📄 n1bv_64.c
字号:
DVK(KP555570233, +0.555570233019602224742830813948532874374937191); DVK(KP831469612, +0.831469612302545237078788377617905756738560812); DVK(KP923879532, +0.923879532511286756128183189396788286822416626); DVK(KP382683432, +0.382683432365089771728459984030398866761344562); DVK(KP707106781, +0.707106781186547524400844362104849039284835938); INT i; const R *xi; R *xo; xi = ii; xo = io; 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, T5u, Tb, T3A, T2q, T3v, T6G, T78, Tq, T3w, T6B, T79, T2l, T3B, T4w; V T5r, TI, T2g, T6u, T74, T3q, T3D, T4E, T5o, TZ, T2h, T6x, T75, T3t, T3E; V T4L, T5p, T23, T2N, T6m, T70, T6p, T71, T2c, T2O, T3i, T3Y, T5f, T5R, T5k; V T5S, T3l, T3Z, T1s, T2K, T6f, T6X, T6i, T6Y, T1B, T2L, T3b, T3V, T4Y, T5O; V T53, T5P, T3e, T3W; { V T3, T4n, T2p, T4o, T6, T5s, T9, T5t; { V T1, T2, T2n, T2o; T1 = LD(&(xi[0]), ivs, &(xi[0])); T2 = LD(&(xi[WS(is, 32)]), ivs, &(xi[0])); T3 = VSUB(T1, T2); T4n = VADD(T1, T2); T2n = LD(&(xi[WS(is, 16)]), ivs, &(xi[0])); T2o = LD(&(xi[WS(is, 48)]), ivs, &(xi[0])); T2p = VSUB(T2n, T2o); T4o = VADD(T2n, T2o); } { 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); T5s = 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); T5t = VADD(T7, T8); } T4p = VSUB(T4n, T4o); T5u = VSUB(T5s, T5t); { V Ta, T2m, T6E, T6F; Ta = VMUL(LDK(KP707106781), VADD(T6, T9)); Tb = VSUB(T3, Ta); T3A = VADD(T3, Ta); T2m = VMUL(LDK(KP707106781), VSUB(T6, T9)); T2q = VSUB(T2m, T2p); T3v = VADD(T2p, T2m); T6E = VADD(T4n, T4o); T6F = VADD(T5s, T5t); T6G = VSUB(T6E, T6F); T78 = VADD(T6E, T6F); } } { V Te, T4q, To, T4t, Th, T4r, Tl, T4u; { 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, 60)]), ivs, &(xi[0])); Tn = LD(&(xi[WS(is, 28)]), ivs, &(xi[0])); To = VSUB(Tm, Tn); T4t = 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, 12)]), ivs, &(xi[0])); Tk = LD(&(xi[WS(is, 44)]), ivs, &(xi[0])); Tl = VSUB(Tj, Tk); T4u = VADD(Tj, Tk); } { V Ti, Tp, T6z, T6A; Ti = VFMA(LDK(KP382683432), Te, VMUL(LDK(KP923879532), Th)); Tp = VFNMS(LDK(KP382683432), To, VMUL(LDK(KP923879532), Tl)); Tq = VSUB(Ti, Tp); T3w = VADD(Ti, Tp); T6z = VADD(T4q, T4r); T6A = VADD(T4t, T4u); T6B = VSUB(T6z, T6A); T79 = VADD(T6z, T6A); } { V T2j, T2k, T4s, T4v; T2j = VFNMS(LDK(KP382683432), Th, VMUL(LDK(KP923879532), Te)); T2k = VFMA(LDK(KP923879532), To, VMUL(LDK(KP382683432), Tl)); T2l = VSUB(T2j, T2k); T3B = VADD(T2j, T2k); T4s = VSUB(T4q, T4r); T4v = VSUB(T4t, T4u); T4w = VMUL(LDK(KP707106781), VADD(T4s, T4v)); T5r = VMUL(LDK(KP707106781), VSUB(T4s, T4v)); } } { V TB, T4z, TF, T4y, Ty, T4C, TG, T4B; { V Tz, TA, TD, TE; Tz = LD(&(xi[WS(is, 18)]), ivs, &(xi[0])); TA = LD(&(xi[WS(is, 50)]), ivs, &(xi[0])); TB = VSUB(Tz, TA); T4z = VADD(Tz, TA); TD = LD(&(xi[WS(is, 2)]), ivs, &(xi[0])); TE = LD(&(xi[WS(is, 34)]), ivs, &(xi[0])); TF = VSUB(TD, TE); T4y = VADD(TD, TE); { V Ts, Tt, Tu, Tv, Tw, Tx; Ts = LD(&(xi[WS(is, 10)]), ivs, &(xi[0])); Tt = LD(&(xi[WS(is, 42)]), ivs, &(xi[0])); Tu = VSUB(Ts, Tt); Tv = LD(&(xi[WS(is, 58)]), ivs, &(xi[0])); Tw = LD(&(xi[WS(is, 26)]), ivs, &(xi[0])); Tx = VSUB(Tv, Tw); Ty = VMUL(LDK(KP707106781), VSUB(Tu, Tx)); T4C = VADD(Tv, Tw); TG = VMUL(LDK(KP707106781), VADD(Tu, Tx)); T4B = VADD(Ts, Tt); } } { V TC, TH, T6s, T6t; TC = VSUB(Ty, TB); TH = VSUB(TF, TG); TI = VFMA(LDK(KP831469612), TC, VMUL(LDK(KP555570233), TH)); T2g = VFNMS(LDK(KP555570233), TC, VMUL(LDK(KP831469612), TH)); T6s = VADD(T4y, T4z); T6t = VADD(T4B, T4C); T6u = VSUB(T6s, T6t); T74 = VADD(T6s, T6t); } { V T3o, T3p, T4A, T4D; T3o = VADD(TB, Ty); T3p = VADD(TF, TG); T3q = VFMA(LDK(KP980785280), T3o, VMUL(LDK(KP195090322), T3p)); T3D = VFNMS(LDK(KP195090322), T3o, VMUL(LDK(KP980785280), T3p)); T4A = VSUB(T4y, T4z); T4D = VSUB(T4B, T4C); T4E = VFMA(LDK(KP382683432), T4A, VMUL(LDK(KP923879532), T4D)); T5o = VFNMS(LDK(KP382683432), T4D, VMUL(LDK(KP923879532), T4A)); } } { V TS, T4J, TW, T4I, TP, T4G, TX, T4F; { V TQ, TR, TU, TV; TQ = LD(&(xi[WS(is, 14)]), ivs, &(xi[0])); TR = LD(&(xi[WS(is, 46)]), ivs, &(xi[0])); TS = VSUB(TQ, TR); T4J = VADD(TQ, TR); TU = LD(&(xi[WS(is, 62)]), ivs, &(xi[0])); TV = LD(&(xi[WS(is, 30)]), ivs, &(xi[0])); TW = VSUB(TU, TV); T4I = VADD(TU, TV); { V TJ, TK, TL, TM, TN, TO; TJ = LD(&(xi[WS(is, 6)]), ivs, &(xi[0])); TK = LD(&(xi[WS(is, 38)]), ivs, &(xi[0])); TL = VSUB(TJ, TK); TM = LD(&(xi[WS(is, 54)]), ivs, &(xi[0])); TN = LD(&(xi[WS(is, 22)]), ivs, &(xi[0])); TO = VSUB(TM, TN); TP = VMUL(LDK(KP707106781), VSUB(TL, TO)); T4G = VADD(TM, TN); TX = VMUL(LDK(KP707106781), VADD(TL, TO)); T4F = VADD(TJ, TK); } } { V TT, TY, T6v, T6w; TT = VSUB(TP, TS); TY = VSUB(TW, TX); TZ = VFNMS(LDK(KP555570233), TY, VMUL(LDK(KP831469612), TT)); T2h = VFMA(LDK(KP555570233), TT, VMUL(LDK(KP831469612), TY)); T6v = VADD(T4I, T4J); T6w = VADD(T4F, T4G); T6x = VSUB(T6v, T6w); T75 = VADD(T6v, T6w); } { V T3r, T3s, T4H, T4K; T3r = VADD(TS, TP); T3s = VADD(TW, TX); T3t = VFNMS(LDK(KP195090322), T3s, VMUL(LDK(KP980785280), T3r)); T3E = VFMA(LDK(KP195090322), T3r, VMUL(LDK(KP980785280), T3s)); T4H = VSUB(T4F, T4G); T4K = VSUB(T4I, T4J); T4L = VFNMS(LDK(KP382683432), T4K, VMUL(LDK(KP923879532), T4H)); T5p = VFMA(LDK(KP923879532), T4K, VMUL(LDK(KP382683432), T4H)); } } { V