hf2_32.c

快速fft变换

			 T48 = T1J - T1U;
			 T4b = T49 - T4a;
			 T4c = T48 + T4b;
			 T5B = T48 - T4b;
		    }
	       }
	       {
		    E T2B, T4m, T2G, T4n, T4l, T4o, T2M, T4q, T2P, T4r, T4s, T4t;
		    {
			 E T2z, T2A, T2D, T2F;
			 T2z = cr[WS(rs, 5)];
			 T2A = ci[WS(rs, 5)];
			 T2B = FMA(T21, T2z, T22 * T2A);
			 T4m = FNMS(T22, T2z, T21 * T2A);
			 T2D = cr[WS(rs, 21)];
			 T2F = ci[WS(rs, 21)];
			 T2G = FMA(T2C, T2D, T2E * T2F);
			 T4n = FNMS(T2E, T2D, T2C * T2F);
		    }
		    T4l = T2B - T2G;
		    T4o = T4m - T4n;
		    {
			 E T2J, T2L, T2N, T2O;
			 T2J = cr[WS(rs, 29)];
			 T2L = ci[WS(rs, 29)];
			 T2M = FMA(T2I, T2J, T2K * T2L);
			 T4q = FNMS(T2K, T2J, T2I * T2L);
			 T2N = cr[WS(rs, 13)];
			 T2O = ci[WS(rs, 13)];
			 T2P = FMA(T1M, T2N, T1P * T2O);
			 T4r = FNMS(T1P, T2N, T1M * T2O);
		    }
		    T4s = T4q - T4r;
		    T4t = T2M - T2P;
		    {
			 E T2H, T2Q, T6C, T6D;
			 T2H = T2B + T2G;
			 T2Q = T2M + T2P;
			 T2R = T2H + T2Q;
			 T6z = T2H - T2Q;
			 T6C = T4q + T4r;
			 T6D = T4m + T4n;
			 T6E = T6C - T6D;
			 T7k = T6D + T6C;
		    }
		    {
			 E T4p, T4u, T4C, T4D;
			 T4p = T4l + T4o;
			 T4u = T4s - T4t;
			 T4v = KP707106781 * (T4p - T4u);
			 T5K = KP707106781 * (T4p + T4u);
			 T4C = T4t + T4s;
			 T4D = T4l - T4o;
			 T4E = KP707106781 * (T4C - T4D);
			 T5H = KP707106781 * (T4D + T4C);
		    }
	       }
	       {
		    E T3k, T4S, T3p, T4T, T4R, T4U, T3t, T4N, T3w, T4O, T4M, T4P;
		    {
			 E T3i, T3j, T3m, T3o;
			 T3i = cr[WS(rs, 3)];
			 T3j = ci[WS(rs, 3)];
			 T3k = FMA(T3, T3i, T6 * T3j);
			 T4S = FNMS(T6, T3i, T3 * T3j);
			 T3m = cr[WS(rs, 19)];
			 T3o = ci[WS(rs, 19)];
			 T3p = FMA(T3l, T3m, T3n * T3o);
			 T4T = FNMS(T3n, T3m, T3l * T3o);
		    }
		    T4R = T3k - T3p;
		    T4U = T4S - T4T;
		    {
			 E T3r, T3s, T3u, T3v;
			 T3r = cr[WS(rs, 27)];
			 T3s = ci[WS(rs, 27)];
			 T3t = FMA(Th, T3r, Tl * T3s);
			 T4N = FNMS(Tl, T3r, Th * T3s);
			 T3u = cr[WS(rs, 11)];
			 T3v = ci[WS(rs, 11)];
			 T3w = FMA(Tg, T3u, Tk * T3v);
			 T4O = FNMS(Tk, T3u, Tg * T3v);
		    }
		    T4M = T3t - T3w;
		    T4P = T4N - T4O;
		    {
			 E T3q, T3x, T6I, T6J;
			 T3q = T3k + T3p;
			 T3x = T3t + T3w;
			 T3y = T3q + T3x;
			 T6P = T3q - T3x;
			 T6I = T4N + T4O;
			 T6J = T4S + T4T;
