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>
44 #include <llvm/Config/llvm-config.h>
46 #include "radv_debug.h"
47 #include "radv_private.h"
48 #include "radv_shader.h"
50 #include "util/disk_cache.h"
54 #include <amdgpu_drm.h>
55 #include "winsys/amdgpu/radv_amdgpu_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_is_mem_type_vram(enum radv_mem_type type
)
135 return type
== RADV_MEM_TYPE_VRAM
||
136 type
== RADV_MEM_TYPE_VRAM_UNCACHED
;
140 radv_is_mem_type_vram_visible(enum radv_mem_type type
)
142 return type
== RADV_MEM_TYPE_VRAM_CPU_ACCESS
||
143 type
== RADV_MEM_TYPE_VRAM_CPU_ACCESS_UNCACHED
;
146 radv_is_mem_type_gtt_wc(enum radv_mem_type type
)
148 return type
== RADV_MEM_TYPE_GTT_WRITE_COMBINE
||
149 type
== RADV_MEM_TYPE_GTT_WRITE_COMBINE_VRAM_UNCACHED
;
153 radv_is_mem_type_gtt_cached(enum radv_mem_type type
)
155 return type
== RADV_MEM_TYPE_GTT_CACHED
||
156 type
== RADV_MEM_TYPE_GTT_CACHED_VRAM_UNCACHED
;
160 radv_is_mem_type_uncached(enum radv_mem_type type
)
162 return type
== RADV_MEM_TYPE_VRAM_UNCACHED
||
163 type
== RADV_MEM_TYPE_VRAM_CPU_ACCESS_UNCACHED
||
164 type
== RADV_MEM_TYPE_GTT_WRITE_COMBINE_VRAM_UNCACHED
||
165 type
== RADV_MEM_TYPE_GTT_CACHED_VRAM_UNCACHED
;
169 radv_physical_device_init_mem_types(struct radv_physical_device
*device
)
171 STATIC_ASSERT(RADV_MEM_HEAP_COUNT
<= VK_MAX_MEMORY_HEAPS
);
172 uint64_t visible_vram_size
= radv_get_visible_vram_size(device
);
173 uint64_t vram_size
= radv_get_vram_size(device
);
174 int vram_index
= -1, visible_vram_index
= -1, gart_index
= -1;
175 device
->memory_properties
.memoryHeapCount
= 0;
177 vram_index
= device
->memory_properties
.memoryHeapCount
++;
178 device
->memory_properties
.memoryHeaps
[vram_index
] = (VkMemoryHeap
) {
180 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
183 if (visible_vram_size
) {
184 visible_vram_index
= device
->memory_properties
.memoryHeapCount
++;
185 device
->memory_properties
.memoryHeaps
[visible_vram_index
] = (VkMemoryHeap
) {
186 .size
= visible_vram_size
,
187 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
190 if (device
->rad_info
.gart_size
> 0) {
191 gart_index
= device
->memory_properties
.memoryHeapCount
++;
192 device
->memory_properties
.memoryHeaps
[gart_index
] = (VkMemoryHeap
) {
193 .size
= device
->rad_info
.gart_size
,
194 .flags
= device
->rad_info
.has_dedicated_vram
? 0 : VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
198 STATIC_ASSERT(RADV_MEM_TYPE_COUNT
<= VK_MAX_MEMORY_TYPES
);
199 unsigned type_count
= 0;
200 if (vram_index
>= 0) {
201 device
->mem_type_indices
[type_count
] = RADV_MEM_TYPE_VRAM
;
202 device
->memory_properties
.memoryTypes
[type_count
++] = (VkMemoryType
) {
203 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
,
204 .heapIndex
= vram_index
,
207 if (gart_index
>= 0 && device
->rad_info
.has_dedicated_vram
) {
208 device
->mem_type_indices
[type_count
] = RADV_MEM_TYPE_GTT_WRITE_COMBINE
;
209 device
->memory_properties
.memoryTypes
[type_count
++] = (VkMemoryType
) {
210 .propertyFlags
= VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
211 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
212 .heapIndex
= gart_index
,
215 if (visible_vram_index
>= 0) {
216 device
->mem_type_indices
[type_count
] = RADV_MEM_TYPE_VRAM_CPU_ACCESS
;
217 device
->memory_properties
.memoryTypes
[type_count
++] = (VkMemoryType
) {
218 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
219 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
220 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
221 .heapIndex
= visible_vram_index
,
224 if (gart_index
>= 0 && !device
->rad_info
.has_dedicated_vram
) {
225 /* Put GTT after visible VRAM for GPUs without dedicated VRAM
226 * as they have identical property flags, and according to the
227 * spec, for types with identical flags, the one with greater
228 * performance must be given a lower index. */
229 device
->mem_type_indices
[type_count
] = RADV_MEM_TYPE_GTT_WRITE_COMBINE
;
230 device
->memory_properties
.memoryTypes
[type_count
++] = (VkMemoryType
) {
231 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
232 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
233 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
234 .heapIndex
= gart_index
,
237 if (gart_index
>= 0) {
238 device
->mem_type_indices
[type_count
] = RADV_MEM_TYPE_GTT_CACHED
;
239 device
->memory_properties
.memoryTypes
[type_count
++] = (VkMemoryType
) {
240 .propertyFlags
= VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
241 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
242 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
|
243 (device
->rad_info
.has_dedicated_vram
? 0 : VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
),
244 .heapIndex
= gart_index
,
247 device
->memory_properties
.memoryTypeCount
= type_count
;
249 if (device
->rad_info
.has_l2_uncached
) {
250 for (int i
= 0; i
< device
->memory_properties
.memoryTypeCount
; i
++) {
251 VkMemoryType mem_type
= device
->memory_properties
.memoryTypes
[i
];
253 if ((mem_type
.propertyFlags
& (VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
254 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
)) ||
255 mem_type
.propertyFlags
== VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
) {
256 enum radv_mem_type mem_type_id
;
258 switch (device
->mem_type_indices
[i
]) {
259 case RADV_MEM_TYPE_VRAM
:
260 mem_type_id
= RADV_MEM_TYPE_VRAM_UNCACHED
;
262 case RADV_MEM_TYPE_VRAM_CPU_ACCESS
:
263 mem_type_id
= RADV_MEM_TYPE_VRAM_CPU_ACCESS_UNCACHED
;
265 case RADV_MEM_TYPE_GTT_WRITE_COMBINE
:
266 mem_type_id
= RADV_MEM_TYPE_GTT_WRITE_COMBINE_VRAM_UNCACHED
;
268 case RADV_MEM_TYPE_GTT_CACHED
:
269 mem_type_id
= RADV_MEM_TYPE_GTT_CACHED_VRAM_UNCACHED
;
272 unreachable("invalid memory type");
275 VkMemoryPropertyFlags property_flags
= mem_type
.propertyFlags
|
276 VK_MEMORY_PROPERTY_DEVICE_COHERENT_BIT_AMD
|
277 VK_MEMORY_PROPERTY_DEVICE_UNCACHED_BIT_AMD
;
279 device
->mem_type_indices
[type_count
] = mem_type_id
;
280 device
->memory_properties
.memoryTypes
[type_count
++] = (VkMemoryType
) {
281 .propertyFlags
= property_flags
,
282 .heapIndex
= mem_type
.heapIndex
,
286 device
->memory_properties
.memoryTypeCount
= type_count
;
291 radv_handle_env_var_force_family(struct radv_physical_device
*device
)
293 const char *family
= getenv("RADV_FORCE_FAMILY");
299 for (i
= CHIP_TAHITI
; i
< CHIP_LAST
; i
++) {
300 if (!strcmp(family
, ac_get_llvm_processor_name(i
))) {
301 /* Override family and chip_class. */
302 device
->rad_info
.family
= i
;
304 if (i
>= CHIP_NAVI10
)
305 device
->rad_info
.chip_class
= GFX10
;
306 else if (i
>= CHIP_VEGA10
)
307 device
->rad_info
.chip_class
= GFX9
;
308 else if (i
>= CHIP_TONGA
)
309 device
->rad_info
.chip_class
= GFX8
;
310 else if (i
>= CHIP_BONAIRE
)
311 device
->rad_info
.chip_class
= GFX7
;
313 device
->rad_info
.chip_class
= GFX6
;
319 fprintf(stderr
, "radv: Unknown family: %s\n", family
);
324 radv_physical_device_init(struct radv_physical_device
*device
,
325 struct radv_instance
*instance
,
326 drmDevicePtr drm_device
)
328 const char *path
= drm_device
->nodes
[DRM_NODE_RENDER
];
330 drmVersionPtr version
;
334 fd
= open(path
, O_RDWR
| O_CLOEXEC
);
336 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
337 radv_logi("Could not open device '%s'", path
);
339 return vk_error(instance
, VK_ERROR_INCOMPATIBLE_DRIVER
);
342 version
= drmGetVersion(fd
);
346 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
347 radv_logi("Could not get the kernel driver version for device '%s'", path
);
349 return vk_errorf(instance
, VK_ERROR_INCOMPATIBLE_DRIVER
,
350 "failed to get version %s: %m", path
);
353 if (strcmp(version
->name
, "amdgpu")) {
354 drmFreeVersion(version
);
357 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
358 radv_logi("Device '%s' is not using the amdgpu kernel driver.", path
);
360 return VK_ERROR_INCOMPATIBLE_DRIVER
;
362 drmFreeVersion(version
);
364 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
365 radv_logi("Found compatible device '%s'.", path
);
367 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
368 device
->instance
= instance
;
370 device
->ws
= radv_amdgpu_winsys_create(fd
, instance
->debug_flags
,
371 instance
->perftest_flags
);
373 result
= vk_error(instance
, VK_ERROR_INCOMPATIBLE_DRIVER
);
377 if (instance
->enabled_extensions
.KHR_display
) {
378 master_fd
= open(drm_device
->nodes
[DRM_NODE_PRIMARY
], O_RDWR
| O_CLOEXEC
);
379 if (master_fd
>= 0) {
380 uint32_t accel_working
= 0;
381 struct drm_amdgpu_info request
= {
382 .return_pointer
= (uintptr_t)&accel_working
,
383 .return_size
= sizeof(accel_working
),
384 .query
= AMDGPU_INFO_ACCEL_WORKING
387 if (drmCommandWrite(master_fd
, DRM_AMDGPU_INFO
, &request
, sizeof (struct drm_amdgpu_info
)) < 0 || !accel_working
) {
394 device
->master_fd
= master_fd
;
395 device
->local_fd
= fd
;
396 device
->ws
->query_info(device
->ws
, &device
->rad_info
);
398 radv_handle_env_var_force_family(device
);
400 device
->use_aco
= instance
->perftest_flags
& RADV_PERFTEST_ACO
;
401 if (device
->rad_info
.chip_class
< GFX8
&& device
->use_aco
) {
402 fprintf(stderr
, "WARNING: disabling ACO on unsupported GPUs.\n");
403 device
->use_aco
= false;
406 snprintf(device
->name
, sizeof(device
->name
),
407 "AMD RADV%s %s (LLVM " MESA_LLVM_VERSION_STRING
")", device
->use_aco
? "/ACO" : "",
408 device
->rad_info
.name
);
410 if (radv_device_get_cache_uuid(device
->rad_info
.family
, device
->cache_uuid
)) {
411 device
->ws
->destroy(device
->ws
);
412 result
= vk_errorf(instance
, VK_ERROR_INITIALIZATION_FAILED
,
413 "cannot generate UUID");
417 /* These flags affect shader compilation. */
418 uint64_t shader_env_flags
=
419 (device
->instance
->perftest_flags
& RADV_PERFTEST_SISCHED
? 0x1 : 0) |
420 (device
->use_aco
? 0x2 : 0);
422 /* The gpu id is already embedded in the uuid so we just pass "radv"
423 * when creating the cache.
425 char buf
[VK_UUID_SIZE
* 2 + 1];
426 disk_cache_format_hex_id(buf
, device
->cache_uuid
, VK_UUID_SIZE
* 2);
427 device
->disk_cache
= disk_cache_create(device
->name
, buf
, shader_env_flags
);
429 if (device
->rad_info
.chip_class
< GFX8
||
430 device
->rad_info
.chip_class
> GFX9
)
431 fprintf(stderr
, "WARNING: radv is not a conformant vulkan implementation, testing use only.\n");
433 radv_get_driver_uuid(&device
->driver_uuid
);
434 radv_get_device_uuid(&device
->rad_info
, &device
->device_uuid
);
436 device
->out_of_order_rast_allowed
= device
->rad_info
.has_out_of_order_rast
&&
437 !(device
->instance
->debug_flags
& RADV_DEBUG_NO_OUT_OF_ORDER
);
439 device
->dcc_msaa_allowed
=
440 (device
->instance
->perftest_flags
& RADV_PERFTEST_DCC_MSAA
);
442 device
->use_shader_ballot
= device
->use_aco
|| (device
->instance
->perftest_flags
& RADV_PERFTEST_SHADER_BALLOT
);
444 device
->use_ngg
= device
->rad_info
.chip_class
>= GFX10
&&
445 device
->rad_info
.family
!= CHIP_NAVI14
&&
446 !(device
->instance
->debug_flags
& RADV_DEBUG_NO_NGG
);
447 if (device
->use_aco
&& device
->use_ngg
) {
448 fprintf(stderr
, "WARNING: disabling NGG because ACO is used.\n");
449 device
->use_ngg
= false;
452 device
->use_ngg_streamout
= false;
454 /* Determine the number of threads per wave for all stages. */
455 device
->cs_wave_size
= 64;
456 device
->ps_wave_size
= 64;
457 device
->ge_wave_size
= 64;
459 if (device
->rad_info
.chip_class
>= GFX10
) {
460 if (device
->instance
->perftest_flags
& RADV_PERFTEST_CS_WAVE_32
)
461 device
->cs_wave_size
= 32;
463 /* For pixel shaders, wave64 is recommanded. */
464 if (device
->instance
->perftest_flags
& RADV_PERFTEST_PS_WAVE_32
)
465 device
->ps_wave_size
= 32;
467 if (device
->instance
->perftest_flags
& RADV_PERFTEST_GE_WAVE_32
)
468 device
->ge_wave_size
= 32;
471 radv_physical_device_init_mem_types(device
);
472 radv_fill_device_extension_table(device
, &device
->supported_extensions
);
474 device
->bus_info
= *drm_device
->businfo
.pci
;
476 if ((device
->instance
->debug_flags
& RADV_DEBUG_INFO
))
477 ac_print_gpu_info(&device
->rad_info
);
479 /* The WSI is structured as a layer on top of the driver, so this has
480 * to be the last part of initialization (at least until we get other
483 result
= radv_init_wsi(device
);
484 if (result
!= VK_SUCCESS
) {
485 device
->ws
->destroy(device
->ws
);
486 vk_error(instance
, result
);
500 radv_physical_device_finish(struct radv_physical_device
*device
)
502 radv_finish_wsi(device
);
503 device
->ws
->destroy(device
->ws
);
504 disk_cache_destroy(device
->disk_cache
);
505 close(device
->local_fd
);
506 if (device
->master_fd
!= -1)
507 close(device
->master_fd
);
511 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
512 VkSystemAllocationScope allocationScope
)
518 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
519 size_t align
, VkSystemAllocationScope allocationScope
)
521 return realloc(pOriginal
, size
);
525 default_free_func(void *pUserData
, void *pMemory
)
530 static const VkAllocationCallbacks default_alloc
= {
532 .pfnAllocation
= default_alloc_func
,
533 .pfnReallocation
= default_realloc_func
,
534 .pfnFree
= default_free_func
,
537 static const struct debug_control radv_debug_options
[] = {
538 {"nofastclears", RADV_DEBUG_NO_FAST_CLEARS
},
539 {"nodcc", RADV_DEBUG_NO_DCC
},
540 {"shaders", RADV_DEBUG_DUMP_SHADERS
},
541 {"nocache", RADV_DEBUG_NO_CACHE
},
542 {"shaderstats", RADV_DEBUG_DUMP_SHADER_STATS
},
543 {"nohiz", RADV_DEBUG_NO_HIZ
},
544 {"nocompute", RADV_DEBUG_NO_COMPUTE_QUEUE
},
545 {"allbos", RADV_DEBUG_ALL_BOS
},
546 {"noibs", RADV_DEBUG_NO_IBS
},
547 {"spirv", RADV_DEBUG_DUMP_SPIRV
},
548 {"vmfaults", RADV_DEBUG_VM_FAULTS
},
549 {"zerovram", RADV_DEBUG_ZERO_VRAM
},
550 {"syncshaders", RADV_DEBUG_SYNC_SHADERS
},
551 {"nosisched", RADV_DEBUG_NO_SISCHED
},
552 {"preoptir", RADV_DEBUG_PREOPTIR
},
553 {"nodynamicbounds", RADV_DEBUG_NO_DYNAMIC_BOUNDS
},
554 {"nooutoforder", RADV_DEBUG_NO_OUT_OF_ORDER
},
555 {"info", RADV_DEBUG_INFO
},
556 {"errors", RADV_DEBUG_ERRORS
},
557 {"startup", RADV_DEBUG_STARTUP
},
558 {"checkir", RADV_DEBUG_CHECKIR
},
559 {"nothreadllvm", RADV_DEBUG_NOTHREADLLVM
},
560 {"nobinning", RADV_DEBUG_NOBINNING
},
561 {"noloadstoreopt", RADV_DEBUG_NO_LOAD_STORE_OPT
},
562 {"nongg", RADV_DEBUG_NO_NGG
},
563 {"noshaderballot", RADV_DEBUG_NO_SHADER_BALLOT
},
564 {"allentrypoints", RADV_DEBUG_ALL_ENTRYPOINTS
},
565 {"metashaders", RADV_DEBUG_DUMP_META_SHADERS
},
566 {"nomemorycache", RADV_DEBUG_NO_MEMORY_CACHE
},
571 radv_get_debug_option_name(int id
)
573 assert(id
< ARRAY_SIZE(radv_debug_options
) - 1);
574 return radv_debug_options
[id
].string
;
577 static const struct debug_control radv_perftest_options
[] = {
578 {"nobatchchain", RADV_PERFTEST_NO_BATCHCHAIN
},
579 {"sisched", RADV_PERFTEST_SISCHED
},
580 {"localbos", RADV_PERFTEST_LOCAL_BOS
},
581 {"dccmsaa", RADV_PERFTEST_DCC_MSAA
},
582 {"bolist", RADV_PERFTEST_BO_LIST
},
583 {"shader_ballot", RADV_PERFTEST_SHADER_BALLOT
},
584 {"tccompatcmask", RADV_PERFTEST_TC_COMPAT_CMASK
},
585 {"cswave32", RADV_PERFTEST_CS_WAVE_32
},
586 {"pswave32", RADV_PERFTEST_PS_WAVE_32
},
587 {"gewave32", RADV_PERFTEST_GE_WAVE_32
},
588 {"dfsm", RADV_PERFTEST_DFSM
},
589 {"aco", RADV_PERFTEST_ACO
},
594 radv_get_perftest_option_name(int id
)
596 assert(id
< ARRAY_SIZE(radv_perftest_options
) - 1);
597 return radv_perftest_options
[id
].string
;
601 radv_handle_per_app_options(struct radv_instance
*instance
,
602 const VkApplicationInfo
*info
)
604 const char *name
= info
? info
->pApplicationName
: NULL
;
609 if (!strcmp(name
, "Talos - Linux - 32bit") ||
610 !strcmp(name
, "Talos - Linux - 64bit")) {
611 if (!(instance
->debug_flags
& RADV_DEBUG_NO_SISCHED
)) {
612 /* Force enable LLVM sisched for Talos because it looks
613 * safe and it gives few more FPS.
615 instance
->perftest_flags
|= RADV_PERFTEST_SISCHED
;
617 } else if (!strcmp(name
, "DOOM_VFR")) {
618 /* Work around a Doom VFR game bug */
619 instance
->debug_flags
|= RADV_DEBUG_NO_DYNAMIC_BOUNDS
;
620 } else if (!strcmp(name
, "MonsterHunterWorld.exe")) {
621 /* Workaround for a WaW hazard when LLVM moves/merges
622 * load/store memory operations.
623 * See https://reviews.llvm.org/D61313
625 if (LLVM_VERSION_MAJOR
< 9)
626 instance
->debug_flags
|= RADV_DEBUG_NO_LOAD_STORE_OPT
;
627 } else if (!strcmp(name
, "Wolfenstein: Youngblood")) {
628 if (!(instance
->debug_flags
& RADV_DEBUG_NO_SHADER_BALLOT
) &&
629 !(instance
->perftest_flags
& RADV_PERFTEST_ACO
)) {
630 /* Force enable VK_AMD_shader_ballot because it looks
631 * safe and it gives a nice boost (+20% on Vega 56 at
632 * this time). It also prevents corruption on LLVM.
634 instance
->perftest_flags
|= RADV_PERFTEST_SHADER_BALLOT
;
636 } else if (!strcmp(name
, "Fledge")) {
638 * Zero VRAM for "The Surge 2"
640 * This avoid a hang when when rendering any level. Likely
641 * uninitialized data in an indirect draw.
643 instance
->debug_flags
|= RADV_DEBUG_ZERO_VRAM
;
647 static int radv_get_instance_extension_index(const char *name
)
649 for (unsigned i
= 0; i
< RADV_INSTANCE_EXTENSION_COUNT
; ++i
) {
650 if (strcmp(name
, radv_instance_extensions
[i
].extensionName
) == 0)
656 static const char radv_dri_options_xml
[] =
658 DRI_CONF_SECTION_PERFORMANCE
659 DRI_CONF_ADAPTIVE_SYNC("true")
660 DRI_CONF_VK_X11_OVERRIDE_MIN_IMAGE_COUNT(0)
661 DRI_CONF_VK_X11_STRICT_IMAGE_COUNT("false")
665 static void radv_init_dri_options(struct radv_instance
*instance
)
667 driParseOptionInfo(&instance
->available_dri_options
, radv_dri_options_xml
);
668 driParseConfigFiles(&instance
->dri_options
,
669 &instance
->available_dri_options
,
671 instance
->engineName
,
672 instance
->engineVersion
);
675 VkResult
radv_CreateInstance(
676 const VkInstanceCreateInfo
* pCreateInfo
,
677 const VkAllocationCallbacks
* pAllocator
,
678 VkInstance
* pInstance
)
680 struct radv_instance
*instance
;
683 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
685 uint32_t client_version
;
686 if (pCreateInfo
->pApplicationInfo
&&
687 pCreateInfo
->pApplicationInfo
->apiVersion
!= 0) {
688 client_version
= pCreateInfo
->pApplicationInfo
->apiVersion
;
690 client_version
= VK_API_VERSION_1_0
;
693 const char *engine_name
= NULL
;
694 uint32_t engine_version
= 0;
695 if (pCreateInfo
->pApplicationInfo
) {
696 engine_name
= pCreateInfo
->pApplicationInfo
->pEngineName
;
697 engine_version
= pCreateInfo
->pApplicationInfo
->engineVersion
;
700 instance
= vk_zalloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
701 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
703 return vk_error(NULL
, VK_ERROR_OUT_OF_HOST_MEMORY
);
705 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
708 instance
->alloc
= *pAllocator
;
710 instance
->alloc
= default_alloc
;
712 instance
->apiVersion
= client_version
;
713 instance
->physicalDeviceCount
= -1;
715 /* Get secure compile thread count. NOTE: We cap this at 32 */
716 #define MAX_SC_PROCS 32
717 char *num_sc_threads
= getenv("RADV_SECURE_COMPILE_THREADS");
719 instance
->num_sc_threads
= MIN2(strtoul(num_sc_threads
, NULL
, 10), MAX_SC_PROCS
);
721 instance
->debug_flags
= parse_debug_string(getenv("RADV_DEBUG"),
724 /* Disable memory cache when secure compile is set */
725 if (radv_device_use_secure_compile(instance
))
726 instance
->debug_flags
|= RADV_DEBUG_NO_MEMORY_CACHE
;
728 instance
->perftest_flags
= parse_debug_string(getenv("RADV_PERFTEST"),
729 radv_perftest_options
);
731 if (instance
->perftest_flags
& RADV_PERFTEST_ACO
)
732 fprintf(stderr
, "WARNING: Experimental compiler backend enabled. Here be dragons! Incorrect rendering, GPU hangs and/or resets are likely\n");
734 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
735 radv_logi("Created an instance");
737 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
738 const char *ext_name
= pCreateInfo
->ppEnabledExtensionNames
[i
];
739 int index
= radv_get_instance_extension_index(ext_name
);
741 if (index
< 0 || !radv_supported_instance_extensions
.extensions
[index
]) {
742 vk_free2(&default_alloc
, pAllocator
, instance
);
743 return vk_error(instance
, VK_ERROR_EXTENSION_NOT_PRESENT
);
746 instance
->enabled_extensions
.extensions
[index
] = true;
749 result
= vk_debug_report_instance_init(&instance
->debug_report_callbacks
);
750 if (result
!= VK_SUCCESS
) {
751 vk_free2(&default_alloc
, pAllocator
, instance
);
752 return vk_error(instance
, result
);
755 instance
->engineName
= vk_strdup(&instance
->alloc
, engine_name
,
756 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
757 instance
->engineVersion
= engine_version
;
759 glsl_type_singleton_init_or_ref();
761 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
763 radv_init_dri_options(instance
);
764 radv_handle_per_app_options(instance
, pCreateInfo
->pApplicationInfo
);
766 *pInstance
= radv_instance_to_handle(instance
);
771 void radv_DestroyInstance(
772 VkInstance _instance
,
773 const VkAllocationCallbacks
* pAllocator
)
775 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
780 for (int i
= 0; i
< instance
->physicalDeviceCount
; ++i
) {
781 radv_physical_device_finish(instance
->physicalDevices
+ i
);
784 vk_free(&instance
->alloc
, instance
->engineName
);
786 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
788 glsl_type_singleton_decref();
790 driDestroyOptionCache(&instance
->dri_options
);
791 driDestroyOptionInfo(&instance
->available_dri_options
);
793 vk_debug_report_instance_destroy(&instance
->debug_report_callbacks
);
795 vk_free(&instance
->alloc
, instance
);
799 radv_enumerate_devices(struct radv_instance
*instance
)
801 /* TODO: Check for more devices ? */
802 drmDevicePtr devices
[8];
803 VkResult result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
806 instance
->physicalDeviceCount
= 0;
808 max_devices
= drmGetDevices2(0, devices
, ARRAY_SIZE(devices
));
810 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
811 radv_logi("Found %d drm nodes", max_devices
);
814 return vk_error(instance
, VK_ERROR_INCOMPATIBLE_DRIVER
);
816 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
817 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
818 devices
[i
]->bustype
== DRM_BUS_PCI
&&
819 devices
[i
]->deviceinfo
.pci
->vendor_id
== ATI_VENDOR_ID
) {
821 result
= radv_physical_device_init(instance
->physicalDevices
+
822 instance
->physicalDeviceCount
,
825 if (result
== VK_SUCCESS
)
826 ++instance
->physicalDeviceCount
;
827 else if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
831 drmFreeDevices(devices
, max_devices
);
836 VkResult
radv_EnumeratePhysicalDevices(
837 VkInstance _instance
,
838 uint32_t* pPhysicalDeviceCount
,
839 VkPhysicalDevice
* pPhysicalDevices
)
841 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
844 if (instance
->physicalDeviceCount
< 0) {
845 result
= radv_enumerate_devices(instance
);
846 if (result
!= VK_SUCCESS
&&
847 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
851 if (!pPhysicalDevices
) {
852 *pPhysicalDeviceCount
= instance
->physicalDeviceCount
;
854 *pPhysicalDeviceCount
= MIN2(*pPhysicalDeviceCount
, instance
->physicalDeviceCount
);
855 for (unsigned i
= 0; i
< *pPhysicalDeviceCount
; ++i
)
856 pPhysicalDevices
[i
] = radv_physical_device_to_handle(instance
->physicalDevices
+ i
);
859 return *pPhysicalDeviceCount
< instance
->physicalDeviceCount
? VK_INCOMPLETE
863 VkResult
radv_EnumeratePhysicalDeviceGroups(
864 VkInstance _instance
,
865 uint32_t* pPhysicalDeviceGroupCount
,
866 VkPhysicalDeviceGroupProperties
* pPhysicalDeviceGroupProperties
)
868 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
871 if (instance
->physicalDeviceCount
< 0) {
872 result
= radv_enumerate_devices(instance
);
873 if (result
!= VK_SUCCESS
&&
874 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
878 if (!pPhysicalDeviceGroupProperties
) {
879 *pPhysicalDeviceGroupCount
= instance
->physicalDeviceCount
;
881 *pPhysicalDeviceGroupCount
= MIN2(*pPhysicalDeviceGroupCount
, instance
->physicalDeviceCount
);
882 for (unsigned i
= 0; i
< *pPhysicalDeviceGroupCount
; ++i
) {
883 pPhysicalDeviceGroupProperties
[i
].physicalDeviceCount
= 1;
884 pPhysicalDeviceGroupProperties
[i
].physicalDevices
[0] = radv_physical_device_to_handle(instance
->physicalDevices
+ i
);
885 pPhysicalDeviceGroupProperties
[i
].subsetAllocation
= false;
888 return *pPhysicalDeviceGroupCount
< instance
->physicalDeviceCount
? VK_INCOMPLETE
892 void radv_GetPhysicalDeviceFeatures(
893 VkPhysicalDevice physicalDevice
,
894 VkPhysicalDeviceFeatures
* pFeatures
)
896 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
897 memset(pFeatures
, 0, sizeof(*pFeatures
));
899 *pFeatures
= (VkPhysicalDeviceFeatures
) {
900 .robustBufferAccess
= true,
901 .fullDrawIndexUint32
= true,
902 .imageCubeArray
= true,
903 .independentBlend
= true,
904 .geometryShader
= true,
905 .tessellationShader
= true,
906 .sampleRateShading
= true,
907 .dualSrcBlend
= true,
909 .multiDrawIndirect
= true,
910 .drawIndirectFirstInstance
= true,
912 .depthBiasClamp
= true,
913 .fillModeNonSolid
= true,
918 .multiViewport
= true,
919 .samplerAnisotropy
= true,
920 .textureCompressionETC2
= radv_device_supports_etc(pdevice
),
921 .textureCompressionASTC_LDR
= false,
922 .textureCompressionBC
= true,
923 .occlusionQueryPrecise
= true,
924 .pipelineStatisticsQuery
= true,
925 .vertexPipelineStoresAndAtomics
= true,
926 .fragmentStoresAndAtomics
= true,
927 .shaderTessellationAndGeometryPointSize
= true,
928 .shaderImageGatherExtended
= true,
929 .shaderStorageImageExtendedFormats
= true,
930 .shaderStorageImageMultisample
= pdevice
->rad_info
.chip_class
>= GFX8
,
931 .shaderUniformBufferArrayDynamicIndexing
= true,
932 .shaderSampledImageArrayDynamicIndexing
= true,
933 .shaderStorageBufferArrayDynamicIndexing
= true,
934 .shaderStorageImageArrayDynamicIndexing
= true,
935 .shaderStorageImageReadWithoutFormat
= true,
936 .shaderStorageImageWriteWithoutFormat
= true,
937 .shaderClipDistance
= true,
938 .shaderCullDistance
= true,
939 .shaderFloat64
= true,
941 .shaderInt16
= pdevice
->rad_info
.chip_class
>= GFX9
&& !pdevice
->use_aco
,
942 .sparseBinding
= true,
943 .variableMultisampleRate
= true,
944 .inheritedQueries
= true,
948 void radv_GetPhysicalDeviceFeatures2(
949 VkPhysicalDevice physicalDevice
,
950 VkPhysicalDeviceFeatures2
*pFeatures
)
952 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
953 vk_foreach_struct(ext
, pFeatures
->pNext
) {
954 switch (ext
->sType
) {
955 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES
: {
956 VkPhysicalDeviceVariablePointersFeatures
*features
= (void *)ext
;
957 features
->variablePointersStorageBuffer
= true;
958 features
->variablePointers
= true;
961 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES
: {
962 VkPhysicalDeviceMultiviewFeatures
*features
= (VkPhysicalDeviceMultiviewFeatures
*)ext
;
963 features
->multiview
= true;
964 features
->multiviewGeometryShader
= true;
965 features
->multiviewTessellationShader
= true;
968 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETERS_FEATURES
: {
969 VkPhysicalDeviceShaderDrawParametersFeatures
*features
=
970 (VkPhysicalDeviceShaderDrawParametersFeatures
*)ext
;
971 features
->shaderDrawParameters
= true;
974 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES
: {
975 VkPhysicalDeviceProtectedMemoryFeatures
*features
=
976 (VkPhysicalDeviceProtectedMemoryFeatures
*)ext
;
977 features
->protectedMemory
= false;
980 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES
: {
981 VkPhysicalDevice16BitStorageFeatures
*features
=
982 (VkPhysicalDevice16BitStorageFeatures
*)ext
;
983 bool enabled
= pdevice
->rad_info
.chip_class
>= GFX8
&& !pdevice
->use_aco
;
984 features
->storageBuffer16BitAccess
= enabled
;
985 features
->uniformAndStorageBuffer16BitAccess
= enabled
;
986 features
->storagePushConstant16
= enabled
;
987 features
->storageInputOutput16
= enabled
&& LLVM_VERSION_MAJOR
>= 9;
990 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES
: {
991 VkPhysicalDeviceSamplerYcbcrConversionFeatures
*features
=
992 (VkPhysicalDeviceSamplerYcbcrConversionFeatures
*)ext
;
993 features
->samplerYcbcrConversion
= true;
996 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT
: {
997 VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*features
=
998 (VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*)ext
;
999 features
->shaderInputAttachmentArrayDynamicIndexing
= true;
1000 features
->shaderUniformTexelBufferArrayDynamicIndexing
= true;
1001 features
->shaderStorageTexelBufferArrayDynamicIndexing
= true;
1002 features
->shaderUniformBufferArrayNonUniformIndexing
= true;
1003 features
->shaderSampledImageArrayNonUniformIndexing
= true;
1004 features
->shaderStorageBufferArrayNonUniformIndexing
= true;
1005 features
->shaderStorageImageArrayNonUniformIndexing
= true;
1006 features
->shaderInputAttachmentArrayNonUniformIndexing
= true;
1007 features
->shaderUniformTexelBufferArrayNonUniformIndexing
= true;
1008 features
->shaderStorageTexelBufferArrayNonUniformIndexing
= true;
1009 features
->descriptorBindingUniformBufferUpdateAfterBind
= true;
1010 features
->descriptorBindingSampledImageUpdateAfterBind
= true;
1011 features
->descriptorBindingStorageImageUpdateAfterBind
= true;
1012 features
->descriptorBindingStorageBufferUpdateAfterBind
= true;
1013 features
->descriptorBindingUniformTexelBufferUpdateAfterBind
= true;
1014 features
->descriptorBindingStorageTexelBufferUpdateAfterBind
= true;
1015 features
->descriptorBindingUpdateUnusedWhilePending
= true;
1016 features
->descriptorBindingPartiallyBound
= true;
1017 features
->descriptorBindingVariableDescriptorCount
= true;
1018 features
->runtimeDescriptorArray
= true;
1021 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT
: {
1022 VkPhysicalDeviceConditionalRenderingFeaturesEXT
*features
=
1023 (VkPhysicalDeviceConditionalRenderingFeaturesEXT
*)ext
;
1024 features
->conditionalRendering
= true;
1025 features
->inheritedConditionalRendering
= false;
1028 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT
: {
1029 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*features
=
1030 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*)ext
;
1031 features
->vertexAttributeInstanceRateDivisor
= VK_TRUE
;
1032 features
->vertexAttributeInstanceRateZeroDivisor
= VK_TRUE
;
1035 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT
: {
1036 VkPhysicalDeviceTransformFeedbackFeaturesEXT
*features
=
1037 (VkPhysicalDeviceTransformFeedbackFeaturesEXT
*)ext
;
1038 features
->transformFeedback
= true;
1039 features
->geometryStreams
= !pdevice
->use_ngg_streamout
;
1042 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES_EXT
: {
1043 VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*features
=
1044 (VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*)ext
;
1045 features
->scalarBlockLayout
= pdevice
->rad_info
.chip_class
>= GFX7
;
1048 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_PRIORITY_FEATURES_EXT
: {
1049 VkPhysicalDeviceMemoryPriorityFeaturesEXT
*features
=
1050 (VkPhysicalDeviceMemoryPriorityFeaturesEXT
*)ext
;
1051 features
->memoryPriority
= VK_TRUE
;
1054 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_EXT
: {
1055 VkPhysicalDeviceBufferDeviceAddressFeaturesEXT
*features
=
1056 (VkPhysicalDeviceBufferDeviceAddressFeaturesEXT
*)ext
;
1057 features
->bufferDeviceAddress
= true;
1058 features
->bufferDeviceAddressCaptureReplay
= false;
1059 features
->bufferDeviceAddressMultiDevice
= false;
1062 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_KHR
: {
1063 VkPhysicalDeviceBufferDeviceAddressFeaturesKHR
*features
=
1064 (VkPhysicalDeviceBufferDeviceAddressFeaturesKHR
*)ext
;
1065 features
->bufferDeviceAddress
= true;
1066 features
->bufferDeviceAddressCaptureReplay
= false;
1067 features
->bufferDeviceAddressMultiDevice
= false;
1070 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_CLIP_ENABLE_FEATURES_EXT
: {
1071 VkPhysicalDeviceDepthClipEnableFeaturesEXT
*features
=
1072 (VkPhysicalDeviceDepthClipEnableFeaturesEXT
*)ext
;
1073 features
->depthClipEnable
= true;
1076 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_QUERY_RESET_FEATURES_EXT
: {
1077 VkPhysicalDeviceHostQueryResetFeaturesEXT
*features
=
1078 (VkPhysicalDeviceHostQueryResetFeaturesEXT
