• Home
  • Raw
  • Download

Lines Matching refs:v16u8

48     out_y = (v16u8)__msa_insert_d((v2i64)zero_m, 0, (int64)y_m); \
49 out_u = (v16u8)__msa_insert_w(zero_m, 0, (int32)u_m); \
50 out_v = (v16u8)__msa_insert_w(zero_m, 0, (int32)v_m); \
127 v16u8 dst0_m, dst1_m; \
130 dst0_m = (v16u8)__msa_ilvr_h(vec1_m, vec0_m); \
131 dst1_m = (v16u8)__msa_ilvl_h(vec1_m, vec0_m); \
139 v16u8 vec0_m, vec1_m, vec2_m, vec3_m; \
142 vec0_m = (v16u8)__msa_pckev_h((v8i16)argb1, (v8i16)argb0); \
143 vec1_m = (v16u8)__msa_pckev_h((v8i16)argb3, (v8i16)argb2); \
144 vec2_m = (v16u8)__msa_pckod_h((v8i16)argb1, (v8i16)argb0); \
145 vec3_m = (v16u8)__msa_pckod_h((v8i16)argb3, (v8i16)argb2); \
154 y_out = (v16u8)__msa_pckev_b((v16i8)reg1_m, (v16i8)reg0_m); \
160 v16u8 src0_m, src1_m, src2_m, src3_m, src4_m, src5_m, src6_m, src7_m; \
161 v16u8 vec0_m, vec1_m, vec2_m, vec3_m, vec4_m, vec5_m, vec6_m, vec7_m; \
162 v16u8 vec8_m, vec9_m; \
166 src0_m = (v16u8)__msa_ld_b((v16i8*)s, 0); \
167 src1_m = (v16u8)__msa_ld_b((v16i8*)s, 16); \
168 src2_m = (v16u8)__msa_ld_b((v16i8*)s, 32); \
169 src3_m = (v16u8)__msa_ld_b((v16i8*)s, 48); \
170 src4_m = (v16u8)__msa_ld_b((v16i8*)t, 0); \
171 src5_m = (v16u8)__msa_ld_b((v16i8*)t, 16); \
172 src6_m = (v16u8)__msa_ld_b((v16i8*)t, 32); \
173 src7_m = (v16u8)__msa_ld_b((v16i8*)t, 48); \
174 vec0_m = (v16u8)__msa_ilvr_b((v16i8)src0_m, (v16i8)src4_m); \
175 vec1_m = (v16u8)__msa_ilvr_b((v16i8)src1_m, (v16i8)src5_m); \
176 vec2_m = (v16u8)__msa_ilvr_b((v16i8)src2_m, (v16i8)src6_m); \
177 vec3_m = (v16u8)__msa_ilvr_b((v16i8)src3_m, (v16i8)src7_m); \
178 vec4_m = (v16u8)__msa_ilvl_b((v16i8)src0_m, (v16i8)src4_m); \
179 vec5_m = (v16u8)__msa_ilvl_b((v16i8)src1_m, (v16i8)src5_m); \
180 vec6_m = (v16u8)__msa_ilvl_b((v16i8)src2_m, (v16i8)src6_m); \
181 vec7_m = (v16u8)__msa_ilvl_b((v16i8)src3_m, (v16i8)src7_m); \
202 argb0 = (v16u8)__msa_pckev_b((v16i8)reg9_m, (v16i8)reg8_m); \
203 argb1 = (v16u8)__msa_pckev_b((v16i8)reg1_m, (v16i8)reg0_m); \
204 src0_m = (v16u8)__msa_ld_b((v16i8*)s, 64); \
205 src1_m = (v16u8)__msa_ld_b((v16i8*)s, 80); \
206 src2_m = (v16u8)__msa_ld_b((v16i8*)s, 96); \
207 src3_m = (v16u8)__msa_ld_b((v16i8*)s, 112); \
208 src4_m = (v16u8)__msa_ld_b((v16i8*)t, 64); \
209 src5_m = (v16u8)__msa_ld_b((v16i8*)t, 80); \
210 src6_m = (v16u8)__msa_ld_b((v16i8*)t, 96); \
211 src7_m = (v16u8)__msa_ld_b((v16i8*)t, 112); \
212 vec2_m = (v16u8)__msa_ilvr_b((v16i8)src0_m, (v16i8)src4_m); \
213 vec3_m = (v16u8)__msa_ilvr_b((v16i8)src1_m, (v16i8)src5_m); \
214 vec4_m = (v16u8)__msa_ilvr_b((v16i8)src2_m, (v16i8)src6_m); \
215 vec5_m = (v16u8)__msa_ilvr_b((v16i8)src3_m, (v16i8)src7_m); \
216 vec6_m = (v16u8)__msa_ilvl_b((v16i8)src0_m, (v16i8)src4_m); \
217 vec7_m = (v16u8)__msa_ilvl_b((v16i8)src1_m, (v16i8)src5_m); \
218 vec8_m = (v16u8)__msa_ilvl_b((v16i8)src2_m, (v16i8)src6_m); \
219 vec9_m = (v16u8)__msa_ilvl_b((v16i8)src3_m, (v16i8)src7_m); \
240 argb2 = (v16u8)__msa_pckev_b((v16i8)reg9_m, (v16i8)reg8_m); \
241 argb3 = (v16u8)__msa_pckev_b((v16i8)reg1_m, (v16i8)reg0_m); \
248 v16u8 vec0_m, vec1_m, vec2_m, vec3_m, vec4_m, vec5_m, vec6_m, vec7_m; \
251 vec0_m = (v16u8)__msa_vshf_b(shf0, (v16i8)argb1, (v16i8)argb0); \
252 vec1_m = (v16u8)__msa_vshf_b(shf0, (v16i8)argb3, (v16i8)argb2); \
253 vec2_m = (v16u8)__msa_vshf_b(shf1, (v16i8)argb1, (v16i8)argb0); \
254 vec3_m = (v16u8)__msa_vshf_b(shf1, (v16i8)argb3, (v16i8)argb2); \
255 vec4_m = (v16u8)__msa_vshf_b(shf2, (v16i8)argb1, (v16i8)argb0); \
256 vec5_m = (v16u8)__msa_vshf_b(shf2, (v16i8)argb3, (v16i8)argb2); \
257 vec6_m = (v16u8)__msa_vshf_b(shf3, (v16i8)argb1, (v16i8)argb0); \
258 vec7_m = (v16u8)__msa_vshf_b(shf3, (v16i8)argb3, (v16i8)argb2); \
271 v_out = (v16u8)__msa_pckod_b((v16i8)reg1_m, (v16i8)reg0_m); \
272 u_out = (v16u8)__msa_pckod_b((v16i8)reg3_m, (v16i8)reg2_m); \
283 out_y = (v16u8)__msa_insert_d(zero_m, 0, (int64)y_m); \
284 out_u = (v16u8)__msa_insert_d(zero_m, 0, (int64)u_m); \
285 out_v = (v16u8)__msa_insert_d(zero_m, 0, (int64)v_m); \
290 v16u8 src0, src1, src2, src3; in MirrorRow_MSA()
291 v16u8 dst0, dst1, dst2, dst3; in MirrorRow_MSA()
307 v16u8 src0, src1, src2, src3; in ARGBMirrorRow_MSA()
308 v16u8 dst0, dst1, dst2, dst3; in ARGBMirrorRow_MSA()
328 v16u8 src_u0, src_v0, src_y0, src_y1, vec_uv0, vec_uv1; in I422ToYUY2Row_MSA()
329 v16u8 dst_yuy2_0, dst_yuy2_1, dst_yuy2_2, dst_yuy2_3; in I422ToYUY2Row_MSA()
352 v16u8 src_u0, src_v0, src_y0, src_y1, vec_uv0, vec_uv1; in I422ToUYVYRow_MSA()
353 v16u8 dst_uyvy0, dst_uyvy1, dst_uyvy2, dst_uyvy3; in I422ToUYVYRow_MSA()
377 v16u8 src0, src1, src2; in I422ToARGBRow_MSA()
381 v16u8 alpha = (v16u8)__msa_ldi_b(ALPHA_VAL); in I422ToARGBRow_MSA()
390 src1 = (v16u8)__msa_ilvr_b((v16i8)src2, (v16i8)src1); in I422ToARGBRow_MSA()
408 v16u8 src0, src1, src2; in I422ToRGBARow_MSA()
412 v16u8 alpha = (v16u8)__msa_ldi_b(ALPHA_VAL); in I422ToRGBARow_MSA()
421 src1 = (v16u8)__msa_ilvr_b((v16i8)src2, (v16i8)src1); in I422ToRGBARow_MSA()
441 v16u8 src0, src1, src2, src3; in I422AlphaToARGBRow_MSA()
455 src1 = (v16u8)__msa_ilvr_b((v16i8)src2, (v16i8)src1); in I422AlphaToARGBRow_MSA()
456 src3 = (v16u8)__msa_insert_d((v2i64)zero, 0, data_a); in I422AlphaToARGBRow_MSA()
459 src3 = (v16u8)__msa_ilvr_b((v16i8)src3, (v16i8)src3); in I422AlphaToARGBRow_MSA()
477 v16u8 src0, src1, src2, src3, src4, dst0, dst1, dst2; in I422ToRGB24Row_MSA()
481 v16u8 reg0, reg1, reg2, reg3; in I422ToRGB24Row_MSA()
494 src0 = (v16u8)__msa_ld_b((v16u8*)src_y, 0); in I422ToRGB24Row_MSA()
497 src1 = (v16u8)__msa_insert_d(zero, 0, data_u); in I422ToRGB24Row_MSA()
498 src2 = (v16u8)__msa_insert_d(zero, 0, data_v); in I422ToRGB24Row_MSA()
499 src1 = (v16u8)__msa_ilvr_b((v16i8)src2, (v16i8)src1); in I422ToRGB24Row_MSA()
500 src3 = (v16u8)__msa_sldi_b((v16i8)src0, (v16i8)src0, 8); in I422ToRGB24Row_MSA()
501 src4 = (v16u8)__msa_sldi_b((v16i8)src1, (v16i8)src1, 8); in I422ToRGB24Row_MSA()
506 reg0 = (v16u8)__msa_ilvev_b((v16i8)vec1, (v16i8)vec0); in I422ToRGB24Row_MSA()
507 reg2 = (v16u8)__msa_ilvev_b((v16i8)vec4, (v16i8)vec3); in I422ToRGB24Row_MSA()
508 reg3 = (v16u8)__msa_pckev_b((v16i8)vec5, (v16i8)vec2); in I422ToRGB24Row_MSA()
509 reg1 = (v16u8)__msa_sldi_b((v16i8)reg2, (v16i8)reg0, 11); in I422ToRGB24Row_MSA()
510 dst0 = (v16u8)__msa_vshf_b(shuffler0, (v16i8)reg3, (v16i8)reg0); in I422ToRGB24Row_MSA()
511 dst1 = (v16u8)__msa_vshf_b(shuffler1, (v16i8)reg3, (v16i8)reg1); in I422ToRGB24Row_MSA()
512 dst2 = (v16u8)__msa_vshf_b(shuffler2, (v16i8)reg3, (v16i8)reg2); in I422ToRGB24Row_MSA()
530 v16u8 src0, src1, src2, dst0; in I422ToRGB565Row_MSA()
542 src1 = (v16u8)__msa_ilvr_b((v16i8)src2, (v16i8)src1); in I422ToRGB565Row_MSA()
