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_private.h"
34 #include "util/disk_cache.h"
35 #include "util/strtod.h"
36 #include "util/vk_util.h"
39 #include <amdgpu_drm.h>
40 #include "amdgpu_id.h"
41 #include "winsys/amdgpu/radv_amdgpu_winsys_public.h"
42 #include "ac_llvm_util.h"
43 #include "vk_format.h"
45 #include "util/debug.h"
48 radv_device_get_cache_uuid(enum radeon_family family
, void *uuid
)
50 uint32_t mesa_timestamp
, llvm_timestamp
;
52 memset(uuid
, 0, VK_UUID_SIZE
);
53 if (!disk_cache_get_function_timestamp(radv_device_get_cache_uuid
, &mesa_timestamp
) ||
54 !disk_cache_get_function_timestamp(LLVMInitializeAMDGPUTargetInfo
, &llvm_timestamp
))
57 memcpy(uuid
, &mesa_timestamp
, 4);
58 memcpy((char*)uuid
+ 4, &llvm_timestamp
, 4);
59 memcpy((char*)uuid
+ 8, &f
, 2);
60 snprintf((char*)uuid
+ 10, VK_UUID_SIZE
- 10, "radv");
64 static const VkExtensionProperties instance_extensions
[] = {
66 .extensionName
= VK_KHR_SURFACE_EXTENSION_NAME
,
69 #ifdef VK_USE_PLATFORM_XCB_KHR
71 .extensionName
= VK_KHR_XCB_SURFACE_EXTENSION_NAME
,
75 #ifdef VK_USE_PLATFORM_XLIB_KHR
77 .extensionName
= VK_KHR_XLIB_SURFACE_EXTENSION_NAME
,
81 #ifdef VK_USE_PLATFORM_WAYLAND_KHR
83 .extensionName
= VK_KHR_WAYLAND_SURFACE_EXTENSION_NAME
,
89 static const VkExtensionProperties common_device_extensions
[] = {
91 .extensionName
= VK_KHR_MAINTENANCE1_EXTENSION_NAME
,
95 .extensionName
= VK_KHR_SAMPLER_MIRROR_CLAMP_TO_EDGE_EXTENSION_NAME
,
99 .extensionName
= VK_KHR_SWAPCHAIN_EXTENSION_NAME
,
103 .extensionName
= VK_AMD_DRAW_INDIRECT_COUNT_EXTENSION_NAME
,
107 .extensionName
= VK_KHR_GET_PHYSICAL_DEVICE_PROPERTIES_2_EXTENSION_NAME
,
111 .extensionName
= VK_KHR_SHADER_DRAW_PARAMETERS_EXTENSION_NAME
,
115 .extensionName
= VK_NV_DEDICATED_ALLOCATION_EXTENSION_NAME
,
121 radv_extensions_register(struct radv_instance
*instance
,
122 struct radv_extensions
*extensions
,
123 const VkExtensionProperties
*new_ext
,
127 VkExtensionProperties
*new_ptr
;
129 assert(new_ext
&& num_ext
> 0);
132 return VK_ERROR_INITIALIZATION_FAILED
;
134 new_size
= (extensions
->num_ext
+ num_ext
) * sizeof(VkExtensionProperties
);
135 new_ptr
= vk_realloc(&instance
->alloc
, extensions
->ext_array
,
136 new_size
, 8, VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
138 /* Old array continues to be valid, update nothing */
140 return VK_ERROR_OUT_OF_HOST_MEMORY
;
142 memcpy(&new_ptr
[extensions
->num_ext
], new_ext
,
143 num_ext
* sizeof(VkExtensionProperties
));
144 extensions
->ext_array
= new_ptr
;
145 extensions
->num_ext
+= num_ext
;
151 radv_extensions_finish(struct radv_instance
*instance
,
152 struct radv_extensions
*extensions
)
157 radv_loge("Attemted to free invalid extension struct\n");
159 if (extensions
->ext_array
)
160 vk_free(&instance
->alloc
, extensions
->ext_array
);
164 is_extension_enabled(const VkExtensionProperties
*extensions
,
168 assert(extensions
&& name
);
170 for (uint32_t i
= 0; i
< num_ext
; i
++) {
171 if (strcmp(name
, extensions
[i
].extensionName
) == 0)
179 radv_physical_device_init(struct radv_physical_device
*device
,
180 struct radv_instance
*instance
,
184 drmVersionPtr version
;
187 fd
= open(path
, O_RDWR
| O_CLOEXEC
);
189 return VK_ERROR_INCOMPATIBLE_DRIVER
;
191 version
= drmGetVersion(fd
);
194 return vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER
,
195 "failed to get version %s: %m", path
);
198 if (strcmp(version
->name
, "amdgpu")) {
199 drmFreeVersion(version
);
201 return VK_ERROR_INCOMPATIBLE_DRIVER
;
203 drmFreeVersion(version
);
205 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
206 device
->instance
= instance
;
207 assert(strlen(path
) < ARRAY_SIZE(device
->path
));
208 strncpy(device
->path
, path
, ARRAY_SIZE(device
->path
));
210 device
->ws
= radv_amdgpu_winsys_create(fd
, instance
->debug_flags
);
212 result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
216 device
->local_fd
= fd
;
217 device
->ws
->query_info(device
->ws
, &device
->rad_info
);
218 result
= radv_init_wsi(device
);
219 if (result
!= VK_SUCCESS
) {
220 device
->ws
->destroy(device
->ws
);
224 if (radv_device_get_cache_uuid(device
->rad_info
.family
, device
->uuid
)) {
225 radv_finish_wsi(device
);
226 device
->ws
->destroy(device
->ws
);
227 result
= vk_errorf(VK_ERROR_INITIALIZATION_FAILED
,
228 "cannot generate UUID");
232 result
= radv_extensions_register(instance
,
234 common_device_extensions
,
235 ARRAY_SIZE(common_device_extensions
));
236 if (result
!= VK_SUCCESS
)
239 fprintf(stderr
, "WARNING: radv is not a conformant vulkan implementation, testing use only.\n");
240 device
->name
= device
->rad_info
.name
;
250 radv_physical_device_finish(struct radv_physical_device
*device
)
252 radv_extensions_finish(device
->instance
, &device
->extensions
);
253 radv_finish_wsi(device
);
254 device
->ws
->destroy(device
->ws
);
255 close(device
->local_fd
);
260 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
261 VkSystemAllocationScope allocationScope
)
267 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
268 size_t align
, VkSystemAllocationScope allocationScope
)
270 return realloc(pOriginal
, size
);
274 default_free_func(void *pUserData
, void *pMemory
)
279 static const VkAllocationCallbacks default_alloc
= {
281 .pfnAllocation
= default_alloc_func
,
282 .pfnReallocation
= default_realloc_func
,
283 .pfnFree
= default_free_func
,
286 static const struct debug_control radv_debug_options
[] = {
287 {"nofastclears", RADV_DEBUG_NO_FAST_CLEARS
},
288 {"nodcc", RADV_DEBUG_NO_DCC
},
289 {"shaders", RADV_DEBUG_DUMP_SHADERS
},
290 {"nocache", RADV_DEBUG_NO_CACHE
},
291 {"shaderstats", RADV_DEBUG_DUMP_SHADER_STATS
},
292 {"nohiz", RADV_DEBUG_NO_HIZ
},
293 {"nocompute", RADV_DEBUG_NO_COMPUTE_QUEUE
},
294 {"unsafemath", RADV_DEBUG_UNSAFE_MATH
},
295 {"allbos", RADV_DEBUG_ALL_BOS
},
296 {"noibs", RADV_DEBUG_NO_IBS
},
300 VkResult
radv_CreateInstance(
301 const VkInstanceCreateInfo
* pCreateInfo
,
302 const VkAllocationCallbacks
* pAllocator
,
303 VkInstance
* pInstance
)
305 struct radv_instance
*instance
;
307 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
309 uint32_t client_version
;
310 if (pCreateInfo
->pApplicationInfo
&&
311 pCreateInfo
->pApplicationInfo
->apiVersion
!= 0) {
312 client_version
= pCreateInfo
->pApplicationInfo
->apiVersion
;
314 client_version
= VK_MAKE_VERSION(1, 0, 0);
317 if (VK_MAKE_VERSION(1, 0, 0) > client_version
||
318 client_version
> VK_MAKE_VERSION(1, 0, 0xfff)) {
319 return vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER
,
320 "Client requested version %d.%d.%d",
321 VK_VERSION_MAJOR(client_version
),
322 VK_VERSION_MINOR(client_version
),
323 VK_VERSION_PATCH(client_version
));
326 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
327 if (!is_extension_enabled(instance_extensions
,
328 ARRAY_SIZE(instance_extensions
),
329 pCreateInfo
->ppEnabledExtensionNames
[i
]))
330 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
333 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
334 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
336 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
338 memset(instance
, 0, sizeof(*instance
));
340 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
343 instance
->alloc
= *pAllocator
;
345 instance
->alloc
= default_alloc
;
347 instance
->apiVersion
= client_version
;
348 instance
->physicalDeviceCount
= -1;
352 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
354 instance
->debug_flags
= parse_debug_string(getenv("RADV_DEBUG"),
357 *pInstance
= radv_instance_to_handle(instance
);
362 void radv_DestroyInstance(
363 VkInstance _instance
,
364 const VkAllocationCallbacks
* pAllocator
)
366 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
371 for (int i
= 0; i
< instance
->physicalDeviceCount
; ++i
) {
372 radv_physical_device_finish(instance
->physicalDevices
+ i
);
375 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
379 vk_free(&instance
->alloc
, instance
);
382 VkResult
radv_EnumeratePhysicalDevices(
383 VkInstance _instance
,
384 uint32_t* pPhysicalDeviceCount
,
385 VkPhysicalDevice
* pPhysicalDevices
)
387 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
390 if (instance
->physicalDeviceCount
< 0) {
392 instance
->physicalDeviceCount
= 0;
393 for (unsigned i
= 0; i
< RADV_MAX_DRM_DEVICES
; i
++) {
394 snprintf(path
, sizeof(path
), "/dev/dri/renderD%d", 128 + i
);
395 result
= radv_physical_device_init(instance
->physicalDevices
+
396 instance
->physicalDeviceCount
,
398 if (result
== VK_SUCCESS
)
399 ++instance
->physicalDeviceCount
;
400 else if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
405 if (!pPhysicalDevices
) {
406 *pPhysicalDeviceCount
= instance
->physicalDeviceCount
;
408 *pPhysicalDeviceCount
= MIN2(*pPhysicalDeviceCount
, instance
->physicalDeviceCount
);
409 for (unsigned i
= 0; i
< *pPhysicalDeviceCount
; ++i
)
410 pPhysicalDevices
[i
] = radv_physical_device_to_handle(instance
->physicalDevices
+ i
);
413 return *pPhysicalDeviceCount
< instance
->physicalDeviceCount
? VK_INCOMPLETE
417 void radv_GetPhysicalDeviceFeatures(
418 VkPhysicalDevice physicalDevice
,
419 VkPhysicalDeviceFeatures
* pFeatures
)
421 // RADV_FROM_HANDLE(radv_physical_device, pdevice, physicalDevice);
423 memset(pFeatures
, 0, sizeof(*pFeatures
));
425 *pFeatures
= (VkPhysicalDeviceFeatures
) {
426 .robustBufferAccess
= true,
427 .fullDrawIndexUint32
