krrishnarraj/clpeak ? reverse-engineered prompt

Build me this GPU and CPU benchmark app from the repo, with a clean command line tool that can measure peak compute, memory bandwidth, matrix throughput, and basic latency tests across the supported backends. I want it to work on whatever hardware is available, like OpenCL, Vulkan, CUDA, ROCm, Metal, oneAPI, and CPU, and to print clear results that make it easy to compare devices or spot regressions from a saved baseline.

Make the build work with CMake and keep the backend detection automatic, so it only uses the SDKs that are actually installed. Also make sure the common flags behave the same no matter which backend is running, and that I can select specific tests or specific devices when needed.

If there are any missing docs or platform details, look up the current official docs online before guessing.