I am back.
 

Hi,

I am now back from Uzbekistan, and apparently with some work to to on LLR!
Many thanks to all people wo worked on these problems, your help will be essential!
Best Regards,
Jean

[QUOTE=Jean Penné;401209]I am now back from Uzbekistan, and apparently with some work to to on LLR![/QUOTE]

Welcome back :smile: Here is a quick summary of the current status:

1) There was a roundoff error "bug" regarding Pentium4 type-0 FFT post #21-#32 and George recommended a fix in [URL="http://www.mersenneforum.org/showpost.php?p=400591&postcount=32"]post #32[/URL]

2) George found a bug in AVX FFT in his recent gwnum fix, so a new gwnum will be available soon

3) There are currently 5 known Lucas errors and 1 known Frobenius error found by Batalov which may or may not be fixed by the new upcoming gwnum version: [URL="http://www.mersenneforum.org/showpost.php?p=400765&postcount=38"]post #38[/URL] + [URL="http://www.mersenneforum.org/showpost.php?p=400810&postcount=42"]post #42[/URL]

[QUOTE=Batalov;400765]Here is an interesting composite strong-Fermat PSP; might be good for tests of different implementations:



There is nothing trivially* wrong with this number. 40216 = 2^3 * 11 * 457, no threes in it.

Actually, with PFGW, it is a PRP in many bases, as well as by N+1/N-1 tests. Looks like another bug.

__________________

*[SIZE="1"]For example, all (19683^p-1)/19682 are strong-Fermat PSP, but it is hardly surprising because 19683 is simply a power of 3, and these are classic false hitters.)[/SIZE][/QUOTE]

Here is my test on P4 and Windows 32bit (indeed, much slower as yours) :

Starting probable prime test of (40216^11719-1)/40215
Using Pentium4 type-1 FFT length 24K, Pass1=96, Pass2=256, a = 3
(40216^11719-1)/40215 is base 3-Strong Fermat PRP! (53955 decimal digits) Time : 271.308 sec.
Starting Lucas sequence
Using Pentium4 type-1 FFT length 24K, Pass1=96, Pass2=256, P = 5, Q = 2
(40216^11719-1)/40215 is strong-Fermat and Lucas PRP, Starting Frobenius test sequence
Using Pentium4 type-1 FFT length 24K, Pass1=96, Pass2=256, Q = 2
(40216^11719-1)/40215 is strong-Fermat, Lucas and Frobenius PRP! (P = 5, Q = 2, D = 17) Time : 1799.318 sec.

So, this problem is also machine dependant!
Jean

Of course. The problem is in the AVX-dependent code. On the non-AVX machine (or with AVX disabled) one will get the correct answer but the run time will be 50-80% more. This is similar to how the 2[SUP]10[/SUP]*3[SUP]n[/SUP]+1 bug behaved.

We need [STRIKE] to wait for[/STRIKE] George's release of GWNUM lib ver.28.7.
Actually it [B]is[/B] posted: [URL]ftp://mersenne.org/gimps/[/URL]
[CODE][URL="ftp://mersenne.org/gimps/p95v287.source.zip"]p95v287.source.zip[/URL] 43738 KB 4/25/2015 11:45:00 PM[/CODE]

[QUOTE=Batalov;401312]Of course. The problem is in the AVX-dependent code. On the non-AVX machine (or with AVX disabled) one will get the correct answer but the run time will be 50-80% more. This is similar to how the 2[SUP]10[/SUP]*3[SUP]n[/SUP]+1 bug behaved.

