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);     \
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()
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()