There are some adjusments I'd like to try to this algorithm for better integration but I cannot find the time for it.
If you are looking for a bachelor thesis or master's project at TU Vienna and you find this forked repository interesting, write me a mail: [email protected]
This repo contains a few branchless modifications to a traditional rans coder, as a result it's performance is comparable to FSE whilst supporting adaptive probabilities and using no platform specific intrinsics or assembly. All of this is public domain, do whatever you want with it.
A small benchmark on enwik8, Ryzen 1700 @ 3.7ghz:
| Implementation | Encode speed | Decode speed |
|---|---|---|
| 4-way interleaved rans trick | 415 MB/s | 460 MB/s |
| 4-way implicit rans trick | 307 MB/s | 410 MB/s |
| 4-way interleaved rans standard | 250 MB/s | 300 MB/s |
| 4-way implicit rans standard | 225 MB/s | 250 MB/s |
| FSE | 325 MB/s | 440 MB/s |