hf_32.c

快速fft变换

		    T2b = FMA(T27, T28, T29 * T2a);
		    T3R = FNMS(T29, T28, T27 * T2a);
	       }
	       T3S = T3Q - T3R;
	       T3T = T26 - T2b;
	       {
		    E T21, T2c, T62, T63;
		    T21 = T1V + T20;
		    T2c = T26 + T2b;
		    T2d = T21 + T2c;
		    T5Z = T21 - T2c;
		    T62 = T3Q + T3R;
		    T63 = T3M + T3N;
		    T64 = T62 - T63;
		    T6K = T63 + T62;
	       }
	       {
		    E T3P, T3U, T42, T43;
		    T3P = T3L + T3O;
		    T3U = T3S - T3T;
		    T3V = KP707106781 * (T3P - T3U);
		    T5a = KP707106781 * (T3P + T3U);
		    T42 = T3T + T3S;
		    T43 = T3L - T3O;
		    T44 = KP707106781 * (T42 - T43);
		    T57 = KP707106781 * (T43 + T42);
	       }
	  }
	  {
	       E T2G, T4i, T2L, T4j, T4h, T4k, T2R, T4d, T2W, T4e, T4c, T4f;
	       {
		    E T2D, T2F, T2C, T2E;
		    T2D = cr[WS(rs, 3)];
		    T2F = ci[WS(rs, 3)];
		    T2C = W[4];
		    T2E = W[5];
		    T2G = FMA(T2C, T2D, T2E * T2F);
		    T4i = FNMS(T2E, T2D, T2C * T2F);
	       }
	       {
		    E T2I, T2K, T2H, T2J;
		    T2I = cr[WS(rs, 19)];
		    T2K = ci[WS(rs, 19)];
		    T2H = W[36];
		    T2J = W[37];
		    T2L = FMA(T2H, T2I, T2J * T2K);
		    T4j = FNMS(T2J, T2I, T2H * T2K);
	       }
	       T4h = T2G - T2L;
	       T4k = T4i - T4j;
	       {
		    E T2O, T2Q, T2N, T2P;
		    T2O = cr[WS(rs, 27)];
		    T2Q = ci[WS(rs, 27)];
		    T2N = W[52];
		    T2P = W[53];
		    T2R = FMA(T2N, T2O, T2P * T2Q);
		    T4d = FNMS(T2P, T2O, T2N * T2Q);
	       }
	       {
		    E T2T, T2V, T2S, T2U;
		    T2T = cr[WS(rs, 11)];
		    T2V = ci[WS(rs, 11)];
		    T2S = W[20];
		    T2U = W[21];
		    T2W = FMA(T2S, T2T, T2U * T2V);
		    T4e = FNMS(T2U, T2T, T2S * T2V);
	       }
	       T4c = T2R - T2W;
	       T4f = T4d - T4e;
	       {
		    E T2M, T2X, T68, T69;
		    T2M = T2G + T2L;
		    T2X = T2R + T2W;
		    T2Y = T2M + T2X;
		    T6f = T2M - T2X;
		    T68 = T4d + T4e;
		    T69 = T4i + T4j;
		    T6a = T68 - T69;
		    T6P = T69 + T68;
	       }
	       {
		    E T4g, T4l, T4t, T4u;
		    T4g = T4c + T4f;
		    T4l = T4h - T4k;
		    T4m = KP707106781 * (T4g - T4l);
		    T5e = KP707106781 * (T4l + T4g);
		    T4t = T4h + T4k;
		    T4u = T4f - T4c;
		    T4v = KP707106781 * (T4t - T4u);
		    T5h = KP707106781 * (T4t + T4u);
	       }
	  }
	  {
	       E T1t, T6X, T7a, T7c, T30, T7b, T70, T71;
	       {
		    E TH, T1s, T72, T79;
		    TH = Tj + TG;
		    T1s = T14 + T1r;
		    T1t = TH + T1s;
		    T6X = TH - T1s;
		    T72 = T6F + T6E;
		    T79 = T73 + T78;
		    T7a = T72 + T79;
		    T7c = T79 - T72;
	       }
	       {
		    E T2e, T2Z, T6Y, T6Z;
		    T2e = T1Q + T2d;
		    T2Z = T2B + T2Y;
		    T30 = T2e + T2Z;
		    T7b = T2Z - T2e;
		    T6Y = T6O + T6P;
		    T6Z = T6J + T6K;
		    T70 = T6Y - T6Z;
		    T71 = T6Z + T6Y;
	       }
	       ci[WS(rs, 15)] = T1t - T30;
	       cr[WS(rs, 24)] = T7b - T7c;
	       ci[WS(rs, 23)] = T7b + T7c;