*)ext
;
1079 features
->hostQueryReset
= true;
1082 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES_KHR
: {
1083 VkPhysicalDevice8BitStorageFeaturesKHR
*features
=
1084 (VkPhysicalDevice8BitStorageFeaturesKHR
*)ext
;
1085 bool enabled
= pdevice
->rad_info
.chip_class
>= GFX8
&& !pdevice
->use_aco
;
1086 features
->storageBuffer8BitAccess
= enabled
;
1087 features
->uniformAndStorageBuffer8BitAccess
= enabled
;
1088 features
->storagePushConstant8
= enabled
;
1091 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_FLOAT16_INT8_FEATURES_KHR
: {
1092 VkPhysicalDeviceShaderFloat16Int8FeaturesKHR
*features
=
1093 (VkPhysicalDeviceShaderFloat16Int8FeaturesKHR
*)ext
;
1094 features
->shaderFloat16
= pdevice
->rad_info
.chip_class
>= GFX8
&& !pdevice
->use_aco
;
1095 features
->shaderInt8
= !pdevice
->use_aco
;
1098 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_INT64_FEATURES_KHR
: {
1099 VkPhysicalDeviceShaderAtomicInt64FeaturesKHR
*features
=
1100 (VkPhysicalDeviceShaderAtomicInt64FeaturesKHR
*)ext
;
1101 features
->shaderBufferInt64Atomics
= LLVM_VERSION_MAJOR
>= 9;
1102 features
->shaderSharedInt64Atomics
= LLVM_VERSION_MAJOR
>= 9;
1105 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DEMOTE_TO_HELPER_INVOCATION_FEATURES_EXT
: {
1106 VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT
*features
=
1107 (VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT
*)ext
;
1108 features
->shaderDemoteToHelperInvocation
= pdevice
->use_aco
;
1111 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_FEATURES_EXT
: {
1112 VkPhysicalDeviceInlineUniformBlockFeaturesEXT
*features
=
1113 (VkPhysicalDeviceInlineUniformBlockFeaturesEXT
*)ext
;
1115 features
->inlineUniformBlock
= true;
1116 features
->descriptorBindingInlineUniformBlockUpdateAfterBind
= true;
1119 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COMPUTE_SHADER_DERIVATIVES_FEATURES_NV
: {
1120 VkPhysicalDeviceComputeShaderDerivativesFeaturesNV
*features
=
1121 (VkPhysicalDeviceComputeShaderDerivativesFeaturesNV
*)ext
;
1122 features
->computeDerivativeGroupQuads
= false;
1123 features
->computeDerivativeGroupLinear
= true;
1126 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_YCBCR_IMAGE_ARRAYS_FEATURES_EXT
: {
1127 VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*features
=
1128 (VkPhysicalDeviceYcbcrImageArraysFeaturesEXT
*)ext
;
1129 features
->ycbcrImageArrays
= true;
1132 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_UNIFORM_BUFFER_STANDARD_LAYOUT_FEATURES_KHR
: {
1133 VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR
*features
=
1134 (VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR
*)ext
;
1135 features
->uniformBufferStandardLayout
= true;
1138 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INDEX_TYPE_UINT8_FEATURES_EXT
: {
1139 VkPhysicalDeviceIndexTypeUint8FeaturesEXT
*features
=
1140 (VkPhysicalDeviceIndexTypeUint8FeaturesEXT
*)ext
;
1141 features
->indexTypeUint8
= pdevice
->rad_info
.chip_class
>= GFX8
;
1144 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGELESS_FRAMEBUFFER_FEATURES_KHR
: {
1145 VkPhysicalDeviceImagelessFramebufferFeaturesKHR
*features
=
1146 (VkPhysicalDeviceImagelessFramebufferFeaturesKHR
*)ext
;
1147 features
->imagelessFramebuffer
= true;
1150 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR
: {
1151 VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR
*features
=
1152 (VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR
*)ext
;
1153 features
->pipelineExecutableInfo
= true;
1156 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CLOCK_FEATURES_KHR
: {
1157 VkPhysicalDeviceShaderClockFeaturesKHR
*features
=
1158 (VkPhysicalDeviceShaderClockFeaturesKHR
*)ext
;
1159 features
->shaderSubgroupClock
= true;
1160 features
->shaderDeviceClock
= false;
1163 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_FEATURES_EXT
: {
1164 VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT
*features
=
1165 (VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT
*)ext
;
1166 features
->texelBufferAlignment
= true;
1169 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_FEATURES_KHR
: {
1170 VkPhysicalDeviceTimelineSemaphoreFeaturesKHR
*features
=
1171 (VkPhysicalDeviceTimelineSemaphoreFeaturesKHR
*) ext
;
1172 features
->timelineSemaphore
= true;
1175 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_FEATURES_EXT
: {
1176 VkPhysicalDeviceSubgroupSizeControlFeaturesEXT
*features
=
1177 (VkPhysicalDeviceSubgroupSizeControlFeaturesEXT
*)ext
;
1178 features
->subgroupSizeControl
= true;
1179 features
->computeFullSubgroups
= true;
1182 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COHERENT_MEMORY_FEATURES_AMD
: {
1183 VkPhysicalDeviceCoherentMemoryFeaturesAMD
*features
=
1184 (VkPhysicalDeviceCoherentMemoryFeaturesAMD
*)ext
;
1185 features
->deviceCoherentMemory
= pdevice
->rad_info
.has_l2_uncached
;
1188 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_SUBGROUP_EXTENDED_TYPES_FEATURES_KHR
: {
1189 VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR
*features
=
1190 (VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR
*)ext
;
1191 features
->shaderSubgroupExtendedTypes
= true;
1198 return radv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
1202 radv_max_descriptor_set_size()
1204 /* make sure that the entire descriptor set is addressable with a signed
1205 * 32-bit int. So the sum of all limits scaled by descriptor size has to
1206 * be at most 2 GiB. the combined image & samples object count as one of
1207 * both. This limit is for the pipeline layout, not for the set layout, but
1208 * there is no set limit, so we just set a pipeline limit. I don't think
1209 * any app is going to hit this soon. */
1210 return ((1ull << 31) - 16 * MAX_DYNAMIC_BUFFERS
1211 - MAX_INLINE_UNIFORM_BLOCK_SIZE
* MAX_INLINE_UNIFORM_BLOCK_COUNT
) /
1212 (32 /* uniform buffer, 32 due to potential space wasted on alignment */ +
1213 32 /* storage buffer, 32 due to potential space wasted on alignment */ +
1214 32 /* sampler, largest when combined with image */ +
1215 64 /* sampled image */ +
1216 64 /* storage image */);
1219 void radv_GetPhysicalDeviceProperties(
1220 VkPhysicalDevice physicalDevice
,
1221 VkPhysicalDeviceProperties
* pProperties
)
1223 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
1224 VkSampleCountFlags sample_counts
= 0xf;
1226 size_t max_descriptor_set_size
= radv_max_descriptor_set_size();
1228 VkPhysicalDeviceLimits limits
= {
1229 .maxImageDimension1D
= (1 << 14),
1230 .maxImageDimension2D
= (1 << 14),
1231 .maxImageDimension3D
= (1 << 11),
1232 .maxImageDimensionCube
= (1 << 14),
1233 .maxImageArrayLayers
= (1 << 11),
1234 .maxTexelBufferElements
= 128 * 1024 * 1024,
1235 .maxUniformBufferRange
= UINT32_MAX
,
1236 .maxStorageBufferRange
= UINT32_MAX
,
1237 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
1238 .maxMemoryAllocationCount
= UINT32_MAX
,
1239 .maxSamplerAllocationCount
= 64 * 1024,
1240 .bufferImageGranularity
= 64, /* A cache line */
1241 .sparseAddressSpaceSize
= 0xffffffffu
, /* buffer max size */
1242 .maxBoundDescriptorSets
= MAX_SETS
,
1243 .maxPerStageDescriptorSamplers
= max_descriptor_set_size
,
1244 .maxPerStageDescriptorUniformBuffers
= max_descriptor_set_size
,
1245 .maxPerStageDescriptorStorageBuffers
= max_descriptor_set_size
,
1246 .maxPerStageDescriptorSampledImages
= max_descriptor_set_size
,
1247 .maxPerStageDescriptorStorageImages
= max_descriptor_set_size
,
1248 .maxPerStageDescriptorInputAttachments
= max_descriptor_set_size
,
1249 .maxPerStageResources
= max_descriptor_set_size
,
1250 .maxDescriptorSetSamplers
= max_descriptor_set_size
,
1251 .maxDescriptorSetUniformBuffers
= max_descriptor_set_size
,
1252 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_UNIFORM_BUFFERS
,
1253 .maxDescriptorSetStorageBuffers
= max_descriptor_set_size
,
1254 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_STORAGE_BUFFERS
,
1255 .maxDescriptorSetSampledImages
= max_descriptor_set_size
,
1256 .maxDescriptorSetStorageImages
= max_descriptor_set_size
,
1257 .maxDescriptorSetInputAttachments
= max_descriptor_set_size
,
1258 .maxVertexInputAttributes
= MAX_VERTEX_ATTRIBS
,
1259 .maxVertexInputBindings
= MAX_VBS
,
1260 .maxVertexInputAttributeOffset
= 2047,
1261 .maxVertexInputBindingStride
= 2048,
1262 .maxVertexOutputComponents
= 128,
1263 .maxTessellationGenerationLevel
= 64,
1264 .maxTessellationPatchSize
= 32,
1265 .maxTessellationControlPerVertexInputComponents
= 128,
1266 .maxTessellationControlPerVertexOutputComponents
= 128,
1267 .maxTessellationControlPerPatchOutputComponents
= 120,
1268 .maxTessellationControlTotalOutputComponents
= 4096,
1269 .maxTessellationEvaluationInputComponents
= 128,
1270 .maxTessellationEvaluationOutputComponents
= 128,
1271 .maxGeometryShaderInvocations
= 127,
1272 .maxGeometryInputComponents
= 64,
1273 .maxGeometryOutputComponents
= 128,
1274 .maxGeometryOutputVertices
= 256,
1275 .maxGeometryTotalOutputComponents
= 1024,
1276 .maxFragmentInputComponents
= 128,
1277 .maxFragmentOutputAttachments
= 8,
1278 .maxFragmentDualSrcAttachments
= 1,
1279 .maxFragmentCombinedOutputResources
= 8,
1280 .maxComputeSharedMemorySize
= 32768,
1281 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
1282 .maxComputeWorkGroupInvocations
= 2048,
1283 .maxComputeWorkGroupSize
= {
1288 .subPixelPrecisionBits
= 8,
1289 .subTexelPrecisionBits
= 8,
1290 .mipmapPrecisionBits
= 8,
1291 .maxDrawIndexedIndexValue
= UINT32_MAX
,
1292 .maxDrawIndirectCount
= UINT32_MAX
,
1293 .maxSamplerLodBias
= 16,
1294 .maxSamplerAnisotropy
= 16,
1295 .maxViewports
= MAX_VIEWPORTS
,
1296 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
1297 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
1298 .viewportSubPixelBits
= 8,
1299 .minMemoryMapAlignment
= 4096, /* A page */
1300 .minTexelBufferOffsetAlignment
= 4,
1301 .minUniformBufferOffsetAlignment
= 4,
1302 .minStorageBufferOffsetAlignment
= 4,
1303 .minTexelOffset
= -32,
1304 .maxTexelOffset
= 31,
1305 .minTexelGatherOffset
= -32,
1306 .maxTexelGatherOffset
= 31,
1307 .minInterpolationOffset
= -2,
1308 .maxInterpolationOffset
= 2,
1309 .subPixelInterpolationOffsetBits
= 8,
1310 .maxFramebufferWidth
= (1 << 14),
1311 .maxFramebufferHeight
= (1 << 14),
1312 .maxFramebufferLayers
= (1 << 10),
1313 .framebufferColorSampleCounts
= sample_counts
,
1314 .framebufferDepthSampleCounts
= sample_counts
,
1315 .framebufferStencilSampleCounts
= sample_counts
,
1316 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
1317 .maxColorAttachments
= MAX_RTS
,
1318 .sampledImageColorSampleCounts
= sample_counts
,
1319 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
1320 .sampledImageDepthSampleCounts
= sample_counts
,
1321 .sampledImageStencilSampleCounts
= sample_counts
,
1322 .storageImageSampleCounts
= pdevice
->rad_info
.chip_class
>= GFX8
? sample_counts
: VK_SAMPLE_COUNT_1_BIT
,
1323 .maxSampleMaskWords
= 1,
1324 .timestampComputeAndGraphics
= true,
1325 .timestampPeriod
= 1000000.0 / pdevice
->rad_info
.clock_crystal_freq
,
1326 .maxClipDistances
= 8,
1327 .maxCullDistances
= 8,
1328 .maxCombinedClipAndCullDistances
= 8,
1329 .discreteQueuePriorities
= 2,
1330 .pointSizeRange
= { 0.0, 8192.0 },
1331 .lineWidthRange
= { 0.0, 7.9921875 },
1332 .pointSizeGranularity
= (1.0 / 8.0),
1333 .lineWidthGranularity
= (1.0 / 128.0),
1334 .strictLines
= false, /* FINISHME */
1335 .standardSampleLocations
= true,
1336 .optimalBufferCopyOffsetAlignment
= 128,
1337 .optimalBufferCopyRowPitchAlignment
= 128,
1338 .nonCoherentAtomSize
= 64,
1341 *pProperties
= (VkPhysicalDeviceProperties
) {
1342 .apiVersion
= radv_physical_device_api_version(pdevice
),
1343 .driverVersion
= vk_get_driver_version(),
1344 .vendorID
= ATI_VENDOR_ID
,
1345 .deviceID
= pdevice
->rad_info
.pci_id
,
1346 .deviceType
= pdevice
->rad_info
.has_dedicated_vram
? VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU
: VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
1348 .sparseProperties
= {0},
1351 strcpy(pProperties
->deviceName
, pdevice
->name
);
1352 memcpy(pProperties
->pipelineCacheUUID
, pdevice
->cache_uuid
, VK_UUID_SIZE
);
1355 void radv_GetPhysicalDeviceProperties2(
1356 VkPhysicalDevice physicalDevice
,
1357 VkPhysicalDeviceProperties2
*pProperties
)
1359 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
1360 radv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
1362 vk_foreach_struct(ext
, pProperties
->pNext
) {
1363 switch (ext
->sType
) {
1364 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
1365 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
1366 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
1367 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
1370 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES
: {
1371 VkPhysicalDeviceIDProperties
*properties
= (VkPhysicalDeviceIDProperties
*)ext
;
1372 memcpy(properties
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
1373 memcpy(properties
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
1374 properties
->deviceLUIDValid
= false;
1377 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES
: {
1378 VkPhysicalDeviceMultiviewProperties
*properties
= (VkPhysicalDeviceMultiviewProperties
*)ext
;
1379 properties
->maxMultiviewViewCount
= MAX_VIEWS
;
1380 properties
->maxMultiviewInstanceIndex
= INT_MAX
;
1383 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES
: {
1384 VkPhysicalDevicePointClippingProperties
*properties
=
1385 (VkPhysicalDevicePointClippingProperties
*)ext
;
1386 properties
->pointClippingBehavior
= VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES
;
1389 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DISCARD_RECTANGLE_PROPERTIES_EXT
: {
1390 VkPhysicalDeviceDiscardRectanglePropertiesEXT
*properties
=
1391 (VkPhysicalDeviceDiscardRectanglePropertiesEXT
*)ext
;
1392 properties
->maxDiscardRectangles
= MAX_DISCARD_RECTANGLES
;
1395 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT
: {
1396 VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*properties
=
1397 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*) ext
;
1398 properties
->minImportedHostPointerAlignment
= 4096;
1401 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES
: {
1402 VkPhysicalDeviceSubgroupProperties
*properties
=
1403 (VkPhysicalDeviceSubgroupProperties
*)ext
;
1404 properties
->subgroupSize
= RADV_SUBGROUP_SIZE
;
1405 properties
->supportedStages
= VK_SHADER_STAGE_ALL
;
1406 properties
->supportedOperations
=
1407 VK_SUBGROUP_FEATURE_BASIC_BIT
|
1408 VK_SUBGROUP_FEATURE_VOTE_BIT
|
1409 VK_SUBGROUP_FEATURE_ARITHMETIC_BIT
|
1410 VK_SUBGROUP_FEATURE_BALLOT_BIT
|
1411 VK_SUBGROUP_FEATURE_CLUSTERED_BIT
|
1412 VK_SUBGROUP_FEATURE_QUAD_BIT
;
1413 if (pdevice
->rad_info
.chip_class
== GFX8
||
1414 pdevice
->rad_info
.chip_class
== GFX9
) {
1415 properties
->supportedOperations
|=
1416 VK_SUBGROUP_FEATURE_SHUFFLE_BIT
|
1417 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT
;
1419 properties
->quadOperationsInAllStages
= true;
1422 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES
: {
1423 VkPhysicalDeviceMaintenance3Properties
*properties
=
1424 (VkPhysicalDeviceMaintenance3Properties
*)ext
;
1425 properties
->maxPerSetDescriptors
= RADV_MAX_PER_SET_DESCRIPTORS
;
1426 properties
->maxMemoryAllocationSize
= RADV_MAX_MEMORY_ALLOCATION_SIZE
;
1429 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT
: {
1430 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*properties
=
1431 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*)ext
;
1432 /* GFX6-8 only support single channel min/max filter. */
1433 properties
->filterMinmaxImageComponentMapping
= pdevice
->rad_info
.chip_class
>= GFX9
;
1434 properties
->filterMinmaxSingleComponentFormats
= true;
1437 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CORE_PROPERTIES_AMD
: {
1438 VkPhysicalDeviceShaderCorePropertiesAMD
*properties
=
1439 (VkPhysicalDeviceShaderCorePropertiesAMD
*)ext
;
1441 /* Shader engines. */
1442 properties
->shaderEngineCount
=
1443 pdevice
->rad_info
.max_se
;
1444 properties
->shaderArraysPerEngineCount
=
1445 pdevice
->rad_info
.max_sh_per_se
;
1446 properties
->computeUnitsPerShaderArray
=
1447 pdevice
->rad_info
.num_good_cu_per_sh
;
1448 properties
->simdPerComputeUnit
= 4;
1449 properties
->wavefrontsPerSimd
=
1450 pdevice
->rad_info
.family
== CHIP_TONGA
||
1451 pdevice
->rad_info
.family
== CHIP_ICELAND
||
1452 pdevice
->rad_info
.family
== CHIP_POLARIS10
||
1453 pdevice
->rad_info
.family
== CHIP_POLARIS11
||
1454 pdevice
->rad_info
.family
== CHIP_POLARIS12
||
1455 pdevice
->rad_info
.family
== CHIP_VEGAM
? 8 : 10;
1456 properties
->wavefrontSize
= 64;
1459 properties
->sgprsPerSimd
=
1460 pdevice
->rad_info
.num_physical_sgprs_per_simd
;
1461 properties
->minSgprAllocation
=
1462 pdevice
->rad_info
.chip_class
>= GFX8
? 16 : 8;
1463 properties
->maxSgprAllocation
=
1464 pdevice
->rad_info
.family
== CHIP_TONGA
||
1465 pdevice
->rad_info
.family
== CHIP_ICELAND
? 96 : 104;
1466 properties
->sgprAllocationGranularity
=
1467 pdevice
->rad_info
.chip_class
>= GFX8
? 16 : 8;
1470 properties
->vgprsPerSimd
= RADV_NUM_PHYSICAL_VGPRS
;
1471 properties
->minVgprAllocation
= 4;
1472 properties
->maxVgprAllocation
= 256;
1473 properties
->vgprAllocationGranularity
= 4;
1476 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CORE_PROPERTIES_2_AMD
: {
1477 VkPhysicalDeviceShaderCoreProperties2AMD
*properties
=
1478 (VkPhysicalDeviceShaderCoreProperties2AMD
*)ext
;
1480 properties
->shaderCoreFeatures
= 0;
1481 properties
->activeComputeUnitCount
=
1482 pdevice
->rad_info
.num_good_compute_units
;
1485 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT
: {
1486 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*properties
=
1487 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*)ext
;
1488 properties
->maxVertexAttribDivisor
= UINT32_MAX
;
1491 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES_EXT
: {
1492 VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*properties
=
1493 (VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*)ext
;
1494 properties
->maxUpdateAfterBindDescriptorsInAllPools
= UINT32_MAX
/ 64;
1495 properties
->shaderUniformBufferArrayNonUniformIndexingNative
= false;
1496 properties
->shaderSampledImageArrayNonUniformIndexingNative
= false;
1497 properties
->shaderStorageBufferArrayNonUniformIndexingNative
= false;
1498 properties
->shaderStorageImageArrayNonUniformIndexingNative
= false;
1499 properties
->shaderInputAttachmentArrayNonUniformIndexingNative
= false;
1500 properties
->robustBufferAccessUpdateAfterBind
= false;
1501 properties
->quadDivergentImplicitLod
= false;
1503 size_t max_descriptor_set_size
= radv_max_descriptor_set_size();
1504 properties
->maxPerStageDescriptorUpdateAfterBindSamplers
= max_descriptor_set_size
;
1505 properties
->maxPerStageDescriptorUpdateAfterBindUniformBuffers
= max_descriptor_set_size
;
1506 properties
->maxPerStageDescriptorUpdateAfterBindStorageBuffers
= max_descriptor_set_size
;
1507 properties
->maxPerStageDescriptorUpdateAfterBindSampledImages
= max_descriptor_set_size
;
1508 properties
->maxPerStageDescriptorUpdateAfterBindStorageImages
= max_descriptor_set_size
;
1509 properties
->maxPerStageDescriptorUpdateAfterBindInputAttachments
= max_descriptor_set_size
;
1510 properties
->maxPerStageUpdateAfterBindResources
= max_descriptor_set_size
;
1511 properties
->maxDescriptorSetUpdateAfterBindSamplers
= max_descriptor_set_size
;
1512 properties
->maxDescriptorSetUpdateAfterBindUniformBuffers
= max_descriptor_set_size
;
1513 properties
->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic
= MAX_DYNAMIC_UNIFORM_BUFFERS
;
1514 properties
->maxDescriptorSetUpdateAfterBindStorageBuffers
= max_descriptor_set_size
;
1515 properties
->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic
= MAX_DYNAMIC_STORAGE_BUFFERS
;
1516 properties
->maxDescriptorSetUpdateAfterBindSampledImages
= max_descriptor_set_size
;
1517 properties
->maxDescriptorSetUpdateAfterBindStorageImages
= max_descriptor_set_size
;
1518 properties
->maxDescriptorSetUpdateAfterBindInputAttachments
= max_descriptor_set_size
;
1521 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES
: {
1522 VkPhysicalDeviceProtectedMemoryProperties
*properties
=
1523 (VkPhysicalDeviceProtectedMemoryProperties
*)ext
;
1524 properties
->protectedNoFault
= false;
1527 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONSERVATIVE_RASTERIZATION_PROPERTIES_EXT
: {
1528 VkPhysicalDeviceConservativeRasterizationPropertiesEXT
*properties
=
1529 (VkPhysicalDeviceConservativeRasterizationPropertiesEXT
*)ext
;
1530 properties
->primitiveOverestimationSize
= 0;
1531 properties
->maxExtraPrimitiveOverestimationSize
= 0;
1532 properties
->extraPrimitiveOverestimationSizeGranularity
= 0;
1533 properties
->primitiveUnderestimation
= VK_FALSE
;
1534 properties
->conservativePointAndLineRasterization
= VK_FALSE
;
1535 properties
->degenerateTrianglesRasterized
= VK_FALSE
;
1536 properties
->degenerateLinesRasterized
= VK_FALSE
;
1537 properties
->fullyCoveredFragmentShaderInputVariable
= VK_FALSE
;
1538 properties
->conservativeRasterizationPostDepthCoverage
= VK_FALSE
;
1541 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT
: {
1542 VkPhysicalDevicePCIBusInfoPropertiesEXT
*properties
=
1543 (VkPhysicalDevicePCIBusInfoPropertiesEXT
*)ext
;
1544 properties
->pciDomain
= pdevice
->bus_info
.domain
;
1545 properties
->pciBus
= pdevice
->bus_info
.bus
;
1546 properties
->pciDevice
= pdevice
->bus_info
.dev
;
1547 properties
->pciFunction
= pdevice
->bus_info
.func
;
1550 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR
: {
1551 VkPhysicalDeviceDriverPropertiesKHR
*driver_props
=
1552 (VkPhysicalDeviceDriverPropertiesKHR
*) ext
;
1554 driver_props
->driverID
= VK_DRIVER_ID_MESA_RADV_KHR
;
1555 snprintf(driver_props
->driverName
, VK_MAX_DRIVER_NAME_SIZE_KHR
, "radv");
1556 snprintf(driver_props
->driverInfo
, VK_MAX_DRIVER_INFO_SIZE_KHR
,
1557 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1
1558 " (LLVM " MESA_LLVM_VERSION_STRING
")");
1560 driver_props
->conformanceVersion
= (VkConformanceVersionKHR
) {
1568 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT
: {
1569 VkPhysicalDeviceTransformFeedbackPropertiesEXT
*properties
=
1570 (VkPhysicalDeviceTransformFeedbackPropertiesEXT
*)ext
;
1571 properties
->maxTransformFeedbackStreams
= MAX_SO_STREAMS
;
1572 properties
->maxTransformFeedbackBuffers
= MAX_SO_BUFFERS
;
1573 properties
->maxTransformFeedbackBufferSize
= UINT32_MAX
;
1574 properties
->maxTransformFeedbackStreamDataSize
= 512;
1575 properties
->maxTransformFeedbackBufferDataSize
= UINT32_MAX
;
1576 properties
->maxTransformFeedbackBufferDataStride
= 512;
1577 properties
->transformFeedbackQueries
= !pdevice
->use_ngg_streamout
;
1578 properties
->transformFeedbackStreamsLinesTriangles
= !pdevice
->use_ngg_streamout
;
1579 properties
->transformFeedbackRasterizationStreamSelect
= false;
1580 properties
->transformFeedbackDraw
= true;
1583 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_PROPERTIES_EXT
: {
1584 VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*props
=
1585 (VkPhysicalDeviceInlineUniformBlockPropertiesEXT
*)ext
;
1587 props
->maxInlineUniformBlockSize
= MAX_INLINE_UNIFORM_BLOCK_SIZE
;
1588 props
->maxPerStageDescriptorInlineUniformBlocks
= MAX_INLINE_UNIFORM_BLOCK_SIZE
* MAX_SETS
;
1589 props
->maxPerStageDescriptorUpdateAfterBindInlineUniformBlocks
= MAX_INLINE_UNIFORM_BLOCK_SIZE
* MAX_SETS
;
1590 props
->maxDescriptorSetInlineUniformBlocks
= MAX_INLINE_UNIFORM_BLOCK_COUNT
;
1591 props
->maxDescriptorSetUpdateAfterBindInlineUniformBlocks
= MAX_INLINE_UNIFORM_BLOCK_COUNT
;
1594 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLE_LOCATIONS_PROPERTIES_EXT
: {
1595 VkPhysicalDeviceSampleLocationsPropertiesEXT
*properties
=
1596 (VkPhysicalDeviceSampleLocationsPropertiesEXT
*)ext
;
1597 properties
->sampleLocationSampleCounts
= VK_SAMPLE_COUNT_2_BIT
|
1598 VK_SAMPLE_COUNT_4_BIT
|
1599 VK_SAMPLE_COUNT_8_BIT
;
1600 properties
->maxSampleLocationGridSize
= (VkExtent2D
){ 2 , 2 };
1601 properties
->sampleLocationCoordinateRange
[0] = 0.0f
;
1602 properties
->sampleLocationCoordinateRange
[1] = 0.9375f
;
1603 properties
->sampleLocationSubPixelBits
= 4;
1604 properties
->variableSampleLocations
= VK_FALSE
;
1607 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_STENCIL_RESOLVE_PROPERTIES_KHR
: {
1608 VkPhysicalDeviceDepthStencilResolvePropertiesKHR
*properties
=
1609 (VkPhysicalDeviceDepthStencilResolvePropertiesKHR
*)ext
;
1611 /* We support all of the depth resolve modes */
1612 properties
->supportedDepthResolveModes
=
1613 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
|
1614 VK_RESOLVE_MODE_AVERAGE_BIT_KHR
|
1615 VK_RESOLVE_MODE_MIN_BIT_KHR
|
1616 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1618 /* Average doesn't make sense for stencil so we don't support that */
1619 properties
->supportedStencilResolveModes
=
1620 VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
|
1621 VK_RESOLVE_MODE_MIN_BIT_KHR
|
1622 VK_RESOLVE_MODE_MAX_BIT_KHR
;
1624 properties
->independentResolveNone
= VK_TRUE
;
1625 properties
->independentResolve
= VK_TRUE
;
1628 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_PROPERTIES_EXT
: {
1629 VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT
*properties
=
1630 (VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT
*)ext
;
1631 properties
->storageTexelBufferOffsetAlignmentBytes
= 4;
1632 properties
->storageTexelBufferOffsetSingleTexelAlignment
= true;
1633 properties
->uniformTexelBufferOffsetAlignmentBytes
= 4;
1634 properties
->uniformTexelBufferOffsetSingleTexelAlignment
= true;
1637 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT_CONTROLS_PROPERTIES_KHR
: {
1638 VkPhysicalDeviceFloatControlsPropertiesKHR
*properties
=
1639 (VkPhysicalDeviceFloatControlsPropertiesKHR
*)ext
;
1641 /* On AMD hardware, denormals and rounding modes for
1642 * fp16/fp64 are controlled by the same config
1645 properties
->denormBehaviorIndependence
= VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_32_BIT_ONLY_KHR
;
1646 properties
->roundingModeIndependence
= VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_32_BIT_ONLY_KHR
;
1648 /* Do not allow both preserving and flushing denorms
1649 * because different shaders in the same pipeline can
1650 * have different settings and this won't work for
1651 * merged shaders. To make it work, this requires LLVM
1652 * support for changing the register. The same logic
1653 * applies for the rounding modes because they are
1654 * configured with the same config register.
1655 * TODO: we can enable a lot of these for ACO when it
1656 * supports all stages
1658 properties
->shaderDenormFlushToZeroFloat32
= true;
1659 properties
->shaderDenormPreserveFloat32
= false;
1660 properties
->shaderRoundingModeRTEFloat32
= true;
1661 properties
->shaderRoundingModeRTZFloat32
= false;
1662 properties
->shaderSignedZeroInfNanPreserveFloat32
= true;
1664 properties
->shaderDenormFlushToZeroFloat16
= false;
1665 properties
->shaderDenormPreserveFloat16
= pdevice
->rad_info
.chip_class
>= GFX8
;
1666 properties
->shaderRoundingModeRTEFloat16
= pdevice
->rad_info
.chip_class
>= GFX8
;
1667 properties
->shaderRoundingModeRTZFloat16
= false;
1668 properties
->shaderSignedZeroInfNanPreserveFloat16
= pdevice
->rad_info
.chip_class
>= GFX8
;
1670 properties
->shaderDenormFlushToZeroFloat64
= false;
1671 properties
->shaderDenormPreserveFloat64
= pdevice
->rad_info
.chip_class
>= GFX8
;
1672 properties
->shaderRoundingModeRTEFloat64
= pdevice
->rad_info
.chip_class
>= GFX8
;
1673 properties
->shaderRoundingModeRTZFloat64
= false;
1674 properties
->shaderSignedZeroInfNanPreserveFloat64
= pdevice
->rad_info
.chip_class
>= GFX8
;
1677 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_PROPERTIES_KHR
: {
1678 VkPhysicalDeviceTimelineSemaphorePropertiesKHR
*props
=
1679 (VkPhysicalDeviceTimelineSemaphorePropertiesKHR
*) ext
;
1680 props
->maxTimelineSemaphoreValueDifference
= UINT64_MAX
;
1683 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_PROPERTIES_EXT
: {
1684 VkPhysicalDeviceSubgroupSizeControlPropertiesEXT
*props
=
1685 (VkPhysicalDeviceSubgroupSizeControlPropertiesEXT
*)ext
;
1686 props
->minSubgroupSize
= 64;
1687 props
->maxSubgroupSize
= 64;
1688 props
->maxComputeWorkgroupSubgroups
= UINT32_MAX
;
1689 props
->requiredSubgroupSizeStages
= 0;
1691 if (pdevice
->rad_info
.chip_class
>= GFX10
) {
1692 /* Only GFX10+ supports wave32. */
1693 props
->minSubgroupSize
= 32;
1694 props
->requiredSubgroupSizeStages
= VK_SHADER_STAGE_COMPUTE_BIT
;
1704 static void radv_get_physical_device_queue_family_properties(
1705 struct radv_physical_device
* pdevice
,
1707 VkQueueFamilyProperties
** pQueueFamilyProperties
)
1709 int num_queue_families
= 1;
1711 if (pdevice
->rad_info
.num_rings
[RING_COMPUTE
] > 0 &&
1712 !(pdevice
->instance
->debug_flags
& RADV_DEBUG_NO_COMPUTE_QUEUE
))
1713 num_queue_families
++;
1715 if (pQueueFamilyProperties
== NULL
) {
1716 *pCount
= num_queue_families
;
1725 *pQueueFamilyProperties
[idx
] = (VkQueueFamilyProperties
) {
1726 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
1727 VK_QUEUE_COMPUTE_BIT
|
1728 VK_QUEUE_TRANSFER_BIT
|
1729 VK_QUEUE_SPARSE_BINDING_BIT
,
1731 .timestampValidBits
= 64,
1732 .minImageTransferGranularity
= (VkExtent3D
) { 1, 1, 1 },
1737 if (pdevice
->rad_info
.num_rings
[RING_COMPUTE
] > 0 &&
1738 !(pdevice
->instance
->debug_flags
& RADV_DEBUG_NO_COMPUTE_QUEUE
)) {
1739 if (*pCount
> idx
) {
1740 *pQueueFamilyProperties
[idx
] = (VkQueueFamilyProperties
) {
1741 .queueFlags
= VK_QUEUE_COMPUTE_BIT
|
1742 VK_QUEUE_TRANSFER_BIT
|
1743 VK_QUEUE_SPARSE_BINDING_BIT
,
1744 .queueCount
= pdevice
->rad_info
.num_rings
[RING_COMPUTE
],
1745 .timestampValidBits
= 64,
1746 .minImageTransferGranularity
= (VkExtent3D
) { 1, 1, 1 },
1754 void radv_GetPhysicalDeviceQueueFamilyProperties(
1755 VkPhysicalDevice physicalDevice
,
1757 VkQueueFamilyProperties
* pQueueFamilyProperties
)
1759 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
1760 if (!pQueueFamilyProperties
) {
1761 radv_get_physical_device_queue_family_properties(pdevice
, pCount
, NULL
);
1764 VkQueueFamilyProperties
*properties
[] = {
1765 pQueueFamilyProperties
+ 0,
1766 pQueueFamilyProperties
+ 1,
1767 pQueueFamilyProperties
+ 2,
1769 radv_get_physical_device_queue_family_properties(pdevice
, pCount
, properties
);
1770 assert(*pCount
<= 3);
1773 void radv_GetPhysicalDeviceQueueFamilyProperties2(
1774 VkPhysicalDevice physicalDevice
,
1776 VkQueueFamilyProperties2
*pQueueFamilyProperties
)
1778 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
1779 if (!pQueueFamilyProperties
) {
1780 radv_get_physical_device_queue_family_properties(pdevice
, pCount
, NULL
);
1783 VkQueueFamilyProperties
*properties
[] = {
1784 &pQueueFamilyProperties
[0].queueFamilyProperties
,
1785 &pQueueFamilyProperties
[1].queueFamilyProperties
,
1786 &pQueueFamilyProperties
[2].queueFamilyProperties
,
1788 radv_get_physical_device_queue_family_properties(pdevice
, pCount
, properties
);
1789 assert(*pCount
<= 3);
1792 void radv_GetPhysicalDeviceMemoryProperties(
1793 VkPhysicalDevice physicalDevice
,
1794 VkPhysicalDeviceMemoryProperties
*pMemoryProperties
)
1796 RADV_FROM_HANDLE(radv_physical_device
, physical_device
, physicalDevice
);
1798 *pMemoryProperties
= physical_device
->memory_properties
;
1802 radv_get_memory_budget_properties(VkPhysicalDevice physicalDevice
,
1803 VkPhysicalDeviceMemoryBudgetPropertiesEXT
*memoryBudget
)
1805 RADV_FROM_HANDLE(radv_physical_device
, device
, physicalDevice
);
1806 VkPhysicalDeviceMemoryProperties
*memory_properties
= &device
->memory_properties
;
1807 uint64_t visible_vram_size
= radv_get_visible_vram_size(device
);
1808 uint64_t vram_size
= radv_get_vram_size(device
);
1809 uint64_t gtt_size
= device
->rad_info
.gart_size
;
1810 uint64_t heap_budget
, heap_usage
;
1812 /* For all memory heaps, the computation of budget is as follow:
1813 * heap_budget = heap_size - global_heap_usage + app_heap_usage
1815 * The Vulkan spec 1.1.97 says that the budget should include any
1816 * currently allocated device memory.
