oxionnx-directml

DirectML execution provider for OxiONNX — Windows D3D12 GPU backend.

This crate provides a feature-gated DirectML execution provider that dispatches selected ONNX operators to the Windows Direct3D 12 compute pipeline. On all other platforms it compiles as a transparent no-op: DirectMLContext::try_new() returns None and try_directml_dispatch() returns Ok(None), allowing the CPU path to handle all operations.

Architecture

oxionnx (root)
  └── oxionnx-directml  (feature = "directml")
       ├── context.rs   — D3D12 device + command queue init (Windows only)
       ├── dispatch.rs  — per-op routing to HLSL compute kernels
       └── kernels/
            ├── matmul.rs     — MatMul HLSL shader dispatch
            └── elementwise.rs — Add, Mul, Relu, Sigmoid HLSL kernels

Enabling the Feature

[dependencies]
oxionnx = { version = "0.1.2", features = ["directml"] }

Usage

use oxionnx::{Session, ExecutionProvider};
use oxionnx::execution_providers::DirectMLExecutionProvider;

// Build a session that prefers DirectML on Windows (falls back to CPU elsewhere)
let session = Session::builder()
    .with_execution_provider(DirectMLExecutionProvider::default().build())
    .load("model.onnx".as_ref())?;

Supported Operators (v0.1.2)

Operator	Status
MatMul	Scaffold (HLSL defined; kernel binding v0.1.6)
Add	Scaffold (HLSL defined; kernel binding v0.1.6)
Mul	Scaffold (HLSL defined; kernel binding v0.1.6)
Relu	Scaffold (HLSL defined; kernel binding v0.1.6)
Sigmoid	Scaffold (HLSL defined; kernel binding v0.1.6)

All unsupported operations fall back transparently to the CPU execution path.

Platform Notes

• Windows: DirectMLContext::try_new() initializes a D3D12 device and returns Some(ctx).
• macOS / Linux / WASM: DirectMLContext::try_new() always returns None; the entire crate compiles as a no-op with zero overhead.
• The windows crate dependency is target-gated: [target.'cfg(target_os = "windows")'.dependencies].

Part of oxionnx

A Pure Rust ONNX inference engine.

License

Apache-2.0