Stack/Pool Allocator for GPU, only allocate on device

[MichaelSt98]

Is your feature request related to a problem? Please describe.

No response

Describe the solution you'd like

Device-only allocation for the pool allocator/stack.

Describe alternatives you've considered

No response

Additional context

Currently the stack is created and allocated on device like that:

REAL(KIND=JPRB), ALLOCATABLE :: ZSTACK(:, :)
ISTSZ = ...
ALLOCATE (ZSTACK(ISTSZ, NGPBLKS))
!$acc data create( ZSTACK )

...

!$acc end data
DEALLOCATE (ZSTACK)

thus the stack is both allocated on host and device, although we only require it to be on the device.

To only allocate it on device we could:

• use declare device_resident

REAL(KIND=JPRB), ALLOCATABLE :: ZSTACK(:, :)
ISTSZ = ...
!$acc declare device_resident(ZSTACK)
allocate(ZSTACK(ISTZ, NGPBLKS))

...

deallocate(ZSTACK)

however, it is not clear whether NVIDIA adheres to the standard: StackOverflow: (NVHPC aka PGI) treats device_resident as a create?!

• use CUDA Fortran REAL(KIND=JPRB), ALLOCATABLE, DEVICE :: ZSTACK(:, :)
  • and sacrifice portability?!
• use OpenACC runtime function acc_malloc()
  • which complicates things as ZSTACK would need to be a c_ptr ....

Organisation

No response