Works for me, the switcher chooses the correct app -
CPUID 1058619, AMD Athlon(tm) 64 X2 Dual Core Processor 4800+ [Family 15 Model 35 Stepping 2] Linux 2.6.22-14-generic
No confirmed results yet but it really steps right along, ~~ 23,000 vs 33,000.
Given how difficult it seems to have been to incorporate the SSE instruction set in the Einstein project, even given the good speedup that it contributes to use it, I have to wonder has any program anywhere ever bothered to use this instruction set?
Now for a really dumb question, why is it so difficult to use? Why is there no math lib that incorporates SSE in an easy enough and good enough way? I'm afraid I have not followed discussions like this here, so maybe this is already well understood by many.
Thanks to all for bringing this nice speedup about, for making it happen.
Now for a really dumb question, why is it so difficult to use? Why is there no math lib that incorporates SSE in an easy enough and good enough way? I'm afraid I have not followed discussions like this here, so maybe this is already well understood by many.
Compilers. These give you cross platform binaries without too much twiddling, however you don't necessarily get optimal solutions for your problem. To conform to specification, presumably making them sale-able products, they only have to emit code that emulates virtual behaviours ( 'as if' ) where outlined in some language standard ( C/C++ in this instance ). There is heaps of implementation latitude right there. The store-forward-stall issue ( mentioned somewhere around here recently ) is an excellent example of incredibly dumb, but nonetheless quite compliant code. Market wise, I suppose you're only going to get what the compiler mechanics choose to aim for in performance terms, given their own constraints. They'll want to sell, so they won't go niche-ing about.
[aside]
For my enjoyment I've used the Greg Comeau line - rock solid on standards compliance - with Plauger's libraries. But C++ Comeau, in effect, emits an intermediate format from your source and then calls a native/existing C compiler to transform that to a binary. While he has written a bucket of ports, I'm unsure what they could offer up to solve E@H's quandary though.
[/aside]
Cheers, Mike.
[ edit - gratuitous MS bash/pummel ] M$ has always leaned away from producing industry standardised software. With Vista they have made a hash of trying, yet again, to lead us over to the BillSide. It's bombing badly, and I think he's left/leaving because he knows that we know that he's not going to get any more Monopoly Money.
I have made this letter longer than usual because I lack the time to make it shorter ...
... and my other CPU is a Ryzen 5950X :-) Blaise Pascal
Works for me, the switcher chooses the correct app -
CPUID 1058619, AMD Athlon(tm) 64 X2 Dual Core Processor 4800+ [Family 15 Model 35 Stepping 2] Linux 2.6.22-14-generic
No confirmed results yet but it really steps right along, ~~ 23,000 vs 33,000.
I have confirmed results on AMD Athlon(tm) 64 X2 (4200,5600,6000), Fedora 7.
Times in the 17000 to 19000 range as opposed to 30000+ for the old app.
Good Job! :)
Your optimised apps version 4.38 and 4.35 works very well on my computers. I have a question. When do you release it as stable ? This question is for 4.38 version. The Alliance Francophone Team members have to choose between your project and others for a race during fifteen days and it will be great if you could provide with your optimised apps as default applications.
Your optimised apps version 4.38 and 4.35 works very well on my computers. I have a question. When do you release it as stable ? This question is for 4.38 version. The Alliance Francophone Team members have to choose between your project and others for a race during fifteen days and it will be great if you could provide with your optimised apps as default applications.
I've just removed winxp and put Debian (Linux kernel 2.6.18-6-686) on my Intel(R) Pentium(R) M processor 1.70GHz model 13. I used beta clients on both systems, but my performance has dropped significantly. Before, average time per unit was about 30000-35000s, now it's about 50000. There is a difference in benchmark results also.
On other systems I have Ubuntu 7.10 (Linux kernel 2.6.22-14-generic) and there I can see no difference between winxp and Linux.
What is the problem and how can I fix it? Does this mean that Debian is in general slower for other applications too (for example Matlab or Mathematica) or is it just E@H?
I've just removed winxp and put Debian (Linux kernel 2.6.18-6-686) on my Intel(R) Pentium(R) M processor 1.70GHz model 13. I used beta clients on both systems, but my performance has dropped significantly. Before, average time per unit was about 30000-35000s, now it's about 50000. There is a difference in benchmark results also.
On other systems I have Ubuntu 7.10 (Linux kernel 2.6.22-14-generic) and there I can see no difference between winxp and Linux.
