"pswave32", RADV_PERFTEST_PS_WAVE_32
},
518 {"gewave32", RADV_PERFTEST_GE_WAVE_32
},
519 {"dfsm", RADV_PERFTEST_DFSM
},
520 {"aco", RADV_PERFTEST_ACO
},
525 radv_get_perftest_option_name(int id
)
527 assert(id
< ARRAY_SIZE(radv_perftest_options
) - 1);
528 return radv_perftest_options
[id
].string
;
532 radv_handle_per_app_options(struct radv_instance
*instance
,
533 const VkApplicationInfo
*info
)
535 const char *name
= info
? info
->pApplicationName
: NULL
;
540 if (!strcmp(name
, "DOOM_VFR")) {
541 /* Work around a Doom VFR game bug */
542 instance
->debug_flags
|= RADV_DEBUG_NO_DYNAMIC_BOUNDS
;
543 } else if (!strcmp(name
, "MonsterHunterWorld.exe")) {
544 /* Workaround for a WaW hazard when LLVM moves/merges
545 * load/store memory operations.
546 * See https://reviews.llvm.org/D61313
548 if (LLVM_VERSION_MAJOR
< 9)
549 instance
->debug_flags
|= RADV_DEBUG_NO_LOAD_STORE_OPT
;
550 } else if (!strcmp(name
, "Wolfenstein: Youngblood")) {
551 if (!(instance
->debug_flags
& RADV_DEBUG_NO_SHADER_BALLOT
) &&
552 !(instance
->perftest_flags
& RADV_PERFTEST_ACO
)) {
553 /* Force enable VK_AMD_shader_ballot because it looks
554 * safe and it gives a nice boost (+20% on Vega 56 at
555 * this time). It also prevents corruption on LLVM.
557 instance
->perftest_flags
|= RADV_PERFTEST_SHADER_BALLOT
;
559 } else if (!strcmp(name
, "Fledge")) {
561 * Zero VRAM for "The Surge 2"
563 * This avoid a hang when when rendering any level. Likely
564 * uninitialized data in an indirect draw.
566 instance
->debug_flags
|= RADV_DEBUG_ZERO_VRAM
;
567 } else if (!strcmp(name
, "No Man's Sky")) {
568 /* Work around a NMS game bug */
569 instance
->debug_flags
|= RADV_DEBUG_DISCARD_TO_DEMOTE
;
573 static const char radv_dri_options_xml
[] =
575 DRI_CONF_SECTION_PERFORMANCE
576 DRI_CONF_ADAPTIVE_SYNC("true")
577 DRI_CONF_VK_X11_OVERRIDE_MIN_IMAGE_COUNT(0)
578 DRI_CONF_VK_X11_STRICT_IMAGE_COUNT("false")
581 DRI_CONF_SECTION_DEBUG
582 DRI_CONF_VK_WSI_FORCE_BGRA8_UNORM_FIRST("false")
586 static void radv_init_dri_options(struct radv_instance
*instance
)
588 driParseOptionInfo(&instance
->available_dri_options
, radv_dri_options_xml
);
589 driParseConfigFiles(&instance
->dri_options
,
590 &instance
->available_dri_options
,
592 instance
->engineName
,
593 instance
->engineVersion
);
596 VkResult
radv_CreateInstance(
597 const VkInstanceCreateInfo
* pCreateInfo
,
598 const VkAllocationCallbacks
* pAllocator
,
599 VkInstance
* pInstance
)
601 struct radv_instance
*instance
;
604 instance
= vk_zalloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
605 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
607 return vk_error(NULL
, VK_ERROR_OUT_OF_HOST_MEMORY
);
609 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
612 instance
->alloc
= *pAllocator
;
614 instance
->alloc
= default_alloc
;
616 if (pCreateInfo
->pApplicationInfo
) {
617 const VkApplicationInfo
*app
= pCreateInfo
->pApplicationInfo
;
619 instance
->engineName
=
620 vk_strdup(&instance
->alloc
, app
->pEngineName
,
621 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
622 instance
->engineVersion
= app
->engineVersion
;
623 instance
->apiVersion
= app
->apiVersion
;
626 if (instance
->apiVersion
== 0)
627 instance
->apiVersion
= VK_API_VERSION_1_0
;
629 /* Get secure compile thread count. NOTE: We cap this at 32 */
630 #define MAX_SC_PROCS 32
631 char *num_sc_threads
= getenv("RADV_SECURE_COMPILE_THREADS");
633 instance
->num_sc_threads
= MIN2(strtoul(num_sc_threads
, NULL
, 10), MAX_SC_PROCS
);
635 instance
->debug_flags
= parse_debug_string(getenv("RADV_DEBUG"),
638 /* Disable memory cache when secure compile is set */
639 if (radv_device_use_secure_compile(instance
))
640 instance
->debug_flags
|= RADV_DEBUG_NO_MEMORY_CACHE
;
642 instance
->perftest_flags
= parse_debug_string(getenv("RADV_PERFTEST"),
643 radv_perftest_options
);
645 if (instance
->perftest_flags
& RADV_PERFTEST_ACO
)
646 fprintf(stderr
, "WARNING: Experimental compiler backend enabled. Here be dragons! Incorrect rendering, GPU hangs and/or resets are likely\n");
648 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
649 radv_logi("Created an instance");
651 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
653 for (idx
= 0; idx
< RADV_INSTANCE_EXTENSION_COUNT
; idx
++) {
654 if (!strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
655 radv_instance_extensions
[idx
].extensionName
))
659 if (idx
>= RADV_INSTANCE_EXTENSION_COUNT
||
660 !radv_supported_instance_extensions
.extensions
[idx
]) {
661 vk_free2(&default_alloc
, pAllocator
, instance
);
662 return vk_error(instance
, VK_ERROR_EXTENSION_NOT_PRESENT
);
665 instance
->enabled_extensions
.extensions
[idx
] = true;
668 bool unchecked
= instance
->debug_flags
& RADV_DEBUG_ALL_ENTRYPOINTS
;
670 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->dispatch
.entrypoints
); i
++) {
671 /* Vulkan requires that entrypoints for extensions which have
672 * not been enabled must not be advertised.
675 !radv_instance_entrypoint_is_enabled(i
, instance
->apiVersion
,
676 &instance
->enabled_extensions
)) {
677 instance
->dispatch
.entrypoints
[i
] = NULL
;
679 instance
->dispatch
.entrypoints
[i
] =
680 radv_instance_dispatch_table
.entrypoints
[i
];
684 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->physical_device_dispatch
.entrypoints
); i
++) {
685 /* Vulkan requires that entrypoints for extensions which have
686 * not been enabled must not be advertised.
689 !radv_physical_device_entrypoint_is_enabled(i
, instance
->apiVersion
,
690 &instance
->enabled_extensions
)) {
691 instance
->physical_device_dispatch
.entrypoints
[i
] = NULL
;
693 instance
->physical_device_dispatch
.entrypoints
[i
] =
694 radv_physical_device_dispatch_table
.entrypoints
[i
];
698 for (unsigned i
= 0; i
< ARRAY_SIZE(instance
->device_dispatch
.entrypoints
); i
++) {
699 /* Vulkan requires that entrypoints for extensions which have
700 * not been enabled must not be advertised.
703 !radv_device_entrypoint_is_enabled(i
, instance
->apiVersion
,
704 &instance
->enabled_extensions
, NULL
)) {
705 instance
->device_dispatch
.entrypoints
[i
] = NULL
;
707 instance
->device_dispatch
.entrypoints
[i
] =
708 radv_device_dispatch_table
.entrypoints
[i
];
712 instance
->physical_devices_enumerated
= false;
713 list_inithead(&instance
->physical_devices
);
715 result
= vk_debug_report_instance_init(&instance
->debug_report_callbacks
);
716 if (result
!= VK_SUCCESS
) {
717 vk_free2(&default_alloc
, pAllocator
, instance
);
718 return vk_error(instance
, result
);
721 glsl_type_singleton_init_or_ref();
723 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
725 radv_init_dri_options(instance
);
726 radv_handle_per_app_options(instance
, pCreateInfo
->pApplicationInfo
);
728 *pInstance
= radv_instance_to_handle(instance
);
733 void radv_DestroyInstance(
734 VkInstance _instance
,
735 const VkAllocationCallbacks
* pAllocator
)
737 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
742 list_for_each_entry_safe(struct radv_physical_device
, pdevice
,
743 &instance
->physical_devices
, link
) {
744 radv_physical_device_destroy(pdevice
);
747 vk_free(&instance
->alloc
, instance
->engineName
);
749 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
751 glsl_type_singleton_decref();
753 driDestroyOptionCache(&instance
->dri_options
);
754 driDestroyOptionInfo(&instance
->available_dri_options
);
756 vk_debug_report_instance_destroy(&instance
->debug_report_callbacks
);
758 vk_free(&instance
->alloc
, instance
);
762 radv_enumerate_physical_devices(struct radv_instance
*instance
)
764 if (instance
->physical_devices_enumerated
)
767 instance
->physical_devices_enumerated
= true;
769 /* TODO: Check for more devices ? */
770 drmDevicePtr devices
[8];
771 VkResult result
= VK_SUCCESS
;
774 if (getenv("RADV_FORCE_FAMILY")) {
775 /* When RADV_FORCE_FAMILY is set, the driver creates a nul
776 * device that allows to test the compiler without having an
779 struct radv_physical_device
*pdevice
;
781 result
= radv_physical_device_try_create(instance
, NULL
, &pdevice
);
782 if (result
!= VK_SUCCESS
)
785 list_addtail(&pdevice
->link
, &instance
->physical_devices
);
789 max_devices
= drmGetDevices2(0, devices
, ARRAY_SIZE(devices
));
791 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
792 radv_logi("Found %d drm nodes", max_devices
);
795 return vk_error(instance
, VK_SUCCESS
);
797 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
798 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
799 devices
[i
]->bustype
== DRM_BUS_PCI
&&
800 devices
[i
]->deviceinfo
.pci
->vendor_id
== ATI_VENDOR_ID
) {
802 struct radv_physical_device
*pdevice
;
803 result
= radv_physical_device_try_create(instance
, devices
[i
],
805 /* Incompatible DRM device, skip. */
806 if (result
== VK_ERROR_INCOMPATIBLE_DRIVER
) {
811 /* Error creating the physical device, report the error. */
812 if (result
!= VK_SUCCESS
)
815 list_addtail(&pdevice
->link
, &instance
->physical_devices
);
818 drmFreeDevices(devices
, max_devices
);
820 /* If we successfully enumerated any devices, call it success */
824 VkResult
radv_EnumeratePhysicalDevices(
825 VkInstance _instance
,
826 uint32_t* pPhysicalDeviceCount
,
827 VkPhysicalDevice
* pPhysicalDevices
)
829 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
830 VK_OUTARRAY_MAKE(out
, pPhysicalDevices
, pPhysicalDeviceCount
);
832 VkResult result
= radv_enumerate_physical_devices(instance
);
833 if (result
!= VK_SUCCESS
)
836 list_for_each_entry(struct radv_physical_device
, pdevice
,
837 &instance
->physical_devices
, link
) {
838 vk_outarray_append(&out
, i
) {
839 *i
= radv_physical_device_to_handle(pdevice
);
843 return vk_outarray_status(&out
);
846 VkResult
radv_EnumeratePhysicalDeviceGroups(
847 VkInstance _instance
,
848 uint32_t* pPhysicalDeviceGroupCount
,
849 VkPhysicalDeviceGroupProperties
* pPhysicalDeviceGroupProperties
)
851 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
852 VK_OUTARRAY_MAKE(out
, pPhysicalDeviceGroupProperties
,
853 pPhysicalDeviceGroupCount
);
855 VkResult result
= radv_enumerate_physical_devices(instance
);
856 if (result
!= VK_SUCCESS
)
859 list_for_each_entry(struct radv_physical_device
, pdevice
,
860 &instance
->physical_devices
, link
) {
861 vk_outarray_append(&out
, p
) {
862 p
->physicalDeviceCount
= 1;
863 memset(p
->physicalDevices
, 0, sizeof(p
->physicalDevices
));
864 p
->physicalDevices
[0] = radv_physical_device_to_handle(pdevice
);
865 p
->subsetAllocation
= false;
869 return vk_outarray_status(&out
);
872 void radv_GetPhysicalDeviceFeatures(
873 VkPhysicalDevice physicalDevice
,
874 VkPhysicalDeviceFeatures
* pFeatures
)
876 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
877 memset(pFeatures
, 0, sizeof(*pFeatures
));
879 *pFeatures
= (VkPhysicalDeviceFeatures
) {
880 .robustBufferAccess
= true,
881 .fullDrawIndexUint32
= true,
882 .imageCubeArray
= true,
883 .independentBlend
= true,
884 .geometryShader
= true,
885 .tessellationShader
= true,
886 .sampleRateShading
= true,
887 .dualSrcBlend
= true,
889 .multiDrawIndirect
= true,
890 .drawIndirectFirstInstance
= true,
892 .depthBiasClamp
= true,
893 .fillModeNonSolid
= true,
898 .multiViewport
= true,
899 .samplerAnisotropy
= true,
900 .textureCompressionETC2
= radv_device_supports_etc(pdevice
),
901 .textureCompressionASTC_LDR
= false,
902 .textureCompressionBC
= true,
903 .occlusionQueryPrecise
= true,
904 .pipelineStatisticsQuery
= true,
905 .vertexPipelineStoresAndAtomics
= true,
906 .fragmentStoresAndAtomics
= true,
907 .shaderTessellationAndGeometryPointSize
= true,
908 .shaderImageGatherExtended
= true,
909 .shaderStorageImageExtendedFormats
= true,
910 .shaderStorageImageMultisample
= true,
911 .shaderUniformBufferArrayDynamicIndexing
= true,
912 .shaderSampledImageArrayDynamicIndexing
= true,
913 .shaderStorageBufferArrayDynamicIndexing
= true,
914 .shaderStorageImageArrayDynamicIndexing
= true,
915 .shaderStorageImageReadWithoutFormat
= true,
916 .shaderStorageImageWriteWithoutFormat
= true,
917 .shaderClipDistance
= true,
918 .shaderCullDistance
= true,
919 .shaderFloat64
= true,
921 .shaderInt16
= !pdevice
->use_aco
|| pdevice
->rad_info
.chip_class
>= GFX8
,
922 .sparseBinding
= true,
923 .variableMultisampleRate
= true,
924 .inheritedQueries
= true,
928 void radv_GetPhysicalDeviceFeatures2(
929 VkPhysicalDevice physicalDevice
,
930 VkPhysicalDeviceFeatures2
*pFeatures
)
932 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
933 vk_foreach_struct(ext
, pFeatures
->pNext
) {
934 switch (ext
->sType
) {
935 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES
: {
936 VkPhysicalDeviceVariablePointersFeatures
*features
= (void *)ext
;
937 features
->variablePointersStorageBuffer
= true;
938 features
->variablePointers
= true;
941 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES
: {
942 VkPhysicalDeviceMultiviewFeatures
*features
= (VkPhysicalDeviceMultiviewFeatures
*)ext
;
943 features
->multiview
= true;
944 features
->multiviewGeometryShader
= true;
945 features
->multiviewTessellationShader
= true;
948 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETERS_FEATURES
: {
949 VkPhysicalDeviceShaderDrawParametersFeatures
*features
=
950 (VkPhysicalDeviceShaderDrawParametersFeatures
*)ext
;
951 features
->shaderDrawParameters
= true;
954 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES
: {
955 VkPhysicalDeviceProtectedMemoryFeatures
*features
=
956 (VkPhysicalDeviceProtectedMemoryFeatures
*)ext
;
957 features
->protectedMemory
= false;
960 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES
: {
961 VkPhysicalDevice16BitStorageFeatures
*features
=
962 (VkPhysicalDevice16BitStorageFeatures
*)ext
;
963 bool enable
= !pdevice
->use_aco
|| pdevice
->rad_info
.chip_class
>= GFX8
;
964 features
->storageBuffer16BitAccess
= enable
;
965 features
->uniformAndStorageBuffer16BitAccess
= enable
;
966 features
->storagePushConstant16
= enable
;
967 features
->storageInputOutput16
= pdevice
->rad_info
.has_double_rate_fp16
&& !pdevice
->use_aco
&& LLVM_VERSION_MAJOR
>= 9;
970 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES
: {
971 VkPhysicalDeviceSamplerYcbcrConversionFeatures
*features
=
972 (VkPhysicalDeviceSamplerYcbcrConversionFeatures
*)ext
;
973 features
->samplerYcbcrConversion
= true;
976 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES
: {
977 VkPhysicalDeviceDescriptorIndexingFeatures
*features
=
978 (VkPhysicalDeviceDescriptorIndexingFeatures
*)ext
;
979 features
->shaderInputAttachmentArrayDynamicIndexing
= true;
980 features
->shaderUniformTexelBufferArrayDynamicIndexing
= true;
981 features
->shaderStorageTexelBufferArrayDynamicIndexing
= true;
982 features
->shaderUniformBufferArrayNonUniformIndexing
= true;
983 features
->shaderSampledImageArrayNonUniformIndexing
= true;
984 features
->shaderStorageBufferArrayNonUniformIndexing
= true;
985 features
->shaderStorageImageArrayNonUniformIndexing
= true;
986 features
->shaderInputAttachmentArrayNonUniformIndexing
= true;
987 features
->shaderUniformTexelBufferArrayNonUniformIndexing
= true;
988 features
->shaderStorageTexelBufferArrayNonUniformIndexing
= true;
989 features
->descriptorBindingUniformBufferUpdateAfterBind
= true;
990 features
->descriptorBindingSampledImageUpdateAfterBind
= true;
991 features
->descriptorBindingStorageImageUpdateAfterBind
= true;
992 features
->descriptorBindingStorageBufferUpdateAfterBind
= true;
993 features
->descriptorBindingUniformTexelBufferUpdateAfterBind
= true;
994 features
->descriptorBindingStorageTexelBufferUpdateAfterBind
= true;
995 features
->descriptorBindingUpdateUnusedWhilePending
= true;
996 features
->descriptorBindingPartiallyBound
= true;
997 features
->descriptorBindingVariableDescriptorCount
= true;
998 features
->runtimeDescriptorArray
= true;
1001 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT
: {
1002 VkPhysicalDeviceConditionalRenderingFeaturesEXT
*features
=
1003 (VkPhysicalDeviceConditionalRenderingFeaturesEXT
*)ext
;
1004 features
->conditionalRendering
= true;
1005 features
->inheritedConditionalRendering
= false;
1008 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT
: {
1009 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*features
=
1010 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*)ext
;
1011 features
->vertexAttributeInstanceRateDivisor
= true;
1012 features
->vertexAttributeInstanceRateZeroDivisor
= true;
1015 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT
: {
1016 VkPhysicalDeviceTransformFeedbackFeaturesEXT
*features
=
1017 (VkPhysicalDeviceTransformFeedbackFeaturesEXT
*)ext
;
1018 features
->transformFeedback
= true;
1019 features
->geometryStreams
= !pdevice
->use_ngg_streamout
;
1022 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES
: {
1023 VkPhysicalDeviceScalarBlockLayoutFeatures
*features
=
1024 (VkPhysicalDeviceScalarBlockLayoutFeatures
*)ext
;
1025 features
->scalarBlockLayout
= pdevice
->rad_info
.chip_class
>= GFX7
;
1028 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_PRIORITY_FEATURES_EXT
: {
1029 VkPhysicalDeviceMemoryPriorityFeaturesEXT
*features
=
1030 (VkPhysicalDeviceMemoryPriorityFeaturesEXT
*)ext
;
1031 features
->memoryPriority
= true;
1034 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_EXT
: {
1035 VkPhysicalDeviceBufferDeviceAddressFeaturesEXT
*features
=
1036 (VkPhysicalDeviceBufferDeviceAddressFeaturesEXT
*)ext
;
1037 features
->bufferDeviceAddress
= true;
1038 features
->bufferDeviceAddressCaptureReplay
= false;
1039 features
->bufferDeviceAddressMultiDevice
= false;
1042 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES
: {
1043 VkPhysicalDeviceBufferDeviceAddressFeatures
*features
=
1044 (VkPhysicalDeviceBufferDeviceAddressFeatures
*)ext
;
1045 features
->bufferDeviceAddress
= true;
1046 features
->bufferDeviceAddressCaptureReplay
= false;
1047 features
->bufferDeviceAddressMultiDevice
= false;
1050 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_CLIP_ENABLE_FEATURES_EXT
: {
1051 VkPhysicalDeviceDepthClipEnableFeaturesEXT
*features
=
1052 (VkPhysicalDeviceDepthClipEnableFeaturesEXT
*)ext
;
1053 features
->depthClipEnable
= true;
1056 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_QUERY_RESET_FEATURES
: {
1057 VkPhysicalDeviceHostQueryResetFeatures
*features
=
1058 (VkPhysicalDeviceHostQueryResetFeatures
*)ext
;
1059 features
->hostQueryReset
= true;
1062 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES
: {
1063 VkPhysicalDevice8BitStorageFeatures
*features
=
1064 (VkPhysicalDevice8BitStorageFeatures
*)ext
;
1065 bool enable
= !pdevice
->use_aco
|| pdevice
->rad_info
.chip_class
>= GFX8
;
1066 features
->storageBuffer8BitAccess
= enable
;
1067 features
->uniformAndStorageBuffer8BitAccess
= enable
;
1068 features
->storagePushConstant8
= enable
;
1071 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_FLOAT16_INT8_FEATURES
: {
1072 VkPhysicalDeviceShaderFloat16Int8Features
*features
=
1073 (VkPhysicalDeviceShaderFloat16Int8Features
*)ext
;
1074 features
->shaderFloat16
= pdevice
->rad_info
.has_double_rate_fp16
&& !pdevice
->use_aco
;
1075 features
->shaderInt8
= !pdevice
->use_aco
|| pdevice
->rad_info
.chip_class
>= GFX8
;
1078 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_INT64_FEATURES
: {
1079 VkPhysicalDeviceShaderAtomicInt64Features
*features
=
1080 (VkPhysicalDeviceShaderAtomicInt64Features
*)ext
;
1081 features
->shaderBufferInt64Atomics
= LLVM_VERSION_MAJOR
>= 9;
1082 features
->shaderSharedInt64Atomics
= LLVM_VERSION_MAJOR
>= 9;
1085 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DEMOTE_TO_HELPER_INVOCATION_FEATURES_EXT
: {
1086 VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT
*features
=
1087 (VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT
*)ext
;
1088 features
->shaderDemoteToHelperInvocation
= pdevice
->use_aco
;
1091 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_FEATURES_EXT
: {
1092 VkPhysicalDeviceInlineUniformBlockFeaturesEXT
*features
=
1093 (VkPhysicalDeviceInlineUniformBlockFeaturesEXT
*)ext
;
1095 features
->inlineUniformBlock
= true;
1096 features
->descriptorBindingInlineUniformBlockUpdateAfterBind
= true;
1099 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COMPUTE_SHADER_DERIVATIVES_FEATURES_NV
: {
1100 VkPhysicalDeviceComputeShaderDerivativesFeaturesNV
*features
=
1101 (VkPhysicalDeviceComputeShaderDerivativesFeaturesNV
*)ext
;
1102 features
->computeDerivativeGroupQuads
= false;
1103 features
->computeDerivativeGroupLinear
= true;
1106 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_YCBCR_IMAGE_ARRAYS_FEATURES_EXT
: {
1107 VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*features
=
1108 (VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*)ext
;
1109 features
->ycbcrImageArrays
= true;
1112 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_UNIFORM_BUFFER_STANDARD_LAYOUT_FEATURES
: {
1113 VkPhysicalDeviceUniformBufferStandardLayoutFeatures
*features
=
1114 (VkPhysicalDeviceUniformBufferStandardLayoutFeatures
*)ext
;
1115 features
->uniformBufferStandardLayout
= true;
1118 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INDEX_TYPE_UINT8_FEATURES_EXT
: {
1119 VkPhysicalDeviceIndexTypeUint8FeaturesEXT
*features
=
1120 (VkPhysicalDeviceIndexTypeUint8FeaturesEXT
*)ext
;
1121 features
->indexTypeUint8
= pdevice
->rad_info
.chip_class
>= GFX8
;
1124 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGELESS_FRAMEBUFFER_FEATURES
: {
1125 VkPhysicalDeviceImagelessFramebufferFeatures
*features
=
1126 (VkPhysicalDeviceImagelessFramebufferFeatures
*)ext
;
1127 features
->imagelessFramebuffer
= true;
1130 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR
: {
1131 VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR
*features
=
1132 (VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR
*)ext
;
1133 features
->pipelineExecutableInfo
= true;
1136 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CLOCK_FEATURES_KHR
: {
1137 VkPhysicalDeviceShaderClockFeaturesKHR
*features
=
1138 (VkPhysicalDeviceShaderClockFeaturesKHR
*)ext
;
1139 features
->shaderSubgroupClock
= true;
1140 features
->shaderDeviceClock
= false;
1143 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_FEATURES_EXT
: {
1144 VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT
*features
=
1145 (VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT
*)ext
;
1146 features
->texelBufferAlignment
= true;
1149 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_FEATURES
: {
1150 VkPhysicalDeviceTimelineSemaphoreFeatures
*features
=
1151 (VkPhysicalDeviceTimelineSemaphoreFeatures
*) ext
;
1152 features
->timelineSemaphore
= true;
1155 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_FEATURES_EXT
: {
1156 VkPhysicalDeviceSubgroupSizeControlFeaturesEXT
*features
=
1157 (VkPhysicalDeviceSubgroupSizeControlFeaturesEXT
*)ext
;
1158 features
->subgroupSizeControl
= true;
1159 features
->computeFullSubgroups
= true;
1162 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COHERENT_MEMORY_FEATURES_AMD
: {
1163 VkPhysicalDeviceCoherentMemoryFeaturesAMD
*features
=
1164 (VkPhysicalDeviceCoherentMemoryFeaturesAMD
*)ext
;
1165 features
->deviceCoherentMemory
= pdevice
->rad_info
.has_l2_uncached
;
1168 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_SUBGROUP_EXTENDED_TYPES_FEATURES
: {
1169 VkPhysicalDeviceShaderSubgroupExtendedTypesFeatures
*features
=
1170 (VkPhysicalDeviceShaderSubgroupExtendedTypesFeatures
*)ext
;
1171 features
->shaderSubgroupExtendedTypes
= !pdevice
->use_aco
;
1174 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SEPARATE_DEPTH_STENCIL_LAYOUTS_FEATURES_KHR
: {
1175 VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR
*features
=
1176 (VkPhysicalDeviceSeparateDepthStencilLayoutsFeaturesKHR
*)ext
;
1177 features
->separateDepthStencilLayouts
= true;
1180 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES
: {
1181 VkPhysicalDeviceVulkan11Features
*features
=
1182 (VkPhysicalDeviceVulkan11Features
*)ext
;
1183 bool storage16_enable
= !pdevice
->use_aco
|| pdevice
->rad_info
.chip_class
>= GFX8
;
1184 features
->storageBuffer16BitAccess
= storage16_enable
;
1185 features
->uniformAndStorageBuffer16BitAccess
= storage16_enable
;
1186 features
->storagePushConstant16
= storage16_enable
;
1187 features
->storageInputOutput16
= pdevice
->rad_info
.has_double_rate_fp16
&& !pdevice
->use_aco
&& LLVM_VERSION_MAJOR
>= 9;
1188 features
->multiview
= true;
1189 features
->multiviewGeometryShader
= true;
1190 features
->multiviewTessellationShader
= true;
1191 features
->variablePointersStorageBuffer
= true;
1192 features
->variablePointers
= true;
1193 features
->protectedMemory
= false;
1194 features
->samplerYcbcrConversion
= true;
1195 features
->shaderDrawParameters
= true;
1198 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES
: {
1199 VkPhysicalDeviceVulkan12Features
*features
=
1200 (VkPhysicalDeviceVulkan12Features
*)ext
;
1201 bool int8_enable
= !pdevice
->use_aco
|| pdevice
->rad_info
.chip_class
>= GFX8
;
1202 features
->samplerMirrorClampToEdge
= true;
1203 features
->drawIndirectCount
= true;
1204 features
->storageBuffer8BitAccess
= int8_enable
;
1205 features
->uniformAndStorageBuffer8BitAccess
= int8_enable
;
1206 features
->storagePushConstant8
= int8_enable
;
1207 features
->shaderBufferInt64Atomics
= LLVM_VERSION_MAJOR
>= 9;
1208 features
->shaderSharedInt64Atomics
= LLVM_VERSION_MAJOR
>= 9;
1209 features
->shaderFloat16
= pdevice
->rad_info
.has_double_rate_fp16
&& !pdevice
->use_aco
;
1210 features
->shaderInt8
= int8_enable
;
1211 features
->descriptorIndexing
= true;
1212 features
->shaderInputAttachmentArrayDynamicIndexing
= true;
1213 features
->shaderUniformTexelBufferArrayDynamicIndexing
= true;
1214 features
->shaderStorageTexelBufferArrayDynamicIndexing
= true;
1215 features
->shaderUniformBufferArrayNonUniformIndexing
= true;
1216 features
->shaderSampledImageArrayNonUniformIndexing
= true;
1217 features
->shaderStorageBufferArrayNonUniformIndexing
= true;
1218 features
->shaderStorageImageArrayNonUniformIndexing
= true;
1219 features
->shaderInputAttachmentArrayNonUniformIndexing
= true;
1220 features
->shaderUniformTexelBufferArrayNonUniformIndexing
= true;
1221 features
->shaderStorageTexelBufferArrayNonUniformIndexing
= true;
1222 features
->descriptorBindingUniformBufferUpdateAfterBind
= true;
1223 features
->descriptorBindingSampledImageUpdateAfterBind
= true;
1224 features
->descriptorBindingStorageImageUpdateAfterBind
= true;
1225 features
->descriptorBindingStorageBufferUpdateAfterBind
= true;
1226 features
->descriptorBindingUniformTexelBufferUpdateAfterBind
= true;
1227 features
->descriptorBindingStorageTexelBufferUpdateAfterBind
= true;
1228 features
->descriptorBindingUpdateUnusedWhilePending
= true;
1229 features
->descriptorBindingPartiallyBound
= true;
1230 features
->descriptorBindingVariableDescriptorCount
= true;
1231 features
->runtimeDescriptorArray
= true;
1232 features
->samplerFilterMinmax
= true;
1233 features
->scalarBlockLayout
= pdevice
->rad_info
.chip_class
>= GFX7
;
1234 features
->imagelessFramebuffer
= true;
1235 features
->uniformBufferStandardLayout
= true;
1236 features
->shaderSubgroupExtendedTypes
= !pdevice
->use_aco
;
1237 features
->separateDepthStencilLayouts
= true;
1238 features
->hostQueryReset
= true;
1239 features
->timelineSemaphore
= pdevice
->rad_info
.has_syncobj_wait_for_submit
;
1240 features
->bufferDeviceAddress
= true;
1241 features
->bufferDeviceAddressCaptureReplay
= false;
1242 features
->bufferDeviceAddressMultiDevice
= false;
1243 features
->vulkanMemoryModel
= false;
1244 features
->vulkanMemoryModelDeviceScope
= false;
1245 features
->vulkanMemoryModelAvailabilityVisibilityChains
= false;
1246 features
->shaderOutputViewportIndex
= true;
1247 features
->shaderOutputLayer
= true;
1248 features
->subgroupBroadcastDynamicId
= true;
1251 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_FEATURES_EXT
: {
1252 VkPhysicalDeviceLineRasterizationFeaturesEXT
*features
=
1253 (VkPhysicalDeviceLineRasterizationFeaturesEXT
*)ext
;
1254 features
->rectangularLines
= false;
1255 features
->bresenhamLines
= true;
1256 features
->smoothLines
= false;
1257 features
->stippledRectangularLines
= false;
1258 features
->stippledBresenhamLines
= true;
1259 features
->stippledSmoothLines
= false;
1262 case VK_STRUCTURE_TYPE_DEVICE_MEMORY_OVERALLOCATION_CREATE_INFO_AMD
: {
1263 VkDeviceMemoryOverallocationCreateInfoAMD
*features
=
1264 (VkDeviceMemoryOverallocationCreateInfoAMD
*)ext
;
1265 features
->overallocationBehavior
= true;
1268 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ROBUSTNESS_2_FEATURES_EXT
: {
1269 VkPhysicalDeviceRobustness2FeaturesEXT
*features
=
1270 (VkPhysicalDeviceRobustness2FeaturesEXT
*)ext
;
1271 features
->robustBufferAccess2
= true;
1272 features
->robustImageAccess2
= true;
1273 features
->nullDescriptor
= true;
1280 return radv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
1284 radv_max_descriptor_set_size()
1286 /* make sure that the entire descriptor set is addressable with a signed
1287 * 32-bit int. So the sum of all limits scaled by descriptor size has to
1288 * be at most 2 GiB. the combined image & samples object count as one of
1289 * both. This limit is for the pipeline layout, not for the set layout, but
1290 * there is no set limit, so we just set a pipeline limit. I don't think
1291 * any app is going to hit this soon. */
1292 return ((1ull << 31) - 16 * MAX_DYNAMIC_BUFFERS
1293 - MAX_INLINE_UNIFORM_BLOCK_SIZE
* MAX_INLINE_UNIFORM_BLOCK_COUNT
) /
1294 (32 /* uniform buffer, 32 due to potential space wasted on alignment */ +
1295 32 /* storage buffer, 32 due to potential space wasted on alignment */ +
1296 32 /* sampler, largest when combined with image */ +
1297 64 /* sampled image */ +
1298 64 /* storage image */);
1301 void radv_GetPhysicalDeviceProperties(
1302 VkPhysicalDevice physicalDevice
,
1303 VkPhysicalDeviceProperties
* pProperties
)
1305 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
1306 VkSampleCountFlags sample_counts
= 0xf;
1308 size_t max_descriptor_set_size
= radv_max_descriptor_set_size();
1310 VkPhysicalDeviceLimits limits
= {
1311 .maxImageDimension1D
= (1 << 14),
1312 .maxImageDimension2D
= (1 << 14),
1313 .maxImageDimension3D
= (1 << 11),
1314 .maxImageDimensionCube
= (1 << 14),
1315 .maxImageArrayLayers
= (1 << 11),
1316 .maxTexelBufferElements
= UINT32_MAX
,
1317 .maxUniformBufferRange
= UINT32_MAX
,
1318 .maxStorageBufferRange
= UINT32_MAX
,
1319 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
1320 .maxMemoryAllocationCount
= UINT32_MAX
,
1321 .maxSamplerAllocationCount
= 64 * 1024,
1322 .bufferImageGranularity
= 64, /* A cache line */
1323 .sparseAddressSpaceSize
= RADV_MAX_MEMORY_ALLOCATION_SIZE
, /* buffer max size */
1324 .maxBoundDescriptorSets
= MAX_SETS
,
1325 .maxPerStageDescriptorSamplers
= max_descriptor_set_size
,
1326 .maxPerStageDescriptorUniformBuffers
= max_descriptor_set_size
,
1327 .maxPerStageDescriptorStorageBuffers
= max_descriptor_set_size
,
1328 .maxPerStageDescriptorSampledImages
= max_descriptor_set_size
,
1329 .maxPerStageDescriptorStorageImages
= max_descriptor_set_size
,
1330 .maxPerStageDescriptorInputAttachments
= max_descriptor_set_size
,
1331 .maxPerStageResources
= max_descriptor_set_size
,
1332 .maxDescriptorSetSamplers
= max_descriptor_set_size
,
1333 .maxDescriptorSetUniformBuffers
= max_descriptor_set_size
,
1334 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_UNIFORM_BUFFERS
,
1335 .maxDescriptorSetStorageBuffers
= max_descriptor_set_size
,
1336 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_STORAGE_BUFFERS
,
1337 .maxDescriptorSetSampledImages
= max_descriptor_set_size
,
1338 .maxDescriptorSetStorageImages
= max_descriptor_set_size
,
1339 .maxDescriptorSetInputAttachments
= max_descriptor_set_size
,
1340 .maxVertexInputAttributes
= MAX_VERTEX_ATTRIBS
,
1341 .maxVertexInputBindings
= MAX_VBS
,
1342 .maxVertexInputAttributeOffset
= 2047,
1343 .maxVertexInputBindingStride
= 2048,
1344 .maxVertexOutputComponents
= 128,
1345 .maxTessellationGenerationLevel
= 64,
1346 .maxTessellationPatchSize
= 32,
1347 .maxTessellationControlPerVertexInputComponents
= 128,
1348 .maxTessellationControlPerVertexOutputComponents
= 128,
1349 .maxTessellationControlPerPatchOutputComponents
= 120,
1350 .maxTessellationControlTotalOutputComponents
= 4096,
1351 .maxTessellationEvaluationInputComponents
= 128,
1352 .maxTessellationEvaluationOutputComponents
= 128,
1353 .maxGeometryShaderInvocations
= 127,
1354 .maxGeometryInputComponents
= 64,
1355 .maxGeometryOutputComponents
= 128,
1356 .maxGeometryOutputVertices
= 256,
1357 .maxGeometryTotalOutputComponents
= 1024,
1358 .maxFragmentInputComponents
= 128,
1359 .maxFragmentOutputAttachments
= 8,
1360 .maxFragmentDualSrcAttachments
= 1,
1361 .maxFragmentCombinedOutputResources
= 8,
1362 .maxComputeSharedMemorySize
= 32768,
1363 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
1364 .maxComputeWorkGroupInvocations
= 1024,
1365 .maxComputeWorkGroupSize
= {
1370 .subPixelPrecisionBits
= 8,
1371 .subTexelPrecisionBits
= 8,
1372 .mipmapPrecisionBits
= 8,
1373 .maxDrawIndexedIndexValue
= UINT32_MAX
,
1374 .maxDrawIndirectCount
= UINT32_MAX
,
1375 .maxSamplerLodBias
= 16,
1376 .maxSamplerAnisotropy
= 16,
1377 .maxViewports
= MAX_VIEWPORTS
,
1378 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
1379 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
1380 .viewportSubPixelBits
= 8,
1381 .minMemoryMapAlignment
= 4096, /* A page */
1382 .minTexelBufferOffsetAlignment
= 4,
1383 .minUniformBufferOffsetAlignment
= 4,
1384 .minStorageBufferOffsetAlignment
= 4,
1385 .minTexelOffset
= -32,
1386 .maxTexelOffset
= 31,
1387 .minTexelGatherOffset
= -32,
1388 .maxTexelGatherOffset
= 31,
1389 .minInterpolationOffset
= -2,
1390 .maxInterpolationOffset
= 2,
1391 .subPixelInterpolationOffsetBits
= 8,
1392 .maxFramebufferWidth
= (1 << 14),
1393 .maxFramebufferHeight
= (1 << 14),
1394 .maxFramebufferLayers
= (1 << 10),
1395 .framebufferColorSampleCounts
= sample_counts
,
1396 .framebufferDepthSampleCounts
= sample_counts
,
1397 .framebufferStencilSampleCounts
= sample_counts
,
1398 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
1399 .maxColorAttachments
= MAX_RTS
,
1400 .sampledImageColorSampleCounts
= sample_counts
,
1401 .sampledImageIntegerSampleCounts
= sample_counts
,
1402 .sampledImageDepthSampleCounts
= sample_counts
,
1403 .sampledImageStencilSampleCounts
= sample_counts
,
1404 .storageImageSampleCounts
= sample_counts
,
1405 .maxSampleMaskWords
= 1,
1406 .timestampComputeAndGraphics
= true,
1407 .timestampPeriod
= 1000000.0 / pdevice
->rad_info
.clock_crystal_freq
,
1408 .maxClipDistances
= 8,
1409 .maxCullDistances
= 8,
1410 .maxCombinedClipAndCullDistances
= 8,
1411 .discreteQueuePriorities
= 2,
1412 .pointSizeRange
= { 0.0, 8192.0 },
1413 .lineWidthRange
= { 0.0, 8192.0 },
1414 .pointSizeGranularity
= (1.0 / 8.0),
1415 .lineWidthGranularity
= (1.0 / 8.0),
1416 .strictLines
= false, /* FINISHME */
1417 .standardSampleLocations
= true,
1418 .optimalBufferCopyOffsetAlignment
= 128,
1419 .optimalBufferCopyRowPitchAlignment
= 128,
1420 .nonCoherentAtomSize
= 64,
1423 *pProperties
= (VkPhysicalDeviceProperties
) {
1424 .apiVersion
= radv_physical_device_api_version(pdevice
),
1425 .driverVersion
= vk_get_driver_version(),
1426 .vendorID
= ATI_VENDOR_ID
,
1427 .deviceID
= pdevice
->rad_info
.pci_id
,
1428 .deviceType
= pdevice
->rad_info
.has_dedicated_vram
? VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU
: VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
1430 .sparseProperties
= {0},
1433 strcpy(pProperties
->deviceName
, pdevice
->name
);
1434 memcpy(pProperties
->pipelineCacheUUID
, pdevice
->cache_uuid
, VK_UUID_SIZE
);
1438 radv_get_physical_device_properties_1_1(struct radv_physical_device
*pdevice
,
1439 VkPhysicalDeviceVulkan11Properties
*p
)
1441 assert(p
->sType
== VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES
);
1443 memcpy(p
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
1444 memcpy(p
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
1445 memset(p
->deviceLUID
, 0, VK_LUID_SIZE
);
1446 /* The LUID is for Windows. */
1447 p
->deviceLUIDValid
= false;
1448 p
->deviceNodeMask
= 0;
1450 p
->subgroupSize
= RADV_SUBGROUP_SIZE
;
1451 p
->subgroupSupportedStages
= VK_SHADER_STAGE_ALL_GRAPHICS
|
1452 VK_SHADER_STAGE_COMPUTE_BIT
;
1453 p
->subgroupSupportedOperations
= VK_SUBGROUP_FEATURE_BASIC_BIT
|
1454 VK_SUBGROUP_FEATURE_VOTE_BIT
|
1455 VK_SUBGROUP_FEATURE_ARITHMETIC_BIT
|
1456 VK_SUBGROUP_FEATURE_BALLOT_BIT
|
1457 VK_SUBGROUP_FEATURE_CLUSTERED_BIT
|
1458 VK_SUBGROUP_FEATURE_QUAD_BIT
;
1460 if (((pdevice
->rad_info
.chip_class
== GFX6
||
1461 pdevice
->rad_info
.chip_class
== GFX7
) && !pdevice
->use_aco
) ||
1462 pdevice
->rad_info
.chip_class
>= GFX8
) {
1463 p
->subgroupSupportedOperations
|= VK_SUBGROUP_FEATURE_SHUFFLE_BIT
|
1464 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT
;
1466 p
->subgroupQuadOperationsInAllStages
= true;
1468 p
->pointClippingBehavior
= VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES
;
1469 p
->maxMultiviewViewCount
= MAX_VIEWS
;
1470 p
->maxMultiviewInstanceIndex
= INT_MAX
;
1471 p
->protectedNoFault
= false;
1472 p
->maxPerSetDescriptors
= RADV_MAX_PER_SET_DESCRIPTORS
;
1473 p
->maxMemoryAllocationSize
= RADV_MAX_MEMORY_ALLOCATION_SIZE
;
1477 radv_get_physical_device_properties_1_2(struct radv_physical_device
*pdevice
,
1478 VkPhysicalDeviceVulkan12Properties
*p
)
1480 assert(p
->sType
== VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES
);
1482 p
->driverID
= VK_DRIVER_ID_MESA_RADV
;
1483 snprintf(p
->driverName
, VK_MAX_DRIVER_NAME_SIZE
, "radv");
1484 snprintf(p
->driverInfo
, VK_MAX_DRIVER_INFO_SIZE
,
1485 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1
" (%s)",
1486 pdevice
->use_aco
? "ACO" : "LLVM " MESA_LLVM_VERSION_STRING
);
1487 p
->conformanceVersion
= (VkConformanceVersion
) {
1494 /* On AMD hardware, denormals and rounding modes for fp16/fp64 are
1495 * controlled by the same config register.
