2 * Copyright © 2016 Red Hat.
3 * Copyright © 2016 Bas Nieuwenhuizen
5 * based in part on anv driver which is:
6 * Copyright © 2015 Intel Corporation
8 * Permission is hereby granted, free of charge, to any person obtaining a
9 * copy of this software and associated documentation files (the "Software"),
10 * to deal in the Software without restriction, including without limitation
11 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
12 * and/or sell copies of the Software, and to permit persons to whom the
13 * Software is furnished to do so, subject to the following conditions:
15 * The above copyright notice and this permission notice (including the next
16 * paragraph) shall be included in all copies or substantial portions of the
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
22 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
23 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
24 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
32 #include "radv_debug.h"
33 #include "radv_private.h"
34 #include "radv_shader.h"
36 #include "util/disk_cache.h"
37 #include "util/strtod.h"
41 #include <amdgpu_drm.h>
42 #include "winsys/amdgpu/radv_amdgpu_winsys_public.h"
43 #include "ac_llvm_util.h"
44 #include "vk_format.h"
47 #include "addrlib/gfx9/chip/gfx9_enum.h"
48 #include "util/build_id.h"
49 #include "util/debug.h"
50 #include "util/mesa-sha1.h"
53 radv_get_build_id(void *ptr
, struct mesa_sha1
*ctx
)
57 #ifdef HAVE_DL_ITERATE_PHDR
58 const struct build_id_note
*note
= NULL
;
59 if ((note
= build_id_find_nhdr_for_addr(ptr
))) {
60 _mesa_sha1_update(ctx
, build_id_data(note
), build_id_length(note
));
63 if (disk_cache_get_function_timestamp(ptr
, ×tamp
)) {
65 fprintf(stderr
, "radv: The provided filesystem timestamp for the cache is bogus!\n");
68 _mesa_sha1_update(ctx
, ×tamp
, sizeof(timestamp
));
75 radv_device_get_cache_uuid(enum radeon_family family
, void *uuid
)
78 unsigned char sha1
[20];
79 unsigned ptr_size
= sizeof(void*);
80 memset(uuid
, 0, VK_UUID_SIZE
);
82 if (!radv_get_build_id(radv_device_get_cache_uuid
, &ctx
) ||
83 !radv_get_build_id(LLVMInitializeAMDGPUTargetInfo
, &ctx
))
86 _mesa_sha1_update(&ctx
, &family
, sizeof(family
));
87 _mesa_sha1_update(&ctx
, &ptr_size
, sizeof(ptr_size
));
88 _mesa_sha1_final(&ctx
, sha1
);
90 memcpy(uuid
, sha1
, VK_UUID_SIZE
);
95 radv_get_driver_uuid(void *uuid
)
97 ac_compute_driver_uuid(uuid
, VK_UUID_SIZE
);
101 radv_get_device_uuid(struct radeon_info
*info
, void *uuid
)
103 ac_compute_device_uuid(info
, uuid
, VK_UUID_SIZE
);
107 radv_get_device_name(enum radeon_family family
, char *name
, size_t name_len
)
109 const char *chip_string
;
110 char llvm_string
[32] = {};
113 case CHIP_TAHITI
: chip_string
= "AMD RADV TAHITI"; break;
114 case CHIP_PITCAIRN
: chip_string
= "AMD RADV PITCAIRN"; break;
115 case CHIP_VERDE
: chip_string
= "AMD RADV CAPE VERDE"; break;
116 case CHIP_OLAND
: chip_string
= "AMD RADV OLAND"; break;
117 case CHIP_HAINAN
: chip_string
= "AMD RADV HAINAN"; break;
118 case CHIP_BONAIRE
: chip_string
= "AMD RADV BONAIRE"; break;
119 case CHIP_KAVERI
: chip_string
= "AMD RADV KAVERI"; break;
120 case CHIP_KABINI
: chip_string
= "AMD RADV KABINI"; break;
121 case CHIP_HAWAII
: chip_string
= "AMD RADV HAWAII"; break;
122 case CHIP_MULLINS
: chip_string
= "AMD RADV MULLINS"; break;
123 case CHIP_TONGA
: chip_string
= "AMD RADV TONGA"; break;
124 case CHIP_ICELAND
: chip_string
= "AMD RADV ICELAND"; break;
125 case CHIP_CARRIZO
: chip_string
= "AMD RADV CARRIZO"; break;
126 case CHIP_FIJI
: chip_string
= "AMD RADV FIJI"; break;
127 case CHIP_POLARIS10
: chip_string
= "AMD RADV POLARIS10"; break;
128 case CHIP_POLARIS11
: chip_string
= "AMD RADV POLARIS11"; break;
129 case CHIP_POLARIS12
: chip_string
= "AMD RADV POLARIS12"; break;
130 case CHIP_STONEY
: chip_string
= "AMD RADV STONEY"; break;
131 case CHIP_VEGAM
: chip_string
= "AMD RADV VEGA M"; break;
132 case CHIP_VEGA10
: chip_string
= "AMD RADV VEGA10"; break;
133 case CHIP_VEGA12
: chip_string
= "AMD RADV VEGA12"; break;
134 case CHIP_RAVEN
: chip_string
= "AMD RADV RAVEN"; break;
135 default: chip_string
= "AMD RADV unknown"; break;
138 snprintf(llvm_string
, sizeof(llvm_string
),
139 " (LLVM %i.%i.%i)", (HAVE_LLVM
>> 8) & 0xff,
140 HAVE_LLVM
& 0xff, MESA_LLVM_VERSION_PATCH
);
141 snprintf(name
, name_len
, "%s%s", chip_string
, llvm_string
);
145 radv_physical_device_init_mem_types(struct radv_physical_device
*device
)
147 STATIC_ASSERT(RADV_MEM_HEAP_COUNT
<= VK_MAX_MEMORY_HEAPS
);
148 uint64_t visible_vram_size
= MIN2(device
->rad_info
.vram_size
,
149 device
->rad_info
.vram_vis_size
);
151 int vram_index
= -1, visible_vram_index
= -1, gart_index
= -1;
152 device
->memory_properties
.memoryHeapCount
= 0;
153 if (device
->rad_info
.vram_size
- visible_vram_size
> 0) {
154 vram_index
= device
->memory_properties
.memoryHeapCount
++;
155 device
->memory_properties
.memoryHeaps
[vram_index
] = (VkMemoryHeap
) {
156 .size
= device
->rad_info
.vram_size
- visible_vram_size
,
157 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
160 if (visible_vram_size
) {
161 visible_vram_index
= device
->memory_properties
.memoryHeapCount
++;
162 device
->memory_properties
.memoryHeaps
[visible_vram_index
] = (VkMemoryHeap
) {
163 .size
= visible_vram_size
,
164 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
167 if (device
->rad_info
.gart_size
> 0) {
168 gart_index
= device
->memory_properties
.memoryHeapCount
++;
169 device
->memory_properties
.memoryHeaps
[gart_index
] = (VkMemoryHeap
) {
170 .size
= device
->rad_info
.gart_size
,
171 .flags
= device
->rad_info
.has_dedicated_vram
? 0 : VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
175 STATIC_ASSERT(RADV_MEM_TYPE_COUNT
<= VK_MAX_MEMORY_TYPES
);
176 unsigned type_count
= 0;
177 if (vram_index
>= 0) {
178 device
->mem_type_indices
[type_count
] = RADV_MEM_TYPE_VRAM
;
179 device
->memory_properties
.memoryTypes
[type_count
++] = (VkMemoryType
) {
180 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
,
181 .heapIndex
= vram_index
,
184 if (gart_index
>= 0) {
185 device
->mem_type_indices
[type_count
] = RADV_MEM_TYPE_GTT_WRITE_COMBINE
;
186 device
->memory_properties
.memoryTypes
[type_count
++] = (VkMemoryType
) {
187 .propertyFlags
= VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
188 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
189 (device
->rad_info
.has_dedicated_vram
? 0 : VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
),
190 .heapIndex
= gart_index
,
193 if (visible_vram_index
>= 0) {
194 device
->mem_type_indices
[type_count
] = RADV_MEM_TYPE_VRAM_CPU_ACCESS
;
195 device
->memory_properties
.memoryTypes
[type_count
++] = (VkMemoryType
) {
196 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
197 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
198 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
199 .heapIndex
= visible_vram_index
,
202 if (gart_index
>= 0) {
203 device
->mem_type_indices
[type_count
] = RADV_MEM_TYPE_GTT_CACHED
;
204 device
->memory_properties
.memoryTypes
[type_count
++] = (VkMemoryType
) {
205 .propertyFlags
= VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
206 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
207 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
|
208 (device
->rad_info
.has_dedicated_vram
? 0 : VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
),
209 .heapIndex
= gart_index
,
212 device
->memory_properties
.memoryTypeCount
= type_count
;
216 radv_handle_env_var_force_family(struct radv_physical_device
*device
)
218 const char *family
= getenv("RADV_FORCE_FAMILY");
224 for (i
= CHIP_TAHITI
; i
< CHIP_LAST
; i
++) {
225 if (!strcmp(family
, ac_get_llvm_processor_name(i
))) {
226 /* Override family and chip_class. */
227 device
->rad_info
.family
= i
;
229 if (i
>= CHIP_VEGA10
)
230 device
->rad_info
.chip_class
= GFX9
;
231 else if (i
>= CHIP_TONGA
)
232 device
->rad_info
.chip_class
= VI
;
233 else if (i
>= CHIP_BONAIRE
)
234 device
->rad_info
.chip_class
= CIK
;
236 device
->rad_info
.chip_class
= SI
;
242 fprintf(stderr
, "radv: Unknown family: %s\n", family
);
247 radv_physical_device_init(struct radv_physical_device
*device
,
248 struct radv_instance
*instance
,
249 drmDevicePtr drm_device
)
251 const char *path
= drm_device
->nodes
[DRM_NODE_RENDER
];
253 drmVersionPtr version
;
257 fd
= open(path
, O_RDWR
| O_CLOEXEC
);
259 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
260 radv_logi("Could not open device '%s'", path
);
262 return vk_error(instance
, VK_ERROR_INCOMPATIBLE_DRIVER
);
265 version
= drmGetVersion(fd
);
269 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
270 radv_logi("Could not get the kernel driver version for device '%s'", path
);
272 return vk_errorf(instance
, VK_ERROR_INCOMPATIBLE_DRIVER
,
273 "failed to get version %s: %m", path
);
276 if (strcmp(version
->name
, "amdgpu")) {
277 drmFreeVersion(version
);
282 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
283 radv_logi("Device '%s' is not using the amdgpu kernel driver.", path
);
285 return VK_ERROR_INCOMPATIBLE_DRIVER
;
287 drmFreeVersion(version
);
289 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
290 radv_logi("Found compatible device '%s'.", path
);
292 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
293 device
->instance
= instance
;
294 assert(strlen(path
) < ARRAY_SIZE(device
->path
));
295 strncpy(device
->path
, path
, ARRAY_SIZE(device
->path
));
297 device
->ws
= radv_amdgpu_winsys_create(fd
, instance
->debug_flags
,
298 instance
->perftest_flags
);
300 result
= vk_error(instance
, VK_ERROR_INCOMPATIBLE_DRIVER
);
304 if (instance
->enabled_extensions
.KHR_display
) {
305 master_fd
= open(drm_device
->nodes
[DRM_NODE_PRIMARY
], O_RDWR
| O_CLOEXEC
);
306 if (master_fd
>= 0) {
307 uint32_t accel_working
= 0;
308 struct drm_amdgpu_info request
= {
309 .return_pointer
= (uintptr_t)&accel_working
,
310 .return_size
= sizeof(accel_working
),
311 .query
= AMDGPU_INFO_ACCEL_WORKING
314 if (drmCommandWrite(master_fd
, DRM_AMDGPU_INFO
, &request
, sizeof (struct drm_amdgpu_info
)) < 0 || !accel_working
) {
321 device
->master_fd
= master_fd
;
322 device
->local_fd
= fd
;
323 device
->ws
->query_info(device
->ws
, &device
->rad_info
);
325 radv_handle_env_var_force_family(device
);
327 radv_get_device_name(device
->rad_info
.family
, device
->name
, sizeof(device
->name
));
329 if (radv_device_get_cache_uuid(device
->rad_info
.family
, device
->cache_uuid
)) {
330 device
->ws
->destroy(device
->ws
);
331 result
= vk_errorf(instance
, VK_ERROR_INITIALIZATION_FAILED
,
332 "cannot generate UUID");
336 /* These flags affect shader compilation. */
337 uint64_t shader_env_flags
=
338 (device
->instance
->perftest_flags
& RADV_PERFTEST_SISCHED
? 0x1 : 0) |
339 (device
->instance
->debug_flags
& RADV_DEBUG_UNSAFE_MATH
? 0x2 : 0);
341 /* The gpu id is already embedded in the uuid so we just pass "radv"
342 * when creating the cache.
344 char buf
[VK_UUID_SIZE
* 2 + 1];
345 disk_cache_format_hex_id(buf
, device
->cache_uuid
, VK_UUID_SIZE
* 2);
346 device
->disk_cache
= disk_cache_create(device
->name
, buf
, shader_env_flags
);
348 if (device
->rad_info
.chip_class
< VI
||
349 device
->rad_info
.chip_class
> GFX9
)
350 fprintf(stderr
, "WARNING: radv is not a conformant vulkan implementation, testing use only.\n");
352 radv_get_driver_uuid(&device
->device_uuid
);
353 radv_get_device_uuid(&device
->rad_info
, &device
->device_uuid
);
355 if (device
->rad_info
.family
== CHIP_STONEY
||
356 device
->rad_info
.chip_class
>= GFX9
) {
357 device
->has_rbplus
= true;
358 device
->rbplus_allowed
= device
->rad_info
.family
== CHIP_STONEY
||
359 device
->rad_info
.family
== CHIP_VEGA12
||
360 device
->rad_info
.family
== CHIP_RAVEN
;
363 /* The mere presence of CLEAR_STATE in the IB causes random GPU hangs
366 device
->has_clear_state
= device
->rad_info
.chip_class
>= CIK
;
368 device
->cpdma_prefetch_writes_memory
= device
->rad_info
.chip_class
<= VI
;
370 /* Vega10/Raven need a special workaround for a hardware bug. */
371 device
->has_scissor_bug
= device
->rad_info
.family
== CHIP_VEGA10
||
372 device
->rad_info
.family
== CHIP_RAVEN
;
374 /* Out-of-order primitive rasterization. */
375 device
->has_out_of_order_rast
= device
->rad_info
.chip_class
>= VI
&&
376 device
->rad_info
.max_se
>= 2;
377 device
->out_of_order_rast_allowed
= device
->has_out_of_order_rast
&&
378 !(device
->instance
->debug_flags
& RADV_DEBUG_NO_OUT_OF_ORDER
);
380 device
->dcc_msaa_allowed
=
381 (device
->instance
->perftest_flags
& RADV_PERFTEST_DCC_MSAA
);
383 radv_physical_device_init_mem_types(device
);
384 radv_fill_device_extension_table(device
, &device
->supported_extensions
);
386 result
= radv_init_wsi(device
);
387 if (result
!= VK_SUCCESS
) {
388 device
->ws
->destroy(device
->ws
);
389 vk_error(instance
, result
);
393 if ((device
->instance
->debug_flags
& RADV_DEBUG_INFO
))
394 ac_print_gpu_info(&device
->rad_info
);
406 radv_physical_device_finish(struct radv_physical_device
*device
)
408 radv_finish_wsi(device
);
409 device
->ws
->destroy(device
->ws
);
410 disk_cache_destroy(device
->disk_cache
);
411 close(device
->local_fd
);
412 if (device
->master_fd
!= -1)
413 close(device
->master_fd
);
417 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
418 VkSystemAllocationScope allocationScope
)
424 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
425 size_t align
, VkSystemAllocationScope allocationScope
)
427 return realloc(pOriginal
, size
);
431 default_free_func(void *pUserData
, void *pMemory
)
436 static const VkAllocationCallbacks default_alloc
= {
438 .pfnAllocation
= default_alloc_func
,
439 .pfnReallocation
= default_realloc_func
,
440 .pfnFree
= default_free_func
,
443 static const struct debug_control radv_debug_options
[] = {
444 {"nofastclears", RADV_DEBUG_NO_FAST_CLEARS
},
445 {"nodcc", RADV_DEBUG_NO_DCC
},
446 {"shaders", RADV_DEBUG_DUMP_SHADERS
},
447 {"nocache", RADV_DEBUG_NO_CACHE
},
448 {"shaderstats", RADV_DEBUG_DUMP_SHADER_STATS
},
449 {"nohiz", RADV_DEBUG_NO_HIZ
},
450 {"nocompute", RADV_DEBUG_NO_COMPUTE_QUEUE
},
451 {"unsafemath", RADV_DEBUG_UNSAFE_MATH
},
452 {"allbos", RADV_DEBUG_ALL_BOS
},
453 {"noibs", RADV_DEBUG_NO_IBS
},
454 {"spirv", RADV_DEBUG_DUMP_SPIRV
},
455 {"vmfaults", RADV_DEBUG_VM_FAULTS
},
456 {"zerovram", RADV_DEBUG_ZERO_VRAM
},
457 {"syncshaders", RADV_DEBUG_SYNC_SHADERS
},
458 {"nosisched", RADV_DEBUG_NO_SISCHED
},
459 {"preoptir", RADV_DEBUG_PREOPTIR
},
460 {"nodynamicbounds", RADV_DEBUG_NO_DYNAMIC_BOUNDS
},
461 {"nooutoforder", RADV_DEBUG_NO_OUT_OF_ORDER
},
462 {"info", RADV_DEBUG_INFO
},
463 {"errors", RADV_DEBUG_ERRORS
},
464 {"startup", RADV_DEBUG_STARTUP
},
465 {"checkir", RADV_DEBUG_CHECKIR
},
466 {"nothreadllvm", RADV_DEBUG_NOTHREADLLVM
},
471 radv_get_debug_option_name(int id
)
473 assert(id
< ARRAY_SIZE(radv_debug_options
) - 1);
474 return radv_debug_options
[id
].string
;
477 static const struct debug_control radv_perftest_options
[] = {
478 {"nobatchchain", RADV_PERFTEST_NO_BATCHCHAIN
},
479 {"sisched", RADV_PERFTEST_SISCHED
},
480 {"localbos", RADV_PERFTEST_LOCAL_BOS
},
481 {"binning", RADV_PERFTEST_BINNING
},
482 {"dccmsaa", RADV_PERFTEST_DCC_MSAA
},
487 radv_get_perftest_option_name(int id
)
489 assert(id
< ARRAY_SIZE(radv_debug_options
) - 1);
490 return radv_perftest_options
[id
].string
;
494 radv_handle_per_app_options(struct radv_instance
*instance
,
495 const VkApplicationInfo
*info
)
497 const char *name
= info
? info
->pApplicationName
: NULL
;
502 if (!strcmp(name
, "Talos - Linux - 32bit") ||
503 !strcmp(name
, "Talos - Linux - 64bit")) {
504 if (!(instance
->debug_flags
& RADV_DEBUG_NO_SISCHED
)) {
505 /* Force enable LLVM sisched for Talos because it looks
506 * safe and it gives few more FPS.
