Enhancement: GLOBAL/LOCAL Loops (OMP=1)

[hikettei]

• AUTO_SCHEDULER=1 is an default behaviour!
• Finding better schedule by ISL
• Fast CLANG=1 is enough for this PR
• Introduce OMP=1
• CLANG: Pragma SIMD
• Parallelise considering the dynamic shape (Introduce IF)
• Memory planner does not conflict
• They should be reconfigurable depending on the device
• APIs for doing classical loop transformations (and docs!)
• BEAM Search
• https://github.com/facebookresearch/TensorComprehensions/blob/master/tc/core/polyhedral/scop.cc#L381-L409
• Reading: https://inria.hal.science/hal-01253322v1/document
• https://repository.dl.itc.u-tokyo.ac.jp/record/6112/files/48136454.pdf
• TODO: Add Reference Sections in the documentation
• Adding primitive schedule commands first.
  • polyir-parallel
  • polyir-tile
  • polyir-vectorize
  • polyir-fuse
  • should not rely on isl algorithm right?
  • Instead adding BEAM Search

[hikettei]

Cool 😎 (but need a chore for accumulation otherwise ISL cannot label coincidence)

• This is not a scalar: float _acc_0 = 0.0;
  

[hikettei]

clang: vectorize by gcc options
with regard to mathematical ops -> replace them with approx ops