Comment by jstimpfle

Yeah... There are libraries like Google Highway but that seems to me like a bulky software abstraction layer even if it is a so-called "zero cost" abstraction. Maybe appropriate in some cases. In my case, I've implemented a couple of lines for AVX2 and AVX512, maybe around 100 lines for the most perf sensitive sections currently, not a big deal. The actual work is in the understanding of the problem and of the phyiscal hardware, and in the software architecture and data layout to even enable vectorized stream processing in the first place. I'm not concerned writing a couple of perf sensitive lines against 2-3 most important concrete vector ISAs. I also fancy adding a GPU backend later (requires D3D12/SM6.0 API level), Highway couldn't help with that.