551 dst0 = (v16u8)(vec2 | vec0); in I422ToRGB565Row_MSA()
568 v16u8 src0, src1, src2, dst0; in I422ToARGB4444Row_MSA()
582 src1 = (v16u8)__msa_ilvr_b((v16i8)src2, (v16i8)src1); in I422ToARGB4444Row_MSA()
592 dst0 = (v16u8)(reg1 | reg0); in I422ToARGB4444Row_MSA()
608 v16u8 src0, src1, src2, dst0; in I422ToARGB1555Row_MSA()
622 src1 = (v16u8)__msa_ilvr_b((v16i8)src2, (v16i8)src1); in I422ToARGB1555Row_MSA()
632 dst0 = (v16u8)(reg1 | reg0); in I422ToARGB1555Row_MSA()
643 v16u8 src0, src1, src2, src3, dst0, dst1; in YUY2ToYRow_MSA()
647 dst0 = (v16u8)__msa_pckev_b((v16i8)src1, (v16i8)src0); in YUY2ToYRow_MSA()
648 dst1 = (v16u8)__msa_pckev_b((v16i8)src3, (v16i8)src2); in YUY2ToYRow_MSA()
662 v16u8 src0, src1, src2, src3, src4, src5, src6, src7; in YUY2ToUVRow_MSA()
663 v16u8 vec0, vec1, dst0, dst1; in YUY2ToUVRow_MSA()
668 src0 = (v16u8)__msa_pckod_b((v16i8)src1, (v16i8)src0); in YUY2ToUVRow_MSA()
669 src1 = (v16u8)__msa_pckod_b((v16i8)src3, (v16i8)src2); in YUY2ToUVRow_MSA()
670 src2 = (v16u8)__msa_pckod_b((v16i8)src5, (v16i8)src4); in YUY2ToUVRow_MSA()
671 src3 = (v16u8)__msa_pckod_b((v16i8)src7, (v16i8)src6); in YUY2ToUVRow_MSA()
674 dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0); in YUY2ToUVRow_MSA()
675 dst1 = (v16u8)__msa_pckod_b((v16i8)vec1, (v16i8)vec0); in YUY2ToUVRow_MSA()
690 v16u8 src0, src1, src2, src3, dst0, dst1; in YUY2ToUV422Row_MSA()
694 src0 = (v16u8)__msa_pckod_b((v16i8)src1, (v16i8)src0); in YUY2ToUV422Row_MSA()
695 src1 = (v16u8)__msa_pckod_b((v16i8)src3, (v16i8)src2); in YUY2ToUV422Row_MSA()
696 dst0 = (v16u8)__msa_pckev_b((v16i8)src1, (v16i8)src0); in YUY2ToUV422Row_MSA()
697 dst1 = (v16u8)__msa_pckod_b((v16i8)src1, (v16i8)src0); in YUY2ToUV422Row_MSA()
708 v16u8 src0, src1, src2, src3, dst0, dst1; in UYVYToYRow_MSA()
712 dst0 = (v16u8)__msa_pckod_b((v16i8)src1, (v16i8)src0); in UYVYToYRow_MSA()
713 dst1 = (v16u8)__msa_pckod_b((v16i8)src3, (v16i8)src2); in UYVYToYRow_MSA()
727 v16u8 src0, src1, src2, src3, src4, src5, src6, src7; in UYVYToUVRow_MSA()
728 v16u8 vec0, vec1, dst0, dst1; in UYVYToUVRow_MSA()
733 src0 = (v16u8)__msa_pckev_b((v16i8)src1, (v16i8)src0); in UYVYToUVRow_MSA()
734 src1 = (v16u8)__msa_pckev_b((v16i8)src3, (v16i8)src2); in UYVYToUVRow_MSA()
735 src2 = (v16u8)__msa_pckev_b((v16i8)src5, (v16i8)src4); in UYVYToUVRow_MSA()
736 src3 = (v16u8)__msa_pckev_b((v16i8)src7, (v16i8)src6); in UYVYToUVRow_MSA()
739 dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0); in UYVYToUVRow_MSA()
740 dst1 = (v16u8)__msa_pckod_b((v16i8)vec1, (v16i8)vec0); in UYVYToUVRow_MSA()
755 v16u8 src0, src1, src2, src3, dst0, dst1; in UYVYToUV422Row_MSA()
759 src0 = (v16u8)__msa_pckev_b((v16i8)src1, (v16i8)src0); in UYVYToUV422Row_MSA()
760 src1 = (v16u8)__msa_pckev_b((v16i8)src3, (v16i8)src2); in UYVYToUV422Row_MSA()
761 dst0 = (v16u8)__msa_pckev_b((v16i8)src1, (v16i8)src0); in UYVYToUV422Row_MSA()
762 dst1 = (v16u8)__msa_pckod_b((v16i8)src1, (v16i8)src0); in UYVYToUV422Row_MSA()
773 v16u8 src0, src1, src2, src3, vec0, vec1, vec2, vec3, dst0; in ARGBToYRow_MSA()
782 src0 = (v16u8)__msa_ld_b((v16u8*)src_argb0, 0); in ARGBToYRow_MSA()
783 src1 = (v16u8)__msa_ld_b((v16u8*)src_argb0, 16); in ARGBToYRow_MSA()
784 src2 = (v16u8)__msa_ld_b((v16u8*)src_argb0, 32); in ARGBToYRow_MSA()
785 src3 = (v16u8)__msa_ld_b((v16u8*)src_argb0, 48); in ARGBToYRow_MSA()
786 vec0 = (v16u8)__msa_pckev_b((v16i8)src1, (v16i8)src0); in ARGBToYRow_MSA()
787 vec1 = (v16u8)__msa_pckev_b((v16i8)src3, (v16i8)src2); in ARGBToYRow_MSA()
788 vec2 = (v16u8)__msa_pckod_b((v16i8)src1, (v16i8)src0); in ARGBToYRow_MSA()
789 vec3 = (v16u8)__msa_pckod_b((v16i8)src3, (v16i8)src2); in ARGBToYRow_MSA()
810 dst0 = (v16u8)__msa_pckev_b((v16i8)reg1, (v16i8)reg0); in ARGBToYRow_MSA()
824 v16u8 src0, src1, src2, src3, src4, src5, src6, src7; in ARGBToUVRow_MSA()
825 v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, vec8, vec9; in ARGBToUVRow_MSA()
827 v16u8 dst0, dst1; in ARGBToUVRow_MSA()
836 src0 = (v16u8)__msa_ld_b((v16u8*)src_argb0, 0); in ARGBToUVRow_MSA()
837 src1 = (v16u8)__msa_ld_b((v16u8*)src_argb0, 16); in ARGBToUVRow_MSA()
838 src2 = (v16u8)__msa_ld_b((v16u8*)src_argb0, 32); in ARGBToUVRow_MSA()
839 src3 = (v16u8)__msa_ld_b((v16u8*)src_argb0, 48); in ARGBToUVRow_MSA()
840 src4 = (v16u8)__msa_ld_b((v16u8*)src_argb0, 64); in ARGBToUVRow_MSA()
841 src5 = (v16u8)__msa_ld_b((v16u8*)src_argb0, 80); in ARGBToUVRow_MSA()
842 src6 = (v16u8)__msa_ld_b((v16u8*)src_argb0, 96); in ARGBToUVRow_MSA()
843 src7 = (v16u8)__msa_ld_b((v16u8*)src_argb0, 112); in ARGBToUVRow_MSA()
844 vec0 = (v16u8)__msa_pckev_b((v16i8)src1, (v16i8)src0); in ARGBToUVRow_MSA()
845 vec1 = (v16u8)__msa_pckev_b((v16i8)src3, (v16i8)src2); in ARGBToUVRow_MSA()
846 vec2 = (v16u8)__msa_pckev_b((v16i8)src5, (v16i8)src4); in ARGBToUVRow_MSA()
847 vec3 = (v16u8)__msa_pckev_b((v16i8)src7, (v16i8)src6); in ARGBToUVRow_MSA()
848 vec4 = (v16u8)__msa_pckod_b((v16i8)src1, (v16i8)src0); in ARGBToUVRow_MSA()
849 vec5 = (v16u8)__msa_pckod_b((v16i8)src3, (v16i8)src2); in ARGBToUVRow_MSA()
850 vec6 = (v16u8)__msa_pckod_b((v16i8)src5, (v16i8)src4); in ARGBToUVRow_MSA()
851 vec7 = (v16u8)__msa_pckod_b((v16i8)src7, (v16i8)src6); in ARGBToUVRow_MSA()
852 vec8 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0); in ARGBToUVRow_MSA()
853 vec9 = (v16u8)__msa_pckev_b((v16i8)vec3, (v16i8)vec2); in ARGBToUVRow_MSA()
854 vec4 = (v16u8)__msa_pckev_b((v16i8)vec5, (v16i8)vec4); in ARGBToUVRow_MSA()
855 vec5 = (v16u8)__msa_pckev_b((v16i8)vec7, (v16i8)vec6); in ARGBToUVRow_MSA()
856 vec0 = (v16u8)__msa_pckod_b((v16i8)vec1, (v16i8)vec0); in ARGBToUVRow_MSA()
857 vec1 = (v16u8)__msa_pckod_b((v16i8)vec3, (v16i8)vec2); in ARGBToUVRow_MSA()
864 src0 = (v16u8)__msa_ld_b((v16u8*)src_argb0_next, 0); in ARGBToUVRow_MSA()
865 src1 = (v16u8)__msa_ld_b((v16u8*)src_argb0_next, 16); in ARGBToUVRow_MSA()
866 src2 = (v16u8)__msa_ld_b((v16u8*)src_argb0_next, 32); in ARGBToUVRow_MSA()
867 src3 = (v16u8)__msa_ld_b((v16u8*)src_argb0_next, 48); in ARGBToUVRow_MSA()
868 src4 = (v16u8)__msa_ld_b((v16u8*)src_argb0_next, 64); in ARGBToUVRow_MSA()
869 src5 = (v16u8)__msa_ld_b((v16u8*)src_argb0_next, 80); in ARGBToUVRow_MSA()
870 src6 = (v16u8)__msa_ld_b((v16u8*)src_argb0_next, 96); in ARGBToUVRow_MSA()
871 src7 = (v16u8)__msa_ld_b((v16u8*)src_argb0_next, 112); in ARGBToUVRow_MSA()
872 vec0 = (v16u8)__msa_pckev_b((v16i8)src1, (v16i8)src0); in ARGBToUVRow_MSA()
873 vec1 = (v16u8)__msa_pckev_b((v16i8)src3, (v16i8)src2); in ARGBToUVRow_MSA()
874 vec2 = (v16u8)__msa_pckev_b((v16i8)src5, (v16i8)src4); in ARGBToUVRow_MSA()
875 vec3 = (v16u8)__msa_pckev_b((v16i8)src7, (v16i8)src6); in ARGBToUVRow_MSA()
876 vec4 = (v16u8)__msa_pckod_b((v16i8)src1, (v16i8)src0); in ARGBToUVRow_MSA()
