Skip to content

Releases

SOLVAX 0.36.2

Choose a tag to compare

@rogeriojorge rogeriojorge released this 13 Nov 03:28
50a9168

block_thomas_solve runs both substitution sweeps as scans, so its program no longer grows with the block count (31 x 150 cold compile about 9 s -> about 1 s, warm solve unchanged). fast_diagonal, flux_form_operator and fast_diagonal_schur accept a (lower, upper) pair of "dirichlet"/"neumann" ends per axis (#200, closes #172). generalized_eigenpairs uses six shift-invert applications per sweep (#193).

SOLVAX 0.36.0

Choose a tag to compare

@rogeriojorge rogeriojorge released this 10 Nov 19:48
7a8baa3

generalized_eigenpairs now extracts pairs by Rayleigh-Ritz for the shift-invert operator, over a block Krylov window of three applications per sweep. A target between an isolated eigenvalue and an accumulating set no longer stalls on Ritz ghosts. On FAX's Alfvén pencils at N=800 that is 10 sweeps instead of 83 iterations. iterations and max_iter now count sweeps.

SOLVAX 0.35.3

Choose a tag to compare

@rogeriojorge rogeriojorge released this 10 Nov 04:25
12c9ca3

Reverts the right-side Schur coupling of 0.35.2, which made the factored and unfactored block-Thomas routes disagree beyond rounding on ill-conditioned bands. Adds implicit differentiation by default for block_thomas and mixed_precision_block_thomas.

SOLVAX 0.35.2

Choose a tag to compare

@rogeriojorge rogeriojorge released this 09 Nov 04:42
5c85176

Block-Thomas factorization forms the dense Schur coupling with right-side triangular solves: 10-25% faster per Schur step on one CPU thread for dense blocks of 64 to 800, equal to rounding.

SOLVAX 0.35.1

Choose a tag to compare

@rogeriojorge rogeriojorge released this 06 Nov 05:15
3e09423

Correct broadcast coefficient bands in tridiagonal solves (#181). Share the checked real Thomas CUDA forward sweep with pivot diagnostics, preserving fallback and differentiation behavior (#180).

SOLVAX 0.35.0

Choose a tag to compare

@rogeriojorge rogeriojorge released this 02 Nov 02:37
59b8c31

Full-node helpers shared by the uwplasma codes: available_cpus, map_independent, shard_batch(check_vma=) and chunk_map(mesh=). Additive; the JAX floor is unchanged.

SOLVAX 0.34.1

Choose a tag to compare

@rogeriojorge rogeriojorge released this 27 Oct 19:14
260309e
  • block_cyclic_reduction_factor(..., factor_dtype=...) keeps the complex kind: complex bands with a real low-precision factor_dtype no longer drop their imaginary parts in the reduction levels (#175).

SOLVAX 0.34.0

Choose a tag to compare

@rogeriojorge rogeriojorge released this 25 Oct 16:04
be5e650

Version bump for #141, #142, #169 and #170.

  • flux_form_operator with a separable fast_diagonal preconditioner, and inverse_iteration with a residual certificate and an eigenvalue-only derivative contract (#169).
  • fast_diagonal_schur: the exact constrained stage by PCG on the Schur complement, with axis-weighted fast_diagonal (#170).
  • pcg_linear_solve returns NaN for a nonfinite cotangent in the transposed solve (#169).
  • Fixed-plan differentiable exponential actions (#142); propagator eigenpair certification rejects zero and nonfinite vectors (#141).

SOLVAX 0.33.0

Choose a tag to compare

@rogeriojorge rogeriojorge released this 24 Oct 18:35
7b8fb23

Version bump for #163, #164 and #165.

  • block_thomas_factor(..., normalized=True) keeps the normalized lower couplings, so primal and transposed solves apply each Schur LU once instead of twice (#163).
  • A real factor_dtype given complex bands now selects the complex dtype of the same width, where it previously discarded the imaginary part (#164).
  • block_cyclic_reduction_factor adds block cyclic reduction with a normalized-Thomas leaf, accepted by block_thomas_solve (#165).

SOLVAX 0.32.1

Choose a tag to compare

@rogeriojorge rogeriojorge released this 24 Oct 06:17
6f1f130

generalized_eigenpairs: the eigenvector tangent through an exactly degenerate cluster is the finite gauge-fixed subspace tangent (no intra-cluster rotation) instead of NaN, so reverse-mode cluster means through ties no longer pick up NaN·0 on some XLA builds (#166).