Lines Matching refs:summ
608 __m128i summ, temp; in FLAC__lpc_restore_signal_wide_intrin_sse41() local
633 summ = _mm_mul_epi32(dat[5], qlp[5]) ; in FLAC__lpc_restore_signal_wide_intrin_sse41()
634 summ = _mm_add_epi64(summ, _mm_mul_epi32(dat[4], qlp[4])); in FLAC__lpc_restore_signal_wide_intrin_sse41()
635 summ = _mm_add_epi64(summ, _mm_mul_epi32(dat[3], qlp[3])); in FLAC__lpc_restore_signal_wide_intrin_sse41()
636 summ = _mm_add_epi64(summ, _mm_mul_epi32(dat[2], qlp[2])); in FLAC__lpc_restore_signal_wide_intrin_sse41()
637 summ = _mm_add_epi64(summ, _mm_mul_epi32(dat[1], qlp[1])); in FLAC__lpc_restore_signal_wide_intrin_sse41()
638 summ = _mm_add_epi64(summ, _mm_mul_epi32(dat[0], qlp[0])); in FLAC__lpc_restore_signal_wide_intrin_sse41()
640 summ = _mm_add_epi64(summ, _mm_srli_si128(summ, 8)); // ?_64 sum_64 in FLAC__lpc_restore_signal_wide_intrin_sse41()
641 …summ = _mm_srl_epi64(summ, cnt); // ?_64 (sum >> lp_quantization)_64 == ?_32 ?_32 ?_32 … in FLAC__lpc_restore_signal_wide_intrin_sse41()
643 temp = _mm_add_epi32(temp, summ); // ? ? ? d[i] in FLAC__lpc_restore_signal_wide_intrin_sse41()
654 summ = _mm_mul_epi32(dat[5], qlp[5]) ; in FLAC__lpc_restore_signal_wide_intrin_sse41()
655 summ = _mm_add_epi64(summ, _mm_mul_epi32(dat[4], qlp[4])); in FLAC__lpc_restore_signal_wide_intrin_sse41()
656 summ = _mm_add_epi64(summ, _mm_mul_epi32(dat[3], qlp[3])); in FLAC__lpc_restore_signal_wide_intrin_sse41()
657 summ = _mm_add_epi64(summ, _mm_mul_epi32(dat[2], qlp[2])); in FLAC__lpc_restore_signal_wide_intrin_sse41()
658 summ = _mm_add_epi64(summ, _mm_mul_epi32(dat[1], qlp[1])); in FLAC__lpc_restore_signal_wide_intrin_sse41()
659 summ = _mm_add_epi64(summ, _mm_mul_epi32(dat[0], qlp[0])); in FLAC__lpc_restore_signal_wide_intrin_sse41()
661 summ = _mm_add_epi64(summ, _mm_srli_si128(summ, 8)); // ?_64 sum_64 in FLAC__lpc_restore_signal_wide_intrin_sse41()
662 …summ = _mm_srl_epi64(summ, cnt); // ?_64 (sum >> lp_quantization)_64 == ?_32 ?_32 ?_32 … in FLAC__lpc_restore_signal_wide_intrin_sse41()
664 temp = _mm_add_epi32(temp, summ); // ? ? ? d[i] in FLAC__lpc_restore_signal_wide_intrin_sse41()
670 __m128i summ, temp; in FLAC__lpc_restore_signal_wide_intrin_sse41() local
692 summ = _mm_mul_epi32(dat[4], qlp[4]) ; in FLAC__lpc_restore_signal_wide_intrin_sse41()
693 summ = _mm_add_epi64(summ, _mm_mul_epi32(dat[3], qlp[3])); in FLAC__lpc_restore_signal_wide_intrin_sse41()
694 summ = _mm_add_epi64(summ, _mm_mul_epi32(dat[2], qlp[2])); in FLAC__lpc_restore_signal_wide_intrin_sse41()
695 summ = _mm_add_epi64(summ, _mm_mul_epi32(dat[1], qlp[1])); in FLAC__lpc_restore_signal_wide_intrin_sse41()
696 summ = _mm_add_epi64(summ, _mm_mul_epi32(dat[0], qlp[0])); in FLAC__lpc_restore_signal_wide_intrin_sse41()
698 summ = _mm_add_epi64(summ, _mm_srli_si128(summ, 8)); in FLAC__lpc_restore_signal_wide_intrin_sse41()
699 summ = _mm_srl_epi64(summ, cnt); in FLAC__lpc_restore_signal_wide_intrin_sse41()
701 temp = _mm_add_epi32(temp, summ); in FLAC__lpc_restore_signal_wide_intrin_sse41()
711 summ = _mm_mul_epi32(dat[4], qlp[4]) ; in FLAC__lpc_restore_signal_wide_intrin_sse41()
712 summ = _mm_add_epi64(summ, _mm_mul_epi32(dat[3], qlp[3])); in FLAC__lpc_restore_signal_wide_intrin_sse41()
713 summ = _mm_add_epi64(summ, _mm_mul_epi32(dat[2], qlp[2])); in FLAC__lpc_restore_signal_wide_intrin_sse41()
714 summ = _mm_add_epi64(summ, _mm_mul_epi32(dat[1], qlp[1])); in FLAC__lpc_restore_signal_wide_intrin_sse41()
715 summ = _mm_add_epi64(summ, _mm_mul_epi32(dat[0], qlp[0])); in FLAC__lpc_restore_signal_wide_intrin_sse41()
717 summ = _mm_add_epi64(summ, _mm_srli_si128(summ, 8)); in FLAC__lpc_restore_signal_wide_intrin_sse41()
718 summ = _mm_srl_epi64(summ, cnt); in FLAC__lpc_restore_signal_wide_intrin_sse41()
720 temp = _mm_add_epi32(temp, summ); in FLAC__lpc_restore_signal_wide_intrin_sse41()