1818 * Note that the application heap usages are not really accurate (eg.
1819 * in presence of shared buffers).
1821 for (int i
= 0; i
< device
->memory_properties
.memoryTypeCount
; i
++) {
1822 uint32_t heap_index
= device
->memory_properties
.memoryTypes
[i
].heapIndex
;
1824 if (radv_is_mem_type_vram(device
->mem_type_indices
[i
])) {
1825 heap_usage
= device
->ws
->query_value(device
->ws
,
1826 RADEON_ALLOCATED_VRAM
);
1828 heap_budget
= vram_size
-
1829 device
->ws
->query_value(device
->ws
, RADEON_VRAM_USAGE
) +
1832 memoryBudget
->heapBudget
[heap_index
] = heap_budget
;
1833 memoryBudget
->heapUsage
[heap_index
] = heap_usage
;
1834 } else if (radv_is_mem_type_vram_visible(device
->mem_type_indices
[i
])) {
1835 heap_usage
= device
->ws
->query_value(device
->ws
,
1836 RADEON_ALLOCATED_VRAM_VIS
);
1838 heap_budget
= visible_vram_size
-
1839 device
->ws
->query_value(device
->ws
, RADEON_VRAM_VIS_USAGE
) +
1842 memoryBudget
->heapBudget
[heap_index
] = heap_budget
;
1843 memoryBudget
->heapUsage
[heap_index
] = heap_usage
;
1844 } else if (radv_is_mem_type_gtt_wc(device
->mem_type_indices
[i
])) {
1845 heap_usage
= device
->ws
->query_value(device
->ws
,
1846 RADEON_ALLOCATED_GTT
);
1848 heap_budget
= gtt_size
-
1849 device
->ws
->query_value(device
->ws
, RADEON_GTT_USAGE
) +
1852 memoryBudget
->heapBudget
[heap_index
] = heap_budget
;
1853 memoryBudget
->heapUsage
[heap_index
] = heap_usage
;
1857 /* The heapBudget and heapUsage values must be zero for array elements
1858 * greater than or equal to
1859 * VkPhysicalDeviceMemoryProperties::memoryHeapCount.
1861 for (uint32_t i
= memory_properties
->memoryHeapCount
; i
< VK_MAX_MEMORY_HEAPS
; i
++) {
1862 memoryBudget
->heapBudget
[i
] = 0;
1863 memoryBudget
->heapUsage
[i
] = 0;
1867 void radv_GetPhysicalDeviceMemoryProperties2(
1868 VkPhysicalDevice physicalDevice
,
1869 VkPhysicalDeviceMemoryProperties2
*pMemoryProperties
)
1871 radv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
1872 &pMemoryProperties
->memoryProperties
);
1874 VkPhysicalDeviceMemoryBudgetPropertiesEXT
*memory_budget
=
1875 vk_find_struct(pMemoryProperties
->pNext
,
1876 PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT
);
1878 radv_get_memory_budget_properties(physicalDevice
, memory_budget
);
1881 VkResult
radv_GetMemoryHostPointerPropertiesEXT(
1883 VkExternalMemoryHandleTypeFlagBits handleType
,
1884 const void *pHostPointer
,
1885 VkMemoryHostPointerPropertiesEXT
*pMemoryHostPointerProperties
)
1887 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1891 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
: {
1892 const struct radv_physical_device
*physical_device
= device
->physical_device
;
1893 uint32_t memoryTypeBits
= 0;
1894 for (int i
= 0; i
< physical_device
->memory_properties
.memoryTypeCount
; i
++) {
1895 if (radv_is_mem_type_gtt_cached(physical_device
->mem_type_indices
[i
])) {
1896 memoryTypeBits
= (1 << i
);
1900 pMemoryHostPointerProperties
->memoryTypeBits
= memoryTypeBits
;
1904 return VK_ERROR_INVALID_EXTERNAL_HANDLE
;
1908 static enum radeon_ctx_priority
1909 radv_get_queue_global_priority(const VkDeviceQueueGlobalPriorityCreateInfoEXT
*pObj
)
1911 /* Default to MEDIUM when a specific global priority isn't requested */
1913 return RADEON_CTX_PRIORITY_MEDIUM
;
1915 switch(pObj
->globalPriority
) {
1916 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT
:
1917 return RADEON_CTX_PRIORITY_REALTIME
;
1918 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT
:
1919 return RADEON_CTX_PRIORITY_HIGH
;
1920 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
:
1921 return RADEON_CTX_PRIORITY_MEDIUM
;
1922 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT
:
1923 return RADEON_CTX_PRIORITY_LOW
;
1925 unreachable("Illegal global priority value");
1926 return RADEON_CTX_PRIORITY_INVALID
;
1931 radv_queue_init(struct radv_device
*device
, struct radv_queue
*queue
,
1932 uint32_t queue_family_index
, int idx
,
1933 VkDeviceQueueCreateFlags flags
,
1934 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*global_priority
)
1936 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1937 queue
->device
= device
;
1938 queue
->queue_family_index
= queue_family_index
;
1939 queue
->queue_idx
= idx
;
1940 queue
->priority
= radv_get_queue_global_priority(global_priority
);
1941 queue
->flags
= flags
;
1943 queue
->hw_ctx
= device
->ws
->ctx_create(device
->ws
, queue
->priority
);
1945 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
1947 list_inithead(&queue
->pending_submissions
);
1948 pthread_mutex_init(&queue
->pending_mutex
, NULL
);
1954 radv_queue_finish(struct radv_queue
*queue
)
1956 pthread_mutex_destroy(&queue
->pending_mutex
);
1959 queue
->device
->ws
->ctx_destroy(queue
->hw_ctx
);
1961 if (queue
->initial_full_flush_preamble_cs
)
1962 queue
->device
->ws
->cs_destroy(queue
->initial_full_flush_preamble_cs
);
1963 if (queue
->initial_preamble_cs
)
1964 queue
->device
->ws
->cs_destroy(queue
->initial_preamble_cs
);
1965 if (queue
->continue_preamble_cs
)
1966 queue
->device
->ws
->cs_destroy(queue
->continue_preamble_cs
);
1967 if (queue
->descriptor_bo
)
1968 queue
->device
->ws
->buffer_destroy(queue
->descriptor_bo
);
1969 if (queue
->scratch_bo
)
1970 queue
->device
->ws
->buffer_destroy(queue
->scratch_bo
);
1971 if (queue
->esgs_ring_bo
)
1972 queue
->device
->ws
->buffer_destroy(queue
->esgs_ring_bo
);
1973 if (queue
->gsvs_ring_bo
)
1974 queue
->device
->ws
->buffer_destroy(queue
->gsvs_ring_bo
);
1975 if (queue
->tess_rings_bo
)
1976 queue
->device
->ws
->buffer_destroy(queue
->tess_rings_bo
);
1978 queue
->device
->ws
->buffer_destroy(queue
->gds_bo
);
1979 if (queue
->gds_oa_bo
)
1980 queue
->device
->ws
->buffer_destroy(queue
->gds_oa_bo
);
1981 if (queue
->compute_scratch_bo
)
1982 queue
->device
->ws
->buffer_destroy(queue
->compute_scratch_bo
);
1986 radv_bo_list_init(struct radv_bo_list
*bo_list
)
1988 pthread_mutex_init(&bo_list
->mutex
, NULL
);
1989 bo_list
->list
.count
= bo_list
->capacity
= 0;
1990 bo_list
->list
.bos
= NULL
;
1994 radv_bo_list_finish(struct radv_bo_list
*bo_list
)
1996 free(bo_list
->list
.bos
);
1997 pthread_mutex_destroy(&bo_list
->mutex
);
2000 static VkResult
radv_bo_list_add(struct radv_device
*device
,
2001 struct radeon_winsys_bo
*bo
)
2003 struct radv_bo_list
*bo_list
= &device
->bo_list
;
2008 if (unlikely(!device
->use_global_bo_list
))
2011 pthread_mutex_lock(&bo_list
->mutex
);
2012 if (bo_list
->list
.count
== bo_list
->capacity
) {
2013 unsigned capacity
= MAX2(4, bo_list
->capacity
* 2);
2014 void *data
= realloc(bo_list
->list
.bos
, capacity
* sizeof(struct radeon_winsys_bo
*));
2017 pthread_mutex_unlock(&bo_list
->mutex
);
2018 return VK_ERROR_OUT_OF_HOST_MEMORY
;
2021 bo_list
->list
.bos
= (struct radeon_winsys_bo
**)data
;
2022 bo_list
->capacity
= capacity
;
2025 bo_list
->list
.bos
[bo_list
->list
.count
++] = bo
;
2026 pthread_mutex_unlock(&bo_list
->mutex
);
2030 static void radv_bo_list_remove(struct radv_device
*device
,
2031 struct radeon_winsys_bo
*bo
)
2033 struct radv_bo_list
*bo_list
= &device
->bo_list
;
2038 if (unlikely(!device
->use_global_bo_list
))
2041 pthread_mutex_lock(&bo_list
->mutex
);
2042 for(unsigned i
= 0; i
< bo_list
->list
.count
; ++i
) {
2043 if (bo_list
->list
.bos
[i
] == bo
) {
2044 bo_list
->list
.bos
[i
] = bo_list
->list
.bos
[bo_list
->list
.count
- 1];
2045 --bo_list
->list
.count
;
2049 pthread_mutex_unlock(&bo_list
->mutex
);
2053 radv_device_init_gs_info(struct radv_device
*device
)
2055 device
->gs_table_depth
= ac_get_gs_table_depth(device
->physical_device
->rad_info
.chip_class
,
2056 device
->physical_device
->rad_info
.family
);
2059 static int radv_get_device_extension_index(const char *name
)
2061 for (unsigned i
= 0; i
< RADV_DEVICE_EXTENSION_COUNT
; ++i
) {
2062 if (strcmp(name
, radv_device_extensions
[i
].extensionName
) == 0)
2069 radv_get_int_debug_option(const char *name
, int default_value
)
2076 result
= default_value
;
2080 result
= strtol(str
, &endptr
, 0);
2081 if (str
== endptr
) {
2082 /* No digits founs. */
2083 result
= default_value
;
2090 static int install_seccomp_filter() {
2092 struct sock_filter filter
[] = {
2093 /* Check arch is 64bit x86 */
2094 BPF_STMT(BPF_LD
+ BPF_W
+ BPF_ABS
, (offsetof(struct seccomp_data
, arch
))),
2095 BPF_JUMP(BPF_JMP
+ BPF_JEQ
+ BPF_K
, AUDIT_ARCH_X86_64
, 0, 12),
2097 /* Futex is required for mutex locks */
2098 #if defined __NR__newselect
2099 BPF_STMT(BPF_LD
+ BPF_W
+ BPF_ABS
, (offsetof(struct seccomp_data
, nr
))),
2100 BPF_JUMP(BPF_JMP
+ BPF_JEQ
+ BPF_K
, __NR__newselect
, 11, 0),
2101 #elif defined __NR_select
2102 BPF_STMT(BPF_LD
+ BPF_W
+ BPF_ABS
, (offsetof(struct seccomp_data
, nr
))),
2103 BPF_JUMP(BPF_JMP
+ BPF_JEQ
+ BPF_K
, __NR_select
, 11, 0),
2105 BPF_STMT(BPF_LD
+ BPF_W
+ BPF_ABS
, (offsetof(struct seccomp_data
, nr
))),
2106 BPF_JUMP(BPF_JMP
+ BPF_JEQ
+ BPF_K
, __NR_pselect6
, 11, 0),
2109 /* Allow system exit calls for the forked process */
2110 BPF_STMT(BPF_LD
+ BPF_W
+ BPF_ABS
, (offsetof(struct seccomp_data
, nr
))),
2111 BPF_JUMP(BPF_JMP
+ BPF_JEQ
+ BPF_K
, __NR_exit_group
, 9, 0),
2113 /* Allow system read calls */
2114 BPF_STMT(BPF_LD
+ BPF_W
+ BPF_ABS
, (offsetof(struct seccomp_data
, nr
))),
2115 BPF_JUMP(BPF_JMP
+ BPF_JEQ
+ BPF_K
, __NR_read
, 7, 0),
2117 /* Allow system write calls */
2118 BPF_STMT(BPF_LD
+ BPF_W
+ BPF_ABS
, (offsetof(struct seccomp_data
, nr
))),
2119 BPF_JUMP(BPF_JMP
+ BPF_JEQ
+ BPF_K
, __NR_write
, 5, 0),
2121 /* Allow system brk calls (we need this for malloc) */
2122 BPF_STMT(BPF_LD
+ BPF_W
+ BPF_ABS
, (offsetof(struct seccomp_data
, nr
))),
2123 BPF_JUMP(BPF_JMP
+ BPF_JEQ
+ BPF_K
, __NR_brk
, 3, 0),
2125 /* Futex is required for mutex locks */
2126 BPF_STMT(BPF_LD
+ BPF_W
+ BPF_ABS
, (offsetof(struct seccomp_data
, nr
))),
2127 BPF_JUMP(BPF_JMP
+ BPF_JEQ
+ BPF_K
, __NR_futex
, 1, 0),
2129 /* Return error if we hit a system call not on the whitelist */
2130 BPF_STMT(BPF_RET
+ BPF_K
, SECCOMP_RET_ERRNO
| (EPERM
& SECCOMP_RET_DATA
)),
2132 /* Allow whitelisted system calls */
2133 BPF_STMT(BPF_RET
+ BPF_K
, SECCOMP_RET_ALLOW
),
2136 struct sock_fprog prog
= {
2137 .len
= (unsigned short)(sizeof(filter
) / sizeof(filter
[0])),
2141 if (prctl(PR_SET_NO_NEW_PRIVS
, 1, 0, 0, 0))
2144 if (prctl(PR_SET_SECCOMP
, SECCOMP_MODE_FILTER
, &prog
))
2150 /* Helper function with timeout support for reading from the pipe between
2151 * processes used for secure compile.
2153 bool radv_sc_read(int fd
, void *buf
, size_t size
, bool timeout
)
2162 /* We can't rely on the value of tv after calling select() so
2163 * we must reset it on each iteration of the loop.
2168 int rval
= select(fd
+ 1, &fds
, NULL
, NULL
, timeout
? &tv
: NULL
);
2174 ssize_t bytes_read
= read(fd
, buf
, size
);
2183 /* select timeout */
2189 static bool radv_close_all_fds(const int *keep_fds
, int keep_fd_count
)
2193 d
= opendir("/proc/self/fd");
2196 int dir_fd
= dirfd(d
);
2198 while ((dir
= readdir(d
)) != NULL
) {
2199 if (dir
->d_name
[0] == '.')
2202 int fd
= atoi(dir
->d_name
);
2207 for (int i
= 0; !keep
&& i
< keep_fd_count
; ++i
)
2208 if (keep_fds
[i
] == fd
)
2220 static bool secure_compile_open_fifo_fds(struct radv_secure_compile_state
*sc
,
2221 int *fd_server
, int *fd_client
,
2222 unsigned process
, bool make_fifo
)
2224 bool result
= false;
2225 char *fifo_server_path
= NULL
;
2226 char *fifo_client_path
= NULL
;
2228 if (asprintf(&fifo_server_path
, "/tmp/radv_server_%s_%u", sc
->uid
, process
) == -1)
2229 goto open_fifo_exit
;
2231 if (asprintf(&fifo_client_path
, "/tmp/radv_client_%s_%u", sc
->uid
, process
) == -1)
2232 goto open_fifo_exit
;
2235 int file1
= mkfifo(fifo_server_path
, 0666);
2237 goto open_fifo_exit
;
2239 int file2
= mkfifo(fifo_client_path
, 0666);
2241 goto open_fifo_exit
;
2244 *fd_server
= open(fifo_server_path
, O_RDWR
);
2246 goto open_fifo_exit
;
2248 *fd_client
= open(fifo_client_path
, O_RDWR
);
2249 if(*fd_client
< 1) {
2251 goto open_fifo_exit
;
2257 free(fifo_server_path
);
2258 free(fifo_client_path
);
2263 static void run_secure_compile_device(struct radv_device
*device
, unsigned process
,
2264 int fd_idle_device_output
)
2266 int fd_secure_input
;
2267 int fd_secure_output
;
2268 bool fifo_result
= secure_compile_open_fifo_fds(device
->sc_state
,
2273 enum radv_secure_compile_type sc_type
;
2275 const int needed_fds
[] = {
2278 fd_idle_device_output
,
2281 if (!fifo_result
|| !radv_close_all_fds(needed_fds
, ARRAY_SIZE(needed_fds
)) ||
2282 install_seccomp_filter() == -1) {
2283 sc_type
= RADV_SC_TYPE_INIT_FAILURE
;
2285 sc_type
= RADV_SC_TYPE_INIT_SUCCESS
;
2286 device
->sc_state
->secure_compile_processes
[process
].fd_secure_input
= fd_secure_input
;
2287 device
->sc_state
->secure_compile_processes
[process
].fd_secure_output
= fd_secure_output
;
2290 write(fd_idle_device_output
, &sc_type
, sizeof(sc_type
));
2292 if (sc_type
== RADV_SC_TYPE_INIT_FAILURE
)
2293 goto secure_compile_exit
;
2296 radv_sc_read(fd_secure_input
, &sc_type
, sizeof(sc_type
), false);
2298 if (sc_type
== RADV_SC_TYPE_COMPILE_PIPELINE
) {
2299 struct radv_pipeline
*pipeline
;
2300 bool sc_read
= true;
2302 pipeline
= vk_zalloc2(&device
->alloc
, NULL
, sizeof(*pipeline
), 8,
2303 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2305 pipeline
->device
= device
;
2307 /* Read pipeline layout */
2308 struct radv_pipeline_layout layout
;
2309 sc_read
= radv_sc_read(fd_secure_input
, &layout
, sizeof(struct radv_pipeline_layout
), true);
2310 sc_read
&= radv_sc_read(fd_secure_input
, &layout
.num_sets
, sizeof(uint32_t), true);
2312 goto secure_compile_exit
;
2314 for (uint32_t set
= 0; set
< layout
.num_sets
; set
++) {
2315 uint32_t layout_size
;
2316 sc_read
&= radv_sc_read(fd_secure_input
, &layout_size
, sizeof(uint32_t), true);
2318 goto secure_compile_exit
;
2320 layout
.set
[set
].layout
= malloc(layout_size
);
2321 layout
.set
[set
].layout
->layout_size
= layout_size
;
2322 sc_read
&= radv_sc_read(fd_secure_input
, layout
.set
[set
].layout
,
2323 layout
.set
[set
].layout
->layout_size
, true);
2326 pipeline
->layout
= &layout
;
2328 /* Read pipeline key */
2329 struct radv_pipeline_key key
;
2330 sc_read
&= radv_sc_read(fd_secure_input
, &key
, sizeof(struct radv_pipeline_key
), true);
2332 /* Read pipeline create flags */
2333 VkPipelineCreateFlags flags
;
2334 sc_read
&= radv_sc_read(fd_secure_input
, &flags
, sizeof(VkPipelineCreateFlags
), true);
2336 /* Read stage and shader information */
2337 uint32_t num_stages
;
2338 const VkPipelineShaderStageCreateInfo
*pStages
[MESA_SHADER_STAGES
] = { 0, };
2339 sc_read
&= radv_sc_read(fd_secure_input
, &num_stages
, sizeof(uint32_t), true);
2341 goto secure_compile_exit
;
2343 for (uint32_t i
= 0; i
< num_stages
; i
++) {
2346 gl_shader_stage stage
;
2347 sc_read
&= radv_sc_read(fd_secure_input
, &stage
, sizeof(gl_shader_stage
), true);
2349 VkPipelineShaderStageCreateInfo
*pStage
= calloc(1, sizeof(VkPipelineShaderStageCreateInfo
));
2351 /* Read entry point name */
2353 sc_read
&= radv_sc_read(fd_secure_input
, &name_size
, sizeof(size_t), true);
2355 goto secure_compile_exit
;
2357 char *ep_name
= malloc(name_size
);
2358 sc_read
&= radv_sc_read(fd_secure_input
, ep_name
, name_size
, true);
2359 pStage
->pName
= ep_name
;
2361 /* Read shader module */
2363 sc_read
&= radv_sc_read(fd_secure_input
, &module_size
, sizeof(size_t), true);
2365 goto secure_compile_exit
;
2367 struct radv_shader_module
*module
= malloc(module_size
);
2368 sc_read
&= radv_sc_read(fd_secure_input
, module
, module_size
, true);
2369 pStage
->module
= radv_shader_module_to_handle(module
);
2371 /* Read specialization info */
2373 sc_read
&= radv_sc_read(fd_secure_input
, &has_spec_info
, sizeof(bool), true);
2375 goto secure_compile_exit
;
2377 if (has_spec_info
) {
2378 VkSpecializationInfo
*specInfo
= malloc(sizeof(VkSpecializationInfo
));
2379 pStage
->pSpecializationInfo
= specInfo
;
2381 sc_read
&= radv_sc_read(fd_secure_input
, &specInfo
->dataSize
, sizeof(size_t), true);
2383 goto secure_compile_exit
;
2385 void *si_data
= malloc(specInfo
->dataSize
);
2386 sc_read
&= radv_sc_read(fd_secure_input
, si_data
, specInfo
->dataSize
, true);
2387 specInfo
->pData
= si_data
;
2389 sc_read
&= radv_sc_read(fd_secure_input
, &specInfo
->mapEntryCount
, sizeof(uint32_t), true);
2391 goto secure_compile_exit
;
2393 VkSpecializationMapEntry
*mapEntries
= malloc(sizeof(VkSpecializationMapEntry
) * specInfo
->mapEntryCount
);
2394 for (uint32_t j
= 0; j
< specInfo
->mapEntryCount
; j
++) {
2395 sc_read
&= radv_sc_read(fd_secure_input
, &mapEntries
[j
], sizeof(VkSpecializationMapEntry
), true);
2397 goto secure_compile_exit
;
2400 specInfo
->pMapEntries
= mapEntries
;
2403 pStages
[stage
] = pStage
;
2406 /* Compile the shaders */
2407 VkPipelineCreationFeedbackEXT
*stage_feedbacks
[MESA_SHADER_STAGES
] = { 0 };
2408 radv_create_shaders(pipeline
, device
, NULL
, &key
, pStages
, flags
, NULL
, stage_feedbacks
);
2410 /* free memory allocated above */
2411 for (uint32_t set
= 0; set
< layout
.num_sets
; set
++)
2412 free(layout
.set
[set
].layout
);
2414 for (uint32_t i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
2418 free((void *) pStages
[i
]->pName
);
2419 free(radv_shader_module_from_handle(pStages
[i
]->module
));
2420 if (pStages
[i
]->pSpecializationInfo
) {
2421 free((void *) pStages
[i
]->pSpecializationInfo
->pData
);
2422 free((void *) pStages
[i
]->pSpecializationInfo
->pMapEntries
);
2423 free((void *) pStages
[i
]->pSpecializationInfo
);
2425 free((void *) pStages
[i
]);
2428 vk_free(&device
->alloc
, pipeline
);
2430 sc_type
= RADV_SC_TYPE_COMPILE_PIPELINE_FINISHED
;
2431 write(fd_secure_output
, &sc_type
, sizeof(sc_type
));
2433 } else if (sc_type
== RADV_SC_TYPE_DESTROY_DEVICE
) {
2434 goto secure_compile_exit
;
2438 secure_compile_exit
:
2439 close(fd_secure_input
);
2440 close(fd_secure_output
);
2441 close(fd_idle_device_output
);
2445 static enum radv_secure_compile_type
fork_secure_compile_device(struct radv_device
*device
, unsigned process
)
2447 int fd_secure_input
[2];
2448 int fd_secure_output
[2];
2450 /* create pipe descriptors (used to communicate between processes) */
2451 if (pipe(fd_secure_input
) == -1 || pipe(fd_secure_output
) == -1)
2452 return RADV_SC_TYPE_INIT_FAILURE
;
2456 if ((sc_pid
= fork()) == 0) {
2457 device
->sc_state
->secure_compile_thread_counter
= process
;
2458 run_secure_compile_device(device
, process
, fd_secure_output
[1]);
2461 return RADV_SC_TYPE_INIT_FAILURE
;
2463 /* Read the init result returned from the secure process */
2464 enum radv_secure_compile_type sc_type
;
2465 bool sc_read
= radv_sc_read(fd_secure_output
[0], &sc_type
, sizeof(sc_type
), true);
2467 if (sc_type
== RADV_SC_TYPE_INIT_FAILURE
|| !sc_read
) {
2468 close(fd_secure_input
[0]);
2469 close(fd_secure_input
[1]);
2470 close(fd_secure_output
[1]);
2471 close(fd_secure_output
[0]);
2473 waitpid(sc_pid
, &status
, 0);
2475 return RADV_SC_TYPE_INIT_FAILURE
;
2477 assert(sc_type
== RADV_SC_TYPE_INIT_SUCCESS
);
2478 write(device
->sc_state
->secure_compile_processes
[process
].fd_secure_output
, &sc_type
, sizeof(sc_type
));
2480 close(fd_secure_input
[0]);
2481 close(fd_secure_input
[1]);
2482 close(fd_secure_output
[1]);
2483 close(fd_secure_output
[0]);
2486 waitpid(sc_pid
, &status
, 0);
2490 return RADV_SC_TYPE_INIT_SUCCESS
;
2493 /* Run a bare bones fork of a device that was forked right after its creation.
2494 * This device will have low overhead when it is forked again before each
2495 * pipeline compilation. This device sits idle and its only job is to fork
2498 static void run_secure_compile_idle_device(struct radv_device
*device
, unsigned process
,
2499 int fd_secure_input
, int fd_secure_output
)
2501 enum radv_secure_compile_type sc_type
= RADV_SC_TYPE_INIT_SUCCESS
;
2502 device
->sc_state
->secure_compile_processes
[process
].fd_secure_input
= fd_secure_input
;
2503 device
->sc_state
->secure_compile_processes
[process
].fd_secure_output
= fd_secure_output
;
2505 write(fd_secure_output
, &sc_type
, sizeof(sc_type
));
2508 radv_sc_read(fd_secure_input
, &sc_type
, sizeof(sc_type
), false);
2510 if (sc_type
== RADV_SC_TYPE_FORK_DEVICE
) {
2511 sc_type
= fork_secure_compile_device(device
, process
);
2513 if (sc_type
== RADV_SC_TYPE_INIT_FAILURE
)
2514 goto secure_compile_exit
;
2516 } else if (sc_type
== RADV_SC_TYPE_DESTROY_DEVICE
) {
2517 goto secure_compile_exit
;
2521 secure_compile_exit
:
2522 close(fd_secure_input
);
2523 close(fd_secure_output
);
2527 static void destroy_secure_compile_device(struct radv_device
*device
, unsigned process
)
2529 int fd_secure_input
= device
->sc_state
->secure_compile_processes
[process
].fd_secure_input
;
2531 enum radv_secure_compile_type sc_type
= RADV_SC_TYPE_DESTROY_DEVICE
;
2532 write(fd_secure_input
, &sc_type
, sizeof(sc_type
));
2534 close(device
->sc_state
->secure_compile_processes
[process
].fd_secure_input
);
2535 close(device
->sc_state
->secure_compile_processes
[process
].fd_secure_output
);
2538 waitpid(device
->sc_state
->secure_compile_processes
[process
].sc_pid
, &status
, 0);
2541 static VkResult
fork_secure_compile_idle_device(struct radv_device
*device
)
2543 device
->sc_state
= vk_zalloc(&device
->alloc
,
2544 sizeof(struct radv_secure_compile_state
),
2545 8, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
2547 mtx_init(&device
->sc_state
->secure_compile_mutex
, mtx_plain
);
2549 pid_t upid
= getpid();
2550 time_t seconds
= time(NULL
);
2553 if (asprintf(&uid
, "%ld_%ld", (long) upid
, (long) seconds
) == -1)
2554 return VK_ERROR_INITIALIZATION_FAILED
;
2556 device
->sc_state
->uid
= uid
;
2558 uint8_t sc_threads
= device
->instance
->num_sc_threads
;
2559 int fd_secure_input
[MAX_SC_PROCS
][2];
2560 int fd_secure_output
[MAX_SC_PROCS
][2];
2562 /* create pipe descriptors (used to communicate between processes) */
2563 for (unsigned i
= 0; i
< sc_threads
; i
++) {
2564 if (pipe(fd_secure_input
[i
]) == -1 ||
2565 pipe(fd_secure_output
[i
]) == -1) {
2566 return VK_ERROR_INITIALIZATION_FAILED
;
2570 device
->sc_state
->secure_compile_processes
= vk_zalloc(&device
->alloc
,
2571 sizeof(struct radv_secure_compile_process
) * sc_threads
, 8,
2572 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
2574 for (unsigned process
= 0; process
< sc_threads
; process
++) {
2575 if ((device
->sc_state
->secure_compile_processes
[process
].sc_pid
= fork()) == 0) {
2576 device
->sc_state
->secure_compile_thread_counter
= process
;
2577 run_secure_compile_idle_device(device
, process
, fd_secure_input
[process
][0], fd_secure_output
[process
][1]);
2579 if (device
->sc_state
->secure_compile_processes
[process
].sc_pid
== -1)
2580 return VK_ERROR_INITIALIZATION_FAILED
;
2582 /* Read the init result returned from the secure process */
2583 enum radv_secure_compile_type sc_type
;
2584 bool sc_read
= radv_sc_read(fd_secure_output
[process
][0], &sc_type
, sizeof(sc_type
), true);
2587 if (sc_read
&& sc_type
== RADV_SC_TYPE_INIT_SUCCESS
) {
2588 fifo_result
= secure_compile_open_fifo_fds(device
->sc_state
,
2589 &device
->sc_state
->secure_compile_processes
[process
].fd_server
,
2590 &device
->sc_state
->secure_compile_processes
[process
].fd_client
,
2593 device
->sc_state
->secure_compile_processes
[process
].fd_secure_input
= fd_secure_input
[process
][1];
2594 device
->sc_state
->secure_compile_processes
[process
].fd_secure_output
= fd_secure_output
[process
][0];
2597 if (sc_type
== RADV_SC_TYPE_INIT_FAILURE
|| !sc_read
|| !fifo_result
) {
2598 close(fd_secure_input
[process
][0]);
2599 close(fd_secure_input
[process
][1]);
2600 close(fd_secure_output
[process
][1]);
2601 close(fd_secure_output
[process
][0]);
2603 waitpid(device
->sc_state
->secure_compile_processes
[process
].sc_pid
, &status
, 0);
2605 /* Destroy any forks that were created sucessfully */
2606 for (unsigned i
= 0; i
< process
; i
++) {
2607 destroy_secure_compile_device(device
, i
);
2610 return VK_ERROR_INITIALIZATION_FAILED
;
2618 radv_create_pthread_cond(pthread_cond_t
*cond
)
2620 pthread_condattr_t condattr
;
2621 if (pthread_condattr_init(&condattr
)) {
2622 return VK_ERROR_INITIALIZATION_FAILED
;
2625 if (pthread_condattr_setclock(&condattr
, CLOCK_MONOTONIC
)) {
2626 pthread_condattr_destroy(&condattr
);
2627 return VK_ERROR_INITIALIZATION_FAILED
;
2629 if (pthread_cond_init(cond
, &condattr
)) {
2630 pthread_condattr_destroy(&condattr
);
2631 return VK_ERROR_INITIALIZATION_FAILED
;
2633 pthread_condattr_destroy(&condattr
);
2637 VkResult
radv_CreateDevice(
2638 VkPhysicalDevice physicalDevice
,
2639 const VkDeviceCreateInfo
* pCreateInfo
,
2640 const VkAllocationCallbacks
* pAllocator
,
2643 RADV_FROM_HANDLE(radv_physical_device
, physical_device
, physicalDevice
);
2645 struct radv_device
*device
;
2647 bool keep_shader_info
= false;
2649 /* Check enabled features */
2650 if (pCreateInfo
->pEnabledFeatures
) {
2651 VkPhysicalDeviceFeatures supported_features
;
2652 radv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
2653 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
2654 VkBool32
*enabled_feature
= (VkBool32
*)pCreateInfo
->pEnabledFeatures
;
2655 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
2656 for (uint32_t i
= 0; i
< num_features
; i
++) {
2657 if (enabled_feature
[i
] && !supported_feature
[i
])
2658 return vk_error(physical_device
->instance
, VK_ERROR_FEATURE_NOT_PRESENT
);
2662 device
= vk_zalloc2(&physical_device
->instance
->alloc
, pAllocator
,
2664 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
2666 return vk_error(physical_device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
2668 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
2669 device
->instance
= physical_device
->instance
;
2670 device
->physical_device
= physical_device
;
2672 device
->ws
= physical_device
->ws
;
2674 device
->alloc
= *pAllocator
;
2676 device
->alloc
= physical_device
->instance
->alloc
;
2678 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
2679 const char *ext_name
= pCreateInfo
->ppEnabledExtensionNames
[i
];
2680 int index
= radv_get_device_extension_index(ext_name
);
2681 if (index
< 0 || !physical_device
->supported_extensions
.extensions
[index
]) {
2682 vk_free(&device
->alloc
, device
);
2683 return vk_error(physical_device
->instance
, VK_ERROR_EXTENSION_NOT_PRESENT
);
2686 device
->enabled_extensions
.extensions
[index
] = true;
2689 keep_shader_info
= device
->enabled_extensions
.AMD_shader_info
;
2691 /* With update after bind we can't attach bo's to the command buffer
2692 * from the descriptor set anymore, so we have to use a global BO list.
