Mercurial > mplayer.hg

comparison postproc/postprocess.c @ 3031:86e1a0f4f0bc

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
cleanup precopy fewer lines from src to dst if possible speedup (due to cleanup of blockcopy)
author michael
date Tue, 20 Nov 2001 17:47:52 +0000
parents 71384f064a3e
children 6de073cf52b5
comparison
equal deleted inserted replaced
3030:1dbc569b6528 3031:86e1a0f4f0bc
60 split this huge file 60 split this huge file
61 border remover 61 border remover
62 optimize c versions 62 optimize c versions
63 try to unroll inner for(x=0 ... loop to avoid these damn if(x ... checks 63 try to unroll inner for(x=0 ... loop to avoid these damn if(x ... checks
64 smart blur 64 smart blur
65 commandline option for the deblock thresholds
65 ... 66 ...
66 */ 67 */
67 68
68 //Changelog: use the CVS log 69 //Changelog: use the CVS log
69 70
856 } 857 }
857 */ 858 */
858 #endif 859 #endif
859 } 860 }
860 861
861 /**
862 * Experimental Filter 1 (Horizontal)
863 * will not damage linear gradients
864 * Flat blocks should look like they where passed through the (1,1,2,2,4,2,2,1,1) 9-Tap filter
865 * can only smooth blocks at the expected locations (it cant smooth them if they did move)
866 * MMX2 version does correct clipping C version doesnt
867 * not identical with the vertical one
868 */
869 static inline void horizX1Filter(uint8_t *src, int stride, int QP)
870 {
871 int y;
872 static uint64_t *lut= NULL;
873 if(lut==NULL)
874 {
875 int i;
876 lut= (uint64_t*)memalign(8, 256*8);
877 for(i=0; i<256; i++)
878 {
879 int v= i < 128 ? 2*i : 2*(i-256);
880 /*
881 //Simulate 112242211 9-Tap filter
882 uint64_t a= (v/16) & 0xFF;
883 uint64_t b= (v/8) & 0xFF;
884 uint64_t c= (v/4) & 0xFF;
885 uint64_t d= (3*v/8) & 0xFF;
886 */
887 //Simulate piecewise linear interpolation
888 uint64_t a= (v/16) & 0xFF;
889 uint64_t b= (v*3/16) & 0xFF;
890 uint64_t c= (v*5/16) & 0xFF;
891 uint64_t d= (7*v/16) & 0xFF;
892 uint64_t A= (0x100 - a)&0xFF;
893 uint64_t B= (0x100 - b)&0xFF;
894 uint64_t C= (0x100 - c)&0xFF;
895 uint64_t D= (0x100 - c)&0xFF;
896
897 lut[i] = (a<<56) | (b<<48) | (c<<40) | (d<<32) |
898 (D<<24) | (C<<16) | (B<<8) | (A);
899 //lut[i] = (v<<32) | (v<<24);
900 }
901 }
902
903 #if 0
904 asm volatile(
905 "pxor %%mm7, %%mm7 \n\t" // 0
906 // "movq b80, %%mm6 \n\t" // MIN_SIGNED_BYTE
907 "leal (%0, %1), %%eax \n\t"
908 "leal (%%eax, %1, 4), %%ebx \n\t"
909
910 "movq b80, %%mm6 \n\t"
911 "movd pQPb, %%mm5 \n\t" // QP
912 "movq %%mm5, %%mm4 \n\t"
913 "paddusb %%mm5, %%mm5 \n\t" // 2QP
914 "paddusb %%mm5, %%mm4 \n\t" // 3QP
915 "pxor %%mm5, %%mm5 \n\t" // 0
916 "psubb %%mm4, %%mm5 \n\t" // -3QP
917 "por bm11111110, %%mm5 \n\t" // ...,FF,FF,-3QP
918 "psllq $24, %%mm5 \n\t"
919
920 // 0 1 2 3 4 5 6 7 8 9
921 // %0 eax eax+%1 eax+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1
922
923 #define HX1old(a) \
924 "movd " #a ", %%mm0 \n\t"\
925 "movd 4" #a ", %%mm1 \n\t"\
926 "punpckldq %%mm1, %%mm0 \n\t"\
927 "movq %%mm0, %%mm1 \n\t"\
928 "movq %%mm0, %%mm2 \n\t"\
929 "psrlq $8, %%mm1 \n\t"\
930 "psubusb %%mm1, %%mm2 \n\t"\
931 "psubusb %%mm0, %%mm1 \n\t"\
932 "por %%mm2, %%mm1 \n\t" /* p´x = |px - p(x+1)| */\
933 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\
934 "pshufw $0x00, %%mm1, %%mm3 \n\t" /* p´5 = |p1 - p2| */\
935 PAVGB(%%mm1, %%mm3) /* p´5 = (|p2-p1| + |p6-p5|)/2 */\
936 "psrlq $16, %%mm3 \n\t" /* p´3 = (|p2-p1| + |p6-p5|)/2 */\
937 "psubusb %%mm3, %%mm1 \n\t" /* |p3-p4|-(|p2-p1| + |p6-p5|)/2 */\
938 "paddb %%mm5, %%mm1 \n\t"\
939 "psubusb %%mm5, %%mm1 \n\t"\
940 PAVGB(%%mm7, %%mm1)\
941 "pxor %%mm2, %%mm1 \n\t"\
942 "psubb %%mm2, %%mm1 \n\t"\
943 "psrlq $24, %%mm1 \n\t"\
944 "movd %%mm1, %%ecx \n\t"\
945 "paddb %%mm6, %%mm0 \n\t"\
946 "paddsb (%3, %%ecx, 8), %%mm0 \n\t"\
947 "paddb %%mm6, %%mm0 \n\t"\
948 "movq %%mm0, " #a " \n\t"\
949
950 /*
951 HX1old((%0))
952 HX1old((%%eax))
953 HX1old((%%eax, %1))
954 HX1old((%%eax, %1, 2))
955 HX1old((%0, %1, 4))
956 HX1old((%%ebx))
957 HX1old((%%ebx, %1))
958 HX1old((%%ebx, %1, 2))
959 */
960
961 //FIXME add some comments, its unreadable ...