508 instance
->perftest_flags
|= RADV_PERFTEST_SISCHED
;
510 } else if (!strcmp(name
, "DOOM_VFR")) {
511 /* Work around a Doom VFR game bug */
512 instance
->debug_flags
|= RADV_DEBUG_NO_DYNAMIC_BOUNDS
;
516 static int radv_get_instance_extension_index(const char *name
)
518 for (unsigned i
= 0; i
< RADV_INSTANCE_EXTENSION_COUNT
; ++i
) {
519 if (strcmp(name
, radv_instance_extensions
[i
].extensionName
) == 0)
526 VkResult
radv_CreateInstance(
527 const VkInstanceCreateInfo
* pCreateInfo
,
528 const VkAllocationCallbacks
* pAllocator
,
529 VkInstance
* pInstance
)
531 struct radv_instance
*instance
;
534 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
536 uint32_t client_version
;
537 if (pCreateInfo
->pApplicationInfo
&&
538 pCreateInfo
->pApplicationInfo
->apiVersion
!= 0) {
539 client_version
= pCreateInfo
->pApplicationInfo
->apiVersion
;
541 radv_EnumerateInstanceVersion(&client_version
);
544 instance
= vk_zalloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
545 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
547 return vk_error(NULL
, VK_ERROR_OUT_OF_HOST_MEMORY
);
549 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
552 instance
->alloc
= *pAllocator
;
554 instance
->alloc
= default_alloc
;
556 instance
->apiVersion
= client_version
;
557 instance
->physicalDeviceCount
= -1;
559 instance
->debug_flags
= parse_debug_string(getenv("RADV_DEBUG"),
562 instance
->perftest_flags
= parse_debug_string(getenv("RADV_PERFTEST"),
563 radv_perftest_options
);
566 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
567 radv_logi("Created an instance");
569 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
570 const char *ext_name
= pCreateInfo
->ppEnabledExtensionNames
[i
];
571 int index
= radv_get_instance_extension_index(ext_name
);
573 if (index
< 0 || !radv_supported_instance_extensions
.extensions
[index
]) {
574 vk_free2(&default_alloc
, pAllocator
, instance
);
575 return vk_error(instance
, VK_ERROR_EXTENSION_NOT_PRESENT
);
578 instance
->enabled_extensions
.extensions
[index
] = true;
581 result
= vk_debug_report_instance_init(&instance
->debug_report_callbacks
);
582 if (result
!= VK_SUCCESS
) {
583 vk_free2(&default_alloc
, pAllocator
, instance
);
584 return vk_error(instance
, result
);
589 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
591 radv_handle_per_app_options(instance
, pCreateInfo
->pApplicationInfo
);
593 *pInstance
= radv_instance_to_handle(instance
);
598 void radv_DestroyInstance(
599 VkInstance _instance
,
600 const VkAllocationCallbacks
* pAllocator
)
602 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
607 for (int i
= 0; i
< instance
->physicalDeviceCount
; ++i
) {
608 radv_physical_device_finish(instance
->physicalDevices
+ i
);
611 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
615 vk_debug_report_instance_destroy(&instance
->debug_report_callbacks
);
617 vk_free(&instance
->alloc
, instance
);
621 radv_enumerate_devices(struct radv_instance
*instance
)
623 /* TODO: Check for more devices ? */
624 drmDevicePtr devices
[8];
625 VkResult result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
628 instance
->physicalDeviceCount
= 0;
630 max_devices
= drmGetDevices2(0, devices
, ARRAY_SIZE(devices
));
632 if (instance
->debug_flags
& RADV_DEBUG_STARTUP
)
633 radv_logi("Found %d drm nodes", max_devices
);
636 return vk_error(instance
, VK_ERROR_INCOMPATIBLE_DRIVER
);
638 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
639 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
640 devices
[i
]->bustype
== DRM_BUS_PCI
&&
641 devices
[i
]->deviceinfo
.pci
->vendor_id
== ATI_VENDOR_ID
) {
643 result
= radv_physical_device_init(instance
->physicalDevices
+
644 instance
->physicalDeviceCount
,
647 if (result
== VK_SUCCESS
)
648 ++instance
->physicalDeviceCount
;
649 else if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
653 drmFreeDevices(devices
, max_devices
);
658 VkResult
radv_EnumeratePhysicalDevices(
659 VkInstance _instance
,
660 uint32_t* pPhysicalDeviceCount
,
661 VkPhysicalDevice
* pPhysicalDevices
)
663 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
666 if (instance
->physicalDeviceCount
< 0) {
667 result
= radv_enumerate_devices(instance
);
668 if (result
!= VK_SUCCESS
&&
669 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
673 if (!pPhysicalDevices
) {
674 *pPhysicalDeviceCount
= instance
->physicalDeviceCount
;
676 *pPhysicalDeviceCount
= MIN2(*pPhysicalDeviceCount
, instance
->physicalDeviceCount
);
677 for (unsigned i
= 0; i
< *pPhysicalDeviceCount
; ++i
)
678 pPhysicalDevices
[i
] = radv_physical_device_to_handle(instance
->physicalDevices
+ i
);
681 return *pPhysicalDeviceCount
< instance
->physicalDeviceCount
? VK_INCOMPLETE
685 VkResult
radv_EnumeratePhysicalDeviceGroups(
686 VkInstance _instance
,
687 uint32_t* pPhysicalDeviceGroupCount
,
688 VkPhysicalDeviceGroupProperties
* pPhysicalDeviceGroupProperties
)
690 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
693 if (instance
->physicalDeviceCount
< 0) {
694 result
= radv_enumerate_devices(instance
);
695 if (result
!= VK_SUCCESS
&&
696 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
700 if (!pPhysicalDeviceGroupProperties
) {
701 *pPhysicalDeviceGroupCount
= instance
->physicalDeviceCount
;
703 *pPhysicalDeviceGroupCount
= MIN2(*pPhysicalDeviceGroupCount
, instance
->physicalDeviceCount
);
704 for (unsigned i
= 0; i
< *pPhysicalDeviceGroupCount
; ++i
) {
705 pPhysicalDeviceGroupProperties
[i
].physicalDeviceCount
= 1;
706 pPhysicalDeviceGroupProperties
[i
].physicalDevices
[0] = radv_physical_device_to_handle(instance
->physicalDevices
+ i
);
707 pPhysicalDeviceGroupProperties
[i
].subsetAllocation
= false;
710 return *pPhysicalDeviceGroupCount
< instance
->physicalDeviceCount
? VK_INCOMPLETE
714 void radv_GetPhysicalDeviceFeatures(
715 VkPhysicalDevice physicalDevice
,
716 VkPhysicalDeviceFeatures
* pFeatures
)
718 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
719 memset(pFeatures
, 0, sizeof(*pFeatures
));
721 *pFeatures
= (VkPhysicalDeviceFeatures
) {
722 .robustBufferAccess
= true,
723 .fullDrawIndexUint32
= true,
724 .imageCubeArray
= true,
725 .independentBlend
= true,
726 .geometryShader
= true,
727 .tessellationShader
= true,
728 .sampleRateShading
= true,
729 .dualSrcBlend
= true,
731 .multiDrawIndirect
= true,
732 .drawIndirectFirstInstance
= true,
734 .depthBiasClamp
= true,
735 .fillModeNonSolid
= true,
740 .multiViewport
= true,
741 .samplerAnisotropy
= true,
742 .textureCompressionETC2
= pdevice
->rad_info
.chip_class
>= GFX9
||
743 pdevice
->rad_info
.family
== CHIP_STONEY
,
744 .textureCompressionASTC_LDR
= false,
745 .textureCompressionBC
= true,
746 .occlusionQueryPrecise
= true,
747 .pipelineStatisticsQuery
= true,
748 .vertexPipelineStoresAndAtomics
= true,
749 .fragmentStoresAndAtomics
= true,
750 .shaderTessellationAndGeometryPointSize
= true,
751 .shaderImageGatherExtended
= true,
752 .shaderStorageImageExtendedFormats
= true,
753 .shaderStorageImageMultisample
= false,
754 .shaderUniformBufferArrayDynamicIndexing
= true,
755 .shaderSampledImageArrayDynamicIndexing
= true,
756 .shaderStorageBufferArrayDynamicIndexing
= true,
757 .shaderStorageImageArrayDynamicIndexing
= true,
758 .shaderStorageImageReadWithoutFormat
= true,
759 .shaderStorageImageWriteWithoutFormat
= true,
760 .shaderClipDistance
= true,
761 .shaderCullDistance
= true,
762 .shaderFloat64
= true,
765 .sparseBinding
= true,
766 .variableMultisampleRate
= true,
767 .inheritedQueries
= true,
771 void radv_GetPhysicalDeviceFeatures2(
772 VkPhysicalDevice physicalDevice
,
773 VkPhysicalDeviceFeatures2KHR
*pFeatures
)
775 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
776 vk_foreach_struct(ext
, pFeatures
->pNext
) {
777 switch (ext
->sType
) {
778 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTER_FEATURES_KHR
: {
779 VkPhysicalDeviceVariablePointerFeaturesKHR
*features
= (void *)ext
;
780 features
->variablePointersStorageBuffer
= true;
781 features
->variablePointers
= false;
784 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES_KHR
: {
785 VkPhysicalDeviceMultiviewFeaturesKHR
*features
= (VkPhysicalDeviceMultiviewFeaturesKHR
*)ext
;
786 features
->multiview
= true;
787 features
->multiviewGeometryShader
= true;
788 features
->multiviewTessellationShader
= true;
791 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETER_FEATURES
: {
792 VkPhysicalDeviceShaderDrawParameterFeatures
*features
=
793 (VkPhysicalDeviceShaderDrawParameterFeatures
*)ext
;
794 features
->shaderDrawParameters
= true;
797 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES
: {
798 VkPhysicalDeviceProtectedMemoryFeatures
*features
=
799 (VkPhysicalDeviceProtectedMemoryFeatures
*)ext
;
800 features
->protectedMemory
= false;
803 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES
: {
804 VkPhysicalDevice16BitStorageFeatures
*features
=
805 (VkPhysicalDevice16BitStorageFeatures
*)ext
;
806 bool enabled
= HAVE_LLVM
>= 0x0700 && pdevice
->rad_info
.chip_class
>= VI
;
807 features
->storageBuffer16BitAccess
= enabled
;
808 features
->uniformAndStorageBuffer16BitAccess
= enabled
;
809 features
->storagePushConstant16
= enabled
;
810 features
->storageInputOutput16
= enabled
;
813 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES
: {
814 VkPhysicalDeviceSamplerYcbcrConversionFeatures
*features
=
815 (VkPhysicalDeviceSamplerYcbcrConversionFeatures
*)ext
;
816 features
->samplerYcbcrConversion
= false;
819 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT
: {
820 VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*features
=
821 (VkPhysicalDeviceDescriptorIndexingFeaturesEXT
*)ext
;
822 features
->shaderInputAttachmentArrayDynamicIndexing
= true;
823 features
->shaderUniformTexelBufferArrayDynamicIndexing
= true;
824 features
->shaderStorageTexelBufferArrayDynamicIndexing
= true;
825 features
->shaderUniformBufferArrayNonUniformIndexing
= false;
826 features
->shaderSampledImageArrayNonUniformIndexing
= false;
827 features
->shaderStorageBufferArrayNonUniformIndexing
= false;
828 features
->shaderStorageImageArrayNonUniformIndexing
= false;
829 features
->shaderInputAttachmentArrayNonUniformIndexing
= false;
830 features
->shaderUniformTexelBufferArrayNonUniformIndexing
= false;
831 features
->shaderStorageTexelBufferArrayNonUniformIndexing
= false;
832 features
->descriptorBindingUniformBufferUpdateAfterBind
= true;
833 features
->descriptorBindingSampledImageUpdateAfterBind
= true;
834 features
->descriptorBindingStorageImageUpdateAfterBind
= true;
835 features
->descriptorBindingStorageBufferUpdateAfterBind
= true;
836 features
->descriptorBindingUniformTexelBufferUpdateAfterBind
= true;
837 features
->descriptorBindingStorageTexelBufferUpdateAfterBind
= true;
838 features
->descriptorBindingUpdateUnusedWhilePending
= true;
839 features
->descriptorBindingPartiallyBound
= true;
840 features
->descriptorBindingVariableDescriptorCount
= true;
841 features
->runtimeDescriptorArray
= true;
844 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT
: {
845 VkPhysicalDeviceConditionalRenderingFeaturesEXT
*features
=
846 (VkPhysicalDeviceConditionalRenderingFeaturesEXT
*)ext
;
847 features
->conditionalRendering
= true;
848 features
->inheritedConditionalRendering
= false;
851 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT
: {
852 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*features
=
853 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT
*)ext
;
854 features
->vertexAttributeInstanceRateDivisor
= VK_TRUE
;
855 features
->vertexAttributeInstanceRateZeroDivisor
= VK_TRUE
;
862 return radv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
865 void radv_GetPhysicalDeviceProperties(
866 VkPhysicalDevice physicalDevice
,
867 VkPhysicalDeviceProperties
* pProperties
)
869 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
870 VkSampleCountFlags sample_counts
= 0xf;
872 /* make sure that the entire descriptor set is addressable with a signed
873 * 32-bit int. So the sum of all limits scaled by descriptor size has to
874 * be at most 2 GiB. the combined image & samples object count as one of
875 * both. This limit is for the pipeline layout, not for the set layout, but
876 * there is no set limit, so we just set a pipeline limit. I don't think
877 * any app is going to hit this soon. */
878 size_t max_descriptor_set_size
= ((1ull << 31) - 16 * MAX_DYNAMIC_BUFFERS
) /
879 (32 /* uniform buffer, 32 due to potential space wasted on alignment */ +
880 32 /* storage buffer, 32 due to potential space wasted on alignment */ +
881 32 /* sampler, largest when combined with image */ +
882 64 /* sampled image */ +
883 64 /* storage image */);
885 VkPhysicalDeviceLimits limits
= {
886 .maxImageDimension1D
= (1 << 14),
887 .maxImageDimension2D
= (1 << 14),
888 .maxImageDimension3D
= (1 << 11),
889 .maxImageDimensionCube
= (1 << 14),
890 .maxImageArrayLayers
= (1 << 11),
891 .maxTexelBufferElements
= 128 * 1024 * 1024,
892 .maxUniformBufferRange
= UINT32_MAX
,
893 .maxStorageBufferRange
= UINT32_MAX
,
894 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
895 .maxMemoryAllocationCount
= UINT32_MAX
,
896 .maxSamplerAllocationCount
= 64 * 1024,
897 .bufferImageGranularity
= 64, /* A cache line */
898 .sparseAddressSpaceSize
= 0xffffffffu
, /* buffer max size */
899 .maxBoundDescriptorSets
= MAX_SETS
,
900 .maxPerStageDescriptorSamplers
= max_descriptor_set_size
,
901 .maxPerStageDescriptorUniformBuffers
= max_descriptor_set_size
,
902 .maxPerStageDescriptorStorageBuffers
= max_descriptor_set_size
,
903 .maxPerStageDescriptorSampledImages
= max_descriptor_set_size
,
904 .maxPerStageDescriptorStorageImages
= max_descriptor_set_size
,
905 .maxPerStageDescriptorInputAttachments
= max_descriptor_set_size
,
906 .maxPerStageResources
= max_descriptor_set_size
,
907 .maxDescriptorSetSamplers
= max_descriptor_set_size
,
908 .maxDescriptorSetUniformBuffers
= max_descriptor_set_size
,
909 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_UNIFORM_BUFFERS
,
910 .maxDescriptorSetStorageBuffers
= max_descriptor_set_size
,
911 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_STORAGE_BUFFERS
,
912 .maxDescriptorSetSampledImages
= max_descriptor_set_size
,
913 .maxDescriptorSetStorageImages
= max_descriptor_set_size
,
914 .maxDescriptorSetInputAttachments
= max_descriptor_set_size
,
915 .maxVertexInputAttributes
= 32,
916 .maxVertexInputBindings
= 32,
917 .maxVertexInputAttributeOffset
= 2047,
918 .maxVertexInputBindingStride
= 2048,
919 .maxVertexOutputComponents
= 128,
920 .maxTessellationGenerationLevel
= 64,
921 .maxTessellationPatchSize
= 32,
922 .maxTessellationControlPerVertexInputComponents
= 128,
923 .maxTessellationControlPerVertexOutputComponents
= 128,
924 .maxTessellationControlPerPatchOutputComponents
= 120,
925 .maxTessellationControlTotalOutputComponents
= 4096,
926 .maxTessellationEvaluationInputComponents
= 128,
927 .maxTessellationEvaluationOutputComponents
= 128,
928 .maxGeometryShaderInvocations
= 127,
929 .maxGeometryInputComponents
= 64,
930 .maxGeometryOutputComponents
= 128,
931 .maxGeometryOutputVertices
= 256,
932 .maxGeometryTotalOutputComponents
= 1024,
933 .maxFragmentInputComponents
= 128,
934 .maxFragmentOutputAttachments
= 8,
935 .maxFragmentDualSrcAttachments
= 1,
936 .maxFragmentCombinedOutputResources
= 8,
937 .maxComputeSharedMemorySize
= 32768,
938 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
939 .maxComputeWorkGroupInvocations
= 2048,
940 .maxComputeWorkGroupSize
= {
945 .subPixelPrecisionBits
= 4 /* FIXME */,
946 .subTexelPrecisionBits
= 4 /* FIXME */,
947 .mipmapPrecisionBits
= 4 /* FIXME */,
948 .maxDrawIndexedIndexValue
= UINT32_MAX
,
949 .maxDrawIndirectCount
= UINT32_MAX
,
950 .maxSamplerLodBias
= 16,
951 .maxSamplerAnisotropy
= 16,
952 .maxViewports
= MAX_VIEWPORTS
,
953 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
954 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
955 .viewportSubPixelBits
= 8,
956 .minMemoryMapAlignment
= 4096, /* A page */
957 .minTexelBufferOffsetAlignment
= 1,
958 .minUniformBufferOffsetAlignment
= 4,
959 .minStorageBufferOffsetAlignment
= 4,
960 .minTexelOffset
= -32,
961 .maxTexelOffset
= 31,
962 .minTexelGatherOffset
= -32,
963 .maxTexelGatherOffset
= 31,
964 .minInterpolationOffset
= -2,
965 .maxInterpolationOffset
= 2,
966 .subPixelInterpolationOffsetBits
= 8,
967 .maxFramebufferWidth
= (1 << 14),
968 .maxFramebufferHeight
= (1 << 14),
969 .maxFramebufferLayers
= (1 << 10),
970 .framebufferColorSampleCounts
= sample_counts
,
971 .framebufferDepthSampleCounts
= sample_counts
,
972 .framebufferStencilSampleCounts
= sample_counts
,
973 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
974 .maxColorAttachments
= MAX_RTS
,
975 .sampledImageColorSampleCounts
= sample_counts
,
976 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
977 .sampledImageDepthSampleCounts
= sample_counts
,
978 .sampledImageStencilSampleCounts
= sample_counts
,
979 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
980 .maxSampleMaskWords
= 1,
981 .timestampComputeAndGraphics
= true,
982 .timestampPeriod
= 1000000.0 / pdevice
->rad_info
.clock_crystal_freq
,
983 .maxClipDistances
= 8,
984 .maxCullDistances
= 8,
985 .maxCombinedClipAndCullDistances
= 8,
986 .discreteQueuePriorities
= 1,
987 .pointSizeRange
= { 0.125, 255.875 },
988 .lineWidthRange
= { 0.0, 7.9921875 },
989 .pointSizeGranularity
= (1.0 / 8.0),
990 .lineWidthGranularity
= (1.0 / 128.0),
991 .strictLines
= false, /* FINISHME */
992 .standardSampleLocations
= true,
993 .optimalBufferCopyOffsetAlignment
= 128,
994 .optimalBufferCopyRowPitchAlignment
= 128,
995 .nonCoherentAtomSize
= 64,
998 *pProperties
= (VkPhysicalDeviceProperties
) {
999 .apiVersion
= radv_physical_device_api_version(pdevice
),
1000 .driverVersion
= vk_get_driver_version(),
1001 .vendorID
= ATI_VENDOR_ID
,
1002 .deviceID
= pdevice
->rad_info
.pci_id
,
1003 .deviceType
= pdevice
->rad_info
.has_dedicated_vram
? VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU
: VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
1005 .sparseProperties
= {0},
1008 strcpy(pProperties
->deviceName
, pdevice
->name
);
1009 memcpy(pProperties
->pipelineCacheUUID
, pdevice
->cache_uuid
, VK_UUID_SIZE
);
1012 void radv_GetPhysicalDeviceProperties2(
1013 VkPhysicalDevice physicalDevice
,
1014 VkPhysicalDeviceProperties2KHR
*pProperties
)
1016 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
1017 radv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
1019 vk_foreach_struct(ext
, pProperties
->pNext
) {
1020 switch (ext
->sType
) {
1021 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
1022 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
1023 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
1024 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
1027 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES_KHR
: {
1028 VkPhysicalDeviceIDPropertiesKHR
*properties
= (VkPhysicalDeviceIDPropertiesKHR
*)ext
;
1029 memcpy(properties
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
1030 memcpy(properties
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
1031 properties
->deviceLUIDValid
= false;
1034 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES_KHR
: {
1035 VkPhysicalDeviceMultiviewPropertiesKHR
*properties
= (VkPhysicalDeviceMultiviewPropertiesKHR
*)ext
;
1036 properties
->maxMultiviewViewCount
= MAX_VIEWS
;
1037 properties
->maxMultiviewInstanceIndex
= INT_MAX
;
1040 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES_KHR
: {
1041 VkPhysicalDevicePointClippingPropertiesKHR
*properties
=
1042 (VkPhysicalDevicePointClippingPropertiesKHR
*)ext
;
1043 properties
->pointClippingBehavior
= VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES_KHR
;
1046 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DISCARD_RECTANGLE_PROPERTIES_EXT
: {
1047 VkPhysicalDeviceDiscardRectanglePropertiesEXT
*properties
=
1048 (VkPhysicalDeviceDiscardRectanglePropertiesEXT
*)ext
;
1049 properties
->maxDiscardRectangles
= MAX_DISCARD_RECTANGLES
;
1052 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT
: {
1053 VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*properties
=
1054 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT
*) ext
;
1055 properties
->minImportedHostPointerAlignment
= 4096;
1058 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES
: {
1059 VkPhysicalDeviceSubgroupProperties
*properties
=
1060 (VkPhysicalDeviceSubgroupProperties
*)ext
;
1061 properties
->subgroupSize
= 64;
1062 properties
->supportedStages
= VK_SHADER_STAGE_ALL
;
1063 properties
->supportedOperations
=
1064 VK_SUBGROUP_FEATURE_ARITHMETIC_BIT
|
1065 VK_SUBGROUP_FEATURE_BASIC_BIT
|
1066 VK_SUBGROUP_FEATURE_BALLOT_BIT
|
1067 VK_SUBGROUP_FEATURE_QUAD_BIT
|
1068 VK_SUBGROUP_FEATURE_VOTE_BIT
;
1069 if (pdevice
->rad_info
.chip_class
>= VI
) {
1070 properties
->supportedOperations
|=
1071 VK_SUBGROUP_FEATURE_SHUFFLE_BIT
|
1072 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT
;
1074 properties
->quadOperationsInAllStages
= true;
1077 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES
: {
1078 VkPhysicalDeviceMaintenance3Properties
*properties
=
1079 (VkPhysicalDeviceMaintenance3Properties
*)ext
;
1080 /* Make sure everything is addressable by a signed 32-bit int, and
1081 * our largest descriptors are 96 bytes. */
1082 properties
->maxPerSetDescriptors
= (1ull << 31) / 96;
1083 /* Our buffer size fields allow only this much */
1084 properties
->maxMemoryAllocationSize
= 0xFFFFFFFFull
;
1087 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_FILTER_MINMAX_PROPERTIES_EXT
: {
1088 VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*properties
=
1089 (VkPhysicalDeviceSamplerFilterMinmaxPropertiesEXT
*)ext
;
1090 /* GFX6-8 only support single channel min/max filter. */
1091 properties
->filterMinmaxImageComponentMapping
= pdevice
->rad_info
.chip_class
>= GFX9
;
1092 properties
->filterMinmaxSingleComponentFormats
= true;
1095 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CORE_PROPERTIES_AMD
: {
1096 VkPhysicalDeviceShaderCorePropertiesAMD
*properties
=
1097 (VkPhysicalDeviceShaderCorePropertiesAMD
*)ext
;
1099 /* Shader engines. */
1100 properties
->shaderEngineCount
=
1101 pdevice
->rad_info
.max_se
;
1102 properties
->shaderArraysPerEngineCount
=
1103 pdevice
->rad_info
.max_sh_per_se
;
1104 properties
->computeUnitsPerShaderArray
=
1105 pdevice
->rad_info
.num_good_compute_units
/
1106 (pdevice
->rad_info
.max_se
*
1107 pdevice
->rad_info
.max_sh_per_se
);
1108 properties
->simdPerComputeUnit
= 4;
1109 properties
->wavefrontsPerSimd
=
1110 pdevice
->rad_info
.family
== CHIP_TONGA
||
1111 pdevice
->rad_info
.family
== CHIP_ICELAND
||
1112 pdevice
->rad_info
.family
== CHIP_POLARIS10
||
1113 pdevice
->rad_info
.family
== CHIP_POLARIS11
||
1114 pdevice
->rad_info
.family
== CHIP_POLARIS12
||
1115 pdevice
->rad_info
.family
== CHIP_VEGAM
? 8 : 10;
1116 properties
->wavefrontSize
= 64;
1119 properties
->sgprsPerSimd
=
1120 radv_get_num_physical_sgprs(pdevice
);
1121 properties
->minSgprAllocation
=
1122 pdevice
->rad_info
.chip_class
>= VI
? 16 : 8;
1123 properties
