get from R to 2^(_W * n) R mod m (aka from one to R in // the Montgomery domain, meaning we can use Montgomery multiplication now). // We could do that by doubling _W * n times, or with a square-and-double // chain log2(_W * n) long. Turns out the fastest thing is to start out with // doublings, and switch to square-and-double once the exponent is large // enough to justify the cost of the multiplications. // The threshold is selected experimentally as a linear function of n. threshold := n / 4 //...

get from R to 2^(_W * n) R mod m (aka from one to R in // the Montgomery domain, meaning we can use Montgomery multiplication now). // We could do that by doubling _W * n times, or with a square-and-double // chain log2(_W * n) long. Turns out the fastest thing is to start out with // doublings, and switch to square-and-double once the exponent is large // enough to justify the cost of the multiplications. // The threshold is selected experimentally as a linear function of n. threshold := n / 4 //...

* Introducing `tf.types.experimental.AtomicFunction` as the fastest way to perform TF computations in Python.

    * Can be accessed through `inference_fn` property of `ConcreteFunction`s
    * Does not support gradients.