emu

The old version of Emu (which used macros) is here.

Overview

Emu is a GPGPU library for Rust with a focus on portability, modularity, and performance.

It's a CUDA-esque compute-specific abstraction over WebGPU providing specific functionality to make WebGPU feel more like CUDA. Here's a quick run-down of highlight features...

• Emu can run anywhere - Emu uses WebGPU to support DirectX, Metal, Vulkan (and also OpenGL and browser eventually) as compile targets. This allows Emu to run on pretty much any user interface including desktop, mobile, and browser. By moving heavy computations to the user's device, you can reduce system latency and improve privacy.

• Emu makes compute easier - Emu makes WebGPU feel like CUDA. It does this by providing...

  • DeviceBox<T> as a wrapper for data that lives on the GPU (thereby ensuring type-safe data movement)
  • DevicePool as a no-config auto-managed pool of devices (similar to CUDA)
  • trait Cache - a no-setup-required LRU cache of JITed compute kernels.

• Emu is transparent - Emu is a fully transparent abstraction. This means, at any point, you can decide to remove the abstraction and work directly with WebGPU constructs with zero overhead. For example, if you want to mix Emu with WebGPU-based graphics, you can do that with zero overhead. You can also swap out the JIT compiler artifact cache with your own cache, manage the device pool if you wish, and define your own compile-to-SPIR-V compiler that interops with Emu.

• Emu is asynchronous - Emu is fully asynchronous. Most API calls will be non-blocking and can be synchronized by calls to DeviceBox::get when data is read back from device.

An example

Here's a quick example of Emu. You can find more in emu_core/examples and most recent documentation here.

First, we just import a bunch of stuff

use emu_glsl::*;
use emu_core::prelude::*;
use zerocopy::*;

We can define types of structures so that they can be safely serialized and deserialized to/from the GPU.

#[repr(C)]
#[derive(AsBytes, FromBytes, Copy, Clone, Default, Debug)]
struct Rectangle {
    x: u32,
    y: u32,
    w: i32,
    h: i32,
}

For this example, we make this entire function async but in reality you will only want small blocks of code to be async (like a bunch of asynchronous memory transfers and computation) and these blocks will be sent off to an executor to execute. You definitely don't want to do something like this where you are blocking (by doing an entire compilation step) in your async code.

…

And last but certainly not least, we use an executor to execute.

fn main() {
    futures::executor::block_on(do_some_stuff()).expect("failed to do stuff on GPU");
}

Built with Emu

Emu is relatively new but has already been used for GPU acceleration in a variety of projects.

• Used in toil for GPU-accelerated linear algebra
• Used in ipl3hasher for hash collision finding
• Used in bigbang for simulating gravitational acceleration (used older version of Emu)

Getting started

The latest stable version is on Crates.io. To start using Emu, simply add the following line to your Cargo.toml.

[dependencies]
emu_core = "0.1.1"

To understand how to start using Emu, check out the docs. If you have any questions, please ask in the Discord.

Contributing

Feedback, discussion, PRs would all very much be appreciated. Some relatively high-priority, non-API-breaking things that have yet to be implemented are the following in rough order of priority.

• Enusre that WebGPU polling is done correctly in `DeviceBox::get
• Add support for WGLSL as input, use Naga for shader compilation
• Add WASM support in Cargo.toml
• Add benchmarks`
• Reuse staging buffers between different DeviceBoxes
• Maybe use uniforms for DeviceBox<T> when T is small (maybe)

If you are interested in any of these or anything else, please don't hesitate to open an issue on GitHub or discuss more on Discord.