Description
The current CUDA-only volume allocation delegates topology construction to NanoVDB's PointsToGrid, which requires host count copies and stream synchronization during construction. This makes it unsuitable for CUDA graph capture.
A new implementation should allow over-provisioned allocations, with later rebuilds writing into fixed-capacity NanoVDB buffers, keeping actual topology counts on device, and reporting capacity or input issues through status bits.
Context
Graph-capturable volume allocation is particularly interesting in the context of Newtons' MPM solver, where new volumes are built from particles at each step
Description
The current CUDA-only volume allocation delegates topology construction to NanoVDB's PointsToGrid, which requires host count copies and stream synchronization during construction. This makes it unsuitable for CUDA graph capture.
A new implementation should allow over-provisioned allocations, with later rebuilds writing into fixed-capacity NanoVDB buffers, keeping actual topology counts on device, and reporting capacity or input issues through status bits.
Context
Graph-capturable volume allocation is particularly interesting in the context of Newtons' MPM solver, where new volumes are built from particles at each step