95 |
cglobal interpolate8x8_halfpel_h_mmx |
cglobal interpolate8x8_halfpel_h_mmx |
96 |
cglobal interpolate8x8_halfpel_v_mmx |
cglobal interpolate8x8_halfpel_v_mmx |
97 |
cglobal interpolate8x8_halfpel_hv_mmx |
cglobal interpolate8x8_halfpel_hv_mmx |
98 |
|
|
99 |
cglobal interpolate8x8_avg4_mmx |
cglobal interpolate8x8_avg4_mmx |
100 |
cglobal interpolate8x8_avg2_mmx |
cglobal interpolate8x8_avg2_mmx |
101 |
|
|
102 |
cglobal interpolate8x8_6tap_lowpass_h_mmx |
cglobal interpolate8x8_6tap_lowpass_h_mmx |
103 |
cglobal interpolate8x8_6tap_lowpass_v_mmx |
cglobal interpolate8x8_6tap_lowpass_v_mmx |
104 |
|
|
105 |
|
cglobal interpolate8x8_halfpel_add_mmx |
106 |
|
cglobal interpolate8x8_halfpel_h_add_mmx |
107 |
|
cglobal interpolate8x8_halfpel_v_add_mmx |
108 |
|
cglobal interpolate8x8_halfpel_hv_add_mmx |
109 |
|
|
110 |
%macro CALC_AVG 6 |
%macro CALC_AVG 6 |
111 |
punpcklbw %3, %6 |
punpcklbw %3, %6 |
112 |
punpckhbw %4, %6 |
punpckhbw %4, %6 |
915 |
|
|
916 |
pop ebx |
pop ebx |
917 |
ret |
ret |
918 |
|
|
919 |
|
;=========================================================================== |
920 |
|
; |
921 |
|
; The next functions combine both source halfpel interpolation step and the |
922 |
|
; averaging (with rouding) step to avoid wasting memory bandwidth computing |
923 |
|
; intermediate halfpel images and then averaging them. |
924 |
|
; |
925 |
|
;=========================================================================== |
926 |
|
|
927 |
|
%macro PROLOG0 0 |
928 |
|
mov ecx, [esp+ 4] ; Dst |
929 |
|
mov eax, [esp+ 8] ; Src |
930 |
|
mov edx, [esp+12] ; BpS |
931 |
|
%endmacro |
932 |
|
|
933 |
|
%macro PROLOG 2 ; %1: Rounder, %2 load Dst-Rounder |
934 |
|
pxor mm6, mm6 |
935 |
|
movq mm7, [%1] ; TODO: dangerous! (eax isn't checked) |
936 |
|
%if %2 |
937 |
|
movq mm5, [rounding1_mmx] |
938 |
|
%endif |
939 |
|
|
940 |
|
PROLOG0 |
941 |
|
%endmacro |
942 |
|
|
943 |
|
; performs: mm0 == (mm0+mm2) mm1 == (mm1+mm3) |
944 |
|
%macro MIX 0 |
945 |
|
punpcklbw mm0, mm6 |
946 |
|
punpcklbw mm2, mm6 |
947 |
|
punpckhbw mm1, mm6 |
948 |
|
punpckhbw mm3, mm6 |
949 |
|
paddusw mm0, mm2 |
950 |
|
paddusw mm1, mm3 |
951 |
|
%endmacro |
952 |
|
|
953 |
|
%macro MIX_DST 0 |
954 |
|
movq mm3, mm2 |
955 |
|
paddusw mm0, mm7 ; rounder |
956 |
|
paddusw mm1, mm7 ; rounder |
957 |
|
punpcklbw mm2, mm6 |
958 |
|
punpckhbw mm3, mm6 |
959 |
|
psrlw mm0, 1 |
960 |
|
psrlw mm1, 1 |
961 |
|
|
962 |
|
paddusw mm0, mm2 ; mix Src(mm0/mm1) with Dst(mm2/mm3) |
963 |
|
paddusw mm1, mm3 |
964 |
|
paddusw mm0, mm5 |
965 |
|
