__attribute__ in separatable Convolution

Hello everyone

I’m looking into the Convolutionseparable example code for studies

I see some Quilifiers in the kernel like,
attribute((reqd_work_group_size(ROWS_BLOCKDIM_X, ROWS_BLOCKDIM_Y, 1))).

The specification says that it helps while compiler compiles the Kernel by giving the size of the group.And no more details…

What kind of help does it mean? like memory coalescing?
Does it lead to performance improvement?