Xyntra

Project Description

Xyntra is an automatic kernel-fusion compiler pass written in safe Rust.
It ingests ONNX / TorchScript graphs, pattern-matches common op-chains, and emits one fused GPU kernel through wgpu (cross-platform WGSL) or optional CUDA PTX.
The project explores graph rewriting, GPU occupancy modelling, and autotuned code-generation while keeping the entire pipeline 100 % unsafe-free.

---

Technologies & Dependencies

🦀 Core Technologies

• Rust 2024 Edition – type-safe IR with comprehensive error handling
• Safe Rust only – zero unsafe blocks in current implementation

📦 Current Dependencies

• Standard library only – no external crates yet
• Planned: egg (e-graphs), wgpu (GPU), clap (CLI)

---

Features & Roadmap

🔧 Core Infrastructure & Foundations

• Type-safe primitives – NodeId, TensorShape, OpKind, Graph in progress
• Error-enum with recoverable vs fatal classes in progress
• Config loader – CLI flags & fusion.toml in progress
• Modular crate layout – xyntra-core, xyntra-cli, xyntra-ir in progress

📡 Graph Ingestion & Export

• ONNX parser – load .onnx into internal IR
• TorchScript loader – parse .pt via tch-rs
• IR serialisation – export to DOT / JSON for debugging
• Fused-graph writer – emit reduced node graph snapshots

🧩 IR, Pattern Matching & Scheduling

• egg-based e-graph integration – rewrite rules & saturation loop
• Declarative fusion DSL – macro for matmul -> gelu -> dropout
• Scheduling heuristics – cost model for fusion candidates
• Fusion legality checker – shape, dtype, broadcast guards

⚡ Kernel Code Generation

• WGSL backend – emit compute shaders for wgpu
• CUDA PTX backend – optional NV path behind --backend ptx
• Shared-memory tiling – configurable tile/block sizes
• Vectorisation pass – vec4<f32> style loads/stores
• Mixed-precision support (FP16/BF16) (stretch)

🚀 Autotuning & Performance

• Parameter search harness – Bayesian optimiser over tile sizes
• GPU occupancy analysis – register & SM utilisation metrics
• Latency histogram – HDR log, p50/p95/p99 prints
• Flamegraphs – CPU-side hotspots with cargo flamegraph
• Roofline model script – FLOP/s vs bandwidth chart (stretch)

🔒 Correctness & Validation

• Golden unit tests – compare fused vs unfused outputs in progress
• Gradient checks – optional back-prop correctness suite
• Edge-case library – broadcast, dynamic shapes, odd strides in progress
• Numerical tolerance config – FP32 / FP16 epsilon thresholds

📊 Observability & Diagnostics

• Structured tracing spans – tracing crate with GPU timestamps
• --trace CLI flag – dump kernel timeline to JSON
• Occupancy dashboard – live CLI table of SM usage (stretch)

🛠️ Bench & Test Harness

• Micro-bench harness – single op-chain latency
• Model-zoo benchmarks – BERT, ResNet, ViT comparison
• Determinism suite – random seeds & output hashes in progress
• CI matrix – MSRV check, clippy, fmt, criterion

🧰 Developer eXperience (DX)

• cargo xtask or justfile – shortcuts (just fuse resnet.onnx)
• Pre-commit hook – cargo fmt && cargo clippy --fix
• make dev alias – spin-up CI-like environment locally

📦 Packaging & Release

• GitHub Release action – build macOS, Linux, Windows binaries
• Publish xyntra-core & xyntra-cli to crates.io
• SemVer policy & CHANGELOG.md generation
• Signed tags + GPG release checklist

🔗 Framework Plugins

• PyTorch 2 torch.xyntra.compile() drop-in backend
• ONNX Runtime execution-provider stub (libxyntra_ep.so)

📚 Docs & Examples

• Quick-start guide – clone → build → fuse tiny MLP
• Architecture diagram – ASCII / Mermaid / SVG
• Fusion logs demo – before/after latency screenshot

🗃️ Model-Zoo Benchmarks

• Scripted download + benchmark of BERT-Base, ResNet-50, ViT-Tiny, GPT-2
• Auto-generated result table in README via CI

🌐 Stretch Goals & Research Paths

• Horizontal fusion across attention blocks
• Dynamic-shape specialisation & cache
• Triton IR interoperability adapter
• WebAssembly demo – run fused WGSL in browser
• Meta-scheduler – ML-predicted tile sizes

🤝 DevOps & Community

• GitHub Actions pipeline – lint, test, benches, release
• Dual MIT / Apache-2 licence – broad adoption
• CONTRIBUTING.md, CODE_OF_CONDUCT.md, issue/PR templates
• GitHub Discussions – Q&A, roadmap voting
• Annotated blog series – graph rewriting, GPU tuning deep-dives

---