FFT — `NVMathNet.Fft`

Memory layout

Unlike the BLAS-style modules, FFT shapes are row-major (cuFFT convention): in a shape such as [batch, n] the last axis is contiguous in memory. axes selects the dimensions to transform; they must be ascending, contiguous and include the first or last dimension. Transforming a leading axis (for example shape [8, 3], axes [0]) uses a strided layout. For R2C the shape is the real input; for C2R it is the real output, and the half-spectrum has n/2 + 1 bins along the last transformed axis.

Enums

enum FftDirection { Forward = -1, Inverse = 1 }
enum FftType { C2C, R2C, C2R }
enum LastAxisParity { Even, Odd }
enum FftNormalization { None, Backward, Forward, Orthonormal }

`FftOptions`

Property	Type	Default	Description
`FftType`	`FftType?`	`null` (auto)	Transform type.
`InPlace`	`bool`	`false`	Execute in-place.
`LastAxisParity`	`LastAxisParity`	`Even`	Parity of the last axis for R2C/C2R.
`DeviceId`	`int`	`0`	GPU device index.
`Blocking`	`bool`	`false`	Synchronise after execution.
`Normalization`	`FftNormalization`	`None`	Output scaling: `None` (cuFFT behaviour), `Backward` (1/N on inverse), `Forward` (1/N on forward), `Orthonormal` (1/√N on both). R2C counts as forward and C2R as inverse.

`Fft` (class)

Stateful cuFFT plan wrapper. Implements IAsyncDisposable, IDisposable.

Member	Description
`Fft(long[] shape, int[]? axes = null, FftOptions? options = null, CudaStream? stream = null, bool doublePrecision = false)`	Creates a plan descriptor. `axes` selects which dimensions to transform (`null` = all); they must be ascending, contiguous and include the first or last dimension (negative values allowed). Batched trailing- and leading-axis layouts are supported.
`Task PlanAsync(CancellationToken ct = default)`	Builds the cuFFT plan.
`void Plan()`	Synchronous plan creation.
`Task ExecuteAsync(FftDirection direction = Forward, CancellationToken ct = default)`	Executes the transform.
`void Execute(FftDirection direction = Forward)`	Synchronous execution.
`void ResetOperand(nint inputPtr, nint outputPtr)`	Changes data pointers without replanning.
`Task SynchronizeAsync(CancellationToken ct = default)`	Synchronises the stream.

Static Helpers

Method	Description
`static Task<DeviceBuffer<Complex>> FftAsync(...)`	One-shot forward C2C FFT. Returns new output buffer.
`static Task<DeviceBuffer<Complex>> IFftAsync(...)`	One-shot inverse C2C FFT.
`static Task<DeviceBuffer<Complex>> RFftAsync(DeviceBuffer<double> input, long[] shape, ...)`	Double-precision real-to-complex FFT. Output length = `n/2+1` along the last axis.
`static Task<DeviceBuffer<float>> RFftAsync(DeviceBuffer<float> input, long[] shape, ...)`	Single-precision real-to-complex FFT. .NET has no single-precision complex type, so the result is interleaved `(re, im)` floats (`2 * (n/2+1)` per row).
`static Task<DeviceBuffer<double>> IRFftAsync(DeviceBuffer<Complex> input, long[] outputShape, ...)`	Double-precision complex-to-real inverse FFT.
`static Task<DeviceBuffer<float>> IRFftAsync(DeviceBuffer<float> input, long[] outputShape, ...)`	Single-precision complex-to-real inverse FFT; input is interleaved `(re, im)` floats.

Table of Contents

FFT — NVMathNet.Fft