微软和Java / Linux之间的正弦函数结果不同

时间:2013-10-28 20:56:11

标签: java c++ math.h sine

我有一个为Windows(Visual Studio)编写的C ++代码,我需要将其移植到Java中。这不是很容易,目前我被困在使用正弦函数。从Linux(经测试进行比较)和Java给出的结果与Windows源代码的结果不同。两种结果都是错误的,但这并不重要。重要的是结果完全相同。

我会将整个来源发布在底部。例如,我需要计算5174852443848405000.0的正弦值。我知道这是一个非常大的,也许是不寻常的数字,但我无法改变这一点。 Linux和Java返回0.153662,Windows返回0.16xx。 函数“random_value_genrator()”使用了大约500,000次,因此结果的差异可能会在某个时候发生。

initial_value_generator将计算random_value_generator函数稍后使用的值。该值由FILETIME对象和三个常量生成。缓冲区溢出正在发生但未处理。 每次使用时,random_value_generator都在修改DWORD64 prng_initial_value。

我能够成功构建initial_value_generator函数。

我想我无法完成这项任务,但感谢任何帮助。

一些全局变量:

DWORD64 prng_initial_value = 0;

DWORD64 CON1_RVG = 0x4F3D859E;
double CON2_RVG = 0.946270391;

DWORD64 CON1_PRNG = 0x2682D10B7;
DWORD64 CON2_PRNG = 0x19254D38000;
DWORD64 CON3_PRNG = 0x0F1E34A09;

此功能在程序启动时使用一次。将大DWORD64写入prng_initial_value,稍后由random_value_generator()使用。 系统时间乘以常数1(缓冲区溢出),除以常数2并加上常数3。

void initial_value_generator ()
{
    SYSTEMTIME systime;
    FILETIME filetime;

    // Systemzeit zu GMT-Format umwandeln
    SystemTimeToFileTime(&systime,&filetime);

    prng_initial_value = (*(DWORD64*)&filetime) * CON1_PRNG / CON2_PRNG + CON3_PRNG;
}

此函数在每次使用时更改DWORD64 prng_initial_value。

int random_value_generator () 
{
    double sin_value;
    double copy_of_prng_initial_value;
    DWORD64 prng_con1;
    double result;

    // Initialen Wert aus dem initial_random_generator in lokaler Variable speichern
    copy_of_prng_initial_value = prng_initial_value;

    // Sinus vom initialen Wert 
    sin_value = sin(copy_of_prng_initial_value);

    // Initialen Wert mulipikation mit einem konstanten Wert (0x4F3D859E)
    prng_con1 = prng_initial_value * CON1_RVG;

进一步的计算变得疯狂:

    result = prng_con1 + sin_value;
    result = result * copy_of_prng_initial_value;
    result = result + CON2_RVG;
    result = result * copy_of_prng_initial_value;

    // Das Ergebnis aus der Logarithmus Rechnung addieren
    result += log(copy_of_prng_initial_value);

    // Das Ergebnis aus den Berechnungen als Pointer in die
    // Speicheradresse von prng_initial_value als double Pointer speichern.
    *(double*)&prng_initial_value = result;

    // Rueckgabe des Berechneten Wert als Integer
    return prng_initial_value;
}

作为参考,我发布了我的Java代码(所有评论都是英文的)。随机函数看起来有点疯狂,因为我测试了很多东西。我很抱歉。但重要的是只使用Math.sin(double x)函数,结果与使用Microsoft C ++编译器的Math.h中的sin函数不同。

private final long initialValue;
private long randomValue;
final BigInteger uint64MaxValue = new BigInteger("18446744073709551616");   //2^64

public ConfickerC() {
    this.initialValue = this.generateInitialValue();
    this.randomValue = this.initialValue;
}

private long generateInitialValue() {
    //We need the actual date without the time from GMT +0 timezone
    Calendar cal = Calendar.getInstance(TimeZone.getTimeZone("GMT"));
    cal.set(Calendar.HOUR_OF_DAY, 0);
    cal.set(Calendar.MINUTE, 0);
    cal.set(Calendar.SECOND, 0);
    cal.set(Calendar.MILLISECOND, 0);

    long systemtimeAsFiletime = cal.getTimeInMillis();

    /*
     * Goal is to get the above created date into Windows FileTime format.
     * The Windows FileTime format has got 100 nano seconds per tick.
     * So one increment of the long value results further 100 nano seconds.
     * Instead of Unix the FileTime format begins with 1st January 1601 - not 1970.
     * 11644473600 is the interval between 1601 and 1970 in seconds.
     * Java has got a resolution of 1 ms per tick unix have got 1 second per
     * tick. So first devide the time by 1000. Then add the interval. 
     * After this we multiply by 10 million to get a resolution of 100
     * nano seconds per tick.
     */
    systemtimeAsFiletime /= 1000;        //divide by 1000 to get seconds instead of milliseconds
    systemtimeAsFiletime += 11644473600L; //add x seconds to add the interval between 1601 and 1970
    systemtimeAsFiletime *= 10000000L;   //Windows FileTime has a resolution of 100 nano seconds per tick; so multiply by 10M

