#include <stdio.h>
#include <arm_neon.h>

unsigned short int a[8] = {1, 2, 3, 4, 5, 6, 7, 8};

int main()
{
    uint16x8_t va = vld1q_u16(a);
    uint16x8_t vb = vrev64q_u16(va);
    uint16_t b[8];
    vst1q_u16(b, vb);
    for (int i = 0; i < 8; i++)
        printf("%d ", b[i]);
    return 0;
}