Linux S5R2 App 4.31 available for Beta test
Bernd Machenschalk
Moderator
Administrator
Joined: 15 Oct 04
Posts: 4312
Credit: 250191663
RAC: 35048
20 Jul 2007 15:11:29 UTC
Topic 192985
(moderation:
A new Linux App is available from our Beta Test page.
This App has the same codebase of the 4.30 Windows App, with the exception of using floor() instead of ftol() in the SIN/COS LUT. My intention is to bring it up to speed with the current official App this way.
(most of the proejct team is travelling to a meeting the next days and may have poor internet connectivity. For the case I may be wrong and this App is even slower than the previous one, I added a link to the previous Beta App on the page.)
BM
BM
Linux S5R2 App 4.31 available for Beta test
)
Have installed it, runs fine so far.
However, from "watching" the stderr stream, I'd say it isn't that much faster with floor(), maybe 5% or so. I guess modf was comparatively fast with Linux because the libm implementations I've seen use bit-manipulation instead of FPU instructions in combination with this not-so-optimal "setting the rounding mode" instruction. I wonder whether it's safe to set the rounding mode at the beginning of the hot-loop (say, to truncation mode) and then use plain inline assembler (2 intructions) to do the truncation. this should work as long as we can rely that anything "in between" does not change the rounding mode (or restores it to its previous (our)) mode. AFAIK this assumption is safe. This approach should have benefits beyond sin/cos function.
CU
BRM
My host has completed 4 WU
)
My host has completed 4 WU mostly with 4.31 and the rest with 4.27:
http://einsteinathome.org/host/948532/tasks&offset=20 (the 4 most recent results)
All four results have validated although crunch time takes just as long as the previous beta.
There are 10^11 stars in the galaxy. That used to be a huge number. But it's only a hundred billion. It's less than the national deficit! We used to call them astronomical numbers. Now we should call them economical numbers. - Richard Feynman
I have completed my first WU
)
I have completed my first WU with 4.31, an another one mixed 4.27/4.31. Both have been validated, with credit granted, no errors.
4.31 is slightly faster than 4.27 om my Duron 1600 Applebred, about a 5%.
But this new beta is still a 17 % slower than the official 4.21 app.
Why are this binaries so much
)
Why are this binaries so much larger than the older version?
The 4.21 binary was 5,1 MiB, the 4.31 is 7,6 MiB and the 4.21 .so was only 1,9 MiB while the 4.31 .so is 4,2 MiB fat.
probably varying amounts of
)
probably varying amounts of debug material trying to run down he last handful of remaining problems.
Plese test the new App. BM
)
Plese test the new App.
BM
BM