llrCUDA - Page 9

msft · 2011-02-02, 14:05

3*2^414840-1 is prime! Time : 210.625 sec.
3*2^382449+1 is prime! Time : 190.457 sec.
3*2^414840-1 is prime! Time : 210.625 sec.

msft · 2011-02-02, 17:05

llrcuda.0.39$ ./llrCUDA -d -q5*2^3569154-1
Starting Lucas Lehmer Riesel prime test of 5*2^3569154-1
Using real irrational base DWT, FFT length = 524288
V1 = 4 ; Computing U0...done.
5*2^3569154-1, iteration : 60000 / 3569154 [1.68%]. Time per iteration : 2.520 ms.

msft · 2011-02-03, 02:37

llrcuda.0.40$ ./llrCUDA -q"5*2^15111112-1" -d
Starting Lucas Lehmer Riesel prime test of 5*2^15111112-1
Using real irrational base DWT, FFT length = 2097152
V1 = 4 ; Computing U0...done.
5*2^15111112-1, iteration : 20000 / 15111112 [0.13%]. Time per iteration : 10.593 ms.

llrcuda.0.40$ ./llrCUDA -q"5*2^15111113+1" -d
Starting Proth prime test of 5*2^15111113+1
5*2^15111113+1, bit: 20000 / 15111115 [0.13%]. Time per bit: 10.661 ms.

Not support save file.

Ken_g6 · 2011-02-03, 04:30

Quote:

Originally Posted by henryzz

Quote:

Originally Posted by frmky

And a larger number where the GPU shines...

Code:

./llrCUDA -d -q938237*2^3752950-1
Starting Lucas Lehmer Riesel prime test of 938237*2^3752950-1
Using real irrational base DWT, FFT length = 1048576
V1 = 4 ; Computing U0...done.
938237*2^3752950-1 is prime!  Time : 10740.747 sec.

How long would that take on a pc?

With the current Linux LLR client from PrimeGrid, on a Core 2@3GHz, I estimate about 27,000 seconds.

When checkpointing is working again, I plan to look into this for Seventeen or Bust!

henryzz · 2011-02-03, 16:50

Quote:

Originally Posted by Ken_g6

With the current Linux LLR client from PrimeGrid, on a Core 2@3GHz, I estimate about 27,000 seconds.

When checkpointing is working again, I plan to look into this for Seventeen or Bust!

Nice speed.

mdettweiler · 2011-02-03, 19:32

Finally--I got llrCUDA working on Gary's GPU!

(Thanks nuggetprime for the pointers on getting the drivers installed!)

Firstly:

Code:

gary@herford:~/Desktop/gpu-stuff/llrcuda$ ./llrCUDA -d -q3*2^382449+1
Starting Proth prime test of 3*2^382449+1
3*2^382449+1 is prime!  Time : 183.300 sec.

And then, with frmky's modifications to allow it to use IBDWT:

Code:

gary@herford:~/Desktop/gpu-stuff/llrcuda/experimental$ ./llrCUDA -d -q1623*2^481564-1
Starting Lucas Lehmer Riesel prime test of 1623*2^481564-1
Using real irrational base DWT, FFT length = 131072
V1 = 3 ; Computing U0...done.
1623*2^481564-1 is prime!  Time : 377.385 sec.
gary@herford:~/Desktop/gpu-stuff/llrcuda/experimental$ ./llrCUDA -d -q1623*2^481564+1
Starting Proth prime test of 1623*2^481564+1
1623*2^481564+1 is not prime.  Proth RES64: 5B6D67B9B4648D0E  Time : 387.913 sec.

The last two both took about 200 seconds on a modern CPU, so there's definitely a slowdown due to the bigger k. But llrCUDA got the reisdual right!

msft · 2011-02-04, 08:55

Quote:

Originally Posted by mdettweiler

And then, with frmky's modifications to allow it to use IBDWT:

How do that ?

mdettweiler · 2011-02-04, 16:36

Quote:

Originally Posted by msft

How do that ?

See here:

Quote:

Originally Posted by frmky

Code:

Starting Lucas Lehmer Riesel prime test of 2303*2^251634-1
Using real irrational base DWT, FFT length = 65536
V1 = 4 ; Computing U0...done.
2303*2^251634-1 is not prime.  LLR Res64: 99B680F2A92ECC27  Time : 112.600 sec.

