v0.96.2 2026/03

Fixes

• Explicitly enforce rank 0 as the source of truth for parallel HDF step attribute writes.

v0.96.1 2026/03

Fixes compilation issues with CUDA 13 and OpenMPI

v0.96 2026/03

Features

• Add option to remove particles for which neighbor search did not converge

Improvements

• Separate halo and MAC peer rank lists
• Fully converge global tree when needed
• Replace focusTransfer and macRefinement with multi-level LET updates and update LET until converged
• Improve performance of segmented reductions with multiple threads per segment
• Eliminate duplicated data structure for CPU and GPU particles data

Fixes

• Add rocthrust and hipcub dependencies in CMake to fix HIP spack package build
• Add custom MPI reduction to prevent overflows during initialization

v0.95 2025/10

LET improvements

• Enforce global octree keys in LET
• find halos with tight interaction boxes
• unify mac and halo flags

Performance improvements

• Avoid duplication of particle data fields on the CPU and GPU
• find halos with a) tight and b) fp interaction boxes
• stackless DFS traversals
• GPU-direct MPI communication for all LET parts, including globals

Features and enhancements:

• Replaced h5part with h5hut
• Improved profiling framework, capturing additional information
• replaced gsl::span with std::span

HIP and Spack

• HIP compatibility without requiring the source code to be hipified
• CMake changes to allow easier integration with Spack

Propagator library

Enhancements:

• Separate library with a translation unit for each propagator to reduce compilation times

Fixes:

• Prevent GPU kernel launches with 0 thread blocks which started to be an issue with CUDA 12.6

CUDA 12.5 compatibility

Fixes:

• Full encapsulation of thrust::device_vector, because starting from CUDA 12.5 inclusion of its in .cpp files is no longer possible

Dynamic LET surface refinement and node pruning

New features:

• Refine LET resolution at surface after domain boundary changes
• Prune LET nodes outside focus that exceed the LET resolution on the owning rank

Hierarchical block time steps

New features:

• Hierarchical block time stepping

Ewald summation

New features:

• Ewald summation on CPUs and GPUs for gravitational forces with periodic boundarys
• New smoothing kernel for SPH: S49

Performance enhancements:

• Improve tree refinement for remote LET nodes such that fewer remote nodes are needed to ensure successful gravity traversal. Improves performance due to smaller amount of communication needed
• injectKeys on GPUs. This tree resolution-enforcement mechanism is needed more frequently than previously thought,
  hence it made sense to port it to GPU.

Fixes:

• Fix compilation issues with CUDA 12.4 related to thrust::device_vector