T21, T5h, T26, T5g, T1Y, T5d, T27, T5c, T55, T56, T1J, T57, T29, T58, T59; V T1Q, T5a, T2a; { V T1Z, T20, T24, T25; T1Z = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)])); T20 = LD(&(xi[WS(is, 47)]), ivs, &(xi[WS(is, 1)])); T21 = VSUB(T1Z, T20); T5h = VADD(T1Z, T20); T24 = LD(&(xi[WS(is, 63)]), ivs, &(xi[WS(is, 1)])); T25 = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)])); T26 = VSUB(T24, T25); T5g = VADD(T24, T25); } { V T1S, T1T, T1U, T1V, T1W, T1X; T1S = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)])); T1T = LD(&(xi[WS(is, 39)]), ivs, &(xi[WS(is, 1)])); T1U = VSUB(T1S, T1T); T1V = LD(&(xi[WS(is, 55)]), ivs, &(xi[WS(is, 1)])); T1W = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)])); T1X = VSUB(T1V, T1W); T1Y = VMUL(LDK(KP707106781), VSUB(T1U, T1X)); T5d = VADD(T1V, T1W); T27 = VMUL(LDK(KP707106781), VADD(T1U, T1X)); T5c = VADD(T1S, T1T); } { V T1F, T1I, T1M, T1P; { V T1D, T1E, T1G, T1H; T1D = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)])); T1E = LD(&(xi[WS(is, 35)]), ivs, &(xi[WS(is, 1)])); T1F = VSUB(T1D, T1E); T55 = VADD(T1D, T1E); T1G = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)])); T1H = LD(&(xi[WS(is, 51)]), ivs, &(xi[WS(is, 1)])); T1I = VSUB(T1G, T1H); T56 = VADD(T1G, T1H); } T1J = VFNMS(LDK(KP382683432), T1I, VMUL(LDK(KP923879532), T1F)); T57 = VSUB(T55, T56); T29 = VFMA(LDK(KP382683432), T1F, VMUL(LDK(KP923879532), T1I)); { V T1K, T1L, T1N, T1O; T1K = LD(&(xi[WS(is, 59)]), ivs, &(xi[WS(is, 1)])); T1L = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)])); T1M = VSUB(T1K, T1L); T58 = VADD(T1K, T1L); T1N = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)])); T1O = LD(&(xi[WS(is, 43)]), ivs, &(xi[WS(is, 1)])); T1P = VSUB(T1N, T1O); T59 = VADD(T1N, T1O); } T1Q = VFMA(LDK(KP923879532), T1M, VMUL(LDK(KP382683432), T1P)); T5a = VSUB(T58, T59); T2a = VFNMS(LDK(KP382683432), T1M, VMUL(LDK(KP923879532), T1P)); } { V T1R, T22, T6k, T6l; T1R = VSUB(T1J, T1Q); T22 = VSUB(T1Y, T21); T23 = VSUB(T1R, T22); T2N = VADD(T22, T1R); T6k = VADD(T5g, T5h); T6l = VADD(T5c, T5d); T6m = VSUB(T6k, T6l); T70 = VADD(T6k, T6l); } { V T6n, T6o, T28, T2b; T6n = VADD(T55, T56); T6o = VADD(T58, T59); T6p = VSUB(T6n, T6o); T71 = VADD(T6n, T6o); T28 = VSUB(T26, T27); T2b = VSUB(T29, T2a); T2c = VSUB(T28, T2b); T2O = VADD(T28, T2b); } { V T3g, T3h, T5b, T5e; T3g = VADD(T26, T27); T3h = VADD(T1J, T1Q); T3i = VADD(T3g, T3h); T3Y = VSUB(T3g, T3h); T5b = VMUL(LDK(KP707106781), VSUB(T57, T5a)); T5e = VSUB(T5c, T5d); T5f = VSUB(T5b, T5e); T5R = VADD(T5e, T5b); } { V T5i, T5j, T3j, T3k; T5i = VSUB(T5g, T5h); T5j = VMUL(LDK(KP707106781), VADD(T57, T5a)); T5k = VSUB(T5i, T5j); T5S = VADD(T5i, T5j); T3j = VADD(T21, T1Y); T3k = VADD(T29, T2a); T3l = VADD(T3j, T3k); T3Z = VSUB(T3k, T3j); } } { V T1q, T50, T1v, T4Z, T1n, T4W, T1w, T4V, T4O, T4P, T18, T4Q, T1y, T4R, T4S; V T1f, T4T, T1z; { 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); T50 = 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); T4Z = VADD(T1t, T1u); } { V T1h, T1i, T1j, T1k, T1l, T1m; T1h = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)])); T1i = LD(&(xi[WS(is, 41)]), ivs, &(xi[WS(is, 1)])); T1j = VSUB(T1h, T1i); T1k = LD(&(xi[WS(is, 57)]), ivs, &(xi[WS(is, 1)])); T1l = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)])); T1m = VSUB(T1k, T1l); T1n = VMUL(LDK(KP707106781), VSUB(T1j, T1m)); T4W = VADD(T1k, T1l); T1w = VMUL(LDK(KP707106781), VADD(T1j, T1m)); T4V = VADD(T1h, T1i); } { V T14, T17, T1b, T1e; { V T12, T13, T15, T16; T12 = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)])); T13 = LD(&(xi[WS(is, 37)]), ivs, &(xi[WS(is, 1)])); T14 = VSUB(T12, T13); T4O = VADD(T12, T13); T15 = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)])); T16 = LD(&(xi[WS(is, 53)]), ivs, &(xi[WS(is, 1)])); T17 = VSUB(T15, T16); T4P = VADD(T15, T16); } T18 = VFNMS(LDK(KP382683432), T17, VMUL(LDK(KP923879532), T14)); T4Q = VSUB(T4O, T4P); T1y = VFMA(LDK(KP382683432), T14, VMUL(LDK(KP923879532), T17)); { V T19, T1a, T1c, T1d; T19 = LD(&(xi[WS(is, 61)]), ivs, &(xi[WS(is, 1)])); T1a = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)])); T1b = VSUB(T19, T1a); T4R = VADD(T19, T1a); T1c = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)])); T1d = LD(&(xi[WS(is, 45)]), ivs, &(xi[WS(is, 1)])); T1e = VSUB(T1c, T1d); T4S = VADD(T1c, T1d); } T1f = VFMA(LDK(KP923879532), T1b, VMUL(LDK(KP382683432), T1e)); T4T = VSUB(T4R, T4S); T1z = VFNMS(LDK(KP382683432), T1b, VMUL(LDK(KP923879532), T1e)); } { V T1g, T1r, T6d, T6e; T1g = VSUB(T18, T1f); T1r = VSUB(T1n, T1q); T1s = VSUB(T1g, T1r); T2K = VADD(T1r, T1g); T6d = VADD(T4Z, T50); T6e = VADD(T4V, T4W); T6f = VSUB(T6d, T6e); T6X = VADD(T6d, T6e); } { V T6g, T6h, T1x, T1A; T6g = VADD(T4O, T4P); T6h = VADD(T4R, T4S); T6i = VSUB(T6g, T6h); T6Y = VADD(T6g, T6h); T1x = VSUB(T1v, T1w); T1A = VSUB(T1y, T1z); T1B = VSUB(T1x, T1A); T2L = VADD(T1x, T1A); } { V T39, T3a, T4U, T4X; T39 = VADD(T1v, T1w); T3a = VADD(T18, T1f); T3b = VADD(T39, T3a); T3V = VSUB(T39, T3a); T4U = VMUL(LDK(KP707106781), VSUB(T4Q, T4T));
⌨️ 快捷键说明
复制代码
Ctrl + C
搜索代码
Ctrl + F
全屏模式
F11
切换主题
Ctrl + Shift + D
显示快捷键
?
增大字号
Ctrl + =
减小字号
Ctrl + -