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"
48 #include "util/build_id.h"
49 #include "util/debug.h"
50 #include "util/mesa-sha1.h"
53 radv_device_get_cache_uuid(enum radeon_family family
, void *uuid
)
56 unsigned char sha1
[20];
57 unsigned ptr_size
= sizeof(void*);
59 memset(uuid
, 0, VK_UUID_SIZE
);
60 _mesa_sha1_init(&ctx
);
62 if (!disk_cache_get_function_identifier(radv_device_get_cache_uuid
, &ctx
) ||
63 !disk_cache_get_function_identifier(LLVMInitializeAMDGPUTargetInfo
, &ctx
))
66 _mesa_sha1_update(&ctx
, &family
, sizeof(family
));
67 _mesa_sha1_update(&ctx
, &ptr_size
, sizeof(ptr_size
));
68 _mesa_sha1_final(&ctx
, sha1
);
70 memcpy(uuid
, sha1
, VK_UUID_SIZE
);
75 radv_get_driver_uuid(void *uuid
)
77 ac_compute_driver_uuid(uuid
, VK_UUID_SIZE
);
81 radv_get_device_uuid(struct radeon_info
*info
, void *uuid
)
83 ac_compute_device_uuid(info
, uuid
, VK_UUID_SIZE
);
87 radv_get_device_name(enum radeon_family family
, char *name
, size_t name_len
)
89 const char *chip_string
;
90 char llvm_string
[32] = {};
93 case CHIP_TAHITI
: chip_string
= "AMD RADV TAHITI"; break;
94 case CHIP_PITCAIRN
: chip_string
= "AMD RADV PITCAIRN"; break;
95 case CHIP_VERDE
: chip_string
= "AMD RADV CAPE VERDE"; break;
96 case CHIP_OLAND
: chip_string
= "AMD RADV OLAND"; break;
97 case CHIP_HAINAN
: chip_string
= "AMD RADV HAINAN"; break;
98 case CHIP_BONAIRE
: chip_string
= "AMD RADV BONAIRE"; break;
99 case CHIP_KAVERI
: chip_string
= "AMD RADV KAVERI"; break;
100 case CHIP_KABINI
: chip_string
= "AMD RADV KABINI"; break;
101 case CHIP_HAWAII
: chip_string
= "AMD RADV HAWAII"; break;
102 case CHIP_MULLINS
: chip_string
= "AMD RADV MULLINS"; break;
103 case CHIP_TONGA
: chip_string
= "AMD RADV TONGA"; break;
104 case CHIP_ICELAND
: chip_string
= "AMD RADV ICELAND"; break;
105 case CHIP_CARRIZO
: chip_string
= "AMD RADV CARRIZO"; break;
106 case CHIP_FIJI
: chip_string
= "AMD RADV FIJI"; break;
107 case CHIP_POLARIS10
: chip_string
= "AMD RADV POLARIS10"; break;
108 case CHIP_POLARIS11
: chip_string
= "AMD RADV POLARIS11"; break;
109 case CHIP_POLARIS12
: chip_string
= "AMD RADV POLARIS12"; break;
110 case CHIP_STONEY
: chip_string
= "AMD RADV STONEY"; break;
111 case CHIP_VEGAM
: chip_string
= "AMD RADV VEGA M"; break;
112 case CHIP_VEGA10
: chip_string
= "AMD RADV VEGA10"; break;
113 case CHIP_VEGA12
: chip_string
= "AMD RADV VEGA12"; break;
114 case CHIP_RAVEN
: chip_string
= "AMD RADV RAVEN"; break;
115 case CHIP_RAVEN2
: chip_string
= "AMD RADV RAVEN2"; break;
116 default: chip_string
= "AMD RADV unknown"; break;
119 snprintf(llvm_string
, sizeof(llvm_string
),
120 " (LLVM %i.%i.%i)", (HAVE_LLVM
>> 8) & 0xff,
121 HAVE_LLVM
& 0xff, MESA_LLVM_VERSION_PATCH
);
122 snprintf(name
, name_len
, "%s%s", chip_string
, llvm_string
);
126 radv_get_visible_vram_size(struct radv_physical_device
*device
)
128 return MIN2(device
->rad_info
.vram_size
, device
->rad_info
.vram_vis_size
);
132 radv_get_vram_size(struct radv_physical_device
*device
)
134 return device
->rad_info
.vram_size
- radv_get_visible_vram_size(device
);
138 radv_physical_device_init_mem_types(struct radv_physical_device
*device
)
140 STATIC_ASSERT(RADV_MEM_HEAP_COUNT
<= VK_MAX_MEMORY_HEAPS
);
141 uint64_t visible_vram_size
= radv_get_visible_vram_size(device
);
142 uint64_t vram_size
= radv_get_vram_size(device
);
143 int vram_index
= -1, visible_vram_index
= -1, gart_index
= -1;
144 device
->memory_properties
.memoryHeapCount
= 0;
146 vram_index
= device
->memory_properties
.memoryHeapCount
++;
147 device
->memory_properties
.memoryHeaps
[vram_index
] = (VkMemoryHeap
) {
149 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
152 if (visible_vram_size
) {
153 visible_vram_index
= device
->memory_properties
.memoryHeapCount
++;
154 device
->memory_properties
.memoryHeaps
[visible_vram_index
] = (VkMemoryHeap
) {
155 .size
= visible_vram_size
,
156 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
159 if (device
->rad_info
.gart_size
> 0) {
160 gart_index
= device
->memory_properties
.memoryHeapCount
++;
161 device
->memory_properties
.memoryHeaps
[gart_index
] = (VkMemoryHeap
) {
162 .size
= device
->rad_info
.gart_size
,
163 .flags
= device
->rad_info
.has_dedicated_vram
? 0 : VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
167 STATIC_ASSERT(RADV_MEM_TYPE_COUNT
<= VK_MAX_MEMORY_TYPES
);
168 unsigned type_count
= 0;
169 if (vram_index
>= 0) {
170 device
->mem_type_indices
[type_count
] = RADV_MEM_TYPE_VRAM
;
171 device
->memory_properties
.memoryTypes
[type_count
++] = (VkMemoryType
) {
172 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
,
173 .heapIndex
= vram_index
,
176 if (gart_index
>= 0) {
177 device
->mem_type_indices
[type_count
] = RADV_MEM_TYPE_GTT_WRITE_COMBINE
;
178 device
->memory_properties
.memoryTypes
[type_count
++] = (VkMemoryType
) {
179 .propertyFlags
= VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
180 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
181 (device
->rad_info
.has_dedicated_vram
? 0 : VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
),
182 .heapIndex
= gart_index
,
185 if (visible_vram_index
>= 0) {
186 device
->mem_type_indices
[type_count
] = RADV_MEM_TYPE_VRAM_CPU_ACCESS
;
187 device
->memory_properties
.memoryTypes
[type_count
++] = (VkMemoryType
) {
188 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
189 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
190 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
191 .heapIndex
= visible_vram_index
,
194 if (gart_index
>= 0) {
195 device
->mem_type_indices
[type_count
] = RADV_MEM_TYPE_GTT_CACHED
;
196 device
->memory_properties
.memoryTypes
[type_count
++] = (VkMemoryType
) {
197 .propertyFlags
= VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
198 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
199 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
|
200 (device
->rad_info
.has_dedicated_vram
? 0 : VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
),
201 .heapIndex
= gart_index
,
204 device
->memory_properties
.memoryTypeCount
= type_count
;
208 radv_handle_env_var_force_family(struct radv_physical_device
*device
)
210 const char *family
= getenv("RADV_FORCE_FAMILY");
216 for (i
= CHIP_TAHITI
; i
< CHIP_LAST
; i
++) {
217 if (!strcmp(family
, ac_get_llvm_processor_name(i
))) {
218 /* Override family and chip_class. */
219 device
->rad_info
.family
= i
;
221 if (i
>= CHIP_VEGA10
)
222 device
->rad_info
.chip_class
= GFX9
;
223 else if (i
>= CHIP_TONGA
)
224 device
->rad_info
.chip_class
= VI
;
225 else if (i
>= CHIP_BONAIRE
)
226 device
->rad_info
.chip_class
= CIK
;
228 device
->rad_info
.chip_class
= SI
;
234 fprintf(stderr
, "radv: Unknown family: %s\n", family
);
239 radv_physical_device_init(struct radv_physical_device
*device
,
240 struct radv_instance
*instance
,
241 drmDevicePtr drm_device
)
243 const char *path
= drm_device
->nodes
[DRM_NODE_RENDER
];
245 drmVersionPtr version
;
249 fd
= open(path
, O_RDWR
| O_CLOEXEC
);
251 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
252 radv_logi("Could not open device '%s'", path
);
254 return vk_error(instance
, VK_ERROR_INCOMPATIBLE_DRIVER
);
257 version
= drmGetVersion(fd
);
261 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
262 radv_logi("Could not get the kernel driver version for device '%s'", path
);
264 return vk_errorf(instance
, VK_ERROR_INCOMPATIBLE_DRIVER
,
265 "failed to get version %s: %m", path
);
268 if (strcmp(version
->name
, "amdgpu")) {
269 drmFreeVersion(version
);
272 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
273 radv_logi("Device '%s' is not using the amdgpu kernel driver.", path
);
275 return VK_ERROR_INCOMPATIBLE_DRIVER
;
277 drmFreeVersion(version
);
279 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
280 radv_logi("Found compatible device '%s'.", path
);
282 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
283 device
->instance
= instance
;
285 device
->ws
= radv_amdgpu_winsys_create(fd
, instance
->debug_flags
,
286 instance
->perftest_flags
);
288 result
= vk_error(instance
, VK_ERROR_INCOMPATIBLE_DRIVER
);
292 if (instance
->enabled_extensions
.KHR_display
) {
293 master_fd
= open(drm_device
->nodes
[DRM_NODE_PRIMARY
], O_RDWR
| O_CLOEXEC
);
294 if (master_fd
>= 0) {
295 uint32_t accel_working
= 0;
296 struct drm_amdgpu_info request
= {
297 .return_pointer
= (uintptr_t)&accel_working
,
298 .return_size
= sizeof(accel_working
),
299 .query
= AMDGPU_INFO_ACCEL_WORKING
302 if (drmCommandWrite(master_fd
, DRM_AMDGPU_INFO
, &request
, sizeof (struct drm_amdgpu_info
)) < 0 || !accel_working
) {
309 device
->master_fd
= master_fd
;
310 device
->local_fd
= fd
;
311 device
->ws
->query_info(device
->ws
, &device
->rad_info
);
313 radv_handle_env_var_force_family(device
);
315 radv_get_device_name(device
->rad_info
.family
, device
->name
, sizeof(device
->name
));
317 if (radv_device_get_cache_uuid(device
->rad_info
.family
, device
->cache_uuid
)) {
318 device
->ws
->destroy(device
->ws
);
319 result
= vk_errorf(instance
, VK_ERROR_INITIALIZATION_FAILED
,
320 "cannot generate UUID");
324 /* These flags affect shader compilation. */
325 uint64_t shader_env_flags
=
326 (device
->instance
->perftest_flags
& RADV_PERFTEST_SISCHED
? 0x1 : 0) |
327 (device
->instance
->debug_flags
& RADV_DEBUG_UNSAFE_MATH
? 0x2 : 0);
329 /* The gpu id is already embedded in the uuid so we just pass "radv"
330 * when creating the cache.
332 char buf
[VK_UUID_SIZE
* 2 + 1];
333 disk_cache_format_hex_id(buf
, device
->cache_uuid
, VK_UUID_SIZE
* 2);
334 device
->disk_cache
= disk_cache_create(device
->name
, buf
, shader_env_flags
);
336 if (device
->rad_info
.chip_class
< VI
||
337 device
->rad_info
.chip_class
> GFX9
)
338 fprintf(stderr
, "WARNING: radv is not a conformant vulkan implementation, testing use only.\n");
340 radv_get_driver_uuid(&device
->device_uuid
);
341 radv_get_device_uuid(&device
->rad_info
, &device
->device_uuid
);
343 if (device
->rad_info
.family
== CHIP_STONEY
||
344 device
->rad_info
.chip_class
>= GFX9
) {
345 device
->has_rbplus
= true;
346 device
->rbplus_allowed
= device
->rad_info
.family
== CHIP_STONEY
||
347 device
->rad_info
.family
== CHIP_VEGA12
||
348 device
->rad_info
.family
== CHIP_RAVEN
||
349 device
->rad_info
.family
== CHIP_RAVEN2
;
352 /* The mere presence of CLEAR_STATE in the IB causes random GPU hangs
355 device
->has_clear_state
= device
->rad_info
.chip_class
>= CIK
;
357 device
->cpdma_prefetch_writes_memory
= device
->rad_info
.chip_class
<= VI
;
359 /* Vega10/Raven need a special workaround for a hardware bug. */
360 device
->has_scissor_bug
= device
->rad_info
.family
== CHIP_VEGA10
||
361 device
->rad_info
.family
== CHIP_RAVEN
;
363 /* Out-of-order primitive rasterization. */
364 device
->has_out_of_order_rast
= device
->rad_info
.chip_class
>= VI
&&
365 device
->rad_info
.max_se
>= 2;
366 device
->out_of_order_rast_allowed
= device
->has_out_of_order_rast
&&
367 !(device
->instance
->debug_flags
& RADV_DEBUG_NO_OUT_OF_ORDER
);
369 device
->dcc_msaa_allowed
=
370 (device
->instance
->perftest_flags
& RADV_PERFTEST_DCC_MSAA
);
372 radv_physical_device_init_mem_types(device
);
373 radv_fill_device_extension_table(device
, &device
->supported_extensions
);
375 device
->bus_info
= *drm_device
->businfo
.pci
;
377 if ((device
->instance
->debug_flags
& RADV_DEBUG_INFO
))
378 ac_print_gpu_info(&device
->rad_info
);
380 /* The WSI is structured as a layer on top of the driver, so this has
381 * to be the last part of initialization (at least until we get other
384 result
= radv_init_wsi(device
);
385 if (result
!= VK_SUCCESS
) {
386 device
->ws
->destroy(device
->ws
);
387 vk_error(instance
, result
);
401 radv_physical_device_finish(struct radv_physical_device
*device
)
403 radv_finish_wsi(device
);
404 device
->ws
->destroy(device
->ws
);
405 disk_cache_destroy(device
->disk_cache
);
406 close(device
->local_fd
);
407 if (device
->master_fd
!= -1)
408 close(device
->master_fd
);
412 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
413 VkSystemAllocationScope allocationScope
)
419 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
420 size_t align
, VkSystemAllocationScope allocationScope
)
422 return realloc(pOriginal
, size
);
426 default_free_func(void *pUserData
, void *pMemory
)
431 static const VkAllocationCallbacks default_alloc
= {
433 .pfnAllocation
= default_alloc_func
,
434 .pfnReallocation
= default_realloc_func
,
435 .pfnFree
= default_free_func
,
438 static const struct debug_control radv_debug_options
[] = {
439 {"nofastclears", RADV_DEBUG_NO_FAST_CLEARS
},
440 {"nodcc", RADV_DEBUG_NO_DCC
},
441 {"shaders", RADV_DEBUG_DUMP_SHADERS
},
442 {"nocache", RADV_DEBUG_NO_CACHE
},
443 {"shaderstats", RADV_DEBUG_DUMP_SHADER_STATS
},
444 {"nohiz", RADV_DEBUG_NO_HIZ
},
445 {"nocompute", RADV_DEBUG_NO_COMPUTE_QUEUE
},
446 {"unsafemath", RADV_DEBUG_UNSAFE_MATH
},
447 {"allbos", RADV_DEBUG_ALL_BOS
},
448 {"noibs", RADV_DEBUG_NO_IBS
},
449 {"spirv", RADV_DEBUG_DUMP_SPIRV
},
450 {"vmfaults", RADV_DEBUG_VM_FAULTS
},
451 {"zerovram", RADV_DEBUG_ZERO_VRAM
},
452 {"syncshaders", RADV_DEBUG_SYNC_SHADERS
},
453 {"nosisched", RADV_DEBUG_NO_SISCHED
},
454 {"preoptir", RADV_DEBUG_PREOPTIR
},
455 {"nodynamicbounds", RADV_DEBUG_NO_DYNAMIC_BOUNDS
},
456 {"nooutoforder", RADV_DEBUG_NO_OUT_OF_ORDER
},
457 {"info", RADV_DEBUG_INFO
},
458 {"errors", RADV_DEBUG_ERRORS
},
459 {"startup", RADV_DEBUG_STARTUP
},
460 {"checkir", RADV_DEBUG_CHECKIR
},
461 {"nothreadllvm", RADV_DEBUG_NOTHREADLLVM
},
462 {"nobinning", RADV_DEBUG_NOBINNING
},
467 radv_get_debug_option_name(int id
)
469 assert(id
< ARRAY_SIZE(radv_debug_options
) - 1);
470 return radv_debug_options
[id
].string
;
473 static const struct debug_control radv_perftest_options
[] = {
474 {"nobatchchain", RADV_PERFTEST_NO_BATCHCHAIN
},
475 {"sisched", RADV_PERFTEST_SISCHED
},
476 {"localbos", RADV_PERFTEST_LOCAL_BOS
},
477 {"dccmsaa", RADV_PERFTEST_DCC_MSAA
},
482 radv_get_perftest_option_name(int id
)
484 assert(id
< ARRAY_SIZE(radv_perftest_options
) - 1);
485 return radv_perftest_options
[id
].string
;
489 radv_handle_per_app_options(struct radv_instance
*instance
,
490 const VkApplicationInfo
*info
)
492 const char *name
= info
? info
->pApplicationName
: NULL
;
497 if (!strcmp(name
, "Talos - Linux - 32bit") ||
498 !strcmp(name
, "Talos - Linux - 64bit")) {
499 if (!(instance
->debug_flags
& RADV_DEBUG_NO_SISCHED
)) {
500 /* Force enable LLVM sisched for Talos because it looks
501 * safe and it gives few more FPS.
