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_physical_device_init_mem_types(struct radv_physical_device
*device
)
128 STATIC_ASSERT(RADV_MEM_HEAP_COUNT
<= VK_MAX_MEMORY_HEAPS
);
129 uint64_t visible_vram_size
= MIN2(device
->rad_info
.vram_size
,
130 device
->rad_info
.vram_vis_size
);
132 int vram_index
= -1, visible_vram_index
= -1, gart_index
= -1;
133 device
->memory_properties
.memoryHeapCount
= 0;
134 if (device
->rad_info
.vram_size
- visible_vram_size
> 0) {
135 vram_index
= device
->memory_properties
.memoryHeapCount
++;
136 device
->memory_properties
.memoryHeaps
[vram_index
] = (VkMemoryHeap
) {
137 .size
= device
->rad_info
.vram_size
- visible_vram_size
,
138 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
141 if (visible_vram_size
) {
142 visible_vram_index
= device
->memory_properties
.memoryHeapCount
++;
143 device
->memory_properties
.memoryHeaps
[visible_vram_index
] = (VkMemoryHeap
) {
144 .size
= visible_vram_size
,
145 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
148 if (device
->rad_info
.gart_size
> 0) {
149 gart_index
= device
->memory_properties
.memoryHeapCount
++;
150 device
->memory_properties
.memoryHeaps
[gart_index
] = (VkMemoryHeap
) {
151 .size
= device
->rad_info
.gart_size
,
152 .flags
= device
->rad_info
.has_dedicated_vram
? 0 : VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
156 STATIC_ASSERT(RADV_MEM_TYPE_COUNT
<= VK_MAX_MEMORY_TYPES
);
157 unsigned type_count
= 0;
158 if (vram_index
>= 0) {
159 device
->mem_type_indices
[type_count
] = RADV_MEM_TYPE_VRAM
;
160 device
->memory_properties
.memoryTypes
[type_count
++] = (VkMemoryType
) {
161 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
,
162 .heapIndex
= vram_index
,
165 if (gart_index
>= 0) {
166 device
->mem_type_indices
[type_count
] = RADV_MEM_TYPE_GTT_WRITE_COMBINE
;
167 device
->memory_properties
.memoryTypes
[type_count
++] = (VkMemoryType
) {
168 .propertyFlags
= VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
169 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
170 (device
->rad_info
.has_dedicated_vram
? 0 : VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
),
171 .heapIndex
= gart_index
,
174 if (visible_vram_index
>= 0) {
175 device
->mem_type_indices
[type_count
] = RADV_MEM_TYPE_VRAM_CPU_ACCESS
;
176 device
->memory_properties
.memoryTypes
[type_count
++] = (VkMemoryType
) {
177 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
178 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
179 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
180 .heapIndex
= visible_vram_index
,
183 if (gart_index
>= 0) {
184 device
->mem_type_indices
[type_count
] = RADV_MEM_TYPE_GTT_CACHED
;
185 device
->memory_properties
.memoryTypes
[type_count
++] = (VkMemoryType
) {
186 .propertyFlags
= VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
187 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
188 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
|
189 (device
->rad_info
.has_dedicated_vram
? 0 : VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
),
190 .heapIndex
= gart_index
,
193 device
->memory_properties
.memoryTypeCount
= type_count
;
197 radv_handle_env_var_force_family(struct radv_physical_device
*device
)
199 const char *family
= getenv("RADV_FORCE_FAMILY");
205 for (i
= CHIP_TAHITI
; i
< CHIP_LAST
; i
++) {
206 if (!strcmp(family
, ac_get_llvm_processor_name(i
))) {
207 /* Override family and chip_class. */
208 device
->rad_info
.family
= i
;
210 if (i
>= CHIP_VEGA10
)
211 device
->rad_info
.chip_class
= GFX9
;
212 else if (i
>= CHIP_TONGA
)
213 device
->rad_info
.chip_class
= VI
;
214 else if (i
>= CHIP_BONAIRE
)
215 device
->rad_info
.chip_class
= CIK
;
217 device
->rad_info
.chip_class
= SI
;
223 fprintf(stderr
, "radv: Unknown family: %s\n", family
);
228 radv_physical_device_init(struct radv_physical_device
*device
,
229 struct radv_instance
*instance
,
230 drmDevicePtr drm_device
)
232 const char *path
= drm_device
->nodes
[DRM_NODE_RENDER
];
234 drmVersionPtr version
;
238 fd
= open(path
, O_RDWR
| O_CLOEXEC
);
240 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
241 radv_logi("Could not open device '%s'", path
);
243 return vk_error(instance
, VK_ERROR_INCOMPATIBLE_DRIVER
);
246 version
= drmGetVersion(fd
);
250 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
251 radv_logi("Could not get the kernel driver version for device '%s'", path
);
253 return vk_errorf(instance
, VK_ERROR_INCOMPATIBLE_DRIVER
,
254 "failed to get version %s: %m", path
);
257 if (strcmp(version
->name
, "amdgpu")) {
258 drmFreeVersion(version
);
261 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
262 radv_logi("Device '%s' is not using the amdgpu kernel driver.", path
);
264 return VK_ERROR_INCOMPATIBLE_DRIVER
;
266 drmFreeVersion(version
);
268 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
269 radv_logi("Found compatible device '%s'.", path
);
271 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
272 device
->instance
= instance
;
273 assert(strlen(path
) < ARRAY_SIZE(device
->path
));
274 strncpy(device
->path
, path
, ARRAY_SIZE(device
->path
));
276 device
->ws
= radv_amdgpu_winsys_create(fd
, instance
->debug_flags
,
277 instance
->perftest_flags
);
279 result
= vk_error(instance
, VK_ERROR_INCOMPATIBLE_DRIVER
);
283 if (instance
->enabled_extensions
.KHR_display
) {
284 master_fd
= open(drm_device
->nodes
[DRM_NODE_PRIMARY
], O_RDWR
| O_CLOEXEC
);
285 if (master_fd
>= 0) {
286 uint32_t accel_working
= 0;
287 struct drm_amdgpu_info request
= {
288 .return_pointer
= (uintptr_t)&accel_working
,
289 .return_size
= sizeof(accel_working
),
290 .query
= AMDGPU_INFO_ACCEL_WORKING
293 if (drmCommandWrite(master_fd
, DRM_AMDGPU_INFO
, &request
, sizeof (struct drm_amdgpu_info
)) < 0 || !accel_working
) {
300 device
->master_fd
= master_fd
;
301 device
->local_fd
= fd
;
302 device
->ws
->query_info(device
->ws
, &device
->rad_info
);
304 radv_handle_env_var_force_family(device
);
306 radv_get_device_name(device
->rad_info
.family
, device
->name
, sizeof(device
->name
));
308 if (radv_device_get_cache_uuid(device
->rad_info
.family
, device
->cache_uuid
)) {
309 device
->ws
->destroy(device
->ws
);
310 result
= vk_errorf(instance
, VK_ERROR_INITIALIZATION_FAILED
,
311 "cannot generate UUID");
315 /* These flags affect shader compilation. */
316 uint64_t shader_env_flags
=
317 (device
->instance
->perftest_flags
& RADV_PERFTEST_SISCHED
? 0x1 : 0) |
318 (device
->instance
->debug_flags
& RADV_DEBUG_UNSAFE_MATH
? 0x2 : 0);
320 /* The gpu id is already embedded in the uuid so we just pass "radv"
321 * when creating the cache.
323 char buf
[VK_UUID_SIZE
* 2 + 1];
324 disk_cache_format_hex_id(buf
, device
->cache_uuid
, VK_UUID_SIZE
* 2);
325 device
->disk_cache
= disk_cache_create(device
->name
, buf
, shader_env_flags
);
327 if (device
->rad_info
.chip_class
< VI
||
328 device
->rad_info
.chip_class
> GFX9
)
329 fprintf(stderr
, "WARNING: radv is not a conformant vulkan implementation, testing use only.\n");
331 radv_get_driver_uuid(&device
->device_uuid
);
332 radv_get_device_uuid(&device
->rad_info
, &device
->device_uuid
);
334 if (device
->rad_info
.family
== CHIP_STONEY
||
335 device
->rad_info
.chip_class
>= GFX9
) {
336 device
->has_rbplus
= true;
337 device
->rbplus_allowed
= device
->rad_info
.family
== CHIP_STONEY
||
338 device
->rad_info
.family
== CHIP_VEGA12
||
339 device
->rad_info
.family
== CHIP_RAVEN
||
340 device
->rad_info
.family
== CHIP_RAVEN2
;
343 /* The mere presence of CLEAR_STATE in the IB causes random GPU hangs
346 device
->has_clear_state
= device
->rad_info
.chip_class
>= CIK
;
348 device
->cpdma_prefetch_writes_memory
= device
->rad_info
.chip_class
<= VI
;
350 /* Vega10/Raven need a special workaround for a hardware bug. */
351 device
->has_scissor_bug
= device
->rad_info
.family
== CHIP_VEGA10
||
352 device
->rad_info
.family
== CHIP_RAVEN
;
354 /* Out-of-order primitive rasterization. */
355 device
->has_out_of_order_rast
= device
->rad_info
.chip_class
>= VI
&&
356 device
->rad_info
.max_se
>= 2;
357 device
->out_of_order_rast_allowed
= device
->has_out_of_order_rast
&&
358 !(device
->instance
->debug_flags
& RADV_DEBUG_NO_OUT_OF_ORDER
);
360 device
->dcc_msaa_allowed
=
361 (device
->instance
->perftest_flags
& RADV_PERFTEST_DCC_MSAA
);
363 radv_physical_device_init_mem_types(device
);
364 radv_fill_device_extension_table(device
, &device
->supported_extensions
);
366 device
->bus_info
= *drm_device
->businfo
.pci
;
368 if ((device
->instance
->debug_flags
& RADV_DEBUG_INFO
))
369 ac_print_gpu_info(&device
->rad_info
);
371 /* The WSI is structured as a layer on top of the driver, so this has
372 * to be the last part of initialization (at least until we get other
375 result
= radv_init_wsi(device
);
376 if (result
!= VK_SUCCESS
) {
377 device
->ws
->destroy(device
->ws
);
378 vk_error(instance
, result
);
392 radv_physical_device_finish(struct radv_physical_device
*device
)
394 radv_finish_wsi(device
);
395 device
->ws
->destroy(device
->ws
);
396 disk_cache_destroy(device
->disk_cache
);
397 close(device
->local_fd
);
398 if (device
->master_fd
!= -1)
399 close(device
->master_fd
);
403 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
404 VkSystemAllocationScope allocationScope
)
410 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
411 size_t align
, VkSystemAllocationScope allocationScope
)
413 return realloc(pOriginal
, size
);
417 default_free_func(void *pUserData
, void *pMemory
)
422 static const VkAllocationCallbacks default_alloc
= {
424 .pfnAllocation
= default_alloc_func
,
425 .pfnReallocation
= default_realloc_func
,
426 .pfnFree
= default_free_func
,
429 static const struct debug_control radv_debug_options
[] = {
430 {"nofastclears", RADV_DEBUG_NO_FAST_CLEARS
},
431 {"nodcc", RADV_DEBUG_NO_DCC
},
432 {"shaders", RADV_DEBUG_DUMP_SHADERS
},
433 {"nocache", RADV_DEBUG_NO_CACHE
},
434 {"shaderstats", RADV_DEBUG_DUMP_SHADER_STATS
},
435 {"nohiz", RADV_DEBUG_NO_HIZ
},
436 {"nocompute", RADV_DEBUG_NO_COMPUTE_QUEUE
},
437 {"unsafemath", RADV_DEBUG_UNSAFE_MATH
},
438 {"allbos", RADV_DEBUG_ALL_BOS
},
439 {"noibs", RADV_DEBUG_NO_IBS
},
440 {"spirv", RADV_DEBUG_DUMP_SPIRV
},
441 {"vmfaults", RADV_DEBUG_VM_FAULTS
},
442 {"zerovram", RADV_DEBUG_ZERO_VRAM
},
443 {"syncshaders", RADV_DEBUG_SYNC_SHADERS
},
444 {"nosisched", RADV_DEBUG_NO_SISCHED
},
445 {"preoptir", RADV_DEBUG_PREOPTIR
},
446 {"nodynamicbounds", RADV_DEBUG_NO_DYNAMIC_BOUNDS
},
447 {"nooutoforder", RADV_DEBUG_NO_OUT_OF_ORDER
},
448 {"info", RADV_DEBUG_INFO
},
449 {"errors", RADV_DEBUG_ERRORS
},
450 {"startup", RADV_DEBUG_STARTUP
},
451 {"checkir", RADV_DEBUG_CHECKIR
},
452 {"nothreadllvm", RADV_DEBUG_NOTHREADLLVM
},
453 {"nobinning", RADV_DEBUG_NOBINNING
},
458 radv_get_debug_option_name(int id
)
460 assert(id
< ARRAY_SIZE(radv_debug_options
) - 1);
461 return radv_debug_options
[id
].string
;
464 static const struct debug_control radv_perftest_options
[] = {
465 {"nobatchchain", RADV_PERFTEST_NO_BATCHCHAIN
},
466 {"sisched", RADV_PERFTEST_SISCHED
},
467 {"localbos", RADV_PERFTEST_LOCAL_BOS
},
468 {"dccmsaa", RADV_PERFTEST_DCC_MSAA
},
473 radv_get_perftest_option_name(int id
)
475 assert(id
< ARRAY_SIZE(radv_perftest_options
) - 1);
476 return radv_perftest_options
[id
].string
;
480 radv_handle_per_app_options(struct radv_instance
*instance
,
481 const VkApplicationInfo
*info
)
483 const char *name
= info
? info
->pApplicationName
: NULL
;
488 if (!strcmp(name
, "Talos - Linux - 32bit") ||
489 !strcmp(name
, "Talos - Linux - 64bit")) {
490 if (!(instance
->debug_flags
& RADV_DEBUG_NO_SISCHED
)) {
491 /* Force enable LLVM sisched for Talos because it looks
492 * safe and it gives few more FPS.
494 instance
->perftest_flags
|= RADV_PERFTEST_SISCHED
;
496 } else if (!strcmp(name
, "DOOM_VFR")) {
497 /* Work around a Doom VFR game bug */
498 instance
->debug_flags
|= RADV_DEBUG_NO_DYNAMIC_BOUNDS
;
502 static int radv_get_instance_extension_index(const char *name
)
504 for (unsigned i
= 0; i
< RADV_INSTANCE_EXTENSION_COUNT
; ++i
) {
505 if (strcmp(name
, radv_instance_extensions
[i
].extensionName
) == 0)
512 VkResult
radv_CreateInstance(
513 const VkInstanceCreateInfo
* pCreateInfo
,
514 const VkAllocationCallbacks
* pAllocator
,
515 VkInstance
* pInstance
)
517 struct radv_instance
*instance
;
520 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
522 uint32_t client_version
;
523 if (pCreateInfo
->pApplicationInfo
&&
524 pCreateInfo
->pApplicationInfo
->apiVersion
!= 0) {
525 client_version
= pCreateInfo
->pApplicationInfo
->apiVersion
;
527 radv_EnumerateInstanceVersion(&client_version
);
530 instance
= vk_zalloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
531 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
533 return vk_error(NULL
, VK_ERROR_OUT_OF_HOST_MEMORY
);
535 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
538 instance
->alloc
= *pAllocator
;
540 instance
->alloc
= default_alloc
;
542 instance
->apiVersion
= client_version
;
543 instance
->physicalDeviceCount
= -1;
545 instance
->debug_flags
= parse_debug_string(getenv("RADV_DEBUG"),
548 instance
->perftest_flags
= parse_debug_string(getenv("RADV_PERFTEST"),
549 radv_perftest_options
);
552 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
553 radv_logi("Created an instance");
555 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
556 const char *ext_name
= pCreateInfo
->ppEnabledExtensionNames
[i
];
557 int index
= radv_get_instance_extension_index(ext_name
);
559 if (index
< 0 || !radv_supported_instance_extensions
.extensions
[index
]) {
560 vk_free2(&default_alloc
, pAllocator
, instance
);
561 return vk_error(instance
, VK_ERROR_EXTENSION_NOT_PRESENT
);
564 instance
->enabled_extensions
.extensions
[index
] = true;
567 result
= vk_debug_report_instance_init(&instance
->debug_report_callbacks
);
568 if (result
!= VK_SUCCESS
) {
569 vk_free2(&default_alloc
, pAllocator
, instance
);
570 return vk_error(instance
, result
);
575 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
577 radv_handle_per_app_options(instance
, pCreateInfo
->pApplicationInfo
);
579 *pInstance
= radv_instance_to_handle(instance
);
584 void radv_DestroyInstance(
585 VkInstance _instance
,
586 const VkAllocationCallbacks
* pAllocator
)
588 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
593 for (int i
= 0; i
< instance
->physicalDeviceCount
; ++i
) {
594 radv_physical_device_finish(instance
->physicalDevices
+ i
);
597 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
601 vk_debug_report_instance_destroy(&instance
->debug_report_callbacks
);
603 vk_free(&instance
->alloc
, instance
);
607 radv_enumerate_devices(struct radv_instance
*instance
)
609 /* TODO: Check for more devices ? */
610 drmDevicePtr devices
[8];
611 VkResult result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
614 instance
->physicalDeviceCount
= 0;
616 max_devices
= drmGetDevices2(0, devices
, ARRAY_SIZE(devices
));
618 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
619 radv_logi("Found %d drm nodes", max_devices
);
622 return vk_error(instance
, VK_ERROR_INCOMPATIBLE_DRIVER
);
624 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
625 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
626 devices
[i
]->bustype
== DRM_BUS_PCI
&&
627 devices
[i
]->deviceinfo
.pci
->vendor_id
== ATI_VENDOR_ID
) {
629 result
= radv_physical_device_init(instance
->physicalDevices
+
630 instance
->physicalDeviceCount
,
633 if (result
== VK_SUCCESS
)
634 ++instance
->physicalDeviceCount
;
635 else if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
639 drmFreeDevices(devices
, max_devices
);
644 VkResult
radv_EnumeratePhysicalDevices(
645 VkInstance _instance
,
646 uint32_t* pPhysicalDeviceCount
,
647 VkPhysicalDevice
* pPhysicalDevices
)
649 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
652 if (instance
->physicalDeviceCount
< 0) {
653 result
= radv_enumerate_devices(instance
);
654 if (result
!= VK_SUCCESS
&&
655 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
659 if (!pPhysicalDevices
) {
660 *pPhysicalDeviceCount
= instance
->physicalDeviceCount
;
662 *pPhysicalDeviceCount
= MIN2(*pPhysicalDeviceCount
, instance
->physicalDeviceCount
);
663 for (unsigned i
= 0; i
< *pPhysicalDeviceCount
; ++i
)
664 pPhysicalDevices
[i
] = radv_physical_device_to_handle(instance
->physicalDevices
+ i
);
667 return *pPhysicalDeviceCount
< instance
->physicalDeviceCount
? VK_INCOMPLETE
671 VkResult
radv_EnumeratePhysicalDeviceGroups(
672 VkInstance _instance
,
673 uint32_t* pPhysicalDeviceGroupCount
,
674 VkPhysicalDeviceGroupProperties
* pPhysicalDeviceGroupProperties
)
676 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
679 if (instance
->physicalDeviceCount
< 0) {
680 result
= radv_enumerate_devices(instance
);
681 if (result
!= VK_SUCCESS
&&
682 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
686 if (!pPhysicalDeviceGroupProperties
) {
687 *pPhysicalDeviceGroupCount
= instance
->physicalDeviceCount
;
689 *pPhysicalDeviceGroupCount
= MIN2(*pPhysicalDeviceGroupCount
, instance
->physicalDeviceCount
);
690 for (unsigned i
= 0; i
< *pPhysicalDeviceGroupCount
; ++i
) {
691 pPhysicalDeviceGroupProperties
[i
].physicalDeviceCount
= 1;
692 pPhysicalDeviceGroupProperties
[i
].physicalDevices
[0] = radv_physical_device_to_handle(instance
->physicalDevices
+ i
);
693 pPhysicalDeviceGroupProperties
[i
].subsetAllocation
= false;
696 return *pPhysicalDeviceGroupCount
< instance
->physicalDeviceCount
? VK_INCOMPLETE
700 void radv_GetPhysicalDeviceFeatures(
701 VkPhysicalDevice physicalDevice
,
702 VkPhysicalDeviceFeatures
* pFeatures
)
704 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
705 memset(pFeatures
, 0, sizeof(*pFeatures
));
707 *pFeatures
= (VkPhysicalDeviceFeatures
) {
708 .robustBufferAccess
= true,
709 .fullDrawIndexUint32
= true,
710 .imageCubeArray
= true,
711 .independentBlend
= true,
712 .geometryShader
= true,
713 .tessellationShader
= true,
714 .sampleRateShading
= true,
715 .dualSrcBlend
= true,
717 .multiDrawIndirect
= true,
718 .drawIndirectFirstInstance
= true,
720 .depthBiasClamp
= true,
721 .fillModeNonSolid
= true,
726 .multiViewport
= true,
727 .samplerAnisotropy
= true,
728 .textureCompressionETC2
= pdevice
->rad_info
.chip_class
>= GFX9
||
729 pdevice
->rad_info
.family
== CHIP_STONEY
,
730 .textureCompressionASTC_LDR
= false,
731 .textureCompressionBC
= true,
732 .occlusionQueryPrecise
= true,
733 .pipelineStatisticsQuery
= true,
734 .vertexPipelineStoresAndAtomics
= true,
735 .fragmentStoresAndAtomics
= true,
736 .shaderTessellationAndGeometryPointSize
= true,
737 .shaderImageGatherExtended
= true,
738 .shaderStorageImageExtendedFormats
= true,
739 .shaderStorageImageMultisample
= false,
740 .shaderUniformBufferArrayDynamicIndexing
= true,
741 .shaderSampledImageArrayDynamicIndexing
= true,
742 .shaderStorageBufferArrayDynamicIndexing
= true,
743 .shaderStorageImageArrayDynamicIndexing
= true,
744 .shaderStorageImageReadWithoutFormat
= true,
745 .shaderStorageImageWriteWithoutFormat
= true,
746 .shaderClipDistance
= true,
747 .shaderCullDistance
= true,
748 .shaderFloat64
= true,
750 .shaderInt16
= pdevice
->rad_info
.chip_class
>= GFX9
,
751 .sparseBinding
= true,
752 .variableMultisampleRate
= true,
753 .inheritedQueries
= true,
757 void radv_GetPhysicalDeviceFeatures2(
758 VkPhysicalDevice physicalDevice
,
759 VkPhysicalDeviceFeatures2KHR
*pFeatures
)
761 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
762 vk_foreach_struct(ext
, pFeatures
->pNext
) {
763 switch (ext
->sType
) {
764 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTER_FEATURES_KHR
: {
765 VkPhysicalDeviceVariablePointerFeaturesKHR
*features
= (void *)ext
;
766 features
->variablePointersStorageBuffer
= true;
767 features
->variablePointers
= false;
770 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES_KHR
: {
771 VkPhysicalDeviceMultiviewFeaturesKHR
*features
= (VkPhysicalDeviceMultiviewFeaturesKHR
*)ext
;
772 features
->multiview
= true;
773 features
->multiviewGeometryShader
= true;
774 features
->multiviewTessellationShader
= true;
