mtsieve

[Jean Penné]

Originally Posted by pepi37

Nope, just for linux , and as I know it is not fast enough ( in my case)

Nevertheless, I need to know if the static binary of llrCUDA works for you, even if it is not fast enough...

Thank you by advance,

Jean

[pepi37]

Originally Posted by Jean Penné

Nevertheless, I need to know if the static binary of llrCUDA works for you, even if it is not fast enough...

Thank you by advance,

Jean

Yes it works :) But one instance "eat one CPU core"
Using trick with libsleep.I can reduce it to 50% of one CPU core. Speed is same as Ryzen7 3700x per core: since both need around 17 minutes for test of 535000 digits candidate

root@OMICRON:~/LLR# ./sllrCUDA -d -q"4569*2^1778899+1"
Starting Proth prime test of 4569*2^1778899+1
Using complex irrational base DWT, FFT length = 262144, a = 5
^Ceration: 160000 / 1778910 [8.49%], ms/iter: 0.596, ETA: 00:16:04
Caught signal. Terminating.
Stopping Proth prime test of 4569*2^1778899+1 at iteration 164342 [9.23%]


root@OMICRON:~/LLR# LD_PRELOAD="/usr/local/lib/libsleep.so" ./sllrCUDA -d -q"4569*2^1778899+1"
libsleep: Sleep time: 50usec
Resuming Proth prime test of 4569*2^1778899+1 at bit 164343 [9.23%]
Using complex irrational base DWT, FFT length = 262144, a = 5
^Ceration: 310000 / 1778910 [16.93%], ms/iter: 0.593, ETA: 00:14:30
Caught signal. Terminating.
Stopping Proth prime test of 4569*2^1778899+1 at iteration 317616 [17.85%]

[Citrix]

I am getting the following error. What settings do I need to change?

srsieve2cl.exe -i sr_2.abcd -W4  -p 10000000000000 -P 11000000000000  -Ofactors.txt -osr_2_new.abcd  -G12 -M100000 -l1000

srsieve2cl v1.6.5, a program to find factors of k*b^n+c numbers for fixed b and variable k and n
Sieving with multi-sequence c=1 logic for p >= 10000000000000
BASE_MULTIPLE = 2, POWER_RESIDUE_LCM = 720, LIMIT_BASE = 720
Assertion failed: m <= HASH_MAX_ELTS, file sierpinski_riesel/AbstractSequenceHelper.cpp, line 272

[rogue]

Originally Posted by Citrix

I am getting the following error. What settings do I need to change?

srsieve2cl.exe -i sr_2.abcd -W4  -p 10000000000000 -P 11000000000000  -Ofactors.txt -osr_2_new.abcd  -G12 -M100000 -l1000

srsieve2cl v1.6.5, a program to find factors of k*b^n+c numbers for fixed b and variable k and n
Sieving with multi-sequence c=1 logic for p >= 10000000000000
BASE_MULTIPLE = 2, POWER_RESIDUE_LCM = 720, LIMIT_BASE = 720
Assertion failed: m <= HASH_MAX_ELTS, file sierpinski_riesel/AbstractSequenceHelper.cpp, line 272

I ran into this in the past week so I have a solution for it. I posted an experimental build over at sourceforge that should address this.

[Citrix]

I get with new

srsieve2cl.exe -i sr_2.abcd -W2  -p 10000000000000 -P 11000000000000  -Ofactors.txt -osr_2_new.abcd   -M1000 -l10000  -w1000 -G12
srsieve2cl v1.6.5, a program to find factors of k*b^n+c numbers for fixed b and variable k and n
Sieving with multi-sequence c=1 logic for p >= 10000000000000
BASE_MULTIPLE = 2, POWER_RESIDUE_LCM = 720, LIMIT_BASE = 720
Split 204 base 2 sequences into 9182 base 2^720 sequences.
Legendre summary:  Approximately 4752 B needed for Legendre tables
       204 total sequences
       204 are eligible for Legendre tables
         0 are not eligible for Legendre tables
       204 have Legendre tables in memory
         0 cannot have Legendre tables in memory
         0 have Legendre tables loaded from files
       204 required building of the Legendre tables
17625600 bytes used for congruent subseq indices
1360000 bytes used for congruent subseqs
Fatal Error:  Must use generic worker if using GPU with multiple sequences by specifying -l0

With generic code

srsieve2cl.exe -i sr_2.abcd -W2  -p 10000000000000 -P 11000000000000  -Ofactors.txt -osr_2_new.abcd   -M1000   -w1000 -G6
srsieve2cl v1.6.5, a program to find factors of k*b^n+c numbers for fixed b and variable k and n
Must use generic sieving logic because -l was not specified for mutiple sequences
Sieving with generic logic for p >= 10000000000000
Split 204 base 2 sequences into 20555 base 2^2880 sequences.
bestQ = 2880 yields bs = 6077, gs = 1, sieveLow = 868, sieveRange = 6077
bestQ = 2880 yields bs = 6077, gs = 1, sieveLow = 868, sieveRange = 6077
GPU primes per worker is 57344
Sieve started: 1e13 < p < 11e12 with 134418 terms (2500875 < n < 20000000, k*2^n-1) (expecting 427 factors)
Increasing worksize to 16000 since each chunk is tested in less than a second

