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;
824 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT
: {
825 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*features
=
826 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*)ext
;
827 features
->vertexAttributeInstanceRateDivisor
= VK_TRUE
;
828 features
->vertexAttributeInstanceRateZeroDivisor
= VK_TRUE
;
835 return radv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
838 void radv_GetPhysicalDeviceProperties(
839 VkPhysicalDevice physicalDevice
,
840 VkPhysicalDeviceProperties
* pProperties
)
842 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
843 VkSampleCountFlags sample_counts
= 0xf;
845 /* make sure that the entire descriptor set is addressable with a signed
846 * 32-bit int. So the sum of all limits scaled by descriptor size has to
847 * be at most 2 GiB. the combined image & samples object count as one of
848 * both. This limit is for the pipeline layout, not for the set layout, but
849 * there is no set limit, so we just set a pipeline limit. I don't think
850 * any app is going to hit this soon. */
851 size_t max_descriptor_set_size
= ((1ull << 31) - 16 * MAX_DYNAMIC_BUFFERS
) /
852 (32 /* uniform buffer, 32 due to potential space wasted on alignment */ +
853 32 /* storage buffer, 32 due to potential space wasted on alignment */ +
854 32 /* sampler, largest when combined with image */ +
855 64 /* sampled image */ +
856 64 /* storage image */);
858 VkPhysicalDeviceLimits limits
= {
859 .maxImageDimension1D
= (1 << 14),
860 .maxImageDimension2D
= (1 << 14),
861 .maxImageDimension3D
= (1 << 11),
862 .maxImageDimensionCube
= (1 << 14),
863 .maxImageArrayLayers
= (1 << 11),
864 .maxTexelBufferElements
= 128 * 1024 * 1024,
865 .maxUniformBufferRange
= UINT32_MAX
,
866 .maxStorageBufferRange
= UINT32_MAX
,
867 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
868 .maxMemoryAllocationCount
= UINT32_MAX
,
869 .maxSamplerAllocationCount
= 64 * 1024,
870 .bufferImageGranularity
= 64, /* A cache line */
871 .sparseAddressSpaceSize
= 0xffffffffu
, /* buffer max size */
872 .maxBoundDescriptorSets
= MAX_SETS
,
873 .maxPerStageDescriptorSamplers
= max_descriptor_set_size
,
874 .maxPerStageDescriptorUniformBuffers
= max_descriptor_set_size
,
875 .maxPerStageDescriptorStorageBuffers
= max_descriptor_set_size
,
876 .maxPerStageDescriptorSampledImages
= max_descriptor_set_size
,
877 .maxPerStageDescriptorStorageImages
= max_descriptor_set_size
,
878 .maxPerStageDescriptorInputAttachments
= max_descriptor_set_size
,
879 .maxPerStageResources
= max_descriptor_set_size
,
880 .maxDescriptorSetSamplers
= max_descriptor_set_size
,
881 .maxDescriptorSetUniformBuffers
= max_descriptor_set_size
,
882 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_UNIFORM_BUFFERS
,
883 .maxDescriptorSetStorageBuffers
= max_descriptor_set_size
,
884 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_STORAGE_BUFFERS
,
885 .maxDescriptorSetSampledImages
= max_descriptor_set_size
,
886 .maxDescriptorSetStorageImages
= max_descriptor_set_size
,
887 .maxDescriptorSetInputAttachments
= max_descriptor_set_size
,
888 .maxVertexInputAttributes
= 32,
889 .maxVertexInputBindings
= 32,
890 .maxVertexInputAttributeOffset
= 2047,
891 .maxVertexInputBindingStride
= 2048,
892 .maxVertexOutputComponents
= 128,
893 .maxTessellationGenerationLevel
= 64,
894 .maxTessellationPatchSize
= 32,
895 .maxTessellationControlPerVertexInputComponents
= 128,
896 .maxTessellationControlPerVertexOutputComponents
= 128,
897 .maxTessellationControlPerPatchOutputComponents
= 120,
898 .maxTessellationControlTotalOutputComponents
= 4096,
899 .maxTessellationEvaluationInputComponents
= 128,
900 .maxTessellationEvaluationOutputComponents
= 128,
901 .maxGeometryShaderInvocations
= 127,
902 .maxGeometryInputComponents
= 64,
903 .maxGeometryOutputComponents
= 128,
904 .maxGeometryOutputVertices
= 256,
905 .maxGeometryTotalOutputComponents
= 1024,
906 .maxFragmentInputComponents
= 128,
907 .maxFragmentOutputAttachments
= 8,
908 .maxFragmentDualSrcAttachments
= 1,
909 .maxFragmentCombinedOutputResources
= 8,
910 .maxComputeSharedMemorySize
= 32768,
911 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
912 .maxComputeWorkGroupInvocations
= 2048,
913 .maxComputeWorkGroupSize
= {
918 .subPixelPrecisionBits
= 4 /* FIXME */,
919 .subTexelPrecisionBits
= 4 /* FIXME */,
920 .mipmapPrecisionBits
= 4 /* FIXME */,
921 .maxDrawIndexedIndexValue
= UINT32_MAX
,
922 .maxDrawIndirectCount
= UINT32_MAX
,
923 .maxSamplerLodBias
= 16,
924 .maxSamplerAnisotropy
= 16,
925 .maxViewports
= MAX_VIEWPORTS
,
926 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
927 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
928 .viewportSubPixelBits
= 8,
929 .minMemoryMapAlignment
= 4096, /* A page */
930 .minTexelBufferOffsetAlignment
= 1,
931 .minUniformBufferOffsetAlignment
= 4,
932 .minStorageBufferOffsetAlignment
= 4,
933 .minTexelOffset
= -32,
934 .maxTexelOffset
= 31,
935 .minTexelGatherOffset
= -32,
936 .maxTexelGatherOffset
= 31,
937 .minInterpolationOffset
= -2,
938 .maxInterpolationOffset
= 2,
939 .subPixelInterpolationOffsetBits
= 8,
940 .maxFramebufferWidth
= (1 << 14),
941 .maxFramebufferHeight
= (1 << 14),
942 .maxFramebufferLayers
= (1 << 10),
943 .framebufferColorSampleCounts
= sample_counts
,
944 .framebufferDepthSampleCounts
= sample_counts
,
945 .framebufferStencilSampleCounts
= sample_counts
,
946 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
947 .maxColorAttachments
= MAX_RTS
,
948 .sampledImageColorSampleCounts
= sample_counts
,
949 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
950 .sampledImageDepthSampleCounts
= sample_counts
,
951 .sampledImageStencilSampleCounts
= sample_counts
,
952 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
953 .maxSampleMaskWords
= 1,
954 .timestampComputeAndGraphics
= true,
955 .timestampPeriod
= 1000000.0 / pdevice
->rad_info
.clock_crystal_freq
,
956 .maxClipDistances
= 8,
957 .maxCullDistances
= 8,
958 .maxCombinedClipAndCullDistances
= 8,
959 .discreteQueuePriorities
= 1,
960 .pointSizeRange
= { 0.125, 255.875 },
961 .lineWidthRange
= { 0.0, 7.9921875 },
962 .pointSizeGranularity
= (1.0 / 8.0),
963 .lineWidthGranularity
= (1.0 / 128.0),
964 .strictLines
= false, /* FINISHME */
965 .standardSampleLocations
= true,
966 .optimalBufferCopyOffsetAlignment
= 128,
967 .optimalBufferCopyRowPitchAlignment
= 128,
968 .nonCoherentAtomSize
= 64,
971 *pProperties
= (VkPhysicalDeviceProperties
) {
972 .apiVersion
= radv_physical_device_api_version(pdevice
),
973 .driverVersion
= vk_get_driver_version(),
974 .vendorID
= ATI_VENDOR_ID
,
975 .deviceID
= pdevice
->rad_info
.pci_id
,
976 .deviceType
= pdevice
->rad_info
.has_dedicated_vram
? VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU
: VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
978 .sparseProperties
= {0},
981 strcpy(pProperties
->deviceName
, pdevice
->name
);
982 memcpy(pProperties
->pipelineCacheUUID
, pdevice
->cache_uuid
, VK_UUID_SIZE
);
985 void radv_GetPhysicalDeviceProperties2(
986 VkPhysicalDevice physicalDevice
,
987 VkPhysicalDeviceProperties2KHR
*pProperties
)
989 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
990 radv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
992 vk_foreach_struct(ext
, pProperties
->pNext
) {
993 switch (ext
->sType
) {
994 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
995 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
996 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
997 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
1000 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES_KHR
: {
1001 VkPhysicalDeviceIDPropertiesKHR
*properties
= (VkPhysicalDeviceIDPropertiesKHR
*)ext
;
1002 memcpy(properties
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
1003 memcpy(properties
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
1004 properties
->deviceLUIDValid
= false;
1007 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES_KHR
: {
1008 VkPhysicalDeviceMultiviewPropertiesKHR
*properties
= (VkPhysicalDeviceMultiviewPropertiesKHR
*)ext
;
1009 properties
->maxMultiviewViewCount
= MAX_VIEWS
;
1010 properties
->maxMultiviewInstanceIndex
= INT_MAX
;
1013 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES_KHR
: {
1014 VkPhysicalDevicePointClippingPropertiesKHR
*properties
=
1015 (VkPhysicalDevicePointClippingPropertiesKHR
*)ext
;
1016 properties
->pointClippingBehavior
= VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES_KHR
;
1019 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DISCARD_RECTANGLE_PROPERTIES_EXT
: {
1020 VkPhysicalDeviceDiscardRectanglePropertiesEXT
*properties
=
1021 (VkPhysicalDeviceDiscardRectanglePropertiesEXT
*)ext
;
1022 properties
->maxDiscardRectangles
= MAX_DISCARD_RECTANGLES
;
1025 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT
: {
1026 VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*properties
=
1027 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*) ext
;
1028 properties
->minImportedHostPointerAlignment
= 4096;
1031 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES
: {
1032 VkPhysicalDeviceSubgroupProperties
*properties
=
1033 (VkPhysicalDeviceSubgroupProperties
*)ext
;
1034 properties
->subgroupSize
= 64;
1035 properties
->supportedStages
= VK_SHADER_STAGE_ALL
;
1036 properties
->supportedOperations
=
1037 VK_SUBGROUP_FEATURE_BASIC_BIT
|
1038 VK_SUBGROUP_FEATURE_BALLOT_BIT
|
1039 VK_SUBGROUP_FEATURE_QUAD_BIT
|
1040 VK_SUBGROUP_FEATURE_VOTE_BIT
;
1041 if (pdevice
->rad_info
.chip_class
>= VI
) {
1042 properties
->supportedOperations
|=
1043 VK_SUBGROUP_FEATURE_SHUFFLE_BIT
|
1044 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT
;
1046 properties
->quadOperationsInAllStages
= true;
1049 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES
: {
1050 VkPhysicalDeviceMaintenance3Properties
*properties
=
1051 (VkPhysicalDeviceMaintenance3Properties
*)ext
;
1052 /* Make sure everything is addressable by a signed 32-bit int, and
1053 * our largest descriptors are 96 bytes. */
1054 properties
->maxPerSetDescriptors
= (1ull << 31) / 96;
1055 /* Our buffer size fields allow only this much */
1056 properties
->maxMemoryAllocationSize
= 0xFFFFFFFFull
;
1059 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT
: {
1060 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*properties
=
1061 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*)ext
;
1062 /* GFX6-8 only support single channel min/max filter. */
1063 properties
->filterMinmaxImageComponentMapping
= pdevice
->rad_info
.chip_class
>= GFX9
;
1064 properties
->filterMinmaxSingleComponentFormats
= true;
1067 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CORE_PROPERTIES_AMD
: {
1068 VkPhysicalDeviceShaderCorePropertiesAMD
*properties
=
1069 (VkPhysicalDeviceShaderCorePropertiesAMD
*)ext
;
1071 /* Shader engines. */
1072 properties
->shaderEngineCount
=
1073 pdevice
->rad_info
.max_se
;
1074 properties
->shaderArraysPerEngineCount
=
1075 pdevice
->rad_info
.max_sh_per_se
;
1076 properties
->computeUnitsPerShaderArray
=
1077 pdevice
->rad_info
.num_good_compute_units
/
1078 (pdevice
->rad_info
.max_se
*
1079 pdevice
->rad_info
.max_sh_per_se
);
1080 properties
->simdPerComputeUnit
= 4;
1081 properties
->wavefrontsPerSimd
=
1082 pdevice
->rad_info
.family
== CHIP_TONGA
||
1083 pdevice
->rad_info
.family
== CHIP_ICELAND
||
1084 pdevice
->rad_info
.family
== CHIP_POLARIS10
||
1085 pdevice
->rad_info
.family
== CHIP_POLARIS11
||
1086 pdevice
->rad_info
.family
== CHIP_POLARIS12
||
1087 pdevice
->rad_info
.family
== CHIP_VEGAM
? 8 : 10;
1088 properties
->wavefrontSize
= 64;
1091 properties
->sgprsPerSimd
=
1092 radv_get_num_physical_sgprs(pdevice
);
1093 properties
->minSgprAllocation
=
1094 pdevice
->rad_info
.chip_class
>= VI
? 16 : 8;
1095 properties
->maxSgprAllocation
=
1096 pdevice
->rad_info
.family
== CHIP_TONGA
||
1097 pdevice
->rad_info
.family
== CHIP_ICELAND
? 96 : 104;
1098 properties
->sgprAllocationGranularity
=
1099 pdevice
->rad_info
.chip_class
>= VI
? 16 : 8;
1102 properties
->vgprsPerSimd
= RADV_NUM_PHYSICAL_VGPRS
;
1103 properties
->minVgprAllocation
= 4;
1104 properties
->maxVgprAllocation
= 256;
1105 properties
->vgprAllocationGranularity
= 4;
1108 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT
: {
1109 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*properties
=
1110 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*)ext
;
1111 properties
->maxVertexAttribDivisor
= UINT32_MAX
;
1114 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES_EXT
: {
1115 VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*properties
=
1116 (VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*)ext
;
1117 properties
->maxUpdateAfterBindDescriptorsInAllPools
= UINT32_MAX
/ 64;
1118 properties
->shaderUniformBufferArrayNonUniformIndexingNative
= false;
1119 properties
->shaderSampledImageArrayNonUniformIndexingNative
= false;
1120 properties
->shaderStorageBufferArrayNonUniformIndexingNative
= false;
1121 properties
->shaderStorageImageArrayNonUniformIndexingNative
= false;
1122 properties
->shaderInputAttachmentArrayNonUniformIndexingNative
= false;
1123 properties
->robustBufferAccessUpdateAfterBind
= false;
1124 properties
->quadDivergentImplicitLod
= false;
1126 size_t max_descriptor_set_size
= ((1ull << 31) - 16 * MAX_DYNAMIC_BUFFERS
) /
1127 (32 /* uniform buffer, 32 due to potential space wasted on alignment */ +
1128 32 /* storage buffer, 32 due to potential space wasted on alignment */ +
1129 32 /* sampler, largest when combined with image */ +
1130 64 /* sampled image */ +
1131 64 /* storage image */);
1132 properties
->maxPerStageDescriptorUpdateAfterBindSamplers
= max_descriptor_set_size
;
1133 properties
->maxPerStageDescriptorUpdateAfterBindUniformBuffers
= max_descriptor_set_size
;
1134 properties
->maxPerStageDescriptorUpdateAfterBindStorageBuffers
= max_descriptor_set_size
;
1135 properties
->maxPerStageDescriptorUpdateAfterBindSampledImages
= max_descriptor_set_size
;
1136 properties
->maxPerStageDescriptorUpdateAfterBindStorageImages
= max_descriptor_set_size
;
1137 properties
->maxPerStageDescriptorUpdateAfterBindInputAttachments
= max_descriptor_set_size
;
1138 properties
->maxPerStageUpdateAfterBindResources
= max_descriptor_set_size
;
1139 properties
->maxDescriptorSetUpdateAfterBindSamplers
= max_descriptor_set_size
;
1140 properties
->maxDescriptorSetUpdateAfterBindUniformBuffers
= max_descriptor_set_size
;
1141 properties
->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic
= MAX_DYNAMIC_UNIFORM_BUFFERS
;
1142 properties
->maxDescriptorSetUpdateAfterBindStorageBuffers
= max_descriptor_set_size
;
1143 properties
->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic
= MAX_DYNAMIC_STORAGE_BUFFERS
;
1144 properties
->maxDescriptorSetUpdateAfterBindSampledImages
= max_descriptor_set_size
;
1145 properties
->maxDescriptorSetUpdateAfterBindStorageImages
= max_descriptor_set_size
;
1146 properties
->maxDescriptorSetUpdateAfterBindInputAttachments
= max_descriptor_set_size
;
1149 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES
: {
1150 VkPhysicalDeviceProtectedMemoryProperties
*properties
=
1151 (VkPhysicalDeviceProtectedMemoryProperties
*)ext
;
1152 properties
->protectedNoFault
= false;
1161 static void radv_get_physical_device_queue_family_properties(
1162 struct radv_physical_device
* pdevice
,
1164 VkQueueFamilyProperties
** pQueueFamilyProperties
)
1166 int num_queue_families
= 1;
1168 if (pdevice
->rad_info
.num_compute_rings
> 0 &&
1169 !(pdevice
->instance
->debug_flags
& RADV_DEBUG_NO_COMPUTE_QUEUE
))
1170 num_queue_families
++;
1172 if (pQueueFamilyProperties
== NULL
) {
1173 *pCount
= num_queue_families
;
1182 *pQueueFamilyProperties
[idx
] = (VkQueueFamilyProperties
) {
1183 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
1184 VK_QUEUE_COMPUTE_BIT
|
1185 VK_QUEUE_TRANSFER_BIT
|
1186 VK_QUEUE_SPARSE_BINDING_BIT
,
1188 .timestampValidBits
= 64,
1189 .minImageTransferGranularity
= (VkExtent3D
) { 1, 1, 1 },
1194 if (pdevice
->rad_info
.num_compute_rings
> 0 &&
1195 !(pdevice
->instance
->debug_flags
& RADV_DEBUG_NO_COMPUTE_QUEUE
)) {
1196 if (*pCount
> idx
) {
1197 *pQueueFamilyProperties
[idx
] = (VkQueueFamilyProperties
) {
1198 .queueFlags
= VK_QUEUE_COMPUTE_BIT
|
1199 VK_QUEUE_TRANSFER_BIT
|
1200 VK_QUEUE_SPARSE_BINDING_BIT
,
1201 .queueCount
= pdevice
->rad_info
.num_compute_rings
,
1202 .timestampValidBits
= 64,
1203 .minImageTransferGranularity
= (VkExtent3D
) { 1, 1, 1 },
1211 void radv_GetPhysicalDeviceQueueFamilyProperties(
1212 VkPhysicalDevice physicalDevice
,
1214 VkQueueFamilyProperties
* pQueueFamilyProperties
)
1216 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
1217 if (!pQueueFamilyProperties
) {
1218 return radv_get_physical_device_queue_family_properties(pdevice