503 instance
->perftest_flags
|= RADV_PERFTEST_SISCHED
;
505 } else if (!strcmp(name
, "DOOM_VFR")) {
506 /* Work around a Doom VFR game bug */
507 instance
->debug_flags
|= RADV_DEBUG_NO_DYNAMIC_BOUNDS
;
511 static int radv_get_instance_extension_index(const char *name
)
513 for (unsigned i
= 0; i
< RADV_INSTANCE_EXTENSION_COUNT
; ++i
) {
514 if (strcmp(name
, radv_instance_extensions
[i
].extensionName
) == 0)
521 VkResult
radv_CreateInstance(
522 const VkInstanceCreateInfo
* pCreateInfo
,
523 const VkAllocationCallbacks
* pAllocator
,
524 VkInstance
* pInstance
)
526 struct radv_instance
*instance
;
529 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
531 uint32_t client_version
;
532 if (pCreateInfo
->pApplicationInfo
&&
533 pCreateInfo
->pApplicationInfo
->apiVersion
!= 0) {
534 client_version
= pCreateInfo
->pApplicationInfo
->apiVersion
;
536 client_version
= VK_API_VERSION_1_0
;
539 instance
= vk_zalloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
540 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
542 return vk_error(NULL
, VK_ERROR_OUT_OF_HOST_MEMORY
);
544 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
547 instance
->alloc
= *pAllocator
;
549 instance
->alloc
= default_alloc
;
551 instance
->apiVersion
= client_version
;
552 instance
->physicalDeviceCount
= -1;
554 instance
->debug_flags
= parse_debug_string(getenv("RADV_DEBUG"),
557 instance
->perftest_flags
= parse_debug_string(getenv("RADV_PERFTEST"),
558 radv_perftest_options
);
561 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
562 radv_logi("Created an instance");
564 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
565 const char *ext_name
= pCreateInfo
->ppEnabledExtensionNames
[i
];
566 int index
= radv_get_instance_extension_index(ext_name
);
568 if (index
< 0 || !radv_supported_instance_extensions
.extensions
[index
]) {
569 vk_free2(&default_alloc
, pAllocator
, instance
);
570 return vk_error(instance
, VK_ERROR_EXTENSION_NOT_PRESENT
);
573 instance
->enabled_extensions
.extensions
[index
] = true;
576 result
= vk_debug_report_instance_init(&instance
->debug_report_callbacks
);
577 if (result
!= VK_SUCCESS
) {
578 vk_free2(&default_alloc
, pAllocator
, instance
);
579 return vk_error(instance
, result
);
584 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
586 radv_handle_per_app_options(instance
, pCreateInfo
->pApplicationInfo
);
588 *pInstance
= radv_instance_to_handle(instance
);
593 void radv_DestroyInstance(
594 VkInstance _instance
,
595 const VkAllocationCallbacks
* pAllocator
)
597 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
602 for (int i
= 0; i
< instance
->physicalDeviceCount
; ++i
) {
603 radv_physical_device_finish(instance
->physicalDevices
+ i
);
606 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
610 vk_debug_report_instance_destroy(&instance
->debug_report_callbacks
);
612 vk_free(&instance
->alloc
, instance
);
616 radv_enumerate_devices(struct radv_instance
*instance
)
618 /* TODO: Check for more devices ? */
619 drmDevicePtr devices
[8];
620 VkResult result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
623 instance
->physicalDeviceCount
= 0;
625 max_devices
= drmGetDevices2(0, devices
, ARRAY_SIZE(devices
));
627 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
628 radv_logi("Found %d drm nodes", max_devices
);
631 return vk_error(instance
, VK_ERROR_INCOMPATIBLE_DRIVER
);
633 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
634 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
635 devices
[i
]->bustype
== DRM_BUS_PCI
&&
636 devices
[i
]->deviceinfo
.pci
->vendor_id
== ATI_VENDOR_ID
) {
638 result
= radv_physical_device_init(instance
->physicalDevices
+
639 instance
->physicalDeviceCount
,
642 if (result
== VK_SUCCESS
)
643 ++instance
->physicalDeviceCount
;
644 else if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
648 drmFreeDevices(devices
, max_devices
);
653 VkResult
radv_EnumeratePhysicalDevices(
654 VkInstance _instance
,
655 uint32_t* pPhysicalDeviceCount
,
656 VkPhysicalDevice
* pPhysicalDevices
)
658 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
661 if (instance
->physicalDeviceCount
< 0) {
662 result
= radv_enumerate_devices(instance
);
663 if (result
!= VK_SUCCESS
&&
664 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
668 if (!pPhysicalDevices
) {
669 *pPhysicalDeviceCount
= instance
->physicalDeviceCount
;
671 *pPhysicalDeviceCount
= MIN2(*pPhysicalDeviceCount
, instance
->physicalDeviceCount
);
672 for (unsigned i
= 0; i
< *pPhysicalDeviceCount
; ++i
)
673 pPhysicalDevices
[i
] = radv_physical_device_to_handle(instance
->physicalDevices
+ i
);
676 return *pPhysicalDeviceCount
< instance
->physicalDeviceCount
? VK_INCOMPLETE
680 VkResult
radv_EnumeratePhysicalDeviceGroups(
681 VkInstance _instance
,
682 uint32_t* pPhysicalDeviceGroupCount
,
683 VkPhysicalDeviceGroupProperties
* pPhysicalDeviceGroupProperties
)
685 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
688 if (instance
->physicalDeviceCount
< 0) {
689 result
= radv_enumerate_devices(instance
);
690 if (result
!= VK_SUCCESS
&&
691 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
695 if (!pPhysicalDeviceGroupProperties
) {
696 *pPhysicalDeviceGroupCount
= instance
->physicalDeviceCount
;
698 *pPhysicalDeviceGroupCount
= MIN2(*pPhysicalDeviceGroupCount
, instance
->physicalDeviceCount
);
699 for (unsigned i
= 0; i
< *pPhysicalDeviceGroupCount
; ++i
) {
700 pPhysicalDeviceGroupProperties
[i
].physicalDeviceCount
= 1;
701 pPhysicalDeviceGroupProperties
[i
].physicalDevices
[0] = radv_physical_device_to_handle(instance
->physicalDevices
+ i
);
702 pPhysicalDeviceGroupProperties
[i
].subsetAllocation
= false;
705 return *pPhysicalDeviceGroupCount
< instance
->physicalDeviceCount
? VK_INCOMPLETE
709 void radv_GetPhysicalDeviceFeatures(
710 VkPhysicalDevice physicalDevice
,
711 VkPhysicalDeviceFeatures
* pFeatures
)
713 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
714 memset(pFeatures
, 0, sizeof(*pFeatures
));
716 *pFeatures
= (VkPhysicalDeviceFeatures
) {
717 .robustBufferAccess
= true,
718 .fullDrawIndexUint32
= true,
719 .imageCubeArray
= true,
720 .independentBlend
= true,
721 .geometryShader
= true,
722 .tessellationShader
= true,
723 .sampleRateShading
= true,
724 .dualSrcBlend
= true,
726 .multiDrawIndirect
= true,
727 .drawIndirectFirstInstance
= true,
729 .depthBiasClamp
= true,
730 .fillModeNonSolid
= true,
735 .multiViewport
= true,
736 .samplerAnisotropy
= true,
737 .textureCompressionETC2
= pdevice
->rad_info
.chip_class
>= GFX9
||
738 pdevice
->rad_info
.family
== CHIP_STONEY
,
739 .textureCompressionASTC_LDR
= false,
740 .textureCompressionBC
= true,
741 .occlusionQueryPrecise
= true,
742 .pipelineStatisticsQuery
= true,
743 .vertexPipelineStoresAndAtomics
= true,
744 .fragmentStoresAndAtomics
= true,
745 .shaderTessellationAndGeometryPointSize
= true,
746 .shaderImageGatherExtended
= true,
747 .shaderStorageImageExtendedFormats
= true,
748 .shaderStorageImageMultisample
= pdevice
->rad_info
.chip_class
>= VI
,
749 .shaderUniformBufferArrayDynamicIndexing
= true,
750 .shaderSampledImageArrayDynamicIndexing
= true,
751 .shaderStorageBufferArrayDynamicIndexing
= true,
752 .shaderStorageImageArrayDynamicIndexing
= true,
753 .shaderStorageImageReadWithoutFormat
= true,
754 .shaderStorageImageWriteWithoutFormat
= true,
755 .shaderClipDistance
= true,
756 .shaderCullDistance
= true,
757 .shaderFloat64
= true,
759 .shaderInt16
= pdevice
->rad_info
.chip_class
>= GFX9
,
760 .sparseBinding
= true,
761 .variableMultisampleRate
= true,
762 .inheritedQueries
= true,
766 void radv_GetPhysicalDeviceFeatures2(
767 VkPhysicalDevice physicalDevice
,
768 VkPhysicalDeviceFeatures2
*pFeatures
)
770 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
771 vk_foreach_struct(ext
, pFeatures
->pNext
) {
772 switch (ext
->sType
) {
773 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTER_FEATURES
: {
774 VkPhysicalDeviceVariablePointerFeatures
*features
= (void *)ext
;
775 features
->variablePointersStorageBuffer
= true;
776 features
->variablePointers
= true;
779 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES
: {
780 VkPhysicalDeviceMultiviewFeatures
*features
= (VkPhysicalDeviceMultiviewFeatures
*)ext
;
781 features
->multiview
= true;
782 features
->multiviewGeometryShader
= true;
783 features
->multiviewTessellationShader
= true;
786 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETER_FEATURES
: {
787 VkPhysicalDeviceShaderDrawParameterFeatures
*features
=
788 (VkPhysicalDeviceShaderDrawParameterFeatures
*)ext
;
789 features
->shaderDrawParameters
= true;
792 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES
: {
793 VkPhysicalDeviceProtectedMemoryFeatures
*features
=
794 (VkPhysicalDeviceProtectedMemoryFeatures
*)ext
;
795 features
->protectedMemory
= false;
798 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES
: {
799 VkPhysicalDevice16BitStorageFeatures
*features
=
800 (VkPhysicalDevice16BitStorageFeatures
*)ext
;
801 bool enabled
= pdevice
->rad_info
.chip_class
>= VI
;
802 features
->storageBuffer16BitAccess
= enabled
;
803 features
->uniformAndStorageBuffer16BitAccess
= enabled
;
804 features
->storagePushConstant16
= enabled
;
805 features
->storageInputOutput16
= enabled
;
808 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES
: {
809 VkPhysicalDeviceSamplerYcbcrConversionFeatures
*features
=
810 (VkPhysicalDeviceSamplerYcbcrConversionFeatures
*)ext
;
811 features
->samplerYcbcrConversion
= false;
814 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT
: {
815 VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*features
=
816 (VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*)ext
;
817 features
->shaderInputAttachmentArrayDynamicIndexing
= true;
818 features
->shaderUniformTexelBufferArrayDynamicIndexing
= true;
819 features
->shaderStorageTexelBufferArrayDynamicIndexing
= true;
820 features
->shaderUniformBufferArrayNonUniformIndexing
= false;
821 features
->shaderSampledImageArrayNonUniformIndexing
= false;
822 features
->shaderStorageBufferArrayNonUniformIndexing
= false;
823 features
->shaderStorageImageArrayNonUniformIndexing
= false;
824 features
->shaderInputAttachmentArrayNonUniformIndexing
= false;
825 features
->shaderUniformTexelBufferArrayNonUniformIndexing
= false;
826 features
->shaderStorageTexelBufferArrayNonUniformIndexing
= false;
827 features
->descriptorBindingUniformBufferUpdateAfterBind
= true;
828 features
->descriptorBindingSampledImageUpdateAfterBind
= true;
829 features
->descriptorBindingStorageImageUpdateAfterBind
= true;
830 features
->descriptorBindingStorageBufferUpdateAfterBind
= true;
831 features
->descriptorBindingUniformTexelBufferUpdateAfterBind
= true;
832 features
->descriptorBindingStorageTexelBufferUpdateAfterBind
= true;
833 features
->descriptorBindingUpdateUnusedWhilePending
= true;
834 features
->descriptorBindingPartiallyBound
= true;
835 features
->descriptorBindingVariableDescriptorCount
= true;
836 features
->runtimeDescriptorArray
= true;
839 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT
: {
840 VkPhysicalDeviceConditionalRenderingFeaturesEXT
*features
=
841 (VkPhysicalDeviceConditionalRenderingFeaturesEXT
*)ext
;
842 features
->conditionalRendering
= true;
843 features
->inheritedConditionalRendering
= false;
846 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT
: {
847 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*features
=
848 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*)ext
;
849 features
->vertexAttributeInstanceRateDivisor
= VK_TRUE
;
850 features
->vertexAttributeInstanceRateZeroDivisor
= VK_TRUE
;
853 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT
: {
854 VkPhysicalDeviceTransformFeedbackFeaturesEXT
*features
=
855 (VkPhysicalDeviceTransformFeedbackFeaturesEXT
*)ext
;
856 features
->transformFeedback
= true;
857 features
->geometryStreams
= true;
860 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES_EXT
: {
861 VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*features
=
862 (VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*)ext
;
863 features
->scalarBlockLayout
= pdevice
->rad_info
.chip_class
>= CIK
;
866 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_PRIORITY_FEATURES_EXT
: {
867 VkPhysicalDeviceMemoryPriorityFeaturesEXT
*features
=
868 (VkPhysicalDeviceMemoryPriorityFeaturesEXT
*)ext
;
869 features
->memoryPriority
= VK_TRUE
;
876 return radv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
879 void radv_GetPhysicalDeviceProperties(
880 VkPhysicalDevice physicalDevice
,
881 VkPhysicalDeviceProperties
* pProperties
)
883 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
884 VkSampleCountFlags sample_counts
= 0xf;
886 /* make sure that the entire descriptor set is addressable with a signed
887 * 32-bit int. So the sum of all limits scaled by descriptor size has to
888 * be at most 2 GiB. the combined image & samples object count as one of
889 * both. This limit is for the pipeline layout, not for the set layout, but
890 * there is no set limit, so we just set a pipeline limit. I don't think
891 * any app is going to hit this soon. */
892 size_t max_descriptor_set_size
= ((1ull << 31) - 16 * MAX_DYNAMIC_BUFFERS
) /
893 (32 /* uniform buffer, 32 due to potential space wasted on alignment */ +
894 32 /* storage buffer, 32 due to potential space wasted on alignment */ +
895 32 /* sampler, largest when combined with image */ +
896 64 /* sampled image */ +
897 64 /* storage image */);
899 VkPhysicalDeviceLimits limits
= {
900 .maxImageDimension1D
= (1 << 14),
901 .maxImageDimension2D
= (1 << 14),
902 .maxImageDimension3D
= (1 << 11),
903 .maxImageDimensionCube
= (1 << 14),
904 .maxImageArrayLayers
= (1 << 11),
905 .maxTexelBufferElements
= 128 * 1024 * 1024,
906 .maxUniformBufferRange
= UINT32_MAX
,
907 .maxStorageBufferRange
= UINT32_MAX
,
908 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
909 .maxMemoryAllocationCount
= UINT32_MAX
,
910 .maxSamplerAllocationCount
= 64 * 1024,
911 .bufferImageGranularity
= 64, /* A cache line */
912 .sparseAddressSpaceSize
= 0xffffffffu
, /* buffer max size */
913 .maxBoundDescriptorSets
= MAX_SETS
,
914 .maxPerStageDescriptorSamplers
= max_descriptor_set_size
,
915 .maxPerStageDescriptorUniformBuffers
= max_descriptor_set_size
,
916 .maxPerStageDescriptorStorageBuffers
= max_descriptor_set_size
,
917 .maxPerStageDescriptorSampledImages
= max_descriptor_set_size
,
918 .maxPerStageDescriptorStorageImages
= max_descriptor_set_size
,
919 .maxPerStageDescriptorInputAttachments
= max_descriptor_set_size
,
920 .maxPerStageResources
= max_descriptor_set_size
,
921 .maxDescriptorSetSamplers
= max_descriptor_set_size
,
922 .maxDescriptorSetUniformBuffers
= max_descriptor_set_size
,
923 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_UNIFORM_BUFFERS
,
924 .maxDescriptorSetStorageBuffers
= max_descriptor_set_size
,
925 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_STORAGE_BUFFERS
,
926 .maxDescriptorSetSampledImages
= max_descriptor_set_size
,
927 .maxDescriptorSetStorageImages
= max_descriptor_set_size
,
928 .maxDescriptorSetInputAttachments
= max_descriptor_set_size
,
929 .maxVertexInputAttributes
= 32,
930 .maxVertexInputBindings
= 32,
931 .maxVertexInputAttributeOffset
= 2047,
932 .maxVertexInputBindingStride
= 2048,
933 .maxVertexOutputComponents
= 128,
934 .maxTessellationGenerationLevel
= 64,
935 .maxTessellationPatchSize
= 32,
936 .maxTessellationControlPerVertexInputComponents
= 128,
937 .maxTessellationControlPerVertexOutputComponents
= 128,
938 .maxTessellationControlPerPatchOutputComponents
= 120,
939 .maxTessellationControlTotalOutputComponents
= 4096,
940 .maxTessellationEvaluationInputComponents
= 128,
941 .maxTessellationEvaluationOutputComponents
= 128,
942 .maxGeometryShaderInvocations
= 127,
943 .maxGeometryInputComponents
= 64,
944 .maxGeometryOutputComponents
= 128,
945 .maxGeometryOutputVertices
= 256,
946 .maxGeometryTotalOutputComponents
= 1024,
947 .maxFragmentInputComponents
= 128,
948 .maxFragmentOutputAttachments
= 8,
949 .maxFragmentDualSrcAttachments
= 1,
950 .maxFragmentCombinedOutputResources
= 8,
951 .maxComputeSharedMemorySize
= 32768,
952 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
953 .maxComputeWorkGroupInvocations
= 2048,
954 .maxComputeWorkGroupSize
= {
959 .subPixelPrecisionBits
= 8,
960 .subTexelPrecisionBits
= 8,
961 .mipmapPrecisionBits
= 8,
962 .maxDrawIndexedIndexValue
= UINT32_MAX
,
963 .maxDrawIndirectCount
= UINT32_MAX
,
964 .maxSamplerLodBias
= 16,
965 .maxSamplerAnisotropy
= 16,
966 .maxViewports
= MAX_VIEWPORTS
,
967 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
968 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
969 .viewportSubPixelBits
= 8,
970 .minMemoryMapAlignment
= 4096, /* A page */
971 .minTexelBufferOffsetAlignment
= 1,
972 .minUniformBufferOffsetAlignment
= 4,
973 .minStorageBufferOffsetAlignment
= 4,
974 .minTexelOffset
= -32,
975 .maxTexelOffset
= 31,
976 .minTexelGatherOffset
= -32,
977 .maxTexelGatherOffset
= 31,
978 .minInterpolationOffset
= -2,
979 .maxInterpolationOffset
= 2,
980 .subPixelInterpolationOffsetBits
= 8,
981 .maxFramebufferWidth
= (1 << 14),
982 .maxFramebufferHeight
= (1 << 14),
983 .maxFramebufferLayers
= (1 << 10),
984 .framebufferColorSampleCounts
= sample_counts
,
985 .framebufferDepthSampleCounts
= sample_counts
,
986 .framebufferStencilSampleCounts
= sample_counts
,
987 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
988 .maxColorAttachments
= MAX_RTS
,
989 .sampledImageColorSampleCounts
= sample_counts
,
990 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
991 .sampledImageDepthSampleCounts
= sample_counts
,
992 .sampledImageStencilSampleCounts
= sample_counts
,
993 .storageImageSampleCounts
= pdevice
->rad_info
.chip_class
>= VI
? sample_counts
: VK_SAMPLE_COUNT_1_BIT
,
994 .maxSampleMaskWords
= 1,
995 .timestampComputeAndGraphics
= true,
996 .timestampPeriod
= 1000000.0 / pdevice
->rad_info
.clock_crystal_freq
,
997 .maxClipDistances
= 8,
998 .maxCullDistances
= 8,
999 .maxCombinedClipAndCullDistances
= 8,
1000 .discreteQueuePriorities
= 2,
1001 .pointSizeRange
= { 0.125, 255.875 },
1002 .lineWidthRange
= { 0.0, 7.9921875 },
1003 .pointSizeGranularity
= (1.0 / 8.0),
1004 .lineWidthGranularity
= (1.0 / 128.0),
1005 .strictLines
= false, /* FINISHME */
1006 .standardSampleLocations
= true,
1007 .optimalBufferCopyOffsetAlignment
= 128,
1008 .optimalBufferCopyRowPitchAlignment
= 128,
1009 .nonCoherentAtomSize
= 64,
1012 *pProperties
= (VkPhysicalDeviceProperties
) {
1013 .apiVersion
= radv_physical_device_api_version(pdevice
),
1014 .driverVersion
= vk_get_driver_version(),
1015 .vendorID
= ATI_VENDOR_ID
,
1016 .deviceID
= pdevice
->rad_info
.pci_id
,
1017 .deviceType
= pdevice
->rad_info
.has_dedicated_vram
? VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU
: VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
1019 .sparseProperties
= {0},
1022 strcpy(pProperties
->deviceName
, pdevice
->name
);
1023 memcpy(pProperties
->pipelineCacheUUID
, pdevice
->cache_uuid
, VK_UUID_SIZE
);
1026 void radv_GetPhysicalDeviceProperties2(
1027 VkPhysicalDevice physicalDevice
,
1028 VkPhysicalDeviceProperties2
*pProperties
)
1030 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
1031 radv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
1033 vk_foreach_struct(ext
, pProperties
->pNext
) {
1034 switch (ext
->sType
) {
1035 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
1036 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
1037 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
1038 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
1041 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES
: {
1042 VkPhysicalDeviceIDProperties
*properties
= (VkPhysicalDeviceIDProperties
*)ext
;
1043 memcpy(properties
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
1044 memcpy(properties
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
1045 properties
->deviceLUIDValid
= false;
1048 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES
: {
1049 VkPhysicalDeviceMultiviewProperties
*properties
= (VkPhysicalDeviceMultiviewProperties
*)ext
;
1050 properties
->maxMultiviewViewCount
= MAX_VIEWS
;
1051 properties
->maxMultiviewInstanceIndex
= INT_MAX
;
1054 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES
: {
1055 VkPhysicalDevicePointClippingProperties
*properties
=
1056 (VkPhysicalDevicePointClippingProperties
*)ext
;
1057 properties
->pointClippingBehavior
= VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES
;
1060 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DISCARD_RECTANGLE_PROPERTIES_EXT
: {
1061 VkPhysicalDeviceDiscardRectanglePropertiesEXT
*properties
=
1062 (VkPhysicalDeviceDiscardRectanglePropertiesEXT
*)ext
;
1063 properties
->maxDiscardRectangles
= MAX_DISCARD_RECTANGLES
;
1066 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT
: {
1067 VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*properties
=
1068 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*) ext
;
1069 properties
->minImportedHostPointerAlignment
= 4096;
1072 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES
: {
1073 VkPhysicalDeviceSubgroupProperties
*properties
=
1074 (VkPhysicalDeviceSubgroupProperties
*)ext
;
1075 properties
->subgroupSize
= 64;
1076 properties
->supportedStages
= VK_SHADER_STAGE_ALL
;
1077 properties
->supportedOperations
=
1078 VK_SUBGROUP_FEATURE_BASIC_BIT
|
1079 VK_SUBGROUP_FEATURE_BALLOT_BIT
|
1080 VK_SUBGROUP_FEATURE_QUAD_BIT
|
1081 VK_SUBGROUP_FEATURE_VOTE_BIT
;
1082 if (pdevice
->rad_info
.chip_class
>= VI
) {
1083 properties
->supportedOperations
|=
1084 VK_SUBGROUP_FEATURE_ARITHMETIC_BIT
|
1085 VK_SUBGROUP_FEATURE_SHUFFLE_BIT
|
1086 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT
;
1088 properties
->quadOperationsInAllStages
= true;
1091 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES
: {
1092 VkPhysicalDeviceMaintenance3Properties
*properties
=
1093 (VkPhysicalDeviceMaintenance3Properties
*)ext
;
1094 /* Make sure everything is addressable by a signed 32-bit int, and
1095 * our largest descriptors are 96 bytes. */
1096 properties
->maxPerSetDescriptors
= (1ull << 31) / 96;
1097 /* Our buffer size fields allow only this much */
1098 properties
->maxMemoryAllocationSize
= 0xFFFFFFFFull
;
1101 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT
: {
1102 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*properties
=
1103 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*)ext
;
1104 /* GFX6-8 only support single channel min/max filter. */
1105 properties
->filterMinmaxImageComponentMapping
= pdevice
->rad_info
.chip_class
>= GFX9
;
1106 properties
->filterMinmaxSingleComponentFormats
= true;
1109 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CORE_PROPERTIES_AMD
: {
1110 VkPhysicalDeviceShaderCorePropertiesAMD
*properties
=
1111 (VkPhysicalDeviceShaderCorePropertiesAMD
*)ext
;
1113 /* Shader engines. */
1114 properties
->shaderEngineCount
=
1115 pdevice
->rad_info
.max_se
;
1116 properties
->shaderArraysPerEngineCount
=
1117 pdevice
->rad_info
.max_sh_per_se
;
1118 properties
->computeUnitsPerShaderArray
=
1119 pdevice
->rad_info
.num_good_cu_per_sh
;
1120 properties
->simdPerComputeUnit
= 4;
1121 properties
->wavefrontsPerSimd
=
1122 pdevice
->rad_info
.family
== CHIP_TONGA
||
1123 pdevice
->rad_info
.family
== CHIP_ICELAND
||
1124 pdevice
->rad_info
.family
== CHIP_POLARIS10
||
1125 pdevice
->rad_info
.family
== CHIP_POLARIS11
||
1126 pdevice
->rad_info
.family
== CHIP_POLARIS12
||
1127 pdevice
->rad_info
.family
== CHIP_VEGAM
? 8 : 10;
1128 properties
->wavefrontSize
= 64;
1131 properties
->sgprsPerSimd
=
1132 ac_get_num_physical_sgprs(pdevice
->rad_info
.chip_class
);
1133 properties
->minSgprAllocation
=
1134 pdevice
->rad_info
.chip_class
>= VI
? 16 : 8;
1135 properties
->maxSgprAllocation
=
1136 pdevice
->rad_info
.family
== CHIP_TONGA
||
1137 pdevice
->rad_info
.family
== CHIP_ICELAND
? 96 : 104;
1138 properties
->sgprAllocationGranularity
=
1139 pdevice
->rad_info
.chip_class
>= VI
? 16 : 8;
1142 properties
->vgprsPerSimd
= RADV_NUM_PHYSICAL_VGPRS
;
1143 properties
->minVgprAllocation
= 4;
1144 properties
->maxVgprAllocation
= 256;
1145 properties
->vgprAllocationGranularity
= 4;
1148 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT
: {
1149 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*properties
=
1150 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*)ext
;
1151 properties
->maxVertexAttribDivisor
= UINT32_MAX
;
1154 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES_EXT
: {
1155 VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*properties
=
1156 (VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*)ext
;
1157 properties
->maxUpdateAfterBindDescriptorsInAllPools
= UINT32_MAX
/ 64;
1158 properties
->shaderUniformBufferArrayNonUniformIndexingNative
= false;
1159 properties
->shaderSampledImageArrayNonUniformIndexingNative
= false;
1160 properties
->shaderStorageBufferArrayNonUniformIndexingNative
= false;
1161 properties
->shaderStorageImageArrayNonUniformIndexingNative
= false;
1162 properties
->shaderInputAttachmentArrayNonUniformIndexingNative
= false;
1163 properties
->robustBufferAccessUpdateAfterBind
= false;
1164 properties
->quadDivergentImplicitLod
= false;
1166 size_t max_descriptor_set_size
= ((1ull << 31) - 16 * MAX_DYNAMIC_BUFFERS
) /
1167 (32 /* uniform buffer, 32 due to potential space wasted on alignment */ +
1168 32 /* storage buffer, 32 due to potential space wasted on alignment */ +
1169 32 /* sampler, largest when combined with image */ +
1170 64 /* sampled image */ +
1171 64 /* storage image */);
1172 properties
->maxPerStageDescriptorUpdateAfterBindSamplers
= max_descriptor_set_size
;
1173 properties
->maxPerStageDescriptorUpdateAfterBindUniformBuffers
= max_descriptor_set_size
;
1174 properties
->maxPerStageDescriptorUpdateAfterBindStorageBuffers
= max_descriptor_set_size
;
1175 properties
->maxPerStageDescriptorUpdateAfterBindSampledImages
= max_descriptor_set_size
;
1176 properties
->maxPerStageDescriptorUpdateAfterBindStorageImages
= max_descriptor_set_size
;
1177 properties
->maxPerStageDescriptorUpdateAfterBindInputAttachments
= max_descriptor_set_size
;
1178 properties
->maxPerStageUpdateAfterBindResources
= max_descriptor_set_size
;
1179 properties
->maxDescriptorSetUpdateAfterBindSamplers
= max_descriptor_set_size
;
1180 properties
->maxDescriptorSetUpdateAfterBindUniformBuffers
= max_descriptor_set_size
;
1181 properties
->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic
= MAX_DYNAMIC_UNIFORM_BUFFERS
;
1182 properties
->maxDescriptorSetUpdateAfterBindStorageBuffers
= max_descriptor_set_size
;
1183 properties
->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic
= MAX_DYNAMIC_STORAGE_BUFFERS
;
1184 properties
->maxDescriptorSetUpdateAfterBindSampledImages
= max_descriptor_set_size
;
1185 properties
->maxDescriptorSetUpdateAfterBindStorageImages
= max_descriptor_set_size
;
1186 properties
->maxDescriptorSetUpdateAfterBindInputAttachments
= max_descriptor_set_size
;
1189 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES
: {
1190 VkPhysicalDeviceProtectedMemoryProperties
*properties
=
1191 (VkPhysicalDeviceProtectedMemoryProperties
*)ext
;
1192 properties
->protectedNoFault
= false;
1195 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONSERVATIVE_RASTERIZATION_PROPERTIES_EXT
: {
1196 VkPhysicalDeviceConservativeRasterizationPropertiesEXT
*properties
=
1197 (VkPhysicalDeviceConservativeRasterizationPropertiesEXT
*)ext
;
1198 properties
->primitiveOverestimationSize
= 0;
1199 properties
->maxExtraPrimitiveOverestimationSize
= 0;
1200 properties
->extraPrimitiveOverestimationSizeGranularity
= 0;
1201 properties
->primitiveUnderestimation
= VK_FALSE
;
1202 properties
->conservativePointAndLineRasterization
= VK_FALSE
;
1203 properties
->degenerateTrianglesRasterized
= VK_FALSE
;
1204 properties
->degenerateLinesRasterized
= VK_FALSE
;
1205 properties
->fullyCoveredFragmentShaderInputVariable
= VK_FALSE
;
1206 properties
->conservativeRasterizationPostDepthCoverage
= VK_FALSE
;
1209 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT
: {
1210 VkPhysicalDevicePCIBusInfoPropertiesEXT
*properties
=
1211 (VkPhysicalDevicePCIBusInfoPropertiesEXT
*)ext
;
1212 properties
->pciDomain
= pdevice
->bus_info
.domain
;
1213 properties
->pciBus
= pdevice
->bus_info
.bus
;
1214 properties
->pciDevice
= pdevice
->bus_info
.dev
;
1215 properties
->pciFunction
= pdevice
->bus_info
.func
;
1218 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR
: {
1219 VkPhysicalDeviceDriverPropertiesKHR
*driver_props
=
1220 (VkPhysicalDeviceDriverPropertiesKHR
*) ext
;
1222 driver_props
->driverID
= VK_DRIVER_ID_MESA_RADV_KHR
;
1223 memset(driver_props
->driverName
, 0, VK_MAX_DRIVER_NAME_SIZE_KHR
);
1224 strcpy(driver_props
->driverName
, "radv");
1226 memset(driver_props
->driverInfo
, 0, VK_MAX_DRIVER_INFO_SIZE_KHR
);
1227 snprintf(driver_props
->driverInfo
, VK_MAX_DRIVER_INFO_SIZE_KHR
,
1228 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1
1230 (HAVE_LLVM
>> 8) & 0xff, HAVE_LLVM
& 0xff,
1231 MESA_LLVM_VERSION_PATCH
);
1233 driver_props
->conformanceVersion
= (VkConformanceVersionKHR
) {
1241 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT
: {
1242 VkPhysicalDeviceTransformFeedbackPropertiesEXT
*properties
=
1243 (VkPhysicalDeviceTransformFeedbackPropertiesEXT
*)ext
;
1244 properties
->maxTransformFeedbackStreams
= MAX_SO_STREAMS
;
1245 properties
->maxTransformFeedbackBuffers
= MAX_SO_BUFFERS
;
1246 properties
->maxTransformFeedbackBufferSize
= UINT32_MAX
;
1247 properties
->maxTransformFeedbackStreamDataSize
= 512;
1248 properties
->maxTransformFeedbackBufferDataSize
= UINT32_MAX
;
1249 properties
->maxTransformFeedbackBufferDataStride
= 512;
1250 properties
->transformFeedbackQueries
= true;
1251 properties
->transformFeedbackStreamsLinesTriangles
= false;
1252 properties
->transformFeedbackRasterizationStreamSelect
= false;
1253 properties
->transformFeedbackDraw
= true;
1262 static void radv_get_physical_device_queue_family_properties(
1263 struct radv_physical_device
* pdevice
,
1265 VkQueueFamilyProperties
** pQueueFamilyProperties
)
1267 int num_queue_families
= 1;
1269 if (pdevice
->rad_info
.num_compute_rings
> 0 &&
1270 !(pdevice
->instance
->debug_flags
& RADV_DEBUG_NO_COMPUTE_QUEUE
))
1271 num_queue_families
++;
1273 if (pQueueFamilyProperties
== NULL
) {
1274 *pCount
= num_queue_families
;
1283 *pQueueFamilyProperties
[idx
] = (VkQueueFamilyProperties
) {
1284 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
1285 VK_QUEUE_COMPUTE_BIT
|
1286 VK_QUEUE_TRANSFER_BIT
|
1287 VK_QUEUE_SPARSE_BINDING_BIT
,
1289 .timestampValidBits
= 64,
1290 .minImageTransferGranularity
= (VkExtent3D
) { 1, 1, 1 },
1295 if (pdevice
->rad_info
.num_compute_rings
> 0 &&
1296 !(pdevice
->instance
->debug_flags
& RADV_DEBUG_NO_COMPUTE_QUEUE
)) {
1297 if (*pCount
> idx
) {
1298 *pQueueFamilyProperties
[idx
] = (VkQueueFamilyProperties
) {
1299 .queueFlags
= VK_QUEUE_COMPUTE_BIT
|
1300 VK_QUEUE_TRANSFER_BIT
|
1301 VK_QUEUE_SPARSE_BINDING_BIT
,
1302 .queueCount
= pdevice
->rad_info
.num_compute_rings
,
1303 .timestampValidBits
= 64,
1304 .minImageTransferGranularity
= (VkExtent3D
) { 1, 1, 1 },
1312 void radv_GetPhysicalDeviceQueueFamilyProperties(
1313 VkPhysicalDevice physicalDevice
,
1315 VkQueueFamilyProperties
* pQueueFamilyProperties
)
1317 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
1318 if (!pQueueFamilyProperties
) {
1319 radv_get_physical_device_queue_family_properties(pdevice
, pCount
, NULL
);
1322 VkQueueFamilyProperties
*properties
[] = {
1323 pQueueFamilyProperties
+ 0,
1324 pQueueFamilyProperties
+ 1,
1325 pQueueFamilyProperties
+ 2,
1327 radv_get_physical_device_queue_family_properties(pdevice
, pCount
, properties
);
1328 assert(*pCount
<= 3);
1331 void radv_GetPhysicalDeviceQueueFamilyProperties2(
1332 VkPhysicalDevice physicalDevice
,
1334 VkQueueFamilyProperties2
*pQueueFamilyProperties
)
1336 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
1337 if (!pQueueFamilyProperties
) {
1338 radv_get_physical_device_queue_family_properties(pdevice
, pCount
, NULL
);
1341 VkQueueFamilyProperties
*properties
[] = {
1342 &pQueueFamilyProperties
[0].queueFamilyProperties
,
1343 &pQueueFamilyProperties
[1].queueFamilyProperties
,
1344 &pQueueFamilyProperties
[2].queueFamilyProperties
,
1346 radv_get_physical_device_queue_family_properties(pdevice
, pCount
, properties
);
1347 assert(*pCount
<= 3);
1350 void radv_GetPhysicalDeviceMemoryProperties(
1351 VkPhysicalDevice physicalDevice
,
1352 VkPhysicalDeviceMemoryProperties
*pMemoryProperties
)
1354 RADV_FROM_HANDLE(radv_physical_device
, physical_device
, physicalDevice
);
1356 *pMemoryProperties
= physical_device
->memory_properties
;
1360 radv_get_memory_budget_properties(VkPhysicalDevice physicalDevice
,
1361 VkPhysicalDeviceMemoryBudgetPropertiesEXT
*memoryBudget
)
1363 RADV_FROM_HANDLE(radv_physical_device
, device
, physicalDevice
);
1364 VkPhysicalDeviceMemoryProperties
*memory_properties
= &device
->memory_properties
;
1365 uint64_t visible_vram_size
= radv_get_visible_vram_size(device
);
1366 uint64_t vram_size
= radv_get_vram_size(device
);
1367 uint64_t gtt_size
= device
->rad_info
.gart_size
;
1368 uint64_t heap_budget
, heap_usage
;
1370 /* For all memory heaps, the computation of budget is as follow:
1371 * heap_budget = heap_size - global_heap_usage + app_heap_usage
1373 * The Vulkan spec 1.1.97 says that the budget should include any
1374 * currently allocated device memory.
