We Keep Coding

How Do I Optimize This Algorithm To Handle Large Integers?

I have two integers x and y.
Rule:
If x > y : x = x - y and y = 2 * y;
If y > x : y = y - x and x = 2 * x;
If y == x : not infinite loop
The question is, will the two integers be an infinite loop or not?
Here is my code:
private static boolean IsPairLoop(int first, int second)
{
boolean loop = false;
HashMap<Integer, Integer> round_record = new HashMap<>();
int[] first_round = new int[]{first, second};
while ((first_round[0] != -1 && first_round[1] != -1))
{
if (round_record.containsKey(first_round[0]))
{
loop = true;
break;
}
round_record.put(first_round[0], first_round[1]);
PlayRound(first_round);
}
return loop;
}
private static void PlayRound(int[] round)
{
if (round[0] > round[1])
{
round[0] -= round[1];
round[1] += round[1];
}
else if (round[0] < round[1])
{
round[1] -= round[0];
round[0] += round[0];
}
else
{
round[0] = -1;
round[1] = -1;
}
}
This works fine for small integers. However, this is painfully slow when the integer difference is really large. Integer range is 1 to 2^30 for both x and y. What can I do to make this fast even when the integers difference is large?

In this problem, the sum x + y is invariant, and if it is odd, x == y is impossible.
Then if x and y have the same parity, after one iteration they become both even and remain so, and the problem is unchanged if you divide both by 2.
Hence the "instantaneous" solution:
while (x + y) & 1 == 0:
if x == y:
print "Not infinite"
break
if x > y:
x= (x - y) / 2
else:
y= (y - x) / 2
else:
print "Infinite"
As one of the arguments loses at least one bit on every iteration, there are never more than 64 iterations for 32 bits integers (and in practice much less, most of the time 0 !).
A variant is possible and can make sense for bignumbers:
if x == y, conclude finite.
if x and y have a different number of trailing zeroes, conclude infinite.
otherwise, discard the trailing zeroes and perform the reduction according to x > y or y > x and loop.

Instead of using a hashmap, use Floyd's cycle detection algorithm. Not only will this avoid large memory use, it will also avoid costly boxing and unboxing between int and Integer.
A second optimisation is to re-write the recurrence relations via a change of variables:
s = x+y
t = x-y
Then the recurrence relations become:
if t=0, stop
if t>0, s'=s, t'=2t-s
if t<0, s'=s, t'=2t+s
Note in this formulation, only the t variable changes.
The code (untested) will look something like this:
private static int step(int s, int t) {
if (t>0) return 2*t - s;
if (t<0) return 2*t + s;
return 0;
}
private static boolean IsPairLoop(int first, int second) {
int s = first+second;
int t_slow = first-second;
int t_fast = t_slow;
while(t_slow != t_fast) {
t_slow = step(s, t_slow);
t_fast = step(s, step(s, s_fast));
}
return t_slow != 0;
}
You may need to replace int with long if first+second can overflow.
I think, given the preconditions, you never get a t that goes off to infinity (since it's always true that |t| < s). But you may wish to double-check, perhaps adding some sort of assertion into your code.

Let’s work backwards a bit. What can produce x == y? The difference between the previous x and y has to be equal to twice the smaller of the two, i.e. the larger has to be three times the smaller. Things that don’t loop infinitely so far:
{n, n}
{n, 3n}
Where can {n, 3n} come from? Either
n is a difference a − b for some a > b, and 3n = 2b
3(a − b) = 2b
3a − 3b = 2b
3a = 5b
a = 5/3 b
A pair {m, 5/3 m} is something that produces {n, 3n} on the next step. (m has to be divisible by 3, but that’s okay.)
3n is a difference a − b for some a > b, and n = 2b
(a − b)/3 = 2b
a − b = 6b
a = 7b
A pair {m, 7m} is the only other thing that can produce {n, 3n} on the next step.
An updated list:
{n, n}
{n, 3n}
{n, 7n}
{n, 5/3 n}
Seems like a good time to generalize those last steps.
{n, qn} happens when:
n is a difference a − b for some a > b, and qn = 2b
q(a − b) = 2b
qa − qb = 2b
qa = (2 + q)b
a = (2 + q)/q b
or qn is a difference a − b for some a > b, and n = 2b
(a − b)/q = 2b
a − b = 2qb
a = (2q + 1)b
So if q = m/n is in the list, these are also in the list:
(2n + m)/m
(2m + n)/n
q = 1 generates:
(2 + 1)/1 = 3
2×1 + 1 = 3
q = 3 generates:
3
(2 + 3)/3 = 5/3
2×3 + 1 = 7
q = 5/3 generates:
3
5/3
7
(2 + 5/3)/(5/3) = (6 + 5)/5 = 11/5
(2×5/3) + 1 = 10/3 + 1 = 13/3
q = 7 generates:
3
5/3
7
11/5
13/3
(2 + 7)/7 = 9/7
2×7 + 1 = 15
Hmm… that’s pretty interesting. Let’s sort the list by the numerator:
3/1
5/3
7/1
9/7
11/5
13/3
15/1
Based on that pattern, I’d expect 17/15 to come next. Generating a list sorted by denominator with a computer:
3/1
7/1
15/1
31/1
63/1
5/3
13/3
29/3
61/3
11/5
27/5
59/5
9/7
25/7
57/7
23/9
55/9
21/11
53/11
19/13
51/13
17/15
49/15
47/17
45/19
43/21
41/23
39/25
37/27
35/29
33/31
Looks an awful lot like m/n where n is odd, m > n, and m + n is a power of two. So one way to optimize your algorithm would be:
private static boolean isPairLoop(int first, int second)
{
if (first == second) return false;
if (first > second) return isPairLoop(second, first);
if (first == 0) return true;
int d = gcd(first, second);
return Integer.bitCount(first / d + second / d) != 1;
}
private static int gcd(int a, int b)
{
return b == 0 ? a : gcd(b, a % b);
}
Takes quadratic time on the number of digits on bigints.
Now you just have to prove that it works. I hope it works.

Understanding the strictMath java library

I got bored and decided to dive into remaking the square root function without referencing any of the Math.java functions. I have gotten to this point:
package sqrt;
public class SquareRoot {
public static void main(String[] args) {
System.out.println(sqrtOf(8));
}
public static double sqrtOf(double n){
double x = log(n,2);
return powerOf(2, x/2);
}
public static double log(double n, double base)
{
return (Math.log(n)/Math.log(base));
}
public static double powerOf(double x, double y) {
return powerOf(e(),y * log(x, e()));
}
public static int factorial(int n){
if(n <= 1){
return 1;
}else{
return n * factorial((n-1));
}
}
public static double e(){
return 1/factorial(1);
}
public static double e(int precision){
return 1/factorial(precision);
}
}
As you may very well see, I came to the point in my powerOf() function that infinitely recalls itself. I could replace that and use Math.exp(y * log(x, e()), so I dived into the Math source code to see how it handled my problem, resulting in a goose chase.
public static double exp(double a) {
return StrictMath.exp(a); // default impl. delegates to StrictMath
}
which leads to:
public static double exp(double x)
{
if (x != x)
return x;
if (x > EXP_LIMIT_H)
return Double.POSITIVE_INFINITY;
if (x < EXP_LIMIT_L)
return 0;
// Argument reduction.
double hi;
double lo;
int k;
double t = abs(x);
if (t > 0.5 * LN2)
{
if (t < 1.5 * LN2)
{
hi = t - LN2_H;
lo = LN2_L;
k = 1;
}
else
{
k = (int) (INV_LN2 * t + 0.5);
hi = t - k * LN2_H;
lo = k * LN2_L;
}
if (x < 0)
{
hi = -hi;
lo = -lo;
k = -k;
}
x = hi - lo;
}
else if (t < 1 / TWO_28)
return 1;
else
lo = hi = k = 0;
// Now x is in primary range.
t = x * x;
double c = x - t * (P1 + t * (P2 + t * (P3 + t * (P4 + t * P5))));
if (k == 0)
return 1 - (x * c / (c - 2) - x);
double y = 1 - (lo - x * c / (2 - c) - hi);
return scale(y, k);
}
Values that are referenced:
LN2 = 0.6931471805599453, // Long bits 0x3fe62e42fefa39efL.
LN2_H = 0.6931471803691238, // Long bits 0x3fe62e42fee00000L.
LN2_L = 1.9082149292705877e-10, // Long bits 0x3dea39ef35793c76L.
INV_LN2 = 1.4426950408889634, // Long bits 0x3ff71547652b82feL.
INV_LN2_H = 1.4426950216293335, // Long bits 0x3ff7154760000000L.
INV_LN2_L = 1.9259629911266175e-8; // Long bits 0x3e54ae0bf85ddf44L.
P1 = 0.16666666666666602, // Long bits 0x3fc555555555553eL.
P2 = -2.7777777777015593e-3, // Long bits 0xbf66c16c16bebd93L.
P3 = 6.613756321437934e-5, // Long bits 0x3f11566aaf25de2cL.
P4 = -1.6533902205465252e-6, // Long bits 0xbebbbd41c5d26bf1L.
P5 = 4.1381367970572385e-8, // Long bits 0x3e66376972bea4d0L.
TWO_28 = 0x10000000, // Long bits 0x41b0000000000000L
Here is where I'm starting to get lost. But I can make a few assumptions that so far the answer is starting to become estimated. I then find myself here:
private static double scale(double x, int n)
{
if (Configuration.DEBUG && abs(n) >= 2048)
throw new InternalError("Assertion failure");
if (x == 0 || x == Double.NEGATIVE_INFINITY
|| ! (x < Double.POSITIVE_INFINITY) || n == 0)
return x;
long bits = Double.doubleToLongBits(x);
int exp = (int) (bits >> 52) & 0x7ff;
if (exp == 0) // Subnormal x.
{
x *= TWO_54;
exp = ((int) (Double.doubleToLongBits(x) >> 52) & 0x7ff) - 54;
}
exp += n;
if (exp > 0x7fe) // Overflow.
return Double.POSITIVE_INFINITY * x;
if (exp > 0) // Normal.
return Double.longBitsToDouble((bits & 0x800fffffffffffffL)
| ((long) exp << 52));
if (exp <= -54)
return 0 * x; // Underflow.
exp += 54; // Subnormal result.
x = Double.longBitsToDouble((bits & 0x800fffffffffffffL)
| ((long) exp << 52));
return x * (1 / TWO_54);
}
TWO_54 = 0x40000000000000L
While I am, I would say, very understanding of math and programming, I hit the point to where I find myself at a Frankenstein monster mix of the two. I noticed the intrinsic switch to bits (which I have little to no experience with), and I was hoping someone could explain to me the processes that are occurring "under the hood" so to speak. Specifically where I got lost is from "Now x is in primary range" in the exp() method on wards and what the values that are being referenced really represent. I'm was asking for someone to help me understand not only the methods themselves, but also how they arrive to the answer. Feel free to go as in depth as needed.
edit:
if someone could maybe make this tag: "strictMath" that would be great. I believe that its size and for the Math library deriving from it justifies its existence.

