How to handle trivial precision differences with different compiler optimizations

[xuzheng97]

Hello All,

Recently I met some result differences when I tried to use compier options from O0 to O3.
Though theyare trvial and even can be ignored, I still want to know more clear about them.

1. one variable value is 2.468057e-18for O0 and 1.676039e-18for O3.

This variable looks very small.
I tried '-ftz/-zero/-no-bss-init' options with O3 but no any influence.

2. variable F=4867459.8367both for O0 and O3
While F-ksi=-9.31323e-10 for O0 and -1.86265e-09 for O3

All data here are double.
In my memory double stype data has 15-16 valid numbers, then F-ksi e-9 should not make sense at all.


But how to eliminate these differences from compiler site?

Thanks & Best Regards

[Om_S_Intel]

Could you please try the following:

Linux and Mac OS X: -fp-modelprecise

Windows: /fp:precise

If the above does not work then you may try

Linux and Mac OS X: -fp-model source

Windows: /fp:source

[xuzheng97]

Om Sachan,

Thanks, I forget fp-model option :-)
In fact fp-model precise worked and eliminated these difference.

To my unstandings, fp-model precise may decrease performance.
Based my upper decription, could these difference be ignored?

Thanks

[Om_S_Intel]

Yes, the fp-model precise may decrease performance.

The error tolerance depends on the application. The application owner should know better.

[mecej4]

Quoting xuzheng97

1. one variable value is 2.468057e-18for O0 and 1.676039e-18for O3.

This variable looks very small.

"Small" compared to what? 1 ? How about compared to 0 ? The diameter of an electron in light-years ?

2. variable F=4867459.8367both for O0 and O3
While F-ksi=-9.31323e-10 for O0 and -1.86265e-09 for O3

... then F-ksi e-9 should not make sense at all.

We have no idea what "F-ksi" is; therefore, any judgements about its value making sense are premature.

But how to eliminate these differences from compiler site?

You cannot, except in a contrived and pointless way. You would benefit by reading books and articles about what to expect from finite-precision calculations. For starters, Google "perfidious polynomial" and read some of the articles related to that.