pytorch recipe

[escorciav]

Create your own layer (forward&backward)

(within python)

• numpy/scipy, check documentation/tutorial.

• cupy
  For GPU acceleration.
  numpy operations. use dlpack as bridge.
  Porting existing kernel as a string. example

• numba.
  speed-up cpu forward&backward.
  write cuda kernel within python. example
  suggestions from numba-core-dev.

• cython
  speed-up CPU forward&backward.