paddusw mm1, mm5 |
966 |
|
psrlw mm0, 1 |
967 |
|
psrlw mm1, 1 |
968 |
|
|
969 |
|
packuswb mm0, mm1 |
970 |
|
%endmacro |
971 |
|
|
972 |
|
%macro MIX2 0 |
973 |
|
punpcklbw mm0, mm6 |
974 |
|
punpcklbw mm2, mm6 |
975 |
|
paddusw mm0, mm2 |
976 |
|
paddusw mm0, mm7 |
977 |
|
punpckhbw mm1, mm6 |
978 |
|
punpckhbw mm3, mm6 |
979 |
|
paddusw mm1, mm7 |
980 |
|
paddusw mm1, mm3 |
981 |
|
psrlw mm0, 1 |
982 |
|
psrlw mm1, 1 |
983 |
|
|
984 |
|
packuswb mm0, mm1 |
985 |
|
%endmacro |
986 |
|
|
987 |
|
;=========================================================================== |
988 |
|
; |
989 |
|
; void interpolate8x8_halfpel_add_mmx(uint8_t * const dst, |
990 |
|
; const uint8_t * const src, |
991 |
|
; const uint32_t stride, |
992 |
|
; const uint32_t rounding); |
993 |
|
; |
994 |
|
; |
995 |
|
;=========================================================================== |
996 |
|
|
997 |
|
%macro ADD_FF_MMX 1 |
998 |
|
movq mm0, [eax] |
999 |
|
movq mm2, [ecx] |
1000 |
|
movq mm1, mm0 |
1001 |
|
movq mm3, mm2 |
1002 |
|
%if (%1!=0) |
1003 |
|
lea eax,[eax+%1*edx] |
1004 |
|
%endif |
1005 |
|
MIX |
1006 |
|
paddusw mm0, mm5 ; rounder |
1007 |
|
paddusw mm1, mm5 ; rounder |
1008 |
|
psrlw mm0, 1 |
1009 |
|
psrlw mm1, 1 |
1010 |
|
|
1011 |
|
packuswb mm0, mm1 |
1012 |
|
movq [ecx], mm0 |
1013 |
|
%if (%1!=0) |
1014 |
|
lea ecx,[ecx+%1*edx] |
1015 |
|
%endif |
1016 |
|
%endmacro |
1017 |
|
|
1018 |
|
ALIGN 16 |
1019 |
|
interpolate8x8_halfpel_add_mmx: |
1020 |
|
PROLOG rounding1_mmx, 1 |
1021 |
|
ADD_FF_MMX 1 |
1022 |
|
ADD_FF_MMX 1 |
1023 |
|
ADD_FF_MMX 1 |
1024 |
|
ADD_FF_MMX 1 |
1025 |
|
ADD_FF_MMX 1 |
1026 |
|
ADD_FF_MMX 1 |
1027 |
|
ADD_FF_MMX 1 |
1028 |
|
ADD_FF_MMX 0 |
1029 |
|
ret |
1030 |
|
|
1031 |
|
;=========================================================================== |
1032 |
|
; |
1033 |
|
; void interpolate8x8_halfpel_h_add_mmx(uint8_t * const dst, |
1034 |
|
; const uint8_t * const src, |
1035 |
|
; const uint32_t stride, |
1036 |
|
; const uint32_t rounding); |
1037 |
|
; |
1038 |
|
; |
1039 |
|
;=========================================================================== |
1040 |
|
|
1041 |
|
%macro ADD_FH_MMX 0 |
1042 |
|
movq mm0, [eax] |
1043 |
|
movq mm2, [eax+1] |
1044 |
|
movq mm1, mm0 |
1045 |
|
movq mm3, mm2 |
1046 |
|
|
1047 |
|
lea eax,[eax+edx] |
1048 |
|
|
1049 |
|
MIX |
1050 |
|
movq mm2, [ecx] ; prepare mix with Dst[0] |
1051 |
|
MIX_DST |
1052 |
|
movq [ecx], mm0 |
1053 |
|
%endmacro |
1054 |
|
|
1055 |
|
ALIGN 16 |
1056 |
|
interpolate8x8_halfpel_h_add_mmx: |
1057 |
|