877 vec5 = (v16u8)__msa_pckod_b((v16i8)src3, (v16i8)src2); in ARGBToUVRow_MSA()
878 vec6 = (v16u8)__msa_pckod_b((v16i8)src5, (v16i8)src4); in ARGBToUVRow_MSA()
879 vec7 = (v16u8)__msa_pckod_b((v16i8)src7, (v16i8)src6); in ARGBToUVRow_MSA()
880 vec8 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0); in ARGBToUVRow_MSA()
881 vec9 = (v16u8)__msa_pckev_b((v16i8)vec3, (v16i8)vec2); in ARGBToUVRow_MSA()
882 vec4 = (v16u8)__msa_pckev_b((v16i8)vec5, (v16i8)vec4); in ARGBToUVRow_MSA()
883 vec5 = (v16u8)__msa_pckev_b((v16i8)vec7, (v16i8)vec6); in ARGBToUVRow_MSA()
884 vec0 = (v16u8)__msa_pckod_b((v16i8)vec1, (v16i8)vec0); in ARGBToUVRow_MSA()
885 vec1 = (v16u8)__msa_pckod_b((v16i8)vec3, (v16i8)vec2); in ARGBToUVRow_MSA()
924 dst0 = (v16u8)__msa_pckev_b((v16i8)reg7, (v16i8)reg6); in ARGBToUVRow_MSA()
925 dst1 = (v16u8)__msa_pckev_b((v16i8)reg5, (v16i8)reg4); in ARGBToUVRow_MSA()
937 v16u8 src0, src1, src2, src3, dst0, dst1, dst2; in ARGBToRGB24Row_MSA()
945 src0 = (v16u8)__msa_ld_b((v16i8*)src_argb, 0); in ARGBToRGB24Row_MSA()
946 src1 = (v16u8)__msa_ld_b((v16i8*)src_argb, 16); in ARGBToRGB24Row_MSA()
947 src2 = (v16u8)__msa_ld_b((v16i8*)src_argb, 32); in ARGBToRGB24Row_MSA()
948 src3 = (v16u8)__msa_ld_b((v16i8*)src_argb, 48); in ARGBToRGB24Row_MSA()
949 dst0 = (v16u8)__msa_vshf_b(shuffler0, (v16i8)src1, (v16i8)src0); in ARGBToRGB24Row_MSA()
950 dst1 = (v16u8)__msa_vshf_b(shuffler1, (v16i8)src2, (v16i8)src1); in ARGBToRGB24Row_MSA()
951 dst2 = (v16u8)__msa_vshf_b(shuffler2, (v16i8)src3, (v16i8)src2); in ARGBToRGB24Row_MSA()
961 v16u8 src0, src1, src2, src3, dst0, dst1, dst2; in ARGBToRAWRow_MSA()
969 src0 = (v16u8)__msa_ld_b((v16i8*)src_argb, 0); in ARGBToRAWRow_MSA()
970 src1 = (v16u8)__msa_ld_b((v16i8*)src_argb, 16); in ARGBToRAWRow_MSA()
971 src2 = (v16u8)__msa_ld_b((v16i8*)src_argb, 32); in ARGBToRAWRow_MSA()
972 src3 = (v16u8)__msa_ld_b((v16i8*)src_argb, 48); in ARGBToRAWRow_MSA()
973 dst0 = (v16u8)__msa_vshf_b(shuffler0, (v16i8)src1, (v16i8)src0); in ARGBToRAWRow_MSA()
974 dst1 = (v16u8)__msa_vshf_b(shuffler1, (v16i8)src2, (v16i8)src1); in ARGBToRAWRow_MSA()
975 dst2 = (v16u8)__msa_vshf_b(shuffler2, (v16i8)src3, (v16i8)src2); in ARGBToRAWRow_MSA()
985 v16u8 src0, src1, dst0; in ARGBToRGB565Row_MSA()
986 v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7; in ARGBToRGB565Row_MSA()
990 src0 = (v16u8)__msa_ld_b((v16i8*)src_argb, 0); in ARGBToRGB565Row_MSA()
991 src1 = (v16u8)__msa_ld_b((v16i8*)src_argb, 16); in ARGBToRGB565Row_MSA()
992 vec0 = (v16u8)__msa_srai_b((v16i8)src0, 3); in ARGBToRGB565Row_MSA()
993 vec1 = (v16u8)__msa_slli_b((v16i8)src0, 3); in ARGBToRGB565Row_MSA()
994 vec2 = (v16u8)__msa_srai_b((v16i8)src0, 5); in ARGBToRGB565Row_MSA()
995 vec4 = (v16u8)__msa_srai_b((v16i8)src1, 3); in ARGBToRGB565Row_MSA()
996 vec5 = (v16u8)__msa_slli_b((v16i8)src1, 3); in ARGBToRGB565Row_MSA()
997 vec6 = (v16u8)__msa_srai_b((v16i8)src1, 5); in ARGBToRGB565Row_MSA()
998 vec1 = (v16u8)__msa_sldi_b(zero, (v16i8)vec1, 1); in ARGBToRGB565Row_MSA()
999 vec2 = (v16u8)__msa_sldi_b(zero, (v16i8)vec2, 1); in ARGBToRGB565Row_MSA()
1000 vec5 = (v16u8)__msa_sldi_b(zero, (v16i8)vec5, 1); in ARGBToRGB565Row_MSA()
1001 vec6 = (v16u8)__msa_sldi_b(zero, (v16i8)vec6, 1); in ARGBToRGB565Row_MSA()
1002 vec3 = (v16u8)__msa_sldi_b(zero, (v16i8)src0, 2); in ARGBToRGB565Row_MSA()
1003 vec7 = (v16u8)__msa_sldi_b(zero, (v16i8)src1, 2); in ARGBToRGB565Row_MSA()
1008 vec0 = (v16u8)__msa_ilvev_b((v16i8)vec1, (v16i8)vec0); in ARGBToRGB565Row_MSA()
1009 vec4 = (v16u8)__msa_ilvev_b((v16i8)vec5, (v16i8)vec4); in ARGBToRGB565Row_MSA()
1010 dst0 = (v16u8)__msa_pckev_h((v8i16)vec4, (v8i16)vec0); in ARGBToRGB565Row_MSA()
1019 v16u8 src0, src1, dst0; in ARGBToARGB1555Row_MSA()
1020 v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, vec8, vec9; in ARGBToARGB1555Row_MSA()
1024 src0 = (v16u8)__msa_ld_b((v16i8*)src_argb, 0); in ARGBToARGB1555Row_MSA()
1025 src1 = (v16u8)__msa_ld_b((v16i8*)src_argb, 16); in ARGBToARGB1555Row_MSA()
1026 vec0 = (v16u8)__msa_srai_b((v16i8)src0, 3); in ARGBToARGB1555Row_MSA()
1027 vec1 = (v16u8)__msa_slli_b((v16i8)src0, 2); in ARGBToARGB1555Row_MSA()
1028 vec2 = (v16u8)__msa_srai_b((v16i8)vec0, 3); in ARGBToARGB1555Row_MSA()
1029 vec1 = (v16u8)__msa_sldi_b(zero, (v16i8)vec1, 1); in ARGBToARGB1555Row_MSA()
1030 vec2 = (v16u8)__msa_sldi_b(zero, (v16i8)vec2, 1); in ARGBToARGB1555Row_MSA()
1031 vec3 = (v16u8)__msa_srai_b((v16i8)src0, 1); in ARGBToARGB1555Row_MSA()
1032 vec5 = (v16u8)__msa_srai_b((v16i8)src1, 3); in ARGBToARGB1555Row_MSA()
1033 vec6 = (v16u8)__msa_slli_b((v16i8)src1, 2); in ARGBToARGB1555Row_MSA()
1034 vec7 = (v16u8)__msa_srai_b((v16i8)vec5, 3); in ARGBToARGB1555Row_MSA()
1035 vec6 = (v16u8)__msa_sldi_b(zero, (v16i8)vec6, 1); in ARGBToARGB1555Row_MSA()
1036 vec7 = (v16u8)__msa_sldi_b(zero, (v16i8)vec7, 1); in ARGBToARGB1555Row_MSA()
1037 vec8 = (v16u8)__msa_srai_b((v16i8)src1, 1); in ARGBToARGB1555Row_MSA()
1038 vec3 = (v16u8)__msa_sldi_b(zero, (v16i8)vec3, 2); in ARGBToARGB1555Row_MSA()
1039 vec8 = (v16u8)__msa_sldi_b(zero, (v16i8)vec8, 2); in ARGBToARGB1555Row_MSA()
1040 vec4 = (v16u8)__msa_sldi_b(zero, (v16i8)src0, 3); in ARGBToARGB1555Row_MSA()
1041 vec9 = (v16u8)__msa_sldi_b(zero, (v16i8)src1, 3); in ARGBToARGB1555Row_MSA()
1048 vec0 = (v16u8)__msa_ilvev_b((v16i8)vec1, (v16i8)vec0); in ARGBToARGB1555Row_MSA()
1049 vec1 = (v16u8)__msa_ilvev_b((v16i8)vec6, (v16i8)vec5); in ARGBToARGB1555Row_MSA()
1050 dst0 = (v16u8)__msa_pckev_h((v8i16)vec1, (v8i16)vec0); in ARGBToARGB1555Row_MSA()
1059 v16u8 src0, src1; in ARGBToARGB4444Row_MSA()
1060 v16u8 vec0, vec1; in ARGBToARGB4444Row_MSA()
1061 v16u8 dst0; in ARGBToARGB4444Row_MSA()
1065 src0 = (v16u8)__msa_ld_b((v16i8*)src_argb, 0); in ARGBToARGB4444Row_MSA()
1066 src1 = (v16u8)__msa_ld_b((v16i8*)src_argb, 16); in ARGBToARGB4444Row_MSA()
1067 vec0 = (v16u8)__msa_srai_b((v16i8)src0, 4); in ARGBToARGB4444Row_MSA()
1068 vec1 = (v16u8)__msa_srai_b((v16i8)src1, 4); in ARGBToARGB4444Row_MSA()
1069 src0 = (v16u8)__msa_sldi_b(zero, (v16i8)src0, 1); in ARGBToARGB4444Row_MSA()
1070 src1 = (v16u8)__msa_sldi_b(zero, (v16i8)src1, 1); in ARGBToARGB4444Row_MSA()
1073 dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0); in ARGBToARGB4444Row_MSA()
1085 v16u8 src0, src1, src2, src3, reg0, reg1, reg2, reg3, dst0, dst1; in ARGBToUV444Row_MSA()
1097 src0 = (v16u8)__msa_ld_b((v16i8*)src_argb, 0); in ARGBToUV444Row_MSA()
1098 src1 = (v16u8)__msa_ld_b((v16i8*)src_argb, 16); in ARGBToUV444Row_MSA()
1099 src2 = (v16u8)__msa_ld_b((v16i8*)src_argb, 32); in ARGBToUV444Row_MSA()
1100 src3 = (v16u8)__msa_ld_b((v16i8*)src_argb, 48); in ARGBToUV444Row_MSA()
