hc2cfdft_20.c

快速fft变换

		    }
		    {
			 E TS, TW, T1G, T1I;
			 TS = W[36];
			 TW = W[37];
			 T10 = FNMS(TW, TZ, TS * TV);
			 T2Q = FMA(TW, TV, TS * TZ);
			 T1G = W[34];
			 T1I = W[35];
			 T1K = FNMS(T1I, T1J, T1G * T1H);
			 T3s = FMA(T1I, T1H, T1G * T1J);
		    }
	       }
	       {
		    E T1P, T27, T1T, T29;
		    {
			 E T1N, T1O, T1R, T1S;
			 T1N = Ip[WS(rs, 6)];
			 T1O = Im[WS(rs, 6)];
			 T1P = T1N + T1O;
			 T27 = T1N - T1O;
			 T1R = Rp[WS(rs, 6)];
			 T1S = Rm[WS(rs, 6)];
			 T1T = T1R - T1S;
			 T29 = T1R + T1S;
		    }
		    {
			 E T1M, T1Q, T26, T28;
			 T1M = W[24];
			 T1Q = W[25];
			 T1U = FNMS(T1Q, T1T, T1M * T1P);
			 T2J = FMA(T1Q, T1P, T1M * T1T);
			 T26 = W[22];
			 T28 = W[23];
			 T2a = FNMS(T28, T29, T26 * T27);
			 T3x = FMA(T28, T27, T26 * T29);
		    }
	       }
	       {
		    E T16, T2k, T1a, T2i;
		    {
			 E T14, T15, T18, T19;
			 T14 = Ip[WS(rs, 8)];
			 T15 = Im[WS(rs, 8)];
			 T16 = T14 - T15;
			 T2k = T14 + T15;
			 T18 = Rp[WS(rs, 8)];
			 T19 = Rm[WS(rs, 8)];
			 T1a = T18 + T19;
			 T2i = T19 - T18;
		    }
		    {
			 E T13, T17, T2h, T2j;
			 T13 = W[30];
			 T17 = W[31];
			 T1b = FNMS(T17, T1a, T13 * T16);
			 T3l = FMA(T13, T1a, T17 * T16);
			 T2h = W[33];
			 T2j = W[32];
			 T2l = FMA(T2h, T2i, T2j * T2k);
			 T2D = FNMS(T2h, T2k, T2j * T2i);
		    }
	       }
	       {
		    E T2g, T2r, T3n, T3o;
		    {
			 E TI, T11, T4m, T4n;
			 TI = TC - TH;
			 T11 = TR - T10;
			 T12 = TI - T11;
			 T2w = TI + T11;
			 T4m = T3g + T3h;
			 T4n = TR + T10;
			 T4o = T4m + T4n;
			 T4V = T4m - T4n;
		    }
		    {
			 E T2F, T2G, T4w, T4x;
			 T2F = T2D - T2E;
			 T2G = T2a + T2f;
			 T2H = T2F - T2G;
			 T3a = T2F + T2G;
			 T4w = T2l + T2q;
			 T4x = T3x + T3y;
			 T4y = T4w + T4x;
			 T4Y = T4x - T4w;
		    }
		    {
			 E T1l, T1y, T1L, T24;
			 T1l = T1b - T1k;
			 T1y = T1q - T1x;
			 T1z = T1l + T1y;
			 T2v = T1y - T1l;
			 T1L = T1F - T1K;
			 T24 = T1U - T23;
			 T25 = T1L - T24;
			 T2y = T1L + T24;
		    }
		    T2g = T2a - T2f;
		    T2r = T2l - T2q;
		    T2s = T2g - T2r;
		    T2z = T2r + T2g;
		    {
			 E T4t, T4u, T4p, T4q;
			 T4t = T3r + T3s;
			 T4u = T1U + T23;
			 T4v = T4t + T4u;
			 T4X = T4t - T4u;
			 T4p = T3l + T3m;
			 T4q = T1q + T1x;
			 T4r = T4p + T4q;
			 T4U = T4p - T4q;
		    }
		    {
			 E T3w, T3z, T2T, T2W;
			 T3w = T2D + T2E;
