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%zd", sizeof(void *));
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
;
483 } else if (!strcmp(name
, "DOOM_VFR")) {
484 /* Work around a Doom VFR game bug */
485 instance
->debug_flags
|= RADV_DEBUG_NO_DYNAMIC_BOUNDS
;
489 static int radv_get_instance_extension_index(const char *name
)
491 for (unsigned i
= 0; i
< RADV_INSTANCE_EXTENSION_COUNT
; ++i
) {
492 if (strcmp(name
, radv_instance_extensions
[i
].extensionName
) == 0)
499 VkResult
radv_CreateInstance(
500 const VkInstanceCreateInfo
* pCreateInfo
,
501 const VkAllocationCallbacks
* pAllocator
,
502 VkInstance
* pInstance
)
504 struct radv_instance
*instance
;
507 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
509 uint32_t client_version
;
510 if (pCreateInfo
->pApplicationInfo
&&
511 pCreateInfo
->pApplicationInfo
->apiVersion
!= 0) {
512 client_version
= pCreateInfo
->pApplicationInfo
->apiVersion
;
514 radv_EnumerateInstanceVersion(&client_version
);
517 instance
= vk_zalloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
518 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
520 return vk_error(NULL
, VK_ERROR_OUT_OF_HOST_MEMORY
);
522 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
525 instance
->alloc
= *pAllocator
;
527 instance
->alloc
= default_alloc
;
529 instance
->apiVersion
= client_version
;
530 instance
->physicalDeviceCount
= -1;
532 instance
->debug_flags
= parse_debug_string(getenv("RADV_DEBUG"),
535 instance
->perftest_flags
= parse_debug_string(getenv("RADV_PERFTEST"),
536 radv_perftest_options
);
539 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
540 radv_logi("Created an instance");
542 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
543 const char *ext_name
= pCreateInfo
->ppEnabledExtensionNames
[i
];
544 int index
= radv_get_instance_extension_index(ext_name
);
546 if (index
< 0 || !radv_supported_instance_extensions
.extensions
[index
]) {
547 vk_free2(&default_alloc
, pAllocator
, instance
);
548 return vk_error(instance
, VK_ERROR_EXTENSION_NOT_PRESENT
);
551 instance
->enabled_extensions
.extensions
[index
] = true;
554 result
= vk_debug_report_instance_init(&instance
->debug_report_callbacks
);
555 if (result
!= VK_SUCCESS
) {
556 vk_free2(&default_alloc
, pAllocator
, instance
);
557 return vk_error(instance
, result
);
562 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
564 radv_handle_per_app_options(instance
, pCreateInfo
->pApplicationInfo
);
566 *pInstance
= radv_instance_to_handle(instance
);
571 void radv_DestroyInstance(
572 VkInstance _instance
,
573 const VkAllocationCallbacks
* pAllocator
)
575 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
580 for (int i
= 0; i
< instance
->physicalDeviceCount
; ++i
) {
581 radv_physical_device_finish(instance
->physicalDevices
+ i
);
584 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
588 vk_debug_report_instance_destroy(&instance
->debug_report_callbacks
);
590 vk_free(&instance
->alloc
, instance
);
594 radv_enumerate_devices(struct radv_instance
*instance
)
596 /* TODO: Check for more devices ? */
597 drmDevicePtr devices
[8];
598 VkResult result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
601 instance
->physicalDeviceCount
= 0;
603 max_devices
= drmGetDevices2(0, devices
, ARRAY_SIZE(devices
));
605 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
606 radv_logi("Found %d drm nodes", max_devices
);
609 return vk_error(instance
, VK_ERROR_INCOMPATIBLE_DRIVER
);
611 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
612 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
613 devices
[i
]->bustype
== DRM_BUS_PCI
&&
614 devices
[i
]->deviceinfo
.pci
->vendor_id
== ATI_VENDOR_ID
) {
616 result
= radv_physical_device_init(instance
->physicalDevices
+
617 instance
->physicalDeviceCount
,
620 if (result
== VK_SUCCESS
)
621 ++instance
->physicalDeviceCount
;
622 else if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
626 drmFreeDevices(devices
, max_devices
);
631 VkResult
radv_EnumeratePhysicalDevices(
632 VkInstance _instance
,
633 uint32_t* pPhysicalDeviceCount
,
634 VkPhysicalDevice
* pPhysicalDevices
)
636 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
639 if (instance
->physicalDeviceCount
< 0) {
640 result
= radv_enumerate_devices(instance
);
641 if (result
!= VK_SUCCESS
&&
642 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
646 if (!pPhysicalDevices
) {
647 *pPhysicalDeviceCount
= instance
->physicalDeviceCount
;
649 *pPhysicalDeviceCount
= MIN2(*pPhysicalDeviceCount
, instance
->physicalDeviceCount
);
650 for (unsigned i
= 0; i
< *pPhysicalDeviceCount
; ++i
)
651 pPhysicalDevices
[i
] = radv_physical_device_to_handle(instance
->physicalDevices
+ i
);
654 return *pPhysicalDeviceCount
< instance
->physicalDeviceCount
? VK_INCOMPLETE
658 VkResult
radv_EnumeratePhysicalDeviceGroups(
659 VkInstance _instance
,
660 uint32_t* pPhysicalDeviceGroupCount
,
661 VkPhysicalDeviceGroupProperties
* pPhysicalDeviceGroupProperties
)
663 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
666 if (instance
->physicalDeviceCount
< 0) {
667 result
= radv_enumerate_devices(instance
);
668 if (result
!= VK_SUCCESS
&&
669 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
673 if (!pPhysicalDeviceGroupProperties
) {
674 *pPhysicalDeviceGroupCount
= instance
->physicalDeviceCount
;
676 *pPhysicalDeviceGroupCount
= MIN2(*pPhysicalDeviceGroupCount
, instance
->physicalDeviceCount
);
677 for (unsigned i
= 0; i
< *pPhysicalDeviceGroupCount
; ++i
) {
678 pPhysicalDeviceGroupProperties
[i
].physicalDeviceCount
= 1;
679 pPhysicalDeviceGroupProperties
[i
].physicalDevices
[0] = radv_physical_device_to_handle(instance
->physicalDevices
+ i
);
680 pPhysicalDeviceGroupProperties
[i
].subsetAllocation
= false;
683 return *pPhysicalDeviceGroupCount
< instance
->physicalDeviceCount
? VK_INCOMPLETE
687 void radv_GetPhysicalDeviceFeatures(
688 VkPhysicalDevice physicalDevice
,
689 VkPhysicalDeviceFeatures
* pFeatures
)
691 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
692 memset(pFeatures
, 0, sizeof(*pFeatures
));
694 *pFeatures
= (VkPhysicalDeviceFeatures
) {
695 .robustBufferAccess
= true,
696 .fullDrawIndexUint32
= true,
697 .imageCubeArray
= true,
698 .independentBlend
= true,
699 .geometryShader
= true,
700 .tessellationShader
= true,
701 .sampleRateShading
= true,
702 .dualSrcBlend
= true,
704 .multiDrawIndirect
= true,
705 .drawIndirectFirstInstance
= true,
707 .depthBiasClamp
= true,
708 .fillModeNonSolid
= true,
713 .multiViewport
= true,
714 .samplerAnisotropy
= true,
715 .textureCompressionETC2
= pdevice
->rad_info
.chip_class
>= GFX9
||
716 pdevice
->rad_info
.family
== CHIP_STONEY
,
717 .textureCompressionASTC_LDR
= false,
718 .textureCompressionBC
= true,
719 .occlusionQueryPrecise
= true,
720 .pipelineStatisticsQuery
= true,
721 .vertexPipelineStoresAndAtomics
= true,
722 .fragmentStoresAndAtomics
= true,
723 .shaderTessellationAndGeometryPointSize
= true,
724 .shaderImageGatherExtended
= true,
725 .shaderStorageImageExtendedFormats
= true,
726 .shaderStorageImageMultisample
= false,
727 .shaderUniformBufferArrayDynamicIndexing
= true,
728 .shaderSampledImageArrayDynamicIndexing
= true,
729 .shaderStorageBufferArrayDynamicIndexing
= true,
730 .shaderStorageImageArrayDynamicIndexing
= true,
731 .shaderStorageImageReadWithoutFormat
= true,
732 .shaderStorageImageWriteWithoutFormat
= true,
733 .shaderClipDistance
= true,
734 .shaderCullDistance
= true,
735 .shaderFloat64
= true,
737 .shaderInt16
= false,
738 .sparseBinding
= true,
739 .variableMultisampleRate
= true,
740 .inheritedQueries
= true,
744 void radv_GetPhysicalDeviceFeatures2(
745 VkPhysicalDevice physicalDevice
,
746 VkPhysicalDeviceFeatures2KHR
*pFeatures
)
748 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
749 vk_foreach_struct(ext
, pFeatures
->pNext
) {
750 switch (ext
->sType
) {
751 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTER_FEATURES_KHR
: {
752 VkPhysicalDeviceVariablePointerFeaturesKHR
*features
= (void *)ext
;
753 features
->variablePointersStorageBuffer
= true;
754 features
->variablePointers
= false;
757 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES_KHR
: {
758 VkPhysicalDeviceMultiviewFeaturesKHR
*features
= (VkPhysicalDeviceMultiviewFeaturesKHR
*)ext
;
759 features
->multiview
= true;
760 features
->multiviewGeometryShader
= true;
761 features
->multiviewTessellationShader
= true;
764 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETER_FEATURES
: {
765 VkPhysicalDeviceShaderDrawParameterFeatures
*features
=
766 (VkPhysicalDeviceShaderDrawParameterFeatures
*)ext
;
767 features
->shaderDrawParameters
= true;
770 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES
: {
771 VkPhysicalDeviceProtectedMemoryFeatures
*features
=
772 (VkPhysicalDeviceProtectedMemoryFeatures
*)ext
;
773 features
->protectedMemory
= false;
776 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES
: {
777 VkPhysicalDevice16BitStorageFeatures
*features
=
778 (VkPhysicalDevice16BitStorageFeatures
*)ext
;
779 bool enabled
= HAVE_LLVM
>= 0x0700 && pdevice
->rad_info
.chip_class
>= VI
;
780 features
->storageBuffer16BitAccess
= enabled
;
781 features
->uniformAndStorageBuffer16BitAccess
= enabled
;
782 features
->storagePushConstant16
= enabled
;
783 features
->storageInputOutput16
= enabled
;
786 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES
: {
787 VkPhysicalDeviceSamplerYcbcrConversionFeatures
*features
=
788 (VkPhysicalDeviceSamplerYcbcrConversionFeatures
*)ext
;
789 features
->samplerYcbcrConversion
= false;
792 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT
: {
793 VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*features
=
794 (VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*)ext
;
795 features
->shaderInputAttachmentArrayDynamicIndexing
= true;
796 features
->shaderUniformTexelBufferArrayDynamicIndexing
= true;
797 features
->shaderStorageTexelBufferArrayDynamicIndexing
= true;
798 features
->shaderUniformBufferArrayNonUniformIndexing
= false;
799 features
->shaderSampledImageArrayNonUniformIndexing
= false;
800 features
->shaderStorageBufferArrayNonUniformIndexing
= false;
801 features
->shaderStorageImageArrayNonUniformIndexing
= false;
802 features
->shaderInputAttachmentArrayNonUniformIndexing
= false;
803 features
->shaderUniformTexelBufferArrayNonUniformIndexing
= false;
804 features
->shaderStorageTexelBufferArrayNonUniformIndexing
= false;
805 features
->descriptorBindingUniformBufferUpdateAfterBind
= true;
806 features
->descriptorBindingSampledImageUpdateAfterBind
= true;
807 features
->descriptorBindingStorageImageUpdateAfterBind
= true;
808 features
->descriptorBindingStorageBufferUpdateAfterBind
= true;
809 features
->descriptorBindingUniformTexelBufferUpdateAfterBind
= true;
810 features
->descriptorBindingStorageTexelBufferUpdateAfterBind
= true;
811 features
->descriptorBindingUpdateUnusedWhilePending
= true;
812 features
->descriptorBindingPartiallyBound
= true;
813 features
->descriptorBindingVariableDescriptorCount
= true;
814 features
->runtimeDescriptorArray
= true;
817 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT
: {
818 VkPhysicalDeviceConditionalRenderingFeaturesEXT
*features
=
819 (VkPhysicalDeviceConditionalRenderingFeaturesEXT
*)ext
;
820 features
->conditionalRendering
= true;
821 features
->inheritedConditionalRendering
= false;
828 return radv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
831 void radv_GetPhysicalDeviceProperties(
832 VkPhysicalDevice physicalDevice
,
833 VkPhysicalDeviceProperties
* pProperties
)
835 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
836 VkSampleCountFlags sample_counts
= 0xf;
838 /* make sure that the entire descriptor set is addressable with a signed
839 * 32-bit int. So the sum of all limits scaled by descriptor size has to
840 * be at most 2 GiB. the combined image & samples object count as one of
841 * both. This limit is for the pipeline layout, not for the set layout, but
842 * there is no set limit, so we just set a pipeline limit. I don't think
843 * any app is going to hit this soon. */
844 size_t max_descriptor_set_size
= ((1ull << 31) - 16 * MAX_DYNAMIC_BUFFERS
) /
845 (32 /* uniform buffer, 32 due to potential space wasted on alignment */ +
846 32 /* storage buffer, 32 due to potential space wasted on alignment */ +
847 32 /* sampler, largest when combined with image */ +
848 64 /* sampled image */ +
849 64 /* storage image */);
851 VkPhysicalDeviceLimits limits
= {
852 .maxImageDimension1D
= (1 << 14),
853 .maxImageDimension2D
= (1 << 14),
854 .maxImageDimension3D
= (1 << 11),
855 .maxImageDimensionCube
= (1 << 14),
856 .maxImageArrayLayers
= (1 << 11),
857 .maxTexelBufferElements
= 128 * 1024 * 1024,
858 .maxUniformBufferRange
= UINT32_MAX
,
859 .maxStorageBufferRange
= UINT32_MAX
,
860 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
861 .maxMemoryAllocationCount
= UINT32_MAX
,
862 .maxSamplerAllocationCount
= 64 * 1024,
863 .bufferImageGranularity
= 64, /* A cache line */
864 .sparseAddressSpaceSize
= 0xffffffffu
, /* buffer max size */
865 .maxBoundDescriptorSets
= MAX_SETS
,
866 .maxPerStageDescriptorSamplers
= max_descriptor_set_size
,
867 .maxPerStageDescriptorUniformBuffers
= max_descriptor_set_size
,
868 .maxPerStageDescriptorStorageBuffers
= max_descriptor_set_size
,
869 .maxPerStageDescriptorSampledImages
= max_descriptor_set_size
,
870 .maxPerStageDescriptorStorageImages
= max_descriptor_set_size
,
871 .maxPerStageDescriptorInputAttachments
= max_descriptor_set_size
,
872 .maxPerStageResources
= max_descriptor_set_size
,
873 .maxDescriptorSetSamplers
= max_descriptor_set_size
,
874 .maxDescriptorSetUniformBuffers
= max_descriptor_set_size
,
875 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_UNIFORM_BUFFERS
,
876 .maxDescriptorSetStorageBuffers
= max_descriptor_set_size
,
877 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_STORAGE_BUFFERS
,
878 .maxDescriptorSetSampledImages
= max_descriptor_set_size
,
879 .maxDescriptorSetStorageImages
= max_descriptor_set_size
,
880 .maxDescriptorSetInputAttachments
= max_descriptor_set_size
,
881 .maxVertexInputAttributes
= 32,
882 .maxVertexInputBindings
= 32,
883 .maxVertexInputAttributeOffset
= 2047,
884 .maxVertexInputBindingStride
= 2048,
885 .maxVertexOutputComponents
= 128,
886 .maxTessellationGenerationLevel
= 64,
887 .maxTessellationPatchSize
= 32,
888 .maxTessellationControlPerVertexInputComponents
= 128,
889 .maxTessellationControlPerVertexOutputComponents
= 128,
890 .maxTessellationControlPerPatchOutputComponents
= 120,
891 .maxTessellationControlTotalOutputComponents
= 4096,
892 .maxTessellationEvaluationInputComponents
= 128,
893 .maxTessellationEvaluationOutputComponents
= 128,
894 .maxGeometryShaderInvocations
= 127,
895 .maxGeometryInputComponents
= 64,
896 .maxGeometryOutputComponents
= 128,
897 .maxGeometryOutputVertices
= 256,
898 .maxGeometryTotalOutputComponents
= 1024,
899 .maxFragmentInputComponents
= 128,
900 .maxFragmentOutputAttachments
= 8,
901 .maxFragmentDualSrcAttachments
= 1,
902 .maxFragmentCombinedOutputResources
= 8,
903 .maxComputeSharedMemorySize
= 32768,
904 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
905 .maxComputeWorkGroupInvocations
= 2048,
906 .maxComputeWorkGroupSize
= {
911 .subPixelPrecisionBits
= 4 /* FIXME */,
912 .subTexelPrecisionBits
= 4 /* FIXME */,
913 .mipmapPrecisionBits
= 4 /* FIXME */,
914 .maxDrawIndexedIndexValue
= UINT32_MAX
,
915 .maxDrawIndirectCount
= UINT32_MAX
,
916 .maxSamplerLodBias
= 16,
917 .maxSamplerAnisotropy
= 16,
918 .maxViewports
= MAX_VIEWPORTS
,
919 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
920 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
921 .viewportSubPixelBits
= 8,
922 .minMemoryMapAlignment
= 4096, /* A page */
923 .minTexelBufferOffsetAlignment
= 1,
924 .minUniformBufferOffsetAlignment
= 4,
925 .minStorageBufferOffsetAlignment
= 4,
926 .minTexelOffset
= -32,
927 .maxTexelOffset
= 31,
928 .minTexelGatherOffset
= -32,
929 .maxTexelGatherOffset
= 31,
930 .minInterpolationOffset
= -2,
931 .maxInterpolationOffset
= 2,
932 .subPixelInterpolationOffsetBits
= 8,
933 .maxFramebufferWidth
= (1 << 14),
934 .maxFramebufferHeight
= (1 << 14),
935 .maxFramebufferLayers
= (1 << 10),
936 .framebufferColorSampleCounts
= sample_counts
,
937 .framebufferDepthSampleCounts
= sample_counts
,
938 .framebufferStencilSampleCounts
= sample_counts
,
939 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
940 .maxColorAttachments
= MAX_RTS
,
941 .sampledImageColorSampleCounts
= sample_counts
,
942 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
943 .sampledImageDepthSampleCounts
= sample_counts
,
944 .sampledImageStencilSampleCounts
= sample_counts
,
945 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
946 .maxSampleMaskWords
= 1,
947 .timestampComputeAndGraphics
= true,
948 .timestampPeriod
= 1000000.0 / pdevice
->rad_info
.clock_crystal_freq
,
949 .maxClipDistances
= 8,
950 .maxCullDistances
= 8,
951 .maxCombinedClipAndCullDistances
= 8,
952 .discreteQueuePriorities
= 1,
953 .pointSizeRange
= { 0.125, 255.875 },
954 .lineWidthRange
= { 0.0, 7.9921875 },
955 .pointSizeGranularity
= (1.0 / 8.0),
956 .lineWidthGranularity
= (1.0 / 128.0),
957 .strictLines
= false, /* FINISHME */
958 .standardSampleLocations
= true,
959 .optimalBufferCopyOffsetAlignment
= 128,
960 .optimalBufferCopyRowPitchAlignment
= 128,
961 .nonCoherentAtomSize
= 64,
964 *pProperties
= (VkPhysicalDeviceProperties
) {
965 .apiVersion
= radv_physical_device_api_version(pdevice
),
966 .driverVersion
= vk_get_driver_version(),
967 .vendorID
= ATI_VENDOR_ID
,
968 .deviceID
= pdevice
->rad_info
.pci_id
,
969 .deviceType
= pdevice
->rad_info
.has_dedicated_vram
? VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU
: VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
971 .sparseProperties
= {0},
974 strcpy(pProperties
->deviceName
, pdevice
->name
);
975 memcpy(pProperties
->pipelineCacheUUID
, pdevice
->cache_uuid
, VK_UUID_SIZE
);
978 void radv_GetPhysicalDeviceProperties2(
979 VkPhysicalDevice physicalDevice
,
980 VkPhysicalDeviceProperties2KHR
*pProperties
)
982 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
983 radv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
985 vk_foreach_struct(ext
, pProperties
->pNext
) {
986 switch (ext
->sType
) {
987 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
