Closed drwells closed 1 year ago
One problem is that deal.II (via its householder transformation code) produces slightly different results in optimized mode, which in and of itself is enough to throw off Triangle. I added some extra regularization to fix this - the combination of regularization and picking compilation flags seems to be enough.
Just clearing compilation flags on a target isn't enough - CMAKE_CXX_FLAGS will still be used to populate it. This change decouples the input C++ flags from those used to compile triangle - that will make triangle's output more reliable since it relies a lot on very precise floating point tricks.
That being said - this isn't done yet since the
surface_tria_01_2d.input
test still produces different output with-march=native
.