Yeah, I think only Power has hardware double of 128 bits, but it's not the same as double-double because the exponent range is equal to double's exponent range.

Yeah, I think only Power has hardware double of 128 bits, but it's not the same as double-double because the exponent range is equal to double's exponent range. I agree this package is not new conceptually, but it has a convenient metaprogramming to generate code for arithmetic for any [hardware ieee number type] x N. It also lists the complexity as a function of N, most operations scale as N^3, so yes, for larger N it might fail to be fast.