2694 device
->use_global_bo_list
=
2695 (device
->instance
->perftest_flags
& RADV_PERFTEST_BO_LIST
) ||
2696 device
->enabled_extensions
.EXT_descriptor_indexing
||
2697 device
->enabled_extensions
.EXT_buffer_device_address
||
2698 device
->enabled_extensions
.KHR_buffer_device_address
;
2700 device
->robust_buffer_access
= pCreateInfo
->pEnabledFeatures
&&
2701 pCreateInfo
->pEnabledFeatures
->robustBufferAccess
;
2703 mtx_init(&device
->shader_slab_mutex
, mtx_plain
);
2704 list_inithead(&device
->shader_slabs
);
2706 radv_bo_list_init(&device
->bo_list
);
2708 for (unsigned i
= 0; i
< pCreateInfo
->queueCreateInfoCount
; i
++) {
2709 const VkDeviceQueueCreateInfo
*queue_create
= &pCreateInfo
->pQueueCreateInfos
[i
];
2710 uint32_t qfi
= queue_create
->queueFamilyIndex
;
2711 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*global_priority
=
2712 vk_find_struct_const(queue_create
->pNext
, DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT
);
2714 assert(!global_priority
|| device
->physical_device
->rad_info
.has_ctx_priority
);
2716 device
->queues
[qfi
] = vk_alloc(&device
->alloc
,
2717 queue_create
->queueCount
* sizeof(struct radv_queue
), 8, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
2718 if (!device
->queues
[qfi
]) {
2719 result
= VK_ERROR_OUT_OF_HOST_MEMORY
;
2723 memset(device
->queues
[qfi
], 0, queue_create
->queueCount
* sizeof(struct radv_queue
));
2725 device
->queue_count
[qfi
] = queue_create
->queueCount
;
2727 for (unsigned q
= 0; q
< queue_create
->queueCount
; q
++) {
2728 result
= radv_queue_init(device
, &device
->queues
[qfi
][q
],
2729 qfi
, q
, queue_create
->flags
,
2731 if (result
!= VK_SUCCESS
)
2736 device
->pbb_allowed
= device
->physical_device
->rad_info
.chip_class
>= GFX9
&&
2737 !(device
->instance
->debug_flags
& RADV_DEBUG_NOBINNING
);
2739 /* Disable DFSM by default. As of 2019-09-15 Talos on Low is still 3% slower on Raven. */
2740 device
->dfsm_allowed
= device
->pbb_allowed
&&
2741 (device
->instance
->perftest_flags
& RADV_PERFTEST_DFSM
);
2743 device
->always_use_syncobj
= device
->physical_device
->rad_info
.has_syncobj_wait_for_submit
;
2745 /* The maximum number of scratch waves. Scratch space isn't divided
2746 * evenly between CUs. The number is only a function of the number of CUs.
2747 * We can decrease the constant to decrease the scratch buffer size.
2749 * sctx->scratch_waves must be >= the maximum possible size of
2750 * 1 threadgroup, so that the hw doesn't hang from being unable
2753 * The recommended value is 4 per CU at most. Higher numbers don't
2754 * bring much benefit, but they still occupy chip resources (think
2755 * async compute). I've seen ~2% performance difference between 4 and 32.
2757 uint32_t max_threads_per_block
= 2048;
2758 device
->scratch_waves
= MAX2(32 * physical_device
->rad_info
.num_good_compute_units
,
2759 max_threads_per_block
/ 64);
2761 device
->dispatch_initiator
= S_00B800_COMPUTE_SHADER_EN(1);
2763 if (device
->physical_device
->rad_info
.chip_class
>= GFX7
) {
2764 /* If the KMD allows it (there is a KMD hw register for it),
2765 * allow launching waves out-of-order.
2767 device
->dispatch_initiator
|= S_00B800_ORDER_MODE(1);
2770 radv_device_init_gs_info(device
);
2772 device
->tess_offchip_block_dw_size
=
2773 device
->physical_device
->rad_info
.family
== CHIP_HAWAII
? 4096 : 8192;
2775 if (getenv("RADV_TRACE_FILE")) {
2776 const char *filename
= getenv("RADV_TRACE_FILE");
2778 keep_shader_info
= true;
2780 if (!radv_init_trace(device
))
2783 fprintf(stderr
, "*****************************************************************************\n");
2784 fprintf(stderr
, "* WARNING: RADV_TRACE_FILE is costly and should only be used for debugging! *\n");
2785 fprintf(stderr
, "*****************************************************************************\n");
2787 fprintf(stderr
, "Trace file will be dumped to %s\n", filename
);
2788 radv_dump_enabled_options(device
, stderr
);
2791 /* Temporarily disable secure compile while we create meta shaders, etc */
2792 uint8_t sc_threads
= device
->instance
->num_sc_threads
;
2794 device
->instance
->num_sc_threads
= 0;
2796 device
->keep_shader_info
= keep_shader_info
;
2797 result
= radv_device_init_meta(device
);
2798 if (result
!= VK_SUCCESS
)
2801 radv_device_init_msaa(device
);
2803 for (int family
= 0; family
< RADV_MAX_QUEUE_FAMILIES
; ++family
) {
2804 device
->empty_cs
[family
] = device
->ws
->cs_create(device
->ws
, family
);
2806 case RADV_QUEUE_GENERAL
:
2807 radeon_emit(device
->empty_cs
[family
], PKT3(PKT3_CONTEXT_CONTROL
, 1, 0));
2808 radeon_emit(device
->empty_cs
[family
], CONTEXT_CONTROL_LOAD_ENABLE(1));
2809 radeon_emit(device
->empty_cs
[family
], CONTEXT_CONTROL_SHADOW_ENABLE(1));
2811 case RADV_QUEUE_COMPUTE
:
2812 radeon_emit(device
->empty_cs
[family
], PKT3(PKT3_NOP
, 0, 0));
2813 radeon_emit(device
->empty_cs
[family
], 0);
2816 device
->ws
->cs_finalize(device
->empty_cs
[family
]);
2819 if (device
->physical_device
->rad_info
.chip_class
>= GFX7
)
2820 cik_create_gfx_config(device
);
2822 VkPipelineCacheCreateInfo ci
;
2823 ci
.sType
= VK_STRUCTURE_TYPE_PIPELINE_CACHE_CREATE_INFO
;
2826 ci
.pInitialData
= NULL
;
2827 ci
.initialDataSize
= 0;
2829 result
= radv_CreatePipelineCache(radv_device_to_handle(device
),
2831 if (result
!= VK_SUCCESS
)
2834 device
->mem_cache
= radv_pipeline_cache_from_handle(pc
);
2836 result
= radv_create_pthread_cond(&device
->timeline_cond
);
2837 if (result
!= VK_SUCCESS
)
2838 goto fail_mem_cache
;
2840 device
->force_aniso
=
2841 MIN2(16, radv_get_int_debug_option("RADV_TEX_ANISO", -1));
2842 if (device
->force_aniso
>= 0) {
2843 fprintf(stderr
, "radv: Forcing anisotropy filter to %ix\n",
2844 1 << util_logbase2(device
->force_aniso
));
2847 /* Fork device for secure compile as required */
2848 device
->instance
->num_sc_threads
= sc_threads
;
2849 if (radv_device_use_secure_compile(device
->instance
)) {
2851 result
= fork_secure_compile_idle_device(device
);
2852 if (result
!= VK_SUCCESS
)
2856 *pDevice
= radv_device_to_handle(device
);
2860 radv_DestroyPipelineCache(radv_device_to_handle(device
), pc
, NULL
);
2862 radv_device_finish_meta(device
);
2864 radv_bo_list_finish(&device
->bo_list
);
2866 if (device
->trace_bo
)
2867 device
->ws
->buffer_destroy(device
->trace_bo
);
2869 if (device
->gfx_init
)
2870 device
->ws
->buffer_destroy(device
->gfx_init
);
2872 for (unsigned i
= 0; i
< RADV_MAX_QUEUE_FAMILIES
; i
++) {
2873 for (unsigned q
= 0; q
< device
->queue_count
[i
]; q
++)
2874 radv_queue_finish(&device
->queues
[i
][q
]);
2875 if (device
->queue_count
[i
])
2876 vk_free(&device
->alloc
, device
->queues
[i
]);
2879 vk_free(&device
->alloc
, device
);
2883 void radv_DestroyDevice(
2885 const VkAllocationCallbacks
* pAllocator
)
2887 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2892 if (device
->trace_bo
)
2893 device
->ws
->buffer_destroy(device
->trace_bo
);
2895 if (device
->gfx_init
)
2896 device
->ws
->buffer_destroy(device
->gfx_init
);
2898 for (unsigned i
= 0; i
< RADV_MAX_QUEUE_FAMILIES
; i
++) {
2899 for (unsigned q
= 0; q
< device
->queue_count
[i
]; q
++)
2900 radv_queue_finish(&device
->queues
[i
][q
]);
2901 if (device
->queue_count
[i
])
2902 vk_free(&device
->alloc
, device
->queues
[i
]);
2903 if (device
->empty_cs
[i
])
2904 device
->ws
->cs_destroy(device
->empty_cs
[i
]);
2906 radv_device_finish_meta(device
);
2908 VkPipelineCache pc
= radv_pipeline_cache_to_handle(device
->mem_cache
);
2909 radv_DestroyPipelineCache(radv_device_to_handle(device
), pc
, NULL
);
2911 radv_destroy_shader_slabs(device
);
2913 pthread_cond_destroy(&device
->timeline_cond
);
2914 radv_bo_list_finish(&device
->bo_list
);
2915 if (radv_device_use_secure_compile(device
->instance
)) {
2916 for (unsigned i
= 0; i
< device
->instance
->num_sc_threads
; i
++ ) {
2917 destroy_secure_compile_device(device
, i
);
2921 if (device
->sc_state
) {
2922 free(device
->sc_state
->uid
);
2923 vk_free(&device
->alloc
, device
->sc_state
->secure_compile_processes
);
2925 vk_free(&device
->alloc
, device
->sc_state
);
2926 vk_free(&device
->alloc
, device
);
2929 VkResult
radv_EnumerateInstanceLayerProperties(
2930 uint32_t* pPropertyCount
,
2931 VkLayerProperties
* pProperties
)
2933 if (pProperties
== NULL
) {
2934 *pPropertyCount
= 0;
2938 /* None supported at this time */
2939 return vk_error(NULL
, VK_ERROR_LAYER_NOT_PRESENT
);
2942 VkResult
radv_EnumerateDeviceLayerProperties(
2943 VkPhysicalDevice physicalDevice
,
2944 uint32_t* pPropertyCount
,
2945 VkLayerProperties
* pProperties
)
2947 if (pProperties
== NULL
) {
2948 *pPropertyCount
= 0;
2952 /* None supported at this time */
2953 return vk_error(NULL
, VK_ERROR_LAYER_NOT_PRESENT
);
2956 void radv_GetDeviceQueue2(
2958 const VkDeviceQueueInfo2
* pQueueInfo
,
2961 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2962 struct radv_queue
*queue
;
2964 queue
= &device
->queues
[pQueueInfo
->queueFamilyIndex
][pQueueInfo
->queueIndex
];
2965 if (pQueueInfo
->flags
!= queue
->flags
) {
2966 /* From the Vulkan 1.1.70 spec:
2968 * "The queue returned by vkGetDeviceQueue2 must have the same
2969 * flags value from this structure as that used at device
2970 * creation time in a VkDeviceQueueCreateInfo instance. If no
2971 * matching flags were specified at device creation time then
2972 * pQueue will return VK_NULL_HANDLE."
2974 *pQueue
= VK_NULL_HANDLE
;
2978 *pQueue
= radv_queue_to_handle(queue
);
2981 void radv_GetDeviceQueue(
2983 uint32_t queueFamilyIndex
,
2984 uint32_t queueIndex
,
2987 const VkDeviceQueueInfo2 info
= (VkDeviceQueueInfo2
) {
2988 .sType
= VK_STRUCTURE_TYPE_DEVICE_QUEUE_INFO_2
,
2989 .queueFamilyIndex
= queueFamilyIndex
,
2990 .queueIndex
= queueIndex
2993 radv_GetDeviceQueue2(_device
, &info
, pQueue
);
2997 fill_geom_tess_rings(struct radv_queue
*queue
,
2999 bool add_sample_positions
,
3000 uint32_t esgs_ring_size
,
3001 struct radeon_winsys_bo
*esgs_ring_bo
,
3002 uint32_t gsvs_ring_size
,
3003 struct radeon_winsys_bo
*gsvs_ring_bo
,
3004 uint32_t tess_factor_ring_size
,
3005 uint32_t tess_offchip_ring_offset
,
3006 uint32_t tess_offchip_ring_size
,
3007 struct radeon_winsys_bo
*tess_rings_bo
)
3009 uint32_t *desc
= &map
[4];
3012 uint64_t esgs_va
= radv_buffer_get_va(esgs_ring_bo
);
3014 /* stride 0, num records - size, add tid, swizzle, elsize4,
3017 desc
[1] = S_008F04_BASE_ADDRESS_HI(esgs_va
>> 32) |
3018 S_008F04_SWIZZLE_ENABLE(true);
3019 desc
[2] = esgs_ring_size
;
3020 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
3021 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
3022 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
3023 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
3024 S_008F0C_INDEX_STRIDE(3) |
3025 S_008F0C_ADD_TID_ENABLE(1);
3027 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
3028 desc
[3] |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT
) |
3029 S_008F0C_OOB_SELECT(2) |
3030 S_008F0C_RESOURCE_LEVEL(1);
3032 desc
[3] |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
3033 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
3034 S_008F0C_ELEMENT_SIZE(1);
3037 /* GS entry for ES->GS ring */
3038 /* stride 0, num records - size, elsize0,
3041 desc
[5] = S_008F04_BASE_ADDRESS_HI(esgs_va
>> 32);
3042 desc
[6] = esgs_ring_size
;
3043 desc
[7] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
3044 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
3045 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
3046 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
);
3048 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
3049 desc
[7] |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT
) |
3050 S_008F0C_OOB_SELECT(2) |
3051 S_008F0C_RESOURCE_LEVEL(1);
3053 desc
[7] |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
3054 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
);
3061 uint64_t gsvs_va
= radv_buffer_get_va(gsvs_ring_bo
);
3063 /* VS entry for GS->VS ring */
3064 /* stride 0, num records - size, elsize0,
3067 desc
[1] = S_008F04_BASE_ADDRESS_HI(gsvs_va
>> 32);
3068 desc
[2] = gsvs_ring_size
;
3069 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
3070 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
3071 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
3072 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
);
3074 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
3075 desc
[3] |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT
) |
3076 S_008F0C_OOB_SELECT(2) |
3077 S_008F0C_RESOURCE_LEVEL(1);
3079 desc
[3] |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
3080 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
);
3083 /* stride gsvs_itemsize, num records 64
3084 elsize 4, index stride 16 */
3085 /* shader will patch stride and desc[2] */
3087 desc
[5] = S_008F04_BASE_ADDRESS_HI(gsvs_va
>> 32) |
3088 S_008F04_SWIZZLE_ENABLE(1);
3090 desc
[7] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
3091 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
3092 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
3093 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
3094 S_008F0C_INDEX_STRIDE(1) |
3095 S_008F0C_ADD_TID_ENABLE(true);
3097 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
3098 desc
[7] |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT
) |
3099 S_008F0C_OOB_SELECT(2) |
3100 S_008F0C_RESOURCE_LEVEL(1);
3102 desc
[7] |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
3103 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
3104 S_008F0C_ELEMENT_SIZE(1);
3111 if (tess_rings_bo
) {
3112 uint64_t tess_va
= radv_buffer_get_va(tess_rings_bo
);
3113 uint64_t tess_offchip_va
= tess_va
+ tess_offchip_ring_offset
;
3116 desc
[1] = S_008F04_BASE_ADDRESS_HI(tess_va
>> 32);
3117 desc
[2] = tess_factor_ring_size
;
3118 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
3119 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
3120 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
3121 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
);
3123 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
3124 desc
[3] |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT
) |
3125 S_008F0C_OOB_SELECT(3) |
3126 S_008F0C_RESOURCE_LEVEL(1);
3128 desc
[3] |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
3129 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
);
3132 desc
[4] = tess_offchip_va
;
3133 desc
[5] = S_008F04_BASE_ADDRESS_HI(tess_offchip_va
>> 32);
3134 desc
[6] = tess_offchip_ring_size
;
3135 desc
[7] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
3136 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
3137 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
3138 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
);
3140 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
3141 desc
[7] |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT
) |
3142 S_008F0C_OOB_SELECT(3) |
3143 S_008F0C_RESOURCE_LEVEL(1);
3145 desc
[7] |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
3146 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
);
3152 if (add_sample_positions
) {
3153 /* add sample positions after all rings */
3154 memcpy(desc
, queue
->device
->sample_locations_1x
, 8);
3156 memcpy(desc
, queue
->device
->sample_locations_2x
, 16);
3158 memcpy(desc
, queue
->device
->sample_locations_4x
, 32);
3160 memcpy(desc
, queue
->device
->sample_locations_8x
, 64);
3165 radv_get_hs_offchip_param(struct radv_device
*device
, uint32_t *max_offchip_buffers_p
)
3167 bool double_offchip_buffers
= device
->physical_device
->rad_info
.chip_class
>= GFX7
&&
3168 device
->physical_device
->rad_info
.family
!= CHIP_CARRIZO
&&
3169 device
->physical_device
->rad_info
.family
!= CHIP_STONEY
;
3170 unsigned max_offchip_buffers_per_se
= double_offchip_buffers
? 128 : 64;
3171 unsigned max_offchip_buffers
;
3172 unsigned offchip_granularity
;
3173 unsigned hs_offchip_param
;
3177 * This must be one less than the maximum number due to a hw limitation.
3178 * Various hardware bugs need thGFX7
3181 * Vega10 should limit max_offchip_buffers to 508 (4 * 127).
3182 * Gfx7 should limit max_offchip_buffers to 508
3183 * Gfx6 should limit max_offchip_buffers to 126 (2 * 63)
3185 * Follow AMDVLK here.
3187 if (device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
3188 max_offchip_buffers_per_se
= 256;
3189 } else if (device
->physical_device
->rad_info
.family
== CHIP_VEGA10
||
3190 device
->physical_device
->rad_info
.chip_class
== GFX7
||
3191 device
->physical_device
->rad_info
.chip_class
== GFX6
)
3192 --max_offchip_buffers_per_se
;
3194 max_offchip_buffers
= max_offchip_buffers_per_se
*
3195 device
->physical_device
->rad_info
.max_se
;
3197 /* Hawaii has a bug with offchip buffers > 256 that can be worked
3198 * around by setting 4K granularity.
3200 if (device
->tess_offchip_block_dw_size
== 4096) {
3201 assert(device
->physical_device
->rad_info
.family
== CHIP_HAWAII
);
3202 offchip_granularity
= V_03093C_X_4K_DWORDS
;
3204 assert(device
->tess_offchip_block_dw_size
== 8192);
3205 offchip_granularity
= V_03093C_X_8K_DWORDS
;
3208 switch (device
->physical_device
->rad_info
.chip_class
) {
3210 max_offchip_buffers
= MIN2(max_offchip_buffers
, 126);
3215 max_offchip_buffers
= MIN2(max_offchip_buffers
, 508);
3223 *max_offchip_buffers_p
= max_offchip_buffers
;
3224 if (device
->physical_device
->rad_info
.chip_class
>= GFX7
) {
3225 if (device
->physical_device
->rad_info
.chip_class
>= GFX8
)
3226 --max_offchip_buffers
;
3228 S_03093C_OFFCHIP_BUFFERING(max_offchip_buffers
) |
3229 S_03093C_OFFCHIP_GRANULARITY(offchip_granularity
);
3232 S_0089B0_OFFCHIP_BUFFERING(max_offchip_buffers
);
3234 return hs_offchip_param
;
3238 radv_emit_gs_ring_sizes(struct radv_queue
*queue
, struct radeon_cmdbuf
*cs
,
3239 struct radeon_winsys_bo
*esgs_ring_bo
,
3240 uint32_t esgs_ring_size
,
3241 struct radeon_winsys_bo
*gsvs_ring_bo
,
3242 uint32_t gsvs_ring_size
)
3244 if (!esgs_ring_bo
&& !gsvs_ring_bo
)
3248 radv_cs_add_buffer(queue
->device
->ws
, cs
, esgs_ring_bo
);
3251 radv_cs_add_buffer(queue
->device
->ws
, cs
, gsvs_ring_bo
);
3253 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX7
) {
3254 radeon_set_uconfig_reg_seq(cs
, R_030900_VGT_ESGS_RING_SIZE
, 2);
3255 radeon_emit(cs
, esgs_ring_size
>> 8);
3256 radeon_emit(cs
, gsvs_ring_size
>> 8);
3258 radeon_set_config_reg_seq(cs
, R_0088C8_VGT_ESGS_RING_SIZE
, 2);
3259 radeon_emit(cs
, esgs_ring_size
>> 8);
3260 radeon_emit(cs
, gsvs_ring_size
>> 8);
3265 radv_emit_tess_factor_ring(struct radv_queue
*queue
, struct radeon_cmdbuf
*cs
,
3266 unsigned hs_offchip_param
, unsigned tf_ring_size
,
3267 struct radeon_winsys_bo
*tess_rings_bo
)
3274 tf_va
= radv_buffer_get_va(tess_rings_bo
);
3276 radv_cs_add_buffer(queue
->device
->ws
, cs
, tess_rings_bo
);
3278 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX7
) {
3279 radeon_set_uconfig_reg(cs
, R_030938_VGT_TF_RING_SIZE
,
3280 S_030938_SIZE(tf_ring_size
/ 4));
3281 radeon_set_uconfig_reg(cs
, R_030940_VGT_TF_MEMORY_BASE
,
3284 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
3285 radeon_set_uconfig_reg(cs
, R_030984_VGT_TF_MEMORY_BASE_HI_UMD
,
3286 S_030984_BASE_HI(tf_va
>> 40));
3287 } else if (queue
->device
->physical_device
->rad_info
.chip_class
== GFX9
) {
3288 radeon_set_uconfig_reg(cs
, R_030944_VGT_TF_MEMORY_BASE_HI
,
3289 S_030944_BASE_HI(tf_va
>> 40));
3291 radeon_set_uconfig_reg(cs
, R_03093C_VGT_HS_OFFCHIP_PARAM
,
3294 radeon_set_config_reg(cs
, R_008988_VGT_TF_RING_SIZE
,
3295 S_008988_SIZE(tf_ring_size
/ 4));
3296 radeon_set_config_reg(cs
, R_0089B8_VGT_TF_MEMORY_BASE
,
3298 radeon_set_config_reg(cs
, R_0089B0_VGT_HS_OFFCHIP_PARAM
,
3304 radv_emit_graphics_scratch(struct radv_queue
*queue
, struct radeon_cmdbuf
*cs
,
3305 uint32_t size_per_wave
, uint32_t waves
,
3306 struct radeon_winsys_bo
*scratch_bo
)
3308 if (queue
->queue_family_index
!= RADV_QUEUE_GENERAL
)
3314 radv_cs_add_buffer(queue
->device
->ws
, cs
, scratch_bo
);
3316 radeon_set_context_reg(cs
, R_0286E8_SPI_TMPRING_SIZE
,
3317 S_0286E8_WAVES(waves
) |
3318 S_0286E8_WAVESIZE(round_up_u32(size_per_wave
, 1024)));
3322 radv_emit_compute_scratch(struct radv_queue
*queue
, struct radeon_cmdbuf
*cs
,
3323 uint32_t size_per_wave
, uint32_t waves
,
3324 struct radeon_winsys_bo
*compute_scratch_bo
)
3326 uint64_t scratch_va
;
3328 if (!compute_scratch_bo
)
3331 scratch_va
= radv_buffer_get_va(compute_scratch_bo
);
3333 radv_cs_add_buffer(queue
->device
->ws
, cs
, compute_scratch_bo
);
3335 radeon_set_sh_reg_seq(cs
, R_00B900_COMPUTE_USER_DATA_0
, 2);
3336 radeon_emit(cs
, scratch_va
);
3337 radeon_emit(cs
, S_008F04_BASE_ADDRESS_HI(scratch_va
>> 32) |
3338 S_008F04_SWIZZLE_ENABLE(1));
3340 radeon_set_sh_reg(cs
, R_00B860_COMPUTE_TMPRING_SIZE
,
3341 S_00B860_WAVES(waves
) |
3342 S_00B860_WAVESIZE(round_up_u32(size_per_wave
, 1024)));
3346 radv_emit_global_shader_pointers(struct radv_queue
*queue
,
3347 struct radeon_cmdbuf
*cs
,
3348 struct radeon_winsys_bo
*descriptor_bo
)
3355 va
= radv_buffer_get_va(descriptor_bo
);
3357 radv_cs_add_buffer(queue
->device
->ws
, cs
, descriptor_bo
);
3359 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
3360 uint32_t regs
[] = {R_00B030_SPI_SHADER_USER_DATA_PS_0
,
3361 R_00B130_SPI_SHADER_USER_DATA_VS_0
,
3362 R_00B208_SPI_SHADER_USER_DATA_ADDR_LO_GS
,
3363 R_00B408_SPI_SHADER_USER_DATA_ADDR_LO_HS
};
3365 for (int i
= 0; i
< ARRAY_SIZE(regs
); ++i
) {
3366 radv_emit_shader_pointer(queue
->device
, cs
, regs
[i
],
3369 } else if (queue
->device
->physical_device
->rad_info
.chip_class
== GFX9
) {
3370 uint32_t regs
[] = {R_00B030_SPI_SHADER_USER_DATA_PS_0
,
3371 R_00B130_SPI_SHADER_USER_DATA_VS_0
,
3372 R_00B208_SPI_SHADER_USER_DATA_ADDR_LO_GS
,
3373 R_00B408_SPI_SHADER_USER_DATA_ADDR_LO_HS
};
3375 for (int i
= 0; i
< ARRAY_SIZE(regs
); ++i
) {
3376 radv_emit_shader_pointer(queue
->device
, cs
, regs
[i
],
3380 uint32_t regs
[] = {R_00B030_SPI_SHADER_USER_DATA_PS_0
,
3381 R_00B130_SPI_SHADER_USER_DATA_VS_0
,
3382 R_00B230_SPI_SHADER_USER_DATA_GS_0
,
3383 R_00B330_SPI_SHADER_USER_DATA_ES_0
,
3384 R_00B430_SPI_SHADER_USER_DATA_HS_0
,
3385 R_00B530_SPI_SHADER_USER_DATA_LS_0
};
3387 for (int i
= 0; i
< ARRAY_SIZE(regs
); ++i
) {
3388 radv_emit_shader_pointer(queue
->device
, cs
, regs
[i
],
3395 radv_init_graphics_state(struct radeon_cmdbuf
*cs
, struct radv_queue
*queue
)
3397 struct radv_device
*device
= queue
->device
;
3399 if (device
->gfx_init
) {
3400 uint64_t va
= radv_buffer_get_va(device
->gfx_init
);
3402 radeon_emit(cs
, PKT3(PKT3_INDIRECT_BUFFER_CIK
, 2, 0));
3403 radeon_emit(cs
, va
);
3404 radeon_emit(cs
, va
>> 32);
3405 radeon_emit(cs
, device
->gfx_init_size_dw
& 0xffff);
3407 radv_cs_add_buffer(device
->ws
, cs
, device
->gfx_init
);
3409 struct radv_physical_device
*physical_device
= device
->physical_device
;
3410 si_emit_graphics(physical_device
, cs
);
3415 radv_init_compute_state(struct radeon_cmdbuf
*cs
, struct radv_queue
*queue
)
3417 struct radv_physical_device
*physical_device
= queue
->device
->physical_device
;
3418 si_emit_compute(physical_device
, cs
);
3422 radv_get_preamble_cs(struct radv_queue
*queue
,
3423 uint32_t scratch_size_per_wave
,
3424 uint32_t scratch_waves
,
3425 uint32_t compute_scratch_size_per_wave
,
3426 uint32_t compute_scratch_waves
,
3427 uint32_t esgs_ring_size
,
3428 uint32_t gsvs_ring_size
,
3429 bool needs_tess_rings
,
3431 bool needs_sample_positions
,
3432 struct radeon_cmdbuf
**initial_full_flush_preamble_cs
,
3433 struct radeon_cmdbuf
**initial_preamble_cs
,
3434 struct radeon_cmdbuf
**continue_preamble_cs
)
3436 struct radeon_winsys_bo
*scratch_bo
= NULL
;
3437 struct radeon_winsys_bo
*descriptor_bo
= NULL
;
3438 struct radeon_winsys_bo
*compute_scratch_bo
= NULL
;
3439 struct radeon_winsys_bo
*esgs_ring_bo
= NULL
;
3440 struct radeon_winsys_bo
*gsvs_ring_bo
= NULL
;
3441 struct radeon_winsys_bo
*tess_rings_bo
= NULL
;
3442 struct radeon_winsys_bo
*gds_bo
= NULL
;
3443 struct radeon_winsys_bo
*gds_oa_bo
= NULL
;
3444 struct radeon_cmdbuf
*dest_cs
[3] = {0};
3445 bool add_tess_rings
= false, add_gds
= false, add_sample_positions
= false;
3446 unsigned tess_factor_ring_size
= 0, tess_offchip_ring_size
= 0;
3447 unsigned max_offchip_buffers
;
3448 unsigned hs_offchip_param
= 0;
3449 unsigned tess_offchip_ring_offset
;
3450 uint32_t ring_bo_flags
= RADEON_FLAG_NO_CPU_ACCESS
| RADEON_FLAG_NO_INTERPROCESS_SHARING
;
3451 if (!queue
->has_tess_rings
) {
3452 if (needs_tess_rings
)
3453 add_tess_rings
= true;
3455 if (!queue
->has_gds
) {
3459 if (!queue
->has_sample_positions
) {
3460 if (needs_sample_positions
)
3461 add_sample_positions
= true;
3463 tess_factor_ring_size
= 32768 * queue
->device
->physical_device
->rad_info
.max_se
;
3464 hs_offchip_param
= radv_get_hs_offchip_param(queue
->device
,
3465 &max_offchip_buffers
);
3466 tess_offchip_ring_offset
= align(tess_factor_ring_size
, 64 * 1024);
3467 tess_offchip_ring_size
= max_offchip_buffers
*
3468 queue
->device
->tess_offchip_block_dw_size
* 4;
3470 scratch_size_per_wave
= MAX2(scratch_size_per_wave
, queue
->scratch_size_per_wave
);
3471 if (scratch_size_per_wave
)
3472 scratch_waves
= MIN2(scratch_waves
, UINT32_MAX
/ scratch_size_per_wave
);
3476 compute_scratch_size_per_wave
= MAX2(compute_scratch_size_per_wave
, queue
->compute_scratch_size_per_wave
);
3477 if (compute_scratch_size_per_wave
)
3478 compute_scratch_waves
= MIN2(compute_scratch_waves
, UINT32_MAX
/ compute_scratch_size_per_wave
);
3480 compute_scratch_waves
= 0;
3482 if (scratch_size_per_wave
<= queue
->scratch_size_per_wave
&&
3483 scratch_waves
<= queue
->scratch_waves
&&
3484 compute_scratch_size_per_wave
<= queue
->compute_scratch_size_per_wave
&&
3485 compute_scratch_waves
<= queue
->compute_scratch_waves
&&
3486 esgs_ring_size
<= queue
->esgs_ring_size
&&
3487 gsvs_ring_size
<= queue
->gsvs_ring_size
&&
3488 !add_tess_rings
&& !add_gds
&& !add_sample_positions
&&
3489 queue
->initial_preamble_cs
) {
3490 *initial_full_flush_preamble_cs
= queue
->initial_full_flush_preamble_cs
;
3491 *initial_preamble_cs
= queue
->initial_preamble_cs
;
3492 *continue_preamble_cs
= queue
->continue_preamble_cs
;
3493 if (!scratch_size_per_wave
&& !compute_scratch_size_per_wave
&&
3494 !esgs_ring_size
&& !gsvs_ring_size
&& !needs_tess_rings
&&
3495 !needs_gds
&& !needs_sample_positions
)
3496 *continue_preamble_cs
= NULL
;
3500 uint32_t scratch_size
= scratch_size_per_wave
* scratch_waves
;
3501 uint32_t queue_scratch_size
= queue
->scratch_size_per_wave
* queue
->scratch_waves
;
3502 if (scratch_size
> queue_scratch_size
) {
3503 scratch_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
3508 RADV_BO_PRIORITY_SCRATCH
);
3512 scratch_bo
= queue
->scratch_bo
;
3514 uint32_t compute_scratch_size
= compute_scratch_size_per_wave
* compute_scratch_waves
;
3515 uint32_t compute_queue_scratch_size
= queue
->compute_scratch_size_per_wave
* queue
->compute_scratch_waves
;
3516 if (compute_scratch_size
> compute_queue_scratch_size
) {
3517 compute_scratch_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
3518 compute_scratch_size
,
3522 RADV_BO_PRIORITY_SCRATCH
);
3523 if (!compute_scratch_bo
)
3527 compute_scratch_bo
= queue
->compute_scratch_bo
;
3529 if (esgs_ring_size
> queue
->esgs_ring_size
) {
3530 esgs_ring_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
3535 RADV_BO_PRIORITY_SCRATCH
);
3539 esgs_ring_bo
= queue
->esgs_ring_bo
;
3540 esgs_ring_size
= queue
->esgs_ring_size
;
3543 if (gsvs_ring_size
> queue
->gsvs_ring_size
) {
3544 gsvs_ring_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
3549 RADV_BO_PRIORITY_SCRATCH
);
3553 gsvs_ring_bo
= queue
->gsvs_ring_bo
;
3554 gsvs_ring_size
= queue
->gsvs_ring_size
;
3557 if (add_tess_rings
) {
3558 tess_rings_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
3559 tess_offchip_ring_offset
+ tess_offchip_ring_size
,
3563 RADV_BO_PRIORITY_SCRATCH
);
3567 tess_rings_bo
= queue
->tess_rings_bo
;
3571 assert(queue
->device
->physical_device
->rad_info
.chip_class
>= GFX10
);
3573 /* 4 streamout GDS counters.
3574 * We need 256B (64 dw) of GDS, otherwise streamout hangs.