1497 if (pdevice
->rad_info
.has_double_rate_fp16
) {
1498 p
->denormBehaviorIndependence
= VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_32_BIT_ONLY_KHR
;
1499 p
->roundingModeIndependence
= VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_32_BIT_ONLY_KHR
;
1501 p
->denormBehaviorIndependence
= VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_ALL_KHR
;
1502 p
->roundingModeIndependence
= VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_ALL_KHR
;
1505 /* Do not allow both preserving and flushing denorms because different
1506 * shaders in the same pipeline can have different settings and this
1507 * won't work for merged shaders. To make it work, this requires LLVM
1508 * support for changing the register. The same logic applies for the
1509 * rounding modes because they are configured with the same config
1510 * register. TODO: we can enable a lot of these for ACO when it
1511 * supports all stages.
1513 p
->shaderDenormFlushToZeroFloat32
= true;
1514 p
->shaderDenormPreserveFloat32
= false;
1515 p
->shaderRoundingModeRTEFloat32
= true;
1516 p
->shaderRoundingModeRTZFloat32
= false;
1517 p
->shaderSignedZeroInfNanPreserveFloat32
= true;
1519 p
->shaderDenormFlushToZeroFloat16
= false;
1520 p
->shaderDenormPreserveFloat16
= pdevice
->rad_info
.has_double_rate_fp16
;
1521 p
->shaderRoundingModeRTEFloat16
= pdevice
->rad_info
.has_double_rate_fp16
;
1522 p
->shaderRoundingModeRTZFloat16
= false;
1523 p
->shaderSignedZeroInfNanPreserveFloat16
= pdevice
->rad_info
.has_double_rate_fp16
;
1525 p
->shaderDenormFlushToZeroFloat64
= false;
1526 p
->shaderDenormPreserveFloat64
= pdevice
->rad_info
.chip_class
>= GFX8
;
1527 p
->shaderRoundingModeRTEFloat64
= pdevice
->rad_info
.chip_class
>= GFX8
;
1528 p
->shaderRoundingModeRTZFloat64
= false;
1529 p
->shaderSignedZeroInfNanPreserveFloat64
= pdevice
->rad_info
.chip_class
>= GFX8
;
1531 p
->maxUpdateAfterBindDescriptorsInAllPools
= UINT32_MAX
/ 64;
1532 p
->shaderUniformBufferArrayNonUniformIndexingNative
= false;
1533 p
->shaderSampledImageArrayNonUniformIndexingNative
= false;
1534 p
->shaderStorageBufferArrayNonUniformIndexingNative
= false;
1535 p
->shaderStorageImageArrayNonUniformIndexingNative
= false;
1536 p
->shaderInputAttachmentArrayNonUniformIndexingNative
= false;
1537 p
->robustBufferAccessUpdateAfterBind
= false;
1538 p
->quadDivergentImplicitLod
= false;
1540 size_t max_descriptor_set_size
= ((1ull << 31) - 16 * MAX_DYNAMIC_BUFFERS
-
1541 MAX_INLINE_UNIFORM_BLOCK_SIZE
* MAX_INLINE_UNIFORM_BLOCK_COUNT
) /
1542 (32 /* uniform buffer, 32 due to potential space wasted on alignment */ +
1543 32 /* storage buffer, 32 due to potential space wasted on alignment */ +
1544 32 /* sampler, largest when combined with image */ +
1545 64 /* sampled image */ +
1546 64 /* storage image */);
1547 p
->maxPerStageDescriptorUpdateAfterBindSamplers
= max_descriptor_set_size
;
1548 p
->maxPerStageDescriptorUpdateAfterBindUniformBuffers
= max_descriptor_set_size
;
1549 p
->maxPerStageDescriptorUpdateAfterBindStorageBuffers
= max_descriptor_set_size
;
1550 p
->maxPerStageDescriptorUpdateAfterBindSampledImages
= max_descriptor_set_size
;
1551 p
->maxPerStageDescriptorUpdateAfterBindStorageImages
= max_descriptor_set_size
;
1552 p
->maxPerStageDescriptorUpdateAfterBindInputAttachments
= max_descriptor_set_size
;
1553 p
->maxPerStageUpdateAfterBindResources
= max_descriptor_set_size
;
1554 p
->maxDescriptorSetUpdateAfterBindSamplers
= max_descriptor_set_size
;
1555 p
->maxDescriptorSetUpdateAfterBindUniformBuffers
= max_descriptor_set_size
;
1556 p
->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic
= MAX_DYNAMIC_UNIFORM_BUFFERS
;
1557 p
->maxDescriptorSetUpdateAfterBindStorageBuffers
= max_descriptor_set_size
;
1558 p
->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic
= MAX_DYNAMIC_STORAGE_BUFFERS
;
1559 p
->maxDescriptorSetUpdateAfterBindSampledImages
= max_descriptor_set_size
;
1560 p
->maxDescriptorSetUpdateAfterBindStorageImages
= max_descriptor_set_size
;
1561 p
->maxDescriptorSetUpdateAfterBindInputAttachments
= max_descriptor_set_size
;
1563 /* We support all of the depth resolve modes */
1564 p
->supportedDepthResolveModes
= VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
|
1565 VK_RESOLVE_MODE_AVERAGE_BIT_KHR
|
1566 VK_RESOLVE_MODE_MIN_BIT_KHR
|
1567 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1569 /* Average doesn't make sense for stencil so we don't support that */
1570 p
->supportedStencilResolveModes
= VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
|
1571 VK_RESOLVE_MODE_MIN_BIT_KHR
|
1572 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1574 p
->independentResolveNone
= true;
1575 p
->independentResolve
= true;
1577 /* GFX6-8 only support single channel min/max filter. */
1578 p
->filterMinmaxImageComponentMapping
= pdevice
->rad_info
.chip_class
>= GFX9
;
1579 p
->filterMinmaxSingleComponentFormats
= true;
1581 p
->maxTimelineSemaphoreValueDifference
= UINT64_MAX
;
1583 p
->framebufferIntegerColorSampleCounts
= VK_SAMPLE_COUNT_1_BIT
;
1586 void radv_GetPhysicalDeviceProperties2(
1587 VkPhysicalDevice physicalDevice
,
1588 VkPhysicalDeviceProperties2
*pProperties
)
1590 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
1591 radv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
1593 VkPhysicalDeviceVulkan11Properties core_1_1
= {
1594 .sType
= VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES
,
1596 radv_get_physical_device_properties_1_1(pdevice
, &core_1_1
);
1598 VkPhysicalDeviceVulkan12Properties core_1_2
= {
1599 .sType
= VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES
,
1601 radv_get_physical_device_properties_1_2(pdevice
, &core_1_2
);
1603 #define CORE_RENAMED_PROPERTY(major, minor, ext_property, core_property) \
1604 memcpy(&properties->ext_property, &core_##major##_##minor.core_property, \
1605 sizeof(core_##major##_##minor.core_property))
1607 #define CORE_PROPERTY(major, minor, property) \
1608 CORE_RENAMED_PROPERTY(major, minor, property, property)
1610 vk_foreach_struct(ext
, pProperties
->pNext
) {
1611 switch (ext
->sType
) {
1612 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
1613 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
1614 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
1615 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
1618 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES
: {
1619 VkPhysicalDeviceIDProperties
*properties
= (VkPhysicalDeviceIDProperties
*)ext
;
1620 CORE_PROPERTY(1, 1, deviceUUID
);
1621 CORE_PROPERTY(1, 1, driverUUID
);
1622 CORE_PROPERTY(1, 1, deviceLUID
);
1623 CORE_PROPERTY(1, 1, deviceLUIDValid
);
1626 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES
: {
1627 VkPhysicalDeviceMultiviewProperties
*properties
= (VkPhysicalDeviceMultiviewProperties
*)ext
;
1628 CORE_PROPERTY(1, 1, maxMultiviewViewCount
);
1629 CORE_PROPERTY(1, 1, maxMultiviewInstanceIndex
);
1632 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES
: {
1633 VkPhysicalDevicePointClippingProperties
*properties
=
1634 (VkPhysicalDevicePointClippingProperties
*)ext
;
1635 CORE_PROPERTY(1, 1, pointClippingBehavior
);
1638 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DISCARD_RECTANGLE_PROPERTIES_EXT
: {
1639 VkPhysicalDeviceDiscardRectanglePropertiesEXT
*properties
=
1640 (VkPhysicalDeviceDiscardRectanglePropertiesEXT
*)ext
;
1641 properties
->maxDiscardRectangles
= MAX_DISCARD_RECTANGLES
;
1644 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT
: {
1645 VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*properties
=
1646 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*) ext
;
1647 properties
->minImportedHostPointerAlignment
= 4096;
1650 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES
: {
1651 VkPhysicalDeviceSubgroupProperties
*properties
=
1652 (VkPhysicalDeviceSubgroupProperties
*)ext
;
1653 CORE_PROPERTY(1, 1, subgroupSize
);
1654 CORE_RENAMED_PROPERTY(1, 1, supportedStages
,
1655 subgroupSupportedStages
);
1656 CORE_RENAMED_PROPERTY(1, 1, supportedOperations
,
1657 subgroupSupportedOperations
);
1658 CORE_RENAMED_PROPERTY(1, 1, quadOperationsInAllStages
,
1659 subgroupQuadOperationsInAllStages
);
1662 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES
: {
1663 VkPhysicalDeviceMaintenance3Properties
*properties
=
1664 (VkPhysicalDeviceMaintenance3Properties
*)ext
;
1665 CORE_PROPERTY(1, 1, maxPerSetDescriptors
);
1666 CORE_PROPERTY(1, 1, maxMemoryAllocationSize
);
1669 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES
: {
1670 VkPhysicalDeviceSamplerFilterMinmaxProperties
*properties
=
1671 (VkPhysicalDeviceSamplerFilterMinmaxProperties
*)ext
;
1672 CORE_PROPERTY(1, 2, filterMinmaxImageComponentMapping
);
1673 CORE_PROPERTY(1, 2, filterMinmaxSingleComponentFormats
);
1676 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CORE_PROPERTIES_AMD
: {
1677 VkPhysicalDeviceShaderCorePropertiesAMD
*properties
=
1678 (VkPhysicalDeviceShaderCorePropertiesAMD
*)ext
;
1680 /* Shader engines. */
1681 properties
->shaderEngineCount
=
1682 pdevice
->rad_info
.max_se
;
1683 properties
->shaderArraysPerEngineCount
=
1684 pdevice
->rad_info
.max_sh_per_se
;
1685 properties
->computeUnitsPerShaderArray
=
1686 pdevice
->rad_info
.num_good_cu_per_sh
;
1687 properties
->simdPerComputeUnit
=
1688 pdevice
->rad_info
.num_simd_per_compute_unit
;
1689 properties
->wavefrontsPerSimd
=
1690 pdevice
->rad_info
.max_wave64_per_simd
;
1691 properties
->wavefrontSize
= 64;
1694 properties
->sgprsPerSimd
=
1695 pdevice
->rad_info
.num_physical_sgprs_per_simd
;
1696 properties
->minSgprAllocation
=
1697 pdevice
->rad_info
.min_sgpr_alloc
;
1698 properties
->maxSgprAllocation
=
1699 pdevice
->rad_info
.max_sgpr_alloc
;
1700 properties
->sgprAllocationGranularity
=
1701 pdevice
->rad_info
.sgpr_alloc_granularity
;
1704 properties
->vgprsPerSimd
=
1705 pdevice
->rad_info
.num_physical_wave64_vgprs_per_simd
;
1706 properties
->minVgprAllocation
=
1707 pdevice
->rad_info
.min_wave64_vgpr_alloc
;
1708 properties
->maxVgprAllocation
=
1709 pdevice
->rad_info
.max_vgpr_alloc
;
1710 properties
->vgprAllocationGranularity
=
1711 pdevice
->rad_info
.wave64_vgpr_alloc_granularity
;
1714 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CORE_PROPERTIES_2_AMD
: {
1715 VkPhysicalDeviceShaderCoreProperties2AMD
*properties
=
1716 (VkPhysicalDeviceShaderCoreProperties2AMD
*)ext
;
1718 properties
->shaderCoreFeatures
= 0;
1719 properties
->activeComputeUnitCount
=
1720 pdevice
->rad_info
.num_good_compute_units
;
1723 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT
: {
1724 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*properties
=
1725 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*)ext
;
1726 properties
->maxVertexAttribDivisor
= UINT32_MAX
;
1729 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES
: {
1730 VkPhysicalDeviceDescriptorIndexingProperties
*properties
=
1731 (VkPhysicalDeviceDescriptorIndexingProperties
*)ext
;
1732 CORE_PROPERTY(1, 2, maxUpdateAfterBindDescriptorsInAllPools
);
1733 CORE_PROPERTY(1, 2, shaderUniformBufferArrayNonUniformIndexingNative
);
1734 CORE_PROPERTY(1, 2, shaderSampledImageArrayNonUniformIndexingNative
);
1735 CORE_PROPERTY(1, 2, shaderStorageBufferArrayNonUniformIndexingNative
);
1736 CORE_PROPERTY(1, 2, shaderStorageImageArrayNonUniformIndexingNative
);
1737 CORE_PROPERTY(1, 2, shaderInputAttachmentArrayNonUniformIndexingNative
);
1738 CORE_PROPERTY(1, 2, robustBufferAccessUpdateAfterBind
);
1739 CORE_PROPERTY(1, 2, quadDivergentImplicitLod
);
1740 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindSamplers
);
1741 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindUniformBuffers
);
1742 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindStorageBuffers
);
1743 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindSampledImages
);
1744 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindStorageImages
);
1745 CORE_PROPERTY(1, 2, maxPerStageDescriptorUpdateAfterBindInputAttachments
);
1746 CORE_PROPERTY(1, 2, maxPerStageUpdateAfterBindResources
);
1747 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindSamplers
);
1748 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindUniformBuffers
);
1749 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindUniformBuffersDynamic
);
1750 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindStorageBuffers
);
1751 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindStorageBuffersDynamic
);
1752 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindSampledImages
);
1753 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindStorageImages
);
1754 CORE_PROPERTY(1, 2, maxDescriptorSetUpdateAfterBindInputAttachments
);
1757 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES
: {
1758 VkPhysicalDeviceProtectedMemoryProperties
*properties
=
1759 (VkPhysicalDeviceProtectedMemoryProperties
*)ext
;
1760 CORE_PROPERTY(1, 1, protectedNoFault
);
1763 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONSERVATIVE_RASTERIZATION_PROPERTIES_EXT
: {
1764 VkPhysicalDeviceConservativeRasterizationPropertiesEXT
*properties
=
1765 (VkPhysicalDeviceConservativeRasterizationPropertiesEXT
*)ext
;
1766 properties
->primitiveOverestimationSize
= 0;
1767 properties
->maxExtraPrimitiveOverestimationSize
= 0;
1768 properties
->extraPrimitiveOverestimationSizeGranularity
= 0;
1769 properties
->primitiveUnderestimation
= false;
1770 properties
->conservativePointAndLineRasterization
= false;
1771 properties
->degenerateTrianglesRasterized
= false;
1772 properties
->degenerateLinesRasterized
= false;
1773 properties
->fullyCoveredFragmentShaderInputVariable
= false;
1774 properties
->conservativeRasterizationPostDepthCoverage
= false;
1777 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT
: {
1778 VkPhysicalDevicePCIBusInfoPropertiesEXT
*properties
=
1779 (VkPhysicalDevicePCIBusInfoPropertiesEXT
*)ext
;
1780 properties
->pciDomain
= pdevice
->bus_info
.domain
;
1781 properties
->pciBus
= pdevice
->bus_info
.bus
;
1782 properties
->pciDevice
= pdevice
->bus_info
.dev
;
1783 properties
->pciFunction
= pdevice
->bus_info
.func
;
1786 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES
: {
1787 VkPhysicalDeviceDriverProperties
*properties
=
1788 (VkPhysicalDeviceDriverProperties
*) ext
;
1789 CORE_PROPERTY(1, 2, driverID
);
1790 CORE_PROPERTY(1, 2, driverName
);
1791 CORE_PROPERTY(1, 2, driverInfo
);
1792 CORE_PROPERTY(1, 2, conformanceVersion
);
1795 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT
: {
1796 VkPhysicalDeviceTransformFeedbackPropertiesEXT
*properties
=
1797 (VkPhysicalDeviceTransformFeedbackPropertiesEXT
*)ext
;
1798 properties
->maxTransformFeedbackStreams
= MAX_SO_STREAMS
;
1799 properties
->maxTransformFeedbackBuffers
= MAX_SO_BUFFERS
;
1800 properties
->maxTransformFeedbackBufferSize
= UINT32_MAX
;
1801 properties
->maxTransformFeedbackStreamDataSize
= 512;
1802 properties
->maxTransformFeedbackBufferDataSize
= UINT32_MAX
;
1803 properties
->maxTransformFeedbackBufferDataStride
= 512;
1804 properties
->transformFeedbackQueries
= !pdevice
->use_ngg_streamout
;
1805 properties
->transformFeedbackStreamsLinesTriangles
= !pdevice
->use_ngg_streamout
;
1806 properties
->transformFeedbackRasterizationStreamSelect
= false;
1807 properties
->transformFeedbackDraw
= true;
1810 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_PROPERTIES_EXT
: {
1811 VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*props
=
1812 (VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*)ext
;
1814 props
->maxInlineUniformBlockSize
= MAX_INLINE_UNIFORM_BLOCK_SIZE
;
1815 props
->maxPerStageDescriptorInlineUniformBlocks
= MAX_INLINE_UNIFORM_BLOCK_SIZE
* MAX_SETS
;
1816 props
->maxPerStageDescriptorUpdateAfterBindInlineUniformBlocks
= MAX_INLINE_UNIFORM_BLOCK_SIZE
* MAX_SETS
;
1817 props
->maxDescriptorSetInlineUniformBlocks
= MAX_INLINE_UNIFORM_BLOCK_COUNT
;
1818 props
->maxDescriptorSetUpdateAfterBindInlineUniformBlocks
= MAX_INLINE_UNIFORM_BLOCK_COUNT
;
1821 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLE_LOCATIONS_PROPERTIES_EXT
: {
1822 VkPhysicalDeviceSampleLocationsPropertiesEXT
*properties
=
1823 (VkPhysicalDeviceSampleLocationsPropertiesEXT
*)ext
;
1824 properties
->sampleLocationSampleCounts
= VK_SAMPLE_COUNT_2_BIT
|
1825 VK_SAMPLE_COUNT_4_BIT
|
1826 VK_SAMPLE_COUNT_8_BIT
;
1827 properties
->maxSampleLocationGridSize
= (VkExtent2D
){ 2 , 2 };
1828 properties
->sampleLocationCoordinateRange
[0] = 0.0f
;
1829 properties
->sampleLocationCoordinateRange
[1] = 0.9375f
;
1830 properties
->sampleLocationSubPixelBits
= 4;
1831 properties
->variableSampleLocations
= false;
1834 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_STENCIL_RESOLVE_PROPERTIES
: {
1835 VkPhysicalDeviceDepthStencilResolveProperties
*properties
=
1836 (VkPhysicalDeviceDepthStencilResolveProperties
*)ext
;
1837 CORE_PROPERTY(1, 2, supportedDepthResolveModes
);
1838 CORE_PROPERTY(1, 2, supportedStencilResolveModes
);
1839 CORE_PROPERTY(1, 2, independentResolveNone
);
1840 CORE_PROPERTY(1, 2, independentResolve
);
1843 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_PROPERTIES_EXT
: {
1844 VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT
*properties
=
1845 (VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT
*)ext
;
1846 properties
->storageTexelBufferOffsetAlignmentBytes
= 4;
1847 properties
->storageTexelBufferOffsetSingleTexelAlignment
= true;
1848 properties
->uniformTexelBufferOffsetAlignmentBytes
= 4;
1849 properties
->uniformTexelBufferOffsetSingleTexelAlignment
= true;
1852 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT_CONTROLS_PROPERTIES
: {
1853 VkPhysicalDeviceFloatControlsProperties
*properties
=
1854 (VkPhysicalDeviceFloatControlsProperties
*)ext
;
1855 CORE_PROPERTY(1, 2, denormBehaviorIndependence
);
1856 CORE_PROPERTY(1, 2, roundingModeIndependence
);
1857 CORE_PROPERTY(1, 2, shaderDenormFlushToZeroFloat16
);
1858 CORE_PROPERTY(1, 2, shaderDenormPreserveFloat16
);
1859 CORE_PROPERTY(1, 2, shaderRoundingModeRTEFloat16
);
1860 CORE_PROPERTY(1, 2, shaderRoundingModeRTZFloat16
);
1861 CORE_PROPERTY(1, 2, shaderSignedZeroInfNanPreserveFloat16
);
1862 CORE_PROPERTY(1, 2, shaderDenormFlushToZeroFloat32
);
1863 CORE_PROPERTY(1, 2, shaderDenormPreserveFloat32
);
1864 CORE_PROPERTY(1, 2, shaderRoundingModeRTEFloat32
);
1865 CORE_PROPERTY(1, 2, shaderRoundingModeRTZFloat32
);
1866 CORE_PROPERTY(1, 2, shaderSignedZeroInfNanPreserveFloat32
);
1867 CORE_PROPERTY(1, 2, shaderDenormFlushToZeroFloat64
);
1868 CORE_PROPERTY(1, 2, shaderDenormPreserveFloat64
);
1869 CORE_PROPERTY(1, 2, shaderRoundingModeRTEFloat64
);
1870 CORE_PROPERTY(1, 2, shaderRoundingModeRTZFloat64
);
1871 CORE_PROPERTY(1, 2, shaderSignedZeroInfNanPreserveFloat64
);
1874 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_PROPERTIES
: {
1875 VkPhysicalDeviceTimelineSemaphoreProperties
*properties
=
1876 (VkPhysicalDeviceTimelineSemaphoreProperties
*) ext
;
1877 CORE_PROPERTY(1, 2, maxTimelineSemaphoreValueDifference
);
1880 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_PROPERTIES_EXT
: {
1881 VkPhysicalDeviceSubgroupSizeControlPropertiesEXT
*props
=
1882 (VkPhysicalDeviceSubgroupSizeControlPropertiesEXT
*)ext
;
1883 props
->minSubgroupSize
= 64;
1884 props
->maxSubgroupSize
= 64;
1885 props
->maxComputeWorkgroupSubgroups
= UINT32_MAX
;
1886 props
->requiredSubgroupSizeStages
= 0;
1888 if (pdevice
->rad_info
.chip_class
>= GFX10
) {
1889 /* Only GFX10+ supports wave32. */
1890 props
->minSubgroupSize
= 32;
1891 props
->requiredSubgroupSizeStages
= VK_SHADER_STAGE_COMPUTE_BIT
;
1895 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES
:
1896 radv_get_physical_device_properties_1_1(pdevice
, (void *)ext
);
1898 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES
:
1899 radv_get_physical_device_properties_1_2(pdevice
, (void *)ext
);
1901 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_PROPERTIES_EXT
: {
1902 VkPhysicalDeviceLineRasterizationPropertiesEXT
*props
=
1903 (VkPhysicalDeviceLineRasterizationPropertiesEXT
*)ext
;
1904 props
->lineSubPixelPrecisionBits
= 4;
1907 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ROBUSTNESS_2_PROPERTIES_EXT
: {
1908 VkPhysicalDeviceRobustness2PropertiesEXT
*properties
=
1909 (VkPhysicalDeviceRobustness2PropertiesEXT
*)ext
;
1910 properties
->robustStorageBufferAccessSizeAlignment
= 4;
1911 properties
->robustUniformBufferAccessSizeAlignment
= 4;
1920 static void radv_get_physical_device_queue_family_properties(
1921 struct radv_physical_device
* pdevice
,
1923 VkQueueFamilyProperties
** pQueueFamilyProperties
)
1925 int num_queue_families
= 1;
1927 if (pdevice
->rad_info
.num_rings
[RING_COMPUTE
] > 0 &&
1928 !(pdevice
->instance
->debug_flags
& RADV_DEBUG_NO_COMPUTE_QUEUE
))
1929 num_queue_families
++;
1931 if (pQueueFamilyProperties
== NULL
) {
1932 *pCount
= num_queue_families
;
1941 *pQueueFamilyProperties
[idx
] = (VkQueueFamilyProperties
) {
1942 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
1943 VK_QUEUE_COMPUTE_BIT
|
1944 VK_QUEUE_TRANSFER_BIT
|
1945 VK_QUEUE_SPARSE_BINDING_BIT
,
1947 .timestampValidBits
= 64,
1948 .minImageTransferGranularity
= (VkExtent3D
) { 1, 1, 1 },
1953 if (pdevice
->rad_info
.num_rings
[RING_COMPUTE
] > 0 &&
1954 !(pdevice
->instance
->debug_flags
& RADV_DEBUG_NO_COMPUTE_QUEUE
)) {
1955 if (*pCount
> idx
) {
1956 *pQueueFamilyProperties
[idx
] = (VkQueueFamilyProperties
) {
1957 .queueFlags
= VK_QUEUE_COMPUTE_BIT
|
1958 VK_QUEUE_TRANSFER_BIT
|
1959 VK_QUEUE_SPARSE_BINDING_BIT
,
1960 .queueCount
= pdevice
->rad_info
.num_rings
[RING_COMPUTE
],
1961 .timestampValidBits
= 64,
1962 .minImageTransferGranularity
= (VkExtent3D
) { 1, 1, 1 },
1970 void radv_GetPhysicalDeviceQueueFamilyProperties(
1971 VkPhysicalDevice physicalDevice
,
1973 VkQueueFamilyProperties
* pQueueFamilyProperties
)
1975 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
1976 if (!pQueueFamilyProperties
) {
1977 radv_get_physical_device_queue_family_properties(pdevice
, pCount
, NULL
);
1980 VkQueueFamilyProperties
*properties
[] = {
1981 pQueueFamilyProperties
+ 0,
1982 pQueueFamilyProperties
+ 1,
1983 pQueueFamilyProperties
+ 2,
1985 radv_get_physical_device_queue_family_properties(pdevice
, pCount
, properties
);
1986 assert(*pCount
<= 3);
1989 void radv_GetPhysicalDeviceQueueFamilyProperties2(
1990 VkPhysicalDevice physicalDevice
,
1992 VkQueueFamilyProperties2
*pQueueFamilyProperties
)
1994 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
1995 if (!pQueueFamilyProperties
) {
1996 radv_get_physical_device_queue_family_properties(pdevice
, pCount
, NULL
);
1999 VkQueueFamilyProperties
*properties
[] = {
2000 &pQueueFamilyProperties
[0].queueFamilyProperties
,
2001 &pQueueFamilyProperties
[1].queueFamilyProperties
,
2002 &pQueueFamilyProperties
[2].queueFamilyProperties
,
2004 radv_get_physical_device_queue_family_properties(pdevice
, pCount
, properties
);
2005 assert(*pCount
<= 3);
2008 void radv_GetPhysicalDeviceMemoryProperties(
2009 VkPhysicalDevice physicalDevice
,
2010 VkPhysicalDeviceMemoryProperties
*pMemoryProperties
)
2012 RADV_FROM_HANDLE(radv_physical_device
, physical_device
, physicalDevice
);
2014 *pMemoryProperties
= physical_device
->memory_properties
;
2018 radv_get_memory_budget_properties(VkPhysicalDevice physicalDevice
,
2019 VkPhysicalDeviceMemoryBudgetPropertiesEXT
*memoryBudget
)
2021 RADV_FROM_HANDLE(radv_physical_device
, device
, physicalDevice
);
2022 VkPhysicalDeviceMemoryProperties
*memory_properties
= &device
->memory_properties
;
2023 uint64_t visible_vram_size
= radv_get_visible_vram_size(device
);
2024 uint64_t vram_size
= radv_get_vram_size(device
);
2025 uint64_t gtt_size
= device
->rad_info
.gart_size
;
2026 uint64_t heap_budget
, heap_usage
;
2028 /* For all memory heaps, the computation of budget is as follow:
2029 * heap_budget = heap_size - global_heap_usage + app_heap_usage
2031 * The Vulkan spec 1.1.97 says that the budget should include any
2032 * currently allocated device memory.
2034 * Note that the application heap usages are not really accurate (eg.
2035 * in presence of shared buffers).
2037 for (int i
= 0; i
< device
->memory_properties
.memoryTypeCount
; i
++) {
2038 uint32_t heap_index
= device
->memory_properties
.memoryTypes
[i
].heapIndex
;
2040 if ((device
->memory_domains
[i
] & RADEON_DOMAIN_VRAM
) && (device
->memory_flags
[i
] & RADEON_FLAG_NO_CPU_ACCESS
)) {
2041 heap_usage
= device
->ws
->query_value(device
->ws
,
2042 RADEON_ALLOCATED_VRAM
);
2044 heap_budget
= vram_size
-
2045 device
->ws
->query_value(device
->ws
, RADEON_VRAM_USAGE
) +
2048 memoryBudget
->heapBudget
[heap_index
] = heap_budget
;
2049 memoryBudget
->heapUsage
[heap_index
] = heap_usage
;
2050 } else if (device
->memory_domains
[i
] & RADEON_DOMAIN_VRAM
) {
2051 heap_usage
= device
->ws
->query_value(device
->ws
,
2052 RADEON_ALLOCATED_VRAM_VIS
);
2054 heap_budget
= visible_vram_size
-
2055 device
->ws
->query_value(device
->ws
, RADEON_VRAM_VIS_USAGE
) +
2058 memoryBudget
->heapBudget
[heap_index
] = heap_budget
;
2059 memoryBudget
->heapUsage
[heap_index
] = heap_usage
;
2061 assert(device
->memory_domains
[i
] & RADEON_DOMAIN_GTT
);
2063 heap_usage
= device
->ws
->query_value(device
->ws
,
2064 RADEON_ALLOCATED_GTT
);
2066 heap_budget
= gtt_size
-
2067 device
->ws
->query_value(device
->ws
, RADEON_GTT_USAGE
) +
2070 memoryBudget
->heapBudget
[heap_index
] = heap_budget
;
2071 memoryBudget
->heapUsage
[heap_index
] = heap_usage
;
2075 /* The heapBudget and heapUsage values must be zero for array elements
2076 * greater than or equal to
2077 * VkPhysicalDeviceMemoryProperties::memoryHeapCount.
