Was I affected by the ECM bug?

[patrik]

From whatsnew.txt

3) A bug in the assembly language add and subtract routines was fixed. The
bug could affect ECM on exponents just below
MAXIMUM_EXPONENT_FOR_FFTLEN - FFTLEN/2.

What are the smaller FFT limits of Prime95? The smallest at the benchmark page is 384K for the 6.465M to 7.69M range, but are there smaller ranges as well? I wonder since I have used it for doing stage 1 of ECM for the following exponents:
[code:1]
M739 858 B1=11e6 B2=11e8
M739 2942 B1=11e6 B2=26e9
M743 7500 B1=11e6 B2=6e9
M757 4399 B1=11e6 B2=6e9
M761 5000 B1=11e6 B2=6e9
M761 3700 B1=11e6 B2=6e9
M763 9480 B1=11e6 B2=6e9
M781 9980 B1=11e6 B2=6e9
[/code:1]
and I want to know if I must run any of them again. (Last one still not reported to George W.)

[Prime95]

You are not affected. The bug probably does not kick in until exponents get larger.

You can verify this claim by running an ECM curve in 22.12 with roundoff checking enabled. If it stays under 0.5, then all is OK.

[patrik]

Originally Posted by Prime95

If it stays under 0.5, then all is OK.

For M739 and M743 I get exactly 0.5, so they seem to be affected. So was M731, for which I found that 48 digit factor, so it doesn't seem to affect every curve. I have also just rerun 10 curves of M739 and compared the output (which I use as an input save file to GMP-ECM) and it matches the old output.

Have you any more information about the bug? Like if it only affects stage 2 and that is why it matches? Or if it was introduced recently and I was not using the latest client? Or if it only affects a small fraction of the curves? How can I otherwise get matching results?

I am mainly concerned that the pages that show how much work has been done won't be correct.

[Prime95]

Originally Posted by patrik

For M739 and M743 I get exactly 0.5, so they seem to be affected.

Have you any more information about the bug? Or if it only affects a small fraction of the curves? How can I otherwise get matching results?

The bug can crops up when

floor ((fftlen / 2 + (maxp - p) - 4) / fftlen)

is different than

floor ((fftlen / 4 + (maxp - p)) / fftlen)

where maxp is the maximum exponent that can be run for that fft length

In your case, assuming a P4, then fftlen = 32 and maxp = 743. Both forumlas above generate zero. So 23.4 should work just the same as 23.3. Which means....prime95 should be changed to use a larger fft length for the exponents you are testing. I can believe this for 743 and 739 which are right at the limit, but 731 surprises me. Will the problem affect every curve, I wouldn't think so. I just ran 100 curves on M741 with B1=100000 and never got an error above 0.375

[patrik]

Originally Posted by Prime95

You can verify this claim by running an ECM curve in 22.12 with roundoff checking enabled. If it stays under 0.5, then all is OK.

Originally Posted by Prime95

I just ran 100 curves on M741 with B1=100000 and never got an error above 0.375

Which version should one use to check this? If I use 22.12 I get 0.5, but if I use 23.2 I get different values, max 0.375.

I let the program (v23.4) continue overnight and compared the save file for GMP-ECM with the old ones (which were done with either v22.13 or 23.2) and all 205 curves of M739 matched. This morning I have run 35 curves of M743, all with matching save values. I also ran one curve of M743 using GMP-ECM, and that value matched as well. So perhaps we can assume all work I reported was correct.

Edit: All three computers I use at home for ECM are Pentium 4.

[Prime95]

From whatsnew.txt:

New features in Version 22.13 of prime95.exe
--------------------------------------------

1) A bug in ECM for exponents below 172,700 and near the limit of an FFT size and using SSE2 code was fixed.


So amend my previous instructions. Compare your version 23.4 result to version 22.13