Struct RealFftPlan 

pub struct RealFftPlan<R: Runtime> { /* private fields */ }

Reusable exact-length F32 FFT plan. The direction and N are fixed; batches may change, in which case the batched workspace is replaced, not grown without bound. Caller-owned output bindings must be non-overlapping. Raw bindings carry no dtype/device tag: pass F32 storage from this client.

Implementations

impl<R: Runtime> RealFftPlan<R>

pub fn new( client: ComputeClient<R>, n: usize, mode: FftMode, ) -> Result<Self, LaunchError>

pub fn len(&self) -> usize

pub fn is_empty(&self) -> bool

pub fn convolution_len(&self) -> Option<usize>

pub fn retained_bytes(&self) -> usize

Bytes retained by this plan, excluding output tensors, driver metadata, compiled kernels and temporary initialization scratch.

pub fn clear_workspace(&mut self)

Release scratch; chirp tables remain reusable. Runtime stream ordering handles deferred release. This is not a device-wide synchronization.

pub fn forward( &mut self, signal: TensorBinding<R>, real: TensorBinding<R>, imag: TensorBinding<R>, dim: usize, used: usize, ) -> Result<(), LaunchError>

Unnormalized N-point RFFT. Samples at used..N are zero, without constructing a padded input tensor. used may be zero.

pub fn inverse( &mut self, real: TensorBinding<R>, imag: TensorBinding<R>, signal: TensorBinding<R>, dim: usize, used: usize, ) -> Result<(), LaunchError>

N-point IRFFT normalized by 1/N. DC, and Nyquist for even N, are real-only. Bins at used..N/2+1 are treated as zero.