I am trying to analyze how reduction (parallel) can be used to add a large array of floating point numbers and precision loss involved in it. Definitely reduction will help in getting more precision compared to serial addition . I'll be really thankful if you can direct me to some detailed source or provide some insight for this analysis. Thanks.
Every primitive floating point operation will have a rounding error; if the result is x then the rounding error is <= c * abs (x) for some rather small constant c > 0.
If you add 1000 numbers, that takes 999 additions. Each addition has a result and a rounding error. The rounding error is small when the result is small. So you want to adjust the order of additions so that the average absolute value of the result is as small as possible. A binary tree is one method. Sorting the values, then adding the smallest two numbers and putting the result back into the sorted list is also quite reasonable. Both methods keep the average result small, and therefore keep the rounding error small.
Related
I have a large array of floating point values that vary widely in magnitude. Does it help rescaling those in [0,1] for precision purposes (e.g. if I want to perform arithmetic operations on the array)? I can think of the smaller values getting truncated if I do so, but on the other hand small values will not contribute much to the absolute error. If I do the rescaling on an array of already computed values, I believe this can only make things worse as I would only introduce additional round-off error. On the other hand, I believe I can decrease the error if the scaling is instead involved at the point when I generate said values.
I am mainly referring to the fact that absolute distance between consecutive error grows 2 times for values in subsequent intervals (i.e. [0,1) vs [1,2) vs [2,4), etc.). Am I interpreting this correctly in the current context? I have seen such effect of floating point errors due to large scaling when trying to render a massively scaled 3D scene versus a less scaled version of it (similar effects occur when a camera in 3D space is too far from the origin, since absolute distances between floats become larger).
Considering the above, is there an optimal way to choose the scaling factor for an array of values I plan to generate (provided I know what the minimum and maximum will be without scaling). I was thinking of just generating it so that all values are within [0,1], however I was worried that the truncation of the smallest element may be an issue. Are there known heuristics based on the largest and smallest elements that allow to find a semi-optimal rescaling wrt precision. On an unrelated note, I am aware of the Kahan summation algorithm and its variants and I do use it for the summation of said array. My question is rather whether a choice of scale can help further, or will this not matter?
Scaling by powers of two in a binary floating-point format (or, generally, by powers of b in a base-b floating-point format) has no error as long as the results stay within normal exponent bounds. That is, for any x, the result of computing x•be has the same significand as x, as long as x•be is in the normal range.
Again as long as results stay in the normal range, the operations of adding, subtracting, multiplying, and dividing scaled numbers produce results with identical significands as the same operations on unscaled numbers that stay in the normal range. Any rounding errors that occur in the unscaled operations are identical to the rounding errors in the scaled operations, as adjusted by the scale.
Therefore, scaling numbers by a power of b, performing the same operations, and undoing the scaling will not improve or alter floating-point rounding errors. (Note that multiplications and divisions will affect the scaling, and this can be compensated for either after each operation, after all the operations, or periodically. For example, given X = x*16 and Y = y*16, X*Y would equal x*16*y*16 = x*y*256. So undoing its scaling requires dividing by 256 rather than 16.)
If other operations are used, the rounding errors may differ. For example, if a square root is performed and the scaling in its operand is not an even power of b, its result will include a scaling that is not an integral power of b, and so the significand must be different from the significand of the corresponding unscaled result, and that allows the rounding errors to be different.
Of course, if sines, cosines, or other trigonometric functions are used on scaled numbers, drastically different results will be obtained, as these functions do not scale in the required way (f(x•s) generally does not equal f(x)•s). However, if the numbers that are being scaled represent points in space, any angles computed between them would be identical in the scaled and unscaled implementations. That is, the computed angles would be free of scaling, and so applying trigonometric functions would produce identical results.
If any intermediate results exceed the normal exponent range in either the scaled or the unscaled computations, then different significands may be produced. This includes the case where the results are subnormal but have not underflowed to zero—subnormal results may have truncated signficands, so some information is lost compared to a differently-scaled computation that produces a result in the normal range.
An alternative to scaling may be translation. When working with points from the origin, the coordinates may be large, and the floating-point resolution may be large relative to distances between the points. If the points are translated to near the origin (a fixed amount is subtracted from each coordinate [fixed per dimension]), the geometric relationships between them are preserved, but the coordinates will be in a finer range of the floating-point format. This can improve the floating-point rounding errors that occur.
