Hi Peter,
Here are the long-awaited precision change modifications I discussed before. In summary, aside from my new implementation I have a unified precisionChange interface that will redirect to precisionChangeFast for double2 etc. precisionChange accepts a workspace for reuse that will speed up the call if used. I've implemented the workspace changes in the mixed-CG, reliable update CG and mixed-prec multishifts and have tested them on both Crusher and my local computer. I've also added a test and benchmark for precisionChange, showing significant reduction in wallclock time for precision change on both.
Best,
Chris
paboyle
commented
1 year ago
Hi Peter, Here are the long-awaited precision change modifications I discussed before. In summary, aside from my new implementation I have a unified precisionChange interface that will redirect to precisionChangeFast for double2 etc. precisionChange accepts a workspace for reuse that will speed up the call if used. I've implemented the workspace changes in the mixed-CG, reliable update CG and mixed-prec multishifts and have tested them on both Crusher and my local computer. I've also added a test and benchmark for precisionChange, showing significant reduction in wallclock time for precision change on both. Best, Chris