gpusolver

CUDA linear solver with python API

This is a general purpose linear solver based on CUDA libraries with python API.

Requirements:

CUDA8.0 (cublas, cudart, cusparse, cusolver)
python2.7( numpy, scipy, cython)

To begin navigate to a directory:

git clone https://github.com/yunfanz/gpusolver.git ./
cd ./gpusolver
python setup.py build_ext -i

Features

Current features include:

1. single precision float solver for 2 dimensional matrix and 1 dimensional right hand side
2. Passing data in either dense or csr format
3. AtAx=Atb solver for general/non-square matrix A; Ax=b solver for square matrices
4. Four options of kernel backends (0:QR decomposition, 1:Cholesky decomposition, 2:LU, 3:SVD)

Features to be added:

1. Complex solver
2. Batched transfer

General structure of usage:

1. Initialize solver by calling solver = gpusolver.DnSolver(num_rows, num_cols)
2. Solve using either solver.solve(n) or solver.solve_Axb(n)
3. Retrieve the result x = solver.retrieve()

Sample usage is given in unit_test.py

Performance suggestions:

GPU code performance is a balance of data transfer and kernel execution. This implementation does not overlap the two, so here's some guidelines in choosing what to use:

size\sparsity	Sparsity>0.5	Sparsity<0.5
small A	Compare cpu vs. from_csr	Compare cpu vs. from_dense
large nrows>>ncols	from_csr, solve<------	------> AtA on cpu, solve_Axb
large A	from_csr	from_csr vs. from_dense