= true,
428 .imageCubeArray
= true,
429 .independentBlend
= true,
430 .geometryShader
= true,
431 .tessellationShader
= false,
432 .sampleRateShading
= false,
433 .dualSrcBlend
= true,
435 .multiDrawIndirect
= true,
436 .drawIndirectFirstInstance
= true,
438 .depthBiasClamp
= true,
439 .fillModeNonSolid
= true,
444 .multiViewport
= true,
445 .samplerAnisotropy
= true,
446 .textureCompressionETC2
= false,
447 .textureCompressionASTC_LDR
= false,
448 .textureCompressionBC
= true,
449 .occlusionQueryPrecise
= true,
450 .pipelineStatisticsQuery
= false,
451 .vertexPipelineStoresAndAtomics
= true,
452 .fragmentStoresAndAtomics
= true,
453 .shaderTessellationAndGeometryPointSize
= true,
454 .shaderImageGatherExtended
= true,
455 .shaderStorageImageExtendedFormats
= true,
456 .shaderStorageImageMultisample
= false,
457 .shaderUniformBufferArrayDynamicIndexing
= true,
458 .shaderSampledImageArrayDynamicIndexing
= true,
459 .shaderStorageBufferArrayDynamicIndexing
= true,
460 .shaderStorageImageArrayDynamicIndexing
= true,
461 .shaderStorageImageReadWithoutFormat
= true,
462 .shaderStorageImageWriteWithoutFormat
= true,
463 .shaderClipDistance
= true,
464 .shaderCullDistance
= true,
465 .shaderFloat64
= true,
466 .shaderInt64
= false,
467 .shaderInt16
= false,
468 .variableMultisampleRate
= false,
469 .inheritedQueries
= false,
473 void radv_GetPhysicalDeviceFeatures2KHR(
474 VkPhysicalDevice physicalDevice
,
475 VkPhysicalDeviceFeatures2KHR
*pFeatures
)
477 return radv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
480 void radv_GetPhysicalDeviceProperties(
481 VkPhysicalDevice physicalDevice
,
482 VkPhysicalDeviceProperties
* pProperties
)
484 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
485 VkSampleCountFlags sample_counts
= 0xf;
486 VkPhysicalDeviceLimits limits
= {
487 .maxImageDimension1D
= (1 << 14),
488 .maxImageDimension2D
= (1 << 14),
489 .maxImageDimension3D
= (1 << 11),
490 .maxImageDimensionCube
= (1 << 14),
491 .maxImageArrayLayers
= (1 << 11),
492 .maxTexelBufferElements
= 128 * 1024 * 1024,
493 .maxUniformBufferRange
= UINT32_MAX
,
494 .maxStorageBufferRange
= UINT32_MAX
,
495 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
496 .maxMemoryAllocationCount
= UINT32_MAX
,
497 .maxSamplerAllocationCount
= 64 * 1024,
498 .bufferImageGranularity
= 64, /* A cache line */
499 .sparseAddressSpaceSize
= 0,
500 .maxBoundDescriptorSets
= MAX_SETS
,
501 .maxPerStageDescriptorSamplers
= 64,
502 .maxPerStageDescriptorUniformBuffers
= 64,
503 .maxPerStageDescriptorStorageBuffers
= 64,
504 .maxPerStageDescriptorSampledImages
= 64,
505 .maxPerStageDescriptorStorageImages
= 64,
506 .maxPerStageDescriptorInputAttachments
= 64,
507 .maxPerStageResources
= 128,
508 .maxDescriptorSetSamplers
= 256,
509 .maxDescriptorSetUniformBuffers
= 256,
510 .maxDescriptorSetUniformBuffersDynamic
= 256,
511 .maxDescriptorSetStorageBuffers
= 256,
512 .maxDescriptorSetStorageBuffersDynamic
= 256,
513 .maxDescriptorSetSampledImages
= 256,
514 .maxDescriptorSetStorageImages
= 256,
515 .maxDescriptorSetInputAttachments
= 256,
516 .maxVertexInputAttributes
= 32,
517 .maxVertexInputBindings
= 32,
518 .maxVertexInputAttributeOffset
= 2047,
519 .maxVertexInputBindingStride
= 2048,
520 .maxVertexOutputComponents
= 128,
521 .maxTessellationGenerationLevel
= 0,
522 .maxTessellationPatchSize
= 0,
523 .maxTessellationControlPerVertexInputComponents
= 0,
524 .maxTessellationControlPerVertexOutputComponents
= 0,
525 .maxTessellationControlPerPatchOutputComponents
= 0,
526 .maxTessellationControlTotalOutputComponents
= 0,
527 .maxTessellationEvaluationInputComponents
= 0,
528 .maxTessellationEvaluationOutputComponents
= 0,
529 .maxGeometryShaderInvocations
= 32,
530 .maxGeometryInputComponents
= 64,
531 .maxGeometryOutputComponents
= 128,
532 .maxGeometryOutputVertices
= 256,
533 .maxGeometryTotalOutputComponents
= 1024,
534 .maxFragmentInputComponents
= 128,
535 .maxFragmentOutputAttachments
= 8,
536 .maxFragmentDualSrcAttachments
= 1,
537 .maxFragmentCombinedOutputResources
= 8,
538 .maxComputeSharedMemorySize
= 32768,
539 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
540 .maxComputeWorkGroupInvocations
= 2048,
541 .maxComputeWorkGroupSize
= {
546 .subPixelPrecisionBits
= 4 /* FIXME */,
547 .subTexelPrecisionBits
= 4 /* FIXME */,
548 .mipmapPrecisionBits
= 4 /* FIXME */,
549 .maxDrawIndexedIndexValue
= UINT32_MAX
,
550 .maxDrawIndirectCount
= UINT32_MAX
,
551 .maxSamplerLodBias
= 16,
552 .maxSamplerAnisotropy
= 16,
553 .maxViewports
= MAX_VIEWPORTS
,
554 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
555 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
556 .viewportSubPixelBits
= 13, /* We take a float? */
557 .minMemoryMapAlignment
= 4096, /* A page */
558 .minTexelBufferOffsetAlignment
= 1,
559 .minUniformBufferOffsetAlignment
= 4,
560 .minStorageBufferOffsetAlignment
= 4,
561 .minTexelOffset
= -32,
562 .maxTexelOffset
= 31,
563 .minTexelGatherOffset
= -32,
564 .maxTexelGatherOffset
= 31,
565 .minInterpolationOffset
= -2,
566 .maxInterpolationOffset
= 2,
567 .subPixelInterpolationOffsetBits
= 8,
568 .maxFramebufferWidth
= (1 << 14),
569 .maxFramebufferHeight
= (1 << 14),
570 .maxFramebufferLayers
= (1 << 10),
571 .framebufferColorSampleCounts
= sample_counts
,
572 .framebufferDepthSampleCounts
= sample_counts
,
573 .framebufferStencilSampleCounts
= sample_counts
,
574 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
575 .maxColorAttachments
= MAX_RTS
,
576 .sampledImageColorSampleCounts
= sample_counts
,
577 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
578 .sampledImageDepthSampleCounts
= sample_counts
,
579 .sampledImageStencilSampleCounts
= sample_counts
,
580 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
581 .maxSampleMaskWords
= 1,
582 .timestampComputeAndGraphics
= false,
583 .timestampPeriod
= 100000.0 / pdevice
->rad_info
.clock_crystal_freq
,
584 .maxClipDistances
= 8,
585 .maxCullDistances
= 8,
586 .maxCombinedClipAndCullDistances
= 8,
587 .discreteQueuePriorities
= 1,
588 .pointSizeRange
= { 0.125, 255.875 },
589 .lineWidthRange
= { 0.0, 7.9921875 },
590 .pointSizeGranularity
= (1.0 / 8.0),
591 .lineWidthGranularity
= (1.0 / 128.0),
592 .strictLines
= false, /* FINISHME */
593 .standardSampleLocations
= true,
594 .optimalBufferCopyOffsetAlignment
= 128,
595 .optimalBufferCopyRowPitchAlignment
= 128,
596 .nonCoherentAtomSize
= 64,
599 *pProperties
= (VkPhysicalDeviceProperties
) {
600 .apiVersion
= VK_MAKE_VERSION(1, 0, 42),
603 .deviceID
= pdevice
->rad_info
.pci_id
,
604 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU
,
606 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
609 strcpy(pProperties
->deviceName
, pdevice
->name
);
610 memcpy(pProperties
->pipelineCacheUUID
, pdevice
->uuid
, VK_UUID_SIZE
);
613 void radv_GetPhysicalDeviceProperties2KHR(
614 VkPhysicalDevice physicalDevice
,
615 VkPhysicalDeviceProperties2KHR
*pProperties
)
617 return radv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
620 static void radv_get_physical_device_queue_family_properties(
621 struct radv_physical_device
* pdevice
,
623 VkQueueFamilyProperties
** pQueueFamilyProperties
)
625 int num_queue_families
= 1;
627 if (pdevice
->rad_info
.compute_rings
> 0 &&
628 pdevice
->rad_info
.chip_class
>= CIK
&&
629 !(pdevice
->instance
->debug_flags
& RADV_DEBUG_NO_COMPUTE_QUEUE
))
630 num_queue_families
++;
632 if (pQueueFamilyProperties
== NULL
) {
633 *pCount
= num_queue_families
;
642 *pQueueFamilyProperties
[idx
] = (VkQueueFamilyProperties
) {
643 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
644 VK_QUEUE_COMPUTE_BIT
|
645 VK_QUEUE_TRANSFER_BIT
,
647 .timestampValidBits
= 64,
648 .minImageTransferGranularity
= (VkExtent3D
) { 1, 1, 1 },
653 if (pdevice
->rad_info
.compute_rings
> 0 &&
654 pdevice
->rad_info
.chip_class
>= CIK
&&
655 !(pdevice
->instance
->debug_flags
& RADV_DEBUG_NO_COMPUTE_QUEUE
)) {
657 *pQueueFamilyProperties
[idx
] = (VkQueueFamilyProperties
) {
658 .queueFlags
= VK_QUEUE_COMPUTE_BIT
| VK_QUEUE_TRANSFER_BIT
,
659 .queueCount
= pdevice
->rad_info
.compute_rings
,
660 .timestampValidBits
= 64,
661 .minImageTransferGranularity
= (VkExtent3D
) { 1, 1, 1 },
669 void radv_GetPhysicalDeviceQueueFamilyProperties(
670 VkPhysicalDevice physicalDevice
,
672 VkQueueFamilyProperties
* pQueueFamilyProperties
)
674 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
675 if (!pQueueFamilyProperties
) {
676 return radv_get_physical_device_queue_family_properties(pdevice
, pCount
, NULL
);
679 VkQueueFamilyProperties
*properties
[] = {
680 pQueueFamilyProperties
+ 0,
681 pQueueFamilyProperties
+ 1,
682 pQueueFamilyProperties
+ 2,
684 radv_get_physical_device_queue_family_properties(pdevice
, pCount
, properties
);
685 assert(*pCount
<= 3);
688 void radv_GetPhysicalDeviceQueueFamilyProperties2KHR(
689 VkPhysicalDevice physicalDevice
,
691 VkQueueFamilyProperties2KHR
*pQueueFamilyProperties
)
693 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
694 if (!pQueueFamilyProperties
) {
695 return radv_get_physical_device_queue_family_properties(pdevice
, pCount
, NULL
);
698 VkQueueFamilyProperties
*properties
[] = {
699 &pQueueFamilyProperties
[0].queueFamilyProperties
,
700 &pQueueFamilyProperties
[1].queueFamilyProperties
,
701 &pQueueFamilyProperties
[2].queueFamilyProperties
,
703 radv_get_physical_device_queue_family_properties(pdevice
