hb2_32.c

快速fft变换

	       }
	       {
		    E TN, T5O, TQ, T5Z, T20, T5P, T1X, T60, TX, T63, T27, T5W, TU, T62, T24;
		    E T5T;
		    {
			 E TL, TM, T1V, T1W;
			 TL = ci[0];
			 TM = cr[WS(rs, 15)];
			 TN = TL + TM;
			 T5O = TL - TM;
			 {
			      E TO, TP, T1Y, T1Z;
			      TO = cr[WS(rs, 7)];
			      TP = ci[WS(rs, 8)];
			      TQ = TO + TP;
			      T5Z = TO - TP;
			      T1Y = ci[WS(rs, 24)];
			      T1Z = cr[WS(rs, 23)];
			      T20 = T1Y - T1Z;
			      T5P = T1Y + T1Z;
			 }
			 T1V = ci[WS(rs, 16)];
			 T1W = cr[WS(rs, 31)];
			 T1X = T1V - T1W;
			 T60 = T1V + T1W;
			 {
			      E TV, TW, T5U, T25, T26, T5V;
			      TV = ci[WS(rs, 4)];
			      TW = cr[WS(rs, 11)];
			      T5U = TV - TW;
			      T25 = ci[WS(rs, 20)];
			      T26 = cr[WS(rs, 27)];
			      T5V = T25 + T26;
			      TX = TV + TW;
			      T63 = T5U + T5V;
			      T27 = T25 - T26;
			      T5W = T5U - T5V;
			 }
			 {
			      E TS, TT, T5R, T22, T23, T5S;
			      TS = cr[WS(rs, 3)];
			      TT = ci[WS(rs, 12)];
			      T5R = TS - TT;
			      T22 = ci[WS(rs, 28)];
			      T23 = cr[WS(rs, 19)];
			      T5S = T22 + T23;
			      TU = TS + TT;
			      T62 = T5R + T5S;
			      T24 = T22 - T23;
			      T5T = T5R - T5S;
			 }
		    }
		    {
			 E TR, TY, T7m, T7n;
			 TR = TN + TQ;
			 TY = TU + TX;
			 TZ = TR + TY;
			 T1U = TR - TY;
			 T7m = KP707106781 * (T5T - T5W);
			 T7n = T5Z + T60;
			 T7o = T7m - T7n;
			 T89 = T7n + T7m;
		    }
		    {
			 E T7p, T7q, T21, T28;
			 T7p = T5O + T5P;
			 T7q = KP707106781 * (T62 + T63);
			 T7r = T7p - T7q;
			 T88 = T7p + T7q;
			 T21 = T1X + T20;
			 T28 = T24 + T27;
			 T29 = T21 - T28;
			 T36 = T21 + T28;
		    }
		    {
			 E T3F, T3G, T5Q, T5X;
			 T3F = T1X - T20;
			 T3G = TU - TX;
			 T3H = T3F - T3G;
			 T4z = T3G + T3F;
			 T5Q = T5O - T5P;
			 T5X = KP707106781 * (T5T + T5W);
			 T5Y = T5Q - T5X;
			 T6M = T5Q + T5X;
		    }
		    {
			 E T61, T64, T3I, T3J;
			 T61 = T5Z - T60;
			 T64 = KP707106781 * (T62 - T63);
			 T65 = T61 - T64;
			 T6N = T61 + T64;
			 T3I = TN - TQ;
			 T3J = T27 - T24;
			 T3K = T3I - T3J;
			 T4A = T3I + T3J;
		    }
	       }
	       {
		    E Tv, T10, T30, T34, T37, T38;
		    Tv = Tf + Tu;
		    T10 = TK + TZ;
		    T30 = Tv - T10;
		    T34 = T32 + T33;
		    T37 = T35 + T36;
		    T38 = T34 - T37;
		    cr[0] = Tv + T10;
