srsieve/sr2sieve enhancements

[yoyo]

If I run sr1sieve version 1.4.5
sr1sieve_1.4.5-x86_64-linux -P 15257050000000000 -i in -o out -f factor1.txt)

I get an out file with this format:

15257500000000000:M:1:10:258
92 1400033
92 1400035
92 1400045
92 1400068
92 1400075
...

If I run version 1.4.7
sr1sieve_1.4.7_aarch64 -P 15257050000000000 -i in -o out -f factor1.txt
I get an out file with this format:

ABCD 92*10^$a-1 [1400033] // Sieved to 15257500000000000
2
10
23
...

How to get in 1.4.7 the same output format as in 1.4.5?
Background is that I run a task on different hosts and compare the results afterwards to ensure valid results.

[LaurV]

You may have some options (was it -w or -G? I don't recall) to convert file formats, or you also can call srfile after the run, to convert to pfgw format, or newpgen format, or any format.

edit: I only have 1.4.5 here, you can call "srfile -g second_file" to get the "first_file" (i.e. convert from abc format to newpgen format). For details, see "srfile -h" and "sr1sieve -h".

[bur]

Does sr1sieve support Riesel and Proth numbers in the same sieve? I combined two seperate sieves using srfile and sr1sieve threw the error

Cannot support multiple sequences in input file `sr_2.abcd'.

I guess that means, no?


I wanted to use it since it was the fastest among sr1sieve, sr2sieve and srsieve2 for my fixed k.


Does it even make sense efficiency-wise to combine sieve files like that or would it be the same if I ran them seperately?

[Happy5214]

sr1sieve doesn't support multiple sequences period. It's optimized for a single sequence (i.e. k). The traditional rule of thumb is to use sr2sieve for 3 or more k's, sr1sieve for 1 or 2 (two separate instances in the latter case), and srsieve/srsieve2 just to start (I go to 1e9 because my workflow is highly automated by e9, but that's probably not optimal), but I think srsieve2cl (GPU, not srsieve2 for CPU) is faster than even sr1sieve for a single k.

[rogue]

Originally Posted by bur

Does sr1sieve support Riesel and Proth numbers in the same sieve? I combined two seperate sieves using srfile and sr1sieve threw the error

Cannot support multiple sequences in input file `sr_2.abcd'.

I guess that means, no?


I wanted to use it since it was the fastest among sr1sieve, sr2sieve and srsieve2 for my fixed k.


Does it even make sense efficiency-wise to combine sieve files like that or would it be the same if I ran them seperately?

sr1sieve cannot be modified to sieve both +1 and -1 for the same k/b concurrently. The sign is important WRT how the Legendre tables are built and used.

[bur]

Generally it's adviced to sieve the whole set of candidates in one sieve as computation time increases only with the sqrt of the number of candidates. But does this rule still apply in the case of +1 and -1 combined?

If not, then sr1sieve run on two seperate sieves would be more efficient.

I cannot test it right now as I don't have access to the computer.

but I think srsieve2cl (GPU, not srsieve2 for CPU) is faster than even sr1sieve for a single k.

Do you know if that's the software PG uses for their Proth Prime Sieves? I tried that on a GTX 730 and it was very slow, likely due to the old card, but it's the only one I have available for sieving.

[Gelly]

../srsieve2cl -P 5e13 -i candy2.cand -o candy.cand -f A -W 4 -G 1
srsieve2cl v1.5.2, a program to find factors of k*b^n+c numbers for fixed b and variable k and n
Sieving one sequence where abs(c) = 1 for p >= 15175098061843
Split 1 base 13 sequence into 19 base 13^120 sequences.
551218 bytes used for congruence tables
351 bytes used for Legendre tables
*** buffer overflow detected ***: terminated
Aborted (core dumped)

srsieve2cl is crying. Perhaps I am missing the exact methodology to run the siever with OpenCL, but it seems like any attempt to introduce a GPU worker with -G 1 causes a buffer overflow. GPU is an RTX 2070.

[rogue]

Originally Posted by Gelly

../srsieve2cl -P 5e13 -i candy2.cand -o candy.cand -f A -W 4 -G 1
srsieve2cl v1.5.2, a program to find factors of k*b^n+c numbers for fixed b and variable k and n
Sieving one sequence where abs(c) = 1 for p >= 15175098061843
Split 1 base 13 sequence into 19 base 13^120 sequences.
551218 bytes used for congruence tables
351 bytes used for Legendre tables
*** buffer overflow detected ***: terminated
Aborted (core dumped)

srsieve2cl is crying. Perhaps I am missing the exact methodology to run the siever with OpenCL, but it seems like any attempt to introduce a GPU worker with -G 1 causes a buffer overflow. GPU is an RTX 2070.

It hasn't started any workers yet. I'm guessing that srsieve2 will fail at the same point. Please post or e-mail your input file and I'll take a look.

[Gelly]

Here's the list of candidates. I have been running it under srsieve2 (and, indeed, even with srsieve2cl with no GPU workers) with no issues, so I do think it's a GPU error on my end.

[rogue]

I did find one bug in CisOneWithOneSequenceHelper.cpp and committed a fix, but that bug prevented compilation on OS X. The sequence in this file works on two different GPUs that I've tried, one AMD and one NVIDIA.

If you get the updated file and rebuild (start with make clean), are you able to do any debugging?

Have you run any of the other GPU-enabled sievers built on the framework? Do any of those work?

[Gelly]

Originally Posted by rogue

I did find one bug in CisOneWithOneSequenceHelper.cpp and committed a fix, but that bug prevented compilation on OS X. The sequence in this file works on two different GPUs that I've tried, one AMD and one NVIDIA.

If you get the updated file and rebuild (start with make clean), are you able to do any debugging?

Have you run any of the other GPU-enabled sievers built on the framework? Do any of those work?

Did what you suggested. Still no dice for srsieve2cl. It seems like most of the other cl sievers work just fine with OpenCL, though I did not make a decimal file for pixsievecl to test with.

Here's another example where I start sieving on a random sequence, leading me to believe it's not a sequence problem, neither:

../srsieve2cl -G 1 -P 1e13 -n 1000 -N 1500 -s "2564*2345^n+1"
srsieve2cl v1.5.2, a program to find factors of k*b^n+c numbers for fixed b and variable k and n
Sieving with generic logic for p >= 3
Creating CPU worker to use until p >= 1000000
GPU primes per worker is 56320
Sieve started: 3 < p < 1e13 with 501 terms (1000 < n < 1500, k*2345^n+c) (expecting 482 factors)
Sieving one sequence where abs(c) = 1 for p >= 2564
Split 1 base 2345 sequence into 11 base 2345^30 sequences.
552614 bytes used for congruence tables
376119 bytes used for Legendre tables
Creating CPU worker to use until p >= 1000000
*** buffer overflow detected ***: terminated
Aborted (core dumped)