728 __m128i summ, temp; in FLAC__lpc_restore_signal_wide_intrin_sse41() local
747 summ = _mm_mul_epi32(dat[3], qlp[3]) ; in FLAC__lpc_restore_signal_wide_intrin_sse41()
748 summ = _mm_add_epi64(summ, _mm_mul_epi32(dat[2], qlp[2])); in FLAC__lpc_restore_signal_wide_intrin_sse41()
749 summ = _mm_add_epi64(summ, _mm_mul_epi32(dat[1], qlp[1])); in FLAC__lpc_restore_signal_wide_intrin_sse41()
750 summ = _mm_add_epi64(summ, _mm_mul_epi32(dat[0], qlp[0])); in FLAC__lpc_restore_signal_wide_intrin_sse41()
752 summ = _mm_add_epi64(summ, _mm_srli_si128(summ, 8)); in FLAC__lpc_restore_signal_wide_intrin_sse41()
753 summ = _mm_srl_epi64(summ, cnt); in FLAC__lpc_restore_signal_wide_intrin_sse41()
755 temp = _mm_add_epi32(temp, summ); in FLAC__lpc_restore_signal_wide_intrin_sse41()
764 summ = _mm_mul_epi32(dat[3], qlp[3]) ; in FLAC__lpc_restore_signal_wide_intrin_sse41()
765 summ = _mm_add_epi64(summ, _mm_mul_epi32(dat[2], qlp[2])); in FLAC__lpc_restore_signal_wide_intrin_sse41()
766 summ = _mm_add_epi64(summ, _mm_mul_epi32(dat[1], qlp[1])); in FLAC__lpc_restore_signal_wide_intrin_sse41()
767 summ = _mm_add_epi64(summ, _mm_mul_epi32(dat[0], qlp[0])); in FLAC__lpc_restore_signal_wide_intrin_sse41()
769 summ = _mm_add_epi64(summ, _mm_srli_si128(summ, 8)); in FLAC__lpc_restore_signal_wide_intrin_sse41()
770 summ = _mm_srl_epi64(summ, cnt); in FLAC__lpc_restore_signal_wide_intrin_sse41()
772 temp = _mm_add_epi32(temp, summ); in FLAC__lpc_restore_signal_wide_intrin_sse41()
778 __m128i summ, temp; in FLAC__lpc_restore_signal_wide_intrin_sse41() local
794 summ = _mm_mul_epi32(dat[2], qlp[2]) ; in FLAC__lpc_restore_signal_wide_intrin_sse41()
795 summ = _mm_add_epi64(summ, _mm_mul_epi32(dat[1], qlp[1])); in FLAC__lpc_restore_signal_wide_intrin_sse41()
796 summ = _mm_add_epi64(summ, _mm_mul_epi32(dat[0], qlp[0])); in FLAC__lpc_restore_signal_wide_intrin_sse41()
798 summ = _mm_add_epi64(summ, _mm_srli_si128(summ, 8)); in FLAC__lpc_restore_signal_wide_intrin_sse41()
799 summ = _mm_srl_epi64(summ, cnt); in FLAC__lpc_restore_signal_wide_intrin_sse41()
801 temp = _mm_add_epi32(temp, summ); in FLAC__lpc_restore_signal_wide_intrin_sse41()
809 summ = _mm_mul_epi32(dat[2], qlp[2]) ; in FLAC__lpc_restore_signal_wide_intrin_sse41()
810 summ = _mm_add_epi64(summ, _mm_mul_epi32(dat[1], qlp[1])); in FLAC__lpc_restore_signal_wide_intrin_sse41()
811 summ = _mm_add_epi64(summ, _mm_mul_epi32(dat[0], qlp[0])); in FLAC__lpc_restore_signal_wide_intrin_sse41()
813 summ = _mm_add_epi64(summ, _mm_srli_si128(summ, 8)); in FLAC__lpc_restore_signal_wide_intrin_sse41()
814 summ = _mm_srl_epi64(summ, cnt); in FLAC__lpc_restore_signal_wide_intrin_sse41()
816 temp = _mm_add_epi32(temp, summ); in FLAC__lpc_restore_signal_wide_intrin_sse41()
824 __m128i summ, temp; in FLAC__lpc_restore_signal_wide_intrin_sse41() local
837 summ = _mm_mul_epi32(dat[1], qlp[1]) ; in FLAC__lpc_restore_signal_wide_intrin_sse41()
838 summ = _mm_add_epi64(summ, _mm_mul_epi32(dat[0], qlp[0])); in FLAC__lpc_restore_signal_wide_intrin_sse41()
840 summ = _mm_add_epi64(summ, _mm_srli_si128(summ, 8)); in FLAC__lpc_restore_signal_wide_intrin_sse41()
841 summ = _mm_srl_epi64(summ, cnt); in FLAC__lpc_restore_signal_wide_intrin_sse41()
843 temp = _mm_add_epi32(temp, summ); in FLAC__lpc_restore_signal_wide_intrin_sse41()
850 summ = _mm_mul_epi32(dat[1], qlp[1]) ; in FLAC__lpc_restore_signal_wide_intrin_sse41()
851 summ = _mm_add_epi64(summ, _mm_mul_epi32(dat[0], qlp[0])); in FLAC__lpc_restore_signal_wide_intrin_sse41()
853 summ = _mm_add_epi64(summ, _mm_srli_si128(summ, 8)); in FLAC__lpc_restore_signal_wide_intrin_sse41()
854 summ = _mm_srl_epi64(summ, cnt); in FLAC__lpc_restore_signal_wide_intrin_sse41()
856 temp = _mm_add_epi32(temp, summ); in FLAC__lpc_restore_signal_wide_intrin_sse41()
