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 "amdgpu_id.h"
43 #include "winsys/amdgpu/radv_amdgpu_winsys_public.h"
44 #include "ac_llvm_util.h"
45 #include "vk_format.h"
48 #include "util/debug.h"
51 radv_device_get_cache_uuid(enum radeon_family family
, void *uuid
)
53 uint32_t mesa_timestamp
, llvm_timestamp
;
55 memset(uuid
, 0, VK_UUID_SIZE
);
56 if (!disk_cache_get_function_timestamp(radv_device_get_cache_uuid
, &mesa_timestamp
) ||
57 !disk_cache_get_function_timestamp(LLVMInitializeAMDGPUTargetInfo
, &llvm_timestamp
))
60 memcpy(uuid
, &mesa_timestamp
, 4);
61 memcpy((char*)uuid
+ 4, &llvm_timestamp
, 4);
62 memcpy((char*)uuid
+ 8, &f
, 2);
63 snprintf((char*)uuid
+ 10, VK_UUID_SIZE
- 10, "radv");
68 radv_get_driver_uuid(void *uuid
)
70 ac_compute_driver_uuid(uuid
, VK_UUID_SIZE
);
74 radv_get_device_uuid(struct radeon_info
*info
, void *uuid
)
76 ac_compute_device_uuid(info
, uuid
, VK_UUID_SIZE
);
79 static const VkExtensionProperties instance_extensions
[] = {
81 .extensionName
= VK_KHR_SURFACE_EXTENSION_NAME
,
84 #ifdef VK_USE_PLATFORM_XCB_KHR
86 .extensionName
= VK_KHR_XCB_SURFACE_EXTENSION_NAME
,
90 #ifdef VK_USE_PLATFORM_XLIB_KHR
92 .extensionName
= VK_KHR_XLIB_SURFACE_EXTENSION_NAME
,
96 #ifdef VK_USE_PLATFORM_WAYLAND_KHR
98 .extensionName
= VK_KHR_WAYLAND_SURFACE_EXTENSION_NAME
,
103 .extensionName
= VK_KHR_GET_PHYSICAL_DEVICE_PROPERTIES_2_EXTENSION_NAME
,
107 .extensionName
= VK_KHR_EXTERNAL_MEMORY_CAPABILITIES_EXTENSION_NAME
,
111 .extensionName
= VK_KHR_EXTERNAL_SEMAPHORE_CAPABILITIES_EXTENSION_NAME
,
116 static const VkExtensionProperties common_device_extensions
[] = {
118 .extensionName
= VK_KHR_DESCRIPTOR_UPDATE_TEMPLATE_EXTENSION_NAME
,
122 .extensionName
= VK_KHR_INCREMENTAL_PRESENT_EXTENSION_NAME
,
126 .extensionName
= VK_KHR_MAINTENANCE1_EXTENSION_NAME
,
130 .extensionName
= VK_KHR_PUSH_DESCRIPTOR_EXTENSION_NAME
,
134 .extensionName
= VK_KHR_SAMPLER_MIRROR_CLAMP_TO_EDGE_EXTENSION_NAME
,
138 .extensionName
= VK_KHR_SWAPCHAIN_EXTENSION_NAME
,
142 .extensionName
= VK_AMD_DRAW_INDIRECT_COUNT_EXTENSION_NAME
,
146 .extensionName
= VK_KHR_SHADER_DRAW_PARAMETERS_EXTENSION_NAME
,
150 .extensionName
= VK_KHR_GET_MEMORY_REQUIREMENTS_2_EXTENSION_NAME
,
154 .extensionName
= VK_KHR_DEDICATED_ALLOCATION_EXTENSION_NAME
,
158 .extensionName
= VK_KHR_EXTERNAL_MEMORY_EXTENSION_NAME
,
162 .extensionName
= VK_KHR_EXTERNAL_MEMORY_FD_EXTENSION_NAME
,
166 .extensionName
= VK_KHR_STORAGE_BUFFER_STORAGE_CLASS_EXTENSION_NAME
,
170 .extensionName
= VK_KHR_VARIABLE_POINTERS_EXTENSION_NAME
,
174 .extensionName
= VK_KHR_IMAGE_FORMAT_LIST_EXTENSION_NAME
,
178 static const VkExtensionProperties ext_sema_device_extensions
[] = {
180 .extensionName
= VK_KHR_EXTERNAL_SEMAPHORE_EXTENSION_NAME
,
184 .extensionName
= VK_KHR_EXTERNAL_SEMAPHORE_FD_EXTENSION_NAME
,
188 .extensionName
= VK_KHX_MULTIVIEW_EXTENSION_NAME
,
194 radv_extensions_register(struct radv_instance
*instance
,
195 struct radv_extensions
*extensions
,
196 const VkExtensionProperties
*new_ext
,
200 VkExtensionProperties
*new_ptr
;
202 assert(new_ext
&& num_ext
> 0);
205 return VK_ERROR_INITIALIZATION_FAILED
;
207 new_size
= (extensions
->num_ext
+ num_ext
) * sizeof(VkExtensionProperties
);
208 new_ptr
= vk_realloc(&instance
->alloc
, extensions
->ext_array
,
209 new_size
, 8, VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
211 /* Old array continues to be valid, update nothing */
213 return VK_ERROR_OUT_OF_HOST_MEMORY
;
215 memcpy(&new_ptr
[extensions
->num_ext
], new_ext
,
216 num_ext
* sizeof(VkExtensionProperties
));
217 extensions
->ext_array
= new_ptr
;
218 extensions
->num_ext
+= num_ext
;
224 radv_extensions_finish(struct radv_instance
*instance
,
225 struct radv_extensions
*extensions
)
230 radv_loge("Attemted to free invalid extension struct\n");
232 if (extensions
->ext_array
)
233 vk_free(&instance
->alloc
, extensions
->ext_array
);
237 is_extension_enabled(const VkExtensionProperties
*extensions
,
241 assert(extensions
&& name
);
243 for (uint32_t i
= 0; i
< num_ext
; i
++) {
244 if (strcmp(name
, extensions
[i
].extensionName
) == 0)
252 get_chip_name(enum radeon_family family
)
255 case CHIP_TAHITI
: return "AMD RADV TAHITI";
256 case CHIP_PITCAIRN
: return "AMD RADV PITCAIRN";
257 case CHIP_VERDE
: return "AMD RADV CAPE VERDE";
258 case CHIP_OLAND
: return "AMD RADV OLAND";
259 case CHIP_HAINAN
: return "AMD RADV HAINAN";
260 case CHIP_BONAIRE
: return "AMD RADV BONAIRE";
261 case CHIP_KAVERI
: return "AMD RADV KAVERI";
262 case CHIP_KABINI
: return "AMD RADV KABINI";
263 case CHIP_HAWAII
: return "AMD RADV HAWAII";
264 case CHIP_MULLINS
: return "AMD RADV MULLINS";
265 case CHIP_TONGA
: return "AMD RADV TONGA";
266 case CHIP_ICELAND
: return "AMD RADV ICELAND";
267 case CHIP_CARRIZO
: return "AMD RADV CARRIZO";
268 case CHIP_FIJI
: return "AMD RADV FIJI";
269 case CHIP_POLARIS10
: return "AMD RADV POLARIS10";
270 case CHIP_POLARIS11
: return "AMD RADV POLARIS11";
271 case CHIP_POLARIS12
: return "AMD RADV POLARIS12";
272 case CHIP_STONEY
: return "AMD RADV STONEY";
273 case CHIP_VEGA10
: return "AMD RADV VEGA";
274 case CHIP_RAVEN
: return "AMD RADV RAVEN";
275 default: return "AMD RADV unknown";
280 radv_physical_device_init(struct radv_physical_device
*device
,
281 struct radv_instance
*instance
,
282 drmDevicePtr drm_device
)
284 const char *path
= drm_device
->nodes
[DRM_NODE_RENDER
];
286 drmVersionPtr version
;
289 fd
= open(path
, O_RDWR
| O_CLOEXEC
);
291 return VK_ERROR_INCOMPATIBLE_DRIVER
;
293 version
= drmGetVersion(fd
);
296 return vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER
,
297 "failed to get version %s: %m", path
);
300 if (strcmp(version
->name
, "amdgpu")) {
301 drmFreeVersion(version
);
303 return VK_ERROR_INCOMPATIBLE_DRIVER
;
305 drmFreeVersion(version
);
307 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
308 device
->instance
= instance
;
309 assert(strlen(path
) < ARRAY_SIZE(device
->path
));
310 strncpy(device
->path
, path
, ARRAY_SIZE(device
->path
));
312 device
->ws
= radv_amdgpu_winsys_create(fd
, instance
->debug_flags
,
313 instance
->perftest_flags
);
315 result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
319 device
->local_fd
= fd
;
320 device
->ws
->query_info(device
->ws
, &device
->rad_info
);
321 result
= radv_init_wsi(device
);
322 if (result
!= VK_SUCCESS
) {
323 device
->ws
->destroy(device
->ws
);
327 if (radv_device_get_cache_uuid(device
->rad_info
.family
, device
->cache_uuid
)) {
328 radv_finish_wsi(device
);
329 device
->ws
->destroy(device
->ws
);
330 result
= vk_errorf(VK_ERROR_INITIALIZATION_FAILED
,
331 "cannot generate UUID");
335 result
= radv_extensions_register(instance
,
337 common_device_extensions
,
338 ARRAY_SIZE(common_device_extensions
));
339 if (result
!= VK_SUCCESS
)
342 if (device
->rad_info
.has_syncobj
) {
343 result
= radv_extensions_register(instance
,
345 ext_sema_device_extensions
,
346 ARRAY_SIZE(ext_sema_device_extensions
));
347 if (result
!= VK_SUCCESS
)
351 fprintf(stderr
, "WARNING: radv is not a conformant vulkan implementation, testing use only.\n");
352 device
->name
= get_chip_name(device
->rad_info
.family
);
354 radv_get_driver_uuid(&device
->device_uuid
);
355 radv_get_device_uuid(&device
->rad_info
, &device
->device_uuid
);
357 if (device
->rad_info
.family
== CHIP_STONEY
||
358 device
->rad_info
.chip_class
>= GFX9
) {
359 device
->has_rbplus
= true;
360 device
->rbplus_allowed
= device
->rad_info
.family
== CHIP_STONEY
;
371 radv_physical_device_finish(struct radv_physical_device
*device
)
373 radv_extensions_finish(device
->instance
, &device
->extensions
);
374 radv_finish_wsi(device
);
375 device
->ws
->destroy(device
->ws
);
376 close(device
->local_fd
);
380 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
381 VkSystemAllocationScope allocationScope
)
387 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
388 size_t align
, VkSystemAllocationScope allocationScope
)
390 return realloc(pOriginal
, size
);
394 default_free_func(void *pUserData
, void *pMemory
)
399 static const VkAllocationCallbacks default_alloc
= {
401 .pfnAllocation
= default_alloc_func
,
402 .pfnReallocation
= default_realloc_func
,
403 .pfnFree
= default_free_func
,
406 static const struct debug_control radv_debug_options
[] = {
407 {"nofastclears", RADV_DEBUG_NO_FAST_CLEARS
},
408 {"nodcc", RADV_DEBUG_NO_DCC
},
409 {"shaders", RADV_DEBUG_DUMP_SHADERS
},
410 {"nocache", RADV_DEBUG_NO_CACHE
},
411 {"shaderstats", RADV_DEBUG_DUMP_SHADER_STATS
},
412 {"nohiz", RADV_DEBUG_NO_HIZ
},
413 {"nocompute", RADV_DEBUG_NO_COMPUTE_QUEUE
},
414 {"unsafemath", RADV_DEBUG_UNSAFE_MATH
},
415 {"allbos", RADV_DEBUG_ALL_BOS
},
416 {"noibs", RADV_DEBUG_NO_IBS
},
417 {"spirv", RADV_DEBUG_DUMP_SPIRV
},
418 {"vmfaults", RADV_DEBUG_VM_FAULTS
},
419 {"zerovram", RADV_DEBUG_ZERO_VRAM
},
420 {"syncshaders", RADV_DEBUG_SYNC_SHADERS
},
425 radv_get_debug_option_name(int id
)
427 assert(id
< ARRAY_SIZE(radv_debug_options
) - 1);
428 return radv_debug_options
[id
].string
;
431 static const struct debug_control radv_perftest_options
[] = {
432 {"nobatchchain", RADV_PERFTEST_NO_BATCHCHAIN
},
433 {"sisched", RADV_PERFTEST_SISCHED
},
438 radv_get_perftest_option_name(int id
)
440 assert(id
< ARRAY_SIZE(radv_debug_options
) - 1);
441 return radv_perftest_options
[id
].string
;
444 VkResult
radv_CreateInstance(
445 const VkInstanceCreateInfo
* pCreateInfo
,
446 const VkAllocationCallbacks
* pAllocator
,
447 VkInstance
* pInstance
)
449 struct radv_instance
*instance
;
451 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
453 uint32_t client_version
;
454 if (pCreateInfo
->pApplicationInfo
&&
455 pCreateInfo
->pApplicationInfo
->apiVersion
!= 0) {
456 client_version
= pCreateInfo
->pApplicationInfo
->apiVersion
;
458 client_version
= VK_MAKE_VERSION(1, 0, 0);
461 if (VK_MAKE_VERSION(1, 0, 0) > client_version
||
462 client_version
> VK_MAKE_VERSION(1, 0, 0xfff)) {
463 return vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER
,
464 "Client requested version %d.%d.%d",
465 VK_VERSION_MAJOR(client_version
),
466 VK_VERSION_MINOR(client_version
),
467 VK_VERSION_PATCH(client_version
));
470 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
471 if (!is_extension_enabled(instance_extensions
,
472 ARRAY_SIZE(instance_extensions
),
473 pCreateInfo
->ppEnabledExtensionNames
[i
]))
474 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
477 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
478 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
480 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
482 memset(instance
, 0, sizeof(*instance
));
484 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
487 instance
->alloc
= *pAllocator
;
489 instance
->alloc
= default_alloc
;
491 instance
->apiVersion
= client_version
;
492 instance
->physicalDeviceCount
= -1;
496 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
498 instance
->debug_flags
= parse_debug_string(getenv("RADV_DEBUG"),
501 instance
->perftest_flags
= parse_debug_string(getenv("RADV_PERFTEST"),
502 radv_perftest_options
);
504 *pInstance
= radv_instance_to_handle(instance
);
509 void radv_DestroyInstance(
510 VkInstance _instance
,
511 const VkAllocationCallbacks
* pAllocator
)
513 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
518 for (int i
= 0; i
< instance
->physicalDeviceCount
; ++i
) {
519 radv_physical_device_finish(instance
->physicalDevices
+ i
);
522 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
526 vk_free(&instance
->alloc
, instance
);
530 radv_enumerate_devices(struct radv_instance
*instance
)
532 /* TODO: Check for more devices ? */
533 drmDevicePtr devices
[8];
534 VkResult result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
537 instance
->physicalDeviceCount
= 0;
539 max_devices
= drmGetDevices2(0, devices
, ARRAY_SIZE(devices
));
541 return VK_ERROR_INCOMPATIBLE_DRIVER
;
543 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
544 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
545 devices
[i
]->bustype
== DRM_BUS_PCI
&&
546 devices
[i
]->deviceinfo
.pci
->vendor_id
== 0x1002) {
548 result
= radv_physical_device_init(instance
->physicalDevices
+
549 instance
->physicalDeviceCount
,
552 if (result
== VK_SUCCESS
)
553 ++instance
->physicalDeviceCount
;
554 else if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
558 drmFreeDevices(devices
, max_devices
);
563 VkResult
radv_EnumeratePhysicalDevices(
564 VkInstance _instance
,
565 uint32_t* pPhysicalDeviceCount
,
566 VkPhysicalDevice
* pPhysicalDevices
)
568 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
571 if (instance
->physicalDeviceCount
< 0) {
572 result
= radv_enumerate_devices(instance
);
573 if (result
!= VK_SUCCESS
&&
574 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
578 if (!pPhysicalDevices
) {
579 *pPhysicalDeviceCount
= instance
->physicalDeviceCount
;
581 *pPhysicalDeviceCount
= MIN2(*pPhysicalDeviceCount
, instance
->physicalDeviceCount
);
582 for (unsigned i
= 0; i
< *pPhysicalDeviceCount
; ++i
)
583 pPhysicalDevices
