Writing wrappers for LAPACK and BLAS routines

[ivanpribec]

@fracton

I am actually running this on a larger machine (a node with 2 CPUs) with about 500Gb and intend to run it with much large size matrices (>50k). However, there’s another limitation using contiguous arrays, since 500Gb is divide into 250Gb for each CPU with NUMA. In that case, the only option to scale this up then would be to store the matrices on disk.

How about using Parallel BLAS?

! Standard BLAS
call  zgemv(trans, m, n, alpha, a, lda, &
                                x, incx, &
                          beta, y, incy)

! Parallel BLAS
call pzgemv(trans, m, n, alpha, a, ia, ja, desca, & 
                                x, ix, jx, descx, incx, &
                          beta, y, iy, jy, descy, incy)

You will need to establish the distributed arrays and descriptors at first, but otherwise your code can remain mostly the same.

More info:

• PBLAS | Netlib
• ScaLAPACK Array Descriptors | Intel MKL