Mercurial > mplayer.hg

comparison postproc/postprocess_template.c @ 7990:f4f9390651b2

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
using fewer registers ... to workaround something
author michael
date Thu, 31 Oct 2002 01:42:59 +0000
parents a57c1fc0c2fc
children 772d6d27fd66
comparison
equal deleted inserted replaced
7989:31d6c10b8887 7990:f4f9390651b2
890 } 890 }
891 } 891 }
892 */ 892 */
893 #elif defined (HAVE_MMX) 893 #elif defined (HAVE_MMX)
894 src+= stride*4; 894 src+= stride*4;
895
896 asm volatile( 895 asm volatile(
897 "pxor %%mm7, %%mm7 \n\t" 896 "pxor %%mm7, %%mm7 \n\t"
898 "leal (%0, %1), %%eax \n\t"
899 "leal (%%eax, %1, 4), %%edx \n\t"
900 "leal -40(%%esp), %%ecx \n\t" // make space for 4 8-byte vars 897 "leal -40(%%esp), %%ecx \n\t" // make space for 4 8-byte vars
901 "andl $0xFFFFFFF8, %%ecx \n\t" // align 898 "andl $0xFFFFFFF8, %%ecx \n\t" // align
902 // 0 1 2 3 4 5 6 7 899 // 0 1 2 3 4 5 6 7
903 // %0 %0+%1 %0+2%1 eax+2%1 %0+4%1 eax+4%1 edx+%1 edx+2%1 900 // %0 %0+%1 %0+2%1 eax+2%1 %0+4%1 eax+4%1 edx+%1 edx+2%1
904 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 901 // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1
906 "movq (%0), %%mm0 \n\t" 903 "movq (%0), %%mm0 \n\t"
907 "movq %%mm0, %%mm1 \n\t" 904 "movq %%mm0, %%mm1 \n\t"
908 "punpcklbw %%mm7, %%mm0 \n\t" // low part of line 0 905 "punpcklbw %%mm7, %%mm0 \n\t" // low part of line 0
909 "punpckhbw %%mm7, %%mm1 \n\t" // high part of line 0 906 "punpckhbw %%mm7, %%mm1 \n\t" // high part of line 0
910 907
911 "movq (%%eax), %%mm2 \n\t" 908 "movq (%0, %1), %%mm2 \n\t"
909 "leal (%0, %1, 2), %%eax \n\t"
912 "movq %%mm2, %%mm3 \n\t" 910 "movq %%mm2, %%mm3 \n\t"
913 "punpcklbw %%mm7, %%mm2 \n\t" // low part of line 1 911 "punpcklbw %%mm7, %%mm2 \n\t" // low part of line 1
914 "punpckhbw %%mm7, %%mm3 \n\t" // high part of line 1 912 "punpckhbw %%mm7, %%mm3 \n\t" // high part of line 1
915 913
916 "movq (%%eax, %1), %%mm4 \n\t" 914 "movq (%%eax), %%mm4 \n\t"
917 "movq %%mm4, %%mm5 \n\t" 915 "movq %%mm4, %%mm5 \n\t"
918 "punpcklbw %%mm7, %%mm4 \n\t" // low part of line 2 916 "punpcklbw %%mm7, %%mm4 \n\t" // low part of line 2
919 "punpckhbw %%mm7, %%mm5 \n\t" // high part of line 2 917 "punpckhbw %%mm7, %%mm5 \n\t" // high part of line 2
920 918
921 "paddw %%mm0, %%mm0 \n\t" // 2L0 919 "paddw %%mm0, %%mm0 \n\t" // 2L0
928 "psllw $2, %%mm2 \n\t" // 4L1 - 4L2 926 "psllw $2, %%mm2 \n\t" // 4L1 - 4L2
929 "psllw $2, %%mm3 \n\t" // 4H1 - 4H2 927 "psllw $2, %%mm3 \n\t" // 4H1 - 4H2
930 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 928 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2
931 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 929 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2