777 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETER_FEATURES
: {
778 VkPhysicalDeviceShaderDrawParameterFeatures
*features
=
779 (VkPhysicalDeviceShaderDrawParameterFeatures
*)ext
;
780 features
->shaderDrawParameters
= true;
783 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES
: {
784 VkPhysicalDeviceProtectedMemoryFeatures
*features
=
785 (VkPhysicalDeviceProtectedMemoryFeatures
*)ext
;
786 features
->protectedMemory
= false;
789 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES
: {
790 VkPhysicalDevice16BitStorageFeatures
*features
=
791 (VkPhysicalDevice16BitStorageFeatures
*)ext
;
792 bool enabled
= pdevice
->rad_info
.chip_class
>= VI
;
793 features
->storageBuffer16BitAccess
= enabled
;
794 features
->uniformAndStorageBuffer16BitAccess
= enabled
;
795 features
->storagePushConstant16
= enabled
;
796 features
->storageInputOutput16
= enabled
;
799 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES
: {
800 VkPhysicalDeviceSamplerYcbcrConversionFeatures
*features
=
801 (VkPhysicalDeviceSamplerYcbcrConversionFeatures
*)ext
;
802 features
->samplerYcbcrConversion
= false;
805 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT
: {
806 VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*features
=
807 (VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*)ext
;
808 features
->shaderInputAttachmentArrayDynamicIndexing
= true;
809 features
->shaderUniformTexelBufferArrayDynamicIndexing
= true;
810 features
->shaderStorageTexelBufferArrayDynamicIndexing
= true;
811 features
->shaderUniformBufferArrayNonUniformIndexing
= false;
812 features
->shaderSampledImageArrayNonUniformIndexing
= false;
813 features
->shaderStorageBufferArrayNonUniformIndexing
= false;
814 features
->shaderStorageImageArrayNonUniformIndexing
= false;
815 features
->shaderInputAttachmentArrayNonUniformIndexing
= false;
816 features
->shaderUniformTexelBufferArrayNonUniformIndexing
= false;
817 features
->shaderStorageTexelBufferArrayNonUniformIndexing
= false;
818 features
->descriptorBindingUniformBufferUpdateAfterBind
= true;
819 features
->descriptorBindingSampledImageUpdateAfterBind
= true;
820 features
->descriptorBindingStorageImageUpdateAfterBind
= true;
821 features
->descriptorBindingStorageBufferUpdateAfterBind
= true;
822 features
->descriptorBindingUniformTexelBufferUpdateAfterBind
= true;
823 features
->descriptorBindingStorageTexelBufferUpdateAfterBind
= true;
824 features
->descriptorBindingUpdateUnusedWhilePending
= true;
825 features
->descriptorBindingPartiallyBound
= true;
826 features
->descriptorBindingVariableDescriptorCount
= true;
827 features
->runtimeDescriptorArray
= true;
830 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT
: {
831 VkPhysicalDeviceConditionalRenderingFeaturesEXT
*features
=
832 (VkPhysicalDeviceConditionalRenderingFeaturesEXT
*)ext
;
833 features
->conditionalRendering
= true;
834 features
->inheritedConditionalRendering
= false;
837 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT
: {
838 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*features
=
839 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*)ext
;
840 features
->vertexAttributeInstanceRateDivisor
= VK_TRUE
;
841 features
->vertexAttributeInstanceRateZeroDivisor
= VK_TRUE
;
844 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT
: {
845 VkPhysicalDeviceTransformFeedbackFeaturesEXT
*features
=
846 (VkPhysicalDeviceTransformFeedbackFeaturesEXT
*)ext
;
847 features
->transformFeedback
= true;
848 features
->geometryStreams
= true;
851 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES_EXT
: {
852 VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*features
=
853 (VkPhysicalDeviceScalarBlockLayoutFeaturesEXT
*)ext
;
854 features
->scalarBlockLayout
= pdevice
->rad_info
.chip_class
>= CIK
;
861 return radv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
864 void radv_GetPhysicalDeviceProperties(
865 VkPhysicalDevice physicalDevice
,
866 VkPhysicalDeviceProperties
* pProperties
)
868 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
869 VkSampleCountFlags sample_counts
= 0xf;
871 /* make sure that the entire descriptor set is addressable with a signed
872 * 32-bit int. So the sum of all limits scaled by descriptor size has to
873 * be at most 2 GiB. the combined image & samples object count as one of
874 * both. This limit is for the pipeline layout, not for the set layout, but
875 * there is no set limit, so we just set a pipeline limit. I don't think
876 * any app is going to hit this soon. */
877 size_t max_descriptor_set_size
= ((1ull << 31) - 16 * MAX_DYNAMIC_BUFFERS
) /
878 (32 /* uniform buffer, 32 due to potential space wasted on alignment */ +
879 32 /* storage buffer, 32 due to potential space wasted on alignment */ +
880 32 /* sampler, largest when combined with image */ +
881 64 /* sampled image */ +
882 64 /* storage image */);
884 VkPhysicalDeviceLimits limits
= {
885 .maxImageDimension1D
= (1 << 14),
886 .maxImageDimension2D
= (1 << 14),
887 .maxImageDimension3D
= (1 << 11),
888 .maxImageDimensionCube
= (1 << 14),
889 .maxImageArrayLayers
= (1 << 11),
890 .maxTexelBufferElements
= 128 * 1024 * 1024,
891 .maxUniformBufferRange
= UINT32_MAX
,
892 .maxStorageBufferRange
= UINT32_MAX
,
893 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
894 .maxMemoryAllocationCount
= UINT32_MAX
,
895 .maxSamplerAllocationCount
= 64 * 1024,
896 .bufferImageGranularity
= 64, /* A cache line */
897 .sparseAddressSpaceSize
= 0xffffffffu
, /* buffer max size */
898 .maxBoundDescriptorSets
= MAX_SETS
,
899 .maxPerStageDescriptorSamplers
= max_descriptor_set_size
,
900 .maxPerStageDescriptorUniformBuffers
= max_descriptor_set_size
,
901 .maxPerStageDescriptorStorageBuffers
= max_descriptor_set_size
,
902 .maxPerStageDescriptorSampledImages
= max_descriptor_set_size
,
903 .maxPerStageDescriptorStorageImages
= max_descriptor_set_size
,
904 .maxPerStageDescriptorInputAttachments
= max_descriptor_set_size
,
905 .maxPerStageResources
= max_descriptor_set_size
,
906 .maxDescriptorSetSamplers
= max_descriptor_set_size
,
907 .maxDescriptorSetUniformBuffers
= max_descriptor_set_size
,
908 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_UNIFORM_BUFFERS
,
909 .maxDescriptorSetStorageBuffers
= max_descriptor_set_size
,
910 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_STORAGE_BUFFERS
,
911 .maxDescriptorSetSampledImages
= max_descriptor_set_size
,
912 .maxDescriptorSetStorageImages
= max_descriptor_set_size
,
913 .maxDescriptorSetInputAttachments
= max_descriptor_set_size
,
914 .maxVertexInputAttributes
= 32,
915 .maxVertexInputBindings
= 32,
916 .maxVertexInputAttributeOffset
= 2047,
917 .maxVertexInputBindingStride
= 2048,
918 .maxVertexOutputComponents
= 128,
919 .maxTessellationGenerationLevel
= 64,
920 .maxTessellationPatchSize
= 32,
921 .maxTessellationControlPerVertexInputComponents
= 128,
922 .maxTessellationControlPerVertexOutputComponents
= 128,
923 .maxTessellationControlPerPatchOutputComponents
= 120,
924 .maxTessellationControlTotalOutputComponents
= 4096,
925 .maxTessellationEvaluationInputComponents
= 128,
926 .maxTessellationEvaluationOutputComponents
= 128,
927 .maxGeometryShaderInvocations
= 127,
928 .maxGeometryInputComponents
= 64,
929 .maxGeometryOutputComponents
= 128,
930 .maxGeometryOutputVertices
= 256,
931 .maxGeometryTotalOutputComponents
= 1024,
932 .maxFragmentInputComponents
= 128,
933 .maxFragmentOutputAttachments
= 8,
934 .maxFragmentDualSrcAttachments
= 1,
935 .maxFragmentCombinedOutputResources
= 8,
936 .maxComputeSharedMemorySize
= 32768,
937 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
938 .maxComputeWorkGroupInvocations
= 2048,
939 .maxComputeWorkGroupSize
= {
944 .subPixelPrecisionBits
= 4 /* FIXME */,
945 .subTexelPrecisionBits
= 4 /* FIXME */,
946 .mipmapPrecisionBits
= 4 /* FIXME */,
947 .maxDrawIndexedIndexValue
= UINT32_MAX
,
948 .maxDrawIndirectCount
= UINT32_MAX
,
949 .maxSamplerLodBias
= 16,
950 .maxSamplerAnisotropy
= 16,
951 .maxViewports
= MAX_VIEWPORTS
,
952 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
953 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
954 .viewportSubPixelBits
= 8,
955 .minMemoryMapAlignment
= 4096, /* A page */
956 .minTexelBufferOffsetAlignment
= 1,
957 .minUniformBufferOffsetAlignment
= 4,
958 .minStorageBufferOffsetAlignment
= 4,
959 .minTexelOffset
= -32,
960 .maxTexelOffset
= 31,
961 .minTexelGatherOffset
= -32,
962 .maxTexelGatherOffset
= 31,
963 .minInterpolationOffset
= -2,
964 .maxInterpolationOffset
= 2,
965 .subPixelInterpolationOffsetBits
= 8,
966 .maxFramebufferWidth
= (1 << 14),
967 .maxFramebufferHeight
= (1 << 14),
968 .maxFramebufferLayers
= (1 << 10),
969 .framebufferColorSampleCounts
= sample_counts
,
970 .framebufferDepthSampleCounts
= sample_counts
,
971 .framebufferStencilSampleCounts
= sample_counts
,
972 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
973 .maxColorAttachments
= MAX_RTS
,
974 .sampledImageColorSampleCounts
= sample_counts
,
975 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
976 .sampledImageDepthSampleCounts
= sample_counts
,
977 .sampledImageStencilSampleCounts
= sample_counts
,
978 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
979 .maxSampleMaskWords
= 1,
980 .timestampComputeAndGraphics
= true,
981 .timestampPeriod
= 1000000.0 / pdevice
->rad_info
.clock_crystal_freq
,
982 .maxClipDistances
= 8,
983 .maxCullDistances
= 8,
984 .maxCombinedClipAndCullDistances
= 8,
985 .discreteQueuePriorities
= 2,
986 .pointSizeRange
= { 0.125, 255.875 },
987 .lineWidthRange
= { 0.0, 7.9921875 },
988 .pointSizeGranularity
= (1.0 / 8.0),
989 .lineWidthGranularity
= (1.0 / 128.0),
990 .strictLines
= false, /* FINISHME */
991 .standardSampleLocations
= true,
992 .optimalBufferCopyOffsetAlignment
= 128,
993 .optimalBufferCopyRowPitchAlignment
= 128,
994 .nonCoherentAtomSize
= 64,
997 *pProperties
= (VkPhysicalDeviceProperties
) {
998 .apiVersion
= radv_physical_device_api_version(pdevice
),
999 .driverVersion
= vk_get_driver_version(),
1000 .vendorID
= ATI_VENDOR_ID
,
1001 .deviceID
= pdevice
->rad_info
.pci_id
,
1002 .deviceType
= pdevice
->rad_info
.has_dedicated_vram
? VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU
: VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
1004 .sparseProperties
= {0},
1007 strcpy(pProperties
->deviceName
, pdevice
->name
);
1008 memcpy(pProperties
->pipelineCacheUUID
, pdevice
->cache_uuid
, VK_UUID_SIZE
);
1011 void radv_GetPhysicalDeviceProperties2(
1012 VkPhysicalDevice physicalDevice
,
1013 VkPhysicalDeviceProperties2KHR
*pProperties
)
1015 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
1016 radv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
1018 vk_foreach_struct(ext
, pProperties
->pNext
) {
1019 switch (ext
->sType
) {
1020 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
1021 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
1022 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
1023 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
1026 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES_KHR
: {
1027 VkPhysicalDeviceIDPropertiesKHR
*properties
= (VkPhysicalDeviceIDPropertiesKHR
*)ext
;
1028 memcpy(properties
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
1029 memcpy(properties
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
1030 properties
->deviceLUIDValid
= false;
1033 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES_KHR
: {
1034 VkPhysicalDeviceMultiviewPropertiesKHR
*properties
= (VkPhysicalDeviceMultiviewPropertiesKHR
*)ext
;
1035 properties
->maxMultiviewViewCount
= MAX_VIEWS
;
1036 properties
->maxMultiviewInstanceIndex
= INT_MAX
;
1039 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES_KHR
: {
1040 VkPhysicalDevicePointClippingPropertiesKHR
*properties
=
1041 (VkPhysicalDevicePointClippingPropertiesKHR
*)ext
;
1042 properties
->pointClippingBehavior
= VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES_KHR
;
1045 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DISCARD_RECTANGLE_PROPERTIES_EXT
: {
1046 VkPhysicalDeviceDiscardRectanglePropertiesEXT
*properties
=
1047 (VkPhysicalDeviceDiscardRectanglePropertiesEXT
*)ext
;
1048 properties
->maxDiscardRectangles
= MAX_DISCARD_RECTANGLES
;
1051 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT
: {
1052 VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*properties
=
1053 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*) ext
;
1054 properties
->minImportedHostPointerAlignment
= 4096;
1057 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES
: {
1058 VkPhysicalDeviceSubgroupProperties
*properties
=
1059 (VkPhysicalDeviceSubgroupProperties
*)ext
;
1060 properties
->subgroupSize
= 64;
1061 properties
->supportedStages
= VK_SHADER_STAGE_ALL
;
1062 properties
->supportedOperations
=
1063 VK_SUBGROUP_FEATURE_BASIC_BIT
|
1064 VK_SUBGROUP_FEATURE_BALLOT_BIT
|
1065 VK_SUBGROUP_FEATURE_QUAD_BIT
|
1066 VK_SUBGROUP_FEATURE_VOTE_BIT
;
1067 if (pdevice
->rad_info
.chip_class
>= VI
) {
1068 properties
->supportedOperations
|=
1069 VK_SUBGROUP_FEATURE_ARITHMETIC_BIT
|
1070 VK_SUBGROUP_FEATURE_SHUFFLE_BIT
|
1071 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT
;
1073 properties
->quadOperationsInAllStages
= true;
1076 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES
: {
1077 VkPhysicalDeviceMaintenance3Properties
*properties
=
1078 (VkPhysicalDeviceMaintenance3Properties
*)ext
;
1079 /* Make sure everything is addressable by a signed 32-bit int, and
1080 * our largest descriptors are 96 bytes. */
1081 properties
->maxPerSetDescriptors
= (1ull << 31) / 96;
1082 /* Our buffer size fields allow only this much */
1083 properties
->maxMemoryAllocationSize
= 0xFFFFFFFFull
;
1086 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT
: {
1087 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*properties
=
1088 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*)ext
;
1089 /* GFX6-8 only support single channel min/max filter. */
1090 properties
->filterMinmaxImageComponentMapping
= pdevice
->rad_info
.chip_class
>= GFX9
;
1091 properties
->filterMinmaxSingleComponentFormats
= true;
1094 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CORE_PROPERTIES_AMD
: {
1095 VkPhysicalDeviceShaderCorePropertiesAMD
*properties
=
1096 (VkPhysicalDeviceShaderCorePropertiesAMD
*)ext
;
1098 /* Shader engines. */
1099 properties
->shaderEngineCount
=
1100 pdevice
->rad_info
.max_se
;
1101 properties
->shaderArraysPerEngineCount
=
1102 pdevice
->rad_info
.max_sh_per_se
;
1103 properties
->computeUnitsPerShaderArray
=
1104 pdevice
->rad_info
.num_good_cu_per_sh
;
1105 properties
->simdPerComputeUnit
= 4;
1106 properties
->wavefrontsPerSimd
=
1107 pdevice
->rad_info
.family
== CHIP_TONGA
||
1108 pdevice
->rad_info
.family
== CHIP_ICELAND
||
1109 pdevice
->rad_info
.family
== CHIP_POLARIS10
||
1110 pdevice
->rad_info
.family
== CHIP_POLARIS11
||
1111 pdevice
->rad_info
.family
== CHIP_POLARIS12
||
1112 pdevice
->rad_info
.family
== CHIP_VEGAM
? 8 : 10;
1113 properties
->wavefrontSize
= 64;
1116 properties
->sgprsPerSimd
=
1117 radv_get_num_physical_sgprs(pdevice
);
1118 properties
->minSgprAllocation
=
1119 pdevice
->rad_info
.chip_class
>= VI
? 16 : 8;
1120 properties
->maxSgprAllocation
=
1121 pdevice
->rad_info
.family
== CHIP_TONGA
||
1122 pdevice
->rad_info
.family
== CHIP_ICELAND
? 96 : 104;
1123 properties
->sgprAllocationGranularity
=
1124 pdevice
->rad_info
.chip_class
>= VI
? 16 : 8;
1127 properties
->vgprsPerSimd
= RADV_NUM_PHYSICAL_VGPRS
;
1128 properties
->minVgprAllocation
= 4;
1129 properties
->maxVgprAllocation
= 256;
1130 properties
->vgprAllocationGranularity
= 4;
1133 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT
: {
1134 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*properties
=
1135 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*)ext
;
1136 properties
->maxVertexAttribDivisor
= UINT32_MAX
;
1139 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES_EXT
: {
1140 VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*properties
=
1141 (VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*)ext
;
1142 properties
->maxUpdateAfterBindDescriptorsInAllPools
= UINT32_MAX
/ 64;
1143 properties
->shaderUniformBufferArrayNonUniformIndexingNative
= false;
1144 properties
->shaderSampledImageArrayNonUniformIndexingNative
= false;
1145 properties
->shaderStorageBufferArrayNonUniformIndexingNative
= false;
1146 properties
->shaderStorageImageArrayNonUniformIndexingNative
= false;
1147 properties
->shaderInputAttachmentArrayNonUniformIndexingNative
= false;
1148 properties
->robustBufferAccessUpdateAfterBind
= false;
1149 properties
->quadDivergentImplicitLod
= false;
1151 size_t max_descriptor_set_size
= ((1ull << 31) - 16 * MAX_DYNAMIC_BUFFERS
) /
1152 (32 /* uniform buffer, 32 due to potential space wasted on alignment */ +
1153 32 /* storage buffer, 32 due to potential space wasted on alignment */ +
1154 32 /* sampler, largest when combined with image */ +
1155 64 /* sampled image */ +
1156 64 /* storage image */);
1157 properties
->maxPerStageDescriptorUpdateAfterBindSamplers
= max_descriptor_set_size
;
1158 properties
->maxPerStageDescriptorUpdateAfterBindUniformBuffers
= max_descriptor_set_size
;
1159 properties
->maxPerStageDescriptorUpdateAfterBindStorageBuffers
= max_descriptor_set_size
;
1160 properties
->maxPerStageDescriptorUpdateAfterBindSampledImages
= max_descriptor_set_size
;
1161 properties
->maxPerStageDescriptorUpdateAfterBindStorageImages
= max_descriptor_set_size
;
1162 properties
->maxPerStageDescriptorUpdateAfterBindInputAttachments
= max_descriptor_set_size
;
1163 properties
->maxPerStageUpdateAfterBindResources
= max_descriptor_set_size
;
1164 properties
->maxDescriptorSetUpdateAfterBindSamplers
= max_descriptor_set_size
;
1165 properties
->maxDescriptorSetUpdateAfterBindUniformBuffers
= max_descriptor_set_size
;
1166 properties
->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic
= MAX_DYNAMIC_UNIFORM_BUFFERS
;
1167 properties
->maxDescriptorSetUpdateAfterBindStorageBuffers
= max_descriptor_set_size
;
1168 properties
->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic
= MAX_DYNAMIC_STORAGE_BUFFERS
;
1169 properties
->maxDescriptorSetUpdateAfterBindSampledImages
= max_descriptor_set_size
;
1170 properties
->maxDescriptorSetUpdateAfterBindStorageImages
= max_descriptor_set_size
;
1171 properties
->maxDescriptorSetUpdateAfterBindInputAttachments
= max_descriptor_set_size
;
1174 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES
: {
1175 VkPhysicalDeviceProtectedMemoryProperties
*properties
=
1176 (VkPhysicalDeviceProtectedMemoryProperties
*)ext
;
1177 properties
->protectedNoFault
= false;
1180 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONSERVATIVE_RASTERIZATION_PROPERTIES_EXT
: {
1181 VkPhysicalDeviceConservativeRasterizationPropertiesEXT
*properties
=
1182 (VkPhysicalDeviceConservativeRasterizationPropertiesEXT
*)ext
;
1183 properties
->primitiveOverestimationSize
= 0;
1184 properties
->maxExtraPrimitiveOverestimationSize
= 0;
1185 properties
->extraPrimitiveOverestimationSizeGranularity
= 0;
1186 properties
->primitiveUnderestimation
= VK_FALSE
;
1187 properties
->conservativePointAndLineRasterization
= VK_FALSE
;
1188 properties
->degenerateTrianglesRasterized
= VK_FALSE
;
1189 properties
->degenerateLinesRasterized
= VK_FALSE
;
1190 properties
->fullyCoveredFragmentShaderInputVariable
= VK_FALSE
;
1191 properties
->conservativeRasterizationPostDepthCoverage
= VK_FALSE
;
1194 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT
: {
1195 VkPhysicalDevicePCIBusInfoPropertiesEXT
*properties
=
1196 (VkPhysicalDevicePCIBusInfoPropertiesEXT
*)ext
;
1197 properties
->pciDomain
= pdevice
->bus_info
.domain
;
1198 properties
->pciBus
= pdevice
->bus_info
.bus
;
1199 properties
->pciDevice
= pdevice
->bus_info
.dev
;
1200 properties
->pciFunction
= pdevice
->bus_info
.func
;
1203 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR
: {
1204 VkPhysicalDeviceDriverPropertiesKHR
*driver_props
=
1205 (VkPhysicalDeviceDriverPropertiesKHR
*) ext
;
1207 driver_props
->driverID
= VK_DRIVER_ID_MESA_RADV_KHR
;
1208 memset(driver_props