[i
] = radv_physical_device_to_handle(instance
->physicalDevices
+ i
);
586 return *pPhysicalDeviceCount
< instance
->physicalDeviceCount
? VK_INCOMPLETE
590 void radv_GetPhysicalDeviceFeatures(
591 VkPhysicalDevice physicalDevice
,
592 VkPhysicalDeviceFeatures
* pFeatures
)
594 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
595 bool is_gfx9
= pdevice
->rad_info
.chip_class
>= GFX9
;
596 memset(pFeatures
, 0, sizeof(*pFeatures
));
598 *pFeatures
= (VkPhysicalDeviceFeatures
) {
599 .robustBufferAccess
= true,
600 .fullDrawIndexUint32
= true,
601 .imageCubeArray
= true,
602 .independentBlend
= true,
603 .geometryShader
= !is_gfx9
,
604 .tessellationShader
= !is_gfx9
,
605 .sampleRateShading
= true,
606 .dualSrcBlend
= true,
608 .multiDrawIndirect
= true,
609 .drawIndirectFirstInstance
= true,
611 .depthBiasClamp
= true,
612 .fillModeNonSolid
= true,
617 .multiViewport
= true,
618 .samplerAnisotropy
= true,
619 .textureCompressionETC2
= false,
620 .textureCompressionASTC_LDR
= false,
621 .textureCompressionBC
= true,
622 .occlusionQueryPrecise
= true,
623 .pipelineStatisticsQuery
= true,
624 .vertexPipelineStoresAndAtomics
= true,
625 .fragmentStoresAndAtomics
= true,
626 .shaderTessellationAndGeometryPointSize
= true,
627 .shaderImageGatherExtended
= true,
628 .shaderStorageImageExtendedFormats
= true,
629 .shaderStorageImageMultisample
= false,
630 .shaderUniformBufferArrayDynamicIndexing
= true,
631 .shaderSampledImageArrayDynamicIndexing
= true,
632 .shaderStorageBufferArrayDynamicIndexing
= true,
633 .shaderStorageImageArrayDynamicIndexing
= true,
634 .shaderStorageImageReadWithoutFormat
= true,
635 .shaderStorageImageWriteWithoutFormat
= true,
636 .shaderClipDistance
= true,
637 .shaderCullDistance
= true,
638 .shaderFloat64
= true,
640 .shaderInt16
= false,
641 .sparseBinding
= true,
642 .variableMultisampleRate
= true,
643 .inheritedQueries
= true,
647 void radv_GetPhysicalDeviceFeatures2KHR(
648 VkPhysicalDevice physicalDevice
,
649 VkPhysicalDeviceFeatures2KHR
*pFeatures
)
651 vk_foreach_struct(ext
, pFeatures
->pNext
) {
652 switch (ext
->sType
) {
653 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTER_FEATURES_KHR
: {
654 VkPhysicalDeviceVariablePointerFeaturesKHR
*features
= (void *)ext
;
655 features
->variablePointersStorageBuffer
= true;
656 features
->variablePointers
= false;
659 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES_KHX
: {
660 VkPhysicalDeviceMultiviewFeaturesKHX
*features
= (VkPhysicalDeviceMultiviewFeaturesKHX
*)ext
;
661 features
->multiview
= true;
662 features
->multiviewGeometryShader
= true;
663 features
->multiviewTessellationShader
= true;
670 return radv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
673 void radv_GetPhysicalDeviceProperties(
674 VkPhysicalDevice physicalDevice
,
675 VkPhysicalDeviceProperties
* pProperties
)
677 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
678 VkSampleCountFlags sample_counts
= 0xf;
680 /* make sure that the entire descriptor set is addressable with a signed
681 * 32-bit int. So the sum of all limits scaled by descriptor size has to
682 * be at most 2 GiB. the combined image & samples object count as one of
683 * both. This limit is for the pipeline layout, not for the set layout, but
684 * there is no set limit, so we just set a pipeline limit. I don't think
685 * any app is going to hit this soon. */
686 size_t max_descriptor_set_size
= ((1ull << 31) - 16 * MAX_DYNAMIC_BUFFERS
) /
687 (32 /* uniform buffer, 32 due to potential space wasted on alignement */ +
688 32 /* storage buffer, 32 due to potential space wasted on alignement */ +
689 32 /* sampler, largest when combined with image */ +
690 64 /* sampled image */ +
691 64 /* storage image */);
693 VkPhysicalDeviceLimits limits
= {
694 .maxImageDimension1D
= (1 << 14),
695 .maxImageDimension2D
= (1 << 14),
696 .maxImageDimension3D
= (1 << 11),
697 .maxImageDimensionCube
= (1 << 14),
698 .maxImageArrayLayers
= (1 << 11),
699 .maxTexelBufferElements
= 128 * 1024 * 1024,
700 .maxUniformBufferRange
= UINT32_MAX
,
701 .maxStorageBufferRange
= UINT32_MAX
,
702 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
703 .maxMemoryAllocationCount
= UINT32_MAX
,
704 .maxSamplerAllocationCount
= 64 * 1024,
705 .bufferImageGranularity
= 64, /* A cache line */
706 .sparseAddressSpaceSize
= 0xffffffffu
, /* buffer max size */
707 .maxBoundDescriptorSets
= MAX_SETS
,
708 .maxPerStageDescriptorSamplers
= max_descriptor_set_size
,
709 .maxPerStageDescriptorUniformBuffers
= max_descriptor_set_size
,
710 .maxPerStageDescriptorStorageBuffers
= max_descriptor_set_size
,
711 .maxPerStageDescriptorSampledImages
= max_descriptor_set_size
,
712 .maxPerStageDescriptorStorageImages
= max_descriptor_set_size
,
713 .maxPerStageDescriptorInputAttachments
= max_descriptor_set_size
,
714 .maxPerStageResources
= max_descriptor_set_size
,
715 .maxDescriptorSetSamplers
= max_descriptor_set_size
,
716 .maxDescriptorSetUniformBuffers
= max_descriptor_set_size
,
717 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
718 .maxDescriptorSetStorageBuffers
= max_descriptor_set_size
,
719 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
720 .maxDescriptorSetSampledImages
= max_descriptor_set_size
,
721 .maxDescriptorSetStorageImages
= max_descriptor_set_size
,
722 .maxDescriptorSetInputAttachments
= max_descriptor_set_size
,
723 .maxVertexInputAttributes
= 32,
724 .maxVertexInputBindings
= 32,
725 .maxVertexInputAttributeOffset
= 2047,
726 .maxVertexInputBindingStride
= 2048,
727 .maxVertexOutputComponents
= 128,
728 .maxTessellationGenerationLevel
= 64,
729 .maxTessellationPatchSize
= 32,
730 .maxTessellationControlPerVertexInputComponents
= 128,
731 .maxTessellationControlPerVertexOutputComponents
= 128,
732 .maxTessellationControlPerPatchOutputComponents
= 120,
733 .maxTessellationControlTotalOutputComponents
= 4096,
734 .maxTessellationEvaluationInputComponents
= 128,
735 .maxTessellationEvaluationOutputComponents
= 128,
736 .maxGeometryShaderInvocations
= 127,
737 .maxGeometryInputComponents
= 64,
738 .maxGeometryOutputComponents
= 128,
739 .maxGeometryOutputVertices
= 256,
740 .maxGeometryTotalOutputComponents
= 1024,
741 .maxFragmentInputComponents
= 128,
742 .maxFragmentOutputAttachments
= 8,
743 .maxFragmentDualSrcAttachments
= 1,
744 .maxFragmentCombinedOutputResources
= 8,
745 .maxComputeSharedMemorySize
= 32768,
746 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
747 .maxComputeWorkGroupInvocations
= 2048,
748 .maxComputeWorkGroupSize
= {
753 .subPixelPrecisionBits
= 4 /* FIXME */,
754 .subTexelPrecisionBits
= 4 /* FIXME */,
755 .mipmapPrecisionBits
= 4 /* FIXME */,
756 .maxDrawIndexedIndexValue
= UINT32_MAX
,
757 .maxDrawIndirectCount
= UINT32_MAX
,
758 .maxSamplerLodBias
= 16,
759 .maxSamplerAnisotropy
= 16,
760 .maxViewports
= MAX_VIEWPORTS
,
761 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
762 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
763 .viewportSubPixelBits
= 13, /* We take a float? */
764 .minMemoryMapAlignment
= 4096, /* A page */
765 .minTexelBufferOffsetAlignment
= 1,
766 .minUniformBufferOffsetAlignment
= 4,
767 .minStorageBufferOffsetAlignment
= 4,
768 .minTexelOffset
= -32,
769 .maxTexelOffset
= 31,
770 .minTexelGatherOffset
= -32,
771 .maxTexelGatherOffset
= 31,
772 .minInterpolationOffset
= -2,
773 .maxInterpolationOffset
= 2,
774 .subPixelInterpolationOffsetBits
= 8,
775 .maxFramebufferWidth
= (1 << 14),
776 .maxFramebufferHeight
= (1 << 14),
777 .maxFramebufferLayers
= (1 << 10),
778 .framebufferColorSampleCounts
= sample_counts
,
779 .framebufferDepthSampleCounts
= sample_counts
,
780 .framebufferStencilSampleCounts
= sample_counts
,
781 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
782 .maxColorAttachments
= MAX_RTS
,
783 .sampledImageColorSampleCounts
= sample_counts
,
784 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
785 .sampledImageDepthSampleCounts
= sample_counts
,
786 .sampledImageStencilSampleCounts
= sample_counts
,
787 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
788 .maxSampleMaskWords
= 1,
789 .timestampComputeAndGraphics
= true,
790 .timestampPeriod
= 1000000.0 / pdevice
->rad_info
.clock_crystal_freq
,
791 .maxClipDistances
= 8,
792 .maxCullDistances
= 8,
793 .maxCombinedClipAndCullDistances
= 8,
794 .discreteQueuePriorities
= 1,
795 .pointSizeRange
= { 0.125, 255.875 },
796 .lineWidthRange
= { 0.0, 7.9921875 },
797 .pointSizeGranularity
= (1.0 / 8.0),
798 .lineWidthGranularity
= (1.0 / 128.0),
799 .strictLines
= false, /* FINISHME */
800 .standardSampleLocations
= true,
801 .optimalBufferCopyOffsetAlignment
= 128,
802 .optimalBufferCopyRowPitchAlignment
= 128,
803 .nonCoherentAtomSize
= 64,
806 *pProperties
= (VkPhysicalDeviceProperties
) {
807 .apiVersion
= VK_MAKE_VERSION(1, 0, 42),
808 .driverVersion
= vk_get_driver_version(),
810 .deviceID
= pdevice
->rad_info
.pci_id
,
811 .deviceType
= pdevice
->rad_info
.has_dedicated_vram
? VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU
: VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
813 .sparseProperties
= {0},
816 strcpy(pProperties
->deviceName
, pdevice
->name
);
817 memcpy(pProperties
->pipelineCacheUUID
, pdevice
->cache_uuid
, VK_UUID_SIZE
);
820 void radv_GetPhysicalDeviceProperties2KHR(
821 VkPhysicalDevice physicalDevice
,
822 VkPhysicalDeviceProperties2KHR
*pProperties
)
824 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
825 radv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
827 vk_foreach_struct(ext
, pProperties
->pNext
) {
828 switch (ext
->sType
) {
829 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
830 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
831 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
832 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
835 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES_KHR
: {
836 VkPhysicalDeviceIDPropertiesKHR
*properties
= (VkPhysicalDeviceIDPropertiesKHR
*)ext
;
837 memcpy(properties
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
838 memcpy(properties
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
839 properties
->deviceLUIDValid
= false;
842 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES_KHX
: {
843 VkPhysicalDeviceMultiviewPropertiesKHX
*properties
= (VkPhysicalDeviceMultiviewPropertiesKHX
*)ext
;
844 properties
->maxMultiviewViewCount
= MAX_VIEWS
;
845 properties
->maxMultiviewInstanceIndex
= INT_MAX
;
854 static void radv_get_physical_device_queue_family_properties(
855 struct radv_physical_device
* pdevice
,
857 VkQueueFamilyProperties
** pQueueFamilyProperties
)
859 int num_queue_families
= 1;
861 if (pdevice
->rad_info
.num_compute_rings
> 0 &&
862 pdevice
->rad_info
.chip_class
>= CIK
&&
863 !(pdevice
->instance
->debug_flags
& RADV_DEBUG_NO_COMPUTE_QUEUE
))
864 num_queue_families
++;
866 if (pQueueFamilyProperties
== NULL
) {
867 *pCount
= num_queue_families
;
876 *pQueueFamilyProperties
[idx
] = (VkQueueFamilyProperties
) {
877 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
878 VK_QUEUE_COMPUTE_BIT
|
879 VK_QUEUE_TRANSFER_BIT
|
880 VK_QUEUE_SPARSE_BINDING_BIT
,
882 .timestampValidBits
= 64,
883 .minImageTransferGranularity
= (VkExtent3D
) { 1, 1, 1 },
888 if (pdevice
->rad_info
.num_compute_rings
> 0 &&
889 pdevice
->rad_info
.chip_class
>= CIK
&&
890 !(pdevice
->instance
->debug_flags
& RADV_DEBUG_NO_COMPUTE_QUEUE
)) {
892 *pQueueFamilyProperties
[idx
] = (VkQueueFamilyProperties
) {
893 .queueFlags
= VK_QUEUE_COMPUTE_BIT
|
894 VK_QUEUE_TRANSFER_BIT
|
895 VK_QUEUE_SPARSE_BINDING_BIT
,
896 .queueCount
= pdevice
->rad_info
.num_compute_rings
,
897 .timestampValidBits
= 64,
898 .minImageTransferGranularity
= (VkExtent3D
) { 1, 1, 1 },
906 void radv_GetPhysicalDeviceQueueFamilyProperties(
907 VkPhysicalDevice physicalDevice
,
909 VkQueueFamilyProperties
* pQueueFamilyProperties
)
911 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
912 if (!pQueueFamilyProperties
) {
913 return radv_get_physical_device_queue_family_properties(pdevice
, pCount
, NULL
);
916 VkQueueFamilyProperties
*properties
[] = {
917 pQueueFamilyProperties
+ 0,
918 pQueueFamilyProperties
+ 1,
919 pQueueFamilyProperties
+ 2,
921 radv_get_physical_device_queue_family_properties(pdevice
, pCount
, properties
);
922 assert(*pCount
<= 3);
925 void radv_GetPhysicalDeviceQueueFamilyProperties2KHR(
926 VkPhysicalDevice physicalDevice
,
928 VkQueueFamilyProperties2KHR
*pQueueFamilyProperties
)
930 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
931 if (!pQueueFamilyProperties
) {
932 return radv_get_physical_device_queue_family_properties(pdevice
, pCount
, NULL
);
935 VkQueueFamilyProperties
*properties
[] = {
936 &pQueueFamilyProperties
[0].queueFamilyProperties
,
937 &pQueueFamilyProperties
[1].queueFamilyProperties
,
938 &pQueueFamilyProperties
[2].queueFamilyProperties
,
940 radv_get_physical_device_queue_family_properties(pdevice
, pCount
, properties
);
941 assert(*pCount
<= 3);
944 void radv_GetPhysicalDeviceMemoryProperties(
945 VkPhysicalDevice physicalDevice
,
