Develop Reference

Check if square root is a perfect integer in Java [duplicate] - java

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.)

Related

I am currently working on a Java math library which will include a variety of correctly rounded functions (i.e. sqrt, cbrt, exp, sin, gamma, and ln). I have already used the Babylonian method to write a square root algorithm that is correct to within 1 ulp of the correct answer. However, I cannot figure out how to properly calculate which way the number should be rounded to represent the best possible approximation to the actual square root of the input. Answers containing principles which can be extended to other functions would be preferred, but I have heard that sqrt is a simpler case than many transcendental functions, and specialized solutions would also be much appreciated.
Also, here is a cleaned-up version of my code as of this question's original submission:
public static double sqrt(double x) {
long bits = Double.doubleToLongBits(x);
// NaN and non-zero negatives:
if (Double.isNaN(x) || x < 0) return Double.NaN;
// +-0 and 1:
if (x == 0d || x == 1d) return x;
// Halving the exponent to come up with a good initial guess:
long exp = bits << 1;
exp = (exp - 0x7fe0000000000000L >> 1) + 0x7fe0000000000000L >>> 1 & 0x7ff0000000000000L;
double guess = Double.longBitsToDouble(bits & 0x800fffffffffffffL | exp);
double nextUp, nextDown, guessSq, nextUpSq, nextDownSq;
// Main loop:
while (true) {
guessSq = guess * guess;
if (guessSq == x) return guess;
nextUp = Math.nextUp(guess);
nextUpSq = nextUp * nextUp;
if (nextUpSq == x) return nextUp;
if (guessSq < x && x < nextUpSq) {
double z = x / nextUp;
if (z * nextUp > x) z = Math.nextDown(z);
return z < nextUp ? nextUp : guess;
}
nextDown = Math.nextDown(guess);
nextDownSq = nextDown * nextDown;
if (nextDownSq == x) return nextDown;
if (nextDownSq < x && x < guessSq) {
double z = x / guess;
if (z * guess > x) z = Math.nextDown(z);
return z < guess ? guess : nextDown;
}
// Babylonian method:
guess = 0.5 * (guess + x / guess);
}
}
As you can see, I was using division as a test. However, I believe that requires the division to round towards 0, which obviously doesn't happen in Java.

By the Taylor theorem, the square root function is locally approximated by a linear function, of slope 1/2√x, which is positive. So you can relate the error to the error in the square, x - (√x)², where √x is understood to be the approximate root. Then you round in the direction that minimizes this error.
Anyway, the computation of x - (√x)² is subjected to catastrophic cancellation and you may need extended accuracy to compute it reliably. Not sure the benefit is worth the effort.

How to find the next lower binary number for an integer (same number of 1s)? For example: if given input number n = 10 (1010), the function should return 9 (1001), or n = 14 (1110) then return 13 (1101), or n = 22 (10110) then return 21 (10101), n = 25 (11001) then return 22 (10110)... etc.

You can do this.
static int nextLower(int n) {
int bc = Integer.bitCount(n);
for (int i = n - 1; i > 0; i--)
if (Integer.bitCount(i) == bc)
return i;
throw new RuntimeException(n+" is the lowest with a bit count of "+bc);
}
Of course if this is homework you are going to have trouble convincing someone you wrote this ;)

