Markus Holzer authored Feb 07, 2019 and Martin Bauer committed Apr 26, 2019

The communication of the ghost layers used to communicate just all
values in between one time step to make sure that everything is correct.
Furthermore the communication was only valid for pull stream steps.
The improved communication distinguishes automatically between pull and
push and communicates only values which are needed. With this improvement
it was possible to implement the EsoTwist streaming scheme.

- typos, more content
- added numba to benchmark comparison

-> replaced by 1e-16

- turned on restrict keyword by default (makes large difference on GPUs)
- smarter block indexing: changing block size depending on domain size
  Example: previously there where (1,1,1) blocks when requested
  block size was (64, 1, 1) and domain size (1, 512, 512), now the
  block size is changed automatically to (1, 64, 1) in this case
- added __lauch_bounds__ to kernels to allow better optimizations from
  the CUDA compiler

- style changes marked by flake
- using newest kerncraft version

This restructuring allows for easier separation of modules into
separate repositories later. Also, now pip install with repo url can be
used.

The setup.py files have also been updated to correctly reference each
other. Module versions are not extracted from git state

- counter-based philox RNG: counter/key is filled with cell coordinate
  and optional external parameters like block position and time step
- works on CPU and GPU - on CPU only for non-vectorized versions

- introduced more flexible "CustomCodeNode" that can inject
  backend-specific hand-written code

boundary conditions can specify how the index list should be built:
- list coordinates of domain or boundary cells (previous always inner)
- list all links or only the first link

- complicated pressure tensor derivation not required

-> last concentration is automatically computed as 1-others
-> makes code more general

- conceptionally not required: force needs to be in transformed coords
- for 3 phase model it makes no difference after bug was fixed in
  extract_gammas

- previous implementation did not work for van der Walls eos
- works now in V-P coordinates

- sweep updates pdf field in place, but does not load all the values
  before overwriting them
-> added a new transformation on AssignmentCollection that loads all
   read values first

- small (length < 5) arrays with shape and stride information had to be
  memcpy'd to the GPU before every kernel call
- instead of passing the information as arrays, the single elements are
  passed
- leads to more function arguments, but simplifies GPU kernel calls

-> changes in all backends required

- correction functions