			 T6K = T6I - T6J;
			 T7p = T6J + T6I;
		    }
		    {
			 E T4Q, T4V, T53, T54;
			 T4Q = T4M + T4P;
			 T4V = T4R - T4U;
			 T4W = KP707106781 * (T4Q - T4V);
			 T5O = KP707106781 * (T4V + T4Q);
			 T53 = T4R + T4U;
			 T54 = T4P - T4M;
			 T55 = KP707106781 * (T53 - T54);
			 T5R = KP707106781 * (T53 + T54);
		    }
	       }
	       {
		    E T2b, T7x, T7K, T7M, T3A, T7L, T7A, T7B;
		    {
			 E T1j, T2a, T7C, T7J;
			 T1j = TL + T1i;
			 T2a = T1E + T29;
			 T2b = T1j + T2a;
			 T7x = T1j - T2a;
			 T7C = T7f + T7e;
			 T7J = T7D + T7I;
			 T7K = T7C + T7J;
			 T7M = T7J - T7C;
		    }
		    {
			 E T2S, T3z, T7y, T7z;
			 T2S = T2y + T2R;
			 T3z = T3h + T3y;
			 T3A = T2S + T3z;
			 T7L = T3z - T2S;
			 T7y = T7o + T7p;
			 T7z = T7j + T7k;
			 T7A = T7y - T7z;
			 T7B = T7z + T7y;
		    }
		    ci[WS(rs, 15)] = T2b - T3A;
		    cr[WS(rs, 24)] = T7L - T7M;
		    ci[WS(rs, 23)] = T7L + T7M;
		    cr[0] = T2b + T3A;
		    cr[WS(rs, 8)] = T7x - T7A;
		    cr[WS(rs, 16)] = T7B - T7K;
		    ci[WS(rs, 31)] = T7B + T7K;
		    ci[WS(rs, 7)] = T7x + T7A;
	       }
	       {
		    E T5x, T5Z, T8d, T8j, T5E, T88, T69, T6d, T5M, T5W, T62, T8i, T66, T6c, T5T;
		    E T5X, T5w, T89;
		    T5w = KP707106781 * (T5u + T5v);
		    T5x = T5t - T5w;
		    T5Z = T5t + T5w;
		    T89 = KP707106781 * (T3K - T3P);
		    T8d = T89 + T8c;
		    T8j = T8c - T89;
		    {
			 E T5A, T5D, T67, T68;
			 T5A = FMA(KP923879532, T5y, KP382683432 * T5z);
			 T5D = FNMS(KP923879532, T5C, KP382683432 * T5B);
			 T5E = T5A + T5D;
			 T88 = T5A - T5D;
			 T67 = T5N + T5O;
			 T68 = T5Q + T5R;
			 T69 = FNMS(KP980785280, T68, KP195090322 * T67);
			 T6d = FMA(KP980785280, T67, KP195090322 * T68);
		    }
		    {
			 E T5I, T5L, T60, T61;
			 T5I = T5G - T5H;
			 T5L = T5J - T5K;
			 T5M = FMA(KP831469612, T5I, KP555570233 * T5L);
			 T5W = FNMS(KP831469612, T5L, KP555570233 * T5I);
			 T60 = FNMS(KP382683432, T5y, KP923879532 * T5z);
			 T61 = FMA(KP382683432, T5C, KP923879532 * T5B);
			 T62 = T60 + T61;
			 T8i = T61 - T60;
		    }
		    {
			 E T64, T65, T5P, T5S;
			 T64 = T5G + T5H;
			 T65 = T5J + T5K;
			 T66 = FMA(KP195090322, T64, KP980785280 * T65);
			 T6c = FNMS(KP195090322, T65, KP980785280 * T64);
			 T5P = T5N - T5O;
			 T5S = T5Q - T5R;
