Most programs that process numerical data consist of loops. But loop overhead is unnecessarily slow. Replacing these loops by hand-optimized, Assembler-written vector functions and matrix routines speeds up program execution by up to a factor of 2-3 and also greatly enhances numerical accuracy.