The key insight behind this implementation is that each viwi term is in {−1,0,1}, and can still be represented in 2 bits in an addition-friendly manner. (This is not strictly necessary for bitwise parallel processing—the partial sum lookup algorithm discussed later handles 3-bit outputs by padding the raw input data to 3 bits per genotype call—but it allows for unusually high efficiency). The exact sequence of operations that we chose to evaluate the dot-product terms in a bitwise parallel fashion is somewhat arbitrary.
