Actually, I just realised we already timed the sqr_basecase.as code for atom and it was faster than the generic x86_64 code. So no need to check that again for atom. Just tuning values for atom will suffice.
Bill. On 24 March 2014 13:51, Bill Hart <goodwillh...@googlemail.com> wrote: > It is actually possible that the bobcat version is faster. I've no idea > whether Jason explicitly superoptimised it for bobcat. It's not without > precedent anyway. > > You can check if the CPU throttling is the problem because tuneup actually > reports the CPU speed before it gives the timings. These should be the same > when comparing. If so, the values should be accurate. > > I noticed the non-monotonicity before. By the way, tuning thresholds > should not affect the performance of sqr_basecase. The code is called > directly, with no adjustment of algorithm related to tuning cutoffs. > > You can also test out to 100 limbs with -s 1-100 and see if the > improvement is maintained to larger sizes. Not that this is relevant in > practice. The sqr_basecase cutoff will fairly low (probably around 20-40), > so any sizes above that are probably not relevant. > > Bill. > > > On 24 March 2014 12:12, leif <not.rea...@online.de> wrote: > >> Bill Hart wrote: >> >>> Leif said he was going to tune on a bobcat, but hasn't yet. >>> >> >> Well, this of course^TM turned out to be a can of worms... ;-) >> >> So far it looks like the bobcat version of mpn_sqr_basecase was actually >> faster, but I don't really trust the figures. (I played a little with the >> "precision" option, but this seems to be logically limited to 2^31. With >> the default precision, I occasionally get non-monotonic numbers; not that >> one value was exceptionally bad -- as one would expect, but timings about >> twice as fast, despite both cores being on "performance", and the machine >> mostly idle.) >> >> Still have to analyze how reliable individual thresholds are. >> >> >> -leif >> >> >> That may be because I mentioned there is an additional step for the >>> bobcat. We need to know if mpn/x86_64/bobcat/sqr_basecase.as >>> <http://sqr_basecase.as> or mpn/x86_64/sqr_basecase.asm is faster. >>> >>> >>> This can be determined by running: >>> >>> ./configure >>> make >>> make clean >>> cd tune >>> make speed >>> ./speed -s 1-40 mpn_sqr_basecase >>> >>> before and after deleting mpn/x86_64/bobcat/sqr_basecase.as >>> <http://sqr_basecase.as> >>> >>> >>> Assuming it is faster after deleting it (smaller numbers are better -- >>> and we assume they will be smaller), run make tune and send the results >>> to us. >>> >>> Actually, I just realised, we'd like to know the same thing about >>> mpn/x86_64/atom/sqr_basecase.as <http://sqr_basecase.as> on the atom. >>> >>> >>> Bill. >>> >>> >>> On 24 March 2014 11:12, Frithjof Schulze <sfrith...@gmail.com >>> <mailto:sfrith...@gmail.com>> wrote: >>> >>> On Friday, March 21, 2014 5:37:16 PM UTC, Bill Hart wrote: >>> >>> Note MPIR will run very slow on these machines without proper >>> tuning values due to significant recent changes. So we'd >>> appreciate any help. >>> >>> We can also tell you how to do the tuning yourself and send us >>> the values (it's really easy). >>> >>> >>> I have a netbook with an E-350 and also an older atom based netbook. >>> I could tune MPIR on both, if nobody has done so yet. >>> >>> --Frithjof >>> >> >> -- >> () The ASCII Ribbon Campaign >> /\ Help Cure HTML E-Mail >> >> -- >> You received this message because you are subscribed to the Google Groups >> "mpir-devel" group. >> To unsubscribe from this group and stop receiving emails from it, send an >> email to mpir-devel+unsubscr...@googlegroups.com. >> To post to this group, send email to mpir-devel@googlegroups.com. >> Visit this group at http://groups.google.com/group/mpir-devel. >> For more options, visit https://groups.google.com/d/optout. >> > > -- You received this message because you are subscribed to the Google Groups "mpir-devel" group. To unsubscribe from this group and stop receiving emails from it, send an email to mpir-devel+unsubscr...@googlegroups.com. To post to this group, send email to mpir-devel@googlegroups.com. Visit this group at http://groups.google.com/group/mpir-devel. For more options, visit https://groups.google.com/d/optout.