		    ci[0] = T34 + T37;
		    cr[WS(rs, 16)] = FNMS(T31, T38, T2Z * T30);
		    ci[WS(rs, 16)] = FMA(T31, T30, T2Z * T38);
	       }
	       {
		    E T3e, T3o, T3k, T3s;
		    {
			 E T3c, T3d, T3i, T3j;
			 T3c = Tf - Tu;
			 T3d = T36 - T35;
			 T3e = T3c - T3d;
			 T3o = T3c + T3d;
			 T3i = T32 - T33;
			 T3j = TK - TZ;
			 T3k = T3i - T3j;
			 T3s = T3j + T3i;
		    }
		    cr[WS(rs, 24)] = FNMS(T3h, T3k, T3b * T3e);
		    ci[WS(rs, 24)] = FMA(T3b, T3k, T3h * T3e);
		    cr[WS(rs, 8)] = FNMS(T3r, T3s, T3n * T3o);
		    ci[WS(rs, 8)] = FMA(T3n, T3s, T3r * T3o);
	       }
	       {
		    E T1C, T2u, T2M, T2G, T2x, T2H, T2b, T2N;
		    T1C = T1m + T1B;
		    T2u = T2e + T2t;
		    T2M = T2t - T2e;
		    T2G = T1m - T1B;
		    {
			 E T2v, T2w, T1T, T2a;
			 T2v = T1D + T1S;
			 T2w = T29 - T1U;
			 T2x = KP707106781 * (T2v + T2w);
			 T2H = KP707106781 * (T2w - T2v);
			 T1T = T1D - T1S;
			 T2a = T1U + T29;
			 T2b = KP707106781 * (T1T + T2a);
			 T2N = KP707106781 * (T1T - T2a);
		    }
		    {
			 E T2c, T2y, T2S, T2W;
			 T2c = T1C - T2b;
			 T2y = T2u - T2x;
			 cr[WS(rs, 20)] = FNMS(T2d, T2y, T1l * T2c);
			 ci[WS(rs, 20)] = FMA(T2d, T2c, T1l * T2y);
			 T2S = T2G + T2H;
			 T2W = T2M + T2N;
			 cr[WS(rs, 12)] = FNMS(T2V, T2W, T2R * T2S);
			 ci[WS(rs, 12)] = FMA(T2R, T2W, T2V * T2S);
		    }
		    {
			 E T2A, T2C, T2I, T2O;
			 T2A = T1C + T2b;
			 T2C = T2u + T2x;
			 cr[WS(rs, 4)] = FNMS(T2B, T2C, T2z * T2A);
			 ci[WS(rs, 4)] = FMA(T2B, T2A, T2z * T2C);
			 T2I = T2G - T2H;
			 T2O = T2M - T2N;
			 cr[WS(rs, 28)] = FNMS(T2L, T2O, T2F * T2I);
			 ci[WS(rs, 28)] = FMA(T2F, T2O, T2L * T2I);
		    }
	       }
	       {
		    E T4y, T4U, T4K, T4Y, T4F, T4Z, T4N, T4V, T4x, T4J;
		    T4x = KP707106781 * (T3Z + T40);
		    T4y = T4w - T4x;
		    T4U = T4w + T4x;
		    T4J = KP707106781 * (T3C + T3z);
		    T4K = T4I - T4J;
		    T4Y = T4I + T4J;
		    {
			 E T4B, T4E, T4L, T4M;
			 T4B = FNMS(KP382683432, T4A, KP923879532 * T4z);
			 T4E = FMA(KP923879532, T4C, KP382683432 * T4D);
			 T4F = T4B - T4E;
			 T4Z = T4E + T4B;
			 T4L = FNMS(KP382683432, T4C, KP923879532 * T4D);
			 T4M = FMA(KP382683432, T4z, KP923879532 * T4A);
			 T4N = T4L - T4M;
			 T4V = T4L + T4M;
		    }
		    {
			 E T4G, T4O, T51, T52;
			 T4G = T4y - T4F;
			 T4O = T4K - T4N;
			 cr[WS(rs, 26)] = FNMS(T4H, T4O, T4v * T4G);
