Comments

• MikeMcMullen

  January 5

  To be expected. The first half dozen to dozen attempts are in single precision which is usually an order of magnitude faster on most machines, but about 7 digit accuracy, but in long sequences of calculations rounding and truncating errors will reduce this to 3 or 4 digits some times less for very modest size matrices and can cause erronous results and failure to converge due to this computational noise. The last pass or two is done in double precision, to correct these errors. Not a big deal, as if mixed precision causes a problem, run it again without mixed precision and come backlater. Just need to be aware that some problems are sensitive.
• Victor

  January 10

  I notice PASTIX_MIXED_PRECISION = 1 (solver message shows "Arithmetic: Float") seems to be the default and I have to turn it off by PASTIX_MIXED_PRECISION = 0. Not sure if that's the case for everybody else but it seems like a bad default. I agree with @disla that mixed precision is dangerous. See this solution of two identical cantilever beams with the same tip load. No error message, just bad solution.