946 VkPhysicalDeviceMemoryProperties
*pMemoryProperties
)
948 RADV_FROM_HANDLE(radv_physical_device
, physical_device
, physicalDevice
);
950 STATIC_ASSERT(RADV_MEM_TYPE_COUNT
<= VK_MAX_MEMORY_TYPES
);
952 pMemoryProperties
->memoryTypeCount
= RADV_MEM_TYPE_COUNT
;
953 pMemoryProperties
->memoryTypes
[RADV_MEM_TYPE_VRAM
] = (VkMemoryType
) {
954 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
,
955 .heapIndex
= RADV_MEM_HEAP_VRAM
,
957 pMemoryProperties
->memoryTypes
[RADV_MEM_TYPE_GTT_WRITE_COMBINE
] = (VkMemoryType
) {
958 .propertyFlags
= VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
959 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
960 .heapIndex
= RADV_MEM_HEAP_GTT
,
962 pMemoryProperties
->memoryTypes
[RADV_MEM_TYPE_VRAM_CPU_ACCESS
] = (VkMemoryType
) {
963 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
964 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
965 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
966 .heapIndex
= RADV_MEM_HEAP_VRAM_CPU_ACCESS
,
968 pMemoryProperties
->memoryTypes
[RADV_MEM_TYPE_GTT_CACHED
] = (VkMemoryType
) {
969 .propertyFlags
= VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
970 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
971 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
972 .heapIndex
= RADV_MEM_HEAP_GTT
,
975 STATIC_ASSERT(RADV_MEM_HEAP_COUNT
<= VK_MAX_MEMORY_HEAPS
);
976 uint64_t visible_vram_size
= MIN2(physical_device
->rad_info
.vram_size
,
977 physical_device
->rad_info
.vram_vis_size
);
979 pMemoryProperties
->memoryHeapCount
= RADV_MEM_HEAP_COUNT
;
980 pMemoryProperties
->memoryHeaps
[RADV_MEM_HEAP_VRAM
] = (VkMemoryHeap
) {
981 .size
= physical_device
->rad_info
.vram_size
-
983 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
985 pMemoryProperties
->memoryHeaps
[RADV_MEM_HEAP_VRAM_CPU_ACCESS
] = (VkMemoryHeap
) {
986 .size
= visible_vram_size
,
987 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
989 pMemoryProperties
->memoryHeaps
[RADV_MEM_HEAP_GTT
] = (VkMemoryHeap
) {
990 .size
= physical_device
->rad_info
.gart_size
,
995 void radv_GetPhysicalDeviceMemoryProperties2KHR(
996 VkPhysicalDevice physicalDevice
,
997 VkPhysicalDeviceMemoryProperties2KHR
*pMemoryProperties
)
999 return radv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
1000 &pMemoryProperties
->memoryProperties
);
1004 radv_queue_init(struct radv_device
*device
, struct radv_queue
*queue
,
1005 int queue_family_index
, int idx
)
1007 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1008 queue
->device
= device
;
1009 queue
->queue_family_index
= queue_family_index
;
1010 queue
->queue_idx
= idx
;
1012 queue
->hw_ctx
= device
->ws
->ctx_create(device
->ws
);
1014 return VK_ERROR_OUT_OF_HOST_MEMORY
;
1020 radv_queue_finish(struct radv_queue
*queue
)
1023 queue
->device
->ws
->ctx_destroy(queue
->hw_ctx
);
1025 if (queue
->initial_full_flush_preamble_cs
)
1026 queue
->device
->ws
->cs_destroy(queue
->initial_full_flush_preamble_cs
);
1027 if (queue
->initial_preamble_cs
)
1028 queue
->device
->ws
->cs_destroy(queue
->initial_preamble_cs
);
1029 if (queue
->continue_preamble_cs
)
1030 queue
->device
->ws
->cs_destroy(queue
->continue_preamble_cs
);
1031 if (queue
->descriptor_bo
)
1032 queue
->device
->ws
->buffer_destroy(queue
->descriptor_bo
);
1033 if (queue
->scratch_bo
)
1034 queue
->device
->ws
->buffer_destroy(queue
->scratch_bo
);
1035 if (queue
->esgs_ring_bo
)
1036 queue
->device
->ws
->buffer_destroy(queue
->esgs_ring_bo
);
1037 if (queue
->gsvs_ring_bo
)
1038 queue
->device
->ws
->buffer_destroy(queue
->gsvs_ring_bo
);
1039 if (queue
->tess_factor_ring_bo
)
1040 queue
->device
->ws
->buffer_destroy(queue
->tess_factor_ring_bo
);
1041 if (queue
->tess_offchip_ring_bo
)
1042 queue
->device
->ws
->buffer_destroy(queue
->tess_offchip_ring_bo
);
1043 if (queue
->compute_scratch_bo
)
1044 queue
->device
->ws
->buffer_destroy(queue
->compute_scratch_bo
);
1048 radv_device_init_gs_info(struct radv_device
*device
)
1050 switch (device
->physical_device
->rad_info
.family
) {
1059 device
->gs_table_depth
= 16;
1068 case CHIP_POLARIS10
:
1069 case CHIP_POLARIS11
:
1070 case CHIP_POLARIS12
:
1073 device
->gs_table_depth
= 32;
1076 unreachable("unknown GPU");
1080 VkResult
radv_CreateDevice(
1081 VkPhysicalDevice physicalDevice
,
1082 const VkDeviceCreateInfo
* pCreateInfo
,
1083 const VkAllocationCallbacks
* pAllocator
,
1086 RADV_FROM_HANDLE(radv_physical_device
, physical_device
, physicalDevice
);
1088 struct radv_device
*device
;
1090 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
1091 if (!is_extension_enabled(physical_device
->extensions
.ext_array
,
1092 physical_device
->extensions
.num_ext
,
1093 pCreateInfo
->ppEnabledExtensionNames
[i
]))
1094 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
1097 /* Check enabled features */
1098 if (pCreateInfo
->pEnabledFeatures
) {
1099 VkPhysicalDeviceFeatures supported_features
;
1100 radv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
1101 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
1102 VkBool32
*enabled_feature
= (VkBool32
*)pCreateInfo
->pEnabledFeatures
;
1103 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
1104 for (uint32_t i
= 0; i
< num_features
; i
++) {
1105 if (enabled_feature
[i
] && !supported_feature
[i
])
1106 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
1110 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
1112 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
1114 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1116 memset(device
, 0, sizeof(*device
));
1118 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1119 device
->instance
= physical_device
->instance
;
1120 device
->physical_device
= physical_device
;
1122 device
->debug_flags
= device
->instance
->debug_flags
;
1124 device
->ws
= physical_device
->ws
;
1126 device
->alloc
= *pAllocator
;
1128 device
->alloc
= physical_device
->instance
->alloc
;
1130 mtx_init(&device
->shader_slab_mutex
, mtx_plain
);
1131 list_inithead(&device
->shader_slabs
);
1133 for (unsigned i
= 0; i
< pCreateInfo
->queueCreateInfoCount
; i
++) {
1134 const VkDeviceQueueCreateInfo
*queue_create
= &pCreateInfo
->pQueueCreateInfos
[i
];
1135 uint32_t qfi
= queue_create
->queueFamilyIndex
;
1137 device
->queues
[qfi
] = vk_alloc(&device
->alloc
,
1138 queue_create
->queueCount
* sizeof(struct radv_queue
), 8, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
1139 if (!device
->queues
[qfi
]) {
1140 result
= VK_ERROR_OUT_OF_HOST_MEMORY
;
1144 memset(device
->queues
[qfi
], 0, queue_create
->queueCount
* sizeof(struct radv_queue
));
1146 device
->queue_count
[qfi
] = queue_create
->queueCount
;
1148 for (unsigned q
= 0; q
< queue_create
->queueCount
; q
++) {
1149 result
= radv_queue_init(device
, &device
->queues
[qfi
][q
], qfi
, q
);
1150 if (result
!= VK_SUCCESS
)
1155 #if HAVE_LLVM < 0x0400
1156 device
->llvm_supports_spill
= false;
1158 device
->llvm_supports_spill
= true;
1161 /* The maximum number of scratch waves. Scratch space isn't divided
1162 * evenly between CUs. The number is only a function of the number of CUs.
1163 * We can decrease the constant to decrease the scratch buffer size.
1165 * sctx->scratch_waves must be >= the maximum posible size of
1166 * 1 threadgroup, so that the hw doesn't hang from being unable
1169 * The recommended value is 4 per CU at most. Higher numbers don't
1170 * bring much benefit, but they still occupy chip resources (think
1171 * async compute). I've seen ~2% performance difference between 4 and 32.
1173 uint32_t max_threads_per_block
= 2048;
1174 device
->scratch_waves
= MAX2(32 * physical_device
->rad_info
.num_good_compute_units
,
1175 max_threads_per_block
/ 64);
1177 radv_device_init_gs_info(device
);
1179 device
->tess_offchip_block_dw_size
=
1180 device
->physical_device
->rad_info
.family
== CHIP_HAWAII
? 4096 : 8192;
1181 device
->has_distributed_tess
=
1182 device
->physical_device
->rad_info
.chip_class
>= VI
&&
1183 device
->physical_device
->rad_info
.max_se
>= 2;
1185 result
= radv_device_init_meta(device
);
1186 if (result
!= VK_SUCCESS
)
1189 radv_device_init_msaa(device
);
1191 for (int family
= 0; family
< RADV_MAX_QUEUE_FAMILIES
; ++family
) {
1192 device
->empty_cs
[family
] = device
->ws
->cs_create(device
->ws
, family
);
1194 case RADV_QUEUE_GENERAL
:
1195 radeon_emit(device
->empty_cs
[family
], PKT3(PKT3_CONTEXT_CONTROL
, 1, 0));
1196 radeon_emit(device
->empty_cs
[family
], CONTEXT_CONTROL_LOAD_ENABLE(1));
1197 radeon_emit(device
->empty_cs
[family
], CONTEXT_CONTROL_SHADOW_ENABLE(1));
1199 case RADV_QUEUE_COMPUTE
:
1200 radeon_emit(device
->empty_cs
[family
], PKT3(PKT3_NOP
, 0, 0));
1201 radeon_emit(device
->empty_cs
[family
], 0);
1204 device
->ws
->cs_finalize(device
->empty_cs
[family
]);
1207 if (getenv("RADV_TRACE_FILE")) {
1208 if (!radv_init_trace(device
))
1212 if (device
->physical_device
->rad_info
.chip_class
>= CIK
)
1213 cik_create_gfx_config(device
);
1215 VkPipelineCacheCreateInfo ci
;
1216 ci
.sType
= VK_STRUCTURE_TYPE_PIPELINE_CACHE_CREATE_INFO
;
1219 ci
.pInitialData
= NULL
;
1220 ci
.initialDataSize
= 0;
1222 result
= radv_CreatePipelineCache(radv_device_to_handle(device
),
1224 if (result
!= VK_SUCCESS
)
1227 device
->mem_cache
= radv_pipeline_cache_from_handle(pc
);
1229 *pDevice
= radv_device_to_handle(device
);
1233 if (device
->trace_bo
)
1234 device
->ws
->buffer_destroy(device
->trace_bo
);
1236 if (device
->gfx_init
)
1237 device
->ws
->buffer_destroy(device
->gfx_init
);
1239 for (unsigned i
= 0; i
< RADV_MAX_QUEUE_FAMILIES
; i
++) {
1240 for (unsigned q
= 0; q
< device
->queue_count
[i
]; q
++)
1241 radv_queue_finish(&device
->queues
[i
][q
]);
1242 if (device
->queue_count
[i
])
1243 vk_free(&device
->alloc
, device
->queues
[i
]);
1246 vk_free(&device
->alloc
, device
);
1250 void radv_DestroyDevice(
1252 const VkAllocationCallbacks
* pAllocator
)
1254 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1259 if (device
->trace_bo
)
1260 device
->ws
->buffer_destroy(device
->trace_bo
);
1262 if (device
->gfx_init
)
1263 device
->ws
->buffer_destroy(device
->gfx_init
);
1265 for (unsigned i
= 0; i
< RADV_MAX_QUEUE_FAMILIES
; i
++) {
1266 for (unsigned q
= 0; q
< device
->queue_count
[i
]; q
++)
1267 radv_queue_finish(&device
->queues
[i
][q
]);
1268 if (device
->queue_count
[i
])
1269 vk_free(&device
->alloc
, device
->queues
[i
]);
1270 if (device
->empty_cs
[i
])
1271 device
->ws
->cs_destroy(device
->empty_cs
[i
]);
1273 radv_device_finish_meta(device
);
1275 VkPipelineCache pc
= radv_pipeline_cache_to_handle(device
->mem_cache
);
1276 radv_DestroyPipelineCache(radv_device_to_handle(device
), pc
, NULL
);
1278 radv_destroy_shader_slabs(device
);
1280 vk_free(&device
->alloc
, device
);
1283 VkResult
radv_EnumerateInstanceExtensionProperties(
1284 const char* pLayerName
,
1285 uint32_t* pPropertyCount
,
1286 VkExtensionProperties
* pProperties
)
1288 if (pProperties
== NULL
) {
1289 *pPropertyCount
= ARRAY_SIZE(instance_extensions
);
1293 *pPropertyCount
= MIN2(*pPropertyCount
, ARRAY_SIZE(instance_extensions
));
1294 typed_memcpy(pProperties
, instance_extensions
, *pPropertyCount
);
1296 if (*pPropertyCount
< ARRAY_SIZE(instance_extensions
))
1297 return VK_INCOMPLETE
;
1302 VkResult
radv_EnumerateDeviceExtensionProperties(
1303 VkPhysicalDevice physicalDevice
,
1304 const char* pLayerName
,
1305 uint32_t* pPropertyCount
,
1306 VkExtensionProperties
* pProperties
)
1308 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
1310 if (pProperties
== NULL
) {
1311 *pPropertyCount
= pdevice
->extensions
.num_ext
;
1315 *pPropertyCount
= MIN2(*pPropertyCount
, pdevice
->extensions
.num_ext
);
1316 typed_memcpy(pProperties
, pdevice
->extensions
.ext_array
, *pPropertyCount
);
1318 if (*pPropertyCount
< pdevice
->extensions
.num_ext
)
1319 return VK_INCOMPLETE
;
1324 VkResult
radv_EnumerateInstanceLayerProperties(
1325 uint32_t* pPropertyCount
,
1326 VkLayerProperties
* pProperties
)
1328 if (pProperties
== NULL
) {
1329 *pPropertyCount
= 0;
1333 /* None supported at this time */
1334 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1337 VkResult
radv_EnumerateDeviceLayerProperties(
1338 VkPhysicalDevice physicalDevice
,
1339 uint32_t* pPropertyCount
,
1340 VkLayerProperties
* pProperties
)
1342 if (pProperties
== NULL
) {
1343 *pPropertyCount
= 0;
1347 /* None supported at this time */
1348 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1351 void radv_GetDeviceQueue(
1353 uint32_t queueFamilyIndex
,
1354 uint32_t queueIndex
,
1357 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1359 *pQueue
= radv_queue_to_handle(&device
->queues
[queueFamilyIndex
][queueIndex
]);
1363 fill_geom_tess_rings(struct radv_queue
*queue
,
1365 bool add_sample_positions
,
1366 uint32_t esgs_ring_size
,
1367 struct radeon_winsys_bo
*esgs_ring_bo
,
1368 uint32_t gsvs_ring_size
,
1369 struct radeon_winsys_bo
*gsvs_ring_bo
,
1370 uint32_t tess_factor_ring_size
,
1371 struct radeon_winsys_bo
*tess_factor_ring_bo
,
1372 uint32_t tess_offchip_ring_size