To the exponential function:
What happens is that
exp(x) = 2^k * exp(x-k*log(2))
is exploited for positive x. Some magic is used to get more consistent results for large x where the reduction x-k*log(2) will introduce cancellation errors.
On the reduced x a rational approximation with minimized maximal error over the interval 0.5..1.5 is used, see Pade approximations and similar. This is based on the symmetric formula
exp(x) = exp(x/2)/exp(-x/2) = (c(x²)+x)/(c(x²)-x)
(note that the c in the code is x+c(x)-2). When using Taylor series, approximations for c(x*x)=x*coth(x/2) are based on
c(u)=2 + 1/6*u - 1/360*u^2 + 1/15120*u^3 - 1/604800*u^4 + 1/23950080*u^5 - 691/653837184000*u^6
The scale(x,n) function implements the multiplication x*2^n by directly manipulating the exponent in the bit assembly of the double floating point format.
Computing square roots
To compute square roots it would be more advantageous to compute them directly. First reduce the interval of approximation arguments via
sqrt(x)=2^k*sqrt(x/4^k)
which can again be done efficiently by directly manipulating the bit format of double.
After x is reduced to the interval 0.5..2.0 one can then employ formulas of the form
u = (x-1)/(x+1)
y = (c(u*u)+u) / (c(u*u)-u)
based on
sqrt(x)=sqrt(1+u)/sqrt(1-u)
and
c(v) = 1+sqrt(1-v) = 2 - 1/2*v - 1/8*v^2 - 1/16*v^3 - 5/128*v^4 - 7/256*v^5 - 21/1024*v^6 - 33/2048*v^7 - ...
In a program without bit manipulations this could look like
double my_sqrt(double x) {
double c,u,v,y,scale=1;
int k=0;
if(x<0) return NaN;
while(x>2 ) { x/=4; scale *=2; k++; }
while(x<0.5) { x*=4; scale /=2; k--; }
// rational approximation of sqrt
u = (x-1)/(x+1);
v = u*u;
c = 2 - v/2*(1 + v/4*(1 + v/2));
y = 1 + 2*u/(c-u); // = (c+u)/(c-u);
// one Halley iteration
y = y*(1+8*x/(3*(3*y*y+x))) // = y*(y*y+3*x)/(3*y*y+x)
// reconstruct original scale
return y*scale;
}
One could replace the Halley step with two Newton steps, or
with a better uniform approximation in c one could replace the Halley step with one Newton step, or ...

Can I make this function more efficient (Project Euler Number 9)?

I just finished Project Euler problem 9 (warning spoilers):
A Pythagorean triplet is a set of three natural numbers, a < b < c, for which,
a^2 + b^2 = c^2
For example, 3^2 + 4^2 = 9 + 16 = 25 = 5^2.
There exists exactly one Pythagorean triplet for which a + b + c = 1000.
Find the product abc.
Here's my solution:
public static int specPyth(int num)
{
for (int a = 1; a < num; a++)
for (int b = 2; b < a; b++)
{
if (a*a +b*b == (num-a-b)*(num-a-b))
return a*b*(num-a-b); //ans = 31875000
}
return -1;
}
I can't help but think that there's a solution that involves only one loop. Anyone have ideas? I'd prefer answers using only one loop, but anything that's more efficient than what I currently have would be nice.

if a + b +c = 1000
then
a + b + sqroot(a² + b²) = 1000
-> (a² + b²) = (1000 - a - b)²
-> a² + b² = 1000000 - 2000*(a+b) + a² + 2*a*b + b²
-> 0 = 1000000 - 2000*(a+b) + 2*a*b
-> ... (easy basic maths)
-> a = (500000 - 1000*b) / (1000 - b)
Then you try every b until you find one that makes a natural number out of a.
public static int specPyth(int num)
{
double a;
for (int b = 1; b < num/2; b++)
{
a=(num*num/2 - num*b)/(num - b);
if (a%1 == 0)
return (int) (a*b*(num-a-b));
}
return -1;
}
EDIT: b can't be higher than 499, because c>b and (b+c) would then be higher than 1000.

I highly recommend reading http://en.wikipedia.org/wiki/Pythagorean_triple#Generating_a_triple and writing a function that will generate the Pythagorean triples one by one.
Not to give too much of a spoiler, but there are a number of other PE problems that this function will come in handy for.
(I don't consider this giving away too much, because part of the purpose of PE is to encourage people to learn about things like this.)

First, since a is the smallest, you need not to count it up to num, num/3 is sufficient, and even num/(2+sqrt(2)).
Second, having a and constraints
a+b+c=num
a^2+b^2=c^2
we can solve this equations and find b and c for given a, which already satisfy this equations and there is no need to check if a^2+b^2=c^2 as you do now. All you need is to check if b and c are integer. And this is done in one loop
for (int a = 1; a < num/3; a++)

Runs in 62 milli seconds
import time
s = time.time()
tag,n=True,1000
for a in xrange (1,n/2):
if tag==False:
break
for b in xrange (1,n/2):
if a*a + b*b - (n-a-b)*(n-a-b) ==0:
print a,b,n-a-b
print a*b*(n-a-b)
tag=False
print time.time() - s

C solution
Warning : solution assumes that GCD(a, b) = 1. It works here but may not always work. I'll fix the solution in some time.
#include <stdio.h>
#include <math.h>
int main(void)
{
int n = 1000; // a + b + c = n
n /= 2;
for(int r = (int) sqrt(n / 2); r <= (int) sqrt(n); r++)
{
if(n % r == 0)
{
int s = (n / r) - r;
printf("%d %d %d\n", r*r - s*s, 2*r*s, r*r + s*s);
printf("Product is %d\n", (2*r*s) * (r*r - s*s) * (r*r + s*s));
}
}
return 0;
}
Solution uses Euclid's formula for triplets which states that any primitive triple is of form a = r^2 - s^2, b = 2rs, c = r^2 + s^2.
Certain restrictions like sqrt(n / 2) <= r <= sqrt(n) can be added based on the fact that s is positive and r > s.
Warning: you may need long long if the product is large

Definitely not the most optimal solution, but my first instinct was to use a modified 3SUM. In Python,
def problem_9(max_value = 1000):
i = 0
range_of_values = [n for n in range(1, max_value + 1)]
while i < max_value - 3:
j = i + 1
k = max_value - 1
while j < k:
a = range_of_values[i]
b = range_of_values[j]
c = range_of_values[k]
if ((a + b + c) == 1000) and (a*a + b*b == c*c):
return a*b*c
elif (a + b + c) < 1000:
j += 1
else:
k -= 1
i += 1
return -1

You say a < b < c, then b must always be bigger than a, so your starting point in the second loop could be b = a + 1; that would lead certainly to fewer iterations.
int specPyth(int num)
{
for (int a = 1; a < num/3; a++)
for (int b = a + 1; b < num/2; b++)
{
int c = num - a - b;
if (a * a + b * b == c * c)
return a * b * c; //ans = 31875000
}
return -1;
}

In the first given equation, you have three variables a, b, c. If you want to find-out matching values for this equation, you have to run 3 dimension loop. Fortunately there is another equation a+b+c=N where N is known number.
Using this, you can reduce down the dimension to two because if you know two among the three, you can calculate the rest. For instance, if you know a and b, c equals N - a - b.
What if you can reduce one more dimension of the loop? It is possible if you fiddle with the two given equations. Get a pen and paper. Once you get the additional equation with two variables and one constant (N), you will be able to acquire the result in O(n). Solve the two equations a+b+c=n; a^2+b^2=c^2 taking n and a to be constant and solve for b and c:
public static void main(String[] args) {
Scanner in = new Scanner(System.in);
int t = in.nextInt();
for(int a0 = 0; a0 < t; a0++){
int n = in.nextInt();
int max=-1;
int multi=0;
int b=1,c=1;
for(int i=1;i<=n;i++)
{
if(2*(i-n)!=0)
b=(2*i*n-(n*n))/(2*(i-n));
c=n-b-i;
if( (i*i+b*b==c*c)&& i+b+c==n && b>0 && c>=0 && i+b>c && c+i>b && b+c>i)
{
multi=i*b*c;
if(max<multi)
max=multi;
}
}
if(max==-1)
System.out.println(-1);
else
System.out.println(max);
}
}

Python:
Before jump into the code, do a small exercise in algebra.
a^2 + b^2 = c ^2
a + b + c = 1000 --> c = 1000 - (a+b)
a^2 + b^2 = (1000 - (a+b))^2
a = 1000*(500-b)/(1000-b)
Now the time to write the code.
for b in range(2, 500):
a = 1000*(500 - b)/(1000 - b)
if a.is_integer():
c = 1000 - (a+b)
print(a, b, c, a*b*c)
break

