Explore whether using pointers to const data as arg to CUDA kernels improves performance

[mattldawson]

Currently, CUDA kernels get passed const structs like LuDecomposeParam which include pointers to arrays of indices. However, the array element types are not const. Explore whether creating something like a ConstLuDecomposeParam struct that has data member pointers to const data, and then having this type be used as the argument to CUDA kernels (with appropriate assignment operator overloads) will result in better pre-fetching of data in memory during the execution of the kernel.

Possible definition of new struct:

struct ConstLuDecomposeParam
{
  const std::pair<const std::size_t, const std::size_t>* niLU_;
  const char* const do_aik_;
  const std::size_t* const aik_;
  const std::pair<const std::size_t, const std::size_t>* uik_nkj_;
  const std::pair<const std::size_t, const std::size_t>* lij_ujk_;
  const char* do_aki_;
  const std::size_t* aki_;
  const std::pair<const std::size_t, const std::size_t>* lki_nkj_;
  const std::pair<const std::size_t, const std::size_t>* lkj_uji_;
  const std::size_t* uii_;
  std::size_t niLU_size_;
  std::size_t do_aik_size_;
  std::size_t aik_size_;
  std::size_t uik_nkj_size_;
  std::size_t lij_ujk_size_;
  std::size_t do_aki_size_;
  std::size_t aki_size_;
  std::size_t lki_nkj_size_;
  std::size_t lkj_uji_size_;
  std::size_t uii_size_;

  ConstLuDecomposeParam(const LuDecomposeParam& other) :
    niLU_(other.niLU_),
    ...
    {}
};

or we can change the existing LuDecomposeParam struct with the new one above (without the initializer) and see if it works too.