Closed ffreyer closed 3 years ago
Apparently rvmul!
and lvmul!
are almost an order of magnitude worse than just vmul!
, so I'll try to phase these out too.
BlockDiagonal
works with the repulsive Hubbard model and results in up to 4x faster matrix multiplications, given large enough matrices (like 8+ sites on my AVX2 machine). The speed up for the whole simulation will be lower, but not drastically. measurements should be slower, due to indexing.
Closes #80
Implements optimizations for #78 and opens up more optimizations such as using
Diagonal
matrices as the interaction matrix or introducing specialized matrix types such asBlockDiagonal
matrices for greens function etc.