CUDA: banking conflicts between various distortions?

Question

I just found out (from Why is only one of the distortions performed by SM in cuda? ) That Kepler GPUs can actually execute instructions from several (apparently 4) deformations at once.

Can a shared memory bank simultaneously serve four requests at once? If not, this would mean that bank conflicts can occur between flows of different distortions that occur simultaneously, even though there are no banking conflicts within any of the individual distortions, is it? Is there any information about this?

Answer 1

The computing power of 3.x devices (Kepler) has 4 skew schedulers on SM. On each cycle, each warp planner selects the basics and issues 1-2 instructions from the basics. An SM has only one storage unit storage unit (LSU) that serves L1 and shared memory requests, so only 1 out of 8 potential instructions can be sent to the LSU, so bank conflicts between skews will not occur.