863 __m128i summ, temp; in FLAC__lpc_restore_signal_wide_intrin_sse41() local
869 summ = _mm_mul_epi32(dat0, qlp0) ; in FLAC__lpc_restore_signal_wide_intrin_sse41()
871 summ = _mm_add_epi64(summ, _mm_srli_si128(summ, 8)); in FLAC__lpc_restore_signal_wide_intrin_sse41()
872 summ = _mm_srl_epi64(summ, cnt); in FLAC__lpc_restore_signal_wide_intrin_sse41()
874 temp = _mm_add_epi32(temp, summ); in FLAC__lpc_restore_signal_wide_intrin_sse41()
880 summ = _mm_mul_epi32(dat0, qlp0) ; in FLAC__lpc_restore_signal_wide_intrin_sse41()
882 summ = _mm_add_epi64(summ, _mm_srli_si128(summ, 8)); in FLAC__lpc_restore_signal_wide_intrin_sse41()
883 summ = _mm_srl_epi64(summ, cnt); in FLAC__lpc_restore_signal_wide_intrin_sse41()
885 temp = _mm_add_epi32(temp, summ); in FLAC__lpc_restore_signal_wide_intrin_sse41()
891 __m128i summ, temp; in FLAC__lpc_restore_signal_wide_intrin_sse41() local
895 summ = _mm_mul_epi32(temp, qlp0); in FLAC__lpc_restore_signal_wide_intrin_sse41()
896 summ = _mm_srl_epi64(summ, cnt); in FLAC__lpc_restore_signal_wide_intrin_sse41()
898 temp = _mm_add_epi32(temp, summ); in FLAC__lpc_restore_signal_wide_intrin_sse41()
902 summ = _mm_mul_epi32(temp, qlp0) ; in FLAC__lpc_restore_signal_wide_intrin_sse41()
903 summ = _mm_srl_epi64(summ, cnt); in FLAC__lpc_restore_signal_wide_intrin_sse41()
905 temp = _mm_add_epi32(temp, summ); in FLAC__lpc_restore_signal_wide_intrin_sse41()
986 __m128i summ, mull; in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41() local
987 summ = _mm_mullo_epi32(q11, _mm_loadu_si128((const __m128i*)(data+i-12))); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
988 …m_mullo_epi32(q10, _mm_loadu_si128((const __m128i*)(data+i-11))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
989 …mm_mullo_epi32(q9, _mm_loadu_si128((const __m128i*)(data+i-10))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
990 …_mm_mullo_epi32(q8, _mm_loadu_si128((const __m128i*)(data+i-9))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
991 …_mm_mullo_epi32(q7, _mm_loadu_si128((const __m128i*)(data+i-8))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
992 …_mm_mullo_epi32(q6, _mm_loadu_si128((const __m128i*)(data+i-7))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
993 …_mm_mullo_epi32(q5, _mm_loadu_si128((const __m128i*)(data+i-6))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
994 …_mm_mullo_epi32(q4, _mm_loadu_si128((const __m128i*)(data+i-5))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
995 …_mm_mullo_epi32(q3, _mm_loadu_si128((const __m128i*)(data+i-4))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
996 …_mm_mullo_epi32(q2, _mm_loadu_si128((const __m128i*)(data+i-3))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
997 …_mm_mullo_epi32(q1, _mm_loadu_si128((const __m128i*)(data+i-2))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
998 …_mm_mullo_epi32(q0, _mm_loadu_si128((const __m128i*)(data+i-1))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
999 summ = _mm_sra_epi32(summ, cnt); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1000 …oreu_si128((__m128i*)(residual+i), _mm_sub_epi32(_mm_loadu_si128((const __m128i*)(data+i)), summ)); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1018 __m128i summ, mull; in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41() local