, pCount
, properties
);
704 assert(*pCount
<= 3);
707 void radv_GetPhysicalDeviceMemoryProperties(
708 VkPhysicalDevice physicalDevice
,
709 VkPhysicalDeviceMemoryProperties
*pMemoryProperties
)
711 RADV_FROM_HANDLE(radv_physical_device
, physical_device
, physicalDevice
);
713 STATIC_ASSERT(RADV_MEM_TYPE_COUNT
<= VK_MAX_MEMORY_TYPES
);
715 pMemoryProperties
->memoryTypeCount
= RADV_MEM_TYPE_COUNT
;
716 pMemoryProperties
->memoryTypes
[RADV_MEM_TYPE_VRAM
] = (VkMemoryType
) {
717 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
,
718 .heapIndex
= RADV_MEM_HEAP_VRAM
,
720 pMemoryProperties
->memoryTypes
[RADV_MEM_TYPE_GTT_WRITE_COMBINE
] = (VkMemoryType
) {
721 .propertyFlags
= VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
722 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
723 .heapIndex
= RADV_MEM_HEAP_GTT
,
725 pMemoryProperties
->memoryTypes
[RADV_MEM_TYPE_VRAM_CPU_ACCESS
] = (VkMemoryType
) {
726 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
727 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
728 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
729 .heapIndex
= RADV_MEM_HEAP_VRAM_CPU_ACCESS
,
731 pMemoryProperties
->memoryTypes
[RADV_MEM_TYPE_GTT_CACHED
] = (VkMemoryType
) {
732 .propertyFlags
= VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
733 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
734 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
735 .heapIndex
= RADV_MEM_HEAP_GTT
,
738 STATIC_ASSERT(RADV_MEM_HEAP_COUNT
<= VK_MAX_MEMORY_HEAPS
);
740 pMemoryProperties
->memoryHeapCount
= RADV_MEM_HEAP_COUNT
;
741 pMemoryProperties
->memoryHeaps
[RADV_MEM_HEAP_VRAM
] = (VkMemoryHeap
) {
742 .size
= physical_device
->rad_info
.vram_size
-
743 physical_device
->rad_info
.visible_vram_size
,
744 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
746 pMemoryProperties
->memoryHeaps
[RADV_MEM_HEAP_VRAM_CPU_ACCESS
] = (VkMemoryHeap
) {
747 .size
= physical_device
->rad_info
.visible_vram_size
,
748 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
750 pMemoryProperties
->memoryHeaps
[RADV_MEM_HEAP_GTT
] = (VkMemoryHeap
) {
751 .size
= physical_device
->rad_info
.gart_size
,
756 void radv_GetPhysicalDeviceMemoryProperties2KHR(
757 VkPhysicalDevice physicalDevice
,
758 VkPhysicalDeviceMemoryProperties2KHR
*pMemoryProperties
)
760 return radv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
761 &pMemoryProperties
->memoryProperties
);
765 radv_queue_init(struct radv_device
*device
, struct radv_queue
*queue
,
766 int queue_family_index
, int idx
)
768 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
769 queue
->device
= device
;
770 queue
->queue_family_index
= queue_family_index
;
771 queue
->queue_idx
= idx
;
773 queue
->hw_ctx
= device
->ws
->ctx_create(device
->ws
);
775 return VK_ERROR_OUT_OF_HOST_MEMORY
;
781 radv_queue_finish(struct radv_queue
*queue
)
784 queue
->device
->ws
->ctx_destroy(queue
->hw_ctx
);
786 if (queue
->initial_preamble_cs
)
787 queue
->device
->ws
->cs_destroy(queue
->initial_preamble_cs
);
788 if (queue
->continue_preamble_cs
)
789 queue
->device
->ws
->cs_destroy(queue
->continue_preamble_cs
);
790 if (queue
->descriptor_bo
)
791 queue
->device
->ws
->buffer_destroy(queue
->descriptor_bo
);
792 if (queue
->scratch_bo
)
793 queue
->device
->ws
->buffer_destroy(queue
->scratch_bo
);
794 if (queue
->esgs_ring_bo
)
795 queue
->device
->ws
->buffer_destroy(queue
->esgs_ring_bo
);
796 if (queue
->gsvs_ring_bo
)
797 queue
->device
->ws
->buffer_destroy(queue
->gsvs_ring_bo
);
798 if (queue
->compute_scratch_bo
)
799 queue
->device
->ws
->buffer_destroy(queue
->compute_scratch_bo
);
803 radv_device_init_gs_info(struct radv_device
*device
)
805 switch (device
->physical_device
->rad_info
.family
) {
814 device
->gs_table_depth
= 16;
825 device
->gs_table_depth
= 32;
828 unreachable("unknown GPU");
832 VkResult
radv_CreateDevice(
833 VkPhysicalDevice physicalDevice
,
834 const VkDeviceCreateInfo
* pCreateInfo
,
835 const VkAllocationCallbacks
* pAllocator
,
838 RADV_FROM_HANDLE(radv_physical_device
, physical_device
, physicalDevice
);
840 struct radv_device
*device
;
842 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
843 if (!is_extension_enabled(physical_device
->extensions
.ext_array
,
844 physical_device
->extensions
.num_ext
,
845 pCreateInfo
->ppEnabledExtensionNames
[i
]))
846 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
849 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
851 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
853 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
855 memset(device
, 0, sizeof(*device
));
857 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
858 device
->instance
= physical_device
->instance
;
859 device
->physical_device
= physical_device
;
861 device
->debug_flags
= device
->instance
->debug_flags
;
863 device
->ws
= physical_device
->ws
;
865 device
->alloc
= *pAllocator
;
867 device
->alloc
= physical_device
->instance
->alloc
;
869 for (unsigned i
= 0; i
< pCreateInfo
->queueCreateInfoCount
; i
++) {
870 const VkDeviceQueueCreateInfo
*queue_create
= &pCreateInfo
->pQueueCreateInfos
[i
];
871 uint32_t qfi
= queue_create
->queueFamilyIndex
;
873 device
->queues
[qfi
] = vk_alloc(&device
->alloc
,
874 queue_create
->queueCount
* sizeof(struct radv_queue
), 8, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
875 if (!device
->queues
[qfi
]) {
876 result
= VK_ERROR_OUT_OF_HOST_MEMORY
;
880 memset(device
->queues
[qfi
], 0, queue_create
->queueCount
* sizeof(struct radv_queue
));
882 device
->queue_count
[qfi
] = queue_create
->queueCount
;
884 for (unsigned q
= 0; q
< queue_create
->queueCount
; q
++) {
885 result
= radv_queue_init(device
, &device
->queues
[qfi
][q
], qfi
, q
);
886 if (result
!= VK_SUCCESS
)
891 #if HAVE_LLVM < 0x0400
892 device
->llvm_supports_spill
= false;
894 device
->llvm_supports_spill
= true;
897 /* The maximum number of scratch waves. Scratch space isn't divided
898 * evenly between CUs. The number is only a function of the number of CUs.
899 * We can decrease the constant to decrease the scratch buffer size.
901 * sctx->scratch_waves must be >= the maximum posible size of
902 * 1 threadgroup, so that the hw doesn't hang from being unable
905 * The recommended value is 4 per CU at most. Higher numbers don't
906 * bring much benefit, but they still occupy chip resources (think
907 * async compute). I've seen ~2% performance difference between 4 and 32.
909 uint32_t max_threads_per_block
= 2048;
910 device
->scratch_waves
= MAX2(32 * physical_device
->rad_info
.num_good_compute_units
,
911 max_threads_per_block
/ 64);
913 radv_device_init_gs_info(device
);
915 result
= radv_device_init_meta(device
);
916 if (result
!= VK_SUCCESS
)
919 radv_device_init_msaa(device
);
921 for (int family
= 0; family
< RADV_MAX_QUEUE_FAMILIES
; ++family
) {
922 device
->empty_cs
[family
] = device
->ws
->cs_create(device
->ws
, family
);
924 case RADV_QUEUE_GENERAL
:
925 radeon_emit(device
->empty_cs
[family
], PKT3(PKT3_CONTEXT_CONTROL
, 1, 0));
926 radeon_emit(device
->empty_cs
[family
], CONTEXT_CONTROL_LOAD_ENABLE(1));
927 radeon_emit(device
->empty_cs
[family
], CONTEXT_CONTROL_SHADOW_ENABLE(1));
929 case RADV_QUEUE_COMPUTE
:
930 radeon_emit(device
->empty_cs
[family
], PKT3(PKT3_NOP
, 0, 0));
931 radeon_emit(device
->empty_cs
[family
], 0);
934 device
->ws
->cs_finalize(device
->empty_cs
[family
]);
936 device
->flush_cs
[family
] = device
->ws
->cs_create(device
->ws
, family
);
938 case RADV_QUEUE_GENERAL
:
939 case RADV_QUEUE_COMPUTE
:
940 si_cs_emit_cache_flush(device
->flush_cs
[family
],
941 device
->physical_device
->rad_info
.chip_class
,
942 family
== RADV_QUEUE_COMPUTE
&& device
->physical_device
->rad_info
.chip_class
>= CIK
,
943 RADV_CMD_FLAG_INV_ICACHE
|
944 RADV_CMD_FLAG_INV_SMEM_L1
|
945 RADV_CMD_FLAG_INV_VMEM_L1
|
946 RADV_CMD_FLAG_INV_GLOBAL_L2
);
949 device
->ws
->cs_finalize(device
->flush_cs
[family
]);
952 if (getenv("RADV_TRACE_FILE")) {
953 device
->trace_bo
= device
->ws
->buffer_create(device
->ws
, 4096, 8,
954 RADEON_DOMAIN_VRAM
, RADEON_FLAG_CPU_ACCESS
);
955 if (!device
->trace_bo
)
958 device
->trace_id_ptr
= device
->ws
->buffer_map(device
->trace_bo
);
959 if (!device
->trace_id_ptr
)
963 if (device
->physical_device
->rad_info
.chip_class
>= CIK
)
964 cik_create_gfx_config(device
);
966 *pDevice
= radv_device_to_handle(device
);
970 if (device
->trace_bo
)
971 device
->ws
->buffer_destroy(device
->trace_bo
);
973 if (device
->gfx_init
)
974 device
->ws
->buffer_destroy(device
->gfx_init
);
976 for (unsigned i
= 0; i
< RADV_MAX_QUEUE_FAMILIES
; i
++) {
977 for (unsigned q
= 0; q
< device
->queue_count
[i
]; q
++)
978 radv_queue_finish(&device
->queues
[i
][q
]);
979 if (device
->queue_count
[i
])
980 vk_free(&device
->alloc
, device
->queues
[i
]);
983 vk_free(&device
->alloc
, device
);
987 void radv_DestroyDevice(
989 const VkAllocationCallbacks
* pAllocator
)
991 RADV_FROM_HANDLE(radv_device
, device
, _device
);
996 if (device
->trace_bo
)
997 device
->ws
->buffer_destroy(device
->trace_bo
);
999 if (device
->gfx_init
)
1000 device
->ws
->buffer_destroy(device
->gfx_init
);
1002 for (unsigned i
= 0; i
< RADV_MAX_QUEUE_FAMILIES
; i
++) {
1003 for (unsigned q
= 0; q
< device
->queue_count
[i
]; q
++)
1004 radv_queue_finish(&device
->queues
[i
][q
]);
1005 if (device
->queue_count
[i
])