, pCount
, NULL
);
1221 VkQueueFamilyProperties
*properties
[] = {
1222 pQueueFamilyProperties
+ 0,
1223 pQueueFamilyProperties
+ 1,
1224 pQueueFamilyProperties
+ 2,
1226 radv_get_physical_device_queue_family_properties(pdevice
, pCount
, properties
);
1227 assert(*pCount
<= 3);
1230 void radv_GetPhysicalDeviceQueueFamilyProperties2(
1231 VkPhysicalDevice physicalDevice
,
1233 VkQueueFamilyProperties2KHR
*pQueueFamilyProperties
)
1235 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
1236 if (!pQueueFamilyProperties
) {
1237 return radv_get_physical_device_queue_family_properties(pdevice
, pCount
, NULL
);
1240 VkQueueFamilyProperties
*properties
[] = {
1241 &pQueueFamilyProperties
[0].queueFamilyProperties
,
1242 &pQueueFamilyProperties
[1].queueFamilyProperties
,
1243 &pQueueFamilyProperties
[2].queueFamilyProperties
,
1245 radv_get_physical_device_queue_family_properties(pdevice
, pCount
, properties
);
1246 assert(*pCount
<= 3);
1249 void radv_GetPhysicalDeviceMemoryProperties(
1250 VkPhysicalDevice physicalDevice
,
1251 VkPhysicalDeviceMemoryProperties
*pMemoryProperties
)
1253 RADV_FROM_HANDLE(radv_physical_device
, physical_device
, physicalDevice
);
1255 *pMemoryProperties
= physical_device
->memory_properties
;
1258 void radv_GetPhysicalDeviceMemoryProperties2(
1259 VkPhysicalDevice physicalDevice
,
1260 VkPhysicalDeviceMemoryProperties2KHR
*pMemoryProperties
)
1262 return radv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
1263 &pMemoryProperties
->memoryProperties
);
1266 VkResult
radv_GetMemoryHostPointerPropertiesEXT(
1268 VkExternalMemoryHandleTypeFlagBitsKHR handleType
,
1269 const void *pHostPointer
,
1270 VkMemoryHostPointerPropertiesEXT
*pMemoryHostPointerProperties
)
1272 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1276 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
: {
1277 const struct radv_physical_device
*physical_device
= device
->physical_device
;
1278 uint32_t memoryTypeBits
= 0;
1279 for (int i
= 0; i
< physical_device
->memory_properties
.memoryTypeCount
; i
++) {
1280 if (physical_device
->mem_type_indices
[i
] == RADV_MEM_TYPE_GTT_CACHED
) {
1281 memoryTypeBits
= (1 << i
);
1285 pMemoryHostPointerProperties
->memoryTypeBits
= memoryTypeBits
;
1289 return VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
;
1293 static enum radeon_ctx_priority
1294 radv_get_queue_global_priority(const VkDeviceQueueGlobalPriorityCreateInfoEXT
*pObj
)
1296 /* Default to MEDIUM when a specific global priority isn't requested */
1298 return RADEON_CTX_PRIORITY_MEDIUM
;
1300 switch(pObj
->globalPriority
) {
1301 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT
:
1302 return RADEON_CTX_PRIORITY_REALTIME
;
1303 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT
:
1304 return RADEON_CTX_PRIORITY_HIGH
;
1305 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
:
1306 return RADEON_CTX_PRIORITY_MEDIUM
;
1307 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT
:
1308 return RADEON_CTX_PRIORITY_LOW
;
1310 unreachable("Illegal global priority value");
1311 return RADEON_CTX_PRIORITY_INVALID
;
1316 radv_queue_init(struct radv_device
*device
, struct radv_queue
*queue
,
1317 uint32_t queue_family_index
, int idx
,
1318 VkDeviceQueueCreateFlags flags
,
1319 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*global_priority
)
1321 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1322 queue
->device
= device
;
1323 queue
->queue_family_index
= queue_family_index
;
1324 queue
->queue_idx
= idx
;
1325 queue
->priority
= radv_get_queue_global_priority(global_priority
);
1326 queue
->flags
= flags
;
1328 queue
->hw_ctx
= device
->ws
->ctx_create(device
->ws
, queue
->priority
);
1330 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
1336 radv_queue_finish(struct radv_queue
*queue
)
1339 queue
->device
->ws
->ctx_destroy(queue
->hw_ctx
);
1341 if (queue
->initial_full_flush_preamble_cs
)
1342 queue
->device
->ws
->cs_destroy(queue
->initial_full_flush_preamble_cs
);
1343 if (queue
->initial_preamble_cs
)
1344 queue
->device
->ws
->cs_destroy(queue
->initial_preamble_cs
);
1345 if (queue
->continue_preamble_cs
)
1346 queue
->device
->ws
->cs_destroy(queue
->continue_preamble_cs
);
1347 if (queue
->descriptor_bo
)
1348 queue
->device
->ws
->buffer_destroy(queue
->descriptor_bo
);
1349 if (queue
->scratch_bo
)
1350 queue
->device
->ws
->buffer_destroy(queue
->scratch_bo
);
1351 if (queue
->esgs_ring_bo
)
1352 queue
->device
->ws
->buffer_destroy(queue
->esgs_ring_bo
);
1353 if (queue
->gsvs_ring_bo
)
1354 queue
->device
->ws
->buffer_destroy(queue
->gsvs_ring_bo
);
1355 if (queue
->tess_rings_bo
)
1356 queue
->device
->ws
->buffer_destroy(queue
->tess_rings_bo
);
1357 if (queue
->compute_scratch_bo
)
1358 queue
->device
->ws
->buffer_destroy(queue
->compute_scratch_bo
);
1362 radv_bo_list_init(struct radv_bo_list
*bo_list
)
1364 pthread_mutex_init(&bo_list
->mutex
, NULL
);
1365 bo_list
->list
.count
= bo_list
->capacity
= 0;
1366 bo_list
->list
.bos
= NULL
;
1370 radv_bo_list_finish(struct radv_bo_list
*bo_list
)
1372 free(bo_list
->list
.bos
);
1373 pthread_mutex_destroy(&bo_list
->mutex
);
1376 static VkResult
radv_bo_list_add(struct radv_device
*device
,
1377 struct radeon_winsys_bo
*bo
)
1379 struct radv_bo_list
*bo_list
= &device
->bo_list
;
1381 if (unlikely(!device
->use_global_bo_list
))
1384 pthread_mutex_lock(&bo_list
->mutex
);
1385 if (bo_list
->list
.count
== bo_list
->capacity
) {
1386 unsigned capacity
= MAX2(4, bo_list
->capacity
* 2);
1387 void *data
= realloc(bo_list
->list
.bos
, capacity
* sizeof(struct radeon_winsys_bo
*));
1390 pthread_mutex_unlock(&bo_list
->mutex
);
1391 return VK_ERROR_OUT_OF_HOST_MEMORY
;
1394 bo_list
->list
.bos
= (struct radeon_winsys_bo
**)data
;
1395 bo_list
->capacity
= capacity
;
1398 bo_list
->list
.bos
[bo_list
->list
.count
++] = bo
;
1399 pthread_mutex_unlock(&bo_list
->mutex
);
1403 static void radv_bo_list_remove(struct radv_device
*device
,
1404 struct radeon_winsys_bo
*bo
)
1406 struct radv_bo_list
*bo_list
= &device
->bo_list
;
1408 if (unlikely(!device
->use_global_bo_list
))
1411 pthread_mutex_lock(&bo_list
->mutex
);
1412 for(unsigned i
= 0; i
< bo_list
->list
.count
; ++i
) {
1413 if (bo_list
->list
.bos
[i
] == bo
) {
1414 bo_list
->list
.bos
[i
] = bo_list
->list
.bos
[bo_list
->list
.count
- 1];
1415 --bo_list
->list
.count
;
1419 pthread_mutex_unlock(&bo_list
->mutex
);
1423 radv_device_init_gs_info(struct radv_device
*device
)
1425 device
->gs_table_depth
= ac_get_gs_table_depth(device
->physical_device
->rad_info
.chip_class
,
1426 device
->physical_device
->rad_info
.family
);
1429 static int radv_get_device_extension_index(const char *name
)
1431 for (unsigned i
= 0; i
< RADV_DEVICE_EXTENSION_COUNT
; ++i
) {
1432 if (strcmp(name
, radv_device_extensions
[i
].extensionName
) == 0)
1438 VkResult
radv_CreateDevice(
1439 VkPhysicalDevice physicalDevice
,
1440 const VkDeviceCreateInfo
* pCreateInfo
,
1441 const VkAllocationCallbacks
* pAllocator
,
1444 RADV_FROM_HANDLE(radv_physical_device
, physical_device
, physicalDevice
);
1446 struct radv_device
*device
;
1448 bool keep_shader_info
= false;
1450 /* Check enabled features */
1451 if (pCreateInfo
->pEnabledFeatures
) {
1452 VkPhysicalDeviceFeatures supported_features
;
1453 radv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
1454 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
1455 VkBool32
*enabled_feature
= (VkBool32
*)pCreateInfo
->pEnabledFeatures
;
1456 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
1457 for (uint32_t i
= 0; i
< num_features
; i
++) {
1458 if (enabled_feature
[i
] && !supported_feature
[i
])
1459 return vk_error(physical_device
->instance
, VK_ERROR_FEATURE_NOT_PRESENT
);
1463 device
= vk_zalloc2(&physical_device
->instance
->alloc
, pAllocator
,
1465 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
1467 return vk_error(physical_device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
1469 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1470 device
->instance
= physical_device
->instance
;
1471 device
->physical_device
= physical_device
;
1473 device
->ws
= physical_device
->ws
;
1475 device
->alloc
= *pAllocator
;
1477 device
->alloc
= physical_device
->instance
->alloc
;
1479 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
1480 const char *ext_name
= pCreateInfo
->ppEnabledExtensionNames
[i
];
1481 int index
= radv_get_device_extension_index(ext_name
);
1482 if (index
< 0 || !physical_device
->supported_extensions
.extensions
[index
]) {
1483 vk_free(&device
->alloc
, device
);
1484 return vk_error(physical_device
->instance
, VK_ERROR_EXTENSION_NOT_PRESENT
);
1487 device
->enabled_extensions
.extensions
[index
] = true;
1490 keep_shader_info
= device
->enabled_extensions
.AMD_shader_info
;
1492 /* With update after bind we can't attach bo's to the command buffer
1493 * from the descriptor set anymore, so we have to use a global BO list.
1495 device
->use_global_bo_list
=
1496 device
->enabled_extensions
.EXT_descriptor_indexing
;
1498 mtx_init(&device
->shader_slab_mutex
, mtx_plain
);
1499 list_inithead(&device
->shader_slabs
);
1501 radv_bo_list_init(&device
->bo_list
);
1503 for (unsigned i
= 0; i
< pCreateInfo
->queueCreateInfoCount
; i
++) {
1504 const VkDeviceQueueCreateInfo
*queue_create
= &pCreateInfo
->pQueueCreateInfos
[i
];
1505 uint32_t qfi
= queue_create
->queueFamilyIndex
;
1506 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*global_priority
=
1507 vk_find_struct_const(queue_create
->pNext
, DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT
);
1509 assert(!global_priority
|| device
->physical_device
->rad_info
.has_ctx_priority
);
1511 device
->queues
[qfi
] = vk_alloc(&device
->alloc
,
1512 queue_create
->queueCount
* sizeof(struct radv_queue
), 8, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
1513 if (!device
->queues
[qfi
]) {
1514 result
= VK_ERROR_OUT_OF_HOST_MEMORY
;
1518 memset(device
->queues
[qfi
], 0, queue_create
->queueCount
* sizeof(struct radv_queue
));
1520 device
->queue_count
[qfi
] = queue_create
->queueCount
;
1522 for (unsigned q
= 0; q
< queue_create
->queueCount
; q
++) {
1523 result
= radv_queue_init(device
, &device
->queues
[qfi
][q
],
1524 qfi
, q
, queue_create
->flags
,
1526 if (result
!= VK_SUCCESS
)
1531 device
->pbb_allowed
= device
->physical_device
->rad_info
.chip_class
>= GFX9
&&
1532 ((device
->instance
->perftest_flags
& RADV_PERFTEST_BINNING
) ||
1533 device
->physical_device
->rad_info
.family
== CHIP_RAVEN
);
1535 /* Disabled and not implemented for now. */
1536 device
->dfsm_allowed
= device
->pbb_allowed
&&
1537 device
->physical_device
->rad_info
.family
== CHIP_RAVEN
;
1540 device
->always_use_syncobj
= device
->physical_device
->rad_info
.has_syncobj_wait_for_submit
;
1543 /* The maximum number of scratch waves. Scratch space isn't divided
1544 * evenly between CUs. The number is only a function of the number of CUs.
1545 * We can decrease the constant to decrease the scratch buffer size.
1547 * sctx->scratch_waves must be >= the maximum possible size of
1548 * 1 threadgroup, so that the hw doesn't hang from being unable
1551 * The recommended value is 4 per CU at most. Higher numbers don't
1552 * bring much benefit, but they still occupy chip resources (think
1553 * async compute). I've seen ~2% performance difference between 4 and 32.
1555 uint32_t max_threads_per_block
= 2048;
1556 device
->scratch_waves
= MAX2(32 * physical_device
->rad_info
.num_good_compute_units
,
1557 max_threads_per_block
/ 64);
1559 device
->dispatch_initiator
= S_00B800_COMPUTE_SHADER_EN(1);
1561 if (device
->physical_device
->rad_info
.chip_class
>= CIK
) {
1562 /* If the KMD allows it (there is a KMD hw register for it),
1563 * allow launching waves out-of-order.
1565 device
->dispatch_initiator
|= S_00B800_ORDER_MODE(1);
1568 radv_device_init_gs_info(device
);
1570 device
->tess_offchip_block_dw_size
=
1571 device
->physical_device
->rad_info
.family
== CHIP_HAWAII
? 4096 : 8192;
1572 device
->has_distributed_tess
=
1573 device
->physical_device
->rad_info
.chip_class
>= VI
&&
1574 device
->physical_device
->rad_info
.max_se
>= 2;
1576 if (getenv("RADV_TRACE_FILE")) {
1577 const char *filename
= getenv("RADV_TRACE_FILE");
1579 keep_shader_info
= true;
1581 if (!radv_init_trace(device
))
1584 fprintf(stderr
, "*****************************************************************************\n");
1585 fprintf(stderr
, "* WARNING: RADV_TRACE_FILE is costly and should only be used for debugging! *\n");
1586 fprintf(stderr
, "*****************************************************************************\n");
1588 fprintf(stderr
, "Trace file will be dumped to %s\n", filename
);
1589 radv_dump_enabled_options(device
, stderr
);
1592 device
->keep_shader_info
= keep_shader_info
;
1594 result
= radv_device_init_meta(device
);
1595 if (result
!= VK_SUCCESS
)
1598 radv_device_init_msaa(device
);
1600 for (int family
= 0; family
< RADV_MAX_QUEUE_FAMILIES
; ++family
) {
1601 device
->empty_cs
[family
] = device
->ws
->cs_create(device
->ws
, family
);
1603 case RADV_QUEUE_GENERAL
:
1604 radeon_emit(device
->empty_cs
[family
], PKT3(PKT3_CONTEXT_CONTROL
, 1, 0));
1605 radeon_emit(device
->empty_cs
[family
], CONTEXT_CONTROL_LOAD_ENABLE(1));
1606 radeon_emit(device
->empty_cs
[family
], CONTEXT_CONTROL_SHADOW_ENABLE(1));
1608 case RADV_QUEUE_COMPUTE
:
1609 radeon_emit(device
->empty_cs
[family
], PKT3(PKT3_NOP
, 0, 0));
1610 radeon_emit(device
->empty_cs
[family
], 0);
1613 device
->ws
->cs_finalize(device
->empty_cs
[family
]);
1616 if (device
->physical_device
->rad_info
.chip_class
>= CIK
)
1617 cik_create_gfx_config(device
);
1619 VkPipelineCacheCreateInfo ci
;
1620 ci
.sType
= VK_STRUCTURE_TYPE_PIPELINE_CACHE_CREATE_INFO
;
1623 ci
.pInitialData
= NULL
;
1624 ci
.initialDataSize
= 0;
1626 result
= radv_CreatePipelineCache(radv_device_to_handle(device
),
1628 if (result
!= VK_SUCCESS
)
1631 device
->mem_cache
= radv_pipeline_cache_from_handle(pc
);
1633 *pDevice
= radv_device_to_handle(device
);
1637 radv_device_finish_meta(device
);
1639 radv_bo_list_finish(&device
->bo_list
);
1641 if (device
->trace_bo
)
1642 device
->ws
->buffer_destroy(device
->trace_bo
);
1644 if (device
->gfx_init
)
1645 device
->ws
->buffer_destroy(device
->gfx_init
);
1647 for (unsigned i
= 0; i
< RADV_MAX_QUEUE_FAMILIES
; i
++) {
1648 for (unsigned q
= 0; q
< device
->queue_count
[i
]; q
++)
1649 radv_queue_finish(&device
->queues
[i
][q
]);
1650 if (device
->queue_count
[i
])
1651 vk_free(&device
->alloc
, device
->queues
[i
]);
1654 vk_free(&device
->alloc
, device
);
1658 void radv_DestroyDevice(
1660 const VkAllocationCallbacks
* pAllocator
)
1662 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1667 if (device
->trace_bo
)
1668 device
->ws
->buffer_destroy(device
->trace_bo
);
1670 if (device
->gfx_init
)
1671 device
->ws
->buffer_destroy(device
->gfx_init
);
1673 for (unsigned i
= 0; i
< RADV_MAX_QUEUE_FAMILIES
; i
++) {
1674 for (unsigned q
= 0; q
< device
->queue_count
[i
]; q
++)
1675 radv_queue_finish(&device
->queues
[i
][q
]);
1676 if (device
->queue_count
[i
])
1677 vk_free(&device
->alloc
, device
->queues
[i
]);
1678 if (device
->empty_cs
[i
])
1679 device
->ws
->cs_destroy(device
->empty_cs
[i
]);
1681 radv_device_finish_meta(device
);
1683 VkPipelineCache pc
= radv_pipeline_cache_to_handle(device
->mem_cache
);
1684 radv_DestroyPipelineCache(radv_device_to_handle(device
), pc
, NULL
);
1686 radv_destroy_shader_slabs(device
);
1688 radv_bo_list_finish(&device
->bo_list
);
1689 vk_free(&device
->alloc
, device
);
1692 VkResult
radv_EnumerateInstanceLayerProperties(
1693 uint32_t* pPropertyCount
,
1694 VkLayerProperties
* pProperties
)
1696 if (pProperties
== NULL
) {
1697 *pPropertyCount
= 0;
1701 /* None supported at this time */
1702 return vk_error(NULL
, VK_ERROR_LAYER_NOT_PRESENT
);
1705 VkResult
radv_EnumerateDeviceLayerProperties(
1706 VkPhysicalDevice physicalDevice
,
1707 uint32_t* pPropertyCount
,
1708 VkLayerProperties
* pProperties
)
1710 if (pProperties
== NULL
) {
1711 *pPropertyCount
= 0;
1715 /* None supported at this time */
1716 return vk_error(NULL
, VK_ERROR_LAYER_NOT_PRESENT
);
1719 void radv_GetDeviceQueue2(
1721 const VkDeviceQueueInfo2
* pQueueInfo
,
1724 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1725 struct radv_queue
*queue
;
1727 queue
= &device
->queues
[pQueueInfo
->queueFamilyIndex
][pQueueInfo
->queueIndex
];
1728 if (pQueueInfo
->flags
!= queue
->flags
) {
1729 /* From the Vulkan 1.1.70 spec:
1731 * "The queue returned by vkGetDeviceQueue2 must have the same
1732 * flags value from this structure as that used at device
1733 * creation time in a VkDeviceQueueCreateInfo instance. If no
1734 * matching flags were specified at device creation time then
1735 * pQueue will return VK_NULL_HANDLE."
