0

I'm trying to write a program that is going to use both Arpack-ng and MKL routines. I was testing the behavior by running a simple code which use a single call to the function dsaupd, which should return ido=-1 and info = 0, but it actually returned ido = 99 and info=4294967286.

This is the Arpack-ng build summary:

-- Configuration summary for arpack-ng-3.9.0:
   -- prefix: /usr/local
   -- MPI: OFF (ICB provided )
   -- ICB: ON
   -- INTERFACE64: 0
   -- FC:      /opt/intel/oneapi/compiler/2023.1.0/linux/bin/intel64/ifort
   -- FCFLAGS: -O3 
   -- CC:      /usr/bin/cc
   -- CCFLAGS: -O3 -DNDEBUG 
   -- CXX:      /usr/bin/c++
   -- CXXFLAGS: -O3 -DNDEBUG 
   -- BLAS:
      -- link:    /opt/intel/oneapi/mkl/2023.1.0/lib/intel64/libmkl_intel_lp64.so
      -- link:    /opt/intel/oneapi/mkl/2023.1.0/lib/intel64/libmkl_intel_thread.so
      -- link:    /opt/intel/oneapi/mkl/2023.1.0/lib/intel64/libmkl_core.so
      -- link:    /opt/intel/oneapi/compiler/2023.1.0/linux/compiler/lib/intel64_lin/libiomp5.so
      -- link:    -lm
      -- link:    -ldl
   -- LAPACK:
      -- link:    /opt/intel/oneapi/mkl/2023.1.0/lib/intel64/libmkl_intel_lp64.so
      -- link:    /opt/intel/oneapi/mkl/2023.1.0/lib/intel64/libmkl_intel_thread.so
      -- link:    /opt/intel/oneapi/mkl/2023.1.0/lib/intel64/libmkl_core.so
      -- link:    /opt/intel/oneapi/compiler/2023.1.0/linux/compiler/lib/intel64_lin/libiomp5.so
      -- link:    -lm
      -- link:    -ldl
      -- link:    -lm
      -- link:    -ldl

This is the code I'm running:

program dsaupd_mkl
    implicit none
    integer, parameter :: n=3, ldv=n, ncv = n
    integer :: ido, nev, iparam(11), ipntr(11), lworkl, info
    real(8) :: tol, resid(n), v(ldv,ncv), workd(3*n), workl(ncv*(ncv+8))
    character ::bmat(1)

    integer :: ishfts, maxitr, mode1

    external :: dsaupd

    ido = 0
    bmat = 'I'
    nev = 1
    tol = 0.d0
    
    ishfts = 1
    maxitr = 1000
    mode1 = 1
    !
    iparam(1) = ishfts
    !
    iparam(3) = maxitr
    !
    iparam(7) = mode1

    lworkl = ncv*(ncv+8)
    info = 0
    print*, 'iparam= ', iparam
    print*, ''

    call dsaupd( ido, bmat, n, 'SA', nev, tol, resid, ncv, v, ldv, iparam, ipntr, workd, workl,&
            lworkl, info )
    print*, ido, info
end program dsaupd_mkl

And last the compiling line

ifort -I"${MKLROOT}/include" -w _results/$(FILENAME).o -o _results/$(FILENAME) -L${MKLROOT}/lib/intel64 -larpack -lmkl_intel_lp64 -lmkl_intel_thread -lmkl_core -liomp5 -lpthread -lm -ldl

By inserting some prints in dsaupd.f, I figure out that actually info was equal to -10, while the value returned to the main program was 2**32-10, as if the info variable in the main is an unsigned int, even though it was not declared as such. Furthermore the info=-10 was related to the fact that the array indexes have been shifted:

  • main/iparam(1) -> dsaupd/iparam(1)
  • main/iparam(2) -> dsaupd/iparam(3)
  • main/iparam(3) -> dsaupd/iparam(5) And so on.

Such problem happens only with the integer arrays (iparam and ipntr), while it does not happen for the arrays of reals (I tried with resid).

I don't know if there is something wrong in the building options of Arpack-ng or in the compiling options, or even worse, it is a package problem.

VINCENZO BISOGNO
  • 1
  • 3

1 Answers1

0

What was happening is that the main program was using 8 bytes integers, while the Arpack Routine was using 4 bytes. I didn't realize that was the mistake cause I already tried to build Arpack with ILP64 interface, but I did not realize that cmake was always linking libmkl_intel_lp64 instead of libmkl_intel_ilp64. What was missing was the following line:

cmake -D BLA_VENDOR=Intel10_64ilp
Aqib Javed
  • 935
  • 1
  • 5
  • 15
VINCENZO BISOGNO
  • 1
  • 3