2079 for (uint32_t i
= memory_properties
->memoryHeapCount
; i
< VK_MAX_MEMORY_HEAPS
; i
++) {
2080 memoryBudget
->heapBudget
[i
] = 0;
2081 memoryBudget
->heapUsage
[i
] = 0;
2085 void radv_GetPhysicalDeviceMemoryProperties2(
2086 VkPhysicalDevice physicalDevice
,
2087 VkPhysicalDeviceMemoryProperties2
*pMemoryProperties
)
2089 radv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
2090 &pMemoryProperties
->memoryProperties
);
2092 VkPhysicalDeviceMemoryBudgetPropertiesEXT
*memory_budget
=
2093 vk_find_struct(pMemoryProperties
->pNext
,
2094 PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT
);
2096 radv_get_memory_budget_properties(physicalDevice
, memory_budget
);
2099 VkResult
radv_GetMemoryHostPointerPropertiesEXT(
2101 VkExternalMemoryHandleTypeFlagBits handleType
,
2102 const void *pHostPointer
,
2103 VkMemoryHostPointerPropertiesEXT
*pMemoryHostPointerProperties
)
2105 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2109 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
: {
2110 const struct radv_physical_device
*physical_device
= device
->physical_device
;
2111 uint32_t memoryTypeBits
= 0;
2112 for (int i
= 0; i
< physical_device
->memory_properties
.memoryTypeCount
; i
++) {
2113 if (physical_device
->memory_domains
[i
] == RADEON_DOMAIN_GTT
&&
2114 !(physical_device
->memory_flags
[i
] & RADEON_FLAG_GTT_WC
)) {
2115 memoryTypeBits
= (1 << i
);
2119 pMemoryHostPointerProperties
->memoryTypeBits
= memoryTypeBits
;
2123 return VK_ERROR_INVALID_EXTERNAL_HANDLE
;
2127 static enum radeon_ctx_priority
2128 radv_get_queue_global_priority(const VkDeviceQueueGlobalPriorityCreateInfoEXT
*pObj
)
2130 /* Default to MEDIUM when a specific global priority isn't requested */
2132 return RADEON_CTX_PRIORITY_MEDIUM
;
2134 switch(pObj
->globalPriority
) {
2135 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT
:
2136 return RADEON_CTX_PRIORITY_REALTIME
;
2137 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT
:
2138 return RADEON_CTX_PRIORITY_HIGH
;
2139 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
:
2140 return RADEON_CTX_PRIORITY_MEDIUM
;
2141 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT
:
2142 return RADEON_CTX_PRIORITY_LOW
;
2144 unreachable("Illegal global priority value");
2145 return RADEON_CTX_PRIORITY_INVALID
;
2150 radv_queue_init(struct radv_device
*device
, struct radv_queue
*queue
,
2151 uint32_t queue_family_index
, int idx
,
2152 VkDeviceQueueCreateFlags flags
,
2153 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*global_priority
)
2155 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
2156 queue
->device
= device
;
2157 queue
->queue_family_index
= queue_family_index
;
2158 queue
->queue_idx
= idx
;
2159 queue
->priority
= radv_get_queue_global_priority(global_priority
);
2160 queue
->flags
= flags
;
2162 queue
->hw_ctx
= device
->ws
->ctx_create(device
->ws
, queue
->priority
);
2164 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
2166 list_inithead(&queue
->pending_submissions
);
2167 pthread_mutex_init(&queue
->pending_mutex
, NULL
);
2173 radv_queue_finish(struct radv_queue
*queue
)
2175 pthread_mutex_destroy(&queue
->pending_mutex
);
2178 queue
->device
->ws
->ctx_destroy(queue
->hw_ctx
);
2180 if (queue
->initial_full_flush_preamble_cs
)
2181 queue
->device
->ws
->cs_destroy(queue
->initial_full_flush_preamble_cs
);
2182 if (queue
->initial_preamble_cs
)
2183 queue
->device
->ws
->cs_destroy(queue
->initial_preamble_cs
);
2184 if (queue
->continue_preamble_cs
)
2185 queue
->device
->ws
->cs_destroy(queue
->continue_preamble_cs
);
2186 if (queue
->descriptor_bo
)
2187 queue
->device
->ws
->buffer_destroy(queue
->descriptor_bo
);
2188 if (queue
->scratch_bo
)
2189 queue
->device
->ws
->buffer_destroy(queue
->scratch_bo
);
2190 if (queue
->esgs_ring_bo
)
2191 queue
->device
->ws
->buffer_destroy(queue
->esgs_ring_bo
);
2192 if (queue
->gsvs_ring_bo
)
2193 queue
->device
->ws
->buffer_destroy(queue
->gsvs_ring_bo
);
2194 if (queue
->tess_rings_bo
)
2195 queue
->device
->ws
->buffer_destroy(queue
->tess_rings_bo
);
2197 queue
->device
->ws
->buffer_destroy(queue
->gds_bo
);
2198 if (queue
->gds_oa_bo
)
2199 queue
->device
->ws
->buffer_destroy(queue
->gds_oa_bo
);
2200 if (queue
->compute_scratch_bo
)
2201 queue
->device
->ws
->buffer_destroy(queue
->compute_scratch_bo
);
2205 radv_bo_list_init(struct radv_bo_list
*bo_list
)
2207 pthread_mutex_init(&bo_list
->mutex
, NULL
);
2208 bo_list
->list
.count
= bo_list
->capacity
= 0;
2209 bo_list
->list
.bos
= NULL
;
2213 radv_bo_list_finish(struct radv_bo_list
*bo_list
)
2215 free(bo_list
->list
.bos
);
2216 pthread_mutex_destroy(&bo_list
->mutex
);
2219 VkResult
radv_bo_list_add(struct radv_device
*device
,
2220 struct radeon_winsys_bo
*bo
)
2222 struct radv_bo_list
*bo_list
= &device
->bo_list
;
2227 if (unlikely(!device
->use_global_bo_list
))
2230 pthread_mutex_lock(&bo_list
->mutex
);
2231 if (bo_list
->list
.count
== bo_list
->capacity
) {
2232 unsigned capacity
= MAX2(4, bo_list
->capacity
* 2);
2233 void *data
= realloc(bo_list
->list
.bos
, capacity
* sizeof(struct radeon_winsys_bo
*));
2236 pthread_mutex_unlock(&bo_list
->mutex
);
2237 return VK_ERROR_OUT_OF_HOST_MEMORY
;
2240 bo_list
->list
.bos
= (struct radeon_winsys_bo
**)data
;
2241 bo_list
->capacity
= capacity
;
2244 bo_list
->list
.bos
[bo_list
->list
.count
++] = bo
;
2245 pthread_mutex_unlock(&bo_list
->mutex
);
2249 void radv_bo_list_remove(struct radv_device
*device
,
2250 struct radeon_winsys_bo
*bo
)
2252 struct radv_bo_list
*bo_list
= &device
->bo_list
;
2257 if (unlikely(!device
->use_global_bo_list
))
2260 pthread_mutex_lock(&bo_list
->mutex
);
2261 /* Loop the list backwards so we find the most recently added
2263 for(unsigned i
= bo_list
->list
.count
; i
-- > 0;) {
2264 if (bo_list
->list
.bos
[i
] == bo
) {
2265 bo_list
->list
.bos
[i
] = bo_list
->list
.bos
[bo_list
->list
.count
- 1];
2266 --bo_list
->list
.count
;
2270 pthread_mutex_unlock(&bo_list
->mutex
);
2274 radv_device_init_gs_info(struct radv_device
*device
)
2276 device
->gs_table_depth
= ac_get_gs_table_depth(device
->physical_device
->rad_info
.chip_class
,
2277 device
->physical_device
->rad_info
.family
);
2280 static int radv_get_device_extension_index(const char *name
)
2282 for (unsigned i
= 0; i
< RADV_DEVICE_EXTENSION_COUNT
; ++i
) {
2283 if (strcmp(name
, radv_device_extensions
[i
].extensionName
) == 0)
2290 radv_get_int_debug_option(const char *name
, int default_value
)
2297 result
= default_value
;
2301 result
= strtol(str
, &endptr
, 0);
2302 if (str
== endptr
) {
2303 /* No digits founs. */
2304 result
= default_value
;
2311 static int install_seccomp_filter() {
2313 struct sock_filter filter
[] = {
2314 /* Check arch is 64bit x86 */
2315 BPF_STMT(BPF_LD
+ BPF_W
+ BPF_ABS
, (offsetof(struct seccomp_data
, arch
))),
2316 BPF_JUMP(BPF_JMP
+ BPF_JEQ
+ BPF_K
, AUDIT_ARCH_X86_64
, 0, 12),
2318 /* Futex is required for mutex locks */
2319 #if defined __NR__newselect
2320 BPF_STMT(BPF_LD
+ BPF_W
+ BPF_ABS
, (offsetof(struct seccomp_data
, nr
))),
2321 BPF_JUMP(BPF_JMP
+ BPF_JEQ
+ BPF_K
, __NR__newselect
, 11, 0),
2322 #elif defined __NR_select
2323 BPF_STMT(BPF_LD
+ BPF_W
+ BPF_ABS
, (offsetof(struct seccomp_data
, nr
))),
2324 BPF_JUMP(BPF_JMP
+ BPF_JEQ
+ BPF_K
, __NR_select
, 11, 0),
2326 BPF_STMT(BPF_LD
+ BPF_W
+ BPF_ABS
, (offsetof(struct seccomp_data
, nr
))),
2327 BPF_JUMP(BPF_JMP
+ BPF_JEQ
+ BPF_K
, __NR_pselect6
, 11, 0),
2330 /* Allow system exit calls for the forked process */
2331 BPF_STMT(BPF_LD
+ BPF_W
+ BPF_ABS
, (offsetof(struct seccomp_data
, nr
))),
2332 BPF_JUMP(BPF_JMP
+ BPF_JEQ
+ BPF_K
, __NR_exit_group
, 9, 0),
2334 /* Allow system read calls */
2335 BPF_STMT(BPF_LD
+ BPF_W
+ BPF_ABS
, (offsetof(struct seccomp_data
, nr
))),
2336 BPF_JUMP(BPF_JMP
+ BPF_JEQ
+ BPF_K
, __NR_read
, 7, 0),
2338 /* Allow system write calls */
2339 BPF_STMT(BPF_LD
+ BPF_W
+ BPF_ABS
, (offsetof(struct seccomp_data
, nr
))),
2340 BPF_JUMP(BPF_JMP
+ BPF_JEQ
+ BPF_K
, __NR_write
, 5, 0),
2342 /* Allow system brk calls (we need this for malloc) */
2343 BPF_STMT(BPF_LD
+ BPF_W
+ BPF_ABS
, (offsetof(struct seccomp_data
, nr
))),
2344 BPF_JUMP(BPF_JMP
+ BPF_JEQ
+ BPF_K
, __NR_brk
, 3, 0),
2346 /* Futex is required for mutex locks */
2347 BPF_STMT(BPF_LD
+ BPF_W
+ BPF_ABS
, (offsetof(struct seccomp_data
, nr
))),
2348 BPF_JUMP(BPF_JMP
+ BPF_JEQ
+ BPF_K
, __NR_futex
, 1, 0),
2350 /* Return error if we hit a system call not on the whitelist */
2351 BPF_STMT(BPF_RET
+ BPF_K
, SECCOMP_RET_ERRNO
| (EPERM
& SECCOMP_RET_DATA
)),
2353 /* Allow whitelisted system calls */
2354 BPF_STMT(BPF_RET
+ BPF_K
, SECCOMP_RET_ALLOW
),
2357 struct sock_fprog prog
= {
2358 .len
= (unsigned short)(sizeof(filter
) / sizeof(filter
[0])),
2362 if (prctl(PR_SET_NO_NEW_PRIVS
, 1, 0, 0, 0))
2365 if (prctl(PR_SET_SECCOMP
, SECCOMP_MODE_FILTER
, &prog
))
2371 /* Helper function with timeout support for reading from the pipe between
2372 * processes used for secure compile.
2374 bool radv_sc_read(int fd
, void *buf
, size_t size
, bool timeout
)
2383 /* We can't rely on the value of tv after calling select() so
2384 * we must reset it on each iteration of the loop.
2389 int rval
= select(fd
+ 1, &fds
, NULL
, NULL
, timeout
? &tv
: NULL
);
2395 ssize_t bytes_read
= read(fd
, buf
, size
);
2404 /* select timeout */
2410 static bool radv_close_all_fds(const int *keep_fds
, int keep_fd_count
)
2414 d
= opendir("/proc/self/fd");
2417 int dir_fd
= dirfd(d
);
2419 while ((dir
= readdir(d
)) != NULL
) {
2420 if (dir
->d_name
[0] == '.')
2423 int fd
= atoi(dir
->d_name
);
2428 for (int i
= 0; !keep
&& i
< keep_fd_count
; ++i
)
2429 if (keep_fds
[i
] == fd
)
2441 static bool secure_compile_open_fifo_fds(struct radv_secure_compile_state
*sc
,
2442 int *fd_server
, int *fd_client
,
2443 unsigned process
, bool make_fifo
)
2445 bool result
= false;
2446 char *fifo_server_path
= NULL
;
2447 char *fifo_client_path
= NULL
;
2449 if (asprintf(&fifo_server_path
, "/tmp/radv_server_%s_%u", sc
->uid
, process
) == -1)
2450 goto open_fifo_exit
;
2452 if (asprintf(&fifo_client_path
, "/tmp/radv_client_%s_%u", sc
->uid
, process
) == -1)
2453 goto open_fifo_exit
;
2456 int file1
= mkfifo(fifo_server_path
, 0666);
2458 goto open_fifo_exit
;
2460 int file2
= mkfifo(fifo_client_path
, 0666);
2462 goto open_fifo_exit
;
2465 *fd_server
= open(fifo_server_path
, O_RDWR
);
2467 goto open_fifo_exit
;
2469 *fd_client
= open(fifo_client_path
, O_RDWR
);
2470 if(*fd_client
< 1) {
2472 goto open_fifo_exit
;
2478 free(fifo_server_path
);
2479 free(fifo_client_path
);
2484 static void run_secure_compile_device(struct radv_device
*device
, unsigned process
,
2485 int fd_idle_device_output
)
2487 int fd_secure_input
;
2488 int fd_secure_output
;
2489 bool fifo_result
= secure_compile_open_fifo_fds(device
->sc_state
,
2494 enum radv_secure_compile_type sc_type
;
2496 const int needed_fds
[] = {
2499 fd_idle_device_output
,
2502 if (!fifo_result
|| !radv_close_all_fds(needed_fds
, ARRAY_SIZE(needed_fds
)) ||
2503 install_seccomp_filter() == -1) {
2504 sc_type
= RADV_SC_TYPE_INIT_FAILURE
;
2506 sc_type
= RADV_SC_TYPE_INIT_SUCCESS
;
2507 device
->sc_state
->secure_compile_processes
[process
].fd_secure_input
= fd_secure_input
;
2508 device
->sc_state
->secure_compile_processes
[process
].fd_secure_output
= fd_secure_output
;
2511 write(fd_idle_device_output
, &sc_type
, sizeof(sc_type
));
2513 if (sc_type
== RADV_SC_TYPE_INIT_FAILURE
)
2514 goto secure_compile_exit
;
2517 radv_sc_read(fd_secure_input
, &sc_type
, sizeof(sc_type
), false);
2519 if (sc_type
== RADV_SC_TYPE_COMPILE_PIPELINE
) {
2520 struct radv_pipeline
*pipeline
;
2521 bool sc_read
= true;
2523 pipeline
= vk_zalloc2(&device
->alloc
, NULL
, sizeof(*pipeline
), 8,
2524 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2526 pipeline
->device
= device
;
2528 /* Read pipeline layout */
2529 struct radv_pipeline_layout layout
;
2530 sc_read
= radv_sc_read(fd_secure_input
, &layout
, sizeof(struct radv_pipeline_layout
), true);
2531 sc_read
&= radv_sc_read(fd_secure_input
, &layout
.num_sets
, sizeof(uint32_t), true);
2533 goto secure_compile_exit
;
2535 for (uint32_t set
= 0; set
< layout
.num_sets
; set
++) {
2536 uint32_t layout_size
;
2537 sc_read
&= radv_sc_read(fd_secure_input
, &layout_size
, sizeof(uint32_t), true);
2539 goto secure_compile_exit
;
2541 layout
.set
[set
].layout
= malloc(layout_size
);
2542 layout
.set
[set
].layout
->layout_size
= layout_size
;
2543 sc_read
&= radv_sc_read(fd_secure_input
, layout
.set
[set
].layout
,
2544 layout
.set
[set
].layout
->layout_size
, true);
2547 pipeline
->layout
= &layout
;
2549 /* Read pipeline key */
2550 struct radv_pipeline_key key
;
2551 sc_read
&= radv_sc_read(fd_secure_input
, &key
, sizeof(struct radv_pipeline_key
), true);
2553 /* Read pipeline create flags */
2554 VkPipelineCreateFlags flags
;
2555 sc_read
&= radv_sc_read(fd_secure_input
, &flags
, sizeof(VkPipelineCreateFlags
), true);
2557 /* Read stage and shader information */
2558 uint32_t num_stages
;
2559 const VkPipelineShaderStageCreateInfo
*pStages
[MESA_SHADER_STAGES
] = { 0, };
2560 sc_read
&= radv_sc_read(fd_secure_input
, &num_stages
, sizeof(uint32_t), true);
2562 goto secure_compile_exit
;
2564 for (uint32_t i
= 0; i
< num_stages
; i
++) {
2567 gl_shader_stage stage
;
2568 sc_read
&= radv_sc_read(fd_secure_input
, &stage
, sizeof(gl_shader_stage
), true);
2570 VkPipelineShaderStageCreateInfo
*pStage
= calloc(1, sizeof(VkPipelineShaderStageCreateInfo
));
2572 /* Read entry point name */
2574 sc_read
&= radv_sc_read(fd_secure_input
, &name_size
, sizeof(size_t), true);
2576 goto secure_compile_exit
;
2578 char *ep_name
= malloc(name_size
);
2579 sc_read
&= radv_sc_read(fd_secure_input
, ep_name
, name_size
, true);
2580 pStage
->pName
= ep_name
;
2582 /* Read shader module */
2584 sc_read
&= radv_sc_read(fd_secure_input
, &module_size
, sizeof(size_t), true);
2586 goto secure_compile_exit
;
2588 struct radv_shader_module
*module
= malloc(module_size
);
2589 sc_read
&= radv_sc_read(fd_secure_input
, module
, module_size
, true);
2590 pStage
->module
= radv_shader_module_to_handle(module
);
2592 /* Read specialization info */
2594 sc_read
&= radv_sc_read(fd_secure_input
, &has_spec_info
, sizeof(bool), true);
2596 goto secure_compile_exit
;
2598 if (has_spec_info
) {
2599 VkSpecializationInfo
*specInfo
= malloc(sizeof(VkSpecializationInfo
));
2600 pStage
->pSpecializationInfo
= specInfo
;
2602 sc_read
&= radv_sc_read(fd_secure_input
, &specInfo
->dataSize
, sizeof(size_t), true);
2604 goto secure_compile_exit
;
2606 void *si_data
= malloc(specInfo
->dataSize
);
2607 sc_read
&= radv_sc_read(fd_secure_input
, si_data
, specInfo
->dataSize
, true);
2608 specInfo
->pData
= si_data
;
2610 sc_read
&= radv_sc_read(fd_secure_input
, &specInfo
->mapEntryCount
, sizeof(uint32_t), true);
2612 goto secure_compile_exit
;
2614 VkSpecializationMapEntry
*mapEntries
= malloc(sizeof(VkSpecializationMapEntry
) * specInfo
->mapEntryCount
);
2615 for (uint32_t j
= 0; j
< specInfo
->mapEntryCount
; j
++) {
2616 sc_read
&= radv_sc_read(fd_secure_input
, &mapEntries
[j
], sizeof(VkSpecializationMapEntry
), true);
2618 goto secure_compile_exit
;
2621 specInfo
->pMapEntries
= mapEntries
;
2624 pStages
[stage
] = pStage
;
2627 /* Compile the shaders */
2628 VkPipelineCreationFeedbackEXT
*stage_feedbacks
[MESA_SHADER_STAGES
] = { 0 };
2629 radv_create_shaders(pipeline
, device
, NULL
, &key
, pStages
, flags
, NULL
, stage_feedbacks
);
2631 /* free memory allocated above */
2632 for (uint32_t set
= 0; set
< layout
.num_sets
; set
++)
2633 free(layout
.set
[set
].layout
);
2635 for (uint32_t i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
2639 free((void *) pStages
[i
]->pName
);
2640 free(radv_shader_module_from_handle(pStages
[i
]->module
));
2641 if (pStages
[i
]->pSpecializationInfo
) {
2642 free((void *) pStages
[i
]->pSpecializationInfo
->pData
);
2643 free((void *) pStages
[i
]->pSpecializationInfo
->pMapEntries
);
2644 free((void *) pStages
[i
]->pSpecializationInfo
);
2646 free((void *) pStages
[i
]);
2649 vk_free(&device
->alloc
, pipeline
);
2651 sc_type
= RADV_SC_TYPE_COMPILE_PIPELINE_FINISHED
;
2652 write(fd_secure_output
, &sc_type
, sizeof(sc_type
));
2654 } else if (sc_type
== RADV_SC_TYPE_DESTROY_DEVICE
) {
2655 goto secure_compile_exit
;
2659 secure_compile_exit
:
2660 close(fd_secure_input
);
2661 close(fd_secure_output
);
2662 close(fd_idle_device_output
);
2666 static enum radv_secure_compile_type
fork_secure_compile_device(struct radv_device
*device
, unsigned process
)
2668 int fd_secure_input
[2];
2669 int fd_secure_output
[2];
2671 /* create pipe descriptors (used to communicate between processes) */
2672 if (pipe(fd_secure_input
) == -1 || pipe(fd_secure_output
) == -1)
2673 return RADV_SC_TYPE_INIT_FAILURE
;
2677 if ((sc_pid
= fork()) == 0) {
2678 device
->sc_state
->secure_compile_thread_counter
= process
;
2679 run_secure_compile_device(device
, process
, fd_secure_output
[1]);
2682 return RADV_SC_TYPE_INIT_FAILURE
;
2684 /* Read the init result returned from the secure process */
2685 enum radv_secure_compile_type sc_type
;
2686 bool sc_read
= radv_sc_read(fd_secure_output
[0], &sc_type
, sizeof(sc_type
), true);
2688 if (sc_type
== RADV_SC_TYPE_INIT_FAILURE
|| !sc_read
) {
2689 close(fd_secure_input
[0]);
2690 close(fd_secure_input
[1]);
2691 close(fd_secure_output
[1]);
2692 close(fd_secure_output
[0]);
2694 waitpid(sc_pid
, &status
, 0);
2696 return RADV_SC_TYPE_INIT_FAILURE
;
2698 assert(sc_type
== RADV_SC_TYPE_INIT_SUCCESS
);
2699 write(device
->sc_state
->secure_compile_processes
[process
].fd_secure_output
, &sc_type
, sizeof(sc_type
));
2701 close(fd_secure_input
[0]);
2702 close(fd_secure_input
[1]);
2703 close(fd_secure_output
[1]);
2704 close(fd_secure_output
[0]);
2707 waitpid(sc_pid
, &status
, 0);
2711 return RADV_SC_TYPE_INIT_SUCCESS
;
2714 /* Run a bare bones fork of a device that was forked right after its creation.
2715 * This device will have low overhead when it is forked again before each
2716 * pipeline compilation. This device sits idle and its only job is to fork
2719 static void run_secure_compile_idle_device(struct radv_device
*device
, unsigned process
,
2720 int fd_secure_input
, int fd_secure_output
)
2722 enum radv_secure_compile_type sc_type
= RADV_SC_TYPE_INIT_SUCCESS
;
2723 device
->sc_state
->secure_compile_processes
[process
].fd_secure_input
= fd_secure_input
;
2724 device
->sc_state
->secure_compile_processes
[process
].fd_secure_output
= fd_secure_output
;
2726 write(fd_secure_output
, &sc_type
, sizeof(sc_type
));
2729 radv_sc_read(fd_secure_input
, &sc_type
, sizeof(sc_type
), false);
2731 if (sc_type
== RADV_SC_TYPE_FORK_DEVICE
) {
2732 sc_type
= fork_secure_compile_device(device
, process
);
2734 if (sc_type
== RADV_SC_TYPE_INIT_FAILURE
)
2735 goto secure_compile_exit
;
2737 } else if (sc_type
== RADV_SC_TYPE_DESTROY_DEVICE
) {
2738 goto secure_compile_exit
;
2742 secure_compile_exit
:
2743 close(fd_secure_input
);
2744 close(fd_secure_output
);
2748 static void destroy_secure_compile_device(struct radv_device
*device
, unsigned process
)
2750 int fd_secure_input
= device
->sc_state
->secure_compile_processes
[process
].fd_secure_input
;
2752 enum radv_secure_compile_type sc_type
= RADV_SC_TYPE_DESTROY_DEVICE
;
2753 write(fd_secure_input
, &sc_type
, sizeof(sc_type
));
2755 close(device
->sc_state
->secure_compile_processes
[process
].fd_secure_input
);
2756 close(device
->sc_state
->secure_compile_processes
[process
].fd_secure_output
);
2759 waitpid(device
->sc_state
->secure_compile_processes
[process
].sc_pid
, &status
, 0);
2762 static VkResult
fork_secure_compile_idle_device(struct radv_device
*device
)
2764 device
->sc_state
= vk_zalloc(&device
->alloc
,
2765 sizeof(struct radv_secure_compile_state
),
2766 8, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
2768 mtx_init(&device
->sc_state
->secure_compile_mutex
, mtx_plain
);
2770 pid_t upid
= getpid();
2771 time_t seconds
= time(NULL
);
2774 if (asprintf(&uid
, "%ld_%ld", (long) upid
, (long) seconds
) == -1)
2775 return VK_ERROR_INITIALIZATION_FAILED
;
2777 device
->sc_state
->uid
= uid
;
2779 uint8_t sc_threads
= device
->instance
->num_sc_threads
;
2780 int fd_secure_input
[MAX_SC_PROCS
][2];
2781 int fd_secure_output
[MAX_SC_PROCS
][2];
2783 /* create pipe descriptors (used to communicate between processes) */
2784 for (unsigned i
= 0; i
< sc_threads
; i
++) {
2785 if (pipe(fd_secure_input
[i
]) == -1 ||
2786 pipe(fd_secure_output
[i
]) == -1) {
2787 return VK_ERROR_INITIALIZATION_FAILED
;
2791 device
->sc_state
->secure_compile_processes
= vk_zalloc(&device
->alloc
,
2792 sizeof(struct radv_secure_compile_process
) * sc_threads
, 8,
2793 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
2795 for (unsigned process
= 0; process
< sc_threads
; process
++) {
2796 if ((device
->sc_state
->secure_compile_processes
[process
].sc_pid
= fork()) == 0) {
2797 device
->sc_state
->secure_compile_thread_counter
= process
;
2798 run_secure_compile_idle_device(device
, process
, fd_secure_input
[process
][0], fd_secure_output
[process
][1]);
2800 if (device
->sc_state
->secure_compile_processes
[process
].sc_pid
== -1)
2801 return VK_ERROR_INITIALIZATION_FAILED
;
2803 /* Read the init result returned from the secure process */
2804 enum radv_secure_compile_type sc_type
;
2805 bool sc_read
= radv_sc_read(fd_secure_output
[process
][0], &sc_type
, sizeof(sc_type
), true);
2808 if (sc_read
&& sc_type
== RADV_SC_TYPE_INIT_SUCCESS
) {
2809 fifo_result
= secure_compile_open_fifo_fds(device
->sc_state
,
2810 &device
->sc_state
->secure_compile_processes
[process
].fd_server
,
2811 &device
->sc_state
->secure_compile_processes
[process
].fd_client
,
2814 device
->sc_state
->secure_compile_processes
[process
].fd_secure_input
= fd_secure_input
[process
][1];
2815 device
->sc_state
->secure_compile_processes
[process
].fd_secure_output
= fd_secure_output
[process
][0];
2818 if (sc_type
== RADV_SC_TYPE_INIT_FAILURE
|| !sc_read
|| !fifo_result
) {
2819 close(fd_secure_input
[process
][0]);
2820 close(fd_secure_input
[process
][1]);
2821 close(fd_secure_output
[process
][1]);
2822 close(fd_secure_output
[process
][0]);
2824 waitpid(device
->sc_state
->secure_compile_processes
[process
].sc_pid
, &status
, 0);
2826 /* Destroy any forks that were created sucessfully */
2827 for (unsigned i
= 0; i
< process
; i
++) {
2828 destroy_secure_compile_device(device
, i
);
2831 return VK_ERROR_INITIALIZATION_FAILED
;
2839 radv_device_init_dispatch(struct radv_device
*device
)
2841 const struct radv_instance
*instance
= device
->physical_device
->instance
;
2842 const struct radv_device_dispatch_table
*dispatch_table_layer
= NULL
;
2843 bool unchecked
= instance
->debug_flags
& RADV_DEBUG_ALL_ENTRYPOINTS
;
2844 int radv_thread_trace
= radv_get_int_debug_option("RADV_THREAD_TRACE", -1);
2846 if (radv_thread_trace
>= 0) {
2847 /* Use device entrypoints from the SQTT layer if enabled. */
2848 dispatch_table_layer
= &sqtt_device_dispatch_table
;
2851 for (unsigned i
= 0; i
< ARRAY_SIZE(device
->dispatch
.entrypoints
); i
++) {
2852 /* Vulkan requires that entrypoints for extensions which have not been
2853 * enabled must not be advertised.
2856 !radv_device_entrypoint_is_enabled(i
, instance
->apiVersion
,
2857 &instance
->enabled_extensions
,
2858 &device
->enabled_extensions
)) {
2859 device
->dispatch
.entrypoints
[i
] = NULL
;
2860 } else if (dispatch_table_layer
&&
2861 dispatch_table_layer
->entrypoints
[i
]) {
2862 device
->dispatch
.entrypoints
[i
] =
2863 dispatch_table_layer
->entrypoints
[i
];
2865 device
->dispatch
.entrypoints
[i
] =
2866 radv_device_dispatch_table
.entrypoints
[i
];
2872 radv_create_pthread_cond(pthread_cond_t
*cond
)
2874 pthread_condattr_t condattr
;
2875 if (pthread_condattr_init(&condattr
)) {
2876 return VK_ERROR_INITIALIZATION_FAILED
;
2879 if (pthread_condattr_setclock(&condattr
, CLOCK_MONOTONIC
)) {
2880 pthread_condattr_destroy(&condattr
);
2881 return VK_ERROR_INITIALIZATION_FAILED
;
2883 if (pthread_cond_init(cond
, &condattr
)) {
2884 pthread_condattr_destroy(&condattr
);
2885 return VK_ERROR_INITIALIZATION_FAILED
;
2887 pthread_condattr_destroy(&condattr
);
2892 check_physical_device_features(VkPhysicalDevice physicalDevice
,
2893 const VkPhysicalDeviceFeatures
*features
)
2895 RADV_FROM_HANDLE(radv_physical_device
, physical_device
, physicalDevice
);
2896 VkPhysicalDeviceFeatures supported_features
;
2897 radv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
2898 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
2899 VkBool32
*enabled_feature
= (VkBool32
*)features
;
2900 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
2901 for (uint32_t i
= 0; i
< num_features
; i
++) {
2902 if (enabled_feature
[i
] && !supported_feature
[i
])
2903 return vk_error(physical_device
->instance
, VK_ERROR_FEATURE_NOT_PRESENT
);
2909 VkResult
radv_CreateDevice(
2910 VkPhysicalDevice physicalDevice
,
2911 const VkDeviceCreateInfo
* pCreateInfo
,
2912 const VkAllocationCallbacks
* pAllocator
,
2915 RADV_FROM_HANDLE(radv_physical_device
, physical_device
, physicalDevice
);
2917 struct radv_device
*device
;
2919 bool keep_shader_info
= false;
2920 bool robust_buffer_access
= false;
2921 bool overallocation_disallowed
= false;
2923 /* Check enabled features */
2924 if (pCreateInfo
->pEnabledFeatures
) {
2925 result
= check_physical_device_features(physicalDevice
,
2926 pCreateInfo
->pEnabledFeatures
);
2927 if (result
!= VK_SUCCESS
)
2930 if (pCreateInfo
->pEnabledFeatures
->robustBufferAccess
)
2931 robust_buffer_access
= true;
2934 vk_foreach_struct_const(ext
, pCreateInfo
->pNext
) {
2935 switch (ext
->sType
) {
2936 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2
: {
2937 const VkPhysicalDeviceFeatures2
*features
= (const void *)ext
;
2938 result
= check_physical_device_features(physicalDevice
,
2939 &features
->features
);
2940 if (result
!= VK_SUCCESS
)
2943 if (features
->features
.robustBufferAccess
)
2944 robust_buffer_access
= true;
2947 case VK_STRUCTURE_TYPE_DEVICE_MEMORY_OVERALLOCATION_CREATE_INFO_AMD
: {
2948 const VkDeviceMemoryOverallocationCreateInfoAMD
*overallocation
= (const void *)ext
;
2949 if (overallocation
->overallocationBehavior
== VK_MEMORY_OVERALLOCATION_BEHAVIOR_DISALLOWED_AMD
)
2950 overallocation_disallowed
= true;
2958 device
= vk_zalloc2(&physical_device
->instance
->alloc
, pAllocator
,
2960 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
2962 return vk_error(physical_device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
2964 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
2965 device
->instance
= physical_device
->instance
;
2966 device
->physical_device
= physical_device
;
2968 device
->ws
= physical_device
->ws
;
2970 device
->alloc
= *pAllocator
;
2972 device
->alloc
= physical_device
->instance
->alloc
;
2974 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
2975 const char *ext_name
= pCreateInfo
->ppEnabledExtensionNames
[i
];
2976 int index
= radv_get_device_extension_index(ext_name
);
2977 if (index
< 0 || !physical_device
->supported_extensions
.extensions
[index
]) {
2978 vk_free(&device
->alloc
, device
);
2979 return vk_error(physical_device
->instance
, VK_ERROR_EXTENSION_NOT_PRESENT
);
2982 device
->enabled_extensions
.extensions
[index
] = true;
2985 radv_device_init_dispatch(device
);
2987 keep_shader_info
= device
->enabled_extensions
.AMD_shader_info
;
2989 /* With update after bind we can't attach bo's to the command buffer
2990 * from the descriptor set anymore, so we have to use a global BO list.
2992 device
->use_global_bo_list
=
2993 (device
->instance
->perftest_flags
& RADV_PERFTEST_BO_LIST
) ||
2994 device
->enabled_extensions
.EXT_descriptor_indexing
||
2995 device
->enabled_extensions
.EXT_buffer_device_address
||
2996 device
->enabled_extensions
.KHR_buffer_device_address
;
2998 device
->robust_buffer_access
= robust_buffer_access
;
3000 mtx_init(&device
->shader_slab_mutex
, mtx_plain
);
3001 list_inithead(&device
->shader_slabs
);
3003 device
->overallocation_disallowed
= overallocation_disallowed
;
3004 mtx_init(&device
->overallocation_mutex
, mtx_plain
);
3006 radv_bo_list_init(&device
->bo_list
);
3008 for (unsigned i
= 0; i
< pCreateInfo
->queueCreateInfoCount
; i
++) {
3009 const VkDeviceQueueCreateInfo
*queue_create
= &pCreateInfo
->pQueueCreateInfos
[i
];
3010 uint32_t qfi
= queue_create
->queueFamilyIndex
;
3011 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*global_priority
=
3012 vk_find_struct_const(queue_create
->pNext
, DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT
);
3014 assert(!global_priority
|| device
->physical_device
->rad_info
.has_ctx_priority
);
3016 device
->queues
[qfi
] = vk_alloc(&device
->alloc
,
3017 queue_create
->queueCount
* sizeof(struct radv_queue
), 8, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
3018 if (!device
->queues
[qfi
]) {
3019 result
= VK_ERROR_OUT_OF_HOST_MEMORY
;
3023 memset(device
->queues
[qfi
], 0, queue_create
->queueCount
* sizeof(struct radv_queue
));
3025 device
->queue_count
[qfi
] = queue_create
->queueCount
;
3027 for (unsigned q
= 0; q
< queue_create
->queueCount
; q
++) {
3028 result
= radv_queue_init(device
, &device
->queues
[qfi
][q
],
3029 qfi
, q
, queue_create
->flags
,
3031 if (result
!= VK_SUCCESS
)
3036 device
->pbb_allowed
= device
->physical_device
->rad_info
.chip_class
>= GFX9
&&
3037 !(device
->instance
->debug_flags
& RADV_DEBUG_NOBINNING
);
3039 /* Disable DFSM by default. As of 2019-09-15 Talos on Low is still 3% slower on Raven. */
3040 device
->dfsm_allowed
= device
->pbb_allowed
&&
3041 (device
->instance
->perftest_flags
& RADV_PERFTEST_DFSM
);
3043 device
->always_use_syncobj
= device
->physical_device
->rad_info
.has_syncobj_wait_for_submit
;
3045 /* The maximum number of scratch waves. Scratch space isn't divided
3046 * evenly between CUs. The number is only a function of the number of CUs.
