Rank 0 choosing device 0 out of 1
Standard lattice layout:
 4 dimensions
 Node remapping: TRIVIAL (no effort made to reorder)

 Sites on node: 32 x 32 x 32 x 32
 Processor layout: 1 x 1 x 1 x 1
Matrix * Matrix: 6.52344ms 
Vector * Matrix: 2.69531 ms 
Vector square sum: 0.771484 ms 
Dirac 4 dirs: 29.375ms 
Dirac: 32.8125ms 
CG: 36.25ms / iteration
 COMMS from node 0: 260 done, 1264(82.9396%) optimized away
