> The routines include precomputed full double precision points and > weights for n = 2, ..., 512, and 1024 Gauss points. If the user selects > any other number, the necessary points and weights are computed > on the fly and remain quite good.
A quick comment from Pavel about the numerical precision in the patch's coefficients... > ... fixed tables included in [the patch] are way beyond full double precision. > Full double precision is about 1e-16, fixed tables are of 1e-25. > This doesn't hurt the double calculations and allows future extensions > e.g. for 'long double' and multi-precision libraries (like gmp or mpfr). > > Tables generated on the fly are of 1e-10, which is pretty much the > maximum precision we can get using double to calculate weights. > So, on-the-fly tables are of full double precision, fixed have much > higher precision. - Rhys