3576 gds_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
3580 RADV_BO_PRIORITY_SCRATCH
);
3584 gds_oa_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
3588 RADV_BO_PRIORITY_SCRATCH
);
3592 gds_bo
= queue
->gds_bo
;
3593 gds_oa_bo
= queue
->gds_oa_bo
;
3596 if (scratch_bo
!= queue
->scratch_bo
||
3597 esgs_ring_bo
!= queue
->esgs_ring_bo
||
3598 gsvs_ring_bo
!= queue
->gsvs_ring_bo
||
3599 tess_rings_bo
!= queue
->tess_rings_bo
||
3600 add_sample_positions
) {
3602 if (gsvs_ring_bo
|| esgs_ring_bo
||
3603 tess_rings_bo
|| add_sample_positions
) {
3604 size
= 112; /* 2 dword + 2 padding + 4 dword * 6 */
3605 if (add_sample_positions
)
3606 size
+= 128; /* 64+32+16+8 = 120 bytes */
3608 else if (scratch_bo
)
3609 size
= 8; /* 2 dword */
3611 descriptor_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
3615 RADEON_FLAG_CPU_ACCESS
|
3616 RADEON_FLAG_NO_INTERPROCESS_SHARING
|
3617 RADEON_FLAG_READ_ONLY
,
3618 RADV_BO_PRIORITY_DESCRIPTOR
);
3622 descriptor_bo
= queue
->descriptor_bo
;
3624 if (descriptor_bo
!= queue
->descriptor_bo
) {
3625 uint32_t *map
= (uint32_t*)queue
->device
->ws
->buffer_map(descriptor_bo
);
3628 uint64_t scratch_va
= radv_buffer_get_va(scratch_bo
);
3629 uint32_t rsrc1
= S_008F04_BASE_ADDRESS_HI(scratch_va
>> 32) |
3630 S_008F04_SWIZZLE_ENABLE(1);
3631 map
[0] = scratch_va
;
3635 if (esgs_ring_bo
|| gsvs_ring_bo
|| tess_rings_bo
|| add_sample_positions
)
3636 fill_geom_tess_rings(queue
, map
, add_sample_positions
,
3637 esgs_ring_size
, esgs_ring_bo
,
3638 gsvs_ring_size
, gsvs_ring_bo
,
3639 tess_factor_ring_size
,
3640 tess_offchip_ring_offset
,
3641 tess_offchip_ring_size
,
3644 queue
->device
->ws
->buffer_unmap(descriptor_bo
);
3647 for(int i
= 0; i
< 3; ++i
) {
3648 struct radeon_cmdbuf
*cs
= NULL
;
3649 cs
= queue
->device
->ws
->cs_create(queue
->device
->ws
,
3650 queue
->queue_family_index
? RING_COMPUTE
: RING_GFX
);
3657 radv_cs_add_buffer(queue
->device
->ws
, cs
, scratch_bo
);
3659 /* Emit initial configuration. */
3660 switch (queue
->queue_family_index
) {
3661 case RADV_QUEUE_GENERAL
:
3662 radv_init_graphics_state(cs
, queue
);
3664 case RADV_QUEUE_COMPUTE
:
3665 radv_init_compute_state(cs
, queue
);
3667 case RADV_QUEUE_TRANSFER
:
3671 if (esgs_ring_bo
|| gsvs_ring_bo
|| tess_rings_bo
) {
3672 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
3673 radeon_emit(cs
, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
3675 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
3676 radeon_emit(cs
, EVENT_TYPE(V_028A90_VGT_FLUSH
) | EVENT_INDEX(0));
3679 radv_emit_gs_ring_sizes(queue
, cs
, esgs_ring_bo
, esgs_ring_size
,
3680 gsvs_ring_bo
, gsvs_ring_size
);
3681 radv_emit_tess_factor_ring(queue
, cs
, hs_offchip_param
,
3682 tess_factor_ring_size
, tess_rings_bo
);
3683 radv_emit_global_shader_pointers(queue
, cs
, descriptor_bo
);
3684 radv_emit_compute_scratch(queue
, cs
, compute_scratch_size_per_wave
,
3685 compute_scratch_waves
, compute_scratch_bo
);
3686 radv_emit_graphics_scratch(queue
, cs
, scratch_size_per_wave
,
3687 scratch_waves
, scratch_bo
);
3690 radv_cs_add_buffer(queue
->device
->ws
, cs
, gds_bo
);
3692 radv_cs_add_buffer(queue
->device
->ws
, cs
, gds_oa_bo
);
3695 si_cs_emit_cache_flush(cs
,
3696 queue
->device
->physical_device
->rad_info
.chip_class
,
3698 queue
->queue_family_index
== RING_COMPUTE
&&
3699 queue
->device
->physical_device
->rad_info
.chip_class
>= GFX7
,
3700 (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
)) |
3701 RADV_CMD_FLAG_INV_ICACHE
|
3702 RADV_CMD_FLAG_INV_SCACHE
|
3703 RADV_CMD_FLAG_INV_VCACHE
|
3704 RADV_CMD_FLAG_INV_L2
|
3705 RADV_CMD_FLAG_START_PIPELINE_STATS
, 0);
3706 } else if (i
== 1) {
3707 si_cs_emit_cache_flush(cs
,
3708 queue
->device
->physical_device
->rad_info
.chip_class
,
3710 queue
->queue_family_index
== RING_COMPUTE
&&
3711 queue
->device
->physical_device
->rad_info
.chip_class
>= GFX7
,
3712 RADV_CMD_FLAG_INV_ICACHE
|
3713 RADV_CMD_FLAG_INV_SCACHE
|
3714 RADV_CMD_FLAG_INV_VCACHE
|
3715 RADV_CMD_FLAG_INV_L2
|
3716 RADV_CMD_FLAG_START_PIPELINE_STATS
, 0);
3719 if (!queue
->device
->ws
->cs_finalize(cs
))
3723 if (queue
->initial_full_flush_preamble_cs
)
3724 queue
->device
->ws
->cs_destroy(queue
->initial_full_flush_preamble_cs
);
3726 if (queue
->initial_preamble_cs
)
3727 queue
->device
->ws
->cs_destroy(queue
->initial_preamble_cs
);
3729 if (queue
->continue_preamble_cs
)
3730 queue
->device
->ws
->cs_destroy(queue
->continue_preamble_cs
);
3732 queue
->initial_full_flush_preamble_cs
= dest_cs
[0];
3733 queue
->initial_preamble_cs
= dest_cs
[1];
3734 queue
->continue_preamble_cs
= dest_cs
[2];
3736 if (scratch_bo
!= queue
->scratch_bo
) {
3737 if (queue
->scratch_bo
)
3738 queue
->device
->ws
->buffer_destroy(queue
->scratch_bo
);
3739 queue
->scratch_bo
= scratch_bo
;
3741 queue
->scratch_size_per_wave
= scratch_size_per_wave
;
3742 queue
->scratch_waves
= scratch_waves
;
3744 if (compute_scratch_bo
!= queue
->compute_scratch_bo
) {
3745 if (queue
->compute_scratch_bo
)
3746 queue
->device
->ws
->buffer_destroy(queue
->compute_scratch_bo
);
3747 queue
->compute_scratch_bo
= compute_scratch_bo
;
3749 queue
->compute_scratch_size_per_wave
= compute_scratch_size_per_wave
;
3750 queue
->compute_scratch_waves
= compute_scratch_waves
;
3752 if (esgs_ring_bo
!= queue
->esgs_ring_bo
) {
3753 if (queue
->esgs_ring_bo
)
3754 queue
->device
->ws
->buffer_destroy(queue
->esgs_ring_bo
);
3755 queue
->esgs_ring_bo
= esgs_ring_bo
;
3756 queue
->esgs_ring_size
= esgs_ring_size
;
3759 if (gsvs_ring_bo
!= queue
->gsvs_ring_bo
) {
3760 if (queue
->gsvs_ring_bo
)
3761 queue
->device
->ws
->buffer_destroy(queue
->gsvs_ring_bo
);
3762 queue
->gsvs_ring_bo
= gsvs_ring_bo
;
3763 queue
->gsvs_ring_size
= gsvs_ring_size
;
3766 if (tess_rings_bo
!= queue
->tess_rings_bo
) {
3767 queue
->tess_rings_bo
= tess_rings_bo
;
3768 queue
->has_tess_rings
= true;
3771 if (gds_bo
!= queue
->gds_bo
) {
3772 queue
->gds_bo
= gds_bo
;
3773 queue
->has_gds
= true;
3776 if (gds_oa_bo
!= queue
->gds_oa_bo
)
3777 queue
->gds_oa_bo
= gds_oa_bo
;
3779 if (descriptor_bo
!= queue
->descriptor_bo
) {
3780 if (queue
->descriptor_bo
)
3781 queue
->device
->ws
->buffer_destroy(queue
->descriptor_bo
);
3783 queue
->descriptor_bo
= descriptor_bo
;
3786 if (add_sample_positions
)
3787 queue
->has_sample_positions
= true;
3789 *initial_full_flush_preamble_cs
= queue
->initial_full_flush_preamble_cs
;
3790 *initial_preamble_cs
= queue
->initial_preamble_cs
;
3791 *continue_preamble_cs
= queue
->continue_preamble_cs
;
3792 if (!scratch_size
&& !compute_scratch_size
&& !esgs_ring_size
&& !gsvs_ring_size
)
3793 *continue_preamble_cs
= NULL
;
3796 for (int i
= 0; i
< ARRAY_SIZE(dest_cs
); ++i
)
3798 queue
->device
->ws
->cs_destroy(dest_cs
[i
]);
3799 if (descriptor_bo
&& descriptor_bo
!= queue
->descriptor_bo
)
3800 queue
->device
->ws
->buffer_destroy(descriptor_bo
);
3801 if (scratch_bo
&& scratch_bo
!= queue
->scratch_bo
)
3802 queue
->device
->ws
->buffer_destroy(scratch_bo
);
3803 if (compute_scratch_bo
&& compute_scratch_bo
!= queue
->compute_scratch_bo
)
3804 queue
->device
->ws
->buffer_destroy(compute_scratch_bo
);
3805 if (esgs_ring_bo
&& esgs_ring_bo
!= queue
->esgs_ring_bo
)
3806 queue
->device
->ws
->buffer_destroy(esgs_ring_bo
);
3807 if (gsvs_ring_bo
&& gsvs_ring_bo
!= queue
->gsvs_ring_bo
)
3808 queue
->device
->ws
->buffer_destroy(gsvs_ring_bo
);
3809 if (tess_rings_bo
&& tess_rings_bo
!= queue
->tess_rings_bo
)
3810 queue
->device
->ws
->buffer_destroy(tess_rings_bo
);
3811 if (gds_bo
&& gds_bo
!= queue
->gds_bo
)
3812 queue
->device
->ws
->buffer_destroy(gds_bo
);
3813 if (gds_oa_bo
&& gds_oa_bo
!= queue
->gds_oa_bo
)
3814 queue
->device
->ws
->buffer_destroy(gds_oa_bo
);
3816 return vk_error(queue
->device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
3819 static VkResult
radv_alloc_sem_counts(struct radv_device
*device
,
3820 struct radv_winsys_sem_counts
*counts
,
3822 struct radv_semaphore_part
**sems
,
3823 const uint64_t *timeline_values
,
3827 int syncobj_idx
= 0, sem_idx
= 0;
3829 if (num_sems
== 0 && _fence
== VK_NULL_HANDLE
)
3832 for (uint32_t i
= 0; i
< num_sems
; i
++) {
3833 switch(sems
[i
]->kind
) {
3834 case RADV_SEMAPHORE_SYNCOBJ
:
3835 counts
->syncobj_count
++;
3837 case RADV_SEMAPHORE_WINSYS
:
3838 counts
->sem_count
++;
3840 case RADV_SEMAPHORE_NONE
:
3842 case RADV_SEMAPHORE_TIMELINE
:
3843 counts
->syncobj_count
++;
3848 if (_fence
!= VK_NULL_HANDLE
) {
3849 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
3850 if (fence
->temp_syncobj
|| fence
->syncobj
)
3851 counts
->syncobj_count
++;
3854 if (counts
->syncobj_count
) {
3855 counts
->syncobj
= (uint32_t *)malloc(sizeof(uint32_t) * counts
->syncobj_count
);
3856 if (!counts
->syncobj
)
3857 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3860 if (counts
->sem_count
) {
3861 counts
->sem
= (struct radeon_winsys_sem
**)malloc(sizeof(struct radeon_winsys_sem
*) * counts
->sem_count
);
3863 free(counts
->syncobj
);
3864 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3868 for (uint32_t i
= 0; i
< num_sems
; i
++) {
3869 switch(sems
[i
]->kind
) {
3870 case RADV_SEMAPHORE_NONE
:
3871 unreachable("Empty semaphore");
3873 case RADV_SEMAPHORE_SYNCOBJ
:
3874 counts
->syncobj
[syncobj_idx
++] = sems
[i
]->syncobj
;
3876 case RADV_SEMAPHORE_WINSYS
:
3877 counts
->sem
[sem_idx
++] = sems
[i
]->ws_sem
;
3879 case RADV_SEMAPHORE_TIMELINE
: {
3880 pthread_mutex_lock(&sems
[i
]->timeline
.mutex
);
3881 struct radv_timeline_point
*point
= NULL
;
3883 point
= radv_timeline_add_point_locked(device
, &sems
[i
]->timeline
, timeline_values
[i
]);
3885 point
= radv_timeline_find_point_at_least_locked(device
, &sems
[i
]->timeline
, timeline_values
[i
]);
3888 pthread_mutex_unlock(&sems
[i
]->timeline
.mutex
);
3891 counts
->syncobj
[syncobj_idx
++] = point
->syncobj
;
3893 /* Explicitly remove the semaphore so we might not find
3894 * a point later post-submit. */
3902 if (_fence
!= VK_NULL_HANDLE
) {
3903 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
3904 if (fence
->temp_syncobj
)
3905 counts
->syncobj
[syncobj_idx
++] = fence
->temp_syncobj
;
3906 else if (fence
->syncobj
)
3907 counts
->syncobj
[syncobj_idx
++] = fence
->syncobj
;
3910 assert(syncobj_idx
<= counts
->syncobj_count
);
3911 counts
->syncobj_count
= syncobj_idx
;
3917 radv_free_sem_info(struct radv_winsys_sem_info
*sem_info
)
3919 free(sem_info
->wait
.syncobj
);
3920 free(sem_info
->wait
.sem
);
3921 free(sem_info
->signal
.syncobj
);
3922 free(sem_info
->signal
.sem
);
3926 static void radv_free_temp_syncobjs(struct radv_device
*device
,
3928 struct radv_semaphore_part
*sems
)
3930 for (uint32_t i
= 0; i
< num_sems
; i
++) {
3931 radv_destroy_semaphore_part(device
, sems
+ i
);
3936 radv_alloc_sem_info(struct radv_device
*device
,
3937 struct radv_winsys_sem_info
*sem_info
,
3939 struct radv_semaphore_part
**wait_sems
,
3940 const uint64_t *wait_values
,
3941 int num_signal_sems
,
3942 struct radv_semaphore_part
**signal_sems
,
3943 const uint64_t *signal_values
,
3947 memset(sem_info
, 0, sizeof(*sem_info
));
3949 ret
= radv_alloc_sem_counts(device
, &sem_info
->wait
, num_wait_sems
, wait_sems
, wait_values
, VK_NULL_HANDLE
, false);
3952 ret
= radv_alloc_sem_counts(device
, &sem_info
->signal
, num_signal_sems
, signal_sems
, signal_values
, fence
, true);
3954 radv_free_sem_info(sem_info
);
3956 /* caller can override these */
3957 sem_info
->cs_emit_wait
= true;
3958 sem_info
->cs_emit_signal
= true;
3963 radv_finalize_timelines(struct radv_device
*device
,
3964 uint32_t num_wait_sems
,
3965 struct radv_semaphore_part
**wait_sems
,
3966 const uint64_t *wait_values
,
3967 uint32_t num_signal_sems
,
3968 struct radv_semaphore_part
**signal_sems
,
3969 const uint64_t *signal_values
,
3970 struct list_head
*processing_list
)
3972 for (uint32_t i
= 0; i
< num_wait_sems
; ++i
) {
3973 if (wait_sems
[i
] && wait_sems
[i
]->kind
== RADV_SEMAPHORE_TIMELINE
) {
3974 pthread_mutex_lock(&wait_sems
[i
]->timeline
.mutex
);
3975 struct radv_timeline_point
*point
=
3976 radv_timeline_find_point_at_least_locked(device
, &wait_sems
[i
]->timeline
, wait_values
[i
]);
3977 point
->wait_count
-= 2;
3978 pthread_mutex_unlock(&wait_sems
[i
]->timeline
.mutex
);
3981 for (uint32_t i
= 0; i
< num_signal_sems
; ++i
) {
3982 if (signal_sems
[i
] && signal_sems
[i
]->kind
== RADV_SEMAPHORE_TIMELINE
) {
3983 pthread_mutex_lock(&signal_sems
[i
]->timeline
.mutex
);
3984 struct radv_timeline_point
*point
=
3985 radv_timeline_find_point_at_least_locked(device
, &signal_sems
[i
]->timeline
, signal_values
[i
]);
3986 signal_sems
[i
]->timeline
.highest_submitted
=
3987 MAX2(signal_sems
[i
]->timeline
.highest_submitted
, point
->value
);
3988 point
->wait_count
-= 2;
3989 radv_timeline_trigger_waiters_locked(&signal_sems
[i
]->timeline
, processing_list
);
3990 pthread_mutex_unlock(&signal_sems
[i
]->timeline
.mutex
);
3996 radv_sparse_buffer_bind_memory(struct radv_device
*device
,
3997 const VkSparseBufferMemoryBindInfo
*bind
)
3999 RADV_FROM_HANDLE(radv_buffer
, buffer
, bind
->buffer
);
4001 for (uint32_t i
= 0; i
< bind
->bindCount
; ++i
) {
4002 struct radv_device_memory
*mem
= NULL
;
4004 if (bind
->pBinds
[i
].memory
!= VK_NULL_HANDLE
)
4005 mem
= radv_device_memory_from_handle(bind
->pBinds
[i
].memory
);
4007 device
->ws
->buffer_virtual_bind(buffer
->bo
,
4008 bind
->pBinds
[i
].resourceOffset
,
4009 bind
->pBinds
[i
].size
,
4010 mem
? mem
->bo
: NULL
,
4011 bind
->pBinds
[i
].memoryOffset
);
4016 radv_sparse_image_opaque_bind_memory(struct radv_device
*device
,
4017 const VkSparseImageOpaqueMemoryBindInfo
*bind
)
4019 RADV_FROM_HANDLE(radv_image
, image
, bind
->image
);
4021 for (uint32_t i
= 0; i
< bind
->bindCount
; ++i
) {
4022 struct radv_device_memory
*mem
= NULL
;
4024 if (bind
->pBinds
[i
].memory
!= VK_NULL_HANDLE
)
4025 mem
= radv_device_memory_from_handle(bind
->pBinds
[i
].memory
);
4027 device
->ws
->buffer_virtual_bind(image
->bo
,
4028 bind
->pBinds
[i
].resourceOffset
,
4029 bind
->pBinds
[i
].size
,
4030 mem
? mem
->bo
: NULL
,
4031 bind
->pBinds
[i
].memoryOffset
);
4036 radv_get_preambles(struct radv_queue
*queue
,
4037 const VkCommandBuffer
*cmd_buffers
,
4038 uint32_t cmd_buffer_count
,
4039 struct radeon_cmdbuf
**initial_full_flush_preamble_cs
,
4040 struct radeon_cmdbuf
**initial_preamble_cs
,
4041 struct radeon_cmdbuf
**continue_preamble_cs
)
4043 uint32_t scratch_size_per_wave
= 0, waves_wanted
= 0;
4044 uint32_t compute_scratch_size_per_wave
= 0, compute_waves_wanted
= 0;
4045 uint32_t esgs_ring_size
= 0, gsvs_ring_size
= 0;
4046 bool tess_rings_needed
= false;
4047 bool gds_needed
= false;
4048 bool sample_positions_needed
= false;
4050 for (uint32_t j
= 0; j
< cmd_buffer_count
; j
++) {
4051 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
,
4054 scratch_size_per_wave
= MAX2(scratch_size_per_wave
, cmd_buffer
->scratch_size_per_wave_needed
);
4055 waves_wanted
= MAX2(waves_wanted
, cmd_buffer
->scratch_waves_wanted
);
4056 compute_scratch_size_per_wave
= MAX2(compute_scratch_size_per_wave
,
4057 cmd_buffer
->compute_scratch_size_per_wave_needed
);
4058 compute_waves_wanted
= MAX2(compute_waves_wanted
,
4059 cmd_buffer
->compute_scratch_waves_wanted
);
4060 esgs_ring_size
= MAX2(esgs_ring_size
, cmd_buffer
->esgs_ring_size_needed
);
4061 gsvs_ring_size
= MAX2(gsvs_ring_size
, cmd_buffer
->gsvs_ring_size_needed
);
4062 tess_rings_needed
|= cmd_buffer
->tess_rings_needed
;
4063 gds_needed
|= cmd_buffer
->gds_needed
;
4064 sample_positions_needed
|= cmd_buffer
->sample_positions_needed
;
4067 return radv_get_preamble_cs(queue
, scratch_size_per_wave
, waves_wanted
,
4068 compute_scratch_size_per_wave
, compute_waves_wanted
,
4069 esgs_ring_size
, gsvs_ring_size
, tess_rings_needed
,
4070 gds_needed
, sample_positions_needed
,
4071 initial_full_flush_preamble_cs
,
4072 initial_preamble_cs
, continue_preamble_cs
);
4075 struct radv_deferred_queue_submission
{
4076 struct radv_queue
*queue
;
4077 VkCommandBuffer
*cmd_buffers
;
4078 uint32_t cmd_buffer_count
;
4080 /* Sparse bindings that happen on a queue. */
4081 VkSparseBufferMemoryBindInfo
*buffer_binds
;
4082 uint32_t buffer_bind_count
;
4083 VkSparseImageOpaqueMemoryBindInfo
*image_opaque_binds
;
4084 uint32_t image_opaque_bind_count
;
4087 VkShaderStageFlags wait_dst_stage_mask
;
4088 struct radv_semaphore_part
**wait_semaphores
;
4089 uint32_t wait_semaphore_count
;
4090 struct radv_semaphore_part
**signal_semaphores
;
4091 uint32_t signal_semaphore_count
;
4094 uint64_t *wait_values
;
4095 uint64_t *signal_values
;
4097 struct radv_semaphore_part
*temporary_semaphore_parts
;
4098 uint32_t temporary_semaphore_part_count
;
4100 struct list_head queue_pending_list
;
4101 uint32_t submission_wait_count
;
4102 struct radv_timeline_waiter
*wait_nodes
;
4104 struct list_head processing_list
;
4107 struct radv_queue_submission
{
4108 const VkCommandBuffer
*cmd_buffers
;
4109 uint32_t cmd_buffer_count
;
4111 /* Sparse bindings that happen on a queue. */
4112 const VkSparseBufferMemoryBindInfo
*buffer_binds
;
4113 uint32_t buffer_bind_count
;
4114 const VkSparseImageOpaqueMemoryBindInfo
*image_opaque_binds
;
4115 uint32_t image_opaque_bind_count
;
4118 VkPipelineStageFlags wait_dst_stage_mask
;
4119 const VkSemaphore
*wait_semaphores
;
4120 uint32_t wait_semaphore_count
;
4121 const VkSemaphore
*signal_semaphores
;
4122 uint32_t signal_semaphore_count
;
4125 const uint64_t *wait_values
;
4126 uint32_t wait_value_count
;
4127 const uint64_t *signal_values
;
4128 uint32_t signal_value_count
;
4132 radv_create_deferred_submission(struct radv_queue
*queue
,
4133 const struct radv_queue_submission
*submission
,
4134 struct radv_deferred_queue_submission
**out
)
4136 struct radv_deferred_queue_submission
*deferred
= NULL
;
4137 size_t size
= sizeof(struct radv_deferred_queue_submission
);
4139 uint32_t temporary_count
= 0;
4140 for (uint32_t i
= 0; i
< submission
->wait_semaphore_count
; ++i
) {
4141 RADV_FROM_HANDLE(radv_semaphore
, semaphore
, submission
->wait_semaphores
[i
]);
4142 if (semaphore
->temporary
.kind
!= RADV_SEMAPHORE_NONE
)
4146 size
+= submission
->cmd_buffer_count
* sizeof(VkCommandBuffer
);
4147 size
+= submission
->buffer_bind_count
* sizeof(VkSparseBufferMemoryBindInfo
);
4148 size
+= submission
->image_opaque_bind_count
* sizeof(VkSparseImageOpaqueMemoryBindInfo
);
4149 size
+= submission
->wait_semaphore_count
* sizeof(struct radv_semaphore_part
*);
4150 size
+= temporary_count
* sizeof(struct radv_semaphore_part
);
4151 size
+= submission
->signal_semaphore_count
* sizeof(struct radv_semaphore_part
*);
4152 size
+= submission
->wait_value_count
* sizeof(uint64_t);
4153 size
+= submission
->signal_value_count
* sizeof(uint64_t);
4154 size
+= submission
->wait_semaphore_count
* sizeof(struct radv_timeline_waiter
);
4156 deferred
= calloc(1, size
);
4158 return VK_ERROR_OUT_OF_HOST_MEMORY
;
4160 deferred
->queue
= queue
;
4162 deferred
->cmd_buffers
= (void*)(deferred
+ 1);
4163 deferred
->cmd_buffer_count
= submission
->cmd_buffer_count
;
4164 memcpy(deferred
->cmd_buffers
, submission
->cmd_buffers
,
4165 submission
->cmd_buffer_count
* sizeof(*deferred
->cmd_buffers
));
4167 deferred
->buffer_binds
= (void*)(deferred
->cmd_buffers
+ submission
->cmd_buffer_count
);
4168 deferred
->buffer_bind_count
= submission
->buffer_bind_count
;
4169 memcpy(deferred
->buffer_binds
, submission
->buffer_binds
,
4170 submission
->buffer_bind_count
* sizeof(*deferred
->buffer_binds
));
4172 deferred
->image_opaque_binds
= (void*)(deferred
->buffer_binds
+ submission
->buffer_bind_count
);
4173 deferred
->image_opaque_bind_count
= submission
->image_opaque_bind_count
;
4174 memcpy(deferred
->image_opaque_binds
, submission
->image_opaque_binds
,
4175 submission
->image_opaque_bind_count
* sizeof(*deferred
->image_opaque_binds
));
4177 deferred
->flush_caches
= submission
->flush_caches
;
4178 deferred
->wait_dst_stage_mask
= submission
->wait_dst_stage_mask
;
4180 deferred
->wait_semaphores
= (void*)(deferred
->image_opaque_binds
+ deferred
->image_opaque_bind_count
);
4181 deferred
->wait_semaphore_count
= submission
->wait_semaphore_count
;
4183 deferred
->signal_semaphores
= (void*)(deferred
->wait_semaphores
+ deferred
->wait_semaphore_count
);
4184 deferred
->signal_semaphore_count
= submission
->signal_semaphore_count
;
4186 deferred
->fence
= submission
->fence
;
4188 deferred
->temporary_semaphore_parts
= (void*)(deferred
->signal_semaphores
+ deferred
->signal_semaphore_count
);
4189 deferred
->temporary_semaphore_part_count
= temporary_count
;
4191 uint32_t temporary_idx
= 0;
4192 for (uint32_t i
= 0; i
< submission
->wait_semaphore_count
; ++i
) {
4193 RADV_FROM_HANDLE(radv_semaphore
, semaphore
, submission
->wait_semaphores
[i
]);
4194 if (semaphore
->temporary
.kind
!= RADV_SEMAPHORE_NONE
) {
4195 deferred
->wait_semaphores
[i
] = &deferred
->temporary_semaphore_parts
[temporary_idx
];
4196 deferred
->temporary_semaphore_parts
[temporary_idx
] = semaphore
->temporary
;
4197 semaphore
->temporary
.kind
= RADV_SEMAPHORE_NONE
;
4200 deferred
->wait_semaphores
[i
] = &semaphore
->permanent
;
4203 for (uint32_t i
= 0; i
< submission
->signal_semaphore_count
; ++i
) {
4204 RADV_FROM_HANDLE(radv_semaphore
, semaphore
, submission
->signal_semaphores
[i
]);
4205 if (semaphore
->temporary
.kind
!= RADV_SEMAPHORE_NONE
) {
4206 deferred
->signal_semaphores
[i
] = &semaphore
->temporary
;
4208 deferred
->signal_semaphores
[i
] = &semaphore
->permanent
;
4212 deferred
->wait_values
= (void*)(deferred
->temporary_semaphore_parts
+ temporary_count
);
4213 memcpy(deferred
->wait_values
, submission
->wait_values
, submission
->wait_value_count
* sizeof(uint64_t));
4214 deferred
->signal_values
= deferred
->wait_values
+ submission
->wait_value_count
;
4215 memcpy(deferred
->signal_values
, submission
->signal_values
, submission
->signal_value_count
* sizeof(uint64_t));
4217 deferred
->wait_nodes
= (void*)(deferred
->signal_values
+ submission
->signal_value_count
);
4218 /* This is worst-case. radv_queue_enqueue_submission will fill in further, but this
4219 * ensure the submission is not accidentally triggered early when adding wait timelines. */
4220 deferred
->submission_wait_count
= 1 + submission
->wait_semaphore_count
;
4227 radv_queue_enqueue_submission(struct radv_deferred_queue_submission
*submission
,
4228 struct list_head
*processing_list
)
4230 uint32_t wait_cnt
= 0;
4231 struct radv_timeline_waiter
*waiter
= submission
->wait_nodes
;
4232 for (uint32_t i
= 0; i
< submission
->wait_semaphore_count
; ++i
) {
4233 if (submission
->wait_semaphores
[i
]->kind
== RADV_SEMAPHORE_TIMELINE
) {
4234 pthread_mutex_lock(&submission
->wait_semaphores
[i
]->timeline
.mutex
);
4235 if (submission
->wait_semaphores
[i
]->timeline
.highest_submitted
< submission
->wait_values
[i
]) {
4237 waiter
->value
= submission
->wait_values
[i
];
4238 waiter
->submission
= submission
;
4239 list_addtail(&waiter
->list
, &submission
->wait_semaphores
[i
]->timeline
.waiters
);
4242 pthread_mutex_unlock(&submission
->wait_semaphores
[i
]->timeline
.mutex
);
4246 pthread_mutex_lock(&submission
->queue
->pending_mutex
);
4248 bool is_first
= list_is_empty(&submission
->queue
->pending_submissions
);
4249 list_addtail(&submission
->queue_pending_list
, &submission
->queue
->pending_submissions
);
4251 pthread_mutex_unlock(&submission
->queue
->pending_mutex
);
4253 /* If there is already a submission in the queue, that will decrement the counter by 1 when
4254 * submitted, but if the queue was empty, we decrement ourselves as there is no previous
4256 uint32_t decrement
= submission
->wait_semaphore_count
- wait_cnt
+ (is_first
? 1 : 0);
4257 if (__atomic_sub_fetch(&submission
->submission_wait_count
, decrement
, __ATOMIC_ACQ_REL
) == 0) {
4258 list_addtail(&submission
->processing_list
, processing_list
);
4263 radv_queue_submission_update_queue(struct radv_deferred_queue_submission
*submission
,
4264 struct list_head
*processing_list
)
4266 pthread_mutex_lock(&submission
->queue
->pending_mutex
);
4267 list_del(&submission
->queue_pending_list
);
4269 /* trigger the next submission in the queue. */
4270 if (!list_is_empty(&submission
->queue
->pending_submissions
)) {
4271 struct radv_deferred_queue_submission
*next_submission
=
4272 list_first_entry(&submission
->queue
->pending_submissions
,
4273 struct radv_deferred_queue_submission
,
4274 queue_pending_list
);
4275 if (p_atomic_dec_zero(&next_submission
->submission_wait_count
)) {
4276 list_addtail(&next_submission
->processing_list
, processing_list
);
4279 pthread_mutex_unlock(&submission
->queue
->pending_mutex
);
4281 pthread_cond_broadcast(&submission
->queue
->device
->timeline_cond
);
4285 radv_queue_submit_deferred(struct radv_deferred_queue_submission
*submission
,
4286 struct list_head
*processing_list
)
4288 RADV_FROM_HANDLE(radv_fence
, fence
, submission
->fence
);
4289 struct radv_queue
*queue
= submission
->queue
;
4290 struct radeon_winsys_ctx
*ctx
= queue
->hw_ctx
;
4291 uint32_t max_cs_submission
= queue
->device
->trace_bo
? 1 : RADV_MAX_IBS_PER_SUBMIT
;
4292 struct radeon_winsys_fence
*base_fence
= fence
? fence
->fence
: NULL
;
4293 bool do_flush
= submission
->flush_caches
|| submission
->wait_dst_stage_mask
;
4294 bool can_patch
= true;
4296 struct radv_winsys_sem_info sem_info
;
4299 struct radeon_cmdbuf
*initial_preamble_cs
= NULL
;
4300 struct radeon_cmdbuf
*initial_flush_preamble_cs
= NULL
;
4301 struct radeon_cmdbuf
*continue_preamble_cs
= NULL
;
4303 result
= radv_get_preambles(queue
, submission
->cmd_buffers
,
4304 submission
->cmd_buffer_count
,
4305 &initial_preamble_cs
,
4306 &initial_flush_preamble_cs
,
4307 &continue_preamble_cs
);
4308 if (result
!= VK_SUCCESS
)
4311 result
= radv_alloc_sem_info(queue
->device
,
4313 submission
->wait_semaphore_count
,
4314 submission
->wait_semaphores
,
4315 submission
->wait_values
,
4316 submission
->signal_semaphore_count
,
4317 submission
->signal_semaphores
,
4318 submission
->signal_values
,
4320 if (result
!= VK_SUCCESS
)
4323 for (uint32_t i
= 0; i
< submission
->buffer_bind_count
; ++i
) {
4324 radv_sparse_buffer_bind_memory(queue
->device
,
4325 submission
->buffer_binds
+ i
);
4328 for (uint32_t i
= 0; i
< submission
->image_opaque_bind_count
; ++i
) {
4329 radv_sparse_image_opaque_bind_memory(queue
->device
,
4330 submission
->image_opaque_binds
+ i
);
4333 if (!submission
->cmd_buffer_count
) {
4334 ret
= queue
->device
->ws
->cs_submit(ctx
, queue
->queue_idx
,
4335 &queue
->device
->empty_cs
[queue
->queue_family_index
],
4340 radv_loge("failed to submit CS\n");
4346 struct radeon_cmdbuf
**cs_array
= malloc(sizeof(struct radeon_cmdbuf
*) *
4347 (submission
->cmd_buffer_count
));
4349 for (uint32_t j
= 0; j
< submission
->cmd_buffer_count
; j
++) {
4350 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, submission
->cmd_buffers
[j
]);
4351 assert(cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_PRIMARY
);
4353 cs_array
[j
] = cmd_buffer
->cs
;
4354 if ((cmd_buffer
->usage_flags
& VK_COMMAND_BUFFER_USAGE_SIMULTANEOUS_USE_BIT
))
4357 cmd_buffer
->status
= RADV_CMD_BUFFER_STATUS_PENDING
;
4360 for (uint32_t j
= 0; j
< submission
->cmd_buffer_count
; j
+= advance
) {
4361 struct radeon_cmdbuf
*initial_preamble
= (do_flush
&& !j
) ? initial_flush_preamble_cs
: initial_preamble_cs
;
4362 const struct radv_winsys_bo_list
*bo_list
= NULL
;
4364 advance
= MIN2(max_cs_submission
,
4365 submission
->cmd_buffer_count
- j
);
4367 if (queue
->device
->trace_bo
)
4368 *queue
->device
->trace_id_ptr
= 0;
4370 sem_info
.cs_emit_wait
= j
== 0;
4371 sem_info
.cs_emit_signal
= j
+ advance
== submission
->cmd_buffer_count
;
4373 if (unlikely(queue
->device
->use_global_bo_list
)) {
4374 pthread_mutex_lock(&queue
->device
->bo_list
.mutex
);
4375 bo_list
= &queue
->device
->bo_list
.list
;
4378 ret
= queue
->device
->ws
->cs_submit(ctx
, queue
->queue_idx
, cs_array
+ j
,
4379 advance
, initial_preamble
, continue_preamble_cs
,
4381 can_patch
, base_fence
);
4383 if (unlikely(queue
->device
->use_global_bo_list
))
4384 pthread_mutex_unlock(&queue
->device
->bo_list
.mutex
);
4387 radv_loge("failed to submit CS\n");
4390 if (queue
->device
->trace_bo
) {
4391 radv_check_gpu_hangs(queue
, cs_array
[j
]);
4399 radv_free_temp_syncobjs(queue
->device
,
4400 submission
->temporary_semaphore_part_count
,
4401 submission
->temporary_semaphore_parts
);
4402 radv_finalize_timelines(queue
->device
,
4403 submission
->wait_semaphore_count
,
4404 submission
->wait_semaphores
,
4405 submission
->wait_values
,
4406 submission
->signal_semaphore_count
,
4407 submission
->signal_semaphores
,
4408 submission
->signal_values
,
4410 /* Has to happen after timeline finalization to make sure the
4411 * condition variable is only triggered when timelines and queue have
4413 radv_queue_submission_update_queue(submission
, processing_list
);
4414 radv_free_sem_info(&sem_info
);
4419 radv_free_temp_syncobjs(queue
->device
,
4420 submission
->temporary_semaphore_part_count
,
4421 submission
->temporary_semaphore_parts
);
4423 return VK_ERROR_DEVICE_LOST
;
4427 radv_process_submissions(struct list_head
*processing_list
)
4429 while(!list_is_empty(processing_list
)) {
4430 struct radv_deferred_queue_submission
*submission
=
4431 list_first_entry(processing_list
, struct radv_deferred_queue_submission
, processing_list
);
4432 list_del(&submission
->processing_list
);
4434 VkResult result
= radv_queue_submit_deferred(submission
, processing_list
);
4435 if (result
!= VK_SUCCESS
)
4441 static VkResult
radv_queue_submit(struct radv_queue
*queue
,
4442 const struct radv_queue_submission
*submission
)
4444 struct radv_deferred_queue_submission
*deferred
= NULL
;
4446 VkResult result
= radv_create_deferred_submission(queue
, submission
, &deferred
);
4447 if (result
!= VK_SUCCESS
)
4450 struct list_head processing_list
;
4451 list_inithead(&processing_list
);
4453 radv_queue_enqueue_submission(deferred
, &processing_list
);
4454 return radv_process_submissions(&processing_list
);
4457 /* Signals fence as soon as all the work currently put on queue is done. */
4458 static VkResult
radv_signal_fence(struct radv_queue
*queue
,
4461 return radv_queue_submit(queue
, &(struct radv_queue_submission
) {
4466 static bool radv_submit_has_effects(const VkSubmitInfo
*info
)
4468 return info
->commandBufferCount
||
4469 info
->waitSemaphoreCount
||
4470 info
->signalSemaphoreCount
;
4473 VkResult
radv_QueueSubmit(
4475 uint32_t submitCount
,
4476 const VkSubmitInfo
* pSubmits
,
4479 RADV_FROM_HANDLE(radv_queue
, queue
, _queue
);
4481 uint32_t fence_idx
= 0;
4482 bool flushed_caches
= false;
4484 if (fence
!= VK_NULL_HANDLE
) {
4485 for (uint32_t i
= 0; i
< submitCount
; ++i
)
4486 if (radv_submit_has_effects(pSubmits
+ i
))
4489 fence_idx
= UINT32_MAX
;
4491 for (uint32_t i
= 0; i
< submitCount
; i
++) {
4492 if (!radv_submit_has_effects(pSubmits
+ i
) && fence_idx
!= i
)
4495 VkPipelineStageFlags wait_dst_stage_mask
= 0;
4496 for (unsigned j
= 0; j
< pSubmits
[i
].waitSemaphoreCount
; ++j
) {
4497 wait_dst_stage_mask
|= pSubmits
[i
].pWaitDstStageMask
[j
];
4500 const VkTimelineSemaphoreSubmitInfoKHR
*timeline_info
=
4501 vk_find_struct_const(pSubmits
[i
].pNext
, TIMELINE_SEMAPHORE_SUBMIT_INFO_KHR
);
4503 result
= radv_queue_submit(queue
, &(struct radv_queue_submission
) {
4504 .cmd_buffers
= pSubmits
[i
].pCommandBuffers
,
4505 .cmd_buffer_count
= pSubmits
[i
].commandBufferCount
,
4506 .wait_dst_stage_mask
= wait_dst_stage_mask
,
4507 .flush_caches
= !flushed_caches
,
4508 .wait_semaphores
= pSubmits
[i
].pWaitSemaphores
,
4509 .wait_semaphore_count
= pSubmits
[i
].waitSemaphoreCount
,
4510 .signal_semaphores
= pSubmits
[i
].pSignalSemaphores
,
4511 .signal_semaphore_count
= pSubmits
[i
].signalSemaphoreCount
,
4512 .fence
= i
== fence_idx
? fence
: VK_NULL_HANDLE
,
4513 .wait_values
= timeline_info
? timeline_info
->pWaitSemaphoreValues
: NULL
,
4514 .wait_value_count
= timeline_info
&& timeline_info
->pWaitSemaphoreValues
? timeline_info
->waitSemaphoreValueCount
: 0,
4515 .signal_values
= timeline_info
? timeline_info
->pSignalSemaphoreValues
: NULL
,
4516 .signal_value_count
= timeline_info
&& timeline_info
->pSignalSemaphoreValues
? timeline_info
->signalSemaphoreValueCount
: 0,
4518 if (result
!= VK_SUCCESS
)
4521 flushed_caches
= true;
4524 if (fence
!= VK_NULL_HANDLE
&& !submitCount
) {
4525 result
= radv_signal_fence(queue
, fence
);
4526 if (result
!= VK_SUCCESS
)
4533 VkResult
radv_QueueWaitIdle(
4536 RADV_FROM_HANDLE(radv_queue
, queue
, _queue
);
4538 pthread_mutex_lock(&queue
->pending_mutex
);
4539 while (!list_is_empty(&queue
->pending_submissions
)) {
4540 pthread_cond_wait(&queue
->device
->timeline_cond
, &queue
->pending_mutex
);
4542 pthread_mutex_unlock(&queue
->pending_mutex
);
4544 queue
->device
->ws
->ctx_wait_idle(queue
->hw_ctx
,
4545 radv_queue_family_to_ring(queue
->queue_family_index
),
4550 VkResult
radv_DeviceWaitIdle(
4553 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4555 for (unsigned i
= 0; i
< RADV_MAX_QUEUE_FAMILIES
; i
++) {
4556 for (unsigned q
= 0; q
< device
->queue_count
[i
]; q
++) {
4557 radv_QueueWaitIdle(radv_queue_to_handle(&device
->queues
[i
][q
]));
4563 VkResult
radv_EnumerateInstanceExtensionProperties(
4564 const char* pLayerName
,
4565 uint32_t* pPropertyCount
,
4566 VkExtensionProperties
* pProperties
)
4568 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
4570 for (int i
= 0; i
< RADV_INSTANCE_EXTENSION_COUNT
; i
++) {
4571 if (radv_supported_instance_extensions
.extensions
[i
]) {
4572 vk_outarray_append(&out
, prop
) {
4573 *prop
= radv_instance_extensions
[i
];
4578 return vk_outarray_status(&out
);
4581 VkResult
radv_EnumerateDeviceExtensionProperties(
4582 VkPhysicalDevice physicalDevice
,
4583 const char* pLayerName
,
4584 uint32_t* pPropertyCount
,
4585 VkExtensionProperties
* pProperties
)
4587 RADV_FROM_HANDLE(radv_physical_device
, device
, physicalDevice
);
4588 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
4590 for (int i
= 0; i
< RADV_DEVICE_EXTENSION_COUNT
; i
++) {
4591 if (device
->supported_extensions
.extensions
[i
]) {
4592 vk_outarray_append(&out
, prop
) {
4593 *prop
= radv_device_extensions
[i
];
4598 return vk_outarray_status(&out
);
4601 PFN_vkVoidFunction
radv_GetInstanceProcAddr(
4602 VkInstance _instance
,
4605 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
4606 bool unchecked
= instance
? instance
->debug_flags
& RADV_DEBUG_ALL_ENTRYPOINTS
: false;
4609 return radv_lookup_entrypoint_unchecked(pName
);
4611 return radv_lookup_entrypoint_checked(pName
,
4612 instance
? instance
->apiVersion
: 0,
4613 instance
? &instance
->enabled_extensions
: NULL
,
4618 /* The loader wants us to expose a second GetInstanceProcAddr function
4619 * to work around certain LD_PRELOAD issues seen in apps.
