⭐ 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

📄 t2_20.c

📁 快速fft变换
💻 C
📖 第 1 页 / 共 2 页
字号:
				   T3Y = FMA(KP559016994, T3X, T3W);				   ri[WS(rs, 9)] = FNMS(KP951056516, T45, T3Y);				   ri[WS(rs, 1)] = FMA(KP951056516, T45, T3Y);				   ri[WS(rs, 17)] = FNMS(KP951056516, T47, T46);				   ri[WS(rs, 13)] = FMA(KP951056516, T47, T46);				   T5e = T5a - T5b;				   T5c = T5a + T5b;			      }			 }		    }	       }	  }	  ii[WS(rs, 5)] = T5c + T59;	  T5d = FNMS(KP250000000, T5c, T59);	  T5j = FNMS(KP559016994, T5e, T5d);	  T5f = FMA(KP559016994, T5e, T5d);	  ii[WS(rs, 9)] = FMA(KP951056516, T5i, T5f);	  ii[WS(rs, 1)] = FNMS(KP951056516, T5i, T5f);	  ii[WS(rs, 17)] = FMA(KP951056516, T5k, T5j);	  ii[WS(rs, 13)] = FNMS(KP951056516, T5k, T5j);     }}static const tw_instr twinstr[] = {     {TW_CEXP, 0, 1},     {TW_CEXP, 0, 3},     {TW_CEXP, 0, 9},     {TW_CEXP, 0, 19},     {TW_NEXT, 1, 0}};static const ct_desc desc = { 20, "t2_20", twinstr, &GENUS, {136, 58, 140, 0}, 0, 0, 0 };void X(codelet_t2_20) (planner *p) {     X(kdft_dit_register) (p, t2_20, &desc);}#else				/* HAVE_FMA *//* Generated by: ../../../genfft/gen_twiddle -compact -variables 4 -pipeline-latency 4 -twiddle-log3 -precompute-twiddles -n 20 -name t2_20 -include t.h *//* * This function contains 276 FP additions, 164 FP multiplications, * (or, 204 additions, 92 multiplications, 72 fused multiply/add), * 123 stack variables, 4 constants, and 80 memory accesses */#include "t.h"static void t2_20(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms){     DK(KP587785252, +0.587785252292473129168705954639072768597652438);     DK(KP951056516, +0.951056516295153572116439333379382143405698634);     DK(KP250000000, +0.250000000000000000000000000000000000000000000);     DK(KP559016994, +0.559016994374947424102293417182819058860154590);     INT m;     for (m = mb, W = W + (mb * 8); m < me; m = m + 1, ri = ri + ms, ii = ii + ms, W = W + 8, MAKE_VOLATILE_STRIDE(rs)) {	  E T2, T5, Tg, Ti, Tk, To, T1h, T1f, T6, T3, T8, T14, T1Q, Tc, T1O;	  E T1v, T18, T1t, T1n, T24, T1j, T22, Tq, Tu, T1E, T1G, Tx, Ty, Tz, TJ;	  E T1Z, TB, T1X, T1A, TZ, TL, T1y, TX;	  {	       E T7, T16, Ta, T13, T4, T17, Tb, T12;	       {		    E Th, Tn, Tj, Tm;		    T2 = W[0];		    T5 = W[1];		    Tg = W[2];		    Ti = W[3];		    Th = T2 * Tg;		    Tn = T5 * Tg;		    Tj = T5 * Ti;		    Tm = T2 * Ti;		    Tk = Th - Tj;		    To = Tm + Tn;		    T1h = Tm - Tn;		    T1f = Th + Tj;		    T6 = W[5];		    T7 = T5 * T6;		    T16 = Tg * T6;		    Ta = T2 * T6;		    T13 = Ti * T6;		    T3 = W[4];		    T4 = T2 * T3;		    T17 = Ti * T3;		    Tb = T5 * T3;		    T12 = Tg * T3;	       }	       T8 = T4 - T7;	       T14 = T12 + T13;	       T1Q = T16 + T17;	       Tc = Ta + Tb;	       T1O = T12 - T13;	       T1v = Ta - Tb;	       T18 = T16 - T17;	       