->driverName
, 0, VK_MAX_DRIVER_NAME_SIZE_KHR
);
1209 strcpy(driver_props
->driverName
, "radv");
1211 memset(driver_props
->driverInfo
, 0, VK_MAX_DRIVER_INFO_SIZE_KHR
);
1212 snprintf(driver_props
->driverInfo
, VK_MAX_DRIVER_INFO_SIZE_KHR
,
1213 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1
1215 (HAVE_LLVM
>> 8) & 0xff, HAVE_LLVM
& 0xff,
1216 MESA_LLVM_VERSION_PATCH
);
1218 driver_props
->conformanceVersion
= (VkConformanceVersionKHR
) {
1226 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT
: {
1227 VkPhysicalDeviceTransformFeedbackPropertiesEXT
*properties
=
1228 (VkPhysicalDeviceTransformFeedbackPropertiesEXT
*)ext
;
1229 properties
->maxTransformFeedbackStreams
= MAX_SO_STREAMS
;
1230 properties
->maxTransformFeedbackBuffers
= MAX_SO_BUFFERS
;
1231 properties
->maxTransformFeedbackBufferSize
= UINT32_MAX
;
1232 properties
->maxTransformFeedbackStreamDataSize
= 512;
1233 properties
->maxTransformFeedbackBufferDataSize
= UINT32_MAX
;
1234 properties
->maxTransformFeedbackBufferDataStride
= 512;
1235 properties
->transformFeedbackQueries
= true;
1236 properties
->transformFeedbackStreamsLinesTriangles
= false;
1237 properties
->transformFeedbackRasterizationStreamSelect
= false;
1238 properties
->transformFeedbackDraw
= true;
1247 static void radv_get_physical_device_queue_family_properties(
1248 struct radv_physical_device
* pdevice
,
1250 VkQueueFamilyProperties
** pQueueFamilyProperties
)
1252 int num_queue_families
= 1;
1254 if (pdevice
->rad_info
.num_compute_rings
> 0 &&
1255 !(pdevice
->instance
->debug_flags
& RADV_DEBUG_NO_COMPUTE_QUEUE
))
1256 num_queue_families
++;
1258 if (pQueueFamilyProperties
== NULL
) {
1259 *pCount
= num_queue_families
;
1268 *pQueueFamilyProperties
[idx
] = (VkQueueFamilyProperties
) {
1269 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
1270 VK_QUEUE_COMPUTE_BIT
|
1271 VK_QUEUE_TRANSFER_BIT
|
1272 VK_QUEUE_SPARSE_BINDING_BIT
,
1274 .timestampValidBits
= 64,
1275 .minImageTransferGranularity
= (VkExtent3D
) { 1, 1, 1 },
1280 if (pdevice
->rad_info
.num_compute_rings
> 0 &&
1281 !(pdevice
->instance
->debug_flags
& RADV_DEBUG_NO_COMPUTE_QUEUE
)) {
1282 if (*pCount
> idx
) {
1283 *pQueueFamilyProperties
[idx
] = (VkQueueFamilyProperties
) {
1284 .queueFlags
= VK_QUEUE_COMPUTE_BIT
|
1285 VK_QUEUE_TRANSFER_BIT
|
1286 VK_QUEUE_SPARSE_BINDING_BIT
,
1287 .queueCount
= pdevice
->rad_info
.num_compute_rings
,
1288 .timestampValidBits
= 64,
1289 .minImageTransferGranularity
= (VkExtent3D
) { 1, 1, 1 },
1297 void radv_GetPhysicalDeviceQueueFamilyProperties(
1298 VkPhysicalDevice physicalDevice
,
1300 VkQueueFamilyProperties
* pQueueFamilyProperties
)
1302 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
1303 if (!pQueueFamilyProperties
) {
1304 return radv_get_physical_device_queue_family_properties(pdevice
, pCount
, NULL
);
1307 VkQueueFamilyProperties
*properties
[] = {
1308 pQueueFamilyProperties
+ 0,
1309 pQueueFamilyProperties
+ 1,
1310 pQueueFamilyProperties
+ 2,
1312 radv_get_physical_device_queue_family_properties(pdevice
, pCount
, properties
);
1313 assert(*pCount
<= 3);
1316 void radv_GetPhysicalDeviceQueueFamilyProperties2(
1317 VkPhysicalDevice physicalDevice
,
1319 VkQueueFamilyProperties2KHR
*pQueueFamilyProperties
)
1321 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
1322 if (!pQueueFamilyProperties
) {
1323 return radv_get_physical_device_queue_family_properties(pdevice
, pCount
, NULL
);
1326 VkQueueFamilyProperties
*properties
[] = {
1327 &pQueueFamilyProperties
[0].queueFamilyProperties
,
1328 &pQueueFamilyProperties
[1].queueFamilyProperties
,
1329 &pQueueFamilyProperties
[2].queueFamilyProperties
,
1331 radv_get_physical_device_queue_family_properties(pdevice
, pCount
, properties
);
1332 assert(*pCount
<= 3);
1335 void radv_GetPhysicalDeviceMemoryProperties(
1336 VkPhysicalDevice physicalDevice
,
1337 VkPhysicalDeviceMemoryProperties
*pMemoryProperties
)
1339 RADV_FROM_HANDLE(radv_physical_device
, physical_device
, physicalDevice
);
1341 *pMemoryProperties
= physical_device
->memory_properties
;
1344 void radv_GetPhysicalDeviceMemoryProperties2(
1345 VkPhysicalDevice physicalDevice
,
1346 VkPhysicalDeviceMemoryProperties2KHR
*pMemoryProperties
)
1348 return radv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
1349 &pMemoryProperties
->memoryProperties
);
1352 VkResult
radv_GetMemoryHostPointerPropertiesEXT(
1354 VkExternalMemoryHandleTypeFlagBitsKHR handleType
,
1355 const void *pHostPointer
,
1356 VkMemoryHostPointerPropertiesEXT
*pMemoryHostPointerProperties
)
1358 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1362 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
: {
1363 const struct radv_physical_device
*physical_device
= device
->physical_device
;
1364 uint32_t memoryTypeBits
= 0;
1365 for (int i
= 0; i
< physical_device
->memory_properties
.memoryTypeCount
; i
++) {
1366 if (physical_device
->mem_type_indices
[i
] == RADV_MEM_TYPE_GTT_CACHED
) {
1367 memoryTypeBits
= (1 << i
);
1371 pMemoryHostPointerProperties
->memoryTypeBits
= memoryTypeBits
;
1375 return VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
;
1379 static enum radeon_ctx_priority
1380 radv_get_queue_global_priority(const VkDeviceQueueGlobalPriorityCreateInfoEXT
*pObj
)
1382 /* Default to MEDIUM when a specific global priority isn't requested */
1384 return RADEON_CTX_PRIORITY_MEDIUM
;
1386 switch(pObj
->globalPriority
) {
1387 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT
:
1388 return RADEON_CTX_PRIORITY_REALTIME
;
1389 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT
:
1390 return RADEON_CTX_PRIORITY_HIGH
;
1391 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
:
1392 return RADEON_CTX_PRIORITY_MEDIUM
;
1393 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT
:
1394 return RADEON_CTX_PRIORITY_LOW
;
1396 unreachable("Illegal global priority value");
1397 return RADEON_CTX_PRIORITY_INVALID
;
1402 radv_queue_init(struct radv_device
*device
, struct radv_queue
*queue
,
1403 uint32_t queue_family_index
, int idx
,
1404 VkDeviceQueueCreateFlags flags
,
1405 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*global_priority
)
1407 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1408 queue
->device
= device
;
1409 queue
->queue_family_index
= queue_family_index
;
1410 queue
->queue_idx
= idx
;
1411 queue
->priority
= radv_get_queue_global_priority(global_priority
);
1412 queue
->flags
= flags
;
1414 queue
->hw_ctx
= device
->ws
->ctx_create(device
->ws
, queue
->priority
);
1416 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
1422 radv_queue_finish(struct radv_queue
*queue
)
1425 queue
->device
->ws
->ctx_destroy(queue
->hw_ctx
);
1427 if (queue
->initial_full_flush_preamble_cs
)
1428 queue
->device
->ws
->cs_destroy(queue
->initial_full_flush_preamble_cs
);
1429 if (queue
->initial_preamble_cs
)
1430 queue
->device
->ws
->cs_destroy(queue
->initial_preamble_cs
);
1431 if (queue
->continue_preamble_cs
)
1432 queue
->device
->ws
->cs_destroy(queue
->continue_preamble_cs
);
1433 if (queue
->descriptor_bo
)
1434 queue
->device
->ws
->buffer_destroy(queue
->descriptor_bo
);
1435 if (queue
->scratch_bo
)
1436 queue
->device
->ws
->buffer_destroy(queue
->scratch_bo
);
1437 if (queue
->esgs_ring_bo
)
1438 queue
->device
->ws
->buffer_destroy(queue
->esgs_ring_bo
);
1439 if (queue
->gsvs_ring_bo
)
1440 queue
->device
->ws
->buffer_destroy(queue
->gsvs_ring_bo
);
1441 if (queue
->tess_rings_bo
)
1442 queue
->device
->ws
->buffer_destroy(queue
->tess_rings_bo
);
1443 if (queue
->compute_scratch_bo
)
1444 queue
->device
->ws
->buffer_destroy(queue
->compute_scratch_bo
);
1448 radv_bo_list_init(struct radv_bo_list
*bo_list
)
1450 pthread_mutex_init(&bo_list
->mutex
, NULL
);
1451 bo_list
->list
.count
= bo_list
->capacity
= 0;
1452 bo_list
->list
.bos
= NULL
;
1456 radv_bo_list_finish(struct radv_bo_list
*bo_list
)
1458 free(bo_list
->list
.bos
);
1459 pthread_mutex_destroy(&bo_list
->mutex
);
1462 static VkResult
radv_bo_list_add(struct radv_device
*device
,
1463 struct radeon_winsys_bo
*bo
)
1465 struct radv_bo_list
*bo_list
= &device
->bo_list
;
1467 if (unlikely(!device
->use_global_bo_list
))
1470 pthread_mutex_lock(&bo_list
->mutex
);
1471 if (bo_list
->list
.count
== bo_list
->capacity
) {
1472 unsigned capacity
= MAX2(4, bo_list
->capacity
* 2);
1473 void *data
= realloc(bo_list
->list
.bos
, capacity
* sizeof(struct radeon_winsys_bo
*));
1476 pthread_mutex_unlock(&bo_list
->mutex
);
1477 return VK_ERROR_OUT_OF_HOST_MEMORY
;
1480 bo_list
->list
.bos
= (struct radeon_winsys_bo
**)data
;
1481 bo_list
->capacity
= capacity
;
1484 bo_list
->list
.bos
[bo_list
->list
.count
++] = bo
;
1485 pthread_mutex_unlock(&bo_list
->mutex
);
1489 static void radv_bo_list_remove(struct radv_device
*device
,
1490 struct radeon_winsys_bo
*bo
)
1492 struct radv_bo_list
*bo_list
= &device
->bo_list
;
1494 if (unlikely(!device
->use_global_bo_list
))
1497 pthread_mutex_lock(&bo_list
->mutex
);
1498 for(unsigned i
= 0; i
< bo_list
->list
.count
; ++i
) {
1499 if (bo_list
->list
.bos
[i
] == bo
) {
1500 bo_list
->list
.bos
[i
] = bo_list
->list
.bos
[bo_list
->list
.count
- 1];
1501 --bo_list
->list
.count
;
1505 pthread_mutex_unlock(&bo_list
->mutex
);
1509 radv_device_init_gs_info(struct radv_device
*device
)
1511 device
->gs_table_depth
= ac_get_gs_table_depth(device
->physical_device
->rad_info
.chip_class
,
1512 device
->physical_device
->rad_info
.family
);
1515 static int radv_get_device_extension_index(const char *name
)
1517 for (unsigned i
= 0; i
< RADV_DEVICE_EXTENSION_COUNT
; ++i
) {
1518 if (strcmp(name
, radv_device_extensions
[i
].extensionName
) == 0)
1525 radv_get_int_debug_option(const char *name
, int default_value
)
1532 result
= default_value
;
1536 result
= strtol(str
, &endptr
, 0);
1537 if (str
== endptr
) {
1538 /* No digits founs. */
1539 result
= default_value
;
1546 VkResult
radv_CreateDevice(
1547 VkPhysicalDevice physicalDevice
,
1548 const VkDeviceCreateInfo
* pCreateInfo
,
1549 const VkAllocationCallbacks
* pAllocator
,
1552 RADV_FROM_HANDLE(radv_physical_device
, physical_device
, physicalDevice
);
1554 struct radv_device
*device
;
1556 bool keep_shader_info
= false;
1558 /* Check enabled features */
1559 if (pCreateInfo
->pEnabledFeatures
) {
1560 VkPhysicalDeviceFeatures supported_features
;
1561 radv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
1562 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
1563 VkBool32
*enabled_feature
= (VkBool32
*)pCreateInfo
->pEnabledFeatures
;
1564 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
1565 for (uint32_t i
= 0; i
< num_features
; i
++) {
1566 if (enabled_feature
[i
] && !supported_feature
[i
])
1567 return vk_error(physical_device
->instance
, VK_ERROR_FEATURE_NOT_PRESENT
);
1571 device
= vk_zalloc2(&physical_device
->instance
->alloc
, pAllocator
,
1573 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
1575 return vk_error(physical_device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
1577 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1578 device
->instance
= physical_device
->instance
;
1579 device
->physical_device
= physical_device
;
1581 device
->ws
= physical_device
->ws
;
1583 device
->alloc
= *pAllocator
;
1585 device
->alloc
= physical_device
->instance
->alloc
;
1587 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
1588 const char *ext_name
= pCreateInfo
->ppEnabledExtensionNames
[i
];
1589 int index
= radv_get_device_extension_index(ext_name
);
1590 if (index
< 0 || !physical_device
->supported_extensions
.extensions
[index
]) {
1591 vk_free(&device
->alloc
, device
);
1592 return vk_error(physical_device
->instance
, VK_ERROR_EXTENSION_NOT_PRESENT
);
1595 device
->enabled_extensions
.extensions
[index
] = true;
1598 keep_shader_info
= device
->enabled_extensions
.AMD_shader_info
;
1600 /* With update after bind we can't attach bo's to the command buffer
1601 * from the descriptor set anymore, so we have to use a global BO list.
1603 device
->use_global_bo_list
=
1604 device
->enabled_extensions
.EXT_descriptor_indexing
;
1606 mtx_init(&device
->shader_slab_mutex
, mtx_plain
);
1607 list_inithead(&device
->shader_slabs
);
1609 radv_bo_list_init(&device
->bo_list
);
1611 for (unsigned i
= 0; i
< pCreateInfo
->queueCreateInfoCount
; i
++) {
1612 const VkDeviceQueueCreateInfo
*queue_create
= &pCreateInfo
->pQueueCreateInfos
[i
];
1613 uint32_t qfi
= queue_create
->queueFamilyIndex
;
1614 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*global_priority
=
1615 vk_find_struct_const(queue_create
->pNext
, DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT
);
1617 assert(!global_priority
|| device
->physical_device
->rad_info
.has_ctx_priority
);
1619 device
->queues
[qfi
] = vk_alloc(&device
->alloc
,
1620 queue_create
->queueCount
* sizeof(struct radv_queue
), 8, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
1621 if (!device
->queues
[qfi
]) {
1622 result
= VK_ERROR_OUT_OF_HOST_MEMORY
;
1626 memset(device
->queues
[qfi
], 0, queue_create
->queueCount
* sizeof(struct radv_queue
));
1628 device
->queue_count
[qfi
] = queue_create
->queueCount
;
1630 for (unsigned q
= 0; q
< queue_create
->queueCount
; q
++) {
1631 result
= radv_queue_init(device
, &device
->queues
[qfi
][q
],
1632 qfi
, q
, queue_create
->flags
,
1634 if (result
!= VK_SUCCESS
)
1639 device
->pbb_allowed
= device
->physical_device
->rad_info
.chip_class
>= GFX9
&&
1640 !(device
->instance
->debug_flags
& RADV_DEBUG_NOBINNING
);
1642 /* Disabled and not implemented for now. */
1643 device
->dfsm_allowed
= device
->pbb_allowed
&&
1644 (device
->physical_device
->rad_info
.family
== CHIP_RAVEN
||
1645 device
->physical_device
->rad_info
.family
== CHIP_RAVEN2
);
1648 device
->always_use_syncobj
= device
->physical_device
->rad_info
.has_syncobj_wait_for_submit
;
1651 /* The maximum number of scratch waves. Scratch space isn't divided
1652 * evenly between CUs. The number is only a function of the number of CUs.
1653 * We can decrease the constant to decrease the scratch buffer size.
1655 * sctx->scratch_waves must be >= the maximum possible size of
1656 * 1 threadgroup, so that the hw doesn't hang from being unable
1659 * The recommended value is 4 per CU at most. Higher numbers don't
1660 * bring much benefit, but they still occupy chip resources (think
1661 * async compute). I've seen ~2% performance difference between 4 and 32.
1663 uint32_t max_threads_per_block
= 2048;
1664 device
->scratch_waves
= MAX2(32 * physical_device
->rad_info
.num_good_compute_units
,
1665 max_threads_per_block
/ 64);
1667 device
->dispatch_initiator
= S_00B800_COMPUTE_SHADER_EN(1);
1669 if (device
->physical_device
->rad_info
.chip_class
>= CIK
) {
1670 /* If the KMD allows it (there is a KMD hw register for it),
1671 * allow launching waves out-of-order.
1673 device
->dispatch_initiator
|= S_00B800_ORDER_MODE(1);
1676 radv_device_init_gs_info(device
);
1678 device
->tess_offchip_block_dw_size
=
1679 device
->physical_device
->rad_info
.family
== CHIP_HAWAII
? 4096 : 8192;
1680 device
->has_distributed_tess
=
1681 device
->physical_device
->rad_info
.chip_class
>= VI
&&
1682 device
->physical_device
->rad_info
.max_se
>= 2;
1684 if (getenv("RADV_TRACE_FILE")) {
1685 const char *filename
= getenv("RADV_TRACE_FILE");
1687 keep_shader_info
= true;
1689 if (!radv_init_trace(device
))
1692 fprintf(stderr
, "*****************************************************************************\n");
1693 fprintf(stderr
, "* WARNING: RADV_TRACE_FILE is costly and should only be used for debugging! *\n");
1694 fprintf(stderr
, "*****************************************************************************\n");
1696 fprintf(stderr
, "Trace file will be dumped to %s\n", filename
);
1697 radv_dump_enabled_options(device
, stderr
);
1700 device
->keep_shader_info
= keep_shader_info
;
1702 result
= radv_device_init_meta(device
);
1703 if (result
!= VK_SUCCESS
)
1706 radv_device_init_msaa(device
);
1708 for (int family
= 0; family
< RADV_MAX_QUEUE_FAMILIES
; ++family
) {
1709 device
->empty_cs
[family
] = device
->ws
->cs_create(device
->ws
, family
);
1711 case RADV_QUEUE_GENERAL
:
1712 radeon_emit(device
->empty_cs
[family
], PKT3(PKT3_CONTEXT_CONTROL
, 1, 0));
1713 radeon_emit(device
->empty_cs
[family
], CONTEXT_CONTROL_LOAD_ENABLE(1));
1714 radeon_emit(device
->empty_cs
[family
], CONTEXT_CONTROL_SHADOW_ENABLE(1));
1716 case RADV_QUEUE_COMPUTE
:
1717 radeon_emit(device
->empty_cs
[family
], PKT3(PKT3_NOP
, 0, 0));
1718 radeon_emit(device
->empty_cs
[family
], 0);
1721 device
->ws
->cs_finalize(device
->empty_cs
[family
]);
1724 if (device
->physical_device
->rad_info
.chip_class
>= CIK
)
1725 cik_create_gfx_config(device
);
1727 VkPipelineCacheCreateInfo ci
;
1728 ci
.sType
= VK_STRUCTURE_TYPE_PIPELINE_CACHE_CREATE_INFO
;
1731 ci
.pInitialData
= NULL
;
1732 ci
.initialDataSize
= 0;
1734 result
= radv_CreatePipelineCache(radv_device_to_handle(device
),
1736 if (result
!= VK_SUCCESS
)
1739 device
->mem_cache
= radv_pipeline_cache_from_handle(pc
);
1741 device
->force_aniso
=
1742 MIN2(16, radv_get_int_debug_option("RADV_TEX_ANISO", -1));
1743 if (device
->force_aniso
>= 0) {
1744 fprintf(stderr
, "radv: Forcing anisotropy filter to %ix\n",
1745 1 << util_logbase2(device
->force_aniso
));
1748 *pDevice
= radv_device_to_handle(device
);
1752 radv_device_finish_meta(device
);
1754 radv_bo_list_finish(&device
->bo_list
);
1756 if (device
->trace_bo
)
1757 device
->ws
->buffer_destroy(device
->trace_bo
);
1759 if (device
->gfx_init
)
1760 device
->ws
->buffer_destroy(device
->gfx_init
);
1762 for (unsigned i
= 0; i
< RADV_MAX_QUEUE_FAMILIES
; i
++) {
1763 for (unsigned q
= 0; q
< device
->queue_count
[i
]; q
++)
1764 radv_queue_finish(&device
->queues
[i
][q
]);
1765 if (device
->queue_count
[i
])
1766 vk_free(&device
->alloc
, device
->queues
[i
]);
1769 vk_free(&device
->alloc
, device
);
1773 void radv_DestroyDevice(
1775 const VkAllocationCallbacks
* pAllocator
)
1777 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1782 if (device
->trace_bo
)
1783 device
->ws
->buffer_destroy(device
->trace_bo
);
1785 if (device
->gfx_init
)
1786 device
->ws
->buffer_destroy(device
->gfx_init
);
1788 for (unsigned i
= 0; i
< RADV_MAX_QUEUE_FAMILIES
; i
++) {
1789 for (unsigned q
= 0; q
< device
->queue_count
[i
]; q
++)
1790 radv_queue_finish(&device
->queues
[i
][q
]);
1791 if (device
->queue_count
[i
])
1792 vk_free(&device
->alloc
, device
->queues
[i
]);
1793 if (device
->empty_cs
[i
])
1794 device
->ws
->cs_destroy(device
->empty_cs
[i
]);
1796 radv_device_finish_meta(device
);
1798 VkPipelineCache pc
= radv_pipeline_cache_to_handle(device
->mem_cache
);
1799 radv_DestroyPipelineCache(radv_device_to_handle(device
), pc
, NULL
);
1801 radv_destroy_shader_slabs(device
);
1803 radv_bo_list_finish(&device
->bo_list
);
1804 vk_free(&device
->alloc
, device
);
1807 VkResult
radv_EnumerateInstanceLayerProperties(
1808 uint32_t* pPropertyCount
,
1809 VkLayerProperties
* pProperties
)
1811 if (pProperties
== NULL
) {
1812 *pPropertyCount
= 0;
1816 /* None supported at this time */
1817 return vk_error(NULL
, VK_ERROR_LAYER_NOT_PRESENT
);
1820 VkResult
radv_EnumerateDeviceLayerProperties(
1821 VkPhysicalDevice physicalDevice
,
1822 uint32_t* pPropertyCount
,
1823 VkLayerProperties
* pProperties
)
1825 if (pProperties
== NULL
) {
1826 *pPropertyCount
= 0;
1830 /* None supported at this time */
1831 return vk_error(NULL
, VK_ERROR_LAYER_NOT_PRESENT
);
1834 void radv_GetDeviceQueue2(
1836 const VkDeviceQueueInfo2
* pQueueInfo
,
1839 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1840 struct radv_queue
*queue
;
1842 queue
= &device
->queues
[pQueueInfo
->queueFamilyIndex
][pQueueInfo
->queueIndex
];
1843 if (pQueueInfo
->flags
!= queue
->flags
) {
1844 /* From the Vulkan 1.1.70 spec:
1846 * "The queue returned by vkGetDeviceQueue2 must have the same
1847 * flags value from this structure as that used at device
1848 * creation time in a VkDeviceQueueCreateInfo instance. If no
1849 * matching flags were specified at device creation time then
1850 * pQueue will return VK_NULL_HANDLE."
