SIMD enhancements

[pfeatherstone]

GCC and Clang support Vector extensions using __attribute__((vector_size(N)). This is a platform independent SIMD extension.
We could enhance dlib's SIMD wrappers to use this stuff when available. I'm not sure if it does runtime dispatch or if the exact instruction is decided at compile time, but still worth an investigation. It could potentially unlock a whole bunch of other performance enhancements without too much work.

[pfeatherstone]

So I tried this. It work well. However, using simple loops or STL (std::transform and co.), and compiling with -O3 -March=native -ffast-math gives much better performance. It auto-vectorises the code better and fast-math is the biggest win. Interestingly fast-math has no effect on the manual instructions. So the moral of the story is: just use STL algorithm and aggressively turn optimisations on.

[pfeatherstone]

I guess you might have to use SIMD in non-trivial loops which the compiler cannot auto-vectorise. In which case this stuff could be cool. Reading the docs though it doesn't work for all instructions.

[davisking]

Yeah it's always a mixed bag. Never know what's really going to work best until you try and compare and benchmark 🤷‍♂️😁