hb_32.c

最新的FFT程序

		    T46 = TZ - TK;
		    T49 = T47 - T48;
		    T4a = T46 + T49;
		    T4e = T49 - T46;
	       }
	       {
		    E T3Z, T45, T4b, T4d;
		    T3Z = W[46];
		    T45 = W[47];
		    rio[WS(ios, 24)] = FNMS(T45, T4a, T3Z * T44);
		    iio[-WS(ios, 7)] = FMA(T45, T44, T3Z * T4a);
		    T4b = W[14];
		    T4d = W[15];
		    rio[WS(ios, 8)] = FNMS(T4d, T4e, T4b * T4c);
		    iio[-WS(ios, 23)] = FMA(T4d, T4c, T4b * T4e);
	       }
	  }
	  {
	       E T3m, T3H, T3T, T3O, T3C, T3P, T3x, T3S;
	       T3m = T3i + T3l;
	       T3H = T3D + T3G;
	       T3T = T3G - T3D;
	       T3O = T3i - T3l;
	       {
		    E T3A, T3B, T3r, T3w;
		    T3A = T3q - T3n;
		    T3B = T3s + T3v;
		    T3C = KP707106781 * (T3A + T3B);
		    T3P = KP707106781 * (T3B - T3A);
		    T3r = T3n + T3q;
		    T3w = T3s - T3v;
		    T3x = KP707106781 * (T3r + T3w);
		    T3S = KP707106781 * (T3r - T3w);
	       }
	       {
		    E T3y, T3I, T3h, T3z;
		    T3y = T3m + T3x;
		    T3I = T3C + T3H;
		    T3h = W[54];
		    T3z = W[55];
		    rio[WS(ios, 28)] = FNMS(T3z, T3I, T3h * T3y);
		    iio[-WS(ios, 3)] = FMA(T3z, T3y, T3h * T3I);
	       }
	       {
		    E T3W, T3Y, T3V, T3X;
		    T3W = T3O - T3P;
		    T3Y = T3T - T3S;
		    T3V = W[38];
		    T3X = W[39];
		    rio[WS(ios, 20)] = FNMS(T3X, T3Y, T3V * T3W);
		    iio[-WS(ios, 11)] = FMA(T3X, T3W, T3V * T3Y);
	       }
	       {
		    E T3K, T3M, T3J, T3L;
		    T3K = T3m - T3x;
		    T3M = T3H - T3C;
		    T3J = W[22];
		    T3L = W[23];
		    rio[WS(ios, 12)] = FNMS(T3L, T3M, T3J * T3K);
		    iio[-WS(ios, 19)] = FMA(T3L, T3K, T3J * T3M);
	       }
	       {
		    E T3Q, T3U, T3N, T3R;
		    T3Q = T3O + T3P;
		    T3U = T3S + T3T;
		    T3N = W[6];
		    T3R = W[7];
		    rio[WS(ios, 4)] = FNMS(T3R, T3U, T3N * T3Q);
		    iio[-WS(ios, 27)] = FMA(T3R, T3Q, T3N * T3U);
	       }
	  }
	  {
	       E T2K, T36, T2Z, T3b, T2R, T3a, T2W, T37, T2J, T2X;
	       T2J = KP707106781 * (T2e - T2d);
	       T2K = T2I + T2J;
	       T36 = T2I - T2J;
	       T2X = KP707106781 * (T1j - T1s);
	       T2Z = T2X + T2Y;
	       T3b = T2Y - T2X;
	       {
		    E T2N, T2Q, T2U, T2V;
		    T2N = FNMS(KP382683432, T2M, KP923879532 * T2L);
		    T2Q = FMA(KP923879532, T2O, KP382683432 * T2P);
		    T2R = T2N + T2Q;
		    T3a = T2Q - T2N;
		    T2U = FMA(KP382683432, T2L, KP923879532 * T2M);
		    T2V = FNMS(KP382683432, T2O, KP923879532 * T2P);
		    T2W = T2U + T2V;
		    T37 = T2U - T2V;
	       }
	       {
		    E T2S, T30, T2H, T2T;
		    T2S = T2K + T2R;
		    T30 = T2W + T2Z;
		    T2H = W[2];
		    T2T = W[3];
		    rio[WS(ios, 2)] = FNMS(T2T, T30, T2H * T2S);
		    iio[-WS(ios, 29)] = FMA(T2T, T2S, T2H * T30);
	       }
	       {
		    E T3e, T3g, T3d, T3f;
		    T3e = T36 - T37;
		    T3g = T3b - T3a;
		    T3d = W[18];
		    T3f = W[19];
		    rio[WS(ios, 10)] = FNMS(T3f, T3g, T3d * T3e);
		    iio[-WS(ios, 21)] = FMA(T3f, T3e, T3d * T3g);
