Backend helpers¶

backend ¶

NumPy/CuPy backend selection and array-module helpers.

A thin wrapper around the array-API differences between NumPy and CuPy so the rest of the package can be written once and run on either CPU or GPU. Also exposes a few low-level helpers (memory probe, FFT-size rounding, dtype pairs) that are shared between the spectrogram and Gabor stages.

get_array_module ¶

get_array_module(use_gpu: bool = False)

Return the active array module (numpy or cupy).

PARAMETER	DESCRIPTION
`use_gpu`	If True and CuPy is available, return `cupy`. Otherwise return `numpy`. TYPE: `bool` DEFAULT: `False`

RETURNS	DESCRIPTION
`module`	Either the `numpy` or `cupy` module object.

get_signal_module ¶

get_signal_module(use_gpu: bool = False)

Return the active signal-processing module.

PARAMETER	DESCRIPTION
`use_gpu`	If True and CuPy is available, return `cupyx.scipy.signal`. Otherwise return `scipy.signal`. TYPE: `bool` DEFAULT: `False`

RETURNS	DESCRIPTION
`module`	Either `scipy.signal` or `cupyx.scipy.signal`.

next_fast_len ¶

next_fast_len(n: int, use_gpu: bool = False) -> int

Round n up to an FFT-friendly length.

CPU uses :func:scipy.fft.next_fast_len (product of small primes). GPU uses the next power of two, which matches cuFFT's optimal path.

PARAMETER	DESCRIPTION
`n`	Minimum length required. TYPE: `int`
`use_gpu`	Switches between the CPU and GPU rounding rules. TYPE: `bool` DEFAULT: `False`

RETURNS	DESCRIPTION
`int`	The smallest fast length `>= n`.

get_available_memory ¶

get_available_memory(use_gpu: bool = False) -> int

Best-effort estimate of free memory in bytes.

On GPU, reports free VRAM via the current CUDA device. On CPU, tries /proc/meminfo first, then :mod:psutil, then falls back to a hard-coded 4 GiB so callers always get a usable number.

PARAMETER	DESCRIPTION
`use_gpu`	If True, report device memory rather than host memory. TYPE: `bool` DEFAULT: `False`

RETURNS	DESCRIPTION
`int`	Available memory in bytes.

get_dtypes ¶

get_dtypes(use_float32: bool = True)

Return a matching (float, complex) dtype pair.

PARAMETER	DESCRIPTION
`use_float32`	If True, return `(np.float32, np.complex64)`; otherwise the double-precision pair. TYPE: `bool` DEFAULT: `True`

RETURNS	DESCRIPTION
`tuple of numpy.dtype`	`(float_dtype, complex_dtype)`.

to_numpy ¶

to_numpy(array) -> ndarray

Return array as a host-side NumPy array, copying from GPU if needed.

PARAMETER	DESCRIPTION
`array`	Input array. May be a numpy array, cupy array, or any array-like accepted by :func:`numpy.asarray`. TYPE: `array_like`

RETURNS	DESCRIPTION
`ndarray`	Host-side numpy array.