More efficient GPU compute about waveorder

[lihaoyeh]

I am aware of this when initially writing up these methods. There are two reasons why I ended up choosing to do so.

First, GPU memory handling with class variables in jupyter may not be optimal. The GPU memory will not be automatically freed up even though the GPU array is reassigned to something else. Deleting the class in the notebook also does not free up the GPU memory. That's why I keep all of the GPU array assignments in the class methods. When I finished using the GPU array, I clean it right away in the method. Even so, there is still some accumulation needed to be sorted out. Doing so ensures that the memory stays as clean as possible for other processes. Otherwise, careless accumulation of the memory space tends to limit the computational capacity. This might not be a major concern if we do not use jupyter often.

Second, taking arrays in and out of the GPU memory takes much less time than the actual computation. I have profiled these operations before. Removing these in-and-out operations does save time but it is marginal to the overall computation time.

[ieivanov]

Second, taking arrays in and out of the GPU memory takes much less time than the actual computation. I have profiled these operations before. Removing these in-and-out operations does save time but it is marginal to the overall computation time.

Ah, that's good to know, thanks for explaining.