XorSAR PRNG Hardware Implementation

Comments

• pmrobert Posts: 669

  2019-01-13 23:44

  If you search on "Random/LFSR on P2" you'll find a long thread of that name in the Propeller 2 forum. I understand very little of it but it sure looks like a long, detailed discussion on developing the PRNG that eventually ended up in the P@ silicon.
• xoroshironot Posts: 309

  2019-02-03 23:50

  @pmrobert Thanks for the reply... I have posted there before regarding my derivative work with xoroshiro (and may post there again).
  
  I was hoping to get some traction on my new original work, but need something more substantial to consider writing a paper.
  
  Currently I'm working on basic matrix equations to express the lowliest 4-bit state version, but have not spent much time on it.
  
  So far, it is harder for me than I first thought for such seemingly simple/fast code, which is part of why I imagine it may be worth the struggle.
  
  Regards,
  
  Chris
• evanh Posts: 15,171

  2019-02-04 02:34 edited 2019-02-04 02:35

  Hi Chris,
  Apologies for not making a comment earlier. I had a great time running the various bits of code for Xoroshiro testing and automating the use of testing suites, Practrand in particular. My maths skills aren't great though so probably can't give much help.
  
  Tony grok'd it better than I did.
  
  Looks good for parallelising over all. A little hard to follow the state flow with the "core" listed separately, imho. I'm mostly interested in easy hardware solutions or low-end microcontroller software so would avoid the ones with multipliers.
  
  The final xor'ing of a constant has a purpose I presume? Melissa tore shreds out of Xoshiro for heavily relying on multiplying a constant outside the iterative engine. I can see you aren't relying on it though.
  
  So that mean A, B and C are the constants of interest for tuning.
  
• xoroshironot Posts: 309

  2019-02-04 04:53 edited 2019-02-04 05:24

  Thanks for replying...
  
  The "core" is confusing at another point, as the 1D versions do not use the return value. I'll likely eliminate it, as it no longer serves its original purpose.
  
  The final xor constant (D) is unnecessary, if you consider a non-zero single missing output acceptable. Xoroshiro (all variants) is missing a single zero, which might be required for some uses. The end user could perform the xor themselves, if it were somehow required.
  
  Embedding the A constant plus or multiply in the iterative engine was certainly a conscious choice, but I found no way to fully digest it without the compare with -1. I have consoled myself to the idea that there is a certain obtuse elegance to it: Why would creating very-good statistical randomness be an artifact of 'clean-looking' code?
  
  Indeed, A, B and C are the only constants required for tuning, but just be careful that you don't accidentally discover X and Y while seeding (which I have done, and is equivalent to seeding xoroshiro with 0,0). I eventually started seeding 0,0 and checked state for 0,0 after one call and, if found, seed 0,1 instead.
  
  Discovering all ABC ++ and +* triplet constants up to 12-bit output/24-bit state (as I have already posted) is fairly trivial. It took a total of ~4.5 days on 16 cores/32 threads. Still working on 26-bit state... need significantly more processing power (~1000x) to discover even one triplet at 32-bit state (assuming they exist). I am hoping there is a better way to discover the triplets, but have yet to find it.
• evanh Posts: 15,171

  2019-02-04 11:43 edited 2019-02-04 11:48

  Discovering, as in finding triplet combinations that produce a full 2**N - 1 period I presume? N == number of state bits. We'll call these the full-period candidate set for a given N. Xoroshiro didn't need near as much computation to find all the full-period triplets because of far fewer possible combinations for a given N. I went up to N == 40 bits at least.
  
  Funnily, as I was busy finding each set, Tony came on board and made contact with one of the Xoroshiro authors, Seba. Seba was able to produce the full-period candidate set of any sized N at the drop of the hat. And with this set came a sort of ranking number for each candidate in the set. This helps priorities the subsequent quality testing.
  
  I've seen Melisa produce similar in her blog for completely unrelated algorithms, and Chris of Practrand fame also demonstrated it somewhere. So there is a non-brute-force mathematical solution. I just have no clue what it is though.
  
• xoroshironot Posts: 309

  2019-02-04 23:02 edited 2019-02-04 23:04

  Yes, 2**N-1 full-period candidates should be calculable, in theory. Perhaps this is trivial for some, but requires more effort for me. With my particular background in assembly language, I found it much easier to visualize simple code that might resolve many common PRNG issues, without constraining the algorithm to my (presumably) limited concept of solvability.
  
  My original hope was that I would be able to quickly find a few full-period 30 or 32-bit state candidates in order to more accurately compare against other similar state size PRNGs using SmallCrush, PractRand and gjrand... perhaps even finding a pattern in the constants themselves. This has been less fruitful than I originally thought, as I have only a few 26-bit state constants (++, though +* are actually ~2x faster to find).
  
  Currently, my preliminary analysis of 24-bit state candidates looks promising enough to move forward and continue work on an algorithm for calculating candidates. However, I also realized that the code (especially considering the ad-hoc rar function) might be more well-suited for hardware implementation.
  
  I went ahead and moved the "core" function into the individual PRNGs, for better readability.
  
  Thanks for the comments so far. I welcome any input on this, as I see some reasonable probability that this PRNG may prove useful, if taken to its logical conclusion.
• evanh Posts: 15,171

  2019-02-05 18:31 edited 2019-02-05 18:36

  Try doing full-period discovery without B, without the xor. This should make a large difference to the brute forcing speed and I get the feeling that operation is contributing very little to the randomness.
  
  EDIT: Maybe I'm wrong, it is in the iteration, any bit flipping is good there. I dunno.
  
• xoroshironot Posts: 309

  2019-02-05 19:40 edited 2019-02-05 21:19

  The B/xor is almost essential for full period above 16-bit state, as I have found no solutions with B=0 on a complete search of 18 to 24-bit state.
  
  My current plan is to attempt to write and solve equations describing the 4-bit state PRNG, which again, in theory should be possible. That being the case, scaling upward from there should be relatively easy.
  
  Currently, with one of the few candidates I have for 26-bit state (and likely not a best-case one), I am hard-failing SmallCrush on only 3 test statistics each when looking at upper 8 bits and lower 8 bits. As I recall for milestones, it should be possible to remove 2 of the 3 failures at about 30-bits (passing 15-bits x 2 at a time) and the final failure should disappear at 32-bits, if all goes well.
  
  BTW, as I recall, it was one of Tony's notes that gave me the idea to force a carry back through to the low bit(s). The hard part was to visualize how that could happen without breaking the equidistribution properties and invoking it through C++ (without resulting in overly complicated asm code).
• xoroshironot Posts: 309

  2019-02-07 19:08 edited 2019-02-07 19:10

  evanh wrote: »

  Try doing full-period discovery without B, without the xor. This should make a large difference to the brute forcing speed and I get the feeling that operation is contributing very little to the randomness...
  
  EDIT: Maybe I'm wrong, it is in the iteration, any bit flipping is good there. I dunno.

  That gave me an idea... perhaps I can move and convert 'B' to a rotation constant, something like: "tmp = ROL( s[0] + s[1], B ) + A;", but that may limit the possible candidate solutions... perhaps to none.
  
  I'll look into it anyway, as (if successful) it would bring much larger state sizes well within reach by my current (exhaustive search) method, hopefully without reducing statistical randomness.