- 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

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

- was not used consistently before
- symbol names are expected to be valid C identifiers
- for complicated field names, the latex_name of field should be used

- 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

- better latex display for indirect accesses
- new field type 'custom': only custom fields can be accessed indirectly
  no static bounds check possible for custom fields

- removed warnings
- added flake8 as CI target

- test run again
- notebooks not yet

- previously all objects where cached by id()
- for waLBerla simulations in each time step a new np.array view
  is created from the waLBerla field. Each of these views has a
  different id -> caching did not work for waLBerla setups
- changed hash for numpy arrays: instead of id, a tuple of
  (dataPtr, strides, shapes) is used as hash input

- scaling interface width eta instead of surface tensions tau to correct
  interface profile & surface tensions

João Victor Tozatti Risso authored Dec 13, 2017 and Martin Bauer committed Jan 19, 2018

Concept: Generate code involving the (un)packing of fields (from)to linear
(1D) arrays, i.e. (de)serialization of the field values for buffered
communication.

A linear index is generated for the buffer, by inferring the strides and
variables of the loops over fields in the AST. In the CPU, this information is
obtained through the makeLoopOverDomain function, in
pystencils/transformations/transformations.py. On CUDA, the strides of
the fields (excluding buffers) are combined with the indexing variables to infer
the indexing of the buffer.

What is supported:
    - code generation for both CPU and GPU
    - (un)packing of fields with all the memory layouts supported by
    pystencils
    - (un)packing slices of fields (from)into the buffer
    - (un)packing subsets of cell values from the fields (from)into the buffer

Limitations:

- assumes that only one buffer and one field are being operated within
each kernel, however multiple equations involving the buffer and the
field are supported.

- (un)packing multiple cell values (from)into the buffer is supported,
however it is limited to the fields with indexDimensions=1. The same
applies to (un)packing subset of cell values of each cell.

Changes in this commit:

- add the FieldType enumeration to pystencils/field.py, to mark fields
of various types. This is replaces and is a generalization of the
isIndexedField boolean flag of the Field class. For now, the types
supported are: generic, indexed and buffer fields.

- add the fieldType property to the Field class, which indicates the
type of the field. Modifications were also performed to the member
functions of the Field class to add this property.

- add resolveBufferAccesses function, which replaces the fields marked
as buffers with the actual field access in the AST traversal.

Miscelaneous changes:

- add blockDim and gridDim variables as CUDA indexing variables.

- renaming because of clashes with types.py from other packages

- cache relied on uniqueness of  python id()
- id may be reused if object is freed
-> object must be held alive
-> kernel keeps all it arguments it was ever called with, alive (problematic in terms of memory consumption)

-> smaller block

- abstraction layer for selecting CUDA block and grid sizes
  - line based (was implemented before)
  - block based (new, more flexible)
-  new conditional (if/else) ast node, which is necessary for indexing schemes (guarding if)

- bugfix for CUDA kernels with variable field sizes
- extended tests for pystencils gpu kernels

- windows support
- automatic caching and creation of shared library with all generated kernels
- restrict keyword and function prefixes are preprocessor macros now -> easier to generate one code for linux, cuda, windows