A First Look at SIMD

Every instruction covered since assembly1-1 — across all three courses — has operated on one piece of data at a time. This chapter previews something genuinely different: instructions that operate on several pieces of data at once. Deliberately light-touch, on purpose — this is honestly a topic large enough to be its own course.

Why Vector Instructions Exist — SIMD in One Idea

SIMD — Single Instruction, Multiple Data. Every ordinary instruction this entire arc has covered works on one value: one register, one memory operand. Adding two 4-element arrays together the ordinary way takes four separate ADD instructions, one pair at a time. A SIMD instruction packs several values into one wide register and performs the same operation on all of them simultaneously, inside a single instruction's own execution — not four instructions doing one thing each, but one instruction doing four things at once.

The Real Lineage — MMX → SSE → AVX

MMX's own register-sharing compromise meant a program couldn't freely mix MMX instructions with ordinary floating-point math without real overhead switching between the two uses of the same physical registers. SSE fixed this by giving vector instructions their own dedicated register file entirely. This whole lineage is one more concrete instance of assembly2-1's own "four decades of accretion" theme — SIMD has its own internal history of layered extensions, each one adding capability on top of the last, mirroring assembly2-2's own RAX/EAX/AX nesting, just for an entirely different feature.

One Worked Example

Adding two 4-element arrays of single-precision floats, the SIMD way:

MOVUPS XMM0, [array1]   ; load 4 floats from array1 into XMM0
MOVUPS XMM1, [array2]   ; load 4 floats from array2 into XMM1
ADDPS  XMM0, XMM1       ; add all 4 pairs simultaneously — ONE instruction
MOVUPS [result], XMM0   ; store all 4 results back at once

ADDPS — ADD Packed Single-precision — the mnemonic itself describes exactly what it does. The equivalent ordinary, scalar version of this same task would need four separate ADD-family instructions, one value pair at a time; ADDPS does all four in one.

Honestly Scoping This Topic

This is genuinely as far as this course goes. A real, complete treatment of SIMD would need to cover the many packed integer/float data-type variants, each instruction set's own feature-detection requirements, real memory-alignment rules (some SIMD instructions fault on unaligned memory operands), and AVX-512's own further masking and broadcasting capabilities. All of that is large enough to be a genuinely separate course on its own — this chapter deliberately stops here rather than pretending a shallow tour is complete coverage.

Hands-On Exercises