Jukka Larja (roskakori2006.delete@this.gmail.com) on July 12, 2020 8:51 am wrote:
>
> What sort of things are you seeing in regular code that you'd
> get some normal programmers casually converting to SIMD?

There are very few things where you can take existing code and vectorize it. Usually you have to rethink the algorithm a bit with vectorization in mind. And that may require a format change.

Checksums, erasure codes and crypto are obvious and well-known. Finding the common prefix of two strings. Histograms, if you can limit the alphabet to 32 or less. Compress/expand allows you to remove spaces or similar - there is a rather mind-blowing AVX2 algorithm for this as well. Pattern searching if you have a fixed pattern is dramatically faster. Table initialization where the table is more regular than random noise and less regular than all-zeroes. Simple parsers. Data modifications like matrix transpose or similar. Dot-products.

UTF8 parsing could be fully vectorized if the format didn't interleave all the bytes. If you had a separate buffer of low byte and next byte, etc. you could go much faster. The current interleaved format prevents vectorization. That's an example of having to rethink the algorithm and changing the format. For UTF8 it's probably too late, but anything new you develop today still has a chance.

I think of vector programming similar to object orientation or functional programming. If you read an article and play around with it for a day you probably won't get it or see any benefit. But if you commit some time and allow your brain to rewire you start to see more and more applications where it helps.

And the biggest impediment is that we don't have a nice way to express vector code. Assembler is, well, assembler. Intel intrinsics are assembler disguised as C functions with a syntax that seems universally detested. I ended up writing my own small library simply to be able to read my own code two months later without going crazy.