1101 reg0 = (v16u8)__msa_pckev_b((v16i8)src1, (v16i8)src0); in ARGBToUV444Row_MSA()
1102 reg1 = (v16u8)__msa_pckev_b((v16i8)src3, (v16i8)src2); in ARGBToUV444Row_MSA()
1103 reg2 = (v16u8)__msa_pckod_b((v16i8)src1, (v16i8)src0); in ARGBToUV444Row_MSA()
1104 reg3 = (v16u8)__msa_pckod_b((v16i8)src3, (v16i8)src2); in ARGBToUV444Row_MSA()
1105 src0 = (v16u8)__msa_pckev_b((v16i8)reg1, (v16i8)reg0); in ARGBToUV444Row_MSA()
1106 src1 = (v16u8)__msa_pckev_b((v16i8)reg3, (v16i8)reg2); in ARGBToUV444Row_MSA()
1107 src2 = (v16u8)__msa_pckod_b((v16i8)reg1, (v16i8)reg0); in ARGBToUV444Row_MSA()
1142 dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0); in ARGBToUV444Row_MSA()
1143 dst1 = (v16u8)__msa_pckev_b((v16i8)vec7, (v16i8)vec6); in ARGBToUV444Row_MSA()
1157 v16u8 src0, src1, dst0; in ARGBMultiplyRow_MSA()
1163 src0 = (v16u8)__msa_ld_b((v16i8*)src_argb0, 0); in ARGBMultiplyRow_MSA()
1164 src1 = (v16u8)__msa_ld_b((v16i8*)src_argb1, 0); in ARGBMultiplyRow_MSA()
1183 dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0); in ARGBMultiplyRow_MSA()
1196 v16u8 src0, src1, src2, src3, dst0, dst1; in ARGBAddRow_MSA()
1199 src0 = (v16u8)__msa_ld_b((v16i8*)src_argb0, 0); in ARGBAddRow_MSA()
1200 src1 = (v16u8)__msa_ld_b((v16i8*)src_argb0, 16); in ARGBAddRow_MSA()
1201 src2 = (v16u8)__msa_ld_b((v16i8*)src_argb1, 0); in ARGBAddRow_MSA()
1202 src3 = (v16u8)__msa_ld_b((v16i8*)src_argb1, 16); in ARGBAddRow_MSA()
1217 v16u8 src0, src1, src2, src3, dst0, dst1; in ARGBSubtractRow_MSA()
1220 src0 = (v16u8)__msa_ld_b((v16i8*)src_argb0, 0); in ARGBSubtractRow_MSA()
1221 src1 = (v16u8)__msa_ld_b((v16i8*)src_argb0, 16); in ARGBSubtractRow_MSA()
1222 src2 = (v16u8)__msa_ld_b((v16i8*)src_argb1, 0); in ARGBSubtractRow_MSA()
1223 src3 = (v16u8)__msa_ld_b((v16i8*)src_argb1, 16); in ARGBSubtractRow_MSA()
1235 v16u8 src0, src1, dst0, dst1; in ARGBAttenuateRow_MSA()
1239 v16u8 mask = {0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255}; in ARGBAttenuateRow_MSA()
1242 src0 = (v16u8)__msa_ld_b((v16i8*)src_argb, 0); in ARGBAttenuateRow_MSA()
1243 src1 = (v16u8)__msa_ld_b((v16i8*)src_argb, 16); in ARGBAttenuateRow_MSA()
1288 dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0); in ARGBAttenuateRow_MSA()
1289 dst1 = (v16u8)__msa_pckev_b((v16i8)vec3, (v16i8)vec2); in ARGBAttenuateRow_MSA()
1303 v16u8 src0, src1, dst0, vec0, vec1; in ARGBToRGB565DitherRow_MSA()
1313 src0 = (v16u8)__msa_ld_b((v16i8*)src_argb, 0); in ARGBToRGB565DitherRow_MSA()
1314 src1 = (v16u8)__msa_ld_b((v16i8*)src_argb, 16); in ARGBToRGB565DitherRow_MSA()
1315 vec0 = (v16u8)__msa_pckev_b((v16i8)src1, (v16i8)src0); in ARGBToRGB565DitherRow_MSA()
1316 vec1 = (v16u8)__msa_pckod_b((v16i8)src1, (v16i8)src0); in ARGBToRGB565DitherRow_MSA()
1335 dst0 = (v16u8)(reg0 | reg2); in ARGBToRGB565DitherRow_MSA()
1347 v16u8 src0, src1, dst0, dst1; in ARGBShuffleRow_MSA()
1356 src0 = (v16u8)__msa_ld_b((v16u8*)src_argb, 0); in ARGBShuffleRow_MSA()
1357 src1 = (v16u8)__msa_ld_b((v16u8*)src_argb, 16); in ARGBShuffleRow_MSA()
1358 dst0 = (v16u8)__msa_vshf_b(shuffler_vec, (v16i8)src0, (v16i8)src0); in ARGBShuffleRow_MSA()
1359 dst1 = (v16u8)__msa_vshf_b(shuffler_vec, (v16i8)src1, (v16i8)src1); in ARGBShuffleRow_MSA()
1371 v16u8 src0, dst0; in ARGBShadeRow_MSA()
1381 src0 = (v16u8)__msa_ld_b((v16u8*)src_argb, 0); in ARGBShadeRow_MSA()
1398 dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0); in ARGBShadeRow_MSA()
1407 v16u8 src0, src1, vec0, vec1, dst0, dst1; in ARGBGrayRow_MSA()
1409 v16u8 const_0x26 = (v16u8)__msa_ldi_h(0x26); in ARGBGrayRow_MSA()
1410 v16u8 const_0x4B0F = (v16u8)__msa_fill_h(0x4B0F); in ARGBGrayRow_MSA()
1413 src0 = (v16u8)__msa_ld_b((v16u8*)src_argb, 0); in ARGBGrayRow_MSA()
1414 src1 = (v16u8)__msa_ld_b((v16u8*)src_argb, 16); in ARGBGrayRow_MSA()
1415 vec0 = (v16u8)__msa_pckev_h((v8i16)src1, (v8i16)src0); in ARGBGrayRow_MSA()
1416 vec1 = (v16u8)__msa_pckod_h((v8i16)src1, (v8i16)src0); in ARGBGrayRow_MSA()
1420 vec0 = (v16u8)__msa_ilvev_b((v16i8)reg0, (v16i8)reg0); in ARGBGrayRow_MSA()
1421 vec1 = (v16u8)__msa_ilvod_b((v16i8)vec1, (v16i8)vec0); in ARGBGrayRow_MSA()
1422 dst0 = (v16u8)__msa_ilvr_b((v16i8)vec1, (v16i8)vec0); in ARGBGrayRow_MSA()
1423 dst1 = (v16u8)__msa_ilvl_b((v16i8)vec1, (v16i8)vec0); in ARGBGrayRow_MSA()
1432 v16u8 src0, src1, dst0, dst1, vec0, vec1, vec2, vec3, vec4, vec5; in ARGBSepiaRow_MSA()
1434 v16u8 const_0x4411 = (v16u8)__msa_fill_h(0x4411); in ARGBSepiaRow_MSA()
1435 v16u8 const_0x23 = (v16u8)__msa_ldi_h(0x23); in ARGBSepiaRow_MSA()
1436 v16u8 const_0x5816 = (v16u8)__msa_fill_h(0x5816); in ARGBSepiaRow_MSA()
1437 v16u8 const_0x2D = (v16u8)__msa_ldi_h(0x2D); in ARGBSepiaRow_MSA()
1438 v16u8 const_0x6218 = (v16u8)__msa_fill_h(0x6218); in ARGBSepiaRow_MSA()
1439 v16u8 const_0x32 = (v16u8)__msa_ldi_h(0x32); in ARGBSepiaRow_MSA()
1443 src0 = (v16u8)__msa_ld_b((v16u8*)dst_argb, 0); in ARGBSepiaRow_MSA()
1444 src1 = (v16u8)__msa_ld_b((v16u8*)dst_argb, 16); in ARGBSepiaRow_MSA()
1445 vec0 = (v16u8)__msa_pckev_h((v8i16)src1, (v8i16)src0); in ARGBSepiaRow_MSA()
1446 vec1 = (v16u8)__msa_pckod_h((v8i16)src1, (v8i16)src0); in ARGBSepiaRow_MSA()
1447 vec3 = (v16u8)__msa_pckod_b((v16i8)vec1, (v16i8)vec1); in ARGBSepiaRow_MSA()
1459 vec0 = (v16u8)__msa_pckev_b((v16i8)reg0, (v16i8)reg0); in ARGBSepiaRow_MSA()
1460 vec1 = (v16u8)__msa_pckev_b((v16i8)reg1, (v16i8)reg1); in ARGBSepiaRow_MSA()
1461 vec2 = (v16u8)__msa_pckev_b((v16i8)reg2, (v16i8)reg2); in ARGBSepiaRow_MSA()
1462 vec4 = (v16u8)__msa_ilvr_b((v16i8)vec2, (v16i8)vec0); in ARGBSepiaRow_MSA()
1463 vec5 = (v16u8)__msa_ilvr_b((v16i8)vec3, (v16i8)vec1); in ARGBSepiaRow_MSA()
1464 dst0 = (v16u8)__msa_ilvr_b((v16i8)vec5, (v16i8)vec4); in ARGBSepiaRow_MSA()
1465 dst1 = (v16u8)__msa_ilvl_b((v16i8)vec5, (v16i8)vec4); in ARGBSepiaRow_MSA()
1475 v16u8 src0, src1; in ARGB4444ToARGBRow_MSA()
1477 v16u8 dst0, dst1, dst2, dst3; in ARGB4444ToARGBRow_MSA()
1480 src0 = (v16u8)__msa_ld_b((v16u8*)src_argb4444, 0); in ARGB4444ToARGBRow_MSA()
1481 src1 = (v16u8)__msa_ld_b((v16u8*)src_argb4444, 16); in ARGB4444ToARGBRow_MSA()
1490 dst0 = (v16u8)__msa_ilvr_b((v16i8)vec2, (v16i8)vec0); in ARGB4444ToARGBRow_MSA()
1491 dst1 = (v16u8)__msa_ilvl_b((v16i8)vec2, (v16i8)vec0); in ARGB4444ToARGBRow_MSA()
1492 dst2 = (v16u8)__msa_ilvr_b((v16i8)vec3, (v16i8)vec1); in ARGB4444ToARGBRow_MSA()
1493 dst3 = (v16u8)__msa_ilvl_b((v16i8)vec3, (v16i8)vec1); in ARGB4444ToARGBRow_MSA()
1506 v16u8 reg0, reg1, reg2, reg3, reg4, reg5, reg6; in ARGB1555ToARGBRow_MSA()
1507 v16u8 dst0, dst1, dst2, dst3; in ARGB1555ToARGBRow_MSA()
1525 reg0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0); in ARGB1555ToARGBRow_MSA()
