2.1.1

2.1.1

Performance and correctness. The library API is unchanged from 2.1.0: no signature moved, so this is a
drop-in update.

Performance

• Small matrix products no longer pack a megabyte. gemm_core allocated its packing panels at the
  largest supported block -- 128x256 plus 256x256 f64, 768 kB zeroed -- on every call, including the
  3x8 by 8x3 product of seventy-two multiplications that element assembly makes 160 times per step.
  Below 1024 multiplications the product is now computed directly, and above that the panels are sized
  by the block in use. A 3x8 by 8x3 product went from 8 600 ns to 60 ns, a 24x24 product to
  798 ns (34.7 GFLOP/s), and 500x500 is unchanged at 6.33 ms (39.5 GFLOP/s).
• The supernodal front eliminates down the columns. The panel is column-major and the update
  walked its rows, one cache miss per element. Swapping the loops took the front factorization on a
  192-dof lattice from 2.989 ms to 0.633 ms, which is what let the supernodal path overtake the
  simplicial one.

Fixed

• The AMG hierarchy builds when a row has no diagonal entry. A structural matrix may have one: a
  constrained dof, a pure drift row. The smoother then divided by zero and the prolongator weighted by
  it. The hierarchy's own copies now carry a unit diagonal where the diagonal is exactly zero (a full
  diagonal is still shared, not copied), and a coarsest operator that comes out semi-definite is
  factorized with a relative diagonal shift between 1e-12 and 1e-6 of its largest entry. A coarsest
  level that still will not factor now relaxes instead of failing the solve.
• krylov_solve solves and measures the caller's matrix. The fill above had been inserted into
  level 0 while the Krylov operator came from the hierarchy, so the method solved a different system
  than the caller handed in: on the benchmark, Goodwin_040 reported a relative residual of 7.8e-11
  with a solution 68 times away from the reference (123x on Goodwin_071, 220x on Goodwin_127).
  The operator is now the caller's matrix -- the hierarchy keeps its adjusted copies for the
  preconditioner only -- and the residual is measured against it. The cost is visible in the iteration
  count: 430 instead of 74 on Goodwin_040, and the honest BiCGStab yield on the 33 real matrices is
  12, not the 13 the defective path reported.

Documentation

• The supernodal factorization was described as slower and experimental. It is faster on the banded
  and lattice patterns that dominate element assembly (0.604 ms against 1.248 ms), so the module
  documentation now carries that measurement instead of the claim, and explains why the two callers
  pass different flop thresholds (LuMethod::auto() 40.0, russell_sparse 1.0).
• The README gains a Performance section with the numbers above, and an Iterative solvers section with
  the yield measured on 33 real SuiteSparse matrices and the matrices where that path earns its place:
  atmosmodd and atmosmodl do not fit in memory for UMFPACK and take 7.78 s and 5.70 s with
  BiCGStab, against 46.2 s and 40.9 s with MUMPS.

Testing

282 tests, 2 ignored, including a coarse operator built from a dependent row and a matrix with an
isolated unknown.

Measurements

Arch, Intel i9-13900KF, release builds. The solver numbers come from
benchmark-russell; the element-assembly numbers come
from pmsim's stress-strain driver.

2.1.0

2.1.0

Documentation and test wiring. The library API is unchanged from 2.0.0.

README

• Describes the 2.0 array-based API: SparseLu over caller-owned CSC arrays, and
  amg_setup/krylov_solve over CSR arrays.
• Provenance moves into two bullets: the sparse LU and its COLAMD ordering come from faer (MIT),
  sparse::amgcl from AMGCL (MIT).
• Adds a SparseLu example that reuses one analysis across value updates, and states that at three
  unknowns the AMGCL hierarchy is a single level the coarse solve handles.
• The conventions name the error type of each half (LapackError, SparseError, amgcl::Error).
• The performance table calls the russell_sparse solver Direct, which is its name from
  russell_sparse 5.0; the numbers were measured with 4.x, where it was called Native.

Testing

• src/lib.rs exposes README.md through #[doc = include_str!(...)] on a #[cfg(doctest)] type,
  so cargo test runs both of the page's examples as doctests.
• The CI Test step runs cargo test instead of cargo test --lib, so those examples cannot rot
  unnoticed.

Full changelog: v2.0.0...v2.1.0