1737 *pQueue
= VK_NULL_HANDLE
;
1741 *pQueue
= radv_queue_to_handle(queue
);
1744 void radv_GetDeviceQueue(
1746 uint32_t queueFamilyIndex
,
1747 uint32_t queueIndex
,
1750 const VkDeviceQueueInfo2 info
= (VkDeviceQueueInfo2
) {
1751 .sType
= VK_STRUCTURE_TYPE_DEVICE_QUEUE_INFO_2
,
1752 .queueFamilyIndex
= queueFamilyIndex
,
1753 .queueIndex
= queueIndex
1756 radv_GetDeviceQueue2(_device
, &info
, pQueue
);
1760 fill_geom_tess_rings(struct radv_queue
*queue
,
1762 bool add_sample_positions
,
1763 uint32_t esgs_ring_size
,
1764 struct radeon_winsys_bo
*esgs_ring_bo
,
1765 uint32_t gsvs_ring_size
,
1766 struct radeon_winsys_bo
*gsvs_ring_bo
,
1767 uint32_t tess_factor_ring_size
,
1768 uint32_t tess_offchip_ring_offset
,
1769 uint32_t tess_offchip_ring_size
,
1770 struct radeon_winsys_bo
*tess_rings_bo
)
1772 uint64_t esgs_va
= 0, gsvs_va
= 0;
1773 uint64_t tess_va
= 0, tess_offchip_va
= 0;
1774 uint32_t *desc
= &map
[4];
1777 esgs_va
= radv_buffer_get_va(esgs_ring_bo
);
1779 gsvs_va
= radv_buffer_get_va(gsvs_ring_bo
);
1780 if (tess_rings_bo
) {
1781 tess_va
= radv_buffer_get_va(tess_rings_bo
);
1782 tess_offchip_va
= tess_va
+ tess_offchip_ring_offset
;
1785 /* stride 0, num records - size, add tid, swizzle, elsize4,
1788 desc
[1] = S_008F04_BASE_ADDRESS_HI(esgs_va
>> 32) |
1789 S_008F04_STRIDE(0) |
1790 S_008F04_SWIZZLE_ENABLE(true);
1791 desc
[2] = esgs_ring_size
;
1792 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1793 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1794 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1795 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1796 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1797 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1798 S_008F0C_ELEMENT_SIZE(1) |
1799 S_008F0C_INDEX_STRIDE(3) |
1800 S_008F0C_ADD_TID_ENABLE(true);
1803 /* GS entry for ES->GS ring */
1804 /* stride 0, num records - size, elsize0,
1807 desc
[1] = S_008F04_BASE_ADDRESS_HI(esgs_va
>> 32)|
1808 S_008F04_STRIDE(0) |
1809 S_008F04_SWIZZLE_ENABLE(false);
1810 desc
[2] = esgs_ring_size
;
1811 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1812 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1813 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1814 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1815 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1816 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1817 S_008F0C_ELEMENT_SIZE(0) |
1818 S_008F0C_INDEX_STRIDE(0) |
1819 S_008F0C_ADD_TID_ENABLE(false);
1822 /* VS entry for GS->VS ring */
1823 /* stride 0, num records - size, elsize0,
1826 desc
[1] = S_008F04_BASE_ADDRESS_HI(gsvs_va
>> 32)|
1827 S_008F04_STRIDE(0) |
1828 S_008F04_SWIZZLE_ENABLE(false);
1829 desc
[2] = gsvs_ring_size
;
1830 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1831 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1832 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1833 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1834 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1835 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1836 S_008F0C_ELEMENT_SIZE(0) |
1837 S_008F0C_INDEX_STRIDE(0) |
1838 S_008F0C_ADD_TID_ENABLE(false);
1841 /* stride gsvs_itemsize, num records 64
1842 elsize 4, index stride 16 */
1843 /* shader will patch stride and desc[2] */
1845 desc
[1] = S_008F04_BASE_ADDRESS_HI(gsvs_va
>> 32)|
1846 S_008F04_STRIDE(0) |
1847 S_008F04_SWIZZLE_ENABLE(true);
1849 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1850 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1851 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1852 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1853 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1854 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1855 S_008F0C_ELEMENT_SIZE(1) |
1856 S_008F0C_INDEX_STRIDE(1) |
1857 S_008F0C_ADD_TID_ENABLE(true);
1861 desc
[1] = S_008F04_BASE_ADDRESS_HI(tess_va
>> 32) |
1862 S_008F04_STRIDE(0) |
1863 S_008F04_SWIZZLE_ENABLE(false);
1864 desc
[2] = tess_factor_ring_size
;
1865 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1866 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1867 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1868 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1869 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1870 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1871 S_008F0C_ELEMENT_SIZE(0) |
1872 S_008F0C_INDEX_STRIDE(0) |
1873 S_008F0C_ADD_TID_ENABLE(false);
1876 desc
[0] = tess_offchip_va
;
1877 desc
[1] = S_008F04_BASE_ADDRESS_HI(tess_offchip_va
>> 32) |
1878 S_008F04_STRIDE(0) |
1879 S_008F04_SWIZZLE_ENABLE(false);
1880 desc
[2] = tess_offchip_ring_size
;
1881 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1882 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1883 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1884 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1885 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1886 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1887 S_008F0C_ELEMENT_SIZE(0) |
1888 S_008F0C_INDEX_STRIDE(0) |
1889 S_008F0C_ADD_TID_ENABLE(false);
1892 /* add sample positions after all rings */
1893 memcpy(desc
, queue
->device
->sample_locations_1x
, 8);
1895 memcpy(desc
, queue
->device
->sample_locations_2x
, 16);
1897 memcpy(desc
, queue
->device
->sample_locations_4x
, 32);
1899 memcpy(desc
, queue
->device
->sample_locations_8x
, 64);
1901 memcpy(desc
, queue
->device
->sample_locations_16x
, 128);
1905 radv_get_hs_offchip_param(struct radv_device
*device
, uint32_t *max_offchip_buffers_p
)
1907 bool double_offchip_buffers
= device
->physical_device
->rad_info
.chip_class
>= CIK
&&
1908 device
->physical_device
->rad_info
.family
!= CHIP_CARRIZO
&&
1909 device
->physical_device
->rad_info
.family
!= CHIP_STONEY
;
1910 unsigned max_offchip_buffers_per_se
= double_offchip_buffers
? 128 : 64;
1911 unsigned max_offchip_buffers
;
1912 unsigned offchip_granularity
;
1913 unsigned hs_offchip_param
;
1917 * This must be one less than the maximum number due to a hw limitation.
1918 * Various hardware bugs in SI, CIK, and GFX9 need this.
1921 * Vega10 should limit max_offchip_buffers to 508 (4 * 127).
1922 * Gfx7 should limit max_offchip_buffers to 508
1923 * Gfx6 should limit max_offchip_buffers to 126 (2 * 63)
1925 * Follow AMDVLK here.
1927 if (device
->physical_device
->rad_info
.family
== CHIP_VEGA10
||
1928 device
->physical_device
->rad_info
.chip_class
== CIK
||
1929 device
->physical_device
->rad_info
.chip_class
== SI
)
1930 --max_offchip_buffers_per_se
;
1932 max_offchip_buffers
= max_offchip_buffers_per_se
*
1933 device
->physical_device
->rad_info
.max_se
;
1935 switch (device
->tess_offchip_block_dw_size
) {
1940 offchip_granularity
= V_03093C_X_8K_DWORDS
;
1943 offchip_granularity
= V_03093C_X_4K_DWORDS
;
1947 switch (device
->physical_device
->rad_info
.chip_class
) {
1949 max_offchip_buffers
= MIN2(max_offchip_buffers
, 126);
1955 max_offchip_buffers
= MIN2(max_offchip_buffers
, 508);
1959 *max_offchip_buffers_p
= max_offchip_buffers
;
1960 if (device
->physical_device
->rad_info
.chip_class
>= CIK
) {
1961 if (device
->physical_device
->rad_info
.chip_class
>= VI
)
1962 --max_offchip_buffers
;
1964 S_03093C_OFFCHIP_BUFFERING(max_offchip_buffers
) |
1965 S_03093C_OFFCHIP_GRANULARITY(offchip_granularity
);
1968 S_0089B0_OFFCHIP_BUFFERING(max_offchip_buffers
);
1970 return hs_offchip_param
;
1974 radv_emit_gs_ring_sizes(struct radv_queue
*queue
, struct radeon_cmdbuf
*cs
,
1975 struct radeon_winsys_bo
*esgs_ring_bo
,
1976 uint32_t esgs_ring_size
,
1977 struct radeon_winsys_bo
*gsvs_ring_bo
,
1978 uint32_t gsvs_ring_size
)
1980 if (!esgs_ring_bo
&& !gsvs_ring_bo
)
1984 radv_cs_add_buffer(queue
->device
->ws
, cs
, esgs_ring_bo
);
1987 radv_cs_add_buffer(queue
->device
->ws
, cs
, gsvs_ring_bo
);
1989 if (queue
->device
->physical_device
->rad_info
.chip_class
>= CIK
) {
1990 radeon_set_uconfig_reg_seq(cs
, R_030900_VGT_ESGS_RING_SIZE
, 2);
1991 radeon_emit(cs
, esgs_ring_size
>> 8);
1992 radeon_emit(cs
, gsvs_ring_size
>> 8);
1994 radeon_set_config_reg_seq(cs
, R_0088C8_VGT_ESGS_RING_SIZE
, 2);
1995 radeon_emit(cs
, esgs_ring_size
>> 8);
1996 radeon_emit(cs
, gsvs_ring_size
>> 8);
2001 radv_emit_tess_factor_ring(struct radv_queue
*queue
, struct radeon_cmdbuf
*cs
,
2002 unsigned hs_offchip_param
, unsigned tf_ring_size
,
2003 struct radeon_winsys_bo
*tess_rings_bo
)
2010 tf_va
= radv_buffer_get_va(tess_rings_bo
);
2012 radv_cs_add_buffer(queue
->device
->ws
, cs
, tess_rings_bo
);
2014 if (queue
->device
->physical_device
->rad_info
.chip_class
>= CIK
) {
2015 radeon_set_uconfig_reg(cs
, R_030938_VGT_TF_RING_SIZE
,
2016 S_030938_SIZE(tf_ring_size
/ 4));
2017 radeon_set_uconfig_reg(cs
, R_030940_VGT_TF_MEMORY_BASE
,
2019 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
2020 radeon_set_uconfig_reg(cs
, R_030944_VGT_TF_MEMORY_BASE_HI
,
2021 S_030944_BASE_HI(tf_va
>> 40));
2023 radeon_set_uconfig_reg(cs
, R_03093C_VGT_HS_OFFCHIP_PARAM
,
2026 radeon_set_config_reg(cs
, R_008988_VGT_TF_RING_SIZE
,
2027 S_008988_SIZE(tf_ring_size
/ 4));
2028 radeon_set_config_reg(cs
, R_0089B8_VGT_TF_MEMORY_BASE
,
2030 radeon_set_config_reg(cs
, R_0089B0_VGT_HS_OFFCHIP_PARAM
,
2036 radv_emit_compute_scratch(struct radv_queue
*queue
, struct radeon_cmdbuf
*cs
,
2037 struct radeon_winsys_bo
*compute_scratch_bo
)
2039 uint64_t scratch_va
;
2041 if (!compute_scratch_bo
)
2044 scratch_va
= radv_buffer_get_va(compute_scratch_bo
);
2046 radv_cs_add_buffer(queue
->device
->ws
, cs
, compute_scratch_bo
);
2048 radeon_set_sh_reg_seq(cs
, R_00B900_COMPUTE_USER_DATA_0
, 2);
2049 radeon_emit(cs
, scratch_va
);
2050 radeon_emit(cs
, S_008F04_BASE_ADDRESS_HI(scratch_va
>> 32) |
2051 S_008F04_SWIZZLE_ENABLE(1));
2055 radv_emit_global_shader_pointers(struct radv_queue
*queue
,
2056 struct radeon_cmdbuf
*cs
,
2057 struct radeon_winsys_bo
*descriptor_bo
)
2064 va
= radv_buffer_get_va(descriptor_bo
);
2066 radv_cs_add_buffer(queue
->device
->ws
, cs
, descriptor_bo
);
2068 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
2069 uint32_t regs
[] = {R_00B030_SPI_SHADER_USER_DATA_PS_0
,
2070 R_00B130_SPI_SHADER_USER_DATA_VS_0
,
2071 R_00B208_SPI_SHADER_USER_DATA_ADDR_LO_GS
,
2072 R_00B408_SPI_SHADER_USER_DATA_ADDR_LO_HS
};
2074 for (int i
= 0; i
< ARRAY_SIZE(regs
); ++i
) {
2075 radv_emit_shader_pointer(queue
->device
, cs
, regs
[i
],
2079 uint32_t regs
[] = {R_00B030_SPI_SHADER_USER_DATA_PS_0
,
2080 R_00B130_SPI_SHADER_USER_DATA_VS_0
,
2081 R_00B230_SPI_SHADER_USER_DATA_GS_0
,
2082 R_00B330_SPI_SHADER_USER_DATA_ES_0
,
2083 R_00B430_SPI_SHADER_USER_DATA_HS_0
,
2084 R_00B530_SPI_SHADER_USER_DATA_LS_0
};
2086 for (int i
= 0; i
< ARRAY_SIZE(regs
); ++i
) {
2087 radv_emit_shader_pointer(queue
->device
, cs
, regs
[i
],
2094 radv_get_preamble_cs(struct radv_queue
*queue
,
2095 uint32_t scratch_size
,
2096 uint32_t compute_scratch_size
,
2097 uint32_t esgs_ring_size
,
2098 uint32_t gsvs_ring_size
,
2099 bool needs_tess_rings
,
2100 bool needs_sample_positions
,
2101 struct radeon_cmdbuf
**initial_full_flush_preamble_cs
,
2102 struct radeon_cmdbuf
**initial_preamble_cs
,
2103 struct radeon_cmdbuf
**continue_preamble_cs
)
2105 struct radeon_winsys_bo
*scratch_bo
= NULL
;
2106 struct radeon_winsys_bo
*descriptor_bo
= NULL
;
2107 struct radeon_winsys_bo
*compute_scratch_bo
= NULL
;
2108 struct radeon_winsys_bo
*esgs_ring_bo
= NULL
;
2109 struct radeon_winsys_bo
*gsvs_ring_bo
= NULL
;
2110 struct radeon_winsys_bo
*tess_rings_bo
= NULL
;
2111 struct radeon_cmdbuf
*dest_cs
[3] = {0};
2112 bool add_tess_rings
= false, add_sample_positions
= false;
2113 unsigned tess_factor_ring_size
= 0, tess_offchip_ring_size
= 0;
2114 unsigned max_offchip_buffers
;
2115 unsigned hs_offchip_param
= 0;
2116 unsigned tess_offchip_ring_offset
;
2117 uint32_t ring_bo_flags
= RADEON_FLAG_NO_CPU_ACCESS
| RADEON_FLAG_NO_INTERPROCESS_SHARING
;
2118 if (!queue
->has_tess_rings
) {
2119 if (needs_tess_rings
)
2120 add_tess_rings
= true;
2122 if (!queue
->has_sample_positions
) {
2123 if (needs_sample_positions
)
2124 add_sample_positions
= true;
2126 tess_factor_ring_size
= 32768 * queue
->device
->physical_device
->rad_info
.max_se
;
2127 hs_offchip_param
= radv_get_hs_offchip_param(queue
->device
,
2128 &max_offchip_buffers
);
2129 tess_offchip_ring_offset
= align(tess_factor_ring_size
, 64 * 1024);
2130 tess_offchip_ring_size
= max_offchip_buffers
*
2131 queue
->device
->tess_offchip_block_dw_size
* 4;
2133 if (scratch_size
<= queue
->scratch_size
&&
2134 compute_scratch_size
<= queue
->compute_scratch_size
&&
2135 esgs_ring_size
<= queue
->esgs_ring_size
&&
2136 gsvs_ring_size
<= queue
->gsvs_ring_size
&&
2137 !add_tess_rings
&& !add_sample_positions
&&
2138 queue
->initial_preamble_cs
) {
2139 *initial_full_flush_preamble_cs
= queue
->initial_full_flush_preamble_cs
;
2140 *initial_preamble_cs
= queue
->initial_preamble_cs
;
2141 *continue_preamble_cs
= queue
->continue_preamble_cs
;
2142 if (!scratch_size
&& !compute_scratch_size
&& !esgs_ring_size
&& !gsvs_ring_size
)
2143 *continue_preamble_cs
= NULL
;
2147 if (scratch_size
> queue
->scratch_size
) {
2148 scratch_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
2156 scratch_bo
= queue
->scratch_bo
;
2158 if (compute_scratch_size
> queue
->compute_scratch_size
) {
2159 compute_scratch_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
2160 compute_scratch_size
,
2164 if (!compute_scratch_bo
)
2168 compute_scratch_bo
= queue
->compute_scratch_bo
;
2170 if (esgs_ring_size
> queue
->esgs_ring_size
) {
2171 esgs_ring_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
2179 esgs_ring_bo
= queue
->esgs_ring_bo
;
2180 esgs_ring_size
= queue
->esgs_ring_size
;
2183 if (gsvs_ring_size
> queue
->gsvs_ring_size
) {
2184 gsvs_ring_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
2192 gsvs_ring_bo
= queue
->gsvs_ring_bo
;
2193 gsvs_ring_size
= queue
->gsvs_ring_size
;
2196 if (add_tess_rings
) {
2197 tess_rings_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
2198 tess_offchip_ring_offset
+ tess_offchip_ring_size
,
2205 tess_rings_bo
= queue
->tess_rings_bo
;
2208 if (scratch_bo
!= queue
->scratch_bo
||
2209 esgs_ring_bo
!= queue
->esgs_ring_bo
||
2210 gsvs_ring_bo
!= queue
->gsvs_ring_bo
||
2211 tess_rings_bo
!= queue
->tess_rings_bo
||
2212 add_sample_positions
) {
2214 if (gsvs_ring_bo
|| esgs_ring_bo
||
2215 tess_rings_bo
|| add_sample_positions
) {
2216 size
= 112; /* 2 dword + 2 padding + 4 dword * 6 */
2217 if (add_sample_positions
)
2218 size
+= 256; /* 32+16+8+4+2+1 samples * 4 * 2 = 248 bytes. */
2220 else if (scratch_bo
)
2221 size
= 8; /* 2 dword */
2223 descriptor_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
2227 RADEON_FLAG_CPU_ACCESS
|
2228 RADEON_FLAG_NO_INTERPROCESS_SHARING
|
2229 RADEON_FLAG_READ_ONLY
);
2233 descriptor_bo
= queue
->descriptor_bo
;
2235 for(int i
= 0; i
< 3; ++i
) {
2236 struct radeon_cmdbuf
*cs
= NULL
;
2237 cs
= queue
->device
->ws
->cs_create(queue
->device
->ws
,
2238 queue
->queue_family_index
? RING_COMPUTE
: RING_GFX
);
2245 radv_cs_add_buffer(queue
->device
->ws
, cs
, scratch_bo
);
2247 if (descriptor_bo
!= queue
->descriptor_bo
) {
2248 uint32_t *map
= (uint32_t*)queue
->device
->ws
->buffer_map(descriptor_bo
);
2251 uint64_t scratch_va
= radv_buffer_get_va(scratch_bo
);
2252 uint32_t rsrc1
= S_008F04_BASE_ADDRESS_HI(scratch_va
>> 32) |
2253 S_008F04_SWIZZLE_ENABLE(1);
2254 map
[0] = scratch_va
;
2258 if (esgs_ring_bo
