Shared-memory implementation of parallel sparse matrix vector product in Julia. We thank Roman Shekhtman (UBC) for providing the Fortran code.
To install on a unix machine, follow these steps
The first line downloads the package and (on unix) compiles the Fortran codes (
gfortran is used by default). Currently there is no automatic build procedure for Windows. Pull requests are welcome.
The second line tests the package.
Currently, we do not overload the matrix vector product in
Base (this might be added in the future). Let
A be a sparse matrix,
beta floating point numbers and
y be real- or complex values vectors of appropriate size. Then, the following commands are equivalent
nproc = 4; # choose number of OMP threads yt = copy(y) y = beta*y + alpha * A*x ParSpMatVec.A_mul_B!( alpha, A, x, beta, yt, nproc)
Similarly, for the transpose matrix-vector product:
yt= copy(y) y = beta*y + alpha * A'*x ParSpMatVec.Ac_mul_B!( alpha, A, x, beta, yt, nproc)
The last input,
nproc, determines how many OpenMP threads are used. Note that, due to the compressed column storage, products with the adjoint of
A are expected to scale better.
A few things to do:
3 months ago