For the sake of clarity, in this answer I will use the term 'cardinality' to indicate the number of 1s in the binary representation of a number.
One (obvious) way is to run a downwards loop, and seek for the first number with the same cardinality as your input (just like Peter Lawrey suggested).
I don't think this is inefficient, because I guess the output number is always pretty close to the input. More precisely, all you have to do is to find the rightmost '10' bit sequence, and change it to '01'. Then replace the right part with a number having all 1s at its left, as many as you can, without breaking the postcondition. This brings us to another solution, which consists in converting the number to a binary string (like user2573153 showed you), performing the replacement (with a regular expression, maybe), and then converting back to int.
A slightly faster version of Peter's algorithm should be the following, which performs on integers the manipulation I proposed you for strings:
static int nextLower(int n) {
int fixPart = 0;
int shiftCount = 0;
while ((n & 3) != 2) {
if (n == 0) {
throw new IllegalArgumentException(
fixPart + " is the lowest number with its cardinality");
}
fixPart |= (n & 1) << shiftCount;
shiftCount += 1;
n /= 2;
}
int fixZeros = shiftCount - Integer.bitCount(fixPart);
return ((n ^ 3) << shiftCount) | (((1 << shiftCount) - 1) & ~((1 << fixZeros) - 1));
}
which is O(log n) rather than O(n), but it's definitely harder to understand, and may also be practically slower, due to its complexity. Anyway, you could only notice a difference if you try with some huge difficult number.
EDIT I tried a little benchmark, and found that this code is 67% faster than Peter Lawrey's when applied consecutively to all numbers from 2 to 100,000,000. I don't think this is enough to justify the increased code complexity.

I like such binary task, so to find next lower number you should find right most 1 followed by 0 and exchange them,. UPDATE: you need to "reorder" the rest part of number with 1s at left and 0s at right
10 1010 ->
9 1001
14 1110 ->
13 1101
25 11001 ->
22 10110
here is sample code:
int originalValue = 25;
int maskToCheck = 2; // in binary 10b
int clearingMask = 1;
int settingMask = 0;
int zeroCount = 0;
while (maskToCheck > 0)
{
if ( (originalValue&(maskToCheck|(maskToCheck>>1))) == maskToCheck ) // we found such
{
int newValue = originalValue&(~maskToCheck); // set 1 with 0
newValue = newValue&(~clearingMask)|(settingMask<<zeroCount); // clear all the rest bits, and set most valuable ones
newValue = newValue|(maskToCheck>>1); // set 0 with 1
System.out.println("for " + originalValue + " we found " + newValue);
break;
}
else
{
if ( (originalValue&(maskToCheck>>1)) > 0) // we have 1 bit in cleared part
settingMask = (settingMask<<1) | 1;
else
zeroCount++;
maskToCheck = maskToCheck<<1; // try next left bits
clearingMask = (clearingMask<<1)|1;
}
}

Java's Random function takes a seed and produces the a sequence of 'psuedo-random' numbers.
(It is implemented based on some algorithm discussed in Donald Knuth, The Art of Computer Programming, Volume 3, Section 3.2.1.), but the article is too technical for me to understand)
Is there an inverse function of it?
That is, given a sequence of numbers, would it be possible to mathematically determine what the seed would be?
(, which means, brute-forcing doesn't count as a valid method)
[Edit]
There seems to be quite a number of comments here... I thought I'd clarify what I am looking for.
So for instance, the function y = f(x) = 3x has an inverse function, which is y = g(x) = x/3.
But the function z = f(x, y) = x * y does not have an inverse function, because (I could give a full mathematical proof here, but I don't want to sidetrack my main question), intuitively speaking, there are more than one pair of (x, y) such that (x * y) == z.
Now back to my question, if you say the function is not inversible, please explain why.
(And I am hoping to get answers from those who have really read to article and understand it. Answers like "It's just not possible" aren't really helping)

