Mixed Precision
This issue provides an overview for the progress towards mixed precision
-
Make real_t=float
usable.
Nils and Dominik decided to exclude the "old", stencil generated kernels for this task since these are hard-coded to double.
-
Create an app/test that uses functions with double
andfloat
and has a assign/copy function to couple them -
Optimize the hand-written vector-vector operations (such as assign
,add
, etc.) and get rid of their generated counterparts since we do not really need to have those generated, even for performance. -
Couple the hfg with pystencils to generate a "full" operator class. This should firstly contain the apply
function
See also #164 (closed)
Edited by Dominik Thoennes