iaea-nds / datp-python

DATP code developed by Wolfgang P. Poenitz translated to Python

DATP - Python

The DATP code preprocesses experimental data sets prior to fitting with the GMAP code. Both codes have been developed by Wolfgang P. Poenitz in Fortran 77. As Fortran 77 is not used that much anymore nowadays, the conversion to Python will facilitate future developments. Due to the large ecosystem of extension modules around Python, the code logic of DATP and GMAP can be formulated on a higher level, thereby reducing code complexity. Moreover, also interfacing with other applications, such as web applications and databases will become easier.

Approach to conversion

Because the code is used for the evaluation of neutron cross section standards, the functional equivalence of the Python and Fortran version is very important. Only then can methodological improvements in the new code be made and their effect on the results studied.

For this reason, the code was translated in a first step on a line by line basis to Python, keeping exactly the same logic and level of abstraction as present in the Fortran code, even keeping goto statements which had to be introduced into Python using a bytecode hack implemented by the goto-statement module. Also the fortranformat module proved invaluably useful due to heavy input/output operations and associated FORMAT edit descriptors which do not have a perfect counterpart in Python.

To ensure the functional equivalence of the Python code and the Fortran code, a test case was written that compares the numerical output of both codes. As Python works natively with double precision floats whereas the Fortran code with single precision and, in addition, Python and Fortran implement different rounding schemes, small numerical differences in the outputs have to be expected. The test case therefore relies on the numdiff tool to compare the outputs, which ignores small numerical differences. In the implemented test case, it is set up to accept relative differences below 5x10^(-5).

Assuming that you cloned this repository, in order to run the test from a command line in Linux, you have to change into the directory tests/test_001. Being there execute:

export NUMDIFF_EXE=<path to numdiff binary>
./run_test.sh

The shell script will return a text message on the console whether outputs are equivalent and also indicate the result with the return code, i.e., 0 for equivalent output and 1 if differences were detected. In the latter case, the file in result/numdiff.out contains details on the location of the differences in result/fortran/DAT.RES and result/python/DAT.RES.

Future developments

Now that a first draft version of the Python code is available, with large probability fully functionally equivalent to the Fortran version due to implementing exactly the same program flow and the equivalence of the output being tested, the Python code is going to be incrementally refactored, always validating the effect of updates against the output of the original Fortran version.