Charlie Debus

Implement Nystrom method for better scaling behaviour

3

Current Spectral Clsutering algorihm is based on full distance matrix calculation and Lanczos algorithm for Eigenvalue decomposition. Whilst this provides accurate result at acceptable speed, the Nystrom method via approximating...

The dndarray ``` __getitem__ ``` function appeared to be rather slow in the Gram-Schmitt orthogonalization of the Lanczos algorithm (linalg.py): ``` for i in Lanczos_Iterations: ... for j in range(i)...

For now, the docs of the spatial.distance module is missing examples.

MPI collectives buffer prep with specified counts and displacements

2

**Related** Enhancement of the MPI buffer preparation. Related to Issue #341 and PR #364 **Feature functionality** Currently, counts and displacements for custom tensor distribution can be passed on but are...

Rework MPI communication Wrappers

1

**Related** Related to Issue #233 : Wrong sorting in dndarray resplit was due to errorous allgatherv **Feature functionality** Refactor MPI global functions: allgather*, alltoall*, scatter*, gather*. So far algorithmic of...

Wiki pages / GitHub pages describing architecture designs/assumptions

3

Create some wiki pages that describe (maybe also with a couple of graphics) the concept of the HeAT design, assumptions made in tensor and MPI handling and architectural decisions.