26 |
|
|
27 |
%macro cglobal 1 |
%macro cglobal 1 |
28 |
%ifdef PREFIX |
%ifdef PREFIX |
29 |
|
%ifdef MARK_FUNCS |
30 |
|
global _%1:function %1.endfunc-%1 |
31 |
|
%define %1 _%1:function %1.endfunc-%1 |
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 %1.endfunc-%1 |
39 |
%else |
%else |
40 |
global %1 |
global %1 |
41 |
%endif |
%endif |
42 |
|
%endif |
43 |
%endmacro |
%endmacro |
44 |
|
|
45 |
;============================================================================= |
;============================================================================= |
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 interpolate8x4_halfpel_h_mmx |
109 |
|
cglobal interpolate8x4_halfpel_v_mmx |
110 |
|
cglobal interpolate8x4_halfpel_hv_mmx |
111 |
|
|
112 |
cglobal interpolate8x8_avg4_mmx |
cglobal interpolate8x8_avg4_mmx |
113 |
cglobal interpolate8x8_avg2_mmx |
cglobal interpolate8x8_avg2_mmx |
114 |
|
|
115 |
cglobal interpolate8x8_6tap_lowpass_h_mmx |
cglobal interpolate8x8_6tap_lowpass_h_mmx |
116 |
cglobal interpolate8x8_6tap_lowpass_v_mmx |
cglobal interpolate8x8_6tap_lowpass_v_mmx |
117 |
|
|
118 |
|
cglobal interpolate8x8_halfpel_add_mmx |
119 |
|
cglobal interpolate8x8_halfpel_h_add_mmx |
120 |
|
cglobal interpolate8x8_halfpel_v_add_mmx |
121 |
|
cglobal interpolate8x8_halfpel_hv_add_mmx |
122 |
|
|
123 |
%macro CALC_AVG 6 |
%macro CALC_AVG 6 |
124 |
punpcklbw %3, %6 |
punpcklbw %3, %6 |
125 |
punpckhbw %4, %6 |
punpckhbw %4, %6 |
189 |
pop esi |
pop esi |
190 |
|
|
191 |
ret |
ret |
192 |
|
.endfunc |
193 |
|
|
194 |
|
|
195 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
249 |
pop esi |
pop esi |
250 |
|
|
251 |
ret |
ret |
252 |
|
.endfunc |
253 |
|
|
254 |
|
|
255 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
341 |
pop esi |
pop esi |
342 |
|
|
343 |
ret |
ret |
344 |
|
.endfunc |
345 |
|
|
346 |
|
;----------------------------------------------------------------------------- |
347 |
|
; |
348 |
|
; void interpolate8x4_halfpel_h_mmx(uint8_t * const dst, |
349 |
|
; const uint8_t * const src, |
350 |
|
; const uint32_t stride, |
351 |
|
; const uint32_t rounding); |
352 |
|
; |
353 |
|
;----------------------------------------------------------------------------- |
354 |
|
|
355 |
|
ALIGN 16 |
356 |
|
interpolate8x4_halfpel_h_mmx: |
357 |
|
|
358 |
|
push esi |
359 |
|
push edi |
360 |
|
mov eax, [esp + 8 + 16] ; rounding |
361 |
|
|
362 |
|
movq mm7, [rounding1_mmx + eax * 8] |
363 |
|
|
364 |
|
