Thanks for the heads up! I will keep a sharper eye on the results, and if they start to not validate I will switch back to D40 (which worked very well).
You can also switch back to S40, that's almost as fast(~0.9% diff.) as S40.03, but faster than D40.
Today I tried the S40.03, it worked fine. Here are the last 3 WU CPU-Times
5,279.44
5,189.41
5,132.86
Before that I used S39L, but I can find no big difference, between them.
Mayby S40.03 is 100 ~ 200 seconds faster
I used a Pentium 2,6GHz
Once you might get WUs of a different frequency and then the scheduler will throw away your results.
My X2 did't have a problem so far, but they may come and the XP also crunched very well, until z1_1373.5 showed up.
Let's see what akos is gonna answer.
Albert uses a very low precision sin/cos look-up table. The rounding at its address generation enlarges the deviations.
example:
before "low precision" operation:
value(S40.03)=0.416723851
value(original)=0.416723849
deviation: 0,00000048%
after "low precision" operation:
value(S40.03)=0.423156123
values(original)=0.40853772
deviation: 3,58% (cca. 7500000 more times bigger difference)
I will modify this part if I will found some free hours.
edit: This rounding problem appears infrequently, but if it appears that means a big problem... (zero credit)
S40.03 works brilliantly for me. S39L for some reason was about 3 times slower than C40 on my machine, so I was running C40 for a while. S40.03 seems to take about two thirds of the time for a WU when compared to C40, so thanks for the optimized app Akosf :)
RE: Thanks for the heads
)
You can also switch back to S40, that's almost as fast(~0.9% diff.) as S40.03, but faster than D40.
cu,
Michael
Today I tried the S40.03, it
)
Today I tried the S40.03, it worked fine. Here are the last 3 WU CPU-Times
5,279.44
5,189.41
5,132.86
Before that I used S39L, but I can find no big difference, between them.
Mayby S40.03 is 100 ~ 200 seconds faster
I used a Pentium 2,6GHz
Or is the S40.03 still for Athlon CPU´s ?
RE: Once you might get WUs
)
Albert uses a very low precision sin/cos look-up table. The rounding at its address generation enlarges the deviations.
example:
before "low precision" operation:
value(S40.03)=0.416723851
value(original)=0.416723849
deviation: 0,00000048%
after "low precision" operation:
value(S40.03)=0.423156123
values(original)=0.40853772
deviation: 3,58% (cca. 7500000 more times bigger difference)
I will modify this part if I will found some free hours.
edit: This rounding problem appears infrequently, but if it appears that means a big problem... (zero credit)
S40.03 works brilliantly for
)
S40.03 works brilliantly for me. S39L for some reason was about 3 times slower than C40 on my machine, so I was running C40 for a while. S40.03 seems to take about two thirds of the time for a WU when compared to C40, so thanks for the optimized app Akosf :)
RE: RE: S40.04 - SSE
)
Akos,
Thank you! (A much better answer than the one Bruce gave me.)
Stick