,
1373 struct radeon_winsys_bo
*tess_offchip_ring_bo
)
1375 uint64_t esgs_va
= 0, gsvs_va
= 0;
1376 uint64_t tess_factor_va
= 0, tess_offchip_va
= 0;
1377 uint32_t *desc
= &map
[4];
1380 esgs_va
= queue
->device
->ws
->buffer_get_va(esgs_ring_bo
);
1382 gsvs_va
= queue
->device
->ws
->buffer_get_va(gsvs_ring_bo
);
1383 if (tess_factor_ring_bo
)
1384 tess_factor_va
= queue
->device
->ws
->buffer_get_va(tess_factor_ring_bo
);
1385 if (tess_offchip_ring_bo
)
1386 tess_offchip_va
= queue
->device
->ws
->buffer_get_va(tess_offchip_ring_bo
);
1388 /* stride 0, num records - size, add tid, swizzle, elsize4,
1391 desc
[1] = S_008F04_BASE_ADDRESS_HI(esgs_va
>> 32) |
1392 S_008F04_STRIDE(0) |
1393 S_008F04_SWIZZLE_ENABLE(true);
1394 desc
[2] = esgs_ring_size
;
1395 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1396 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1397 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1398 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1399 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1400 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1401 S_008F0C_ELEMENT_SIZE(1) |
1402 S_008F0C_INDEX_STRIDE(3) |
1403 S_008F0C_ADD_TID_ENABLE(true);
1406 /* GS entry for ES->GS ring */
1407 /* stride 0, num records - size, elsize0,
1410 desc
[1] = S_008F04_BASE_ADDRESS_HI(esgs_va
>> 32)|
1411 S_008F04_STRIDE(0) |
1412 S_008F04_SWIZZLE_ENABLE(false);
1413 desc
[2] = esgs_ring_size
;
1414 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1415 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1416 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1417 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1418 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1419 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1420 S_008F0C_ELEMENT_SIZE(0) |
1421 S_008F0C_INDEX_STRIDE(0) |
1422 S_008F0C_ADD_TID_ENABLE(false);
1425 /* VS entry for GS->VS ring */
1426 /* stride 0, num records - size, elsize0,
1429 desc
[1] = S_008F04_BASE_ADDRESS_HI(gsvs_va
>> 32)|
1430 S_008F04_STRIDE(0) |
1431 S_008F04_SWIZZLE_ENABLE(false);
1432 desc
[2] = gsvs_ring_size
;
1433 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1434 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1435 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1436 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1437 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1438 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1439 S_008F0C_ELEMENT_SIZE(0) |
1440 S_008F0C_INDEX_STRIDE(0) |
1441 S_008F0C_ADD_TID_ENABLE(false);
1444 /* stride gsvs_itemsize, num records 64
1445 elsize 4, index stride 16 */
1446 /* shader will patch stride and desc[2] */
1448 desc
[1] = S_008F04_BASE_ADDRESS_HI(gsvs_va
>> 32)|
1449 S_008F04_STRIDE(0) |
1450 S_008F04_SWIZZLE_ENABLE(true);
1452 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1453 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1454 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1455 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1456 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1457 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1458 S_008F0C_ELEMENT_SIZE(1) |
1459 S_008F0C_INDEX_STRIDE(1) |
1460 S_008F0C_ADD_TID_ENABLE(true);
1463 desc
[0] = tess_factor_va
;
1464 desc
[1] = S_008F04_BASE_ADDRESS_HI(tess_factor_va
>> 32) |
1465 S_008F04_STRIDE(0) |
1466 S_008F04_SWIZZLE_ENABLE(false);
1467 desc
[2] = tess_factor_ring_size
;
1468 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1469 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1470 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1471 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1472 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1473 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1474 S_008F0C_ELEMENT_SIZE(0) |
1475 S_008F0C_INDEX_STRIDE(0) |
1476 S_008F0C_ADD_TID_ENABLE(false);
1479 desc
[0] = tess_offchip_va
;
1480 desc
[1] = S_008F04_BASE_ADDRESS_HI(tess_offchip_va
>> 32) |
1481 S_008F04_STRIDE(0) |
1482 S_008F04_SWIZZLE_ENABLE(false);
1483 desc
[2] = tess_offchip_ring_size
;
1484 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1485 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1486 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1487 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1488 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1489 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1490 S_008F0C_ELEMENT_SIZE(0) |
1491 S_008F0C_INDEX_STRIDE(0) |
1492 S_008F0C_ADD_TID_ENABLE(false);
1495 /* add sample positions after all rings */
1496 memcpy(desc
, queue
->device
->sample_locations_1x
, 8);
1498 memcpy(desc
, queue
->device
->sample_locations_2x
, 16);
1500 memcpy(desc
, queue
->device
->sample_locations_4x
, 32);
1502 memcpy(desc
, queue
->device
->sample_locations_8x
, 64);
1504 memcpy(desc
, queue
->device
->sample_locations_16x
, 128);
1508 radv_get_hs_offchip_param(struct radv_device
*device
, uint32_t *max_offchip_buffers_p
)
1510 bool double_offchip_buffers
= device
->physical_device
->rad_info
.chip_class
>= CIK
&&
1511 device
->physical_device
->rad_info
.family
!= CHIP_CARRIZO
&&
1512 device
->physical_device
->rad_info
.family
!= CHIP_STONEY
;
1513 unsigned max_offchip_buffers_per_se
= double_offchip_buffers
? 128 : 64;
1514 unsigned max_offchip_buffers
= max_offchip_buffers_per_se
*
1515 device
->physical_device
->rad_info
.max_se
;
1516 unsigned offchip_granularity
;
1517 unsigned hs_offchip_param
;
1518 switch (device
->tess_offchip_block_dw_size
) {
1523 offchip_granularity
= V_03093C_X_8K_DWORDS
;
1526 offchip_granularity
= V_03093C_X_4K_DWORDS
;
1530 switch (device
->physical_device
->rad_info
.chip_class
) {
1532 max_offchip_buffers
= MIN2(max_offchip_buffers
, 126);
1538 max_offchip_buffers
= MIN2(max_offchip_buffers
, 508);
1542 *max_offchip_buffers_p
= max_offchip_buffers
;
1543 if (device
->physical_device
->rad_info
.chip_class
>= CIK
) {
1544 if (device
->physical_device
->rad_info
.chip_class
>= VI
)
1545 --max_offchip_buffers
;
1547 S_03093C_OFFCHIP_BUFFERING(max_offchip_buffers
) |
1548 S_03093C_OFFCHIP_GRANULARITY(offchip_granularity
);
1551 S_0089B0_OFFCHIP_BUFFERING(max_offchip_buffers
);
1553 return hs_offchip_param
;
1557 radv_get_preamble_cs(struct radv_queue
*queue
,
1558 uint32_t scratch_size
,
1559 uint32_t compute_scratch_size
,
1560 uint32_t esgs_ring_size
,
1561 uint32_t gsvs_ring_size
,
1562 bool needs_tess_rings
,
1563 bool needs_sample_positions
,
1564 struct radeon_winsys_cs
**initial_full_flush_preamble_cs
,
1565 struct radeon_winsys_cs
**initial_preamble_cs
,
1566 struct radeon_winsys_cs
**continue_preamble_cs
)
1568 struct radeon_winsys_bo
*scratch_bo
= NULL
;
1569 struct radeon_winsys_bo
*descriptor_bo
= NULL
;
1570 struct radeon_winsys_bo
*compute_scratch_bo
= NULL
;
1571 struct radeon_winsys_bo
*esgs_ring_bo
= NULL
;
1572 struct radeon_winsys_bo
*gsvs_ring_bo
= NULL
;
1573 struct radeon_winsys_bo
*tess_factor_ring_bo
= NULL
;
1574 struct radeon_winsys_bo
*tess_offchip_ring_bo
= NULL
;
1575 struct radeon_winsys_cs
*dest_cs
[3] = {0};
1576 bool add_tess_rings
= false, add_sample_positions
= false;
1577 unsigned tess_factor_ring_size
= 0, tess_offchip_ring_size
= 0;
1578 unsigned max_offchip_buffers
;
1579 unsigned hs_offchip_param
= 0;
1580 if (!queue
->has_tess_rings
) {
1581 if (needs_tess_rings
)
1582 add_tess_rings
= true;
1584 if (!queue
->has_sample_positions
) {
1585 if (needs_sample_positions
)
1586 add_sample_positions
= true;
1588 tess_factor_ring_size
= 32768 * queue
->device
->physical_device
->rad_info
.max_se
;
1589 hs_offchip_param
= radv_get_hs_offchip_param(queue
->device
,
1590 &max_offchip_buffers
);
1591 tess_offchip_ring_size
= max_offchip_buffers
*
1592 queue
->device
->tess_offchip_block_dw_size
* 4;
1594 if (scratch_size
<= queue
->scratch_size
&&
1595 compute_scratch_size
<= queue
->compute_scratch_size
&&
1596 esgs_ring_size
<= queue
->esgs_ring_size
&&
1597 gsvs_ring_size
<= queue
->gsvs_ring_size
&&
1598 !add_tess_rings
&& !add_sample_positions
&&
1599 queue
->initial_preamble_cs
) {
1600 *initial_full_flush_preamble_cs
= queue
->initial_full_flush_preamble_cs
;
1601 *initial_preamble_cs
= queue
->initial_preamble_cs
;
1602 *continue_preamble_cs
= queue
->continue_preamble_cs
;
1603 if (!scratch_size
&& !compute_scratch_size
&& !esgs_ring_size
&& !gsvs_ring_size
)
1604 *continue_preamble_cs
= NULL
;
1608 if (scratch_size
> queue
->scratch_size
) {
1609 scratch_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
1613 RADEON_FLAG_NO_CPU_ACCESS
);
1617 scratch_bo
= queue
->scratch_bo
;
1619 if (compute_scratch_size
> queue
->compute_scratch_size
) {
1620 compute_scratch_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
1621 compute_scratch_size
,
1624 RADEON_FLAG_NO_CPU_ACCESS
);
1625 if (!compute_scratch_bo
)
1629 compute_scratch_bo
= queue
->compute_scratch_bo
;
1631 if (esgs_ring_size
> queue
->esgs_ring_size
) {
1632 esgs_ring_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
1636 RADEON_FLAG_NO_CPU_ACCESS
);
1640 esgs_ring_bo
= queue
->esgs_ring_bo
;
1641 esgs_ring_size
= queue
->esgs_ring_size
;
1644 if (gsvs_ring_size
> queue
->gsvs_ring_size
) {
1645 gsvs_ring_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
1649 RADEON_FLAG_NO_CPU_ACCESS
);
1653 gsvs_ring_bo
= queue
->gsvs_ring_bo
;
1654 gsvs_ring_size
= queue
->gsvs_ring_size
;
1657 if (add_tess_rings
) {
1658 tess_factor_ring_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
1659 tess_factor_ring_size
,
1662 RADEON_FLAG_NO_CPU_ACCESS
);
1663 if (!tess_factor_ring_bo
)
1665 tess_offchip_ring_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
1666 tess_offchip_ring_size
,
1669 RADEON_FLAG_NO_CPU_ACCESS
);
1670 if (!tess_offchip_ring_bo
)
1673 tess_factor_ring_bo
= queue
->tess_factor_ring_bo
;
1674 tess_offchip_ring_bo
= queue
->tess_offchip_ring_bo
;
1677 if (scratch_bo
!= queue
->scratch_bo
||
1678 esgs_ring_bo
!= queue
->esgs_ring_bo
||
1679 gsvs_ring_bo
!= queue
->gsvs_ring_bo
||
1680 tess_factor_ring_bo
!= queue
->tess_factor_ring_bo
||
1681 tess_offchip_ring_bo
!= queue
->tess_offchip_ring_bo
|| add_sample_positions
) {
1683 if (gsvs_ring_bo
|| esgs_ring_bo
||
1684 tess_factor_ring_bo
|| tess_offchip_ring_bo
|| add_sample_positions
) {
1685 size
= 112; /* 2 dword + 2 padding + 4 dword * 6 */
1686 if (add_sample_positions
)
1687 size
+= 256; /* 32+16+8+4+2+1 samples * 4 * 2 = 248 bytes. */
1689 else if (scratch_bo
)
1690 size
= 8; /* 2 dword */
1692 descriptor_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
1696 RADEON_FLAG_CPU_ACCESS
);
1700 descriptor_bo
= queue
->descriptor_bo
;
1702 for(int i
= 0; i
< 3; ++i
) {
1703 struct radeon_winsys_cs
*cs
= NULL
;
1704 cs
= queue
->device
->ws
->cs_create(queue
->device
->ws
,
1705 queue
->queue_family_index
? RING_COMPUTE
: RING_GFX
);
1712 queue
->device
->ws
->cs_add_buffer(cs
, scratch_bo
, 8);
1715 queue
->device
->ws
->cs_add_buffer(cs
, esgs_ring_bo
, 8);
1718 queue
->device
->ws
->cs_add_buffer(cs
, gsvs_ring_bo
, 8);
1720 if (tess_factor_ring_bo
)
1721 queue
->device
->ws
->cs_add_buffer(cs
, tess_factor_ring_bo
, 8);
1723 if (tess_offchip_ring_bo
)
1724 queue
->device
->ws
->cs_add_buffer(cs
, tess_offchip_ring_bo
, 8);
1727 queue
->device
->ws
->cs_add_buffer(cs
, descriptor_bo
, 8);
1729 if (descriptor_bo
!= queue
->descriptor_bo
) {
1730 uint32_t *map
= (uint32_t*)queue
->device
->ws
->buffer_map(descriptor_bo
);
1733 uint64_t scratch_va
= queue
->device
->ws
->buffer_get_va(scratch_bo
);
1734 uint32_t rsrc1
= S_008F04_BASE_ADDRESS_HI(scratch_va
>> 32) |
1735 S_008F04_SWIZZLE_ENABLE(1);
1736 map
[0] = scratch_va
;
1740 if (esgs_ring_bo
|| gsvs_ring_bo
|| tess_factor_ring_bo
|| tess_offchip_ring_bo
||
1741 add_sample_positions
)
1742 fill_geom_tess_rings(queue
, map
, add_sample_positions
,
1743 esgs_ring_size
, esgs_ring_bo
,
1744 gsvs_ring_size
, gsvs_ring_bo
,
1745 tess_factor_ring_size
, tess_factor_ring_bo
,
1746 tess_offchip_ring_size
, tess_offchip_ring_bo
);
1748 queue
->device
->ws
->buffer_unmap(descriptor_bo
);
1751 if (esgs_ring_bo
|| gsvs_ring_bo
|| tess_factor_ring_bo
|| tess_offchip_ring_bo
) {
1752 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1753 radeon_emit(cs
, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
1754 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1755 radeon_emit(cs
, EVENT_TYPE(V_028A90_VGT_FLUSH
) | EVENT_INDEX(0));
1758 if (esgs_ring_bo
|| gsvs_ring_bo
) {
1759 if (queue
->device
->physical_device
->rad_info
.chip_class
>= CIK
) {
1760 radeon_set_uconfig_reg_seq(cs
, R_030900_VGT_ESGS_RING_SIZE
, 2);
1761 radeon_emit(cs
, esgs_ring_size
>> 8);
1762 radeon_emit(cs
, gsvs_ring_size
>> 8);
1764 radeon_set_config_reg_seq(cs
, R_0088C8_VGT_ESGS_RING_SIZE
, 2);
1765 radeon_emit(cs
, esgs_ring_size
>> 8);
1766 radeon_emit(cs
, gsvs_ring_size
>> 8);
1770 if (tess_factor_ring_bo
) {
1771 uint64_t tf_va
= queue
->device
->ws
->buffer_get_va(tess_factor_ring_bo
);
1772 if (queue