988 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
989 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
990 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
993 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES_KHR
: {
994 VkPhysicalDeviceIDPropertiesKHR
*properties
= (VkPhysicalDeviceIDPropertiesKHR
*)ext
;
995 memcpy(properties
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
996 memcpy(properties
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
997 properties
->deviceLUIDValid
= false;
1000 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES_KHR
: {
1001 VkPhysicalDeviceMultiviewPropertiesKHR
*properties
= (VkPhysicalDeviceMultiviewPropertiesKHR
*)ext
;
1002 properties
->maxMultiviewViewCount
= MAX_VIEWS
;
1003 properties
->maxMultiviewInstanceIndex
= INT_MAX
;
1006 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES_KHR
: {
1007 VkPhysicalDevicePointClippingPropertiesKHR
*properties
=
1008 (VkPhysicalDevicePointClippingPropertiesKHR
*)ext
;
1009 properties
->pointClippingBehavior
= VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES_KHR
;
1012 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DISCARD_RECTANGLE_PROPERTIES_EXT
: {
1013 VkPhysicalDeviceDiscardRectanglePropertiesEXT
*properties
=
1014 (VkPhysicalDeviceDiscardRectanglePropertiesEXT
*)ext
;
1015 properties
->maxDiscardRectangles
= MAX_DISCARD_RECTANGLES
;
1018 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT
: {
1019 VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*properties
=
1020 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*) ext
;
1021 properties
->minImportedHostPointerAlignment
= 4096;
1024 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES
: {
1025 VkPhysicalDeviceSubgroupProperties
*properties
=
1026 (VkPhysicalDeviceSubgroupProperties
*)ext
;
1027 properties
->subgroupSize
= 64;
1028 properties
->supportedStages
= VK_SHADER_STAGE_ALL
;
1029 properties
->supportedOperations
=
1030 VK_SUBGROUP_FEATURE_BASIC_BIT
|
1031 VK_SUBGROUP_FEATURE_BALLOT_BIT
|
1032 VK_SUBGROUP_FEATURE_QUAD_BIT
|
1033 VK_SUBGROUP_FEATURE_VOTE_BIT
;
1034 if (pdevice
->rad_info
.chip_class
>= VI
) {
1035 properties
->supportedOperations
|=
1036 VK_SUBGROUP_FEATURE_SHUFFLE_BIT
|
1037 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT
;
1039 properties
->quadOperationsInAllStages
= true;
1042 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES
: {
1043 VkPhysicalDeviceMaintenance3Properties
*properties
=
1044 (VkPhysicalDeviceMaintenance3Properties
*)ext
;
1045 /* Make sure everything is addressable by a signed 32-bit int, and
1046 * our largest descriptors are 96 bytes. */
1047 properties
->maxPerSetDescriptors
= (1ull << 31) / 96;
1048 /* Our buffer size fields allow only this much */
1049 properties
->maxMemoryAllocationSize
= 0xFFFFFFFFull
;
1052 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT
: {
1053 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*properties
=
1054 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*)ext
;
1055 /* GFX6-8 only support single channel min/max filter. */
1056 properties
->filterMinmaxImageComponentMapping
= pdevice
->rad_info
.chip_class
>= GFX9
;
1057 properties
->filterMinmaxSingleComponentFormats
= true;
1060 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CORE_PROPERTIES_AMD
: {
1061 VkPhysicalDeviceShaderCorePropertiesAMD
*properties
=
1062 (VkPhysicalDeviceShaderCorePropertiesAMD
*)ext
;
1064 /* Shader engines. */
1065 properties
->shaderEngineCount
=
1066 pdevice
->rad_info
.max_se
;
1067 properties
->shaderArraysPerEngineCount
=
1068 pdevice
->rad_info
.max_sh_per_se
;
1069 properties
->computeUnitsPerShaderArray
=
1070 pdevice
->rad_info
.num_good_compute_units
/
1071 (pdevice
->rad_info
.max_se
*
1072 pdevice
->rad_info
.max_sh_per_se
);
1073 properties
->simdPerComputeUnit
= 4;
1074 properties
->wavefrontsPerSimd
=
1075 pdevice
->rad_info
.family
== CHIP_TONGA
||
1076 pdevice
->rad_info
.family
== CHIP_ICELAND
||
1077 pdevice
->rad_info
.family
== CHIP_POLARIS10
||
1078 pdevice
->rad_info
.family
== CHIP_POLARIS11
||
1079 pdevice
->rad_info
.family
== CHIP_POLARIS12
||
1080 pdevice
->rad_info
.family
== CHIP_VEGAM
? 8 : 10;
1081 properties
->wavefrontSize
= 64;
1084 properties
->sgprsPerSimd
=
1085 radv_get_num_physical_sgprs(pdevice
);
1086 properties
->minSgprAllocation
=
1087 pdevice
->rad_info
.chip_class
>= VI
? 16 : 8;
1088 properties
->maxSgprAllocation
=
1089 pdevice
->rad_info
.family
== CHIP_TONGA
||
1090 pdevice
->rad_info
.family
== CHIP_ICELAND
? 96 : 104;
1091 properties
->sgprAllocationGranularity
=
1092 pdevice
->rad_info
.chip_class
>= VI
? 16 : 8;
1095 properties
->vgprsPerSimd
= RADV_NUM_PHYSICAL_VGPRS
;
1096 properties
->minVgprAllocation
= 4;
1097 properties
->maxVgprAllocation
= 256;
1098 properties
->vgprAllocationGranularity
= 4;
1101 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT
: {
1102 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*properties
=
1103 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*)ext
;
1104 properties
->maxVertexAttribDivisor
= UINT32_MAX
;
1107 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES_EXT
: {
1108 VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*properties
=
1109 (VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*)ext
;
1110 properties
->maxUpdateAfterBindDescriptorsInAllPools
= UINT32_MAX
/ 64;
1111 properties
->shaderUniformBufferArrayNonUniformIndexingNative
= false;
1112 properties
->shaderSampledImageArrayNonUniformIndexingNative
= false;
1113 properties
->shaderStorageBufferArrayNonUniformIndexingNative
= false;
1114 properties
->shaderStorageImageArrayNonUniformIndexingNative
= false;
1115 properties
->shaderInputAttachmentArrayNonUniformIndexingNative
= false;
1116 properties
->robustBufferAccessUpdateAfterBind
= false;
1117 properties
->quadDivergentImplicitLod
= false;
1119 size_t max_descriptor_set_size
= ((1ull << 31) - 16 * MAX_DYNAMIC_BUFFERS
) /
1120 (32 /* uniform buffer, 32 due to potential space wasted on alignment */ +
1121 32 /* storage buffer, 32 due to potential space wasted on alignment */ +
1122 32 /* sampler, largest when combined with image */ +
1123 64 /* sampled image */ +
1124 64 /* storage image */);
1125 properties
->maxPerStageDescriptorUpdateAfterBindSamplers
= max_descriptor_set_size
;
1126 properties
->maxPerStageDescriptorUpdateAfterBindUniformBuffers
= max_descriptor_set_size
;
1127 properties
->maxPerStageDescriptorUpdateAfterBindStorageBuffers
= max_descriptor_set_size
;
1128 properties
->maxPerStageDescriptorUpdateAfterBindSampledImages
= max_descriptor_set_size
;
1129 properties
->maxPerStageDescriptorUpdateAfterBindStorageImages
= max_descriptor_set_size
;
1130 properties
->maxPerStageDescriptorUpdateAfterBindInputAttachments
= max_descriptor_set_size
;
1131 properties
->maxPerStageUpdateAfterBindResources
= max_descriptor_set_size
;
1132 properties
->maxDescriptorSetUpdateAfterBindSamplers
= max_descriptor_set_size
;
1133 properties
->maxDescriptorSetUpdateAfterBindUniformBuffers
= max_descriptor_set_size
;
1134 properties
->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic
= MAX_DYNAMIC_UNIFORM_BUFFERS
;
1135 properties
->maxDescriptorSetUpdateAfterBindStorageBuffers
= max_descriptor_set_size
;
1136 properties
->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic
= MAX_DYNAMIC_STORAGE_BUFFERS
;
1137 properties
->maxDescriptorSetUpdateAfterBindSampledImages
= max_descriptor_set_size
;
1138 properties
->maxDescriptorSetUpdateAfterBindStorageImages
= max_descriptor_set_size
;
1139 properties
->maxDescriptorSetUpdateAfterBindInputAttachments
= max_descriptor_set_size
;
1142 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES
: {
1143 VkPhysicalDeviceProtectedMemoryProperties
*properties
=
1144 (VkPhysicalDeviceProtectedMemoryProperties
*)ext
;
1145 properties
->protectedNoFault
= false;
1154 static void radv_get_physical_device_queue_family_properties(
1155 struct radv_physical_device
* pdevice
,
1157 VkQueueFamilyProperties
** pQueueFamilyProperties
)
1159 int num_queue_families
= 1;
1161 if (pdevice
->rad_info
.num_compute_rings
> 0 &&
1162 !(pdevice
->instance
->debug_flags
& RADV_DEBUG_NO_COMPUTE_QUEUE
))
1163 num_queue_families
++;
1165 if (pQueueFamilyProperties
== NULL
) {
1166 *pCount
= num_queue_families
;
1175 *pQueueFamilyProperties
[idx
] = (VkQueueFamilyProperties
) {
1176 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
1177 VK_QUEUE_COMPUTE_BIT
|
1178 VK_QUEUE_TRANSFER_BIT
|
1179 VK_QUEUE_SPARSE_BINDING_BIT
,
1181 .timestampValidBits
= 64,
1182 .minImageTransferGranularity
= (VkExtent3D
) { 1, 1, 1 },
1187 if (pdevice
->rad_info
.num_compute_rings
> 0 &&
1188 !(pdevice
->instance
->debug_flags
& RADV_DEBUG_NO_COMPUTE_QUEUE
)) {
1189 if (*pCount
> idx
) {
1190 *pQueueFamilyProperties
[idx
] = (VkQueueFamilyProperties
) {
1191 .queueFlags
= VK_QUEUE_COMPUTE_BIT
|
1192 VK_QUEUE_TRANSFER_BIT
|
1193 VK_QUEUE_SPARSE_BINDING_BIT
,
1194 .queueCount
= pdevice
->rad_info
.num_compute_rings
,
1195 .timestampValidBits
= 64,
1196 .minImageTransferGranularity
= (VkExtent3D
) { 1, 1, 1 },
1204 void radv_GetPhysicalDeviceQueueFamilyProperties(
1205 VkPhysicalDevice physicalDevice
,
1207 VkQueueFamilyProperties
* pQueueFamilyProperties
)
1209 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
1210 if (!pQueueFamilyProperties
) {
1211 return radv_get_physical_device_queue_family_properties(pdevice
, pCount
, NULL
);
1214 VkQueueFamilyProperties
*properties
[] = {
1215 pQueueFamilyProperties
+ 0,
1216 pQueueFamilyProperties
+ 1,
1217 pQueueFamilyProperties
+ 2,
1219 radv_get_physical_device_queue_family_properties(pdevice
, pCount
, properties
);
1220 assert(*pCount
<= 3);
1223 void radv_GetPhysicalDeviceQueueFamilyProperties2(
1224 VkPhysicalDevice physicalDevice
,
1226 VkQueueFamilyProperties2KHR
*pQueueFamilyProperties
)
1228 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
1229 if (!pQueueFamilyProperties
) {
1230 return radv_get_physical_device_queue_family_properties(pdevice
, pCount
, NULL
);
1233 VkQueueFamilyProperties
*properties
[] = {
1234 &pQueueFamilyProperties
[0].queueFamilyProperties
,
1235 &pQueueFamilyProperties
[1].queueFamilyProperties
,
1236 &pQueueFamilyProperties
[2].queueFamilyProperties
,
1238 radv_get_physical_device_queue_family_properties(pdevice
, pCount
, properties
);
1239 assert(*pCount
<= 3);
1242 void radv_GetPhysicalDeviceMemoryProperties(
1243 VkPhysicalDevice physicalDevice
,
1244 VkPhysicalDeviceMemoryProperties
*pMemoryProperties
)
1246 RADV_FROM_HANDLE(radv_physical_device
, physical_device
, physicalDevice
);
1248 *pMemoryProperties
= physical_device
->memory_properties
;
1251 void radv_GetPhysicalDeviceMemoryProperties2(
1252 VkPhysicalDevice physicalDevice
,
1253 VkPhysicalDeviceMemoryProperties2KHR
*pMemoryProperties
)
1255 return radv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
1256 &pMemoryProperties
->memoryProperties
);
1259 VkResult
radv_GetMemoryHostPointerPropertiesEXT(
1261 VkExternalMemoryHandleTypeFlagBitsKHR handleType
,
1262 const void *pHostPointer
,
1263 VkMemoryHostPointerPropertiesEXT
*pMemoryHostPointerProperties
)
1265 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1269 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
: {
1270 const struct radv_physical_device
*physical_device
= device
->physical_device
;
1271 uint32_t memoryTypeBits
= 0;
1272 for (int i
= 0; i
< physical_device
->memory_properties
.memoryTypeCount
; i
++) {
1273 if (physical_device
->mem_type_indices
[i
] == RADV_MEM_TYPE_GTT_CACHED
) {
1274 memoryTypeBits
= (1 << i
);
1278 pMemoryHostPointerProperties
->memoryTypeBits
= memoryTypeBits
;
1282 return VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
;
1286 static enum radeon_ctx_priority
1287 radv_get_queue_global_priority(const VkDeviceQueueGlobalPriorityCreateInfoEXT
*pObj
)
1289 /* Default to MEDIUM when a specific global priority isn't requested */
1291 return RADEON_CTX_PRIORITY_MEDIUM
;
1293 switch(pObj
->globalPriority
) {
1294 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT
:
1295 return RADEON_CTX_PRIORITY_REALTIME
;
1296 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT
:
1297 return RADEON_CTX_PRIORITY_HIGH
;
1298 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
:
1299 return RADEON_CTX_PRIORITY_MEDIUM
;
1300 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT
:
1301 return RADEON_CTX_PRIORITY_LOW
;
1303 unreachable("Illegal global priority value");
1304 return RADEON_CTX_PRIORITY_INVALID
;
1309 radv_queue_init(struct radv_device
*device
, struct radv_queue
*queue
,
1310 uint32_t queue_family_index
, int idx
,
1311 VkDeviceQueueCreateFlags flags
,
1312 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*global_priority
)
1314 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1315 queue
->device
= device
;
1316 queue
->queue_family_index
= queue_family_index
;
1317 queue
->queue_idx
= idx
;
1318 queue
->priority
= radv_get_queue_global_priority(global_priority
);
1319 queue
->flags
= flags
;
1321 queue
->hw_ctx
= device
->ws
->ctx_create(device
->ws
, queue
->priority
);
1323 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
1329 radv_queue_finish(struct radv_queue
*queue
)
1332 queue
->device
->ws
->ctx_destroy(queue
->hw_ctx
);
1334 if (queue
->initial_full_flush_preamble_cs
)
1335 queue
->device
->ws
->cs_destroy(queue
->initial_full_flush_preamble_cs
);
1336 if (queue
->initial_preamble_cs
)
1337 queue
->device
->ws
->cs_destroy(queue
->initial_preamble_cs
);
1338 if (queue
->continue_preamble_cs
)
1339 queue
->device
->ws
->cs_destroy(queue
->continue_preamble_cs
);
1340 if (queue
->descriptor_bo
)
1341 queue
->device
->ws
->buffer_destroy(queue
->descriptor_bo
);
1342 if (queue
->scratch_bo
)
1343 queue
->device
->ws
->buffer_destroy(queue
->scratch_bo
);
1344 if (queue
->esgs_ring_bo
)
1345 queue
->device
->ws
->buffer_destroy(queue
->esgs_ring_bo
);
1346 if (queue
->gsvs_ring_bo
)
1347 queue
->device
->ws
->buffer_destroy(queue
->gsvs_ring_bo
);
1348 if (queue
->tess_rings_bo
)
1349 queue
->device
->ws
->buffer_destroy(queue
->tess_rings_bo
);
1350 if (queue
->compute_scratch_bo
)
1351 queue
->device
->ws
->buffer_destroy(queue
->compute_scratch_bo
);
1355 radv_bo_list_init(struct radv_bo_list
*bo_list
)
1357 pthread_mutex_init(&bo_list
->mutex
, NULL
);
1358 bo_list
->list
.count
= bo_list
->capacity
= 0;
1359 bo_list
->list
.bos
= NULL
;
1363 radv_bo_list_finish(struct radv_bo_list
*bo_list
)
1365 free(bo_list
->list
.bos
);
1366 pthread_mutex_destroy(&bo_list
->mutex
);
1369 static VkResult
radv_bo_list_add(struct radv_device
*device
,
1370 struct radeon_winsys_bo
*bo
)
1372 struct radv_bo_list
*bo_list
= &device
->bo_list
;
1374 if (unlikely(!device
->use_global_bo_list
))
1377 pthread_mutex_lock(&bo_list
->mutex
);
1378 if (bo_list
->list
.count
== bo_list
->capacity
) {
1379 unsigned capacity
= MAX2(4, bo_list
->capacity
* 2);
1380 void *data
= realloc(bo_list
->list
.bos
, capacity
* sizeof(struct radeon_winsys_bo
*));
1383 pthread_mutex_unlock(&bo_list
->mutex
);
1384 return VK_ERROR_OUT_OF_HOST_MEMORY
;
1387 bo_list
->list
.bos
= (struct radeon_winsys_bo
**)data
;
1388 bo_list
->capacity
= capacity
;
1391 bo_list
->list
.bos
[bo_list
->list
.count
++] = bo
;
1392 pthread_mutex_unlock(&bo_list
->mutex
);
1396 static void radv_bo_list_remove(struct radv_device
*device
,
1397 struct radeon_winsys_bo
*bo
)
1399 struct radv_bo_list
*bo_list
= &device
->bo_list
;
1401 if (unlikely(!device
->use_global_bo_list
))
1404 pthread_mutex_lock(&bo_list
->mutex
);
1405 for(unsigned i
= 0; i
< bo_list
->list
.count
; ++i
) {
1406 if (bo_list
->list
.bos
[i
] == bo
) {
1407 bo_list
->list
.bos
[i
] = bo_list
->list
.bos
[bo_list
->list
.count
- 1];
1408 --bo_list
->list
.count
;
1412 pthread_mutex_unlock(&bo_list
->mutex
);
1416 radv_device_init_gs_info(struct radv_device
*device
)
1418 device
->gs_table_depth
= ac_get_gs_table_depth(device
->physical_device
->rad_info
.chip_class
,
1419 device
->physical_device
->rad_info
.family
);
1422 static int radv_get_device_extension_index(const char *name
)
1424 for (unsigned i
= 0; i
< RADV_DEVICE_EXTENSION_COUNT
; ++i
) {
1425 if (strcmp(name
, radv_device_extensions
[i
].extensionName
) == 0)
1431 VkResult
radv_CreateDevice(
1432 VkPhysicalDevice physicalDevice
,
1433 const VkDeviceCreateInfo
* pCreateInfo
,
1434 const VkAllocationCallbacks
* pAllocator
,
1437 RADV_FROM_HANDLE(radv_physical_device
, physical_device
, physicalDevice
);
1439 struct radv_device
*device
;
1441 bool keep_shader_info
= false;
1443 /* Check enabled features */
1444 if (pCreateInfo
->pEnabledFeatures
) {
1445 VkPhysicalDeviceFeatures supported_features
;
1446 radv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
1447 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
1448 VkBool32
*enabled_feature
= (VkBool32
*)pCreateInfo
->pEnabledFeatures
;
1449 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
1450 for (uint32_t i
= 0; i
< num_features
; i
++) {
1451 if (enabled_feature
[i
] && !supported_feature
[i
])
1452 return vk_error(physical_device
->instance
, VK_ERROR_FEATURE_NOT_PRESENT
);
1456 device
= vk_zalloc2(&physical_device
->instance
->alloc
, pAllocator
,
1458 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
1460 return vk_error(physical_device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
1462 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1463 device
->instance
= physical_device
->instance
;
1464 device
->physical_device
= physical_device
;
1466 device
->ws
= physical_device
->ws
;
1468 device
->alloc
= *pAllocator
;
1470 device
->alloc
= physical_device
->instance
->alloc
;
1472 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
1473 const char *ext_name
= pCreateInfo
->ppEnabledExtensionNames
[i
];
1474 int index
= radv_get_device_extension_index(ext_name
);
1475 if (index
< 0 || !physical_device
->supported_extensions
.extensions
[index
]) {
1476 vk_free(&device
->alloc
, device
);
1477 return vk_error(physical_device
->instance
, VK_ERROR_EXTENSION_NOT_PRESENT
);
1480 device
->enabled_extensions
.extensions
[index
] = true;
1483 keep_shader_info
= device
->enabled_extensions
.AMD_shader_info
;
1485 /* With update after bind we can't attach bo's to the command buffer
1486 * from the descriptor set anymore, so we have to use a global BO list.