1006 vk_free(&device
->alloc
, device
->queues
[i
]);
1007 if (device
->empty_cs
[i
])
1008 device
->ws
->cs_destroy(device
->empty_cs
[i
]);
1009 if (device
->flush_cs
[i
])
1010 device
->ws
->cs_destroy(device
->flush_cs
[i
]);
1012 radv_device_finish_meta(device
);
1014 vk_free(&device
->alloc
, device
);
1017 VkResult
radv_EnumerateInstanceExtensionProperties(
1018 const char* pLayerName
,
1019 uint32_t* pPropertyCount
,
1020 VkExtensionProperties
* pProperties
)
1022 if (pProperties
== NULL
) {
1023 *pPropertyCount
= ARRAY_SIZE(instance_extensions
);
1027 *pPropertyCount
= MIN2(*pPropertyCount
, ARRAY_SIZE(instance_extensions
));
1028 typed_memcpy(pProperties
, instance_extensions
, *pPropertyCount
);
1030 if (*pPropertyCount
< ARRAY_SIZE(instance_extensions
))
1031 return VK_INCOMPLETE
;
1036 VkResult
radv_EnumerateDeviceExtensionProperties(
1037 VkPhysicalDevice physicalDevice
,
1038 const char* pLayerName
,
1039 uint32_t* pPropertyCount
,
1040 VkExtensionProperties
* pProperties
)
1042 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
1044 if (pProperties
== NULL
) {
1045 *pPropertyCount
= pdevice
->extensions
.num_ext
;
1049 *pPropertyCount
= MIN2(*pPropertyCount
, pdevice
->extensions
.num_ext
);
1050 typed_memcpy(pProperties
, pdevice
->extensions
.ext_array
, *pPropertyCount
);
1052 if (*pPropertyCount
< pdevice
->extensions
.num_ext
)
1053 return VK_INCOMPLETE
;
1058 VkResult
radv_EnumerateInstanceLayerProperties(
1059 uint32_t* pPropertyCount
,
1060 VkLayerProperties
* pProperties
)
1062 if (pProperties
== NULL
) {
1063 *pPropertyCount
= 0;
1067 /* None supported at this time */
1068 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1071 VkResult
radv_EnumerateDeviceLayerProperties(
1072 VkPhysicalDevice physicalDevice
,
1073 uint32_t* pPropertyCount
,
1074 VkLayerProperties
* pProperties
)
1076 if (pProperties
== NULL
) {
1077 *pPropertyCount
= 0;
1081 /* None supported at this time */
1082 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1085 void radv_GetDeviceQueue(
1087 uint32_t queueFamilyIndex
,
1088 uint32_t queueIndex
,
1091 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1093 *pQueue
= radv_queue_to_handle(&device
->queues
[queueFamilyIndex
][queueIndex
]);
1096 static void radv_dump_trace(struct radv_device
*device
,
1097 struct radeon_winsys_cs
*cs
)
1099 const char *filename
= getenv("RADV_TRACE_FILE");
1100 FILE *f
= fopen(filename
, "w");
1102 fprintf(stderr
, "Failed to write trace dump to %s\n", filename
);
1106 fprintf(f
, "Trace ID: %x\n", *device
->trace_id_ptr
);
1107 device
->ws
->cs_dump(cs
, f
, *device
->trace_id_ptr
);
1112 fill_geom_rings(struct radv_queue
*queue
,
1114 uint32_t esgs_ring_size
,
1115 struct radeon_winsys_bo
*esgs_ring_bo
,
1116 uint32_t gsvs_ring_size
,
1117 struct radeon_winsys_bo
*gsvs_ring_bo
)
1119 uint64_t esgs_va
= 0, gsvs_va
= 0;
1120 uint32_t *desc
= &map
[4];
1123 esgs_va
= queue
->device
->ws
->buffer_get_va(esgs_ring_bo
);
1125 gsvs_va
= queue
->device
->ws
->buffer_get_va(gsvs_ring_bo
);
1127 /* stride 0, num records - size, add tid, swizzle, elsize4,
1130 desc
[1] = S_008F04_BASE_ADDRESS_HI(esgs_va
>> 32) |
1131 S_008F04_STRIDE(0) |
1132 S_008F04_SWIZZLE_ENABLE(true);
1133 desc
[2] = esgs_ring_size
;
1134 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1135 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1136 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1137 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1138 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1139 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1140 S_008F0C_ELEMENT_SIZE(1) |
1141 S_008F0C_INDEX_STRIDE(3) |
1142 S_008F0C_ADD_TID_ENABLE(true);
1145 /* GS entry for ES->GS ring */
1146 /* stride 0, num records - size, elsize0,
1149 desc
[1] = S_008F04_BASE_ADDRESS_HI(esgs_va
>> 32)|
1150 S_008F04_STRIDE(0) |
1151 S_008F04_SWIZZLE_ENABLE(false);
1152 desc
[2] = esgs_ring_size
;
1153 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1154 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1155 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1156 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1157 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1158 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1159 S_008F0C_ELEMENT_SIZE(0) |
1160 S_008F0C_INDEX_STRIDE(0) |
1161 S_008F0C_ADD_TID_ENABLE(false);
1164 /* VS entry for GS->VS ring */
1165 /* stride 0, num records - size, elsize0,
1168 desc
[1] = S_008F04_BASE_ADDRESS_HI(gsvs_va
>> 32)|
1169 S_008F04_STRIDE(0) |
1170 S_008F04_SWIZZLE_ENABLE(false);
1171 desc
[2] = gsvs_ring_size
;
1172 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1173 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1174 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1175 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1176 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1177 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1178 S_008F0C_ELEMENT_SIZE(0) |
1179 S_008F0C_INDEX_STRIDE(0) |
1180 S_008F0C_ADD_TID_ENABLE(false);
1183 /* stride gsvs_itemsize, num records 64
1184 elsize 4, index stride 16 */
1185 /* shader will patch stride and desc[2] */
1187 desc
[1] = S_008F04_BASE_ADDRESS_HI(gsvs_va
>> 32)|
1188 S_008F04_STRIDE(0) |
1189 S_008F04_SWIZZLE_ENABLE(true);
1191 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1192 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1193 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1194 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1195 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1196 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1197 S_008F0C_ELEMENT_SIZE(1) |
1198 S_008F0C_INDEX_STRIDE(1) |
1199 S_008F0C_ADD_TID_ENABLE(true);
1203 radv_get_preamble_cs(struct radv_queue
*queue
,
1204 uint32_t scratch_size
,
1205 uint32_t compute_scratch_size
,
1206 uint32_t esgs_ring_size
,
1207 uint32_t gsvs_ring_size
,
1208 struct radeon_winsys_cs
**initial_preamble_cs
,
1209 struct radeon_winsys_cs
**continue_preamble_cs
)
1211 struct radeon_winsys_bo
*scratch_bo
= NULL
;
1212 struct radeon_winsys_bo
*descriptor_bo
= NULL
;
1213 struct radeon_winsys_bo
*compute_scratch_bo
= NULL
;
1214 struct radeon_winsys_bo
*esgs_ring_bo
= NULL
;
1215 struct radeon_winsys_bo
*gsvs_ring_bo
= NULL
;
1216 struct radeon_winsys_cs
*dest_cs
[2] = {0};
1218 if (scratch_size
<= queue
->scratch_size
&&
1219 compute_scratch_size
<= queue
->compute_scratch_size
&&
1220 esgs_ring_size
<= queue
->esgs_ring_size
&&
1221 gsvs_ring_size
<= queue
->gsvs_ring_size
&&
1222 queue
->initial_preamble_cs
) {
1223 *initial_preamble_cs
= queue
->initial_preamble_cs
;
1224 *continue_preamble_cs
= queue
->continue_preamble_cs
;
1225 if (!scratch_size
&& !compute_scratch_size
&& !esgs_ring_size
&& !gsvs_ring_size
)
1226 *continue_preamble_cs
= NULL
;
1230 if (scratch_size
> queue
->scratch_size
) {
1231 scratch_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
1235 RADEON_FLAG_NO_CPU_ACCESS
);
1239 scratch_bo
= queue
->scratch_bo
;
1241 if (compute_scratch_size
> queue
->compute_scratch_size
) {
1242 compute_scratch_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
1243 compute_scratch_size
,
1246 RADEON_FLAG_NO_CPU_ACCESS
);
1247 if (!compute_scratch_bo
)
1251 compute_scratch_bo
= queue
->compute_scratch_bo
;
1253 if (esgs_ring_size
> queue
->esgs_ring_size
) {
1254 esgs_ring_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
1258 RADEON_FLAG_NO_CPU_ACCESS
);
1262 esgs_ring_bo
= queue
->esgs_ring_bo
;
1263 esgs_ring_size
= queue
->esgs_ring_size
;
1266 if (gsvs_ring_size
> queue
->gsvs_ring_size
) {
1267 gsvs_ring_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
1271 RADEON_FLAG_NO_CPU_ACCESS
);
1275 gsvs_ring_bo
= queue
->gsvs_ring_bo
;
1276 gsvs_ring_size
= queue
->gsvs_ring_size
;
1279 if (scratch_bo
!= queue
->scratch_bo
||
1280 esgs_ring_bo
!= queue
->esgs_ring_bo
||
1281 gsvs_ring_bo
!= queue
->gsvs_ring_bo
) {
1283 if (gsvs_ring_bo
|| esgs_ring_bo
)
1284 size
= 80; /* 2 dword + 2 padding + 4 dword * 4 */
1285 else if (scratch_bo
)
1286 size
= 8; /* 2 dword */
1288 descriptor_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
1292 RADEON_FLAG_CPU_ACCESS
);
1296 descriptor_bo
= queue
->descriptor_bo
;
1298 for(int i
= 0; i
< 2; ++i
) {
1299 struct radeon_winsys_cs
*cs
= NULL
;
1300 cs
= queue
->device
->ws
->cs_create(queue
->device
->ws
,
1301 queue
->queue_family_index
? RING_COMPUTE
: RING_GFX
);
1308 queue
->device
->ws
->cs_add_buffer(cs
, scratch_bo
, 8);
1311 queue
->device
->ws
->cs_add_buffer(cs
, esgs_ring_bo
, 8);
1314 queue
->device
->ws
->cs_add_buffer(cs
, gsvs_ring_bo
, 8);
1317 queue
->device
->ws
->cs_add_buffer(cs
, descriptor_bo
, 8);
1319 if (descriptor_bo
!= queue
->descriptor_bo
) {
1320 uint32_t *map
= (uint32_t*)queue
->device
->ws
->buffer_map(descriptor_bo
);
1323 uint64_t scratch_va
= queue
->device
->ws
->buffer_get_va(scratch_bo
);
1324 uint32_t rsrc1
= S_008F04_BASE_ADDRESS_HI(scratch_va
>> 32) |
1325 S_008F04_SWIZZLE_ENABLE(1);
1326 map
[0] = scratch_va
;
1330 if (esgs_ring_bo
|| gsvs_ring_bo
)
1331 fill_geom_rings(queue
, map
, esgs_ring_size
, esgs_ring_bo
, gsvs_ring_size
, gsvs_ring_bo
);
1333 queue
->device
->ws
->buffer_unmap(descriptor_bo
);
1336 if (esgs_ring_bo
|| gsvs_ring_bo
) {
1337 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1338 radeon_emit(cs
, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
1339 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1340 radeon_emit(cs
, EVENT_TYPE(V_028A90_VGT_FLUSH
) | EVENT_INDEX(0));
1342 if (queue
->device
->physical_device
->rad_info
.chip_class
>= CIK
) {
1343 radeon_set_uconfig_reg_seq(cs
, R_030900_VGT_ESGS_RING_SIZE
, 2);
1344 radeon_emit(cs
, esgs_ring_size
>> 8);
1345 radeon_emit(cs
, gsvs_ring_size
>> 8);
1347 radeon_set_config_reg_seq(cs
, R_0088C8_VGT_ESGS_RING_SIZE
, 2);
1348 radeon_emit(cs
, esgs_ring_size
>> 8);
1349 radeon_emit(cs
, gsvs_ring_size
>> 8);
1353 if (descriptor_bo
) {
1354 uint32_t regs
[] = {R_00B030_SPI_SHADER_USER_DATA_PS_0
,
1355 R_00B130_SPI_SHADER_USER_DATA_VS_0
,
1356 R_00B230_SPI_SHADER_USER_DATA_GS_0
,
1357 R_00B330_SPI_SHADER_USER_DATA_ES_0
,
1358 R_00B430_SPI_SHADER_USER_DATA_HS_0
,
1359 R_00B530_SPI_SHADER_USER_DATA_LS_0
};
1361 uint64_t va
= queue
->device
->ws
->buffer_get_va(descriptor_bo
);
1363 for (int i
= 0; i
< ARRAY_SIZE(regs
); ++i
) {
1364 radeon_set_sh_reg_seq(cs
, regs
[i
], 2);
1365 radeon_emit(cs
, va
);
1366 radeon_emit(cs
, va
>> 32);
1370 if (compute_scratch_bo
) {
1371 uint64_t scratch_va
= queue
->device
->ws
->buffer_get_va(compute_scratch_bo
);
1372 uint32_t rsrc1
= S_008F04_BASE_ADDRESS_HI(scratch_va
>> 32) |
1373 S_008F04_SWIZZLE_ENABLE(1);
1375 queue
->device
->ws
->cs_add_buffer(cs
, compute_scratch_bo
, 8);
1377 radeon_set_sh_reg_seq(cs
, R_00B900_COMPUTE_USER_DATA_0
, 2);
1378 radeon_emit(cs
, scratch_va
);
1379 radeon_emit(cs
, rsrc1
);
1383 si_cs_emit_cache_flush(cs
,
1384 queue
->device
->physical_device
->rad_info
.chip_class
,
1385 queue
->queue_family_index
== RING_COMPUTE
&&
1386 queue
->device
->physical_device
->rad_info
.chip_class
>= CIK
,
1387 RADV_CMD_FLAG_INV_ICACHE
|
1388 RADV_CMD_FLAG_INV_SMEM_L1
|
1389 RADV_CMD_FLAG_INV_VMEM_L1
|
1390 RADV_CMD_FLAG_INV_GLOBAL_L2
);
1393 if (!queue
->device
->ws
->cs_finalize(cs
))
1397 if (queue
->initial_preamble_cs
)
1398 queue
->device
->ws
->cs_destroy(queue
->initial_preamble_cs
);
1400 if (queue
->continue_preamble_cs
)
1401 queue
->device
->ws
->cs_destroy(queue
->continue_preamble_cs
);
1403 queue
->initial_preamble_cs
= dest_cs
[0];
1404 queue
->continue_preamble_cs
= dest_cs
[1];
1406 if (scratch_bo
!= queue
->scratch_bo
) {
1407 if (queue
->scratch_bo
)
1408 queue
->device
->ws
->buffer_destroy(queue
->scratch_bo
);
1409 queue
->scratch_bo
= scratch_bo
;
1410 queue
->scratch_size
= scratch_size
;
1413 if (compute_scratch_bo
!= queue
->compute_scratch_bo
) {
1414 if (queue
->compute_scratch_bo
)
1415 queue
->device
->ws
->buffer_destroy(queue
->compute_scratch_bo
);
1416 queue
->compute_scratch_bo
= compute_scratch_bo
;
1417 queue
->compute_scratch_size
= compute_scratch_size
;
1420 if (esgs_ring_bo
!= queue
->esgs_ring_bo
) {
1421 if (queue
->esgs_ring_bo
)
1422 queue
->device
->ws
->buffer_destroy(queue
->esgs_ring_bo
);
1423 queue
->esgs_ring_bo
= esgs_ring_bo
;
1424 queue
->esgs_ring_size
= esgs_ring_size
;
1427 if (gsvs_ring_bo
!= queue
->gsvs_ring_bo
) {
1428 if (queue
->gsvs_ring_bo
)
1429 queue
->device
->ws
->buffer_destroy(queue
->gsvs_ring_bo
);
1430 queue
->gsvs_ring_bo
= gsvs_ring_bo
;
1431 queue
->gsvs_ring_size
= gsvs_ring_size
;
1434 if (descriptor_bo
!= queue
->descriptor_bo
) {
1435 if (queue
->descriptor_bo
)
1436 queue
->device
->ws
->buffer_destroy(queue
->descriptor_bo
);
1438 queue
->descriptor_bo
= descriptor_bo
;
1441 *initial_preamble_cs
= queue
->initial_preamble_cs
;
1442 *continue_preamble_cs
= queue
->continue_preamble_cs
;
1443 if (!scratch_size
&& !compute_scratch_size
&& !esgs_ring_size
&& !gsvs_ring_size
)
1444 *continue_preamble_cs
= NULL
;
1447 for (int i
= 0; i
< ARRAY_SIZE(dest_cs
); ++i
)
1449 queue
->device
->ws
->cs_destroy(dest_cs
[i
]);
1450 if (descriptor_bo
&& descriptor_bo
!= queue
->descriptor_bo
)
1451 queue
->device
->ws
->buffer_destroy(descriptor_bo
);
1452 if (scratch_bo
&& scratch_bo
!= queue
->scratch_bo
)
1453 queue
->device
->ws
->buffer_destroy(scratch_bo
);
1454 if (compute_scratch_bo
&& compute_scratch_bo
!= queue
->compute_scratch_bo
)
1455 queue
->device
->ws
->buffer_destroy(compute_scratch_bo
);
1456 if (esgs_ring_bo
&& esgs_ring_bo
!= queue
->esgs_ring_bo
)
1457 queue
->device
->ws
->buffer_destroy(esgs_ring_bo
);
1458 if (gsvs_ring_bo
&& gsvs_ring_bo
!= queue
->gsvs_ring_bo
)
1459 queue
->device
->ws
->buffer_destroy(gsvs_ring_bo
);
1460 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
1463 VkResult
radv_QueueSubmit(
1465 uint32_t submitCount
,
1466 const VkSubmitInfo
* pSubmits
,
1469 RADV_FROM_HANDLE(radv_queue
, queue
, _queue
);
1470 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
1471 struct radeon_winsys_fence
*base_fence
= fence
? fence
->fence
: NULL
;
1472 struct radeon_winsys_ctx
*ctx
= queue
->hw_ctx
;
1474 uint32_t max_cs_submission
= queue
->device
->trace_bo
? 1 : UINT32_MAX
;
1475 uint32_t scratch_size
= 0;
1476 uint32_t compute_scratch_size
= 0;
1477 uint32_t esgs_ring_size
= 0, gsvs_ring_size
= 0;
1478 struct radeon_winsys_cs
*initial_preamble_cs
= NULL
, *continue_preamble_cs
= NULL
;
1480 bool fence_emitted
= false;
1482 /* Do this first so failing to allocate scratch buffers can't result in
1483 * partially executed submissions. */
1484 for (uint32_t i
= 0; i
< submitCount
; i
++) {
1485 for (uint32_t j
= 0; j
< pSubmits
[i
].commandBufferCount
; j
++) {
1486 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
,
1487 pSubmits
[i
].pCommandBuffers
[j
]);
1489 scratch_size
= MAX2(scratch_size
, cmd_buffer
->scratch_size_needed
);
1490 compute_scratch_size
= MAX2(compute_scratch_size
,
1491 cmd_buffer
->compute_scratch_size_needed
);
1492 esgs_ring_size
= MAX2(esgs_ring_size
, cmd_buffer
->esgs_ring_size_needed
);
1493 gsvs_ring_size
= MAX2(gsvs_ring_size
, cmd_buffer
->gsvs_ring_size_needed
);
1497 result
= radv_get_preamble_cs(queue
, scratch_size
, compute_scratch_size
,
1498 esgs_ring_size
, gsvs_ring_size
,
1499 &initial_preamble_cs
, &continue_preamble_cs
);
1500 if (result
!= VK_SUCCESS
)
1503 for (uint32_t i
= 0; i
< submitCount
; i
++) {
1504 struct radeon_winsys_cs
**cs_array
;
1505 bool has_flush
= !submitCount
;
1506 bool can_patch
= !has_flush
;
1509 if (!pSubmits
[i
].commandBufferCount
) {
1510 if (pSubmits
[i
].waitSemaphoreCount
|| pSubmits
[i
].signalSemaphoreCount
) {
1511 ret
= queue
->device
->ws
->cs_submit(ctx
, queue
->queue_idx
,
1512 &queue
->device
->empty_cs
[queue
->queue_family_index
],
1514 (struct radeon_winsys_sem
**)pSubmits
[i
].pWaitSemaphores
,
1515 pSubmits
[i
].waitSemaphoreCount
,
1516 (struct radeon_winsys_sem
**)pSubmits
[i
].pSignalSemaphores
,
1517 pSubmits
[i
].signalSemaphoreCount
,
1520 radv_loge("failed to submit CS %d\n", i
);
1523 fence_emitted
= true;
1528 cs_array
= malloc(sizeof(struct radeon_winsys_cs
*) *
1529 (pSubmits
[i
].commandBufferCount
+ has_flush
));
1532 cs_array
[0] = queue
->device
->flush_cs
[queue
->queue_family_index
];
1534 for (uint32_t j
= 0; j
< pSubmits
[i
].commandBufferCount
; j
++) {
1535 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
,
1536 pSubmits
[i
].pCommandBuffers
[j
]);
1537 assert(cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_PRIMARY
);
1539 cs_array
[j
+ has_flush
] = cmd_buffer
->cs
;
1540 if ((cmd_buffer
->usage_flags
& VK_COMMAND_BUFFER_USAGE_SIMULTANEOUS_USE_BIT
))
1544 for (uint32_t j
= 0; j
< pSubmits
[i
].commandBufferCount
+ has_flush
; j
+= advance
) {
1545 advance
= MIN2(max_cs_submission
,
1546 pSubmits
[i
].commandBufferCount
+ has_flush
- j
);
1548 bool e
= j
+ advance
== pSubmits
[i
].commandBufferCount
+ has_flush
;
1550 if (queue
->device
->trace_bo
)
1551 *queue
->device
->trace_id_ptr
= 0;
1553 ret
= queue
->device
->ws
->cs_submit(ctx
, queue
->queue_idx
, cs_array
+ j
,
1554 advance
, initial_preamble_cs
, continue_preamble_cs
,
1555 (struct radeon_winsys_sem
**)pSubmits
[i
].pWaitSemaphores
,
1556 b
? pSubmits
[i
].waitSemaphoreCount
: 0,
1557 (struct radeon_winsys_sem
**)pSubmits
[i
].pSignalSemaphores
,
1558 e
? pSubmits
[i
].signalSemaphoreCount
: 0,
1559 can_patch
, base_fence
);
1562 radv_loge("failed to submit CS %d\n", i
);
1565 fence_emitted
= true;
1566 if (queue
->device
->trace_bo
) {
1567 bool success
= queue
->device
->ws
->ctx_wait_idle(
1569 radv_queue_family_to_ring(
1570 queue
->queue_family_index
),
1573 if (!success
) { /* Hang */
1574 radv_dump_trace(queue
->device
, cs_array
[j
]);
1584 ret
= queue
->device
->ws
->cs_submit(ctx
, queue
->queue_idx
,
1585 &queue
->device
->empty_cs
[queue
->queue_family_index
],
1586 1, NULL
, NULL
, NULL
, 0, NULL
, 0,
1589 fence
->submitted
= true;
1595 VkResult
radv_QueueWaitIdle(
1598 RADV_FROM_HANDLE(radv_queue
, queue
, _queue
);
1600 queue
->device
->ws
->ctx_wait_idle(queue
->hw_ctx
,
1601 radv_queue_family_to_ring(queue
->queue_family_index
),
1606 VkResult
radv_DeviceWaitIdle(
1609 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1611 for (unsigned i
= 0; i
< RADV_MAX_QUEUE_FAMILIES
; i
++) {
1612 for (unsigned q
= 0; q
< device
->queue_count
[i
]; q
++) {
1613 radv_QueueWaitIdle(radv_queue_to_handle(&device
->queues
[i
][q
]));
1619 PFN_vkVoidFunction
radv_GetInstanceProcAddr(
1620 VkInstance instance
,
1623 return radv_lookup_entrypoint(pName
);
1626 /* The loader wants us to expose a second GetInstanceProcAddr function
1627 * to work around certain LD_PRELOAD issues seen in apps.