3047 * We can decrease the constant to decrease the scratch buffer size.
3049 * sctx->scratch_waves must be >= the maximum possible size of
3050 * 1 threadgroup, so that the hw doesn't hang from being unable
3053 * The recommended value is 4 per CU at most. Higher numbers don't
3054 * bring much benefit, but they still occupy chip resources (think
3055 * async compute). I've seen ~2% performance difference between 4 and 32.
3057 uint32_t max_threads_per_block
= 2048;
3058 device
->scratch_waves
= MAX2(32 * physical_device
->rad_info
.num_good_compute_units
,
3059 max_threads_per_block
/ 64);
3061 device
->dispatch_initiator
= S_00B800_COMPUTE_SHADER_EN(1);
3063 if (device
->physical_device
->rad_info
.chip_class
>= GFX7
) {
3064 /* If the KMD allows it (there is a KMD hw register for it),
3065 * allow launching waves out-of-order.
3067 device
->dispatch_initiator
|= S_00B800_ORDER_MODE(1);
3070 radv_device_init_gs_info(device
);
3072 device
->tess_offchip_block_dw_size
=
3073 device
->physical_device
->rad_info
.family
== CHIP_HAWAII
? 4096 : 8192;
3075 if (getenv("RADV_TRACE_FILE")) {
3076 const char *filename
= getenv("RADV_TRACE_FILE");
3078 keep_shader_info
= true;
3080 if (!radv_init_trace(device
))
3083 fprintf(stderr
, "*****************************************************************************\n");
3084 fprintf(stderr
, "* WARNING: RADV_TRACE_FILE is costly and should only be used for debugging! *\n");
3085 fprintf(stderr
, "*****************************************************************************\n");
3087 fprintf(stderr
, "Trace file will be dumped to %s\n", filename
);
3088 radv_dump_enabled_options(device
, stderr
);
3091 int radv_thread_trace
= radv_get_int_debug_option("RADV_THREAD_TRACE", -1);
3092 if (radv_thread_trace
>= 0) {
3093 fprintf(stderr
, "*************************************************\n");
3094 fprintf(stderr
, "* WARNING: Thread trace support is experimental *\n");
3095 fprintf(stderr
, "*************************************************\n");
3097 if (device
->physical_device
->rad_info
.chip_class
< GFX8
) {
3098 fprintf(stderr
, "GPU hardware not supported: refer to "
3099 "the RGP documentation for the list of "
3100 "supported GPUs!\n");
3104 /* Default buffer size set to 1MB per SE. */
3105 device
->thread_trace_buffer_size
=
3106 radv_get_int_debug_option("RADV_THREAD_TRACE_BUFFER_SIZE", 1024 * 1024);
3107 device
->thread_trace_start_frame
= radv_thread_trace
;
3109 if (!radv_thread_trace_init(device
))
3113 /* Temporarily disable secure compile while we create meta shaders, etc */
3114 uint8_t sc_threads
= device
->instance
->num_sc_threads
;
3116 device
->instance
->num_sc_threads
= 0;
3118 device
->keep_shader_info
= keep_shader_info
;
3119 result
= radv_device_init_meta(device
);
3120 if (result
!= VK_SUCCESS
)
3123 radv_device_init_msaa(device
);
3125 for (int family
= 0; family
< RADV_MAX_QUEUE_FAMILIES
; ++family
) {
3126 device
->empty_cs
[family
] = device
->ws
->cs_create(device
->ws
, family
);
3128 case RADV_QUEUE_GENERAL
:
3129 radeon_emit(device
->empty_cs
[family
], PKT3(PKT3_CONTEXT_CONTROL
, 1, 0));
3130 radeon_emit(device
->empty_cs
[family
], CONTEXT_CONTROL_LOAD_ENABLE(1));
3131 radeon_emit(device
->empty_cs
[family
], CONTEXT_CONTROL_SHADOW_ENABLE(1));
3133 case RADV_QUEUE_COMPUTE
:
3134 radeon_emit(device
->empty_cs
[family
], PKT3(PKT3_NOP
, 0, 0));
3135 radeon_emit(device
->empty_cs
[family
], 0);
3138 device
->ws
->cs_finalize(device
->empty_cs
[family
]);
3141 if (device
->physical_device
->rad_info
.chip_class
>= GFX7
)
3142 cik_create_gfx_config(device
);
3144 VkPipelineCacheCreateInfo ci
;
3145 ci
.sType
= VK_STRUCTURE_TYPE_PIPELINE_CACHE_CREATE_INFO
;
3148 ci
.pInitialData
= NULL
;
3149 ci
.initialDataSize
= 0;
3151 result
= radv_CreatePipelineCache(radv_device_to_handle(device
),
3153 if (result
!= VK_SUCCESS
)
3156 device
->mem_cache
= radv_pipeline_cache_from_handle(pc
);
3158 result
= radv_create_pthread_cond(&device
->timeline_cond
);
3159 if (result
!= VK_SUCCESS
)
3160 goto fail_mem_cache
;
3162 device
->force_aniso
=
3163 MIN2(16, radv_get_int_debug_option("RADV_TEX_ANISO", -1));
3164 if (device
->force_aniso
>= 0) {
3165 fprintf(stderr
, "radv: Forcing anisotropy filter to %ix\n",
3166 1 << util_logbase2(device
->force_aniso
));
3169 /* Fork device for secure compile as required */
3170 device
->instance
->num_sc_threads
= sc_threads
;
3171 if (radv_device_use_secure_compile(device
->instance
)) {
3173 result
= fork_secure_compile_idle_device(device
);
3174 if (result
!= VK_SUCCESS
)
3178 *pDevice
= radv_device_to_handle(device
);
3182 radv_DestroyPipelineCache(radv_device_to_handle(device
), pc
, NULL
);
3184 radv_device_finish_meta(device
);
3186 radv_bo_list_finish(&device
->bo_list
);
3188 radv_thread_trace_finish(device
);
3190 if (device
->trace_bo
)
3191 device
->ws
->buffer_destroy(device
->trace_bo
);
3193 if (device
->gfx_init
)
3194 device
->ws
->buffer_destroy(device
->gfx_init
);
3196 for (unsigned i
= 0; i
< RADV_MAX_QUEUE_FAMILIES
; i
++) {
3197 for (unsigned q
= 0; q
< device
->queue_count
[i
]; q
++)
3198 radv_queue_finish(&device
->queues
[i
][q
]);
3199 if (device
->queue_count
[i
])
3200 vk_free(&device
->alloc
, device
->queues
[i
]);
3203 vk_free(&device
->alloc
, device
);
3207 void radv_DestroyDevice(
3209 const VkAllocationCallbacks
* pAllocator
)
3211 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3216 if (device
->trace_bo
)
3217 device
->ws
->buffer_destroy(device
->trace_bo
);
3219 if (device
->gfx_init
)
3220 device
->ws
->buffer_destroy(device
->gfx_init
);
3222 for (unsigned i
= 0; i
< RADV_MAX_QUEUE_FAMILIES
; i
++) {
3223 for (unsigned q
= 0; q
< device
->queue_count
[i
]; q
++)
3224 radv_queue_finish(&device
->queues
[i
][q
]);
3225 if (device
->queue_count
[i
])
3226 vk_free(&device
->alloc
, device
->queues
[i
]);
3227 if (device
->empty_cs
[i
])
3228 device
->ws
->cs_destroy(device
->empty_cs
[i
]);
3230 radv_device_finish_meta(device
);
3232 VkPipelineCache pc
= radv_pipeline_cache_to_handle(device
->mem_cache
);
3233 radv_DestroyPipelineCache(radv_device_to_handle(device
), pc
, NULL
);
3235 radv_destroy_shader_slabs(device
);
3237 pthread_cond_destroy(&device
->timeline_cond
);
3238 radv_bo_list_finish(&device
->bo_list
);
3240 radv_thread_trace_finish(device
);
3242 if (radv_device_use_secure_compile(device
->instance
)) {
3243 for (unsigned i
= 0; i
< device
->instance
->num_sc_threads
; i
++ ) {
3244 destroy_secure_compile_device(device
, i
);
3248 if (device
->sc_state
) {
3249 free(device
->sc_state
->uid
);
3250 vk_free(&device
->alloc
, device
->sc_state
->secure_compile_processes
);
3252 vk_free(&device
->alloc
, device
->sc_state
);
3253 vk_free(&device
->alloc
, device
);
3256 VkResult
radv_EnumerateInstanceLayerProperties(
3257 uint32_t* pPropertyCount
,
3258 VkLayerProperties
* pProperties
)
3260 if (pProperties
== NULL
) {
3261 *pPropertyCount
= 0;
3265 /* None supported at this time */
3266 return vk_error(NULL
, VK_ERROR_LAYER_NOT_PRESENT
);
3269 VkResult
radv_EnumerateDeviceLayerProperties(
3270 VkPhysicalDevice physicalDevice
,
3271 uint32_t* pPropertyCount
,
3272 VkLayerProperties
* pProperties
)
3274 if (pProperties
== NULL
) {
3275 *pPropertyCount
= 0;
3279 /* None supported at this time */
3280 return vk_error(NULL
, VK_ERROR_LAYER_NOT_PRESENT
);
3283 void radv_GetDeviceQueue2(
3285 const VkDeviceQueueInfo2
* pQueueInfo
,
3288 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3289 struct radv_queue
*queue
;
3291 queue
= &device
->queues
[pQueueInfo
->queueFamilyIndex
][pQueueInfo
->queueIndex
];
3292 if (pQueueInfo
->flags
!= queue
->flags
) {
3293 /* From the Vulkan 1.1.70 spec:
3295 * "The queue returned by vkGetDeviceQueue2 must have the same
3296 * flags value from this structure as that used at device
3297 * creation time in a VkDeviceQueueCreateInfo instance. If no
3298 * matching flags were specified at device creation time then
3299 * pQueue will return VK_NULL_HANDLE."
3301 *pQueue
= VK_NULL_HANDLE
;
3305 *pQueue
= radv_queue_to_handle(queue
);
3308 void radv_GetDeviceQueue(
3310 uint32_t queueFamilyIndex
,
3311 uint32_t queueIndex
,
3314 const VkDeviceQueueInfo2 info
= (VkDeviceQueueInfo2
) {
3315 .sType
= VK_STRUCTURE_TYPE_DEVICE_QUEUE_INFO_2
,
3316 .queueFamilyIndex
= queueFamilyIndex
,
3317 .queueIndex
= queueIndex
3320 radv_GetDeviceQueue2(_device
, &info
, pQueue
);
3324 fill_geom_tess_rings(struct radv_queue
*queue
,
3326 bool add_sample_positions
,
3327 uint32_t esgs_ring_size
,
3328 struct radeon_winsys_bo
*esgs_ring_bo
,
3329 uint32_t gsvs_ring_size
,
3330 struct radeon_winsys_bo
*gsvs_ring_bo
,
3331 uint32_t tess_factor_ring_size
,
3332 uint32_t tess_offchip_ring_offset
,
3333 uint32_t tess_offchip_ring_size
,
3334 struct radeon_winsys_bo
*tess_rings_bo
)
3336 uint32_t *desc
= &map
[4];
3339 uint64_t esgs_va
= radv_buffer_get_va(esgs_ring_bo
);
3341 /* stride 0, num records - size, add tid, swizzle, elsize4,
3344 desc
[1] = S_008F04_BASE_ADDRESS_HI(esgs_va
>> 32) |
3345 S_008F04_SWIZZLE_ENABLE(true);
3346 desc
[2] = esgs_ring_size
;
3347 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
3348 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
3349 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
3350 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
3351 S_008F0C_INDEX_STRIDE(3) |
3352 S_008F0C_ADD_TID_ENABLE(1);
3354 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
3355 desc
[3] |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT
) |
3356 S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED
) |
3357 S_008F0C_RESOURCE_LEVEL(1);
3359 desc
[3] |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
3360 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
3361 S_008F0C_ELEMENT_SIZE(1);
3364 /* GS entry for ES->GS ring */
3365 /* stride 0, num records - size, elsize0,
3368 desc
[5] = S_008F04_BASE_ADDRESS_HI(esgs_va
>> 32);
3369 desc
[6] = esgs_ring_size
;
3370 desc
[7] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
3371 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
3372 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
3373 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
);
3375 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
3376 desc
[7] |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT
) |
3377 S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED
) |
3378 S_008F0C_RESOURCE_LEVEL(1);
3380 desc
[7] |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
3381 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
);
3388 uint64_t gsvs_va
= radv_buffer_get_va(gsvs_ring_bo
);
3390 /* VS entry for GS->VS ring */
3391 /* stride 0, num records - size, elsize0,
3394 desc
[1] = S_008F04_BASE_ADDRESS_HI(gsvs_va
>> 32);
3395 desc
[2] = gsvs_ring_size
;
3396 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
3397 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
3398 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
3399 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
);
3401 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
3402 desc
[3] |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT
) |
3403 S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED
) |
3404 S_008F0C_RESOURCE_LEVEL(1);
3406 desc
[3] |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
3407 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
);
3410 /* stride gsvs_itemsize, num records 64
3411 elsize 4, index stride 16 */
3412 /* shader will patch stride and desc[2] */
3414 desc
[5] = S_008F04_BASE_ADDRESS_HI(gsvs_va
>> 32) |
3415 S_008F04_SWIZZLE_ENABLE(1);
3417 desc
[7] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
3418 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
3419 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
3420 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
3421 S_008F0C_INDEX_STRIDE(1) |
3422 S_008F0C_ADD_TID_ENABLE(true);
3424 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
3425 desc
[7] |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT
) |
3426 S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED
) |
3427 S_008F0C_RESOURCE_LEVEL(1);
3429 desc
[7] |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
3430 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
3431 S_008F0C_ELEMENT_SIZE(1);
3438 if (tess_rings_bo
) {
3439 uint64_t tess_va
= radv_buffer_get_va(tess_rings_bo
);
3440 uint64_t tess_offchip_va
= tess_va
+ tess_offchip_ring_offset
;
3443 desc
[1] = S_008F04_BASE_ADDRESS_HI(tess_va
>> 32);
3444 desc
[2] = tess_factor_ring_size
;
3445 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
3446 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
3447 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
3448 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
);
3450 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
3451 desc
[3] |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT
) |
3452 S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_RAW
) |
3453 S_008F0C_RESOURCE_LEVEL(1);
3455 desc
[3] |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
3456 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
);
3459 desc
[4] = tess_offchip_va
;
3460 desc
[5] = S_008F04_BASE_ADDRESS_HI(tess_offchip_va
>> 32);
3461 desc
[6] = tess_offchip_ring_size
;
3462 desc
[7] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
3463 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
3464 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
3465 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
);
3467 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
3468 desc
[7] |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT
) |
3469 S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_RAW
) |
3470 S_008F0C_RESOURCE_LEVEL(1);
3472 desc
[7] |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
3473 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
);
3479 if (add_sample_positions
) {
3480 /* add sample positions after all rings */
3481 memcpy(desc
, queue
->device
->sample_locations_1x
, 8);
3483 memcpy(desc
, queue
->device
->sample_locations_2x
, 16);
3485 memcpy(desc
, queue
->device
->sample_locations_4x
, 32);
3487 memcpy(desc
, queue
->device
->sample_locations_8x
, 64);
3492 radv_get_hs_offchip_param(struct radv_device
*device
, uint32_t *max_offchip_buffers_p
)
3494 bool double_offchip_buffers
= device
->physical_device
->rad_info
.chip_class
>= GFX7
&&
3495 device
->physical_device
->rad_info
.family
!= CHIP_CARRIZO
&&
3496 device
->physical_device
->rad_info
.family
!= CHIP_STONEY
;
3497 unsigned max_offchip_buffers_per_se
= double_offchip_buffers
? 128 : 64;
3498 unsigned max_offchip_buffers
;
3499 unsigned offchip_granularity
;
3500 unsigned hs_offchip_param
;
3504 * This must be one less than the maximum number due to a hw limitation.
3505 * Various hardware bugs need thGFX7
3508 * Vega10 should limit max_offchip_buffers to 508 (4 * 127).
3509 * Gfx7 should limit max_offchip_buffers to 508
3510 * Gfx6 should limit max_offchip_buffers to 126 (2 * 63)
3512 * Follow AMDVLK here.
3514 if (device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
3515 max_offchip_buffers_per_se
= 256;
3516 } else if (device
->physical_device
->rad_info
.family
== CHIP_VEGA10
||
3517 device
->physical_device
->rad_info
.chip_class
== GFX7
||
3518 device
->physical_device
->rad_info
.chip_class
== GFX6
)
3519 --max_offchip_buffers_per_se
;
3521 max_offchip_buffers
= max_offchip_buffers_per_se
*
3522 device
->physical_device
->rad_info
.max_se
;
3524 /* Hawaii has a bug with offchip buffers > 256 that can be worked
3525 * around by setting 4K granularity.
3527 if (device
->tess_offchip_block_dw_size
== 4096) {
3528 assert(device
->physical_device
->rad_info
.family
== CHIP_HAWAII
);
3529 offchip_granularity
= V_03093C_X_4K_DWORDS
;
3531 assert(device
->tess_offchip_block_dw_size
== 8192);
3532 offchip_granularity
= V_03093C_X_8K_DWORDS
;
3535 switch (device
->physical_device
->rad_info
.chip_class
) {
3537 max_offchip_buffers
= MIN2(max_offchip_buffers
, 126);
3542 max_offchip_buffers
= MIN2(max_offchip_buffers
, 508);
3550 *max_offchip_buffers_p
= max_offchip_buffers
;
3551 if (device
->physical_device
->rad_info
.chip_class
>= GFX7
) {
3552 if (device
->physical_device
->rad_info
.chip_class
>= GFX8
)
3553 --max_offchip_buffers
;
3555 S_03093C_OFFCHIP_BUFFERING(max_offchip_buffers
) |
3556 S_03093C_OFFCHIP_GRANULARITY(offchip_granularity
);
3559 S_0089B0_OFFCHIP_BUFFERING(max_offchip_buffers
);
3561 return hs_offchip_param
;
3565 radv_emit_gs_ring_sizes(struct radv_queue
*queue
, struct radeon_cmdbuf
*cs
,
3566 struct radeon_winsys_bo
*esgs_ring_bo
,
3567 uint32_t esgs_ring_size
,
3568 struct radeon_winsys_bo
*gsvs_ring_bo
,
3569 uint32_t gsvs_ring_size
)
3571 if (!esgs_ring_bo
&& !gsvs_ring_bo
)
3575 radv_cs_add_buffer(queue
->device
->ws
, cs
, esgs_ring_bo
);
3578 radv_cs_add_buffer(queue
->device
->ws
, cs
, gsvs_ring_bo
);
3580 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX7
) {
3581 radeon_set_uconfig_reg_seq(cs
, R_030900_VGT_ESGS_RING_SIZE
, 2);
3582 radeon_emit(cs
, esgs_ring_size
>> 8);
3583 radeon_emit(cs
, gsvs_ring_size
>> 8);
3585 radeon_set_config_reg_seq(cs
, R_0088C8_VGT_ESGS_RING_SIZE
, 2);
3586 radeon_emit(cs
, esgs_ring_size
>> 8);
3587 radeon_emit(cs
, gsvs_ring_size
>> 8);
3592 radv_emit_tess_factor_ring(struct radv_queue
*queue
, struct radeon_cmdbuf
*cs
,
3593 unsigned hs_offchip_param
, unsigned tf_ring_size
,
3594 struct radeon_winsys_bo
*tess_rings_bo
)
3601 tf_va
= radv_buffer_get_va(tess_rings_bo
);
3603 radv_cs_add_buffer(queue
->device
->ws
, cs
, tess_rings_bo
);
3605 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX7
) {
3606 radeon_set_uconfig_reg(cs
, R_030938_VGT_TF_RING_SIZE
,
3607 S_030938_SIZE(tf_ring_size
/ 4));
3608 radeon_set_uconfig_reg(cs
, R_030940_VGT_TF_MEMORY_BASE
,
3611 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
3612 radeon_set_uconfig_reg(cs
, R_030984_VGT_TF_MEMORY_BASE_HI_UMD
,
3613 S_030984_BASE_HI(tf_va
>> 40));
3614 } else if (queue
->device
->physical_device
->rad_info
.chip_class
== GFX9
) {
3615 radeon_set_uconfig_reg(cs
, R_030944_VGT_TF_MEMORY_BASE_HI
,
3616 S_030944_BASE_HI(tf_va
>> 40));
3618 radeon_set_uconfig_reg(cs
, R_03093C_VGT_HS_OFFCHIP_PARAM
,
3621 radeon_set_config_reg(cs
, R_008988_VGT_TF_RING_SIZE
,
3622 S_008988_SIZE(tf_ring_size
/ 4));
3623 radeon_set_config_reg(cs
, R_0089B8_VGT_TF_MEMORY_BASE
,
3625 radeon_set_config_reg(cs
, R_0089B0_VGT_HS_OFFCHIP_PARAM
,
3631 radv_emit_graphics_scratch(struct radv_queue
*queue
, struct radeon_cmdbuf
*cs
,
3632 uint32_t size_per_wave
, uint32_t waves
,
3633 struct radeon_winsys_bo
*scratch_bo
)
3635 if (queue
->queue_family_index
!= RADV_QUEUE_GENERAL
)
3641 radv_cs_add_buffer(queue
->device
->ws
, cs
, scratch_bo
);
3643 radeon_set_context_reg(cs
, R_0286E8_SPI_TMPRING_SIZE
,
3644 S_0286E8_WAVES(waves
) |
3645 S_0286E8_WAVESIZE(round_up_u32(size_per_wave
, 1024)));
3649 radv_emit_compute_scratch(struct radv_queue
*queue
, struct radeon_cmdbuf
*cs
,
3650 uint32_t size_per_wave
, uint32_t waves
,
3651 struct radeon_winsys_bo
*compute_scratch_bo
)
3653 uint64_t scratch_va
;
3655 if (!compute_scratch_bo
)
3658 scratch_va
= radv_buffer_get_va(compute_scratch_bo
);
3660 radv_cs_add_buffer(queue
->device
->ws
, cs
, compute_scratch_bo
);
3662 radeon_set_sh_reg_seq(cs
, R_00B900_COMPUTE_USER_DATA_0
, 2);
3663 radeon_emit(cs
, scratch_va
);
3664 radeon_emit(cs
, S_008F04_BASE_ADDRESS_HI(scratch_va
>> 32) |
3665 S_008F04_SWIZZLE_ENABLE(1));
3667 radeon_set_sh_reg(cs
, R_00B860_COMPUTE_TMPRING_SIZE
,
3668 S_00B860_WAVES(waves
) |
3669 S_00B860_WAVESIZE(round_up_u32(size_per_wave
, 1024)));
3673 radv_emit_global_shader_pointers(struct radv_queue
*queue
,
3674 struct radeon_cmdbuf
*cs
,
3675 struct radeon_winsys_bo
*descriptor_bo
)
3682 va
= radv_buffer_get_va(descriptor_bo
);
3684 radv_cs_add_buffer(queue
->device
->ws
, cs
, descriptor_bo
);
3686 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
3687 uint32_t regs
[] = {R_00B030_SPI_SHADER_USER_DATA_PS_0
,
3688 R_00B130_SPI_SHADER_USER_DATA_VS_0
,
3689 R_00B208_SPI_SHADER_USER_DATA_ADDR_LO_GS
,
3690 R_00B408_SPI_SHADER_USER_DATA_ADDR_LO_HS
};
3692 for (int i
= 0; i
< ARRAY_SIZE(regs
); ++i
) {
3693 radv_emit_shader_pointer(queue
->device
, cs
, regs
[i
],
3696 } else if (queue
->device
->physical_device
->rad_info
.chip_class
== GFX9
) {
3697 uint32_t regs
[] = {R_00B030_SPI_SHADER_USER_DATA_PS_0
,
3698 R_00B130_SPI_SHADER_USER_DATA_VS_0
,
3699 R_00B208_SPI_SHADER_USER_DATA_ADDR_LO_GS
,
3700 R_00B408_SPI_SHADER_USER_DATA_ADDR_LO_HS
};
3702 for (int i
= 0; i
< ARRAY_SIZE(regs
); ++i
) {
3703 radv_emit_shader_pointer(queue
->device
, cs
, regs
[i
],
3707 uint32_t regs
[] = {R_00B030_SPI_SHADER_USER_DATA_PS_0
,
3708 R_00B130_SPI_SHADER_USER_DATA_VS_0
,
3709 R_00B230_SPI_SHADER_USER_DATA_GS_0
,
3710 R_00B330_SPI_SHADER_USER_DATA_ES_0
,
3711 R_00B430_SPI_SHADER_USER_DATA_HS_0
,
3712 R_00B530_SPI_SHADER_USER_DATA_LS_0
};
3714 for (int i
= 0; i
< ARRAY_SIZE(regs
); ++i
) {
3715 radv_emit_shader_pointer(queue
->device
, cs
, regs
[i
],
3722 radv_init_graphics_state(struct radeon_cmdbuf
*cs
, struct radv_queue
*queue
)
3724 struct radv_device
*device
= queue
->device
;
3726 if (device
->gfx_init
) {
3727 uint64_t va
= radv_buffer_get_va(device
->gfx_init
);
3729 radeon_emit(cs
, PKT3(PKT3_INDIRECT_BUFFER_CIK
, 2, 0));
3730 radeon_emit(cs
, va
);
3731 radeon_emit(cs
, va
>> 32);
3732 radeon_emit(cs
, device
->gfx_init_size_dw
& 0xffff);
3734 radv_cs_add_buffer(device
->ws
, cs
, device
->gfx_init
);
3736 si_emit_graphics(device
, cs
);
3741 radv_init_compute_state(struct radeon_cmdbuf
*cs
, struct radv_queue
*queue
)
3743 struct radv_physical_device
*physical_device
= queue
->device
->physical_device
;
3744 si_emit_compute(physical_device
, cs
);
3748 radv_get_preamble_cs(struct radv_queue
*queue
,
3749 uint32_t scratch_size_per_wave
,
3750 uint32_t scratch_waves
,
3751 uint32_t compute_scratch_size_per_wave
,
3752 uint32_t compute_scratch_waves
,
3753 uint32_t esgs_ring_size
,
3754 uint32_t gsvs_ring_size
,
3755 bool needs_tess_rings
,
3758 bool needs_sample_positions
,
3759 struct radeon_cmdbuf
**initial_full_flush_preamble_cs
,
3760 struct radeon_cmdbuf
**initial_preamble_cs
,
3761 struct radeon_cmdbuf
**continue_preamble_cs
)
3763 struct radeon_winsys_bo
*scratch_bo
= NULL
;
3764 struct radeon_winsys_bo
*descriptor_bo
= NULL
;
3765 struct radeon_winsys_bo
*compute_scratch_bo
= NULL
;
3766 struct radeon_winsys_bo
*esgs_ring_bo
= NULL
;
3767 struct radeon_winsys_bo
*gsvs_ring_bo
= NULL
;
3768 struct radeon_winsys_bo
*tess_rings_bo
= NULL
;
3769 struct radeon_winsys_bo
*gds_bo
= NULL
;
3770 struct radeon_winsys_bo
*gds_oa_bo
= NULL
;
3771 struct radeon_cmdbuf
*dest_cs
[3] = {0};
3772 bool add_tess_rings
= false, add_gds
= false, add_gds_oa
= false, add_sample_positions
= false;
3773 unsigned tess_factor_ring_size
= 0, tess_offchip_ring_size
= 0;
3774 unsigned max_offchip_buffers
;
3775 unsigned hs_offchip_param
= 0;
3776 unsigned tess_offchip_ring_offset
;
3777 uint32_t ring_bo_flags
= RADEON_FLAG_NO_CPU_ACCESS
| RADEON_FLAG_NO_INTERPROCESS_SHARING
;
3778 if (!queue
->has_tess_rings
) {
3779 if (needs_tess_rings
)
3780 add_tess_rings
= true;
3782 if (!queue
->has_gds
) {
3786 if (!queue
->has_gds_oa
) {
3790 if (!queue
->has_sample_positions
) {
3791 if (needs_sample_positions
)
3792 add_sample_positions
= true;
3794 tess_factor_ring_size
= 32768 * queue
->device
->physical_device
->rad_info
.max_se
;
3795 hs_offchip_param
= radv_get_hs_offchip_param(queue
->device
,
3796 &max_offchip_buffers
);
3797 tess_offchip_ring_offset
= align(tess_factor_ring_size
, 64 * 1024);
3798 tess_offchip_ring_size
= max_offchip_buffers
*
3799 queue
->device
->tess_offchip_block_dw_size
* 4;
3801 scratch_size_per_wave
= MAX2(scratch_size_per_wave
, queue
->scratch_size_per_wave
);
3802 if (scratch_size_per_wave
)
3803 scratch_waves
= MIN2(scratch_waves
, UINT32_MAX
/ scratch_size_per_wave
);
3807 compute_scratch_size_per_wave
= MAX2(compute_scratch_size_per_wave
, queue
->compute_scratch_size_per_wave
);
3808 if (compute_scratch_size_per_wave
)
3809 compute_scratch_waves
= MIN2(compute_scratch_waves
, UINT32_MAX
/ compute_scratch_size_per_wave
);
3811 compute_scratch_waves
= 0;
3813 if (scratch_size_per_wave
<= queue
->scratch_size_per_wave
&&
3814 scratch_waves
<= queue
->scratch_waves
&&
3815 compute_scratch_size_per_wave
<= queue
->compute_scratch_size_per_wave
&&
3816 compute_scratch_waves
<= queue
->compute_scratch_waves
&&
3817 esgs_ring_size
<= queue
->esgs_ring_size
&&
3818 gsvs_ring_size
<= queue
->gsvs_ring_size
&&
3819 !add_tess_rings
&& !add_gds
&& !add_gds_oa
&& !add_sample_positions
&&
3820 queue
->initial_preamble_cs
) {
3821 *initial_full_flush_preamble_cs
= queue
->initial_full_flush_preamble_cs
;
3822 *initial_preamble_cs
= queue
->initial_preamble_cs
;
3823 *continue_preamble_cs
= queue
->continue_preamble_cs
;
3824 if (!scratch_size_per_wave
&& !compute_scratch_size_per_wave
&&
3825 !esgs_ring_size
&& !gsvs_ring_size
&& !needs_tess_rings
&&
3826 !needs_gds
&& !needs_gds_oa
&& !needs_sample_positions
)
3827 *continue_preamble_cs
= NULL
;
3831 uint32_t scratch_size
= scratch_size_per_wave
* scratch_waves
;
3832 uint32_t queue_scratch_size
= queue
->scratch_size_per_wave
* queue
->scratch_waves
;
3833 if (scratch_size
> queue_scratch_size
) {
3834 scratch_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
3839 RADV_BO_PRIORITY_SCRATCH
);
3843 scratch_bo
= queue
->scratch_bo
;
3845 uint32_t compute_scratch_size
= compute_scratch_size_per_wave
* compute_scratch_waves
;
3846 uint32_t compute_queue_scratch_size
= queue
->compute_scratch_size_per_wave
* queue
->compute_scratch_waves
;
3847 if (compute_scratch_size
> compute_queue_scratch_size
) {
3848 compute_scratch_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
3849 compute_scratch_size
,
3853 RADV_BO_PRIORITY_SCRATCH
);
3854 if (!compute_scratch_bo
)
3858 compute_scratch_bo
= queue
->compute_scratch_bo
;
3860 if (esgs_ring_size
> queue
->esgs_ring_size
) {
3861 esgs_ring_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
3866 RADV_BO_PRIORITY_SCRATCH
);
3870 esgs_ring_bo
= queue
->esgs_ring_bo
;
3871 esgs_ring_size
= queue
->esgs_ring_size
;
3874 if (gsvs_ring_size
> queue
->gsvs_ring_size
) {
3875 gsvs_ring_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
3880 RADV_BO_PRIORITY_SCRATCH
);
3884 gsvs_ring_bo
= queue
->gsvs_ring_bo
;
3885 gsvs_ring_size
= queue
->gsvs_ring_size
;
3888 if (add_tess_rings
) {
3889 tess_rings_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
3890 tess_offchip_ring_offset
+ tess_offchip_ring_size
,
3894 RADV_BO_PRIORITY_SCRATCH
);
3898 tess_rings_bo
= queue
->tess_rings_bo
;
3902 assert(queue
->device
->physical_device
->rad_info
.chip_class
>= GFX10
);
3904 /* 4 streamout GDS counters.
3905 * We need 256B (64 dw) of GDS, otherwise streamout hangs.
