Struct FftRequest 

pub struct FftRequest<T: FftSupported, I = u8, O = u8> {
    pub plan_key: PlanKey,
    pub direction: FftDirection,
    pub input: GpuRef<I>,
    pub output: GpuRef<O>,
    pub reply: Sender<Result<(), GpuError>>,
    /* private fields */
}

Typed cuFFT request — the canonical Phase-1 entry point.

Type parameters

• T is the scalar dtype of the transform (f32 for the float lane, f64 for the double lane). Used for T::KIND reporting on the FftDispatch trait.
• I is the input buffer’s element type (default u8 for the raw-byte Path B). Phase 1.5++ adds typed Path A by using e.g. I = f32 (R2C input), I = C32 (C2R / C2C input).
• O is the output buffer’s element type (default u8). Path A uses O = C32 (R2C / C2C output) or O = f32 (C2R output).

Plan resolution is performed by the actor on receipt of the FftMsg::Exec message — the request only carries a PlanKey. Repeat calls with the same key hit the LRU cache on the actor.

In-place transforms: output may alias input (cuFFT supports this when shapes line up — only meaningful when I = O).

Fields

plan_key: PlanKey

direction: FftDirection

input: GpuRef<I>

output: GpuRef<O>

reply: Sender<Result<(), GpuError>>

Implementations

impl<T: FftSupported, I, O> FftRequest<T, I, O>

pub fn new( plan_key: PlanKey, direction: FftDirection, input: GpuRef<I>, output: GpuRef<O>, reply: Sender<Result<(), GpuError>>, ) -> Self

Construct a request. For Path B, I = O = u8 and the buffers are byte-cast on the caller side. For Path A, I and O are the per-side element types (e.g. f32/C32 for R2C).

Trait Implementations

impl<T, I, O> FftDispatch for FftRequest<T, I, O>

where T: FftSupported, I: Send + Sync + 'static, O: Send + Sync + 'static,

fn dtype_kind(&self) -> DType

fn plan_key(&self) -> PlanKey

fn dispatch(self: Box<Self>, ctx: &FftDispatchCtx<'_>)