t1bv_64.c

快速fft变换

				   Tx = VADD(Tl, Tw);
				   T38 = VSUB(Tl, Tw);
				   T1I = BYTW(&(W[TWVL * 120]), T1H);
				   T1K = BYTW(&(W[TWVL * 56]), T1J);
				   T1F = VADD(T1C, T1E);
				   T5H = VSUB(T1C, T1E);
				   T2k = BYTW(&(W[TWVL * 4]), T2j);
				   T2l = LD(&(x[WS(rs, 35)]), ms, &(x[WS(rs, 1)]));
				   T2z = LD(&(x[WS(rs, 11)]), ms, &(x[WS(rs, 1)]));
				   T2B = LD(&(x[WS(rs, 43)]), ms, &(x[WS(rs, 1)]));
				   {
					V T5I, T5R, T44, T1G, T2m, T2A, T2C, T5S, T5L, T1R, T45, T2o, T5J, T1L;
					T2o = LD(&(x[WS(rs, 19)]), ms, &(x[WS(rs, 1)]));
					T5J = VSUB(T1I, T1K);
					T1L = VADD(T1I, T1K);
					T5I = VFNMS(LDK(KP414213562), T5H, T5G);
					T5R = VFMA(LDK(KP414213562), T5G, T5H);
					T44 = VADD(T1A, T1F);
					T1G = VSUB(T1A, T1F);
					T2m = BYTW(&(W[TWVL * 68]), T2l);
					T2A = BYTW(&(W[TWVL * 20]), T2z);
					T2C = BYTW(&(W[TWVL * 84]), T2B);
					T5S = VFNMS(LDK(KP414213562), T5J, T5K);
					T5L = VFMA(LDK(KP414213562), T5K, T5J);
					T1R = VSUB(T1L, T1Q);
					T45 = VADD(T1L, T1Q);
					T2p = BYTW(&(W[TWVL * 36]), T2o);
					T61 = VSUB(T2k, T2m);
					T2n = VADD(T2k, T2m);
					T65 = VSUB(T2C, T2A);
					T2D = VADD(T2A, T2C);
					T7p = VSUB(T5I, T5L);
					T5M = VADD(T5I, T5L);
					T7m = VSUB(T5R, T5S);
					T5T = VADD(T5R, T5S);
					T4G = VSUB(T44, T45);
					T46 = VADD(T44, T45);
					T25 = VSUB(T1G, T1R);
					T1S = VADD(T1G, T1R);
					T2q = LD(&(x[WS(rs, 51)]), ms, &(x[WS(rs, 1)]));
				   }
				   T2u = LD(&(x[WS(rs, 59)]), ms, &(x[WS(rs, 1)]));
				   T2w = LD(&(x[WS(rs, 27)]), ms, &(x[WS(rs, 1)]));
			      }
			 }
		    }
	       }
	       {
		    V T67, T7w, T6e, T7t, T3s, T3E, T39, T3D, T1k, T3k, T3t, T3c, T1T, T3v, T3w;
		    V T26, T2G, T3y, T3z, T2T;
		    {
			 V T4A, T4N, T47, T4v, T2r, T2v, T2x, T4s, T40, T3W, T3Z;
			 T4A = VSUB(T3U, T3V);
			 T3W = VADD(T3U, T3V);
			 T3Z = VADD(T3X, T3Y);
			 T4N = VSUB(T3X, T3Y);
			 T47 = VSUB(T43, T46);
			 T4v = VADD(T43, T46);
			 T2r = BYTW(&(W[TWVL * 100]), T2q);
			 T2v = BYTW(&(W[TWVL * 116]), T2u);
			 T2x = BYTW(&(W[TWVL * 52]), T2w);
			 T4s = VADD(T3W, T3Z);
			 T40 = VSUB(T3W, T3Z);
			 {
			      V T4O, T4n, T4Q, T4H, T4E, T4W, T4u, T4y, T4d, T4J, T2F, T2S;
			      {
				   V T6c, T63, T2t, T4b, T6d, T66, T2E, T4c;
				   {
					V T4D, T62, T2s, T64, T2y, T4t;
					T4O = VSUB(T4B, T4C);
					T4D = VADD(T4B, T4C);
					T62 = VSUB(T2r, T2p);
					T2s = VADD(T2p, T2r);
					T64 = VSUB(T2v, T2x);