3907 gds_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
3911 RADV_BO_PRIORITY_SCRATCH
);
3915 gds_bo
= queue
->gds_bo
;
3919 assert(queue
->device
->physical_device
->rad_info
.chip_class
>= GFX10
);
3921 gds_oa_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
3925 RADV_BO_PRIORITY_SCRATCH
);
3929 gds_oa_bo
= queue
->gds_oa_bo
;
3932 if (scratch_bo
!= queue
->scratch_bo
||
3933 esgs_ring_bo
!= queue
->esgs_ring_bo
||
3934 gsvs_ring_bo
!= queue
->gsvs_ring_bo
||
3935 tess_rings_bo
!= queue
->tess_rings_bo
||
3936 add_sample_positions
) {
3938 if (gsvs_ring_bo
|| esgs_ring_bo
||
3939 tess_rings_bo
|| add_sample_positions
) {
3940 size
= 112; /* 2 dword + 2 padding + 4 dword * 6 */
3941 if (add_sample_positions
)
3942 size
+= 128; /* 64+32+16+8 = 120 bytes */
3944 else if (scratch_bo
)
3945 size
= 8; /* 2 dword */
3947 descriptor_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
3951 RADEON_FLAG_CPU_ACCESS
|
3952 RADEON_FLAG_NO_INTERPROCESS_SHARING
|
3953 RADEON_FLAG_READ_ONLY
,
3954 RADV_BO_PRIORITY_DESCRIPTOR
);
3958 descriptor_bo
= queue
->descriptor_bo
;
3960 if (descriptor_bo
!= queue
->descriptor_bo
) {
3961 uint32_t *map
= (uint32_t*)queue
->device
->ws
->buffer_map(descriptor_bo
);
3964 uint64_t scratch_va
= radv_buffer_get_va(scratch_bo
);
3965 uint32_t rsrc1
= S_008F04_BASE_ADDRESS_HI(scratch_va
>> 32) |
3966 S_008F04_SWIZZLE_ENABLE(1);
3967 map
[0] = scratch_va
;
3971 if (esgs_ring_bo
|| gsvs_ring_bo
|| tess_rings_bo
|| add_sample_positions
)
3972 fill_geom_tess_rings(queue
, map
, add_sample_positions
,
3973 esgs_ring_size
, esgs_ring_bo
,
3974 gsvs_ring_size
, gsvs_ring_bo
,
3975 tess_factor_ring_size
,
3976 tess_offchip_ring_offset
,
3977 tess_offchip_ring_size
,
3980 queue
->device
->ws
->buffer_unmap(descriptor_bo
);
3983 for(int i
= 0; i
< 3; ++i
) {
3984 struct radeon_cmdbuf
*cs
= NULL
;
3985 cs
= queue
->device
->ws
->cs_create(queue
->device
->ws
,
3986 queue
->queue_family_index
? RING_COMPUTE
: RING_GFX
);
3993 radv_cs_add_buffer(queue
->device
->ws
, cs
, scratch_bo
);
3995 /* Emit initial configuration. */
3996 switch (queue
->queue_family_index
) {
3997 case RADV_QUEUE_GENERAL
:
3998 radv_init_graphics_state(cs
, queue
);
4000 case RADV_QUEUE_COMPUTE
:
4001 radv_init_compute_state(cs
, queue
);
4003 case RADV_QUEUE_TRANSFER
:
4007 if (esgs_ring_bo
|| gsvs_ring_bo
|| tess_rings_bo
) {
4008 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
4009 radeon_emit(cs
, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
4011 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
4012 radeon_emit(cs
, EVENT_TYPE(V_028A90_VGT_FLUSH
) | EVENT_INDEX(0));
4015 radv_emit_gs_ring_sizes(queue
, cs
, esgs_ring_bo
, esgs_ring_size
,
4016 gsvs_ring_bo
, gsvs_ring_size
);
4017 radv_emit_tess_factor_ring(queue
, cs
, hs_offchip_param
,
4018 tess_factor_ring_size
, tess_rings_bo
);
4019 radv_emit_global_shader_pointers(queue
, cs
, descriptor_bo
);
4020 radv_emit_compute_scratch(queue
, cs
, compute_scratch_size_per_wave
,
4021 compute_scratch_waves
, compute_scratch_bo
);
4022 radv_emit_graphics_scratch(queue
, cs
, scratch_size_per_wave
,
4023 scratch_waves
, scratch_bo
);
4026 radv_cs_add_buffer(queue
->device
->ws
, cs
, gds_bo
);
4028 radv_cs_add_buffer(queue
->device
->ws
, cs
, gds_oa_bo
);
4030 if (queue
->device
->trace_bo
)
4031 radv_cs_add_buffer(queue
->device
->ws
, cs
, queue
->device
->trace_bo
);
4034 si_cs_emit_cache_flush(cs
,
4035 queue
->device
->physical_device
->rad_info
.chip_class
,
4037 queue
->queue_family_index
== RING_COMPUTE
&&
4038 queue
->device
->physical_device
->rad_info
.chip_class
>= GFX7
,
4039 (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
)) |
4040 RADV_CMD_FLAG_INV_ICACHE
|
4041 RADV_CMD_FLAG_INV_SCACHE
|
4042 RADV_CMD_FLAG_INV_VCACHE
|
4043 RADV_CMD_FLAG_INV_L2
|
4044 RADV_CMD_FLAG_START_PIPELINE_STATS
, 0);
4045 } else if (i
== 1) {
4046 si_cs_emit_cache_flush(cs
,
4047 queue
->device
->physical_device
->rad_info
.chip_class
,
4049 queue
->queue_family_index
== RING_COMPUTE
&&
4050 queue
->device
->physical_device
->rad_info
.chip_class
>= GFX7
,
4051 RADV_CMD_FLAG_INV_ICACHE
|
4052 RADV_CMD_FLAG_INV_SCACHE
|
4053 RADV_CMD_FLAG_INV_VCACHE
|
4054 RADV_CMD_FLAG_INV_L2
|
4055 RADV_CMD_FLAG_START_PIPELINE_STATS
, 0);
4058 if (!queue
->device
->ws
->cs_finalize(cs
))
4062 if (queue
->initial_full_flush_preamble_cs
)
4063 queue
->device
->ws
->cs_destroy(queue
->initial_full_flush_preamble_cs
);
4065 if (queue
->initial_preamble_cs
)
4066 queue
->device
->ws
->cs_destroy(queue
->initial_preamble_cs
);
4068 if (queue
->continue_preamble_cs
)
4069 queue
->device
->ws
->cs_destroy(queue
->continue_preamble_cs
);
4071 queue
->initial_full_flush_preamble_cs
= dest_cs
[0];
4072 queue
->initial_preamble_cs
= dest_cs
[1];
4073 queue
->continue_preamble_cs
= dest_cs
[2];
4075 if (scratch_bo
!= queue
->scratch_bo
) {
4076 if (queue
->scratch_bo
)
4077 queue
->device
->ws
->buffer_destroy(queue
->scratch_bo
);
4078 queue
->scratch_bo
= scratch_bo
;
4080 queue
->scratch_size_per_wave
= scratch_size_per_wave
;
4081 queue
->scratch_waves
= scratch_waves
;
4083 if (compute_scratch_bo
!= queue
->compute_scratch_bo
) {
4084 if (queue
->compute_scratch_bo
)
4085 queue
->device
->ws
->buffer_destroy(queue
->compute_scratch_bo
);
4086 queue
->compute_scratch_bo
= compute_scratch_bo
;
4088 queue
->compute_scratch_size_per_wave
= compute_scratch_size_per_wave
;
4089 queue
->compute_scratch_waves
= compute_scratch_waves
;
4091 if (esgs_ring_bo
!= queue
->esgs_ring_bo
) {
4092 if (queue
->esgs_ring_bo
)
4093 queue
->device
->ws
->buffer_destroy(queue
->esgs_ring_bo
);
4094 queue
->esgs_ring_bo
= esgs_ring_bo
;
4095 queue
->esgs_ring_size
= esgs_ring_size
;
4098 if (gsvs_ring_bo
!= queue
->gsvs_ring_bo
) {
4099 if (queue
->gsvs_ring_bo
)
4100 queue
->device
->ws
->buffer_destroy(queue
->gsvs_ring_bo
);
4101 queue
->gsvs_ring_bo
= gsvs_ring_bo
;
4102 queue
->gsvs_ring_size
= gsvs_ring_size
;
4105 if (tess_rings_bo
!= queue
->tess_rings_bo
) {
4106 queue
->tess_rings_bo
= tess_rings_bo
;
4107 queue
->has_tess_rings
= true;
4110 if (gds_bo
!= queue
->gds_bo
) {
4111 queue
->gds_bo
= gds_bo
;
4112 queue
->has_gds
= true;
4115 if (gds_oa_bo
!= queue
->gds_oa_bo
) {
4116 queue
->gds_oa_bo
= gds_oa_bo
;
4117 queue
->has_gds_oa
= true;
4120 if (descriptor_bo
!= queue
->descriptor_bo
) {
4121 if (queue
->descriptor_bo
)
4122 queue
->device
->ws
->buffer_destroy(queue
->descriptor_bo
);
4124 queue
->descriptor_bo
= descriptor_bo
;
4127 if (add_sample_positions
)
4128 queue
->has_sample_positions
= true;
4130 *initial_full_flush_preamble_cs
= queue
->initial_full_flush_preamble_cs
;
4131 *initial_preamble_cs
= queue
->initial_preamble_cs
;
4132 *continue_preamble_cs
= queue
->continue_preamble_cs
;
4133 if (!scratch_size
&& !compute_scratch_size
&& !esgs_ring_size
&& !gsvs_ring_size
)
4134 *continue_preamble_cs
= NULL
;
4137 for (int i
= 0; i
< ARRAY_SIZE(dest_cs
); ++i
)
4139 queue
->device
->ws
->cs_destroy(dest_cs
[i
]);
4140 if (descriptor_bo
&& descriptor_bo
!= queue
->descriptor_bo
)
4141 queue
->device
->ws
->buffer_destroy(descriptor_bo
);
4142 if (scratch_bo
&& scratch_bo
!= queue
->scratch_bo
)
4143 queue
->device
->ws
->buffer_destroy(scratch_bo
);
4144 if (compute_scratch_bo
&& compute_scratch_bo
!= queue
->compute_scratch_bo
)
4145 queue
->device
->ws
->buffer_destroy(compute_scratch_bo
);
4146 if (esgs_ring_bo
&& esgs_ring_bo
!= queue
->esgs_ring_bo
)
4147 queue
->device
->ws
->buffer_destroy(esgs_ring_bo
);
4148 if (gsvs_ring_bo
&& gsvs_ring_bo
!= queue
->gsvs_ring_bo
)
4149 queue
->device
->ws
->buffer_destroy(gsvs_ring_bo
);
4150 if (tess_rings_bo
&& tess_rings_bo
!= queue
->tess_rings_bo
)
4151 queue
->device
->ws
->buffer_destroy(tess_rings_bo
);
4152 if (gds_bo
&& gds_bo
!= queue
->gds_bo
)
4153 queue
->device
->ws
->buffer_destroy(gds_bo
);
4154 if (gds_oa_bo
&& gds_oa_bo
!= queue
->gds_oa_bo
)
4155 queue
->device
->ws
->buffer_destroy(gds_oa_bo
);
4157 return vk_error(queue
->device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
4160 static VkResult
radv_alloc_sem_counts(struct radv_device
*device
,
4161 struct radv_winsys_sem_counts
*counts
,
4163 struct radv_semaphore_part
**sems
,
4164 const uint64_t *timeline_values
,
4168 int syncobj_idx
= 0, sem_idx
= 0;
4170 if (num_sems
== 0 && _fence
== VK_NULL_HANDLE
)
4173 for (uint32_t i
= 0; i
< num_sems
; i
++) {
4174 switch(sems
[i
]->kind
) {
4175 case RADV_SEMAPHORE_SYNCOBJ
:
4176 counts
->syncobj_count
++;
4178 case RADV_SEMAPHORE_WINSYS
:
4179 counts
->sem_count
++;
4181 case RADV_SEMAPHORE_NONE
:
4183 case RADV_SEMAPHORE_TIMELINE
:
4184 counts
->syncobj_count
++;
4189 if (_fence
!= VK_NULL_HANDLE
) {
4190 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
4191 if (fence
->temp_syncobj
|| fence
->syncobj
)
4192 counts
->syncobj_count
++;
4195 if (counts
->syncobj_count
) {
4196 counts
->syncobj
= (uint32_t *)malloc(sizeof(uint32_t) * counts
->syncobj_count
);
4197 if (!counts
->syncobj
)
4198 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
4201 if (counts
->sem_count
) {
4202 counts
->sem
= (struct radeon_winsys_sem
**)malloc(sizeof(struct radeon_winsys_sem
*) * counts
->sem_count
);
4204 free(counts
->syncobj
);
4205 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
4209 for (uint32_t i
= 0; i
< num_sems
; i
++) {
4210 switch(sems
[i
]->kind
) {
4211 case RADV_SEMAPHORE_NONE
:
4212 unreachable("Empty semaphore");
4214 case RADV_SEMAPHORE_SYNCOBJ
:
4215 counts
->syncobj
[syncobj_idx
++] = sems
[i
]->syncobj
;
4217 case RADV_SEMAPHORE_WINSYS
:
4218 counts
->sem
[sem_idx
++] = sems
[i
]->ws_sem
;
4220 case RADV_SEMAPHORE_TIMELINE
: {
4221 pthread_mutex_lock(&sems
[i
]->timeline
.mutex
);
4222 struct radv_timeline_point
*point
= NULL
;
4224 point
= radv_timeline_add_point_locked(device
, &sems
[i
]->timeline
, timeline_values
[i
]);
4226 point
= radv_timeline_find_point_at_least_locked(device
, &sems
[i
]->timeline
, timeline_values
[i
]);
4229 pthread_mutex_unlock(&sems
[i
]->timeline
.mutex
);
4232 counts
->syncobj
[syncobj_idx
++] = point
->syncobj
;
4234 /* Explicitly remove the semaphore so we might not find
4235 * a point later post-submit. */
4243 if (_fence
!= VK_NULL_HANDLE
) {
4244 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
4245 if (fence
->temp_syncobj
)
4246 counts
->syncobj
[syncobj_idx
++] = fence
->temp_syncobj
;
4247 else if (fence
->syncobj
)
4248 counts
->syncobj
[syncobj_idx
++] = fence
->syncobj
;
4251 assert(syncobj_idx
<= counts
->syncobj_count
);
4252 counts
->syncobj_count
= syncobj_idx
;
4258 radv_free_sem_info(struct radv_winsys_sem_info
*sem_info
)
4260 free(sem_info
->wait
.syncobj
);
4261 free(sem_info
->wait
.sem
);
4262 free(sem_info
->signal
.syncobj
);
4263 free(sem_info
->signal
.sem
);
4267 static void radv_free_temp_syncobjs(struct radv_device
*device
,
4269 struct radv_semaphore_part
*sems
)
4271 for (uint32_t i
= 0; i
< num_sems
; i
++) {
4272 radv_destroy_semaphore_part(device
, sems
+ i
);
4277 radv_alloc_sem_info(struct radv_device
*device
,
4278 struct radv_winsys_sem_info
*sem_info
,
4280 struct radv_semaphore_part
**wait_sems
,
4281 const uint64_t *wait_values
,
4282 int num_signal_sems
,
4283 struct radv_semaphore_part
**signal_sems
,
4284 const uint64_t *signal_values
,
4288 memset(sem_info
, 0, sizeof(*sem_info
));
4290 ret
= radv_alloc_sem_counts(device
, &sem_info
->wait
, num_wait_sems
, wait_sems
, wait_values
, VK_NULL_HANDLE
, false);
4293 ret
= radv_alloc_sem_counts(device
, &sem_info
->signal
, num_signal_sems
, signal_sems
, signal_values
, fence
, true);
4295 radv_free_sem_info(sem_info
);
4297 /* caller can override these */
4298 sem_info
->cs_emit_wait
= true;
4299 sem_info
->cs_emit_signal
= true;
4304 radv_finalize_timelines(struct radv_device
*device
,
4305 uint32_t num_wait_sems
,
4306 struct radv_semaphore_part
**wait_sems
,
4307 const uint64_t *wait_values
,
4308 uint32_t num_signal_sems
,
4309 struct radv_semaphore_part
**signal_sems
,
4310 const uint64_t *signal_values
,
4311 struct list_head
*processing_list
)
4313 for (uint32_t i
= 0; i
< num_wait_sems
; ++i
) {
4314 if (wait_sems
[i
] && wait_sems
[i
]->kind
== RADV_SEMAPHORE_TIMELINE
) {
4315 pthread_mutex_lock(&wait_sems
[i
]->timeline
.mutex
);
4316 struct radv_timeline_point
*point
=
4317 radv_timeline_find_point_at_least_locked(device
, &wait_sems
[i
]->timeline
, wait_values
[i
]);
4318 point
->wait_count
-= 2;
4319 pthread_mutex_unlock(&wait_sems
[i
]->timeline
.mutex
);
4322 for (uint32_t i
= 0; i
< num_signal_sems
; ++i
) {
4323 if (signal_sems
[i
] && signal_sems
[i
]->kind
== RADV_SEMAPHORE_TIMELINE
) {
4324 pthread_mutex_lock(&signal_sems
[i
]->timeline
.mutex
);
4325 struct radv_timeline_point
*point
=
4326 radv_timeline_find_point_at_least_locked(device
, &signal_sems
[i
]->timeline
, signal_values
[i
]);
4327 signal_sems
[i
]->timeline
.highest_submitted
=
4328 MAX2(signal_sems
[i
]->timeline
.highest_submitted
, point
->value
);
4329 point
->wait_count
-= 2;
4330 radv_timeline_trigger_waiters_locked(&signal_sems
[i
]->timeline
, processing_list
);
4331 pthread_mutex_unlock(&signal_sems
[i
]->timeline
.mutex
);
4337 radv_sparse_buffer_bind_memory(struct radv_device
*device
,
4338 const VkSparseBufferMemoryBindInfo
*bind
)
4340 RADV_FROM_HANDLE(radv_buffer
, buffer
, bind
->buffer
);
4342 for (uint32_t i
= 0; i
< bind
->bindCount
; ++i
) {
4343 struct radv_device_memory
*mem
= NULL
;
4345 if (bind
->pBinds
[i
].memory
!= VK_NULL_HANDLE
)
4346 mem
= radv_device_memory_from_handle(bind
->pBinds
[i
].memory
);
4348 device
->ws
->buffer_virtual_bind(buffer
->bo
,
4349 bind
->pBinds
[i
].resourceOffset
,
4350 bind
->pBinds
[i
].size
,
4351 mem
? mem
->bo
: NULL
,
4352 bind
->pBinds
[i
].memoryOffset
);
4357 radv_sparse_image_opaque_bind_memory(struct radv_device
*device
,
4358 const VkSparseImageOpaqueMemoryBindInfo
*bind
)
4360 RADV_FROM_HANDLE(radv_image
, image
, bind
->image
);
4362 for (uint32_t i
= 0; i
< bind
->bindCount
; ++i
) {
4363 struct radv_device_memory
*mem
= NULL
;
4365 if (bind
->pBinds
[i
].memory
!= VK_NULL_HANDLE
)
4366 mem
= radv_device_memory_from_handle(bind
->pBinds
[i
].memory
);
4368 device
->ws
->buffer_virtual_bind(image
->bo
,
4369 bind
->pBinds
[i
].resourceOffset
,
4370 bind
->pBinds
[i
].size
,
4371 mem
? mem
->bo
: NULL
,
4372 bind
->pBinds
[i
].memoryOffset
);
4377 radv_get_preambles(struct radv_queue
*queue
,
4378 const VkCommandBuffer
*cmd_buffers
,
4379 uint32_t cmd_buffer_count
,
4380 struct radeon_cmdbuf
**initial_full_flush_preamble_cs
,
4381 struct radeon_cmdbuf
**initial_preamble_cs
,
4382 struct radeon_cmdbuf
**continue_preamble_cs
)
4384 uint32_t scratch_size_per_wave
= 0, waves_wanted
= 0;
4385 uint32_t compute_scratch_size_per_wave
= 0, compute_waves_wanted
= 0;
4386 uint32_t esgs_ring_size
= 0, gsvs_ring_size
= 0;
4387 bool tess_rings_needed
= false;
4388 bool gds_needed
= false;
4389 bool gds_oa_needed
= false;
4390 bool sample_positions_needed
= false;
4392 for (uint32_t j
= 0; j
< cmd_buffer_count
; j
++) {
4393 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
,
4396 scratch_size_per_wave
= MAX2(scratch_size_per_wave
, cmd_buffer
->scratch_size_per_wave_needed
);
4397 waves_wanted
= MAX2(waves_wanted
, cmd_buffer
->scratch_waves_wanted
);
4398 compute_scratch_size_per_wave
= MAX2(compute_scratch_size_per_wave
,
4399 cmd_buffer
->compute_scratch_size_per_wave_needed
);
4400 compute_waves_wanted
= MAX2(compute_waves_wanted
,
4401 cmd_buffer
->compute_scratch_waves_wanted
);
4402 esgs_ring_size
= MAX2(esgs_ring_size
, cmd_buffer
->esgs_ring_size_needed
);
4403 gsvs_ring_size
= MAX2(gsvs_ring_size
, cmd_buffer
->gsvs_ring_size_needed
);
4404 tess_rings_needed
|= cmd_buffer
->tess_rings_needed
;
4405 gds_needed
|= cmd_buffer
->gds_needed
;
4406 gds_oa_needed
|= cmd_buffer
->gds_oa_needed
;
4407 sample_positions_needed
|= cmd_buffer
->sample_positions_needed
;
4410 return radv_get_preamble_cs(queue
, scratch_size_per_wave
, waves_wanted
,
4411 compute_scratch_size_per_wave
, compute_waves_wanted
,
4412 esgs_ring_size
, gsvs_ring_size
, tess_rings_needed
,
4413 gds_needed
, gds_oa_needed
, sample_positions_needed
,
4414 initial_full_flush_preamble_cs
,
4415 initial_preamble_cs
, continue_preamble_cs
);
4418 struct radv_deferred_queue_submission
{
4419 struct radv_queue
*queue
;
4420 VkCommandBuffer
*cmd_buffers
;
4421 uint32_t cmd_buffer_count
;
4423 /* Sparse bindings that happen on a queue. */
4424 VkSparseBufferMemoryBindInfo
*buffer_binds
;
4425 uint32_t buffer_bind_count
;
4426 VkSparseImageOpaqueMemoryBindInfo
*image_opaque_binds
;
4427 uint32_t image_opaque_bind_count
;
4430 VkShaderStageFlags wait_dst_stage_mask
;
4431 struct radv_semaphore_part
**wait_semaphores
;
4432 uint32_t wait_semaphore_count
;
4433 struct radv_semaphore_part
**signal_semaphores
;
4434 uint32_t signal_semaphore_count
;
4437 uint64_t *wait_values
;
4438 uint64_t *signal_values
;
4440 struct radv_semaphore_part
*temporary_semaphore_parts
;
4441 uint32_t temporary_semaphore_part_count
;
4443 struct list_head queue_pending_list
;
4444 uint32_t submission_wait_count
;
4445 struct radv_timeline_waiter
*wait_nodes
;
4447 struct list_head processing_list
;
4450 struct radv_queue_submission
{
4451 const VkCommandBuffer
*cmd_buffers
;
4452 uint32_t cmd_buffer_count
;
4454 /* Sparse bindings that happen on a queue. */
4455 const VkSparseBufferMemoryBindInfo
*buffer_binds
;
4456 uint32_t buffer_bind_count
;
4457 const VkSparseImageOpaqueMemoryBindInfo
*image_opaque_binds
;
4458 uint32_t image_opaque_bind_count
;
4461 VkPipelineStageFlags wait_dst_stage_mask
;
4462 const VkSemaphore
*wait_semaphores
;
4463 uint32_t wait_semaphore_count
;
4464 const VkSemaphore
*signal_semaphores
;
4465 uint32_t signal_semaphore_count
;
4468 const uint64_t *wait_values
;
4469 uint32_t wait_value_count
;
4470 const uint64_t *signal_values
;
4471 uint32_t signal_value_count
;
4475 radv_create_deferred_submission(struct radv_queue
*queue
,
4476 const struct radv_queue_submission
*submission
,
4477 struct radv_deferred_queue_submission
**out
)
4479 struct radv_deferred_queue_submission
*deferred
= NULL
;
4480 size_t size
= sizeof(struct radv_deferred_queue_submission
);
4482 uint32_t temporary_count
= 0;
4483 for (uint32_t i
= 0; i
< submission
->wait_semaphore_count
; ++i
) {
4484 RADV_FROM_HANDLE(radv_semaphore
, semaphore
, submission
->wait_semaphores
[i
]);
4485 if (semaphore
->temporary
.kind
!= RADV_SEMAPHORE_NONE
)
4489 size
+= submission
->cmd_buffer_count
* sizeof(VkCommandBuffer
);
4490 size
+= submission
->buffer_bind_count
* sizeof(VkSparseBufferMemoryBindInfo
);
4491 size
+= submission
->image_opaque_bind_count
* sizeof(VkSparseImageOpaqueMemoryBindInfo
);
4492 size
+= submission
->wait_semaphore_count
* sizeof(struct radv_semaphore_part
*);
4493 size
+= temporary_count
* sizeof(struct radv_semaphore_part
);
4494 size
+= submission
->signal_semaphore_count
* sizeof(struct radv_semaphore_part
*);
4495 size
+= submission
->wait_value_count
* sizeof(uint64_t);
4496 size
+= submission
->signal_value_count
* sizeof(uint64_t);
4497 size
+= submission
->wait_semaphore_count
* sizeof(struct radv_timeline_waiter
);
4499 deferred
= calloc(1, size
);
4501 return VK_ERROR_OUT_OF_HOST_MEMORY
;
4503 deferred
->queue
= queue
;
4505 deferred
->cmd_buffers
= (void*)(deferred
+ 1);
4506 deferred
->cmd_buffer_count
= submission
->cmd_buffer_count
;
4507 memcpy(deferred
->cmd_buffers
, submission
->cmd_buffers
,
4508 submission
->cmd_buffer_count
* sizeof(*deferred
->cmd_buffers
));
4510 deferred
->buffer_binds
= (void*)(deferred
->cmd_buffers
+ submission
->cmd_buffer_count
);
4511 deferred
->buffer_bind_count
= submission
->buffer_bind_count
;
4512 memcpy(deferred
->buffer_binds
, submission
->buffer_binds
,
4513 submission
->buffer_bind_count
* sizeof(*deferred
->buffer_binds
));
4515 deferred
->image_opaque_binds
= (void*)(deferred
->buffer_binds
+ submission
->buffer_bind_count
);
4516 deferred
->image_opaque_bind_count
= submission
->image_opaque_bind_count
;
4517 memcpy(deferred
->image_opaque_binds
, submission
->image_opaque_binds
,
4518 submission
->image_opaque_bind_count
* sizeof(*deferred
->image_opaque_binds
));
4520 deferred
->flush_caches
= submission
->flush_caches
;
4521 deferred
->wait_dst_stage_mask
= submission
->wait_dst_stage_mask
;
4523 deferred
->wait_semaphores
= (void*)(deferred
->image_opaque_binds
+ deferred
->image_opaque_bind_count
);
4524 deferred
->wait_semaphore_count
= submission
->wait_semaphore_count
;
4526 deferred
->signal_semaphores
= (void*)(deferred
->wait_semaphores
+ deferred
->wait_semaphore_count
);
4527 deferred
->signal_semaphore_count
= submission
->signal_semaphore_count
;
4529 deferred
->fence
= submission
->fence
;
4531 deferred
->temporary_semaphore_parts
= (void*)(deferred
->signal_semaphores
+ deferred
->signal_semaphore_count
);
4532 deferred
->temporary_semaphore_part_count
= temporary_count
;
4534 uint32_t temporary_idx
= 0;
4535 for (uint32_t i
= 0; i
< submission
->wait_semaphore_count
; ++i
) {
4536 RADV_FROM_HANDLE(radv_semaphore
, semaphore
, submission
->wait_semaphores
[i
]);
4537 if (semaphore
->temporary
.kind
!= RADV_SEMAPHORE_NONE
) {
4538 deferred
->wait_semaphores
[i
] = &deferred
->temporary_semaphore_parts
[temporary_idx
];
4539 deferred
->temporary_semaphore_parts
[temporary_idx
] = semaphore
->temporary
;
4540 semaphore
->temporary
.kind
= RADV_SEMAPHORE_NONE
;
4543 deferred
->wait_semaphores
[i
] = &semaphore
->permanent
;
4546 for (uint32_t i
= 0; i
< submission
->signal_semaphore_count
; ++i
) {
4547 RADV_FROM_HANDLE(radv_semaphore
, semaphore
, submission
->signal_semaphores
[i
]);
4548 if (semaphore
->temporary
.kind
!= RADV_SEMAPHORE_NONE
) {
4549 deferred
->signal_semaphores
[i
] = &semaphore
->temporary
;
4551 deferred
->signal_semaphores
[i
] = &semaphore
->permanent
;
4555 deferred
->wait_values
= (void*)(deferred
->temporary_semaphore_parts
+ temporary_count
);
4556 memcpy(deferred
->wait_values
, submission
->wait_values
, submission
->wait_value_count
* sizeof(uint64_t));
4557 deferred
->signal_values
= deferred
->wait_values
+ submission
->wait_value_count
;
4558 memcpy(deferred
->signal_values
, submission
->signal_values
, submission
->signal_value_count
* sizeof(uint64_t));
4560 deferred
->wait_nodes
= (void*)(deferred
->signal_values
+ submission
->signal_value_count
);
4561 /* This is worst-case. radv_queue_enqueue_submission will fill in further, but this
4562 * ensure the submission is not accidentally triggered early when adding wait timelines. */
4563 deferred
->submission_wait_count
= 1 + submission
->wait_semaphore_count
;
4570 radv_queue_enqueue_submission(struct radv_deferred_queue_submission
*submission
,
4571 struct list_head
*processing_list
)
4573 uint32_t wait_cnt
= 0;
4574 struct radv_timeline_waiter
*waiter
= submission
->wait_nodes
;
4575 for (uint32_t i
= 0; i
< submission
->wait_semaphore_count
; ++i
) {
4576 if (submission
->wait_semaphores
[i
]->kind
== RADV_SEMAPHORE_TIMELINE
) {
4577 pthread_mutex_lock(&submission
->wait_semaphores
[i
]->timeline
.mutex
);
4578 if (submission
->wait_semaphores
[i
]->timeline
.highest_submitted
< submission
->wait_values
[i
]) {
4580 waiter
->value
= submission
->wait_values
[i
];
4581 waiter
->submission
= submission
;
4582 list_addtail(&waiter
->list
, &submission
->wait_semaphores
[i
]->timeline
.waiters
);
4585 pthread_mutex_unlock(&submission
->wait_semaphores
[i
]->timeline
.mutex
);
4589 pthread_mutex_lock(&submission
->queue
->pending_mutex
);
4591 bool is_first
= list_is_empty(&submission
->queue
->pending_submissions
);
4592 list_addtail(&submission
->queue_pending_list
, &submission
->queue
->pending_submissions
);
4594 pthread_mutex_unlock(&submission
->queue
->pending_mutex
);
4596 /* If there is already a submission in the queue, that will decrement the counter by 1 when
4597 * submitted, but if the queue was empty, we decrement ourselves as there is no previous
4599 uint32_t decrement
= submission
->wait_semaphore_count
- wait_cnt
+ (is_first
? 1 : 0);
4600 if (__atomic_sub_fetch(&submission
->submission_wait_count
, decrement
, __ATOMIC_ACQ_REL
) == 0) {
4601 list_addtail(&submission
->processing_list
, processing_list
);
4606 radv_queue_submission_update_queue(struct radv_deferred_queue_submission
*submission
,
4607 struct list_head
*processing_list
)
4609 pthread_mutex_lock(&submission
->queue
->pending_mutex
);
4610 list_del(&submission
->queue_pending_list
);
4612 /* trigger the next submission in the queue. */
4613 if (!list_is_empty(&submission
->queue
->pending_submissions
)) {
4614 struct radv_deferred_queue_submission
*next_submission
=
4615 list_first_entry(&submission
->queue
->pending_submissions
,
4616 struct radv_deferred_queue_submission
,
4617 queue_pending_list
);
4618 if (p_atomic_dec_zero(&next_submission
->submission_wait_count
)) {
4619 list_addtail(&next_submission
->processing_list
, processing_list
);
4622 pthread_mutex_unlock(&submission
->queue
->pending_mutex
);
4624 pthread_cond_broadcast(&submission
->queue
->device
->timeline_cond
);
4628 radv_queue_submit_deferred(struct radv_deferred_queue_submission
*submission
,
4629 struct list_head
*processing_list
)
4631 RADV_FROM_HANDLE(radv_fence
, fence
, submission
->fence
);
4632 struct radv_queue
*queue
= submission
->queue
;
4633 struct radeon_winsys_ctx
*ctx
= queue
->hw_ctx
;
4634 uint32_t max_cs_submission
= queue
->device
->trace_bo
? 1 : RADV_MAX_IBS_PER_SUBMIT
;
4635 struct radeon_winsys_fence
*base_fence
= fence
? fence
->fence
: NULL
;
4636 bool do_flush
= submission
->flush_caches
|| submission
->wait_dst_stage_mask
;
4637 bool can_patch
= true;
4639 struct radv_winsys_sem_info sem_info
;
4642 struct radeon_cmdbuf
*initial_preamble_cs
= NULL
;
4643 struct radeon_cmdbuf
*initial_flush_preamble_cs
= NULL
;
4644 struct radeon_cmdbuf
*continue_preamble_cs
= NULL
;
4646 result
= radv_get_preambles(queue
, submission
->cmd_buffers
,
4647 submission
->cmd_buffer_count
,
4648 &initial_preamble_cs
,
4649 &initial_flush_preamble_cs
,
4650 &continue_preamble_cs
);
4651 if (result
!= VK_SUCCESS
)
4654 result
= radv_alloc_sem_info(queue
->device
,
4656 submission
->wait_semaphore_count
,
4657 submission
->wait_semaphores
,
4658 submission
->wait_values
,
4659 submission
->signal_semaphore_count
,
4660 submission
->signal_semaphores
,
4661 submission
->signal_values
,
4663 if (result
!= VK_SUCCESS
)
4666 for (uint32_t i
= 0; i
< submission
->buffer_bind_count
; ++i
) {
4667 radv_sparse_buffer_bind_memory(queue
->device
,
4668 submission
->buffer_binds
+ i
);
4671 for (uint32_t i
= 0; i
< submission
->image_opaque_bind_count
; ++i
) {
4672 radv_sparse_image_opaque_bind_memory(queue
->device
,
4673 submission
->image_opaque_binds
+ i
);
4676 if (!submission
->cmd_buffer_count
) {
4677 ret
= queue
->device
->ws
->cs_submit(ctx
, queue
->queue_idx
,
4678 &queue
->device
->empty_cs
[queue
->queue_family_index
],
4683 radv_loge("failed to submit CS\n");
4689 struct radeon_cmdbuf
**cs_array
= malloc(sizeof(struct radeon_cmdbuf
*) *
4690 (submission
->cmd_buffer_count
));
4692 for (uint32_t j
= 0; j
< submission
->cmd_buffer_count
; j
++) {
4693 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, submission
->cmd_buffers
[j
]);
4694 assert(cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_PRIMARY
);
4696 cs_array
[j
] = cmd_buffer
->cs
;
4697 if ((cmd_buffer
->usage_flags
& VK_COMMAND_BUFFER_USAGE_SIMULTANEOUS_USE_BIT
))
4700 cmd_buffer
->status
= RADV_CMD_BUFFER_STATUS_PENDING
;
4703 for (uint32_t j
= 0; j
< submission
->cmd_buffer_count
; j
+= advance
) {
4704 struct radeon_cmdbuf
*initial_preamble
= (do_flush
&& !j
) ? initial_flush_preamble_cs
: initial_preamble_cs
;
4705 const struct radv_winsys_bo_list
*bo_list
= NULL
;
4707 advance
= MIN2(max_cs_submission
,
4708 submission
->cmd_buffer_count
- j
);
4710 if (queue
->device
->trace_bo
)
4711 *queue
->device
->trace_id_ptr
= 0;
4713 sem_info
.cs_emit_wait
= j
== 0;
4714 sem_info
.cs_emit_signal
= j
+ advance
== submission
->cmd_buffer_count
;
4716 if (unlikely(queue
->device
->use_global_bo_list
)) {
4717 pthread_mutex_lock(&queue
->device
->bo_list
.mutex
);
4718 bo_list
= &queue
->device
->bo_list
.list
;
4721 ret
= queue
->device
->ws
->cs_submit(ctx
, queue
->queue_idx
, cs_array
+ j
,
4722 advance
, initial_preamble
, continue_preamble_cs
,
4724 can_patch
, base_fence
);
4726 if (unlikely(queue
->device
->use_global_bo_list
))
4727 pthread_mutex_unlock(&queue
->device
->bo_list
.mutex
);
4730 radv_loge("failed to submit CS\n");
4733 if (queue
->device
->trace_bo
) {
4734 radv_check_gpu_hangs(queue
, cs_array
[j
]);
4742 radv_free_temp_syncobjs(queue
->device
,
4743 submission
->temporary_semaphore_part_count
,
4744 submission
->temporary_semaphore_parts
);
4745 radv_finalize_timelines(queue
->device
,
4746 submission
->wait_semaphore_count
,
4747 submission
->wait_semaphores
,
4748 submission
->wait_values
,
4749 submission
->signal_semaphore_count
,
4750 submission
->signal_semaphores
,
4751 submission
->signal_values
,
4753 /* Has to happen after timeline finalization to make sure the
4754 * condition variable is only triggered when timelines and queue have
4756 radv_queue_submission_update_queue(submission
, processing_list
);
4757 radv_free_sem_info(&sem_info
);
4762 radv_free_temp_syncobjs(queue
->device
,
4763 submission
->temporary_semaphore_part_count
,
4764 submission
->temporary_semaphore_parts
);
4766 return VK_ERROR_DEVICE_LOST
;
4770 radv_process_submissions(struct list_head
*processing_list
)
4772 while(!list_is_empty(processing_list
)) {
4773 struct radv_deferred_queue_submission
*submission
=
4774 list_first_entry(processing_list
, struct radv_deferred_queue_submission
, processing_list
);
4775 list_del(&submission
->processing_list
);
4777 VkResult result
= radv_queue_submit_deferred(submission
, processing_list
);
4778 if (result
!= VK_SUCCESS
)
4784 static VkResult
radv_queue_submit(struct radv_queue
*queue
,
4785 const struct radv_queue_submission
*submission
)
4787 struct radv_deferred_queue_submission
*deferred
= NULL
;
4789 VkResult result
= radv_create_deferred_submission(queue
, submission
, &deferred
);
4790 if (result
!= VK_SUCCESS
)
4793 struct list_head processing_list
;
4794 list_inithead(&processing_list
);
4796 radv_queue_enqueue_submission(deferred
, &processing_list
);
4797 return radv_process_submissions(&processing_list
);
4801 radv_queue_internal_submit(struct radv_queue
*queue
, struct radeon_cmdbuf
*cs
)
4803 struct radeon_winsys_ctx
*ctx
= queue
->hw_ctx
;
4804 struct radv_winsys_sem_info sem_info
;
4808 result
= radv_alloc_sem_info(queue
->device
, &sem_info
, 0, NULL
, 0, 0,
4809 0, NULL
, VK_NULL_HANDLE
);
4810 if (result
!= VK_SUCCESS
)
4813 ret
= queue
->device
->ws
->cs_submit(ctx
, queue
->queue_idx
, &cs
, 1, NULL
,
4814 NULL
, &sem_info
, NULL
, false, NULL
);
4815 radv_free_sem_info(&sem_info
);
4819 /* Signals fence as soon as all the work currently put on queue is done. */
4820 static VkResult
radv_signal_fence(struct radv_queue
*queue
,
4823 return radv_queue_submit(queue
, &(struct radv_queue_submission
) {
4828 static bool radv_submit_has_effects(const VkSubmitInfo
*info
)
4830 return info
->commandBufferCount
||
4831 info
->waitSemaphoreCount
||
4832 info
->signalSemaphoreCount
;
4835 VkResult
radv_QueueSubmit(
4837 uint32_t submitCount
,
4838 const VkSubmitInfo
* pSubmits
,
4841 RADV_FROM_HANDLE(radv_queue
, queue
, _queue
);
4843 uint32_t fence_idx
= 0;
4844 bool flushed_caches
= false;
4846 if (fence
!= VK_NULL_HANDLE
) {
4847 for (uint32_t i
= 0; i
< submitCount
; ++i
)
4848 if (radv_submit_has_effects(pSubmits
+ i
))
4851 fence_idx
= UINT32_MAX
;
4853 for (uint32_t i
= 0; i
< submitCount
; i
++) {
4854 if (!radv_submit_has_effects(pSubmits
+ i
) && fence_idx
!= i
)
4857 VkPipelineStageFlags wait_dst_stage_mask
= 0;
4858 for (unsigned j
= 0; j
< pSubmits
[i
].waitSemaphoreCount
; ++j
) {
4859 wait_dst_stage_mask
|= pSubmits
[i
].pWaitDstStageMask
[j
];
4862 const VkTimelineSemaphoreSubmitInfo
*timeline_info
=
4863 vk_find_struct_const(pSubmits
[i
].pNext
, TIMELINE_SEMAPHORE_SUBMIT_INFO
);
4865 result
= radv_queue_submit(queue
, &(struct radv_queue_submission
) {
4866 .cmd_buffers
= pSubmits
[i
].pCommandBuffers
,
4867 .cmd_buffer_count
= pSubmits
[i
].commandBufferCount
,
4868 .wait_dst_stage_mask
= wait_dst_stage_mask
,
4869 .flush_caches
= !flushed_caches
,
4870 .wait_semaphores
= pSubmits
[i
].pWaitSemaphores
,
4871 .wait_semaphore_count
= pSubmits
[i
].waitSemaphoreCount
,
4872 .signal_semaphores
= pSubmits
[i
].pSignalSemaphores
,
4873 .signal_semaphore_count
= pSubmits
[i
].signalSemaphoreCount
,
4874 .fence
= i
== fence_idx
? fence
: VK_NULL_HANDLE
,
4875 .wait_values
= timeline_info
? timeline_info
->pWaitSemaphoreValues
: NULL
,
4876 .wait_value_count
= timeline_info
&& timeline_info
->pWaitSemaphoreValues
? timeline_info
->waitSemaphoreValueCount
: 0,
4877 .signal_values
= timeline_info
? timeline_info
->pSignalSemaphoreValues
: NULL
,
4878 .signal_value_count
= timeline_info
&& timeline_info
->pSignalSemaphoreValues
? timeline_info
->signalSemaphoreValueCount
: 0,
4880 if (result
!= VK_SUCCESS
)
4883 flushed_caches
= true;
4886 if (fence
!= VK_NULL_HANDLE
&& !submitCount
) {
4887 result
= radv_signal_fence(queue
, fence
);
4888 if (result
!= VK_SUCCESS
)
4895 VkResult
radv_QueueWaitIdle(
4898 RADV_FROM_HANDLE(radv_queue
, queue
, _queue
);
4900 pthread_mutex_lock(&queue
->pending_mutex
);
4901 while (!list_is_empty(&queue
->pending_submissions
)) {
4902 pthread_cond_wait(&queue
->device
->timeline_cond
, &queue
->pending_mutex
);
4904 pthread_mutex_unlock(&queue
->pending_mutex
);
4906 queue
->device
->ws
->ctx_wait_idle(queue
->hw_ctx
,
4907 radv_queue_family_to_ring(queue
->queue_family_index
),
4912 VkResult
radv_DeviceWaitIdle(
4915 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4917 for (unsigned i
= 0; i
< RADV_MAX_QUEUE_FAMILIES
; i
++) {
4918 for (unsigned q
= 0; q
< device
->queue_count
[i
]; q
++) {
4919 radv_QueueWaitIdle(radv_queue_to_handle(&device
->queues
[i
][q
]));
4925 VkResult
radv_EnumerateInstanceExtensionProperties(
4926 const char* pLayerName
,
4927 uint32_t* pPropertyCount
,
4928 VkExtensionProperties
* pProperties
)
4930 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
4932 for (int i
= 0; i
< RADV_INSTANCE_EXTENSION_COUNT
; i
++) {
4933 if (radv_supported_instance_extensions
.extensions
[i
]) {
4934 vk_outarray_append(&out
, prop
) {
4935 *prop
= radv_instance_extensions
[i
];
4940 return vk_outarray_status(&out
);
4943 VkResult
radv_EnumerateDeviceExtensionProperties(
4944 VkPhysicalDevice physicalDevice
,
4945 const char* pLayerName
,
4946 uint32_t* pPropertyCount
,
4947 VkExtensionProperties
* pProperties
)
4949 RADV_FROM_HANDLE(radv_physical_device
, device
, physicalDevice
);
4950 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
4952 for (int i
= 0; i
< RADV_DEVICE_EXTENSION_COUNT
; i
++) {
4953 if (device
->supported_extensions
.extensions
[i
]) {
4954 vk_outarray_append(&out
, prop
) {
4955 *prop
= radv_device_extensions
[i
];
4960 return vk_outarray_status(&out
);
4963 PFN_vkVoidFunction
radv_GetInstanceProcAddr(
4964 VkInstance _instance
,
4967 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
4969 /* The Vulkan 1.0 spec for vkGetInstanceProcAddr has a table of exactly
4970 * when we have to return valid function pointers, NULL, or it's left
4971 * undefined. See the table for exact details.
4976 #define LOOKUP_RADV_ENTRYPOINT(entrypoint) \
4977 if (strcmp(pName, "vk" #entrypoint) == 0) \
4978 return (PFN_vkVoidFunction)radv_##entrypoint
4980 LOOKUP_RADV_ENTRYPOINT(EnumerateInstanceExtensionProperties
);
4981 LOOKUP_RADV_ENTRYPOINT(EnumerateInstanceLayerProperties
);
4982 LOOKUP_RADV_ENTRYPOINT(EnumerateInstanceVersion
);
4983 LOOKUP_RADV_ENTRYPOINT(CreateInstance
);
4985 /* GetInstanceProcAddr() can also be called with a NULL instance.
4986 * See https://gitlab.khronos.org/vulkan/vulkan/issues/2057
4988 LOOKUP_RADV_ENTRYPOINT(GetInstanceProcAddr
);
4990 #undef LOOKUP_RADV_ENTRYPOINT
4992 if (instance
== NULL
)
4995 int idx
= radv_get_instance_entrypoint_index(pName
);
4997 return instance
->dispatch
.entrypoints
[idx
];
4999 idx
= radv_get_physical_device_entrypoint_index(pName
);
5001 return instance
->physical_device_dispatch
.entrypoints
[idx
];
5003 idx
= radv_get_device_entrypoint_index(pName
);
5005 return instance
->device_dispatch
.entrypoints
[idx
];
5010 /* The loader wants us to expose a second GetInstanceProcAddr function
5011 * to work around certain LD_PRELOAD issues seen in apps.
5014 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
5015 VkInstance instance
,
5019 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
5020 VkInstance instance
,
5023 return radv_GetInstanceProcAddr(instance
, pName
);
5027 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetPhysicalDeviceProcAddr(
5028 VkInstance _instance
,
5032 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetPhysicalDeviceProcAddr(
5033 VkInstance _instance
,
5036 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
5038 if (!pName
|| !instance
)
5041 int idx
= radv_get_physical_device_entrypoint_index(pName
);
5045 return instance
->physical_device_dispatch
.entrypoints
[idx
];
5048 PFN_vkVoidFunction
radv_GetDeviceProcAddr(
5052 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5054 if (!device
|| !pName
)
5057 int idx
= radv_get_device_entrypoint_index(pName
);
5061 return device
->dispatch
.entrypoints
[idx
];
5064 bool radv_get_memory_fd(struct radv_device
*device
,
5065 struct radv_device_memory
*memory
,
5068 struct radeon_bo_metadata metadata
;
5070 if (memory
->image
) {
5071 if (memory
->image
->tiling
!= VK_IMAGE_TILING_LINEAR
)
5072 radv_init_metadata(device
, memory
->image
, &metadata
);
5073 device
->ws
->buffer_set_metadata(memory
->bo
, &metadata
);
5076 return device
->ws
->buffer_get_fd(device
->ws
, memory
->bo
,
5081 static void radv_free_memory(struct radv_device
*device
,
5082 const VkAllocationCallbacks
* pAllocator
,
5083 struct radv_device_memory
*mem
)
5088 #if RADV_SUPPORT_ANDROID_HARDWARE_BUFFER
5089 if (mem
->android_hardware_buffer
)
5090 AHardwareBuffer_release(mem
->android_hardware_buffer
);
5094 if (device
->overallocation_disallowed
) {
5095 mtx_lock(&device
->overallocation_mutex
);
5096 device
->allocated_memory_size
[mem
->heap_index
] -= mem
->alloc_size
;
5097 mtx_unlock(&device
->overallocation_mutex
);
5100 radv_bo_list_remove(device
, mem
->bo
);
5101 device
->ws
->buffer_destroy(mem
->bo
);
5105 vk_free2(&device
->alloc
, pAllocator
, mem
);
5108 static VkResult
radv_alloc_memory(struct radv_device
*device
,
5109 const VkMemoryAllocateInfo
* pAllocateInfo
,
5110 const VkAllocationCallbacks
* pAllocator
,
5111 VkDeviceMemory
* pMem
)
5113 struct radv_device_memory
*mem
;
5115 enum radeon_bo_domain domain
;
5118 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
5120 const VkImportMemoryFdInfoKHR
*import_info
=
5121 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_FD_INFO_KHR
);
5122 const VkMemoryDedicatedAllocateInfo
*dedicate_info
=
5123 vk_find_struct_const(pAllocateInfo
->pNext
, MEMORY_DEDICATED_ALLOCATE_INFO
);
5124 const VkExportMemoryAllocateInfo
*export_info
=
5125 vk_find_struct_const(pAllocateInfo
->pNext
, EXPORT_MEMORY_ALLOCATE_INFO
);
5126 const struct VkImportAndroidHardwareBufferInfoANDROID
*ahb_import_info
=
5127 vk_find_struct_const(pAllocateInfo
->pNext
,
5128 IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID
);
5129 const VkImportMemoryHostPointerInfoEXT
*host_ptr_info
=
5130 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_HOST_POINTER_INFO_EXT
);
5132 const struct wsi_memory_allocate_info
*wsi_info
=
5133 vk_find_struct_const(pAllocateInfo
->pNext
, WSI_MEMORY_ALLOCATE_INFO_MESA
);
5135 if (pAllocateInfo
->allocationSize
== 0 && !ahb_import_info
&&
5136 !(export_info
&& (export_info
->handleTypes
& VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID
))) {
5137 /* Apparently, this is allowed */
5138 *pMem
= VK_NULL_HANDLE
;
5142 mem
= vk_zalloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
5143 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
5145 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
5147 if (wsi_info
&& wsi_info
->implicit_sync
)
5148 flags
|= RADEON_FLAG_IMPLICIT_SYNC
;
5150 if (dedicate_info
) {
5151 mem
->image
= radv_image_from_handle(dedicate_info
->image
);
5152 mem
->buffer
= radv_buffer_from_handle(dedicate_info
->buffer
);
5158 float priority_float
= 0.5;
5159 const struct VkMemoryPriorityAllocateInfoEXT
*priority_ext
=
5160 vk_find_struct_const(pAllocateInfo
->pNext
,
5161 MEMORY_PRIORITY_ALLOCATE_INFO_EXT
);
5163 priority_float
= priority_ext
->priority
;
5165 unsigned priority
= MIN2(RADV_BO_PRIORITY_APPLICATION_MAX
- 1,
5166 (int)(priority_float
* RADV_BO_PRIORITY_APPLICATION_MAX
));
5168 mem
->user_ptr
= NULL
;
5171 #if RADV_SUPPORT_ANDROID_HARDWARE_BUFFER
5172 mem
->android_hardware_buffer
= NULL
;
5175 if (ahb_import_info
) {
5176 result
= radv_import_ahb_memory(device
, mem
, priority
, ahb_import_info
);
5177 if (result
!= VK_SUCCESS
)
5179 } else if(export_info
&& (export_info
->handleTypes
& VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID
)) {
5180 result
= radv_create_ahb_memory(device
, mem
, priority
, pAllocateInfo
);
5181 if (result
!= VK_SUCCESS
)
5183 } else if (import_info
) {
5184 assert(import_info
->handleType
==
5185 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
5186 import_info
->handleType
==
5187 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
5188 mem
->bo
= device
->ws
->buffer_from_fd(device
->ws
, import_info
->fd
,
5191 result
= VK_ERROR_INVALID_EXTERNAL_HANDLE
;
5194 close(import_info
->fd
);
5196 } else if (host_ptr_info
) {
5197 assert(host_ptr_info
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
);
5198 mem
->bo
= device
->ws
->buffer_from_ptr(device
->ws
, host_ptr_info
->pHostPointer
,
5199 pAllocateInfo
->allocationSize
,
5202 result
= VK_ERROR_INVALID_EXTERNAL_HANDLE
;
5205 mem
->user_ptr
= host_ptr_info
->pHostPointer
;
5208 uint64_t alloc_size
= align_u64(pAllocateInfo
->allocationSize
, 4096);
5209 uint32_t heap_index
;
5211 heap_index
= device
->physical_device
->memory_properties
.memoryTypes
[pAllocateInfo
->memoryTypeIndex
].heapIndex
;
5212 domain
= device
->physical_device
->memory_domains
[pAllocateInfo
->memoryTypeIndex
];
5213 flags
|= device
->physical_device
->memory_flags
[pAllocateInfo
->memoryTypeIndex
];
5215 if (!dedicate_info
&& !import_info
&& (!export_info
|| !export_info
->handleTypes
)) {
5216 flags
|= RADEON_FLAG_NO_INTERPROCESS_SHARING
;
5217 if (device
->use_global_bo_list
) {
5218 flags
|= RADEON_FLAG_PREFER_LOCAL_BO
;
5222 if (device
->overallocation_disallowed
) {
5223 uint64_t total_size
=
5224 device
->physical_device
->memory_properties
.memoryHeaps
[heap_index
].size
;
5226 mtx_lock(&device
->overallocation_mutex
);
5227 if (device
->allocated_memory_size
[heap_index
] + alloc_size
> total_size
) {
5228 mtx_unlock(&device
->overallocation_mutex
);
5229 result
= VK_ERROR_OUT_OF_DEVICE_MEMORY
;
5232 device
->allocated_memory_size
[heap_index
] += alloc_size
;
5233 mtx_unlock(&device
->overallocation_mutex
);
5236 mem
->bo
= device
->ws
->buffer_create(device
->ws
, alloc_size
, device
->physical_device
->rad_info
.max_alignment
,
5237 domain
, flags
, priority
);
5240 if (device
->overallocation_disallowed
) {
5241 mtx_lock(&device
->overallocation_mutex
);
5242 device
->allocated_memory_size
[heap_index
] -= alloc_size
;
5243 mtx_unlock(&device
->overallocation_mutex
);
5245 result
= VK_ERROR_OUT_OF_DEVICE_MEMORY
;
5249 mem
->heap_index
= heap_index
;
5250 mem
->alloc_size
= alloc_size
;
5254 result
= radv_bo_list_add(device
, mem
->bo
);
5255 if (result
!= VK_SUCCESS
)
5259 *pMem
= radv_device_memory_to_handle(mem
);
5264 radv_free_memory(device
, pAllocator
,mem
);
5269 VkResult
radv_AllocateMemory(
5271 const VkMemoryAllocateInfo
* pAllocateInfo
,
5272 const VkAllocationCallbacks
* pAllocator
,
5273 VkDeviceMemory
* pMem
)
5275 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5276 return radv_alloc_memory(device
, pAllocateInfo
, pAllocator
, pMem
);
5279 void radv_FreeMemory(
5281 VkDeviceMemory _mem
,
5282 const VkAllocationCallbacks
* pAllocator
)
5284 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5285 RADV_FROM_HANDLE(radv_device_memory
, mem
, _mem
);
5287 radv_free_memory(device
, pAllocator
, mem
);
5290 VkResult
radv_MapMemory(
5292 VkDeviceMemory _memory
,
5293 VkDeviceSize offset
,
5295 VkMemoryMapFlags flags
,
5298 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5299 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
5307 *ppData
= mem
->user_ptr
;
5309 *ppData
= device
->ws
->buffer_map(mem
->bo
);
5316 return vk_error(device
->instance
, VK_ERROR_MEMORY_MAP_FAILED
);
5319 void radv_UnmapMemory(
5321 VkDeviceMemory _memory
)
5323 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5324 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
5329 if (mem
->user_ptr
== NULL
)
5330 device
->ws
->buffer_unmap(mem
->bo
);
5333 VkResult
radv_FlushMappedMemoryRanges(
5335 uint32_t memoryRangeCount
,
5336 const VkMappedMemoryRange
* pMemoryRanges
)
5341 VkResult
radv_InvalidateMappedMemoryRanges(
5343 uint32_t memoryRangeCount
,
5344 const VkMappedMemoryRange
* pMemoryRanges
)
5349 void radv_GetBufferMemoryRequirements(
5352 VkMemoryRequirements
* pMemoryRequirements
)
5354 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5355 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
5357 pMemoryRequirements
->memoryTypeBits
= (1u << device
->physical_device
->memory_properties
.memoryTypeCount
) - 1;
5359 if (buffer
->flags
& VK_BUFFER_CREATE_SPARSE_BINDING_BIT
)
5360 pMemoryRequirements
->alignment
= 4096;
5362 pMemoryRequirements
->alignment
= 16;
5364 pMemoryRequirements
->size
= align64(buffer
->size
, pMemoryRequirements
->alignment
);
5367 void radv_GetBufferMemoryRequirements2(
5369 const VkBufferMemoryRequirementsInfo2
*pInfo
,
5370 VkMemoryRequirements2
*pMemoryRequirements
)
5372 radv_GetBufferMemoryRequirements(device
, pInfo
->buffer
,
5373 &pMemoryRequirements
->memoryRequirements
);
5374 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
5375 switch (ext
->sType
) {
5376 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
5377 VkMemoryDedicatedRequirements
*req
=
5378 (VkMemoryDedicatedRequirements
*) ext
;
5379 req
->requiresDedicatedAllocation
= false;
5380 req
->prefersDedicatedAllocation
= req
->requiresDedicatedAllocation
;
5389 void radv_GetImageMemoryRequirements(
5392 VkMemoryRequirements
* pMemoryRequirements
)
5394 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5395 RADV_FROM_HANDLE(radv_image
, image
, _image
);
5397 pMemoryRequirements
->memoryTypeBits
= (1u << device
->physical_device
->memory_properties
.memoryTypeCount
) - 1;
5399 pMemoryRequirements
->size
= image
->size
;
5400 pMemoryRequirements
->alignment
= image
->alignment
;
5403 void radv_GetImageMemoryRequirements2(
5405 const VkImageMemoryRequirementsInfo2
*pInfo
,
5406 VkMemoryRequirements2
*pMemoryRequirements
)
5408 radv_GetImageMemoryRequirements(device
, pInfo
->image
,
5409 &pMemoryRequirements
->memoryRequirements
);
5411 RADV_FROM_HANDLE(radv_image
, image
, pInfo
->image
);
5413 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
5414 switch (ext
->sType
) {
5415 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
5416 VkMemoryDedicatedRequirements
*req
=
5417 (VkMemoryDedicatedRequirements
*) ext
;
5418 req
->requiresDedicatedAllocation
= image
->shareable
&&
5419 image
->tiling
!= VK_IMAGE_TILING_LINEAR
;
5420 req
->prefersDedicatedAllocation
= req
->requiresDedicatedAllocation
;
5429 void radv_GetImageSparseMemoryRequirements(
5432 uint32_t* pSparseMemoryRequirementCount
,
5433 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
5438 void radv_GetImageSparseMemoryRequirements2(
5440 const VkImageSparseMemoryRequirementsInfo2
*pInfo
,
5441 uint32_t* pSparseMemoryRequirementCount
,
5442 VkSparseImageMemoryRequirements2
*pSparseMemoryRequirements
)
5447 void radv_GetDeviceMemoryCommitment(
5449 VkDeviceMemory memory
,
5450 VkDeviceSize
* pCommittedMemoryInBytes
)
5452 *pCommittedMemoryInBytes
= 0;
5455 VkResult
radv_BindBufferMemory2(VkDevice device
,
5456 uint32_t bindInfoCount
,
5457 const VkBindBufferMemoryInfo
*pBindInfos
)
5459 for (uint32_t i
= 0; i
< bindInfoCount
; ++i
) {
5460 RADV_FROM_HANDLE(radv_device_memory
, mem
, pBindInfos
[i
].memory
);
5461 RADV_FROM_HANDLE(radv_buffer
, buffer
, pBindInfos
[i
].buffer
);
5464 buffer
->bo
= mem
->bo
;
5465 buffer
->offset
= pBindInfos
[i
].memoryOffset
;
5473 VkResult
radv_BindBufferMemory(
5476 VkDeviceMemory memory
,
5477 VkDeviceSize memoryOffset
)
5479 const VkBindBufferMemoryInfo info
= {
5480 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
,
5483 .memoryOffset
= memoryOffset
5486 return radv_BindBufferMemory2(device
, 1, &info
);
5489 VkResult
radv_BindImageMemory2(VkDevice device
,
5490 uint32_t bindInfoCount
,
5491 const VkBindImageMemoryInfo
*pBindInfos
)
5493 for (uint32_t i
= 0; i
< bindInfoCount
; ++i
) {
5494 RADV_FROM_HANDLE(radv_device_memory
, mem
, pBindInfos
[i
].memory
);
5495 RADV_FROM_HANDLE(radv_image
, image
, pBindInfos
[i
].image
);
5498 image
->bo
= mem
->bo
;
5499 image
->offset
= pBindInfos
[i
].memoryOffset
;
5509 VkResult
radv_BindImageMemory(
5512 VkDeviceMemory memory
,
5513 VkDeviceSize memoryOffset
)
5515 const VkBindImageMemoryInfo info
= {
5516 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
,
5519 .memoryOffset
= memoryOffset
5522 return radv_BindImageMemory2(device
, 1, &info
);
5525 static bool radv_sparse_bind_has_effects(const VkBindSparseInfo
*info
)
5527 return info
->bufferBindCount
||
5528 info
->imageOpaqueBindCount
||
5529 info
->imageBindCount
||
5530 info
->waitSemaphoreCount
||
5531 info
->signalSemaphoreCount
;
5534 VkResult
radv_QueueBindSparse(
5536 uint32_t bindInfoCount
,
5537 const VkBindSparseInfo
* pBindInfo
,
5540 RADV_FROM_HANDLE(radv_queue
, queue
, _queue
);
5542 uint32_t fence_idx
= 0;
5544 if (fence
!= VK_NULL_HANDLE
) {
5545 for (uint32_t i
= 0; i
< bindInfoCount
; ++i
)
5546 if (radv_sparse_bind_has_effects(pBindInfo
+ i
))
5549 fence_idx
= UINT32_MAX
;
5551 for (uint32_t i
= 0; i
< bindInfoCount
; ++i
) {
5552 if (i
!= fence_idx
&& !radv_sparse_bind_has_effects(pBindInfo
+ i
))
5555 const VkTimelineSemaphoreSubmitInfo
*timeline_info
=
5556 vk_find_struct_const(pBindInfo
[i
].pNext
, TIMELINE_SEMAPHORE_SUBMIT_INFO
);
5558 VkResult result
= radv_queue_submit(queue
, &(struct radv_queue_submission
) {
5559 .buffer_binds
= pBindInfo
[i
].pBufferBinds
,
5560 .buffer_bind_count
= pBindInfo
[i
].bufferBindCount
,
5561 .image_opaque_binds
= pBindInfo
[i
].pImageOpaqueBinds
,
5562 .image_opaque_bind_count
= pBindInfo
[i
].imageOpaqueBindCount
,
5563 .wait_semaphores
= pBindInfo
[i
].pWaitSemaphores
,
5564 .wait_semaphore_count
= pBindInfo
[i
].waitSemaphoreCount
,
5565 .signal_semaphores
= pBindInfo
[i
].pSignalSemaphores
,
5566 .signal_semaphore_count
= pBindInfo
[i
].signalSemaphoreCount
,
5567 .fence
= i
== fence_idx
? fence
: VK_NULL_HANDLE
,
5568 .wait_values
= timeline_info
? timeline_info
->pWaitSemaphoreValues
: NULL
,
5569 .wait_value_count
= timeline_info
&& timeline_info
->pWaitSemaphoreValues
? timeline_info
->waitSemaphoreValueCount
: 0,
5570 .signal_values
= timeline_info
? timeline_info
->pSignalSemaphoreValues
: NULL
,
5571 .signal_value_count
= timeline_info
&& timeline_info
->pSignalSemaphoreValues
? timeline_info
->signalSemaphoreValueCount
: 0,
5574 if (result
!= VK_SUCCESS
)
5578 if (fence
!= VK_NULL_HANDLE
&& !bindInfoCount
) {
5579 result
= radv_signal_fence(queue
, fence
);
5580 if (result
!= VK_SUCCESS
)
5587 VkResult
radv_CreateFence(
5589 const VkFenceCreateInfo
* pCreateInfo
,
5590 const VkAllocationCallbacks
* pAllocator
,
5593 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5594 const VkExportFenceCreateInfo
*export
=
5595 vk_find_struct_const(pCreateInfo
->pNext
, EXPORT_FENCE_CREATE_INFO
);
5596 VkExternalFenceHandleTypeFlags handleTypes
=
5597 export
? export
->handleTypes
: 0;
5599 struct radv_fence
*fence
= vk_alloc2(&device
->alloc
, pAllocator
,
5601 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
5604 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
5606 fence
->fence_wsi
= NULL
;
5607 fence
->temp_syncobj
= 0;
5608 if (device
->always_use_syncobj
|| handleTypes
) {
5609 int ret
= device
->ws
->create_syncobj(device
->ws
, &fence
->syncobj
);
5611 vk_free2(&device
->alloc
, pAllocator
, fence
);
5612 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
5614 if (pCreateInfo
->flags
& VK_FENCE_CREATE_SIGNALED_BIT
) {
5615 device
->ws
->signal_syncobj(device
->ws
, fence
->syncobj
);
5617 fence
->fence
= NULL
;
5619 fence
->fence
= device
->ws
->create_fence();
5620 if (!fence
->fence
) {
5621 vk_free2(&device
->alloc
, pAllocator
, fence
);
5622 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
5625 if (pCreateInfo
->flags
& VK_FENCE_CREATE_SIGNALED_BIT
)
5626 device
->ws
->signal_fence(fence
->fence
);
5629 *pFence
= radv_fence_to_handle(fence
);
5634 void radv_DestroyFence(
5637 const VkAllocationCallbacks
* pAllocator
)
5639 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5640 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
5645 if (fence
->temp_syncobj
)
5646 device
->ws
->destroy_syncobj(device
->ws
, fence
->temp_syncobj
);
5648 device
->ws
->destroy_syncobj(device
->ws
, fence
->syncobj
);
5650 device
->ws
->destroy_fence(fence
->fence
);
5651 if (fence
->fence_wsi
)
5652 fence
->fence_wsi
->destroy(fence
->fence_wsi
);
5653 vk_free2(&device
->alloc
, pAllocator
, fence
);
5657 uint64_t radv_get_current_time(void)
5660 clock_gettime(CLOCK_MONOTONIC
, &tv
);
5661 return tv
.tv_nsec
+ tv
.tv_sec
*1000000000ull;
5664 static uint64_t radv_get_absolute_timeout(uint64_t timeout
)
5666 uint64_t current_time
= radv_get_current_time();
5668 timeout
= MIN2(UINT64_MAX
- current_time
, timeout
);
5670 return current_time
+ timeout
;
5674 static bool radv_all_fences_plain_and_submitted(struct radv_device
*device
,
5675 uint32_t fenceCount
, const VkFence
*pFences
)
5677 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
5678 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
5679 if (fence
->fence
== NULL
|| fence
->syncobj
||
5680 fence
->temp_syncobj
|| fence
->fence_wsi
||
5681 (!device
->ws
->is_fence_waitable(fence
->fence
)))
5687 static bool radv_all_fences_syncobj(uint32_t fenceCount
, const VkFence
*pFences
)
5689 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
5690 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
5691 if (fence
->syncobj
== 0 && fence
->temp_syncobj
== 0)
5697 VkResult
radv_WaitForFences(
5699 uint32_t fenceCount
,
5700 const VkFence
* pFences
,
5704 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5705 timeout
= radv_get_absolute_timeout(timeout
);
5707 if (device
->always_use_syncobj
&&
5708 radv_all_fences_syncobj(fenceCount
, pFences
))
5710 uint32_t *handles
= malloc(sizeof(uint32_t) * fenceCount
);
5712 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
5714 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
5715 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
5716 handles
[i
] = fence
->temp_syncobj
? fence
->temp_syncobj
: fence
->syncobj
;
5719 bool success
= device
->ws
->wait_syncobj(device
->ws
, handles
, fenceCount
, waitAll
, timeout
);
5722 return success
? VK_SUCCESS
: VK_TIMEOUT
;
5725 if (!waitAll
&& fenceCount
> 1) {
5726 /* Not doing this by default for waitAll, due to needing to allocate twice. */
5727 if (device
->physical_device
->rad_info
.drm_minor
>= 10 && radv_all_fences_plain_and_submitted(device
, fenceCount
, pFences
)) {
5728 uint32_t wait_count
= 0;
5729 struct radeon_winsys_fence
**fences
= malloc(sizeof(struct radeon_winsys_fence
*) * fenceCount
);
5731 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
5733 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
5734 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
5736 if (device
->ws
->fence_wait(device
->ws
, fence
->fence
, false, 0)) {
5741 fences
[wait_count
++] = fence
->fence
;
5744 bool success
= device
->ws
->fences_wait(device
->ws
, fences
, wait_count
,
5745 waitAll
, timeout
- radv_get_current_time());
5748 return success
? VK_SUCCESS
: VK_TIMEOUT
;
5751 while(radv_get_current_time() <= timeout
) {
5752 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
5753 if (radv_GetFenceStatus(_device
, pFences
[i
]) == VK_SUCCESS
)
5760 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
5761 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
5762 bool expired
= false;
5764 if (fence
->temp_syncobj
) {
5765 if (!device
->ws
->wait_syncobj(device
->ws
, &fence
->temp_syncobj
, 1, true, timeout
))
5770 if (fence
->syncobj
) {
5771 if (!device
->ws
->wait_syncobj(device
->ws
, &fence
->syncobj
, 1, true, timeout
))
5777 if (!device
->ws
->is_fence_waitable(fence
->fence
)) {
5778 while(!device
->ws
->is_fence_waitable(fence
->fence
) &&
5779 radv_get_current_time() <= timeout
)
5783 expired
= device
->ws
->fence_wait(device
->ws
,
5790 if (fence
->fence_wsi
) {
5791 VkResult result
= fence
->fence_wsi
->wait(fence
->fence_wsi
, timeout
);
5792 if (result
!= VK_SUCCESS
)
5800 VkResult
radv_ResetFences(VkDevice _device
,
5801 uint32_t fenceCount
,
5802 const VkFence
*pFences
)
5804 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5806 for (unsigned i
= 0; i
< fenceCount
; ++i
) {
5807 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
5809 device
->ws
->reset_fence(fence
->fence
);
5811 /* Per spec, we first restore the permanent payload, and then reset, so
5812 * having a temp syncobj should not skip resetting the permanent syncobj. */
5813 if (fence
->temp_syncobj
) {
5814 device
->ws
->destroy_syncobj(device
->ws
, fence
->temp_syncobj
);
5815 fence
->temp_syncobj
= 0;
5818 if (fence
->syncobj
) {
5819 device
->ws
->reset_syncobj(device
->ws
, fence
->syncobj
);
5826 VkResult
radv_GetFenceStatus(VkDevice _device
, VkFence _fence
)
5828 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5829 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
5831 if (fence
->temp_syncobj
) {
5832 bool success
= device
->ws
->wait_syncobj(device
->ws
, &fence
->temp_syncobj
, 1, true, 0);
5833 return success
? VK_SUCCESS
: VK_NOT_READY
;
5836 if (fence
->syncobj
) {
5837 bool success
= device
->ws
->wait_syncobj(device
->ws
, &fence
->syncobj
, 1, true, 0);
5838 return success
? VK_SUCCESS
: VK_NOT_READY
;
5842 if (!device
->ws
->fence_wait(device
->ws
, fence
->fence
, false, 0))
5843 return VK_NOT_READY
;
5845 if (fence
->fence_wsi
) {
5846 VkResult result
= fence
->fence_wsi
->wait(fence
->fence_wsi
, 0);
5848 if (result
!= VK_SUCCESS
) {
5849 if (result
== VK_TIMEOUT
)
5850 return VK_NOT_READY
;
5858 // Queue semaphore functions
5861 radv_create_timeline(struct radv_timeline
*timeline
, uint64_t value
)
5863 timeline
->highest_signaled
= value
;
5864 timeline
->highest_submitted
= value
;
5865 list_inithead(&timeline
->points
);
5866 list_inithead(&timeline
->free_points
);
5867 list_inithead(&timeline
->waiters
);
5868 pthread_mutex_init(&timeline
->mutex
, NULL
);
5872 radv_destroy_timeline(struct radv_device
*device
,
5873 struct radv_timeline
*timeline
)
5875 list_for_each_entry_safe(struct radv_timeline_point
, point
,
5876 &timeline
->free_points
, list
) {
5877 list_del(&point
->list
);
5878 device
->ws
->destroy_syncobj(device
->ws
, point
->syncobj
);
5881 list_for_each_entry_safe(struct radv_timeline_point
, point
,
5882 &timeline
->points
, list
) {
5883 list_del(&point
->list
);
5884 device
->ws
->destroy_syncobj(device
->ws
, point
->syncobj
);
5887 pthread_mutex_destroy(&timeline
->mutex
);
5891 radv_timeline_gc_locked(struct radv_device
*device
,
5892 struct radv_timeline
*timeline
)
5894 list_for_each_entry_safe(struct radv_timeline_point
, point
,
5895 &timeline
->points
, list
) {
5896 if (point
->wait_count
|| point
->value
> timeline
->highest_submitted
)
5899 if (device
->ws
->wait_syncobj(device
->ws
, &point
->syncobj
, 1, true, 0)) {
5900 timeline
->highest_signaled
= point
->value
;
5901 list_del(&point
->list
);
5902 list_add(&point
->list
, &timeline
->free_points
);
5907 static struct radv_timeline_point
*
5908 radv_timeline_find_point_at_least_locked(struct radv_device
*device
,
5909 struct radv_timeline
*timeline
,
5912 radv_timeline_gc_locked(device
, timeline
);
5914 if (p
<= timeline
->highest_signaled
)
5917 list_for_each_entry(struct radv_timeline_point
, point
,
5918 &timeline
->points
, list
) {
5919 if (point
->value
>= p
) {
5920 ++point
->wait_count
;
5927 static struct radv_timeline_point
*
5928 radv_timeline_add_point_locked(struct radv_device
*device
,
5929 struct radv_timeline
*timeline
,
5932 radv_timeline_gc_locked(device
, timeline
);
5934 struct radv_timeline_point
*ret
= NULL
;
5935 struct radv_timeline_point
*prev
= NULL
;
5937 if (p
<= timeline
->highest_signaled
)
5940 list_for_each_entry(struct radv_timeline_point
