26 |
|
|
27 |
%macro cglobal 1 |
%macro cglobal 1 |
28 |
%ifdef PREFIX |
%ifdef PREFIX |
29 |
|
%ifdef MARK_FUNCS |
30 |
|
global _%1:function |
31 |
|
%define %1 _%1:function |
32 |
|
%else |
33 |
global _%1 |
global _%1 |
34 |
%define %1 _%1 |
%define %1 _%1 |
35 |
|
%endif |
36 |
|
%else |
37 |
|
%ifdef MARK_FUNCS |
38 |
|
global %1:function |
39 |
%else |
%else |
40 |
global %1 |
global %1 |
41 |
%endif |
%endif |
42 |
|
%endif |
43 |
%endmacro |
%endmacro |
44 |
|
|
45 |
;============================================================================= |
;============================================================================= |
46 |
; Read only data |
; Read only data |
47 |
;============================================================================= |
;============================================================================= |
48 |
|
|
49 |
|
%ifdef FORMAT_COFF |
50 |
SECTION .rodata |
SECTION .rodata |
51 |
|
%else |
52 |
|
SECTION .rodata align=16 |
53 |
|
%endif |
54 |
|
|
55 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
56 |
; (16 - r) rounding table |
; (16 - r) rounding table |
104 |
cglobal interpolate8x8_halfpel_h_mmx |
cglobal interpolate8x8_halfpel_h_mmx |
105 |
cglobal interpolate8x8_halfpel_v_mmx |
cglobal interpolate8x8_halfpel_v_mmx |
106 |
cglobal interpolate8x8_halfpel_hv_mmx |
cglobal interpolate8x8_halfpel_hv_mmx |
107 |
|
|
108 |
cglobal interpolate8x8_avg4_mmx |
cglobal interpolate8x8_avg4_mmx |
109 |
cglobal interpolate8x8_avg2_mmx |
cglobal interpolate8x8_avg2_mmx |
110 |
|
|
111 |
cglobal interpolate8x8_6tap_lowpass_h_mmx |
cglobal interpolate8x8_6tap_lowpass_h_mmx |
112 |
cglobal interpolate8x8_6tap_lowpass_v_mmx |
cglobal interpolate8x8_6tap_lowpass_v_mmx |
113 |
|
|
114 |
|
cglobal interpolate8x8_halfpel_add_mmx |
115 |
|
cglobal interpolate8x8_halfpel_h_add_mmx |
116 |
|
cglobal interpolate8x8_halfpel_v_add_mmx |
117 |
|
cglobal interpolate8x8_halfpel_hv_add_mmx |
118 |
|
|
119 |
%macro CALC_AVG 6 |
%macro CALC_AVG 6 |
120 |
punpcklbw %3, %6 |
punpcklbw %3, %6 |
121 |
punpckhbw %4, %6 |
punpckhbw %4, %6 |
924 |
|
|
925 |
pop ebx |
pop ebx |
926 |
ret |
ret |
927 |
|
|
928 |
|
;=========================================================================== |
929 |
|
; |
930 |
|
; The next functions combine both source halfpel interpolation step and the |
931 |
|
; averaging (with rouding) step to avoid wasting memory bandwidth computing |
932 |
|
; intermediate halfpel images and then averaging them. |
933 |
|
; |
934 |
|
;=========================================================================== |
935 |
|
|
936 |
|
%macro PROLOG0 0 |
937 |
|
mov ecx, [esp+ 4] ; Dst |
938 |
|
mov eax, [esp+ 8] ; Src |
939 |
|
mov edx, [esp+12] ; BpS |
940 |
|
%endmacro |
941 |
|
|
942 |
|
%macro PROLOG 2 ; %1: Rounder, %2 load Dst-Rounder |
943 |
|
pxor mm6, mm6 |