OpenCL Error: Out of host memory
       in call to clEnqueueNDRangeOpenCLKernel
       kernelName: generic_kernel  globalworksize 57344  localworksize 256

[rogue]

Originally Posted by Citrix

With generic code

srsieve2cl.exe -i sr_2.abcd -W2  -p 10000000000000 -P 11000000000000  -Ofactors.txt -osr_2_new.abcd   -M1000   -w1000 -G6
srsieve2cl v1.6.5, a program to find factors of k*b^n+c numbers for fixed b and variable k and n
Must use generic sieving logic because -l was not specified for mutiple sequences
Sieving with generic logic for p >= 10000000000000
Split 204 base 2 sequences into 20555 base 2^2880 sequences.
bestQ = 2880 yields bs = 6077, gs = 1, sieveLow = 868, sieveRange = 6077
bestQ = 2880 yields bs = 6077, gs = 1, sieveLow = 868, sieveRange = 6077
GPU primes per worker is 57344
Sieve started: 1e13 < p < 11e12 with 134418 terms (2500875 < n < 20000000, k*2^n-1) (expecting 427 factors)
Increasing worksize to 16000 since each chunk is tested in less than a second

OpenCL Error: Out of host memory
       in call to clEnqueueNDRangeOpenCLKernel
       kernelName: generic_kernel  globalworksize 57344  localworksize 256

You should use -g to increase GPU primes per worker as opposed to the number of GPU threads. The framework, at this time, does not support one executable running concurrently on multiple GPUs.

Using -G impacts GPU memory usage, but with that many subsequences I suggest that you use -b (a value less than 1.0) to reduce the size of the hash table that the GPU will use. You might also want to use -K to split the sequences across multiple chunks. This will require some trial and error on your part. There is no way (that I am aware of) to compute the memory required for a kernel so the code cannot "auto-tune" these parameters.

You cannot use -l > 0 with the GPU when you have multiple sequences. srsieve2cl does not support it at this time.

I also do not recommend mixing -W and -G. The factor rate calculation does not work correctly when using both CPU and GPU workers.

You can use -p10e12 -P11e12 if that is easier to read.

[storm5510]

@rogue

Q.: Does srsieve2cl generate an exit code when it finishes? Running small sieves from a batch sometimes would fail because I had the -M set too low. It was at 3,500. Now, it is 10,000. It varied based on what the k value was. Some k's caused problems and others did not. All used the same values for -n, -N, and -P.

[rogue]

Originally Posted by storm5510

@rogue

Q.: Does srsieve2cl generate an exit code when it finishes? Running small sieves from a batch sometimes would fail because I had the -M set too low. It was at 3,500. Now, it is 10,000. It varied based on what the k value was. Some k's caused problems and others did not. All used the same values for -n, -N, and -P.

For normal completion it will output the number of terms written to the output file and the time it took to run.

SEGFAULTs will just give you the command prompt without any of that. If that happens let me know.

[storm5510]

Originally Posted by rogue

For normal completion it will output the number of terms written to the output file and the time it took to run.

SEGFAULTs will just give you the command prompt without any of that. If that happens let me know.

Forgive me, but I didn't specify it correctly. An error code?

For a normal program run and exit, an error code of zero is expected. If there is an error, a non-zero code is returned.

Originally Posted by Jean Penné

Nevertheless, I need to know if the static binary of llrCUDA works for you, even if it is not fast enough...

Thank you by advance,

Jean

Off-topic: I am running it as a test. According to nvidia-smi, it is using about 30% of the GPU's capability. I am running "1955*2^n+1" for the test. The k is my birth year. The n's are around 102K presently. Despite not being all that fast, it is quite stable in my case. Ubuntu 20.04.4 LTS using a GTX 1080. The iteration time holds steady at 0.14 seconds. The overall time is increasing gradually.

[rogue]

Originally Posted by storm5510

Forgive me, but I didn't specify it correctly. An error code?

For a normal program run and exit, an error code of zero is expected. If there is an error, a non-zero code is returned.

It will be zero upon successful completion. A FatalError (caught and output to the console) is -1. I'm not certain what assert() with exit with.

I do not understand why you care. The error code is not output to the console.

[Citrix]

@Rogue

I can get the program to work but it is extremely slow without the Legendre tables.

Couple of other questions/thoughts

1. I get the following error with the CPU code as well (srsieve2). Can you release a fix.

Assertion failed: m <= HASH_MAX_ELTS, file sierpinski_riesel/AbstractSequenceHelper.cpp, line 272

2. For BASE_MULTIPLE there is a limit of 60 ... can this be increased to 256 or higher.

3. Possible bug:- The GPU code seems to crash if the n range is large (~15M); seems to produce false factors if n range is large and LIMIT_BASE is huge.

4. For what type of sequences is it best to use GPU and for which ones should you stick to CPU.

Thanks