created and tested bindings for hyperreduction

[pravija12345]

S_OPT

The input argument std::vector<int> *init_samples cannot have a default value nullptr. See the pybind11 documentation.

Although, init_samples needs not be a pointer in S_OPT implementation. Therefore, we switched it into std::vector<int> &, and set the default value to be std::vector<int>(0). This does not impact the internal execution.

The arguments f_sampled_row and f_sampled_rows_per_proc are STL-type output variables, which cannot change its value if it is passed as an input argument. Therefore it is removed from the input argument list, and the python function return these two variables as tuple:

import pylibROM.hyperreduction as hyperreduction
f_sampled_row, f_sampled_rows_per_proc = hyperreduction.S_OPT(u, num_basis_vectors, f_basis_sampled_inv, myid, num_procs, num_samples)

Note that Matrix& f_basis_sampled_inv is both input and output, and it is not a STL-type. This allows us passing it as an input argument which can modify its value within the function.

[dreamer2368]

Based on the pybind11 documentation, copy is the default action for STL-type arguments. This means that input/output variables of type std::vector<T> will essentially always involve a deep copy, unless we set a custom casting action.

@chldkdtn , do you have an idea about the typical size of std::vector<int> variables passed into libROM's hyperreduction routines, such as DEIM, GNAT, S_OPT, QDEIM? If they are not so large, then we may not need to do anything particular.

[dreamer2368]

This is for a future reference. If we do not want to have the 'out-of-box' copy action for std::vector, we have to construct py::list or py::array_t and return them, not directly returning std::vector. See this post.

[chldkdtn]

Based on the pybind11 documentation, copy is the default action for STL-type arguments. This means that input/output variables of type std::vector<T> will essentially always involve a deep copy, unless we set a custom casting action.

@chldkdtn , do you have an idea about the typical size of std::vector<int> variables passed into libROM's hyperreduction routines, such as DEIM, GNAT, S_OPT, QDEIM? If they are not so large, then we may not need to do anything particular.

Ideally, we do not want it to be big because the whole purpose of the hyper-reduction is to be able to represent a big size of nonlinear vector function with a small number of samples, say 50 or 100 (or perhaps in worst case 1000 at maximum). By the way, this deep copy is only for the output index only, right? which will be part of offline training cost. This should be okay for now, but let's keep it in mind so that in the future, in the context of "on-the-fly," we might need to avoid deep copy to make the efficiency optimized.

[pravija12345]

Created bindings for the following classes in hyper reduction module:

• DEIM
• GNAT
• QDEIM
• S_OPT
• STSampling
• Utilities

In these classes a key modification has been made to the init_samples parameter. The input argument std::vector *init_samples cannot have a default value nullptr, so, we switched it into std::vector &, and set the default value to be std::vector(0) (Note: These changes doesn't impact the internal execution).

Further more another key modification is made to the parameters f_sampled_row, f_sampled_rows_per_proc, and t_samples which were originally passed as input arguments. As these arguments f_sampled_row, f_sampled_rows_per_proc, and t_samples are STL-type output variables, which cannot change its value if it is passed as an input argument. Therefore it is removed from the input argument list, and the python function returns these variables as tuple if necessary.