|| gsvs_ring_bo
|| tess_rings_bo
||
2259 add_sample_positions
)
2260 fill_geom_tess_rings(queue
, map
, add_sample_positions
,
2261 esgs_ring_size
, esgs_ring_bo
,
2262 gsvs_ring_size
, gsvs_ring_bo
,
2263 tess_factor_ring_size
,
2264 tess_offchip_ring_offset
,
2265 tess_offchip_ring_size
,
2268 queue
->device
->ws
->buffer_unmap(descriptor_bo
);
2271 if (esgs_ring_bo
|| gsvs_ring_bo
|| tess_rings_bo
) {
2272 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
2273 radeon_emit(cs
, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
2274 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
2275 radeon_emit(cs
, EVENT_TYPE(V_028A90_VGT_FLUSH
) | EVENT_INDEX(0));
2278 radv_emit_gs_ring_sizes(queue
, cs
, esgs_ring_bo
, esgs_ring_size
,
2279 gsvs_ring_bo
, gsvs_ring_size
);
2280 radv_emit_tess_factor_ring(queue
, cs
, hs_offchip_param
,
2281 tess_factor_ring_size
, tess_rings_bo
);
2282 radv_emit_global_shader_pointers(queue
, cs
, descriptor_bo
);
2283 radv_emit_compute_scratch(queue
, cs
, compute_scratch_bo
);
2286 si_cs_emit_cache_flush(cs
,
2287 queue
->device
->physical_device
->rad_info
.chip_class
,
2289 queue
->queue_family_index
== RING_COMPUTE
&&
2290 queue
->device
->physical_device
->rad_info
.chip_class
>= CIK
,
2291 (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
)) |
2292 RADV_CMD_FLAG_INV_ICACHE
|
2293 RADV_CMD_FLAG_INV_SMEM_L1
|
2294 RADV_CMD_FLAG_INV_VMEM_L1
|
2295 RADV_CMD_FLAG_INV_GLOBAL_L2
|
2296 RADV_CMD_FLAG_START_PIPELINE_STATS
, 0);
2297 } else if (i
== 1) {
2298 si_cs_emit_cache_flush(cs
,
2299 queue
->device
->physical_device
->rad_info
.chip_class
,
2301 queue
->queue_family_index
== RING_COMPUTE
&&
2302 queue
->device
->physical_device
->rad_info
.chip_class
>= CIK
,
2303 RADV_CMD_FLAG_INV_ICACHE
|
2304 RADV_CMD_FLAG_INV_SMEM_L1
|
2305 RADV_CMD_FLAG_INV_VMEM_L1
|
2306 RADV_CMD_FLAG_INV_GLOBAL_L2
|
2307 RADV_CMD_FLAG_START_PIPELINE_STATS
, 0);
2310 if (!queue
->device
->ws
->cs_finalize(cs
))
2314 if (queue
->initial_full_flush_preamble_cs
)
2315 queue
->device
->ws
->cs_destroy(queue
->initial_full_flush_preamble_cs
);
2317 if (queue
->initial_preamble_cs
)
2318 queue
->device
->ws
->cs_destroy(queue
->initial_preamble_cs
);
2320 if (queue
->continue_preamble_cs
)
2321 queue
->device
->ws
->cs_destroy(queue
->continue_preamble_cs
);
2323 queue
->initial_full_flush_preamble_cs
= dest_cs
[0];
2324 queue
->initial_preamble_cs
= dest_cs
[1];
2325 queue
->continue_preamble_cs
= dest_cs
[2];
2327 if (scratch_bo
!= queue
->scratch_bo
) {
2328 if (queue
->scratch_bo
)
2329 queue
->device
->ws
->buffer_destroy(queue
->scratch_bo
);
2330 queue
->scratch_bo
= scratch_bo
;
2331 queue
->scratch_size
= scratch_size
;
2334 if (compute_scratch_bo
!= queue
->compute_scratch_bo
) {
2335 if (queue
->compute_scratch_bo
)
2336 queue
->device
->ws
->buffer_destroy(queue
->compute_scratch_bo
);
2337 queue
->compute_scratch_bo
= compute_scratch_bo
;
2338 queue
->compute_scratch_size
= compute_scratch_size
;
2341 if (esgs_ring_bo
!= queue
->esgs_ring_bo
) {
2342 if (queue
->esgs_ring_bo
)
2343 queue
->device
->ws
->buffer_destroy(queue
->esgs_ring_bo
);
2344 queue
->esgs_ring_bo
= esgs_ring_bo
;
2345 queue
->esgs_ring_size
= esgs_ring_size
;
2348 if (gsvs_ring_bo
!= queue
->gsvs_ring_bo
) {
2349 if (queue
->gsvs_ring_bo
)
2350 queue
->device
->ws
->buffer_destroy(queue
->gsvs_ring_bo
);
2351 queue
->gsvs_ring_bo
= gsvs_ring_bo
;
2352 queue
->gsvs_ring_size
= gsvs_ring_size
;
2355 if (tess_rings_bo
!= queue
->tess_rings_bo
) {
2356 queue
->tess_rings_bo
= tess_rings_bo
;
2357 queue
->has_tess_rings
= true;
2360 if (descriptor_bo
!= queue
->descriptor_bo
) {
2361 if (queue
->descriptor_bo
)
2362 queue
->device
->ws
->buffer_destroy(queue
->descriptor_bo
);
2364 queue
->descriptor_bo
= descriptor_bo
;
2367 if (add_sample_positions
)
2368 queue
->has_sample_positions
= true;
2370 *initial_full_flush_preamble_cs
= queue
->initial_full_flush_preamble_cs
;
2371 *initial_preamble_cs
= queue
->initial_preamble_cs
;
2372 *continue_preamble_cs
= queue
->continue_preamble_cs
;
2373 if (!scratch_size
&& !compute_scratch_size
&& !esgs_ring_size
&& !gsvs_ring_size
)
2374 *continue_preamble_cs
= NULL
;
2377 for (int i
= 0; i
< ARRAY_SIZE(dest_cs
); ++i
)
2379 queue
->device
->ws
->cs_destroy(dest_cs
[i
]);
2380 if (descriptor_bo
&& descriptor_bo
!= queue
->descriptor_bo
)
2381 queue
->device
->ws
->buffer_destroy(descriptor_bo
);
2382 if (scratch_bo
&& scratch_bo
!= queue
->scratch_bo
)
2383 queue
->device
->ws
->buffer_destroy(scratch_bo
);
2384 if (compute_scratch_bo
&& compute_scratch_bo
!= queue
->compute_scratch_bo
)
2385 queue
->device
->ws
->buffer_destroy(compute_scratch_bo
);
2386 if (esgs_ring_bo
&& esgs_ring_bo
!= queue
->esgs_ring_bo
)
2387 queue
->device
->ws
->buffer_destroy(esgs_ring_bo
);
2388 if (gsvs_ring_bo
&& gsvs_ring_bo
!= queue
->gsvs_ring_bo
)
2389 queue
->device
->ws
->buffer_destroy(gsvs_ring_bo
);
2390 if (tess_rings_bo
&& tess_rings_bo
!= queue
->tess_rings_bo
)
2391 queue
->device
->ws
->buffer_destroy(tess_rings_bo
);
2392 return vk_error(queue
->device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
2395 static VkResult
radv_alloc_sem_counts(struct radv_instance
*instance
,
2396 struct radv_winsys_sem_counts
*counts
,
2398 const VkSemaphore
*sems
,
2402 int syncobj_idx
= 0, sem_idx
= 0;
2404 if (num_sems
== 0 && _fence
== VK_NULL_HANDLE
)
2407 for (uint32_t i
= 0; i
< num_sems
; i
++) {
2408 RADV_FROM_HANDLE(radv_semaphore
, sem
, sems
[i
]);
2410 if (sem
->temp_syncobj
|| sem
->syncobj
)
2411 counts
->syncobj_count
++;
2413 counts
->sem_count
++;
2416 if (_fence
!= VK_NULL_HANDLE
) {
2417 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
2418 if (fence
->temp_syncobj
|| fence
->syncobj
)
2419 counts
->syncobj_count
++;
2422 if (counts
->syncobj_count
) {
2423 counts
->syncobj
= (uint32_t *)malloc(sizeof(uint32_t) * counts
->syncobj_count
);
2424 if (!counts
->syncobj
)
2425 return vk_error(instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
2428 if (counts
->sem_count
) {
2429 counts
->sem
= (struct radeon_winsys_sem
**)malloc(sizeof(struct radeon_winsys_sem
*) * counts
->sem_count
);
2431 free(counts
->syncobj
);
2432 return vk_error(instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
2436 for (uint32_t i
= 0; i
< num_sems
; i
++) {
2437 RADV_FROM_HANDLE(radv_semaphore
, sem
, sems
[i
]);
2439 if (sem
->temp_syncobj
) {
2440 counts
->syncobj
[syncobj_idx
++] = sem
->temp_syncobj
;
2442 else if (sem
->syncobj
)
2443 counts
->syncobj
[syncobj_idx
++] = sem
->syncobj
;
2446 counts
->sem
[sem_idx
++] = sem
->sem
;
2450 if (_fence
!= VK_NULL_HANDLE
) {
2451 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
2452 if (fence
->temp_syncobj
)
2453 counts
->syncobj
[syncobj_idx
++] = fence
->temp_syncobj
;
2454 else if (fence
->syncobj
)
2455 counts
->syncobj
[syncobj_idx
++] = fence
->syncobj
;
2462 radv_free_sem_info(struct radv_winsys_sem_info
*sem_info
)
2464 free(sem_info
->wait
.syncobj
);
2465 free(sem_info
->wait
.sem
);
2466 free(sem_info
->signal
.syncobj
);
2467 free(sem_info
->signal
.sem
);
2471 static void radv_free_temp_syncobjs(struct radv_device
*device
,
2473 const VkSemaphore
*sems
)
2475 for (uint32_t i
= 0; i
< num_sems
; i
++) {
2476 RADV_FROM_HANDLE(radv_semaphore
, sem
, sems
[i
]);
2478 if (sem
->temp_syncobj
) {
2479 device
->ws
->destroy_syncobj(device
->ws
, sem
->temp_syncobj
);
2480 sem
->temp_syncobj
= 0;
2486 radv_alloc_sem_info(struct radv_instance
*instance
,
2487 struct radv_winsys_sem_info
*sem_info
,
2489 const VkSemaphore
*wait_sems
,
2490 int num_signal_sems
,
2491 const VkSemaphore
*signal_sems
,
2495 memset(sem_info
, 0, sizeof(*sem_info
));
2497 ret
= radv_alloc_sem_counts(instance
, &sem_info
->wait
, num_wait_sems
, wait_sems
, VK_NULL_HANDLE
, true);
2500 ret
= radv_alloc_sem_counts(instance
, &sem_info
->signal
, num_signal_sems
, signal_sems
, fence
, false);
2502 radv_free_sem_info(sem_info
);
2504 /* caller can override these */
2505 sem_info
->cs_emit_wait
= true;
2506 sem_info
->cs_emit_signal
= true;
2510 /* Signals fence as soon as all the work currently put on queue is done. */
2511 static VkResult
radv_signal_fence(struct radv_queue
*queue
,
2512 struct radv_fence
*fence
)
2516 struct radv_winsys_sem_info sem_info
;
2518 result
= radv_alloc_sem_info(queue
->device
->instance
, &sem_info
, 0, NULL
, 0, NULL
,
2519 radv_fence_to_handle(fence
));
2520 if (result
!= VK_SUCCESS
)
2523 ret
= queue
->device
->ws
->cs_submit(queue
->hw_ctx
, queue
->queue_idx
,
2524 &queue
->device
->empty_cs
[queue
->queue_family_index
],
2525 1, NULL
, NULL
, &sem_info
, NULL
,
2526 false, fence
->fence
);
2527 radv_free_sem_info(&sem_info
);
2530 return vk_error(queue
->device
->instance
, VK_ERROR_DEVICE_LOST
);
2535 VkResult
radv_QueueSubmit(
2537 uint32_t submitCount
,
2538 const VkSubmitInfo
* pSubmits
,
2541 RADV_FROM_HANDLE(radv_queue
, queue
, _queue
);
2542 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
2543 struct radeon_winsys_fence
*base_fence
= fence
? fence
->fence
: NULL
;
2544 struct radeon_winsys_ctx
*ctx
= queue
->hw_ctx
;
2546 uint32_t max_cs_submission
= queue
->device
->trace_bo
? 1 : UINT32_MAX
;
2547 uint32_t scratch_size
= 0;
2548 uint32_t compute_scratch_size
= 0;
2549 uint32_t esgs_ring_size
= 0, gsvs_ring_size
= 0;
2550 struct radeon_cmdbuf
*initial_preamble_cs
= NULL
, *initial_flush_preamble_cs
= NULL
, *continue_preamble_cs
= NULL
;
2552 bool fence_emitted
= false;
2553 bool tess_rings_needed
= false;
2554 bool sample_positions_needed
= false;
2556 /* Do this first so failing to allocate scratch buffers can't result in
2557 * partially executed submissions. */
2558 for (uint32_t i
= 0; i
< submitCount
; i
++) {
2559 for (uint32_t j
= 0; j
< pSubmits
[i
].commandBufferCount
; j
++) {
2560 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
,
2561 pSubmits
[i
].pCommandBuffers
[j
]);
2563 scratch_size
= MAX2(scratch_size
, cmd_buffer
->scratch_size_needed
);
2564 compute_scratch_size
= MAX2(compute_scratch_size
,
2565 cmd_buffer
->compute_scratch_size_needed
);
2566 esgs_ring_size
= MAX2(esgs_ring_size
, cmd_buffer
->esgs_ring_size_needed
);
2567 gsvs_ring_size
= MAX2(gsvs_ring_size
, cmd_buffer
->gsvs_ring_size_needed
);
2568 tess_rings_needed
|= cmd_buffer
->tess_rings_needed
;
2569 sample_positions_needed
|= cmd_buffer
->sample_positions_needed
;
2573 result
= radv_get_preamble_cs(queue
, scratch_size
, compute_scratch_size
,
2574 esgs_ring_size
, gsvs_ring_size
, tess_rings_needed
,
2575 sample_positions_needed
, &initial_flush_preamble_cs
,
2576 &initial_preamble_cs
, &continue_preamble_cs
);
2577 if (result
!= VK_SUCCESS
)
2580 for (uint32_t i
= 0; i
< submitCount
; i
++) {
2581 struct radeon_cmdbuf
**cs_array
;
2582 bool do_flush
= !i
|| pSubmits
[i
].pWaitDstStageMask
;
2583 bool can_patch
= true;
2585 struct radv_winsys_sem_info sem_info
;
2587 result
= radv_alloc_sem_info(queue
->device
->instance
,
2589 pSubmits
[i
].waitSemaphoreCount
,
2590 pSubmits
[i
].pWaitSemaphores
,
2591 pSubmits
[i
].signalSemaphoreCount
,
2592 pSubmits
[i
].pSignalSemaphores
,
2594 if (result
!= VK_SUCCESS
)
2597 if (!pSubmits
[i
].commandBufferCount
) {
2598 if (pSubmits
[i
].waitSemaphoreCount
|| pSubmits
[i
].signalSemaphoreCount
) {
2599 ret
= queue
->device
->ws
->cs_submit(ctx
, queue
->queue_idx
,
2600 &queue
->device
->empty_cs
[queue
->queue_family_index
],
2605 radv_loge("failed to submit CS %d\n", i
);
2608 fence_emitted
= true;
2610 radv_free_sem_info(&sem_info
);
2614 cs_array
= malloc(sizeof(struct radeon_cmdbuf
*) *
2615 (pSubmits
[i
].commandBufferCount
));
2617 for (uint32_t j
= 0; j
< pSubmits
[i
].commandBufferCount
; j
++) {
2618 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
,
2619 pSubmits
[i
].pCommandBuffers
[j
]);
2620 assert(cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_PRIMARY
);
2622 cs_array
[j
] = cmd_buffer
->cs
;
2623 if ((cmd_buffer
->usage_flags
& VK_COMMAND_BUFFER_USAGE_SIMULTANEOUS_USE_BIT
))
2626 cmd_buffer
->status
= RADV_CMD_BUFFER_STATUS_PENDING
;
2629 for (uint32_t j
= 0; j
< pSubmits
[i
].commandBufferCount
; j
+= advance
) {
2630 struct radeon_cmdbuf
*initial_preamble
= (do_flush
&& !j
) ? initial_flush_preamble_cs
: initial_preamble_cs
;
2631 const struct radv_winsys_bo_list
*bo_list
= NULL
;
2633 advance
= MIN2(max_cs_submission
,
2634 pSubmits
[i
].commandBufferCount
- j
);
2636 if (queue
->device
->trace_bo
)
2637 *queue
->device
->trace_id_ptr
= 0;
2639 sem_info
.cs_emit_wait
= j
== 0;
2640 sem_info
.cs_emit_signal
= j
+ advance
== pSubmits
[i
].commandBufferCount
;
2642 if (unlikely(queue
->device
->use_global_bo_list
)) {
2643 pthread_mutex_lock(&queue
->device
->bo_list
.mutex
);
2644 bo_list
= &queue
->device
->bo_list
.list
;
2647 ret
= queue
->device
->ws
->cs_submit(ctx
, queue
->queue_idx
, cs_array
+ j
,
2648 advance
, initial_preamble
, continue_preamble_cs
,
2650 can_patch
, base_fence
);
2652 if (unlikely(queue
->device
->use_global_bo_list
))
2653 pthread_mutex_unlock(&queue
->device
->bo_list
.mutex
);
2656 radv_loge("failed to submit CS %d\n", i
);
2659 fence_emitted
= true;
2660 if (queue
->device
->trace_bo
) {
2661 radv_check_gpu_hangs(queue
, cs_array
[j
]);
2665 radv_free_temp_syncobjs(queue
->device
,
2666 pSubmits
[i
].waitSemaphoreCount
,
2667 pSubmits
[i
].pWaitSemaphores
);
2668 radv_free_sem_info(&sem_info
);
2673 if (!fence_emitted
) {
2674 result
= radv_signal_fence(queue
, fence
);
2675 if (result
!= VK_SUCCESS
)
2678 fence
->submitted
= true;
2684 VkResult
radv_QueueWaitIdle(
2687 RADV_FROM_HANDLE(radv_queue
, queue
, _queue
);
2689 queue
->device
->ws
->ctx_wait_idle(queue
->hw_ctx
,
2690 radv_queue_family_to_ring(queue
->queue_family_index
),
2695 VkResult
radv_DeviceWaitIdle(
2698 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2700 for (unsigned i
= 0; i
< RADV_MAX_QUEUE_FAMILIES
; i
++) {
2701 for (unsigned q
= 0; q
< device
->queue_count
[i
]; q
++) {
2702 radv_QueueWaitIdle(radv_queue_to_handle(&device
->queues
[i
][q
]));
2708 VkResult
radv_EnumerateInstanceExtensionProperties(
2709 const char* pLayerName
,
2710 uint32_t* pPropertyCount
,
2711 VkExtensionProperties
* pProperties
)
2713 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
2715 for (int i
= 0; i
< RADV_INSTANCE_EXTENSION_COUNT
; i
++) {
2716 if (radv_supported_instance_extensions
.extensions
[i
]) {
2717 vk_outarray_append(&out
, prop
) {
2718 *prop
= radv_instance_extensions
[i
];
2723 return vk_outarray_status(&out
);
2726 VkResult
radv_EnumerateDeviceExtensionProperties(
2727 VkPhysicalDevice physicalDevice
,
2728 const char* pLayerName
,
2729 uint32_t* pPropertyCount
,
2730 VkExtensionProperties
* pProperties
)
2732 RADV_FROM_HANDLE(radv_physical_device
, device
, physicalDevice
);
2733 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
2735 for (int i
= 0; i
< RADV_DEVICE_EXTENSION_COUNT
; i
++) {
2736 if (device
->supported_extensions
.extensions
[i
]) {
2737 vk_outarray_append(&out
, prop
) {
2738 *prop
= radv_device_extensions
[i
];
2743 return vk_outarray_status(&out
);
2746 PFN_vkVoidFunction
radv_GetInstanceProcAddr(
2747 VkInstance _instance
,
2750 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
2752 return radv_lookup_entrypoint_checked(pName
,
2753 instance
? instance
->apiVersion
: 0,
2754 instance
? &instance
->enabled_extensions
: NULL
,
2758 /* The loader wants us to expose a second GetInstanceProcAddr function
2759 * to work around certain LD_PRELOAD issues seen in apps.