PROLOG rounding1_mmx, 1 |
1058 |
|
|
1059 |
|
ADD_FH_MMX |
1060 |
|
lea ecx,[ecx+edx] |
1061 |
|
ADD_FH_MMX |
1062 |
|
lea ecx,[ecx+edx] |
1063 |
|
ADD_FH_MMX |
1064 |
|
lea ecx,[ecx+edx] |
1065 |
|
ADD_FH_MMX |
1066 |
|
lea ecx,[ecx+edx] |
1067 |
|
ADD_FH_MMX |
1068 |
|
lea ecx,[ecx+edx] |
1069 |
|
ADD_FH_MMX |
1070 |
|
lea ecx,[ecx+edx] |
1071 |
|
ADD_FH_MMX |
1072 |
|
lea ecx,[ecx+edx] |
1073 |
|
ADD_FH_MMX |
1074 |
|
ret |
1075 |
|
|
1076 |
|
;=========================================================================== |
1077 |
|
; |
1078 |
|
; void interpolate8x8_halfpel_v_add_mmx(uint8_t * const dst, |
1079 |
|
; const uint8_t * const src, |
1080 |
|
; const uint32_t stride, |
1081 |
|
; const uint32_t rounding); |
1082 |
|
; |
1083 |
|
; |
1084 |
|
;=========================================================================== |
1085 |
|
|
1086 |
|
%macro ADD_HF_MMX 0 |
1087 |
|
movq mm0, [eax] |
1088 |
|
movq mm2, [eax+edx] |
1089 |
|
movq mm1, mm0 |
1090 |
|
movq mm3, mm2 |
1091 |
|
|
1092 |
|
lea eax,[eax+edx] |
1093 |
|
|
1094 |
|
MIX |
1095 |
|
movq mm2, [ecx] ; prepare mix with Dst[0] |
1096 |
|
MIX_DST |
1097 |
|
movq [ecx], mm0 |
1098 |
|
|
1099 |
|
%endmacro |
1100 |
|
|
1101 |
|
ALIGN 16 |
1102 |
|
interpolate8x8_halfpel_v_add_mmx: |
1103 |
|
PROLOG rounding1_mmx, 1 |
1104 |
|
|
1105 |
|
ADD_HF_MMX |
1106 |
|
lea ecx,[ecx+edx] |
1107 |
|
ADD_HF_MMX |
1108 |
|
lea ecx,[ecx+edx] |
1109 |
|
ADD_HF_MMX |
1110 |
|
lea ecx,[ecx+edx] |
1111 |
|
ADD_HF_MMX |
1112 |
|
lea ecx,[ecx+edx] |
1113 |
|
ADD_HF_MMX |
1114 |
|
lea ecx,[ecx+edx] |
1115 |
|
ADD_HF_MMX |
1116 |
|
lea ecx,[ecx+edx] |
1117 |
|
ADD_HF_MMX |
1118 |
|
lea ecx,[ecx+edx] |
1119 |
|
ADD_HF_MMX |
1120 |
|
ret |
1121 |
|
|
1122 |
|
; The trick is to correct the result of 'pavgb' with some combination of the |
1123 |
|
; lsb's of the 4 input values i,j,k,l, and their intermediate 'pavgb' (s and t). |
1124 |
|
; The boolean relations are: |
1125 |
|
; (i+j+k+l+3)/4 = (s+t+1)/2 - (ij&kl)&st |
1126 |
|
; (i+j+k+l+2)/4 = (s+t+1)/2 - (ij|kl)&st |
1127 |
|
; (i+j+k+l+1)/4 = (s+t+1)/2 - (ij&kl)|st |
1128 |
|
; (i+j+k+l+0)/4 = (s+t+1)/2 - (ij|kl)|st |
1129 |
|
; with s=(i+j+1)/2, t=(k+l+1)/2, ij = i^j, kl = k^l, st = s^t. |
1130 |
|
|
1131 |
|
; Moreover, we process 2 lines at a times, for better overlapping (~15% faster). |
1132 |
|
|
1133 |
|
;=========================================================================== |
1134 |
|
; |
1135 |
|
; void interpolate8x8_halfpel_hv_add_mmx(uint8_t * const dst, |
1136 |
|
; const uint8_t * const src, |
1137 |
|
