> I also improved a bit the asm (mainly the csum_copy that wasn't
> unrolling enough according to the new numbers).
Unrolling more eats more instruction cache for incredibly low gain, I
don't think it's a good idea except maybe for dedicated routers. As for
the other changes, such as changing testl %esi,%esi to clc, that hurts the
current code (we spent a LOT of time putting the right instructions in
there, and got some very non-obvious results a lot of times, due to
alignment issues and other PPro weirdness I guess).
I also noticed that you removed the code that handles 2-byte aligned
areas, I don't think that's a very good idea, all the adcl instructions
will probably slow down a lot if they don't operate on aligned data.
If there is interesting in improving the checksum codes, I can try to dig
out my old test program which let you benchmark with various "real-life"
packet sizes with both warm and cold i-cache and d-cache.
- Finn Arne
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/