->device
->physical_device
->rad_info
.chip_class
>= CIK
) {
1773 radeon_set_uconfig_reg(cs
, R_030938_VGT_TF_RING_SIZE
,
1774 S_030938_SIZE(tess_factor_ring_size
/ 4));
1775 radeon_set_uconfig_reg(cs
, R_030940_VGT_TF_MEMORY_BASE
,
1777 if (queue
->device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
1778 radeon_set_uconfig_reg(cs
, R_030944_VGT_TF_MEMORY_BASE_HI
,
1781 radeon_set_uconfig_reg(cs
, R_03093C_VGT_HS_OFFCHIP_PARAM
, hs_offchip_param
);
1783 radeon_set_config_reg(cs
, R_008988_VGT_TF_RING_SIZE
,
1784 S_008988_SIZE(tess_factor_ring_size
/ 4));
1785 radeon_set_config_reg(cs
, R_0089B8_VGT_TF_MEMORY_BASE
,
1787 radeon_set_config_reg(cs
, R_0089B0_VGT_HS_OFFCHIP_PARAM
,
1792 if (descriptor_bo
) {
1793 uint32_t regs
[] = {R_00B030_SPI_SHADER_USER_DATA_PS_0
,
1794 R_00B130_SPI_SHADER_USER_DATA_VS_0
,
1795 R_00B230_SPI_SHADER_USER_DATA_GS_0
,
1796 R_00B330_SPI_SHADER_USER_DATA_ES_0
,
1797 R_00B430_SPI_SHADER_USER_DATA_HS_0
,
1798 R_00B530_SPI_SHADER_USER_DATA_LS_0
};
1800 uint64_t va
= queue
->device
->ws
->buffer_get_va(descriptor_bo
);
1802 for (int i
= 0; i
< ARRAY_SIZE(regs
); ++i
) {
1803 radeon_set_sh_reg_seq(cs
, regs
[i
], 2);
1804 radeon_emit(cs
, va
);
1805 radeon_emit(cs
, va
>> 32);
1809 if (compute_scratch_bo
) {
1810 uint64_t scratch_va
= queue
->device
->ws
->buffer_get_va(compute_scratch_bo
);
1811 uint32_t rsrc1
= S_008F04_BASE_ADDRESS_HI(scratch_va
>> 32) |
1812 S_008F04_SWIZZLE_ENABLE(1);
1814 queue
->device
->ws
->cs_add_buffer(cs
, compute_scratch_bo
, 8);
1816 radeon_set_sh_reg_seq(cs
, R_00B900_COMPUTE_USER_DATA_0
, 2);
1817 radeon_emit(cs
, scratch_va
);
1818 radeon_emit(cs
, rsrc1
);
1822 si_cs_emit_cache_flush(cs
,
1824 queue
->device
->physical_device
->rad_info
.chip_class
,
1826 queue
->queue_family_index
== RING_COMPUTE
&&
1827 queue
->device
->physical_device
->rad_info
.chip_class
>= CIK
,
1828 (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
)) |
1829 RADV_CMD_FLAG_INV_ICACHE
|
1830 RADV_CMD_FLAG_INV_SMEM_L1
|
1831 RADV_CMD_FLAG_INV_VMEM_L1
|
1832 RADV_CMD_FLAG_INV_GLOBAL_L2
);
1833 } else if (i
== 1) {
1834 si_cs_emit_cache_flush(cs
,
1836 queue
->device
->physical_device
->rad_info
.chip_class
,
1838 queue
->queue_family_index
== RING_COMPUTE
&&
1839 queue
->device
->physical_device
->rad_info
.chip_class
>= CIK
,
1840 RADV_CMD_FLAG_INV_ICACHE
|
1841 RADV_CMD_FLAG_INV_SMEM_L1
|
1842 RADV_CMD_FLAG_INV_VMEM_L1
|
1843 RADV_CMD_FLAG_INV_GLOBAL_L2
);
1846 if (!queue
->device
->ws
->cs_finalize(cs
))
1850 if (queue
->initial_full_flush_preamble_cs
)
1851 queue
->device
->ws
->cs_destroy(queue
->initial_full_flush_preamble_cs
);
1853 if (queue
->initial_preamble_cs
)
1854 queue
->device
->ws
->cs_destroy(queue
->initial_preamble_cs
);
1856 if (queue
->continue_preamble_cs
)
1857 queue
->device
->ws
->cs_destroy(queue
->continue_preamble_cs
);
1859 queue
->initial_full_flush_preamble_cs
= dest_cs
[0];
1860 queue
->initial_preamble_cs
= dest_cs
[1];
1861 queue
->continue_preamble_cs
= dest_cs
[2];
1863 if (scratch_bo
!= queue
->scratch_bo
) {
1864 if (queue
->scratch_bo
)
1865 queue
->device
->ws
->buffer_destroy(queue
->scratch_bo
);
1866 queue
->scratch_bo
= scratch_bo
;
1867 queue
->scratch_size
= scratch_size
;
1870 if (compute_scratch_bo
!= queue
->compute_scratch_bo
) {
1871 if (queue
->compute_scratch_bo
)
1872 queue
->device
->ws
->buffer_destroy(queue
->compute_scratch_bo
);
1873 queue
->compute_scratch_bo
= compute_scratch_bo
;
1874 queue
->compute_scratch_size
= compute_scratch_size
;
1877 if (esgs_ring_bo
!= queue
->esgs_ring_bo
) {
1878 if (queue
->esgs_ring_bo
)
1879 queue
->device
->ws
->buffer_destroy(queue
->esgs_ring_bo
);
1880 queue
->esgs_ring_bo
= esgs_ring_bo
;
1881 queue
->esgs_ring_size
= esgs_ring_size
;
1884 if (gsvs_ring_bo
!= queue
->gsvs_ring_bo
) {
1885 if (queue
->gsvs_ring_bo
)
1886 queue
->device
->ws
->buffer_destroy(queue
->gsvs_ring_bo
);
1887 queue
->gsvs_ring_bo
= gsvs_ring_bo
;
1888 queue
->gsvs_ring_size
= gsvs_ring_size
;
1891 if (tess_factor_ring_bo
!= queue
->tess_factor_ring_bo
) {
1892 queue
->tess_factor_ring_bo
= tess_factor_ring_bo
;
1895 if (tess_offchip_ring_bo
!= queue
->tess_offchip_ring_bo
) {
1896 queue
->tess_offchip_ring_bo
= tess_offchip_ring_bo
;
1897 queue
->has_tess_rings
= true;
1900 if (descriptor_bo
!= queue
->descriptor_bo
) {
1901 if (queue
->descriptor_bo
)
1902 queue
->device
->ws
->buffer_destroy(queue
->descriptor_bo
);
1904 queue
->descriptor_bo
= descriptor_bo
;
1907 if (add_sample_positions
)
1908 queue
->has_sample_positions
= true;
1910 *initial_full_flush_preamble_cs
= queue
->initial_full_flush_preamble_cs
;
1911 *initial_preamble_cs
= queue
->initial_preamble_cs
;
1912 *continue_preamble_cs
= queue
->continue_preamble_cs
;
1913 if (!scratch_size
&& !compute_scratch_size
&& !esgs_ring_size
&& !gsvs_ring_size
)
1914 *continue_preamble_cs
= NULL
;
1917 for (int i
= 0; i
< ARRAY_SIZE(dest_cs
); ++i
)
1919 queue
->device
->ws
->cs_destroy(dest_cs
[i
]);
1920 if (descriptor_bo
&& descriptor_bo
!= queue
->descriptor_bo
)
1921 queue
->device
->ws
->buffer_destroy(descriptor_bo
);
1922 if (scratch_bo
&& scratch_bo
!= queue
->scratch_bo
)
1923 queue
->device
->ws
->buffer_destroy(scratch_bo
);
1924 if (compute_scratch_bo
&& compute_scratch_bo
!= queue
->compute_scratch_bo
)
1925 queue
->device
->ws
->buffer_destroy(compute_scratch_bo
);
1926 if (esgs_ring_bo
&& esgs_ring_bo
!= queue
->esgs_ring_bo
)
1927 queue
->device
->ws
->buffer_destroy(esgs_ring_bo
);
1928 if (gsvs_ring_bo
&& gsvs_ring_bo
!= queue
->gsvs_ring_bo
)
1929 queue
->device
->ws
->buffer_destroy(gsvs_ring_bo
);
1930 if (tess_factor_ring_bo
&& tess_factor_ring_bo
!= queue
->tess_factor_ring_bo
)
1931 queue
->device
->ws
->buffer_destroy(tess_factor_ring_bo
);
1932 if (tess_offchip_ring_bo
&& tess_offchip_ring_bo
!= queue
->tess_offchip_ring_bo
)
1933 queue
->device
->ws
->buffer_destroy(tess_offchip_ring_bo
);
1934 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
1937 static VkResult
radv_alloc_sem_counts(struct radv_winsys_sem_counts
*counts
,
1939 const VkSemaphore
*sems
,
1942 int syncobj_idx
= 0, sem_idx
= 0;
1946 for (uint32_t i
= 0; i
< num_sems
; i
++) {
1947 RADV_FROM_HANDLE(radv_semaphore
, sem
, sems
[i
]);
1949 if (sem
->temp_syncobj
|| sem
->syncobj
)
1950 counts
->syncobj_count
++;
1952 counts
->sem_count
++;
1955 if (counts
->syncobj_count
) {
1956 counts
->syncobj
= (uint32_t *)malloc(sizeof(uint32_t) * counts
->syncobj_count
);
1957 if (!counts
->syncobj
)
1958 return VK_ERROR_OUT_OF_HOST_MEMORY
;
1961 if (counts
->sem_count
) {
1962 counts
->sem
= (struct radeon_winsys_sem
**)malloc(sizeof(struct radeon_winsys_sem
*) * counts
->sem_count
);
1964 free(counts
->syncobj
);
1965 return VK_ERROR_OUT_OF_HOST_MEMORY
;
1969 for (uint32_t i
= 0; i
< num_sems
; i
++) {
1970 RADV_FROM_HANDLE(radv_semaphore
, sem
, sems
[i
]);
1972 if (sem
->temp_syncobj
) {
1973 counts
->syncobj
[syncobj_idx
++] = sem
->temp_syncobj
;
1975 /* after we wait on a temp import - drop it */
1976 sem
->temp_syncobj
= 0;
1979 else if (sem
->syncobj
)
1980 counts
->syncobj
[syncobj_idx
++] = sem
->syncobj
;
1983 counts
->sem
[sem_idx
++] = sem
->sem
;
1990 void radv_free_sem_info(struct radv_winsys_sem_info
*sem_info
)
1992 free(sem_info
->wait
.syncobj
);
1993 free(sem_info
->wait
.sem
);
1994 free(sem_info
->signal
.syncobj
);
1995 free(sem_info
->signal
.sem
);
1998 VkResult
radv_alloc_sem_info(struct radv_winsys_sem_info
*sem_info
,
2000 const VkSemaphore
*wait_sems
,
2001 int num_signal_sems
,
2002 const VkSemaphore
*signal_sems
)
2005 memset(sem_info
, 0, sizeof(*sem_info
));
2007 ret
= radv_alloc_sem_counts(&sem_info
->wait
, num_wait_sems
, wait_sems
, true);
2010 ret
= radv_alloc_sem_counts(&sem_info
->signal
, num_signal_sems
, signal_sems
, false);
2012 radv_free_sem_info(sem_info
);
2014 /* caller can override these */
2015 sem_info
->cs_emit_wait
= true;
2016 sem_info
->cs_emit_signal
= true;
2020 VkResult
radv_QueueSubmit(
2022 uint32_t submitCount
,
2023 const VkSubmitInfo
* pSubmits
,
2026 RADV_FROM_HANDLE(radv_queue
, queue
, _queue
);
2027 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
2028 struct radeon_winsys_fence
*base_fence
= fence
? fence
->fence
: NULL
;
2029 struct radeon_winsys_ctx
*ctx
= queue
->hw_ctx
;
2031 uint32_t max_cs_submission
= queue
->device
->trace_bo
? 1 : UINT32_MAX
;
2032 uint32_t scratch_size
= 0;
2033 uint32_t compute_scratch_size
= 0;
2034 uint32_t esgs_ring_size
= 0, gsvs_ring_size
= 0;
2035 struct radeon_winsys_cs
*initial_preamble_cs
= NULL
, *initial_flush_preamble_cs
= NULL
, *continue_preamble_cs
= NULL
;
2037 bool fence_emitted
= false;
2038 bool tess_rings_needed
= false;
2039 bool sample_positions_needed
= false;
2041 /* Do this first so failing to allocate scratch buffers can't result in
2042 * partially executed submissions. */
2043 for (uint32_t i
= 0; i
< submitCount
; i
++) {
2044 for (uint32_t j
= 0; j
< pSubmits
[i
].commandBufferCount
; j
++) {
2045 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
,
2046 pSubmits
[i
].pCommandBuffers
[j
]);
2048 scratch_size
= MAX2(scratch_size
, cmd_buffer
->scratch_size_needed
);
2049 compute_scratch_size
= MAX2(compute_scratch_size
,
2050 cmd_buffer
->compute_scratch_size_needed
);
2051 esgs_ring_size
= MAX2(esgs_ring_size
, cmd_buffer
->esgs_ring_size_needed
);
2052 gsvs_ring_size
= MAX2(gsvs_ring_size
, cmd_buffer
->gsvs_ring_size_needed
);
2053 tess_rings_needed
|= cmd_buffer
->tess_rings_needed
;
2054 sample_positions_needed
|= cmd_buffer
->sample_positions_needed
;
2058 result
= radv_get_preamble_cs(queue
, scratch_size
, compute_scratch_size
,
2059 esgs_ring_size
, gsvs_ring_size
, tess_rings_needed
,
2060 sample_positions_needed
, &initial_flush_preamble_cs
,
2061 &initial_preamble_cs
, &continue_preamble_cs
);
2062 if (result
!= VK_SUCCESS
)
2065 for (uint32_t i
= 0; i
< submitCount
; i
++) {
2066 struct radeon_winsys_cs
**cs_array
;
2067 bool do_flush
= !i
|| pSubmits
[i
].pWaitDstStageMask
;
2068 bool can_patch
= true;
2070 struct radv_winsys_sem_info sem_info
;
2072 result
= radv_alloc_sem_info(&sem_info
,
2073 pSubmits
[i
].waitSemaphoreCount
,
2074 pSubmits
[i
].pWaitSemaphores
,
2075 pSubmits
[i
].signalSemaphoreCount
,
2076 pSubmits
[i
].pSignalSemaphores
);
2077 if (result
!= VK_SUCCESS
)
2080 if (!pSubmits
[i
].commandBufferCount
) {
2081 if (pSubmits
[i
].waitSemaphoreCount
|| pSubmits
[i
].signalSemaphoreCount
) {
2082 ret
= queue
->device
->ws
->cs_submit(ctx
, queue
->queue_idx
,
2083 &queue
->device
->empty_cs
[queue
->queue_family_index
],
2088 radv_loge("failed to submit CS %d\n", i
);
2091 fence_emitted
= true;
2093 radv_free_sem_info(&sem_info
);
2097 cs_array
= malloc(sizeof(struct radeon_winsys_cs
*) *
2098 (pSubmits
[i
].commandBufferCount
));
2100 for (uint32_t j
= 0; j
< pSubmits
[i
].commandBufferCount
; j
++) {
2101 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
,
2102 pSubmits
[i
].pCommandBuffers
[j
]);
2103 assert(cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_PRIMARY
);
2105 cs_array
[j
] = cmd_buffer
->cs
;
2106 if ((cmd_buffer
->usage_flags
& VK_COMMAND_BUFFER_USAGE_SIMULTANEOUS_USE_BIT
))
2110 for (uint32_t j
= 0; j
< pSubmits
[i
].commandBufferCount
; j
+= advance
) {
2111 struct radeon_winsys_cs
*initial_preamble
= (do_flush
&& !j
) ? initial_flush_preamble_cs
: initial_preamble_cs
;
2112 advance
= MIN2(max_cs_submission
,
2113 pSubmits
[i
].commandBufferCount
- j
);
2115 if (queue
->device
->trace_bo
)
2116 *queue
->device
->trace_id_ptr
= 0;
2118 sem_info
.cs_emit_wait
= j
== 0;
2119 sem_info
.cs_emit_signal
= j
+ advance
== pSubmits
[i
].commandBufferCount
;
2121 ret
= queue
->device
->ws
->cs_submit(ctx
, queue
->queue_idx
, cs_array
+ j
,
2122 advance
, initial_preamble
, continue_preamble_cs
,
2124 can_patch
, base_fence
);
2127 radv_loge("failed to submit CS %d\n", i
);
2130 fence_emitted
= true;
2131 if (queue
->device
->trace_bo
) {
2132 radv_check_gpu_hangs(queue
, cs_array
[j
]);
2136 radv_free_sem_info(&sem_info
);
2141 if (!fence_emitted
) {
2142 struct radv_winsys_sem_info sem_info
= {0};
2143 ret
= queue
->device
->ws
->cs_submit(ctx
, queue
->queue_idx
,
2144 &queue
->device
->empty_cs
[queue
->queue_family_index
],
2145 1, NULL
, NULL
, &sem_info
,
2148 fence
->submitted
= true;
2154 VkResult
radv_QueueWaitIdle(
2157 RADV_FROM_HANDLE(radv_queue
, queue
, _queue
);
2159 queue
->device
->ws
->ctx_wait_idle(queue
->hw_ctx
,
2160 radv_queue_family_to_ring(queue
->queue_family_index
),
2165 VkResult
radv_DeviceWaitIdle(
2168 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2170 for (unsigned i
= 0; i
< RADV_MAX_QUEUE_FAMILIES
; i
++) {
2171 for (unsigned q
= 0; q
< device
->queue_count
[i
]; q
++) {
2172 radv_QueueWaitIdle(radv_queue_to_handle(&device
->queues
[i
][q
]));
2178 PFN_vkVoidFunction
radv_GetInstanceProcAddr(
2179 VkInstance instance
,
2182 return radv_lookup_entrypoint(pName
);
2185 /* The loader wants us to expose a second GetInstanceProcAddr function
2186 * to work around certain LD_PRELOAD issues seen in apps.
