thanks :) the braiding approach is super clever too, this was one of those weird moments where you find something and then have to triple check your results because how could i accidentally find something better than the algorithm that hasn't been touched in decades...
the part i really like is that it gives us small improvement on the pclmul too, as the non-accelerated algorithm doesn't really stand a chance against the accelerated opcode on newer hardware so it probably isn't going to see much use in practice. however... i think hardware solutions could possibly benefit (e.g. ethernet cards)
I forget, these were the outputs of my test scripts and I wasn't exactly fastidious here, at least the axes all start from zero :)
IIRC they're off by maybe a factor of 10 or 100, the test scripts just generate a bunch of (seeded) random data and then execute the CRC of X bits Y times and that's where the number comes from, it's consistent across the different tests even if the units are wrong
the part i really like is that it gives us small improvement on the pclmul too, as the non-accelerated algorithm doesn't really stand a chance against the accelerated opcode on newer hardware so it probably isn't going to see much use in practice. however... i think hardware solutions could possibly benefit (e.g. ethernet cards)