	       }
	       {
		    E T32, T34, T31, T33;
		    T32 = T2K - T2R;
		    T34 = T2Z - T2W;
		    T31 = W[34];
		    T33 = W[35];
		    rio[WS(ios, 18)] = FNMS(T33, T34, T31 * T32);
		    iio[-WS(ios, 13)] = FMA(T33, T32, T31 * T34);
	       }
	       {
		    E T38, T3c, T35, T39;
		    T38 = T36 + T37;
		    T3c = T3a + T3b;
		    T35 = W[50];
		    T39 = W[51];
		    rio[WS(ios, 26)] = FNMS(T39, T3c, T35 * T38);
		    iio[-WS(ios, 5)] = FMA(T39, T38, T35 * T3c);
	       }
	  }
	  {
	       E T1u, T2w, T2p, T2B, T27, T2A, T2c, T2x, T1t, T2f;
	       T1t = KP707106781 * (T1j + T1s);
	       T1u = T1a + T1t;
	       T2w = T1a - T1t;
	       T2f = KP707106781 * (T2d + T2e);
	       T2p = T2f + T2o;
	       T2B = T2o - T2f;
	       {
		    E T1N, T26, T2a, T2b;
		    T1N = FMA(KP923879532, T1D, KP382683432 * T1M);
		    T26 = FNMS(KP382683432, T25, KP923879532 * T1W);
		    T27 = T1N + T26;
		    T2A = T1N - T26;
		    T2a = FNMS(KP382683432, T1D, KP923879532 * T1M);
		    T2b = FMA(KP382683432, T1W, KP923879532 * T25);
		    T2c = T2a + T2b;
		    T2x = T2b - T2a;
	       }
	       {
		    E T28, T2q, T11, T29;
		    T28 = T1u + T27;
		    T2q = T2c + T2p;
		    T11 = W[58];
		    T29 = W[59];
		    rio[WS(ios, 30)] = FNMS(T29, T2q, T11 * T28);
		    iio[-WS(ios, 1)] = FMA(T29, T28, T11 * T2q);
	       }
	       {
		    E T2E, T2G, T2D, T2F;
		    T2E = T2w - T2x;
		    T2G = T2B - T2A;
		    T2D = W[42];
		    T2F = W[43];
		    rio[WS(ios, 22)] = FNMS(T2F, T2G, T2D * T2E);
		    iio[-WS(ios, 9)] = FMA(T2F, T2E, T2D * T2G);
	       }
	       {
		    E T2s, T2u, T2r, T2t;
		    T2s = T1u - T27;
		    T2u = T2p - T2c;
		    T2r = W[26];
		    T2t = W[27];
		    rio[WS(ios, 14)] = FNMS(T2t, T2u, T2r * T2s);
		    iio[-WS(ios, 17)] = FMA(T2t, T2s, T2r * T2u);
	       }
	       {
		    E T2y, T2C, T2v, T2z;
		    T2y = T2w + T2x;
		    T2C = T2A + T2B;
		    T2v = W[10];
		    T2z = W[11];
		    rio[WS(ios, 6)] = FNMS(T2z, T2C, T2v * T2y);
		    iio[-WS(ios, 25)] = FMA(T2z, T2y, T2v * T2C);
	       }
	  }
	  {
	       E T4Y, T5N, T5F, T5Q, T5p, T5R, T5C, T5M;
	       {
		    E T4E, T4X, T5D, T5E;
		    T4E = FNMS(KP195090322, T4D, KP980785280 * T4w);
		    T4X = FMA(KP195090322, T4P, KP980785280 * T4W);
		    T4Y = T4E + T4X;
		    T5N = T4X - T4E;
		    T5D = FMA(KP980785280, T4D, KP195090322 * T4w);
		    T5E = FNMS(KP195090322, T4W, KP980785280 * T4P);
		    T5F = T5D + T5E;
		    T5Q = T5D - T5E;
	       }
	       {
		    E T5d, T5o, T5y, T5B;
		    T5d = T55 + T5c;
		    T5o = T5k + T5n;
		    T5p = T5d + T5o;
		    T5R = T5o - T5d;
		    T5y = T5u + T5x;
		    T5B = T5z + T5A;
		    T5C = T5y + T5B;
		    T5M = T5y - T5B;
	       }
	       {
		    E T5q, T5G, T4l, T5r;
		    T5q = T4Y + T5p;
		    T5G = T5C + T5F;
		    T4l = W[0];
		    T5r = W[1];
		    rio[WS(ios, 1)] = FNMS(T5r, T5G, T4l * T5q);
		    iio[-WS(ios, 30)] = FMA(T4l, T5G, T5r * T5q);
