Currently the max is calculated by looping over the elements of the two registers and calculating the max of both elements. The number of elements of each registers is 2^p, so at most of 2^16 (or 2^18, Google provides bias correction data up to p = 18).
It might be useful to parallelize this calculation.
Currently the max is calculated by looping over the elements of the two registers and calculating the max of both elements. The number of elements of each registers is 2^p, so at most of 2^16 (or 2^18, Google provides bias correction data up to p = 18).
It might be useful to parallelize this calculation.