Lines Matching refs:mm4
53 "movq %%mm3, %%mm4\n\t"
55 "psllq $40, %%mm4\n\t"
56 "por %%mm4, %%mm3\n\t"
57 "movq %%mm3, %%mm4\n\t"
59 "punpckhbw %%mm0, %%mm4\n\t"
65 "movq %%mm3, %%mm4\n\t"
67 "punpckhbw %%mm0, %%mm4\n\t"
86 "paddw %%mm1, %%mm4\n\t"
88 "psraw $4, %%mm4\n\t"
94 "movq %%mm4, %%mm7\n\t"
109 "paddw %%mm5, %%mm4 \n\t"\
112 "psraw $3, %%mm4 \n\t"\
113 "packuswb %%mm4, %%mm3 \n\t"
118 "movq %%mm7, %%mm4\n\t"
137 "movq %%mm3, %%mm4 \n\t"\
149 "paddw %%mm5, %%mm4 \n\t"\
192 "pxor %%mm4, %%mm4 \n\t"\
209 "paddd %%mm0, %%mm4 \n\t"\
224 "psrad $16, %%mm4 \n\t"\
229 "packssdw %%mm5, %%mm4 \n\t"\
231 "paddw %%mm0, %%mm4 \n\t"\
233 "movq %%mm4, "U_TEMP"(%0) \n\t"\
248 "movq (%%"REG_S", %%"REG_a", 2), %%mm4 \n\t" /* Y1srcData */\
250 "punpcklwd %%mm4, %%mm0 \n\t"\
251 "punpckhwd %%mm4, %%mm3 \n\t"\
252 "movq "STR(APCK_COEF)"(%%"REG_d"), %%mm4 \n\t" /* filterCoeff */\
253 "pmaddwd %%mm4, %%mm0 \n\t"\
254 "pmaddwd %%mm4, %%mm3 \n\t"\
264 "pmaddwd %%mm4, %%mm2 \n\t"\
265 "pmaddwd %%mm4, %%mm0 \n\t"\
279 "movq "V_TEMP"(%0), %%mm4 \n\t"\
287 "psubw "V_OFFSET"(%0), %%mm4 \n\t" /* (V-128)8*/\
289 "movq %%mm4, %%mm5 \n\t" /* (V-128)8*/\
291 "pmulhw "VG_COEFF"(%0), %%mm4 \n\t"\
292 /* mm2=(U-128)8, mm3=ug, mm4=vg mm5=(V-128)8 */\
299 /* mm1= Y1, mm2=ub, mm3=ug, mm4=vg mm5=vr, mm7=Y2 */\
300 "paddw %%mm3, %%mm4 \n\t"\
303 "movq %%mm4, %%mm3 \n\t"\
306 "punpcklwd %%mm4, %%mm4 \n\t"\
309 "paddw %%mm1, %%mm4 \n\t"\
316 /* mm0=B1, mm2=B2, mm3=G2, mm4=G1, mm5=R1, mm6=R2 */\
319 "packuswb %%mm3, %%mm4 \n\t"\
360 "movq %%mm4, "V_TEMP"(%0) \n\t"
367 WRITEBGR32(%4, "%5", %%REGa, %%mm3, %%mm4, %%mm5, %%mm1, %%mm0, %%mm7, %%mm2, %%mm6)
373 WRITEBGR32(%4, "%5", %%REGa, %%mm2, %%mm4, %%mm5, %%mm7, %%mm0, %%mm1, %%mm3, %%mm6)
396 WRITEBGR32(%4, "%5", %%REGa, %%mm2, %%mm4, %%mm5, %%mm1, %%mm0, %%mm7, %%mm3, %%mm6)
402 WRITEBGR32(%4, "%5", %%REGa, %%mm2, %%mm4, %%mm5, %%mm7, %%mm0, %%mm1, %%mm3, %%mm6)
409 "pand "MANGLE(bFC)", %%mm4 \n\t" /* G */\
414 "movq %%mm4, %%mm3 \n\t"\
418 "punpckhbw %%mm7, %%mm4 \n\t"\
422 "psllq $3, %%mm4 \n\t"\
425 "por %%mm4, %%mm1 \n\t"\
449 /* mm2=B, %%mm4=G, %%mm5=R, %%mm7=0 */
452 "paddusb "GREEN_DITHER"(%0), %%mm4\n\t"
473 /* mm2=B, %%mm4=G, %%mm5=R, %%mm7=0 */
476 "paddusb "GREEN_DITHER"(%0), %%mm4 \n\t"
485 "pand "MANGLE(bF8)", %%mm4 \n\t" /* G */\
491 "movq %%mm4, %%mm3 \n\t"\
495 "punpckhbw %%mm7, %%mm4 \n\t"\
499 "psllq $2, %%mm4 \n\t"\
502 "por %%mm4, %%mm1 \n\t"\
