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
32 #include "radv_debug.h"
33 #include "radv_private.h"
34 #include "radv_shader.h"
36 #include "util/disk_cache.h"
37 #include "util/strtod.h"
41 #include <amdgpu_drm.h>
42 #include "winsys/amdgpu/radv_amdgpu_winsys_public.h"
43 #include "ac_llvm_util.h"
44 #include "vk_format.h"
47 #include "addrlib/gfx9/chip/gfx9_enum.h"
48 #include "util/debug.h"
51 radv_device_get_cache_uuid(enum radeon_family family
, void *uuid
)
53 uint32_t mesa_timestamp
, llvm_timestamp
;
55 memset(uuid
, 0, VK_UUID_SIZE
);
56 if (!disk_cache_get_function_timestamp(radv_device_get_cache_uuid
, &mesa_timestamp
) ||
57 !disk_cache_get_function_timestamp(LLVMInitializeAMDGPUTargetInfo
, &llvm_timestamp
))
60 memcpy(uuid
, &mesa_timestamp
, 4);
61 memcpy((char*)uuid
+ 4, &llvm_timestamp
, 4);
62 memcpy((char*)uuid
+ 8, &f
, 2);
63 snprintf((char*)uuid
+ 10, VK_UUID_SIZE
- 10, "radv");
68 radv_get_driver_uuid(void *uuid
)
70 ac_compute_driver_uuid(uuid
, VK_UUID_SIZE
);
74 radv_get_device_uuid(struct radeon_info
*info
, void *uuid
)
76 ac_compute_device_uuid(info
, uuid
, VK_UUID_SIZE
);
80 radv_get_device_name(enum radeon_family family
, char *name
, size_t name_len
)
82 const char *chip_string
;
83 char llvm_string
[32] = {};
86 case CHIP_TAHITI
: chip_string
= "AMD RADV TAHITI"; break;
87 case CHIP_PITCAIRN
: chip_string
= "AMD RADV PITCAIRN"; break;
88 case CHIP_VERDE
: chip_string
= "AMD RADV CAPE VERDE"; break;
89 case CHIP_OLAND
: chip_string
= "AMD RADV OLAND"; break;
90 case CHIP_HAINAN
: chip_string
= "AMD RADV HAINAN"; break;
91 case CHIP_BONAIRE
: chip_string
= "AMD RADV BONAIRE"; break;
92 case CHIP_KAVERI
: chip_string
= "AMD RADV KAVERI"; break;
93 case CHIP_KABINI
: chip_string
= "AMD RADV KABINI"; break;
94 case CHIP_HAWAII
: chip_string
= "AMD RADV HAWAII"; break;
95 case CHIP_MULLINS
: chip_string
= "AMD RADV MULLINS"; break;
96 case CHIP_TONGA
: chip_string
= "AMD RADV TONGA"; break;
97 case CHIP_ICELAND
: chip_string
= "AMD RADV ICELAND"; break;
98 case CHIP_CARRIZO
: chip_string
= "AMD RADV CARRIZO"; break;
99 case CHIP_FIJI
: chip_string
= "AMD RADV FIJI"; break;
100 case CHIP_POLARIS10
: chip_string
= "AMD RADV POLARIS10"; break;
101 case CHIP_POLARIS11
: chip_string
= "AMD RADV POLARIS11"; break;
102 case CHIP_POLARIS12
: chip_string
= "AMD RADV POLARIS12"; break;
103 case CHIP_STONEY
: chip_string
= "AMD RADV STONEY"; break;
104 case CHIP_VEGAM
: chip_string
= "AMD RADV VEGA M"; break;
105 case CHIP_VEGA10
: chip_string
= "AMD RADV VEGA10"; break;
106 case CHIP_VEGA12
: chip_string
= "AMD RADV VEGA12"; break;
107 case CHIP_RAVEN
: chip_string
= "AMD RADV RAVEN"; break;
108 default: chip_string
= "AMD RADV unknown"; break;
111 snprintf(llvm_string
, sizeof(llvm_string
),
112 " (LLVM %i.%i.%i)", (HAVE_LLVM
>> 8) & 0xff,
113 HAVE_LLVM
& 0xff, MESA_LLVM_VERSION_PATCH
);
114 snprintf(name
, name_len
, "%s%s", chip_string
, llvm_string
);
118 radv_physical_device_init_mem_types(struct radv_physical_device
*device
)
120 STATIC_ASSERT(RADV_MEM_HEAP_COUNT
<= VK_MAX_MEMORY_HEAPS
);
121 uint64_t visible_vram_size
= MIN2(device
->rad_info
.vram_size
,
122 device
->rad_info
.vram_vis_size
);
124 int vram_index
= -1, visible_vram_index
= -1, gart_index
= -1;
125 device
->memory_properties
.memoryHeapCount
= 0;
126 if (device
->rad_info
.vram_size
- visible_vram_size
> 0) {
127 vram_index
= device
->memory_properties
.memoryHeapCount
++;
128 device
->memory_properties
.memoryHeaps
[vram_index
] = (VkMemoryHeap
) {
129 .size
= device
->rad_info
.vram_size
- visible_vram_size
,
130 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
133 if (visible_vram_size
) {
134 visible_vram_index
= device
->memory_properties
.memoryHeapCount
++;
135 device
->memory_properties
.memoryHeaps
[visible_vram_index
] = (VkMemoryHeap
) {
136 .size
= visible_vram_size
,
137 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
140 if (device
->rad_info
.gart_size
> 0) {
141 gart_index
= device
->memory_properties
.memoryHeapCount
++;
142 device
->memory_properties
.memoryHeaps
[gart_index
] = (VkMemoryHeap
) {
143 .size
= device
->rad_info
.gart_size
,
144 .flags
= device
->rad_info
.has_dedicated_vram
? 0 : VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
148 STATIC_ASSERT(RADV_MEM_TYPE_COUNT
<= VK_MAX_MEMORY_TYPES
);
149 unsigned type_count
= 0;
150 if (vram_index
>= 0) {
151 device
->mem_type_indices
[type_count
] = RADV_MEM_TYPE_VRAM
;
152 device
->memory_properties
.memoryTypes
[type_count
++] = (VkMemoryType
) {
153 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
,
154 .heapIndex
= vram_index
,
157 if (gart_index
>= 0) {
158 device
->mem_type_indices
[type_count
] = RADV_MEM_TYPE_GTT_WRITE_COMBINE
;
159 device
->memory_properties
.memoryTypes
[type_count
++] = (VkMemoryType
) {
160 .propertyFlags
= VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
161 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
162 (device
->rad_info
.has_dedicated_vram
? 0 : VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
),
163 .heapIndex
= gart_index
,
166 if (visible_vram_index
>= 0) {
167 device
->mem_type_indices
[type_count
] = RADV_MEM_TYPE_VRAM_CPU_ACCESS
;
168 device
->memory_properties
.memoryTypes
[type_count
++] = (VkMemoryType
) {
169 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
170 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
171 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
172 .heapIndex
= visible_vram_index
,
175 if (gart_index
>= 0) {
176 device
->mem_type_indices
[type_count
] = RADV_MEM_TYPE_GTT_CACHED
;
177 device
->memory_properties
.memoryTypes
[type_count
++] = (VkMemoryType
) {
178 .propertyFlags
= VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
179 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
180 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
|
181 (device
->rad_info
.has_dedicated_vram
? 0 : VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
),
182 .heapIndex
= gart_index
,
185 device
->memory_properties
.memoryTypeCount
= type_count
;
189 radv_handle_env_var_force_family(struct radv_physical_device
*device
)
191 const char *family
= getenv("RADV_FORCE_FAMILY");
197 for (i
= CHIP_TAHITI
; i
< CHIP_LAST
; i
++) {
198 if (!strcmp(family
, ac_get_llvm_processor_name(i
))) {
199 /* Override family and chip_class. */
200 device
->rad_info
.family
= i
;
202 if (i
>= CHIP_VEGA10
)
203 device
->rad_info
.chip_class
= GFX9
;
204 else if (i
>= CHIP_TONGA
)
205 device
->rad_info
.chip_class
= VI
;
206 else if (i
>= CHIP_BONAIRE
)
207 device
->rad_info
.chip_class
= CIK
;
209 device
->rad_info
.chip_class
= SI
;
215 fprintf(stderr
, "radv: Unknown family: %s\n", family
);
220 radv_physical_device_init(struct radv_physical_device
*device
,
221 struct radv_instance
*instance
,
222 drmDevicePtr drm_device
)
224 const char *path
= drm_device
->nodes
[DRM_NODE_RENDER
];
226 drmVersionPtr version
;
230 fd
= open(path
, O_RDWR
| O_CLOEXEC
);
232 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
233 radv_logi("Could not open device '%s'", path
);
235 return vk_error(instance
, VK_ERROR_INCOMPATIBLE_DRIVER
);
238 version
= drmGetVersion(fd
);
242 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
243 radv_logi("Could not get the kernel driver version for device '%s'", path
);
245 return vk_errorf(instance
, VK_ERROR_INCOMPATIBLE_DRIVER
,
246 "failed to get version %s: %m", path
);
249 if (strcmp(version
->name
, "amdgpu")) {
250 drmFreeVersion(version
);
255 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
256 radv_logi("Device '%s' is not using the amdgpu kernel driver.", path
);
258 return VK_ERROR_INCOMPATIBLE_DRIVER
;
260 drmFreeVersion(version
);
262 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
263 radv_logi("Found compatible device '%s'.", path
);
265 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
266 device
->instance
= instance
;
267 assert(strlen(path
) < ARRAY_SIZE(device
->path
));
268 strncpy(device
->path
, path
, ARRAY_SIZE(device
->path
));
270 device
->ws
= radv_amdgpu_winsys_create(fd
, instance
->debug_flags
,
271 instance
->perftest_flags
);
273 result
= vk_error(instance
, VK_ERROR_INCOMPATIBLE_DRIVER
);
277 if (instance
->enabled_extensions
.KHR_display
) {
278 master_fd
= open(drm_device
->nodes
[DRM_NODE_PRIMARY
], O_RDWR
| O_CLOEXEC
);
279 if (master_fd
>= 0) {
280 uint32_t accel_working
= 0;
281 struct drm_amdgpu_info request
= {
282 .return_pointer
= (uintptr_t)&accel_working
,
283 .return_size
= sizeof(accel_working
),
284 .query
= AMDGPU_INFO_ACCEL_WORKING
287 if (drmCommandWrite(master_fd
, DRM_AMDGPU_INFO
, &request
, sizeof (struct drm_amdgpu_info
)) < 0 || !accel_working
) {
294 device
->master_fd
= master_fd
;
295 device
->local_fd
= fd
;
296 device
->ws
->query_info(device
->ws
, &device
->rad_info
);
298 radv_handle_env_var_force_family(device
);
300 radv_get_device_name(device
->rad_info
.family
, device
->name
, sizeof(device
->name
));
302 if (radv_device_get_cache_uuid(device
->rad_info
.family
, device
->cache_uuid
)) {
303 device
->ws
->destroy(device
->ws
);
304 result
= vk_errorf(instance
, VK_ERROR_INITIALIZATION_FAILED
,
305 "cannot generate UUID");
309 /* These flags affect shader compilation. */
310 uint64_t shader_env_flags
=
311 (device
->instance
->perftest_flags
& RADV_PERFTEST_SISCHED
? 0x1 : 0) |
312 (device
->instance
->debug_flags
& RADV_DEBUG_UNSAFE_MATH
? 0x2 : 0);
314 /* The gpu id is already embedded in the uuid so we just pass "radv"
315 * when creating the cache.
317 char buf
[VK_UUID_SIZE
* 2 + 1];
318 disk_cache_format_hex_id(buf
, device
->cache_uuid
, VK_UUID_SIZE
* 2);
319 device
->disk_cache
= disk_cache_create(device
->name
, buf
, shader_env_flags
);
321 if (device
->rad_info
.chip_class
< VI
||
322 device
->rad_info
.chip_class
> GFX9
)
323 fprintf(stderr
, "WARNING: radv is not a conformant vulkan implementation, testing use only.\n");
325 radv_get_driver_uuid(&device
->device_uuid
);
326 radv_get_device_uuid(&device
->rad_info
, &device
->device_uuid
);
328 if (device
->rad_info
.family
== CHIP_STONEY
||
329 device
->rad_info
.chip_class
>= GFX9
) {
330 device
->has_rbplus
= true;
331 device
->rbplus_allowed
= device
->rad_info
.family
== CHIP_STONEY
||
332 device
->rad_info
.family
== CHIP_VEGA12
||
333 device
->rad_info
.family
== CHIP_RAVEN
;
336 /* The mere presence of CLEAR_STATE in the IB causes random GPU hangs
339 device
->has_clear_state
= device
->rad_info
.chip_class
>= CIK
;
341 device
->cpdma_prefetch_writes_memory
= device
->rad_info
.chip_class
<= VI
;
343 /* Vega10/Raven need a special workaround for a hardware bug. */
344 device
->has_scissor_bug
= device
->rad_info
.family
== CHIP_VEGA10
||
345 device
->rad_info
.family
== CHIP_RAVEN
;
347 /* Out-of-order primitive rasterization. */
348 device
->has_out_of_order_rast
= device
->rad_info
.chip_class
>= VI
&&
349 device
->rad_info
.max_se
>= 2;
350 device
->out_of_order_rast_allowed
= device
->has_out_of_order_rast
&&
351 !(device
->instance
->debug_flags
& RADV_DEBUG_NO_OUT_OF_ORDER
);
353 device
->dcc_msaa_allowed
=
354 (device
->instance
->perftest_flags
& RADV_PERFTEST_DCC_MSAA
);
356 radv_physical_device_init_mem_types(device
);
357 radv_fill_device_extension_table(device
, &device
->supported_extensions
);
359 result
= radv_init_wsi(device
);
360 if (result
!= VK_SUCCESS
) {
361 device
->ws
->destroy(device
->ws
);
362 vk_error(instance
, result
);
366 if ((device
->instance
->debug_flags
& RADV_DEBUG_INFO
))
367 ac_print_gpu_info(&device
->rad_info
);
379 radv_physical_device_finish(struct radv_physical_device
*device
)
381 radv_finish_wsi(device
);
382 device
->ws
->destroy(device
->ws
);
383 disk_cache_destroy(device
->disk_cache
);
384 close(device
->local_fd
);
385 if (device
->master_fd
!= -1)
386 close(device
->master_fd
);
390 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
391 VkSystemAllocationScope allocationScope
)
397 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
398 size_t align
, VkSystemAllocationScope allocationScope
)
400 return realloc(pOriginal
, size
);
404 default_free_func(void *pUserData
, void *pMemory
)
409 static const VkAllocationCallbacks default_alloc
= {
411 .pfnAllocation
= default_alloc_func
,
412 .pfnReallocation
= default_realloc_func
,
413 .pfnFree
= default_free_func
,
416 static const struct debug_control radv_debug_options
[] = {
417 {"nofastclears", RADV_DEBUG_NO_FAST_CLEARS
},
418 {"nodcc", RADV_DEBUG_NO_DCC
},
419 {"shaders", RADV_DEBUG_DUMP_SHADERS
},
420 {"nocache", RADV_DEBUG_NO_CACHE
},
421 {"shaderstats", RADV_DEBUG_DUMP_SHADER_STATS
},
422 {"nohiz", RADV_DEBUG_NO_HIZ
},
423 {"nocompute", RADV_DEBUG_NO_COMPUTE_QUEUE
},
424 {"unsafemath", RADV_DEBUG_UNSAFE_MATH
},
425 {"allbos", RADV_DEBUG_ALL_BOS
},
426 {"noibs", RADV_DEBUG_NO_IBS
},
427 {"spirv", RADV_DEBUG_DUMP_SPIRV
},
428 {"vmfaults", RADV_DEBUG_VM_FAULTS
},
429 {"zerovram", RADV_DEBUG_ZERO_VRAM
},
430 {"syncshaders", RADV_DEBUG_SYNC_SHADERS
},
431 {"nosisched", RADV_DEBUG_NO_SISCHED
},
432 {"preoptir", RADV_DEBUG_PREOPTIR
},
433 {"nodynamicbounds", RADV_DEBUG_NO_DYNAMIC_BOUNDS
},
434 {"nooutoforder", RADV_DEBUG_NO_OUT_OF_ORDER
},
435 {"info", RADV_DEBUG_INFO
},
436 {"errors", RADV_DEBUG_ERRORS
},
437 {"startup", RADV_DEBUG_STARTUP
},
438 {"checkir", RADV_DEBUG_CHECKIR
},
439 {"nothreadllvm", RADV_DEBUG_NOTHREADLLVM
},
444 radv_get_debug_option_name(int id
)
446 assert(id
< ARRAY_SIZE(radv_debug_options
) - 1);
447 return radv_debug_options
[id
].string
;
450 static const struct debug_control radv_perftest_options
[] = {
451 {"nobatchchain", RADV_PERFTEST_NO_BATCHCHAIN
},
452 {"sisched", RADV_PERFTEST_SISCHED
},
453 {"localbos", RADV_PERFTEST_LOCAL_BOS
},
454 {"binning", RADV_PERFTEST_BINNING
},
455 {"dccmsaa", RADV_PERFTEST_DCC_MSAA
},
460 radv_get_perftest_option_name(int id
)
462 assert(id
< ARRAY_SIZE(radv_debug_options
) - 1);
463 return radv_perftest_options
[id
].string
;
467 radv_handle_per_app_options(struct radv_instance
*instance
,
468 const VkApplicationInfo
*info
)
470 const char *name
= info
? info
->pApplicationName
: NULL
;
475 if (!strcmp(name
, "Talos - Linux - 32bit") ||
476 !strcmp(name
, "Talos - Linux - 64bit")) {
477 if (!(instance
->debug_flags
& RADV_DEBUG_NO_SISCHED
)) {
478 /* Force enable LLVM sisched for Talos because it looks
479 * safe and it gives few more FPS.
