| > I know that SIMD doesn't imply particular instruction set but a pattern of concurrent data transformation. Eh. It's both. You don't have one without the other. SIMD is pretty much entirely extensions to base processor ISA; there might be a few architectures that have SIMD as a basic part of them (GPUs are one of them, taken to an extreme order), but for the most part you have to know which instruction set you're working with. SIMD is basically "pack multiple data points into a single CPU register, then to another, and perform some operation on them as if you ran that op on all of the pairwise data points individually". Some ops are binary (arithmetic, bitwise, etc) and some are unary. Some have "gates" whereby you can do a comparison, the boolean output of which is stored in a bit packed integer. Then you can run conditional instructions after that that only perform the instruction if the bit in that variable is set, creating "constant time" SIMD instructions with what amounts to branching. Really depends on the instruction set's capabilities. AVX512 is far and away one of the most extensive extensions, with a ton of super niche instructions meant for enterprise number crunching. It has weird stuff, like swapping bytes, collating them, doing all sorts of weird manipulations. But the number of x86 CPUs that support that instruction set are small. You can't emit code that has e.g. AVX512 and just run it on a CPU that doesn't have that extension. You get a CPU exception and it crashes the process (or, if this is in kernel/driver land, your machine). So they're kind of tied together. Anyway if you want to see a list of them, Intel's SIMD intrinsics site has always been really nice to browse IMO. https://www.intel.com/content/www/us/en/docs/intrinsics-guid... |