While working on the Make It Random library for Unity, I took up an engineering challenge: generation of random floating point numbers between zero and one quickly and with perfect uniform distribution. The most common method, dividing a random integer by the full range of possible integers, has a few substantial flaws that I hoped to avoid:

- First, it involves a division operation, which could naturally slow things down a bit.
- Secondly, although the distribution on a high level is uniform, it’s more clumpy for higher numbers closer to one, and more diffusely spread for lower numbers closer to zero, so the uniformity is imperfect.
- Third, the quantity of possible integers generated rarely maps cleanly to the possible number of floating point values available, and so different floating point values are likely to have at least slightly different probabilities of occurring, further weakening the uniformity.
- And fourth, even when the mapping from integers to floating point values is carefully managed, if the range of values needed is not a power of two, this adds further problems to any attempt to get perfect uniformity without sacrificing performance, due to how typical random engines generate data as chunks of bits.

With the help of various sources around the internet plus some clever use of probability mathematics, I was able to conquer all of these difficulties, providing perfectly uniform and fast generation of floating point numbers in the unit range. This includes all four variants of whether the lower and upper bounds are inclusive or exclusive. The techniques involved are explained below.

The first two are not at all new, but are included because I have not seen them discussed as often as I think they deserve, and because they help build the context for the third technique. This last technique is something I devised on my own, and although I have no doubt other smart people have already discovered it or something similar, I never ran across it while researching, so I’m eager to share it with others in this post. Plus, I suspect that the general technique can be usefully applied to other random value generation beyond just floating point numbers, so the more people are aware of it, the better.

## Contents

- Clarification on Perfect Uniformity
- The Setup
- Step 1: Integers to Floating Point Without Division
- Step 2: Open Range and the Rejection Method
- Step 3: Closed Range and Fancy Probability
- High Precision Alternative Distribution
- Using the Precise Method to Gain an Extra Bit
- Performance Measurements
- Full C++ Source for Benchmark

## Clarification on Perfect Uniformity

Before I proceed, let me clarify a bit what I mean by the second problem listed above, the issue of non-uniform clumping. Given the way floating point numbers are most commonly represented (according to the IEEE 754 Standard for Floating-Point Arithmetic), there isn’t a nice clean uniformity of values between zero and one. Instead, due to their exponential nature, the density of unique values is higher for ranges closer to zero. This image roughly exemplifies what is happening:

In this hypothetical format that behaves similarly to IEEE 754 floating point numbers, the overall range is divided into sub-ranges, starting with [0.5, 1) on the right, and reducing to half the size of each prior sub-range as the sub-ranges proceed to the left toward zero. So after the first is [0.25, 0.5), followed by [0.125, 0.25), and so on. This reduction goes on until the exponent cannot get any smaller, when the whole process is finally terminated by the value 0 (ignoring the minor detail of subnormal numbers). And yet, despite the ranges getting smaller and smaller, every single one of them contains the exact same quantity of unique values; eight of them, in the example shown above. Or equivalently, the density keeps getting higher, starting at 16 values per unit range, up to 32, then 64, doubling each time until the capacity of the exponent is exhausted.

When using division to generate a random floating point number between zero and one, this is the sort of distribution that is produced, and it is the result of the non-uniform clumpiness that I mentioned earlier. The reason is that before division, the random integers generated can and should be perfectly uniform. Each possible integer in the target range has an equal probability of being generated as any other. But the division operation maps those integers to unique floating point values in such a way that each unique floating point value will have non-uniform number of integers mapped to it.

Consider if we generate 16-bit integers, in the range [0, 2^{16} = 65536), and map those to the range [0, 1) using the low-precision floating point format shown above. The upper half of the 65536 integers, 32768 of them, will map to the 8 unique values between 0.5 and 1, or 4096 possible integers mapped to each of the 8 final values, resulting in a 4096/65536 chance (6.25%) that any one of these values will be generated. In the second range from the right, [0.25, 0.5), there are only 16384 possible integers, but they still map to 8 unique floating point values. Each one is mapped from 2048 possible integers and has a 2048/65536 chance (3.125%) of being generated. Following the same pattern, each value in the range [0.125, 0.25) has a 1024/65536 chance (1.5625%), followed by 512/65536 (0.78125%) and so on.

The shading in the above images indicates this probability. Many different integers map to the same unique values in the upper ranges, and so those values each have a higher probability of occurring. Very few integers map to same values in the lower ranges, and they each have a lower probability of occurring. The combination of non-uniform density with non-uniform probability allows the overall effect to be consistent uniformity. Low density combined with high probability in the upper ranges perfectly matches the high density and low probability in the lower ranges.

This tends to be good enough for many purposes, but there could be some subtle biases inadvertently introduced into a system because of the heavier clustering of larger numbers and finer distribution of smaller numbers. Instead, I want a distribution of floating point numbers that looks more like the following:

There are trade-offs with this model. The benefit is obviously that uniformity is far more consistent, and applies not just at a high level, but also in the details. Every single unique value that can be produced has an exactly equal probability of being generated as any other. The downside is that there are far fewer unique values possible, and a corresponding loss of absolute precision in some sub-ranges. For example, it is impossible for 0.03125 to be generated by the second model, whereas it is entirely possible in the first. But the first model has a nice symmetry that the second does not: It is also impossible for 0.96875 ( = 1.0 – 0.03125) to be generated in either model, which could be taken as a form of consistency in the second model, but feels sort of inconsistent in the first.

You may sometimes explicitly want the distribution of the first model, maximizing the nuance of numbers near zero, and there is at least one way other than using a division operation to generate such a distribution, so I will come back around to briefly cover that technique at the end. However, the majority of this article is aimed at generating the perfectly uniform model, so let’s dive in!

## The Setup

First of all, let me describe the foundation from which I’m starting. I’ll presume we already have a pseudo-random number generator capable of generating high-quality uniformly distributed random bits. This might be something like the Mersenne Twister, or perhaps a more recent PCG or XorShift random engine. You could even use a cryptographically secure engine, though that would defeat the performance objectives of this post. Regardless of the specific engine chosen, these types of engines all tend to produce integers in the range [0, 2^{n}), for some integer n. Just be careful to avoid engines that produce integers in a range whose size is not a power of two, such as a multiply-with-carry engine with 2^{32} – 1 chosen as its base.

I’m going to further suppose we have two functions built on an engine of choice, `rand32()`

and `rand64()`

, which return a `uint32_t`

and `uint64_t`

respectively. (I’ll be working in C/C++ land for this post.) These two functions fully saturate their number types with 32 or 64 fully uniform bits. Some of my implementations below will only use some but not all of those bits, but others will use all 32 or 64 bits to maximize performance, so it is presumed that all bits are available for such use.

One last bit of machinery, I’ll presume two functions for performing type punning from integers to floats and doubles, `as_float(uint32_t)`

and `as_double(uint64_t)`

. Depending on your language, and your willingness to violate certain principles of the language and complicate things for your compiler, you might wish to be selective about the method used to perform this process. Here’s an example implementation in C++ using a union; it technically violates the C++ rules and results in the dreaded “undefined behavior”, but it is my understanding that it will nonetheless work reliably on nearly all compilers and platforms.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 |
float as_float(uint32_t i) { union { uint32_t i; float f; } pun = { i }; return pun.f; } double as_double(uint64_t i) { union { uint64_t i; double f; } pun = { i }; return pun.f; } |

As the product of this post, I’ll write eight functions making use of `rand32()`

and `rand64()`

. The first four will return 32-bit floats, and I’ll name them `rand_float_oo()`

, `rand_float_oc()`

, `rand_float_co()`

, and `rand_float_cc()`

. The ‘o’ and ‘c’ suffixes stand for “open” and “closed”, referring to whether the lower and upper bounds of the range are exclusive (open) or inclusive (closed). I’ll likewise implement four more for 64-bit doubles.

## Step 1: Integers to Floating Point Without Division

The first functions I will implement are `rand_float_co()`

and `rand_double_co()`

, as they only makes use of the first technique that I will explain. These functions generate a floating point number between 0 and 1, *including* the possibility of 0, but *excluding* the possibility of 1.

The above diagrams show the bit layout of the single and double precision formats of IEEE 754. The left bit is the sign bit (0 for positive, 1 for negative). Following that are the exponent and mantissa, which roughly follow the formula (1 + M) × 2^{E}, where E is a signed integer, and M is a fixed point number in the range [0, 1). In other words, the mantissa describes the *details* of the number, while the exponent indicates the *magnitude*. There are a number of useful ways that these binary formats can be manipulated, either directly as raw bits or interpreted as integers.

It just so happens that for the IEEE 754 binary formats, all floating point numbers from 1 up to but not including 2 have the same constant exponent. That’s exactly the size of range we’re looking for, just offset by 1. And conveniently, if you consider all the 2^{23} or 2^{52} possible bit patterns of the mantissa when paired with this constant exponent, they perfectly map to all the possible floating point values possible between 1 and 2, with perfect uniformity and identical “clumpiness”. By the latter, I mean that the difference between any given number in this range and the next larger representable number is always exactly the same amount.

So to generate a random number, all we have to do is set the sign bit to 0, the exponent to the appropriate constant, and fill the mantissa with random bits. (We can use a shift rather than a mask to get the needed bits in the right place; this is because some random engines tend to have higher quality randomness in the high bits compared to the low bit or bits.) And then to convert that to the desired range [0, 1), we can simply interpret the constructed value as a floating point number in the range [1, 2) and subtract 1 using standard floating point arithmetic.

1 2 3 4 5 6 7 8 9 |
float rand_float_co() { return as_float(0x3F800000U | (rand32() >> 9)) - 1.0f; } double rand_double_co() { return as_double(0x3FF0000000000000ULL | (rand64() >> 12)) - 1.0; } |

Conveniently, the range (0, 1] which does not include 0 but does include 1 is nearly as easy to generate. The only difference is that instead of subtracting 1, we’ll subtract *from* 2. Because the inner floating point generation can generate the value 1, then 2 – 1 = 1, meaning that 1 is a possible value generated by the final computation. But the inner generation can produce floating point numbers close to but not including 2. So, for example, 2 – 1.9999999 = 0.0000001 is really close to 0, but an actual value of 0 is not possible.

1 2 3 4 5 6 7 8 9 |
float rand_float_oc() { return 2.0f - as_float(0x3F800000U | (rand32() >> 9)); } double rand_double_oc() { return 2.0 - as_double(0x3FF0000000000000ULL | (rand64() >> 12)); } |

## Step 2: Open Range and the Rejection Method

Next let us consider the open range (0, 1), from which both 0 and 1 are excluded. In the half-open ranges above, the number of possible values was an integer power of 2, 2^{23} for floats and 2^{52} for doubles. This is why we could get away with simply grabbing the requisite number of bits directly from the random engine, since we know that the individual bits are uniformly distributed.

But with this new range, we’ll have a non-power of 2 possible values, 2^{23} – 1 and 2^{52} – 1, respectively. The typical solution is to add 1 to the randomly generated integer of n bits, and then perform a floating point division by 2^{n} + 1. But as noted at the beginning, this is guaranteed to produce a less than perfectly uniform distribution, because you are mapping 2^{n} – 1 intermediate values to 2^{m} final values, and the final values will be targets for inconsistent quantities of the intermediate values. (Try mapping 2^{4} + 1 = 17 values to 2^{2} = 4 final values; at best, one of the final values will occur 25% more often than each of the other three.)

When dealing with large quantities, this non-uniformity is admittedly quite minor, but it is less than perfect. Besides, it does not work with the bit manipulation shown in the above section, incurring the performance cost of a floating point division instead. An alternative approach, important even when just working with integer ranges of non-power of 2 sizes, is the rejection method. Simply put, random numbers are generated repeatedly and invalid ones are rejected until one is generated that is within the desired range.

Yeah, at first consideration it sounds terrible. Pay the cost of generating potentially multiple random integers, with no definite cap on how many get rejected, plus the cost of a conditional for each iteration. But if the range of the initial random integer is properly controlled, things actually work out quite well under some circumstances, and this is fortunately just about the best circumstance possible.