962 #define HX1b(a, c, b, d) \
963 "movd " #a ", %%mm0 \n\t"\
964 "movd 4" #a ", %%mm1 \n\t"\
965 "punpckldq %%mm1, %%mm0 \n\t"\
966 "movd " #b ", %%mm4 \n\t"\
967 "movq %%mm0, %%mm1 \n\t"\
968 "movq %%mm0, %%mm2 \n\t"\
969 "psrlq $8, %%mm1 \n\t"\
970 "movd 4" #b ", %%mm3 \n\t"\
971 "psubusb %%mm1, %%mm2 \n\t"\
972 "psubusb %%mm0, %%mm1 \n\t"\
973 "por %%mm2, %%mm1 \n\t" /* p´x = |px - p(x+1)| */\
974 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\
975 "punpckldq %%mm3, %%mm4 \n\t"\
976 "movq %%mm1, %%mm3 \n\t"\
977 "psllq $32, %%mm3 \n\t" /* p´5 = |p1 - p2| */\
978 PAVGB(%%mm1, %%mm3) /* p´5 = (|p2-p1| + |p6-p5|)/2 */\
979 "paddb %%mm6, %%mm0 \n\t"\
980 "psrlq $16, %%mm3 \n\t" /* p´3 = (|p2-p1| + |p6-p5|)/2 */\
981 "psubusb %%mm3, %%mm1 \n\t" /* |p3-p4|-(|p2-p1| + |p6-p5|)/2 */\
982 "movq %%mm4, %%mm3 \n\t"\
983 "paddb %%mm5, %%mm1 \n\t"\
984 "psubusb %%mm5, %%mm1 \n\t"\
985 "psrlq $8, %%mm3 \n\t"\
986 PAVGB(%%mm7, %%mm1)\
987 "pxor %%mm2, %%mm1 \n\t"\
988 "psubb %%mm2, %%mm1 \n\t"\
989 "movq %%mm4, %%mm2 \n\t"\
990 "psrlq $24, %%mm1 \n\t"\
991 "psubusb %%mm3, %%mm2 \n\t"\
992 "movd %%mm1, %%ecx \n\t"\
993 "psubusb %%mm4, %%mm3 \n\t"\
994 "paddsb (%2, %%ecx, 8), %%mm0 \n\t"\
995 "por %%mm2, %%mm3 \n\t" /* p´x = |px - p(x+1)| */\
996 "paddb %%mm6, %%mm0 \n\t"\
997 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\
998 "movq %%mm3, %%mm1 \n\t"\
999 "psllq $32, %%mm1 \n\t" /* p´5 = |p1 - p2| */\
1000 "movq %%mm0, " #a " \n\t"\
1001 PAVGB(%%mm3, %%mm1) /* p´5 = (|p2-p1| + |p6-p5|)/2 */\
1002 "paddb %%mm6, %%mm4 \n\t"\
1003 "psrlq $16, %%mm1 \n\t" /* p´3 = (|p2-p1| + |p6-p5|)/2 */\
1004 "psubusb %%mm1, %%mm3 \n\t" /* |p3-p4|-(|p2-p1| + |p6-p5|)/2 */\
1005 "paddb %%mm5, %%mm3 \n\t"\
1006 "psubusb %%mm5, %%mm3 \n\t"\
1007 PAVGB(%%mm7, %%mm3)\
1008 "pxor %%mm2, %%mm3 \n\t"\
1009 "psubb %%mm2, %%mm3 \n\t"\
1010 "psrlq $24, %%mm3 \n\t"\
1011 "movd " #c ", %%mm0 \n\t"\
1012 "movd 4" #c ", %%mm1 \n\t"\
1013 "punpckldq %%mm1, %%mm0 \n\t"\
1014 "paddb %%mm6, %%mm0 \n\t"\
1015 "paddsb (%2, %%ecx, 8), %%mm0 \n\t"\
1016 "paddb %%mm6, %%mm0 \n\t"\
1017 "movq %%mm0, " #c " \n\t"\
1018 "movd %%mm3, %%ecx \n\t"\
1019 "movd " #d ", %%mm0 \n\t"\
1020 "paddsb (%2, %%ecx, 8), %%mm4 \n\t"\
1021 "movd 4" #d ", %%mm1 \n\t"\
1022 "paddb %%mm6, %%mm4 \n\t"\
1023 "punpckldq %%mm1, %%mm0 \n\t"\
1024 "movq %%mm4, " #b " \n\t"\
1025 "paddb %%mm6, %%mm0 \n\t"\
1026 "paddsb (%2, %%ecx, 8), %%mm0 \n\t"\
1027 "paddb %%mm6, %%mm0 \n\t"\
1028 "movq %%mm0, " #d " \n\t"\
1029
1030 HX1b((%0),(%%eax),(%%eax, %1),(%%eax, %1, 2))
1031 HX1b((%0, %1, 4),(%%ebx),(%%ebx, %1),(%%ebx, %1, 2))
1032
1033
1034 :
1035 : "r" (src), "r" (stride), "r" (lut)
1036 : "%eax", "%ebx", "%ecx"
1037 );
1038 #else
1039
1040 //FIXME (has little in common with the mmx2 version)
1041 for(y=0; y<BLOCK_SIZE; y++)
1042 {
1043 int a= src[1] - src[2];
1044 int b= src[3] - src[4];
1045 int c= src[5] - src[6];
1046
1047 int d= MAX(ABS(b) - (ABS(a) + ABS(c))/2, 0);
1048
1049 if(d < QP)
1050 {
1051 int v = d * SIGN(-b);
1052
1053 src[1] +=v/8;
1054 src[2] +=v/4;
1055 src[3] +=3*v/8;
1056 src[4] -=3*v/8;
1057 src[5] -=v/4;
1058 src[6] -=v/8;
1059
1060 }
1061 src+=stride;
1062 }
1063 #endif
1064 }
1065
1066
1067 static inline void doVertDefFilter(uint8_t src[], int stride, int QP) 862 static inline void doVertDefFilter(uint8_t src[], int stride, int QP)
1068 { 863 {
1069 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW) 864 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1070 /* 865 /*
1071 uint8_t tmp[16]; 866 uint8_t tmp[16];
1637 1432
1638 src[l4]-= d; 1433 src[l4]-= d;
1639 src[l5]+= d; 1434 src[l5]+= d;
1640 } 1435 }
1641 src++; 1436 src++;
1642 }
1643 #endif
1644 }
1645
1646 //FIXME? |255-0| = 1
1647 /**
1648 * Check if the given 8x8 Block is mostly "flat"
1649 */
1650 static inline int isHorizDC(uint8_t src[], int stride)
1651 {
1652 // src++;
1653 int numEq= 0;
1654 #if 0
1655 asm volatile (
1656 // "int $3 \n\t"
1657 "leal (%1, %2), %%ecx \n\t"
1658 "leal (%%ecx, %2, 4), %%ebx \n\t"
1659 // 0 1 2 3 4 5 6 7 8 9
1660 // %1 ecx ecx+%2 ecx+2%2 %1+4%2 ebx ebx+%2 ebx+2%2 %1+8%2 ebx+4%2
1661 "movq b7E, %%mm7 \n\t" // mm7 = 0x7F
1662 "movq b7C, %%mm6 \n\t" // mm6 = 0x7D
1663 "pxor %%mm0, %%mm0 \n\t"
1664 "movl %1, %%eax \n\t"
1665 "andl $0x1F, %%eax \n\t"
1666 "cmpl $24, %%eax \n\t"
1667 "leal tempBlock, %%eax \n\t"
1668 "jb 1f \n\t"
1669
1670 #define HDC_CHECK_AND_CPY(src, dst) \
1671 "movd " #src ", %%mm2 \n\t"\
1672 "punpckldq 4" #src ", %%mm2 \n\t" /* (%1) */\
1673 "movq %%mm2, %%mm1 \n\t"\
1674 "psrlq $8, %%mm2 \n\t"\
1675 "psubb %%mm1, %%mm2 \n\t"\
1676 "paddb %%mm7, %%mm2 \n\t"\
1677 "pcmpgtb %%mm6, %%mm2 \n\t"\
1678 "paddb %%mm2, %%mm0 \n\t"\
1679 "movq %%mm1," #dst "(%%eax) \n\t"
1680
1681 HDC_CHECK_AND_CPY((%1),0)
1682 HDC_CHECK_AND_CPY((%%ecx),8)
1683 HDC_CHECK_AND_CPY((%%ecx, %2),16)
1684 HDC_CHECK_AND_CPY((%%ecx, %2, 2),24)
1685 HDC_CHECK_AND_CPY((%1, %2, 4),32)
1686 HDC_CHECK_AND_CPY((%%ebx),40)