			 T3z = T3x - T3y;
			 T3A = T3w + T3z;
			 T3Z = T3z - T3w;
			 T2T = T1b + T1k;
			 T2W = T2U + T2V;
			 T2X = T2T + T2W;
			 T37 = T2T - T2W;
		    }
		    {
			 E T3i, T3j, T2I, T2L;
			 T3i = T3g - T3h;
			 T3j = T2Q - T2P;
			 T3k = T3i + T3j;
			 T41 = T3i - T3j;
			 T2I = T1F + T1K;
			 T2L = T2J + T2K;
			 T2M = T2I + T2L;
			 T39 = T2I - T2L;
		    }
		    {
			 E T3t, T3u, T2O, T2R;
			 T3t = T3r - T3s;
			 T3u = T2K - T2J;
			 T3v = T3t + T3u;
			 T3Y = T3t - T3u;
			 T2O = TC + TH;
			 T2R = T2P + T2Q;
			 T2S = T2O + T2R;
			 T36 = T2O - T2R;
		    }
		    T3n = T3l - T3m;
		    T3o = T2U - T2V;
		    T3p = T3n + T3o;
		    T42 = T3n - T3o;
		    {
			 E Tc, T3M, T4, T8;
			 T4 = W[18];
			 T8 = W[19];
			 Tc = FNMS(T8, Tb, T4 * T7);
			 T3M = FMA(T4, Tb, T8 * T7);
			 Td = T3 - Tc;
			 T4G = T3L + T3M;
			 T33 = Tc + T3;
			 T3N = T3L - T3M;
		    }
		    {
			 E Tm, T30, Tv, T31;
			 {
			      E Te, Ti, Tn, Tr;
			      Te = W[8];
			      Ti = W[9];
			      Tm = FNMS(Ti, Tl, Te * Th);
			      T30 = FMA(Ti, Th, Te * Tl);
			      Tn = W[28];
			      Tr = W[29];
			      Tv = FNMS(Tr, Tu, Tn * Tq);
			      T31 = FMA(Tr, Tq, Tn * Tu);
			 }
			 Tw = Tm - Tv;
			 T4H = Tm + Tv;
			 T32 = T30 + T31;
			 T3O = T31 - T30;
		    }
	       }
	  }
	  {
	       E T3C, T3E, Tx, T2u, T3d, T3e, T3D, T3f;
	       {
		    E T3q, T3B, T1A, T2t;
		    T3q = T3k - T3p;
		    T3B = T3v - T3A;
		    T3C = FMA(KP475528258, T3q, KP293892626 * T3B);
		    T3E = FNMS(KP293892626, T3q, KP475528258 * T3B);
		    Tx = Td - Tw;
		    T1A = T12 + T1z;
		    T2t = T25 + T2s;
		    T2u = T1A + T2t;
		    T3d = KP279508497 * (T1A - T2t);
		    T3e = FNMS(KP125000000, T2u, KP500000000 * Tx);
	       }
	       Ip[WS(rs, 5)] = KP500000000 * (Tx + T2u);
	       T3D = T3d - T3e;
	       Im[WS(rs, 2)] = T3D - T3E;
	       Im[WS(rs, 6)] = T3D + T3E;
	       T3f = T3d + T3e;
	       Ip[WS(rs, 1)] = T3f - T3C;
	       Ip[WS(rs, 9)] = T3f + T3C;
	  }
	  {
	       E T3H, T3T, T3P, T3Q, T3K, T3R, T3U, T3S;
	       {
		    E T3F, T3G, T3I, T3J;
		    T3F = T12 - T1z;
		    T3G = T25 - T2s;
		    T3H = FMA(KP475528258, T3F, KP293892626 * T3G);
		    T3T = FNMS(KP293892626, T3F, KP475528258 * T3G);
		    T3P = T3N + T3O;
		    T3I = T3k + T3p;
		    T3J = T3v + T3A;
		    T3Q = T3I + T3J;
		    T3K = KP279508497 * (T3I - T3J);
