llrCUDA version 3.8.3 is released!

Jean Penné · 2021-02-12, 20:30

Hi All,

I released now a new GPU version of the LLR program on my personal page : jpenne.free.fr

No much new feature, but some improvements related to reliability and speed.
- By default, all tests on base two numbers use Gerbicz error checking.
This is the case for PRP Fermat and SPRP tests as in Prime95 or Mprime,
but also for the deterministic prime tests of Proth numbers.

LLR tests on Riesel numbers are only done after a positive Fermat PRP result.
Also, if b==2, k==+1 and abs(c)==1, a random shift on the PRP base is done.
It is especially interesting for the prime test of Gaussian Mersenne norms.
Like the previous versions, this code is fully C and C++ written, no Assembler code.

Large numbers (at least 1 mega digits) benefit more from the GPU
parallelism, but this program may also be used on smaller positive
results for verification...

For more details, would you see the Readme.txt file.

Please, let me know if you have any problem to run the binary on Linux and/or to build it on your system.

I wish you many successes in prime hunting!
Best Regards,
Jean

rogue · 2021-02-12, 22:46

Any plans for an OpenCL version? I can't run CUDA on AMD.

pepi37 · 2021-02-12, 23:16

Quote:

Originally Posted by Jean Penné

Hi All,

I released now a new GPU version of the LLR program on my personal page : jpenne.free.fr

No much new feature, but some improvements related to reliability and speed.
- By default, all tests on base two numbers use Gerbicz error checking.
This is the case for PRP Fermat and SPRP tests as in Prime95 or Mprime,
but also for the deterministic prime tests of Proth numbers.

LLR tests on Riesel numbers are only done after a positive Fermat PRP result.
Also, if b==2, k==+1 and abs(c)==1, a random shift on the PRP base is done.
It is especially interesting for the prime test of Gaussian Mersenne norms.
Like the previous versions, this code is fully C and C++ written, no Assembler code.

Large numbers (at least 1 mega digits) benefit more from the GPU
parallelism, but this program may also be used on smaller positive
results for verification...

For more details, would you see the Readme.txt file.

Please, let me know if you have any problem to run the binary on Linux and/or to build it on your system.

I wish you many successes in prime hunting!
Best Regards,
Jean

Any timings, how fast or slow? I will try it but there is no binaries for windows...

Jean Penné · 2021-02-13, 07:18

Quote:

Originally Posted by pepi37

Any timings, how fast or slow? I will try it but there is no binaries for windows...

I cannot presently build this program on windows, but it is open source and any developper on a Windows 64 platform would be welcome.

Best Regards,
Jean

Jean Penné · 2021-02-15, 20:09

Hi All,

While Fermat testing k*b^n+c large numbers with c<0, and Gerbicz error checking activated, the final computation of the residue used a call to the invg() function in the giants.c code.
This code being CPU only, it is very time consuming for mega_digits numbers and so, its use must be avoided.

That is the only fix done in this new build of llrCUDA.

Would you excuse me for this drawback, and Best Regards,
Jean

Trilo · 2021-02-15, 21:05

Hi, I have a machine with 3 cards: 2x gtx1060 6gb and a 980 ti. How would I ensure that I can run LLRcuda instances on all 3 cards? Just run 3 instances of it?

paulunderwood · 2021-02-28, 15:28

If you have tried llrCuda out please post how well it does. How many AVX2 cores is it equivalent to on a top end nVidia card?

storm5510 · 2021-02-28, 15:56

Quote:

Originally Posted by Jean Penné

I cannot presently build this program on windows, but it is open source and any developper on a Windows 64 platform would be welcome.

Best Regards,
Jean

What specific language would be required for a Windows 10 x64 build?

Jeff Gilchrist · 2021-03-03, 02:22

What kind of speedup should be expected using llrCUDA compared to using a CPU?

I just tried with a P100 GPU:

ABC$a*$b^$c$d
1 2 13377491 -31

Using complex rational base DWT and generic reduction, FFT length = 1867776, a = 3
2^13377491-31 is not prime. RES64: B7A37D7DABBAAC31. Time : -1614496548690.000 ms.

Some kind of bug with time output but it took about 38 hours to complete with the 1 P100 GPU so considerably slower than using a CPU unless I am doing something wrong. This is with beta2 ( Primality Testing of k*b^n+/-1 Program - GPU Version 3.8.3 ; linked with CUDA Version 8.0.44 ).

Jean Penné · 2021-03-10, 20:32

Hi All,

When a too large round off error was encountered, the test was restarted from the beginning in "cufft only" mode, which was very penalizing about elapsed time.
I fixed that in this new build, so the restart is now from the last saved intermediate file.

Sorry for this drawback and Best Regards,

Jean

Jeff Gilchrist · 2021-03-12, 18:10

Just did another test using our latest PRP with beta3:

llrCUDA - GPU Version 3.8.3b3 ; linked with CUDA Version 8.0.44
2^13380298-27 is a Fermat Probable prime! (4027872 decimal digits) Time : 136632.210 sec.

