device to device assignment

Accelerated Computing HPC Compilers Legacy PGI Compilers
1

It says on p. 20 of the user guide:

An assignment statement with a device variable or device array or array section on both sides of the assignment statement will copy data between two device variables or arrays.

When I try this, however, I get the error:

more than one device-resident object in assignment

Example code:

      program dev2dev

      use cudafor

      integer, device :: d1
      integer, device :: d2
      integer         :: h

      h = 1

      d1 = h

      d2 = d1

      h = d2

      write (*,*) h

      end program dev2dev

[/quote]

2

Hi BeachHut,

Unfortunately, implementation is slightly behind the CUDA Fortran spec and this is one of the few features left. I’ll update this post once this is supported.

Thanks,
Mat

3

Hi BeachHut,

FYI, device to device copies will be supported in the 10.5 release.

  • Mat
4

This seems only partially fixed. If I use allocatables I still get an error.

dev2dev.cuf:

program dev2dev

      use cudafor

      integer, allocatable, device, dimension(:) :: d1
      integer, allocatable, device, dimension(:) :: d2

      allocate(d1(10))
      allocate(d2(10))

      d1(5) = 10

      d2(5) = d1(5)

end program dev2dev




$ pgfortran dev2dev.cuf

PGF90-S-0155-more than one device-resident object in assignment  (dev2dev.cuf: 13)
  0 inform,   0 warnings,   1 severes, 0 fatal for dev2dev
5

This should work. I have duplicated the error and have filed a bug.