1852 *pQueue
= VK_NULL_HANDLE
;
1856 *pQueue
= radv_queue_to_handle(queue
);
1859 void radv_GetDeviceQueue(
1861 uint32_t queueFamilyIndex
,
1862 uint32_t queueIndex
,
1865 const VkDeviceQueueInfo2 info
= (VkDeviceQueueInfo2
) {
1866 .sType
= VK_STRUCTURE_TYPE_DEVICE_QUEUE_INFO_2
,
1867 .queueFamilyIndex
= queueFamilyIndex
,
1868 .queueIndex
= queueIndex
1871 radv_GetDeviceQueue2(_device
, &info
, pQueue
);
1875 fill_geom_tess_rings(struct radv_queue
*queue
,
1877 bool add_sample_positions
,
1878 uint32_t esgs_ring_size
,
1879 struct radeon_winsys_bo
*esgs_ring_bo
,
1880 uint32_t gsvs_ring_size
,
1881 struct radeon_winsys_bo
*gsvs_ring_bo
,
1882 uint32_t tess_factor_ring_size
,
1883 uint32_t tess_offchip_ring_offset
,
1884 uint32_t tess_offchip_ring_size
,
1885 struct radeon_winsys_bo
*tess_rings_bo
)
1887 uint64_t esgs_va
= 0, gsvs_va
= 0;
1888 uint64_t tess_va
= 0, tess_offchip_va
= 0;
1889 uint32_t *desc
= &map
[4];
1892 esgs_va
= radv_buffer_get_va(esgs_ring_bo
);
1894 gsvs_va
= radv_buffer_get_va(gsvs_ring_bo
);
1895 if (tess_rings_bo
) {
1896 tess_va
= radv_buffer_get_va(tess_rings_bo
);
1897 tess_offchip_va
= tess_va
+ tess_offchip_ring_offset
;
1900 /* stride 0, num records - size, add tid, swizzle, elsize4,
1903 desc
[1] = S_008F04_BASE_ADDRESS_HI(esgs_va
>> 32) |
1904 S_008F04_STRIDE(0) |
1905 S_008F04_SWIZZLE_ENABLE(true);
1906 desc
[2] = esgs_ring_size
;
1907 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1908 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1909 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1910 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1911 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1912 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1913 S_008F0C_ELEMENT_SIZE(1) |
1914 S_008F0C_INDEX_STRIDE(3) |
1915 S_008F0C_ADD_TID_ENABLE(true);
1918 /* GS entry for ES->GS ring */
1919 /* stride 0, num records - size, elsize0,
1922 desc
[1] = S_008F04_BASE_ADDRESS_HI(esgs_va
>> 32)|
1923 S_008F04_STRIDE(0) |
1924 S_008F04_SWIZZLE_ENABLE(false);
1925 desc
[2] = esgs_ring_size
;
1926 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1927 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1928 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1929 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1930 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1931 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1932 S_008F0C_ELEMENT_SIZE(0) |
1933 S_008F0C_INDEX_STRIDE(0) |
1934 S_008F0C_ADD_TID_ENABLE(false);
1937 /* VS entry for GS->VS ring */
1938 /* stride 0, num records - size, elsize0,
1941 desc
[1] = S_008F04_BASE_ADDRESS_HI(gsvs_va
>> 32)|
1942 S_008F04_STRIDE(0) |
1943 S_008F04_SWIZZLE_ENABLE(false);
1944 desc
[2] = gsvs_ring_size
;
1945 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1946 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1947 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1948 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1949 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1950 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1951 S_008F0C_ELEMENT_SIZE(0) |
1952 S_008F0C_INDEX_STRIDE(0) |
1953 S_008F0C_ADD_TID_ENABLE(false);
1956 /* stride gsvs_itemsize, num records 64
1957 elsize 4, index stride 16 */
1958 /* shader will patch stride and desc[2] */
1960 desc
[1] = S_008F04_BASE_ADDRESS_HI(gsvs_va
>> 32)|
1961 S_008F04_STRIDE(0) |
1962 S_008F04_SWIZZLE_ENABLE(true);
1964 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1965 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1966 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1967 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1968 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1969 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1970 S_008F0C_ELEMENT_SIZE(1) |
1971 S_008F0C_INDEX_STRIDE(1) |
1972 S_008F0C_ADD_TID_ENABLE(true);
1976 desc
[1] = S_008F04_BASE_ADDRESS_HI(tess_va
>> 32) |
1977 S_008F04_STRIDE(0) |
1978 S_008F04_SWIZZLE_ENABLE(false);
1979 desc
[2] = tess_factor_ring_size
;
1980 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1981 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1982 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1983 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1984 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1985 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1986 S_008F0C_ELEMENT_SIZE(0) |
1987 S_008F0C_INDEX_STRIDE(0) |
1988 S_008F0C_ADD_TID_ENABLE(false);
1991 desc
[0] = tess_offchip_va
;
1992 desc
[1] = S_008F04_BASE_ADDRESS_HI(tess_offchip_va
>> 32) |
1993 S_008F04_STRIDE(0) |
1994 S_008F04_SWIZZLE_ENABLE(false);
1995 desc
[2] = tess_offchip_ring_size
;
1996 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1997 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1998 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1999 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
2000 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
2001 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
2002 S_008F0C_ELEMENT_SIZE(0) |
2003 S_008F0C_INDEX_STRIDE(0) |
2004 S_008F0C_ADD_TID_ENABLE(false);
2007 /* add sample positions after all rings */
2008 memcpy(desc
, queue
->device
->sample_locations_1x
, 8);
2010 memcpy(desc
, queue
->device
->sample_locations_2x
, 16);
2012 memcpy(desc
, queue
->device
->sample_locations_4x
, 32);
2014 memcpy(desc
, queue
->device
->sample_locations_8x
, 64);
2016 memcpy(desc
, queue
->device
->sample_locations_16x
, 128);
2020 radv_get_hs_offchip_param(struct radv_device
*device
, uint32_t *max_offchip_buffers_p
)
2022 bool double_offchip_buffers
= device
->physical_device
->rad_info
.chip_class
>= CIK
&&
2023 device
->physical_device
->rad_info
.family
!= CHIP_CARRIZO
&&
2024 device
->physical_device
->rad_info
.family
!= CHIP_STONEY
;
2025 unsigned max_offchip_buffers_per_se
= double_offchip_buffers
? 128 : 64;
2026 unsigned max_offchip_buffers
;
2027 unsigned offchip_granularity
;
2028 unsigned hs_offchip_param
;
2032 * This must be one less than the maximum number due to a hw limitation.
2033 * Various hardware bugs in SI, CIK, and GFX9 need this.
2036 * Vega10 should limit max_offchip_buffers to 508 (4 * 127).
2037 * Gfx7 should limit max_offchip_buffers to 508
2038 * Gfx6 should limit max_offchip_buffers to 126 (2 * 63)
2040 * Follow AMDVLK here.
2042 if (device
->physical_device
->rad_info
.family
== CHIP_VEGA10
||
2043 device
->physical_device
->rad_info
.chip_class
== CIK
||
2044 device
->physical_device
->rad_info
.chip_class
== SI
)
2045 --max_offchip_buffers_per_se
;
2047 max_offchip_buffers
= max_offchip_buffers_per_se
*
2048 device
->physical_device
->rad_info
.max_se
;
2050 /* Hawaii has a bug with offchip buffers > 256 that can be worked
2051 * around by setting 4K granularity.
2053 if (device
->tess_offchip_block_dw_size
== 4096) {
2054 assert(device
->physical_device
->rad_info
.family
== CHIP_HAWAII
);
2055 offchip_granularity
= V_03093C_X_4K_DWORDS
;
2057 assert(device
->tess_offchip_block_dw_size
== 8192);
2058 offchip_granularity
= V_03093C_X_8K_DWORDS
;
2061 switch (device
->physical_device
->rad_info
.chip_class
) {
2063 max_offchip_buffers
= MIN2(max_offchip_buffers
, 126);
2069 max_offchip_buffers
= MIN2(max_offchip_buffers
, 508);
2073 *max_offchip_buffers_p
= max_offchip_buffers
;
2074 if (device
->physical_device
->rad_info
.chip_class
>= CIK
) {
2075 if (device
->physical_device
->rad_info
.chip_class
>= VI
)
2076 --max_offchip_buffers
;
2078 S_03093C_OFFCHIP_BUFFERING(max_offchip_buffers
) |
2079 S_03093C_OFFCHIP_GRANULARITY(offchip_granularity
);
2082 S_0089B0_OFFCHIP_BUFFERING(max_offchip_buffers
);
2084 return hs_offchip_param
;
2088 radv_emit_gs_ring_sizes(struct radv_queue
*queue
, struct radeon_cmdbuf
*cs
,
2089 struct radeon_winsys_bo
*esgs_ring_bo
,
2090 uint32_t esgs_ring_size
,
2091 struct radeon_winsys_bo
*gsvs_ring_bo
,
2092 uint32_t gsvs_ring_size
)
2094 if (!esgs_ring_bo
&& !gsvs_ring_bo
)
2098 radv_cs_add_buffer(queue
->device
->ws
, cs
, esgs_ring_bo
);
2101 radv_cs_add_buffer(queue
->device
->ws
, cs
, gsvs_ring_bo
);
2103 if (queue
->device
->physical_device
->rad_info
.chip_class
>= CIK
) {
2104 radeon_set_uconfig_reg_seq(cs
, R_030900_VGT_ESGS_RING_SIZE
, 2);
2105 radeon_emit(cs
, esgs_ring_size
>> 8);
2106 radeon_emit(cs
, gsvs_ring_size
>> 8);
2108 radeon_set_config_reg_seq(cs
, R_0088C8_VGT_ESGS_RING_SIZE
, 2);
2109 radeon_emit(cs
, esgs_ring_size
>> 8);
2110 radeon_emit(cs
, gsvs_ring_size
>> 8);
2115 radv_emit_tess_factor_ring(struct radv_queue
*queue
, struct radeon_cmdbuf
*cs
,
2116 unsigned hs_offchip_param
, unsigned tf_ring_size
,
2117 struct radeon_winsys_bo
*tess_rings_bo
)
2124 tf_va
= radv_buffer_get_va(tess_rings_bo
);
2126 radv_cs_add_buffer(queue
->device
->ws
, cs
, tess_rings_bo
);
2128 if (queue
->device
->physical_device
->rad_info
.chip_class
>= CIK
) {
2129 radeon_set_uconfig_reg(cs
, R_030938_VGT_TF_RING_SIZE
,
2130 S_030938_SIZE(tf_ring_size
/ 4));
2131 radeon_set_uconfig_reg(cs
, R_030940_VGT_TF_MEMORY_BASE
,
2133 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
2134 radeon_set_uconfig_reg(cs
, R_030944_VGT_TF_MEMORY_BASE_HI
,
2135 S_030944_BASE_HI(tf_va
>> 40));
2137 radeon_set_uconfig_reg(cs
, R_03093C_VGT_HS_OFFCHIP_PARAM
,
2140 radeon_set_config_reg(cs
, R_008988_VGT_TF_RING_SIZE
,
2141 S_008988_SIZE(tf_ring_size
/ 4));
2142 radeon_set_config_reg(cs
, R_0089B8_VGT_TF_MEMORY_BASE
,
2144 radeon_set_config_reg(cs
, R_0089B0_VGT_HS_OFFCHIP_PARAM
,
2150 radv_emit_compute_scratch(struct radv_queue
*queue
, struct radeon_cmdbuf
*cs
,
2151 struct radeon_winsys_bo
*compute_scratch_bo
)
2153 uint64_t scratch_va
;
2155 if (!compute_scratch_bo
)
2158 scratch_va
= radv_buffer_get_va(compute_scratch_bo
);
2160 radv_cs_add_buffer(queue
->device
->ws
, cs
, compute_scratch_bo
);
2162 radeon_set_sh_reg_seq(cs
, R_00B900_COMPUTE_USER_DATA_0
, 2);
2163 radeon_emit(cs
, scratch_va
);
2164 radeon_emit(cs
, S_008F04_BASE_ADDRESS_HI(scratch_va
>> 32) |
2165 S_008F04_SWIZZLE_ENABLE(1));
2169 radv_emit_global_shader_pointers(struct radv_queue
*queue
,
2170 struct radeon_cmdbuf
*cs
,
2171 struct radeon_winsys_bo
*descriptor_bo
)
2178 va
= radv_buffer_get_va(descriptor_bo
);
2180 radv_cs_add_buffer(queue
->device
->ws
, cs
, descriptor_bo
);
2182 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
2183 uint32_t regs
[] = {R_00B030_SPI_SHADER_USER_DATA_PS_0
,
2184 R_00B130_SPI_SHADER_USER_DATA_VS_0
,
2185 R_00B208_SPI_SHADER_USER_DATA_ADDR_LO_GS
,
2186 R_00B408_SPI_SHADER_USER_DATA_ADDR_LO_HS
};
2188 for (int i
= 0; i
< ARRAY_SIZE(regs
); ++i
) {
2189 radv_emit_shader_pointer(queue
->device
, cs
, regs
[i
],
2193 uint32_t regs
[] = {R_00B030_SPI_SHADER_USER_DATA_PS_0
,
2194 R_00B130_SPI_SHADER_USER_DATA_VS_0
,
2195 R_00B230_SPI_SHADER_USER_DATA_GS_0
,
2196 R_00B330_SPI_SHADER_USER_DATA_ES_0
,
2197 R_00B430_SPI_SHADER_USER_DATA_HS_0
,
2198 R_00B530_SPI_SHADER_USER_DATA_LS_0
};
2200 for (int i
= 0; i
< ARRAY_SIZE(regs
); ++i
) {
2201 radv_emit_shader_pointer(queue
->device
, cs
, regs
[i
],
2208 radv_init_graphics_state(struct radeon_cmdbuf
*cs
, struct radv_queue
*queue
)
2210 struct radv_device
*device
= queue
->device
;
2212 if (device
->gfx_init
) {
2213 uint64_t va
= radv_buffer_get_va(device
->gfx_init
);
2215 radeon_emit(cs
, PKT3(PKT3_INDIRECT_BUFFER_CIK
, 2, 0));
2216 radeon_emit(cs
, va
);
2217 radeon_emit(cs
, va
>> 32);
2218 radeon_emit(cs
, device
->gfx_init_size_dw
& 0xffff);
2220 radv_cs_add_buffer(device
->ws
, cs
, device
->gfx_init
);
2222 struct radv_physical_device
*physical_device
= device
->physical_device
;
2223 si_emit_graphics(physical_device
, cs
);
2228 radv_init_compute_state(struct radeon_cmdbuf
*cs
, struct radv_queue
*queue
)
2230 struct radv_physical_device
*physical_device
= queue
->device
->physical_device
;
2231 si_emit_compute(physical_device
, cs
);
2235 radv_get_preamble_cs(struct radv_queue
*queue
,
2236 uint32_t scratch_size
,
2237 uint32_t compute_scratch_size
,
2238 uint32_t esgs_ring_size
,
2239 uint32_t gsvs_ring_size
,
2240 bool needs_tess_rings
,
2241 bool needs_sample_positions
,
2242 struct radeon_cmdbuf
**initial_full_flush_preamble_cs
,
2243 struct radeon_cmdbuf
**initial_preamble_cs
,
2244 struct radeon_cmdbuf
**continue_preamble_cs
)
2246 struct radeon_winsys_bo
*scratch_bo
= NULL
;
2247 struct radeon_winsys_bo
*descriptor_bo
= NULL
;
2248 struct radeon_winsys_bo
*compute_scratch_bo
= NULL
;
2249 struct radeon_winsys_bo
*esgs_ring_bo
= NULL
;
2250 struct radeon_winsys_bo
*gsvs_ring_bo
= NULL
;
2251 struct radeon_winsys_bo
*tess_rings_bo
= NULL
;
2252 struct radeon_cmdbuf
*dest_cs
[3] = {0};
2253 bool add_tess_rings
= false, add_sample_positions
= false;
2254 unsigned tess_factor_ring_size
= 0, tess_offchip_ring_size
= 0;
2255 unsigned max_offchip_buffers
;
2256 unsigned hs_offchip_param
= 0;
2257 unsigned tess_offchip_ring_offset
;
2258 uint32_t ring_bo_flags
= RADEON_FLAG_NO_CPU_ACCESS
| RADEON_FLAG_NO_INTERPROCESS_SHARING
;
2259 if (!queue
->has_tess_rings
) {
2260 if (needs_tess_rings
)
2261 add_tess_rings
= true;
2263 if (!queue
->has_sample_positions
) {
2264 if (needs_sample_positions
)
2265 add_sample_positions
= true;
2267 tess_factor_ring_size
= 32768 * queue
->device
->physical_device
->rad_info
.max_se
;
2268 hs_offchip_param
= radv_get_hs_offchip_param(queue
->device
,
2269 &max_offchip_buffers
);
2270 tess_offchip_ring_offset
= align(tess_factor_ring_size
, 64 * 1024);
2271 tess_offchip_ring_size
= max_offchip_buffers
*
2272 queue
->device
->tess_offchip_block_dw_size
* 4;
2274 if (scratch_size
<= queue
->scratch_size
&&
2275 compute_scratch_size
<= queue
->compute_scratch_size
&&
2276 esgs_ring_size
<= queue
->esgs_ring_size
&&
2277 gsvs_ring_size
<= queue
->gsvs_ring_size
&&
2278 !add_tess_rings
&& !add_sample_positions
&&
2279 queue
->initial_preamble_cs
) {
2280 *initial_full_flush_preamble_cs
= queue
->initial_full_flush_preamble_cs
;
2281 *initial_preamble_cs
= queue
->initial_preamble_cs
;
2282 *continue_preamble_cs
= queue
->continue_preamble_cs
;
2283 if (!scratch_size
&& !compute_scratch_size
&& !esgs_ring_size
&& !gsvs_ring_size
)
2284 *continue_preamble_cs
= NULL
;
2288 if (scratch_size
> queue
->scratch_size
) {
2289 scratch_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
2297 scratch_bo
= queue
->scratch_bo
;
2299 if (compute_scratch_size
> queue
->compute_scratch_size
) {
2300 compute_scratch_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
2301 compute_scratch_size
,
2305 if (!compute_scratch_bo
)
2309 compute_scratch_bo
= queue
->compute_scratch_bo
;
2311 if (esgs_ring_size
> queue
->esgs_ring_size
) {
2312 esgs_ring_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
2320 esgs_ring_bo
= queue
->esgs_ring_bo
;
2321 esgs_ring_size
= queue
->esgs_ring_size
;
2324 if (gsvs_ring_size
> queue
->gsvs_ring_size
) {
2325 gsvs_ring_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
2333 gsvs_ring_bo
= queue
->gsvs_ring_bo
;
2334 gsvs_ring_size
= queue
->gsvs_ring_size
;
2337 if (add_tess_rings
) {
2338 tess_rings_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
2339 tess_offchip_ring_offset
+ tess_offchip_ring_size
,
2346 tess_rings_bo
= queue
->tess_rings_bo
;
2349 if (scratch_bo
!= queue
->scratch_bo
||
2350 esgs_ring_bo
!= queue
->esgs_ring_bo
||
2351 gsvs_ring_bo
!= queue
->gsvs_ring_bo
||
2352 tess_rings_bo
!= queue
->tess_rings_bo
||
2353 add_sample_positions
) {
2355 if (gsvs_ring_bo
|| esgs_ring_bo
||
2356 tess_rings_bo
|| add_sample_positions
) {
2357 size
= 112; /* 2 dword + 2 padding + 4 dword * 6 */
2358 if (add_sample_positions
)
2359 size
+= 256; /* 32+16+8+4+2+1 samples * 4 * 2 = 248 bytes. */
2361 else if (scratch_bo
)
2362 size
= 8; /* 2 dword */
2364 descriptor_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
2368 RADEON_FLAG_CPU_ACCESS
|
2369 RADEON_FLAG_NO_INTERPROCESS_SHARING
|
2370 RADEON_FLAG_READ_ONLY
);
2374 descriptor_bo
= queue
->descriptor_bo
;
2376 for(int i
= 0; i
< 3; ++i
) {
2377 struct radeon_cmdbuf
*cs
= NULL
;
2378 cs
= queue
->device
->ws
->cs_create(queue
->device
->ws
,
2379 queue
->queue_family_index
? RING_COMPUTE
: RING_GFX
);
2386 radv_cs_add_buffer(queue
->device
->ws
, cs
, scratch_bo
);
2388 /* Emit initial configuration. */
2389 switch (queue
->queue_family_index
) {
2390 case RADV_QUEUE_GENERAL
:
2391 radv_init_graphics_state(cs
, queue
);
2393 case RADV_QUEUE_COMPUTE
:
2394 radv_init_compute_state(cs
, queue
);
2396 case RADV_QUEUE_TRANSFER
:
2400 if (descriptor_bo
!= queue
->descriptor_bo
) {
2401 uint32_t *map
= (uint32_t*)queue
->device
->ws
->buffer_map(descriptor_bo
);
2404 uint64_t scratch_va
= radv_buffer_get_va(scratch_bo
);
2405 uint32_t rsrc1
= S_008F04_BASE_ADDRESS_HI(scratch_va
>> 32) |
2406 S_008F04_SWIZZLE_ENABLE(1);
2407 map
[0] = scratch_va
;
2411 if (esgs_ring_bo
|| gsvs_ring_bo
|| tess_rings_bo
||
2412 add_sample_positions
)
2413 fill_geom_tess_rings(queue
, map
, add_sample_positions
,
2414 esgs_ring_size
, esgs_ring_bo
,
2415 gsvs_ring_size
, gsvs_ring_bo
,
2416 tess_factor_ring_size
,
2417 tess_offchip_ring_offset
,
2418 tess_offchip_ring_size
,
2421 queue
->device
->ws
->buffer_unmap(descriptor_bo
);
2424 if (esgs_ring_bo
|| gsvs_ring_bo
|| tess_rings_bo
) {
2425 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
2426 radeon_emit(cs
, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
2427 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
2428 radeon_emit(cs
, EVENT_TYPE(V_028A90_VGT_FLUSH
) | EVENT_INDEX(0));
2431 radv_emit_gs_ring_sizes(queue
, cs
, esgs_ring_bo
, esgs_ring_size
,
2432 gsvs_ring_bo
, gsvs_ring_size
);
2433 radv_emit_tess_factor_ring(queue
, cs
, hs_offchip_param
,
2434 tess_factor_ring_size
, tess_rings_bo
);
2435 radv_emit_global_shader_pointers(queue
, cs
, descriptor_bo
);
2436 radv_emit_compute_scratch(queue
, cs
, compute_scratch_bo
);
2439 si_cs_emit_cache_flush(cs
,
2440 queue
->device
->physical_device
->rad_info
.chip_class
,
2442 queue
->queue_family_index
== RING_COMPUTE
&&
2443 queue
->device
->physical_device
->rad_info
.chip_class
>= CIK
,
2444 (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
)) |
2445 RADV_CMD_FLAG_INV_ICACHE
|
2446 RADV_CMD_FLAG_INV_SMEM_L1
|
2447 RADV_CMD_FLAG_INV_VMEM_L1
|
2448 RADV_CMD_FLAG_INV_GLOBAL_L2
|
2449 RADV_CMD_FLAG_START_PIPELINE_STATS
, 0);
2450 } else if (i
== 1) {
2451 si_cs_emit_cache_flush(cs
,
2452 queue
->device
->physical_device
->rad_info
.chip_class
,
2454 queue
->queue_family_index
== RING_COMPUTE
&&
2455 queue
->device
->physical_device
->rad_info
.chip_class
>= CIK
,
2456 RADV_CMD_FLAG_INV_ICACHE
|
2457 RADV_CMD_FLAG_INV_SMEM_L1
|
2458 RADV_CMD_FLAG_INV_VMEM_L1
|
2459 RADV_CMD_FLAG_INV_GLOBAL_L2
|
2460 RADV_CMD_FLAG_START_PIPELINE_STATS
, 0);
2463 if (!queue
->device
->ws
->cs_finalize(cs
))
2467 if (queue
->initial_full_flush_preamble_cs
)
2468 queue
->device
->ws
->cs_destroy(queue
->initial_full_flush_preamble_cs
);
2470 if (queue
->initial_preamble_cs
)
2471 queue
->device
->ws
->cs_destroy(queue
->initial_preamble_cs
);
2473 if (queue
->continue_preamble_cs
)
2474 queue
->device
->ws
->cs_destroy(queue
->continue_preamble_cs
);
2476 queue
->initial_full_flush_preamble_cs
= dest_cs
[0];
2477 queue
->initial_preamble_cs
= dest_cs
[1];
2478 queue
->continue_preamble_cs
= dest_cs
[2];
2480 if (scratch_bo
!= queue
->scratch_bo
) {
2481 if (queue
->scratch_bo
)
2482 queue
->device
->ws
->buffer_destroy(queue
->scratch_bo
);
2483 queue
->scratch_bo
= scratch_bo
;
2484 queue
->scratch_size
= scratch_size
;
2487 if (compute_scratch_bo
!= queue
->compute_scratch_bo
) {
2488 if (queue
->compute_scratch_bo
)
2489 queue
->device
->ws
->buffer_destroy(queue
->compute_scratch_bo
);
2490 queue
->compute_scratch_bo
= compute_scratch_bo
;
2491 queue
