Parametrizando bloco e largura

To let autotuning vary a value, it has to be a compile-time parameter, not a plain runtime variable. Parameters can be specialized per candidate.

A tile size controls how big a block of work you process at once. Make it a parameter and the tuner can try 16, 32, 64, and more.

SIMD width sets how many numbers you crunch per instruction. Different chips favor different widths, so expose it as a parameter too.

Put tunable values inside the function's bracket list. Each entry becomes a knob the autotuner is allowed to change.

Inside the body, treat the parameter like a constant. The compiler bakes its value in for each specialized version it builds.

Pick candidate values that make hardware sense: powers of two like 8, 16, 32 for width, and cache-friendly blocks for tile size.

Each knob should mean one clear thing. Tangled parameters make the search confusing and the results hard to explain.

Use a constraint to reject impossible combinations early, so the tuner never wastes time building candidates that cannot work.

Give each parameter a reasonable default. Code stays usable without tuning, and the search starts from a known-good baseline.

Every extra parameter multiplies the candidates to test. Expose the few knobs that truly matter, not every number you can find.

With tile and width as parameters, your kernel is ready: the autotuner has clear dials to turn while searching for peak throughput. 🔧

Confirm what makes a value tunable by the autotuner.

You exposed tile and width as compile-time parameters, added a constraint and defaults, and gave the autotuner clean dials to turn. 🎯