Mercurial > libavcodec.hg

comparison i386/dsputil_mmx.c @ 402:92d143c2d5a8 libavcodec

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
removed unused code
author glantau
date Mon, 20 May 2002 16:25:09 +0000
parents bf164fce2c14
children 1c3f42442fba
comparison
equal deleted inserted replaced
401:e20655449d4a 402:92d143c2d5a8
905 pix += line_size; 905 pix += line_size;
906 p += line_size; 906 p += line_size;
907 } while(--h); 907 } while(--h);
908 } 908 }
909 909
910 static void sub_pixels_mmx( DCTELEM *block, const UINT8 *pixels, int line_size, int h)
911 {
912 DCTELEM *p;
913 const UINT8 *pix;
914 p = block;
915 pix = pixels;
916 MOVQ_ZERO(mm7);
917 do {
918 __asm __volatile(
919 "movq %0, %%mm0\n\t"
920 "movq %1, %%mm2\n\t"
921 "movq 8%0, %%mm1\n\t"
922 "movq %%mm2, %%mm3\n\t"
923 "punpcklbw %%mm7, %%mm2\n\t"
924 "punpckhbw %%mm7, %%mm3\n\t"
925 "psubsw %%mm2, %%mm0\n\t"
926 "psubsw %%mm3, %%mm1\n\t"
927 "movq %%mm0, %0\n\t"
928 "movq %%mm1, 8%0\n\t"
929 :"+m"(*p)
930 :"m"(*pix)
931 :"memory");
932 pix += line_size;
933 p += 8;
934 } while (--h);
935 }
936
937 static void sub_pixels_x2_mmx( DCTELEM *block, const UINT8 *pixels, int line_size, int h)
938 {
939 DCTELEM *p;
940 const UINT8 *pix;
941 p = block;
942 pix = pixels;
943 MOVQ_ZERO(mm7);
944 MOVQ_WONE(mm6);
945 JUMPALIGN();
946 do {
947 __asm __volatile(
948 "movq %0, %%mm0\n\t"
949 "movq %1, %%mm2\n\t"
950 "movq 8%0, %%mm1\n\t"
951 "movq 1%1, %%mm4\n\t"
952 "movq %%mm2, %%mm3\n\t"
953 "movq %%mm4, %%mm5\n\t"
954 "punpcklbw %%mm7, %%mm2\n\t"
955 "punpckhbw %%mm7, %%mm3\n\t"
956 "punpcklbw %%mm7, %%mm4\n\t"
957 "punpckhbw %%mm7, %%mm5\n\t"
958 "paddusw %%mm4, %%mm2\n\t"
959 "paddusw %%mm5, %%mm3\n\t"
960 "paddusw %%mm6, %%mm2\n\t"
961 "paddusw %%mm6, %%mm3\n\t"
962 "psrlw $1, %%mm2\n\t"
963 "psrlw $1, %%mm3\n\t"
964 "psubsw %%mm2, %%mm0\n\t"
965 "psubsw %%mm3, %%mm1\n\t"
966 "movq %%mm0, %0\n\t"
967 "movq %%mm1, 8%0\n\t"
968 :"+m"(*p)
969 :"m"(*pix)
970 :"memory");
971 pix += line_size;
972 p += 8;
973 } while (--h);
974 }
975
976 static void sub_pixels_y2_mmx( DCTELEM *block, const UINT8 *pixels, int line_size, int h)
977 {
978 DCTELEM *p;
979 const UINT8 *pix;
980 p = block;
981 pix = pixels;
982 MOVQ_ZERO(mm7);
983 MOVQ_WONE(mm6);
984 do {
985 __asm __volatile(
986 "movq %0, %%mm0\n\t"
987 "movq %1, %%mm2\n\t"
988 "movq 8%0, %%mm1\n\t"
989 "movq %2, %%mm4\n\t"
990 "movq %%mm2, %%mm3\n\t"
991 "movq %%mm4, %%mm5\n\t"
992 "punpcklbw %%mm7, %%mm2\n\t"
993 "punpckhbw %%mm7, %%mm3\n\t"
994 "punpcklbw %%mm7, %%mm4\n\t"
995 "punpckhbw %%mm7, %%mm5\n\t"
996 "paddusw %%mm4, %%mm2\n\t"
997 "paddusw %%mm5, %%mm3\n\t"
998 "paddusw %%mm6, %%mm2\n\t"
999 "paddusw %%mm6, %%mm3\n\t"
1000 "psrlw $1, %%mm2\n\t"
1001 "psrlw $1, %%mm3\n\t"
1002 "psubsw %%mm2, %%mm0\n\t"
1003 "psubsw %%mm3, %%mm1\n\t"
1004 "movq %%mm0, %0\n\t"
1005 "movq %%mm1, 8%0\n\t"
1006 :"+m"(*p)
1007 :"m"(*pix), "m"(*(pix+line_size))
1008 :"memory");
1009 pix += line_size;
1010 p += 8;
1011 } while (--h);
1012 }
1013
1014 static void sub_pixels_xy2_mmx( DCTELEM *block, const UINT8 *pixels, int line_size, int h)
1015 {
1016 DCTELEM *p;
1017 const UINT8 *pix;
1018 p = block;
1019 pix = pixels;
1020 MOVQ_ZERO(mm7);
1021 MOVQ_WTWO(mm6);
1022 JUMPALIGN();
1023 do {
1024 __asm __volatile(
1025 "movq %1, %%mm0\n\t"
1026 "movq %2, %%mm1\n\t"
1027 "movq 1%1, %%mm4\n\t"
1028 "movq 1%2, %%mm5\n\t"
1029 "movq %%mm0, %%mm2\n\t"
1030 "movq %%mm1, %%mm3\n\t"
1031 "punpcklbw %%mm7, %%mm0\n\t"
1032 "punpcklbw %%mm7, %%mm1\n\t"
1033 "punpckhbw %%mm7, %%mm2\n\t"
1034 "punpckhbw %%mm7, %%mm3\n\t"
1035 "paddusw %%mm1, %%mm0\n\t"
1036 "paddusw %%mm3, %%mm2\n\t"
1037 "movq %%mm4, %%mm1\n\t"
1038 "movq %%mm5, %%mm3\n\t"
1039 "punpcklbw %%mm7, %%mm4\n\t"
1040 "punpcklbw %%mm7, %%mm5\n\t"
1041 "punpckhbw %%mm7, %%mm1\n\t"
1042 "punpckhbw %%mm7, %%mm3\n\t"
1043 "paddusw %%mm5, %%mm4\n\t"
1044 "paddusw %%mm3, %%mm1\n\t"
1045 "paddusw %%mm6, %%mm4\n\t"
1046 "paddusw %%mm6, %%mm1\n\t"
1047 "paddusw %%mm4, %%mm0\n\t"
1048 "paddusw %%mm1, %%mm2\n\t"
1049 "movq %0, %%mm1\n\t"
1050 "movq 8%0, %%mm3\n\t"
1051 "psrlw $2, %%mm0\n\t"
1052 "psrlw $2, %%mm2\n\t"
1053 "psubsw %%mm0, %%mm1\n\t"
1054 "psubsw %%mm2, %%mm3\n\t"
1055 "movq %%mm1, %0\n\t"
1056 "movq %%mm3, 8%0\n\t"
1057 :"+m"(*p)
1058 :"m"(*pix),
1059 "m"(*(pix+line_size))
1060 :"memory");
1061 pix += line_size;
1062 p += 8 ;
1063 } while(--h);
1064 }
1065
1066 static void clear_blocks_mmx(DCTELEM *blocks) 910 static void clear_blocks_mmx(DCTELEM *blocks)
1067 { 911 {
1068 asm volatile( 912 asm volatile(
1069 "pxor %%mm7, %%mm7 \n\t" 913 "pxor %%mm7, %%mm7 \n\t"
1070 "movl $-128*6, %%eax \n\t" 914 "movl $-128*6, %%eax \n\t"
1137 avg_no_rnd_pixels_tab[0] = avg_no_rnd_pixels_mmx; 981 avg_no_rnd_pixels_tab[0] = avg_no_rnd_pixels_mmx;
1138 avg_no_rnd_pixels_tab[1] = avg_no_rnd_pixels_x2_mmx; 982 avg_no_rnd_pixels_tab[1] = avg_no_rnd_pixels_x2_mmx;
1139 avg_no_rnd_pixels_tab[2] = avg_no_rnd_pixels_y2_mmx; 983 avg_no_rnd_pixels_tab[2] = avg_no_rnd_pixels_y2_mmx;
1140 avg_no_rnd_pixels_tab[3] = avg_no_rnd_pixels_xy2_mmx; 984 avg_no_rnd_pixels_tab[3] = avg_no_rnd_pixels_xy2_mmx;
1141 985
1142 sub_pixels_tab[0] = sub_pixels_mmx;
1143 sub_pixels_tab[1] = sub_pixels_x2_mmx;
1144 sub_pixels_tab[2] = sub_pixels_y2_mmx;
1145 sub_pixels_tab[3] = sub_pixels_xy2_mmx;
1146
1147 if (mm_flags & MM_MMXEXT) { 986 if (mm_flags & MM_MMXEXT) {
1148 pix_abs16x16 = pix_abs16x16_mmx2; 987 pix_abs16x16 = pix_abs16x16_mmx2;
1149 pix_abs16x16_x2 = pix_abs16x16_x2_mmx2; 988 pix_abs16x16_x2 = pix_abs16x16_x2_mmx2;
1150 pix_abs16x16_y2 = pix_abs16x16_y2_mmx2; 989 pix_abs16x16_y2 = pix_abs16x16_y2_mmx2;
1151 pix_abs16x16_xy2= pix_abs16x16_xy2_mmx2; 990 pix_abs16x16_xy2= pix_abs16x16_xy2_mmx2;
1162 1001
1163 avg_pixels_tab[0] = avg_pixels_mmx2; 1002 avg_pixels_tab[0] = avg_pixels_mmx2;
