تقليل حركة الذاكرة

Modern CPUs compute far faster than they can fetch data. Often a kernel waits on memory, not on the math itself.

Memory traffic is the total bytes your kernel reads and writes. Less traffic per result usually means a faster kernel.

Reading the same value many times wastes bandwidth. Load it once, do all the work, then reuse it from a register.

Two separate loops over the same array read it twice. Fusing them into one pass reads each element only once.

A value held in a CPU register needs no memory access. Reuse intermediate results instead of writing them out and back.

Extra temporary arrays add both stores and loads. Skip them when you can and compute straight into the final output.

Reading memory in order lets the CPU prefetch ahead. Jumping around defeats prefetching and stalls the loop.

Memory arrives in fixed-size cache lines. Using every byte of a line you fetched gives you free, already-loaded data.

Arithmetic intensity is work done per byte loaded. Raising it means each fetched value earns more compute before you move on.

Stores cost bandwidth too. Accumulate in a local and write the final result once rather than updating memory repeatedly.

When the kernel waits on data, cutting reads and writes is the biggest win, often beating clever arithmetic tweaks.

Your kernel reads the same array in two separate loops. What single change cuts its memory traffic most?

Cut memory traffic by touching data once, fusing loops, reusing registers, streaming in order, and writing results just once. 💾