Secret seeds and primes

[tansy]

First a question - as far as I understand these secret seeds are meant to be primes. Why is that, or where can I read about that?

Now to the point.
You said that secret seeds are more balanced. What you mean by that?

Also checking 2^32 division reminders is rather ineffective. checking 2^64 - 59 took me, literally 99 seconds.
There is this algorithm - Miller-Rabin primality test, that checks it in 1.3 s, and all four wyhash secret's seeds in 1.03 ms, yes millisecond. There is, not only one, example of it here (last C version).

[wangyi-fudan]

The secrets should be prime. because in wyrand, we add the secret to the seed. a prime secret ensure that the period is 2^64. For other part of wyhash library, it is a "feeling to be perfect". And non-prime secret seem to fail smhasher.
A balanced secret meant all its 8 bytes have popcount==4. The early version of secret are popcount==32 for the whole seed which is not fully balanced.
Thank you for your Miller-Rabin primality test. I am looking for the code. However, I need some time to port it to wyhash.
usually, we generate secret offline and hard code it to the .h file. A de novo make_secret is indeed time consuming. That's why someone propose a PR to remove the prime-test. And I didn't notice this change. Now, the prime-test is back as normal and intended.

[wangyi-fudan]

Miller-Rabin test is done by copying :-)

[wangyi-fudan]

The cleaned code for Miller Rabin test is available here https://github.com/wangyi-fudan/MillerRabin64

[tansy]

However, I need some time to port it to wyhash.
usually, we generate secret offline and hard code it to the .h file

It could be in separate tool/file. I already proposed it with hash map (#120). I don't get why would it have to be in hash itself. If you just mentioned in the hash that secret was generated with function moved to make-secret.c it would be fine and everyone wanting to dig it further would find the tool and used it.

May I propose PR #140 that does that.