如题,今天我们来看下java.math.BigDecimal是如何解决浮点数的精度问题的,在那之前当然得先了解下浮点数精度问题是什么问题了。下面我们先从IEEE 754说起。
IEEE二进制浮点数算术标准(IEEE 754)是20世纪80年代以来最广泛使用的浮点数运算标准,为许多CPU与浮点运算器所采用。这个标准定义了表示浮点数的格式(包括负零-0)与反常值(denormal number)),一些特殊数值(无穷(Inf)与非数值(NaN)),以及这些数值的“浮点数运算符”;它也指明了四种数值舍入规则和五种异常状况(包括异常发生的时机与处理方式)。
下面我们就以双精度,也就是double类型,为例来看看浮点数的格式。
sign | exponent | fraction |
---|---|---|
1位 | 11位 | 52位 |
63 | 62-52 实际的指数大小+1023 |
51-0 |
下面看个栗子,直接输出double类型的二进制表示,
public static void main(String[] args) {
printBits(3.5);
}
private static void printBits(double d) {
System.out.println("##"+d);
long l = Double.doubleToLongBits(d);
String bits = Long.toBinaryString(l);
int len = bits.length();
System.out.println(bits+"#"+len);
if(len == 64) {
System.out.println("[63]"+bits.charAt(0));
System.out.println("[62-52]"+bits.substring(1,12));
System.out.println("[51-0]"+bits.substring(12, 64));
} else {
System.out.println("[63]0");
System.out.println("[62-52]"+ pad(bits.substring(0, len - 52)));
System.out.println("[51-0]"+bits.substring(len-52, len));
}
}
private static String pad(String exp) {
int len = exp.length();
if(len == 11) {
return exp;
} else {
StringBuilder sb = new StringBuilder();
for (int i = 11-len; i > 0; i--) {
sb.append("0");
}
sb.append(exp);
return sb.toString();
}
}
##3.5
100000000001100000000000000000000000000000000000000000000000000#63
[63]0
[62-52]10000000000
[51-0]1100000000000000000000000000000000000000000000000000
指数大小为10000000000B-1023=1
,尾数为1.11B
,所以实际数值大小为11.1B=3.5
,妥妥的。
有一点需要注意的是上述格式为归约形式,所以尾数的整数部分为1,而当非归约形式时,尾数的整数部分是为0的。
上面我们使用的浮点数3.5刚好可以准确的用二进制来表示,
public static void main(String[] args) {
printBits(0.1);
}
##0.1
11111110111001100110011001100110011001100110011001100110011010#62
[63]0
[62-52]01111111011
[51-0]1001100110011001100110011001100110011001100110011010
0.1无法表示成0.1 + 0.1 + 0.1 == 0.3
将会返回false,
public static void main(String[] args) {
System.out.println(0.1 + 0.1 == 0.2); // true
System.out.println(0.1 + 0.1 + 0.1 == 0.3); // false
}
那么BigDecimal又是如何解决这个问题的?
BigDecimal的解决方案就是,不使用二进制,而是使用十进制(BigInteger)+小数点位置(scale)来表示小数,
public static void main(String[] args) {
BigDecimal bd = new BigDecimal("100.001");
System.out.println(bd.scale());
System.out.println(bd.unscaledValue());
}
输出,
3
100001
也就是100.001 = 100001 * 0.1^3
。这种表示方式下,避免了小数的出现,当然也就不会有精度问题了。十进制,也就是整数部分使用了BigInteger来表示,小数点位置只需要一个整数scale来表示就OK了。
当使用BigDecimal来进行运算时,也就可以分解成两部分,BigInteger间的运算,以及小数点位置scale的更新,下面先看下运算过程中scale的更新。
加法运算时,根据下面的公式scale更新为两个BigDecimal中较大的那个scale即可。
X*
相应的代码如下,
/**
* Returns a {@code BigDecimal} whose value is {@code (this +
* augend)}, and whose scale is {@code max(this.scale(),
* augend.scale())}.
*
* @param augend value to be added to this {@code BigDecimal}.
