Worth loading all to shared memory?

Accelerated Computing CUDA CUDA Programming and Performance
1

I have a matrix to which I need to add a certain quantity. Let’s say 2 to each element.
In order for a kernel to do this, and since the algorithm is basically:

1-readElement
2-Add 2
3-writeElement

the question is: Is it worth it to load each submatrix into shared memory to perform the operations or can I just read modify and write the submatrix through registers? I am assuming coalesced global memory reads in both cases.
My opinion is that doing it with registers I can save shared memory.

2

No, it is not worth it.

3

With such simple processing you won’t benefit from using shared memory. Just make sure your global memory reads and writes are coalesced.