Goldy: GPU runtime for the Fondaco Machine

Goldy is a Rust GPU library that realizes the Fondaco Machine.

Maturity: Goldy 0.2 is the Fondaco Machine public API. SemVer applies within 0.2.x; expect breaking changes at 0.3.

The Fondaco model in Goldy

Goldy abstracts where bytes live (descriptor slots, residency, relocation) but exposes what access costs (access patterns, resize cost, readback path). See the design thesis.

Typed bindless shaders

Shaders use Slang with goldy_exp virtual entry points. Resources are typed parameters — Goldy resolves bindless slots automatically:

import goldy_exp;

[goldy_compute]
[numthreads(64, 1, 1)]
void cs_main(MyUniforms cfg, Scattered<uint> data, ThreadId id) {
    data[id.x] = data[id.x] + cfg.base;
}

Scheme

Record once, submit every frame. Goldy inserts barriers, parallelizes independent nodes, and aliases transient resources:

#![allow(unused)]
fn main() {
let mut scheme = Scheme::new(&ctx);
scheme
    .node("simulate", &sim_pipeline)
    .with_parcel(&particles, NodeAccess::ReadWrite)
    .dispatch(group_count, 1, 1);
let submission = scheme.submit()?;
}

Compute-to-surface

Compute shaders can write swapchain drawables directly — no graphics pipeline or raster pass:

#![allow(unused)]
fn main() {
let surface = SurfaceExchange::new(&ctx, &window, SurfaceConfig::default())?;
let mut scheme = Scheme::new(&ctx);
let (lease, present) = surface.bind_destination(&mut scheme)?;
scheme
    .node("render", &compute_pipeline)
    .with_parcel(&uniforms, NodeAccess::Read)
    .with_present(&lease)
    .dispatch(wg_x, wg_y, 1);

let mut submission = scheme.submit()?;
present.claim(&mut submission)?.consume()?;
}

Backends and bindings

CUDA and WebGPU backends are in progress. A Tenstorrent backend (Torus) is planned. See Backend Architecture.

Bindings: Python, .NET, C++, Rust FFI Client.

Quick links

• Installation
• Your First Triangle
• Your First Compute Shader
• Fondaco overview
• Goldy runtime mapping
• GitHub

License

MIT — see License.