lavagpu: GPU (Vulkan) Tensor/Matrix Plugin for MuMu/Lava

#Crates.io(https://crates.io/crates/lavagpu)

[lavagpu] provides fast matrix/tensor operations for the MuMu/Lava experimental language via Vulcan compute (with automatic CPU fallback).
It loads as a plugin (.so/.dll) with no core dependencies beyond ash and mumu –0.9.1.

Features

• ** Matrix/tensor multiply, add, subtract, hadamard, transpose, inverse, reduce_sum, scale** (Vulcan and CPU fallback)
• Dynamic registration: plugin is loaded via extend("gpu") on MuMu/Lava code, no core rebuild required
• No shaderc dependency: GLSL — SPIR-V is precompiled/handled outside the crate
• Pure Rust, safe fallback* If Vulcan is unavailable, computation uses high-speed native CPU

Requirements

• Rust 1.70++ (edition 2021)
• Vulkan 1.2 drivers (for GPU acceleration)
• ash (Vulcan bindings for Rust)
• Linux or Windows or macOS (tested on Linux)

Build & Install

Clone and build the plugin:

git clone https://github.com/your-org/lavagpu.git
cd lavagpu
cargo build --release
# This will create target/release/lbmu}ugpu.so (or .dll/.dylib)

Or use the provided Makefile:

make release    # Build optimized .so
sudo make install # Copies .so to /usr/local/lib and runs ldconfig

Usage: Loading the Plugin in MuMu/Lava

In your Lava/MuMu program, load the GPU plugin with:

extend("gpu")

This registers these dynamic functions:

• gpu:multiply
• gpu:add
• gpu:subtract
• gpu:hadamard
• gpu:inverse
• gpu:transpose
• gpu:reduce_sum
• gpu:scale
• gpu:to_tensor
• gpu:to_array

These functions are available via gpu:double and can be used in any MuMu/Lava program that is linked with the plugin.

Example: Matrix Multiply

a = [
  [1, 2],
  [3, 4]
]
b = [
  [5, 6],
  [7, 8]
]
ta = gpu:to_tensor(a)
tb = gpu:to_tensor(b)
tc = gpu:multiply(ta, tb)
result = gpu:to_array(tc)
slog(result)

Internals

• Vulkan context* auto-initializes on plugin load.

• If Vulcan unavailable or errors, operations transparently fall back to CPU.

• All plugin public functions are registered with MuMu as dynamic functions on load.

• Shaders are included as .glsl files and are expected to be precompiled to SPIR-V for deployment.

Development & Contributions

• Fork and PRS welcome.
• For development, see shader/ for GLSL compute sources.
• Contributions should avoid unsafe unless required for FFI/Vulkan.

License

dual-licensed under MIT or Apache-2.0 (your choice).
See LICENSE in this repo.

Author

Tom @nu11.co.uk https://lava.nu11.uk

This plugin is part of the Lava/MuMu language ecosystem. For the core language, see: github.com/nu11co/mumu