High-performance, multidimensional arrays in modern C++.

• Read the codebase again and write an overview of the major concepts to make it easier to understand later.

• Revise the codebase:

  • Change the line width to 100; remove headers from files.
  • Think about any simplifications to the overall design (coord, index, array, etc.).
    • Do we really distinctions among all the different types of "constant"?
  • Examine the assembly dump for range_iterator again to see if there really is a degradation.

• Update range module.

  • Update range_iterator.hpp so that the direction attribute is taken into account.
  • Update index_wrapper and array_wrapper so that we use the macro decltype deducting trick instead of traits classes.
  • Add range.size()

• Replace ccbase mpl parsing with metaparse library; just include metaparse in ccbase.

  • Replace parse_natural, parse_ratio, parse_wsv, etc.
  • Update ratio_literal.hpp.
  • Delete the unused headers.
  • Update range_literal.hpp?

• Finish array module.

• GPU support.

  • Test compilation with Cuda 7.5.
  • Support for GPU allocation and transfer.
  • Support for GPU execution using heterogenerous lambdas.

• Simplifications in the future:

  • Constexpr lambdas (C++17) could be used to greatly simplify the code in many places (e.g. elementwise views).

• Consider using the new C++14 named operators (e.g. <repeat>) where appropriate.

• Come up with the syntax to write loops using generic packets instead of the same scalar type of the array. This would allow the loop operation to be expressed using SIMD operations. Combined with the syntax for unrolling loops, this can become very powerful.