My guess would be stalls due to unavailability of data from the cache in time for the pipeline. Some insertions of nop's would probably make this go away. That may or may not affect the timings on Opteron.
I suspect it is only an issue on cheaper processors like the Athlon 64 x2 and Turion 64 x2. I have one of the latter, and it would be interesting to try out. I have linux on this machine so I could possibly give it a go. But it might be tomorrow before I get to it. I also have an Athlon 64 x2 at home. It's not clear how much effort we should expend on optimisation for home market PC's. I'm not saying we shouldn't do it. I'm just saying it is not clear how much we should do for it. Bill. 2009/3/4 <ja...@njkfrudils.plus.com>: > - Show quoted text - > On Wednesday 04 March 2009 15:57:25 Cactus wrote: >> On Mar 4, 3:49 pm, ja...@njkfrudils.plus.com wrote: >> > On Wednesday 04 March 2009 15:40:04 Bill Hart wrote: >> > > This is on a K10. The runs differ significantly. Interesting. >> > > >> > > It might have to do with the almost completely unpredictable >> > > scheduling on the K10. Certainly the differences at limb n are almost >> > > always made up again, or nearly so, at limb n+1, and by the end of 40 >> > > limbs the times are usually within a cycle or so. On average >> > > performance doesn't vary. >> > > >> > > Bill. >> > >> > My K8 is very stable, the K10(cuda1) a bit less so , and I found the >> > core2 (sage.math) to be all over the place . You could try increasing the >> > precision , yet again. >> > >> > > 2009/3/4 Jeff Gilchrist <jeff.gilchr...@gmail.com>: >> > > > On Wed, Mar 4, 2009 at 10:27 AM, Bill Hart >> > > > <goodwillh...@googlemail.com> >> > >> > wrote: >> > > >> I've just noticed that the times from speed are not consistent >> > > >> between runs!! I thought that was almost impossible. >> > > > >> > > > I did several runs on my systems and they all agreed for me, even >> > > > when the system was loaded or not. >> > > > >> > > > Jeff. >> > > > - Show quoted text - >> >> I do see some variation on my Core 2 portable but my Athlon X2 gives >> consistent results. The variation on the Core 2 is not large enough to >> worry about. >> >> By the way, does anyone have a theory about the mpn_mul_1 curve (2nd >> to bottom) that goes up and down regularly? > > thats the green line , your windows K8 > > mpn_mul_1 and mpn_addmul_1 on K8 are the same function(practically) , but the > addmul doesn't display that behavour. > - Show quoted text - > >> >> Brian >> >> >> >> > > > > > --~--~---------~--~----~------------~-------~--~----~ You received this message because you are subscribed to the Google Groups "mpir-devel" group. To post to this group, send email to mpir-devel@googlegroups.com To unsubscribe from this group, send email to mpir-devel+unsubscr...@googlegroups.com For more options, visit this group at http://groups.google.com/group/mpir-devel?hl=en -~----------~----~----~----~------~----~------~--~---