T1t = T4 + T7;	       {		    E T1l, T1m, T1g, T1i;		    T1l = T1f * T6;		    T1m = T1h * T3;		    T1n = T1l + T1m;		    T24 = T1l - T1m;		    T1g = T1f * T3;		    T1i = T1h * T6;		    T1j = T1g - T1i;		    T22 = T1g + T1i;		    {			 E Tl, Tp, Ts, Tt;			 Tl = Tk * T3;			 Tp = To * T6;			 Tq = Tl + Tp;			 Ts = Tk * T6;			 Tt = To * T3;			 Tu = Ts - Tt;			 T1E = Tl - Tp;			 T1G = Ts + Tt;			 Tx = W[6];			 Ty = W[7];			 Tz = FMA(Tk, Tx, To * Ty);			 TJ = FMA(Tq, Tx, Tu * Ty);			 T1Z = FNMS(T1h, Tx, T1f * Ty);			 TB = FNMS(To, Tx, Tk * Ty);			 T1X = FMA(T1f, Tx, T1h * Ty);			 T1A = FNMS(T5, Tx, T2 * Ty);			 TZ = FNMS(Ti, Tx, Tg * Ty);			 TL = FNMS(Tu, Tx, Tq * Ty);			 T1y = FMA(T2, Tx, T5 * Ty);			 TX = FMA(Tg, Tx, Ti * Ty);		    }	       }	  }	  {	       E TF, T2b, T4A, T4J, T2K, T3r, T4a, T4m, T1N, T28, T29, T3C, T3F, T4o, T3X;	       E T3Y, T44, T2f, T2g, T2h, T2n, T2s, T4L, T3g, T3h, T4w, T3n, T3o, T3p, T30;	       E T35, T36, TW, T1r, T1s, T3J, T3M, T4n, T3U, T3V, T43, T2c, T2d, T2e, T2y;	       E T2D, T4K, T3d, T3e, T4v, T3k, T3l, T3m, T2P, T2U, T2V;	       {		    E T1, T48, Te, T47, Tw, T2H, TD, T2I, T9, Td;		    T1 = ri[0];		    T48 = ii[0];		    T9 = ri[WS(rs, 10)];		    Td = ii[WS(rs, 10)];		    Te = FMA(T8, T9, Tc * Td);		    T47 = FNMS(Tc, T9, T8 * Td);		    {			 E Tr, Tv, TA, TC;			 Tr = ri[WS(rs, 5)];			 Tv = ii[WS(rs, 5)];			 Tw = FMA(Tq, Tr, Tu * Tv);			 T2H = FNMS(Tu, Tr, Tq * Tv);			 TA = ri[WS(rs, 15)];			 TC = ii[WS(rs, 15)];			 TD = FMA(Tz, TA, TB * TC);			 T2I = FNMS(TB, TA, Tz * TC);		    }		    {			 E Tf, TE, T4y, T4z;			 Tf = T1 + Te;			 TE = Tw + TD;			 TF = Tf - TE;			 T2b = Tf + TE;			 T4y = T48 - T47;			 T4z = Tw - TD;			 T4A = T4y - T4z;			 T4J = T4z + T4y;		    }		    {			 E T2G, T2J, T46, T49;			 T2G = T1 - Te;			 T2J = T2H - T2I;			 T2K = T2G - T2J;			 T3r = T2G + T2J;			 T46 = T2H + T2I;			 T49 = T47 + T48;			 T4a = T46 + T49;			 T4m = T49 - T46;		    }	       }	       {		    E T1D, T3A, T2l, T2W, T27, T3E, T2r, T34, T1M, T3B, T2m, T2Z, T1W, T3D, T2q;		    E T31;		    {			 E T1x, T2j, T1C, T2k;			 {			      E T1u, T1w, T1z, T1B;			      T1u = ri[WS(rs, 8)];			      T1w = ii[WS(rs, 8)];			      T1x = FMA(T1t, T1u, T1v * T1w);			      T2j = FNMS(T1v, T1u, T1t * T1w);			      T1z = ri[WS(rs, 18)];			      T1B = ii[WS(rs, 18)];			      T1C = FMA(T1y, T1z, T1A * T1B);			      T2k = FNMS(T1A, T1z, T1y * T1B);			 }			 T1D = T1x + T1C;			 T3A = T2j + T2k;			 T2l = T2j - T2k;			 T2W = T1x - T1C;		    }		    {			 E T21, T32, T26, T33;			 {			      E T1Y, T20, T23, T25;			      T1Y = ri[WS(rs, 17)];			      T20 = ii[WS(rs, 17)];			      T21 = FMA(T1X, T1Y, T1Z * T20);			      T32 = FNMS(T1Z, T1Y, T1X * T20);			      T23 = ri[WS(rs, 7)];			      T25 = ii[WS(rs, 7)];			      T26 = FMA(T22, T23, T24 * T25);			      T33 = FNMS(T24, T23, T22 * T25);			 }			 T27 = T21 + T26;			 T3E = T32 + T33;			 T2r = T21 - T26;			 T34 = T32 - T33;		    }		    {			 E T1I, T2X, T1L, T2Y;			 {			      E T1F, T1H, T1J, T1K;			      T1F = ri[WS(rs, 13)];			      T1H = ii[WS(rs, 13)];			      T1I = FMA(T1E, T1F, T1G * T1H);			      T2X = FNMS(T1G, T1F, T1E * T1H);			      T1J = ri[WS(rs, 3)];			      T1K = ii[WS(rs, 3)];			      T1L = FMA(Tg, T1J, Ti * T1K);			      T2Y = FNMS(Ti, T1J, Tg * T1K);			 }			 T1M = T1I + T1L;			 T3B = T2X + T2Y;			 T2m = T1I - T1L;			 T2Z = T2X - T2Y;		    }		    {			 E T1S, T2o, T1V, T2p;			 {			      E T1P, T1R, T1T, T1U;			      T1P = ri[WS(rs, 12)];			      T1R = ii[WS(rs, 12)];			      T1S = FMA(T1O, T1P, T1Q * T1R);			      T2o = FNMS(T1Q, T1P, T1O * T1R);			      T1T = ri[WS(rs, 2)];			      T1U = ii[WS(rs, 2)];			      T1V = FMA(T1f, T1T, T1h * T1U);			      T2p = FNMS(T1h, T1T, T1f * T1U);			 }			 T1W = T1S + T1V;			 T3D = T2o + T2p;			 T2q = T2o - T2p;			 T31 = T1S - T1V;		    }		    T1N = T1D - T1M;		    T28 = T1W - T27;		    T29 = T1N + T28;		    T3C = T3A - T3B;		    T3F = T3D - T3E;		    T4o = T3C + T3F;		    T3X = T3A + T3B;		    T3Y = T3D + T3E;		    T44 = T3X + T3Y;		    T2f = T1D + T1M;		    T2g = T1W + T27;		    T2h = T2f + T2g;		    T2n = T2l + T2m;		    T2s = T2q + T2r;		    T4L = T2n + T2s;		    T3g = T2l - T2m;		    T3h = T2q - T2r;		    T4w = T3g + T3h;		    T3n = T2W + T2Z;		    T3o = T31 + T34;		    T3p = T3n + T3o;		    T30 = T2W - T2Z;		    T35 = T31 - T34;		    T36 = T30 + T35;	       }	       {		    E TO, T3H, T2w, T2L, T1q, T3L, T2C, T2T, TV, T3I, T2x, T2O, T1b, T3K, T2B;		    E T2Q;		    {			 E TI, T2u, TN, T2v;			 {			      E TG, TH, TK, TM;			      TG = ri[WS(rs, 4)];			      TH = ii[WS(rs, 4)];			      TI = FMA(Tk, TG, To * TH);			      T2u = FNMS(To, TG, Tk * TH);			      TK = ri[WS(rs, 14)];			      TM = ii[WS(rs, 14)];			      TN = FMA(TJ, TK, TL * TM);			      T2v = FNMS(TL, TK, TJ * TM);			 }			 TO = TI + TN;			 T3H = T2u + T2v;			 T2w = T2u - T2v;			 T2L = TI - TN;		    }		    {			 E T1e, T2R, T1p, T2S;			 {			      E T1c, T1d, T1k, T1o;			      T1c = ri[WS(rs, 1)];			      T1d = ii[WS(rs, 1)];			      T1e = FMA(T2, T1c, T5 * T1d);			      T2R = FNMS(T5, T1c, T2 * T1d);			      T1k = ri[WS(rs, 11)];			      T1o = ii[WS(rs, 11)];			      T1p = FMA(T1j, T1k, T1n * T1o);			      T2S = FNMS(T1n, T1k, T1j * T1o);			 }			 T1q = T1e + T1p;			 T3L = T2R + T2S;			 T2C = T1e - T1p;			 T2T = T2R - T2S;		    }		    {			 E TR, T2M, TU, T2N;			 {			      E TP, TQ, TS, TT;			      TP = ri[WS(rs, 9)];			      TQ = ii[WS(rs, 9)];			      TR = FMA(T3, TP, T6 * TQ);			      T2M = FNMS(T6, TP, T3 * TQ);			      TS = ri[WS(rs, 19)];			      TT = ii[WS(rs, 19)];			      TU = FMA(Tx, TS, Ty * TT);			      T2N = FNMS(Ty, TS, Tx * TT);			 }			 TV = TR + TU;			 