Computing RD(sqrt(x)) with a FPU in RU mode - c

Intervals of floating-point bounds can be used to over-approximate sets of reals, as long as the upper bound of any result interval is computed in round-upwards and the lower bound in round-downwards.
One recommended trick is to actually compute the negation of the lower bound. This allows to keep the FPU in round-upwards at all times (for instance, “Handbook of Floating-Point Arithmetic”, 2.9.2).
This works well for addition and for multiplication. The square root operation, on the other hand, is not symmetrical in the ways addition and multiplication are.
It occurs to me that in order to compute sqrtRD, for the lower bound, the following idiom, despite its complications, might be faster on an ordinary platform with IEEE 754 double-precision and FLT_EVAL_METHOD defined to 0 than changing the rounding mode twice:
#include <fenv.h>
#include <math.h>
#pragma STDC FENV_ACCESS ON
…
/* assumes round-upwards */
double sqrt_rd(double l) {
feclearexcept(FE_INEXACT);
double candidate = sqrt(l);
if (fetestexcept(FE_INEXACT))
return nextafter(candidate, 0);
return candidate;
}
I am wondering whether this is better, and whether if it is the fastest yet. As one possible alternative, but still not necessarily the fastest, it seems to me that FMARU(candidate, candidate, -l) might perhaps not be always exact (because of the directed rounding) but might be accurate enough around 0 for the following to work:
/* assumes round-upwards */
double sqrt_rd(double l) {
double candidate = sqrt(l);
if (fma(candidate, candidate, -l) != 0.0)
return nextafter(candidate, 0);
return candidate;
}
What other inexpensive ways are there to detect that sqrt was inexact?
What combination of floating-point operations leads to the fastest computation of sqrt_rd on a modern FPU set to round upwards?

I think you should be able to use:
/* assumes round-upwards */
double sqrt_rd(double l) {
double u = sqrt(l);
double w = u*u;
if (w != l)
return nextafter(u, 0);
return u;
}
The justification here being that if u is inexact, then it will be strictly greater than √l, which in turn implies that w >= u2 > l (since w is also calculated in RU mode). And if u is exact, then so is w (since we know it must be representable as a double).

fma calculates the result with infinite precision, then applies the rounding mode.
If your candidate is too large, then the infinitely precise result is greater than 0, and since you are rounding up, it will be rounded up. Even if it is only a tiny little bit larger than zero. To verify this, first try l = 1 + 2eps, where (1 + eps) = sqrt (1 + 2eps + eps^2) is just a tiny bit too large; then scale l down by a negative power of 4 so that the eps^2 is way beyond the resolution of denormalised numbers, and check that as well.

