Gompute

A Zig 0.16.0 library for defining a one-dimensional GPU/CPU kernel once and specializing the whole abstraction at compile time. CUDA (PTX), HIP (HSACO) and native CPU backends.

The backend is part of the type, and the CPU path compiles down to the loop you would have written by hand. That is checked on every build, not just claimed: zig build test compares the emitted assembly and fails on a difference.

📖 Documentation · Guide · Build integration · Reference · API

I am aware, as the zig library evolves that this library will need to be updated greatly. Hence, I'll attempt to maintain the interface, only ADDING features, rather than removing any features.

Install

zig fetch --save git+https://github.com/OmarSiwy/Gompute.git

Requires Zig 0.16.0 exactly. Linux and macOS; Windows and wasm are CPU-only and the GPU backends do not compile there. The CUDA/HIP backends need exe.root_module.linkSystemLibrary("c", .{}) — the drivers are dlopen'd.

Example

// src/kernels.zig
const g = @import("gompute");

pub const Params = struct { scale: f32 };

fn scaleRelu(x: f32, p: Params) f32 {
    const y = x * p.scale;
    return if (y > 0) y else 0;
}

pub const scale_relu = g.map("scale_relu", f32, Params, scaleRelu, .{});

comptime { g.exportKernels(@This()); }

// src/main.zig
var data = [_]f32{ -1, 2, -3, 4 };

var kernel = try g.Kernel(kernels.scale_relu, .cpu).init(0);
defer kernel.deinit();

try kernel.run(&data, .{ .scale = 2 });
// data == .{ 0, 4, 0, 8 }

Change .cpu to .cuda or .hip without touching the call site, or use AutoKernel to probe at run time. One emitKernels call in your build.zig wires up the device compilation — see the guide.

Commands

zig build test    # library tests, including the codegen comparison
zig build docs    # documentation site -> zig-out/docs
zig build docs -Dopen   # ...and open it in a browser

cd examples/basic     && zig build run   # CPU path
cd examples/exhaustive && zig build run  # GPU, needs libc + a driver

Documentation

The site under docs/ is the full documentation; it is published to GitHub Pages on every push to main, together with the generated API reference.

Page
Guide	Install, defining kernels, the build call, selecting a backend
Build integration	.auto pinning, multiple executables, kernel roots with their own imports, the artifact pipeline
Reference	Platform support, behaviour, the generated ABI, device math, backend status
Advanced	Allocation-free launches, fusion, vectorized CPU kernels, raw kernels, custom ABIs

License

MIT. See LICENSE.