Now that AP26 has ended, my three PS3 need something else to do. Seeing as how they got me to an amethyst badge on AP26, I'm hoping to improve my badge here on PSA.
I do have my first PS3 trying out a GCW13 task but I'm looking for recommendations for which servers have work they do the best at.

The following servers (in order of WU length) mentioned by rogue are available to phrot which is used by the PS3 :

• server=PPSElow:0:10:uwin.mine.nu:11000
  
• server=PPSEhigh:0:5:uwin.mine.nu:10000
  
• server=SGS:0:1:prpnet.primegrid.com:12000
  
• server=ESP:0:1:pgllr.mine.nu:9000
  
• server=27121:0:1:prpnet.primegrid.com:12006
  
• server=GCW13:0:1:prpnet.primegrid.com:12004
  

The GFN servers are also mentioned but currently the PS3 build does not have a generfer option:

• server=GFN32768:0:1:prpnet.primegrid.com:12005
  
• server=GFN65536:0:1:prpnet.primegrid.com:12003
  

More information on each port:

PPSElow: Proth Prime Search Extended: k*2^n=1; 1200<k<10000 for n<500K
PPSEhigh: Proth Prime Search Extended: k*2^n=1; 1200<k<10000 for n>500K
SGS: Sophie Germain Prime Search: (k*2^666666-1)
ESP: The extended Sierpinski problem
27121: 27121 Prime Search: k=27 & 121 for k*2^n+-1 for n<10M
GCW: Generalized Woodall & Cullen; b=13 (n*13^n-/+1)
GFN32768: Generalized Fermat Number Prime Search (b^2^32768+1)
GFN65536: Generalized Fermat Number Prime Search (b^2^65536+1)
____________

[quote
The GFN servers are also mentioned but currently the PS3 builds do not have a generfer option:

• server=GFN32768:0:1:prpnet.primegrid.com:12005
  
• server=GFN65536:0:1:prpnet.primegrid.com:12003
  

[/quote]

They don't? That is something that needs to be rectified. I don't know if I provided anyone with the RISC source to genefer, but it should build on PS3. It uses its own FFT, so it isn't reliant on a third party library for that. If anyone with a PS3 want to try to build it, please send me an e-mail (not a PM) and we'll work offline to get it working.

I don't know if I provided anyone with the RISC source to genefer, but it should build on PS3. It uses its own FFT, so it isn't reliant on a third party library for that. If anyone with a PS3 want to try to build it, please send me an e-mail (not a PM) and we'll work offline to get it working.

I don't know if I provided anyone with the RISC source to genefer, but it should build on PS3. It uses its own FFT, so it isn't reliant on a third party library for that. If anyone with a PS3 want to try to build it, please send me an e-mail (not a PM) and we'll work offline to get it working.

I have a genefer-2.2 build for PS3 which also uses the 6 SPU's, but it is not very stable. Further I have a CUDA version, but only tested in CUDAEMU.
If someone has time to look into it more deeply I'll send you the sources.

I have a genefer-2.2 build for PS3 which also uses the 6 SPU's, but it is not very stable. Further I have a CUDA version, but only tested in CUDAEMU.
If someone has time to look into it more deeply I'll send you the sources.

It is probably my sources that you are working with. If you e-mail me we can investigate the stability issues.

I have a genefer-2.2 build for PS3 which also uses the 6 SPU's, but it is not very stable. Further I have a CUDA version, but only tested in CUDAEMU.
If someone has time to look into it more deeply I'll send you the sources.

It is probably my sources that you are working with. If you e-mail me we can investigate the stability issues.

Yes, its based on your sources of genefer 1.3 with its FFT routines ripped out and replaced by Syoichiro Yamada's FFTW implementation then forward-ported to the checkpointing and checks of genefer 2.2.
Since FFTW gained seemless SPU support in version 3.3.1 it now utilizes the SPU's without any change.

The CUDA version is quite similar, instead of the FFT done with FFTW it uses the CUDAFFT libs.
At least in simulation and for smaller exponents it works OK. But there sure is more work to be done for boundary checking. I'll send you both sources by PM.