Consider 32-bit floats first. We can almost use exactly the same process as in `float rand_float_co()`

above. The only thing we need to do is reject the bit pattern of 23 zeros, as that would produce 1.0f as an intermediate float, and consequently a final value of 0.0f after the subtraction. All other bit patterns of 23 bits are entirely valid. This means that the loop conditional is false around 99.99999% of the time. Thanks to branch prediction on modern processors, this will minimize the impact of the conditional on instruction pipelining, keeping things flowing nice and smoothly. For doubles it is even better, with the conditional being false over 99.9999999999999% of the time!

How shall we perform the rejection test? The obvious approach is to just mask off the unwanted bits, and then compare the remaining bits to 0. But that’s two operations per iteration. It just so happens that because we are using the high bits instead of the low bits, we can use a cleverly defined direct comparison operation and skip the mask entirely. Once we have an integer that passes our check, a single shift operation is enough to get the desired bits in the right arrangement, replacing the bit mask entirely. With all the above combined, here are the implementations for our next two functions:

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 |
float rand_float_oo() { uint32_t n; do { n = rand32(); } while (n <= 0x000001FFU); // If true, then the highest 23 bits must all be zero. return as_float(0x3F800000U | (n >> 9)) - 1.0f; } double rand_double_oo() { uint64_t n; do { n = rand64(); } while (n <= 0x0000000000000FFFULL); // If true, then the highest 52 bits must all be zero. return as_double(0x3FF0000000000000ULL | (n >> 12)) - 1.0; } |

## Step 3: Closed Range and Fancy Probability

And now we are ready to tackle the final two functions, generating numbers in the *closed* range [0, 1], within which both 0 *and* 1 are included. We might be tempted to implement this function using the same rejection method concept as used above, but in contrast to how the above case was nearly ideal concerning the predictability of the conditional, this case is the exact opposite.

The reason why is that there is one *more* possible output value, 2^{n} + 1, rather than one less. To utilize the rejection method, we must generate numbers in a *larger* range than needed, and the next power of 2 is 2^{(n + 1)}, which is nearly twice as large as 2^{n} + 1. The consequence is that almost half of our random integers will be rejected, meaning that on average, two random integers need to be generated for every floating point number produced. Further, CPU branch predication becomes worthless, because their is about a 50/50 chance of the branch going either way.

Fortunately, there’s a clever way around this, to turn the worst case into nearly the best case again, if one has some spare bits to work with. And we conveniently do indeed have spare bits to work with: 9 bits when generating floats, since we’re only using 23 out of 32, and 12 with doubles, after subtracting the 52 we’re using from the 64 generated.

So how does this work? We start once again by considering the half-open range [0, 1). All of these are perfectly valid for generating a number in the closed range [0, 1]; the only issue is that 1 itself is missing. So there needs to be a small chance, when generating a number, that 1 will be produced instead of something from the range [0, 1). And that small chance needs to be perfectly balanced so that 1 has an equal likelihood of being generated as any single value from [0, 1).

Let’s consider the 32-bit float case, so that we can work with concrete numbers. There are 2^{23} + 1 possible numbers in our output range [0, 1]. Naturally, that means that each individual value has a 1 in 2^{23} + 1 chance of being generated. Now we can very quickly generate a number in the range [0, 1), and there are exactly 2^{23} numbers in this range, each of which of course has a 1 in 2^{23} chance of being generated. This is *almost* the probability that we want them to have, but just a tiny bit more likely than it should be.

How can we adjust that probability? We combine it with an additional probability. The likelihood of two independent random events both happening is just the product of each of their individual probabilities. We know the probability of one event: Generating any one of the values in the range [0, 1) is 1 in 2^{23}; we’ll label that p. We also know the probability we want: 1 in 2^{23} + 1; we’ll call that r. So what we are looking for is a second random event q such that the probability of both p and q happening together is r:

Solving for q we get:

$q=\frac{{2}^{23}}{{2}^{23}+1}$Given that, here is the concept of our adjusted algorithm thus far: Generate a number in the range [0, 1). Do an additional probability check that has a 2^{23} in 2^{23} + 1 of succeeding. If it succeeds, which it almost always will, return the first random number generated. Otherwise, in that rare 1 in 2^{23} + 1 case, return 1.

Frustratingly, this still obviously requires performing an additional random check. And the probability of this second random check still has that same terribly uncooperative denominator which we were hoping to avoid. But if we further divide this second random check itself into two distinct random checks, one which is super cheap and usually true, and a second one which is expensive but rarely needs to be performed, we can once again get great performance in the common case, and only rarely deal with the nasty stuff. This is where the spare bits mentioned above come into play.

We have 9 spare bits which we can use to perform a check with 2^{9} – 1 in 2^{9} chance of passing (511 in 512 chance). I would say that reasonably counts as “almost always true”, but it’s not quite enough to reach the desired 2^{23} in 2^{23} + 1 probability. And to combine that 9-bit usually-true chance with a second random event such that the overall probability is *higher*, we would need to use a logical *OR* relationship, which in probability mathematics becomes a bit more awkward than an *AND* relationship.

To keep things simple with the math, let us invert the logic a bit. Instead of performing a usually true probability check and only returning 1 if it fails, we will perform a usually *false* probability check and return 1 in the rare case that it *succeeds*. The result is that we can use the 9 bits to perform a 1 in 2^{9} check. It’s a small likelihood, but not as small as 1 in 2^{23} + 1 that we are now looking for after inverting the earlier probability. And as intended, making the probability even smaller is as simple as requiring yet another probability check to succeed. Calculating this new probability follows the same process that we did above. Our first probability (p) is 1 in 512; our desired probability (r) is 1 in 2^{23} + 1. We can put that into the same formula p ⋅ q = r from above:

Solving this new equation for q we get:

$q=\frac{{2}^{9}}{{2}^{23}+1}$The rearranged and expanded algorithm is as follows: Generate a number in the range [0, 1). Do an additional probability check that has just a 1 in 2^{9} chance of succeeding. If it does succeed, then do a further probability check with a 2^{9} in 2^{23} + 1 chance of succeeding. If it also succeeds, return 1. Otherwise, as soon as either of those two checks fails, return the random number first generated. This ensures that 1 has exactly a 1 in 2^{23} + 1 chance of being returned, which by implication means that all the other values from 0 up to but not including 1 collectively have a 2^{23} in 2^{23} + 1 chance. And since there are exactly 2^{23} values in this range less than 1, and since we know that these already occur with equal probability, we can assert that each one individually has a 1 in 2^{23} + 1 probability of occurring, just like 1.

As for performance, most of the time, the first random check (which just uses the spare 9 bits already generated) will fail, and we can immediately return the random number generated. Only occasionally will we need to perform the awkward 2^{9} in 2^{23} + 1 check. This means that we will only need to generate one random integer under most circumstances. Additionally, branch prediction should work quite well too, since the conditional is highly predictable.

For doubles, everything is much the same, except that the first check is now 1 in 2^{12} (even better!) thanks to having 12 spare bits, and the messy check becomes 2^{12} in 2^{52} + 1.

The only detail remaining is the implementation of the awkward non-power of 2 probability check. That can be done simply enough using the same rejection method concept used earlier with the open range, just in a slightly different code form. I’ll provide this as separate functions, since doing so will simplify the structure of the main functions. I’ll also utilize a bit manipulation trick to get a mask appropriate for minimizing the number of rejections.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 |
bool rand_probability(uint32_t numerator, uint32_t denominator) { uint32_t mask = denominator - 1U; mask |= mask >> 1; mask |= mask >> 2; mask |= mask >> 4; mask |= mask >> 8; mask |= mask >> 16; uint32_t n; do { n = rand32() & mask; } while (n >= denominator); return n < numerator; } bool rand_probability(uint64_t numerator, uint64_t denominator) { uint64_t mask = denominator - 1ULL; mask |= mask >> 1; mask |= mask >> 2; mask |= mask >> 4; mask |= mask >> 8; mask |= mask >> 16; mask |= mask >> 32; uint64_t n; do { n = rand64() & mask; } while (n >= denominator); return n < numerator; } |

With those helper functions in place, our final two random floating point functions look like this:

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 |
float rand_float_cc() { uint32_t n = rand32(); // First check if the upper 9 bits are all set. // If not, short circuit to the else block. // If so, carry on with the messy check. if (n >= 0xFF800000U && rand_probability(0x00000200U, 0x00800001U)) { return 1.0f; } else { return as_float(0x3F800000U | (n & 0x007FFFFFU)) - 1.0f; } } double rand_double_cc() { uint64_t n = rand64(); // First check if the upper 12 bits are all set. // If not, short circuit to the else block. // If so, carry on with the messy check. if (n >= 0xFFF0000000000000ULL && rand_probability(0x00001000ULL, 0x0010000000000001ULL)) { return 1.0; } else { return as_double(0x3FF0000000000000ULL | (n & 0x000FFFFFFFFFFFFFULL)) - 1.0; } } |

## High Precision Alternative Distribution

As mentioned near the beginning of the article, all of the above functions sacrifice possible precision that floating point formats are capable of representing for the sake of more perfect uniformity. If you would rather get the higher precision for numbers close to zero, but wish to still avoid division and take advantage of the techniques above, there is a way to do this.

The key is to use standard integer-to-floating-point conversion to turn a randomly generated integer into the nearest equivalent floating point representation, and then to subtract the appropriate amount from the floating point number’s exponent. By subtracting n from the exponent, this is essentially the same as dividing the number by 2^{n}, but is faster, with the tradeoff that the cast operation is more complex than simple bit manipulations. This process can also be interpretted as converting a fixed point value to floating point.

Because this requires a back-and-forth interpretation of bits as integers and floating point values, it is something that needs to be done within the context of the type punning functionality introduced earlier.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 |
float as_float_precise(uint32_t i) { if (i == 0U) return 0.0f; union { float f; uint32_t i; } pun = { (float)i }; // Subtract 32 from exponent (32 << 23) // Same as dividing by 2^32 pun.i -= 0x10000000U; return pun.f; } float as_float_precise(uint64_t i) { if (i == 0ULL) return 0.0f; union { float f; uint32_t i; } pun = { (float)i }; // Subtract 64 from exponent (64 << 23) // Same as dividing by 2^64 pun.i -= 0x20000000U; return pun.f; } double as_double_precise(uint64_t i) { if (i == 0ULL) return 0.0; union { double f; uint64_t i; } pun = { (double)i }; // Subtract 64 from exponent (64 << 52) // Same as dividing by 2^64 pun.i -= 0x0400000000000000ULL; return pun.f; } |

These functions are capable of taking a full random sequence of 32 or 64 bits respectively, and converting them into floats and doubles with the extra precision required. Note that they check for the input equaling zero beforehand, because zero as a floating point number does not have the appropriate bitwise format to behave properly with the exponential subtraction and thus needs to be handled explicitly.

Also note that there is an extra function, the second one, that takes 64 bits as input but still only returns a single precision float. This would have been irrelevant with the earlier function, because only 23 bits were ever used or useful in that case. In this case, however, more bits can indeed be useful, for numbers that get closer to zero. For larger numbers, the lower bits will just get truncated away, while with smaller numbers, all the upper bits will be zero anyway, so the lower bits can be utilized for extra precision. Using 32 bits is better than nothing, because the extra 9 bits provide some degree of higher absolute precision for smaller numbers. But if you are willing to generate 64 bits, they can also be put to use. In fact, even 128 bits could be useful if you wanted your numbers to be able to get *really* tiny, but do not know what options there are on various platforms for quickly converting a 128-bit integer to a floating point number.

One final note: Because standard conversion is used, the floating point rounding rules come into play, so you should be aware of how they can affect the results. In particular, if you want to produce floating point numbers up to be not including one, you will need to use the rejection method to keep your random integers capped to a certain maximum, close to but a little less than the maximum possible for the number of bits being generated. Anything above this maximum would still produce a number technically less than one after the exponent subtraction is performed if an infinite-precision format were used, but in practice might be rounded up to one depending on the active rounding mode.

## Using the Precise Method to Gain an Extra Bit

The original bit-manipulation method happens to leave a single bit of available precision unused. This is due to the fact that we are restricting ourselves to the precision offered within the range [1, 2) when we first assign the random bits combined with the fixed exponent, even though the range [0.5, 1) offers double the absolute precision, and ranges below that offer even more.

