v0.4.2

Added

• (bench) a benchmark harness: fixed problems timed at a stated accuracy, and photonoxide bench (#135)
• (fdfd) GMRES preconditioned by a multigrid cycle for high-contrast 3D problems (#144)
• (job) job files in JSON and YAML as well as TOML, the same job in each (#145)
• (fdfd) QMR for complex symmetric matrices on the curl-curl operator's symmetric similarity, twice as fast (#149)
• (bench) the bandwidth each iterative solve reaches, its bytes counted by the kernels, against the triad (#152)
• (bench) the direct solvers' systems exported as Matrix Market files, and photonoxide's factorization timed on them (#153)

Fixed

• (fdfd) a 3D port's guided mode whose plane crosses PMLs goes forward, not backward (#136)

Performance

• (job) a sweep's points solved side by side on rayon's threads, recorded in order (#146)
• (fdfd) the 3D direct solver orders its LU by nested dissection, twice as fast as COLAMD at 40³ cells (#147)
• (fdfd) QMR's vector work fused into two passes on rayon's threads, deterministic sums, nothing allocated an iteration (#148)
• the examples that sweep by hand solve their points side by side, through parallel::map_in_order (#151)

Documentation

• (fdfd) parallel ILU(0) measured on our matrices, both halves, and not used (#150)
• photonoxide's direct solvers against PARDISO and MUMPS, where the difference comes from, and what would close it in pure Rust (#155)