Only a few short benchmarks, skipping the 64K and above for now,
can't waste that much CPU time during a challenge ;-)


CUDA Emulation Mode sm_13 on a Mac mini 2006 Core Duo 1.66GHz without any NVIDIA chipset
./genefer-2.2cuda-1 genefer.work

5683936^256+1 is a probable composite. (RES=893994a255a3326f) (1730 digits) (err = 0.3750) (time = 0:03:16)
…this is so slow, someone on a faster machine and/or real NVidia card should to this benchmark...



PS3 FFTW-SPU version (using 1 PPU, 6 SPU, on Cell-Blade it should automatically use more SPU's):
./genefer-2.2-ps3 genefer.work

5683936^256+1 is a probable composite. (RES=893994a255a3326f) (1730 digits) (err = 0.2500) (time = 0:00:00)
4616790^512+1 is a probable composite. (RES=05704de8d08c2a0a) (3413 digits) (err = 0.2500) (time = 0:00:00)
3750000^1024+1 is a probable composite. (RES=0f807c291b252057) (6732 digits) (err = 0.2500) (time = 0:00:02)
3045946^2048+1 is a probable composite. (RES=bd72b5713f909aed) (13279 digits) (err = 0.2188) (time = 0:00:08)
2474076^4096+1 is a probable composite. (RES=0b53a7da1c7c9181) (26188 digits) (err = 0.2500) (time = 0:00:26)
2009574^8192+1 is a probable composite. (RES=ab752d28c1e60445) (51636 digits) (err = 0.2500) (time = 0:01:42)
1632282^16384+1 is a probable composite. (RES=eee7f094cb5f4f86) (101791 digits) (err = 0.2500) (time = 0:06:29)
1325824^32768+1 is a probable composite. (RES=7aea6cae5b1b0904) (200622 digits) (err = 0.2500) (time = 0:24:42)
With bigger FFTs PS3 should become really good for this, see http://www.fftw.org/cell/ps3/



Linux 32-bit LLVM-2.7/Clang FFTW version on a Pentium 4 2.8GHz:
./genefer-clang genefer.work

5683936^256+1 is a probable composite. (RES=893994a255a3326f) (1730 digits) (err = 0.2500) (time = 0:00:01)
4616790^512+1 is a probable composite. (RES=05704de8d08c2a0a) (3413 digits) (err = 0.2500) (time = 0:00:00)
3750000^1024+1 is a probable composite. (RES=0f807c291b252057) (6732 digits) (err = 0.3125) (time = 0:00:02)
3045946^2048+1 is a probable composite. (RES=bd72b5713f909aed) (13279 digits) (err = 0.2500) (time = 0:00:06)
2474076^4096+1 is a probable composite. (RES=0b53a7da1c7c9181) (26188 digits) (err = 0.2812) (time = 0:00:27)
2009574^8192+1 is a probable composite. (RES=ab752d28c1e60445) (51636 digits) (err = 0.2812) (time = 0:01:52)
1632282^16384+1 is a probable composite. (RES=eee7f094cb5f4f86) (101791 digits) (err = 0.3125) (time = 0:07:51)
1325824^32768+1 is a probable composite. (RES=7aea6cae5b1b0904) (200622 digits) (err = 0.2812) (time = 0:39:31)
FFTW on Pentium 4 has its highlight on 8K FFTs, see http://www.fftw.org/speed/Pentium4-2.4GHz-gcc/
the Core2 processors seem to shine up to 64K FFTs, see http://www.fftw.org/speed/CoreDuo-3.0GHz-icc/



Linux 32-bit GCC-4.4.2 x86-generic-32 version on a Pentium 4 2.8GHz:
./genefer_x86-gcc genefer.work

5683936^256+1 is a probable composite. (RES=893994a255a3326f) (1730 digits) (err = 0.2767) (time = 0:00:00)
4616790^512+1 is a probable composite. (RES=05704de8d08c2a0a) (3413 digits) (err = 0.3314) (time = 0:00:00)
3750000^1024+1 is a probable composite. (RES=0f807c291b252057) (6732 digits) (err = 0.3495) (time = 0:00:02)
3045946^2048+1 is a probable composite. (RES=bd72b5713f909aed) (13279 digits) (err = 0.3695) (time = 0:00:06)
2474076^4096+1 is a probable composite. (RES=0b53a7da1c7c9181) (26188 digits) (err = 0.3952) (time = 0:00:24)
2009574^8192+1 is a probable composite. (RES=ab752d28c1e60445) (51636 digits) (err = 0.3533) (time = 0:01:42)
1632282^16384+1 is a probable composite. (RES=eee7f094cb5f4f86) (101791 digits) (err = 0.3636) (time = 0:07:10)
1325824^32768+1 is a probable composite. (RES=7aea6cae5b1b0904) (200622 digits) (err = 0.3804) (time = 0:30:23)
the native version is faster so far, but FFTW has a lot of options for tuning and also different CPU models
show completely different speedups
____________