					T2y = VADD(T2v, T2x);
					T4t = VADD(T4j, T4m);
					T4n = VSUB(T4j, T4m);
					T4Q = VFMA(LDK(KP414213562), T4F, T4G);
					T4H = VFNMS(LDK(KP414213562), T4G, T4F);
					T4E = VFMA(LDK(KP707106781), T4D, T4A);
					T4W = VFNMS(LDK(KP707106781), T4D, T4A);
					T6c = VFNMS(LDK(KP414213562), T61, T62);
					T63 = VFMA(LDK(KP414213562), T62, T61);
					T2t = VSUB(T2n, T2s);
					T4b = VADD(T2n, T2s);
					T6d = VFMA(LDK(KP414213562), T64, T65);
					T66 = VFNMS(LDK(KP414213562), T65, T64);
					T2E = VSUB(T2y, T2D);
					T4c = VADD(T2y, T2D);
					T4u = VSUB(T4s, T4t);
					T4y = VADD(T4s, T4t);
				   }
				   T67 = VADD(T63, T66);
				   T7w = VSUB(T66, T63);
				   T6e = VADD(T6c, T6d);
				   T7t = VSUB(T6d, T6c);
				   T4d = VADD(T4b, T4c);
				   T4J = VSUB(T4c, T4b);
				   T2F = VADD(T2t, T2E);
				   T2S = VSUB(T2E, T2t);
			      }
			      {
				   V Ty, T1j, T4R, T4K;
				   Ty = VFMA(LDK(KP707106781), Tx, Ta);
				   T3s = VFNMS(LDK(KP707106781), Tx, Ta);
				   T3E = VSUB(TV, T1i);
				   T1j = VADD(TV, T1i);
				   T39 = VFMA(LDK(KP707106781), T38, T37);
				   T3D = VFNMS(LDK(KP707106781), T38, T37);
				   T4R = VFMA(LDK(KP414213562), T4I, T4J);
				   T4K = VFNMS(LDK(KP414213562), T4J, T4I);
				   {
					V T4w, T4e, T4P, T4Z;
					T4w = VADD(T4a, T4d);
					T4e = VSUB(T4a, T4d);
					T4P = VFMA(LDK(KP707106781), T4O, T4N);
					T4Z = VFNMS(LDK(KP707106781), T4O, T4N);
					T1k = VFMA(LDK(KP923879532), T1j, Ty);
					T3k = VFNMS(LDK(KP923879532), T1j, Ty);
					{
					     V T4L, T50, T4S, T4X;
					     T4L = VADD(T4H, T4K);
					     T50 = VSUB(T4H, T4K);
					     T4S = VSUB(T4Q, T4R);
					     T4X = VADD(T4Q, T4R);
					     {
						  V T4f, T4o, T4x, T4z;
						  T4f = VADD(T47, T4e);
						  T4o = VSUB(T47, T4e);
						  T4x = VSUB(T4v, T4w);
						  T4z = VADD(T4v, T4w);
						  {
						       V T53, T51, T4M, T4U;
						       T53 = VFNMS(LDK(KP923879532), T50, T4Z);
						       T51 = VFMA(LDK(KP923879532), T50, T4Z);
						       T4M = VFNMS(LDK(KP923879532), T4L, T4E);
						       T4U = VFMA(LDK(KP923879532), T4L, T4E);
						       {
							    V T52, T4Y, T4T, T4V;
							    T52 = VFMA(LDK(KP923879532), T4X, T4W);
							    T4Y = VFNMS(LDK(KP923879532), T4X, T4W);
							    T4T = VFNMS(LDK(KP923879532), T4S, T4P);
							    T4V = VFMA(LDK(KP923879532), T4S, T4P);
							    {
								 V T4p, T4r, T4g, T4q;
								 T4p = VFNMS(LDK(KP707106781), T4o, T4n);