How to re-implement sin() method in Java ? (to have results close to Math.sin() )

I know Math.sin() can work but I need to implement it myself using factorial(int) I have a factorial method already below are my sin method but I can't get the same result as Math.sin():
public static double factorial(double n) {
if (n <= 1) // base case
return 1;
else
return n * factorial(n - 1);
}
public static double sin(int n) {
double sum = 0.0;
for (int i = 1; i <= n; i++) {
if (i % 2 == 0) {
sum += Math.pow(1, i) / factorial(2 * i + 1);
} else {
sum += Math.pow(-1, i) / factorial(2 * i + 1);
}
}
return sum;
}

You should use the Taylor series. A great tutorial here
I can see that you've tried but your sin method is incorrect
public static sin(int n) {
// angle to radians
double rad = n*1./180.*Math.PI;
// the first element of the taylor series
double sum = rad;
// add them up until a certain precision (eg. 10)
for (int i = 1; i <= PRECISION; i++) {
if (i % 2 == 0)
sum += Math.pow(rad, 2*i+1) / factorial(2 * i + 1);
else
sum -= Math.pow(rad, 2*i+1) / factorial(2 * i + 1);
}
return sum;
}
A working example of calculating the sin function. Sorry I've jotted it down in C++, but hope you get the picture. It's not that different :)

Your formula is wrong and you are getting a rough result of sin(1) and all you're doing by changing n is changing the accuracy of this calculation. You should look the formula up in Wikipedia and there you'll see that your n is in the wrong place and shouldn't be used as the limit of the for loop but rather in the numerator of the fraction, in the Math.pow(...) method. Check out Taylor Series

It looks like you are trying to use the taylor series expansion for sin, but have not included the term for x. Therefore, your method will always attempt to approximate sin(1) regardless of argument.
The method parameter only controls accuracy. In a good implementation, a reasonable value for that parameter is auto-detected, preventing the caller from passing to low a value, which can result in highly inaccurate results for large x. Moreover, to assist fast convergence (and prevent unnecessary loss of significance) of the series, implementations usually use that sin(x + k * 2 * PI) = sin(x) to first move x into the range [-PI, PI].
Also, your method is not very efficient, due to the repeated evaluations of factorials. (To evaluate factorial(5) you compute factorial(3), which you have already computed in the previous iteration of the for-loop).
Finally, note that your factorial implementation accepts an argument of type double, but is only correct for integers, and your sin method should probably receive the angle as double.

Sin (x) can be represented as Taylor series:
Sin (x) = (x/1!) – (x3/3!) + (x5/5!) - (x7/7!) + …
So you can write your code like this:
public static double getSine(double x) {
double result = 0;
for (int i = 0, j = 1, k = 1; i < 100; i++, j = j + 2, k = k * -1) {
result = result + ((Math.pow(x, j) / factorial (j)) * k);
}
return result;
}
Here we have run our loop only 100 times. If you want to run more than that you need to change your base equation (otherwise infinity value will occur).
I have learned a very good trick from the book “How to solve it by computer” by R.G.Dromey. He explain it like this way:
(x3/3! ) = (x X x X x)/(3 X 2 X 1) = (x2/(3 X 2)) X (x1/1!) i = 3
(x5/5! ) = (x X x X x X x X x)/(5 X 4 X 3 X 2 X 1) = (x2/(5 X 4)) X (x3/3!) i = 5
(x7/7! ) = (x X x X x X x X x X x X x)/(7 X 6 X 5 X 4 X 3 X 2 X 1) = (x2/(7 X 6)) X (x5/5!) i = 7
So the terms (x2/(3 X 2)) , (x2/(5 X 4)), (x2/(7 X 6)) can be expressed as x2/(i X (i - 1)) for i = 3,5,7,…
Therefore to generate consecutive terms of the sine series we can write:
current ith term = (x2 / ( i X (i - 1)) ) X (previous term)
The code is following:
public static double getSine(double x) {
double result = 0;
double term = x;
result = x;
for (int i = 3, j = -1; i < 100000000; i = i + 2, j = j * -1) {
term = x * x * term / (i * (i - 1));
result = result + term * j;
}
return result;
}
Note that j variable used to alternate the sign of the term .

Fastest way to determine if an integer's square root is an integer

I'm looking for the fastest way to determine if a long value is a perfect square (i.e. its square root is another integer):
I've done it the easy way, by using the built-in Math.sqrt()
function, but I'm wondering if there is a way to do it faster by
restricting yourself to integer-only domain.
Maintaining a lookup table is impractical (since there are about
231.5 integers whose square is less than 263).
Here is the very simple and straightforward way I'm doing it now:
public final static boolean isPerfectSquare(long n)
{
if (n < 0)
return false;
long tst = (long)(Math.sqrt(n) + 0.5);
return tst*tst == n;
}
Note: I'm using this function in many Project Euler problems. So no one else will ever have to maintain this code. And this kind of micro-optimization could actually make a difference, since part of the challenge is to do every algorithm in less than a minute, and this function will need to be called millions of times in some problems.
I've tried the different solutions to the problem:
After exhaustive testing, I found that adding 0.5 to the result of Math.sqrt() is not necessary, at least not on my machine.
The fast inverse square root was faster, but it gave incorrect results for n >= 410881. However, as suggested by BobbyShaftoe, we can use the FISR hack for n < 410881.
Newton's method was a good bit slower than Math.sqrt(). This is probably because Math.sqrt() uses something similar to Newton's Method, but implemented in the hardware so it's much faster than in Java. Also, Newton's Method still required use of doubles.
A modified Newton's method, which used a few tricks so that only integer math was involved, required some hacks to avoid overflow (I want this function to work with all positive 64-bit signed integers), and it was still slower than Math.sqrt().
Binary chop was even slower. This makes sense because the binary chop will on average require 16 passes to find the square root of a 64-bit number.
According to John's tests, using or statements is faster in C++ than using a switch, but in Java and C# there appears to be no difference between or and switch.
I also tried making a lookup table (as a private static array of 64 boolean values). Then instead of either switch or or statement, I would just say if(lookup[(int)(n&0x3F)]) { test } else return false;. To my surprise, this was (just slightly) slower. This is because array bounds are checked in Java.