mov edi, [esp + 8 + 4] ; dst |
365 |
|
mov esi, [esp + 8 + 8] ; src |
366 |
|
mov edx, [esp + 8 + 12] ; stride |
367 |
|
|
368 |
|
pxor mm6, mm6 ; zero |
369 |
|
|
370 |
|
COPY_H_MMX |
371 |
|
COPY_H_MMX |
372 |
|
COPY_H_MMX |
373 |
|
COPY_H_MMX |
374 |
|
|
375 |
|
pop edi |
376 |
|
pop esi |
377 |
|
|
378 |
|
ret |
379 |
|
.endfunc |
380 |
|
|
381 |
|
|
382 |
|
;----------------------------------------------------------------------------- |
383 |
|
; |
384 |
|
; void interpolate8x4_halfpel_v_mmx(uint8_t * const dst, |
385 |
|
; const uint8_t * const src, |
386 |
|
; const uint32_t stride, |
387 |
|
; const uint32_t rounding); |
388 |
|
; |
389 |
|
;----------------------------------------------------------------------------- |
390 |
|
|
391 |
|
ALIGN 16 |
392 |
|
interpolate8x4_halfpel_v_mmx: |
393 |
|
|
394 |
|
push esi |
395 |
|
push edi |
396 |
|
|
397 |
|
mov eax, [esp + 8 + 16] ; rounding |
398 |
|
|
399 |
|
movq mm7, [rounding1_mmx + eax * 8] |
400 |
|
|
401 |
|
mov edi, [esp + 8 + 4] ; dst |
402 |
|
mov esi, [esp + 8 + 8] ; src |
403 |
|
mov edx, [esp + 8 + 12] ; stride |
404 |
|
|
405 |
|
pxor mm6, mm6 ; zero |
406 |
|
|
407 |
|
|
408 |
|
COPY_V_MMX |
409 |
|
COPY_V_MMX |
410 |
|
COPY_V_MMX |
411 |
|
COPY_V_MMX |
412 |
|
|
413 |
|
pop edi |
414 |
|
pop esi |
415 |
|
|
416 |
|
ret |
417 |
|
.endfunc |
418 |
|
|
419 |
|
|
420 |
|
;----------------------------------------------------------------------------- |
421 |
|
; |
422 |
|
; void interpolate8x4_halfpel_hv_mmx(uint8_t * const dst, |
423 |
|
; const uint8_t * const src, |
424 |
|
; const uint32_t stride, |
425 |
|
; const uint32_t rounding); |
426 |
|
; |
427 |
|
; |
428 |
|
;----------------------------------------------------------------------------- |
429 |
|
|
430 |
|
ALIGN 16 |
431 |
|
interpolate8x4_halfpel_hv_mmx: |
432 |
|
|
433 |
|
push esi |
434 |
|
push edi |
435 |
|
|
436 |
|
mov eax, [esp + 8 + 16] ; rounding |
437 |
|
|
438 |
|
movq mm7, [rounding2_mmx + eax * 8] |
439 |
|
|
440 |
|
mov edi, [esp + 8 + 4] ; dst |
441 |
|
mov esi, [esp + 8 + 8] ; src |
442 |
|
|
443 |
|
mov eax, 8 |
444 |
|
|
445 |
|
pxor mm6, mm6 ; zero |
446 |
|
|
447 |
|
mov edx, [esp + 8 + 12] ; stride |
448 |
|
|
449 |
|
COPY_HV_MMX |
450 |
|
COPY_HV_MMX |
451 |
|
COPY_HV_MMX |
452 |
|
COPY_HV_MMX |
453 |
|
|
454 |
|
pop edi |
455 |
|
pop esi |
456 |
|
|
457 |
|
ret |
458 |
|
.endfunc |
459 |
|
|
460 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
461 |
; |
; |
630 |
|
|
631 |
pop ebx |
pop ebx |