1376 * Note that the application heap usages are not really accurate (eg.
1377 * in presence of shared buffers).
1380 heap_usage
= device
->ws
->query_value(device
->ws
,
1381 RADEON_ALLOCATED_VRAM
);
1383 heap_budget
= vram_size
-
1384 device
->ws
->query_value(device
->ws
, RADEON_VRAM_USAGE
) +
1387 memoryBudget
->heapBudget
[RADV_MEM_HEAP_VRAM
] = heap_budget
;
1388 memoryBudget
->heapUsage
[RADV_MEM_HEAP_VRAM
] = heap_usage
;
1391 if (visible_vram_size
) {
1392 heap_usage
= device
->ws
->query_value(device
->ws
,
1393 RADEON_ALLOCATED_VRAM_VIS
);
1395 heap_budget
= visible_vram_size
-
1396 device
->ws
->query_value(device
->ws
, RADEON_VRAM_VIS_USAGE
) +
1399 memoryBudget
->heapBudget
[RADV_MEM_HEAP_VRAM_CPU_ACCESS
] = heap_budget
;
1400 memoryBudget
->heapUsage
[RADV_MEM_HEAP_VRAM_CPU_ACCESS
] = heap_usage
;
1404 heap_usage
= device
->ws
->query_value(device
->ws
,
1405 RADEON_ALLOCATED_GTT
);
1407 heap_budget
= gtt_size
-
1408 device
->ws
->query_value(device
->ws
, RADEON_GTT_USAGE
) +
1411 memoryBudget
->heapBudget
[RADV_MEM_HEAP_GTT
] = heap_budget
;
1412 memoryBudget
->heapUsage
[RADV_MEM_HEAP_GTT
] = heap_usage
;
1415 /* The heapBudget and heapUsage values must be zero for array elements
1416 * greater than or equal to
1417 * VkPhysicalDeviceMemoryProperties::memoryHeapCount.
1419 for (uint32_t i
= memory_properties
->memoryHeapCount
; i
< VK_MAX_MEMORY_HEAPS
; i
++) {
1420 memoryBudget
->heapBudget
[i
] = 0;
1421 memoryBudget
->heapUsage
[i
] = 0;
1425 void radv_GetPhysicalDeviceMemoryProperties2(
1426 VkPhysicalDevice physicalDevice
,
1427 VkPhysicalDeviceMemoryProperties2
*pMemoryProperties
)
1429 radv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
1430 &pMemoryProperties
->memoryProperties
);
1432 VkPhysicalDeviceMemoryBudgetPropertiesEXT
*memory_budget
=
1433 vk_find_struct(pMemoryProperties
->pNext
,
1434 PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT
);
1436 radv_get_memory_budget_properties(physicalDevice
, memory_budget
);
1439 VkResult
radv_GetMemoryHostPointerPropertiesEXT(
1441 VkExternalMemoryHandleTypeFlagBits handleType
,
1442 const void *pHostPointer
,
1443 VkMemoryHostPointerPropertiesEXT
*pMemoryHostPointerProperties
)
1445 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1449 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
: {
1450 const struct radv_physical_device
*physical_device
= device
->physical_device
;
1451 uint32_t memoryTypeBits
= 0;
1452 for (int i
= 0; i
< physical_device
->memory_properties
.memoryTypeCount
; i
++) {
1453 if (physical_device
->mem_type_indices
[i
] == RADV_MEM_TYPE_GTT_CACHED
) {
1454 memoryTypeBits
= (1 << i
);
1458 pMemoryHostPointerProperties
->memoryTypeBits
= memoryTypeBits
;
1462 return VK_ERROR_INVALID_EXTERNAL_HANDLE
;
1466 static enum radeon_ctx_priority
1467 radv_get_queue_global_priority(const VkDeviceQueueGlobalPriorityCreateInfoEXT
*pObj
)
1469 /* Default to MEDIUM when a specific global priority isn't requested */
1471 return RADEON_CTX_PRIORITY_MEDIUM
;
1473 switch(pObj
->globalPriority
) {
1474 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT
:
1475 return RADEON_CTX_PRIORITY_REALTIME
;
1476 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT
:
1477 return RADEON_CTX_PRIORITY_HIGH
;
1478 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
:
1479 return RADEON_CTX_PRIORITY_MEDIUM
;
1480 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT
:
1481 return RADEON_CTX_PRIORITY_LOW
;
1483 unreachable("Illegal global priority value");
1484 return RADEON_CTX_PRIORITY_INVALID
;
1489 radv_queue_init(struct radv_device
*device
, struct radv_queue
*queue
,
1490 uint32_t queue_family_index
, int idx
,
1491 VkDeviceQueueCreateFlags flags
,
1492 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*global_priority
)
1494 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1495 queue
->device
= device
;
1496 queue
->queue_family_index
= queue_family_index
;
1497 queue
->queue_idx
= idx
;
1498 queue
->priority
= radv_get_queue_global_priority(global_priority
);
1499 queue
->flags
= flags
;
1501 queue
->hw_ctx
= device
->ws
->ctx_create(device
->ws
, queue
->priority
);
1503 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
1509 radv_queue_finish(struct radv_queue
*queue
)
1512 queue
->device
->ws
->ctx_destroy(queue
->hw_ctx
);
1514 if (queue
->initial_full_flush_preamble_cs
)
1515 queue
->device
->ws
->cs_destroy(queue
->initial_full_flush_preamble_cs
);
1516 if (queue
->initial_preamble_cs
)
1517 queue
->device
->ws
->cs_destroy(queue
->initial_preamble_cs
);
1518 if (queue
->continue_preamble_cs
)
1519 queue
->device
->ws
->cs_destroy(queue
->continue_preamble_cs
);
1520 if (queue
->descriptor_bo
)
1521 queue
->device
->ws
->buffer_destroy(queue
->descriptor_bo
);
1522 if (queue
->scratch_bo
)
1523 queue
->device
->ws
->buffer_destroy(queue
->scratch_bo
);
1524 if (queue
->esgs_ring_bo
)
1525 queue
->device
->ws
->buffer_destroy(queue
->esgs_ring_bo
);
1526 if (queue
->gsvs_ring_bo
)
1527 queue
->device
->ws
->buffer_destroy(queue
->gsvs_ring_bo
);
1528 if (queue
->tess_rings_bo
)
1529 queue
->device
->ws
->buffer_destroy(queue
->tess_rings_bo
);
1530 if (queue
->compute_scratch_bo
)
1531 queue
->device
->ws
->buffer_destroy(queue
->compute_scratch_bo
);
1535 radv_bo_list_init(struct radv_bo_list
*bo_list
)
1537 pthread_mutex_init(&bo_list
->mutex
, NULL
);
1538 bo_list
->list
.count
= bo_list
->capacity
= 0;
1539 bo_list
->list
.bos
= NULL
;
1543 radv_bo_list_finish(struct radv_bo_list
*bo_list
)
1545 free(bo_list
->list
.bos
);
1546 pthread_mutex_destroy(&bo_list
->mutex
);
1549 static VkResult
radv_bo_list_add(struct radv_device
*device
,
1550 struct radeon_winsys_bo
*bo
)
1552 struct radv_bo_list
*bo_list
= &device
->bo_list
;
1557 if (unlikely(!device
->use_global_bo_list
))
1560 pthread_mutex_lock(&bo_list
->mutex
);
1561 if (bo_list
->list
.count
== bo_list
->capacity
) {
1562 unsigned capacity
= MAX2(4, bo_list
->capacity
* 2);
1563 void *data
= realloc(bo_list
->list
.bos
, capacity
* sizeof(struct radeon_winsys_bo
*));
1566 pthread_mutex_unlock(&bo_list
->mutex
);
1567 return VK_ERROR_OUT_OF_HOST_MEMORY
;
1570 bo_list
->list
.bos
= (struct radeon_winsys_bo
**)data
;
1571 bo_list
->capacity
= capacity
;
1574 bo_list
->list
.bos
[bo_list
->list
.count
++] = bo
;
1575 pthread_mutex_unlock(&bo_list
->mutex
);
1579 static void radv_bo_list_remove(struct radv_device
*device
,
1580 struct radeon_winsys_bo
*bo
)
1582 struct radv_bo_list
*bo_list
= &device
->bo_list
;
1587 if (unlikely(!device
->use_global_bo_list
))
1590 pthread_mutex_lock(&bo_list
->mutex
);
1591 for(unsigned i
= 0; i
< bo_list
->list
.count
; ++i
) {
1592 if (bo_list
->list
.bos
[i
] == bo
) {
1593 bo_list
->list
.bos
[i
] = bo_list
->list
.bos
[bo_list
->list
.count
- 1];
1594 --bo_list
->list
.count
;
1598 pthread_mutex_unlock(&bo_list
->mutex
);
1602 radv_device_init_gs_info(struct radv_device
*device
)
1604 device
->gs_table_depth
= ac_get_gs_table_depth(device
->physical_device
->rad_info
.chip_class
,
1605 device
->physical_device
->rad_info
.family
);
1608 static int radv_get_device_extension_index(const char *name
)
1610 for (unsigned i
= 0; i
< RADV_DEVICE_EXTENSION_COUNT
; ++i
) {
1611 if (strcmp(name
, radv_device_extensions
[i
].extensionName
) == 0)
1618 radv_get_int_debug_option(const char *name
, int default_value
)
1625 result
= default_value
;
1629 result
= strtol(str
, &endptr
, 0);
1630 if (str
== endptr
) {
1631 /* No digits founs. */
1632 result
= default_value
;
1639 VkResult
radv_CreateDevice(
1640 VkPhysicalDevice physicalDevice
,
1641 const VkDeviceCreateInfo
* pCreateInfo
,
1642 const VkAllocationCallbacks
* pAllocator
,
1645 RADV_FROM_HANDLE(radv_physical_device
, physical_device
, physicalDevice
);
1647 struct radv_device
*device
;
1649 bool keep_shader_info
= false;
1651 /* Check enabled features */
1652 if (pCreateInfo
->pEnabledFeatures
) {
1653 VkPhysicalDeviceFeatures supported_features
;
1654 radv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
1655 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
1656 VkBool32
*enabled_feature
= (VkBool32
*)pCreateInfo
->pEnabledFeatures
;
1657 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
1658 for (uint32_t i
= 0; i
< num_features
; i
++) {
1659 if (enabled_feature
[i
] && !supported_feature
[i
])
1660 return vk_error(physical_device
->instance
, VK_ERROR_FEATURE_NOT_PRESENT
);
1664 device
= vk_zalloc2(&physical_device
->instance
->alloc
, pAllocator
,
1666 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
1668 return vk_error(physical_device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
1670 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1671 device
->instance
= physical_device
->instance
;
1672 device
->physical_device
= physical_device
;
1674 device
->ws
= physical_device
->ws
;
1676 device
->alloc
= *pAllocator
;
1678 device
->alloc
= physical_device
->instance
->alloc
;
1680 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
1681 const char *ext_name
= pCreateInfo
->ppEnabledExtensionNames
[i
];
1682 int index
= radv_get_device_extension_index(ext_name
);
1683 if (index
< 0 || !physical_device
->supported_extensions
.extensions
[index
]) {
1684 vk_free(&device
->alloc
, device
);
1685 return vk_error(physical_device
->instance
, VK_ERROR_EXTENSION_NOT_PRESENT
);
1688 device
->enabled_extensions
.extensions
[index
] = true;
1691 keep_shader_info
= device
->enabled_extensions
.AMD_shader_info
;
1693 /* With update after bind we can't attach bo's to the command buffer
1694 * from the descriptor set anymore, so we have to use a global BO list.
1696 device
->use_global_bo_list
=
1697 device
->enabled_extensions
.EXT_descriptor_indexing
;
1699 mtx_init(&device
->shader_slab_mutex
, mtx_plain
);
1700 list_inithead(&device
->shader_slabs
);
1702 radv_bo_list_init(&device
->bo_list
);
1704 for (unsigned i
= 0; i
< pCreateInfo
->queueCreateInfoCount
; i
++) {
1705 const VkDeviceQueueCreateInfo
*queue_create
= &pCreateInfo
->pQueueCreateInfos
[i
];
1706 uint32_t qfi
= queue_create
->queueFamilyIndex
;
1707 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*global_priority
=
1708 vk_find_struct_const(queue_create
->pNext
, DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT
);
1710 assert(!global_priority
|| device
->physical_device
->rad_info
.has_ctx_priority
);
1712 device
->queues
[qfi
] = vk_alloc(&device
->alloc
,
1713 queue_create
->queueCount
* sizeof(struct radv_queue
), 8, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
1714 if (!device
->queues
[qfi
]) {
1715 result
= VK_ERROR_OUT_OF_HOST_MEMORY
;
1719 memset(device
->queues
[qfi
], 0, queue_create
->queueCount
* sizeof(struct radv_queue
));
1721 device
->queue_count
[qfi
] = queue_create
->queueCount
;
1723 for (unsigned q
= 0; q
< queue_create
->queueCount
; q
++) {
1724 result
= radv_queue_init(device
, &device
->queues
[qfi
][q
],
1725 qfi
, q
, queue_create
->flags
,
1727 if (result
!= VK_SUCCESS
)
1732 device
->pbb_allowed
= device
->physical_device
->rad_info
.chip_class
>= GFX9
&&
1733 !(device
->instance
->debug_flags
& RADV_DEBUG_NOBINNING
);
1735 /* Disabled and not implemented for now. */
1736 device
->dfsm_allowed
= device
->pbb_allowed
&&
1737 (device
->physical_device
->rad_info
.family
== CHIP_RAVEN
||
1738 device
->physical_device
->rad_info
.family
== CHIP_RAVEN2
);
1741 device
->always_use_syncobj
= device
->physical_device
->rad_info
.has_syncobj_wait_for_submit
;
1744 /* The maximum number of scratch waves. Scratch space isn't divided
1745 * evenly between CUs. The number is only a function of the number of CUs.
1746 * We can decrease the constant to decrease the scratch buffer size.
1748 * sctx->scratch_waves must be >= the maximum possible size of
1749 * 1 threadgroup, so that the hw doesn't hang from being unable
1752 * The recommended value is 4 per CU at most. Higher numbers don't
1753 * bring much benefit, but they still occupy chip resources (think
1754 * async compute). I've seen ~2% performance difference between 4 and 32.
1756 uint32_t max_threads_per_block
= 2048;
1757 device
->scratch_waves
= MAX2(32 * physical_device
->rad_info
.num_good_compute_units
,
1758 max_threads_per_block
/ 64);
1760 device
->dispatch_initiator
= S_00B800_COMPUTE_SHADER_EN(1);
1762 if (device
->physical_device
->rad_info
.chip_class
>= CIK
) {
1763 /* If the KMD allows it (there is a KMD hw register for it),
1764 * allow launching waves out-of-order.
1766 device
->dispatch_initiator
|= S_00B800_ORDER_MODE(1);
1769 radv_device_init_gs_info(device
);
1771 device
->tess_offchip_block_dw_size
=
1772 device
->physical_device
->rad_info
.family
== CHIP_HAWAII
? 4096 : 8192;
1773 device
->has_distributed_tess
=
1774 device
->physical_device
->rad_info
.chip_class
>= VI
&&
1775 device
->physical_device
->rad_info
.max_se
>= 2;
1777 if (getenv("RADV_TRACE_FILE")) {
1778 const char *filename
= getenv("RADV_TRACE_FILE");
1780 keep_shader_info
= true;
1782 if (!radv_init_trace(device
))
1785 fprintf(stderr
, "*****************************************************************************\n");
1786 fprintf(stderr
, "* WARNING: RADV_TRACE_FILE is costly and should only be used for debugging! *\n");
1787 fprintf(stderr
, "*****************************************************************************\n");
1789 fprintf(stderr
, "Trace file will be dumped to %s\n", filename
);
1790 radv_dump_enabled_options(device
, stderr
);
1793 device
->keep_shader_info
= keep_shader_info
;
1795 result
= radv_device_init_meta(device
);
1796 if (result
!= VK_SUCCESS
)
1799 radv_device_init_msaa(device
);
1801 for (int family
= 0; family
< RADV_MAX_QUEUE_FAMILIES
; ++family
) {
1802 device
->empty_cs
[family
] = device
->ws
->cs_create(device
->ws
, family
);
1804 case RADV_QUEUE_GENERAL
:
1805 radeon_emit(device
->empty_cs
[family
], PKT3(PKT3_CONTEXT_CONTROL
, 1, 0));
1806 radeon_emit(device
->empty_cs
[family
], CONTEXT_CONTROL_LOAD_ENABLE(1));
1807 radeon_emit(device
->empty_cs
[family
], CONTEXT_CONTROL_SHADOW_ENABLE(1));
1809 case RADV_QUEUE_COMPUTE
:
1810 radeon_emit(device
->empty_cs
[family
], PKT3(PKT3_NOP
, 0, 0));
1811 radeon_emit(device
->empty_cs
[family
], 0);
1814 device
->ws
->cs_finalize(device
->empty_cs
[family
]);
1817 if (device
->physical_device
->rad_info
.chip_class
>= CIK
)
1818 cik_create_gfx_config(device
);
1820 VkPipelineCacheCreateInfo ci
;
1821 ci
.sType
= VK_STRUCTURE_TYPE_PIPELINE_CACHE_CREATE_INFO
;
1824 ci
.pInitialData
= NULL
;
1825 ci
.initialDataSize
= 0;
1827 result
= radv_CreatePipelineCache(radv_device_to_handle(device
),
1829 if (result
!= VK_SUCCESS
)
1832 device
->mem_cache
= radv_pipeline_cache_from_handle(pc
);
1834 device
->force_aniso
=
1835 MIN2(16, radv_get_int_debug_option("RADV_TEX_ANISO", -1));
1836 if (device
->force_aniso
>= 0) {
1837 fprintf(stderr
, "radv: Forcing anisotropy filter to %ix\n",
1838 1 << util_logbase2(device
->force_aniso
));
1841 *pDevice
= radv_device_to_handle(device
);
1845 radv_device_finish_meta(device
);
1847 radv_bo_list_finish(&device
->bo_list
);
1849 if (device
->trace_bo
)
1850 device
->ws
->buffer_destroy(device
->trace_bo
);
1852 if (device
->gfx_init
)
1853 device
->ws
->buffer_destroy(device
->gfx_init
);
1855 for (unsigned i
= 0; i
< RADV_MAX_QUEUE_FAMILIES
; i
++) {
1856 for (unsigned q
= 0; q
< device
->queue_count
[i
]; q
++)
1857 radv_queue_finish(&device
->queues
[i
][q
]);
1858 if (device
->queue_count
[i
])
1859 vk_free(&device
->alloc
, device
->queues
[i
]);
1862 vk_free(&device
->alloc
, device
);
1866 void radv_DestroyDevice(
1868 const VkAllocationCallbacks
* pAllocator
)
1870 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1875 if (device
->trace_bo
)
1876 device
->ws
->buffer_destroy(device
->trace_bo
);
1878 if (device
->gfx_init
)
1879 device
->ws
->buffer_destroy(device
->gfx_init
);
1881 for (unsigned i
= 0; i
< RADV_MAX_QUEUE_FAMILIES
; i
++) {
1882 for (unsigned q
= 0; q
< device
->queue_count
[i
]; q
++)
1883 radv_queue_finish(&device
->queues
[i
][q
]);
1884 if (device
->queue_count
[i
])
1885 vk_free(&device
->alloc
, device
->queues
[i
]);
1886 if (device
->empty_cs
[i
])
1887 device
->ws
->cs_destroy(device
->empty_cs
[i
]);
1889 radv_device_finish_meta(device
);
1891 VkPipelineCache pc
= radv_pipeline_cache_to_handle(device
->mem_cache
);
1892 radv_DestroyPipelineCache(radv_device_to_handle(device
), pc
, NULL
);
1894 radv_destroy_shader_slabs(device
);
1896 radv_bo_list_finish(&device
->bo_list
);
1897 vk_free(&device
->alloc
, device
);
1900 VkResult
radv_EnumerateInstanceLayerProperties(
1901 uint32_t* pPropertyCount
,
1902 VkLayerProperties
* pProperties
)
1904 if (pProperties
== NULL
) {
1905 *pPropertyCount
= 0;
1909 /* None supported at this time */
1910 return vk_error(NULL
, VK_ERROR_LAYER_NOT_PRESENT
);
1913 VkResult
radv_EnumerateDeviceLayerProperties(
1914 VkPhysicalDevice physicalDevice
,
1915 uint32_t* pPropertyCount
,
1916 VkLayerProperties
* pProperties
)
1918 if (pProperties
== NULL
) {
1919 *pPropertyCount
= 0;
1923 /* None supported at this time */
1924 return vk_error(NULL
, VK_ERROR_LAYER_NOT_PRESENT
);
1927 void radv_GetDeviceQueue2(
1929 const VkDeviceQueueInfo2
* pQueueInfo
,
1932 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1933 struct radv_queue
*queue
;
1935 queue
= &device
->queues
[pQueueInfo
->queueFamilyIndex
][pQueueInfo
->queueIndex
];
1936 if (pQueueInfo
->flags
!= queue
->flags
) {
1937 /* From the Vulkan 1.1.70 spec:
1939 * "The queue returned by vkGetDeviceQueue2 must have the same
1940 * flags value from this structure as that used at device
1941 * creation time in a VkDeviceQueueCreateInfo instance. If no
1942 * matching flags were specified at device creation time then
1943 * pQueue will return VK_NULL_HANDLE."
