💡 [REQUEST] - The CPU kernels for hamiltonian related operator kernel

[hzhangxyz]

Description

Operators about Hamiltonian are only implemented in cuda now, which makes we cannot execute the program without cuda device.

• Implement the proto of CPU code based on the current GPU code. (Dev: cpu support for c extension #34)
• Conditional compilation in hamiltonian.py, if users are using CPU, skip to compile GPU code, and vice versa. (hamiltonian扩展条件编译 #42)
• Make CPU kernels better for CPU, such as, do not double check in add_into_heap, which is designed for mutex previously, and replace thrust::sort by std::sort.
  • Remove duplicated check in add_into_heap (Remove duplicated if checking in add_into_heap for cpu. #49)
  • Use std::sort(from C++ standard library) instead of thrust::sort(from CUDA indeed). Currently, we also use thrust::sort_by_key but there is not such equivalent in C++ standard library. (Remove thrust from dependencies of the cpu kernel. #46)

[hzhangxyz]

@windy-pig

For conditional compilation, you can add another argument device for function _load_module in hamiltonian.py. There are four callers for _load_module currently, the caller function _prepare should always use cpu(torch.device(type="cpu", index=0)), and as for other three callers, use the argument passed to self._prepare_data. BTW, please use different build_directory(such as ... / device.type / name) in _load_module for different device, which help to reduce duplicated compilations.

[hzhangxyz]

@windy-pig
As for the third optimization, you need to

• Remove duplicated check in add_into_heap
• Use std::sort(from C++ standard library) instead of thrust::sort(from CUDA indeed). Currently, we also use thrust::sort_by_key but there is not such equivalent in C++ standard library.

[hzhangxyz]

Everything completed.