Hi! On 09/04/2012 05:58 PM, Javier Ros wrote:

I've been exporting C code from GiNaC using "csrc_double" output format during a long time.

Now we are dealing with rather long real expressions, that involve only +-*\ and pow. In a benchmark and for comparison purposes we've end up comparing the results of evaluate the same function through "dflt" (used to generate output for MATLAB) and csrc_double (used to generate same code for C).

Numerical discrepancies where found and traced back to to small symbolic differences in the output obtained through both formats.

To avoid the problem and due to the symbolical simplicity of our expressions (only +-*\ and pow) we've developed our own -very simple- exporting format for C. Basically we derive our format from "dflt" and change the way in which natural numbers and powers are written (we append a .0 to natural numbers) and we change a^b with pow(a,b).

Right now we are obtaining a correct behaviour the export based on our new format, coherent with "dflt" numerical results (in a particular case), and obtaining different results with csrc_double (I think too with csrc alone).

So I think I'm facing a bug in csrc output format.

I wonder if anybody wants me to report further on this (I don't feel confident as to dig into GiNaC's implementation of crsc). If so please let me know the way in which I can report back the error (or what I think is an error) with the required detail.

You don't have to dig into GiNaC's implementation in order to find the reason why your compiler produces different numerical results. Most likely, the different results are due to cancellations in an ill-conditioned expression. On the other hand, if you can prove that GiNaC does produce wrong code, please be so kind and show us the expression where it occurs. -richy. -- Richard B. Kreckel <http://www.ginac.de/~kreckel/>