1630 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1631 VkInstance instance
,
1635 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1636 VkInstance instance
,
1639 return radv_GetInstanceProcAddr(instance
, pName
);
1642 PFN_vkVoidFunction
radv_GetDeviceProcAddr(
1646 return radv_lookup_entrypoint(pName
);
1649 bool radv_get_memory_fd(struct radv_device
*device
,
1650 struct radv_device_memory
*memory
,
1653 struct radeon_bo_metadata metadata
;
1655 if (memory
->image
) {
1656 radv_init_metadata(device
, memory
->image
, &metadata
);
1657 device
->ws
->buffer_set_metadata(memory
->bo
, &metadata
);
1660 return device
->ws
->buffer_get_fd(device
->ws
, memory
->bo
,
1664 VkResult
radv_AllocateMemory(
1666 const VkMemoryAllocateInfo
* pAllocateInfo
,
1667 const VkAllocationCallbacks
* pAllocator
,
1668 VkDeviceMemory
* pMem
)
1670 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1671 struct radv_device_memory
*mem
;
1673 enum radeon_bo_domain domain
;
1675 const VkDedicatedAllocationMemoryAllocateInfoNV
*dedicate_info
= NULL
;
1676 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
1678 if (pAllocateInfo
->allocationSize
== 0) {
1679 /* Apparently, this is allowed */
1680 *pMem
= VK_NULL_HANDLE
;
1684 vk_foreach_struct(ext
, pAllocateInfo
->pNext
) {
1685 switch (ext
->sType
) {
1686 case VK_STRUCTURE_TYPE_DEDICATED_ALLOCATION_MEMORY_ALLOCATE_INFO_NV
:
1687 dedicate_info
= (const VkDedicatedAllocationMemoryAllocateInfoNV
*)ext
;
1694 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
1695 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1697 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1699 if (dedicate_info
) {
1700 mem
->image
= radv_image_from_handle(dedicate_info
->image
);
1701 mem
->buffer
= radv_buffer_from_handle(dedicate_info
->buffer
);
1707 uint64_t alloc_size
= align_u64(pAllocateInfo
->allocationSize
, 4096);
1708 if (pAllocateInfo
->memoryTypeIndex
== RADV_MEM_TYPE_GTT_WRITE_COMBINE
||
1709 pAllocateInfo
->memoryTypeIndex
== RADV_MEM_TYPE_GTT_CACHED
)
1710 domain
= RADEON_DOMAIN_GTT
;
1712 domain
= RADEON_DOMAIN_VRAM
;
1714 if (pAllocateInfo
->memoryTypeIndex
== RADV_MEM_TYPE_VRAM
)
1715 flags
|= RADEON_FLAG_NO_CPU_ACCESS
;
1717 flags
|= RADEON_FLAG_CPU_ACCESS
;
1719 if (pAllocateInfo
->memoryTypeIndex
== RADV_MEM_TYPE_GTT_WRITE_COMBINE
)
1720 flags
|= RADEON_FLAG_GTT_WC
;
1722 mem
->bo
= device
->ws
->buffer_create(device
->ws
, alloc_size
, 65536,
1726 result
= VK_ERROR_OUT_OF_DEVICE_MEMORY
;
1729 mem
->type_index
= pAllocateInfo
->memoryTypeIndex
;
1731 *pMem
= radv_device_memory_to_handle(mem
);
1736 vk_free2(&device
->alloc
, pAllocator
, mem
);
1741 void radv_FreeMemory(
1743 VkDeviceMemory _mem
,
1744 const VkAllocationCallbacks
* pAllocator
)
1746 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1747 RADV_FROM_HANDLE(radv_device_memory
, mem
, _mem
);
1752 device
->ws
->buffer_destroy(mem
->bo
);
1755 vk_free2(&device
->alloc
, pAllocator
, mem
);
1758 VkResult
radv_MapMemory(
1760 VkDeviceMemory _memory
,
1761 VkDeviceSize offset
,
1763 VkMemoryMapFlags flags
,
1766 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1767 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
1774 *ppData
= device
->ws
->buffer_map(mem
->bo
);
1780 return VK_ERROR_MEMORY_MAP_FAILED
;
1783 void radv_UnmapMemory(
1785 VkDeviceMemory _memory
)
1787 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1788 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
1793 device
->ws
->buffer_unmap(mem
->bo
);
1796 VkResult
radv_FlushMappedMemoryRanges(
1798 uint32_t memoryRangeCount
,
1799 const VkMappedMemoryRange
* pMemoryRanges
)
1804 VkResult
radv_InvalidateMappedMemoryRanges(
1806 uint32_t memoryRangeCount
,
1807 const VkMappedMemoryRange
* pMemoryRanges
)
1812 void radv_GetBufferMemoryRequirements(
1815 VkMemoryRequirements
* pMemoryRequirements
)
1817 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
1819 pMemoryRequirements
->memoryTypeBits
= (1u << RADV_MEM_TYPE_COUNT
) - 1;
1821 pMemoryRequirements
->size
= buffer
->size
;
1822 pMemoryRequirements
->alignment
= 16;
1825 void radv_GetImageMemoryRequirements(
1828 VkMemoryRequirements
* pMemoryRequirements
)
1830 RADV_FROM_HANDLE(radv_image
, image
, _image
);
1832 pMemoryRequirements
->memoryTypeBits
= (1u << RADV_MEM_TYPE_COUNT
) - 1;
1834 pMemoryRequirements
->size
= image
->size
;
1835 pMemoryRequirements
->alignment
= image
->alignment
;
1838 void radv_GetImageSparseMemoryRequirements(
1841 uint32_t* pSparseMemoryRequirementCount
,
1842 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
1847 void radv_GetDeviceMemoryCommitment(
1849 VkDeviceMemory memory
,
1850 VkDeviceSize
* pCommittedMemoryInBytes
)
1852 *pCommittedMemoryInBytes
= 0;
1855 VkResult
radv_BindBufferMemory(
1858 VkDeviceMemory _memory
,
1859 VkDeviceSize memoryOffset
)
1861 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
1862 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
1865 buffer
->bo
= mem
->bo
;
1866 buffer
->offset
= memoryOffset
;
1875 VkResult
radv_BindImageMemory(
1878 VkDeviceMemory _memory
,
1879 VkDeviceSize memoryOffset
)
1881 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
1882 RADV_FROM_HANDLE(radv_image
, image
, _image
);
1885 image
->bo
= mem
->bo
;
1886 image
->offset
= memoryOffset
;
1895 VkResult
radv_QueueBindSparse(
1897 uint32_t bindInfoCount
,
1898 const VkBindSparseInfo
* pBindInfo
,
1901 stub_return(VK_ERROR_INCOMPATIBLE_DRIVER
);
1904 VkResult
radv_CreateFence(
1906 const VkFenceCreateInfo
* pCreateInfo
,
1907 const VkAllocationCallbacks
* pAllocator
,
1910 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1911 struct radv_fence
*fence
= vk_alloc2(&device
->alloc
, pAllocator
,
1913 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1916 return VK_ERROR_OUT_OF_HOST_MEMORY
;
1918 memset(fence
, 0, sizeof(*fence
));
1919 fence
->submitted
= false;
1920 fence
->signalled
= !!(pCreateInfo
->flags
& VK_FENCE_CREATE_SIGNALED_BIT
);
1921 fence
->fence
= device
->ws
->create_fence();
1922 if (!fence
->fence
) {
1923 vk_free2(&device
->alloc
, pAllocator
, fence
);
1924 return VK_ERROR_OUT_OF_HOST_MEMORY
;
1927 *pFence
= radv_fence_to_handle(fence
);
1932 void radv_DestroyFence(
1935 const VkAllocationCallbacks
* pAllocator
)
1937 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1938 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
1942 device
->ws
->destroy_fence(fence
->fence
);
1943 vk_free2(&device
->alloc
, pAllocator
, fence
);
1946 static uint64_t radv_get_absolute_timeout(uint64_t timeout
)
1948 uint64_t current_time
;
1951 clock_gettime(CLOCK_MONOTONIC
, &tv
);
1952 current_time
= tv
.tv_nsec
+ tv
.tv_sec
*1000000000ull;
1954 timeout
= MIN2(UINT64_MAX
- current_time
, timeout
);
1956 return current_time
+ timeout
;
1959 VkResult
radv_WaitForFences(
1961 uint32_t fenceCount
,
1962 const VkFence
* pFences
,
1966 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1967 timeout
= radv_get_absolute_timeout(timeout
);
1969 if (!waitAll
&& fenceCount
> 1) {
1970 fprintf(stderr
, "radv: WaitForFences without waitAll not implemented yet\n");
1973 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
1974 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
1975 bool expired
= false;
1977 if (fence
->signalled
)
1980 if (!fence
->submitted
)
1983 expired
= device
->ws
->fence_wait(device
->ws
, fence
->fence
, true, timeout
);
1987 fence
->signalled
= true;
1993 VkResult
radv_ResetFences(VkDevice device
,
1994 uint32_t fenceCount
,
1995 const VkFence
*pFences
)
1997 for (unsigned i
= 0; i
< fenceCount
; ++i
) {
1998 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
1999 fence
->submitted
= fence
->signalled
= false;
2005 VkResult
radv_GetFenceStatus(VkDevice _device
, VkFence _fence
)
2007 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2008 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
2010 if (fence
->signalled
)
2012 if (!fence
->submitted
)
2013 return VK_NOT_READY
;
2015 if (!device
->ws
->fence_wait(device
->ws
, fence
->fence
, false, 0))
2016 return VK_NOT_READY
;
2022 // Queue semaphore functions
2024 VkResult
radv_CreateSemaphore(
2026 const VkSemaphoreCreateInfo
* pCreateInfo
,
2027 const VkAllocationCallbacks
* pAllocator
,
2028 VkSemaphore
* pSemaphore
)
2030 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2031 struct radeon_winsys_sem
*sem
;
2033 sem
= device
->ws
->create_sem(device
->ws
);
2035 return VK_ERROR_OUT_OF_HOST_MEMORY
;
2037 *pSemaphore
= radeon_winsys_sem_to_handle(sem
);
2041 void radv_DestroySemaphore(
2043 VkSemaphore _semaphore
,
2044 const VkAllocationCallbacks
* pAllocator
)
2046 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2047 RADV_FROM_HANDLE(radeon_winsys_sem
, sem
, _semaphore
);
2051 device
->ws
->destroy_sem(sem
);
2054 VkResult
radv_CreateEvent(
2056 const VkEventCreateInfo
* pCreateInfo
,
2057 const VkAllocationCallbacks
* pAllocator
,
2060 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2061 struct radv_event
*event
= vk_alloc2(&device
->alloc
, pAllocator
,
2063 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2066 return VK_ERROR_OUT_OF_HOST_MEMORY
;
2068 event
->bo
= device
->ws
->buffer_create(device
->ws
, 8, 8,
2070 RADEON_FLAG_CPU_ACCESS
);
2072 vk_free2(&device
->alloc
, pAllocator
, event
);
2073 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
2076 event
->map
= (uint64_t*)device
->ws
->buffer_map(event
->bo
);
2078 *pEvent
= radv_event_to_handle(event
);
2083 void radv_DestroyEvent(
2086 const VkAllocationCallbacks
* pAllocator
)
2088 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2089 RADV_FROM_HANDLE(radv_event
, event
, _event
);
2093 device
->ws
->buffer_destroy(event
->bo
);
2094 vk_free2(&device
->alloc
, pAllocator
, event
);
2097 VkResult
radv_GetEventStatus(
2101 RADV_FROM_HANDLE(radv_event
, event
, _event
);
2103 if (*event
->map
== 1)
2104 return VK_EVENT_SET
;
2105 return VK_EVENT_RESET
;
2108 VkResult
radv_SetEvent(
2112 RADV_FROM_HANDLE(radv_event
, event
, _event
);
2118 VkResult
radv_ResetEvent(
2122 RADV_FROM_HANDLE(radv_event
, event
, _event
);
2128 VkResult
radv_CreateBuffer(
2130 const VkBufferCreateInfo
* pCreateInfo
,
2131 const VkAllocationCallbacks
* pAllocator
,
2134 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2135 struct radv_buffer
*buffer
;
2137 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
2139 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
2140 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2142 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2144 buffer
->size
= pCreateInfo
->size
;
2145 buffer
->usage
= pCreateInfo
->usage
;
2149 *pBuffer
= radv_buffer_to_handle(buffer
