Lines Matching refs:VxxV

389 …packo(Vu32.b)", "Unpack byte to odd bytes ",       fVARRAY_ELEMENT_ACCESS(VxxV, uh, i) |= fZE8_16(…
390 …packo(Vu32.h)", "Unpack halves to odd halves", fVARRAY_ELEMENT_ACCESS(VxxV, uw, i) |= fZE16_32…
507 VxxV.v[0].h[i] += fMPY8US(fGETUBYTE(0, VuuV.v[0].uh[i]),fGETBYTE((2*i) % 4, RtV));
508 VxxV.v[0].h[i] += fMPY8US(fGETUBYTE(1, VuuV.v[0].uh[i]),fGETBYTE((2*i+1)%4, RtV));
510 VxxV.v[1].h[i] += fMPY8US(fGETUBYTE(1, VuuV.v[0].uh[i]),fGETBYTE((2*i) % 4, RtV));
511 VxxV.v[1].h[i] += fMPY8US(fGETUBYTE(0, VuuV.v[1].uh[i]),fGETBYTE((2*i+1)%4, RtV)))
541 VxxV.v[0].w[i] += fMPY16SS(fGETHALF(0, VuuV.v[0].w[i]),fGETBYTE((2*i+0)%4, RtV));
542 VxxV.v[0].w[i] += fMPY16SS(fGETHALF(1, VuuV.v[0].w[i]),fGETBYTE((2*i+1)%4, RtV));
544 VxxV.v[1].w[i] += fMPY16SS(fGETHALF(1, VuuV.v[0].w[i]),fGETBYTE((2*i+0)%4, RtV));
545 VxxV.v[1].w[i] += fMPY16SS(fGETHALF(0, VuuV.v[1].w[i]),fGETBYTE((2*i+1)%4, RtV)))
659 VxxV.v[0].h[i] += fMPY8SS(fGETBYTE(0,VuuV.v[0].h[i]), fGETBYTE((2*i )%4, RtV));
660 VxxV.v[0].h[i] += fMPY8SS(fGETBYTE(1,VuuV.v[0].h[i]), fGETBYTE((2*i+1)%4, RtV));
661 VxxV.v[0].h[i] += fGETBYTE(0,VuuV.v[1].h[i]);
663 VxxV.v[1].h[i] += fMPY8SS(fGETBYTE(1,VuuV.v[0].h[i]), fGETBYTE((2*i )%4, RtV));
664 VxxV.v[1].h[i] += fMPY8SS(fGETBYTE(0,VuuV.v[1].h[i]), fGETBYTE((2*i+1)%4, RtV));
665 VxxV.v[1].h[i] += fGETBYTE(1,VuuV.v[1].h[i]))
681 VxxV.v[0].h[i] += fMPY8US(fGETUBYTE(0,VuuV.v[0].uh[i]), fGETBYTE((2*i )%4, RtV));
682 VxxV.v[0].h[i] += fMPY8US(fGETUBYTE(1,VuuV.v[0].uh[i]), fGETBYTE((2*i+1)%4, RtV));
683 VxxV.v[0].h[i] += fGETUBYTE(0,VuuV.v[1].uh[i]);
685 VxxV.v[1].h[i] += fMPY8US(fGETUBYTE(1,VuuV.v[0].uh[i]), fGETBYTE((2*i )%4, RtV));
686 VxxV.v[1].h[i] += fMPY8US(fGETUBYTE(0,VuuV.v[1].uh[i]), fGETBYTE((2*i+1)%4, RtV));
687 VxxV.v[1].h[i] += fGETUBYTE(1,VuuV.v[1].uh[i]))
702 VxxV.v[0].w[i]+= fMPY16SS(fGETHALF(0,VuuV.v[0].w[i]), fSE8_16(fGETBYTE((2*i+0)%4, RtV)));
703 VxxV.v[0].w[i]+= fMPY16SS(fGETHALF(1,VuuV.v[0].w[i]), fSE8_16(fGETBYTE((2*i+1)%4, RtV)));
704 VxxV.v[0].w[i]+= fGETHALF(0,VuuV.v[1].w[i]);
706 VxxV.v[1].w[i]+= fMPY16SS(fGETHALF(1,VuuV.v[0].w[i]), fSE8_16(fGETBYTE((2*i+0)%4, RtV)));
707 VxxV.v[1].w[i]+= fMPY16SS(fGETHALF(0,VuuV.v[1].w[i]), fSE8_16(fGETBYTE((2*i+1)%4, RtV)));
708 VxxV.v[1].w[i]+= fGETHALF(1,VuuV.v[1].w[i]))
777 VxxV.v[0].uw[i] += fMPY8UU(fGETUBYTE(0, VuuV.v[uiV ? 1:0].uw[i]),fGETUBYTE((0-uiV) & 0x3,RtV));
778 VxxV.v[0].uw[i] += fMPY8UU(fGETUBYTE(1, VuuV.v[0 ].uw[i]),fGETUBYTE((1-uiV) & 0x3,RtV));
779 VxxV.v[0].uw[i] += fMPY8UU(fGETUBYTE(2, VuuV.v[0 ].uw[i]),fGETUBYTE((2-uiV) & 0x3,RtV));
780 VxxV.v[0].uw[i] += fMPY8UU(fGETUBYTE(3, VuuV.v[0 ].uw[i]),fGETUBYTE((3-uiV) & 0x3,RtV));
782 VxxV.v[1].uw[i] += fMPY8UU(fGETUBYTE(0, VuuV.v[1 ].uw[i]),fGETUBYTE((2-uiV) & 0x3,RtV));
783 VxxV.v[1].uw[i] += fMPY8UU(fGETUBYTE(1, VuuV.v[1 ].uw[i]),fGETUBYTE((3-uiV) & 0x3,RtV));