2189 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
2190 VkInstance instance
,
2194 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
2195 VkInstance instance
,
2198 return radv_GetInstanceProcAddr(instance
, pName
);
2201 PFN_vkVoidFunction
radv_GetDeviceProcAddr(
2205 return radv_lookup_entrypoint(pName
);
2208 bool radv_get_memory_fd(struct radv_device
*device
,
2209 struct radv_device_memory
*memory
,
2212 struct radeon_bo_metadata metadata
;
2214 if (memory
->image
) {
2215 radv_init_metadata(device
, memory
->image
, &metadata
);
2216 device
->ws
->buffer_set_metadata(memory
->bo
, &metadata
);
2219 return device
->ws
->buffer_get_fd(device
->ws
, memory
->bo
,
2223 VkResult
radv_AllocateMemory(
2225 const VkMemoryAllocateInfo
* pAllocateInfo
,
2226 const VkAllocationCallbacks
* pAllocator
,
2227 VkDeviceMemory
* pMem
)
2229 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2230 struct radv_device_memory
*mem
;
2232 enum radeon_bo_domain domain
;
2235 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
2237 if (pAllocateInfo
->allocationSize
== 0) {
2238 /* Apparently, this is allowed */
2239 *pMem
= VK_NULL_HANDLE
;
2243 const VkImportMemoryFdInfoKHR
*import_info
=
2244 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_FD_INFO_KHR
);
2245 const VkMemoryDedicatedAllocateInfoKHR
*dedicate_info
=
2246 vk_find_struct_const(pAllocateInfo
->pNext
, MEMORY_DEDICATED_ALLOCATE_INFO_KHR
);
2248 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
2249 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2251 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2253 if (dedicate_info
) {
2254 mem
->image
= radv_image_from_handle(dedicate_info
->image
);
2255 mem
->buffer
= radv_buffer_from_handle(dedicate_info
->buffer
);
2262 assert(import_info
->handleType
==
2263 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
);
2264 mem
->bo
= device
->ws
->buffer_from_fd(device
->ws
, import_info
->fd
,
2267 result
= VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
;
2270 close(import_info
->fd
);
2275 uint64_t alloc_size
= align_u64(pAllocateInfo
->allocationSize
, 4096);
2276 if (pAllocateInfo
->memoryTypeIndex
== RADV_MEM_TYPE_GTT_WRITE_COMBINE
||
2277 pAllocateInfo
->memoryTypeIndex
== RADV_MEM_TYPE_GTT_CACHED
)
2278 domain
= RADEON_DOMAIN_GTT
;
2280 domain
= RADEON_DOMAIN_VRAM
;
2282 if (pAllocateInfo
->memoryTypeIndex
== RADV_MEM_TYPE_VRAM
)
2283 flags
|= RADEON_FLAG_NO_CPU_ACCESS
;
2285 flags
|= RADEON_FLAG_CPU_ACCESS
;
2287 if (pAllocateInfo
->memoryTypeIndex
== RADV_MEM_TYPE_GTT_WRITE_COMBINE
)
2288 flags
|= RADEON_FLAG_GTT_WC
;
2290 mem
->bo
= device
->ws
->buffer_create(device
->ws
, alloc_size
, device
->physical_device
->rad_info
.max_alignment
,
2294 result
= VK_ERROR_OUT_OF_DEVICE_MEMORY
;
2297 mem
->type_index
= pAllocateInfo
->memoryTypeIndex
;
2299 *pMem
= radv_device_memory_to_handle(mem
);
2304 vk_free2(&device
->alloc
, pAllocator
, mem
);
2309 void radv_FreeMemory(
2311 VkDeviceMemory _mem
,
2312 const VkAllocationCallbacks
* pAllocator
)
2314 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2315 RADV_FROM_HANDLE(radv_device_memory
, mem
, _mem
);
2320 device
->ws
->buffer_destroy(mem
->bo
);
2323 vk_free2(&device
->alloc
, pAllocator
, mem
);
2326 VkResult
radv_MapMemory(
2328 VkDeviceMemory _memory
,
2329 VkDeviceSize offset
,
2331 VkMemoryMapFlags flags
,
2334 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2335 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
2342 *ppData
= device
->ws
->buffer_map(mem
->bo
);
2348 return VK_ERROR_MEMORY_MAP_FAILED
;
2351 void radv_UnmapMemory(
2353 VkDeviceMemory _memory
)
2355 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2356 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
2361 device
->ws
->buffer_unmap(mem
->bo
);
2364 VkResult
radv_FlushMappedMemoryRanges(
2366 uint32_t memoryRangeCount
,
2367 const VkMappedMemoryRange
* pMemoryRanges
)
2372 VkResult
radv_InvalidateMappedMemoryRanges(
2374 uint32_t memoryRangeCount
,
2375 const VkMappedMemoryRange
* pMemoryRanges
)
2380 void radv_GetBufferMemoryRequirements(
2383 VkMemoryRequirements
* pMemoryRequirements
)
2385 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
2387 pMemoryRequirements
->memoryTypeBits
= (1u << RADV_MEM_TYPE_COUNT
) - 1;
2389 if (buffer
->flags
& VK_BUFFER_CREATE_SPARSE_BINDING_BIT
)
2390 pMemoryRequirements
->alignment
= 4096;
2392 pMemoryRequirements
->alignment
= 16;
2394 pMemoryRequirements
->size
= align64(buffer
->size
, pMemoryRequirements
->alignment
);
2397 void radv_GetBufferMemoryRequirements2KHR(
2399 const VkBufferMemoryRequirementsInfo2KHR
* pInfo
,
2400 VkMemoryRequirements2KHR
* pMemoryRequirements
)
2402 radv_GetBufferMemoryRequirements(device
, pInfo
->buffer
,
2403 &pMemoryRequirements
->memoryRequirements
);
2405 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
2406 switch (ext
->sType
) {
2407 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS_KHR
: {
2408 VkMemoryDedicatedRequirementsKHR
*req
=
2409 (VkMemoryDedicatedRequirementsKHR
*) ext
;
2410 req
->requiresDedicatedAllocation
= false;
2411 req
->prefersDedicatedAllocation
= req
->requiresDedicatedAllocation
;
2420 void radv_GetImageMemoryRequirements(
2423 VkMemoryRequirements
* pMemoryRequirements
)
2425 RADV_FROM_HANDLE(radv_image
, image
, _image
);
2427 pMemoryRequirements
->memoryTypeBits
= (1u << RADV_MEM_TYPE_COUNT
) - 1;
2429 pMemoryRequirements
->size
= image
->size
;
2430 pMemoryRequirements
->alignment
= image
->alignment
;
2433 void radv_GetImageMemoryRequirements2KHR(
2435 const VkImageMemoryRequirementsInfo2KHR
* pInfo
,
2436 VkMemoryRequirements2KHR
* pMemoryRequirements
)
2438 radv_GetImageMemoryRequirements(device
, pInfo
->image
,
2439 &pMemoryRequirements
->memoryRequirements
);
2441 RADV_FROM_HANDLE(radv_image
, image
, pInfo
->image
);
2443 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
2444 switch (ext
->sType
) {
2445 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS_KHR
: {
2446 VkMemoryDedicatedRequirementsKHR
*req
=
2447 (VkMemoryDedicatedRequirementsKHR
*) ext
;
2448 req
->requiresDedicatedAllocation
= image
->shareable
;
2449 req
->prefersDedicatedAllocation
= req
->requiresDedicatedAllocation
;
2458 void radv_GetImageSparseMemoryRequirements(
2461 uint32_t* pSparseMemoryRequirementCount
,
2462 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
2467 void radv_GetImageSparseMemoryRequirements2KHR(
2469 const VkImageSparseMemoryRequirementsInfo2KHR
* pInfo
,
2470 uint32_t* pSparseMemoryRequirementCount
,
2471 VkSparseImageMemoryRequirements2KHR
* pSparseMemoryRequirements
)
2476 void radv_GetDeviceMemoryCommitment(
2478 VkDeviceMemory memory
,
2479 VkDeviceSize
* pCommittedMemoryInBytes
)
2481 *pCommittedMemoryInBytes
= 0;
2484 VkResult
radv_BindBufferMemory(
2487 VkDeviceMemory _memory
,
2488 VkDeviceSize memoryOffset
)
2490 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
2491 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
2494 buffer
->bo
= mem
->bo
;
2495 buffer
->offset
= memoryOffset
;
2504 VkResult
radv_BindImageMemory(
2507 VkDeviceMemory _memory
,
2508 VkDeviceSize memoryOffset
)
2510 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
2511 RADV_FROM_HANDLE(radv_image
, image
, _image
);
2514 image
->bo
= mem
->bo
;
2515 image
->offset
= memoryOffset
;
2526 radv_sparse_buffer_bind_memory(struct radv_device
*device
,
2527 const VkSparseBufferMemoryBindInfo
*bind
)
2529 RADV_FROM_HANDLE(radv_buffer
, buffer
, bind
->buffer
);
2531 for (uint32_t i
= 0; i
< bind
->bindCount
; ++i
) {
2532 struct radv_device_memory
*mem
= NULL
;
2534 if (bind
->pBinds
[i
].memory
!= VK_NULL_HANDLE
)
2535 mem
= radv_device_memory_from_handle(bind
->pBinds
[i
].memory
);
2537 device
->ws
->buffer_virtual_bind(buffer
->bo
,
2538 bind
->pBinds
[i
].resourceOffset
,
2539 bind
->pBinds
[i
].size
,
2540 mem
? mem
->bo
: NULL
,
2541 bind
->pBinds
[i
].memoryOffset
);
2546 radv_sparse_image_opaque_bind_memory(struct radv_device
*device
,
2547 const VkSparseImageOpaqueMemoryBindInfo
*bind
)
2549 RADV_FROM_HANDLE(radv_image
, image
, bind
->image
);
2551 for (uint32_t i
= 0; i
< bind
->bindCount
; ++i
) {
2552 struct radv_device_memory
*mem
= NULL
;
2554 if (bind
->pBinds
[i
].memory
!= VK_NULL_HANDLE
)
2555 mem
= radv_device_memory_from_handle(bind
->pBinds
[i
].memory
);
2557 device
->ws
->buffer_virtual_bind(image
->bo
,
2558 bind
->pBinds
[i
].resourceOffset
,
2559 bind
->pBinds
[i
].size
,
2560 mem
? mem
->bo
: NULL
,
2561 bind
->pBinds
[i
].memoryOffset
);
2565 VkResult
radv_QueueBindSparse(
2567 uint32_t bindInfoCount
,
2568 const VkBindSparseInfo
* pBindInfo
,
2571 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
2572 RADV_FROM_HANDLE(radv_queue
, queue
, _queue
);
2573 struct radeon_winsys_fence
*base_fence
= fence
? fence
->fence
: NULL
;
2574 bool fence_emitted
= false;
2576 for (uint32_t i
= 0; i
< bindInfoCount
; ++i
) {
2577 struct radv_winsys_sem_info sem_info
;
2578 for (uint32_t j
= 0; j
< pBindInfo
[i
].bufferBindCount
; ++j
) {
2579 radv_sparse_buffer_bind_memory(queue
->device
,
2580 pBindInfo
[i
].pBufferBinds
+ j
);
2583 for (uint32_t j
= 0; j
< pBindInfo
[i
].imageOpaqueBindCount
; ++j
) {
2584 radv_sparse_image_opaque_bind_memory(queue
->device
,
2585 pBindInfo
[i
].pImageOpaqueBinds
+ j
);
2589 result
= radv_alloc_sem_info(&sem_info
,
2590 pBindInfo
[i
].waitSemaphoreCount
,
2591 pBindInfo
[i
].pWaitSemaphores
,
2592 pBindInfo
[i
].signalSemaphoreCount
,
2593 pBindInfo
[i
].pSignalSemaphores
);
2594 if (result
!= VK_SUCCESS
)
2597 if (pBindInfo
[i
].waitSemaphoreCount
|| pBindInfo
[i
].signalSemaphoreCount
) {
2598 queue
->device
->ws
->cs_submit(queue
->hw_ctx
, queue
->queue_idx
,
2599 &queue
->device
->empty_cs
[queue
->queue_family_index
],
2603 fence_emitted
= true;
2605 fence
->submitted
= true;
2608 radv_free_sem_info(&sem_info
);
2612 if (fence
&& !fence_emitted
) {
2613 fence
->signalled
= true;
2619 VkResult
radv_CreateFence(
2621 const VkFenceCreateInfo
* pCreateInfo
,
2622 const VkAllocationCallbacks
* pAllocator
,
2625 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2626 struct radv_fence
*fence
= vk_alloc2(&device
->alloc
, pAllocator
,
2628 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2631 return VK_ERROR_OUT_OF_HOST_MEMORY
;
2633 memset(fence
, 0, sizeof(*fence
));
2634 fence
->submitted
= false;
2635 fence
->signalled
= !!(pCreateInfo
->flags
& VK_FENCE_CREATE_SIGNALED_BIT
);
2636 fence
->fence
= device
->ws
->create_fence();
2637 if (!fence
->fence
) {
2638 vk_free2(&device
->alloc
, pAllocator
, fence
);
2639 return VK_ERROR_OUT_OF_HOST_MEMORY
;
2642 *pFence
= radv_fence_to_handle(fence
);
2647 void radv_DestroyFence(
2650 const VkAllocationCallbacks
* pAllocator
)
2652 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2653 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
2657 device
->ws
->destroy_fence(fence
->fence
);
2658 vk_free2(&device
->alloc
, pAllocator
, fence
);
2661 static uint64_t radv_get_absolute_timeout(uint64_t timeout
)
2663 uint64_t current_time
;
2666 clock_gettime(CLOCK_MONOTONIC
, &tv
);
2667 current_time
= tv
.tv_nsec
+ tv
.tv_sec
*1000000000ull;
2669 timeout
= MIN2(UINT64_MAX
- current_time
, timeout
);
2671 return current_time
+ timeout
;
2674 VkResult
radv_WaitForFences(
2676 uint32_t fenceCount
,
2677 const VkFence
* pFences
,
2681 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2682 timeout
= radv_get_absolute_timeout(timeout
);
2684 if (!waitAll
&& fenceCount
> 1) {
2685 fprintf(stderr
, "radv: WaitForFences without waitAll not implemented yet\n");
2688 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
2689 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
2690 bool expired
= false;
2692 if (fence
->signalled
)
2695 if (!fence
->submitted
)
2698 expired
= device
->ws
->fence_wait(device
->ws
, fence
->fence
, true, timeout
);
2702 fence
->signalled
= true;
2708 VkResult
radv_ResetFences(VkDevice device
,
2709 uint32_t fenceCount
,
2710 const VkFence
*pFences
)
2712 for (unsigned i
= 0; i
< fenceCount
; ++i
) {
2713 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
2714 fence
->submitted
= fence
->signalled
= false;
2720 VkResult
radv_GetFenceStatus(VkDevice _device
, VkFence _fence
)
2722 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2723 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
2725 if (fence
->signalled
)
2727 if (!fence
->submitted
)
2728 return VK_NOT_READY
;
2730 if (!device
->ws
->fence_wait(device
->ws
, fence
->fence
, false, 0))
2731 return VK_NOT_READY
;
2737 // Queue semaphore functions
2739 VkResult
radv_CreateSemaphore(
2741 const VkSemaphoreCreateInfo
* pCreateInfo
,
2742 const VkAllocationCallbacks
* pAllocator
,
2743 VkSemaphore
* pSemaphore
)
2745 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2746 const VkExportSemaphoreCreateInfoKHR
*export
=
2747 vk_find_struct_const(pCreateInfo
->pNext
, EXPORT_SEMAPHORE_CREATE_INFO_KHR
);
2748 VkExternalSemaphoreHandleTypeFlagsKHR handleTypes
=
2749 export
? export
->handleTypes
: 0;
2751 struct radv_semaphore
*sem
= vk_alloc2(&device
->alloc
, pAllocator
,
2753 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2755 return VK_ERROR_OUT_OF_HOST_MEMORY
;
2757 sem
->temp_syncobj
= 0;
2758 /* create a syncobject if we are going to export this semaphore */
2760 assert (device
->physical_device
->rad_info
.has_syncobj
);
2761 assert (handleTypes
== VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
);
2762 int ret
= device
->ws
->create_syncobj(device
->ws
, &sem
->syncobj
);
2764 vk_free2(&device
->alloc
, pAllocator
, sem
);
2765 return VK_ERROR_OUT_OF_HOST_MEMORY
;
2769 sem
->sem
= device
->ws
->create_sem(device
->ws
);
2771 vk_free2(&device
->alloc
, pAllocator
, sem
);
2772 return VK_ERROR_OUT_OF_HOST_MEMORY
;
2777 *pSemaphore
= radv_semaphore_to_handle(sem
);
2781 void radv_DestroySemaphore(
2783 VkSemaphore _semaphore
,
2784 const VkAllocationCallbacks
* pAllocator
)
2786 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2787 RADV_FROM_HANDLE(radv_semaphore
, sem
, _semaphore
);
2792 device
->ws
->destroy_syncobj(device
->ws
, sem
->syncobj
);
2794 device
->ws
->destroy_sem(sem
->sem
);
2795 vk_free2(&device
->alloc
, pAllocator
, sem
);
2798 VkResult
radv_CreateEvent(
2800 const VkEventCreateInfo
* pCreateInfo
,
2801 const VkAllocationCallbacks
* pAllocator
,
2804 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2805 struct radv_event
*event
= vk_alloc2(&device
->alloc
, pAllocator
,
2807 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2810 return VK_ERROR_OUT_OF_HOST_MEMORY
;
2812 event
->bo
= device
->ws
->buffer_create(device
->ws
, 8, 8,
2814 RADEON_FLAG_VA_UNCACHED
| RADEON_FLAG_CPU_ACCESS
);
2816 vk_free2(&device
->alloc
, pAllocator
, event
);
2817 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
2820 event
->map
= (uint64_t*)device
->ws
->buffer_map(event
->bo
);
2822 *pEvent
= radv_event_to_handle(event
);
2827 void radv_DestroyEvent(
2830 const VkAllocationCallbacks
* pAllocator
)
2832 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2833 RADV_FROM_HANDLE(radv_event
, event
, _event
);
2837 device
->ws
->buffer_destroy(event
->bo
);
2838 vk_free2(&device
->alloc
, pAllocator
, event
);
2841 VkResult
radv_GetEventStatus(
2845 RADV_FROM_HANDLE(radv_event
, event
, _event
);
2847 if (*event
->map
== 1)
2848 return VK_EVENT_SET
;
2849 return VK_EVENT_RESET
;
2852 VkResult
radv_SetEvent(
2856 RADV_FROM_HANDLE(radv_event
, event
, _event
);
2862 VkResult
radv_ResetEvent(
2866 RADV_FROM_HANDLE(radv_event
, event
, _event
);
2872 VkResult
radv_CreateBuffer(
2874 const VkBufferCreateInfo
* pCreateInfo
,
2875 const VkAllocationCallbacks
* pAllocator
,
2878 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2879 struct radv_buffer
*buffer
;
2881 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
2883 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
2884 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2886 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2888 buffer
->size
= pCreateInfo
->size
;
2889 buffer
->usage
= pCreateInfo
->usage
;
2892 buffer
->flags
= pCreateInfo
->flags
;
2894 if (pCreateInfo
->flags
& VK_BUFFER_CREATE_SPARSE_BINDING_BIT
) {
2895 buffer
->bo
= device
->ws
->buffer_create(device
->ws
,
2896 align64(buffer
->size
, 4096),
2897 4096, 0, RADEON_FLAG_VIRTUAL
);
2899 vk_free2(&device
->alloc
, pAllocator
, buffer
);
2900 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
2904 *pBuffer
= radv_buffer_to_handle(buffer
);
2909 void radv_DestroyBuffer(
2912 const VkAllocationCallbacks
* pAllocator
)
2914 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2915 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
2920 if (buffer
->flags
& VK_BUFFER_CREATE_SPARSE_BINDING_BIT
)
2921 device
->ws
->buffer_destroy(buffer
->bo
);
2923 vk_free2(&device
->alloc
, pAllocator
, buffer
);
2926 static inline unsigned
2927 si_tile_mode_index(const struct radv_image
*image
, unsigned level
, bool stencil
)
2930 return image
->surface
.u
.legacy
.stencil_tiling_index
[level
];
2932 return image
->surface
.u
.legacy
.tiling_index
[level
];
2935 static uint32_t radv_surface_layer_count(struct radv_image_view
*iview
)
2937 return iview
->type
== VK_IMAGE_VIEW_TYPE_3D
? iview
->extent
.depth
: iview
->layer_count
;
2941 radv_initialise_color_surface(struct radv_device
*device
,
2942 struct radv_color_buffer_info
*cb
,
2943 struct radv_image_view
*iview
)
2945 const struct vk_format_description
*desc
;
2946 unsigned ntype
, format
, swap
, endian
;
2947 unsigned blend_clamp
= 0, blend_bypass
= 0;
2949 const struct radeon_surf
*surf
= &iview
->image
->surface
;
2951 desc
= vk_format_description(iview
->vk_format
);
2953 memset(cb
, 0, sizeof(*cb
));
2955 /* Intensity is implemented as Red, so treat it that way. */
2956 cb
->cb_color_attrib
= S_028C74_FORCE_DST_ALPHA_1(desc
->swizzle
[3] == VK_SWIZZLE_1
);
2958 va
= device
->ws
->buffer_get_va(iview
->bo
) + iview
->image
->offset
;
2960 cb
->cb_color_base
= va
>> 8;
2962 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
2963 struct gfx9_surf_meta_flags meta
;
2964 if (iview
->image
->dcc_offset
)
2965 meta
= iview
->image
->surface
.u
.gfx9
.dcc
;
2967 meta
= iview
->image
->surface
.u
.gfx9
.cmask
;
2969 cb
->cb_color_attrib
|= S_028C74_COLOR_SW_MODE(iview
->image
->surface
.u
.gfx9
.surf
.swizzle_mode
) |
2970 S_028C74_FMASK_SW_MODE(iview
->image
->surface
.u
.gfx9
.fmask
.swizzle_mode
) |
2971 S_028C74_RB_ALIGNED(meta
.rb_aligned
) |
2972 S_028C74_PIPE_ALIGNED(meta
.pipe_aligned
);
2974 cb
->cb_color_base
+= iview
->image
->surface
.u
.gfx9
.surf_offset
>> 8;
2975 cb
->cb_color_base
|= iview
->image
->surface
.tile_swizzle
;
2977 const struct legacy_surf_level
*level_info
= &surf
->u
.legacy
.level
[iview
->base_mip
];
2978 unsigned pitch_tile_max
, slice_tile_max
, tile_mode_index
;
2980 cb
->cb_color_base
+= level_info
->offset
>> 8;
2981 if (level_info
->mode
== RADEON_SURF_MODE_2D
)
2982 cb
->cb_color_base
|= iview
->image
->surface
.tile_swizzle
;
2984 pitch_tile_max
= level_info
->nblk_x
/ 8 - 1;
2985 slice_tile_max
= (level_info
->nblk_x
* level_info
->nblk_y
) / 64 - 1;
2986 tile_mode_index
= si_tile_mode_index(iview
->image
, iview
->base_mip
, false);
2988 cb
->cb_color_pitch
= S_028C64_TILE_MAX(pitch_tile_max
);
2989 cb
->cb_color_slice
= S_028C68_TILE_MAX(slice_tile_max
);
2990 cb
->cb_color_cmask_slice
= iview
->image
->cmask
.slice_tile_max
;
2992 cb
->cb_color_attrib
|= S_028C74_TILE_MODE_INDEX(tile_mode_index
);
2993 cb
->micro_tile_mode
= iview
->image
->surface
.micro_tile_mode
;
2995 if (iview
->image
->fmask
.size
) {
2996 if (device
->physical_device
->rad_info
.chip_class
>= CIK
)
2997 cb
->cb_color_pitch
|= S_028C64_FMASK_TILE_MAX(iview
->image
->fmask
.pitch_in_pixels
/ 8 - 1);
2998 cb
->cb_color_attrib
|= S_028C74_FMASK_TILE_MODE_INDEX(iview
->image
->fmask
.tile_mode_index
);
2999 cb
->cb_color_fmask_slice
= S_028C88_TILE_MAX(iview
->image
->fmask
.slice_tile_max
);
3001 /* This must be set for fast clear to work without FMASK. */
3002 if (device
->physical_device
->rad_info
.chip_class
>= CIK
)
3003 cb
->cb_color_pitch
|= S_028C64_FMASK_TILE_MAX(pitch_tile_max
);
3004 cb
->cb_color_attrib
|= S_028C74_FMASK_TILE_MODE_INDEX(tile_mode_index
);
3005 cb
->cb_color_fmask_slice
= S_028C88_TILE_MAX(slice_tile_max
);
3009 /* CMASK variables */
3010 va
= device
->ws
->buffer_get_va(iview
->bo
) + iview
->image
->offset
;
3011 va
+= iview
->image
->cmask
.offset
;
3012 cb
->cb_color_cmask
= va
>> 8;
3014 va
= device
->ws
->buffer_get_va(iview
->bo
) + iview
->image
->offset
;
3015 va
+= iview
->image
->dcc_offset
;
3016 cb
->cb_dcc_base
= va
>> 8;
3017 cb
->cb_dcc_base
|= iview
->image
->surface
.tile_swizzle
;
3019 uint32_t max_slice
= radv_surface_layer_count(iview
);
3020 cb
->cb_color_view
= S_028C6C_SLICE_START(iview
->base_layer
) |
3021 S_028C6C_SLICE_MAX(iview
->base_layer
+ max_slice
- 1);
3023 if (iview
->image
->info
.samples
> 1) {
3024 unsigned log_samples
= util_logbase2(iview
->image
->info
.samples
);
3026 cb
->cb_color_attrib
|= S_028C74_NUM_SAMPLES(log_samples
) |
3027 S_028C74_NUM_FRAGMENTS(log_samples
);
3030 if (iview
->image
->fmask
.size
) {
3031 va
= device
->ws
->buffer_get_va(iview
->bo
) + iview
->image
->offset
+ iview
->image
->fmask
.offset
;
3032 cb
->cb_color_fmask
= va
>> 8;
3033 cb
->cb_color_fmask
|= iview
->image
->fmask
.tile_swizzle
;
3035 cb
->cb_color_fmask
= cb
->cb_color_base
;
3038 ntype
= radv_translate_color_numformat(iview
->vk_format
,
3040 vk_format_get_first_non_void_channel(iview
->vk_format
));
3041 format
= radv_translate_colorformat(iview
->vk_format
);
3042 if (format
== V_028C70_COLOR_INVALID
|| ntype
== ~0u)
3043 radv_finishme("Illegal color\n");
3044 swap
= radv_translate_colorswap(iview
->vk_format
, FALSE
);
3045 endian
= radv_colorformat_endian_swap(format
);
3047 /* blend clamp should be set for all NORM/SRGB types */
3048 if (ntype
== V_028C70_NUMBER_UNORM
||
3049 ntype
== V_028C70_NUMBER_SNORM
||
3050 ntype
== V_028C70_NUMBER_SRGB
)
3053 /* set blend bypass according to docs if SINT/UINT or
3054 8/24 COLOR variants */
3055 if (ntype
== V_028C70_NUMBER_UINT
|| ntype
== V_028C70_NUMBER_SINT
||
3056 format
== V_028C70_COLOR_8_24
|| format
== V_028C70_COLOR_24_8
||
3057 format
== V_028C70_COLOR_X24_8_32_FLOAT
) {
3062 if ((ntype
== V_028C70_NUMBER_UINT
|| ntype
== V_028C70_NUMBER_SINT
) &&
3063 (format
== V_028C70_COLOR_8
||
3064 format
== V_028C70_COLOR_8_8
||
3065 format
== V_028C70_COLOR_8_8_8_8
))
3066 ->color_is_int8
= true;
3068 cb
->cb_color_info
= S_028C70_FORMAT(format
) |
3069 S_028C70_COMP_SWAP(swap
) |
3070 S_028C70_BLEND_CLAMP(blend_clamp
) |
3071 S_028C70_BLEND_BYPASS(blend_bypass
) |
3072 S_028C70_SIMPLE_FLOAT(1) |
3073 S_028C70_ROUND_MODE(ntype
!= V_028C70_NUMBER_UNORM
&&
3074 ntype
!= V_028C70_NUMBER_SNORM
&&
3075 ntype
!= V_028C70_NUMBER_SRGB
&&
3076 format
!= V_028C70_COLOR_8_24
&&
3077 format
!= V_028C70_COLOR_24_8
) |
3078 S_028C70_NUMBER_TYPE(ntype
) |
3079 S_028C70_ENDIAN(endian
);
3080 if ((iview
->image
->info
.samples
> 1) && iview
->image
->fmask
.size
) {
3081 cb
->cb_color_info
|= S_028C70_COMPRESSION(1);
3082 if (device
->physical_device
->rad_info
.chip_class
== SI
) {
3083 unsigned fmask_bankh
= util_logbase2(iview
->image
->fmask
.bank_height
);
3084 cb
->cb_color_attrib
|= S_028C74_FMASK_BANK_HEIGHT(fmask_bankh
);
3088 if (iview
->image
->cmask
.size
&&
3089 !(device
->debug_flags
& RADV_DEBUG_NO_FAST_CLEARS
))
3090 cb
->cb_color_info
|= S_028C70_FAST_CLEAR(1);
3092 if (iview
->image
->surface
.dcc_size
&& iview
->base_mip
< surf
->num_dcc_levels
)
3093 cb
->cb_color_info
|= S_028C70_DCC_ENABLE(1);
3095 if (device
->physical_device
->rad_info
.chip_class
>= VI
) {
3096 unsigned max_uncompressed_block_size
= 2;
3097 if (iview
->image
->info
.samples
> 1) {
3098 if (iview
->image
->surface
.bpe
== 1)
3099 max_uncompressed_block_size
= 0;
3100 else if (iview
->image
->surface
.bpe
== 2)
3101 max_uncompressed_block_size
= 1;
3104 cb
->cb_dcc_control
= S_028C78_MAX_UNCOMPRESSED_BLOCK_SIZE(max_uncompressed_block_size
) |
3105 S_028C78_INDEPENDENT_64B_BLOCKS(1);
3108 /* This must be set for fast clear to work without FMASK. */
3109 if (!iview
->image
->fmask
.size
&&
3110 device
->physical_device
->rad_info
.chip_class
== SI
) {
3111 unsigned bankh
= util_logbase2(iview
->image
->surface
.u
.legacy
.bankh
);
3112 cb
->cb_color_attrib
|= S_028C74_FMASK_BANK_HEIGHT(bankh
);
3115 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
3116 unsigned mip0_depth
= iview
->image
->type
== VK_IMAGE_TYPE_3D
?
3117 (iview
->extent
.depth
- 1) : (iview
->image
->info
.array_size
- 1);
3119 cb
->cb_color_view
|= S_028C6C_MIP_LEVEL(iview
->base_mip
);
3120 cb
->cb_color_attrib
|= S_028C74_MIP0_DEPTH(mip0_depth
) |
3121 S_028C74_RESOURCE_TYPE(iview
->image
->surface
.u
.gfx9
.resource_type
);
3122 cb
->cb_color_attrib2
= S_028C68_MIP0_WIDTH(iview
->extent
.width
- 1) |
3123 S_028C68_MIP0_HEIGHT(iview
->extent
.height
- 1) |
3124 S_028C68_MAX_MIP(iview
->image
->info
.levels
- 1);
3126 cb
->gfx9_epitch
= S_0287A0_EPITCH(iview
->image
->surface
.u
.gfx9
.surf
.epitch
);
3132 radv_initialise_ds_surface(struct radv_device
*device
,
3133 struct radv_ds_buffer_info
*ds
,
3134 struct radv_image_view
*iview
)
3136 unsigned level
= iview
->base_mip
;
3137 unsigned format
, stencil_format
;
3138 uint64_t va
, s_offs
, z_offs
;
3139 bool stencil_only
= false;
3140 memset(ds
, 0, sizeof(*ds
));
3141 switch (iview
->image
->vk_format
) {
3142 case VK_FORMAT_D24_UNORM_S8_UINT
:
3143 case VK_FORMAT_X8_D24_UNORM_PACK32
:
3144 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-24);
3145 ds
->offset_scale
= 2.0f
;
3147 case VK_FORMAT_D16_UNORM
:
3148 case VK_FORMAT_D16_UNORM_S8_UINT
:
3149 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-16);
3150 ds
->offset_scale
= 4.0f
;
3152 case VK_FORMAT_D32_SFLOAT
:
3153 case VK_FORMAT_D32_SFLOAT_S8_UINT
:
3154 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-23) |
3155 S_028B78_POLY_OFFSET_DB_IS_FLOAT_FMT(1);
3156 ds
->offset_scale
= 1.0f
;
3158 case VK_FORMAT_S8_UINT
:
3159 stencil_only
= true;
3165 format
= radv_translate_dbformat(iview
->image
->vk_format
);
3166 stencil_format
= iview
->image
->surface
.has_stencil
?