	       cr[0] = T1t + T30;
	       cr[WS(rs, 8)] = T6X - T70;
	       cr[WS(rs, 16)] = T71 - T7a;
	       ci[WS(rs, 31)] = T71 + T7a;
	       ci[WS(rs, 7)] = T6X + T70;
	  }
	  {
	       E T4X, T5p, T7D, T7J, T54, T7y, T5z, T5D, T5c, T5m, T5s, T7I, T5w, T5C, T5j;
	       E T5n, T4W, T7z;
	       T4W = KP707106781 * (T4U + T4V);
	       T4X = T4T - T4W;
	       T5p = T4T + T4W;
	       T7z = KP707106781 * (T3a - T3f);
	       T7D = T7z + T7C;
	       T7J = T7C - T7z;
	       {
		    E T50, T53, T5x, T5y;
		    T50 = FMA(KP923879532, T4Y, KP382683432 * T4Z);
		    T53 = FNMS(KP923879532, T52, KP382683432 * T51);
		    T54 = T50 + T53;
		    T7y = T50 - T53;
		    T5x = T5d + T5e;
		    T5y = T5g + T5h;
		    T5z = FNMS(KP980785280, T5y, KP195090322 * T5x);
		    T5D = FMA(KP980785280, T5x, KP195090322 * T5y);
	       }
	       {
		    E T58, T5b, T5q, T5r;
		    T58 = T56 - T57;
		    T5b = T59 - T5a;
		    T5c = FMA(KP831469612, T58, KP555570233 * T5b);
		    T5m = FNMS(KP831469612, T5b, KP555570233 * T58);
		    T5q = FNMS(KP382683432, T4Y, KP923879532 * T4Z);
		    T5r = FMA(KP382683432, T52, KP923879532 * T51);
		    T5s = T5q + T5r;
		    T7I = T5r - T5q;
	       }
	       {
		    E T5u, T5v, T5f, T5i;
		    T5u = T56 + T57;
		    T5v = T59 + T5a;
		    T5w = FMA(KP195090322, T5u, KP980785280 * T5v);
		    T5C = FNMS(KP195090322, T5v, KP980785280 * T5u);
		    T5f = T5d - T5e;
		    T5i = T5g - T5h;
		    T5j = FNMS(KP555570233, T5i, KP831469612 * T5f);
		    T5n = FMA(KP555570233, T5f, KP831469612 * T5i);
	       }
	       {
		    E T55, T5k, T7H, T7K;
		    T55 = T4X + T54;
		    T5k = T5c + T5j;
		    ci[WS(rs, 12)] = T55 - T5k;
		    cr[WS(rs, 3)] = T55 + T5k;
		    T7H = T5n - T5m;
		    T7K = T7I + T7J;
		    cr[WS(rs, 19)] = T7H - T7K;
		    ci[WS(rs, 28)] = T7H + T7K;
	       }
	       {
		    E T7L, T7M, T5l, T5o;
		    T7L = T5j - T5c;
		    T7M = T7J - T7I;
		    cr[WS(rs, 27)] = T7L - T7M;
		    ci[WS(rs, 20)] = T7L + T7M;
		    T5l = T4X - T54;
		    T5o = T5m + T5n;
		    cr[WS(rs, 11)] = T5l - T5o;
		    ci[WS(rs, 4)] = T5l + T5o;
	       }
	       {
		    E T5t, T5A, T7x, T7E;
		    T5t = T5p - T5s;
		    T5A = T5w + T5z;
		    ci[WS(rs, 8)] = T5t - T5A;
		    cr[WS(rs, 7)] = T5t + T5A;
		    T7x = T5z - T5w;
		    T7E = T7y + T7D;
		    cr[WS(rs, 31)] = T7x - T7E;
		    ci[WS(rs, 16)] = T7x + T7E;
	       }
	       {
		    E T7F, T7G, T5B, T5E;
		    T7F = T5D - T5C;
		    T7G = T7D - T7y;
		    cr[WS(rs, 23)] = T7F - T7G;
		    ci[WS(rs, 24)] = T7F + T7G;
		    T5B = T5p + T5s;
		    T5E = T5C + T5D;
		    cr[WS(rs, 15)] = T5B - T5E;
		    ci[0] = T5B + T5E;
	       }
	  }
	  {
	       E T6H, T6T, T7g, T7i, T6M, T6U, T6R, T6V;
	       {
		    E T6D, T6G, T7e, T7f;
		    T6D = Tj - TG;
		    T6G = T6E - T6F;
		    T6H = T6D - T6G;
		    T6T = T6D + T6G;
		    T7e = T14 - T1r;
		    T7f = T78 - T73;
