This will be necessary to avoid code bloat with more than 2 devices
(execution types).

- cudaMemcpy is slower than our ParallelFor kernel for CUDA
- use std::copy and std::equal instead of memcpy and memcmp, but only as
  sequential fallbacks
- use parallel algorithms for containsValue and containsOnlyValue (again
  with sequential fallbacks)

It has nothing to do with devices.

Also set the buffer size to 1 MiB, because larger buffer size slows down
memory copies significantly (e.g. MeshTest would take about 10x longer).

Addresses #26

Moved synchronization of smart pointers from Devices::Cuda into TNL::Pointers namespace as free functions

synchronizeDevice() was renamed to synchronizeSmartPointersOnDevice()
for clarity - there are many similarly named functions in CUDA (e.g.
cudaDeviceSynchronize()).

Moved (most of) static methods from TNL::Devices::Cuda as free functions into separate namespace TNL::Cuda

The class TNL::Devices::Cuda was too bloated, breaking the Single
Responsibility Principle. It should be used only for template
specializations and other things common to all devices.

The functions in MemoryHelpers.h are deprecated, smart pointers should
be used instead.

The functions in LaunchHelpers.h are temporary, more refactoring is
needed with respect to execution policies and custom launch parameters.

Fixes #46

The implementation for std::string (which is a base class of
TNL::String) is perfectly sufficient.

Fixes #11

Matouš Fencl authored 6 years ago and Tomáš Oberhuber committed 5 years ago

MPI implemented for CPU and GPU in 2D but meshFunction.template synchronize< Communicator >(); doesn't copy overlaps.

StaticArrayAssignment expects the arguments passed as reference.