CUDA kernel¶

gammatone_kernel ¶

Custom CUDA kernel for batched IIR (SOS) filtering.

Applies a different SOS filter to the same input signal on every channel in a single kernel launch, replacing a Python-level loop of N independent cupyx.scipy.signal.sosfilt calls.

Falls back gracefully: is_available() returns False whenever CuPy or nvrtc cannot be used, and callers should keep the original loop as a fallback path.

is_available ¶

is_available() -> bool

Return True iff CuPy + nvrtc can compile this kernel on this machine.

Triggers a one-time JIT compile of a tiny stub kernel; subsequent calls hit the kernel cache and are essentially free.

RETURNS	DESCRIPTION
`bool`	True if the kernel can be used. False on CPU-only systems or when nvrtc is missing.

batched_sosfilt ¶

batched_sosfilt(sos: 'cp.ndarray', x: 'cp.ndarray', gain: float = 1.0, out: Optional['cp.ndarray'] = None, precision: str = 'float64') -> 'cp.ndarray'

Apply a per-channel SOS cascade to the same input in one kernel launch.

Equivalent to a loop of cupyx.scipy.signal.sosfilt calls, each with its own SOS coefficients but a shared input signal, fused into a single CUDA kernel.

PARAMETER	DESCRIPTION
`sos`	SOS coefficients of shape `(n_channels, n_sections, 6)`. Dtype must match `precision` (float64 for `"float64"`, float32 for `"float32"`). TYPE: `ndarray`
`x`	Shared 1D input signal of shape `(n_samples,)`. Must be float32. TYPE: `ndarray`
`gain`	Per-channel scalar applied to the output before write-back. TYPE: `float` DEFAULT: `1.0`
`out`	Pre-allocated output buffer of shape `(n_channels, n_samples)` and dtype float32. Allocated by the caller when reused across many invocations to avoid per-call allocation. TYPE: `ndarray` DEFAULT: `None`
`precision`	Internal compute precision. `"float32"` is ~8x faster on consumer Ampere (3090ti, Jetson Orin) where FP64 throughput is throttled to 1/64 of FP32. The float32 path matches scipy to ~1e-3 worst-case relative error; the float64 path to ~1e-9. TYPE: `('float64', 'float32')` DEFAULT: `"float64"`

RETURNS	DESCRIPTION
`ndarray`	Filtered output of shape `(n_channels, n_samples)`, dtype float32. If `out` was provided, returns that same buffer.

RAISES	DESCRIPTION
`RuntimeError`	If CuPy is not installed.
`ValueError`	If `precision` is not one of the supported values.
`TypeError`	If `sos` or `x` does not match the expected dtype.