1526 reg1 = (v16u8)__msa_pckev_b((v16i8)vec3, (v16i8)vec2); in ARGB1555ToARGBRow_MSA()
1527 reg2 = (v16u8)__msa_pckev_b((v16i8)vec5, (v16i8)vec4); in ARGB1555ToARGBRow_MSA()
1528 reg3 = (v16u8)__msa_pckev_b((v16i8)src1, (v16i8)src0); in ARGB1555ToARGBRow_MSA()
1529 reg4 = (v16u8)__msa_slli_b((v16i8)reg0, 3); in ARGB1555ToARGBRow_MSA()
1530 reg5 = (v16u8)__msa_slli_b((v16i8)reg1, 3); in ARGB1555ToARGBRow_MSA()
1531 reg6 = (v16u8)__msa_slli_b((v16i8)reg2, 3); in ARGB1555ToARGBRow_MSA()
1532 reg4 |= (v16u8)__msa_srai_b((v16i8)reg0, 2); in ARGB1555ToARGBRow_MSA()
1533 reg5 |= (v16u8)__msa_srai_b((v16i8)reg1, 2); in ARGB1555ToARGBRow_MSA()
1534 reg6 |= (v16u8)__msa_srai_b((v16i8)reg2, 2); in ARGB1555ToARGBRow_MSA()
1536 reg0 = (v16u8)__msa_ilvr_b((v16i8)reg6, (v16i8)reg4); in ARGB1555ToARGBRow_MSA()
1537 reg1 = (v16u8)__msa_ilvl_b((v16i8)reg6, (v16i8)reg4); in ARGB1555ToARGBRow_MSA()
1538 reg2 = (v16u8)__msa_ilvr_b((v16i8)reg3, (v16i8)reg5); in ARGB1555ToARGBRow_MSA()
1539 reg3 = (v16u8)__msa_ilvl_b((v16i8)reg3, (v16i8)reg5); in ARGB1555ToARGBRow_MSA()
1540 dst0 = (v16u8)__msa_ilvr_b((v16i8)reg2, (v16i8)reg0); in ARGB1555ToARGBRow_MSA()
1541 dst1 = (v16u8)__msa_ilvl_b((v16i8)reg2, (v16i8)reg0); in ARGB1555ToARGBRow_MSA()
1542 dst2 = (v16u8)__msa_ilvr_b((v16i8)reg3, (v16i8)reg1); in ARGB1555ToARGBRow_MSA()
1543 dst3 = (v16u8)__msa_ilvl_b((v16i8)reg3, (v16i8)reg1); in ARGB1555ToARGBRow_MSA()
1554 v16u8 res0, res1, res2, res3, dst0, dst1, dst2, dst3; in RGB565ToARGBRow_MSA()
1555 v16u8 alpha = (v16u8)__msa_ldi_b(ALPHA_VAL); in RGB565ToARGBRow_MSA()
1581 res0 = (v16u8)__msa_ilvev_b((v16i8)reg2, (v16i8)reg0); in RGB565ToARGBRow_MSA()
1582 res1 = (v16u8)__msa_ilvev_b((v16i8)alpha, (v16i8)reg1); in RGB565ToARGBRow_MSA()
1583 res2 = (v16u8)__msa_ilvev_b((v16i8)reg5, (v16i8)reg3); in RGB565ToARGBRow_MSA()
1584 res3 = (v16u8)__msa_ilvev_b((v16i8)alpha, (v16i8)reg4); in RGB565ToARGBRow_MSA()
1585 dst0 = (v16u8)__msa_ilvr_b((v16i8)res1, (v16i8)res0); in RGB565ToARGBRow_MSA()
1586 dst1 = (v16u8)__msa_ilvl_b((v16i8)res1, (v16i8)res0); in RGB565ToARGBRow_MSA()
1587 dst2 = (v16u8)__msa_ilvr_b((v16i8)res3, (v16i8)res2); in RGB565ToARGBRow_MSA()
1588 dst3 = (v16u8)__msa_ilvl_b((v16i8)res3, (v16i8)res2); in RGB565ToARGBRow_MSA()
1597 v16u8 src0, src1, src2; in RGB24ToARGBRow_MSA()
1598 v16u8 vec0, vec1, vec2; in RGB24ToARGBRow_MSA()
1599 v16u8 dst0, dst1, dst2, dst3; in RGB24ToARGBRow_MSA()
1600 v16u8 alpha = (v16u8)__msa_ldi_b(ALPHA_VAL); in RGB24ToARGBRow_MSA()
1604 src0 = (v16u8)__msa_ld_b((v16i8*)src_rgb24, 0); in RGB24ToARGBRow_MSA()
1605 src1 = (v16u8)__msa_ld_b((v16i8*)src_rgb24, 16); in RGB24ToARGBRow_MSA()
1606 src2 = (v16u8)__msa_ld_b((v16i8*)src_rgb24, 32); in RGB24ToARGBRow_MSA()
1607 vec0 = (v16u8)__msa_sldi_b((v16i8)src1, (v16i8)src0, 12); in RGB24ToARGBRow_MSA()
1608 vec1 = (v16u8)__msa_sldi_b((v16i8)src2, (v16i8)src1, 8); in RGB24ToARGBRow_MSA()
1609 vec2 = (v16u8)__msa_sldi_b((v16i8)src2, (v16i8)src2, 4); in RGB24ToARGBRow_MSA()
1610 dst0 = (v16u8)__msa_vshf_b(shuffler, (v16i8)alpha, (v16i8)src0); in RGB24ToARGBRow_MSA()
1611 dst1 = (v16u8)__msa_vshf_b(shuffler, (v16i8)alpha, (v16i8)vec0); in RGB24ToARGBRow_MSA()
1612 dst2 = (v16u8)__msa_vshf_b(shuffler, (v16i8)alpha, (v16i8)vec1); in RGB24ToARGBRow_MSA()
1613 dst3 = (v16u8)__msa_vshf_b(shuffler, (v16i8)alpha, (v16i8)vec2); in RGB24ToARGBRow_MSA()
1622 v16u8 src0, src1, src2; in RAWToARGBRow_MSA()
1623 v16u8 vec0, vec1, vec2; in RAWToARGBRow_MSA()
1624 v16u8 dst0, dst1, dst2, dst3; in RAWToARGBRow_MSA()
1625 v16u8 alpha = (v16u8)__msa_ldi_b(ALPHA_VAL); in RAWToARGBRow_MSA()
1629 src0 = (v16u8)__msa_ld_b((v16i8*)src_raw, 0); in RAWToARGBRow_MSA()
1630 src1 = (v16u8)__msa_ld_b((v16i8*)src_raw, 16); in RAWToARGBRow_MSA()
1631 src2 = (v16u8)__msa_ld_b((v16i8*)src_raw, 32); in RAWToARGBRow_MSA()
1632 vec0 = (v16u8)__msa_sldi_b((v16i8)src1, (v16i8)src0, 12); in RAWToARGBRow_MSA()
1633 vec1 = (v16u8)__msa_sldi_b((v16i8)src2, (v16i8)src1, 8); in RAWToARGBRow_MSA()
1634 vec2 = (v16u8)__msa_sldi_b((v16i8)src2, (v16i8)src2, 4); in RAWToARGBRow_MSA()
1635 dst0 = (v16u8)__msa_vshf_b(mask, (v16i8)alpha, (v16i8)src0); in RAWToARGBRow_MSA()
1636 dst1 = (v16u8)__msa_vshf_b(mask, (v16i8)alpha, (v16i8)vec0); in RAWToARGBRow_MSA()
1637 dst2 = (v16u8)__msa_vshf_b(mask, (v16i8)alpha, (v16i8)vec1); in RAWToARGBRow_MSA()
1638 dst3 = (v16u8)__msa_vshf_b(mask, (v16i8)alpha, (v16i8)vec2); in RAWToARGBRow_MSA()
1649 v16u8 dst0; in ARGB1555ToYRow_MSA()
1695 dst0 = (v16u8)__msa_pckev_b((v16i8)reg1, (v16i8)reg0); in ARGB1555ToYRow_MSA()
1707 v16u8 dst0; in RGB565ToYRow_MSA()
1758 dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0); in RGB565ToYRow_MSA()
1767 v16u8 src0, src1, src2, reg0, reg1, reg2, reg3, dst0; in RGB24ToYRow_MSA()
1780 src0 = (v16u8)__msa_ld_b((v16i8*)src_argb0, 0); in RGB24ToYRow_MSA()
1781 src1 = (v16u8)__msa_ld_b((v16i8*)src_argb0, 16); in RGB24ToYRow_MSA()
1782 src2 = (v16u8)__msa_ld_b((v16i8*)src_argb0, 32); in RGB24ToYRow_MSA()
1783 reg0 = (v16u8)__msa_vshf_b(mask0, zero, (v16i8)src0); in RGB24ToYRow_MSA()
1784 reg1 = (v16u8)__msa_vshf_b(mask1, (v16i8)src1, (v16i8)src0); in RGB24ToYRow_MSA()
1785 reg2 = (v16u8)__msa_vshf_b(mask2, (v16i8)src2, (v16i8)src1); in RGB24ToYRow_MSA()
1786 reg3 = (v16u8)__msa_vshf_b(mask3, zero, (v16i8)src2); in RGB24ToYRow_MSA()
1791 vec0 = __msa_dotp_u_h((v16u8)vec0, (v16u8)const_0x8119); in RGB24ToYRow_MSA()
1792 vec1 = __msa_dotp_u_h((v16u8)vec1, (v16u8)const_0x8119); in RGB24ToYRow_MSA()
1793 vec0 = __msa_dpadd_u_h(vec0, (v16u8)vec2, (v16u8)const_0x42); in RGB24ToYRow_MSA()
1794 vec1 = __msa_dpadd_u_h(vec1, (v16u8)vec3, (v16u8)const_0x42); in RGB24ToYRow_MSA()
1799 dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0); in RGB24ToYRow_MSA()
1808 v16u8 src0, src1, src2, reg0, reg1, reg2, reg3, dst0; in RAWToYRow_MSA()
1821 src0 = (v16u8)__msa_ld_b((v16i8*)src_argb0, 0); in RAWToYRow_MSA()
1822 src1 = (v16u8)__msa_ld_b((v16i8*)src_argb0, 16); in RAWToYRow_MSA()
1823 src2 = (v16u8)__msa_ld_b((v16i8*)src_argb0, 32); in RAWToYRow_MSA()
1824 reg0 = (v16u8)__msa_vshf_b(mask0, zero, (v16i8)src0); in RAWToYRow_MSA()
1825 reg1 = (v16u8)__msa_vshf_b(mask1, (v16i8)src1, (v16i8)src0); in RAWToYRow_MSA()
1826 reg2 = (v16u8)__msa_vshf_b(mask2, (v16i8)src2, (v16i8)src1); in RAWToYRow_MSA()
1827 reg3 = (v16u8)__msa_vshf_b(mask3, zero, (v16i8)src2); in RAWToYRow_MSA()
1832 vec0 = __msa_dotp_u_h((v16u8)vec0, (v16u8)const_0x8142); in RAWToYRow_MSA()
1833 vec1 = __msa_dotp_u_h((v16u8)vec1, (v16u8)const_0x8142); in RAWToYRow_MSA()
1834 vec0 = __msa_dpadd_u_h(vec0, (v16u8)vec2, (v16u8)const_0x19); in RAWToYRow_MSA()
1835 vec1 = __msa_dpadd_u_h(vec1, (v16u8)vec3, (v16u8)const_0x19); in RAWToYRow_MSA()