632 |
ret |
ret |
633 |
|
.endfunc |
634 |
|
|
635 |
|
|
636 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
819 |
pop edi |
pop edi |
820 |
pop ebx |
pop ebx |
821 |
ret |
ret |
822 |
|
.endfunc |
823 |
|
|
824 |
|
|
825 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
930 |
LOWPASS_6TAP_H_MMX |
LOWPASS_6TAP_H_MMX |
931 |
|
|
932 |
ret |
ret |
933 |
|
.endfunc |
934 |
|
|
935 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
936 |
; |
; |
1048 |
|
|
1049 |
pop ebx |
pop ebx |
1050 |
ret |
ret |
1051 |
|
.endfunc |
1052 |
|
|
1053 |
|
;=========================================================================== |
1054 |
|
; |
1055 |
|
; The next functions combine both source halfpel interpolation step and the |
1056 |
|
; averaging (with rouding) step to avoid wasting memory bandwidth computing |
1057 |
|
; intermediate halfpel images and then averaging them. |
1058 |
|
; |
1059 |
|
;=========================================================================== |
1060 |
|
|
1061 |
|
%macro PROLOG0 0 |
1062 |
|
mov ecx, [esp+ 4] ; Dst |
1063 |
|
mov eax, [esp+ 8] ; Src |
1064 |
|
mov edx, [esp+12] ; BpS |
1065 |
|
%endmacro |
1066 |
|
|
1067 |
|
%macro PROLOG 2 ; %1: Rounder, %2 load Dst-Rounder |
1068 |
|
pxor mm6, mm6 |
1069 |
|
movq mm7, [%1] ; TODO: dangerous! (eax isn't checked) |
1070 |
|
%if %2 |
1071 |
|
movq mm5, [rounding1_mmx] |
1072 |
|
%endif |
1073 |
|
|
1074 |
|
PROLOG0 |
1075 |
|
%endmacro |
1076 |
|
|
1077 |
|
; performs: mm0 == (mm0+mm2) mm1 == (mm1+mm3) |
1078 |
|
%macro MIX 0 |
1079 |
|
punpcklbw mm0, mm6 |
1080 |
|
punpcklbw mm2, mm6 |
1081 |
|
punpckhbw mm1, mm6 |
1082 |
|
punpckhbw mm3, mm6 |
1083 |
|
paddusw mm0, mm2 |
1084 |
|
paddusw mm1, mm3 |
1085 |
|
%endmacro |
1086 |
|
|
1087 |
|
%macro MIX_DST 0 |
1088 |
|
movq mm3, mm2 |
1089 |
|
paddusw mm0, mm7 ; rounder |
1090 |
|
paddusw mm1, mm7 ; rounder |
1091 |
|
punpcklbw mm2, mm6 |
1092 |
|
punpckhbw mm3, mm6 |
1093 |
|
psrlw mm0, 1 |
1094 |
|
psrlw mm1, 1 |
1095 |
|
|
1096 |
|
paddusw mm0, mm2 ; mix Src(mm0/mm1) with Dst(mm2/mm3) |
1097 |
|
paddusw mm1, mm3 |
1098 |
|
paddusw mm0, mm5 |
1099 |
|
paddusw mm1, mm5 |
1100 |
|
psrlw mm0, 1 |
1101 |
|
psrlw mm1, 1 |
1102 |
|
|
1103 |
|
packuswb mm0, mm1 |
1104 |
|
%endmacro |
1105 |
|
|
1106 |
|
%macro MIX2 0 |
1107 |
|
punpcklbw mm0, mm6 |
1108 |
|
punpcklbw mm2, mm6 |
1109 |
|
paddusw mm0, mm2 |
1110 |
|
paddusw mm0, mm7 |
1111 |
|
punpckhbw mm1, mm6 |
1112 |
|
punpckhbw mm3, mm6 |
1113 |
|