If we're talking about the Oracle (née Sun) implementation of java.util.Random, then yes, it is possible once you know enough bits.
Random uses a 48-bit seed and a linear congruential generator. These are not cryptographically safe generators, because of the tiny state size (bruteforceable!) and the fact that the output just isn't that random (many generators will exhibit small cycle length in certain bits, meaning that those bits can be easily predicted even if the other bits seem random).
Random's seed update is as follows:
nextseed = (seed * 0x5DEECE66DL + 0xBL) & ((1L << 48) - 1)
This is a very simple function, and it can be inverted if you know all the bits of the seed by calculating
seed = ((nextseed - 0xBL) * 0xdfe05bcb1365L) & ((1L << 48) - 1)
since 0x5DEECE66DL * 0xdfe05bcb1365L = 1 mod 248. With this, a single seed value at any point in time suffices to recover all past and future seeds.
Random has no functions that reveal the whole seed, though, so we'll have to be a bit clever.
Now, obviously, with a 48-bit seed, you have to observe at least 48 bits of output or you clearly don't have an injective (and thus invertible) function to work with. We're in luck: nextLong returns ((long)(next(32)) << 32) + next(32);, so it produces 64 bits of output (more than we need). Indeed, we could probably make do with nextDouble (which produces 53 bits), or just repeated calls of any other function. Note that these functions cannot output more than 248 unique values because of the seed's limited size (hence, for example, there are 264-248 longs that nextLong will never produce).
Let's specifically look at nextLong. It returns a number (a << 32) + b where a and b are both 32-bit quantities. Let s be the seed before nextLong is called. Then, let t = s * 0x5DEECE66DL + 0xBL, so that a is the high 32 bits of t, and let u = t * 0x5DEECE66DL + 0xBL so that b is the high 32 bits of u. Let c and d be the low 16 bits of t and u respectively.
Note that since c and d are 16-bit quantities, we can just bruteforce them (since we only need one) and be done with it. That's pretty cheap, since 216 is only 65536 -- tiny for a computer. But let's be a bit more clever and see if there's a faster way.
We have (b << 16) + d = ((a << 16) + c) * 0x5DEECE66DL + 11. Thus, doing some algebra, we obtain (b << 16) - 11 - (a << 16)*0x5DEECE66DL = c*0x5DEECE66DL - d, mod 248. Since c and d are both 16-bit quantities, c*0x5DEECE66DL has at most 51 bits. This usefully means that
(b << 16) - 11 - (a << 16)*0x5DEECE66DL + (k<<48)
is equal to c*0x5DEECE66DL - d for some k at most 6. (There are more sophisticated ways to compute c and d, but because the bound on k is so tiny, it's easier to just bruteforce).
We can just test all the possible values for k until we get a value whos negated remainder mod 0x5DEECE66DL is 16 bits (mod 248 again), so that we recover the lower 16 bits of both t and u. At that point, we have a full seed, so we can either find future seeds using the first equation, or past seeds using the second equation.
Code demonstrating the approach:
import java.util.Random;
public class randhack {
public static long calcSeed(long nextLong) {
final long x = 0x5DEECE66DL;
final long xinv = 0xdfe05bcb1365L;
final long y = 0xBL;
final long mask = ((1L << 48)-1);
long a = nextLong >>> 32;
long b = nextLong & ((1L<<32)-1);
if((b & 0x80000000) != 0)
a++; // b had a sign bit, so we need to restore a
long q = ((b << 16) - y - (a << 16)*x) & mask;
for(long k=0; k<=5; k++) {
long rem = (x - (q + (k<<48))) % x;
long d = (rem + x)%x; // force positive
if(d < 65536) {
long c = ((q + d) * xinv) & mask;
if(c < 65536) {
return ((((a << 16) + c) - y) * xinv) & mask;
}
}
}
throw new RuntimeException("Failed!!");
}
public static void main(String[] args) {
Random r = new Random();
long next = r.nextLong();
System.out.println("Next long value: " + next);
long seed = calcSeed(next);
System.out.println("Seed " + seed);
// setSeed mangles the input, so demangle it here to get the right output
Random r2 = new Random((seed ^ 0x5DEECE66DL) & ((1L << 48)-1));
System.out.println("Next long value from seed: " + r2.nextLong());
}
}

I normally wouldn't just link articles... But I found a site where someone looks into this in some depth and thought it was worth posting. http://jazzy.id.au/default/2010/09/20/cracking_random_number_generators_part_1.html
It seems that you can calculate a seed this way:
seed = (seed * multiplier + addend) mod (2 ^ precision)
where multiplier is 25214903917, addend is 11, and precision is 48 (bits). You can't calculate what the seed was with only 1 number, but you can with 2.
EDIT: As nhahtdh said there's a part 2 where he delves into more of the math behind the seeds.