								 T4r = VFMA(LDK(KP707106781), T4o, T4n);
								 T4g = VFNMS(LDK(KP707106781), T4f, T40);
								 T4q = VFMA(LDK(KP707106781), T4f, T40);
								 ST(&(x[0]), VADD(T4y, T4z), ms, &(x[0]));
								 ST(&(x[WS(rs, 32)]), VSUB(T4y, T4z), ms, &(x[0]));
								 ST(&(x[WS(rs, 16)]), VFMAI(T4x, T4u), ms, &(x[0]));
								 ST(&(x[WS(rs, 48)]), VFNMSI(T4x, T4u), ms, &(x[0]));
								 ST(&(x[WS(rs, 44)]), VFNMSI(T51, T4Y), ms, &(x[0]));
								 ST(&(x[WS(rs, 20)]), VFMAI(T51, T4Y), ms, &(x[0]));
								 ST(&(x[WS(rs, 52)]), VFMAI(T53, T52), ms, &(x[0]));
								 ST(&(x[WS(rs, 12)]), VFNMSI(T53, T52), ms, &(x[0]));
								 ST(&(x[WS(rs, 4)]), VFMAI(T4V, T4U), ms, &(x[0]));
								 ST(&(x[WS(rs, 60)]), VFNMSI(T4V, T4U), ms, &(x[0]));
								 ST(&(x[WS(rs, 36)]), VFMAI(T4T, T4M), ms, &(x[0]));
								 ST(&(x[WS(rs, 28)]), VFNMSI(T4T, T4M), ms, &(x[0]));
								 ST(&(x[WS(rs, 56)]), VFNMSI(T4r, T4q), ms, &(x[0]));
								 ST(&(x[WS(rs, 8)]), VFMAI(T4r, T4q), ms, &(x[0]));
								 ST(&(x[WS(rs, 40)]), VFMAI(T4p, T4g), ms, &(x[0]));
								 ST(&(x[WS(rs, 24)]), VFNMSI(T4p, T4g), ms, &(x[0]));
								 T3t = VADD(T3a, T3b);
								 T3c = VSUB(T3a, T3b);
							    }
						       }
						  }
					     }
					}
				   }
				   T1T = VFMA(LDK(KP707106781), T1S, T1v);
				   T3v = VFNMS(LDK(KP707106781), T1S, T1v);
				   T3w = VFNMS(LDK(KP707106781), T25, T24);
				   T26 = VFMA(LDK(KP707106781), T25, T24);
				   T2G = VFMA(LDK(KP707106781), T2F, T2i);
				   T3y = VFNMS(LDK(KP707106781), T2F, T2i);
				   T3z = VFNMS(LDK(KP707106781), T2S, T2R);
				   T2T = VFMA(LDK(KP707106781), T2S, T2R);
			      }
			 }
		    }
		    {
			 V T3u, T3M, T3F, T3P, T3x, T3G, T3q, T3m, T3h, T3j, T3r, T3p, T2W, T3i;
			 {
			      V T3d, T3n, T27, T3e, T2U, T3f;
			      T3d = VFMA(LDK(KP923879532), T3c, T39);
			      T3n = VFNMS(LDK(KP923879532), T3c, T39);
			      T27 = VFNMS(LDK(KP198912367), T26, T1T);
			      T3e = VFMA(LDK(KP198912367), T1T, T26);
			      T2U = VFNMS(LDK(KP198912367), T2T, T2G);
			      T3f = VFMA(LDK(KP198912367), T2G, T2T);
			      T3u = VFMA(LDK(KP923879532), T3t, T3s);
			      T3M = VFNMS(LDK(KP923879532), T3t, T3s);
			      {
				   V T3g, T3l, T2V, T3o;
				   T3g = VSUB(T3e, T3f);
				   T3l = VADD(T3e, T3f);
				   T2V = VADD(T27, T2U);
				   T3o = VSUB(T27, T2U);
				   T3F = VFNMS(LDK(KP923879532), T3E, T3D);
				   T3P = VFMA(LDK(KP923879532), T3E, T3D);
				   T3x = VFMA(LDK(KP668178637), T3w, T3v);
