40     out0 = (v16u8)__msa_ilvr_w((v4i32)in1, (v4i32)in0);     \
41     out1 = (v16u8)__msa_ilvl_w((v4i32)in1, (v4i32)in0);     \
42     out2 = (v16u8)__msa_ilvr_w((v4i32)in3, (v4i32)in2);     \
43     out3 = (v16u8)__msa_ilvl_w((v4i32)in3, (v4i32)in2);     \
132     res8 = (v16u8)__msa_ilvr_w((v4i32)reg4, (v4i32)reg0);  in TransposeWx16_MSA()
133     res9 = (v16u8)__msa_ilvl_w((v4i32)reg4, (v4i32)reg0);  in TransposeWx16_MSA()
137     res8 = (v16u8)__msa_ilvr_w((v4i32)reg5, (v4i32)reg1);  in TransposeWx16_MSA()
138     res9 = (v16u8)__msa_ilvl_w((v4i32)reg5, (v4i32)reg1);  in TransposeWx16_MSA()
142     res8 = (v16u8)__msa_ilvr_w((v4i32)reg6, (v4i32)reg2);  in TransposeWx16_MSA()
143     res9 = (v16u8)__msa_ilvl_w((v4i32)reg6, (v4i32)reg2);  in TransposeWx16_MSA()
147     res8 = (v16u8)__msa_ilvr_w((v4i32)reg7, (v4i32)reg3);  in TransposeWx16_MSA()
148     res9 = (v16u8)__msa_ilvl_w((v4i32)reg7, (v4i32)reg3);  in TransposeWx16_MSA()
213     res8 = (v16u8)__msa_ilvr_w((v4i32)reg4, (v4i32)reg0);  in TransposeUVWx16_MSA()
214     res9 = (v16u8)__msa_ilvl_w((v4i32)reg4, (v4i32)reg0);  in TransposeUVWx16_MSA()
220     res8 = (v16u8)__msa_ilvr_w((v4i32)reg5, (v4i32)reg1);  in TransposeUVWx16_MSA()
221     res9 = (v16u8)__msa_ilvl_w((v4i32)reg5, (v4i32)reg1);  in TransposeUVWx16_MSA()
227     res8 = (v16u8)__msa_ilvr_w((v4i32)reg6, (v4i32)reg2);  in TransposeUVWx16_MSA()
228     res9 = (v16u8)__msa_ilvl_w((v4i32)reg6, (v4i32)reg2);  in TransposeUVWx16_MSA()
234     res8 = (v16u8)__msa_ilvr_w((v4i32)reg7, (v4i32)reg3);  in TransposeUVWx16_MSA()
235     res9 = (v16u8)__msa_ilvl_w((v4i32)reg7, (v4i32)reg3);  in TransposeUVWx16_MSA()