I would like to present an implementation to reverse a sequence of integers generated by nextInt().
The program will brute force on the lower 16-bit discarded by nextInt(), use the algorithm provided in the blog by James Roper to find previous seed, then check that upper 32 bit of the 48-bit seed are the same as the previous number. We need at least 2 integers to derive the previous seed. Otherwise, there will be 216 possibilities for the previous seed, and all of them are equally valid until we have at least one more number.
It can be extended for nextLong() easily, and 1 long number is enough to find the seed, since we have 2 pieces of upper 32-bit of the seed in one long, due to the way it is generated.
Note that there are cases where the result is not the same as what you set as secret seed in the SEED variable. If the number you set as secret seed occupies more than 48-bit (which is the number of bits used for generating random numbers internally), then the upper 16 bits of 64 bit of long will be removed in the setSeed() method. In such cases, the result returned will not be the same as what you have set initially, it is likely that the lower 48-bit will be the same.
I would like to give most the credit to James Roper, the author of this blog article which makes the sample code below possible:
import java.util.Random;
import java.util.Arrays;
class TestRandomReverse {
// The secret seed that we want to find
private static long SEED = 782634283105L;
// Number of random numbers to be generated
private static int NUM_GEN = 5;
private static int[] genNum(long seed) {
Random rand = new Random(seed);
int arr[] = new int[NUM_GEN];
for (int i = 0; i < arr.length; i++) {
arr[i] = rand.nextInt();
}
return arr;
}
public static void main(String args[]) {
int arr[] = genNum(SEED);
System.out.println(Arrays.toString(arr));
Long result = reverse(arr);
if (result != null) {
System.out.println(Arrays.toString(genNum(result)));
} else {
System.out.println("Seed not found");
}
}
private static long combine(int rand, int suffix) {
return (unsignedIntToLong(rand) << 16) | (suffix & ((1L << 16) - 1));
}
private static long unsignedIntToLong(int num) {
return num & ((1L << 32) - 1);
}
// This function finds the seed of a sequence of integer,
// generated by nextInt()
// Can be easily modified to find the seed of a sequence
// of long, generated by nextLong()
private static Long reverse(int arr[]) {
// Need at least 2 numbers.
assert (arr.length > 1);
int end = arr.length - 1;
// Brute force lower 16 bits, then compare
// upper 32 bit of the previous seed generated
// to the previous number.
for (int i = 0; i < (1 << 16); i++) {
long candidateSeed = combine(arr[end], i);
long previousSeed = getPreviousSeed(candidateSeed);
if ((previousSeed >>> 16) == unsignedIntToLong(arr[end - 1])) {
System.out.println("Testing seed: " +
previousSeed + " --> " + candidateSeed);
for (int j = end - 1; j >= 0; j--) {
candidateSeed = previousSeed;
previousSeed = getPreviousSeed(candidateSeed);
if (j > 0 &&
(previousSeed >>> 16) == unsignedIntToLong(arr[j - 1])) {
System.out.println("Verifying: " +
previousSeed + " --> " + candidateSeed);
} else if (j == 0) {
// The XOR is done when the seed is set, need to reverse it
System.out.println("Seed found: " + (previousSeed ^ MULTIPLIER));
return previousSeed ^ MULTIPLIER;
} else {
System.out.println("Failed");
break;
}
}
}
}
return null;
}
private static long ADDEND = 0xBL;
private static long MULTIPLIER = 0x5DEECE66DL;
// Credit to James Roper
// http://jazzy.id.au/default/2010/09/21/cracking_random_number_generators_part_2.html
private static long getPreviousSeed(long currentSeed) {
long seed = currentSeed;
// reverse the addend from the seed
seed -= ADDEND; // reverse the addend
long result = 0;
// iterate through the seeds bits
for (int i = 0; i < 48; i++)
{
long mask = 1L << i;
// find the next bit
long bit = seed & mask;
// add it to the result
result |= bit;
if (bit == mask)
{
// if the bit was 1, subtract its effects from the seed
seed -= MULTIPLIER << i;
}
}
return result & ((1L << 48) - 1);
}
}

I need to work out a very large power modulo (2^32), i.e. I want the result of:
y = (p^n) mod (2^32)
p is a prime number
n is a large integer
Is there a trick to doing this efficiently in Java?
Or am I stuck with doing it in a loop with n iterations?