		    T3R = FNMS(KP125000000, T3Q, KP500000000 * T3P);
	       }
	       Rp[WS(rs, 5)] = KP500000000 * (T3P + T3Q);
	       T3U = T3R - T3K;
	       Rm[WS(rs, 6)] = T3T + T3U;
	       Rm[WS(rs, 2)] = T3U - T3T;
	       T3S = T3K + T3R;
	       Rp[WS(rs, 1)] = T3H + T3S;
	       Rp[WS(rs, 9)] = T3S - T3H;
	  }
	  {
	       E T44, T46, T2C, T2B, T3V, T3W, T45, T3X;
	       {
		    E T40, T43, T2x, T2A;
		    T40 = T3Y - T3Z;
		    T43 = T41 - T42;
		    T44 = FNMS(KP293892626, T43, KP475528258 * T40);
		    T46 = FMA(KP475528258, T43, KP293892626 * T40);
		    T2C = Tw + Td;
		    T2x = T2v - T2w;
		    T2A = T2y + T2z;
		    T2B = T2x - T2A;
		    T3V = FMA(KP500000000, T2C, KP125000000 * T2B);
		    T3W = KP279508497 * (T2x + T2A);
	       }
	       Im[WS(rs, 4)] = KP500000000 * (T2B - T2C);
	       T45 = T3W - T3V;
	       Im[0] = T45 - T46;
	       Im[WS(rs, 8)] = T45 + T46;
	       T3X = T3V + T3W;
	       Ip[WS(rs, 3)] = T3X - T44;
	       Ip[WS(rs, 7)] = T3X + T44;
	  }
	  {
	       E T49, T4h, T4a, T4d, T4e, T4f, T4i, T4g;
	       {
		    E T47, T48, T4b, T4c;
		    T47 = T2y - T2z;
		    T48 = T2w + T2v;
		    T49 = FNMS(KP293892626, T48, KP475528258 * T47);
		    T4h = FMA(KP475528258, T48, KP293892626 * T47);
		    T4a = T3N - T3O;
		    T4b = T41 + T42;
		    T4c = T3Y + T3Z;
		    T4d = T4b + T4c;
		    T4e = FNMS(KP125000000, T4d, KP500000000 * T4a);
		    T4f = KP279508497 * (T4b - T4c);
	       }
	       Rm[WS(rs, 4)] = KP500000000 * (T4a + T4d);
	       T4i = T4f + T4e;
	       Rm[WS(rs, 8)] = T4h + T4i;
	       Rm[0] = T4i - T4h;
	       T4g = T4e - T4f;
	       Rp[WS(rs, 3)] = T49 + T4g;
	       Rp[WS(rs, 7)] = T4g - T49;
	  }
	  {
	       E T50, T52, T34, T2Z, T4R, T4S, T51, T4T;
	       {
		    E T4W, T4Z, T2N, T2Y;
		    T4W = T4U - T4V;
		    T4Z = T4X - T4Y;
		    T50 = FNMS(KP293892626, T4Z, KP475528258 * T4W);
		    T52 = FMA(KP293892626, T4W, KP475528258 * T4Z);
		    T34 = T32 + T33;
		    T2N = T2H - T2M;
		    T2Y = T2S + T2X;
		    T2Z = T2N - T2Y;
		    T4R = FMA(KP500000000, T34, KP125000000 * T2Z);
		    T4S = KP279508497 * (T2Y + T2N);
	       }
	       Im[WS(rs, 9)] = KP500000000 * (T2Z - T34);
	       T51 = T4R - T4S;
	       Ip[WS(rs, 2)] = T51 + T52;
	       Im[WS(rs, 1)] = T52 - T51;
	       T4T = T4R + T4S;
	       Ip[WS(rs, 6)] = T4T + T50;
	       Im[WS(rs, 5)] = T50 - T4T;
	  }
	  {