);
2154 void radv_DestroyBuffer(
2157 const VkAllocationCallbacks
* pAllocator
)
2159 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2160 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
2165 vk_free2(&device
->alloc
, pAllocator
, buffer
);
2168 static inline unsigned
2169 si_tile_mode_index(const struct radv_image
*image
, unsigned level
, bool stencil
)
2172 return image
->surface
.stencil_tiling_index
[level
];
2174 return image
->surface
.tiling_index
[level
];
2177 static uint32_t radv_surface_layer_count(struct radv_image_view
*iview
)
2179 return iview
->type
== VK_IMAGE_VIEW_TYPE_3D
? iview
->extent
.depth
: iview
->layer_count
;
2183 radv_initialise_color_surface(struct radv_device
*device
,
2184 struct radv_color_buffer_info
*cb
,
2185 struct radv_image_view
*iview
)
2187 const struct vk_format_description
*desc
;
2188 unsigned ntype
, format
, swap
, endian
;
2189 unsigned blend_clamp
= 0, blend_bypass
= 0;
2190 unsigned pitch_tile_max
, slice_tile_max
, tile_mode_index
;
2192 const struct radeon_surf
*surf
= &iview
->image
->surface
;
2193 const struct radeon_surf_level
*level_info
= &surf
->level
[iview
->base_mip
];
2195 desc
= vk_format_description(iview
->vk_format
);
2197 memset(cb
, 0, sizeof(*cb
));
2199 va
= device
->ws
->buffer_get_va(iview
->bo
) + iview
->image
->offset
;
2200 va
+= level_info
->offset
;
2201 cb
->cb_color_base
= va
>> 8;
2203 /* CMASK variables */
2204 va
= device
->ws
->buffer_get_va(iview
->bo
) + iview
->image
->offset
;
2205 va
+= iview
->image
->cmask
.offset
;
2206 cb
->cb_color_cmask
= va
>> 8;
2207 cb
->cb_color_cmask_slice
= iview
->image
->cmask
.slice_tile_max
;
2209 va
= device
->ws
->buffer_get_va(iview
->bo
) + iview
->image
->offset
;
2210 va
+= iview
->image
->dcc_offset
;
2211 cb
->cb_dcc_base
= va
>> 8;
2213 uint32_t max_slice
= radv_surface_layer_count(iview
);
2214 cb
->cb_color_view
= S_028C6C_SLICE_START(iview
->base_layer
) |
2215 S_028C6C_SLICE_MAX(iview
->base_layer
+ max_slice
- 1);
2217 cb
->micro_tile_mode
= iview
->image
->surface
.micro_tile_mode
;
2218 pitch_tile_max
= level_info
->nblk_x
/ 8 - 1;
2219 slice_tile_max
= (level_info
->nblk_x
* level_info
->nblk_y
) / 64 - 1;
2220 tile_mode_index
= si_tile_mode_index(iview
->image
, iview
->base_mip
, false);
2222 cb
->cb_color_pitch
= S_028C64_TILE_MAX(pitch_tile_max
);
2223 cb
->cb_color_slice
= S_028C68_TILE_MAX(slice_tile_max
);
2225 /* Intensity is implemented as Red, so treat it that way. */
2226 cb
->cb_color_attrib
= S_028C74_FORCE_DST_ALPHA_1(desc
->swizzle
[3] == VK_SWIZZLE_1
) |
2227 S_028C74_TILE_MODE_INDEX(tile_mode_index
);
2229 if (iview
->image
->samples
> 1) {
2230 unsigned log_samples
= util_logbase2(iview
->image
->samples
);
2232 cb
->cb_color_attrib
|= S_028C74_NUM_SAMPLES(log_samples
) |
2233 S_028C74_NUM_FRAGMENTS(log_samples
);
2236 if (iview
->image
->fmask
.size
) {
2237 va
= device
->ws
->buffer_get_va(iview
->bo
) + iview
->image
->offset
+ iview
->image
->fmask
.offset
;
2238 if (device
->physical_device
->rad_info
.chip_class
>= CIK
)
2239 cb
->cb_color_pitch
|= S_028C64_FMASK_TILE_MAX(iview
->image
->fmask
.pitch_in_pixels
/ 8 - 1);
2240 cb
->cb_color_attrib
|= S_028C74_FMASK_TILE_MODE_INDEX(iview
->image
->fmask
.tile_mode_index
);
2241 cb
->cb_color_fmask
= va
>> 8;
2242 cb
->cb_color_fmask_slice
= S_028C88_TILE_MAX(iview
->image
->fmask
.slice_tile_max
);
2244 /* This must be set for fast clear to work without FMASK. */
2245 if (device
->physical_device
->rad_info
.chip_class
>= CIK
)
2246 cb
->cb_color_pitch
|= S_028C64_FMASK_TILE_MAX(pitch_tile_max
);
2247 cb
->cb_color_attrib
|= S_028C74_FMASK_TILE_MODE_INDEX(tile_mode_index
);
2248 cb
->cb_color_fmask
= cb
->cb_color_base
;
2249 cb
->cb_color_fmask_slice
= S_028C88_TILE_MAX(slice_tile_max
);
2252 ntype
= radv_translate_color_numformat(iview
->vk_format
,
2254 vk_format_get_first_non_void_channel(iview
->vk_format
));
2255 format
= radv_translate_colorformat(iview
->vk_format
);
2256 if (format
== V_028C70_COLOR_INVALID
|| ntype
== ~0u)
2257 radv_finishme("Illegal color\n");
2258 swap
= radv_translate_colorswap(iview
->vk_format
, FALSE
);
2259 endian
= radv_colorformat_endian_swap(format
);
2261 /* blend clamp should be set for all NORM/SRGB types */
2262 if (ntype
== V_028C70_NUMBER_UNORM
||
2263 ntype
== V_028C70_NUMBER_SNORM
||
2264 ntype
== V_028C70_NUMBER_SRGB
)
2267 /* set blend bypass according to docs if SINT/UINT or
2268 8/24 COLOR variants */
2269 if (ntype
== V_028C70_NUMBER_UINT
|| ntype
== V_028C70_NUMBER_SINT
||
2270 format
== V_028C70_COLOR_8_24
|| format
== V_028C70_COLOR_24_8
||
2271 format
== V_028C70_COLOR_X24_8_32_FLOAT
) {
2276 if ((ntype
== V_028C70_NUMBER_UINT
|| ntype
== V_028C70_NUMBER_SINT
) &&
2277 (format
== V_028C70_COLOR_8
||
2278 format
== V_028C70_COLOR_8_8
||
2279 format
== V_028C70_COLOR_8_8_8_8
))
2280 ->color_is_int8
= true;
2282 cb
->cb_color_info
= S_028C70_FORMAT(format
) |
2283 S_028C70_COMP_SWAP(swap
) |
2284 S_028C70_BLEND_CLAMP(blend_clamp
) |
2285 S_028C70_BLEND_BYPASS(blend_bypass
) |
2286 S_028C70_SIMPLE_FLOAT(1) |
2287 S_028C70_ROUND_MODE(ntype
!= V_028C70_NUMBER_UNORM
&&
2288 ntype
!= V_028C70_NUMBER_SNORM
&&
2289 ntype
!= V_028C70_NUMBER_SRGB
&&
2290 format
!= V_028C70_COLOR_8_24
&&
2291 format
!= V_028C70_COLOR_24_8
) |
2292 S_028C70_NUMBER_TYPE(ntype
) |
2293 S_028C70_ENDIAN(endian
);
2294 if (iview
->image
->samples
> 1)
2295 if (iview
->image
->fmask
.size
)
2296 cb
->cb_color_info
|= S_028C70_COMPRESSION(1);
2298 if (iview
->image
->cmask
.size
&&
2299 !(device
->debug_flags
& RADV_DEBUG_NO_FAST_CLEARS
))
2300 cb
->cb_color_info
|= S_028C70_FAST_CLEAR(1);
2302 if (iview
->image
->surface
.dcc_size
&& level_info
->dcc_enabled
)
2303 cb
->cb_color_info
|= S_028C70_DCC_ENABLE(1);
2305 if (device
->physical_device
->rad_info
.chip_class
>= VI
) {
2306 unsigned max_uncompressed_block_size
= 2;
2307 if (iview
->image
->samples
> 1) {
2308 if (iview
->image
->surface
.bpe
== 1)
2309 max_uncompressed_block_size
= 0;
2310 else if (iview
->image
->surface
.bpe
== 2)
2311 max_uncompressed_block_size
= 1;
2314 cb
->cb_dcc_control
= S_028C78_MAX_UNCOMPRESSED_BLOCK_SIZE(max_uncompressed_block_size
) |
2315 S_028C78_INDEPENDENT_64B_BLOCKS(1);
2318 /* This must be set for fast clear to work without FMASK. */
2319 if (!iview
->image
->fmask
.size
&&
2320 device
->physical_device
->rad_info
.chip_class
== SI
) {
2321 unsigned bankh
= util_logbase2(iview
->image
->surface
.bankh
);
2322 cb
->cb_color_attrib
|= S_028C74_FMASK_BANK_HEIGHT(bankh
);
2327 radv_initialise_ds_surface(struct radv_device
*device
,
2328 struct radv_ds_buffer_info
*ds
,
2329 struct radv_image_view
*iview
)
2331 unsigned level
= iview
->base_mip
;
2333 uint64_t va
, s_offs
, z_offs
;
2334 const struct radeon_surf_level
*level_info
= &iview
->image
->surface
.level
[level
];
2335 memset(ds
, 0, sizeof(*ds
));
2336 switch (iview
->vk_format
) {
2337 case VK_FORMAT_D24_UNORM_S8_UINT
:
2338 case VK_FORMAT_X8_D24_UNORM_PACK32
:
2339 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-24);
2340 ds
->offset_scale
= 2.0f
;
2342 case VK_FORMAT_D16_UNORM
:
2343 case VK_FORMAT_D16_UNORM_S8_UINT
:
2344 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-16);
2345 ds
->offset_scale
= 4.0f
;
2347 case VK_FORMAT_D32_SFLOAT
:
2348 case VK_FORMAT_D32_SFLOAT_S8_UINT
:
2349 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-23) |
2350 S_028B78_POLY_OFFSET_DB_IS_FLOAT_FMT(1);
2351 ds
->offset_scale
= 1.0f
;
2357 format
= radv_translate_dbformat(iview
->vk_format
);
2358 if (format
== V_028040_Z_INVALID
) {
2359 fprintf(stderr
, "Invalid DB format: %d, disabling DB.\n", iview
->vk_format
);
2362 va
= device
->ws
->buffer_get_va(iview
->bo
) + iview
->image
->offset
;
2363 s_offs
= z_offs
= va
;
2364 z_offs
+= iview
->image
->surface
.level
[level
].offset
;
2365 s_offs
+= iview
->image
->surface
.stencil_level
[level
].offset
;
2367 uint32_t max_slice
= radv_surface_layer_count(iview
);
2368 ds
->db_depth_view
= S_028008_SLICE_START(iview
->base_layer
) |
2369 S_028008_SLICE_MAX(iview
->base_layer
+ max_slice
- 1);
2370 ds
->db_depth_info
= S_02803C_ADDR5_SWIZZLE_MASK(1);
2371 ds
->db_z_info
= S_028040_FORMAT(format
) | S_028040_ZRANGE_PRECISION(1);
2373 if (iview
->image
->samples
> 1)
2374 ds
->db_z_info
|= S_028040_NUM_SAMPLES(util_logbase2(iview
->image
->samples
));
2376 if (iview
->image
->surface
.flags
& RADEON_SURF_SBUFFER
)
2377 ds
->db_stencil_info
= S_028044_FORMAT(V_028044_STENCIL_8
);
2379 ds
->db_stencil_info
= S_028044_FORMAT(V_028044_STENCIL_INVALID
);
2381 if (device
->physical_device
->rad_info
.chip_class
>= CIK
) {
2382 struct radeon_info
*info
= &device
->physical_device
->rad_info
;
2383 unsigned tiling_index
= iview
->image
->surface
.tiling_index
[level
];
2384 unsigned stencil_index
= iview
->image
->surface
.stencil_tiling_index
[level
];
2385 unsigned macro_index
= iview
->image
->surface
.macro_tile_index
;
2386 unsigned tile_mode
= info
->si_tile_mode_array
[tiling_index
];
2387 unsigned stencil_tile_mode
= info
->si_tile_mode_array
[stencil_index
];
2388 unsigned macro_mode
= info
->cik_macrotile_mode_array
[macro_index
];
2390 ds
->db_depth_info
|=
2391 S_02803C_ARRAY_MODE(G_009910_ARRAY_MODE(tile_mode
)) |
2392 S_02803C_PIPE_CONFIG(G_009910_PIPE_CONFIG(tile_mode
)) |
2393 S_02803C_BANK_WIDTH(G_009990_BANK_WIDTH(macro_mode
)) |
2394 S_02803C_BANK_HEIGHT(G_009990_BANK_HEIGHT(macro_mode
)) |
2395 S_02803C_MACRO_TILE_ASPECT(G_009990_MACRO_TILE_ASPECT(macro_mode
)) |
2396 S_02803C_NUM_BANKS(G_009990_NUM_BANKS(macro_mode
));
2397 ds
->db_z_info
|= S_028040_TILE_SPLIT(G_009910_TILE_SPLIT(tile_mode
));
2398 ds
->db_stencil_info
|= S_028044_TILE_SPLIT(G_009910_TILE_SPLIT(stencil_tile_mode
));
2400 unsigned tile_mode_index
= si_tile_mode_index(iview
->image
, level
, false);
2401 ds
->db_z_info
|= S_028040_TILE_MODE_INDEX(tile_mode_index
);
2402 tile_mode_index
= si_tile_mode_index(iview
->image
, level
, true);
2403 ds
->db_stencil_info
|= S_028044_TILE_MODE_INDEX(tile_mode_index
);
2406 if (iview
->image
->surface
.htile_size
&& !level
) {
2407 ds
->db_z_info
|= S_028040_TILE_SURFACE_ENABLE(1) |
2408 S_028040_ALLOW_EXPCLEAR(1);
2410 if (iview
->image
->surface
.flags
& RADEON_SURF_SBUFFER
) {
2411 /* Workaround: For a not yet understood reason, the
2412 * combination of MSAA, fast stencil clear and stencil
2413 * decompress messes with subsequent stencil buffer
2414 * uses. Problem was reproduced on Verde, Bonaire,
2415 * Tonga, and Carrizo.
2417 * Disabling EXPCLEAR works around the problem.
2419 * Check piglit's arb_texture_multisample-stencil-clear
2420 * test if you want to try changing this.