1945 *pQueue
= VK_NULL_HANDLE
;
1949 *pQueue
= radv_queue_to_handle(queue
);
1952 void radv_GetDeviceQueue(
1954 uint32_t queueFamilyIndex
,
1955 uint32_t queueIndex
,
1958 const VkDeviceQueueInfo2 info
= (VkDeviceQueueInfo2
) {
1959 .sType
= VK_STRUCTURE_TYPE_DEVICE_QUEUE_INFO_2
,
1960 .queueFamilyIndex
= queueFamilyIndex
,
1961 .queueIndex
= queueIndex
1964 radv_GetDeviceQueue2(_device
, &info
, pQueue
);
1968 fill_geom_tess_rings(struct radv_queue
*queue
,
1970 bool add_sample_positions
,
1971 uint32_t esgs_ring_size
,
1972 struct radeon_winsys_bo
*esgs_ring_bo
,
1973 uint32_t gsvs_ring_size
,
1974 struct radeon_winsys_bo
*gsvs_ring_bo
,
1975 uint32_t tess_factor_ring_size
,
1976 uint32_t tess_offchip_ring_offset
,
1977 uint32_t tess_offchip_ring_size
,
1978 struct radeon_winsys_bo
*tess_rings_bo
)
1980 uint32_t *desc
= &map
[4];
1983 uint64_t esgs_va
= radv_buffer_get_va(esgs_ring_bo
);
1985 /* stride 0, num records - size, add tid, swizzle, elsize4,
1988 desc
[1] = S_008F04_BASE_ADDRESS_HI(esgs_va
>> 32) |
1989 S_008F04_STRIDE(0) |
1990 S_008F04_SWIZZLE_ENABLE(true);
1991 desc
[2] = esgs_ring_size
;
1992 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1993 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1994 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1995 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1996 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1997 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1998 S_008F0C_ELEMENT_SIZE(1) |
1999 S_008F0C_INDEX_STRIDE(3) |
2000 S_008F0C_ADD_TID_ENABLE(true);
2002 /* GS entry for ES->GS ring */
2003 /* stride 0, num records - size, elsize0,
2006 desc
[5] = S_008F04_BASE_ADDRESS_HI(esgs_va
>> 32)|
2007 S_008F04_STRIDE(0) |
2008 S_008F04_SWIZZLE_ENABLE(false);
2009 desc
[6] = esgs_ring_size
;
2010 desc
[7] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
2011 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
2012 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
2013 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
2014 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
2015 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
2016 S_008F0C_ELEMENT_SIZE(0) |
2017 S_008F0C_INDEX_STRIDE(0) |
2018 S_008F0C_ADD_TID_ENABLE(false);
2024 uint64_t gsvs_va
= radv_buffer_get_va(gsvs_ring_bo
);
2026 /* VS entry for GS->VS ring */
2027 /* stride 0, num records - size, elsize0,
2030 desc
[1] = S_008F04_BASE_ADDRESS_HI(gsvs_va
>> 32)|
2031 S_008F04_STRIDE(0) |
2032 S_008F04_SWIZZLE_ENABLE(false);
2033 desc
[2] = gsvs_ring_size
;
2034 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
2035 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
2036 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
2037 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
2038 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
2039 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
2040 S_008F0C_ELEMENT_SIZE(0) |
2041 S_008F0C_INDEX_STRIDE(0) |
2042 S_008F0C_ADD_TID_ENABLE(false);
2044 /* stride gsvs_itemsize, num records 64
2045 elsize 4, index stride 16 */
2046 /* shader will patch stride and desc[2] */
2048 desc
[5] = S_008F04_BASE_ADDRESS_HI(gsvs_va
>> 32)|
2049 S_008F04_STRIDE(0) |
2050 S_008F04_SWIZZLE_ENABLE(true);
2052 desc
[7] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
2053 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
2054 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
2055 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
2056 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
2057 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
2058 S_008F0C_ELEMENT_SIZE(1) |
2059 S_008F0C_INDEX_STRIDE(1) |
2060 S_008F0C_ADD_TID_ENABLE(true);
2065 if (tess_rings_bo
) {
2066 uint64_t tess_va
= radv_buffer_get_va(tess_rings_bo
);
2067 uint64_t tess_offchip_va
= tess_va
+ tess_offchip_ring_offset
;
2070 desc
[1] = S_008F04_BASE_ADDRESS_HI(tess_va
>> 32) |
2071 S_008F04_STRIDE(0) |
2072 S_008F04_SWIZZLE_ENABLE(false);
2073 desc
[2] = tess_factor_ring_size
;
2074 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
2075 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
2076 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
2077 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
2078 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
2079 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
2080 S_008F0C_ELEMENT_SIZE(0) |
2081 S_008F0C_INDEX_STRIDE(0) |
2082 S_008F0C_ADD_TID_ENABLE(false);
2084 desc
[4] = tess_offchip_va
;
2085 desc
[5] = S_008F04_BASE_ADDRESS_HI(tess_offchip_va
>> 32) |
2086 S_008F04_STRIDE(0) |
2087 S_008F04_SWIZZLE_ENABLE(false);
2088 desc
[6] = tess_offchip_ring_size
;
2089 desc
[7] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
2090 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
2091 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
2092 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
2093 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
2094 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
2095 S_008F0C_ELEMENT_SIZE(0) |
2096 S_008F0C_INDEX_STRIDE(0) |
2097 S_008F0C_ADD_TID_ENABLE(false);
2102 if (add_sample_positions
) {
2103 /* add sample positions after all rings */
2104 memcpy(desc
, queue
->device
->sample_locations_1x
, 8);
2106 memcpy(desc
, queue
->device
->sample_locations_2x
, 16);
2108 memcpy(desc
, queue
->device
->sample_locations_4x
, 32);
2110 memcpy(desc
, queue
->device
->sample_locations_8x
, 64);
2115 radv_get_hs_offchip_param(struct radv_device
*device
, uint32_t *max_offchip_buffers_p
)
2117 bool double_offchip_buffers
= device
->physical_device
->rad_info
.chip_class
>= CIK
&&
2118 device
->physical_device
->rad_info
.family
!= CHIP_CARRIZO
&&
2119 device
->physical_device
->rad_info
.family
!= CHIP_STONEY
;
2120 unsigned max_offchip_buffers_per_se
= double_offchip_buffers
? 128 : 64;
2121 unsigned max_offchip_buffers
;
2122 unsigned offchip_granularity
;
2123 unsigned hs_offchip_param
;
2127 * This must be one less than the maximum number due to a hw limitation.
2128 * Various hardware bugs in SI, CIK, and GFX9 need this.
2131 * Vega10 should limit max_offchip_buffers to 508 (4 * 127).
2132 * Gfx7 should limit max_offchip_buffers to 508
2133 * Gfx6 should limit max_offchip_buffers to 126 (2 * 63)
2135 * Follow AMDVLK here.
2137 if (device
->physical_device
->rad_info
.family
== CHIP_VEGA10
||
2138 device
->physical_device
->rad_info
.chip_class
== CIK
||
2139 device
->physical_device
->rad_info
.chip_class
== SI
)
2140 --max_offchip_buffers_per_se
;
2142 max_offchip_buffers
= max_offchip_buffers_per_se
*
2143 device
->physical_device
->rad_info
.max_se
;
2145 /* Hawaii has a bug with offchip buffers > 256 that can be worked
2146 * around by setting 4K granularity.
2148 if (device
->tess_offchip_block_dw_size
== 4096) {
2149 assert(device
->physical_device
->rad_info
.family
== CHIP_HAWAII
);
2150 offchip_granularity
= V_03093C_X_4K_DWORDS
;
2152 assert(device
->tess_offchip_block_dw_size
== 8192);
2153 offchip_granularity
= V_03093C_X_8K_DWORDS
;
2156 switch (device
->physical_device
->rad_info
.chip_class
) {
2158 max_offchip_buffers
= MIN2(max_offchip_buffers
, 126);
2164 max_offchip_buffers
= MIN2(max_offchip_buffers
, 508);
2168 *max_offchip_buffers_p
= max_offchip_buffers
;
2169 if (device
->physical_device
->rad_info
.chip_class
>= CIK
) {
2170 if (device
->physical_device
->rad_info
.chip_class
>= VI
)
2171 --max_offchip_buffers
;
2173 S_03093C_OFFCHIP_BUFFERING(max_offchip_buffers
) |
2174 S_03093C_OFFCHIP_GRANULARITY(offchip_granularity
);
2177 S_0089B0_OFFCHIP_BUFFERING(max_offchip_buffers
);
2179 return hs_offchip_param
;
2183 radv_emit_gs_ring_sizes(struct radv_queue
*queue
, struct radeon_cmdbuf
*cs
,
2184 struct radeon_winsys_bo
*esgs_ring_bo
,
2185 uint32_t esgs_ring_size
,
2186 struct radeon_winsys_bo
*gsvs_ring_bo
,
2187 uint32_t gsvs_ring_size
)
2189 if (!esgs_ring_bo
&& !gsvs_ring_bo
)
2193 radv_cs_add_buffer(queue
->device
->ws
, cs
, esgs_ring_bo
);
2196 radv_cs_add_buffer(queue
->device
->ws
, cs
, gsvs_ring_bo
);
2198 if (queue
->device
->physical_device
->rad_info
.chip_class
>= CIK
) {
2199 radeon_set_uconfig_reg_seq(cs
, R_030900_VGT_ESGS_RING_SIZE
, 2);
2200 radeon_emit(cs
, esgs_ring_size
>> 8);
2201 radeon_emit(cs
, gsvs_ring_size
>> 8);
2203 radeon_set_config_reg_seq(cs
, R_0088C8_VGT_ESGS_RING_SIZE
, 2);
2204 radeon_emit(cs
, esgs_ring_size
>> 8);
2205 radeon_emit(cs
, gsvs_ring_size
>> 8);
2210 radv_emit_tess_factor_ring(struct radv_queue
*queue
, struct radeon_cmdbuf
*cs
,
2211 unsigned hs_offchip_param
, unsigned tf_ring_size
,
2212 struct radeon_winsys_bo
*tess_rings_bo
)
2219 tf_va
= radv_buffer_get_va(tess_rings_bo
);
2221 radv_cs_add_buffer(queue
->device
->ws
, cs
, tess_rings_bo
);
2223 if (queue
->device
->physical_device
->rad_info
.chip_class
>= CIK
) {
2224 radeon_set_uconfig_reg(cs
, R_030938_VGT_TF_RING_SIZE
,
2225 S_030938_SIZE(tf_ring_size
/ 4));
2226 radeon_set_uconfig_reg(cs
, R_030940_VGT_TF_MEMORY_BASE
,
2228 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
2229 radeon_set_uconfig_reg(cs
, R_030944_VGT_TF_MEMORY_BASE_HI
,
2230 S_030944_BASE_HI(tf_va
>> 40));
2232 radeon_set_uconfig_reg(cs
, R_03093C_VGT_HS_OFFCHIP_PARAM
,
2235 radeon_set_config_reg(cs
, R_008988_VGT_TF_RING_SIZE
,
2236 S_008988_SIZE(tf_ring_size
/ 4));
2237 radeon_set_config_reg(cs
, R_0089B8_VGT_TF_MEMORY_BASE
,
2239 radeon_set_config_reg(cs
, R_0089B0_VGT_HS_OFFCHIP_PARAM
,
2245 radv_emit_compute_scratch(struct radv_queue
*queue
, struct radeon_cmdbuf
*cs
,
2246 struct radeon_winsys_bo
*compute_scratch_bo
)
2248 uint64_t scratch_va
;
2250 if (!compute_scratch_bo
)
2253 scratch_va
= radv_buffer_get_va(compute_scratch_bo
);
2255 radv_cs_add_buffer(queue
->device
->ws
, cs
, compute_scratch_bo
);
2257 radeon_set_sh_reg_seq(cs
, R_00B900_COMPUTE_USER_DATA_0
, 2);
2258 radeon_emit(cs
, scratch_va
);
2259 radeon_emit(cs
, S_008F04_BASE_ADDRESS_HI(scratch_va
>> 32) |
2260 S_008F04_SWIZZLE_ENABLE(1));
2264 radv_emit_global_shader_pointers(struct radv_queue
*queue
,
2265 struct radeon_cmdbuf
*cs
,
2266 struct radeon_winsys_bo
*descriptor_bo
)
2273 va
= radv_buffer_get_va(descriptor_bo
);
2275 radv_cs_add_buffer(queue
->device
->ws
, cs
, descriptor_bo
);
2277 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
2278 uint32_t regs
[] = {R_00B030_SPI_SHADER_USER_DATA_PS_0
,
2279 R_00B130_SPI_SHADER_USER_DATA_VS_0
,
2280 R_00B208_SPI_SHADER_USER_DATA_ADDR_LO_GS
,
2281 R_00B408_SPI_SHADER_USER_DATA_ADDR_LO_HS
};
2283 for (int i
= 0; i
< ARRAY_SIZE(regs
); ++i
) {
2284 radv_emit_shader_pointer(queue
->device
, cs
, regs
[i
],
2288 uint32_t regs
[] = {R_00B030_SPI_SHADER_USER_DATA_PS_0
,
2289 R_00B130_SPI_SHADER_USER_DATA_VS_0
,
2290 R_00B230_SPI_SHADER_USER_DATA_GS_0
,
2291 R_00B330_SPI_SHADER_USER_DATA_ES_0
,
2292 R_00B430_SPI_SHADER_USER_DATA_HS_0
,
2293 R_00B530_SPI_SHADER_USER_DATA_LS_0
};
2295 for (int i
= 0; i
< ARRAY_SIZE(regs
); ++i
) {
2296 radv_emit_shader_pointer(queue
->device
, cs
, regs
[i
],
2303 radv_init_graphics_state(struct radeon_cmdbuf
*cs
, struct radv_queue
*queue
)
2305 struct radv_device
*device
= queue
->device
;
2307 if (device
->gfx_init
) {
2308 uint64_t va
= radv_buffer_get_va(device
->gfx_init
);
2310 radeon_emit(cs
, PKT3(PKT3_INDIRECT_BUFFER_CIK
, 2, 0));
2311 radeon_emit(cs
, va
);
2312 radeon_emit(cs
, va
>> 32);
2313 radeon_emit(cs
, device
->gfx_init_size_dw
& 0xffff);
2315 radv_cs_add_buffer(device
->ws
, cs
, device
->gfx_init
);
2317 struct radv_physical_device
*physical_device
= device
->physical_device
;
2318 si_emit_graphics(physical_device
, cs
);
2323 radv_init_compute_state(struct radeon_cmdbuf
*cs
, struct radv_queue
*queue
)
2325 struct radv_physical_device
*physical_device
= queue
->device
->physical_device
;
2326 si_emit_compute(physical_device
, cs
);
2330 radv_get_preamble_cs(struct radv_queue
*queue
,
2331 uint32_t scratch_size
,
2332 uint32_t compute_scratch_size
,
2333 uint32_t esgs_ring_size
,
2334 uint32_t gsvs_ring_size
,
2335 bool needs_tess_rings
,
2336 bool needs_sample_positions
,
2337 struct radeon_cmdbuf
**initial_full_flush_preamble_cs
,
2338 struct radeon_cmdbuf
**initial_preamble_cs
,
2339 struct radeon_cmdbuf
**continue_preamble_cs
)
2341 struct radeon_winsys_bo
*scratch_bo
= NULL
;
2342 struct radeon_winsys_bo
*descriptor_bo
= NULL
;
2343 struct radeon_winsys_bo
*compute_scratch_bo
= NULL
;
2344 struct radeon_winsys_bo
*esgs_ring_bo
= NULL
;
2345 struct radeon_winsys_bo
*gsvs_ring_bo
= NULL
;
2346 struct radeon_winsys_bo
*tess_rings_bo
= NULL
;
2347 struct radeon_cmdbuf
*dest_cs
[3] = {0};
2348 bool add_tess_rings
= false, add_sample_positions
= false;
2349 unsigned tess_factor_ring_size
= 0, tess_offchip_ring_size
= 0;
2350 unsigned max_offchip_buffers
;
2351 unsigned hs_offchip_param
= 0;
2352 unsigned tess_offchip_ring_offset
;
2353 uint32_t ring_bo_flags
= RADEON_FLAG_NO_CPU_ACCESS
| RADEON_FLAG_NO_INTERPROCESS_SHARING
;
2354 if (!queue
->has_tess_rings
) {
2355 if (needs_tess_rings
)
2356 add_tess_rings
= true;
2358 if (!queue
->has_sample_positions
) {
2359 if (needs_sample_positions
)
2360 add_sample_positions
= true;
2362 tess_factor_ring_size
= 32768 * queue
->device
->physical_device
->rad_info
.max_se
;
2363 hs_offchip_param
= radv_get_hs_offchip_param(queue
->device
,
2364 &max_offchip_buffers
);
2365 tess_offchip_ring_offset
= align(tess_factor_ring_size
, 64 * 1024);
2366 tess_offchip_ring_size
= max_offchip_buffers
*
2367 queue
->device
->tess_offchip_block_dw_size
* 4;
2369 if (scratch_size
<= queue
->scratch_size
&&
2370 compute_scratch_size
<= queue
->compute_scratch_size
&&
2371 esgs_ring_size
<= queue
->esgs_ring_size
&&
2372 gsvs_ring_size
<= queue
->gsvs_ring_size
&&
2373 !add_tess_rings
&& !add_sample_positions
&&
2374 queue
->initial_preamble_cs
) {
2375 *initial_full_flush_preamble_cs
= queue
->initial_full_flush_preamble_cs
;
2376 *initial_preamble_cs
= queue
->initial_preamble_cs
;
2377 *continue_preamble_cs
= queue
->continue_preamble_cs
;
2378 if (!scratch_size
&& !compute_scratch_size
&& !esgs_ring_size
&& !gsvs_ring_size
)
2379 *continue_preamble_cs
= NULL
;
2383 if (scratch_size
> queue
->scratch_size
) {
2384 scratch_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
2389 RADV_BO_PRIORITY_SCRATCH
);
2393 scratch_bo
= queue
->scratch_bo
;
2395 if (compute_scratch_size
> queue
->compute_scratch_size
) {
2396 compute_scratch_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
2397 compute_scratch_size
,
2401 RADV_BO_PRIORITY_SCRATCH
);
2402 if (!compute_scratch_bo
)
2406 compute_scratch_bo
= queue
->compute_scratch_bo
;
2408 if (esgs_ring_size
> queue
->esgs_ring_size
) {
2409 esgs_ring_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
2414 RADV_BO_PRIORITY_SCRATCH
);
2418 esgs_ring_bo
= queue
->esgs_ring_bo
;
2419 esgs_ring_size
= queue
->esgs_ring_size
;
2422 if (gsvs_ring_size
> queue
->gsvs_ring_size
) {
2423 gsvs_ring_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
2428 RADV_BO_PRIORITY_SCRATCH
);
2432 gsvs_ring_bo
= queue
->gsvs_ring_bo
;
2433 gsvs_ring_size
= queue
->gsvs_ring_size
;
2436 if (add_tess_rings
) {
2437 tess_rings_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
2438 tess_offchip_ring_offset
+ tess_offchip_ring_size
,
2442 RADV_BO_PRIORITY_SCRATCH
);
2446 tess_rings_bo
= queue
->tess_rings_bo
;
2449 if (scratch_bo
!= queue
->scratch_bo
||
2450 esgs_ring_bo
!= queue
->esgs_ring_bo
||
2451 gsvs_ring_bo
!= queue
->gsvs_ring_bo
||
2452 tess_rings_bo
!= queue
->tess_rings_bo
||
2453 add_sample_positions
) {
2455 if (gsvs_ring_bo
|| esgs_ring_bo
||
2456 tess_rings_bo
|| add_sample_positions
) {
2457 size
= 112; /* 2 dword + 2 padding + 4 dword * 6 */
2458 if (add_sample_positions
)
2459 size
+= 128; /* 64+32+16+8 = 120 bytes */
2461 else if (scratch_bo
)
2462 size
= 8; /* 2 dword */
2464 descriptor_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
2468 RADEON_FLAG_CPU_ACCESS
|
2469 RADEON_FLAG_NO_INTERPROCESS_SHARING
|
2470 RADEON_FLAG_READ_ONLY
,
2471 RADV_BO_PRIORITY_DESCRIPTOR
);
2475 descriptor_bo
= queue
->descriptor_bo
;
2477 if (descriptor_bo
!= queue
->descriptor_bo
) {
2478 uint32_t *map
= (uint32_t*)queue
->device
->ws
->buffer_map(descriptor_bo
