feat(src): add README.md

src/assemblers/dof_based/README.md | 50 ++++++++++++++++++++++++++++++++++++++ 1 file changed, 50 insertions(+)
2026-09-12 06:22:00 +00:00 · 2026-05-09 16:30:31 +03:00
parent e5b80fd171
commit b87aa90a9c
1 changed files with 50 additions and 0 deletions
@@ -0,0 +1,50 @@
+# src/assemblers/dof_based/
+
+DOF-by-DOF assembler. The driver loops over global DOFs (rather than
+elements) and writes one row of the global system at a time. This
+removes the per-element gather/scatter overhead, eliminates atomic
+adds for GPU execution, and lets the same kernel run on CPU and on
+every KernelAbstractions backend without code duplication.
+
+## Files
+
+- `dof_based_coo.jl` — CPU implementation. Defines `DOFBasedCOOAssembler`, `DOFBasedCOOCache`, and the matrix-free entry points (`apply_K!`, `apply_K_contributions!`, `apply_M!`, `assemble_M!`, `extract_system`). Built around the microkernel contract in `assemblers/microkernel.jl` and the DOF connectivity in `dofs/dof_connectivity.jl`.
+- `dof_based_coo_ka.jl` — Backend-agnostic GPU port via `KernelAbstractions.jl`. Defines `DOFBasedCOOCacheKA`, `sync_from_cpu!`, and the precision helpers used by the Float32 GPU pipeline. The same kernel runs on `CPU()`, `CUDABackend()`, `MetalBackend()`, `AMDGPUBackend()`, `oneAPIBackend()`; the in-tree CI validates the CPU() backend.
+
+## Design notes
+
+The hot path is intentionally a tight loop over global DOFs; a Pass 1
+fills shape-function values, derivatives and detJ for every quadrature
+point in the element batch, and Pass 2 evaluates the per-row entries
+via `evaluate_entry`. The constraint and load infrastructure in
+`assemblers/matrix_free/` plugs into this same Pass 1 / Pass 2 split.
+
+Zero-allocation invariants are enforced by
+`test/assemblers/test_dof_based_zero_alloc.jl`.
+
+## Partitioned element sets (distributed matvec stepping stone)
+
+`apply_K!` visits each global DOF once and sums every incident element;
+`apply_K_contributions!` loops only over a subset of elements and adds into
+`y` (`+=`). If two subsets are disjoint and their union is all elements,
+summing the partial vectors equals `apply_K!` on the same `x` (reference
+test: `test/assemblers/test_dof_based_partitioned_matvec.jl`). If subsets
+overlap, contributions double-count until the overlap is handled elsewhere.
+
+Lightweight partition metadata (`MeshPartitionLayout`,
+`element_indices_for_part`, `brick_hex_partition_slabs`, `referenced_global_dofs`,
+`element_counts_by_part`, ...) lives in `assemblers/partitioning.jl`. For MPI-style
+staging without GC traffic, preallocate a DOF-sized [`BitVector`](@ref) and use
+`mark_referenced_dofs!`, `collect_true_indices!`, `fill_referenced_dof_indices!`,
+`ghost_dof_mask!`, `node_partition_owner_min!`, and `mark_owned_vertex_field_dofs!`
+(zero allocations once buffers exist — see `test/assemblers/test_partitioning_zero_alloc.jl`).
+Neighbor partitions and MPI-style send/recv DOF lists live in `assemblers/halo_exchange.jl`
+(`build_partition_adjacency`, `build_rank_halo_exchanges`, `ReferenceMaskMultiplyLayout`).
+For `MatrixFreeOperator`, the buffer filled before `apply_K!` is controlled
+by `AbstractMultiplyGhostLayout` / `prepare_multiply_workspace!` (default:
+copy `x`), which is the extension point for ghost DOFs under MPI.
+
+MPI helpers (weak dependency `MPI`, extension `ext/JuliaFEMMPIExt.jl`): halo exchange
+`exchange_matvec_halos_mpi!`, owned-row product `mpi_partitioned_operator_matvec_owned!`,
+workspace `partitioned_mpi_owned_matvec_workspace`, and reusable request buffers via
+`allocate_exchange_matvec_halo_mpi_requests` / keyword `mpi_requests` (see `AGENTS.md` §3.6).