    /*
     * The virus is calulating for getting the initial value: time * con1 / con2 + con3
     * Originaly there occurs a buffer overflow which is not handled in the C++ code.
     * The funny thing is that Java does not have a DWORD64 (unsinged _int64). So because of this bit missing (and so the overflow is different) we need BigInteger. 
     * Because BigInteger has no 2^64 limitation we need the maximul value of DWORD64.
     * This is used to "simulate" the buffer overflow by calculating ((time * con1) % 2^64) / con2 + con3
     * modulo 2^64 will result a value which is equal to the C++ calculation
     */

    final BigInteger CONSTANT_1 = new BigInteger("10337718455");                //Original: 0x2682D10B7
    final BigInteger CONSTANT_2 = new BigInteger("1728000000000");              //Original: 0x19254D38000
    final BigInteger CONSTANT_3 = new BigInteger("4058204681");                 //Original: 0x0F1E34A09

    BigInteger bigIntSystemTime = BigInteger.valueOf(systemtimeAsFiletime);

    //Return as long value: ((time * con1) % 2^64) / con2 + con3
    return bigIntSystemTime.multiply(CONSTANT_1).divideAndRemainder(uint64MaxValue)[1].divide(CONSTANT_2).add(CONSTANT_3).longValue();
}

private int generateRandomValue() {
    final long CONSTANT_1 = 1329431966L;
    final double CONSTANT_2 = 0.946270391;
    double result = 0.0;
    double copyOfInit = this.randomValue;

    System.out.println(System.getProperty("line.separator") + "Copy of init: " + copyOfInit);
    System.out.printf("Copy of init: %f\n", copyOfInit);
    double sinInit = Math.sin(copyOfInit); System.out.println("Sinus: " + sinInit);  
    System.out.printf("Sinus: %f\n", sinInit);
    System.out.println("Sinus gerundet: " + Math.round(sinInit*1000000)/1000000.0d);
    BigInteger b =        BigInteger.valueOf(this.randomValue).multiply(BigInteger.valueOf(CONSTANT_1)).divideAndRemainder(uint64MaxValue)[1];

    System.out.println("Init * Konstante 1: " + b);
    BigDecimal bd = new BigDecimal(b.toString());
    //bd.add(BigDecimal.valueOf(sinInit));

    //result = t + sinInit; System.out.println("Multi + Sinus: " + result);
    result = bd.add(BigDecimal.valueOf(sinInit)).doubleValue(); System.out.println("Multi + Sinus: " + result);
    result *= (long) this.randomValue; System.out.println("Zwischenergebnis * init: " + result);
    result += CONSTANT_2; System.out.println("Konstante 2 addiert: " + result);
    System.out.printf("BigD: %s", BigDecimal.valueOf(result).multiply(BigDecimal.valueOf(randomValue)));
    result *= this.randomValue; System.out.printf("Erneut mit init multipliziert: %f", result);
    double l = Math.log((long)this.randomValue); System.out.println("Log von init: " + l);
    result += l; System.out.printf("+= mit Log: %f\n", result);

    this.randomValue = (long)result; System.out.printf("Ende: %d\n", this.randomValue);

    this.randomValue = Double.doubleToRawLongBits(result);

    return (int)this.randomValue;   
}

1 个答案:

答案 0 :(得分:2)

三角函数是具有合理模糊规范的库函数。例如,以下是C标准就此主题所说的内容(7.12.4.6):

  

罪的功能

     

概要

#include <math.h>
double sin(double x);
float sinf(float x);
long double sinl(long double x);
  

描述

     

sin函数计算x的正弦值(以弧度为单位)。

     

返回

     

sin函数返回sin x

因此,他们将使用不同的算法和不同的准确度,即使用库版本,您将得不到完全相同的结果。例如,不同的库可以在准确度和计算速度之间进行不同的权衡。即使库实现完全相同,您也可能无法在不同系统上获得完全相同的结果,因为在整个计算过程中,值可能会在不同点处舍入。要在不同平台之间获得相当接近的结果,您可能需要在这些平台上实现相同的算法。

请注意,sin(x)明确提供[0, π/2]范围内的最佳结果。将一个巨大的数字传递给sin(x)可能会产生一个相当糟糕的近似值,尽管我希望大多数实现都是在进行任何计算之前将x映射到上面给出的范围内。理想情况下,您从一开始就避免使用大值,并以π的倍数表示它们。但是,即使x属于上述范围,您也可能会从不同的实现中获得不同的结果。