T3I = T2M + T2N;			 T2x = TR - TU;			 T2O = T2M - T2N;		    }		    {			 E T11, T2z, T1a, T2A;			 {			      E TY, T10, T15, T19;			      TY = ri[WS(rs, 16)];			      T10 = ii[WS(rs, 16)];			      T11 = FMA(TX, TY, TZ * T10);			      T2z = FNMS(TZ, TY, TX * T10);			      T15 = ri[WS(rs, 6)];			      T19 = ii[WS(rs, 6)];			      T1a = FMA(T14, T15, T18 * T19);			      T2A = FNMS(T18, T15, T14 * T19);			 }			 T1b = T11 + T1a;			 T3K = T2z + T2A;			 T2B = T2z - T2A;			 T2Q = T11 - T1a;		    }		    TW = TO - TV;		    T1r = T1b - T1q;		    T1s = TW + T1r;		    T3J = T3H - T3I;		    T3M = T3K - T3L;		    T4n = T3J + T3M;		    T3U = T3H + T3I;		    T3V = T3K + T3L;		    T43 = T3U + T3V;		    T2c = TO + TV;		    T2d = T1b + T1q;		    T2e = T2c + T2d;		    T2y = T2w + T2x;		    T2D = T2B + T2C;		    T4K = T2y + T2D;		    T3d = T2w - T2x;		    T3e = T2B - T2C;		    T4v = T3d + T3e;		    T3k = T2L + T2O;		    T3l = T2Q + T2T;		    T3m = T3k + T3l;		    T2P = T2L - T2O;		    T2U = T2Q - T2T;		    T2V = T2P + T2U;	       }	       {		    E T3y, T2a, T3x, T3O, T3Q, T3G, T3N, T3P, T3z;		    T3y = KP559016994 * (T1s - T29);		    T2a = T1s + T29;		    T3x = FNMS(KP250000000, T2a, TF);		    T3G = T3C - T3F;		    T3N = T3J - T3M;		    T3O = FNMS(KP587785252, T3N, KP951056516 * T3G);		    T3Q = FMA(KP951056516, T3N, KP587785252 * T3G);		    ri[WS(rs, 10)] = TF + T2a;		    T3P = T3y + T3x;		    ri[WS(rs, 14)] = T3P - T3Q;		    ri[WS(rs, 6)] = T3P + T3Q;		    T3z = T3x - T3y;		    ri[WS(rs, 2)] = T3z - T3O;		    ri[WS(rs, 18)] = T3z + T3O;	       }	       {		    E T4r, T4p, T4q, T4l, T4u, T4j, T4k, T4t, T4s;		    T4r = KP559016994 * (T4n - T4o);		    T4p = T4n + T4o;		    T4q = FNMS(KP250000000, T4p, T4m);		    T4j = T1N - T28;		    T4k = TW - T1r;		    T4l = FNMS(KP587785252, T4k, KP951056516 * T4j);		    T4u = FMA(KP951056516, T4k, KP587785252 * T4j);		    ii[WS(rs, 10)] = T4p + T4m;		    T4t = T4r + T4q;		    ii[WS(rs, 6)] = T4t - T4u;		    ii[WS(rs, 14)] = T4u + T4t;		    T4s = T4q - T4r;		    ii[WS(rs, 2)] = T4l + T4s;		    ii[WS(rs, 18)] = T4s - T4l;	       }	       {		    E T3R, T2i, T3S, T40, T42, T3W, T3Z, T41, T3T;		    T3R = KP559016994 * (T2e - T2h);		    T2i = T2e + T2h;		    T3S = FNMS(KP250000000, T2i, T2b);		    T3W = T3U - T3V;		    T3Z = T3X - T3Y;		    T40 = FMA(KP951056516, T3W, KP587785252 * T3Z);		    T42 = FNMS(KP587785252, T3W, KP951056516 * T3Z);		    ri[0] = T2b + T2i;		    T41 = T3S - T3R;		    ri[WS(rs, 12)] = T41 - T42;		    ri[WS(rs, 8)] = T41 + T42;		    T3T = T3R + T3S;		    ri[WS(rs, 4)] = T3T - T40;		    ri[WS(rs, 16)] = T3T + T40;	       }	       {		    E T4e, T45, T4f, T4d, T4i, T4b, T4c, T4h, T4g;		    T4e = KP559016994 * (T43 - T44);		    T45 = T43 + T44;		    T4f = FNMS(KP250000000, T45, T4a);		    T4b = T2c - T2d;		    T4c = T2f - T2g;		    T4d = FMA(KP951056516, T4b, KP587785252 * T4c);		    T4i = FNMS(KP587785252, T4b, KP951056516 * T4c);		    ii[0] = T45 + T4a;		    T4h = T4f - T4e;		    ii[WS(rs, 8)] = T4h - T4i;		    ii[WS(rs, 12)] = T4i + T4h;		    T4g = T4e + T4f;		    ii[WS(rs, 4)] = T4d + T4g;		    ii[WS(rs, 16)] = T4g - T4d;	       }	       {		    E T39, T37, T38, T2F, T3b, T2t, T2E, T3c, T3a;		    T39 = KP559016994 * (T2V - T36);		    T37 = T2V + T36;		    T38 = FNMS(KP250000000, T37, T2K);		    T2t = T2n - T2s;		    T2E = T2y - T2D;		    T2F = FNMS(KP587785252, T2E, KP951056516 * T2t);		    T3b = FMA(KP951056516, T2E, KP587785252 * T2t);		    ri[WS(rs, 15)] = T2K + T37;		    T3c = T39 + T38;		    ri[WS(rs, 11)] = T3b + T3c;		    ri[WS(rs, 19)] = T3c - T3b;		    T3a = T38 - T39;		    ri[WS(rs, 3)] = T2F + T3a;		    ri[WS(rs, 7)] = T3a - T2F;	       }	       {		    E T4O, T4M, T4N, T4S, T4U, T4Q, T4R, T4T, T4P;		    T4O = KP559016994 * (T4K - T4L);		    T4M = T4K + T4L;		    T4N = FNMS(KP250000000, T4M, T4J);		    T4Q = T30 - T35;		    T4R = T2P - T2U;		    T4S = FNMS(KP587785252, T4R, KP951056516 * T4Q);		    T4U = FMA(KP951056516, T4R, KP587785252 * T4Q);		    ii[WS(rs, 15)] = T4M + T4J;		    T4T = T4O + T4N;		    ii[WS(rs, 11)] = T4T - T4U;		    ii[WS(rs, 19)] = T4U + T4T;		    T4P = T4N - T4O;		    ii[WS(rs, 3)] = T4P - T4S;		    ii[WS(rs, 7)] = T4S + T4P;	       }	       {		    E T3q, T3s, T3t, T3j, T3v, T3f, T3i, T3w, T3u;		    T3q = KP559016994 * (T3m - T3p);		    T3s = T3m + T3p;		    T3t = FNMS(KP250000000, T3s, T3r);		    T3f = T3d - T3e;		    T3i = T3g - T3h;		    T3j = FMA(KP951056516, T3f, KP587785252 * T3i);		    T3v = FNMS(KP587785252, T3f, KP951056516 * T3i);		    ri[WS(rs, 5)] = T3r + T3s;		    T3w = T3t - T3q;		    ri[WS(rs, 13)] = T3v + T3w;		    ri[WS(rs, 17)] = T3w - T3v;		    T3u = T3q + T3t;		    ri[WS(rs, 1)] = T3j + T3u;		    ri[WS(rs, 9)] = T3u - T3j;	       }	       {		    E T4x, T4B, T4C, T4G, T4I, T4E, T4F, T4H, T4D;		    T4x = KP559016994 * (T4v - T4w);		    T4B = T4v + T4w;		    T4C = FNMS(KP250000000, T4B, T4A);		    T4E = T3k - T3l;		    T4F = T3n - T3o;		    T4G = FMA(KP951056516, T4E, KP587785252 * T4F);		    T4I = FNMS(KP587785252, T4E, KP951056516 * T4F);		    ii[WS(rs, 5)] = T4B + T4A;		    T4H = T4C - T4x;		    ii[WS(rs, 13)] = T4H - T4I;		    ii[WS(rs, 17)] = T4I + T4H;		    T4D = T4x + T4C;		    ii[WS(rs, 1)] = T4D - T4G;		    ii[WS(rs, 9)] = T4G + T4D;	       }	  }     }}static const tw_instr twinstr[] = {     {TW_CEXP, 0, 1},     {TW_CEXP, 0, 3},     {TW_CEXP, 0, 9},     {TW_CEXP, 0, 19},     {TW_NEXT, 1, 0}};static const ct_desc desc = { 20, "t2_20", twinstr, &GENUS, {204, 92, 72, 0}, 0, 0, 0 };void X(codelet_t2_20) (planner *p) {     X(kdft_dit_register) (p, t2_20, &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 + -