Lines Matching refs:out1

37                                    out0, out1)                              \
42 DOTP_SB2_SH(vec0_m, vec1_m, filt0, filt0, out0, out1); \
44 DPADD_SB2_SH(vec2_m, vec3_m, filt1, filt1, out0, out1); \
46 DPADD_SB2_SH(vec4_m, vec5_m, filt2, filt2, out0, out1); \
48 DPADD_SB2_SH(vec6_m, vec7_m, filt3, filt3, out0, out1); \
54 out0, out1, out2, out3) \
61 out0, out1, out2, out3); \
65 out0, out1, out2, out3); \
69 out0, out1, out2, out3); \
73 out0, out1, out2, out3); \
78 out0, out1) \
83 DOTP_SB2_SH(vec0_m, vec1_m, filt0, filt0, out0, out1); \
85 DPADD_SB2_SH(vec2_m, vec3_m, filt1, filt1, out0, out1); \
90 out0, out1, out2, out3) \
97 out0, out1, out2, out3); \
101 out0, out1, out2, out3); \
109 uint64_t out0, out1, out2, out3, out4, out5, out6, out7;
112 LD2(src, src_stride, out0, out1);
115 SD(out1, dst);
117 LD4(src, src_stride, out0, out1, out2, out3);
119 SD4(out0, out1, out2, out3, dst, dst_stride);
121 LD2(src, src_stride, out0, out1);
124 SD(out1, dst);
127 LD4(src, src_stride, out0, out1, out2, out3);
131 SD4(out0, out1, out2, out3, dst, dst_stride);
138 LD4(src, src_stride, out0, out1, out2, out3);
140 SD4(out0, out1, out2, out3, dst, dst_stride);
202 uint64_t out0, out1, out2, out3, out4, out5, out6, out7;
206 LD4(src + 16, src_stride, out0, out1, out2, out3);
212 SD4(out0, out1, out2, out3, dst + 16, dst_stride);
292 v8i16 filt, out0, out1;
308 mask3, filt0, filt1, filt2, filt3, out0, out1);
309 SRARI_H2_SH(out0, out1, 6);
310 SAT_SH2_SH(out0, out1, 7);
311 out = PCKEV_XORI128_UB(out0, out1);
322 v8i16 filt, out0, out1, out2, out3;
339 mask3, filt0, filt1, filt2, filt3, out0, out1);
344 SRARI_H4_SH(out0, out1, out2, out3, 6);
345 SAT_SH4_SH(out0, out1, out2, out3, 7);
346 out = PCKEV_XORI128_UB(out0, out1);
358 v8i16 filt, out0, out1, out2, out3;
375 mask3, filt0, filt1, filt2, filt3, out0, out1);
381 SRARI_H4_SH(out0, out1, out2, out3, 6);
382 SAT_SH4_SH(out0, out1, out2, out3, 7);
383 out = PCKEV_XORI128_UB(out0, out1);
393 mask3, filt0, filt1, filt2, filt3, out0, out1);
400 SRARI_H4_SH(out0, out1, out2, out3, 6);
401 SAT_SH4_SH(out0, out1, out2, out3, 7);
402 out = PCKEV_XORI128_UB(out0, out1);
429 v8i16 filt, out0, out1, out2, out3;
450 out0, out1, out2, out3);
454 out0, out1, out2, out3);
458 out0, out1, out2, out3);
462 out0, out1, out2, out3);
464 SRARI_H4_SH(out0, out1, out2, out3, 6);
465 SAT_SH4_SH(out0, out1, out2, out3, 7);
466 tmp0 = PCKEV_XORI128_UB(out0, out1);
483 v8i16 filt, out0, out1, out2, out3, out4, out5;
514 out1, out2, out3);
518 out1, out2, out3);
522 out1, out2, out3);
526 out1, out2, out3);
538 SRARI_H4_SH(out0, out1, out2, out3, 6);
540 SAT_SH4_SH(out0, out1, out2, out3, 7);
542 tmp0 = PCKEV_XORI128_UB(out0, out1);
560 v8i16 filt, out0, out1, out2, out3;
586 out1, out2, out3);
587 SRARI_H4_SH(out0, out1, out2, out3, 6);
588 SAT_SH4_SH(out0, out1, out2, out3, 7);
589 out = PCKEV_XORI128_UB(out0, out1);
598 out1, out2, out3);
599 SRARI_H4_SH(out0, out1, out2, out3, 6);
600 SAT_SH4_SH(out0, out1, out2, out3, 7);
601 out = PCKEV_XORI128_UB(out0, out1);
619 v8i16 out0, out1, out2, out3, out8, out9, filt;
