The codegenerated TMVA MLP reader classes have some
float *fWeights to store the intermediate node values during the network evaluation. These are not threadsafe. Once making them local C-arrays in the evaluation function helgrind is happy (test done in https://gitlab.cern.ch/pseyfert/tmva-threadsafety).
to be done:
check with @raaij that was the only reason for the mutex
check if I didn't break the computation