		    T7g = T7e + T7f;
		    T7i = T7f - T7e;
	       }
	       {
		    E T6I, T6L, T6N, T6Q;
		    T6I = T1Q - T2d;
		    T6L = T6J - T6K;
		    T6M = T6I + T6L;
		    T6U = T6I - T6L;
		    T6N = T2B - T2Y;
		    T6Q = T6O - T6P;
		    T6R = T6N - T6Q;
		    T6V = T6N + T6Q;
	       }
	       {
		    E T6S, T7h, T6W, T7d;
		    T6S = KP707106781 * (T6M + T6R);
		    ci[WS(rs, 11)] = T6H - T6S;
		    cr[WS(rs, 4)] = T6H + T6S;
		    T7h = KP707106781 * (T6V - T6U);
		    cr[WS(rs, 20)] = T7h - T7i;
		    ci[WS(rs, 27)] = T7h + T7i;
		    T6W = KP707106781 * (T6U + T6V);
		    cr[WS(rs, 12)] = T6T - T6W;
		    ci[WS(rs, 3)] = T6T + T6W;
		    T7d = KP707106781 * (T6R - T6M);
		    cr[WS(rs, 28)] = T7d - T7g;
		    ci[WS(rs, 19)] = T7d + T7g;
	       }
	  }
	  {
	       E T5J, T7n, T7t, T6n, T5U, T7k, T6x, T6B, T6q, T7s, T66, T6k, T6u, T6A, T6h;
	       E T6l;
	       {
		    E T5O, T5T, T60, T65;
		    T5J = T5F - T5I;
		    T7n = T7l + T7m;
		    T7t = T7m - T7l;
		    T6n = T5F + T5I;
		    T5O = T5K + T5N;
		    T5T = T5P - T5S;
		    T5U = KP707106781 * (T5O + T5T);
		    T7k = KP707106781 * (T5O - T5T);
		    {
			 E T6v, T6w, T6o, T6p;
			 T6v = T6e + T6f;
			 T6w = T67 + T6a;
			 T6x = FMA(KP382683432, T6v, KP923879532 * T6w);
			 T6B = FNMS(KP923879532, T6v, KP382683432 * T6w);
			 T6o = T5K - T5N;
			 T6p = T5P + T5S;
			 T6q = KP707106781 * (T6o + T6p);
			 T7s = KP707106781 * (T6p - T6o);
		    }
		    T60 = T5Y - T5Z;
		    T65 = T61 - T64;
		    T66 = FMA(KP382683432, T60, KP923879532 * T65);
		    T6k = FNMS(KP923879532, T60, KP382683432 * T65);
		    {
			 E T6s, T6t, T6b, T6g;
			 T6s = T61 + T64;
			 T6t = T5Y + T5Z;
			 T6u = FNMS(KP382683432, T6t, KP923879532 * T6s);
			 T6A = FMA(KP923879532, T6t, KP382683432 * T6s);
			 T6b = T67 - T6a;
			 T6g = T6e - T6f;
			 T6h = FNMS(KP382683432, T6g, KP923879532 * T6b);
			 T6l = FMA(KP923879532, T6g, KP382683432 * T6b);
		    }
	       }
	       {
		    E T5V, T6i, T7r, T7u;
		    T5V = T5J + T5U;
		    T6i = T66 + T6h;
		    ci[WS(rs, 13)] = T5V - T6i;
		    cr[WS(rs, 2)] = T5V + T6i;
		    T7r = T6l - T6k;
		    T7u = T7s + T7t;
		    cr[WS(rs, 18)] = T7r - T7u;
		    ci[WS(rs, 29)] = T7r + T7u;
	       }
	       {
		    E T7v, T7w, T6j, T6m;
		    T7v = T6h - T66;
		    T7w = T7t - T7s;
		    cr[WS(rs, 26)] = T7v - T7w;
		    ci[WS(rs, 21)] = T7v + T7w;
		    T6j = T5J - T5U;
		    T6m = T6k + T6l;
		    cr[WS(rs, 10)] = T6j - T6m;
		    ci[WS(rs, 5)] = T6j + T6m;
	       }
	       {
		    E T6r, T6y, T7j, T7o;
		    T6r = T6n + T6q;
		    T6y = T6u + T6x;
		    cr[WS(rs, 14)] = T6r - T6y;
		    ci[WS(rs, 1)] = T6r + T6y;
		    T7j = T6B - T6A;
		    T7o = T7k + T7n;
		    cr[WS(rs, 30)] = T7j - T7o;
		    ci[WS(rs, 17)] = T7j + T7o;
	       }
	       {
		    E T7p, T7q, T6z, T6C;
		    T7p = T6x - T6u;
		    T7q = T7n - T7k;
		    cr[WS(rs, 22)] = T7p - T7q;
