static INLINE void write_2x4(unsigned char *dst, int pitch,
const uint8x8x2_t result) {
/*
* uint8x8x2_t result
00 01 02 03 | 04 05 06 07
10 11 12 13 | 14 15 16 17
---
* after vtrn_u8
00 10 02 12 | 04 14 06 16
01 11 03 13 | 05 15 07 17
*/
const uint8x8x2_t r01_u8 = vtrn_u8(result.val[0],
result.val[1]);
const uint16x4_t x_0_4 = vreinterpret_u16_u8(r01_u8.val[0]);
const uint16x4_t x_1_5 = vreinterpret_u16_u8(r01_u8.val[1]);
vst1_lane_u16((uint16_t *)dst, x_0_4, 0);
dst += pitch;
vst1_lane_u16((uint16_t *)dst, x_1_5, 0);
dst += pitch;
vst1_lane_u16((uint16_t *)dst, x_0_4, 1);
dst += pitch;
vst1_lane_u16((uint16_t *)dst, x_1_5, 1);
dst += pitch;
vst1_lane_u16((uint16_t *)dst, x_0_4, 2);
dst += pitch;
vst1_lane_u16((uint16_t *)dst, x_1_5, 2);
dst += pitch;
vst1_lane_u16((uint16_t *)dst, x_0_4, 3);
dst += pitch;
vst1_lane_u16((uint16_t *)dst, x_1_5, 3);
}
void test_vst1_laneu16 (void)
{
uint16_t *arg0_uint16_t;
uint16x4_t arg1_uint16x4_t;
vst1_lane_u16 (arg0_uint16_t, arg1_uint16x4_t, 1);
}