784 VxxV.v[1].uw[i] += fMPY8UU(fGETUBYTE(2, VuuV.v[uiV ? 1:0].uw[i]),fGETUBYTE((0-uiV) & 0x3,RtV));
785 VxxV.v[1].uw[i] += fMPY8UU(fGETUBYTE(3, VuuV.v[0 ].uw[i]),fGETUBYTE((1-uiV) & 0x3,RtV)))
825 VxxV.v[0].w[i] += fMPY8US(fGETUBYTE(0, VuuV.v[uiV ? 1:0].uw[i]),fGETBYTE((0-uiV) & 0x3,RtV));
826 VxxV.v[0].w[i] += fMPY8US(fGETUBYTE(1, VuuV.v[0 ].uw[i]),fGETBYTE((1-uiV) & 0x3,RtV));
827 VxxV.v[0].w[i] += fMPY8US(fGETUBYTE(2, VuuV.v[0 ].uw[i]),fGETBYTE((2-uiV) & 0x3,RtV));
828 VxxV.v[0].w[i] += fMPY8US(fGETUBYTE(3, VuuV.v[0 ].uw[i]),fGETBYTE((3-uiV) & 0x3,RtV));
830 VxxV.v[1].w[i] += fMPY8US(fGETUBYTE(0, VuuV.v[1 ].uw[i]),fGETBYTE((2-uiV) & 0x3,RtV));
831 VxxV.v[1].w[i] += fMPY8US(fGETUBYTE(1, VuuV.v[1 ].uw[i]),fGETBYTE((3-uiV) & 0x3,RtV));
832 VxxV.v[1].w[i] += fMPY8US(fGETUBYTE(2, VuuV.v[uiV ? 1:0].uw[i]),fGETBYTE((0-uiV) & 0x3,RtV));
833 VxxV.v[1].w[i] += fMPY8US(fGETUBYTE(3, VuuV.v[0 ].uw[i]),fGETBYTE((1-uiV) & 0x3,RtV)))
876 VxxV.v[0].uw[i] += fABS(fGETUHALF(0, VuuV.v[0].uw[i]) - fGETUHALF(0,RtV));
877 VxxV.v[0].uw[i] += fABS(fGETUHALF(1, VuuV.v[0].uw[i]) - fGETUHALF(1,RtV));
878 VxxV.v[1].uw[i] += fABS(fGETUHALF(1, VuuV.v[0].uw[i]) - fGETUHALF(0,RtV));
879 VxxV.v[1].uw[i] += fABS(fGETUHALF(0, VuuV.v[1].uw[i]) - fGETUHALF(1,RtV)))
904VxxV.v[0].uw[i] += fABS(fZE8_16(fGETUBYTE(0, VuuV.v[uiV?1:0].uw[i])) - fZE8_16(fGETUBYTE((0-uiV)&3…
905VxxV.v[0].uw[i] += fABS(fZE8_16(fGETUBYTE(1, VuuV.v[0 ].uw[i])) - fZE8_16(fGETUBYTE((1-uiV)&3…
906VxxV.v[0].uw[i] += fABS(fZE8_16(fGETUBYTE(2, VuuV.v[0 ].uw[i])) - fZE8_16(fGETUBYTE((2-uiV)&3…
907VxxV.v[0].uw[i] += fABS(fZE8_16(fGETUBYTE(3, VuuV.v[0 ].uw[i])) - fZE8_16(fGETUBYTE((3-uiV)&3…
909VxxV.v[1].uw[i] += fABS(fZE8_16(fGETUBYTE(0, VuuV.v[1 ].uw[i])) - fZE8_16(fGETUBYTE((2-uiV)&3…
910VxxV.v[1].uw[i] += fABS(fZE8_16(fGETUBYTE(1, VuuV.v[1 ].uw[i])) - fZE8_16(fGETUBYTE((3-uiV)&3…
911VxxV.v[1].uw[i] += fABS(fZE8_16(fGETUBYTE(2, VuuV.v[uiV?1:0].uw[i])) - fZE8_16(fGETUBYTE((0-uiV)&3…
912VxxV.v[1].uw[i] += fABS(fZE8_16(fGETUBYTE(3, VuuV.v[0 ].uw[i])) - fZE8_16(fGETUBYTE((1-uiV)&3…
949 fHIDE(int64_t ) mask = (((fSE32_64(VxxV.v[0].w[i])) << 32) | fZE32_64(VxxV.v[0].w[i]));
953 VxxV.v[1].w[i] = ((result >> 32) & 0xffffffff);
954 VxxV.v[0].w[i] = (result & 0xffffffff))
1254 VxxV.v[0].w[i] += fGETHALF(0, VuV.w[i]) + fGETHALF(0, VvV.w[i]);
1255 VxxV.v[1].w[i] += fGETHALF(1, VuV.w[i]) + fGETHALF(1, VvV.w[i]))
1259 VxxV.v[0].w[i] += fGETUHALF(0, VuV.w[i]) + fGETUHALF(0, VvV.w[i]);
1260 VxxV.v[1].w[i] += fGETUHALF(1, VuV.w[i]) + fGETUHALF(1, VvV.w[i]))
1264 VxxV.v[0].h[i] += fGETUBYTE(0, VuV.h[i]) + fGETUBYTE(0, VvV.h[i]);
1265 VxxV.v[1].h[i] += fGETUBYTE(1, VuV.h[i]) + fGETUBYTE(1, VvV.h[i]))
1309 VxxV.v[0].h[i] += fMPY8SS(fGETBYTE(0, VuV.h[i]), fGETBYTE(0, VvV.h[i]));
1310 VxxV.v[1].h[i] += fMPY8SS(fGETBYTE(1, VuV.h[i]), fGETBYTE(1, VvV.h[i])))
1320 VxxV.v[0].uh[i] += fMPY8UU(fGETUBYTE(0, VuV.uh[i]), fGETUBYTE(0, VvV.uh[i]) );