1840 dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0); in RAWToYRow_MSA()
1858 v16u8 dst0; in ARGB1555ToUVRow_MSA()
1895 vec0 = __msa_hadd_u_h((v16u8)vec0, (v16u8)vec0); in ARGB1555ToUVRow_MSA()
1896 vec2 = __msa_hadd_u_h((v16u8)vec2, (v16u8)vec2); in ARGB1555ToUVRow_MSA()
1897 vec4 = __msa_hadd_u_h((v16u8)vec4, (v16u8)vec4); in ARGB1555ToUVRow_MSA()
1916 dst0 = (v16u8)__msa_pckev_b((v16i8)reg2, (v16i8)reg0); in ARGB1555ToUVRow_MSA()
1939 v16u8 dst0; in RGB565ToUVRow_MSA()
1977 vec0 = __msa_hadd_u_h((v16u8)vec0, (v16u8)vec0); in RGB565ToUVRow_MSA()
1978 vec1 = __msa_hadd_u_h((v16u8)vec1, (v16u8)vec1); in RGB565ToUVRow_MSA()
1979 vec2 = __msa_hadd_u_h((v16u8)vec2, (v16u8)vec2); in RGB565ToUVRow_MSA()
1996 dst0 = (v16u8)__msa_pckev_b((v16i8)reg2, (v16i8)reg0); in RGB565ToUVRow_MSA()
2017 v16u8 src0, src1, src2, src3, src4, src5, src6, src7; in RGB24ToUVRow_MSA()
2018 v16u8 inp0, inp1, inp2, inp3, inp4, inp5; in RGB24ToUVRow_MSA()
2021 v16u8 dst0; in RGB24ToUVRow_MSA()
2032 inp0 = (v16u8)__msa_ld_b((v16i8*)s, 0); in RGB24ToUVRow_MSA()
2033 inp1 = (v16u8)__msa_ld_b((v16i8*)s, 16); in RGB24ToUVRow_MSA()
2034 inp2 = (v16u8)__msa_ld_b((v16i8*)s, 32); in RGB24ToUVRow_MSA()
2035 inp3 = (v16u8)__msa_ld_b((v16i8*)t, 0); in RGB24ToUVRow_MSA()
2036 inp4 = (v16u8)__msa_ld_b((v16i8*)t, 16); in RGB24ToUVRow_MSA()
2037 inp5 = (v16u8)__msa_ld_b((v16i8*)t, 32); in RGB24ToUVRow_MSA()
2038 src1 = (v16u8)__msa_sldi_b((v16i8)inp1, (v16i8)inp0, 12); in RGB24ToUVRow_MSA()
2039 src5 = (v16u8)__msa_sldi_b((v16i8)inp4, (v16i8)inp3, 12); in RGB24ToUVRow_MSA()
2040 src2 = (v16u8)__msa_sldi_b((v16i8)inp2, (v16i8)inp1, 8); in RGB24ToUVRow_MSA()
2041 src6 = (v16u8)__msa_sldi_b((v16i8)inp5, (v16i8)inp4, 8); in RGB24ToUVRow_MSA()
2042 src3 = (v16u8)__msa_sldi_b((v16i8)inp2, (v16i8)inp2, 4); in RGB24ToUVRow_MSA()
2043 src7 = (v16u8)__msa_sldi_b((v16i8)inp5, (v16i8)inp5, 4); in RGB24ToUVRow_MSA()
2044 src0 = (v16u8)__msa_vshf_b(mask, (v16i8)zero, (v16i8)inp0); in RGB24ToUVRow_MSA()
2045 src1 = (v16u8)__msa_vshf_b(mask, (v16i8)zero, (v16i8)src1); in RGB24ToUVRow_MSA()
2046 src2 = (v16u8)__msa_vshf_b(mask, (v16i8)zero, (v16i8)src2); in RGB24ToUVRow_MSA()
2047 src3 = (v16u8)__msa_vshf_b(mask, (v16i8)zero, (v16i8)src3); in RGB24ToUVRow_MSA()
2048 src4 = (v16u8)__msa_vshf_b(mask, (v16i8)zero, (v16i8)inp3); in RGB24ToUVRow_MSA()
2049 src5 = (v16u8)__msa_vshf_b(mask, (v16i8)zero, (v16i8)src5); in RGB24ToUVRow_MSA()
2050 src6 = (v16u8)__msa_vshf_b(mask, (v16i8)zero, (v16i8)src6); in RGB24ToUVRow_MSA()
2051 src7 = (v16u8)__msa_vshf_b(mask, (v16i8)zero, (v16i8)src7); in RGB24ToUVRow_MSA()
2060 vec0 = (v8u16)__msa_hadd_u_h((v16u8)vec0, (v16u8)vec0); in RGB24ToUVRow_MSA()
2061 vec1 = (v8u16)__msa_hadd_u_h((v16u8)vec1, (v16u8)vec1); in RGB24ToUVRow_MSA()
2062 vec2 = (v8u16)__msa_hadd_u_h((v16u8)vec2, (v16u8)vec2); in RGB24ToUVRow_MSA()
2063 vec3 = (v8u16)__msa_hadd_u_h((v16u8)vec3, (v16u8)vec3); in RGB24ToUVRow_MSA()
2064 vec4 = (v8u16)__msa_hadd_u_h((v16u8)vec4, (v16u8)vec4); in RGB24ToUVRow_MSA()
2065 vec5 = (v8u16)__msa_hadd_u_h((v16u8)vec5, (v16u8)vec5); in RGB24ToUVRow_MSA()
2066 vec6 = (v8u16)__msa_hadd_u_h((v16u8)vec6, (v16u8)vec6); in RGB24ToUVRow_MSA()
2067 vec7 = (v8u16)__msa_hadd_u_h((v16u8)vec7, (v16u8)vec7); in RGB24ToUVRow_MSA()
2101 dst0 = (v16u8)__msa_pckev_b((v16i8)reg2, (v16i8)reg0); in RGB24ToUVRow_MSA()
2122 v16u8 inp0, inp1, inp2, inp3, inp4, inp5; in RAWToUVRow_MSA()
2123 v16u8 src0, src1, src2, src3, src4, src5, src6, src7; in RAWToUVRow_MSA()
2126 v16u8 dst0; in RAWToUVRow_MSA()
2137 inp0 = (v16u8)__msa_ld_b((v16i8*)s, 0); in RAWToUVRow_MSA()
2138 inp1 = (v16u8)__msa_ld_b((v16i8*)s, 16); in RAWToUVRow_MSA()
2139 inp2 = (v16u8)__msa_ld_b((v16i8*)s, 32); in RAWToUVRow_MSA()
2140 inp3 = (v16u8)__msa_ld_b((v16i8*)t, 0); in RAWToUVRow_MSA()
2141 inp4 = (v16u8)__msa_ld_b((v16i8*)t, 16); in RAWToUVRow_MSA()
2142 inp5 = (v16u8)__msa_ld_b((v16i8*)t, 32); in RAWToUVRow_MSA()
2143 src1 = (v16u8)__msa_sldi_b((v16i8)inp1, (v16i8)inp0, 12); in RAWToUVRow_MSA()
2144 src5 = (v16u8)__msa_sldi_b((v16i8)inp4, (v16i8)inp3, 12); in RAWToUVRow_MSA()
2145 src2 = (v16u8)__msa_sldi_b((v16i8)inp2, (v16i8)inp1, 8); in RAWToUVRow_MSA()
2146 src6 = (v16u8)__msa_sldi_b((v16i8)inp5, (v16i8)inp4, 8); in RAWToUVRow_MSA()
2147 src3 = (v16u8)__msa_sldi_b((v16i8)inp2, (v16i8)inp2, 4); in RAWToUVRow_MSA()
2148 src7 = (v16u8)__msa_sldi_b((v16i8)inp5, (v16i8)inp5, 4); in RAWToUVRow_MSA()
2149 src0 = (v16u8)__msa_vshf_b(mask, (v16i8)zero, (v16i8)inp0); in RAWToUVRow_MSA()
2150 src1 = (v16u8)__msa_vshf_b(mask, (v16i8)zero, (v16i8)src1); in RAWToUVRow_MSA()
2151 src2 = (v16u8)__msa_vshf_b(mask, (v16i8)zero, (v16i8)src2); in RAWToUVRow_MSA()
2152 src3 = (v16u8)__msa_vshf_b(mask, (v16i8)zero, (v16i8)src3); in RAWToUVRow_MSA()
2153 src4 = (v16u8)__msa_vshf_b(mask, (v16i8)zero, (v16i8)inp3); in RAWToUVRow_MSA()
2154 src5 = (v16u8)__msa_vshf_b(mask, (v16i8)zero, (v16i8)src5); in RAWToUVRow_MSA()
2155 src6 = (v16u8)__msa_vshf_b(mask, (v16i8)zero, (v16i8)src6); in RAWToUVRow_MSA()
2156 src7 = (v16u8)__msa_vshf_b(mask, (v16i8)zero, (v16i8)src7); in RAWToUVRow_MSA()
2165 vec0 = (v8u16)__msa_hadd_u_h((v16u8)vec0, (v16u8)vec0); in RAWToUVRow_MSA()
2166 vec1 = (v8u16)__msa_hadd_u_h((v16u8)vec1, (v16u8)vec1); in RAWToUVRow_MSA()
2167 vec2 = (v8u16)__msa_hadd_u_h((v16u8)vec2, (v16u8)vec2); in RAWToUVRow_MSA()
2168 vec3 = (v8u16)__msa_hadd_u_h((v16u8)vec3, (v16u8)vec3); in RAWToUVRow_MSA()
2169 vec4 = (v8u16)__msa_hadd_u_h((v16u8)vec4, (v16u8)vec4); in RAWToUVRow_MSA()
2170 vec5 = (v8u16)__msa_hadd_u_h((v16u8)vec5, (v16u8)vec5); in RAWToUVRow_MSA()
2171 vec6 = (v8u16)__msa_hadd_u_h((v16u8)vec6, (v16u8)vec6); in RAWToUVRow_MSA()
2172 vec7 = (v8u16)__msa_hadd_u_h((v16u8)vec7, (v16u8)vec7); in RAWToUVRow_MSA()
2206 dst0 = (v16u8)__msa_pckev_b((v16i8)reg2, (v16i8)reg0); in RAWToUVRow_MSA()
2225 v16u8 src0, src1, res0, res1, dst0, dst1; in NV12ToARGBRow_MSA()
2229 v16u8 zero = {0}; in NV12ToARGBRow_MSA()
2230 v16u8 alpha = (v16u8)__msa_ldi_b(ALPHA_VAL); in NV12ToARGBRow_MSA()
2240 src0 = (v16u8)__msa_insert_d((v2i64)zero, 0, val0); in NV12ToARGBRow_MSA()
2241 src1 = (v16u8)__msa_insert_d((v2i64)zero, 0, val1); in NV12ToARGBRow_MSA()
2244 res0 = (v16u8)__msa_ilvev_b((v16i8)vec2, (v16i8)vec0); in NV12ToARGBRow_MSA()
2245 res1 = (v16u8)__msa_ilvev_b((v16i8)alpha, (v16i8)vec1); in NV12ToARGBRow_MSA()
2246 dst0 = (v16u8)__msa_ilvr_b((v16i8)res1, (v16i8)res0); in NV12ToARGBRow_MSA()
2247 dst1 = (v16u8)__msa_ilvl_b((v16i8)res1, (v16i8)res0); in NV12ToARGBRow_MSA()
2262 v16u8 src0, src1, dst0; in NV12ToRGB565Row_MSA()
2266 v16u8 zero = {0}; in NV12ToRGB565Row_MSA()