1019 summ = _mm_mullo_epi32(q10, _mm_loadu_si128((const __m128i*)(data+i-11))); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1020 …mm_mullo_epi32(q9, _mm_loadu_si128((const __m128i*)(data+i-10))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1021 …_mm_mullo_epi32(q8, _mm_loadu_si128((const __m128i*)(data+i-9))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1022 …_mm_mullo_epi32(q7, _mm_loadu_si128((const __m128i*)(data+i-8))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1023 …_mm_mullo_epi32(q6, _mm_loadu_si128((const __m128i*)(data+i-7))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1024 …_mm_mullo_epi32(q5, _mm_loadu_si128((const __m128i*)(data+i-6))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1025 …_mm_mullo_epi32(q4, _mm_loadu_si128((const __m128i*)(data+i-5))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1026 …_mm_mullo_epi32(q3, _mm_loadu_si128((const __m128i*)(data+i-4))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1027 …_mm_mullo_epi32(q2, _mm_loadu_si128((const __m128i*)(data+i-3))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1028 …_mm_mullo_epi32(q1, _mm_loadu_si128((const __m128i*)(data+i-2))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1029 …_mm_mullo_epi32(q0, _mm_loadu_si128((const __m128i*)(data+i-1))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1030 summ = _mm_sra_epi32(summ, cnt); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1031 …oreu_si128((__m128i*)(residual+i), _mm_sub_epi32(_mm_loadu_si128((const __m128i*)(data+i)), summ)); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1050 __m128i summ, mull; in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41() local
1051 summ = _mm_mullo_epi32(q9, _mm_loadu_si128((const __m128i*)(data+i-10))); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1052 …_mm_mullo_epi32(q8, _mm_loadu_si128((const __m128i*)(data+i-9))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1053 …_mm_mullo_epi32(q7, _mm_loadu_si128((const __m128i*)(data+i-8))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1054 …_mm_mullo_epi32(q6, _mm_loadu_si128((const __m128i*)(data+i-7))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1055 …_mm_mullo_epi32(q5, _mm_loadu_si128((const __m128i*)(data+i-6))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1056 …_mm_mullo_epi32(q4, _mm_loadu_si128((const __m128i*)(data+i-5))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1057 …_mm_mullo_epi32(q3, _mm_loadu_si128((const __m128i*)(data+i-4))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1058 …_mm_mullo_epi32(q2, _mm_loadu_si128((const __m128i*)(data+i-3))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1059 …_mm_mullo_epi32(q1, _mm_loadu_si128((const __m128i*)(data+i-2))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1060 …_mm_mullo_epi32(q0, _mm_loadu_si128((const __m128i*)(data+i-1))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1061 summ = _mm_sra_epi32(summ, cnt); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1062 …oreu_si128((__m128i*)(residual+i), _mm_sub_epi32(_mm_loadu_si128((const __m128i*)(data+i)), summ)); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1078 __m128i summ, mull; in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41() local
1079 summ = _mm_mullo_epi32(q8, _mm_loadu_si128((const __m128i*)(data+i-9))); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1080 …_mm_mullo_epi32(q7, _mm_loadu_si128((const __m128i*)(data+i-8))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1081 …_mm_mullo_epi32(q6, _mm_loadu_si128((const __m128i*)(data+i-7))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1082 …_mm_mullo_epi32(q5, _mm_loadu_si128((const __m128i*)(data+i-6))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1083 …_mm_mullo_epi32(q4, _mm_loadu_si128((const __m128i*)(data+i-5))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1084 …_mm_mullo_epi32(q3, _mm_loadu_si128((const __m128i*)(data+i-4))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1085 …_mm_mullo_epi32(q2, _mm_loadu_si128((const __m128i*)(data+i-3))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1086 …_mm_mullo_epi32(q1, _mm_loadu_si128((const __m128i*)(data+i-2))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1087 …_mm_mullo_epi32(q0, _mm_loadu_si128((const __m128i*)(data+i-1))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1088 summ = _mm_sra_epi32(summ, cnt); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1089 …oreu_si128((__m128i*)(residual+i), _mm_sub_epi32(_mm_loadu_si128((const __m128i*)(data+i)), summ)); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1108 __m128i summ, mull; in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41() local