->maxSgprAllocation
=
1124 pdevice
->rad_info
.family
== CHIP_TONGA
||
1125 pdevice
->rad_info
.family
== CHIP_ICELAND
? 96 : 104;
1126 properties
->sgprAllocationGranularity
=
1127 pdevice
->rad_info
.chip_class
>= VI
? 16 : 8;
1130 properties
->vgprsPerSimd
= RADV_NUM_PHYSICAL_VGPRS
;
1131 properties
->minVgprAllocation
= 4;
1132 properties
->maxVgprAllocation
= 256;
1133 properties
->vgprAllocationGranularity
= 4;
1136 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT
: {
1137 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*properties
=
1138 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT
*)ext
;
1139 properties
->maxVertexAttribDivisor
= UINT32_MAX
;
1142 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES_EXT
: {
1143 VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*properties
=
1144 (VkPhysicalDeviceDescriptorIndexingPropertiesEXT
*)ext
;
1145 properties
->maxUpdateAfterBindDescriptorsInAllPools
= UINT32_MAX
/ 64;
1146 properties
->shaderUniformBufferArrayNonUniformIndexingNative
= false;
1147 properties
->shaderSampledImageArrayNonUniformIndexingNative
= false;
1148 properties
->shaderStorageBufferArrayNonUniformIndexingNative
= false;
1149 properties
->shaderStorageImageArrayNonUniformIndexingNative
= false;
1150 properties
->shaderInputAttachmentArrayNonUniformIndexingNative
= false;
1151 properties
->robustBufferAccessUpdateAfterBind
= false;
1152 properties
->quadDivergentImplicitLod
= false;
1154 size_t max_descriptor_set_size
= ((1ull << 31) - 16 * MAX_DYNAMIC_BUFFERS
) /
1155 (32 /* uniform buffer, 32 due to potential space wasted on alignment */ +
1156 32 /* storage buffer, 32 due to potential space wasted on alignment */ +
1157 32 /* sampler, largest when combined with image */ +
1158 64 /* sampled image */ +
1159 64 /* storage image */);
1160 properties
->maxPerStageDescriptorUpdateAfterBindSamplers
= max_descriptor_set_size
;
1161 properties
->maxPerStageDescriptorUpdateAfterBindUniformBuffers
= max_descriptor_set_size
;
1162 properties
->maxPerStageDescriptorUpdateAfterBindStorageBuffers
= max_descriptor_set_size
;
1163 properties
->maxPerStageDescriptorUpdateAfterBindSampledImages
= max_descriptor_set_size
;
1164 properties
->maxPerStageDescriptorUpdateAfterBindStorageImages
= max_descriptor_set_size
;
1165 properties
->maxPerStageDescriptorUpdateAfterBindInputAttachments
= max_descriptor_set_size
;
1166 properties
->maxPerStageUpdateAfterBindResources
= max_descriptor_set_size
;
1167 properties
->maxDescriptorSetUpdateAfterBindSamplers
= max_descriptor_set_size
;
1168 properties
->maxDescriptorSetUpdateAfterBindUniformBuffers
= max_descriptor_set_size
;
1169 properties
->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic
= MAX_DYNAMIC_UNIFORM_BUFFERS
;
1170 properties
->maxDescriptorSetUpdateAfterBindStorageBuffers
= max_descriptor_set_size
;
1171 properties
->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic
= MAX_DYNAMIC_STORAGE_BUFFERS
;
1172 properties
->maxDescriptorSetUpdateAfterBindSampledImages
= max_descriptor_set_size
;
1173 properties
->maxDescriptorSetUpdateAfterBindStorageImages
= max_descriptor_set_size
;
1174 properties
->maxDescriptorSetUpdateAfterBindInputAttachments
= max_descriptor_set_size
;
1177 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES
: {
1178 VkPhysicalDeviceProtectedMemoryProperties
*properties
=
1179 (VkPhysicalDeviceProtectedMemoryProperties
*)ext
;
1180 properties
->protectedNoFault
= false;
1183 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONSERVATIVE_RASTERIZATION_PROPERTIES_EXT
: {
1184 VkPhysicalDeviceConservativeRasterizationPropertiesEXT
*properties
=
1185 (VkPhysicalDeviceConservativeRasterizationPropertiesEXT
*)ext
;
1186 properties
->primitiveOverestimationSize
= 0;
1187 properties
->maxExtraPrimitiveOverestimationSize
= 0;
1188 properties
->extraPrimitiveOverestimationSizeGranularity
= 0;
1189 properties
->primitiveUnderestimation
= VK_FALSE
;
1190 properties
->conservativePointAndLineRasterization
= VK_FALSE
;
1191 properties
->degenerateTrianglesRasterized
= VK_FALSE
;
1192 properties
->degenerateLinesRasterized
= VK_FALSE
;
1193 properties
->fullyCoveredFragmentShaderInputVariable
= VK_FALSE
;
1194 properties
->conservativeRasterizationPostDepthCoverage
= VK_FALSE
;
1203 static void radv_get_physical_device_queue_family_properties(
1204 struct radv_physical_device
* pdevice
,
1206 VkQueueFamilyProperties
** pQueueFamilyProperties
)
1208 int num_queue_families
= 1;
1210 if (pdevice
->rad_info
.num_compute_rings
> 0 &&
1211 !(pdevice
->instance
->debug_flags
& RADV_DEBUG_NO_COMPUTE_QUEUE
))
1212 num_queue_families
++;
1214 if (pQueueFamilyProperties
== NULL
) {
1215 *pCount
= num_queue_families
;
1224 *pQueueFamilyProperties
[idx
] = (VkQueueFamilyProperties
) {
1225 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
1226 VK_QUEUE_COMPUTE_BIT
|
1227 VK_QUEUE_TRANSFER_BIT
|
1228 VK_QUEUE_SPARSE_BINDING_BIT
,
1230 .timestampValidBits
= 64,
1231 .minImageTransferGranularity
= (VkExtent3D
) { 1, 1, 1 },
1236 if (pdevice
->rad_info
.num_compute_rings
> 0 &&
1237 !(pdevice
->instance
->debug_flags
& RADV_DEBUG_NO_COMPUTE_QUEUE
)) {
1238 if (*pCount
> idx
) {
1239 *pQueueFamilyProperties
[idx
] = (VkQueueFamilyProperties
) {
1240 .queueFlags
= VK_QUEUE_COMPUTE_BIT
|
1241 VK_QUEUE_TRANSFER_BIT
|
1242 VK_QUEUE_SPARSE_BINDING_BIT
,
1243 .queueCount
= pdevice
->rad_info
.num_compute_rings
,
1244 .timestampValidBits
= 64,
1245 .minImageTransferGranularity
= (VkExtent3D
) { 1, 1, 1 },
1253 void radv_GetPhysicalDeviceQueueFamilyProperties(
1254 VkPhysicalDevice physicalDevice
,
1256 VkQueueFamilyProperties
* pQueueFamilyProperties
)
1258 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
1259 if (!pQueueFamilyProperties
) {
1260 return radv_get_physical_device_queue_family_properties(pdevice
, pCount
, NULL
);
1263 VkQueueFamilyProperties
*properties
[] = {
1264 pQueueFamilyProperties
+ 0,
1265 pQueueFamilyProperties
+ 1,
1266 pQueueFamilyProperties
+ 2,
1268 radv_get_physical_device_queue_family_properties(pdevice
, pCount
, properties
);
1269 assert(*pCount
<= 3);
1272 void radv_GetPhysicalDeviceQueueFamilyProperties2(
1273 VkPhysicalDevice physicalDevice
,
1275 VkQueueFamilyProperties2KHR
*pQueueFamilyProperties
)
1277 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
1278 if (!pQueueFamilyProperties
) {
1279 return radv_get_physical_device_queue_family_properties(pdevice
, pCount
, NULL
);
1282 VkQueueFamilyProperties
*properties
[] = {
1283 &pQueueFamilyProperties
[0].queueFamilyProperties
,
1284 &pQueueFamilyProperties
[1].queueFamilyProperties
,
1285 &pQueueFamilyProperties
[2].queueFamilyProperties
,
1287 radv_get_physical_device_queue_family_properties(pdevice
, pCount
, properties
);
1288 assert(*pCount
<= 3);
1291 void radv_GetPhysicalDeviceMemoryProperties(
1292 VkPhysicalDevice physicalDevice
,
1293 VkPhysicalDeviceMemoryProperties
*pMemoryProperties
)
1295 RADV_FROM_HANDLE(radv_physical_device
, physical_device
, physicalDevice
);
1297 *pMemoryProperties
= physical_device
->memory_properties
;
1300 void radv_GetPhysicalDeviceMemoryProperties2(
1301 VkPhysicalDevice physicalDevice
,
1302 VkPhysicalDeviceMemoryProperties2KHR
*pMemoryProperties
)
1304 return radv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
1305 &pMemoryProperties
->memoryProperties
);
1308 VkResult
radv_GetMemoryHostPointerPropertiesEXT(
1310 VkExternalMemoryHandleTypeFlagBitsKHR handleType
,
1311 const void *pHostPointer
,
1312 VkMemoryHostPointerPropertiesEXT
*pMemoryHostPointerProperties
)
1314 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1318 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
: {
1319 const struct radv_physical_device
*physical_device
= device
->physical_device
;
1320 uint32_t memoryTypeBits
= 0;
1321 for (int i
= 0; i
< physical_device
->memory_properties
.memoryTypeCount
; i
++) {
1322 if (physical_device
->mem_type_indices
[i
] == RADV_MEM_TYPE_GTT_CACHED
) {
1323 memoryTypeBits
= (1 << i
);
1327 pMemoryHostPointerProperties
->memoryTypeBits
= memoryTypeBits
;
1331 return VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
;
1335 static enum radeon_ctx_priority
1336 radv_get_queue_global_priority(const VkDeviceQueueGlobalPriorityCreateInfoEXT
*pObj
)
1338 /* Default to MEDIUM when a specific global priority isn't requested */
1340 return RADEON_CTX_PRIORITY_MEDIUM
;
1342 switch(pObj
->globalPriority
) {
1343 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT
:
1344 return RADEON_CTX_PRIORITY_REALTIME
;
1345 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT
:
1346 return RADEON_CTX_PRIORITY_HIGH
;
1347 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT
:
1348 return RADEON_CTX_PRIORITY_MEDIUM
;
1349 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT
:
1350 return RADEON_CTX_PRIORITY_LOW
;
1352 unreachable("Illegal global priority value");
1353 return RADEON_CTX_PRIORITY_INVALID
;
1358 radv_queue_init(struct radv_device
*device
, struct radv_queue
*queue
,
1359 uint32_t queue_family_index
, int idx
,
1360 VkDeviceQueueCreateFlags flags
,
1361 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*global_priority
)
1363 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1364 queue
->device
= device
;
1365 queue
->queue_family_index
= queue_family_index
;
1366 queue
->queue_idx
= idx
;
1367 queue
->priority
= radv_get_queue_global_priority(global_priority
);
1368 queue
->flags
= flags
;
1370 queue
->hw_ctx
= device
->ws
->ctx_create(device
->ws
, queue
->priority
);
1372 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
1378 radv_queue_finish(struct radv_queue
*queue
)
1381 queue
->device
->ws
->ctx_destroy(queue
->hw_ctx
);
1383 if (queue
->initial_full_flush_preamble_cs
)
1384 queue
->device
->ws
->cs_destroy(queue
->initial_full_flush_preamble_cs
);
1385 if (queue
->initial_preamble_cs
)
1386 queue
->device
->ws
->cs_destroy(queue
->initial_preamble_cs
);
1387 if (queue
->continue_preamble_cs
)
1388 queue
->device
->ws
->cs_destroy(queue
->continue_preamble_cs
);
1389 if (queue
->descriptor_bo
)
1390 queue
->device
->ws
->buffer_destroy(queue
->descriptor_bo
);
1391 if (queue
->scratch_bo
)
1392 queue
->device
->ws
->buffer_destroy(queue
->scratch_bo
);
1393 if (queue
->esgs_ring_bo
)
1394 queue
->device
->ws
->buffer_destroy(queue
->esgs_ring_bo
);
1395 if (queue
->gsvs_ring_bo
)
1396 queue
->device
->ws
->buffer_destroy(queue
->gsvs_ring_bo
);
1397 if (queue
->tess_rings_bo
)
1398 queue
->device
->ws
->buffer_destroy(queue
->tess_rings_bo
);
1399 if (queue
->compute_scratch_bo
)
1400 queue
->device
->ws
->buffer_destroy(queue
->compute_scratch_bo
);
1404 radv_bo_list_init(struct radv_bo_list
*bo_list
)
1406 pthread_mutex_init(&bo_list
->mutex
, NULL
);
1407 bo_list
->list
.count
= bo_list
->capacity
= 0;
1408 bo_list
->list
.bos
= NULL
;
1412 radv_bo_list_finish(struct radv_bo_list
*bo_list
)
1414 free(bo_list
->list
.bos
);
1415 pthread_mutex_destroy(&bo_list
->mutex
);
1418 static VkResult
radv_bo_list_add(struct radv_device
*device
,
1419 struct radeon_winsys_bo
*bo
)
1421 struct radv_bo_list
*bo_list
= &device
->bo_list
;
1423 if (unlikely(!device
->use_global_bo_list
))
1426 pthread_mutex_lock(&bo_list
->mutex
);
1427 if (bo_list
->list
.count
== bo_list
->capacity
) {
1428 unsigned capacity
= MAX2(4, bo_list
->capacity
* 2);
1429 void *data
= realloc(bo_list
->list
.bos
, capacity
* sizeof(struct radeon_winsys_bo
*));
1432 pthread_mutex_unlock(&bo_list
->mutex
);
1433 return VK_ERROR_OUT_OF_HOST_MEMORY
;
1436 bo_list
->list
.bos
= (struct radeon_winsys_bo
**)data
;
1437 bo_list
->capacity
= capacity
;
1440 bo_list
->list
.bos
[bo_list
->list
.count
++] = bo
;
1441 pthread_mutex_unlock(&bo_list
->mutex
);
1445 static void radv_bo_list_remove(struct radv_device
*device
,
1446 struct radeon_winsys_bo
*bo
)
1448 struct radv_bo_list
*bo_list
= &device
->bo_list
;
1450 if (unlikely(!device
->use_global_bo_list
))
1453 pthread_mutex_lock(&bo_list
->mutex
);
1454 for(unsigned i
= 0; i
< bo_list
->list
.count
; ++i
) {
1455 if (bo_list
->list
.bos
[i
] == bo
) {
1456 bo_list
->list
.bos
[i
] = bo_list
->list
.bos
[bo_list
->list
.count
- 1];
1457 --bo_list
->list
.count
;
1461 pthread_mutex_unlock(&bo_list
->mutex
);
1465 radv_device_init_gs_info(struct radv_device
*device
)
1467 device
->gs_table_depth
= ac_get_gs_table_depth(device
->physical_device
->rad_info
.chip_class
,
1468 device
->physical_device
->rad_info
.family
);
1471 static int radv_get_device_extension_index(const char *name
)
1473 for (unsigned i
= 0; i
< RADV_DEVICE_EXTENSION_COUNT
; ++i
) {
1474 if (strcmp(name
, radv_device_extensions
[i
].extensionName
) == 0)
1481 radv_get_int_debug_option(const char *name
, int default_value
)
1488 result
= default_value
;
1492 result
= strtol(str
, &endptr
, 0);
1493 if (str
== endptr
) {
1494 /* No digits founs. */
1495 result
= default_value
;
1502 VkResult
radv_CreateDevice(
1503 VkPhysicalDevice physicalDevice
,
1504 const VkDeviceCreateInfo
* pCreateInfo
,
1505 const VkAllocationCallbacks
* pAllocator
,
1508 RADV_FROM_HANDLE(radv_physical_device
, physical_device
, physicalDevice
);
1510 struct radv_device
*device
;
1512 bool keep_shader_info
= false;
1514 /* Check enabled features */
1515 if (pCreateInfo
->pEnabledFeatures
) {
1516 VkPhysicalDeviceFeatures supported_features
;
1517 radv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
1518 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
1519 VkBool32
*enabled_feature
= (VkBool32
*)pCreateInfo
->pEnabledFeatures
;
1520 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
1521 for (uint32_t i
= 0; i
< num_features
; i
++) {
1522 if (enabled_feature
[i
] && !supported_feature
[i
])
1523 return vk_error(physical_device
->instance
, VK_ERROR_FEATURE_NOT_PRESENT
);
1527 device
= vk_zalloc2(&physical_device
->instance
->alloc
, pAllocator
,
1529 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
1531 return vk_error(physical_device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
1533 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1534 device
->instance
= physical_device
->instance
;
1535 device
->physical_device
= physical_device
;
1537 device
->ws
= physical_device
->ws
;
1539 device
->alloc
= *pAllocator
;
1541 device
->alloc
= physical_device
->instance
->alloc
;
1543 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
1544 const char *ext_name
= pCreateInfo
->ppEnabledExtensionNames
[i
];
1545 int index
= radv_get_device_extension_index(ext_name
);
1546 if (index
< 0 || !physical_device
->supported_extensions
.extensions
[index
]) {
1547 vk_free(&device
->alloc
, device
);
1548 return vk_error(physical_device
->instance
, VK_ERROR_EXTENSION_NOT_PRESENT
);
1551 device
->enabled_extensions
.extensions
[index
] = true;
1554 keep_shader_info
= device
->enabled_extensions
.AMD_shader_info
;
1556 /* With update after bind we can't attach bo's to the command buffer
1557 * from the descriptor set anymore, so we have to use a global BO list.
1559 device
->use_global_bo_list
=
1560 device
->enabled_extensions
.EXT_descriptor_indexing
;
1562 mtx_init(&device
->shader_slab_mutex
, mtx_plain
);
1563 list_inithead(&device
->shader_slabs
);
1565 radv_bo_list_init(&device
->bo_list
);
1567 for (unsigned i
= 0; i
< pCreateInfo
->queueCreateInfoCount
; i
++) {
1568 const VkDeviceQueueCreateInfo
*queue_create
= &pCreateInfo
->pQueueCreateInfos
[i
];
1569 uint32_t qfi
= queue_create
->queueFamilyIndex
;
1570 const VkDeviceQueueGlobalPriorityCreateInfoEXT
*global_priority
=
1571 vk_find_struct_const(queue_create
->pNext
, DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT
);
1573 assert(!global_priority
|| device
->physical_device
->rad_info
.has_ctx_priority
);
1575 device
->queues
[qfi
] = vk_alloc(&device
->alloc
,
1576 queue_create
->queueCount
* sizeof(struct radv_queue
), 8, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
1577 if (!device
->queues
[qfi
]) {
1578 result
= VK_ERROR_OUT_OF_HOST_MEMORY
;
1582 memset(device
->queues
[qfi
], 0, queue_create
->queueCount
* sizeof(struct radv_queue
));
1584 device
->queue_count
[qfi
] = queue_create
->queueCount
;
1586 for (unsigned q
= 0; q
< queue_create
->queueCount
; q
++) {
1587 result
= radv_queue_init(device
, &device
->queues
[qfi
][q
],
1588 qfi
, q
, queue_create
->flags
,
1590 if (result
!= VK_SUCCESS
)
1595 device
->pbb_allowed
= device
->physical_device
->rad_info
.chip_class
>= GFX9
&&
1596 ((device
->instance
->perftest_flags
& RADV_PERFTEST_BINNING
) ||
1597 device
->physical_device
->rad_info
.family
== CHIP_RAVEN
);
1599 /* Disabled and not implemented for now. */
1600 device
->dfsm_allowed
= device
->pbb_allowed
&&
1601 device
->physical_device
->rad_info
.family
== CHIP_RAVEN
;
1604 device
->always_use_syncobj
= device
->physical_device
->rad_info
.has_syncobj_wait_for_submit
;
1607 /* The maximum number of scratch waves. Scratch space isn't divided
1608 * evenly between CUs. The number is only a function of the number of CUs.
1609 * We can decrease the constant to decrease the scratch buffer size.
1611 * sctx->scratch_waves must be >= the maximum possible size of
1612 * 1 threadgroup, so that the hw doesn't hang from being unable
1615 * The recommended value is 4 per CU at most. Higher numbers don't
1616 * bring much benefit, but they still occupy chip resources (think
1617 * async compute). I've seen ~2% performance difference between 4 and 32.
1619 uint32_t max_threads_per_block
= 2048;
1620 device
->scratch_waves
= MAX2(32 * physical_device
->rad_info
.num_good_compute_units
,
1621 max_threads_per_block
/ 64);
1623 device
->dispatch_initiator
= S_00B800_COMPUTE_SHADER_EN(1);
1625 if (device
->physical_device
->rad_info
.chip_class
>= CIK
) {
1626 /* If the KMD allows it (there is a KMD hw register for it),
1627 * allow launching waves out-of-order.
1629 device
->dispatch_initiator
|= S_00B800_ORDER_MODE(1);
1632 radv_device_init_gs_info(device
);
1634 device
->tess_offchip_block_dw_size
=
1635 device
->physical_device
->rad_info
.family
== CHIP_HAWAII
? 4096 : 8192;
1636 device
->has_distributed_tess
=
1637 device
->physical_device
->rad_info
.chip_class
>= VI
&&
1638 device
->physical_device
->rad_info
.max_se
>= 2;
1640 if (getenv("RADV_TRACE_FILE")) {
1641 const char *filename
= getenv("RADV_TRACE_FILE");
1643 keep_shader_info
= true;
1645 if (!radv_init_trace(device
))
1648 fprintf(stderr
, "*****************************************************************************\n");
1649 fprintf(stderr
, "* WARNING: RADV_TRACE_FILE is costly and should only be used for debugging! *\n");
1650 fprintf(stderr
, "*****************************************************************************\n");
1652 fprintf(stderr
, "Trace file will be dumped to %s\n", filename
);
1653 radv_dump_enabled_options(device
, stderr
);
1656 device
->keep_shader_info
= keep_shader_info
;
1658 result
= radv_device_init_meta(device
);
1659 if (result
!= VK_SUCCESS
)
1662 radv_device_init_msaa(device
);
1664 for (int family
= 0; family
< RADV_MAX_QUEUE_FAMILIES
; ++family
) {
1665 device
->empty_cs
[family
] = device
->ws
->cs_create(device
->ws
, family
);
1667 case RADV_QUEUE_GENERAL
:
1668 radeon_emit(device
->empty_cs
[family
], PKT3(PKT3_CONTEXT_CONTROL
, 1, 0));
1669 radeon_emit(device
->empty_cs
[family
], CONTEXT_CONTROL_LOAD_ENABLE(1));
1670 radeon_emit(device
->empty_cs
[family
], CONTEXT_CONTROL_SHADOW_ENABLE(1));
1672 case RADV_QUEUE_COMPUTE
:
1673 radeon_emit(device
->empty_cs
[family
], PKT3(PKT3_NOP
, 0, 0));
1674 radeon_emit(device
->empty_cs
[family
], 0);
1677 device
->ws
->cs_finalize(device
->empty_cs
[family
]);
1680 if (device
->physical_device
->rad_info
.chip_class
>= CIK
)
1681 cik_create_gfx_config(device
);
1683 VkPipelineCacheCreateInfo ci
;
1684 ci
.sType
= VK_STRUCTURE_TYPE_PIPELINE_CACHE_CREATE_INFO
;
1687 ci
.pInitialData
= NULL
;
1688 ci
.initialDataSize
= 0;
1690 result
= radv_CreatePipelineCache(radv_device_to_handle(device
),
1692 if (result
!= VK_SUCCESS
)
1695 device
->mem_cache
= radv_pipeline_cache_from_handle(pc
);
1697 device
->force_aniso
=
1698 MIN2(16, radv_get_int_debug_option("RADV_TEX_ANISO", -1));
1699 if (device
->force_aniso
>= 0) {
1700 fprintf(stderr
, "radv: Forcing anisotropy filter to %ix\n",
1701 1 << util_logbase2(device
->force_aniso
));
1704 *pDevice
= radv_device_to_handle(device
);
1708 radv_device_finish_meta(device
);
1710 radv_bo_list_finish(&device
->bo_list
);
1712 if (device
->trace_bo
)
1713 device
->ws
->buffer_destroy(device
->trace_bo
);
1715 if (device
->gfx_init
)
1716 device
->ws
->buffer_destroy(device
->gfx_init
);
1718 for (unsigned i
= 0; i
< RADV_MAX_QUEUE_FAMILIES
; i
++) {
1719 for (unsigned q
= 0; q
< device
->queue_count
[i
]; q
++)
1720 radv_queue_finish(&device
->queues
[i
][q
]);
1721 if (device
->queue_count
[i
])
1722 vk_free(&device
->alloc
, device
->queues
[i
]);
1725 vk_free(&device
->alloc
, device
);
1729 void radv_DestroyDevice(
1731 const VkAllocationCallbacks
* pAllocator
)
1733 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1738 if (device
->trace_bo
)
1739 device
->ws
->buffer_destroy(device
->trace_bo
);
1741 if (device
->gfx_init
)
1742 device
->ws
->buffer_destroy(device
->gfx_init
);
1744 for (unsigned i
= 0; i
< RADV_MAX_QUEUE_FAMILIES
; i
++) {
1745 for (unsigned q
= 0; q
< device
->queue_count
[i
]; q
++)
1746 radv_queue_finish(&device
->queues
[i
][q
]);
1747 if (device
->queue_count
[i
])
1748 vk_free(&device
->alloc
, device
->queues
[i
]);
1749 if (device
->empty_cs
[i
])
1750 device
->ws
->cs_destroy(device
->empty_cs
[i
]);
1752 radv_device_finish_meta(device
);
1754 VkPipelineCache pc
= radv_pipeline_cache_to_handle(device
->mem_cache
);
1755 radv_DestroyPipelineCache(radv_device_to_handle(device
), pc
, NULL
);
1757 radv_destroy_shader_slabs(device
);
1759 radv_bo_list_finish(&device
->bo_list
);
1760 vk_free(&device
->alloc
, device
);
1763 VkResult
radv_EnumerateInstanceLayerProperties(
1764 uint32_t* pPropertyCount
,
1765 VkLayerProperties
* pProperties
)
1767 if (pProperties
== NULL
) {
1768 *pPropertyCount
= 0;
1772 /* None supported at this time */
1773 return vk_error(NULL
, VK_ERROR_LAYER_NOT_PRESENT
);
1776 VkResult
radv_EnumerateDeviceLayerProperties(
1777 VkPhysicalDevice physicalDevice
,
1778 uint32_t* pPropertyCount
,
1779 VkLayerProperties
* pProperties
)
1781 if (pProperties
== NULL
) {
1782 *pPropertyCount
= 0;
1786 /* None supported at this time */
1787 return vk_error(NULL
, VK_ERROR_LAYER_NOT_PRESENT
);
1790 void radv_GetDeviceQueue2(
1792 const VkDeviceQueueInfo2
* pQueueInfo
,
1795 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1796 struct radv_queue
*queue
;
1798 queue
= &device
->queues
[pQueueInfo
->queueFamilyIndex
][pQueueInfo
->queueIndex
];
1799 if (pQueueInfo
->flags
!= queue
->flags
) {
1800 /* From the Vulkan 1.1.70 spec:
1802 * "The queue returned by vkGetDeviceQueue2 must have the same
1803 * flags value from this structure as that used at device
1804 * creation time in a VkDeviceQueueCreateInfo instance. If no
1805 * matching flags were specified at device creation time then
1806 * pQueue will return VK_NULL_HANDLE."
