mr2hc_64.c

fftw-3.0.1

		    T4W = T2u + T2v;
	       }
	       {
		    E T2n, T2o, T2r, T2s;
		    T2n = I[WS(is, 19)];
		    T2o = I[WS(is, 51)];
		    T2p = T2n - T2o;
		    T4T = T2n + T2o;
		    T2r = I[WS(is, 59)];
		    T2s = I[WS(is, 27)];
		    T2t = T2r - T2s;
		    T4V = T2r + T2s;
	       }
	       {
		    E T2q, T2x, T5Q, T5R;
		    T2q = FNMS(KP382683432, T2p, KP923879532 * T2m);
		    T2x = FMA(KP923879532, T2t, KP382683432 * T2w);
		    T2y = T2q + T2x;
		    T3J = T2x - T2q;
		    T5Q = T4S + T4T;
		    T5R = T4V + T4W;
		    T5S = T5Q + T5R;
		    T6a = T5R - T5Q;
	       }
	       {
		    E T2A, T2B, T4U, T4X;
		    T2A = FNMS(KP923879532, T2w, KP382683432 * T2t);
		    T2B = FMA(KP382683432, T2m, KP923879532 * T2p);
		    T2C = T2A - T2B;
		    T3G = T2B + T2A;
		    T4U = T4S - T4T;
		    T4X = T4V - T4W;
		    T4Y = KP707106781 * (T4U + T4X);
		    T53 = KP707106781 * (T4X - T4U);
	       }
	  }
	  {
	       E Tv, T10, T5X, T5Y, T5Z, T60;
	       Tv = Tf + Tu;
	       T10 = TK + TZ;
	       T5X = Tv + T10;
	       T5Y = T5I + T5L;
	       T5Z = T5P + T5S;
	       T60 = T5Y + T5Z;
	       ro[WS(ros, 16)] = Tv - T10;
	       io[WS(ios, 16)] = T5Z - T5Y;
	       ro[WS(ros, 32)] = T5X - T60;
	       ro[0] = T5X + T60;
	  }
	  {
	       E T5F, T5V, T5U, T5W, T5M, T5T;
	       T5F = Tf - Tu;
	       T5V = TZ - TK;
	       T5M = T5I - T5L;
	       T5T = T5P - T5S;
	       T5U = KP707106781 * (T5M + T5T);
	       T5W = KP707106781 * (T5T - T5M);
	       ro[WS(ros, 24)] = T5F - T5U;
	       io[WS(ios, 24)] = T5W - T5V;
	       ro[WS(ros, 8)] = T5F + T5U;
	       io[WS(ios, 8)] = T5V + T5W;
	  }
	  {
	       E T65, T6l, T6k, T6m, T6c, T6g, T6f, T6h;
	       {
		    E T61, T64, T6i, T6j;
		    T61 = T7 - Te;
		    T64 = KP707106781 * (T62 + T63);
		    T65 = T61 + T64;
		    T6l = T61 - T64;
		    T6i = FNMS(KP382683432, T66, KP923879532 * T67);
		    T6j = FMA(KP382683432, T69, KP923879532 * T6a);
		    T6k = T6i + T6j;
		    T6m = T6j - T6i;
	       }
	       {
		    E T68, T6b, T6d, T6e;
		    T68 = FMA(KP923879532, T66, KP382683432 * T67);
		    T6b = FNMS(KP382683432, T6a, KP923879532 * T69);
		    T6c = T68 + T6b;
		    T6g = T6b - T68;
		    T6d = KP707106781 * (T63 - T62);
		    T6e = Tt - Tm;
		    T6f = T6d - T6e;
		    T6h = T6e + T6d;
	       }
	       ro[WS(ros, 28)] = T65 - T6c;
	       io[WS(ios, 28)] = T6k - T6h;
	       ro[WS(ros, 4)] = T65 + T6c;
	       io[WS(ios, 4)] = T6h + T6k;