paddusw mm1, mm7 |
1114 |
|
paddusw mm1, mm3 |
1115 |
|
psrlw mm0, 1 |
1116 |
|
psrlw mm1, 1 |
1117 |
|
|
1118 |
|
packuswb mm0, mm1 |
1119 |
|
%endmacro |
1120 |
|
|
1121 |
|
;=========================================================================== |
1122 |
|
; |
1123 |
|
; void interpolate8x8_halfpel_add_mmx(uint8_t * const dst, |
1124 |
|
; const uint8_t * const src, |
1125 |
|
; const uint32_t stride, |
1126 |
|
; const uint32_t rounding); |
1127 |
|
; |
1128 |
|
; |
1129 |
|
;=========================================================================== |
1130 |
|
|
1131 |
|
%macro ADD_FF_MMX 1 |
1132 |
|
movq mm0, [eax] |
1133 |
|
movq mm2, [ecx] |
1134 |
|
movq mm1, mm0 |
1135 |
|
movq mm3, mm2 |
1136 |
|
%if (%1!=0) |
1137 |
|
lea eax,[eax+%1*edx] |
1138 |
|
%endif |
1139 |
|
MIX |
1140 |
|
paddusw mm0, mm5 ; rounder |
1141 |
|
paddusw mm1, mm5 ; rounder |
1142 |
|
psrlw mm0, 1 |
1143 |
|
psrlw mm1, 1 |
1144 |
|
|
1145 |
|
packuswb mm0, mm1 |
1146 |
|
movq [ecx], mm0 |
1147 |
|
%if (%1!=0) |
1148 |
|
lea ecx,[ecx+%1*edx] |
1149 |
|
%endif |
1150 |
|
%endmacro |
1151 |
|
|
1152 |
|
ALIGN 16 |
1153 |
|
interpolate8x8_halfpel_add_mmx: |
1154 |
|
PROLOG rounding1_mmx, 1 |
1155 |
|
ADD_FF_MMX 1 |
1156 |
|
ADD_FF_MMX 1 |
1157 |
|
ADD_FF_MMX 1 |
1158 |
|
ADD_FF_MMX 1 |
1159 |
|
ADD_FF_MMX 1 |
1160 |
|
ADD_FF_MMX 1 |
1161 |
|
ADD_FF_MMX 1 |
1162 |
|
ADD_FF_MMX 0 |
1163 |
|
ret |
1164 |
|
.endfunc |
1165 |
|
|
1166 |
|
;=========================================================================== |
1167 |
|
; |
1168 |
|
; void interpolate8x8_halfpel_h_add_mmx(uint8_t * const dst, |
1169 |
|
; const uint8_t * const src, |
1170 |
|
; const uint32_t stride, |
1171 |
|
; const uint32_t rounding); |
1172 |
|
; |
1173 |
|
; |
1174 |
|
;=========================================================================== |
1175 |
|
|
1176 |
|
%macro ADD_FH_MMX 0 |
1177 |
|
movq mm0, [eax] |
1178 |
|
movq mm2, [eax+1] |
1179 |
|
movq mm1, mm0 |
1180 |
|
movq mm3, mm2 |
1181 |
|
|
1182 |
|
lea eax,[eax+edx] |
1183 |
|
|
1184 |
|
MIX |
1185 |
|
movq mm2, [ecx] ; prepare mix with Dst[0] |
1186 |
|
MIX_DST |
1187 |
|
movq [ecx], mm0 |
1188 |
|
%endmacro |
1189 |
|
|
1190 |
|
ALIGN 16 |
1191 |
|
interpolate8x8_halfpel_h_add_mmx: |
1192 |
|
PROLOG rounding1_mmx, 1 |
1193 |
|
|
1194 |
|
ADD_FH_MMX |
1195 |
|
lea ecx,[ecx+edx] |
1196 |
|
ADD_FH_MMX |
1197 |
|
lea ecx,[ecx+edx] |
1198 |
|
ADD_FH_MMX |
1199 |
|
lea ecx,[ecx+edx] |
1200 |
|
ADD_FH_MMX |
1201 |
|
