CuVec

• fix clang builds (#11)
• fix builds without SWIG (#11)
• optimal test thread spawning

• fix CMake export target (#10)
• document SWIG usage

• make CUDA optional (#9)
  • make __half/e optional
• add CPU versions of examples
• add & update tests

• major SWIG updates (#8)
  • expose SwigCuVec<T>
  • separate core (swvec.i) & example (example_swig.i) modules
  • update tests & examples
• update & tidy performance tests

• add SWIG support via cuvec.i (#7)
• add & document example_swig
• improve & document example_mod
• fix external #includes
  • multiple definitions
  • missing cuda runtime header
  • use of static & templates
  • namespace clashes
• statically link CUDA libraries
• use macros to reduce code duplication
• add and fix tests
• add performance checks
• misc code tidy

• add __cuda_array_interface__ (#4)
• add tests
• update documentation

• fix asarray not copying on ndarray.view(CuVec) (#2)
• fix potential name clashes with namespaces
• consistent name capitalisation in documentation

• add half-precision (fp16/float16/__half/'e') (#1)
• fix unneeded asarray memcopy on raw objects
• fix external #include "cuvec.cuh"
  • remove unneeded cuhelpers
• update supported typecodes
• and external module example
• add performance test
• add asarray tests

• add proper asarray equivalent
  • add tests
• minor tidy

• fix external allocation (https://github.com/python/cpython/pull/24283)

• catch exceptions
• update __init__ keywords
• build updates
  • add CUVEC_DEBUG flag (default: OFF)
  • use CMake add_compile_definitions
• test updates
  • increase verbosity
• documentation updates
  • add contributing notes
    • using editable install & pytest
• misc tidy
  • clean asarray implementation

• add asarray convenience for wrapping external libraries
• minor code tidy

• add external project usage documentation
• add more accurate build warning message

• add seamless numpy-like functionality
  • rename cuvec.from_numpy() => cuvec.copy()
  • make cuvec.zeros() and cuvec.copy() return numpy.ndarray-like objects
• fix potential installation errors
• add complete documentation
• add citation information

Initial proof of concept.

Unifying Python/C++/CUDA memory: Python buffered array <-> C++11 std::vector <-> CUDA managed memory.

• Provides creation functions
  • Python object
    • zeros
    • from_numpy
  • C++ vector
    • CuVec<T> with identical interface as std::vector<T>
  • CPython API (castable to PyObject *)
    • PyCuVec<T> *PyCuVec_zeros(std::vector<Py_ssize_t> shape);
    • PyCuVec<T> *PyCuVec_zeros_like(PyCuVec<T> *other);
    • PyCuVec<T> *PyCuVec_deepcopy(PyCuVec<T> *other);
• Quick start documentation
• 100% test coverage
• Allows CUDA-free installation for sdist

Requires a C++11 compiler, CUDA compiler, and an NVIDIA GPU with compute capability 3.5 or greater.
Python functionality requires Python 3.6 or greater.