3167 V_028044_STENCIL_8
: V_028044_STENCIL_INVALID
;
3169 uint32_t max_slice
= radv_surface_layer_count(iview
);
3170 ds
->db_depth_view
= S_028008_SLICE_START(iview
->base_layer
) |
3171 S_028008_SLICE_MAX(iview
->base_layer
+ max_slice
- 1);
3173 ds
->db_htile_data_base
= 0;
3174 ds
->db_htile_surface
= 0;
3176 va
= device
->ws
->buffer_get_va(iview
->bo
) + iview
->image
->offset
;
3177 s_offs
= z_offs
= va
;
3179 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
3180 assert(iview
->image
->surface
.u
.gfx9
.surf_offset
== 0);
3181 s_offs
+= iview
->image
->surface
.u
.gfx9
.stencil_offset
;
3183 ds
->db_z_info
= S_028038_FORMAT(format
) |
3184 S_028038_NUM_SAMPLES(util_logbase2(iview
->image
->info
.samples
)) |
3185 S_028038_SW_MODE(iview
->image
->surface
.u
.gfx9
.surf
.swizzle_mode
) |
3186 S_028038_MAXMIP(iview
->image
->info
.levels
- 1);
3187 ds
->db_stencil_info
= S_02803C_FORMAT(stencil_format
) |
3188 S_02803C_SW_MODE(iview
->image
->surface
.u
.gfx9
.stencil
.swizzle_mode
);
3190 ds
->db_z_info2
= S_028068_EPITCH(iview
->image
->surface
.u
.gfx9
.surf
.epitch
);
3191 ds
->db_stencil_info2
= S_02806C_EPITCH(iview
->image
->surface
.u
.gfx9
.stencil
.epitch
);
3192 ds
->db_depth_view
|= S_028008_MIPID(level
);
3194 ds
->db_depth_size
= S_02801C_X_MAX(iview
->image
->info
.width
- 1) |
3195 S_02801C_Y_MAX(iview
->image
->info
.height
- 1);
3197 /* Only use HTILE for the first level. */
3198 if (iview
->image
->surface
.htile_size
&& !level
) {
3199 ds
->db_z_info
|= S_028038_TILE_SURFACE_ENABLE(1);
3201 if (!iview
->image
->surface
.has_stencil
)
3202 /* Use all of the htile_buffer for depth if there's no stencil. */
3203 ds
->db_stencil_info
|= S_02803C_TILE_STENCIL_DISABLE(1);
3204 va
= device
->ws
->buffer_get_va(iview
->bo
) + iview
->image
->offset
+
3205 iview
->image
->htile_offset
;
3206 ds
->db_htile_data_base
= va
>> 8;
3207 ds
->db_htile_surface
= S_028ABC_FULL_CACHE(1) |
3208 S_028ABC_PIPE_ALIGNED(iview
->image
->surface
.u
.gfx9
.htile
.pipe_aligned
) |
3209 S_028ABC_RB_ALIGNED(iview
->image
->surface
.u
.gfx9
.htile
.rb_aligned
);
3212 const struct legacy_surf_level
*level_info
= &iview
->image
->surface
.u
.legacy
.level
[level
];
3215 level_info
= &iview
->image
->surface
.u
.legacy
.stencil_level
[level
];
3217 z_offs
+= iview
->image
->surface
.u
.legacy
.level
[level
].offset
;
3218 s_offs
+= iview
->image
->surface
.u
.legacy
.stencil_level
[level
].offset
;
3220 ds
->db_depth_info
= S_02803C_ADDR5_SWIZZLE_MASK(1);
3221 ds
->db_z_info
= S_028040_FORMAT(format
) | S_028040_ZRANGE_PRECISION(1);
3222 ds
->db_stencil_info
= S_028044_FORMAT(stencil_format
);
3224 if (iview
->image
->info
.samples
> 1)
3225 ds
->db_z_info
|= S_028040_NUM_SAMPLES(util_logbase2(iview
->image
->info
.samples
));
3227 if (device
->physical_device
->rad_info
.chip_class
>= CIK
) {
3228 struct radeon_info
*info
= &device
->physical_device
->rad_info
;
3229 unsigned tiling_index
= iview
->image
->surface
.u
.legacy
.tiling_index
[level
];
3230 unsigned stencil_index
= iview
->image
->surface
.u
.legacy
.stencil_tiling_index
[level
];
3231 unsigned macro_index
= iview
->image
->surface
.u
.legacy
.macro_tile_index
;
3232 unsigned tile_mode
= info
->si_tile_mode_array
[tiling_index
];
3233 unsigned stencil_tile_mode
= info
->si_tile_mode_array
[stencil_index
];
3234 unsigned macro_mode
= info
->cik_macrotile_mode_array
[macro_index
];
3237 tile_mode
= stencil_tile_mode
;
3239 ds
->db_depth_info
|=
3240 S_02803C_ARRAY_MODE(G_009910_ARRAY_MODE(tile_mode
)) |
3241 S_02803C_PIPE_CONFIG(G_009910_PIPE_CONFIG(tile_mode
)) |
3242 S_02803C_BANK_WIDTH(G_009990_BANK_WIDTH(macro_mode
)) |
3243 S_02803C_BANK_HEIGHT(G_009990_BANK_HEIGHT(macro_mode
)) |
3244 S_02803C_MACRO_TILE_ASPECT(G_009990_MACRO_TILE_ASPECT(macro_mode
)) |
3245 S_02803C_NUM_BANKS(G_009990_NUM_BANKS(macro_mode
));
3246 ds
->db_z_info
|= S_028040_TILE_SPLIT(G_009910_TILE_SPLIT(tile_mode
));
3247 ds
->db_stencil_info
|= S_028044_TILE_SPLIT(G_009910_TILE_SPLIT(stencil_tile_mode
));
3249 unsigned tile_mode_index
= si_tile_mode_index(iview
->image
, level
, false);
3250 ds
->db_z_info
|= S_028040_TILE_MODE_INDEX(tile_mode_index
);
3251 tile_mode_index
= si_tile_mode_index(iview
->image
, level
, true);
3252 ds
->db_stencil_info
|= S_028044_TILE_MODE_INDEX(tile_mode_index
);
3254 ds
->db_z_info
|= S_028040_TILE_MODE_INDEX(tile_mode_index
);
3257 ds
->db_depth_size
= S_028058_PITCH_TILE_MAX((level_info
->nblk_x
/ 8) - 1) |
3258 S_028058_HEIGHT_TILE_MAX((level_info
->nblk_y
/ 8) - 1);
3259 ds
->db_depth_slice
= S_02805C_SLICE_TILE_MAX((level_info
->nblk_x
* level_info
->nblk_y
) / 64 - 1);
3261 if (iview
->image
->surface
.htile_size
&& !level
) {
3262 ds
->db_z_info
|= S_028040_TILE_SURFACE_ENABLE(1);
3264 if (!iview
->image
->surface
.has_stencil
)
3265 /* Use all of the htile_buffer for depth if there's no stencil. */
3266 ds
->db_stencil_info
|= S_028044_TILE_STENCIL_DISABLE(1);
3268 va
= device
->ws
->buffer_get_va(iview
->bo
) + iview
->image
->offset
+
3269 iview
->image
->htile_offset
;
3270 ds
->db_htile_data_base
= va
>> 8;
3271 ds
->db_htile_surface
= S_028ABC_FULL_CACHE(1);
3275 ds
->db_z_read_base
= ds
->db_z_write_base
= z_offs
>> 8;
3276 ds
->db_stencil_read_base
= ds
->db_stencil_write_base
= s_offs
>> 8;
3279 VkResult
radv_CreateFramebuffer(
3281 const VkFramebufferCreateInfo
* pCreateInfo
,
3282 const VkAllocationCallbacks
* pAllocator
,
3283 VkFramebuffer
* pFramebuffer
)
3285 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3286 struct radv_framebuffer
*framebuffer
;
3288 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
3290 size_t size
= sizeof(*framebuffer
) +
3291 sizeof(struct radv_attachment_info
) * pCreateInfo
->attachmentCount
;
3292 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
3293 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3294 if (framebuffer
== NULL
)
3295 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3297 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
3298 framebuffer
->width
= pCreateInfo
->width
;
3299 framebuffer
->height
= pCreateInfo
->height
;
3300 framebuffer
->layers
= pCreateInfo
->layers
;
3301 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
3302 VkImageView _iview
= pCreateInfo
->pAttachments
[i
];
3303 struct radv_image_view
*iview
= radv_image_view_from_handle(_iview
);
3304 framebuffer
->attachments
[i
].attachment
= iview
;
3305 if (iview
->aspect_mask
& VK_IMAGE_ASPECT_COLOR_BIT
) {
3306 radv_initialise_color_surface(device
, &framebuffer
->attachments
[i
].cb
, iview
);
3307 } else if (iview
->aspect_mask
& (VK_IMAGE_ASPECT_DEPTH_BIT
| VK_IMAGE_ASPECT_STENCIL_BIT
)) {
3308 radv_initialise_ds_surface(device
, &framebuffer
->attachments
[i
].ds
, iview
);
3310 framebuffer
->width
= MIN2(framebuffer
->width
, iview
->extent
.width
);
3311 framebuffer
->height
= MIN2(framebuffer
->height
, iview
->extent
.height
);
3312 framebuffer
->layers
= MIN2(framebuffer
->layers
, radv_surface_layer_count(iview
));
3315 *pFramebuffer
= radv_framebuffer_to_handle(framebuffer
);
3319 void radv_DestroyFramebuffer(
3322 const VkAllocationCallbacks
* pAllocator
)
3324 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3325 RADV_FROM_HANDLE(radv_framebuffer
, fb
, _fb
);
3329 vk_free2(&device
->alloc
, pAllocator
, fb
);
3332 static unsigned radv_tex_wrap(VkSamplerAddressMode address_mode
)
3334 switch (address_mode
) {
3335 case VK_SAMPLER_ADDRESS_MODE_REPEAT
:
3336 return V_008F30_SQ_TEX_WRAP
;
3337 case VK_SAMPLER_ADDRESS_MODE_MIRRORED_REPEAT
:
3338 return V_008F30_SQ_TEX_MIRROR
;
3339 case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE
:
3340 return V_008F30_SQ_TEX_CLAMP_LAST_TEXEL
;
3341 case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_BORDER
:
3342 return V_008F30_SQ_TEX_CLAMP_BORDER
;
3343 case VK_SAMPLER_ADDRESS_MODE_MIRROR_CLAMP_TO_EDGE
:
3344 return V_008F30_SQ_TEX_MIRROR_ONCE_LAST_TEXEL
;
3346 unreachable("illegal tex wrap mode");
3352 radv_tex_compare(VkCompareOp op
)
3355 case VK_COMPARE_OP_NEVER
:
3356 return V_008F30_SQ_TEX_DEPTH_COMPARE_NEVER
;
3357 case VK_COMPARE_OP_LESS
:
3358 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESS
;
3359 case VK_COMPARE_OP_EQUAL
:
3360 return V_008F30_SQ_TEX_DEPTH_COMPARE_EQUAL
;
3361 case VK_COMPARE_OP_LESS_OR_EQUAL
:
3362 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESSEQUAL
;
3363 case VK_COMPARE_OP_GREATER
:
3364 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATER
;
3365 case VK_COMPARE_OP_NOT_EQUAL
:
3366 return V_008F30_SQ_TEX_DEPTH_COMPARE_NOTEQUAL
;
3367 case VK_COMPARE_OP_GREATER_OR_EQUAL
:
3368 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATEREQUAL
;
3369 case VK_COMPARE_OP_ALWAYS
:
3370 return V_008F30_SQ_TEX_DEPTH_COMPARE_ALWAYS
;
3372 unreachable("illegal compare mode");
3378 radv_tex_filter(VkFilter filter
, unsigned max_ansio
)
3381 case VK_FILTER_NEAREST
:
3382 return (max_ansio
> 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_POINT
:
3383 V_008F38_SQ_TEX_XY_FILTER_POINT
);
3384 case VK_FILTER_LINEAR
:
3385 return (max_ansio
> 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_BILINEAR
:
3386 V_008F38_SQ_TEX_XY_FILTER_BILINEAR
);
3387 case VK_FILTER_CUBIC_IMG
:
3389 fprintf(stderr
, "illegal texture filter");
3395 radv_tex_mipfilter(VkSamplerMipmapMode mode
)
3398 case VK_SAMPLER_MIPMAP_MODE_NEAREST
:
3399 return V_008F38_SQ_TEX_Z_FILTER_POINT
;
3400 case VK_SAMPLER_MIPMAP_MODE_LINEAR
:
3401 return V_008F38_SQ_TEX_Z_FILTER_LINEAR
;
3403 return V_008F38_SQ_TEX_Z_FILTER_NONE
;
3408 radv_tex_bordercolor(VkBorderColor bcolor
)
3411 case VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
:
3412 case VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
:
3413 return V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK
;
3414 case VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
:
3415 case VK_BORDER_COLOR_INT_OPAQUE_BLACK
:
3416 return V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_BLACK
;
3417 case VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
:
3418 case VK_BORDER_COLOR_INT_OPAQUE_WHITE
:
3419 return V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_WHITE
;
3427 radv_tex_aniso_filter(unsigned filter
)
3441 radv_init_sampler(struct radv_device
*device
,
3442 struct radv_sampler
*sampler
,
3443 const VkSamplerCreateInfo
*pCreateInfo
)
3445 uint32_t max_aniso
= pCreateInfo
->anisotropyEnable
&& pCreateInfo
->maxAnisotropy
> 1.0 ?
3446 (uint32_t) pCreateInfo
->maxAnisotropy
: 0;
3447 uint32_t max_aniso_ratio
= radv_tex_aniso_filter(max_aniso
);
3448 bool is_vi
= (device
->physical_device
->rad_info
.chip_class
>= VI
);
3450 sampler
->state
[0] = (S_008F30_CLAMP_X(radv_tex_wrap(pCreateInfo
->addressModeU
)) |
3451 S_008F30_CLAMP_Y(radv_tex_wrap(pCreateInfo
->addressModeV
)) |
3452 S_008F30_CLAMP_Z(radv_tex_wrap(pCreateInfo
->addressModeW
)) |
3453 S_008F30_MAX_ANISO_RATIO(max_aniso_ratio
) |
3454 S_008F30_DEPTH_COMPARE_FUNC(radv_tex_compare(pCreateInfo
->compareOp
)) |
3455 S_008F30_FORCE_UNNORMALIZED(pCreateInfo
->unnormalizedCoordinates
? 1 : 0) |
3456 S_008F30_ANISO_THRESHOLD(max_aniso_ratio
>> 1) |
3457 S_008F30_ANISO_BIAS(max_aniso_ratio
) |
3458 S_008F30_DISABLE_CUBE_WRAP(0) |
3459 S_008F30_COMPAT_MODE(is_vi
));
3460 sampler
->state
[1] = (S_008F34_MIN_LOD(S_FIXED(CLAMP(pCreateInfo
->minLod
, 0, 15), 8)) |
3461 S_008F34_MAX_LOD(S_FIXED(CLAMP(pCreateInfo
->maxLod
, 0, 15), 8)) |
3462 S_008F34_PERF_MIP(max_aniso_ratio
? max_aniso_ratio
+ 6 : 0));
3463 sampler
->state
[2] = (S_008F38_LOD_BIAS(S_FIXED(CLAMP(pCreateInfo
->mipLodBias
, -16, 16), 8)) |
3464 S_008F38_XY_MAG_FILTER(radv_tex_filter(pCreateInfo
->magFilter
, max_aniso
)) |
3465 S_008F38_XY_MIN_FILTER(radv_tex_filter(pCreateInfo
->minFilter
, max_aniso
)) |
3466 S_008F38_MIP_FILTER(radv_tex_mipfilter(pCreateInfo
->mipmapMode
)) |
3467 S_008F38_MIP_POINT_PRECLAMP(0) |
3468 S_008F38_DISABLE_LSB_CEIL(1) |
3469 S_008F38_FILTER_PREC_FIX(1) |
3470 S_008F38_ANISO_OVERRIDE(is_vi
));
3471 sampler
->state
[3] = (S_008F3C_BORDER_COLOR_PTR(0) |
3472 S_008F3C_BORDER_COLOR_TYPE(radv_tex_bordercolor(pCreateInfo
->borderColor
)));
3475 VkResult
radv_CreateSampler(
3477 const VkSamplerCreateInfo
* pCreateInfo
,
3478 const VkAllocationCallbacks
* pAllocator
,
3479 VkSampler
* pSampler
)
3481 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3482 struct radv_sampler
*sampler
;
3484 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO
);
3486 sampler
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*sampler
), 8,
3487 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
3489 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
3491 radv_init_sampler(device
, sampler
, pCreateInfo
);
3492 *pSampler
= radv_sampler_to_handle(sampler
);
3497 void radv_DestroySampler(
3500 const VkAllocationCallbacks
* pAllocator
)
3502 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3503 RADV_FROM_HANDLE(radv_sampler
, sampler
, _sampler
);
3507 vk_free2(&device
->alloc
, pAllocator
, sampler
);
3510 /* vk_icd.h does not declare this function, so we declare it here to
3511 * suppress Wmissing-prototypes.
3513 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
3514 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t *pSupportedVersion
);
3516 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
3517 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t *pSupportedVersion
)
3519 /* For the full details on loader interface versioning, see
3520 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
3521 * What follows is a condensed summary, to help you navigate the large and
3522 * confusing official doc.
3524 * - Loader interface v0 is incompatible with later versions. We don't
3527 * - In loader interface v1:
3528 * - The first ICD entrypoint called by the loader is
3529 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
3531 * - The ICD must statically expose no other Vulkan symbol unless it is
3532 * linked with -Bsymbolic.
3533 * - Each dispatchable Vulkan handle created by the ICD must be
3534 * a pointer to a struct whose first member is VK_LOADER_DATA. The
3535 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
3536 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
3537 * vkDestroySurfaceKHR(). The ICD must be capable of working with
3538 * such loader-managed surfaces.
3540 * - Loader interface v2 differs from v1 in:
3541 * - The first ICD entrypoint called by the loader is
3542 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
3543 * statically expose this entrypoint.
3545 * - Loader interface v3 differs from v2 in:
3546 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
3547 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
3548 * because the loader no longer does so.
3550 *pSupportedVersion
= MIN2(*pSupportedVersion
, 3u);
3554 VkResult
radv_GetMemoryFdKHR(VkDevice _device
,
3555 const VkMemoryGetFdInfoKHR
*pGetFdInfo
,
3558 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3559 RADV_FROM_HANDLE(radv_device_memory
, memory
, pGetFdInfo
->memory
);
3561 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
3563 /* We support only one handle type. */
3564 assert(pGetFdInfo
->handleType
==
3565 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
);
3567 bool ret
= radv_get_memory_fd(device
, memory
, pFD
);
3569 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
3573 VkResult
radv_GetMemoryFdPropertiesKHR(VkDevice _device
,
3574 VkExternalMemoryHandleTypeFlagBitsKHR handleType
,
3576 VkMemoryFdPropertiesKHR
*pMemoryFdProperties
)
3578 /* The valid usage section for this function says:
3580 * "handleType must not be one of the handle types defined as opaque."
3582 * Since we only handle opaque handles for now, there are no FD properties.
3584 return VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
;
3587 VkResult
radv_ImportSemaphoreFdKHR(VkDevice _device
,
3588 const VkImportSemaphoreFdInfoKHR
*pImportSemaphoreFdInfo
)
3590 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3591 RADV_FROM_HANDLE(radv_semaphore
, sem
, pImportSemaphoreFdInfo
->semaphore
);
3592 uint32_t syncobj_handle
= 0;
3593 assert(pImportSemaphoreFdInfo
->handleType
== VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
);
3595 int ret
= device
->ws
->import_syncobj(device
->ws
, pImportSemaphoreFdInfo
->fd
, &syncobj_handle
);
3597 return VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
;
3599 if (pImportSemaphoreFdInfo
->flags
& VK_SEMAPHORE_IMPORT_TEMPORARY_BIT_KHR
) {
3600 sem
->temp_syncobj
= syncobj_handle
;
3602 sem
->syncobj
= syncobj_handle
;
3604 close(pImportSemaphoreFdInfo
->fd
);
3608 VkResult
radv_GetSemaphoreFdKHR(VkDevice _device
,
3609 const VkSemaphoreGetFdInfoKHR
*pGetFdInfo
,
3612 RADV_FROM_HANDLE(radv_device
, device
, _device
);
3613 RADV_FROM_HANDLE(radv_semaphore
, sem
, pGetFdInfo
->semaphore
);
3615 uint32_t syncobj_handle
;
3617 assert(pGetFdInfo
->handleType
== VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
);
3618 if (sem
->temp_syncobj
)
3619 syncobj_handle
= sem
->temp_syncobj
;
3621 syncobj_handle
= sem
->syncobj
;
3622 ret
= device
->ws
->export_syncobj(device
->ws
, syncobj_handle
, pFd
);
3624 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
3628 void radv_GetPhysicalDeviceExternalSemaphorePropertiesKHR(
3629 VkPhysicalDevice physicalDevice
,
3630 const VkPhysicalDeviceExternalSemaphoreInfoKHR
* pExternalSemaphoreInfo
,
3631 VkExternalSemaphorePropertiesKHR
* pExternalSemaphoreProperties
)
3633 if (pExternalSemaphoreInfo
->handleType
== VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
) {
3634 pExternalSemaphoreProperties
->exportFromImportedHandleTypes
= VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
;
3635 pExternalSemaphoreProperties
->compatibleHandleTypes
= VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
;
3636 pExternalSemaphoreProperties
->externalSemaphoreFeatures
= VK_EXTERNAL_SEMAPHORE_FEATURE_EXPORTABLE_BIT_KHR
|
3637 VK_EXTERNAL_SEMAPHORE_FEATURE_IMPORTABLE_BIT_KHR
;
3639 pExternalSemaphoreProperties
->exportFromImportedHandleTypes
= 0;
3640 pExternalSemaphoreProperties
->compatibleHandleTypes
= 0;
3641 pExternalSemaphoreProperties
->externalSemaphoreFeatures
= 0;