1321 VxxV.v[1].uh[i] += fMPY8UU(fGETUBYTE(1, VuV.uh[i]), fGETUBYTE(1, VvV.uh[i]) ))
1335 VxxV.v[0].h[i] += fMPY8US(fGETUBYTE(0, VuV.uh[i]), fGETBYTE(0, VvV.h[i]));
1336 VxxV.v[1].h[i] += fMPY8US(fGETUBYTE(1, VuV.uh[i]), fGETBYTE(1, VvV.h[i])))
1364 VxxV.v[0].w[i] += fMPY16SS(fGETHALF(0, VuV.w[i]), fGETHALF(0, VvV.w[i]));
1365 VxxV.v[1].w[i] += fMPY16SS(fGETHALF(1, VuV.w[i]), fGETHALF(1, VvV.w[i])))
1374 VxxV.v[0].uw[i] += fMPY16UU(fGETUHALF(0, VuV.uw[i]), fGETUHALF(0, VvV.uw[i]));
1375 VxxV.v[1].uw[i] += fMPY16UU(fGETUHALF(1, VuV.uw[i]), fGETUHALF(1, VvV.uw[i])))
1399 VxxV.v[0].w[i] += fMPY16SU(fGETHALF(0, VuV.w[i]), fGETUHALF(0, VvV.uw[i]));
1400 VxxV.v[1].w[i] += fMPY16SU(fGETHALF(1, VuV.w[i]), fGETUHALF(1, VvV.uw[i])))
1440 prod = fMPY32SS(VuV.w[i],fGETHALF(1,VvV.w[i])) + fSE32_64(VxxV.v[1].w[i]);
1441 VxxV.v[1].w[i] = prod >> 16;
1442 fSETHALF(0, VxxV.v[0].w[i], VxxV.v[0].w[i] >> 16);
1443 fSETHALF(1, VxxV.v[0].w[i], prod & 0x0000ffff))
1490 VxxV.v[0].uh[i] += fMPY8UU(fGETUBYTE(0, VuV.uh[i]), fGETUBYTE((2*i+0)%4, RtV));
1491 VxxV.v[1].uh[i] += fMPY8UU(fGETUBYTE(1, VuV.uh[i]), fGETUBYTE((2*i+1)%4, RtV)))
1501 VxxV.v[0].h[i] += fMPY8US(fGETUBYTE(0, VuV.uh[i]), fGETBYTE((2*i+0)%4, RtV));
1502 VxxV.v[1].h[i] += fMPY8US(fGETUBYTE(1, VuV.uh[i]), fGETBYTE((2*i+1)%4, RtV)))
1512VxxV.v[0].h[i] += fMPY8US(fGETUBYTE(0, VuuV.v[0].uh[i]), fGETBYTE(0, RtV)) + fMPY16SS(fGETUBYTE(0,…
1513VxxV.v[1].h[i] += fMPY8US(fGETUBYTE(1, VuuV.v[0].uh[i]), fGETBYTE(2, RtV)) + fMPY16SS(fGETUBYTE(1,…
1524VxxV.v[0].uh[i] += fMPY8UU(fGETUBYTE(0, VuuV.v[0].uh[i]), fGETUBYTE(0, RtV)) + fMPY8UU(fGETUBYTE(0…
1525VxxV.v[1].uh[i] += fMPY8UU(fGETUBYTE(1, VuuV.v[0].uh[i]), fGETUBYTE(2, RtV)) + fMPY8UU(fGETUBYTE(1…
1538VxxV.v[0].w[i] += fMPY16SS(fGETHALF(0, VuuV.v[0].w[i]), fSE8_16(fGETBYTE(0, RtV))) + fMPY16SS(fGET…
1539VxxV.v[1].w[i] += fMPY16SS(fGETHALF(1, VuuV.v[0].w[i]), fSE8_16(fGETBYTE(2, RtV))) + fMPY16SS(fGET…
1549VxxV.v[0].w[i] += fMPY16US(fGETUHALF(0, VuuV.v[0].w[i]), fSE8_16(fGETBYTE(0, RtV))) + fMPY16US(fGE…
1550VxxV.v[1].w[i] += fMPY16US(fGETUHALF(1, VuuV.v[0].w[i]), fSE8_16(fGETBYTE(2, RtV))) + fMPY16US(fGE…
1566 VxxV.v[0].w[i] = fCAST8s(VxxV.v[0].w[i]) + fMPY16SS(fGETHALF(0, VuV.w[i]), fGETHALF(0, RtV));
1567 VxxV.v[1].w[i] = fCAST8s(VxxV.v[1].w[i]) + fMPY16SS(fGETHALF(1, VuV.w[i]), fGETHALF(1, RtV)))
1572VxxV.v[0].w[i] = fVSATW(fCAST8s(VxxV.v[0].w[i]) + fMPY16SS(fGETHALF(0, VuV.w[i]), fGETHALF(0, RtV…
1573VxxV.v[1].w[i] = fVSATW(fCAST8s(VxxV.v[1].w[i]) + fMPY16SS(fGETHALF(1, VuV.w[i]), fGETHALF(1, RtV…
1598 VxxV.v[0].uw[i] += fMPY16UU(fGETUHALF(0, VuV.uw[i]),fGETUHALF(0,RtV));
1599 VxxV.v[1].uw[i] += fMPY16UU(fGETUHALF(1, VuV.uw[i]),fGETUHALF(1,RtV)))
2251 VxxV.v[0].h[i] |= ((idx & 0xF0) == (matchval << 4)) ? fGETHALF(oddhalf,VvV.w[idx % fVELEM(32)]) : 0;
2253 VxxV.v[1].h[i] |= ((idx & 0xF0) == (matchval << 4)) ? fGETHALF(oddhalf,VvV.w[idx % fVELEM(32)]) : 0)