Those are great numbers. How do they compare to genefx64 and genefer80? Have you run it with the -l option to see what the upper limit is for b?

I will find some time in the next few days to see how PPC compares when using FFTW.

I did some testing with the FFTW version of genefer on MacPPC. It performs poorly compared to the RISC version. I've added some of the RISC optimizations for unrolling loops and although that improved it by about 20%, it is still about 20% slower than the code I've been using. What's even worse is that as the FFT size increases, the performance gets worse compared to the RISC version.

Here are the comparable benchmarks:

genefer_risc -b
Generalized Fermat Number Bench
5683936^256+1 Time: 4.16 us/mul. Err: 0.0000 1730 digits
4616790^512+1 Time: 9.04 us/mul. Err: 0.0000 3413 digits
3750000^1024+1 Time: 19.8 us/mul. Err: 0.0000 6732 digits
3045946^2048+1 Time: 42.8 us/mul. Err: 0.0000 13279 digits
2474076^4096+1 Time: 97.3 us/mul. Err: 0.0000 26188 digits
2009574^8192+1 Time: 213 us/mul. Err: 0.0000 51636 digits
1632282^16384+1 Time: 439 us/mul. Err: 0.0000 101791 digits
1325824^32768+1 Time: 965 us/mul. Err: 0.0000 200622 digits
1076904^65536+1 Time: 2.58 ms/mul. Err: 0.0000 395325 digits
874718^131072+1 Time: 6.38 ms/mul. Err: 0.0000 778813 digits
710492^262144+1 Time: 14.4 ms/mul. Err: 0.0000 1533952 digits
577098^524288+1 Time: 33.1 ms/mul. Err: 0.0000 3020555 digits
468750^1048576+1 Time: 77.1 ms/mul. Err: 0.0000 5946413 digits
380742^2097152+1 Time: 170 ms/mul. Err: 0.0000 11703432 digits

./genefer_fftw -b
Generalized Fermat Number Bench
5683936^256+1 Time: 4.54 us/mul. Err: 0.0000 1730 digits
4616790^512+1 Time: 9.99 us/mul. Err: 0.0000 3413 digits
3750000^1024+1 Time: 21.7 us/mul. Err: 0.0000 6732 digits
3045946^2048+1 Time: 48.3 us/mul. Err: 0.0000 13279 digits
2474076^4096+1 Time: 103 us/mul. Err: 0.0000 26188 digits
2009574^8192+1 Time: 262 us/mul. Err: 0.0000 51636 digits
1632282^16384+1 Time: 641 us/mul. Err: 0.0000 101791 digits
1325824^32768+1 Time: 2.04 ms/mul. Err: 0.0000 200622 digits
1076904^65536+1 Time: 8.58 ms/mul. Err: 0.0000 395325 digits
874718^131072+1 Time: 49.7 ms/mul. Err: 0.0000 778813 digits
710492^262144+1 Time: 113 ms/mul. Err: 0.0000 1533952 digits
577098^524288+1 Time: 234 ms/mul. Err: 0.0000 3020555 digits
468750^1048576+1 Time: 344 ms/mul. Err: 0.0000 5946413 digits
380742^2097152+1 Time: 879 ms/mul. Err: 0.0000 11703432 digits

I realized that --fast-math causes problems, so I removed it. That doesn't impact the relative benchmarks, but does impact performance. I need to investigate.

I realized that --fast-math causes problems, so I removed it.

The one causing the problems is -funsafe-math-optimizations, which is part of -ffast-math. Oddly, this does not cause any problems with phrot on my G5. Removing that option is a performance killer, but using it leads to invalid results.