We need [STRIKE] to wait for[/STRIKE] George's release of GWNUM lib ver.28.7.
Actually it [B]is[/B] posted: [URL]ftp://mersenne.org/gimps/[/URL]
[CODE][URL="ftp://mersenne.org/gimps/p95v287.source.zip"]p95v287.source.zip[/URL] 43738 KB 4/25/2015 11:45:00 PM[/CODE][/QUOTE]

Hi Serge,

I have completed the tests of your five problematic GRU candidates, and cannot find any composite result while using my 32bit and 64bit machines, so, you are right, it is a second AVX bug ; but, also we have still to wait for the fix, because George said he only fixed a tiny cosmetic title bug, in the new released p95v287.source...
Regards,
Jean

[QUOTE=ATH;401211]Welcome back :smile: Here is a quick summary of the current status:

1) There was a roundoff error "bug" regarding Pentium4 type-0 FFT post #21-#32 and George recommended a fix in [URL="http://www.mersenneforum.org/showpost.php?p=400591&postcount=32"]post #32[/URL]

2) George found a bug in AVX FFT in his recent gwnum fix, so a new gwnum will be available soon

3) There are currently 5 known Lucas errors and 1 known Frobenius error found by Batalov which may or may not be fixed by the new upcoming gwnum version: [URL="http://www.mersenneforum.org/showpost.php?p=400765&postcount=38"]post #38[/URL] + [URL="http://www.mersenneforum.org/showpost.php?p=400810&postcount=42"]post #42[/URL][/QUOTE]

1) In the future Llr.c, I force "echk" to one if gwnear_fft_limit (gwdata, 0.5) returns true.
2) Waiting for the new fix.
3) I verified all these candidates on my 32bit and 64bit (non AVX) machines, and found no composite PSPs.
4) Thanks to axn, I fixed the trivial bug in process_num function.
Thank you for your help!
Jean

[QUOTE=Jean Penné;401315] George said he only fixed a tiny cosmetic title bug, in the new released p95v287.source...
[/QUOTE]

The released 28.7 source has the AVX fix you need.

[QUOTE=Prime95;401323]The released 28.7 source has the AVX fix you need.[/QUOTE]

George, would you excuse me for this misunderstanding, and many thanks for informing me of that!
I shall download the new source now.
Jean

I built the 3.8.14 + gwnum.28.7 linux64 binary here, and [STRIKE]all five[/STRIKE] three small GRUs now pass all three PRP tests without a hitch.
No, wait; two still don't pass:
[CODE](50322^11317-1)/50321 is base 3-Strong Fermat PRP! (53206 decimal digits) Time : 27.603 sec.
(50322^11317-1)/50321 is strong-Fermat, Lucas and Frobenius PRP! (P = 5, Q = 2, D = 17) Time : 115.507 sec.
(48591^11329-1)/48590 is base 3-Strong Fermat PRP! (53090 decimal digits) Time : 26.935 sec.
(48591^11329-1)/48590 is strong-Fermat, Lucas and Frobenius PRP! (P = 7, Q = 4, D = 33) Time : 188.224 sec.
(57042^11491-1)/57041 is base 3-Strong Fermat PRP! (54649 decimal digits) Time : 33.937 sec.
[COLOR="DarkRed"](57042^11491-1)/57041 is strong-Fermat PSP, but composite!! (P = 5, Q = 2), Lucas RES64: 8D2F6A09C977A0CD Time : 104.920 sec.
(57352^11491-1)/57351 is base 3-Strong Fermat PRP! (54676 decimal digits) Time : 35.093 sec.
(57352^11491-1)/57351 is strong-Fermat PSP, but composite!! (P = 5, Q = 2), Lucas RES64: 5295A221C17D767A Time : 137.133 sec.[/COLOR]
(40216^11719-1)/40215 is base 3-Strong Fermat PRP! (53955 decimal digits) Time : 27.439 sec.
(40216^11719-1)/40215 is strong-Fermat, Lucas and Frobenius PRP! (P = 5, Q = 2, D = 17) Time : 143.414 sec.