	       io[WS(ios, 12)] = T6f + T6g;
	       ro[WS(ros, 12)] = T6l + T6m;
	       io[WS(ios, 20)] = T6g - T6f;
	       ro[WS(ros, 20)] = T6l - T6m;
	  }
	  {
	       E T5n, T5D, T5x, T5z, T5q, T5A, T5t, T5B;
	       {
		    E T5l, T5m, T5v, T5w;
		    T5l = T4l - T4o;
		    T5m = T58 - T57;
		    T5n = T5l + T5m;
		    T5D = T5l - T5m;
		    T5v = T4v - T4s;
		    T5w = T5b - T5a;
		    T5x = T5v - T5w;
		    T5z = T5w + T5v;
	       }
	       {
		    E T5o, T5p, T5r, T5s;
		    T5o = T4A - T4H;
		    T5p = T4M - T4L;
		    T5q = FMA(KP831469612, T5o, KP555570233 * T5p);
		    T5A = FNMS(KP555570233, T5o, KP831469612 * T5p);
		    T5r = T4R - T4Y;
		    T5s = T53 - T52;
		    T5t = FNMS(KP555570233, T5s, KP831469612 * T5r);
		    T5B = FMA(KP555570233, T5r, KP831469612 * T5s);
	       }
	       {
		    E T5u, T5C, T5y, T5E;
		    T5u = T5q + T5t;
		    ro[WS(ros, 26)] = T5n - T5u;
		    ro[WS(ros, 6)] = T5n + T5u;
		    T5C = T5A + T5B;
		    io[WS(ios, 6)] = T5z + T5C;
		    io[WS(ios, 26)] = T5C - T5z;
		    T5y = T5t - T5q;
		    io[WS(ios, 10)] = T5x + T5y;
		    io[WS(ios, 22)] = T5y - T5x;
		    T5E = T5B - T5A;
		    ro[WS(ros, 22)] = T5D - T5E;
		    ro[WS(ros, 10)] = T5D + T5E;
	       }
	  }
	  {
	       E T4x, T5j, T5d, T5f, T4O, T5g, T55, T5h;
	       {
		    E T4p, T4w, T59, T5c;
		    T4p = T4l + T4o;
		    T4w = T4s + T4v;
		    T4x = T4p + T4w;
		    T5j = T4p - T4w;
		    T59 = T57 + T58;
		    T5c = T5a + T5b;
		    T5d = T59 - T5c;
		    T5f = T5c + T59;
	       }
	       {
		    E T4I, T4N, T4Z, T54;
		    T4I = T4A + T4H;
		    T4N = T4L + T4M;
		    T4O = FMA(KP980785280, T4I, KP195090322 * T4N);
		    T5g = FNMS(KP195090322, T4I, KP980785280 * T4N);
		    T4Z = T4R + T4Y;
		    T54 = T52 + T53;
		    T55 = FNMS(KP195090322, T54, KP980785280 * T4Z);
		    T5h = FMA(KP195090322, T4Z, KP980785280 * T54);
	       }
	       {
		    E T56, T5i, T5e, T5k;
		    T56 = T4O + T55;
		    ro[WS(ros, 30)] = T4x - T56;
		    ro[WS(ros, 2)] = T4x + T56;
		    T5i = T5g + T5h;
		    io[WS(ios, 2)] = T5f + T5i;
		    io[WS(ios, 30)] = T5i - T5f;
		    T5e = T55 - T4O;
		    io[WS(ios, 14)] = T5d + T5e;
		    io[WS(ios, 18)] = T5e - T5d;
		    T5k = T5h - T5g;
		    ro[WS(ros, 18)] = T5j - T5k;
		    ro[WS(ros, 14)] = T5j + T5k;
	       }
	  }
	  {
	       E T3p, T41, T4c, T3S, T3w, T4b, T49, T4h, T3P, T42, T3E, T3W, T46, T4g, T3L;
	       E T3X;
	       {
		    E T3s, T3v, T3A, T3D;
		    T3p = T3n + T3o;
		    T41 = T3n - T3o;