, point
,
5941 &timeline
->points
, list
) {
5942 if (point
->value
== p
) {
5946 if (point
->value
< p
)
5950 if (list_is_empty(&timeline
->free_points
)) {
5951 ret
= malloc(sizeof(struct radv_timeline_point
));
5952 device
->ws
->create_syncobj(device
->ws
, &ret
->syncobj
);
5954 ret
= list_first_entry(&timeline
->free_points
, struct radv_timeline_point
, list
);
5955 list_del(&ret
->list
);
5957 device
->ws
->reset_syncobj(device
->ws
, ret
->syncobj
);
5961 ret
->wait_count
= 1;
5964 list_add(&ret
->list
, &prev
->list
);
5966 list_addtail(&ret
->list
, &timeline
->points
);
5973 radv_timeline_wait_locked(struct radv_device
*device
,
5974 struct radv_timeline
*timeline
,
5976 uint64_t abs_timeout
)
5978 while(timeline
->highest_submitted
< value
) {
5979 struct timespec abstime
;
5980 timespec_from_nsec(&abstime
, abs_timeout
);
5982 pthread_cond_timedwait(&device
->timeline_cond
, &timeline
->mutex
, &abstime
);
5984 if (radv_get_current_time() >= abs_timeout
&& timeline
->highest_submitted
< value
)
5988 struct radv_timeline_point
*point
= radv_timeline_find_point_at_least_locked(device
, timeline
, value
);
5992 pthread_mutex_unlock(&timeline
->mutex
);
5994 bool success
= device
->ws
->wait_syncobj(device
->ws
, &point
->syncobj
, 1, true, abs_timeout
);
5996 pthread_mutex_lock(&timeline
->mutex
);
5997 point
->wait_count
--;
5998 return success
? VK_SUCCESS
: VK_TIMEOUT
;
6002 radv_timeline_trigger_waiters_locked(struct radv_timeline
*timeline
,
6003 struct list_head
*processing_list
)
6005 list_for_each_entry_safe(struct radv_timeline_waiter
, waiter
,
6006 &timeline
->waiters
, list
) {
6007 if (waiter
->value
> timeline
->highest_submitted
)
6010 if (p_atomic_dec_zero(&waiter
->submission
->submission_wait_count
)) {
6011 list_addtail(&waiter
->submission
->processing_list
, processing_list
);
6013 list_del(&waiter
->list
);
6018 void radv_destroy_semaphore_part(struct radv_device
*device
,
6019 struct radv_semaphore_part
*part
)
6021 switch(part
->kind
) {
6022 case RADV_SEMAPHORE_NONE
:
6024 case RADV_SEMAPHORE_WINSYS
:
6025 device
->ws
->destroy_sem(part
->ws_sem
);
6027 case RADV_SEMAPHORE_TIMELINE
:
6028 radv_destroy_timeline(device
, &part
->timeline
);
6030 case RADV_SEMAPHORE_SYNCOBJ
:
6031 device
->ws
->destroy_syncobj(device
->ws
, part
->syncobj
);
6034 part
->kind
= RADV_SEMAPHORE_NONE
;
6037 static VkSemaphoreTypeKHR
6038 radv_get_semaphore_type(const void *pNext
, uint64_t *initial_value
)
6040 const VkSemaphoreTypeCreateInfo
*type_info
=
6041 vk_find_struct_const(pNext
, SEMAPHORE_TYPE_CREATE_INFO
);
6044 return VK_SEMAPHORE_TYPE_BINARY
;
6047 *initial_value
= type_info
->initialValue
;
6048 return type_info
->semaphoreType
;
6051 VkResult
radv_CreateSemaphore(
6053 const VkSemaphoreCreateInfo
* pCreateInfo
,
6054 const VkAllocationCallbacks
* pAllocator
,
6055 VkSemaphore
* pSemaphore
)
6057 RADV_FROM_HANDLE(radv_device
, device
, _device
);
6058 const VkExportSemaphoreCreateInfo
*export
=
6059 vk_find_struct_const(pCreateInfo
->pNext
, EXPORT_SEMAPHORE_CREATE_INFO
);
6060 VkExternalSemaphoreHandleTypeFlags handleTypes
=
6061 export
? export
->handleTypes
: 0;
6062 uint64_t initial_value
= 0;
6063 VkSemaphoreTypeKHR type
= radv_get_semaphore_type(pCreateInfo
->pNext
, &initial_value
);
6065 struct radv_semaphore
*sem
= vk_alloc2(&device
->alloc
, pAllocator
,
6067 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
6069 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
6071 sem
->temporary
.kind
= RADV_SEMAPHORE_NONE
;
6072 sem
->permanent
.kind
= RADV_SEMAPHORE_NONE
;
6074 if (type
== VK_SEMAPHORE_TYPE_TIMELINE
) {
6075 radv_create_timeline(&sem
->permanent
.timeline
, initial_value
);
6076 sem
->permanent
.kind
= RADV_SEMAPHORE_TIMELINE
;
6077 } else if (device
->always_use_syncobj
|| handleTypes
) {
6078 assert (device
->physical_device
->rad_info
.has_syncobj
);
6079 int ret
= device
->ws
->create_syncobj(device
->ws
, &sem
->permanent
.syncobj
);
6081 vk_free2(&device
->alloc
, pAllocator
, sem
);
6082 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
6084 sem
->permanent
.kind
= RADV_SEMAPHORE_SYNCOBJ
;
6086 sem
->permanent
.ws_sem
= device
->ws
->create_sem(device
->ws
);
6087 if (!sem
->permanent
.ws_sem
) {
6088 vk_free2(&device
->alloc
, pAllocator
, sem
);
6089 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
6091 sem
->permanent
.kind
= RADV_SEMAPHORE_WINSYS
;
6094 *pSemaphore
= radv_semaphore_to_handle(sem
);
6098 void radv_DestroySemaphore(
6100 VkSemaphore _semaphore
,
6101 const VkAllocationCallbacks
* pAllocator
)
6103 RADV_FROM_HANDLE(radv_device
, device
, _device
);
6104 RADV_FROM_HANDLE(radv_semaphore
, sem
, _semaphore
);
6108 radv_destroy_semaphore_part(device
, &sem
->temporary
);
6109 radv_destroy_semaphore_part(device
, &sem
->permanent
);
6110 vk_free2(&device
->alloc
, pAllocator
, sem
);
6114 radv_GetSemaphoreCounterValue(VkDevice _device
,
6115 VkSemaphore _semaphore
,
6118 RADV_FROM_HANDLE(radv_device
, device
, _device
);
6119 RADV_FROM_HANDLE(radv_semaphore
, semaphore
, _semaphore
);
6121 struct radv_semaphore_part
*part
=
6122 semaphore
->temporary
.kind
!= RADV_SEMAPHORE_NONE
? &semaphore
->temporary
: &semaphore
->permanent
;
6124 switch (part
->kind
) {
6125 case RADV_SEMAPHORE_TIMELINE
: {
6126 pthread_mutex_lock(&part
->timeline
.mutex
);
6127 radv_timeline_gc_locked(device
, &part
->timeline
);
6128 *pValue
= part
->timeline
.highest_signaled
;
6129 pthread_mutex_unlock(&part
->timeline
.mutex
);
6132 case RADV_SEMAPHORE_NONE
:
6133 case RADV_SEMAPHORE_SYNCOBJ
:
6134 case RADV_SEMAPHORE_WINSYS
:
6135 unreachable("Invalid semaphore type");
6137 unreachable("Unhandled semaphore type");
6142 radv_wait_timelines(struct radv_device
*device
,
6143 const VkSemaphoreWaitInfo
* pWaitInfo
,
6144 uint64_t abs_timeout
)
6146 if ((pWaitInfo
->flags
& VK_SEMAPHORE_WAIT_ANY_BIT_KHR
) && pWaitInfo
->semaphoreCount
> 1) {
6148 for(uint32_t i
= 0; i
< pWaitInfo
->semaphoreCount
; ++i
) {
6149 RADV_FROM_HANDLE(radv_semaphore
, semaphore
, pWaitInfo
->pSemaphores
[i
]);
6150 pthread_mutex_lock(&semaphore
->permanent
.timeline
.mutex
);
6151 VkResult result
= radv_timeline_wait_locked(device
, &semaphore
->permanent
.timeline
, pWaitInfo
->pValues
[i
], 0);
6152 pthread_mutex_unlock(&semaphore
->permanent
.timeline
.mutex
);
6154 if (result
== VK_SUCCESS
)
6157 if (radv_get_current_time() > abs_timeout
)
6162 for(uint32_t i
= 0; i
< pWaitInfo
->semaphoreCount
; ++i
) {
6163 RADV_FROM_HANDLE(radv_semaphore
, semaphore
, pWaitInfo
->pSemaphores
[i
]);
6164 pthread_mutex_lock(&semaphore
->permanent
.timeline
.mutex
);
6165 VkResult result
= radv_timeline_wait_locked(device
, &semaphore
->permanent
.timeline
, pWaitInfo
->pValues
[i
], abs_timeout
);
6166 pthread_mutex_unlock(&semaphore
->permanent
.timeline
.mutex
);
6168 if (result
!= VK_SUCCESS
)
6174 radv_WaitSemaphores(VkDevice _device
,
6175 const VkSemaphoreWaitInfo
* pWaitInfo
,
6178 RADV_FROM_HANDLE(radv_device
, device
, _device
);
6179 uint64_t abs_timeout
= radv_get_absolute_timeout(timeout
);
6180 return radv_wait_timelines(device
, pWaitInfo
, abs_timeout
);
6184 radv_SignalSemaphore(VkDevice _device
,
6185 const VkSemaphoreSignalInfo
* pSignalInfo
)
6187 RADV_FROM_HANDLE(radv_device
, device
, _device
);
6188 RADV_FROM_HANDLE(radv_semaphore
, semaphore
, pSignalInfo
->semaphore
);
6190 struct radv_semaphore_part
*part
=
6191 semaphore
->temporary
.kind
!= RADV_SEMAPHORE_NONE
? &semaphore
->temporary
: &semaphore
->permanent
;
6193 switch(part
->kind
) {
6194 case RADV_SEMAPHORE_TIMELINE
: {
6195 pthread_mutex_lock(&part
->timeline
.mutex
);
6196 radv_timeline_gc_locked(device
, &part
->timeline
);
6197 part
->timeline
.highest_submitted
= MAX2(part
->timeline
.highest_submitted
, pSignalInfo
->value
);
6198 part
->timeline
.highest_signaled
= MAX2(part
->timeline
.highest_signaled
, pSignalInfo
->value
);
6200 struct list_head processing_list
;
6201 list_inithead(&processing_list
);
6202 radv_timeline_trigger_waiters_locked(&part
->timeline
, &processing_list
);
6203 pthread_mutex_unlock(&part
->timeline
.mutex
);
6205 return radv_process_submissions(&processing_list
);
6207 case RADV_SEMAPHORE_NONE
:
6208 case RADV_SEMAPHORE_SYNCOBJ
:
6209 case RADV_SEMAPHORE_WINSYS
:
6210 unreachable("Invalid semaphore type");
6217 VkResult
radv_CreateEvent(
6219 const VkEventCreateInfo
* pCreateInfo
,
6220 const VkAllocationCallbacks
* pAllocator
,
6223 RADV_FROM_HANDLE(radv_device
, device
, _device
);
6224 struct radv_event
*event
= vk_alloc2(&device
->alloc
, pAllocator
,
6226 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
6229 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
6231 event
->bo
= device
->ws
->buffer_create(device
->ws
, 8, 8,
6233 RADEON_FLAG_VA_UNCACHED
| RADEON_FLAG_CPU_ACCESS
| RADEON_FLAG_NO_INTERPROCESS_SHARING
,
6234 RADV_BO_PRIORITY_FENCE
);
6236 vk_free2(&device
->alloc
, pAllocator
, event
);
6237 return vk_error(device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
6240 event
->map
= (uint64_t*)device
->ws
->buffer_map(event
->bo
);
6242 *pEvent
= radv_event_to_handle(event
);
6247 void radv_DestroyEvent(
6250 const VkAllocationCallbacks
* pAllocator
)
6252 RADV_FROM_HANDLE(radv_device
, device
, _device
);
6253 RADV_FROM_HANDLE(radv_event
, event
, _event
);
6257 device
->ws
->buffer_destroy(event
->bo
);
6258 vk_free2(&device
->alloc
, pAllocator
, event
);
6261 VkResult
radv_GetEventStatus(
6265 RADV_FROM_HANDLE(radv_event
, event
, _event
);
6267 if (*event
->map
== 1)
6268 return VK_EVENT_SET
;
6269 return VK_EVENT_RESET
;
6272 VkResult
radv_SetEvent(
6276 RADV_FROM_HANDLE(radv_event
, event
, _event
);
6282 VkResult
radv_ResetEvent(
6286 RADV_FROM_HANDLE(radv_event
, event
, _event
);
6292 VkResult
radv_CreateBuffer(
6294 const VkBufferCreateInfo
* pCreateInfo
,
6295 const VkAllocationCallbacks
* pAllocator
,
6298 RADV_FROM_HANDLE(radv_device
, device
, _device
);
6299 struct radv_buffer
*buffer
;
6301 if (pCreateInfo
->size
> RADV_MAX_MEMORY_ALLOCATION_SIZE
)
6302 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
6304 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
6306 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
6307 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
6309 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
6311 buffer
->size
= pCreateInfo
->size
;
6312 buffer
->usage
= pCreateInfo
->usage
;
6315 buffer
->flags
= pCreateInfo
->flags
;
6317 buffer
->shareable
= vk_find_struct_const(pCreateInfo
->pNext
,
6318 EXTERNAL_MEMORY_BUFFER_CREATE_INFO
) != NULL
;
6320 if (pCreateInfo
->flags
& VK_BUFFER_CREATE_SPARSE_BINDING_BIT
) {
6321 buffer
->bo
= device
->ws
->buffer_create(device
->ws
,
6322 align64(buffer
->size
, 4096),
6323 4096, 0, RADEON_FLAG_VIRTUAL
,
6324 RADV_BO_PRIORITY_VIRTUAL
);
6326 vk_free2(&device
->alloc
, pAllocator
, buffer
);
6327 return vk_error(device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
6331 *pBuffer
= radv_buffer_to_handle(buffer
);
6336 void radv_DestroyBuffer(
6339 const VkAllocationCallbacks
* pAllocator
)
6341 RADV_FROM_HANDLE(radv_device
, device
, _device
);
6342 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
6347 if (buffer
->flags
& VK_BUFFER_CREATE_SPARSE_BINDING_BIT
)
6348 device
->ws
->buffer_destroy(buffer
->bo
);
6350 vk_free2(&device
->alloc
, pAllocator
, buffer
);
6353 VkDeviceAddress
radv_GetBufferDeviceAddress(
6355 const VkBufferDeviceAddressInfo
* pInfo
)
6357 RADV_FROM_HANDLE(radv_buffer
, buffer
, pInfo
->buffer
);
6358 return radv_buffer_get_va(buffer
->bo
) + buffer
->offset
;
6362 uint64_t radv_GetBufferOpaqueCaptureAddress(VkDevice device
,
6363 const VkBufferDeviceAddressInfo
* pInfo
)
6368 uint64_t radv_GetDeviceMemoryOpaqueCaptureAddress(VkDevice device
,
6369 const VkDeviceMemoryOpaqueCaptureAddressInfo
* pInfo
)
6374 static inline unsigned
6375 si_tile_mode_index(const struct radv_image_plane
*plane
, unsigned level
, bool stencil
)
6378 return plane
->surface
.u
.legacy
.stencil_tiling_index
[level
];
6380 return plane
->surface
.u
.legacy
.tiling_index
[level
];
6383 static uint32_t radv_surface_max_layer_count(struct radv_image_view
*iview
)
6385 return iview
->type
== VK_IMAGE_VIEW_TYPE_3D
? iview
->extent
.depth
: (iview
->base_layer
+ iview
->layer_count
);
6389 radv_init_dcc_control_reg(struct radv_device
*device
,
6390 struct radv_image_view
*iview
)
6392 unsigned max_uncompressed_block_size
= V_028C78_MAX_BLOCK_SIZE_256B
;
6393 unsigned min_compressed_block_size
= V_028C78_MIN_BLOCK_SIZE_32B
;
6394 unsigned max_compressed_block_size
;
6395 unsigned independent_128b_blocks
;
6396 unsigned independent_64b_blocks
;
6398 if (!radv_dcc_enabled(iview
->image
, iview
->base_mip
))
6401 if (!device
->physical_device
->rad_info
.has_dedicated_vram
) {
6402 /* amdvlk: [min-compressed-block-size] should be set to 32 for
6403 * dGPU and 64 for APU because all of our APUs to date use
6404 * DIMMs which have a request granularity size of 64B while all
6405 * other chips have a 32B request size.
6407 min_compressed_block_size
= V_028C78_MIN_BLOCK_SIZE_64B
;
6410 if (device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
6411 max_compressed_block_size
= V_028C78_MAX_BLOCK_SIZE_128B
;
6412 independent_64b_blocks
= 0;
6413 independent_128b_blocks
= 1;
6415 independent_128b_blocks
= 0;
6417 if (iview
->image
->info
.samples
> 1) {
6418 if (iview
->image
->planes
[0].surface
.bpe
== 1)
6419 max_uncompressed_block_size
= V_028C78_MAX_BLOCK_SIZE_64B
;
6420 else if (iview
->image
->planes
[0].surface
.bpe
== 2)
6421 max_uncompressed_block_size
= V_028C78_MAX_BLOCK_SIZE_128B
;
6424 if (iview
->image
->usage
& (VK_IMAGE_USAGE_SAMPLED_BIT
|
6425 VK_IMAGE_USAGE_TRANSFER_SRC_BIT
|
6426 VK_IMAGE_USAGE_INPUT_ATTACHMENT_BIT
)) {
6427 /* If this DCC image is potentially going to be used in texture
6428 * fetches, we need some special settings.
6430 independent_64b_blocks
= 1;
6431 max_compressed_block_size
= V_028C78_MAX_BLOCK_SIZE_64B
;
6433 /* MAX_UNCOMPRESSED_BLOCK_SIZE must be >=
6434 * MAX_COMPRESSED_BLOCK_SIZE. Set MAX_COMPRESSED_BLOCK_SIZE as
6435 * big as possible for better compression state.
6437 independent_64b_blocks
= 0;
6438 max_compressed_block_size
= max_uncompressed_block_size
;
6442 return S_028C78_MAX_UNCOMPRESSED_BLOCK_SIZE(max_uncompressed_block_size
) |
6443 S_028C78_MAX_COMPRESSED_BLOCK_SIZE(max_compressed_block_size
) |
6444 S_028C78_MIN_COMPRESSED_BLOCK_SIZE(min_compressed_block_size
) |
6445 S_028C78_INDEPENDENT_64B_BLOCKS(independent_64b_blocks
) |
6446 S_028C78_INDEPENDENT_128B_BLOCKS(independent_128b_blocks
);
6450 radv_initialise_color_surface(struct radv_device
*device
,
6451 struct radv_color_buffer_info
*cb
,
6452 struct radv_image_view
*iview
)
6454 const struct vk_format_description
*desc
;
6455 unsigned ntype
, format
, swap
, endian
;
6456 unsigned blend_clamp
= 0, blend_bypass
= 0;
6458 const struct radv_image_plane
*plane
= &iview
->image
->planes
[iview
->plane_id
];
6459 const struct radeon_surf
*surf
= &plane
->surface
;
6461 desc
= vk_format_description(iview
->vk_format
);
6463 memset(cb
, 0, sizeof(*cb
));
6465 /* Intensity is implemented as Red, so treat it that way. */
6466 cb
->cb_color_attrib
= S_028C74_FORCE_DST_ALPHA_1(desc
->swizzle
[3] == VK_SWIZZLE_1
);
6468 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
+ plane
->offset
;
6470 cb
->cb_color_base
= va
>> 8;
6472 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
6473 if (device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
6474 cb
->cb_color_attrib3
|= S_028EE0_COLOR_SW_MODE(surf
->u
.gfx9
.surf
.swizzle_mode
) |
6475 S_028EE0_FMASK_SW_MODE(surf
->u
.gfx9
.fmask
.swizzle_mode
) |
6476 S_028EE0_CMASK_PIPE_ALIGNED(1) |
6477 S_028EE0_DCC_PIPE_ALIGNED(surf
->u
.gfx9
.dcc
.pipe_aligned
);
6479 struct gfx9_surf_meta_flags meta
= {
6484 if (iview
->image
->dcc_offset
)
6485 meta
= surf
->u
.gfx9
.dcc
;
6487 cb
->cb_color_attrib
|= S_028C74_COLOR_SW_MODE(surf
->u
.gfx9
.surf
.swizzle_mode
) |
6488 S_028C74_FMASK_SW_MODE(surf
->u
.gfx9
.fmask
.swizzle_mode
) |
6489 S_028C74_RB_ALIGNED(meta
.rb_aligned
) |
6490 S_028C74_PIPE_ALIGNED(meta
.pipe_aligned
);
6491 cb
->cb_mrt_epitch
= S_0287A0_EPITCH(surf
->u
.gfx9
.surf
.epitch
);
6494 cb
->cb_color_base
+= surf
->u
.gfx9
.surf_offset
>> 8;
6495 cb
->cb_color_base
|= surf
->tile_swizzle
;
6497 const struct legacy_surf_level
*level_info
= &surf
->u
.legacy
.level
[iview
->base_mip
];
6498 unsigned pitch_tile_max
, slice_tile_max
, tile_mode_index
;
6500 cb
->cb_color_base
+= level_info
->offset
>> 8;
6501 if (level_info
->mode
== RADEON_SURF_MODE_2D
)
6502 cb
->cb_color_base
|= surf
->tile_swizzle
;
6504 pitch_tile_max
= level_info
->nblk_x
/ 8 - 1;
6505 slice_tile_max
= (level_info
->nblk_x
* level_info
->nblk_y
) / 64 - 1;
6506 tile_mode_index
= si_tile_mode_index(plane
, iview
->base_mip
, false);
6508 cb
->cb_color_pitch
= S_028C64_TILE_MAX(pitch_tile_max
);
6509 cb
->cb_color_slice
= S_028C68_TILE_MAX(slice_tile_max
);
6510 cb
->cb_color_cmask_slice
= surf
->u
.legacy
.cmask_slice_tile_max
;
6512 cb
->cb_color_attrib
|= S_028C74_TILE_MODE_INDEX(tile_mode_index
);
6514 if (radv_image_has_fmask(iview
->image
)) {
6515 if (device
->physical_device
->rad_info
.chip_class
>= GFX7
)
6516 cb
->cb_color_pitch
|= S_028C64_FMASK_TILE_MAX(surf
->u
.legacy
.fmask
.pitch_in_pixels
/ 8 - 1);
6517 cb
->cb_color_attrib
|= S_028C74_FMASK_TILE_MODE_INDEX(surf
->u
.legacy
.fmask
.tiling_index
);
6518 cb
->cb_color_fmask_slice
= S_028C88_TILE_MAX(surf
->u
.legacy
.fmask
.slice_tile_max
);
6520 /* This must be set for fast clear to work without FMASK. */
6521 if (device
->physical_device
->rad_info
.chip_class
>= GFX7
)
6522 cb
->cb_color_pitch
|= S_028C64_FMASK_TILE_MAX(pitch_tile_max
);
6523 cb
->cb_color_attrib
|= S_028C74_FMASK_TILE_MODE_INDEX(tile_mode_index
);
6524 cb
->cb_color_fmask_slice
= S_028C88_TILE_MAX(slice_tile_max
);
6528 /* CMASK variables */
6529 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
;
6530 va
+= iview
->image
->cmask_offset
;
6531 cb
->cb_color_cmask
= va
>> 8;
6533 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
;
6534 va
+= iview
->image
->dcc_offset
;
6536 if (radv_dcc_enabled(iview
->image
, iview
->base_mip
) &&
6537 device
->physical_device
->rad_info
.chip_class
<= GFX8
)
6538 va
+= plane
->surface
.u
.legacy
.level
[iview
->base_mip
].dcc_offset
;
6540 unsigned dcc_tile_swizzle
= surf
->tile_swizzle
;
6541 dcc_tile_swizzle
&= (surf
->dcc_alignment
- 1) >> 8;
6543 cb
->cb_dcc_base
= va
>> 8;
6544 cb
->cb_dcc_base
|= dcc_tile_swizzle
;
6546 /* GFX10 field has the same base shift as the GFX6 field. */
6547 uint32_t max_slice
= radv_surface_max_layer_count(iview
) - 1;
6548 cb
->cb_color_view
= S_028C6C_SLICE_START(iview
->base_layer
) |
6549 S_028C6C_SLICE_MAX_GFX10(max_slice
);
6551 if (iview
->image
->info
.samples
> 1) {
6552 unsigned log_samples
= util_logbase2(iview
->image
->info
.samples
);
6554 cb
->cb_color_attrib
|= S_028C74_NUM_SAMPLES(log_samples
) |
6555 S_028C74_NUM_FRAGMENTS(log_samples
);
6558 if (radv_image_has_fmask(iview
->image
)) {
6559 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
+ iview
->image
->fmask_offset
;
6560 cb
->cb_color_fmask
= va
>> 8;
6561 cb
->cb_color_fmask
|= surf
->fmask_tile_swizzle
;
6563 cb
->cb_color_fmask
= cb
->cb_color_base
;
6566 ntype
= radv_translate_color_numformat(iview
->vk_format
,
6568 vk_format_get_first_non_void_channel(iview
->vk_format
));
6569 format
= radv_translate_colorformat(iview
->vk_format
);
6570 if (format
== V_028C70_COLOR_INVALID
|| ntype
== ~0u)
6571 radv_finishme("Illegal color\n");
6572 swap
= radv_translate_colorswap(iview
->vk_format
, false);
6573 endian
= radv_colorformat_endian_swap(format
);
6575 /* blend clamp should be set for all NORM/SRGB types */
6576 if (ntype
== V_028C70_NUMBER_UNORM
||
6577 ntype
== V_028C70_NUMBER_SNORM
||
6578 ntype
== V_028C70_NUMBER_SRGB
)
6581 /* set blend bypass according to docs if SINT/UINT or
6582 8/24 COLOR variants */
6583 if (ntype
== V_028C70_NUMBER_UINT
|| ntype
== V_028C70_NUMBER_SINT
||
6584 format
== V_028C70_COLOR_8_24
|| format
== V_028C70_COLOR_24_8
||
6585 format
== V_028C70_COLOR_X24_8_32_FLOAT
) {
6590 if ((ntype
== V_028C70_NUMBER_UINT
|| ntype
== V_028C70_NUMBER_SINT
) &&
6591 (format
== V_028C70_COLOR_8
||
6592 format
== V_028C70_COLOR_8_8
||
6593 format
== V_028C70_COLOR_8_8_8_8
))
6594 ->color_is_int8
= true;
6596 cb
->cb_color_info
= S_028C70_FORMAT(format
) |
6597 S_028C70_COMP_SWAP(swap
) |
6598 S_028C70_BLEND_CLAMP(blend_clamp
) |
6599 S_028C70_BLEND_BYPASS(blend_bypass
) |
6600 S_028C70_SIMPLE_FLOAT(1) |
6601 S_028C70_ROUND_MODE(ntype
!= V_028C70_NUMBER_UNORM
&&
6602 ntype
!= V_028C70_NUMBER_SNORM
&&
6603 ntype
!= V_028C70_NUMBER_SRGB
&&
6604 format
!= V_028C70_COLOR_8_24
&&
6605 format
!= V_028C70_COLOR_24_8
) |
6606 S_028C70_NUMBER_TYPE(ntype
) |
6607 S_028C70_ENDIAN(endian
);
6608 if (radv_image_has_fmask(iview
->image
)) {
6609 cb
->cb_color_info
|= S_028C70_COMPRESSION(1);
6610 if (device
->physical_device
->rad_info
.chip_class
== GFX6
) {
6611 unsigned fmask_bankh
= util_logbase2(surf
->u
.legacy
.fmask
.bankh
);
6612 cb
->cb_color_attrib
|= S_028C74_FMASK_BANK_HEIGHT(fmask_bankh
);
6615 if (radv_image_is_tc_compat_cmask(iview
->image
)) {
6616 /* Allow the texture block to read FMASK directly
6617 * without decompressing it. This bit must be cleared
6618 * when performing FMASK_DECOMPRESS or DCC_COMPRESS,
6619 * otherwise the operation doesn't happen.
6621 cb
->cb_color_info
|= S_028C70_FMASK_COMPRESS_1FRAG_ONLY(1);
6623 /* Set CMASK into a tiling format that allows the
6624 * texture block to read it.
6626 cb
->cb_color_info
|= S_028C70_CMASK_ADDR_TYPE(2);
6630 if (radv_image_has_cmask(iview
->image
) &&
6631 !(device
->instance
->debug_flags
& RADV_DEBUG_NO_FAST_CLEARS
))
6632 cb
->cb_color_info
|= S_028C70_FAST_CLEAR(1);
6634 if (radv_dcc_enabled(iview
->image
, iview
->base_mip
))
6635 cb
->cb_color_info
|= S_028C70_DCC_ENABLE(1);
6637 cb
->cb_dcc_control
= radv_init_dcc_control_reg(device
, iview
);
6639 /* This must be set for fast clear to work without FMASK. */
6640 if (!radv_image_has_fmask(iview
->image
) &&
6641 device
->physical_device
->rad_info
.chip_class
== GFX6
) {
6642 unsigned bankh
= util_logbase2(surf
->u
.legacy
.bankh
);
6643 cb
->cb_color_attrib
|= S_028C74_FMASK_BANK_HEIGHT(bankh
);
6646 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
6647 const struct vk_format_description
*format_desc
= vk_format_description(iview
->image
->vk_format
);
6649 unsigned mip0_depth
= iview
->image
->type
== VK_IMAGE_TYPE_3D
?
6650 (iview
->extent
.depth
- 1) : (iview
->image
->info
.array_size
- 1);
6651 unsigned width
= iview
->extent
.width
/ (iview
->plane_id
? format_desc
->width_divisor
: 1);
6652 unsigned height
= iview
->extent
.height
/ (iview
->plane_id
? format_desc
->height_divisor
: 1);
6654 if (device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
6655 cb
->cb_color_view
|= S_028C6C_MIP_LEVEL_GFX10(iview
->base_mip
);
6657 cb
->cb_color_attrib3
|= S_028EE0_MIP0_DEPTH(mip0_depth
) |
6658 S_028EE0_RESOURCE_TYPE(surf
->u
.gfx9
.resource_type
) |
6659 S_028EE0_RESOURCE_LEVEL(1);
6661 cb
->cb_color_view
|= S_028C6C_MIP_LEVEL_GFX9(iview
->base_mip
);
6662 cb
->cb_color_attrib
|= S_028C74_MIP0_DEPTH(mip0_depth
) |
6663 S_028C74_RESOURCE_TYPE(surf
->u
.gfx9
.resource_type
);
6666 cb
->cb_color_attrib2
= S_028C68_MIP0_WIDTH(width
- 1) |
6667 S_028C68_MIP0_HEIGHT(height
- 1) |
6668 S_028C68_MAX_MIP(iview
->image
->info
.levels
- 1);
6673 radv_calc_decompress_on_z_planes(struct radv_device
*device
,
6674 struct radv_image_view
*iview
)
6676 unsigned max_zplanes
= 0;
6678 assert(radv_image_is_tc_compat_htile(iview
->image
));
6680 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
6681 /* Default value for 32-bit depth surfaces. */
6684 if (iview
->vk_format
== VK_FORMAT_D16_UNORM
&&
6685 iview
->image
->info
.samples
> 1)
6688 max_zplanes
= max_zplanes
+ 1;
6690 if (iview
->vk_format
== VK_FORMAT_D16_UNORM
) {
6691 /* Do not enable Z plane compression for 16-bit depth
6692 * surfaces because isn't supported on GFX8. Only
6693 * 32-bit depth surfaces are supported by the hardware.
6694 * This allows to maintain shader compatibility and to
6695 * reduce the number of depth decompressions.
6699 if (iview
->image
->info
.samples
<= 1)
6701 else if (iview
->image
->info
.samples
<= 4)
6712 radv_initialise_ds_surface(struct radv_device
*device
,
6713 struct radv_ds_buffer_info
*ds
,
6714 struct radv_image_view
*iview
)
6716 unsigned level
= iview
->base_mip
;
6717 unsigned format
, stencil_format
;
6718 uint64_t va
, s_offs
, z_offs
;
6719 bool stencil_only
= false;
6720 const struct radv_image_plane
*plane
= &iview
->image
->planes
[0];
6721 const struct radeon_surf
*surf
= &plane
->surface
;
6723 assert(vk_format_get_plane_count(iview
->image
->vk_format
) == 1);
6725 memset(ds
, 0, sizeof(*ds
));
6726 switch (iview
->image
->vk_format
) {
6727 case VK_FORMAT_D24_UNORM_S8_UINT
:
6728 case VK_FORMAT_X8_D24_UNORM_PACK32
:
6729 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-24);
6730 ds
->offset_scale
= 2.0f
;
6732 case VK_FORMAT_D16_UNORM
:
6733 case VK_FORMAT_D16_UNORM_S8_UINT
:
6734 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-16);
6735 ds
->offset_scale
= 4.0f
;
6737 case VK_FORMAT_D32_SFLOAT
:
6738 case VK_FORMAT_D32_SFLOAT_S8_UINT
:
6739 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-23) |
6740 S_028B78_POLY_OFFSET_DB_IS_FLOAT_FMT(1);
6741 ds
->offset_scale
= 1.0f
;
6743 case VK_FORMAT_S8_UINT
:
6744 stencil_only
= true;
6750 format
= radv_translate_dbformat(iview
->image
->vk_format
);
6751 stencil_format
= surf
->has_stencil
?
