mpirun on CUDA GPU

Dear experts,

I am using AMBER12 software on 4 units of TESLA-2075 GPU cards in CENTOS5.7 with intel processor. I have installed MPICH2 package to run my simulation in parallel.
Recently I could install all the software properly and it is running fine.

Now my intention is to use all the GPUs for bench-marking. Currently I am running a job in one GPU at a time. I would like to run one job in two or three GPUs at a time.

How this can be achieved? Is there any script specially for this? I need some explanation if can.

I appreciate any help in advance.