		    T4c = T3R - T3Q;
		    T3S = T3Q + T3R;
		    T3s = FMA(KP831469612, T3q, KP555570233 * T3r);
		    T3v = FNMS(KP555570233, T3u, KP831469612 * T3t);
		    T3w = T3s + T3v;
		    T4b = T3v - T3s;
		    {
			 E T47, T48, T3N, T3O;
			 T47 = T3F - T3G;
			 T48 = T3J - T3I;
			 T49 = FNMS(KP471396736, T48, KP881921264 * T47);
			 T4h = FMA(KP471396736, T47, KP881921264 * T48);
			 T3N = FNMS(KP555570233, T3q, KP831469612 * T3r);
			 T3O = FMA(KP555570233, T3t, KP831469612 * T3u);
			 T3P = T3N + T3O;
			 T42 = T3O - T3N;
		    }
		    T3A = T3y + T3z;
		    T3D = T3B + T3C;
		    T3E = FMA(KP956940335, T3A, KP290284677 * T3D);
		    T3W = FNMS(KP290284677, T3A, KP956940335 * T3D);
		    {
			 E T44, T45, T3H, T3K;
			 T44 = T3y - T3z;
			 T45 = T3C - T3B;
			 T46 = FMA(KP881921264, T44, KP471396736 * T45);
			 T4g = FNMS(KP471396736, T44, KP881921264 * T45);
			 T3H = T3F + T3G;
			 T3K = T3I + T3J;
			 T3L = FNMS(KP290284677, T3K, KP956940335 * T3H);
			 T3X = FMA(KP290284677, T3H, KP956940335 * T3K);
		    }
	       }
	       {
		    E T3x, T3M, T3V, T3Y;
		    T3x = T3p + T3w;
		    T3M = T3E + T3L;
		    ro[WS(ros, 29)] = T3x - T3M;
		    ro[WS(ros, 3)] = T3x + T3M;
		    T3V = T3S + T3P;
		    T3Y = T3W + T3X;
		    io[WS(ios, 3)] = T3V + T3Y;
		    io[WS(ios, 29)] = T3Y - T3V;
	       }
	       {
		    E T3T, T3U, T3Z, T40;
		    T3T = T3P - T3S;
		    T3U = T3L - T3E;
		    io[WS(ios, 13)] = T3T + T3U;
		    io[WS(ios, 19)] = T3U - T3T;
		    T3Z = T3p - T3w;
		    T40 = T3X - T3W;
		    ro[WS(ros, 19)] = T3Z - T40;
		    ro[WS(ros, 13)] = T3Z + T40;
	       }
	       {
		    E T43, T4a, T4f, T4i;
		    T43 = T41 + T42;
		    T4a = T46 + T49;
		    ro[WS(ros, 27)] = T43 - T4a;
		    ro[WS(ros, 5)] = T43 + T4a;
		    T4f = T4c + T4b;
		    T4i = T4g + T4h;
		    io[WS(ios, 5)] = T4f + T4i;
		    io[WS(ios, 27)] = T4i - T4f;
	       }
	       {
		    E T4d, T4e, T4j, T4k;
		    T4d = T4b - T4c;
		    T4e = T49 - T46;
		    io[WS(ios, 11)] = T4d + T4e;
		    io[WS(ios, 21)] = T4e - T4d;
		    T4j = T41 - T42;
		    T4k = T4h - T4g;
		    ro[WS(ros, 21)] = T4j - T4k;
		    ro[WS(ros, 11)] = T4j + T4k;
	       }
	  }
	  {
	       E T1d, T33, T3e, T2U, T1w, T3d, T3b, T3j, T2N, T34, T28, T2Y, T38, T3i, T2J;
	       E T2Z;
	       {
		    E T1m, T1v, T1Y, T27;
		    T1d = T15 - T1c;
		    T33 = T15 + T1c;
		    T3e = T2T + T2Q;
		    T2U = T2Q - T2T;
		    T1m = FMA(KP195090322, T1i, KP980785280 * T1l);
