24     out0 = (v16u8)__msa_ilvr_b((v16i8)in1, (v16i8)in0);     \
25     out1 = (v16u8)__msa_ilvl_b((v16i8)in1, (v16i8)in0);     \
26     out2 = (v16u8)__msa_ilvr_b((v16i8)in3, (v16i8)in2);     \
27     out3 = (v16u8)__msa_ilvl_b((v16i8)in3, (v16i8)in2);     \
32     out0 = (v16u8)__msa_ilvr_h((v8i16)in1, (v8i16)in0);     \
33     out1 = (v16u8)__msa_ilvl_h((v8i16)in1, (v8i16)in0);     \
34     out2 = (v16u8)__msa_ilvr_h((v8i16)in3, (v8i16)in2);     \
35     out3 = (v16u8)__msa_ilvl_h((v8i16)in3, (v8i16)in2);     \
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);     \
48     out0 = (v16u8)__msa_ilvr_d((v2i64)in1, (v2i64)in0);     \
49     out1 = (v16u8)__msa_ilvl_d((v2i64)in1, (v2i64)in0);     \
50     out2 = (v16u8)__msa_ilvr_d((v2i64)in3, (v2i64)in2);     \
51     out3 = (v16u8)__msa_ilvl_d((v2i64)in3, (v2i64)in2);     \
84   v16u8 src0, src1, src2, src3, dst0, dst1, dst2, dst3, vec0, vec1, vec2, vec3;  in TransposeWx16_MSA()
85   v16u8 reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7;  in TransposeWx16_MSA()
86   v16u8 res0, res1, res2, res3, res4, res5, res6, res7, res8, res9;  in TransposeWx16_MSA()
90     src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);  in TransposeWx16_MSA()
92     src1 = (v16u8)__msa_ld_b((v16i8*)s, 0);  in TransposeWx16_MSA()
94     src2 = (v16u8)__msa_ld_b((v16i8*)s, 0);  in TransposeWx16_MSA()
96     src3 = (v16u8)__msa_ld_b((v16i8*)s, 0);  in TransposeWx16_MSA()
100     src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);  in TransposeWx16_MSA()
102     src1 = (v16u8)__msa_ld_b((v16i8*)s, 0);  in TransposeWx16_MSA()
104     src2 = (v16u8)__msa_ld_b((v16i8*)s, 0);  in TransposeWx16_MSA()
106     src3 = (v16u8)__msa_ld_b((v16i8*)s, 0);  in TransposeWx16_MSA()
112     src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);  in TransposeWx16_MSA()
114     src1 = (v16u8)__msa_ld_b((v16i8*)s, 0);  in TransposeWx16_MSA()
116     src2 = (v16u8)__msa_ld_b((v16i8*)s, 0);  in TransposeWx16_MSA()
118     src3 = (v16u8)__msa_ld_b((v16i8*)s, 0);  in TransposeWx16_MSA()
122     src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);  in TransposeWx16_MSA()
124     src1 = (v16u8)__msa_ld_b((v16i8*)s, 0);  in TransposeWx16_MSA()
126     src2 = (v16u8)__msa_ld_b((v16i8*)s, 0);  in TransposeWx16_MSA()
128     src3 = (v16u8)__msa_ld_b((v16i8*)s, 0);  in TransposeWx16_MSA()
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()
165   v16u8 src0, src1, src2, src3, dst0, dst1, dst2, dst3, vec0, vec1, vec2, vec3;  in TransposeUVWx16_MSA()
166   v16u8 reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7;  in TransposeUVWx16_MSA()
167   v16u8 res0, res1, res2, res3, res4, res5, res6, res7, res8, res9;  in TransposeUVWx16_MSA()
171     src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);  in TransposeUVWx16_MSA()
173     src1 = (v16u8)__msa_ld_b((v16i8*)s, 0);  in TransposeUVWx16_MSA()
175     src2 = (v16u8)__msa_ld_b((v16i8*)s, 0);  in TransposeUVWx16_MSA()
177     src3 = (v16u8)__msa_ld_b((v16i8*)s, 0);  in TransposeUVWx16_MSA()
181     src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);  in TransposeUVWx16_MSA()
183     src1 = (v16u8)__msa_ld_b((v16i8*)s, 0);  in TransposeUVWx16_MSA()
185     src2 = (v16u8)__msa_ld_b((v16i8*)s, 0);  in TransposeUVWx16_MSA()
187     src3 = (v16u8)__msa_ld_b((v16i8*)s, 0);  in TransposeUVWx16_MSA()
193     src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);  in TransposeUVWx16_MSA()
195     src1 = (v16u8)__msa_ld_b((v16i8*)s, 0);  in TransposeUVWx16_MSA()
197     src2 = (v16u8)__msa_ld_b((v16i8*)s, 0);  in TransposeUVWx16_MSA()
199     src3 = (v16u8)__msa_ld_b((v16i8*)s, 0);  in TransposeUVWx16_MSA()
203     src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);  in TransposeUVWx16_MSA()
205     src1 = (v16u8)__msa_ld_b((v16i8*)s, 0);  in TransposeUVWx16_MSA()
207     src2 = (v16u8)__msa_ld_b((v16i8*)s, 0);  in TransposeUVWx16_MSA()
209     src3 = (v16u8)__msa_ld_b((v16i8*)s, 0);  in TransposeUVWx16_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()