646 DOTP_SB2_SH(vec1, vec3, filt0, filt0, out1, out3);
652 DPADD_SB2_SH(vec1, vec3, filt2, filt2, out1, out3);
658 DPADD_SB2_SH(vec5, vec7, filt1, filt1, out1, out3);
664 DPADD_SB2_SH(vec5, vec7, filt3, filt3, out1, out3);
666 SRARI_H2_SH(out1, out3, 6);
668 SAT_SH2_SH(out1, out3, 7);
671 out = PCKEV_XORI128_UB(out0, out1);
688 v8i16 filt, out0, out1, out2, out3;
718 out1, out2, out3);
719 SRARI_H4_SH(out0, out1, out2, out3, 6);
720 SAT_SH4_SH(out0, out1, out2, out3, 7);
722 out = PCKEV_XORI128_UB(out0, out1);
730 out1, out2, out3);
731 SRARI_H4_SH(out0, out1, out2, out3, 6);
732 SAT_SH4_SH(out0, out1, out2, out3, 7);
733 out = PCKEV_XORI128_UB(out0, out1);
749 v8i16 filt, out0, out1, out2, out3;
779 DOTP_SB3_SH(vec0, vec1, vec2, filt0, filt0, filt0, out0, out1, out2);
782 DPADD_SB2_SH(vec0, vec1, filt1, filt1, out0, out1);
786 DPADD_SB2_SH(vec0, vec1, filt2, filt2, out0, out1);
791 DPADD_SB2_SH(vec0, vec1, filt3, filt3, out0, out1);
794 SRARI_H2_SH(out0, out1, 6);
796 SAT_SH3_SH(out0, out1, out3, 7);
797 out = PCKEV_XORI128_UB(out0, out1);
802 DOTP_SB3_SH(vec0, vec1, vec2, filt0, filt0, filt0, out0, out1, out2);
805 DPADD_SB2_SH(vec0, vec1, filt1, filt1, out0, out1);
809 DPADD_SB2_SH(vec0, vec1, filt2, filt2, out0, out1);
813 DPADD_SB2_SH(vec0, vec1, filt3, filt3, out0, out1);
816 SRARI_H2_SH(out0, out1, 6);
818 SAT_SH3_SH(out0, out1, out2, 7);
821 out = PCKEV_XORI128_UB(out1, out2);
911 v16u8 out0, out1;
964 out1 = PCKEV_XORI128_UB(out54, out76);
965 ST_W8(out0, out1, 0, 1, 2, 3, 0, 1, 2, 3, dst, dst_stride);
1036 uint64_t out0, out1;
1096 out1 = __msa_copy_u_d((v2i64) tmp1, 0);
1102 SD(out1, dst);
1106 out1 = __msa_copy_u_d((v2i64) tmp3, 0);
1112 SD(out1, dst);
1344 v16u8 out0, out1;
1455 out1 = PCKEV_XORI128_UB(dst4_r, dst5_r);
1456 ST_W8(out0, out1, 0, 1, 2, 3, 0, 1, 2, 3, dst, dst_stride);
1630 v16u8 out0, out1;
1840 out1 = PCKEV_XORI128_UB(dst4_r, dst5_r);
1841 ST_W8(out0, out1, 0, 1, 2, 3, 0, 1, 2, 3, dst, dst_stride);
1946 v8i16 filt, out0, out1;
1961 filt0, filt1, out0, out1);
1962 SRARI_H2_SH(out0, out1, 6);
1963 SAT_SH2_SH(out0, out1, 7);
1964 out = PCKEV_XORI128_UB(out0, out1);
1974 v8i16 filt, out0, out1, out2, out3;
1990 filt0, filt1, out0, out1);
1995 SRARI_H4_SH(out0, out1, out2, out3, 6);
1996 SAT_SH4_SH(out0, out1, out2, out3, 7);
1997 out = PCKEV_XORI128_UB(out0, out1);
2010 v8i16 filt, out0, out1, out2, out3;
2025 filt0, filt1, out0, out1);
2028 SRARI_H4_SH(out0, out1, out2, out3, 6);
2029 SAT_SH4_SH(out0, out1, out2, out3, 7);
2030 out = PCKEV_XORI128_UB(out0, out1);
2040 filt0, filt1, out0, out1);
2043 SRARI_H4_SH(out0, out1, out2, out3, 6);
2044 SAT_SH4_SH(out0, out1, out2, out3, 7);
2045 out = PCKEV_XORI128_UB(out0, out1);
2072 v8i16 filt, out0, out1, out2, out3;
2088 filt1, out0, out1, out2, out3);
2089 SRARI_H4_SH(out0, out1, out2, out3, 6);
2090 SAT_SH4_SH(out0, out1, out2, out3, 7);
2091 out4 = PCKEV_XORI128_UB(out0, out1);
2104 filt1, out0, out1, out2, out3);
2105 SRARI_H4_SH(out0, out1, out2, out3, 6);
2106 SAT_SH4_SH(out0, out1, out2, out3, 7);
2107 out4 = PCKEV_XORI128_UB(out0, out1);
2156 v8i16 filt, out0, out1, out2, out3;