I figured out a method that works ~35% faster than your 6bits+Carmack+sqrt code, at least with my CPU (x86) and programming language (C/C++). Your results may vary, especially because I don't know how the Java factor will play out.
My approach is threefold:
First, filter out obvious answers. This includes negative numbers and looking at the last 4 bits. (I found looking at the last six didn't help.) I also answer yes for 0. (In reading the code below, note that my input is int64 x.)
if( x < 0 || (x&2) || ((x & 7) == 5) || ((x & 11) == 8) )
return false;
if( x == 0 )
return true;
Next, check if it's a square modulo 255 = 3 * 5 * 17. Because that's a product of three distinct primes, only about 1/8 of the residues mod 255 are squares. However, in my experience, calling the modulo operator (%) costs more than the benefit one gets, so I use bit tricks involving 255 = 2^8-1 to compute the residue. (For better or worse, I am not using the trick of reading individual bytes out of a word, only bitwise-and and shifts.)
int64 y = x;
y = (y & 4294967295LL) + (y >> 32);
y = (y & 65535) + (y >> 16);
y = (y & 255) + ((y >> 8) & 255) + (y >> 16);
// At this point, y is between 0 and 511. More code can reduce it farther.
To actually check if the residue is a square, I look up the answer in a precomputed table.
if( bad255[y] )
return false;
// However, I just use a table of size 512
Finally, try to compute the square root using a method similar to Hensel's lemma. (I don't think it's applicable directly, but it works with some modifications.) Before doing that, I divide out all powers of 2 with a binary search:
if((x & 4294967295LL) == 0)
x >>= 32;
if((x & 65535) == 0)
x >>= 16;
if((x & 255) == 0)
x >>= 8;
if((x & 15) == 0)
x >>= 4;
if((x & 3) == 0)
x >>= 2;
At this point, for our number to be a square, it must be 1 mod 8.
if((x & 7) != 1)
return false;
The basic structure of Hensel's lemma is the following. (Note: untested code; if it doesn't work, try t=2 or 8.)
int64 t = 4, r = 1;
t <<= 1; r += ((x - r * r) & t) >> 1;
t <<= 1; r += ((x - r * r) & t) >> 1;
t <<= 1; r += ((x - r * r) & t) >> 1;
// Repeat until t is 2^33 or so. Use a loop if you want.
The idea is that at each iteration, you add one bit onto r, the "current" square root of x; each square root is accurate modulo a larger and larger power of 2, namely t/2. At the end, r and t/2-r will be square roots of x modulo t/2. (Note that if r is a square root of x, then so is -r. This is true even modulo numbers, but beware, modulo some numbers, things can have even more than 2 square roots; notably, this includes powers of 2.) Because our actual square root is less than 2^32, at that point we can actually just check if r or t/2-r are real square roots. In my actual code, I use the following modified loop:
int64 r, t, z;
r = start[(x >> 3) & 1023];
do {
z = x - r * r;
if( z == 0 )
return true;
if( z < 0 )
return false;
t = z & (-z);
r += (z & t) >> 1;
if( r > (t >> 1) )
r = t - r;
} while( t <= (1LL << 33) );
The speedup here is obtained in three ways: precomputed start value (equivalent to ~10 iterations of the loop), earlier exit of the loop, and skipping some t values. For the last part, I look at z = r - x * x, and set t to be the largest power of 2 dividing z with a bit trick. This allows me to skip t values that wouldn't have affected the value of r anyway. The precomputed start value in my case picks out the "smallest positive" square root modulo 8192.
Even if this code doesn't work faster for you, I hope you enjoy some of the ideas it contains. Complete, tested code follows, including the precomputed tables.
typedef signed long long int int64;
int start[1024] =
{1,3,1769,5,1937,1741,7,1451,479,157,9,91,945,659,1817,11,
1983,707,1321,1211,1071,13,1479,405,415,1501,1609,741,15,339,1703,203,
129,1411,873,1669,17,1715,1145,1835,351,1251,887,1573,975,19,1127,395,
1855,1981,425,453,1105,653,327,21,287,93,713,1691,1935,301,551,587,
257,1277,23,763,1903,1075,1799,1877,223,1437,1783,859,1201,621,25,779,
1727,573,471,1979,815,1293,825,363,159,1315,183,27,241,941,601,971,
385,131,919,901,273,435,647,1493,95,29,1417,805,719,1261,1177,1163,
1599,835,1367,315,1361,1933,1977,747,31,1373,1079,1637,1679,1581,1753,1355,
513,1539,1815,1531,1647,205,505,1109,33,1379,521,1627,1457,1901,1767,1547,
1471,1853,1833,1349,559,1523,967,1131,97,35,1975,795,497,1875,1191,1739,
641,1149,1385,133,529,845,1657,725,161,1309,375,37,463,1555,615,1931,
1343,445,937,1083,1617,883,185,1515,225,1443,1225,869,1423,1235,39,1973,
769,259,489,1797,1391,1485,1287,341,289,99,1271,1701,1713,915,537,1781,
1215,963,41,581,303,243,1337,1899,353,1245,329,1563,753,595,1113,1589,
897,1667,407,635,785,1971,135,43,417,1507,1929,731,207,275,1689,1397,
1087,1725,855,1851,1873,397,1607,1813,481,163,567,101,1167,45,1831,1205,
1025,1021,1303,1029,1135,1331,1017,427,545,1181,1033,933,1969,365,1255,1013,
959,317,1751,187,47,1037,455,1429,609,1571,1463,1765,1009,685,679,821,
1153,387,1897,1403,1041,691,1927,811,673,227,137,1499,49,1005,103,629,
831,1091,1449,1477,1967,1677,697,1045,737,1117,1737,667,911,1325,473,437,
1281,1795,1001,261,879,51,775,1195,801,1635,759,165,1871,1645,1049,245,
703,1597,553,955,209,1779,1849,661,865,291,841,997,1265,1965,1625,53,
1409,893,105,1925,1297,589,377,1579,929,1053,1655,1829,305,1811,1895,139,
575,189,343,709,1711,1139,1095,277,993,1699,55,1435,655,1491,1319,331,
1537,515,791,507,623,1229,1529,1963,1057,355,1545,603,1615,1171,743,523,
447,1219,1239,1723,465,499,57,107,1121,989,951,229,1521,851,167,715,
1665,1923,1687,1157,1553,1869,1415,1749,1185,1763,649,1061,561,531,409,907,
319,1469,1961,59,1455,141,1209,491,1249,419,1847,1893,399,211,985,1099,
1793,765,1513,1275,367,1587,263,1365,1313,925,247,1371,1359,109,1561,1291,
191,61,1065,1605,721,781,1735,875,1377,1827,1353,539,1777,429,1959,1483,
1921,643,617,389,1809,947,889,981,1441,483,1143,293,817,749,1383,1675,
63,1347,169,827,1199,1421,583,1259,1505,861,457,1125,143,1069,807,1867,
2047,2045,279,2043,111,307,2041,597,1569,1891,2039,1957,1103,1389,231,2037,
65,1341,727,837,977,2035,569,1643,1633,547,439,1307,2033,1709,345,1845,
1919,637,1175,379,2031,333,903,213,1697,797,1161,475,1073,2029,921,1653,
193,67,1623,1595,943,1395,1721,2027,1761,1955,1335,357,113,1747,1497,1461,
1791,771,2025,1285,145,973,249,171,1825,611,265,1189,847,1427,2023,1269,
321,1475,1577,69,1233,755,1223,1685,1889,733,1865,2021,1807,1107,1447,1077,
1663,1917,1129,1147,1775,1613,1401,555,1953,2019,631,1243,1329,787,871,885,
449,1213,681,1733,687,115,71,1301,2017,675,969,411,369,467,295,693,
1535,509,233,517,401,1843,1543,939,2015,669,1527,421,591,147,281,501,
577,195,215,699,1489,525,1081,917,1951,2013,73,1253,1551,173,857,309,
1407,899,663,1915,1519,1203,391,1323,1887,739,1673,2011,1585,493,1433,117,
705,1603,1111,965,431,1165,1863,533,1823,605,823,1179,625,813,2009,75,
1279,1789,1559,251,657,563,761,1707,1759,1949,777,347,335,1133,1511,267,
833,1085,2007,1467,1745,1805,711,149,1695,803,1719,485,1295,1453,935,459,
1151,381,1641,1413,1263,77,1913,2005,1631,541,119,1317,1841,1773,359,651,
961,323,1193,197,175,1651,441,235,1567,1885,1481,1947,881,2003,217,843,
1023,1027,745,1019,913,717,1031,1621,1503,867,1015,1115,79,1683,793,1035,
1089,1731,297,1861,2001,1011,1593,619,1439,477,585,283,1039,1363,1369,1227,
895,1661,151,645,1007,1357,121,1237,1375,1821,1911,549,1999,1043,1945,1419,
1217,957,599,571,81,371,1351,1003,1311,931,311,1381,1137,723,1575,1611,
767,253,1047,1787,1169,1997,1273,853,1247,413,1289,1883,177,403,999,1803,
1345,451,1495,1093,1839,269,199,1387,1183,1757,1207,1051,783,83,423,1995,
639,1155,1943,123,751,1459,1671,469,1119,995,393,219,1743,237,153,1909,
1473,1859,1705,1339,337,909,953,1771,1055,349,1993,613,1393,557,729,1717,
511,1533,1257,1541,1425,819,519,85,991,1693,503,1445,433,877,1305,1525,
1601,829,809,325,1583,1549,1991,1941,927,1059,1097,1819,527,1197,1881,1333,
383,125,361,891,495,179,633,299,863,285,1399,987,1487,1517,1639,1141,
1729,579,87,1989,593,1907,839,1557,799,1629,201,155,1649,1837,1063,949,
255,1283,535,773,1681,461,1785,683,735,1123,1801,677,689,1939,487,757,
1857,1987,983,443,1327,1267,313,1173,671,221,695,1509,271,1619,89,565,
127,1405,1431,1659,239,1101,1159,1067,607,1565,905,1755,1231,1299,665,373,
1985,701,1879,1221,849,627,1465,789,543,1187,1591,923,1905,979,1241,181};
bool bad255[512] =
{0,0,1,1,0,1,1,1,1,0,1,1,1,1,1,0,0,1,1,0,1,0,1,1,1,0,1,1,1,1,0,1,
1,1,0,1,0,1,1,1,1,1,1,1,1,1,1,1,1,0,1,0,1,1,1,0,1,1,1,1,0,1,1,1,
0,1,0,1,1,0,0,1,1,1,1,1,0,1,1,1,1,0,1,1,0,0,1,1,1,1,1,1,1,1,0,1,
1,1,1,1,0,1,1,1,1,1,0,1,1,1,1,0,1,1,1,0,1,1,1,1,0,0,1,1,1,1,1,1,
1,1,1,1,1,1,1,0,0,1,1,1,1,1,1,1,0,0,1,1,1,1,1,0,1,1,0,1,1,1,1,1,
1,1,1,1,1,1,0,1,1,0,1,0,1,1,0,1,1,1,1,1,1,1,1,1,1,1,0,1,1,0,1,1,
1,1,1,0,0,1,1,1,1,1,1,1,0,0,1,1,1,1,1,1,1,1,1,1,1,1,1,0,0,1,1,1,
1,0,1,1,1,0,1,1,1,1,0,1,1,1,1,1,0,1,1,1,1,1,0,1,1,1,1,1,1,1,1,
0,0,1,1,0,1,1,1,1,0,1,1,1,1,1,0,0,1,1,0,1,0,1,1,1,0,1,1,1,1,0,1,
1,1,0,1,0,1,1,1,1,1,1,1,1,1,1,1,1,0,1,0,1,1,1,0,1,1,1,1,0,1,1,1,
0,1,0,1,1,0,0,1,1,1,1,1,0,1,1,1,1,0,1,1,0,0,1,1,1,1,1,1,1,1,0,1,
1,1,1,1,0,1,1,1,1,1,0,1,1,1,1,0,1,1,1,0,1,1,1,1,0,0,1,1,1,1,1,1,
1,1,1,1,1,1,1,0,0,1,1,1,1,1,1,1,0,0,1,1,1,1,1,0,1,1,0,1,1,1,1,1,
1,1,1,1,1,1,0,1,1,0,1,0,1,1,0,1,1,1,1,1,1,1,1,1,1,1,0,1,1,0,1,1,
1,1,1,0,0,1,1,1,1,1,1,1,0,0,1,1,1,1,1,1,1,1,1,1,1,1,1,0,0,1,1,1,
1,0,1,1,1,0,1,1,1,1,0,1,1,1,1,1,0,1,1,1,1,1,0,1,1,1,1,1,1,1,1,
0,0};
inline bool square( int64 x ) {
// Quickfail
if( x < 0 || (x&2) || ((x & 7) == 5) || ((x & 11) == 8) )
return false;
if( x == 0 )
return true;
// Check mod 255 = 3 * 5 * 17, for fun
int64 y = x;
y = (y & 4294967295LL) + (y >> 32);
y = (y & 65535) + (y >> 16);
y = (y & 255) + ((y >> 8) & 255) + (y >> 16);
if( bad255[y] )
return false;
// Divide out powers of 4 using binary search
if((x & 4294967295LL) == 0)
x >>= 32;
if((x & 65535) == 0)
x >>= 16;
if((x & 255) == 0)
x >>= 8;
if((x & 15) == 0)
x >>= 4;
if((x & 3) == 0)
x >>= 2;
if((x & 7) != 1)
return false;
// Compute sqrt using something like Hensel's lemma
int64 r, t, z;
r = start[(x >> 3) & 1023];
do {
z = x - r * r;
if( z == 0 )
return true;
if( z < 0 )
return false;
t = z & (-z);
r += (z & t) >> 1;
if( r > (t >> 1) )
r = t - r;
} while( t <= (1LL << 33) );
return false;
}