1488 device
->use_global_bo_list
=
1489 device
->enabled_extensions
.EXT_descriptor_indexing
;
1491 mtx_init(&device
->shader_slab_mutex
, mtx_plain
);
1492 list_inithead(&device
->shader_slabs
);
1494 radv_bo_list_init(&device
->bo_list
);
1496 for (unsigned i
= 0; i
< pCreateInfo
->queueCreateInfoCount
; i
++) {
1497 const VkDeviceQueueCreateInfo
*queue_create
= &pCreateInfo
->pQueueCreateInfos
[i
];
1498 uint32_t qfi
= queue_create
->queueFamilyIndex
;
1499 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*global_priority
=
1500 vk_find_struct_const(queue_create
->pNext
, DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT
);
1502 assert(!global_priority
|| device
->physical_device
->rad_info
.has_ctx_priority
);
1504 device
->queues
[qfi
] = vk_alloc(&device
->alloc
,
1505 queue_create
->queueCount
* sizeof(struct radv_queue
), 8, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
1506 if (!device
->queues
[qfi
]) {
1507 result
= VK_ERROR_OUT_OF_HOST_MEMORY
;
1511 memset(device
->queues
[qfi
], 0, queue_create
->queueCount
* sizeof(struct radv_queue
));
1513 device
->queue_count
[qfi
] = queue_create
->queueCount
;
1515 for (unsigned q
= 0; q
< queue_create
->queueCount
; q
++) {
1516 result
= radv_queue_init(device
, &device
->queues
[qfi
][q
],
1517 qfi
, q
, queue_create
->flags
,
1519 if (result
!= VK_SUCCESS
)
1524 device
->pbb_allowed
= device
->physical_device
->rad_info
.chip_class
>= GFX9
&&
1525 ((device
->instance
->perftest_flags
& RADV_PERFTEST_BINNING
) ||
1526 device
->physical_device
->rad_info
.family
== CHIP_RAVEN
);
1528 /* Disabled and not implemented for now. */
1529 device
->dfsm_allowed
= device
->pbb_allowed
&&
1530 device
->physical_device
->rad_info
.family
== CHIP_RAVEN
;
1533 device
->always_use_syncobj
= device
->physical_device
->rad_info
.has_syncobj_wait_for_submit
;
1536 /* The maximum number of scratch waves. Scratch space isn't divided
1537 * evenly between CUs. The number is only a function of the number of CUs.
1538 * We can decrease the constant to decrease the scratch buffer size.
1540 * sctx->scratch_waves must be >= the maximum possible size of
1541 * 1 threadgroup, so that the hw doesn't hang from being unable
1544 * The recommended value is 4 per CU at most. Higher numbers don't
1545 * bring much benefit, but they still occupy chip resources (think
1546 * async compute). I've seen ~2% performance difference between 4 and 32.
1548 uint32_t max_threads_per_block
= 2048;
1549 device
->scratch_waves
= MAX2(32 * physical_device
->rad_info
.num_good_compute_units
,
1550 max_threads_per_block
/ 64);
1552 device
->dispatch_initiator
= S_00B800_COMPUTE_SHADER_EN(1);
1554 if (device
->physical_device
->rad_info
.chip_class
>= CIK
) {
1555 /* If the KMD allows it (there is a KMD hw register for it),
1556 * allow launching waves out-of-order.
1558 device
->dispatch_initiator
|= S_00B800_ORDER_MODE(1);
1561 radv_device_init_gs_info(device
);
1563 device
->tess_offchip_block_dw_size
=
1564 device
->physical_device
->rad_info
.family
== CHIP_HAWAII
? 4096 : 8192;
1565 device
->has_distributed_tess
=
1566 device
->physical_device
->rad_info
.chip_class
>= VI
&&
1567 device
->physical_device
->rad_info
.max_se
>= 2;
1569 if (getenv("RADV_TRACE_FILE")) {
1570 const char *filename
= getenv("RADV_TRACE_FILE");
1572 keep_shader_info
= true;
1574 if (!radv_init_trace(device
))
1577 fprintf(stderr
, "*****************************************************************************\n");
1578 fprintf(stderr
, "* WARNING: RADV_TRACE_FILE is costly and should only be used for debugging! *\n");
1579 fprintf(stderr
, "*****************************************************************************\n");
1581 fprintf(stderr
, "Trace file will be dumped to %s\n", filename
);
1582 radv_dump_enabled_options(device
, stderr
);
1585 device
->keep_shader_info
= keep_shader_info
;
1587 result
= radv_device_init_meta(device
);
1588 if (result
!= VK_SUCCESS
)
1591 radv_device_init_msaa(device
);
1593 for (int family
= 0; family
< RADV_MAX_QUEUE_FAMILIES
; ++family
) {
1594 device
->empty_cs
[family
] = device
->ws
->cs_create(device
->ws
, family
);
1596 case RADV_QUEUE_GENERAL
:
1597 radeon_emit(device
->empty_cs
[family
], PKT3(PKT3_CONTEXT_CONTROL
, 1, 0));
1598 radeon_emit(device
->empty_cs
[family
], CONTEXT_CONTROL_LOAD_ENABLE(1));
1599 radeon_emit(device
->empty_cs
[family
], CONTEXT_CONTROL_SHADOW_ENABLE(1));
1601 case RADV_QUEUE_COMPUTE
:
1602 radeon_emit(device
->empty_cs
[family
], PKT3(PKT3_NOP
, 0, 0));
1603 radeon_emit(device
->empty_cs
[family
], 0);
1606 device
->ws
->cs_finalize(device
->empty_cs
[family
]);
1609 if (device
->physical_device
->rad_info
.chip_class
>= CIK
)
1610 cik_create_gfx_config(device
);
1612 VkPipelineCacheCreateInfo ci
;
1613 ci
.sType
= VK_STRUCTURE_TYPE_PIPELINE_CACHE_CREATE_INFO
;
1616 ci
.pInitialData
= NULL
;
1617 ci
.initialDataSize
= 0;
1619 result
= radv_CreatePipelineCache(radv_device_to_handle(device
),
1621 if (result
!= VK_SUCCESS
)
1624 device
->mem_cache
= radv_pipeline_cache_from_handle(pc
);
1626 *pDevice
= radv_device_to_handle(device
);
1630 radv_device_finish_meta(device
);
1632 radv_bo_list_finish(&device
->bo_list
);
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
]);
1647 vk_free(&device
->alloc
, device
);
1651 void radv_DestroyDevice(
1653 const VkAllocationCallbacks
* pAllocator
)
1655 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1660 if (device
->trace_bo
)
1661 device
->ws
->buffer_destroy(device
->trace_bo
);
1663 if (device
->gfx_init
)
1664 device
->ws
->buffer_destroy(device
->gfx_init
);
1666 for (unsigned i
= 0; i
< RADV_MAX_QUEUE_FAMILIES
; i
++) {
1667 for (unsigned q
= 0; q
< device
->queue_count
[i
]; q
++)
1668 radv_queue_finish(&device
->queues
[i
][q
]);
1669 if (device
->queue_count
[i
])
1670 vk_free(&device
->alloc
, device
->queues
[i
]);
1671 if (device
->empty_cs
[i
])
1672 device
->ws
->cs_destroy(device
->empty_cs
[i
]);
1674 radv_device_finish_meta(device
);
1676 VkPipelineCache pc
= radv_pipeline_cache_to_handle(device
->mem_cache
);
1677 radv_DestroyPipelineCache(radv_device_to_handle(device
), pc
, NULL
);
1679 radv_destroy_shader_slabs(device
);
1681 radv_bo_list_finish(&device
->bo_list
);
1682 vk_free(&device
->alloc
, device
);
1685 VkResult
radv_EnumerateInstanceLayerProperties(
1686 uint32_t* pPropertyCount
,
1687 VkLayerProperties
* pProperties
)
1689 if (pProperties
== NULL
) {
1690 *pPropertyCount
= 0;
1694 /* None supported at this time */
1695 return vk_error(NULL
, VK_ERROR_LAYER_NOT_PRESENT
);
1698 VkResult
radv_EnumerateDeviceLayerProperties(
1699 VkPhysicalDevice physicalDevice
,
1700 uint32_t* pPropertyCount
,
1701 VkLayerProperties
* pProperties
)
1703 if (pProperties
== NULL
) {
1704 *pPropertyCount
= 0;
1708 /* None supported at this time */
1709 return vk_error(NULL
, VK_ERROR_LAYER_NOT_PRESENT
);
1712 void radv_GetDeviceQueue2(
1714 const VkDeviceQueueInfo2
* pQueueInfo
,
1717 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1718 struct radv_queue
*queue
;
1720 queue
= &device
->queues
[pQueueInfo
->queueFamilyIndex
][pQueueInfo
->queueIndex
];
1721 if (pQueueInfo
->flags
!= queue
->flags
) {
1722 /* From the Vulkan 1.1.70 spec:
1724 * "The queue returned by vkGetDeviceQueue2 must have the same
1725 * flags value from this structure as that used at device
1726 * creation time in a VkDeviceQueueCreateInfo instance. If no
1727 * matching flags were specified at device creation time then
1728 * pQueue will return VK_NULL_HANDLE."
1730 *pQueue
= VK_NULL_HANDLE
;
1734 *pQueue
= radv_queue_to_handle(queue
);
1737 void radv_GetDeviceQueue(
1739 uint32_t queueFamilyIndex
,
1740 uint32_t queueIndex
,
1743 const VkDeviceQueueInfo2 info
= (VkDeviceQueueInfo2
) {
1744 .sType
= VK_STRUCTURE_TYPE_DEVICE_QUEUE_INFO_2
,
1745 .queueFamilyIndex
= queueFamilyIndex
,
1746 .queueIndex
= queueIndex
1749 radv_GetDeviceQueue2(_device
, &info
, pQueue
);
1753 fill_geom_tess_rings(struct radv_queue
*queue
,
1755 bool add_sample_positions
,
1756 uint32_t esgs_ring_size
,
1757 struct radeon_winsys_bo
*esgs_ring_bo
,
1758 uint32_t gsvs_ring_size
,
1759 struct radeon_winsys_bo
*gsvs_ring_bo
,
1760 uint32_t tess_factor_ring_size
,
1761 uint32_t tess_offchip_ring_offset
,
1762 uint32_t tess_offchip_ring_size
,
1763 struct radeon_winsys_bo
*tess_rings_bo
)
1765 uint64_t esgs_va
= 0, gsvs_va
= 0;
1766 uint64_t tess_va
= 0, tess_offchip_va
= 0;
1767 uint32_t *desc
= &map
[4];
1770 esgs_va
= radv_buffer_get_va(esgs_ring_bo
);
1772 gsvs_va
= radv_buffer_get_va(gsvs_ring_bo
);
1773 if (tess_rings_bo
) {
1774 tess_va
= radv_buffer_get_va(tess_rings_bo
);
1775 tess_offchip_va
= tess_va
+ tess_offchip_ring_offset
;
1778 /* stride 0, num records - size, add tid, swizzle, elsize4,
1781 desc
[1] = S_008F04_BASE_ADDRESS_HI(esgs_va
>> 32) |
1782 S_008F04_STRIDE(0) |
1783 S_008F04_SWIZZLE_ENABLE(true);
1784 desc
[2] = esgs_ring_size
;
1785 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1786 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1787 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1788 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1789 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1790 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1791 S_008F0C_ELEMENT_SIZE(1) |
1792 S_008F0C_INDEX_STRIDE(3) |
1793 S_008F0C_ADD_TID_ENABLE(true);
1796 /* GS entry for ES->GS ring */
1797 /* stride 0, num records - size, elsize0,
1800 desc
[1] = S_008F04_BASE_ADDRESS_HI(esgs_va
>> 32)|
1801 S_008F04_STRIDE(0) |
1802 S_008F04_SWIZZLE_ENABLE(false);
1803 desc
[2] = esgs_ring_size
;
1804 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1805 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1806 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1807 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1808 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1809 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1810 S_008F0C_ELEMENT_SIZE(0) |
1811 S_008F0C_INDEX_STRIDE(0) |
1812 S_008F0C_ADD_TID_ENABLE(false);
1815 /* VS entry for GS->VS ring */
1816 /* stride 0, num records - size, elsize0,
1819 desc
[1] = S_008F04_BASE_ADDRESS_HI(gsvs_va
>> 32)|
1820 S_008F04_STRIDE(0) |
1821 S_008F04_SWIZZLE_ENABLE(false);
1822 desc
[2] = gsvs_ring_size
;
1823 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1824 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1825 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1826 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1827 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1828 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1829 S_008F0C_ELEMENT_SIZE(0) |
1830 S_008F0C_INDEX_STRIDE(0) |
1831 S_008F0C_ADD_TID_ENABLE(false);
1834 /* stride gsvs_itemsize, num records 64
1835 elsize 4, index stride 16 */
1836 /* shader will patch stride and desc[2] */
1838 desc
[1] = S_008F04_BASE_ADDRESS_HI(gsvs_va
>> 32)|
1839 S_008F04_STRIDE(0) |
1840 S_008F04_SWIZZLE_ENABLE(true);
1842 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1843 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1844 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1845 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1846 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1847 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1848 S_008F0C_ELEMENT_SIZE(1) |
1849 S_008F0C_INDEX_STRIDE(1) |
1850 S_008F0C_ADD_TID_ENABLE(true);
1854 desc
[1] = S_008F04_BASE_ADDRESS_HI(tess_va
>> 32) |
1855 S_008F04_STRIDE(0) |
1856 S_008F04_SWIZZLE_ENABLE(false);
1857 desc
[2] = tess_factor_ring_size
;
1858 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1859 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1860 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1861 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1862 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1863 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1864 S_008F0C_ELEMENT_SIZE(0) |
1865 S_008F0C_INDEX_STRIDE(0) |
1866 S_008F0C_ADD_TID_ENABLE(false);
1869 desc
[0] = tess_offchip_va
;
1870 desc
[1] = S_008F04_BASE_ADDRESS_HI(tess_offchip_va
>> 32) |
1871 S_008F04_STRIDE(0) |
1872 S_008F04_SWIZZLE_ENABLE(false);
1873 desc
[2] = tess_offchip_ring_size
;
1874 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1875 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1876 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1877 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1878 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1879 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1880 S_008F0C_ELEMENT_SIZE(0) |
1881 S_008F0C_INDEX_STRIDE(0) |
1882 S_008F0C_ADD_TID_ENABLE(false);
1885 /* add sample positions after all rings */
1886 memcpy(desc
, queue
->device
->sample_locations_1x
, 8);
1888 memcpy(desc
, queue
->device
->sample_locations_2x
, 16);
1890 memcpy(desc
, queue
->device
->sample_locations_4x
, 32);
1892 memcpy(desc
, queue
->device
->sample_locations_8x
, 64);
1894 memcpy(desc
, queue
->device
->sample_locations_16x
, 128);
1898 radv_get_hs_offchip_param(struct radv_device
*device
, uint32_t *max_offchip_buffers_p
)
1900 bool double_offchip_buffers
= device
->physical_device
->rad_info
.chip_class
>= CIK
&&
1901 device
->physical_device
->rad_info
.family
!= CHIP_CARRIZO
&&
1902 device
->physical_device
->rad_info
.family
!= CHIP_STONEY
;
1903 unsigned max_offchip_buffers_per_se
= double_offchip_buffers
? 128 : 64;
1904 unsigned max_offchip_buffers
= max_offchip_buffers_per_se
*
1905 device
->physical_device
->rad_info
.max_se
;
1906 unsigned offchip_granularity
;
1907 unsigned hs_offchip_param
;
1908 switch (device
->tess_offchip_block_dw_size
) {
1913 offchip_granularity
= V_03093C_X_8K_DWORDS
;
1916 offchip_granularity
= V_03093C_X_4K_DWORDS
;
1920 switch (device
->physical_device
->rad_info
.chip_class
) {
1922 max_offchip_buffers
= MIN2(max_offchip_buffers
, 126);
1928 max_offchip_buffers
= MIN2(max_offchip_buffers
, 508);
1932 *max_offchip_buffers_p
= max_offchip_buffers
;
1933 if (device
->physical_device
->rad_info
.chip_class
>= CIK
) {
1934 if (device
->physical_device
->rad_info
.chip_class
>= VI
)
1935 --max_offchip_buffers
;
1937 S_03093C_OFFCHIP_BUFFERING(max_offchip_buffers
) |
1938 S_03093C_OFFCHIP_GRANULARITY(offchip_granularity
);
1941 S_0089B0_OFFCHIP_BUFFERING(max_offchip_buffers
);
1943 return hs_offchip_param
;
1947 radv_emit_gs_ring_sizes(struct radv_queue
*queue
, struct radeon_cmdbuf
*cs
,
1948 struct radeon_winsys_bo
*esgs_ring_bo
,
1949 uint32_t esgs_ring_size
,
1950 struct radeon_winsys_bo
*gsvs_ring_bo
,
1951 uint32_t gsvs_ring_size
)
1953 if (!esgs_ring_bo
&& !gsvs_ring_bo
)
1957 radv_cs_add_buffer(queue
->device
->ws
, cs
, esgs_ring_bo
);
1960 radv_cs_add_buffer(queue
->device
->ws
, cs
, gsvs_ring_bo
);
1962 if (queue
->device
->physical_device
->rad_info
.chip_class
>= CIK
) {
1963 radeon_set_uconfig_reg_seq(cs
, R_030900_VGT_ESGS_RING_SIZE
, 2);
1964 radeon_emit(cs
, esgs_ring_size
>> 8);
1965 radeon_emit(cs
, gsvs_ring_size
>> 8);
1967 radeon_set_config_reg_seq(cs
, R_0088C8_VGT_ESGS_RING_SIZE
, 2);
1968 radeon_emit(cs
, esgs_ring_size
>> 8);
1969 radeon_emit(cs
, gsvs_ring_size
>> 8);
1974 radv_emit_tess_factor_ring(struct radv_queue
*queue
, struct radeon_cmdbuf
*cs
,
1975 unsigned hs_offchip_param
, unsigned tf_ring_size
,
1976 struct radeon_winsys_bo
*tess_rings_bo
)
1983 tf_va
= radv_buffer_get_va(tess_rings_bo
);
1985 radv_cs_add_buffer(queue
->device
->ws
, cs
, tess_rings_bo
);
1987 if (queue
->device
->physical_device
->rad_info
.chip_class
>= CIK
) {
1988 radeon_set_uconfig_reg(cs
, R_030938_VGT_TF_RING_SIZE
,
1989 S_030938_SIZE(tf_ring_size
/ 4));
1990 radeon_set_uconfig_reg(cs
, R_030940_VGT_TF_MEMORY_BASE
,
1992 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
1993 radeon_set_uconfig_reg(cs
, R_030944_VGT_TF_MEMORY_BASE_HI
,
1994 S_030944_BASE_HI(tf_va
>> 40));
1996 radeon_set_uconfig_reg(cs
, R_03093C_VGT_HS_OFFCHIP_PARAM
,
1999 radeon_set_config_reg(cs
, R_008988_VGT_TF_RING_SIZE
,
2000 S_008988_SIZE(tf_ring_size
/ 4));
2001 radeon_set_config_reg(cs
, R_0089B8_VGT_TF_MEMORY_BASE
,
2003 radeon_set_config_reg(cs
, R_0089B0_VGT_HS_OFFCHIP_PARAM
,
2009 radv_emit_compute_scratch(struct radv_queue