I'm just curious, why in IEEE-754 any non zero float number divided by zero results in infinite value? It's a nonsense from the mathematical perspective. So I think that correct result for this operation is NaN.
Function f(x) = 1/x is not defined when x=0, if x is a real number. For example, function sqrt is not defined for any negative number and sqrt(-1.0f) if IEEE-754 produces a NaN value. But 1.0f/0 is Inf.
But for some reason this is not the case in IEEE-754. There must be a reason for this, maybe some optimization or compatibility reasons.
So what's the point?
It's a nonsense from the mathematical perspective.
Yes. No. Sort of.
The thing is: Floating-point numbers are approximations. You want to use a wide range of exponents and a limited number of digits and get results which are not completely wrong. :)
The idea behind IEEE-754 is that every operation could trigger "traps" which indicate possible problems. They are
Illegal (senseless operation like sqrt of negative number)
Overflow (too big)
Underflow (too small)
Division by zero (The thing you do not like)
Inexact (This operation may give you wrong results because you are losing precision)
Now many people like scientists and engineers do not want to be bothered with writing trap routines. So Kahan, the inventor of IEEE-754, decided that every operation should also return a sensible default value if no trap routines exist.
They are
NaN for illegal values
signed infinities for Overflow
signed zeroes for Underflow
NaN for indeterminate results (0/0) and infinities for (x/0 x != 0)
normal operation result for Inexact
The thing is that in 99% of all cases zeroes are caused by underflow and therefore in 99%
of all times Infinity is "correct" even if wrong from a mathematical perspective.
I'm not sure why you would believe this to be nonsense.
The simplistic definition of a / b, at least for non-zero b, is the unique number of bs that has to be subtracted from a before you get to zero.
Expanding that to the case where b can be zero, the number that has to be subtracted from any non-zero number to get to zero is indeed infinite, because you'll never get to zero.
Another way to look at it is to talk in terms of limits. As a positive number n approaches zero, the expression 1 / n approaches "infinity". You'll notice I've quoted that word because I'm a firm believer in not propagating the delusion that infinity is actually a concrete number :-)
NaN is reserved for situations where the number cannot be represented (even approximately) by any other value (including the infinities), it is considered distinct from all those other values.
For example, 0 / 0 (using our simplistic definition above) can have any amount of bs subtracted from a to reach 0. Hence the result is indeterminate - it could be 1, 7, 42, 3.14159 or any other value.
Similarly things like the square root of a negative number, which has no value in the real plane used by IEEE754 (you have to go to the complex plane for that), cannot be represented.
In mathematics, division by zero is undefined because zero has no sign, therefore two results are equally possible, and exclusive: negative infinity or positive infinity (but not both).
In (most) computing, 0.0 has a sign. Therefore we know what direction we are approaching from, and what sign infinity would have. This is especially true when 0.0 represents a non-zero value too small to be expressed by the system, as it frequently the case.
The only time NaN would be appropriate is if the system knows with certainty that the denominator is truly, exactly zero. And it can't unless there is a special way to designate that, which would add overhead.
NOTE:
I re-wrote this following a valuable comment from #Cubic.
I think the correct answer to this has to come from calculus and the notion of limits. Consider the limit of f(x)/g(x) as x->0 under the assumption that g(0) == 0. There are two broad cases that are interesting here:
If f(0) != 0, then the limit as x->0 is either plus or minus infinity, or it's undefined. If g(x) takes both signs in the neighborhood of x==0, then the limit is undefined (left and right limits don't agree). If g(x) has only one sign near 0, however, the limit will be defined and be either positive or negative infinity. More on this later.
If f(0) == 0 as well, then the limit can be anything, including positive infinity, negative infinity, a finite number, or undefined.
In the second case, generally speaking, you cannot say anything at all. Arguably, in the second case NaN is the only viable answer.
Now in the first case, why choose one particular sign when either is possible or it might be undefined? As a practical matter, it gives you more flexibility in cases where you do know something about the sign of the denominator, at relatively little cost in the cases where you don't. You may have a formula, for example, where you know analytically that g(x) >= 0 for all x, say, for example, g(x) = x*x. In that case the limit is defined and it's infinity with sign equal to the sign of f(0). You might want to take advantage of that as a convenience in your code. In other cases, where you don't know anything about the sign of g, you cannot generally take advantage of it, but the cost here is just that you need to trap for a few extra cases - positive and negative infinity - in addition to NaN if you want to fully error check your code. There is some price there, but it's not large compared to the flexibility gained in other cases.