			 ci[WS(rs, 26)] = FMA(T4H, T4G, T4v * T4O);
			 T51 = T4U + T4V;
			 T52 = T4Y + T4Z;
			 cr[WS(rs, 2)] = FNMS(T1c, T52, T17 * T51);
			 ci[WS(rs, 2)] = FMA(T17, T52, T1c * T51);
		    }
		    {
			 E T4Q, T4S, T4W, T50;
			 T4Q = T4y + T4F;
			 T4S = T4K + T4N;
			 cr[WS(rs, 10)] = FNMS(T4R, T4S, T4P * T4Q);
			 ci[WS(rs, 10)] = FMA(T4R, T4Q, T4P * T4S);
			 T4W = T4U - T4V;
			 T50 = T4Y - T4Z;
			 cr[WS(rs, 18)] = FNMS(T4X, T50, T4T * T4W);
			 ci[WS(rs, 18)] = FMA(T4T, T50, T4X * T4W);
		    }
	       }
	       {
		    E T3E, T4k, T42, T4o, T3T, T4p, T45, T4l, T3D, T41;
		    T3D = KP707106781 * (T3z - T3C);
		    T3E = T3w - T3D;
		    T4k = T3w + T3D;
		    T41 = KP707106781 * (T3Z - T40);
		    T42 = T3Y - T41;
		    T4o = T3Y + T41;
		    {
			 E T3L, T3S, T43, T44;
			 T3L = FNMS(KP923879532, T3K, KP382683432 * T3H);
			 T3S = FMA(KP382683432, T3O, KP923879532 * T3R);
			 T3T = T3L - T3S;
			 T4p = T3S + T3L;
			 T43 = FNMS(KP923879532, T3O, KP382683432 * T3R);
			 T44 = FMA(KP923879532, T3H, KP382683432 * T3K);
			 T45 = T43 - T44;
			 T4l = T43 + T44;
		    }
		    {
			 E T3U, T46, T4s, T4u;
			 T3U = T3E - T3T;
			 T46 = T42 - T45;
			 cr[WS(rs, 30)] = FNMS(T3V, T46, T3t * T3U);
			 ci[WS(rs, 30)] = FMA(T3V, T3U, T3t * T46);
			 T4s = T4k + T4l;
			 T4u = T4o + T4p;
			 cr[WS(rs, 6)] = FNMS(T4t, T4u, T4r * T4s);
			 ci[WS(rs, 6)] = FMA(T4r, T4u, T4t * T4s);
		    }
		    {
			 E T4e, T4g, T4m, T4q;
			 T4e = T3E + T3T;
			 T4g = T42 + T45;
			 cr[WS(rs, 14)] = FNMS(T4f, T4g, T4d * T4e);
			 ci[WS(rs, 14)] = FMA(T4f, T4e, T4d * T4g);
			 T4m = T4k - T4l;
			 T4q = T4o - T4p;
			 cr[WS(rs, 22)] = FNMS(T4n, T4q, T4j * T4m);
			 ci[WS(rs, 22)] = FMA(T4j, T4q, T4n * T4m);
		    }
	       }
	       {
		    E T6I, T72, T6X, T73, T6P, T77, T6U, T76;
		    {
			 E T6G, T6H, T6V, T6W;
			 T6G = T56 + T5d;
			 T6H = T6h + T6i;
			 T6I = T6G + T6H;
			 T72 = T6G - T6H;
			 T6V = FMA(KP195090322, T6J, KP980785280 * T6K);
			 T6W = FNMS(KP195090322, T6M, KP980785280 * T6N);
			 T6X = T6V + T6W;
			 T73 = T6W - T6V;
		    }
		    {
			 E T6L, T6O, T6S, T6T;
			 T6L = FNMS(KP195090322, T6K, KP980785280 * T6J);
			 T6O = FMA(KP980785280, T6M, KP195090322 * T6N);
			 T6P = T6L + T6O;
			 T77 = T6L - T6O;
			 T6S = T6c + T6f;
			 T6T = T5s + T5l;
			 T6U = T6S + T6T;
			 T76 = T6S - T6T;
		    }
		    {
			 E T6Q, T6Y, T79, T7a;