; const uint32_t stride, |
1138 |
|
; const uint32_t rounding); |
1139 |
|
; |
1140 |
|
; |
1141 |
|
;=========================================================================== |
1142 |
|
|
1143 |
|
%macro ADD_HH_MMX 0 |
1144 |
|
lea eax,[eax+edx] |
1145 |
|
|
1146 |
|
; transfert prev line to mm0/mm1 |
1147 |
|
movq mm0, mm2 |
1148 |
|
movq mm1, mm3 |
1149 |
|
|
1150 |
|
; load new line in mm2/mm3 |
1151 |
|
movq mm2, [eax] |
1152 |
|
movq mm4, [eax+1] |
1153 |
|
movq mm3, mm2 |
1154 |
|
movq mm5, mm4 |
1155 |
|
|
1156 |
|
punpcklbw mm2, mm6 |
1157 |
|
punpcklbw mm4, mm6 |
1158 |
|
paddusw mm2, mm4 |
1159 |
|
punpckhbw mm3, mm6 |
1160 |
|
punpckhbw mm5, mm6 |
1161 |
|
paddusw mm3, mm5 |
1162 |
|
|
1163 |
|
; mix current line (mm2/mm3) with previous (mm0,mm1); |
1164 |
|
; we'll preserve mm2/mm3 for next line... |
1165 |
|
|
1166 |
|
paddusw mm0, mm2 |
1167 |
|
paddusw mm1, mm3 |
1168 |
|
|
1169 |
|
movq mm4, [ecx] ; prepare mix with Dst[0] |
1170 |
|
movq mm5, mm4 |
1171 |
|
|
1172 |
|
paddusw mm0, mm7 ; finish mixing current line |
1173 |
|
paddusw mm1, mm7 |
1174 |
|
|
1175 |
|
punpcklbw mm4, mm6 |
1176 |
|
punpckhbw mm5, mm6 |
1177 |
|
|
1178 |
|
psrlw mm0, 2 |
1179 |
|
psrlw mm1, 2 |
1180 |
|
|
1181 |
|
paddusw mm0, mm4 ; mix Src(mm0/mm1) with Dst(mm2/mm3) |
1182 |
|
paddusw mm1, mm5 |
1183 |
|
|
1184 |
|
paddusw mm0, [rounding1_mmx] |
1185 |
|
paddusw mm1, [rounding1_mmx] |
1186 |
|
|
1187 |
|
psrlw mm0, 1 |
1188 |
|
psrlw mm1, 1 |
1189 |
|
|
1190 |
|
packuswb mm0, mm1 |
1191 |
|
|
1192 |
|
movq [ecx], mm0 |
1193 |
|
%endmacro |
1194 |
|
|
1195 |
|
ALIGN 16 |
1196 |
|
interpolate8x8_halfpel_hv_add_mmx: |
1197 |
|
PROLOG rounding2_mmx, 0 ; mm5 is busy. Don't load dst-rounder |
1198 |
|
|
1199 |
|
; preprocess first line |
1200 |
|
movq mm0, [eax] |
1201 |
|
movq mm2, [eax+1] |
1202 |
|
movq mm1, mm0 |
1203 |
|
movq mm3, mm2 |
1204 |
|
|
1205 |
|
punpcklbw mm0, mm6 |
1206 |
|
punpcklbw mm2, mm6 |
1207 |
|
punpckhbw mm1, mm6 |
1208 |
|
punpckhbw mm3, mm6 |
1209 |
|
paddusw mm2, mm0 |
1210 |
|
paddusw mm3, mm1 |
1211 |
|
|
1212 |
|
; Input: mm2/mm3 contains the value (Src[0]+Src[1]) of previous line |
1213 |
|
|
1214 |
|
ADD_HH_MMX |
1215 |
|
lea ecx,[ecx+edx] |
1216 |
|
ADD_HH_MMX |
1217 |
|
lea ecx,[ecx+edx] |
1218 |
|
ADD_HH_MMX |
1219 |
|
lea ecx,[ecx+edx] |
1220 |
|
ADD_HH_MMX |
1221 |
|
lea ecx,[ecx+edx] |
1222 |
|
ADD_HH_MMX |
1223 |
|
lea ecx,[ecx+edx] |
1224 |
|
ADD_HH_MMX |
1225 |
|
lea ecx,[ecx+edx] |
1226 |
|
ADD_HH_MMX |
1227 |
|
lea ecx,[ecx+edx] |
1228 |
|
ADD_HH_MMX |
1229 |
|
|
1230 |
|
ret |
1231 |
|
|