1687 HDC_CHECK_AND_CPY((%%ebx, %2),48)
1688 HDC_CHECK_AND_CPY((%%ebx, %2, 2),56)
1689 "jmp 2f \n\t"
1690 "1: \n\t"
1691 // src does not cross a 32 byte cache line so dont waste time with alignment
1692 #define HDC_CHECK_AND_CPY2(src, dst) \
1693 "movq " #src ", %%mm2 \n\t"\
1694 "movq " #src ", %%mm1 \n\t"\
1695 "psrlq $8, %%mm2 \n\t"\
1696 "psubb %%mm1, %%mm2 \n\t"\
1697 "paddb %%mm7, %%mm2 \n\t"\
1698 "pcmpgtb %%mm6, %%mm2 \n\t"\
1699 "paddb %%mm2, %%mm0 \n\t"\
1700 "movq %%mm1," #dst "(%%eax) \n\t"
1701
1702 HDC_CHECK_AND_CPY2((%1),0)
1703 HDC_CHECK_AND_CPY2((%%ecx),8)
1704 HDC_CHECK_AND_CPY2((%%ecx, %2),16)
1705 HDC_CHECK_AND_CPY2((%%ecx, %2, 2),24)
1706 HDC_CHECK_AND_CPY2((%1, %2, 4),32)
1707 HDC_CHECK_AND_CPY2((%%ebx),40)
1708 HDC_CHECK_AND_CPY2((%%ebx, %2),48)
1709 HDC_CHECK_AND_CPY2((%%ebx, %2, 2),56)
1710 "2: \n\t"
1711 "psllq $8, %%mm0 \n\t" // remove dummy value
1712 "movq %%mm0, %%mm1 \n\t"
1713 "psrlw $8, %%mm0 \n\t"
1714 "paddb %%mm1, %%mm0 \n\t"
1715 "movq %%mm0, %%mm1 \n\t"
1716 "psrlq $16, %%mm0 \n\t"
1717 "paddb %%mm1, %%mm0 \n\t"
1718 "movq %%mm0, %%mm1 \n\t"
1719 "psrlq $32, %%mm0 \n\t"
1720 "paddb %%mm1, %%mm0 \n\t"
1721 "movd %%mm0, %0 \n\t"
1722 : "=r" (numEq)
1723 : "r" (src), "r" (stride)
1724 : "%eax", "%ebx", "%ecx"
1725 );
1726 // printf("%d\n", numEq);
1727 numEq= (256 - numEq) &0xFF;
1728 #else
1729 int y;
1730 for(y=0; y<BLOCK_SIZE; y++)
1731 {
1732 if(((src[0] - src[1] + 1) & 0xFFFF) < 3) numEq++;
1733 if(((src[1] - src[2] + 1) & 0xFFFF) < 3) numEq++;
1734 if(((src[2] - src[3] + 1) & 0xFFFF) < 3) numEq++;
1735 if(((src[3] - src[4] + 1) & 0xFFFF) < 3) numEq++;
1736 if(((src[4] - src[5] + 1) & 0xFFFF) < 3) numEq++;
1737 if(((src[5] - src[6] + 1) & 0xFFFF) < 3) numEq++;
1738 if(((src[6] - src[7] + 1) & 0xFFFF) < 3) numEq++;
1739 src+= stride;
1740 }
1741 #endif
1742 /* if(abs(numEq - asmEq) > 0)
1743 {
1744 // printf("\nasm:%d c:%d\n", asmEq, numEq);
1745 for(int y=0; y<8; y++)
1746 {
1747 for(int x=0; x<8; x++)
1748 {
1749 printf("%d ", src[x + y*stride]);
1750 }
1751 printf("\n");
1752 }
1753 }
1754 */
1755 // printf("%d\n", numEq);
1756 return numEq > hFlatnessThreshold;
1757 }
1758
1759 static inline int isHorizMinMaxOk(uint8_t src[], int stride, int QP)
1760 {
1761 if(abs(src[0] - src[7]) > 2*QP) return 0;
1762
1763 return 1;
1764 }
1765
1766 static inline void doHorizDefFilter(uint8_t dst[], int stride, int QP)
1767 {
1768 #if 0
1769 asm volatile(
1770 "leal (%0, %1), %%ecx \n\t"
1771 "leal (%%ecx, %1, 4), %%ebx \n\t"
1772 // 0 1 2 3 4 5 6 7 8 9
1773 // %0 ecx ecx+%1 ecx+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1
1774 "pxor %%mm7, %%mm7 \n\t"
1775 "movq bm00001000, %%mm6 \n\t"
1776 "movd %2, %%mm5 \n\t" // QP
1777 "movq %%mm5, %%mm4 \n\t"
1778 "paddusb %%mm5, %%mm5 \n\t" // 2QP
1779 "paddusb %%mm5, %%mm4 \n\t" // 3QP
1780 "psllq $24, %%mm4 \n\t"
1781 "pxor %%mm5, %%mm5 \n\t" // 0
1782 "psubb %%mm4, %%mm5 \n\t" // -QP
1783 "leal tempBlock, %%eax \n\t"
1784
1785 //FIXME? "unroll by 2" and mix
1786 #ifdef HAVE_MMX2
1787 #define HDF(src, dst) \
1788 "movq " #src "(%%eax), %%mm0 \n\t"\
1789 "movq " #src "(%%eax), %%mm1 \n\t"\
1790 "movq " #src "(%%eax), %%mm2 \n\t"\
1791 "psrlq $8, %%mm1 \n\t"\
1792 "psubusb %%mm1, %%mm2 \n\t"\
1793 "psubusb %%mm0, %%mm1 \n\t"\
1794 "por %%mm2, %%mm1 \n\t" /* p´x = |px - p(x+1)| */\
1795 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\
1796 "pshufw $0x00, %%mm1, %%mm3 \n\t" /* p´5 = |p1 - p2| */\
1797 "pminub %%mm1, %%mm3 \n\t" /* p´5 = min(|p2-p1|, |p6-p5|)*/\
1798 "psrlq $16, %%mm3 \n\t" /* p´3 = min(|p2-p1|, |p6-p5|)*/\
1799 "psubusb %%mm3, %%mm1 \n\t" /* |p3-p4|-min(|p1-p2|,|p5-p6|) */\
1800 "paddb %%mm5, %%mm1 \n\t"\
1801 "psubusb %%mm5, %%mm1 \n\t"\
1802 "psrlw $2, %%mm1 \n\t"\
1803 "pxor %%mm2, %%mm1 \n\t"\
1804 "psubb %%mm2, %%mm1 \n\t"\
1805 "pand %%mm6, %%mm1 \n\t"\
1806 "psubb %%mm1, %%mm0 \n\t"\
1807 "psllq $8, %%mm1 \n\t"\
1808 "paddb %%mm1, %%mm0 \n\t"\
1809 "movd %%mm0, " #dst" \n\t"\
1810 "psrlq $32, %%mm0 \n\t"\
1811 "movd %%mm0, 4" #dst" \n\t"
1812 #else
1813 #define HDF(src, dst)\
1814 "movq " #src "(%%eax), %%mm0 \n\t"\
1815 "movq %%mm0, %%mm1 \n\t"\
1816 "movq %%mm0, %%mm2 \n\t"\
1817 "psrlq $8, %%mm1 \n\t"\
1818 "psubusb %%mm1, %%mm2 \n\t"\
1819 "psubusb %%mm0, %%mm1 \n\t"\
1820 "por %%mm2, %%mm1 \n\t" /* p´x = |px - p(x+1)| */\
1821 "pcmpeqb %%mm7, %%mm2 \n\t" /* p´x = sgn[px - p(x+1)] */\
1822 "movq %%mm1, %%mm3 \n\t"\
1823 "psllq $32, %%mm3 \n\t"\
1824 "movq %%mm3, %%mm4 \n\t"\
1825 "psubusb %%mm1, %%mm4 \n\t"\
1826 "psubb %%mm4, %%mm3 \n\t"\
1827 "psrlq $16, %%mm3 \n\t" /* p´3 = min(|p2-p1|, |p6-p5|)*/\
1828 "psubusb %%mm3, %%mm1 \n\t" /* |p3-p4|-min(|p1-p2|,|p5,ü6|) */\
1829 "paddb %%mm5, %%mm1 \n\t"\
1830 "psubusb %%mm5, %%mm1 \n\t"\
1831 "psrlw $2, %%mm1 \n\t"\
1832 "pxor %%mm2, %%mm1 \n\t"\
1833 "psubb %%mm2, %%mm1 \n\t"\
1834 "pand %%mm6, %%mm1 \n\t"\
1835 "psubb %%mm1, %%mm0 \n\t"\
1836 "psllq $8, %%mm1 \n\t"\
1837 "paddb %%mm1, %%mm0 \n\t"\
1838 "movd %%mm0, " #dst " \n\t"\