->compute_scratch_size
= compute_scratch_size
;
2494 if (esgs_ring_bo
!= queue
->esgs_ring_bo
) {
2495 if (queue
->esgs_ring_bo
)
2496 queue
->device
->ws
->buffer_destroy(queue
->esgs_ring_bo
);
2497 queue
->esgs_ring_bo
= esgs_ring_bo
;
2498 queue
->esgs_ring_size
= esgs_ring_size
;
2501 if (gsvs_ring_bo
!= queue
->gsvs_ring_bo
) {
2502 if (queue
->gsvs_ring_bo
)
2503 queue
->device
->ws
->buffer_destroy(queue
->gsvs_ring_bo
);
2504 queue
->gsvs_ring_bo
= gsvs_ring_bo
;
2505 queue
->gsvs_ring_size
= gsvs_ring_size
;
2508 if (tess_rings_bo
!= queue
->tess_rings_bo
) {
2509 queue
->tess_rings_bo
= tess_rings_bo
;
2510 queue
->has_tess_rings
= true;
2513 if (descriptor_bo
!= queue
->descriptor_bo
) {
2514 if (queue
->descriptor_bo
)
2515 queue
->device
->ws
->buffer_destroy(queue
->descriptor_bo
);
2517 queue
->descriptor_bo
= descriptor_bo
;
2520 if (add_sample_positions
)
2521 queue
->has_sample_positions
= true;
2523 *initial_full_flush_preamble_cs
= queue
->initial_full_flush_preamble_cs
;
2524 *initial_preamble_cs
= queue
->initial_preamble_cs
;
2525 *continue_preamble_cs
= queue
->continue_preamble_cs
;
2526 if (!scratch_size
&& !compute_scratch_size
&& !esgs_ring_size
&& !gsvs_ring_size
)
2527 *continue_preamble_cs
= NULL
;
2530 for (int i
= 0; i
< ARRAY_SIZE(dest_cs
); ++i
)
2532 queue
->device
->ws
->cs_destroy(dest_cs
[i
]);
2533 if (descriptor_bo
&& descriptor_bo
!= queue
->descriptor_bo
)
2534 queue
->device
->ws
->buffer_destroy(descriptor_bo
);
2535 if (scratch_bo
&& scratch_bo
!= queue
->scratch_bo
)
2536 queue
->device
->ws
->buffer_destroy(scratch_bo
);
2537 if (compute_scratch_bo
&& compute_scratch_bo
!= queue
->compute_scratch_bo
)
2538 queue
->device
->ws
->buffer_destroy(compute_scratch_bo
);
2539 if (esgs_ring_bo
&& esgs_ring_bo
!= queue
->esgs_ring_bo
)
2540 queue
->device
->ws
->buffer_destroy(esgs_ring_bo
);
2541 if (gsvs_ring_bo
&& gsvs_ring_bo
!= queue
->gsvs_ring_bo
)
2542 queue
->device
->ws
->buffer_destroy(gsvs_ring_bo
);
2543 if (tess_rings_bo
&& tess_rings_bo
!= queue
->tess_rings_bo
)
2544 queue
->device
->ws
->buffer_destroy(tess_rings_bo
);
2545 return vk_error(queue
->device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
2548 static VkResult
radv_alloc_sem_counts(struct radv_instance
*instance
,
2549 struct radv_winsys_sem_counts
*counts
,
2551 const VkSemaphore
*sems
,
2555 int syncobj_idx
= 0, sem_idx
= 0;
2557 if (num_sems
== 0 && _fence
== VK_NULL_HANDLE
)
2560 for (uint32_t i
= 0; i
< num_sems
; i
++) {
2561 RADV_FROM_HANDLE(radv_semaphore
, sem
, sems
[i
]);
2563 if (sem
->temp_syncobj
|| sem
->syncobj
)
2564 counts
->syncobj_count
++;
2566 counts
->sem_count
++;
2569 if (_fence
!= VK_NULL_HANDLE
) {
2570 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
2571 if (fence
->temp_syncobj
|| fence
->syncobj
)
2572 counts
->syncobj_count
++;
2575 if (counts
->syncobj_count
) {
2576 counts
->syncobj
= (uint32_t *)malloc(sizeof(uint32_t) * counts
->syncobj_count
);
2577 if (!counts
->syncobj
)
2578 return vk_error(instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
2581 if (counts
->sem_count
) {
2582 counts
->sem
= (struct radeon_winsys_sem
**)malloc(sizeof(struct radeon_winsys_sem
*) * counts
->sem_count
);
2584 free(counts
->syncobj
);
2585 return vk_error(instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
2589 for (uint32_t i
= 0; i
< num_sems
; i
++) {
2590 RADV_FROM_HANDLE(radv_semaphore
, sem
, sems
[i
]);
2592 if (sem
->temp_syncobj
) {
2593 counts
->syncobj
[syncobj_idx
++] = sem
->temp_syncobj
;
2595 else if (sem
->syncobj
)
2596 counts
->syncobj
[syncobj_idx
++] = sem
->syncobj
;
2599 counts
->sem
[sem_idx
++] = sem
->sem
;
2603 if (_fence
!= VK_NULL_HANDLE
) {
2604 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
2605 if (fence
->temp_syncobj
)
2606 counts
->syncobj
[syncobj_idx
++] = fence
->temp_syncobj
;
2607 else if (fence
->syncobj
)
2608 counts
->syncobj
[syncobj_idx
++] = fence
->syncobj
;
2615 radv_free_sem_info(struct radv_winsys_sem_info
*sem_info
)
2617 free(sem_info
->wait
.syncobj
);
2618 free(sem_info
->wait
.sem
);
2619 free(sem_info
->signal
.syncobj
);
2620 free(sem_info
->signal
.sem
);
2624 static void radv_free_temp_syncobjs(struct radv_device
*device
,
2626 const VkSemaphore
*sems
)
2628 for (uint32_t i
= 0; i
< num_sems
; i
++) {
2629 RADV_FROM_HANDLE(radv_semaphore
, sem
, sems
[i
]);
2631 if (sem
->temp_syncobj
) {
2632 device
->ws
->destroy_syncobj(device
->ws
, sem
->temp_syncobj
);
2633 sem
->temp_syncobj
= 0;
2639 radv_alloc_sem_info(struct radv_instance
*instance
,
2640 struct radv_winsys_sem_info
*sem_info
,
2642 const VkSemaphore
*wait_sems
,
2643 int num_signal_sems
,
2644 const VkSemaphore
*signal_sems
,
2648 memset(sem_info
, 0, sizeof(*sem_info
));
2650 ret
= radv_alloc_sem_counts(instance
, &sem_info
->wait
, num_wait_sems
, wait_sems
, VK_NULL_HANDLE
, true);
2653 ret
= radv_alloc_sem_counts(instance
, &sem_info
->signal
, num_signal_sems
, signal_sems
, fence
, false);
2655 radv_free_sem_info(sem_info
);
2657 /* caller can override these */
2658 sem_info
->cs_emit_wait
= true;
2659 sem_info
->cs_emit_signal
= true;
2663 /* Signals fence as soon as all the work currently put on queue is done. */
2664 static VkResult
radv_signal_fence(struct radv_queue
*queue
,
2665 struct radv_fence
*fence
)
2669 struct radv_winsys_sem_info sem_info
;
2671 result
= radv_alloc_sem_info(queue
->device
->instance
, &sem_info
, 0, NULL
, 0, NULL
,
2672 radv_fence_to_handle(fence
));
2673 if (result
!= VK_SUCCESS
)
2676 ret
= queue
->device
->ws
->cs_submit(queue
->hw_ctx
, queue
->queue_idx
,
2677 &queue
->device
->empty_cs
[queue
->queue_family_index
],
2678 1, NULL
, NULL
, &sem_info
, NULL
,
2679 false, fence
->fence
);
2680 radv_free_sem_info(&sem_info
);
2683 return vk_error(queue
->device
->instance
, VK_ERROR_DEVICE_LOST
);
2688 VkResult
radv_QueueSubmit(
2690 uint32_t submitCount
,
2691 const VkSubmitInfo
* pSubmits
,
2694 RADV_FROM_HANDLE(radv_queue
, queue
, _queue
);
2695 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
2696 struct radeon_winsys_fence
*base_fence
= fence
? fence
->fence
: NULL
;
2697 struct radeon_winsys_ctx
*ctx
= queue
->hw_ctx
;
2699 uint32_t max_cs_submission
= queue
->device
->trace_bo
? 1 : UINT32_MAX
;
2700 uint32_t scratch_size
= 0;
2701 uint32_t compute_scratch_size
= 0;
2702 uint32_t esgs_ring_size
= 0, gsvs_ring_size
= 0;
2703 struct radeon_cmdbuf
*initial_preamble_cs
= NULL
, *initial_flush_preamble_cs
= NULL
, *continue_preamble_cs
= NULL
;
2705 bool fence_emitted
= false;
2706 bool tess_rings_needed
= false;
2707 bool sample_positions_needed
= false;
2709 /* Do this first so failing to allocate scratch buffers can't result in
2710 * partially executed submissions. */
2711 for (uint32_t i
= 0; i
< submitCount
; i
++) {
2712 for (uint32_t j
= 0; j
< pSubmits
[i
].commandBufferCount
; j
++) {
2713 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
,
2714 pSubmits
[i
].pCommandBuffers
[j
]);
2716 scratch_size
= MAX2(scratch_size
, cmd_buffer
->scratch_size_needed
);
2717 compute_scratch_size
= MAX2(compute_scratch_size
,
2718 cmd_buffer
->compute_scratch_size_needed
);
2719 esgs_ring_size
= MAX2(esgs_ring_size
, cmd_buffer
->esgs_ring_size_needed
);
2720 gsvs_ring_size
= MAX2(gsvs_ring_size
, cmd_buffer
->gsvs_ring_size_needed
);
2721 tess_rings_needed
|= cmd_buffer
->tess_rings_needed
;
2722 sample_positions_needed
|= cmd_buffer
->sample_positions_needed
;
2726 result
= radv_get_preamble_cs(queue
, scratch_size
, compute_scratch_size
,
2727 esgs_ring_size
, gsvs_ring_size
, tess_rings_needed
,
2728 sample_positions_needed
, &initial_flush_preamble_cs
,
2729 &initial_preamble_cs
, &continue_preamble_cs
);
2730 if (result
!= VK_SUCCESS
)
2733 for (uint32_t i
= 0; i
< submitCount
; i
++) {
2734 struct radeon_cmdbuf
**cs_array
;
2735 bool do_flush
= !i
|| pSubmits
[i
].pWaitDstStageMask
;
2736 bool can_patch
= true;
2738 struct radv_winsys_sem_info sem_info
;
2740 result
= radv_alloc_sem_info(queue
->device
->instance
,
2742 pSubmits
[i
].waitSemaphoreCount
,
2743 pSubmits
[i
].pWaitSemaphores
,
2744 pSubmits
[i
].signalSemaphoreCount
,
2745 pSubmits
[i
].pSignalSemaphores
,
2747 if (result
!= VK_SUCCESS
)
2750 if (!pSubmits
[i
].commandBufferCount
) {
2751 if (pSubmits
[i
].waitSemaphoreCount
|| pSubmits
[i
].signalSemaphoreCount
) {
2752 ret
= queue
->device
->ws
->cs_submit(ctx
, queue
->queue_idx
,
2753 &queue
->device
->empty_cs
[queue
->queue_family_index
],
2758 radv_loge("failed to submit CS %d\n", i
);
2761 fence_emitted
= true;
2763 radv_free_sem_info(&sem_info
);
2767 cs_array
= malloc(sizeof(struct radeon_cmdbuf
*) *
2768 (pSubmits
[i
].commandBufferCount
));
2770 for (uint32_t j
= 0; j
< pSubmits
[i
].commandBufferCount
; j
++) {
2771 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
,
2772 pSubmits
[i
].pCommandBuffers
[j
]);
2773 assert(cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_PRIMARY
);
2775 cs_array
[j
] = cmd_buffer
->cs
;
2776 if ((cmd_buffer
->usage_flags
& VK_COMMAND_BUFFER_USAGE_SIMULTANEOUS_USE_BIT
))
2779 cmd_buffer
->status
= RADV_CMD_BUFFER_STATUS_PENDING
;
2782 for (uint32_t j
= 0; j
< pSubmits
[i
].commandBufferCount
; j
+= advance
) {
2783 struct radeon_cmdbuf
*initial_preamble
= (do_flush
&& !j
) ? initial_flush_preamble_cs
: initial_preamble_cs
;
2784 const struct radv_winsys_bo_list
*bo_list
= NULL
;
2786 advance
= MIN2(max_cs_submission
,
2787 pSubmits
[i
].commandBufferCount
- j
);
2789 if (queue
->device
->trace_bo
)
2790 *queue
->device
->trace_id_ptr
= 0;
2792 sem_info
.cs_emit_wait
= j
== 0;
2793 sem_info
.cs_emit_signal
= j
+ advance
== pSubmits
[i
].commandBufferCount
;
2795 if (unlikely(queue
->device
->use_global_bo_list
)) {
2796 pthread_mutex_lock(&queue
->device
->bo_list
.mutex
);
2797 bo_list
= &queue
->device
->bo_list
.list
;
2800 ret
= queue
->device
->ws
->cs_submit(ctx
, queue
->queue_idx
, cs_array
+ j
,
2801 advance
, initial_preamble
, continue_preamble_cs
,
2803 can_patch
, base_fence
);
2805 if (unlikely(queue
->device
->use_global_bo_list
))
2806 pthread_mutex_unlock(&queue
->device
->bo_list
.mutex
);
2809 radv_loge("failed to submit CS %d\n", i
);
2812 fence_emitted
= true;
2813 if (queue
->device
->trace_bo
) {
2814 radv_check_gpu_hangs(queue
, cs_array
[j
]);
2818 radv_free_temp_syncobjs(queue
->device
,
2819 pSubmits
[i
].waitSemaphoreCount
,
2820 pSubmits
[i
].pWaitSemaphores
);
2821 radv_free_sem_info(&sem_info
);
2826 if (!fence_emitted
) {
2827 result
= radv_signal_fence(queue
, fence
);
2828 if (result
!= VK_SUCCESS
)
2831 fence
->submitted
= true;
2837 VkResult
radv_QueueWaitIdle(
2840 RADV_FROM_HANDLE(radv_queue
, queue
, _queue
);
2842 queue
->device
->ws
->ctx_wait_idle(queue
->hw_ctx
,
2843 radv_queue_family_to_ring(queue
->queue_family_index
),
2848 VkResult
radv_DeviceWaitIdle(
2851 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2853 for (unsigned i
= 0; i
< RADV_MAX_QUEUE_FAMILIES
; i
++) {
2854 for (unsigned q
= 0; q
< device
->queue_count
[i
]; q
++) {
2855 radv_QueueWaitIdle(radv_queue_to_handle(&device
->queues
[i
][q
]));
2861 VkResult
radv_EnumerateInstanceExtensionProperties(
2862 const char* pLayerName
,
2863 uint32_t* pPropertyCount
,
2864 VkExtensionProperties
* pProperties
)
2866 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
2868 for (int i
= 0; i
< RADV_INSTANCE_EXTENSION_COUNT
; i
++) {
2869 if (radv_supported_instance_extensions
.extensions
[i
]) {
2870 vk_outarray_append(&out
, prop
) {
2871 *prop
= radv_instance_extensions
[i
];
2876 return vk_outarray_status(&out
);
2879 VkResult
radv_EnumerateDeviceExtensionProperties(
2880 VkPhysicalDevice physicalDevice
,
2881 const char* pLayerName
,
2882 uint32_t* pPropertyCount
,
2883 VkExtensionProperties
* pProperties
)
2885 RADV_FROM_HANDLE(radv_physical_device
, device
, physicalDevice
);
2886 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
2888 for (int i
= 0; i
< RADV_DEVICE_EXTENSION_COUNT
; i
++) {
2889 if (device
->supported_extensions
.extensions
[i
]) {
2890 vk_outarray_append(&out
, prop
) {
2891 *prop
= radv_device_extensions
[i
];
2896 return vk_outarray_status(&out
);
2899 PFN_vkVoidFunction
radv_GetInstanceProcAddr(
2900 VkInstance _instance
,
2903 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
2905 return radv_lookup_entrypoint_checked(pName
,
2906 instance
? instance
->apiVersion
: 0,
2907 instance
? &instance
->enabled_extensions
: NULL
,
2911 /* The loader wants us to expose a second GetInstanceProcAddr function
2912 * to work around certain LD_PRELOAD issues seen in apps.
2915 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
2916 VkInstance instance
,
2920 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
2921 VkInstance instance
,
2924 return radv_GetInstanceProcAddr(instance
, pName
);
2927 PFN_vkVoidFunction
radv_GetDeviceProcAddr(
2931 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2933 return radv_lookup_entrypoint_checked(pName
,
2934 device
->instance
->apiVersion
,
2935 &device
->instance
->enabled_extensions
,
2936 &device
->enabled_extensions
);
2939 bool radv_get_memory_fd(struct radv_device
*device
,
2940 struct radv_device_memory
*memory
,
2943 struct radeon_bo_metadata metadata
;
2945 if (memory
->image
) {
2946 radv_init_metadata(device
, memory
->image
, &metadata
);
2947 device
->ws
->buffer_set_metadata(memory
->bo
, &metadata
);
2950 return device
->ws
->buffer_get_fd(device
->ws
, memory
->bo
,
2954 static VkResult
radv_alloc_memory(struct radv_device
*device
,
2955 const VkMemoryAllocateInfo
* pAllocateInfo
,
2956 const VkAllocationCallbacks
* pAllocator
,
2957 VkDeviceMemory
* pMem
)
2959 struct radv_device_memory
*mem
;
2961 enum radeon_bo_domain domain
;
2963 enum radv_mem_type mem_type_index
= device
->physical_device
->mem_type_indices
[pAllocateInfo
->memoryTypeIndex
];
2965 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
2967 if (pAllocateInfo
->allocationSize
== 0) {
2968 /* Apparently, this is allowed */
2969 *pMem
= VK_NULL_HANDLE
;
2973 const VkImportMemoryFdInfoKHR
*import_info
=
2974 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_FD_INFO_KHR
);
2975 const VkMemoryDedicatedAllocateInfoKHR
*dedicate_info
=
2976 vk_find_struct_const(pAllocateInfo
->pNext
, MEMORY_DEDICATED_ALLOCATE_INFO_KHR
);
2977 const VkExportMemoryAllocateInfoKHR
*export_info
=
2978 vk_find_struct_const(pAllocateInfo
->pNext
, EXPORT_MEMORY_ALLOCATE_INFO_KHR
);
2979 const VkImportMemoryHostPointerInfoEXT
*host_ptr_info
=
2980 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_HOST_POINTER_INFO_EXT
);
2982 const struct wsi_memory_allocate_info
*wsi_info
=
2983 vk_find_struct_const(pAllocateInfo
->pNext
, WSI_MEMORY_ALLOCATE_INFO_MESA
);
2985 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
2986 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2988 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
2990 if (wsi_info
&& wsi_info
->implicit_sync
)
2991 flags
|= RADEON_FLAG_IMPLICIT_SYNC
;
2993 if (dedicate_info
) {
2994 mem
->image
= radv_image_from_handle(dedicate_info
->image
);
2995 mem
->buffer
= radv_buffer_from_handle(dedicate_info
->buffer
);
3001 mem
->user_ptr
= NULL
;
3004 assert(import_info
->handleType
==
3005 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
||
3006 import_info
->handleType
==
3007 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
3008 mem
->bo
= device
->ws
->buffer_from_fd(device
->ws
, import_info
->fd
,
3011 result
= VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
;
3014 close(import_info
->fd
);
3016 } else if (host_ptr_info
) {
3017 assert(host_ptr_info
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
);
3018 assert(mem_type_index
== RADV_MEM_TYPE_GTT_CACHED
);
3019 mem
->bo
= device
->ws
->buffer_from_ptr(device
->ws
, host_ptr_info
->pHostPointer
,
3020 pAllocateInfo
->allocationSize
);
3022 result
= VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
;
3025 mem
->user_ptr
= host_ptr_info
->pHostPointer
;
3028 uint64_t alloc_size
= align_u64(pAllocateInfo
->allocationSize
, 4096);
3029 if (mem_type_index
== RADV_MEM_TYPE_GTT_WRITE_COMBINE
||
3030 mem_type_index
== RADV_MEM_TYPE_GTT_CACHED
)
3031 domain
= RADEON_DOMAIN_GTT
;
3033 domain
= RADEON_DOMAIN_VRAM
;
3035 if (mem_type_index
== RADV_MEM_TYPE_VRAM
)
3036 flags
|= RADEON_FLAG_NO_CPU_ACCESS
;
3038 flags
|= RADEON_FLAG_CPU_ACCESS
;
3040 if (mem_type_index
== RADV_MEM_TYPE_GTT_WRITE_COMBINE
)
3041 flags
|= RADEON_FLAG_GTT_WC
;
3043 if (!dedicate_info
&& !import_info
&& (!export_info
|| !export_info
->handleTypes
))
3044 flags
|= RADEON_FLAG_NO_INTERPROCESS_SHARING
;
3046 mem
->bo
= device
->ws
->buffer_create(device
->ws
, alloc_size
, device
->physical_device
->rad_info
.max_alignment
,
3050 result
= VK_ERROR_OUT_OF_DEVICE_MEMORY
;
3053 mem
->type_index
= mem_type_index
;
3056 result
= radv_bo_list_add(device
, mem
->bo
);
3057 if (result
!= VK_SUCCESS
)
3060 *pMem
= radv_device_memory_to_handle(mem
);
3065 device
->ws
->buffer_destroy(mem
->bo
);
3067 vk_free2(&device
->alloc
, pAllocator
, mem
);
3072 VkResult
radv_AllocateMemory(
3074 const VkMemoryAllocateInfo
* pAllocateInfo
,
3075 const VkAllocationCallbacks
* pAllocator
,
3076 VkDeviceMemory
* pMem
)
3078 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3079 return radv_alloc_memory(device
, pAllocateInfo
, pAllocator
, pMem
);
3082 void radv_FreeMemory(
3084 VkDeviceMemory _mem
,
3085 const VkAllocationCallbacks
* pAllocator
)
3087 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3088 RADV_FROM_HANDLE(radv_device_memory
, mem
, _mem
);
3093 radv_bo_list_remove(device
, mem
->bo
);
3094 device
->ws
->buffer_destroy(mem
->bo
);
3097 vk_free2(&device
->alloc
, pAllocator
, mem
);
3100 VkResult
radv_MapMemory(
3102 VkDeviceMemory _memory
,
3103 VkDeviceSize offset
,
3105 VkMemoryMapFlags flags
,
3108 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3109 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
3117 *ppData
= mem
->user_ptr
;
3119 *ppData
= device
->ws
->buffer_map(mem
->bo
);
3126 return vk_error(device
->instance
, VK_ERROR_MEMORY_MAP_FAILED
);
3129 void radv_UnmapMemory(
3131 VkDeviceMemory _memory
)
3133 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3134 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
3139 if (mem
->user_ptr
== NULL
)
3140 device
->ws
->buffer_unmap(mem
->bo
);
3143 VkResult
radv_FlushMappedMemoryRanges(
3145 uint32_t memoryRangeCount
,
3146 const VkMappedMemoryRange
* pMemoryRanges
)
3151 VkResult
radv_InvalidateMappedMemoryRanges(
3153 uint32_t memoryRangeCount
,
3154 const VkMappedMemoryRange
* pMemoryRanges
)
3159 void radv_GetBufferMemoryRequirements(
3162 VkMemoryRequirements
* pMemoryRequirements
)
3164 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3165 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
3167 pMemoryRequirements
->memoryTypeBits
= (1u << device
->physical_device
->memory_properties
.memoryTypeCount
) - 1;
3169 if (buffer
->flags
& VK_BUFFER_CREATE_SPARSE_BINDING_BIT
)
3170 pMemoryRequirements
->alignment
= 4096;
3172 pMemoryRequirements
->alignment
= 16;
3174 pMemoryRequirements
->size
= align64(buffer
->size
, pMemoryRequirements
->alignment
);
3177 void radv_GetBufferMemoryRequirements2(
3179 const VkBufferMemoryRequirementsInfo2KHR
* pInfo
,
3180 VkMemoryRequirements2KHR
* pMemoryRequirements
)
3182 radv_GetBufferMemoryRequirements(device
, pInfo
->buffer
,
3183 &pMemoryRequirements
->memoryRequirements
);
3184 RADV_FROM_HANDLE(radv_buffer
, buffer
, pInfo
->buffer
);
3185 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3186 switch (ext
->sType
) {
3187 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS_KHR
: {
3188 VkMemoryDedicatedRequirementsKHR
*req
=
3189 (VkMemoryDedicatedRequirementsKHR
*) ext
;
3190 req
->requiresDedicatedAllocation
