#include <arm_neon.h>

void test(signed int *output)
{ 
	int16_t * out = (int16_t *) output;
	{
		int16x4_t a = {0, };
		int16x4_t b = {1, };
		int16x4x4_t z;
		z.val[0] = vadd_s16(a, b); 
		z.val[1] = vadd_s16(a, b); 
		z.val[2] = vsub_s16(a, b); 
		z.val[3] = vadd_s16(a, b); 
		
		vst4_lane_s16(&out[0]+0*4, z, 0);
		vst4_lane_s16(&out[8]+0*4, z, 1);
		
	}
}