mp-numerics roadmap
Milestone 0: composition layer bootstrapped (done)
Section titled “Milestone 0: composition layer bootstrapped (done)”- CMake scaffold replicated from the mp-* template (mp-iterative / mp-ir):
INTERFACE library
sw::mp_numerics, find_package -> FetchContent fallback for MTL5 + Universal, config-package install, CI matrix (MSVC/GCC/Clang/AppleClang). - Shared header
include/sw/mp_numerics/integration.hpp: composite Simpson and trapezoidal quadrature, generic over the arithmetic type. - Smoke test: composite Simpson integrates known functions (x^2, 1/(1+x^2)) in
double,float,posit<32,2>,cfloat<32,8>to each type’s accuracy floor. - Demo application
simpson_precision(quadrature error vs subinterval count across number types) and benchmarksimpson_benchmark(accuracy per unit time). - Repo organized by numerical-method category:
integration/,roots/,interpolation/,optimization/,approximation/. - Exact quire accumulation: pass
quire<T>as the MTL5 accumulator directly. Universal defines the quire’saccumulator_*functions and MTL5’saccumulator_traitsfinds them by argument-dependent lookup, so neither library includes the other. This replaced thequire_accumulator.hppbridge header formerly copied into this repo (stillwater-sc/mtl5#582).
Milestone 1: migrate the Universal numerics experiments
Section titled “Milestone 1: migrate the Universal numerics experiments”Migrate from Universal mixedprecision/{roots,integration,interpolation, optimization,approximation}, re-expressed on Universal precisions (and MTL5
kernels where the method uses linear algebra, e.g. Chebyshev approximation):
-
roots— scalar root finding (bisection/Newton/secant,include/sw/mp_numerics/roots.hpp) across precisions:root_precisionapp +test_root_smoke. The Jenkins-Traub polynomial (all-roots) finderrpolyis ported ininclude/sw/mp_numerics/polyroots.hpp:rpoly_precisionapp +test_rpoly(universal#1211). The engine is the canonical ak1 / Netlib TOMS-493 reference implementation templatized overReal(include/sw/mp_numerics/detail/rpoly_ak1.hpp, mp-numerics#7), with a precision-aware coefficient-scaling gate so tapered types (posit) reach their precision floor (posit<64,2> is exact on well-separated integer roots; double 3.6e-15 vs the earlier port’s 1.6e-2 on the same case). integration— extend the quadrature study (adaptive rules, Gauss).-
interpolation— Newton divided-difference interpolation (include/sw/mp_numerics/interpolation.hpp):interp_precisionapp +test_interp_smoke. High-degree interpolation amplifies low-precision roundoff (posit<16,2>/cfloat<16,5> degrade as node count grows). -
optimization— golden-section 1-D minimization (include/sw/mp_numerics/optimization.hpp):optim_precisionapp +test_optim_smoke. Minimizer accuracy floors at ~sqrt(eps) of the type. -
approximation— truncated Chebyshev-series approximation (include/sw/mp_numerics/approximation.hpp, coefficients + Clenshaw eval):cheb_precisionapp +test_cheb_smoke. double converges geometrically; narrow types stall at their coefficient-accumulation floor. (A least-squares variant using MTL5 QR is a possible follow-up.)
Milestone 2: mixed-precision numerics studies
Section titled “Milestone 2: mixed-precision numerics studies”- Where can low precision live inside an iterative numerical method without stalling? (mirrors the mp-ir / mp-iterative residual-precision question.)
- Quire (exact accumulation) inside quadrature / polynomial evaluation via the
accumulator_traitsbridge, vs naive same-precision accumulation.
Milestone 3: control and stability numerics
Section titled “Milestone 3: control and stability numerics”Kalman filtering and control-algorithm failure modes under mixed precision.
See docs/control.md for the full taxonomy.
-
applications/control/kalman_scalar— scalar Kalman filter (1-D position/velocity) acrossdouble,float,posit<32,2>, andposit<32,2>with quire accumulation. Measures filter divergence, covariance SPD preservation, and innovation consistency. -
benchmarks/control/kalman_precision— precision sweep: error vs number type over N timesteps. -
benchmarks/control/spd_preservation— does P remain positive definite? Failure mode: negative eigenvalue from rounding. -
benchmarks/control/consistency— normalized innovation chi-squared test across precisions. - Integrate LDL* factorization for gain solve (avoids Cholesky square-root rounding per John Gustafson’s note).