2284 VxxV.v[0].h[i] |= ((idx & 0xF0) == (matchval << 4)) ? fGETHALF(oddhalf,VvV.w[idx % fVELEM(32)]) : 0;
2286 VxxV.v[1].h[i] |= ((idx & 0xF0) == (matchval << 4)) ? fGETHALF(oddhalf,VvV.w[idx % fVELEM(32)]) : 0)
2570 VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[0].uw[i]), c10);
2571 VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[1].uw[i]), c11);
2572 VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[1].uw[i]), c12);
2574 VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[0].uw[i]), c00);
2575 VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[1].uw[i]), c01);
2576 VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[1].uw[i]), c02);
2578 VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[0].uw[i]), c10);
2579 VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[1].uw[i]), c11);
2580 VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[1].uw[i]), c12);
2583 VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[0].uw[i]), c00);
2584 VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[1].uw[i]), c01);
2585 VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[1].uw[i]), c02);
2587 VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[0].uw[i]), c10);
2588 VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[1].uw[i]), c11);
2589 VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[1].uw[i]), c12);
2591 VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[0].uw[i]), c00);
2592 VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[1].uw[i]), c01);
2593 VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[1].uw[i]), c02);
2596 VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[0].uw[i]), c10);
2597 VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[0].uw[i]), c11);
2598 VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[1].uw[i]), c12);
2600 VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[0].uw[i]), c00);
2601 VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[0].uw[i]), c01);
2602 VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[1].uw[i]), c02);
2604 VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[0].uw[i]), c10);
2605 VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[0].uw[i]), c11);
2606 VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[1].uw[i]), c12);
2609 VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[0].uw[i]), c00);
2610 VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[0].uw[i]), c01);
2611 VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[1].uw[i]), c02);
2613 VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[0].uw[i]), c10);
2614 VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[0].uw[i]), c11);
2615 VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[1].uw[i]), c12);
2617 VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[0].uw[i]), c00);
2618 VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[0].uw[i]), c01);