2762 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
2763 VkInstance instance
,
2767 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
2768 VkInstance instance
,
2771 return radv_GetInstanceProcAddr(instance
, pName
);
2774 PFN_vkVoidFunction
radv_GetDeviceProcAddr(
2778 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2780 return radv_lookup_entrypoint_checked(pName
,
2781 device
->instance
->apiVersion
,
2782 &device
->instance
->enabled_extensions
,
2783 &device
->enabled_extensions
);
2786 bool radv_get_memory_fd(struct radv_device
*device
,
2787 struct radv_device_memory
*memory
,
2790 struct radeon_bo_metadata metadata
;
2792 if (memory
->image
) {
2793 radv_init_metadata(device
, memory
->image
, &metadata
);
2794 device
->ws
->buffer_set_metadata(memory
->bo
, &metadata
);
2797 return device
->ws
->buffer_get_fd(device
->ws
, memory
->bo
,
2801 static VkResult
radv_alloc_memory(struct radv_device
*device
,
2802 const VkMemoryAllocateInfo
* pAllocateInfo
,
2803 const VkAllocationCallbacks
* pAllocator
,
2804 VkDeviceMemory
* pMem
)
2806 struct radv_device_memory
*mem
;
2808 enum radeon_bo_domain domain
;
2810 enum radv_mem_type mem_type_index
= device
->physical_device
->mem_type_indices
[pAllocateInfo
->memoryTypeIndex
];
2812 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
2814 if (pAllocateInfo
->allocationSize
== 0) {
2815 /* Apparently, this is allowed */
2816 *pMem
= VK_NULL_HANDLE
;
2820 const VkImportMemoryFdInfoKHR
*import_info
=
2821 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_FD_INFO_KHR
);
2822 const VkMemoryDedicatedAllocateInfoKHR
*dedicate_info
=
2823 vk_find_struct_const(pAllocateInfo
->pNext
, MEMORY_DEDICATED_ALLOCATE_INFO_KHR
);
2824 const VkExportMemoryAllocateInfoKHR
*export_info
=
2825 vk_find_struct_const(pAllocateInfo
->pNext
, EXPORT_MEMORY_ALLOCATE_INFO_KHR
);
2826 const VkImportMemoryHostPointerInfoEXT
*host_ptr_info
=
2827 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_HOST_POINTER_INFO_EXT
);
2829 const struct wsi_memory_allocate_info
*wsi_info
=
2830 vk_find_struct_const(pAllocateInfo
->pNext
, WSI_MEMORY_ALLOCATE_INFO_MESA
);
2832 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
2833 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2835 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
2837 if (wsi_info
&& wsi_info
->implicit_sync
)
2838 flags
|= RADEON_FLAG_IMPLICIT_SYNC
;
2840 if (dedicate_info
) {
2841 mem
->image
= radv_image_from_handle(dedicate_info
->image
);
2842 mem
->buffer
= radv_buffer_from_handle(dedicate_info
->buffer
);
2848 mem
->user_ptr
= NULL
;
2851 assert(import_info
->handleType
==
2852 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
||
2853 import_info
->handleType
==
2854 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
2855 mem
->bo
= device
->ws
->buffer_from_fd(device
->ws
, import_info
->fd
,
2858 result
= VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
;
2861 close(import_info
->fd
);
2863 } else if (host_ptr_info
) {
2864 assert(host_ptr_info
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
);
2865 assert(mem_type_index
== RADV_MEM_TYPE_GTT_CACHED
);
2866 mem
->bo
= device
->ws
->buffer_from_ptr(device
->ws
, host_ptr_info
->pHostPointer
,
2867 pAllocateInfo
->allocationSize
);
2869 result
= VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
;
2872 mem
->user_ptr
= host_ptr_info
->pHostPointer
;
2875 uint64_t alloc_size
= align_u64(pAllocateInfo
->allocationSize
, 4096);
2876 if (mem_type_index
== RADV_MEM_TYPE_GTT_WRITE_COMBINE
||
2877 mem_type_index
== RADV_MEM_TYPE_GTT_CACHED
)
2878 domain
= RADEON_DOMAIN_GTT
;
2880 domain
= RADEON_DOMAIN_VRAM
;
2882 if (mem_type_index
== RADV_MEM_TYPE_VRAM
)
2883 flags
|= RADEON_FLAG_NO_CPU_ACCESS
;
2885 flags
|= RADEON_FLAG_CPU_ACCESS
;
2887 if (mem_type_index
== RADV_MEM_TYPE_GTT_WRITE_COMBINE
)
2888 flags
|= RADEON_FLAG_GTT_WC
;
2890 if (!dedicate_info
&& !import_info
&& (!export_info
|| !export_info
->handleTypes
))
2891 flags
|= RADEON_FLAG_NO_INTERPROCESS_SHARING
;
2893 mem
->bo
= device
->ws
->buffer_create(device
->ws
, alloc_size
, device
->physical_device
->rad_info
.max_alignment
,
2897 result
= VK_ERROR_OUT_OF_DEVICE_MEMORY
;
2900 mem
->type_index
= mem_type_index
;
2903 result
= radv_bo_list_add(device
, mem
->bo
);
2904 if (result
!= VK_SUCCESS
)
2907 *pMem
= radv_device_memory_to_handle(mem
);
2912 device
->ws
->buffer_destroy(mem
->bo
);
2914 vk_free2(&device
->alloc
, pAllocator
, mem
);
2919 VkResult
radv_AllocateMemory(
2921 const VkMemoryAllocateInfo
* pAllocateInfo
,
2922 const VkAllocationCallbacks
* pAllocator
,
2923 VkDeviceMemory
* pMem
)
2925 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2926 return radv_alloc_memory(device
, pAllocateInfo
, pAllocator
, pMem
);
2929 void radv_FreeMemory(
2931 VkDeviceMemory _mem
,
2932 const VkAllocationCallbacks
* pAllocator
)
2934 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2935 RADV_FROM_HANDLE(radv_device_memory
, mem
, _mem
);
2940 radv_bo_list_remove(device
, mem
->bo
);
2941 device
->ws
->buffer_destroy(mem
->bo
);
2944 vk_free2(&device
->alloc
, pAllocator
, mem
);
2947 VkResult
radv_MapMemory(
2949 VkDeviceMemory _memory
,
2950 VkDeviceSize offset
,
2952 VkMemoryMapFlags flags
,
2955 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2956 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
2964 *ppData
= mem
->user_ptr
;
2966 *ppData
= device
->ws
->buffer_map(mem
->bo
);
2973 return vk_error(device
->instance
, VK_ERROR_MEMORY_MAP_FAILED
);
2976 void radv_UnmapMemory(
2978 VkDeviceMemory _memory
)
2980 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2981 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
2986 if (mem
->user_ptr
== NULL
)
2987 device
->ws
->buffer_unmap(mem
->bo
);
2990 VkResult
radv_FlushMappedMemoryRanges(
2992 uint32_t memoryRangeCount
,
2993 const VkMappedMemoryRange
* pMemoryRanges
)
2998 VkResult
radv_InvalidateMappedMemoryRanges(
3000 uint32_t memoryRangeCount
,
3001 const VkMappedMemoryRange
* pMemoryRanges
)
3006 void radv_GetBufferMemoryRequirements(
3009 VkMemoryRequirements
* pMemoryRequirements
)
3011 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3012 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
3014 pMemoryRequirements
->memoryTypeBits
= (1u << device
->physical_device
->memory_properties
.memoryTypeCount
) - 1;
3016 if (buffer
->flags
& VK_BUFFER_CREATE_SPARSE_BINDING_BIT
)
3017 pMemoryRequirements
->alignment
= 4096;
3019 pMemoryRequirements
->alignment
= 16;
3021 pMemoryRequirements
->size
= align64(buffer
->size
, pMemoryRequirements
->alignment
);
3024 void radv_GetBufferMemoryRequirements2(
3026 const VkBufferMemoryRequirementsInfo2KHR
* pInfo
,
3027 VkMemoryRequirements2KHR
* pMemoryRequirements
)
3029 radv_GetBufferMemoryRequirements(device
, pInfo
->buffer
,
3030 &pMemoryRequirements
->memoryRequirements
);
3031 RADV_FROM_HANDLE(radv_buffer
, buffer
, pInfo
->buffer
);
3032 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3033 switch (ext
->sType
) {
3034 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS_KHR
: {
3035 VkMemoryDedicatedRequirementsKHR
*req
=
3036 (VkMemoryDedicatedRequirementsKHR
*) ext
;
3037 req
->requiresDedicatedAllocation
= buffer
->shareable
;
3038 req
->prefersDedicatedAllocation
= req
->requiresDedicatedAllocation
;
3047 void radv_GetImageMemoryRequirements(
3050 VkMemoryRequirements
* pMemoryRequirements
)
3052 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3053 RADV_FROM_HANDLE(radv_image
, image
, _image
);
3055 pMemoryRequirements
->memoryTypeBits
= (1u << device
->physical_device
->memory_properties
.memoryTypeCount
) - 1;
3057 pMemoryRequirements
->size
= image
->size
;
3058 pMemoryRequirements
->alignment
= image
->alignment
;
3061 void radv_GetImageMemoryRequirements2(
3063 const VkImageMemoryRequirementsInfo2KHR
* pInfo
,
3064 VkMemoryRequirements2KHR
* pMemoryRequirements
)
3066 radv_GetImageMemoryRequirements(device
, pInfo
->image
,
3067 &pMemoryRequirements
->memoryRequirements
);
3069 RADV_FROM_HANDLE(radv_image
, image
, pInfo
->image
);
3071 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3072 switch (ext
->sType
) {
3073 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS_KHR
: {
3074 VkMemoryDedicatedRequirementsKHR
*req
=
3075 (VkMemoryDedicatedRequirementsKHR
*) ext
;
3076 req
->requiresDedicatedAllocation
= image
->shareable
;
3077 req
->prefersDedicatedAllocation
= req
->requiresDedicatedAllocation
;
3086 void radv_GetImageSparseMemoryRequirements(
3089 uint32_t* pSparseMemoryRequirementCount
,
3090 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
3095 void radv_GetImageSparseMemoryRequirements2(
3097 const VkImageSparseMemoryRequirementsInfo2KHR
* pInfo
,
3098 uint32_t* pSparseMemoryRequirementCount
,
3099 VkSparseImageMemoryRequirements2KHR
* pSparseMemoryRequirements
)
3104 void radv_GetDeviceMemoryCommitment(
3106 VkDeviceMemory memory
,
3107 VkDeviceSize
* pCommittedMemoryInBytes
)
3109 *pCommittedMemoryInBytes
= 0;
3112 VkResult
radv_BindBufferMemory2(VkDevice device
,
3113 uint32_t bindInfoCount
,
3114 const VkBindBufferMemoryInfoKHR
*pBindInfos
)
3116 for (uint32_t i
= 0; i
< bindInfoCount
; ++i
) {
3117 RADV_FROM_HANDLE(radv_device_memory
, mem
, pBindInfos
[i
].memory
);
3118 RADV_FROM_HANDLE(radv_buffer
, buffer
, pBindInfos
[i
].buffer
);
3121 buffer
->bo
= mem
->bo
;
3122 buffer
->offset
= pBindInfos
[i
].memoryOffset
;
3130 VkResult
radv_BindBufferMemory(
3133 VkDeviceMemory memory
,
3134 VkDeviceSize memoryOffset
)
3136 const VkBindBufferMemoryInfoKHR info
= {
3137 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO_KHR
,
3140 .memoryOffset
= memoryOffset
3143 return radv_BindBufferMemory2(device
, 1, &info
);
3146 VkResult
radv_BindImageMemory2(VkDevice device
,
3147 uint32_t bindInfoCount
,
3148 const VkBindImageMemoryInfoKHR
*pBindInfos
)
3150 for (uint32_t i
= 0; i
< bindInfoCount
; ++i
) {
3151 RADV_FROM_HANDLE(radv_device_memory
, mem
, pBindInfos
[i
].memory
);
3152 RADV_FROM_HANDLE(radv_image
, image
, pBindInfos
[i
].image
);
3155 image
->bo
= mem
->bo
;
3156 image
->offset
= pBindInfos
[i
].memoryOffset
;
3166 VkResult
radv_BindImageMemory(
3169 VkDeviceMemory memory
,
3170 VkDeviceSize memoryOffset
)
3172 const VkBindImageMemoryInfoKHR info
= {
3173 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO_KHR
,
3176 .memoryOffset
= memoryOffset
3179 return radv_BindImageMemory2(device
, 1, &info
);
3184 radv_sparse_buffer_bind_memory(struct radv_device
*device
,
3185 const VkSparseBufferMemoryBindInfo
*bind
)
3187 RADV_FROM_HANDLE(radv_buffer
, buffer
, bind
->buffer
);
3189 for (uint32_t i
= 0; i
< bind
->bindCount
; ++i
) {
3190 struct radv_device_memory
*mem
= NULL
;
3192 if (bind
->pBinds
[i
].memory
!= VK_NULL_HANDLE
)
3193 mem
= radv_device_memory_from_handle(bind
->pBinds
[i
].memory
);
3195 device
->ws
->buffer_virtual_bind(buffer
->bo
,
3196 bind
->pBinds
[i
].resourceOffset
,
3197 bind
->pBinds
[i
].size
,
3198 mem
? mem
->bo
: NULL
,
3199 bind
->pBinds
[i
].memoryOffset
);
3204 radv_sparse_image_opaque_bind_memory(struct radv_device
*device
,
3205 const VkSparseImageOpaqueMemoryBindInfo
*bind
)
3207 RADV_FROM_HANDLE(radv_image
, image
, bind
->image
);
3209 for (uint32_t i
= 0; i
< bind
->bindCount
; ++i
) {
3210 struct radv_device_memory
*mem
= NULL
;
3212 if (bind
->pBinds
[i
].memory
!= VK_NULL_HANDLE
)
3213 mem
= radv_device_memory_from_handle(bind
->pBinds
[i
].memory
);
3215 device
->ws
->buffer_virtual_bind(image
->bo
,
3216 bind
->pBinds
[i
].resourceOffset
,
3217 bind
->pBinds
[i
].size
,
3218 mem
? mem
->bo
: NULL
,
3219 bind
->pBinds
[i
].memoryOffset
);
3223 VkResult
radv_QueueBindSparse(
3225 uint32_t bindInfoCount
,
3226 const VkBindSparseInfo
* pBindInfo
,
3229 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
3230 RADV_FROM_HANDLE(radv_queue
, queue
, _queue
);
3231 struct radeon_winsys_fence
*base_fence
= fence
? fence
->fence
: NULL
;
3232 bool fence_emitted
= false;
3236 for (uint32_t i
= 0; i
< bindInfoCount
; ++i
) {
3237 struct radv_winsys_sem_info sem_info
;
3238 for (uint32_t j
= 0; j
< pBindInfo
[i
].bufferBindCount
; ++j
) {
3239 radv_sparse_buffer_bind_memory(queue
->device
,
3240 pBindInfo
[i
].pBufferBinds
+ j
);
3243 for (uint32_t j
= 0; j
< pBindInfo
[i
].imageOpaqueBindCount
; ++j
) {
3244 radv_sparse_image_opaque_bind_memory(queue
->device
,
3245 pBindInfo
[i
].pImageOpaqueBinds
+ j
);
3249 result
= radv_alloc_sem_info(queue
->device
->instance
,
3251 pBindInfo
[i
].waitSemaphoreCount
,
3252 pBindInfo
[i
].pWaitSemaphores
,
3253 pBindInfo
[i
].signalSemaphoreCount
,
3254 pBindInfo
[i
].pSignalSemaphores
,
3256 if (result
!= VK_SUCCESS
)
3259 if (pBindInfo
[i
].waitSemaphoreCount
|| pBindInfo
[i
].signalSemaphoreCount
) {
3260 ret
= queue
->device
->ws
->cs_submit(queue
->hw_ctx
, queue
->queue_idx
,
3261 &queue
->device
->empty_cs
[queue
->queue_family_index
],
3266 radv_loge("failed to submit CS %d\n", i
);
3270 fence_emitted
= true;
3272 fence
->submitted
= true;
3275 radv_free_sem_info(&sem_info
);
3280 if (!fence_emitted
) {
3281 result
= radv_signal_fence(queue
, fence
);
3282 if (result
!= VK_SUCCESS
)
3285 fence
->submitted
= true;
3291 VkResult
radv_CreateFence(
3293 const VkFenceCreateInfo
* pCreateInfo
,
3294 const VkAllocationCallbacks
* pAllocator
,
3297 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3298 const VkExportFenceCreateInfoKHR
*export
=
3299 vk_find_struct_const(pCreateInfo
->pNext
, EXPORT_FENCE_CREATE_INFO_KHR
);
3300 VkExternalFenceHandleTypeFlagsKHR handleTypes
=
3301 export
? export
->handleTypes
: 0;
3303 struct radv_fence
*fence
= vk_alloc2(&device
->alloc
, pAllocator
,
3305 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3308 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3310 fence
->fence_wsi
= NULL
;
3311 fence
->submitted
= false;
3312 fence
->signalled
= !!(pCreateInfo
->flags
& VK_FENCE_CREATE_SIGNALED_BIT
);
3313 fence
->temp_syncobj
= 0;
3314 if (device
->always_use_syncobj
|| handleTypes
) {
3315 int ret
= device
->ws
->create_syncobj(device
->ws
, &fence
->syncobj
);
3317 vk_free2(&device
->alloc
, pAllocator
, fence
);
3318 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3320 if (pCreateInfo
->flags
& VK_FENCE_CREATE_SIGNALED_BIT
) {
3321 device
->ws
->signal_syncobj(device
->ws
, fence
->syncobj
);
3323 fence
->fence
= NULL
;
3325 fence
->fence
= device
->ws
->create_fence();
3326 if (!fence
->fence
) {
3327 vk_free2(&device
->alloc
, pAllocator
, fence
);
3328 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3333 *pFence
= radv_fence_to_handle(fence
);
3338 void radv_DestroyFence(
3341 const VkAllocationCallbacks
* pAllocator
)
3343 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3344 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
3349 if (fence
->temp_syncobj
)
3350 device
->ws
->destroy_syncobj(device
->ws
, fence
->temp_syncobj
);
3352 device
->ws
->destroy_syncobj(device
->ws
, fence
->syncobj
);
3354 device
->ws
->destroy_fence(fence
->fence
);
3355 if (fence
->fence_wsi
)
3356 fence
->fence_wsi
->destroy(fence
->fence_wsi
);
3357 vk_free2(&device
->alloc
, pAllocator
, fence
);
3361 static uint64_t radv_get_current_time()
3364 clock_gettime(CLOCK_MONOTONIC
, &tv
);
3365 return tv
.tv_nsec
+ tv
.tv_sec
*1000000000ull;
3368 static uint64_t radv_get_absolute_timeout(uint64_t timeout
)
3370 uint64_t current_time
= radv_get_current_time();
3372 timeout
= MIN2(UINT64_MAX
- current_time
, timeout
);
3374 return current_time
+ timeout
;
3378 static bool radv_all_fences_plain_and_submitted(uint32_t fenceCount
, const VkFence
*pFences
)
3380 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
3381 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
3382 if (fence
->fence
== NULL
|| fence
->syncobj
||
3383 fence
->temp_syncobj
||
3384 (!fence
->signalled
&& !fence
->submitted
))
3390 static bool radv_all_fences_syncobj(uint32_t fenceCount
, const VkFence
*pFences
)
3392 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
3393 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
3394 if (fence
->syncobj
== 0 && fence
->temp_syncobj
== 0)
3400 VkResult
radv_WaitForFences(
3402 uint32_t fenceCount
,
3403 const VkFence
* pFences
,
3407 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3408 timeout
= radv_get_absolute_timeout(timeout
);
3410 if (device
->always_use_syncobj
&&
3411 radv_all_fences_syncobj(fenceCount
, pFences
))
3413 uint32_t *handles
= malloc(sizeof(uint32_t) * fenceCount
);
3415 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3417 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
3418 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
3419 handles
[i
] = fence
->temp_syncobj
? fence
->temp_syncobj
: fence
->syncobj
;
3422 bool success
= device
->ws
->wait_syncobj(device
->ws
, handles
, fenceCount
, waitAll
, timeout
);
3425 return success
? VK_SUCCESS
: VK_TIMEOUT
;
3428 if (!waitAll
&& fenceCount
> 1) {
3429 /* Not doing this by default for waitAll, due to needing to allocate twice. */
3430 if (device
->physical_device
