在cuda中使用和不使用共享内存的矩阵乘法中经过的时间之间的差异

I am new in cuda programming. My program is matrix multiplication with and without shared memory . I use the codes in Cuda_C_Programming_Guide ebook . In book we see the program which uses shared memory has less elapsed time than the elapse time in none_shared program. But when I run the programs it’s vice versa. Does anyone know why? Or am I wrong? none-shared memory :I am new in cuda programming. My program is mat