[Numpy-discussion] Non-deterministic floating point behavior in numpy 1.5.1 ?

[Grégory Guyomarc'h]

Hello,

I have noticed strange non-deterministic behaviours with numpy 1.5.1 when using 
floating point arrays. The following script were run on 6 different machines, 
all Intel Core i7:
- 3 of them running numpy 1.5.1 with either Python 2.7.1 (x86) or 2.5.2 (x86) 
and,
- 3 of them running numpy 1.3 and Python 2.5.2.(x86).

   import numpy

   x = numpy.array([[0.00010876945607980702], [0.22568137594619658], 
[5.6435218858623557]])

   for i in range(10):
      m = numpy.array([[36.0 * 36.0, 36.0, 1.0] for j in range(6)])
      y = (numpy.dot(m, x) - 13.90901663) * 1000.0
      print y[0]

The output on each machine running 1.5.1 are similar to this one:
   [  5.00486230e-06]
   [  5.00486408e-06]
   [  5.00486230e-06]
   [  5.00486408e-06]
   [  5.00486230e-06]
   [  5.00486408e-06]
   [  5.00486230e-06]
   [  5.00486408e-06]
   [  5.00486230e-06]
   [  5.00486408e-06]

I cannot make sense of the changes of the least significant digits across 
different iterations of the for loop since its body is actually constant. Note 
that this behavior is hard to reproduce: on some machines I had to insert dummy 
print statements here and there to reproduce the bug or increase the number of 
iterations or the length of the arrays inside the loop. 

Also, I could not reproduce it with older versions of numpy such as 1.3. Is this 
behavior expected ? Is there a way to make sure the results of a numpy floating 
point computations remains the same for multiple runs ?

Thanks,
Gregory.