);
2481 uint64_t scratch_va
= radv_buffer_get_va(scratch_bo
);
2482 uint32_t rsrc1
= S_008F04_BASE_ADDRESS_HI(scratch_va
>> 32) |
2483 S_008F04_SWIZZLE_ENABLE(1);
2484 map
[0] = scratch_va
;
2488 if (esgs_ring_bo
|| gsvs_ring_bo
|| tess_rings_bo
|| add_sample_positions
)
2489 fill_geom_tess_rings(queue
, map
, add_sample_positions
,
2490 esgs_ring_size
, esgs_ring_bo
,
2491 gsvs_ring_size
, gsvs_ring_bo
,
2492 tess_factor_ring_size
,
2493 tess_offchip_ring_offset
,
2494 tess_offchip_ring_size
,
2497 queue
->device
->ws
->buffer_unmap(descriptor_bo
);
2500 for(int i
= 0; i
< 3; ++i
) {
2501 struct radeon_cmdbuf
*cs
= NULL
;
2502 cs
= queue
->device
->ws
->cs_create(queue
->device
->ws
,
2503 queue
->queue_family_index
? RING_COMPUTE
: RING_GFX
);
2510 radv_cs_add_buffer(queue
->device
->ws
, cs
, scratch_bo
);
2512 /* Emit initial configuration. */
2513 switch (queue
->queue_family_index
) {
2514 case RADV_QUEUE_GENERAL
:
2515 radv_init_graphics_state(cs
, queue
);
2517 case RADV_QUEUE_COMPUTE
:
2518 radv_init_compute_state(cs
, queue
);
2520 case RADV_QUEUE_TRANSFER
:
2524 if (esgs_ring_bo
|| gsvs_ring_bo
|| tess_rings_bo
) {
2525 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
2526 radeon_emit(cs
, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
2527 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
2528 radeon_emit(cs
, EVENT_TYPE(V_028A90_VGT_FLUSH
) | EVENT_INDEX(0));
2531 radv_emit_gs_ring_sizes(queue
, cs
, esgs_ring_bo
, esgs_ring_size
,
2532 gsvs_ring_bo
, gsvs_ring_size
);
2533 radv_emit_tess_factor_ring(queue
, cs
, hs_offchip_param
,
2534 tess_factor_ring_size
, tess_rings_bo
);
2535 radv_emit_global_shader_pointers(queue
, cs
, descriptor_bo
);
2536 radv_emit_compute_scratch(queue
, cs
, compute_scratch_bo
);
2539 si_cs_emit_cache_flush(cs
,
2540 queue
->device
->physical_device
->rad_info
.chip_class
,
2542 queue
->queue_family_index
== RING_COMPUTE
&&
2543 queue
->device
->physical_device
->rad_info
.chip_class
>= CIK
,
2544 (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
)) |
2545 RADV_CMD_FLAG_INV_ICACHE
|
2546 RADV_CMD_FLAG_INV_SMEM_L1
|
2547 RADV_CMD_FLAG_INV_VMEM_L1
|
2548 RADV_CMD_FLAG_INV_GLOBAL_L2
|
2549 RADV_CMD_FLAG_START_PIPELINE_STATS
, 0);
2550 } else if (i
== 1) {
2551 si_cs_emit_cache_flush(cs
,
2552 queue
->device
->physical_device
->rad_info
.chip_class
,
2554 queue
->queue_family_index
== RING_COMPUTE
&&
2555 queue
->device
->physical_device
->rad_info
.chip_class
>= CIK
,
2556 RADV_CMD_FLAG_INV_ICACHE
|
2557 RADV_CMD_FLAG_INV_SMEM_L1
|
2558 RADV_CMD_FLAG_INV_VMEM_L1
|
2559 RADV_CMD_FLAG_INV_GLOBAL_L2
|
2560 RADV_CMD_FLAG_START_PIPELINE_STATS
, 0);
2563 if (!queue
->device
->ws
->cs_finalize(cs
))
2567 if (queue
->initial_full_flush_preamble_cs
)
2568 queue
->device
->ws
->cs_destroy(queue
->initial_full_flush_preamble_cs
);
2570 if (queue
->initial_preamble_cs
)
2571 queue
->device
->ws
->cs_destroy(queue
->initial_preamble_cs
);
2573 if (queue
->continue_preamble_cs
)
2574 queue
->device
->ws
->cs_destroy(queue
->continue_preamble_cs
);
2576 queue
->initial_full_flush_preamble_cs
= dest_cs
[0];
2577 queue
->initial_preamble_cs
= dest_cs
[1];
2578 queue
->continue_preamble_cs
= dest_cs
[2];
2580 if (scratch_bo
!= queue
->scratch_bo
) {
2581 if (queue
->scratch_bo
)
2582 queue
->device
->ws
->buffer_destroy(queue
->scratch_bo
);
2583 queue
->scratch_bo
= scratch_bo
;
2584 queue
->scratch_size
= scratch_size
;
2587 if (compute_scratch_bo
!= queue
->compute_scratch_bo
) {
2588 if (queue
->compute_scratch_bo
)
2589 queue
->device
->ws
->buffer_destroy(queue
->compute_scratch_bo
);
2590 queue
->compute_scratch_bo
= compute_scratch_bo
;
2591 queue
->compute_scratch_size
= compute_scratch_size
;
2594 if (esgs_ring_bo
!= queue
->esgs_ring_bo
) {
2595 if (queue
->esgs_ring_bo
)
2596 queue
->device
->ws
->buffer_destroy(queue
->esgs_ring_bo
);
2597 queue
->esgs_ring_bo
= esgs_ring_bo
;
2598 queue
->esgs_ring_size
= esgs_ring_size
;
2601 if (gsvs_ring_bo
!= queue
->gsvs_ring_bo
) {
2602 if (queue
->gsvs_ring_bo
)
2603 queue
->device
->ws
->buffer_destroy(queue
->gsvs_ring_bo
);
2604 queue
->gsvs_ring_bo
= gsvs_ring_bo
;
2605 queue
->gsvs_ring_size
= gsvs_ring_size
;
2608 if (tess_rings_bo
!= queue
->tess_rings_bo
) {
2609 queue
->tess_rings_bo
= tess_rings_bo
;
2610 queue
->has_tess_rings
= true;
2613 if (descriptor_bo
!= queue
->descriptor_bo
) {
2614 if (queue
->descriptor_bo
)
2615 queue
->device
->ws
->buffer_destroy(queue
->descriptor_bo
);
2617 queue
->descriptor_bo
= descriptor_bo
;
2620 if (add_sample_positions
)
2621 queue
->has_sample_positions
= true;
2623 *initial_full_flush_preamble_cs
= queue
->initial_full_flush_preamble_cs
;
2624 *initial_preamble_cs
= queue
->initial_preamble_cs
;
2625 *continue_preamble_cs
= queue
->continue_preamble_cs
;
2626 if (!scratch_size
&& !compute_scratch_size
&& !esgs_ring_size
&& !gsvs_ring_size
)
2627 *continue_preamble_cs
= NULL
;
2630 for (int i
= 0; i
< ARRAY_SIZE(dest_cs
); ++i
)
2632 queue
->device
->ws
->cs_destroy(dest_cs
[i
]);
2633 if (descriptor_bo
&& descriptor_bo
!= queue
->descriptor_bo
)
2634 queue
->device
->ws
->buffer_destroy(descriptor_bo
);
2635 if (scratch_bo
&& scratch_bo
!= queue
->scratch_bo
)
2636 queue
->device
->ws
->buffer_destroy(scratch_bo
);
2637 if (compute_scratch_bo
&& compute_scratch_bo
!= queue
->compute_scratch_bo
)
2638 queue
->device
->ws
->buffer_destroy(compute_scratch_bo
);
2639 if (esgs_ring_bo
&& esgs_ring_bo
!= queue
->esgs_ring_bo
)
2640 queue
->device
->ws
->buffer_destroy(esgs_ring_bo
);
2641 if (gsvs_ring_bo
&& gsvs_ring_bo
!= queue
->gsvs_ring_bo
)
2642 queue
->device
->ws
->buffer_destroy(gsvs_ring_bo
);
2643 if (tess_rings_bo
&& tess_rings_bo
!= queue
->tess_rings_bo
)
2644 queue
->device
->ws
->buffer_destroy(tess_rings_bo
);
2645 return vk_error(queue
->device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
2648 static VkResult
radv_alloc_sem_counts(struct radv_instance
*instance
,
2649 struct radv_winsys_sem_counts
*counts
,
2651 const VkSemaphore
*sems
,
2655 int syncobj_idx
= 0, sem_idx
= 0;
2657 if (num_sems
== 0 && _fence
== VK_NULL_HANDLE
)
2660 for (uint32_t i
= 0; i
< num_sems
; i
++) {
2661 RADV_FROM_HANDLE(radv_semaphore
, sem
, sems
[i
]);
2663 if (sem
->temp_syncobj
|| sem
->syncobj
)
2664 counts
->syncobj_count
++;
2666 counts
->sem_count
++;
2669 if (_fence
!= VK_NULL_HANDLE
) {
2670 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
2671 if (fence
->temp_syncobj
|| fence
->syncobj
)
2672 counts
->syncobj_count
++;
2675 if (counts
->syncobj_count
) {
2676 counts
->syncobj
= (uint32_t *)malloc(sizeof(uint32_t) * counts
->syncobj_count
);
2677 if (!counts
->syncobj
)
2678 return vk_error(instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
2681 if (counts
->sem_count
) {
2682 counts
->sem
= (struct radeon_winsys_sem
**)malloc(sizeof(struct radeon_winsys_sem
*) * counts
->sem_count
);
2684 free(counts
->syncobj
);
2685 return vk_error(instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
2689 for (uint32_t i
= 0; i
< num_sems
; i
++) {
2690 RADV_FROM_HANDLE(radv_semaphore
, sem
, sems
[i
]);
2692 if (sem
->temp_syncobj
) {
2693 counts
->syncobj
[syncobj_idx
++] = sem
->temp_syncobj
;
2695 else if (sem
->syncobj
)
2696 counts
->syncobj
[syncobj_idx
++] = sem
->syncobj
;
2699 counts
->sem
[sem_idx
++] = sem
->sem
;
2703 if (_fence
!= VK_NULL_HANDLE
) {
2704 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
2705 if (fence
->temp_syncobj
)
2706 counts
->syncobj
[syncobj_idx
++] = fence
->temp_syncobj
;
2707 else if (fence
->syncobj
)
2708 counts
->syncobj
[syncobj_idx
++] = fence
->syncobj
;
2715 radv_free_sem_info(struct radv_winsys_sem_info
*sem_info
)
2717 free(sem_info
->wait
.syncobj
);
2718 free(sem_info
->wait
.sem
);
2719 free(sem_info
->signal
.syncobj
);
2720 free(sem_info
->signal
.sem
);
2724 static void radv_free_temp_syncobjs(struct radv_device
*device
,
2726 const VkSemaphore
*sems
)
2728 for (uint32_t i
= 0; i
< num_sems
; i
++) {
2729 RADV_FROM_HANDLE(radv_semaphore
, sem
, sems
[i
]);
2731 if (sem
->temp_syncobj
) {
2732 device
->ws
->destroy_syncobj(device
->ws
, sem
->temp_syncobj
);
2733 sem
->temp_syncobj
= 0;
2739 radv_alloc_sem_info(struct radv_instance
*instance
,
2740 struct radv_winsys_sem_info
*sem_info
,
2742 const VkSemaphore
*wait_sems
,
2743 int num_signal_sems
,
2744 const VkSemaphore
*signal_sems
,
2748 memset(sem_info
, 0, sizeof(*sem_info
));
2750 ret
= radv_alloc_sem_counts(instance
, &sem_info
->wait
, num_wait_sems
, wait_sems
, VK_NULL_HANDLE
, true);
2753 ret
= radv_alloc_sem_counts(instance
, &sem_info
->signal
, num_signal_sems
, signal_sems
, fence
, false);
2755 radv_free_sem_info(sem_info
);
2757 /* caller can override these */
2758 sem_info
->cs_emit_wait
= true;
2759 sem_info
->cs_emit_signal
= true;
2763 /* Signals fence as soon as all the work currently put on queue is done. */
2764 static VkResult
radv_signal_fence(struct radv_queue
*queue
,
2765 struct radv_fence
*fence
)
2769 struct radv_winsys_sem_info sem_info
;
2771 result
= radv_alloc_sem_info(queue
->device
->instance
, &sem_info
, 0, NULL
, 0, NULL
,
2772 radv_fence_to_handle(fence
));
2773 if (result
!= VK_SUCCESS
)
2776 ret
= queue
->device
->ws
->cs_submit(queue
->hw_ctx
, queue
->queue_idx
,
2777 &queue
->device
->empty_cs
[queue
->queue_family_index
],
2778 1, NULL
, NULL
, &sem_info
, NULL
,
2779 false, fence
->fence
);
2780 radv_free_sem_info(&sem_info
);
2783 return vk_error(queue
->device
->instance
, VK_ERROR_DEVICE_LOST
);
2788 VkResult
radv_QueueSubmit(
2790 uint32_t submitCount
,
2791 const VkSubmitInfo
* pSubmits
,
2794 RADV_FROM_HANDLE(radv_queue
, queue
, _queue
);
2795 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
2796 struct radeon_winsys_fence
*base_fence
= fence
? fence
->fence
: NULL
;
2797 struct radeon_winsys_ctx
*ctx
= queue
->hw_ctx
;
2799 uint32_t max_cs_submission
= queue
->device
->trace_bo
? 1 : UINT32_MAX
;
2800 uint32_t scratch_size
= 0;
2801 uint32_t compute_scratch_size
= 0;
2802 uint32_t esgs_ring_size
= 0, gsvs_ring_size
= 0;
2803 struct radeon_cmdbuf
*initial_preamble_cs
= NULL
, *initial_flush_preamble_cs
= NULL
, *continue_preamble_cs
= NULL
;
2805 bool fence_emitted
= false;
2806 bool tess_rings_needed
= false;
2807 bool sample_positions_needed
= false;
2809 /* Do this first so failing to allocate scratch buffers can't result in
2810 * partially executed submissions. */
2811 for (uint32_t i
= 0; i
< submitCount
; i
++) {
2812 for (uint32_t j
= 0; j
< pSubmits
[i
].commandBufferCount
; j
++) {
2813 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
,
2814 pSubmits
[i
].pCommandBuffers
[j
]);
2816 scratch_size
= MAX2(scratch_size
, cmd_buffer
->scratch_size_needed
);
2817 compute_scratch_size
= MAX2(compute_scratch_size
,
2818 cmd_buffer
->compute_scratch_size_needed
);
2819 esgs_ring_size
= MAX2(esgs_ring_size
, cmd_buffer
->esgs_ring_size_needed
);
2820 gsvs_ring_size
= MAX2(gsvs_ring_size
, cmd_buffer
->gsvs_ring_size_needed
);
2821 tess_rings_needed
|= cmd_buffer
->tess_rings_needed
;
2822 sample_positions_needed
|= cmd_buffer
->sample_positions_needed
;
2826 result
= radv_get_preamble_cs(queue
, scratch_size
, compute_scratch_size
,
2827 esgs_ring_size
, gsvs_ring_size
, tess_rings_needed
,
2828 sample_positions_needed
, &initial_flush_preamble_cs
,
2829 &initial_preamble_cs
, &continue_preamble_cs
);
2830 if (result
!= VK_SUCCESS
)
2833 for (uint32_t i
= 0; i
< submitCount
; i
++) {
2834 struct radeon_cmdbuf
**cs_array
;
2835 bool do_flush
= !i
|| pSubmits
[i
].pWaitDstStageMask
;
2836 bool can_patch
= true;
2838 struct radv_winsys_sem_info sem_info
;
2840 result
= radv_alloc_sem_info(queue
->device
->instance
,
2842 pSubmits
[i
].waitSemaphoreCount
,
2843 pSubmits
[i
].pWaitSemaphores
,
2844 pSubmits
[i
].signalSemaphoreCount
,
2845 pSubmits
[i
].pSignalSemaphores
,
2847 if (result
!= VK_SUCCESS
)
2850 if (!pSubmits
[i
].commandBufferCount
) {
2851 if (pSubmits
[i
].waitSemaphoreCount
|| pSubmits
[i
].signalSemaphoreCount
) {
2852 ret
= queue
->device
->ws
->cs_submit(ctx
, queue
->queue_idx
,
2853 &queue
->device
->empty_cs
[queue
->queue_family_index
],
2858 radv_loge("failed to submit CS %d\n", i
);
2861 fence_emitted
= true;
2863 radv_free_sem_info(&sem_info
);
2867 cs_array
= malloc(sizeof(struct radeon_cmdbuf
*) *
2868 (pSubmits
[i
].commandBufferCount
));
2870 for (uint32_t j
= 0; j
< pSubmits
[i
].commandBufferCount
; j
++) {
2871 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
,
2872 pSubmits
[i
].pCommandBuffers
[j
]);
2873 assert(cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_PRIMARY
);
2875 cs_array
[j
] = cmd_buffer
->cs
;
2876 if ((cmd_buffer
->usage_flags
& VK_COMMAND_BUFFER_USAGE_SIMULTANEOUS_USE_BIT
))
2879 cmd_buffer
->status
= RADV_CMD_BUFFER_STATUS_PENDING
;
2882 for (uint32_t j
= 0; j
< pSubmits
[i
].commandBufferCount
; j
+= advance
) {
2883 struct radeon_cmdbuf
*initial_preamble
= (do_flush
&& !j
) ? initial_flush_preamble_cs
: initial_preamble_cs
;
2884 const struct radv_winsys_bo_list
*bo_list
= NULL
;
2886 advance
= MIN2(max_cs_submission
,
2887 pSubmits
[i
].commandBufferCount
- j
);
2889 if (queue
->device
->trace_bo
)
2890 *queue
->device
->trace_id_ptr
= 0;
2892 sem_info
.cs_emit_wait
= j
== 0;
2893 sem_info
.cs_emit_signal
= j
+ advance
== pSubmits
[i
].commandBufferCount
;
2895 if (unlikely(queue
->device
->use_global_bo_list
)) {
2896 pthread_mutex_lock(&queue
->device
->bo_list
.mutex
);
2897 bo_list
= &queue
->device
->bo_list
.list
;
2900 ret
= queue
->device
->ws
->cs_submit(ctx
, queue
->queue_idx
, cs_array
+ j
,
2901 advance
, initial_preamble
, continue_preamble_cs
,
2903 can_patch
, base_fence
);
2905 if (unlikely(queue
->device
->use_global_bo_list
))
2906 pthread_mutex_unlock(&queue
->device
->bo_list
.mutex
);
2909 radv_loge("failed to submit CS %d\n", i
);
2912 fence_emitted
= true;
2913 if (queue
->device
->trace_bo
) {
2914 radv_check_gpu_hangs(queue
, cs_array
[j
]);
2918 radv_free_temp_syncobjs(queue
->device
,
2919 pSubmits
[i
].waitSemaphoreCount
,
2920 pSubmits
[i
].pWaitSemaphores
);
2921 radv_free_sem_info(&sem_info
);
2926 if (!fence_emitted
) {
2927 result
= radv_signal_fence(queue
, fence
);
2928 if (result
!= VK_SUCCESS
)
2931 fence
->submitted
= true;
2937 VkResult
radv_QueueWaitIdle(
2940 RADV_FROM_HANDLE(radv_queue
, queue
, _queue
);
2942 queue
->device
->ws
->ctx_wait_idle(queue
->hw_ctx
,
2943 radv_queue_family_to_ring(queue
->queue_family_index
),
2948 VkResult
radv_DeviceWaitIdle(
2951 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2953 for (unsigned i
= 0; i
< RADV_MAX_QUEUE_FAMILIES
; i
++) {
2954 for (unsigned q
= 0; q
< device
->queue_count
[i
]; q
++) {
2955 radv_QueueWaitIdle(radv_queue_to_handle(&device
->queues
[i
][q
]));
2961 VkResult
radv_EnumerateInstanceExtensionProperties(
2962 const char* pLayerName
,
2963 uint32_t* pPropertyCount
,
2964 VkExtensionProperties
* pProperties
)
2966 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
2968 for (int i
= 0; i
< RADV_INSTANCE_EXTENSION_COUNT
; i
++) {
2969 if (radv_supported_instance_extensions
.extensions
[i
]) {
2970 vk_outarray_append(&out
, prop
) {
2971 *prop
= radv_instance_extensions
[i
];
2976 return vk_outarray_status(&out
);
2979 VkResult
radv_EnumerateDeviceExtensionProperties(
2980 VkPhysicalDevice physicalDevice
,
2981 const char* pLayerName
,
2982 uint32_t* pPropertyCount
,
2983 VkExtensionProperties
* pProperties
)
2985 RADV_FROM_HANDLE(radv_physical_device
, device
, physicalDevice
);
2986 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
2988 for (int i
= 0; i
< RADV_DEVICE_EXTENSION_COUNT
; i
++) {
2989 if (device
->supported_extensions
.extensions
[i
]) {
2990 vk_outarray_append(&out
, prop
) {
2991 *prop
= radv_device_extensions
[i
];
2996 return vk_outarray_status(&out
);
2999 PFN_vkVoidFunction
radv_GetInstanceProcAddr(
3000 VkInstance _instance
,
3003 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
3005 return radv_lookup_entrypoint_checked(pName
,
3006 instance
? instance
->apiVersion
: 0,
3007 instance
? &instance
->enabled_extensions
: NULL
,
3011 /* The loader wants us to expose a second GetInstanceProcAddr function
3012 * to work around certain LD_PRELOAD issues seen in apps.