What is the problem and how can I fix it? Does this mean that Debian is in general slower for other applications too (for example Matlab or Mathematica) or is it just E@H?
Check whether Linux is running your processor at the full 1.70GHz - it may have dropped down to a power-save mode if there is little high-priority load on the CPU.
ubuntu linux 7.10 stderr
)
ubuntu linux 7.10
stderr out:
5.10.28
Detected CPU type 1
2008-03-02 15:35:16.6591 [normal]: Built at: Feb 21 2008 15:57:05
CPU time: 20601.49
resultid=93050952
switching between different apps works, (seti, astropulse, milkyway)
Matthias
I can confirm that.
)
I can confirm that.
Works for me, the switcher
)
Works for me, the switcher chooses the correct app -
CPUID 1058619, AMD Athlon(tm) 64 X2 Dual Core Processor 4800+ [Family 15 Model 35 Stepping 2] Linux 2.6.22-14-generic
No confirmed results yet but it really steps right along, ~~ 23,000 vs 33,000.
Given how difficult it seems to have been to incorporate the SSE instruction set in the Einstein project, even given the good speedup that it contributes to use it, I have to wonder has any program anywhere ever bothered to use this instruction set?
Now for a really dumb question, why is it so difficult to use? Why is there no math lib that incorporates SSE in an easy enough and good enough way? I'm afraid I have not followed discussions like this here, so maybe this is already well understood by many.
Thanks to all for bringing this nice speedup about, for making it happen.
All the best.
RE: Now for a really dumb
)
Compilers. These give you cross platform binaries without too much twiddling, however you don't necessarily get optimal solutions for your problem. To conform to specification, presumably making them sale-able products, they only have to emit code that emulates virtual behaviours ( 'as if' ) where outlined in some language standard ( C/C++ in this instance ). There is heaps of implementation latitude right there. The store-forward-stall issue ( mentioned somewhere around here recently ) is an excellent example of incredibly dumb, but nonetheless quite compliant code. Market wise, I suppose you're only going to get what the compiler mechanics choose to aim for in performance terms, given their own constraints. They'll want to sell, so they won't go niche-ing about.
[aside]
For my enjoyment I've used the Greg Comeau line - rock solid on standards compliance - with Plauger's libraries. But C++ Comeau, in effect, emits an intermediate format from your source and then calls a native/existing C compiler to transform that to a binary. While he has written a bucket of ports, I'm unsure what they could offer up to solve E@H's quandary though.
[/aside]
Cheers, Mike.
[ edit - gratuitous MS bash/pummel ] M$ has always leaned away from producing industry standardised software. With Vista they have made a hash of trying, yet again, to lead us over to the BillSide. It's bombing badly, and I think he's left/leaving because he knows that we know that he's not going to get any more Monopoly Money.
I have made this letter longer than usual because I lack the time to make it shorter ...
... and my other CPU is a Ryzen 5950X :-) Blaise Pascal
Working fine on Fedora 8 as
)
Working fine on Fedora 8 as well, Core 2 Quad Q6600
RE: Works for me, the
)
I have confirmed results on AMD Athlon(tm) 64 X2 (4200,5600,6000), Fedora 7.
Times in the 17000 to 19000 range as opposed to 30000+ for the old app.
Good Job! :)
Hi, Your optimised apps
)
Hi,
Your optimised apps version 4.38 and 4.35 works very well on my computers. I have a question. When do you release it as stable ? This question is for 4.38 version. The Alliance Francophone Team members have to choose between your project and others for a race during fifteen days and it will be great if you could provide with your optimised apps as default applications.
Thanks.
elgrande71
RE: Hi, Your optimised
)
I'll do this in the next few hours.
BM
BM
I've just removed winxp and
)
I've just removed winxp and put Debian (Linux kernel 2.6.18-6-686) on my Intel(R) Pentium(R) M processor 1.70GHz model 13. I used beta clients on both systems, but my performance has dropped significantly. Before, average time per unit was about 30000-35000s, now it's about 50000. There is a difference in benchmark results also.
On other systems I have Ubuntu 7.10 (Linux kernel 2.6.22-14-generic) and there I can see no difference between winxp and Linux.
What is the problem and how can I fix it? Does this mean that Debian is in general slower for other applications too (for example Matlab or Mathematica) or is it just E@H?
RE: I've just removed winxp
)
Check whether Linux is running your processor at the full 1.70GHz - it may have dropped down to a power-save mode if there is little high-priority load on the CPU.