* @return {@code this + augend}
*/
public BigDecimal add(BigDecimal augend) {
long xs = this.intCompact;
long ys = augend.intCompact;
BigInteger fst = (xs != INFLATED) ? null : this.intVal;
BigInteger snd = (ys != INFLATED) ? null : augend.intVal;
int rscale = this.scale;
long sdiff = (long)rscale - augend.scale;
if (sdiff != 0) {
if (sdiff < 0) {
int raise = checkScale(-sdiff);
rscale = augend.scale;
if (xs == INFLATED ||
(xs = longMultiplyPowerTen(xs, raise)) == INFLATED)
fst = bigMultiplyPowerTen(raise);
} else {
int raise = augend.checkScale(sdiff);
if (ys == INFLATED ||
(ys = longMultiplyPowerTen(ys, raise)) == INFLATED)
snd = augend.bigMultiplyPowerTen(raise);
}
}
if (xs != INFLATED && ys != INFLATED) {
long sum = xs + ys;
// See "Hacker‘s Delight" section 2-12 for explanation of
// the overflow test.
if ( (((sum ^ xs) & (sum ^ ys))) >= 0L) // not overflowed
return BigDecimal.valueOf(sum, rscale);
}
if (fst == null)
fst = BigInteger.valueOf(xs);
if (snd == null)
snd = BigInteger.valueOf(ys);
BigInteger sum = fst.add(snd);
return (fst.signum == snd.signum) ?
new BigDecimal(sum, INFLATED, rscale, 0) :
new BigDecimal(sum, rscale);
}
乘法运算根据下面的公式也可以确定scale更新为两个scale之和。
X*
相应的代码,
/**
* Returns a {@code BigDecimal} whose value is <tt>(this ×
* multiplicand)</tt>, and whose scale is {@code (this.scale() +
* multiplicand.scale())}.
*
* @param multiplicand value to be multiplied by this {@code BigDecimal}.
* @return {@code this * multiplicand}
*/
public BigDecimal multiply(BigDecimal multiplicand) {
long x = this.intCompact;
long y = multiplicand.intCompact;
int productScale = checkScale((long)scale + multiplicand.scale);
// Might be able to do a more clever check incorporating the
// inflated check into the overflow computation.
if (x != INFLATED && y != INFLATED) {
/*
* If the product is not an overflowed value, continue
* to use the compact representation. if either of x or y
* is INFLATED, the product should also be regarded as
* an overflow. Before using the overflow test suggested in
* "Hacker‘s Delight" section 2-12, we perform quick checks
* using the precision information to see whether the overflow
* would occur since division is expensive on most CPUs.
*/
long product = x * y;
long prec = this.precision() + multiplicand.precision();
if (prec < 19 || (prec < 21 && (y == 0 || product / y == x)))
return BigDecimal.valueOf(product, productScale);
return new BigDecimal(BigInteger.valueOf(x).multiply(y), INFLATED,
productScale, 0);
}
BigInteger rb;
if (x == INFLATED && y == INFLATED)
rb = this.intVal.multiply(multiplicand.intVal);
else if (x != INFLATED)
rb = multiplicand.intVal.multiply(x);
else
rb = this.intVal.multiply(y);
return new BigDecimal(rb, INFLATED, productScale, 0);
}
BigInteger可以表示任意精度的整数。当你使用long类型进行运算,可能会产生溢出时就要考虑使用BigInteger了。BigDecimal就使用了BigInteger作为backend。
那么BigInteger是如何做到可以表示任意精度的整数的?答案是使用数组来表示,看下面这个栗子就很直观了,
public static void main(String[] args) {
byte[] mag = {
2, 1 // 10 00000001 == 513
};
System.out.println(new BigInteger(mag));
}
通过byte[]来当作底层的二进制表示,例如栗子中的[2, 1],也就是[00000010B, 00000001B],就是表示二进制的10 00000001B这个数,也就是513了。
BigInteger内部会将这个byte[]转换成int[]保存,代码在stripLeadingZeroBytes方法,
/**
* Translates a byte array containing the two‘s-complement binary
* representation of a BigInteger into a BigInteger. The input array is
* assumed to be in <i>big-endian</i> byte-order: the most significant
* byte is in the zeroth element.
*
* @param val big-endian two‘s-complement binary representation of
* BigInteger.
* @throws NumberFormatException {@code val} is zero bytes long.