1808 *pQueue
= VK_NULL_HANDLE
;
1812 *pQueue
= radv_queue_to_handle(queue
);
1815 void radv_GetDeviceQueue(
1817 uint32_t queueFamilyIndex
,
1818 uint32_t queueIndex
,
1821 const VkDeviceQueueInfo2 info
= (VkDeviceQueueInfo2
) {
1822 .sType
= VK_STRUCTURE_TYPE_DEVICE_QUEUE_INFO_2
,
1823 .queueFamilyIndex
= queueFamilyIndex
,
1824 .queueIndex
= queueIndex
1827 radv_GetDeviceQueue2(_device
, &info
, pQueue
);
1831 fill_geom_tess_rings(struct radv_queue
*queue
,
1833 bool add_sample_positions
,
1834 uint32_t esgs_ring_size
,
1835 struct radeon_winsys_bo
*esgs_ring_bo
,
1836 uint32_t gsvs_ring_size
,
1837 struct radeon_winsys_bo
*gsvs_ring_bo
,
1838 uint32_t tess_factor_ring_size
,
1839 uint32_t tess_offchip_ring_offset
,
1840 uint32_t tess_offchip_ring_size
,
1841 struct radeon_winsys_bo
*tess_rings_bo
)
1843 uint64_t esgs_va
= 0, gsvs_va
= 0;
1844 uint64_t tess_va
= 0, tess_offchip_va
= 0;
1845 uint32_t *desc
= &map
[4];
1848 esgs_va
= radv_buffer_get_va(esgs_ring_bo
);
1850 gsvs_va
= radv_buffer_get_va(gsvs_ring_bo
);
1851 if (tess_rings_bo
) {
1852 tess_va
= radv_buffer_get_va(tess_rings_bo
);
1853 tess_offchip_va
= tess_va
+ tess_offchip_ring_offset
;
1856 /* stride 0, num records - size, add tid, swizzle, elsize4,
1859 desc
[1] = S_008F04_BASE_ADDRESS_HI(esgs_va
>> 32) |
1860 S_008F04_STRIDE(0) |
1861 S_008F04_SWIZZLE_ENABLE(true);
1862 desc
[2] = esgs_ring_size
;
1863 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1864 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1865 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1866 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1867 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1868 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1869 S_008F0C_ELEMENT_SIZE(1) |
1870 S_008F0C_INDEX_STRIDE(3) |
1871 S_008F0C_ADD_TID_ENABLE(true);
1874 /* GS entry for ES->GS ring */
1875 /* stride 0, num records - size, elsize0,
1878 desc
[1] = S_008F04_BASE_ADDRESS_HI(esgs_va
>> 32)|
1879 S_008F04_STRIDE(0) |
1880 S_008F04_SWIZZLE_ENABLE(false);
1881 desc
[2] = esgs_ring_size
;
1882 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1883 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1884 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1885 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1886 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1887 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1888 S_008F0C_ELEMENT_SIZE(0) |
1889 S_008F0C_INDEX_STRIDE(0) |
1890 S_008F0C_ADD_TID_ENABLE(false);
1893 /* VS entry for GS->VS ring */
1894 /* stride 0, num records - size, elsize0,
1897 desc
[1] = S_008F04_BASE_ADDRESS_HI(gsvs_va
>> 32)|
1898 S_008F04_STRIDE(0) |
1899 S_008F04_SWIZZLE_ENABLE(false);
1900 desc
[2] = gsvs_ring_size
;
1901 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1902 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1903 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1904 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1905 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1906 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1907 S_008F0C_ELEMENT_SIZE(0) |
1908 S_008F0C_INDEX_STRIDE(0) |
1909 S_008F0C_ADD_TID_ENABLE(false);
1912 /* stride gsvs_itemsize, num records 64
1913 elsize 4, index stride 16 */
1914 /* shader will patch stride and desc[2] */
1916 desc
[1] = S_008F04_BASE_ADDRESS_HI(gsvs_va
>> 32)|
1917 S_008F04_STRIDE(0) |
1918 S_008F04_SWIZZLE_ENABLE(true);
1920 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1921 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1922 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1923 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1924 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1925 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1926 S_008F0C_ELEMENT_SIZE(1) |
1927 S_008F0C_INDEX_STRIDE(1) |
1928 S_008F0C_ADD_TID_ENABLE(true);
1932 desc
[1] = S_008F04_BASE_ADDRESS_HI(tess_va
>> 32) |
1933 S_008F04_STRIDE(0) |
1934 S_008F04_SWIZZLE_ENABLE(false);
1935 desc
[2] = tess_factor_ring_size
;
1936 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1937 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1938 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1939 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1940 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1941 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1942 S_008F0C_ELEMENT_SIZE(0) |
1943 S_008F0C_INDEX_STRIDE(0) |
1944 S_008F0C_ADD_TID_ENABLE(false);
1947 desc
[0] = tess_offchip_va
;
1948 desc
[1] = S_008F04_BASE_ADDRESS_HI(tess_offchip_va
>> 32) |
1949 S_008F04_STRIDE(0) |
1950 S_008F04_SWIZZLE_ENABLE(false);
1951 desc
[2] = tess_offchip_ring_size
;
1952 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1953 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1954 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1955 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1956 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1957 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1958 S_008F0C_ELEMENT_SIZE(0) |
1959 S_008F0C_INDEX_STRIDE(0) |
1960 S_008F0C_ADD_TID_ENABLE(false);
1963 /* add sample positions after all rings */
1964 memcpy(desc
, queue
->device
->sample_locations_1x
, 8);
1966 memcpy(desc
, queue
->device
->sample_locations_2x
, 16);
1968 memcpy(desc
, queue
->device
->sample_locations_4x
, 32);
1970 memcpy(desc
, queue
->device
->sample_locations_8x
, 64);
1972 memcpy(desc
, queue
->device
->sample_locations_16x
, 128);
1976 radv_get_hs_offchip_param(struct radv_device
*device
, uint32_t *max_offchip_buffers_p
)
1978 bool double_offchip_buffers
= device
->physical_device
->rad_info
.chip_class
>= CIK
&&
1979 device
->physical_device
->rad_info
.family
!= CHIP_CARRIZO
&&
1980 device
->physical_device
->rad_info
.family
!= CHIP_STONEY
;
1981 unsigned max_offchip_buffers_per_se
= double_offchip_buffers
? 128 : 64;
1982 unsigned max_offchip_buffers
;
1983 unsigned offchip_granularity
;
1984 unsigned hs_offchip_param
;
1988 * This must be one less than the maximum number due to a hw limitation.
1989 * Various hardware bugs in SI, CIK, and GFX9 need this.
1992 * Vega10 should limit max_offchip_buffers to 508 (4 * 127).
1993 * Gfx7 should limit max_offchip_buffers to 508
1994 * Gfx6 should limit max_offchip_buffers to 126 (2 * 63)
1996 * Follow AMDVLK here.
1998 if (device
->physical_device
->rad_info
.family
== CHIP_VEGA10
||
1999 device
->physical_device
->rad_info
.chip_class
== CIK
||
2000 device
->physical_device
->rad_info
.chip_class
== SI
)
2001 --max_offchip_buffers_per_se
;
2003 max_offchip_buffers
= max_offchip_buffers_per_se
*
2004 device
->physical_device
->rad_info
.max_se
;
2006 switch (device
->tess_offchip_block_dw_size
) {
2011 offchip_granularity
= V_03093C_X_8K_DWORDS
;
2014 offchip_granularity
= V_03093C_X_4K_DWORDS
;
2018 switch (device
->physical_device
->rad_info
.chip_class
) {
2020 max_offchip_buffers
= MIN2(max_offchip_buffers
, 126);
2026 max_offchip_buffers
= MIN2(max_offchip_buffers
, 508);
2030 *max_offchip_buffers_p
= max_offchip_buffers
;
2031 if (device
->physical_device
->rad_info
.chip_class
>= CIK
) {
2032 if (device
->physical_device
->rad_info
.chip_class
>= VI
)
2033 --max_offchip_buffers
;
2035 S_03093C_OFFCHIP_BUFFERING(max_offchip_buffers
) |
2036 S_03093C_OFFCHIP_GRANULARITY(offchip_granularity
);
2039 S_0089B0_OFFCHIP_BUFFERING(max_offchip_buffers
);
2041 return hs_offchip_param
;
2045 radv_emit_gs_ring_sizes(struct radv_queue
*queue
, struct radeon_cmdbuf
*cs
,
2046 struct radeon_winsys_bo
*esgs_ring_bo
,
2047 uint32_t esgs_ring_size
,
2048 struct radeon_winsys_bo
*gsvs_ring_bo
,
2049 uint32_t gsvs_ring_size
)
2051 if (!esgs_ring_bo
&& !gsvs_ring_bo
)
2055 radv_cs_add_buffer(queue
->device
->ws
, cs
, esgs_ring_bo
);
2058 radv_cs_add_buffer(queue
->device
->ws
, cs
, gsvs_ring_bo
);
2060 if (queue
->device
->physical_device
->rad_info
.chip_class
>= CIK
) {
2061 radeon_set_uconfig_reg_seq(cs
, R_030900_VGT_ESGS_RING_SIZE
, 2);
2062 radeon_emit(cs
, esgs_ring_size
>> 8);
2063 radeon_emit(cs
, gsvs_ring_size
>> 8);
2065 radeon_set_config_reg_seq(cs
, R_0088C8_VGT_ESGS_RING_SIZE
, 2);
2066 radeon_emit(cs
, esgs_ring_size
>> 8);
2067 radeon_emit(cs
, gsvs_ring_size
>> 8);
2072 radv_emit_tess_factor_ring(struct radv_queue
*queue
, struct radeon_cmdbuf
*cs
,
2073 unsigned hs_offchip_param
, unsigned tf_ring_size
,
2074 struct radeon_winsys_bo
*tess_rings_bo
)
2081 tf_va
= radv_buffer_get_va(tess_rings_bo
);
2083 radv_cs_add_buffer(queue
->device
->ws
, cs
, tess_rings_bo
);
2085 if (queue
->device
->physical_device
->rad_info
.chip_class
>= CIK
) {
2086 radeon_set_uconfig_reg(cs
, R_030938_VGT_TF_RING_SIZE
,
2087 S_030938_SIZE(tf_ring_size
/ 4));
2088 radeon_set_uconfig_reg(cs
, R_030940_VGT_TF_MEMORY_BASE
,
2090 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
2091 radeon_set_uconfig_reg(cs
, R_030944_VGT_TF_MEMORY_BASE_HI
,
2092 S_030944_BASE_HI(tf_va
>> 40));
2094 radeon_set_uconfig_reg(cs
, R_03093C_VGT_HS_OFFCHIP_PARAM
,
2097 radeon_set_config_reg(cs
, R_008988_VGT_TF_RING_SIZE
,
2098 S_008988_SIZE(tf_ring_size
/ 4));
2099 radeon_set_config_reg(cs
, R_0089B8_VGT_TF_MEMORY_BASE
,
2101 radeon_set_config_reg(cs
, R_0089B0_VGT_HS_OFFCHIP_PARAM
,
2107 radv_emit_compute_scratch(struct radv_queue
*queue
, struct radeon_cmdbuf
*cs
,
2108 struct radeon_winsys_bo
*compute_scratch_bo
)
2110 uint64_t scratch_va
;
2112 if (!compute_scratch_bo
)
2115 scratch_va
= radv_buffer_get_va(compute_scratch_bo
);
2117 radv_cs_add_buffer(queue
->device
->ws
, cs
, compute_scratch_bo
);
2119 radeon_set_sh_reg_seq(cs
, R_00B900_COMPUTE_USER_DATA_0
, 2);
2120 radeon_emit(cs
, scratch_va
);
2121 radeon_emit(cs
, S_008F04_BASE_ADDRESS_HI(scratch_va
>> 32) |
2122 S_008F04_SWIZZLE_ENABLE(1));
2126 radv_emit_global_shader_pointers(struct radv_queue
*queue
,
2127 struct radeon_cmdbuf
*cs
,
2128 struct radeon_winsys_bo
*descriptor_bo
)
2135 va
= radv_buffer_get_va(descriptor_bo
);
2137 radv_cs_add_buffer(queue
->device
->ws
, cs
, descriptor_bo
);
2139 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
2140 uint32_t regs
[] = {R_00B030_SPI_SHADER_USER_DATA_PS_0
,
2141 R_00B130_SPI_SHADER_USER_DATA_VS_0
,
2142 R_00B208_SPI_SHADER_USER_DATA_ADDR_LO_GS
,
2143 R_00B408_SPI_SHADER_USER_DATA_ADDR_LO_HS
};
2145 for (int i
= 0; i
< ARRAY_SIZE(regs
); ++i
) {
2146 radv_emit_shader_pointer(queue
->device
, cs
, regs
[i
],
2150 uint32_t regs
[] = {R_00B030_SPI_SHADER_USER_DATA_PS_0
,
2151 R_00B130_SPI_SHADER_USER_DATA_VS_0
,
2152 R_00B230_SPI_SHADER_USER_DATA_GS_0
,
2153 R_00B330_SPI_SHADER_USER_DATA_ES_0
,
2154 R_00B430_SPI_SHADER_USER_DATA_HS_0
,
2155 R_00B530_SPI_SHADER_USER_DATA_LS_0
};
2157 for (int i
= 0; i
< ARRAY_SIZE(regs
); ++i
) {
2158 radv_emit_shader_pointer(queue
->device
, cs
, regs
[i
],
2165 radv_init_graphics_state(struct radeon_cmdbuf
*cs
, struct radv_queue
*queue
)
2167 struct radv_device
*device
= queue
->device
;
2169 if (device
->gfx_init
) {
2170 uint64_t va
= radv_buffer_get_va(device
->gfx_init
);
2172 radeon_emit(cs
, PKT3(PKT3_INDIRECT_BUFFER_CIK
, 2, 0));
2173 radeon_emit(cs
, va
);
2174 radeon_emit(cs
, va
>> 32);
2175 radeon_emit(cs
, device
->gfx_init_size_dw
& 0xffff);
2177 radv_cs_add_buffer(device
->ws
, cs
, device
->gfx_init
);
2179 struct radv_physical_device
*physical_device
= device
->physical_device
;
2180 si_emit_graphics(physical_device
, cs
);
2185 radv_init_compute_state(struct radeon_cmdbuf
*cs
, struct radv_queue
*queue
)
2187 struct radv_physical_device
*physical_device
= queue
->device
->physical_device
;
2188 si_emit_compute(physical_device
, cs
);
2192 radv_get_preamble_cs(struct radv_queue
*queue
,
2193 uint32_t scratch_size
,
2194 uint32_t compute_scratch_size
,
2195 uint32_t esgs_ring_size
,
2196 uint32_t gsvs_ring_size
,
2197 bool needs_tess_rings
,
2198 bool needs_sample_positions
,
2199 struct radeon_cmdbuf
**initial_full_flush_preamble_cs
,
2200 struct radeon_cmdbuf
**initial_preamble_cs
,
2201 struct radeon_cmdbuf
**continue_preamble_cs
)
2203 struct radeon_winsys_bo
*scratch_bo
= NULL
;
2204 struct radeon_winsys_bo
*descriptor_bo
= NULL
;
2205 struct radeon_winsys_bo
*compute_scratch_bo
= NULL
;
2206 struct radeon_winsys_bo
*esgs_ring_bo
= NULL
;
2207 struct radeon_winsys_bo
*gsvs_ring_bo
= NULL
;
2208 struct radeon_winsys_bo
*tess_rings_bo
= NULL
;
2209 struct radeon_cmdbuf
*dest_cs
[3] = {0};
2210 bool add_tess_rings
= false, add_sample_positions
= false;
2211 unsigned tess_factor_ring_size
= 0, tess_offchip_ring_size
= 0;
2212 unsigned max_offchip_buffers
;
2213 unsigned hs_offchip_param
= 0;
2214 unsigned tess_offchip_ring_offset
;
2215 uint32_t ring_bo_flags
= RADEON_FLAG_NO_CPU_ACCESS
| RADEON_FLAG_NO_INTERPROCESS_SHARING
;
2216 if (!queue
->has_tess_rings
) {
2217 if (needs_tess_rings
)
2218 add_tess_rings
= true;
2220 if (!queue
->has_sample_positions
) {
2221 if (needs_sample_positions
)
2222 add_sample_positions
= true;
2224 tess_factor_ring_size
= 32768 * queue
->device
->physical_device
->rad_info
.max_se
;
2225 hs_offchip_param
= radv_get_hs_offchip_param(queue
->device
,
2226 &max_offchip_buffers
);
2227 tess_offchip_ring_offset
= align(tess_factor_ring_size
, 64 * 1024);
2228 tess_offchip_ring_size
= max_offchip_buffers
*
2229 queue
->device
->tess_offchip_block_dw_size
* 4;
2231 if (scratch_size
<= queue
->scratch_size
&&
2232 compute_scratch_size
<= queue
->compute_scratch_size
&&
2233 esgs_ring_size
<= queue
->esgs_ring_size
&&
2234 gsvs_ring_size
<= queue
->gsvs_ring_size
&&
2235 !add_tess_rings
&& !add_sample_positions
&&
2236 queue
->initial_preamble_cs
) {
2237 *initial_full_flush_preamble_cs
= queue
->initial_full_flush_preamble_cs
;
2238 *initial_preamble_cs
= queue
->initial_preamble_cs
;
2239 *continue_preamble_cs
= queue
->continue_preamble_cs
;
2240 if (!scratch_size
&& !compute_scratch_size
&& !esgs_ring_size
&& !gsvs_ring_size
)
2241 *continue_preamble_cs
= NULL
;
2245 if (scratch_size
> queue
->scratch_size
) {
2246 scratch_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
2254 scratch_bo
= queue
->scratch_bo
;
2256 if (compute_scratch_size
> queue
->compute_scratch_size
) {
2257 compute_scratch_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
2258 compute_scratch_size
,
2262 if (!compute_scratch_bo
)
2266 compute_scratch_bo
= queue
->compute_scratch_bo
;
2268 if (esgs_ring_size
> queue
->esgs_ring_size
) {
2269 esgs_ring_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
2277 esgs_ring_bo
= queue
->esgs_ring_bo
;
2278 esgs_ring_size
= queue
->esgs_ring_size
;
2281 if (gsvs_ring_size
> queue
->gsvs_ring_size
) {
2282 gsvs_ring_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
2290 gsvs_ring_bo
= queue
->gsvs_ring_bo
;
2291 gsvs_ring_size
= queue
->gsvs_ring_size
;
2294 if (add_tess_rings
) {
2295 tess_rings_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
2296 tess_offchip_ring_offset
+ tess_offchip_ring_size
,
2303 tess_rings_bo
= queue
->tess_rings_bo
;
2306 if (scratch_bo
!= queue
->scratch_bo
||
2307 esgs_ring_bo
!= queue
->esgs_ring_bo
||
2308 gsvs_ring_bo
!= queue
->gsvs_ring_bo
||
2309 tess_rings_bo
!= queue
->tess_rings_bo
||
2310 add_sample_positions
) {
2312 if (gsvs_ring_bo
|| esgs_ring_bo
||
2313 tess_rings_bo
|| add_sample_positions
) {
2314 size
= 112; /* 2 dword + 2 padding + 4 dword * 6 */
2315 if (add_sample_positions
)
2316 size
+= 256; /* 32+16+8+4+2+1 samples * 4 * 2 = 248 bytes. */
2318 else if (scratch_bo
)
2319 size
= 8; /* 2 dword */
2321 descriptor_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
2325 RADEON_FLAG_CPU_ACCESS
|
2326 RADEON_FLAG_NO_INTERPROCESS_SHARING
|
2327 RADEON_FLAG_READ_ONLY
);
2331 descriptor_bo
= queue
->descriptor_bo
;
2333 for(int i
= 0; i
< 3; ++i
) {
2334 struct radeon_cmdbuf
*cs
= NULL
;
2335 cs
= queue
->device
->ws
->cs_create(queue
->device
->ws
,
2336 queue
->queue_family_index
? RING_COMPUTE
: RING_GFX
);
2343 radv_cs_add_buffer(queue
->device
->ws
, cs
, scratch_bo
);
2345 /* Emit initial configuration. */
2346 switch (queue
