Especially if it's meant to be an educational and in any case non-realtime encoder, intrinsics are not really necessary and make the code less portable.

I also wonder how much the compiler can do autovectorisation on code like this --- it's pretty much exactly the type of code that autovectorisation is intended for.

Edit: I noticed in the benchmark that it compressed the 10s foreman.cif (demo video) in half a second, so it's already 20x faster than realtime on that small resolution.

It's also single core comparison. Multi-threaded encoding is supported too: with 4 threads I get 0,190s for minih264 and 0,135s for x264.