Inconsistent return types of random_generator functions in cpu and gpu

[ChaiBapchya]

The file - /incubator-mxnet/include/mxnet/random_generator.h includes CPU and GPU implementation of random number generator functions.

Close look at cpu implementation

MSHADOW_XINLINE FType uniform() {...}
MSHADOW_XINLINE FType normal() {...}

While in case of gpu

MSHADOW_FORCE_INLINE __device__ float uniform() {...}
MSHADOW_FORCE_INLINE __device__ float normal() {...}

Why is it hardcoded in gpu? Shouldn't it be generic?

[leleamol]

Hello @ChaiBapchya
For questions, please submit on MXNet discussion forum (https://discuss.mxnet.io), where it will get a wider audience and allow other to learn as well.
I'm closing this issue now in favor of the discussion forum issue you will file, please feel free to re-open if closed in error.
Thanks!

@mxnet-label-bot add [Question]

[leleamol]

@mxnet-label-bot add [Question]

[asitstands]

There is a template specialization for double in gpu. Generic implementation without specialization is not so simple because the underlying curand API is not generic. It requires an implementation of a new generic random number generator. Actually I have tried to address some issues in random number generation by implementing a new generator common for cpu and gpu (dmlc/mshadow#338), but no one might be interested in the proposal 😁