PGI User Forum

[SarahA]

Hello,

I have a fortran 90 code to try with CUDA Fortran (2.3). It has a lot of array aliasing, where they thought it would be nice to refer to double complex 5d arrays sometimes with double reduced-dimension arrays. This was done with argument aliasing, for example:

real array(18, N, 2, 4)
call subr( array )
...
subroutine subr( array )
complex array(3,3,N,2,4)

This doesn't work with fortran modules, and that seems to be the best way to write sets of CUDA device arrays and routines that use the arrays. There is much whining about formal parameter mismatches, etc.

It doesn't help that EQUIVALENCE is documented as unsupported for CUDA 2.3 ... it does compile without error, but it doesn't seem to work. Of course, I could have another error.

So, my long winded questions are

1) Is there a way to relax type conformance on array copies? ( host = device, etc ).

2) I have tried using the cudaMemcpy() routine, but it fails with errors like:
copyin Memcpy FAILED:4 Perhaps it is because these arrays are multi-dimensional.

3) The runtime routines are not real well documented. Is there a somewhere more details are provided on things like "cudaMemcpyToArray"? For instance, what are the dstx, dsty parameters?

4) Is device array equivalence really supposed to work? Section 3.2.1 of the v1.2 CUDA Fortran guide says no.


Thanks,
Sarah

[mkcolg]

Hi Sarah,

[SarahA]

Thanks for the idea.

Any pointers on documentation for the CUDA Fortran run time memory copy routines? The manual, even the newest one, is a little thin there. I'll try matching them up with the CUDA C equivalents to decode the "dstx" and "dsty" arguments and so on.

Perhaps a runtime call to do copies would solve my problem.

[mkcolg]

Hi SarahA,

[SarahA]