- previous method did not work with kernels generated for walberla where
  block size changes are made at runtime
- device query does not always work, since the compile system may have
  no GPU or not the same GPU
-> max block size is passed as parameter and only optionally determined
   by a device query

-> different fields need different shape/stride symbols

Block.subs method tried to be too smart:
a = field[..]
b = a + b

was "simplified" incorrectly to
b = field[...] + b

- stencil plotting & transformation now in ps.stencil
- additional documentation & notebooks

- updates in readme text
- added release script

- typos, more content
- added numba to benchmark comparison

Nils Kohl authored 5 years ago and Martin Bauer committed 5 years ago

- introduced AbstractField and AbstractAccess

Fixes #28

When two loops have assignments to the same symbol with different
rhs and both are pulled before the loops, one of them is now renamed.
Previously one of them was left inside the loop.

Fixes #27

- 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

see https://github.com/sympy/sympy/issues/16662

- replace_density_and_velocity simplification produced terms like
  0 * omega, because sympy's auto-eval is turned off
- sympys CSE routine can apparently only handle evaluated terms
- solution: evaluate multiplications with zero (i.e. replace them by 0)

- style changes marked by flake
- using newest kerncraft version