*queue
, struct radeon_cmdbuf
*cs
,
2010 struct radeon_winsys_bo
*compute_scratch_bo
)
2012 uint64_t scratch_va
;
2014 if (!compute_scratch_bo
)
2017 scratch_va
= radv_buffer_get_va(compute_scratch_bo
);
2019 radv_cs_add_buffer(queue
->device
->ws
, cs
, compute_scratch_bo
);
2021 radeon_set_sh_reg_seq(cs
, R_00B900_COMPUTE_USER_DATA_0
, 2);
2022 radeon_emit(cs
, scratch_va
);
2023 radeon_emit(cs
, S_008F04_BASE_ADDRESS_HI(scratch_va
>> 32) |
2024 S_008F04_SWIZZLE_ENABLE(1));
2028 radv_emit_global_shader_pointers(struct radv_queue
*queue
,
2029 struct radeon_cmdbuf
*cs
,
2030 struct radeon_winsys_bo
*descriptor_bo
)
2037 va
= radv_buffer_get_va(descriptor_bo
);
2039 radv_cs_add_buffer(queue
->device
->ws
, cs
, descriptor_bo
);
2041 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
2042 uint32_t regs
[] = {R_00B030_SPI_SHADER_USER_DATA_PS_0
,
2043 R_00B130_SPI_SHADER_USER_DATA_VS_0
,
2044 R_00B208_SPI_SHADER_USER_DATA_ADDR_LO_GS
,
2045 R_00B408_SPI_SHADER_USER_DATA_ADDR_LO_HS
};
2047 for (int i
= 0; i
< ARRAY_SIZE(regs
); ++i
) {
2048 radv_emit_shader_pointer(queue
->device
, cs
, regs
[i
],
2052 uint32_t regs
[] = {R_00B030_SPI_SHADER_USER_DATA_PS_0
,
2053 R_00B130_SPI_SHADER_USER_DATA_VS_0
,
2054 R_00B230_SPI_SHADER_USER_DATA_GS_0
,
2055 R_00B330_SPI_SHADER_USER_DATA_ES_0
,
2056 R_00B430_SPI_SHADER_USER_DATA_HS_0
,
2057 R_00B530_SPI_SHADER_USER_DATA_LS_0
};
2059 for (int i
= 0; i
< ARRAY_SIZE(regs
); ++i
) {
2060 radv_emit_shader_pointer(queue
->device
, cs
, regs
[i
],
2067 radv_get_preamble_cs(struct radv_queue
*queue
,
2068 uint32_t scratch_size
,
2069 uint32_t compute_scratch_size
,
2070 uint32_t esgs_ring_size
,
2071 uint32_t gsvs_ring_size
,
2072 bool needs_tess_rings
,
2073 bool needs_sample_positions
,
2074 struct radeon_cmdbuf
**initial_full_flush_preamble_cs
,
2075 struct radeon_cmdbuf
**initial_preamble_cs
,
2076 struct radeon_cmdbuf
**continue_preamble_cs
)
2078 struct radeon_winsys_bo
*scratch_bo
= NULL
;
2079 struct radeon_winsys_bo
*descriptor_bo
= NULL
;
2080 struct radeon_winsys_bo
*compute_scratch_bo
= NULL
;
2081 struct radeon_winsys_bo
*esgs_ring_bo
= NULL
;
2082 struct radeon_winsys_bo
*gsvs_ring_bo
= NULL
;
2083 struct radeon_winsys_bo
*tess_rings_bo
= NULL
;
2084 struct radeon_cmdbuf
*dest_cs
[3] = {0};
2085 bool add_tess_rings
= false, add_sample_positions
= false;
2086 unsigned tess_factor_ring_size
= 0, tess_offchip_ring_size
= 0;
2087 unsigned max_offchip_buffers
;
2088 unsigned hs_offchip_param
= 0;
2089 unsigned tess_offchip_ring_offset
;
2090 uint32_t ring_bo_flags
= RADEON_FLAG_NO_CPU_ACCESS
| RADEON_FLAG_NO_INTERPROCESS_SHARING
;
2091 if (!queue
->has_tess_rings
) {
2092 if (needs_tess_rings
)
2093 add_tess_rings
= true;
2095 if (!queue
->has_sample_positions
) {
2096 if (needs_sample_positions
)
2097 add_sample_positions
= true;
2099 tess_factor_ring_size
= 32768 * queue
->device
->physical_device
->rad_info
.max_se
;
2100 hs_offchip_param
= radv_get_hs_offchip_param(queue
->device
,
2101 &max_offchip_buffers
);
2102 tess_offchip_ring_offset
= align(tess_factor_ring_size
, 64 * 1024);
2103 tess_offchip_ring_size
= max_offchip_buffers
*
2104 queue
->device
->tess_offchip_block_dw_size
* 4;
2106 if (scratch_size
<= queue
->scratch_size
&&
2107 compute_scratch_size
<= queue
->compute_scratch_size
&&
2108 esgs_ring_size
<= queue
->esgs_ring_size
&&
2109 gsvs_ring_size
<= queue
->gsvs_ring_size
&&
2110 !add_tess_rings
&& !add_sample_positions
&&
2111 queue
->initial_preamble_cs
) {
2112 *initial_full_flush_preamble_cs
= queue
->initial_full_flush_preamble_cs
;
2113 *initial_preamble_cs
= queue
->initial_preamble_cs
;
2114 *continue_preamble_cs
= queue
->continue_preamble_cs
;
2115 if (!scratch_size
&& !compute_scratch_size
&& !esgs_ring_size
&& !gsvs_ring_size
)
2116 *continue_preamble_cs
= NULL
;
2120 if (scratch_size
> queue
->scratch_size
) {
2121 scratch_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
2129 scratch_bo
= queue
->scratch_bo
;
2131 if (compute_scratch_size
> queue
->compute_scratch_size
) {
2132 compute_scratch_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
2133 compute_scratch_size
,
2137 if (!compute_scratch_bo
)
2141 compute_scratch_bo
= queue
->compute_scratch_bo
;
2143 if (esgs_ring_size
> queue
->esgs_ring_size
) {
2144 esgs_ring_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
2152 esgs_ring_bo
= queue
->esgs_ring_bo
;
2153 esgs_ring_size
= queue
->esgs_ring_size
;
2156 if (gsvs_ring_size
> queue
->gsvs_ring_size
) {
2157 gsvs_ring_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
2165 gsvs_ring_bo
= queue
->gsvs_ring_bo
;
2166 gsvs_ring_size
= queue
->gsvs_ring_size
;
2169 if (add_tess_rings
) {
2170 tess_rings_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
2171 tess_offchip_ring_offset
+ tess_offchip_ring_size
,
2178 tess_rings_bo
= queue
->tess_rings_bo
;
2181 if (scratch_bo
!= queue
->scratch_bo
||
2182 esgs_ring_bo
!= queue
->esgs_ring_bo
||
2183 gsvs_ring_bo
!= queue
->gsvs_ring_bo
||
2184 tess_rings_bo
!= queue
->tess_rings_bo
||
2185 add_sample_positions
) {
2187 if (gsvs_ring_bo
|| esgs_ring_bo
||
2188 tess_rings_bo
|| add_sample_positions
) {
2189 size
= 112; /* 2 dword + 2 padding + 4 dword * 6 */
2190 if (add_sample_positions
)
2191 size
+= 256; /* 32+16+8+4+2+1 samples * 4 * 2 = 248 bytes. */
2193 else if (scratch_bo
)
2194 size
= 8; /* 2 dword */
2196 descriptor_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
2200 RADEON_FLAG_CPU_ACCESS
|
2201 RADEON_FLAG_NO_INTERPROCESS_SHARING
|
2202 RADEON_FLAG_READ_ONLY
);
2206 descriptor_bo
= queue
->descriptor_bo
;
2208 for(int i
= 0; i
< 3; ++i
) {
2209 struct radeon_cmdbuf
*cs
= NULL
;
2210 cs
= queue
->device
->ws
->cs_create(queue
->device
->ws
,
2211 queue
->queue_family_index
? RING_COMPUTE
: RING_GFX
);
2218 radv_cs_add_buffer(queue
->device
->ws
, cs
, scratch_bo
);
2220 if (descriptor_bo
!= queue
->descriptor_bo
) {
2221 uint32_t *map
= (uint32_t*)queue
->device
->ws
->buffer_map(descriptor_bo
);
2224 uint64_t scratch_va
= radv_buffer_get_va(scratch_bo
);
2225 uint32_t rsrc1
= S_008F04_BASE_ADDRESS_HI(scratch_va
>> 32) |
2226 S_008F04_SWIZZLE_ENABLE(1);
2227 map
[0] = scratch_va
;
2231 if (esgs_ring_bo
|| gsvs_ring_bo
|| tess_rings_bo
||
2232 add_sample_positions
)
2233 fill_geom_tess_rings(queue
, map
, add_sample_positions
,
2234 esgs_ring_size
, esgs_ring_bo
,
2235 gsvs_ring_size
, gsvs_ring_bo
,
2236 tess_factor_ring_size
,
2237 tess_offchip_ring_offset
,
2238 tess_offchip_ring_size
,
2241 queue
->device
->ws
->buffer_unmap(descriptor_bo
);
2244 if (esgs_ring_bo
|| gsvs_ring_bo
|| tess_rings_bo
) {
2245 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
2246 radeon_emit(cs
, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
2247 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
2248 radeon_emit(cs
, EVENT_TYPE(V_028A90_VGT_FLUSH
) | EVENT_INDEX(0));
2251 radv_emit_gs_ring_sizes(queue
, cs
, esgs_ring_bo
, esgs_ring_size
,
2252 gsvs_ring_bo
, gsvs_ring_size
);
2253 radv_emit_tess_factor_ring(queue
, cs
, hs_offchip_param
,
2254 tess_factor_ring_size
, tess_rings_bo
);
2255 radv_emit_global_shader_pointers(queue
, cs
, descriptor_bo
);
2256 radv_emit_compute_scratch(queue
, cs
, compute_scratch_bo
);
2259 si_cs_emit_cache_flush(cs
,
2260 queue
->device
->physical_device
->rad_info
.chip_class
,
2262 queue
->queue_family_index
== RING_COMPUTE
&&
2263 queue
->device
->physical_device
->rad_info
.chip_class
>= CIK
,
2264 (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
)) |
2265 RADV_CMD_FLAG_INV_ICACHE
|
2266 RADV_CMD_FLAG_INV_SMEM_L1
|
2267 RADV_CMD_FLAG_INV_VMEM_L1
|
2268 RADV_CMD_FLAG_INV_GLOBAL_L2
|
2269 RADV_CMD_FLAG_START_PIPELINE_STATS
, 0);
2270 } else if (i
== 1) {
2271 si_cs_emit_cache_flush(cs
,
2272 queue
->device
->physical_device
->rad_info
.chip_class
,
2274 queue
->queue_family_index
== RING_COMPUTE
&&
2275 queue
->device
->physical_device
->rad_info
.chip_class
>= CIK
,
2276 RADV_CMD_FLAG_INV_ICACHE
|
2277 RADV_CMD_FLAG_INV_SMEM_L1
|
2278 RADV_CMD_FLAG_INV_VMEM_L1
|
2279 RADV_CMD_FLAG_INV_GLOBAL_L2
|
2280 RADV_CMD_FLAG_START_PIPELINE_STATS
, 0);
2283 if (!queue
->device
->ws
->cs_finalize(cs
))
2287 if (queue
->initial_full_flush_preamble_cs
)
2288 queue
->device
->ws
->cs_destroy(queue
->initial_full_flush_preamble_cs
);
2290 if (queue
->initial_preamble_cs
)
2291 queue
->device
->ws
->cs_destroy(queue
->initial_preamble_cs
);
2293 if (queue
->continue_preamble_cs
)
2294 queue
->device
->ws
->cs_destroy(queue
->continue_preamble_cs
);
2296 queue
->initial_full_flush_preamble_cs
= dest_cs
[0];
2297 queue
->initial_preamble_cs
= dest_cs
[1];
2298 queue
->continue_preamble_cs
= dest_cs
[2];
2300 if (scratch_bo
!= queue
->scratch_bo
) {
2301 if (queue
->scratch_bo
)
2302 queue
->device
->ws
->buffer_destroy(queue
->scratch_bo
);
2303 queue
->scratch_bo
= scratch_bo
;
2304 queue
->scratch_size
= scratch_size
;
2307 if (compute_scratch_bo
!= queue
->compute_scratch_bo
) {
2308 if (queue
->compute_scratch_bo
)
2309 queue
->device
->ws
->buffer_destroy(queue
->compute_scratch_bo
);
2310 queue
->compute_scratch_bo
= compute_scratch_bo
;
2311 queue
->compute_scratch_size
= compute_scratch_size
;
2314 if (esgs_ring_bo
!= queue
->esgs_ring_bo
) {
2315 if (queue
->esgs_ring_bo
)
2316 queue
->device
->ws
->buffer_destroy(queue
->esgs_ring_bo
);
2317 queue
->esgs_ring_bo
= esgs_ring_bo
;
2318 queue
->esgs_ring_size
= esgs_ring_size
;
2321 if (gsvs_ring_bo
!= queue
->gsvs_ring_bo
) {
2322 if (queue
->gsvs_ring_bo
)
2323 queue
->device
->ws
->buffer_destroy(queue
->gsvs_ring_bo
);
2324 queue
->gsvs_ring_bo
= gsvs_ring_bo
;
2325 queue
->gsvs_ring_size
= gsvs_ring_size
;
2328 if (tess_rings_bo
!= queue
->tess_rings_bo
) {
2329 queue
->tess_rings_bo
= tess_rings_bo
;
2330 queue
->has_tess_rings
= true;
2333 if (descriptor_bo
!= queue
->descriptor_bo
) {
2334 if (queue
->descriptor_bo
)
2335 queue
->device
->ws
->buffer_destroy(queue
->descriptor_bo
);
2337 queue
->descriptor_bo
= descriptor_bo
;
2340 if (add_sample_positions
)
2341 queue
->has_sample_positions
= true;
2343 *initial_full_flush_preamble_cs
= queue
->initial_full_flush_preamble_cs
;
2344 *initial_preamble_cs
= queue
->initial_preamble_cs
;
2345 *continue_preamble_cs
= queue
->continue_preamble_cs
;
2346 if (!scratch_size
&& !compute_scratch_size
&& !esgs_ring_size
&& !gsvs_ring_size
)
2347 *continue_preamble_cs
= NULL
;
2350 for (int i
= 0; i
< ARRAY_SIZE(dest_cs
); ++i
)
2352 queue
->device
->ws
->cs_destroy(dest_cs
[i
]);
2353 if (descriptor_bo
&& descriptor_bo
!= queue
->descriptor_bo
)
2354 queue
->device
->ws
->buffer_destroy(descriptor_bo
);
2355 if (scratch_bo
&& scratch_bo
!= queue
->scratch_bo
)
2356 queue
->device
->ws
->buffer_destroy(scratch_bo
);
2357 if (compute_scratch_bo
&& compute_scratch_bo
!= queue
->compute_scratch_bo
)
2358 queue
->device
->ws
->buffer_destroy(compute_scratch_bo
);
2359 if (esgs_ring_bo
&& esgs_ring_bo
!= queue
->esgs_ring_bo
)
2360 queue
->device
->ws
->buffer_destroy(esgs_ring_bo
);
2361 if (gsvs_ring_bo
&& gsvs_ring_bo
!= queue
->gsvs_ring_bo
)
2362 queue
->device
->ws
->buffer_destroy(gsvs_ring_bo
);
2363 if (tess_rings_bo
&& tess_rings_bo
!= queue
->tess_rings_bo
)
2364 queue
->device
->ws
->buffer_destroy(tess_rings_bo
);
2365 return vk_error(queue
->device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
2368 static VkResult
radv_alloc_sem_counts(struct radv_instance
*instance
,
2369 struct radv_winsys_sem_counts
*counts
,
2371 const VkSemaphore
*sems
,
2375 int syncobj_idx
= 0, sem_idx
= 0;
2377 if (num_sems
== 0 && _fence
== VK_NULL_HANDLE
)
2380 for (uint32_t i
= 0; i
< num_sems
; i
++) {
2381 RADV_FROM_HANDLE(radv_semaphore
, sem
, sems
[i
]);
2383 if (sem
->temp_syncobj
|| sem
->syncobj
)
2384 counts
->syncobj_count
++;
2386 counts
->sem_count
++;
2389 if (_fence
!= VK_NULL_HANDLE
) {
2390 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
2391 if (fence
->temp_syncobj
|| fence
->syncobj
)
2392 counts
->syncobj_count
++;
2395 if (counts
->syncobj_count
) {
2396 counts
->syncobj
= (uint32_t *)malloc(sizeof(uint32_t) * counts
->syncobj_count
);
2397 if (!counts
->syncobj
)
2398 return vk_error(instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
2401 if (counts
->sem_count
) {
2402 counts
->sem
= (struct radeon_winsys_sem
**)malloc(sizeof(struct radeon_winsys_sem
*) * counts
->sem_count
);
2404 free(counts
->syncobj
);
2405 return vk_error(instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
2409 for (uint32_t i
= 0; i
< num_sems
; i
++) {
2410 RADV_FROM_HANDLE(radv_semaphore
, sem
, sems
[i
]);
2412 if (sem
->temp_syncobj
) {
2413 counts
->syncobj
[syncobj_idx
++] = sem
->temp_syncobj
;
2415 else if (sem
->syncobj
)
2416 counts
->syncobj
[syncobj_idx
++] = sem
->syncobj
;
2419 counts
->sem
[sem_idx
++] = sem
->sem
;
2423 if (_fence
!= VK_NULL_HANDLE
) {
2424 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
2425 if (fence
->temp_syncobj
)
2426 counts
->syncobj
[syncobj_idx
++] = fence
->temp_syncobj
;
2427 else if (fence
->syncobj
)
2428 counts
->syncobj
[syncobj_idx
++] = fence
->syncobj
;
2435 radv_free_sem_info(struct radv_winsys_sem_info
*sem_info
)
2437 free(sem_info
->wait
.syncobj
);
2438 free(sem_info
->wait
.sem
);
2439 free(sem_info
->signal
.syncobj
);
2440 free(sem_info
->signal
.sem
);
2444 static void radv_free_temp_syncobjs(struct radv_device
*device
,
2446 const VkSemaphore
*sems
)
2448 for (uint32_t i
= 0; i
< num_sems
; i
++) {
2449 RADV_FROM_HANDLE(radv_semaphore
, sem
, sems
[i
]);
2451 if (sem
->temp_syncobj
) {
2452 device
->ws
->destroy_syncobj(device
->ws
, sem
->temp_syncobj
);
2453 sem
->temp_syncobj
= 0;
2459 radv_alloc_sem_info(struct radv_instance
*instance
,
2460 struct radv_winsys_sem_info
*sem_info
,
2462 const VkSemaphore
*wait_sems
,
2463 int num_signal_sems
,
2464 const VkSemaphore
*signal_sems
,
2468 memset(sem_info
, 0, sizeof(*sem_info
));
2470 ret
= radv_alloc_sem_counts(instance
, &sem_info
->wait
, num_wait_sems
, wait_sems
, VK_NULL_HANDLE
, true);
2473 ret
= radv_alloc_sem_counts(instance
, &sem_info
->signal
, num_signal_sems
, signal_sems
, fence
, false);
2475 radv_free_sem_info(sem_info
);
2477 /* caller can override these */
2478 sem_info
->cs_emit_wait
= true;
2479 sem_info
->cs_emit_signal
= true;
2483 /* Signals fence as soon as all the work currently put on queue is done. */
2484 static VkResult
radv_signal_fence(struct radv_queue
*queue
,
2485 struct radv_fence
*fence
)
2489 struct radv_winsys_sem_info sem_info
;
2491 result
= radv_alloc_sem_info(queue
->device
->instance
, &sem_info
, 0, NULL
, 0, NULL
,
2492 radv_fence_to_handle(fence
));
2493 if (result
!= VK_SUCCESS
)
2496 ret
= queue
->device
->ws
->cs_submit(queue
->hw_ctx
, queue
->queue_idx
,
2497 &queue
->device
->empty_cs
[queue
->queue_family_index
],
2498 1, NULL
, NULL
, &sem_info
, NULL
,
2499 false, fence
->fence
);
2500 radv_free_sem_info(&sem_info
);
2503 return vk_error(queue
->device
->instance
, VK_ERROR_DEVICE_LOST
);
2508 VkResult
radv_QueueSubmit(
2510 uint32_t submitCount
,
2511 const VkSubmitInfo
* pSubmits
,
2514 RADV_FROM_HANDLE(radv_queue
, queue
, _queue
);
2515 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
2516 struct radeon_winsys_fence
*base_fence
= fence
? fence
->fence
: NULL
;
2517 struct radeon_winsys_ctx
*ctx
= queue
->hw_ctx
;
2519 uint32_t max_cs_submission
= queue
->device
->trace_bo
? 1 : UINT32_MAX
;
2520 uint32_t scratch_size
= 0;
2521 uint32_t compute_scratch_size
= 0;
2522 uint32_t esgs_ring_size
= 0, gsvs_ring_size
= 0;
2523 struct radeon_cmdbuf
*initial_preamble_cs
= NULL
, *initial_flush_preamble_cs
= NULL
, *continue_preamble_cs
= NULL
;
2525 bool fence_emitted
= false;
2526 bool tess_rings_needed
= false;
2527 bool sample_positions_needed
= false;
2529 /* Do this first so failing to allocate scratch buffers can't result in
2530 * partially executed submissions. */
2531 for (uint32_t i
= 0; i
< submitCount
; i
++) {
2532 for (uint32_t j
= 0; j
< pSubmits
[i
].commandBufferCount
; j
++) {
2533 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
,
2534 pSubmits
[i
].pCommandBuffers
[j
]);
2536 scratch_size
= MAX2(scratch_size
, cmd_buffer
->scratch_size_needed
);
2537 compute_scratch_size
= MAX2(compute_scratch_size
,
