pyifb

Python bindings for Fortran's ISO_Fortran_binding.h descriptors (CFI_cdesc_t).

pyifb provides:

• A low-level extension module (pyifb.ifb) exposing C descriptor structs, constants, and CFI helper functions.
• A high-level Python wrapper (pyifb.CFI_cdesc) that makes descriptor allocation and array exchange feel NumPy-native.

This is useful when interoperating with Fortran code that uses bind(C) and assumed-shape/descriptor-based interfaces.

Features

• Wraps CFI_cdesc_t and CFI_dim_t from ISO_Fortran_binding.h
• Exposes CFI constants/macros (for example: CFI_SUCCESS, CFI_type_int, CFI_MAX_RANK)
• Supports creating, serializing, and deserializing descriptors
• Supports descriptor operations such as allocation, deallocation, sectioning, pointer setup, and contiguity checks
• Includes a high-level CFI_cdesc.value property for exchanging data with NumPy arrays

Requirements

• Python 3.10+
• NumPy (installed automatically as a dependency)
• A C compiler toolchain for building the extension
• A Fortran runtime available at link/runtime
  • Default build path targets GNU Fortran runtime (libgfortran)
  • Alternative compiler/runtime behavior is handled in setup.py

Install

Install from source:

python -m pip install .

For editable development install:

python -m pip install -e .

Build

Build wheel/sdist:

python -m build

Quick Start

High-level descriptor wrapper

import numpy as np
import pyifb

# Create rank-2 descriptor
cdesc = pyifb.CFI_cdesc(rank=2)

# Assign NumPy data (allocates descriptor storage if needed)
cdesc.value = np.array([[1, 2, 3], [4, 5, 6]], dtype=np.int32)

print(cdesc.rank)      # 2
print(cdesc.shape)     # (2, 3)
print(cdesc.elem_len)  # 4

# Round-trip back to NumPy
arr = cdesc.value
print(arr)

Low-level API access

import pyifb

# Access raw constants and structs from the extension
print(pyifb.ifb.CFI_MAX_RANK)
print(pyifb.ifb.CFI_SUCCESS)

cdesc_t = pyifb.ifb.CFI_cdesc_t(rank=1)
status = cdesc_t.allocate([0], [9], 4)
print(status == pyifb.ifb.CFI_SUCCESS)

Using pyifb With Fortran Code

pyifb is designed to work with Fortran procedures exposed through bind(C). The typical workflow is:

1. Write a bind(C) Fortran routine.
2. Build it as a shared library (.so/.dylib/.dll).
3. Create a CFI_cdesc in Python and pass its serialized descriptor to the Fortran routine with ctypes.

Fortran side (bind(C) + assumed-shape)

module finterop
  use iso_c_binding
  implicit none
contains

  integer(c_int) function sum_desc(x) bind(C, name="sum_desc")
    integer(c_int), intent(in) :: x(:)
    integer(c_int) :: i

    sum_desc = 0
    do i = 1, size(x)
      sum_desc = sum_desc + x(i)
    end do
  end function sum_desc

end module finterop

Build a shared library (GNU example):

gfortran -shared -fPIC -O2 finterop.f90 -o libfinterop.so

Python side (pyifb + ctypes)

import ctypes
import numpy as np
import pyifb

lib = ctypes.CDLL("./libfinterop.so")
sum_desc = lib.sum_desc
sum_desc.restype = ctypes.c_int
sum_desc.argtypes = [ctypes.c_void_p]

# Build a rank-1 descriptor and populate from NumPy.
cdesc = pyifb.CFI_cdesc(rank=1)
cdesc.value = np.array([10, 20, 30], dtype=np.int32)

# Pass descriptor bytes by pointer.
raw = cdesc.to_bytes()
buf = ctypes.create_string_buffer(raw, len(raw))

result = sum_desc(buf)
print(result)  # 60

Notes and pitfalls

• Use Fortran interoperable kinds (integer(c_int), real(c_double), etc.) from iso_c_binding.
• Match NumPy dtype to Fortran kind (for example, np.int32 <-> integer(c_int)).
• CFI_cdesc.value stores arrays in Fortran (column-major) order.
• For scalar descriptors, use CFI_cdesc(rank=0) and assign a scalar NumPy value.
• For routines that expect allocatable/pointer dummy arguments, descriptor attributes must also match (CFI_attribute_allocatable/ CFI_attribute_pointer).

Testing

The test suite builds a small Fortran shared library in tests/ (via CMake) and then runs pytest.

Run tests:

pytest

Or use the helper script:

./build_and_test.sh

Development

Useful scripts in the repository root:

• build.sh
• build_and_test.sh
• lint.sh
• coverage.sh

Optional dependency groups (from pyproject.toml):

• test
• coverage
• dev
• perfomance

Install development extras (example):

python -m pip install -e ".[dev,test]"

Compilers

By default we compile with gcc/gfortran and that is the version shipped in PyPI. We support other compiliers:

• Intel ICC/IFX
• Clang

These must be built locally, to do set the environement variables:

export CC=icx
export FC=ifx

or

export CC=clang
export FC=flang

Then install the package

python -m pip install -e .

Other compilers can be support on request, if they have a publicly readable ISO_Fortran_Bindings.h file. Please open a bug request for a new compiler.

License

GPL-2.0-or-later. See COPYING.txt.