oxicuda-graph 0.4.1

OxiCUDA Graph — CUDA Graph execution engine with operator fusion, buffer lifetime analysis, stream partitioning, and optimized execution planning
Documentation

oxicuda-graph

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

Overview

oxicuda-graph is a high-level CUDA Graph execution engine that models GPU workloads as directed acyclic graphs (DAGs) of operations. It applies analysis and optimisation passes — operator fusion, buffer lifetime colouring, and multi-stream partitioning — and lowers the result to an oxicuda_driver::graph::Graph for low-overhead CUDA Graph submission.

Features

  • Ergonomic GraphBuilder API for constructing kernel, memcpy, and memset nodes with explicit buffer dependencies
  • Analysis passes: topological ordering (ASAP/ALAP scheduling, slack, levels), buffer liveness intervals, and a Lengauer–Tarjan dominator tree
  • Optimisation passes: element-wise kernel fusion, interval-graph buffer colouring (minimises peak allocation), and multi-stream partitioning
  • ExecutionPlan lowering to ordered PlanStep sequences ready for SequentialExecutor (CPU simulation) or CudaGraphExecutor
  • Execution statistics: kernels launched, nodes visited, fused ops count
  • #![forbid(unsafe_code)] — fully safe Rust

Usage

Add to your Cargo.toml:

[dependencies]
oxicuda-graph = "0.4.1"

use oxicuda_graph::builder::GraphBuilder;
use oxicuda_graph::executor::{ExecutionPlan, SequentialExecutor};
use oxicuda_graph::node::MemcpyDir;

let mut b = GraphBuilder::new();
let buf_in  = b.alloc_buffer("input",  4096);
let buf_out = b.alloc_buffer("output", 4096);
let upload  = b.add_memcpy("upload", MemcpyDir::HostToDevice, 4096);
let kernel  = b.add_kernel("relu", 4, 256, 0)
               .fusible(true).inputs([buf_in]).outputs([buf_out]).finish();
let download = b.add_memcpy("download", MemcpyDir::DeviceToHost, 4096);
b.chain(&[upload, kernel, download]);

let graph = b.build().unwrap();
let plan  = ExecutionPlan::build(&graph, 4).unwrap();
let stats = SequentialExecutor::new(&plan).run().unwrap();
println!("kernels launched: {}", stats.kernels_launched);

Status

v0.3.0 (2026-06-25) — 299 tests passing

License

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