Related

Efficient floating-point division with constant integer divisors

A recent question, whether compilers are allowed to replace floating-point division with floating-point multiplication, inspired me to ask this question.
Under the stringent requirement, that the results after code transformation shall be bit-wise identical to the actual division operation,
it is trivial to see that for binary IEEE-754 arithmetic, this is possible for divisors that are a power of two. As long as the reciprocal
of the divisor is representable, multiplying by the reciprocal of the divisor delivers results identical to the division. For example, multiplication by 0.5 can replace division by 2.0.
One then wonders for what other divisors such replacements work, assuming we allow any short instruction sequence that replaces division but runs significantly faster, while delivering bit-identical results. In particular allow fused multiply-add operations in addition to plain multiplication.
In comments I pointed to the following relevant paper:
Nicolas Brisebarre, Jean-Michel Muller, and Saurabh Kumar Raina. Accelerating correctly rounded floating-point division when the divisor is known in advance. IEEE Transactions on Computers, Vol. 53, No. 8, August 2004, pp. 1069-1072.
The technique advocated by the authors of the paper precomputes the reciprocal of the divisor y as a normalized head-tail pair zh:zl as follows: zh = 1 / y, zl = fma (-y, zh, 1) / y. Later, the division q = x / y is then computed as q = fma (zh, x, zl * x). The paper derives various conditions that divisor y must satisfy for this algorithm to work. As one readily observes, this algorithm has problems with infinities and zero when the signs of head and tail differ. More importantly, it will fail to deliver correct results for dividends x that are very small in magnitude, because computation of the quotient tail, zl * x, suffers from underflow.
The paper also makes a passing reference to an alternative FMA-based division algorithm, pioneered by Peter Markstein when he was at IBM. The relevant reference is:
P. W. Markstein. Computation of elementary functions on the IBM RISC System/6000 processor. IBM Journal of Research & Development, Vol. 34, No. 1, January 1990, pp. 111-119
In Markstein's algorithm, one first computes a reciprocal rc, from which an initial quotient q = x * rc is formed. Then, the remainder of the division is computed accurately with an FMA as r = fma (-y, q, x), and an improved, more accurate quotient is finally computed as q = fma (r, rc, q).
This algorithm also has issues for x that are zeroes or infinities (easily worked around with appropriate conditional execution), but exhaustive testing using IEEE-754 single-precision float data shows that it delivers the correct quotient across all possibe dividends x for many divisors y, among these many small integers. This C code implements it:
/* precompute reciprocal */
rc = 1.0f / y;
/* compute quotient q=x/y */
q = x * rc;
if ((x != 0) && (!isinf(x))) {
r = fmaf (-y, q, x);
q = fmaf (r, rc, q);
}
On most processor architectures, this should translate into a branchless sequence of instructions, using either predication, conditional moves, or select-type instructions. To give a concrete example: For division by 3.0f, the nvcc compiler of CUDA 7.5 generates the following machine code for a Kepler-class GPU:
LDG.E R5, [R2]; // load x
FSETP.NEU.AND P0, PT, |R5|, +INF , PT; // pred0 = fabsf(x) != INF
FMUL32I R2, R5, 0.3333333432674408; // q = x * (1.0f/3.0f)
FSETP.NEU.AND P0, PT, R5, RZ, P0; // pred0 = (x != 0.0f) && (fabsf(x) != INF)
FMA R5, R2, -3, R5; // r = fmaf (q, -3.0f, x);
MOV R4, R2 // q
#P0 FFMA R4, R5, c[0x2][0x0], R2; // if (pred0) q = fmaf (r, (1.0f/3.0f), q)
ST.E [R6], R4; // store q
For my experiments, I wrote the tiny C test program shown below that steps through integer divisors in increasing order and for each of them exhaustively tests the above code sequence against the proper division. It prints a list of the divisors that passed this exhaustive test. Partial output looks as follows:
PASS: 1, 2, 3, 4, 5, 7, 8, 9, 11, 13, 15, 16, 17, 19, 21, 23, 25, 27, 29, 31, 32, 33, 35, 37, 39, 41, 43, 45, 47, 49, 51, 53, 55, 57, 59, 61, 63, 64, 65, 67, 69,
To incorporate the replacement algorithm into a compiler as an optimization, a whitelist of divisors to which the above code transformation can safely be applied is impractical. The output of the program so far (at a rate of about one result per minute) suggests that the fast code works correctly across all possible encodings of x for those divisors y that are odd integers or are powers of two. Anecdotal evidence, not a proof, of course.