2538 cmd_buffer
->compute_scratch_size_needed
);
2539 esgs_ring_size
= MAX2(esgs_ring_size
, cmd_buffer
->esgs_ring_size_needed
);
2540 gsvs_ring_size
= MAX2(gsvs_ring_size
, cmd_buffer
->gsvs_ring_size_needed
);
2541 tess_rings_needed
|= cmd_buffer
->tess_rings_needed
;
2542 sample_positions_needed
|= cmd_buffer
->sample_positions_needed
;
2546 result
= radv_get_preamble_cs(queue
, scratch_size
, compute_scratch_size
,
2547 esgs_ring_size
, gsvs_ring_size
, tess_rings_needed
,
2548 sample_positions_needed
, &initial_flush_preamble_cs
,
2549 &initial_preamble_cs
, &continue_preamble_cs
);
2550 if (result
!= VK_SUCCESS
)
2553 for (uint32_t i
= 0; i
< submitCount
; i
++) {
2554 struct radeon_cmdbuf
**cs_array
;
2555 bool do_flush
= !i
|| pSubmits
[i
].pWaitDstStageMask
;
2556 bool can_patch
= true;
2558 struct radv_winsys_sem_info sem_info
;
2560 result
= radv_alloc_sem_info(queue
->device
->instance
,
2562 pSubmits
[i
].waitSemaphoreCount
,
2563 pSubmits
[i
].pWaitSemaphores
,
2564 pSubmits
[i
].signalSemaphoreCount
,
2565 pSubmits
[i
].pSignalSemaphores
,
2567 if (result
!= VK_SUCCESS
)
2570 if (!pSubmits
[i
].commandBufferCount
) {
2571 if (pSubmits
[i
].waitSemaphoreCount
|| pSubmits
[i
].signalSemaphoreCount
) {
2572 ret
= queue
->device
->ws
->cs_submit(ctx
, queue
->queue_idx
,
2573 &queue
->device
->empty_cs
[queue
->queue_family_index
],
2578 radv_loge("failed to submit CS %d\n", i
);
2581 fence_emitted
= true;
2583 radv_free_sem_info(&sem_info
);
2587 cs_array
= malloc(sizeof(struct radeon_cmdbuf
*) *
2588 (pSubmits
[i
].commandBufferCount
));
2590 for (uint32_t j
= 0; j
< pSubmits
[i
].commandBufferCount
; j
++) {
2591 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
,
2592 pSubmits
[i
].pCommandBuffers
[j
]);
2593 assert(cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_PRIMARY
);
2595 cs_array
[j
] = cmd_buffer
->cs
;
2596 if ((cmd_buffer
->usage_flags
& VK_COMMAND_BUFFER_USAGE_SIMULTANEOUS_USE_BIT
))
2599 cmd_buffer
->status
= RADV_CMD_BUFFER_STATUS_PENDING
;
2602 for (uint32_t j
= 0; j
< pSubmits
[i
].commandBufferCount
; j
+= advance
) {
2603 struct radeon_cmdbuf
*initial_preamble
= (do_flush
&& !j
) ? initial_flush_preamble_cs
: initial_preamble_cs
;
2604 const struct radv_winsys_bo_list
*bo_list
= NULL
;
2606 advance
= MIN2(max_cs_submission
,
2607 pSubmits
[i
].commandBufferCount
- j
);
2609 if (queue
->device
->trace_bo
)
2610 *queue
->device
->trace_id_ptr
= 0;
2612 sem_info
.cs_emit_wait
= j
== 0;
2613 sem_info
.cs_emit_signal
= j
+ advance
== pSubmits
[i
].commandBufferCount
;
2615 if (unlikely(queue
->device
->use_global_bo_list
)) {
2616 pthread_mutex_lock(&queue
->device
->bo_list
.mutex
);
2617 bo_list
= &queue
->device
->bo_list
.list
;
2620 ret
= queue
->device
->ws
->cs_submit(ctx
, queue
->queue_idx
, cs_array
+ j
,
2621 advance
, initial_preamble
, continue_preamble_cs
,
2623 can_patch
, base_fence
);
2625 if (unlikely(queue
->device
->use_global_bo_list
))
2626 pthread_mutex_unlock(&queue
->device
->bo_list
.mutex
);
2629 radv_loge("failed to submit CS %d\n", i
);
2632 fence_emitted
= true;
2633 if (queue
->device
->trace_bo
) {
2634 radv_check_gpu_hangs(queue
, cs_array
[j
]);
2638 radv_free_temp_syncobjs(queue
->device
,
2639 pSubmits
[i
].waitSemaphoreCount
,
2640 pSubmits
[i
].pWaitSemaphores
);
2641 radv_free_sem_info(&sem_info
);
2646 if (!fence_emitted
) {
2647 result
= radv_signal_fence(queue
, fence
);
2648 if (result
!= VK_SUCCESS
)
2651 fence
->submitted
= true;
2657 VkResult
radv_QueueWaitIdle(
2660 RADV_FROM_HANDLE(radv_queue
, queue
, _queue
);
2662 queue
->device
->ws
->ctx_wait_idle(queue
->hw_ctx
,
2663 radv_queue_family_to_ring(queue
->queue_family_index
),
2668 VkResult
radv_DeviceWaitIdle(
2671 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2673 for (unsigned i
= 0; i
< RADV_MAX_QUEUE_FAMILIES
; i
++) {
2674 for (unsigned q
= 0; q
< device
->queue_count
[i
]; q
++) {
2675 radv_QueueWaitIdle(radv_queue_to_handle(&device
->queues
[i
][q
]));
2681 VkResult
radv_EnumerateInstanceExtensionProperties(
2682 const char* pLayerName
,
2683 uint32_t* pPropertyCount
,
2684 VkExtensionProperties
* pProperties
)
2686 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
2688 for (int i
= 0; i
< RADV_INSTANCE_EXTENSION_COUNT
; i
++) {
2689 if (radv_supported_instance_extensions
.extensions
[i
]) {
2690 vk_outarray_append(&out
, prop
) {
2691 *prop
= radv_instance_extensions
[i
];
2696 return vk_outarray_status(&out
);
2699 VkResult
radv_EnumerateDeviceExtensionProperties(
2700 VkPhysicalDevice physicalDevice
,
2701 const char* pLayerName
,
2702 uint32_t* pPropertyCount
,
2703 VkExtensionProperties
* pProperties
)
2705 RADV_FROM_HANDLE(radv_physical_device
, device
, physicalDevice
);
2706 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
2708 for (int i
= 0; i
< RADV_DEVICE_EXTENSION_COUNT
; i
++) {
2709 if (device
->supported_extensions
.extensions
[i
]) {
2710 vk_outarray_append(&out
, prop
) {
2711 *prop
= radv_device_extensions
[i
];
2716 return vk_outarray_status(&out
);
2719 PFN_vkVoidFunction
radv_GetInstanceProcAddr(
2720 VkInstance _instance
,
2723 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
2725 return radv_lookup_entrypoint_checked(pName
,
2726 instance
? instance
->apiVersion
: 0,
2727 instance
? &instance
->enabled_extensions
: NULL
,
2731 /* The loader wants us to expose a second GetInstanceProcAddr function
2732 * to work around certain LD_PRELOAD issues seen in apps.
2735 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
2736 VkInstance instance
,
2740 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
2741 VkInstance instance
,
2744 return radv_GetInstanceProcAddr(instance
, pName
);
2747 PFN_vkVoidFunction
radv_GetDeviceProcAddr(
2751 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2753 return radv_lookup_entrypoint_checked(pName
,
2754 device
->instance
->apiVersion
,
2755 &device
->instance
->enabled_extensions
,
2756 &device
->enabled_extensions
);
2759 bool radv_get_memory_fd(struct radv_device
*device
,
2760 struct radv_device_memory
*memory
,
2763 struct radeon_bo_metadata metadata
;
2765 if (memory
->image
) {
2766 radv_init_metadata(device
, memory
->image
, &metadata
);
2767 device
->ws
->buffer_set_metadata(memory
->bo
, &metadata
);
2770 return device
->ws
->buffer_get_fd(device
->ws
, memory
->bo
,
2774 static VkResult
radv_alloc_memory(struct radv_device
*device
,
2775 const VkMemoryAllocateInfo
* pAllocateInfo
,
2776 const VkAllocationCallbacks
* pAllocator
,
2777 VkDeviceMemory
* pMem
)
2779 struct radv_device_memory
*mem
;
2781 enum radeon_bo_domain domain
;
2783 enum radv_mem_type mem_type_index
= device
->physical_device
->mem_type_indices
[pAllocateInfo
->memoryTypeIndex
];
2785 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
2787 if (pAllocateInfo
->allocationSize
== 0) {
2788 /* Apparently, this is allowed */
2789 *pMem
= VK_NULL_HANDLE
;
2793 const VkImportMemoryFdInfoKHR
*import_info
=
2794 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_FD_INFO_KHR
);
2795 const VkMemoryDedicatedAllocateInfoKHR
*dedicate_info
=
2796 vk_find_struct_const(pAllocateInfo
->pNext
, MEMORY_DEDICATED_ALLOCATE_INFO_KHR
);
2797 const VkExportMemoryAllocateInfoKHR
*export_info
=
2798 vk_find_struct_const(pAllocateInfo
->pNext
, EXPORT_MEMORY_ALLOCATE_INFO_KHR
);
2799 const VkImportMemoryHostPointerInfoEXT
*host_ptr_info
=
2800 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_HOST_POINTER_INFO_EXT
);
2802 const struct wsi_memory_allocate_info
*wsi_info
=
2803 vk_find_struct_const(pAllocateInfo
->pNext
, WSI_MEMORY_ALLOCATE_INFO_MESA
);
2805 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
2806 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2808 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
2810 if (wsi_info
&& wsi_info
->implicit_sync
)
2811 flags
|= RADEON_FLAG_IMPLICIT_SYNC
;
2813 if (dedicate_info
) {
2814 mem
->image
= radv_image_from_handle(dedicate_info
->image
);
2815 mem
->buffer
= radv_buffer_from_handle(dedicate_info
->buffer
);
2821 mem
->user_ptr
= NULL
;
2824 assert(import_info
->handleType
==
2825 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
||
2826 import_info
->handleType
==
2827 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
2828 mem
->bo
= device
->ws
->buffer_from_fd(device
->ws
, import_info
->fd
,
2831 result
= VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
;
2834 close(import_info
->fd
);
2836 } else if (host_ptr_info
) {
2837 assert(host_ptr_info
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
);
2838 assert(mem_type_index
== RADV_MEM_TYPE_GTT_CACHED
);
2839 mem
->bo
= device
->ws
->buffer_from_ptr(device
->ws
, host_ptr_info
->pHostPointer
,
2840 pAllocateInfo
->allocationSize
);
2842 result
= VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
;
2845 mem
->user_ptr
= host_ptr_info
->pHostPointer
;
2848 uint64_t alloc_size
= align_u64(pAllocateInfo
->allocationSize
, 4096);
2849 if (mem_type_index
== RADV_MEM_TYPE_GTT_WRITE_COMBINE
||
2850 mem_type_index
== RADV_MEM_TYPE_GTT_CACHED
)
2851 domain
= RADEON_DOMAIN_GTT
;
2853 domain
= RADEON_DOMAIN_VRAM
;
2855 if (mem_type_index
== RADV_MEM_TYPE_VRAM
)
2856 flags
|= RADEON_FLAG_NO_CPU_ACCESS
;
2858 flags
|= RADEON_FLAG_CPU_ACCESS
;
2860 if (mem_type_index
== RADV_MEM_TYPE_GTT_WRITE_COMBINE
)
2861 flags
|= RADEON_FLAG_GTT_WC
;
2863 if (!dedicate_info
&& !import_info
&& (!export_info
|| !export_info
->handleTypes
))
2864 flags
|= RADEON_FLAG_NO_INTERPROCESS_SHARING
;
2866 mem
->bo
= device
->ws
->buffer_create(device
->ws
, alloc_size
, device
->physical_device
->rad_info
.max_alignment
,
2870 result
= VK_ERROR_OUT_OF_DEVICE_MEMORY
;
2873 mem
->type_index
= mem_type_index
;
2876 result
= radv_bo_list_add(device
, mem
->bo
);
2877 if (result
!= VK_SUCCESS
)
2880 *pMem
= radv_device_memory_to_handle(mem
);
2885 device
->ws
->buffer_destroy(mem
->bo
);
2887 vk_free2(&device
->alloc
, pAllocator
, mem
);
2892 VkResult
radv_AllocateMemory(
2894 const VkMemoryAllocateInfo
* pAllocateInfo
,
2895 const VkAllocationCallbacks
* pAllocator
,
2896 VkDeviceMemory
* pMem
)
2898 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2899 return radv_alloc_memory(device
, pAllocateInfo
, pAllocator
, pMem
);
2902 void radv_FreeMemory(
2904 VkDeviceMemory _mem
,
2905 const VkAllocationCallbacks
* pAllocator
)
2907 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2908 RADV_FROM_HANDLE(radv_device_memory
, mem
, _mem
);
2913 radv_bo_list_remove(device
, mem
->bo
);
2914 device
->ws
->buffer_destroy(mem
->bo
);
2917 vk_free2(&device
->alloc
, pAllocator
, mem
);
2920 VkResult
radv_MapMemory(
2922 VkDeviceMemory _memory
,
2923 VkDeviceSize offset
,
2925 VkMemoryMapFlags flags
,
2928 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2929 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
2937 *ppData
= mem
->user_ptr
;
2939 *ppData
= device
->ws
->buffer_map(mem
->bo
);
2946 return vk_error(device
->instance
, VK_ERROR_MEMORY_MAP_FAILED
);
2949 void radv_UnmapMemory(
2951 VkDeviceMemory _memory
)
2953 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2954 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
2959 if (mem
->user_ptr
== NULL
)
2960 device
->ws
->buffer_unmap(mem
->bo
);
2963 VkResult
radv_FlushMappedMemoryRanges(
2965 uint32_t memoryRangeCount
,
2966 const VkMappedMemoryRange
* pMemoryRanges
)
2971 VkResult
radv_InvalidateMappedMemoryRanges(
2973 uint32_t memoryRangeCount
,
2974 const VkMappedMemoryRange
* pMemoryRanges
)
2979 void radv_GetBufferMemoryRequirements(
2982 VkMemoryRequirements
* pMemoryRequirements
)
2984 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2985 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
2987 pMemoryRequirements
->memoryTypeBits
= (1u << device
->physical_device
->memory_properties
.memoryTypeCount
) - 1;
2989 if (buffer
->flags
& VK_BUFFER_CREATE_SPARSE_BINDING_BIT
)
2990 pMemoryRequirements
->alignment
= 4096;
2992 pMemoryRequirements
->alignment
= 16;
2994 pMemoryRequirements
->size
= align64(buffer
->size
, pMemoryRequirements
->alignment
);
2997 void radv_GetBufferMemoryRequirements2(
2999 const VkBufferMemoryRequirementsInfo2KHR
* pInfo
,
3000 VkMemoryRequirements2KHR
* pMemoryRequirements
)
3002 radv_GetBufferMemoryRequirements(device
, pInfo
->buffer
,
3003 &pMemoryRequirements
->memoryRequirements
);
3004 RADV_FROM_HANDLE(radv_buffer
, buffer
, pInfo
->buffer
);
3005 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3006 switch (ext
->sType
) {
3007 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS_KHR
: {
3008 VkMemoryDedicatedRequirementsKHR
*req
=
3009 (VkMemoryDedicatedRequirementsKHR
*) ext
;
3010 req
->requiresDedicatedAllocation
= buffer
->shareable
;
3011 req
->prefersDedicatedAllocation
= req
->requiresDedicatedAllocation
;
3020 void radv_GetImageMemoryRequirements(
3023 VkMemoryRequirements
* pMemoryRequirements
)
3025 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3026 RADV_FROM_HANDLE(radv_image
, image
, _image
);
3028 pMemoryRequirements
->memoryTypeBits
= (1u << device
->physical_device
->memory_properties
.memoryTypeCount
) - 1;
3030 pMemoryRequirements
->size
= image
->size
;
3031 pMemoryRequirements
->alignment
= image
->alignment
;
3034 void radv_GetImageMemoryRequirements2(
3036 const VkImageMemoryRequirementsInfo2KHR
* pInfo
,
3037 VkMemoryRequirements2KHR
* pMemoryRequirements
)
3039 radv_GetImageMemoryRequirements(device
, pInfo
->image
,
3040 &pMemoryRequirements
->memoryRequirements
);
3042 RADV_FROM_HANDLE(radv_image
, image
, pInfo
->image
);
3044 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3045 switch (ext
->sType
) {
3046 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS_KHR
: {
3047 VkMemoryDedicatedRequirementsKHR
*req
=
3048 (VkMemoryDedicatedRequirementsKHR
*) ext
;
3049 req
->requiresDedicatedAllocation
= image
->shareable
;
3050 req
->prefersDedicatedAllocation
= req
->requiresDedicatedAllocation
;
3059 void radv_GetImageSparseMemoryRequirements(
3062 uint32_t* pSparseMemoryRequirementCount
,
3063 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
3068 void radv_GetImageSparseMemoryRequirements2(
3070 const VkImageSparseMemoryRequirementsInfo2KHR
* pInfo
,
3071 uint32_t* pSparseMemoryRequirementCount
,
3072 VkSparseImageMemoryRequirements2KHR
* pSparseMemoryRequirements
)
3077 void radv_GetDeviceMemoryCommitment(
3079 VkDeviceMemory memory
,
3080 VkDeviceSize
* pCommittedMemoryInBytes
)
3082 *pCommittedMemoryInBytes
= 0;
3085 VkResult
radv_BindBufferMemory2(VkDevice device
,
3086 uint32_t bindInfoCount
,
3087 const VkBindBufferMemoryInfoKHR
*pBindInfos
)
3089 for (uint32_t i
= 0; i
< bindInfoCount
; ++i
) {
3090 RADV_FROM_HANDLE(radv_device_memory
, mem
, pBindInfos
[i
].memory
);
3091 RADV_FROM_HANDLE(radv_buffer
, buffer
, pBindInfos
[i
].buffer
);
3094 buffer
->bo
= mem
->bo
;
3095 buffer
->offset
= pBindInfos
[i
].memoryOffset
;
3103 VkResult
radv_BindBufferMemory(
3106 VkDeviceMemory memory
,
3107 VkDeviceSize memoryOffset
)
3109 const VkBindBufferMemoryInfoKHR info
= {
3110 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO_KHR
,
3113 .memoryOffset
= memoryOffset
3116 return radv_BindBufferMemory2(device
, 1, &info
);
3119 VkResult
radv_BindImageMemory2(VkDevice device
,
3120 uint32_t bindInfoCount
,
3121 const VkBindImageMemoryInfoKHR
*pBindInfos
)
3123 for (uint32_t i
= 0; i
< bindInfoCount
; ++i
) {
3124 RADV_FROM_HANDLE(radv_device_memory
, mem
, pBindInfos
[i
].memory
);
3125 RADV_FROM_HANDLE(radv_image
, image
, pBindInfos
[i
].image
);
3128 image
->bo
= mem
->bo
;
3129 image
->offset
= pBindInfos
[i
].memoryOffset
;
3139 VkResult
radv_BindImageMemory(
3142 VkDeviceMemory memory
,
3143 VkDeviceSize memoryOffset
)
3145 const VkBindImageMemoryInfoKHR info
= {
3146 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO_KHR
,
3149 .memoryOffset
= memoryOffset
3152 return radv_BindImageMemory2(device
, 1, &info
);
3157 radv_sparse_buffer_bind_memory(struct radv_device
*device
,
3158 const VkSparseBufferMemoryBindInfo
*bind
)
3160 RADV_FROM_HANDLE(radv_buffer
, buffer
, bind
->buffer
);
3162 for (uint32_t i
= 0; i
< bind
->bindCount
; ++i
) {
3163 struct radv_device_memory
*mem
= NULL
;
3165 if (bind
->pBinds
[i
].memory
!= VK_NULL_HANDLE
)
