Double-double: 31 digits of precision without leaving the FPU
Summary
The article explains double-double arithmetic by glueing two doubles together to achieve ~31 digits of precision with about 9x the cost of a plain double. It covers error-free transformations (Knuth's two-sum, Dekker's quickTwoSum, and twoProduct with FMA), benchmarks against MPFR, and practical guidance for using fixed-width floating-point types in fractal rendering, including code samples and appendices.