Get the latest tech news

Faster Argmin on Floats

If you know that you are working with "regular" floats, you can make argmin faster.

Consider the following problem: you are given a dynamically large array of NNN floating point numbers, and you are asked to find the index of the smallest one (commonly called arg min⁡\argminargmin) as fast as possible. When optimizing this, my first instinct was to implement our own comparator function, using the natural partial order of floats. Based on this, we can use a very elegant property of floating point representation: you can sort the f32 values as u32, if you have only positive numbers.

Get the Android app