			 T5T = FNMS(KP555570233, T5S, KP831469612 * T5P);
			 T5X = FMA(KP555570233, T5P, KP831469612 * T5S);
		    }
		    {
			 E T5F, T5U, T8h, T8k;
			 T5F = T5x + T5E;
			 T5U = T5M + T5T;
			 ci[WS(rs, 12)] = T5F - T5U;
			 cr[WS(rs, 3)] = T5F + T5U;
			 T8h = T5X - T5W;
			 T8k = T8i + T8j;
			 cr[WS(rs, 19)] = T8h - T8k;
			 ci[WS(rs, 28)] = T8h + T8k;
		    }
		    {
			 E T8l, T8m, T5V, T5Y;
			 T8l = T5T - T5M;
			 T8m = T8j - T8i;
			 cr[WS(rs, 27)] = T8l - T8m;
			 ci[WS(rs, 20)] = T8l + T8m;
			 T5V = T5x - T5E;
			 T5Y = T5W + T5X;
			 cr[WS(rs, 11)] = T5V - T5Y;
			 ci[WS(rs, 4)] = T5V + T5Y;
		    }
		    {
			 E T63, T6a, T87, T8e;
			 T63 = T5Z - T62;
			 T6a = T66 + T69;
			 ci[WS(rs, 8)] = T63 - T6a;
			 cr[WS(rs, 7)] = T63 + T6a;
			 T87 = T69 - T66;
			 T8e = T88 + T8d;
			 cr[WS(rs, 31)] = T87 - T8e;
			 ci[WS(rs, 16)] = T87 + T8e;
		    }
		    {
			 E T8f, T8g, T6b, T6e;
			 T8f = T6d - T6c;
			 T8g = T8d - T88;
			 cr[WS(rs, 23)] = T8f - T8g;
			 ci[WS(rs, 24)] = T8f + T8g;
			 T6b = T5Z + T62;
			 T6e = T6c + T6d;
			 cr[WS(rs, 15)] = T6b - T6e;
			 ci[0] = T6b + T6e;
		    }
	       }
	       {
		    E T7h, T7t, T7Q, T7S, T7m, T7u, T7r, T7v;
		    {
			 E T7d, T7g, T7O, T7P;
			 T7d = TL - T1i;
			 T7g = T7e - T7f;
			 T7h = T7d - T7g;
			 T7t = T7d + T7g;
			 T7O = T1E - T29;
			 T7P = T7I - T7D;
			 T7Q = T7O + T7P;
			 T7S = T7P - T7O;
		    }
		    {
			 E T7i, T7l, T7n, T7q;
			 T7i = T2y - T2R;
			 T7l = T7j - T7k;
			 T7m = T7i + T7l;
			 T7u = T7i - T7l;
			 T7n = T3h - T3y;
			 T7q = T7o - T7p;
			 T7r = T7n - T7q;
			 T7v = T7n + T7q;
		    }
		    {
			 E T7s, T7R, T7w, T7N;
			 T7s = KP707106781 * (T7m + T7r);
			 ci[WS(rs, 11)] = T7h - T7s;
			 cr[WS(rs, 4)] = T7h + T7s;
			 T7R = KP707106781 * (T7v - T7u);
			 cr[WS(rs, 20)] = T7R - T7S;
			 ci[WS(rs, 27)] = T7R + T7S;
			 T7w = KP707106781 * (T7u + T7v);
			 cr[WS(rs, 12)] = T7t - T7w;
			 ci[WS(rs, 3)] = T7t + T7w;
			 T7N = KP707106781 * (T7r - T7m);
			 cr[WS(rs, 28)] = T7N - T7Q;
			 ci[WS(rs, 19)] = T7N + T7Q;
		    }
	       }
	       {
		    E T6j, T7X, T83, T6X, T6u, T7U, T77, T7b, T70, T82, T6G, T6U, T74, T7a, T6R;
		    E T6V;
		    {
			 E T6o, T6t, T6A, T6F;
			 T6j = T6f - T6i;
			 T7X = T7V + T7W;
			 T83 = T7W - T7V;
			 T6X = T6f + T6i;
