65     p2_out = __msa_bmnz_v(p2, (v16u8)p2_filt8_r, flat);  in hz_lpf_t4_and_t8_16w()
66     p1_out = __msa_bmnz_v(p1_out, (v16u8)p1_filt8_r, flat);  in hz_lpf_t4_and_t8_16w()
67     p0_out = __msa_bmnz_v(p0_out, (v16u8)p0_filt8_r, flat);  in hz_lpf_t4_and_t8_16w()
68     q0_out = __msa_bmnz_v(q0_out, (v16u8)q0_filt8_r, flat);  in hz_lpf_t4_and_t8_16w()
69     q1_out = __msa_bmnz_v(q1_out, (v16u8)q1_filt8_r, flat);  in hz_lpf_t4_and_t8_16w()
70     q2_out = __msa_bmnz_v(q2, (v16u8)q2_filt8_r, flat);  in hz_lpf_t4_and_t8_16w()
149     p6 = __msa_bmnz_v(p6, (v16u8)r_out, flat2);  in hz_lpf_t16_16w()
169     p5 = __msa_bmnz_v(p5, (v16u8)r_out, flat2);  in hz_lpf_t16_16w()
189     p4 = __msa_bmnz_v(p4, (v16u8)r_out, flat2);  in hz_lpf_t16_16w()
209     p3 = __msa_bmnz_v(p3, (v16u8)r_out, flat2);  in hz_lpf_t16_16w()
230     filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);  in hz_lpf_t16_16w()
251     filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);  in hz_lpf_t16_16w()
272     filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);  in hz_lpf_t16_16w()
293     filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);  in hz_lpf_t16_16w()
312     filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);  in hz_lpf_t16_16w()
331     filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);  in hz_lpf_t16_16w()
349     q3 = __msa_bmnz_v(q3, (v16u8)r_out, flat2);  in hz_lpf_t16_16w()
367     q4 = __msa_bmnz_v(q4, (v16u8)r_out, flat2);  in hz_lpf_t16_16w()
385     q5 = __msa_bmnz_v(q5, (v16u8)r_out, flat2);  in hz_lpf_t16_16w()
403     q6 = __msa_bmnz_v(q6, (v16u8)r_out, flat2);  in hz_lpf_t16_16w()
479       p2_out = __msa_bmnz_v(p2, (v16u8)p2_filter8, flat);  in mb_lpf_horizontal_edge()
480       p1_out = __msa_bmnz_v(p1_out, (v16u8)p1_filter8, flat);  in mb_lpf_horizontal_edge()
481       p0_out = __msa_bmnz_v(p0_out, (v16u8)p0_filter8, flat);  in mb_lpf_horizontal_edge()
482       q0_out = __msa_bmnz_v(q0_out, (v16u8)q0_filter8, flat);  in mb_lpf_horizontal_edge()
483       q1_out = __msa_bmnz_v(q1_out, (v16u8)q1_filter8, flat);  in mb_lpf_horizontal_edge()
484       q2_out = __msa_bmnz_v(q2, (v16u8)q2_filter8, flat);  in mb_lpf_horizontal_edge()
526         p0_filter16 = __msa_bmnz_v(p6, p0_filter16, flat2);  in mb_lpf_horizontal_edge()
527         p1_filter16 = __msa_bmnz_v(p5, p1_filter16, flat2);  in mb_lpf_horizontal_edge()
544         p0_filter16 = __msa_bmnz_v(p4, p0_filter16, flat2);  in mb_lpf_horizontal_edge()
545         p1_filter16 = __msa_bmnz_v(p3, p1_filter16, flat2);  in mb_lpf_horizontal_edge()
562         p0_filter16 = __msa_bmnz_v(p2_out, p0_filter16, flat2);  in mb_lpf_horizontal_edge()
563         p1_filter16 = __msa_bmnz_v(p1_out, p1_filter16, flat2);  in mb_lpf_horizontal_edge()
580         p0_filter16 = __msa_bmnz_v(p0_out, p0_filter16, flat2);  in mb_lpf_horizontal_edge()
581         p1_filter16 = __msa_bmnz_v(q0_out, p1_filter16, flat2);  in mb_lpf_horizontal_edge()
598         p0_filter16 = __msa_bmnz_v(q1_out, p0_filter16, flat2);  in mb_lpf_horizontal_edge()
599         p1_filter16 = __msa_bmnz_v(q2_out, p1_filter16, flat2);  in mb_lpf_horizontal_edge()
616         p0_filter16 = __msa_bmnz_v(q3, p0_filter16, flat2);  in mb_lpf_horizontal_edge()
617         p1_filter16 = __msa_bmnz_v(q4, p1_filter16, flat2);  in mb_lpf_horizontal_edge()
634         p0_filter16 = __msa_bmnz_v(q5, p0_filter16, flat2);  in mb_lpf_horizontal_edge()
