Skip to content

Feature single precision

Mattijs Janssens requested to merge feature-single-precision into develop

Summary

Replace all linear solver (lduSolver) variables with new type 'solveScalar' instead of 'scalar'. In double precision this is the same as scalar (i.e. double), in single precision it is either double or scalar (=float).

Resolved bugs (If applicable)

#1086

Details of new models (If applicable)

Currently: WM_PRECISION_OPTION:

  • DP as usual
  • SP as usual
  • SPDP single precision except for linear solver, which is double precision

When using WM_SPDP argList now prints in Arch: label=32;scalar=32;solveScalar=64

Risks

Following tests before merging:

  • DP = same performance
  • SP = pure single precision same performance as current
  • SPDP = mixed precision performance may need case-by-case evaluation
Edited by Mark OLESEN

Merge request reports