oxicuda-runtime

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

Overview

oxicuda-runtime is a Pure Rust implementation of the CUDA Runtime API (libcudart) surface, built on top of oxicuda-driver's dynamic driver loader. It exposes ergonomic Rust types for streams, events, device pointers, and kernel dimensions, covering the full breadth of commonly used cudaXxx functions — without any CUDA SDK build-time dependency.

Features

• Device management — cudaGetDeviceCount, cudaSetDevice, cudaGetDeviceProperties, cudaDeviceSynchronize, cudaDeviceReset
• Memory — cudaMalloc, cudaFree, cudaMallocHost, cudaMallocManaged, cudaMallocPitch, cudaMemcpy, cudaMemcpyAsync, cudaMemset, cudaMemGetInfo
• Streams — Create/destroy/synchronize CUDA streams with priority and flag support
• Events — cudaEventCreate, cudaEventRecord, cudaEventSynchronize, cudaEventElapsedTime
• Kernel launch — cudaLaunchKernel, cudaFuncGetAttributes, PTX module loading/unloading
• Peer access — Multi-GPU peer memory copies and access enable/disable
• Profiler — cudaProfilerStart/cudaProfilerStop with a ProfilerGuard RAII wrapper
• Texture & surface objects — Full CudaTextureObject, CudaSurfaceObject, array descriptors

Usage

Add to your Cargo.toml:

[dependencies]
oxicuda-runtime = "0.1.3"

use oxicuda_runtime::{device, memory, stream, event};

// Select device 0
device::set_device(0)?;

// Allocate 1 MiB of device memory and zero it
let d_buf = memory::malloc(1 << 20)?;
memory::memset(d_buf, 0, 1 << 20)?;

// Create a stream and record an event
let s = stream::stream_create()?;
let e = event::event_create()?;
event::event_record(e, s)?;
event::event_synchronize(e)?;

// Cleanup
event::event_destroy(e)?;
stream::stream_destroy(s)?;
memory::free(d_buf)?;

Status

License

Apache-2.0 — © 2026 COOLJAPAN OU (Team KitaSan)