2276 src0 = (v16u8)__msa_insert_d((v2i64)zero, 0, val0); in NV12ToRGB565Row_MSA()
2277 src1 = (v16u8)__msa_insert_d((v2i64)zero, 0, val1); in NV12ToRGB565Row_MSA()
2283 dst0 = (v16u8)(vec0 | vec1 | vec2); in NV12ToRGB565Row_MSA()
2298 v16u8 src0, src1, res0, res1, dst0, dst1; in NV21ToARGBRow_MSA()
2302 v16u8 alpha = (v16u8)__msa_ldi_b(ALPHA_VAL); in NV21ToARGBRow_MSA()
2303 v16u8 zero = {0}; in NV21ToARGBRow_MSA()
2314 src0 = (v16u8)__msa_insert_d((v2i64)zero, 0, val0); in NV21ToARGBRow_MSA()
2315 src1 = (v16u8)__msa_insert_d((v2i64)zero, 0, val1); in NV21ToARGBRow_MSA()
2316 src1 = (v16u8)__msa_vshf_b(shuffler, (v16i8)src1, (v16i8)src1); in NV21ToARGBRow_MSA()
2319 res0 = (v16u8)__msa_ilvev_b((v16i8)vec2, (v16i8)vec0); in NV21ToARGBRow_MSA()
2320 res1 = (v16u8)__msa_ilvev_b((v16i8)alpha, (v16i8)vec1); in NV21ToARGBRow_MSA()
2321 dst0 = (v16u8)__msa_ilvr_b((v16i8)res1, (v16i8)res0); in NV21ToARGBRow_MSA()
2322 dst1 = (v16u8)__msa_ilvl_b((v16i8)res1, (v16i8)res0); in NV21ToARGBRow_MSA()
2335 v16u8 src0, src1, vec0, dst0, dst1, dst2, dst3; in SobelRow_MSA()
2341 v16u8 alpha = (v16u8)__msa_ldi_b(ALPHA_VAL); in SobelRow_MSA()
2344 src0 = (v16u8)__msa_ld_b((v16i8*)src_sobelx, 0); in SobelRow_MSA()
2345 src1 = (v16u8)__msa_ld_b((v16i8*)src_sobely, 0); in SobelRow_MSA()
2347 dst0 = (v16u8)__msa_vshf_b(mask0, (v16i8)alpha, (v16i8)vec0); in SobelRow_MSA()
2348 dst1 = (v16u8)__msa_vshf_b(mask1, (v16i8)alpha, (v16i8)vec0); in SobelRow_MSA()
2349 dst2 = (v16u8)__msa_vshf_b(mask2, (v16i8)alpha, (v16i8)vec0); in SobelRow_MSA()
2350 dst3 = (v16u8)__msa_vshf_b(mask3, (v16i8)alpha, (v16i8)vec0); in SobelRow_MSA()
2363 v16u8 src0, src1, src2, src3, dst0, dst1; in SobelToPlaneRow_MSA()
2366 src0 = (v16u8)__msa_ld_b((v16i8*)src_sobelx, 0); in SobelToPlaneRow_MSA()
2367 src1 = (v16u8)__msa_ld_b((v16i8*)src_sobelx, 16); in SobelToPlaneRow_MSA()
2368 src2 = (v16u8)__msa_ld_b((v16i8*)src_sobely, 0); in SobelToPlaneRow_MSA()
2369 src3 = (v16u8)__msa_ld_b((v16i8*)src_sobely, 16); in SobelToPlaneRow_MSA()
2384 v16u8 src0, src1, vec0, vec1, vec2; in SobelXYRow_MSA()
2385 v16u8 reg0, reg1, dst0, dst1, dst2, dst3; in SobelXYRow_MSA()
2386 v16u8 alpha = (v16u8)__msa_ldi_b(ALPHA_VAL); in SobelXYRow_MSA()
2389 src0 = (v16u8)__msa_ld_b((v16i8*)src_sobelx, 0); in SobelXYRow_MSA()
2390 src1 = (v16u8)__msa_ld_b((v16i8*)src_sobely, 0); in SobelXYRow_MSA()
2392 vec1 = (v16u8)__msa_ilvr_b((v16i8)src0, (v16i8)src1); in SobelXYRow_MSA()
2393 vec2 = (v16u8)__msa_ilvl_b((v16i8)src0, (v16i8)src1); in SobelXYRow_MSA()
2394 reg0 = (v16u8)__msa_ilvr_b((v16i8)alpha, (v16i8)vec0); in SobelXYRow_MSA()
2395 reg1 = (v16u8)__msa_ilvl_b((v16i8)alpha, (v16i8)vec0); in SobelXYRow_MSA()
2396 dst0 = (v16u8)__msa_ilvr_b((v16i8)reg0, (v16i8)vec1); in SobelXYRow_MSA()
2397 dst1 = (v16u8)__msa_ilvl_b((v16i8)reg0, (v16i8)vec1); in SobelXYRow_MSA()
2398 dst2 = (v16u8)__msa_ilvr_b((v16i8)reg1, (v16i8)vec2); in SobelXYRow_MSA()
2399 dst3 = (v16u8)__msa_ilvl_b((v16i8)reg1, (v16i8)vec2); in SobelXYRow_MSA()
2409 v16u8 src0, src1, src2, src3, dst0; in ARGBToYJRow_MSA()
2410 v16u8 const_0x4B0F = (v16u8)__msa_fill_h(0x4B0F); in ARGBToYJRow_MSA()
2411 v16u8 const_0x26 = (v16u8)__msa_fill_h(0x26); in ARGBToYJRow_MSA()
2415 src0 = (v16u8)__msa_ld_b((v16i8*)src_argb0, 0); in ARGBToYJRow_MSA()
2416 src1 = (v16u8)__msa_ld_b((v16i8*)src_argb0, 16); in ARGBToYJRow_MSA()
2417 src2 = (v16u8)__msa_ld_b((v16i8*)src_argb0, 32); in ARGBToYJRow_MSA()
2418 src3 = (v16u8)__msa_ld_b((v16i8*)src_argb0, 48); in ARGBToYJRow_MSA()
2429 v16u8 src0, src1, src2, src3, dst0; in BGRAToYRow_MSA()
2430 v16u8 const_0x4200 = (v16u8)__msa_fill_h(0x4200); in BGRAToYRow_MSA()
2431 v16u8 const_0x1981 = (v16u8)__msa_fill_h(0x1981); in BGRAToYRow_MSA()
2435 src0 = (v16u8)__msa_ld_b((v16i8*)src_argb0, 0); in BGRAToYRow_MSA()
2436 src1 = (v16u8)__msa_ld_b((v16i8*)src_argb0, 16); in BGRAToYRow_MSA()
2437 src2 = (v16u8)__msa_ld_b((v16i8*)src_argb0, 32); in BGRAToYRow_MSA()
2438 src3 = (v16u8)__msa_ld_b((v16i8*)src_argb0, 48); in BGRAToYRow_MSA()
2449 v16u8 src0, src1, src2, src3, dst0; in ABGRToYRow_MSA()
2450 v16u8 const_0x8142 = (v16u8)__msa_fill_h(0x8142); in ABGRToYRow_MSA()
2451 v16u8 const_0x19 = (v16u8)__msa_fill_h(0x19); in ABGRToYRow_MSA()
2455 src0 = (v16u8)__msa_ld_b((v16i8*)src_argb0, 0); in ABGRToYRow_MSA()
2456 src1 = (v16u8)__msa_ld_b((v16i8*)src_argb0, 16); in ABGRToYRow_MSA()
2457 src2 = (v16u8)__msa_ld_b((v16i8*)src_argb0, 32); in ABGRToYRow_MSA()
2458 src3 = (v16u8)__msa_ld_b((v16i8*)src_argb0, 48); in ABGRToYRow_MSA()
2469 v16u8 src0, src1, src2, src3, dst0; in RGBAToYRow_MSA()
2470 v16u8 const_0x1900 = (v16u8)__msa_fill_h(0x1900); in RGBAToYRow_MSA()
2471 v16u8 const_0x4281 = (v16u8)__msa_fill_h(0x4281); in RGBAToYRow_MSA()
2475 src0 = (v16u8)__msa_ld_b((v16i8*)src_argb0, 0); in RGBAToYRow_MSA()
2476 src1 = (v16u8)__msa_ld_b((v16i8*)src_argb0, 16); in RGBAToYRow_MSA()
2477 src2 = (v16u8)__msa_ld_b((v16i8*)src_argb0, 32); in RGBAToYRow_MSA()
2478 src3 = (v16u8)__msa_ld_b((v16i8*)src_argb0, 48); in RGBAToYRow_MSA()
2495 v16u8 src0, src1, src2, src3, src4, src5, src6, src7; in ARGBToUVJRow_MSA()
2496 v16u8 vec0, vec1, vec2, vec3; in ARGBToUVJRow_MSA()
2497 v16u8 dst0, dst1; in ARGBToUVJRow_MSA()
2503 v16u8 const_0x7F = (v16u8)__msa_fill_h(0x7F); in ARGBToUVJRow_MSA()
2504 v16u8 const_0x6B14 = (v16u8)__msa_fill_h(0x6B14); in ARGBToUVJRow_MSA()
2505 v16u8 const_0x2B54 = (v16u8)__msa_fill_h(0x2B54); in ARGBToUVJRow_MSA()
2509 src0 = (v16u8)__msa_ld_b((v16i8*)s, 0); in ARGBToUVJRow_MSA()
2510 src1 = (v16u8)__msa_ld_b((v16i8*)s, 16); in ARGBToUVJRow_MSA()
2511 src2 = (v16u8)__msa_ld_b((v16i8*)s, 32); in ARGBToUVJRow_MSA()
2512 src3 = (v16u8)__msa_ld_b((v16i8*)s, 48); in ARGBToUVJRow_MSA()
2513 src4 = (v16u8)__msa_ld_b((v16i8*)t, 0); in ARGBToUVJRow_MSA()
2514 src5 = (v16u8)__msa_ld_b((v16i8*)t, 16); in ARGBToUVJRow_MSA()
2515 src6 = (v16u8)__msa_ld_b((v16i8*)t, 32); in ARGBToUVJRow_MSA()
2516 src7 = (v16u8)__msa_ld_b((v16i8*)t, 48); in ARGBToUVJRow_MSA()
2521 src4 = (v16u8)__msa_pckev_w((v4i32)src1, (v4i32)src0); in ARGBToUVJRow_MSA()
2522 src5 = (v16u8)__msa_pckev_w((v4i32)src3, (v4i32)src2); in ARGBToUVJRow_MSA()
2523 src6 = (v16u8)__msa_pckod_w((v4i32)src1, (v4i32)src0); in ARGBToUVJRow_MSA()
2524 src7 = (v16u8)__msa_pckod_w((v4i32)src3, (v4i32)src2); in ARGBToUVJRow_MSA()
2527 src0 = (v16u8)__msa_ld_b((v16i8*)s, 64); in ARGBToUVJRow_MSA()
2528 src1 = (v16u8)__msa_ld_b((v16i8*)s, 80); in ARGBToUVJRow_MSA()
2529 src2 = (v16u8)__msa_ld_b((v16i8*)s, 96); in ARGBToUVJRow_MSA()
2530 src3 = (v16u8)__msa_ld_b((v16i8*)s, 112); in ARGBToUVJRow_MSA()
2531 src4 = (v16u8)__msa_ld_b((v16i8*)t, 64); in ARGBToUVJRow_MSA()
2532 src5 = (v16u8)__msa_ld_b((v16i8*)t, 80); in ARGBToUVJRow_MSA()
