slatec

Safe, family-featured Rust APIs over validated original SLATEC Fortran routines.

Use this crate when an owned, checked Rust interface exists for the required family. The safe surface is intentionally selective and is much smaller than the 812-routine raw API in slatec-sys. A native family requires exactly one backend (bundled, source-build, system, or external-backend). The canonical bundled default provides ten reviewed special/root families on x86_64-pc-windows-gnu and x86_64-unknown-linux-gnu from hash-verified carriers, without compiler, cache, system, or network access. Builds never download source. GNU MinGW 14.2 on Windows and GNU Fortran 11.4 on Linux are the validated native profiles.

The hosted least-squares-nonlinear-easy family provides residual-only, finite-difference nonlinear least-squares fitting through the original SNLS1E and DNLS1E easy drivers. The separate least-squares-nonlinear-expert feature provides reviewed SNLS1 and DNLS1 finite-difference and dense analytic-Jacobian modes. Both require std, an explicit native backend, and the validated GNU MinGW profile; the independent least-squares-covariance feature provides reviewed SCOV and DCOV covariance estimation under the same hosted-profile requirement. The distinct least-squares-linear-nonnegative feature wraps WNNLS and DWNNLS for equality-constrained linear least squares with selected nonnegative variables; it also requires std, an explicit backend, and the validated GNU MinGW profile. The separate least-squares-linear-bounded feature wraps SBOLS and DBOLS for dense linear least squares with typed closed per-variable bounds under the same hosted-profile requirement; it does not provide general equalities, inequalities, or linear programming. least-squares-linear-bounded-constrained separately wraps SBOCLS and DBOCLS: it bounds original variables and the constraint expressions C x, and therefore is neither a generic combination of the other drivers nor a linear-programming interface.

The hosted ode-sdrive-expert feature provides owned continuation sessions over reviewed SDRIV1/DDRIV1, SDRIV2/DDRIV2, CDRIV1/CDRIV2, and the existing expert SDRIV3/DDRIV3 drivers. Driv2Session and ComplexDriv2Session expose zero-based indexed root events; all callback sessions are panic-contained, process-serialized, and preserve same-direction continuation workspace. The real expert sessions additionally support functional, internal dense/banded finite-difference, and checked analytic dense/banded Jacobian iterations; mass matrices, event lifecycle, DAEs, interpolation, and CDRIV3 remain deliberately deferred.

The quadrature-piecewise-polynomial feature adds DPFQAD multiplicative integration over the checked PiecewisePolynomial<f64> representation. The nonlinear-systems feature adds scalar-equation SOS/DSOS solvers with typed termination reports. Both reuse the existing callback runtime and require std, an explicit native backend, and the validated GNU MinGW profile.

The hosted dassl feature separately provides owned real residual-only DAE sessions over SDASSL/DDASSL for index-1 G(t, y, y') = 0 problems. It requires caller-supplied sufficiently consistent (y, y'), uses internal dense or checked-banded finite differences, contains residual errors and panics, and keeps all native calls process-serialized. Analytic user Jacobians remain deferred because their native callback has no abort protocol; events and automatic initial-condition calculation remain deferred; see the DASSL guide.

The hosted optimization-linear-programming-in-memory feature provides typed owned sparse-column LPs over original SPLP/DSPLP. It rejects problems that would require paging and never exposes Fortran-unit, save/restore, or printing controls. Optimal results provide checked typed basis records, source-audited row multipliers and reduced costs, and independently recomputed KKT diagnostics; reviewed controls are limited to iteration count, feasibility tolerances, and pricing.

The hosted approximation-polynomial-fitting feature provides immutable f32/f64 weighted polynomial least-squares models over POLFIT/DPOLFT. It supports source-defined degree selection, value and derivative evaluation, and origin power coefficients without exposing native workspace. It is a fitting API, not global interpolation; see the polynomial-fitting guide.

The hosted fftpack-real feature provides reusable f32 plans over the reviewed real FFTPACK families: periodic packed real FFT, easy Fourier series, full sine/cosine, and quarter-wave sine/cosine transforms. Each plan owns its initialized work array and preserves the original in-place storage and native normalization. The selected snapshot contains no reviewed f64 real FFTPACK counterparts. See the real FFTPACK guide.

The hosted fftpack-complex feature provides a one-dimensional in-place ComplexFftPlan32 over public num_complex::Complex32. It calls the reviewed standard real-array CFFTI1/CFFTF1/CFFTB1 interface, preserves the native negative/positive exponent directions and N round-trip scale, and has no Complex64 plan because the selected snapshot contains no complex f64 roots. See the complex FFTPACK guide.

The hosted fishpack-cartesian-2d feature provides an owned f32 Cartesian Poisson/Helmholtz finite-difference solve over reviewed HWSCRT. It has typed periodic, value, derivative, and mixed edge conditions; explicit perturbation and non-uniqueness reporting; private checked workspace; and process-global native serialization. Three-dimensional and non-Cartesian FISHPACK families remain deferred. See the Cartesian FISHPACK guide.

The hosted fishpack-pois3d feature provides a separate owned f32 facade over the structured discrete POIS3D system. It has typed transverse ghost-node rules plus checked cyclic or noncyclic third-axis coefficients; it is not an arbitrary six-face three-dimensional PDE interface. See the structured POIS3D guide.

The hosted banded-linear-systems feature provides compact general-band f32/f64 LU factors, reusable direct and transpose solves, SGBCO/DGBCO reciprocal 1-norm condition estimates, and SGBDI/DGBDI base-ten scaled determinant metadata. Input is copied only into the private expanded LINPACK layout; no dense conversion or matrix dependency is introduced. Exact singular pivots remain errors, while a successful zero reciprocal estimate is preserved as a numerical diagnostic. See the banded-system guide.

The hosted pchip feature provides owned f32 and f64 PiecewiseCubicHermite curves over reviewed PCHIP. Monotone, controlled, and PCHSP endpoint derivative construction feed native value/derivative evaluation and definite integration. Knots remain strictly ordered as supplied, native endpoint extrapolation is opt-in, and the shared XERROR runtime keeps calls process-serialized. See the PCHIP guide.

The hosted bspline feature provides owned f32 and f64 BSpline curves. Reviewed BINTK/DBINTK construct exact interpolants from strictly increasing nodes, values, a complete caller-supplied knot sequence, and order; BVALU/DBVALU evaluate values or derivatives and BSQAD/DBSQAD integrate. There is no hidden knot-generation policy, sorting, coefficient conversion, or caller-managed workspace. See the B-spline guide. The additive bspline-cubic-interpolation feature supplies typed fixed-cubic BINT4/DBINT4 construction with explicit endpoint derivative and knot-placement policies.

The hosted roots-polynomial feature supplies owned single-precision complex polynomial roots over RPZERO/CPZERO and RPQR79/CPQR79. Iterative drivers preserve documented best roots on their iteration limit; companion-QR nonconvergence is an error because the source does not promise partial output. See the roots guide.

The hosted piecewise-polynomial feature provides owned f32 and f64 right-Taylor PP curves over reviewed PPVAL/DPPVAL evaluation and PPQAD/DPPQAD exact integration. With bspline, it also performs exact BSPPP/DBSPPP B-spline-to-PP conversion. Breakpoints are strictly increasing and never reordered; native extrapolation is not used. PCHIP and PP-to-B-spline conversion remain deferred. See the piecewise-polynomial guide.

The hosted tabulated-data feature provides TabulatedData<f32> and TabulatedData<f64> for finite, strictly increasing sampled data. It creates private Newton polynomial representations with POLINT/DPLINT, evaluates values and derivatives with POLYVL/DPOLVL, derives Taylor coefficients with POLCOF/DPOLCF, and integrates arbitrary-spacing samples with AVINT/DAVINT. It has no callback, caller-owned workspace, or implicit sorting policy. See the tabulated-data guide.

The hosted special-scalar-expanded feature provides real scalar logarithmic/Spence integrals and Carlson symmetric elliptic integrals in both precisions. It has no caller workspace or external numerical dependency; native saved initialization and XERROR keep it globally serialized. Complex, sequence, arbitrary-workspace, and translated special-function APIs remain deferred. See the special-functions guide.

The narrower special-airy feature exposes checked f64 and f32 real Ai and Bi values plus their explicitly exponentially scaled variants through slatec::special::airy. On non-positive inputs the scaled values equal their unscaled counterparts. On positive inputs Ai is multiplied by exp(2 x^(3/2) / 3) and Bi by exp(-2 x^(3/2) / 3). The selected FNLIB set has no separately promoted real derivative driver; complex Amos drivers and Airy subsidiaries remain deferred. See the Airy guide.

Native concurrency is conservative. APIs that use the hosted legacy runtime are process-serialized; the existing no_std/alloc BLAS and Jacobian-check features remain backend-dependent and do not promise parallel native safety. Matrix and vector storage is passed only under each function's documented Fortran layout contract, with no implicit repacking or transposition. See the repository runtime concurrency and storage policy.

Native implementations use exactly one provider: bundled, source-build, system, or external-backend. bundled is the canonical default feature, and distributes the ten reviewed special/root families on both supported bundled targets. Select an expert provider explicitly for an unavailable family.

Safe family features do not create a numerical implementation root. On the supported GNU MinGW source-build release profile, each safe operation is laid out to retain only its wrapper, checked shared support, and genuine native closure. BLAS Level 1 is isolated per operation and precision; sampled Level 2/3 operations provide the same regression boundary. See the safe-facade link-granularity policy.