6752 V_028044_STENCIL_8
: V_028044_STENCIL_INVALID
;
6754 uint32_t max_slice
= radv_surface_max_layer_count(iview
) - 1;
6755 ds
->db_depth_view
= S_028008_SLICE_START(iview
->base_layer
) |
6756 S_028008_SLICE_MAX(max_slice
);
6757 if (device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
6758 ds
->db_depth_view
|= S_028008_SLICE_START_HI(iview
->base_layer
>> 11) |
6759 S_028008_SLICE_MAX_HI(max_slice
>> 11);
6762 ds
->db_htile_data_base
= 0;
6763 ds
->db_htile_surface
= 0;
6765 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
;
6766 s_offs
= z_offs
= va
;
6768 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
6769 assert(surf
->u
.gfx9
.surf_offset
== 0);
6770 s_offs
+= surf
->u
.gfx9
.stencil_offset
;
6772 ds
->db_z_info
= S_028038_FORMAT(format
) |
6773 S_028038_NUM_SAMPLES(util_logbase2(iview
->image
->info
.samples
)) |
6774 S_028038_SW_MODE(surf
->u
.gfx9
.surf
.swizzle_mode
) |
6775 S_028038_MAXMIP(iview
->image
->info
.levels
- 1) |
6776 S_028038_ZRANGE_PRECISION(1);
6777 ds
->db_stencil_info
= S_02803C_FORMAT(stencil_format
) |
6778 S_02803C_SW_MODE(surf
->u
.gfx9
.stencil
.swizzle_mode
);
6780 if (device
->physical_device
->rad_info
.chip_class
== GFX9
) {
6781 ds
->db_z_info2
= S_028068_EPITCH(surf
->u
.gfx9
.surf
.epitch
);
6782 ds
->db_stencil_info2
= S_02806C_EPITCH(surf
->u
.gfx9
.stencil
.epitch
);
6785 ds
->db_depth_view
|= S_028008_MIPID(level
);
6786 ds
->db_depth_size
= S_02801C_X_MAX(iview
->image
->info
.width
- 1) |
6787 S_02801C_Y_MAX(iview
->image
->info
.height
- 1);
6789 if (radv_htile_enabled(iview
->image
, level
)) {
6790 ds
->db_z_info
|= S_028038_TILE_SURFACE_ENABLE(1);
6792 if (radv_image_is_tc_compat_htile(iview
->image
)) {
6793 unsigned max_zplanes
=
6794 radv_calc_decompress_on_z_planes(device
, iview
);
6796 ds
->db_z_info
|= S_028038_DECOMPRESS_ON_N_ZPLANES(max_zplanes
);
6798 if (device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
6799 ds
->db_z_info
|= S_028040_ITERATE_FLUSH(1);
6800 ds
->db_stencil_info
|= S_028044_ITERATE_FLUSH(1);
6802 ds
->db_z_info
|= S_028038_ITERATE_FLUSH(1);
6803 ds
->db_stencil_info
|= S_02803C_ITERATE_FLUSH(1);
6807 if (!surf
->has_stencil
)
6808 /* Use all of the htile_buffer for depth if there's no stencil. */
6809 ds
->db_stencil_info
|= S_02803C_TILE_STENCIL_DISABLE(1);
6810 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
+
6811 iview
->image
->htile_offset
;
6812 ds
->db_htile_data_base
= va
>> 8;
6813 ds
->db_htile_surface
= S_028ABC_FULL_CACHE(1) |
6814 S_028ABC_PIPE_ALIGNED(1);
6816 if (device
->physical_device
->rad_info
.chip_class
== GFX9
) {
6817 ds
->db_htile_surface
|= S_028ABC_RB_ALIGNED(1);
6821 const struct legacy_surf_level
*level_info
= &surf
->u
.legacy
.level
[level
];
6824 level_info
= &surf
->u
.legacy
.stencil_level
[level
];
6826 z_offs
+= surf
->u
.legacy
.level
[level
].offset
;
6827 s_offs
+= surf
->u
.legacy
.stencil_level
[level
].offset
;
6829 ds
->db_depth_info
= S_02803C_ADDR5_SWIZZLE_MASK(!radv_image_is_tc_compat_htile(iview
->image
));
6830 ds
->db_z_info
= S_028040_FORMAT(format
) | S_028040_ZRANGE_PRECISION(1);
6831 ds
->db_stencil_info
= S_028044_FORMAT(stencil_format
);
6833 if (iview
->image
->info
.samples
> 1)
6834 ds
->db_z_info
|= S_028040_NUM_SAMPLES(util_logbase2(iview
->image
->info
.samples
));
6836 if (device
->physical_device
->rad_info
.chip_class
>= GFX7
) {
6837 struct radeon_info
*info
= &device
->physical_device
->rad_info
;
6838 unsigned tiling_index
= surf
->u
.legacy
.tiling_index
[level
];
6839 unsigned stencil_index
= surf
->u
.legacy
.stencil_tiling_index
[level
];
6840 unsigned macro_index
= surf
->u
.legacy
.macro_tile_index
;
6841 unsigned tile_mode
= info
->si_tile_mode_array
[tiling_index
];
6842 unsigned stencil_tile_mode
= info
->si_tile_mode_array
[stencil_index
];
6843 unsigned macro_mode
= info
->cik_macrotile_mode_array
[macro_index
];
6846 tile_mode
= stencil_tile_mode
;
6848 ds
->db_depth_info
|=
6849 S_02803C_ARRAY_MODE(G_009910_ARRAY_MODE(tile_mode
)) |
6850 S_02803C_PIPE_CONFIG(G_009910_PIPE_CONFIG(tile_mode
)) |
6851 S_02803C_BANK_WIDTH(G_009990_BANK_WIDTH(macro_mode
)) |
6852 S_02803C_BANK_HEIGHT(G_009990_BANK_HEIGHT(macro_mode
)) |
6853 S_02803C_MACRO_TILE_ASPECT(G_009990_MACRO_TILE_ASPECT(macro_mode
)) |
6854 S_02803C_NUM_BANKS(G_009990_NUM_BANKS(macro_mode
));
6855 ds
->db_z_info
|= S_028040_TILE_SPLIT(G_009910_TILE_SPLIT(tile_mode
));
6856 ds
->db_stencil_info
|= S_028044_TILE_SPLIT(G_009910_TILE_SPLIT(stencil_tile_mode
));
6858 unsigned tile_mode_index
= si_tile_mode_index(&iview
->image
->planes
[0], level
, false);
6859 ds
->db_z_info
|= S_028040_TILE_MODE_INDEX(tile_mode_index
);
6860 tile_mode_index
= si_tile_mode_index(&iview
->image
->planes
[0], level
, true);
6861 ds
->db_stencil_info
|= S_028044_TILE_MODE_INDEX(tile_mode_index
);
6863 ds
->db_z_info
|= S_028040_TILE_MODE_INDEX(tile_mode_index
);
6866 ds
->db_depth_size
= S_028058_PITCH_TILE_MAX((level_info
->nblk_x
/ 8) - 1) |
6867 S_028058_HEIGHT_TILE_MAX((level_info
->nblk_y
/ 8) - 1);
6868 ds
->db_depth_slice
= S_02805C_SLICE_TILE_MAX((level_info
->nblk_x
* level_info
->nblk_y
) / 64 - 1);
6870 if (radv_htile_enabled(iview
->image
, level
)) {
6871 ds
->db_z_info
|= S_028040_TILE_SURFACE_ENABLE(1);
6873 if (!surf
->has_stencil
&&
6874 !radv_image_is_tc_compat_htile(iview
->image
))
6875 /* Use all of the htile_buffer for depth if there's no stencil. */
6876 ds
->db_stencil_info
|= S_028044_TILE_STENCIL_DISABLE(1);
6878 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
+
6879 iview
->image
->htile_offset
;
6880 ds
->db_htile_data_base
= va
>> 8;
6881 ds
->db_htile_surface
= S_028ABC_FULL_CACHE(1);
6883 if (radv_image_is_tc_compat_htile(iview
->image
)) {
6884 unsigned max_zplanes
=
6885 radv_calc_decompress_on_z_planes(device
, iview
);
6887 ds
->db_htile_surface
|= S_028ABC_TC_COMPATIBLE(1);
6888 ds
->db_z_info
|= S_028040_DECOMPRESS_ON_N_ZPLANES(max_zplanes
);
6893 ds
->db_z_read_base
= ds
->db_z_write_base
= z_offs
>> 8;
6894 ds
->db_stencil_read_base
= ds
->db_stencil_write_base
= s_offs
>> 8;
6897 VkResult
radv_CreateFramebuffer(
6899 const VkFramebufferCreateInfo
* pCreateInfo
,
6900 const VkAllocationCallbacks
* pAllocator
,
6901 VkFramebuffer
* pFramebuffer
)
6903 RADV_FROM_HANDLE(radv_device
, device
, _device
);
6904 struct radv_framebuffer
*framebuffer
;
6905 const VkFramebufferAttachmentsCreateInfo
*imageless_create_info
=
6906 vk_find_struct_const(pCreateInfo
->pNext
,
6907 FRAMEBUFFER_ATTACHMENTS_CREATE_INFO
);
6909 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
6911 size_t size
= sizeof(*framebuffer
);
6912 if (!imageless_create_info
)
6913 size
+= sizeof(struct radv_image_view
*) * pCreateInfo
->attachmentCount
;
6914 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
6915 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
6916 if (framebuffer
== NULL
)
6917 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
6919 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
6920 framebuffer
->width
= pCreateInfo
->width
;
6921 framebuffer
->height
= pCreateInfo
->height
;
6922 framebuffer
->layers
= pCreateInfo
->layers
;
6923 if (imageless_create_info
) {
6924 for (unsigned i
= 0; i
< imageless_create_info
->attachmentImageInfoCount
; ++i
) {
6925 const VkFramebufferAttachmentImageInfo
*attachment
=
6926 imageless_create_info
->pAttachmentImageInfos
+ i
;
6927 framebuffer
->width
= MIN2(framebuffer
->width
, attachment
->width
);
6928 framebuffer
->height
= MIN2(framebuffer
->height
, attachment
->height
);
6929 framebuffer
->layers
= MIN2(framebuffer
->layers
, attachment
->layerCount
);
6932 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
6933 VkImageView _iview
= pCreateInfo
->pAttachments
[i
];
6934 struct radv_image_view
*iview
= radv_image_view_from_handle(_iview
);
6935 framebuffer
->attachments
[i
] = iview
;
6936 framebuffer
->width
= MIN2(framebuffer
->width
, iview
->extent
.width
);
6937 framebuffer
->height
= MIN2(framebuffer
->height
, iview
->extent
.height
);
6938 framebuffer
->layers
= MIN2(framebuffer
->layers
, radv_surface_max_layer_count(iview
));
6942 *pFramebuffer
= radv_framebuffer_to_handle(framebuffer
);
6946 void radv_DestroyFramebuffer(
6949 const VkAllocationCallbacks
* pAllocator
)
6951 RADV_FROM_HANDLE(radv_device
, device
, _device
);
6952 RADV_FROM_HANDLE(radv_framebuffer
, fb
, _fb
);
6956 vk_free2(&device
->alloc
, pAllocator
, fb
);
6959 static unsigned radv_tex_wrap(VkSamplerAddressMode address_mode
)
6961 switch (address_mode
) {
6962 case VK_SAMPLER_ADDRESS_MODE_REPEAT
:
6963 return V_008F30_SQ_TEX_WRAP
;
6964 case VK_SAMPLER_ADDRESS_MODE_MIRRORED_REPEAT
:
6965 return V_008F30_SQ_TEX_MIRROR
;
6966 case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE
:
6967 return V_008F30_SQ_TEX_CLAMP_LAST_TEXEL
;
6968 case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_BORDER
:
6969 return V_008F30_SQ_TEX_CLAMP_BORDER
;
6970 case VK_SAMPLER_ADDRESS_MODE_MIRROR_CLAMP_TO_EDGE
:
6971 return V_008F30_SQ_TEX_MIRROR_ONCE_LAST_TEXEL
;
6973 unreachable("illegal tex wrap mode");
6979 radv_tex_compare(VkCompareOp op
)
6982 case VK_COMPARE_OP_NEVER
:
6983 return V_008F30_SQ_TEX_DEPTH_COMPARE_NEVER
;
6984 case VK_COMPARE_OP_LESS
:
6985 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESS
;
6986 case VK_COMPARE_OP_EQUAL
:
6987 return V_008F30_SQ_TEX_DEPTH_COMPARE_EQUAL
;
6988 case VK_COMPARE_OP_LESS_OR_EQUAL
:
6989 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESSEQUAL
;
6990 case VK_COMPARE_OP_GREATER
:
6991 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATER
;
6992 case VK_COMPARE_OP_NOT_EQUAL
:
6993 return V_008F30_SQ_TEX_DEPTH_COMPARE_NOTEQUAL
;
6994 case VK_COMPARE_OP_GREATER_OR_EQUAL
:
6995 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATEREQUAL
;
6996 case VK_COMPARE_OP_ALWAYS
:
6997 return V_008F30_SQ_TEX_DEPTH_COMPARE_ALWAYS
;
6999 unreachable("illegal compare mode");
7005 radv_tex_filter(VkFilter filter
, unsigned max_ansio
)
7008 case VK_FILTER_NEAREST
:
7009 return (max_ansio
> 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_POINT
:
7010 V_008F38_SQ_TEX_XY_FILTER_POINT
);
7011 case VK_FILTER_LINEAR
:
7012 return (max_ansio
> 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_BILINEAR
:
7013 V_008F38_SQ_TEX_XY_FILTER_BILINEAR
);
7014 case VK_FILTER_CUBIC_IMG
:
7016 fprintf(stderr
, "illegal texture filter");
7022 radv_tex_mipfilter(VkSamplerMipmapMode mode
)
7025 case VK_SAMPLER_MIPMAP_MODE_NEAREST
:
7026 return V_008F38_SQ_TEX_Z_FILTER_POINT
;
7027 case VK_SAMPLER_MIPMAP_MODE_LINEAR
:
7028 return V_008F38_SQ_TEX_Z_FILTER_LINEAR
;
7030 return V_008F38_SQ_TEX_Z_FILTER_NONE
;
7035 radv_tex_bordercolor(VkBorderColor bcolor
)
7038 case VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
:
7039 case VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
:
7040 return V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK
;
7041 case VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
:
7042 case VK_BORDER_COLOR_INT_OPAQUE_BLACK
:
7043 return V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_BLACK
;
7044 case VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
:
7045 case VK_BORDER_COLOR_INT_OPAQUE_WHITE
:
7046 return V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_WHITE
;
7054 radv_tex_aniso_filter(unsigned filter
)
7068 radv_tex_filter_mode(VkSamplerReductionMode mode
)
7071 case VK_SAMPLER_REDUCTION_MODE_WEIGHTED_AVERAGE_EXT
:
7072 return V_008F30_SQ_IMG_FILTER_MODE_BLEND
;
7073 case VK_SAMPLER_REDUCTION_MODE_MIN_EXT
:
7074 return V_008F30_SQ_IMG_FILTER_MODE_MIN
;
7075 case VK_SAMPLER_REDUCTION_MODE_MAX_EXT
:
7076 return V_008F30_SQ_IMG_FILTER_MODE_MAX
;
7084 radv_get_max_anisotropy(struct radv_device
*device
,
7085 const VkSamplerCreateInfo
*pCreateInfo
)
7087 if (device
->force_aniso
>= 0)
7088 return device
->force_aniso
;
7090 if (pCreateInfo
->anisotropyEnable
&&
7091 pCreateInfo
->maxAnisotropy
> 1.0f
)
7092 return (uint32_t)pCreateInfo
->maxAnisotropy
;
7097 static inline int S_FIXED(float value
, unsigned frac_bits
)
7099 return value
* (1 << frac_bits
);
7103 radv_init_sampler(struct radv_device
*device
,
7104 struct radv_sampler
*sampler
,
7105 const VkSamplerCreateInfo
*pCreateInfo
)
7107 uint32_t max_aniso
= radv_get_max_anisotropy(device
, pCreateInfo
);
7108 uint32_t max_aniso_ratio
= radv_tex_aniso_filter(max_aniso
);
7109 bool compat_mode
= device
->physical_device
->rad_info
.chip_class
== GFX8
||
7110 device
->physical_device
->rad_info
.chip_class
== GFX9
;
7111 unsigned filter_mode
= V_008F30_SQ_IMG_FILTER_MODE_BLEND
;
7112 unsigned depth_compare_func
= V_008F30_SQ_TEX_DEPTH_COMPARE_NEVER
;
7113 bool trunc_coord
= pCreateInfo
->minFilter
== VK_FILTER_NEAREST
&& pCreateInfo
->magFilter
== VK_FILTER_NEAREST
;
7115 const struct VkSamplerReductionModeCreateInfo
*sampler_reduction
=
7116 vk_find_struct_const(pCreateInfo
->pNext
,
7117 SAMPLER_REDUCTION_MODE_CREATE_INFO
);
7118 if (sampler_reduction
)
7119 filter_mode
= radv_tex_filter_mode(sampler_reduction
->reductionMode
);
7121 if (pCreateInfo
->compareEnable
)
7122 depth_compare_func
= radv_tex_compare(pCreateInfo
->compareOp
);
7124 sampler
->state
[0] = (S_008F30_CLAMP_X(radv_tex_wrap(pCreateInfo
->addressModeU
)) |
7125 S_008F30_CLAMP_Y(radv_tex_wrap(pCreateInfo
->addressModeV
)) |
7126 S_008F30_CLAMP_Z(radv_tex_wrap(pCreateInfo
->addressModeW
)) |
7127 S_008F30_MAX_ANISO_RATIO(max_aniso_ratio
) |
7128 S_008F30_DEPTH_COMPARE_FUNC(depth_compare_func
) |
7129 S_008F30_FORCE_UNNORMALIZED(pCreateInfo
->unnormalizedCoordinates
? 1 : 0) |
7130 S_008F30_ANISO_THRESHOLD(max_aniso_ratio
>> 1) |
7131 S_008F30_ANISO_BIAS(max_aniso_ratio
) |
7132 S_008F30_DISABLE_CUBE_WRAP(0) |
7133 S_008F30_COMPAT_MODE(compat_mode
) |
7134 S_008F30_FILTER_MODE(filter_mode
) |
7135 S_008F30_TRUNC_COORD(trunc_coord
));
7136 sampler
->state
[1] = (S_008F34_MIN_LOD(S_FIXED(CLAMP(pCreateInfo
->minLod
, 0, 15), 8)) |
7137 S_008F34_MAX_LOD(S_FIXED(CLAMP(pCreateInfo
->maxLod
, 0, 15), 8)) |
7138 S_008F34_PERF_MIP(max_aniso_ratio
? max_aniso_ratio
+ 6 : 0));
7139 sampler
->state
[2] = (S_008F38_LOD_BIAS(S_FIXED(CLAMP(pCreateInfo
->mipLodBias
, -16, 16), 8)) |
7140 S_008F38_XY_MAG_FILTER(radv_tex_filter(pCreateInfo
->magFilter
, max_aniso
)) |
7141 S_008F38_XY_MIN_FILTER(radv_tex_filter(pCreateInfo
->minFilter
, max_aniso
)) |
7142 S_008F38_MIP_FILTER(radv_tex_mipfilter(pCreateInfo
->mipmapMode
)) |
7143 S_008F38_MIP_POINT_PRECLAMP(0));
7144 sampler
->state
[3] = (S_008F3C_BORDER_COLOR_PTR(0) |
7145 S_008F3C_BORDER_COLOR_TYPE(radv_tex_bordercolor(pCreateInfo
->borderColor
)));
7147 if (device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
7148 sampler
->state
[2] |= S_008F38_ANISO_OVERRIDE_GFX10(1);
7150 sampler
->state
[2] |=
7151 S_008F38_DISABLE_LSB_CEIL(device
->physical_device
->rad_info
.chip_class
<= GFX8
) |
7152 S_008F38_FILTER_PREC_FIX(1) |
7153 S_008F38_ANISO_OVERRIDE_GFX6(device
->physical_device
->rad_info
.chip_class
>= GFX8
);
7157 VkResult
radv_CreateSampler(
7159 const VkSamplerCreateInfo
* pCreateInfo
,
7160 const VkAllocationCallbacks
* pAllocator
,
7161 VkSampler
* pSampler
)
7163 RADV_FROM_HANDLE(radv_device
, device
, _device
);
7164 struct radv_sampler
*sampler
;
7166 const struct VkSamplerYcbcrConversionInfo
*ycbcr_conversion
=
7167 vk_find_struct_const(pCreateInfo
->pNext
,
7168 SAMPLER_YCBCR_CONVERSION_INFO
);
7170 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO
);
7172 sampler
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*sampler
), 8,
7173 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
7175 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
7177 radv_init_sampler(device
, sampler
, pCreateInfo
);
7179 sampler
->ycbcr_sampler
= ycbcr_conversion
? radv_sampler_ycbcr_conversion_from_handle(ycbcr_conversion
->conversion
): NULL
;
7180 *pSampler
= radv_sampler_to_handle(sampler
);
7185 void radv_DestroySampler(
7188 const VkAllocationCallbacks
* pAllocator
)
7190 RADV_FROM_HANDLE(radv_device
, device
, _device
);
7191 RADV_FROM_HANDLE(radv_sampler
, sampler
, _sampler
);
7195 vk_free2(&device
->alloc
, pAllocator
, sampler
);
7198 /* vk_icd.h does not declare this function, so we declare it here to
7199 * suppress Wmissing-prototypes.
7201 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
7202 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t *pSupportedVersion
);
7204 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
7205 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t *pSupportedVersion
)
7207 /* For the full details on loader interface versioning, see
7208 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
7209 * What follows is a condensed summary, to help you navigate the large and
7210 * confusing official doc.
7212 * - Loader interface v0 is incompatible with later versions. We don't
7215 * - In loader interface v1:
7216 * - The first ICD entrypoint called by the loader is
7217 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
7219 * - The ICD must statically expose no other Vulkan symbol unless it is
7220 * linked with -Bsymbolic.
7221 * - Each dispatchable Vulkan handle created by the ICD must be
7222 * a pointer to a struct whose first member is VK_LOADER_DATA. The
7223 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
7224 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
7225 * vkDestroySurfaceKHR(). The ICD must be capable of working with
7226 * such loader-managed surfaces.
7228 * - Loader interface v2 differs from v1 in:
7229 * - The first ICD entrypoint called by the loader is
7230 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
7231 * statically expose this entrypoint.
7233 * - Loader interface v3 differs from v2 in:
7234 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
7235 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
7236 * because the loader no longer does so.
7238 *pSupportedVersion
= MIN2(*pSupportedVersion
, 4u);
7242 VkResult
radv_GetMemoryFdKHR(VkDevice _device
,
7243 const VkMemoryGetFdInfoKHR
*pGetFdInfo
,
7246 RADV_FROM_HANDLE(radv_device
, device
, _device
);
7247 RADV_FROM_HANDLE(radv_device_memory
, memory
, pGetFdInfo
->memory
);
7249 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
7251 /* At the moment, we support only the below handle types. */
7252 assert(pGetFdInfo
->handleType
==
7253 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
7254 pGetFdInfo
->handleType
==
7255 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
7257 bool ret
= radv_get_memory_fd(device
, memory
, pFD
);
7259 return vk_error(device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
7263 static uint32_t radv_compute_valid_memory_types_attempt(struct radv_physical_device
*dev
,
7264 enum radeon_bo_domain domains
,
7265 enum radeon_bo_flag flags
,
7266 enum radeon_bo_flag ignore_flags
)
7268 /* Don't count GTT/CPU as relevant:
7270 * - We're not fully consistent between the two.
7271 * - Sometimes VRAM gets VRAM|GTT.
7273 const enum radeon_bo_domain relevant_domains
= RADEON_DOMAIN_VRAM
|
7277 for (unsigned i
= 0; i
< dev
->memory_properties
.memoryTypeCount
; ++i
) {
7278 if ((domains
& relevant_domains
) != (dev
->memory_domains
[i
] & relevant_domains
))
7281 if ((flags
& ~ignore_flags
) != (dev
->memory_flags
[i
] & ~ignore_flags
))
7290 static uint32_t radv_compute_valid_memory_types(struct radv_physical_device
*dev
,
7291 enum radeon_bo_domain domains
,
7292 enum radeon_bo_flag flags
)
7294 enum radeon_bo_flag ignore_flags
= ~(RADEON_FLAG_NO_CPU_ACCESS
| RADEON_FLAG_GTT_WC
);
7295 uint32_t bits
= radv_compute_valid_memory_types_attempt(dev
, domains
, flags
, ignore_flags
);
7298 ignore_flags
|= RADEON_FLAG_NO_CPU_ACCESS
;
7299 bits
= radv_compute_valid_memory_types_attempt(dev
, domains
, flags
, ignore_flags
);
7304 VkResult
radv_GetMemoryFdPropertiesKHR(VkDevice _device
,
7305 VkExternalMemoryHandleTypeFlagBits handleType
,
7307 VkMemoryFdPropertiesKHR
*pMemoryFdProperties
)
7309 RADV_FROM_HANDLE(radv_device
, device
, _device
);
7311 switch (handleType
) {
7312 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
: {
7313 enum radeon_bo_domain domains
;
7314 enum radeon_bo_flag flags
;
7315 if (!device
->ws
->buffer_get_flags_from_fd(device
->ws
, fd
, &domains
, &flags
))
7316 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE
);
7318 pMemoryFdProperties
->memoryTypeBits
= radv_compute_valid_memory_types(device
->physical_device
, domains
, flags
);
7322 /* The valid usage section for this function says:
7324 * "handleType must not be one of the handle types defined as
7327 * So opaque handle types fall into the default "unsupported" case.
7329 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE
);
7333 static VkResult
radv_import_opaque_fd(struct radv_device
*device
,
7337 uint32_t syncobj_handle
= 0;
7338 int ret
= device
->ws
->import_syncobj(device
->ws
, fd
, &syncobj_handle
);
7340 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE
);
7343 device
->ws
->destroy_syncobj(device
->ws
, *syncobj
);
7345 *syncobj
= syncobj_handle
;
7351 static VkResult
radv_import_sync_fd(struct radv_device
*device
,
7355 /* If we create a syncobj we do it locally so that if we have an error, we don't
7356 * leave a syncobj in an undetermined state in the fence. */
7357 uint32_t syncobj_handle
= *syncobj
;
7358 if (!syncobj_handle
) {
7359 int ret
= device
->ws
->create_syncobj(device
->ws
, &syncobj_handle
);
7361 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE
);
7366 device
->ws
->signal_syncobj(device
->ws
, syncobj_handle
);
7368 int ret
= device
->ws
->import_syncobj_from_sync_file(device
->ws
, syncobj_handle
, fd
);
7370 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE
);
7373 *syncobj
= syncobj_handle
;
7380 VkResult
radv_ImportSemaphoreFdKHR(VkDevice _device
,
7381 const VkImportSemaphoreFdInfoKHR
*pImportSemaphoreFdInfo
)
7383 RADV_FROM_HANDLE(radv_device
, device
, _device
);
7384 RADV_FROM_HANDLE(radv_semaphore
, sem
, pImportSemaphoreFdInfo
->semaphore
);
7386 struct radv_semaphore_part
*dst
= NULL
;
7388 if (pImportSemaphoreFdInfo
->flags
& VK_SEMAPHORE_IMPORT_TEMPORARY_BIT
) {
7389 dst
= &sem
->temporary
;
7391 dst
= &sem
->permanent
;
7394 uint32_t syncobj
= dst
->kind
== RADV_SEMAPHORE_SYNCOBJ
? dst
->syncobj
: 0;
7396 switch(pImportSemaphoreFdInfo
->handleType
) {
7397 case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT
:
7398 result
= radv_import_opaque_fd(device
, pImportSemaphoreFdInfo
->fd
, &syncobj
);
7400 case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT
:
7401 result
= radv_import_sync_fd(device
, pImportSemaphoreFdInfo
->fd
, &syncobj
);
7404 unreachable("Unhandled semaphore handle type");
7407 if (result
== VK_SUCCESS
) {
7408 dst
->syncobj
= syncobj
;
7409 dst
->kind
= RADV_SEMAPHORE_SYNCOBJ
;
7415 VkResult
radv_GetSemaphoreFdKHR(VkDevice _device
,
7416 const VkSemaphoreGetFdInfoKHR
*pGetFdInfo
,
7419 RADV_FROM_HANDLE(radv_device
, device
, _device
);
7420 RADV_FROM_HANDLE(radv_semaphore
, sem
, pGetFdInfo
->semaphore
);
7422 uint32_t syncobj_handle
;
7424 if (sem
->temporary
.kind
!= RADV_SEMAPHORE_NONE
) {
7425 assert(sem
->temporary
.kind
== RADV_SEMAPHORE_SYNCOBJ
);
7426 syncobj_handle
= sem
->temporary
.syncobj
;
7428 assert(sem
->permanent
.kind
== RADV_SEMAPHORE_SYNCOBJ
);
7429 syncobj_handle
= sem
->permanent
.syncobj
;
7432 switch(pGetFdInfo
->handleType
) {
7433 case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT
:
7434 ret
= device
->ws
->export_syncobj(device
->ws
, syncobj_handle
, pFd
);
7436 case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT
:
7437 ret
= device
->ws
->export_syncobj_to_sync_file(device
->ws
, syncobj_handle
, pFd
);
7439 if (sem
->temporary
.kind
!= RADV_SEMAPHORE_NONE
) {
7440 radv_destroy_semaphore_part(device
, &sem
->temporary
);
7442 device
->ws
->reset_syncobj(device
->ws
, syncobj_handle
);
7447 unreachable("Unhandled semaphore handle type");
7451 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE
);
7455 void radv_GetPhysicalDeviceExternalSemaphoreProperties(
7456 VkPhysicalDevice physicalDevice
,
7457 const VkPhysicalDeviceExternalSemaphoreInfo
*pExternalSemaphoreInfo
,
7458 VkExternalSemaphoreProperties
*pExternalSemaphoreProperties
)
7460 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
7461 VkSemaphoreTypeKHR type
= radv_get_semaphore_type(pExternalSemaphoreInfo
->pNext
, NULL
);
7463 if (type
== VK_SEMAPHORE_TYPE_TIMELINE
) {
7464 pExternalSemaphoreProperties
->exportFromImportedHandleTypes
= 0;
7465 pExternalSemaphoreProperties
->compatibleHandleTypes
= 0;
7466 pExternalSemaphoreProperties
->externalSemaphoreFeatures
= 0;
7468 /* Require has_syncobj_wait_for_submit for the syncobj signal ioctl introduced at virtually the same time */
7469 } else if (pdevice
->rad_info
.has_syncobj_wait_for_submit
&&
7470 (pExternalSemaphoreInfo
->handleType
== VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT
||
7471 pExternalSemaphoreInfo
->handleType
== VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT
)) {
7472 pExternalSemaphoreProperties
->exportFromImportedHandleTypes
= VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT
| VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT
;
7473 pExternalSemaphoreProperties
->compatibleHandleTypes
= VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT
| VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT
;
7474 pExternalSemaphoreProperties
->externalSemaphoreFeatures
= VK_EXTERNAL_SEMAPHORE_FEATURE_EXPORTABLE_BIT
|
7475 VK_EXTERNAL_SEMAPHORE_FEATURE_IMPORTABLE_BIT
;
7476 } else if (pExternalSemaphoreInfo
->handleType
== VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT
) {
7477 pExternalSemaphoreProperties
->exportFromImportedHandleTypes
= VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT
;
7478 pExternalSemaphoreProperties
->compatibleHandleTypes
= VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT
;
7479 pExternalSemaphoreProperties
->externalSemaphoreFeatures
= VK_EXTERNAL_SEMAPHORE_FEATURE_EXPORTABLE_BIT
|
7480 VK_EXTERNAL_SEMAPHORE_FEATURE_IMPORTABLE_BIT
;
7482 pExternalSemaphoreProperties
->exportFromImportedHandleTypes
= 0;
7483 pExternalSemaphoreProperties
->compatibleHandleTypes
= 0;
7484 pExternalSemaphoreProperties
->externalSemaphoreFeatures
= 0;
7488 VkResult
radv_ImportFenceFdKHR(VkDevice _device
,
7489 const VkImportFenceFdInfoKHR
*pImportFenceFdInfo
)
7491 RADV_FROM_HANDLE(radv_device
, device
, _device
);
7492 RADV_FROM_HANDLE(radv_fence
, fence
, pImportFenceFdInfo
->fence
);
7493 uint32_t *syncobj_dst
= NULL
;
7496 if (pImportFenceFdInfo
->flags
& VK_FENCE_IMPORT_TEMPORARY_BIT
) {
7497 syncobj_dst
= &fence
->temp_syncobj
;
7499 syncobj_dst
= &fence
->syncobj
;
7502 switch(pImportFenceFdInfo
->handleType
) {
7503 case VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT
:
7504 return radv_import_opaque_fd(device
, pImportFenceFdInfo
->fd
, syncobj_dst
);
7505 case VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT
:
7506 return radv_import_sync_fd(device
, pImportFenceFdInfo
->fd
, syncobj_dst
);
7508 unreachable("Unhandled fence handle type");
7512 VkResult
radv_GetFenceFdKHR(VkDevice _device
,
7513 const VkFenceGetFdInfoKHR
*pGetFdInfo
,
7516 RADV_FROM_HANDLE(radv_device
, device
, _device
);
7517 RADV_FROM_HANDLE(radv_fence
, fence
, pGetFdInfo
->fence
);
7519 uint32_t syncobj_handle
;
7521 if (fence
->temp_syncobj
)
7522 syncobj_handle
= fence
->temp_syncobj
;
7524 syncobj_handle
= fence
->syncobj
;
7526 switch(pGetFdInfo
->handleType
) {
7527 case VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT
:
7528 ret
= device
->ws
->export_syncobj(device
->ws
, syncobj_handle
, pFd
);
7530 case VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT
:
7531 ret
= device
->ws
->export_syncobj_to_sync_file(device
->ws
, syncobj_handle
, pFd
);
7533 if (fence
->temp_syncobj
) {
7534 close (fence
->temp_syncobj
);
7535 fence
->temp_syncobj
= 0;
7537 device
->ws
->reset_syncobj(device
->ws
, syncobj_handle
);
7542 unreachable("Unhandled fence handle type");
7546 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE
);
7550 void radv_GetPhysicalDeviceExternalFenceProperties(
7551 VkPhysicalDevice physicalDevice
,
7552 const VkPhysicalDeviceExternalFenceInfo
*pExternalFenceInfo
,
7553 VkExternalFenceProperties
*pExternalFenceProperties
)
7555 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
7557 if (pdevice
->rad_info
.has_syncobj_wait_for_submit
&&
7558 (pExternalFenceInfo
->handleType
== VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT
||
7559 pExternalFenceInfo
->handleType
== VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT
)) {
7560 pExternalFenceProperties
->exportFromImportedHandleTypes
= VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT
| VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT
;
7561 pExternalFenceProperties
->compatibleHandleTypes
= VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT
| VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT
;
7562 pExternalFenceProperties
->externalFenceFeatures
= VK_EXTERNAL_FENCE_FEATURE_EXPORTABLE_BIT
|
7563 VK_EXTERNAL_SEMAPHORE_FEATURE_IMPORTABLE_BIT
;
7565 pExternalFenceProperties
->exportFromImportedHandleTypes
= 0;
7566 pExternalFenceProperties
->compatibleHandleTypes
= 0;
7567 pExternalFenceProperties
->externalFenceFeatures
= 0;
7572 radv_CreateDebugReportCallbackEXT(VkInstance _instance
,
7573 const VkDebugReportCallbackCreateInfoEXT
* pCreateInfo
,
7574 const VkAllocationCallbacks
* pAllocator
,
7575 VkDebugReportCallbackEXT
* pCallback
)
7577 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
7578 return vk_create_debug_report_callback(&instance
->debug_report_callbacks
,
7579 pCreateInfo
, pAllocator
, &instance
->alloc
,
7584 radv_DestroyDebugReportCallbackEXT(VkInstance _instance
,
7585 VkDebugReportCallbackEXT _callback
,
7586 const VkAllocationCallbacks
* pAllocator
)
7588 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
7589 vk_destroy_debug_report_callback(&instance
->debug_report_callbacks
,
7590 _callback
, pAllocator
, &instance
->alloc
);
7594 radv_DebugReportMessageEXT(VkInstance _instance
,
7595 VkDebugReportFlagsEXT flags
,
7596 VkDebugReportObjectTypeEXT objectType
,
7599 int32_t messageCode
,
7600 const char* pLayerPrefix
,
7601 const char* pMessage
)
7603 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
7604 vk_debug_report(&instance
->debug_report_callbacks
, flags
, objectType
,
7605 object
, location
, messageCode
, pLayerPrefix
, pMessage
);
7609 radv_GetDeviceGroupPeerMemoryFeatures(
7612 uint32_t localDeviceIndex
,
7613 uint32_t remoteDeviceIndex
,
7614 VkPeerMemoryFeatureFlags
* pPeerMemoryFeatures
)
7616 assert(localDeviceIndex
== remoteDeviceIndex
);
7618 *pPeerMemoryFeatures
= VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT
|
7619 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT
|
7620 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT
|
7621 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT
;
7624 static const VkTimeDomainEXT radv_time_domains
[] = {
7625 VK_TIME_DOMAIN_DEVICE_EXT
,
7626 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
,
7627 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
,
7630 VkResult
radv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
7631 VkPhysicalDevice physicalDevice
,
7632 uint32_t *pTimeDomainCount
,
7633 VkTimeDomainEXT
*pTimeDomains
)
7636 VK_OUTARRAY_MAKE(out
, pTimeDomains
, pTimeDomainCount
);
7638 for (d
= 0; d
< ARRAY_SIZE(radv_time_domains
); d
++) {
7639 vk_outarray_append(&out
, i
) {
7640 *i
= radv_time_domains
[d
];
7644 return vk_outarray_status(&out
);
7648 radv_clock_gettime(clockid_t clock_id
)
7650 struct timespec current
;
7653 ret
= clock_gettime(clock_id
, ¤t
);
7654 if (ret
< 0 && clock_id
== CLOCK_MONOTONIC_RAW
)
7655 ret
= clock_gettime(CLOCK_MONOTONIC
, ¤t
);
7659 return (uint64_t) current
.tv_sec
* 1000000000ULL + current
.tv_nsec
;
7662 VkResult
radv_GetCalibratedTimestampsEXT(
7664 uint32_t timestampCount
,
7665 const VkCalibratedTimestampInfoEXT
*pTimestampInfos
,
7666 uint64_t *pTimestamps
,
7667 uint64_t *pMaxDeviation
)
7669 RADV_FROM_HANDLE(radv_device
, device
, _device
);
7670 uint32_t clock_crystal_freq
= device
->physical_device
->rad_info
.clock_crystal_freq
;
7672 uint64_t begin
, end
;
7673 uint64_t max_clock_period
= 0;
7675 begin
= radv_clock_gettime(CLOCK_MONOTONIC_RAW
);
7677 for (d
= 0; d
< timestampCount
; d
++) {
7678 switch (pTimestampInfos
[d
].timeDomain
) {
7679 case VK_TIME_DOMAIN_DEVICE_EXT
:
7680 pTimestamps
[d
] = device
->ws
->query_value(device
->ws
,
7682 uint64_t device_period
= DIV_ROUND_UP(1000000, clock_crystal_freq
);
7683 max_clock_period
= MAX2(max_clock_period
, device_period
);
7685 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
:
7686 pTimestamps
[d
] = radv_clock_gettime(CLOCK_MONOTONIC
);
7687 max_clock_period
= MAX2(max_clock_period
, 1);
7690 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
:
7691 pTimestamps
[d
] = begin
;
7699 end
= radv_clock_gettime(CLOCK_MONOTONIC_RAW
);
7702 * The maximum deviation is the sum of the interval over which we
7703 * perform the sampling and the maximum period of any sampled
7704 * clock. That's because the maximum skew between any two sampled
7705 * clock edges is when the sampled clock with the largest period is
7706 * sampled at the end of that period but right at the beginning of the
7707 * sampling interval and some other clock is sampled right at the
7708 * begining of its sampling period and right at the end of the
7709 * sampling interval. Let's assume the GPU has the longest clock
7710 * period and that the application is sampling GPU and monotonic:
7713 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
7714 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
7718 * GPU -----_____-----_____-----_____-----_____
7721 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
7722 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
7724 * Interval <----------------->
7725 * Deviation <-------------------------->
7729 * m = read(monotonic) 2
7732 * We round the sample interval up by one tick to cover sampling error
7733 * in the interval clock
7736 uint64_t sample_interval
= end
- begin
+ 1;
7738 *pMaxDeviation
= sample_interval
+ max_clock_period
;
7743 void radv_GetPhysicalDeviceMultisamplePropertiesEXT(
7744 VkPhysicalDevice physicalDevice
,
7745 VkSampleCountFlagBits samples
,
7746 VkMultisamplePropertiesEXT
* pMultisampleProperties
)
7748 if (samples
& (VK_SAMPLE_COUNT_2_BIT
|
7749 VK_SAMPLE_COUNT_4_BIT
|
7750 VK_SAMPLE_COUNT_8_BIT
)) {
7751 pMultisampleProperties
->maxSampleLocationGridSize
= (VkExtent2D
){ 2, 2 };
7753 pMultisampleProperties
->maxSampleLocationGridSize
= (VkExtent2D
){ 0, 0 };