481 instance
->perftest_flags
|= RADV_PERFTEST_SISCHED
;
486 static int radv_get_instance_extension_index(const char *name
)
488 for (unsigned i
= 0; i
< RADV_INSTANCE_EXTENSION_COUNT
; ++i
) {
489 if (strcmp(name
, radv_instance_extensions
[i
].extensionName
) == 0)
496 VkResult
radv_CreateInstance(
497 const VkInstanceCreateInfo
* pCreateInfo
,
498 const VkAllocationCallbacks
* pAllocator
,
499 VkInstance
* pInstance
)
501 struct radv_instance
*instance
;
504 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
506 uint32_t client_version
;
507 if (pCreateInfo
->pApplicationInfo
&&
508 pCreateInfo
->pApplicationInfo
->apiVersion
!= 0) {
509 client_version
= pCreateInfo
->pApplicationInfo
->apiVersion
;
511 radv_EnumerateInstanceVersion(&client_version
);
514 instance
= vk_zalloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
515 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
517 return vk_error(NULL
, VK_ERROR_OUT_OF_HOST_MEMORY
);
519 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
522 instance
->alloc
= *pAllocator
;
524 instance
->alloc
= default_alloc
;
526 instance
->apiVersion
= client_version
;
527 instance
->physicalDeviceCount
= -1;
529 instance
->debug_flags
= parse_debug_string(getenv("RADV_DEBUG"),
532 instance
->perftest_flags
= parse_debug_string(getenv("RADV_PERFTEST"),
533 radv_perftest_options
);
536 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
537 radv_logi("Created an instance");
539 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
540 const char *ext_name
= pCreateInfo
->ppEnabledExtensionNames
[i
];
541 int index
= radv_get_instance_extension_index(ext_name
);
543 if (index
< 0 || !radv_supported_instance_extensions
.extensions
[index
]) {
544 vk_free2(&default_alloc
, pAllocator
, instance
);
545 return vk_error(instance
, VK_ERROR_EXTENSION_NOT_PRESENT
);
548 instance
->enabled_extensions
.extensions
[index
] = true;
551 result
= vk_debug_report_instance_init(&instance
->debug_report_callbacks
);
552 if (result
!= VK_SUCCESS
) {
553 vk_free2(&default_alloc
, pAllocator
, instance
);
554 return vk_error(instance
, result
);
559 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
561 radv_handle_per_app_options(instance
, pCreateInfo
->pApplicationInfo
);
563 *pInstance
= radv_instance_to_handle(instance
);
568 void radv_DestroyInstance(
569 VkInstance _instance
,
570 const VkAllocationCallbacks
* pAllocator
)
572 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
577 for (int i
= 0; i
< instance
->physicalDeviceCount
; ++i
) {
578 radv_physical_device_finish(instance
->physicalDevices
+ i
);
581 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
585 vk_debug_report_instance_destroy(&instance
->debug_report_callbacks
);
587 vk_free(&instance
->alloc
, instance
);
591 radv_enumerate_devices(struct radv_instance
*instance
)
593 /* TODO: Check for more devices ? */
594 drmDevicePtr devices
[8];
595 VkResult result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
598 instance
->physicalDeviceCount
= 0;
600 max_devices
= drmGetDevices2(0, devices
, ARRAY_SIZE(devices
));
602 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
603 radv_logi("Found %d drm nodes", max_devices
);
606 return vk_error(instance
, VK_ERROR_INCOMPATIBLE_DRIVER
);
608 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
609 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
610 devices
[i
]->bustype
== DRM_BUS_PCI
&&
611 devices
[i
]->deviceinfo
.pci
->vendor_id
== ATI_VENDOR_ID
) {
613 result
= radv_physical_device_init(instance
->physicalDevices
+
614 instance
->physicalDeviceCount
,
617 if (result
== VK_SUCCESS
)
618 ++instance
->physicalDeviceCount
;
619 else if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
623 drmFreeDevices(devices
, max_devices
);
628 VkResult
radv_EnumeratePhysicalDevices(
629 VkInstance _instance
,
630 uint32_t* pPhysicalDeviceCount
,
631 VkPhysicalDevice
* pPhysicalDevices
)
633 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
636 if (instance
->physicalDeviceCount
< 0) {
637 result
= radv_enumerate_devices(instance
);
638 if (result
!= VK_SUCCESS
&&
639 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
643 if (!pPhysicalDevices
) {
644 *pPhysicalDeviceCount
= instance
->physicalDeviceCount
;
646 *pPhysicalDeviceCount
= MIN2(*pPhysicalDeviceCount
, instance
->physicalDeviceCount
);
647 for (unsigned i
= 0; i
< *pPhysicalDeviceCount
; ++i
)
648 pPhysicalDevices
[i
] = radv_physical_device_to_handle(instance
->physicalDevices
+ i
);
651 return *pPhysicalDeviceCount
< instance
->physicalDeviceCount
? VK_INCOMPLETE
655 VkResult
radv_EnumeratePhysicalDeviceGroups(
656 VkInstance _instance
,
657 uint32_t* pPhysicalDeviceGroupCount
,
658 VkPhysicalDeviceGroupProperties
* pPhysicalDeviceGroupProperties
)
660 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
663 if (instance
->physicalDeviceCount
< 0) {
664 result
= radv_enumerate_devices(instance
);
665 if (result
!= VK_SUCCESS
&&
666 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
670 if (!pPhysicalDeviceGroupProperties
) {
671 *pPhysicalDeviceGroupCount
= instance
->physicalDeviceCount
;
673 *pPhysicalDeviceGroupCount
= MIN2(*pPhysicalDeviceGroupCount
, instance
->physicalDeviceCount
);
674 for (unsigned i
= 0; i
< *pPhysicalDeviceGroupCount
; ++i
) {
675 pPhysicalDeviceGroupProperties
[i
].physicalDeviceCount
= 1;
676 pPhysicalDeviceGroupProperties
[i
].physicalDevices
[0] = radv_physical_device_to_handle(instance
->physicalDevices
+ i
);
677 pPhysicalDeviceGroupProperties
[i
].subsetAllocation
= false;
680 return *pPhysicalDeviceGroupCount
< instance
->physicalDeviceCount
? VK_INCOMPLETE
684 void radv_GetPhysicalDeviceFeatures(
685 VkPhysicalDevice physicalDevice
,
686 VkPhysicalDeviceFeatures
* pFeatures
)
688 memset(pFeatures
, 0, sizeof(*pFeatures
));
690 *pFeatures
= (VkPhysicalDeviceFeatures
) {
691 .robustBufferAccess
= true,
692 .fullDrawIndexUint32
= true,
693 .imageCubeArray
= true,
694 .independentBlend
= true,
695 .geometryShader
= true,
696 .tessellationShader
= true,
697 .sampleRateShading
= true,
698 .dualSrcBlend
= true,
700 .multiDrawIndirect
= true,
701 .drawIndirectFirstInstance
= true,
703 .depthBiasClamp
= true,
704 .fillModeNonSolid
= true,
709 .multiViewport
= true,
710 .samplerAnisotropy
= true,
711 .textureCompressionETC2
= false,
712 .textureCompressionASTC_LDR
= false,
713 .textureCompressionBC
= true,
714 .occlusionQueryPrecise
= true,
715 .pipelineStatisticsQuery
= true,
716 .vertexPipelineStoresAndAtomics
= true,
717 .fragmentStoresAndAtomics
= true,
718 .shaderTessellationAndGeometryPointSize
= true,
719 .shaderImageGatherExtended
= true,
720 .shaderStorageImageExtendedFormats
= true,
721 .shaderStorageImageMultisample
= false,
722 .shaderUniformBufferArrayDynamicIndexing
= true,
723 .shaderSampledImageArrayDynamicIndexing
= true,
724 .shaderStorageBufferArrayDynamicIndexing
= true,
725 .shaderStorageImageArrayDynamicIndexing
= true,
726 .shaderStorageImageReadWithoutFormat
= true,
727 .shaderStorageImageWriteWithoutFormat
= true,
728 .shaderClipDistance
= true,
729 .shaderCullDistance
= true,
730 .shaderFloat64
= true,
732 .shaderInt16
= false,
733 .sparseBinding
= true,
734 .variableMultisampleRate
= true,
735 .inheritedQueries
= true,
739 void radv_GetPhysicalDeviceFeatures2(
740 VkPhysicalDevice physicalDevice
,
741 VkPhysicalDeviceFeatures2KHR
*pFeatures
)
743 vk_foreach_struct(ext
, pFeatures
->pNext
) {
744 switch (ext
->sType
) {
745 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTER_FEATURES_KHR
: {
746 VkPhysicalDeviceVariablePointerFeaturesKHR
*features
= (void *)ext
;
747 features
->variablePointersStorageBuffer
= true;
748 features
->variablePointers
= false;
751 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES_KHR
: {
752 VkPhysicalDeviceMultiviewFeaturesKHR
*features
= (VkPhysicalDeviceMultiviewFeaturesKHR
*)ext
;
753 features
->multiview
= true;
754 features
->multiviewGeometryShader
= true;
755 features
->multiviewTessellationShader
= true;
758 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETER_FEATURES
: {
759 VkPhysicalDeviceShaderDrawParameterFeatures
*features
=
760 (VkPhysicalDeviceShaderDrawParameterFeatures
*)ext
;
761 features
->shaderDrawParameters
= true;
764 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES
: {
765 VkPhysicalDeviceProtectedMemoryFeatures
*features
=
766 (VkPhysicalDeviceProtectedMemoryFeatures
*)ext
;
767 features
->protectedMemory
= false;
770 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES
: {
771 VkPhysicalDevice16BitStorageFeatures
*features
=
772 (VkPhysicalDevice16BitStorageFeatures
*)ext
;
773 features
->storageBuffer16BitAccess
= false;
774 features
->uniformAndStorageBuffer16BitAccess
= false;
775 features
->storagePushConstant16
= false;
776 features
->storageInputOutput16
= false;
779 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES
: {
780 VkPhysicalDeviceSamplerYcbcrConversionFeatures
*features
=
781 (VkPhysicalDeviceSamplerYcbcrConversionFeatures
*)ext
;
782 features
->samplerYcbcrConversion
= false;
785 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT
: {
786 VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*features
=
787 (VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*)ext
;
788 features
->shaderInputAttachmentArrayDynamicIndexing
= true;
789 features
->shaderUniformTexelBufferArrayDynamicIndexing
= true;
790 features
->shaderStorageTexelBufferArrayDynamicIndexing
= true;
791 features
->shaderUniformBufferArrayNonUniformIndexing
= false;
792 features
->shaderSampledImageArrayNonUniformIndexing
= false;
793 features
->shaderStorageBufferArrayNonUniformIndexing
= false;
794 features
->shaderStorageImageArrayNonUniformIndexing
= false;
795 features
->shaderInputAttachmentArrayNonUniformIndexing
= false;
796 features
->shaderUniformTexelBufferArrayNonUniformIndexing
= false;
797 features
->shaderStorageTexelBufferArrayNonUniformIndexing
= false;
798 features
->descriptorBindingUniformBufferUpdateAfterBind
= true;
799 features
->descriptorBindingSampledImageUpdateAfterBind
= true;
800 features
->descriptorBindingStorageImageUpdateAfterBind
= true;
801 features
->descriptorBindingStorageBufferUpdateAfterBind
= true;
802 features
->descriptorBindingUniformTexelBufferUpdateAfterBind
= true;
803 features
->descriptorBindingStorageTexelBufferUpdateAfterBind
= true;
804 features
->descriptorBindingUpdateUnusedWhilePending
= true;
805 features
->descriptorBindingPartiallyBound
= true;
806 features
->descriptorBindingVariableDescriptorCount
= true;
807 features
->runtimeDescriptorArray
= true;
814 return radv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
817 void radv_GetPhysicalDeviceProperties(
818 VkPhysicalDevice physicalDevice
,
819 VkPhysicalDeviceProperties
* pProperties
)
821 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
822 VkSampleCountFlags sample_counts
= 0xf;
824 /* make sure that the entire descriptor set is addressable with a signed
825 * 32-bit int. So the sum of all limits scaled by descriptor size has to
826 * be at most 2 GiB. the combined image & samples object count as one of
827 * both. This limit is for the pipeline layout, not for the set layout, but
828 * there is no set limit, so we just set a pipeline limit. I don't think
829 * any app is going to hit this soon. */
830 size_t max_descriptor_set_size
= ((1ull << 31) - 16 * MAX_DYNAMIC_BUFFERS
) /
831 (32 /* uniform buffer, 32 due to potential space wasted on alignment */ +
832 32 /* storage buffer, 32 due to potential space wasted on alignment */ +
833 32 /* sampler, largest when combined with image */ +
834 64 /* sampled image */ +
835 64 /* storage image */);
837 VkPhysicalDeviceLimits limits
= {
838 .maxImageDimension1D
= (1 << 14),
839 .maxImageDimension2D
= (1 << 14),
840 .maxImageDimension3D
= (1 << 11),
841 .maxImageDimensionCube
= (1 << 14),
842 .maxImageArrayLayers
= (1 << 11),
843 .maxTexelBufferElements
= 128 * 1024 * 1024,
844 .maxUniformBufferRange
= UINT32_MAX
,
845 .maxStorageBufferRange
= UINT32_MAX
,
846 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
847 .maxMemoryAllocationCount
= UINT32_MAX
,
848 .maxSamplerAllocationCount
= 64 * 1024,
849 .bufferImageGranularity
= 64, /* A cache line */
850 .sparseAddressSpaceSize
= 0xffffffffu
, /* buffer max size */
851 .maxBoundDescriptorSets
= MAX_SETS
,
852 .maxPerStageDescriptorSamplers
= max_descriptor_set_size
,
853 .maxPerStageDescriptorUniformBuffers
= max_descriptor_set_size
,
854 .maxPerStageDescriptorStorageBuffers
= max_descriptor_set_size
,
855 .maxPerStageDescriptorSampledImages
= max_descriptor_set_size
,
856 .maxPerStageDescriptorStorageImages
= max_descriptor_set_size
,
857 .maxPerStageDescriptorInputAttachments
= max_descriptor_set_size
,
858 .maxPerStageResources
= max_descriptor_set_size
,
859 .maxDescriptorSetSamplers
= max_descriptor_set_size
,
860 .maxDescriptorSetUniformBuffers
= max_descriptor_set_size
,
861 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_UNIFORM_BUFFERS
,
862 .maxDescriptorSetStorageBuffers
= max_descriptor_set_size
,
863 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_STORAGE_BUFFERS
,
864 .maxDescriptorSetSampledImages
= max_descriptor_set_size
,
865 .maxDescriptorSetStorageImages
= max_descriptor_set_size
,
866 .maxDescriptorSetInputAttachments
= max_descriptor_set_size
,
867 .maxVertexInputAttributes
= 32,
868 .maxVertexInputBindings
= 32,
869 .maxVertexInputAttributeOffset
= 2047,
870 .maxVertexInputBindingStride
= 2048,
871 .maxVertexOutputComponents
= 128,
872 .maxTessellationGenerationLevel
= 64,
873 .maxTessellationPatchSize
= 32,
874 .maxTessellationControlPerVertexInputComponents
= 128,
875 .maxTessellationControlPerVertexOutputComponents
= 128,
876 .maxTessellationControlPerPatchOutputComponents
= 120,
877 .maxTessellationControlTotalOutputComponents
= 4096,
878 .maxTessellationEvaluationInputComponents
= 128,
879 .maxTessellationEvaluationOutputComponents
= 128,
880 .maxGeometryShaderInvocations
= 127,
881 .maxGeometryInputComponents
= 64,
882 .maxGeometryOutputComponents
= 128,
883 .maxGeometryOutputVertices
= 256,
884 .maxGeometryTotalOutputComponents
= 1024,
885 .maxFragmentInputComponents
= 128,
886 .maxFragmentOutputAttachments
= 8,
887 .maxFragmentDualSrcAttachments
= 1,
888 .maxFragmentCombinedOutputResources
= 8,
889 .maxComputeSharedMemorySize
= 32768,
890 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
891 .maxComputeWorkGroupInvocations
= 2048,
892 .maxComputeWorkGroupSize
= {
897 .subPixelPrecisionBits
= 4 /* FIXME */,
898 .subTexelPrecisionBits
= 4 /* FIXME */,
899 .mipmapPrecisionBits
= 4 /* FIXME */,
900 .maxDrawIndexedIndexValue
= UINT32_MAX
,
901 .maxDrawIndirectCount
= UINT32_MAX
,
902 .maxSamplerLodBias
= 16,
903 .maxSamplerAnisotropy
= 16,
904 .maxViewports
= MAX_VIEWPORTS
,
905 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
906 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
907 .viewportSubPixelBits
= 8,
908 .minMemoryMapAlignment
= 4096, /* A page */
909 .minTexelBufferOffsetAlignment
= 1,
910 .minUniformBufferOffsetAlignment
= 4,
911 .minStorageBufferOffsetAlignment
= 4,
912 .minTexelOffset
= -32,
913 .maxTexelOffset
= 31,
914 .minTexelGatherOffset
= -32,
915 .maxTexelGatherOffset
= 31,
916 .minInterpolationOffset
= -2,
917 .maxInterpolationOffset
= 2,
918 .subPixelInterpolationOffsetBits
= 8,
919 .maxFramebufferWidth
= (1 << 14),
920 .maxFramebufferHeight
= (1 << 14),
921 .maxFramebufferLayers
= (1 << 10),
922 .framebufferColorSampleCounts
= sample_counts
,
923 .framebufferDepthSampleCounts
= sample_counts
,
924 .framebufferStencilSampleCounts
= sample_counts
,
925 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
926 .maxColorAttachments
= MAX_RTS
,
927 .sampledImageColorSampleCounts
= sample_counts
,
928 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
929 .sampledImageDepthSampleCounts
= sample_counts
,
930 .sampledImageStencilSampleCounts
= sample_counts
,
931 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
932 .maxSampleMaskWords
= 1,
933 .timestampComputeAndGraphics
= true,
934 .timestampPeriod
= 1000000.0 / pdevice
->rad_info
.clock_crystal_freq
,
935 .maxClipDistances
= 8,
936 .maxCullDistances
= 8,
937 .maxCombinedClipAndCullDistances
= 8,
938 .discreteQueuePriorities
= 1,
939 .pointSizeRange
= { 0.125, 255.875 },
940 .lineWidthRange
= { 0.0, 7.9921875 },
941 .pointSizeGranularity
= (1.0 / 8.0),
942 .lineWidthGranularity
= (1.0 / 128.0),
943 .strictLines
= false, /* FINISHME */
944 .standardSampleLocations
= true,
945 .optimalBufferCopyOffsetAlignment
= 128,
946 .optimalBufferCopyRowPitchAlignment
= 128,
947 .nonCoherentAtomSize
= 64,
950 *pProperties
= (VkPhysicalDeviceProperties
) {
951 .apiVersion
= radv_physical_device_api_version(pdevice
),
952 .driverVersion
= vk_get_driver_version(),
953 .vendorID
= ATI_VENDOR_ID
,
954 .deviceID
= pdevice
->rad_info
.pci_id
,
955 .deviceType
= pdevice
->rad_info
.has_dedicated_vram
? VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU
: VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
957 .sparseProperties
= {0},
960 strcpy(pProperties
->deviceName
, pdevice
->name
);
961 memcpy(pProperties
->pipelineCacheUUID
, pdevice
->cache_uuid
, VK_UUID_SIZE
);
964 void radv_GetPhysicalDeviceProperties2(
965 VkPhysicalDevice physicalDevice
,
966 VkPhysicalDeviceProperties2KHR
*pProperties
)
968 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
969 radv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
971 vk_foreach_struct(ext
, pProperties
->pNext
) {
972 switch (ext
->sType
) {
973 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
974 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
975 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
976 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
979 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES_KHR
: {
980 VkPhysicalDeviceIDPropertiesKHR
*properties
= (VkPhysicalDeviceIDPropertiesKHR
*)ext
;
981 memcpy(properties
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
982 memcpy(properties
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
983 properties
->deviceLUIDValid
= false;
986 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES_KHR
: {
987 VkPhysicalDeviceMultiviewPropertiesKHR
*properties
= (VkPhysicalDeviceMultiviewPropertiesKHR
*)ext
;
988 properties
->maxMultiviewViewCount
= MAX_VIEWS
;
989 properties
->maxMultiviewInstanceIndex
= INT_MAX
;
992 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES_KHR
: {
993 VkPhysicalDevicePointClippingPropertiesKHR
*properties
=
994 (VkPhysicalDevicePointClippingPropertiesKHR
*)ext
;
995 properties
->pointClippingBehavior
= VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES_KHR
;
998 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DISCARD_RECTANGLE_PROPERTIES_EXT
: {
999 VkPhysicalDeviceDiscardRectanglePropertiesEXT
*properties
=
1000 (VkPhysicalDeviceDiscardRectanglePropertiesEXT
*)ext
;
1001 properties
->maxDiscardRectangles
= MAX_DISCARD_RECTANGLES
;
1004 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT
: {
1005 VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*properties
=
1006 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*) ext
;
1007 properties
->minImportedHostPointerAlignment
= 4096;
1010 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES
: {
1011 VkPhysicalDeviceSubgroupProperties
*properties
=
1012 (VkPhysicalDeviceSubgroupProperties
*)ext
;
1013 properties
->subgroupSize
= 64;
1014 properties
->supportedStages
= VK_SHADER_STAGE_ALL
;
1015 properties
->supportedOperations
=
1016 VK_SUBGROUP_FEATURE_BASIC_BIT
|
1017 VK_SUBGROUP_FEATURE_BALLOT_BIT
|
1018 VK_SUBGROUP_FEATURE_QUAD_BIT
|
1019 VK_SUBGROUP_FEATURE_VOTE_BIT
;
1020 if (pdevice
->rad_info
.chip_class
>= VI
) {
1021 properties
->supportedOperations
|=
1022 VK_SUBGROUP_FEATURE_SHUFFLE_BIT
|
1023 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT
;
1025 properties
->quadOperationsInAllStages
= true;
1028 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES
: {
1029 VkPhysicalDeviceMaintenance3Properties
*properties
=
1030 (VkPhysicalDeviceMaintenance3Properties
*)ext
;
1031 /* Make sure everything is addressable by a signed 32-bit int, and
1032 * our largest descriptors are 96 bytes. */
1033 properties
->maxPerSetDescriptors
= (1ull << 31) / 96;
1034 /* Our buffer size fields allow only this much */
1035 properties
->maxMemoryAllocationSize
= 0xFFFFFFFFull
;
1038 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT
: {
1039 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*properties
=
1040 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*)ext
;
1041 /* GFX6-8 only support single channel min/max filter. */
1042 properties
->filterMinmaxImageComponentMapping
= pdevice
->rad_info
.chip_class
>= GFX9
;
1043 properties
->filterMinmaxSingleComponentFormats
= true;
1046 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CORE_PROPERTIES_AMD
: {
1047 VkPhysicalDeviceShaderCorePropertiesAMD
*properties
=
1048 (VkPhysicalDeviceShaderCorePropertiesAMD
*)ext
;
1050 /* Shader engines. */
1051 properties
->shaderEngineCount
=
1052 pdevice
->rad_info
.max_se
;
1053 properties
->shaderArraysPerEngineCount
=
1054 pdevice
->rad_info
.max_sh_per_se
;
1055 properties
->computeUnitsPerShaderArray
=
1056 pdevice
->rad_info
.num_good_compute_units
/
1057 (pdevice
->rad_info
.max_se
*
1058 pdevice
->rad_info
.max_sh_per_se
);
1059 properties
->simdPerComputeUnit
= 4;
1060 properties
->wavefrontsPerSimd
=
1061 pdevice
->rad_info
.family
== CHIP_TONGA
||
1062 pdevice
->rad_info
.family
== CHIP_ICELAND
||
1063 pdevice
