Martin Mares <mj@ucw.cz> wrote on 2010/04/23 23:02:29:
Hello!
How did you come to the conclusion the the current code was better than the previous version? Seems like "hand waving" to me.
Did I claim anywhere that the old code is better? I only pointed out
You did when you commited it and I asked twice how you came to that conlusion. Pretty much all recent tests shows otherwise.
the lack of arguments about the new code being better, which is a reason to stay with the old, tested code.
I told told you I had benched the "add carry in C" before and it wasn't any better(worse actually).
Actually, back in the ages when I wrote the old checksum function, I have checked that it performs better than a trivial implementation, and now you claim otherwise, so I naturally want to see new data which show that modern hardware behaves differently.
Santiago benched it too and it was better or just as good as before. Only the MIPS had a regression.
If I recall his results correctly, he has performed three tests:
In the 1st one, your code was 20% faster. In the 2nd one, it was of the same speed. In the 3rd one, it was 20% slower.
Maybe I wear different glasses from yours, but I clearly see that on average, there is no improvement.
You have ignored my tests which also show an improvement.
So what now? what more proof do you need?
First of all, I want at least a rudimentary proof that IT MATTERS AT ALL. We are spending lots of time talking about a minor (20%) speedup in a small chunk of code, without having any clue about how often it gets called and what fraction of the total time is really spent there.
20% is a lot and I not going to bench it any further. if you are not happy with the results so far, nothing I can do will change that.