mpi and cublas problem

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi all,

I am trying to use cublas within an mpi application.

the mpi code is as follow :

[codebox]

…

if(rank==0){

rc = run_gpu(0);

if (rc != 0){

  printf ("Error in run_gpu. (%i)\n",rc);

  MPI_Abort(MPI_COMM_WORLD, rc);

}

}

…

[/codebox]

and the function :

[codebox]

int run_gpu(int gpu){

…

cublasStatus toto;

if( cudaSetDevice(gpu)!=cudaSuccess) return 1;

toto=cublasInit();

if(toto!=CUBLAS_STATUS_SUCCESS ){

printf("TOTO = %u\n",toto);

return 2;

}

…

}

[/codebox]

this code exit because the function cublasInit() returned CUBLAS_STATUS_NOT_INITIALIZED. The exact same code runing outside mpi execute well…

Does anyone knows why it is not working under MPI? are there special things to do in this case?

Thanks,

Jonathan

2

Issue solved,

the gpu I was using with mpi was in protected mode, which prevent me from running…

Sorry for the noise,

Jonathan

3

hi
how can you change the “protected mode”?
I just face the same problem ,and do not know your solution~~~
kai