gnuradio/volk ? reverse-engineered prompt

Build me a C++ library for portable SIMD math kernels, something like VOLK, where the same function can use different optimized implementations depending on the CPU at runtime.

I want it to support common x86 and ARM targets, with a clean fallback path when no special instructions are available. Include a simple way to detect the best kernel on the current machine, plus a small command line tool that profiles the system and picks the fastest option. Make it build with CMake, install cleanly, and come with a solid test suite so I can verify the kernels work across platforms.

Please keep the design practical and easy to extend, so new kernels can be added without touching everything else. If you need to check current docs for SIMD or CPU feature detection details, go ahead and look them up online. Also include basic documentation showing how to build, run tests, install, and profile the system.