The four encodings

The last field of every FRD block header is a format code, and there are four of them: two ASCII widths and two binary ones. CalculiX writes binary unconditionally from *REFINE MESH and on demand from a DOUBLE output card, and a reader that handles only the ASCII codes returns an empty mesh for those files without raising anything.

This package reads and writes all four, and can convert between them without going through a mesh at all.

from pathlib import Path
import tempfile

import matplotlib.pyplot as plt

import pyvista_frd

# Sphinx-Gallery runs each example from its own directory; the decks and
# their solved output live one level up, in ``doc/_data``.
DATA = Path('../_data').resolve()
tmp = Path(tempfile.mkdtemp())

pyvista_frd.convert() reads a document and writes it out again in another encoding. Nothing is re-derived: the records are re-spelled and every line the parser does not need to understand is copied through untouched.

source = DATA / 'cantilever.frd'
sizes = {'long ASCII (source)': source.stat().st_size}

for label, kwargs in (
    ('binary float64', {'binary': True}),
    ('long ASCII', {'binary': False}),
):
    target = tmp / f'{label.replace(" ", "_")}.frd'
    pyvista_frd.convert(source, target, **kwargs)
    sizes[label] = target.stat().st_size

for label, size in sizes.items():
    print(f'{label:22s} {size / 1024:8.0f} KB')

long ASCII (source)         330 KB
binary float64              181 KB
long ASCII                  330 KB

Binary is about a third of the size, and exact: the values are the stored doubles rather than a six-significant-digit rendering of them.

fig, ax = plt.subplots(figsize=(6, 3))
ax.barh(list(sizes), [v / 1024 for v in sizes.values()], color=['#4c72b0', '#dd8452', '#55a868'])
ax.set_xlabel('KB')
ax.set_title('Same document, three encodings')
fig.tight_layout()
plt.show()

The conversion is lossless in the direction that matters and the mesh is unchanged either way.

original = pyvista_frd.read(source)
converted = pyvista_frd.read(tmp / 'binary_float64.frd')
print(f'points  {original.n_points} -> {converted.n_points}')
print(f'cells   {original.n_cells} -> {converted.n_cells}')
print(f'arrays  {len(original.point_data)} -> {len(converted.point_data)}')

points  900 -> 900
cells   600 -> 600
arrays  15 -> 15

Going the other way – binary to ASCII – is the conversion the format code has always implied and no tool offered. A binary FRD that no ASCII-only reader can open becomes one any of them can, including CalculiX’s own.

One caveat, recorded in doc/divergences.md: CalculiX casts every value to float before printing it, and this writer does not. The text it produces is the nearest six-digit decimal to the number actually held, which differs from CalculiX’s own at a rounding tie – about 3% of record lines.