What set of mathematical conditions can determine a-priori whether the transformation of division into the above code sequence is safe? Answers can assume that all the floating-point operations are performed in the default rounding mode of "round to nearest or even".
#include <stdlib.h>
#include <stdio.h>
#include <math.h>
int main (void)
{
float r, q, x, y, rc;
volatile union {
float f;
unsigned int i;
} arg, res, ref;
int err;
y = 1.0f;
printf ("PASS: ");
while (1) {
/* precompute reciprocal */
rc = 1.0f / y;
arg.i = 0x80000000;
err = 0;
do {
/* do the division, fast */
x = arg.f;
q = x * rc;
if ((x != 0) && (!isinf(x))) {
r = fmaf (-y, q, x);
q = fmaf (r, rc, q);
}
res.f = q;
/* compute the reference, slowly */
ref.f = x / y;
if (res.i != ref.i) {
err = 1;
break;
}
arg.i--;
} while (arg.i != 0x80000000);
if (!err) printf ("%g, ", y);
y += 1.0f;
}
return EXIT_SUCCESS;
}
Let me restart for the third time. We are trying to accelerate
q = x / y
where y is an integer constant, and q, x, and y are all IEEE 754-2008 binary32 floating-point values. Below, fmaf(a,b,c) indicates a fused multiply-add a * b + c using binary32 values.
The naive algorithm is via a precalculated reciprocal,
C = 1.0f / y
so that at runtime a (much faster) multiplication suffices:
q = x * C
The Brisebarre-Muller-Raina acceleration uses two precalculated constants,
zh = 1.0f / y
zl = -fmaf(zh, y, -1.0f) / y
so that at runtime, one multiplication and one fused multiply-add suffices:
q = fmaf(x, zh, x * zl)
The Markstein algorithm combines the naive approach with two fused multiply-adds that yields the correct result if the naive approach yields a result within 1 unit in the least significant place, by precalculating
C1 = 1.0f / y
C2 = -y
so that the divison can be approximated using
t1 = x * C1
t2 = fmaf(C1, t1, x)
q = fmaf(C2, t2, t1)
The naive approach works for all powers of two y, but otherwise it is pretty bad. For example, for divisors 7, 14, 15, 28, and 30, it yields an incorrect result for more than half of all possible x.
The Brisebarre-Muller-Raina approach similarly fails for almost all non-power of two y, but much fewer x yield the incorrect result (less than half a percent of all possible x, varies depending on y).
The Brisebarre-Muller-Raina article shows that the maximum error in the naive approach is ±1.5 ULPs.
The Markstein approach yields correct results for powers of two y, and also for odd integer y. (I have not found a failing odd integer divisor for the Markstein approach.)
For the Markstein approach, I have analysed divisors 1 - 19700 (raw data here).
Plotting the number of failure cases (divisor in the horizontal axis, the number of values of x where Markstein approach fails for said divisor), we can see a simple pattern occur:
(source: nominal-animal.net)
Note that these plots have both horizontal and vertical axes logarithmic. There are no dots for odd divisors, as the approach yields correct results for all odd divisors I've tested.
If we change the x axis to the bit reverse (binary digits in reverse order, i.e. 0b11101101 → 0b10110111, data) of the divisors, we have a very clear pattern:
(source: nominal-animal.net)
If we draw a straight line through the center of the point sets, we get curve 4194304/x. (Remember, the plot considers only half the possible floats, so when considering all possible floats, double it.)
8388608/x and 2097152/x bracket the entire error pattern completely.
Thus, if we use rev(y) to compute the bit reverse of divisor y, then 8388608/rev(y) is a good first order approximation of the number of cases (out of all possible float) where the Markstein approach yields an incorrect result for an even, non-power-of-two divisor y. (Or, 16777216/rev(x) for the upper limit.)
Added 2016-02-28: I found an approximation for the number of error cases using the Markstein approach, given any integer (binary32) divisor. Here it is as pseudocode:
function markstein_failure_estimate(divisor):
if (divisor is zero)
return no estimate
if (divisor is not an integer)
return no estimate
if (divisor is negative)
negate divisor
# Consider, for avoiding underflow cases,
if (divisor is very large, say 1e+30 or larger)
return no estimate - do as division
while (divisor > 16777216)
divisor = divisor / 2
if (divisor is a power of two)
return 0
if (divisor is odd)
return 0
while (divisor is not odd)
divisor = divisor / 2
# Use return (1 + 83833608 / divisor) / 2
# if only nonnegative finite float divisors are counted!
return 1 + 8388608 / divisor
