Messages in this thread | | | Subject | RE: [PATCH] add slice by 8 algorithm to crc32.c | From | Joakim Tjernlund <> | Date | Mon, 8 Aug 2011 09:11:38 +0200 |
| |
"Bob Pearson" <rpearson@systemfabricworks.com> wrote on 2011/08/05 17:51:00: > > > > version. While I haven't done the experiment you suggest there is > > something > > > to the point that the second q computation in the new version can be > > moved > > > ahead of the table lookups from the first q computation . My guess is > that > > > the unrolled version will be significantly slower. > > > > Ah, didn't see that. Don't understand how this works though. > > Why do you do two 32 bit loads instead of one 64 bit load? > > > > > > > The two expression trees can be computed in parallel and combined with the > final xor. If the compiler/instruction scheduler are smart enough and can > process enough instructions per cycle they overlap well and you get some > speedup. I did try a 64 bit load on Nehalem but got about 2 cycles per byte > which is a little worse than doing two loads and better than the 32 bit > version. I'm not really sure why.
I see. There is one thing that strikes me as a bit odd. Why do 8 byte alignment when you do 32 bit accesses? Surely the current 4 byte alignment will do?
Jocke
| |