		    ci[WS(rs, 25)] = T7p + T7q;
		    T6z = T6n - T6q;
		    T6C = T6A + T6B;
		    ci[WS(rs, 9)] = T6z - T6C;
		    cr[WS(rs, 6)] = T6z + T6C;
	       }
	  }
	  {
	       E T3h, T4D, T7R, T7X, T3E, T7O, T4N, T4R, T46, T4A, T4G, T7W, T4K, T4Q, T4x;
	       E T4B, T3g, T7P;
	       T3g = KP707106781 * (T3a + T3f);
	       T3h = T35 - T3g;
	       T4D = T35 + T3g;
	       T7P = KP707106781 * (T4V - T4U);
	       T7R = T7P + T7Q;
	       T7X = T7Q - T7P;
	       {
		    E T3s, T3D, T4L, T4M;
		    T3s = FNMS(KP923879532, T3r, KP382683432 * T3m);
		    T3D = FMA(KP923879532, T3x, KP382683432 * T3C);
		    T3E = T3s + T3D;
		    T7O = T3D - T3s;
		    T4L = T4s + T4v;
		    T4M = T4b + T4m;
		    T4N = FNMS(KP195090322, T4M, KP980785280 * T4L);
		    T4R = FMA(KP980785280, T4M, KP195090322 * T4L);
	       }
	       {
		    E T3W, T45, T4E, T4F;
		    T3W = T3K - T3V;
		    T45 = T41 - T44;
		    T46 = FNMS(KP555570233, T45, KP831469612 * T3W);
		    T4A = FMA(KP831469612, T45, KP555570233 * T3W);
		    T4E = FMA(KP382683432, T3r, KP923879532 * T3m);
		    T4F = FNMS(KP382683432, T3x, KP923879532 * T3C);
		    T4G = T4E + T4F;
		    T7W = T4E - T4F;
	       }
	       {
		    E T4I, T4J, T4n, T4w;
		    T4I = T41 + T44;
		    T4J = T3K + T3V;
		    T4K = FMA(KP195090322, T4I, KP980785280 * T4J);
		    T4Q = FNMS(KP980785280, T4I, KP195090322 * T4J);
		    T4n = T4b - T4m;
		    T4w = T4s - T4v;
		    T4x = FMA(KP555570233, T4n, KP831469612 * T4w);
		    T4B = FNMS(KP831469612, T4n, KP555570233 * T4w);
	       }
	       {
		    E T3F, T4y, T7V, T7Y;
		    T3F = T3h + T3E;
		    T4y = T46 + T4x;
		    cr[WS(rs, 13)] = T3F - T4y;
		    ci[WS(rs, 2)] = T3F + T4y;
		    T7V = T4B - T4A;
		    T7Y = T7W + T7X;
		    cr[WS(rs, 29)] = T7V - T7Y;
		    ci[WS(rs, 18)] = T7V + T7Y;
	       }
	       {
		    E T7Z, T80, T4z, T4C;
		    T7Z = T4x - T46;
		    T80 = T7X - T7W;
		    cr[WS(rs, 21)] = T7Z - T80;
		    ci[WS(rs, 26)] = T7Z + T80;
		    T4z = T3h - T3E;
		    T4C = T4A + T4B;
		    ci[WS(rs, 10)] = T4z - T4C;
		    cr[WS(rs, 5)] = T4z + T4C;
	       }
	       {
		    E T4H, T4O, T7N, T7S;
		    T4H = T4D + T4G;
		    T4O = T4K + T4N;
		    ci[WS(rs, 14)] = T4H - T4O;
		    cr[WS(rs, 1)] = T4H + T4O;
		    T7N = T4R - T4Q;
		    T7S = T7O + T7R;
		    cr[WS(rs, 17)] = T7N - T7S;
		    ci[WS(rs, 30)] = T7N + T7S;
	       }
	       {
		    E T7T, T7U, T4P, T4S;
		    T7T = T4N - T4K;
		    T7U = T7R - T7O;
		    cr[WS(rs, 25)] = T7T - T7U;
		    ci[WS(rs, 22)] = T7T + T7U;
		    T4P = T4D - T4G;
		    T4S = T4Q + T4R;
		    cr[WS(rs, 9)] = T4P - T4S;
		    ci[WS(rs, 6)] = T4P + T4S;
	       }
	  }
     }
}

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

static const hc2hc_desc desc = { 32, "hf_32", twinstr, &GENUS, {340, 114, 94, 0} };

void X(codelet_hf_32) (planner *p) {
     X(khc2hc_register) (p, hf_32, &desc);
}
#endif				/* HAVE_FMA */