This yields a correct error estimate to within ±1 on the Markstein failure cases I have tested (but I have not yet adequately tested divisors larger than 8388608). The final division should be such that it reports no false zeroes, but I cannot guarantee it (yet). It does not take into account very large divisors (say 0x1p100, or 1e+30, and larger in magnitude) which have underflow issues -- I would definitely exclude such divisors from acceleration anyway.
In preliminary testing, the estimate seems uncannily accurate. I did not draw a plot comparing the estimates and the actual errors for divisors 1 to 20000, because the points all coincide exactly in the plots. (Within this range, the estimate is exact, or one too large.) Essentially, the estimates reproduce the first plot in this answer exactly.
The pattern of failures for the Markstein approach is regular, and very interesting. The approach works for all power of two divisors, and all odd integer divisors.
For divisors greater than 16777216, I consistently see the same errors as for a divisor that is divided by the smallest power of two to yield a value less than 16777216. For example, 0x1.3cdfa4p+23 and 0x1.3cdfa4p+41, 0x1.d8874p+23 and 0x1.d8874p+32, 0x1.cf84f8p+23 and 0x1.cf84f8p+34, 0x1.e4a7fp+23 and 0x1.e4a7fp+37. (Within each pair, the mantissa is the same, and only the power of two varies.)
Assuming my test bench is not in error, this means that the Markstein approach also works divisors larger than 16777216 in magnitude (but smaller than, say, 1e+30), if the divisor is such that when divided by the smallest power of two that yields a quotient of less than 16777216 in magnitude, and the quotient is odd.
This question asks for a way to identify the values of the constant Y that make it safe to transform x / Y into a cheaper computation using FMA for all possible values of x. Another approach is to use static analysis to determine an over-approximation of the values x can take, so that the generally unsound transformation can be applied in the knowledge that the values for which the transformed code differs from the original division do not happen.
Using representations of sets of floating-point values that are well adapted to the problems of floating-point computations, even a forwards analysis starting from the beginning of the function can produce useful information. For instance:
float f(float z) {
float x = 1.0f + z;
float r = x / Y;
return r;
}
Assuming the default round-to-nearest mode(*), in the above function x can only be NaN (if the input is NaN), +0.0f, or a number larger than 2-24 in magnitude, but not -0.0f or anything closer to zero than 2-24. This justifies the transformation into one of the two forms shown in the question for many values of the constant Y.
(*) assumption without which many optimizations are impossible and that C compilers already make unless the program explicitly uses #pragma STDC FENV_ACCESS ON
A forwards static analysis that predicts the information for x above can be based on a representation of sets of floating-point values an expression can take as a tuple of:
a representation for the sets of possible NaN values (Since behaviors of NaN are underspecified, a choice is to use only a boolean, with true meaning some NaNs can be present, and false indicating no NaN is present.),
four boolean flags indicating respectively the presence of +inf, -inf, +0.0, -0.0,
an inclusive interval of negative finite floating-point values, and
an inclusive interval of positive finite floating-point values.
In order to follow this approach, all the floating-point operations that can occur in a C program must be understood by the static analyzer. To illustrate, the addition betweens sets of values U and V, to be used to handle + in the analyzed code, can be implemented as:
If NaN is present in one of the operands, or if the operands can be infinities of opposite signs, NaN is present in the result.
If 0 cannot be a result of the addition of a value of U and a value of V, use standard interval arithmetic. The upper bound of the result is obtained for the round-to-nearest addition of the largest value in U and the largest value in V, so these bounds should be computed with round-to-nearest.
If 0 can be a result of the addition of a positive value of U and a negative value of V, then let M be the smallest positive value in U such that -M is present in V.
if succ(M) is present in U, then this pair of values contributes succ(M) - M to the positive values of the result.
if -succ(M) is present in V, then this pair of values contributes the negative value M - succ(M) to the negative values of the result.
