The integration layer
mp-numerics — Mixed-Precision Numerical Methods
Where does each number system stop being usable? mp-numerics runs classic numerical methods in IEEE, posit, takum, cfloat, lns, fixed-point and multi-component arithmetic, and measures the answer exactly.
Measured exactly
Computed values are read out of the number type as exact expansions, so a reported error is the error of the arithmetic, not of the measurement.
Reproducible
Every committed result carries a .sysinfo sidecar with the platform,
compiler and the commits of mp-numerics, MTL5 and Universal it was run
against.
Organized by method
Quadrature, root finding, interpolation, optimization, approximation and linear algebra, each with applications, benchmarks and tests.
Studies
Section titled “Studies” Givens rotations and Householder reflectors One rotation or one reflector at a time, across 65 types: where each form loses orthogonality, and why (#13).
QR and least squares The factorizations built from them: where each type stops being usable, and whether the method or the type decides (#14).