⭐ 欢迎来到虫虫下载站! | 📦 资源下载 📁 资源专辑 ℹ️ 关于我们
⭐ 虫虫下载站

📄 n2fv_32.c

📁 快速fft变换
💻 C
📖 第 1 页 / 共 2 页
字号:
 */#include "n2f.h"static void n2fv_32(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs){     DVK(KP195090322, +0.195090322016128267848284868477022240927691618);     DVK(KP980785280, +0.980785280403230449126182236134239036973933731);     DVK(KP555570233, +0.555570233019602224742830813948532874374937191);     DVK(KP831469612, +0.831469612302545237078788377617905756738560812);     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 T1T, T1W, T2K, T2x, T16, T1A, Tb, T1p, TT, T1v, TY, T1w, T27, T2a, T2b;	  V T2H, T2O, TC, T1s, TH, T1t, T20, T23, T24, T2E, T2N, T2g, T2j, Tq, T1B;	  V T19, T1q, T2A, T2L;	  {	       V T3, T1R, T15, T1S, T6, T1U, T9, T1V, T12, Ta;	       {		    V T1, T2, T13, T14;		    T1 = LD(&(xi[0]), ivs, &(xi[0]));		    T2 = LD(&(xi[WS(is, 16)]), ivs, &(xi[0]));		    T3 = VSUB(T1, T2);		    T1R = VADD(T1, T2);		    T13 = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));		    T14 = LD(&(xi[WS(is, 24)]), ivs, &(xi[0]));		    T15 = VSUB(T13, T14);		    T1S = VADD(T13, T14);	       }	       {		    V T4, T5, T7, T8;		    T4 = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));		    T5 = LD(&(xi[WS(is, 20)]), ivs, &(xi[0]));		    T6 = VSUB(T4, T5);		    T1U = VADD(T4, T5);		    T7 = LD(&(xi[WS(is, 28)]), ivs, &(xi[0]));		    T8 = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));		    T9 = VSUB(T7, T8);		    T1V = VADD(T7, T8);	       }	       T1T = VADD(T1R, T1S);	       T1W = VADD(T1U, T1V);	       T2K = VSUB(T1V, T1U);	       T2x = VSUB(T1R, T1S);	       T12 = VMUL(LDK(KP707106781), VSUB(T9, T6));	       T16 = VSUB(T12, T15);	       T1A = VADD(T15, T12);	       Ta = VMUL(LDK(KP707106781), VADD(T6, T9));	       Tb = VADD(T3, Ta);	       T1p = VSUB(T3, Ta);	  }	  {	       V TL, T25, TX, T26, TO, T28, TR, T29;	       {		    V TJ, TK, TV, TW;		    TJ = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)]));		    TK = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));		    TL = VSUB(TJ, TK);		    T25 = VADD(TJ, TK);		    TV = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));		    TW = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)]));		    TX = VSUB(TV, TW);		    T26 = VADD(TV, TW);	       }	       {		    V TM, TN, TP, TQ;		    TM = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));		    TN = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));		    TO = VSUB(TM, TN);		    T28 = VADD(TM, TN);		    TP = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)]));		    TQ = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));		    TR = VSUB(TP, TQ);		    T29 = VADD(TP, TQ);	       }	       {		    V TS, TU, T2F, T2G;		    TS = VMUL(LDK(KP707106781), VADD(TO, TR));		    TT = VADD(TL, TS);		    T1v = VSUB(TL, TS);		    TU = VMUL(LDK(KP707106781), VSUB(TR, TO));		    TY = VSUB(TU, TX);		    T1w = VADD(TX, TU);		    T27 = VADD(T25, T26);		    T2a = VADD(T28, T29);		    T2b = VSUB(T27, T2a);		    T2F = VSUB(T25, T26);		    T2G = VSUB(T29, T28);		    T2H = VFNMS(LDK(KP382683432), T2G, VMUL(LDK(KP923879532), T2F));		    T2O = VFMA(LDK(KP382683432), T2F, VMUL(LDK(KP923879532), T2G));	       }	  }	  {	       V Tu, T1Y, TG, T1Z, Tx, T21, TA, T22;	       {		    V Ts, Tt, TE, TF;		    Ts = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));		    Tt = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));		    Tu = VSUB(Ts, Tt);		    T1Y = VADD(Ts, Tt);		    TE = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));		    TF = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)]));		    TG = VSUB(TE, TF);		    T1Z = VADD(TE, TF);	       }	       {		    V Tv, Tw, Ty, Tz;		    Tv = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));		    Tw = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)]));		    Tx = VSUB(Tv, Tw);		    T21 = VADD(Tv, Tw);		    Ty = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)]));		    Tz = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));		    TA = VSUB(Ty, Tz);		    T22 = VADD(Ty, Tz);	       }	       {		    V TB, TD, T2C, T2D;		    TB = VMUL(LDK(KP707106781), VADD(Tx, TA));		    TC = VADD(Tu, TB);		    T1s = VSUB(Tu, TB);		    TD = VMUL(LDK(KP707106781), VSUB(TA, Tx));		    TH = VSUB(TD, TG);		    T1t = VADD(TG, TD);		    T20 = VADD(T1Y, T1Z);		    T23 = VADD(T21, T22);		    T24 = VSUB(T20, T23);		    T2C = VSUB(T1Y, T1Z);		    T2D = VSUB(T22, T21);		    T2E = VFMA(LDK(KP923879532), T2C, VMUL(LDK(KP382683432), T2D));		    T2N = VFNMS(LDK(KP382683432), T2C, VMUL(LDK(KP923879532), T2D));	       }	  }	  {	       V Te, T2h, To, T2f, Th, T2i, Tl, T2e, Ti, Tp;	       {		    V Tc, Td, Tm, Tn;		    Tc = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));		    Td = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));		    Te = VSUB(Tc, Td);		    T2h = VADD(Tc, Td);		    Tm = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));		    Tn = LD(&(xi[WS(is, 22)]), ivs, &(xi[0]));		    To = VSUB(Tm, Tn);		    T2f = VADD(Tm, Tn);	       }	       {		    V Tf, Tg, Tj, Tk;		    Tf = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));		    Tg = LD(&(xi[WS(is, 26)]), ivs, &(xi[0]));		    Th = VSUB(Tf, Tg);		    T2i = VADD(Tf, Tg);		    Tj = LD(&(xi[WS(is, 30)]), ivs, &(xi[0]));		    Tk = LD(&(xi[WS(is, 14)]), ivs, &(xi[0]));		    Tl = VSUB(Tj, Tk);		    T2e = VADD(Tj, Tk);	       }	       T2g = VADD(T2e, T2f);	       T2j = VADD(T2h, T2i);	       Ti = VFNMS(LDK(KP382683432), Th, VMUL(LDK(KP923879532), Te));	       Tp = VFMA(LDK(KP923879532), Tl, VMUL(LDK(KP382683432), To));	       Tq = VADD(Ti, Tp);	       T1B = VSUB(Tp, Ti);	       {		    V T17, T18, T2y, T2z;		    T17 = VFNMS(LDK(KP923879532), To, VMUL(LDK(KP382683432), Tl));		    T18 = VFMA(LDK(KP382683432), Te, VMUL(LDK(KP923879532), Th));		    T19 = VSUB(T17, T18);		    T1q = VADD(T18, T17);		    T2y = VSUB(T2h, T2i);		    T2z = VSUB(T2e, T2f);		    T2A = VMUL(LDK(KP707106781), VADD(T2y, T2z));		    T2L = VMUL(LDK(KP707106781), VSUB(T2z, T2y));	       }	  }	  {	       V T31, T32, T33, T34, T35, T36, T37, T38, T39, T3a, T3b, T3c;	       {		    V T2d, T2n, T2m, T2o;		    {			 V T1X, T2c, T2k, T2l;			 T1X = VSUB(T1T, T1W);			 T2c = VMUL(LDK(KP707106781), VADD(T24, T2b));			 T2d = VADD(T1X, T2c);			 T2n = VSUB(T1X, T2c);			 T2k = VSUB(T2g, T2j);			 T2l = VMUL(LDK(KP707106781), VSUB(T2b, T24));			 T2m = VBYI(VADD(T2k, T2l));			 T2o = VBYI(VSUB(T2l, T2k));		    }		    T31 = VSUB(T2d, T2m);		    STM2(&(xo[56]), T31, ovs, &(xo[0]));		    T32 = VADD(T2n, T2o);		    STM2(&(xo[24]), T32, ovs, &(xo[0]));		    T33 = VADD(T2d, T2m);		    STM2(&(xo[8]), T33, ovs, &(xo[0]));		    T34 = VSUB(T2n, T2o);		    STM2(&(xo[40]), T34, ovs, &(xo[0]));	       }	       {		    V T2r, T2v, T2u, T2w;		    {			 V T2p, T2q, T2s, T2t;			 T2p = VADD(T1T, T1W);			 T2q = VADD(T2j, T2g);			 T2r = VADD(T2p, T2q);			 T2v = VSUB(T2p, T2q);			 T2s = VADD(T20, T23);			 T2t = VADD(T27, T2a);			 T2u = VADD(T2s, T2t);			 T2w = VBYI(VSUB(T2t, T2s));		    }		    T35 = VSUB(T2r, T2u);		    STM2(&(xo[32]), T35, ovs, &(xo[0]));		    T36 = VADD(T2v, T2w);		    STM2(&(xo[16]), T36, ovs, &(xo[0]));		    T37 = VADD(T2r, T2u);		    STM2(&(xo[0]), T37, ovs, &(xo[0]));		    T38 = VSUB(T2v, T2w);		    STM2(&(xo[48]), T38, ovs, &(xo[0]));	       }	       {		    V T2V, T2Z, T2Y, T30;		    {			 V T2T, T2U, T2W, T2X;			 T2T = VSUB(T2H, T2E);			 T2U = VSUB(T2L, T2K);			 T2V = VBYI(VSUB(T2T, T2U));			 T2Z = VBYI(VADD(T2U, T2T));			 T2W = VSUB(T2x, T2A);			 T2X = VSUB(T2O, T2N);			 T2Y = VSUB(T2W, T2X);			 T30 = VADD(T2W, T2X);		    }		    T39 = VADD(T2V, T2Y);		    STM2(&(xo[20]), T39, ovs, &(xo[0]));		    T3a = VSUB(T30, T2Z);		    STM2(&(xo[52]), T3a, ovs, &(xo[0]));		    T3b = VSUB(T2Y, T2V);		    STM2(&(xo[44]), T3b, ovs, &(xo[0]));		    T3c = VADD(T2Z, T30);		    STM2(&(xo[12]), T3c, ovs, &(xo[0]));	       }	       {		    V T3d, T3e, T3f, T3g;		    {			 V T2J, T2R, T2Q, T2S;			 {			      V T2B, T2I, T2M, T2P;			      T2B = VADD(T2x, T2A);			      T2I = VADD(T2E, T2H);			      T2J = VADD(T2B, T2I);			      T2R = VSUB(T2B, T2I);			      T2M = VADD(T2K, T2L);			      T2P = VADD(T2N, T2O);			      T2Q = VBYI(VADD(T2M, T2P));			      T2S = VBYI(VSUB(T2P, T2M));			 }			 T3d = VSUB(T2J, T2Q);			 STM2(&(xo[60]), T3d, ovs, &(xo[0]));			 T3e = VADD(T2R, T2S);			 STM2(&(xo[28]), T3e, ovs, &(xo[0]));			 T3f = VADD(T2J, T2Q);			 STM2(&(xo[4]), T3f, ovs, &(xo[0]));			 T3g = VSUB(T2R, T2S);			 STM2(&(xo[36]), T3g, ovs, &(xo[0]));		    }		    {			 V T1r, T1C, T1M, T1K, T1F, T1N, T1y, T1J;			 T1r = VADD(T1p, T1q);			 T1C = VADD(T1A, T1B);			 T1M = VSUB(T1p, T1q);			 T1K = VSUB(T1B, T1A);			 {			      V T1D, T1E, T1u, T1x;			      T1D = VFNMS(LDK(KP555570233), T1s, VMUL(LDK(KP831469612), T1t));			      T1E = VFMA(LDK(KP555570233), T1v, VMUL(LDK(KP831469612), T1w));			      T1F = VADD(T1D, T1E);			      T1N = VSUB(T1E, T1D);			      T1u = VFMA(LDK(KP831469612), T1s, VMUL(LDK(KP555570233), T1t));			      T1x = VFNMS(LDK(KP555570233), T1w, VMUL(LDK(KP831469612), T1v));			      T1y = VADD(T1u, T1x);			      T1J = VSUB(T1x, T1u);			 }			 {			      V T1z, T1G, T3h, T3i;			      T1z = VADD(T1r, T1y);			      T1G = VBYI(VADD(T1C, T1F));			      T3h = VSUB(T1z, T1G);			      STM2(&(xo[58]), T3h, ovs, &(xo[2]));			      STN2(&(xo[56]), T31, T3h, ovs);			      T3i = VADD(T1z, T1G);			      STM2(&(xo[6]), T3i, ovs, &(xo[2]));			      STN2(&(xo[4]), T3f, T3i, ovs);			 }			 {			      V T1P, T1Q, T3j, T3k;			      T1P = VBYI(VADD(T1K, T1J));			      T1Q = VADD(T1M, T1N);			      T3j = VADD(T1P, T1Q);			      STM2(&(xo[10]), T3j, ovs, &(xo[2]));			      STN2(&(xo[8]), T33, T3j, ovs);			      T3k = VSUB(T1Q, T1P);			      STM2(&(xo[54]), T3k, ovs, &(xo[2]));			      STN2(&(xo[52]), T3a, T3k, ovs);			 }			 {			      V T1H, T1I, T3l, T3m;			      T1H = VSUB(T1r, T1y);			      T1I = VBYI(VSUB(T1F, T1C));			      T3l = VSUB(T1H, T1I);			      STM2(&(xo[38]), T3l, ovs, &(xo[2]));			      STN2(&(xo[36]), T3g, T3l, ovs);			      T3m = VADD(T1H, T1I);			      STM2(&(xo[26]), T3m, ovs, &(xo[2]));			      STN2(&(xo[24]), T32, T3m, ovs);			 }			 {			      V T1L, T1O, T3n, T3o;			      T1L = VBYI(VSUB(T1J, T1K));			      T1O = VSUB(T1M, T1N);			      T3n = VADD(T1L, T1O);			      STM2(&(xo[22]), T3n, ovs, &(xo[2]));			      STN2(&(xo[20]), T39, T3n, ovs);			      T3o = VSUB(T1O, T1L);			      STM2(&(xo[42]), T3o, ovs, &(xo[2]));			      STN2(&(xo[40]), T34, T3o, ovs);			 }		    }		    {			 V Tr, T1a, T1k, T1i, T1d, T1l, T10, T1h;			 Tr = VADD(Tb, Tq);			 T1a = VADD(T16, T19);			 T1k = VSUB(Tb, Tq);			 T1i = VSUB(T19, T16);			 {			      V T1b, T1c, TI, TZ;			      T1b = VFNMS(LDK(KP195090322), TC, VMUL(LDK(KP980785280), TH));			      T1c = VFMA(LDK(KP195090322), TT, VMUL(LDK(KP980785280), TY));			      T1d = VADD(T1b, T1c);			      T1l = VSUB(T1c, T1b);			      TI = VFMA(LDK(KP980785280), TC, VMUL(LDK(KP195090322), TH));			      TZ = VFNMS(LDK(KP195090322), TY, VMUL(LDK(KP980785280), TT));			      T10 = VADD(TI, TZ);			      T1h = VSUB(TZ, TI);			 }			 {			      V T11, T1e, T3p, T3q;			      T11 = VADD(Tr, T10);			      T1e = VBYI(VADD(T1a, T1d));			      T3p = VSUB(T11, T1e);			      STM2(&(xo[62]), T3p, ovs, &(xo[2]));			      STN2(&(xo[60]), T3d, T3p, ovs);			      T3q = VADD(T11, T1e);			      STM2(&(xo[2]), T3q, ovs, &(xo[2]));			      STN2(&(xo[0]), T37, T3q, ovs);			 }			 {			      V T1n, T1o, T3r, T3s;			      T1n = VBYI(VADD(T1i, T1h));			      T1o = VADD(T1k, T1l);			      T3r = VADD(T1n, T1o);			      STM2(&(xo[14]), T3r, ovs, &(xo[2]));			      STN2(&(xo[12]), T3c, T3r, ovs);			      T3s = VSUB(T1o, T1n);			      STM2(&(xo[50]), T3s, ovs, &(xo[2]));			      STN2(&(xo[48]), T38, T3s, ovs);			 }			 {			      V T1f, T1g, T3t, T3u;			      T1f = VSUB(Tr, T10);			      T1g = VBYI(VSUB(T1d, T1a));			      T3t = VSUB(T1f, T1g);			      STM2(&(xo[34]), T3t, ovs, &(xo[2]));			      STN2(&(xo[32]), T35, T3t, ovs);			      T3u = VADD(T1f, T1g);			      STM2(&(xo[30]), T3u, ovs, &(xo[2]));			      STN2(&(xo[28]), T3e, T3u, ovs);			 }			 {			      V T1j, T1m, T3v, T3w;			      T1j = VBYI(VSUB(T1h, T1i));			      T1m = VSUB(T1k, T1l);			      T3v = VADD(T1j, T1m);			      STM2(&(xo[18]), T3v, ovs, &(xo[2]));			      STN2(&(xo[16]), T36, T3v, ovs);			      T3w = VSUB(T1m, T1j);			      STM2(&(xo[46]), T3w, ovs, &(xo[2]));			      STN2(&(xo[44]), T3b, T3w, ovs);			 }		    }	       }	  }     }}static const kdft_desc desc = { 32, "n2fv_32", {170, 26, 16, 0}, &GENUS, 0, 2, 0, 0 };void X(codelet_n2fv_32) (planner *p) {     X(kdft_register) (p, n2fv_32, &desc);}#endif				/* HAVE_FMA */

⌨️ 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 + -