Running Glucas correctly? - Page 2

rogue · 2003-12-14, 02:06

Comparing the speed of Glucas to Prime95 is like comparing apples and oranges. Prime95 has hand optimized SSE2 code for the PIV. Neither Glucas nor Mlucas can use Altivec on the G4 because Altivec doesn't support double precision, which is a requirement for doing DWTs. The G5 is a different matter. Although there won't much of a gain from 64-bit vs 32-bit, there would be a significant gain from having two FPUs on each chip. Compile Glucas or Mlucas with xlc and you see that the G5 with Glucas/Mlucas is much closer in speed to the PIV, and possibly even faster, although nobody has done this yet. Another thread does have speed of Glucas on a G5, but that was compiled with gcc. xlc is much better at optimizing for the G5 than gcc.

GP2 · 2003-12-14, 02:58

Well then, it would be extremely cool to have an Altivec-aware binary version of Glucas for the G5, compiled with xlc.

We really need a binary version of Glucas for G5, so that it can use the best optimizing compiler (probably proprietary), rather than expecting each user to compile it with the free gcc compiler. Making it work automatically with PrimeNet would be the icing on the cake.

bayanne · 2003-12-14, 15:29

I hope so because my G5 duallie is due to arrive sometime in Jan and I would really love to do some serious crunching

rogue · 2003-12-14, 17:08

Unfortunately, Altivec is relatively unchanged on the G5. In other words, it still doesn't support doubles. I have heard rumors of a future version of Altivec (256 bit vs 128 bit) that will support doubles, but unless Apple pushes for it, it won't happen.

On the other hand, there might be pieces of code in Glucas that could benefit from Altivec, but I really don't know.

Making it (and other clients) work with primenet is supposed to be one of the more important features of the new server code, whenever that happens.

GP2 · 2003-12-14, 20:27

Quote:

Originally posted by rogue
Making it (and other clients) work with primenet is supposed to be one of the more important features of the new server code, whenever that happens.

I believe it could be done with the existing server as well.

I made a preliminary effort to familiarize myself with Glucas and mprime source code, to see how difficult it would be. I'm having a problem running my compiled version of mprime... it segmentation faults in gwsquare. This is using gcc and mprime version 23.6... anybody else tried to compile the mprime source code under Linux?

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Are these polynomials being generated correctly?	Sam Kennedy	Factoring	0	2016-12-02 19:01
LLR not doing checkpoints correctly on some numbers	Puzzle-Peter	Conjectures 'R Us	0	2013-08-15 13:45
OS X Glucas build	rtharper	Software	3	2007-06-13 23:28
Glucas and GIMPS	optim	Software	6	2004-04-05 21:32
GLucas....	bayanne	Software	5	2003-08-15 16:14

2003-12-14, 02:06	#12
rogue "Mark" Apr 2003 Between here and the 6354₁₀ Posts	Comparing the speed of Glucas to Prime95 is like comparing apples and oranges. Prime95 has hand optimized SSE2 code for the PIV. Neither Glucas nor Mlucas can use Altivec on the G4 because Altivec doesn't support double precision, which is a requirement for doing DWTs. The G5 is a different matter. Although there won't much of a gain from 64-bit vs 32-bit, there would be a significant gain from having two FPUs on each chip. Compile Glucas or Mlucas with xlc and you see that the G5 with Glucas/Mlucas is much closer in speed to the PIV, and possibly even faster, although nobody has done this yet. Another thread does have speed of Glucas on a G5, but that was compiled with gcc. xlc is much better at optimizing for the G5 than gcc.

2003-12-14, 02:58	#13
GP2 Sep 2003 5031₈ Posts	Well then, it would be extremely cool to have an Altivec-aware binary version of Glucas for the G5, compiled with xlc. We really need a binary version of Glucas for G5, so that it can use the best optimizing compiler (probably proprietary), rather than expecting each user to compile it with the free gcc compiler. Making it work automatically with PrimeNet would be the icing on the cake.

2003-12-14, 15:29	#14
bayanne "Tony Gott" Aug 2002 Yell, Shetland, UK 2²·83 Posts	I hope so because my G5 duallie is due to arrive sometime in Jan and I would really love to do some serious crunching

2003-12-14, 17:08	#15
rogue "Mark" Apr 2003 Between here and the 2·3²·353 Posts	Unfortunately, Altivec is relatively unchanged on the G5. In other words, it still doesn't support doubles. I have heard rumors of a future version of Altivec (256 bit vs 128 bit) that will support doubles, but unless Apple pushes for it, it won't happen. On the other hand, there might be pieces of code in Glucas that could benefit from Altivec, but I really don't know. Making it (and other clients) work with primenet is supposed to be one of the more important features of the new server code, whenever that happens.