rustnn exposes a C ABI in rustnn.h and a header-only C++17 wrapper in rustnn.hpp. Both APIs are experimental and can change without backward compatibility. Pin the release or Git commit used by your application.

Browse the C and C++ API reference for the generated declarations and source browser.

Build the reference

Install Doxygen with your system package manager and cbindgen with cargo install cbindgen --locked, then run from the repository root:

make docs-capi

Open target/doxygen/html/index.html to browse the C functions, structures and enums alongside the rustnn namespace and C++ classes. The target regenerates target/doxygen/include/rustnn/rustnn.h directly from src/capi.rs using cbindgen.toml before running Doxygen. It needs no backend SDK or library build. The generated header and HTML stay under the ignored target/ directory.

C documentation comes from Rust doc comments in src/capi.rs; C++ documentation comes from assets/capi/include/rustnn.hpp. Doxygen expands the C++ operation macros to include the named methods and their overloads.

The website build (make docs-build or make ci-docs, also mkdocs build directly) runs the same target and embeds the HTML at site/c-api/index.html. Serve site/ after building to browse both references together. The Doxygen pages have their own search index; MkDocs search covers the guide pages.

C++ usage

Include <rustnn/rustnn.hpp> and link against the installed rustnn library. The wrapper owns its handles, releases them in destructors, and supports moving them. Handles cannot be copied. Calls that return a failing C status throw rustnn::Error.

#include <rustnn/rustnn.hpp>
#include <iostream>
 
int main() {
  try {
    rustnn::MLContextOptions options;
    options.accelerated = false;
    rustnn::MLContext context(options);
    rustnn::MLGraphBuilder builder(context);
    rustnn::MLOperandDescriptor descriptor(rustnn::MLOperandDataType::Float32, {2});
    auto input = builder.input("input", descriptor);
    auto output = builder.relu(input);
    auto graph = builder.build({{"output", &output}});
 
    rustnn::MLTensorDescriptor input_descriptor(
        rustnn::MLOperandDataType::Float32, {2}, false, true);
    rustnn::MLTensorDescriptor output_descriptor(
        rustnn::MLOperandDataType::Float32, {2}, true, false);
    auto input_tensor = context.createTensor(input_descriptor);
    auto output_tensor = context.createTensor(output_descriptor);
    context.writeTensor<float>(input_tensor, {-1.0f, 2.0f});
    context.dispatch(graph, {{"input", &input_tensor}}, {{"output", &output_tensor}});
    auto values = context.readTensor<float>(output_tensor, 2);
    return values == std::vector<float>{0.0f, 2.0f} ? 0 : 1;
  } catch (const rustnn::Error &error) {
    std::cerr << "rustnn error: " << error.what() << '\n';
    return 1;
  }
}

Keep the context alive while using its builder, graphs and tensors. A default rustnn::MLGraphBuilder supports graph construction and text export without a runtime; compiling with build requires a builder created with a context. Building consumes the builder. Subsequent builder operations are invalid.

C ownership and errors

Include <rustnn/rustnn.h>. Check each RustnnStatus result against RustnnStatus_Success; rustnn_last_error_message() supplies thread-local failure details. Its returned pointer is borrowed and must not be freed.

Release every owned handle exactly once with its matching rustnn_*_destroy function. A successful rustnn_graph_builder_build consumes the builder and sets the supplied builder pointer to null. Release text returned by rustnn_graph_builder_webnn_text with rustnn_string_destroy.

Pointers must refer to live values of the declared type for the duration of each call. Arrays and byte buffers must contain at least the supplied number of elements or bytes. Tensor data is passed as raw bytes; use the descriptor's element type, shape and packing when sizing buffers. Host reads and writes require the tensor's readable and writable flags respectively.

For installation, CMake integration and complete C/C++ dispatch examples, see the C and C++ examples. The WebNN specification describes the graph operations; available operations and options follow the declarations in the headers.

Consuming rustnn from another CMake project

Point CMAKE_PREFIX_PATH at the cargo-c installation and import the package:

cmake_minimum_required(VERSION 3.16)
project(my_rustnn_app LANGUAGES C CXX)
 
find_package(rustnn CONFIG REQUIRED)
 
add_executable(my_rustnn_app main.cpp)
target_compile_features(my_rustnn_app PRIVATE cxx_std_17)
target_link_libraries(my_rustnn_app PRIVATE rustnn::rustnn)

Configure it with:

cmake -S . -B build \
  -DCMAKE_PREFIX_PATH=/absolute/path/to/rustnn-capi-install/
cmake --build build

You can create cargo-c build of RustNN and configure it with your preferred runtimes (e.g. with onnx-runtime and trtx-runtime)

cargo cbuild --features capi,trtx-runtime,onnx-runtime --prefix /absolute/path/to/rustnn-capi-install/

See https://github.com/lu-zero/cargo-c#usage for more detailed usage of cargo-c (e.g. –prefix/–libdir parameters). The prefix path (e.g. ~/.local) must exist.