I'm pretty late to the party, but I hope to provide a better answer; shorter and (assuming my benchmark is correct) also much faster.
long goodMask; // 0xC840C04048404040 computed below
{
for (int i=0; i<64; ++i) goodMask |= Long.MIN_VALUE >>> (i*i);
}
public boolean isSquare(long x) {
// This tests if the 6 least significant bits are right.
// Moving the to be tested bit to the highest position saves us masking.
if (goodMask << x >= 0) return false;
final int numberOfTrailingZeros = Long.numberOfTrailingZeros(x);
// Each square ends with an even number of zeros.
if ((numberOfTrailingZeros & 1) != 0) return false;
x >>= numberOfTrailingZeros;
// Now x is either 0 or odd.
// In binary each odd square ends with 001.
// Postpone the sign test until now; handle zero in the branch.
if ((x&7) != 1 | x <= 0) return x == 0;
// Do it in the classical way.
// The correctness is not trivial as the conversion from long to double is lossy!
final long tst = (long) Math.sqrt(x);
return tst * tst == x;
}
The first test catches most non-squares quickly. It uses a 64-item table packed in a long, so there's no array access cost (indirection and bounds checks). For a uniformly random long, there's a 81.25% probability of ending here.
The second test catches all numbers having an odd number of twos in their factorization. The method Long.numberOfTrailingZeros is very fast as it gets JIT-ed into a single i86 instruction.
After dropping the trailing zeros, the third test handles numbers ending with 011, 101, or 111 in binary, which are no perfect squares. It also cares about negative numbers and also handles 0.
The final test falls back to double arithmetic. As double has only 53 bits mantissa,
the conversion from long to double includes rounding for big values. Nonetheless, the test is correct (unless the proof is wrong).
Trying to incorporate the mod255 idea wasn't successful.

You'll have to do some benchmarking. The best algorithm will depend on the distribution of your inputs.
Your algorithm may be nearly optimal, but you might want to do a quick check to rule out some possibilities before calling your square root routine. For example, look at the last digit of your number in hex by doing a bit-wise "and." Perfect squares can only end in 0, 1, 4, or 9 in base 16, So for 75% of your inputs (assuming they are uniformly distributed) you can avoid a call to the square root in exchange for some very fast bit twiddling.
Kip benchmarked the following code implementing the hex trick. When testing numbers 1 through 100,000,000, this code ran twice as fast as the original.
public final static boolean isPerfectSquare(long n)
{
if (n < 0)
return false;
switch((int)(n & 0xF))
{
case 0: case 1: case 4: case 9:
long tst = (long)Math.sqrt(n);
return tst*tst == n;
default:
return false;
}
}
When I tested the analogous code in C++, it actually ran slower than the original. However, when I eliminated the switch statement, the hex trick once again make the code twice as fast.
int isPerfectSquare(int n)
{
int h = n & 0xF; // h is the last hex "digit"
if (h > 9)
return 0;
// Use lazy evaluation to jump out of the if statement as soon as possible
if (h != 2 && h != 3 && h != 5 && h != 6 && h != 7 && h != 8)
{
int t = (int) floor( sqrt((double) n) + 0.5 );
return t*t == n;
}
return 0;
}
Eliminating the switch statement had little effect on the C# code.

I was thinking about the horrible times I've spent in Numerical Analysis course.
And then I remember, there was this function circling around the 'net from the Quake Source code:
float Q_rsqrt( float number )
{
long i;
float x2, y;
const float threehalfs = 1.5F;
x2 = number * 0.5F;
y = number;
i = * ( long * ) &y; // evil floating point bit level hacking
i = 0x5f3759df - ( i >> 1 ); // wtf?
y = * ( float * ) &i;
y = y * ( threehalfs - ( x2 * y * y ) ); // 1st iteration
// y = y * ( threehalfs - ( x2 * y * y ) ); // 2nd iteration, this can be removed
#ifndef Q3_VM
#ifdef __linux__
assert( !isnan(y) ); // bk010122 - FPE?
#endif
#endif
return y;
}
Which basically calculates a square root, using Newton's approximation function (cant remember the exact name).
It should be usable and might even be faster, it's from one of the phenomenal id software's game!
It's written in C++ but it should not be too hard to reuse the same technique in Java once you get the idea:
I originally found it at: http://www.codemaestro.com/reviews/9
Newton's method explained at wikipedia: http://en.wikipedia.org/wiki/Newton%27s_method
You can follow the link for more explanation of how it works, but if you don't care much, then this is roughly what I remember from reading the blog and from taking the Numerical Analysis course:
the * (long*) &y is basically a fast convert-to-long function so integer operations can be applied on the raw bytes.
the 0x5f3759df - (i >> 1); line is a pre-calculated seed value for the approximation function.
the * (float*) &i converts the value back to floating point.
the y = y * ( threehalfs - ( x2 * y * y ) ) line bascially iterates the value over the function again.
The approximation function gives more precise values the more you iterate the function over the result. In Quake's case, one iteration is "good enough", but if it wasn't for you... then you could add as much iteration as you need.
This should be faster because it reduces the number of division operations done in naive square rooting down to a simple divide by 2 (actually a * 0.5F multiply operation) and replace it with a few fixed number of multiplication operations instead.

I'm not sure if it would be faster, or even accurate, but you could use John Carmack's Magical Square Root, algorithm to solve the square root faster. You could probably easily test this for all possible 32 bit integers, and validate that you actually got correct results, as it's only an appoximation. However, now that I think about it, using doubles is approximating also, so I'm not sure how that would come into play.

If you do a binary chop to try to find the "right" square root, you can fairly easily detect if the value you've got is close enough to tell:
(n+1)^2 = n^2 + 2n + 1
(n-1)^2 = n^2 - 2n + 1
So having calculated n^2, the options are:
n^2 = target: done, return true
n^2 + 2n + 1 > target > n^2 : you're close, but it's not perfect: return false
n^2 - 2n + 1 < target < n^2 : ditto
target < n^2 - 2n + 1 : binary chop on a lower n
target > n^2 + 2n + 1 : binary chop on a higher n
(Sorry, this uses n as your current guess, and target for the parameter. Apologise for the confusion!)
I don't know whether this will be faster or not, but it's worth a try.
EDIT: The binary chop doesn't have to take in the whole range of integers, either (2^x)^2 = 2^(2x), so once you've found the top set bit in your target (which can be done with a bit-twiddling trick; I forget exactly how) you can quickly get a range of potential answers. Mind you, a naive binary chop is still only going to take up to 31 or 32 iterations.