944 |
|
movq mm7, [%1] ; TODO: dangerous! (eax isn't checked) |
945 |
|
%if %2 |
946 |
|
movq mm5, [rounding1_mmx] |
947 |
|
%endif |
948 |
|
|
949 |
|
PROLOG0 |
950 |
|
%endmacro |
951 |
|
|
952 |
|
; performs: mm0 == (mm0+mm2) mm1 == (mm1+mm3) |
953 |
|
%macro MIX 0 |
954 |
|
punpcklbw mm0, mm6 |
955 |
|
punpcklbw mm2, mm6 |
956 |
|
punpckhbw mm1, mm6 |
957 |
|
punpckhbw mm3, mm6 |
958 |
|
paddusw mm0, mm2 |
959 |
|
paddusw mm1, mm3 |
960 |
|
%endmacro |
961 |
|
|
962 |
|
%macro MIX_DST 0 |
963 |
|
movq mm3, mm2 |
964 |
|
paddusw mm0, mm7 ; rounder |
965 |
|
paddusw mm1, mm7 ; rounder |
966 |
|
punpcklbw mm2, mm6 |
967 |
|
punpckhbw mm3, mm6 |
968 |
|
psrlw mm0, 1 |
969 |
|
psrlw mm1, 1 |
970 |
|
|
971 |
|
paddusw mm0, mm2 ; mix Src(mm0/mm1) with Dst(mm2/mm3) |
972 |
|
paddusw mm1, mm3 |
973 |
|
paddusw mm0, mm5 |
974 |
|
paddusw mm1, mm5 |
975 |
|
psrlw mm0, 1 |
976 |
|
psrlw mm1, 1 |
977 |
|
|
978 |
|
packuswb mm0, mm1 |
979 |
|
%endmacro |
980 |
|
|
981 |
|
%macro MIX2 0 |
982 |
|
punpcklbw mm0, mm6 |
983 |
|
punpcklbw mm2, mm6 |
984 |
|
paddusw mm0, mm2 |
985 |
|
paddusw mm0, mm7 |
986 |
|
punpckhbw mm1, mm6 |
987 |
|
punpckhbw mm3, mm6 |
988 |
|
paddusw mm1, mm7 |
989 |
|
paddusw mm1, mm3 |
990 |
|
psrlw mm0, 1 |
991 |
|
psrlw mm1, 1 |
992 |
|
|
993 |
|
packuswb mm0, mm1 |
994 |
|
%endmacro |
995 |
|
|
996 |
|
;=========================================================================== |
997 |
|
; |
998 |
|
; void interpolate8x8_halfpel_add_mmx(uint8_t * const dst, |
999 |
|
; const uint8_t * const src, |
1000 |
|
; const uint32_t stride, |
1001 |
|
; const uint32_t rounding); |
1002 |
|
; |
1003 |
|
; |
1004 |
|
;=========================================================================== |
1005 |
|
|
1006 |
|
%macro ADD_FF_MMX 1 |
1007 |
|
movq mm0, [eax] |
1008 |
|
movq mm2, [ecx] |
1009 |
|
movq mm1, mm0 |
1010 |
|
movq mm3, mm2 |
1011 |
|
%if (%1!=0) |
1012 |
|
lea eax,[eax+%1*edx] |
1013 |
|
%endif |
1014 |
|
MIX |
1015 |
|
paddusw mm0, mm5 ; rounder |
1016 |
|
paddusw mm1, mm5 ; rounder |
1017 |
|
psrlw mm0, 1 |
1018 |
|
psrlw mm1, 1 |
1019 |
|
|
1020 |
|
packuswb mm0, mm1 |
1021 |
|
movq [ecx], mm0 |
1022 |
|
%if (%1!=0) |
1023 |
|
lea ecx,[ecx+%1*edx] |
1024 |
|
%endif |
1025 |
|
%endmacro |
1026 |
|
|
1027 |
|
ALIGN 16 |
1028 |
|
interpolate8x8_halfpel_add_mmx: |
1029 |
|
PROLOG rounding1_mmx, 1 |
1030 |
|
ADD_FF_MMX 1 |
1031 |
|
ADD_FF_MMX 1 |
1032 |
|
ADD_FF_MMX 1 |
1033 |
|
ADD_FF_MMX 1 |
1034 |
|
ADD_FF_MMX 1 |
1035 |
|
ADD_FF_MMX 1 |
1036 |
|
ADD_FF_MMX 1 |
1037 |
|
ADD_FF_MMX 0 |
1038 |
|
ret |
1039 |
|
|
1040 |
|
;=========================================================================== |
1041 |
|
; |
1042 |
|
; void interpolate8x8_halfpel_h_add_mmx(uint8_t * const dst, |
1043 |
|
; const uint8_t * const src, |
1044 |
|
; const uint32_t stride, |
1045 |
|
