Performance comparison and document

Hi, I have a few questions about this repo.

1. How much faster is your implementation compare to the original pytorch/TensorFlow version ?
2. Can you explain more about how do you speed up the operation?
3. If I want to implement a faster TensorFlow version of pointnet++, do you have any suggestion?

Thanks.

I didn't compare the running time systematically. I just tested each ops separately and generally it will be 2+ times faster than original version, and it also depended on the number of input points, so it's hard to say how much faster than original version. You could compare with the original codes to find the differences. CUDA codes could be shared between TensorFlow and PyTorch, so you just need to copy my CUDA kernel and write the TensorFlow related cpp interface files.