->rad_info
.family
== CHIP_POLARIS10
||
1064 pdevice
->rad_info
.family
== CHIP_POLARIS11
||
1065 pdevice
->rad_info
.family
== CHIP_POLARIS12
||
1066 pdevice
->rad_info
.family
== CHIP_VEGAM
? 8 : 10;
1067 properties
->wavefrontSize
= 64;
1070 properties
->sgprsPerSimd
=
1071 radv_get_num_physical_sgprs(pdevice
);
1072 properties
->minSgprAllocation
=
1073 pdevice
->rad_info
.chip_class
>= VI
? 16 : 8;
1074 properties
->maxSgprAllocation
=
1075 pdevice
->rad_info
.family
== CHIP_TONGA
||
1076 pdevice
->rad_info
.family
== CHIP_ICELAND
? 96 : 104;
1077 properties
->sgprAllocationGranularity
=
1078 pdevice
->rad_info
.chip_class
>= VI
? 16 : 8;
1081 properties
->vgprsPerSimd
= RADV_NUM_PHYSICAL_VGPRS
;
1082 properties
->minVgprAllocation
= 4;
1083 properties
->maxVgprAllocation
= 256;
1084 properties
->vgprAllocationGranularity
= 4;
1087 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT
: {
1088 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*properties
=
1089 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*)ext
;
1090 properties
->maxVertexAttribDivisor
= UINT32_MAX
;
1093 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES_EXT
: {
1094 VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*properties
=
1095 (VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*)ext
;
1096 properties
->maxUpdateAfterBindDescriptorsInAllPools
= UINT32_MAX
/ 64;
1097 properties
->shaderUniformBufferArrayNonUniformIndexingNative
= false;
1098 properties
->shaderSampledImageArrayNonUniformIndexingNative
= false;
1099 properties
->shaderStorageBufferArrayNonUniformIndexingNative
= false;
1100 properties
->shaderStorageImageArrayNonUniformIndexingNative
= false;
1101 properties
->shaderInputAttachmentArrayNonUniformIndexingNative
= false;
1102 properties
->robustBufferAccessUpdateAfterBind
= false;
1103 properties
->quadDivergentImplicitLod
= false;
1105 size_t max_descriptor_set_size
= ((1ull << 31) - 16 * MAX_DYNAMIC_BUFFERS
) /
1106 (32 /* uniform buffer, 32 due to potential space wasted on alignment */ +
1107 32 /* storage buffer, 32 due to potential space wasted on alignment */ +
1108 32 /* sampler, largest when combined with image */ +
1109 64 /* sampled image */ +
1110 64 /* storage image */);
1111 properties
->maxPerStageDescriptorUpdateAfterBindSamplers
= max_descriptor_set_size
;
1112 properties
->maxPerStageDescriptorUpdateAfterBindUniformBuffers
= max_descriptor_set_size
;
1113 properties
->maxPerStageDescriptorUpdateAfterBindStorageBuffers
= max_descriptor_set_size
;
1114 properties
->maxPerStageDescriptorUpdateAfterBindSampledImages
= max_descriptor_set_size
;
1115 properties
->maxPerStageDescriptorUpdateAfterBindStorageImages
= max_descriptor_set_size
;
1116 properties
->maxPerStageDescriptorUpdateAfterBindInputAttachments
= max_descriptor_set_size
;
1117 properties
->maxPerStageUpdateAfterBindResources
= max_descriptor_set_size
;
1118 properties
->maxDescriptorSetUpdateAfterBindSamplers
= max_descriptor_set_size
;
1119 properties
->maxDescriptorSetUpdateAfterBindUniformBuffers
= max_descriptor_set_size
;
1120 properties
->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic
= MAX_DYNAMIC_UNIFORM_BUFFERS
;
1121 properties
->maxDescriptorSetUpdateAfterBindStorageBuffers
= max_descriptor_set_size
;
1122 properties
->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic
= MAX_DYNAMIC_STORAGE_BUFFERS
;
1123 properties
->maxDescriptorSetUpdateAfterBindSampledImages
= max_descriptor_set_size
;
1124 properties
->maxDescriptorSetUpdateAfterBindStorageImages
= max_descriptor_set_size
;
1125 properties
->maxDescriptorSetUpdateAfterBindInputAttachments
= max_descriptor_set_size
;
1134 static void radv_get_physical_device_queue_family_properties(
1135 struct radv_physical_device
* pdevice
,
1137 VkQueueFamilyProperties
** pQueueFamilyProperties
)
1139 int num_queue_families
= 1;
1141 if (pdevice
->rad_info
.num_compute_rings
> 0 &&
1142 !(pdevice
->instance
->debug_flags
& RADV_DEBUG_NO_COMPUTE_QUEUE
))
1143 num_queue_families
++;
1145 if (pQueueFamilyProperties
== NULL
) {
1146 *pCount
= num_queue_families
;
1155 *pQueueFamilyProperties
[idx
] = (VkQueueFamilyProperties
) {
1156 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
1157 VK_QUEUE_COMPUTE_BIT
|
1158 VK_QUEUE_TRANSFER_BIT
|
1159 VK_QUEUE_SPARSE_BINDING_BIT
,
1161 .timestampValidBits
= 64,
1162 .minImageTransferGranularity
= (VkExtent3D
) { 1, 1, 1 },
1167 if (pdevice
->rad_info
.num_compute_rings
> 0 &&
1168 !(pdevice
->instance
->debug_flags
& RADV_DEBUG_NO_COMPUTE_QUEUE
)) {
1169 if (*pCount
> idx
) {
1170 *pQueueFamilyProperties
[idx
] = (VkQueueFamilyProperties
) {
1171 .queueFlags
= VK_QUEUE_COMPUTE_BIT
|
1172 VK_QUEUE_TRANSFER_BIT
|
1173 VK_QUEUE_SPARSE_BINDING_BIT
,
1174 .queueCount
= pdevice
->rad_info
.num_compute_rings
,
1175 .timestampValidBits
= 64,
1176 .minImageTransferGranularity
= (VkExtent3D
) { 1, 1, 1 },
1184 void radv_GetPhysicalDeviceQueueFamilyProperties(
1185 VkPhysicalDevice physicalDevice
,
1187 VkQueueFamilyProperties
* pQueueFamilyProperties
)
1189 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
1190 if (!pQueueFamilyProperties
) {
1191 return radv_get_physical_device_queue_family_properties(pdevice
, pCount
, NULL
);
1194 VkQueueFamilyProperties
*properties
[] = {
1195 pQueueFamilyProperties
+ 0,
1196 pQueueFamilyProperties
+ 1,
1197 pQueueFamilyProperties
+ 2,
1199 radv_get_physical_device_queue_family_properties(pdevice
, pCount
, properties
);
1200 assert(*pCount
<= 3);
1203 void radv_GetPhysicalDeviceQueueFamilyProperties2(
1204 VkPhysicalDevice physicalDevice
,
1206 VkQueueFamilyProperties2KHR
*pQueueFamilyProperties
)
1208 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
1209 if (!pQueueFamilyProperties
) {
1210 return radv_get_physical_device_queue_family_properties(pdevice
, pCount
, NULL
);
1213 VkQueueFamilyProperties
*properties
[] = {
1214 &pQueueFamilyProperties
[0].queueFamilyProperties
,
1215 &pQueueFamilyProperties
[1].queueFamilyProperties
,
1216 &pQueueFamilyProperties
[2].queueFamilyProperties
,
1218 radv_get_physical_device_queue_family_properties(pdevice
, pCount
, properties
);
1219 assert(*pCount
<= 3);
1222 void radv_GetPhysicalDeviceMemoryProperties(
1223 VkPhysicalDevice physicalDevice
,
1224 VkPhysicalDeviceMemoryProperties
*pMemoryProperties
)
1226 RADV_FROM_HANDLE(radv_physical_device
, physical_device
, physicalDevice
);
1228 *pMemoryProperties
= physical_device
->memory_properties
;
1231 void radv_GetPhysicalDeviceMemoryProperties2(
1232 VkPhysicalDevice physicalDevice
,
1233 VkPhysicalDeviceMemoryProperties2KHR
*pMemoryProperties
)
1235 return radv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
1236 &pMemoryProperties
->memoryProperties
);
1239 VkResult
radv_GetMemoryHostPointerPropertiesEXT(
1241 VkExternalMemoryHandleTypeFlagBitsKHR handleType
,
1242 const void *pHostPointer
,
1243 VkMemoryHostPointerPropertiesEXT
*pMemoryHostPointerProperties
)
1245 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1249 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
: {
1250 const struct radv_physical_device
*physical_device
= device
->physical_device
;
1251 uint32_t memoryTypeBits
= 0;
1252 for (int i
= 0; i
< physical_device
->memory_properties
.memoryTypeCount
; i
++) {
1253 if (physical_device
->mem_type_indices
[i
] == RADV_MEM_TYPE_GTT_CACHED
) {
1254 memoryTypeBits
= (1 << i
);
1258 pMemoryHostPointerProperties
->memoryTypeBits
= memoryTypeBits
;
1262 return VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
;
1266 static enum radeon_ctx_priority
1267 radv_get_queue_global_priority(const VkDeviceQueueGlobalPriorityCreateInfoEXT
*pObj
)
1269 /* Default to MEDIUM when a specific global priority isn't requested */
1271 return RADEON_CTX_PRIORITY_MEDIUM
;
1273 switch(pObj
->globalPriority
) {
1274 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT
:
1275 return RADEON_CTX_PRIORITY_REALTIME
;
1276 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT
:
1277 return RADEON_CTX_PRIORITY_HIGH
;
1278 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
:
1279 return RADEON_CTX_PRIORITY_MEDIUM
;
1280 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT
:
1281 return RADEON_CTX_PRIORITY_LOW
;
1283 unreachable("Illegal global priority value");
1284 return RADEON_CTX_PRIORITY_INVALID
;
1289 radv_queue_init(struct radv_device
*device
, struct radv_queue
*queue
,
1290 uint32_t queue_family_index
, int idx
,
1291 VkDeviceQueueCreateFlags flags
,
1292 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*global_priority
)
1294 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1295 queue
->device
= device
;
1296 queue
->queue_family_index
= queue_family_index
;
1297 queue
->queue_idx
= idx
;
1298 queue
->priority
= radv_get_queue_global_priority(global_priority
);
1299 queue
->flags
= flags
;
1301 queue
->hw_ctx
= device
->ws
->ctx_create(device
->ws
, queue
->priority
);
1303 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
1309 radv_queue_finish(struct radv_queue
*queue
)
1312 queue
->device
->ws
->ctx_destroy(queue
->hw_ctx
);
1314 if (queue
->initial_full_flush_preamble_cs
)
1315 queue
->device
->ws
->cs_destroy(queue
->initial_full_flush_preamble_cs
);
1316 if (queue
->initial_preamble_cs
)
1317 queue
->device
->ws
->cs_destroy(queue
->initial_preamble_cs
);
1318 if (queue
->continue_preamble_cs
)
1319 queue
->device
->ws
->cs_destroy(queue
->continue_preamble_cs
);
1320 if (queue
->descriptor_bo
)
1321 queue
->device
->ws
->buffer_destroy(queue
->descriptor_bo
);
1322 if (queue
->scratch_bo
)
1323 queue
->device
->ws
->buffer_destroy(queue
->scratch_bo
);
1324 if (queue
->esgs_ring_bo
)
1325 queue
->device
->ws
->buffer_destroy(queue
->esgs_ring_bo
);
1326 if (queue
->gsvs_ring_bo
)
1327 queue
->device
->ws
->buffer_destroy(queue
->gsvs_ring_bo
);
1328 if (queue
->tess_rings_bo
)
1329 queue
->device
->ws
->buffer_destroy(queue
->tess_rings_bo
);
1330 if (queue
->compute_scratch_bo
)
1331 queue
->device
->ws
->buffer_destroy(queue
->compute_scratch_bo
);
1335 radv_bo_list_init(struct radv_bo_list
*bo_list
)
1337 pthread_mutex_init(&bo_list
->mutex
, NULL
);
1338 bo_list
->list
.count
= bo_list
->capacity
= 0;
1339 bo_list
->list
.bos
= NULL
;
1343 radv_bo_list_finish(struct radv_bo_list
*bo_list
)
1345 free(bo_list
->list
.bos
);
1346 pthread_mutex_destroy(&bo_list
->mutex
);
1349 static VkResult
radv_bo_list_add(struct radv_device
*device
,
1350 struct radeon_winsys_bo
*bo
)
1352 struct radv_bo_list
*bo_list
= &device
->bo_list
;
1354 if (unlikely(!device
->use_global_bo_list
))
1357 pthread_mutex_lock(&bo_list
->mutex
);
1358 if (bo_list
->list
.count
== bo_list
->capacity
) {
1359 unsigned capacity
= MAX2(4, bo_list
->capacity
* 2);
1360 void *data
= realloc(bo_list
->list
.bos
, capacity
* sizeof(struct radeon_winsys_bo
*));
1363 pthread_mutex_unlock(&bo_list
->mutex
);
1364 return VK_ERROR_OUT_OF_HOST_MEMORY
;
1367 bo_list
->list
.bos
= (struct radeon_winsys_bo
**)data
;
1368 bo_list
->capacity
= capacity
;
1371 bo_list
->list
.bos
[bo_list
->list
.count
++] = bo
;
1372 pthread_mutex_unlock(&bo_list
->mutex
);
1376 static void radv_bo_list_remove(struct radv_device
*device
,
1377 struct radeon_winsys_bo
*bo
)
1379 struct radv_bo_list
*bo_list
= &device
->bo_list
;
1381 if (unlikely(!device
->use_global_bo_list
))
1384 pthread_mutex_lock(&bo_list
->mutex
);
1385 for(unsigned i
= 0; i
< bo_list
->list
.count
; ++i
) {
1386 if (bo_list
->list
.bos
[i
] == bo
) {
1387 bo_list
->list
.bos
[i
] = bo_list
->list
.bos
[bo_list
->list
.count
- 1];
1388 --bo_list
->list
.count
;
1392 pthread_mutex_unlock(&bo_list
->mutex
);
1396 radv_device_init_gs_info(struct radv_device
*device
)
1398 device
->gs_table_depth
= ac_get_gs_table_depth(device
->physical_device
->rad_info
.chip_class
,
1399 device
->physical_device
->rad_info
.family
);
1402 static int radv_get_device_extension_index(const char *name
)
1404 for (unsigned i
= 0; i
< RADV_DEVICE_EXTENSION_COUNT
; ++i
) {
1405 if (strcmp(name
, radv_device_extensions
[i
].extensionName
) == 0)
1411 VkResult
radv_CreateDevice(
1412 VkPhysicalDevice physicalDevice
,
1413 const VkDeviceCreateInfo
* pCreateInfo
,
1414 const VkAllocationCallbacks
* pAllocator
,
1417 RADV_FROM_HANDLE(radv_physical_device
, physical_device
, physicalDevice
);
1419 struct radv_device
*device
;
1421 bool keep_shader_info
= false;
1423 /* Check enabled features */
1424 if (pCreateInfo
->pEnabledFeatures
) {
1425 VkPhysicalDeviceFeatures supported_features
;
1426 radv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
1427 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
1428 VkBool32
*enabled_feature
= (VkBool32
*)pCreateInfo
->pEnabledFeatures
;
1429 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
1430 for (uint32_t i
= 0; i
< num_features
; i
++) {
1431 if (enabled_feature
[i
] && !supported_feature
[i
])
1432 return vk_error(physical_device
->instance
, VK_ERROR_FEATURE_NOT_PRESENT
);
1436 device
= vk_zalloc2(&physical_device
->instance
->alloc
, pAllocator
,
1438 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
1440 return vk_error(physical_device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
1442 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1443 device
->instance
= physical_device
->instance
;
1444 device
->physical_device
= physical_device
;
1446 device
->ws
= physical_device
->ws
;
1448 device
->alloc
= *pAllocator
;
1450 device
->alloc
= physical_device
->instance
->alloc
;
1452 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
1453 const char *ext_name
= pCreateInfo
->ppEnabledExtensionNames
[i
];
1454 int index
= radv_get_device_extension_index(ext_name
);
1455 if (index
< 0 || !physical_device
->supported_extensions
.extensions
[index
]) {
1456 vk_free(&device
->alloc
, device
);
1457 return vk_error(physical_device
->instance
, VK_ERROR_EXTENSION_NOT_PRESENT
);
1460 device
->enabled_extensions
.extensions
[index
] = true;
1463 keep_shader_info
= device
->enabled_extensions
.AMD_shader_info
;
1465 /* With update after bind we can't attach bo's to the command buffer
1466 * from the descriptor set anymore, so we have to use a global BO list.
1468 device
->use_global_bo_list
=
1469 device
->enabled_extensions
.EXT_descriptor_indexing
;
1471 mtx_init(&device
->shader_slab_mutex
, mtx_plain
);
1472 list_inithead(&device
->shader_slabs
);
1474 radv_bo_list_init(&device
->bo_list
);
1476 for (unsigned i
= 0; i
< pCreateInfo
->queueCreateInfoCount
; i
++) {
1477 const VkDeviceQueueCreateInfo
*queue_create
= &pCreateInfo
->pQueueCreateInfos
[i
];
1478 uint32_t qfi
= queue_create
->queueFamilyIndex
;
1479 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*global_priority
=
1480 vk_find_struct_const(queue_create
->pNext
, DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT
);
1482 assert(!global_priority
|| device
->physical_device
->rad_info
.has_ctx_priority
);
1484 device
->queues
[qfi
] = vk_alloc(&device
->alloc
,
1485 queue_create
->queueCount
* sizeof(struct radv_queue
), 8, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
1486 if (!device
->queues
[qfi
]) {
1487 result
= VK_ERROR_OUT_OF_HOST_MEMORY
;
1491 memset(device
->queues
[qfi
], 0, queue_create
->queueCount
* sizeof(struct radv_queue
));
1493 device
->queue_count
[qfi
] = queue_create
->queueCount
;
1495 for (unsigned q
= 0; q
< queue_create
->queueCount
; q
++) {
1496 result
= radv_queue_init(device
, &device
->queues
[qfi
][q
],
1497 qfi
, q
, queue_create
->flags
,
1499 if (result
!= VK_SUCCESS
)
1504 device
->pbb_allowed
= device
->physical_device
->rad_info
.chip_class
>= GFX9
&&
1505 (device
->instance
->perftest_flags
& RADV_PERFTEST_BINNING
);
1507 /* Disabled and not implemented for now. */
1508 device
->dfsm_allowed
= device
->pbb_allowed
&& false;
1511 device
->always_use_syncobj
= device
->physical_device
->rad_info
.has_syncobj_wait_for_submit
;
1514 /* The maximum number of scratch waves. Scratch space isn't divided
1515 * evenly between CUs. The number is only a function of the number of CUs.
1516 * We can decrease the constant to decrease the scratch buffer size.
1518 * sctx->scratch_waves must be >= the maximum possible size of
1519 * 1 threadgroup, so that the hw doesn't hang from being unable
1522 * The recommended value is 4 per CU at most. Higher numbers don't
1523 * bring much benefit, but they still occupy chip resources (think
1524 * async compute). I've seen ~2% performance difference between 4 and 32.
1526 uint32_t max_threads_per_block
= 2048;
1527 device
->scratch_waves
= MAX2(32 * physical_device
->rad_info
.num_good_compute_units
,
1528 max_threads_per_block
/ 64);
1530 device
->dispatch_initiator
= S_00B800_COMPUTE_SHADER_EN(1);
1532 if (device
->physical_device
->rad_info
.chip_class
>= CIK
) {
1533 /* If the KMD allows it (there is a KMD hw register for it),
1534 * allow launching waves out-of-order.
1536 device
->dispatch_initiator
|= S_00B800_ORDER_MODE(1);
1539 radv_device_init_gs_info(device
);
1541 device
->tess_offchip_block_dw_size
=
1542 device
->physical_device
->rad_info
.family
== CHIP_HAWAII
? 4096 : 8192;
1543 device
->has_distributed_tess
=
1544 device
->physical_device
->rad_info
.chip_class
>= VI
&&
1545 device
->physical_device
->rad_info
.max_se
>= 2;
1547 if (getenv("RADV_TRACE_FILE")) {
1548 const char *filename
= getenv("RADV_TRACE_FILE");
1550 keep_shader_info
= true;
1552 if (!radv_init_trace(device
))
1555 fprintf(stderr
, "Trace file will be dumped to %s\n", filename
);
1556 radv_dump_enabled_options(device
, stderr
);
1559 device
->keep_shader_info
= keep_shader_info
;
1561 result
= radv_device_init_meta(device
);
1562 if (result
!= VK_SUCCESS
)
1565 radv_device_init_msaa(device
);
1567 for (int family
= 0; family
< RADV_MAX_QUEUE_FAMILIES
; ++family
) {
1568 device
->empty_cs
[family
] = device
->ws
->cs_create(device
->ws
, family
);
1570 case RADV_QUEUE_GENERAL
:
1571 radeon_emit(device
->empty_cs
[family
], PKT3(PKT3_CONTEXT_CONTROL
, 1, 0));
1572 radeon_emit(device
->empty_cs
[family
], CONTEXT_CONTROL_LOAD_ENABLE(1));
1573 radeon_emit(device
->empty_cs
[family
], CONTEXT_CONTROL_SHADOW_ENABLE(1));
1575 case RADV_QUEUE_COMPUTE
:
1576 radeon_emit(device
->empty_cs
[family
], PKT3(PKT3_NOP
, 0, 0));
1577 radeon_emit(device
->empty_cs
[family
], 0);
1580 device
->ws
->cs_finalize(device
->empty_cs
[family
]);
1583 if (device
->physical_device
->rad_info
.chip_class
>= CIK
)
1584 cik_create_gfx_config(device
);
1586 VkPipelineCacheCreateInfo ci
;
1587 ci
.sType
= VK_STRUCTURE_TYPE_PIPELINE_CACHE_CREATE_INFO
;
1590 ci
.pInitialData
= NULL
;
1591 ci
.initialDataSize
= 0;
1593 result
= radv_CreatePipelineCache(radv_device_to_handle(device
),
1595 if (result
!= VK_SUCCESS
)
1598 device
->mem_cache
= radv_pipeline_cache_from_handle(pc
);
1600 *pDevice
= radv_device_to_handle(device
);
1604 radv_device_finish_meta(device
);
1606 radv_bo_list_finish(&device
->bo_list
);
1608 if (device
->trace_bo
)
1609 device
->ws
->buffer_destroy(device
->trace_bo
);
1611 if (device
->gfx_init
)
1612 device
->ws
->buffer_destroy(device
->gfx_init
);
1614 for (unsigned i
= 0; i
< RADV_MAX_QUEUE_FAMILIES
; i
++) {
1615 for (unsigned q
= 0; q
< device
->queue_count
[i
]; q
++)
1616 radv_queue_finish(&device
->queues
[i
][q
]);
1617 if (device
->queue_count
[i
])
1618 vk_free(&device
->alloc
, device
->queues
[i
]);
1621 vk_free(&device
->alloc
, device
);
1625 void radv_DestroyDevice(
1627 const VkAllocationCallbacks
* pAllocator
)
1629 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1634 if (device
->trace_bo
)
1635 device
->ws
->buffer_destroy(device
->trace_bo
);
1637 if (device
->gfx_init
)
1638 device
->ws
->buffer_destroy(device
->gfx_init
);
1640 for (unsigned i
= 0; i
< RADV_MAX_QUEUE_FAMILIES
; i
++) {
1641 for (unsigned q
= 0; q
< device
->queue_count
[i
]; q
++)
1642 radv_queue_finish(&device
->queues
[i
][q
]);
1643 if (device
->queue_count
[i
])
1644 vk_free(&device
->alloc
, device
->queues
[i
]);
1645 if (device
->empty_cs
[i
])
1646 device
->ws
->cs_destroy(device
->empty_cs
[i
]);
1648 radv_device_finish_meta(device
);
1650 VkPipelineCache pc
= radv_pipeline_cache_to_handle(device
->mem_cache
);
1651 radv_DestroyPipelineCache(radv_device_to_handle(device
), pc
, NULL
);
1653 radv_destroy_shader_slabs(device
);
1655 radv_bo_list_finish(&device
->bo_list
);
1656 vk_free(&device
->alloc
, device
);
1659 VkResult
radv_EnumerateInstanceLayerProperties(
1660 uint32_t* pPropertyCount
,
1661 VkLayerProperties
* pProperties
)
1663 if (pProperties
== NULL
) {
1664 *pPropertyCount
= 0;
1668 /* None supported at this time */
1669 return vk_error(NULL
, VK_ERROR_LAYER_NOT_PRESENT
);
1672 VkResult
radv_EnumerateDeviceLayerProperties(
1673 VkPhysicalDevice physicalDevice
,
1674 uint32_t* pPropertyCount
,
1675 VkLayerProperties
* pProperties
)
1677 if (pProperties
== NULL
) {
1678 *pPropertyCount
= 0;
1682 /* None supported at this time */
1683 return vk_error(NULL
, VK_ERROR_LAYER_NOT_PRESENT
);
1686 void radv_GetDeviceQueue2(
1688 const VkDeviceQueueInfo2
* pQueueInfo
,
1691 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1692 struct radv_queue
*queue
;
1694 queue
= &device
->queues
[pQueueInfo
->queueFamilyIndex
][pQueueInfo
->queueIndex
];
1695 if (pQueueInfo
->flags
!= queue
->flags
) {
1696 /* From the Vulkan 1.1.70 spec:
1698 * "The queue returned by vkGetDeviceQueue2 must have the same
1699 * flags value from this structure as that used at device
1700 * creation time in a VkDeviceQueueCreateInfo instance. If no
1701 * matching flags were specified at device creation time then
1702 * pQueue will return VK_NULL_HANDLE."
1704 *pQueue
= VK_NULL_HANDLE
;
1708 *pQueue
= radv_queue_to_handle(queue
);
1711 void radv_GetDeviceQueue(
1713 uint32_t queueFamilyIndex
,
1714 uint32_t queueIndex
,
1717 const VkDeviceQueueInfo2 info
= (VkDeviceQueueInfo2
) {
1718 .sType
= VK_STRUCTURE_TYPE_DEVICE_QUEUE_INFO_2
,
1719 .queueFamilyIndex
= queueFamilyIndex
,
1720 .queueIndex
= queueIndex
1723 radv_GetDeviceQueue2(_device
, &info
, pQueue
);
1727 fill_geom_tess_rings(struct radv_queue
*queue
,
1729 bool add_sample_positions
,
1730 uint32_t esgs_ring_size
,
1731 struct radeon_winsys_bo
*esgs_ring_bo
,
1732 uint32_t gsvs_ring_size
,
1733 struct radeon_winsys_bo
*gsvs_ring_bo
,
1734 uint32_t tess_factor_ring_size
,
1735 uint32_t tess_offchip_ring_offset
,
1736 uint32_t tess_offchip_ring_size
,
1737 struct radeon_winsys_bo
*tess_rings_bo
)
1739 uint64_t esgs_va
= 0, gsvs_va
= 0;
1740 uint64_t tess_va
= 0, tess_offchip_va
= 0;
1741 uint32_t *desc
= &map
[4];
1744 esgs_va
= radv_buffer_get_va(esgs_ring_bo
);
1746 gsvs_va
= radv_buffer_get_va(gsvs_ring_bo
);
1747 if (tess_rings_bo
) {
1748 tess_va
= radv_buffer_get_va(tess_rings_bo
);
1749 tess_offchip_va
= tess_va
+ tess_offchip_ring_offset
;
1752 /* stride 0, num records - size, add tid, swizzle, elsize4,
1755 desc
[1] = S_008F04_BASE_ADDRESS_HI(esgs_va
>> 32) |
1756 S_008F04_STRIDE(0) |
1757 S_008F04_SWIZZLE_ENABLE(true);
1758 desc
[2] = esgs_ring_size
;
1759 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1760 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1761 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1762 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1763 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1764 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1765 S_008F0C_ELEMENT_SIZE(1) |
1766 S_008F0C_INDEX_STRIDE(3) |
1767 S_008F0C_ADD_TID_ENABLE(true);
1770 /* GS entry for ES->GS ring */
1771 /* stride 0, num records - size, elsize0,
1774 desc
[1] = S_008F04_BASE_ADDRESS_HI(esgs_va
>> 32)|
1775 S_008F04_STRIDE(0) |
1776 S_008F04_SWIZZLE_ENABLE(false);
1777 desc
[2] = esgs_ring_size
;
1778 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1779 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1780 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1781 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1782 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1783 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1784 S_008F0C_ELEMENT_SIZE(0) |
1785 S_008F0C_INDEX_STRIDE(0) |
1786 S_008F0C_ADD_TID_ENABLE(false);
1789 /* VS entry for GS->VS ring */
1790 /* stride 0, num records - size, elsize0,
1793 desc
[1] = S_008F04_BASE_ADDRESS_HI(gsvs_va
>> 32)|
1794 S_008F04_STRIDE(0) |
1795 S_008F04_SWIZZLE_ENABLE(false);
1796 desc
[2] = gsvs_ring_size
;
1797 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1798 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1799 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1800 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1801 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1802 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1803 S_008F0C_ELEMENT_SIZE(0) |
1804 S_008F0C_INDEX_STRIDE(0) |
1805 S_008F0C_ADD_TID_ENABLE(false);
1808 /* stride gsvs_itemsize, num records 64
1809 elsize 4, index stride 16 */
1810 /* shader will patch stride and desc[2] */
1812 desc
[1] = S_008F04_BASE_ADDRESS_HI(gsvs_va
>> 32)|
1813 S_008F04_STRIDE(0) |
1814 S_008F04_SWIZZLE_ENABLE(true);
1816 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1817 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1818 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1819 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1820 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1821 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1822 S_008F0C_ELEMENT_SIZE(1) |
1823 S_008F0C_INDEX_STRIDE(1) |
1824 S_008F0C_ADD_TID_ENABLE(true);
1828 desc
[1] = S_008F04_BASE_ADDRESS_HI(tess_va
>> 32) |
1829 S_008F04_STRIDE(0) |
1830 S_008F04_SWIZZLE_ENABLE(false);
1831 desc
[2] = tess_factor_ring_size
;
1832 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1833 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1834 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1835 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1836 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1837 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1838 S_008F0C_ELEMENT_SIZE(0) |
1839 S_008F0C_INDEX_STRIDE(0) |
1840 S_008F0C_ADD_TID_ENABLE(false);
1843 desc
[0] = tess_offchip_va
;
1844 desc
[1] = S_008F04_BASE_ADDRESS_HI(tess_offchip_va
>> 32) |
1845 S_008F04_STRIDE(0) |
1846 S_008F04_SWIZZLE_ENABLE(false);
1847 desc
[2] = tess_offchip_ring_size
;
1848 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1849 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1850 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1851 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1852 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1853 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1854 S_008F0C_ELEMENT_SIZE(0) |
1855 S_008F0C_INDEX_STRIDE(0) |
1856 S_008F0C_ADD_TID_ENABLE(false);
1859 /* add sample positions after all rings */
1860 memcpy(desc
, queue
->device
->sample_locations_1x
, 8);
1862 memcpy(desc
, queue
->device
->sample_locations_2x
, 16);
1864 memcpy(desc
, queue
->device
->sample_locations_4x
, 32);
1866 memcpy(desc
, queue
->device
->sample_locations_8x
, 64);
1868 memcpy(desc
, queue
->device
->sample_locations_16x
, 128);
1872 radv_get_hs_offchip_param(struct radv_device
*device
, uint32_t *max_offchip_buffers_p
)
1874 bool double_offchip_buffers
= device
->physical_device
->rad_info
.chip_class
>= CIK
&&
1875 device
->physical_device
->rad_info
.family
!= CHIP_CARRIZO
&&
1876 device
->physical_device
->rad_info
.family
!= CHIP_STONEY
;
1877 unsigned max_offchip_buffers_per_se
= double_offchip_buffers
? 128 : 64;
1878 unsigned max_offchip_buffers
= max_offchip_buffers_per_se
*
1879 device
->physical_device
->rad_info
.max_se
;
1880 unsigned offchip_granularity
;
1881 unsigned hs_offchip_param
;
1882 switch (device
->tess_offchip_block_dw_size
) {
1887 offchip_granularity
= V_03093C_X_8K_DWORDS
;
1890 offchip_granularity
= V_03093C_X_4K_DWORDS
;
1894 switch (device
->physical_device
->rad_info
.chip_class
) {
1896 max_offchip_buffers
= MIN2(max_offchip_buffers
, 126);
1902 max_offchip_buffers
= MIN2(max_offchip_buffers
, 508);
1906 *max_offchip_buffers_p
= max_offchip_buffers
;
1907 if (device
->physical_device
->rad_info
.chip_class
>= CIK
) {
1908 if (device
->physical_device
->rad_info
.chip_class
>= VI
)
1909 --max_offchip_buffers
;
1911 S_03093C_OFFCHIP_BUFFERING(max_offchip_buffers
) |
1912 S_03093C_OFFCHIP_GRANULARITY(offchip_granularity
);
1915 S_0089B0_OFFCHIP_BUFFERING(max_offchip_buffers
);
1917 return hs_offchip_param
;
1921 radv_emit_gs_ring_sizes(struct radv_queue
*queue
, struct radeon_cmdbuf
*cs
,
1922 struct radeon_winsys_bo
*esgs_ring_bo
,
1923 uint32_t esgs_ring_size
,
1924 struct radeon_winsys_bo
*gsvs_ring_bo
,
1925 uint32_t gsvs_ring_size
)
1927 if (!esgs_ring_bo
&& !gsvs_ring_bo
)
1931 radv_cs_add_buffer(queue
->device
->ws
, cs
, esgs_ring_bo
, 8);
1934 radv_cs_add_buffer(queue
->device
->ws
, cs
, gsvs_ring_bo
, 8);
1936 if (queue
->device
->physical_device
->rad_info
.chip_class
>= CIK
) {
1937 radeon_set_uconfig_reg_seq(cs
, R_030900_VGT_ESGS_RING_SIZE
, 2);
1938 radeon_emit(cs
, esgs_ring_size
>> 8);
1939 radeon_emit(cs
, gsvs_ring_size
>> 8);
1941 radeon_set_config_reg_seq(cs
, R_0088C8_VGT_ESGS_RING_SIZE
, 2);
1942 radeon_emit(cs
, esgs_ring_size
>> 8);
1943 radeon_emit(cs
, gsvs_ring_size
>> 8);
1948 radv_emit_tess_factor_ring(struct radv_queue
*queue
, struct radeon_cmdbuf
*cs
,
1949 unsigned hs_offchip_param
, unsigned tf_ring_size
,
1950 struct radeon_winsys_bo
*tess_rings_bo
)
1957 tf_va
= radv_buffer_get_va(tess_rings_bo
);
1959 radv_cs_add_buffer(queue
->device
->ws
, cs
, tess_rings_bo
, 8);
1961 if (queue
->device
->physical_device
->rad_info
.chip_class
>= CIK
) {
1962 radeon_set_uconfig_reg(cs
, R_030938_VGT_TF_RING_SIZE
,
1963 S_030938_SIZE(tf_ring_size
/ 4));
1964 radeon_set_uconfig_reg(cs
, R_030940_VGT_TF_MEMORY_BASE
,
1966 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
1967 radeon_set_uconfig_reg(cs
, R_030944_VGT_TF_MEMORY_BASE_HI
,
1968 S_030944_BASE_HI(tf_va
>> 40));
1970 radeon_set_uconfig_reg(cs
, R_03093C_VGT_HS_OFFCHIP_PARAM
,
1973 radeon_set_config_reg(cs
, R_008988_VGT_TF_RING_SIZE
,
1974 S_008988_SIZE(tf_ring_size
/ 4));
1975 radeon_set_config_reg(cs
, R_0089B8_VGT_TF_MEMORY_BASE
,
1977 radeon_set_config_reg(cs
, R_0089B0_VGT_HS_OFFCHIP_PARAM
,
1983 radv_emit_compute_scratch(struct radv_queue
*queue
, struct radeon_cmdbuf
*cs
,
1984 struct radeon_winsys_bo
*compute_scratch_bo
)
1986 uint64_t scratch_va
;
1988 if (!compute_scratch_bo
)
1991 scratch_va
= radv_buffer_get_va(compute_scratch_bo
);
1993 radv_cs_add_buffer(queue
->device
->ws
, cs
, compute_scratch_bo
, 8);
1995 radeon_set_sh_reg_seq(cs
, R_00B900_COMPUTE_USER_DATA_0
, 2);
1996 radeon_emit(cs
, scratch_va
);
1997 radeon_emit(cs
, S_008F04_BASE_ADDRESS_HI(scratch_va
>> 32) |
1998 S_008F04_SWIZZLE_ENABLE(1));
2002 radv_emit_global_shader_pointers(struct radv_queue
*queue
,
2003 struct radeon_cmdbuf
*cs
,
2004 struct radeon_winsys_bo
*descriptor_bo
)
2011 va
= radv_buffer_get_va(descriptor_bo
);
2013 radv_cs_add_buffer(queue
->device
->ws
, cs
, descriptor_bo
, 8);
2015 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
2016 uint32_t regs
[] = {R_00B030_SPI_SHADER_USER_DATA_PS_0
,
2017 R_00B130_SPI_SHADER_USER_DATA_VS_0
,
2018 R_00B208_SPI_SHADER_USER_DATA_ADDR_LO_GS
,
2019 R_00B408_SPI_SHADER_USER_DATA_ADDR_LO_HS
};
2021 for (int i
= 0; i
< ARRAY_SIZE(regs
); ++i
) {
2022 radv_emit_shader_pointer(queue
->device
, cs
, regs
[i
],
2026 uint32_t regs
[] = {R_00B030_SPI_SHADER_USER_DATA_PS_0
,
2027 R_00B130_SPI_SHADER_USER_DATA_VS_0
,
2028 R_00B230_SPI_SHADER_USER_DATA_GS_0
,
2029 R_00B330_SPI_SHADER_USER_DATA_ES_0
,
2030 R_00B430_SPI_SHADER_USER_DATA_HS_0
,
2031 R_00B530_SPI_SHADER_USER_DATA_LS_0
};
2033 for (int i
= 0; i
< ARRAY_SIZE(regs
); ++i
) {
2034 radv_emit_shader_pointer(queue
->device
, cs
, regs
[i
],
2041 radv_get_preamble_cs(struct radv_queue
*queue
,
2042 uint32_t scratch_size
,
2043 uint32_t compute_scratch_size
,
2044 uint32_t esgs_ring_size
,
2045 uint32_t gsvs_ring_size
,
2046 bool needs_tess_rings
,
2047 bool needs_sample_positions
,
2048 struct radeon_cmdbuf
**initial_full_flush_preamble_cs
,
2049 struct radeon_cmdbuf
**initial_preamble_cs
,
2050 struct radeon_cmdbuf
**continue_preamble_cs
)
2052 struct radeon_winsys_bo
*scratch_bo
= NULL
;
2053 struct radeon_winsys_bo
*descriptor_bo
= NULL
;
2054 struct radeon_winsys_bo
*compute_scratch_bo
= NULL
;
2055 struct radeon_winsys_bo
*esgs_ring_bo
= NULL
;
2056 struct radeon_winsys_bo
*gsvs_ring_bo
= NULL
;
2057 struct radeon_winsys_bo
*tess_rings_bo
= NULL
;
2058 struct radeon_cmdbuf
*dest_cs
[3] = {0};
2059 bool add_tess_rings
= false, add_sample_positions
= false;
2060 unsigned tess_factor_ring_size
= 0, tess_offchip_ring_size
= 0;
2061 unsigned max_offchip_buffers
;
2062 unsigned hs_offchip_param
= 0;
2063 unsigned tess_offchip_ring_offset
;
2064 uint32_t ring_bo_flags
= RADEON_FLAG_NO_CPU_ACCESS
| RADEON_FLAG_NO_INTERPROCESS_SHARING
;
2065 if (!queue
->has_tess_rings
) {
2066 if (needs_tess_rings
)
2067 add_tess_rings
= true;
2069 if (!queue
->has_sample_positions
) {
2070 if (needs_sample_positions
)
2071 add_sample_positions
= true;
2073 tess_factor_ring_size
= 32768 * queue
->device
->physical_device
->rad_info
.max_se
;
2074 hs_offchip_param
= radv_get_hs_offchip_param(queue
->device
,
2075 &max_offchip_buffers
);
2076 tess_offchip_ring_offset
= align(tess_factor_ring_size
, 64 * 1024);
2077 tess_offchip_ring_size
= max_offchip_buffers
*
2078 queue
->device
->tess_offchip_block_dw_size
* 4;
2080 if (scratch_size
<= queue
->scratch_size
&&
2081 compute_scratch_size
<= queue
->compute_scratch_size
&&
2082 esgs_ring_size
<= queue
->esgs_ring_size
&&
2083 gsvs_ring_size
<= queue
->gsvs_ring_size
&&
2084 !add_tess_rings
&& !add_sample_positions
&&
2085 queue
->initial_preamble_cs
) {
2086 *initial_full_flush_preamble_cs
= queue
->initial_full_flush_preamble_cs
;
2087 *initial_preamble_cs
= queue
->initial_preamble_cs
;
2088 *continue_preamble_cs
= queue
->continue_preamble_cs
;
2089 if (!scratch_size
&& !compute_scratch_size
&& !esgs_ring_size
&& !gsvs_ring_size
)
2090 *continue_preamble_cs
= NULL
;
2094 if (scratch_size
> queue
->scratch_size
) {
2095 scratch_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
2103 scratch_bo
= queue
->scratch_bo
;
2105 if (compute_scratch_size
> queue
->compute_scratch_size
) {
2106 compute_scratch_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
2107 compute_scratch_size
,
2111 if (!compute_scratch_bo
)
2115 compute_scratch_bo
= queue
->compute_scratch_bo
;
2117 if (esgs_ring_size
> queue
->esgs_ring_size
) {
2118 esgs_ring_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
2126 esgs_ring_bo
= queue
->esgs_ring_bo
;
2127 esgs_ring_size
= queue
->esgs_ring_size
;
2130 if (gsvs_ring_size
> queue
->gsvs_ring_size
) {
2131 gsvs_ring_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
2139 gsvs_ring_bo
= queue
->gsvs_ring_bo
;
2140 gsvs_ring_size
= queue
->gsvs_ring_size
;
2143 if (add_tess_rings
) {
2144 tess_rings_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
2145 tess_offchip_ring_offset
+ tess_offchip_ring_size
,
2152 tess_rings_bo
= queue
->tess_rings_bo
;
2155 if (scratch_bo
!= queue
->scratch_bo
||
2156 esgs_ring_bo
!= queue
->esgs_ring_bo
||
2157 gsvs_ring_bo
!= queue
->gsvs_ring_bo
||
2158 tess_rings_bo
!= queue
->tess_rings_bo
||
2159 add_sample_positions
) {
2161 if (gsvs_ring_bo
|| esgs_ring_bo
||
2162 tess_rings_bo
|| add_sample_positions
) {
2163 size
= 112; /* 2 dword + 2 padding + 4 dword * 6 */
2164 if (add_sample_positions
)
2165 size
+= 256; /* 32+16+8+4+2+1 samples * 4 * 2 = 248 bytes. */
2167 else if (scratch_bo
)
2168 size
= 8; /* 2 dword */
2170 descriptor_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
2174 RADEON_FLAG_CPU_ACCESS
|
2175 RADEON_FLAG_NO_INTERPROCESS_SHARING
|
2176 RADEON_FLAG_READ_ONLY
);
2180 descriptor_bo
= queue
->descriptor_bo
;
2182 for(int i
= 0; i
< 3; ++i
) {
2183 struct radeon_cmdbuf
*cs
= NULL
;
2184 cs
= queue
->device
->ws
->cs_create(queue
->device
->ws
,
2185 queue
->queue_family_index
? RING_COMPUTE
: RING_GFX
);
2192 radv_cs_add_buffer(queue
->device
->ws
, cs
, scratch_bo
, 8);
2194 if (descriptor_bo
!= queue
->descriptor_bo
) {
2195 uint32_t *map
= (uint32_t*)queue
->device
->ws
->buffer_map(descriptor_bo
);
2198 uint64_t scratch_va
= radv_buffer_get_va(scratch_bo
);
2199 uint32_t rsrc1
= S_008F04_BASE_ADDRESS_HI(scratch_va
>> 32) |
2200 S_008F04_SWIZZLE_ENABLE(1);
2201 map
[0] = scratch_va
;
2205 if (esgs_ring_bo
|| gsvs_ring_bo
|| tess_rings_bo
||
2206 add_sample_positions
)
2207 fill_geom_tess_rings(queue
, map
, add_sample_positions
,
2208 esgs_ring_size
, esgs_ring_bo
,
2209 gsvs_ring_size
, gsvs_ring_bo
,
2210 tess_factor_ring_size
,
2211 tess_offchip_ring_offset
,
2212 tess_offchip_ring_size
,
2215 queue
->device
->ws
->buffer_unmap(descriptor_bo
);
2218 if (esgs_ring_bo
|| gsvs_ring_bo
|| tess_rings_bo
) {
2219 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
2220 radeon_emit(cs
, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
2221 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
2222 radeon_emit(cs
, EVENT_TYPE(V_028A90_VGT_FLUSH
) | EVENT_INDEX(0));
2225 radv_emit_gs_ring_sizes(queue
, cs
, esgs_ring_bo
, esgs_ring_size
,
2226 gsvs_ring_bo
, gsvs_ring_size
);
2227 radv_emit_tess_factor_ring(queue
, cs
, hs_offchip_param
,
2228 tess_factor_ring_size
, tess_rings_bo
);
2229 radv_emit_global_shader_pointers(queue
, cs
, descriptor_bo
);
2230 radv_emit_compute_scratch(queue
, cs
, compute_scratch_bo
);
2233 si_cs_emit_cache_flush(cs
,
2234 queue
->device
->physical_device
->rad_info
.chip_class
,
2236 queue
->queue_family_index
== RING_COMPUTE
&&
2237 queue
->device
->physical_device
->rad_info
.chip_class
>= CIK
,
2238 (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
)) |
2239 RADV_CMD_FLAG_INV_ICACHE
|
2240 RADV_CMD_FLAG_INV_SMEM_L1
|
2241 RADV_CMD_FLAG_INV_VMEM_L1
|
2242 RADV_CMD_FLAG_INV_GLOBAL_L2
|
2243 RADV_CMD_FLAG_START_PIPELINE_STATS
, 0);
2244 } else if (i
== 1) {
2245 si_cs_emit_cache_flush(cs
,
2246 queue
->device
->physical_device
->rad_info
.chip_class
,
2248 queue
->queue_family_index
== RING_COMPUTE
&&
2249 queue
->device
->physical_device
->rad_info
.chip_class
>= CIK
,
2250 RADV_CMD_FLAG_INV_ICACHE
|
2251 RADV_CMD_FLAG_INV_SMEM_L1
|
2252 RADV_CMD_FLAG_INV_VMEM_L1
|
2253 RADV_CMD_FLAG_INV_GLOBAL_L2
|
2254 RADV_CMD_FLAG_START_PIPELINE_STATS
, 0);
2257 if (!queue
->device
->ws
->cs_finalize(cs
))
2261 if (queue
->initial_full_flush_preamble_cs
)
2262 queue
->device
->ws
->cs_destroy(queue
->initial_full_flush_preamble_cs
);
2264 if (queue
->initial_preamble_cs
)
2265 queue
->device
->ws
->cs_destroy(queue
->initial_preamble_cs
);
2267 if (queue
->continue_preamble_cs
)
2268 queue
->device
->ws
->cs_destroy(queue
->continue_preamble_cs
);
2270 queue
->initial_full_flush_preamble_cs
= dest_cs
[0];
2271 queue
->initial_preamble_cs
= dest_cs
[1];
2272 queue
->continue_preamble_cs
= dest_cs
[2];
2274 if (scratch_bo
!= queue
->scratch_bo
) {
2275 if (queue
->scratch_bo
)
2276 queue
->device
->ws
->buffer_destroy(queue
->scratch_bo
);
2277 queue
->scratch_bo
= scratch_bo
;
2278 queue
->scratch_size
= scratch_size
;
2281 if (compute_scratch_bo
!= queue
->compute_scratch_bo
) {
2282 if (queue
->compute_scratch_bo
)
2283 queue
->device
->ws
->buffer_destroy(queue
->compute_scratch_bo
);
2284 queue
->compute_scratch_bo
= compute_scratch_bo
;
2285 queue
->compute_scratch_size
= compute_scratch_size
;
2288 if (esgs_ring_bo
!= queue
->esgs_ring_bo
) {
2289 if (queue
->esgs_ring_bo
)
2290 queue
->device
->ws
->buffer_destroy(queue
->esgs_ring_bo
);
2291 queue
->esgs_ring_bo
= esgs_ring_bo
;
2292 queue
->esgs_ring_size
= esgs_ring_size
;
2295 if (gsvs_ring_bo
!= queue
->gsvs_ring_bo
) {
2296 if (queue
->gsvs_ring_bo
)
2297 queue
->device
->ws
->buffer_destroy(queue
->gsvs_ring_bo
);
2298 queue
->gsvs_ring_bo
= gsvs_ring_bo
;
2299 queue
->gsvs_ring_size
= gsvs_ring_size
;
2302 if (tess_rings_bo
!= queue
->tess_rings_bo
) {
2303 queue
->tess_rings_bo
= tess_rings_bo
;
2304 queue
->has_tess_rings
= true;
2307 if (descriptor_bo
!= queue
->descriptor_bo
) {
2308 if (queue
->descriptor_bo
)
2309 queue
->device
->ws
->buffer_destroy(queue
->descriptor_bo
);
2311 queue
->descriptor_bo
= descriptor_bo
;
2314 if (add_sample_positions
)
2315 queue
->has_sample_positions
= true;
2317 *initial_full_flush_preamble_cs
= queue
->initial_full_flush_preamble_cs
;
2318 *initial_preamble_cs
= queue
->initial_preamble_cs
;
2319 *continue_preamble_cs
