Fix typo in ordering of fields in ZGMV

[samhatfield]

The calculation of the ordering of fields in zgmv is incorrect. First spotted by @marsdeno when he tried running with --uvders --scders --vordiv all enabled (I had never thought of testing all three at once!). The benchmark crashed with a "partially present on device" error referring to PGP3A.

UV-like fields and other 3D fields are both stored in the same array: zgmv. zgpuv and zgp3a are pointers pointing to different slices of this array, and these are subsequently passed to inv_trans and dir_trans:

zgpuv => zgmv(:,:,1:jend_vder_EW,:)
zgp3a => zgmv(:,:,jbegin_sc:jend_scder_EW,:)

(Here)

Without this fix (assuming nfld == 1): jend_vder_EW = 6, jbegin_sc = 6, jend_scder_EW = 8: UV-like fields are arrange in entries 1:6, other 3D fields in entries 6:8. This is obviously wrong because the first entry of zgp3a is pointing to the same address as the last entry of the zgpuv.

The typo is here. jbegin_vder_EW should be jend_vder_EW. With this change the two pointer slices do not overlap.

Lessons learned:

• We should add a test case with ALL benchmark arguments specified, but this still wouldn't have caught this bug because...
• We should also find a way to execute the tests on a GPU-equipped machine. Is this possible with AC?

[wdeconinck]

Is this also a change that needs to be in develop?

[samhatfield]

Yes good point.