->queue_family_index
) {
2347 case RADV_QUEUE_GENERAL
:
2348 radv_init_graphics_state(cs
, queue
);
2350 case RADV_QUEUE_COMPUTE
:
2351 radv_init_compute_state(cs
, queue
);
2353 case RADV_QUEUE_TRANSFER
:
2357 if (descriptor_bo
!= queue
->descriptor_bo
) {
2358 uint32_t *map
= (uint32_t*)queue
->device
->ws
->buffer_map(descriptor_bo
);
2361 uint64_t scratch_va
= radv_buffer_get_va(scratch_bo
);
2362 uint32_t rsrc1
= S_008F04_BASE_ADDRESS_HI(scratch_va
>> 32) |
2363 S_008F04_SWIZZLE_ENABLE(1);
2364 map
[0] = scratch_va
;
2368 if (esgs_ring_bo
|| gsvs_ring_bo
|| tess_rings_bo
||
2369 add_sample_positions
)
2370 fill_geom_tess_rings(queue
, map
, add_sample_positions
,
2371 esgs_ring_size
, esgs_ring_bo
,
2372 gsvs_ring_size
, gsvs_ring_bo
,
2373 tess_factor_ring_size
,
2374 tess_offchip_ring_offset
,
2375 tess_offchip_ring_size
,
2378 queue
->device
->ws
->buffer_unmap(descriptor_bo
);
2381 if (esgs_ring_bo
|| gsvs_ring_bo
|| tess_rings_bo
) {
2382 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
2383 radeon_emit(cs
, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
2384 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
2385 radeon_emit(cs
, EVENT_TYPE(V_028A90_VGT_FLUSH
) | EVENT_INDEX(0));
2388 radv_emit_gs_ring_sizes(queue
, cs
, esgs_ring_bo
, esgs_ring_size
,
2389 gsvs_ring_bo
, gsvs_ring_size
);
2390 radv_emit_tess_factor_ring(queue
, cs
, hs_offchip_param
,
2391 tess_factor_ring_size
, tess_rings_bo
);
2392 radv_emit_global_shader_pointers(queue
, cs
, descriptor_bo
);
2393 radv_emit_compute_scratch(queue
, cs
, compute_scratch_bo
);
2396 si_cs_emit_cache_flush(cs
,
2397 queue
->device
->physical_device
->rad_info
.chip_class
,
2399 queue
->queue_family_index
== RING_COMPUTE
&&
2400 queue
->device
->physical_device
->rad_info
.chip_class
>= CIK
,
2401 (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
)) |
2402 RADV_CMD_FLAG_INV_ICACHE
|
2403 RADV_CMD_FLAG_INV_SMEM_L1
|
2404 RADV_CMD_FLAG_INV_VMEM_L1
|
2405 RADV_CMD_FLAG_INV_GLOBAL_L2
|
2406 RADV_CMD_FLAG_START_PIPELINE_STATS
, 0);
2407 } else if (i
== 1) {
2408 si_cs_emit_cache_flush(cs
,
2409 queue
->device
->physical_device
->rad_info
.chip_class
,
2411 queue
->queue_family_index
== RING_COMPUTE
&&
2412 queue
->device
->physical_device
->rad_info
.chip_class
>= CIK
,
2413 RADV_CMD_FLAG_INV_ICACHE
|
2414 RADV_CMD_FLAG_INV_SMEM_L1
|
2415 RADV_CMD_FLAG_INV_VMEM_L1
|
2416 RADV_CMD_FLAG_INV_GLOBAL_L2
|
2417 RADV_CMD_FLAG_START_PIPELINE_STATS
, 0);
2420 if (!queue
->device
->ws
->cs_finalize(cs
))
2424 if (queue
->initial_full_flush_preamble_cs
)
2425 queue
->device
->ws
->cs_destroy(queue
->initial_full_flush_preamble_cs
);
2427 if (queue
->initial_preamble_cs
)
2428 queue
->device
->ws
->cs_destroy(queue
->initial_preamble_cs
);
2430 if (queue
->continue_preamble_cs
)
2431 queue
->device
->ws
->cs_destroy(queue
->continue_preamble_cs
);
2433 queue
->initial_full_flush_preamble_cs
= dest_cs
[0];
2434 queue
->initial_preamble_cs
= dest_cs
[1];
2435 queue
->continue_preamble_cs
= dest_cs
[2];
2437 if (scratch_bo
!= queue
->scratch_bo
) {
2438 if (queue
->scratch_bo
)
2439 queue
->device
->ws
->buffer_destroy(queue
->scratch_bo
);
2440 queue
->scratch_bo
= scratch_bo
;
2441 queue
->scratch_size
= scratch_size
;
2444 if (compute_scratch_bo
!= queue
->compute_scratch_bo
) {
2445 if (queue
->compute_scratch_bo
)
2446 queue
->device
->ws
->buffer_destroy(queue
->compute_scratch_bo
);
2447 queue
->compute_scratch_bo
= compute_scratch_bo
;
2448 queue
->compute_scratch_size
= compute_scratch_size
;
2451 if (esgs_ring_bo
!= queue
->esgs_ring_bo
) {
2452 if (queue
->esgs_ring_bo
)
2453 queue
->device
->ws
->buffer_destroy(queue
->esgs_ring_bo
);
2454 queue
->esgs_ring_bo
= esgs_ring_bo
;
2455 queue
->esgs_ring_size
= esgs_ring_size
;
2458 if (gsvs_ring_bo
!= queue
->gsvs_ring_bo
) {
2459 if (queue
->gsvs_ring_bo
)
2460 queue
->device
->ws
->buffer_destroy(queue
->gsvs_ring_bo
);
2461 queue
->gsvs_ring_bo
= gsvs_ring_bo
;
2462 queue
->gsvs_ring_size
= gsvs_ring_size
;
2465 if (tess_rings_bo
!= queue
->tess_rings_bo
) {
2466 queue
->tess_rings_bo
= tess_rings_bo
;
2467 queue
->has_tess_rings
= true;
2470 if (descriptor_bo
!= queue
->descriptor_bo
) {
2471 if (queue
->descriptor_bo
)
2472 queue
->device
->ws
->buffer_destroy(queue
->descriptor_bo
);
2474 queue
->descriptor_bo
= descriptor_bo
;
2477 if (add_sample_positions
)
2478 queue
->has_sample_positions
= true;
2480 *initial_full_flush_preamble_cs
= queue
->initial_full_flush_preamble_cs
;
2481 *initial_preamble_cs
= queue
->initial_preamble_cs
;
2482 *continue_preamble_cs
= queue
->continue_preamble_cs
;
2483 if (!scratch_size
&& !compute_scratch_size
&& !esgs_ring_size
&& !gsvs_ring_size
)
2484 *continue_preamble_cs
= NULL
;
2487 for (int i
= 0; i
< ARRAY_SIZE(dest_cs
); ++i
)
2489 queue
->device
->ws
->cs_destroy(dest_cs
[i
]);
2490 if (descriptor_bo
&& descriptor_bo
!= queue
->descriptor_bo
)
2491 queue
->device
->ws
->buffer_destroy(descriptor_bo
);
2492 if (scratch_bo
&& scratch_bo
!= queue
->scratch_bo
)
2493 queue
->device
->ws
->buffer_destroy(scratch_bo
);
2494 if (compute_scratch_bo
&& compute_scratch_bo
!= queue
->compute_scratch_bo
)
2495 queue
->device
->ws
->buffer_destroy(compute_scratch_bo
);
2496 if (esgs_ring_bo
&& esgs_ring_bo
!= queue
->esgs_ring_bo
)
2497 queue
->device
->ws
->buffer_destroy(esgs_ring_bo
);
2498 if (gsvs_ring_bo
&& gsvs_ring_bo
!= queue
->gsvs_ring_bo
)
2499 queue
->device
->ws
->buffer_destroy(gsvs_ring_bo
);
2500 if (tess_rings_bo
&& tess_rings_bo
!= queue
->tess_rings_bo
)
2501 queue
->device
->ws
->buffer_destroy(tess_rings_bo
);
2502 return vk_error(queue
->device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
2505 static VkResult
radv_alloc_sem_counts(struct radv_instance
*instance
,
2506 struct radv_winsys_sem_counts
*counts
,
2508 const VkSemaphore
*sems
,
2512 int syncobj_idx
= 0, sem_idx
= 0;
2514 if (num_sems
== 0 && _fence
== VK_NULL_HANDLE
)
2517 for (uint32_t i
= 0; i
< num_sems
; i
++) {
2518 RADV_FROM_HANDLE(radv_semaphore
, sem
, sems
[i
]);
2520 if (sem
->temp_syncobj
|| sem
->syncobj
)
2521 counts
->syncobj_count
++;
2523 counts
->sem_count
++;
2526 if (_fence
!= VK_NULL_HANDLE
) {
2527 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
2528 if (fence
->temp_syncobj
|| fence
->syncobj
)
2529 counts
->syncobj_count
++;
2532 if (counts
->syncobj_count
) {
2533 counts
->syncobj
= (uint32_t *)malloc(sizeof(uint32_t) * counts
->syncobj_count
);
2534 if (!counts
->syncobj
)
2535 return vk_error(instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
2538 if (counts
->sem_count
) {
2539 counts
->sem
= (struct radeon_winsys_sem
**)malloc(sizeof(struct radeon_winsys_sem
*) * counts
->sem_count
);
2541 free(counts
->syncobj
);
2542 return vk_error(instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
2546 for (uint32_t i
= 0; i
< num_sems
; i
++) {
2547 RADV_FROM_HANDLE(radv_semaphore
, sem
, sems
[i
]);
2549 if (sem
->temp_syncobj
) {
2550 counts
->syncobj
[syncobj_idx
++] = sem
->temp_syncobj
;
2552 else if (sem
->syncobj
)
2553 counts
->syncobj
[syncobj_idx
++] = sem
->syncobj
;
2556 counts
->sem
[sem_idx
++] = sem
->sem
;
2560 if (_fence
!= VK_NULL_HANDLE
) {
2561 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
2562 if (fence
->temp_syncobj
)
2563 counts
->syncobj
[syncobj_idx
++] = fence
->temp_syncobj
;
2564 else if (fence
->syncobj
)
2565 counts
->syncobj
[syncobj_idx
++] = fence
->syncobj
;
2572 radv_free_sem_info(struct radv_winsys_sem_info
*sem_info
)
2574 free(sem_info
->wait
.syncobj
);
2575 free(sem_info
->wait
.sem
);
2576 free(sem_info
->signal
.syncobj
);
2577 free(sem_info
->signal
.sem
);
2581 static void radv_free_temp_syncobjs(struct radv_device
*device
,
2583 const VkSemaphore
*sems
)
2585 for (uint32_t i
= 0; i
< num_sems
; i
++) {
2586 RADV_FROM_HANDLE(radv_semaphore
, sem
, sems
[i
]);
2588 if (sem
->temp_syncobj
) {
2589 device
->ws
->destroy_syncobj(device
->ws
, sem
->temp_syncobj
);
2590 sem
->temp_syncobj
= 0;
2596 radv_alloc_sem_info(struct radv_instance
*instance
,
2597 struct radv_winsys_sem_info
*sem_info
,
2599 const VkSemaphore
*wait_sems
,
2600 int num_signal_sems
,
2601 const VkSemaphore
*signal_sems
,
2605 memset(sem_info
, 0, sizeof(*sem_info
));
2607 ret
= radv_alloc_sem_counts(instance
, &sem_info
->wait
, num_wait_sems
, wait_sems
, VK_NULL_HANDLE
, true);
2610 ret
= radv_alloc_sem_counts(instance
, &sem_info
->signal
, num_signal_sems
, signal_sems
, fence
, false);
2612 radv_free_sem_info(sem_info
);
2614 /* caller can override these */
2615 sem_info
->cs_emit_wait
= true;
2616 sem_info
->cs_emit_signal
= true;
2620 /* Signals fence as soon as all the work currently put on queue is done. */
2621 static VkResult
radv_signal_fence(struct radv_queue
*queue
,
2622 struct radv_fence
*fence
)
2626 struct radv_winsys_sem_info sem_info
;
2628 result
= radv_alloc_sem_info(queue
->device
->instance
, &sem_info
, 0, NULL
, 0, NULL
,
2629 radv_fence_to_handle(fence
));
2630 if (result
!= VK_SUCCESS
)
2633 ret
= queue
->device
->ws
->cs_submit(queue
->hw_ctx
, queue
->queue_idx
,
2634 &queue
->device
->empty_cs
[queue
->queue_family_index
],
2635 1, NULL
, NULL
, &sem_info
, NULL
,
2636 false, fence
->fence
);
2637 radv_free_sem_info(&sem_info
);
2640 return vk_error(queue
->device
->instance
, VK_ERROR_DEVICE_LOST
);
2645 VkResult
radv_QueueSubmit(
2647 uint32_t submitCount
,
2648 const VkSubmitInfo
* pSubmits
,
2651 RADV_FROM_HANDLE(radv_queue
, queue
, _queue
);
2652 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
2653 struct radeon_winsys_fence
*base_fence
= fence
? fence
->fence
: NULL
;
2654 struct radeon_winsys_ctx
*ctx
= queue
->hw_ctx
;
2656 uint32_t max_cs_submission
= queue
->device
->trace_bo
? 1 : UINT32_MAX
;
2657 uint32_t scratch_size
= 0;
2658 uint32_t compute_scratch_size
= 0;
2659 uint32_t esgs_ring_size
= 0, gsvs_ring_size
= 0;
2660 struct radeon_cmdbuf
*initial_preamble_cs
= NULL
, *initial_flush_preamble_cs
= NULL
, *continue_preamble_cs
= NULL
;
2662 bool fence_emitted
= false;
2663 bool tess_rings_needed
= false;
2664 bool sample_positions_needed
= false;
2666 /* Do this first so failing to allocate scratch buffers can't result in
2667 * partially executed submissions. */
2668 for (uint32_t i
= 0; i
< submitCount
; i
++) {
2669 for (uint32_t j
= 0; j
< pSubmits
[i
].commandBufferCount
; j
++) {
2670 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
,
2671 pSubmits
[i
].pCommandBuffers
[j
]);
2673 scratch_size
= MAX2(scratch_size
, cmd_buffer
->scratch_size_needed
);
2674 compute_scratch_size
= MAX2(compute_scratch_size
,
2675 cmd_buffer
->compute_scratch_size_needed
);
2676 esgs_ring_size
= MAX2(esgs_ring_size
, cmd_buffer
->esgs_ring_size_needed
);
2677 gsvs_ring_size
= MAX2(gsvs_ring_size
, cmd_buffer
->gsvs_ring_size_needed
);
2678 tess_rings_needed
|= cmd_buffer
->tess_rings_needed
;
2679 sample_positions_needed
|= cmd_buffer
->sample_positions_needed
;
2683 result
= radv_get_preamble_cs(queue
, scratch_size
, compute_scratch_size
,
2684 esgs_ring_size
, gsvs_ring_size
, tess_rings_needed
,
2685 sample_positions_needed
, &initial_flush_preamble_cs
,
2686 &initial_preamble_cs
, &continue_preamble_cs
);
2687 if (result
!= VK_SUCCESS
)
2690 for (uint32_t i
= 0; i
< submitCount
; i
++) {
2691 struct radeon_cmdbuf
**cs_array
;
2692 bool do_flush
= !i
|| pSubmits
[i
].pWaitDstStageMask
;
2693 bool can_patch
= true;
2695 struct radv_winsys_sem_info sem_info
;
2697 result
= radv_alloc_sem_info(queue
->device
->instance
,
2699 pSubmits
[i
].waitSemaphoreCount
,
2700 pSubmits
[i
].pWaitSemaphores
,
2701 pSubmits
[i
].signalSemaphoreCount
,
2702 pSubmits
[i
].pSignalSemaphores
,
2704 if (result
!= VK_SUCCESS
)
2707 if (!pSubmits
[i
].commandBufferCount
) {
2708 if (pSubmits
[i
].waitSemaphoreCount
|| pSubmits
[i
].signalSemaphoreCount
) {
2709 ret
= queue
->device
->ws
->cs_submit(ctx
, queue
->queue_idx
,
2710 &queue
->device
->empty_cs
[queue
->queue_family_index
],
2715 radv_loge("failed to submit CS %d\n", i
);
2718 fence_emitted
= true;
2720 radv_free_sem_info(&sem_info
);
2724 cs_array
= malloc(sizeof(struct radeon_cmdbuf
*) *
2725 (pSubmits
[i
].commandBufferCount
));
2727 for (uint32_t j
= 0; j
< pSubmits
[i
].commandBufferCount
; j
++) {
2728 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
,
2729 pSubmits
[i
].pCommandBuffers
[j
]);
2730 assert(cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_PRIMARY
);
2732 cs_array
[j
] = cmd_buffer
->cs
;
2733 if ((cmd_buffer
->usage_flags
& VK_COMMAND_BUFFER_USAGE_SIMULTANEOUS_USE_BIT
))
2736 cmd_buffer
->status
= RADV_CMD_BUFFER_STATUS_PENDING
;
2739 for (uint32_t j
= 0; j
< pSubmits
[i
].commandBufferCount
; j
+= advance
) {
2740 struct radeon_cmdbuf
*initial_preamble
= (do_flush
&& !j
) ? initial_flush_preamble_cs
: initial_preamble_cs
;
2741 const struct radv_winsys_bo_list
*bo_list
= NULL
;
2743 advance
= MIN2(max_cs_submission
,
2744 pSubmits
[i
].commandBufferCount
- j
);
2746 if (queue
->device
->trace_bo
)
2747 *queue
->device
->trace_id_ptr
= 0;
2749 sem_info
.cs_emit_wait
= j
== 0;
2750 sem_info
.cs_emit_signal
= j
+ advance
== pSubmits
[i
].commandBufferCount
;
2752 if (unlikely(queue
->device
->use_global_bo_list
)) {
2753 pthread_mutex_lock(&queue
->device
->bo_list
.mutex
);
2754 bo_list
= &queue
->device
->bo_list
.list
;
2757 ret
= queue
->device
->ws
->cs_submit(ctx
, queue
->queue_idx
, cs_array
+ j
,
2758 advance
, initial_preamble
, continue_preamble_cs
,
2760 can_patch
, base_fence
);
2762 if (unlikely(queue
->device
->use_global_bo_list
))
2763 pthread_mutex_unlock(&queue
->device
->bo_list
.mutex
);
2766 radv_loge("failed to submit CS %d\n", i
);
2769 fence_emitted
= true;
2770 if (queue
->device
->trace_bo
) {
2771 radv_check_gpu_hangs(queue
, cs_array
[j
]);
2775 radv_free_temp_syncobjs(queue
->device
,
2776 pSubmits
[i
].waitSemaphoreCount
,
2777 pSubmits
[i
].pWaitSemaphores
);
2778 radv_free_sem_info(&sem_info
);
2783 if (!fence_emitted
) {
2784 result
= radv_signal_fence(queue
, fence
);
2785 if (result
!= VK_SUCCESS
)
2788 fence
->submitted
= true;
2794 VkResult
radv_QueueWaitIdle(
2797 RADV_FROM_HANDLE(radv_queue
, queue
, _queue
);
2799 queue
->device
->ws
->ctx_wait_idle(queue
->hw_ctx
,
2800 radv_queue_family_to_ring(queue
->queue_family_index
),
2805 VkResult
radv_DeviceWaitIdle(
2808 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2810 for (unsigned i
= 0; i
< RADV_MAX_QUEUE_FAMILIES
; i
++) {
2811 for (unsigned q
= 0; q
< device
->queue_count
[i
]; q
++) {
2812 radv_QueueWaitIdle(radv_queue_to_handle(&device
->queues
[i
][q
]));
2818 VkResult
radv_EnumerateInstanceExtensionProperties(
2819 const char* pLayerName
,
2820 uint32_t* pPropertyCount
,
2821 VkExtensionProperties
* pProperties
)
2823 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
2825 for (int i
= 0; i
< RADV_INSTANCE_EXTENSION_COUNT
; i
++) {
2826 if (radv_supported_instance_extensions
.extensions
[i
]) {
2827 vk_outarray_append(&out
, prop
) {
2828 *prop
= radv_instance_extensions
[i
];
2833 return vk_outarray_status(&out
);
2836 VkResult
radv_EnumerateDeviceExtensionProperties(
2837 VkPhysicalDevice physicalDevice
,
2838 const char* pLayerName
,
2839 uint32_t* pPropertyCount
,
2840 VkExtensionProperties
* pProperties
)
2842 RADV_FROM_HANDLE(radv_physical_device
, device
, physicalDevice
);
2843 VK_OUTARRAY_MAKE(out
, pProperties
, pPropertyCount
);
2845 for (int i
= 0; i
< RADV_DEVICE_EXTENSION_COUNT
; i
++) {
2846 if (device
->supported_extensions
.extensions
[i
]) {
2847 vk_outarray_append(&out
, prop
) {
2848 *prop
= radv_device_extensions
[i
];
2853 return vk_outarray_status(&out
);
2856 PFN_vkVoidFunction
radv_GetInstanceProcAddr(
2857 VkInstance _instance
,
2860 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
2862 return radv_lookup_entrypoint_checked(pName
,
2863 instance
? instance
->apiVersion
: 0,
2864 instance
? &instance
->enabled_extensions
: NULL
,
2868 /* The loader wants us to expose a second GetInstanceProcAddr function
2869 * to work around certain LD_PRELOAD issues seen in apps.