; const uint32_t rounding); |
1046 |
|
; |
1047 |
|
; |
1048 |
|
;=========================================================================== |
1049 |
|
|
1050 |
|
%macro ADD_FH_MMX 0 |
1051 |
|
movq mm0, [eax] |
1052 |
|
movq mm2, [eax+1] |
1053 |
|
movq mm1, mm0 |
1054 |
|
movq mm3, mm2 |
1055 |
|
|
1056 |
|
lea eax,[eax+edx] |
1057 |
|
|
1058 |
|
MIX |
1059 |
|
movq mm2, [ecx] ; prepare mix with Dst[0] |
1060 |
|
MIX_DST |
1061 |
|
movq [ecx], mm0 |
1062 |
|
%endmacro |
1063 |
|
|
1064 |
|
ALIGN 16 |
1065 |
|
interpolate8x8_halfpel_h_add_mmx: |
1066 |
|
PROLOG rounding1_mmx, 1 |
1067 |
|
|
1068 |
|
ADD_FH_MMX |
1069 |
|
lea ecx,[ecx+edx] |
1070 |
|
ADD_FH_MMX |
1071 |
|
lea ecx,[ecx+edx] |
1072 |
|
ADD_FH_MMX |
1073 |
|
lea ecx,[ecx+edx] |
1074 |
|
ADD_FH_MMX |
1075 |
|
lea ecx,[ecx+edx] |
1076 |
|
ADD_FH_MMX |
1077 |
|
lea ecx,[ecx+edx] |
1078 |
|
ADD_FH_MMX |
1079 |
|
lea ecx,[ecx+edx] |
1080 |
|
ADD_FH_MMX |
1081 |
|
lea ecx,[ecx+edx] |
1082 |
|
ADD_FH_MMX |
1083 |
|
ret |
1084 |
|
|
1085 |
|
;=========================================================================== |
1086 |
|
; |
1087 |
|
; void interpolate8x8_halfpel_v_add_mmx(uint8_t * const dst, |
1088 |
|
; const uint8_t * const src, |
1089 |
|
; const uint32_t stride, |
1090 |
|
; const uint32_t rounding); |
1091 |
|
; |
1092 |
|
; |
1093 |
|
;=========================================================================== |
1094 |
|
|
1095 |
|
%macro ADD_HF_MMX 0 |
1096 |
|
movq mm0, [eax] |
1097 |
|
movq mm2, [eax+edx] |
1098 |
|
movq mm1, mm0 |
1099 |
|
movq mm3, mm2 |
1100 |
|
|
1101 |
|
lea eax,[eax+edx] |
1102 |
|
|
1103 |
|
MIX |
1104 |
|
movq mm2, [ecx] ; prepare mix with Dst[0] |
1105 |
|
MIX_DST |
1106 |
|
movq [ecx], mm0 |
1107 |
|
|
1108 |
|
%endmacro |
1109 |
|
|
1110 |
|
ALIGN 16 |
1111 |
|
interpolate8x8_halfpel_v_add_mmx: |
1112 |
|
PROLOG rounding1_mmx, 1 |
1113 |
|
|
1114 |
|
ADD_HF_MMX |
1115 |
|
lea ecx,[ecx+edx] |
1116 |
|
ADD_HF_MMX |
1117 |
|
lea ecx,[ecx+edx] |
1118 |
|
ADD_HF_MMX |
1119 |
|
lea ecx,[ecx+edx] |
1120 |
|
ADD_HF_MMX |
1121 |
|
lea ecx,[ecx+edx] |
1122 |
|
ADD_HF_MMX |
1123 |
|
lea ecx,[ecx+edx] |
1124 |
|
ADD_HF_MMX |
1125 |
|
lea ecx,[ecx+edx] |
1126 |
|
ADD_HF_MMX |
1127 |
|
lea ecx,[ecx+edx] |
1128 |
|
ADD_HF_MMX |
1129 |
|
ret |
1130 |
|
|
1131 |
|
; The trick is to correct the result of 'pavgb' with some combination of the |
1132 |
|
; lsb's of the 4 input values i,j,k,l, and their intermediate 'pavgb' (s and t). |
1133 |
|
; The boolean relations are: |
1134 |
|
; (i+j+k+l+3)/4 = (s+t+1)/2 - (ij&kl)&st |
1135 |
|
; (i+j+k+l+2)/4 = (s+t+1)/2 - (ij|kl)&st |
1136 |
|
; (i+j+k+l+1)/4 = (s+t+1)/2 - (ij&kl)|st |
1137 |
|
; (i+j+k+l+0)/4 = (s+t+1)/2 - (ij|kl)|st |
1138 |
|
; with s=(i+j+1)/2, t=(k+l+1)/2, ij = i^j, kl = k^l, st = s^t. |
1139 |
|
|
1140 |
|
; Moreover, we process 2 lines at a times, for better overlapping (~15% faster). |
