You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
This repository has been archived by the owner on Nov 9, 2023. It is now read-only.
The inverse square root operations is fundamental to the Ristretto operations. This is due to division. As division and exponentiation are expensive for a CPU to perform, inverse modular operations are used when any such operations need to be programmed.
For example taking 1/(√x), there is both a division and an exponentiation. For this, a modular inverse square root computes 1 multiplied by inverse mod of square root x.
There are multiple ways to compute the aforementioned calculation. This issue is opened to explore the various ways we can implement inverse square root. In general, the non-Euclidian methods provide shorter operation times but it doesn't necessarily mean this will be the case for us. This is a result of which sign choices we assign as well as which canonical endings we need to omit.
The issue will document which properties we seek from computed as well as benchmarks of various implemented inverse square root algorithms.
The text was updated successfully, but these errors were encountered:
The inverse square root operations is fundamental to the Ristretto operations. This is due to division. As division and exponentiation are expensive for a CPU to perform, inverse modular operations are used when any such operations need to be programmed.
For example taking 1/(√x), there is both a division and an exponentiation. For this, a modular inverse square root computes 1 multiplied by inverse mod of square root x.
There are multiple ways to compute the aforementioned calculation. This issue is opened to explore the various ways we can implement inverse square root. In general, the non-Euclidian methods provide shorter operation times but it doesn't necessarily mean this will be the case for us. This is a result of which sign choices we assign as well as which canonical endings we need to omit.
The issue will document which properties we seek from computed as well as benchmarks of various implemented inverse square root algorithms.
The text was updated successfully, but these errors were encountered: