fortran-to-python

This skill specializes in converting legacy Fortran code into maintainable, modern Python. It prioritizes the use of numpy for array operations and scipy for scientific computing to ensure performance and clarity.

Refactoring Workflow

1. Analysis phase

• Identify the Fortran version (e.g., F77 fixed format or F90/F95 free format).
• Extract mathematical formulas and core logic.
• Identify external dependencies (e.g., BLAS, LAPACK) that can be replaced by scipy.linalg.

2. Implementation phase

• Array Indexing: Convert 1-based Fortran indexing to 0-based Python indexing.
• Array Order: Note that Fortran is column-major. When using numpy, consider order='F' if direct memory mapping is needed, or transpose for row-major idiomatic Python.
• Control Structures:
  • DO loops -> for i in range(...) or, preferably, vectorized numpy operations.
  • IF blocks -> Python if/elif/else.
  • GOTO -> Refactor into logical loops or functions (DO NOT use goto-like hacks in Python).
• Subroutines & Functions: Map Fortran SUBROUTINE and FUNCTION to Python def. Handle "pass-by-reference" by returning modified values or using mutable objects (like numpy arrays).
• Types: Map REAL*8 or DOUBLE PRECISION to np.float64, INTEGER to int (or np.int32/64), etc.

3. Verification phase

• Ensure that numpy vectorization is used wherever possible instead of explicit loops.
• Check that the logic for boundary conditions (often tricky during index conversion) is preserved.
• Suggest unit tests using pytest to compare numerical results if sample data is available.

Best Practices

• Prefer Vectorization: If a Fortran loop performs element-wise operations, use NumPy broadcasting/vectorization.
• Docstrings: Add descriptive docstrings to the resulting Python functions, including the original Fortran logic if complex.
• Error Handling: Use Python exceptions instead of Fortran error flags or STOP statements.
• I/O: Replace OPEN/READ/WRITE with with open(...) or np.loadtxt/np.savetxt.

Reference Mappings

• MATMUL(A, B) -> A @ B or np.dot(A, B)
• TRANSPOSE(A) -> A.T
• SQRT, SIN, COS, etc. -> np.sqrt, np.sin, np.cos
• ALLOCATE(A(N)) -> A = np.zeros(N) or np.empty(N)
• DEALLOCATE(A) -> del A (or let Python handle GC)