Why worry about general functions when the question was about "simple division"? One common reason is that if you're computing your numerator and denominator through other arithmetic operations, you accumulate round-off errors. The presence of those errors can be abstracted into the general formula format shown above. For example f(x) = x + e, where x is the analytically correct, exact answer, e represents the error from round-off, and f(x) is the floating point number that you actually have on the machine at execution.
Suppose we have a data set of numbers, with which we want to do some calculations using addition/subtraction/multiplication/division using a computer.
The coverage of the real numbers by the floating point representation varies a lot, depending on the number being represented:
In terms of absolute precision in the real->FP mapping the "holes" grow towards the bigger numbers, with a weird hole around 0, depending on the architecture. Due to this, the add/sub precision towards the bigger numbers will drop.
If we divide 2 consecutive numbers which are represented in our floating point representation, the result of the division will be bigger both while going to the bigger numbers and when going to smaller and smaller fractions.
So, my question is:
Is there a "sweet interval" for floats on an ordinary PC today, where the results for the arithmetics with the said operators (add/sub/mul/div) are just more precise?
If I have a data set of many-significant-digit numbers like "123123123123123", "134534513412351151", etc., with which I want to do some arithmetics, which floating point interval should it be converted to, to have the best precision for the result?
Since floating points are something like 1.xxx*10^yyy, 2.xxx*10^yyy, ..., 9.xxx*10^yyy, I would assume, converting my numbers into the [1, 9] interval would give the best results for the memory consumed, but I may be terribly wrong...
Suppose I use C, can such conversion even be made? Is there a best-practice to do that? Before an operation, C will convert the operands to the same format, so I guess I would have to use a string representation, inject a "." somewhere and parse that as float.
Please note:
This is a theoretical question, I don't have an actual data set on my hand that would decide what is best. On the same note, the mentioning of C was random, I am also interested in responses like "forget C, I would use this and this, BECAUSE it supports this and this".
Please spare me from answers like "this cannot be answered, because it depends on the actual operations, since the results may be in another magnitude range than the original data, etc., etc.". Let's suppose that the results of the calculation is more or less in the same interval, as the operands. Sure, when dividing the "more-or-less the same magnitude" operands, the result will be somewhere between 1-10, maybe 0.1-100, ... , but that is probably exactly the best interval they can be in.
Of course, if the answer includes some explanation, other than a brush-off, I will be happy to read it!
The absolute precision of floating-point numbers changes with the magnitude of the numbers because the exponent changes. The relative precision does not change, except for numbers near the bottom of the exponent range, where underflow occurs. If you multiply binary floating-point numbers by a power of two, perform arithmetic (suitably adjusted for the scaling), and reverse the scaling, the results will be identical to doing the arithmetic without scaling, barring effects from overflow and underflow. If your arithmetic does involve underflow or overflow, then scaling could help avoid that. For example, if your precision is suffering because your numbers are so small that some intermediate results are below the normal range of the floating-point format, then scaling by a power of two can avoid the loss of precision from underflow.
If you scale by something other than a power of two, the results can be different, due to changes in the significands. The effects will generally be tiny, and whether the results are better or worse will effectively be random chance, except in carefully engineered special situations.
I used AMD's two-stage reduction example to compute the sum of all numbers from 0 to 65 536 using floating point precision. Unfortunately, the result is not correct. However, when I modify my code, so that I compute the sum of 65 536 smaller numbers (for example 1), the result is correct.
I couldn't find any error in the code. Is it possible that I am getting wrong results, because of the float type? If this is the case, what is the best approach to solve the issue?
This is a "side effect" of summing floating point numbers using finite precision CPU's or GPU's. The accuracy depends the algorithm and the order the values are summed. The theory and practice behind is explained in Nicholas J, Higham's paper
The Accuracy of Floating Point Summation
http://citeseerx.ist.psu.edu/viewdoc/download;jsessionid=7AECC0D6458288CD6E4488AD63A33D5D?doi=10.1.1.43.3535&rep=rep1&type=pdf
The fix is to use a smarter algorithm like the Kahan Summation Algorithm
https://en.wikipedia.org/wiki/Kahan_summation_algorithm
And the Higham paper has some alternatives too.
This problem illustrates the nature of benchmarking, the first rule of the benchmark is to get the
right answer, using realistic data!
There is probably no error in the coding of your kernel or host application. The issue is with the single-precision floating point.
The correct sum is: 65537 * 32768 = 2147516416, and it takes 31 bits to represent it in binary (10000000000000001000000000000000). 32-bit floats can only hold integers accurately up to 2^24.
"Any integer with absolute value less than [2^24] can be exactly represented in the single precision format"
"Floating Point" article, wikipedia
This is why you are getting the correct sum when it is less than or equal to 2^24. If you are doing a complete sum using single-precision, you will eventually lose accuracy no matter which device you are executing the kernel on. There are a few things you can do to get the correct answer:
use double instead of float if your platform supports it
use int or unsigned int
sum a smaller set of numbers eg: 0+1+2+...+4095+4096 = (2^23 + 2^11)
Read more about single precision here.
I have been researching the log-sum-exp problem. I have a list of numbers stored as logarithms which I would like to sum and store in a logarithm.
the naive algorithm is
def naive(listOfLogs):
return math.log10(sum(10**x for x in listOfLogs))
many websites including:
logsumexp implementation in C?
and
http://machineintelligence.tumblr.com/post/4998477107/
recommend using
def recommend(listOfLogs):
maxLog = max(listOfLogs)
return maxLog + math.log10(sum(10**(x-maxLog) for x in listOfLogs))
aka
def recommend(listOfLogs):
maxLog = max(listOfLogs)
return maxLog + naive((x-maxLog) for x in listOfLogs)
what I don't understand is if recommended algorithm is better why should we call it recursively?
would that provide even more benefit?
def recursive(listOfLogs):
maxLog = max(listOfLogs)
return maxLog + recursive((x-maxLog) for x in listOfLogs)
while I'm asking are there other tricks to make this calculation more numerically stable?
Some background for others: when you're computing an expression of the following type directly
ln( exp(x_1) + exp(x_2) + ... )
you can run into two kinds of problems:
exp(x_i) can overflow (x_i is too big), resulting in numbers that you can't add together
exp(x_i) can underflow (x_i is too small), resulting in a bunch of zeroes
If all the values are big, or all are small, we can divide by some exp(const) and add const to the outside of the ln to get the same value. Thus if we can pick the right const, we can shift the values into some range to prevent overflow/underflow.
The OP's question is, why do we pick max(x_i) for this const instead of any other value? Why don't we recursively do this calculation, picking the max out of each subset and computing the logarithm repeatedly?
The answer: because it doesn't matter.
The reason? Let's say x_1 = 10 is big, and x_2 = -10 is small. (These numbers aren't even very large in magnitude, right?) The expression
ln( exp(10) + exp(-10) )
will give you a value very close to 10. If you don't believe me, go try it. In fact, in general, ln( exp(x_1) + exp(x_2) + ... ) will give be very close to max(x_i) if some particular x_i is much bigger than all the others. (As an aside, this functional form, asymptotically, actually lets you mathematically pick the maximum from a set of numbers.)
Hence, the reason we pick the max instead of any other value is because the smaller values will hardly affect the result. If they underflow, they would have been too small to affect the sum anyway, because it would be dominated by the largest number and anything close to it. In computing terms, the contribution of the small numbers will be less than an ulp after computing the ln. So there's no reason to waste time computing the expression for the smaller values recursively if they will be lost in your final result anyway.
If you wanted to be really persnickety about implementing this, you'd divide by exp(max(x_i) - some_constant) or so to 'center' the resulting values around 1 to avoid both overflow and underflow, and that might give you a few extra digits of precision in the result. But avoiding overflow is much more important about avoiding underflow, because the former determines the result and the latter doesn't, so it's much simpler just to do it this way.
Not really any better to do it recursively. The problem's just that you want to make sure your finite-precision arithmetic doesn't swamp the answer in noise. By dealing with the max on its own, you ensure that any junk is kept small in the final answer because the most significant component of it is guaranteed to get through.
Apologies for the waffly explanation. Try it with some numbers yourself (a sensible list to start with might be [1E-5,1E25,1E-5]) and see what happens to get a feel for it.
As you have defined it, your recursive function will never terminate. That's because ((x-maxlog) for x in listOfLogs) still has the same number of elements as listOfLogs.
I don't think that this is easily fixable either, without significantly impacting either the performance or the precision (compared to the non-recursive version).