4622 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
4623 VkInstance instance
,
4627 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
4628 VkInstance instance
,
4631 return radv_GetInstanceProcAddr(instance
, pName
);
4635 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetPhysicalDeviceProcAddr(
4636 VkInstance _instance
,
4640 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetPhysicalDeviceProcAddr(
4641 VkInstance _instance
,
4644 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
4646 return radv_lookup_physical_device_entrypoint_checked(pName
,
4647 instance
? instance
->apiVersion
: 0,
4648 instance
? &instance
->enabled_extensions
: NULL
);
4651 PFN_vkVoidFunction
radv_GetDeviceProcAddr(
4655 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4656 bool unchecked
= device
? device
->instance
->debug_flags
& RADV_DEBUG_ALL_ENTRYPOINTS
: false;
4659 return radv_lookup_entrypoint_unchecked(pName
);
4661 return radv_lookup_entrypoint_checked(pName
,
4662 device
->instance
->apiVersion
,
4663 &device
->instance
->enabled_extensions
,
4664 &device
->enabled_extensions
);
4668 bool radv_get_memory_fd(struct radv_device
*device
,
4669 struct radv_device_memory
*memory
,
4672 struct radeon_bo_metadata metadata
;
4674 if (memory
->image
) {
4675 radv_init_metadata(device
, memory
->image
, &metadata
);
4676 device
->ws
->buffer_set_metadata(memory
->bo
, &metadata
);
4679 return device
->ws
->buffer_get_fd(device
->ws
, memory
->bo
,
4684 static void radv_free_memory(struct radv_device
*device
,
4685 const VkAllocationCallbacks
* pAllocator
,
4686 struct radv_device_memory
*mem
)
4691 #if RADV_SUPPORT_ANDROID_HARDWARE_BUFFER
4692 if (mem
->android_hardware_buffer
)
4693 AHardwareBuffer_release(mem
->android_hardware_buffer
);
4697 radv_bo_list_remove(device
, mem
->bo
);
4698 device
->ws
->buffer_destroy(mem
->bo
);
4702 vk_free2(&device
->alloc
, pAllocator
, mem
);
4705 static VkResult
radv_alloc_memory(struct radv_device
*device
,
4706 const VkMemoryAllocateInfo
* pAllocateInfo
,
4707 const VkAllocationCallbacks
* pAllocator
,
4708 VkDeviceMemory
* pMem
)
4710 struct radv_device_memory
*mem
;
4712 enum radeon_bo_domain domain
;
4714 enum radv_mem_type mem_type_index
= device
->physical_device
->mem_type_indices
[pAllocateInfo
->memoryTypeIndex
];
4716 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
4718 const VkImportMemoryFdInfoKHR
*import_info
=
4719 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_FD_INFO_KHR
);
4720 const VkMemoryDedicatedAllocateInfo
*dedicate_info
=
4721 vk_find_struct_const(pAllocateInfo
->pNext
, MEMORY_DEDICATED_ALLOCATE_INFO
);
4722 const VkExportMemoryAllocateInfo
*export_info
=
4723 vk_find_struct_const(pAllocateInfo
->pNext
, EXPORT_MEMORY_ALLOCATE_INFO
);
4724 const struct VkImportAndroidHardwareBufferInfoANDROID
*ahb_import_info
=
4725 vk_find_struct_const(pAllocateInfo
->pNext
,
4726 IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID
);
4727 const VkImportMemoryHostPointerInfoEXT
*host_ptr_info
=
4728 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_HOST_POINTER_INFO_EXT
);
4730 const struct wsi_memory_allocate_info
*wsi_info
=
4731 vk_find_struct_const(pAllocateInfo
->pNext
, WSI_MEMORY_ALLOCATE_INFO_MESA
);
4733 if (pAllocateInfo
->allocationSize
== 0 && !ahb_import_info
&&
4734 !(export_info
&& (export_info
->handleTypes
& VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID
))) {
4735 /* Apparently, this is allowed */
4736 *pMem
= VK_NULL_HANDLE
;
4740 mem
= vk_zalloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
4741 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
4743 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
4745 if (wsi_info
&& wsi_info
->implicit_sync
)
4746 flags
|= RADEON_FLAG_IMPLICIT_SYNC
;
4748 if (dedicate_info
) {
4749 mem
->image
= radv_image_from_handle(dedicate_info
->image
);
4750 mem
->buffer
= radv_buffer_from_handle(dedicate_info
->buffer
);
4756 float priority_float
= 0.5;
4757 const struct VkMemoryPriorityAllocateInfoEXT
*priority_ext
=
4758 vk_find_struct_const(pAllocateInfo
->pNext
,
4759 MEMORY_PRIORITY_ALLOCATE_INFO_EXT
);
4761 priority_float
= priority_ext
->priority
;
4763 unsigned priority
= MIN2(RADV_BO_PRIORITY_APPLICATION_MAX
- 1,
4764 (int)(priority_float
* RADV_BO_PRIORITY_APPLICATION_MAX
));
4766 mem
->user_ptr
= NULL
;
4769 #if RADV_SUPPORT_ANDROID_HARDWARE_BUFFER
4770 mem
->android_hardware_buffer
= NULL
;
4773 if (ahb_import_info
) {
4774 result
= radv_import_ahb_memory(device
, mem
, priority
, ahb_import_info
);
4775 if (result
!= VK_SUCCESS
)
4777 } else if(export_info
&& (export_info
->handleTypes
& VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID
)) {
4778 result
= radv_create_ahb_memory(device
, mem
, priority
, pAllocateInfo
);
4779 if (result
!= VK_SUCCESS
)
4781 } else if (import_info
) {
4782 assert(import_info
->handleType
==
4783 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
4784 import_info
->handleType
==
4785 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
4786 mem
->bo
= device
->ws
->buffer_from_fd(device
->ws
, import_info
->fd
,
4789 result
= VK_ERROR_INVALID_EXTERNAL_HANDLE
;
4792 close(import_info
->fd
);
4794 } else if (host_ptr_info
) {
4795 assert(host_ptr_info
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
);
4796 assert(radv_is_mem_type_gtt_cached(mem_type_index
));
4797 mem
->bo
= device
->ws
->buffer_from_ptr(device
->ws
, host_ptr_info
->pHostPointer
,
4798 pAllocateInfo
->allocationSize
,
4801 result
= VK_ERROR_INVALID_EXTERNAL_HANDLE
;
4804 mem
->user_ptr
= host_ptr_info
->pHostPointer
;
4807 uint64_t alloc_size
= align_u64(pAllocateInfo
->allocationSize
, 4096);
4808 if (radv_is_mem_type_gtt_wc(mem_type_index
) ||
4809 radv_is_mem_type_gtt_cached(mem_type_index
))
4810 domain
= RADEON_DOMAIN_GTT
;
4812 domain
= RADEON_DOMAIN_VRAM
;
4814 if (radv_is_mem_type_vram(mem_type_index
))
4815 flags
|= RADEON_FLAG_NO_CPU_ACCESS
;
4817 flags
|= RADEON_FLAG_CPU_ACCESS
;
4819 if (radv_is_mem_type_gtt_wc(mem_type_index
))
4820 flags
|= RADEON_FLAG_GTT_WC
;
4822 if (!dedicate_info
&& !import_info
&& (!export_info
|| !export_info
->handleTypes
)) {
4823 flags
|= RADEON_FLAG_NO_INTERPROCESS_SHARING
;
4824 if (device
->use_global_bo_list
) {
4825 flags
|= RADEON_FLAG_PREFER_LOCAL_BO
;
4829 if (radv_is_mem_type_uncached(mem_type_index
)) {
4830 assert(device
->physical_device
->rad_info
.has_l2_uncached
);
4831 flags
|= RADEON_FLAG_VA_UNCACHED
;
4834 mem
->bo
= device
->ws
->buffer_create(device
->ws
, alloc_size
, device
->physical_device
->rad_info
.max_alignment
,
4835 domain
, flags
, priority
);
4838 result
= VK_ERROR_OUT_OF_DEVICE_MEMORY
;
4841 mem
->type_index
= mem_type_index
;
4844 result
= radv_bo_list_add(device
, mem
->bo
);
4845 if (result
!= VK_SUCCESS
)
4848 *pMem
= radv_device_memory_to_handle(mem
);
4853 radv_free_memory(device
, pAllocator
,mem
);
4854 vk_free2(&device
->alloc
, pAllocator
, mem
);
4859 VkResult
radv_AllocateMemory(
4861 const VkMemoryAllocateInfo
* pAllocateInfo
,
4862 const VkAllocationCallbacks
* pAllocator
,
4863 VkDeviceMemory
* pMem
)
4865 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4866 return radv_alloc_memory(device
, pAllocateInfo
, pAllocator
, pMem
);
4869 void radv_FreeMemory(
4871 VkDeviceMemory _mem
,
4872 const VkAllocationCallbacks
* pAllocator
)
4874 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4875 RADV_FROM_HANDLE(radv_device_memory
, mem
, _mem
);
4877 radv_free_memory(device
, pAllocator
, mem
);
4880 VkResult
radv_MapMemory(
4882 VkDeviceMemory _memory
,
4883 VkDeviceSize offset
,
4885 VkMemoryMapFlags flags
,
4888 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4889 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
4897 *ppData
= mem
->user_ptr
;
4899 *ppData
= device
->ws
->buffer_map(mem
->bo
);
4906 return vk_error(device
->instance
, VK_ERROR_MEMORY_MAP_FAILED
);
4909 void radv_UnmapMemory(
4911 VkDeviceMemory _memory
)
4913 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4914 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
4919 if (mem
->user_ptr
== NULL
)
4920 device
->ws
->buffer_unmap(mem
->bo
);
4923 VkResult
radv_FlushMappedMemoryRanges(
4925 uint32_t memoryRangeCount
,
4926 const VkMappedMemoryRange
* pMemoryRanges
)
4931 VkResult
radv_InvalidateMappedMemoryRanges(
4933 uint32_t memoryRangeCount
,
4934 const VkMappedMemoryRange
* pMemoryRanges
)
4939 void radv_GetBufferMemoryRequirements(
4942 VkMemoryRequirements
* pMemoryRequirements
)
4944 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4945 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
4947 pMemoryRequirements
->memoryTypeBits
= (1u << device
->physical_device
->memory_properties
.memoryTypeCount
) - 1;
4949 if (buffer
->flags
& VK_BUFFER_CREATE_SPARSE_BINDING_BIT
)
4950 pMemoryRequirements
->alignment
= 4096;
4952 pMemoryRequirements
->alignment
= 16;
4954 pMemoryRequirements
->size
= align64(buffer
->size
, pMemoryRequirements
->alignment
);
4957 void radv_GetBufferMemoryRequirements2(
4959 const VkBufferMemoryRequirementsInfo2
*pInfo
,
4960 VkMemoryRequirements2
*pMemoryRequirements
)
4962 radv_GetBufferMemoryRequirements(device
, pInfo
->buffer
,
4963 &pMemoryRequirements
->memoryRequirements
);
4964 RADV_FROM_HANDLE(radv_buffer
, buffer
, pInfo
->buffer
);
4965 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
4966 switch (ext
->sType
) {
4967 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
4968 VkMemoryDedicatedRequirements
*req
=
4969 (VkMemoryDedicatedRequirements
*) ext
;
4970 req
->requiresDedicatedAllocation
= buffer
->shareable
;
4971 req
->prefersDedicatedAllocation
= req
->requiresDedicatedAllocation
;
4980 void radv_GetImageMemoryRequirements(
4983 VkMemoryRequirements
* pMemoryRequirements
)
4985 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4986 RADV_FROM_HANDLE(radv_image
, image
, _image
);
4988 pMemoryRequirements
->memoryTypeBits
= (1u << device
->physical_device
->memory_properties
.memoryTypeCount
) - 1;
4990 pMemoryRequirements
->size
= image
->size
;
4991 pMemoryRequirements
->alignment
= image
->alignment
;
4994 void radv_GetImageMemoryRequirements2(
4996 const VkImageMemoryRequirementsInfo2
*pInfo
,
4997 VkMemoryRequirements2
*pMemoryRequirements
)
4999 radv_GetImageMemoryRequirements(device
, pInfo
->image
,
5000 &pMemoryRequirements
->memoryRequirements
);
5002 RADV_FROM_HANDLE(radv_image
, image
, pInfo
->image
);
5004 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
5005 switch (ext
->sType
) {
5006 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
5007 VkMemoryDedicatedRequirements
*req
=
5008 (VkMemoryDedicatedRequirements
*) ext
;
5009 req
->requiresDedicatedAllocation
= image
->shareable
;
5010 req
->prefersDedicatedAllocation
= req
->requiresDedicatedAllocation
;
5019 void radv_GetImageSparseMemoryRequirements(
5022 uint32_t* pSparseMemoryRequirementCount
,
5023 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
5028 void radv_GetImageSparseMemoryRequirements2(
5030 const VkImageSparseMemoryRequirementsInfo2
*pInfo
,
5031 uint32_t* pSparseMemoryRequirementCount
,
5032 VkSparseImageMemoryRequirements2
*pSparseMemoryRequirements
)
5037 void radv_GetDeviceMemoryCommitment(
5039 VkDeviceMemory memory
,
5040 VkDeviceSize
* pCommittedMemoryInBytes
)
5042 *pCommittedMemoryInBytes
= 0;
5045 VkResult
radv_BindBufferMemory2(VkDevice device
,
5046 uint32_t bindInfoCount
,
5047 const VkBindBufferMemoryInfo
*pBindInfos
)
5049 for (uint32_t i
= 0; i
< bindInfoCount
; ++i
) {
5050 RADV_FROM_HANDLE(radv_device_memory
, mem
, pBindInfos
[i
].memory
);
5051 RADV_FROM_HANDLE(radv_buffer
, buffer
, pBindInfos
[i
].buffer
);
5054 buffer
->bo
= mem
->bo
;
5055 buffer
->offset
= pBindInfos
[i
].memoryOffset
;
5063 VkResult
radv_BindBufferMemory(
5066 VkDeviceMemory memory
,
5067 VkDeviceSize memoryOffset
)
5069 const VkBindBufferMemoryInfo info
= {
5070 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
,
5073 .memoryOffset
= memoryOffset
5076 return radv_BindBufferMemory2(device
, 1, &info
);
5079 VkResult
radv_BindImageMemory2(VkDevice device
,
5080 uint32_t bindInfoCount
,
5081 const VkBindImageMemoryInfo
*pBindInfos
)
5083 for (uint32_t i
= 0; i
< bindInfoCount
; ++i
) {
5084 RADV_FROM_HANDLE(radv_device_memory
, mem
, pBindInfos
[i
].memory
);
5085 RADV_FROM_HANDLE(radv_image
, image
, pBindInfos
[i
].image
);
5088 image
->bo
= mem
->bo
;
5089 image
->offset
= pBindInfos
[i
].memoryOffset
;
5099 VkResult
radv_BindImageMemory(
5102 VkDeviceMemory memory
,
5103 VkDeviceSize memoryOffset
)
5105 const VkBindImageMemoryInfo info
= {
5106 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
,
5109 .memoryOffset
= memoryOffset
5112 return radv_BindImageMemory2(device
, 1, &info
);
5115 static bool radv_sparse_bind_has_effects(const VkBindSparseInfo
*info
)
5117 return info
->bufferBindCount
||
5118 info
->imageOpaqueBindCount
||
5119 info
->imageBindCount
||
5120 info
->waitSemaphoreCount
||
5121 info
->signalSemaphoreCount
;
5124 VkResult
radv_QueueBindSparse(
5126 uint32_t bindInfoCount
,
5127 const VkBindSparseInfo
* pBindInfo
,
5130 RADV_FROM_HANDLE(radv_queue
, queue
, _queue
);
5132 uint32_t fence_idx
= 0;
5134 if (fence
!= VK_NULL_HANDLE
) {
5135 for (uint32_t i
= 0; i
< bindInfoCount
; ++i
)
5136 if (radv_sparse_bind_has_effects(pBindInfo
+ i
))
5139 fence_idx
= UINT32_MAX
;
5141 for (uint32_t i
= 0; i
< bindInfoCount
; ++i
) {
5142 if (i
!= fence_idx
&& !radv_sparse_bind_has_effects(pBindInfo
+ i
))
5145 const VkTimelineSemaphoreSubmitInfoKHR
*timeline_info
=
5146 vk_find_struct_const(pBindInfo
[i
].pNext
, TIMELINE_SEMAPHORE_SUBMIT_INFO_KHR
);
5148 VkResult result
= radv_queue_submit(queue
, &(struct radv_queue_submission
) {
5149 .buffer_binds
= pBindInfo
[i
].pBufferBinds
,
5150 .buffer_bind_count
= pBindInfo
[i
].bufferBindCount
,
5151 .image_opaque_binds
= pBindInfo
[i
].pImageOpaqueBinds
,
5152 .image_opaque_bind_count
= pBindInfo
[i
].imageOpaqueBindCount
,
5153 .wait_semaphores
= pBindInfo
[i
].pWaitSemaphores
,
5154 .wait_semaphore_count
= pBindInfo
[i
].waitSemaphoreCount
,
5155 .signal_semaphores
= pBindInfo
[i
].pSignalSemaphores
,
5156 .signal_semaphore_count
= pBindInfo
[i
].signalSemaphoreCount
,
5157 .fence
= i
== fence_idx
? fence
: VK_NULL_HANDLE
,
5158 .wait_values
= timeline_info
? timeline_info
->pWaitSemaphoreValues
: NULL
,
5159 .wait_value_count
= timeline_info
&& timeline_info
->pWaitSemaphoreValues
? timeline_info
->waitSemaphoreValueCount
: 0,
5160 .signal_values
= timeline_info
? timeline_info
->pSignalSemaphoreValues
: NULL
,
5161 .signal_value_count
= timeline_info
&& timeline_info
->pSignalSemaphoreValues
? timeline_info
->signalSemaphoreValueCount
: 0,
5164 if (result
!= VK_SUCCESS
)
5168 if (fence
!= VK_NULL_HANDLE
&& !bindInfoCount
) {
5169 result
= radv_signal_fence(queue
, fence
);
5170 if (result
!= VK_SUCCESS
)
5177 VkResult
radv_CreateFence(
5179 const VkFenceCreateInfo
* pCreateInfo
,
5180 const VkAllocationCallbacks
* pAllocator
,
5183 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5184 const VkExportFenceCreateInfo
*export
=
5185 vk_find_struct_const(pCreateInfo
->pNext
, EXPORT_FENCE_CREATE_INFO
);
5186 VkExternalFenceHandleTypeFlags handleTypes
=
5187 export
? export
->handleTypes
: 0;
5189 struct radv_fence
*fence
= vk_alloc2(&device
->alloc
, pAllocator
,
5191 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
5194 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
5196 fence
->fence_wsi
= NULL
;
5197 fence
->temp_syncobj
= 0;
5198 if (device
->always_use_syncobj
|| handleTypes
) {
5199 int ret
= device
->ws
->create_syncobj(device
->ws
, &fence
->syncobj
);
5201 vk_free2(&device
->alloc
, pAllocator
, fence
);
5202 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
5204 if (pCreateInfo
->flags
& VK_FENCE_CREATE_SIGNALED_BIT
) {
5205 device
->ws
->signal_syncobj(device
->ws
, fence
->syncobj
);
5207 fence
->fence
= NULL
;
5209 fence
->fence
= device
->ws
->create_fence();
5210 if (!fence
->fence
) {
5211 vk_free2(&device
->alloc
, pAllocator
, fence
);
5212 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
5215 if (pCreateInfo
->flags
& VK_FENCE_CREATE_SIGNALED_BIT
)
5216 device
->ws
->signal_fence(fence
->fence
);
5219 *pFence
= radv_fence_to_handle(fence
);
5224 void radv_DestroyFence(
5227 const VkAllocationCallbacks
* pAllocator
)
5229 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5230 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
5235 if (fence
->temp_syncobj
)
5236 device
->ws
->destroy_syncobj(device
->ws
, fence
->temp_syncobj
);
5238 device
->ws
->destroy_syncobj(device
->ws
, fence
->syncobj
);
5240 device
->ws
->destroy_fence(fence
->fence
);
5241 if (fence
->fence_wsi
)
5242 fence
->fence_wsi
->destroy(fence
->fence_wsi
);
5243 vk_free2(&device
->alloc
, pAllocator
, fence
);
5247 uint64_t radv_get_current_time(void)
5250 clock_gettime(CLOCK_MONOTONIC
, &tv
);
5251 return tv
.tv_nsec
+ tv
.tv_sec
*1000000000ull;
5254 static uint64_t radv_get_absolute_timeout(uint64_t timeout
)
5256 uint64_t current_time
= radv_get_current_time();
5258 timeout
= MIN2(UINT64_MAX
- current_time
, timeout
);
5260 return current_time
+ timeout
;
5264 static bool radv_all_fences_plain_and_submitted(struct radv_device
*device
,
5265 uint32_t fenceCount
, const VkFence
*pFences
)
5267 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
5268 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
5269 if (fence
->fence
== NULL
|| fence
->syncobj
||
5270 fence
->temp_syncobj
|| fence
->fence_wsi
||
5271 (!device
->ws
->is_fence_waitable(fence
->fence
)))
5277 static bool radv_all_fences_syncobj(uint32_t fenceCount
, const VkFence
*pFences
)
5279 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
5280 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
5281 if (fence
->syncobj
== 0 && fence
->temp_syncobj
== 0)
5287 VkResult
radv_WaitForFences(
5289 uint32_t fenceCount
,
5290 const VkFence
* pFences
,
5294 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5295 timeout
= radv_get_absolute_timeout(timeout
);
5297 if (device
->always_use_syncobj
&&
5298 radv_all_fences_syncobj(fenceCount
, pFences
))
5300 uint32_t *handles
= malloc(sizeof(uint32_t) * fenceCount
);
5302 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
5304 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
5305 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
5306 handles
[i
] = fence
->temp_syncobj
? fence
->temp_syncobj
: fence
->syncobj
;
5309 bool success
= device
->ws
->wait_syncobj(device
->ws
, handles
, fenceCount
, waitAll
, timeout
);
5312 return success
? VK_SUCCESS
: VK_TIMEOUT
;
5315 if (!waitAll
&& fenceCount
> 1) {
5316 /* Not doing this by default for waitAll, due to needing to allocate twice. */
5317 if (device
->physical_device
->rad_info
.drm_minor
>= 10 && radv_all_fences_plain_and_submitted(device
, fenceCount
, pFences
)) {
5318 uint32_t wait_count
= 0;
5319 struct radeon_winsys_fence
**fences
= malloc(sizeof(struct radeon_winsys_fence
*) * fenceCount
);
5321 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
5323 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
5324 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
5326 if (device
->ws
->fence_wait(device
->ws
, fence
->fence
, false, 0)) {
5331 fences
[wait_count
++] = fence
->fence
;
5334 bool success
= device
->ws
->fences_wait(device
->ws
, fences
, wait_count
,
5335 waitAll
, timeout
- radv_get_current_time());
5338 return success
? VK_SUCCESS
: VK_TIMEOUT
;
5341 while(radv_get_current_time() <= timeout
) {
5342 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
5343 if (radv_GetFenceStatus(_device
, pFences
[i
]) == VK_SUCCESS
)
5350 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
5351 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
5352 bool expired
= false;
5354 if (fence
->temp_syncobj
) {
5355 if (!device
->ws
->wait_syncobj(device
->ws
, &fence
->temp_syncobj
, 1, true, timeout
))
5360 if (fence
->syncobj
) {
5361 if (!device
->ws
->wait_syncobj(device
->ws
, &fence
->syncobj
, 1, true, timeout
))
5367 if (!device
->ws
->is_fence_waitable(fence
->fence
)) {
5368 while(!device
->ws
->is_fence_waitable(fence
->fence
) &&
5369 radv_get_current_time() <= timeout
)
5373 expired
= device
->ws
->fence_wait(device
->ws
,
5380 if (fence
->fence_wsi
) {
5381 VkResult result
= fence
->fence_wsi
->wait(fence
->fence_wsi
, timeout
);
5382 if (result
!= VK_SUCCESS
)
5390 VkResult
radv_ResetFences(VkDevice _device
,
5391 uint32_t fenceCount
,
5392 const VkFence
*pFences
)
5394 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5396 for (unsigned i
= 0; i
< fenceCount
; ++i
) {
5397 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
5399 device
->ws
->reset_fence(fence
->fence
);
5401 /* Per spec, we first restore the permanent payload, and then reset, so
5402 * having a temp syncobj should not skip resetting the permanent syncobj. */
5403 if (fence
->temp_syncobj
) {
5404 device
->ws
->destroy_syncobj(device
->ws
, fence
->temp_syncobj
);
5405 fence
->temp_syncobj
= 0;
5408 if (fence
->syncobj
) {
5409 device
->ws
->reset_syncobj(device
->ws
, fence
->syncobj
);
5416 VkResult
radv_GetFenceStatus(VkDevice _device
, VkFence _fence
)
5418 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5419 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
5421 if (fence
->temp_syncobj
) {
5422 bool success
= device
->ws
->wait_syncobj(device
->ws
, &fence
->temp_syncobj
, 1, true, 0);
5423 return success
? VK_SUCCESS
: VK_NOT_READY
;
5426 if (fence
->syncobj
) {
5427 bool success
= device
->ws
->wait_syncobj(device
->ws
, &fence
->syncobj
, 1, true, 0);
5428 return success
? VK_SUCCESS
: VK_NOT_READY
;
5432 if (!device
->ws
->fence_wait(device
->ws
, fence
->fence
, false, 0))
5433 return VK_NOT_READY
;
5435 if (fence
->fence_wsi
) {
5436 VkResult result
= fence
->fence_wsi
->wait(fence
->fence_wsi
, 0);
5438 if (result
!= VK_SUCCESS
) {
5439 if (result
== VK_TIMEOUT
)
5440 return VK_NOT_READY
;
5448 // Queue semaphore functions
5451 radv_create_timeline(struct radv_timeline
*timeline
, uint64_t value
)
5453 timeline
->highest_signaled
= value
;
5454 timeline
->highest_submitted
= value
;
5455 list_inithead(&timeline
->points
);
5456 list_inithead(&timeline
->free_points
);
5457 list_inithead(&timeline
->waiters
);
5458 pthread_mutex_init(&timeline
->mutex
, NULL
);
5462 radv_destroy_timeline(struct radv_device
*device
,
5463 struct radv_timeline
*timeline
)
5465 list_for_each_entry_safe(struct radv_timeline_point
, point
,
5466 &timeline
->free_points
, list
) {
5467 list_del(&point
->list
);
5468 device
->ws
->destroy_syncobj(device
->ws
, point
->syncobj
);
5471 list_for_each_entry_safe(struct radv_timeline_point
, point
,
5472 &timeline
->points
, list
) {
5473 list_del(&point
->list
);
5474 device
->ws
->destroy_syncobj(device
->ws
, point
->syncobj
);
5477 pthread_mutex_destroy(&timeline
->mutex
);
5481 radv_timeline_gc_locked(struct radv_device
*device
,
5482 struct radv_timeline
*timeline
)
5484 list_for_each_entry_safe(struct radv_timeline_point
, point
,
5485 &timeline
->points
, list
) {
5486 if (point
->wait_count
|| point
->value
> timeline
->highest_submitted
)
5489 if (device
->ws
->wait_syncobj(device
->ws
, &point
->syncobj
, 1, true, 0)) {
5490 timeline
->highest_signaled
= point
->value
;
5491 list_del(&point
->list
);
5492 list_add(&point
->list
, &timeline
->free_points
);
5497 static struct radv_timeline_point
*
5498 radv_timeline_find_point_at_least_locked(struct radv_device
*device
,
5499 struct radv_timeline
*timeline
,
5502 radv_timeline_gc_locked(device
, timeline
);
5504 if (p
<= timeline
->highest_signaled
)
5507 list_for_each_entry(struct radv_timeline_point
, point
,
5508 &timeline
->points
, list
) {
5509 if (point
->value
>= p
) {
5510 ++point
->wait_count
;
5517 static struct radv_timeline_point
*
5518 radv_timeline_add_point_locked(struct radv_device
*device
,
5519 struct radv_timeline
*timeline
,
5522 radv_timeline_gc_locked(device
, timeline
);
5524 struct radv_timeline_point
*ret
= NULL
;
5525 struct radv_timeline_point
*prev
= NULL
;
5527 if (p
<= timeline
->highest_signaled
)
5530 list_for_each_entry(struct radv_timeline_point
, point
,
5531 &timeline
->points
, list
) {
5532 if (point
->value
== p
) {
5536 if (point
->value
< p
)
5540 if (list_is_empty(&timeline
->free_points
)) {
5541 ret
= malloc(sizeof(struct radv_timeline_point
));
5542 device
->ws
->create_syncobj(device
->ws
, &ret
->syncobj
);
5544 ret
= list_first_entry(&timeline
->free_points
, struct radv_timeline_point
, list
);
5545 list_del(&ret
->list
);
5547 device
->ws
->reset_syncobj(device
->ws
, ret
->syncobj
);
5551 ret
->wait_count
= 1;
5554 list_add(&ret
->list
, &prev
->list
);
5556 list_addtail(&ret
->list
, &timeline
->points
);
5563 radv_timeline_wait_locked(struct radv_device
*device
,
5564 struct radv_timeline
*timeline
,
5566 uint64_t abs_timeout
)
5568 while(timeline
->highest_submitted
