v0.12.5

MultiGridBarrier v0.12.5

Diff since v0.12.4

v0.12.4

MultiGridBarrier v0.12.4

Diff since v0.12.3

v0.12.3

MultiGridBarrier v0.12.3

Diff since v0.12.2

v0.12.2

MultiGridBarrier v0.12.2

Diff since v0.12.1

v0.12.1

MultiGridBarrier v0.12.1

Diff since v0.12.0

v0.12.0

MultiGridBarrier v0.12.0

Diff since v0.11.41

Breaking changes

• Geometry and MultiGrid are now separate types. A mesh constructor returns a single-level Geometry;
  the multigrid hierarchy is a distinct MultiGrid built from it. Build the hierarchy with amg(geom)
  (algebraic multigrid, the new default) or the legacy geometric_mg(geom, L).
• Solver renamed: amgb → mgb_solve, and AMGBSOL → MGBSOL. The solver now takes a MultiGrid, e.g.
  mgb_solve(amg(fem2d_P2()); p=1.5).
• fem2d renamed to fem2d_P2; new fem2d_P1 added. Mesh constructors are now fem1d, fem2d_P1,
  fem2d_P2, fem3d, spectral1d, spectral2d, each returning a Geometry.
• Geometry.x is now an order-3 tensor (V, N, D) — V vertices per element, N elements, D spatial
  dimensions — replacing the flat (V*N, D) matrix.
• find_boundary(geom) now returns Vector{Tuple{Int,Int}} of (vertex, element) pairs, and amg(; dirichlet_nodes=…) takes the same tuple-pair format (enables mixed Dirichlet/Neumann boundary conditions).
• One-shot *_solve convenience functions removed (fem1d_solve, fem2d_solve, fem3d_solve,
  spectral1d_solve, spectral2d_solve). Use mgb_solve(amg(<constructor>()); …) instead.
• GPU API simplified. The *_cuda / *_cuda_solve exports and the multi-GPU distributed code path were
  removed. To run on GPU, using CUDA, CUDSS_jll first, then convert with native_to_cuda(geom_or_mg) and call
  mgb_solve.
• Phase 1 (feasibility/centering) now runs entirely on the fine quadrature rule. The hierarchical
  coarse-quadrature centering was removed; the feasibility subproblem path-follows on the fine rule and
  early-stops on strict feasibility. Internal AMG fields (R_coarse, D_levels, refine_u, coarsen_u,
  refine_z, coarsen_z) and mgb_phase1 were removed.

New features

• amg(geom) algebraic-multigrid hierarchy — solve on a single user-supplied or externally generated
  mesh, with no geometric refinement schedule required.
• Zoo submodule with six convex variational test problems: elasto-plastic torsion, minimal surface,
  vector p-harmonic maps, Norton–Hoff power-law elasticity, ROF total-variation denoising, and the two-sided
  obstacle problem.
• subdivide(geom, L) mesh-refinement utility; convex_piecewise constraint;
  native_to_cuda(::NamedTuple) for running Zoo problems on GPU.
• Structured block-diagonal Hessian assembly shared across CPU and GPU (cuDSS sparse-direct solve on GPU),
  giving a large GPU speedup and ~2× on CPU.

Merged pull requests:

• Bump julia-actions/cache from 2 to 3 (#23) (@dependabot[bot])
• Bump codecov/codecov-action from 5 to 6 (#24) (@dependabot[bot])
• Bump julia-actions/setup-julia from 2 to 3 (#25) (@dependabot[bot])

v0.11.41

MultiGridBarrier v0.11.41

Diff since v0.11.40

v0.11.40

MultiGridBarrier v0.11.40

Diff since v0.11.39

v0.11.39

MultiGridBarrier v0.11.39

Diff since v0.11.38

v0.11.38

MultiGridBarrier v0.11.38

Diff since v0.11.37