Selecting rows using 'log10' in the query returns wrong results

[153957]

I have PyTables (v3.2.2) running on multiple systems (OS X, Scientific Linux CERN, and Cent OS).
On the Mac numpy 1.9.4 and numexpr 2.4.4 are installed
On the SLC machine numpy 1.10.4 and numexpr 2.4.6 are installed.

I have a table (sims) with a column (energy) containing values ranging from 1e12 to 1e18. Most values are at full powers of ten (e.g 1e13 and 1e14) or in between (e.g. 10**13.5 and 10**15.5). So to easily query these values I use the log10 option for PyTables queries. The following query works on the Mac, but returns the wrong result on the SLC machine:

import numpy
numpy.log10(sims.read_where('log10(energy) == 12')['energy'][:9])
# On Mac: array([ 12.,  12.,  12.,  12.,  12.,  12.,  12.,  12.,  12.], dtype=float32)
# On SLC: array([ 15.,  12.,  16.,  15.,  17.,  13.,  15.,  15.,  15.], dtype=float32)

So this may not be a PyTables bug perse, perhaps the fault lies with Numexpr. However, I encountered it while using PyTables. It would be good for PyTables to require a properly working version of Numexpr. Also, what is the source of this bug? Numexpr, Numpy, or something else?

We encountered a similar bug/inconsistency recently concerning the accuracy of the query when using log10: HiSPARC/sapphire#114

(Our issue on this problem HiSPARC/sapphire#116)

[153957]

It appears that Numexpr 2.4.6 is the culprit. We can downgrade to 2.4.4, unfortunately upgrading to 2.5.0 is not an option because that also breaks some parts of the code: HiSPARC/publicdb#131

[FrancescAlted]

The problem seems to be in the conda package . Closing this.

[tomkooij]

This needs reopening.

When using Intel MKL with numexpr (default in conda from numexpr=2.4.6) all table.read_where() queries with (VML) functions are broken.

Numexpr uses VML functions which need contiguous memory, but tables are numpy.rec.array()s which are non-contiguous.

This can be fixed by getting and passing the right arguments to NumExpr.run(). I have a fix, will make PR.