gpu_runtime

A C++ library that detects available (GP)GPU backends and loads them in runtime. It also provides Device information for detected GPUs and their backends.

Building

# After clone, fetch the third-party header submodules:
git submodule update --init --recursive

cmake -B build -DBUILD_SAMPLE_APP=1 -DCMAKE_BUILD_TYPE=Release
cmake --build build
./build/app/gpu_runtime_query/gpu_runtime_query

Third-party headers

Three GPGPU API header sets are pulled in as git submodules under external/. None of the vendor runtimes is required at build time - the libraries are dlopen'd at runtime.

Submodule	Upstream	License
external/OpenCL-Headers	KhronosGroup/OpenCL-Headers	Apache-2.0
external/Vulkan-Headers	KhronosGroup/Vulkan-Headers	Apache-2.0 (a few files MIT)
external/level-zero	oneapi-src/level-zero	MIT (a few files Apache-2.0)

Each submodule ships its own LICENSE. To pin a different upstream revision, cd into the submodule, git checkout <ref>, then commit the submodule SHA update in this repository.

Not vendored

• NVIDIA CUDA headers: proprietary license, redistribution not permitted. The CUDA backend (src/backends/cuda.cpp) declares the small Driver-API subset it uses inline.
• AMD HIP / ROCm headers: MIT-licensed, but hip_runtime_api.h is 10 K+ lines and transitively pulls in nvidia_detail/ and amd_detail/ paths that require full CUDA / ROCm SDK installations. The HIP backend (src/backends/rocm.cpp) declares the small Runtime-API subset it uses inline.
• Apple Metal: <Metal/Metal.h> is part of the macOS system SDK and is #imported directly when building on Apple platforms.