526 /* mm2=B, %%mm4=G, %%mm5=R, %%mm7=0 */
529 "paddusb "GREEN_DITHER"(%0), %%mm4\n\t"
550 /* mm2=B, %%mm4=G, %%mm5=R, %%mm7=0 */
553 "paddusb "GREEN_DITHER"(%0), %%mm4 \n\t"
561 /* mm2=B, %%mm4=G, %%mm5=R, %%mm7=0 */\
564 "punpcklbw %%mm4, %%mm2 \n\t" /* GBGBGBGB 0 */\
566 "punpckhbw %%mm4, %%mm1 \n\t" /* GBGBGBGB 2 */\
575 "movq %%mm0, %%mm4 \n\t" /* 0RGB0RGB 0 */\
585 "punpckhdq %%mm4, %%mm0 \n\t" /* 0RGBRGB0 0 */\
614 /* mm2=B, %%mm4=G, %%mm5=R, %%mm7=0 */\
618 "pshufw $0x50, %%mm4, %%mm3 \n\t" /* G3 G2 G3 G2 G1 G0 G1 G0 */\
630 "psrlq $8, %%mm4 \n\t" /* 00 G7 G6 G5 G4 G3 G2 G1 */\
632 "pshufw $0x55, %%mm4, %%mm3 \n\t" /* G4 G3 G4 G3 G4 G3 G4 G3 */\
644 "pshufw $0xFA, %%mm4, %%mm3 \n\t" /* 00 G7 00 G7 G6 G5 G6 G5 */\
723 "packuswb %%mm4, %%mm4 \n\t"\
725 "punpcklbw %%mm4, %%mm3 \n\t"\
750 /* mm2=B, %%mm4=G, %%mm5=R, %%mm7=0 */
752 "psraw $3, %%mm4 \n\t"
771 /* mm2=B, %%mm4=G, %%mm5=R, %%mm7=0 */
773 "psraw $3, %%mm4 \n\t"
788 "movq (%3, "#index"), %%mm4 \n\t" /* uvbuf1[eax+2048]*/\
791 "psubw %%mm4, %%mm5 \n\t" /* uvbuf0[eax+2048] - uvbuf1[eax+2048]*/\
796 "psraw $4, %%mm4 \n\t" /* uvbuf0[eax+2048] - uvbuf1[eax+2048] >>4*/\
798 "paddw %%mm5, %%mm4 \n\t" /* uvbuf0[eax+2048]uvalpha1 - uvbuf1[eax+2048](1-uvalpha1)*/\
800 "psubw "V_OFFSET"("#c"), %%mm4 \n\t" /* (V-128)8*/\
802 "movq %%mm4, %%mm5 \n\t" /* (V-128)8*/\
804 "pmulhw "VG_COEFF"("#c"), %%mm4 \n\t"\
805 /* mm2=(U-128)8, mm3=ug, mm4=vg mm5=(V-128)8 */\
828 /* mm1= Y1, mm2=ub, mm3=ug, mm4=vg mm5=vr, mm7=Y2 */\
829 "paddw %%mm3, %%mm4 \n\t"\
832 "movq %%mm4, %%mm3 \n\t"\
835 "punpcklwd %%mm4, %%mm4 \n\t"\
838 "paddw %%mm1, %%mm4 \n\t"\
845 /* mm0=B1, mm2=B2, mm3=G2, mm4=G1, mm5=R1, mm6=R2 */\
848 "packuswb %%mm3, %%mm4 \n\t"\
877 WRITEBGR32(%4, DSTW_OFFSET"(%5)", %%r8, %%mm2, %%mm4, %%mm5, %%mm1, %%mm0, %%mm7, %%mm3, %%mm6)
901 WRITEBGR32(%%REGb, DSTW_OFFSET"(%5)", %%REGBP, %%mm2, %%mm4, %%mm5, %%mm1, %%mm0, %%mm7, %%mm3, %%mm6)
915 WRITEBGR32(%%REGb, DSTW_OFFSET"(%5)", %%REGBP, %%mm2, %%mm4, %%mm5, %%mm7, %%mm0, %%mm1, %%mm3, %%mm6)
961 /* mm2=B, %%mm4=G, %%mm5=R, %%mm7=0 */
964 "paddusb "GREEN_DITHER"(%5), %%mm4 \n\t"
990 /* mm2=B, %%mm4=G, %%mm5=R, %%mm7=0 */
993 "paddusb "GREEN_DITHER"(%5), %%mm4 \n\t"
1019 "movq (%3, "#index"), %%mm4 \n\t" /* uvbuf1[eax+2048]*/\
1022 "psubw %%mm4, %%mm5 \n\t" /* uvbuf0[eax+2048] - uvbuf1[eax+2048]*/\
1027 "psraw $7, %%mm4 \n\t" /* uvbuf0[eax+2048] - uvbuf1[eax+2048] >>4*/\
