如何“修复” java 中的浮点错误？

Question

I realize that the real answer is "don't use floating point numbers, use BigDecimal" but that will be a longer-term change in a large codebase.

My problem is relatively simple - I have numbers (doubles), both positive and negative, that I want to round towards zero to the nearest whole number. However, in the case of floating point imprecision errors, I need them to be rounded to the "actual" number.

Examples:

• 351.365 rounds to 351
• 3.9991532 rounds to 3
• 3.9999999999999996 rounds to 4
• -4.5 rounds to -4
• -123.9999999999993 rounds to -124

I guess for the purposes of this question, we can just ignore the fact that the rounding errors are caused by floating point imprecision, and rephrase the requirements as:

"Round towards zero to the nearest whole number, unless the value is less than 1e-6 away from a whole number N, in which case round to N."

My code is gross (particularly the "% 1" feels wrong, and I know that there could be a loss of precision casting double to int, but for simplicity just assume that the double values are all less than Integer.MAX_VALUE):

private int getRoundedValue(double d) {
    double fractionalPart = d % 1;
    if (Math.abs(fractionalPart) < 1e-6 || (1 - Math.abs(fractionalPart) < 1e-6)) {
        
        // here we are close enough to a whole number so just round to it (up or down)
        return (int) Math.round(d);
    }

    // otherwise just return the nearest whole number towards zero
    return (int) d;  
} 

How can I improve the above method?

Answer 1

Here's a simpler version:

private static int getRoundedValue(double d) {
    return Math.toIntExact((long)(Math.round(d * 1e6) / 1e6));
}

Math.toIntExact throws ArithmeticException on overflows.