oxicuda-levelzero

Part of the OxiCUDA ecosystem — Pure Rust CUDA replacement for the COOLJAPAN ecosystem.

Overview

oxicuda-levelzero provides a LevelZeroBackend that implements the ComputeBackend trait from oxicuda-backend using Intel's oneAPI Level Zero API. It targets Intel Arc, Xe, and integrated GPUs on Linux and Windows by loading libze_loader.so / ze_loader.dll at runtime via libloading, with no C SDK dependency at compile time.

Features

LevelZeroBackend implementing the full ComputeBackend trait (GEMM, conv2D, attention, reduce, unary/binary ops, memory management)
Runtime driver loading via libloading — zero link-time dependency on the Intel GPU driver
SPIR-V kernel support via the spirv module for cross-vendor shader compilation
Intel Xe Matrix Extensions (XMX) SPIR-V generators via the spirv_xmx module — cooperative-matrix GEMM for Xe/Arc/Ponte Vecchio GPUs
Sub-group optimized kernels via the spirv_subgroup module — warp-level reduction, scan, and GEMM using GroupNonUniform opcodes
Neural-network SPIR-V kernels via the spirv_nn module — Conv2D and scaled dot-product attention
Multi-tile / multi-device dispatch via the multi_tile module — distributes workloads across Intel Max series GPU sub-device tiles
Device enumeration and selection via the device module
Memory management with unified shared memory (USM) via the memory module
Compiles on all platforms; returns UnsupportedPlatform on macOS so cross-platform workspaces are unaffected

Platform Support

Platform	Status
Linux (Intel GPU)	Full support via `libze_loader.so`
Windows (Intel GPU)	Full support via `ze_loader.dll`
macOS	Compile-only; returns `UnsupportedPlatform` at runtime

Usage

Add to your Cargo.toml:

[dependencies]
oxicuda-levelzero = "0.1.5"

use oxicuda_levelzero::LevelZeroBackend;
use oxicuda_backend::ComputeBackend;

let mut backend = LevelZeroBackend::new();
backend.init()?;

let ptr = backend.alloc(1024)?;
// ... copy data, run kernels ...
backend.free(ptr)?;

Status

Version: 0.1.5 (2026-05-01)
Tests: 104 passing