How to do fast bitwise operations on a scalar with the `cupy` library

Question

I am building a reinforcement learning model on the GPU, so I am using chainer which has the cupy backend. cupy is intended to be a duplicate of numpy except that it operates on the GPU.

I asked this question earlier on how to do fast bit shifting on a scalar in numpy, and the answer was easy: I need to do my bit shifting on the actual numpy.uint64 object and not on a numpy.array object. It would be nice if I could transfer my bit shifting methods over to cupy to get the same speedups.

However, cupy documentation requires that scalars be on the GPU instead of the CPU (source). This means that either I...

represent my scalar as an array, which brings up my original problem in my linked question above, or
push my scalar integer to the CPU, do my computations, and push it back to the GPU, which is also slow.

If I want to do hundreds of thousands of bit-shifts on a scalar value, although this would take less than a second in numpy, this takes way too long in cupy. How do I speed up bit shifting of scalars in cupy?

Shifting a single scalar boxed in an array is slow, but if you put hundreds of thousands of scalars in an array and bitshift it will be fast. Do you have to do the shifts one at a time? — Steven Rumbalski, Oct 12 '18 at 20:43
Yes, I cannot batch my bit shifts. They must be done one at a time. — Paul Terwilliger, Oct 12 '18 at 20:44

How to do fast bitwise operations on a scalar with the `cupy` library

0 Answers0