3166 mem
= radv_device_memory_from_handle(bind
->pBinds
[i
].memory
);
3168 device
->ws
->buffer_virtual_bind(buffer
->bo
,
3169 bind
->pBinds
[i
].resourceOffset
,
3170 bind
->pBinds
[i
].size
,
3171 mem
? mem
->bo
: NULL
,
3172 bind
->pBinds
[i
].memoryOffset
);
3177 radv_sparse_image_opaque_bind_memory(struct radv_device
*device
,
3178 const VkSparseImageOpaqueMemoryBindInfo
*bind
)
3180 RADV_FROM_HANDLE(radv_image
, image
, bind
->image
);
3182 for (uint32_t i
= 0; i
< bind
->bindCount
; ++i
) {
3183 struct radv_device_memory
*mem
= NULL
;
3185 if (bind
->pBinds
[i
].memory
!= VK_NULL_HANDLE
)
3186 mem
= radv_device_memory_from_handle(bind
->pBinds
[i
].memory
);
3188 device
->ws
->buffer_virtual_bind(image
->bo
,
3189 bind
->pBinds
[i
].resourceOffset
,
3190 bind
->pBinds
[i
].size
,
3191 mem
? mem
->bo
: NULL
,
3192 bind
->pBinds
[i
].memoryOffset
);
3196 VkResult
radv_QueueBindSparse(
3198 uint32_t bindInfoCount
,
3199 const VkBindSparseInfo
* pBindInfo
,
3202 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
3203 RADV_FROM_HANDLE(radv_queue
, queue
, _queue
);
3204 struct radeon_winsys_fence
*base_fence
= fence
? fence
->fence
: NULL
;
3205 bool fence_emitted
= false;
3209 for (uint32_t i
= 0; i
< bindInfoCount
; ++i
) {
3210 struct radv_winsys_sem_info sem_info
;
3211 for (uint32_t j
= 0; j
< pBindInfo
[i
].bufferBindCount
; ++j
) {
3212 radv_sparse_buffer_bind_memory(queue
->device
,
3213 pBindInfo
[i
].pBufferBinds
+ j
);
3216 for (uint32_t j
= 0; j
< pBindInfo
[i
].imageOpaqueBindCount
; ++j
) {
3217 radv_sparse_image_opaque_bind_memory(queue
->device
,
3218 pBindInfo
[i
].pImageOpaqueBinds
+ j
);
3222 result
= radv_alloc_sem_info(queue
->device
->instance
,
3224 pBindInfo
[i
].waitSemaphoreCount
,
3225 pBindInfo
[i
].pWaitSemaphores
,
3226 pBindInfo
[i
].signalSemaphoreCount
,
3227 pBindInfo
[i
].pSignalSemaphores
,
3229 if (result
!= VK_SUCCESS
)
3232 if (pBindInfo
[i
].waitSemaphoreCount
|| pBindInfo
[i
].signalSemaphoreCount
) {
3233 ret
= queue
->device
->ws
->cs_submit(queue
->hw_ctx
, queue
->queue_idx
,
3234 &queue
->device
->empty_cs
[queue
->queue_family_index
],
3239 radv_loge("failed to submit CS %d\n", i
);
3243 fence_emitted
= true;
3245 fence
->submitted
= true;
3248 radv_free_sem_info(&sem_info
);
3253 if (!fence_emitted
) {
3254 result
= radv_signal_fence(queue
, fence
);
3255 if (result
!= VK_SUCCESS
)
3258 fence
->submitted
= true;
3264 VkResult
radv_CreateFence(
3266 const VkFenceCreateInfo
* pCreateInfo
,
3267 const VkAllocationCallbacks
* pAllocator
,
3270 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3271 const VkExportFenceCreateInfoKHR
*export
=
3272 vk_find_struct_const(pCreateInfo
->pNext
, EXPORT_FENCE_CREATE_INFO_KHR
);
3273 VkExternalFenceHandleTypeFlagsKHR handleTypes
=
3274 export
? export
->handleTypes
: 0;
3276 struct radv_fence
*fence
= vk_alloc2(&device
->alloc
, pAllocator
,
3278 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3281 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3283 fence
->fence_wsi
= NULL
;
3284 fence
->submitted
= false;
3285 fence
->signalled
= !!(pCreateInfo
->flags
& VK_FENCE_CREATE_SIGNALED_BIT
);
3286 fence
->temp_syncobj
= 0;
3287 if (device
->always_use_syncobj
|| handleTypes
) {
3288 int ret
= device
->ws
->create_syncobj(device
->ws
, &fence
->syncobj
);
3290 vk_free2(&device
->alloc
, pAllocator
, fence
);
3291 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3293 if (pCreateInfo
->flags
& VK_FENCE_CREATE_SIGNALED_BIT
) {
3294 device
->ws
->signal_syncobj(device
->ws
, fence
->syncobj
);
3296 fence
->fence
= NULL
;
3298 fence
->fence
= device
->ws
->create_fence();
3299 if (!fence
->fence
) {
3300 vk_free2(&device
->alloc
, pAllocator
, fence
);
3301 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3306 *pFence
= radv_fence_to_handle(fence
);
3311 void radv_DestroyFence(
3314 const VkAllocationCallbacks
* pAllocator
)
3316 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3317 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
3322 if (fence
->temp_syncobj
)
3323 device
->ws
->destroy_syncobj(device
->ws
, fence
->temp_syncobj
);
3325 device
->ws
->destroy_syncobj(device
->ws
, fence
->syncobj
);
3327 device
->ws
->destroy_fence(fence
->fence
);
3328 if (fence
->fence_wsi
)
3329 fence
->fence_wsi
->destroy(fence
->fence_wsi
);
3330 vk_free2(&device
->alloc
, pAllocator
, fence
);
3334 static uint64_t radv_get_current_time()
3337 clock_gettime(CLOCK_MONOTONIC
, &tv
);
3338 return tv
.tv_nsec
+ tv
.tv_sec
*1000000000ull;
3341 static uint64_t radv_get_absolute_timeout(uint64_t timeout
)
3343 uint64_t current_time
= radv_get_current_time();
3345 timeout
= MIN2(UINT64_MAX
- current_time
, timeout
);
3347 return current_time
+ timeout
;
3351 static bool radv_all_fences_plain_and_submitted(uint32_t fenceCount
, const VkFence
*pFences
)
3353 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
3354 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
3355 if (fence
->fence
== NULL
|| fence
->syncobj
||
3356 fence
->temp_syncobj
||
3357 (!fence
->signalled
&& !fence
->submitted
))
3363 static bool radv_all_fences_syncobj(uint32_t fenceCount
, const VkFence
*pFences
)
3365 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
3366 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
3367 if (fence
->syncobj
== 0 && fence
->temp_syncobj
== 0)
3373 VkResult
radv_WaitForFences(
3375 uint32_t fenceCount
,
3376 const VkFence
* pFences
,
3380 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3381 timeout
= radv_get_absolute_timeout(timeout
);
3383 if (device
->always_use_syncobj
&&
3384 radv_all_fences_syncobj(fenceCount
, pFences
))
3386 uint32_t *handles
= malloc(sizeof(uint32_t) * fenceCount
);
3388 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3390 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
3391 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
3392 handles
[i
] = fence
->temp_syncobj
? fence
->temp_syncobj
: fence
->syncobj
;
3395 bool success
= device
->ws
->wait_syncobj(device
->ws
, handles
, fenceCount
, waitAll
, timeout
);
3398 return success
? VK_SUCCESS
: VK_TIMEOUT
;
3401 if (!waitAll
&& fenceCount
> 1) {
3402 /* Not doing this by default for waitAll, due to needing to allocate twice. */
3403 if (device
->physical_device
->rad_info
.drm_minor
>= 10 && radv_all_fences_plain_and_submitted(fenceCount
, pFences
)) {
3404 uint32_t wait_count
= 0;
3405 struct radeon_winsys_fence
**fences
= malloc(sizeof(struct radeon_winsys_fence
*) * fenceCount
);
3407 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3409 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
3410 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
3412 if (fence
->signalled
) {
3417 fences
[wait_count
++] = fence
->fence
;
3420 bool success
= device
->ws
->fences_wait(device
->ws
, fences
, wait_count
,
3421 waitAll
, timeout
- radv_get_current_time());
3424 return success
? VK_SUCCESS
: VK_TIMEOUT
;
3427 while(radv_get_current_time() <= timeout
) {
3428 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
3429 if (radv_GetFenceStatus(_device
, pFences
[i
]) == VK_SUCCESS
)
3436 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
3437 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
3438 bool expired
= false;
3440 if (fence
->temp_syncobj
) {
3441 if (!device
->ws
->wait_syncobj(device
->ws
, &fence
->temp_syncobj
, 1, true, timeout
))
3446 if (fence
->syncobj
) {
3447 if (!device
->ws
->wait_syncobj(device
->ws
, &fence
->syncobj
, 1, true, timeout
))
3452 if (fence
->signalled
)
3456 if (!fence
->submitted
) {
3457 while(radv_get_current_time() <= timeout
&&
3461 if (!fence
->submitted
)
3464 /* Recheck as it may have been set by
3465 * submitting operations. */
3467 if (fence
->signalled
)
3471 expired
= device
->ws
->fence_wait(device
->ws
,
3478 if (fence
->fence_wsi
) {
3479 VkResult result
= fence
->fence_wsi
->wait(fence
->fence_wsi
, timeout
);
3480 if (result
!= VK_SUCCESS
)
3484 fence
->signalled
= true;
3490 VkResult
radv_ResetFences(VkDevice _device
,
3491 uint32_t fenceCount
,
3492 const VkFence
*pFences
)
3494 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3496 for (unsigned i
= 0; i
< fenceCount
; ++i
) {
3497 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
3498 fence
->submitted
= fence
->signalled
= false;
3500 /* Per spec, we first restore the permanent payload, and then reset, so
3501 * having a temp syncobj should not skip resetting the permanent syncobj. */
3502 if (fence
->temp_syncobj
) {
3503 device
->ws
->destroy_syncobj(device
->ws
, fence
->temp_syncobj
);
3504 fence
->temp_syncobj
= 0;
3507 if (fence
->syncobj
) {
3508 device
->ws
->reset_syncobj(device
->ws
, fence
->syncobj
);
3515 VkResult
radv_GetFenceStatus(VkDevice _device
, VkFence _fence
)
3517 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3518 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
3520 if (fence
->temp_syncobj
) {
3521 bool success
= device
->ws
->wait_syncobj(device
->ws
, &fence
->temp_syncobj
, 1, true, 0);
3522 return success
? VK_SUCCESS
: VK_NOT_READY
;
3525 if (fence
->syncobj
) {
3526 bool success
= device
->ws
->wait_syncobj(device
->ws
, &fence
->syncobj
, 1, true, 0);
3527 return success
? VK_SUCCESS
: VK_NOT_READY
;
3530 if (fence
->signalled
)
3532 if (!fence
->submitted
)
3533 return VK_NOT_READY
;
3535 if (!device
->ws
->fence_wait(device
->ws
, fence
->fence
, false, 0))
3536 return VK_NOT_READY
;
3538 if (fence
->fence_wsi
) {
3539 VkResult result
= fence
->fence_wsi
->wait(fence
->fence_wsi
, 0);
3541 if (result
!= VK_SUCCESS
) {
3542 if (result
== VK_TIMEOUT
)
3543 return VK_NOT_READY
;
3551 // Queue semaphore functions
3553 VkResult
radv_CreateSemaphore(
3555 const VkSemaphoreCreateInfo
* pCreateInfo
,
3556 const VkAllocationCallbacks
* pAllocator
,
3557 VkSemaphore
* pSemaphore
)
3559 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3560 const VkExportSemaphoreCreateInfoKHR
*export
=
3561 vk_find_struct_const(pCreateInfo
->pNext
, EXPORT_SEMAPHORE_CREATE_INFO_KHR
);
3562 VkExternalSemaphoreHandleTypeFlagsKHR handleTypes
=
3563 export
? export
->handleTypes
: 0;
3565 struct radv_semaphore
*sem
= vk_alloc2(&device
->alloc
, pAllocator
,
3567 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3569 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3571 sem
->temp_syncobj
= 0;
3572 /* create a syncobject if we are going to export this semaphore */
3573 if (device
->always_use_syncobj
|| handleTypes
) {
3574 assert (device
->physical_device
->rad_info
.has_syncobj
);
3575 int ret
= device
->ws
->create_syncobj(device
->ws
, &sem
->syncobj
);
3577 vk_free2(&device
->alloc
, pAllocator
, sem
);
3578 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3582 sem
->sem
= device
->ws
->create_sem(device
->ws
);
3584 vk_free2(&device
->alloc
, pAllocator
, sem
);
3585 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3590 *pSemaphore
= radv_semaphore_to_handle(sem
);
3594 void radv_DestroySemaphore(
3596 VkSemaphore _semaphore
,
3597 const VkAllocationCallbacks
* pAllocator
)
3599 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3600 RADV_FROM_HANDLE(radv_semaphore
, sem
, _semaphore
);
3605 device
->ws
->destroy_syncobj(device
->ws
, sem
->syncobj
);
3607 device
->ws
->destroy_sem(sem
->sem
);
3608 vk_free2(&device
->alloc
, pAllocator
, sem
);
3611 VkResult
radv_CreateEvent(
3613 const VkEventCreateInfo
* pCreateInfo
,
3614 const VkAllocationCallbacks
* pAllocator
,
3617 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3618 struct radv_event
*event
= vk_alloc2(&device
->alloc
, pAllocator
,
3620 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3623 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3625 event
->bo
= device
->ws
->buffer_create(device
->ws
, 8, 8,
3627 RADEON_FLAG_VA_UNCACHED
| RADEON_FLAG_CPU_ACCESS
| RADEON_FLAG_NO_INTERPROCESS_SHARING
);
3629 vk_free2(&device
->alloc
, pAllocator
, event
);
3630 return vk_error(device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
3633 event
->map
= (uint64_t*)device
->ws
->buffer_map(event
->bo
);
3635 *pEvent
= radv_event_to_handle(event
);
3640 void radv_DestroyEvent(
3643 const VkAllocationCallbacks
* pAllocator
)
3645 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3646 RADV_FROM_HANDLE(radv_event
, event
, _event
);
3650 device
->ws
->buffer_destroy(event
->bo
);
3651 vk_free2(&device
->alloc
, pAllocator
, event
);
3654 VkResult
radv_GetEventStatus(
3658 RADV_FROM_HANDLE(radv_event
, event
, _event
);
3660 if (*event
->map
== 1)
3661 return VK_EVENT_SET
;
3662 return VK_EVENT_RESET
;
3665 VkResult
radv_SetEvent(
3669 RADV_FROM_HANDLE(radv_event
, event
, _event
);
3675 VkResult
radv_ResetEvent(
3679 RADV_FROM_HANDLE(radv_event
, event
, _event
);
3685 VkResult
radv_CreateBuffer(
3687 const VkBufferCreateInfo
* pCreateInfo
,
3688 const VkAllocationCallbacks
* pAllocator
,
3691 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3692 struct radv_buffer
*buffer
;
3694 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
3696 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
3697 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3699 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3701 buffer
->size
= pCreateInfo
->size
;
3702 buffer
->usage
= pCreateInfo
->usage
;
3705 buffer
->flags
= pCreateInfo
->flags
;
3707 buffer
->shareable
= vk_find_struct_const(pCreateInfo
->pNext
,
3708 EXTERNAL_MEMORY_BUFFER_CREATE_INFO_KHR
) != NULL
;
3710 if (pCreateInfo
->flags
& VK_BUFFER_CREATE_SPARSE_BINDING_BIT
) {
3711 buffer
->bo
= device
->ws
->buffer_create(device
->ws
,
3712 align64(buffer
->size
, 4096),
3713 4096, 0, RADEON_FLAG_VIRTUAL
);
3715 vk_free2(&device
->alloc
, pAllocator
, buffer
);
3716 return vk_error(device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
3720 *pBuffer
= radv_buffer_to_handle(buffer
);
3725 void radv_DestroyBuffer(
3728 const VkAllocationCallbacks
* pAllocator
)
3730 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3731 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
3736 if (buffer
->flags
& VK_BUFFER_CREATE_SPARSE_BINDING_BIT
)
3737 device
->ws
->buffer_destroy(buffer
->bo
);
3739 vk_free2(&device
->alloc
, pAllocator
, buffer
);
3742 static inline unsigned
3743 si_tile_mode_index(const struct radv_image
*image
, unsigned level
, bool stencil
)
3746 return image
->surface
.u
.legacy
.stencil_tiling_index
[level
];
3748 return image
->surface
.u
.legacy
.tiling_index
[level
];
3751 static uint32_t radv_surface_max_layer_count(struct radv_image_view
*iview
)
3753 return iview
->type
== VK_IMAGE_VIEW_TYPE_3D
? iview
->extent
.depth
: (iview
->base_layer
+ iview
->layer_count
);
3757 radv_init_dcc_control_reg(struct radv_device
*device
,
3758 struct radv_image_view
*iview
)
3760 unsigned max_uncompressed_block_size
= V_028C78_MAX_BLOCK_SIZE_256B
;
3761 unsigned min_compressed_block_size
= V_028C78_MIN_BLOCK_SIZE_32B
;
3762 unsigned max_compressed_block_size
;
3763 unsigned independent_64b_blocks
;
3765 if (device
->physical_device
->rad_info
.chip_class
< VI
)
3768 if (iview
->image
->info
.samples
> 1) {
3769 if (iview
->image
->surface
.bpe
== 1)
3770 max_uncompressed_block_size
= V_028C78_MAX_BLOCK_SIZE_64B
;
3771 else if (iview
->image
->surface
.bpe
== 2)
3772 max_uncompressed_block_size
= V_028C78_MAX_BLOCK_SIZE_128B
;
3775 if (!device
->physical_device
->rad_info
.has_dedicated_vram
) {
3776 /* amdvlk: [min-compressed-block-size] should be set to 32 for
3777 * dGPU and 64 for APU because all of our APUs to date use
3778 * DIMMs which have a request granularity size of 64B while all
3779 * other chips have a 32B request size.
3781 min_compressed_block_size
= V_028C78_MIN_BLOCK_SIZE_64B
;
3784 if (iview
->image
->usage
& (VK_IMAGE_USAGE_SAMPLED_BIT
|
3785 VK_IMAGE_USAGE_TRANSFER_SRC_BIT
|
3786 VK_IMAGE_USAGE_INPUT_ATTACHMENT_BIT
)) {
3787 /* If this DCC image is potentially going to be used in texture
3788 * fetches, we need some special settings.
3790 independent_64b_blocks
= 1;
3791 max_compressed_block_size
= V_028C78_MAX_BLOCK_SIZE_64B
;
3793 /* MAX_UNCOMPRESSED_BLOCK_SIZE must be >=
3794 * MAX_COMPRESSED_BLOCK_SIZE. Set MAX_COMPRESSED_BLOCK_SIZE as
3795 * big as possible for better compression state.
