SIMD型を知る

A normal Int holds a single number, so math happens one value at a time. For AI you often need to crunch many numbers at once.

SIMD means Single Instruction, Multiple Data: one operation applied to a whole pack of numbers in a single step.

In Mojo the SIMD type holds several values of the same kind together, like a tiny fixed-size vector handled as one unit.

A SIMD type takes two parameters in brackets: the element DType and how many lanes it holds, like SIMD[DType.int32, 4].

The first parameter is a DType such as float32 or int32. It decides what each lane stores and how big it is.

The second parameter is the width: the number of values packed side by side. A width of 4 means four numbers travel together.

SIMD widths are powers of two like 1, 2, 4, 8, or 16. This maps cleanly onto the lanes your CPU offers.

Even a plain number is really a SIMD of width 1 in Mojo. Scalars and vectors share the same underlying type.

You can read a single lane by index just like a list. Lane 0 is the first value in the vector.

Passing one value fills all lanes with it. This is a quick way to make a SIMD vector of constant values.

Because the CPU acts on a whole pack in one instruction, SIMD does more work per cycle. That is the heart of vectorization.

You declare SIMD[DType.float32, 4]. What do those two parameters mean?

SIMD packs several same-typed values into one vector set by a DType and a power-of-two width, so one instruction touches them all. 🚀