Matrix Division on the 28C/S (and later)

[Thomas Okken]

(03-02-2020 05:48 PM)Werner Wrote:  No, what I meant was that since LU-decomp uses DOT products as well, some accuracy gains can be expected there as well, even without an iterative refinement step. Test cases for DOT products shouldn’t be too hard, for Sys solve (and MATINV) a bit harder.

The LU decomposition in Free42 is not implemented using the DOT function, though, so having it benefit from a more accurate dot product algorithm would be a separate exercise. The same thing is true of all the other functions that perform dot products under the hood, like matrix-matrix multiplication.