webnn-onnx-utils

Shared ONNX/WebNN conversion utilities used by:

• rust-webnn-graph (WebNN → ONNX)
• webnn-wg (ONNX → WebNN)

This crate is intentionally small and modular. It provides:

• Data type mapping - Bidirectional conversion between WebNN and ONNX data types (Float32, Float16, Int32, Int64, etc.)
• Operation name mapping - 90+ operation mappings between WebNN and ONNX (matmul↔MatMul, conv2d↔Conv, etc.)
• Attribute parsing/building - Type-safe attribute handling for ONNX NodeProto (int, float, string, arrays)
• Tensor data handling - Conversion between ONNX TensorProto and typed data (with all data type support)
• Identifier sanitization - WebNN DSL-compatible identifier generation
• Shape inference - Comprehensive shape inference for common operations (matmul, transpose, reduce, concat, etc.)

Features

Data Types

All common ONNX and WebNN data types are supported:

• Float32, Float16
• Int8, Int32, Int64
• Uint8, Uint32, Uint64

Operations

90+ operations mapped including:

• Matrix operations (MatMul, Gemm)
• Convolutions (Conv, ConvTranspose)
• Pooling (AveragePool, MaxPool, GlobalAveragePool)
• Activations (Relu, Sigmoid, Tanh, Softmax, Gelu)
• Elementwise (Add, Sub, Mul, Div)
• Reductions (ReduceSum, ReduceMean, ReduceMax, etc.)
• Tensor manipulation (Concat, Reshape, Transpose, Squeeze, Unsqueeze)
• And many more...

Shape Inference

Shape inference context with support for:

• Unary and binary operations with broadcasting
• Matrix multiplication (2D and batched)
• Transpose with permutation
• Reduction operations (with keepdims support)
• Concat along any axis
• Reshape (with -1 inference)
• Squeeze and Unsqueeze
• Dynamic dimensions with overrides

Usage

use webnn_onnx_utils::{
    data_types::{DataType, webnn_to_onnx, onnx_to_webnn},
    operation_names::mapper,
    attributes::{AttrBuilder, AttrParser},
    tensor_data::TensorData,
    shape_inference::{ShapeInferenceContext, TensorShape},
};

// Data type conversion
let onnx_type = webnn_to_onnx(DataType::Float32);
let webnn_type = onnx_to_webnn(1).unwrap(); // 1 = FLOAT

// Operation mapping
let m = mapper();
assert_eq!(m.webnn_to_onnx("matmul"), Some("MatMul"));
assert_eq!(m.onnx_to_webnn("MatMul"), Some("matmul"));

// Attribute building
let attrs = AttrBuilder::new()
    .add_int("axis", 1)
    .add_float("epsilon", 1e-5)
    .add_ints("perm", vec![0, 2, 1, 3])
    .build();

// Tensor data
let data = TensorData::scalar(DataType::Float32, 3.14);
let proto = data.to_tensor_proto("weight".to_string(), DataType::Float32, vec![1]);

// Shape inference
let mut ctx = ShapeInferenceContext::new();
ctx.set_shape("a".to_string(), TensorShape::from_static(vec![2, 3, 4]));
ctx.set_shape("b".to_string(), TensorShape::from_static(vec![2, 4, 5]));
let result = ctx.infer_matmul("a", "b").unwrap();
// result shape: [2, 3, 5]

Testing

Comprehensive test coverage with 57+ tests:

cargo test

All tests pass covering:

• Data type round-trip conversions
• Operation name mappings (case-insensitive)
• Attribute parsing and building (all types)
• Tensor data operations
• Shape inference for all supported operations
• Broadcasting rules
• Identifier sanitization

Protobufs

ONNX protobufs are compiled with prost-build from protos/onnx/onnx.proto3.

License

Apache-2.0