I ran my own analysis of several of the algorithms in this thread and came up with some new results. You can see those old results in the edit history of this answer, but they're not accurate, as I made a mistake, and wasted time analyzing several algorithms which aren't close. However, pulling lessons from several different answers, I now have two algorithms that crush the "winner" of this thread. Here's the core thing I do differently than everyone else:
// This is faster because a number is divisible by 2^4 or more only 6% of the time
// and more than that a vanishingly small percentage.
while((x & 0x3) == 0) x >>= 2;
// This is effectively the same as the switch-case statement used in the original
// answer.
if((x & 0x7) != 1) return false;
However, this simple line, which most of the time adds one or two very fast instructions, greatly simplifies the switch-case statement into one if statement. However, it can add to the runtime if many of the tested numbers have significant power-of-two factors.
The algorithms below are as follows:
Internet - Kip's posted answer
Durron - My modified answer using the one-pass answer as a base
DurronTwo - My modified answer using the two-pass answer (by #JohnnyHeggheim), with some other slight modifications.
Here is a sample runtime if the numbers are generated using Math.abs(java.util.Random.nextLong())
0% Scenario{vm=java, trial=0, benchmark=Internet} 39673.40 ns; ?=378.78 ns # 3 trials
33% Scenario{vm=java, trial=0, benchmark=Durron} 37785.75 ns; ?=478.86 ns # 10 trials
67% Scenario{vm=java, trial=0, benchmark=DurronTwo} 35978.10 ns; ?=734.10 ns # 10 trials
benchmark us linear runtime
Internet 39.7 ==============================
Durron 37.8 ============================
DurronTwo 36.0 ===========================
vm: java
trial: 0
And here is a sample runtime if it's run on the first million longs only:
0% Scenario{vm=java, trial=0, benchmark=Internet} 2933380.84 ns; ?=56939.84 ns # 10 trials
33% Scenario{vm=java, trial=0, benchmark=Durron} 2243266.81 ns; ?=50537.62 ns # 10 trials
67% Scenario{vm=java, trial=0, benchmark=DurronTwo} 3159227.68 ns; ?=10766.22 ns # 3 trials
benchmark ms linear runtime
Internet 2.93 ===========================
Durron 2.24 =====================
DurronTwo 3.16 ==============================
vm: java
trial: 0
As you can see, DurronTwo does better for large inputs, because it gets to use the magic trick very very often, but gets clobbered compared to the first algorithm and Math.sqrt because the numbers are so much smaller. Meanwhile, the simpler Durron is a huge winner because it never has to divide by 4 many many times in the first million numbers.
Here's Durron:
public final static boolean isPerfectSquareDurron(long n) {
if(n < 0) return false;
if(n == 0) return true;
long x = n;
// This is faster because a number is divisible by 16 only 6% of the time
// and more than that a vanishingly small percentage.
while((x & 0x3) == 0) x >>= 2;
// This is effectively the same as the switch-case statement used in the original
// answer.
if((x & 0x7) == 1) {
long sqrt;
if(x < 410881L)
{
int i;
float x2, y;
x2 = x * 0.5F;
y = x;
i = Float.floatToRawIntBits(y);
i = 0x5f3759df - ( i >> 1 );
y = Float.intBitsToFloat(i);
y = y * ( 1.5F - ( x2 * y * y ) );
sqrt = (long)(1.0F/y);
} else {
sqrt = (long) Math.sqrt(x);
}
return sqrt*sqrt == x;
}
return false;
}
And DurronTwo
public final static boolean isPerfectSquareDurronTwo(long n) {
if(n < 0) return false;
// Needed to prevent infinite loop
if(n == 0) return true;
long x = n;
while((x & 0x3) == 0) x >>= 2;
if((x & 0x7) == 1) {
long sqrt;
if (x < 41529141369L) {
int i;
float x2, y;
x2 = x * 0.5F;
y = x;
i = Float.floatToRawIntBits(y);
//using the magic number from
//http://www.lomont.org/Math/Papers/2003/InvSqrt.pdf
//since it more accurate
i = 0x5f375a86 - (i >> 1);
y = Float.intBitsToFloat(i);
y = y * (1.5F - (x2 * y * y));
y = y * (1.5F - (x2 * y * y)); //Newton iteration, more accurate
sqrt = (long) ((1.0F/y) + 0.2);
} else {
//Carmack hack gives incorrect answer for n >= 41529141369.
sqrt = (long) Math.sqrt(x);
}
return sqrt*sqrt == x;
}
return false;
}
And my benchmark harness: (Requires Google caliper 0.1-rc5)
public class SquareRootBenchmark {
public static class Benchmark1 extends SimpleBenchmark {
private static final int ARRAY_SIZE = 10000;
long[] trials = new long[ARRAY_SIZE];
#Override
protected void setUp() throws Exception {
Random r = new Random();
for (int i = 0; i < ARRAY_SIZE; i++) {
trials[i] = Math.abs(r.nextLong());
}
}
public int timeInternet(int reps) {
int trues = 0;
for(int i = 0; i < reps; i++) {
for(int j = 0; j < ARRAY_SIZE; j++) {
if(SquareRootAlgs.isPerfectSquareInternet(trials[j])) trues++;
}
}
return trues;
}
public int timeDurron(int reps) {
int trues = 0;
for(int i = 0; i < reps; i++) {
for(int j = 0; j < ARRAY_SIZE; j++) {
if(SquareRootAlgs.isPerfectSquareDurron(trials[j])) trues++;
}
}
return trues;
}
public int timeDurronTwo(int reps) {
int trues = 0;
for(int i = 0; i < reps; i++) {
for(int j = 0; j < ARRAY_SIZE; j++) {
if(SquareRootAlgs.isPerfectSquareDurronTwo(trials[j])) trues++;
}
}
return trues;
}
}
public static void main(String... args) {
Runner.main(Benchmark1.class, args);
}
}
UPDATE: I've made a new algorithm that is faster in some scenarios, slower in others, I've gotten different benchmarks based on different inputs. If we calculate modulo 0xFFFFFF = 3 x 3 x 5 x 7 x 13 x 17 x 241, we can eliminate 97.82% of numbers that cannot be squares. This can be (sort of) done in one line, with 5 bitwise operations:
if (!goodLookupSquares[(int) ((n & 0xFFFFFFl) + ((n >> 24) & 0xFFFFFFl) + (n >> 48))]) return false;
The resulting index is either 1) the residue, 2) the residue + 0xFFFFFF, or 3) the residue + 0x1FFFFFE. Of course, we need to have a lookup table for residues modulo 0xFFFFFF, which is about a 3mb file (in this case stored as ascii text decimal numbers, not optimal but clearly improvable with a ByteBuffer and so forth. But since that is precalculation it doesn't matter so much. You can find the file here (or generate it yourself):
public final static boolean isPerfectSquareDurronThree(long n) {
if(n < 0) return false;
if(n == 0) return true;
long x = n;
while((x & 0x3) == 0) x >>= 2;
if((x & 0x7) == 1) {
if (!goodLookupSquares[(int) ((n & 0xFFFFFFl) + ((n >> 24) & 0xFFFFFFl) + (n >> 48))]) return false;
long sqrt;
if(x < 410881L)
{
int i;
float x2, y;
x2 = x * 0.5F;
y = x;
i = Float.floatToRawIntBits(y);
i = 0x5f3759df - ( i >> 1 );
y = Float.intBitsToFloat(i);
y = y * ( 1.5F - ( x2 * y * y ) );
sqrt = (long)(1.0F/y);
} else {
sqrt = (long) Math.sqrt(x);
}
return sqrt*sqrt == x;
}
return false;
}
I load it into a boolean array like this:
private static boolean[] goodLookupSquares = null;
public static void initGoodLookupSquares() throws Exception {
Scanner s = new Scanner(new File("24residues_squares.txt"));
goodLookupSquares = new boolean[0x1FFFFFE];
while(s.hasNextLine()) {
int residue = Integer.valueOf(s.nextLine());
goodLookupSquares[residue] = true;
goodLookupSquares[residue + 0xFFFFFF] = true;
goodLookupSquares[residue + 0x1FFFFFE] = true;
}
s.close();
}
Example runtime. It beat Durron (version one) in every trial I ran.
0% Scenario{vm=java, trial=0, benchmark=Internet} 40665.77 ns; ?=566.71 ns # 10 trials
33% Scenario{vm=java, trial=0, benchmark=Durron} 38397.60 ns; ?=784.30 ns # 10 trials
67% Scenario{vm=java, trial=0, benchmark=DurronThree} 36171.46 ns; ?=693.02 ns # 10 trials
benchmark us linear runtime
Internet 40.7 ==============================
Durron 38.4 ============================
DurronThree 36.2 ==========================
vm: java
trial: 0

It should be much faster to use Newton's method to calculate the Integer Square Root, then square this number and check, as you do in your current solution. Newton's method is the basis for the Carmack solution mentioned in some other answers. You should be able to get a faster answer since you're only interested in the integer part of the root, allowing you to stop the approximation algorithm sooner.
Another optimization that you can try: If the Digital Root of a number doesn't end in
1, 4, 7, or 9 the number is not a perfect square. This can be used as a quick way to eliminate 60% of your inputs before applying the slower square root algorithm.

I want this function to work with all
positive 64-bit signed integers
Math.sqrt() works with doubles as input parameters, so you won't get accurate results for integers bigger than 2^53.

An integer problem deserves an integer solution. Thus
Do binary search on the (non-negative) integers to find the greatest integer t such that t**2 <= n. Then test whether r**2 = n exactly. This takes time O(log n).
If you don't know how to binary search the positive integers because the set is unbounded, it's easy. You starting by computing your increasing function f (above f(t) = t**2 - n) on powers of two. When you see it turn positive, you've found an upper bound. Then you can do standard binary search.

Just for the record, another approach is to use the prime decomposition. If every factor of the decomposition is even, then the number is a perfect square. So what you want is to see if a number can be decomposed as a product of squares of prime numbers. Of course, you don't need to obtain such a decomposition, just to see if it exists.
First build a table of squares of prime numbers which are lower than 2^32. This is far smaller than a table of all integers up to this limit.
A solution would then be like this:
boolean isPerfectSquare(long number)
{
if (number < 0) return false;
if (number < 2) return true;
for (int i = 0; ; i++)
{
long square = squareTable[i];
if (square > number) return false;
while (number % square == 0)
{
number /= square;
}
if (number == 1) return true;
}
}
I guess it's a bit cryptic. What it does is checking in every step that the square of a prime number divide the input number. If it does then it divides the number by the square as long as it is possible, to remove this square from the prime decomposition.
If by this process, we came to 1, then the input number was a decomposition of square of prime numbers. If the square becomes larger than the number itself, then there is no way this square, or any larger squares, can divide it, so the number can not be a decomposition of squares of prime numbers.
Given nowadays' sqrt done in hardware and the need to compute prime numbers here, I guess this solution is way slower. But it should give better results than solution with sqrt which won't work over 2^54, as says mrzl in his answer.

It's been pointed out that the last d digits of a perfect square can only take on certain values. The last d digits (in base b) of a number n is the same as the remainder when n is divided by bd, ie. in C notation n % pow(b, d).
This can be generalized to any modulus m, ie. n % m can be used to rule out some percentage of numbers from being perfect squares. The modulus you are currently using is 64, which allows 12, ie. 19% of remainders, as possible squares. With a little coding I found the modulus 110880, which allows only 2016, ie. 1.8% of remainders as possible squares. So depending on the cost of a modulus operation (ie. division) and a table lookup versus a square root on your machine, using this modulus might be faster.
By the way if Java has a way to store a packed array of bits for the lookup table, don't use it. 110880 32-bit words is not much RAM these days and fetching a machine word is going to be faster than fetching a single bit.

The following simplification of maaartinus's solution appears to shave a few percentage points off the runtime, but I'm not good enough at benchmarking to produce a benchmark I can trust:
long goodMask; // 0xC840C04048404040 computed below
{
for (int i=0; i<64; ++i) goodMask |= Long.MIN_VALUE >>> (i*i);
}
public boolean isSquare(long x) {
// This tests if the 6 least significant bits are right.
// Moving the to be tested bit to the highest position saves us masking.
if (goodMask << x >= 0) return false;
// Remove an even number of trailing zeros, leaving at most one.
x >>= (Long.numberOfTrailingZeros(x) & (-2);
// Repeat the test on the 6 least significant remaining bits.
if (goodMask << x >= 0 | x <= 0) return x == 0;
// Do it in the classical way.
// The correctness is not trivial as the conversion from long to double is lossy!
final long tst = (long) Math.sqrt(x);
return tst * tst == x;
}
It would be worth checking how omitting the first test,
if (goodMask << x >= 0) return false;
would affect performance.