The simple way to mod 2^32 is to use & 0xFFFFFFFFL. Also, there happens to be a type which naturally keeps the lowest 32-bit called int ;) If you use that you don't even need to perform the & until you have the result (so the answer is unsigned) For this reason you only need to keep the last 32 bit of the answer. To speed up the ^n you can calculate the square, it's square and it's square etc, e.g if n is 0b11111 then you need to multiply p^16 * p^8 * p^4 * p^2 * p.
In short, you can use plain int as you only need 32-bit of accuracy and values with a cost of O(ln n) where n is the power.
int prime = 2106945901;
for (int i = 0; i < 10; i++) {
long start = System.nanoTime();
long answer1 = BigInteger.valueOf(prime)
.modPow(
BigInteger.valueOf(prime),
BigInteger.valueOf(2).pow(32)).longValue();
long mid = System.nanoTime();
int answer2 = 1;
int p = prime;
for (int n = prime; n > 0; n >>>= 1) {
if ((n & 1) != 0)
answer2 *= p;
p *= p;
}
long end = System.nanoTime();
System.out.printf("True answer %,d took %.3f ms, quick answer %,d took %.3f ms%n",
answer1, (mid - start) / 1e6, answer2 & 0xFFFFFFFFL, (end - mid) / 1e6);
}
prints finally
True answer 4,169,684,317 took 0.233 ms, quick answer 4,169,684,317 took 0.002 ms

You can utilize exponentiation by squaring. Firstly, break it down into powers of two for your given n. Since p^n (mod x) == p^(k1) (mod x) . p^(k2) (mod x) . ... p^(kn) (mod x) where sum k_i = n, you can utilize this and successive powers of two to calculate this in O(log n) steps.

In addition to the other answers you can use some elementary number theory to reduce the time needed to compute an mod 232 for a an odd integer to O(1). The Euler Phi function together with Euler's Theorem allows you to discard all but the low-order 31 bits of n.
φ(232) = 231, and aφ(232) = 1 mod 232.
Thus if n = q*(231) + r, 0 <= r < 231, then an mod 232 = ar mod 232
r is simply the low-order 31 bits of n, i.e. n & 0x7fffffff. In fact, by Carmichael's Theorem you can do a bit better (literally), and you only need to consider the low-order 30 bits of n, i.e. n & 0x3fffffff. You can precompute these once and store them in a table of size 4GB for a given base a. Here is some java code as an example.
import java.math.BigInteger;
public class PowMod2_32 {
private static final long MASK32 = 0xffffffffL;
public static long pow32(final int a, final int exponent)
{
int prod = 1;
for (int i = 29; i>=0; i--) {
prod *= prod; // square
if (((exponent >> i) & 1) == 1) {
prod *= a; // multiply
}
}
return prod & MASK32;
}
public static long pow32(BigInteger a, BigInteger exponent) {
return pow32(a.intValue(), exponent.intValue());
}
}

There are no tricks in java that I know of but rather there are some tricks in maths.
If you implement these as an algorithm it should speed up computation.
Look at 5 and 6. Look at 4 also if power of two is always even

Use the Class Bigintiger. here´s an example how to work / pow with it
public String higherPow() {
BigIntiger i = new Bigintger("2");
// doing a power(2^32)
i = i.pow(32);
// after 2^32 was made, do mod 100
i = i.mod(new Bigintiger("100"));
return i.toString();
}

This question already has answers here:
Closed 10 years ago.
Possible Duplicate:
Best algorithm to count the number of set bits in a 32-bit integer?
How do I count the number of 1's a number will have in binary?
So let's say I have the number 45, which is equal to 101101 in binary and has 4 1's in it. What's the most efficient way to write an algorithm to do this?

