Skip to content

usage.md

Latest commit

 

History

History
82 lines (63 loc) · 3.41 KB

usage.md

File metadata and controls

82 lines (63 loc) · 3.41 KB
Raw

Detailed usage and background. See the README for the current candidate quickstart and support boundary.

onnxruntime-m

ONNX Runtime 1.25.1 C and C++ APIs through mcpplibs.mwrap modules. Upstream names remain unchanged: Ort::Env, Ort::Session, Ort::Value, and the C API retain their original namespaces and types.

[dependencies."mcpplibs.mwrap"]
onnxruntime = "1.25.1"

[target.x86_64-linux-gnu]
cxx_runtime = "toolchain-coupled"
import mcpplibs.mwrap.onnxruntime;
Ort::Env env(ORT_LOGGING_LEVEL_WARNING, "example");

The Linux official SDK itself loads libstdc++. GCC applications should use the toolchain-coupled C++ runtime shown above to share one implementation. The SDK provider supplies its declared runtime closure; no host library search is used.

The package depends on compat.onnxruntime, which supplies the matching official CPU SDK. Initial targets are Linux x86_64 with glibc, Windows x86_64 with the MSVC ABI, and macOS arm64. GPU execution providers, musl and MinGW require separate SDK builds and are not supplied by this version.

Quickstart template

mcpp new inference --template mcpplibs.mwrap.onnxruntime@1.25.1
cd inference
mcpp run

The template contains a reproducible ONNX Identity model and verifies real CPU inference. It needs no external model download or device.

import mcpplibs.mwrap.onnxruntime.macros; exposes the API version as mcpplibs::mwrap::onnxruntime::macros::ORT_API_VERSION. Original preprocessing macros, including ORT_TSTR, remain available through #include <mwrap/onnxruntime/macros.hpp> before imports. A preprocessing macro still expands inside a qualified identifier; use the mapping and raw-macro forms in separate contexts.

The root also exposes the CPU provider factory, plugin EP API declarations, Ort::Custom lightweight custom-op types and onnxruntime::ProviderOptions. API availability does not install an execution provider that is absent from the CPU SDK.

Configuration keys follow the SDK's separate-header grouping:

import mcpplibs.mwrap.onnxruntime;
import mcpplibs.mwrap.onnxruntime.config;
Ort::SessionOptions options;
options.AddConfigEntry(kOrtSessionOptionsConfigDisablePrepacking, "1");

The config module preserves global key names and pointer types. Upstream gives these header variables internal linkage, which GCC cannot export directly, so the module supplies equivalent inline constants. Use either this module or the raw config headers in a translation unit; combining both creates duplicate key names. Raw config headers can be mixed with the root module, as the tests verify.

Development

Use mcpp 2026.10.24.1 or newer. Tests exercise module-only inference, output shape and values, and header/module coexistence. tools/generate_model.py regenerates the checked-in model, and tools/generate_exports.py <sdk-include-directory> regenerates exports from Clang's declaration AST. SDK source pins and the public declaration inventory live under upstream/.

For an unpublished provider, use a local compat index during development, then run mcpp test --toolchain llvm@23.1.3 and mcpp test --toolchain gcc@16.1.0. python3 tools/test_template.py --index /path/to/mcpp-index runs the actual mcpp template renderer and builds/runs the generated project against the candidate.

The wrapper is MIT licensed. The SDK is distributed separately with Microsoft's MIT license and its ThirdPartyNotices.txt.