For performance, you very often have to do some compromsies. Others have expressed various methods, however, you noted Carmack's hack was faster up to certain values of N. Then, you should check the "n" and if it is less than that number N, use Carmack's hack, else use some other method described in the answers here.

This is the fastest Java implementation I could come up with, using a combination of techniques suggested by others in this thread.
Mod-256 test
Inexact mod-3465 test (avoids integer division at the cost of some false positives)
Floating-point square root, round and compare with input value
I also experimented with these modifications but they did not help performance:
Additional mod-255 test
Dividing the input value by powers of 4
Fast Inverse Square Root (to work for high values of N it needs 3 iterations, enough to make it slower than the hardware square root function.)
public class SquareTester {
public static boolean isPerfectSquare(long n) {
if (n < 0) {
return false;
} else {
switch ((byte) n) {
case -128: case -127: case -124: case -119: case -112:
case -111: case -103: case -95: case -92: case -87:
case -79: case -71: case -64: case -63: case -60:
case -55: case -47: case -39: case -31: case -28:
case -23: case -15: case -7: case 0: case 1:
case 4: case 9: case 16: case 17: case 25:
case 33: case 36: case 41: case 49: case 57:
case 64: case 65: case 68: case 73: case 81:
case 89: case 97: case 100: case 105: case 113:
case 121:
long i = (n * INV3465) >>> 52;
if (! good3465[(int) i]) {
return false;
} else {
long r = round(Math.sqrt(n));
return r*r == n;
}
default:
return false;
}
}
}
private static int round(double x) {
return (int) Double.doubleToRawLongBits(x + (double) (1L << 52));
}
/** 3465<sup>-1</sup> modulo 2<sup>64</sup> */
private static final long INV3465 = 0x8ffed161732e78b9L;
private static final boolean[] good3465 =
new boolean[0x1000];
static {
for (int r = 0; r < 3465; ++ r) {
int i = (int) ((r * r * INV3465) >>> 52);
good3465[i] = good3465[i+1] = true;
}
}
}

You should get rid of the 2-power part of N right from the start.
2nd Edit
The magical expression for m below should be
m = N - (N & (N-1));
and not as written
End of 2nd edit
m = N & (N-1); // the lawest bit of N
N /= m;
byte = N & 0x0F;
if ((m % 2) || (byte !=1 && byte !=9))
return false;
1st Edit:
Minor improvement:
m = N & (N-1); // the lawest bit of N
N /= m;
if ((m % 2) || (N & 0x07 != 1))
return false;
End of 1st edit
Now continue as usual. This way, by the time you get to the floating point part, you already got rid of all the numbers whose 2-power part is odd (about half), and then you only consider 1/8 of whats left. I.e. you run the floating point part on 6% of the numbers.

Project Euler is mentioned in the tags and many of the problems in it require checking numbers >> 2^64. Most of the optimizations mentioned above don't work easily when you are working with an 80 byte buffer.
I used java BigInteger and a slightly modified version of Newton's method, one that works better with integers. The problem was that exact squares n^2 converged to (n-1) instead of n because n^2-1 = (n-1)(n+1) and the final error was just one step below the final divisor and the algorithm terminated. It was easy to fix by adding one to the original argument before computing the error. (Add two for cube roots, etc.)
One nice attribute of this algorithm is that you can immediately tell if the number is a perfect square - the final error (not correction) in Newton's method will be zero. A simple modification also lets you quickly calculate floor(sqrt(x)) instead of the closest integer. This is handy with several Euler problems.

The sqrt call is not perfectly accurate, as has been mentioned, but it's interesting and instructive that it doesn't blow away the other answers in terms of speed. After all, the sequence of assembly language instructions for a sqrt is tiny. Intel has a hardware instruction, which isn't used by Java I believe because it doesn't conform to IEEE.
So why is it slow? Because Java is actually calling a C routine through JNI, and it's actually slower to do so than to call a Java subroutine, which itself is slower than doing it inline. This is very annoying, and Java should have come up with a better solution, ie building in floating point library calls if necessary. Oh well.
In C++, I suspect all the complex alternatives would lose on speed, but I haven't checked them all.
What I did, and what Java people will find usefull, is a simple hack, an extension of the special case testing suggested by A. Rex. Use a single long value as a bit array, which isn't bounds checked. That way, you have 64 bit boolean lookup.
typedef unsigned long long UVLONG
UVLONG pp1,pp2;
void init2() {
for (int i = 0; i < 64; i++) {
for (int j = 0; j < 64; j++)
if (isPerfectSquare(i * 64 + j)) {
pp1 |= (1 << j);
pp2 |= (1 << i);
break;
}
}
cout << "pp1=" << pp1 << "," << pp2 << "\n";
}
inline bool isPerfectSquare5(UVLONG x) {
return pp1 & (1 << (x & 0x3F)) ? isPerfectSquare(x) : false;
}
The routine isPerfectSquare5 runs in about 1/3 the time on my core2 duo machine. I suspect that further tweaks along the same lines could reduce the time further on average, but every time you check, you are trading off more testing for more eliminating, so you can't go too much farther on that road.
Certainly, rather than having a separate test for negative, you could check the high 6 bits the same way.
Note that all I'm doing is eliminating possible squares, but when I have a potential case I have to call the original, inlined isPerfectSquare.
The init2 routine is called once to initialize the static values of pp1 and pp2.
Note that in my implementation in C++, I'm using unsigned long long, so since you're signed, you'd have to use the >>> operator.
There is no intrinsic need to bounds check the array, but Java's optimizer has to figure this stuff out pretty quickly, so I don't blame them for that.

I like the idea to use an almost correct method on some of the input. Here is a version with a higher "offset". The code seems to work and passes my simple test case.
Just replace your:
if(n < 410881L){...}
code with this one:
if (n < 11043908100L) {
//John Carmack hack, converted to Java.
// See: http://www.codemaestro.com/reviews/9
int i;
float x2, y;
x2 = n * 0.5F;
y = n;
i = Float.floatToRawIntBits(y);
//using the magic number from
//http://www.lomont.org/Math/Papers/2003/InvSqrt.pdf
//since it more accurate
i = 0x5f375a86 - (i >> 1);
y = Float.intBitsToFloat(i);
y = y * (1.5F - (x2 * y * y));
y = y * (1.5F - (x2 * y * y)); //Newton iteration, more accurate
sqrt = Math.round(1.0F / y);
} else {
//Carmack hack gives incorrect answer for n >= 11043908100.
sqrt = (long) Math.sqrt(n);
}

Considering for general bit length (though I have used specific type here), I tried to design simplistic algo as below. Simple and obvious check for 0,1,2 or <0 is required initially.
Following is simple in sense that it doesn't try to use any existing maths functions. Most of the operator can be replaced with bit-wise operators. I haven't tested with any bench mark data though. I'm neither expert at maths or computer algorithm design in particular, I would love to see you pointing out problem. I know there is lots of improvement chances there.
int main()
{
unsigned int c1=0 ,c2 = 0;
unsigned int x = 0;
unsigned int p = 0;
int k1 = 0;
scanf("%d",&p);
if(p % 2 == 0) {
x = p/2;
}
else {
x = (p/2) +1;
}
while(x)
{
if((x*x) > p) {
c1 = x;
x = x/2;
}else {
c2 = x;
break;
}
}
if((p%2) != 0)
c2++;
while(c2 < c1)
{
if((c2 * c2 ) == p) {
k1 = 1;
break;
}
c2++;
}
if(k1)
printf("\n Perfect square for %d", c2);
else
printf("\n Not perfect but nearest to :%d :", c2);
return 0;
}

This a rework from decimal to binary of the old Marchant calculator algorithm (sorry, I don't have a reference), in Ruby, adapted specifically for this question:
def isexactsqrt(v)
value = v.abs
residue = value
root = 0
onebit = 1
onebit <<= 8 while (onebit < residue)
onebit >>= 2 while (onebit > residue)
while (onebit > 0)
x = root + onebit
if (residue >= x) then
residue -= x
root = x + onebit
end
root >>= 1
onebit >>= 2
end
return (residue == 0)
end
Here's a workup of something similar (there may be coding style/smells or clunky O/O - it's the algorithm that counts, and C++ is not my home language). In this case, we're looking for residue == 0:
#include <iostream>
using namespace std;
typedef unsigned long long int llint;
class ISqrt { // Integer Square Root
llint value; // Integer whose square root is required
llint root; // Result: floor(sqrt(value))
llint residue; // Result: value-root*root
llint onebit, x; // Working bit, working value
public:
ISqrt(llint v = 2) { // Constructor
Root(v); // Take the root
};
llint Root(llint r) { // Resets and calculates new square root
value = r; // Store input
residue = value; // Initialise for subtracting down
root = 0; // Clear root accumulator
onebit = 1; // Calculate start value of counter
onebit <<= (8*sizeof(llint)-2); // Set up counter bit as greatest odd power of 2
while (onebit > residue) {onebit >>= 2; }; // Shift down until just < value
while (onebit > 0) {
x = root ^ onebit; // Will check root+1bit (root bit corresponding to onebit is always zero)
if (residue >= x) { // Room to subtract?
residue -= x; // Yes - deduct from residue
root = x + onebit; // and step root
};
root >>= 1;
onebit >>= 2;
};
return root;
};
llint Residue() { // Returns residue from last calculation
return residue;
};
};
int main() {
llint big, i, q, r, v, delta;
big = 0; big = (big-1); // Kludge for "big number"
ISqrt b; // Make q sqrt generator
for ( i = big; i > 0 ; i /= 7 ) { // for several numbers
q = b.Root(i); // Get the square root
r = b.Residue(); // Get the residue
v = q*q+r; // Recalc original value
delta = v-i; // And diff, hopefully 0
cout << i << ": " << q << " ++ " << r << " V: " << v << " Delta: " << delta << "\n";
};
return 0;
};