Instead of writing an algorithm to do this its best to use the built in function. Integer.bitCount()
What makes this especially efficient is that the JVM can treat this as an intrinsic. i.e. recognise and replace the whole thing with a single machine code instruction on a platform which supports it e.g. Intel/AMD
To demonstrate how effective this optimisation is
public static void main(String... args) {
perfTestIntrinsic();
perfTestACopy();
}
private static void perfTestIntrinsic() {
long start = System.nanoTime();
long countBits = 0;
for (int i = 0; i < Integer.MAX_VALUE; i++)
countBits += Integer.bitCount(i);
long time = System.nanoTime() - start;
System.out.printf("Intrinsic: Each bit count took %.1f ns, countBits=%d%n", (double) time / Integer.MAX_VALUE, countBits);
}
private static void perfTestACopy() {
long start2 = System.nanoTime();
long countBits2 = 0;
for (int i = 0; i < Integer.MAX_VALUE; i++)
countBits2 += myBitCount(i);
long time2 = System.nanoTime() - start2;
System.out.printf("Copy of same code: Each bit count took %.1f ns, countBits=%d%n", (double) time2 / Integer.MAX_VALUE, countBits2);
}
// Copied from Integer.bitCount()
public static int myBitCount(int i) {
// HD, Figure 5-2
i = i - ((i >>> 1) & 0x55555555);
i = (i & 0x33333333) + ((i >>> 2) & 0x33333333);
i = (i + (i >>> 4)) & 0x0f0f0f0f;
i = i + (i >>> 8);
i = i + (i >>> 16);
return i & 0x3f;
}
prints
Intrinsic: Each bit count took 0.4 ns, countBits=33285996513
Copy of same code: Each bit count took 2.4 ns, countBits=33285996513
Each bit count using the intrinsic version and loop takes just 0.4 nano-second on average. Using a copy of the same code takes 6x longer (gets the same result)

The most efficient way to count the number of 1's in a 32-bit variable v I know of is:
v = v - ((v >> 1) & 0x55555555);
v = (v & 0x33333333) + ((v >> 2) & 0x33333333);
c = ((v + (v >> 4) & 0xF0F0F0F) * 0x1010101) >> 24; // c is the result
Updated: I want to make clear that it's not my code, actually it's older than me. According to Donald Knuth (The Art of Computer Programming Vol IV, p 11), the code first appeared in the first textbook on programming, The Preparation of Programs for an Electronic Digital Computer by Wilkes, Wheeler and Gill (2nd Ed 1957, reprinted 1984). Pages 191–193 of the 2nd edition of the book presented Nifty Parallel Count by D B Gillies and J C P Miller.

See Bit Twidling Hacks and study all the 'counting bits set' algorithms. In particular, Brian Kernighan's way is simple and quite fast if you expect a small answer. If you expect an evenly distributed answer, lookup table might be better.

This is called Hamming weight. It is also called the population count, popcount or sideways sum.

The following is either from "Bit Twiddling Hacks" page or Knuth's books (I don't remember). It is adapted to unsigned 64 bit integers and works on C#. I don't know if the lack of unsigned values in Java creates a problem.
By the way, I write the code only for reference; the best answer is using Integer.bitCount() as #Lawrey said; since there is a specific machine code operation for this operation in some (but not all) CPUs.
const UInt64 m1 = 0x5555555555555555;
const UInt64 m2 = 0x3333333333333333;
const UInt64 m4 = 0x0f0f0f0f0f0f0f0f;
const UInt64 h01 = 0x0101010101010101;
public int Count(UInt64 x)
{
x -= (x >> 1) & m1;
x = (x & m2) + ((x >> 2) & m2);
x = (x + (x >> 4)) & m4;
return (int) ((x * h01) >> 56);
}

public int f(int n)
{
int result = 0;
for(;n > 0; n = n >> 1)
result += ((n & 1) == 1 ? 1 : 0);
return result;
}

The following Ruby code works for positive numbers.
count = 0
while num > 1
count = (num % 2 == 1) ? count + 1 : count
num = num >> 1
end
count += 1
return count

The fastest I have used and also seen in a practical implementation (in the open source Sphinx Search Engine) is the MIT HAKMEM algorithm. It runs superfast over a very large stream of 1's and 0's.