1141 |
|
|
1142 |
|
;=========================================================================== |
1143 |
|
; |
1144 |
|
; void interpolate8x8_halfpel_hv_add_mmx(uint8_t * const dst, |
1145 |
|
; const uint8_t * const src, |
1146 |
|
; const uint32_t stride, |
1147 |
|
; const uint32_t rounding); |
1148 |
|
; |
1149 |
|
; |
1150 |
|
;=========================================================================== |
1151 |
|
|
1152 |
|
%macro ADD_HH_MMX 0 |
1153 |
|
lea eax,[eax+edx] |
1154 |
|
|
1155 |
|
; transfert prev line to mm0/mm1 |
1156 |
|
movq mm0, mm2 |
1157 |
|
movq mm1, mm3 |
1158 |
|
|
1159 |
|
; load new line in mm2/mm3 |
1160 |
|
movq mm2, [eax] |
1161 |
|
movq mm4, [eax+1] |
1162 |
|
movq mm3, mm2 |
1163 |
|
movq mm5, mm4 |
1164 |
|
|
1165 |
|
punpcklbw mm2, mm6 |
1166 |
|
punpcklbw mm4, mm6 |
1167 |
|
paddusw mm2, mm4 |
1168 |
|
punpckhbw mm3, mm6 |
1169 |
|
punpckhbw mm5, mm6 |
1170 |
|
paddusw mm3, mm5 |
1171 |
|
|
1172 |
|
; mix current line (mm2/mm3) with previous (mm0,mm1); |
1173 |
|
; we'll preserve mm2/mm3 for next line... |
1174 |
|
|
1175 |
|
paddusw mm0, mm2 |
1176 |
|
paddusw mm1, mm3 |
1177 |
|
|
1178 |
|
movq mm4, [ecx] ; prepare mix with Dst[0] |
1179 |
|
movq mm5, mm4 |
1180 |
|
|
1181 |
|
paddusw mm0, mm7 ; finish mixing current line |
1182 |
|
paddusw mm1, mm7 |
1183 |
|
|
1184 |
|
punpcklbw mm4, mm6 |
1185 |
|
punpckhbw mm5, mm6 |
1186 |
|
|
1187 |
|
psrlw mm0, 2 |
1188 |
|
psrlw mm1, 2 |
1189 |
|
|
1190 |
|
paddusw mm0, mm4 ; mix Src(mm0/mm1) with Dst(mm2/mm3) |
1191 |
|
paddusw mm1, mm5 |
1192 |
|
|
1193 |
|
paddusw mm0, [rounding1_mmx] |
1194 |
|
paddusw mm1, [rounding1_mmx] |
1195 |
|
|
1196 |
|
psrlw mm0, 1 |
1197 |
|
psrlw mm1, 1 |
1198 |
|
|
1199 |
|
packuswb mm0, mm1 |
1200 |
|
|
1201 |
|
movq [ecx], mm0 |
1202 |
|
%endmacro |
1203 |
|
|
1204 |
|
ALIGN 16 |
1205 |
|
interpolate8x8_halfpel_hv_add_mmx: |
1206 |
|
PROLOG rounding2_mmx, 0 ; mm5 is busy. Don't load dst-rounder |
1207 |
|
|
1208 |
|
; preprocess first line |
1209 |
|
movq mm0, [eax] |
1210 |
|
movq mm2, [eax+1] |
1211 |
|
movq mm1, mm0 |
1212 |
|
movq mm3, mm2 |
1213 |
|
|
1214 |
|
punpcklbw mm0, mm6 |
1215 |
|
punpcklbw mm2, mm6 |
1216 |
|
punpckhbw mm1, mm6 |
1217 |
|
punpckhbw mm3, mm6 |
1218 |
|
paddusw mm2, mm0 |
1219 |
|
paddusw mm3, mm1 |
1220 |
|
|
1221 |
|
; Input: mm2/mm3 contains the value (Src[0]+Src[1]) of previous line |
1222 |
|
|
1223 |
|
ADD_HH_MMX |
1224 |
|
lea ecx,[ecx+edx] |
1225 |
|
ADD_HH_MMX |
1226 |
|
lea ecx,[ecx+edx] |
1227 |
|
ADD_HH_MMX |
1228 |
|
lea ecx,[ecx+edx] |
1229 |
|
ADD_HH_MMX |
1230 |
|
lea ecx,[ecx+edx] |
1231 |
|
ADD_HH_MMX |
1232 |
|
lea ecx,[ecx+edx] |
1233 |
|
ADD_HH_MMX |
1234 |
|
lea ecx,[ecx+edx] |
1235 |
|
ADD_HH_MMX |
1236 |
|
lea ecx,[ecx+edx] |
1237 |
|
ADD_HH_MMX |
1238 |
|
|
1239 |
|
ret |
1240 |
|
|