On modern hardware, RAM latency is often the main reason why CPU-bound code is slow. The fix is often changing RAM layout, i.e. modifying data structures. And when the code uses SIMD to process the payload portion of these structures, alignment requirements cause them to become really custom pretty soon.
> that’s still a very small amount of the overall code in a given project
That’s not universal across all software. Lately, I’m working in the area of CAD/CAM/CAE. Before that I worked on videogames, HPC, multimedia. In many of these projects, data structures were not a small amount, even relatively speaking.
On modern hardware, RAM latency is often the main reason why CPU-bound code is slow. The fix is often changing RAM layout, i.e. modifying data structures. And when the code uses SIMD to process the payload portion of these structures, alignment requirements cause them to become really custom pretty soon.
> that’s still a very small amount of the overall code in a given project
That’s not universal across all software. Lately, I’m working in the area of CAD/CAM/CAE. Before that I worked on videogames, HPC, multimedia. In many of these projects, data structures were not a small amount, even relatively speaking.