945     add4(int32x4_t row1_low, int32x4_t row1_high, int32x4_t row2_low, int32x4_t row2_high)  in add4()  function
1097     out[1].val[0] = add4(tmp1.val[0], tmp2.val[0], tmp5.val[0], tmp6.val[0]);  in dct_body_16_32()
1098     out[1].val[1] = add4(tmp1.val[1], tmp2.val[1], tmp5.val[1], tmp6.val[1]);  in dct_body_16_32()
1100     out[15].val[0] = add4(tmp0.val[0], tmp3.val[0], tmp4.val[0], tmp7.val[0]);  in dct_body_16_32()
1101     out[15].val[1] = add4(tmp0.val[1], tmp3.val[1], tmp4.val[1], tmp7.val[1]);  in dct_body_16_32()
1108     out[3].val[0] = add4(tmp0.val[0], tmp3.val[0], tmp4.val[0], tmp7.val[0]);  in dct_body_16_32()
1109     out[3].val[1] = add4(tmp0.val[1], tmp3.val[1], tmp4.val[1], tmp7.val[1]);  in dct_body_16_32()
1111     out[13].val[0] = add4(tmp1.val[0], tmp2.val[0], tmp5.val[0], tmp6.val[0]);  in dct_body_16_32()
1112     out[13].val[1] = add4(tmp1.val[1], tmp2.val[1], tmp5.val[1], tmp6.val[1]);  in dct_body_16_32()
1119     out[5].val[0] = add4(tmp1.val[0], tmp2.val[0], tmp5.val[0], tmp6.val[0]);  in dct_body_16_32()
1120     out[5].val[1] = add4(tmp1.val[1], tmp2.val[1], tmp5.val[1], tmp6.val[1]);  in dct_body_16_32()
1122     out[11].val[0] = add4(tmp0.val[0], tmp3.val[0], tmp4.val[0], tmp7.val[0]);  in dct_body_16_32()
1123     out[11].val[1] = add4(tmp0.val[1], tmp3.val[1], tmp4.val[1], tmp7.val[1]);  in dct_body_16_32()
1130     out[7].val[0] = add4(tmp0.val[0], tmp3.val[0], tmp4.val[0], tmp7.val[0]);  in dct_body_16_32()
1131     out[7].val[1] = add4(tmp0.val[1], tmp3.val[1], tmp4.val[1], tmp7.val[1]);  in dct_body_16_32()
1133     out[9].val[0] = add4(tmp1.val[0], tmp2.val[0], tmp5.val[0], tmp6.val[0]);  in dct_body_16_32()
1134     out[9].val[1] = add4(tmp1.val[1], tmp2.val[1], tmp5.val[1], tmp6.val[1]);  in dct_body_16_32()
1195     out[1].val[0] = add4(tmp1.val[0], tmp2.val[0], tmp5.val[0], tmp6.val[0]);  in dct_body_32_32()
1196     out[1].val[1] = add4(tmp1.val[1], tmp2.val[1], tmp5.val[1], tmp6.val[1]);  in dct_body_32_32()
1198     out[15].val[0] = add4(tmp0.val[0], tmp3.val[0], tmp4.val[0], tmp7.val[0]);  in dct_body_32_32()
1199     out[15].val[1] = add4(tmp0.val[1], tmp3.val[1], tmp4.val[1], tmp7.val[1]);  in dct_body_32_32()
1206     out[3].val[0] = add4(tmp0.val[0], tmp3.val[0], tmp4.val[0], tmp7.val[0]);  in dct_body_32_32()
1207     out[3].val[1] = add4(tmp0.val[1], tmp3.val[1], tmp4.val[1], tmp7.val[1]);  in dct_body_32_32()
1209     out[13].val[0] = add4(tmp1.val[0], tmp2.val[0], tmp5.val[0], tmp6.val[0]);  in dct_body_32_32()
1210     out[13].val[1] = add4(tmp1.val[1], tmp2.val[1], tmp5.val[1], tmp6.val[1]);  in dct_body_32_32()
1217     out[5].val[0] = add4(tmp1.val[0], tmp2.val[0], tmp5.val[0], tmp6.val[0]);  in dct_body_32_32()
1218     out[5].val[1] = add4(tmp1.val[1], tmp2.val[1], tmp5.val[1], tmp6.val[1]);  in dct_body_32_32()
1220     out[11].val[0] = add4(tmp0.val[0], tmp3.val[0], tmp4.val[0], tmp7.val[0]);  in dct_body_32_32()
1221     out[11].val[1] = add4(tmp0.val[1], tmp3.val[1], tmp4.val[1], tmp7.val[1]);  in dct_body_32_32()
1228     out[7].val[0] = add4(tmp0.val[0], tmp3.val[0], tmp4.val[0], tmp7.val[0]);  in dct_body_32_32()
1229     out[7].val[1] = add4(tmp0.val[1], tmp3.val[1], tmp4.val[1], tmp7.val[1]);  in dct_body_32_32()
1231     out[9].val[0] = add4(tmp1.val[0], tmp2.val[0], tmp5.val[0], tmp6.val[0]);  in dct_body_32_32()
1232     out[9].val[1] = add4(tmp1.val[1], tmp2.val[1], tmp5.val[1], tmp6.val[1]);  in dct_body_32_32()