635         p1_filter16 = __msa_bmnz_v(q6, p1_filter16, flat2);  in mb_lpf_horizontal_edge()
808     p2_out = __msa_bmnz_v(p2, (v16u8)p2_r, flat);  in vt_lpf_t4_and_t8_8w()
809     p1_out = __msa_bmnz_v(p1_out, (v16u8)p1_r, flat);  in vt_lpf_t4_and_t8_8w()
810     p0_out = __msa_bmnz_v(p0_out, (v16u8)p0_r, flat);  in vt_lpf_t4_and_t8_8w()
811     q0_out = __msa_bmnz_v(q0_out, (v16u8)q0_r, flat);  in vt_lpf_t4_and_t8_8w()
812     q1_out = __msa_bmnz_v(q1_out, (v16u8)q1_r, flat);  in vt_lpf_t4_and_t8_8w()
813     q2_out = __msa_bmnz_v(q2, (v16u8)q2_r, flat);  in vt_lpf_t4_and_t8_8w()
882     p6 = __msa_bmnz_v(p6, (v16u8)r_out, flat2);  in vt_lpf_t16_8w()
894     p5 = __msa_bmnz_v(p5, (v16u8)r_out, flat2);  in vt_lpf_t16_8w()
906     p4 = __msa_bmnz_v(p4, (v16u8)r_out, flat2);  in vt_lpf_t16_8w()
918     p3 = __msa_bmnz_v(p3, (v16u8)r_out, flat2);  in vt_lpf_t16_8w()
931     filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);  in vt_lpf_t16_8w()
944     filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);  in vt_lpf_t16_8w()
957     filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);  in vt_lpf_t16_8w()
970     filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);  in vt_lpf_t16_8w()
982     filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);  in vt_lpf_t16_8w()
994     filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);  in vt_lpf_t16_8w()
1005     q3 = __msa_bmnz_v(q3, (v16u8)r_out, flat2);  in vt_lpf_t16_8w()
1016     q4 = __msa_bmnz_v(q4, (v16u8)r_out, flat2);  in vt_lpf_t16_8w()
1027     q5 = __msa_bmnz_v(q5, (v16u8)r_out, flat2);  in vt_lpf_t16_8w()
1038     q6 = __msa_bmnz_v(q6, (v16u8)r_out, flat2);  in vt_lpf_t16_8w()
1129     p2_out = __msa_bmnz_v(p2, (v16u8)p2_filt8_r, flat);  in vt_lpf_t4_and_t8_16w()
1130     p1_out = __msa_bmnz_v(p1_out, (v16u8)p1_filt8_r, flat);  in vt_lpf_t4_and_t8_16w()
1131     p0_out = __msa_bmnz_v(p0_out, (v16u8)p0_filt8_r, flat);  in vt_lpf_t4_and_t8_16w()
1132     q0_out = __msa_bmnz_v(q0_out, (v16u8)q0_filt8_r, flat);  in vt_lpf_t4_and_t8_16w()
1133     q1_out = __msa_bmnz_v(q1_out, (v16u8)q1_filt8_r, flat);  in vt_lpf_t4_and_t8_16w()
1134     q2_out = __msa_bmnz_v(q2, (v16u8)q2_filt8_r, flat);  in vt_lpf_t4_and_t8_16w()
1232     p6 = __msa_bmnz_v(p6, (v16u8)r_out, flat2);  in vt_lpf_t16_16w()
1250     p5 = __msa_bmnz_v(p5, (v16u8)r_out, flat2);  in vt_lpf_t16_16w()
1268     p4 = __msa_bmnz_v(p4, (v16u8)r_out, flat2);  in vt_lpf_t16_16w()
1286     p3 = __msa_bmnz_v(p3, (v16u8)r_out, flat2);  in vt_lpf_t16_16w()
1305     filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);  in vt_lpf_t16_16w()
1324     filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);  in vt_lpf_t16_16w()
1343     filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);  in vt_lpf_t16_16w()
1362     filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);  in vt_lpf_t16_16w()
1379     filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);  in vt_lpf_t16_16w()
1396     filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);  in vt_lpf_t16_16w()
1412     q3 = __msa_bmnz_v(q3, (v16u8)r_out, flat2);  in vt_lpf_t16_16w()
1428     q4 = __msa_bmnz_v(q4, (v16u8)r_out, flat2);  in vt_lpf_t16_16w()
1444     q5 = __msa_bmnz_v(q5, (v16u8)r_out, flat2);  in vt_lpf_t16_16w()
1460     q6 = __msa_bmnz_v(q6, (v16u8)r_out, flat2);  in vt_lpf_t16_16w()