= buffer
->shareable
;
3191 req
->prefersDedicatedAllocation
= req
->requiresDedicatedAllocation
;
3200 void radv_GetImageMemoryRequirements(
3203 VkMemoryRequirements
* pMemoryRequirements
)
3205 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3206 RADV_FROM_HANDLE(radv_image
, image
, _image
);
3208 pMemoryRequirements
->memoryTypeBits
= (1u << device
->physical_device
->memory_properties
.memoryTypeCount
) - 1;
3210 pMemoryRequirements
->size
= image
->size
;
3211 pMemoryRequirements
->alignment
= image
->alignment
;
3214 void radv_GetImageMemoryRequirements2(
3216 const VkImageMemoryRequirementsInfo2KHR
* pInfo
,
3217 VkMemoryRequirements2KHR
* pMemoryRequirements
)
3219 radv_GetImageMemoryRequirements(device
, pInfo
->image
,
3220 &pMemoryRequirements
->memoryRequirements
);
3222 RADV_FROM_HANDLE(radv_image
, image
, pInfo
->image
);
3224 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3225 switch (ext
->sType
) {
3226 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS_KHR
: {
3227 VkMemoryDedicatedRequirementsKHR
*req
=
3228 (VkMemoryDedicatedRequirementsKHR
*) ext
;
3229 req
->requiresDedicatedAllocation
= image
->shareable
;
3230 req
->prefersDedicatedAllocation
= req
->requiresDedicatedAllocation
;
3239 void radv_GetImageSparseMemoryRequirements(
3242 uint32_t* pSparseMemoryRequirementCount
,
3243 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
3248 void radv_GetImageSparseMemoryRequirements2(
3250 const VkImageSparseMemoryRequirementsInfo2KHR
* pInfo
,
3251 uint32_t* pSparseMemoryRequirementCount
,
3252 VkSparseImageMemoryRequirements2KHR
* pSparseMemoryRequirements
)
3257 void radv_GetDeviceMemoryCommitment(
3259 VkDeviceMemory memory
,
3260 VkDeviceSize
* pCommittedMemoryInBytes
)
3262 *pCommittedMemoryInBytes
= 0;
3265 VkResult
radv_BindBufferMemory2(VkDevice device
,
3266 uint32_t bindInfoCount
,
3267 const VkBindBufferMemoryInfoKHR
*pBindInfos
)
3269 for (uint32_t i
= 0; i
< bindInfoCount
; ++i
) {
3270 RADV_FROM_HANDLE(radv_device_memory
, mem
, pBindInfos
[i
].memory
);
3271 RADV_FROM_HANDLE(radv_buffer
, buffer
, pBindInfos
[i
].buffer
);
3274 buffer
->bo
= mem
->bo
;
3275 buffer
->offset
= pBindInfos
[i
].memoryOffset
;
3283 VkResult
radv_BindBufferMemory(
3286 VkDeviceMemory memory
,
3287 VkDeviceSize memoryOffset
)
3289 const VkBindBufferMemoryInfoKHR info
= {
3290 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO_KHR
,
3293 .memoryOffset
= memoryOffset
3296 return radv_BindBufferMemory2(device
, 1, &info
);
3299 VkResult
radv_BindImageMemory2(VkDevice device
,
3300 uint32_t bindInfoCount
,
3301 const VkBindImageMemoryInfoKHR
*pBindInfos
)
3303 for (uint32_t i
= 0; i
< bindInfoCount
; ++i
) {
3304 RADV_FROM_HANDLE(radv_device_memory
, mem
, pBindInfos
[i
].memory
);
3305 RADV_FROM_HANDLE(radv_image
, image
, pBindInfos
[i
].image
);
3308 image
->bo
= mem
->bo
;
3309 image
->offset
= pBindInfos
[i
].memoryOffset
;
3319 VkResult
radv_BindImageMemory(
3322 VkDeviceMemory memory
,
3323 VkDeviceSize memoryOffset
)
3325 const VkBindImageMemoryInfoKHR info
= {
3326 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO_KHR
,
3329 .memoryOffset
= memoryOffset
3332 return radv_BindImageMemory2(device
, 1, &info
);
3337 radv_sparse_buffer_bind_memory(struct radv_device
*device
,
3338 const VkSparseBufferMemoryBindInfo
*bind
)
3340 RADV_FROM_HANDLE(radv_buffer
, buffer
, bind
->buffer
);
3342 for (uint32_t i
= 0; i
< bind
->bindCount
; ++i
) {
3343 struct radv_device_memory
*mem
= NULL
;
3345 if (bind
->pBinds
[i
].memory
!= VK_NULL_HANDLE
)
3346 mem
= radv_device_memory_from_handle(bind
->pBinds
[i
].memory
);
3348 device
->ws
->buffer_virtual_bind(buffer
->bo
,
3349 bind
->pBinds
[i
].resourceOffset
,
3350 bind
->pBinds
[i
].size
,
3351 mem
? mem
->bo
: NULL
,
3352 bind
->pBinds
[i
].memoryOffset
);
3357 radv_sparse_image_opaque_bind_memory(struct radv_device
*device
,
3358 const VkSparseImageOpaqueMemoryBindInfo
*bind
)
3360 RADV_FROM_HANDLE(radv_image
, image
, bind
->image
);
3362 for (uint32_t i
= 0; i
< bind
->bindCount
; ++i
) {
3363 struct radv_device_memory
*mem
= NULL
;
3365 if (bind
->pBinds
[i
].memory
!= VK_NULL_HANDLE
)
3366 mem
= radv_device_memory_from_handle(bind
->pBinds
[i
].memory
);
3368 device
->ws
->buffer_virtual_bind(image
->bo
,
3369 bind
->pBinds
[i
].resourceOffset
,
3370 bind
->pBinds
[i
].size
,
3371 mem
? mem
->bo
: NULL
,
3372 bind
->pBinds
[i
].memoryOffset
);
3376 VkResult
radv_QueueBindSparse(
3378 uint32_t bindInfoCount
,
3379 const VkBindSparseInfo
* pBindInfo
,
3382 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
3383 RADV_FROM_HANDLE(radv_queue
, queue
, _queue
);
3384 struct radeon_winsys_fence
*base_fence
= fence
? fence
->fence
: NULL
;
3385 bool fence_emitted
= false;
3389 for (uint32_t i
= 0; i
< bindInfoCount
; ++i
) {
3390 struct radv_winsys_sem_info sem_info
;
3391 for (uint32_t j
= 0; j
< pBindInfo
[i
].bufferBindCount
; ++j
) {
3392 radv_sparse_buffer_bind_memory(queue
->device
,
3393 pBindInfo
[i
].pBufferBinds
+ j
);
3396 for (uint32_t j
= 0; j
< pBindInfo
[i
].imageOpaqueBindCount
; ++j
) {
3397 radv_sparse_image_opaque_bind_memory(queue
->device
,
3398 pBindInfo
[i
].pImageOpaqueBinds
+ j
);
3402 result
= radv_alloc_sem_info(queue
->device
->instance
,
3404 pBindInfo
[i
].waitSemaphoreCount
,
3405 pBindInfo
[i
].pWaitSemaphores
,
3406 pBindInfo
[i
].signalSemaphoreCount
,
3407 pBindInfo
[i
].pSignalSemaphores
,
3409 if (result
!= VK_SUCCESS
)
3412 if (pBindInfo
[i
].waitSemaphoreCount
|| pBindInfo
[i
].signalSemaphoreCount
) {
3413 ret
= queue
->device
->ws
->cs_submit(queue
->hw_ctx
, queue
->queue_idx
,
3414 &queue
->device
->empty_cs
[queue
->queue_family_index
],
3419 radv_loge("failed to submit CS %d\n", i
);
3423 fence_emitted
= true;
3425 fence
->submitted
= true;
3428 radv_free_sem_info(&sem_info
);
3433 if (!fence_emitted
) {
3434 result
= radv_signal_fence(queue
, fence
);
3435 if (result
!= VK_SUCCESS
)
3438 fence
->submitted
= true;
3444 VkResult
radv_CreateFence(
3446 const VkFenceCreateInfo
* pCreateInfo
,
3447 const VkAllocationCallbacks
* pAllocator
,
3450 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3451 const VkExportFenceCreateInfoKHR
*export
=
3452 vk_find_struct_const(pCreateInfo
->pNext
, EXPORT_FENCE_CREATE_INFO_KHR
);
3453 VkExternalFenceHandleTypeFlagsKHR handleTypes
=
3454 export
? export
->handleTypes
: 0;
3456 struct radv_fence
*fence
= vk_alloc2(&device
->alloc
, pAllocator
,
3458 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3461 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3463 fence
->fence_wsi
= NULL
;
3464 fence
->submitted
= false;
3465 fence
->signalled
= !!(pCreateInfo
->flags
& VK_FENCE_CREATE_SIGNALED_BIT
);
3466 fence
->temp_syncobj
= 0;
3467 if (device
->always_use_syncobj
|| handleTypes
) {
3468 int ret
= device
->ws
->create_syncobj(device
->ws
, &fence
->syncobj
);
3470 vk_free2(&device
->alloc
, pAllocator
, fence
);
3471 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3473 if (pCreateInfo
->flags
& VK_FENCE_CREATE_SIGNALED_BIT
) {
3474 device
->ws
->signal_syncobj(device
->ws
, fence
->syncobj
);
3476 fence
->fence
= NULL
;
3478 fence
->fence
= device
->ws
->create_fence();
3479 if (!fence
->fence
) {
3480 vk_free2(&device
->alloc
, pAllocator
, fence
);
3481 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3486 *pFence
= radv_fence_to_handle(fence
);
3491 void radv_DestroyFence(
3494 const VkAllocationCallbacks
* pAllocator
)
3496 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3497 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
3502 if (fence
->temp_syncobj
)
3503 device
->ws
->destroy_syncobj(device
->ws
, fence
->temp_syncobj
);
3505 device
->ws
->destroy_syncobj(device
->ws
, fence
->syncobj
);
3507 device
->ws
->destroy_fence(fence
->fence
);
3508 if (fence
->fence_wsi
)
3509 fence
->fence_wsi
->destroy(fence
->fence_wsi
);
3510 vk_free2(&device
->alloc
, pAllocator
, fence
);
3514 static uint64_t radv_get_current_time()
3517 clock_gettime(CLOCK_MONOTONIC
, &tv
);
3518 return tv
.tv_nsec
+ tv
.tv_sec
*1000000000ull;
3521 static uint64_t radv_get_absolute_timeout(uint64_t timeout
)
3523 uint64_t current_time
= radv_get_current_time();
3525 timeout
= MIN2(UINT64_MAX
- current_time
, timeout
);
3527 return current_time
+ timeout
;
3531 static bool radv_all_fences_plain_and_submitted(uint32_t fenceCount
, const VkFence
*pFences
)
3533 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
3534 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
3535 if (fence
->fence
== NULL
|| fence
->syncobj
||
3536 fence
->temp_syncobj
||
3537 (!fence
->signalled
&& !fence
->submitted
))
3543 static bool radv_all_fences_syncobj(uint32_t fenceCount
, const VkFence
*pFences
)
3545 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
3546 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
3547 if (fence
->syncobj
== 0 && fence
->temp_syncobj
== 0)
3553 VkResult
radv_WaitForFences(
3555 uint32_t fenceCount
,
3556 const VkFence
* pFences
,
3560 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3561 timeout
= radv_get_absolute_timeout(timeout
);
3563 if (device
->always_use_syncobj
&&
3564 radv_all_fences_syncobj(fenceCount
, pFences
))
3566 uint32_t *handles
= malloc(sizeof(uint32_t) * fenceCount
);
3568 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3570 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
3571 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
3572 handles
[i
] = fence
->temp_syncobj
? fence
->temp_syncobj
: fence
->syncobj
;
3575 bool success
= device
->ws
->wait_syncobj(device
->ws
, handles
, fenceCount
, waitAll
, timeout
);
3578 return success
? VK_SUCCESS
: VK_TIMEOUT
;
3581 if (!waitAll
&& fenceCount
> 1) {
3582 /* Not doing this by default for waitAll, due to needing to allocate twice. */
3583 if (device
->physical_device
->rad_info
.drm_minor
>= 10 && radv_all_fences_plain_and_submitted(fenceCount
, pFences
)) {
3584 uint32_t wait_count
= 0;
3585 struct radeon_winsys_fence
**fences
= malloc(sizeof(struct radeon_winsys_fence
*) * fenceCount
);
3587 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3589 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
3590 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
3592 if (fence
->signalled
) {
3597 fences
[wait_count
++] = fence
->fence
;
3600 bool success
= device
->ws
->fences_wait(device
->ws
, fences
, wait_count
,
3601 waitAll
, timeout
- radv_get_current_time());
3604 return success
? VK_SUCCESS
: VK_TIMEOUT
;
3607 while(radv_get_current_time() <= timeout
) {
3608 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
3609 if (radv_GetFenceStatus(_device
, pFences
[i
]) == VK_SUCCESS
)
3616 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
3617 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
3618 bool expired
= false;
3620 if (fence
->temp_syncobj
) {
3621 if (!device
->ws
->wait_syncobj(device
->ws
, &fence
->temp_syncobj
, 1, true, timeout
))
3626 if (fence
->syncobj
) {
3627 if (!device
->ws
->wait_syncobj(device
->ws
, &fence
->syncobj
, 1, true, timeout
))
3632 if (fence
->signalled
)
3636 if (!fence
->submitted
) {
3637 while(radv_get_current_time() <= timeout
&&
3641 if (!fence
->submitted
)
3644 /* Recheck as it may have been set by
3645 * submitting operations. */
3647 if (fence
->signalled
)
3651 expired
= device
->ws
->fence_wait(device
->ws
,
3658 if (fence
->fence_wsi
) {
3659 VkResult result
= fence
->fence_wsi
->wait(fence
->fence_wsi
, timeout
);
3660 if (result
!= VK_SUCCESS
)
3664 fence
->signalled
= true;
3670 VkResult
radv_ResetFences(VkDevice _device
,
3671 uint32_t fenceCount
,
3672 const VkFence
*pFences
)
3674 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3676 for (unsigned i
= 0; i
< fenceCount
; ++i
) {
3677 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
3678 fence
->submitted
= fence
->signalled
= false;
3680 /* Per spec, we first restore the permanent payload, and then reset, so
3681 * having a temp syncobj should not skip resetting the permanent syncobj. */
3682 if (fence
->temp_syncobj
) {
3683 device
->ws
->destroy_syncobj(device
->ws
, fence
->temp_syncobj
);
3684 fence
->temp_syncobj
= 0;
3687 if (fence
->syncobj
) {
3688 device
->ws
->reset_syncobj(device
->ws
, fence
->syncobj
);
3695 VkResult
radv_GetFenceStatus(VkDevice _device
, VkFence _fence
)
3697 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3698 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
3700 if (fence
->temp_syncobj
) {
3701 bool success
= device
->ws
->wait_syncobj(device
->ws
, &fence
->temp_syncobj
, 1, true, 0);
3702 return success
? VK_SUCCESS
: VK_NOT_READY
;
3705 if (fence
->syncobj
) {
3706 bool success
= device
->ws
->wait_syncobj(device
->ws
, &fence
->syncobj
, 1, true, 0);
3707 return success
? VK_SUCCESS
: VK_NOT_READY
;
3710 if (fence
->signalled
)
3712 if (!fence
->submitted
)
3713 return VK_NOT_READY
;
3715 if (!device
->ws
->fence_wait(device
->ws
, fence
->fence
, false, 0))
3716 return VK_NOT_READY
;
3718 if (fence
->fence_wsi
) {
3719 VkResult result
= fence
->fence_wsi
->wait(fence
->fence_wsi
, 0);
3721 if (result
!= VK_SUCCESS
) {
3722 if (result
== VK_TIMEOUT
)
3723 return VK_NOT_READY
;
3731 // Queue semaphore functions
3733 VkResult
radv_CreateSemaphore(
3735 const VkSemaphoreCreateInfo
* pCreateInfo
,
3736 const VkAllocationCallbacks
* pAllocator
,
3737 VkSemaphore
* pSemaphore
)
3739 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3740 const VkExportSemaphoreCreateInfoKHR
*export
=
3741 vk_find_struct_const(pCreateInfo
->pNext
, EXPORT_SEMAPHORE_CREATE_INFO_KHR
);
3742 VkExternalSemaphoreHandleTypeFlagsKHR handleTypes
=
3743 export
? export
->handleTypes
: 0;
3745 struct radv_semaphore
*sem
= vk_alloc2(&device
->alloc
, pAllocator
,
3747 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3749 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3751 sem
->temp_syncobj
= 0;
3752 /* create a syncobject if we are going to export this semaphore */
3753 if (device
->always_use_syncobj
|| handleTypes
) {
3754 assert (device
->physical_device
->rad_info
.has_syncobj
);
3755 int ret
= device
->ws
->create_syncobj(device
->ws
, &sem
->syncobj
);
3757 vk_free2(&device
->alloc
, pAllocator
, sem
);
3758 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3762 sem
->sem
= device
->ws
->create_sem(device
->ws
);
3764 vk_free2(&device
->alloc
, pAllocator
, sem
);
3765 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3770 *pSemaphore
= radv_semaphore_to_handle(sem
);
3774 void radv_DestroySemaphore(
3776 VkSemaphore _semaphore
,
3777 const VkAllocationCallbacks
* pAllocator
)
3779 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3780 RADV_FROM_HANDLE(radv_semaphore
, sem
, _semaphore
);
3785 device
->ws
->destroy_syncobj(device
->ws
, sem
->syncobj
);
3787 device
->ws
->destroy_sem(sem
->sem
);
3788 vk_free2(&device
->alloc
, pAllocator
, sem
);
3791 VkResult
radv_CreateEvent(
3793 const VkEventCreateInfo
* pCreateInfo
,
3794 const VkAllocationCallbacks
* pAllocator
,
3797 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3798 struct radv_event
*event
= vk_alloc2(&device
->alloc
, pAllocator
,
3800 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3803 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3805 event
->bo
= device
->ws
->buffer_create(device
->ws
, 8, 8,
3807 RADEON_FLAG_VA_UNCACHED
| RADEON_FLAG_CPU_ACCESS
| RADEON_FLAG_NO_INTERPROCESS_SHARING
);
3809 vk_free2(&device
->alloc
, pAllocator
, event
);
3810 return vk_error(device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
3813 event
->map
= (uint64_t*)device
->ws
->buffer_map(event
->bo
);
3815 *pEvent
= radv_event_to_handle(event
);
3820 void radv_DestroyEvent(
3823 const VkAllocationCallbacks
* pAllocator
)
3825 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3826 RADV_FROM_HANDLE(radv_event
, event
, _event
);
3830 device
->ws
->buffer_destroy(event
->bo
);
3831 vk_free2(&device
->alloc
, pAllocator
, event
);
3834 VkResult
radv_GetEventStatus(
3838 RADV_FROM_HANDLE(radv_event
, event
, _event
);
3840 if (*event
->map
== 1)
3841 return VK_EVENT_SET
;
3842 return VK_EVENT_RESET
;
3845 VkResult
radv_SetEvent(
3849 RADV_FROM_HANDLE(radv_event
, event
, _event
);
3855 VkResult
radv_ResetEvent(
3859 RADV_FROM_HANDLE(radv_event
, event
, _event
);
3865 VkResult
radv_CreateBuffer(
3867 const VkBufferCreateInfo
* pCreateInfo
,
3868 const VkAllocationCallbacks
* pAllocator
,
3871 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3872 struct radv_buffer
*buffer
;
3874 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
3876 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
3877 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3879 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3881 buffer
->size
= pCreateInfo
->size
;
3882 buffer
->usage
= pCreateInfo
->usage
;
3885 buffer
->flags
= pCreateInfo
->flags
;
3887 buffer
->shareable
= vk_find_struct_const(pCreateInfo
->pNext
,
3888 EXTERNAL_MEMORY_BUFFER_CREATE_INFO_KHR
) != NULL
;
3890 if (pCreateInfo
->flags
& VK_BUFFER_CREATE_SPARSE_BINDING_BIT
) {
3891 buffer
->bo
= device
->ws
->buffer_create(device
->ws
,
3892 align64(buffer
->size
, 4096),
3893 4096, 0, RADEON_FLAG_VIRTUAL
);
3895 vk_free2(&device
->alloc
, pAllocator
, buffer
);
3896 return vk_error(device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
3900 *pBuffer
= radv_buffer_to_handle(buffer
);
3905 void radv_DestroyBuffer(
3908 const VkAllocationCallbacks
* pAllocator
)
3910 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3911 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
3916 if (buffer
->flags
& VK_BUFFER_CREATE_SPARSE_BINDING_BIT
)
3917 device
->ws
->buffer_destroy(buffer
->bo
);
3919 vk_free2(&device
->alloc
, pAllocator
, buffer
);
3922 static inline unsigned
3923 si_tile_mode_index(const struct radv_image
*image
, unsigned level
, bool stencil
)
3926 return image
->surface
.u
.legacy
.stencil_tiling_index
[level
];
3928 return image
->surface
.u
.legacy
.tiling_index
[level
];
3931 static uint32_t radv_surface_max_layer_count(struct radv_image_view
*iview
)
3933 return iview
->type
== VK_IMAGE_VIEW_TYPE_3D
? iview
->extent
.depth
: (iview
->base_layer
+ iview
->layer_count
);
3937 radv_init_dcc_control_reg(struct radv_device
*device
,
3938 struct radv_image_view
*iview
)
3940 unsigned max_uncompressed_block_size
= V_028C78_MAX_BLOCK_SIZE_256B
;
3941 unsigned min_compressed_block_size
= V_028C78_MIN_BLOCK_SIZE_32B
;
3942 unsigned max_compressed_block_size
;
3943 unsigned independent_64b_blocks
;
3945 if (!radv_image_has_dcc(iview
->image
))
3948 if (iview
->image
->info
.samples
> 1) {
3949 if (iview
->image
->surface
.bpe
== 1)
3950 max_uncompressed_block_size
= V_028C78_MAX_BLOCK_SIZE_64B
;
3951 else if (iview
->image
->surface
.bpe
== 2)
3952 max_uncompressed_block_size
= V_028C78_MAX_BLOCK_SIZE_128B
;
3955 if (!device
->physical_device
->rad_info
.has_dedicated_vram
) {
3956 /* amdvlk: [min-compressed-block-size] should be set to 32 for
3957 * dGPU and 64 for APU because all of our APUs to date use
3958 * DIMMs which have a request granularity size of 64B while all
3959 * other chips have a 32B request size.
3961 min_compressed_block_size
= V_028C78_MIN_BLOCK_SIZE_64B
;
3964 if (iview
->image
->usage
& (VK_IMAGE_USAGE_SAMPLED_BIT
|
3965 VK_IMAGE_USAGE_TRANSFER_SRC_BIT
|
3966 VK_IMAGE_USAGE_INPUT_ATTACHMENT_BIT
)) {
3967 /* If this DCC image is potentially going to be used in texture
3968 * fetches, we need some special settings.
3970 independent_64b_blocks
= 1;
3971 max_compressed_block_size
= V_028C78_MAX_BLOCK_SIZE_64B
;
3973 /* MAX_UNCOMPRESSED_BLOCK_SIZE must be >=
3974 * MAX_COMPRESSED_BLOCK_SIZE. Set MAX_COMPRESSED_BLOCK_SIZE as
3975 * big as possible for better compression state.
