CUDA共享内存效率为50％？

I have the following code that performs a tiled matrix transpose using shared memory to improve performance. The shared memory is padded with 1 column to avoid bank conflict for a 32x32 thread block.I have the following code that performs a tiled