Have you tried to go beyond 64 channels?

[HJLebbink]

Hi,

I'm considering a rewrite such that up to 512 channels can be used (on AVX-512F systems). I don't have time to maintain such port of SorterHunter. Would it make sense to commit that into your repo instead of my fork?

Have you tried before to go beyond 64 channels?

[HJLebbink]

I've updated (my fork) such that upto 256 channels can be used!

[55199789]

If it's not too much trouble, I would be very grateful if you could share your approach to 128/256-channel sorting networks. I understand that this does require significant computational efforts, but the sorting network (SN) I’ve developed so far still underperforms compared to bitonic sort

[HJLebbink]

@55199789 In my main branch on https://github.com/HJLebbink/SorterHunter, there's a file called htypes.h with a constexpr called NMAX. Originally, this was set to 64, and it used uint64_t to store the swap bits. I have added support for 128 and 256, which use slim::uint128_t and slim::uint256_t respectively. The thing is, the original code does not make it easy to switch between different types for what they call the SortWord_t. Worth noting though, once you get beyond a certain number of channels (around 48 or so), the genetic algorithm approach starts to lose its effectiveness...

[55199789]

Many thanks!