if pred(M) is present in U, then this pair of values contributes the negative value pred(M) - M to the negative values of the result.
if -pred(M) is present in V, then this pair of values contributes the value M - pred(M) to the positive values of the result.
Do the same work if 0 can be the result of the addition of a negative value of U and a positive value of V.
Acknowledgement: the above borrows ideas from “Improving the Floating Point Addition and Subtraction Constraints”, Bruno Marre & Claude Michel
Example: compilation of the function f below:
float f(float z, float t) {
float x = 1.0f + z;
if (x + t == 0.0f) {
float r = x / 6.0f;
return r;
}
return 0.0f;
}
The approach in the question refuses to transform the division in function f into an alternate form, because 6 is not one of the value for which the division can be unconditionally transformed. Instead, what I am suggesting is to apply a simple value analysis starting from the beginning of the function which, in this case, determines that x is a finite float either +0.0f or at least 2-24 in magnitude, and to use this information to apply Brisebarre et al's transformation, confident in the knowledge that x * C2 does not underflow.
To be explicit, I am suggesting to use an algorithm such as the one below to decide whether or not to transform the division into something simpler:
Is Y one of the values that can be transformed using Brisebarre et al's method according to their algorithm?
Do C1 and C2 from their method have the same sign, or is it possible to exclude the possibility that the dividend is infinite?
Do C1 and C2 from their method have the same sign, or can x take only one of the two representations of 0? If in the case where C1 and C2 have different signs and x can only be one representation of zero, remember to fiddle(**) with the signs of the FMA-based computation to make it produce the correct zero when x is zero.
Can the magnitude of the dividend be guaranteed to be large enough to exclude the possibility that x * C2 underflows?
If the answer to the four questions is “yes”, then the division can be transformed into a multiplication and an FMA in the context of the function being compiled. The static analysis described above serves to answer questions 2., 3. and 4.
(**) “fiddling with the signs” means using -FMA(-C1, x, (-C2)*x) in place of FMA(C1, x, C2*x) when this is necessary to make the result come out correctly when x can only be one of the two signed zeroes
I love #Pascal's answer but in optimization it's often better to have a simple and well-understood subset of transformations rather than a perfect solution.
All current and common historical floating point formats had one thing in common: a binary mantissa.
Therefore, all fractions were rational numbers of the form:
x / 2n
This is in contrast to the constants in the program (and all possible base-10 fractions) which are rational numbers of the form:
x / (2n * 5m)
So, one optimization would simply test the input and reciprocal for m == 0, since those numbers are represented exactly in the FP format and operations with them should produce numbers that are accurate within the format.
So, for example, within the (decimal 2-digit) range of .01 to 0.99 dividing or multiplying by the following numbers would be optimized:
.25 .50 .75
And everything else would not. (I think, do test it first, lol.)
The result of a floating point division is:
a sign flag
a significand
an exponent
a set of flags (overflow, underflow, inexact, etc - see fenv())
Getting the first 3 pieces correct (but the set of flags incorrect) is not enough. Without further knowledge (e.g. which parts of which pieces of the result actually matter, the possible values of the dividend, etc) I would assume that replacing division by a constant with multiplication by a constant (and/or a convoluted FMA mess) is almost never safe.
In addition; for modern CPUs I also wouldn't assume that replacing a division with 2 FMAs is always an improvement. For example, if the bottleneck is instruction fetch/decode, then this "optimisation" would make performance worse. For another example, if subsequent instructions don't depend on the result (the CPU can do many other instructions in parallel while waiting for the result) the FMA version may introduce multiple dependency stalls and make performance worse. For a third example, if all registers are being used then the FMA version (which requires additional "live" variables) may increase "spilling" and make performance worse.
Note that (in many but not all cases) division or multiplication by a constant multiple of 2 can be done with addition alone (specifically, adding a shift count to the exponent).