->rad_info
.drm_minor
>= 10 && radv_all_fences_plain_and_submitted(fenceCount
, pFences
)) {
3431 uint32_t wait_count
= 0;
3432 struct radeon_winsys_fence
**fences
= malloc(sizeof(struct radeon_winsys_fence
*) * fenceCount
);
3434 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3436 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
3437 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
3439 if (fence
->signalled
) {
3444 fences
[wait_count
++] = fence
->fence
;
3447 bool success
= device
->ws
->fences_wait(device
->ws
, fences
, wait_count
,
3448 waitAll
, timeout
- radv_get_current_time());
3451 return success
? VK_SUCCESS
: VK_TIMEOUT
;
3454 while(radv_get_current_time() <= timeout
) {
3455 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
3456 if (radv_GetFenceStatus(_device
, pFences
[i
]) == VK_SUCCESS
)
3463 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
3464 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
3465 bool expired
= false;
3467 if (fence
->temp_syncobj
) {
3468 if (!device
->ws
->wait_syncobj(device
->ws
, &fence
->temp_syncobj
, 1, true, timeout
))
3473 if (fence
->syncobj
) {
3474 if (!device
->ws
->wait_syncobj(device
->ws
, &fence
->syncobj
, 1, true, timeout
))
3479 if (fence
->signalled
)
3483 if (!fence
->submitted
) {
3484 while(radv_get_current_time() <= timeout
&&
3488 if (!fence
->submitted
)
3491 /* Recheck as it may have been set by
3492 * submitting operations. */
3494 if (fence
->signalled
)
3498 expired
= device
->ws
->fence_wait(device
->ws
,
3505 if (fence
->fence_wsi
) {
3506 VkResult result
= fence
->fence_wsi
->wait(fence
->fence_wsi
, timeout
);
3507 if (result
!= VK_SUCCESS
)
3511 fence
->signalled
= true;
3517 VkResult
radv_ResetFences(VkDevice _device
,
3518 uint32_t fenceCount
,
3519 const VkFence
*pFences
)
3521 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3523 for (unsigned i
= 0; i
< fenceCount
; ++i
) {
3524 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
3525 fence
->submitted
= fence
->signalled
= false;
3527 /* Per spec, we first restore the permanent payload, and then reset, so
3528 * having a temp syncobj should not skip resetting the permanent syncobj. */
3529 if (fence
->temp_syncobj
) {
3530 device
->ws
->destroy_syncobj(device
->ws
, fence
->temp_syncobj
);
3531 fence
->temp_syncobj
= 0;
3534 if (fence
->syncobj
) {
3535 device
->ws
->reset_syncobj(device
->ws
, fence
->syncobj
);
3542 VkResult
radv_GetFenceStatus(VkDevice _device
, VkFence _fence
)
3544 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3545 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
3547 if (fence
->temp_syncobj
) {
3548 bool success
= device
->ws
->wait_syncobj(device
->ws
, &fence
->temp_syncobj
, 1, true, 0);
3549 return success
? VK_SUCCESS
: VK_NOT_READY
;
3552 if (fence
->syncobj
) {
3553 bool success
= device
->ws
->wait_syncobj(device
->ws
, &fence
->syncobj
, 1, true, 0);
3554 return success
? VK_SUCCESS
: VK_NOT_READY
;
3557 if (fence
->signalled
)
3559 if (!fence
->submitted
)
3560 return VK_NOT_READY
;
3562 if (!device
->ws
->fence_wait(device
->ws
, fence
->fence
, false, 0))
3563 return VK_NOT_READY
;
3565 if (fence
->fence_wsi
) {
3566 VkResult result
= fence
->fence_wsi
->wait(fence
->fence_wsi
, 0);
3568 if (result
!= VK_SUCCESS
) {
3569 if (result
== VK_TIMEOUT
)
3570 return VK_NOT_READY
;
3578 // Queue semaphore functions
3580 VkResult
radv_CreateSemaphore(
3582 const VkSemaphoreCreateInfo
* pCreateInfo
,
3583 const VkAllocationCallbacks
* pAllocator
,
3584 VkSemaphore
* pSemaphore
)
3586 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3587 const VkExportSemaphoreCreateInfoKHR
*export
=
3588 vk_find_struct_const(pCreateInfo
->pNext
, EXPORT_SEMAPHORE_CREATE_INFO_KHR
);
3589 VkExternalSemaphoreHandleTypeFlagsKHR handleTypes
=
3590 export
? export
->handleTypes
: 0;
3592 struct radv_semaphore
*sem
= vk_alloc2(&device
->alloc
, pAllocator
,
3594 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3596 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3598 sem
->temp_syncobj
= 0;
3599 /* create a syncobject if we are going to export this semaphore */
3600 if (device
->always_use_syncobj
|| handleTypes
) {
3601 assert (device
->physical_device
->rad_info
.has_syncobj
);
3602 int ret
= device
->ws
->create_syncobj(device
->ws
, &sem
->syncobj
);
3604 vk_free2(&device
->alloc
, pAllocator
, sem
);
3605 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3609 sem
->sem
= device
->ws
->create_sem(device
->ws
);
3611 vk_free2(&device
->alloc
, pAllocator
, sem
);
3612 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3617 *pSemaphore
= radv_semaphore_to_handle(sem
);
3621 void radv_DestroySemaphore(
3623 VkSemaphore _semaphore
,
3624 const VkAllocationCallbacks
* pAllocator
)
3626 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3627 RADV_FROM_HANDLE(radv_semaphore
, sem
, _semaphore
);
3632 device
->ws
->destroy_syncobj(device
->ws
, sem
->syncobj
);
3634 device
->ws
->destroy_sem(sem
->sem
);
3635 vk_free2(&device
->alloc
, pAllocator
, sem
);
3638 VkResult
radv_CreateEvent(
3640 const VkEventCreateInfo
* pCreateInfo
,
3641 const VkAllocationCallbacks
* pAllocator
,
3644 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3645 struct radv_event
*event
= vk_alloc2(&device
->alloc
, pAllocator
,
3647 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3650 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3652 event
->bo
= device
->ws
->buffer_create(device
->ws
, 8, 8,
3654 RADEON_FLAG_VA_UNCACHED
| RADEON_FLAG_CPU_ACCESS
| RADEON_FLAG_NO_INTERPROCESS_SHARING
);
3656 vk_free2(&device
->alloc
, pAllocator
, event
);
3657 return vk_error(device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
3660 event
->map
= (uint64_t*)device
->ws
->buffer_map(event
->bo
);
3662 *pEvent
= radv_event_to_handle(event
);
3667 void radv_DestroyEvent(
3670 const VkAllocationCallbacks
* pAllocator
)
3672 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3673 RADV_FROM_HANDLE(radv_event
, event
, _event
);
3677 device
->ws
->buffer_destroy(event
->bo
);
3678 vk_free2(&device
->alloc
, pAllocator
, event
);
3681 VkResult
radv_GetEventStatus(
3685 RADV_FROM_HANDLE(radv_event
, event
, _event
);
3687 if (*event
->map
== 1)
3688 return VK_EVENT_SET
;
3689 return VK_EVENT_RESET
;
3692 VkResult
radv_SetEvent(
3696 RADV_FROM_HANDLE(radv_event
, event
, _event
);
3702 VkResult
radv_ResetEvent(
3706 RADV_FROM_HANDLE(radv_event
, event
, _event
);
3712 VkResult
radv_CreateBuffer(
3714 const VkBufferCreateInfo
* pCreateInfo
,
3715 const VkAllocationCallbacks
* pAllocator
,
3718 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3719 struct radv_buffer
*buffer
;
3721 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
3723 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
3724 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3726 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3728 buffer
->size
= pCreateInfo
->size
;
3729 buffer
->usage
= pCreateInfo
->usage
;
3732 buffer
->flags
= pCreateInfo
->flags
;
3734 buffer
->shareable
= vk_find_struct_const(pCreateInfo
->pNext
,
3735 EXTERNAL_MEMORY_BUFFER_CREATE_INFO_KHR
) != NULL
;
3737 if (pCreateInfo
->flags
& VK_BUFFER_CREATE_SPARSE_BINDING_BIT
) {
3738 buffer
->bo
= device
->ws
->buffer_create(device
->ws
,
3739 align64(buffer
->size
, 4096),
3740 4096, 0, RADEON_FLAG_VIRTUAL
);
3742 vk_free2(&device
->alloc
, pAllocator
, buffer
);
3743 return vk_error(device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
3747 *pBuffer
= radv_buffer_to_handle(buffer
);
3752 void radv_DestroyBuffer(
3755 const VkAllocationCallbacks
* pAllocator
)
3757 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3758 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
3763 if (buffer
->flags
& VK_BUFFER_CREATE_SPARSE_BINDING_BIT
)
3764 device
->ws
->buffer_destroy(buffer
->bo
);
3766 vk_free2(&device
->alloc
, pAllocator
, buffer
);
3769 static inline unsigned
3770 si_tile_mode_index(const struct radv_image
*image
, unsigned level
, bool stencil
)
3773 return image
->surface
.u
.legacy
.stencil_tiling_index
[level
];
3775 return image
->surface
.u
.legacy
.tiling_index
[level
];
3778 static uint32_t radv_surface_max_layer_count(struct radv_image_view
*iview
)
3780 return iview
->type
== VK_IMAGE_VIEW_TYPE_3D
? iview
->extent
.depth
: (iview
->base_layer
+ iview
->layer_count
);
3784 radv_init_dcc_control_reg(struct radv_device
*device
,
3785 struct radv_image_view
*iview
)
3787 unsigned max_uncompressed_block_size
= V_028C78_MAX_BLOCK_SIZE_256B
;
3788 unsigned min_compressed_block_size
= V_028C78_MIN_BLOCK_SIZE_32B
;
3789 unsigned max_compressed_block_size
;
3790 unsigned independent_64b_blocks
;
3792 if (device
->physical_device
->rad_info
.chip_class
< VI
)
3795 if (iview
->image
->info
.samples
> 1) {
3796 if (iview
->image
->surface
.bpe
== 1)
3797 max_uncompressed_block_size
= V_028C78_MAX_BLOCK_SIZE_64B
;
3798 else if (iview
->image
->surface
.bpe
== 2)
3799 max_uncompressed_block_size
= V_028C78_MAX_BLOCK_SIZE_128B
;
3802 if (!device
->physical_device
->rad_info
.has_dedicated_vram
) {
3803 /* amdvlk: [min-compressed-block-size] should be set to 32 for
3804 * dGPU and 64 for APU because all of our APUs to date use
3805 * DIMMs which have a request granularity size of 64B while all
3806 * other chips have a 32B request size.
3808 min_compressed_block_size
= V_028C78_MIN_BLOCK_SIZE_64B
;
3811 if (iview
->image
->usage
& (VK_IMAGE_USAGE_SAMPLED_BIT
|
3812 VK_IMAGE_USAGE_TRANSFER_SRC_BIT
|
3813 VK_IMAGE_USAGE_INPUT_ATTACHMENT_BIT
)) {
3814 /* If this DCC image is potentially going to be used in texture
3815 * fetches, we need some special settings.
3817 independent_64b_blocks
= 1;
3818 max_compressed_block_size
= V_028C78_MAX_BLOCK_SIZE_64B
;
3820 /* MAX_UNCOMPRESSED_BLOCK_SIZE must be >=
3821 * MAX_COMPRESSED_BLOCK_SIZE. Set MAX_COMPRESSED_BLOCK_SIZE as
3822 * big as possible for better compression state.
3824 independent_64b_blocks
= 0;
3825 max_compressed_block_size
= max_uncompressed_block_size
;
3828 return S_028C78_MAX_UNCOMPRESSED_BLOCK_SIZE(max_uncompressed_block_size
) |
3829 S_028C78_MAX_COMPRESSED_BLOCK_SIZE(max_compressed_block_size
) |
3830 S_028C78_MIN_COMPRESSED_BLOCK_SIZE(min_compressed_block_size
) |
3831 S_028C78_INDEPENDENT_64B_BLOCKS(independent_64b_blocks
);
3835 radv_initialise_color_surface(struct radv_device
*device
,
3836 struct radv_color_buffer_info
*cb
,
3837 struct radv_image_view
*iview
)
3839 const struct vk_format_description
*desc
;
3840 unsigned ntype
, format
, swap
, endian
;
3841 unsigned blend_clamp
= 0, blend_bypass
= 0;
3843 const struct radeon_surf
*surf
= &iview
->image
->surface
;
3845 desc
= vk_format_description(iview
->vk_format
);
3847 memset(cb
, 0, sizeof(*cb
));
3849 /* Intensity is implemented as Red, so treat it that way. */
3850 cb
->cb_color_attrib
= S_028C74_FORCE_DST_ALPHA_1(desc
->swizzle
[3] == VK_SWIZZLE_1
);
3852 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
;
3854 cb
->cb_color_base
= va
>> 8;
3856 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
3857 struct gfx9_surf_meta_flags meta
;
3858 if (iview
->image
->dcc_offset
)
3859 meta
= iview
->image
->surface
.u
.gfx9
.dcc
;
3861 meta
= iview
->image
->surface
.u
.gfx9
.cmask
;
3863 cb
->cb_color_attrib
|= S_028C74_COLOR_SW_MODE(iview
->image
->surface
.u
.gfx9
.surf
.swizzle_mode
) |
3864 S_028C74_FMASK_SW_MODE(iview
->image
->surface
.u
.gfx9
.fmask
.swizzle_mode
) |
3865 S_028C74_RB_ALIGNED(meta
.rb_aligned
) |
3866 S_028C74_PIPE_ALIGNED(meta
.pipe_aligned
);
3868 cb
->cb_color_base
+= iview
->image
->surface
.u
.gfx9
.surf_offset
>> 8;
3869 cb
->cb_color_base
|= iview
->image
->surface
.tile_swizzle
;
3871 const struct legacy_surf_level
*level_info
= &surf
->u
.legacy
.level
[iview
->base_mip
];
3872 unsigned pitch_tile_max
, slice_tile_max
, tile_mode_index
;
3874 cb
->cb_color_base
+= level_info
->offset
>> 8;
3875 if (level_info
->mode
== RADEON_SURF_MODE_2D
)
3876 cb
->cb_color_base
|= iview
->image
->surface
.tile_swizzle
;
3878 pitch_tile_max
= level_info
->nblk_x
/ 8 - 1;
3879 slice_tile_max
= (level_info
->nblk_x
* level_info
->nblk_y
) / 64 - 1;
3880 tile_mode_index
= si_tile_mode_index(iview
->image
, iview
->base_mip
, false);
3882 cb
->cb_color_pitch
= S_028C64_TILE_MAX(pitch_tile_max
);
3883 cb
->cb_color_slice
= S_028C68_TILE_MAX(slice_tile_max
);
3884 cb
->cb_color_cmask_slice
= iview
->image
->cmask
.slice_tile_max
;
3886 cb
->cb_color_attrib
|= S_028C74_TILE_MODE_INDEX(tile_mode_index
);
3888 if (radv_image_has_fmask(iview
->image
)) {
3889 if (device
->physical_device
->rad_info
.chip_class
>= CIK
)
3890 cb
->cb_color_pitch
|= S_028C64_FMASK_TILE_MAX(iview
->image
->fmask
.pitch_in_pixels
/ 8 - 1);
3891 cb
->cb_color_attrib
|= S_028C74_FMASK_TILE_MODE_INDEX(iview
->image
->fmask
.tile_mode_index
);
3892 cb
->cb_color_fmask_slice
= S_028C88_TILE_MAX(iview
->image
->fmask
.slice_tile_max
);
3894 /* This must be set for fast clear to work without FMASK. */
3895 if (device
->physical_device
->rad_info
.chip_class
>= CIK
)
3896 cb
->cb_color_pitch
|= S_028C64_FMASK_TILE_MAX(pitch_tile_max
);
3897 cb
->cb_color_attrib
|= S_028C74_FMASK_TILE_MODE_INDEX(tile_mode_index
);
3898 cb
->cb_color_fmask_slice
= S_028C88_TILE_MAX(slice_tile_max
);
3902 /* CMASK variables */
3903 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
;
3904 va
+= iview
->image
->cmask
.offset
;
3905 cb
->cb_color_cmask
= va
>> 8;
3907 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
;
3908 va
+= iview
->image
->dcc_offset
;
3909 cb
->cb_dcc_base
= va
>> 8;
3910 cb
->cb_dcc_base
|= iview
->image
->surface
.tile_swizzle
;
3912 uint32_t max_slice
= radv_surface_max_layer_count(iview
) - 1;
3913 cb
->cb_color_view
= S_028C6C_SLICE_START(iview
->base_layer
) |
3914 S_028C6C_SLICE_MAX(max_slice
);
3916 if (iview
->image
->info
.samples
> 1) {
3917 unsigned log_samples
= util_logbase2(iview
->image
->info
.samples
);
3919 cb
->cb_color_attrib
|= S_028C74_NUM_SAMPLES(log_samples
) |
3920 S_028C74_NUM_FRAGMENTS(log_samples
);
3923 if (radv_image_has_fmask(iview
->image
)) {
3924 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
+ iview
->image
->fmask
.offset
;
3925 cb
->cb_color_fmask
= va
>> 8;
3926 cb
->cb_color_fmask
|= iview
->image
->fmask
.tile_swizzle
;
3928 cb
->cb_color_fmask
= cb
->cb_color_base
;
3931 ntype
= radv_translate_color_numformat(iview
->vk_format
,
3933 vk_format_get_first_non_void_channel(iview
->vk_format
));
3934 format
= radv_translate_colorformat(iview
->vk_format
);
3935 if (format
== V_028C70_COLOR_INVALID
|| ntype
== ~0u)
3936 radv_finishme("Illegal color\n");
3937 swap
= radv_translate_colorswap(iview
->vk_format
, FALSE
);
3938 endian
= radv_colorformat_endian_swap(format
);
3940 /* blend clamp should be set for all NORM/SRGB types */
3941 if (ntype
== V_028C70_NUMBER_UNORM
||
3942 ntype
== V_028C70_NUMBER_SNORM
||
3943 ntype
== V_028C70_NUMBER_SRGB
)
3946 /* set blend bypass according to docs if SINT/UINT or
3947 8/24 COLOR variants */
3948 if (ntype
== V_028C70_NUMBER_UINT
|| ntype
== V_028C70_NUMBER_SINT
||
3949 format
== V_028C70_COLOR_8_24
|| format
== V_028C70_COLOR_24_8
||
3950 format
== V_028C70_COLOR_X24_8_32_FLOAT
) {
3955 if ((ntype
== V_028C70_NUMBER_UINT
|| ntype
== V_028C70_NUMBER_SINT
) &&
3956 (format
== V_028C70_COLOR_8
||
3957 format
== V_028C70_COLOR_8_8
||
3958 format
== V_028C70_COLOR_8_8_8_8
))
3959 ->color_is_int8
= true;
3961 cb
->cb_color_info
= S_028C70_FORMAT(format
) |
3962 S_028C70_COMP_SWAP(swap
) |
3963 S_028C70_BLEND_CLAMP(blend_clamp
) |
3964 S_028C70_BLEND_BYPASS(blend_bypass
) |
3965 S_028C70_SIMPLE_FLOAT(1) |
3966 S_028C70_ROUND_MODE(ntype
!= V_028C70_NUMBER_UNORM
&&
3967 ntype
!= V_028C70_NUMBER_SNORM
&&
3968 ntype
!= V_028C70_NUMBER_SRGB
&&
3969 format
!= V_028C70_COLOR_8_24
&&
3970 format
!= V_028C70_COLOR_24_8
) |
3971 S_028C70_NUMBER_TYPE(ntype
) |
3972 S_028C70_ENDIAN(endian
);
3973 if (radv_image_has_fmask(iview
->image
)) {
3974 cb
->cb_color_info
|= S_028C70_COMPRESSION(1);
3975 if (device
->physical_device
->rad_info
.chip_class
== SI
) {
3976 unsigned fmask_bankh
= util_logbase2(iview
->image
->fmask
.bank_height
);
3977 cb
->cb_color_attrib
|= S_028C74_FMASK_BANK_HEIGHT(fmask_bankh
);
3981 if (radv_image_has_cmask(iview
->image
) &&
3982 !(device
->instance
->debug_flags
& RADV_DEBUG_NO_FAST_CLEARS
))
3983 cb
->cb_color_info
|= S_028C70_FAST_CLEAR(1);
3985 if (radv_dcc_enabled(iview
->image
, iview
->base_mip
))
3986 cb
->cb_color_info
|= S_028C70_DCC_ENABLE(1);
3988 cb
->cb_dcc_control
= radv_init_dcc_control_reg(device
, iview
);
3990 /* This must be set for fast clear to work without FMASK. */
3991 if (!radv_image_has_fmask(iview
->image
) &&
3992 device
->physical_device
->rad_info
.chip_class
== SI
) {
3993 unsigned bankh
= util_logbase2(iview
->image
->surface
.u
.legacy
.bankh
);
3994 cb
->cb_color_attrib
|= S_028C74_FMASK_BANK_HEIGHT(bankh
);
3997 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
3998 unsigned mip0_depth
= iview
->image
->type
== VK_IMAGE_TYPE_3D
?
