dda-rs

Pure Rust Delay Differential Analysis (DDA) engine and the ddalab command-line tool. The same crate runs in the browser as WebAssembly in the DDALAB web application.

All analysis runs through the Rust engine directly; nothing shells out to a native DDA binary.

Command-line tool

Install it as a Python tool (the wheel carries the ddalab executable only) or download an archive and its .sha256 file from GitHub Releases:

uv tool install ddalab
ddalab inspect --file recording.edf
ddalab run --file recording.edf --variants ST --wl 1000 --ws 200 --delays 7 10 --bundle out

It reads EDF, BDF, BrainVision and delimited text. ddalab --help lists the other commands (batch, variants, validate, samples, info).

Features

• Pure Rust execution for DDA variants and CCD extensions
• Typed request/response API with serde
• Built-in variant metadata and SELECT-mask helpers
• Configurable solver backend:
  • RobustSvd for the default numerically stable path
  • NativeCompatSvd for the native-compatible SVD path
• Optional batched CUDA acceleration for ST, CT, CD, DE, and SY

Rust library

[dependencies]
dda-rs = "2.1.0"

Quick Start

use dda_rs::{
    run_request_on_matrix, AlgorithmSelection, DDARequest, DelayParameters, PreprocessingOptions,
    TimeRange, WindowParameters,
};

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let request = DDARequest {
        file_path: "synthetic.csv".to_string(),
        channels: Some(vec![0, 1, 2]),
        time_range: TimeRange {
            start: 0.0,
            end: 511.0,
        },
        preprocessing_options: PreprocessingOptions {
            highpass: None,
            lowpass: None,
        },
        algorithm_selection: AlgorithmSelection {
            enabled_variants: vec!["ST".to_string(), "CD".to_string()],
            select_mask: None,
        },
        window_parameters: WindowParameters {
            window_length: 128,
            window_step: 64,
            ct_window_length: None,
            ct_window_step: None,
        },
        delay_parameters: DelayParameters { delays: vec![1, 2] },
        ct_channel_pairs: None,
        cd_channel_pairs: None,
        model_parameters: None,
        model_terms: None,
        variant_configs: None,
        sampling_rate: None,
    };

    let samples = (0..512)
        .map(|t| {
            let x = t as f64 * 0.03;
            vec![x.sin(), (1.7 * x).cos(), (0.5 * x).sin()]
        })
        .collect::<Vec<_>>();

    let result = run_request_on_matrix(&request, &samples, None)?;
    println!("analysis id: {}", result.id);
    Ok(())
}

Solver Backends

The Rust engine exposes two SVD backends through PureRustOptions::svd_backend:

• SvdBackend::RobustSvd
  • default
  • uses nalgebra SVD
  • preferred for real analysis
• SvdBackend::NativeCompatSvd
  • uses the native-compatible SVD kernel ported into Rust
  • useful when comparing against historical native-binary behavior

CUDA

CPU execution remains the default. Build the optional CUDA backend and select a device with:

cargo run --release --features cuda -- run --file data.csv --channels 0 1 --device cuda

--device cuda:N selects device N. The CUDA build dynamically loads the NVIDIA driver and cuBLAS. Compilation does not require a local CUDA toolkit; execution requires a working NVIDIA CUDA installation. The implementation batches same-shaped f64 regressions and computes predictions and residual norms on the GPU. Rank-deficient regressions fall back to the CPU SVD solver. Rust-only CCD extensions continue to use their existing CPU implementations.

Library users configure the same backend through PureRustOptions:

use dda_rs::{ComputeDevice, PureRustOptions, PureRustRunner};

let runner = PureRustRunner::new(PureRustOptions {
    compute_device: ComputeDevice::Cuda(0),
    ..PureRustOptions::default()
});

Modules

• engine: core DDA and CCD execution
• types: request/response structures
• variants: variant metadata and SELECT-mask utilities
• network_motifs: motif analysis helpers
• profiling: profiling helpers
• error: error types

License

MIT