But trying to get access to this extra bit of precision is tricky, as it isn’t at the same position for all numbers. When the floating point subtraction is performed to shift the [1, 2) number into the [0, 1) range, a left bit shift is essentially being done on the mantissa, based on how small the exponent gets, which is in turn based on how close to zero the number gets. There will always be at least one bit’s worth of shifting, which is how the extra available bit is introduced, but before the shift happens there is nowhere to store that bit, and after the shift we cannot (quickly) know which bit to set.

The more precise method just introduced, of casting the integer to float and then reducing the exponent, gets around this problem, as the left shift occurs during the cast and has access to potentially more bits than will fit into the final mantissa. In the form presented above, we take full advantage of those extra bits, which both induces possible rounding to account for the excess of bits and loses the perfect uniformity. But if we are careful to limit the number of bits used, we can avoid both the rounding and the loss of uniformity.

To do so, we use just 24 bits for floats (one better than the 23 bits used by the first method), and 53 bits for doubles (instead of just 52 bits). Then we subtract 24 or 53 from the exponent, instead of the 32 or 64 when using all the available bits for maximum precision. When compared with the first method, we basically trade a floating point addition with a cast from integer to floating point, and gain a bit of precision in the exchange. Unfortunately, as we already saw with the casting method above, zero is a special case that I could find no clever way around, so it also introduces a conditional; one that is almost always true, but still a complication that could impact performance compared to the first method.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 |
float as_float_cast(uint32_t i) { union { float f; uint32_t i; } pun = { (float)i }; pun.i -= 0x0C000000U; return pun.f; } double as_double_cast(uint64_t i) { union { double f; uint64_t i; } pun = { (double)i }; pun.i -= 0x0350000000000000ULL; return pun.f; } float rand_float_co_cast() { auto n = rand32() >> 8; return n != 0U ? as_float_cast(n) : 0.0f; } double rand_double_co_cast() { auto n = rand64() >> 11; return n != 0ULL ? as_double_cast(n) : 0.0; } float rand_float_oc_cast() { auto n = rand32() >> 8; return n != 0U ? as_float_cast(n) : 1.0f; } double rand_double_oc_cast() { auto n = rand64() >> 11; return n != 0ULL ? as_double_cast(n) : 1.0; } float rand_float_oo_cast() { uint32_t n; do { n = rand32(); } while (n <= 0x000000FFU); // If true, then the highest 24 bits must all be zero. return as_float_cast(n >> 8); } double rand_double_oo_cast() { uint64_t n; do { n = rand64(); } while (n <= 0x00000000000007FFULL); // If true, then the highest 53 bits must all be zero. return as_double_cast(n >> 11); } float rand_float_cc_cast() { uint32_t n = rand32(); // First check if the upper 8 bits are all set. // If not, short circuit to the else block. // If so, carry on with the messy check. if (n >= 0xFF000000U && rand_probability(random, 0x00000100U, 0x01000001U)) { return 0.0f; } else { return as_float_cast((n & 0x00FFFFFFU) + 1U); } } double rand_double_cc_cast() { uint64_t n = rand64(); // First check if the upper 11 bits are all set. // If not, short circuit to the else block. // If so, carry on with the messy check. if (n >= 0xFFE0000000000000ULL && rand_probability(random, 0x00000800ULL, 0x0020000000000001ULL)) { return 0.0; } else { return as_double_cast((n & 0x001FFFFFFFFFFFFFULL) + 1ULL); } } |

## Performance Measurements

No post on writing high performance code is complete without some performance numbers, so I have written a quick program to test these functions out and compare them with their division-based counterparts. Out of curiosity, I also included the a few measurements of some C++11 utilities. It measures partially unrolled loops for around a given number of seconds, and then records the average nanoseconds required per call. It also includes a warmup phase at the begining to get all the caches prepared. The data below is from executing 16 calls per loop iteration, with a half-second warmup and ten second measurement period.

Here is the data from my x86-64 desktop machine with an AMD A10-7850K processor, with the benchmark application built by Visual Studio 2015 in 64-bit mode with settings configured to favor speed over size and to inline any suitable functions. The fourth column is just the time it takes to construct a floating point value, calculated by subtracting the average time it takes to just generate a single random integer (the first row in each table) from the total time of the operation. The fifth column is that same overhead, expressed as a percentage.

32-bit Operation | Total ops/s | Total ns/op | Overhead ns/op | Overhead % |
---|---|---|---|---|

rand_next32 | 630,128,871 | 1.587 | ||

rand_float_co | 455,120,444 | 2.197 | 0.610 | 38.5% |

rand_float_oc | 435,561,244 | 2.296 | 0.709 | 44.7% |

rand_float_oo | 339,964,172 | 2.941 | 1.355 | 85.4% |

rand_float_cc | 239,279,454 | 4.179 | 2.592 | 163.3% |

rand_float_co_cast | 306,611,060 | 3.261 | 1.674 | 105.5% |

rand_float_oc_cast | 306,326,767 | 3.264 | 1.678 | 105.7% |

rand_float_oo_cast | 313,051,776 | 3.194 | 1.607 | 101.3% |

rand_float_cc_cast | 264,845,545 | 3.776 | 2.189 | 137.9% |

rand_float_co_div | 352,447,100 | 2.837 | 1.250 | 78.8% |

rand_float_oc_div | 350,996,447 | 2.849 | 1.262 | 79.5% |

rand_float_oo_div | 327,496,401 | 3.053 | 1.466 | 92.4% |

rand_float_cc_div | 363,223,592 | 2.753 | 1.166 | 73.5% |

64-bit Operation | Total ops/s | Total ns/op | Overhead ns/op | Overhead % |
---|---|---|---|---|

rand_next64 | 622,868,711 | 1.605 | ||

rand_double_co | 439,437,295 | 2.276 | 0.670 | 41.7% |

rand_double_oc | 446,584,345 | 2.239 | 0.634 | 39.5% |

rand_double_oo | 339,282,071 | 2.947 | 1.342 | 83.6% |

rand_double_cc | 250,406,659 | 3.994 | 2.388 | 148.7% |

rand_double_co_cast | 318,517,898 | 3.140 | 1.534 | 95.6% |

rand_double_oc_cast | 319,119,101 | 3.134 | 1.528 | 95.2% |

rand_double_oo_cast | 313,245,056 | 3.192 | 1.587 | 98.8% |

rand_double_cc_cast | 171,230,955 | 5.840 | 4.235 | 263.8% |

rand_double_co_div | 91,477,340 | 10.932 | 9.326 | 580.9% |

rand_double_oc_div | 133,058,891 | 7.515 | 5.910 | 368.1% |

rand_double_oo_div | 86,363,972 | 11.579 | 9.973 | 621.2% |

rand_double_cc_div | 141,402,060 | 7.072 | 5.467 | 340.5% |

For comparison, here are times for generating random floats and doubles in the range [0, 1) using standard C++ utilities. Two of the rows just measure the raw bit generation of the `std::ranlux24`

and `std::ranlux48`

engines, while the other two include the use of `std::uniform_real<T>`

to generate floating point values. This utility is substantially slower when compared to any of my custom methods (as implemented in Visual Studio 2015 anyway), though I would not be surprised if the distribution quality of the C++ utilities are in some way better; it’s not something I’ve investigated yet.

Standard Operation | Total ops/s | Total ns/op | Overhead ns/op | Overhead % |
---|---|---|---|---|

ranlux24 | 116,584,472 | 8.577 | ||

flt_co(ranlux24) | 16,766,934 | 59.641 | 51.064 | 595.3% |

ranlux48 | 98,892,721 | 10.112 | ||

dbl_co(ranlux48) | 13,442,138 | 74.393 | 64.281 | 635.7% |

These are the results on my laptop’s Intel Core i7-4700HQ, also an x86-64 machine, and using the exact same executable file for the benchmark application. Not only is it faster overall, but it also seems to suffer less overhead beyond the generation of bits when constructing the final floating point value. Might be the result of better pipelining behavior or floating point performance compared to my AMD processor?

32-bit Operation | Total ops/s | Total ns/op | Overhead ns/op | Overhead % |
---|---|---|---|---|

rand_next32 | 617,465,578 | 1.620 | ||

rand_float_co | 533,629,167 | 1.874 | 0.254 | 15.7% |

rand_float_oc | 524,876,401 | 1.905 | 0.286 | 17.6% |

rand_float_oo | 416,715,822 | 2.400 | 0.780 | 48.2% |

rand_float_cc | 288,415,301 | 3.467 | 1.848 | 114.1% |

rand_float_co_cast | 308,418,977 | 3.242 | 1.623 | 100.2% |

rand_float_oc_cast | 315,350,235 | 3.171 | 1.552 | 95.8% |

rand_float_oo_cast | 393,888,418 | 2.539 | 0.919 | 56.8% |

rand_float_cc_cast | 276,605,718 | 3.615 | 1.996 | 123.2% |

rand_float_co_div | 322,211,971 | 3.104 | 1.484 | 91.6% |

rand_float_oc_div | 319,745,100 | 3.127 | 1.508 | 93.1% |

rand_float_oo_div | 312,832,881 | 3.197 | 1.577 | 97.4% |

rand_float_cc_div | 324,085,939 | 3.086 | 1.466 | 90.5% |

64-bit Operation | Total ops/s | Total ns/op | Overhead ns/op | Overhead % |
---|---|---|---|---|

rand_next64 | 619,162,882 | 1.615 | ||

rand_double_co | 525,873,328 | 1.902 | 0.287 | 17.7% |

rand_double_oc | 517,698,877 | 1.932 | 0.317 | 19.6% |

rand_double_oo | 412,640,202 | 2.423 | 0.808 | 50.0% |

rand_double_cc | 288,140,814 | 3.471 | 1.855 | 114.9% |

rand_double_co_cast | 404,888,226 | 2.470 | 0.855 | 52.9% |

rand_double_oc_cast | 405,158,394 | 2.468 | 0.853 | 52.8% |

rand_double_oo_cast | 414,983,760 | 2.410 | 0.795 | 49.2% |

rand_double_cc_cast | 235,530,408 | 4.246 | 2.631 | 162.9% |

rand_double_co_div | 216,825,435 | 4.612 | 2.997 | 185.6% |

rand_double_oc_div | 216,306,277 | 4.623 | 3.008 | 186.2% |

rand_double_oo_div | 214,925,793 | 4.653 | 3.038 | 188.1% |

rand_double_cc_div | 220,499,270 | 4.535 | 2.920 | 180.8% |

Standard Operation | Total ops/s | Total ns/op | Overhead ns/op | Overhead % |
---|---|---|---|---|

ranlux24 | 113,273,936 | 8.828 | ||

flt_co(ranlux24) | 20,950,600 | 47.731 | 38.903 | 440.7% |

ranlux48 | 114,318,182 | 8.748 | ||

dbl_co(ranlux48) | 23,091,134 | 43.307 | 34.559 | 395.1% |

The overall pattern is that the primary bit-manipulating method introduced tends to be fastest, while the cast and division methods differ based on the strength of the CPU’s floating point capabilities. Also, the half-open ranges tend to be faster because they can more often be performed without any branches, though the division method is more consistent due to giving up perfect uniformity in favor of simple branchless implementation for all four of the ranges.

And I must state the usual caveat that effective benchmarking is hard, and I’m no expert. I repeatedly wound up with strange results and multiple times rewrote parts of the measurement code trying to make things more reliable. In particular, making the compiler perform all legitimate optimizations while not eliminating what it thinks is dead code is awkward. I ended up using an exclusive-or accumulation that ultimately gets printed out along with the timing results, so that all computations *must* be performed, with the hope that the exclusive-or operation itself would have a negligible effect on the timing. I also had to forcibly prevent a couple of functions (`rand_probability`

) from getting inlined, because that was causing a rarely executed slow path to be optimized in a way that harmed the performance of the fast path.

Because I originally did this work for my Make It Random Unity asset, it was originally in C#, running on Unity’s forked version of the Mono compiler and runtime. Within that environment, I definitely benefited from the engineering effort of implementing (and where necessary figuring out) these techniques. But I never directly measured them against the division method, since I had purposefully started out wanting to get the cleaner uniform distribution. Nor have I yet tested the casting method that gains an additional bit of precision, because I only devised that method in the course of wrirting this post. I *did* measure them extensively against equivalent methods in `UnityEngine.Random`

and `System.Random`

, since that was my practical competition, but the differences there extend well beyond the process of turning random bits into a floating point value.