I checked all of the possible results when the last n bits of a square is observed. By successively examining more bits, up to 5/6th of inputs can be eliminated. I actually designed this to implement Fermat's Factorization algorithm, and it is very fast there.
public static boolean isSquare(final long val) {
if ((val & 2) == 2 || (val & 7) == 5) {
return false;
}
if ((val & 11) == 8 || (val & 31) == 20) {
return false;
}
if ((val & 47) == 32 || (val & 127) == 80) {
return false;
}
if ((val & 191) == 128 || (val & 511) == 320) {
return false;
}
// if((val & a == b) || (val & c == d){
// return false;
// }
if (!modSq[(int) (val % modSq.length)]) {
return false;
}
final long root = (long) Math.sqrt(val);
return root * root == val;
}
The last bit of pseudocode can be used to extend the tests to eliminate more values. The tests above are for k = 0, 1, 2, 3
a is of the form (3 << 2k) - 1
b is of the form (2 << 2k)
c is of the form (2 << 2k + 2) - 1
d is of the form (2 << 2k - 1) * 10
It first tests whether it has a square residual with moduli of power of two, then it tests based on a final modulus, then it uses the Math.sqrt to do a final test. I came up with the idea from the top post, and attempted to extend upon it. I appreciate any comments or suggestions.
Update: Using the test by a modulus, (modSq) and a modulus base of 44352, my test runs in 96% of the time of the one in the OP's update for numbers up to 1,000,000,000.

Here is a divide and conquer solution.
If the square root of a natural number (number) is a natural number (solution), you can easily determine a range for solution based on the number of digits of number:
number has 1 digit: solution in range = 1 - 4
number has 2 digits: solution in range = 3 - 10
number has 3 digits: solution in range = 10 - 40
number has 4 digits: solution in range = 30 - 100
number has 5 digits: solution in range = 100 - 400
Notice the repetition?
You can use this range in a binary search approach to see if there is a solution for which:
number == solution * solution
Here is the code
Here is my class SquareRootChecker
public class SquareRootChecker {
private long number;
private long initialLow;
private long initialHigh;
public SquareRootChecker(long number) {
this.number = number;
initialLow = 1;
initialHigh = 4;
if (Long.toString(number).length() % 2 == 0) {
initialLow = 3;
initialHigh = 10;
}
for (long i = 0; i < Long.toString(number).length() / 2; i++) {
initialLow *= 10;
initialHigh *= 10;
}
if (Long.toString(number).length() % 2 == 0) {
initialLow /= 10;
initialHigh /=10;
}
}
public boolean checkSquareRoot() {
return findSquareRoot(initialLow, initialHigh, number);
}
private boolean findSquareRoot(long low, long high, long number) {
long check = low + (high - low) / 2;
if (high >= low) {
if (number == check * check) {
return true;
}
else if (number < check * check) {
high = check - 1;
return findSquareRoot(low, high, number);
}
else {
low = check + 1;
return findSquareRoot(low, high, number);
}
}
return false;
}
}
And here is an example on how to use it.
long number = 1234567;
long square = number * number;
SquareRootChecker squareRootChecker = new SquareRootChecker(square);
System.out.println(square + ": " + squareRootChecker.checkSquareRoot()); //Prints "1524155677489: true"
long notSquare = square + 1;
squareRootChecker = new SquareRootChecker(notSquare);
System.out.println(notSquare + ": " + squareRootChecker.checkSquareRoot()); //Prints "1524155677490: false"

Newton's Method with integer arithmetic
If you wish to avoid non-integer operations you could use the method below. It basically uses Newton's Method modified for integer arithmetic.
/**
* Test if the given number is a perfect square.
* #param n Must be greater than 0 and less
* than Long.MAX_VALUE.
* #return <code>true</code> if n is a perfect
* square, or <code>false</code> otherwise.
*/
public static boolean isSquare(long n)
{
long x1 = n;
long x2 = 1L;
while (x1 > x2)
{
x1 = (x1 + x2) / 2L;
x2 = n / x1;
}
return x1 == x2 && n % x1 == 0L;
}
This implementation can not compete with solutions that use Math.sqrt. However, its performance can be improved by using the filtering mechanisms described in some of the other posts.

Square Root of a number, given that the number is a perfect square.
The complexity is log(n)
/**
* Calculate square root if the given number is a perfect square.
*
* Approach: Sum of n odd numbers is equals to the square root of n*n, given
* that n is a perfect square.
*
* #param number
* #return squareRoot
*/
public static int calculateSquareRoot(int number) {
int sum=1;
int count =1;
int squareRoot=1;
while(sum<number) {
count+=2;
sum+=count;
squareRoot++;
}
return squareRoot;
}

Here is the simplest and most concise way, although I do not know how it compares in terms of CPU cycles. This works great if you only wish to know if the root is a whole number. If you really care if it is an integer, you can also figure that out. Here is a simple (and pure) function:
private static final MathContext precision = new MathContext(20);
private static final Function<Long, Boolean> isRootWhole = (n) -> {
long digit = n % 10;
if (digit == 2 || digit == 3 || digit == 7 || digit == 8) {
return false;
}
return new BigDecimal(n).sqrt(precision).scale() == 0;
};
If you do not need micro-optimization, this answer is better in terms of simplicity and maintainability. If you will be calculating negative numbers, you will need to handle that accordingly, and send the absolute value into the function. I have included a minor optimization because no perfect squares have a tens digit of 2, 3, 7, or 8 due to quadratic residues mod 10.
On my CPU, a run of this algorithm on 0 - 10,000,000 took an average of 1000 - 1100 nanoseconds per calculation.
If you are performing a lesser number of calculations, the earlier calculations take a bit longer.
I had a negative comment that my previous edit did not work for large numbers. The OP mentioned Longs, and the largest perfect square that is a Long is 9223372030926249001, so this method works for all Longs.

This question got me wondering, so I did some simple coding and I'm presenting it here because I think it's interesting, relevant, but I don't know how useful. There's a simple algorithm
a_n+1 = (a_n + x/a_n)/2
for calculating square roots, but it's meant to be used for decimals. I wondered what would happen if I just coded the same algorithm using integer maths. Would it even converge on the right answer? I didn't know, so I wrote a program...
#include <stdio.h>
#include <stdint.h>
#include <stdlib.h>
#include <math.h>
_Bool isperfectsquare(uint64_t x, uint64_t *isqrtx) {
// NOTE: isqrtx approximate for non-squares. (benchmarked at 162ns 3GHz i5)
uint32_t i;
uint64_t ai;
ai = 1 + ((x & 0xffff000000000000) >> 32) + ((x & 0xffff00000000) >> 24) + ((x & 0xffff0000) >> 16);
ai = (ai + x/ai)/2;
ai = (ai + x/ai)/2;
ai = (ai + x/ai)/2;
ai = (ai + x/ai)/2;
ai = (ai + x/ai)/2;
ai = (ai + x/ai)/2;
ai = (ai + x/ai)/2;
ai = (ai + x/ai)/2;
ai = (ai + x/ai)/2;
ai = (ai + x/ai)/2;
ai = (ai + x/ai)/2;
ai = (ai + x/ai)/2;
ai = ai & 0xffffffff;
if (isqrtx != NULL) isqrtx[0] = ai;
return ai*ai == x;
}
void main() {
uint64_t x, isqrtx;
uint64_t i;
for (i=1; i<0x100000000; i++) {
if (!isperfectsquare(i*i, &isqrtx)) {
printf("Failed at %li", i);
exit(1);
}
}
printf("All OK.\n");
}
So, it turns out that 12 iterations of the formula is enough to give correct results for all 64 bit unsigned longs that are perfect squares, and of course, non-squares will return false.
simon#simon-Inspiron-N5040:~$ time ./isqrt.bin
All OK.
real 11m37.096s
user 11m35.053s
sys 0m0.272s
So 697s/2^32 is approx 162ns. As it is, the function will have the same runtime for all inputs. Some of the measures detailed elsewhere in the discussion could speed it up for non-squares by checking the last four bits etc. Hope someone finds this interesting as I did.

If speed is a concern, why not partition off the most commonly used set of inputs and their values to a lookup table and then do whatever optimized magic algorithm you have come up with for the exceptional cases?

"I'm looking for the fastest way to determine if a long value is a perfect square (i.e. its square root is another integer)."
The answers are impressive, but I failed to see a simple check :
check whether the first number on the right of the long it a member of the set (0,1,4,5,6,9) . If it is not, then it cannot possibly be a 'perfect square' .
eg.
4567 - cannot be a perfect square.

It ought to be possible to pack the 'cannot be a perfect square if the last X digits are N' much more efficiently than that! I'll use java 32 bit ints, and produce enough data to check the last 16 bits of the number - that's 2048 hexadecimal int values.
...
Ok. Either I have run into some number theory that is a little beyond me, or there is a bug in my code. In any case, here is the code:
public static void main(String[] args) {
final int BITS = 16;
BitSet foo = new BitSet();
for(int i = 0; i< (1<<BITS); i++) {
int sq = (i*i);
sq = sq & ((1<<BITS)-1);
foo.set(sq);
}
System.out.println("int[] mayBeASquare = {");
for(int i = 0; i< 1<<(BITS-5); i++) {
int kk = 0;
for(int j = 0; j<32; j++) {
if(foo.get((i << 5) | j)) {
kk |= 1<<j;
}
}
System.out.print("0x" + Integer.toHexString(kk) + ", ");
if(i%8 == 7) System.out.println();
}
System.out.println("};");
}
and here are the results:
(ed: elided for poor performance in prettify.js; view revision history to see.)