	       }
	       {
		    E T5U, T5W, T5T, T5V;
		    T5U = T5R - T5Q;
		    T5W = T5M - T5N;
		    T5T = W[16];
		    T5V = W[17];
		    rio[WS(ios, 9)] = FNMS(T5V, T5W, T5T * T5U);
		    iio[-WS(ios, 22)] = FMA(T5T, T5W, T5V * T5U);
	       }
	       {
		    E T5I, T5K, T5H, T5J;
		    T5I = T5C - T5F;
		    T5K = T5p - T4Y;
		    T5H = W[32];
		    T5J = W[33];
		    iio[-WS(ios, 14)] = FMA(T5H, T5I, T5J * T5K);
		    rio[WS(ios, 17)] = FNMS(T5J, T5I, T5H * T5K);
	       }
	       {
		    E T5O, T5S, T5L, T5P;
		    T5O = T5M + T5N;
		    T5S = T5Q + T5R;
		    T5L = W[48];
		    T5P = W[49];
		    iio[-WS(ios, 6)] = FMA(T5L, T5O, T5P * T5S);
		    rio[WS(ios, 25)] = FNMS(T5P, T5O, T5L * T5S);
	       }
	  }
	  {
	       E T60, T6q, T6f, T6n, T67, T6m, T6c, T6r;
	       {
		    E T5Y, T5Z, T6d, T6e;
		    T5Y = T5u - T5x;
		    T5Z = T5c - T55;
		    T60 = T5Y + T5Z;
		    T6q = T5Y - T5Z;
		    T6d = T5z - T5A;
		    T6e = T5n - T5k;
		    T6f = T6d + T6e;
		    T6n = T6e - T6d;
	       }
	       {
		    E T63, T66, T6a, T6b;
		    T63 = FNMS(KP555570233, T62, KP831469612 * T61);
		    T66 = FMA(KP831469612, T64, KP555570233 * T65);
		    T67 = T63 + T66;
		    T6m = T66 - T63;
		    T6a = FMA(KP555570233, T61, KP831469612 * T62);
		    T6b = FNMS(KP555570233, T64, KP831469612 * T65);
		    T6c = T6a + T6b;
		    T6r = T6a - T6b;
	       }
	       {
		    E T68, T6g, T5X, T69;
		    T68 = T60 + T67;
		    T6g = T6c + T6f;
		    T5X = W[56];
		    T69 = W[57];
		    iio[-WS(ios, 2)] = FMA(T5X, T68, T69 * T6g);
		    rio[WS(ios, 29)] = FNMS(T69, T68, T5X * T6g);
	       }
	       {
		    E T6u, T6w, T6t, T6v;
		    T6u = T6q - T6r;
		    T6w = T6n - T6m;
		    T6t = W[40];
		    T6v = W[41];
		    iio[-WS(ios, 10)] = FMA(T6t, T6u, T6v * T6w);
		    rio[WS(ios, 21)] = FNMS(T6v, T6u, T6t * T6w);
	       }
	       {
		    E T6i, T6k, T6h, T6j;
		    T6i = T6f - T6c;
		    T6k = T60 - T67;
		    T6h = W[24];
		    T6j = W[25];
		    rio[WS(ios, 13)] = FNMS(T6j, T6k, T6h * T6i);
		    iio[-WS(ios, 18)] = FMA(T6h, T6k, T6j * T6i);
	       }
	       {
		    E T6o, T6s, T6l, T6p;
		    T6o = T6m + T6n;
		    T6s = T6q + T6r;
		    T6l = W[8];
		    T6p = W[9];
		    rio[WS(ios, 5)] = FNMS(T6p, T6s, T6l * T6o);
		    iio[-WS(ios, 26)] = FMA(T6l, T6s, T6p * T6o);
	       }
	  }
	  {
	       E T7y, T7R, T7J, T7U, T7B, T7V, T7G, T7Q;
	       {
		    E T7u, T7x, T7H, T7I;
		    T7u = FNMS(KP555570233, T7t, KP831469612 * T7s);
		    T7x = FNMS(KP555570233, T7w, KP831469612 * T7v);
		    T7y = T7u + T7x;
		    T7R = T7x - T7u;
		    T7H = FMA(KP831469612, T7t, KP555570233 * T7s);
		    T7I = FMA(KP831469612, T7w, KP555570233 * T7v);
		    T7J = T7H - T7I;
		    T7U = T7H + T7I;
	       }
	       {
		    E T7z, T7A, T7E, T7F;
		    T7z = T6G - T6D;
		    T7A = T77 - T76;
		    T7B = T7z + T7A;
		    T7V = T7A - T7z;
		    T7E = T6y - T6z;