## Full C++ Source for Benchmark

Below I have included the full source code of the benchmark application (plus some additional validation code), if you would like to try it on your own system. I included some additional C++ features that are not part of the code above, such as packaging the random engines into structs rather than using a global state in order to easily multithread the validation code, but all the operational details are identical.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 |
/* Written in 2016 by Andy Gainey (againey@experilous.com) To the extent possible under law, the author has dedicated all copyright and related and neighboring rights to this software to the public domain worldwide. This software is distributed without any warranty. See <http://creativecommons.org/publicdomain/zero/1.0/>. */ #include <map> #include <string> #include <iostream> #include <iomanip> #include <sstream> #include <algorithm> #include <cstdint> #include <random> #include <future> using std::uint32_t; using std::uint64_t; // OS-specific functions; add your own for other environments/platforms. #ifdef _WIN32 #define NOMINMAX #define WIN32_LEAN_AND_MEAN #include <Windows.h> #define NOINLINE __declspec(noinline) uint64_t get_current_tick() { uint64_t tick; QueryPerformanceCounter(reinterpret_cast<LARGE_INTEGER*>(&tick)); return tick; } uint64_t get_tick_count(double seconds) { uint64_t freq; QueryPerformanceFrequency(reinterpret_cast<LARGE_INTEGER*>(&freq)); return static_cast<uint64_t>(static_cast<double>(freq) * seconds); } void prepare_thread() { HANDLE hProc = GetCurrentProcess(); HANDLE hThread = GetCurrentThread(); DWORD_PTR proc_affinity_mask, sys_affinity_mask; GetProcessAffinityMask(hProc, &proc_affinity_mask, &sys_affinity_mask); proc_affinity_mask = 0x1ULL << 63; while ((proc_affinity_mask & sys_affinity_mask) == 0ULL && proc_affinity_mask != 0ULL) { proc_affinity_mask = proc_affinity_mask >> 1; } if (proc_affinity_mask == 0ULL) proc_affinity_mask = sys_affinity_mask; SetProcessAffinityMask(hProc, proc_affinity_mask); SetThreadAffinityMask(hThread, proc_affinity_mask); SetPriorityClass(hProc, ABOVE_NORMAL_PRIORITY_CLASS); SetThreadPriority(hThread, THREAD_PRIORITY_ABOVE_NORMAL); } #endif float as_float(uint32_t i) { union { uint32_t i; float f; } pun = { i }; return pun.f; } float as_float_cast(uint32_t i) { union { float f; uint32_t i; } pun = { (float)i }; pun.i -= 0x0C000000U; return pun.f; } double as_double(uint64_t i) { union { uint64_t i; double f; } pun = { i }; return pun.f; } double as_double_cast(uint64_t i) { union { double f; uint64_t i; } pun = { (double)i }; pun.i -= 0x0350000000000000ULL; return pun.f; } uint32_t as_int(float f) { union { float f; uint32_t i; } pun = { f }; return pun.i; } uint64_t as_int(double f) { union { double f; uint64_t i; } pun = { f }; return pun.i; } // http://xoroshiro.di.unimi.it/xoroshiro128plus.c class xoroshiro_128_plus { private: uint64_t state0; uint64_t state1; static inline uint64_t rotl(const uint64_t x, int k) { return (x << k) | (x >> (64 - k)); } public: xoroshiro_128_plus() = default; xoroshiro_128_plus(uint64_t state0, uint64_t state1) : state0(state0), state1(state1) {} uint64_t next64() { const uint64_t s0 = state0; uint64_t s1 = state1; uint64_t r = s0 + s1; s1 ^= s0; state0 = rotl(s0, 55) ^ s1 ^ (s1 << 14); state1 = rotl(s1, 36); return r; } uint32_t next32() { const uint64_t s0 = state0; uint64_t s1 = state1; uint64_t r = s0 + s1; s1 ^= s0; state0 = rotl(s0, 55) ^ s1 ^ (s1 << 14); state1 = rotl(s1, 36); return static_cast<uint32_t>(r); } }; // http://xoroshiro.di.unimi.it/xorshift128plus.c class xorshift_128_plus { private: uint64_t state0; uint64_t state1; public: xorshift_128_plus() = default; xorshift_128_plus(uint64_t state0, uint64_t state1) : state0(state0), state1(state1) {} uint64_t next64() { uint64_t x = state0; uint64_t y = state1; uint64_t r = x + y; state0 = y; x ^= x << 23; state1 = x ^ y ^ (x >> 18) ^ (y >> 5); return r; } uint32_t next32() { uint64_t x = state0; uint64_t y = state1; uint64_t r = x + y; state0 = y; x ^= x << 23; state1 = x ^ y ^ (x >> 18) ^ (y >> 5); return static_cast<uint32_t>(r); } }; template <typename TRandom> float rand_float_co(TRandom& random) { return as_float(0x3F800000U | (random.next32() >> 9)) - 1.0f; } template <typename TRandom> double rand_double_co(TRandom& random) { return as_double(0x3FF0000000000000ULL | (random.next64() >> 12)) - 1.0; } template <typename TRandom> float rand_float_oc(TRandom& random) { return 2.0f - as_float(0x3F800000U | (random.next32() >> 9)); } template <typename TRandom> double rand_double_oc(TRandom& random) { return 2.0 - as_double(0x3FF0000000000000ULL | (random.next64() >> 12)); } template <typename TRandom> float rand_float_oo(TRandom& random) { uint32_t n; do { n = random.next32(); } while (n <= 0x000001FFU); // If true, then the highest 23 bits must all be zero. return as_float(0x3F800000U | (n >> 9)) - 1.0f; } template <typename TRandom> double rand_double_oo(TRandom& random) { uint64_t n; do { n = random.next64(); } while (n <= 0x0000000000000FFFULL); // If true, then the highest 52 bits must all be zero. return as_double(0x3FF0000000000000ULL | (n >> 12)) - 1.0; } // Force no-inline, because otherwise, the compiler might be inclined to inline this // into the non-critical slow path of the calling function, which in turns makes that // function too large to inline, hampering its fast path with an unnecessary function // call indirection. template <typename TRandom> NOINLINE bool rand_probability(TRandom& random, uint32_t numerator, uint32_t denominator) { uint32_t mask = denominator - 1U; mask |= mask >> 1; mask |= mask >> 2; mask |= mask >> 4; mask |= mask >> 8; mask |= mask >> 16; uint32_t n; do { n = random.next32() & mask; } while (n >= denominator); return n < numerator; } template <typename TRandom> NOINLINE bool rand_probability(TRandom& random, uint64_t numerator, uint64_t denominator) { uint64_t mask = denominator - 1ULL; mask |= mask >> 1; mask |= mask >> 2; mask |= mask >> 4; mask |= mask >> 8; mask |= mask >> 16; mask |= mask >> 32; uint64_t n; do { n = random.next64() & mask; } while (n >= denominator); return n < numerator; } template <typename TRandom> float rand_float_cc(TRandom& random) { uint32_t n = random.next32(); // First check if the upper 9 bits are all set. // If not, short circuit to the else block. // If so, carry on with the messy check. if (n >= 0xFF800000U && rand_probability(random, 0x00000200U, 0x00800001U)) { return 1.0f; } else { return as_float(0x3F800000U | (n & 0x007FFFFFU)) - 1.0f; } } template <typename TRandom> double rand_double_cc(TRandom& random) { uint64_t n = random.next64(); // First check if the upper 12 bits are all set. // If not, short circuit to the else block. // If so, carry on with the messy check. if (n >= 0xFFF0000000000000ULL && rand_probability(random, 0x00001000ULL, 0x0010000000000001ULL)) { return 1.0; } else { return as_double(0x3FF0000000000000ULL | (n & 0x000FFFFFFFFFFFFFULL)) - 1.0; } } template <typename TRandom> float rand_float_co_cast(TRandom& random) { auto n = random.next32() >> 8; return n != 0U ? as_float_cast(n) : 0.0f; } template <typename TRandom> double rand_double_co_cast(TRandom& random) { auto n = random.next64() >> 11; return n != 0ULL ? as_double_cast(n) : 0.0; } template <typename TRandom> float rand_float_oc_cast(TRandom& random) { auto n = random.next32() >> 8; return n != 0U ? as_float_cast(n) : 1.0f; } template <typename TRandom> double rand_double_oc_cast(TRandom& random) { auto n = random.next64() >> 11; return n != 0ULL ? as_double_cast(n) : 1.0; } template <typename TRandom> float rand_float_oo_cast(TRandom& random) { uint32_t n; do { n = random.next32(); } while (n <= 0x000000FFU); // If true, then the highest 24 bits must all be zero. return as_float_cast(n >> 8); } template <typename TRandom> double rand_double_oo_cast(TRandom& random) { uint64_t n; do { n = random.next64(); } while (n <= 0x00000000000007FFULL); // If true, then the highest 53 bits must all be zero. return as_double_cast(n >> 11); } template <typename TRandom> float rand_float_cc_cast(TRandom& random) { uint32_t n = random.next32(); // First check if the upper 8 bits are all set. // If not, short circuit to the else block. // If so, carry on with the messy check. if (n >= 0xFF000000U && rand_probability(random, 0x00000100U, 0x01000001U)) { return 0.0f; } else { return as_float_cast((n & 0x00FFFFFFU) + 1U); } } template <typename TRandom> double rand_double_cc_cast(TRandom& random) { uint64_t n = random.next64(); // First check if the upper 11 bits are all set. // If not, short circuit to the else block. // If so, carry on with the messy check. if (n >= 0xFFE0000000000000ULL && rand_probability(random, 0x00000800ULL, 0x0020000000000001ULL)) { return 0.0; } else { return as_double_cast((n & 0x001FFFFFFFFFFFFFULL) + 1ULL); } } template <typename TRandom> float rand_float_co_div(TRandom& random) { return static_cast<float>(random.next32()) / 4294967808.0f; } template <typename TRandom> double rand_double_co_div(TRandom& random) { return static_cast<double>(random.next64()) / 18446744073709555712.0; } template <typename TRandom> float rand_float_oc_div(TRandom& random) { return (static_cast<float>(random.next32()) + 1.0f) / 4294967296.0f; } template <typename TRandom> double rand_double_oc_div(TRandom& random) { return (static_cast<double>(random.next64()) + 1.0) / 18446744073709551616.0; } template <typename TRandom> float rand_float_oo_div(TRandom& random) { return (static_cast<float>(random.next32()) + 1.0f) / 4294967808.0f; } template <typename TRandom> double rand_double_oo_div(TRandom& random) { return (static_cast<double>(random.next64()) + 1.0) / 18446744073709555712.0; } template <typename TRandom> float rand_float_cc_div(TRandom& random) { return static_cast<float>(random.next32()) / 4294967296.0f; } template <typename TRandom> double rand_double_cc_div(TRandom& random) { return static_cast<double>(random.next64()) / 18446744073709551616.0; } template <typename TFunc, typename TRandom> std::string validate_float_distribution(TFunc rand_func, TRandom& random, int loop_count, int iterations_per_loop, bool inclusive_lower, bool inclusive_upper, bool cast) { std::map<float, int> f_count; std::vector<float> f_invalid; float superZero; float half; float subOne; float min_value; float max_value; int min_count; int max_count; double average_count; double deviance_sum; if (cast) { if (inclusive_lower) { f_count.insert(std::make_pair(0.0f, 0)); } for (uint32_t i = 1U; i < 0x01000000U; ++i) { f_count.insert(std::make_pair(as_float_cast(i), 0)); } if (inclusive_upper) { f_count.insert(std::make_pair(1.0f, 0)); } superZero = as_float_cast(0x00000001U); half = as_float_cast(0x00800000U); subOne = as_float_cast(0x00FFFFFFU); } else { for (uint32_t i = inclusive_lower ? 0U : 1U; i < 0x00800000U; ++i) { f_count.insert(std::make_pair(as_float(0x3F800000U | i) - 1.0f, 0)); } if (inclusive_upper) { f_count.insert(std::make_pair(1.0f, 0)); } superZero = as_float(0x3F800001U) - 1.0f; half = 0.5f; subOne = as_float(0x3FFFFFFFU) - 1.0f; } double unique_values = (double)f_count.size(); for (int loop = 0; loop < loop_count; ++loop) { std::cout << "Loop " << loop << std::endl; for (int i = 0; i < iterations_per_loop; ++i) { float f = rand_func(random); auto h = f_count.find(f); if (h != f_count.end()) { h->second = h->second + 1; } else { f_invalid.push_back(f); } } } min_value = f_count.begin()->first; max_value = min_value; min_count = f_count.begin()->second; max_count = min_count; average_count = (double)loop_count * (double)iterations_per_loop / unique_values; deviance_sum = 0.0; for (const auto count : f_count) { if (count.second < min_count) { min_value = count.first; min_count = count.second; } if (count.second > max_count) { max_value = count.first; max_count = count.second; } double delta = (double)count.second - average_count; deviance_sum += delta * delta; } std::stringstream ss; ss << "Validating rand_float_" << (inclusive_lower ? 'c' : 'o') << (inclusive_upper ? 