1st call to rand(1.0) generates predictable values < 1e-4 for seeds in 0..<26844, unlike C++, python, D, ...

[timotheecour]

Example

this shows that whatever (small) seed is picked, the 1st call to rand() gives a predictably small value < 1e-4, ie, skews the distribution.

import std/random
proc main()=
  let n = 100_000
  for i in 1..<n:
    randomize(i)
    let x = rand(1.0)
    doAssert x < 1e-4, $(x, i) # this passes for i in 0..<26844, but should not
main()

This is a bug, other languages don't have this issue:

nim: fails

by fail, I mean it produces skewed results
(see example above)

python: works

import random
for i in range(0,10):
  random.seed(a=i)
  print(random.random())

prints:

0.8444218515250481
0.13436424411240122
0.9560342718892494
0.23796462709189137
0.23604808973743452
0.6229016948897019
0.793340083761663
0.32383276483316237
0.2267058593810488
0.46300735781502145

D: works

import std.random,std.conv,std.stdio;
void main(){
  //auto seed = Clock.currTime().toUnixTime;
  for(int i=0;i<10;i++) {
    auto rnd = Random(i.to!int);
    uniform(0.0, 1.0, rnd).writeln;
  }
}

prints:

0.548814
0.417022
0.435995
0.550798
0.96703
0.221993
0.89286
0.0763083
0.873429
0.0103742

C++ works

#include <random>
#include <iostream>

void fn2() {
  // std::random_device rd;
  std::uniform_real_distribution<double> dist(0.0, 1.0);
  for (int i=0; i<10; ++i){
    std::mt19937 mt(i);
    std::cout << dist(mt) << "\n";
  }
}

int main() {
  // fn1();
  fn2();
  return 0;
}

prints:

0.592845
0.997185
0.185082
0.0707249
0.900621
0.0551801
0.947476
0.227339
0.0111144
0.364461

Additional Information

1.5.1 d6a1602

links

https://stackoverflow.com/questions/19665818/generate-random-numbers-using-c11-random-library
https://channel9.msdn.com/Events/GoingNative/2013/rand-Considered-Harmful

this is also related to an issue I had previously raised: #8589