We are investigating issues related to interfacing ScaLAPACK with High Performance Fortran (HPF) . As a part of this effort, we have provided prototype interfaces to some of the ScaLAPACK routines. We are collecting user feedback on these codes, as well as allowing additional time for compiler maturation, before providing a more complete interface.
Initially, interfaces are provided for the following ScaLAPACK routines: the general and symmetric positive definite linear equation solvers (PxGESV and PxPOSV), the linear least squares solver (PxGELS), and the PBLAS matrix multiply routine (PxGEMM).
LA_GESV(A, B, IPIV, INFO) TYPE, intent(inout), dimension(:,:) :: A, B integer, optional, intent(out) :: IPIV(:), INFO LA_POSV(A, B, UPLO, INFO) TYPE, intent(inout), dimension(:,:) :: A, B character(LEN=1), optional, intent(in) :: UPLO integer, optional, intent(out) :: INFO LA_GELS(A, B, TRANS, INFO) TYPE, intent(inout), dimension(:,:) :: A, B character(LEN=1), optional, intent(in) :: TRANS integer, optional, intent(out) :: INFO LA_GEMM(A, B, C, transA, transB, alpha, beta) TYPE, intent(in), dimension(:,:) :: A, B TYPE, intent(inout), dimension(:,:) :: C character(LEN=1), optional, intent(in) :: transA, transB TYPE, optional, intent(in) :: alpha, beta
With this interface, all matrices are inherited, and query functions are used to determine the distribution of the matrices. Only when ScaLAPACK cannot handle the user's distribution are the matrices redistributed. In such a case, it is done transparently to the user, and only performance will show that it has occurred.
The prototype interfaces can be downloaded from netlib at the following URL:
Questions or comments on these routines may be mailed to firstname.lastname@example.org.
The following example code is a complete HPF code calling and testing the ScaLAPACK LU factorization/solve in HPF.
This program is also available in the scalapack directory on netlib
program simplegesv use HPF_LAPACK integer, parameter :: N=500, NRHS=20, NB=64, NBRHS=64, P=1, Q=3 integer, parameter :: DP=kind(0.0D0) integer :: IPIV(N) real(DP) :: A(N, N), X(N, NRHS), B(N, NRHS) !HPF$ PROCESSORS PROC(P,Q) !HPF$ DISTRIBUTE A(cyclic(NB), cyclic(NB)) ONTO PROC !HPF$ DISTRIBUTE (cyclic(NB), cyclic(NBRHS)) ONTO PROC :: B, X ! ! Randomly generate the coefficient matrix A and the solution ! matrix X. Set the right hand side matrix B such that B = A * X. ! call random_number(A) call random_number(X) B = matmul(A, X) ! ! Solve the linear system; the computed solution overwrites B ! call la_gesv(A, B, IPIV) ! ! As a simple test, print the largest difference (in absolute value) ! between the computed solution (B) and the generated solution (X). ! print*,'MAX( ABS(X~ - X) ) = ',maxval( abs(B - X) ) ! ! Shutdown the ScaLAPACK system, I'm done ! call SLhpf_exit() stop end