= queue
->continue_preamble_cs
;
2320 if (!scratch_size
&& !compute_scratch_size
&& !esgs_ring_size
&& !gsvs_ring_size
)
2321 *continue_preamble_cs
= NULL
;
2324 for (int i
= 0; i
< ARRAY_SIZE(dest_cs
); ++i
)
2326 queue
->device
->ws
->cs_destroy(dest_cs
[i
]);
2327 if (descriptor_bo
&& descriptor_bo
!= queue
->descriptor_bo
)
2328 queue
->device
->ws
->buffer_destroy(descriptor_bo
);
2329 if (scratch_bo
&& scratch_bo
!= queue
->scratch_bo
)
2330 queue
->device
->ws
->buffer_destroy(scratch_bo
);
2331 if (compute_scratch_bo
&& compute_scratch_bo
!= queue
->compute_scratch_bo
)
2332 queue
->device
->ws
->buffer_destroy(compute_scratch_bo
);
2333 if (esgs_ring_bo
&& esgs_ring_bo
!= queue
->esgs_ring_bo
)
2334 queue
->device
->ws
->buffer_destroy(esgs_ring_bo
);
2335 if (gsvs_ring_bo
&& gsvs_ring_bo
!= queue
->gsvs_ring_bo
)
2336 queue
->device
->ws
->buffer_destroy(gsvs_ring_bo
);
2337 if (tess_rings_bo
&& tess_rings_bo
!= queue
->tess_rings_bo
)
2338 queue
->device
->ws
->buffer_destroy(tess_rings_bo
);
2339 return vk_error(queue
->device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
2342 static VkResult
radv_alloc_sem_counts(struct radv_instance
*instance
,
2343 struct radv_winsys_sem_counts
*counts
,
2345 const VkSemaphore
*sems
,
2349 int syncobj_idx
= 0, sem_idx
= 0;
2351 if (num_sems
== 0 && _fence
== VK_NULL_HANDLE
)
2354 for (uint32_t i
= 0; i
< num_sems
; i
++) {
2355 RADV_FROM_HANDLE(radv_semaphore
, sem
, sems
[i
]);
2357 if (sem
->temp_syncobj
|| sem
->syncobj
)
2358 counts
->syncobj_count
++;
2360 counts
->sem_count
++;
2363 if (_fence
!= VK_NULL_HANDLE
) {
2364 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
2365 if (fence
->temp_syncobj
|| fence
->syncobj
)
2366 counts
->syncobj_count
++;
2369 if (counts
->syncobj_count
) {
2370 counts
->syncobj
= (uint32_t *)malloc(sizeof(uint32_t) * counts
->syncobj_count
);
2371 if (!counts
->syncobj
)
2372 return vk_error(instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
2375 if (counts
->sem_count
) {
2376 counts
->sem
= (struct radeon_winsys_sem
**)malloc(sizeof(struct radeon_winsys_sem
*) * counts
->sem_count
);
2378 free(counts
->syncobj
);
2379 return vk_error(instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
2383 for (uint32_t i
= 0; i
< num_sems
; i
++) {
2384 RADV_FROM_HANDLE(radv_semaphore
, sem
, sems
[i
]);
2386 if (sem
->temp_syncobj
) {
2387 counts
->syncobj
[syncobj_idx
++] = sem
->temp_syncobj
;
2389 else if (sem
->syncobj
)
2390 counts
->syncobj
[syncobj_idx
++] = sem
->syncobj
;
2393 counts
->sem
[sem_idx
++] = sem
->sem
;
2397 if (_fence
!= VK_NULL_HANDLE
) {
2398 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
2399 if (fence
->temp_syncobj
)
2400 counts
->syncobj
[syncobj_idx
++] = fence
->temp_syncobj
;
2401 else if (fence
->syncobj
)
2402 counts
->syncobj
[syncobj_idx
++] = fence
->syncobj
;
2409 radv_free_sem_info(struct radv_winsys_sem_info
*sem_info
)
2411 free(sem_info
->wait
.syncobj
);
2412 free(sem_info
->wait
.sem
);
2413 free(sem_info
->signal
.syncobj
);
2414 free(sem_info
->signal
.sem
);
2418 static void radv_free_temp_syncobjs(struct radv_device
*device
,
2420 const VkSemaphore
*sems
)
2422 for (uint32_t i
= 0; i
< num_sems
; i
++) {
2423 RADV_FROM_HANDLE(radv_semaphore
, sem
, sems
[i
]);
2425 if (sem
->temp_syncobj
) {
2426 device
->ws
->destroy_syncobj(device
->ws
, sem
->temp_syncobj
);
2427 sem
->temp_syncobj
= 0;
2433 radv_alloc_sem_info(struct radv_instance
*instance
,
2434 struct radv_winsys_sem_info
*sem_info
,
2436 const VkSemaphore
*wait_sems
,
2437 int num_signal_sems
,
2438 const VkSemaphore
*signal_sems
,
2442 memset(sem_info
, 0, sizeof(*sem_info
));
2444 ret
= radv_alloc_sem_counts(instance
, &sem_info
->wait
, num_wait_sems
, wait_sems
, VK_NULL_HANDLE
, true);
2447 ret
= radv_alloc_sem_counts(instance
, &sem_info
->signal
, num_signal_sems
, signal_sems
, fence
, false);
2449 radv_free_sem_info(sem_info
);
2451 /* caller can override these */
2452 sem_info
->cs_emit_wait
= true;
2453 sem_info
->cs_emit_signal
= true;
2457 /* Signals fence as soon as all the work currently put on queue is done. */
2458 static VkResult
radv_signal_fence(struct radv_queue
*queue
,
2459 struct radv_fence
*fence
)
2463 struct radv_winsys_sem_info sem_info
;
2465 result
= radv_alloc_sem_info(queue
->device
->instance
, &sem_info
, 0, NULL
, 0, NULL
,
2466 radv_fence_to_handle(fence
));
2467 if (result
!= VK_SUCCESS
)
2470 ret
= queue
->device
->ws
->cs_submit(queue
->hw_ctx
, queue
->queue_idx
,
2471 &queue
->device
->empty_cs
[queue
->queue_family_index
],
2472 1, NULL
, NULL
, &sem_info
, NULL
,
2473 false, fence
->fence
);
2474 radv_free_sem_info(&sem_info
);
2477 return vk_error(queue
->device
->instance
, VK_ERROR_DEVICE_LOST
);
2482 VkResult
radv_QueueSubmit(
2484 uint32_t submitCount
,
2485 const VkSubmitInfo
* pSubmits
,
2488 RADV_FROM_HANDLE(radv_queue
, queue
, _queue
);
2489 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
2490 struct radeon_winsys_fence
*base_fence
= fence
? fence
->fence
: NULL
;
2491 struct radeon_winsys_ctx
*ctx
= queue
->hw_ctx
;
2493 uint32_t max_cs_submission
= queue
->device
->trace_bo
? 1 : UINT32_MAX
;
2494 uint32_t scratch_size
= 0;
2495 uint32_t compute_scratch_size
= 0;
2496 uint32_t esgs_ring_size
= 0, gsvs_ring_size
= 0;
2497 struct radeon_cmdbuf
*initial_preamble_cs
= NULL
, *initial_flush_preamble_cs
= NULL
, *continue_preamble_cs
= NULL
;
2499 bool fence_emitted
= false;
2500 bool tess_rings_needed
= false;
2501 bool sample_positions_needed
= false;
2503 /* Do this first so failing to allocate scratch buffers can't result in
2504 * partially executed submissions. */
2505 for (uint32_t i
= 0; i
< submitCount
; i
++) {
2506 for (uint32_t j
= 0; j
< pSubmits
[i
].commandBufferCount
; j
++) {
2507 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
,
2508 pSubmits
[i
].pCommandBuffers
[j
]);
2510 scratch_size
= MAX2(scratch_size
, cmd_buffer
->scratch_size_needed
);
2511 compute_scratch_size
= MAX2(compute_scratch_size
,
2512 cmd_buffer
->compute_scratch_size_needed
);
2513 esgs_ring_size
= MAX2(esgs_ring_size
, cmd_buffer
->esgs_ring_size_needed
);
2514 gsvs_ring_size
= MAX2(gsvs_ring_size
, cmd_buffer
->gsvs_ring_size_needed
);
2515 tess_rings_needed
|= cmd_buffer
->tess_rings_needed
;
2516 sample_positions_needed
|= cmd_buffer
->sample_positions_needed
;
2520 result
= radv_get_preamble_cs(queue
, scratch_size
, compute_scratch_size
,
2521 esgs_ring_size
, gsvs_ring_size
, tess_rings_needed
,
2522 sample_positions_needed
, &initial_flush_preamble_cs
,
2523 &initial_preamble_cs
, &continue_preamble_cs
);
2524 if (result
!= VK_SUCCESS
)
2527 for (uint32_t i
= 0; i
< submitCount
; i
++) {
2528 struct radeon_cmdbuf
**cs_array
;
2529 bool do_flush
= !i
|| pSubmits
[i
].pWaitDstStageMask
;
2530 bool can_patch
= true;
2532 struct radv_winsys_sem_info sem_info
;
2534 result
= radv_alloc_sem_info(queue
->device
->instance
,
2536 pSubmits
[i
].waitSemaphoreCount
,
2537 pSubmits
[i
].pWaitSemaphores
,
2538 pSubmits
[i
].signalSemaphoreCount
,
2539 pSubmits
[i
].pSignalSemaphores
,
2541 if (result
!= VK_SUCCESS
)
2544 if (!pSubmits
[i
].commandBufferCount
) {
2545 if (pSubmits
[i
].waitSemaphoreCount
|| pSubmits
[i
].signalSemaphoreCount
) {
2546 ret
= queue
->device
->ws
->cs_submit(ctx
, queue
->queue_idx
,
2547 &queue
->device
->empty_cs
[queue
->queue_family_index
],
2552 radv_loge("failed to submit CS %d\n", i
);
2555 fence_emitted
= true;
2557 radv_free_sem_info(&sem_info
);
2561 cs_array
= malloc(sizeof(struct radeon_cmdbuf
*) *
2562 (pSubmits
[i
].commandBufferCount
));
2564 for (uint32_t j
= 0; j
< pSubmits
[i
].commandBufferCount
; j
++) {
2565 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
,
2566 pSubmits
[i
].pCommandBuffers
[j
]);
2567 assert(cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_PRIMARY
);
2569 cs_array
[j
] = cmd_buffer
->cs
;
2570 if ((cmd_buffer
->usage_flags
& VK_COMMAND_BUFFER_USAGE_SIMULTANEOUS_USE_BIT
))
2573 cmd_buffer
->status
= RADV_CMD_BUFFER_STATUS_PENDING
;
2576 for (uint32_t j
= 0; j
< pSubmits
[i
].commandBufferCount
; j
+= advance
) {
2577 struct radeon_cmdbuf
*initial_preamble
= (do_flush
&& !j
) ? initial_flush_preamble_cs
: initial_preamble_cs
;
2578 const struct radv_winsys_bo_list
*bo_list
= NULL
;
2580 advance
= MIN2(max_cs_submission
,
2581 pSubmits
[i
].commandBufferCount
- j
);
2583 if (queue
->device
->trace_bo
)
2584 *queue
->device
->trace_id_ptr
= 0;
2586 sem_info
.cs_emit_wait
= j
== 0;
2587 sem_info
.cs_emit_signal
= j
+ advance
== pSubmits
[i
].commandBufferCount
;
2589 if (unlikely(queue
->device
->use_global_bo_list
)) {
2590 pthread_mutex_lock(&queue
->device
->bo_list
.mutex
);
2591 bo_list
= &queue
->device
->bo_list
.list
;
2594 ret
= queue
->device
->ws
->cs_submit(ctx
, queue
->queue_idx
, cs_array
+ j
,
2595 advance
, initial_preamble
, continue_preamble_cs
,
2597 can_patch
, base_fence
);
2599 if (unlikely(queue
->device
->use_global_bo_list
))
2600 pthread_mutex_unlock(&queue
->device
->bo_list
.mutex
);
2603 radv_loge("failed to submit CS %d\n", i
);
2606 fence_emitted
= true;
2607 if (queue
->device
->trace_bo
) {
2608 radv_check_gpu_hangs(queue
, cs_array
[j
]);
2612 radv_free_temp_syncobjs(queue
->device
,
2613 pSubmits
[i
].waitSemaphoreCount
,
2614 pSubmits
[i
].pWaitSemaphores
);
2615 radv_free_sem_info(&sem_info
);
2620 if (!fence_emitted
) {
2621 result
= radv_signal_fence(queue
, fence
);
2622 if (result
!= VK_SUCCESS
)
2625 fence
->submitted
= true;
2631 VkResult
radv_QueueWaitIdle(
2634 RADV_FROM_HANDLE(radv_queue
, queue
, _queue
);
2636 queue
->device
->ws
->ctx_wait_idle(queue
->hw_ctx
,
2637 radv_queue_family_to_ring(queue
->queue_family_index
),
2642 VkResult
radv_DeviceWaitIdle(
2645 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2647 for (unsigned i
= 0; i
< RADV_MAX_QUEUE_FAMILIES
; i
++) {
2648 for (unsigned q
= 0; q
< device
->queue_count
[i
]; q
++) {
2649 radv_QueueWaitIdle(radv_queue_to_handle(&device
->queues
[i
][q
]));
2655 VkResult
radv_EnumerateInstanceExtensionProperties(
2656 const char* pLayerName
,
2657 uint32_t* pPropertyCount
,
2658 VkExtensionProperties
* pProperties
)
2660 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
2662 for (int i
= 0; i
< RADV_INSTANCE_EXTENSION_COUNT
; i
++) {
2663 if (radv_supported_instance_extensions
.extensions
[i
]) {
2664 vk_outarray_append(&out
, prop
) {
2665 *prop
= radv_instance_extensions
[i
];
2670 return vk_outarray_status(&out
);
2673 VkResult
radv_EnumerateDeviceExtensionProperties(
2674 VkPhysicalDevice physicalDevice
,
2675 const char* pLayerName
,
2676 uint32_t* pPropertyCount
,
2677 VkExtensionProperties
* pProperties
)
2679 RADV_FROM_HANDLE(radv_physical_device
, device
, physicalDevice
);
2680 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
2682 for (int i
= 0; i
< RADV_DEVICE_EXTENSION_COUNT
; i
++) {
2683 if (device
->supported_extensions
.extensions
[i
]) {
2684 vk_outarray_append(&out
, prop
) {
2685 *prop
= radv_device_extensions
[i
];
2690 return vk_outarray_status(&out
);
2693 PFN_vkVoidFunction
radv_GetInstanceProcAddr(
2694 VkInstance _instance
,
2697 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
2699 return radv_lookup_entrypoint_checked(pName
,
2700 instance
? instance
->apiVersion
: 0,
2701 instance
? &instance
->enabled_extensions
: NULL
,
2705 /* The loader wants us to expose a second GetInstanceProcAddr function
2706 * to work around certain LD_PRELOAD issues seen in apps.
2709 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
2710 VkInstance instance
,
2714 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
2715 VkInstance instance
,
2718 return radv_GetInstanceProcAddr(instance
, pName
);
2721 PFN_vkVoidFunction
radv_GetDeviceProcAddr(
2725 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2727 return radv_lookup_entrypoint_checked(pName
,
2728 device
->instance
->apiVersion
,
2729 &device
->instance
->enabled_extensions
,
2730 &device
->enabled_extensions
);
2733 bool radv_get_memory_fd(struct radv_device
*device
,
2734 struct radv_device_memory
*memory
,
2737 struct radeon_bo_metadata metadata
;
2739 if (memory
->image
) {
2740 radv_init_metadata(device
, memory
->image
, &metadata
);
2741 device
->ws
->buffer_set_metadata(memory
->bo
, &metadata
);
2744 return device
->ws
->buffer_get_fd(device
->ws
, memory
->bo
,
2748 static VkResult
radv_alloc_memory(struct radv_device
*device
,
2749 const VkMemoryAllocateInfo
* pAllocateInfo
,
2750 const VkAllocationCallbacks
* pAllocator
,
2751 VkDeviceMemory
* pMem
)
2753 struct radv_device_memory
*mem
;
2755 enum radeon_bo_domain domain
;
2757 enum radv_mem_type mem_type_index
= device
->physical_device
->mem_type_indices
[pAllocateInfo
->memoryTypeIndex
];
2759 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
2761 if (pAllocateInfo
->allocationSize
== 0) {
2762 /* Apparently, this is allowed */
2763 *pMem
= VK_NULL_HANDLE
;
2767 const VkImportMemoryFdInfoKHR
*import_info
=
2768 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_FD_INFO_KHR
);
2769 const VkMemoryDedicatedAllocateInfoKHR
*dedicate_info
=
2770 vk_find_struct_const(pAllocateInfo
->pNext
, MEMORY_DEDICATED_ALLOCATE_INFO_KHR
);
2771 const VkExportMemoryAllocateInfoKHR
*export_info
=
2772 vk_find_struct_const(pAllocateInfo
->pNext
, EXPORT_MEMORY_ALLOCATE_INFO_KHR
);
2773 const VkImportMemoryHostPointerInfoEXT
*host_ptr_info
=
2774 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_HOST_POINTER_INFO_EXT
);
2776 const struct wsi_memory_allocate_info
*wsi_info
=
2777 vk_find_struct_const(pAllocateInfo
->pNext
, WSI_MEMORY_ALLOCATE_INFO_MESA
);
2779 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
2780 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2782 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
2784 if (wsi_info
&& wsi_info
->implicit_sync
)
2785 flags
|= RADEON_FLAG_IMPLICIT_SYNC
;
2787 if (dedicate_info
) {
2788 mem
->image
= radv_image_from_handle(dedicate_info
->image
);
2789 mem
->buffer
= radv_buffer_from_handle(dedicate_info
->buffer
);
2795 mem
->user_ptr
= NULL
;
2798 assert(import_info
->handleType
==
2799 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
||
2800 import_info
->handleType
==
2801 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
2802 mem
->bo
= device
->ws
->buffer_from_fd(device
->ws
, import_info
->fd
,
2805 result
= VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
;
2808 close(import_info
->fd
);
2810 } else if (host_ptr_info
) {
2811 assert(host_ptr_info
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
);
2812 assert(mem_type_index
== RADV_MEM_TYPE_GTT_CACHED
);
2813 mem
->bo
= device
->ws
->buffer_from_ptr(device
->ws
, host_ptr_info
->pHostPointer
,
2814 pAllocateInfo
->allocationSize
);
2816 result
= VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
;
2819 mem
->user_ptr
= host_ptr_info
->pHostPointer
;
2822 uint64_t alloc_size
= align_u64(pAllocateInfo
->allocationSize
, 4096);
2823 if (mem_type_index
== RADV_MEM_TYPE_GTT_WRITE_COMBINE
||
2824 mem_type_index
== RADV_MEM_TYPE_GTT_CACHED
)
2825 domain
= RADEON_DOMAIN_GTT
;
2827 domain
= RADEON_DOMAIN_VRAM
;
2829 if (mem_type_index
== RADV_MEM_TYPE_VRAM
)
2830 flags
|= RADEON_FLAG_NO_CPU_ACCESS
;
2832 flags
|= RADEON_FLAG_CPU_ACCESS
;
2834 if (mem_type_index
== RADV_MEM_TYPE_GTT_WRITE_COMBINE
)
2835 flags
|= RADEON_FLAG_GTT_WC
;
2837 if (!dedicate_info
&& !import_info
&& (!export_info
|| !export_info
->handleTypes
))
2838 flags
|= RADEON_FLAG_NO_INTERPROCESS_SHARING
;
2840 mem
->bo
= device
->ws
->buffer_create(device
->ws
, alloc_size
, device
->physical_device
->rad_info
.max_alignment
,
2844 result
= VK_ERROR_OUT_OF_DEVICE_MEMORY
;
2847 mem
->type_index
= mem_type_index
;
2850 result
= radv_bo_list_add(device
, mem
->bo
);
2851 if (result
!= VK_SUCCESS
)
2854 *pMem
= radv_device_memory_to_handle(mem
);
2859 device
->ws
->buffer_destroy(mem
->bo
);
2861 vk_free2(&device
->alloc
, pAllocator
, mem
);
2866 VkResult
radv_AllocateMemory(
2868 const VkMemoryAllocateInfo
* pAllocateInfo
,
2869 const VkAllocationCallbacks
* pAllocator
,
2870 VkDeviceMemory
* pMem
)
2872 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2873 return radv_alloc_memory(device
, pAllocateInfo
, pAllocator
, pMem
);
2876 void radv_FreeMemory(
2878 VkDeviceMemory _mem
,
2879 const VkAllocationCallbacks
* pAllocator
)
2881 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2882 RADV_FROM_HANDLE(radv_device_memory
, mem
, _mem
);
2887 radv_bo_list_remove(device
, mem
->bo
);
2888 device
->ws
->buffer_destroy(mem
->bo
);
2891 vk_free2(&device
->alloc
, pAllocator
, mem
);
2894 VkResult
radv_MapMemory(
2896 VkDeviceMemory _memory
,
2897 VkDeviceSize offset
,
2899 VkMemoryMapFlags flags
,
2902 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2903 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
2911 *ppData
= mem
->user_ptr
;
2913 *ppData
= device
->ws
->buffer_map(mem
->bo
);
2920 return vk_error(device
->instance
, VK_ERROR_MEMORY_MAP_FAILED
);
2923 void radv_UnmapMemory(
2925 VkDeviceMemory _memory
)
2927 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2928 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
2933 if (mem
->user_ptr
== NULL
)
2934 device
->ws
->buffer_unmap(mem
->bo
);
2937 VkResult
radv_FlushMappedMemoryRanges(
2939 uint32_t memoryRangeCount
,
2940 const VkMappedMemoryRange
* pMemoryRanges
)
2945 VkResult
radv_InvalidateMappedMemoryRanges(
2947 uint32_t memoryRangeCount
,
2948 const VkMappedMemoryRange
* pMemoryRanges
)
2953 void radv_GetBufferMemoryRequirements(
2956 VkMemoryRequirements
* pMemoryRequirements
)
2958 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2959 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
2961 pMemoryRequirements
->memoryTypeBits
= (1u << device
->physical_device
->memory_properties
.memoryTypeCount
) - 1;
2963 if (buffer
->flags
& VK_BUFFER_CREATE_SPARSE_BINDING_BIT
)
2964 pMemoryRequirements
->alignment
= 4096;
2966 pMemoryRequirements
->alignment
= 16;
2968 pMemoryRequirements
->size
= align64(buffer
->size
, pMemoryRequirements
->alignment
);
2971 void radv_GetBufferMemoryRequirements2(
2973 const VkBufferMemoryRequirementsInfo2KHR
* pInfo
,
2974 VkMemoryRequirements2KHR
* pMemoryRequirements
)
2976 radv_GetBufferMemoryRequirements(device
, pInfo
->buffer
,
2977 &pMemoryRequirements
->memoryRequirements
);
2978 RADV_FROM_HANDLE(radv_buffer
, buffer
, pInfo
->buffer
);
2979 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
2980 switch (ext
->sType
) {
2981 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS_KHR
: {
2982 VkMemoryDedicatedRequirementsKHR
*req
=
2983 (VkMemoryDedicatedRequirementsKHR
*) ext
;
2984 req
->requiresDedicatedAllocation
= buffer
->shareable
;
2985 req
->prefersDedicatedAllocation
= req
->requiresDedicatedAllocation
;
2994 void radv_GetImageMemoryRequirements(
2997 VkMemoryRequirements
* pMemoryRequirements
)
2999 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3000 RADV_FROM_HANDLE(radv_image
