Features

GPU backends

The maintained GPU backend is FNL, one source tree with two device paths:

Modes	Macro	Compiler	Parallelism model
*-fnl-nvf	_FNL	NVIDIA nvfortran	OpenACC + FUNDAL memory management (NVIDIA GPUs)
*-fnl-omp-amd	_FNL, DEV_OMP	AMD amdflang	OpenMP target offloading + FUNDAL (AMD GPUs)

The kernels carry both directive sets (!$acc and !$omp twins). The src/lib/nvf (CUDA Fortran) and src/lib/gmp (OpenMP offloading) libraries are in the tree, but no PRISM or FLUME build mode uses them.

Each application (prism, flume) compiles to separate executables per backend, sharing a common source tree under src/app/<name>/common/.

---

Adaptive Mesh Refinement

ADAM uses a Morton-order linearized octree/quadtree for AMR:

• Blocks are sorted by Z-order (Morton) curve for cache-friendly traversal
• FLUME's refinement and coarsening conserve the state: coarsening restricts by the cell mean, refinement prolongs with a conservative (MC-limited) reconstruction, and the 2:1 faces are refluxed
• Ghost cell exchange via MPI before every stencil operation
• Configurable ghost cell width ngc (WENO5: 3, WENO7: 4)

---

Immersed Boundary Method

Complex and moving geometries are handled without body-fitted meshes:

• Solid body positions are tracked on a background Cartesian grid
• IB forcing reconstructs boundary conditions at cut cells
• Eikonal-based signed distance field for geometry representation
• Compatible with AMR: in FLUME an AMR_GEO_SOLID marker refines around a solid (PRISM does not implement it)

---

WENO Schemes

Weighted Essentially Non-Oscillatory finite difference reconstructions:

Scheme	Order	Ghost cells
weno-u-1	1st (upwind)	ngc = 1
weno-u-3	3rd (upwind)	ngc = 2
weno-u-5	5th (upwind)	ngc = 3
weno-u-7	7th (upwind)	ngc = 4
weno-u-9	9th (upwind)	ngc = 5
weno-c-2 … weno-c-8	2nd–8th (centred)	ngc = order/2

Nonlinear weights: Jiang–Shu (js, the default) or scale-invariant (si). Reconstruction variables: conservative or characteristic. ROR iterations lower the stencil where a variable check fails, and the order is reduced close to immersed solids.

---

Temporal Integration

Explicit Runge-Kutta multi-stage methods:

• runge-kutta-1, -2, -3; SSP runge-kutta-ssp-11, -22, -33, -54; the 4-stage symplectic runge-kutta-yoshida
• CFL-based time step control
• Leapfrog, Blanes–Moan (blanes-moan4) and commutator-free Magnus (CFM-4) integrator objects

---

Field Memory Layout

5D field arrays (nv, ni, nj, nk, nb) with Fortran column-major storage:

• nv — number of variables (density, momentum, energy, …)
• ni, nj, nk — grid cells per block (including ghost cells)
• nb — number of blocks (AMR + domain decomposition)

The leftmost index nv is contiguous in memory and is the innermost loop for CPU vectorization and GPU coalesced access.

---

Parallel I/O

• Field output: one HDF5 file per MPI rank (<basename>-procNNNNNN.h5) plus one XDMF file (<basename>.xdmf) for ParaView/VisIt
• Restart: Fortran stream files, the blocks per rank (-procNNNNNN.fbd), the tree nodes (.tnd) and the time (.time)
• Slices: raw binary written with MPI-IO (.mat), converted by ASCOT or slices_converter.py

---

Compiler Support

Compiler	GPU backends	CPU
NVIDIA nvfortran 26.9 (tested)	FNL (OpenACC)	Yes
GNU gfortran 16.2.0 + Open MPI 5.0.11 (tested)	—	Yes
AMD amdflang	FNL (OpenMP offloading, *-fnl-omp-amd)	Yes (prism-cpu-amd)

---

Design Principles

• KISS — keep it simple; prefer explicit Fortran over macro magic
• Portable — a single source tree, per-backend preprocessor guards only where unavoidable
• Evidence-based — profile first, optimize with benchmarks, document trade-offs
• Standard-conforming — Fortran 2003+ OOP; implicit none everywhere; explicit intent on all arguments