1839 "psrlq $32, %%mm0 \n\t"\
1840 "movd %%mm0, 4" #dst " \n\t"
1841 #endif
1842 HDF(0,(%0))
1843 HDF(8,(%%ecx))
1844 HDF(16,(%%ecx, %1))
1845 HDF(24,(%%ecx, %1, 2))
1846 HDF(32,(%0, %1, 4))
1847 HDF(40,(%%ebx))
1848 HDF(48,(%%ebx, %1))
1849 HDF(56,(%%ebx, %1, 2))
1850 :
1851 : "r" (dst), "r" (stride), "r" (QP)
1852 : "%eax", "%ebx", "%ecx"
1853 );
1854 #else
1855 int y;
1856 for(y=0; y<BLOCK_SIZE; y++)
1857 {
1858 const int middleEnergy= 5*(dst[4] - dst[5]) + 2*(dst[2] - dst[5]);
1859
1860 if(ABS(middleEnergy) < 8*QP)
1861 {
1862 const int q=(dst[3] - dst[4])/2;
1863 const int leftEnergy= 5*(dst[2] - dst[1]) + 2*(dst[0] - dst[3]);
1864 const int rightEnergy= 5*(dst[6] - dst[5]) + 2*(dst[4] - dst[7]);
1865
1866 int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) );
1867 d= MAX(d, 0);
1868
1869 d= (5*d + 32) >> 6;
1870 d*= SIGN(-middleEnergy);
1871
1872 if(q>0)
1873 {
1874 d= d<0 ? 0 : d;
1875 d= d>q ? q : d;
1876 }
1877 else
1878 {
1879 d= d>0 ? 0 : d;
1880 d= d<q ? q : d;
1881 }
1882
1883 dst[3]-= d;
1884 dst[4]+= d;
1885 }
1886 dst+= stride;
1887 }
1888 #endif
1889 }
1890
1891 /**
1892 * Do a horizontal low pass filter on the 10x8 block (dst points to middle 8x8 Block)
1893 * using the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16 (C version)
1894 * using the 7-Tap Filter (2,2,2,4,2,2,2)/16 (MMX2/3DNOW version)
1895 */
1896 static inline void doHorizLowPass(uint8_t dst[], int stride, int QP)
1897 {
1898
1899 #if 0
1900 asm volatile(
1901 "leal (%0, %1), %%ecx \n\t"
1902 "leal (%%ecx, %1, 4), %%ebx \n\t"
1903 // 0 1 2 3 4 5 6 7 8 9
1904 // %0 ecx ecx+%1 ecx+2%1 %0+4%1 ebx ebx+%1 ebx+2%1 %0+8%1 ebx+4%1
1905 "pxor %%mm7, %%mm7 \n\t"
1906 "leal tempBlock, %%eax \n\t"
1907 /*
1908 #define HLP1 "movq (%0), %%mm0 \n\t"\
1909 "movq %%mm0, %%mm1 \n\t"\
1910 "psllq $8, %%mm0 \n\t"\
1911 PAVGB(%%mm1, %%mm0)\
1912 "psrlw $8, %%mm0 \n\t"\
1913 "pxor %%mm1, %%mm1 \n\t"\
1914 "packuswb %%mm1, %%mm0 \n\t"\
1915 "movq %%mm0, %%mm1 \n\t"\
1916 "movq %%mm0, %%mm2 \n\t"\
1917 "psllq $32, %%mm0 \n\t"\
1918 "paddb %%mm0, %%mm1 \n\t"\
1919 "psllq $16, %%mm2 \n\t"\
1920 PAVGB(%%mm2, %%mm0)\
1921 "movq %%mm0, %%mm3 \n\t"\
1922 "pand bm11001100, %%mm0 \n\t"\
1923 "paddusb %%mm0, %%mm3 \n\t"\
1924 "psrlq $8, %%mm3 \n\t"\
1925 PAVGB(%%mm1, %%mm4)\
1926 PAVGB(%%mm3, %%mm2)\
1927 "psrlq $16, %%mm2 \n\t"\
1928 "punpcklbw %%mm2, %%mm2 \n\t"\
1929 "movq %%mm2, (%0) \n\t"\
1930
1931 #define HLP2 "movq (%0), %%mm0 \n\t"\
1932 "movq %%mm0, %%mm1 \n\t"\
1933 "psllq $8, %%mm0 \n\t"\
1934 PAVGB(%%mm1, %%mm0)\
1935 "psrlw $8, %%mm0 \n\t"\
1936 "pxor %%mm1, %%mm1 \n\t"\
1937 "packuswb %%mm1, %%mm0 \n\t"\
1938 "movq %%mm0, %%mm2 \n\t"\
1939 "psllq $32, %%mm0 \n\t"\
1940 "psllq $16, %%mm2 \n\t"\
1941 PAVGB(%%mm2, %%mm0)\
1942 "movq %%mm0, %%mm3 \n\t"\
1943 "pand bm11001100, %%mm0 \n\t"\
1944 "paddusb %%mm0, %%mm3 \n\t"\
1945 "psrlq $8, %%mm3 \n\t"\
1946 PAVGB(%%mm3, %%mm2)\
1947 "psrlq $16, %%mm2 \n\t"\
1948 "punpcklbw %%mm2, %%mm2 \n\t"\
1949 "movq %%mm2, (%0) \n\t"\
1950 */
1951 // approximately a 7-Tap Filter with Vector (1,2,3,4,3,2,1)/16
1952 /*
1953 Implemented Exact 7-Tap
1954 9421 A321
1955 36421 64321
1956 334321 =
1957 1234321 =
1958 1234321 =
1959 123433 =
1960 12463 12346
1961 1249 123A
1962
1963 */
1964
1965 #ifdef HAVE_MMX2
1966 #define HLP3(i) "movq " #i "(%%eax), %%mm0 \n\t"\
1967 "movq %%mm0, %%mm1 \n\t"\
1968 "movq %%mm0, %%mm2 \n\t"\
1969 "movq %%mm0, %%mm3 \n\t"\
1970 "movq %%mm0, %%mm4 \n\t"\
1971 "psllq $8, %%mm1 \n\t"\
1972 "psrlq $8, %%mm2 \n\t"\
1973 "pand bm00000001, %%mm3 \n\t"\
1974 "pand bm10000000, %%mm4 \n\t"\
1975 "por %%mm3, %%mm1 \n\t"\
1976 "por %%mm4, %%mm2 \n\t"\
1977 PAVGB(%%mm2, %%mm1)\
1978 PAVGB(%%mm1, %%mm0)\
1979 \
1980 "pshufw $0xF9, %%mm0, %%mm3 \n\t"\
1981 "pshufw $0x90, %%mm0, %%mm4 \n\t"\
1982 PAVGB(%%mm3, %%mm4)\
1983 PAVGB(%%mm4, %%mm0)\
1984 "movd %%mm0, (%0) \n\t"\
1985 "psrlq $32, %%mm0 \n\t"\
1986 "movd %%mm0, 4(%0) \n\t"
1987 #else
1988 #define HLP3(i) "movq " #i "(%%eax), %%mm0 \n\t"\
1989 "movq %%mm0, %%mm1 \n\t"\
1990 "movq %%mm0, %%mm2 \n\t"\
1991 "movd -4(%0), %%mm3 \n\t" /*0001000*/\
1992 "movd 8(%0), %%mm4 \n\t" /*0001000*/\
1993 "psllq $8, %%mm1 \n\t"\
1994 "psrlq $8, %%mm2 \n\t"\
1995 "psrlq $24, %%mm3 \n\t"\
1996 "psllq $56, %%mm4 \n\t"\
1997 "por %%mm3, %%mm1 \n\t"\
1998 "por %%mm4, %%mm2 \n\t"\
1999 PAVGB(%%mm2, %%mm1)\
2000 PAVGB(%%mm1, %%mm0)\
2001 \
2002 "movq %%mm0, %%mm3 \n\t"\
2003 "movq %%mm0, %%mm4 \n\t"\
2004 "movq %%mm0, %%mm5 \n\t"\
2005 "psrlq $16, %%mm3 \n\t"\
2006 "psllq $16, %%mm4 \n\t"\
2007 "pand bm11000000, %%mm5 \n\t"\
2008 "por %%mm5, %%mm3 \n\t"\
2009 "movq %%mm0, %%mm5 \n\t"\
2010 "pand bm00000011, %%mm5 \n\t"\
2011 "por %%mm5, %%mm4 \n\t"\
2012 PAVGB(%%mm3, %%mm4)\
2013 PAVGB(%%mm4, %%mm0)\
2014 "movd %%mm0, (%0) \n\t"\
2015 "psrlq $32, %%mm0 \n\t"\
2016 "movd %%mm0, 4(%0) \n\t"
2017 #endif
2018
2019 /* uses the 7-Tap Filter: 1112111 */
2020 #define NEW_HLP(src, dst)\
