[SciPy-user] minimizers don't work - d1mach problem

[Robert Kern]

Christian Kristukat wrote:
> Robert Kern wrote:
> 
>>Christian Kristukat wrote:
>>
>>>Hi,
>>>sorry for posting this again, but I'd really like to have the minimizers work
>>>again. Every call to any of the miminizers scipy.fmin* fails with:
>>>
>>> Adjust D1MACH by uncommenting data statements
>>> appropriate for your machine.
>>>STOP 779
>>>
>>>I already looked at Lib/special/d1mach.f but I don't see what I should change
>>>there. Anyway, d1mach.f hasn't been changed for ages so I doubt that the problem
>>>is there.
>>>
>>>Here's my configuration:
>>>
>>>P4
>>>
>>>Python 2.4.1 (#1, Sep 13 2005, 00:39:20)
>>>[GCC 4.0.2 20050901 (prerelease) (SUSE Linux)] on linux2
>>>
>>>numpy/scipy from svn
>>>ATLAS built with gfortran
>>
>>Odd. For your configuration, D1MACH ought to be taking the IEEE LITTLE ENDIAN
>>branch earlier on, and the STOP 779 statements should never be reached. Try
>>adding these lines to diagnose the issue:
>>
>>Index: Lib/special/mach/d1mach.f
>>===================================================================
>>--- Lib/special/mach/d1mach.f   (revision 1630)
>>+++ Lib/special/mach/d1mach.f   (working copy)
>>@@ -60,6 +60,9 @@
>> C     ON FIRST CALL, IF NO DATA UNCOMMENTED, TEST MACHINE TYPES.
>>       IF (SC .NE. 987) THEN
>>          DMACH(1) = 1.D13
>>+         write(*,*) 'DMACH(1) == ', DMACH(1)
>>+         write(*,*) 'SMALL(1) == ', SMALL(1)
>>+         write(*,*) 'SMALL(2) == ', SMALL(2)
>>          IF (      SMALL(1) .EQ. 1117925532
>>      *       .AND. SMALL(2) .EQ. -448790528) THEN
>> *           *** IEEE BIG ENDIAN ***
> 
> There's no additional output, so the if condition seems to be False.

In that case, the STOP statements wouldn't be executed either. Add more
write(*,*) statements to explore it, if you like. But:

> There's
> another d1mach.f in Lib/integrate/mach. Which one is the relevant here?

Couldn't say. Try adding some different write(*,*) statements to it in order to
find out.

> Btw, the minimizers partly run, maybe in case when the gradients are far from
> IEEE numeric limits?

That wouldn't trigger anything in d1mach, though.

What is your function? Are you using anything in scipy.special or scipy.integrate?

-- 
Robert Kern
robert.kern at gmail.com

"In the fields of hell where the grass grows high
 Are the graves of dreams allowed to die."
  -- Richard Harter