fortran cudamalloc and cublassetmatrix wrapper problem - cuda

Question

I maybe a problem, either with "cudamalloc" or "cublassetmatrix" not appropriately called from fortran:
module cuda
interface
integer (C_INT) function cudaMallocHost(buffer, size) bind(C,name="cudaMallocHost")
use iso_c_binding
implicit none
type (C_PTR) :: buffer
integer (C_SIZE_T), value :: size
end function cudaMallocHost
integer (C_INT) function cudaFreeHost(buffer) bind(C,name="cudaFreeHost")
use iso_c_binding
implicit none
type (C_PTR), value :: buffer
end function cudaFreeHost
integer (C_INT) function cudaMalloc(buffer, size) bind(C,name="cudaMalloc")
use iso_c_binding
implicit none
type (C_PTR) :: buffer
integer (C_SIZE_T), value :: size
end function cudaMalloc
integer (C_INT) function cudaFree(buffer) bind(C,name="cudaFree")
use iso_c_binding
implicit none
type (C_PTR), value :: buffer
end function cudaFree
integer (C_INT) function cublassetmatrix(M,N,size,A_host,lda_h&
&,A_dev,lda_d) bind(C,name="cublasSetMatrix")
use iso_c_binding
implicit none
type (C_PTR) :: A_dev
type(c_ptr) :: A_host
integer (C_SIZE_T), value :: size
integer(C_Int) :: M,N,lda_h,lda_d
end function cublassetmatrix
Type(c_ptr) function cudaGetErrorString(err) bind(C,name="cudaGetErrorString")
use iso_c_binding
implicit none
integer (C_SIZE_T), value :: err
end function cudaGetErrorString
integer (C_INT) function cublasCreate(handle) bind(C,name="cublasCreate_v2")
use iso_c_binding
implicit none
Type(C_Ptr) :: handle
end function cublasCreate
end interface
end module cuda
program test
use iso_c_binding
use cuda
implicit none
integer, parameter :: fp_kind = kind(0.0)
type(C_PTR) :: cptr_A, cptr_A_D
real (fp_kind), dimension(:,:), pointer :: A=>null()
real :: time_start,time_end
integer:: i,j, res, m1
integer(c_int) :: x
type(c_ptr) :: handle
logical:: lsexit
CHARACTER(len=50), POINTER :: errchar
m1=500
res=cublasCreate(handle)
if(res/=0) Then
write(*,*) "ERROR 1 ",res;
end if
res = cudaMallocHost ( cptr_A, m1*m1*sizeof(fp_kind) )
if(res/=0) Then
write(*,*) "ERROR 2 ",res;
end if
call c_f_pointer ( cptr_A, A, (/ m1, m1 /) )
A=1._fp_kind
res = cudaMalloc ( cptr_A_D, m1*m1*sizeof(fp_kind) )
if(res/=0) Then
write(*,*) "ERROR 3 ",res;
end if
res=cublasSetMatrix (m1,m1,sizeof(fp_kind),cptr_A,m1,cptr_A_D,m1)
if(res/=0) Then
write(*,*) "ERROR 4 ",res,sizeof(fp_kind)
call c_f_pointer ( cudageterrorstring(int(res,kind=8)),&
& errchar, [ len(errchar) ] )
write(*,*) trim(adjustl(errchar))
end if
end program test
The make command is:
tmp:
ifort -O3 -o tmp $(FFLAGS) tmp.f90 -L/opt/cuda/lib64 -lcublas -lcudart
clean:
rm tmp cuda.mod
Although "cudamallochost" expects "void ** ptr" it seems to work as the fortran pointer is usable, and when put in a loop with "cudafree" does not produce memory leaks.
The code fails at "cublassetmatrix" function either with error code 7 ("too many resources") or 11 ("invalid argument").
Any idea?