Use hcat to concatenate basis.snapshots or explicity compute matrix elements?

[pbrehmer]

In compress and estimate_gs we encounter cases where we need to concatenate the basis snapshot vectors into one matrix in order to perform matrix multiplications.

The questions is whether it is better to:

• use hcat and then use BLAS Level 3 operations but allocate more memory
• or perform explicit matrix-vector BLAS Level 2 operations but perform less allocations

We need to benchmark this for realistic cases.