October 14th, 2010, 7:48 am
QuoteOriginally posted by: outrunGreat! This is very simple to the points! and allows for error analysis.And R33 is optimized for z in [0,1]?I have a deadline tonight, working my *** off, but will contribute soon.. I'll do the Chebyshev polynomial approx. for various number of terms.One quick thing... Why does Intel choose for optimizing the (rescalable) region for [0, 1/2] instead of the more common [0,1] or [-1,1]? Has that to do with the storage of the exponent in floating point notation? Is so, that we should to that too, otherwise we need to waste cycles on processing that part too (the exponent, K).R33 at z = 1 is accurate up to 4/5 decimal places; even for z = 2, R33 = 7.4 while exp(z) = 7.38.I am not sure why in [0,1/2]. The rationale behind the numerical analysis is not clear, as I mentioned.BTW you NTL library for boost Remez.Maybe R22 is OK as well