< value
) {
5569 struct timespec abstime
;
5570 timespec_from_nsec(&abstime
, abs_timeout
);
5572 pthread_cond_timedwait(&device
->timeline_cond
, &timeline
->mutex
, &abstime
);
5574 if (radv_get_current_time() >= abs_timeout
&& timeline
->highest_submitted
< value
)
5578 struct radv_timeline_point
*point
= radv_timeline_find_point_at_least_locked(device
, timeline
, value
);
5582 pthread_mutex_unlock(&timeline
->mutex
);
5584 bool success
= device
->ws
->wait_syncobj(device
->ws
, &point
->syncobj
, 1, true, abs_timeout
);
5586 pthread_mutex_lock(&timeline
->mutex
);
5587 point
->wait_count
--;
5588 return success
? VK_SUCCESS
: VK_TIMEOUT
;
5592 radv_timeline_trigger_waiters_locked(struct radv_timeline
*timeline
,
5593 struct list_head
*processing_list
)
5595 list_for_each_entry_safe(struct radv_timeline_waiter
, waiter
,
5596 &timeline
->waiters
, list
) {
5597 if (waiter
->value
> timeline
->highest_submitted
)
5600 if (p_atomic_dec_zero(&waiter
->submission
->submission_wait_count
)) {
5601 list_addtail(&waiter
->submission
->processing_list
, processing_list
);
5603 list_del(&waiter
->list
);
5608 void radv_destroy_semaphore_part(struct radv_device
*device
,
5609 struct radv_semaphore_part
*part
)
5611 switch(part
->kind
) {
5612 case RADV_SEMAPHORE_NONE
:
5614 case RADV_SEMAPHORE_WINSYS
:
5615 device
->ws
->destroy_sem(part
->ws_sem
);
5617 case RADV_SEMAPHORE_TIMELINE
:
5618 radv_destroy_timeline(device
, &part
->timeline
);
5620 case RADV_SEMAPHORE_SYNCOBJ
:
5621 device
->ws
->destroy_syncobj(device
->ws
, part
->syncobj
);
5624 part
->kind
= RADV_SEMAPHORE_NONE
;
5627 static VkSemaphoreTypeKHR
5628 radv_get_semaphore_type(const void *pNext
, uint64_t *initial_value
)
5630 const VkSemaphoreTypeCreateInfoKHR
*type_info
=
5631 vk_find_struct_const(pNext
, SEMAPHORE_TYPE_CREATE_INFO_KHR
);
5634 return VK_SEMAPHORE_TYPE_BINARY_KHR
;
5637 *initial_value
= type_info
->initialValue
;
5638 return type_info
->semaphoreType
;
5641 VkResult
radv_CreateSemaphore(
5643 const VkSemaphoreCreateInfo
* pCreateInfo
,
5644 const VkAllocationCallbacks
* pAllocator
,
5645 VkSemaphore
* pSemaphore
)
5647 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5648 const VkExportSemaphoreCreateInfo
*export
=
5649 vk_find_struct_const(pCreateInfo
->pNext
, EXPORT_SEMAPHORE_CREATE_INFO
);
5650 VkExternalSemaphoreHandleTypeFlags handleTypes
=
5651 export
? export
->handleTypes
: 0;
5652 uint64_t initial_value
= 0;
5653 VkSemaphoreTypeKHR type
= radv_get_semaphore_type(pCreateInfo
->pNext
, &initial_value
);
5655 struct radv_semaphore
*sem
= vk_alloc2(&device
->alloc
, pAllocator
,
5657 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
5659 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
5661 sem
->temporary
.kind
= RADV_SEMAPHORE_NONE
;
5662 sem
->permanent
.kind
= RADV_SEMAPHORE_NONE
;
5664 if (type
== VK_SEMAPHORE_TYPE_TIMELINE_KHR
) {
5665 radv_create_timeline(&sem
->permanent
.timeline
, initial_value
);
5666 sem
->permanent
.kind
= RADV_SEMAPHORE_TIMELINE
;
5667 } else if (device
->always_use_syncobj
|| handleTypes
) {
5668 assert (device
->physical_device
->rad_info
.has_syncobj
);
5669 int ret
= device
->ws
->create_syncobj(device
->ws
, &sem
->permanent
.syncobj
);
5671 vk_free2(&device
->alloc
, pAllocator
, sem
);
5672 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
5674 sem
->permanent
.kind
= RADV_SEMAPHORE_SYNCOBJ
;
5676 sem
->permanent
.ws_sem
= device
->ws
->create_sem(device
->ws
);
5677 if (!sem
->permanent
.ws_sem
) {
5678 vk_free2(&device
->alloc
, pAllocator
, sem
);
5679 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
5681 sem
->permanent
.kind
= RADV_SEMAPHORE_WINSYS
;
5684 *pSemaphore
= radv_semaphore_to_handle(sem
);
5688 void radv_DestroySemaphore(
5690 VkSemaphore _semaphore
,
5691 const VkAllocationCallbacks
* pAllocator
)
5693 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5694 RADV_FROM_HANDLE(radv_semaphore
, sem
, _semaphore
);
5698 radv_destroy_semaphore_part(device
, &sem
->temporary
);
5699 radv_destroy_semaphore_part(device
, &sem
->permanent
);
5700 vk_free2(&device
->alloc
, pAllocator
, sem
);
5704 radv_GetSemaphoreCounterValueKHR(VkDevice _device
,
5705 VkSemaphore _semaphore
,
5708 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5709 RADV_FROM_HANDLE(radv_semaphore
, semaphore
, _semaphore
);
5711 struct radv_semaphore_part
*part
=
5712 semaphore
->temporary
.kind
!= RADV_SEMAPHORE_NONE
? &semaphore
->temporary
: &semaphore
->permanent
;
5714 switch (part
->kind
) {
5715 case RADV_SEMAPHORE_TIMELINE
: {
5716 pthread_mutex_lock(&part
->timeline
.mutex
);
5717 radv_timeline_gc_locked(device
, &part
->timeline
);
5718 *pValue
= part
->timeline
.highest_signaled
;
5719 pthread_mutex_unlock(&part
->timeline
.mutex
);
5722 case RADV_SEMAPHORE_NONE
:
5723 case RADV_SEMAPHORE_SYNCOBJ
:
5724 case RADV_SEMAPHORE_WINSYS
:
5725 unreachable("Invalid semaphore type");
5727 unreachable("Unhandled semaphore type");
5732 radv_wait_timelines(struct radv_device
*device
,
5733 const VkSemaphoreWaitInfoKHR
* pWaitInfo
,
5734 uint64_t abs_timeout
)
5736 if ((pWaitInfo
->flags
& VK_SEMAPHORE_WAIT_ANY_BIT_KHR
) && pWaitInfo
->semaphoreCount
> 1) {
5738 for(uint32_t i
= 0; i
< pWaitInfo
->semaphoreCount
; ++i
) {
5739 RADV_FROM_HANDLE(radv_semaphore
, semaphore
, pWaitInfo
->pSemaphores
[i
]);
5740 pthread_mutex_lock(&semaphore
->permanent
.timeline
.mutex
);
5741 VkResult result
= radv_timeline_wait_locked(device
, &semaphore
->permanent
.timeline
, pWaitInfo
->pValues
[i
], 0);
5742 pthread_mutex_unlock(&semaphore
->permanent
.timeline
.mutex
);
5744 if (result
== VK_SUCCESS
)
5747 if (radv_get_current_time() > abs_timeout
)
5752 for(uint32_t i
= 0; i
< pWaitInfo
->semaphoreCount
; ++i
) {
5753 RADV_FROM_HANDLE(radv_semaphore
, semaphore
, pWaitInfo
->pSemaphores
[i
]);
5754 pthread_mutex_lock(&semaphore
->permanent
.timeline
.mutex
);
5755 VkResult result
= radv_timeline_wait_locked(device
, &semaphore
->permanent
.timeline
, pWaitInfo
->pValues
[i
], abs_timeout
);
5756 pthread_mutex_unlock(&semaphore
->permanent
.timeline
.mutex
);
5758 if (result
!= VK_SUCCESS
)
5764 radv_WaitSemaphoresKHR(VkDevice _device
,
5765 const VkSemaphoreWaitInfoKHR
* pWaitInfo
,
5768 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5769 uint64_t abs_timeout
= radv_get_absolute_timeout(timeout
);
5770 return radv_wait_timelines(device
, pWaitInfo
, abs_timeout
);
5774 radv_SignalSemaphoreKHR(VkDevice _device
,
5775 const VkSemaphoreSignalInfoKHR
* pSignalInfo
)
5777 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5778 RADV_FROM_HANDLE(radv_semaphore
, semaphore
, pSignalInfo
->semaphore
);
5780 struct radv_semaphore_part
*part
=
5781 semaphore
->temporary
.kind
!= RADV_SEMAPHORE_NONE
? &semaphore
->temporary
: &semaphore
->permanent
;
5783 switch(part
->kind
) {
5784 case RADV_SEMAPHORE_TIMELINE
: {
5785 pthread_mutex_lock(&part
->timeline
.mutex
);
5786 radv_timeline_gc_locked(device
, &part
->timeline
);
5787 part
->timeline
.highest_submitted
= MAX2(part
->timeline
.highest_submitted
, pSignalInfo
->value
);
5788 part
->timeline
.highest_signaled
= MAX2(part
->timeline
.highest_signaled
, pSignalInfo
->value
);
5790 struct list_head processing_list
;
5791 list_inithead(&processing_list
);
5792 radv_timeline_trigger_waiters_locked(&part
->timeline
, &processing_list
);
5793 pthread_mutex_unlock(&part
->timeline
.mutex
);
5795 return radv_process_submissions(&processing_list
);
5797 case RADV_SEMAPHORE_NONE
:
5798 case RADV_SEMAPHORE_SYNCOBJ
:
5799 case RADV_SEMAPHORE_WINSYS
:
5800 unreachable("Invalid semaphore type");
5807 VkResult
radv_CreateEvent(
5809 const VkEventCreateInfo
* pCreateInfo
,
5810 const VkAllocationCallbacks
* pAllocator
,
5813 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5814 struct radv_event
*event
= vk_alloc2(&device
->alloc
, pAllocator
,
5816 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
5819 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
5821 event
->bo
= device
->ws
->buffer_create(device
->ws
, 8, 8,
5823 RADEON_FLAG_VA_UNCACHED
| RADEON_FLAG_CPU_ACCESS
| RADEON_FLAG_NO_INTERPROCESS_SHARING
,
5824 RADV_BO_PRIORITY_FENCE
);
5826 vk_free2(&device
->alloc
, pAllocator
, event
);
5827 return vk_error(device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
5830 event
->map
= (uint64_t*)device
->ws
->buffer_map(event
->bo
);
5832 *pEvent
= radv_event_to_handle(event
);
5837 void radv_DestroyEvent(
5840 const VkAllocationCallbacks
* pAllocator
)
5842 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5843 RADV_FROM_HANDLE(radv_event
, event
, _event
);
5847 device
->ws
->buffer_destroy(event
->bo
);
5848 vk_free2(&device
->alloc
, pAllocator
, event
);
5851 VkResult
radv_GetEventStatus(
5855 RADV_FROM_HANDLE(radv_event
, event
, _event
);
5857 if (*event
->map
== 1)
5858 return VK_EVENT_SET
;
5859 return VK_EVENT_RESET
;
5862 VkResult
radv_SetEvent(
5866 RADV_FROM_HANDLE(radv_event
, event
, _event
);
5872 VkResult
radv_ResetEvent(
5876 RADV_FROM_HANDLE(radv_event
, event
, _event
);
5882 VkResult
radv_CreateBuffer(
5884 const VkBufferCreateInfo
* pCreateInfo
,
5885 const VkAllocationCallbacks
* pAllocator
,
5888 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5889 struct radv_buffer
*buffer
;
5891 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
5893 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
5894 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
5896 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
5898 buffer
->size
= pCreateInfo
->size
;
5899 buffer
->usage
= pCreateInfo
->usage
;
5902 buffer
->flags
= pCreateInfo
->flags
;
5904 buffer
->shareable
= vk_find_struct_const(pCreateInfo
->pNext
,
5905 EXTERNAL_MEMORY_BUFFER_CREATE_INFO
) != NULL
;
5907 if (pCreateInfo
->flags
& VK_BUFFER_CREATE_SPARSE_BINDING_BIT
) {
5908 buffer
->bo
= device
->ws
->buffer_create(device
->ws
,
5909 align64(buffer
->size
, 4096),
5910 4096, 0, RADEON_FLAG_VIRTUAL
,
5911 RADV_BO_PRIORITY_VIRTUAL
);
5913 vk_free2(&device
->alloc
, pAllocator
, buffer
);
5914 return vk_error(device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
5918 *pBuffer
= radv_buffer_to_handle(buffer
);
5923 void radv_DestroyBuffer(
5926 const VkAllocationCallbacks
* pAllocator
)
5928 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5929 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
5934 if (buffer
->flags
& VK_BUFFER_CREATE_SPARSE_BINDING_BIT
)
5935 device
->ws
->buffer_destroy(buffer
->bo
);
5937 vk_free2(&device
->alloc
, pAllocator
, buffer
);
5940 VkDeviceAddress
radv_GetBufferDeviceAddressKHR(
5942 const VkBufferDeviceAddressInfoKHR
* pInfo
)
5944 RADV_FROM_HANDLE(radv_buffer
, buffer
, pInfo
->buffer
);
5945 return radv_buffer_get_va(buffer
->bo
) + buffer
->offset
;
5949 uint64_t radv_GetBufferOpaqueCaptureAddressKHR(VkDevice device
,
5950 const VkBufferDeviceAddressInfoKHR
* pInfo
)
5955 uint64_t radv_GetDeviceMemoryOpaqueCaptureAddressKHR(VkDevice device
,
5956 const VkDeviceMemoryOpaqueCaptureAddressInfoKHR
* pInfo
)
5961 static inline unsigned
5962 si_tile_mode_index(const struct radv_image_plane
*plane
, unsigned level
, bool stencil
)
5965 return plane
->surface
.u
.legacy
.stencil_tiling_index
[level
];
5967 return plane
->surface
.u
.legacy
.tiling_index
[level
];
5970 static uint32_t radv_surface_max_layer_count(struct radv_image_view
*iview
)
5972 return iview
->type
== VK_IMAGE_VIEW_TYPE_3D
? iview
->extent
.depth
: (iview
->base_layer
+ iview
->layer_count
);
5976 radv_init_dcc_control_reg(struct radv_device
*device
,
5977 struct radv_image_view
*iview
)
5979 unsigned max_uncompressed_block_size
= V_028C78_MAX_BLOCK_SIZE_256B
;
5980 unsigned min_compressed_block_size
= V_028C78_MIN_BLOCK_SIZE_32B
;
5981 unsigned max_compressed_block_size
;
5982 unsigned independent_128b_blocks
;
5983 unsigned independent_64b_blocks
;
5985 if (!radv_dcc_enabled(iview
->image
, iview
->base_mip
))
5988 if (!device
->physical_device
->rad_info
.has_dedicated_vram
) {
5989 /* amdvlk: [min-compressed-block-size] should be set to 32 for
5990 * dGPU and 64 for APU because all of our APUs to date use
5991 * DIMMs which have a request granularity size of 64B while all
5992 * other chips have a 32B request size.
5994 min_compressed_block_size
= V_028C78_MIN_BLOCK_SIZE_64B
;
5997 if (device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
5998 max_compressed_block_size
= V_028C78_MAX_BLOCK_SIZE_128B
;
5999 independent_64b_blocks
= 0;
6000 independent_128b_blocks
= 1;
6002 independent_128b_blocks
= 0;
6004 if (iview
->image
->info
.samples
> 1) {
6005 if (iview
->image
->planes
[0].surface
.bpe
== 1)
6006 max_uncompressed_block_size
= V_028C78_MAX_BLOCK_SIZE_64B
;
6007 else if (iview
->image
->planes
[0].surface
.bpe
== 2)
6008 max_uncompressed_block_size
= V_028C78_MAX_BLOCK_SIZE_128B
;
6011 if (iview
->image
->usage
& (VK_IMAGE_USAGE_SAMPLED_BIT
|
6012 VK_IMAGE_USAGE_TRANSFER_SRC_BIT
|
6013 VK_IMAGE_USAGE_INPUT_ATTACHMENT_BIT
)) {
6014 /* If this DCC image is potentially going to be used in texture
6015 * fetches, we need some special settings.
6017 independent_64b_blocks
= 1;
6018 max_compressed_block_size
= V_028C78_MAX_BLOCK_SIZE_64B
;
6020 /* MAX_UNCOMPRESSED_BLOCK_SIZE must be >=
6021 * MAX_COMPRESSED_BLOCK_SIZE. Set MAX_COMPRESSED_BLOCK_SIZE as
6022 * big as possible for better compression state.
6024 independent_64b_blocks
= 0;
6025 max_compressed_block_size
= max_uncompressed_block_size
;
6029 return S_028C78_MAX_UNCOMPRESSED_BLOCK_SIZE(max_uncompressed_block_size
) |
6030 S_028C78_MAX_COMPRESSED_BLOCK_SIZE(max_compressed_block_size
) |
6031 S_028C78_MIN_COMPRESSED_BLOCK_SIZE(min_compressed_block_size
) |
6032 S_028C78_INDEPENDENT_64B_BLOCKS(independent_64b_blocks
) |
6033 S_028C78_INDEPENDENT_128B_BLOCKS(independent_128b_blocks
);
6037 radv_initialise_color_surface(struct radv_device
*device
,
6038 struct radv_color_buffer_info
*cb
,
6039 struct radv_image_view
*iview
)
6041 const struct vk_format_description
*desc
;
6042 unsigned ntype
, format
, swap
, endian
;
6043 unsigned blend_clamp
= 0, blend_bypass
= 0;
6045 const struct radv_image_plane
*plane
= &iview
->image
->planes
[iview
->plane_id
];
6046 const struct radeon_surf
*surf
= &plane
->surface
;
6048 desc
= vk_format_description(iview
->vk_format
);
6050 memset(cb
, 0, sizeof(*cb
));
6052 /* Intensity is implemented as Red, so treat it that way. */
6053 cb
->cb_color_attrib
= S_028C74_FORCE_DST_ALPHA_1(desc
->swizzle
[3] == VK_SWIZZLE_1
);
6055 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
+ plane
->offset
;
6057 cb
->cb_color_base
= va
>> 8;
6059 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
6060 struct gfx9_surf_meta_flags meta
;
6061 if (iview
->image
->dcc_offset
)
6062 meta
= surf
->u
.gfx9
.dcc
;
6064 meta
= surf
->u
.gfx9
.cmask
;
6066 if (device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
6067 cb
->cb_color_attrib3
|= S_028EE0_COLOR_SW_MODE(surf
->u
.gfx9
.surf
.swizzle_mode
) |
6068 S_028EE0_FMASK_SW_MODE(surf
->u
.gfx9
.fmask
.swizzle_mode
) |
6069 S_028EE0_CMASK_PIPE_ALIGNED(surf
->u
.gfx9
.cmask
.pipe_aligned
) |
6070 S_028EE0_DCC_PIPE_ALIGNED(surf
->u
.gfx9
.dcc
.pipe_aligned
);
6072 cb
->cb_color_attrib
|= S_028C74_COLOR_SW_MODE(surf
->u
.gfx9
.surf
.swizzle_mode
) |
6073 S_028C74_FMASK_SW_MODE(surf
->u
.gfx9
.fmask
.swizzle_mode
) |
6074 S_028C74_RB_ALIGNED(meta
.rb_aligned
) |
6075 S_028C74_PIPE_ALIGNED(meta
.pipe_aligned
);
6076 cb
->cb_mrt_epitch
= S_0287A0_EPITCH(surf
->u
.gfx9
.surf
.epitch
);
6079 cb
->cb_color_base
+= surf
->u
.gfx9
.surf_offset
>> 8;
6080 cb
->cb_color_base
|= surf
->tile_swizzle
;
6082 const struct legacy_surf_level
*level_info
= &surf
->u
.legacy
.level
[iview
->base_mip
];
6083 unsigned pitch_tile_max
, slice_tile_max
, tile_mode_index
;
6085 cb
->cb_color_base
+= level_info
->offset
>> 8;
6086 if (level_info
->mode
== RADEON_SURF_MODE_2D
)
6087 cb
->cb_color_base
|= surf
->tile_swizzle
;
6089 pitch_tile_max
= level_info
->nblk_x
/ 8 - 1;
6090 slice_tile_max
= (level_info
->nblk_x
* level_info
->nblk_y
) / 64 - 1;
6091 tile_mode_index
= si_tile_mode_index(plane
, iview
->base_mip
, false);
6093 cb
->cb_color_pitch
= S_028C64_TILE_MAX(pitch_tile_max
);
6094 cb
->cb_color_slice
= S_028C68_TILE_MAX(slice_tile_max
);
6095 cb
->cb_color_cmask_slice
= surf
->u
.legacy
.cmask_slice_tile_max
;
6097 cb
->cb_color_attrib
|= S_028C74_TILE_MODE_INDEX(tile_mode_index
);
6099 if (radv_image_has_fmask(iview
->image
)) {
6100 if (device
->physical_device
->rad_info
.chip_class
>= GFX7
)
6101 cb
->cb_color_pitch
|= S_028C64_FMASK_TILE_MAX(surf
->u
.legacy
.fmask
.pitch_in_pixels
/ 8 - 1);
6102 cb
->cb_color_attrib
|= S_028C74_FMASK_TILE_MODE_INDEX(surf
->u
.legacy
.fmask
.tiling_index
);
6103 cb
->cb_color_fmask_slice
= S_028C88_TILE_MAX(surf
->u
.legacy
.fmask
.slice_tile_max
);
6105 /* This must be set for fast clear to work without FMASK. */
6106 if (device
->physical_device
->rad_info
.chip_class
>= GFX7
)
6107 cb
->cb_color_pitch
|= S_028C64_FMASK_TILE_MAX(pitch_tile_max
);
6108 cb
->cb_color_attrib
|= S_028C74_FMASK_TILE_MODE_INDEX(tile_mode_index
);
6109 cb
->cb_color_fmask_slice
= S_028C88_TILE_MAX(slice_tile_max
);
6113 /* CMASK variables */
6114 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
;
6115 va
+= iview
->image
->cmask_offset
;
6116 cb
->cb_color_cmask
= va
>> 8;
6118 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
;
6119 va
+= iview
->image
->dcc_offset
;
6121 if (radv_dcc_enabled(iview
->image
, iview
->base_mip
) &&
6122 device
->physical_device
->rad_info
.chip_class
<= GFX8
)
6123 va
+= plane
->surface
.u
.legacy
.level
[iview
->base_mip
].dcc_offset
;
6125 unsigned dcc_tile_swizzle
= surf
->tile_swizzle
;
6126 dcc_tile_swizzle
&= (surf
->dcc_alignment
- 1) >> 8;
6128 cb
->cb_dcc_base
= va
>> 8;
6129 cb
->cb_dcc_base
|= dcc_tile_swizzle
;
6131 /* GFX10 field has the same base shift as the GFX6 field. */
6132 uint32_t max_slice
= radv_surface_max_layer_count(iview
) - 1;
6133 cb
->cb_color_view
= S_028C6C_SLICE_START(iview
->base_layer
) |
6134 S_028C6C_SLICE_MAX_GFX10(max_slice
);
6136 if (iview
->image
->info
.samples
> 1) {
6137 unsigned log_samples
= util_logbase2(iview
->image
->info
.samples
);
6139 cb
->cb_color_attrib
|= S_028C74_NUM_SAMPLES(log_samples
) |
6140 S_028C74_NUM_FRAGMENTS(log_samples
);
6143 if (radv_image_has_fmask(iview
->image
)) {
6144 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
+ iview
->image
->fmask_offset
;
6145 cb
->cb_color_fmask
= va
>> 8;
6146 cb
->cb_color_fmask
|= surf
->fmask_tile_swizzle
;
6148 cb
->cb_color_fmask
= cb
->cb_color_base
;
6151 ntype
= radv_translate_color_numformat(iview
->vk_format
,
6153 vk_format_get_first_non_void_channel(iview
->vk_format
));
6154 format
= radv_translate_colorformat(iview
->vk_format
);
6155 if (format
== V_028C70_COLOR_INVALID
|| ntype
== ~0u)
6156 radv_finishme("Illegal color\n");
6157 swap
= radv_translate_colorswap(iview
->vk_format
, false);
6158 endian
= radv_colorformat_endian_swap(format
);
6160 /* blend clamp should be set for all NORM/SRGB types */
6161 if (ntype
== V_028C70_NUMBER_UNORM
||
6162 ntype
== V_028C70_NUMBER_SNORM
||
6163 ntype
== V_028C70_NUMBER_SRGB
)
6166 /* set blend bypass according to docs if SINT/UINT or
6167 8/24 COLOR variants */
6168 if (ntype
== V_028C70_NUMBER_UINT
|| ntype
== V_028C70_NUMBER_SINT
||
6169 format
== V_028C70_COLOR_8_24
|| format
== V_028C70_COLOR_24_8
||
6170 format
== V_028C70_COLOR_X24_8_32_FLOAT
) {
6175 if ((ntype
== V_028C70_NUMBER_UINT
|| ntype
== V_028C70_NUMBER_SINT
) &&
6176 (format
== V_028C70_COLOR_8
||
6177 format
== V_028C70_COLOR_8_8
||
6178 format
== V_028C70_COLOR_8_8_8_8
))
6179 ->color_is_int8
= true;
6181 cb
->cb_color_info
= S_028C70_FORMAT(format
) |
6182 S_028C70_COMP_SWAP(swap
) |
6183 S_028C70_BLEND_CLAMP(blend_clamp
) |
6184 S_028C70_BLEND_BYPASS(blend_bypass
) |
6185 S_028C70_SIMPLE_FLOAT(1) |
6186 S_028C70_ROUND_MODE(ntype
!= V_028C70_NUMBER_UNORM
&&
6187 ntype
!= V_028C70_NUMBER_SNORM
&&
6188 ntype
!= V_028C70_NUMBER_SRGB
&&
6189 format
!= V_028C70_COLOR_8_24
&&
6190 format
!= V_028C70_COLOR_24_8
) |
6191 S_028C70_NUMBER_TYPE(ntype
) |
6192 S_028C70_ENDIAN(endian
);
6193 if (radv_image_has_fmask(iview
->image
)) {
6194 cb
->cb_color_info
|= S_028C70_COMPRESSION(1);
6195 if (device
->physical_device
->rad_info
.chip_class
== GFX6
) {
6196 unsigned fmask_bankh
= util_logbase2(surf
->u
.legacy
.fmask
.bankh
);
6197 cb
->cb_color_attrib
|= S_028C74_FMASK_BANK_HEIGHT(fmask_bankh
);
6200 if (radv_image_is_tc_compat_cmask(iview
->image
)) {
6201 /* Allow the texture block to read FMASK directly
6202 * without decompressing it. This bit must be cleared
6203 * when performing FMASK_DECOMPRESS or DCC_COMPRESS,
6204 * otherwise the operation doesn't happen.
6206 cb
->cb_color_info
|= S_028C70_FMASK_COMPRESS_1FRAG_ONLY(1);
6208 /* Set CMASK into a tiling format that allows the
6209 * texture block to read it.
6211 cb
->cb_color_info
|= S_028C70_CMASK_ADDR_TYPE(2);
6215 if (radv_image_has_cmask(iview
->image
) &&
6216 !(device
->instance
->debug_flags
& RADV_DEBUG_NO_FAST_CLEARS
))
6217 cb
->cb_color_info
|= S_028C70_FAST_CLEAR(1);
6219 if (radv_dcc_enabled(iview
->image
, iview
->base_mip
))
6220 cb
->cb_color_info
|= S_028C70_DCC_ENABLE(1);
6222 cb
->cb_dcc_control
= radv_init_dcc_control_reg(device
, iview
);
6224 /* This must be set for fast clear to work without FMASK. */
6225 if (!radv_image_has_fmask(iview
->image
) &&
6226 device
->physical_device
->rad_info
.chip_class
== GFX6
) {
6227 unsigned bankh
= util_logbase2(surf
->u
.legacy
.bankh
);
6228 cb
->cb_color_attrib
|= S_028C74_FMASK_BANK_HEIGHT(bankh
);
6231 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
6232 const struct vk_format_description
*format_desc
= vk_format_description(iview
->image
->vk_format
);
6234 unsigned mip0_depth
= iview
->image
->type
== VK_IMAGE_TYPE_3D
?
6235 (iview
->extent
.depth
- 1) : (iview
->image
->info
.array_size
- 1);
6236 unsigned width
= iview
->extent
.width
/ (iview
->plane_id
? format_desc
->width_divisor
: 1);
6237 unsigned height
= iview
->extent
.height
/ (iview
->plane_id
? format_desc
->height_divisor
: 1);
6239 if (device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
6240 cb
->cb_color_view
|= S_028C6C_MIP_LEVEL_GFX10(iview
->base_mip
);
6242 cb
->cb_color_attrib3
|= S_028EE0_MIP0_DEPTH(mip0_depth
) |
6243 S_028EE0_RESOURCE_TYPE(surf
->u
.gfx9
.resource_type
) |
6244 S_028EE0_RESOURCE_LEVEL(1);
6246 cb
->cb_color_view
|= S_028C6C_MIP_LEVEL_GFX9(iview
->base_mip
);
6247 cb
->cb_color_attrib
|= S_028C74_MIP0_DEPTH(mip0_depth
) |
6248 S_028C74_RESOURCE_TYPE(surf
->u
.gfx9
.resource_type
);
6251 cb
->cb_color_attrib2
= S_028C68_MIP0_WIDTH(width
- 1) |
6252 S_028C68_MIP0_HEIGHT(height
- 1) |
6253 S_028C68_MAX_MIP(iview
->image
->info
.levels
- 1);
6258 radv_calc_decompress_on_z_planes(struct radv_device
*device
,
6259 struct radv_image_view
*iview
)
6261 unsigned max_zplanes
= 0;
6263 assert(radv_image_is_tc_compat_htile(iview
->image
));
6265 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
6266 /* Default value for 32-bit depth surfaces. */
6269 if (iview
->vk_format
== VK_FORMAT_D16_UNORM
&&
6270 iview
->image
->info
.samples
> 1)
6273 max_zplanes
= max_zplanes
+ 1;
6275 if (iview
->vk_format
== VK_FORMAT_D16_UNORM
) {
6276 /* Do not enable Z plane compression for 16-bit depth
6277 * surfaces because isn't supported on GFX8. Only
6278 * 32-bit depth surfaces are supported by the hardware.
6279 * This allows to maintain shader compatibility and to
6280 * reduce the number of depth decompressions.
6284 if (iview
->image
->info
.samples
<= 1)
6286 else if (iview
->image
->info
.samples
<= 4)
6297 radv_initialise_ds_surface(struct radv_device
*device
,
6298 struct radv_ds_buffer_info
*ds
,
6299 struct radv_image_view
*iview
)
6301 unsigned level
= iview
->base_mip
;
6302 unsigned format
, stencil_format
;
6303 uint64_t va
, s_offs
, z_offs
;
6304 bool stencil_only
= false;
6305 const struct radv_image_plane
*plane
= &iview
->image
->planes
[0];
6306 const struct radeon_surf
*surf
= &plane
->surface
;
6308 assert(vk_format_get_plane_count(iview
->image
->vk_format
) == 1);
6310 memset(ds
, 0, sizeof(*ds
));
6311 switch (iview
->image
->vk_format
) {
6312 case VK_FORMAT_D24_UNORM_S8_UINT
:
6313 case VK_FORMAT_X8_D24_UNORM_PACK32
:
6314 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-24);
6315 ds
->offset_scale
= 2.0f
;
6317 case VK_FORMAT_D16_UNORM
:
6318 case VK_FORMAT_D16_UNORM_S8_UINT
:
6319 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-16);
6320 ds
->offset_scale
= 4.0f
;
6322 case VK_FORMAT_D32_SFLOAT
:
6323 case VK_FORMAT_D32_SFLOAT_S8_UINT
:
6324 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-23) |
6325 S_028B78_POLY_OFFSET_DB_IS_FLOAT_FMT(1);
6326 ds
->offset_scale
= 1.0f
;
6328 case VK_FORMAT_S8_UINT
:
6329 stencil_only
= true;
6335 format
= radv_translate_dbformat(iview
->image
->vk_format
);
6336 stencil_format
= surf
->has_stencil
?