3015 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
3016 VkInstance instance
,
3020 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
3021 VkInstance instance
,
3024 return radv_GetInstanceProcAddr(instance
, pName
);
3027 PFN_vkVoidFunction
radv_GetDeviceProcAddr(
3031 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3033 return radv_lookup_entrypoint_checked(pName
,
3034 device
->instance
->apiVersion
,
3035 &device
->instance
->enabled_extensions
,
3036 &device
->enabled_extensions
);
3039 bool radv_get_memory_fd(struct radv_device
*device
,
3040 struct radv_device_memory
*memory
,
3043 struct radeon_bo_metadata metadata
;
3045 if (memory
->image
) {
3046 radv_init_metadata(device
, memory
->image
, &metadata
);
3047 device
->ws
->buffer_set_metadata(memory
->bo
, &metadata
);
3050 return device
->ws
->buffer_get_fd(device
->ws
, memory
->bo
,
3054 static VkResult
radv_alloc_memory(struct radv_device
*device
,
3055 const VkMemoryAllocateInfo
* pAllocateInfo
,
3056 const VkAllocationCallbacks
* pAllocator
,
3057 VkDeviceMemory
* pMem
)
3059 struct radv_device_memory
*mem
;
3061 enum radeon_bo_domain domain
;
3063 enum radv_mem_type mem_type_index
= device
->physical_device
->mem_type_indices
[pAllocateInfo
->memoryTypeIndex
];
3065 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
3067 if (pAllocateInfo
->allocationSize
== 0) {
3068 /* Apparently, this is allowed */
3069 *pMem
= VK_NULL_HANDLE
;
3073 const VkImportMemoryFdInfoKHR
*import_info
=
3074 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_FD_INFO_KHR
);
3075 const VkMemoryDedicatedAllocateInfo
*dedicate_info
=
3076 vk_find_struct_const(pAllocateInfo
->pNext
, MEMORY_DEDICATED_ALLOCATE_INFO
);
3077 const VkExportMemoryAllocateInfo
*export_info
=
3078 vk_find_struct_const(pAllocateInfo
->pNext
, EXPORT_MEMORY_ALLOCATE_INFO
);
3079 const VkImportMemoryHostPointerInfoEXT
*host_ptr_info
=
3080 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_HOST_POINTER_INFO_EXT
);
3082 const struct wsi_memory_allocate_info
*wsi_info
=
3083 vk_find_struct_const(pAllocateInfo
->pNext
, WSI_MEMORY_ALLOCATE_INFO_MESA
);
3085 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
3086 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3088 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3090 if (wsi_info
&& wsi_info
->implicit_sync
)
3091 flags
|= RADEON_FLAG_IMPLICIT_SYNC
;
3093 if (dedicate_info
) {
3094 mem
->image
= radv_image_from_handle(dedicate_info
->image
);
3095 mem
->buffer
= radv_buffer_from_handle(dedicate_info
->buffer
);
3101 float priority_float
= 0.5;
3102 const struct VkMemoryPriorityAllocateInfoEXT
*priority_ext
=
3103 vk_find_struct_const(pAllocateInfo
->pNext
,
3104 MEMORY_PRIORITY_ALLOCATE_INFO_EXT
);
3106 priority_float
= priority_ext
->priority
;
3108 unsigned priority
= MIN2(RADV_BO_PRIORITY_APPLICATION_MAX
- 1,
3109 (int)(priority_float
* RADV_BO_PRIORITY_APPLICATION_MAX
));
3111 mem
->user_ptr
= NULL
;
3114 assert(import_info
->handleType
==
3115 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
3116 import_info
->handleType
==
3117 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
3118 mem
->bo
= device
->ws
->buffer_from_fd(device
->ws
, import_info
->fd
,
3119 priority
, NULL
, NULL
);
3121 result
= VK_ERROR_INVALID_EXTERNAL_HANDLE
;
3124 close(import_info
->fd
);
3126 } else if (host_ptr_info
) {
3127 assert(host_ptr_info
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
);
3128 assert(mem_type_index
== RADV_MEM_TYPE_GTT_CACHED
);
3129 mem
->bo
= device
->ws
->buffer_from_ptr(device
->ws
, host_ptr_info
->pHostPointer
,
3130 pAllocateInfo
->allocationSize
,
3133 result
= VK_ERROR_INVALID_EXTERNAL_HANDLE
;
3136 mem
->user_ptr
= host_ptr_info
->pHostPointer
;
3139 uint64_t alloc_size
= align_u64(pAllocateInfo
->allocationSize
, 4096);
3140 if (mem_type_index
== RADV_MEM_TYPE_GTT_WRITE_COMBINE
||
3141 mem_type_index
== RADV_MEM_TYPE_GTT_CACHED
)
3142 domain
= RADEON_DOMAIN_GTT
;
3144 domain
= RADEON_DOMAIN_VRAM
;
3146 if (mem_type_index
== RADV_MEM_TYPE_VRAM
)
3147 flags
|= RADEON_FLAG_NO_CPU_ACCESS
;
3149 flags
|= RADEON_FLAG_CPU_ACCESS
;
3151 if (mem_type_index
== RADV_MEM_TYPE_GTT_WRITE_COMBINE
)
3152 flags
|= RADEON_FLAG_GTT_WC
;
3154 if (!dedicate_info
&& !import_info
&& (!export_info
|| !export_info
->handleTypes
))
3155 flags
|= RADEON_FLAG_NO_INTERPROCESS_SHARING
;
3157 mem
->bo
= device
->ws
->buffer_create(device
->ws
, alloc_size
, device
->physical_device
->rad_info
.max_alignment
,
3158 domain
, flags
, priority
);
3161 result
= VK_ERROR_OUT_OF_DEVICE_MEMORY
;
3164 mem
->type_index
= mem_type_index
;
3167 result
= radv_bo_list_add(device
, mem
->bo
);
3168 if (result
!= VK_SUCCESS
)
3171 *pMem
= radv_device_memory_to_handle(mem
);
3176 device
->ws
->buffer_destroy(mem
->bo
);
3178 vk_free2(&device
->alloc
, pAllocator
, mem
);
3183 VkResult
radv_AllocateMemory(
3185 const VkMemoryAllocateInfo
* pAllocateInfo
,
3186 const VkAllocationCallbacks
* pAllocator
,
3187 VkDeviceMemory
* pMem
)
3189 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3190 return radv_alloc_memory(device
, pAllocateInfo
, pAllocator
, pMem
);
3193 void radv_FreeMemory(
3195 VkDeviceMemory _mem
,
3196 const VkAllocationCallbacks
* pAllocator
)
3198 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3199 RADV_FROM_HANDLE(radv_device_memory
, mem
, _mem
);
3204 radv_bo_list_remove(device
, mem
->bo
);
3205 device
->ws
->buffer_destroy(mem
->bo
);
3208 vk_free2(&device
->alloc
, pAllocator
, mem
);
3211 VkResult
radv_MapMemory(
3213 VkDeviceMemory _memory
,
3214 VkDeviceSize offset
,
3216 VkMemoryMapFlags flags
,
3219 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3220 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
3228 *ppData
= mem
->user_ptr
;
3230 *ppData
= device
->ws
->buffer_map(mem
->bo
);
3237 return vk_error(device
->instance
, VK_ERROR_MEMORY_MAP_FAILED
);
3240 void radv_UnmapMemory(
3242 VkDeviceMemory _memory
)
3244 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3245 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
3250 if (mem
->user_ptr
== NULL
)
3251 device
->ws
->buffer_unmap(mem
->bo
);
3254 VkResult
radv_FlushMappedMemoryRanges(
3256 uint32_t memoryRangeCount
,
3257 const VkMappedMemoryRange
* pMemoryRanges
)
3262 VkResult
radv_InvalidateMappedMemoryRanges(
3264 uint32_t memoryRangeCount
,
3265 const VkMappedMemoryRange
* pMemoryRanges
)
3270 void radv_GetBufferMemoryRequirements(
3273 VkMemoryRequirements
* pMemoryRequirements
)
3275 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3276 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
3278 pMemoryRequirements
->memoryTypeBits
= (1u << device
->physical_device
->memory_properties
.memoryTypeCount
) - 1;
3280 if (buffer
->flags
& VK_BUFFER_CREATE_SPARSE_BINDING_BIT
)
3281 pMemoryRequirements
->alignment
= 4096;
3283 pMemoryRequirements
->alignment
= 16;
3285 pMemoryRequirements
->size
= align64(buffer
->size
, pMemoryRequirements
->alignment
);
3288 void radv_GetBufferMemoryRequirements2(
3290 const VkBufferMemoryRequirementsInfo2
*pInfo
,
3291 VkMemoryRequirements2
*pMemoryRequirements
)
3293 radv_GetBufferMemoryRequirements(device
, pInfo
->buffer
,
3294 &pMemoryRequirements
->memoryRequirements
);
3295 RADV_FROM_HANDLE(radv_buffer
, buffer
, pInfo
->buffer
);
3296 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3297 switch (ext
->sType
) {
3298 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
3299 VkMemoryDedicatedRequirements
*req
=
3300 (VkMemoryDedicatedRequirements
*) ext
;
3301 req
->requiresDedicatedAllocation
= buffer
->shareable
;
3302 req
->prefersDedicatedAllocation
= req
->requiresDedicatedAllocation
;
3311 void radv_GetImageMemoryRequirements(
3314 VkMemoryRequirements
* pMemoryRequirements
)
3316 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3317 RADV_FROM_HANDLE(radv_image
, image
, _image
);
3319 pMemoryRequirements
->memoryTypeBits
= (1u << device
->physical_device
->memory_properties
.memoryTypeCount
) - 1;
3321 pMemoryRequirements
->size
= image
->size
;
3322 pMemoryRequirements
->alignment
= image
->alignment
;
3325 void radv_GetImageMemoryRequirements2(
3327 const VkImageMemoryRequirementsInfo2
*pInfo
,
3328 VkMemoryRequirements2
*pMemoryRequirements
)
3330 radv_GetImageMemoryRequirements(device
, pInfo
->image
,
3331 &pMemoryRequirements
->memoryRequirements
);
3333 RADV_FROM_HANDLE(radv_image
, image
, pInfo
->image
);
3335 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3336 switch (ext
->sType
) {
3337 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS
: {
3338 VkMemoryDedicatedRequirements
*req
=
3339 (VkMemoryDedicatedRequirements
*) ext
;
3340 req
->requiresDedicatedAllocation
= image
->shareable
;
3341 req
->prefersDedicatedAllocation
= req
->requiresDedicatedAllocation
;
3350 void radv_GetImageSparseMemoryRequirements(
3353 uint32_t* pSparseMemoryRequirementCount
,
3354 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
3359 void radv_GetImageSparseMemoryRequirements2(
3361 const VkImageSparseMemoryRequirementsInfo2
*pInfo
,
3362 uint32_t* pSparseMemoryRequirementCount
,
3363 VkSparseImageMemoryRequirements2
*pSparseMemoryRequirements
)
3368 void radv_GetDeviceMemoryCommitment(
3370 VkDeviceMemory memory
,
3371 VkDeviceSize
* pCommittedMemoryInBytes
)
3373 *pCommittedMemoryInBytes
= 0;
3376 VkResult
radv_BindBufferMemory2(VkDevice device
,
3377 uint32_t bindInfoCount
,
3378 const VkBindBufferMemoryInfo
*pBindInfos
)
3380 for (uint32_t i
= 0; i
< bindInfoCount
; ++i
) {
3381 RADV_FROM_HANDLE(radv_device_memory
, mem
, pBindInfos
[i
].memory
);
3382 RADV_FROM_HANDLE(radv_buffer
, buffer
, pBindInfos
[i
].buffer
);
3385 buffer
->bo
= mem
->bo
;
3386 buffer
->offset
= pBindInfos
[i
].memoryOffset
;
3394 VkResult
radv_BindBufferMemory(
3397 VkDeviceMemory memory
,
3398 VkDeviceSize memoryOffset
)
3400 const VkBindBufferMemoryInfo info
= {
3401 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
,
3404 .memoryOffset
= memoryOffset
3407 return radv_BindBufferMemory2(device
, 1, &info
);
3410 VkResult
radv_BindImageMemory2(VkDevice device
,
3411 uint32_t bindInfoCount
,
3412 const VkBindImageMemoryInfo
*pBindInfos
)
3414 for (uint32_t i
= 0; i
< bindInfoCount
; ++i
) {
3415 RADV_FROM_HANDLE(radv_device_memory
, mem
, pBindInfos
[i
].memory
);
3416 RADV_FROM_HANDLE(radv_image
, image
, pBindInfos
[i
].image
);
3419 image
->bo
= mem
->bo
;
3420 image
->offset
= pBindInfos
[i
].memoryOffset
;
3430 VkResult
radv_BindImageMemory(
3433 VkDeviceMemory memory
,
3434 VkDeviceSize memoryOffset
)
3436 const VkBindImageMemoryInfo info
= {
3437 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO
,
3440 .memoryOffset
= memoryOffset
3443 return radv_BindImageMemory2(device
, 1, &info
);
3448 radv_sparse_buffer_bind_memory(struct radv_device
*device
,
3449 const VkSparseBufferMemoryBindInfo
*bind
)
3451 RADV_FROM_HANDLE(radv_buffer
, buffer
, bind
->buffer
);
3453 for (uint32_t i
= 0; i
< bind
->bindCount
; ++i
) {
3454 struct radv_device_memory
*mem
= NULL
;
3456 if (bind
->pBinds
[i
].memory
!= VK_NULL_HANDLE
)
3457 mem
= radv_device_memory_from_handle(bind
->pBinds
[i
].memory
);
3459 device
->ws
->buffer_virtual_bind(buffer
->bo
,
3460 bind
->pBinds
[i
].resourceOffset
,
3461 bind
->pBinds
[i
].size
,
3462 mem
? mem
->bo
: NULL
,
3463 bind
->pBinds
[i
].memoryOffset
);
3468 radv_sparse_image_opaque_bind_memory(struct radv_device
*device
,
3469 const VkSparseImageOpaqueMemoryBindInfo
*bind
)
3471 RADV_FROM_HANDLE(radv_image
, image
, bind
->image
);
3473 for (uint32_t i
= 0; i
< bind
->bindCount
; ++i
) {
3474 struct radv_device_memory
*mem
= NULL
;
3476 if (bind
->pBinds
[i
].memory
!= VK_NULL_HANDLE
)
3477 mem
= radv_device_memory_from_handle(bind
->pBinds
[i
].memory
);
3479 device
->ws
->buffer_virtual_bind(image
->bo
,
3480 bind
->pBinds
[i
].resourceOffset
,
3481 bind
->pBinds
[i
].size
,
3482 mem
? mem
->bo
: NULL
,
3483 bind
->pBinds
[i
].memoryOffset
);
3487 VkResult
radv_QueueBindSparse(
3489 uint32_t bindInfoCount
,
3490 const VkBindSparseInfo
* pBindInfo
,
3493 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
3494 RADV_FROM_HANDLE(radv_queue
, queue
, _queue
);
3495 struct radeon_winsys_fence
*base_fence
= fence
? fence
->fence
: NULL
;
3496 bool fence_emitted
= false;
3500 for (uint32_t i
= 0; i
< bindInfoCount
; ++i
) {
3501 struct radv_winsys_sem_info sem_info
;
3502 for (uint32_t j
= 0; j
< pBindInfo
[i
].bufferBindCount
; ++j
) {
3503 radv_sparse_buffer_bind_memory(queue
->device
,
3504 pBindInfo
[i
].pBufferBinds
+ j
);
3507 for (uint32_t j
= 0; j
< pBindInfo
[i
].imageOpaqueBindCount
; ++j
) {
3508 radv_sparse_image_opaque_bind_memory(queue
->device
,
3509 pBindInfo
[i
].pImageOpaqueBinds
+ j
);
3513 result
= radv_alloc_sem_info(queue
->device
->instance
,
3515 pBindInfo
[i
].waitSemaphoreCount
,
3516 pBindInfo
[i
].pWaitSemaphores
,
3517 pBindInfo
[i
].signalSemaphoreCount
,
3518 pBindInfo
[i
].pSignalSemaphores
,
3520 if (result
!= VK_SUCCESS
)
3523 if (pBindInfo
[i
].waitSemaphoreCount
|| pBindInfo
[i
].signalSemaphoreCount
) {
3524 ret
= queue
->device
->ws
->cs_submit(queue
->hw_ctx
, queue
->queue_idx
,
3525 &queue
->device
->empty_cs
[queue
->queue_family_index
],
3530 radv_loge("failed to submit CS %d\n", i
);
3534 fence_emitted
= true;
3536 fence
->submitted
= true;
3539 radv_free_sem_info(&sem_info
);
3544 if (!fence_emitted
) {
3545 result
= radv_signal_fence(queue
, fence
);
3546 if (result
!= VK_SUCCESS
)
3549 fence
->submitted
= true;
3555 VkResult
radv_CreateFence(
3557 const VkFenceCreateInfo
* pCreateInfo
,
3558 const VkAllocationCallbacks
* pAllocator
,
3561 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3562 const VkExportFenceCreateInfo
*export
=
3563 vk_find_struct_const(pCreateInfo
->pNext
, EXPORT_FENCE_CREATE_INFO
);
3564 VkExternalFenceHandleTypeFlags handleTypes
=
3565 export
? export
->handleTypes
: 0;
3567 struct radv_fence
*fence
= vk_alloc2(&device
->alloc
, pAllocator
,
3569 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3572 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3574 fence
->fence_wsi
= NULL
;
3575 fence
->submitted
= false;
3576 fence
->signalled
= !!(pCreateInfo
->flags
& VK_FENCE_CREATE_SIGNALED_BIT
);
3577 fence
->temp_syncobj
= 0;
3578 if (device
->always_use_syncobj
|| handleTypes
) {
3579 int ret
= device
->ws
->create_syncobj(device
->ws
, &fence
->syncobj
);
3581 vk_free2(&device
->alloc
, pAllocator
, fence
);
3582 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3584 if (pCreateInfo
->flags
& VK_FENCE_CREATE_SIGNALED_BIT
) {
3585 device
->ws
->signal_syncobj(device
->ws
, fence
->syncobj
);
3587 fence
->fence
= NULL
;
3589 fence
->fence
= device
->ws
->create_fence();
3590 if (!fence
->fence
) {
3591 vk_free2(&device
->alloc
, pAllocator
, fence
);
3592 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3597 *pFence
= radv_fence_to_handle(fence
);
3602 void radv_DestroyFence(
3605 const VkAllocationCallbacks
* pAllocator
)
3607 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3608 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
3613 if (fence
->temp_syncobj
)
3614 device
->ws
->destroy_syncobj(device
->ws
, fence
->temp_syncobj
);
3616 device
->ws
->destroy_syncobj(device
->ws
, fence
->syncobj
);
3618 device
->ws
->destroy_fence(fence
->fence
);
3619 if (fence
->fence_wsi
)
3620 fence
->fence_wsi
->destroy(fence
->fence_wsi
);
3621 vk_free2(&device
->alloc
, pAllocator
, fence
);
3625 static uint64_t radv_get_current_time()
3628 clock_gettime(CLOCK_MONOTONIC
, &tv
);
3629 return tv
.tv_nsec
+ tv
.tv_sec
*1000000000ull;
3632 static uint64_t radv_get_absolute_timeout(uint64_t timeout
)
3634 uint64_t current_time
= radv_get_current_time();
3636 timeout
= MIN2(UINT64_MAX
- current_time
, timeout
);
3638 return current_time
+ timeout
;
3642 static bool radv_all_fences_plain_and_submitted(uint32_t fenceCount
, const VkFence
*pFences
)
3644 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
3645 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
3646 if (fence
->fence
== NULL
|| fence
->syncobj
||
3647 fence
->temp_syncobj
||
3648 (!fence
->signalled
&& !fence
->submitted
))
3654 static bool radv_all_fences_syncobj(uint32_t fenceCount
, const VkFence
*pFences
)
3656 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
3657 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
3658 if (fence
->syncobj
== 0 && fence
->temp_syncobj
== 0)
3664 VkResult
radv_WaitForFences(
3666 uint32_t fenceCount
,
3667 const VkFence
* pFences
,
3671 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3672 timeout
= radv_get_absolute_timeout(timeout
);
3674 if (device
->always_use_syncobj
&&
3675 radv_all_fences_syncobj(fenceCount
, pFences
))
3677 uint32_t *handles
= malloc(sizeof(uint32_t) * fenceCount
);
3679 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3681 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
3682 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
3683 handles
[i
] = fence
->temp_syncobj
? fence
->temp_syncobj
: fence
->syncobj
;
3686 bool success
= device
->ws
->wait_syncobj(device
->ws
, handles
, fenceCount
, waitAll
, timeout
);
3689 return success
? VK_SUCCESS
: VK_TIMEOUT
;
3692 if (!waitAll
&& fenceCount
> 1) {
3693 /* Not doing this by default for waitAll, due to needing to allocate twice. */
3694 if (device
->physical_device
->rad_info
.drm_minor
>= 10 && radv_all_fences_plain_and_submitted(fenceCount
, pFences
)) {
3695 uint32_t wait_count
= 0;
3696 struct radeon_winsys_fence
**fences
= malloc(sizeof(struct radeon_winsys_fence
*) * fenceCount
);
3698 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3700 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
3701 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
3703 if (fence
->signalled
) {
3708 fences
[wait_count
++] = fence
->fence
;
3711 bool success
= device
->ws
->fences_wait(device
->ws
, fences
, wait_count
,
3712 waitAll
, timeout
- radv_get_current_time());
3715 return success
? VK_SUCCESS
: VK_TIMEOUT
;
3718 while(radv_get_current_time() <= timeout
) {
3719 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
3720 if (radv_GetFenceStatus(_device
, pFences
[i
]) == VK_SUCCESS
)
3727 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
3728 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
3729 bool expired
= false;
3731 if (fence
->temp_syncobj
) {
3732 if (!device
->ws
->wait_syncobj(device
->ws
, &fence
->temp_syncobj
, 1, true, timeout
))
3737 if (fence
->syncobj
) {
3738 if (!device
->ws
->wait_syncobj(device
->ws
, &fence
->syncobj
, 1, true, timeout
))
3743 if (fence
->signalled
)
3747 if (!fence
->submitted
) {
3748 while(radv_get_current_time() <= timeout
&&
3752 if (!fence
->submitted
)
3755 /* Recheck as it may have been set by
3756 * submitting operations. */
3758 if (fence
->signalled
)
3762 expired
= device
->ws
->fence_wait(device
->ws
,
3769 if (fence
->fence_wsi
) {
3770 VkResult result
= fence
->fence_wsi
->wait(fence
->fence_wsi
, timeout
);
3771 if (result
!= VK_SUCCESS
)
3775 fence
->signalled
= true;
3781 VkResult
radv_ResetFences(VkDevice _device
,
3782 uint32_t fenceCount
,
3783 const VkFence
*pFences
)
3785 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3787 for (unsigned i
= 0; i
< fenceCount
; ++i
) {
3788 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
3789 fence
->submitted
= fence
->signalled
= false;
3791 /* Per spec, we first restore the permanent payload, and then reset, so
3792 * having a temp syncobj should not skip resetting the permanent syncobj. */
3793 if (fence
->temp_syncobj
) {
3794 device
->ws
->destroy_syncobj(device
->ws
, fence
->temp_syncobj
);
3795 fence
->temp_syncobj
= 0;
3798 if (fence
->syncobj
) {
3799 device
->ws
->reset_syncobj(device
->ws
, fence
->syncobj
);
3806 VkResult
radv_GetFenceStatus(VkDevice _device
, VkFence _fence
)
3808 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3809 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
3811 if (fence
->temp_syncobj
) {
3812 bool success
= device
->ws
->wait_syncobj(device
->ws
, &fence
->temp_syncobj
, 1, true, 0);
3813 return success
? VK_SUCCESS
: VK_NOT_READY
;
3816 if (fence
->syncobj
) {
3817 bool success
= device
->ws
->wait_syncobj(device
->ws
, &fence
->syncobj
, 1, true, 0);
3818 return success
? VK_SUCCESS
: VK_NOT_READY
;
3821 if (fence
->signalled
)
3823 if (!fence
->submitted
)
3824 return VK_NOT_READY
;
3826 if (!device
->ws
->fence_wait(device
->ws
, fence
->fence
, false, 0))
3827 return VK_NOT_READY
;
3829 if (fence
->fence_wsi
) {
3830 VkResult result
= fence
->fence_wsi
->wait(fence
->fence_wsi
, 0);
3832 if (result
!= VK_SUCCESS
) {
3833 if (result
== VK_TIMEOUT
)
3834 return VK_NOT_READY
;
3842 // Queue semaphore functions
3844 VkResult
radv_CreateSemaphore(
3846 const VkSemaphoreCreateInfo
* pCreateInfo
,
3847 const VkAllocationCallbacks
* pAllocator
,
3848 VkSemaphore
* pSemaphore
)
3850 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3851 const VkExportSemaphoreCreateInfo
*export
=
3852 vk_find_struct_const(pCreateInfo
->pNext
, EXPORT_SEMAPHORE_CREATE_INFO
);
3853 VkExternalSemaphoreHandleTypeFlags handleTypes
=
3854 export
? export
->handleTypes
: 0;
3856 struct radv_semaphore
*sem
= vk_alloc2(&device
->alloc
, pAllocator
,
3858 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3860 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3862 sem
->temp_syncobj
= 0;
3863 /* create a syncobject if we are going to export this semaphore */
3864 if (device
->always_use_syncobj
|| handleTypes
) {
3865 assert (device
->physical_device
->rad_info
.has_syncobj
);
3866 int ret
= device
->ws
->create_syncobj(device
->ws
, &sem
->syncobj
);
3868 vk_free2(&device
->alloc
, pAllocator
, sem
);
3869 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3873 sem
->sem
= device
->ws
->create_sem(device
->ws
);
3875 vk_free2(&device
->alloc
, pAllocator
, sem
);
3876 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3881 *pSemaphore
= radv_semaphore_to_handle(sem
);
3885 void radv_DestroySemaphore(
3887 VkSemaphore _semaphore
,
3888 const VkAllocationCallbacks
* pAllocator
)
3890 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3891 RADV_FROM_HANDLE(radv_semaphore
, sem
, _semaphore
);
3896 device
->ws
->destroy_syncobj(device
->ws
, sem
->syncobj
);
3898 device
->ws
->destroy_sem(sem
->sem
);
3899 vk_free2(&device
->alloc
, pAllocator
, sem
);
3902 VkResult
radv_CreateEvent(
3904 const VkEventCreateInfo
* pCreateInfo
,
3905 const VkAllocationCallbacks
* pAllocator
,
3908 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3909 struct radv_event
*event
= vk_alloc2(&device
->alloc
, pAllocator
,
3911 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3914 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3916 event
->bo
= device
->ws
->buffer_create(device
->ws
, 8, 8,
3918 RADEON_FLAG_VA_UNCACHED
| RADEON_FLAG_CPU_ACCESS
| RADEON_FLAG_NO_INTERPROCESS_SHARING
,
3919 RADV_BO_PRIORITY_FENCE
);
3921 vk_free2(&device
->alloc
, pAllocator
, event
);
3922 return vk_error(device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
3925 event
->map
= (uint64_t*)device
->ws
->buffer_map(event
->bo
);
3927 *pEvent
= radv_event_to_handle(event
);
3932 void radv_DestroyEvent(
3935 const VkAllocationCallbacks
* pAllocator
)
3937 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3938 RADV_FROM_HANDLE(radv_event
, event
, _event
);
3942 device
->ws
->buffer_destroy(event
->bo
);
3943 vk_free2(&device
->alloc
, pAllocator
, event
);
3946 VkResult
radv_GetEventStatus(
3950 RADV_FROM_HANDLE(radv_event
, event
, _event
);
3952 if (*event
->map
== 1)
3953 return VK_EVENT_SET
;
3954 return VK_EVENT_RESET
;
3957 VkResult
radv_SetEvent(
3961 RADV_FROM_HANDLE(radv_event
, event
, _event
);
3967 VkResult
radv_ResetEvent(
3971 RADV_FROM_HANDLE(radv_event
, event
, _event
);
3977 VkResult
radv_CreateBuffer(
3979 const VkBufferCreateInfo
* pCreateInfo
,
3980 const VkAllocationCallbacks
* pAllocator
,
3983 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3984 struct radv_buffer
*buffer
;
3986 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
3988 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
3989 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3991 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3993 buffer
->size
= pCreateInfo
->size
;
3994 buffer
->usage
= pCreateInfo
->usage
;
3997 buffer
->flags
= pCreateInfo
->flags
;
3999 buffer
->shareable
= vk_find_struct_const(pCreateInfo
->pNext
,
4000 EXTERNAL_MEMORY_BUFFER_CREATE_INFO
) != NULL
;
4002 if (pCreateInfo
->flags
& VK_BUFFER_CREATE_SPARSE_BINDING_BIT
) {
4003 buffer
->bo
= device
->ws
->buffer_create(device
->ws
,
4004 align64(buffer
->size
, 4096),
4005 4096, 0, RADEON_FLAG_VIRTUAL
,
4006 RADV_BO_PRIORITY_VIRTUAL
);
4008 vk_free2(&device
->alloc
, pAllocator
, buffer
);
4009 return vk_error(device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
4013 *pBuffer
= radv_buffer_to_handle(buffer
);
4018 void radv_DestroyBuffer(
4021 const VkAllocationCallbacks
* pAllocator
)
4023 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4024 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
4029 if (buffer
->flags
& VK_BUFFER_CREATE_SPARSE_BINDING_BIT
)
4030 device
->ws
->buffer_destroy(buffer
->bo
);
4032 vk_free2(&device
->alloc
, pAllocator
, buffer
);
4035 static inline unsigned
4036 si_tile_mode_index(const struct radv_image
*image
, unsigned level
, bool stencil
)
4039 return image
->surface
.u
.legacy
.stencil_tiling_index
[level
];
4041 return image
->surface
.u
.legacy
.tiling_index
[level
];
4044 static uint32_t radv_surface_max_layer_count(struct radv_image_view
*iview
)
4046 return iview
->type
== VK_IMAGE_VIEW_TYPE_3D
? iview
->extent
.depth
: (iview
->base_layer
+ iview
->layer_count
);
4050 radv_init_dcc_control_reg(struct radv_device
*device
,
4051 struct radv_image_view
*iview
)
4053 unsigned max_uncompressed_block_size
= V_028C78_MAX_BLOCK_SIZE_256B
;
4054 unsigned min_compressed_block_size
= V_028C78_MIN_BLOCK_SIZE_32B
;
4055 unsigned max_compressed_block_size
;
4056 unsigned independent_64b_blocks
;
4058 if (!radv_image_has_dcc(iview
->image
))
4061 if (iview
->image
->info
.samples
> 1) {
4062 if (iview
->image
->surface
.bpe
== 1)
4063 max_uncompressed_block_size
= V_028C78_MAX_BLOCK_SIZE_64B
;
4064 else if (iview
->image
->surface
.bpe
== 2)
4065 max_uncompressed_block_size
= V_028C78_MAX_BLOCK_SIZE_128B
;
4068 if (!device
->physical_device
->rad_info
.has_dedicated_vram
) {
4069 /* amdvlk: [min-compressed-block-size] should be set to 32 for
4070 * dGPU and 64 for APU because all of our APUs to date use
4071 * DIMMs which have a request granularity size of 64B while all
4072 * other chips have a 32B request size.
