Hacker News Reader

Is it possible to do a SIMD version?

I know compute shaders exist, but I have found that with huge datasets and a wide variety of end-user hardware, it tends to end up with many lines of code and brittleness. So I’m curious about high performance CPU-only implementations.

View on HN

You'd probably want to process multiple lines at once and use scatter to do the pixel drawing, I guess? But if you have one long line and a bunch of short ones, the perf would tank due to the lanes all dying, like gpu shader divergence.

Reply on news.ycombinator.com