Radix sort has linear time complexity, O(kN), where k is the number of radices per value, and N is the number of values.
How is this possible? The theoretical lower bound of O(N log N) only applies to comparison-based sorting algorithms, whereas radix sort doesn't actually perform any comparisons on the input data.
var sort = radixsort(), data = new Float32Array([…]); var sorted = sort(data); // You can also preallocate the auxiliary array… sorted = sort(data, new Float32Array(data.length));
Note that radix sort modifies the input array, even though it uses an auxiliary array too. In fact, the sorted result will be the input array when an even number of radixes are in use, which is currently always the case.
The sorter always returns a buffer representing the sorted result, so you can pass this to the appropriate typed array constructor, as in the example above.
The most common usage scenario for this will probably be sorting 32-bit floats e.g. for geometry algorithms. My informal benchmark repeatedly sorts an array of 65,536 random 32-bit floats.
- Radixsort.js: ~411 sorts per second.
Radixsort.js is roughly 16x faster! The speed difference gets even larger as you increase the input size.