CUBLAS Fortran issues

Accelerated Computing CUDA CUDA Programming and Performance
1

I’m trying to use the CUBLAS Fortran wrapper (fortran.c). The code runs OK on my desktop (Intel Core2 Duo, GTX285, Ubuntu10.04 64bit). But the same code will return “CUBLAS_STATUS_MAPPING_ERROR” at the cublas_set_vector, if run on one node of a GPU cluster (6-core Intel Xeon, Tesla M2050, Linux x86-64).

I also test the C version of the code and it runs OK on both machines.

And it seems that there are some similar issues posted before but didn’t get solved:

Here is the Fortran code:

program main

	implicit none

	integer, parameter :: N=8

	integer, parameter :: sizeof_real=4

	real(kind=4), dimension(:),pointer :: h_p

	integer :: d_p

	integer stat

	integer cublas_init, cublas_alloc, cublas_set_vector, cublas_free, cublas_shutdown

	allocate(h_p(N))

	h_p = 0.0d0

	stat = cublas_init()

	if (stat .NE. 0) then

		write(*,*) "cublas_int failed"

		write(*,*) "stat=",stat

		stop

	endif

	stat = cublas_alloc(N, sizeof_real, d_p)

	if (stat .NE. 0) then

		write(*,*) " cublas_alloc failed"

		write(*,*) "stat=",stat

		stop

	endif

	write(*,*) "d_p =", d_p

	stat = cublas_set_vector (N, sizeof_real, h_p, 1, d_p, 1)

	if (stat .NE. 0) then

		write(*,*) "cublas_set_vector failed"

		write(*,*) "stat=",stat

		stop

	endif

	deallocate(h_p)

	stat = cublas_free (d_p)

	stat = cublas_shutdown()

	write(*,*) "test pass..."

	stop

end

Here is C version code:

#include <stdio.h>

#include <stdint.h>

#include <cublas.h>

int main(int argc, char *argv[])

{

	int N=8;

        float* h_p;

	float* d_p;

	int devPtr;

        cublasStatus stat;

h_p = (float *)malloc (N * sizeof(float));

        for (int i = 0; i < N; i++)  h_p[i] = 0;

stat = cublasInit();

        if (stat != 0) {

                printf ("cublasInit failed\n");

		printf ("stat=%d\n",stat);

                return 1;

        }

stat = cublasAlloc (N, sizeof(float), (void**)&d_p);

        if (stat != 0) {

                printf ("cublasAlloc failed");

		printf ("stat=%d\n",stat);

                return 1;

        }

	devPtr = (int)(uintptr_t)d_p;

	printf("devPtr=%d\n",devPtr);

stat = cublasSetVector (N, sizeof(float), h_p, 1, d_p, 1);

        if (stat != 0) {

                printf ("cublasSetVector failed");

		printf ("stat=%d\n",stat);

                return 1;

        }

cublasFree (d_p);

        cublasShutdown();

	printf("test pass...\n");

        return 0;

}

The output of the Fortran code on GTX 285:

d_p =     1048576

 test pass...

The output of the C code on GTX 285:

devPtr=1048576

test pass...

The output of the Fortran code on M2050:

d_p =           0

 cublas_set_vector failed

 stat=          11

The output of the C code on M2050:

devPtr=0

test pass...

For all the cases, the compilers and options are:

gcc -O3

gfortran -cpp -O3 -ffree-form -ffree-line-length-0 -fno-second-underscore

nvcc -O3 -arch sm_13 -use_fast_math

Any clue?

Thanks,

Kong