By comparison using 4 threads on an Intel Core i7-6700K took (for base 3-Fermat PRP test): 12456.743 sec.

So the GPU version is running about 11 times slower. Is that what you would expect for speed with a fermat number like this? Are there specific types of numbers that should be a lot faster with the GPU code?

2021-02-12, 20:30	#1
Jean Penné May 2004 FRANCE 1173₈ Posts	llrCUDA version 3.8.3 is released! Hi All, I released now a new GPU version of the LLR program on my personal page : jpenne.free.fr No much new feature, but some improvements related to reliability and speed. - By default, all tests on base two numbers use Gerbicz error checking. This is the case for PRP Fermat and SPRP tests as in Prime95 or Mprime, but also for the deterministic prime tests of Proth numbers. LLR tests on Riesel numbers are only done after a positive Fermat PRP result. Also, if b==2, k==+1 and abs(c)==1, a random shift on the PRP base is done. It is especially interesting for the prime test of Gaussian Mersenne norms. Like the previous versions, this code is fully C and C++ written, no Assembler code. Large numbers (at least 1 mega digits) benefit more from the GPU parallelism, but this program may also be used on smaller positive results for verification... For more details, would you see the Readme.txt file. Please, let me know if you have any problem to run the binary on Linux and/or to build it on your system. I wish you many successes in prime hunting! Best Regards, Jean

2021-02-15, 20:09	#5
Jean Penné May 2004 FRANCE 5·127 Posts	llrCUDA version 3.8.3 Build 2 is released! Hi All, While Fermat testing k*b^n+c large numbers with c<0, and Gerbicz error checking activated, the final computation of the residue used a call to the invg() function in the giants.c code. This code being CPU only, it is very time consuming for mega_digits numbers and so, its use must be avoided. That is the only fix done in this new build of llrCUDA. Would you excuse me for this drawback, and Best Regards, Jean

2021-03-10, 20:32	#10
Jean Penné May 2004 FRANCE 5×127 Posts	llrCUDA version 3.8.3 Build 3 is released! Hi All, When a too large round off error was encountered, the test was restarted from the beginning in "cufft only" mode, which was very penalizing about elapsed time. I fixed that in this new build, so the restart is now from the last saved intermediate file. Sorry for this drawback and Best Regards, Jean

2021-03-12, 18:10	#11
Jeff Gilchrist Jun 2003 Ottawa, Canada 3×17×23 Posts	Just did another test using our latest PRP with beta3: llrCUDA - GPU Version 3.8.3b3 ; linked with CUDA Version 8.0.44 2^13380298-27 is a Fermat Probable prime! (4027872 decimal digits) Time : 136632.210 sec. By comparison using 4 threads on an Intel Core i7-6700K took (for base 3-Fermat PRP test): 12456.743 sec. So the GPU version is running about 11 times slower. Is that what you would expect for speed with a fermat number like this? Are there specific types of numbers that should be a lot faster with the GPU code? Last fiddled with by Jeff Gilchrist on 2021-03-12 at 18:11

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
LLR Version 3.8.21 Released	Jean Penné	Software	26	2019-07-08 16:54
LLR Version 3.8.16 released	Jean Penné	Software	38	2015-12-10 07:31
LLR Version 3.8.15 released	Jean Penné	Software	28	2015-08-04 04:51
LLR Version 3.8.9 released	Jean Penné	Software	37	2013-10-31 08:45
llr 3.8.2 released as dev-version	opyrt	Prime Sierpinski Project	11	2010-11-18 18:24

2021-02-12, 22:46	#2
rogue "Mark" Apr 2003 Between here and the 2⁴×467 Posts	Any plans for an OpenCL version? I can't run CUDA on AMD.

2021-02-15, 21:05	#6
Trilo "W. Byerly" Aug 2013 81*2^3174353-1 7·19 Posts	Hi, I have a machine with 3 cards: 2x gtx1060 6gb and a 980 ti. How would I ensure that I can run LLRcuda instances on all 3 cards? Just run 3 instances of it?

2021-02-28, 15:28	#7
paulunderwood Sep 2002 Database er0rr 2·7⁴ Posts	If you have tried llrCuda out please post how well it does. How many AVX2 cores is it equivalent to on a top end nVidia card?

2021-03-03, 02:22	#9
Jeff Gilchrist Jun 2003 Ottawa, Canada 3×17×23 Posts	What kind of speedup should be expected using llrCUDA compared to using a CPU? I just tried with a P100 GPU: ABC$a$b^$c$d 1 2 13377491 -31 Using complex rational base DWT and generic reduction, FFT length = 1867776, a = 3 2^13377491-31 is not prime. RES64: B7A37D7DABBAAC31. Time : -1614496548690.000 ms. Some kind of bug with time output but it took about 38 hours to complete with the 1 P100 GPU so considerably slower than using a CPU unless I am doing something wrong. This is with beta2 ( Primality Testing of kb^n+/-1 Program - GPU Version 3.8.3 ; linked with CUDA Version 8.0.44 ).