Lightweight mock that stands in for a real GPU buffer during tests.

reason: enum variants represent all possible solver strategies; not all are constructed in every build configuration

An aligned buffer that guarantees proper alignment for SIMD operations. This is a wrapper around a raw allocation that ensures the data is aligned to

Bluestein (Chirp-Z) solver for arbitrary sizes. Bluestein's algorithm uses the identity: nk = (n² + k² - (k-n)²) / 2 This allows rewriting the DFT a

Buffered solver for improved cache locality. Wraps another FFT solver and adds buffering for strided access patterns.

Convert complex format to half-complex format. This is the inverse operation of `Hc2cSolver::execute`.

Complex-to-Real FFT solver. For N/2+1 complex inputs satisfying conjugate symmetry, produces N real outputs. This is the inverse of R2C (without norm

Cache-oblivious FFT solver. Uses the four-step FFT algorithm with recursive decomposition. Decomposes N = N1 × N2 and performs: 1. N2 column-FFTs of

Builder for programmatic codelet generation without proc-macros. # Example ```no_run use oxifft_codegen_impl::CodeletBuilder; let ts = CodeletBuilde

Trait for compile-time FFT operations. Implementations are provided for power-of-2 sizes up to 1024.

Marker struct for const FFT implementations.

Constant folder that applies algebraic simplifications to fixpoint. This wraps [`StrengthReducer`] and applies it repeatedly until the expression no

Cooley-Tukey FFT solver. Implements the Cooley-Tukey algorithm with radix-2, radix-4, radix-8, and split-radix variants: - DIT (Decimation-in-Time):

CUDA FFT plan wrapper holding real oxicuda resources. Note: `CudaFftPlan` is NOT `Clone` because `FftHandle` contains a non-cloneable `Stream`.

Chirp-Z Transform plan. Precomputes all chirp sequences and the FFT of the convolution kernel so that repeated calls to [`execute`](Self::execute) ru

DFT plan implementation.

Differentiable FFT plan for automatic differentiation. Wraps a standard FFT plan and provides differentiation capabilities.

Direct DFT solver using O(n²) algorithm. This solver computes the DFT directly from the definition: ```text X[k] = Σ_{n=0}^{N-1} x[n] * e^(-2πink/N)

Extension trait adding FFT methods to ndarray arrays of [`Complex<T>`]. Implemented for: - `ArrayBase<S, Ix1>` — 1D arrays (N-point DFT) - `ArrayBase

Trait for floating-point types supported by OxiFFT. This trait combines numeric operations with trigonometric functions needed for FFT computation.

Fractional Fourier Transform plan. Precomputes chirp factors for efficient repeated transforms.

Generic mixed-radix recursive solver. This solver factorizes N and applies the Cooley-Tukey decomposition recursively. For prime factors, it falls ba

Internal cache holding type-separated twiddle tables.

Trait for executing N independent FFTs of the same size as a batch. The batch is represented as a slice of input-slice references and a corresponding

Thread-safe, LRU-evicting pool of GPU buffers. Acquire a buffer via [`acquire`][GpuBufferPool::acquire], use it, then return it with [`release`][GpuB

High-level GPU FFT interface. Provides a simple API for GPU-accelerated FFT with automatic backend selection and memory management.

Trait for GPU FFT implementations.

Guru interface for maximum flexibility. Allows specifying arbitrary dimensions, strides, and batch parameters. This is the most general FFT interface

Half-Complex to Complex solver. Converts FFTW half-complex format to standard interleaved complex format.

Half-Complex to Half-Complex solver. Operates on data in FFTW half-complex format: - hc\[0\] = Re(X\[0\]) (DC) - hc\[1\], hc\[2\] = Re(X\[1\]), Im(X\

Indirect solver for non-contiguous memory layouts. Wraps a contiguous FFT solver and handles gather/scatter for strided data. This is used internally

Local recursive CSE for RDFT expressions. Counts subexpression usages across all outputs, then rewrites shared subexpressions (used ≥ 2 times) as `Te

Parsed arguments from `gen_dispatcher_codelet!(size = 4, ty = f32)`.

Parsed arguments from `gen_multi_transform_codelet!(size=4, v=8, isa=avx2, ty=f32)`.