GPU-accelerate dIdV (no tilt)

[GianmarcoMidena]

• Module: ProbeSTM

General subtasks

• Static analysis of PP-STM code.
• Review other projects embracing GPU acceleration, e.g., PP-AFM.
• Test GUI on GPU.
• Evaluate GPU frameworks: PyOpenCL, Kokos, PyTorch, CuPy, Numba, RAPIDS (cuDF), GPU4PySCF, ...

Specific subtasks

• Automatic test
  • Output tensor dimensionality
  • Pointwise independence
  • Even-symmetry
  • Match behavior of NumPy backend
    • spd sample orbitals
      • s tip orbital
      • pz tip orbital
      • pxy tip orbital
      • CO tip orbital
    • sp sample orbitals
      • s tip orbital
      • pz tip orbital
      • pxy tip orbital
      • CO tip orbital
      • dz2 tip orbital
      • dxzyz tip orbital
• GPU-accelerate (OpenCL, PyTorch)
  • spd sample orbitals
    • s tip orbital
    • py tip orbital
    • pz tip orbital
    • px tip orbital
  • sp sample orbitals
    • s tip orbital
    • py tip orbital
    • pz tip orbital
    • px tip orbital
    • dyz tip orbital
    • dz2 tip orbital
    • dxz tip orbital

[ondrejkrejci]

ReadSTM cannot be probabky GPU accelerated, but should be CPU accelerated

Similar to GridUtils. Those are also changing in AFM, maybe we can put ppafm as necessary dependency and use theirs.

IETS is not working as it is not working in PPAFM now. I will know later this month if there is even any market for this.

Kokos could be an aditional GPU framework.