4074 min_compressed_block_size
= V_028C78_MIN_BLOCK_SIZE_64B
;
4077 if (iview
->image
->usage
& (VK_IMAGE_USAGE_SAMPLED_BIT
|
4078 VK_IMAGE_USAGE_TRANSFER_SRC_BIT
|
4079 VK_IMAGE_USAGE_INPUT_ATTACHMENT_BIT
)) {
4080 /* If this DCC image is potentially going to be used in texture
4081 * fetches, we need some special settings.
4083 independent_64b_blocks
= 1;
4084 max_compressed_block_size
= V_028C78_MAX_BLOCK_SIZE_64B
;
4086 /* MAX_UNCOMPRESSED_BLOCK_SIZE must be >=
4087 * MAX_COMPRESSED_BLOCK_SIZE. Set MAX_COMPRESSED_BLOCK_SIZE as
4088 * big as possible for better compression state.
4090 independent_64b_blocks
= 0;
4091 max_compressed_block_size
= max_uncompressed_block_size
;
4094 return S_028C78_MAX_UNCOMPRESSED_BLOCK_SIZE(max_uncompressed_block_size
) |
4095 S_028C78_MAX_COMPRESSED_BLOCK_SIZE(max_compressed_block_size
) |
4096 S_028C78_MIN_COMPRESSED_BLOCK_SIZE(min_compressed_block_size
) |
4097 S_028C78_INDEPENDENT_64B_BLOCKS(independent_64b_blocks
);
4101 radv_initialise_color_surface(struct radv_device
*device
,
4102 struct radv_color_buffer_info
*cb
,
4103 struct radv_image_view
*iview
)
4105 const struct vk_format_description
*desc
;
4106 unsigned ntype
, format
, swap
, endian
;
4107 unsigned blend_clamp
= 0, blend_bypass
= 0;
4109 const struct radeon_surf
*surf
= &iview
->image
->surface
;
4111 desc
= vk_format_description(iview
->vk_format
);
4113 memset(cb
, 0, sizeof(*cb
));
4115 /* Intensity is implemented as Red, so treat it that way. */
4116 cb
->cb_color_attrib
= S_028C74_FORCE_DST_ALPHA_1(desc
->swizzle
[3] == VK_SWIZZLE_1
);
4118 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
;
4120 cb
->cb_color_base
= va
>> 8;
4122 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
4123 struct gfx9_surf_meta_flags meta
;
4124 if (iview
->image
->dcc_offset
)
4125 meta
= iview
->image
->surface
.u
.gfx9
.dcc
;
4127 meta
= iview
->image
->surface
.u
.gfx9
.cmask
;
4129 cb
->cb_color_attrib
|= S_028C74_COLOR_SW_MODE(iview
->image
->surface
.u
.gfx9
.surf
.swizzle_mode
) |
4130 S_028C74_FMASK_SW_MODE(iview
->image
->surface
.u
.gfx9
.fmask
.swizzle_mode
) |
4131 S_028C74_RB_ALIGNED(meta
.rb_aligned
) |
4132 S_028C74_PIPE_ALIGNED(meta
.pipe_aligned
);
4134 cb
->cb_color_base
+= iview
->image
->surface
.u
.gfx9
.surf_offset
>> 8;
4135 cb
->cb_color_base
|= iview
->image
->surface
.tile_swizzle
;
4137 const struct legacy_surf_level
*level_info
= &surf
->u
.legacy
.level
[iview
->base_mip
];
4138 unsigned pitch_tile_max
, slice_tile_max
, tile_mode_index
;
4140 cb
->cb_color_base
+= level_info
->offset
>> 8;
4141 if (level_info
->mode
== RADEON_SURF_MODE_2D
)
4142 cb
->cb_color_base
|= iview
->image
->surface
.tile_swizzle
;
4144 pitch_tile_max
= level_info
->nblk_x
/ 8 - 1;
4145 slice_tile_max
= (level_info
->nblk_x
* level_info
->nblk_y
) / 64 - 1;
4146 tile_mode_index
= si_tile_mode_index(iview
->image
, iview
->base_mip
, false);
4148 cb
->cb_color_pitch
= S_028C64_TILE_MAX(pitch_tile_max
);
4149 cb
->cb_color_slice
= S_028C68_TILE_MAX(slice_tile_max
);
4150 cb
->cb_color_cmask_slice
= iview
->image
->cmask
.slice_tile_max
;
4152 cb
->cb_color_attrib
|= S_028C74_TILE_MODE_INDEX(tile_mode_index
);
4154 if (radv_image_has_fmask(iview
->image
)) {
4155 if (device
->physical_device
->rad_info
.chip_class
>= CIK
)
4156 cb
->cb_color_pitch
|= S_028C64_FMASK_TILE_MAX(iview
->image
->fmask
.pitch_in_pixels
/ 8 - 1);
4157 cb
->cb_color_attrib
|= S_028C74_FMASK_TILE_MODE_INDEX(iview
->image
->fmask
.tile_mode_index
);
4158 cb
->cb_color_fmask_slice
= S_028C88_TILE_MAX(iview
->image
->fmask
.slice_tile_max
);
4160 /* This must be set for fast clear to work without FMASK. */
4161 if (device
->physical_device
->rad_info
.chip_class
>= CIK
)
4162 cb
->cb_color_pitch
|= S_028C64_FMASK_TILE_MAX(pitch_tile_max
);
4163 cb
->cb_color_attrib
|= S_028C74_FMASK_TILE_MODE_INDEX(tile_mode_index
);
4164 cb
->cb_color_fmask_slice
= S_028C88_TILE_MAX(slice_tile_max
);
4168 /* CMASK variables */
4169 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
;
4170 va
+= iview
->image
->cmask
.offset
;
4171 cb
->cb_color_cmask
= va
>> 8;
4173 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
;
4174 va
+= iview
->image
->dcc_offset
;
4175 cb
->cb_dcc_base
= va
>> 8;
4176 cb
->cb_dcc_base
|= iview
->image
->surface
.tile_swizzle
;
4178 uint32_t max_slice
= radv_surface_max_layer_count(iview
) - 1;
4179 cb
->cb_color_view
= S_028C6C_SLICE_START(iview
->base_layer
) |
4180 S_028C6C_SLICE_MAX(max_slice
);
4182 if (iview
->image
->info
.samples
> 1) {
4183 unsigned log_samples
= util_logbase2(iview
->image
->info
.samples
);
4185 cb
->cb_color_attrib
|= S_028C74_NUM_SAMPLES(log_samples
) |
4186 S_028C74_NUM_FRAGMENTS(log_samples
);
4189 if (radv_image_has_fmask(iview
->image
)) {
4190 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
+ iview
->image
->fmask
.offset
;
4191 cb
->cb_color_fmask
= va
>> 8;
4192 cb
->cb_color_fmask
|= iview
->image
->fmask
.tile_swizzle
;
4194 cb
->cb_color_fmask
= cb
->cb_color_base
;
4197 ntype
= radv_translate_color_numformat(iview
->vk_format
,
4199 vk_format_get_first_non_void_channel(iview
->vk_format
));
4200 format
= radv_translate_colorformat(iview
->vk_format
);
4201 if (format
== V_028C70_COLOR_INVALID
|| ntype
== ~0u)
4202 radv_finishme("Illegal color\n");
4203 swap
= radv_translate_colorswap(iview
->vk_format
, FALSE
);
4204 endian
= radv_colorformat_endian_swap(format
);
4206 /* blend clamp should be set for all NORM/SRGB types */
4207 if (ntype
== V_028C70_NUMBER_UNORM
||
4208 ntype
== V_028C70_NUMBER_SNORM
||
4209 ntype
== V_028C70_NUMBER_SRGB
)
4212 /* set blend bypass according to docs if SINT/UINT or
4213 8/24 COLOR variants */
4214 if (ntype
== V_028C70_NUMBER_UINT
|| ntype
== V_028C70_NUMBER_SINT
||
4215 format
== V_028C70_COLOR_8_24
|| format
== V_028C70_COLOR_24_8
||
4216 format
== V_028C70_COLOR_X24_8_32_FLOAT
) {
4221 if ((ntype
== V_028C70_NUMBER_UINT
|| ntype
== V_028C70_NUMBER_SINT
) &&
4222 (format
== V_028C70_COLOR_8
||
4223 format
== V_028C70_COLOR_8_8
||
4224 format
== V_028C70_COLOR_8_8_8_8
))
4225 ->color_is_int8
= true;
4227 cb
->cb_color_info
= S_028C70_FORMAT(format
) |
4228 S_028C70_COMP_SWAP(swap
) |
4229 S_028C70_BLEND_CLAMP(blend_clamp
) |
4230 S_028C70_BLEND_BYPASS(blend_bypass
) |
4231 S_028C70_SIMPLE_FLOAT(1) |
4232 S_028C70_ROUND_MODE(ntype
!= V_028C70_NUMBER_UNORM
&&
4233 ntype
!= V_028C70_NUMBER_SNORM
&&
4234 ntype
!= V_028C70_NUMBER_SRGB
&&
4235 format
!= V_028C70_COLOR_8_24
&&
4236 format
!= V_028C70_COLOR_24_8
) |
4237 S_028C70_NUMBER_TYPE(ntype
) |
4238 S_028C70_ENDIAN(endian
);
4239 if (radv_image_has_fmask(iview
->image
)) {
4240 cb
->cb_color_info
|= S_028C70_COMPRESSION(1);
4241 if (device
->physical_device
->rad_info
.chip_class
== SI
) {
4242 unsigned fmask_bankh
= util_logbase2(iview
->image
->fmask
.bank_height
);
4243 cb
->cb_color_attrib
|= S_028C74_FMASK_BANK_HEIGHT(fmask_bankh
);
4247 if (radv_image_has_cmask(iview
->image
) &&
4248 !(device
->instance
->debug_flags
& RADV_DEBUG_NO_FAST_CLEARS
))
4249 cb
->cb_color_info
|= S_028C70_FAST_CLEAR(1);
4251 if (radv_dcc_enabled(iview
->image
, iview
->base_mip
))
4252 cb
->cb_color_info
|= S_028C70_DCC_ENABLE(1);
4254 cb
->cb_dcc_control
= radv_init_dcc_control_reg(device
, iview
);
4256 /* This must be set for fast clear to work without FMASK. */
4257 if (!radv_image_has_fmask(iview
->image
) &&
4258 device
->physical_device
->rad_info
.chip_class
== SI
) {
4259 unsigned bankh
= util_logbase2(iview
->image
->surface
.u
.legacy
.bankh
);
4260 cb
->cb_color_attrib
|= S_028C74_FMASK_BANK_HEIGHT(bankh
);
4263 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
4264 unsigned mip0_depth
= iview
->image
->type
== VK_IMAGE_TYPE_3D
?
4265 (iview
->extent
.depth
- 1) : (iview
->image
->info
.array_size
- 1);
4267 cb
->cb_color_view
|= S_028C6C_MIP_LEVEL(iview
->base_mip
);
4268 cb
->cb_color_attrib
|= S_028C74_MIP0_DEPTH(mip0_depth
) |
4269 S_028C74_RESOURCE_TYPE(iview
->image
->surface
.u
.gfx9
.resource_type
);
4270 cb
->cb_color_attrib2
= S_028C68_MIP0_WIDTH(iview
->extent
.width
- 1) |
4271 S_028C68_MIP0_HEIGHT(iview
->extent
.height
- 1) |
4272 S_028C68_MAX_MIP(iview
->image
->info
.levels
- 1);
4277 radv_calc_decompress_on_z_planes(struct radv_device
*device
,
4278 struct radv_image_view
*iview
)
4280 unsigned max_zplanes
= 0;
4282 assert(radv_image_is_tc_compat_htile(iview
->image
));
4284 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
4285 /* Default value for 32-bit depth surfaces. */
4288 if (iview
->vk_format
== VK_FORMAT_D16_UNORM
&&
4289 iview
->image
->info
.samples
> 1)
4292 max_zplanes
= max_zplanes
+ 1;
4294 if (iview
->vk_format
== VK_FORMAT_D16_UNORM
) {
4295 /* Do not enable Z plane compression for 16-bit depth
4296 * surfaces because isn't supported on GFX8. Only
4297 * 32-bit depth surfaces are supported by the hardware.
4298 * This allows to maintain shader compatibility and to
4299 * reduce the number of depth decompressions.
4303 if (iview
->image
->info
.samples
<= 1)
4305 else if (iview
->image
->info
.samples
<= 4)
4316 radv_initialise_ds_surface(struct radv_device
*device
,
4317 struct radv_ds_buffer_info
*ds
,
4318 struct radv_image_view
*iview
)
4320 unsigned level
= iview
->base_mip
;
4321 unsigned format
, stencil_format
;
4322 uint64_t va
, s_offs
, z_offs
;
4323 bool stencil_only
= false;
4324 memset(ds
, 0, sizeof(*ds
));
4325 switch (iview
->image
->vk_format
) {
4326 case VK_FORMAT_D24_UNORM_S8_UINT
:
4327 case VK_FORMAT_X8_D24_UNORM_PACK32
:
4328 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-24);
4329 ds
->offset_scale
= 2.0f
;
4331 case VK_FORMAT_D16_UNORM
:
4332 case VK_FORMAT_D16_UNORM_S8_UINT
:
4333 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-16);
4334 ds
->offset_scale
= 4.0f
;
4336 case VK_FORMAT_D32_SFLOAT
:
4337 case VK_FORMAT_D32_SFLOAT_S8_UINT
:
4338 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-23) |
4339 S_028B78_POLY_OFFSET_DB_IS_FLOAT_FMT(1);
4340 ds
->offset_scale
= 1.0f
;
4342 case VK_FORMAT_S8_UINT
:
4343 stencil_only
= true;
4349 format
= radv_translate_dbformat(iview
->image
->vk_format
);
4350 stencil_format
= iview
->image
->surface
.has_stencil
?
4351 V_028044_STENCIL_8
: V_028044_STENCIL_INVALID
;
4353 uint32_t max_slice
= radv_surface_max_layer_count(iview
) - 1;
4354 ds
->db_depth_view
= S_028008_SLICE_START(iview
->base_layer
) |
4355 S_028008_SLICE_MAX(max_slice
);
4357 ds
->db_htile_data_base
= 0;
4358 ds
->db_htile_surface
= 0;
4360 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
;
4361 s_offs
= z_offs
= va
;
4363 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
4364 assert(iview
->image
->surface
.u
.gfx9
.surf_offset
== 0);
4365 s_offs
+= iview
->image
->surface
.u
.gfx9
.stencil_offset
;
4367 ds
->db_z_info
= S_028038_FORMAT(format
) |
4368 S_028038_NUM_SAMPLES(util_logbase2(iview
->image
->info
.samples
)) |
4369 S_028038_SW_MODE(iview
->image
->surface
.u
.gfx9
.surf
.swizzle_mode
) |
4370 S_028038_MAXMIP(iview
->image
->info
.levels
- 1) |
4371 S_028038_ZRANGE_PRECISION(1);
4372 ds
->db_stencil_info
= S_02803C_FORMAT(stencil_format
) |
4373 S_02803C_SW_MODE(iview
->image
->surface
.u
.gfx9
.stencil
.swizzle_mode
);
4375 ds
->db_z_info2
= S_028068_EPITCH(iview
->image
->surface
.u
.gfx9
.surf
.epitch
);
4376 ds
->db_stencil_info2
= S_02806C_EPITCH(iview
->image
->surface
.u
.gfx9
.stencil
.epitch
);
4377 ds
->db_depth_view
|= S_028008_MIPID(level
);
4379 ds
->db_depth_size
= S_02801C_X_MAX(iview
->image
->info
.width
- 1) |
4380 S_02801C_Y_MAX(iview
->image
->info
.height
- 1);
4382 if (radv_htile_enabled(iview
->image
, level
)) {
4383 ds
->db_z_info
|= S_028038_TILE_SURFACE_ENABLE(1);
4385 if (radv_image_is_tc_compat_htile(iview
->image
)) {
4386 unsigned max_zplanes
=
4387 radv_calc_decompress_on_z_planes(device
, iview
);
4389 ds
->db_z_info
|= S_028038_DECOMPRESS_ON_N_ZPLANES(max_zplanes
) |
4390 S_028038_ITERATE_FLUSH(1);
4391 ds
->db_stencil_info
|= S_02803C_ITERATE_FLUSH(1);
4394 if (!iview
->image
->surface
.has_stencil
)
4395 /* Use all of the htile_buffer for depth if there's no stencil. */
4396 ds
->db_stencil_info
|= S_02803C_TILE_STENCIL_DISABLE(1);
4397 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
+
4398 iview
->image
->htile_offset
;
4399 ds
->db_htile_data_base
= va
>> 8;
4400 ds
->db_htile_surface
= S_028ABC_FULL_CACHE(1) |
4401 S_028ABC_PIPE_ALIGNED(iview
->image
->surface
.u
.gfx9
.htile
.pipe_aligned
) |
4402 S_028ABC_RB_ALIGNED(iview
->image
->surface
.u
.gfx9
.htile
.rb_aligned
);
4405 const struct legacy_surf_level
*level_info
= &iview
->image
->surface
.u
.legacy
.level
[level
];
4408 level_info
= &iview
->image
->surface
.u
.legacy
.stencil_level
[level
];
4410 z_offs
+= iview
->image
->surface
.u
.legacy
.level
[level
].offset
;
4411 s_offs
+= iview
->image
->surface
.u
.legacy
.stencil_level
[level
].offset
;
4413 ds
->db_depth_info
= S_02803C_ADDR5_SWIZZLE_MASK(!radv_image_is_tc_compat_htile(iview
->image
));
4414 ds
->db_z_info
= S_028040_FORMAT(format
) | S_028040_ZRANGE_PRECISION(1);
4415 ds
->db_stencil_info
= S_028044_FORMAT(stencil_format
);
4417 if (iview
->image
->info
.samples
> 1)
4418 ds
->db_z_info
|= S_028040_NUM_SAMPLES(util_logbase2(iview
->image
->info
.samples
));
4420 if (device
->physical_device
->rad_info
.chip_class
>= CIK
) {
4421 struct radeon_info
*info
= &device
->physical_device
->rad_info
;
4422 unsigned tiling_index
= iview
->image
->surface
.u
.legacy
.tiling_index
[level
];
4423 unsigned stencil_index
= iview
->image
->surface
.u
.legacy
.stencil_tiling_index
[level
];
4424 unsigned macro_index
= iview
->image
->surface
.u
.legacy
.macro_tile_index
;
4425 unsigned tile_mode
= info
->si_tile_mode_array
[tiling_index
];
4426 unsigned stencil_tile_mode
= info
->si_tile_mode_array
[stencil_index
];
4427 unsigned macro_mode
= info
->cik_macrotile_mode_array
[macro_index
];
4430 tile_mode
= stencil_tile_mode
;
4432 ds
->db_depth_info
|=
4433 S_02803C_ARRAY_MODE(G_009910_ARRAY_MODE(tile_mode
)) |
4434 S_02803C_PIPE_CONFIG(G_009910_PIPE_CONFIG(tile_mode
)) |
4435 S_02803C_BANK_WIDTH(G_009990_BANK_WIDTH(macro_mode
)) |
4436 S_02803C_BANK_HEIGHT(G_009990_BANK_HEIGHT(macro_mode
)) |
4437 S_02803C_MACRO_TILE_ASPECT(G_009990_MACRO_TILE_ASPECT(macro_mode
)) |
4438 S_02803C_NUM_BANKS(G_009990_NUM_BANKS(macro_mode
));
4439 ds
->db_z_info
|= S_028040_TILE_SPLIT(G_009910_TILE_SPLIT(tile_mode
));
4440 ds
->db_stencil_info
|= S_028044_TILE_SPLIT(G_009910_TILE_SPLIT(stencil_tile_mode
));
4442 unsigned tile_mode_index
= si_tile_mode_index(iview
->image
, level
, false);
4443 ds
->db_z_info
|= S_028040_TILE_MODE_INDEX(tile_mode_index
);
4444 tile_mode_index
= si_tile_mode_index(iview
->image
, level
, true);
4445 ds
->db_stencil_info
|= S_028044_TILE_MODE_INDEX(tile_mode_index
);
4447 ds
->db_z_info
|= S_028040_TILE_MODE_INDEX(tile_mode_index
);
4450 ds
->db_depth_size
= S_028058_PITCH_TILE_MAX((level_info
->nblk_x
/ 8) - 1) |
4451 S_028058_HEIGHT_TILE_MAX((level_info
->nblk_y
/ 8) - 1);
4452 ds
->db_depth_slice
= S_02805C_SLICE_TILE_MAX((level_info
->nblk_x
* level_info
->nblk_y
) / 64 - 1);
4454 if (radv_htile_enabled(iview
->image
, level
)) {
4455 ds
->db_z_info
|= S_028040_TILE_SURFACE_ENABLE(1);
4457 if (!iview
->image
->surface
.has_stencil
&&
4458 !radv_image_is_tc_compat_htile(iview
->image
))
4459 /* Use all of the htile_buffer for depth if there's no stencil. */
4460 ds
->db_stencil_info
|= S_028044_TILE_STENCIL_DISABLE(1);
4462 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
+
4463 iview
->image
->htile_offset
;
4464 ds
->db_htile_data_base
= va
>> 8;
4465 ds
->db_htile_surface
= S_028ABC_FULL_CACHE(1);
4467 if (radv_image_is_tc_compat_htile(iview
->image
)) {
4468 unsigned max_zplanes
=
4469 radv_calc_decompress_on_z_planes(device
, iview
);
4471 ds
->db_htile_surface
|= S_028ABC_TC_COMPATIBLE(1);
4472 ds
->db_z_info
|= S_028040_DECOMPRESS_ON_N_ZPLANES(max_zplanes
);
4477 ds
->db_z_read_base
= ds
->db_z_write_base
= z_offs
>> 8;
4478 ds
->db_stencil_read_base
= ds
->db_stencil_write_base
= s_offs
>> 8;
4481 VkResult
radv_CreateFramebuffer(
4483 const VkFramebufferCreateInfo
* pCreateInfo
,
4484 const VkAllocationCallbacks
* pAllocator
,
4485 VkFramebuffer
* pFramebuffer
)
4487 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4488 struct radv_framebuffer
*framebuffer
;
4490 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
4492 size_t size
= sizeof(*framebuffer
) +
4493 sizeof(struct radv_attachment_info
) * pCreateInfo
->attachmentCount
;
4494 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
4495 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
4496 if (framebuffer
== NULL
)
4497 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
4499 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
4500 framebuffer
->width
= pCreateInfo
->width
;
4501 framebuffer
->height
= pCreateInfo
->height
;
4502 framebuffer
->layers
= pCreateInfo
->layers
;
4503 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
4504 VkImageView _iview
= pCreateInfo
->pAttachments
[i
];
4505 struct radv_image_view
*iview
= radv_image_view_from_handle(_iview
);
4506 framebuffer
->attachments
[i
].attachment
= iview
;
4507 if (iview
->aspect_mask
& VK_IMAGE_ASPECT_COLOR_BIT
) {
4508 radv_initialise_color_surface(device
, &framebuffer
->attachments
[i
].cb
, iview
);
4509 } else if (iview
->aspect_mask
& (VK_IMAGE_ASPECT_DEPTH_BIT
| VK_IMAGE_ASPECT_STENCIL_BIT
)) {
4510 radv_initialise_ds_surface(device
, &framebuffer
->attachments
[i
].ds
, iview
);
4512 framebuffer
->width
= MIN2(framebuffer
->width
, iview
->extent
.width
);
4513 framebuffer
->height
= MIN2(framebuffer
->height
, iview
->extent
.height
);
4514 framebuffer
->layers
= MIN2(framebuffer
->layers
, radv_surface_max_layer_count(iview
));
4517 *pFramebuffer
= radv_framebuffer_to_handle(framebuffer
);
4521 void radv_DestroyFramebuffer(
4524 const VkAllocationCallbacks
* pAllocator
)
4526 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4527 RADV_FROM_HANDLE(radv_framebuffer
, fb
, _fb
);
4531 vk_free2(&device
->alloc
, pAllocator
, fb
);
4534 static unsigned radv_tex_wrap(VkSamplerAddressMode address_mode
)
4536 switch (address_mode
) {
4537 case VK_SAMPLER_ADDRESS_MODE_REPEAT
:
4538 return V_008F30_SQ_TEX_WRAP
;
4539 case VK_SAMPLER_ADDRESS_MODE_MIRRORED_REPEAT
:
4540 return V_008F30_SQ_TEX_MIRROR
;
4541 case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE
:
4542 return V_008F30_SQ_TEX_CLAMP_LAST_TEXEL
;
4543 case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_BORDER
:
4544 return V_008F30_SQ_TEX_CLAMP_BORDER
;
4545 case VK_SAMPLER_ADDRESS_MODE_MIRROR_CLAMP_TO_EDGE
:
4546 return V_008F30_SQ_TEX_MIRROR_ONCE_LAST_TEXEL
;
4548 unreachable("illegal tex wrap mode");
4554 radv_tex_compare(VkCompareOp op
)
4557 case VK_COMPARE_OP_NEVER
:
4558 return V_008F30_SQ_TEX_DEPTH_COMPARE_NEVER
;
4559 case VK_COMPARE_OP_LESS
:
4560 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESS
;
4561 case VK_COMPARE_OP_EQUAL
:
4562 return V_008F30_SQ_TEX_DEPTH_COMPARE_EQUAL
;
4563 case VK_COMPARE_OP_LESS_OR_EQUAL
:
4564 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESSEQUAL
;
4565 case VK_COMPARE_OP_GREATER
:
4566 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATER
;
4567 case VK_COMPARE_OP_NOT_EQUAL
:
4568 return V_008F30_SQ_TEX_DEPTH_COMPARE_NOTEQUAL
;
4569 case VK_COMPARE_OP_GREATER_OR_EQUAL
:
4570 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATEREQUAL
;
4571 case VK_COMPARE_OP_ALWAYS
:
4572 return V_008F30_SQ_TEX_DEPTH_COMPARE_ALWAYS
;
4574 unreachable("illegal compare mode");
4580 radv_tex_filter(VkFilter filter
, unsigned max_ansio
)
4583 case VK_FILTER_NEAREST
:
4584 return (max_ansio
> 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_POINT
:
4585 V_008F38_SQ_TEX_XY_FILTER_POINT
);
4586 case VK_FILTER_LINEAR
:
4587 return (max_ansio
> 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_BILINEAR
:
4588 V_008F38_SQ_TEX_XY_FILTER_BILINEAR
);
4589 case VK_FILTER_CUBIC_IMG
:
4591 fprintf(stderr
, "illegal texture filter");
4597 radv_tex_mipfilter(VkSamplerMipmapMode mode
)
4600 case VK_SAMPLER_MIPMAP_MODE_NEAREST
:
4601 return V_008F38_SQ_TEX_Z_FILTER_POINT
;
4602 case VK_SAMPLER_MIPMAP_MODE_LINEAR
:
4603 return V_008F38_SQ_TEX_Z_FILTER_LINEAR
;
4605 return V_008F38_SQ_TEX_Z_FILTER_NONE
;
4610 radv_tex_bordercolor(VkBorderColor bcolor
)
4613 case VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
