RDataFrame EnableImplicitMT() results in wrong results

ROOT
1

I am new to RDataFrame and trying to learn it step by step.
I could represent my results from regular pyROOT loop using:

    df_fit = filtered_df.Define("tof", "tof_fit(tNewCalHit, dToRefPointNewCalHit, dToLineNewCalHit, layerNewCalHit)")\

where the function is:

double tof_fit(const RVec<double> &t_hit, const RVec<double> &r_hit, const RVec<double> &d_hit, const RVec<int> &layer_hit){
    int n_hits = t_hit.size();
    map <int, vector <Hit> > layer_hits;
    for (int i=0; i < n_hits; ++i){
        Hit hit(t_hit[i], r_hit[i], d_hit[i], layer_hit[i]);
        layer_hits[hit.layer].push_back(hit);
    }
    // TCanvas canvas_debug;

    vector<Hit> selected_hits;
    for (int layer=0; layer<10; ++layer){
        vector <Hit>& arr = layer_hits[layer];
        if (arr.size() == 0) continue;
        Hit closest_hit = *min_element(layer_hits[layer].begin(), layer_hits[layer].end(), [](const Hit &hit1, const Hit &hit2){return hit1.d < hit2.d;} );
        selected_hits.push_back(closest_hit);
    }

    const int n_sel_hits = selected_hits.size();
    if (n_sel_hits <= 1) return 0.;

    double r[n_sel_hits];
    double t[n_sel_hits];
    double r_err[n_sel_hits];
    double t_err[n_sel_hits];
    for(int i=0; i < n_sel_hits; ++i){
        r[i] = selected_hits[i].r;
        t[i] = selected_hits[i].t;
        r_err[i] = 0.;
        t_err[i] = 10.;
    }
    TGraphErrors gr(n_sel_hits, r, t, r_err, t_err);

    gr.Fit("pol1", "Q");
    // gr.Draw("AP");
    // gr.SetMarkerStyle(20);
    // gr.SetTitle("Fit of a TOF vs distance to impact point; d, [mm];TOF, [ns]");
    // canvas_debug.Update();
    // sleep(3);
    TF1 fit = *gr.GetFunction("pol1");
    // if (fit.GetChisquare() > 1.) continue;
    double tof = fit.GetParameter(0);
    if (tof <= 0.) return 0.;
    return tof;
}

But if I add
ROOT.EnableImplicitMT()
in the beginning in my code it outputs wrong results! (It seems half of the points just missing in my final histo).

I assume there is something not “thread-safe” with my function?

cheers

ROOT Version: 6.22
Platform: RedHat CentOS 7.8.2003
Compiler: g++ 9.2.0
pyROOT: python3

2

Hi @FoxWise,
I believe you are right and that function is not thread-safe due to the Fit call. @moneta, can you confirm whether TGraphErrors::Fit is thread-safe?

Cheers,
Enrico

P.S.
@FoxWise at the cost of some performance you can add a std::mutex and a std::lock_guard to tof_fit to make it thread-safe. You can actually figure out which line(s) are unsafe by reducing the section protected by the lock until you recover the broken behavior.