Fortran version of magma_zgesv_gpu fails on GPU1 works on GPU0

[Timothy Kaiser]

Running the Fortran and C++ examples magma-2.7 2.0/example/[example_f.F90,example_v1.cpp]
I note that the Fortran example fails if I try to run GPU 1 instead of GPU 0.  The C++
example works fine on either GPU.  These examples call  the Fortran and C++ versions of
magma_zgesv_gpu.

These were run on a machine that has two V100 cards, Intel compilers, gcc 10.x, Cuda 7.x.

I was able to see that the Fortran version started a process on each of the GPUs but the
C++ version only started a process on either 0 or 1.  It looks like something in this
routine has GPU 0 hard coded.  

I modified the two codes to read the GPU id and then call the Fortran and C++ versions of
magma_setdevice(id).  Can someone try to replicate this?

Here are the mods and results:


**** example_f.F90 ****
    integer id
    call magmaf_init()

    read(*,*)id
    write(*,*)"setting gpu to ",id
    call magmaf_setdevice(id)

[tkaiser2@r103u21 example]$ ./example_f
0
 setting gpu to            0
using MAGMA CPU interface
using MAGMA GPU interface
[tkaiser2@r103u21 example]$ ./example_f
1
 setting gpu to            1
using MAGMA CPU interface
using MAGMA GPU interface
magma_zgesv_gpu failed with info= -112
 Error: magmaf_free( dA ) failed:         -115
 Error: magmaf_free( dX ) failed:         -115
[tkaiser2@r103u21 example]$


**** example_v1.cpp ****

    magma_device_t id;
    id=1;
    scanf("%d",&id);
    printf("setting gpu to %d\n",id);
    magma_setdevice(id);

[tkaiser2@r103u21 example]$ ./example_v1
1
setting gpu to 1
using MAGMA CPU interface
using MAGMA GPU interface
[tkaiser2@r103u21 example]$ ./example_f
0
 setting gpu to            0
using MAGMA CPU interface
using MAGMA GPU interface
[tkaiser2@r103u21 example]$