5 Replies Latest reply on Sep 18, 2013 3:28 AM by rvstebbing

    Non-deterministic results from dpotrf when re-running problem on the same machine

    rvstebbing

      Dear List,

       

      I am using ACML 5.3.1 for Windows (ifort64, 32-bit int build) with Microsoft Visual Studio 2012 (Update 3). I have a program to do the in-place Cholesky factorisation of a 126x126 matrix using dpotrf and I am linking against libacml_mp_dll.lib. When I run the program and dump the raw doubles (as hex) to file, I don't get consistent output between runs on the same machine. Specifically, there are differences in the lower-bits of some entries. However, when I run the factorisation multiple times within a single program instance the results are consistent.

       

      Is this considered a bug?

       

      The C++ to generate the dumps is here: ACML 5.3.1 dpotrf example - 8b959cb7

      The Python code to run new instances until a difference is found is here: run_till_difference.py - dc2d3dca

      Example output files are ACML_dpotrf_0.txt - 6f0435ad and ACML_dpotrf_1.txt - e8d61509

       

      I appreciate any feedback you can give on this issue.

       

      Cheers,

       

      Richard