accuracy of sqrt of integers

I have a loop like this:
for(uint64_t i=0; i*i<n; i++) {
This requires doing a multiplication every iteration. If I could calculate the sqrt before the loop then I could avoid this.
unsigned cut = sqrt(n)
for(uint64_t i=0; i<cut; i++) {
In my case it's okay if the sqrt function rounds up to the next integer but it's not okay if it rounds down.
My question is: is the sqrt function accurate enough to do this for all cases?
Edit: Let me list some cases. If n is a perfect square so that n = y^2 my question would be - is cut=sqrt(n)>=y for all n? If cut=y-1 then there is a problem. E.g. if n = 120 and cut = 10 it's okay but if n=121 (11^2) and cut is still 10 then it won't work.
My first concern was the fractional part of float only has 23 bits and double 52 so they can't store all the digits of some 32-bit or 64-bit integers. However, I don't think this is a problem. Let's assume we want the sqrt of some number y but we can't store all the digits of y. If we let the fraction of y we can store be x we can write y = x + dx then we want to make sure that whatever dx we choose does not move us to the next integer.
sqrt(x+dx) < sqrt(x) + 1 //solve
dx < 2*sqrt(x) + 1
// e.g for x = 100 dx < 21
// sqrt(100+20) < sqrt(100) + 1
Float can store 23 bits so we let y = 2^23 + 2^9. This is more than sufficient since 2^9 < 2*sqrt(2^23) + 1. It's easy to show this for double as well with 64-bit integers. So although they can't store all the digits as long as the sqrt of what they can store is accurate then the sqrt(fraction) should be sufficient. Now let's look at what happens for integers close to INT_MAX and the sqrt:
unsigned xi = -1-1;
printf("%u %u\n", xi, (unsigned)(float)xi); //4294967294 4294967295
printf("%u %u\n", (unsigned)sqrt(xi), (unsigned)sqrtf(xi)); //65535 65536
Since float can't store all the digits of 2^31-2 and double can they get different results for the sqrt. But the float version of the sqrt is one integer larger. This is what I want. For 64-bit integers as long as the sqrt of the double always rounds up it's okay.
First, integer multiplication is really quite cheap. So long as you have more than a few cycles of work per loop iteration and one spare execute slot, it should be entirely hidden by reorder on most non-tiny processors.
If you did have a processor with dramatically slow integer multiply, a truly clever compiler might transform your loop to:
for (uint64_t i = 0, j = 0; j < cut; j += 2*i+1, i++)
replacing the multiply with an lea or a shift and two adds.
Those notes aside, let’s look at your question as stated. No, you can’t just use i < sqrt(n). Counter-example: n = 0x20000000000000. Assuming adherence to IEEE-754, you will have cut = 0x5a82799, and cut*cut is 0x1ffffff8eff971.
However, a basic floating-point error analysis shows that the error in computing sqrt(n) (before conversion to integer) is bounded by 3/4 of an ULP. So you can safely use:
uint32_t cut = sqrt(n) + 1;
and you’ll perform at most one extra loop iteration, which is probably acceptable. If you want to be totally precise, instead use:
uint32_t cut = sqrt(n);
cut += (uint64_t)cut*cut < n;
Edit: z boson clarifies that for his purposes, this only matters when n is an exact square (otherwise, getting a value of cut that is “too small by one” is acceptable). In that case, there is no need for the adjustment and on can safely just use:
uint32_t cut = sqrt(n);
Why is this true? It’s pretty simple to see, actually. Converting n to double introduces a perturbation:
double_n = n*(1 + e)
which satisfies |e| < 2^-53. The mathematical square root of this value can be expanded as follows:
square_root(double_n) = square_root(n)*square_root(1+e)
Now, since n is assumed to be a perfect square with at most 64 bits, square_root(n) is an exact integer with at most 32 bits, and is the mathematically precise value that we hope to compute. To analyze the square_root(1+e) term, use a taylor series about 1:
square_root(1+e) = 1 + e/2 + O(e^2)
= 1 + d with |d| <~ 2^-54
Thus, the mathematically exact value square_root(double_n) is less than half an ULP away from[1] the desired exact answer, and necessarily rounds to that value.
[1] I’m being fast and loose here in my abuse of relative error estimates, where the relative size of an ULP actually varies across a binade — I’m trying to give a bit of the flavor of the proof without getting too bogged down in details. This can all be made perfectly rigorous, it just gets to be a bit wordy for Stack Overflow.
All my answer is useless if you have access to IEEE 754 double precision floating point, since Stephen Canon demonstrated both
a simple way to avoid imul in loop
a simple way to compute the ceiling sqrt
Otherwise, if for some reason you have a non IEEE 754 compliant platform, or only single precision, you could get the integer part of square root with a simple Newton-Raphson loop. For example in Squeak Smalltalk we have this method in Integer:
sqrtFloor
"Return the integer part of the square root of self"
| guess delta |
guess := 1 bitShift: (self highBit + 1) // 2.
[
delta := (guess squared - self) // (guess + guess).
delta = 0 ] whileFalse: [
guess := guess - delta ].
^guess - 1
Where // is operator for quotient of integer division.
Final guard guess*guess <= self ifTrue: [^guess]. can be avoided if initial guess is fed in excess of exact solution as is the case here.
Initializing with approximate float sqrt was not an option because integers are arbitrarily large and might overflow
But here, you could seed the initial guess with floating point sqrt approximation, and my bet is that the exact solution will be found in very few loops. In C that would be:
uint32_t sqrtFloor(uint64_t n)
{
int64_t diff;
int64_t delta;
uint64_t guess=sqrt(n); /* implicit conversions here... */
while( (delta = (diff=guess*guess-n) / (guess+guess)) != 0 )
guess -= delta;
return guess-(diff>0);
}
That's a few integer multiplications and divisions, but outside the main loop.
What you are looking for is a way to calculate a rational upper bound of the square root of a natural number. Continued fraction is what you need see wikipedia.
For x>0, there is
.
To make the notation more compact, rewriting the above formula as
Truncate the continued fraction by removing the tail term (x-1)/2's at each recursion depth, one gets a sequence of approximations of sqrt(x) as below:
Upper bounds appear at lines with odd line numbers, and gets tighter. When distance between an upper bound and its neighboring lower bound is less than 1, that approximation is what you need. Using that value as the value of cut, here cut must be a float number, solves the problem.
For very large number, rational number should be used, so no precision is lost during conversion between integer and floating point number.