:
4614 case VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
:
4615 return V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK
;
4616 case VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
:
4617 case VK_BORDER_COLOR_INT_OPAQUE_BLACK
:
4618 return V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_BLACK
;
4619 case VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
:
4620 case VK_BORDER_COLOR_INT_OPAQUE_WHITE
:
4621 return V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_WHITE
;
4629 radv_tex_aniso_filter(unsigned filter
)
4643 radv_tex_filter_mode(VkSamplerReductionModeEXT mode
)
4646 case VK_SAMPLER_REDUCTION_MODE_WEIGHTED_AVERAGE_EXT
:
4647 return V_008F30_SQ_IMG_FILTER_MODE_BLEND
;
4648 case VK_SAMPLER_REDUCTION_MODE_MIN_EXT
:
4649 return V_008F30_SQ_IMG_FILTER_MODE_MIN
;
4650 case VK_SAMPLER_REDUCTION_MODE_MAX_EXT
:
4651 return V_008F30_SQ_IMG_FILTER_MODE_MAX
;
4659 radv_get_max_anisotropy(struct radv_device
*device
,
4660 const VkSamplerCreateInfo
*pCreateInfo
)
4662 if (device
->force_aniso
>= 0)
4663 return device
->force_aniso
;
4665 if (pCreateInfo
->anisotropyEnable
&&
4666 pCreateInfo
->maxAnisotropy
> 1.0f
)
4667 return (uint32_t)pCreateInfo
->maxAnisotropy
;
4673 radv_init_sampler(struct radv_device
*device
,
4674 struct radv_sampler
*sampler
,
4675 const VkSamplerCreateInfo
*pCreateInfo
)
4677 uint32_t max_aniso
= radv_get_max_anisotropy(device
, pCreateInfo
);
4678 uint32_t max_aniso_ratio
= radv_tex_aniso_filter(max_aniso
);
4679 bool is_vi
= (device
->physical_device
->rad_info
.chip_class
>= VI
);
4680 unsigned filter_mode
= V_008F30_SQ_IMG_FILTER_MODE_BLEND
;
4682 const struct VkSamplerReductionModeCreateInfoEXT
*sampler_reduction
=
4683 vk_find_struct_const(pCreateInfo
->pNext
,
4684 SAMPLER_REDUCTION_MODE_CREATE_INFO_EXT
);
4685 if (sampler_reduction
)
4686 filter_mode
= radv_tex_filter_mode(sampler_reduction
->reductionMode
);
4688 sampler
->state
[0] = (S_008F30_CLAMP_X(radv_tex_wrap(pCreateInfo
->addressModeU
)) |
4689 S_008F30_CLAMP_Y(radv_tex_wrap(pCreateInfo
->addressModeV
)) |
4690 S_008F30_CLAMP_Z(radv_tex_wrap(pCreateInfo
->addressModeW
)) |
4691 S_008F30_MAX_ANISO_RATIO(max_aniso_ratio
) |
4692 S_008F30_DEPTH_COMPARE_FUNC(radv_tex_compare(pCreateInfo
->compareOp
)) |
4693 S_008F30_FORCE_UNNORMALIZED(pCreateInfo
->unnormalizedCoordinates
? 1 : 0) |
4694 S_008F30_ANISO_THRESHOLD(max_aniso_ratio
>> 1) |
4695 S_008F30_ANISO_BIAS(max_aniso_ratio
) |
4696 S_008F30_DISABLE_CUBE_WRAP(0) |
4697 S_008F30_COMPAT_MODE(is_vi
) |
4698 S_008F30_FILTER_MODE(filter_mode
));
4699 sampler
->state
[1] = (S_008F34_MIN_LOD(S_FIXED(CLAMP(pCreateInfo
->minLod
, 0, 15), 8)) |
4700 S_008F34_MAX_LOD(S_FIXED(CLAMP(pCreateInfo
->maxLod
, 0, 15), 8)) |
4701 S_008F34_PERF_MIP(max_aniso_ratio
? max_aniso_ratio
+ 6 : 0));
4702 sampler
->state
[2] = (S_008F38_LOD_BIAS(S_FIXED(CLAMP(pCreateInfo
->mipLodBias
, -16, 16), 8)) |
4703 S_008F38_XY_MAG_FILTER(radv_tex_filter(pCreateInfo
->magFilter
, max_aniso
)) |
4704 S_008F38_XY_MIN_FILTER(radv_tex_filter(pCreateInfo
->minFilter
, max_aniso
)) |
4705 S_008F38_MIP_FILTER(radv_tex_mipfilter(pCreateInfo
->mipmapMode
)) |
4706 S_008F38_MIP_POINT_PRECLAMP(0) |
4707 S_008F38_DISABLE_LSB_CEIL(device
->physical_device
->rad_info
.chip_class
<= VI
) |
4708 S_008F38_FILTER_PREC_FIX(1) |
4709 S_008F38_ANISO_OVERRIDE(is_vi
));
4710 sampler
->state
[3] = (S_008F3C_BORDER_COLOR_PTR(0) |
4711 S_008F3C_BORDER_COLOR_TYPE(radv_tex_bordercolor(pCreateInfo
->borderColor
)));
4714 VkResult
radv_CreateSampler(
4716 const VkSamplerCreateInfo
* pCreateInfo
,
4717 const VkAllocationCallbacks
* pAllocator
,
4718 VkSampler
* pSampler
)
4720 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4721 struct radv_sampler
*sampler
;
4723 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO
);
4725 sampler
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*sampler
), 8,
4726 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
4728 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
4730 radv_init_sampler(device
, sampler
, pCreateInfo
);
4731 *pSampler
= radv_sampler_to_handle(sampler
);
4736 void radv_DestroySampler(
4739 const VkAllocationCallbacks
* pAllocator
)
4741 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4742 RADV_FROM_HANDLE(radv_sampler
, sampler
, _sampler
);
4746 vk_free2(&device
->alloc
, pAllocator
, sampler
);
4749 /* vk_icd.h does not declare this function, so we declare it here to
4750 * suppress Wmissing-prototypes.
4752 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4753 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t *pSupportedVersion
);
4755 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4756 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t *pSupportedVersion
)
4758 /* For the full details on loader interface versioning, see
4759 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
4760 * What follows is a condensed summary, to help you navigate the large and
4761 * confusing official doc.
4763 * - Loader interface v0 is incompatible with later versions. We don't
4766 * - In loader interface v1:
4767 * - The first ICD entrypoint called by the loader is
4768 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
4770 * - The ICD must statically expose no other Vulkan symbol unless it is
4771 * linked with -Bsymbolic.
4772 * - Each dispatchable Vulkan handle created by the ICD must be
4773 * a pointer to a struct whose first member is VK_LOADER_DATA. The
4774 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
4775 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
4776 * vkDestroySurfaceKHR(). The ICD must be capable of working with
4777 * such loader-managed surfaces.
4779 * - Loader interface v2 differs from v1 in:
4780 * - The first ICD entrypoint called by the loader is
4781 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
4782 * statically expose this entrypoint.
4784 * - Loader interface v3 differs from v2 in:
4785 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
4786 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
4787 * because the loader no longer does so.
4789 *pSupportedVersion
= MIN2(*pSupportedVersion
, 3u);
4793 VkResult
radv_GetMemoryFdKHR(VkDevice _device
,
4794 const VkMemoryGetFdInfoKHR
*pGetFdInfo
,
4797 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4798 RADV_FROM_HANDLE(radv_device_memory
, memory
, pGetFdInfo
->memory
);
4800 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
4802 /* At the moment, we support only the below handle types. */
4803 assert(pGetFdInfo
->handleType
==
4804 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT
||
4805 pGetFdInfo
->handleType
==
4806 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
4808 bool ret
= radv_get_memory_fd(device
, memory
, pFD
);
4810 return vk_error(device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
4814 VkResult
radv_GetMemoryFdPropertiesKHR(VkDevice _device
,
4815 VkExternalMemoryHandleTypeFlagBits handleType
,
4817 VkMemoryFdPropertiesKHR
*pMemoryFdProperties
)
4819 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4821 switch (handleType
) {
4822 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
:
4823 pMemoryFdProperties
->memoryTypeBits
= (1 << RADV_MEM_TYPE_COUNT
) - 1;
4827 /* The valid usage section for this function says:
4829 * "handleType must not be one of the handle types defined as
4832 * So opaque handle types fall into the default "unsupported" case.
4834 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE
);
4838 static VkResult
radv_import_opaque_fd(struct radv_device
*device
,
4842 uint32_t syncobj_handle
= 0;
4843 int ret
= device
->ws
->import_syncobj(device
->ws
, fd
, &syncobj_handle
);
4845 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE
);
4848 device
->ws
->destroy_syncobj(device
->ws
, *syncobj
);
4850 *syncobj
= syncobj_handle
;
4856 static VkResult
radv_import_sync_fd(struct radv_device
*device
,
4860 /* If we create a syncobj we do it locally so that if we have an error, we don't
4861 * leave a syncobj in an undetermined state in the fence. */
4862 uint32_t syncobj_handle
= *syncobj
;
4863 if (!syncobj_handle
) {
4864 int ret
= device
->ws
->create_syncobj(device
->ws
, &syncobj_handle
);
4866 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE
);
4871 device
->ws
->signal_syncobj(device
->ws
, syncobj_handle
);
4873 int ret
= device
->ws
->import_syncobj_from_sync_file(device
->ws
, syncobj_handle
, fd
);
4875 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE
);
4878 *syncobj
= syncobj_handle
;
4885 VkResult
radv_ImportSemaphoreFdKHR(VkDevice _device
,
4886 const VkImportSemaphoreFdInfoKHR
*pImportSemaphoreFdInfo
)
4888 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4889 RADV_FROM_HANDLE(radv_semaphore
, sem
, pImportSemaphoreFdInfo
->semaphore
);
4890 uint32_t *syncobj_dst
= NULL
;
4892 if (pImportSemaphoreFdInfo
->flags
& VK_SEMAPHORE_IMPORT_TEMPORARY_BIT
) {
4893 syncobj_dst
= &sem
->temp_syncobj
;
4895 syncobj_dst
= &sem
->syncobj
;
4898 switch(pImportSemaphoreFdInfo
->handleType
) {
4899 case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT
:
4900 return radv_import_opaque_fd(device
, pImportSemaphoreFdInfo
->fd
, syncobj_dst
);
4901 case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT
:
4902 return radv_import_sync_fd(device
, pImportSemaphoreFdInfo
->fd
, syncobj_dst
);
4904 unreachable("Unhandled semaphore handle type");
4908 VkResult
radv_GetSemaphoreFdKHR(VkDevice _device
,
4909 const VkSemaphoreGetFdInfoKHR
*pGetFdInfo
,
4912 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4913 RADV_FROM_HANDLE(radv_semaphore
, sem
, pGetFdInfo
->semaphore
);
4915 uint32_t syncobj_handle
;
4917 if (sem
->temp_syncobj
)
4918 syncobj_handle
= sem
->temp_syncobj
;
4920 syncobj_handle
= sem
->syncobj
;
4922 switch(pGetFdInfo
->handleType
) {
4923 case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT
:
4924 ret
= device
->ws
->export_syncobj(device
->ws
, syncobj_handle
, pFd
);
4926 case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT
:
4927 ret
= device
->ws
->export_syncobj_to_sync_file(device
->ws
, syncobj_handle
, pFd
);
4929 if (sem
->temp_syncobj
) {
4930 close (sem
->temp_syncobj
);
4931 sem
->temp_syncobj
= 0;
4933 device
->ws
->reset_syncobj(device
->ws
, syncobj_handle
);
4938 unreachable("Unhandled semaphore handle type");
4942 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE
);
4946 void radv_GetPhysicalDeviceExternalSemaphoreProperties(
4947 VkPhysicalDevice physicalDevice
,
4948 const VkPhysicalDeviceExternalSemaphoreInfo
*pExternalSemaphoreInfo
,
4949 VkExternalSemaphoreProperties
*pExternalSemaphoreProperties
)
4951 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
4953 /* Require has_syncobj_wait_for_submit for the syncobj signal ioctl introduced at virtually the same time */
4954 if (pdevice
->rad_info
.has_syncobj_wait_for_submit
&&
4955 (pExternalSemaphoreInfo
->handleType
== VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT
||
4956 pExternalSemaphoreInfo
->handleType
== VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT
)) {
4957 pExternalSemaphoreProperties
->exportFromImportedHandleTypes
= VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT
| VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT
;
4958 pExternalSemaphoreProperties
->compatibleHandleTypes
= VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT
| VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT
;
4959 pExternalSemaphoreProperties
->externalSemaphoreFeatures
= VK_EXTERNAL_SEMAPHORE_FEATURE_EXPORTABLE_BIT
|
4960 VK_EXTERNAL_SEMAPHORE_FEATURE_IMPORTABLE_BIT
;
4961 } else if (pExternalSemaphoreInfo
->handleType
== VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT
) {
4962 pExternalSemaphoreProperties
->exportFromImportedHandleTypes
= VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT
;
4963 pExternalSemaphoreProperties
->compatibleHandleTypes
= VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT
;
4964 pExternalSemaphoreProperties
->externalSemaphoreFeatures
= VK_EXTERNAL_SEMAPHORE_FEATURE_EXPORTABLE_BIT
|
4965 VK_EXTERNAL_SEMAPHORE_FEATURE_IMPORTABLE_BIT
;
4967 pExternalSemaphoreProperties
->exportFromImportedHandleTypes
= 0;
4968 pExternalSemaphoreProperties
->compatibleHandleTypes
= 0;
4969 pExternalSemaphoreProperties
->externalSemaphoreFeatures
= 0;
4973 VkResult
radv_ImportFenceFdKHR(VkDevice _device
,
4974 const VkImportFenceFdInfoKHR
*pImportFenceFdInfo
)
4976 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4977 RADV_FROM_HANDLE(radv_fence
, fence
, pImportFenceFdInfo
->fence
);
4978 uint32_t *syncobj_dst
= NULL
;
4981 if (pImportFenceFdInfo
->flags
& VK_FENCE_IMPORT_TEMPORARY_BIT
) {
4982 syncobj_dst
= &fence
->temp_syncobj
;
4984 syncobj_dst
= &fence
->syncobj
;
4987 switch(pImportFenceFdInfo
->handleType
) {
4988 case VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT
:
4989 return radv_import_opaque_fd(device
, pImportFenceFdInfo
->fd
, syncobj_dst
);
4990 case VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT
:
4991 return radv_import_sync_fd(device
, pImportFenceFdInfo
->fd
, syncobj_dst
);
4993 unreachable("Unhandled fence handle type");
4997 VkResult
radv_GetFenceFdKHR(VkDevice _device
,
4998 const VkFenceGetFdInfoKHR
*pGetFdInfo
,
5001 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5002 RADV_FROM_HANDLE(radv_fence
, fence
, pGetFdInfo
->fence
);
5004 uint32_t syncobj_handle
;
5006 if (fence
->temp_syncobj
)
5007 syncobj_handle
= fence
->temp_syncobj
;
5009 syncobj_handle
= fence
->syncobj
;
5011 switch(pGetFdInfo
->handleType
) {
5012 case VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT
:
5013 ret
= device
->ws
->export_syncobj(device
->ws
, syncobj_handle
, pFd
);
5015 case VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT
:
5016 ret
= device
->ws
->export_syncobj_to_sync_file(device
->ws
, syncobj_handle
, pFd
);
5018 if (fence
->temp_syncobj
) {
5019 close (fence
->temp_syncobj
);
5020 fence
->temp_syncobj
= 0;
5022 device
->ws
->reset_syncobj(device
->ws
, syncobj_handle
);
5027 unreachable("Unhandled fence handle type");
5031 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE
);
5035 void radv_GetPhysicalDeviceExternalFenceProperties(
5036 VkPhysicalDevice physicalDevice
,
5037 const VkPhysicalDeviceExternalFenceInfo
*pExternalFenceInfo
,
5038 VkExternalFenceProperties
*pExternalFenceProperties
)
5040 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
5042 if (pdevice
->rad_info
.has_syncobj_wait_for_submit
&&
5043 (pExternalFenceInfo
->handleType
== VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT
||
5044 pExternalFenceInfo
->handleType
== VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT
)) {
5045 pExternalFenceProperties
->exportFromImportedHandleTypes
= VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT
| VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT
;
5046 pExternalFenceProperties
->compatibleHandleTypes
= VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT
| VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT
;
5047 pExternalFenceProperties
->externalFenceFeatures
= VK_EXTERNAL_FENCE_FEATURE_EXPORTABLE_BIT
|
5048 VK_EXTERNAL_SEMAPHORE_FEATURE_IMPORTABLE_BIT
;
5050 pExternalFenceProperties
->exportFromImportedHandleTypes
= 0;
5051 pExternalFenceProperties
->compatibleHandleTypes
= 0;
5052 pExternalFenceProperties
->externalFenceFeatures
= 0;
5057 radv_CreateDebugReportCallbackEXT(VkInstance _instance
,
5058 const VkDebugReportCallbackCreateInfoEXT
* pCreateInfo
,
5059 const VkAllocationCallbacks
* pAllocator
,
5060 VkDebugReportCallbackEXT
* pCallback
)
5062 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
5063 return vk_create_debug_report_callback(&instance
->debug_report_callbacks
,
5064 pCreateInfo
, pAllocator
, &instance
->alloc
,
5069 radv_DestroyDebugReportCallbackEXT(VkInstance _instance
,
5070 VkDebugReportCallbackEXT _callback
,
5071 const VkAllocationCallbacks
* pAllocator
)
5073 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
5074 vk_destroy_debug_report_callback(&instance
->debug_report_callbacks
,
5075 _callback
, pAllocator
, &instance
->alloc
);
5079 radv_DebugReportMessageEXT(VkInstance _instance
,
5080 VkDebugReportFlagsEXT flags
,
5081 VkDebugReportObjectTypeEXT objectType
,
5084 int32_t messageCode
,
5085 const char* pLayerPrefix
,
5086 const char* pMessage
)
5088 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
5089 vk_debug_report(&instance
->debug_report_callbacks
, flags
, objectType
,
5090 object
, location
, messageCode
, pLayerPrefix
, pMessage
);
5094 radv_GetDeviceGroupPeerMemoryFeatures(
5097 uint32_t localDeviceIndex
,
5098 uint32_t remoteDeviceIndex
,
5099 VkPeerMemoryFeatureFlags
* pPeerMemoryFeatures
)
5101 assert(localDeviceIndex
== remoteDeviceIndex
);
5103 *pPeerMemoryFeatures
= VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT
|
5104 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT
|
5105 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT
|
5106 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT
;
5109 static const VkTimeDomainEXT radv_time_domains
[] = {
5110 VK_TIME_DOMAIN_DEVICE_EXT
,
5111 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
,
5112 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
,
5115 VkResult
radv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
5116 VkPhysicalDevice physicalDevice
,
5117 uint32_t *pTimeDomainCount
,
5118 VkTimeDomainEXT
*pTimeDomains
)
5121 VK_OUTARRAY_MAKE(out
, pTimeDomains
, pTimeDomainCount
);
5123 for (d
= 0; d
< ARRAY_SIZE(radv_time_domains
); d
++) {
5124 vk_outarray_append(&out
, i
) {
5125 *i
= radv_time_domains
[d
];
5129 return vk_outarray_status(&out
);
5133 radv_clock_gettime(clockid_t clock_id
)
5135 struct timespec current
;
5138 ret
= clock_gettime(clock_id
, ¤t
);
5139 if (ret
< 0 && clock_id
== CLOCK_MONOTONIC_RAW
)
5140 ret
= clock_gettime(CLOCK_MONOTONIC
, ¤t
);
5144 return (uint64_t) current
.tv_sec
* 1000000000ULL + current
.tv_nsec
;
5147 VkResult
radv_GetCalibratedTimestampsEXT(
5149 uint32_t timestampCount
,
5150 const VkCalibratedTimestampInfoEXT
*pTimestampInfos
,
5151 uint64_t *pTimestamps
,
5152 uint64_t *pMaxDeviation
)
5154 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5155 uint32_t clock_crystal_freq
= device
->physical_device
->rad_info
.clock_crystal_freq
;
5157 uint64_t begin
, end
;
5158 uint64_t max_clock_period
= 0;
5160 begin
= radv_clock_gettime(CLOCK_MONOTONIC_RAW
);
5162 for (d
= 0; d
< timestampCount
; d
++) {
5163 switch (pTimestampInfos
[d
].timeDomain
) {
5164 case VK_TIME_DOMAIN_DEVICE_EXT
:
5165 pTimestamps
[d
] = device
->ws
->query_value(device
->ws
,
5167 uint64_t device_period
= DIV_ROUND_UP(1000000, clock_crystal_freq
);
5168 max_clock_period
= MAX2(max_clock_period
, device_period
);
5170 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
:
5171 pTimestamps
[d
] = radv_clock_gettime(CLOCK_MONOTONIC
);
5172 max_clock_period
= MAX2(max_clock_period
, 1);
5175 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
:
5176 pTimestamps
[d
] = begin
;
5184 end
= radv_clock_gettime(CLOCK_MONOTONIC_RAW
);
5187 * The maximum deviation is the sum of the interval over which we
5188 * perform the sampling and the maximum period of any sampled
5189 * clock. That's because the maximum skew between any two sampled
5190 * clock edges is when the sampled clock with the largest period is
5191 * sampled at the end of that period but right at the beginning of the
5192 * sampling interval and some other clock is sampled right at the
5193 * begining of its sampling period and right at the end of the
5194 * sampling interval. Let's assume the GPU has the longest clock
5195 * period and that the application is sampling GPU and monotonic:
5198 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
5199 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
5203 * GPU -----_____-----_____-----_____-----_____
5206 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
5207 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
5209 * Interval <----------------->
5210 * Deviation <-------------------------->
5214 * m = read(monotonic) 2
5217 * We round the sample interval up by one tick to cover sampling error
5218 * in the interval clock
5221 uint64_t sample_interval
= end
- begin
+ 1;
5223 *pMaxDeviation
= sample_interval
+ max_clock_period
;