As one more comment, because I've made the same mistake in the past, it is often a bad idea to use the length in the initialization vector of a general purpose hash because it means you can't make an incremental version of the hashing algorithm e.g. a stream of data of unknown size. This is a trivial change but it has side effects in how you pad partial blocks.

If the length is not part of the initialization vector, then the padding vector for partial blocks should vary as a function of partial data length. Otherwise it is trivial to create collisions by having partial blocks of different sizes that mimic the padding vector. Fixing this can be as easy as using _mm_set1_epi8(partial_length) as your padding vector.