?? n1fv_64.c
字號:
DVK(KP980785280, +0.980785280403230449126182236134239036973933731); DVK(KP195090322, +0.195090322016128267848284868477022240927691618); DVK(KP382683432, +0.382683432365089771728459984030398866761344562); DVK(KP923879532, +0.923879532511286756128183189396788286822416626); DVK(KP707106781, +0.707106781186547524400844362104849039284835938); INT i; const R *xi; R *xo; xi = ri; xo = ro; for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(is), MAKE_VOLATILE_STRIDE(os)) { V T4p, T5q, Tb, T39, T2n, T3A, T6f, T6T, Tq, T3B, T6i, T76, T2i, T3a, T4w; V T5r, TI, T2p, T6C, T6V, T3h, T3E, T4L, T5u, TZ, T2q, T6F, T6U, T3e, T3D; V T4E, T5t, T23, T2N, T6t, T71, T6w, T72, T2c, T2O, T3t, T41, T5f, T5R, T5k; V T5S, T3w, T42, T1s, T2K, T6m, T6Y, T6p, T6Z, T1B, T2L, T3m, T3Y, T4Y, T5O; V T53, T5P, T3p, T3Z; { V T3, T4n, T2m, T4o, T6, T5p, T9, T5o; { V T1, T2, T2k, T2l; T1 = LD(&(xi[0]), ivs, &(xi[0])); T2 = LD(&(xi[WS(is, 32)]), ivs, &(xi[0])); T3 = VSUB(T1, T2); T4n = VADD(T1, T2); T2k = LD(&(xi[WS(is, 16)]), ivs, &(xi[0])); T2l = LD(&(xi[WS(is, 48)]), ivs, &(xi[0])); T2m = VSUB(T2k, T2l); T4o = VADD(T2k, T2l); } { V T4, T5, T7, T8; T4 = LD(&(xi[WS(is, 8)]), ivs, &(xi[0])); T5 = LD(&(xi[WS(is, 40)]), ivs, &(xi[0])); T6 = VSUB(T4, T5); T5p = VADD(T4, T5); T7 = LD(&(xi[WS(is, 56)]), ivs, &(xi[0])); T8 = LD(&(xi[WS(is, 24)]), ivs, &(xi[0])); T9 = VSUB(T7, T8); T5o = VADD(T7, T8); } T4p = VSUB(T4n, T4o); T5q = VSUB(T5o, T5p); { V Ta, T2j, T6d, T6e; Ta = VMUL(LDK(KP707106781), VADD(T6, T9)); Tb = VADD(T3, Ta); T39 = VSUB(T3, Ta); T2j = VMUL(LDK(KP707106781), VSUB(T9, T6)); T2n = VSUB(T2j, T2m); T3A = VADD(T2m, T2j); T6d = VADD(T4n, T4o); T6e = VADD(T5p, T5o); T6f = VADD(T6d, T6e); T6T = VSUB(T6d, T6e); } } { V Te, T4q, To, T4u, Th, T4r, Tl, T4t; { V Tc, Td, Tm, Tn; Tc = LD(&(xi[WS(is, 4)]), ivs, &(xi[0])); Td = LD(&(xi[WS(is, 36)]), ivs, &(xi[0])); Te = VSUB(Tc, Td); T4q = VADD(Tc, Td); Tm = LD(&(xi[WS(is, 12)]), ivs, &(xi[0])); Tn = LD(&(xi[WS(is, 44)]), ivs, &(xi[0])); To = VSUB(Tm, Tn); T4u = VADD(Tm, Tn); } { V Tf, Tg, Tj, Tk; Tf = LD(&(xi[WS(is, 20)]), ivs, &(xi[0])); Tg = LD(&(xi[WS(is, 52)]), ivs, &(xi[0])); Th = VSUB(Tf, Tg); T4r = VADD(Tf, Tg); Tj = LD(&(xi[WS(is, 60)]), ivs, &(xi[0])); Tk = LD(&(xi[WS(is, 28)]), ivs, &(xi[0])); Tl = VSUB(Tj, Tk); T4t = VADD(Tj, Tk); } { V Ti, Tp, T6g, T6h; Ti = VFNMS(LDK(KP382683432), Th, VMUL(LDK(KP923879532), Te)); Tp = VFMA(LDK(KP923879532), Tl, VMUL(LDK(KP382683432), To)); Tq = VADD(Ti, Tp); T3B = VSUB(Tp, Ti); T6g = VADD(T4q, T4r); T6h = VADD(T4t, T4u); T6i = VADD(T6g, T6h); T76 = VSUB(T6h, T6g); } { V T2g, T2h, T4s, T4v; T2g = VFNMS(LDK(KP923879532), To, VMUL(LDK(KP382683432), Tl)); T2h = VFMA(LDK(KP382683432), Te, VMUL(LDK(KP923879532), Th)); T2i = VSUB(T2g, T2h); T3a = VADD(T2h, T2g); T4s = VSUB(T4q, T4r); T4v = VSUB(T4t, T4u); T4w = VMUL(LDK(KP707106781), VADD(T4s, T4v)); T5r = VMUL(LDK(KP707106781), VSUB(T4v, T4s)); } } { V Tu, T4F, TG, T4G, TB, T4J, TD, T4I; { V Ts, Tt, TE, TF; Ts = LD(&(xi[WS(is, 62)]), ivs, &(xi[0])); Tt = LD(&(xi[WS(is, 30)]), ivs, &(xi[0])); Tu = VSUB(Ts, Tt); T4F = VADD(Ts, Tt); TE = LD(&(xi[WS(is, 14)]), ivs, &(xi[0])); TF = LD(&(xi[WS(is, 46)]), ivs, &(xi[0])); TG = VSUB(TE, TF); T4G = VADD(TE, TF); { V Tv, Tw, Tx, Ty, Tz, TA; Tv = LD(&(xi[WS(is, 6)]), ivs, &(xi[0])); Tw = LD(&(xi[WS(is, 38)]), ivs, &(xi[0])); Tx = VSUB(Tv, Tw); Ty = LD(&(xi[WS(is, 54)]), ivs, &(xi[0])); Tz = LD(&(xi[WS(is, 22)]), ivs, &(xi[0])); TA = VSUB(Ty, Tz); TB = VMUL(LDK(KP707106781), VADD(Tx, TA)); T4J = VADD(Tv, Tw); TD = VMUL(LDK(KP707106781), VSUB(TA, Tx)); T4I = VADD(Ty, Tz); } } { V TC, TH, T6A, T6B; TC = VADD(Tu, TB); TH = VSUB(TD, TG); TI = VFMA(LDK(KP195090322), TC, VMUL(LDK(KP980785280), TH)); T2p = VFNMS(LDK(KP195090322), TH, VMUL(LDK(KP980785280), TC)); T6A = VADD(T4F, T4G); T6B = VADD(T4J, T4I); T6C = VADD(T6A, T6B); T6V = VSUB(T6A, T6B); } { V T3f, T3g, T4H, T4K; T3f = VSUB(Tu, TB); T3g = VADD(TG, TD); T3h = VFNMS(LDK(KP555570233), T3g, VMUL(LDK(KP831469612), T3f)); T3E = VFMA(LDK(KP555570233), T3f, VMUL(LDK(KP831469612), T3g)); T4H = VSUB(T4F, T4G); T4K = VSUB(T4I, T4J); T4L = VFNMS(LDK(KP382683432), T4K, VMUL(LDK(KP923879532), T4H)); T5u = VFMA(LDK(KP382683432), T4H, VMUL(LDK(KP923879532), T4K)); } } { V TS, T4z, TW, T4y, TP, T4C, TX, T4B; { V TQ, TR, TU, TV; TQ = LD(&(xi[WS(is, 18)]), ivs, &(xi[0])); TR = LD(&(xi[WS(is, 50)]), ivs, &(xi[0])); TS = VSUB(TQ, TR); T4z = VADD(TQ, TR); TU = LD(&(xi[WS(is, 2)]), ivs, &(xi[0])); TV = LD(&(xi[WS(is, 34)]), ivs, &(xi[0])); TW = VSUB(TU, TV); T4y = VADD(TU, TV); { V TJ, TK, TL, TM, TN, TO; TJ = LD(&(xi[WS(is, 58)]), ivs, &(xi[0])); TK = LD(&(xi[WS(is, 26)]), ivs, &(xi[0])); TL = VSUB(TJ, TK); TM = LD(&(xi[WS(is, 10)]), ivs, &(xi[0])); TN = LD(&(xi[WS(is, 42)]), ivs, &(xi[0])); TO = VSUB(TM, TN); TP = VMUL(LDK(KP707106781), VSUB(TL, TO)); T4C = VADD(TM, TN); TX = VMUL(LDK(KP707106781), VADD(TO, TL)); T4B = VADD(TJ, TK); } } { V TT, TY, T6D, T6E; TT = VSUB(TP, TS); TY = VADD(TW, TX); TZ = VFNMS(LDK(KP195090322), TY, VMUL(LDK(KP980785280), TT)); T2q = VFMA(LDK(KP980785280), TY, VMUL(LDK(KP195090322), TT)); T6D = VADD(T4y, T4z); T6E = VADD(T4C, T4B); T6F = VADD(T6D, T6E); T6U = VSUB(T6D, T6E); } { V T3c, T3d, T4A, T4D; T3c = VSUB(TW, TX); T3d = VADD(TS, TP); T3e = VFMA(LDK(KP831469612), T3c, VMUL(LDK(KP555570233), T3d)); T3D = VFNMS(LDK(KP555570233), T3c, VMUL(LDK(KP831469612), T3d)); T4A = VSUB(T4y, T4z); T4D = VSUB(T4B, T4C); T4E = VFMA(LDK(KP923879532), T4A, VMUL(LDK(KP382683432), T4D)); T5t = VFNMS(LDK(KP382683432), T4A, VMUL(LDK(KP923879532), T4D)); } } { V T1F, T55, T2a, T56, T1M, T5h, T27, T5g, T58, T59, T1U, T5a, T25, T5b, T5c; V T21, T5d, T24; { V T1D, T1E, T28, T29; T1D = LD(&(xi[WS(is, 63)]), ivs, &(xi[WS(is, 1)])); T1E = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)])); T1F = VSUB(T1D, T1E); T55 = VADD(T1D, T1E); T28 = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)])); T29 = LD(&(xi[WS(is, 47)]), ivs, &(xi[WS(is, 1)])); T2a = VSUB(T28, T29); T56 = VADD(T28, T29); } { V T1G, T1H, T1I, T1J, T1K, T1L; T1G = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)])); T1H = LD(&(xi[WS(is, 39)]), ivs, &(xi[WS(is, 1)])); T1I = VSUB(T1G, T1H); T1J = LD(&(xi[WS(is, 55)]), ivs, &(xi[WS(is, 1)])); T1K = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)])); T1L = VSUB(T1J, T1K); T1M = VMUL(LDK(KP707106781), VADD(T1I, T1L)); T5h = VADD(T1G, T1H); T27 = VMUL(LDK(KP707106781), VSUB(T1L, T1I)); T5g = VADD(T1J, T1K); } { V T1Q, T1T, T1X, T20; { V T1O, T1P, T1R, T1S; T1O = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)])); T1P = LD(&(xi[WS(is, 35)]), ivs, &(xi[WS(is, 1)])); T1Q = VSUB(T1O, T1P); T58 = VADD(T1O, T1P); T1R = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)])); T1S = LD(&(xi[WS(is, 51)]), ivs, &(xi[WS(is, 1)])); T1T = VSUB(T1R, T1S); T59 = VADD(T1R, T1S); } T1U = VFNMS(LDK(KP382683432), T1T, VMUL(LDK(KP923879532), T1Q)); T5a = VSUB(T58, T59); T25 = VFMA(LDK(KP382683432), T1Q, VMUL(LDK(KP923879532), T1T)); { V T1V, T1W, T1Y, T1Z; T1V = LD(&(xi[WS(is, 59)]), ivs, &(xi[WS(is, 1)])); T1W = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)])); T1X = VSUB(T1V, T1W); T5b = VADD(T1V, T1W); T1Y = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)])); T1Z = LD(&(xi[WS(is, 43)]), ivs, &(xi[WS(is, 1)])); T20 = VSUB(T1Y, T1Z); T5c = VADD(T1Y, T1Z); } T21 = VFMA(LDK(KP923879532), T1X, VMUL(LDK(KP382683432), T20)); T5d = VSUB(T5b, T5c); T24 = VFNMS(LDK(KP923879532), T20, VMUL(LDK(KP382683432), T1X)); } { V T1N, T22, T6r, T6s; T1N = VADD(T1F, T1M); T22 = VADD(T1U, T21); T23 = VSUB(T1N, T22); T2N = VADD(T1N, T22); T6r = VADD(T55, T56); T6s = VADD(T5h, T5g); T6t = VADD(T6r, T6s); T71 = VSUB(T6r, T6s); } { V T6u, T6v, T26, T2b; T6u = VADD(T58, T59); T6v = VADD(T5b, T5c); T6w = VADD(T6u, T6v); T72 = VSUB(T6v, T6u); T26 = VSUB(T24, T25); T2b = VSUB(T27, T2a); T2c = VSUB(T26, T2b); T2O = VADD(T2b, T26); } { V T3r, T3s, T57, T5e; T3r = VSUB(T1F, T1M); T3s = VADD(T25, T24); T3t = VADD(T3r, T3s); T41 = VSUB(T3r, T3s); T57 = VSUB(T55, T56); T5e = VMUL(LDK(KP707106781), VADD(T5a, T5d)); T5f = VADD(T57, T5e); T5R = VSUB(T57, T5e); } { V T5i, T5j, T3u, T3v; T5i = VSUB(T5g, T5h); T5j = VMUL(LDK(KP707106781), VSUB(T5d, T5a)); T5k = VADD(T5i, T5j); T5S = VSUB(T5j, T5i); T3u = VADD(T2a, T27); T3v = VSUB(T21, T1U); T3w = VADD(T3u, T3v); T42 = VSUB(T3v, T3u); } } { V T1q, T4P, T1v, T4O, T1n, T50, T1w, T4Z, T4U, T4V, T18, T4W, T1z, T4R, T4S; V T1f, T4T, T1y; { V T1o, T1p, T1t, T1u; T1o = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)])); T1p = LD(&(xi[WS(is, 49)]), ivs, &(xi[WS(is, 1)])); T1q = VSUB(T1o, T1p); T4P = VADD(T1o, T1p); T1t = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)])); T1u = LD(&(xi[WS(is, 33)]), ivs, &(xi[WS(is, 1)])); T1v = VSUB(T1t, T1u); T4O = VADD(T1t, T1u); } { V T1h, T1i, T1j, T1k, T1l, T1m; T1h = LD(&(xi[WS(is, 57)]), ivs, &(xi[WS(is, 1)])); T1i = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)])); T1j = VSUB(T1h, T1i); T1k = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)])); T1l = LD(&(xi[WS(is, 41)]), ivs, &(xi[WS(is, 1)])); T1m = VSUB(T1k, T1l); T1n = VMUL(LDK(KP707106781), VSUB(T1j, T1m)); T50 = VADD(T1k, T1l); T1w = VMUL(LDK(KP707106781), VADD(T1m, T1j)); T4Z = VADD(T1h, T1i); } { V T14, T17, T1b, T1e; { V T12, T13, T15, T16; T12 = LD(&(xi[WS(is, 61)]), ivs, &(xi[WS(is, 1)])); T13 = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)])); T14 = VSUB(T12, T13); T4U = VADD(T12, T13); T15 = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)])); T16 = LD(&(xi[WS(is, 45)]), ivs, &(xi[WS(is, 1)])); T17 = VSUB(T15, T16); T4V = VADD(T15, T16); } T18 = VFNMS(LDK(KP923879532), T17, VMUL(LDK(KP382683432), T14)); T4W = VSUB(T4U, T4V); T1z = VFMA(LDK(KP923879532), T14, VMUL(LDK(KP382683432), T17)); { V T19, T1a, T1c, T1d; T19 = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)])); T1a = LD(&(xi[WS(is, 37)]), ivs, &(xi[WS(is, 1)])); T1b = VSUB(T19, T1a); T4R = VADD(T19, T1a); T1c = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)])); T1d = LD(&(xi[WS(is, 53)]), ivs, &(xi[WS(is, 1)])); T1e = VSUB(T1c, T1d); T4S = VADD(T1c, T1d); } T1f = VFMA(LDK(KP382683432), T1b, VMUL(LDK(KP923879532), T1e)); T4T = VSUB(T4R, T4S); T1y = VFNMS(LDK(KP382683432), T1e, VMUL(LDK(KP923879532), T1b)); } { V T1g, T1r, T6k, T6l; T1g = VSUB(T18, T1f); T1r = VSUB(T1n, T1q); T1s = VSUB(T1g, T1r); T2K = VADD(T1r, T1g); T6k = VADD(T4O, T4P); T6l = VADD(T50, T4Z); T6m = VADD(T6k, T6l); T6Y = VSUB(T6k, T6l); } { V T6n, T6o, T1x, T1A; T6n = VADD(T4R, T4S); T6o = VADD(T4U, T4V); T6p = VADD(T6n, T6o); T6Z = VSUB(T6o, T6n); T1x = VADD(T1v, T1w); T1A = VADD(T1y, T1z); T1B = VSUB(T1x, T1A); T2L = VADD(T1x, T1A); } { V T3k, T3l, T4Q, T4X; T3k = VSUB(T1v, T1w); T3l = VADD(T1f, T18); T3m = VADD(T3k, T3l); T3Y = VSUB(T3k, T3l); T4Q = VSUB(T4O, T4P);
?? 快捷鍵說明
復制代碼
Ctrl + C
搜索代碼
Ctrl + F
全屏模式
F11
切換主題
Ctrl + Shift + D
顯示快捷鍵
?
增大字號
Ctrl + =
減小字號
Ctrl + -