3797 independent_64b_blocks
= 0;
3798 max_compressed_block_size
= max_uncompressed_block_size
;
3801 return S_028C78_MAX_UNCOMPRESSED_BLOCK_SIZE(max_uncompressed_block_size
) |
3802 S_028C78_MAX_COMPRESSED_BLOCK_SIZE(max_compressed_block_size
) |
3803 S_028C78_MIN_COMPRESSED_BLOCK_SIZE(min_compressed_block_size
) |
3804 S_028C78_INDEPENDENT_64B_BLOCKS(independent_64b_blocks
);
3808 radv_initialise_color_surface(struct radv_device
*device
,
3809 struct radv_color_buffer_info
*cb
,
3810 struct radv_image_view
*iview
)
3812 const struct vk_format_description
*desc
;
3813 unsigned ntype
, format
, swap
, endian
;
3814 unsigned blend_clamp
= 0, blend_bypass
= 0;
3816 const struct radeon_surf
*surf
= &iview
->image
->surface
;
3818 desc
= vk_format_description(iview
->vk_format
);
3820 memset(cb
, 0, sizeof(*cb
));
3822 /* Intensity is implemented as Red, so treat it that way. */
3823 cb
->cb_color_attrib
= S_028C74_FORCE_DST_ALPHA_1(desc
->swizzle
[3] == VK_SWIZZLE_1
);
3825 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
;
3827 cb
->cb_color_base
= va
>> 8;
3829 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
3830 struct gfx9_surf_meta_flags meta
;
3831 if (iview
->image
->dcc_offset
)
3832 meta
= iview
->image
->surface
.u
.gfx9
.dcc
;
3834 meta
= iview
->image
->surface
.u
.gfx9
.cmask
;
3836 cb
->cb_color_attrib
|= S_028C74_COLOR_SW_MODE(iview
->image
->surface
.u
.gfx9
.surf
.swizzle_mode
) |
3837 S_028C74_FMASK_SW_MODE(iview
->image
->surface
.u
.gfx9
.fmask
.swizzle_mode
) |
3838 S_028C74_RB_ALIGNED(meta
.rb_aligned
) |
3839 S_028C74_PIPE_ALIGNED(meta
.pipe_aligned
);
3841 cb
->cb_color_base
+= iview
->image
->surface
.u
.gfx9
.surf_offset
>> 8;
3842 cb
->cb_color_base
|= iview
->image
->surface
.tile_swizzle
;
3844 const struct legacy_surf_level
*level_info
= &surf
->u
.legacy
.level
[iview
->base_mip
];
3845 unsigned pitch_tile_max
, slice_tile_max
, tile_mode_index
;
3847 cb
->cb_color_base
+= level_info
->offset
>> 8;
3848 if (level_info
->mode
== RADEON_SURF_MODE_2D
)
3849 cb
->cb_color_base
|= iview
->image
->surface
.tile_swizzle
;
3851 pitch_tile_max
= level_info
->nblk_x
/ 8 - 1;
3852 slice_tile_max
= (level_info
->nblk_x
* level_info
->nblk_y
) / 64 - 1;
3853 tile_mode_index
= si_tile_mode_index(iview
->image
, iview
->base_mip
, false);
3855 cb
->cb_color_pitch
= S_028C64_TILE_MAX(pitch_tile_max
);
3856 cb
->cb_color_slice
= S_028C68_TILE_MAX(slice_tile_max
);
3857 cb
->cb_color_cmask_slice
= iview
->image
->cmask
.slice_tile_max
;
3859 cb
->cb_color_attrib
|= S_028C74_TILE_MODE_INDEX(tile_mode_index
);
3861 if (radv_image_has_fmask(iview
->image
)) {
3862 if (device
->physical_device
->rad_info
.chip_class
>= CIK
)
3863 cb
->cb_color_pitch
|= S_028C64_FMASK_TILE_MAX(iview
->image
->fmask
.pitch_in_pixels
/ 8 - 1);
3864 cb
->cb_color_attrib
|= S_028C74_FMASK_TILE_MODE_INDEX(iview
->image
->fmask
.tile_mode_index
);
3865 cb
->cb_color_fmask_slice
= S_028C88_TILE_MAX(iview
->image
->fmask
.slice_tile_max
);
3867 /* This must be set for fast clear to work without FMASK. */
3868 if (device
->physical_device
->rad_info
.chip_class
>= CIK
)
3869 cb
->cb_color_pitch
|= S_028C64_FMASK_TILE_MAX(pitch_tile_max
);
3870 cb
->cb_color_attrib
|= S_028C74_FMASK_TILE_MODE_INDEX(tile_mode_index
);
3871 cb
->cb_color_fmask_slice
= S_028C88_TILE_MAX(slice_tile_max
);
3875 /* CMASK variables */
3876 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
;
3877 va
+= iview
->image
->cmask
.offset
;
3878 cb
->cb_color_cmask
= va
>> 8;
3880 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
;
3881 va
+= iview
->image
->dcc_offset
;
3882 cb
->cb_dcc_base
= va
>> 8;
3883 cb
->cb_dcc_base
|= iview
->image
->surface
.tile_swizzle
;
3885 uint32_t max_slice
= radv_surface_max_layer_count(iview
) - 1;
3886 cb
->cb_color_view
= S_028C6C_SLICE_START(iview
->base_layer
) |
3887 S_028C6C_SLICE_MAX(max_slice
);
3889 if (iview
->image
->info
.samples
> 1) {
3890 unsigned log_samples
= util_logbase2(iview
->image
->info
.samples
);
3892 cb
->cb_color_attrib
|= S_028C74_NUM_SAMPLES(log_samples
) |
3893 S_028C74_NUM_FRAGMENTS(log_samples
);
3896 if (radv_image_has_fmask(iview
->image
)) {
3897 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
+ iview
->image
->fmask
.offset
;
3898 cb
->cb_color_fmask
= va
>> 8;
3899 cb
->cb_color_fmask
|= iview
->image
->fmask
.tile_swizzle
;
3901 cb
->cb_color_fmask
= cb
->cb_color_base
;
3904 ntype
= radv_translate_color_numformat(iview
->vk_format
,
3906 vk_format_get_first_non_void_channel(iview
->vk_format
));
3907 format
= radv_translate_colorformat(iview
->vk_format
);
3908 if (format
== V_028C70_COLOR_INVALID
|| ntype
== ~0u)
3909 radv_finishme("Illegal color\n");
3910 swap
= radv_translate_colorswap(iview
->vk_format
, FALSE
);
3911 endian
= radv_colorformat_endian_swap(format
);
3913 /* blend clamp should be set for all NORM/SRGB types */
3914 if (ntype
== V_028C70_NUMBER_UNORM
||
3915 ntype
== V_028C70_NUMBER_SNORM
||
3916 ntype
== V_028C70_NUMBER_SRGB
)
3919 /* set blend bypass according to docs if SINT/UINT or
3920 8/24 COLOR variants */
3921 if (ntype
== V_028C70_NUMBER_UINT
|| ntype
== V_028C70_NUMBER_SINT
||
3922 format
== V_028C70_COLOR_8_24
|| format
== V_028C70_COLOR_24_8
||
3923 format
== V_028C70_COLOR_X24_8_32_FLOAT
) {
3928 if ((ntype
== V_028C70_NUMBER_UINT
|| ntype
== V_028C70_NUMBER_SINT
) &&
3929 (format
== V_028C70_COLOR_8
||
3930 format
== V_028C70_COLOR_8_8
||
3931 format
== V_028C70_COLOR_8_8_8_8
))
3932 ->color_is_int8
= true;
3934 cb
->cb_color_info
= S_028C70_FORMAT(format
) |
3935 S_028C70_COMP_SWAP(swap
) |
3936 S_028C70_BLEND_CLAMP(blend_clamp
) |
3937 S_028C70_BLEND_BYPASS(blend_bypass
) |
3938 S_028C70_SIMPLE_FLOAT(1) |
3939 S_028C70_ROUND_MODE(ntype
!= V_028C70_NUMBER_UNORM
&&
3940 ntype
!= V_028C70_NUMBER_SNORM
&&
3941 ntype
!= V_028C70_NUMBER_SRGB
&&
3942 format
!= V_028C70_COLOR_8_24
&&
3943 format
!= V_028C70_COLOR_24_8
) |
3944 S_028C70_NUMBER_TYPE(ntype
) |
3945 S_028C70_ENDIAN(endian
);
3946 if (radv_image_has_fmask(iview
->image
)) {
3947 cb
->cb_color_info
|= S_028C70_COMPRESSION(1);
3948 if (device
->physical_device
->rad_info
.chip_class
== SI
) {
3949 unsigned fmask_bankh
= util_logbase2(iview
->image
->fmask
.bank_height
);
3950 cb
->cb_color_attrib
|= S_028C74_FMASK_BANK_HEIGHT(fmask_bankh
);
3954 if (radv_image_has_cmask(iview
->image
) &&
3955 !(device
->instance
->debug_flags
& RADV_DEBUG_NO_FAST_CLEARS
))
3956 cb
->cb_color_info
|= S_028C70_FAST_CLEAR(1);
3958 if (radv_dcc_enabled(iview
->image
, iview
->base_mip
))
3959 cb
->cb_color_info
|= S_028C70_DCC_ENABLE(1);
3961 cb
->cb_dcc_control
= radv_init_dcc_control_reg(device
, iview
);
3963 /* This must be set for fast clear to work without FMASK. */
3964 if (!radv_image_has_fmask(iview
->image
) &&
3965 device
->physical_device
->rad_info
.chip_class
== SI
) {
3966 unsigned bankh
= util_logbase2(iview
->image
->surface
.u
.legacy
.bankh
);
3967 cb
->cb_color_attrib
|= S_028C74_FMASK_BANK_HEIGHT(bankh
);
3970 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
3971 unsigned mip0_depth
= iview
->image
->type
== VK_IMAGE_TYPE_3D
?
3972 (iview
->extent
.depth
- 1) : (iview
->image
->info
.array_size
- 1);
3974 cb
->cb_color_view
|= S_028C6C_MIP_LEVEL(iview
->base_mip
);
3975 cb
->cb_color_attrib
|= S_028C74_MIP0_DEPTH(mip0_depth
) |
3976 S_028C74_RESOURCE_TYPE(iview
->image
->surface
.u
.gfx9
.resource_type
);
3977 cb
->cb_color_attrib2
= S_028C68_MIP0_WIDTH(iview
->extent
.width
- 1) |
3978 S_028C68_MIP0_HEIGHT(iview
->extent
.height
- 1) |
3979 S_028C68_MAX_MIP(iview
->image
->info
.levels
- 1);
3984 radv_calc_decompress_on_z_planes(struct radv_device
*device
,
3985 struct radv_image_view
*iview
)
3987 unsigned max_zplanes
= 0;
3989 assert(radv_image_is_tc_compat_htile(iview
->image
));
3991 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
3992 /* Default value for 32-bit depth surfaces. */
3995 if (iview
->vk_format
== VK_FORMAT_D16_UNORM
&&
3996 iview
->image
->info
.samples
> 1)
3999 max_zplanes
= max_zplanes
+ 1;
4001 if (iview
->vk_format
== VK_FORMAT_D16_UNORM
) {
4002 /* Do not enable Z plane compression for 16-bit depth
4003 * surfaces because isn't supported on GFX8. Only
4004 * 32-bit depth surfaces are supported by the hardware.
4005 * This allows to maintain shader compatibility and to
4006 * reduce the number of depth decompressions.
4010 if (iview
->image
->info
.samples
<= 1)
4012 else if (iview
->image
->info
.samples
<= 4)
4023 radv_initialise_ds_surface(struct radv_device
*device
,
4024 struct radv_ds_buffer_info
*ds
,
4025 struct radv_image_view
*iview
)
4027 unsigned level
= iview
->base_mip
;
4028 unsigned format
, stencil_format
;
4029 uint64_t va
, s_offs
, z_offs
;
4030 bool stencil_only
= false;
4031 memset(ds
, 0, sizeof(*ds
));
4032 switch (iview
->image
->vk_format
) {
4033 case VK_FORMAT_D24_UNORM_S8_UINT
:
4034 case VK_FORMAT_X8_D24_UNORM_PACK32
:
4035 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-24);
4036 ds
->offset_scale
= 2.0f
;
4038 case VK_FORMAT_D16_UNORM
:
4039 case VK_FORMAT_D16_UNORM_S8_UINT
:
4040 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-16);
4041 ds
->offset_scale
= 4.0f
;
4043 case VK_FORMAT_D32_SFLOAT
:
4044 case VK_FORMAT_D32_SFLOAT_S8_UINT
:
4045 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-23) |
4046 S_028B78_POLY_OFFSET_DB_IS_FLOAT_FMT(1);
4047 ds
->offset_scale
= 1.0f
;
4049 case VK_FORMAT_S8_UINT
:
4050 stencil_only
= true;
4056 format
= radv_translate_dbformat(iview
->image
->vk_format
);
4057 stencil_format
= iview
->image
->surface
.has_stencil
?
4058 V_028044_STENCIL_8
: V_028044_STENCIL_INVALID
;
4060 uint32_t max_slice
= radv_surface_max_layer_count(iview
) - 1;
4061 ds
->db_depth_view
= S_028008_SLICE_START(iview
->base_layer
) |
4062 S_028008_SLICE_MAX(max_slice
);
4064 ds
->db_htile_data_base
= 0;
4065 ds
->db_htile_surface
= 0;
4067 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
;
4068 s_offs
= z_offs
= va
;
4070 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
4071 assert(iview
->image
->surface
.u
.gfx9
.surf_offset
== 0);
4072 s_offs
+= iview
->image
->surface
.u
.gfx9
.stencil_offset
;
4074 ds
->db_z_info
= S_028038_FORMAT(format
) |
4075 S_028038_NUM_SAMPLES(util_logbase2(iview
->image
->info
.samples
)) |
4076 S_028038_SW_MODE(iview
->image
->surface
.u
.gfx9
.surf
.swizzle_mode
) |
4077 S_028038_MAXMIP(iview
->image
->info
.levels
- 1) |
4078 S_028038_ZRANGE_PRECISION(1);
4079 ds
->db_stencil_info
= S_02803C_FORMAT(stencil_format
) |
4080 S_02803C_SW_MODE(iview
->image
->surface
.u
.gfx9
.stencil
.swizzle_mode
);
4082 ds
->db_z_info2
= S_028068_EPITCH(iview
->image
->surface
.u
.gfx9
.surf
.epitch
);
4083 ds
->db_stencil_info2
= S_02806C_EPITCH(iview
->image
->surface
.u
.gfx9
.stencil
.epitch
);
4084 ds
->db_depth_view
|= S_028008_MIPID(level
);
4086 ds
->db_depth_size
= S_02801C_X_MAX(iview
->image
->info
.width
- 1) |
4087 S_02801C_Y_MAX(iview
->image
->info
.height
- 1);
4089 if (radv_htile_enabled(iview
->image
, level
)) {
4090 ds
->db_z_info
|= S_028038_TILE_SURFACE_ENABLE(1);
4092 if (radv_image_is_tc_compat_htile(iview
->image
)) {
4093 unsigned max_zplanes
=
4094 radv_calc_decompress_on_z_planes(device
, iview
);
4096 ds
->db_z_info
|= S_028038_DECOMPRESS_ON_N_ZPLANES(max_zplanes
) |
4097 S_028038_ITERATE_FLUSH(1);
4098 ds
->db_stencil_info
|= S_02803C_ITERATE_FLUSH(1);
4101 if (!iview
->image
->surface
.has_stencil
)
4102 /* Use all of the htile_buffer for depth if there's no stencil. */
4103 ds
->db_stencil_info
|= S_02803C_TILE_STENCIL_DISABLE(1);
4104 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
+
4105 iview
->image
->htile_offset
;
4106 ds
->db_htile_data_base
= va
>> 8;
4107 ds
->db_htile_surface
= S_028ABC_FULL_CACHE(1) |
4108 S_028ABC_PIPE_ALIGNED(iview
->image
->surface
.u
.gfx9
.htile
.pipe_aligned
) |
4109 S_028ABC_RB_ALIGNED(iview
->image
->surface
.u
.gfx9
.htile
.rb_aligned
);
4112 const struct legacy_surf_level
*level_info
= &iview
->image
->surface
.u
.legacy
.level
[level
];
4115 level_info
= &iview
->image
->surface
.u
.legacy
.stencil_level
[level
];
4117 z_offs
+= iview
->image
->surface
.u
.legacy
.level
[level
].offset
;
4118 s_offs
+= iview
->image
->surface
.u
.legacy
.stencil_level
[level
].offset
;
4120 ds
->db_depth_info
= S_02803C_ADDR5_SWIZZLE_MASK(!radv_image_is_tc_compat_htile(iview
->image
));
4121 ds
->db_z_info
= S_028040_FORMAT(format
) | S_028040_ZRANGE_PRECISION(1);
4122 ds
->db_stencil_info
= S_028044_FORMAT(stencil_format
);
4124 if (iview
->image
->info
.samples
> 1)
4125 ds
->db_z_info
|= S_028040_NUM_SAMPLES(util_logbase2(iview
->image
->info
.samples
));
4127 if (device
->physical_device
->rad_info
.chip_class
>= CIK
) {
4128 struct radeon_info
*info
= &device
->physical_device
->rad_info
;
4129 unsigned tiling_index
= iview
->image
->surface
.u
.legacy
.tiling_index
[level
];
4130 unsigned stencil_index
= iview
->image
->surface
.u
.legacy
.stencil_tiling_index
[level
];
4131 unsigned macro_index
= iview
->image
->surface
.u
.legacy
.macro_tile_index
;
4132 unsigned tile_mode
= info
->si_tile_mode_array
[tiling_index
];
4133 unsigned stencil_tile_mode
= info
->si_tile_mode_array
[stencil_index
];
4134 unsigned macro_mode
= info
->cik_macrotile_mode_array
[macro_index
];
4137 tile_mode
= stencil_tile_mode
;
4139 ds
->db_depth_info
|=
4140 S_02803C_ARRAY_MODE(G_009910_ARRAY_MODE(tile_mode
)) |
4141 S_02803C_PIPE_CONFIG(G_009910_PIPE_CONFIG(tile_mode
)) |
4142 S_02803C_BANK_WIDTH(G_009990_BANK_WIDTH(macro_mode
)) |
4143 S_02803C_BANK_HEIGHT(G_009990_BANK_HEIGHT(macro_mode
)) |
4144 S_02803C_MACRO_TILE_ASPECT(G_009990_MACRO_TILE_ASPECT(macro_mode
)) |
4145 S_02803C_NUM_BANKS(G_009990_NUM_BANKS(macro_mode
));
4146 ds
->db_z_info
|= S_028040_TILE_SPLIT(G_009910_TILE_SPLIT(tile_mode
));
4147 ds
->db_stencil_info
|= S_028044_TILE_SPLIT(G_009910_TILE_SPLIT(stencil_tile_mode
));
4149 unsigned tile_mode_index
= si_tile_mode_index(iview
->image
, level
, false);
4150 ds
->db_z_info
|= S_028040_TILE_MODE_INDEX(tile_mode_index
);
4151 tile_mode_index
= si_tile_mode_index(iview
->image
, level
, true);
4152 ds
->db_stencil_info
|= S_028044_TILE_MODE_INDEX(tile_mode_index
);
4154 ds
->db_z_info
|= S_028040_TILE_MODE_INDEX(tile_mode_index
);
4157 ds
->db_depth_size
= S_028058_PITCH_TILE_MAX((level_info
->nblk_x
/ 8) - 1) |
4158 S_028058_HEIGHT_TILE_MAX((level_info
->nblk_y
/ 8) - 1);
4159 ds
->db_depth_slice
= S_02805C_SLICE_TILE_MAX((level_info
->nblk_x
* level_info
->nblk_y
) / 64 - 1);
4161 if (radv_htile_enabled(iview
->image
, level
)) {
4162 ds
->db_z_info
|= S_028040_TILE_SURFACE_ENABLE(1);
4164 if (!iview
->image
->surface
.has_stencil
&&
4165 !radv_image_is_tc_compat_htile(iview
->image
))
4166 /* Use all of the htile_buffer for depth if there's no stencil. */
4167 ds
->db_stencil_info
|= S_028044_TILE_STENCIL_DISABLE(1);
4169 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
+
4170 iview
->image
->htile_offset
;
4171 ds
->db_htile_data_base
= va
>> 8;
4172 ds
->db_htile_surface
= S_028ABC_FULL_CACHE(1);
4174 if (radv_image_is_tc_compat_htile(iview
->image
)) {
4175 unsigned max_zplanes
=
4176 radv_calc_decompress_on_z_planes(device
, iview
);
4178 ds
->db_htile_surface
|= S_028ABC_TC_COMPATIBLE(1);
4179 ds
->db_z_info
|= S_028040_DECOMPRESS_ON_N_ZPLANES(max_zplanes
);
4184 ds
->db_z_read_base
= ds
->db_z_write_base
= z_offs
>> 8;
4185 ds
->db_stencil_read_base
= ds
->db_stencil_write_base
= s_offs
>> 8;
4188 VkResult
radv_CreateFramebuffer(
4190 const VkFramebufferCreateInfo
* pCreateInfo
,
4191 const VkAllocationCallbacks
* pAllocator
,
4192 VkFramebuffer
* pFramebuffer
)
4194 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4195 struct radv_framebuffer
*framebuffer
;
4197 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
4199 size_t size
= sizeof(*framebuffer
) +
4200 sizeof(struct radv_attachment_info
) * pCreateInfo
->attachmentCount
;
4201 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
4202 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
4203 if (framebuffer
== NULL
)
4204 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
4206 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
4207 framebuffer
->width
= pCreateInfo
->width
;
4208 framebuffer
->height
= pCreateInfo
->height
;
4209 framebuffer
->layers
= pCreateInfo
->layers
;
4210 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
4211 VkImageView _iview
= pCreateInfo
->pAttachments
[i
];
4212 struct radv_image_view
*iview
= radv_image_view_from_handle(_iview
);
4213 framebuffer
->attachments
[i
].attachment
= iview
;
4214 if (iview
->aspect_mask
& VK_IMAGE_ASPECT_COLOR_BIT
) {
4215 radv_initialise_color_surface(device
, &framebuffer
->attachments
[i
].cb
, iview
);
4216 } else if (iview
->aspect_mask
& (VK_IMAGE_ASPECT_DEPTH_BIT
| VK_IMAGE_ASPECT_STENCIL_BIT
)) {
4217 radv_initialise_ds_surface(device
, &framebuffer
->attachments
[i
].ds
, iview
);
4219 framebuffer
->width
= MIN2(framebuffer
->width
, iview
->extent
.width
);
4220 framebuffer
->height
= MIN2(framebuffer
->height
, iview
->extent
.height
);
4221 framebuffer
->layers
= MIN2(framebuffer
->layers
, radv_surface_max_layer_count(iview
));
4224 *pFramebuffer
= radv_framebuffer_to_handle(framebuffer
);
4228 void radv_DestroyFramebuffer(
4231 const VkAllocationCallbacks
* pAllocator
)
4233 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4234 RADV_FROM_HANDLE(radv_framebuffer
, fb
, _fb
);
4238 vk_free2(&device
->alloc
, pAllocator
, fb
);
4241 static unsigned radv_tex_wrap(VkSamplerAddressMode address_mode
)
4243 switch (address_mode
) {
4244 case VK_SAMPLER_ADDRESS_MODE_REPEAT
:
4245 return V_008F30_SQ_TEX_WRAP
;
4246 case VK_SAMPLER_ADDRESS_MODE_MIRRORED_REPEAT
:
4247 return V_008F30_SQ_TEX_MIRROR
;
4248 case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE
:
4249 return V_008F30_SQ_TEX_CLAMP_LAST_TEXEL
;
4250 case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_BORDER
:
4251 return V_008F30_SQ_TEX_CLAMP_BORDER
;
4252 case VK_SAMPLER_ADDRESS_MODE_MIRROR_CLAMP_TO_EDGE
:
4253 return V_008F30_SQ_TEX_MIRROR_ONCE_LAST_TEXEL
;
4255 unreachable("illegal tex wrap mode");
4261 radv_tex_compare(VkCompareOp op
)
4264 case VK_COMPARE_OP_NEVER
:
4265 return V_008F30_SQ_TEX_DEPTH_COMPARE_NEVER
;
4266 case VK_COMPARE_OP_LESS
:
4267 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESS
;
4268 case VK_COMPARE_OP_EQUAL
:
4269 return V_008F30_SQ_TEX_DEPTH_COMPARE_EQUAL
;
4270 case VK_COMPARE_OP_LESS_OR_EQUAL
:
4271 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESSEQUAL
;
4272 case VK_COMPARE_OP_GREATER
:
4273 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATER
;
4274 case VK_COMPARE_OP_NOT_EQUAL
:
4275 return V_008F30_SQ_TEX_DEPTH_COMPARE_NOTEQUAL
;
4276 case VK_COMPARE_OP_GREATER_OR_EQUAL
:
4277 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATEREQUAL
;
4278 case VK_COMPARE_OP_ALWAYS
:
4279 return V_008F30_SQ_TEX_DEPTH_COMPARE_ALWAYS
;
4281 unreachable("illegal compare mode");
4287 radv_tex_filter(VkFilter filter
, unsigned max_ansio
)
4290 case VK_FILTER_NEAREST
:
4291 return (max_ansio
> 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_POINT
:
4292 V_008F38_SQ_TEX_XY_FILTER_POINT
);
4293 case VK_FILTER_LINEAR
:
4294 return (max_ansio
> 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_BILINEAR
:
4295 V_008F38_SQ_TEX_XY_FILTER_BILINEAR
);
4296 case VK_FILTER_CUBIC_IMG
:
4298 fprintf(stderr
, "illegal texture filter");
4304 radv_tex_mipfilter(VkSamplerMipmapMode mode
)
4307 case VK_SAMPLER_MIPMAP_MODE_NEAREST
:
4308 return V_008F38_SQ_TEX_Z_FILTER_POINT
;
4309 case VK_SAMPLER_MIPMAP_MODE_LINEAR
:
4310 return V_008F38_SQ_TEX_Z_FILTER_LINEAR
;
4312 return V_008F38_SQ_TEX_Z_FILTER_NONE
;
4317 radv_tex_bordercolor(VkBorderColor bcolor
)
4320 case VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
:
4321 case VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
:
4322 return V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK
;
4323 case VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
:
4324 case VK_BORDER_COLOR_INT_OPAQUE_BLACK
:
4325 return V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_BLACK
;
4326 case VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
:
4327 case VK_BORDER_COLOR_INT_OPAQUE_WHITE
:
4328 return V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_WHITE
;
4336 radv_tex_aniso_filter(unsigned filter
)
4350 radv_tex_filter_mode(VkSamplerReductionModeEXT mode
)
4353 case VK_SAMPLER_REDUCTION_MODE_WEIGHTED_AVERAGE_EXT
:
4354 return SQ_IMG_FILTER_MODE_BLEND
;
4355 case VK_SAMPLER_REDUCTION_MODE_MIN_EXT
:
4356 return SQ_IMG_FILTER_MODE_MIN
;
4357 case VK_SAMPLER_REDUCTION_MODE_MAX_EXT
:
4358 return SQ_IMG_FILTER_MODE_MAX
;
4366 radv_init_sampler(struct radv_device
*device
,
4367 struct radv_sampler
*sampler
,
4368 const VkSamplerCreateInfo
*pCreateInfo
)
4370 uint32_t max_aniso
= pCreateInfo
->anisotropyEnable
&& pCreateInfo
->maxAnisotropy
> 1.0 ?
