consider using THC

[soumith]

Hi there devs, very good work!

I see that a lot of numpy functions are yet to be implemented for the GPU.
Instead of writing Tensor Op CUDA kernels from scratch, I suggest that you take a look at https://github.com/pytorch/pytorch/tree/master/torch/lib/THC
It is a C99 library with a lot of tensor functions implemented and might be helpful in saving time and effort.

This library currently powers torch and pytorch and has been developed and optimized for years.
The documentation for the functions is here: http://pytorch.org/docs/torch.html

Just a suggestion that might help accelerate your progress :)

cc: @tqchen

[lryta]

Thank you very much @soumith ! We will definitely look into THC.

[tqchen]

I opened an issue in MXNet apache/mxnet#4735 to discuss this