pcclark4 / cggml Goto Github PK

Not everything should be in the top-level src/ directory, obviously. Sorting is a completely different realm than geometry or graph operations for example.

https://en.wikipedia.org/wiki/Cycle_sort
https://rosettacode.org/wiki/Sorting_algorithms/Cycle_sort#C

Keep in mind the "swap" that happens isn't an in-place swap within the list. It's a swap with a tmp variable, and this is required for the algorithm to work. This is why we need the caller to pass us a pointer to store the tmp value in. Otherwise we could just do a simple byte-wise swap.

LSD radix sort is do-able on embedded platforms because it is recursion-free, unlike MSD radix sort.

Using keygen_func like the generic counting sort.

This reduces the number of outer loop iterations from 32 to 4, with only mild space increase. This should give a good speedup in performance. However, I wonder if the extra 253 bytes of storage needed could be detrimental to embedded support? Maybe we need a flag or an overload for the other strategy instead?

Big if true. Let's call it phalloc

On non-embedded systems it can simply proxy to malloc, calloc or alloca

This requires splitting the input array into negative and nonnegative numbers, sorting them independently, and then merging it back together at the end.