The first thing I tried was 128 bit additions of 64 bit quantities. The problem is that it doesn't run any faster than fletcher4 on a 32 bit machine, and fletcher4 on 32bit machines is almost 7 times slower than fletcher2.
I did some further testing, and fletcher2 does not look as bad as in my last estimate - no single bit error detection failures in my tests. I have confirmed failures where two 64 bit MSBs are flipped. I expect an error detection failure when any even number of 64 bit MSBs in the same vertical stripe are flipped from what they should be. With "fletcher8", I have yet to discover any dual bit failures and I don't expect any. -- This message posted from opensolaris.org
