Comments on: Approximation of sqrt(x) in Java http://martin.ankerl.com/2009/01/05/approximation-of-sqrtx-in-java/ Chunky bacon!! Wed, 08 Feb 2012 16:00:41 +0000 hourly 1 By: Russ Calvert http://martin.ankerl.com/2009/01/05/approximation-of-sqrtx-in-java/#comment-736 Russ Calvert Fri, 15 Oct 2010 03:32:24 +0000 http://martin.ankerl.com/?p=195#comment-736 I've been checking this out and have a few observations. The method doesn't deal sensibly with a negative, infinite, or NaN parameter. The errors become larger than stated for parameters below about 1.9E-308, and become extremely large below about 1.0E-310. For positive parameters above 1.9E-308 I agree with the error and speedup figures, but find that I get an additional 10% speed improvement (for zero repetitions) by replacing 'doubleToLongBits' with 'doubleToRawLongBits'. An 'invsqrt' method can be created in a similar way from the 'pow' approximation, replacing 'b' with -0.5 instead of 0.5. This has similar errors, speedup, and limitations to the 'sqrt' method. I’ve been checking this out and have a few observations.

The method doesn’t deal sensibly with a negative, infinite, or NaN parameter.

The errors become larger than stated for parameters below about 1.9E-308, and become extremely large below about 1.0E-310.

For positive parameters above 1.9E-308 I agree with the error and speedup figures, but find that I get an additional 10% speed improvement (for zero repetitions) by replacing ‘doubleToLongBits’ with ‘doubleToRawLongBits’.

An ‘invsqrt’ method can be created in a similar way from the ‘pow’ approximation, replacing ‘b’ with -0.5 instead of 0.5. This has similar errors, speedup, and limitations to the ‘sqrt’ method.

]]>