Shared memory references that are part of a GPU program signature are counted as belonging to the grid

[bastian-koepcke]

Shared memory references should only belong to each block.
Instead they are currently part of the grid as being passed to a GPU program via the grid-level function as a parameter.

Instead:

• for static allocations, find a way to make the allocation on the block-level
• for dynamic allocations, is there a way to pass in a dynamic allocation to the GPU program? Or can the same system as for static allocations be used somehow?