2872 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
2873 VkInstance instance
,
2877 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
2878 VkInstance instance
,
2881 return radv_GetInstanceProcAddr(instance
, pName
);
2884 PFN_vkVoidFunction
radv_GetDeviceProcAddr(
2888 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2890 return radv_lookup_entrypoint_checked(pName
,
2891 device
->instance
->apiVersion
,
2892 &device
->instance
->enabled_extensions
,
2893 &device
->enabled_extensions
);
2896 bool radv_get_memory_fd(struct radv_device
*device
,
2897 struct radv_device_memory
*memory
,
2900 struct radeon_bo_metadata metadata
;
2902 if (memory
->image
) {
2903 radv_init_metadata(device
, memory
->image
, &metadata
);
2904 device
->ws
->buffer_set_metadata(memory
->bo
, &metadata
);
2907 return device
->ws
->buffer_get_fd(device
->ws
, memory
->bo
,
2911 static VkResult
radv_alloc_memory(struct radv_device
*device
,
2912 const VkMemoryAllocateInfo
* pAllocateInfo
,
2913 const VkAllocationCallbacks
* pAllocator
,
2914 VkDeviceMemory
* pMem
)
2916 struct radv_device_memory
*mem
;
2918 enum radeon_bo_domain domain
;
2920 enum radv_mem_type mem_type_index
= device
->physical_device
->mem_type_indices
[pAllocateInfo
->memoryTypeIndex
];
2922 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
2924 if (pAllocateInfo
->allocationSize
== 0) {
2925 /* Apparently, this is allowed */
2926 *pMem
= VK_NULL_HANDLE
;
2930 const VkImportMemoryFdInfoKHR
*import_info
=
2931 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_FD_INFO_KHR
);
2932 const VkMemoryDedicatedAllocateInfoKHR
*dedicate_info
=
2933 vk_find_struct_const(pAllocateInfo
->pNext
, MEMORY_DEDICATED_ALLOCATE_INFO_KHR
);
2934 const VkExportMemoryAllocateInfoKHR
*export_info
=
2935 vk_find_struct_const(pAllocateInfo
->pNext
, EXPORT_MEMORY_ALLOCATE_INFO_KHR
);
2936 const VkImportMemoryHostPointerInfoEXT
*host_ptr_info
=
2937 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_HOST_POINTER_INFO_EXT
);
2939 const struct wsi_memory_allocate_info
*wsi_info
=
2940 vk_find_struct_const(pAllocateInfo
->pNext
, WSI_MEMORY_ALLOCATE_INFO_MESA
);
2942 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
2943 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2945 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
2947 if (wsi_info
&& wsi_info
->implicit_sync
)
2948 flags
|= RADEON_FLAG_IMPLICIT_SYNC
;
2950 if (dedicate_info
) {
2951 mem
->image
= radv_image_from_handle(dedicate_info
->image
);
2952 mem
->buffer
= radv_buffer_from_handle(dedicate_info
->buffer
);
2958 mem
->user_ptr
= NULL
;
2961 assert(import_info
->handleType
==
2962 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
||
2963 import_info
->handleType
==
2964 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
2965 mem
->bo
= device
->ws
->buffer_from_fd(device
->ws
, import_info
->fd
,
2968 result
= VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
;
2971 close(import_info
->fd
);
2973 } else if (host_ptr_info
) {
2974 assert(host_ptr_info
->handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT
);
2975 assert(mem_type_index
== RADV_MEM_TYPE_GTT_CACHED
);
2976 mem
->bo
= device
->ws
->buffer_from_ptr(device
->ws
, host_ptr_info
->pHostPointer
,
2977 pAllocateInfo
->allocationSize
);
2979 result
= VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
;
2982 mem
->user_ptr
= host_ptr_info
->pHostPointer
;
2985 uint64_t alloc_size
= align_u64(pAllocateInfo
->allocationSize
, 4096);
2986 if (mem_type_index
== RADV_MEM_TYPE_GTT_WRITE_COMBINE
||
2987 mem_type_index
== RADV_MEM_TYPE_GTT_CACHED
)
2988 domain
= RADEON_DOMAIN_GTT
;
2990 domain
= RADEON_DOMAIN_VRAM
;
2992 if (mem_type_index
== RADV_MEM_TYPE_VRAM
)
2993 flags
|= RADEON_FLAG_NO_CPU_ACCESS
;
2995 flags
|= RADEON_FLAG_CPU_ACCESS
;
2997 if (mem_type_index
== RADV_MEM_TYPE_GTT_WRITE_COMBINE
)
2998 flags
|= RADEON_FLAG_GTT_WC
;
3000 if (!dedicate_info
&& !import_info
&& (!export_info
|| !export_info
->handleTypes
))
3001 flags
|= RADEON_FLAG_NO_INTERPROCESS_SHARING
;
3003 mem
->bo
= device
->ws
->buffer_create(device
->ws
, alloc_size
, device
->physical_device
->rad_info
.max_alignment
,
3007 result
= VK_ERROR_OUT_OF_DEVICE_MEMORY
;
3010 mem
->type_index
= mem_type_index
;
3013 result
= radv_bo_list_add(device
, mem
->bo
);
3014 if (result
!= VK_SUCCESS
)
3017 *pMem
= radv_device_memory_to_handle(mem
);
3022 device
->ws
->buffer_destroy(mem
->bo
);
3024 vk_free2(&device
->alloc
, pAllocator
, mem
);
3029 VkResult
radv_AllocateMemory(
3031 const VkMemoryAllocateInfo
* pAllocateInfo
,
3032 const VkAllocationCallbacks
* pAllocator
,
3033 VkDeviceMemory
* pMem
)
3035 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3036 return radv_alloc_memory(device
, pAllocateInfo
, pAllocator
, pMem
);
3039 void radv_FreeMemory(
3041 VkDeviceMemory _mem
,
3042 const VkAllocationCallbacks
* pAllocator
)
3044 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3045 RADV_FROM_HANDLE(radv_device_memory
, mem
, _mem
);
3050 radv_bo_list_remove(device
, mem
->bo
);
3051 device
->ws
->buffer_destroy(mem
->bo
);
3054 vk_free2(&device
->alloc
, pAllocator
, mem
);
3057 VkResult
radv_MapMemory(
3059 VkDeviceMemory _memory
,
3060 VkDeviceSize offset
,
3062 VkMemoryMapFlags flags
,
3065 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3066 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
3074 *ppData
= mem
->user_ptr
;
3076 *ppData
= device
->ws
->buffer_map(mem
->bo
);
3083 return vk_error(device
->instance
, VK_ERROR_MEMORY_MAP_FAILED
);
3086 void radv_UnmapMemory(
3088 VkDeviceMemory _memory
)
3090 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3091 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
3096 if (mem
->user_ptr
== NULL
)
3097 device
->ws
->buffer_unmap(mem
->bo
);
3100 VkResult
radv_FlushMappedMemoryRanges(
3102 uint32_t memoryRangeCount
,
3103 const VkMappedMemoryRange
* pMemoryRanges
)
3108 VkResult
radv_InvalidateMappedMemoryRanges(
3110 uint32_t memoryRangeCount
,
3111 const VkMappedMemoryRange
* pMemoryRanges
)
3116 void radv_GetBufferMemoryRequirements(
3119 VkMemoryRequirements
* pMemoryRequirements
)
3121 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3122 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
3124 pMemoryRequirements
->memoryTypeBits
= (1u << device
->physical_device
->memory_properties
.memoryTypeCount
) - 1;
3126 if (buffer
->flags
& VK_BUFFER_CREATE_SPARSE_BINDING_BIT
)
3127 pMemoryRequirements
->alignment
= 4096;
3129 pMemoryRequirements
->alignment
= 16;
3131 pMemoryRequirements
->size
= align64(buffer
->size
, pMemoryRequirements
->alignment
);
3134 void radv_GetBufferMemoryRequirements2(
3136 const VkBufferMemoryRequirementsInfo2KHR
* pInfo
,
3137 VkMemoryRequirements2KHR
* pMemoryRequirements
)
3139 radv_GetBufferMemoryRequirements(device
, pInfo
->buffer
,
3140 &pMemoryRequirements
->memoryRequirements
);
3141 RADV_FROM_HANDLE(radv_buffer
, buffer
, pInfo
->buffer
);
3142 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3143 switch (ext
->sType
) {
3144 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS_KHR
: {
3145 VkMemoryDedicatedRequirementsKHR
*req
=
3146 (VkMemoryDedicatedRequirementsKHR
*) ext
;
3147 req
->requiresDedicatedAllocation
= buffer
->shareable
;
3148 req
->prefersDedicatedAllocation
= req
->requiresDedicatedAllocation
;
3157 void radv_GetImageMemoryRequirements(
3160 VkMemoryRequirements
* pMemoryRequirements
)
3162 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3163 RADV_FROM_HANDLE(radv_image
, image
, _image
);
3165 pMemoryRequirements
->memoryTypeBits
= (1u << device
->physical_device
->memory_properties
.memoryTypeCount
) - 1;
3167 pMemoryRequirements
->size
= image
->size
;
3168 pMemoryRequirements
->alignment
= image
->alignment
;
3171 void radv_GetImageMemoryRequirements2(
3173 const VkImageMemoryRequirementsInfo2KHR
* pInfo
,
3174 VkMemoryRequirements2KHR
* pMemoryRequirements
)
3176 radv_GetImageMemoryRequirements(device
, pInfo
->image
,
3177 &pMemoryRequirements
->memoryRequirements
);
3179 RADV_FROM_HANDLE(radv_image
, image
, pInfo
->image
);
3181 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
3182 switch (ext
->sType
) {
3183 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS_KHR
: {
3184 VkMemoryDedicatedRequirementsKHR
*req
=
3185 (VkMemoryDedicatedRequirementsKHR
*) ext
;
3186 req
->requiresDedicatedAllocation
= image
->shareable
;
3187 req
->prefersDedicatedAllocation
= req
->requiresDedicatedAllocation
;
3196 void radv_GetImageSparseMemoryRequirements(
3199 uint32_t* pSparseMemoryRequirementCount
,
3200 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
3205 void radv_GetImageSparseMemoryRequirements2(
3207 const VkImageSparseMemoryRequirementsInfo2KHR
* pInfo
,
3208 uint32_t* pSparseMemoryRequirementCount
,
3209 VkSparseImageMemoryRequirements2KHR
* pSparseMemoryRequirements
)
3214 void radv_GetDeviceMemoryCommitment(
3216 VkDeviceMemory memory
,
3217 VkDeviceSize
* pCommittedMemoryInBytes
)
3219 *pCommittedMemoryInBytes
= 0;
3222 VkResult
radv_BindBufferMemory2(VkDevice device
,
3223 uint32_t bindInfoCount
,
3224 const VkBindBufferMemoryInfoKHR
*pBindInfos
)
3226 for (uint32_t i
= 0; i
< bindInfoCount
; ++i
) {
3227 RADV_FROM_HANDLE(radv_device_memory
, mem
, pBindInfos
[i
].memory
);
3228 RADV_FROM_HANDLE(radv_buffer
, buffer
, pBindInfos
[i
].buffer
);
3231 buffer
->bo
= mem
->bo
;
3232 buffer
->offset
= pBindInfos
[i
].memoryOffset
;
3240 VkResult
radv_BindBufferMemory(
3243 VkDeviceMemory memory
,
3244 VkDeviceSize memoryOffset
)
3246 const VkBindBufferMemoryInfoKHR info
= {
3247 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO_KHR
,
3250 .memoryOffset
= memoryOffset
3253 return radv_BindBufferMemory2(device
, 1, &info
);
3256 VkResult
radv_BindImageMemory2(VkDevice device
,
3257 uint32_t bindInfoCount
,
3258 const VkBindImageMemoryInfoKHR
*pBindInfos
)
3260 for (uint32_t i
= 0; i
< bindInfoCount
; ++i
) {
3261 RADV_FROM_HANDLE(radv_device_memory
, mem
, pBindInfos
[i
].memory
);
3262 RADV_FROM_HANDLE(radv_image
, image
, pBindInfos
[i
].image
);
3265 image
->bo
= mem
->bo
;
3266 image
->offset
= pBindInfos
[i
].memoryOffset
;
3276 VkResult
radv_BindImageMemory(
3279 VkDeviceMemory memory
,
3280 VkDeviceSize memoryOffset
)
3282 const VkBindImageMemoryInfoKHR info
= {
3283 .sType
= VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO_KHR
,
3286 .memoryOffset
= memoryOffset
3289 return radv_BindImageMemory2(device
, 1, &info
);
3294 radv_sparse_buffer_bind_memory(struct radv_device
*device
,
3295 const VkSparseBufferMemoryBindInfo
*bind
)
3297 RADV_FROM_HANDLE(radv_buffer
, buffer
, bind
->buffer
);
3299 for (uint32_t i
= 0; i
< bind
->bindCount
; ++i
) {
3300 struct radv_device_memory
*mem
= NULL
;
3302 if (bind
->pBinds
[i
].memory
!= VK_NULL_HANDLE
)
3303 mem
= radv_device_memory_from_handle(bind
->pBinds
[i
].memory
);
3305 device
->ws
->buffer_virtual_bind(buffer
->bo
,
3306 bind
->pBinds
[i
].resourceOffset
,
3307 bind
->pBinds
[i
].size
,
3308 mem
? mem
->bo
: NULL
,
3309 bind
->pBinds
[i
].memoryOffset
);
3314 radv_sparse_image_opaque_bind_memory(struct radv_device
*device
,
3315 const VkSparseImageOpaqueMemoryBindInfo
*bind
)
3317 RADV_FROM_HANDLE(radv_image
, image
, bind
->image
);
3319 for (uint32_t i
= 0; i
< bind
->bindCount
; ++i
) {
3320 struct radv_device_memory
*mem
= NULL
;
3322 if (bind
->pBinds
[i
].memory
!= VK_NULL_HANDLE
)
3323 mem
= radv_device_memory_from_handle(bind
->pBinds
[i
].memory
);
3325 device
->ws
->buffer_virtual_bind(image
->bo
,
3326 bind
->pBinds
[i
].resourceOffset
,
3327 bind
->pBinds
[i
].size
,
3328 mem
? mem
->bo
: NULL
,
3329 bind
->pBinds
[i
].memoryOffset
);
3333 VkResult
radv_QueueBindSparse(
3335 uint32_t bindInfoCount
,
3336 const VkBindSparseInfo
* pBindInfo
,
3339 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
3340 RADV_FROM_HANDLE(radv_queue
, queue
, _queue
);
3341 struct radeon_winsys_fence
*base_fence
= fence
? fence
->fence
: NULL
;
3342 bool fence_emitted
= false;
3346 for (uint32_t i
= 0; i
< bindInfoCount
; ++i
) {
3347 struct radv_winsys_sem_info sem_info
;
3348 for (uint32_t j
= 0; j
< pBindInfo
[i
].bufferBindCount
; ++j
) {
3349 radv_sparse_buffer_bind_memory(queue
->device
,
3350 pBindInfo
[i
].pBufferBinds
+ j
);
3353 for (uint32_t j
= 0; j
< pBindInfo
[i
].imageOpaqueBindCount
; ++j
) {
3354 radv_sparse_image_opaque_bind_memory(queue
->device
,
3355 pBindInfo
[i
].pImageOpaqueBinds
+ j
);
3359 result
= radv_alloc_sem_info(queue
->device
->instance
,
3361 pBindInfo
[i
].waitSemaphoreCount
,
3362 pBindInfo
[i
].pWaitSemaphores
,
3363 pBindInfo
[i
].signalSemaphoreCount
,
3364 pBindInfo
[i
].pSignalSemaphores
,
3366 if (result
!= VK_SUCCESS
)
3369 if (pBindInfo
[i
].waitSemaphoreCount
|| pBindInfo
[i
].signalSemaphoreCount
) {
3370 ret
= queue
->device
->ws
->cs_submit(queue
->hw_ctx
, queue
->queue_idx
,
3371 &queue
->device
->empty_cs
[queue
->queue_family_index
],
3376 radv_loge("failed to submit CS %d\n", i
);
3380 fence_emitted
= true;
3382 fence
->submitted
= true;
3385 radv_free_sem_info(&sem_info
);
3390 if (!fence_emitted
) {
3391 result
= radv_signal_fence(queue
, fence
);
3392 if (result
!= VK_SUCCESS
)
3395 fence
->submitted
= true;
3401 VkResult
radv_CreateFence(
3403 const VkFenceCreateInfo
* pCreateInfo
,
3404 const VkAllocationCallbacks
* pAllocator
,
3407 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3408 const VkExportFenceCreateInfoKHR
*export
=
3409 vk_find_struct_const(pCreateInfo
->pNext
, EXPORT_FENCE_CREATE_INFO_KHR
);
3410 VkExternalFenceHandleTypeFlagsKHR handleTypes
=
3411 export
? export
->handleTypes
: 0;
3413 struct radv_fence
*fence
= vk_alloc2(&device
->alloc
, pAllocator
,
3415 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3418 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3420 fence
->fence_wsi
= NULL
;
3421 fence
->submitted
= false;
3422 fence
->signalled
= !!(pCreateInfo
->flags
& VK_FENCE_CREATE_SIGNALED_BIT
);
3423 fence
->temp_syncobj
= 0;
3424 if (device
->always_use_syncobj
|| handleTypes
) {
3425 int ret
= device
->ws
->create_syncobj(device
->ws
, &fence
->syncobj
);
3427 vk_free2(&device
->alloc
, pAllocator
, fence
);
3428 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3430 if (pCreateInfo
->flags
& VK_FENCE_CREATE_SIGNALED_BIT
) {
3431 device
->ws
->signal_syncobj(device
->ws
, fence
->syncobj
);
3433 fence
->fence
= NULL
;
3435 fence
->fence
= device
->ws
->create_fence();
3436 if (!fence
->fence
) {
3437 vk_free2(&device
->alloc
, pAllocator
, fence
);
3438 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3443 *pFence
= radv_fence_to_handle(fence
);
3448 void radv_DestroyFence(
3451 const VkAllocationCallbacks
* pAllocator
)
3453 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3454 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
3459 if (fence
->temp_syncobj
)
3460 device
->ws
->destroy_syncobj(device
->ws
, fence
->temp_syncobj
);
3462 device
->ws
->destroy_syncobj(device
->ws
, fence
->syncobj
);
3464 device
->ws
->destroy_fence(fence
->fence
);
3465 if (fence
->fence_wsi
)
3466 fence
->fence_wsi
->destroy(fence
->fence_wsi
);
3467 vk_free2(&device
->alloc
, pAllocator
, fence
);
3471 static uint64_t radv_get_current_time()
3474 clock_gettime(CLOCK_MONOTONIC
, &tv
);
3475 return tv
.tv_nsec
+ tv
.tv_sec
*1000000000ull;
3478 static uint64_t radv_get_absolute_timeout(uint64_t timeout
)
3480 uint64_t current_time
= radv_get_current_time();
3482 timeout
= MIN2(UINT64_MAX
- current_time
, timeout
);
3484 return current_time
+ timeout
;
3488 static bool radv_all_fences_plain_and_submitted(uint32_t fenceCount
, const VkFence
*pFences
)
3490 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
3491 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
3492 if (fence
->fence
== NULL
|| fence
->syncobj
||
3493 fence
->temp_syncobj
||
3494 (!fence
->signalled
&& !fence
->submitted
))
3500 static bool radv_all_fences_syncobj(uint32_t fenceCount
, const VkFence
*pFences
)
3502 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
3503 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
3504 if (fence
->syncobj
== 0 && fence
->temp_syncobj
== 0)
3510 VkResult
radv_WaitForFences(
3512 uint32_t fenceCount
,
3513 const VkFence
* pFences
,
3517 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3518 timeout
= radv_get_absolute_timeout(timeout
);
3520 if (device
->always_use_syncobj
&&
3521 radv_all_fences_syncobj(fenceCount
, pFences
))
3523 uint32_t *handles
= malloc(sizeof(uint32_t) * fenceCount
);
3525 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3527 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
3528 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
3529 handles
[i
] = fence
->temp_syncobj
? fence
->temp_syncobj
: fence
->syncobj
;
3532 bool success
= device
->ws
->wait_syncobj(device
->ws
, handles
, fenceCount
, waitAll
, timeout
);
3535 return success
? VK_SUCCESS
: VK_TIMEOUT
;
3538 if (!waitAll
&& fenceCount
> 1) {
3539 /* Not doing this by default for waitAll, due to needing to allocate twice. */
3540 if (device
->physical_device
->rad_info
.drm_minor
>= 10 && radv_all_fences_plain_and_submitted(fenceCount
, pFences
)) {
3541 uint32_t wait_count
= 0;
3542 struct radeon_winsys_fence
**fences
= malloc(sizeof(struct radeon_winsys_fence
*) * fenceCount
);
3544 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3546 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
3547 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
3549 if (fence
->signalled
) {
3554 fences
[wait_count
++] = fence
->fence
;
3557 bool success
= device
->ws
->fences_wait(device
->ws
, fences
, wait_count
,
3558 waitAll
, timeout
- radv_get_current_time());
3561 return success
? VK_SUCCESS
: VK_TIMEOUT
;
3564 while(radv_get_current_time() <= timeout
) {
3565 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
3566 if (radv_GetFenceStatus(_device
, pFences
[i
]) == VK_SUCCESS
)
3573 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
3574 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
3575 bool expired
= false;
3577 if (fence
->temp_syncobj
) {
3578 if (!device
->ws
->wait_syncobj(device
->ws
, &fence
->temp_syncobj
, 1, true, timeout
))
3583 if (fence
->syncobj
) {
3584 if (!device
->ws
->wait_syncobj(device
->ws
, &fence
->syncobj
, 1, true, timeout
))
3589 if (fence
->signalled
)
3593 if (!fence
->submitted
) {
3594 while(radv_get_current_time() <= timeout
&&
3598 if (!fence
->submitted
)
3601 /* Recheck as it may have been set by
3602 * submitting operations. */
3604 if (fence
->signalled
)
3608 expired
= device
->ws
->fence_wait(device
->ws
,
3615 if (fence
->fence_wsi
) {
3616 VkResult result
= fence
->fence_wsi
->wait(fence
->fence_wsi
, timeout
);
3617 if (result
!= VK_SUCCESS
)
3621 fence
->signalled
= true;
3627 VkResult
radv_ResetFences(VkDevice _device
,
3628 uint32_t fenceCount
,
3629 const VkFence
*pFences
)
3631 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3633 for (unsigned i
= 0; i
< fenceCount
; ++i
) {
3634 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
3635 fence
->submitted
= fence
->signalled
= false;
3637 /* Per spec, we first restore the permanent payload, and then reset, so
3638 * having a temp syncobj should not skip resetting the permanent syncobj. */
3639 if (fence
->temp_syncobj
) {
3640 device
->ws
->destroy_syncobj(device
->ws
, fence
->temp_syncobj
);
3641 fence
->temp_syncobj
= 0;
3644 if (fence
->syncobj
) {
3645 device
->ws
->reset_syncobj(device
->ws
, fence
->syncobj
);
3652 VkResult
radv_GetFenceStatus(VkDevice _device
, VkFence _fence
)
3654 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3655 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
3657 if (fence
->temp_syncobj
) {
3658 bool success
= device
->ws
->wait_syncobj(device
->ws
, &fence
->temp_syncobj
, 1, true, 0);
3659 return success
? VK_SUCCESS
: VK_NOT_READY
;
3662 if (fence
->syncobj
) {
3663 bool success
= device
->ws
->wait_syncobj(device
->ws
, &fence
->syncobj
, 1, true, 0);
3664 return success
? VK_SUCCESS
: VK_NOT_READY
;
3667 if (fence
->signalled
)
3669 if (!fence
->submitted
)
3670 return VK_NOT_READY
;
3672 if (!device
->ws
->fence_wait(device
->ws
, fence
->fence
, false, 0))
3673 return VK_NOT_READY
;
3675 if (fence
->fence_wsi
) {
3676 VkResult result
= fence
->fence_wsi
->wait(fence
->fence_wsi
, 0);
3678 if (result
!= VK_SUCCESS
) {
3679 if (result
== VK_TIMEOUT
)
3680 return VK_NOT_READY
;
3688 // Queue semaphore functions
3690 VkResult
radv_CreateSemaphore(
3692 const VkSemaphoreCreateInfo
* pCreateInfo
,
3693 const VkAllocationCallbacks
* pAllocator
,
3694 VkSemaphore
* pSemaphore
)
3696 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3697 const VkExportSemaphoreCreateInfoKHR
*export
=
3698 vk_find_struct_const(pCreateInfo
->pNext
, EXPORT_SEMAPHORE_CREATE_INFO_KHR
);
3699 VkExternalSemaphoreHandleTypeFlagsKHR handleTypes
=
3700 export
? export
->handleTypes
: 0;
3702 struct radv_semaphore
*sem
= vk_alloc2(&device
->alloc
, pAllocator
,
3704 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3706 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3708 sem
->temp_syncobj
= 0;
3709 /* create a syncobject if we are going to export this semaphore */
3710 if (device
->always_use_syncobj
|| handleTypes
) {
3711 assert (device
->physical_device
->rad_info
.has_syncobj
);
3712 int ret
= device
->ws
->create_syncobj(device
->ws
, &sem
->syncobj
);
3714 vk_free2(&device
->alloc
, pAllocator
, sem
);
3715 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3719 sem
->sem
= device
->ws
->create_sem(device
->ws
);
3721 vk_free2(&device
->alloc
, pAllocator
, sem
);
3722 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3727 *pSemaphore
= radv_semaphore_to_handle(sem
);
3731 void radv_DestroySemaphore(
3733 VkSemaphore _semaphore
,
3734 const VkAllocationCallbacks
* pAllocator
)
3736 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3737 RADV_FROM_HANDLE(radv_semaphore
, sem
, _semaphore
);
3742 device
->ws
->destroy_syncobj(device
->ws
, sem
->syncobj
);
3744 device
->ws
->destroy_sem(sem
->sem
);
3745 vk_free2(&device
->alloc
, pAllocator
, sem
);
3748 VkResult
radv_CreateEvent(
3750 const VkEventCreateInfo
* pCreateInfo
,
3751 const VkAllocationCallbacks
* pAllocator
,
3754 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3755 struct radv_event
*event
= vk_alloc2(&device
->alloc
, pAllocator
,
3757 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3760 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3762 event
->bo
= device
->ws
->buffer_create(device
->ws
, 8, 8,
3764 RADEON_FLAG_VA_UNCACHED
| RADEON_FLAG_CPU_ACCESS
| RADEON_FLAG_NO_INTERPROCESS_SHARING
);
3766 vk_free2(&device
->alloc
, pAllocator
, event
);
3767 return vk_error(device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
3770 event
->map
= (uint64_t*)device
->ws
->buffer_map(event
->bo
);
3772 *pEvent
= radv_event_to_handle(event
);
3777 void radv_DestroyEvent(
3780 const VkAllocationCallbacks
* pAllocator
)
3782 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3783 RADV_FROM_HANDLE(radv_event
, event
, _event
);
3787 device
->ws
->buffer_destroy(event
->bo
);
3788 vk_free2(&device
->alloc
, pAllocator
, event
);
3791 VkResult
radv_GetEventStatus(
3795 RADV_FROM_HANDLE(radv_event
, event
, _event
);
3797 if (*event
->map
== 1)
3798 return VK_EVENT_SET
;
3799 return VK_EVENT_RESET
;
3802 VkResult
radv_SetEvent(
3806 RADV_FROM_HANDLE(radv_event
, event
, _event
);
3812 VkResult
radv_ResetEvent(
3816 RADV_FROM_HANDLE(radv_event
, event
, _event
);
3822 VkResult
radv_CreateBuffer(
3824 const VkBufferCreateInfo
* pCreateInfo
,
3825 const VkAllocationCallbacks
* pAllocator
,
3828 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3829 struct radv_buffer
*buffer
;
3831 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
3833 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
3834 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3836 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
3838 buffer
->size
= pCreateInfo
->size
;
3839 buffer
->usage
= pCreateInfo
->usage
;
3842 buffer
->flags
= pCreateInfo
->flags
;
3844 buffer
->shareable
= vk_find_struct_const(pCreateInfo
->pNext
,
3845 EXTERNAL_MEMORY_BUFFER_CREATE_INFO_KHR
) != NULL
;
3847 if (pCreateInfo
->flags
& VK_BUFFER_CREATE_SPARSE_BINDING_BIT
) {
3848 buffer
->bo
= device
->ws
->buffer_create(device
->ws
,
3849 align64(buffer
->size
, 4096),
3850 4096, 0, RADEON_FLAG_VIRTUAL
);
3852 vk_free2(&device
->alloc
, pAllocator
, buffer
);
3853 return vk_error(device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
3857 *pBuffer
= radv_buffer_to_handle(buffer
);
3862 void radv_DestroyBuffer(
3865 const VkAllocationCallbacks
* pAllocator
)
3867 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3868 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
3873 if (buffer
->flags
& VK_BUFFER_CREATE_SPARSE_BINDING_BIT
)
3874 device
->ws
->buffer_destroy(buffer
->bo
);
3876 vk_free2(&device
->alloc
, pAllocator
, buffer
);
3879 static inline unsigned
3880 si_tile_mode_index(const struct radv_image
*image
, unsigned level
, bool stencil
)
3883 return image
->surface
.u
.legacy
.stencil_tiling_index
[level
];
3885 return image
->surface
.u
.legacy
.tiling_index
[level
];
3888 static uint32_t radv_surface_max_layer_count(struct radv_image_view
*iview
)
3890 return iview
->type
== VK_IMAGE_VIEW_TYPE_3D
? iview
->extent
.depth
: (iview
->base_layer
+ iview
->layer_count
);
3894 radv_init_dcc_control_reg(struct radv_device
*device
,
3895 struct radv_image_view
*iview
)
3897 unsigned max_uncompressed_block_size
= V_028C78_MAX_BLOCK_SIZE_256B
;
3898 unsigned min_compressed_block_size
= V_028C78_MIN_BLOCK_SIZE_32B
;
3899 unsigned max_compressed_block_size
;
3900 unsigned independent_64b_blocks
;
3902 if (device
->physical_device
->rad_info
.chip_class
< VI
)
3905 if (iview
->image
->info
.samples
> 1) {
3906 if (iview
->image
->surface
.bpe
== 1)
3907 max_uncompressed_block_size
= V_028C78_MAX_BLOCK_SIZE_64B
;
3908 else if (iview
->image
->surface
.bpe
== 2)
3909 max_uncompressed_block_size
= V_028C78_MAX_BLOCK_SIZE_128B
;
3912 if (!device
->physical_device
->rad_info
.has_dedicated_vram
) {
3913 /* amdvlk: [min-compressed-block-size] should be set to 32 for
3914 * dGPU and 64 for APU because all of our APUs to date use
3915 * DIMMs which have a request granularity size of 64B while all
3916 * other chips have a 32B request size.
