oxicuda-fft

GPU-accelerated Fast Fourier Transform operations -- pure Rust cuFFT equivalent.

Part of the OxiCUDA project.

Overview

oxicuda-fft provides a complete GPU-accelerated FFT library implemented entirely in Rust, targeting feature parity with NVIDIA's cuFFT. It generates PTX kernels at runtime using the Stockham auto-sort algorithm, which is optimal for GPU execution because it avoids bit-reversal permutations entirely.

The crate supports 1-D, 2-D, and 3-D transforms across complex-to-complex (C2C), real-to-complex (R2C), and complex-to-real (C2R) modes, with both in-place and out-of-place execution. For sizes up to 4096, a single-kernel strategy uses shared memory for maximum throughput. Larger transforms employ a multi-stage ping-pong approach with explicit transpose kernels.

Arbitrary-size FFTs are handled via the Bluestein/Chirp-Z algorithm, which reduces any size N to a power-of-two convolution. Mixed-radix support (radix-3, 5, 7) is also available for composite sizes.

Modules

Supported Transforms

• C2C -- Complex-to-complex, forward and inverse, in-place or out-of-place
• R2C -- Real-to-complex with Hermitian-symmetric output (N/2+1 complex values)
• C2R -- Complex-to-real inverse transform
• 2-D FFT -- Row-wise FFT + transpose + column-wise FFT
• 3-D FFT -- Extension of 2-D to volumetric data
• Batched FFT -- Multiple independent transforms in a single launch

Quick Start

use oxicuda_fft::prelude::*;

// Create a 1-D complex-to-complex plan for 1024 elements
let plan = FftPlan::new_1d(1024, FftType::C2C, 1).expect("plan creation failed");

// Create a 2-D plan for a 256x256 grid
let plan_2d = FftPlan::new_2d(256, 256, FftType::C2C).expect("2d plan");

// With a GPU context:
// let handle = FftHandle::new(&ctx)?;
// handle.execute(&plan, input, output, FftDirection::Forward)?;

Feature Flags

Status

License

Apache-2.0 -- (C) 2026 COOLJAPAN OU (Team KitaSan)