'c' : 'o') << (cast ? "" : "_cast") << "()" << std::endl; ss << "Unique Values = " << f_count.size() << std::endl; for (auto f : f_invalid) { ss << "Invalid number generated: " << std::setprecision(8) << std::fixed << f << std::endl; } ss << "Average Count = " << std::setprecision(16) << std::fixed << (double)loop_count * (double)iterations_per_loop / unique_values << std::endl; ss << "Std Deviation = " << std::setprecision(16) << std::fixed << std::sqrt(deviance_sum / (double)loop_count / (double)iterations_per_loop) << std::endl; ss << "Minimum Count = " << min_count << " (" << std::setprecision(8) << std::fixed << min_value << ")" << std::endl; ss << "Maximum Count = " << max_count << " (" << std::setprecision(8) << std::fixed << max_value << ")" << std::endl; if (inclusive_lower) ss << "Zero Count = " << f_count[0.0f] << std::endl; ss << "Sup-Zro Count = " << f_count[superZero] << std::endl; ss << "Half Count = " << f_count[half] << std::endl; ss << "Sub-One Count = " << f_count[subOne] << std::endl; if (inclusive_upper) ss << "One Count = " << f_count[1.0f] << std::endl; ss << std::endl; return ss.str(); } template <typename TSource> struct sink_type_traits { typedef TSource sink_type; static sink_type sink(TSource source) { return source; } }; template <> struct sink_type_traits<float> { typedef uint32_t sink_type; static sink_type sink(float source) { return reinterpret_cast<sink_type&&>(source); } }; template <> struct sink_type_traits<double> { typedef uint64_t sink_type; static sink_type sink(double source) { return reinterpret_cast<sink_type&&>(source); } }; template <typename TFunc> auto rand_loop(TFunc rand_func, uint64_t iterations) -> typename sink_type_traits<decltype(rand_func())>::sink_type { typename sink_type_traits<decltype(rand_func())>::sink_type sink; uint64_t outer_loops = iterations >> 4; for (uint64_t i = 0ULL; i < outer_loops; ++i) { sink ^= sink_type_traits<decltype(rand_func())>::sink(rand_func()); sink ^= sink_type_traits<decltype(rand_func())>::sink(rand_func()); sink ^= sink_type_traits<decltype(rand_func())>::sink(rand_func()); sink ^= sink_type_traits<decltype(rand_func())>::sink(rand_func()); sink ^= sink_type_traits<decltype(rand_func())>::sink(rand_func()); sink ^= sink_type_traits<decltype(rand_func())>::sink(rand_func()); sink ^= sink_type_traits<decltype(rand_func())>::sink(rand_func()); sink ^= sink_type_traits<decltype(rand_func())>::sink(rand_func()); sink ^= sink_type_traits<decltype(rand_func())>::sink(rand_func()); sink ^= sink_type_traits<decltype(rand_func())>::sink(rand_func()); sink ^= sink_type_traits<decltype(rand_func())>::sink(rand_func()); sink ^= sink_type_traits<decltype(rand_func())>::sink(rand_func()); sink ^= sink_type_traits<decltype(rand_func())>::sink(rand_func()); sink ^= sink_type_traits<decltype(rand_func())>::sink(rand_func()); sink ^= sink_type_traits<decltype(rand_func())>::sink(rand_func()); sink ^= sink_type_traits<decltype(rand_func())>::sink(rand_func()); } return sink; } template <typename TFunc> double measure_operations(TFunc rand_func, double warmup_time, double measure_time, std::string op_name, double baseline_ops_per_second = 0.0) { uint64_t warmup_ticks = get_tick_count(warmup_time); uint64_t measure_ticks = get_tick_count(measure_time); uint64_t target_warmup_batch_ticks = std::min(get_tick_count(1.0), warmup_ticks / 4ULL); uint64_t target_measure_batch_ticks = std::min(get_tick_count(1.0), warmup_ticks / 16ULL); uint64_t batch_iteration_count = 1024ULL; uint64_t warmup_iteration_count = 0ULL; uint64_t measure_iteration_count = 0ULL; uint64_t current_tick; uint64_t elapsed_ticks; typename sink_type_traits<decltype(rand_func())>::sink_type sink = 0; uint64_t warmup_start_tick = get_current_tick(); while (true) { sink ^= rand_loop(rand_func, batch_iteration_count); warmup_iteration_count += batch_iteration_count; current_tick = get_current_tick(); elapsed_ticks = current_tick - warmup_start_tick; if (elapsed_ticks >= warmup_ticks) { break; } batch_iteration_count = batch_iteration_count << 1; uint64_t expected_batch_ticks = batch_iteration_count * elapsed_ticks / warmup_iteration_count; uint64_t next_batch_ticks = std::min(expected_batch_ticks, target_warmup_batch_ticks); next_batch_ticks = std::min(next_batch_ticks, warmup_ticks - elapsed_ticks); batch_iteration_count = next_batch_ticks * warmup_iteration_count / elapsed_ticks; batch_iteration_count = (batch_iteration_count + 15ULL) & ~0xFULL; } batch_iteration_count = target_measure_batch_ticks * warmup_iteration_count / elapsed_ticks; batch_iteration_count = (batch_iteration_count + 15ULL) & ~0xFULL; sink = 0; uint64_t measure_start_tick = get_current_tick(); while (true) { sink ^= rand_loop(rand_func, batch_iteration_count); measure_iteration_count += batch_iteration_count; current_tick = get_current_tick(); elapsed_ticks = current_tick - measure_start_tick; if (elapsed_ticks >= measure_ticks) { break; } batch_iteration_count = batch_iteration_count << 1; uint64_t expected_batch_ticks = batch_iteration_count * elapsed_ticks / measure_iteration_count; uint64_t next_batch_ticks = std::min(expected_batch_ticks, target_measure_batch_ticks); next_batch_ticks = std::min(next_batch_ticks, measure_ticks - elapsed_ticks); batch_iteration_count = next_batch_ticks * measure_iteration_count / elapsed_ticks; batch_iteration_count = (batch_iteration_count + 15ULL) & ~0xFULL; } double ops_per_second = static_cast<double>(measure_iteration_count) * static_cast<double>(get_tick_count(1.0)) / static_cast<double>(elapsed_ticks); double nanoseconds_per_op = 1000000000.0 / ops_per_second; double baseline_nanoseconds_per_op = 1000000000.0 / baseline_ops_per_second; std::cout << std::setw(20) << op_name << ", " << std::setprecision(0) << std::fixed << std::setw(9) << ops_per_second << " ops/s" << ", " << std::setprecision(5) << std::fixed << std::setw(9) << nanoseconds_per_op << " ns/op"; if (baseline_ops_per_second != 0.0) { double overhead = baseline_ops_per_second / ops_per_second - 1.0; std::cout << ", " << std::setprecision(5) << std::fixed << std::setw(9) << (nanoseconds_per_op - baseline_nanoseconds_per_op) << " ns/op" << ", " << std::setprecision(3) << std::fixed << std::setw(7) << overhead * 100.0 << "%"; } std::cout << ", (sink = " << sink << ")"; std::cout << std::endl; return ops_per_second; } int main() { // Use these functions to statistically validate distributions. std::vector<std::future<std::string>> validation_results; // To avoid false sharing with threads, spread out some random engines in memory. std::vector<xorshift_128_plus> random_engines; random_engines.resize(64); random_engines[0] = xorshift_128_plus(0xA6E9377DAF75BDFEULL, 0x863F5CB508510D95ULL); random_engines[8] = xorshift_128_plus(0xA6E9377DAF75BDFEULL, ~0x863F5CB508510D95ULL); random_engines[16] = xorshift_128_plus(~0xA6E9377DAF75BDFEULL, 0x863F5CB508510D95ULL); random_engines[24] = xorshift_128_plus(~0xA6E9377DAF75BDFEULL, ~0x863F5CB508510D95ULL); random_engines[32] = xorshift_128_plus(0xA6E9377DAF75BDFEULL, 0x863F5CB508510D95ULL); random_engines[40] = xorshift_128_plus(0xA6E9377DAF75BDFEULL, ~0x863F5CB508510D95ULL); random_engines[48] = xorshift_128_plus(~0xA6E9377DAF75BDFEULL, 0x863F5CB508510D95ULL); random_engines[56] = xorshift_128_plus(~0xA6E9377DAF75BDFEULL, ~0x863F5CB508510D95ULL); //validation_results.push_back(std::async(std::launch::async, [&random_engines]() -> std::string { return validate_float_distribution(rand_float_co<xorshift_128_plus>, random_engines[0], 10, 100000000, true, false, false); })); //validation_results.push_back(std::async(std::launch::async, [&random_engines]() -> std::string { return validate_float_distribution(rand_float_oc<xorshift_128_plus>, random_engines[8], 10, 100000000, false, true, false); })); //validation_results.push_back(std::async(std::launch::async, [&random_engines]() -> std::string { return validate_float_distribution(rand_float_oo<xorshift_128_plus>, random_engines[16], 10, 100000000, false, false, false); })); //validation_results.push_back(std::async(std::launch::async, [&random_engines]() -> std::string { return validate_float_distribution(rand_float_cc<xorshift_128_plus>, random_engines[24], 10, 100000000, true, true, false); })); //validation_results.push_back(std::async(std::launch::async, [&random_engines]() -> std::string { return validate_float_distribution(rand_float_co_cast<xorshift_128_plus>, random_engines[32], 10, 100000000, true, false, true); })); //validation_results.push_back(std::async(std::launch::async, [&random_engines]() -> std::string { return validate_float_distribution(rand_float_oc_cast<xorshift_128_plus>, random_engines[40], 10, 100000000, false, true, true); })); //validation_results.push_back(std::async(std::launch::async, [&random_engines]() -> std::string { return validate_float_distribution(rand_float_oo_cast<xorshift_128_plus>, random_engines[48], 10, 100000000, false, false, true); })); //validation_results.push_back(std::async(std::launch::async, [&random_engines]() -> std::string { return validate_float_distribution(rand_float_cc_cast<xorshift_128_plus>, random_engines[56], 10, 100000000, true, true, true); })); bool ready; do { ready = true; for (auto&& validation_result : validation_results) { if (validation_result.wait_for(std::chrono::milliseconds(100)) != std::future_status::ready) { ready = false; break; } } } while (!ready); for (auto&& validation_result : validation_results) { std::cout << validation_result.get(); } // Use these functions to measure performance. std::cout << "Press any key to begin." << std::endl; std::cin.get(); prepare_thread(); double baseline; auto random_engine = xorshift_128_plus(0xA6E9377DAF75BDFEULL, 0x863F5CB508510D95ULL); std::ranlux24_base random_engine_32; std::ranlux48_base random_engine_64; std::uniform_real<float> flt_co; std::uniform_real<double> dbl_co; double warmup_duration = 1.0; double measure_duration = 60.0; baseline = measure_operations([&random_engine]() -> uint32_t { return random_engine.next32(); }, warmup_duration, measure_duration, "rand_next32"); std::cout << std::endl; measure_operations([&random_engine]() -> float { return rand_float_co(random_engine); }, warmup_duration, measure_duration, "rand_float_co", baseline); measure_operations([&random_engine]() -> float { return rand_float_oc(random_engine); }, warmup_duration, measure_duration, "rand_float_oc", baseline); measure_operations([&random_engine]() -> float { return rand_float_oo(random_engine); }, warmup_duration, measure_duration, "rand_float_oo", baseline); measure_operations([&random_engine]() -> float { return rand_float_cc(random_engine); }, warmup_duration, measure_duration, "rand_float_cc", baseline); std::cout << std::endl; measure_operations([&random_engine]() -> float { return rand_float_co_cast(random_engine); }, warmup_duration, measure_duration, "rand_float_co_cast", baseline); measure_operations([&random_engine]() -> float { return rand_float_oc_cast(random_engine); }, warmup_duration, measure_duration, "rand_float_oc_cast", baseline); measure_operations([&random_engine]() -> float { return rand_float_oo_cast(random_engine); }, warmup_duration, measure_duration, "rand_float_oo_cast", baseline); measure_operations([&random_engine]() -> float { return rand_float_cc_cast(random_engine); }, warmup_duration, measure_duration, "rand_float_cc_cast", baseline); std::cout << std::endl; measure_operations([&random_engine]() -> float { return rand_float_co_div(random_engine); }, warmup_duration, measure_duration, "rand_float_co_div", baseline); measure_operations([&random_engine]() -> float { return rand_float_oc_div(random_engine); }, warmup_duration, measure_duration, "rand_float_oc_div", baseline); measure_operations([&random_engine]() -> float { return rand_float_oo_div(random_engine); }, warmup_duration, measure_duration, "rand_float_oo_div", baseline); measure_operations([&random_engine]() -> float { return rand_float_cc_div(random_engine); }, warmup_duration, measure_duration, "rand_float_cc_div", baseline); std::cout << std::endl; baseline = measure_operations([&]() -> uint32_t { return random_engine_32(); }, warmup_duration, measure_duration, "ranlux24"); measure_operations([&random_engine_32, flt_co]() -> float { return flt_co(random_engine_32); }, warmup_duration, measure_duration, "flt_co(ranlux24)", baseline); std::cout << std::endl; std::cout << std::endl; std::cout << std::endl; baseline = measure_operations([&random_engine]() -> uint64_t { return random_engine.next64(); }, warmup_duration, measure_duration, "rand_next64"); std::cout << std::endl; measure_operations([&random_engine]() -> double { return rand_double_co(random_engine); }, warmup_duration, measure_duration, "rand_double_co", baseline); measure_operations([&random_engine]() -> double { return rand_double_oc(random_engine); }, warmup_duration, measure_duration, "rand_double_oc", baseline); measure_operations([&random_engine]() -> double { return rand_double_oo(random_engine); }, warmup_duration, measure_duration, "rand_double_oo", baseline); measure_operations([&random_engine]() -> double { return rand_double_cc(random_engine); }, warmup_duration, measure_duration, "rand_double_cc", baseline); std::cout << std::endl; measure_operations([&random_engine]() -> double { return rand_double_co_cast(random_engine); }, warmup_duration, measure_duration, "rand_double_co_cast", baseline); measure_operations([&random_engine]() -> double { return rand_double_oc_cast(random_engine); }, warmup_duration, measure_duration, "rand_double_oc_cast", baseline); measure_operations([&random_engine]() -> double { return rand_double_oo_cast(random_engine); }, warmup_duration, measure_duration, "rand_double_oo_cast", baseline); measure_operations([&random_engine]() -> double { return rand_double_cc_cast(random_engine); }, warmup_duration, measure_duration, "rand_double_cc_cast", baseline); std::cout << std::endl; measure_operations([&random_engine]() -> double { return rand_double_co_div(random_engine); }, warmup_duration, measure_duration, "rand_double_co_div", baseline); measure_operations([&random_engine]() -> double { return rand_double_oc_div(random_engine); }, warmup_duration, measure_duration, "rand_double_oc_div", baseline); measure_operations([&random_engine]() -> double { return rand_double_oo_div(random_engine); }, warmup_duration, measure_duration, "rand_double_oo_div", baseline); measure_operations([&random_engine]() -> double { return rand_double_cc_div(random_engine); }, warmup_duration, measure_duration, "rand_double_cc_div", baseline); std::cout << std::endl; baseline = measure_operations([&]() -> uint64_t { return random_engine_64(); }, warmup_duration, measure_duration, "ranlux48"); measure_operations([&random_engine_64, dbl_co]() -> double { return dbl_co(random_engine_64); }, warmup_duration, measure_duration, "dbl_co(ranlux48)", baseline); std::cout << std::endl; std::cout << std::endl; std::cout << std::endl; std::cout << "Measurements complete." << std::endl; std::cin.get(); return 0; } |

