Permit INV_TRANS with OpenMP offload

[samhatfield]

Now that we can compile with OpenMP offload, I will gradually add back this capability to ecTrans until we can run the benchmark program.

The purpose of this PR is to allow, at least in a minimal fashion, OpenMP offload to be used with the inverse transform. This PR can be merged once this program produces the correct behaviour:

PROGRAM TEST_PROGRAM

USE PARKIND1, ONLY: JPIM, JPRB

IMPLICIT NONE

! Spectral truncation
INTEGER(JPIM) :: TRUNC = 79

! Arrays for storing our field in spectral space and grid point space
REAL(KIND=JPRB), ALLOCATABLE :: SPECTRAL_FIELD(:,:)
REAL(KIND=JPRB), ALLOCATABLE :: GRID_POINT_FIELD(:,:,:)

! Dimensions of our arrays in spectral space and grid point space
INTEGER(KIND=JPIM) :: NSPEC2
INTEGER(KIND=JPIM) :: NGPTOT

#include "setup_trans0.h"
#include "setup_trans.h"
#include "trans_inq.h"
#include "inv_trans.h"

!CALL DR_HOOK_INIT()

! Initialise ecTrans (resolution-agnostic aspects)
CALL SETUP_TRANS0(LDMPOFF=.TRUE., KPRINTLEV=2)

! Initialise ecTrans (resolution-specific aspects)
CALL SETUP_TRANS(KSMAX=TRUNC, KDGL=2 * (TRUNC + 1))

! Inquire about the dimensions in spectral space and grid point space
CALL TRANS_INQ(KSPEC2=NSPEC2, KGPTOT=NGPTOT)

! Allocate our work arrays
ALLOCATE(SPECTRAL_FIELD(1,NSPEC2))
ALLOCATE(GRID_POINT_FIELD(NGPTOT,1,1))

! Initialise our spectral field array
SPECTRAL_FIELD(:,:) = 0.0_JPRB

! Perform an inverse transform
CALL INV_TRANS(PSPSCALAR=SPECTRAL_FIELD, PGP=GRID_POINT_FIELD)

END PROGRAM TEST_PROGRAM

This just performs an inverse transform on a single 2D scalar field of zeroes.

[samhatfield]

Current state of play: the test program crashes in LEINV on the first instance of

!$OMP TARGET DATA USE_DEVICE_PTR(ZAA,ZINP,ZOUTA)

with error

Memory access fault by GPU node-4 (Agent handle: 0xf82500) on address 0xe2b000. Reason: Unknown.

The similar statement above for the 0-mode double-precision arrays (ZAA0 etc.) does not produce this error.

[samhatfield]

After further investigation, it seems that ZAA etc. are correctly transferred to device, which is good news. The issues is that the first call to a routine in HICBLAS_MOD (HIP_DGEMM_BATCHED_OVERLOAD) somehow leaves the device in a corrupted state. Accesses to ZAA before this call are successful (the sum of the array computed on device matches that computed on host) but the exact same computation after the call triggers the above Memory access fault error.

[samhatfield]

#210 should be merged before this.