1109 summ = _mm_mullo_epi32(q7, _mm_loadu_si128((const __m128i*)(data+i-8))); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1110 …_mm_mullo_epi32(q6, _mm_loadu_si128((const __m128i*)(data+i-7))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1111 …_mm_mullo_epi32(q5, _mm_loadu_si128((const __m128i*)(data+i-6))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1112 …_mm_mullo_epi32(q4, _mm_loadu_si128((const __m128i*)(data+i-5))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1113 …_mm_mullo_epi32(q3, _mm_loadu_si128((const __m128i*)(data+i-4))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1114 …_mm_mullo_epi32(q2, _mm_loadu_si128((const __m128i*)(data+i-3))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1115 …_mm_mullo_epi32(q1, _mm_loadu_si128((const __m128i*)(data+i-2))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1116 …_mm_mullo_epi32(q0, _mm_loadu_si128((const __m128i*)(data+i-1))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1117 summ = _mm_sra_epi32(summ, cnt); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1118 …oreu_si128((__m128i*)(residual+i), _mm_sub_epi32(_mm_loadu_si128((const __m128i*)(data+i)), summ)); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1132 __m128i summ, mull; in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41() local
1133 summ = _mm_mullo_epi32(q6, _mm_loadu_si128((const __m128i*)(data+i-7))); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1134 …_mm_mullo_epi32(q5, _mm_loadu_si128((const __m128i*)(data+i-6))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1135 …_mm_mullo_epi32(q4, _mm_loadu_si128((const __m128i*)(data+i-5))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1136 …_mm_mullo_epi32(q3, _mm_loadu_si128((const __m128i*)(data+i-4))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1137 …_mm_mullo_epi32(q2, _mm_loadu_si128((const __m128i*)(data+i-3))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1138 …_mm_mullo_epi32(q1, _mm_loadu_si128((const __m128i*)(data+i-2))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1139 …_mm_mullo_epi32(q0, _mm_loadu_si128((const __m128i*)(data+i-1))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1140 summ = _mm_sra_epi32(summ, cnt); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1141 …oreu_si128((__m128i*)(residual+i), _mm_sub_epi32(_mm_loadu_si128((const __m128i*)(data+i)), summ)); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1156 __m128i summ, mull; in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41() local
1157 summ = _mm_mullo_epi32(q5, _mm_loadu_si128((const __m128i*)(data+i-6))); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1158 …_mm_mullo_epi32(q4, _mm_loadu_si128((const __m128i*)(data+i-5))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1159 …_mm_mullo_epi32(q3, _mm_loadu_si128((const __m128i*)(data+i-4))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1160 …_mm_mullo_epi32(q2, _mm_loadu_si128((const __m128i*)(data+i-3))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1161 …_mm_mullo_epi32(q1, _mm_loadu_si128((const __m128i*)(data+i-2))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1162 …_mm_mullo_epi32(q0, _mm_loadu_si128((const __m128i*)(data+i-1))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1163 summ = _mm_sra_epi32(summ, cnt); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1164 …oreu_si128((__m128i*)(residual+i), _mm_sub_epi32(_mm_loadu_si128((const __m128i*)(data+i)), summ)); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1176 __m128i summ, mull; in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41() local