932 930
933 "movq (%%eax, %1, 2), %%mm2 \n\t" 931 "movq (%%eax, %1), %%mm2 \n\t"
934 "movq %%mm2, %%mm3 \n\t" 932 "movq %%mm2, %%mm3 \n\t"
935 "punpcklbw %%mm7, %%mm2 \n\t" // L3 933 "punpcklbw %%mm7, %%mm2 \n\t" // L3
936 "punpckhbw %%mm7, %%mm3 \n\t" // H3 934 "punpckhbw %%mm7, %%mm3 \n\t" // H3
937 935
938 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - L3 936 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - L3
940 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 938 "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3
941 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 939 "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3
942 "movq %%mm0, (%%ecx) \n\t" // 2L0 - 5L1 + 5L2 - 2L3 940 "movq %%mm0, (%%ecx) \n\t" // 2L0 - 5L1 + 5L2 - 2L3
943 "movq %%mm1, 8(%%ecx) \n\t" // 2H0 - 5H1 + 5H2 - 2H3 941 "movq %%mm1, 8(%%ecx) \n\t" // 2H0 - 5H1 + 5H2 - 2H3
944 942
945 "movq (%0, %1, 4), %%mm0 \n\t" 943 "movq (%%eax, %1, 2), %%mm0 \n\t"
946 "movq %%mm0, %%mm1 \n\t" 944 "movq %%mm0, %%mm1 \n\t"
947 "punpcklbw %%mm7, %%mm0 \n\t" // L4 945 "punpcklbw %%mm7, %%mm0 \n\t" // L4
948 "punpckhbw %%mm7, %%mm1 \n\t" // H4 946 "punpckhbw %%mm7, %%mm1 \n\t" // H4
949 947
950 "psubw %%mm0, %%mm2 \n\t" // L3 - L4 948 "psubw %%mm0, %%mm2 \n\t" // L3 - L4
954 "paddw %%mm4, %%mm4 \n\t" // 2L2 952 "paddw %%mm4, %%mm4 \n\t" // 2L2
955 "paddw %%mm5, %%mm5 \n\t" // 2H2 953 "paddw %%mm5, %%mm5 \n\t" // 2H2
956 "psubw %%mm2, %%mm4 \n\t" // 2L2 - L3 + L4 954 "psubw %%mm2, %%mm4 \n\t" // 2L2 - L3 + L4
957 "psubw %%mm3, %%mm5 \n\t" // 2H2 - H3 + H4 955 "psubw %%mm3, %%mm5 \n\t" // 2H2 - H3 + H4
958 956
957 "leal (%%eax, %1), %0 \n\t"
959 "psllw $2, %%mm2 \n\t" // 4L3 - 4L4 958 "psllw $2, %%mm2 \n\t" // 4L3 - 4L4
960 "psllw $2, %%mm3 \n\t" // 4H3 - 4H4 959 "psllw $2, %%mm3 \n\t" // 4H3 - 4H4
961 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 960 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4
962 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 961 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4
963 //50 opcodes so far 962 //50 opcodes so far
964 "movq (%%edx), %%mm2 \n\t" 963 "movq (%0, %1, 2), %%mm2 \n\t"
965 "movq %%mm2, %%mm3 \n\t" 964 "movq %%mm2, %%mm3 \n\t"
966 "punpcklbw %%mm7, %%mm2 \n\t" // L5 965 "punpcklbw %%mm7, %%mm2 \n\t" // L5
967 "punpckhbw %%mm7, %%mm3 \n\t" // H5 966 "punpckhbw %%mm7, %%mm3 \n\t" // H5
968 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - L5 967 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - L5
969 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - H5 968 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - H5