lea ecx,[ecx+edx] |
1202 |
|
ADD_FH_MMX |
1203 |
|
lea ecx,[ecx+edx] |
1204 |
|
ADD_FH_MMX |
1205 |
|
lea ecx,[ecx+edx] |
1206 |
|
ADD_FH_MMX |
1207 |
|
lea ecx,[ecx+edx] |
1208 |
|
ADD_FH_MMX |
1209 |
|
ret |
1210 |
|
.endfunc |
1211 |
|
|
1212 |
|
;=========================================================================== |
1213 |
|
; |
1214 |
|
; void interpolate8x8_halfpel_v_add_mmx(uint8_t * const dst, |
1215 |
|
; const uint8_t * const src, |
1216 |
|
; const uint32_t stride, |
1217 |
|
; const uint32_t rounding); |
1218 |
|
; |
1219 |
|
; |
1220 |
|
;=========================================================================== |
1221 |
|
|
1222 |
|
%macro ADD_HF_MMX 0 |
1223 |
|
movq mm0, [eax] |
1224 |
|
movq mm2, [eax+edx] |
1225 |
|
movq mm1, mm0 |
1226 |
|
movq mm3, mm2 |
1227 |
|
|
1228 |
|
lea eax,[eax+edx] |
1229 |
|
|
1230 |
|
MIX |
1231 |
|
movq mm2, [ecx] ; prepare mix with Dst[0] |
1232 |
|
MIX_DST |
1233 |
|
movq [ecx], mm0 |
1234 |
|
|
1235 |
|
%endmacro |
1236 |
|
|
1237 |
|
ALIGN 16 |
1238 |
|
interpolate8x8_halfpel_v_add_mmx: |
1239 |
|
PROLOG rounding1_mmx, 1 |
1240 |
|
|
1241 |
|
ADD_HF_MMX |
1242 |
|
lea ecx,[ecx+edx] |
1243 |
|
ADD_HF_MMX |
1244 |
|
lea ecx,[ecx+edx] |
1245 |
|
ADD_HF_MMX |
1246 |
|
lea ecx,[ecx+edx] |
1247 |
|
ADD_HF_MMX |
1248 |
|
lea ecx,[ecx+edx] |
1249 |
|
ADD_HF_MMX |
1250 |
|
lea ecx,[ecx+edx] |
1251 |
|
ADD_HF_MMX |
1252 |
|
lea ecx,[ecx+edx] |
1253 |
|
ADD_HF_MMX |
1254 |
|
lea ecx,[ecx+edx] |
1255 |
|
ADD_HF_MMX |
1256 |
|
ret |
1257 |
|
.endfunc |
1258 |
|
|
1259 |
|
; The trick is to correct the result of 'pavgb' with some combination of the |
1260 |
|
; lsb's of the 4 input values i,j,k,l, and their intermediate 'pavgb' (s and t). |
1261 |
|
; The boolean relations are: |
1262 |
|
; (i+j+k+l+3)/4 = (s+t+1)/2 - (ij&kl)&st |
1263 |
|
; (i+j+k+l+2)/4 = (s+t+1)/2 - (ij|kl)&st |
1264 |
|
; (i+j+k+l+1)/4 = (s+t+1)/2 - (ij&kl)|st |
1265 |
|
; (i+j+k+l+0)/4 = (s+t+1)/2 - (ij|kl)|st |
1266 |
|
; with s=(i+j+1)/2, t=(k+l+1)/2, ij = i^j, kl = k^l, st = s^t. |
1267 |
|
|
1268 |
|
; Moreover, we process 2 lines at a times, for better overlapping (~15% faster). |
1269 |
|
|
1270 |
|
;=========================================================================== |
1271 |
|
; |
1272 |
|
; void interpolate8x8_halfpel_hv_add_mmx(uint8_t * const dst, |
1273 |
|
; const uint8_t * const src, |
1274 |
|
; const uint32_t stride, |
1275 |
|
; const uint32_t rounding); |
1276 |
|
; |
1277 |
|
; |
1278 |
|
