Floating point precision
IncompressibleNavierStokes generates efficient code for different floating point precisions, such as
- Double precision (
Float64) - Single precision (
Float32) - Half precision (
Float16)
To use single or half precision, all user input floats should be converted to the desired type. Mixing different precisions causes unnecessary conversions and may break the code. See the tutorial for an example on how to enforce floating point type hygiene.
For GPUs, single precision is preferred. CUDA.jls cu converts to single precision.
SparseArrays.jls sparse matrix factorizations only support double precision. psolver_direct only works for Float64. Consider using an iterative solver such as psolver_cg when using single or half precision.