2533 src6 = (v16u8)__msa_ld_b((v16i8*)t, 96); in ARGBToUVJRow_MSA()
2534 src7 = (v16u8)__msa_ld_b((v16i8*)t, 112); in ARGBToUVJRow_MSA()
2539 src4 = (v16u8)__msa_pckev_w((v4i32)src1, (v4i32)src0); in ARGBToUVJRow_MSA()
2540 src5 = (v16u8)__msa_pckev_w((v4i32)src3, (v4i32)src2); in ARGBToUVJRow_MSA()
2541 src6 = (v16u8)__msa_pckod_w((v4i32)src1, (v4i32)src0); in ARGBToUVJRow_MSA()
2542 src7 = (v16u8)__msa_pckod_w((v4i32)src3, (v4i32)src2); in ARGBToUVJRow_MSA()
2565 v16u8 dst0, dst1, vec0, vec1, vec2, vec3; in BGRAToUVRow_MSA()
2571 v16u8 const_0x125E = (v16u8)__msa_fill_h(0x125E); in BGRAToUVRow_MSA()
2572 v16u8 const_0x7000 = (v16u8)__msa_fill_h(0x7000); in BGRAToUVRow_MSA()
2573 v16u8 const_0x264A = (v16u8)__msa_fill_h(0x264A); in BGRAToUVRow_MSA()
2598 v16u8 src0, src1, src2, src3; in ABGRToUVRow_MSA()
2599 v16u8 dst0, dst1; in ABGRToUVRow_MSA()
2605 v16u8 const_0x4A26 = (v16u8)__msa_fill_h(0x4A26); in ABGRToUVRow_MSA()
2606 v16u8 const_0x0070 = (v16u8)__msa_fill_h(0x0070); in ABGRToUVRow_MSA()
2607 v16u8 const_0x125E = (v16u8)__msa_fill_h(0x125E); in ABGRToUVRow_MSA()
2632 v16u8 dst0, dst1, vec0, vec1, vec2, vec3; in RGBAToUVRow_MSA()
2638 v16u8 const_0x125E = (v16u8)__msa_fill_h(0x264A); in RGBAToUVRow_MSA()
2639 v16u8 const_0x7000 = (v16u8)__msa_fill_h(0x7000); in RGBAToUVRow_MSA()
2640 v16u8 const_0x264A = (v16u8)__msa_fill_h(0x125E); in RGBAToUVRow_MSA()
2664 v16u8 src0, src1, src2, dst0, dst1; in I444ToARGBRow_MSA()
2668 v16u8 alpha = (v16u8)__msa_ldi_b(ALPHA_VAL); in I444ToARGBRow_MSA()
2715 dst0 = (v16u8)__msa_ilvr_h((v8i16)vec1, (v8i16)vec0); in I444ToARGBRow_MSA()
2716 dst1 = (v16u8)__msa_ilvl_h((v8i16)vec1, (v8i16)vec0); in I444ToARGBRow_MSA()
2727 v16u8 src0, res0, res1, res2, res3, res4, dst0, dst1, dst2, dst3; in I400ToARGBRow_MSA()
2732 v16u8 alpha = (v16u8)__msa_ldi_b(ALPHA_VAL); in I400ToARGBRow_MSA()
2737 src0 = (v16u8)__msa_ld_b((v16i8*)src_y, 0); in I400ToARGBRow_MSA()
2762 res0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0); in I400ToARGBRow_MSA()
2763 res1 = (v16u8)__msa_ilvr_b((v16i8)res0, (v16i8)res0); in I400ToARGBRow_MSA()
2764 res2 = (v16u8)__msa_ilvl_b((v16i8)res0, (v16i8)res0); in I400ToARGBRow_MSA()
2765 res3 = (v16u8)__msa_ilvr_b((v16i8)alpha, (v16i8)res0); in I400ToARGBRow_MSA()
2766 res4 = (v16u8)__msa_ilvl_b((v16i8)alpha, (v16i8)res0); in I400ToARGBRow_MSA()
2767 dst0 = (v16u8)__msa_ilvr_b((v16i8)res3, (v16i8)res1); in I400ToARGBRow_MSA()
2768 dst1 = (v16u8)__msa_ilvl_b((v16i8)res3, (v16i8)res1); in I400ToARGBRow_MSA()
2769 dst2 = (v16u8)__msa_ilvr_b((v16i8)res4, (v16i8)res2); in I400ToARGBRow_MSA()
2770 dst3 = (v16u8)__msa_ilvl_b((v16i8)res4, (v16i8)res2); in I400ToARGBRow_MSA()
2779 v16u8 src0, vec0, vec1, vec2, vec3, dst0, dst1, dst2, dst3; in J400ToARGBRow_MSA()
2780 v16u8 alpha = (v16u8)__msa_ldi_b(ALPHA_VAL); in J400ToARGBRow_MSA()
2783 src0 = (v16u8)__msa_ld_b((v16i8*)src_y, 0); in J400ToARGBRow_MSA()
2784 vec0 = (v16u8)__msa_ilvr_b((v16i8)src0, (v16i8)src0); in J400ToARGBRow_MSA()
2785 vec1 = (v16u8)__msa_ilvl_b((v16i8)src0, (v16i8)src0); in J400ToARGBRow_MSA()
2786 vec2 = (v16u8)__msa_ilvr_b((v16i8)alpha, (v16i8)src0); in J400ToARGBRow_MSA()
2787 vec3 = (v16u8)__msa_ilvl_b((v16i8)alpha, (v16i8)src0); in J400ToARGBRow_MSA()
2788 dst0 = (v16u8)__msa_ilvr_b((v16i8)vec2, (v16i8)vec0); in J400ToARGBRow_MSA()
2789 dst1 = (v16u8)__msa_ilvl_b((v16i8)vec2, (v16i8)vec0); in J400ToARGBRow_MSA()
2790 dst2 = (v16u8)__msa_ilvr_b((v16i8)vec3, (v16i8)vec1); in J400ToARGBRow_MSA()
2791 dst3 = (v16u8)__msa_ilvl_b((v16i8)vec3, (v16i8)vec1); in J400ToARGBRow_MSA()
2803 v16u8 src0, src1, src2; in YUY2ToARGBRow_MSA()
2807 v16u8 alpha = (v16u8)__msa_ldi_b(ALPHA_VAL); in YUY2ToARGBRow_MSA()
2815 src0 = (v16u8)__msa_ld_b((v16i8*)src_yuy2, 0); in YUY2ToARGBRow_MSA()
2816 src1 = (v16u8)__msa_pckev_b((v16i8)src0, (v16i8)src0); in YUY2ToARGBRow_MSA()
2817 src2 = (v16u8)__msa_pckod_b((v16i8)src0, (v16i8)src0); in YUY2ToARGBRow_MSA()
2831 v16u8 src0, src1, src2; in UYVYToARGBRow_MSA()
2835 v16u8 alpha = (v16u8)__msa_ldi_b(ALPHA_VAL); in UYVYToARGBRow_MSA()
2843 src0 = (v16u8)__msa_ld_b((v16i8*)src_uyvy, 0); in UYVYToARGBRow_MSA()
2844 src1 = (v16u8)__msa_pckod_b((v16i8)src0, (v16i8)src0); in UYVYToARGBRow_MSA()
2845 src2 = (v16u8)__msa_pckev_b((v16i8)src0, (v16i8)src0); in UYVYToARGBRow_MSA()
2865 v16u8 src0, src1, src2, src3, dst0, dst1; in InterpolateRow_MSA()
2875 src0 = (v16u8)__msa_ld_b((v16i8*)s, 0); in InterpolateRow_MSA()
2876 src1 = (v16u8)__msa_ld_b((v16i8*)s, 16); in InterpolateRow_MSA()
2877 src2 = (v16u8)__msa_ld_b((v16i8*)t, 0); in InterpolateRow_MSA()
2878 src3 = (v16u8)__msa_ld_b((v16i8*)t, 16); in InterpolateRow_MSA()
2893 src0 = (v16u8)__msa_ld_b((v16i8*)s, 0); in InterpolateRow_MSA()
2894 src1 = (v16u8)__msa_ld_b((v16i8*)s, 16); in InterpolateRow_MSA()
2895 src2 = (v16u8)__msa_ld_b((v16i8*)t, 0); in InterpolateRow_MSA()
2896 src3 = (v16u8)__msa_ld_b((v16i8*)t, 16); in InterpolateRow_MSA()
2901 vec0 = (v8u16)__msa_dotp_u_h((v16u8)vec0, (v16u8)y_frac); in InterpolateRow_MSA()
2902 vec1 = (v8u16)__msa_dotp_u_h((v16u8)vec1, (v16u8)y_frac); in InterpolateRow_MSA()
2903 vec2 = (v8u16)__msa_dotp_u_h((v16u8)vec2, (v16u8)y_frac); in InterpolateRow_MSA()
2904 vec3 = (v8u16)__msa_dotp_u_h((v16u8)vec3, (v16u8)y_frac); in InterpolateRow_MSA()
2909 dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0); in InterpolateRow_MSA()
2910 dst1 = (v16u8)__msa_pckev_b((v16i8)vec3, (v16i8)vec2); in InterpolateRow_MSA()
2920 v16u8 dst0 = (v16u8)__msa_fill_w(v32); in ARGBSetRow_MSA()
2930 v16u8 src0, src1, src2, src3, src4, dst0, dst1, dst2; in RAWToRGB24Row_MSA()
2938 src0 = (v16u8)__msa_ld_b((v16i8*)src_raw, 0); in RAWToRGB24Row_MSA()
2939 src1 = (v16u8)__msa_ld_b((v16i8*)src_raw, 16); in RAWToRGB24Row_MSA()
2940 src2 = (v16u8)__msa_ld_b((v16i8*)src_raw, 32); in RAWToRGB24Row_MSA()
2941 src3 = (v16u8)__msa_sldi_b((v16i8)src1, (v16i8)src0, 8); in RAWToRGB24Row_MSA()
2942 src4 = (v16u8)__msa_sldi_b((v16i8)src2, (v16i8)src1, 8); in RAWToRGB24Row_MSA()
2943 dst0 = (v16u8)__msa_vshf_b(shuffler0, (v16i8)src1, (v16i8)src0); in RAWToRGB24Row_MSA()
2944 dst1 = (v16u8)__msa_vshf_b(shuffler1, (v16i8)src4, (v16i8)src3); in RAWToRGB24Row_MSA()
2945 dst2 = (v16u8)__msa_vshf_b(shuffler2, (v16i8)src2, (v16i8)src1); in RAWToRGB24Row_MSA()
2958 v16u8 src0, src1, dst0, dst1; in MergeUVRow_MSA()
2961 src0 = (v16u8)__msa_ld_b((v16i8*)src_u, 0); in MergeUVRow_MSA()
2962 src1 = (v16u8)__msa_ld_b((v16i8*)src_v, 0); in MergeUVRow_MSA()
2963 dst0 = (v16u8)__msa_ilvr_b((v16i8)src1, (v16i8)src0); in MergeUVRow_MSA()
2964 dst1 = (v16u8)__msa_ilvl_b((v16i8)src1, (v16i8)src0); in MergeUVRow_MSA()