Mixed prec CG algorithm has been modified to precompute precision change workspaces
As the original Test_dwf_mixedcg_prec has been coopted to do a performance stability and reproducibility test, requiring the single-prec CG to be run 200 times, I have created a new version of Test_dwf_mixedcg_prec in the solver subdirectory that just does the mixed vs double CG test
Added a new, much faster implementation of precision change that uses (optionally) a precomputed workspace containing pointer offsets that is device resident, such that all lattice copying occurs only on the device and no host<->device transfer is required, other than the pointer table. It also avoids the need to unpack and repack the fields using explicit lane copying. When this new precisionChange is called without a workspace, one will be computed on-the-fly; however it is still considerably faster than the original implementation.
In the special case of using double2 and when the Grids are the same, calls to the new precisionChange will automatically use precisionChangeFast, such that there is a single API call for all precision changes.
Reliable update and mixed-prec multishift have been modified to precompute precision change workspaces
Renamed the original precisionChange as precisionChangeOrig
Fixed incorrect pointer offset bug in copyLane
Added a test and a benchmark for precisionChange
Added a test for reliable update CG
Cshift is allocated in Unified memory and used
in the LambdaApply kernels but also populated
from the host. This creates a lot of Unified HtoD
and DtoH mem operations and has a negative effect
in performance. With this commit we populate the
Cshift table in the device with the
populate_Cshift_table() kernel.
