> How is it faster than std: It uses the fact that popcount of a uniformly distributed integer follows a binomial distribution. By rescaling that binomial distribution and adding some linear fill we can achieve a curve very similar to a proper normal distribution in just a few instructions. While that level of precision is not suitable for a general use, in instances where quality is not particularly important (gamedev, fuzzing) this is perhaps the fastest possible way of generating normally distributed floats.
> How is it faster than std: It uses the fact that popcount of a uniformly distributed integer follows a binomial distribution. By rescaling that binomial distribution and adding some linear fill we can achieve a curve very similar to a proper normal distribution in just a few instructions. While that level of precision is not suitable for a general use, in instances where quality is not particularly important (gamedev, fuzzing) this is perhaps the fastest possible way of generating normally distributed floats.