2021 "movq " #src "(%%eax), %%mm1 \n\t"\
2022 "movq " #src "(%%eax), %%mm2 \n\t"\
2023 "psllq $8, %%mm1 \n\t"\
2024 "psrlq $8, %%mm2 \n\t"\
2025 "movd -4" #dst ", %%mm3 \n\t" /*0001000*/\
2026 "movd 8" #dst ", %%mm4 \n\t" /*0001000*/\
2027 "psrlq $24, %%mm3 \n\t"\
2028 "psllq $56, %%mm4 \n\t"\
2029 "por %%mm3, %%mm1 \n\t"\
2030 "por %%mm4, %%mm2 \n\t"\
2031 "movq %%mm1, %%mm5 \n\t"\
2032 PAVGB(%%mm2, %%mm1)\
2033 "movq " #src "(%%eax), %%mm0 \n\t"\
2034 PAVGB(%%mm1, %%mm0)\
2035 "psllq $8, %%mm5 \n\t"\
2036 "psrlq $8, %%mm2 \n\t"\
2037 "por %%mm3, %%mm5 \n\t"\
2038 "por %%mm4, %%mm2 \n\t"\
2039 "movq %%mm5, %%mm1 \n\t"\
2040 PAVGB(%%mm2, %%mm5)\
2041 "psllq $8, %%mm1 \n\t"\
2042 "psrlq $8, %%mm2 \n\t"\
2043 "por %%mm3, %%mm1 \n\t"\
2044 "por %%mm4, %%mm2 \n\t"\
2045 PAVGB(%%mm2, %%mm1)\
2046 PAVGB(%%mm1, %%mm5)\
2047 PAVGB(%%mm5, %%mm0)\
2048 "movd %%mm0, " #dst " \n\t"\
2049 "psrlq $32, %%mm0 \n\t"\
2050 "movd %%mm0, 4" #dst " \n\t"
2051
2052 /* uses the 9-Tap Filter: 112242211 */
2053 #define NEW_HLP2(i)\
2054 "movq " #i "(%%eax), %%mm0 \n\t" /*0001000*/\
2055 "movq %%mm0, %%mm1 \n\t" /*0001000*/\
2056 "movq %%mm0, %%mm2 \n\t" /*0001000*/\
2057 "movd -4(%0), %%mm3 \n\t" /*0001000*/\
2058 "movd 8(%0), %%mm4 \n\t" /*0001000*/\
2059 "psllq $8, %%mm1 \n\t"\
2060 "psrlq $8, %%mm2 \n\t"\
2061 "psrlq $24, %%mm3 \n\t"\
2062 "psllq $56, %%mm4 \n\t"\
2063 "por %%mm3, %%mm1 \n\t" /*0010000*/\
2064 "por %%mm4, %%mm2 \n\t" /*0000100*/\
2065 "movq %%mm1, %%mm5 \n\t" /*0010000*/\
2066 PAVGB(%%mm2, %%mm1) /*0010100*/\
2067 PAVGB(%%mm1, %%mm0) /*0012100*/\
2068 "psllq $8, %%mm5 \n\t"\
2069 "psrlq $8, %%mm2 \n\t"\
2070 "por %%mm3, %%mm5 \n\t" /*0100000*/\
2071 "por %%mm4, %%mm2 \n\t" /*0000010*/\
2072 "movq %%mm5, %%mm1 \n\t" /*0100000*/\
2073 PAVGB(%%mm2, %%mm5) /*0100010*/\
2074 "psllq $8, %%mm1 \n\t"\
2075 "psrlq $8, %%mm2 \n\t"\
2076 "por %%mm3, %%mm1 \n\t" /*1000000*/\
2077 "por %%mm4, %%mm2 \n\t" /*0000001*/\
2078 "movq %%mm1, %%mm6 \n\t" /*1000000*/\
2079 PAVGB(%%mm2, %%mm1) /*1000001*/\
2080 "psllq $8, %%mm6 \n\t"\
2081 "psrlq $8, %%mm2 \n\t"\
2082 "por %%mm3, %%mm6 \n\t"/*100000000*/\
2083 "por %%mm4, %%mm2 \n\t"/*000000001*/\
2084 PAVGB(%%mm2, %%mm6) /*100000001*/\
2085 PAVGB(%%mm6, %%mm1) /*110000011*/\
2086 PAVGB(%%mm1, %%mm5) /*112000211*/\
2087 PAVGB(%%mm5, %%mm0) /*112242211*/\
2088 "movd %%mm0, (%0) \n\t"\
2089 "psrlq $32, %%mm0 \n\t"\
2090 "movd %%mm0, 4(%0) \n\t"
2091
2092 #define HLP(src, dst) NEW_HLP(src, dst)
2093
2094 HLP(0, (%0))
2095 HLP(8, (%%ecx))
2096 HLP(16, (%%ecx, %1))
2097 HLP(24, (%%ecx, %1, 2))
2098 HLP(32, (%0, %1, 4))
2099 HLP(40, (%%ebx))
2100 HLP(48, (%%ebx, %1))
2101 HLP(56, (%%ebx, %1, 2))
2102
2103 :
2104 : "r" (dst), "r" (stride)
2105 : "%eax", "%ebx", "%ecx"
2106 );
2107
2108 #else
2109 int y;
2110 for(y=0; y<BLOCK_SIZE; y++)
2111 {
2112 const int first= ABS(dst[-1] - dst[0]) < QP ? dst[-1] : dst[0];
2113 const int last= ABS(dst[8] - dst[7]) < QP ? dst[8] : dst[7];
2114
2115 int sums[9];
2116 sums[0] = first + dst[0];
2117 sums[1] = dst[0] + dst[1];
2118 sums[2] = dst[1] + dst[2];
2119 sums[3] = dst[2] + dst[3];
2120 sums[4] = dst[3] + dst[4];
2121 sums[5] = dst[4] + dst[5];
2122 sums[6] = dst[5] + dst[6];
2123 sums[7] = dst[6] + dst[7];
2124 sums[8] = dst[7] + last;
2125
2126 dst[0]= ((sums[0]<<2) + ((first + sums[2])<<1) + sums[4] + 8)>>4;
2127 dst[1]= ((dst[1]<<2) + ((first + sums[0] + sums[3])<<1) + sums[5] + 8)>>4;
2128 dst[2]= ((dst[2]<<2) + ((first + sums[1] + sums[4])<<1) + sums[6] + 8)>>4;
2129 dst[3]= ((dst[3]<<2) + ((sums[2] + sums[5])<<1) + sums[0] + sums[7] + 8)>>4;
2130 dst[4]= ((dst[4]<<2) + ((sums[3] + sums[6])<<1) + sums[1] + sums[8] + 8)>>4;
2131 dst[5]= ((dst[5]<<2) + ((last + sums[7] + sums[4])<<1) + sums[2] + 8)>>4;
2132 dst[6]= (((last + dst[6])<<2) + ((dst[7] + sums[5])<<1) + sums[3] + 8)>>4;
2133 dst[7]= ((sums[8]<<2) + ((last + sums[6])<<1) + sums[4] + 8)>>4;
2134
2135 dst+= stride;
2136 } 1437 }
2137 #endif 1438 #endif
2138 } 1439 }
2139 1440
2140 static inline void dering(uint8_t src[], int stride, int QP) 1441 static inline void dering(uint8_t src[], int stride, int QP)
3531 2832
3532 horizontal_size >>= 1; 2833 horizontal_size >>= 1;
3533 vertical_size >>= 1; 2834 vertical_size >>= 1;
3534 src_stride >>= 1; 2835 src_stride >>= 1;
3535 dst_stride >>= 1; 2836 dst_stride >>= 1;
3536 // mode&= ~(LINEAR_IPOL_DEINT_FILTER | LINEAR_BLEND_DEINT_FILTER |
3537 // MEDIAN_DEINT_FILTER | CUBIC_IPOL_DEINT_FILTER);
3538 2837
3539 if(1) 2838 if(1)
3540 { 2839 {
3541 postProcess(src[1], src_stride, dst[1], dst_stride, 2840 postProcess(src[1], src_stride, dst[1], dst_stride,
3542 horizontal_size, vertical_size, QP_store, QP_stride, 1, &ppMode); 2841 horizontal_size, vertical_size, QP_store, QP_stride, 1, &ppMode);
3636 * Copies a block from src to dst and fixes the blacklevel 2935 * Copies a block from src to dst and fixes the blacklevel
3637 * numLines must be a multiple of 4 2936 * numLines must be a multiple of 4