1164 avg_pixels_tab[1] = avg_pixels_x2_mmx2; 1003 avg_pixels_tab[1] = avg_pixels_x2_mmx2;
1165 avg_pixels_tab[2] = avg_pixels_y2_mmx2; 1004 avg_pixels_tab[2] = avg_pixels_y2_mmx2;
1166 avg_pixels_tab[3] = avg_pixels_xy2_mmx2; 1005 avg_pixels_tab[3] = avg_pixels_xy2_mmx2;
1167
1168 sub_pixels_tab[1] = sub_pixels_x2_mmx2;
1169 sub_pixels_tab[2] = sub_pixels_y2_mmx2;
1170 } else if (mm_flags & MM_3DNOW) { 1006 } else if (mm_flags & MM_3DNOW) {
1171 put_pixels_tab[1] = put_pixels_x2_3dnow; 1007 put_pixels_tab[1] = put_pixels_x2_3dnow;
1172 put_pixels_tab[2] = put_pixels_y2_3dnow; 1008 put_pixels_tab[2] = put_pixels_y2_3dnow;
1173 put_no_rnd_pixels_tab[1] = put_no_rnd_pixels_x2_3dnow; 1009 put_no_rnd_pixels_tab[1] = put_no_rnd_pixels_x2_3dnow;
1174 put_no_rnd_pixels_tab[2] = put_no_rnd_pixels_y2_3dnow; 1010 put_no_rnd_pixels_tab[2] = put_no_rnd_pixels_y2_3dnow;
1175 1011
1176 avg_pixels_tab[0] = avg_pixels_3dnow; 1012 avg_pixels_tab[0] = avg_pixels_3dnow;
1177 avg_pixels_tab[1] = avg_pixels_x2_3dnow; 1013 avg_pixels_tab[1] = avg_pixels_x2_3dnow;
1178 avg_pixels_tab[2] = avg_pixels_y2_3dnow; 1014 avg_pixels_tab[2] = avg_pixels_y2_3dnow;
1179 avg_pixels_tab[3] = avg_pixels_xy2_3dnow; 1015 avg_pixels_tab[3] = avg_pixels_xy2_3dnow;
1180
1181 sub_pixels_tab[1] = sub_pixels_x2_3dnow;
1182 sub_pixels_tab[2] = sub_pixels_y2_3dnow;
1183 } 1016 }
1184 1017
1185 /* idct */ 1018 /* idct */
1186 if (mm_flags & MM_MMXEXT) { 1019 if (mm_flags & MM_MMXEXT) {
1187 ff_idct = ff_mmxext_idct; 1020 ff_idct = ff_mmxext_idct;
1223 avg_no_rnd_pixels_tab[0] = just_return; 1056 avg_no_rnd_pixels_tab[0] = just_return;
1224 avg_no_rnd_pixels_tab[1] = just_return; 1057 avg_no_rnd_pixels_tab[1] = just_return;
1225 avg_no_rnd_pixels_tab[2] = just_return; 1058 avg_no_rnd_pixels_tab[2] = just_return;
1226 avg_no_rnd_pixels_tab[3] = just_return; 1059 avg_no_rnd_pixels_tab[3] = just_return;
1227 1060
1228 sub_pixels_tab[0] = just_return;
1229 sub_pixels_tab[1] = just_return;
1230 sub_pixels_tab[2] = just_return;
1231 sub_pixels_tab[3] = just_return;
1232
1233 //av_fdct = just_return; 1061 //av_fdct = just_return;
1234 //ff_idct = just_return; 1062 //ff_idct = just_return;
1235 #endif 1063 #endif
1236 } 1064 }
1065
1066 /* remove any non bit exact operation (testing purpose). NOTE that
1067 this function should be kept as small as possible because it is
1068 always difficult to test automatically non bit exact cases. */
1069 void dsputil_set_bit_exact_mmx(void)
1070 {
1071 if (mm_flags & MM_MMX) {
1072 if (mm_flags & MM_MMXEXT) {
1073 put_no_rnd_pixels_tab[1] = put_no_rnd_pixels_x2_mmx;
1074 put_no_rnd_pixels_tab[2] = put_no_rnd_pixels_y2_mmx;
1075 avg_pixels_tab[3] = avg_pixels_xy2_mmx;
1076 } else if (mm_flags & MM_3DNOW) {
1077 put_no_rnd_pixels_tab[1] = put_no_rnd_pixels_x2_mmx;
1078 put_no_rnd_pixels_tab[2] = put_no_rnd_pixels_y2_mmx;
1079 avg_pixels_tab[3] = avg_pixels_xy2_mmx;
1080 }
1081 }
1082 }