Game of GO benchmark

[ViralBShah]

Issue to track the Game of GO benchmark

https://groups.google.com/d/topic/julia-dev/8uIjpx-YTKw/discussion
Code from https://gist.github.com/3373404

[ViralBShah]

Added to the perf2 suite:

https://github.com/JuliaLang/julia/tree/master/test/perf2

[quinnj]

So I played around with this last night incorporating the pending @inbounds macro and using the profiler. My modified code runs in about 9.2s, which is 1.7x gcc -O0 and 4.8x gcc -O3. It's also more than 2x faster than the original Julia code.
One interesting note was the apparent slowness of mod, as shown in the profiling results below:

635 ...pbox/go.jl; ...dditional_liberty; line: 123
191 ...pbox/go.jl; ...dditional_liberty; line: 124

Where lines 123 and 124 correspond to:

ai = 1 + mod(pos - 1, board.size)
aj = 1 + fld(pos - 1, board.size)

I guess I wouldn't expect mod to be that much slower than fld (I ran the profiler a few times just to check it wasn't a sampling thing).

Gist of my modified code: https://gist.github.com/karbarcca/5815251

[StefanKarpinski]

It's deeply unfortunate that processors implement rem in hardware but not mod since mod is generally the better choice. This forces mod to be implemented in terms of rem and it ends up being significantly slower. Given the absurd excess of transistors modern CPUs have and the crazy number of instructions that the x86_64 architecture has, you would think they could add a frigging mod instruction.

[JeffBezanson]

Much faster after #4042.