Contains optimizations of SWT matrix-vector multiply, on order of 3-4X for the example I tried. Speedup comes from avoiding repeated calculation of some q-independent data and from eliminating allocations due to reshape. Vectorization is impeded by memory access patterns. Future work could either be directed toward fine-grained memory work to improve access patterns, or possibly by going to real space and doing a sparse matrix-vector multiply.
The allocating functions multiply_by_hamiltonian_dipole(x, swt, qs) and multiply_by_hamiltonian_SUN(x, swt, qs) have been added. These should not be benchmarked directly. Instead, the contents of these functions after allocations should be benched, specifically these lines and their SU(N) analogs:
qphase = map(qs_reshaped) do q
(exp(2π*im*q[1]), exp(2π*im*q[2]), exp(2π*im*q[3]))
end
multiply_by_hamiltonian_dipole_aux!(reshape(y, (length(qs_reshaped), size(x, 2))), x, phasebuf, qphase, swt)
Contains optimizations of SWT matrix-vector multiply, on order of 3-4X for the example I tried. Speedup comes from avoiding repeated calculation of some q-independent data and from eliminating allocations due to
reshape. Vectorization is impeded by memory access patterns. Future work could either be directed toward fine-grained memory work to improve access patterns, or possibly by going to real space and doing a sparse matrix-vector multiply.The allocating functions
multiply_by_hamiltonian_dipole(x, swt, qs)andmultiply_by_hamiltonian_SUN(x, swt, qs)have been added. These should not be benchmarked directly. Instead, the contents of these functions after allocations should be benched, specifically these lines and their SU(N) analogs: