devernay/cminpack v1.3.15

cminpack 1.3.15

on GitHub

Performance

• qrfac and qform are about 2× faster (up to 2.8× in some builds): each Householder transformation is now applied to several columns at a time (contributed by @xizhexin, #80). All solvers that factor the Jacobian benefit: lmdif, lmder, lmstr, hybrd, hybrj. The number of columns is set at compile time with CMINPACK_BLOCK (default 4; -DCMINPACK_BLOCK=8 is faster on arm64).

  qrfac with pivoting, ms, before → after:

  m × n	x86_64, gcc 13 -O2	x86_64, gcc 13 -O3	arm64, gcc 16 -O2	arm64, Apple clang -O2
  2000 × 100	20.4 → 9.3	17.0 → 7.3	11.0 → 4.3	7.4 → 4.8
  500 × 200	17.8 → 7.7	14.9 → 5.9	9.6 → 3.9	5.8 → 3.8

  qform on an n × n matrix, ms, before → after:

  n	x86_64, gcc 13 -O2	x86_64, gcc 13 -O3	arm64, gcc 16 -O2	arm64, Apple clang -O2
  100	0.69 → 0.29	0.65 → 0.23	0.59 → 0.25	0.37 → 0.19
  300	18.1 → 7.4	15.3 → 5.6	9.27 → 3.29	5.11 → 2.90
  600	145 → 62	122 → 47	73.4 → 26.0	40.8 → 21.9

• New examples tlmderfd.c and thybrjfd.c show how to compute the finite-difference Jacobian in parallel with OpenMP, by passing an lmdif/hybrd function to lmder/hybrj. No change to the library is needed. On an 18-core arm64 machine, with m = 20000 and n = 30, lmder takes 100 ms with 8 threads, where lmdif takes 309 ms.

Library fixes

• lmstr now returns the user's negative iflag when the user aborts during the Jacobian (#81).
• fdjac1 and fdjac2 now restore x when the user aborts (#89).
• hybrd and fdjac1 no longer skip the Jacobian when ml or mu is huge, e.g. INT_MAX (#88).
• The work-array size checks of hybrd, hybrj and the lmdif1, lmder1, lmstr1, hybrd1, hybrj1 drivers no longer overflow, and reject a negative size (#87).
• In long double, enorm now scales correctly on platforms where long double has the range of double (#82).
• USE_LAPACK: qrfac no longer reads past a when m < n (#84), works when wa and rdiag are the same array (#85), and handles allocation failures and LAPACK errors.
• USE_BLAS: the integers passed to BLAS now have the BLAS integer type (#86). Generic BLAS/LAPACK libraries can be used on macOS (#157, #158).
• The Windows DLL now exports the low-level routines (#91).
• The library sources can be built with CMAKE_UNITY_BUILD (#90) and with -Werror=comment (#83).
• CUDA port: it builds without warnings with recent nvcc versions (#190).

C++ wrapper (cminpackcpp.hpp)

• Accept a plain function passed by name (#92).
• Call a const callable as const (#93).
• Rethrow the callable's exceptions after the solver returns, instead of letting them unwind through C code (#94).

CMake and pkg-config

• C99 is required (#162).
• An invalid CMINPACK_PRECISION is rejected with a clear message (#109). An empty CMAKE_BUILD_TYPE is treated as unset (#110).
• The tests now run for every precision that is built (#108, #168).
• The pkg-config file lists the BLAS/LAPACK libraries correctly (#106), defines CMINPACK_NO_DLL for static builds (#107), and is relocatable, including on Windows (#155, #186).

Tests and examples

• Fix bugs in the FORTRAN and C test drivers and in the Jacobians used by the tests (#95, #96, #97, #98, #99, #100). The FORTRAN references were regenerated from the pure-FORTRAN drivers.
• The cross-check against FORTRAN MINPACK is stricter: it checks the least-squares residuals (#103) and the chkder error vectors (#99), and reports NaN residuals (#152). cmpfiles now accepts FORTRAN D exponents (#151) and alternative values a|b (#154), and has a smaller tolerance (#102).
• Remove the unused and buggy fortran/minpack.f90 (#115).

CI

• GitHub Actions only, with one job per configuration. New: macOS jobs (Apple clang and GCC, with BLAS/LAPACK) and a CUDA build (CUDA 13.0 and 12.4). AppVeyor and Travis are removed.

Thanks

Thanks to @xizhexin for the blocked Householder transformations in qrfac (#80), which started this release's speed-ups.