				   T3G = VFNMS(LDK(KP668178637), T3v, T3w);
				   T3q = VFMA(LDK(KP980785280), T3l, T3k);
				   T3m = VFNMS(LDK(KP980785280), T3l, T3k);
				   T3h = VFNMS(LDK(KP980785280), T3g, T3d);
				   T3j = VFMA(LDK(KP980785280), T3g, T3d);
				   T3r = VFNMS(LDK(KP980785280), T3o, T3n);
				   T3p = VFMA(LDK(KP980785280), T3o, T3n);
				   T2W = VFNMS(LDK(KP980785280), T2V, T1k);
				   T3i = VFMA(LDK(KP980785280), T2V, T1k);
			      }
			 }
			 {
			      V T7n, T7Z, T8j, T89, T7k, T7O, T8g, T7Y, T7H, T7R, T80, T7q, T7u, T82, T83;
			      V T7x;
			      {
				   V T7c, T7W, T7D, T87, T7f, T7E, T3A, T3H, T7F, T7i;
				   T7c = VFNMS(LDK(KP923879532), T7b, T7a);
				   T7W = VFMA(LDK(KP923879532), T7b, T7a);
				   T7D = VFMA(LDK(KP923879532), T7C, T7B);
				   T87 = VFNMS(LDK(KP923879532), T7C, T7B);
				   T7f = VFNMS(LDK(KP668178637), T7e, T7d);
				   T7E = VFMA(LDK(KP668178637), T7d, T7e);
				   ST(&(x[WS(rs, 46)]), VFNMSI(T3p, T3m), ms, &(x[0]));
				   ST(&(x[WS(rs, 18)]), VFMAI(T3p, T3m), ms, &(x[0]));
				   ST(&(x[WS(rs, 50)]), VFMAI(T3r, T3q), ms, &(x[0]));
				   ST(&(x[WS(rs, 14)]), VFNMSI(T3r, T3q), ms, &(x[0]));
				   ST(&(x[WS(rs, 2)]), VFMAI(T3j, T3i), ms, &(x[0]));
				   ST(&(x[WS(rs, 62)]), VFNMSI(T3j, T3i), ms, &(x[0]));
				   ST(&(x[WS(rs, 34)]), VFMAI(T3h, T2W), ms, &(x[0]));
				   ST(&(x[WS(rs, 30)]), VFNMSI(T3h, T2W), ms, &(x[0]));
				   T3A = VFMA(LDK(KP668178637), T3z, T3y);
				   T3H = VFNMS(LDK(KP668178637), T3y, T3z);
				   T7F = VFMA(LDK(KP668178637), T7g, T7h);
				   T7i = VFNMS(LDK(KP668178637), T7h, T7g);
				   T7n = VFNMS(LDK(KP923879532), T7m, T7l);
				   T7Z = VFMA(LDK(KP923879532), T7m, T7l);
				   {
					V T3I, T3N, T3B, T3Q;
					T3I = VSUB(T3G, T3H);
					T3N = VADD(T3G, T3H);
					T3B = VADD(T3x, T3A);
					T3Q = VSUB(T3x, T3A);
					{
					     V T7j, T88, T7G, T7X;
					     T7j = VADD(T7f, T7i);
					     T88 = VSUB(T7f, T7i);
					     T7G = VSUB(T7E, T7F);
					     T7X = VADD(T7E, T7F);
					     {
						  V T3S, T3O, T3J, T3L;
						  T3S = VFNMS(LDK(KP831469612), T3N, T3M);
						  T3O = VFMA(LDK(KP831469612), T3N, T3M);
						  T3J = VFNMS(LDK(KP831469612), T3I, T3F);
						  T3L = VFMA(LDK(KP831469612), T3I, T3F);
						  {
						       V T3T, T3R, T3C, T3K;
						       T3T = VFMA(LDK(KP831469612), T3Q, T3P);
						       T3R = VFNMS(LDK(KP831469612), T3Q, T3P);
						       T3C = VFNMS(LDK(KP831469612), T3B, T3u);
