solvr

WebGPU computes in F32 only. The CUDA backend needs the CUDA toolchain that numr requires.

Quick Start

use numr::dtype::DType;
use numr::ops::RandomOps;
use numr::runtime::RuntimeClient;
use numr::runtime::cpu::{CpuClient, CpuDevice};
use solvr::signal::{ConvMode, ConvolutionAlgorithms};
use solvr::window::WindowFunctions;

fn main() -> solvr::Result<()> {
    let device = CpuDevice::new();
    let client = CpuClient::new(device.clone());

    // Create a Hann window
    let _window = client.hann_window(256, DType::F32, &device)?;

    // Convolve two random signals
    let signal = client.randn(&[256], DType::F32)?;
    let kernel = client.randn(&[32], DType::F32)?;
    let _result = client.convolve(&signal, &kernel, ConvMode::Same)?;

    Ok(())
}

The algorithm calls stay the same on any backend that implements the operation. Only the device and client types change.

Module Catalog

The full algorithm list for each module is in docs/modules.md.

The crate re-exports Tensor, Runtime, RuntimeClient, DType, Result, and Error from numr.

For SciPy Users

The Stack

• numr: tensors, FFT, linear algebra, and the Runtime abstraction.
• solvr: scientific and numerical algorithms built on numr.
• boostr: ML primitives built on numr.

Contributing

Contributions are welcome. See CONTRIBUTING.md for conventions and checks.

Questions and design discussion happen on Discord.

License

Licensed under Apache-2.0. See LICENSE.