1177 summ = _mm_mullo_epi32(q4, _mm_loadu_si128((const __m128i*)(data+i-5))); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1178 …_mm_mullo_epi32(q3, _mm_loadu_si128((const __m128i*)(data+i-4))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1179 …_mm_mullo_epi32(q2, _mm_loadu_si128((const __m128i*)(data+i-3))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1180 …_mm_mullo_epi32(q1, _mm_loadu_si128((const __m128i*)(data+i-2))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1181 …_mm_mullo_epi32(q0, _mm_loadu_si128((const __m128i*)(data+i-1))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1182 summ = _mm_sra_epi32(summ, cnt); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1183 …oreu_si128((__m128i*)(residual+i), _mm_sub_epi32(_mm_loadu_si128((const __m128i*)(data+i)), summ)); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1198 __m128i summ, mull; in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41() local
1199 summ = _mm_mullo_epi32(q3, _mm_loadu_si128((const __m128i*)(data+i-4))); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1200 …_mm_mullo_epi32(q2, _mm_loadu_si128((const __m128i*)(data+i-3))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1201 …_mm_mullo_epi32(q1, _mm_loadu_si128((const __m128i*)(data+i-2))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1202 …_mm_mullo_epi32(q0, _mm_loadu_si128((const __m128i*)(data+i-1))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1203 summ = _mm_sra_epi32(summ, cnt); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1204 …oreu_si128((__m128i*)(residual+i), _mm_sub_epi32(_mm_loadu_si128((const __m128i*)(data+i)), summ)); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1214 __m128i summ, mull; in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41() local
1215 summ = _mm_mullo_epi32(q2, _mm_loadu_si128((const __m128i*)(data+i-3))); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1216 …_mm_mullo_epi32(q1, _mm_loadu_si128((const __m128i*)(data+i-2))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1217 …_mm_mullo_epi32(q0, _mm_loadu_si128((const __m128i*)(data+i-1))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1218 summ = _mm_sra_epi32(summ, cnt); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1219 …oreu_si128((__m128i*)(residual+i), _mm_sub_epi32(_mm_loadu_si128((const __m128i*)(data+i)), summ)); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1230 __m128i summ, mull; in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41() local
1231 summ = _mm_mullo_epi32(q1, _mm_loadu_si128((const __m128i*)(data+i-2))); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1232 …_mm_mullo_epi32(q0, _mm_loadu_si128((const __m128i*)(data+i-1))); summ = _mm_add_epi32(summ, mull); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1233 summ = _mm_sra_epi32(summ, cnt); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1234 …oreu_si128((__m128i*)(residual+i), _mm_sub_epi32(_mm_loadu_si128((const __m128i*)(data+i)), summ)); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1242 __m128i summ; in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41() local
1243 summ = _mm_mullo_epi32(q0, _mm_loadu_si128((const __m128i*)(data+i-1))); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1244 summ = _mm_sra_epi32(summ, cnt); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()
1245 …oreu_si128((__m128i*)(residual+i), _mm_sub_epi32(_mm_loadu_si128((const __m128i*)(data+i)), summ)); in FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41()