These examples compile against installed packages and exercise the public interface. scalar.cc checks imports, policy choice and mathematical special values; environment.cc checks admission and complete thread-state restoration; vectors.cc checks typed memory, masks, owning swizzles, register arrays and wide math for both m32 and h32.

Configure with the same LLVM 23/C++26 toolchain as the installed packages:

cmake -S tests/api -B build/api -G Ninja \
  -Dftz_DIR="/path/to/ftz/lib/cmake/ftz" \
  -Dnative_DIR="/path/to/native/lib/cmake/native" \
  -DAPI_PROFILES=AVX2
cmake --build build/api --parallel
ctest --test-dir build/api --output-on-failure

Select only architectures admitted on the test host: AVX2, AVX512 or NEON. A semicolon-separated list builds each requested vector example. The installed CMake targets supply the module providers and numerical compile settings.

shader.hlsl is a complete HLSL 2021 compute entry point. It uses one compile-time shader policy, unlike the two coexisting C++ types. Dispatch only records backed by the input/output buffers (64 records per workgroup in this small example). Compile it with the include directories exported by ftz::hlsl, entry shader, target cs_6_6, and FTZ_FP32_HARDWARE_FTZ=0 or 1. Hardware shader arithmetic requires separate device admission. Compiling an example does not establish it.

Doxygen uses EXAMPLE_PATH=tests/api and the named snippet markers in these files. They check API usage; the validation guide describes the separate numerical and device checks.

api.native_types checks FTZ's scalar and vector mask traits, default one-lane architecture and rejection of foreign-family architecture tags through the installed module graph.