			 T6o = T6k + T6n;
			 T6t = T6p - T6s;
			 T6u = KP707106781 * (T6o + T6t);
			 T7U = KP707106781 * (T6o - T6t);
			 {
			      E T75, T76, T6Y, T6Z;
			      T75 = T6O + T6P;
			      T76 = T6H + T6K;
			      T77 = FMA(KP382683432, T75, KP923879532 * T76);
			      T7b = FNMS(KP923879532, T75, KP382683432 * T76);
			      T6Y = T6k - T6n;
			      T6Z = T6p + T6s;
			      T70 = KP707106781 * (T6Y + T6Z);
			      T82 = KP707106781 * (T6Z - T6Y);
			 }
			 T6A = T6y - T6z;
			 T6F = T6B - T6E;
			 T6G = FMA(KP382683432, T6A, KP923879532 * T6F);
			 T6U = FNMS(KP923879532, T6A, KP382683432 * T6F);
			 {
			      E T72, T73, T6L, T6Q;
			      T72 = T6B + T6E;
			      T73 = T6y + T6z;
			      T74 = FNMS(KP382683432, T73, KP923879532 * T72);
			      T7a = FMA(KP923879532, T73, KP382683432 * T72);
			      T6L = T6H - T6K;
			      T6Q = T6O - T6P;
			      T6R = FNMS(KP382683432, T6Q, KP923879532 * T6L);
			      T6V = FMA(KP923879532, T6Q, KP382683432 * T6L);
			 }
		    }
		    {
			 E T6v, T6S, T81, T84;
			 T6v = T6j + T6u;
			 T6S = T6G + T6R;
			 ci[WS(rs, 13)] = T6v - T6S;
			 cr[WS(rs, 2)] = T6v + T6S;
			 T81 = T6V - T6U;
			 T84 = T82 + T83;
			 cr[WS(rs, 18)] = T81 - T84;
			 ci[WS(rs, 29)] = T81 + T84;
		    }
		    {
			 E T85, T86, T6T, T6W;
			 T85 = T6R - T6G;
			 T86 = T83 - T82;
			 cr[WS(rs, 26)] = T85 - T86;
			 ci[WS(rs, 21)] = T85 + T86;
			 T6T = T6j - T6u;
			 T6W = T6U + T6V;
			 cr[WS(rs, 10)] = T6T - T6W;
			 ci[WS(rs, 5)] = T6T + T6W;
		    }
		    {
			 E T71, T78, T7T, T7Y;
			 T71 = T6X + T70;
			 T78 = T74 + T77;
			 cr[WS(rs, 14)] = T71 - T78;
			 ci[WS(rs, 1)] = T71 + T78;
			 T7T = T7b - T7a;
			 T7Y = T7U + T7X;
			 cr[WS(rs, 30)] = T7T - T7Y;
			 ci[WS(rs, 17)] = T7T + T7Y;
		    }
		    {
			 E T7Z, T80, T79, T7c;
			 T7Z = T77 - T74;
			 T80 = T7X - T7U;
			 cr[WS(rs, 22)] = T7Z - T80;
			 ci[WS(rs, 25)] = T7Z + T80;
			 T79 = T6X - T70;
			 T7c = T7a + T7b;
			 ci[WS(rs, 9)] = T79 - T7c;
			 cr[WS(rs, 6)] = T79 + T7c;
		    }
	       }
	       {
		    E T3R, T5d, T8r, T8x, T4e, T8o, T5n, T5r, T4G, T5a, T5g, T8w, T5k, T5q, T57;
		    E T5b, T3Q, T8p;
		    T3Q = KP707106781 * (T3K + T3P);
		    T3R = T3F - T3Q;
		    T5d = T3F + T3Q;
		    T8p = KP707106781 * (T5v - T5u);
		    T8r = T8p + T8q;
