Question on RooFit's batch mode

lmeyerga · June 6, 2025, 8:23am

Dear experts,

If I run a fit with multiple PDF components e.g. f = f_a + f_b + f_c, do all PDF’s need to implement evaluateSpan in order to benefit from BatchMode? (I’m using root 6.24)

Thank you in advance.

Danilo · June 6, 2025, 1:52pm

Hi,

A lot changed in that area in the last few releases. Let me add in the loop @jonas .

D

lmeyerga · June 16, 2025, 3:09pm

Let me ping @jonas again to prevent the topic from closing automatically.

jonas · June 18, 2025, 6:16am

Hi! No, this has never been the case for the BatchMode, even as far back as 6.24.

There is a fallback RooAbsReal::evaluateSpan() implementation:

github.com/root-project/root

roofit/roofitcore/src/RooAbsReal.cxx

v6-24-00-patches


      
          ///   without recomputing those.
          /// - Create a new batch in `evalData` of the same size as those returned from the servers.
          /// - Use the input data to perform the computations, and store those in the batch.
          ///
          /// \note Error checking and normalisation (of PDFs) will be performed in getValues().
          ///
          /// \param[in,out] evalData Object holding data that should be used in computations.
          /// Computation results have to be stored here.
          /// \param[in]  normSet  Optional normalisation set passed down to the servers of this object.
          /// \return     Span pointing to the results. The memory is owned by `evalData`.
          RooSpan<double> RooAbsReal::evaluateSpan(RooBatchCompute::RunContext& evalData, const RooArgSet* normSet) const {
          
            // Find leaves of the computation graph. Assign known data values to these.
            //
            // We can't use RooAbsArg::leafNodeServerList to find all leaves, sometimes a
            // RooAbsReal sits on top of a leaf in the computation graph but it doesn't
            // depend on it's values. The example here is a RooRealIntegral, which sets
            // the leaf values itself to integrate over them. That's why we only add the
            // parameters and observables here.
            RooArgSet allLeafs;
            RooArgSet parameters;

So if one component doesn’t implement evaluateSpan you will be able to use BatchMode and likely get a speedup.

Note that as Danilo said, the BatchMode was still experimental in 6.24 and not all fits might work correctly. But starting with ROOT 6.32, it is enabled by default and got renamed to the “new CPU likelihood evaluation backend”. The name BatchMode was confusing because it’s already used in other aspects of ROOT, e.g. the batch option for the ROOT interpreter (root -b).

lmeyerga · June 18, 2025, 8:30am

Thanks!