, image
, _image
);
3002 pMemoryRequirements
->memoryTypeBits
= (1u << device
->physical_device
->memory_properties
.memoryTypeCount
) - 1;
3004 pMemoryRequirements
->size
= image
->size
;
3005 pMemoryRequirements
->alignment
= image
->alignment
;
3008 void radv_GetImageMemoryRequirements2(
3010 const VkImageMemoryRequirementsInfo2KHR
* pInfo
,
3011 VkMemoryRequirements2KHR
* pMemoryRequirements
)
3013 radv_GetImageMemoryRequirements(device
, pInfo
->image
,
3014 &pMemoryRequirements
->memoryRequirements
);
3016 RADV_FROM_HANDLE(radv_image
, image
, pInfo
->image
);
3018 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3019 switch (ext
->sType
) {
3020 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS_KHR
: {
3021 VkMemoryDedicatedRequirementsKHR
*req
=
3022 (VkMemoryDedicatedRequirementsKHR
*) ext
;
3023 req
->requiresDedicatedAllocation
= image
->shareable
;
3024 req
->prefersDedicatedAllocation
= req
->requiresDedicatedAllocation
;
3033 void radv_GetImageSparseMemoryRequirements(
3036 uint32_t* pSparseMemoryRequirementCount
,
3037 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
3042 void radv_GetImageSparseMemoryRequirements2(
3044 const VkImageSparseMemoryRequirementsInfo2KHR
* pInfo
,
3045 uint32_t* pSparseMemoryRequirementCount
,
3046 VkSparseImageMemoryRequirements2KHR
* pSparseMemoryRequirements
)
3051 void radv_GetDeviceMemoryCommitment(
3053 VkDeviceMemory memory
,
3054 VkDeviceSize
* pCommittedMemoryInBytes
)
3056 *pCommittedMemoryInBytes
= 0;
3059 VkResult
radv_BindBufferMemory2(VkDevice device
,
3060 uint32_t bindInfoCount
,
3061 const VkBindBufferMemoryInfoKHR
*pBindInfos
)
3063 for (uint32_t i
= 0; i
< bindInfoCount
; ++i
) {
3064 RADV_FROM_HANDLE(radv_device_memory
, mem
, pBindInfos
[i
].memory
);
3065 RADV_FROM_HANDLE(radv_buffer
, buffer
, pBindInfos
[i
].buffer
);
3068 buffer
->bo
= mem
->bo
;
3069 buffer
->offset
= pBindInfos
[i
].memoryOffset
;
3077 VkResult
radv_BindBufferMemory(
3080 VkDeviceMemory memory
,
3081 VkDeviceSize memoryOffset
)
3083 const VkBindBufferMemoryInfoKHR info
= {
3084 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO_KHR
,
3087 .memoryOffset
= memoryOffset
3090 return radv_BindBufferMemory2(device
, 1, &info
);
3093 VkResult
radv_BindImageMemory2(VkDevice device
,
3094 uint32_t bindInfoCount
,
3095 const VkBindImageMemoryInfoKHR
*pBindInfos
)
3097 for (uint32_t i
= 0; i
< bindInfoCount
; ++i
) {
3098 RADV_FROM_HANDLE(radv_device_memory
, mem
, pBindInfos
[i
].memory
);
3099 RADV_FROM_HANDLE(radv_image
, image
, pBindInfos
[i
].image
);
3102 image
->bo
= mem
->bo
;
3103 image
->offset
= pBindInfos
[i
].memoryOffset
;
3113 VkResult
radv_BindImageMemory(
3116 VkDeviceMemory memory
,
3117 VkDeviceSize memoryOffset
)
3119 const VkBindImageMemoryInfoKHR info
= {
3120 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO_KHR
,
3123 .memoryOffset
= memoryOffset
3126 return radv_BindImageMemory2(device
, 1, &info
);
3131 radv_sparse_buffer_bind_memory(struct radv_device
*device
,
3132 const VkSparseBufferMemoryBindInfo
*bind
)
3134 RADV_FROM_HANDLE(radv_buffer
, buffer
, bind
->buffer
);
3136 for (uint32_t i
= 0; i
< bind
->bindCount
; ++i
) {
3137 struct radv_device_memory
*mem
= NULL
;
3139 if (bind
->pBinds
[i
].memory
!= VK_NULL_HANDLE
)
3140 mem
= radv_device_memory_from_handle(bind
->pBinds
[i
].memory
);
3142 device
->ws
->buffer_virtual_bind(buffer
->bo
,
3143 bind
->pBinds
[i
].resourceOffset
,
3144 bind
->pBinds
[i
].size
,
3145 mem
? mem
->bo
: NULL
,
3146 bind
->pBinds
[i
].memoryOffset
);
3151 radv_sparse_image_opaque_bind_memory(struct radv_device
*device
,
3152 const VkSparseImageOpaqueMemoryBindInfo
*bind
)
3154 RADV_FROM_HANDLE(radv_image
, image
, bind
->image
);
3156 for (uint32_t i
= 0; i
< bind
->bindCount
; ++i
) {
3157 struct radv_device_memory
*mem
= NULL
;
3159 if (bind
->pBinds
[i
].memory
!= VK_NULL_HANDLE
)
3160 mem
= radv_device_memory_from_handle(bind
->pBinds
[i
].memory
);
3162 device
->ws
->buffer_virtual_bind(image
->bo
,
3163 bind
->pBinds
[i
].resourceOffset
,
3164 bind
->pBinds
[i
].size
,
3165 mem
? mem
->bo
: NULL
,
3166 bind
->pBinds
[i
].memoryOffset
);
3170 VkResult
radv_QueueBindSparse(
3172 uint32_t bindInfoCount
,
3173 const VkBindSparseInfo
* pBindInfo
,
3176 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
3177 RADV_FROM_HANDLE(radv_queue
, queue
, _queue
);
3178 struct radeon_winsys_fence
*base_fence
= fence
? fence
->fence
: NULL
;
3179 bool fence_emitted
= false;
3183 for (uint32_t i
= 0; i
< bindInfoCount
; ++i
) {
3184 struct radv_winsys_sem_info sem_info
;
3185 for (uint32_t j
= 0; j
< pBindInfo
[i
].bufferBindCount
; ++j
) {
3186 radv_sparse_buffer_bind_memory(queue
->device
,
3187 pBindInfo
[i
].pBufferBinds
+ j
);
3190 for (uint32_t j
= 0; j
< pBindInfo
[i
].imageOpaqueBindCount
; ++j
) {
3191 radv_sparse_image_opaque_bind_memory(queue
->device
,
3192 pBindInfo
[i
].pImageOpaqueBinds
+ j
);
3196 result
= radv_alloc_sem_info(queue
->device
->instance
,
3198 pBindInfo
[i
].waitSemaphoreCount
,
3199 pBindInfo
[i
].pWaitSemaphores
,
3200 pBindInfo
[i
].signalSemaphoreCount
,
3201 pBindInfo
[i
].pSignalSemaphores
,
3203 if (result
!= VK_SUCCESS
)
3206 if (pBindInfo
[i
].waitSemaphoreCount
|| pBindInfo
[i
].signalSemaphoreCount
) {
3207 ret
= queue
->device
->ws
->cs_submit(queue
->hw_ctx
, queue
->queue_idx
,
3208 &queue
->device
->empty_cs
[queue
->queue_family_index
],
3213 radv_loge("failed to submit CS %d\n", i
);
3217 fence_emitted
= true;
3219 fence
->submitted
= true;
3222 radv_free_sem_info(&sem_info
);
3227 if (!fence_emitted
) {
3228 result
= radv_signal_fence(queue
, fence
);
3229 if (result
!= VK_SUCCESS
)
3232 fence
->submitted
= true;
3238 VkResult
radv_CreateFence(
3240 const VkFenceCreateInfo
* pCreateInfo
,
3241 const VkAllocationCallbacks
* pAllocator
,
3244 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3245 const VkExportFenceCreateInfoKHR
*export
=
3246 vk_find_struct_const(pCreateInfo
->pNext
, EXPORT_FENCE_CREATE_INFO_KHR
);
3247 VkExternalFenceHandleTypeFlagsKHR handleTypes
=
3248 export
? export
->handleTypes
: 0;
3250 struct radv_fence
*fence
= vk_alloc2(&device
->alloc
, pAllocator
,
3252 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3255 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3257 fence
->fence_wsi
= NULL
;
3258 fence
->submitted
= false;
3259 fence
->signalled
= !!(pCreateInfo
->flags
& VK_FENCE_CREATE_SIGNALED_BIT
);
3260 fence
->temp_syncobj
= 0;
3261 if (device
->always_use_syncobj
|| handleTypes
) {
3262 int ret
= device
->ws
->create_syncobj(device
->ws
, &fence
->syncobj
);
3264 vk_free2(&device
->alloc
, pAllocator
, fence
);
3265 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3267 if (pCreateInfo
->flags
& VK_FENCE_CREATE_SIGNALED_BIT
) {
3268 device
->ws
->signal_syncobj(device
->ws
, fence
->syncobj
);
3270 fence
->fence
= NULL
;
3272 fence
->fence
= device
->ws
->create_fence();
3273 if (!fence
->fence
) {
3274 vk_free2(&device
->alloc
, pAllocator
, fence
);
3275 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3280 *pFence
= radv_fence_to_handle(fence
);
3285 void radv_DestroyFence(
3288 const VkAllocationCallbacks
* pAllocator
)
3290 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3291 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
3296 if (fence
->temp_syncobj
)
3297 device
->ws
->destroy_syncobj(device
->ws
, fence
->temp_syncobj
);
3299 device
->ws
->destroy_syncobj(device
->ws
, fence
->syncobj
);
3301 device
->ws
->destroy_fence(fence
->fence
);
3302 if (fence
->fence_wsi
)
3303 fence
->fence_wsi
->destroy(fence
->fence_wsi
);
3304 vk_free2(&device
->alloc
, pAllocator
, fence
);
3308 static uint64_t radv_get_current_time()
3311 clock_gettime(CLOCK_MONOTONIC
, &tv
);
3312 return tv
.tv_nsec
+ tv
.tv_sec
*1000000000ull;
3315 static uint64_t radv_get_absolute_timeout(uint64_t timeout
)
3317 uint64_t current_time
= radv_get_current_time();
3319 timeout
= MIN2(UINT64_MAX
- current_time
, timeout
);
3321 return current_time
+ timeout
;
3325 static bool radv_all_fences_plain_and_submitted(uint32_t fenceCount
, const VkFence
*pFences
)
3327 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
3328 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
3329 if (fence
->fence
== NULL
|| fence
->syncobj
||
3330 fence
->temp_syncobj
||
3331 (!fence
->signalled
&& !fence
->submitted
))
3337 static bool radv_all_fences_syncobj(uint32_t fenceCount
, const VkFence
*pFences
)
3339 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
3340 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
3341 if (fence
->syncobj
== 0 && fence
->temp_syncobj
== 0)
3347 VkResult
radv_WaitForFences(
3349 uint32_t fenceCount
,
3350 const VkFence
* pFences
,
3354 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3355 timeout
= radv_get_absolute_timeout(timeout
);
3357 if (device
->always_use_syncobj
&&
3358 radv_all_fences_syncobj(fenceCount
, pFences
))
3360 uint32_t *handles
= malloc(sizeof(uint32_t) * fenceCount
);
3362 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3364 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
3365 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
3366 handles
[i
] = fence
->temp_syncobj
? fence
->temp_syncobj
: fence
->syncobj
;
3369 bool success
= device
->ws
->wait_syncobj(device
->ws
, handles
, fenceCount
, waitAll
, timeout
);
3372 return success
? VK_SUCCESS
: VK_TIMEOUT
;
3375 if (!waitAll
&& fenceCount
> 1) {
3376 /* Not doing this by default for waitAll, due to needing to allocate twice. */
3377 if (device
->physical_device
->rad_info
.drm_minor
>= 10 && radv_all_fences_plain_and_submitted(fenceCount
, pFences
)) {
3378 uint32_t wait_count
= 0;
3379 struct radeon_winsys_fence
**fences
= malloc(sizeof(struct radeon_winsys_fence
*) * fenceCount
);
3381 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3383 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
3384 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
3386 if (fence
->signalled
) {
3391 fences
[wait_count
++] = fence
->fence
;
3394 bool success
= device
->ws
->fences_wait(device
->ws
, fences
, wait_count
,
3395 waitAll
, timeout
- radv_get_current_time());
3398 return success
? VK_SUCCESS
: VK_TIMEOUT
;
3401 while(radv_get_current_time() <= timeout
) {
3402 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
3403 if (radv_GetFenceStatus(_device
, pFences
[i
]) == VK_SUCCESS
)
3410 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
3411 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
3412 bool expired
= false;
3414 if (fence
->temp_syncobj
) {
3415 if (!device
->ws
->wait_syncobj(device
->ws
, &fence
->temp_syncobj
, 1, true, timeout
))
3420 if (fence
->syncobj
) {
3421 if (!device
->ws
->wait_syncobj(device
->ws
, &fence
->syncobj
, 1, true, timeout
))
3426 if (fence
->signalled
)
3430 if (!fence
->submitted
) {
3431 while(radv_get_current_time() <= timeout
&&
3435 if (!fence
->submitted
)
3438 /* Recheck as it may have been set by
3439 * submitting operations. */
3441 if (fence
->signalled
)
3445 expired
= device
->ws
->fence_wait(device
->ws
,
3452 if (fence
->fence_wsi
) {
3453 VkResult result
= fence
->fence_wsi
->wait(fence
->fence_wsi
, timeout
);
3454 if (result
!= VK_SUCCESS
)
3458 fence
->signalled
= true;
3464 VkResult
radv_ResetFences(VkDevice _device
,
3465 uint32_t fenceCount
,
3466 const VkFence
*pFences
)
3468 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3470 for (unsigned i
= 0; i
< fenceCount
; ++i
) {
3471 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
3472 fence
->submitted
= fence
->signalled
= false;
3474 /* Per spec, we first restore the permanent payload, and then reset, so
3475 * having a temp syncobj should not skip resetting the permanent syncobj. */
3476 if (fence
->temp_syncobj
) {
3477 device
->ws
->destroy_syncobj(device
->ws
, fence
->temp_syncobj
);
3478 fence
->temp_syncobj
= 0;
3481 if (fence
->syncobj
) {
3482 device
->ws
->reset_syncobj(device
->ws
, fence
->syncobj
);
3489 VkResult
radv_GetFenceStatus(VkDevice _device
, VkFence _fence
)
3491 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3492 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
3494 if (fence
->temp_syncobj
) {
3495 bool success
= device
->ws
->wait_syncobj(device
->ws
, &fence
->temp_syncobj
, 1, true, 0);
3496 return success
? VK_SUCCESS
: VK_NOT_READY
;
3499 if (fence
->syncobj
) {
3500 bool success
= device
->ws
->wait_syncobj(device
->ws
, &fence
->syncobj
, 1, true, 0);
3501 return success
? VK_SUCCESS
: VK_NOT_READY
;
3504 if (fence
->signalled
)
3506 if (!fence
->submitted
)
3507 return VK_NOT_READY
;
3509 if (!device
->ws
->fence_wait(device
->ws
, fence
->fence
, false, 0))
3510 return VK_NOT_READY
;
3512 if (fence
->fence_wsi
) {
3513 VkResult result
= fence
->fence_wsi
->wait(fence
->fence_wsi
, 0);
3515 if (result
!= VK_SUCCESS
) {
3516 if (result
== VK_TIMEOUT
)
3517 return VK_NOT_READY
;
3525 // Queue semaphore functions
3527 VkResult
radv_CreateSemaphore(
3529 const VkSemaphoreCreateInfo
* pCreateInfo
,
3530 const VkAllocationCallbacks
* pAllocator
,
3531 VkSemaphore
* pSemaphore
)
3533 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3534 const VkExportSemaphoreCreateInfoKHR
*export
=
3535 vk_find_struct_const(pCreateInfo
->pNext
, EXPORT_SEMAPHORE_CREATE_INFO_KHR
);
3536 VkExternalSemaphoreHandleTypeFlagsKHR handleTypes
=
3537 export
? export
->handleTypes
: 0;
3539 struct radv_semaphore
*sem
= vk_alloc2(&device
->alloc
, pAllocator
,
3541 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3543 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3545 sem
->temp_syncobj
= 0;
3546 /* create a syncobject if we are going to export this semaphore */
3547 if (device
->always_use_syncobj
|| handleTypes
) {
3548 assert (device
->physical_device
->rad_info
.has_syncobj
);
3549 int ret
= device
->ws
->create_syncobj(device
->ws
, &sem
->syncobj
);
3551 vk_free2(&device
->alloc
, pAllocator
, sem
);
3552 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3556 sem
->sem
= device
->ws
->create_sem(device
->ws
);
3558 vk_free2(&device
->alloc
, pAllocator
, sem
);
3559 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3564 *pSemaphore
= radv_semaphore_to_handle(sem
);
3568 void radv_DestroySemaphore(
3570 VkSemaphore _semaphore
,
3571 const VkAllocationCallbacks
* pAllocator
)
3573 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3574 RADV_FROM_HANDLE(radv_semaphore
, sem
, _semaphore
);
3579 device
->ws
->destroy_syncobj(device
->ws
, sem
->syncobj
);
3581 device
->ws
->destroy_sem(sem
->sem
);
3582 vk_free2(&device
->alloc
, pAllocator
, sem
);
3585 VkResult
radv_CreateEvent(
3587 const VkEventCreateInfo
* pCreateInfo
,
3588 const VkAllocationCallbacks
* pAllocator
,
3591 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3592 struct radv_event
*event
= vk_alloc2(&device
->alloc
, pAllocator
,
3594 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3597 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3599 event
->bo
= device
->ws
->buffer_create(device
->ws
, 8, 8,
3601 RADEON_FLAG_VA_UNCACHED
| RADEON_FLAG_CPU_ACCESS
| RADEON_FLAG_NO_INTERPROCESS_SHARING
);
3603 vk_free2(&device
->alloc
, pAllocator
, event
);
3604 return vk_error(device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
3607 event
->map
= (uint64_t*)device
->ws
->buffer_map(event
->bo
);
3609 *pEvent
= radv_event_to_handle(event
);
3614 void radv_DestroyEvent(
3617 const VkAllocationCallbacks
* pAllocator
)
3619 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3620 RADV_FROM_HANDLE(radv_event
, event
, _event
);
3624 device
->ws
->buffer_destroy(event
->bo
);
3625 vk_free2(&device
->alloc
, pAllocator
, event
);
3628 VkResult
radv_GetEventStatus(
3632 RADV_FROM_HANDLE(radv_event
, event
, _event
);
3634 if (*event
->map
== 1)
3635 return VK_EVENT_SET
;
3636 return VK_EVENT_RESET
;
3639 VkResult
radv_SetEvent(
3643 RADV_FROM_HANDLE(radv_event
, event
, _event
);
3649 VkResult
radv_ResetEvent(
3653 RADV_FROM_HANDLE(radv_event
, event
, _event
);
3659 VkResult
radv_CreateBuffer(
3661 const VkBufferCreateInfo
* pCreateInfo
,
3662 const VkAllocationCallbacks
* pAllocator
,
3665 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3666 struct radv_buffer
*buffer
;
3668 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
3670 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
3671 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3673 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3675 buffer
->size
= pCreateInfo
->size
;
3676 buffer
->usage
= pCreateInfo
->usage
;
3679 buffer
->flags
= pCreateInfo
->flags
;
3681 buffer
->shareable
= vk_find_struct_const(pCreateInfo
->pNext
,
3682 EXTERNAL_MEMORY_BUFFER_CREATE_INFO_KHR
) != NULL
;
3684 if (pCreateInfo
->flags
& VK_BUFFER_CREATE_SPARSE_BINDING_BIT
) {
3685 buffer
->bo
= device
->ws
->buffer_create(device
->ws
,
3686 align64(buffer
->size
, 4096),
3687 4096, 0, RADEON_FLAG_VIRTUAL
);
3689 vk_free2(&device
->alloc
, pAllocator
, buffer
);
3690 return vk_error(device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
3694 *pBuffer
= radv_buffer_to_handle(buffer
);
3699 void radv_DestroyBuffer(
3702 const VkAllocationCallbacks
* pAllocator
)
3704 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3705 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
3710 if (buffer
->flags
& VK_BUFFER_CREATE_SPARSE_BINDING_BIT
)
3711 device
->ws
->buffer_destroy(buffer
->bo
);
3713 vk_free2(&device
->alloc
, pAllocator
, buffer
);
3716 static inline unsigned
3717 si_tile_mode_index(const struct radv_image
*image
, unsigned level
, bool stencil
)
3720 return image
->surface
.u
.legacy
.stencil_tiling_index
[level
];
3722 return image
->surface
.u
.legacy
.tiling_index
[level
];
3725 static uint32_t radv_surface_max_layer_count(struct radv_image_view
*iview
)
3727 return iview
->type
== VK_IMAGE_VIEW_TYPE_3D
? iview
->extent
.depth
: (iview
->base_layer
+ iview
->layer_count
);
3731 radv_init_dcc_control_reg(struct radv_device
*device
,
3732 struct radv_image_view
*iview
)
3734 unsigned max_uncompressed_block_size
= V_028C78_MAX_BLOCK_SIZE_256B
;
3735 unsigned min_compressed_block_size
= V_028C78_MIN_BLOCK_SIZE_32B
;
3736 unsigned max_compressed_block_size
;
3737 unsigned independent_64b_blocks
;
3739 if (device
->physical_device
->rad_info
.chip_class
< VI
)
3742 if (iview
->image
->info
.samples
> 1) {
3743 if (iview
->image
->surface
.bpe
== 1)
3744 max_uncompressed_block_size
= V_028C78_MAX_BLOCK_SIZE_64B
;
3745 else if (iview
->image
->surface
.bpe
== 2)
3746 max_uncompressed_block_size
= V_028C78_MAX_BLOCK_SIZE_128B
;
3749 if (!device
->physical_device
->rad_info
.has_dedicated_vram
) {
3750 /* amdvlk: [min-compressed-block-size] should be set to 32 for
3751 * dGPU and 64 for APU because all of our APUs to date use
3752 * DIMMs which have a request granularity size of 64B while all
3753 * other chips have a 32B request size.
