Fijar la mejor configuración

Once the search names a winner, you commit to it. Locking in means specializing your kernel on that exact configuration for real use.

Set the winning numbers as compile-time aliases so the compiler builds one optimized version instead of searching again at runtime.

Call your kernel with the chosen parameters in the brackets. From now on it runs as that single, fully specialized variant.

The tuned program does zero searching while it runs. All that work happened at compile time, so each execution is pure fast kernel.

Confirm the specialized version still produces the right answers. A faster kernel is worthless if tuning changed its output.

Re-time the locked-in build against the baseline to prove the gain is real. The measured speedup is your evidence the tuning paid off.

The best values belong to the machine they were found on. Note the hardware so nobody assumes the same numbers win everywhere.

Moving to a different CPU or GPU? Run the search again there. Same source, fresh tuning, best settings for that target.

Store the chosen values and the workload they were tuned on. A clear record makes results reproducible and easy to revisit.

Tune on inputs that match production. Settings tuned to a toy workload can disappoint on the real sizes your app actually handles.

Expose knobs, search, then lock in the winner: that cycle turns one generic kernel into a machine-specific peak performer. 🏁

Pin down what locking in the best config means.

You baked the winning values in as aliases, verified correctness and the speedup, recorded the hardware, and learned to retune on new targets. 🎯