## 194 Comments

## Comment by Sebastiano Vigna — 2017/06/07 @ 05:20

I’m trying to replicate your tests as my results are very different—the division method is only slower 10% on modern CPUs. But I get dozens of errors when trying to compile with gcc your source code. You you publish the exact compilation line, or provide a makefile? Thanks!

## Trackback by fake van cleef style necklace — 2017/08/27 @ 06:38

fake van cleef style necklaceI always used to read post in news papers but now as I am a user of internet thus from now I am using net for posts, thanks to web.|

## Trackback by URL — 2017/12/14 @ 11:43

… [Trackback][…] Read More Infos here: experilous.com/1/blog/post/perfect-fast-random-floating-point-numbers […]

## Trackback by iraqi Aws law — 2017/12/26 @ 01:55

… [Trackback][…] Read More here|Read More|Read More Informations here|Here you can find 96155 additional Informations|Informations on that Topic: experilous.com/1/blog/post/perfect-fast-random-floating-point-numbers […]

## Trackback by freelance web developer — 2017/12/27 @ 14:33

… [Trackback][…] Read More here|Read More|Find More Informations here|There you will find 70268 additional Informations|Infos on that Topic: experilous.com/1/blog/post/perfect-fast-random-floating-point-numbers […]

## Trackback by arab Ceohuman — 2017/12/27 @ 18:05

… [Trackback][…] Find More on|Find More|Find More Infos here|There you can find 64429 more Infos|Infos on that Topic: experilous.com/1/blog/post/perfect-fast-random-floating-point-numbers […]

## Trackback by house for sale — 2018/01/14 @ 00:53

… [Trackback][…] Find More on|Find More|Read More Informations here|There you will find 47267 additional Informations|Infos to that Topic: experilous.com/1/blog/post/perfect-fast-random-floating-point-numbers […]

## Trackback by GVK BIO Latest News — 2018/01/19 @ 05:47

… [Trackback][…] Find More on|Find More|Read More Informations here|Here you can find 58502 more Informations|Informations on that Topic: experilous.com/1/blog/post/perfect-fast-random-floating-point-numbers […]

## Trackback by cpns 2018 umum — 2018/01/27 @ 11:06

… [Trackback][…] Find More on|Find More|Find More Informations here|Here you will find 79057 additional Informations|Informations on that Topic: experilous.com/1/blog/post/perfect-fast-random-floating-point-numbers […]

## Trackback by how to make money with a iphone — 2018/01/29 @ 18:25

… [Trackback][…] Find More on|Find More|Find More Infos here|There you can find 12921 more Infos|Informations to that Topic: experilous.com/1/blog/post/perfect-fast-random-floating-point-numbers […]

## Trackback by iq option review — 2018/02/03 @ 17:14

… [Trackback][…] Find More here|Find More|Read More Infos here|There you can find 25274 additional Infos|Infos to that Topic: experilous.com/1/blog/post/perfect-fast-random-floating-point-numbers […]

## Trackback by آنتی ویروس — 2018/02/04 @ 17:21

… [Trackback][…] Find More on|Find More|Read More Informations here|Here you will find 11087 more Informations|Informations to that Topic: experilous.com/1/blog/post/perfect-fast-random-floating-point-numbers […]

## Trackback by click here for more — 2018/02/05 @ 08:03

… [Trackback][…] Find More on|Find More|Read More Infos here|There you will find 58073 more Infos|Informations on that Topic: experilous.com/1/blog/post/perfect-fast-random-floating-point-numbers […]

## Trackback by Free UK Chat — 2018/02/05 @ 15:46

… [Trackback][…] Read More on|Read More|Read More Infos here|There you will find 20119 more Infos|Informations to that Topic: experilous.com/1/blog/post/perfect-fast-random-floating-point-numbers […]

## Trackback by pur bola — 2018/02/06 @ 17:10

… [Trackback][…] Find More here|Find More|Find More Informations here|There you can find 99269 additional Informations|Informations on that Topic: experilous.com/1/blog/post/perfect-fast-random-floating-point-numbers […]

## Trackback by EDM Jana Gana Mana - Indian National Anthem — 2018/02/06 @ 17:17

… [Trackback][…] Read More here|Read More|Read More Informations here|There you will find 16319 additional Informations|Informations to that Topic: experilous.com/1/blog/post/perfect-fast-random-floating-point-numbers […]

## Trackback by emergency keysmith near me richmond va — 2018/02/06 @ 17:17

… [Trackback][…] Find More on|Find More|Find More Informations here|Here you will find 55816 additional Informations|Informations on that Topic: experilous.com/1/blog/post/perfect-fast-random-floating-point-numbers […]

## Trackback by guaranteedppc — 2018/02/07 @ 05:03

… [Trackback][…] Find More on|Find More|Read More Infos here|Here you can find 32208 more Infos|Infos to that Topic: experilous.com/1/blog/post/perfect-fast-random-floating-point-numbers […]

## Trackback by emergency locked out of house richmond va — 2018/02/10 @ 17:16

… [Trackback][…] Read More on|Read More|Read More Informations here|There you can find 29672 additional Informations|Informations to that Topic: experilous.com/1/blog/post/perfect-fast-random-floating-point-numbers […]

## Trackback by https://www.gkandjobs.in/ — 2018/02/10 @ 17:28

… [Trackback][…] Find More on|Find More|Find More Informations here|Here you will find 79052 additional Informations|Informations on that Topic: experilous.com/1/blog/post/perfect-fast-random-floating-point-numbers […]

## Trackback by jayne herring — 2018/02/11 @ 23:51

… [Trackback][…] Find More here|Find More|Read More Infos here|Here you will find 43687 more Infos|Infos to that Topic: experilous.com/1/blog/post/perfect-fast-random-floating-point-numbers […]

## Trackback by discover this — 2018/02/12 @ 12:54

… [Trackback][…] Read More on|Read More|Find More Informations here|Here you will find 52561 more Informations|Informations on that Topic: experilous.com/1/blog/post/perfect-fast-random-floating-point-numbers […]

## Trackback by online spirit wear apparel store — 2018/02/16 @ 17:14

… [Trackback][…] Find More on|Find More|Read More Informations here|There you will find 9205 more Informations|Infos to that Topic: experilous.com/1/blog/post/perfect-fast-random-floating-point-numbers […]

## Trackback by morar em portugal — 2018/02/16 @ 17:17

… [Trackback][…] Read More here|Read More|Find More Infos here|There you will find 88393 additional Infos|Infos on that Topic: experilous.com/1/blog/post/perfect-fast-random-floating-point-numbers […]

## Trackback by 먹튀검증 — 2018/02/18 @ 00:00

… [Trackback][…] Read More on|Read More|Read More Infos here|There you will find 32752 additional Infos|Informations to that Topic: experilous.com/1/blog/post/perfect-fast-random-floating-point-numbers […]

## Comment by Allen Downey — 2018/02/20 @ 11:55

Looks good. I think it is similar to a method I proposed here: http://allendowney.com/research/rand/

## Trackback by casino online news — 2018/02/21 @ 17:36

… [Trackback][…] Read More here|Read More|Find More Informations here|There you will find 73677 more Informations|Informations on that Topic: experilous.com/1/blog/post/perfect-fast-random-floating-point-numbers […]

## Trackback by Asbestos Testing in Logan City — 2018/02/22 @ 02:33

… [Trackback][…] Find More here|Find More|Read More Informations here|Here you can find 95123 more Informations|Informations on that Topic: experilous.com/1/blog/post/perfect-fast-random-floating-point-numbers […]

## Trackback by http://guaranteedppc.com — 2018/02/23 @ 08:31

… [Trackback][…] Find More on|Find More|Read More Informations here|Here you can find 67339 more Informations|Informations on that Topic: experilous.com/1/blog/post/perfect-fast-random-floating-point-numbers […]

## Trackback by http://www.guaranteedppc.com — 2018/02/23 @ 08:50

… [Trackback][…] Find More on|Find More|Find More Infos here|There you will find 97776 more Infos|Informations on that Topic: experilous.com/1/blog/post/perfect-fast-random-floating-point-numbers […]

