Roundoff error

bcp19 · 2013-02-14, 02:24

Sat down to my computer and see "4 roundoff > .4 of which 2 were repeatable". I have the roundoff reporting turned on, and after this 4th error (which was the 2nd repeatable) it shows 0.000000000 to 0.375000000 (had been .2031250000). Should this exponent be stopped and rerun from the beginning?

LaurV · 2013-02-14, 03:08

If it says something about "using a slower method to test", or "increasing the FFT", then you don't need to repeat anything. Sometime the FFT used is too tough, or too strict (too small). This happens at the "borderlines", when the FFT should be increased, but the smaller one, which is faster, it is still good for the most of the iterations, but not for all, therefore, few iterations will give you a higher error, and they will be repeated using a "slower" method of multiplication. If they turn out right, then the normal FFT style is resumed for the next iterations, and you don't need to repeat all the test. With the older versions of P95 this was happening around 47M expos. I don't know for newer versions, since I upgraded the program, also the range increased. Currently LL around 50M uses a bit higher FFT and gives very small errors, like around 0.1 (I have roundoff turned ON, too).

bcp19 · 2013-02-14, 03:29

This is a 27.9M exp using a 1440k FFT, and yes, it said something about redoing with a slower method.

LaurV · 2013-02-14, 04:12

Quote:

Originally Posted by bcp19

This is a 27.9M exp using a 1440k FFT, and yes, it said something about redoing with a slower method.

Yes, that is at the borderline too. You don't need to repeat the test, the final result will be right. Except in the case he is using the "slower method" at (almost) every iteration, and you still have many iterations to do, in this case it may be/would be faster to repeat the test with a bit larger FFT (you have to calculate by yourself, and specify FFT2= part in the assignment line, or for the newest P95, you can use SoftCrossovers=n and SoftCrossoverAdjust=n, those are very good additions to P95, see undoc.txt).

bcp19 · 2013-02-14, 21:23

The residue successfully completed the double check, so it looks like it was just a borderline FFT. Thanks for the help.

2013-02-14, 02:24	#1
bcp19 Oct 2011 2A7₁₆ Posts	Roundoff error Sat down to my computer and see "4 roundoff > .4 of which 2 were repeatable". I have the roundoff reporting turned on, and after this 4th error (which was the 2nd repeatable) it shows 0.000000000 to 0.375000000 (had been .2031250000). Should this exponent be stopped and rerun from the beginning?

2013-02-14, 03:29	#3
bcp19 Oct 2011 2A7₁₆ Posts	This is a 27.9M exp using a 1440k FFT, and yes, it said something about redoing with a slower method. Last fiddled with by bcp19 on 2013-02-14 at 03:31

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
roundoff = 0.224. Normal? Yes. [solved]	Fred	Software	3	2016-02-05 00:01
Lots of roundoff errors	TheMawn	Software	18	2014-08-16 03:54
Prime95 roundoff errors	pjaj	Software	18	2011-07-20 03:04
Roundoff Error Penalty	nevarcds	Software	5	2004-08-28 14:29
Roundoff Error Message	Teseo77Madrid	Hardware	21	2004-06-02 14:59

2013-02-14, 03:08	#2
LaurV Romulan Interpreter Jun 2011 Thailand 7·1,373 Posts	If it says something about "using a slower method to test", or "increasing the FFT", then you don't need to repeat anything. Sometime the FFT used is too tough, or too strict (too small). This happens at the "borderlines", when the FFT should be increased, but the smaller one, which is faster, it is still good for the most of the iterations, but not for all, therefore, few iterations will give you a higher error, and they will be repeated using a "slower" method of multiplication. If they turn out right, then the normal FFT style is resumed for the next iterations, and you don't need to repeat all the test. With the older versions of P95 this was happening around 47M expos. I don't know for newer versions, since I upgraded the program, also the range increased. Currently LL around 50M uses a bit higher FFT and gives very small errors, like around 0.1 (I have roundoff turned ON, too).

2013-02-14, 21:23	#5
bcp19 Oct 2011 7×97 Posts	The residue successfully completed the double check, so it looks like it was just a borderline FFT. Thanks for the help.