Floating-point division - bias to avoid a result less than an 'exact' value

I am currently tightening floating-point numerics for an estimate of a value. (It's: p(k,t) for those who are interested.) Essentially, the utility can never yield an under-estimate of this value: the security of probable prime generation depends on a numerically robust implementation. While output results agree with the published values, I have used the DBL_EPSILON value to ensure that division, in particular, yields a result that is never less than the true value:
Consider: double x, y; /* assigned some values... */
The evaluation: r = x / y; occurs frequently, but these (finite precision) results may truncate significant digits from the true result - a possibly infinite precision rational expansion. I currently try to mitigate this by applying a bias to the numerator, i.e.,
r = ((1.0 + DBL_EPSILON) * x) / y;
If you know anything about this subject, p(k,t) is typically much smaller than most estimates - but it's simply not good enough to dismiss the issue with this "observation". I can of course state:
(((1.0 + DBL_EPSILON) * x) / y) >= (x / y)
Of course, I need to ensure that the 'biased' result is greater than, or equal to, the 'exact' value. While I am certain it has to do with manipulating or scaling DBL_EPSILON, I obviously want the 'biased' result to exceed the 'exact' result by a minimum - demonstrable under IEEE-754 arithmetic assumptions.
Yes, I've looked though Goldberg's paper, and I've searched for a robust solution. Please don't suggest manipulation of rounding modes. Ideally, I'm after an answer by someone with a very good grasp on floating-point theorems, or knows of a very well illustrated example.
EDIT: To clarify, (((1.0 + DBL_EPSILON) * x) / y) or a form (((1.0 + c) * x) / y), is not a prerequisite. This was simply an approach I was using as 'probably good enough', without having provided a solid basis for it. I can state that the numerator and denominator will not be special values: NaNs, Infs, etc., nor will the denominator be zero.
First: I know that you don't want to set the rounding mode, but it really should be said that
in terms of precision, as others have noted, setting the rounding mode will produce as good of an answer as possible. Specifically, assuming that x and y are both positive (which seems to be the case, but hasn't been explicitly stated in your question), the following is a standard C snippet with the desired effect[1]:
#include <math.h>
#pragma STDC FENV_ACCESS on
int OldRoundingMode = fegetround();
fesetround(FE_UPWARD);
r = x/y;
fesetround(OldRoundingMode);
Now, that aside, there are legitimate reasons not to want to change the rounding mode (some platforms don't support round-to-plus-infinity, on some platforms changing the rounding mode introduces a large serializing stall, etc etc), and your desire not to do so shouldn't be brushed aside so casually. So, respecting your question, what else can we do?
If your platform supports fused multiply-add, there's a very elegant solution available to you:
#include <math.h>
r = x/y;
if (fma(r,y,-x) < 0) r = nextafter(r, INFINITY);
On platforms with hardware fma support, this is very efficient. Even if fma( ) is implemented in software, it may be acceptable. This approach has the virtue that it will deliver the same result as would changing the rounding mode; that is, the tightest bound possible.
If your platform's C library is antediluvian and does not provide fma, there is still hope. Your claimed statement is correct (assuming no denormal values, at least -- I would need to think more about what happens for denormals); (1.0+DBL_EPSILON)*x/y really is always greater than or equal to the infinitely precise x/y. It will sometimes be one ulp larger than the smallest value with this property, but that's a very small and probably acceptable margin. The proof of these claims is pretty fussy, and probably not suitable for StackOverflow, but I'll give a quick sketch:
Ignoring denormals, it suffices to restrict ourselves to x, y in [1.0, 2.0).
(1.0 + eps)*x >= x + eps > x. To see this, observe:
(1.0 + eps)*x = x + x*eps >= x + eps > x.
Let P be the mathematically precise x/y. We have:
(1.0 + eps)*x/y >= (x + eps)/y = x/y + eps/y = P + eps/y
Now, y is bounded above by 2, so this gives us:
(1.0 + eps)*x/y > P + eps/2
which is sufficient to guarantee that the result rounds to a value >= P. This also shows us the way to a tighter bound. We could instead use nextafter(x,INFINITY)/y to get the desired effect with a tighter bound in many cases. (nextafter(x,INFINITY) is always x + ulp, whereas (1.0 + eps)*x will be x + 2ulp half of the time. If you want to avoid calling the nextafter library function, you can use (x + (0.75*DBL_EPSILON)*x) instead to get the same result, under the working assumption of positive normal values).
In order to be really pedantically correct, this would become significantly more complicated. No one really writes code like this, but it would be along these lines:
#include <math.h>
#pragma STDC FENV_ACCESS on
#if defined FE_UPWARD
int OldRoundingMode = fegetround();
if (OldRoundingMode < 0) goto Error;
if (fesetround(FE_UPWARD)) goto Error;
r = x/y;
if (fesetround(OldRoundingMode)) goto TrulyHosed;
return r;
TrulyHosed:
// we established the desired rounding mode and did our computation,
// but now we can't set it back to the original mode. I have no idea
// how you handle this gracefully.
Error:
#else
// we can't establish the desired rounding mode, so fall back on
// something else.

Fast float to int conversion (truncate)

I'm looking for a way to truncate a float into an int in a fast and portable (IEEE 754) way. The reason is because in this function 50% of the time is spent in the cast:
float fm_sinf(float x) {
const float a = 0.00735246819687011731341356165096815f;
const float b = -0.16528911397014738207016302002888890f;
const float c = 0.99969198629596757779830113868360584f;
float r, x2;
int k;
/* bring x in range */
k = (int) (F_1_PI * x + copysignf(0.5f, x)); /* <-- 50% of time is spent in cast */
x -= k * F_PI;
/* if x is in an odd pi count we must flip */
r = 1 - 2 * (k & 1); /* trick for r = (k % 2) == 0 ? 1 : -1; */
x2 = x * x;
return r * x*(c + x2*(b + a*x2));
}
The slowness of float->int casts mainly occurs when using x87 FPU instructions on x86. To do the truncation, the rounding mode in the FPU control word needs to be changed to round-to-zero and back, which tends to be very slow.
When using SSE instead of x87 instructions, a truncation is available without control word changes. You can do this using compiler options (like -mfpmath=sse -msse -msse2 in GCC) or by compiling the code as 64-bit.
The SSE3 instruction set has the FISTTP instruction to convert to integer with truncation without changing the control word. A compiler may generate this instruction if instructed to assume SSE3.
Alternatively, the C99 lrint() function will convert to integer with the current rounding mode (round-to-nearest unless you changed it). You can use this if you remove the copysignf term. Unfortunately, this function is still not ubiquitous after more than ten years.
I found a fast truncate method by Sree Kotay which provides exactly the optimization that I needed.
to be portable you would have to add some directives and learn a couple assembler languages but you could theoretically could use some inline assembly to move portions of the floating point register into eax/rax ebx/rbx and convert what you would need by hand, floating point specification though is a pain in the butt, but I am pretty certain that if you do it with assembly you will be way faster, as your needs are very specific and the system method is probably more generic and less efficient for your purpose
You could skip the conversion to int altogether by using frexpf to get the mantissa and exponent, and inspect the raw mantissa (use a union) at the appropriate bit position (calculated using the exponent) to determine (the quadrant dependent) r.