All use the same FFT routine? no!
Using AVX FFT length 24K, Pass1=128, Pass2=192, a = 3
Using AVX FFT length 24K, Pass1=128, Pass2=192, P = 5, Q = 2
Using AVX FFT length 24K, Pass1=128, Pass2=192, Q = 2
Using AVX FFT length 24K, Pass1=128, Pass2=192, a = 3
Using AVX FFT length 24K, Pass1=128, Pass2=192, P = 7, Q = 4
Using AVX FFT length 24K, Pass1=128, Pass2=192, Q = 4
Using AVX FFT length 24K, Pass1=128, Pass2=192, a = 3
Using AVX FFT length 24K, Pass1=128, Pass2=192, P = 5, Q = 2
Using AVX FFT length 24K, Pass1=128, Pass2=192, Q = 2
[COLOR="DarkRed"]Using AVX FFT length 25K, Pass1=320, Pass2=80, a = 3
Using AVX FFT length 25K, Pass1=320, Pass2=80, P = 5, Q = 2
Using AVX FFT length 25K, Pass1=320, Pass2=80, a = 3
Using AVX FFT length 25K, Pass1=320, Pass2=80, P = 5, Q = 2[/COLOR][/CODE]
Now, I will recheck some old Generalized Woodall's for the extended test (in a forced PRPcheck mode, and only then, if anything is found, - with LLR test).

[QUOTE=Batalov;401372]I built the 3.8.14 + gwnum.28.7 linux64 binary here, and [STRIKE]all five[/STRIKE] three small GRUs now pass all three PRP tests without a hitch.
No, wait; two still don't pass:
[CODE](50322^11317-1)/50321 is base 3-Strong Fermat PRP! (53206 decimal digits) Time : 27.603 sec.
(50322^11317-1)/50321 is strong-Fermat, Lucas and Frobenius PRP! (P = 5, Q = 2, D = 17) Time : 115.507 sec.
(48591^11329-1)/48590 is base 3-Strong Fermat PRP! (53090 decimal digits) Time : 26.935 sec.
(48591^11329-1)/48590 is strong-Fermat, Lucas and Frobenius PRP! (P = 7, Q = 4, D = 33) Time : 188.224 sec.
(57042^11491-1)/57041 is base 3-Strong Fermat PRP! (54649 decimal digits) Time : 33.937 sec.
[COLOR="DarkRed"](57042^11491-1)/57041 is strong-Fermat PSP, but composite!! (P = 5, Q = 2), Lucas RES64: 8D2F6A09C977A0CD Time : 104.920 sec.
(57352^11491-1)/57351 is base 3-Strong Fermat PRP! (54676 decimal digits) Time : 35.093 sec.
(57352^11491-1)/57351 is strong-Fermat PSP, but composite!! (P = 5, Q = 2), Lucas RES64: 5295A221C17D767A Time : 137.133 sec.[/COLOR]
(40216^11719-1)/40215 is base 3-Strong Fermat PRP! (53955 decimal digits) Time : 27.439 sec.
(40216^11719-1)/40215 is strong-Fermat, Lucas and Frobenius PRP! (P = 5, Q = 2, D = 17) Time : 143.414 sec.

All use the same FFT routine? no!
Using AVX FFT length 24K, Pass1=128, Pass2=192, a = 3
Using AVX FFT length 24K, Pass1=128, Pass2=192, P = 5, Q = 2
Using AVX FFT length 24K, Pass1=128, Pass2=192, Q = 2
Using AVX FFT length 24K, Pass1=128, Pass2=192, a = 3
Using AVX FFT length 24K, Pass1=128, Pass2=192, P = 7, Q = 4
Using AVX FFT length 24K, Pass1=128, Pass2=192, Q = 4
Using AVX FFT length 24K, Pass1=128, Pass2=192, a = 3
Using AVX FFT length 24K, Pass1=128, Pass2=192, P = 5, Q = 2
Using AVX FFT length 24K, Pass1=128, Pass2=192, Q = 2
[COLOR="DarkRed"]Using AVX FFT length 25K, Pass1=320, Pass2=80, a = 3
Using AVX FFT length 25K, Pass1=320, Pass2=80, P = 5, Q = 2
Using AVX FFT length 25K, Pass1=320, Pass2=80, a = 3
Using AVX FFT length 25K, Pass1=320, Pass2=80, P = 5, Q = 2[/COLOR][/CODE]
Now, I will recheck some old Generalized Woodall's for the extended test (in a forced PRPcheck mode, and only then, if anything is found, - with LLR test).[/QUOTE]

Thank you for all this work, Serge!
I hope that gwnear_fft_limit (gwdata, 0.5) usage will help to resolve the remaining issues...
Regards,
Jean

Please can you make that LLR remove processed candidates like Prime95 do?