## Trackback by stumbleupon ads — 2018/02/23 @ 09:15

… [Trackback][…] Find More here|Find More|Find More Infos here|Here you can find 29977 more Infos|Infos to that Topic: experilous.com/1/blog/post/perfect-fast-random-floating-point-numbers […]

## Trackback by pay per click — 2018/02/23 @ 09:48

… [Trackback][…] Read More here|Read More|Find More Informations here|There you can find 16110 more Informations|Informations to that Topic: experilous.com/1/blog/post/perfect-fast-random-floating-point-numbers […]

## Trackback by Pune escorts service — 2018/02/26 @ 09:11

… [Trackback][…] Read More on|Read More|Read More Infos here|Here you can find 38379 more Infos|Informations to that Topic: experilous.com/1/blog/post/perfect-fast-random-floating-point-numbers […]

## Trackback by seo salt lake city — 2018/02/26 @ 21:45

… [Trackback][…] Read More on|Read More|Find More Infos here|Here you will find 37537 more Infos|Informations to that Topic: experilous.com/1/blog/post/perfect-fast-random-floating-point-numbers […]

## Trackback by treppenhausreinigung-hamburg — 2018/02/28 @ 10:58

… [Trackback][…] Read More on|Read More|Find More Informations here|Here you will find 16338 additional Informations|Informations to that Topic: experilous.com/1/blog/post/perfect-fast-random-floating-point-numbers […]

## Trackback by visit my homepage — 2018/03/03 @ 04:13

… [Trackback][…] Read More on|Read More|Find More Informations here|Here you will find 34363 more Informations|Informations to that Topic: experilous.com/1/blog/post/perfect-fast-random-floating-point-numbers […]

## Trackback by servicos informatica — 2018/03/07 @ 10:22

… [Trackback][…] Find More on|Find More|Find More Informations here|There you will find 35244 additional Informations|Infos to that Topic: experilous.com/1/blog/post/perfect-fast-random-floating-point-numbers […]

## Trackback by essayforme — 2018/05/27 @ 12:09

write an essay for me http://gsgfsdgfdhjhjhj.com/Reliable advice. Thanks a lot!

## Trackback by cartier love anello falso — 2018/06/03 @ 01:34

cartier love anello falsoThene there are the chemicals used to make plastic. Many of these are toxic and can leach out. Research is showing that chemicals absorbed by the plastic are transferred to the fish.

## Trackback by fotballdrakter barn — 2018/11/03 @ 22:03

fotballdrakter barnjustGoogle.com 9a6rq3 wdkin1 fotbollströjor eafxbc a5lo8r2kvc\n 7u3abm ye76ocmg fotbollströjor barn fcm3zt57kg hte7lmgd1k\n l26w17 y5n9ruled fotballdrakter barn 8su952m0la 45v0p2nwi\n

## Trackback by jordan 12 — 2019/05/06 @ 13:27

jordan 12Thank you for all of the work on this site. My aunt take interest in getting into investigation and it is easy to see why. A number of us know all relating to the lively form you provide priceless secrets on your website and even increase response from…

## Trackback by jordan 12 — 2019/05/13 @ 15:22

jordan 12I must show some thanks to this writer for bailing me out of this type of setting. After exploring through the online world and seeing things which are not pleasant, I figured my entire life was done. Existing devoid of the solutions to the difficultie…

## Trackback by 먹튀검증 — 2019/06/01 @ 21:01

먹튀Experilous: Perfect and Fast Random Floating Point Numbers

## Trackback by URL — 2019/08/15 @ 07:13

… [Trackback][…] Informations on that Topic: experilous.com/1/blog/post/perfect-fast-random-floating-point-numbers/trackback […]

## Trackback by 먹튀검증 — 2019/08/15 @ 18:33

먹튀검증Experilous: Perfect and Fast Random Floating Point Numbers

## Trackback by jordan 11 — 2019/09/14 @ 03:01

jordan 11My spouse and i have been absolutely ecstatic Ervin managed to finish up his investigation from the precious recommendations he gained out of the web page. It is now and again perplexing just to find yourself offering concepts which usually some others…

## Trackback by jordan 6 — 2019/09/16 @ 13:55

jordan 6My spouse and i were fulfilled when Emmanuel could finish off his inquiry out of the precious recommendations he acquired through the weblog. It’s not at all simplistic just to happen to be giving away strategies that some others have been making mone…

## Trackback by deltasone 10mg — 2020/03/11 @ 17:51

deltasone 10mgExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by chloroquine phosphate — 2020/03/29 @ 11:28

chloroquine phosphateExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by ciprofloxacin hydrochloride tablet — 2020/04/24 @ 19:39

ciprofloxacin hydrochloride tabletExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by naltrexone canada — 2020/04/29 @ 09:23

naltrexone canadaExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by walmart tylenol 500 mg — 2020/05/08 @ 06:03

walmart tylenol 500 mgExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by buy chloroquine — 2020/05/08 @ 20:46

buy chloroquineExperilous: Perfect and Fast Random Floating Point Numbers

## Comment by Data entry jobs — 2020/05/19 @ 20:50

Good content in our site and also beautiful blog

## Comment by Data entry jobs — 2020/05/19 @ 20:50

Good content in our site

## Trackback by buy hydroxychloroquine online uk — 2020/06/02 @ 20:41

buy hydroxychloroquine online ukExperilous: Perfect and Fast Random Floating Point Numbers

## Comment by sakshi khera — 2020/06/05 @ 05:36

Thanks for the post! Visit our blog Data entry jobs

## Comment by sanjay singh — 2020/06/30 @ 04:06

RKSA Packers & Movers Company offers reliable and hassle-free relocation services at an affordable rate. Contact us or read more information on our website. https://bit.ly/3hv26os

## Comment by London Kids — 2020/06/30 @ 07:09

London Kids Preschool Franchise is the best place for your children’s education, so contact us today for your children’s admission. https://bit.ly/3dNyM9R

## Trackback by bimatoprost ophthalmic solution 0 03 % — 2020/07/13 @ 10:23

bimatoprost ophthalmic solution 0 03 %Experilous: Perfect and Fast Random Floating Point Numbers

## Trackback by ed meds online without doctor prescription — 2020/07/14 @ 06:12

ed meds online without doctor prescriptionExperilous: Perfect and Fast Random Floating Point Numbers

## Comment by Braiding Shop — 2020/07/20 @ 06:38

Your article is very nice and informative. Thanks for sharing it!

## Comment by Beauty Blog — 2020/07/24 @ 02:57

Hello, This is Priyansh Shrivastava, Your article is very Nice, thanks for sharing… Also Check https://sauvewomen.com/beauty-guest-post/

## Comment by Beauty Blog — 2020/07/24 @ 07:11

Thanks for sharing such information. This is really helpful for me. You can also visit blog. Beauty Blog

## Comment by Priyanka shroti — 2020/07/25 @ 00:12

This article is very nice thanks for sharing this blog Beauty blog.

## Comment by Priyanka shroti — 2020/07/25 @ 00:13

Hii

This is Priyanka shroti

Nice article thanks for sharing this article Beauty blog.

## Comment by Parul Chourasiya — 2020/07/25 @ 01:36

Hey I’m Parul Chourasiya

This is very good article.

Visit our blog too https://www.google.com/url?q=https://sauvewomen.com/beauty-guest-post/&sa=D&source=hangouts&ust=1595743292232000&usg=AFQjCNHk-tukOSZ8q9WRRh8YqMZnlzUEPQ

## Comment by JEFFREY PAMTHIED — 2020/07/25 @ 03:49

If you’re searching for Beauty Blog then please apply here.

## Comment by sudeep — 2020/07/25 @ 04:57

Your article is very nice & informative . Thanks for sharing it. if you search for Beauty Blogthen please apply hear.

## Comment by Harsh — 2020/07/25 @ 05:25

Your website is so amazing

Visit my blog

“Beauty Blog “

## Comment by Beauty blog — 2020/07/27 @ 07:38

Hii

This is Priyanka shroti nice article thanks for sharing this article also check

https://sauvewomen.com/beauty-guest-post/

## Comment by Preeti Advani — 2020/08/07 @ 02:37

This article is amazing.Thanks foe sharing.Visit our blog too Jobs for Pregnant Women

## Trackback by buy chloroquine phosphate — 2020/08/08 @ 14:20

buy chloroquine phosphateExperilous: Perfect and Fast Random Floating Point Numbers

## Comment by Fashion write for us — 2020/08/17 @ 02:26

Hii,This is priyanka shroti and it’s my first comment here so I just wanted to give a quick shout out and say I genuinely enjoy reading your blog post can you recommended by other Fashion Write For Usblogs that go over the same topics ? Thanks a ton !

## Comment by Beauty guest post — 2020/08/18 @ 01:54

Hii,This is priyanka shroti and it’s my first comment here so I just wanted to give a quick shout out and say I genuinely enjoy reading your blog post can you recommended by other Beauty Guest Post blogs that go over the same topics ? Thanks a ton !

## Trackback by 롤듀오 롤스터디 — 2020/09/27 @ 13:53

롤듀오 롤스터디Experilous: Perfect and Fast Random Floating Point Numbers

## Comment by jackdanial219 — 2021/03/31 @ 11:33

May I put part of this on my page if I include a reference to this web page?

Hi dude, you may have a great blog site. I enjoy it. And I feel several folks will enjoy too. Why not attempt more challenging to optimize your site, so it https://www.seo4tricks.com

## Trackback by uspharmus.com — 2021/04/27 @ 19:52

online pharmacy without a prescriptionExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by canadianeve21.com — 2021/04/30 @ 05:00

canada prescription plus pharmacyExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by xlnpharmacy.com — 2021/05/08 @ 10:38

best online pharmacy that does not require a prescriptionExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by pharmacylo.com — 2021/05/08 @ 18:30

pharmacy technician online programExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by pharmacyhrn.com — 2021/05/21 @ 10:45

mail order prescription drugsExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by pharmacyken.com — 2021/05/21 @ 13:23

online pharmacies in usaExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by cjepharmacy.com — 2021/05/22 @ 06:27

online pharmacy no scriptsExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by Samridhisocialhelp official website — 2021/06/16 @ 12:23

Samridhisocialhelp official websiteExperilous: Perfect and Fast Random Floating Point Numbers

## Comment by 안전놀이터 — 2021/06/17 @ 04:24

Greetings! Very helpful advice in this particular article! It’s the little changes which will make the largest changes. Thanks for sharing 안전놀이터

## Comment by 우리카지노사이트 — 2021/07/01 @ 23:59

https://www.ekffo150.com 우리카지노사이트

## Comment by 더킹카지노 — 2021/07/02 @ 00:16

https://www.ajp4949.com/theking 더킹카지노

## Comment by 메리트카지노 — 2021/07/02 @ 00:21

https://www.ajp4949.com/merit 메리트카지노

## Comment by 샌즈카지노 — 2021/07/02 @ 00:25

https://www.ajp4949.com/sands 샌즈카지노

## Comment by 퍼스트카지노 — 2021/07/02 @ 00:46

https://www.ajp4949.com/first 퍼스트카지노

## Comment by 33카지노 — 2021/07/02 @ 00:52

https://www.ajp4949.com/33casino 33카지노

## Comment by 월드카지노 — 2021/07/02 @ 00:59

https://www.ajp4949.com/worldcasino 월드카지노

## Comment by 코리아카지노 — 2021/07/02 @ 01:06

https://www.ajp4949.com/korea 코리아카지노

## Trackback by storpharmon.com — 2021/08/04 @ 19:28

abused prescription drugsExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by pharmpharms.com — 2021/08/05 @ 04:12

canadian online pharmacies legitimate by aarpExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by inpharmxx.com — 2021/08/05 @ 13:13

recreditable canadian pharmaciesExperilous: Perfect and Fast Random Floating Point Numbers

## Comment by Kèonhàcái — 2021/08/06 @ 05:32

Your article was very impressive to me. It was unexpected information,but after reading it like this Kèonhàcái, I found it very interesting.