3638 * levelFix == 0 -> dont touch the brighness & contrast 2937 * levelFix == 0 -> dont touch the brighness & contrast
3639 */ 2938 */
3640 static inline void blockCopy(uint8_t dst[], int dstStride, uint8_t src[], int srcStride, 2939 static inline void blockCopy(uint8_t dst[], int dstStride, uint8_t src[], int srcStride,
3641 int numLines, int levelFix) 2940 int levelFix)
3642 { 2941 {
3643 #ifndef HAVE_MMX 2942 #ifndef HAVE_MMX
3644 int i; 2943 int i;
3645 #endif 2944 #endif
3646 if(levelFix) 2945 if(levelFix)
3693 :"r" (srcStride), 2992 :"r" (srcStride),
3694 "r" (dstStride) 2993 "r" (dstStride)
3695 : "%eax", "%ebx" 2994 : "%eax", "%ebx"
3696 ); 2995 );
3697 #else 2996 #else
3698 for(i=0; i<numLines; i++) 2997 for(i=0; i<8; i++)
3699 memcpy( &(dst[dstStride*i]), 2998 memcpy( &(dst[dstStride*i]),
3700 &(src[srcStride*i]), BLOCK_SIZE); 2999 &(src[srcStride*i]), BLOCK_SIZE);
3701 #endif 3000 #endif
3702 } 3001 }
3703 else 3002 else
3704 { 3003 {
3705 #ifdef HAVE_MMX 3004 #ifdef HAVE_MMX
3706 asm volatile( 3005 asm volatile(
3707 "movl %4, %%eax \n\t"
3708 "movl %%eax, temp0\n\t"
3709 "pushl %0 \n\t" 3006 "pushl %0 \n\t"
3710 "pushl %1 \n\t" 3007 "pushl %1 \n\t"
3711 "leal (%2,%2), %%eax \n\t" 3008 "leal (%2,%2), %%eax \n\t"
3712 "leal (%3,%3), %%ebx \n\t" 3009 "leal (%3,%3), %%ebx \n\t"
3713 "movq packedYOffset, %%mm2 \n\t"
3714 "movq packedYScale, %%mm3 \n\t"
3715 3010
3716 #define SIMPLE_CPY \ 3011 #define SIMPLE_CPY \
3717 "movq (%0), %%mm0 \n\t"\ 3012 "movq (%0), %%mm0 \n\t"\
3718 "movq (%0,%2), %%mm1 \n\t"\ 3013 "movq (%0,%2), %%mm1 \n\t"\
3719 "movq %%mm0, (%1) \n\t"\ 3014 "movq %%mm0, (%1) \n\t"\
3720 "movq %%mm1, (%1, %3) \n\t"\ 3015 "movq %%mm1, (%1, %3) \n\t"\
3721 3016
3722 "1: \n\t"
3723 SIMPLE_CPY 3017 SIMPLE_CPY
3724 "addl %%eax, %0 \n\t" 3018 "addl %%eax, %0 \n\t"
3725 "addl %%ebx, %1 \n\t" 3019 "addl %%ebx, %1 \n\t"
3726 SIMPLE_CPY 3020 SIMPLE_CPY
3727 "addl %%eax, %0 \n\t" 3021 "addl %%eax, %0 \n\t"
3728 "addl %%ebx, %1 \n\t" 3022 "addl %%ebx, %1 \n\t"
3729 "decl temp0 \n\t" 3023 SIMPLE_CPY
3730 "jnz 1b \n\t" 3024 "addl %%eax, %0 \n\t"
3025 "addl %%ebx, %1 \n\t"
3026 SIMPLE_CPY
3731 3027
3732 "popl %1 \n\t" 3028 "popl %1 \n\t"
3733 "popl %0 \n\t" 3029 "popl %0 \n\t"
3734 : : "r" (src), 3030 : : "r" (src),
3735 "r" (dst), 3031 "r" (dst),
3736 "r" (srcStride), 3032 "r" (srcStride),
3737 "r" (dstStride), 3033 "r" (dstStride)
3738 "m" (numLines>>2)
3739 : "%eax", "%ebx" 3034 : "%eax", "%ebx"
3740 ); 3035 );
3741 #else 3036 #else
3742 for(i=0; i<numLines; i++) 3037 for(i=0; i<8; i++)
3743 memcpy( &(dst[dstStride*i]), 3038 memcpy( &(dst[dstStride*i]),
3744 &(src[srcStride*i]), BLOCK_SIZE); 3039 &(src[srcStride*i]), BLOCK_SIZE);
3745 #endif 3040 #endif
3746 } 3041 }
3747 } 3042 }
3772 3067
3773 /* Temporal noise reducing buffers */ 3068 /* Temporal noise reducing buffers */
3774 static uint8_t *tempBlured[3]= {NULL,NULL,NULL}; 3069 static uint8_t *tempBlured[3]= {NULL,NULL,NULL};
3775 static uint32_t *tempBluredPast[3]= {NULL,NULL,NULL}; 3070 static uint32_t *tempBluredPast[3]= {NULL,NULL,NULL};
3776 3071
3072 int copyAhead;
3073
3777 #ifdef PP_FUNNY_STRIDE 3074 #ifdef PP_FUNNY_STRIDE
3778 uint8_t *dstBlockPtrBackup; 3075 uint8_t *dstBlockPtrBackup;
3779 uint8_t *srcBlockPtrBackup; 3076 uint8_t *srcBlockPtrBackup;
3780 #endif 3077 #endif
3781 3078
3790 #ifdef HAVE_MMX 3087 #ifdef HAVE_MMX
3791 maxTmpNoise[0]= ppMode->maxTmpNoise[0]; 3088 maxTmpNoise[0]= ppMode->maxTmpNoise[0];
3792 maxTmpNoise[1]= ppMode->maxTmpNoise[1]; 3089 maxTmpNoise[1]= ppMode->maxTmpNoise[1];
3793 maxTmpNoise[2]= ppMode->maxTmpNoise[2]; 3090 maxTmpNoise[2]= ppMode->maxTmpNoise[2];
3794 #endif 3091 #endif
3092
3093 if(mode & CUBIC_IPOL_DEINT_FILTER) copyAhead=16;
3094 else if(mode & LINEAR_BLEND_DEINT_FILTER) copyAhead=14;
3095 else if( (mode & V_DEBLOCK)
3096 || (mode & LINEAR_IPOL_DEINT_FILTER)
3097 || (mode & MEDIAN_DEINT_FILTER)) copyAhead=13;
3098 else if(mode & V_X1_FILTER) copyAhead=11;
3099 else if(mode & V_RK1_FILTER) copyAhead=10;
3100 else if(mode & DERING) copyAhead=9;
3101 else copyAhead=8;
3102
3103 copyAhead-= 8;
3795 3104
3796 if(tempDst==NULL) 3105 if(tempDst==NULL)
3797 { 3106 {
3798 tempDst= (uint8_t*)memalign(8, 1024*24); 3107 tempDst= (uint8_t*)memalign(8, 1024*24);
3799 tempSrc= (uint8_t*)memalign(8, 1024*24); 3108 tempSrc= (uint8_t*)memalign(8, 1024*24);
3896 for(x=0; x<width; x+=BLOCK_SIZE) 3205 for(x=0; x<width; x+=BLOCK_SIZE)
3897 { 3206 {
3898 3207
3899 #ifdef HAVE_MMX2 3208 #ifdef HAVE_MMX2
3900 /* 3209 /*
3901 prefetchnta(srcBlock + (((x>>3)&3) + 5)*srcStride + 32);
3902 prefetchnta(srcBlock + (((x>>3)&3) + 9)*srcStride + 32);
3903 prefetcht0(dstBlock + (((x>>3)&3) + 5)*dstStride + 32);
3904 prefetcht0(dstBlock + (((x>>3)&3) + 9)*dstStride + 32);
3905 */
3906 /*
3907 prefetchnta(srcBlock + (((x>>2)&6) + 5)*srcStride + 32); 3210 prefetchnta(srcBlock + (((x>>2)&6) + 5)*srcStride + 32);
3908 prefetchnta(srcBlock + (((x>>2)&6) + 6)*srcStride + 32); 3211 prefetchnta(srcBlock + (((x>>2)&6) + 6)*srcStride + 32);
3909 prefetcht0(dstBlock + (((x>>2)&6) + 5)*dstStride + 32); 3212 prefetcht0(dstBlock + (((x>>2)&6) + 5)*dstStride + 32);
3910 prefetcht0(dstBlock + (((x>>2)&6) + 6)*dstStride + 32); 3213 prefetcht0(dstBlock + (((x>>2)&6) + 6)*dstStride + 32);
3911 */ 3214 */
3912 3215
3913 asm( 3216 asm(
3914 "movl %4, %%eax \n\t" 3217 "movl %4, %%eax \n\t"
3915 "shrl $2, %%eax \n\t" 3218 "shrl $2, %%eax \n\t"
3916 "andl $6, %%eax \n\t" 3219 "andl $6, %%eax \n\t"
3917 "addl $8, %%eax \n\t" 3220 "addl %5, %%eax \n\t"
3918 "movl %%eax, %%ebx \n\t" 3221 "movl %%eax, %%ebx \n\t"
3919 "imul %1, %%eax \n\t" 3222 "imul %1, %%eax \n\t"
3920 "imul %3, %%ebx \n\t" 3223 "imul %3, %%ebx \n\t"
3921 "prefetchnta 32(%%eax, %0) \n\t" 3224 "prefetchnta 32(%%eax, %0) \n\t"
3922 "prefetcht0 32(%%ebx, %2) \n\t" 3225 "prefetcht0 32(%%ebx, %2) \n\t"
3923 "addl %1, %%eax \n\t" 3226 "addl %1, %%eax \n\t"
3924 "addl %3, %%ebx \n\t" 3227 "addl %3, %%ebx \n\t"
3925 "prefetchnta 32(%%eax, %0) \n\t" 3228 "prefetchnta 32(%%eax, %0) \n\t"
3926 "prefetcht0 32(%%ebx, %2) \n\t" 3229 "prefetcht0 32(%%ebx, %2) \n\t"
3927 :: "r" (srcBlock), "r" (srcStride), "r" (dstBlock), "r" (dstStride), 3230 :: "r" (srcBlock), "r" (srcStride), "r" (dstBlock), "r" (dstStride),
3928 "m" (x) 3231 "m" (x), "m" (copyAhead)
3929 : "%eax", "%ebx" 3232 : "%eax", "%ebx"
3930 ); 3233 );
3931 3234
3932 #elif defined(HAVE_3DNOW) 3235 #elif defined(HAVE_3DNOW)
3933 //FIXME check if this is faster on an 3dnow chip or if its faster without the prefetch or ... 3236 //FIXME check if this is faster on an 3dnow chip or if its faster without the prefetch or ...
3936 prefetchw(dstBlock + (((x>>3)&3) + 5)*dstStride + 32); 3239 prefetchw(dstBlock + (((x>>3)&3) + 5)*dstStride + 32);
3937 prefetchw(dstBlock + (((x>>3)&3) + 9)*dstStride + 32); 3240 prefetchw(dstBlock + (((x>>3)&3) + 9)*dstStride + 32);
3938 */ 3241 */
3939 #endif 3242 #endif
3940 3243
3941 blockCopy(dstBlock + dstStride*8, dstStride, 3244 blockCopy(dstBlock + dstStride*copyAhead, dstStride,
3942 srcBlock + srcStride*8, srcStride, 8, mode & LEVEL_FIX); 3245 srcBlock + srcStride*copyAhead, srcStride, mode & LEVEL_FIX);
3943 3246
3944 if(mode & LINEAR_IPOL_DEINT_FILTER) 3247 if(mode & LINEAR_IPOL_DEINT_FILTER)
3945 deInterlaceInterpolateLinear(dstBlock, dstStride); 3248 deInterlaceInterpolateLinear(dstBlock, dstStride);
3946 else if(mode & LINEAR_BLEND_DEINT_FILTER) 3249 else if(mode & LINEAR_BLEND_DEINT_FILTER)
3947 deInterlaceBlendLinear(dstBlock, dstStride); 3250 deInterlaceBlendLinear(dstBlock, dstStride);
3953 deInterlaceBlendCubic(dstBlock, dstStride); 3256 deInterlaceBlendCubic(dstBlock, dstStride);
3954 */ 3257 */
3955 dstBlock+=8; 3258 dstBlock+=8;
3956 srcBlock+=8; 3259 srcBlock+=8;
3957 } 3260 }
3958 memcpy(&(dst[y*dstStride]) + 8*dstStride, tempDst + 9*dstStride, 8*dstStride ); 3261 memcpy(&(dst[y*dstStride]) + 8*dstStride, tempDst + 9*dstStride, copyAhead*dstStride );
3959 } 3262 }
3960 3263
3961 for(y=0; y<height; y+=BLOCK_SIZE) 3264 for(y=0; y<height; y+=BLOCK_SIZE)
3962 { 3265 {
3963 //1% speedup if these are here instead of the inner loop 3266 //1% speedup if these are here instead of the inner loop
3974 /* can we mess with a 8x16 block from srcBlock/dstBlock downwards and 1 line upwards 3277 /* can we mess with a 8x16 block from srcBlock/dstBlock downwards and 1 line upwards
3975 if not than use a temporary buffer */ 3278 if not than use a temporary buffer */
3976 if(y+15 >= height) 3279 if(y+15 >= height)
3977 { 3280 {
3978 int i; 3281 int i;
3979 /* copy from line 8 to 15 of src, these will be copied with 3282 /* copy from line (copyAhead) to (copyAhead+7) of src, these will be copied with
3980 blockcopy to dst later */ 3283 blockcopy to dst later */
3981 memcpy(tempSrc + srcStride*8, srcBlock + srcStride*8, 3284 memcpy(tempSrc + srcStride*copyAhead, srcBlock + srcStride*copyAhead,
3982 srcStride*MAX(height-y-8, 0) ); 3285 srcStride*MAX(height-y-copyAhead, 0) );
3983 3286
3984 /* duplicate last line of src to fill the void upto line 15 */ 3287 /* duplicate last line of src to fill the void upto line (copyAhead+7) */
3985 for(i=MAX(height-y, 8); i<=15; i++) 3288 for(i=MAX(height-y, 8); i<copyAhead+8; i++)
3986 memcpy(tempSrc + srcStride*i, src + srcStride*(height-1), srcStride); 3289 memcpy(tempSrc + srcStride*i, src + srcStride*(height-1), srcStride);
3987 3290
3988 /* copy up to 9 lines of dst (line -1 to 7)*/ 3291 /* copy up to (copyAhead+1) lines of dst (line -1 to (copyAhead-1))*/
3989 memcpy(tempDst, dstBlock - dstStride, dstStride*MIN(height-y+1, 9) ); 3292 memcpy(tempDst, dstBlock - dstStride, dstStride*MIN(height-y+1, copyAhead+1) );
3990 3293
3991 /* duplicate last line of dst to fill the void upto line 8 */ 3294 /* duplicate last line of dst to fill the void upto line (copyAhead) */
3992 for(i=height-y+1; i<=8; i++) 3295 for(i=height-y+1; i<=copyAhead; i++)
3993 memcpy(tempDst + dstStride*i, dst + dstStride*(height-1), dstStride); 3296 memcpy(tempDst + dstStride*i, dst + dstStride*(height-1), dstStride);