		    T1v = FNMS(KP195090322, T1u, KP980785280 * T1r);
		    T1w = T1m - T1v;
		    T3d = T1v + T1m;
		    {
			 E T39, T3a, T2L, T2M;
			 T39 = T2j + T2y;
			 T3a = T2H + T2C;
			 T3b = FNMS(KP098017140, T3a, KP995184726 * T39);
			 T3j = FMA(KP995184726, T3a, KP098017140 * T39);
			 T2L = FNMS(KP195090322, T1l, KP980785280 * T1i);
			 T2M = FMA(KP980785280, T1u, KP195090322 * T1r);
			 T2N = T2L - T2M;
			 T34 = T2M + T2L;
		    }
		    T1Y = T1M - T1X;
		    T27 = T23 - T26;
		    T28 = FMA(KP634393284, T1Y, KP773010453 * T27);
		    T2Y = FNMS(KP634393284, T27, KP773010453 * T1Y);
		    {
			 E T36, T37, T2z, T2I;
			 T36 = T1X + T1M;
			 T37 = T23 + T26;
			 T38 = FMA(KP098017140, T36, KP995184726 * T37);
			 T3i = FNMS(KP098017140, T37, KP995184726 * T36);
			 T2z = T2j - T2y;
			 T2I = T2C - T2H;
			 T2J = FNMS(KP634393284, T2I, KP773010453 * T2z);
			 T2Z = FMA(KP773010453, T2I, KP634393284 * T2z);
		    }
	       }
	       {
		    E T1x, T2K, T2X, T30;
		    T1x = T1d + T1w;
		    T2K = T28 + T2J;
		    ro[WS(ros, 25)] = T1x - T2K;
		    ro[WS(ros, 7)] = T1x + T2K;
		    T2X = T2U + T2N;
		    T30 = T2Y + T2Z;
		    io[WS(ios, 7)] = T2X + T30;
		    io[WS(ios, 25)] = T30 - T2X;
	       }
	       {
		    E T2V, T2W, T31, T32;
		    T2V = T2N - T2U;
		    T2W = T2J - T28;
		    io[WS(ios, 9)] = T2V + T2W;
		    io[WS(ios, 23)] = T2W - T2V;
		    T31 = T1d - T1w;
		    T32 = T2Z - T2Y;
		    ro[WS(ros, 23)] = T31 - T32;
		    ro[WS(ros, 9)] = T31 + T32;
	       }
	       {
		    E T35, T3c, T3h, T3k;
		    T35 = T33 + T34;
		    T3c = T38 + T3b;
		    ro[WS(ros, 31)] = T35 - T3c;
		    ro[WS(ros, 1)] = T35 + T3c;
		    T3h = T3e + T3d;
		    T3k = T3i + T3j;
		    io[WS(ios, 1)] = T3h + T3k;
		    io[WS(ios, 31)] = T3k - T3h;
	       }
	       {
		    E T3f, T3g, T3l, T3m;
		    T3f = T3d - T3e;
		    T3g = T3b - T38;
		    io[WS(ios, 15)] = T3f + T3g;
		    io[WS(ios, 17)] = T3g - T3f;
		    T3l = T33 - T34;
		    T3m = T3j - T3i;
		    ro[WS(ros, 17)] = T3l - T3m;
		    ro[WS(ros, 15)] = T3l + T3m;
	       }
	  }
     }
}

static void mr2hc_64(const R *I, R *ro, R *io, stride is, stride ros, stride ios, int v, int ivs, int ovs)
{
     int i;
     for (i = v; i > 0; --i) {
	  mr2hc_64_0(I, ro, io, is, ros, ios);
	  I += ivs;
	  ro += ovs;
	  io += ovs;
     }
}

static const kr2hc_desc desc = { 64, "mr2hc_64", {342, 72, 52, 0}, &GENUS, 0, 0, 0, 0, 0 };

void X(codelet_mr2hc_64) (planner *p) {
     X(kr2hc_register) (p, mr2hc_64, &desc);
}