		    T7F = T73 - T74;
		    T7G = T7E + T7F;
		    T7Q = T7E - T7F;
	       }
	       {
		    E T7C, T7K, T7r, T7D;
		    T7C = T7y + T7B;
		    T7K = T7G + T7J;
		    T7r = W[4];
		    T7D = W[5];
		    rio[WS(ios, 3)] = FNMS(T7D, T7K, T7r * T7C);
		    iio[-WS(ios, 28)] = FMA(T7r, T7K, T7D * T7C);
	       }
	       {
		    E T7Y, T80, T7X, T7Z;
		    T7Y = T7V - T7U;
		    T80 = T7Q - T7R;
		    T7X = W[20];
		    T7Z = W[21];
		    rio[WS(ios, 11)] = FNMS(T7Z, T80, T7X * T7Y);
		    iio[-WS(ios, 20)] = FMA(T7X, T80, T7Z * T7Y);
	       }
	       {
		    E T7M, T7O, T7L, T7N;
		    T7M = T7G - T7J;
		    T7O = T7B - T7y;
		    T7L = W[36];
		    T7N = W[37];
		    iio[-WS(ios, 12)] = FMA(T7L, T7M, T7N * T7O);
		    rio[WS(ios, 19)] = FNMS(T7N, T7M, T7L * T7O);
	       }
	       {
		    E T7S, T7W, T7P, T7T;
		    T7S = T7Q + T7R;
		    T7W = T7U + T7V;
		    T7P = W[52];
		    T7T = W[53];
		    iio[-WS(ios, 4)] = FMA(T7P, T7S, T7T * T7W);
		    rio[WS(ios, 27)] = FNMS(T7T, T7S, T7P * T7W);
	       }
	  }
	  {
	       E T6I, T7k, T79, T7h, T6X, T7g, T72, T7l;
	       {
		    E T6A, T6H, T75, T78;
		    T6A = T6y + T6z;
		    T6H = T6D + T6G;
		    T6I = T6A + T6H;
		    T7k = T6A - T6H;
		    T75 = T73 + T74;
		    T78 = T76 + T77;
		    T79 = T75 + T78;
		    T7h = T78 - T75;
	       }
	       {
		    E T6P, T6W, T70, T71;
		    T6P = FNMS(KP195090322, T6O, KP980785280 * T6L);
		    T6W = FMA(KP980785280, T6S, KP195090322 * T6V);
		    T6X = T6P + T6W;
		    T7g = T6W - T6P;
		    T70 = FMA(KP195090322, T6L, KP980785280 * T6O);
		    T71 = FNMS(KP195090322, T6S, KP980785280 * T6V);
		    T72 = T70 + T71;
		    T7l = T70 - T71;
	       }
	       {
		    E T6Y, T7a, T6x, T6Z;
		    T6Y = T6I + T6X;
		    T7a = T72 + T79;
		    T6x = W[60];
		    T6Z = W[61];
		    iio[0] = FMA(T6x, T6Y, T6Z * T7a);
		    rio[WS(ios, 31)] = FNMS(T6Z, T6Y, T6x * T7a);
	       }
	       {
		    E T7o, T7q, T7n, T7p;
		    T7o = T7k - T7l;
		    T7q = T7h - T7g;
		    T7n = W[44];
		    T7p = W[45];
		    iio[-WS(ios, 8)] = FMA(T7n, T7o, T7p * T7q);
		    rio[WS(ios, 23)] = FNMS(T7p, T7o, T7n * T7q);
	       }
	       {
		    E T7c, T7e, T7b, T7d;
		    T7c = T79 - T72;
		    T7e = T6I - T6X;
		    T7b = W[28];
		    T7d = W[29];
		    rio[WS(ios, 15)] = FNMS(T7d, T7e, T7b * T7c);
		    iio[-WS(ios, 16)] = FMA(T7b, T7e, T7d * T7c);
	       }
	       {
		    E T7i, T7m, T7f, T7j;
		    T7i = T7g + T7h;
		    T7m = T7k + T7l;
		    T7f = W[12];
		    T7j = W[13];
		    rio[WS(ios, 7)] = FNMS(T7j, T7m, T7f * T7i);
		    iio[-WS(ios, 24)] = FMA(T7f, T7m, T7j * T7i);
	       }
	  }
     }
     return W;
}

static const tw_instr twinstr[] = {
     {TW_FULL, 0, 32},
     {TW_NEXT, 1, 0}
};

static const hc2hc_desc desc = { 32, "hb_32", twinstr, &GENUS, {340, 114, 94, 0}, 0, 0, 0 };

void X(codelet_hb_32) (planner *p) {
     X(khc2hc_register) (p, hb_32, &desc);
}
#endif				/* HAVE_FMA */