3977 independent_64b_blocks
= 0;
3978 max_compressed_block_size
= max_uncompressed_block_size
;
3981 return S_028C78_MAX_UNCOMPRESSED_BLOCK_SIZE(max_uncompressed_block_size
) |
3982 S_028C78_MAX_COMPRESSED_BLOCK_SIZE(max_compressed_block_size
) |
3983 S_028C78_MIN_COMPRESSED_BLOCK_SIZE(min_compressed_block_size
) |
3984 S_028C78_INDEPENDENT_64B_BLOCKS(independent_64b_blocks
);
3988 radv_initialise_color_surface(struct radv_device
*device
,
3989 struct radv_color_buffer_info
*cb
,
3990 struct radv_image_view
*iview
)
3992 const struct vk_format_description
*desc
;
3993 unsigned ntype
, format
, swap
, endian
;
3994 unsigned blend_clamp
= 0, blend_bypass
= 0;
3996 const struct radeon_surf
*surf
= &iview
->image
->surface
;
3998 desc
= vk_format_description(iview
->vk_format
);
4000 memset(cb
, 0, sizeof(*cb
));
4002 /* Intensity is implemented as Red, so treat it that way. */
4003 cb
->cb_color_attrib
= S_028C74_FORCE_DST_ALPHA_1(desc
->swizzle
[3] == VK_SWIZZLE_1
);
4005 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
;
4007 cb
->cb_color_base
= va
>> 8;
4009 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
4010 struct gfx9_surf_meta_flags meta
;
4011 if (iview
->image
->dcc_offset
)
4012 meta
= iview
->image
->surface
.u
.gfx9
.dcc
;
4014 meta
= iview
->image
->surface
.u
.gfx9
.cmask
;
4016 cb
->cb_color_attrib
|= S_028C74_COLOR_SW_MODE(iview
->image
->surface
.u
.gfx9
.surf
.swizzle_mode
) |
4017 S_028C74_FMASK_SW_MODE(iview
->image
->surface
.u
.gfx9
.fmask
.swizzle_mode
) |
4018 S_028C74_RB_ALIGNED(meta
.rb_aligned
) |
4019 S_028C74_PIPE_ALIGNED(meta
.pipe_aligned
);
4021 cb
->cb_color_base
+= iview
->image
->surface
.u
.gfx9
.surf_offset
>> 8;
4022 cb
->cb_color_base
|= iview
->image
->surface
.tile_swizzle
;
4024 const struct legacy_surf_level
*level_info
= &surf
->u
.legacy
.level
[iview
->base_mip
];
4025 unsigned pitch_tile_max
, slice_tile_max
, tile_mode_index
;
4027 cb
->cb_color_base
+= level_info
->offset
>> 8;
4028 if (level_info
->mode
== RADEON_SURF_MODE_2D
)
4029 cb
->cb_color_base
|= iview
->image
->surface
.tile_swizzle
;
4031 pitch_tile_max
= level_info
->nblk_x
/ 8 - 1;
4032 slice_tile_max
= (level_info
->nblk_x
* level_info
->nblk_y
) / 64 - 1;
4033 tile_mode_index
= si_tile_mode_index(iview
->image
, iview
->base_mip
, false);
4035 cb
->cb_color_pitch
= S_028C64_TILE_MAX(pitch_tile_max
);
4036 cb
->cb_color_slice
= S_028C68_TILE_MAX(slice_tile_max
);
4037 cb
->cb_color_cmask_slice
= iview
->image
->cmask
.slice_tile_max
;
4039 cb
->cb_color_attrib
|= S_028C74_TILE_MODE_INDEX(tile_mode_index
);
4041 if (radv_image_has_fmask(iview
->image
)) {
4042 if (device
->physical_device
->rad_info
.chip_class
>= CIK
)
4043 cb
->cb_color_pitch
|= S_028C64_FMASK_TILE_MAX(iview
->image
->fmask
.pitch_in_pixels
/ 8 - 1);
4044 cb
->cb_color_attrib
|= S_028C74_FMASK_TILE_MODE_INDEX(iview
->image
->fmask
.tile_mode_index
);
4045 cb
->cb_color_fmask_slice
= S_028C88_TILE_MAX(iview
->image
->fmask
.slice_tile_max
);
4047 /* This must be set for fast clear to work without FMASK. */
4048 if (device
->physical_device
->rad_info
.chip_class
>= CIK
)
4049 cb
->cb_color_pitch
|= S_028C64_FMASK_TILE_MAX(pitch_tile_max
);
4050 cb
->cb_color_attrib
|= S_028C74_FMASK_TILE_MODE_INDEX(tile_mode_index
);
4051 cb
->cb_color_fmask_slice
= S_028C88_TILE_MAX(slice_tile_max
);
4055 /* CMASK variables */
4056 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
;
4057 va
+= iview
->image
->cmask
.offset
;
4058 cb
->cb_color_cmask
= va
>> 8;
4060 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
;
4061 va
+= iview
->image
->dcc_offset
;
4062 cb
->cb_dcc_base
= va
>> 8;
4063 cb
->cb_dcc_base
|= iview
->image
->surface
.tile_swizzle
;
4065 uint32_t max_slice
= radv_surface_max_layer_count(iview
) - 1;
4066 cb
->cb_color_view
= S_028C6C_SLICE_START(iview
->base_layer
) |
4067 S_028C6C_SLICE_MAX(max_slice
);
4069 if (iview
->image
->info
.samples
> 1) {
4070 unsigned log_samples
= util_logbase2(iview
->image
->info
.samples
);
4072 cb
->cb_color_attrib
|= S_028C74_NUM_SAMPLES(log_samples
) |
4073 S_028C74_NUM_FRAGMENTS(log_samples
);
4076 if (radv_image_has_fmask(iview
->image
)) {
4077 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
+ iview
->image
->fmask
.offset
;
4078 cb
->cb_color_fmask
= va
>> 8;
4079 cb
->cb_color_fmask
|= iview
->image
->fmask
.tile_swizzle
;
4081 cb
->cb_color_fmask
= cb
->cb_color_base
;
4084 ntype
= radv_translate_color_numformat(iview
->vk_format
,
4086 vk_format_get_first_non_void_channel(iview
->vk_format
));
4087 format
= radv_translate_colorformat(iview
->vk_format
);
4088 if (format
== V_028C70_COLOR_INVALID
|| ntype
== ~0u)
4089 radv_finishme("Illegal color\n");
4090 swap
= radv_translate_colorswap(iview
->vk_format
, FALSE
);
4091 endian
= radv_colorformat_endian_swap(format
);
4093 /* blend clamp should be set for all NORM/SRGB types */
4094 if (ntype
== V_028C70_NUMBER_UNORM
||
4095 ntype
== V_028C70_NUMBER_SNORM
||
4096 ntype
== V_028C70_NUMBER_SRGB
)
4099 /* set blend bypass according to docs if SINT/UINT or
4100 8/24 COLOR variants */
4101 if (ntype
== V_028C70_NUMBER_UINT
|| ntype
== V_028C70_NUMBER_SINT
||
4102 format
== V_028C70_COLOR_8_24
|| format
== V_028C70_COLOR_24_8
||
4103 format
== V_028C70_COLOR_X24_8_32_FLOAT
) {
4108 if ((ntype
== V_028C70_NUMBER_UINT
|| ntype
== V_028C70_NUMBER_SINT
) &&
4109 (format
== V_028C70_COLOR_8
||
4110 format
== V_028C70_COLOR_8_8
||
4111 format
== V_028C70_COLOR_8_8_8_8
))
4112 ->color_is_int8
= true;
4114 cb
->cb_color_info
= S_028C70_FORMAT(format
) |
4115 S_028C70_COMP_SWAP(swap
) |
4116 S_028C70_BLEND_CLAMP(blend_clamp
) |
4117 S_028C70_BLEND_BYPASS(blend_bypass
) |
4118 S_028C70_SIMPLE_FLOAT(1) |
4119 S_028C70_ROUND_MODE(ntype
!= V_028C70_NUMBER_UNORM
&&
4120 ntype
!= V_028C70_NUMBER_SNORM
&&
4121 ntype
!= V_028C70_NUMBER_SRGB
&&
4122 format
!= V_028C70_COLOR_8_24
&&
4123 format
!= V_028C70_COLOR_24_8
) |
4124 S_028C70_NUMBER_TYPE(ntype
) |
4125 S_028C70_ENDIAN(endian
);
4126 if (radv_image_has_fmask(iview
->image
)) {
4127 cb
->cb_color_info
|= S_028C70_COMPRESSION(1);
4128 if (device
->physical_device
->rad_info
.chip_class
== SI
) {
4129 unsigned fmask_bankh
= util_logbase2(iview
->image
->fmask
.bank_height
);
4130 cb
->cb_color_attrib
|= S_028C74_FMASK_BANK_HEIGHT(fmask_bankh
);
4134 if (radv_image_has_cmask(iview
->image
) &&
4135 !(device
->instance
->debug_flags
& RADV_DEBUG_NO_FAST_CLEARS
))
4136 cb
->cb_color_info
|= S_028C70_FAST_CLEAR(1);
4138 if (radv_dcc_enabled(iview
->image
, iview
->base_mip
))
4139 cb
->cb_color_info
|= S_028C70_DCC_ENABLE(1);
4141 cb
->cb_dcc_control
= radv_init_dcc_control_reg(device
, iview
);
4143 /* This must be set for fast clear to work without FMASK. */
4144 if (!radv_image_has_fmask(iview
->image
) &&
4145 device
->physical_device
->rad_info
.chip_class
== SI
) {
4146 unsigned bankh
= util_logbase2(iview
->image
->surface
.u
.legacy
.bankh
);
4147 cb
->cb_color_attrib
|= S_028C74_FMASK_BANK_HEIGHT(bankh
);
4150 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
4151 unsigned mip0_depth
= iview
->image
->type
== VK_IMAGE_TYPE_3D
?
4152 (iview
->extent
.depth
- 1) : (iview
->image
->info
.array_size
- 1);
4154 cb
->cb_color_view
|= S_028C6C_MIP_LEVEL(iview
->base_mip
);
4155 cb
->cb_color_attrib
|= S_028C74_MIP0_DEPTH(mip0_depth
) |
4156 S_028C74_RESOURCE_TYPE(iview
->image
->surface
.u
.gfx9
.resource_type
);
4157 cb
->cb_color_attrib2
= S_028C68_MIP0_WIDTH(iview
->extent
.width
- 1) |
4158 S_028C68_MIP0_HEIGHT(iview
->extent
.height
- 1) |
4159 S_028C68_MAX_MIP(iview
->image
->info
.levels
- 1);
4164 radv_calc_decompress_on_z_planes(struct radv_device
*device
,
4165 struct radv_image_view
*iview
)
4167 unsigned max_zplanes
= 0;
4169 assert(radv_image_is_tc_compat_htile(iview
->image
));
4171 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
4172 /* Default value for 32-bit depth surfaces. */
4175 if (iview
->vk_format
== VK_FORMAT_D16_UNORM
&&
4176 iview
->image
->info
.samples
> 1)
4179 max_zplanes
= max_zplanes
+ 1;
4181 if (iview
->vk_format
== VK_FORMAT_D16_UNORM
) {
4182 /* Do not enable Z plane compression for 16-bit depth
4183 * surfaces because isn't supported on GFX8. Only
4184 * 32-bit depth surfaces are supported by the hardware.
4185 * This allows to maintain shader compatibility and to
4186 * reduce the number of depth decompressions.
4190 if (iview
->image
->info
.samples
<= 1)
4192 else if (iview
->image
->info
.samples
<= 4)
4203 radv_initialise_ds_surface(struct radv_device
*device
,
4204 struct radv_ds_buffer_info
*ds
,
4205 struct radv_image_view
*iview
)
4207 unsigned level
= iview
->base_mip
;
4208 unsigned format
, stencil_format
;
4209 uint64_t va
, s_offs
, z_offs
;
4210 bool stencil_only
= false;
4211 memset(ds
, 0, sizeof(*ds
));
4212 switch (iview
->image
->vk_format
) {
4213 case VK_FORMAT_D24_UNORM_S8_UINT
:
4214 case VK_FORMAT_X8_D24_UNORM_PACK32
:
4215 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-24);
4216 ds
->offset_scale
= 2.0f
;
4218 case VK_FORMAT_D16_UNORM
:
4219 case VK_FORMAT_D16_UNORM_S8_UINT
:
4220 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-16);
4221 ds
->offset_scale
= 4.0f
;
4223 case VK_FORMAT_D32_SFLOAT
:
4224 case VK_FORMAT_D32_SFLOAT_S8_UINT
:
4225 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-23) |
4226 S_028B78_POLY_OFFSET_DB_IS_FLOAT_FMT(1);
4227 ds
->offset_scale
= 1.0f
;
4229 case VK_FORMAT_S8_UINT
:
4230 stencil_only
= true;
4236 format
= radv_translate_dbformat(iview
->image
->vk_format
);
4237 stencil_format
= iview
->image
->surface
.has_stencil
?
4238 V_028044_STENCIL_8
: V_028044_STENCIL_INVALID
;
4240 uint32_t max_slice
= radv_surface_max_layer_count(iview
) - 1;
4241 ds
->db_depth_view
= S_028008_SLICE_START(iview
->base_layer
) |
4242 S_028008_SLICE_MAX(max_slice
);
4244 ds
->db_htile_data_base
= 0;
4245 ds
->db_htile_surface
= 0;
4247 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
;
4248 s_offs
= z_offs
= va
;
4250 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
4251 assert(iview
->image
->surface
.u
.gfx9
.surf_offset
== 0);
4252 s_offs
+= iview
->image
->surface
.u
.gfx9
.stencil_offset
;
4254 ds
->db_z_info
= S_028038_FORMAT(format
) |
4255 S_028038_NUM_SAMPLES(util_logbase2(iview
->image
->info
.samples
)) |
4256 S_028038_SW_MODE(iview
->image
->surface
.u
.gfx9
.surf
.swizzle_mode
) |
4257 S_028038_MAXMIP(iview
->image
->info
.levels
- 1) |
4258 S_028038_ZRANGE_PRECISION(1);
4259 ds
->db_stencil_info
= S_02803C_FORMAT(stencil_format
) |
4260 S_02803C_SW_MODE(iview
->image
->surface
.u
.gfx9
.stencil
.swizzle_mode
);
4262 ds
->db_z_info2
= S_028068_EPITCH(iview
->image
->surface
.u
.gfx9
.surf
.epitch
);
4263 ds
->db_stencil_info2
= S_02806C_EPITCH(iview
->image
->surface
.u
.gfx9
.stencil
.epitch
);
4264 ds
->db_depth_view
|= S_028008_MIPID(level
);
4266 ds
->db_depth_size
= S_02801C_X_MAX(iview
->image
->info
.width
- 1) |
4267 S_02801C_Y_MAX(iview
->image
->info
.height
- 1);
4269 if (radv_htile_enabled(iview
->image
, level
)) {
4270 ds
->db_z_info
|= S_028038_TILE_SURFACE_ENABLE(1);
4272 if (radv_image_is_tc_compat_htile(iview
->image
)) {
4273 unsigned max_zplanes
=
4274 radv_calc_decompress_on_z_planes(device
, iview
);
4276 ds
->db_z_info
|= S_028038_DECOMPRESS_ON_N_ZPLANES(max_zplanes
) |
4277 S_028038_ITERATE_FLUSH(1);
4278 ds
->db_stencil_info
|= S_02803C_ITERATE_FLUSH(1);
4281 if (!iview
->image
->surface
.has_stencil
)
4282 /* Use all of the htile_buffer for depth if there's no stencil. */
4283 ds
->db_stencil_info
|= S_02803C_TILE_STENCIL_DISABLE(1);
4284 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
+
4285 iview
->image
->htile_offset
;
4286 ds
->db_htile_data_base
= va
>> 8;
4287 ds
->db_htile_surface
= S_028ABC_FULL_CACHE(1) |
4288 S_028ABC_PIPE_ALIGNED(iview
->image
->surface
.u
.gfx9
.htile
.pipe_aligned
) |
4289 S_028ABC_RB_ALIGNED(iview
->image
->surface
.u
.gfx9
.htile
.rb_aligned
);
4292 const struct legacy_surf_level
*level_info
= &iview
->image
->surface
.u
.legacy
.level
[level
];
4295 level_info
= &iview
->image
->surface
.u
.legacy
.stencil_level
[level
];
4297 z_offs
+= iview
->image
->surface
.u
.legacy
.level
[level
].offset
;
4298 s_offs
+= iview
->image
->surface
.u
.legacy
.stencil_level
[level
].offset
;
4300 ds
->db_depth_info
= S_02803C_ADDR5_SWIZZLE_MASK(!radv_image_is_tc_compat_htile(iview
->image
));
4301 ds
->db_z_info
= S_028040_FORMAT(format
) | S_028040_ZRANGE_PRECISION(1);
4302 ds
->db_stencil_info
= S_028044_FORMAT(stencil_format
);
4304 if (iview
->image
->info
.samples
> 1)
4305 ds
->db_z_info
|= S_028040_NUM_SAMPLES(util_logbase2(iview
->image
->info
.samples
));
4307 if (device
->physical_device
->rad_info
.chip_class
>= CIK
) {
4308 struct radeon_info
*info
= &device
->physical_device
->rad_info
;
4309 unsigned tiling_index
= iview
->image
->surface
.u
.legacy
.tiling_index
[level
];
4310 unsigned stencil_index
= iview
->image
->surface
.u
.legacy
.stencil_tiling_index
[level
];
4311 unsigned macro_index
= iview
->image
->surface
.u
.legacy
.macro_tile_index
;
4312 unsigned tile_mode
= info
->si_tile_mode_array
[tiling_index
];
4313 unsigned stencil_tile_mode
= info
->si_tile_mode_array
[stencil_index
];
4314 unsigned macro_mode
= info
->cik_macrotile_mode_array
[macro_index
];
4317 tile_mode
= stencil_tile_mode
;
4319 ds
->db_depth_info
|=
4320 S_02803C_ARRAY_MODE(G_009910_ARRAY_MODE(tile_mode
)) |
4321 S_02803C_PIPE_CONFIG(G_009910_PIPE_CONFIG(tile_mode
)) |
4322 S_02803C_BANK_WIDTH(G_009990_BANK_WIDTH(macro_mode
)) |
4323 S_02803C_BANK_HEIGHT(G_009990_BANK_HEIGHT(macro_mode
)) |
4324 S_02803C_MACRO_TILE_ASPECT(G_009990_MACRO_TILE_ASPECT(macro_mode
)) |
4325 S_02803C_NUM_BANKS(G_009990_NUM_BANKS(macro_mode
));
4326 ds
->db_z_info
|= S_028040_TILE_SPLIT(G_009910_TILE_SPLIT(tile_mode
));
4327 ds
->db_stencil_info
|= S_028044_TILE_SPLIT(G_009910_TILE_SPLIT(stencil_tile_mode
));
4329 unsigned tile_mode_index
= si_tile_mode_index(iview
->image
, level
, false);
4330 ds
->db_z_info
|= S_028040_TILE_MODE_INDEX(tile_mode_index
);
4331 tile_mode_index
= si_tile_mode_index(iview
->image
, level
, true);
4332 ds
->db_stencil_info
|= S_028044_TILE_MODE_INDEX(tile_mode_index
);
4334 ds
->db_z_info
|= S_028040_TILE_MODE_INDEX(tile_mode_index
);
4337 ds
->db_depth_size
= S_028058_PITCH_TILE_MAX((level_info
->nblk_x
/ 8) - 1) |
4338 S_028058_HEIGHT_TILE_MAX((level_info
->nblk_y
/ 8) - 1);
4339 ds
->db_depth_slice
= S_02805C_SLICE_TILE_MAX((level_info
->nblk_x
* level_info
->nblk_y
) / 64 - 1);
4341 if (radv_htile_enabled(iview
->image
, level
)) {
4342 ds
->db_z_info
|= S_028040_TILE_SURFACE_ENABLE(1);
4344 if (!iview
->image
->surface
.has_stencil
&&
4345 !radv_image_is_tc_compat_htile(iview
->image
))
4346 /* Use all of the htile_buffer for depth if there's no stencil. */
4347 ds
->db_stencil_info
|= S_028044_TILE_STENCIL_DISABLE(1);
4349 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
+
4350 iview
->image
->htile_offset
;
4351 ds
->db_htile_data_base
= va
>> 8;
4352 ds
->db_htile_surface
= S_028ABC_FULL_CACHE(1);
4354 if (radv_image_is_tc_compat_htile(iview
->image
)) {
4355 unsigned max_zplanes
=
4356 radv_calc_decompress_on_z_planes(device
, iview
);
4358 ds
->db_htile_surface
|= S_028ABC_TC_COMPATIBLE(1);
4359 ds
->db_z_info
|= S_028040_DECOMPRESS_ON_N_ZPLANES(max_zplanes
);
4364 ds
->db_z_read_base
= ds
->db_z_write_base
= z_offs
>> 8;
4365 ds
->db_stencil_read_base
= ds
->db_stencil_write_base
= s_offs
>> 8;
4368 VkResult
radv_CreateFramebuffer(
4370 const VkFramebufferCreateInfo
* pCreateInfo
,
4371 const VkAllocationCallbacks
* pAllocator
,
4372 VkFramebuffer
* pFramebuffer
)
4374 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4375 struct radv_framebuffer
*framebuffer
;
4377 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
4379 size_t size
= sizeof(*framebuffer
) +
4380 sizeof(struct radv_attachment_info
) * pCreateInfo
->attachmentCount
;
4381 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
4382 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
4383 if (framebuffer
== NULL
)
4384 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
4386 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
4387 framebuffer
->width
= pCreateInfo
->width
;
4388 framebuffer
->height
= pCreateInfo
->height
;
4389 framebuffer
->layers
= pCreateInfo
->layers
;
4390 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
4391 VkImageView _iview
= pCreateInfo
->pAttachments
[i
];
4392 struct radv_image_view
*iview
= radv_image_view_from_handle(_iview
);
4393 framebuffer
->attachments
[i
].attachment
= iview
;
4394 if (iview
->aspect_mask
& VK_IMAGE_ASPECT_COLOR_BIT
) {
4395 radv_initialise_color_surface(device
, &framebuffer
->attachments
[i
].cb
, iview
);
4396 } else if (iview
->aspect_mask
& (VK_IMAGE_ASPECT_DEPTH_BIT
| VK_IMAGE_ASPECT_STENCIL_BIT
)) {
4397 radv_initialise_ds_surface(device
, &framebuffer
->attachments
[i
].ds
, iview
);
4399 framebuffer
->width
= MIN2(framebuffer
->width
, iview
->extent
.width
);
4400 framebuffer
->height
= MIN2(framebuffer
->height
, iview
->extent
.height
);
4401 framebuffer
->layers
= MIN2(framebuffer
->layers
, radv_surface_max_layer_count(iview
));
4404 *pFramebuffer
= radv_framebuffer_to_handle(framebuffer
);
4408 void radv_DestroyFramebuffer(
4411 const VkAllocationCallbacks
* pAllocator
)
4413 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4414 RADV_FROM_HANDLE(radv_framebuffer
, fb
, _fb
);
4418 vk_free2(&device
->alloc
, pAllocator
, fb
);
4421 static unsigned radv_tex_wrap(VkSamplerAddressMode address_mode
)
4423 switch (address_mode
) {
4424 case VK_SAMPLER_ADDRESS_MODE_REPEAT
:
4425 return V_008F30_SQ_TEX_WRAP
;
4426 case VK_SAMPLER_ADDRESS_MODE_MIRRORED_REPEAT
:
4427 return V_008F30_SQ_TEX_MIRROR
;
4428 case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE
:
4429 return V_008F30_SQ_TEX_CLAMP_LAST_TEXEL
;
4430 case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_BORDER
:
4431 return V_008F30_SQ_TEX_CLAMP_BORDER
;
4432 case VK_SAMPLER_ADDRESS_MODE_MIRROR_CLAMP_TO_EDGE
:
4433 return V_008F30_SQ_TEX_MIRROR_ONCE_LAST_TEXEL
;
4435 unreachable("illegal tex wrap mode");
4441 radv_tex_compare(VkCompareOp op
)
4444 case VK_COMPARE_OP_NEVER
:
4445 return V_008F30_SQ_TEX_DEPTH_COMPARE_NEVER
;
4446 case VK_COMPARE_OP_LESS
:
4447 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESS
;
4448 case VK_COMPARE_OP_EQUAL
:
4449 return V_008F30_SQ_TEX_DEPTH_COMPARE_EQUAL
;
4450 case VK_COMPARE_OP_LESS_OR_EQUAL
:
4451 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESSEQUAL
;
4452 case VK_COMPARE_OP_GREATER
:
4453 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATER
;
4454 case VK_COMPARE_OP_NOT_EQUAL
:
4455 return V_008F30_SQ_TEX_DEPTH_COMPARE_NOTEQUAL
;
4456 case VK_COMPARE_OP_GREATER_OR_EQUAL
:
4457 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATEREQUAL
;
4458 case VK_COMPARE_OP_ALWAYS
:
4459 return V_008F30_SQ_TEX_DEPTH_COMPARE_ALWAYS
;
4461 unreachable("illegal compare mode");
4467 radv_tex_filter(VkFilter filter
, unsigned max_ansio
)
4470 case VK_FILTER_NEAREST
:
4471 return (max_ansio
> 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_POINT
:
4472 V_008F38_SQ_TEX_XY_FILTER_POINT
);
4473 case VK_FILTER_LINEAR
:
4474 return (max_ansio
> 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_BILINEAR
:
4475 V_008F38_SQ_TEX_XY_FILTER_BILINEAR
);
4476 case VK_FILTER_CUBIC_IMG
:
4478 fprintf(stderr
, "illegal texture filter");
4484 radv_tex_mipfilter(VkSamplerMipmapMode mode
)
4487 case VK_SAMPLER_MIPMAP_MODE_NEAREST
:
4488 return V_008F38_SQ_TEX_Z_FILTER_POINT
;
4489 case VK_SAMPLER_MIPMAP_MODE_LINEAR
:
4490 return V_008F38_SQ_TEX_Z_FILTER_LINEAR
;
4492 return V_008F38_SQ_TEX_Z_FILTER_NONE
;
4497 radv_tex_bordercolor(VkBorderColor bcolor
)
4500 case VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
:
4501 case VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
:
4502 return V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK
;
4503 case VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
:
4504 case VK_BORDER_COLOR_INT_OPAQUE_BLACK
:
4505 return V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_BLACK
