Are there any extant implementations of the groestl that Monero PoW uses that utilize AesNi instructions and byte-slicing?

I’ve been trying to put together a byte-sliced groestl (small) that use AesNi. I’ve seen plenty of groestl (large) implementations that do this, and have been using them for inspiration, but when you’re only dealing with 64 instead of 128 bytes in the p and q rounds, the transpositions, byte offsets, and finite-field arithmetic is a little trickier. I was wondering if anyone knows of a github that has such a thing that I can use for reference.

Thanks in advance!