SIMDとループの組み合わせ

To accelerate a kernel, swap the one-at-a-time loop for one that handles a whole pack of values per step with SIMD.

Choose how many elements fit in one vector. The width is the number of lanes each step processes together.

Grab several elements at once into a SIMD value with a vector load instead of reading them one by one.

Run the kernel's math on both chunks at once. The whole pack is added element-wise in a single operation.

Write the full result back with a vector store, covering every lane you just computed in one move.

The vectorized loop advances by the width, not by one. Each pass covers a full pack of elements.

Mojo's vectorize helper sweeps a closure across the range in vector steps so you skip the manual bookkeeping.

You define a small parameterized function that handles one chunk. Mojo calls it with the right width as it sweeps.

Call vectorize with your closure, the width, and the size. It loops and even handles the leftover tail for you.

When n is not a multiple of the width, a few elements remain. vectorize cleans up that tail so nothing is missed.

The vectorized kernel produces identical results but moves through data in big steps, so it finishes much sooner.

You vectorize a kernel with width 4 but n is 10. What handles the last two elements?

Vectorize a kernel by loading and storing packs, stepping by the width, and letting vectorize handle the sweep and the tail. 🚀