			 T6Q = T6I - T6P;
			 T6Y = T6U - T6X;
			 cr[WS(rs, 17)] = FNMS(T6R, T6Y, T6F * T6Q);
			 ci[WS(rs, 17)] = FMA(T6R, T6Q, T6F * T6Y);
			 T79 = T72 + T73;
			 T7a = T76 + T77;
			 cr[WS(rs, 9)] = FNMS(T1d, T7a, T18 * T79);
			 ci[WS(rs, 9)] = FMA(T18, T7a, T1d * T79);
		    }
		    {
			 E T6Z, T70, T74, T78;
			 T6Z = T6I + T6P;
			 T70 = T6U + T6X;
			 cr[WS(rs, 1)] = FNMS(T14, T70, T11 * T6Z);
			 ci[WS(rs, 1)] = FMA(T14, T6Z, T11 * T70);
			 T74 = T72 - T73;
			 T78 = T76 - T77;
			 cr[WS(rs, 25)] = FNMS(T75, T78, T71 * T74);
			 ci[WS(rs, 25)] = FMA(T71, T78, T75 * T74);
		    }
	       }
	       {
		    E T84, T8q, T8l, T8r, T8b, T8v, T8i, T8u;
		    {
			 E T82, T83, T8j, T8k;
			 T82 = T7b + T7c;
			 T83 = T7F + T7G;
			 T84 = T82 - T83;
			 T8q = T82 + T83;
			 T8j = FMA(KP195090322, T86, KP980785280 * T85);
			 T8k = FMA(KP195090322, T89, KP980785280 * T88);
			 T8l = T8j - T8k;
			 T8r = T8j + T8k;
		    }
		    {
			 E T87, T8a, T8g, T8h;
			 T87 = FNMS(KP980785280, T86, KP195090322 * T85);
			 T8a = FNMS(KP980785280, T89, KP195090322 * T88);
			 T8b = T87 + T8a;
			 T8v = T87 - T8a;
			 T8g = T7C - T7D;
			 T8h = T7g - T7j;
			 T8i = T8g + T8h;
			 T8u = T8g - T8h;
		    }
		    {
			 E T8c, T8m, T8y, T8A;
			 T8c = T84 - T8b;
			 T8m = T8i - T8l;
			 cr[WS(rs, 23)] = FNMS(T8f, T8m, T81 * T8c);
			 ci[WS(rs, 23)] = FMA(T8f, T8c, T81 * T8m);
			 T8y = T8q + T8r;
			 T8A = T8u - T8v;
			 cr[WS(rs, 31)] = FNMS(T8z, T8A, T8x * T8y);
			 ci[WS(rs, 31)] = FMA(T8x, T8A, T8z * T8y);
		    }
		    {
			 E T8n, T8o, T8s, T8w;
			 T8n = T84 + T8b;
			 T8o = T8i + T8l;
			 cr[WS(rs, 7)] = FNMS(T1j, T8o, T1f * T8n);
			 ci[WS(rs, 7)] = FMA(T1j, T8n, T1f * T8o);
			 T8s = T8q - T8r;
			 T8w = T8u + T8v;
			 cr[WS(rs, 15)] = FNMS(T8t, T8w, T8p * T8s);
			 ci[WS(rs, 15)] = FMA(T8p, T8w, T8t * T8s);
		    }
	       }
	       {
		    E T5u, T6u, T6n, T6v, T67, T6B, T6k, T6A;
		    {
			 E T5e, T5t, T6l, T6m;
			 T5e = T56 - T5d;
			 T5t = T5l - T5s;
			 T5u = T5e + T5t;
			 T6u = T5e - T5t;
			 T6l = FMA(KP831469612, T5F, KP555570233 * T5M);
			 T6m = FNMS(KP831469612, T5Y, KP555570233 * T65);
			 T6n = T6l + T6m;
			 T6v = T6m - T6l;
		    }
		    {
			 E T5N, T66, T6g, T6j;
			 T5N = FNMS(KP831469612, T5M, KP555570233 * T5F);
			 T66 = FMA(KP555570233, T5Y, KP831469612 * T65);
			 T67 = T5N + T66;