3999 (iview
->extent
.depth
- 1) : (iview
->image
->info
.array_size
- 1);
4001 cb
->cb_color_view
|= S_028C6C_MIP_LEVEL(iview
->base_mip
);
4002 cb
->cb_color_attrib
|= S_028C74_MIP0_DEPTH(mip0_depth
) |
4003 S_028C74_RESOURCE_TYPE(iview
->image
->surface
.u
.gfx9
.resource_type
);
4004 cb
->cb_color_attrib2
= S_028C68_MIP0_WIDTH(iview
->extent
.width
- 1) |
4005 S_028C68_MIP0_HEIGHT(iview
->extent
.height
- 1) |
4006 S_028C68_MAX_MIP(iview
->image
->info
.levels
- 1);
4011 radv_calc_decompress_on_z_planes(struct radv_device
*device
,
4012 struct radv_image_view
*iview
)
4014 unsigned max_zplanes
= 0;
4016 assert(radv_image_is_tc_compat_htile(iview
->image
));
4018 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
4019 /* Default value for 32-bit depth surfaces. */
4022 if (iview
->vk_format
== VK_FORMAT_D16_UNORM
&&
4023 iview
->image
->info
.samples
> 1)
4026 max_zplanes
= max_zplanes
+ 1;
4028 if (iview
->vk_format
== VK_FORMAT_D16_UNORM
) {
4029 /* Do not enable Z plane compression for 16-bit depth
4030 * surfaces because isn't supported on GFX8. Only
4031 * 32-bit depth surfaces are supported by the hardware.
4032 * This allows to maintain shader compatibility and to
4033 * reduce the number of depth decompressions.
4037 if (iview
->image
->info
.samples
<= 1)
4039 else if (iview
->image
->info
.samples
<= 4)
4050 radv_initialise_ds_surface(struct radv_device
*device
,
4051 struct radv_ds_buffer_info
*ds
,
4052 struct radv_image_view
*iview
)
4054 unsigned level
= iview
->base_mip
;
4055 unsigned format
, stencil_format
;
4056 uint64_t va
, s_offs
, z_offs
;
4057 bool stencil_only
= false;
4058 memset(ds
, 0, sizeof(*ds
));
4059 switch (iview
->image
->vk_format
) {
4060 case VK_FORMAT_D24_UNORM_S8_UINT
:
4061 case VK_FORMAT_X8_D24_UNORM_PACK32
:
4062 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-24);
4063 ds
->offset_scale
= 2.0f
;
4065 case VK_FORMAT_D16_UNORM
:
4066 case VK_FORMAT_D16_UNORM_S8_UINT
:
4067 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-16);
4068 ds
->offset_scale
= 4.0f
;
4070 case VK_FORMAT_D32_SFLOAT
:
4071 case VK_FORMAT_D32_SFLOAT_S8_UINT
:
4072 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-23) |
4073 S_028B78_POLY_OFFSET_DB_IS_FLOAT_FMT(1);
4074 ds
->offset_scale
= 1.0f
;
4076 case VK_FORMAT_S8_UINT
:
4077 stencil_only
= true;
4083 format
= radv_translate_dbformat(iview
->image
->vk_format
);
4084 stencil_format
= iview
->image
->surface
.has_stencil
?
4085 V_028044_STENCIL_8
: V_028044_STENCIL_INVALID
;
4087 uint32_t max_slice
= radv_surface_max_layer_count(iview
) - 1;
4088 ds
->db_depth_view
= S_028008_SLICE_START(iview
->base_layer
) |
4089 S_028008_SLICE_MAX(max_slice
);
4091 ds
->db_htile_data_base
= 0;
4092 ds
->db_htile_surface
= 0;
4094 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
;
4095 s_offs
= z_offs
= va
;
4097 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
4098 assert(iview
->image
->surface
.u
.gfx9
.surf_offset
== 0);
4099 s_offs
+= iview
->image
->surface
.u
.gfx9
.stencil_offset
;
4101 ds
->db_z_info
= S_028038_FORMAT(format
) |
4102 S_028038_NUM_SAMPLES(util_logbase2(iview
->image
->info
.samples
)) |
4103 S_028038_SW_MODE(iview
->image
->surface
.u
.gfx9
.surf
.swizzle_mode
) |
4104 S_028038_MAXMIP(iview
->image
->info
.levels
- 1) |
4105 S_028038_ZRANGE_PRECISION(1);
4106 ds
->db_stencil_info
= S_02803C_FORMAT(stencil_format
) |
4107 S_02803C_SW_MODE(iview
->image
->surface
.u
.gfx9
.stencil
.swizzle_mode
);
4109 ds
->db_z_info2
= S_028068_EPITCH(iview
->image
->surface
.u
.gfx9
.surf
.epitch
);
4110 ds
->db_stencil_info2
= S_02806C_EPITCH(iview
->image
->surface
.u
.gfx9
.stencil
.epitch
);
4111 ds
->db_depth_view
|= S_028008_MIPID(level
);
4113 ds
->db_depth_size
= S_02801C_X_MAX(iview
->image
->info
.width
- 1) |
4114 S_02801C_Y_MAX(iview
->image
->info
.height
- 1);
4116 if (radv_htile_enabled(iview
->image
, level
)) {
4117 ds
->db_z_info
|= S_028038_TILE_SURFACE_ENABLE(1);
4119 if (radv_image_is_tc_compat_htile(iview
->image
)) {
4120 unsigned max_zplanes
=
4121 radv_calc_decompress_on_z_planes(device
, iview
);
4123 ds
->db_z_info
|= S_028038_DECOMPRESS_ON_N_ZPLANES(max_zplanes
) |
4124 S_028038_ITERATE_FLUSH(1);
4125 ds
->db_stencil_info
|= S_02803C_ITERATE_FLUSH(1);
4128 if (!iview
->image
->surface
.has_stencil
)
4129 /* Use all of the htile_buffer for depth if there's no stencil. */
4130 ds
->db_stencil_info
|= S_02803C_TILE_STENCIL_DISABLE(1);
4131 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
+
4132 iview
->image
->htile_offset
;
4133 ds
->db_htile_data_base
= va
>> 8;
4134 ds
->db_htile_surface
= S_028ABC_FULL_CACHE(1) |
4135 S_028ABC_PIPE_ALIGNED(iview
->image
->surface
.u
.gfx9
.htile
.pipe_aligned
) |
4136 S_028ABC_RB_ALIGNED(iview
->image
->surface
.u
.gfx9
.htile
.rb_aligned
);
4139 const struct legacy_surf_level
*level_info
= &iview
->image
->surface
.u
.legacy
.level
[level
];
4142 level_info
= &iview
->image
->surface
.u
.legacy
.stencil_level
[level
];
4144 z_offs
+= iview
->image
->surface
.u
.legacy
.level
[level
].offset
;
4145 s_offs
+= iview
->image
->surface
.u
.legacy
.stencil_level
[level
].offset
;
4147 ds
->db_depth_info
= S_02803C_ADDR5_SWIZZLE_MASK(!radv_image_is_tc_compat_htile(iview
->image
));
4148 ds
->db_z_info
= S_028040_FORMAT(format
) | S_028040_ZRANGE_PRECISION(1);
4149 ds
->db_stencil_info
= S_028044_FORMAT(stencil_format
);
4151 if (iview
->image
->info
.samples
> 1)
4152 ds
->db_z_info
|= S_028040_NUM_SAMPLES(util_logbase2(iview
->image
->info
.samples
));
4154 if (device
->physical_device
->rad_info
.chip_class
>= CIK
) {
4155 struct radeon_info
*info
= &device
->physical_device
->rad_info
;
4156 unsigned tiling_index
= iview
->image
->surface
.u
.legacy
.tiling_index
[level
];
4157 unsigned stencil_index
= iview
->image
->surface
.u
.legacy
.stencil_tiling_index
[level
];
4158 unsigned macro_index
= iview
->image
->surface
.u
.legacy
.macro_tile_index
;
4159 unsigned tile_mode
= info
->si_tile_mode_array
[tiling_index
];
4160 unsigned stencil_tile_mode
= info
->si_tile_mode_array
[stencil_index
];
4161 unsigned macro_mode
= info
->cik_macrotile_mode_array
[macro_index
];
4164 tile_mode
= stencil_tile_mode
;
4166 ds
->db_depth_info
|=
4167 S_02803C_ARRAY_MODE(G_009910_ARRAY_MODE(tile_mode
)) |
4168 S_02803C_PIPE_CONFIG(G_009910_PIPE_CONFIG(tile_mode
)) |
4169 S_02803C_BANK_WIDTH(G_009990_BANK_WIDTH(macro_mode
)) |
4170 S_02803C_BANK_HEIGHT(G_009990_BANK_HEIGHT(macro_mode
)) |
4171 S_02803C_MACRO_TILE_ASPECT(G_009990_MACRO_TILE_ASPECT(macro_mode
)) |
4172 S_02803C_NUM_BANKS(G_009990_NUM_BANKS(macro_mode
));
4173 ds
->db_z_info
|= S_028040_TILE_SPLIT(G_009910_TILE_SPLIT(tile_mode
));
4174 ds
->db_stencil_info
|= S_028044_TILE_SPLIT(G_009910_TILE_SPLIT(stencil_tile_mode
));
4176 unsigned tile_mode_index
= si_tile_mode_index(iview
->image
, level
, false);
4177 ds
->db_z_info
|= S_028040_TILE_MODE_INDEX(tile_mode_index
);
4178 tile_mode_index
= si_tile_mode_index(iview
->image
, level
, true);
4179 ds
->db_stencil_info
|= S_028044_TILE_MODE_INDEX(tile_mode_index
);
4181 ds
->db_z_info
|= S_028040_TILE_MODE_INDEX(tile_mode_index
);
4184 ds
->db_depth_size
= S_028058_PITCH_TILE_MAX((level_info
->nblk_x
/ 8) - 1) |
4185 S_028058_HEIGHT_TILE_MAX((level_info
->nblk_y
/ 8) - 1);
4186 ds
->db_depth_slice
= S_02805C_SLICE_TILE_MAX((level_info
->nblk_x
* level_info
->nblk_y
) / 64 - 1);
4188 if (radv_htile_enabled(iview
->image
, level
)) {
4189 ds
->db_z_info
|= S_028040_TILE_SURFACE_ENABLE(1);
4191 if (!iview
->image
->surface
.has_stencil
&&
4192 !radv_image_is_tc_compat_htile(iview
->image
))
4193 /* Use all of the htile_buffer for depth if there's no stencil. */
4194 ds
->db_stencil_info
|= S_028044_TILE_STENCIL_DISABLE(1);
4196 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
+
4197 iview
->image
->htile_offset
;
4198 ds
->db_htile_data_base
= va
>> 8;
4199 ds
->db_htile_surface
= S_028ABC_FULL_CACHE(1);
4201 if (radv_image_is_tc_compat_htile(iview
->image
)) {
4202 unsigned max_zplanes
=
4203 radv_calc_decompress_on_z_planes(device
, iview
);
4205 ds
->db_htile_surface
|= S_028ABC_TC_COMPATIBLE(1);
4206 ds
->db_z_info
|= S_028040_DECOMPRESS_ON_N_ZPLANES(max_zplanes
);
4211 ds
->db_z_read_base
= ds
->db_z_write_base
= z_offs
>> 8;
4212 ds
->db_stencil_read_base
= ds
->db_stencil_write_base
= s_offs
>> 8;
4215 VkResult
radv_CreateFramebuffer(
4217 const VkFramebufferCreateInfo
* pCreateInfo
,
4218 const VkAllocationCallbacks
* pAllocator
,
4219 VkFramebuffer
* pFramebuffer
)
4221 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4222 struct radv_framebuffer
*framebuffer
;
4224 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
4226 size_t size
= sizeof(*framebuffer
) +
4227 sizeof(struct radv_attachment_info
) * pCreateInfo
->attachmentCount
;
4228 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
4229 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
4230 if (framebuffer
== NULL
)
4231 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
4233 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
4234 framebuffer
->width
= pCreateInfo
->width
;
4235 framebuffer
->height
= pCreateInfo
->height
;
4236 framebuffer
->layers
= pCreateInfo
->layers
;
4237 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
4238 VkImageView _iview
= pCreateInfo
->pAttachments
[i
];
4239 struct radv_image_view
*iview
= radv_image_view_from_handle(_iview
);
4240 framebuffer
->attachments
[i
].attachment
= iview
;
4241 if (iview
->aspect_mask
& VK_IMAGE_ASPECT_COLOR_BIT
) {
4242 radv_initialise_color_surface(device
, &framebuffer
->attachments
[i
].cb
, iview
);
4243 } else if (iview
->aspect_mask
& (VK_IMAGE_ASPECT_DEPTH_BIT
| VK_IMAGE_ASPECT_STENCIL_BIT
)) {
4244 radv_initialise_ds_surface(device
, &framebuffer
->attachments
[i
].ds
, iview
);
4246 framebuffer
->width
= MIN2(framebuffer
->width
, iview
->extent
.width
);
4247 framebuffer
->height
= MIN2(framebuffer
->height
, iview
->extent
.height
);
4248 framebuffer
->layers
= MIN2(framebuffer
->layers
, radv_surface_max_layer_count(iview
));
4251 *pFramebuffer
= radv_framebuffer_to_handle(framebuffer
);
4255 void radv_DestroyFramebuffer(
4258 const VkAllocationCallbacks
* pAllocator
)
4260 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4261 RADV_FROM_HANDLE(radv_framebuffer
, fb
, _fb
);
4265 vk_free2(&device
->alloc
, pAllocator
, fb
);
4268 static unsigned radv_tex_wrap(VkSamplerAddressMode address_mode
)
4270 switch (address_mode
) {
4271 case VK_SAMPLER_ADDRESS_MODE_REPEAT
:
4272 return V_008F30_SQ_TEX_WRAP
;
4273 case VK_SAMPLER_ADDRESS_MODE_MIRRORED_REPEAT
:
4274 return V_008F30_SQ_TEX_MIRROR
;
4275 case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE
:
4276 return V_008F30_SQ_TEX_CLAMP_LAST_TEXEL
;
4277 case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_BORDER
:
4278 return V_008F30_SQ_TEX_CLAMP_BORDER
;
4279 case VK_SAMPLER_ADDRESS_MODE_MIRROR_CLAMP_TO_EDGE
:
4280 return V_008F30_SQ_TEX_MIRROR_ONCE_LAST_TEXEL
;
4282 unreachable("illegal tex wrap mode");
4288 radv_tex_compare(VkCompareOp op
)
4291 case VK_COMPARE_OP_NEVER
:
4292 return V_008F30_SQ_TEX_DEPTH_COMPARE_NEVER
;
4293 case VK_COMPARE_OP_LESS
:
4294 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESS
;
4295 case VK_COMPARE_OP_EQUAL
:
4296 return V_008F30_SQ_TEX_DEPTH_COMPARE_EQUAL
;
4297 case VK_COMPARE_OP_LESS_OR_EQUAL
:
4298 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESSEQUAL
;
4299 case VK_COMPARE_OP_GREATER
:
4300 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATER
;
4301 case VK_COMPARE_OP_NOT_EQUAL
:
4302 return V_008F30_SQ_TEX_DEPTH_COMPARE_NOTEQUAL
;
4303 case VK_COMPARE_OP_GREATER_OR_EQUAL
:
4304 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATEREQUAL
;
4305 case VK_COMPARE_OP_ALWAYS
:
4306 return V_008F30_SQ_TEX_DEPTH_COMPARE_ALWAYS
;
4308 unreachable("illegal compare mode");
4314 radv_tex_filter(VkFilter filter
, unsigned max_ansio
)
4317 case VK_FILTER_NEAREST
:
4318 return (max_ansio
> 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_POINT
:
4319 V_008F38_SQ_TEX_XY_FILTER_POINT
);
4320 case VK_FILTER_LINEAR
:
4321 return (max_ansio
> 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_BILINEAR
:
4322 V_008F38_SQ_TEX_XY_FILTER_BILINEAR
);
4323 case VK_FILTER_CUBIC_IMG
:
4325 fprintf(stderr
, "illegal texture filter");
4331 radv_tex_mipfilter(VkSamplerMipmapMode mode
)
4334 case VK_SAMPLER_MIPMAP_MODE_NEAREST
:
4335 return V_008F38_SQ_TEX_Z_FILTER_POINT
;
4336 case VK_SAMPLER_MIPMAP_MODE_LINEAR
:
4337 return V_008F38_SQ_TEX_Z_FILTER_LINEAR
;
4339 return V_008F38_SQ_TEX_Z_FILTER_NONE
;
4344 radv_tex_bordercolor(VkBorderColor bcolor
)
4347 case VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
:
4348 case VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
:
4349 return V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK
;
4350 case VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
:
4351 case VK_BORDER_COLOR_INT_OPAQUE_BLACK
:
4352 return V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_BLACK
;
4353 case VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
:
4354 case VK_BORDER_COLOR_INT_OPAQUE_WHITE
:
4355 return V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_WHITE
;
4363 radv_tex_aniso_filter(unsigned filter
)
4377 radv_tex_filter_mode(VkSamplerReductionModeEXT mode
)
4380 case VK_SAMPLER_REDUCTION_MODE_WEIGHTED_AVERAGE_EXT
:
4381 return SQ_IMG_FILTER_MODE_BLEND
;
4382 case VK_SAMPLER_REDUCTION_MODE_MIN_EXT
:
4383 return SQ_IMG_FILTER_MODE_MIN
;
4384 case VK_SAMPLER_REDUCTION_MODE_MAX_EXT
:
4385 return SQ_IMG_FILTER_MODE_MAX
;
4393 radv_init_sampler(struct radv_device
*device
,
4394 struct radv_sampler
*sampler
,
4395 const VkSamplerCreateInfo
*pCreateInfo
)
4397 uint32_t max_aniso
= pCreateInfo
->anisotropyEnable
&& pCreateInfo
->maxAnisotropy
> 1.0 ?
