Numeric backends#

The core numeric path handles ordinary NumPy arrays and ordinary complex numbers. Plugins can provide additional execution branches for specialized array types without adding those dependencies to the core.

A plugin can register:

  • ndmath.execution_branch to identify a plugin-owned numeric branch;

  • ndmath.execute to run an operation for that branch;

  • ndmath.numpy_method.<name> to override selected NumPy-like methods;

  • fft.encoding for domain-specific FFT acquisition encodings.

  • fft.validate_input to reject unsupported FFT/IFFT requests before preparation can modify the dataset.

Example:

def register_handlers(self) -> dict:
    return {
        "ndmath.execution_branch": execution_branch,
        "ndmath.execute": execute,
    }

Handlers should return None when they do not recognize the data so the core default path remains available.

FFT input validation#

fft.validate_input(dataset, *, dim, inv, size, sizeff, **kwargs) receives the original dataset and a resolved dimension name. It runs after the core’s domain checks, before copying, swapping axes, effective-size zeroing, zero-filling, truncation, numerical transformation or coordinate updates. The original keyword options, including inplace, are available in kwargs. This applies to method and functional FFT/IFFT entry points.

Validators must treat the dataset as read-only. Return None to accept the request and allow other registered validators to run, or raise an exception to reject it. Exceptions propagate to the caller before preparation; this is not a rollback mechanism for changes made by a validator itself. With no registered validator, the ordinary core path is unchanged.

Plugins can detect this contract using FFT_PREVALIDATION_API >= 1 in spectrochempy.processing.fft.fft. Older cores do not expose that marker; plugins requiring prevalidation must refuse the affected workflow on them.