Faster way of finding multiple of double

If have the following C function, used to determine if one number is a multiple of another to an arbirary tolerance
#include <math.h>
#define TOLERANCE 0.0001
int IsMultipleOf(double x,double mod)
{
return(fabs(fmod(x, mod)) < TOLERANCE);
}
It works fine, but profiling shows it to be very slow, to the extent that it has become a candidate for optimization. About 75% of the time is spent in modulo and the remaining in fabs. I'm trying to figure a way of speeding things up, using something like a look-up table. The parameter x changes regularly, whereas mod changes infrequently. The number of possible values of x is small enough that the space for a look-up would not be an issue, typically it will be one of a few hundred possible values. I can get rid of the fabs easily enough, but can't figure out a reasonable alternative to the modulo. Any ideas on how to optimize the above?
Edit The code will be running on a wide range of Windows desktop and mobile devices, hence processors could include Intel, AMD on desktop, and ARM or SH4 on mobile devices. VisualStudio 2008 is the compiler.
Do you really have to use modulo for this?
Wouldn't it be possible to just result = x / mod and then check if the decimal part of result is close to 0. For instance:
11 / 5.4999 = 2.000003 ==> 0.000003 < TOLERANCE
Or something like that.
Division (floating point or not, fmod in your case) is often an operation where the execution time varies a lot depending on the cpu and compiler:
gcc has a builtin replacement for
that if you give it the right compile
flags or if you use __builtin_fmod
explicitly. This then might map the
operation on a small number of
assembler instructions.
there may be special units like SSE
on intel processors where this
operation is implemented more
efficiently
By such tricks, depending on your environment (you didn't tell which) the time may vary from some clock cycles to some hundred. I think best is to look into the documentation of your compiler and cpu for that particular operation.
The following is probably overkill, and sub-optimal. But for what it is worth here is one way on how to do it.
We know the format of the double ...
1 bit for the sign
11 bits for the biased exponent
52 fraction bits
Let ...
value = x / mod;
exp = exponent bits of value - BIAS;
lsb = least sig bit of value's fraction bits;
Once you have that ...
/*
* If applying the exponent would eliminate the fraction bits
* then for double precision resolution it is a multiple.
* Note: lsb may require some massaging.
*/
if (exp > lsb)
return (true);
if (exp < 0)
return (false);
The only case remaining is the tolerance case. Build your double so that you are getting rid of all the digits to the left of the decimal.
sign bit is zero (positive)
exponent is the BIAS (1023 I think ... look it up to be sure)
shift the fraction bits as appropriate
Now compare it against your tolerance.
I think you need to inspect the bowels of your C RTL fmod() function: X86 FPU's have 'FPREM/FPREM1' instructions which computes remainders by repeated subtraction.
While floating point division is a single instruction, it seems you may need to call FPREM repeatedly to get the right answer for modulus, so your RTL may not use it.
I have not tested this at all, but from the way I understand fmod this should be equivalent inlined, which might let the compiler optimize it better, though I would have thought that the compiler's math library (or builtins) would work just as well. (also, I don't even know for sure if this is correct).
#include <math.h>
int IsMultipleOf(double x, double mod) {
long n = x / mod; // You should probably test for /0 or NAN result here
double new_x = mod * n;
double delta = x - new_x;
return fabs(delta) < TOLERANCE; // and for NAN result from fabs
}
Maybe you can get away with long long instead of double if you have comparable scale of data. For example long long would be enough for over 60 astronomical units in micrometer resolution.
Does it need to be double precision ? Depending on how good your math library is, this ought to be faster:
#include <math.h>
#define TOLERANCE 0.0001f
bool IsMultipleOf(float x, float mod)
{
return(fabsf(fmodf(x, mod)) < TOLERANCE);
}
I presume modulo looks a little like this on the inside:
mod(x,m) {
while (x > m) {
x = x - m
}
return x
}
I think that through some sort of search i could be optimised: eg:
fastmod(x,m) {
q = 1
while (m * q < x) {
q = q * 2
}
return mod((x - (q / 2) * m), m)
}
You might even choose to replace the finall call to mod with annother call to fastmod, adding the condition that if x < m then to return x.

Resources