1029 "paddw %%mm5, %%mm4 \n\t" /* uvbuf0[eax+2048]uvalpha1 - uvbuf1[eax+2048](1-uvalpha1)*/\
1072 "movq (%2, "#index"), %%mm4 \n\t" /* uvbuf0[eax+2048]*/\
1075 "psraw $4, %%mm4 \n\t" /* uvbuf0[eax+2048] - uvbuf1[eax+2048] >>4*/\
1077 "psubw "V_OFFSET"("#c"), %%mm4 \n\t" /* (V-128)8*/\
1079 "movq %%mm4, %%mm5 \n\t" /* (V-128)8*/\
1081 "pmulhw "VG_COEFF"("#c"), %%mm4 \n\t"\
1082 /* mm2=(U-128)8, mm3=ug, mm4=vg mm5=(V-128)8 */\
1093 /* mm1= Y1, mm2=ub, mm3=ug, mm4=vg mm5=vr, mm7=Y2 */\
1094 "paddw %%mm3, %%mm4 \n\t"\
1097 "movq %%mm4, %%mm3 \n\t"\
1100 "punpcklwd %%mm4, %%mm4 \n\t"\
1103 "paddw %%mm1, %%mm4 \n\t"\
1110 /* mm0=B1, mm2=B2, mm3=G2, mm4=G1, mm5=R1, mm6=R2 */\
1113 "packuswb %%mm3, %%mm4 \n\t"\
1126 "movq (%3, "#index"), %%mm4 \n\t" /* uvbuf1[eax+2048]*/\
1129 "paddw %%mm5, %%mm4 \n\t" /* uvbuf0[eax+2048] + uvbuf1[eax+2048]*/\
1131 "psrlw $5, %%mm4 \n\t" /*FIXME might overflow*/\
1133 "psubw "V_OFFSET"("#c"), %%mm4 \n\t" /* (V-128)8*/\
1135 "movq %%mm4, %%mm5 \n\t" /* (V-128)8*/\
1137 "pmulhw "VG_COEFF"("#c"), %%mm4 \n\t"\
1138 /* mm2=(U-128)8, mm3=ug, mm4=vg mm5=(V-128)8 */\
1149 /* mm1= Y1, mm2=ub, mm3=ug, mm4=vg mm5=vr, mm7=Y2 */\
1150 "paddw %%mm3, %%mm4 \n\t"\
1153 "movq %%mm4, %%mm3 \n\t"\
1156 "punpcklwd %%mm4, %%mm4 \n\t"\
1159 "paddw %%mm1, %%mm4 \n\t"\
1166 /* mm0=B1, mm2=B2, mm3=G2, mm4=G1, mm5=R1, mm6=R2 */\
1169 "packuswb %%mm3, %%mm4 \n\t"\
1201 WRITEBGR32(%%REGb, DSTW_OFFSET"(%5)", %%REGBP, %%mm2, %%mm4, %%mm5, %%mm7, %%mm0, %%mm1, %%mm3, %%mm6)
1214 WRITEBGR32(%%REGb, DSTW_OFFSET"(%5)", %%REGBP, %%mm2, %%mm4, %%mm5, %%mm7, %%mm0, %%mm1, %%mm3, %%mm6)
1230 WRITEBGR32(%%REGb, DSTW_OFFSET"(%5)", %%REGBP, %%mm2, %%mm4, %%mm5, %%mm7, %%mm0, %%mm1, %%mm3, %%mm6)
1243 WRITEBGR32(%%REGb, DSTW_OFFSET"(%5)", %%REGBP, %%mm2, %%mm4, %%mm5, %%mm7, %%mm0, %%mm1, %%mm3, %%mm6)
1310 /* mm2=B, %%mm4=G, %%mm5=R, %%mm7=0 */
1313 "paddusb "GREEN_DITHER"(%5), %%mm4 \n\t"
1331 /* mm2=B, %%mm4=G, %%mm5=R, %%mm7=0 */
1334 "paddusb "GREEN_DITHER"(%5), %%mm4 \n\t"
1363 /* mm2=B, %%mm4=G, %%mm5=R, %%mm7=0 */
1366 "paddusb "GREEN_DITHER"(%5), %%mm4 \n\t"
1384 /* mm2=B, %%mm4=G, %%mm5=R, %%mm7=0 */
1387 "paddusb "GREEN_DITHER"(%5), %%mm4 \n\t"
1406 "movq (%2, "#index"), %%mm4 \n\t" /* uvbuf0[eax+2048]*/\
1409 "psraw $7, %%mm4 \n\t" \
1425 "movq (%3, "#index"), %%mm4 \n\t" /* uvbuf1[eax+2048]*/\
1428 "paddw %%mm5, %%mm4 \n\t" /* uvbuf0[eax+2048] + uvbuf1[eax+2048]*/\
1430 "psrlw $8, %%mm4 \n\t" \