The purpose of this PR is to enable fields with a last dimension (usually the block dimension) different from 1; in particular, we test gather/scatter and reshuffle functionnalities on these fields.
Tested in ARPEGE t798, double precision : we run this test case on 12 NVIDIA V100. Using temporaries with block range allowed us to decrease the amount of required memory by a factor of 2. Each MPI task has 72058 = 2252 x 32 points, we process the 2252 blocks in 4 batches of 640 blocks.
The purpose of this PR is to enable fields with a last dimension (usually the block dimension) different from 1; in particular, we test gather/scatter and reshuffle functionnalities on these fields.
Tested in ARPEGE t798, double precision : we run this test case on 12 NVIDIA V100. Using temporaries with block range allowed us to decrease the amount of required memory by a factor of 2. Each MPI task has 72058 = 2252 x 32 points, we process the 2252 blocks in 4 batches of 640 blocks.