[Bug-gsl] gsl_multifit_linear's output differs on 32 bit vs 64 bit linux

[Ahmed Sulaiman (BLOOMBERG/ 731 LEX)]

Hello,

Please refer to {https://github.com/sulaimn2/gsltest/blob/master/gsltest.cpp} 
for code.
Noticed weird output when I ran my library on 64 bit linux in dev, so I tried 
to hunt down where the problem was.
In my library I am using gsl_multifit_linear from libgsl. I have to tried to 
reproduce the problem with a simpler setup as follows.

Setup:
We have an X matrix with 26 coefficients and 151 observations. See x_matrix [] 
in gsltest.cpp
We also have a Y vector with 151 observations. See y_matrix [] in gsltest.cpp
We run gsl_multifit_linear on X and Y and produce a c vector as the output.
In 32 bit, the output was: 

c = [2.70409, 9.5864, -0.670502, 24.2728, 13.9393, 3.23134, 8.22729, 9.4926, 
-0.213974, 6.85192, 0.215701, 7.49826, 7.22713, -3.39073, 13.5852, 17.3663, 
-0.559321, -1.70083, 18.8802, 5.18303, 3.48551, 1.42608, 9.17107, 22.2, 
-5.66091, 2.83766] //formatted for human readability.

In 64 bit, the output was:
c = [2.87233e+13, 2.87233e+13, 2.87233e+13, 2.87233e+13, 2.87233e+13, 
2.87233e+13, 2.87233e+13, 2.87233e+13, 2.87233e+13, 2.87233e+13, 2.87233e+13, 
2.87233e+13, 2.87233e+13, 2.87233e+13, -2.87233e+13, -2.87233e+13, 
-2.87233e+13, -2.87233e+13, -2.87233e+13, -2.87233e+13, -2.87233e+13, 
-2.87233e+13, -2.87233e+13, -2.87233e+13, -2.87233e+13, -2.87233e+13]

gsl_multifit_linear works when I had smaller test cases (with less 
observations) so I tried to hunt down and see if there was a threshold. In 
gsltest.cpp line 171, all outputs matched between 32 bit and 64 bit for all 
number of observations until n=100. The moment n=101, the outputs started being 
all garbage values for 64 bit linux. This number might just be a coincidence 
but I am not sure.

Thanks and regards,
Ahmed


<< Not so young padawan >>