#define BUFF_SIZE 0x1000
#define OFFS_QNTY 16

uint8_t src_buff[ BUFF_SIZE + OFFS_QNTY ];
uint8_t dst_buff[ BUFF_SIZE + OFFS_QNTY ];
uint32_t results[OFFS_QNTY][OFFS_QNTY];

void memcpy_test()
{
    uint8_t src_offset, dst_offset;
    uint32_t ticks_start, ticks_stop;

    CoreDebug->DEMCR |= 0x01000000;
    DWT->CYCCNT = 0;
    DWT->CTRL = 1;

    for( src_offset = 0; src_offset < OFFS_QNTY; src_offset++ )
        for( dst_offset = 0; dst_offset < OFFS_QNTY; dst_offset++ )
        {
            ticks_start = DWT->CYCCNT;
            memcpy( &dst_buff[ dst_offset ], &src_buff[ src_offset ], BUFF_SIZE );
            ticks_stop = DWT->CYCCNT;
            results[ src_offset ][ dst_offset ] = ticks_stop - ticks_start;
        }
}