4.2 KiB
Minutes from 83rd OpenVDB TSC meeting, March 16th, 2021, (EDT)
Attendees: Nick A., Jeff L., Ken M., Dan B, Andre P.
Additional Attendees: Johannes Meng (Intel), Robin Rowe (Cinepaint) Bruce Chernaik (Intel), JT Nelson (Blender)
Agenda:
-
Confirm Quorum
-
Secretary
-
Forum
-
Website
-
VDB Unit test speeds
-
CMake multi config changes PR # 1003
-
VDB Points mem leak
-
Point statistics PR # 997
-
Header vs CMake defines
-
PR994 / Release 7.2.3
-
DynamicNodeManager and Const Trees (PR996)
-
AX and Active Tiles
-
Half Limits
-
Exrapolate SOP
-
Next Meeting
-
Confirm Quorum
Quorum is present.
- Secretary
Secretary is Nick Avramoussis.
- Forum
RFE #1002 has been filed about possibly extending VDB tree structures with custom allocators. Ken, has investigated such techniques in the past and used to have an implementation, but was essentially unable to get better performance than using a custom malloc (i.e. jemalloc). Jeff, opposed, such extensions introduce a lot of complexity (e.g. the stl...) and have historically been introduced due to limitations in malloc. Jeff, this technique is better suited to custom hardware allocations (host vs device). Ken, nanovdb uses its own structure now so this wouldn't be as useful. Summary, TSC would be interested to see if an implementation can be constructed which demos significant improvements over allocators such as jemalloc, but otherwise will most likely not be actively targeted by the TSC.
- Website
Ken has updated some logos. We should updated the about section to reference the TSC.
- VDB Unit test speeds
The VDB Unit tests take ~3.5 hours to run in Debug. 50% of this time is the FastSweeping unit tests. The top three slowest tests are TestFastSweeping (1.5h), TestGridTransformer (30min) and TestDiagnostics (25min). We should improve these.
- CMake multi config changes PR # 1003
Nick, work to improve OpenVDB's CMake build system for generators that support multiple configurations. Jeff, recently discovered this issue in Houdini where release libs are linking against debug libs and conversely different CRTs on Windows. Nick to finalize PR and merge.
- VDB Points mem leak
Nick, discovered potential memory leak in VDB Points when delay loading if off. Dan to investigate and confirm - suggested fix is a one liner which should hopefully be applicable.
- Point statistics PR # 997
Nick, point statistics PR is up which is a pre-requisite for the surfacing. Dan, has commented, main comment to discuss is precision of intermediate arithmetic on average/accumulates. All agreed that this should be done at the highest applicable precision (double for floating point, 64bit ints otherwise). Nick to address.
- Header vs CMake defines
OpenVDB CMake build system adds a variety of compile time defines, as does the FindOpenVDB.cmake module. Excluding these defines would otherwise cause build failures. This, however, makes downstream software reliant on using the FindOpenVDB.cmake module (i.e. downstream software has to build with cmake). This is an unnecessary limitation and we should instead look to move some of these defines to headers (PlatformConfig/BuildConfig etc). Nick to investigate.
- PR994 / Release 7.2.3
Dan, looking to release 7.2.3, Jeff to approve and then can merge.
- DynamicNodeManager and Const Trees (PR996)
Dan, updated the dynamic node manager to work with const trees. Observed a minor performance improvement whilst doing so, but main instigation was to be able to port existing const node manager methods (such as activeVoxelCount).
- AX and Active Tiles
Dan, observed significant performance loss on heavily tiled VDB grids using AX. This is due to the full expansion/pruning pre and post processing steps. A better way would be to stream the tiled data and use better heuristics to avoid this streaming. Nick, already have a roadmap for siggraph but will see if we can look to add this too.
- Half Limits
Dan, should we be including the half limits header in Types.h and what's the correct name-spacing to use. Jeff, not sure, we should add a unit test to check we don't run into ODR/name lookup issues.
- Extrapolate SOP
Jeff to review.
- Next Meeting
Next meeting is March 23rd, 2021. 1pm-2pm EST (GMT-4).