3755 min_compressed_block_size
= V_028C78_MIN_BLOCK_SIZE_64B
;
3758 if (iview
->image
->usage
& (VK_IMAGE_USAGE_SAMPLED_BIT
|
3759 VK_IMAGE_USAGE_TRANSFER_SRC_BIT
|
3760 VK_IMAGE_USAGE_INPUT_ATTACHMENT_BIT
)) {
3761 /* If this DCC image is potentially going to be used in texture
3762 * fetches, we need some special settings.
3764 independent_64b_blocks
= 1;
3765 max_compressed_block_size
= V_028C78_MAX_BLOCK_SIZE_64B
;
3767 /* MAX_UNCOMPRESSED_BLOCK_SIZE must be >=
3768 * MAX_COMPRESSED_BLOCK_SIZE. Set MAX_COMPRESSED_BLOCK_SIZE as
3769 * big as possible for better compression state.
3771 independent_64b_blocks
= 0;
3772 max_compressed_block_size
= max_uncompressed_block_size
;
3775 return S_028C78_MAX_UNCOMPRESSED_BLOCK_SIZE(max_uncompressed_block_size
) |
3776 S_028C78_MAX_COMPRESSED_BLOCK_SIZE(max_compressed_block_size
) |
3777 S_028C78_MIN_COMPRESSED_BLOCK_SIZE(min_compressed_block_size
) |
3778 S_028C78_INDEPENDENT_64B_BLOCKS(independent_64b_blocks
);
3782 radv_initialise_color_surface(struct radv_device
*device
,
3783 struct radv_color_buffer_info
*cb
,
3784 struct radv_image_view
*iview
)
3786 const struct vk_format_description
*desc
;
3787 unsigned ntype
, format
, swap
, endian
;
3788 unsigned blend_clamp
= 0, blend_bypass
= 0;
3790 const struct radeon_surf
*surf
= &iview
->image
->surface
;
3792 desc
= vk_format_description(iview
->vk_format
);
3794 memset(cb
, 0, sizeof(*cb
));
3796 /* Intensity is implemented as Red, so treat it that way. */
3797 cb
->cb_color_attrib
= S_028C74_FORCE_DST_ALPHA_1(desc
->swizzle
[3] == VK_SWIZZLE_1
);
3799 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
;
3801 cb
->cb_color_base
= va
>> 8;
3803 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
3804 struct gfx9_surf_meta_flags meta
;
3805 if (iview
->image
->dcc_offset
)
3806 meta
= iview
->image
->surface
.u
.gfx9
.dcc
;
3808 meta
= iview
->image
->surface
.u
.gfx9
.cmask
;
3810 cb
->cb_color_attrib
|= S_028C74_COLOR_SW_MODE(iview
->image
->surface
.u
.gfx9
.surf
.swizzle_mode
) |
3811 S_028C74_FMASK_SW_MODE(iview
->image
->surface
.u
.gfx9
.fmask
.swizzle_mode
) |
3812 S_028C74_RB_ALIGNED(meta
.rb_aligned
) |
3813 S_028C74_PIPE_ALIGNED(meta
.pipe_aligned
);
3815 cb
->cb_color_base
+= iview
->image
->surface
.u
.gfx9
.surf_offset
>> 8;
3816 cb
->cb_color_base
|= iview
->image
->surface
.tile_swizzle
;
3818 const struct legacy_surf_level
*level_info
= &surf
->u
.legacy
.level
[iview
->base_mip
];
3819 unsigned pitch_tile_max
, slice_tile_max
, tile_mode_index
;
3821 cb
->cb_color_base
+= level_info
->offset
>> 8;
3822 if (level_info
->mode
== RADEON_SURF_MODE_2D
)
3823 cb
->cb_color_base
|= iview
->image
->surface
.tile_swizzle
;
3825 pitch_tile_max
= level_info
->nblk_x
/ 8 - 1;
3826 slice_tile_max
= (level_info
->nblk_x
* level_info
->nblk_y
) / 64 - 1;
3827 tile_mode_index
= si_tile_mode_index(iview
->image
, iview
->base_mip
, false);
3829 cb
->cb_color_pitch
= S_028C64_TILE_MAX(pitch_tile_max
);
3830 cb
->cb_color_slice
= S_028C68_TILE_MAX(slice_tile_max
);
3831 cb
->cb_color_cmask_slice
= iview
->image
->cmask
.slice_tile_max
;
3833 cb
->cb_color_attrib
|= S_028C74_TILE_MODE_INDEX(tile_mode_index
);
3835 if (radv_image_has_fmask(iview
->image
)) {
3836 if (device
->physical_device
->rad_info
.chip_class
>= CIK
)
3837 cb
->cb_color_pitch
|= S_028C64_FMASK_TILE_MAX(iview
->image
->fmask
.pitch_in_pixels
/ 8 - 1);
3838 cb
->cb_color_attrib
|= S_028C74_FMASK_TILE_MODE_INDEX(iview
->image
->fmask
.tile_mode_index
);
3839 cb
->cb_color_fmask_slice
= S_028C88_TILE_MAX(iview
->image
->fmask
.slice_tile_max
);
3841 /* This must be set for fast clear to work without FMASK. */
3842 if (device
->physical_device
->rad_info
.chip_class
>= CIK
)
3843 cb
->cb_color_pitch
|= S_028C64_FMASK_TILE_MAX(pitch_tile_max
);
3844 cb
->cb_color_attrib
|= S_028C74_FMASK_TILE_MODE_INDEX(tile_mode_index
);
3845 cb
->cb_color_fmask_slice
= S_028C88_TILE_MAX(slice_tile_max
);
3849 /* CMASK variables */
3850 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
;
3851 va
+= iview
->image
->cmask
.offset
;
3852 cb
->cb_color_cmask
= va
>> 8;
3854 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
;
3855 va
+= iview
->image
->dcc_offset
;
3856 cb
->cb_dcc_base
= va
>> 8;
3857 cb
->cb_dcc_base
|= iview
->image
->surface
.tile_swizzle
;
3859 uint32_t max_slice
= radv_surface_max_layer_count(iview
) - 1;
3860 cb
->cb_color_view
= S_028C6C_SLICE_START(iview
->base_layer
) |
3861 S_028C6C_SLICE_MAX(max_slice
);
3863 if (iview
->image
->info
.samples
> 1) {
3864 unsigned log_samples
= util_logbase2(iview
->image
->info
.samples
);
3866 cb
->cb_color_attrib
|= S_028C74_NUM_SAMPLES(log_samples
) |
3867 S_028C74_NUM_FRAGMENTS(log_samples
);
3870 if (radv_image_has_fmask(iview
->image
)) {
3871 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
+ iview
->image
->fmask
.offset
;
3872 cb
->cb_color_fmask
= va
>> 8;
3873 cb
->cb_color_fmask
|= iview
->image
->fmask
.tile_swizzle
;
3875 cb
->cb_color_fmask
= cb
->cb_color_base
;
3878 ntype
= radv_translate_color_numformat(iview
->vk_format
,
3880 vk_format_get_first_non_void_channel(iview
->vk_format
));
3881 format
= radv_translate_colorformat(iview
->vk_format
);
3882 if (format
== V_028C70_COLOR_INVALID
|| ntype
== ~0u)
3883 radv_finishme("Illegal color\n");
3884 swap
= radv_translate_colorswap(iview
->vk_format
, FALSE
);
3885 endian
= radv_colorformat_endian_swap(format
);
3887 /* blend clamp should be set for all NORM/SRGB types */
3888 if (ntype
== V_028C70_NUMBER_UNORM
||
3889 ntype
== V_028C70_NUMBER_SNORM
||
3890 ntype
== V_028C70_NUMBER_SRGB
)
3893 /* set blend bypass according to docs if SINT/UINT or
3894 8/24 COLOR variants */
3895 if (ntype
== V_028C70_NUMBER_UINT
|| ntype
== V_028C70_NUMBER_SINT
||
3896 format
== V_028C70_COLOR_8_24
|| format
== V_028C70_COLOR_24_8
||
3897 format
== V_028C70_COLOR_X24_8_32_FLOAT
) {
3902 if ((ntype
== V_028C70_NUMBER_UINT
|| ntype
== V_028C70_NUMBER_SINT
) &&
3903 (format
== V_028C70_COLOR_8
||
3904 format
== V_028C70_COLOR_8_8
||
3905 format
== V_028C70_COLOR_8_8_8_8
))
3906 ->color_is_int8
= true;
3908 cb
->cb_color_info
= S_028C70_FORMAT(format
) |
3909 S_028C70_COMP_SWAP(swap
) |
3910 S_028C70_BLEND_CLAMP(blend_clamp
) |
3911 S_028C70_BLEND_BYPASS(blend_bypass
) |
3912 S_028C70_SIMPLE_FLOAT(1) |
3913 S_028C70_ROUND_MODE(ntype
!= V_028C70_NUMBER_UNORM
&&
3914 ntype
!= V_028C70_NUMBER_SNORM
&&
3915 ntype
!= V_028C70_NUMBER_SRGB
&&
3916 format
!= V_028C70_COLOR_8_24
&&
3917 format
!= V_028C70_COLOR_24_8
) |
3918 S_028C70_NUMBER_TYPE(ntype
) |
3919 S_028C70_ENDIAN(endian
);
3920 if (radv_image_has_fmask(iview
->image
)) {
3921 cb
->cb_color_info
|= S_028C70_COMPRESSION(1);
3922 if (device
->physical_device
->rad_info
.chip_class
== SI
) {
3923 unsigned fmask_bankh
= util_logbase2(iview
->image
->fmask
.bank_height
);
3924 cb
->cb_color_attrib
|= S_028C74_FMASK_BANK_HEIGHT(fmask_bankh
);
3928 if (radv_image_has_cmask(iview
->image
) &&
3929 !(device
->instance
->debug_flags
& RADV_DEBUG_NO_FAST_CLEARS
))
3930 cb
->cb_color_info
|= S_028C70_FAST_CLEAR(1);
3932 if (radv_dcc_enabled(iview
->image
, iview
->base_mip
))
3933 cb
->cb_color_info
|= S_028C70_DCC_ENABLE(1);
3935 cb
->cb_dcc_control
= radv_init_dcc_control_reg(device
, iview
);
3937 /* This must be set for fast clear to work without FMASK. */
3938 if (!radv_image_has_fmask(iview
->image
) &&
3939 device
->physical_device
->rad_info
.chip_class
== SI
) {
3940 unsigned bankh
= util_logbase2(iview
->image
->surface
.u
.legacy
.bankh
);
3941 cb
->cb_color_attrib
|= S_028C74_FMASK_BANK_HEIGHT(bankh
);
3944 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
3945 unsigned mip0_depth
= iview
->image
->type
== VK_IMAGE_TYPE_3D
?
3946 (iview
->extent
.depth
- 1) : (iview
->image
->info
.array_size
- 1);
3948 cb
->cb_color_view
|= S_028C6C_MIP_LEVEL(iview
->base_mip
);
3949 cb
->cb_color_attrib
|= S_028C74_MIP0_DEPTH(mip0_depth
) |
3950 S_028C74_RESOURCE_TYPE(iview
->image
->surface
.u
.gfx9
.resource_type
);
3951 cb
->cb_color_attrib2
= S_028C68_MIP0_WIDTH(iview
->extent
.width
- 1) |
3952 S_028C68_MIP0_HEIGHT(iview
->extent
.height
- 1) |
3953 S_028C68_MAX_MIP(iview
->image
->info
.levels
- 1);
3958 radv_calc_decompress_on_z_planes(struct radv_device
*device
,
3959 struct radv_image_view
*iview
)
3961 unsigned max_zplanes
= 0;
3963 assert(radv_image_is_tc_compat_htile(iview
->image
));
3965 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
3966 /* Default value for 32-bit depth surfaces. */
3969 if (iview
->vk_format
== VK_FORMAT_D16_UNORM
&&
3970 iview
->image
->info
.samples
> 1)
3973 max_zplanes
= max_zplanes
+ 1;
3975 if (iview
->vk_format
== VK_FORMAT_D16_UNORM
) {
3976 /* Do not enable Z plane compression for 16-bit depth
3977 * surfaces because isn't supported on GFX8. Only
3978 * 32-bit depth surfaces are supported by the hardware.
3979 * This allows to maintain shader compatibility and to
3980 * reduce the number of depth decompressions.
3984 if (iview
->image
->info
.samples
<= 1)
3986 else if (iview
->image
->info
.samples
<= 4)
3997 radv_initialise_ds_surface(struct radv_device
*device
,
3998 struct radv_ds_buffer_info
*ds
,
3999 struct radv_image_view
*iview
)
4001 unsigned level
= iview
->base_mip
;
4002 unsigned format
, stencil_format
;
4003 uint64_t va
, s_offs
, z_offs
;
4004 bool stencil_only
= false;
4005 memset(ds
, 0, sizeof(*ds
));
4006 switch (iview
->image
->vk_format
) {
4007 case VK_FORMAT_D24_UNORM_S8_UINT
:
4008 case VK_FORMAT_X8_D24_UNORM_PACK32
:
4009 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-24);
4010 ds
->offset_scale
= 2.0f
;
4012 case VK_FORMAT_D16_UNORM
:
4013 case VK_FORMAT_D16_UNORM_S8_UINT
:
4014 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-16);
4015 ds
->offset_scale
= 4.0f
;
4017 case VK_FORMAT_D32_SFLOAT
:
4018 case VK_FORMAT_D32_SFLOAT_S8_UINT
:
4019 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-23) |
4020 S_028B78_POLY_OFFSET_DB_IS_FLOAT_FMT(1);
4021 ds
->offset_scale
= 1.0f
;
4023 case VK_FORMAT_S8_UINT
:
4024 stencil_only
= true;
4030 format
= radv_translate_dbformat(iview
->image
->vk_format
);
4031 stencil_format
= iview
->image
->surface
.has_stencil
?
4032 V_028044_STENCIL_8
: V_028044_STENCIL_INVALID
;
4034 uint32_t max_slice
= radv_surface_max_layer_count(iview
) - 1;
4035 ds
->db_depth_view
= S_028008_SLICE_START(iview
->base_layer
) |
4036 S_028008_SLICE_MAX(max_slice
);
4038 ds
->db_htile_data_base
= 0;
4039 ds
->db_htile_surface
= 0;
4041 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
;
4042 s_offs
= z_offs
= va
;
4044 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
4045 assert(iview
->image
->surface
.u
.gfx9
.surf_offset
== 0);
4046 s_offs
+= iview
->image
->surface
.u
.gfx9
.stencil_offset
;
4048 ds
->db_z_info
= S_028038_FORMAT(format
) |
4049 S_028038_NUM_SAMPLES(util_logbase2(iview
->image
->info
.samples
)) |
4050 S_028038_SW_MODE(iview
->image
->surface
.u
.gfx9
.surf
.swizzle_mode
) |
4051 S_028038_MAXMIP(iview
->image
->info
.levels
- 1) |
4052 S_028038_ZRANGE_PRECISION(1);
4053 ds
->db_stencil_info
= S_02803C_FORMAT(stencil_format
) |
4054 S_02803C_SW_MODE(iview
->image
->surface
.u
.gfx9
.stencil
.swizzle_mode
);
4056 ds
->db_z_info2
= S_028068_EPITCH(iview
->image
->surface
.u
.gfx9
.surf
.epitch
);
4057 ds
->db_stencil_info2
= S_02806C_EPITCH(iview
->image
->surface
.u
.gfx9
.stencil
.epitch
);
4058 ds
->db_depth_view
|= S_028008_MIPID(level
);
4060 ds
->db_depth_size
= S_02801C_X_MAX(iview
->image
->info
.width
- 1) |
4061 S_02801C_Y_MAX(iview
->image
->info
.height
- 1);
4063 if (radv_htile_enabled(iview
->image
, level
)) {
4064 ds
->db_z_info
|= S_028038_TILE_SURFACE_ENABLE(1);
4066 if (radv_image_is_tc_compat_htile(iview
->image
)) {
4067 unsigned max_zplanes
=
4068 radv_calc_decompress_on_z_planes(device
, iview
);
4070 ds
->db_z_info
|= S_028038_DECOMPRESS_ON_N_ZPLANES(max_zplanes
) |
4071 S_028038_ITERATE_FLUSH(1);
4072 ds
->db_stencil_info
|= S_02803C_ITERATE_FLUSH(1);
4075 if (!iview
->image
->surface
.has_stencil
)
4076 /* Use all of the htile_buffer for depth if there's no stencil. */
4077 ds
->db_stencil_info
|= S_02803C_TILE_STENCIL_DISABLE(1);
4078 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
+
4079 iview
->image
->htile_offset
;
4080 ds
->db_htile_data_base
= va
>> 8;
4081 ds
->db_htile_surface
= S_028ABC_FULL_CACHE(1) |
4082 S_028ABC_PIPE_ALIGNED(iview
->image
->surface
.u
.gfx9
.htile
.pipe_aligned
) |
4083 S_028ABC_RB_ALIGNED(iview
->image
->surface
.u
.gfx9
.htile
.rb_aligned
);
4086 const struct legacy_surf_level
*level_info
= &iview
->image
->surface
.u
.legacy
.level
[level
];
4089 level_info
= &iview
->image
->surface
.u
.legacy
.stencil_level
[level
];
4091 z_offs
+= iview
->image
->surface
.u
.legacy
.level
[level
].offset
;
4092 s_offs
+= iview
->image
->surface
.u
.legacy
.stencil_level
[level
].offset
;
4094 ds
->db_depth_info
= S_02803C_ADDR5_SWIZZLE_MASK(!radv_image_is_tc_compat_htile(iview
->image
));
4095 ds
->db_z_info
= S_028040_FORMAT(format
) | S_028040_ZRANGE_PRECISION(1);
4096 ds
->db_stencil_info
= S_028044_FORMAT(stencil_format
);
4098 if (iview
->image
->info
.samples
> 1)
4099 ds
->db_z_info
|= S_028040_NUM_SAMPLES(util_logbase2(iview
->image
->info
.samples
));
4101 if (device
->physical_device
->rad_info
.chip_class
>= CIK
) {
4102 struct radeon_info
*info
= &device
->physical_device
->rad_info
;
4103 unsigned tiling_index
= iview
->image
->surface
.u
.legacy
.tiling_index
[level
];
4104 unsigned stencil_index
= iview
->image
->surface
.u
.legacy
.stencil_tiling_index
[level
];
4105 unsigned macro_index
= iview
->image
->surface
.u
.legacy
.macro_tile_index
;
4106 unsigned tile_mode
= info
->si_tile_mode_array
[tiling_index
];
4107 unsigned stencil_tile_mode
= info
->si_tile_mode_array
[stencil_index
];
4108 unsigned macro_mode
= info
->cik_macrotile_mode_array
[macro_index
];
4111 tile_mode
= stencil_tile_mode
;
4113 ds
->db_depth_info
|=
4114 S_02803C_ARRAY_MODE(G_009910_ARRAY_MODE(tile_mode
)) |
4115 S_02803C_PIPE_CONFIG(G_009910_PIPE_CONFIG(tile_mode
)) |
4116 S_02803C_BANK_WIDTH(G_009990_BANK_WIDTH(macro_mode
)) |
4117 S_02803C_BANK_HEIGHT(G_009990_BANK_HEIGHT(macro_mode
)) |
4118 S_02803C_MACRO_TILE_ASPECT(G_009990_MACRO_TILE_ASPECT(macro_mode
)) |
4119 S_02803C_NUM_BANKS(G_009990_NUM_BANKS(macro_mode
));
4120 ds
->db_z_info
|= S_028040_TILE_SPLIT(G_009910_TILE_SPLIT(tile_mode
));
4121 ds
->db_stencil_info
|= S_028044_TILE_SPLIT(G_009910_TILE_SPLIT(stencil_tile_mode
));
4123 unsigned tile_mode_index
= si_tile_mode_index(iview
->image
, level
, false);
4124 ds
->db_z_info
|= S_028040_TILE_MODE_INDEX(tile_mode_index
);
4125 tile_mode_index
= si_tile_mode_index(iview
->image
, level
, true);
4126 ds
->db_stencil_info
|= S_028044_TILE_MODE_INDEX(tile_mode_index
);
4128 ds
->db_z_info
|= S_028040_TILE_MODE_INDEX(tile_mode_index
);
4131 ds
->db_depth_size
= S_028058_PITCH_TILE_MAX((level_info
->nblk_x
/ 8) - 1) |
4132 S_028058_HEIGHT_TILE_MAX((level_info
->nblk_y
/ 8) - 1);
4133 ds
->db_depth_slice
= S_02805C_SLICE_TILE_MAX((level_info
->nblk_x
* level_info
->nblk_y
) / 64 - 1);
4135 if (radv_htile_enabled(iview
->image
, level
)) {
4136 ds
->db_z_info
|= S_028040_TILE_SURFACE_ENABLE(1);
4138 if (!iview
->image
->surface
.has_stencil
&&
4139 !radv_image_is_tc_compat_htile(iview
->image
))
4140 /* Use all of the htile_buffer for depth if there's no stencil. */
4141 ds
->db_stencil_info
|= S_028044_TILE_STENCIL_DISABLE(1);
4143 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
+
4144 iview
->image
->htile_offset
;
4145 ds
->db_htile_data_base
= va
>> 8;
4146 ds
->db_htile_surface
= S_028ABC_FULL_CACHE(1);
4148 if (radv_image_is_tc_compat_htile(iview
->image
)) {
4149 unsigned max_zplanes
=
4150 radv_calc_decompress_on_z_planes(device
, iview
);
4152 ds
->db_htile_surface
|= S_028ABC_TC_COMPATIBLE(1);
4153 ds
->db_z_info
|= S_028040_DECOMPRESS_ON_N_ZPLANES(max_zplanes
);
4158 ds
->db_z_read_base
= ds
->db_z_write_base
= z_offs
>> 8;
4159 ds
->db_stencil_read_base
= ds
->db_stencil_write_base
= s_offs
>> 8;
4162 VkResult
radv_CreateFramebuffer(
4164 const VkFramebufferCreateInfo
* pCreateInfo
,
4165 const VkAllocationCallbacks
* pAllocator
,
4166 VkFramebuffer
* pFramebuffer
)
4168 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4169 struct radv_framebuffer
*framebuffer
;
4171 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
4173 size_t size
= sizeof(*framebuffer
) +
4174 sizeof(struct radv_attachment_info
) * pCreateInfo
->attachmentCount
;
4175 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
4176 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
4177 if (framebuffer
== NULL
)
4178 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
4180 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
4181 framebuffer
->width
= pCreateInfo
->width
;
4182 framebuffer
->height
= pCreateInfo
->height
;
4183 framebuffer
->layers
= pCreateInfo
->layers
;
4184 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
4185 VkImageView _iview
= pCreateInfo
->pAttachments
[i
];
4186 struct radv_image_view
*iview
= radv_image_view_from_handle(_iview
);
4187 framebuffer
->attachments
[i
].attachment
= iview
;
4188 if (iview
->aspect_mask
& VK_IMAGE_ASPECT_COLOR_BIT
) {
4189 radv_initialise_color_surface(device
, &framebuffer
->attachments
[i
].cb
, iview
);
4190 } else if (iview
->aspect_mask
& (VK_IMAGE_ASPECT_DEPTH_BIT
| VK_IMAGE_ASPECT_STENCIL_BIT
)) {
4191 radv_initialise_ds_surface(device
, &framebuffer
->attachments
[i
].ds
, iview
);
4193 framebuffer
->width
= MIN2(framebuffer
->width
, iview
->extent
.width
);
4194 framebuffer
->height
= MIN2(framebuffer
->height
, iview
->extent
.height
);
4195 framebuffer
->layers
= MIN2(framebuffer
->layers
, radv_surface_max_layer_count(iview
));
4198 *pFramebuffer
= radv_framebuffer_to_handle(framebuffer
);
4202 void radv_DestroyFramebuffer(
4205 const VkAllocationCallbacks
* pAllocator
)
4207 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4208 RADV_FROM_HANDLE(radv_framebuffer
, fb
, _fb
);
4212 vk_free2(&device
->alloc
, pAllocator
, fb
);
4215 static unsigned radv_tex_wrap(VkSamplerAddressMode address_mode
)
4217 switch (address_mode
) {
4218 case VK_SAMPLER_ADDRESS_MODE_REPEAT
:
4219 return V_008F30_SQ_TEX_WRAP
;
4220 case VK_SAMPLER_ADDRESS_MODE_MIRRORED_REPEAT
:
4221 return V_008F30_SQ_TEX_MIRROR
;
4222 case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE
:
4223 return V_008F30_SQ_TEX_CLAMP_LAST_TEXEL
;
4224 case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_BORDER
:
4225 return V_008F30_SQ_TEX_CLAMP_BORDER
;
4226 case VK_SAMPLER_ADDRESS_MODE_MIRROR_CLAMP_TO_EDGE
:
4227 return V_008F30_SQ_TEX_MIRROR_ONCE_LAST_TEXEL
;
4229 unreachable("illegal tex wrap mode");
4235 radv_tex_compare(VkCompareOp op
)
4238 case VK_COMPARE_OP_NEVER
:
4239 return V_008F30_SQ_TEX_DEPTH_COMPARE_NEVER
;
4240 case VK_COMPARE_OP_LESS
:
4241 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESS
;
4242 case VK_COMPARE_OP_EQUAL
:
4243 return V_008F30_SQ_TEX_DEPTH_COMPARE_EQUAL
;
4244 case VK_COMPARE_OP_LESS_OR_EQUAL
:
4245 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESSEQUAL
;
4246 case VK_COMPARE_OP_GREATER
:
4247 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATER
;
4248 case VK_COMPARE_OP_NOT_EQUAL
:
4249 return V_008F30_SQ_TEX_DEPTH_COMPARE_NOTEQUAL
;
4250 case VK_COMPARE_OP_GREATER_OR_EQUAL
:
4251 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATEREQUAL
;
4252 case VK_COMPARE_OP_ALWAYS
:
4253 return V_008F30_SQ_TEX_DEPTH_COMPARE_ALWAYS
;
4255 unreachable("illegal compare mode");
4261 radv_tex_filter(VkFilter filter
, unsigned max_ansio
)
4264 case VK_FILTER_NEAREST
:
4265 return (max_ansio
> 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_POINT
:
4266 V_008F38_SQ_TEX_XY_FILTER_POINT
);
4267 case VK_FILTER_LINEAR
:
4268 return (max_ansio
> 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_BILINEAR
:
4269 V_008F38_SQ_TEX_XY_FILTER_BILINEAR
);
4270 case VK_FILTER_CUBIC_IMG
:
4272 fprintf(stderr
, "illegal texture filter");
4278 radv_tex_mipfilter(VkSamplerMipmapMode mode
)
4281 case VK_SAMPLER_MIPMAP_MODE_NEAREST
:
4282 return V_008F38_SQ_TEX_Z_FILTER_POINT
;
4283 case VK_SAMPLER_MIPMAP_MODE_LINEAR
:
4284 return V_008F38_SQ_TEX_Z_FILTER_LINEAR
;
4286 return V_008F38_SQ_TEX_Z_FILTER_NONE
;
4291 radv_tex_bordercolor(VkBorderColor bcolor
)
4294 case VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
:
4295 case VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
:
4296 return V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK
;
4297 case VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
:
4298 case VK_BORDER_COLOR_INT_OPAQUE_BLACK
:
4299 return V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_BLACK
;
4300 case VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
:
4301 case VK_BORDER_COLOR_INT_OPAQUE_WHITE
:
4302 return V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_WHITE
;
4310 radv_tex_aniso_filter(unsigned filter
)
4324 radv_tex_filter_mode(VkSamplerReductionModeEXT mode
)
4327 case VK_SAMPLER_REDUCTION_MODE_WEIGHTED_AVERAGE_EXT
:
4328 return SQ_IMG_FILTER_MODE_BLEND
;
4329 case VK_SAMPLER_REDUCTION_MODE_MIN_EXT
:
4330 return SQ_IMG_FILTER_MODE_MIN
;
4331 case VK_SAMPLER_REDUCTION_MODE_MAX_EXT
:
4332 return SQ_IMG_FILTER_MODE_MAX
;
4340 radv_init_sampler(struct radv_device
*device
,
4341 struct radv_sampler
*sampler
,
4342 const VkSamplerCreateInfo
*pCreateInfo
)
4344 uint32_t max_aniso
= pCreateInfo
->anisotropyEnable
&& pCreateInfo
->maxAnisotropy
> 1.0 ?