						       T3K = VFMA(LDK(KP831469612), T3B, T3u);
						       T8j = VFNMS(LDK(KP831469612), T88, T87);
						       T89 = VFMA(LDK(KP831469612), T88, T87);
						       T7k = VFNMS(LDK(KP831469612), T7j, T7c);
						       T7O = VFMA(LDK(KP831469612), T7j, T7c);
						       T8g = VFNMS(LDK(KP831469612), T7X, T7W);
						       T7Y = VFMA(LDK(KP831469612), T7X, T7W);
						       T7H = VFMA(LDK(KP831469612), T7G, T7D);
						       T7R = VFNMS(LDK(KP831469612), T7G, T7D);
						       ST(&(x[WS(rs, 42)]), VFMAI(T3R, T3O), ms, &(x[0]));
						       ST(&(x[WS(rs, 22)]), VFNMSI(T3R, T3O), ms, &(x[0]));
						       ST(&(x[WS(rs, 54)]), VFNMSI(T3T, T3S), ms, &(x[0]));
						       ST(&(x[WS(rs, 10)]), VFMAI(T3T, T3S), ms, &(x[0]));
						       ST(&(x[WS(rs, 58)]), VFMAI(T3L, T3K), ms, &(x[0]));
						       ST(&(x[WS(rs, 6)]), VFNMSI(T3L, T3K), ms, &(x[0]));
						       ST(&(x[WS(rs, 26)]), VFMAI(T3J, T3C), ms, &(x[0]));
						       ST(&(x[WS(rs, 38)]), VFNMSI(T3J, T3C), ms, &(x[0]));
						       T80 = VFNMS(LDK(KP923879532), T7p, T7o);
						       T7q = VFMA(LDK(KP923879532), T7p, T7o);
						  }
					     }
					}
				   }
				   T7u = VFNMS(LDK(KP923879532), T7t, T7s);
				   T82 = VFMA(LDK(KP923879532), T7t, T7s);
				   T83 = VFNMS(LDK(KP923879532), T7w, T7v);
				   T7x = VFMA(LDK(KP923879532), T7w, T7v);
			      }
			      {
				   V T5g, T6I, T6p, T6T, T5p, T6q, T6r, T5y;
				   T5g = VFMA(LDK(KP923879532), T5f, T58);
				   T6I = VFNMS(LDK(KP923879532), T5f, T58);
				   {
					V T7r, T7I, T7y, T7J;
					T7r = VFNMS(LDK(KP534511135), T7q, T7n);
					T7I = VFMA(LDK(KP534511135), T7n, T7q);
					T7y = VFNMS(LDK(KP534511135), T7x, T7u);
					T7J = VFMA(LDK(KP534511135), T7u, T7x);
					{
					     V T81, T8a, T84, T8b;
					     T81 = VFMA(LDK(KP303346683), T80, T7Z);
					     T8a = VFNMS(LDK(KP303346683), T7Z, T80);
					     T84 = VFMA(LDK(KP303346683), T83, T82);
					     T8b = VFNMS(LDK(KP303346683), T82, T83);
					     T6p = VFMA(LDK(KP923879532), T6o, T6l);
					     T6T = VFNMS(LDK(KP923879532), T6o, T6l);
					     T5p = VFNMS(LDK(KP198912367), T5o, T5l);
					     T6q = VFMA(LDK(KP198912367), T5l, T5o);
					     {
						  V T7K, T7P, T7z, T7S;
						  T7K = VSUB(T7I, T7J);
						  T7P = VADD(T7I, T7J);
						  T7z = VADD(T7r, T7y);
						  T7S = VSUB(T7r, T7y);
						  {
						       V T8c, T8h, T85, T8k;
						       T8c = VSUB(T8a, T8b);
						       T8h = VADD(T8a, T8b);