;
4506 case VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
:
4507 case VK_BORDER_COLOR_INT_OPAQUE_WHITE
:
4508 return V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_WHITE
;
4516 radv_tex_aniso_filter(unsigned filter
)
4530 radv_tex_filter_mode(VkSamplerReductionModeEXT mode
)
4533 case VK_SAMPLER_REDUCTION_MODE_WEIGHTED_AVERAGE_EXT
:
4534 return V_008F30_SQ_IMG_FILTER_MODE_BLEND
;
4535 case VK_SAMPLER_REDUCTION_MODE_MIN_EXT
:
4536 return V_008F30_SQ_IMG_FILTER_MODE_MIN
;
4537 case VK_SAMPLER_REDUCTION_MODE_MAX_EXT
:
4538 return V_008F30_SQ_IMG_FILTER_MODE_MAX
;
4546 radv_get_max_anisotropy(struct radv_device
*device
,
4547 const VkSamplerCreateInfo
*pCreateInfo
)
4549 if (device
->force_aniso
>= 0)
4550 return device
->force_aniso
;
4552 if (pCreateInfo
->anisotropyEnable
&&
4553 pCreateInfo
->maxAnisotropy
> 1.0f
)
4554 return (uint32_t)pCreateInfo
->maxAnisotropy
;
4560 radv_init_sampler(struct radv_device
*device
,
4561 struct radv_sampler
*sampler
,
4562 const VkSamplerCreateInfo
*pCreateInfo
)
4564 uint32_t max_aniso
= radv_get_max_anisotropy(device
, pCreateInfo
);
4565 uint32_t max_aniso_ratio
= radv_tex_aniso_filter(max_aniso
);
4566 bool is_vi
= (device
->physical_device
->rad_info
.chip_class
>= VI
);
4567 unsigned filter_mode
= V_008F30_SQ_IMG_FILTER_MODE_BLEND
;
4569 const struct VkSamplerReductionModeCreateInfoEXT
*sampler_reduction
=
4570 vk_find_struct_const(pCreateInfo
->pNext
,
4571 SAMPLER_REDUCTION_MODE_CREATE_INFO_EXT
);
4572 if (sampler_reduction
)
4573 filter_mode
= radv_tex_filter_mode(sampler_reduction
->reductionMode
);
4575 sampler
->state
[0] = (S_008F30_CLAMP_X(radv_tex_wrap(pCreateInfo
->addressModeU
)) |
4576 S_008F30_CLAMP_Y(radv_tex_wrap(pCreateInfo
->addressModeV
)) |
4577 S_008F30_CLAMP_Z(radv_tex_wrap(pCreateInfo
->addressModeW
)) |
4578 S_008F30_MAX_ANISO_RATIO(max_aniso_ratio
) |
4579 S_008F30_DEPTH_COMPARE_FUNC(radv_tex_compare(pCreateInfo
->compareOp
)) |
4580 S_008F30_FORCE_UNNORMALIZED(pCreateInfo
->unnormalizedCoordinates
? 1 : 0) |
4581 S_008F30_ANISO_THRESHOLD(max_aniso_ratio
>> 1) |
4582 S_008F30_ANISO_BIAS(max_aniso_ratio
) |
4583 S_008F30_DISABLE_CUBE_WRAP(0) |
4584 S_008F30_COMPAT_MODE(is_vi
) |
4585 S_008F30_FILTER_MODE(filter_mode
));
4586 sampler
->state
[1] = (S_008F34_MIN_LOD(S_FIXED(CLAMP(pCreateInfo
->minLod
, 0, 15), 8)) |
4587 S_008F34_MAX_LOD(S_FIXED(CLAMP(pCreateInfo
->maxLod
, 0, 15), 8)) |
4588 S_008F34_PERF_MIP(max_aniso_ratio
? max_aniso_ratio
+ 6 : 0));
4589 sampler
->state
[2] = (S_008F38_LOD_BIAS(S_FIXED(CLAMP(pCreateInfo
->mipLodBias
, -16, 16), 8)) |
4590 S_008F38_XY_MAG_FILTER(radv_tex_filter(pCreateInfo
->magFilter
, max_aniso
)) |
4591 S_008F38_XY_MIN_FILTER(radv_tex_filter(pCreateInfo
->minFilter
, max_aniso
)) |
4592 S_008F38_MIP_FILTER(radv_tex_mipfilter(pCreateInfo
->mipmapMode
)) |
4593 S_008F38_MIP_POINT_PRECLAMP(0) |
4594 S_008F38_DISABLE_LSB_CEIL(device
->physical_device
->rad_info
.chip_class
<= VI
) |
4595 S_008F38_FILTER_PREC_FIX(1) |
4596 S_008F38_ANISO_OVERRIDE(is_vi
));
4597 sampler
->state
[3] = (S_008F3C_BORDER_COLOR_PTR(0) |
4598 S_008F3C_BORDER_COLOR_TYPE(radv_tex_bordercolor(pCreateInfo
->borderColor
)));
4601 VkResult
radv_CreateSampler(
4603 const VkSamplerCreateInfo
* pCreateInfo
,
4604 const VkAllocationCallbacks
* pAllocator
,
4605 VkSampler
* pSampler
)
4607 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4608 struct radv_sampler
*sampler
;
4610 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO
);
4612 sampler
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*sampler
), 8,
4613 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
4615 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
4617 radv_init_sampler(device
, sampler
, pCreateInfo
);
4618 *pSampler
= radv_sampler_to_handle(sampler
);
4623 void radv_DestroySampler(
4626 const VkAllocationCallbacks
* pAllocator
)
4628 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4629 RADV_FROM_HANDLE(radv_sampler
, sampler
, _sampler
);
4633 vk_free2(&device
->alloc
, pAllocator
, sampler
);
4636 /* vk_icd.h does not declare this function, so we declare it here to
4637 * suppress Wmissing-prototypes.
4639 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4640 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t *pSupportedVersion
);
4642 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4643 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t *pSupportedVersion
)
4645 /* For the full details on loader interface versioning, see
4646 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
4647 * What follows is a condensed summary, to help you navigate the large and
4648 * confusing official doc.
4650 * - Loader interface v0 is incompatible with later versions. We don't
4653 * - In loader interface v1:
4654 * - The first ICD entrypoint called by the loader is
4655 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
4657 * - The ICD must statically expose no other Vulkan symbol unless it is
4658 * linked with -Bsymbolic.
4659 * - Each dispatchable Vulkan handle created by the ICD must be
4660 * a pointer to a struct whose first member is VK_LOADER_DATA. The
4661 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
4662 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
4663 * vkDestroySurfaceKHR(). The ICD must be capable of working with
4664 * such loader-managed surfaces.
4666 * - Loader interface v2 differs from v1 in:
4667 * - The first ICD entrypoint called by the loader is
4668 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
4669 * statically expose this entrypoint.
4671 * - Loader interface v3 differs from v2 in:
4672 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
4673 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
4674 * because the loader no longer does so.
4676 *pSupportedVersion
= MIN2(*pSupportedVersion
, 3u);
4680 VkResult
radv_GetMemoryFdKHR(VkDevice _device
,
4681 const VkMemoryGetFdInfoKHR
*pGetFdInfo
,
4684 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4685 RADV_FROM_HANDLE(radv_device_memory
, memory
, pGetFdInfo
->memory
);
4687 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
4689 /* At the moment, we support only the below handle types. */
4690 assert(pGetFdInfo
->handleType
==
4691 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
||
4692 pGetFdInfo
->handleType
==
4693 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
4695 bool ret
= radv_get_memory_fd(device
, memory
, pFD
);
4697 return vk_error(device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
4701 VkResult
radv_GetMemoryFdPropertiesKHR(VkDevice _device
,
4702 VkExternalMemoryHandleTypeFlagBitsKHR handleType
,
4704 VkMemoryFdPropertiesKHR
*pMemoryFdProperties
)
4706 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4708 switch (handleType
) {
4709 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
:
4710 pMemoryFdProperties
->memoryTypeBits
= (1 << RADV_MEM_TYPE_COUNT
) - 1;
4714 /* The valid usage section for this function says:
4716 * "handleType must not be one of the handle types defined as
4719 * So opaque handle types fall into the default "unsupported" case.
4721 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
4725 static VkResult
radv_import_opaque_fd(struct radv_device
*device
,
4729 uint32_t syncobj_handle
= 0;
4730 int ret
= device
->ws
->import_syncobj(device
->ws
, fd
, &syncobj_handle
);
4732 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
4735 device
->ws
->destroy_syncobj(device
->ws
, *syncobj
);
4737 *syncobj
= syncobj_handle
;
4743 static VkResult
radv_import_sync_fd(struct radv_device
*device
,
4747 /* If we create a syncobj we do it locally so that if we have an error, we don't
4748 * leave a syncobj in an undetermined state in the fence. */
4749 uint32_t syncobj_handle
= *syncobj
;
4750 if (!syncobj_handle
) {
4751 int ret
= device
->ws
->create_syncobj(device
->ws
, &syncobj_handle
);
4753 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
4758 device
->ws
->signal_syncobj(device
->ws
, syncobj_handle
);
4760 int ret
= device
->ws
->import_syncobj_from_sync_file(device
->ws
, syncobj_handle
, fd
);
4762 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
4765 *syncobj
= syncobj_handle
;
4772 VkResult
radv_ImportSemaphoreFdKHR(VkDevice _device
,
4773 const VkImportSemaphoreFdInfoKHR
*pImportSemaphoreFdInfo
)
4775 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4776 RADV_FROM_HANDLE(radv_semaphore
, sem
, pImportSemaphoreFdInfo
->semaphore
);
4777 uint32_t *syncobj_dst
= NULL
;
4779 if (pImportSemaphoreFdInfo
->flags
& VK_SEMAPHORE_IMPORT_TEMPORARY_BIT_KHR
) {
4780 syncobj_dst
= &sem
->temp_syncobj
;
4782 syncobj_dst
= &sem
->syncobj
;
4785 switch(pImportSemaphoreFdInfo
->handleType
) {
4786 case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
:
4787 return radv_import_opaque_fd(device
, pImportSemaphoreFdInfo
->fd
, syncobj_dst
);
4788 case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT_KHR
:
4789 return radv_import_sync_fd(device
, pImportSemaphoreFdInfo
->fd
, syncobj_dst
);
4791 unreachable("Unhandled semaphore handle type");
4795 VkResult
radv_GetSemaphoreFdKHR(VkDevice _device
,
4796 const VkSemaphoreGetFdInfoKHR
*pGetFdInfo
,
4799 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4800 RADV_FROM_HANDLE(radv_semaphore
, sem
, pGetFdInfo
->semaphore
);
4802 uint32_t syncobj_handle
;
4804 if (sem
->temp_syncobj
)
4805 syncobj_handle
= sem
->temp_syncobj
;
4807 syncobj_handle
= sem
->syncobj
;
4809 switch(pGetFdInfo
->handleType
) {
4810 case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
:
4811 ret
= device
->ws
->export_syncobj(device
->ws
, syncobj_handle
, pFd
);
4813 case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT_KHR
:
4814 ret
= device
->ws
->export_syncobj_to_sync_file(device
->ws
, syncobj_handle
, pFd
);
4816 if (sem
->temp_syncobj
) {
4817 close (sem
->temp_syncobj
);
4818 sem
->temp_syncobj
= 0;
4820 device
->ws
->reset_syncobj(device
->ws
, syncobj_handle
);
4825 unreachable("Unhandled semaphore handle type");
4829 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
4833 void radv_GetPhysicalDeviceExternalSemaphoreProperties(
4834 VkPhysicalDevice physicalDevice
,
4835 const VkPhysicalDeviceExternalSemaphoreInfoKHR
* pExternalSemaphoreInfo
,
4836 VkExternalSemaphorePropertiesKHR
* pExternalSemaphoreProperties
)
4838 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
4840 /* Require has_syncobj_wait_for_submit for the syncobj signal ioctl introduced at virtually the same time */
4841 if (pdevice
->rad_info
.has_syncobj_wait_for_submit
&&
4842 (pExternalSemaphoreInfo
->handleType
== VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
||
4843 pExternalSemaphoreInfo
->handleType
== VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT_KHR
)) {
4844 pExternalSemaphoreProperties
->exportFromImportedHandleTypes
= VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
| VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT_KHR
;
4845 pExternalSemaphoreProperties
->compatibleHandleTypes
= VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
| VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT_KHR
;
4846 pExternalSemaphoreProperties
->externalSemaphoreFeatures
= VK_EXTERNAL_SEMAPHORE_FEATURE_EXPORTABLE_BIT_KHR
|
4847 VK_EXTERNAL_SEMAPHORE_FEATURE_IMPORTABLE_BIT_KHR
;
4848 } else if (pExternalSemaphoreInfo
->handleType
== VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
) {
4849 pExternalSemaphoreProperties
->exportFromImportedHandleTypes
= VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
;
4850 pExternalSemaphoreProperties
->compatibleHandleTypes
= VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
;
4851 pExternalSemaphoreProperties
->externalSemaphoreFeatures
= VK_EXTERNAL_SEMAPHORE_FEATURE_EXPORTABLE_BIT_KHR
|
4852 VK_EXTERNAL_SEMAPHORE_FEATURE_IMPORTABLE_BIT_KHR
;
4854 pExternalSemaphoreProperties
->exportFromImportedHandleTypes
= 0;
4855 pExternalSemaphoreProperties
->compatibleHandleTypes
= 0;
4856 pExternalSemaphoreProperties
->externalSemaphoreFeatures
= 0;
4860 VkResult
radv_ImportFenceFdKHR(VkDevice _device
,
4861 const VkImportFenceFdInfoKHR
*pImportFenceFdInfo
)
4863 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4864 RADV_FROM_HANDLE(radv_fence
, fence
, pImportFenceFdInfo
->fence
);
4865 uint32_t *syncobj_dst
= NULL
;
4868 if (pImportFenceFdInfo
->flags
& VK_FENCE_IMPORT_TEMPORARY_BIT_KHR
) {
4869 syncobj_dst
= &fence
->temp_syncobj
;
4871 syncobj_dst
= &fence
->syncobj
;
4874 switch(pImportFenceFdInfo
->handleType
) {
4875 case VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
:
4876 return radv_import_opaque_fd(device
, pImportFenceFdInfo
->fd
, syncobj_dst
);
4877 case VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT_KHR
:
4878 return radv_import_sync_fd(device
, pImportFenceFdInfo
->fd
, syncobj_dst
);
4880 unreachable("Unhandled fence handle type");
4884 VkResult
radv_GetFenceFdKHR(VkDevice _device
,
4885 const VkFenceGetFdInfoKHR
*pGetFdInfo
,
4888 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4889 RADV_FROM_HANDLE(radv_fence
, fence
, pGetFdInfo
->fence
);
4891 uint32_t syncobj_handle
;
4893 if (fence
->temp_syncobj
)
4894 syncobj_handle
= fence
->temp_syncobj
;
4896 syncobj_handle
= fence
->syncobj
;
4898 switch(pGetFdInfo
->handleType
) {
4899 case VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
:
4900 ret
= device
->ws
->export_syncobj(device
->ws
, syncobj_handle
, pFd
);
4902 case VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT_KHR
:
4903 ret
= device
->ws
->export_syncobj_to_sync_file(device
->ws
, syncobj_handle
, pFd
);
4905 if (fence
->temp_syncobj
) {
4906 close (fence
->temp_syncobj
);
4907 fence
->temp_syncobj
= 0;
4909 device
->ws
->reset_syncobj(device
->ws
, syncobj_handle
);
4914 unreachable("Unhandled fence handle type");
4918 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
4922 void radv_GetPhysicalDeviceExternalFenceProperties(
4923 VkPhysicalDevice physicalDevice
,
4924 const VkPhysicalDeviceExternalFenceInfoKHR
* pExternalFenceInfo
,
4925 VkExternalFencePropertiesKHR
* pExternalFenceProperties
)
4927 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
4929 if (pdevice
->rad_info
.has_syncobj_wait_for_submit
&&
4930 (pExternalFenceInfo
->handleType
== VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
||
4931 pExternalFenceInfo
->handleType
== VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT_KHR
)) {
4932 pExternalFenceProperties
->exportFromImportedHandleTypes
= VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
| VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT_KHR
;
4933 pExternalFenceProperties
->compatibleHandleTypes
= VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
| VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT_KHR
;
4934 pExternalFenceProperties
->externalFenceFeatures
= VK_EXTERNAL_FENCE_FEATURE_EXPORTABLE_BIT_KHR
|
4935 VK_EXTERNAL_SEMAPHORE_FEATURE_IMPORTABLE_BIT_KHR
;
4937 pExternalFenceProperties
->exportFromImportedHandleTypes
= 0;
4938 pExternalFenceProperties
->compatibleHandleTypes
= 0;
4939 pExternalFenceProperties
->externalFenceFeatures
= 0;
4944 radv_CreateDebugReportCallbackEXT(VkInstance _instance
,
4945 const VkDebugReportCallbackCreateInfoEXT
* pCreateInfo
,
4946 const VkAllocationCallbacks
* pAllocator
,
4947 VkDebugReportCallbackEXT
* pCallback
)
4949 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
4950 return vk_create_debug_report_callback(&instance
->debug_report_callbacks
,
4951 pCreateInfo
, pAllocator
, &instance
->alloc
,
4956 radv_DestroyDebugReportCallbackEXT(VkInstance _instance
,
4957 VkDebugReportCallbackEXT _callback
,
4958 const VkAllocationCallbacks
* pAllocator
)
4960 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
4961 vk_destroy_debug_report_callback(&instance
->debug_report_callbacks
,
4962 _callback
, pAllocator
, &instance
->alloc
);
4966 radv_DebugReportMessageEXT(VkInstance _instance
,
4967 VkDebugReportFlagsEXT flags
,
4968 VkDebugReportObjectTypeEXT objectType
,
4971 int32_t messageCode
,
4972 const char* pLayerPrefix
,
4973 const char* pMessage
)
4975 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
4976 vk_debug_report(&instance
->debug_report_callbacks
, flags
, objectType
,
4977 object
, location
, messageCode
, pLayerPrefix
, pMessage
);
4981 radv_GetDeviceGroupPeerMemoryFeatures(
4984 uint32_t localDeviceIndex
,
4985 uint32_t remoteDeviceIndex
,
4986 VkPeerMemoryFeatureFlags
* pPeerMemoryFeatures
)
4988 assert(localDeviceIndex
== remoteDeviceIndex
);
4990 *pPeerMemoryFeatures
= VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT
|
4991 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT
|
4992 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT
|
4993 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT
;
4996 static const VkTimeDomainEXT radv_time_domains
[] = {
4997 VK_TIME_DOMAIN_DEVICE_EXT
,
4998 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
,
4999 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
,
5002 VkResult
radv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
5003 VkPhysicalDevice physicalDevice
,
5004 uint32_t *pTimeDomainCount
,
5005 VkTimeDomainEXT
*pTimeDomains
)
5008 VK_OUTARRAY_MAKE(out
, pTimeDomains
, pTimeDomainCount
);
5010 for (d
= 0; d
< ARRAY_SIZE(radv_time_domains
); d
++) {
5011 vk_outarray_append(&out
, i
) {
5012 *i
= radv_time_domains
[d
];
5016 return vk_outarray_status(&out
);
5020 radv_clock_gettime(clockid_t clock_id
)
5022 struct timespec current
;
5025 ret
= clock_gettime(clock_id
, ¤t
);
5026 if (ret
< 0 && clock_id
== CLOCK_MONOTONIC_RAW
)
5027 ret
= clock_gettime(CLOCK_MONOTONIC
, ¤t
);
5031 return (uint64_t) current
.tv_sec
* 1000000000ULL + current
.tv_nsec
;
5034 VkResult
radv_GetCalibratedTimestampsEXT(
5036 uint32_t timestampCount
,
5037 const VkCalibratedTimestampInfoEXT
*pTimestampInfos
,
5038 uint64_t *pTimestamps
,
5039 uint64_t *pMaxDeviation
)
5041 RADV_FROM_HANDLE(radv_device
, device
, _device
);
5042 uint32_t clock_crystal_freq
= device
->physical_device
->rad_info
.clock_crystal_freq
;
5044 uint64_t begin
, end
;
5045 uint64_t max_clock_period
= 0;
5047 begin
= radv_clock_gettime(CLOCK_MONOTONIC_RAW
);
5049 for (d
= 0; d
< timestampCount
; d
++) {
5050 switch (pTimestampInfos
[d
].timeDomain
) {
5051 case VK_TIME_DOMAIN_DEVICE_EXT
:
5052 pTimestamps
[d
] = device
->ws
->query_value(device
->ws
,
5054 uint64_t device_period
= DIV_ROUND_UP(1000000, clock_crystal_freq
);
5055 max_clock_period
= MAX2(max_clock_period
, device_period
);
5057 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT
:
5058 pTimestamps
[d
] = radv_clock_gettime(CLOCK_MONOTONIC
);
5059 max_clock_period
= MAX2(max_clock_period
, 1);
5062 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT
:
5063 pTimestamps
[d
] = begin
;
5071 end
= radv_clock_gettime(CLOCK_MONOTONIC_RAW
);
5074 * The maximum deviation is the sum of the interval over which we
5075 * perform the sampling and the maximum period of any sampled
5076 * clock. That's because the maximum skew between any two sampled
5077 * clock edges is when the sampled clock with the largest period is
5078 * sampled at the end of that period but right at the beginning of the
5079 * sampling interval and some other clock is sampled right at the
5080 * begining of its sampling period and right at the end of the
5081 * sampling interval. Let's assume the GPU has the longest clock
5082 * period and that the application is sampling GPU and monotonic:
5085 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
5086 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
5090 * GPU -----_____-----_____-----_____-----_____
5093 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
5094 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
5096 * Interval <----------------->
5097 * Deviation <-------------------------->
5101 * m = read(monotonic) 2
5104 * We round the sample interval up by one tick to cover sampling error
5105 * in the interval clock
5108 uint64_t sample_interval
= end
- begin
+ 1;
5110 *pMaxDeviation
= sample_interval
+ max_clock_period
;