The workspaces are now used by the plaq+rectangle gauge action resulting in a further 2x performance improvement as measured on a 16^4 local volume for 2 nodes (16 ranks) of Crusher
Fixed const correctness on PaddedCell methods
Fixed compile issues on Crusher
Added timing breakdowns for PaddedCell::Expand and the padded implementations of the staples, visible under --log Performance
Optimized kernel for StaplePadded
Test_iwasaki_action_newstaple now repeats the calculation 10 times and reports average timings
Moved versions of the padded-cell implementations of staple and rect-staple from test code to WilsonLoops header
Added StapleAndRectStapleAll which is now called by the plaq+rectangle action class. Under the hood it uses the padded cell implementations with maximal reuse of the padded gauge links
Added test code tests/debug/Test_optimized_staple_gaugebc demonstrating equivalence of above to RectStapleUnoptimised for cconj gauge BCs
Removed optimized staple only being used for periodic gauge BCs; it is now always used
Fixed bug in padded staple code where extract was being called on the result before the GPU view was closed
Fixed compile issue with pointer cast in padded staple code
Added timing summaries of padded staple code and timing breakdown of staple implementation to Test_padded_cell_staple
Fixed const-correctness for PaddedCell and ConjugateGimpl::setDirections
Modified test code for padded-cell implementation of staple, rect-staple to use cconj BCs
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
Added a conjugate momentum "filter" that applies a phase to each site. With sites set to 1.0 or 0.0 this acts as a mask and enables, for example, the freezing of inactive gauge links in DDHMC
Added tests/forces/Test_momentum_filter demonstrating the use of the filter to freeze boundary links
Christopher Kelly