oxicuda-signal

Part of the OxiCUDA ecosystem — Pure Rust CUDA replacement for the COOLJAPAN ecosystem.

Overview

oxicuda-signal is a GPU-accelerated signal, audio, and image processing library for the OxiCUDA ecosystem. It provides CPU reference implementations alongside PTX-generating GPU kernels for a wide range of DSP operations — from spectral transforms and wavelet decompositions to audio feature extraction and computer vision primitives. Kernels are emitted as WGSL/PTX strings at runtime and compiled via the CUDA JIT; no pre-compiled .ptx files are shipped.

Features

• Spectral transforms — DCT-II/III/IV, MDCT/IMDCT via the dct module; Haar, Daubechies db2–db10, Symlets, Biorthogonal, and Coiflet wavelets with multi-level DWT via the dwt module
• Audio processing — STFT, mel filterbank, MFCC, chroma features, and spectrogram variants (magnitude, power, mel)
• Filtering — FIR (windowed-sinc, raised-cosine), IIR (biquad SOS cascade), and Wiener filter
• Correlation — Autocorrelation, cross-correlation, GCC-PHAT time-delay estimation, circular convolution, PACF, Ljung-Box Q
• Image processing — Non-Maximum Suppression (greedy, soft, heatmap), morphological operations (erode, dilate, open, close, gradient), Gaussian blur, Sobel edge detection
• Window functions — Hann, Hamming, Blackman, Kaiser, flat-top and others with ENBW/CG metrics
• Zero-copy PTX — All GPU kernels generated as String at runtime for JIT compilation

Usage

Add to your Cargo.toml:

[dependencies]
oxicuda-signal = "0.1.5"

use oxicuda_signal::prelude::*;

// Compute MFCC features from a raw audio frame
let samples: Vec<f32> = vec![0.0; 16000];
let config = MfccConfig::default();
let mfcc_features = mfcc(&samples, config).unwrap();

// Multi-level DWT decomposition
let signal = vec![1.0_f32, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0];
let decomp = multilevel_forward(&signal, WaveletFamily::Haar, 2).unwrap();
let reconstructed = multilevel_inverse(&decomp, WaveletFamily::Haar).unwrap();

Status

v0.1.5 (2026-05-01) — 240 tests passing

License

Apache-2.0 — © 2026 COOLJAPAN OU (Team KitaSan)