4398 (uint32_t) pCreateInfo
->maxAnisotropy
: 0;
4399 uint32_t max_aniso_ratio
= radv_tex_aniso_filter(max_aniso
);
4400 bool is_vi
= (device
->physical_device
->rad_info
.chip_class
>= VI
);
4401 unsigned filter_mode
= SQ_IMG_FILTER_MODE_BLEND
;
4403 const struct VkSamplerReductionModeCreateInfoEXT
*sampler_reduction
=
4404 vk_find_struct_const(pCreateInfo
->pNext
,
4405 SAMPLER_REDUCTION_MODE_CREATE_INFO_EXT
);
4406 if (sampler_reduction
)
4407 filter_mode
= radv_tex_filter_mode(sampler_reduction
->reductionMode
);
4409 sampler
->state
[0] = (S_008F30_CLAMP_X(radv_tex_wrap(pCreateInfo
->addressModeU
)) |
4410 S_008F30_CLAMP_Y(radv_tex_wrap(pCreateInfo
->addressModeV
)) |
4411 S_008F30_CLAMP_Z(radv_tex_wrap(pCreateInfo
->addressModeW
)) |
4412 S_008F30_MAX_ANISO_RATIO(max_aniso_ratio
) |
4413 S_008F30_DEPTH_COMPARE_FUNC(radv_tex_compare(pCreateInfo
->compareOp
)) |
4414 S_008F30_FORCE_UNNORMALIZED(pCreateInfo
->unnormalizedCoordinates
? 1 : 0) |
4415 S_008F30_ANISO_THRESHOLD(max_aniso_ratio
>> 1) |
4416 S_008F30_ANISO_BIAS(max_aniso_ratio
) |
4417 S_008F30_DISABLE_CUBE_WRAP(0) |
4418 S_008F30_COMPAT_MODE(is_vi
) |
4419 S_008F30_FILTER_MODE(filter_mode
));
4420 sampler
->state
[1] = (S_008F34_MIN_LOD(S_FIXED(CLAMP(pCreateInfo
->minLod
, 0, 15), 8)) |
4421 S_008F34_MAX_LOD(S_FIXED(CLAMP(pCreateInfo
->maxLod
, 0, 15), 8)) |
4422 S_008F34_PERF_MIP(max_aniso_ratio
? max_aniso_ratio
+ 6 : 0));
4423 sampler
->state
[2] = (S_008F38_LOD_BIAS(S_FIXED(CLAMP(pCreateInfo
->mipLodBias
, -16, 16), 8)) |
4424 S_008F38_XY_MAG_FILTER(radv_tex_filter(pCreateInfo
->magFilter
, max_aniso
)) |
4425 S_008F38_XY_MIN_FILTER(radv_tex_filter(pCreateInfo
->minFilter
, max_aniso
)) |
4426 S_008F38_MIP_FILTER(radv_tex_mipfilter(pCreateInfo
->mipmapMode
)) |
4427 S_008F38_MIP_POINT_PRECLAMP(0) |
4428 S_008F38_DISABLE_LSB_CEIL(device
->physical_device
->rad_info
.chip_class
<= VI
) |
4429 S_008F38_FILTER_PREC_FIX(1) |
4430 S_008F38_ANISO_OVERRIDE(is_vi
));
4431 sampler
->state
[3] = (S_008F3C_BORDER_COLOR_PTR(0) |
4432 S_008F3C_BORDER_COLOR_TYPE(radv_tex_bordercolor(pCreateInfo
->borderColor
)));
4435 VkResult
radv_CreateSampler(
4437 const VkSamplerCreateInfo
* pCreateInfo
,
4438 const VkAllocationCallbacks
* pAllocator
,
4439 VkSampler
* pSampler
)
4441 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4442 struct radv_sampler
*sampler
;
4444 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO
);
4446 sampler
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*sampler
), 8,
4447 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
4449 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
4451 radv_init_sampler(device
, sampler
, pCreateInfo
);
4452 *pSampler
= radv_sampler_to_handle(sampler
);
4457 void radv_DestroySampler(
4460 const VkAllocationCallbacks
* pAllocator
)
4462 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4463 RADV_FROM_HANDLE(radv_sampler
, sampler
, _sampler
);
4467 vk_free2(&device
->alloc
, pAllocator
, sampler
);
4470 /* vk_icd.h does not declare this function, so we declare it here to
4471 * suppress Wmissing-prototypes.
4473 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4474 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t *pSupportedVersion
);
4476 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4477 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t *pSupportedVersion
)
4479 /* For the full details on loader interface versioning, see
4480 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
4481 * What follows is a condensed summary, to help you navigate the large and
4482 * confusing official doc.
4484 * - Loader interface v0 is incompatible with later versions. We don't
4487 * - In loader interface v1:
4488 * - The first ICD entrypoint called by the loader is
4489 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
4491 * - The ICD must statically expose no other Vulkan symbol unless it is
4492 * linked with -Bsymbolic.
4493 * - Each dispatchable Vulkan handle created by the ICD must be
4494 * a pointer to a struct whose first member is VK_LOADER_DATA. The
4495 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
4496 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
4497 * vkDestroySurfaceKHR(). The ICD must be capable of working with
4498 * such loader-managed surfaces.
4500 * - Loader interface v2 differs from v1 in:
4501 * - The first ICD entrypoint called by the loader is
4502 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
4503 * statically expose this entrypoint.
4505 * - Loader interface v3 differs from v2 in:
4506 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
4507 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
4508 * because the loader no longer does so.
4510 *pSupportedVersion
= MIN2(*pSupportedVersion
, 3u);
4514 VkResult
radv_GetMemoryFdKHR(VkDevice _device
,
4515 const VkMemoryGetFdInfoKHR
*pGetFdInfo
,
4518 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4519 RADV_FROM_HANDLE(radv_device_memory
, memory
, pGetFdInfo
->memory
);
4521 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
4523 /* At the moment, we support only the below handle types. */
4524 assert(pGetFdInfo
->handleType
==
4525 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
||
4526 pGetFdInfo
->handleType
==
4527 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
4529 bool ret
= radv_get_memory_fd(device
, memory
, pFD
);
4531 return vk_error(device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
4535 VkResult
radv_GetMemoryFdPropertiesKHR(VkDevice _device
,
4536 VkExternalMemoryHandleTypeFlagBitsKHR handleType
,
4538 VkMemoryFdPropertiesKHR
*pMemoryFdProperties
)
4540 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4542 switch (handleType
) {
4543 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
:
4544 pMemoryFdProperties
->memoryTypeBits
= (1 << RADV_MEM_TYPE_COUNT
) - 1;
4548 /* The valid usage section for this function says:
4550 * "handleType must not be one of the handle types defined as
4553 * So opaque handle types fall into the default "unsupported" case.
4555 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
4559 static VkResult
radv_import_opaque_fd(struct radv_device
*device
,
4563 uint32_t syncobj_handle
= 0;
4564 int ret
= device
->ws
->import_syncobj(device
->ws
, fd
, &syncobj_handle
);
4566 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
4569 device
->ws
->destroy_syncobj(device
->ws
, *syncobj
);
4571 *syncobj
= syncobj_handle
;
4577 static VkResult
radv_import_sync_fd(struct radv_device
*device
,
4581 /* If we create a syncobj we do it locally so that if we have an error, we don't
4582 * leave a syncobj in an undetermined state in the fence. */
4583 uint32_t syncobj_handle
= *syncobj
;
4584 if (!syncobj_handle
) {
4585 int ret
= device
->ws
->create_syncobj(device
->ws
, &syncobj_handle
);
4587 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
4592 device
->ws
->signal_syncobj(device
->ws
, syncobj_handle
);
4594 int ret
= device
->ws
->import_syncobj_from_sync_file(device
->ws
, syncobj_handle
, fd
);
4596 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
4599 *syncobj
= syncobj_handle
;
4606 VkResult
radv_ImportSemaphoreFdKHR(VkDevice _device
,
4607 const VkImportSemaphoreFdInfoKHR
*pImportSemaphoreFdInfo
)
4609 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4610 RADV_FROM_HANDLE(radv_semaphore
, sem
, pImportSemaphoreFdInfo
->semaphore
);
4611 uint32_t *syncobj_dst
= NULL
;
4613 if (pImportSemaphoreFdInfo
->flags
& VK_SEMAPHORE_IMPORT_TEMPORARY_BIT_KHR
) {
4614 syncobj_dst
= &sem
->temp_syncobj
;
4616 syncobj_dst
= &sem
->syncobj
;
4619 switch(pImportSemaphoreFdInfo
->handleType
) {
4620 case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
:
4621 return radv_import_opaque_fd(device
, pImportSemaphoreFdInfo
->fd
, syncobj_dst
);
4622 case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT_KHR
:
4623 return radv_import_sync_fd(device
, pImportSemaphoreFdInfo
->fd
, syncobj_dst
);
4625 unreachable("Unhandled semaphore handle type");
4629 VkResult
radv_GetSemaphoreFdKHR(VkDevice _device
,
4630 const VkSemaphoreGetFdInfoKHR
*pGetFdInfo
,
4633 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4634 RADV_FROM_HANDLE(radv_semaphore
, sem
, pGetFdInfo
->semaphore
);
4636 uint32_t syncobj_handle
;
4638 if (sem
->temp_syncobj
)
4639 syncobj_handle
= sem
->temp_syncobj
;
4641 syncobj_handle
= sem
->syncobj
;
4643 switch(pGetFdInfo
->handleType
) {
4644 case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
:
4645 ret
= device
->ws
->export_syncobj(device
->ws
, syncobj_handle
, pFd
);
4647 case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT_KHR
:
4648 ret
= device
->ws
->export_syncobj_to_sync_file(device
->ws
, syncobj_handle
, pFd
);
4650 if (sem
->temp_syncobj
) {
4651 close (sem
->temp_syncobj
);
4652 sem
->temp_syncobj
= 0;
4654 device
->ws
->reset_syncobj(device
->ws
, syncobj_handle
);
4659 unreachable("Unhandled semaphore handle type");
4663 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
4667 void radv_GetPhysicalDeviceExternalSemaphoreProperties(
4668 VkPhysicalDevice physicalDevice
,
4669 const VkPhysicalDeviceExternalSemaphoreInfoKHR
* pExternalSemaphoreInfo
,
4670 VkExternalSemaphorePropertiesKHR
* pExternalSemaphoreProperties
)
4672 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
4674 /* Require has_syncobj_wait_for_submit for the syncobj signal ioctl introduced at virtually the same time */
4675 if (pdevice
->rad_info
.has_syncobj_wait_for_submit
&&
4676 (pExternalSemaphoreInfo
->handleType
== VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
||
4677 pExternalSemaphoreInfo
->handleType
== VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT_KHR
)) {
4678 pExternalSemaphoreProperties
->exportFromImportedHandleTypes
= VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
| VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT_KHR
;
4679 pExternalSemaphoreProperties
->compatibleHandleTypes
= VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
| VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT_KHR
;
4680 pExternalSemaphoreProperties
->externalSemaphoreFeatures
= VK_EXTERNAL_SEMAPHORE_FEATURE_EXPORTABLE_BIT_KHR
|
4681 VK_EXTERNAL_SEMAPHORE_FEATURE_IMPORTABLE_BIT_KHR
;
4682 } else if (pExternalSemaphoreInfo
->handleType
== VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
) {
4683 pExternalSemaphoreProperties
->exportFromImportedHandleTypes
= VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
;
4684 pExternalSemaphoreProperties
->compatibleHandleTypes
= VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
;
4685 pExternalSemaphoreProperties
->externalSemaphoreFeatures
= VK_EXTERNAL_SEMAPHORE_FEATURE_EXPORTABLE_BIT_KHR
|
4686 VK_EXTERNAL_SEMAPHORE_FEATURE_IMPORTABLE_BIT_KHR
;
4688 pExternalSemaphoreProperties
->exportFromImportedHandleTypes
= 0;
4689 pExternalSemaphoreProperties
->compatibleHandleTypes
= 0;
4690 pExternalSemaphoreProperties
->externalSemaphoreFeatures
= 0;
4694 VkResult
radv_ImportFenceFdKHR(VkDevice _device
,
4695 const VkImportFenceFdInfoKHR
*pImportFenceFdInfo
)
4697 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4698 RADV_FROM_HANDLE(radv_fence
, fence
, pImportFenceFdInfo
->fence
);
4699 uint32_t *syncobj_dst
= NULL
;
4702 if (pImportFenceFdInfo
->flags
& VK_FENCE_IMPORT_TEMPORARY_BIT_KHR
) {
4703 syncobj_dst
= &fence
->temp_syncobj
;
4705 syncobj_dst
= &fence
->syncobj
;
4708 switch(pImportFenceFdInfo
->handleType
) {
4709 case VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
:
4710 return radv_import_opaque_fd(device
, pImportFenceFdInfo
->fd
, syncobj_dst
);
4711 case VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT_KHR
:
4712 return radv_import_sync_fd(device
, pImportFenceFdInfo
->fd
, syncobj_dst
);
4714 unreachable("Unhandled fence handle type");
4718 VkResult
radv_GetFenceFdKHR(VkDevice _device
,
4719 const VkFenceGetFdInfoKHR
*pGetFdInfo
,
4722 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4723 RADV_FROM_HANDLE(radv_fence
, fence
, pGetFdInfo
->fence
);
4725 uint32_t syncobj_handle
;
4727 if (fence
->temp_syncobj
)
4728 syncobj_handle
= fence
->temp_syncobj
;
4730 syncobj_handle
= fence
->syncobj
;
4732 switch(pGetFdInfo
->handleType
) {
4733 case VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
:
4734 ret
= device
->ws
->export_syncobj(device
->ws
, syncobj_handle
, pFd
);
4736 case VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT_KHR
:
4737 ret
= device
->ws
->export_syncobj_to_sync_file(device
->ws
, syncobj_handle
, pFd
);
4739 if (fence
->temp_syncobj
) {
4740 close (fence
->temp_syncobj
);
4741 fence
->temp_syncobj
= 0;
4743 device
->ws
->reset_syncobj(device
->ws
, syncobj_handle
);
4748 unreachable("Unhandled fence handle type");
4752 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
4756 void radv_GetPhysicalDeviceExternalFenceProperties(
4757 VkPhysicalDevice physicalDevice
,
4758 const VkPhysicalDeviceExternalFenceInfoKHR
* pExternalFenceInfo
,
4759 VkExternalFencePropertiesKHR
* pExternalFenceProperties
)
4761 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
4763 if (pdevice
->rad_info
.has_syncobj_wait_for_submit
&&
4764 (pExternalFenceInfo
->handleType
== VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
||
4765 pExternalFenceInfo
->handleType
== VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT_KHR
)) {
4766 pExternalFenceProperties
->exportFromImportedHandleTypes
= VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
| VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT_KHR
;
4767 pExternalFenceProperties
->compatibleHandleTypes
= VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
| VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT_KHR
;
4768 pExternalFenceProperties
->externalFenceFeatures
= VK_EXTERNAL_FENCE_FEATURE_EXPORTABLE_BIT_KHR
|
4769 VK_EXTERNAL_SEMAPHORE_FEATURE_IMPORTABLE_BIT_KHR
;
4771 pExternalFenceProperties
->exportFromImportedHandleTypes
= 0;
4772 pExternalFenceProperties
->compatibleHandleTypes
= 0;
4773 pExternalFenceProperties
->externalFenceFeatures
= 0;
4778 radv_CreateDebugReportCallbackEXT(VkInstance _instance
,
4779 const VkDebugReportCallbackCreateInfoEXT
* pCreateInfo
,
4780 const VkAllocationCallbacks
* pAllocator
,
4781 VkDebugReportCallbackEXT
* pCallback
)
4783 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
4784 return vk_create_debug_report_callback(&instance
->debug_report_callbacks
,
4785 pCreateInfo
, pAllocator
, &instance
->alloc
,
4790 radv_DestroyDebugReportCallbackEXT(VkInstance _instance
,
4791 VkDebugReportCallbackEXT _callback
,
4792 const VkAllocationCallbacks
* pAllocator
)
4794 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
4795 vk_destroy_debug_report_callback(&instance
->debug_report_callbacks
,
4796 _callback
, pAllocator
, &instance
->alloc
);
4800 radv_DebugReportMessageEXT(VkInstance _instance
,
4801 VkDebugReportFlagsEXT flags
,
4802 VkDebugReportObjectTypeEXT objectType
,
4805 int32_t messageCode
,
4806 const char* pLayerPrefix
,
4807 const char* pMessage
)
4809 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
4810 vk_debug_report(&instance
->debug_report_callbacks
, flags
, objectType
,
4811 object
, location
, messageCode
, pLayerPrefix
, pMessage
);
4815 radv_GetDeviceGroupPeerMemoryFeatures(
4818 uint32_t localDeviceIndex
,
4819 uint32_t remoteDeviceIndex
,
4820 VkPeerMemoryFeatureFlags
* pPeerMemoryFeatures
)
4822 assert(localDeviceIndex
== remoteDeviceIndex
);
4824 *pPeerMemoryFeatures
= VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT
|
4825 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT
|
4826 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT
|
4827 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT
;