Skip to content

perf(amr): packed-GPU diffusion kernels and conservative regrid transfer #59

Closed

Description

@kylebeggs

Stage 4 of #56. The single-launch packed-GPU kernels for the compact flux-form Diffusion leaf, and the conservative-transfer work regrid! needs before a conserved field can ride an adaptive forest. Depends on stage 3.

Packed-GPU kernels

Add _forest_sweep! and _forest_adjoint_sweep! overrides for Diffusion over a PackedBlockField, in src/operators/forest_packed.jl. Follow the file's established shape:

  • open with the standard guard — backend isa KernelAbstractions.GPU || return _forest_sweep_leaves!(...) — because broadcast fusion beats KernelAbstractions CPU codegen by roughly 1.6× measured, so the CPU path must keep falling back;
  • reuse the same per-cell stencil body the broadcast path uses (_diff_at / diffusion_stencil) so the numerical definition cannot fork, exactly as the existing kernels reuse _lap_at;
  • recompute per-leaf geometry from the levels SoA via _leaf_inv_h2 / _halo_cell, and read κ at the halo-offset interior index the way _scalefield_forest_kernel! does.

Note that Advection currently has no packed adjoint kernel and degrades to the per-leaf sweep; Diffusion should ship both directions rather than inherit that gap.

PackedBlockField coefficient packing lands in stage 3 via _prepare_tree; this stage makes the kernels use it.

AMR conservation (#54 §4)

Independent of the kernels, and the reason this stage is not purely a performance change. Today regrid! coarsens with a conservative 2^-N child mean but prolongs with linear interpolation, which does not preserve the volume-weighted mean of the children.

  • Add mean-preserving prolongation: the volume-weighted mean of the new children must equal the old parent exactly, including at block and physical boundaries.
  • Keep the existing linear-exact interpolation available for non-conserved smooth fields — it is the right transfer for those, and this is not a replacement.
  • Add a slope-limited prolongation policy before using this path for discontinuous solutions.
  • Make conservation semantics explicit per field, so geometric indicators, coefficients, and conserved state do not all silently inherit the same transfer policy. This is the piece most likely to cause a quiet wrong answer later; it should not be a default.
  • The policy must not add runtime branching to operator hot paths — attach it to the field, resolve it at transfer time.

Acceptance

  • CPU/GPU and packed/reference parity within existing tolerances, for forward and adjoint, both averaging policies. Gated tests only run under MFO_TEST_GPU=true with CUDA available, so a green local suite proves nothing here — read the CI matrix.
  • Prepared applications stay allocation-free on the established CPU hot paths and keep a leaf-count-independent launch structure on packed GPUs.
  • Σ V·u is preserved to roundoff across refinement and coarsening for fields using conservative transfer, including balance-induced refinements (the 2:1 balance pass can refine blocks the user never marked).
  • The adapted Niederer-style benchmark performs no more halo exchanges per diffusion application than before, with a phase breakdown comparable to meta(amr): forest performance — the coarse–fine exchange, not regridding, is the AMR tax #49.

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions