Lines Matching refs:temp2
104 __m128i temp1, temp2, temp3, temp4, temp5, temp6; in ih264_deblk_luma_vert_bs4_ssse3() local
123 temp2 = _mm_unpacklo_epi8(line3, line4); in ih264_deblk_luma_vert_bs4_ssse3()
127 line1 = _mm_unpacklo_epi16(temp1, temp2); in ih264_deblk_luma_vert_bs4_ssse3()
128 line2 = _mm_unpackhi_epi16(temp1, temp2); in ih264_deblk_luma_vert_bs4_ssse3()
147 temp2 = _mm_unpacklo_epi8(line3, line4); in ih264_deblk_luma_vert_bs4_ssse3()
151 line1 = _mm_unpacklo_epi16(temp1, temp2); in ih264_deblk_luma_vert_bs4_ssse3()
152 line2 = _mm_unpackhi_epi16(temp1, temp2); in ih264_deblk_luma_vert_bs4_ssse3()
157 temp2 = _mm_unpackhi_epi32(line1, line3); in ih264_deblk_luma_vert_bs4_ssse3()
165 p1_16x8 = _mm_unpacklo_epi64(p0_8x16, temp2); in ih264_deblk_luma_vert_bs4_ssse3()
166 p0_16x8 = _mm_unpackhi_epi64(p0_8x16, temp2); in ih264_deblk_luma_vert_bs4_ssse3()
172 temp2 = _mm_subs_epu8(p0_16x8, q0_16x8); in ih264_deblk_luma_vert_bs4_ssse3()
173 temp1 = _mm_add_epi8(temp1, temp2); in ih264_deblk_luma_vert_bs4_ssse3()
175 temp2 = _mm_unpacklo_epi8(temp1, zero); in ih264_deblk_luma_vert_bs4_ssse3()
178 temp2 = _mm_cmpgt_epi16(Alpha_8x16, temp2); in ih264_deblk_luma_vert_bs4_ssse3()
181 flag1_16x8 = _mm_packs_epi16(temp2, temp1); in ih264_deblk_luma_vert_bs4_ssse3()
185 temp2 = _mm_subs_epu8(q1_16x8, q0_16x8); in ih264_deblk_luma_vert_bs4_ssse3()
186 temp1 = _mm_add_epi8(temp1, temp2); in ih264_deblk_luma_vert_bs4_ssse3()
188 temp2 = _mm_unpacklo_epi8(temp1, zero); in ih264_deblk_luma_vert_bs4_ssse3()
191 temp2 = _mm_cmpgt_epi16(Beta_8x16, temp2); in ih264_deblk_luma_vert_bs4_ssse3()
194 flag2_16x8 = _mm_packs_epi16(temp2, temp1); in ih264_deblk_luma_vert_bs4_ssse3()
200 temp2 = _mm_subs_epu8(p1_16x8, p0_16x8); in ih264_deblk_luma_vert_bs4_ssse3()
201 temp1 = _mm_add_epi8(temp1, temp2); in ih264_deblk_luma_vert_bs4_ssse3()
203 temp2 = _mm_unpacklo_epi8(temp1, zero); in ih264_deblk_luma_vert_bs4_ssse3()
206 temp2 = _mm_cmpgt_epi16(Beta_8x16, temp2); in ih264_deblk_luma_vert_bs4_ssse3()
209 flag2_16x8 = _mm_packs_epi16(temp2, temp1); in ih264_deblk_luma_vert_bs4_ssse3()
216 temp2 = _mm_subs_epu8(q0_16x8, p0_16x8); in ih264_deblk_luma_vert_bs4_ssse3()
217 temp1 = _mm_add_epi8(temp1, temp2); in ih264_deblk_luma_vert_bs4_ssse3()
221 temp2 = _mm_unpacklo_epi8(temp1, zero); in ih264_deblk_luma_vert_bs4_ssse3()
223 temp2 = _mm_cmpgt_epi16(Alpha_8x16, temp2); in ih264_deblk_luma_vert_bs4_ssse3()
226 flag2_16x8 = _mm_packs_epi16(temp2, temp1); in ih264_deblk_luma_vert_bs4_ssse3()
231 temp2 = _mm_subs_epu8(p2_16x8, p0_16x8); in ih264_deblk_luma_vert_bs4_ssse3()
232 temp1 = _mm_add_epi8(temp1, temp2); in ih264_deblk_luma_vert_bs4_ssse3()
234 temp2 = _mm_unpacklo_epi8(temp1, zero); in ih264_deblk_luma_vert_bs4_ssse3()
236 temp2 = _mm_cmpgt_epi16(Beta_8x16, temp2); in ih264_deblk_luma_vert_bs4_ssse3()
239 flag3_16x8 = _mm_packs_epi16(temp2, temp1); in ih264_deblk_luma_vert_bs4_ssse3()
244 temp2 = _mm_subs_epu8(q2_16x8, q0_16x8); in ih264_deblk_luma_vert_bs4_ssse3()
245 temp1 = _mm_add_epi8(temp1, temp2); in ih264_deblk_luma_vert_bs4_ssse3()
247 temp2 = _mm_unpacklo_epi8(temp1, zero); in ih264_deblk_luma_vert_bs4_ssse3()
249 temp2 = _mm_cmpgt_epi16(Beta_8x16, temp2); in ih264_deblk_luma_vert_bs4_ssse3()
252 flag4_16x8 = _mm_packs_epi16(temp2, temp1); in ih264_deblk_luma_vert_bs4_ssse3()
267 temp2 = _mm_add_epi16(p1_8x16, q0_8x16); in ih264_deblk_luma_vert_bs4_ssse3()
269 temp6 = _mm_add_epi16(temp2, const_val2_16x8); in ih264_deblk_luma_vert_bs4_ssse3()
273 temp2 = _mm_add_epi16(temp6, temp4); in ih264_deblk_luma_vert_bs4_ssse3()
275 q0_16x8_1 = _mm_srai_epi16(temp2, 2); in ih264_deblk_luma_vert_bs4_ssse3()
281 temp2 = _mm_add_epi16(temp5, q2_8x16); in ih264_deblk_luma_vert_bs4_ssse3()
283 q1_16x8_2 = _mm_srai_epi16(temp2, 2); in ih264_deblk_luma_vert_bs4_ssse3()
287 temp2 = _mm_add_epi16(temp4, q2_8x16); in ih264_deblk_luma_vert_bs4_ssse3()
289 temp2 = _mm_add_epi16(temp2, p1_8x16); in ih264_deblk_luma_vert_bs4_ssse3()
293 temp2 = _mm_add_epi16(temp2, temp3); in ih264_deblk_luma_vert_bs4_ssse3()
295 temp2 = _mm_add_epi16(temp2, _mm_set1_epi16(4)); in ih264_deblk_luma_vert_bs4_ssse3()
297 q0_16x8_2 = _mm_srai_epi16(temp2, 3); in ih264_deblk_luma_vert_bs4_ssse3()
301 temp2 = _mm_add_epi16(temp5, const_val2_16x8); in ih264_deblk_luma_vert_bs4_ssse3()
309 temp2 = _mm_add_epi16(temp2, temp4); in ih264_deblk_luma_vert_bs4_ssse3()
311 temp2 = _mm_add_epi16(temp2, temp6); in ih264_deblk_luma_vert_bs4_ssse3()
313 q2_16x8_2 = _mm_srai_epi16(temp2, 3); in ih264_deblk_luma_vert_bs4_ssse3()
327 temp2 = _mm_add_epi16(p1_8x16, q0_8x16); in ih264_deblk_luma_vert_bs4_ssse3()
329 temp6 = _mm_add_epi16(temp2, const_val2_16x8); in ih264_deblk_luma_vert_bs4_ssse3()
333 temp2 = _mm_add_epi16(temp6, temp4); in ih264_deblk_luma_vert_bs4_ssse3()
335 temp2 = _mm_srai_epi16(temp2, 2); in ih264_deblk_luma_vert_bs4_ssse3()
337 q0_16x8_1 = _mm_packus_epi16(q0_16x8_1, temp2); in ih264_deblk_luma_vert_bs4_ssse3()
343 temp2 = _mm_add_epi16(temp5, q2_8x16); in ih264_deblk_luma_vert_bs4_ssse3()
345 temp2 = _mm_srai_epi16(temp2, 2); in ih264_deblk_luma_vert_bs4_ssse3()
347 q1_16x8_2 = _mm_packus_epi16(q1_16x8_2, temp2); in ih264_deblk_luma_vert_bs4_ssse3()
351 temp2 = _mm_add_epi16(temp4, q2_8x16); in ih264_deblk_luma_vert_bs4_ssse3()
353 temp2 = _mm_add_epi16(temp2, p1_8x16); in ih264_deblk_luma_vert_bs4_ssse3()
357 temp2 = _mm_add_epi16(temp2, temp3); in ih264_deblk_luma_vert_bs4_ssse3()
359 temp2 = _mm_add_epi16(temp2, _mm_set1_epi16(4)); in ih264_deblk_luma_vert_bs4_ssse3()
361 temp2 = _mm_srai_epi16(temp2, 3); in ih264_deblk_luma_vert_bs4_ssse3()
363 q0_16x8_2 = _mm_packus_epi16(q0_16x8_2, temp2); in ih264_deblk_luma_vert_bs4_ssse3()
367 temp2 = _mm_add_epi16(temp5, const_val2_16x8); in ih264_deblk_luma_vert_bs4_ssse3()
375 temp2 = _mm_add_epi16(temp2, temp4); in ih264_deblk_luma_vert_bs4_ssse3()
377 temp2 = _mm_add_epi16(temp2, temp6); in ih264_deblk_luma_vert_bs4_ssse3()
379 temp2 = _mm_srai_epi16(temp2, 3); in ih264_deblk_luma_vert_bs4_ssse3()
381 q2_16x8_2 = _mm_packus_epi16(q2_16x8_2, temp2); in ih264_deblk_luma_vert_bs4_ssse3()
424 temp2 = _mm_unpacklo_epi8(p1_16x8, p0_16x8); in ih264_deblk_luma_vert_bs4_ssse3()
428 p3_8x16 = _mm_unpacklo_epi16(temp1, temp2); in ih264_deblk_luma_vert_bs4_ssse3()
429 p2_8x16 = _mm_unpackhi_epi16(temp1, temp2); in ih264_deblk_luma_vert_bs4_ssse3()
452 temp2 = _mm_unpackhi_epi8(p1_16x8, p0_16x8); in ih264_deblk_luma_vert_bs4_ssse3()
456 p3_8x16 = _mm_unpacklo_epi16(temp1, temp2); in ih264_deblk_luma_vert_bs4_ssse3()
457 p2_8x16 = _mm_unpackhi_epi16(temp1, temp2); in ih264_deblk_luma_vert_bs4_ssse3()
528 __m128i temp1, temp2, temp3, temp4, temp5, temp6; in ih264_deblk_luma_horz_bs4_ssse3() local
557 temp2 = _mm_subs_epu8(p0_16x8, q0_16x8); in ih264_deblk_luma_horz_bs4_ssse3()
558 temp1 = _mm_add_epi8(temp1, temp2); in ih264_deblk_luma_horz_bs4_ssse3()
560 temp2 = _mm_unpacklo_epi8(temp1, zero); in ih264_deblk_luma_horz_bs4_ssse3()
563 temp2 = _mm_cmpgt_epi16(Alpha_8x16, temp2); in ih264_deblk_luma_horz_bs4_ssse3()
566 flag1_16x8 = _mm_packs_epi16(temp2, temp1); in ih264_deblk_luma_horz_bs4_ssse3()
570 temp2 = _mm_subs_epu8(q1_16x8, q0_16x8); in ih264_deblk_luma_horz_bs4_ssse3()
571 temp1 = _mm_add_epi8(temp1, temp2); in ih264_deblk_luma_horz_bs4_ssse3()
573 temp2 = _mm_unpacklo_epi8(temp1, zero); in ih264_deblk_luma_horz_bs4_ssse3()
576 temp2 = _mm_cmpgt_epi16(Beta_8x16, temp2); in ih264_deblk_luma_horz_bs4_ssse3()
579 flag2_16x8 = _mm_packs_epi16(temp2, temp1); in ih264_deblk_luma_horz_bs4_ssse3()
585 temp2 = _mm_subs_epu8(p1_16x8, p0_16x8); in ih264_deblk_luma_horz_bs4_ssse3()
586 temp1 = _mm_add_epi8(temp1, temp2); in ih264_deblk_luma_horz_bs4_ssse3()
588 temp2 = _mm_unpacklo_epi8(temp1, zero); in ih264_deblk_luma_horz_bs4_ssse3()
591 temp2 = _mm_cmpgt_epi16(Beta_8x16, temp2); in ih264_deblk_luma_horz_bs4_ssse3()
594 flag2_16x8 = _mm_packs_epi16(temp2, temp1); in ih264_deblk_luma_horz_bs4_ssse3()
601 temp2 = _mm_subs_epu8(q0_16x8, p0_16x8); in ih264_deblk_luma_horz_bs4_ssse3()
602 temp1 = _mm_add_epi8(temp1, temp2); in ih264_deblk_luma_horz_bs4_ssse3()
606 temp2 = _mm_unpacklo_epi8(temp1, zero); in ih264_deblk_luma_horz_bs4_ssse3()
608 temp2 = _mm_cmpgt_epi16(Alpha_8x16, temp2); in ih264_deblk_luma_horz_bs4_ssse3()
611 flag2_16x8 = _mm_packs_epi16(temp2, temp1); in ih264_deblk_luma_horz_bs4_ssse3()
616 temp2 = _mm_subs_epu8(p2_16x8, p0_16x8); in ih264_deblk_luma_horz_bs4_ssse3()
617 temp1 = _mm_add_epi8(temp1, temp2); in ih264_deblk_luma_horz_bs4_ssse3()
619 temp2 = _mm_unpacklo_epi8(temp1, zero); in ih264_deblk_luma_horz_bs4_ssse3()
621 temp2 = _mm_cmpgt_epi16(Beta_8x16, temp2); in ih264_deblk_luma_horz_bs4_ssse3()
624 flag3_16x8 = _mm_packs_epi16(temp2, temp1); in ih264_deblk_luma_horz_bs4_ssse3()
629 temp2 = _mm_subs_epu8(q2_16x8, q0_16x8); in ih264_deblk_luma_horz_bs4_ssse3()
630 temp1 = _mm_add_epi8(temp1, temp2); in ih264_deblk_luma_horz_bs4_ssse3()
632 temp2 = _mm_unpacklo_epi8(temp1, zero); in ih264_deblk_luma_horz_bs4_ssse3()
634 temp2 = _mm_cmpgt_epi16(Beta_8x16, temp2); in ih264_deblk_luma_horz_bs4_ssse3()
637 flag4_16x8 = _mm_packs_epi16(temp2, temp1); in ih264_deblk_luma_horz_bs4_ssse3()
652 temp2 = _mm_add_epi16(p1_8x16, q0_8x16); in ih264_deblk_luma_horz_bs4_ssse3()
654 temp6 = _mm_add_epi16(temp2, const_val2_16x8); in ih264_deblk_luma_horz_bs4_ssse3()
658 temp2 = _mm_add_epi16(temp6, temp4); in ih264_deblk_luma_horz_bs4_ssse3()
660 q0_16x8_1 = _mm_srai_epi16(temp2, 2); in ih264_deblk_luma_horz_bs4_ssse3()
666 temp2 = _mm_add_epi16(temp5, q2_8x16); in ih264_deblk_luma_horz_bs4_ssse3()
668 q1_16x8_2 = _mm_srai_epi16(temp2, 2); in ih264_deblk_luma_horz_bs4_ssse3()
672 temp2 = _mm_add_epi16(temp4, q2_8x16); in ih264_deblk_luma_horz_bs4_ssse3()
674 temp2 = _mm_add_epi16(temp2, p1_8x16); in ih264_deblk_luma_horz_bs4_ssse3()
678 temp2 = _mm_add_epi16(temp2, temp3); in ih264_deblk_luma_horz_bs4_ssse3()
680 temp2 = _mm_add_epi16(temp2, _mm_set1_epi16(4)); in ih264_deblk_luma_horz_bs4_ssse3()
682 q0_16x8_2 = _mm_srai_epi16(temp2, 3); in ih264_deblk_luma_horz_bs4_ssse3()
686 temp2 = _mm_add_epi16(temp5, const_val2_16x8); in ih264_deblk_luma_horz_bs4_ssse3()
694 temp2 = _mm_add_epi16(temp2, temp4); in ih264_deblk_luma_horz_bs4_ssse3()
696 temp2 = _mm_add_epi16(temp2, temp6); in ih264_deblk_luma_horz_bs4_ssse3()
698 q2_16x8_2 = _mm_srai_epi16(temp2, 3); in ih264_deblk_luma_horz_bs4_ssse3()
712 temp2 = _mm_add_epi16(p1_8x16, q0_8x16); in ih264_deblk_luma_horz_bs4_ssse3()
714 temp6 = _mm_add_epi16(temp2, const_val2_16x8); in ih264_deblk_luma_horz_bs4_ssse3()
718 temp2 = _mm_add_epi16(temp6, temp4); in ih264_deblk_luma_horz_bs4_ssse3()
720 temp2 = _mm_srai_epi16(temp2, 2); in ih264_deblk_luma_horz_bs4_ssse3()
722 q0_16x8_1 = _mm_packus_epi16(q0_16x8_1, temp2); in ih264_deblk_luma_horz_bs4_ssse3()
728 temp2 = _mm_add_epi16(temp5, q2_8x16); in ih264_deblk_luma_horz_bs4_ssse3()
730 temp2 = _mm_srai_epi16(temp2, 2); in ih264_deblk_luma_horz_bs4_ssse3()
732 q1_16x8_2 = _mm_packus_epi16(q1_16x8_2, temp2); in ih264_deblk_luma_horz_bs4_ssse3()
736 temp2 = _mm_add_epi16(temp4, q2_8x16); in ih264_deblk_luma_horz_bs4_ssse3()
738 temp2 = _mm_add_epi16(temp2, p1_8x16); in ih264_deblk_luma_horz_bs4_ssse3()
742 temp2 = _mm_add_epi16(temp2, temp3); in ih264_deblk_luma_horz_bs4_ssse3()
744 temp2 = _mm_add_epi16(temp2, _mm_set1_epi16(4)); in ih264_deblk_luma_horz_bs4_ssse3()
746 temp2 = _mm_srai_epi16(temp2, 3); in ih264_deblk_luma_horz_bs4_ssse3()
748 q0_16x8_2 = _mm_packus_epi16(q0_16x8_2, temp2); in ih264_deblk_luma_horz_bs4_ssse3()
752 temp2 = _mm_add_epi16(temp5, const_val2_16x8); in ih264_deblk_luma_horz_bs4_ssse3()
760 temp2 = _mm_add_epi16(temp2, temp4); in ih264_deblk_luma_horz_bs4_ssse3()
762 temp2 = _mm_add_epi16(temp2, temp6); in ih264_deblk_luma_horz_bs4_ssse3()
764 temp2 = _mm_srai_epi16(temp2, 3); in ih264_deblk_luma_horz_bs4_ssse3()
766 q2_16x8_2 = _mm_packus_epi16(q2_16x8_2, temp2); in ih264_deblk_luma_horz_bs4_ssse3()
1157 __m128i temp1, temp2; in ih264_deblk_luma_horz_bslt4_ssse3() local
1207 temp2 = _mm_subs_epu8(p0_16x8, q0_16x8); in ih264_deblk_luma_horz_bslt4_ssse3()
1208 temp1 = _mm_add_epi8(temp1, temp2); in ih264_deblk_luma_horz_bslt4_ssse3()
1210 temp2 = _mm_unpacklo_epi8(temp1, zero); in ih264_deblk_luma_horz_bslt4_ssse3()
1213 temp2 = _mm_cmpgt_epi16(Alpha_8x16, temp2); in ih264_deblk_luma_horz_bslt4_ssse3()
1216 flag1_16x8 = _mm_packs_epi16(temp2, temp1); in ih264_deblk_luma_horz_bslt4_ssse3()
1221 temp2 = _mm_subs_epu8(q1_16x8, q0_16x8); in ih264_deblk_luma_horz_bslt4_ssse3()
1222 temp1 = _mm_add_epi8(temp1, temp2); in ih264_deblk_luma_horz_bslt4_ssse3()
1224 temp2 = _mm_unpacklo_epi8(temp1, zero); in ih264_deblk_luma_horz_bslt4_ssse3()
1227 temp2 = _mm_cmpgt_epi16(Beta_8x16, temp2); in ih264_deblk_luma_horz_bslt4_ssse3()
1230 flag2_16x8 = _mm_packs_epi16(temp2, temp1); in ih264_deblk_luma_horz_bslt4_ssse3()
1236 temp2 = _mm_subs_epu8(p1_16x8, p0_16x8); in ih264_deblk_luma_horz_bslt4_ssse3()
1237 temp1 = _mm_add_epi8(temp1, temp2); in ih264_deblk_luma_horz_bslt4_ssse3()
1239 temp2 = _mm_unpacklo_epi8(temp1, zero); in ih264_deblk_luma_horz_bslt4_ssse3()
1242 temp2 = _mm_cmpgt_epi16(Beta_8x16, temp2); in ih264_deblk_luma_horz_bslt4_ssse3()
1245 flag2_16x8 = _mm_packs_epi16(temp2, temp1); in ih264_deblk_luma_horz_bslt4_ssse3()
1252 temp2 = _mm_subs_epu8(p2_16x8, p0_16x8); in ih264_deblk_luma_horz_bslt4_ssse3()
1253 temp1 = _mm_add_epi8(temp1, temp2); in ih264_deblk_luma_horz_bslt4_ssse3()
1255 temp2 = _mm_unpacklo_epi8(temp1, zero); in ih264_deblk_luma_horz_bslt4_ssse3()
1257 temp2 = _mm_cmpgt_epi16(Beta_8x16, temp2); in ih264_deblk_luma_horz_bslt4_ssse3()
1260 flag2_16x8 = _mm_packs_epi16(temp2, temp1); in ih264_deblk_luma_horz_bslt4_ssse3()
1263 temp2 = _mm_subs_epi16(zero, temp2); in ih264_deblk_luma_horz_bslt4_ssse3()
1266 C_8x16 = _mm_add_epi16(C0_8x16, temp2); in ih264_deblk_luma_horz_bslt4_ssse3()
1271 temp2 = _mm_subs_epu8(q2_16x8, q0_16x8); in ih264_deblk_luma_horz_bslt4_ssse3()
1272 temp1 = _mm_add_epi8(temp1, temp2); in ih264_deblk_luma_horz_bslt4_ssse3()
1274 temp2 = _mm_unpacklo_epi8(temp1, zero); in ih264_deblk_luma_horz_bslt4_ssse3()
1276 temp2 = _mm_cmpgt_epi16(Beta_8x16, temp2); in ih264_deblk_luma_horz_bslt4_ssse3()
1279 flag3_16x8 = _mm_packs_epi16(temp2, temp1); in ih264_deblk_luma_horz_bslt4_ssse3()
1282 temp2 = _mm_subs_epi16(zero, temp2); in ih264_deblk_luma_horz_bslt4_ssse3()
1285 C_8x16 = _mm_add_epi16(C_8x16, temp2); in ih264_deblk_luma_horz_bslt4_ssse3()
1291 temp2 = _mm_subs_epi16(_mm_unpacklo_epi8(p1_16x8, zero), in ih264_deblk_luma_horz_bslt4_ssse3()
1294 temp1 = _mm_add_epi16(temp1, temp2); in ih264_deblk_luma_horz_bslt4_ssse3()
1300 temp2 = _mm_subs_epi16(_mm_unpackhi_epi8(p1_16x8, zero), in ih264_deblk_luma_horz_bslt4_ssse3()
1303 temp1 = _mm_add_epi16(temp1, temp2); in ih264_deblk_luma_horz_bslt4_ssse3()
1315 temp2 = _mm_add_epi16(_mm_unpackhi_epi8(p0_16x8, zero), in_macro_hi_16x8); in ih264_deblk_luma_horz_bslt4_ssse3()
1317 temp1 = _mm_packus_epi16(temp1, temp2); in ih264_deblk_luma_horz_bslt4_ssse3()
1320 temp2 = _mm_and_si128(p0_16x8, in ih264_deblk_luma_horz_bslt4_ssse3()
1323 temp1 = _mm_add_epi8(temp1, temp2); in ih264_deblk_luma_horz_bslt4_ssse3()
1328 temp2 = _mm_sub_epi16(_mm_unpackhi_epi8(q0_16x8, zero), in_macro_hi_16x8); in ih264_deblk_luma_horz_bslt4_ssse3()
1330 temp1 = _mm_packus_epi16(temp1, temp2); in ih264_deblk_luma_horz_bslt4_ssse3()
1333 temp2 = _mm_and_si128(q0_16x8, in ih264_deblk_luma_horz_bslt4_ssse3()
1336 temp1 = _mm_add_epi8(temp1, temp2); in ih264_deblk_luma_horz_bslt4_ssse3()
1342 temp2 = _mm_slli_epi16(_mm_unpacklo_epi8(p1_16x8, zero), 1); in ih264_deblk_luma_horz_bslt4_ssse3()
1344 temp2 = _mm_subs_epi16(_mm_unpacklo_epi8(p2_16x8, zero), temp2); in ih264_deblk_luma_horz_bslt4_ssse3()
1345 temp2 = _mm_add_epi16(temp1, temp2); in ih264_deblk_luma_horz_bslt4_ssse3()
1346 in_macro_16x8 = _mm_srai_epi16(temp2, 1); in ih264_deblk_luma_horz_bslt4_ssse3()
1350 temp2 = _mm_slli_epi16(_mm_unpackhi_epi8(p1_16x8, zero), 1); in ih264_deblk_luma_horz_bslt4_ssse3()
1352 temp2 = _mm_subs_epi16(_mm_unpackhi_epi8(p2_16x8, zero), temp2); in ih264_deblk_luma_horz_bslt4_ssse3()
1353 temp2 = _mm_add_epi16(temp1, temp2); in ih264_deblk_luma_horz_bslt4_ssse3()
1354 in_macro_hi_16x8 = _mm_srai_epi16(temp2, 1); in ih264_deblk_luma_horz_bslt4_ssse3()
1364 temp2 = _mm_add_epi16(_mm_unpackhi_epi8(p1_16x8, zero), in_macro_hi_16x8); in ih264_deblk_luma_horz_bslt4_ssse3()
1366 temp1 = _mm_packus_epi16(temp1, temp2); in ih264_deblk_luma_horz_bslt4_ssse3()
1369 temp2 = _mm_and_si128(p1_16x8, in ih264_deblk_luma_horz_bslt4_ssse3()
1371 temp1 = _mm_add_epi8(temp1, temp2); in ih264_deblk_luma_horz_bslt4_ssse3()
1377 temp2 = _mm_slli_epi16(_mm_unpacklo_epi8(q1_16x8, zero), 1); in ih264_deblk_luma_horz_bslt4_ssse3()
1379 temp2 = _mm_subs_epi16(_mm_unpacklo_epi8(q2_16x8, zero), temp2); in ih264_deblk_luma_horz_bslt4_ssse3()
1380 temp2 = _mm_add_epi16(temp1, temp2); in ih264_deblk_luma_horz_bslt4_ssse3()
1381 in_macro_16x8 = _mm_srai_epi16(temp2, 1); in ih264_deblk_luma_horz_bslt4_ssse3()
1385 temp2 = _mm_slli_epi16(_mm_unpackhi_epi8(q1_16x8, zero), 1); in ih264_deblk_luma_horz_bslt4_ssse3()
1387 temp2 = _mm_subs_epi16(_mm_unpackhi_epi8(q2_16x8, zero), temp2); in ih264_deblk_luma_horz_bslt4_ssse3()
1388 temp2 = _mm_add_epi16(temp1, temp2); in ih264_deblk_luma_horz_bslt4_ssse3()
1389 in_macro_hi_16x8 = _mm_srai_epi16(temp2, 1); in ih264_deblk_luma_horz_bslt4_ssse3()
1399 temp2 = _mm_add_epi16(_mm_unpackhi_epi8(q1_16x8, zero), in_macro_hi_16x8); in ih264_deblk_luma_horz_bslt4_ssse3()
1401 temp1 = _mm_packus_epi16(temp1, temp2); in ih264_deblk_luma_horz_bslt4_ssse3()
1404 temp2 = _mm_and_si128(q1_16x8, in ih264_deblk_luma_horz_bslt4_ssse3()
1406 temp1 = _mm_add_epi8(temp1, temp2); in ih264_deblk_luma_horz_bslt4_ssse3()
1456 __m128i temp1, temp2, temp3, temp4, temp5, temp6; in ih264_deblk_luma_vert_bs4_mbaff_ssse3() local
1475 temp2 = _mm_unpacklo_epi8(line3, line4); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1479 line1 = _mm_unpacklo_epi16(temp1, temp2); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1480 line2 = _mm_unpackhi_epi16(temp1, temp2); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1500 temp2 = _mm_subs_epu8(p0_16x8, q0_16x8); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1501 temp1 = _mm_add_epi8(temp1, temp2); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1503 temp2 = _mm_unpacklo_epi8(temp1, zero); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1506 temp2 = _mm_cmpgt_epi16(Alpha_8x16, temp2); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1509 flag1_16x8 = _mm_packs_epi16(temp2, temp1); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1513 temp2 = _mm_subs_epu8(q1_16x8, q0_16x8); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1514 temp1 = _mm_add_epi8(temp1, temp2); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1516 temp2 = _mm_unpacklo_epi8(temp1, zero); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1519 temp2 = _mm_cmpgt_epi16(Beta_8x16, temp2); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1522 flag2_16x8 = _mm_packs_epi16(temp2, temp1); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1528 temp2 = _mm_subs_epu8(p1_16x8, p0_16x8); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1529 temp1 = _mm_add_epi8(temp1, temp2); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1531 temp2 = _mm_unpacklo_epi8(temp1, zero); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1534 temp2 = _mm_cmpgt_epi16(Beta_8x16, temp2); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1537 flag2_16x8 = _mm_packs_epi16(temp2, temp1); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1544 temp2 = _mm_subs_epu8(q0_16x8, p0_16x8); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1545 temp1 = _mm_add_epi8(temp1, temp2); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1549 temp2 = _mm_unpacklo_epi8(temp1, zero); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1551 temp2 = _mm_cmpgt_epi16(Alpha_8x16, temp2); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1554 flag2_16x8 = _mm_packs_epi16(temp2, temp1); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1559 temp2 = _mm_subs_epu8(p2_16x8, p0_16x8); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1560 temp1 = _mm_add_epi8(temp1, temp2); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1562 temp2 = _mm_unpacklo_epi8(temp1, zero); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1564 temp2 = _mm_cmpgt_epi16(Beta_8x16, temp2); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1567 flag3_16x8 = _mm_packs_epi16(temp2, temp1); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1572 temp2 = _mm_subs_epu8(q2_16x8, q0_16x8); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1573 temp1 = _mm_add_epi8(temp1, temp2); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1575 temp2 = _mm_unpacklo_epi8(temp1, zero); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1577 temp2 = _mm_cmpgt_epi16(Beta_8x16, temp2); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1580 flag4_16x8 = _mm_packs_epi16(temp2, temp1); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1595 temp2 = _mm_add_epi16(p1_8x16, q0_8x16); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1597 temp6 = _mm_add_epi16(temp2, const_val2_16x8); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1601 temp2 = _mm_add_epi16(temp6, temp4); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1603 q0_16x8_1 = _mm_srai_epi16(temp2, 2); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1609 temp2 = _mm_add_epi16(temp5, q2_8x16); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1611 q1_16x8_2 = _mm_srai_epi16(temp2, 2); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1615 temp2 = _mm_add_epi16(temp4, q2_8x16); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1617 temp2 = _mm_add_epi16(temp2, p1_8x16); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1621 temp2 = _mm_add_epi16(temp2, temp3); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1623 temp2 = _mm_add_epi16(temp2, _mm_set1_epi16(4)); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1625 q0_16x8_2 = _mm_srai_epi16(temp2, 3); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1629 temp2 = _mm_add_epi16(temp5, const_val2_16x8); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1637 temp2 = _mm_add_epi16(temp2, temp4); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1639 temp2 = _mm_add_epi16(temp2, temp6); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1641 q2_16x8_2 = _mm_srai_epi16(temp2, 3); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1700 temp2 = _mm_unpacklo_epi8(p1_16x8, p0_16x8); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1704 p3_8x16 = _mm_unpacklo_epi16(temp1, temp2); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1705 p2_8x16 = _mm_unpackhi_epi16(temp1, temp2); in ih264_deblk_luma_vert_bs4_mbaff_ssse3()
1772 __m128i temp1, temp2, temp3, temp4; in ih264_deblk_luma_vert_bslt4_mbaff_ssse3() local
1792 temp2 = _mm_unpacklo_epi8(line3, line4); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1796 line1 = _mm_unpacklo_epi16(temp1, temp2); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1797 line2 = _mm_unpackhi_epi16(temp1, temp2); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1802 temp2 = _mm_unpackhi_epi32(line1, line3); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1810 p1_16x8 = _mm_unpacklo_epi64(temp2, zero); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1811 p0_16x8 = _mm_unpackhi_epi64(temp2, zero); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1839 temp2 = _mm_subs_epu8(p0_16x8, q0_16x8); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1840 temp1 = _mm_add_epi8(temp1, temp2); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1842 temp2 = _mm_unpacklo_epi8(temp1, zero); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1843 temp2 = _mm_cmpgt_epi16(Alpha_8x16, temp2); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1845 flag1_16x8 = _mm_packs_epi16(temp2, zero); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1850 temp2 = _mm_subs_epu8(q1_16x8, q0_16x8); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1851 temp1 = _mm_add_epi8(temp1, temp2); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1853 temp2 = _mm_unpacklo_epi8(temp1, zero); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1854 temp2 = _mm_cmpgt_epi16(Beta_8x16, temp2); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1856 flag2_16x8 = _mm_packs_epi16(temp2, zero); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1861 temp2 = _mm_subs_epu8(p1_16x8, p0_16x8); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1862 temp1 = _mm_add_epi8(temp1, temp2); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1864 temp2 = _mm_unpacklo_epi8(temp1, zero); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1865 temp2 = _mm_cmpgt_epi16(Beta_8x16, temp2); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1867 flag2_16x8 = _mm_packs_epi16(temp2, zero); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1874 temp2 = _mm_subs_epu8(p2_16x8, p0_16x8); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1875 temp1 = _mm_add_epi8(temp1, temp2); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1877 temp2 = _mm_unpacklo_epi8(temp1, zero); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1878 temp2 = _mm_cmpgt_epi16(Beta_8x16, temp2); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1880 flag2_16x8 = _mm_packs_epi16(temp2, zero); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1883 temp2 = _mm_subs_epi16(zero, temp2); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1885 C_8x16 = _mm_add_epi16(C0_8x16, temp2); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1889 temp2 = _mm_subs_epu8(q2_16x8, q0_16x8); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1890 temp1 = _mm_add_epi8(temp1, temp2); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1892 temp2 = _mm_unpacklo_epi8(temp1, zero); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1893 temp2 = _mm_cmpgt_epi16(Beta_8x16, temp2); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1895 flag3_16x8 = _mm_packs_epi16(temp2, zero); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1898 temp2 = _mm_subs_epi16(zero, temp2); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1900 C_8x16 = _mm_add_epi16(C_8x16, temp2); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1905 temp2 = _mm_subs_epi16(_mm_unpacklo_epi8(p1_16x8, zero), in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1908 temp1 = _mm_add_epi16(temp1, temp2); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1941 temp2 = _mm_slli_epi16(_mm_unpacklo_epi8(p1_16x8, zero), 1); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1943 temp2 = _mm_subs_epi16(_mm_unpacklo_epi8(p2_16x8, zero), temp2); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1944 temp2 = _mm_add_epi16(temp1, temp2); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1945 in_macro_16x8 = _mm_srai_epi16(temp2, 1); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1964 temp2 = _mm_slli_epi16(_mm_unpacklo_epi8(q1_16x8, zero), 1); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1966 temp2 = _mm_subs_epi16(_mm_unpacklo_epi8(q2_16x8, zero), temp2); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1967 temp2 = _mm_add_epi16(temp1, temp2); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1968 in_macro_16x8 = _mm_srai_epi16(temp2, 1); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1985 temp2 = _mm_unpacklo_epi8(p1_16x8, p0_16x8_1); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1989 line7 = _mm_unpacklo_epi16(temp1, temp2); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1990 temp1 = _mm_unpackhi_epi16(temp1, temp2); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1992 temp2 = _mm_unpackhi_epi16(temp3, temp4); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
1998 line5 = _mm_unpacklo_epi32(temp1, temp2); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()
2000 line7 = _mm_unpackhi_epi32(temp1, temp2); in ih264_deblk_luma_vert_bslt4_mbaff_ssse3()