## Comment by Majortotosite Pro — 2021/08/06 @ 23:33

One of our visitors recently recommended the following website. 토토

## Comment by Racesite Pro — 2021/08/06 @ 23:33

Very informative blog article. Thanks Again. Keep writing. 카지노사이트

## Comment by Oncasinosite Net — 2021/08/06 @ 23:33

Some truly superb blog posts on this website , thanks for contribution. 온라인경마

## Comment by Totopick Pro — 2021/08/06 @ 23:34

Your style is very unique in comparison to other folks I ave read stuff from. Many thanks for posting when you have the opportunity, Guess I all just book mark this page. 토토사이트

## Trackback by Borgata Casino Online Support — 2021/08/13 @ 04:37

Borgata Casino Online SupportExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by resorts casino online Customer service — 2021/08/13 @ 04:48

resorts casino online Customer serviceExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by https://billionairesclub.entrepreneurdailydigest.com/ — 2021/08/13 @ 04:50

https://billionairesclub.entrepreneurdailydigest.com/Experilous: Perfect and Fast Random Floating Point Numbers

## Trackback by play live casino Online — 2021/08/13 @ 04:51

play live casino OnlineExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by Md live casino online — 2021/08/13 @ 05:31

Md live casino onlineExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by view — 2021/08/13 @ 05:32

viewExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by casino online no deposit — 2021/08/13 @ 05:32

casino online no depositExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by Https://ohorganization.com/2021/07/26/sports-betting-online-top-tips-on-how-to-bet-wisely-online — 2021/08/13 @ 05:46

https://ohorganization.com/2021/07/26/sports-betting-online-top-tips-on-how-to-bet-wisely-onlineExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by casino Online slots vs — 2021/08/13 @ 05:52

casino Online slots vsExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by Borgata casino Online — 2021/08/13 @ 05:56

Borgata casino OnlineExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by 스포츠토토 — 2021/08/13 @ 05:56

스포츠토토Experilous: Perfect and Fast Random Floating Point Numbers

## Trackback by 승인전화없는 토토사이트 — 2021/08/13 @ 06:13

승인전화없는 토토사이트Experilous: Perfect and Fast Random Floating Point Numbers

## Trackback by tropicana Casino online — 2021/08/13 @ 06:17

tropicana Casino onlineExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by casino online free Play — 2021/08/13 @ 07:01

casino online free PlayExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by just click the up coming page — 2021/08/13 @ 07:28

just click the up coming pageExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by 사설토토사이트 — 2021/08/13 @ 07:29

사설토토사이트Experilous: Perfect and Fast Random Floating Point Numbers

## Trackback by Wwii-Archives.org — 2021/08/13 @ 08:04

Wwii-Archives.orgExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by Http://Desf.Or.Kr/Board/Index.Php?Mid=M92&Document_Srl=2745838 — 2021/08/13 @ 08:23

http://Desf.Or.Kr/Board/Index.Php?Mid=M92&Document_Srl=2745838Experilous: Perfect and Fast Random Floating Point Numbers

## Trackback by Resorts Casino Online — 2021/08/13 @ 08:24

Resorts Casino OnlineExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by Live Casino online slots — 2021/08/13 @ 08:24

Live Casino online slotsExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by casino online new free bonus codes usa — 2021/08/13 @ 09:30

casino online new free bonus codes usaExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by casino Online youtube — 2021/08/13 @ 09:57

casino Online youtubeExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by 사다리게임 — 2021/08/13 @ 10:36

사다리게임Experilous: Perfect and Fast Random Floating Point Numbers

## Trackback by 승인전화없는사이트 — 2021/08/13 @ 11:16

승인전화없는사이트Experilous: Perfect and Fast Random Floating Point Numbers

## Trackback by Casino Online No Deposit Bonus Codes — 2021/08/13 @ 11:21

Casino Online No Deposit Bonus CodesExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by http://Seosanmarket.Co.Kr/?document_srl=3786862 — 2021/08/13 @ 12:10

http://Seosanmarket.Co.Kr/?document_srl=3786862Experilous: Perfect and Fast Random Floating Point Numbers

## Trackback by tropicana Casino Online Slots — 2021/08/13 @ 12:12

tropicana Casino Online SlotsExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by borgata casino Online reviews — 2021/08/13 @ 12:28

borgata casino Online reviewsExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by borgata casino online betting — 2021/08/13 @ 13:33

borgata casino online bettingExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by Tropicana casino online review — 2021/08/13 @ 13:36

Tropicana casino online reviewExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by slots of vegas casino Online — 2021/08/13 @ 13:51

slots of vegas casino OnlineExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by casino online usa player no Deposit codes — 2021/08/13 @ 14:22

casino online usa player no Deposit codesExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by https://www.gupshuptoinspire.com/groups/casino-guide-for-washington-state-east-of-the-cascade-mountains-1401244439 — 2021/08/13 @ 14:38

https://www.gupshuptoinspire.com/groups/casino-guide-for-washington-state-east-of-the-cascade-mountains-1401244439Experilous: Perfect and Fast Random Floating Point Numbers

## Trackback by casino online Gclub royal — 2021/08/13 @ 15:51

casino online Gclub royalExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by boind.co.kr — 2021/08/13 @ 17:37

boind.co.krExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by Https://Meisstudio.gr/ — 2021/08/13 @ 18:55

https://Meisstudio.gr/Experilous: Perfect and Fast Random Floating Point Numbers

## Trackback by http://Innoeduvation.Org/ — 2021/08/13 @ 23:10

http://Innoeduvation.Org/Experilous: Perfect and Fast Random Floating Point Numbers

## Trackback by casino Online free credit — 2021/08/13 @ 23:36

casino Online free creditExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by My Site — 2021/08/14 @ 01:10

My SiteExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by try these guys out — 2021/08/14 @ 04:39

try these guys outExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by from the Wwii-archives.org blog — 2021/08/14 @ 05:42

from the Wwii-archives.org blogExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by Casino Online Slots Free Play — 2021/08/14 @ 06:38

Casino Online Slots Free PlayExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by Play Resorts Casino Online — 2021/08/14 @ 06:39

Play Resorts Casino OnlineExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by Read This method — 2021/08/14 @ 07:10

Read This methodExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by mouse click the up coming web site — 2021/08/14 @ 07:45

mouse click the up coming web siteExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by 안전검증 — 2021/08/14 @ 10:23

안전검증Experilous: Perfect and Fast Random Floating Point Numbers

## Trackback by Www.mick0711.com — 2021/08/14 @ 10:57

http://Www.mick0711.comExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by borgata casino Online Promo code — 2021/08/14 @ 11:27

borgata casino Online Promo codeExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by click through the next internet site — 2021/08/17 @ 02:16

click through the next internet siteExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by Http://ttlink.com/Joellen774/all — 2021/08/17 @ 02:31

http://ttlink.com/Joellen774/allExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by Borgata Casino Online Download — 2021/08/17 @ 04:45

Borgata Casino Online DownloadExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by they said — 2021/08/17 @ 05:36

they saidExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by casino online no deposit sign up bonus — 2021/08/17 @ 05:52

casino online no deposit sign up bonusExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by vegas casino online free slots play — 2021/08/17 @ 07:49

vegas casino online free slots playExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by casino online slots free — 2021/08/17 @ 07:50

casino online slots freeExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by casino Online no deposit Bonus — 2021/08/17 @ 09:16

casino Online no deposit BonusExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by Play Caesars Casino Online — 2021/08/17 @ 09:41

Play Caesars Casino OnlineExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by https://Wwii-archives.org/ArcherjqCharlestonkg — 2021/08/20 @ 00:23

https://Wwii-archives.org/ArcherjqCharlestonkgExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by casino online free bonus No deposit — 2021/08/20 @ 01:31

casino online free bonus No depositExperilous: Perfect and Fast Random Floating Point Numbers

## Comment by UFA168 — 2021/08/22 @ 13:19

Wonderful article, thanks for putting this together! This is obviously one great post. Thanks for the valuable information and insights you have so provided here. ufa168

## Trackback by caesars Casino online free slots — 2021/08/24 @ 23:54

caesars Casino online free slotsExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by Casino online free slots — 2021/08/24 @ 23:57

Casino online free slotsExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by please click the following webpage — 2021/08/25 @ 00:17

please click the following webpageExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by Https://Ohorganization.com/ — 2021/08/25 @ 00:18

https://Ohorganization.com/Experilous: Perfect and Fast Random Floating Point Numbers

## Trackback by http://climateaction.Re.kr — 2021/08/25 @ 00:21

http://climateaction.Re.krExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by casino online slots free Play Apps — 2021/09/02 @ 18:53

casino online slots free Play AppsExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by norvascamlodipinegh.com — 2021/09/03 @ 13:19

amlodipine maximum dosageExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by iphone — 2021/09/04 @ 02:50

iphoneExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by lipitoratorvastatinfg.com — 2021/09/04 @ 10:33

lipitor drug interactionsExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by Casino online usa legal — 2021/09/06 @ 01:21

Casino online usa legalExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by prilosecomeprazolehu.com — 2021/09/06 @ 04:14

dexilant vs omeprazoleExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by http://www.techno-humanities.kr/member/260326 — 2021/09/06 @ 05:08

http://www.techno-humanities.kr/member/260326Experilous: Perfect and Fast Random Floating Point Numbers

## Trackback by lyricapregabalintv.com — 2021/09/09 @ 04:12

lyrica garrett ageExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by lets have a party — 2021/09/09 @ 08:53

lets have a partyExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by casino online slots apps — 2021/09/10 @ 00:48

casino online slots appsExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by lisinoprilhydrochlorothiazidegh.com — 2021/09/11 @ 14:46

hydrochlorothiazide other namesExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by lipitoratorvastatinla.com — 2021/09/12 @ 22:20

atorvastatin tabs 40mgExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by tour packages — 2021/09/14 @ 08:59

tour packagesExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by click the next web page — 2021/09/29 @ 11:05

click the next web pageExperilous: Perfect and Fast Random Floating Point Numbers

## Comment by sportstototop — 2021/10/11 @ 05:57

토토사이트 Hey There. I found your blog using msn. This is an extremely well written article.I will be sure to bookmark it and return to read more of your useful information. Thanks for the post

## Comment by toto365pro — 2021/10/11 @ 05:58

Greetings! I know this is kinda off topic however , I’d figured I’d ask.Would you be interested in trading links or maybe guest writing a blog post or vice-versa? My site addresses a lot of the same topics as yours and I believe we could greatly benefit from each other. 토토사이트

## Comment by casinositehomecom — 2021/10/11 @ 05:59

Hey There. I found your blog using msn. This is an extremely well written article.I will be sure to bookmark it and return to read more of your useful information. Thanks for the post 카지노

## Trackback by atstromectolaw.com — 2021/10/11 @ 11:08

ivermectin injectable for horsesExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by anstromectol4t.com — 2021/10/11 @ 12:18

ivermectin tablets 12mgExperilous: Perfect and Fast Random Floating Point Numbers

## Trackback by instromectol6y.com — 2021/10/11 @ 12:23

valley vet ivermectinExperilous: Perfect and Fast Random Floating Point Numbers

## Comment by 메이저사이트 — 2021/10/16 @ 02:37

First of all, thank you for your post. 메이저사이트 Your posts are neatly organized with the information I want, so there are plenty of resources to reference. I bookmark this site and will find your posts frequently in the future. Thanks again ^^

## Trackback by jam sholat mushola — 2021/10/17 @ 06:52

jam sholat musholaExperilous: Perfect and Fast Random Floating Point Numbers

## Comment by 토토사이트추천 — 2021/10/17 @ 22:40

Really no matter if someone doesn’t be aware of after that its up to other users that they will help, so here it takes place 토토사이트추천.

## Comment by ihihi99 — 2021/10/19 @ 11:45

I really loved reading your blog. It was very well authored and easy to understand. Unlike other blogs I have read which are really not that good.Thanks alot!

송송넷

## Trackback by garut — 2021/10/19 @ 17:14

garutExperilous: Perfect and Fast Random Floating Point Numbers

## Comment by 사설토토사이트 — 2021/10/19 @ 22:29

Exceptional post however , I was wanting to know if you could write a litte more on this topic? I’d be very thankful if you could elaborate a little bit further. Thanks 사설토토사이트

## Comment by oworldsmewep — 2021/10/20 @ 05:14

Wow, amazing blog layout! How long have you been blogging for?

you made blogging look easy. The overall look of your site is magnificent, let alone the content!

오피월드

oworldsmewep

## Leave a comment