			 T6B = T5N - T66;
			 T6g = T6c - T6f;
			 T6j = T6h - T6i;
			 T6k = T6g + T6j;
			 T6A = T6g - T6j;
		    }
		    {
			 E T68, T6o, T6D, T6E;
			 T68 = T5u - T67;
			 T6o = T6k - T6n;
			 cr[WS(rs, 21)] = FNMS(T69, T6o, T53 * T68);
			 ci[WS(rs, 21)] = FMA(T69, T68, T53 * T6o);
			 T6D = T6u + T6v;
			 T6E = T6A + T6B;
			 cr[WS(rs, 13)] = FNMS(T4c, T6E, T49 * T6D);
			 ci[WS(rs, 13)] = FMA(T49, T6E, T4c * T6D);
		    }
		    {
			 E T6p, T6q, T6w, T6C;
			 T6p = T5u + T67;
			 T6q = T6k + T6n;
			 cr[WS(rs, 5)] = FNMS(T4i, T6q, T4h * T6p);
			 ci[WS(rs, 5)] = FMA(T4i, T6p, T4h * T6q);
			 T6w = T6u - T6v;
			 T6C = T6A - T6B;
			 cr[WS(rs, 29)] = FNMS(T6z, T6C, T6t * T6w);
			 ci[WS(rs, 29)] = FMA(T6t, T6C, T6z * T6w);
		    }
	       }
	       {
		    E T7l, T7Q, T7L, T7R, T7A, T7V, T7I, T7U;
		    {
			 E T7d, T7k, T7J, T7K;
			 T7d = T7b - T7c;
			 T7k = T7g + T7j;
			 T7l = T7d - T7k;
			 T7Q = T7d + T7k;
			 T7J = FNMS(KP555570233, T7v, KP831469612 * T7y);
			 T7K = FMA(KP555570233, T7o, KP831469612 * T7r);
			 T7L = T7J - T7K;
			 T7R = T7J + T7K;
		    }
		    {
			 E T7s, T7z, T7E, T7H;
			 T7s = FNMS(KP555570233, T7r, KP831469612 * T7o);
			 T7z = FMA(KP831469612, T7v, KP555570233 * T7y);
			 T7A = T7s - T7z;
			 T7V = T7z + T7s;
			 T7E = T7C + T7D;
			 T7H = T7F - T7G;
			 T7I = T7E - T7H;
			 T7U = T7E + T7H;
		    }
		    {
			 E T7B, T7M, T7X, T7Y;
			 T7B = T7l - T7A;
			 T7M = T7I - T7L;
			 cr[WS(rs, 27)] = FNMS(T1k, T7M, T1g * T7B);
			 ci[WS(rs, 27)] = FMA(T1k, T7B, T1g * T7M);
			 T7X = T7Q + T7R;
			 T7Y = T7U + T7V;
			 cr[WS(rs, 3)] = FNMS(T15, T7Y, T12 * T7X);
			 ci[WS(rs, 3)] = FMA(T12, T7Y, T15 * T7X);
		    }
		    {
			 E T7N, T7O, T7S, T7W;
			 T7N = T7l + T7A;
			 T7O = T7I + T7L;
			 cr[WS(rs, 11)] = FNMS(T2Y, T7O, T2X * T7N);
			 ci[WS(rs, 11)] = FMA(T2Y, T7N, T2X * T7O);
			 T7S = T7Q - T7R;
			 T7W = T7U - T7V;
			 cr[WS(rs, 19)] = FNMS(T7T, T7W, T7P * T7S);
			 ci[WS(rs, 19)] = FMA(T7P, T7W, T7T * T7S);
		    }
	       }
	  }
     }
}

static const tw_instr twinstr[] = {
     {TW_CEXP, 1, 1},
     {TW_CEXP, 1, 3},
     {TW_CEXP, 1, 9},
     {TW_CEXP, 1, 27},
     {TW_NEXT, 1, 0}
};

static const hc2hc_desc desc = { 32, "hb2_32", twinstr, &GENUS, {376, 168, 112, 0} };

void X(codelet_hb2_32) (planner *p) {
     X(khc2hc_register) (p, hb2_32, &desc);
}
#endif				/* HAVE_FMA */