;=========================================================================== |
1279 |
|
|
1280 |
|
%macro ADD_HH_MMX 0 |
1281 |
|
lea eax,[eax+edx] |
1282 |
|
|
1283 |
|
; transfert prev line to mm0/mm1 |
1284 |
|
movq mm0, mm2 |
1285 |
|
movq mm1, mm3 |
1286 |
|
|
1287 |
|
; load new line in mm2/mm3 |
1288 |
|
movq mm2, [eax] |
1289 |
|
movq mm4, [eax+1] |
1290 |
|
movq mm3, mm2 |
1291 |
|
movq mm5, mm4 |
1292 |
|
|
1293 |
|
punpcklbw mm2, mm6 |
1294 |
|
punpcklbw mm4, mm6 |
1295 |
|
paddusw mm2, mm4 |
1296 |
|
punpckhbw mm3, mm6 |
1297 |
|
punpckhbw mm5, mm6 |
1298 |
|
paddusw mm3, mm5 |
1299 |
|
|
1300 |
|
; mix current line (mm2/mm3) with previous (mm0,mm1); |
1301 |
|
; we'll preserve mm2/mm3 for next line... |
1302 |
|
|
1303 |
|
paddusw mm0, mm2 |
1304 |
|
paddusw mm1, mm3 |
1305 |
|
|
1306 |
|
movq mm4, [ecx] ; prepare mix with Dst[0] |
1307 |
|
movq mm5, mm4 |
1308 |
|
|
1309 |
|
paddusw mm0, mm7 ; finish mixing current line |
1310 |
|
paddusw mm1, mm7 |
1311 |
|
|
1312 |
|
punpcklbw mm4, mm6 |
1313 |
|
punpckhbw mm5, mm6 |
1314 |
|
|
1315 |
|
psrlw mm0, 2 |
1316 |
|
psrlw mm1, 2 |
1317 |
|
|
1318 |
|
paddusw mm0, mm4 ; mix Src(mm0/mm1) with Dst(mm2/mm3) |
1319 |
|
paddusw mm1, mm5 |
1320 |
|
|
1321 |
|
paddusw mm0, [rounding1_mmx] |
1322 |
|
paddusw mm1, [rounding1_mmx] |
1323 |
|
|
1324 |
|
psrlw mm0, 1 |
1325 |
|
psrlw mm1, 1 |
1326 |
|
|
1327 |
|
packuswb mm0, mm1 |
1328 |
|
|
1329 |
|
movq [ecx], mm0 |
1330 |
|
%endmacro |
1331 |
|
|
1332 |
|
ALIGN 16 |
1333 |
|
interpolate8x8_halfpel_hv_add_mmx: |
1334 |
|
PROLOG rounding2_mmx, 0 ; mm5 is busy. Don't load dst-rounder |
1335 |
|
|
1336 |
|
; preprocess first line |
1337 |
|
movq mm0, [eax] |
1338 |
|
movq mm2, [eax+1] |
1339 |
|
movq mm1, mm0 |
1340 |
|
movq mm3, mm2 |
1341 |
|
|
1342 |
|
punpcklbw mm0, mm6 |
1343 |
|
punpcklbw mm2, mm6 |
1344 |
|
punpckhbw mm1, mm6 |
1345 |
|
punpckhbw mm3, mm6 |
1346 |
|
paddusw mm2, mm0 |
1347 |
|
paddusw mm3, mm1 |
1348 |
|
|
1349 |
|
; Input: mm2/mm3 contains the value (Src[0]+Src[1]) of previous line |
1350 |
|
|
1351 |
|
ADD_HH_MMX |
1352 |
|
lea ecx,[ecx+edx] |
1353 |
|
ADD_HH_MMX |
1354 |
|
lea ecx,[ecx+edx] |
1355 |
|
ADD_HH_MMX |
1356 |
|
lea ecx,[ecx+edx] |
1357 |
|
ADD_HH_MMX |
1358 |
|
lea ecx,[ecx+edx] |
1359 |
|
ADD_HH_MMX |
1360 |
|
lea ecx,[ecx+edx] |
1361 |
|
ADD_HH_MMX |
1362 |
|
lea ecx,[ecx+edx] |
1363 |
|
ADD_HH_MMX |
1364 |
|
lea ecx,[ecx+edx] |
1365 |
|
ADD_HH_MMX |
1366 |
|
|
1367 |
|
ret |
1368 |
|
.endfunc |
1369 |
|
|