2422 if (iview
->image
->samples
<= 1)
2423 ds
->db_stencil_info
|= S_028044_ALLOW_EXPCLEAR(1);
2425 /* Use all of the htile_buffer for depth if there's no stencil. */
2426 ds
->db_stencil_info
|= S_028044_TILE_STENCIL_DISABLE(1);
2428 va
= device
->ws
->buffer_get_va(iview
->bo
) + iview
->image
->offset
+
2429 iview
->image
->htile_offset
;
2430 ds
->db_htile_data_base
= va
>> 8;
2431 ds
->db_htile_surface
= S_028ABC_FULL_CACHE(1);
2433 ds
->db_htile_data_base
= 0;
2434 ds
->db_htile_surface
= 0;
2437 ds
->db_z_read_base
= ds
->db_z_write_base
= z_offs
>> 8;
2438 ds
->db_stencil_read_base
= ds
->db_stencil_write_base
= s_offs
>> 8;
2440 ds
->db_depth_size
= S_028058_PITCH_TILE_MAX((level_info
->nblk_x
/ 8) - 1) |
2441 S_028058_HEIGHT_TILE_MAX((level_info
->nblk_y
/ 8) - 1);
2442 ds
->db_depth_slice
= S_02805C_SLICE_TILE_MAX((level_info
->nblk_x
* level_info
->nblk_y
) / 64 - 1);
2445 VkResult
radv_CreateFramebuffer(
2447 const VkFramebufferCreateInfo
* pCreateInfo
,
2448 const VkAllocationCallbacks
* pAllocator
,
2449 VkFramebuffer
* pFramebuffer
)
2451 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2452 struct radv_framebuffer
*framebuffer
;
2454 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
2456 size_t size
= sizeof(*framebuffer
) +
2457 sizeof(struct radv_attachment_info
) * pCreateInfo
->attachmentCount
;
2458 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
2459 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2460 if (framebuffer
== NULL
)
2461 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2463 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
2464 framebuffer
->width
= pCreateInfo
->width
;
2465 framebuffer
->height
= pCreateInfo
->height
;
2466 framebuffer
->layers
= pCreateInfo
->layers
;
2467 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
2468 VkImageView _iview
= pCreateInfo
->pAttachments
[i
];
2469 struct radv_image_view
*iview
= radv_image_view_from_handle(_iview
);
2470 framebuffer
->attachments
[i
].attachment
= iview
;
2471 if (iview
->aspect_mask
& VK_IMAGE_ASPECT_COLOR_BIT
) {
2472 radv_initialise_color_surface(device
, &framebuffer
->attachments
[i
].cb
, iview
);
2473 } else if (iview
->aspect_mask
& (VK_IMAGE_ASPECT_DEPTH_BIT
| VK_IMAGE_ASPECT_STENCIL_BIT
)) {
2474 radv_initialise_ds_surface(device
, &framebuffer
->attachments
[i
].ds
, iview
);
2476 framebuffer
->width
= MIN2(framebuffer
->width
, iview
->extent
.width
);
2477 framebuffer
->height
= MIN2(framebuffer
->height
, iview
->extent
.height
);
2478 framebuffer
->layers
= MIN2(framebuffer
->layers
, radv_surface_layer_count(iview
));
2481 *pFramebuffer
= radv_framebuffer_to_handle(framebuffer
);
2485 void radv_DestroyFramebuffer(
2488 const VkAllocationCallbacks
* pAllocator
)
2490 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2491 RADV_FROM_HANDLE(radv_framebuffer
, fb
, _fb
);
2495 vk_free2(&device
->alloc
, pAllocator
, fb
);
2498 static unsigned radv_tex_wrap(VkSamplerAddressMode address_mode
)
2500 switch (address_mode
) {
2501 case VK_SAMPLER_ADDRESS_MODE_REPEAT
:
2502 return V_008F30_SQ_TEX_WRAP
;
2503 case VK_SAMPLER_ADDRESS_MODE_MIRRORED_REPEAT
:
2504 return V_008F30_SQ_TEX_MIRROR
;
2505 case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE
:
2506 return V_008F30_SQ_TEX_CLAMP_LAST_TEXEL
;
2507 case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_BORDER
:
2508 return V_008F30_SQ_TEX_CLAMP_BORDER
;
2509 case VK_SAMPLER_ADDRESS_MODE_MIRROR_CLAMP_TO_EDGE
:
2510 return V_008F30_SQ_TEX_MIRROR_ONCE_LAST_TEXEL
;
2512 unreachable("illegal tex wrap mode");
2518 radv_tex_compare(VkCompareOp op
)
2521 case VK_COMPARE_OP_NEVER
:
2522 return V_008F30_SQ_TEX_DEPTH_COMPARE_NEVER
;
2523 case VK_COMPARE_OP_LESS
:
2524 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESS
;
2525 case VK_COMPARE_OP_EQUAL
:
2526 return V_008F30_SQ_TEX_DEPTH_COMPARE_EQUAL
;
2527 case VK_COMPARE_OP_LESS_OR_EQUAL
:
2528 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESSEQUAL
;
2529 case VK_COMPARE_OP_GREATER
:
2530 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATER
;
2531 case VK_COMPARE_OP_NOT_EQUAL
:
2532 return V_008F30_SQ_TEX_DEPTH_COMPARE_NOTEQUAL
;
2533 case VK_COMPARE_OP_GREATER_OR_EQUAL
:
2534 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATEREQUAL
;
2535 case VK_COMPARE_OP_ALWAYS
:
2536 return V_008F30_SQ_TEX_DEPTH_COMPARE_ALWAYS
;
2538 unreachable("illegal compare mode");
2544 radv_tex_filter(VkFilter filter
, unsigned max_ansio
)
2547 case VK_FILTER_NEAREST
:
2548 return (max_ansio
> 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_POINT
:
2549 V_008F38_SQ_TEX_XY_FILTER_POINT
);
2550 case VK_FILTER_LINEAR
:
2551 return (max_ansio
> 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_BILINEAR
:
2552 V_008F38_SQ_TEX_XY_FILTER_BILINEAR
);
2553 case VK_FILTER_CUBIC_IMG
:
2555 fprintf(stderr
, "illegal texture filter");
2561 radv_tex_mipfilter(VkSamplerMipmapMode mode
)
2564 case VK_SAMPLER_MIPMAP_MODE_NEAREST
:
2565 return V_008F38_SQ_TEX_Z_FILTER_POINT
;
2566 case VK_SAMPLER_MIPMAP_MODE_LINEAR
:
2567 return V_008F38_SQ_TEX_Z_FILTER_LINEAR
;
2569 return V_008F38_SQ_TEX_Z_FILTER_NONE
;
2574 radv_tex_bordercolor(VkBorderColor bcolor
)
2577 case VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
:
2578 case VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
:
2579 return V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK
;
2580 case VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
:
2581 case VK_BORDER_COLOR_INT_OPAQUE_BLACK
:
2582 return V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_BLACK
;
2583 case VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
:
2584 case VK_BORDER_COLOR_INT_OPAQUE_WHITE
:
2585 return V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_WHITE
;
2593 radv_tex_aniso_filter(unsigned filter
)
2607 radv_init_sampler(struct radv_device
*device
,
2608 struct radv_sampler
*sampler
,
2609 const VkSamplerCreateInfo
*pCreateInfo
)
2611 uint32_t max_aniso
= pCreateInfo
->anisotropyEnable
&& pCreateInfo
->maxAnisotropy
> 1.0 ?
2612 (uint32_t) pCreateInfo
->maxAnisotropy
: 0;
2613 uint32_t max_aniso_ratio
= radv_tex_aniso_filter(max_aniso
);
2614 bool is_vi
= (device
->physical_device
->rad_info
.chip_class
>= VI
);
2616 sampler
->state
[0] = (S_008F30_CLAMP_X(radv_tex_wrap(pCreateInfo
->addressModeU
)) |
2617 S_008F30_CLAMP_Y(radv_tex_wrap(pCreateInfo
->addressModeV
)) |
2618 S_008F30_CLAMP_Z(radv_tex_wrap(pCreateInfo
->addressModeW
)) |
2619 S_008F30_MAX_ANISO_RATIO(max_aniso_ratio
) |
2620 S_008F30_DEPTH_COMPARE_FUNC(radv_tex_compare(pCreateInfo
->compareOp
)) |
2621 S_008F30_FORCE_UNNORMALIZED(pCreateInfo
->unnormalizedCoordinates
? 1 : 0) |
2622 S_008F30_ANISO_THRESHOLD(max_aniso_ratio
>> 1) |
2623 S_008F30_ANISO_BIAS(max_aniso_ratio
) |
2624 S_008F30_DISABLE_CUBE_WRAP(0) |
2625 S_008F30_COMPAT_MODE(is_vi
));
2626 sampler
->state
[1] = (S_008F34_MIN_LOD(S_FIXED(CLAMP(pCreateInfo
->minLod
, 0, 15), 8)) |
2627 S_008F34_MAX_LOD(S_FIXED(CLAMP(pCreateInfo
->maxLod
, 0, 15), 8)) |
2628 S_008F34_PERF_MIP(max_aniso_ratio
? max_aniso_ratio
+ 6 : 0));
2629 sampler
->state
[2] = (S_008F38_LOD_BIAS(S_FIXED(CLAMP(pCreateInfo
->mipLodBias
, -16, 16), 8)) |
2630 S_008F38_XY_MAG_FILTER(radv_tex_filter(pCreateInfo
->magFilter
, max_aniso
)) |
2631 S_008F38_XY_MIN_FILTER(radv_tex_filter(pCreateInfo
->minFilter
, max_aniso
)) |
2632 S_008F38_MIP_FILTER(radv_tex_mipfilter(pCreateInfo
->mipmapMode
)) |
2633 S_008F38_MIP_POINT_PRECLAMP(0) |
2634 S_008F38_DISABLE_LSB_CEIL(1) |
2635 S_008F38_FILTER_PREC_FIX(1) |
2636 S_008F38_ANISO_OVERRIDE(is_vi
));
2637 sampler
->state
[3] = (S_008F3C_BORDER_COLOR_PTR(0) |
2638 S_008F3C_BORDER_COLOR_TYPE(radv_tex_bordercolor(pCreateInfo
->borderColor
)));
2641 VkResult
radv_CreateSampler(
2643 const VkSamplerCreateInfo
* pCreateInfo
,
2644 const VkAllocationCallbacks
* pAllocator
,
2645 VkSampler
* pSampler
)
2647 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2648 struct radv_sampler
*sampler
;
2650 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO
);
2652 sampler
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*sampler
), 8,
2653 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2655 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2657 radv_init_sampler(device
, sampler
, pCreateInfo
);
2658 *pSampler
= radv_sampler_to_handle(sampler
);
2663 void radv_DestroySampler(
2666 const VkAllocationCallbacks
* pAllocator
)
2668 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2669 RADV_FROM_HANDLE(radv_sampler
, sampler
, _sampler
);
2673 vk_free2(&device
->alloc
, pAllocator
, sampler
);
2677 /* vk_icd.h does not declare this function, so we declare it here to
2678 * suppress Wmissing-prototypes.
2680 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2681 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t *pSupportedVersion
);
2683 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2684 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t *pSupportedVersion
)
2686 /* For the full details on loader interface versioning, see
2687 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
2688 * What follows is a condensed summary, to help you navigate the large and
2689 * confusing official doc.
2691 * - Loader interface v0 is incompatible with later versions. We don't
2694 * - In loader interface v1:
2695 * - The first ICD entrypoint called by the loader is
2696 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
2698 * - The ICD must statically expose no other Vulkan symbol unless it is
2699 * linked with -Bsymbolic.
2700 * - Each dispatchable Vulkan handle created by the ICD must be
2701 * a pointer to a struct whose first member is VK_LOADER_DATA. The
2702 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
2703 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
2704 * vkDestroySurfaceKHR(). The ICD must be capable of working with
2705 * such loader-managed surfaces.
2707 * - Loader interface v2 differs from v1 in:
2708 * - The first ICD entrypoint called by the loader is
2709 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
2710 * statically expose this entrypoint.
2712 * - Loader interface v3 differs from v2 in:
2713 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
2714 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
2715 * because the loader no longer does so.
2717 *pSupportedVersion
= MIN2(*pSupportedVersion
, 3u);