970 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - 2L5 969 "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - 2L5
971 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - 2H5 970 "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - 2H5
972 971
973 "movq (%%edx, %1), %%mm6 \n\t" 972 "movq (%%eax, %1, 4), %%mm6 \n\t"
974 "punpcklbw %%mm7, %%mm6 \n\t" // L6 973 "punpcklbw %%mm7, %%mm6 \n\t" // L6
975 "psubw %%mm6, %%mm2 \n\t" // L5 - L6 974 "psubw %%mm6, %%mm2 \n\t" // L5 - L6
976 "movq (%%edx, %1), %%mm6 \n\t" 975 "movq (%%eax, %1, 4), %%mm6 \n\t"
977 "punpckhbw %%mm7, %%mm6 \n\t" // H6 976 "punpckhbw %%mm7, %%mm6 \n\t" // H6
978 "psubw %%mm6, %%mm3 \n\t" // H5 - H6 977 "psubw %%mm6, %%mm3 \n\t" // H5 - H6
979 978
980 "paddw %%mm0, %%mm0 \n\t" // 2L4 979 "paddw %%mm0, %%mm0 \n\t" // 2L4
981 "paddw %%mm1, %%mm1 \n\t" // 2H4 980 "paddw %%mm1, %%mm1 \n\t" // 2H4
985 "psllw $2, %%mm2 \n\t" // 4L5 - 4L6 984 "psllw $2, %%mm2 \n\t" // 4L5 - 4L6
986 "psllw $2, %%mm3 \n\t" // 4H5 - 4H6 985 "psllw $2, %%mm3 \n\t" // 4H5 - 4H6
987 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 986 "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6
988 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 987 "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6
989 988
990 "movq (%%edx, %1, 2), %%mm2 \n\t" 989 "movq (%0, %1, 4), %%mm2 \n\t"
991 "movq %%mm2, %%mm3 \n\t" 990 "movq %%mm2, %%mm3 \n\t"
992 "punpcklbw %%mm7, %%mm2 \n\t" // L7 991 "punpcklbw %%mm7, %%mm2 \n\t" // L7
993 "punpckhbw %%mm7, %%mm3 \n\t" // H7 992 "punpckhbw %%mm7, %%mm3 \n\t" // H7
994 993
995 "paddw %%mm2, %%mm2 \n\t" // 2L7 994 "paddw %%mm2, %%mm2 \n\t" // 2L7
1108 "pxor %%mm6, %%mm4 \n\t" 1107 "pxor %%mm6, %%mm4 \n\t"
1109 "pxor %%mm7, %%mm5 \n\t" 1108 "pxor %%mm7, %%mm5 \n\t"
1110 "psubw %%mm6, %%mm4 \n\t" 1109 "psubw %%mm6, %%mm4 \n\t"
1111 "psubw %%mm7, %%mm5 \n\t" 1110 "psubw %%mm7, %%mm5 \n\t"
1112 "packsswb %%mm5, %%mm4 \n\t" 1111 "packsswb %%mm5, %%mm4 \n\t"
1113 "movq (%%eax, %1, 2), %%mm0 \n\t" 1112 "movq (%0), %%mm0 \n\t"
1114 "paddb %%mm4, %%mm0 \n\t" 1113 "paddb %%mm4, %%mm0 \n\t"
1115 "movq %%mm0, (%%eax, %1, 2) \n\t" 1114 "movq %%mm0, (%0) \n\t"
1116 "movq (%0, %1, 4), %%mm0 \n\t" 1115 "movq (%0, %1), %%mm0 \n\t"
1117 "psubb %%mm4, %%mm0 \n\t" 1116 "psubb %%mm4, %%mm0 \n\t"
1118 "movq %%mm0, (%0, %1, 4) \n\t" 1117 "movq %%mm0, (%0, %1) \n\t"
1119 1118
1120 : 1119 : "+r" (src)
1121 : "r" (src), "r" (stride), "m" (c->pQPb) 1120 : "r" (stride), "m" (c->pQPb)
1122 : "%eax", "%edx", "%ecx" 1121 : "%eax", "%ecx"
1123 ); 1122 );
1124 #else 1123 #else
1125 const int l1= stride; 1124 const int l1= stride;
1126 const int l2= stride + l1; 1125 const int l2= stride + l1;
1127 const int l3= stride + l2; 1126 const int l3= stride + l2;