Messages in this thread | | | Date | Tue, 17 Nov 1998 16:11:19 +0100 | From | Erik Corry <> | Subject | Re: The Kommunity vs. Dick Johnson |
| |
Marc Lehmann <pcg@goof.com> wrote: > Actually, I get:
> Change : 1.86 times faster
> with -O6 (pgcc), and
> Change : 0.54 times faster
Marc P-II 333 pgcc -O2 1.86 times faster Marc P-II 333 pgcc -O6 -funroll-all-loops 0.54 times faster Erik PPro 180 gcc-2.7.2.3 -O2 13149 2.5 times faster Erik PPro 180 gcc-2.7.2.3 -O6 13261 2.5 times faster Erik PPro 180 gcc-2.7.2.3 -O6 -funroll-all-loops 14125 1.45 times faster
Using the i386 checksum code from 2.0.34/2.1.107 instead of the C code:
Erik PPro 180 gcc-2.7.2.3 -O2 13354 1.29 times faster Erik PPro 180 gcc-2.7.2.3 -O6 -funroll-all-loops 14778 0.08 times faster
(You can get the code for this from http://www.image.dk/~ehcorry/linux/chksum.c It slots into the files Richard distributed.)
I conclude that -funroll-all-loops is ruining the test, probably optimising the timing loop out of existence. Either that, or it is unrolling a loop inside an inline asm statement, which I don't believe.
Also, on a PPro, Richard's code is probably 1.29 times faster than what we already have.
This whole discussion has been a little unfocused. Richard seems to think he is up against people who don't like assembler, but the real kernel developers have nothing against assembler in places where it is useful. The proof is that the routine he chose as an example is already written in assembler on most architectures. This seems to have been missed by most of the people who have attacked Richard.
Now that Richard has proven that he is a good assembler hacker, it would be nice if he (or someone else) made a patch, so we all got more out of it than a warm glow by proxy. Of course one will have to adapt his code so it does partial checksums of fragments and also does the copy-and-checksum tricks, and it remains to be seen whether the benefits are still visible in real usage (or at least in lmbench) and on x86 processors that are non-Intel
-- "It's not so much an afterlife, more a sort of apres-vie." --DNA -- Erik Corry erik@arbat.com Ceterum censeo, Microsoftem esse delendam!
- 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/
| |