						       T85 = VADD(T81, T84);
						       T8k = VSUB(T81, T84);
						       {
							    V T7Q, T7U, T7L, T7N;
							    T7Q = VFNMS(LDK(KP881921264), T7P, T7O);
							    T7U = VFMA(LDK(KP881921264), T7P, T7O);
							    T7L = VFNMS(LDK(KP881921264), T7K, T7H);
							    T7N = VFMA(LDK(KP881921264), T7K, T7H);
							    {
								 V T7T, T7V, T7A, T7M;
								 T7T = VFMA(LDK(KP881921264), T7S, T7R);
								 T7V = VFNMS(LDK(KP881921264), T7S, T7R);
								 T7A = VFNMS(LDK(KP881921264), T7z, T7k);
								 T7M = VFMA(LDK(KP881921264), T7z, T7k);
								 {
								      V T8i, T8m, T8d, T8f;
								      T8i = VFMA(LDK(KP956940335), T8h, T8g);
								      T8m = VFNMS(LDK(KP956940335), T8h, T8g);
								      T8d = VFNMS(LDK(KP956940335), T8c, T89);
								      T8f = VFMA(LDK(KP956940335), T8c, T89);
								      {
									   V T8l, T8n, T86, T8e;
									   T8l = VFNMS(LDK(KP956940335), T8k, T8j);
									   T8n = VFMA(LDK(KP956940335), T8k, T8j);
									   T86 = VFNMS(LDK(KP956940335), T85, T7Y);
									   T8e = VFMA(LDK(KP956940335), T85, T7Y);
									   ST(&(x[WS(rs, 53)]), VFMAI(T7V, T7U), ms, &(x[WS(rs, 1)]));
									   ST(&(x[WS(rs, 11)]), VFNMSI(T7V, T7U), ms, &(x[WS(rs, 1)]));
									   ST(&(x[WS(rs, 43)]), VFNMSI(T7T, T7Q), ms, &(x[WS(rs, 1)]));
									   ST(&(x[WS(rs, 21)]), VFMAI(T7T, T7Q), ms, &(x[WS(rs, 1)]));
									   ST(&(x[WS(rs, 5)]), VFMAI(T7N, T7M), ms, &(x[WS(rs, 1)]));
									   ST(&(x[WS(rs, 59)]), VFNMSI(T7N, T7M), ms, &(x[WS(rs, 1)]));
									   ST(&(x[WS(rs, 37)]), VFMAI(T7L, T7A), ms, &(x[WS(rs, 1)]));
									   ST(&(x[WS(rs, 27)]), VFNMSI(T7L, T7A), ms, &(x[WS(rs, 1)]));
									   ST(&(x[WS(rs, 51)]), VFNMSI(T8n, T8m), ms, &(x[WS(rs, 1)]));
									   ST(&(x[WS(rs, 13)]), VFMAI(T8n, T8m), ms, &(x[WS(rs, 1)]));
									   ST(&(x[WS(rs, 45)]), VFMAI(T8l, T8i), ms, &(x[WS(rs, 1)]));
									   ST(&(x[WS(rs, 19)]), VFNMSI(T8l, T8i), ms, &(x[WS(rs, 1)]));
									   ST(&(x[WS(rs, 61)]), VFMAI(T8f, T8e), ms, &(x[WS(rs, 1)]));
									   ST(&(x[WS(rs, 3)]), VFNMSI(T8f, T8e), ms, &(x[WS(rs, 1)]));
									   ST(&(x[WS(rs, 29)]), VFMAI(T8d, T86), ms, &(x[WS(rs, 1)]));
									   ST(&(x[WS(rs, 35)]), VFNMSI(T8d, T86), ms, &(x[WS(rs, 1)]));
									   T6r = VFMA(LDK(KP198912367), T5u, T5x);
									   T5y = VFNMS(LDK(KP198912367), T5x, T5u);
								      }
								 }
							    }
						       }
						  }
					     }
					}
				   }
				   {