3994 3297
3995 dstBlock= tempDst + dstStride; 3298 dstBlock= tempDst + dstStride;
3996 srcBlock= tempSrc; 3299 srcBlock= tempSrc;
3997 } 3300 }
4039 T0= rdtsc(); 3342 T0= rdtsc();
4040 #endif 3343 #endif
4041 3344
4042 #ifdef HAVE_MMX2 3345 #ifdef HAVE_MMX2
4043 /* 3346 /*
4044 prefetchnta(srcBlock + (((x>>3)&3) + 5)*srcStride + 32);
4045 prefetchnta(srcBlock + (((x>>3)&3) + 9)*srcStride + 32);
4046 prefetcht0(dstBlock + (((x>>3)&3) + 5)*dstStride + 32);
4047 prefetcht0(dstBlock + (((x>>3)&3) + 9)*dstStride + 32);
4048 */
4049 /*
4050 prefetchnta(srcBlock + (((x>>2)&6) + 5)*srcStride + 32); 3347 prefetchnta(srcBlock + (((x>>2)&6) + 5)*srcStride + 32);
4051 prefetchnta(srcBlock + (((x>>2)&6) + 6)*srcStride + 32); 3348 prefetchnta(srcBlock + (((x>>2)&6) + 6)*srcStride + 32);
4052 prefetcht0(dstBlock + (((x>>2)&6) + 5)*dstStride + 32); 3349 prefetcht0(dstBlock + (((x>>2)&6) + 5)*dstStride + 32);
4053 prefetcht0(dstBlock + (((x>>2)&6) + 6)*dstStride + 32); 3350 prefetcht0(dstBlock + (((x>>2)&6) + 6)*dstStride + 32);
4054 */ 3351 */
4055 3352
4056 asm( 3353 asm(
4057 "movl %4, %%eax \n\t" 3354 "movl %4, %%eax \n\t"
4058 "shrl $2, %%eax \n\t" 3355 "shrl $2, %%eax \n\t"
4059 "andl $6, %%eax \n\t" 3356 "andl $6, %%eax \n\t"
4060 "addl $8, %%eax \n\t" 3357 "addl %5, %%eax \n\t"
4061 "movl %%eax, %%ebx \n\t" 3358 "movl %%eax, %%ebx \n\t"
4062 "imul %1, %%eax \n\t" 3359 "imul %1, %%eax \n\t"
4063 "imul %3, %%ebx \n\t" 3360 "imul %3, %%ebx \n\t"
4064 "prefetchnta 32(%%eax, %0) \n\t" 3361 "prefetchnta 32(%%eax, %0) \n\t"
4065 "prefetcht0 32(%%ebx, %2) \n\t" 3362 "prefetcht0 32(%%ebx, %2) \n\t"
4066 "addl %1, %%eax \n\t" 3363 "addl %1, %%eax \n\t"
4067 "addl %3, %%ebx \n\t" 3364 "addl %3, %%ebx \n\t"
4068 "prefetchnta 32(%%eax, %0) \n\t" 3365 "prefetchnta 32(%%eax, %0) \n\t"
4069 "prefetcht0 32(%%ebx, %2) \n\t" 3366 "prefetcht0 32(%%ebx, %2) \n\t"
4070 :: "r" (srcBlock), "r" (srcStride), "r" (dstBlock), "r" (dstStride), 3367 :: "r" (srcBlock), "r" (srcStride), "r" (dstBlock), "r" (dstStride),
4071 "m" (x) 3368 "m" (x), "m" (copyAhead)
4072 : "%eax", "%ebx" 3369 : "%eax", "%ebx"
4073 ); 3370 );
4074 3371
4075 #elif defined(HAVE_3DNOW) 3372 #elif defined(HAVE_3DNOW)
4076 //FIXME check if this is faster on an 3dnow chip or if its faster without the prefetch or ... 3373 //FIXME check if this is faster on an 3dnow chip or if its faster without the prefetch or ...
4098 dstBlock= tempDstBlock; 3395 dstBlock= tempDstBlock;
4099 srcBlock= tempSrcBlock; 3396 srcBlock= tempSrcBlock;
4100 } 3397 }
4101 #endif 3398 #endif
4102 3399
4103 blockCopy(dstBlock + dstStride*8, dstStride, 3400 blockCopy(dstBlock + dstStride*copyAhead, dstStride,
4104 srcBlock + srcStride*8, srcStride, 8, mode & LEVEL_FIX); 3401 srcBlock + srcStride*copyAhead, srcStride, mode & LEVEL_FIX);
4105 3402
4106 if(mode & LINEAR_IPOL_DEINT_FILTER) 3403 if(mode & LINEAR_IPOL_DEINT_FILTER)
4107 deInterlaceInterpolateLinear(dstBlock, dstStride); 3404 deInterlaceInterpolateLinear(dstBlock, dstStride);
4108 else if(mode & LINEAR_BLEND_DEINT_FILTER) 3405 else if(mode & LINEAR_BLEND_DEINT_FILTER)
4109 deInterlaceBlendLinear(dstBlock, dstStride); 3406 deInterlaceBlendLinear(dstBlock, dstStride);
4158 vertRK1Filter(tempBlock1, 16, QP); 3455 vertRK1Filter(tempBlock1, 16, QP);
4159 else if(mode & H_X1_FILTER) 3456 else if(mode & H_X1_FILTER)
4160 vertX1Filter(tempBlock1, 16, QP); 3457 vertX1Filter(tempBlock1, 16, QP);
4161 else if(mode & H_DEBLOCK) 3458 else if(mode & H_DEBLOCK)
4162 { 3459 {
4163 if( isVertDC(tempBlock1, 16)) 3460 if( isVertDC(tempBlock1, 16) )
4164 { 3461 {
4165 if(isVertMinMaxOk(tempBlock1, 16, QP)) 3462 if(isVertMinMaxOk(tempBlock1, 16, QP))
4166 doVertLowPass(tempBlock1, 16, QP); 3463 doVertLowPass(tempBlock1, 16, QP);
4167 } 3464 }
4168 else 3465 else
4250 memcpy(dstBlock, tempDst + dstStride, dstStride*(height-y) ); 3547 memcpy(dstBlock, tempDst + dstStride, dstStride*(height-y) );
4251 } 3548 }
4252 /* 3549 /*
4253 for(x=0; x<width; x+=32) 3550 for(x=0; x<width; x+=32)
4254 { 3551 {
4255 int i; 3552 volatile int i;
4256 i+= + dstBlock[x + 7*dstStride] + dstBlock[x + 8*dstStride] 3553 i+= + dstBlock[x + 7*dstStride] + dstBlock[x + 8*dstStride]
4257 + dstBlock[x + 9*dstStride] + dstBlock[x +10*dstStride] 3554 + dstBlock[x + 9*dstStride] + dstBlock[x +10*dstStride]
4258 + dstBlock[x +11*dstStride] + dstBlock[x +12*dstStride] 3555 + dstBlock[x +11*dstStride] + dstBlock[x +12*dstStride];
4259 + dstBlock[x +13*dstStride] + dstBlock[x +14*dstStride] 3556 // + dstBlock[x +13*dstStride]
4260 + dstBlock[x +15*dstStride]; 3557 // + dstBlock[x +14*dstStride] + dstBlock[x +15*dstStride];
4261 } 3558 }*/
4262 */ } 3559 }
4263 #ifdef HAVE_3DNOW 3560 #ifdef HAVE_3DNOW
4264 asm volatile("femms"); 3561 asm volatile("femms");
4265 #elif defined (HAVE_MMX) 3562 #elif defined (HAVE_MMX)
4266 asm volatile("emms"); 3563 asm volatile("emms");
4267 #endif 3564 #endif