		    T8x = T8q - T8p;
		    {
			 E T42, T4d, T5l, T5m;
			 T42 = FNMS(KP923879532, T41, KP382683432 * T3W);
			 T4d = FMA(KP923879532, T47, KP382683432 * T4c);
			 T4e = T42 + T4d;
			 T8o = T4d - T42;
			 T5l = T52 + T55;
			 T5m = T4L + T4W;
			 T5n = FNMS(KP195090322, T5m, KP980785280 * T5l);
			 T5r = FMA(KP980785280, T5m, KP195090322 * T5l);
		    }
		    {
			 E T4w, T4F, T5e, T5f;
			 T4w = T4k - T4v;
			 T4F = T4B - T4E;
			 T4G = FNMS(KP555570233, T4F, KP831469612 * T4w);
			 T5a = FMA(KP831469612, T4F, KP555570233 * T4w);
			 T5e = FMA(KP382683432, T41, KP923879532 * T3W);
			 T5f = FNMS(KP382683432, T47, KP923879532 * T4c);
			 T5g = T5e + T5f;
			 T8w = T5e - T5f;
		    }
		    {
			 E T5i, T5j, T4X, T56;
			 T5i = T4B + T4E;
			 T5j = T4k + T4v;
			 T5k = FMA(KP195090322, T5i, KP980785280 * T5j);
			 T5q = FNMS(KP980785280, T5i, KP195090322 * T5j);
			 T4X = T4L - T4W;
			 T56 = T52 - T55;
			 T57 = FMA(KP555570233, T4X, KP831469612 * T56);
			 T5b = FNMS(KP831469612, T4X, KP555570233 * T56);
		    }
		    {
			 E T4f, T58, T8v, T8y;
			 T4f = T3R + T4e;
			 T58 = T4G + T57;
			 cr[WS(rs, 13)] = T4f - T58;
			 ci[WS(rs, 2)] = T4f + T58;
			 T8v = T5b - T5a;
			 T8y = T8w + T8x;
			 cr[WS(rs, 29)] = T8v - T8y;
			 ci[WS(rs, 18)] = T8v + T8y;
		    }
		    {
			 E T8z, T8A, T59, T5c;
			 T8z = T57 - T4G;
			 T8A = T8x - T8w;
			 cr[WS(rs, 21)] = T8z - T8A;
			 ci[WS(rs, 26)] = T8z + T8A;
			 T59 = T3R - T4e;
			 T5c = T5a + T5b;
			 ci[WS(rs, 10)] = T59 - T5c;
			 cr[WS(rs, 5)] = T59 + T5c;
		    }
		    {
			 E T5h, T5o, T8n, T8s;
			 T5h = T5d + T5g;
			 T5o = T5k + T5n;
			 ci[WS(rs, 14)] = T5h - T5o;
			 cr[WS(rs, 1)] = T5h + T5o;
			 T8n = T5r - T5q;
			 T8s = T8o + T8r;
			 cr[WS(rs, 17)] = T8n - T8s;
			 ci[WS(rs, 30)] = T8n + T8s;
		    }
		    {
			 E T8t, T8u, T5p, T5s;
			 T8t = T5n - T5k;
			 T8u = T8r - T8o;
			 cr[WS(rs, 25)] = T8t - T8u;
			 ci[WS(rs, 22)] = T8t + T8u;
			 T5p = T5d - T5g;
			 T5s = T5q + T5r;
			 cr[WS(rs, 9)] = T5p - T5s;
			 ci[WS(rs, 6)] = T5p + T5s;
		    }
	       }
	  }
     }
}

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, "hf2_32", twinstr, &GENUS, {376, 168, 112, 0} };

void X(codelet_hf2_32) (planner *p) {
     X(khc2hc_register) (p, hf2_32, &desc);
}
#endif				/* HAVE_FMA */