2173 filt1, out0, out1, out2, out3);
2174 SRARI_H4_SH(out0, out1, out2, out3, 6);
2175 SAT_SH4_SH(out0, out1, out2, out3, 7);
2176 tmp0 = PCKEV_XORI128_UB(out0, out1);
2205 v8i16 filt, out0, out1, out2, out3, out4, out5;
2225 DOTP_SB2_SH(vec0, vec1, filt0, filt0, out0, out1);
2227 DPADD_SB2_SH(vec2, vec3, filt1, filt1, out0, out1);
2228 SRARI_H2_SH(out0, out1, 6);
2229 SAT_SH2_SH(out0, out1, 7);
2230 tmp0 = PCKEV_XORI128_UB(out0, out1);
2258 v8i16 filt, out0, out1, out2, out3, out4, out5, out6, out7;
2280 out0, out1, out2, out3);
2284 out0, out1, out2, out3);
2285 SRARI_H4_SH(out0, out1, out2, out3, 6);
2286 SAT_SH4_SH(out0, out1, out2, out3, 7);
2287 out = PCKEV_XORI128_UB(out0, out1);
2322 v8i16 filt, out0, out1, out2, out3;
2347 out0, out1, out2, out3);
2349 out0, out1, out2, out3);
2350 SRARI_H4_SH(out0, out1, out2, out3, 6);
2351 SAT_SH4_SH(out0, out1, out2, out3, 7);
2352 tmp0 = PCKEV_XORI128_UB(out0, out1);
2364 out0, out1, out2, out3);
2366 out0, out1, out2, out3);
2367 SRARI_H4_SH(out0, out1, out2, out3, 6);
2368 SAT_SH4_SH(out0, out1, out2, out3, 7);
2369 tmp0 = PCKEV_XORI128_UB(out0, out1);
2383 out0, out1, out2, out3);
2385 out0, out1, out2, out3);
2387 SRARI_H4_SH(out0, out1, out2, out3, 6);
2388 SAT_SH4_SH(out0, out1, out2, out3, 7);
2389 tmp0 = PCKEV_XORI128_UB(out0, out1);
2405 v8i16 filt, out0, out1, out2, out3, out4, out5, out6, out7;
2433 out0, out1, out2, out3);
2437 out0, out1, out2, out3);
2447 SRARI_H4_SH(out0, out1, out2, out3, 6);
2449 SAT_SH4_SH(out0, out1, out2, out3, 7);
2451 out = PCKEV_XORI128_UB(out0, out1);
2557 v16u8 out0, out1;
2591 out1 = PCKEV_XORI128_UB(dst2_r, dst3_r);
2594 ST_W2(out1, 0, 2, dst + 2 * dst_stride, dst_stride);
2595 ST_H2(out1, 2, 6, dst + 2 * dst_stride + 4, dst_stride);
2617 out1 = PCKEV_XORI128_UB(dst2_r, dst3_r);
2620 ST_W2(out1, 0, 2, dst + 2 * dst_stride, dst_stride);
2621 ST_H2(out1, 2, 6, dst + 2 * dst_stride + 4, dst_stride);
2655 uint64_t out0, out1, out2;
2688 out1 = __msa_copy_u_d((v2i64) tmp0, 1);
2692 SD(out1, dst);
2768 v16u8 out0, out1;
2812 out1 = PCKEV_XORI128_UB(dst2_r, dst3_r);
2813 ST_D4(out0, out1, 0, 1, 0, 1, dst, dst_stride);
2888 uint64_t out0, out1;
2946 out1 = __msa_copy_u_d((v2i64) out3_r, 0);
2951 SD(out1, dst + 16);
3210 v16u8 out0, out1;
3289 out1 = PCKEV_XORI128_UB(tmp2, tmp3);
3290 ST_W8(out0, out1, 0, 1, 2, 3, 0, 1, 2, 3, dst, dst_stride);
3327 v16u8 out0, out1, out2;
3432 out1 = PCKEV_XORI128_UB(tmp2, tmp3);
3434 ST_W8(out0, out1, 0, 1, 2, 3, 0, 1, 2, 3, dst, dst_stride);
3509 v16u8 out0, out1;
3579 out1 = PCKEV_XORI128_UB(tmp2, tmp3);
3580 ST_D4(out0, out1, 0, 1, 0, 1, dst, dst_stride);
3592 v16u8 out0, out1, out2;
3681 out1 = PCKEV_XORI128_UB(out2_r, out3_r);
3684 ST_D4(out0, out1, 0, 1, 0, 1, dst, dst_stride);
3700 v16u8 out0, out1;
3786 out1 = PCKEV_XORI128_UB(out2_r, out3_r);
3787 ST_D4(out0, out1, 0, 1, 0, 1, dst_tmp, dst_stride);
3835 v16u8 out0, out1;
3917 out1 = PCKEV_XORI128_UB(tmp2, tmp3);
3918 ST_D4(out0, out1, 0, 1, 0, 1, dst_tmp, dst_stride);
3983 out1 = PCKEV_XORI128_UB(tmp2, tmp3);
3984 ST_W8(out0, out1, 0, 1, 2, 3, 0, 1, 2, 3, dst, dst_stride);