4371 (uint32_t) pCreateInfo
->maxAnisotropy
: 0;
4372 uint32_t max_aniso_ratio
= radv_tex_aniso_filter(max_aniso
);
4373 bool is_vi
= (device
->physical_device
->rad_info
.chip_class
>= VI
);
4374 unsigned filter_mode
= SQ_IMG_FILTER_MODE_BLEND
;
4376 const struct VkSamplerReductionModeCreateInfoEXT
*sampler_reduction
=
4377 vk_find_struct_const(pCreateInfo
->pNext
,
4378 SAMPLER_REDUCTION_MODE_CREATE_INFO_EXT
);
4379 if (sampler_reduction
)
4380 filter_mode
= radv_tex_filter_mode(sampler_reduction
->reductionMode
);
4382 sampler
->state
[0] = (S_008F30_CLAMP_X(radv_tex_wrap(pCreateInfo
->addressModeU
)) |
4383 S_008F30_CLAMP_Y(radv_tex_wrap(pCreateInfo
->addressModeV
)) |
4384 S_008F30_CLAMP_Z(radv_tex_wrap(pCreateInfo
->addressModeW
)) |
4385 S_008F30_MAX_ANISO_RATIO(max_aniso_ratio
) |
4386 S_008F30_DEPTH_COMPARE_FUNC(radv_tex_compare(pCreateInfo
->compareOp
)) |
4387 S_008F30_FORCE_UNNORMALIZED(pCreateInfo
->unnormalizedCoordinates
? 1 : 0) |
4388 S_008F30_ANISO_THRESHOLD(max_aniso_ratio
>> 1) |
4389 S_008F30_ANISO_BIAS(max_aniso_ratio
) |
4390 S_008F30_DISABLE_CUBE_WRAP(0) |
4391 S_008F30_COMPAT_MODE(is_vi
) |
4392 S_008F30_FILTER_MODE(filter_mode
));
4393 sampler
->state
[1] = (S_008F34_MIN_LOD(S_FIXED(CLAMP(pCreateInfo
->minLod
, 0, 15), 8)) |
4394 S_008F34_MAX_LOD(S_FIXED(CLAMP(pCreateInfo
->maxLod
, 0, 15), 8)) |
4395 S_008F34_PERF_MIP(max_aniso_ratio
? max_aniso_ratio
+ 6 : 0));
4396 sampler
->state
[2] = (S_008F38_LOD_BIAS(S_FIXED(CLAMP(pCreateInfo
->mipLodBias
, -16, 16), 8)) |
4397 S_008F38_XY_MAG_FILTER(radv_tex_filter(pCreateInfo
->magFilter
, max_aniso
)) |
4398 S_008F38_XY_MIN_FILTER(radv_tex_filter(pCreateInfo
->minFilter
, max_aniso
)) |
4399 S_008F38_MIP_FILTER(radv_tex_mipfilter(pCreateInfo
->mipmapMode
)) |
4400 S_008F38_MIP_POINT_PRECLAMP(0) |
4401 S_008F38_DISABLE_LSB_CEIL(device
->physical_device
->rad_info
.chip_class
<= VI
) |
4402 S_008F38_FILTER_PREC_FIX(1) |
4403 S_008F38_ANISO_OVERRIDE(is_vi
));
4404 sampler
->state
[3] = (S_008F3C_BORDER_COLOR_PTR(0) |
4405 S_008F3C_BORDER_COLOR_TYPE(radv_tex_bordercolor(pCreateInfo
->borderColor
)));
4408 VkResult
radv_CreateSampler(
4410 const VkSamplerCreateInfo
* pCreateInfo
,
4411 const VkAllocationCallbacks
* pAllocator
,
4412 VkSampler
* pSampler
)
4414 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4415 struct radv_sampler
*sampler
;
4417 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO
);
4419 sampler
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*sampler
), 8,
4420 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
4422 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
4424 radv_init_sampler(device
, sampler
, pCreateInfo
);
4425 *pSampler
= radv_sampler_to_handle(sampler
);
4430 void radv_DestroySampler(
4433 const VkAllocationCallbacks
* pAllocator
)
4435 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4436 RADV_FROM_HANDLE(radv_sampler
, sampler
, _sampler
);
4440 vk_free2(&device
->alloc
, pAllocator
, sampler
);
4443 /* vk_icd.h does not declare this function, so we declare it here to
4444 * suppress Wmissing-prototypes.
4446 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4447 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t *pSupportedVersion
);
4449 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4450 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t *pSupportedVersion
)
4452 /* For the full details on loader interface versioning, see
4453 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
4454 * What follows is a condensed summary, to help you navigate the large and
4455 * confusing official doc.
4457 * - Loader interface v0 is incompatible with later versions. We don't
4460 * - In loader interface v1:
4461 * - The first ICD entrypoint called by the loader is
4462 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
4464 * - The ICD must statically expose no other Vulkan symbol unless it is
4465 * linked with -Bsymbolic.
4466 * - Each dispatchable Vulkan handle created by the ICD must be
4467 * a pointer to a struct whose first member is VK_LOADER_DATA. The
4468 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
4469 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
4470 * vkDestroySurfaceKHR(). The ICD must be capable of working with
4471 * such loader-managed surfaces.
4473 * - Loader interface v2 differs from v1 in:
4474 * - The first ICD entrypoint called by the loader is
4475 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
4476 * statically expose this entrypoint.
4478 * - Loader interface v3 differs from v2 in:
4479 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
4480 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
4481 * because the loader no longer does so.
4483 *pSupportedVersion
= MIN2(*pSupportedVersion
, 3u);
4487 VkResult
radv_GetMemoryFdKHR(VkDevice _device
,
4488 const VkMemoryGetFdInfoKHR
*pGetFdInfo
,
4491 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4492 RADV_FROM_HANDLE(radv_device_memory
, memory
, pGetFdInfo
->memory
);
4494 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
4496 /* At the moment, we support only the below handle types. */
4497 assert(pGetFdInfo
->handleType
==
4498 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
||
4499 pGetFdInfo
->handleType
==
4500 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
4502 bool ret
= radv_get_memory_fd(device
, memory
, pFD
);
4504 return vk_error(device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
4508 VkResult
radv_GetMemoryFdPropertiesKHR(VkDevice _device
,
4509 VkExternalMemoryHandleTypeFlagBitsKHR handleType
,
4511 VkMemoryFdPropertiesKHR
*pMemoryFdProperties
)
4513 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4515 switch (handleType
) {
4516 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
:
4517 pMemoryFdProperties
->memoryTypeBits
= (1 << RADV_MEM_TYPE_COUNT
) - 1;
4521 /* The valid usage section for this function says:
4523 * "handleType must not be one of the handle types defined as
4526 * So opaque handle types fall into the default "unsupported" case.
4528 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
4532 static VkResult
radv_import_opaque_fd(struct radv_device
*device
,
4536 uint32_t syncobj_handle
= 0;
4537 int ret
= device
->ws
->import_syncobj(device
->ws
, fd
, &syncobj_handle
);
4539 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
4542 device
->ws
->destroy_syncobj(device
->ws
, *syncobj
);
4544 *syncobj
= syncobj_handle
;
4550 static VkResult
radv_import_sync_fd(struct radv_device
*device
,
4554 /* If we create a syncobj we do it locally so that if we have an error, we don't
4555 * leave a syncobj in an undetermined state in the fence. */
4556 uint32_t syncobj_handle
= *syncobj
;
4557 if (!syncobj_handle
) {
4558 int ret
= device
->ws
->create_syncobj(device
->ws
, &syncobj_handle
);
4560 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
4565 device
->ws
->signal_syncobj(device
->ws
, syncobj_handle
);
4567 int ret
= device
->ws
->import_syncobj_from_sync_file(device
->ws
, syncobj_handle
, fd
);
4569 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
4572 *syncobj
= syncobj_handle
;
4579 VkResult
radv_ImportSemaphoreFdKHR(VkDevice _device
,
4580 const VkImportSemaphoreFdInfoKHR
*pImportSemaphoreFdInfo
)
4582 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4583 RADV_FROM_HANDLE(radv_semaphore
, sem
, pImportSemaphoreFdInfo
->semaphore
);
4584 uint32_t *syncobj_dst
= NULL
;
4586 if (pImportSemaphoreFdInfo
->flags
& VK_SEMAPHORE_IMPORT_TEMPORARY_BIT_KHR
) {
4587 syncobj_dst
= &sem
->temp_syncobj
;
4589 syncobj_dst
= &sem
->syncobj
;
4592 switch(pImportSemaphoreFdInfo
->handleType
) {
4593 case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
:
4594 return radv_import_opaque_fd(device
, pImportSemaphoreFdInfo
->fd
, syncobj_dst
);
4595 case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT_KHR
:
4596 return radv_import_sync_fd(device
, pImportSemaphoreFdInfo
->fd
, syncobj_dst
);
4598 unreachable("Unhandled semaphore handle type");
4602 VkResult
radv_GetSemaphoreFdKHR(VkDevice _device
,
4603 const VkSemaphoreGetFdInfoKHR
*pGetFdInfo
,
4606 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4607 RADV_FROM_HANDLE(radv_semaphore
, sem
, pGetFdInfo
->semaphore
);
4609 uint32_t syncobj_handle
;
4611 if (sem
->temp_syncobj
)
4612 syncobj_handle
= sem
->temp_syncobj
;
4614 syncobj_handle
= sem
->syncobj
;
4616 switch(pGetFdInfo
->handleType
) {
4617 case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
:
4618 ret
= device
->ws
->export_syncobj(device
->ws
, syncobj_handle
, pFd
);
4620 case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT_KHR
:
4621 ret
= device
->ws
->export_syncobj_to_sync_file(device
->ws
, syncobj_handle
, pFd
);
4623 if (sem
->temp_syncobj
) {
4624 close (sem
->temp_syncobj
);
4625 sem
->temp_syncobj
= 0;
4627 device
->ws
->reset_syncobj(device
->ws
, syncobj_handle
);
4632 unreachable("Unhandled semaphore handle type");
4636 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
4640 void radv_GetPhysicalDeviceExternalSemaphoreProperties(
4641 VkPhysicalDevice physicalDevice
,
4642 const VkPhysicalDeviceExternalSemaphoreInfoKHR
* pExternalSemaphoreInfo
,
4643 VkExternalSemaphorePropertiesKHR
* pExternalSemaphoreProperties
)
4645 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
4647 /* Require has_syncobj_wait_for_submit for the syncobj signal ioctl introduced at virtually the same time */
4648 if (pdevice
->rad_info
.has_syncobj_wait_for_submit
&&
4649 (pExternalSemaphoreInfo
->handleType
== VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
||
4650 pExternalSemaphoreInfo
->handleType
== VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT_KHR
)) {
4651 pExternalSemaphoreProperties
->exportFromImportedHandleTypes
= VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
| VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT_KHR
;
4652 pExternalSemaphoreProperties
->compatibleHandleTypes
= VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
| VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT_KHR
;
4653 pExternalSemaphoreProperties
->externalSemaphoreFeatures
= VK_EXTERNAL_SEMAPHORE_FEATURE_EXPORTABLE_BIT_KHR
|
4654 VK_EXTERNAL_SEMAPHORE_FEATURE_IMPORTABLE_BIT_KHR
;
4655 } else if (pExternalSemaphoreInfo
->handleType
== VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
) {
4656 pExternalSemaphoreProperties
->exportFromImportedHandleTypes
= VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
;
4657 pExternalSemaphoreProperties
->compatibleHandleTypes
= VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
;
4658 pExternalSemaphoreProperties
->externalSemaphoreFeatures
= VK_EXTERNAL_SEMAPHORE_FEATURE_EXPORTABLE_BIT_KHR
|
4659 VK_EXTERNAL_SEMAPHORE_FEATURE_IMPORTABLE_BIT_KHR
;
4661 pExternalSemaphoreProperties
->exportFromImportedHandleTypes
= 0;
4662 pExternalSemaphoreProperties
->compatibleHandleTypes
= 0;
4663 pExternalSemaphoreProperties
->externalSemaphoreFeatures
= 0;
4667 VkResult
radv_ImportFenceFdKHR(VkDevice _device
,
4668 const VkImportFenceFdInfoKHR
*pImportFenceFdInfo
)
4670 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4671 RADV_FROM_HANDLE(radv_fence
, fence
, pImportFenceFdInfo
->fence
);
4672 uint32_t *syncobj_dst
= NULL
;
4675 if (pImportFenceFdInfo
->flags
& VK_FENCE_IMPORT_TEMPORARY_BIT_KHR
) {
4676 syncobj_dst
= &fence
->temp_syncobj
;
4678 syncobj_dst
= &fence
->syncobj
;
4681 switch(pImportFenceFdInfo
->handleType
) {
4682 case VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
:
4683 return radv_import_opaque_fd(device
, pImportFenceFdInfo
->fd
, syncobj_dst
);
4684 case VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT_KHR
:
4685 return radv_import_sync_fd(device
, pImportFenceFdInfo
->fd
, syncobj_dst
);
4687 unreachable("Unhandled fence handle type");
4691 VkResult
radv_GetFenceFdKHR(VkDevice _device
,
4692 const VkFenceGetFdInfoKHR
*pGetFdInfo
,
4695 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4696 RADV_FROM_HANDLE(radv_fence
, fence
, pGetFdInfo
->fence
);
4698 uint32_t syncobj_handle
;
4700 if (fence
->temp_syncobj
)
4701 syncobj_handle
= fence
->temp_syncobj
;
4703 syncobj_handle
= fence
->syncobj
;
4705 switch(pGetFdInfo
->handleType
) {
4706 case VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
:
4707 ret
= device
->ws
->export_syncobj(device
->ws
, syncobj_handle
, pFd
);
4709 case VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT_KHR
:
4710 ret
= device
->ws
->export_syncobj_to_sync_file(device
->ws
, syncobj_handle
, pFd
);
4712 if (fence
->temp_syncobj
) {
4713 close (fence
->temp_syncobj
);
4714 fence
->temp_syncobj
= 0;
4716 device
->ws
->reset_syncobj(device
->ws
, syncobj_handle
);
4721 unreachable("Unhandled fence handle type");
4725 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
4729 void radv_GetPhysicalDeviceExternalFenceProperties(
4730 VkPhysicalDevice physicalDevice
,
4731 const VkPhysicalDeviceExternalFenceInfoKHR
* pExternalFenceInfo
,
4732 VkExternalFencePropertiesKHR
* pExternalFenceProperties
)
4734 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
4736 if (pdevice
->rad_info
.has_syncobj_wait_for_submit
&&
4737 (pExternalFenceInfo
->handleType
== VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
||
4738 pExternalFenceInfo
->handleType
== VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT_KHR
)) {
4739 pExternalFenceProperties
->exportFromImportedHandleTypes
= VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
| VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT_KHR
;
4740 pExternalFenceProperties
->compatibleHandleTypes
= VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
| VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT_KHR
;
4741 pExternalFenceProperties
->externalFenceFeatures
= VK_EXTERNAL_FENCE_FEATURE_EXPORTABLE_BIT_KHR
|
4742 VK_EXTERNAL_SEMAPHORE_FEATURE_IMPORTABLE_BIT_KHR
;
4744 pExternalFenceProperties
->exportFromImportedHandleTypes
= 0;
4745 pExternalFenceProperties
->compatibleHandleTypes
= 0;
4746 pExternalFenceProperties
->externalFenceFeatures
= 0;
4751 radv_CreateDebugReportCallbackEXT(VkInstance _instance
,
4752 const VkDebugReportCallbackCreateInfoEXT
* pCreateInfo
,
4753 const VkAllocationCallbacks
* pAllocator
,
4754 VkDebugReportCallbackEXT
* pCallback
)
4756 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
4757 return vk_create_debug_report_callback(&instance
->debug_report_callbacks
,
4758 pCreateInfo
, pAllocator
, &instance
->alloc
,
4763 radv_DestroyDebugReportCallbackEXT(VkInstance _instance
,
4764 VkDebugReportCallbackEXT _callback
,
4765 const VkAllocationCallbacks
* pAllocator
)
4767 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
4768 vk_destroy_debug_report_callback(&instance
->debug_report_callbacks
,
4769 _callback
, pAllocator
, &instance
->alloc
);
4773 radv_DebugReportMessageEXT(VkInstance _instance
,
4774 VkDebugReportFlagsEXT flags
,
4775 VkDebugReportObjectTypeEXT objectType
,
4778 int32_t messageCode
,
4779 const char* pLayerPrefix
,
4780 const char* pMessage
)
4782 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
4783 vk_debug_report(&instance
->debug_report_callbacks
, flags
, objectType
,
4784 object
, location
, messageCode
, pLayerPrefix
, pMessage
);
4788 radv_GetDeviceGroupPeerMemoryFeatures(
4791 uint32_t localDeviceIndex
,
4792 uint32_t remoteDeviceIndex
,
4793 VkPeerMemoryFeatureFlags
* pPeerMemoryFeatures
)
4795 assert(localDeviceIndex
== remoteDeviceIndex
);
4797 *pPeerMemoryFeatures
= VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT
|
4798 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT
|
4799 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT
|
4800 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT
;