/Hubbard-GPU

Calculate the groundstate energy of 1D and 2D Fermi-Hubbard model on the GPU with Cuda.

Primary LanguageC++GNU General Public License v3.0GPL-3.0

Fermi-Hubbard on the GPU with Cuda

This program can calculate the groundstate energy of 1 dimensional and 2 dimension Fermi-Hubbard model on the GPU (and the CPU). It is based on Exact diagonalization of the Hubbard model on graphics processing units by Siro and Harju.

There are 2 programs: main and main2D. The former is for 1D Hubbard, the latter is for 2D Hubbard. Everything is split up in several classes to make reusing code easy: The Hamiltonian and HubHam2D classes build the full (dense) Hamiltonian Matrix. SparseHamiltonian stores the Hubbard Hamiltonian in parts: an spin up and a spin down part. The matrix themselves are storred in the ELL format. The SparseHamiltonian2D does the same but for 2D Hubbard. However, here we make a detour: we first store the matrices in the CRS format (the SparseHamiltonian2DCSR class) and then convert it in the ELL format. The reason is that for ELL, we need to know the maximum number of non-zero elements (nnz) of a row.

There are several branches in the git repo: the master contains only the CPU version. The branch 'GPU' constains the GPU version and the branch 'PRIMME' used the PRIMME library to find the eigenvalues and eigenvectors. You can find PRIMME at http://www.cs.wm.edu/~andreas/software/

All code is under the GPLv3.

Symmetry

I've added support for momentum symmetry and spin symmetry. The momentum works great but spin symmetry is slow because of the many operations needed to build the hamiltonian. The GPU currently doesn't make use of any of these symmetries.

Documentation

All code is documented with doxygen. The full docs can be generate or read online.