3918 min_compressed_block_size
= V_028C78_MIN_BLOCK_SIZE_64B
;
3921 if (iview
->image
->usage
& (VK_IMAGE_USAGE_SAMPLED_BIT
|
3922 VK_IMAGE_USAGE_TRANSFER_SRC_BIT
|
3923 VK_IMAGE_USAGE_INPUT_ATTACHMENT_BIT
)) {
3924 /* If this DCC image is potentially going to be used in texture
3925 * fetches, we need some special settings.
3927 independent_64b_blocks
= 1;
3928 max_compressed_block_size
= V_028C78_MAX_BLOCK_SIZE_64B
;
3930 /* MAX_UNCOMPRESSED_BLOCK_SIZE must be >=
3931 * MAX_COMPRESSED_BLOCK_SIZE. Set MAX_COMPRESSED_BLOCK_SIZE as
3932 * big as possible for better compression state.
3934 independent_64b_blocks
= 0;
3935 max_compressed_block_size
= max_uncompressed_block_size
;
3938 return S_028C78_MAX_UNCOMPRESSED_BLOCK_SIZE(max_uncompressed_block_size
) |
3939 S_028C78_MAX_COMPRESSED_BLOCK_SIZE(max_compressed_block_size
) |
3940 S_028C78_MIN_COMPRESSED_BLOCK_SIZE(min_compressed_block_size
) |
3941 S_028C78_INDEPENDENT_64B_BLOCKS(independent_64b_blocks
);
3945 radv_initialise_color_surface(struct radv_device
*device
,
3946 struct radv_color_buffer_info
*cb
,
3947 struct radv_image_view
*iview
)
3949 const struct vk_format_description
*desc
;
3950 unsigned ntype
, format
, swap
, endian
;
3951 unsigned blend_clamp
= 0, blend_bypass
= 0;
3953 const struct radeon_surf
*surf
= &iview
->image
->surface
;
3955 desc
= vk_format_description(iview
->vk_format
);
3957 memset(cb
, 0, sizeof(*cb
));
3959 /* Intensity is implemented as Red, so treat it that way. */
3960 cb
->cb_color_attrib
= S_028C74_FORCE_DST_ALPHA_1(desc
->swizzle
[3] == VK_SWIZZLE_1
);
3962 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
;
3964 cb
->cb_color_base
= va
>> 8;
3966 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
3967 struct gfx9_surf_meta_flags meta
;
3968 if (iview
->image
->dcc_offset
)
3969 meta
= iview
->image
->surface
.u
.gfx9
.dcc
;
3971 meta
= iview
->image
->surface
.u
.gfx9
.cmask
;
3973 cb
->cb_color_attrib
|= S_028C74_COLOR_SW_MODE(iview
->image
->surface
.u
.gfx9
.surf
.swizzle_mode
) |
3974 S_028C74_FMASK_SW_MODE(iview
->image
->surface
.u
.gfx9
.fmask
.swizzle_mode
) |
3975 S_028C74_RB_ALIGNED(meta
.rb_aligned
) |
3976 S_028C74_PIPE_ALIGNED(meta
.pipe_aligned
);
3978 cb
->cb_color_base
+= iview
->image
->surface
.u
.gfx9
.surf_offset
>> 8;
3979 cb
->cb_color_base
|= iview
->image
->surface
.tile_swizzle
;
3981 const struct legacy_surf_level
*level_info
= &surf
->u
.legacy
.level
[iview
->base_mip
];
3982 unsigned pitch_tile_max
, slice_tile_max
, tile_mode_index
;
3984 cb
->cb_color_base
+= level_info
->offset
>> 8;
3985 if (level_info
->mode
== RADEON_SURF_MODE_2D
)
3986 cb
->cb_color_base
|= iview
->image
->surface
.tile_swizzle
;
3988 pitch_tile_max
= level_info
->nblk_x
/ 8 - 1;
3989 slice_tile_max
= (level_info
->nblk_x
* level_info
->nblk_y
) / 64 - 1;
3990 tile_mode_index
= si_tile_mode_index(iview
->image
, iview
->base_mip
, false);
3992 cb
->cb_color_pitch
= S_028C64_TILE_MAX(pitch_tile_max
);
3993 cb
->cb_color_slice
= S_028C68_TILE_MAX(slice_tile_max
);
3994 cb
->cb_color_cmask_slice
= iview
->image
->cmask
.slice_tile_max
;
3996 cb
->cb_color_attrib
|= S_028C74_TILE_MODE_INDEX(tile_mode_index
);
3998 if (radv_image_has_fmask(iview
->image
)) {
3999 if (device
->physical_device
->rad_info
.chip_class
>= CIK
)
4000 cb
->cb_color_pitch
|= S_028C64_FMASK_TILE_MAX(iview
->image
->fmask
.pitch_in_pixels
/ 8 - 1);
4001 cb
->cb_color_attrib
|= S_028C74_FMASK_TILE_MODE_INDEX(iview
->image
->fmask
.tile_mode_index
);
4002 cb
->cb_color_fmask_slice
= S_028C88_TILE_MAX(iview
->image
->fmask
.slice_tile_max
);
4004 /* This must be set for fast clear to work without FMASK. */
4005 if (device
->physical_device
->rad_info
.chip_class
>= CIK
)
4006 cb
->cb_color_pitch
|= S_028C64_FMASK_TILE_MAX(pitch_tile_max
);
4007 cb
->cb_color_attrib
|= S_028C74_FMASK_TILE_MODE_INDEX(tile_mode_index
);
4008 cb
->cb_color_fmask_slice
= S_028C88_TILE_MAX(slice_tile_max
);
4012 /* CMASK variables */
4013 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
;
4014 va
+= iview
->image
->cmask
.offset
;
4015 cb
->cb_color_cmask
= va
>> 8;
4017 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
;
4018 va
+= iview
->image
->dcc_offset
;
4019 cb
->cb_dcc_base
= va
>> 8;
4020 cb
->cb_dcc_base
|= iview
->image
->surface
.tile_swizzle
;
4022 uint32_t max_slice
= radv_surface_max_layer_count(iview
) - 1;
4023 cb
->cb_color_view
= S_028C6C_SLICE_START(iview
->base_layer
) |
4024 S_028C6C_SLICE_MAX(max_slice
);
4026 if (iview
->image
->info
.samples
> 1) {
4027 unsigned log_samples
= util_logbase2(iview
->image
->info
.samples
);
4029 cb
->cb_color_attrib
|= S_028C74_NUM_SAMPLES(log_samples
) |
4030 S_028C74_NUM_FRAGMENTS(log_samples
);
4033 if (radv_image_has_fmask(iview
->image
)) {
4034 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
+ iview
->image
->fmask
.offset
;
4035 cb
->cb_color_fmask
= va
>> 8;
4036 cb
->cb_color_fmask
|= iview
->image
->fmask
.tile_swizzle
;
4038 cb
->cb_color_fmask
= cb
->cb_color_base
;
4041 ntype
= radv_translate_color_numformat(iview
->vk_format
,
4043 vk_format_get_first_non_void_channel(iview
->vk_format
));
4044 format
= radv_translate_colorformat(iview
->vk_format
);
4045 if (format
== V_028C70_COLOR_INVALID
|| ntype
== ~0u)
4046 radv_finishme("Illegal color\n");
4047 swap
= radv_translate_colorswap(iview
->vk_format
, FALSE
);
4048 endian
= radv_colorformat_endian_swap(format
);
4050 /* blend clamp should be set for all NORM/SRGB types */
4051 if (ntype
== V_028C70_NUMBER_UNORM
||
4052 ntype
== V_028C70_NUMBER_SNORM
||
4053 ntype
== V_028C70_NUMBER_SRGB
)
4056 /* set blend bypass according to docs if SINT/UINT or
4057 8/24 COLOR variants */
4058 if (ntype
== V_028C70_NUMBER_UINT
|| ntype
== V_028C70_NUMBER_SINT
||
4059 format
== V_028C70_COLOR_8_24
|| format
== V_028C70_COLOR_24_8
||
4060 format
== V_028C70_COLOR_X24_8_32_FLOAT
) {
4065 if ((ntype
== V_028C70_NUMBER_UINT
|| ntype
== V_028C70_NUMBER_SINT
) &&
4066 (format
== V_028C70_COLOR_8
||
4067 format
== V_028C70_COLOR_8_8
||
4068 format
== V_028C70_COLOR_8_8_8_8
))
4069 ->color_is_int8
= true;
4071 cb
->cb_color_info
= S_028C70_FORMAT(format
) |
4072 S_028C70_COMP_SWAP(swap
) |
4073 S_028C70_BLEND_CLAMP(blend_clamp
) |
4074 S_028C70_BLEND_BYPASS(blend_bypass
) |
4075 S_028C70_SIMPLE_FLOAT(1) |
4076 S_028C70_ROUND_MODE(ntype
!= V_028C70_NUMBER_UNORM
&&
4077 ntype
!= V_028C70_NUMBER_SNORM
&&
4078 ntype
!= V_028C70_NUMBER_SRGB
&&
4079 format
!= V_028C70_COLOR_8_24
&&
4080 format
!= V_028C70_COLOR_24_8
) |
4081 S_028C70_NUMBER_TYPE(ntype
) |
4082 S_028C70_ENDIAN(endian
);
4083 if (radv_image_has_fmask(iview
->image
)) {
4084 cb
->cb_color_info
|= S_028C70_COMPRESSION(1);
4085 if (device
->physical_device
->rad_info
.chip_class
== SI
) {
4086 unsigned fmask_bankh
= util_logbase2(iview
->image
->fmask
.bank_height
);
4087 cb
->cb_color_attrib
|= S_028C74_FMASK_BANK_HEIGHT(fmask_bankh
);
4091 if (radv_image_has_cmask(iview
->image
) &&
4092 !(device
->instance
->debug_flags
& RADV_DEBUG_NO_FAST_CLEARS
))
4093 cb
->cb_color_info
|= S_028C70_FAST_CLEAR(1);
4095 if (radv_dcc_enabled(iview
->image
, iview
->base_mip
))
4096 cb
->cb_color_info
|= S_028C70_DCC_ENABLE(1);
4098 cb
->cb_dcc_control
= radv_init_dcc_control_reg(device
, iview
);
4100 /* This must be set for fast clear to work without FMASK. */
4101 if (!radv_image_has_fmask(iview
->image
) &&
4102 device
->physical_device
->rad_info
.chip_class
== SI
) {
4103 unsigned bankh
= util_logbase2(iview
->image
->surface
.u
.legacy
.bankh
);
4104 cb
->cb_color_attrib
|= S_028C74_FMASK_BANK_HEIGHT(bankh
);
4107 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
4108 unsigned mip0_depth
= iview
->image
->type
== VK_IMAGE_TYPE_3D
?
4109 (iview
->extent
.depth
- 1) : (iview
->image
->info
.array_size
- 1);
4111 cb
->cb_color_view
|= S_028C6C_MIP_LEVEL(iview
->base_mip
);
4112 cb
->cb_color_attrib
|= S_028C74_MIP0_DEPTH(mip0_depth
) |
4113 S_028C74_RESOURCE_TYPE(iview
->image
->surface
.u
.gfx9
.resource_type
);
4114 cb
->cb_color_attrib2
= S_028C68_MIP0_WIDTH(iview
->extent
.width
- 1) |
4115 S_028C68_MIP0_HEIGHT(iview
->extent
.height
- 1) |
4116 S_028C68_MAX_MIP(iview
->image
->info
.levels
- 1);
4121 radv_calc_decompress_on_z_planes(struct radv_device
*device
,
4122 struct radv_image_view
*iview
)
4124 unsigned max_zplanes
= 0;
4126 assert(radv_image_is_tc_compat_htile(iview
->image
));
4128 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
4129 /* Default value for 32-bit depth surfaces. */
4132 if (iview
->vk_format
== VK_FORMAT_D16_UNORM
&&
4133 iview
->image
->info
.samples
> 1)
4136 max_zplanes
= max_zplanes
+ 1;
4138 if (iview
->vk_format
== VK_FORMAT_D16_UNORM
) {
4139 /* Do not enable Z plane compression for 16-bit depth
4140 * surfaces because isn't supported on GFX8. Only
4141 * 32-bit depth surfaces are supported by the hardware.
4142 * This allows to maintain shader compatibility and to
4143 * reduce the number of depth decompressions.
4147 if (iview
->image
->info
.samples
<= 1)
4149 else if (iview
->image
->info
.samples
<= 4)
4160 radv_initialise_ds_surface(struct radv_device
*device
,
4161 struct radv_ds_buffer_info
*ds
,
4162 struct radv_image_view
*iview
)
4164 unsigned level
= iview
->base_mip
;
4165 unsigned format
, stencil_format
;
4166 uint64_t va
, s_offs
, z_offs
;
4167 bool stencil_only
= false;
4168 memset(ds
, 0, sizeof(*ds
));
4169 switch (iview
->image
->vk_format
) {
4170 case VK_FORMAT_D24_UNORM_S8_UINT
:
4171 case VK_FORMAT_X8_D24_UNORM_PACK32
:
4172 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-24);
4173 ds
->offset_scale
= 2.0f
;
4175 case VK_FORMAT_D16_UNORM
:
4176 case VK_FORMAT_D16_UNORM_S8_UINT
:
4177 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-16);
4178 ds
->offset_scale
= 4.0f
;
4180 case VK_FORMAT_D32_SFLOAT
:
4181 case VK_FORMAT_D32_SFLOAT_S8_UINT
:
4182 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-23) |
4183 S_028B78_POLY_OFFSET_DB_IS_FLOAT_FMT(1);
4184 ds
->offset_scale
= 1.0f
;
4186 case VK_FORMAT_S8_UINT
:
4187 stencil_only
= true;
4193 format
= radv_translate_dbformat(iview
->image
->vk_format
);
4194 stencil_format
= iview
->image
->surface
.has_stencil
?
4195 V_028044_STENCIL_8
: V_028044_STENCIL_INVALID
;
4197 uint32_t max_slice
= radv_surface_max_layer_count(iview
) - 1;
4198 ds
->db_depth_view
= S_028008_SLICE_START(iview
->base_layer
) |
4199 S_028008_SLICE_MAX(max_slice
);
4201 ds
->db_htile_data_base
= 0;
4202 ds
->db_htile_surface
= 0;
4204 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
;
4205 s_offs
= z_offs
= va
;
4207 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
4208 assert(iview
->image
->surface
.u
.gfx9
.surf_offset
== 0);
4209 s_offs
+= iview
->image
->surface
.u
.gfx9
.stencil_offset
;
4211 ds
->db_z_info
= S_028038_FORMAT(format
) |
4212 S_028038_NUM_SAMPLES(util_logbase2(iview
->image
->info
.samples
)) |
4213 S_028038_SW_MODE(iview
->image
->surface
.u
.gfx9
.surf
.swizzle_mode
) |
4214 S_028038_MAXMIP(iview
->image
->info
.levels
- 1) |
4215 S_028038_ZRANGE_PRECISION(1);
4216 ds
->db_stencil_info
= S_02803C_FORMAT(stencil_format
) |
4217 S_02803C_SW_MODE(iview
->image
->surface
.u
.gfx9
.stencil
.swizzle_mode
);
4219 ds
->db_z_info2
= S_028068_EPITCH(iview
->image
->surface
.u
.gfx9
.surf
.epitch
);
4220 ds
->db_stencil_info2
= S_02806C_EPITCH(iview
->image
->surface
.u
.gfx9
.stencil
.epitch
);
4221 ds
->db_depth_view
|= S_028008_MIPID(level
);
4223 ds
->db_depth_size
= S_02801C_X_MAX(iview
->image
->info
.width
- 1) |
4224 S_02801C_Y_MAX(iview
->image
->info
.height
- 1);
4226 if (radv_htile_enabled(iview
->image
, level
)) {
4227 ds
->db_z_info
|= S_028038_TILE_SURFACE_ENABLE(1);
4229 if (radv_image_is_tc_compat_htile(iview
->image
)) {
4230 unsigned max_zplanes
=
4231 radv_calc_decompress_on_z_planes(device
, iview
);
4233 ds
->db_z_info
|= S_028038_DECOMPRESS_ON_N_ZPLANES(max_zplanes
) |
4234 S_028038_ITERATE_FLUSH(1);
4235 ds
->db_stencil_info
|= S_02803C_ITERATE_FLUSH(1);
4238 if (!iview
->image
->surface
.has_stencil
)
4239 /* Use all of the htile_buffer for depth if there's no stencil. */
4240 ds
->db_stencil_info
|= S_02803C_TILE_STENCIL_DISABLE(1);
4241 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
+
4242 iview
->image
->htile_offset
;
4243 ds
->db_htile_data_base
= va
>> 8;
4244 ds
->db_htile_surface
= S_028ABC_FULL_CACHE(1) |
4245 S_028ABC_PIPE_ALIGNED(iview
->image
->surface
.u
.gfx9
.htile
.pipe_aligned
) |
4246 S_028ABC_RB_ALIGNED(iview
->image
->surface
.u
.gfx9
.htile
.rb_aligned
);
4249 const struct legacy_surf_level
*level_info
= &iview
->image
->surface
.u
.legacy
.level
[level
];
4252 level_info
= &iview
->image
->surface
.u
.legacy
.stencil_level
[level
];
4254 z_offs
+= iview
->image
->surface
.u
.legacy
.level
[level
].offset
;
4255 s_offs
+= iview
->image
->surface
.u
.legacy
.stencil_level
[level
].offset
;
4257 ds
->db_depth_info
= S_02803C_ADDR5_SWIZZLE_MASK(!radv_image_is_tc_compat_htile(iview
->image
));
4258 ds
->db_z_info
= S_028040_FORMAT(format
) | S_028040_ZRANGE_PRECISION(1);
4259 ds
->db_stencil_info
= S_028044_FORMAT(stencil_format
);
4261 if (iview
->image
->info
.samples
> 1)
4262 ds
->db_z_info
|= S_028040_NUM_SAMPLES(util_logbase2(iview
->image
->info
.samples
));
4264 if (device
->physical_device
->rad_info
.chip_class
>= CIK
) {
4265 struct radeon_info
*info
= &device
->physical_device
->rad_info
;
4266 unsigned tiling_index
= iview
->image
->surface
.u
.legacy
.tiling_index
[level
];
4267 unsigned stencil_index
= iview
->image
->surface
.u
.legacy
.stencil_tiling_index
[level
];
4268 unsigned macro_index
= iview
->image
->surface
.u
.legacy
.macro_tile_index
;
4269 unsigned tile_mode
= info
->si_tile_mode_array
[tiling_index
];
4270 unsigned stencil_tile_mode
= info
->si_tile_mode_array
[stencil_index
];
4271 unsigned macro_mode
= info
->cik_macrotile_mode_array
[macro_index
];
4274 tile_mode
= stencil_tile_mode
;
4276 ds
->db_depth_info
|=
4277 S_02803C_ARRAY_MODE(G_009910_ARRAY_MODE(tile_mode
)) |
4278 S_02803C_PIPE_CONFIG(G_009910_PIPE_CONFIG(tile_mode
)) |
4279 S_02803C_BANK_WIDTH(G_009990_BANK_WIDTH(macro_mode
)) |
4280 S_02803C_BANK_HEIGHT(G_009990_BANK_HEIGHT(macro_mode
)) |
4281 S_02803C_MACRO_TILE_ASPECT(G_009990_MACRO_TILE_ASPECT(macro_mode
)) |
4282 S_02803C_NUM_BANKS(G_009990_NUM_BANKS(macro_mode
));
4283 ds
->db_z_info
|= S_028040_TILE_SPLIT(G_009910_TILE_SPLIT(tile_mode
));
4284 ds
->db_stencil_info
|= S_028044_TILE_SPLIT(G_009910_TILE_SPLIT(stencil_tile_mode
));
4286 unsigned tile_mode_index
= si_tile_mode_index(iview
->image
, level
, false);
4287 ds
->db_z_info
|= S_028040_TILE_MODE_INDEX(tile_mode_index
);
4288 tile_mode_index
= si_tile_mode_index(iview
->image
, level
, true);
4289 ds
->db_stencil_info
|= S_028044_TILE_MODE_INDEX(tile_mode_index
);
4291 ds
->db_z_info
|= S_028040_TILE_MODE_INDEX(tile_mode_index
);
4294 ds
->db_depth_size
= S_028058_PITCH_TILE_MAX((level_info
->nblk_x
/ 8) - 1) |
4295 S_028058_HEIGHT_TILE_MAX((level_info
->nblk_y
/ 8) - 1);
4296 ds
->db_depth_slice
= S_02805C_SLICE_TILE_MAX((level_info
->nblk_x
* level_info
->nblk_y
) / 64 - 1);
4298 if (radv_htile_enabled(iview
->image
, level
)) {
4299 ds
->db_z_info
|= S_028040_TILE_SURFACE_ENABLE(1);
4301 if (!iview
->image
->surface
.has_stencil
&&
4302 !radv_image_is_tc_compat_htile(iview
->image
))
4303 /* Use all of the htile_buffer for depth if there's no stencil. */
4304 ds
->db_stencil_info
|= S_028044_TILE_STENCIL_DISABLE(1);
4306 va
= radv_buffer_get_va(iview
->bo
) + iview
->image
->offset
+
4307 iview
->image
->htile_offset
;
4308 ds
->db_htile_data_base
= va
>> 8;
4309 ds
->db_htile_surface
= S_028ABC_FULL_CACHE(1);
4311 if (radv_image_is_tc_compat_htile(iview
->image
)) {
4312 unsigned max_zplanes
=
4313 radv_calc_decompress_on_z_planes(device
, iview
);
4315 ds
->db_htile_surface
|= S_028ABC_TC_COMPATIBLE(1);
4316 ds
->db_z_info
|= S_028040_DECOMPRESS_ON_N_ZPLANES(max_zplanes
);
4321 ds
->db_z_read_base
= ds
->db_z_write_base
= z_offs
>> 8;
4322 ds
->db_stencil_read_base
= ds
->db_stencil_write_base
= s_offs
>> 8;
4325 VkResult
radv_CreateFramebuffer(
4327 const VkFramebufferCreateInfo
* pCreateInfo
,
4328 const VkAllocationCallbacks
* pAllocator
,
4329 VkFramebuffer
* pFramebuffer
)
4331 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4332 struct radv_framebuffer
*framebuffer
;
4334 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
4336 size_t size
= sizeof(*framebuffer
) +
4337 sizeof(struct radv_attachment_info
) * pCreateInfo
->attachmentCount
;
4338 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
4339 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
4340 if (framebuffer
== NULL
)
4341 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
4343 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
4344 framebuffer
->width
= pCreateInfo
->width
;
4345 framebuffer
->height
= pCreateInfo
->height
;
4346 framebuffer
->layers
= pCreateInfo
->layers
;
4347 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
4348 VkImageView _iview
= pCreateInfo
->pAttachments
[i
];
4349 struct radv_image_view
*iview
= radv_image_view_from_handle(_iview
);
4350 framebuffer
->attachments
[i
].attachment
= iview
;
4351 if (iview
->aspect_mask
& VK_IMAGE_ASPECT_COLOR_BIT
) {
4352 radv_initialise_color_surface(device
, &framebuffer
->attachments
[i
].cb
, iview
);
4353 } else if (iview
->aspect_mask
& (VK_IMAGE_ASPECT_DEPTH_BIT
| VK_IMAGE_ASPECT_STENCIL_BIT
)) {
4354 radv_initialise_ds_surface(device
, &framebuffer
->attachments
[i
].ds
, iview
);
4356 framebuffer
->width
= MIN2(framebuffer
->width
, iview
->extent
.width
);
4357 framebuffer
->height
= MIN2(framebuffer
->height
, iview
->extent
.height
);
4358 framebuffer
->layers
= MIN2(framebuffer
->layers
, radv_surface_max_layer_count(iview
));
4361 *pFramebuffer
= radv_framebuffer_to_handle(framebuffer
);
4365 void radv_DestroyFramebuffer(
4368 const VkAllocationCallbacks
* pAllocator
)
4370 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4371 RADV_FROM_HANDLE(radv_framebuffer
, fb
, _fb
);
4375 vk_free2(&device
->alloc
, pAllocator
, fb
);
4378 static unsigned radv_tex_wrap(VkSamplerAddressMode address_mode
)
4380 switch (address_mode
) {
4381 case VK_SAMPLER_ADDRESS_MODE_REPEAT
:
4382 return V_008F30_SQ_TEX_WRAP
;
4383 case VK_SAMPLER_ADDRESS_MODE_MIRRORED_REPEAT
:
4384 return V_008F30_SQ_TEX_MIRROR
;
4385 case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE
:
4386 return V_008F30_SQ_TEX_CLAMP_LAST_TEXEL
;
4387 case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_BORDER
:
4388 return V_008F30_SQ_TEX_CLAMP_BORDER
;
4389 case VK_SAMPLER_ADDRESS_MODE_MIRROR_CLAMP_TO_EDGE
:
4390 return V_008F30_SQ_TEX_MIRROR_ONCE_LAST_TEXEL
;
4392 unreachable("illegal tex wrap mode");
4398 radv_tex_compare(VkCompareOp op
)
4401 case VK_COMPARE_OP_NEVER
:
4402 return V_008F30_SQ_TEX_DEPTH_COMPARE_NEVER
;
4403 case VK_COMPARE_OP_LESS
:
4404 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESS
;
4405 case VK_COMPARE_OP_EQUAL
:
4406 return V_008F30_SQ_TEX_DEPTH_COMPARE_EQUAL
;
4407 case VK_COMPARE_OP_LESS_OR_EQUAL
:
4408 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESSEQUAL
;
4409 case VK_COMPARE_OP_GREATER
:
4410 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATER
;
4411 case VK_COMPARE_OP_NOT_EQUAL
:
4412 return V_008F30_SQ_TEX_DEPTH_COMPARE_NOTEQUAL
;
4413 case VK_COMPARE_OP_GREATER_OR_EQUAL
:
4414 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATEREQUAL
;
4415 case VK_COMPARE_OP_ALWAYS
:
4416 return V_008F30_SQ_TEX_DEPTH_COMPARE_ALWAYS
;
4418 unreachable("illegal compare mode");
4424 radv_tex_filter(VkFilter filter
, unsigned max_ansio
)
4427 case VK_FILTER_NEAREST
:
4428 return (max_ansio
> 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_POINT
:
4429 V_008F38_SQ_TEX_XY_FILTER_POINT
);
4430 case VK_FILTER_LINEAR
:
4431 return (max_ansio
> 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_BILINEAR
:
4432 V_008F38_SQ_TEX_XY_FILTER_BILINEAR
);
4433 case VK_FILTER_CUBIC_IMG
:
4435 fprintf(stderr
, "illegal texture filter");
4441 radv_tex_mipfilter(VkSamplerMipmapMode mode
)
4444 case VK_SAMPLER_MIPMAP_MODE_NEAREST
:
4445 return V_008F38_SQ_TEX_Z_FILTER_POINT
;
4446 case VK_SAMPLER_MIPMAP_MODE_LINEAR
:
4447 return V_008F38_SQ_TEX_Z_FILTER_LINEAR
;
4449 return V_008F38_SQ_TEX_Z_FILTER_NONE
;
4454 radv_tex_bordercolor(VkBorderColor bcolor
)
4457 case VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
:
4458 case VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
:
4459 return V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK
;
4460 case VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
:
4461 case VK_BORDER_COLOR_INT_OPAQUE_BLACK
:
4462 return V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_BLACK
;
4463 case VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
:
4464 case VK_BORDER_COLOR_INT_OPAQUE_WHITE
:
4465 return V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_WHITE
;
4473 radv_tex_aniso_filter(unsigned filter
)
4487 radv_tex_filter_mode(VkSamplerReductionModeEXT mode
)
4490 case VK_SAMPLER_REDUCTION_MODE_WEIGHTED_AVERAGE_EXT
:
4491 return SQ_IMG_FILTER_MODE_BLEND
;
4492 case VK_SAMPLER_REDUCTION_MODE_MIN_EXT
:
4493 return SQ_IMG_FILTER_MODE_MIN
;
4494 case VK_SAMPLER_REDUCTION_MODE_MAX_EXT
:
4495 return SQ_IMG_FILTER_MODE_MAX
;
4503 radv_get_max_anisotropy(struct radv_device
*device
,
4504 const VkSamplerCreateInfo
*pCreateInfo
)
4506 if (device
->force_aniso
>= 0)
4507 return device
->force_aniso
;
4509 if (pCreateInfo
->anisotropyEnable
&&
4510 pCreateInfo
->maxAnisotropy
> 1.0f
)
4511 return (uint32_t)pCreateInfo
->maxAnisotropy
;
4517 radv_init_sampler(struct radv_device
*device
,
4518 struct radv_sampler
*sampler
,
4519 const VkSamplerCreateInfo
*pCreateInfo
)
4521 uint32_t max_aniso
= radv_get_max_anisotropy(device
, pCreateInfo
);
4522 uint32_t max_aniso_ratio
= radv_tex_aniso_filter(max_aniso
);
4523 bool is_vi
= (device
->physical_device
->rad_info
.chip_class
>= VI
);
4524 unsigned filter_mode
= SQ_IMG_FILTER_MODE_BLEND
;
4526 const struct VkSamplerReductionModeCreateInfoEXT
*sampler_reduction
=
4527 vk_find_struct_const(pCreateInfo
->pNext
,
4528 SAMPLER_REDUCTION_MODE_CREATE_INFO_EXT
);
4529 if (sampler_reduction
)
4530 filter_mode
= radv_tex_filter_mode(sampler_reduction
->reductionMode
);
4532 sampler
->state
[0] = (S_008F30_CLAMP_X(radv_tex_wrap(pCreateInfo
->addressModeU
)) |
4533 S_008F30_CLAMP_Y(radv_tex_wrap(pCreateInfo
->addressModeV
)) |
4534 S_008F30_CLAMP_Z(radv_tex_wrap(pCreateInfo
->addressModeW
)) |
4535 S_008F30_MAX_ANISO_RATIO(max_aniso_ratio
) |
4536 S_008F30_DEPTH_COMPARE_FUNC(radv_tex_compare(pCreateInfo
->compareOp
)) |
4537 S_008F30_FORCE_UNNORMALIZED(pCreateInfo
->unnormalizedCoordinates
? 1 : 0) |
4538 S_008F30_ANISO_THRESHOLD(max_aniso_ratio
>> 1) |
4539 S_008F30_ANISO_BIAS(max_aniso_ratio
) |
4540 S_008F30_DISABLE_CUBE_WRAP(0) |
4541 S_008F30_COMPAT_MODE(is_vi
) |
4542 S_008F30_FILTER_MODE(filter_mode
));
4543 sampler
->state
[1] = (S_008F34_MIN_LOD(S_FIXED(CLAMP(pCreateInfo
->minLod
, 0, 15), 8)) |
4544 S_008F34_MAX_LOD(S_FIXED(CLAMP(pCreateInfo
->maxLod
, 0, 15), 8)) |
4545 S_008F34_PERF_MIP(max_aniso_ratio
? max_aniso_ratio
+ 6 : 0));
4546 sampler
->state
[2] = (S_008F38_LOD_BIAS(S_FIXED(CLAMP(pCreateInfo
->mipLodBias
, -16, 16), 8)) |
4547 S_008F38_XY_MAG_FILTER(radv_tex_filter(pCreateInfo
->magFilter
, max_aniso
)) |
4548 S_008F38_XY_MIN_FILTER(radv_tex_filter(pCreateInfo
->minFilter
, max_aniso
)) |
4549 S_008F38_MIP_FILTER(radv_tex_mipfilter(pCreateInfo
->mipmapMode
)) |
4550 S_008F38_MIP_POINT_PRECLAMP(0) |
4551 S_008F38_DISABLE_LSB_CEIL(device
->physical_device
->rad_info
.chip_class
<= VI
) |
4552 S_008F38_FILTER_PREC_FIX(1) |
4553 S_008F38_ANISO_OVERRIDE(is_vi
));
4554 sampler
->state
[3] = (S_008F3C_BORDER_COLOR_PTR(0) |
4555 S_008F3C_BORDER_COLOR_TYPE(radv_tex_bordercolor(pCreateInfo
->borderColor
)));
4558 VkResult
radv_CreateSampler(
4560 const VkSamplerCreateInfo
* pCreateInfo
,
4561 const VkAllocationCallbacks
* pAllocator
,
4562 VkSampler
* pSampler
)
4564 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4565 struct radv_sampler
*sampler
;
4567 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO
);
4569 sampler
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*sampler
), 8,
4570 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
4572 return vk_error(device
->instance
, VK_ERROR_OUT_OF_HOST_MEMORY
);
4574 radv_init_sampler(device
, sampler
, pCreateInfo
);
4575 *pSampler
= radv_sampler_to_handle(sampler
);
4580 void radv_DestroySampler(
4583 const VkAllocationCallbacks
* pAllocator
)
4585 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4586 RADV_FROM_HANDLE(radv_sampler
, sampler
, _sampler
);
4590 vk_free2(&device
->alloc
, pAllocator
, sampler
);
4593 /* vk_icd.h does not declare this function, so we declare it here to
4594 * suppress Wmissing-prototypes.
4596 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4597 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t *pSupportedVersion
);
4599 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4600 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t *pSupportedVersion
)
4602 /* For the full details on loader interface versioning, see
4603 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
4604 * What follows is a condensed summary, to help you navigate the large and
4605 * confusing official doc.
4607 * - Loader interface v0 is incompatible with later versions. We don't
4610 * - In loader interface v1:
4611 * - The first ICD entrypoint called by the loader is
4612 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
4614 * - The ICD must statically expose no other Vulkan symbol unless it is
4615 * linked with -Bsymbolic.
4616 * - Each dispatchable Vulkan handle created by the ICD must be
4617 * a pointer to a struct whose first member is VK_LOADER_DATA. The
4618 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
4619 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
4620 * vkDestroySurfaceKHR(). The ICD must be capable of working with
4621 * such loader-managed surfaces.
4623 * - Loader interface v2 differs from v1 in:
4624 * - The first ICD entrypoint called by the loader is
4625 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
4626 * statically expose this entrypoint.
4628 * - Loader interface v3 differs from v2 in:
4629 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
4630 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
4631 * because the loader no longer does so.
4633 *pSupportedVersion
= MIN2(*pSupportedVersion
, 3u);
4637 VkResult
radv_GetMemoryFdKHR(VkDevice _device
,
4638 const VkMemoryGetFdInfoKHR
*pGetFdInfo
,
4641 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4642 RADV_FROM_HANDLE(radv_device_memory
, memory
, pGetFdInfo
->memory
);
4644 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
4646 /* At the moment, we support only the below handle types. */
4647 assert(pGetFdInfo
->handleType
==
4648 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
||
4649 pGetFdInfo
->handleType
==
4650 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
);
4652 bool ret
= radv_get_memory_fd(device
, memory
, pFD
);
4654 return vk_error(device
->instance
, VK_ERROR_OUT_OF_DEVICE_MEMORY
);
4658 VkResult
radv_GetMemoryFdPropertiesKHR(VkDevice _device
,
4659 VkExternalMemoryHandleTypeFlagBitsKHR handleType
,
4661 VkMemoryFdPropertiesKHR
*pMemoryFdProperties
)
4663 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4665 switch (handleType
) {
4666 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT
:
4667 pMemoryFdProperties
->memoryTypeBits
= (1 << RADV_MEM_TYPE_COUNT
) - 1;
4671 /* The valid usage section for this function says:
4673 * "handleType must not be one of the handle types defined as
4676 * So opaque handle types fall into the default "unsupported" case.
4678 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
4682 static VkResult
radv_import_opaque_fd(struct radv_device
*device
,
4686 uint32_t syncobj_handle
= 0;
4687 int ret
= device
->ws
->import_syncobj(device
->ws
, fd
, &syncobj_handle
);
4689 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
4692 device
->ws
->destroy_syncobj(device
->ws
, *syncobj
);
4694 *syncobj
= syncobj_handle
;
4700 static VkResult
radv_import_sync_fd(struct radv_device
*device
,
4704 /* If we create a syncobj we do it locally so that if we have an error, we don't
4705 * leave a syncobj in an undetermined state in the fence. */
4706 uint32_t syncobj_handle
= *syncobj
;
4707 if (!syncobj_handle
) {
4708 int ret
= device
->ws
->create_syncobj(device
->ws
, &syncobj_handle
);
4710 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
4715 device
->ws
->signal_syncobj(device
->ws
, syncobj_handle
);
4717 int ret
= device
->ws
->import_syncobj_from_sync_file(device
->ws
, syncobj_handle
, fd
);
4719 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
4722 *syncobj
= syncobj_handle
;
4729 VkResult
radv_ImportSemaphoreFdKHR(VkDevice _device
,
4730 const VkImportSemaphoreFdInfoKHR
*pImportSemaphoreFdInfo
)
4732 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4733 RADV_FROM_HANDLE(radv_semaphore
, sem
, pImportSemaphoreFdInfo
->semaphore
);
4734 uint32_t *syncobj_dst
= NULL
;
4736 if (pImportSemaphoreFdInfo
->flags
& VK_SEMAPHORE_IMPORT_TEMPORARY_BIT_KHR
) {
4737 syncobj_dst
= &sem
->temp_syncobj
;
4739 syncobj_dst
= &sem
->syncobj
;
4742 switch(pImportSemaphoreFdInfo
->handleType
) {
4743 case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
:
4744 return radv_import_opaque_fd(device
, pImportSemaphoreFdInfo
->fd
, syncobj_dst
);
4745 case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT_KHR
:
4746 return radv_import_sync_fd(device
, pImportSemaphoreFdInfo
->fd
, syncobj_dst
);
4748 unreachable("Unhandled semaphore handle type");
4752 VkResult
radv_GetSemaphoreFdKHR(VkDevice _device
,
4753 const VkSemaphoreGetFdInfoKHR
*pGetFdInfo
,
4756 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4757 RADV_FROM_HANDLE(radv_semaphore
, sem
, pGetFdInfo
->semaphore
);
4759 uint32_t syncobj_handle
;
4761 if (sem
->temp_syncobj
)
4762 syncobj_handle
= sem
->temp_syncobj
;
4764 syncobj_handle
= sem
->syncobj
;
4766 switch(pGetFdInfo
->handleType
) {
4767 case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
:
4768 ret
= device
->ws
->export_syncobj(device
->ws
, syncobj_handle
, pFd
);
4770 case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT_KHR
:
4771 ret
= device
->ws
->export_syncobj_to_sync_file(device
->ws
, syncobj_handle
, pFd
);
4773 if (sem
->temp_syncobj
) {
4774 close (sem
->temp_syncobj
);
4775 sem
->temp_syncobj
= 0;
4777 device
->ws
->reset_syncobj(device
->ws
, syncobj_handle
);
4782 unreachable("Unhandled semaphore handle type");
4786 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
4790 void radv_GetPhysicalDeviceExternalSemaphoreProperties(
4791 VkPhysicalDevice physicalDevice
,
4792 const VkPhysicalDeviceExternalSemaphoreInfoKHR
* pExternalSemaphoreInfo
,
4793 VkExternalSemaphorePropertiesKHR
* pExternalSemaphoreProperties
)
4795 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
4797 /* Require has_syncobj_wait_for_submit for the syncobj signal ioctl introduced at virtually the same time */
4798 if (pdevice
->rad_info
.has_syncobj_wait_for_submit
&&
4799 (pExternalSemaphoreInfo
->handleType
== VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
||
4800 pExternalSemaphoreInfo
->handleType
== VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT_KHR
)) {
4801 pExternalSemaphoreProperties
->exportFromImportedHandleTypes
= VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
| VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT_KHR
;
4802 pExternalSemaphoreProperties
->compatibleHandleTypes
= VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
| VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT_KHR
;
4803 pExternalSemaphoreProperties
->externalSemaphoreFeatures
= VK_EXTERNAL_SEMAPHORE_FEATURE_EXPORTABLE_BIT_KHR
|
4804 VK_EXTERNAL_SEMAPHORE_FEATURE_IMPORTABLE_BIT_KHR
;
4805 } else if (pExternalSemaphoreInfo
->handleType
== VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
) {
4806 pExternalSemaphoreProperties
->exportFromImportedHandleTypes
= VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
;
4807 pExternalSemaphoreProperties
->compatibleHandleTypes
= VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
;
4808 pExternalSemaphoreProperties
->externalSemaphoreFeatures
= VK_EXTERNAL_SEMAPHORE_FEATURE_EXPORTABLE_BIT_KHR
|
4809 VK_EXTERNAL_SEMAPHORE_FEATURE_IMPORTABLE_BIT_KHR
;
4811 pExternalSemaphoreProperties
->exportFromImportedHandleTypes
= 0;
4812 pExternalSemaphoreProperties
->compatibleHandleTypes
= 0;
4813 pExternalSemaphoreProperties
->externalSemaphoreFeatures
= 0;
4817 VkResult
radv_ImportFenceFdKHR(VkDevice _device
,
4818 const VkImportFenceFdInfoKHR
*pImportFenceFdInfo
)
4820 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4821 RADV_FROM_HANDLE(radv_fence
, fence
, pImportFenceFdInfo
->fence
);
4822 uint32_t *syncobj_dst
= NULL
;
4825 if (pImportFenceFdInfo
->flags
& VK_FENCE_IMPORT_TEMPORARY_BIT_KHR
) {
4826 syncobj_dst
= &fence
->temp_syncobj
;
4828 syncobj_dst
= &fence
->syncobj
;
4831 switch(pImportFenceFdInfo
->handleType
) {
4832 case VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
:
4833 return radv_import_opaque_fd(device
, pImportFenceFdInfo
->fd
, syncobj_dst
);
4834 case VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT_KHR
:
4835 return radv_import_sync_fd(device
, pImportFenceFdInfo
->fd
, syncobj_dst
);
4837 unreachable("Unhandled fence handle type");
4841 VkResult
radv_GetFenceFdKHR(VkDevice _device
,
4842 const VkFenceGetFdInfoKHR
*pGetFdInfo
,
4845 RADV_FROM_HANDLE(radv_device
, device
, _device
);
4846 RADV_FROM_HANDLE(radv_fence
, fence
, pGetFdInfo
->fence
);
4848 uint32_t syncobj_handle
;
4850 if (fence
->temp_syncobj
)
4851 syncobj_handle
= fence
->temp_syncobj
;
4853 syncobj_handle
= fence
->syncobj
;
4855 switch(pGetFdInfo
->handleType
) {
4856 case VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
:
4857 ret
= device
->ws
->export_syncobj(device
->ws
, syncobj_handle
, pFd
);
4859 case VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT_KHR
:
4860 ret
= device
->ws
->export_syncobj_to_sync_file(device
->ws
, syncobj_handle
, pFd
);
4862 if (fence
->temp_syncobj
) {
4863 close (fence
->temp_syncobj
);
4864 fence
->temp_syncobj
= 0;
4866 device
->ws
->reset_syncobj(device
->ws
, syncobj_handle
);
4871 unreachable("Unhandled fence handle type");
4875 return vk_error(device
->instance
, VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
4879 void radv_GetPhysicalDeviceExternalFenceProperties(
4880 VkPhysicalDevice physicalDevice
,
4881 const VkPhysicalDeviceExternalFenceInfoKHR
* pExternalFenceInfo
,
4882 VkExternalFencePropertiesKHR
* pExternalFenceProperties
)
4884 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
4886 if (pdevice
->rad_info
.has_syncobj_wait_for_submit
&&
4887 (pExternalFenceInfo
->handleType
== VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
||
4888 pExternalFenceInfo
->handleType
== VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT_KHR
)) {
4889 pExternalFenceProperties
->exportFromImportedHandleTypes
= VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
| VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT_KHR
;
4890 pExternalFenceProperties
->compatibleHandleTypes
= VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
| VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT_KHR
;
4891 pExternalFenceProperties
->externalFenceFeatures
= VK_EXTERNAL_FENCE_FEATURE_EXPORTABLE_BIT_KHR
|
4892 VK_EXTERNAL_SEMAPHORE_FEATURE_IMPORTABLE_BIT_KHR
;
4894 pExternalFenceProperties
->exportFromImportedHandleTypes
= 0;
4895 pExternalFenceProperties
->compatibleHandleTypes
= 0;
4896 pExternalFenceProperties
->externalFenceFeatures
= 0;
4901 radv_CreateDebugReportCallbackEXT(VkInstance _instance
,
4902 const VkDebugReportCallbackCreateInfoEXT
* pCreateInfo
,
4903 const VkAllocationCallbacks
* pAllocator
,
4904 VkDebugReportCallbackEXT
* pCallback
)
4906 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
4907 return vk_create_debug_report_callback(&instance
->debug_report_callbacks
,
4908 pCreateInfo
, pAllocator
, &instance
->alloc
,
4913 radv_DestroyDebugReportCallbackEXT(VkInstance _instance
,
4914 VkDebugReportCallbackEXT _callback
,
4915 const VkAllocationCallbacks
* pAllocator
)
4917 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
4918 vk_destroy_debug_report_callback(&instance
->debug_report_callbacks
,
4919 _callback
, pAllocator
, &instance
->alloc
);
4923 radv_DebugReportMessageEXT(VkInstance _instance
,
4924 VkDebugReportFlagsEXT flags
,
4925 VkDebugReportObjectTypeEXT objectType
,
4928 int32_t messageCode
,
4929 const char* pLayerPrefix
,
4930 const char* pMessage
)
4932 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
4933 vk_debug_report(&instance
->debug_report_callbacks
, flags
, objectType
,
4934 object
, location
, messageCode
, pLayerPrefix
, pMessage
);
4938 radv_GetDeviceGroupPeerMemoryFeatures(
4941 uint32_t localDeviceIndex
,
4942 uint32_t remoteDeviceIndex
,
4943 VkPeerMemoryFeatureFlags
* pPeerMemoryFeatures
)
4945 assert(localDeviceIndex
== remoteDeviceIndex
);
4947 *pPeerMemoryFeatures
= VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT
|
4948 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT
|
4949 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT
|
4950 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT
;