Why matmul CUDA is slower with shared memory than naive?

Question

I tried to understand why the optimized version take 30% more time than the normal one:

__global__
void kernel_matmul(DeviceMatrixD::In a, DeviceMatrixD::In b, De