4345 (uint32_t) pCreateInfo
->maxAnisotropy
: 0;
4346 uint32_t max_aniso_ratio
= radv_tex_aniso_filter(max_aniso
);
4347 bool is_vi
= (device
->physical_device
->rad_info
.chip_class
>= VI
);
4348 unsigned filter_mode
= SQ_IMG_FILTER_MODE_BLEND
;
4350 const struct VkSamplerReductionModeCreateInfoEXT
*sampler_reduction
=
4351 vk_find_struct_const(pCreateInfo
->pNext
,
4352 SAMPLER_REDUCTION_MODE_CREATE_INFO_EXT
);
4353 if (sampler_reduction
)
4354 filter_mode
= radv_tex_filter_mode(sampler_reduction
->reductionMode
);
4356 sampler
->state
[0] = (S_008F30_CLAMP_X(radv_tex_wrap(pCreateInfo
->addressModeU
)) |
4357 S_008F30_CLAMP_Y(radv_tex_wrap(pCreateInfo
->addressModeV
)) |
4358 S_008F30_CLAMP_Z(radv_tex_wrap(pCreateInfo
->addressModeW
)) |
4359 S_008F30_MAX_ANISO_RATIO(max_aniso_ratio
) |
4360 S_008F30_DEPTH_COMPARE_FUNC(radv_tex_compare(pCreateInfo
->compareOp
)) |
4361 S_008F30_FORCE_UNNORMALIZED(pCreateInfo
->unnormalizedCoordinates
? 1 : 0) |
4362 S_008F30_ANISO_THRESHOLD(max_aniso_ratio
>> 1) |
4363 S_008F30_ANISO_BIAS(max_aniso_ratio
) |
4364 S_008F30_DISABLE_CUBE_WRAP(0) |
4365 S_008F30_COMPAT_MODE(is_vi
) |
4366 S_008F30_FILTER_MODE(filter_mode
));
4367 sampler
->state
[1] = (S_008F34_MIN_LOD(S_FIXED(CLAMP(pCreateInfo
->minLod
, 0, 15), 8)) |
4368 S_008F34_MAX_LOD(S_FIXED(CLAMP(pCreateInfo
->maxLod
, 0, 15), 8)) |
4369 S_008F34_PERF_MIP(max_aniso_ratio
? max_aniso_ratio
+ 6 : 0));
4370 sampler
->state
[2] = (S_008F38_LOD_BIAS(S_FIXED(CLAMP(pCreateInfo
->mipLodBias
, -16, 16), 8)) |
4371 S_008F38_XY_MAG_FILTER(radv_tex_filter(pCreateInfo
->magFilter
, max_aniso
)) |
4372 S_008F38_XY_MIN_FILTER(radv_tex_filter(pCreateInfo
->minFilter
, max_aniso
)) |
4373 S_008F38_MIP_FILTER(radv_tex_mipfilter(pCreateInfo
->mipmapMode
)) |
4374 S_008F38_MIP_POINT_PRECLAMP(0) |
4375 S_008F38_DISABLE_LSB_CEIL(device
->physical_device
->rad_info
.chip_class
<= VI
) |
4376 S_008F38_FILTER_PREC_FIX(1) |
4377 S_008F38_ANISO_OVERRIDE(is_vi
));
4378 sampler
->state
[3] = (S_008F3C_BORDER_COLOR_PTR(0) |
4379 S_008F3C_BORDER_COLOR_TYPE(radv_tex_bordercolor(pCreateInfo
->borderColor
)));
4382 VkResult
radv_CreateSampler(
4384 const VkSamplerCreateInfo
* pCreateInfo
,
4385 const VkAllocationCallbacks
* pAllocator
,
4386 VkSampler
* pSampler
)
4388 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4389 struct radv_sampler
*sampler
;
4391 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO
);
4393 sampler
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*sampler
), 8,
4394 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
4396 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
4398 radv_init_sampler(device
, sampler
, pCreateInfo
);
4399 *pSampler
= radv_sampler_to_handle(sampler
);
4404 void radv_DestroySampler(
4407 const VkAllocationCallbacks
* pAllocator
)
4409 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4410 RADV_FROM_HANDLE(radv_sampler
, sampler
, _sampler
);
4414 vk_free2(&device
->alloc
, pAllocator
, sampler
);
4417 /* vk_icd.h does not declare this function, so we declare it here to
4418 * suppress Wmissing-prototypes.
4420 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4421 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t *pSupportedVersion
);
4423 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4424 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t *pSupportedVersion
)
4426 /* For the full details on loader interface versioning, see
4427 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
4428 * What follows is a condensed summary, to help you navigate the large and
4429 * confusing official doc.
4431 * - Loader interface v0 is incompatible with later versions. We don't
4434 * - In loader interface v1:
4435 * - The first ICD entrypoint called by the loader is
4436 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
4438 * - The ICD must statically expose no other Vulkan symbol unless it is
4439 * linked with -Bsymbolic.
4440 * - Each dispatchable Vulkan handle created by the ICD must be
4441 * a pointer to a struct whose first member is VK_LOADER_DATA. The
4442 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
4443 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
4444 * vkDestroySurfaceKHR(). The ICD must be capable of working with
4445 * such loader-managed surfaces.
4447 * - Loader interface v2 differs from v1 in:
4448 * - The first ICD entrypoint called by the loader is
4449 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
4450 * statically expose this entrypoint.
4452 * - Loader interface v3 differs from v2 in:
4453 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
4454 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
4455 * because the loader no longer does so.
4457 *pSupportedVersion
= MIN2(*pSupportedVersion
, 3u);
4461 VkResult
radv_GetMemoryFdKHR(VkDevice _device
,
4462 const VkMemoryGetFdInfoKHR
*pGetFdInfo
,
4465 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4466 RADV_FROM_HANDLE(radv_device_memory
, memory
, pGetFdInfo
->memory
);
4468 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
4470 /* At the moment, we support only the below handle types. */
4471 assert(pGetFdInfo
->handleType
==
4472 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
||
4473 pGetFdInfo
->handleType
==
4474 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
4476 bool ret
= radv_get_memory_fd(device
, memory
, pFD
);
4478 return vk_error(device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
4482 VkResult
radv_GetMemoryFdPropertiesKHR(VkDevice _device
,
4483 VkExternalMemoryHandleTypeFlagBitsKHR handleType
,
4485 VkMemoryFdPropertiesKHR
*pMemoryFdProperties
)
4487 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4489 switch (handleType
) {
4490 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
:
4491 pMemoryFdProperties
->memoryTypeBits
= (1 << RADV_MEM_TYPE_COUNT
) - 1;
4495 /* The valid usage section for this function says:
4497 * "handleType must not be one of the handle types defined as
4500 * So opaque handle types fall into the default "unsupported" case.
4502 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
4506 static VkResult
radv_import_opaque_fd(struct radv_device
*device
,
4510 uint32_t syncobj_handle
= 0;
4511 int ret
= device
->ws
->import_syncobj(device
->ws
, fd
, &syncobj_handle
);
4513 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
4516 device
->ws
->destroy_syncobj(device
->ws
, *syncobj
);
4518 *syncobj
= syncobj_handle
;
4524 static VkResult
radv_import_sync_fd(struct radv_device
*device
,
4528 /* If we create a syncobj we do it locally so that if we have an error, we don't
4529 * leave a syncobj in an undetermined state in the fence. */
4530 uint32_t syncobj_handle
= *syncobj
;
4531 if (!syncobj_handle
) {
4532 int ret
= device
->ws
->create_syncobj(device
->ws
, &syncobj_handle
);
4534 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
4539 device
->ws
->signal_syncobj(device
->ws
, syncobj_handle
);
4541 int ret
= device
->ws
->import_syncobj_from_sync_file(device
->ws
, syncobj_handle
, fd
);
4543 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
4546 *syncobj
= syncobj_handle
;
4553 VkResult
radv_ImportSemaphoreFdKHR(VkDevice _device
,
4554 const VkImportSemaphoreFdInfoKHR
*pImportSemaphoreFdInfo
)
4556 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4557 RADV_FROM_HANDLE(radv_semaphore
, sem
, pImportSemaphoreFdInfo
->semaphore
);
4558 uint32_t *syncobj_dst
= NULL
;
4560 if (pImportSemaphoreFdInfo
->flags
& VK_SEMAPHORE_IMPORT_TEMPORARY_BIT_KHR
) {
4561 syncobj_dst
= &sem
->temp_syncobj
;
4563 syncobj_dst
= &sem
->syncobj
;
4566 switch(pImportSemaphoreFdInfo
->handleType
) {
4567 case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
:
4568 return radv_import_opaque_fd(device
, pImportSemaphoreFdInfo
->fd
, syncobj_dst
);
4569 case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT_KHR
:
4570 return radv_import_sync_fd(device
, pImportSemaphoreFdInfo
->fd
, syncobj_dst
);
4572 unreachable("Unhandled semaphore handle type");
4576 VkResult
radv_GetSemaphoreFdKHR(VkDevice _device
,
4577 const VkSemaphoreGetFdInfoKHR
*pGetFdInfo
,
4580 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4581 RADV_FROM_HANDLE(radv_semaphore
, sem
, pGetFdInfo
->semaphore
);
4583 uint32_t syncobj_handle
;
4585 if (sem
->temp_syncobj
)
4586 syncobj_handle
= sem
->temp_syncobj
;
4588 syncobj_handle
= sem
->syncobj
;
4590 switch(pGetFdInfo
->handleType
) {
4591 case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
:
4592 ret
= device
->ws
->export_syncobj(device
->ws
, syncobj_handle
, pFd
);
4594 case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT_KHR
:
4595 ret
= device
->ws
->export_syncobj_to_sync_file(device
->ws
, syncobj_handle
, pFd
);
4597 if (sem
->temp_syncobj
) {
4598 close (sem
->temp_syncobj
);
4599 sem
->temp_syncobj
= 0;
4601 device
->ws
->reset_syncobj(device
->ws
, syncobj_handle
);
4606 unreachable("Unhandled semaphore handle type");
4610 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
4614 void radv_GetPhysicalDeviceExternalSemaphoreProperties(
4615 VkPhysicalDevice physicalDevice
,
4616 const VkPhysicalDeviceExternalSemaphoreInfoKHR
* pExternalSemaphoreInfo
,
4617 VkExternalSemaphorePropertiesKHR
* pExternalSemaphoreProperties
)
4619 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
4621 /* Require has_syncobj_wait_for_submit for the syncobj signal ioctl introduced at virtually the same time */
4622 if (pdevice
->rad_info
.has_syncobj_wait_for_submit
&&
4623 (pExternalSemaphoreInfo
->handleType
== VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
||
4624 pExternalSemaphoreInfo
->handleType
== VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT_KHR
)) {
4625 pExternalSemaphoreProperties
->exportFromImportedHandleTypes
= VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
| VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT_KHR
;
4626 pExternalSemaphoreProperties
->compatibleHandleTypes
= VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
| VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT_KHR
;
4627 pExternalSemaphoreProperties
->externalSemaphoreFeatures
= VK_EXTERNAL_SEMAPHORE_FEATURE_EXPORTABLE_BIT_KHR
|
4628 VK_EXTERNAL_SEMAPHORE_FEATURE_IMPORTABLE_BIT_KHR
;
4629 } else if (pExternalSemaphoreInfo
->handleType
== VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
) {
4630 pExternalSemaphoreProperties
->exportFromImportedHandleTypes
= VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
;
4631 pExternalSemaphoreProperties
->compatibleHandleTypes
= VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
;
4632 pExternalSemaphoreProperties
->externalSemaphoreFeatures
= VK_EXTERNAL_SEMAPHORE_FEATURE_EXPORTABLE_BIT_KHR
|
4633 VK_EXTERNAL_SEMAPHORE_FEATURE_IMPORTABLE_BIT_KHR
;
4635 pExternalSemaphoreProperties
->exportFromImportedHandleTypes
= 0;
4636 pExternalSemaphoreProperties
->compatibleHandleTypes
= 0;
4637 pExternalSemaphoreProperties
->externalSemaphoreFeatures
= 0;
4641 VkResult
radv_ImportFenceFdKHR(VkDevice _device
,
4642 const VkImportFenceFdInfoKHR
*pImportFenceFdInfo
)
4644 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4645 RADV_FROM_HANDLE(radv_fence
, fence
, pImportFenceFdInfo
->fence
);
4646 uint32_t *syncobj_dst
= NULL
;
4649 if (pImportFenceFdInfo
->flags
& VK_FENCE_IMPORT_TEMPORARY_BIT_KHR
) {
4650 syncobj_dst
= &fence
->temp_syncobj
;
4652 syncobj_dst
= &fence
->syncobj
;
4655 switch(pImportFenceFdInfo
->handleType
) {
4656 case VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
:
4657 return radv_import_opaque_fd(device
, pImportFenceFdInfo
->fd
, syncobj_dst
);
4658 case VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT_KHR
:
4659 return radv_import_sync_fd(device
, pImportFenceFdInfo
->fd
, syncobj_dst
);
4661 unreachable("Unhandled fence handle type");
4665 VkResult
radv_GetFenceFdKHR(VkDevice _device
,
4666 const VkFenceGetFdInfoKHR
*pGetFdInfo
,
4669 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4670 RADV_FROM_HANDLE(radv_fence
, fence
, pGetFdInfo
->fence
);
4672 uint32_t syncobj_handle
;
4674 if (fence
->temp_syncobj
)
4675 syncobj_handle
= fence
->temp_syncobj
;
4677 syncobj_handle
= fence
->syncobj
;
4679 switch(pGetFdInfo
->handleType
) {
4680 case VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
:
4681 ret
= device
->ws
->export_syncobj(device
->ws
, syncobj_handle
, pFd
);
4683 case VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT_KHR
:
4684 ret
= device
->ws
->export_syncobj_to_sync_file(device
->ws
, syncobj_handle
, pFd
);
4686 if (fence
->temp_syncobj
) {
4687 close (fence
->temp_syncobj
);
4688 fence
->temp_syncobj
= 0;
4690 device
->ws
->reset_syncobj(device
->ws
, syncobj_handle
);
4695 unreachable("Unhandled fence handle type");
4699 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
4703 void radv_GetPhysicalDeviceExternalFenceProperties(
4704 VkPhysicalDevice physicalDevice
,
4705 const VkPhysicalDeviceExternalFenceInfoKHR
* pExternalFenceInfo
,
4706 VkExternalFencePropertiesKHR
* pExternalFenceProperties
)
4708 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
4710 if (pdevice
->rad_info
.has_syncobj_wait_for_submit
&&
4711 (pExternalFenceInfo
->handleType
== VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
||
4712 pExternalFenceInfo
->handleType
== VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT_KHR
)) {
4713 pExternalFenceProperties
->exportFromImportedHandleTypes
= VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
| VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT_KHR
;
4714 pExternalFenceProperties
->compatibleHandleTypes
= VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
| VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT_KHR
;
4715 pExternalFenceProperties
->externalFenceFeatures
= VK_EXTERNAL_FENCE_FEATURE_EXPORTABLE_BIT_KHR
|
4716 VK_EXTERNAL_SEMAPHORE_FEATURE_IMPORTABLE_BIT_KHR
;
4718 pExternalFenceProperties
->exportFromImportedHandleTypes
= 0;
4719 pExternalFenceProperties
->compatibleHandleTypes
= 0;
4720 pExternalFenceProperties
->externalFenceFeatures
= 0;
4725 radv_CreateDebugReportCallbackEXT(VkInstance _instance
,
4726 const VkDebugReportCallbackCreateInfoEXT
* pCreateInfo
,
4727 const VkAllocationCallbacks
* pAllocator
,
4728 VkDebugReportCallbackEXT
* pCallback
)
4730 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
4731 return vk_create_debug_report_callback(&instance
->debug_report_callbacks
,
4732 pCreateInfo
, pAllocator
, &instance
->alloc
,
4737 radv_DestroyDebugReportCallbackEXT(VkInstance _instance
,
4738 VkDebugReportCallbackEXT _callback
,
4739 const VkAllocationCallbacks
* pAllocator
)
4741 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
4742 vk_destroy_debug_report_callback(&instance
->debug_report_callbacks
,
4743 _callback
, pAllocator
, &instance
->alloc
);
4747 radv_DebugReportMessageEXT(VkInstance _instance
,
4748 VkDebugReportFlagsEXT flags
,
4749 VkDebugReportObjectTypeEXT objectType
,
4752 int32_t messageCode
,
4753 const char* pLayerPrefix
,
4754 const char* pMessage
)
4756 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
4757 vk_debug_report(&instance
->debug_report_callbacks
, flags
, objectType
,
4758 object
, location
, messageCode
, pLayerPrefix
, pMessage
);
4762 radv_GetDeviceGroupPeerMemoryFeatures(
4765 uint32_t localDeviceIndex
,
4766 uint32_t remoteDeviceIndex
,
4767 VkPeerMemoryFeatureFlags
* pPeerMemoryFeatures
)
4769 assert(localDeviceIndex
== remoteDeviceIndex
);
4771 *pPeerMemoryFeatures
= VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT
|
4772 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT
|
4773 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT
|
4774 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT
;