Merge pull request #1649 from Idclip/tsc_notes

TSC minutes

Author: Idclip

tsc/meetings/2023-07-25.md

@@ -0,0 +1,78 @@
+Minutes from OpenVDB TSC meeting, July 25th, 2023
+
+Attendees: *Ken* M., *Jeff* L., *Andre* P, *Dan* B., *Rich* J., *Nick* A.
+
+Regrets:
+Attendees: *Greg* H.,  *Rich* J.
+
+Additional Attendees: John Mather (SideFX)
+
+Agenda:
+
+1) Confirm quorum
+2) Secretary
+3) PR #1645
+4) PR #1525
+5) Read Only Grid
+6) PR #1634
+7) PR #1648
+8) Next meeting
+
+------------
+
+1) Confirm quorum
+
+Quorum is present.
+
+2) Secretary
+
+Secretary is Nick Avramoussis
+
+3) PR #1645
+
+We're emitting x86 compiler flags on ARM which obviously doesn't work. This PR
+disables this flag but keeps the defines. But the defines are X86 specific so it's
+a bit odd. What's really odd is that we're guarding compiler builtins behind the
+defines which doesn't make any sense. We should choose to either use the X86 intrinsic
+directly with -mpopcnt or always use the compiler builtins (if available) and
+otherwise fallback to the software impl. Some compiler builtins have different
+impls, for example popcount will fall back to multiple shifts rather than a
+lookup table (which is what we have as our software fallback, this may be faster).
+Additionally we don't propagate these defines downstream so they're a bit useless.
+Anyway, this PR works fine but we should revisit this.
+
+4) PR #1525
+
+Hollow mesh stuff, ready to go. Nick has some minor comments but looks good,
+should be merged.
+
+5) Read Only Grid
+
+Ken working on pushing up his Read Only grid implementation at request.
+
+6) PR #1634
+
+Anisotropic surfacing kernels and PCA methods. Nick will be talking about this
+at Siggraph along with vectorization stuff. Idea from Jeff to build covariance
+matrices per voxel rather than per point and then lerp/box sample onto the
+points to help with performance issues. Definitely something to consider
+as it might be a better default.
+
+7) PR #1648
+
+Vectorization framework for X86 which incorporates Agner Fogs' SIMD intrinsic
+wrappers and vectorizes the majority of the PointTransfer framework. John
+provided a link to some other library comparisons:
+
+https://www.mnm-team.org/pub/Fopras/rock23/PDF-Version/rock23.pdf
+
+Mentioned Highway as an alternative to also support ARM. Nick mentioned that
+the current use case is very much a manual and instruction specific approach
+to vectorizaton where as Highway looked more like a complete framework. Would
+like to focus on X86 for now and have accepted there will be some duplicate
+work in the future to support ARM, but we can plan for this with some API
+design.
+
+8) Next meeting
+
+Next meeting is on Aug 1st, 2023. 2pm-3pm EDT.