However, to get this I had to modify gwpnumi.cu, set_fftlen(), to remove the checks fftdwt > fftzpad and bpw < 5.0. The values of these parameters are fftdwt = 65536, fftzpad = 32768, and bpw = 4.83979, so in version 0.34, both these checks fail and zero-padded rational base DWT, which runs about 20x slower, is used.

Edit: A second run:

Code:

./llrCUDA -d -q1000065*2^390927-1
Starting Lucas Lehmer Riesel prime test of 1000065*2^390927-1
Using real irrational base DWT, FFT length = 131072
V1 = 5 ; Computing U0...done.
1000065*2^390927-1, iteration : 40000 / 390927 [10.23%].  Time per iteration : 0.627 ms
1000065*2^390927-1 is prime!  Time : 246.393 sec.

Of course Jean Penné can now tell me what I have broken by doing this.

Edit 2: And a Proth test:

Code:

./llrCUDA -d -q9999*2^458051+1   
Starting Proth prime test of 9999*2^458051+1
9999*2^458051+1 is prime!  Time : 313.916 sec.

Jean Penné · 2011-02-04, 19:53

Hi,

Presently, you have a strong interest that set_fftlen() chooses IBDWT FFT!
So, you may remove the fftdwt > fftzpad test... For the second test, I think you still need to verify that bpw goes not too small ; but you may have a minimum value smaller than 5.0. It is necessary to avoid entering an endless loop in which fftdwt goes larger and larger and bpw goes smaller and smaller...
Indeed, bpw must be at least one!
Also, be aware that forcing IBDWT increases the risk of getting a round off error ; so, I suggest you set the ErrorCheck=1 option.
But the very future solution is that modred() and generic_modred() does benefit from CUDA parallelism!

I hope all that may help you!

Jean

ltd · 2011-02-04, 22:07

Trying to get it to compile with VS2008.
So far I get it to compile but when I run it after a while I get an error in
gwpnumi.cu at line 1688:

cutilSafeCall(cudaMemcpy(x,g_x,sizeof(double)*(N),cudaMemcpyDeviceToHost));

I will try to find out what happens tomorrow.

Ken_g6 · 2011-02-04, 22:49

Let me guess, you're running 260.99 or higher drivers? And an "unknown error" is returned? Sounds very reminiscent of this known error on Linux. I think that's the same error that Einstein@Home is waiting on to use less CPU.

I know of two ways to deal with it. One, don't use blocking sync. Or, two, if when it errors out with that particular error you retry the memcpy again and again until it doesn't, that seems to work too.

2011-02-04, 19:53	#97
Jean Penné May 2004 FRANCE 624₁₀ Posts	Forcing IBDWT ? Hi, Presently, you have a strong interest that set_fftlen() chooses IBDWT FFT! So, you may remove the fftdwt > fftzpad test... For the second test, I think you still need to verify that bpw goes not too small ; but you may have a minimum value smaller than 5.0. It is necessary to avoid entering an endless loop in which fftdwt goes larger and larger and bpw goes smaller and smaller... Indeed, bpw must be at least one! Also, be aware that forcing IBDWT increases the risk of getting a round off error ; so, I suggest you set the ErrorCheck=1 option. But the very future solution is that modred() and generic_modred() does benefit from CUDA parallelism! I hope all that may help you! Jean

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
LLRcuda	shanecruise	Riesel Prime Search	8	2014-09-16 02:09
LLRCUDA - getting it to work	diep	GPU Computing	1	2013-10-02 12:12

2011-02-04, 22:07	#98
ltd Apr 2003 2²·193 Posts	Trying to get it to compile with VS2008. So far I get it to compile but when I run it after a while I get an error in gwpnumi.cu at line 1688: cutilSafeCall(cudaMemcpy(x,g_x,sizeof(double)*(N),cudaMemcpyDeviceToHost)); I will try to find out what happens tomorrow.

2011-02-04, 22:49	#99
Ken_g6 Jan 2005 Caught in a sieve 5·79 Posts	Let me guess, you're running 260.99 or higher drivers? And an "unknown error" is returned? Sounds very reminiscent of this known error on Linux. I think that's the same error that Einstein@Home is waiting on to use less CPU. I know of two ways to deal with it. One, don't use blocking sync. Or, two, if when it errors out with that particular error you retry the memcpy again and again until it doesn't, that seems to work too.