PFGW latest well-tested version

[Prime95]

I will look at this in the next release which is a long way off. Almost all of the FFT code is being rewritten. This will take a while to complete.

[henryzz]

Originally Posted by Prime95

I will look at this in the next release which is a long way off. Almost all of the FFT code is being rewritten. This will take a while to complete.

that sounds like 26.x
am i correct?

[PCZ]

Thanks
Seeing that error pop up was troubling.

[Prime95]

Originally Posted by henryzz

that sounds like 26.x
am i correct?

Yes.

[rogue]

Originally Posted by Prime95

I will look at this in the next release which is a long way off. Almost all of the FFT code is being rewritten. This will take a while to complete.

Wow. That sounds like quite a large task. Would you mind sharing some of the improvements you are working on?

[Prime95]

Originally Posted by rogue

Would you mind sharing some of the improvements you are working on?

Right now the code is organized such that it can only be optimized for two different CPUs. I've reorganized so that it can be optimized for several different CPUs.

Right now it only supports one FFT style - a non-standard radix-4 with Bailey multipliers between passes. I've added a mechanism to have several different FFT styles in the same gwnum library. This lets me compare the current non-standard to a traditional radix-4, a radix-8, maybe a split-radix, etc. Some architectures might be faster with different FFT styles. For example, the Pentium 4 pays a big penalty for memory accesses so my non-standard FFT was optimized for minimum memory accesses. The Core i5/i7 has much less memory penalty - so it makes sense to do fewer complex multiplies at the cost of more memory accesses.

D.J.Bernstein's "exponent-1" FFTs which reduces the amount of memory required for sin/cos twiddles.

Some other goodies which I'll keep secret for now :)

[Mini-Geek]

Out of nearly 2.5 million primes verified for a part of Riesel base 3, these 6 all return PRP, return composite with -tp but no -a1, and return prime with -tp and -a1:

800157112*3^64-1
800829524*3^47-1
801245282*3^50-1
802324048*3^490-1
802829672*3^59-1
804875072*3^71-1

I doubt that anybody cares or that anything needs to be done, just felt like reporting that little fact.

[rogue]

Originally Posted by Mini-Geek

Out of nearly 2.5 million primes verified for a part of Riesel base 3, these 6 all return PRP, return composite with -tp but no -a1, and return prime with -tp and -a1:

800157112*3^64-1
800829524*3^47-1
801245282*3^50-1
802324048*3^490-1
802829672*3^59-1
804875072*3^71-1

I doubt that anybody cares or that anything needs to be done, just felt like reporting that little fact.

We should keep track of failures for testing the next release of gwnum/pfgw.

[PCZ]

I am currently testing base 7 numbers using prpnet and pfgw without issue.
Have done about 5000 pairs so far and no round off errors at all.

Seems very speedy.
Almost as fast as LLR would do a base 2 number with an equivalent number of digits.
I am keeping the k's very small though guess this helps speed things up.
So no i'm not working in the 7001 7999 range that was just a test.

In the past i have quickly abandoned any non base 2 work because the available software was so slow.

George i have a few i7's and i5's here, and a pile of work if you need a beta tester :)

PS
The error i reported earlier was caused by a PC with bad memory.
Fortunately i weeded it out quickly and no real work was lost.

Rogue.
I have prpnet loaded up with the riesels, if i add the serps as well can you forsee any issues ?.
Can i mix them in the candidates file ?

[rogue]

Originally Posted by PCZ

I have prpnet loaded up with the riesels, if i add the serps as well can you forsee any issues ?.
Can i mix them in the candidates file ?

There shouldn't be any. Just use the correct servertype in the prpserver.ini file. You should also decide if you want to hand out by k or by length, which is set in the ini file. That feature was added in 2.4.7.

[PCZ]

2.4.7 ?
I have 2.4.6.

Want the work handed out by value of n, smallest to largest.
Type 1 i guess.

When i added the serps using prpadmin the server started handing out work at 286k and up.
Not at 250k which is the lowest value of n.

Only after stopping and starting the server did it start handing out serps at 250k, the riesels have already been processed up to 270k.

Edit
Got 2.4.7 downloaded now.