[NumPy-Tickets] [NumPy] #2209: Inconsistent dot product with BLAS

NumPy Trac numpy-tickets@scipy....
Sat Sep 1 06:00:57 CDT 2012


#2209: Inconsistent dot product with BLAS
-------------------------+--------------------------------------------------
  Reporter:  law         |       Owner:  somebody                          
      Type:  defect      |      Status:  closed                            
  Priority:  normal      |   Milestone:  Unscheduled                       
 Component:  numpy.core  |     Version:  1.6.2                             
Resolution:  wontfix     |    Keywords:  dot product blas dotblas sse2 sse3
-------------------------+--------------------------------------------------
Changes (by pv):

  * status:  new => closed
  * resolution:  => wontfix


Comment:

 The rounding error depends on data alignment, because compilers do
 arithmetic operations especially with SSE in an alignment-dependent order.
 As the alignment is not deterministic across different runs, the results
 are also not.

 See e.g. http://article.gmane.org/gmane.comp.python.numeric.general/48179

 The issue occurs on the BLAS level. There is a performance tradeoff in
 ensuring fully deterministic results, so I do not think we will fix this
 issue.

 If full reproducibility is extremely important for you, you need to
 recompile all libraries (BLAS, LAPACK, etc.) involved with a compiler that
 has a mode for enforcing floating point reproducibility --- I believe the
 Intel compilers have this.

-- 
Ticket URL: <http://projects.scipy.org/numpy/ticket/2209#comment:1>
NumPy <http://projects.scipy.org/numpy>
My example project


More information about the NumPy-Tickets mailing list