sgels function return different answers for mkl 18.5 and 19.5

[gn164]

The following code that calls sgels returns different answers depending on whether I link with mkl 18.5 or 19.5. Also mkl 18.5 seems return different results depending on different machines. 

My link command :

/opt/intel/19/bin/ifort reproduce.f90 -Wl,--start-group ${MKLROOT}/lib/intel64/libmkl_intel_lp64.a ${MKLROOT}/lib/intel64/libmkl_intel_thread.a ${MKLROOT}/lib/intel64/libmkl_core.a -Wl,--end-group -liomp5 -lpthread -lm -ldl -I${MKLROOT}/include

      program reproduce
        implicit none

        integer :: row_A, col_A, row_B, col_B, info
        integer :: i
        real, allocatable :: A(:,:), B(:,:), work(:,:)

        ! allocate input matrices
        row_A = 2
        col_A = 2
        row_B = 2
        col_B = 1

        allocate(A(row_A,col_A))
        allocate(B(row_B,col_B))
        allocate(work(row_A*row_B,col_A*col_B))

        A = 1
        B = 1490.890

        ! sgels
        call sgels('N', size(A,1), size(A,2), size(B,2), A, size(A,1), B,    &
                              size(B,1), work, size(work), info)

        ! check
        write(*,*) '--> A matrix'
        do i = 1, ubound(A, 1)
           write(*,*) A(i, :)
        end do
        write(*,*) '--> B matrix'
        do i = 1, ubound(B, 1)
           write(*,*) B(i, :)
        end do
        write(*,*) '--> Info'
        write(*,*) info

        deallocate(A, B, work)

      end program

Output for mkl 19 on skylake (Intel(R) Xeon(R) W-2145 CPU @ 3.70GHz)

--> A matrix
-1.414214 -1.414214
0.4142135 0.0000000E+00
--> B matrix
-2108.437
8.2589657E-05
--> Info

2

Output for mkl 18 on skylake (Intel(R) Xeon(R) W-2145 CPU @ 3.70GHz)

--> A matrix
-1.414214 -1.414214
0.4142135 1.9078982E-08
--> B matrix
-2837.939
4328.829
--> Info
0

Output for mkl 18 on sandybridge  ( Intel(R) Xeon(R) CPU E5-2640 0 @ 2.50GHz)

--> A matrix
-1.414214 -1.414214
0.4142135 0.0000000E+00
--> B matrix
-2108.437
1.2207031E-04
--> Info
2

[mecej4]

The MKL documentation for ?gels says, "It is assumed that A has full rank". Your matrix is of size 2 X 2 but its rank is only 1, so the input to the routine is improper. In such cases, what is the point of comparing the output, which could well be meaningless?