How to calculate FFT lengths of candidates

pepi37 · 2014-04-13, 16:27

I get this data from first megabit drive

554 candidates (144k: 49.8%, 160k: 50.2%)

Can someone tell me how to calculate K weight of candidate? What formula is used ( I sow that AVX uses some different approach)

Thomas11 · 2014-04-14, 08:26

In principle the procedure is quite simple:
For each candidate in the input file the FFT length is obtained and then the percentages are computed.

In practise I'm using a small tool (see attachment) based on my "llrtools".

Thomas11 · 2014-04-14, 13:27

The attached file contains a binary for Windows machines.

Note that this is a command line tool. You'll need to run it from your DOS prompt:

Code:

fft_percentage.exe input.txt

Also note: The maxlen.txt file is for non-AVX machines.

Thomas11 · 2014-04-17, 11:18

Due to the recent interest (and also since the original LLR tools didn't contain the parameters for AVX machines), I'm posting here a Javascript version of the FFT length calculator - just a proof of concept, so don't expect any sophisticated piece of software...

It's a simple HTML file which you should be able to open in your favourite (Javascript enabled) web browser.

Usage is quite simple:
Enter the values for k, nmin and nmax, select the proper cpu type (avx, sse2, or x87) and hit the "Get FFT lengths" button.
Note that zero-padded FFTs use a different algorithm which is not yet implemented. Thus, for K>1000000 the points where the FFT lengths are changing are slightly different.

Feel free to further modify the code.

pepi37 · 2014-04-17, 14:24

I dont wont to be rude but since I ask this for candidates on base 10 this is pretty unusable for me :(

Batalov · 2014-04-17, 18:03

Quote:

Originally Posted by pepi37

I get this data from first megabit drive

554 candidates (144k: 49.8%, 160k: 50.2%)

Can someone tell me how to calculate K weight of candidate? What formula is used ( I sow that AVX uses some different approach)

Quote:

Originally Posted by pepi37

I dont wont to be rude but since I ask this for candidates on base 10 this is pretty unusable for me :(

Eh?

Must be the language barrier.

Anyway, just run this

Code:

# ./sllr64 -d -q"2*10^1059002-1"
Base factorized as : 2*5
Base prime factor(s) taken : 5
Starting N+1 prime test of 2*10^1059002-1
Using AMD K8 FFT length 224K, Pass1=896, Pass2=256, a = 2

...and kill it (Ctrl-C)! Problem solved.
(In fact, if you set up llr.ini properly and wait for a few seconds, you will also get the time per iteration, ... and then you kill it.)

On a different CPU, the FFT kernel and size may be different.

Repeat for random samples from your future work files. Plot. Scratch your forehead. Analyze. Plan ahead. Keep it simple.

pepi37 · 2014-04-17, 19:05

Batalov, I truly love yours answers, and in fact that is way I found FFT length , start LLR , look at ,and stop it. But I was thinking if there is another ( faster way)....

Batalov · 2014-04-17, 20:50

You can of course,
- understand the LLR source (and GWNUM library that it calls)
- get the appropriate parts of the code and
- repackage them in a new program or a script that will "predict" what particular FFT size will be used.

You will have to update this accessory program every time the LLR+GWNUM changes, when there's a relevant change.

There is no magic single formula! - there are decision paths (if-else-if-else-if-else...) that the GWNUM library uses when it is called to initialize the FFT control structures for a number. Find and read the PrimeGrid topic about that: they initially considered this to be too tedious, but then surprisingly this is what they exactly did: pre-run each number for a split second to know exactly what FFT size will be used and based their credit system on that knowledge.

Thomas11 · 2014-04-17, 20:51

There is an undocumented feature in LLR. Just add the following line to your llr.ini file:

Code:

SetupOnly=1

Then you will no longer need to kill it...

However, for some forms the keyword is just ignored.

2014-04-13, 16:27	#1
pepi37 Dec 2011 After 1.58M nines:) 2⁴·3·37 Posts	How to calculate FFT lengths of candidates I get this data from first megabit drive 554 candidates (144k: 49.8%, 160k: 50.2%) Can someone tell me how to calculate K weight of candidate? What formula is used ( I sow that AVX uses some different approach)

2014-04-17, 20:51	#9
Thomas11 Feb 2003 780₁₆ Posts	There is an undocumented feature in LLR. Just add the following line to your llr.ini file: Code: SetupOnly=1 Then you will no longer need to kill it... However, for some forms the keyword is just ignored.

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Area of a Triangle as a Function of the Lengths of its 3 Sides	a1call	Puzzles	17	2018-02-23 16:28
Lucas–Lehmer sequences and Heronian triangles with consecutive side lengths	a nicol	Math	1	2016-11-08 05:32
List of FFT Lengths (27.7 b2)?	Belial88	Software	1	2012-11-12 16:37
new candidates for M...46 and M48	cochet	Miscellaneous Math	4	2008-10-24 14:33
Hardware Failures at different FFT lengths?	CADavis	Hardware	8	2005-11-29 10:29

2014-04-17, 14:24	#5
pepi37 Dec 2011 After 1.58M nines:) 2⁴×3×37 Posts	I dont wont to be rude but since I ask this for candidates on base 10 this is pretty unusable for me :(

2014-04-17, 19:05	#7
pepi37 Dec 2011 After 1.58M nines:) 2⁴·3·37 Posts	Batalov, I truly love yours answers, and in fact that is way I found FFT length , start LLR , look at ,and stop it. But I was thinking if there is another ( faster way)....

2014-04-17, 20:50	#8
Batalov "Serge" Mar 2008 San Diego, Calif. 101×103 Posts	You can of course, - understand the LLR source (and GWNUM library that it calls) - get the appropriate parts of the code and - repackage them in a new program or a script that will "predict" what particular FFT size will be used. You will have to update this accessory program every time the LLR+GWNUM changes, when there's a relevant change. There is no magic single formula! - there are decision paths (if-else-if-else-if-else...) that the GWNUM library uses when it is called to initialize the FFT control structures for a number. Find and read the PrimeGrid topic about that: they initially considered this to be too tedious, but then surprisingly this is what they exactly did: pre-run each number for a split second to know exactly what FFT size will be used and based their credit system on that knowledge.