2619 VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[1].uw[i]), c02);
2637 VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[1].uw[i]), c10);
2638 VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[1].uw[i]), c11);
2639 VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[0].uw[i]), c12);
2641 VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[1].uw[i]), c00);
2642 VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[1].uw[i]), c01);
2643 VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[0].uw[i]), c02);
2645 VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[1].uw[i]), c10);
2646 VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[1].uw[i]), c11);
2647 VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[0].uw[i]), c12);
2650 VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[1].uw[i]), c00);
2651 VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[1].uw[i]), c01);
2652 VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[0].uw[i]), c02);
2654 VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[1].uw[i]), c10);
2655 VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[1].uw[i]), c11);
2656 VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[0].uw[i]), c12);
2658 VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[1].uw[i]), c00);
2659 VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[1].uw[i]), c01);
2660 VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[0].uw[i]), c02);
2663 VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[1].uw[i]), c10);
2664 VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[0].uw[i]), c11);
2665 VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[0].uw[i]), c12);
2667 VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[1].uw[i]), c00);
2668 VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[0].uw[i]), c01);
2669 VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[0].uw[i]), c02);
2671 VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[1].uw[i]), c10);
2672 VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[0].uw[i]), c11);
2673 VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[0].uw[i]), c12);
2676 VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[1].uw[i]), c00);
2677 VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[0].uw[i]), c01);
2678 VxxV.v[1].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[0].uw[i]), c02);
2680 VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[1].uw[i]), c10);
2681 VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(3,VuuV.v[0].uw[i]), c11);
2682 VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(1,VuuV.v[0].uw[i]), c12);
2684 VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[1].uw[i]), c00);
2685 VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(2,VuuV.v[0].uw[i]), c01);
2686 VxxV.v[0].w[i] += fMPY16US(fGETUBYTE(0,VuuV.v[0].uw[i]), c02);