[Feature Request] Merge CUDA and ROCm codes

[njzjz]

Summary

Merge CUDA and ROCm codes into one file and distinguish them using macros.

Detailed Description

There are lots of duplicated codes between CUDA and ROCm. Most of them are the same. Adding a new GPU method is troublesome as we need to add twice to both CUDA and ROCm codes.
I notice that TensorFlow implements CUDA and ROCm in only one file, such as https://github.com/tensorflow/tensorflow/blob/00a17d7451a789a4df994dac7d616ce2f4438ff0/tensorflow/core/kernels/relu_op_gpu.cu.cc#L16. Some different codes between them are controlled using macros, i.e., GOOGLE_CUDA and TENSORFLOW_USE_ROCM. A universal launch function, GpuLaunchKernel, is used for CUDA and ROCm.

Below is the roadmap for DeePMD-kit:

• Migrate to universal functions for CUDA and ROCM created in gpu_cuda.h and gpu_rocm.h merge cuda and rocm files #2844
• Make the method have the same name between CUDA and ROCM, i.e., no _rocm suffix anymore remove _cuda or _rocm suffix #2839
• Clean the duplicated codes and use GOOGLE_CUDA || TENSORFLOW_USE_ROCM for the same code
  • lib/cuda and lib/rocm merge cuda and rocm files #2844
  • op merge CUDA and ROCm codes in op #2847
  • lib/tests merge CUDA and ROCm codes in tests #2846
  • lib/include merge CUDA and ROCm in header files #2845

Further Information, Files, and Links

No response