Cmdstan 2.30.1 shows warning when compiling with 'reduce_sum'

Interfaces CmdStan
1

Hi,

I am using 'Cmdstan ‘2.30.1’ (cmdstanr version ‘0.5.2’) as backend with brms (version ‘2.17.3’). During compilation, following warnings are shown but model compiles successfully and samples as expected.

&, const double&>::recursive_reducer::args_tuple_' [-Wreorder]
     std::tuple<Args...> args_tuple_;
                         ^~~~~~~~~~~
stan/lib/stan_math/stan/math/rev/functor/reduce_sum.hpp:78:5: warning:   when initialized here [-Wreorder]
     recursive_reducer(recursive_reducer& other, tbb::split)
     ^~~~~~~~~~~~~~~~~

I guess these warnings are harmless but still I want to resolve the issue. I would really appreciate your help.

Thanks

2

I understand the concern, but can assure you this warning is harmless.

The reorder warnings (i.e., -Wreorder) occur when the declaration of member variables is in a different order than their initialization. The reason it’s a warning is that the calls are executed in order or declaration, not in the order they’re written in a C++ constructor.

The problematic source file is https://github.com/stan-dev/math/blob/develop/stan/math/rev/functor/reduce_sum.hpp. The issue is that the member variables are declared as:

    const size_t num_vars_per_term_;
    const size_t num_vars_shared_terms_;  // Number of vars in shared arguments
    double* sliced_partials_;  // Points to adjoints of the partial calculations
    Vec vmapped_;
    std::stringstream msgs_;
    std::tuple<Args...> args_tuple_;
    scoped_args_tuple local_args_tuple_scope_;
    double sum_{0.0};
    Eigen::VectorXd args_adjoints_{0};

but the constructor throwing the error initializes this way:

          num_vars_per_term_(other.num_vars_per_term_),
          num_vars_shared_terms_(other.num_vars_shared_terms_),
          sliced_partials_(other.sliced_partials_),
          vmapped_(other.vmapped_),
          local_args_tuple_scope_(),
          args_tuple_(other.args_tuple_)

The last two are out of order—args_tuple_ should’ve come before local_args_tuple_scope_. Given there’s no dependent computation, this won’t cause a problem.

I opened an issue in Stan math to rearrange the initializers, which will not change behavior, but will get rid of the warning:

2 Likes
3

Thank you for looking into the matter and opening an issue.

4

The fix is done and will be in the next release. Here’s the GitHub pull request with the patch.

5

Thanks for the update.

Regards