	       E T5c, T5d, T53, T56, T57, T58, T5e, T59;
	       {
		    E T5a, T5b, T54, T55;
		    T5a = T2M + T2H;
		    T5b = T2S - T2X;
		    T5c = FNMS(KP293892626, T5b, KP475528258 * T5a);
		    T5d = FMA(KP475528258, T5b, KP293892626 * T5a);
		    T53 = T4G - T4H;
		    T54 = T4V + T4U;
		    T55 = T4X + T4Y;
		    T56 = T54 + T55;
		    T57 = FNMS(KP125000000, T56, KP500000000 * T53);
		    T58 = KP279508497 * (T54 - T55);
	       }
	       Rm[WS(rs, 9)] = KP500000000 * (T53 + T56);
	       T5e = T58 + T57;
	       Rp[WS(rs, 6)] = T5d + T5e;
	       Rm[WS(rs, 5)] = T5e - T5d;
	       T59 = T57 - T58;
	       Rp[WS(rs, 2)] = T59 - T5c;
	       Rm[WS(rs, 1)] = T5c + T59;
	  }
	  {
	       E T4A, T4C, T35, T3c, T4j, T4k, T4B, T4l;
	       {
		    E T4s, T4z, T38, T3b;
		    T4s = T4o - T4r;
		    T4z = T4v - T4y;
		    T4A = FNMS(KP475528258, T4z, KP293892626 * T4s);
		    T4C = FMA(KP475528258, T4s, KP293892626 * T4z);
		    T35 = T33 - T32;
		    T38 = T36 + T37;
		    T3b = T39 + T3a;
		    T3c = T38 + T3b;
		    T4j = FNMS(KP125000000, T3c, KP500000000 * T35);
		    T4k = KP279508497 * (T38 - T3b);
	       }
	       Ip[0] = KP500000000 * (T35 + T3c);
	       T4B = T4k + T4j;
	       Ip[WS(rs, 4)] = T4B + T4C;
	       Im[WS(rs, 3)] = T4C - T4B;
	       T4l = T4j - T4k;
	       Ip[WS(rs, 8)] = T4l + T4A;
	       Im[WS(rs, 7)] = T4A - T4l;
	  }
	  {
	       E T4O, T4P, T4I, T4J, T4F, T4K, T4Q, T4L;
	       {
		    E T4M, T4N, T4D, T4E;
		    T4M = T36 - T37;
		    T4N = T39 - T3a;
		    T4O = FMA(KP475528258, T4M, KP293892626 * T4N);
		    T4P = FNMS(KP293892626, T4M, KP475528258 * T4N);
		    T4I = T4G + T4H;
		    T4D = T4o + T4r;
		    T4E = T4v + T4y;
		    T4J = T4D + T4E;
		    T4F = KP279508497 * (T4D - T4E);
		    T4K = FNMS(KP125000000, T4J, KP500000000 * T4I);
	       }
	       Rp[0] = KP500000000 * (T4I + T4J);
	       T4Q = T4K - T4F;
	       Rp[WS(rs, 8)] = T4P + T4Q;
	       Rm[WS(rs, 7)] = T4Q - T4P;
	       T4L = T4F + T4K;
	       Rp[WS(rs, 4)] = T4L - T4O;
	       Rm[WS(rs, 3)] = T4O + T4L;
	  }
     }
}

static const tw_instr twinstr[] = {
     {TW_FULL, 1, 20},
     {TW_NEXT, 1, 0}
};

static const hc2c_desc desc = { 20, "hc2cfdft_20", twinstr, &GENUS, {224, 78, 62, 0} };

void X(codelet_hc2cfdft_20) (planner *p) {
     X(khc2c_register) (p, hc2cfdft_20, &desc, HC2C_VIA_DFT);
}
#endif				/* HAVE_FMA */