Building

All builds run in Docker.

Configure

docker compose run --rm dev cmake -S . -B /tmp/feature_elm_build -G Ninja

Build

docker compose run --rm dev cmake --build /tmp/feature_elm_build

Common build options

Option	Default	Notes
CMAKE_BUILD_TYPE	Debug from Compose	Use Release for benchmarks
ENABLE_CUDA	ON	Turns CUDA backend support on when CUDA is found
BUILD_TESTING	ON via CTest	Enables test and benchmark targets
CMAKE_CUDA_ARCHITECTURES	sm_75 to sm_120 fat binary	The dev entrypoint uses native with a GPU and 75 without; set e.g. 89 for faster local builds
FEATURE_ELM_OPENMP	ON	Multithread the CPU hot loops with OpenMP. Results are bit-identical to a serial build; set OMP_NUM_THREADS to control the thread count
FEATURE_ELM_STATIC_CUDA	OFF	Link cuBLAS/cuSOLVER statically (used by the GPU demo image)
FEATURE_ELM_BUILD_CAPI	OFF	Build libfeature_elm_capi.so, the C API used by the ZeroGPU Space (turns on position-independent code)

Disable CUDA on CPU-only hosts

docker compose run --rm dev cmake -S . -B /tmp/feature_elm_build_cpu \
  -G Ninja \
  -DENABLE_CUDA=OFF \
  -DCMAKE_BUILD_TYPE=Debug

CUDA stubs keep the public API compileable when no toolkit is available.

Build outputs

• Libraries: /tmp/feature_elm_build/
• Test executables: /tmp/feature_elm_build/tests/
• Benchmark executables: /tmp/feature_elm_build/bench/

Use FEATURE_ELM_BUILD_DIR when a script needs a non-default build directory.