6337 V_028044_STENCIL_8
: V_028044_STENCIL_INVALID
;
6339 uint32_t max_slice
= radv_surface_max_layer_count(iview
) - 1;
6340 ds
->db_depth_view
= S_028008_SLICE_START(iview
->base_layer
) |
6341 S_028008_SLICE_MAX(max_slice
);
6342 if (device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
6343 ds
->db_depth_view
|= S_028008_SLICE_START_HI(iview
->base_layer
>> 11) |
6344 S_028008_SLICE_MAX_HI(max_slice
>> 11);
6347 ds
->db_htile_data_base
= 0;
6348 ds
->db_htile_surface
= 0;
6350 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
;
6351 s_offs
= z_offs
= va
;
6353 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
6354 assert(surf
->u
.gfx9
.surf_offset
== 0);
6355 s_offs
+= surf
->u
.gfx9
.stencil_offset
;
6357 ds
->db_z_info
= S_028038_FORMAT(format
) |
6358 S_028038_NUM_SAMPLES(util_logbase2(iview
->image
->info
.samples
)) |
6359 S_028038_SW_MODE(surf
->u
.gfx9
.surf
.swizzle_mode
) |
6360 S_028038_MAXMIP(iview
->image
->info
.levels
- 1) |
6361 S_028038_ZRANGE_PRECISION(1);
6362 ds
->db_stencil_info
= S_02803C_FORMAT(stencil_format
) |
6363 S_02803C_SW_MODE(surf
->u
.gfx9
.stencil
.swizzle_mode
);
6365 if (device
->physical_device
->rad_info
.chip_class
== GFX9
) {
6366 ds
->db_z_info2
= S_028068_EPITCH(surf
->u
.gfx9
.surf
.epitch
);
6367 ds
->db_stencil_info2
= S_02806C_EPITCH(surf
->u
.gfx9
.stencil
.epitch
);
6370 ds
->db_depth_view
|= S_028008_MIPID(level
);
6371 ds
->db_depth_size
= S_02801C_X_MAX(iview
->image
->info
.width
- 1) |
6372 S_02801C_Y_MAX(iview
->image
->info
.height
- 1);
6374 if (radv_htile_enabled(iview
->image
, level
)) {
6375 ds
->db_z_info
|= S_028038_TILE_SURFACE_ENABLE(1);
6377 if (radv_image_is_tc_compat_htile(iview
->image
)) {
6378 unsigned max_zplanes
=
6379 radv_calc_decompress_on_z_planes(device
, iview
);
6381 ds
->db_z_info
|= S_028038_DECOMPRESS_ON_N_ZPLANES(max_zplanes
);
6383 if (device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
6384 ds
->db_z_info
|= S_028040_ITERATE_FLUSH(1);
6385 ds
->db_stencil_info
|= S_028044_ITERATE_FLUSH(1);
6387 ds
->db_z_info
|= S_028038_ITERATE_FLUSH(1);
6388 ds
->db_stencil_info
|= S_02803C_ITERATE_FLUSH(1);
6392 if (!surf
->has_stencil
)
6393 /* Use all of the htile_buffer for depth if there's no stencil. */
6394 ds
->db_stencil_info
|= S_02803C_TILE_STENCIL_DISABLE(1);
6395 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
+
6396 iview
->image
->htile_offset
;
6397 ds
->db_htile_data_base
= va
>> 8;
6398 ds
->db_htile_surface
= S_028ABC_FULL_CACHE(1) |
6399 S_028ABC_PIPE_ALIGNED(surf
->u
.gfx9
.htile
.pipe_aligned
);
6401 if (device
->physical_device
->rad_info
.chip_class
== GFX9
) {
6402 ds
->db_htile_surface
|= S_028ABC_RB_ALIGNED(surf
->u
.gfx9
.htile
.rb_aligned
);
6406 const struct legacy_surf_level
*level_info
= &surf
->u
.legacy
.level
[level
];
6409 level_info
= &surf
->u
.legacy
.stencil_level
[level
];
6411 z_offs
+= surf
->u
.legacy
.level
[level
].offset
;
6412 s_offs
+= surf
->u
.legacy
.stencil_level
[level
].offset
;
6414 ds
->db_depth_info
= S_02803C_ADDR5_SWIZZLE_MASK(!radv_image_is_tc_compat_htile(iview
->image
));
6415 ds
->db_z_info
= S_028040_FORMAT(format
) | S_028040_ZRANGE_PRECISION(1);
6416 ds
->db_stencil_info
= S_028044_FORMAT(stencil_format
);
6418 if (iview
->image
->info
.samples
> 1)
6419 ds
->db_z_info
|= S_028040_NUM_SAMPLES(util_logbase2(iview
->image
->info
.samples
));
6421 if (device
->physical_device
->rad_info
.chip_class
>= GFX7
) {
6422 struct radeon_info
*info
= &device
->physical_device
->rad_info
;
6423 unsigned tiling_index
= surf
->u
.legacy
.tiling_index
[level
];
6424 unsigned stencil_index
= surf
->u
.legacy
.stencil_tiling_index
[level
];
6425 unsigned macro_index
= surf
->u
.legacy
.macro_tile_index
;
6426 unsigned tile_mode
= info
->si_tile_mode_array
[tiling_index
];
6427 unsigned stencil_tile_mode
= info
->si_tile_mode_array
[stencil_index
];
6428 unsigned macro_mode
= info
->cik_macrotile_mode_array
[macro_index
];
6431 tile_mode
= stencil_tile_mode
;
6433 ds
->db_depth_info
|=
6434 S_02803C_ARRAY_MODE(G_009910_ARRAY_MODE(tile_mode
)) |
6435 S_02803C_PIPE_CONFIG(G_009910_PIPE_CONFIG(tile_mode
)) |
6436 S_02803C_BANK_WIDTH(G_009990_BANK_WIDTH(macro_mode
)) |
6437 S_02803C_BANK_HEIGHT(G_009990_BANK_HEIGHT(macro_mode
)) |
6438 S_02803C_MACRO_TILE_ASPECT(G_009990_MACRO_TILE_ASPECT(macro_mode
)) |
6439 S_02803C_NUM_BANKS(G_009990_NUM_BANKS(macro_mode
));
6440 ds
->db_z_info
|= S_028040_TILE_SPLIT(G_009910_TILE_SPLIT(tile_mode
));
6441 ds
->db_stencil_info
|= S_028044_TILE_SPLIT(G_009910_TILE_SPLIT(stencil_tile_mode
));
6443 unsigned tile_mode_index
= si_tile_mode_index(&iview
->image
->planes
[0], level
, false);
6444 ds
->db_z_info
|= S_028040_TILE_MODE_INDEX(tile_mode_index
);
6445 tile_mode_index
= si_tile_mode_index(&iview
->image
->planes
[0], level
, true);
6446 ds
->db_stencil_info
|= S_028044_TILE_MODE_INDEX(tile_mode_index
);
6448 ds
->db_z_info
|= S_028040_TILE_MODE_INDEX(tile_mode_index
);
6451 ds
->db_depth_size
= S_028058_PITCH_TILE_MAX((level_info
->nblk_x
/ 8) - 1) |
6452 S_028058_HEIGHT_TILE_MAX((level_info
->nblk_y
/ 8) - 1);
6453 ds
->db_depth_slice
= S_02805C_SLICE_TILE_MAX((level_info
->nblk_x
* level_info
->nblk_y
) / 64 - 1);
6455 if (radv_htile_enabled(iview
->image
, level
)) {
6456 ds
->db_z_info
|= S_028040_TILE_SURFACE_ENABLE(1);
6458 if (!surf
->has_stencil
&&
6459 !radv_image_is_tc_compat_htile(iview
->image
))
6460 /* Use all of the htile_buffer for depth if there's no stencil. */
6461 ds
->db_stencil_info
|= S_028044_TILE_STENCIL_DISABLE(1);
6463 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
+
6464 iview
->image
->htile_offset
;
6465 ds
->db_htile_data_base
= va
>> 8;
6466 ds
->db_htile_surface
= S_028ABC_FULL_CACHE(1);
6468 if (radv_image_is_tc_compat_htile(iview
->image
)) {
6469 unsigned max_zplanes
=
6470 radv_calc_decompress_on_z_planes(device
, iview
);
6472 ds
->db_htile_surface
|= S_028ABC_TC_COMPATIBLE(1);
6473 ds
->db_z_info
|= S_028040_DECOMPRESS_ON_N_ZPLANES(max_zplanes
);
6478 ds
->db_z_read_base
= ds
->db_z_write_base
= z_offs
>> 8;
6479 ds
->db_stencil_read_base
= ds
->db_stencil_write_base
= s_offs
>> 8;
6482 VkResult
radv_CreateFramebuffer(
6484 const VkFramebufferCreateInfo
* pCreateInfo
,
6485 const VkAllocationCallbacks
* pAllocator
,
6486 VkFramebuffer
* pFramebuffer
)
6488 RADV_FROM_HANDLE(radv_device
, device
, _device
);
6489 struct radv_framebuffer
*framebuffer
;
6490 const VkFramebufferAttachmentsCreateInfoKHR
*imageless_create_info
=
6491 vk_find_struct_const(pCreateInfo
->pNext
,
6492 FRAMEBUFFER_ATTACHMENTS_CREATE_INFO_KHR
);
6494 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
6496 size_t size
= sizeof(*framebuffer
);
6497 if (!imageless_create_info
)
6498 size
+= sizeof(struct radv_image_view
*) * pCreateInfo
->attachmentCount
;
6499 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
6500 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
6501 if (framebuffer
== NULL
)
6502 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
6504 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
6505 framebuffer
->width
= pCreateInfo
->width
;
6506 framebuffer
->height
= pCreateInfo
->height
;
6507 framebuffer
->layers
= pCreateInfo
->layers
;
6508 if (imageless_create_info
) {
6509 for (unsigned i
= 0; i
< imageless_create_info
->attachmentImageInfoCount
; ++i
) {
6510 const VkFramebufferAttachmentImageInfoKHR
*attachment
=
6511 imageless_create_info
->pAttachmentImageInfos
+ i
;
6512 framebuffer
->width
= MIN2(framebuffer
->width
, attachment
->width
);
6513 framebuffer
->height
= MIN2(framebuffer
->height
, attachment
->height
);
6514 framebuffer
->layers
= MIN2(framebuffer
->layers
, attachment
->layerCount
);
6517 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
6518 VkImageView _iview
= pCreateInfo
->pAttachments
[i
];
6519 struct radv_image_view
*iview
= radv_image_view_from_handle(_iview
);
6520 framebuffer
->attachments
[i
] = iview
;
6521 framebuffer
->width
= MIN2(framebuffer
->width
, iview
->extent
.width
);
6522 framebuffer
->height
= MIN2(framebuffer
->height
, iview
->extent
.height
);
6523 framebuffer
->layers
= MIN2(framebuffer
->layers
, radv_surface_max_layer_count(iview
));
6527 *pFramebuffer
= radv_framebuffer_to_handle(framebuffer
);
6531 void radv_DestroyFramebuffer(
6534 const VkAllocationCallbacks
* pAllocator
)
6536 RADV_FROM_HANDLE(radv_device
, device
, _device
);
6537 RADV_FROM_HANDLE(radv_framebuffer
, fb
, _fb
);
6541 vk_free2(&device
->alloc
, pAllocator
, fb
);
6544 static unsigned radv_tex_wrap(VkSamplerAddressMode address_mode
)
6546 switch (address_mode
) {
6547 case VK_SAMPLER_ADDRESS_MODE_REPEAT
:
6548 return V_008F30_SQ_TEX_WRAP
;
6549 case VK_SAMPLER_ADDRESS_MODE_MIRRORED_REPEAT
:
6550 return V_008F30_SQ_TEX_MIRROR
;
6551 case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE
:
6552 return V_008F30_SQ_TEX_CLAMP_LAST_TEXEL
;
6553 case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_BORDER
:
6554 return V_008F30_SQ_TEX_CLAMP_BORDER
;
6555 case VK_SAMPLER_ADDRESS_MODE_MIRROR_CLAMP_TO_EDGE
:
6556 return V_008F30_SQ_TEX_MIRROR_ONCE_LAST_TEXEL
;
6558 unreachable("illegal tex wrap mode");
6564 radv_tex_compare(VkCompareOp op
)
6567 case VK_COMPARE_OP_NEVER
:
6568 return V_008F30_SQ_TEX_DEPTH_COMPARE_NEVER
;
6569 case VK_COMPARE_OP_LESS
:
6570 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESS
;
6571 case VK_COMPARE_OP_EQUAL
:
6572 return V_008F30_SQ_TEX_DEPTH_COMPARE_EQUAL
;
6573 case VK_COMPARE_OP_LESS_OR_EQUAL
:
6574 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESSEQUAL
;
6575 case VK_COMPARE_OP_GREATER
:
6576 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATER
;
6577 case VK_COMPARE_OP_NOT_EQUAL
:
6578 return V_008F30_SQ_TEX_DEPTH_COMPARE_NOTEQUAL
;
6579 case VK_COMPARE_OP_GREATER_OR_EQUAL
:
6580 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATEREQUAL
;
6581 case VK_COMPARE_OP_ALWAYS
:
6582 return V_008F30_SQ_TEX_DEPTH_COMPARE_ALWAYS
;
6584 unreachable("illegal compare mode");
6590 radv_tex_filter(VkFilter filter
, unsigned max_ansio
)
6593 case VK_FILTER_NEAREST
:
6594 return (max_ansio
> 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_POINT
:
6595 V_008F38_SQ_TEX_XY_FILTER_POINT
);
6596 case VK_FILTER_LINEAR
:
6597 return (max_ansio
> 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_BILINEAR
:
6598 V_008F38_SQ_TEX_XY_FILTER_BILINEAR
);
6599 case VK_FILTER_CUBIC_IMG
:
6601 fprintf(stderr
, "illegal texture filter");
6607 radv_tex_mipfilter(VkSamplerMipmapMode mode
)
6610 case VK_SAMPLER_MIPMAP_MODE_NEAREST
:
6611 return V_008F38_SQ_TEX_Z_FILTER_POINT
;
6612 case VK_SAMPLER_MIPMAP_MODE_LINEAR
:
6613 return V_008F38_SQ_TEX_Z_FILTER_LINEAR
;
6615 return V_008F38_SQ_TEX_Z_FILTER_NONE
;
6620 radv_tex_bordercolor(VkBorderColor bcolor
)
6623 case VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
:
6624 case VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
:
6625 return V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK
;
6626 case VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
:
6627 case VK_BORDER_COLOR_INT_OPAQUE_BLACK
:
6628 return V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_BLACK
;
6629 case VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
:
6630 case VK_BORDER_COLOR_INT_OPAQUE_WHITE
:
6631 return V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_WHITE
;
6639 radv_tex_aniso_filter(unsigned filter
)
6653 radv_tex_filter_mode(VkSamplerReductionModeEXT mode
)
6656 case VK_SAMPLER_REDUCTION_MODE_WEIGHTED_AVERAGE_EXT
:
6657 return V_008F30_SQ_IMG_FILTER_MODE_BLEND
;
6658 case VK_SAMPLER_REDUCTION_MODE_MIN_EXT
:
6659 return V_008F30_SQ_IMG_FILTER_MODE_MIN
;
6660 case VK_SAMPLER_REDUCTION_MODE_MAX_EXT
:
6661 return V_008F30_SQ_IMG_FILTER_MODE_MAX
;
6669 radv_get_max_anisotropy(struct radv_device
*device
,
6670 const VkSamplerCreateInfo
*pCreateInfo
)
6672 if (device
->force_aniso
>= 0)
6673 return device
->force_aniso
;
6675 if (pCreateInfo
->anisotropyEnable
&&
6676 pCreateInfo
->maxAnisotropy
> 1.0f
)
6677 return (uint32_t)pCreateInfo
->maxAnisotropy
;
6683 radv_init_sampler(struct radv_device
*device
,
6684 struct radv_sampler
*sampler
,
6685 const VkSamplerCreateInfo
*pCreateInfo
)
6687 uint32_t max_aniso
= radv_get_max_anisotropy(device
, pCreateInfo
);
6688 uint32_t max_aniso_ratio
= radv_tex_aniso_filter(max_aniso
);
6689 bool compat_mode
= device
->physical_device
->rad_info
.chip_class
== GFX8
||
6690 device
->physical_device
->rad_info
.chip_class
== GFX9
;
6691 unsigned filter_mode
= V_008F30_SQ_IMG_FILTER_MODE_BLEND
;
6693 const struct VkSamplerReductionModeCreateInfoEXT
*sampler_reduction
=
6694 vk_find_struct_const(pCreateInfo
->pNext
,
6695 SAMPLER_REDUCTION_MODE_CREATE_INFO_EXT
);
6696 if (sampler_reduction
)
6697 filter_mode
= radv_tex_filter_mode(sampler_reduction
->reductionMode
);
6699 sampler
->state
[0] = (S_008F30_CLAMP_X(radv_tex_wrap(pCreateInfo
->addressModeU
)) |
6700 S_008F30_CLAMP_Y(radv_tex_wrap(pCreateInfo
->addressModeV
)) |
6701 S_008F30_CLAMP_Z(radv_tex_wrap(pCreateInfo
->addressModeW
)) |
6702 S_008F30_MAX_ANISO_RATIO(max_aniso_ratio
) |
6703 S_008F30_DEPTH_COMPARE_FUNC(radv_tex_compare(pCreateInfo
->compareOp
)) |
6704 S_008F30_FORCE_UNNORMALIZED(pCreateInfo
->unnormalizedCoordinates
? 1 : 0) |
6705 S_008F30_ANISO_THRESHOLD(max_aniso_ratio
>> 1) |
6706 S_008F30_ANISO_BIAS(max_aniso_ratio
) |
6707 S_008F30_DISABLE_CUBE_WRAP(0) |
6708 S_008F30_COMPAT_MODE(compat_mode
) |
6709 S_008F30_FILTER_MODE(filter_mode
));
6710 sampler
->state
[1] = (S_008F34_MIN_LOD(S_FIXED(CLAMP(pCreateInfo
->minLod
, 0, 15), 8)) |
6711 S_008F34_MAX_LOD(S_FIXED(CLAMP(pCreateInfo
->maxLod
, 0, 15), 8)) |
6712 S_008F34_PERF_MIP(max_aniso_ratio
? max_aniso_ratio
+ 6 : 0));
6713 sampler
->state
[2] = (S_008F38_LOD_BIAS(S_FIXED(CLAMP(pCreateInfo
->mipLodBias
, -16, 16), 8)) |
6714 S_008F38_XY_MAG_FILTER(radv_tex_filter(pCreateInfo
->magFilter
, max_aniso
)) |
6715 S_008F38_XY_MIN_FILTER(radv_tex_filter(pCreateInfo
->minFilter
, max_aniso
)) |
6716 S_008F38_MIP_FILTER(radv_tex_mipfilter(pCreateInfo
->mipmapMode
)) |
6717 S_008F38_MIP_POINT_PRECLAMP(0));
6718 sampler
->state
[3] = (S_008F3C_BORDER_COLOR_PTR(0) |
6719 S_008F3C_BORDER_COLOR_TYPE(radv_tex_bordercolor(pCreateInfo
->borderColor
)));
6721 if (device
->physical_device
->rad_info
.chip_class
>= GFX10
) {
6722 sampler
->state
[2] |= S_008F38_ANISO_OVERRIDE_GFX10(1);
6724 sampler
->state
[2] |=
6725 S_008F38_DISABLE_LSB_CEIL(device
->physical_device
->rad_info
.chip_class
<= GFX8
) |
6726 S_008F38_FILTER_PREC_FIX(1) |
6727 S_008F38_ANISO_OVERRIDE_GFX6(device
->physical_device
->rad_info
.chip_class
>= GFX8
);
6731 VkResult
radv_CreateSampler(
6733 const VkSamplerCreateInfo
* pCreateInfo
,
6734 const VkAllocationCallbacks
* pAllocator
,
6735 VkSampler
* pSampler
)
6737 RADV_FROM_HANDLE(radv_device
, device
, _device
);
6738 struct radv_sampler
*sampler
;
6740 const struct VkSamplerYcbcrConversionInfo
*ycbcr_conversion
=
6741 vk_find_struct_const(pCreateInfo
->pNext
,
6742 SAMPLER_YCBCR_CONVERSION_INFO
);
6744 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO
);
6746 sampler
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*sampler
), 8,
6747 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
6749 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
6751 radv_init_sampler(device
, sampler
, pCreateInfo
);
6753 sampler
->ycbcr_sampler
= ycbcr_conversion
? radv_sampler_ycbcr_conversion_from_handle(ycbcr_conversion
->conversion
): NULL
;
6754 *pSampler
= radv_sampler_to_handle(sampler
);
6759 void radv_DestroySampler(
6762 const VkAllocationCallbacks
* pAllocator
)
6764 RADV_FROM_HANDLE(radv_device
, device
, _device
);
6765 RADV_FROM_HANDLE(radv_sampler
, sampler
, _sampler
);
6769 vk_free2(&device
->alloc
, pAllocator
, sampler
);
6772 /* vk_icd.h does not declare this function, so we declare it here to
6773 * suppress Wmissing-prototypes.
6775 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
6776 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t *pSupportedVersion
);
6778 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
6779 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t *pSupportedVersion
)
6781 /* For the full details on loader interface versioning, see
6782 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
6783 * What follows is a condensed summary, to help you navigate the large and
6784 * confusing official doc.
6786 * - Loader interface v0 is incompatible with later versions. We don't
6789 * - In loader interface v1:
6790 * - The first ICD entrypoint called by the loader is
6791 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
6793 * - The ICD must statically expose no other Vulkan symbol unless it is
6794 * linked with -Bsymbolic.
6795 * - Each dispatchable Vulkan handle created by the ICD must be
6796 * a pointer to a struct whose first member is VK_LOADER_DATA. The
6797 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
6798 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
6799 * vkDestroySurfaceKHR(). The ICD must be capable of working with
6800 * such loader-managed surfaces.
6802 * - Loader interface v2 differs from v1 in:
6803 * - The first ICD entrypoint called by the loader is
6804 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
6805 * statically expose this entrypoint.
6807 * - Loader interface v3 differs from v2 in:
6808 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
6809 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
6810 * because the loader no longer does so.
6812 *pSupportedVersion
= MIN2(*pSupportedVersion
, 4u);
6816 VkResult
radv_GetMemoryFdKHR(VkDevice _device
,
6817 const VkMemoryGetFdInfoKHR
*pGetFdInfo
,
6820 RADV_FROM_HANDLE(radv_device
, device
, _device
);
6821 RADV_FROM_HANDLE(radv_device_memory
, memory
, pGetFdInfo
->memory
);
6823 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
6825 /* At the moment, we support only the below handle types. */
6826 assert(pGetFdInfo
->handleType
==
6827 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
6828 pGetFdInfo
->handleType
==
6829 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
6831 bool ret
= radv_get_memory_fd(device
, memory
, pFD
);
6833 return vk_error(device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
6837 VkResult
radv_GetMemoryFdPropertiesKHR(VkDevice _device
,
6838 VkExternalMemoryHandleTypeFlagBits handleType
,
6840 VkMemoryFdPropertiesKHR
*pMemoryFdProperties
)
6842 RADV_FROM_HANDLE(radv_device
, device
, _device
);
6844 switch (handleType
) {
6845 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
:
6846 pMemoryFdProperties
->memoryTypeBits
= (1 << RADV_MEM_TYPE_COUNT
) - 1;
6850 /* The valid usage section for this function says:
6852 * "handleType must not be one of the handle types defined as
6855 * So opaque handle types fall into the default "unsupported" case.
6857 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE
);
6861 static VkResult
radv_import_opaque_fd(struct radv_device
*device
,
6865 uint32_t syncobj_handle
= 0;
6866 int ret
= device
->ws
->import_syncobj(device
->ws
, fd
, &syncobj_handle
);
6868 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE
);
6871 device
->ws
->destroy_syncobj(device
->ws
, *syncobj
);
6873 *syncobj
= syncobj_handle
;
6879 static VkResult
radv_import_sync_fd(struct radv_device
*device
,
6883 /* If we create a syncobj we do it locally so that if we have an error, we don't
6884 * leave a syncobj in an undetermined state in the fence. */
6885 uint32_t syncobj_handle
= *syncobj
;
6886 if (!syncobj_handle
) {
6887 int ret
= device
->ws
->create_syncobj(device
->ws
, &syncobj_handle
);
6889 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE
);
6894 device
->ws
->signal_syncobj(device
->ws
, syncobj_handle
);
6896 int ret
= device
->ws
->import_syncobj_from_sync_file(device
->ws
, syncobj_handle
, fd
);
6898 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE
);
6901 *syncobj
= syncobj_handle
;
6908 VkResult
radv_ImportSemaphoreFdKHR(VkDevice _device
,
6909 const VkImportSemaphoreFdInfoKHR
*pImportSemaphoreFdInfo
)
6911 RADV_FROM_HANDLE(radv_device
, device
, _device
);
6912 RADV_FROM_HANDLE(radv_semaphore
, sem
, pImportSemaphoreFdInfo
->semaphore
);
6914 struct radv_semaphore_part
*dst
= NULL
;
6916 if (pImportSemaphoreFdInfo
->flags
& VK_SEMAPHORE_IMPORT_TEMPORARY_BIT
) {
6917 dst
= &sem
->temporary
;
6919 dst
= &sem
->permanent
;
6922 uint32_t syncobj
= dst
->kind
== RADV_SEMAPHORE_SYNCOBJ
? dst
->syncobj
: 0;
6924 switch(pImportSemaphoreFdInfo
->handleType
) {
6925 case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT
:
6926 result
= radv_import_opaque_fd(device
, pImportSemaphoreFdInfo
->fd
, &syncobj
);
6928 case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT
:
6929 result
= radv_import_sync_fd(device
, pImportSemaphoreFdInfo
->fd
, &syncobj
);
6932 unreachable("Unhandled semaphore handle type");
6935 if (result
== VK_SUCCESS
) {
6936 dst
->syncobj
= syncobj
;
6937 dst
->kind
= RADV_SEMAPHORE_SYNCOBJ
;
6943 VkResult
radv_GetSemaphoreFdKHR(VkDevice _device
,
6944 const VkSemaphoreGetFdInfoKHR
*pGetFdInfo
,
6947 RADV_FROM_HANDLE(radv_device
, device
, _device
);
6948 RADV_FROM_HANDLE(radv_semaphore
, sem
, pGetFdInfo
->semaphore
);
6950 uint32_t syncobj_handle
;
6952 if (sem
->temporary
.kind
!= RADV_SEMAPHORE_NONE
) {
6953 assert(sem
->temporary
.kind
== RADV_SEMAPHORE_SYNCOBJ
);
6954 syncobj_handle
= sem
->temporary
.syncobj
;
6956 assert(sem
->permanent
.kind
== RADV_SEMAPHORE_SYNCOBJ
);
6957 syncobj_handle
= sem
->permanent
.syncobj
;
6960 switch(pGetFdInfo
->handleType
) {
6961 case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT
:
6962 ret
= device
->ws
->export_syncobj(device
->ws
, syncobj_handle
, pFd
);
6964 case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT
:
6965 ret
= device
->ws
->export_syncobj_to_sync_file(device
->ws
, syncobj_handle
, pFd
);
6967 if (sem
->temporary
.kind
!= RADV_SEMAPHORE_NONE
) {
6968 radv_destroy_semaphore_part(device
, &sem
->temporary
);
6970 device
->ws
->reset_syncobj(device
->ws
, syncobj_handle
);
6975 unreachable("Unhandled semaphore handle type");
6979 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE
);
6983 void radv_GetPhysicalDeviceExternalSemaphoreProperties(
6984 VkPhysicalDevice physicalDevice
,
6985 const VkPhysicalDeviceExternalSemaphoreInfo
*pExternalSemaphoreInfo
,
6986 VkExternalSemaphoreProperties
*pExternalSemaphoreProperties
)
6988 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
6989 VkSemaphoreTypeKHR type
= radv_get_semaphore_type(pExternalSemaphoreInfo
->pNext
, NULL
);
6991 if (type
== VK_SEMAPHORE_TYPE_TIMELINE_KHR
) {
6992 pExternalSemaphoreProperties
->exportFromImportedHandleTypes
= 0;
6993 pExternalSemaphoreProperties
->compatibleHandleTypes
= 0;
6994 pExternalSemaphoreProperties
->externalSemaphoreFeatures
= 0;
6996 /* Require has_syncobj_wait_for_submit for the syncobj signal ioctl introduced at virtually the same time */
6997 } else if (pdevice
->rad_info
.has_syncobj_wait_for_submit
&&
6998 (pExternalSemaphoreInfo
->handleType
== VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT
||
6999 pExternalSemaphoreInfo
->handleType
== VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT
)) {
7000 pExternalSemaphoreProperties
->exportFromImportedHandleTypes
= VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT
| VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT
;
7001 pExternalSemaphoreProperties
->compatibleHandleTypes
= VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT
| VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT
;
7002 pExternalSemaphoreProperties
->externalSemaphoreFeatures
= VK_EXTERNAL_SEMAPHORE_FEATURE_EXPORTABLE_BIT
|
7003 VK_EXTERNAL_SEMAPHORE_FEATURE_IMPORTABLE_BIT
;
7004 } else if (pExternalSemaphoreInfo
->handleType
== VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT
) {
7005 pExternalSemaphoreProperties
->exportFromImportedHandleTypes
= VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT
;
7006 pExternalSemaphoreProperties
->compatibleHandleTypes
= VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT
;
7007 pExternalSemaphoreProperties
->externalSemaphoreFeatures
= VK_EXTERNAL_SEMAPHORE_FEATURE_EXPORTABLE_BIT
|
7008 VK_EXTERNAL_SEMAPHORE_FEATURE_IMPORTABLE_BIT
;
7010 pExternalSemaphoreProperties
->exportFromImportedHandleTypes
= 0;
7011 pExternalSemaphoreProperties
->compatibleHandleTypes
= 0;
7012 pExternalSemaphoreProperties
->externalSemaphoreFeatures
= 0;
7016 VkResult
radv_ImportFenceFdKHR(VkDevice _device
,
7017 const VkImportFenceFdInfoKHR
*pImportFenceFdInfo
)
7019 RADV_FROM_HANDLE(radv_device
, device
, _device
);
7020 RADV_FROM_HANDLE(radv_fence
, fence
, pImportFenceFdInfo
->fence
);
7021 uint32_t *syncobj_dst
= NULL
;
7024 if (pImportFenceFdInfo
->flags
& VK_FENCE_IMPORT_TEMPORARY_BIT
) {
7025 syncobj_dst
= &fence
->temp_syncobj
;
7027 syncobj_dst
= &fence
->syncobj
;
7030 switch(pImportFenceFdInfo
->handleType
) {
7031 case VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT
:
7032 return radv_import_opaque_fd(device
, pImportFenceFdInfo
->fd
, syncobj_dst
);
7033 case VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT
:
7034 return radv_import_sync_fd(device
, pImportFenceFdInfo
->fd
, syncobj_dst
);
7036 unreachable("Unhandled fence handle type");
7040 VkResult
radv_GetFenceFdKHR(VkDevice _device
,
7041 const VkFenceGetFdInfoKHR
*pGetFdInfo
,
7044 RADV_FROM_HANDLE(radv_device
, device
, _device
);
7045 RADV_FROM_HANDLE(radv_fence
, fence
, pGetFdInfo
->fence
);
7047 uint32_t syncobj_handle
;
7049 if (fence
->temp_syncobj
)
7050 syncobj_handle
= fence
->temp_syncobj
;
7052 syncobj_handle
= fence
->syncobj
;
7054 switch(pGetFdInfo
->handleType
) {
7055 case VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT
:
7056 ret
= device
->ws
->export_syncobj(device
->ws
, syncobj_handle
, pFd
);
7058 case VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT
:
7059 ret
= device
->ws
->export_syncobj_to_sync_file(device
->ws
, syncobj_handle
, pFd
);
7061 if (fence
->temp_syncobj
) {
7062 close (fence
->temp_syncobj
);
7063 fence
->temp_syncobj
= 0;
7065 device
->ws
->reset_syncobj(device
->ws
, syncobj_handle
);
7070 unreachable("Unhandled fence handle type");
7074 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE
);
7078 void radv_GetPhysicalDeviceExternalFenceProperties(
7079 VkPhysicalDevice physicalDevice
,
7080 const VkPhysicalDeviceExternalFenceInfo
*pExternalFenceInfo
,
7081 VkExternalFenceProperties
*pExternalFenceProperties
)
7083 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
7085 if (pdevice
->rad_info
.has_syncobj_wait_for_submit
&&
7086 (pExternalFenceInfo
->handleType
== VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT
||
7087 pExternalFenceInfo
->handleType
== VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT
)) {
7088 pExternalFenceProperties
->exportFromImportedHandleTypes
= VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT
| VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT
;
7089 pExternalFenceProperties
->compatibleHandleTypes
= VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT
| VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT
;
7090 pExternalFenceProperties
->externalFenceFeatures
= VK_EXTERNAL_FENCE_FEATURE_EXPORTABLE_BIT
|
7091 VK_EXTERNAL_SEMAPHORE_FEATURE_IMPORTABLE_BIT
;
7093 pExternalFenceProperties
->exportFromImportedHandleTypes
= 0;
7094 pExternalFenceProperties
->compatibleHandleTypes
= 0;
7095 pExternalFenceProperties
->externalFenceFeatures
= 0;
7100 radv_CreateDebugReportCallbackEXT(VkInstance _instance
,
7101 const VkDebugReportCallbackCreateInfoEXT
* pCreateInfo
,
7102 const VkAllocationCallbacks
* pAllocator
,
7103 VkDebugReportCallbackEXT
* pCallback
)
7105 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
7106 return vk_create_debug_report_callback(&instance
->debug_report_callbacks
,
7107 pCreateInfo
, pAllocator
, &instance
->alloc
,
7112 radv_DestroyDebugReportCallbackEXT(VkInstance _instance
,
7113 VkDebugReportCallbackEXT _callback
,
7114 const VkAllocationCallbacks
* pAllocator
)
7116 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
7117 vk_destroy_debug_report_callback(&instance
->debug_report_callbacks
,
7118 _callback
, pAllocator
, &instance
->alloc
);
7122 radv_DebugReportMessageEXT(VkInstance _instance
,
7123 VkDebugReportFlagsEXT flags
,
7124 VkDebugReportObjectTypeEXT objectType
,
7127 int32_t messageCode
,
7128 const char* pLayerPrefix
,
7129 const char* pMessage
)
7131 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
7132 vk_debug_report(&instance
->debug_report_callbacks
, flags
, objectType
,
7133 object
, location
, messageCode
, pLayerPrefix
, pMessage
);
7137 radv_GetDeviceGroupPeerMemoryFeatures(
7140 uint32_t localDeviceIndex
,
7141 uint32_t remoteDeviceIndex
,
7142 VkPeerMemoryFeatureFlags
* pPeerMemoryFeatures
)
7144 assert(localDeviceIndex
== remoteDeviceIndex
);
7146 *pPeerMemoryFeatures
= VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT
|
7147 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT
|
7148 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT
|
7149 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT
;
7152 static const VkTimeDomainEXT radv_time_domains
[] = {
7153 VK_TIME_DOMAIN_DEVICE_EXT
,
7154 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
,
7155 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
,
7158 VkResult
radv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
7159 VkPhysicalDevice physicalDevice
,
7160 uint32_t *pTimeDomainCount
,
7161 VkTimeDomainEXT
*pTimeDomains
)
7164 VK_OUTARRAY_MAKE(out
, pTimeDomains
, pTimeDomainCount
);
7166 for (d
= 0; d
< ARRAY_SIZE(radv_time_domains
); d
++) {
7167 vk_outarray_append(&out
, i
) {
7168 *i
= radv_time_domains
[d
];
7172 return vk_outarray_status(&out
);
7176 radv_clock_gettime(clockid_t clock_id
)
7178 struct timespec current
;
7181 ret
= clock_gettime(clock_id
, ¤t
);
7182 if (ret
< 0 && clock_id
== CLOCK_MONOTONIC_RAW
)
7183 ret
= clock_gettime(CLOCK_MONOTONIC
, ¤t
);
7187 return (uint64_t) current
.tv_sec
* 1000000000ULL + current
.tv_nsec
;
7190 VkResult
radv_GetCalibratedTimestampsEXT(
7192 uint32_t timestampCount
,
7193 const VkCalibratedTimestampInfoEXT
*pTimestampInfos
,
7194 uint64_t *pTimestamps
,
7195 uint64_t *pMaxDeviation
)
7197 RADV_FROM_HANDLE(radv_device
, device
, _device
);
7198 uint32_t clock_crystal_freq
= device
->physical_device
->rad_info
.clock_crystal_freq
;
7200 uint64_t begin
, end
;
7201 uint64_t max_clock_period
= 0;
7203 begin
= radv_clock_gettime(CLOCK_MONOTONIC_RAW
);
7205 for (d
= 0; d
< timestampCount
; d
++) {
7206 switch (pTimestampInfos
[d
].timeDomain
) {
7207 case VK_TIME_DOMAIN_DEVICE_EXT
:
7208 pTimestamps
[d
] = device
->ws
->query_value(device
->ws
,
7210 uint64_t device_period
= DIV_ROUND_UP(1000000, clock_crystal_freq
);
7211 max_clock_period
= MAX2(max_clock_period
, device_period
);
7213 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
:
7214 pTimestamps
[d
] = radv_clock_gettime(CLOCK_MONOTONIC
);
7215 max_clock_period
= MAX2(max_clock_period
, 1);
7218 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
:
7219 pTimestamps
[d
] = begin
;
7227 end
= radv_clock_gettime(CLOCK_MONOTONIC_RAW
);
7230 * The maximum deviation is the sum of the interval over which we
7231 * perform the sampling and the maximum period of any sampled
7232 * clock. That's because the maximum skew between any two sampled
7233 * clock edges is when the sampled clock with the largest period is
7234 * sampled at the end of that period but right at the beginning of the
7235 * sampling interval and some other clock is sampled right at the
7236 * begining of its sampling period and right at the end of the
7237 * sampling interval. Let's assume the GPU has the longest clock
7238 * period and that the application is sampling GPU and monotonic:
7241 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
7242 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
7246 * GPU -----_____-----_____-----_____-----_____
7249 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
7250 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
7252 * Interval <----------------->
7253 * Deviation <-------------------------->
7257 * m = read(monotonic) 2
7260 * We round the sample interval up by one tick to cover sampling error
7261 * in the interval clock
7264 uint64_t sample_interval
= end
- begin
+ 1;
7266 *pMaxDeviation
= sample_interval
+ max_clock_period
;
7271 void radv_GetPhysicalDeviceMultisamplePropertiesEXT(
7272 VkPhysicalDevice physicalDevice
,
7273 VkSampleCountFlagBits samples
,
7274 VkMultisamplePropertiesEXT
* pMultisampleProperties
)
7276 if (samples
& (VK_SAMPLE_COUNT_2_BIT
|
7277 VK_SAMPLE_COUNT_4_BIT
|
7278 VK_SAMPLE_COUNT_8_BIT
)) {
7279 pMultisampleProperties
->maxSampleLocationGridSize
= (VkExtent2D
){ 2, 2 };
7281 pMultisampleProperties
->maxSampleLocationGridSize
= (VkExtent2D
){ 0, 0 };