- RHICreate{Vertex, Index, Structured}Buffer
- RHICreate{ShaderResource, UnorderedAccess}View
- RHIUpdateUniformBuffer
- Various initialization / locking methods for helper buffer types in RHIUtilities.h
The goal is to continue to force resource creation through command lists to avoid surprises with moving things off the render thread.
#rb christopher.waters
[CL 26183746 by zach bethel in 5.3 branch]
- added new runtime field (NumStreamedClusters) to Nanite::FResources
- modified logic in Nanite::FStreamingManager to update number of streamted clusters when pages are installed/uninstalled.
- modified NaniteStreamOut to output number of clusters to be able to check for incorrect count on CPU side.
- On CitySample, peak NaniteRayTracing.StagingAuxiliaryDataBuffer size is reduced from close to 1GB to about 200MB.
#rb rune.stubbe
#preflight 646e3399f85111e06ca30efa
[CL 25607114 by tiago costa in ue5-main branch]
Fix crash in FNaniteGeometryCollectionSceneProxy creation.
Issue seems to have been from bad merge of work in other stream.
#preflight 63f905fac35a141980d6a86c
[CL 24405393 by Jeremy Moore in ue5-main branch]
Fixes crash caused by scene proxy recreating without the required render data update.
#preflight 63efee4713586f2c103e2673
[CL 24295291 by jeremy moore in ue5-main branch]
This is built as part of geometry collection DDC contents.
Reworked scene proxy to use this data. Some big changes there:
* Tidied up hit proxy code so that all hit proxy clients use the same path.
* Removing of internal faces are now handled without an additional index buffer by always storing internal faces at the end of a section and providing different section descriptions for including internal faces.
* Dynamic geometry hiding in editor is now done by zeroing transforms instead of rebuilding the index buffer.
#rb cedric.caillaud
#preflight 63edc2e55c7bd278c11efe51
[CL 24261113 by jeremy moore in ue5-main branch]
* Primitives are only tracked by Lumen scene if they can be traced by the current ray tracing method. E.g. if we use HWRT then only primitives with valid BLAS will generate surface cache.
* Instead of recomputing VisibleInLumenScene on every add/remove/update this flag is now cached on proxy
* Added GPU side filtering of draws which should be skipped when capturing surface cache
#preflight 63dd8637323e7eaa30b71680
#jira UE-176073
#rb Daniel.Wright
[CL 24031439 by krzysztof narkowicz in ue5-main branch]
- Removing MaterialAttributeDefinitionMap and MaterialRenderProxy from MaterialShared.h
- Removing MaterialShared from Material.h
#preflight 639cbb35776b61ba3b82f03e
[CL 23541603 by christopher waters in ue5-main branch]
* Remove the need to call BuildInstanceMaskAndFlags(), BuildRayTracingInstanceMaskAndFlags() manually in GetXXRayTracingInstance(). They will be called in the renderer module. All functions manually called has been removed.
* Remove the RENDER_API public api for those functions, as they cause link error in the Engine module (where the header is included in the Engine build file for static library link. the RENDERER_API is interpreted as dllimport, which will not link the functions in RayTracingInstance.cpp) and there is no need to call them outside of renderer module.
* A callback function is added to FRayTracingInstance, use it for RaytracingInstanceMaskandFlags caching.
Note: this is a first step to separate the instance mask between RayTracing and PathTracing, where PathTracing uses a totally different instance mask. This cleanup is un-avoidable if we want to keep rendering related code in the renderer module.
#rb Yuriy.ODonnell
#jira UE-172742
#preflight 6399002f2960b732202acc7c
[CL 23514248 by tiantian xie in ue5-main branch]
On local testing, average performance of the function went from 27ms to 11ms
- Changed UVs data layout in FGeometryCollectionConstantData to be an array of UVChannels instead of an array of vertices containing each multiple channels, this reduces allocations quite a lot
- Avoid memcpy by passing array by reference to BuildMeshSections
- Use batched parallelFor for copying bone colors and UVChannels
- Rewrite indices remap to be O(N + M) instead of O(N * M)
Most of the remaining cost is in copying the arrays now
Future work could include making a copy per thread ( but may have to keep allocation on the calling thread )
#rb jimmy.andrews, brice.criswell, Bryan Sefcik
#preflight 6387971f3377450900d27163
[CL 23340834 by cedric caillaud in ue5-main branch]
- Optimize managed array collection NumElement to only fetch once from the group map
- Make sure to skip IncrementSleepTimer and ImcrementBreakTimer method is GC is not broken yet
- Early out in RefreshEmbeddedGeometry if no examplar is present
- Optimize BuildGeometry to reduce cache missed when running too small batch across large number of threads ( 2.5 time perf boost )
#rb benn.gallagher
#preflight 637531c6953c19d43519551b
[CL 23166264 by cedric caillaud in ue5-main branch]
* GeometryCollectionComponent.h - Most files are related to the removal of includes in this file. Removed 7 headers
- Moved GeometryCollectionDamagePropagationData in to its own file
- Changed COPY_ON_WRITE_ATTRIBUTE to be set in a way so implementations can be placed in cpp file.
- Moved some function implementations to cpp file
* Removed include in FieldSystemActor.h
* Removed include in FieldSystem.h
* Removed include in CommonUISettings.h
#preflight 636b325c376a9cd6a8818aae
#rb cedric.caillaud (for the actual changes, not additional includes and a couple include removals outside GeometryCollection)
[CL 23069399 by henrik karlsson in ue5-main branch]