*/
public BigInteger(byte[] val) {
if (val.length == 0)
throw new NumberFormatException("Zero length BigInteger");
if (val[0] < 0) {
mag = makePositive(val);
signum = -1;
} else {
mag = stripLeadingZeroBytes(val);
signum = (mag.length == 0 ? 0 : 1);
}
}
/**
* Returns a copy of the input array stripped of any leading zero bytes.
*/
private static int[] stripLeadingZeroBytes(byte a[]) {
int byteLength = a.length;
int keep;
// Find first nonzero byte
for (keep = 0; keep < byteLength && a[keep]==0; keep++)
;
// Allocate new array and copy relevant part of input array
int intLength = ((byteLength - keep) + 3) >>> 2;
int[] result = new int[intLength];
int b = byteLength - 1;
for (int i = intLength-1; i >= 0; i--) {
result[i] = a[b--] & 0xff;
int bytesRemaining = b - keep + 1;
int bytesToTransfer = Math.min(3, bytesRemaining);
for (int j=8; j <= (bytesToTransfer << 3); j += 8)
result[i] |= ((a[b--] & 0xff) << j);
}
return result;
}
上面也可以看到这个byte[]应该是big-endian two‘s-complement binary representation
。
那么为什么构造函数不直接让我们扔一个int[]进去就得了呢,还要这么转换一下?答案是因为Java的整数都是有符号整数,举个栗子,int类型没办法表示(byte)255,(byte)255,(byte)255,(byte)255
,这样才能表示32个1。
最后来看看BigInteger间的加法与乘法运算。
代码如下,
private static int[] add(int[] x, int[] y) {
// If x is shorter, swap the two arrays
if (x.length < y.length) {
int[] tmp = x;
x = y;
y = tmp;
}
int xIndex = x.length;
int yIndex = y.length;
int result[] = new int[xIndex];
long sum = 0;
// Add common parts of both numbers
while(yIndex > 0) {
// 最低位对齐再开始加
sum = (x[--xIndex] & LONG_MASK) +
(y[--yIndex] & LONG_MASK) + (sum >>> 32); // sum>>>32 是高32位,也就是进位
result[xIndex] = (int)sum; // 低32位直接保存
}
// Copy remainder of longer number while carry propagation is required
boolean carry = (sum >>> 32 != 0);
while (xIndex > 0 && carry) // x比y长,且最后还有进位
carry = ((result[--xIndex] = x[xIndex] + 1) == 0); // 一位一位往前进位,直到没有产生进位
// Copy remainder of longer number
while (xIndex > 0)
result[--xIndex] = x[xIndex];
// Grow result if necessary
if (carry) {
int bigger[] = new int[result.length + 1];
System.arraycopy(result, 0, bigger, 1, result.length);
bigger[0] = 0x01;
return bigger;
}
return result;
}
加法运算比较简单,就是模拟十进制加法运算的过程,从两个加数的最低位开始加,如果有进位就进位。
代码如下,
private int[] multiplyToLen(int[] x, int xlen, int[] y, int ylen, int[] z) {
int xstart = xlen - 1;
int ystart = ylen - 1;
if (z == null || z.length < (xlen+ ylen))
z = new int[xlen+ylen];
long carry = 0;
for (int j=ystart, k=ystart+1+xstart; j>=0; j--, k--) {
long product = (y[j] & LONG_MASK) *
(x[xstart] & LONG_MASK) + carry;
z[k] = (int)product;
carry = product >>> 32;
}
z[xstart] = (int)carry;
for (int i = xstart-1; i >= 0; i--) {
carry = 0;
for (int j=ystart, k=ystart+1+i; j>=0; j--, k--) {
long product = (y[j] & LONG_MASK) *
(x[i] & LONG_MASK) +
(z[k] & LONG_MASK) + carry;
z[k] = (int)product;
carry = product >>> 32;
}
z[i] = (int)carry;
}
return z;
}
乘法运算要复杂一点,不过也一样是模拟十进制乘法运算,也就是一个乘数的每一位与另一个乘数的每一位相乘再相加(乘法运算可以拆成加法运算),所以才有那个双重的for循环。
最后的最后,想说的一点是,其实BigInteger可以看成是
Java之道系列:BigDecimal如何解决浮点数精度问题
原文地址:http://blog.csdn.net/kisimple/article/details/43773899