t1sv_32.c

用于DFT计算的c语言的库的最新版本,包含丰富的函数库.

				   T1Q = LD(&(ri[WS(rs, 17)]), ms, &(ri[WS(rs, 1)]));
				   T1T = LD(&(ii[WS(rs, 17)]), ms, &(ii[WS(rs, 1)]));
				   T1P = LDW(&(W[TWVL * 32]));
				   T1S = LDW(&(W[TWVL * 33]));
				   T1X = LD(&(ri[WS(rs, 9)]), ms, &(ri[WS(rs, 1)]));
				   T20 = LD(&(ii[WS(rs, 9)]), ms, &(ii[WS(rs, 1)]));
				   T4Q = VFNMS(T25, T23, T4P);
				   T27 = VFMA(T25, T26, T24);
				   T1W = LDW(&(W[TWVL * 16]));
				   T1Z = LDW(&(W[TWVL * 17]));
				   T4r = VMUL(T1P, T1T);
				   T1R = VMUL(T1P, T1Q);
			      }
			      {
				   V T56, T38, T3a, T3d, T3c, T3h, T3k, T3g, T3j, T5g, T3r, T57, T3b;
				   {
					V T3n, T3q, T6V, T4t, T1V, T4M, T6W, T4R, T28, T4u, T3m, T3p, T5f, T3o, T39;
					{
					     V T34, T37, T33, T36, T4N, T1Y, T4s, T1U;
					     T34 = LD(&(ri[WS(rs, 3)]), ms, &(ri[WS(rs, 1)]));
					     T37 = LD(&(ii[WS(rs, 3)]), ms, &(ii[WS(rs, 1)]));
					     T33 = LDW(&(W[TWVL * 4]));
					     T36 = LDW(&(W[TWVL * 5]));
					     T4N = VMUL(T1W, T20);
					     T1Y = VMUL(T1W, T1X);
					     T4s = VFNMS(T1S, T1Q, T4r);
					     T1U = VFMA(T1S, T1T, T1R);
					     T3n = LD(&(ri[WS(rs, 11)]), ms, &(ri[WS(rs, 1)]));
					     T3q = LD(&(ii[WS(rs, 11)]), ms, &(ii[WS(rs, 1)]));
					     {
						  V T55, T35, T4O, T21;
						  T55 = VMUL(T33, T37);
						  T35 = VMUL(T33, T34);
						  T4O = VFNMS(T1Z, T1X, T4N);
						  T21 = VFMA(T1Z, T20, T1Y);
						  T6V = VADD(T4q, T4s);
						  T4t = VSUB(T4q, T4s);
						  T1V = VADD(T1O, T1U);
						  T4M = VSUB(T1O, T1U);
						  T56 = VFNMS(T36, T34, T55);
						  T38 = VFMA(T36, T37, T35);
						  T6W = VADD(T4O, T4Q);
						  T4R = VSUB(T4O, T4Q);
						  T28 = VADD(T21, T27);
						  T4u = VSUB(T21, T27);
						  T3m = LDW(&(W[TWVL * 20]));
					     }
					     T3p = LDW(&(W[TWVL * 21]));
					}
					T3a = LD(&(ri[WS(rs, 19)]), ms, &(ri[WS(rs, 1)]));
					T3d = LD(&(ii[WS(rs, 19)]), ms, &(ii[WS(rs, 1)]));
					T4S = VADD(T4M, T4R);
					T65 = VSUB(T4M, T4R);
					T6X = VSUB(T6V, T6W);
					T7I = VADD(T6V, T6W);
					T4v = VSUB(T4t, T4u);
					T68 = VADD(T4t, T4u);
					T29 = VADD(T1V, T28);
					T70 = VSUB(T1V, T28);
					T5f = VMUL(T3m, T3q);
					T3o = VMUL(T3m, T3n);
					T39 = LDW(&(W[TWVL * 36]));
					T3c = LDW(&(W[TWVL * 37]));
					T3h = LD(&(ri[WS(rs, 27)]), ms, &(ri[WS(rs, 1)]));
					T3k = LD(&(ii[WS(rs, 27)]), ms, &(ii[WS(rs, 1)]));
					T3g = LDW(&(W[TWVL * 52]));
					T3j = LDW(&(W[TWVL * 53]));
					T5g = VFNMS(T3p, T3n, T5f);
					T3r = VFMA(T3p, T3q, T3o);
					T57 = VMUL(T39, T3d);
					T3b = VMUL(T39, T3a);
				   }
				   {
					V T2u, T2x, T2w, T4G, T2v, T2g;
					{
					     V T2b, T2e, T5d, T3i, T2a, T2d, T58, T3e, T2t;
					     T2b = LD(&(ri[WS(rs, 5)]), ms, &(ri[WS(rs, 1)]));
					     T2e = LD(&(ii[WS(rs, 5)]), ms, &(ii[WS(rs, 1)]));
					     T5d = VMUL(T3g, T3k);
					     T3i = VMUL(T3g, T3h);
					     T2a = LDW(&(W[TWVL * 8]));
					     T2d = LDW(&(W[TWVL * 9]));
					     T58 = VFNMS(T3c, T3a, T57);
					     T3e = VFMA(T3c, T3d, T3b);
					     T2u = LD(&(ri[WS(rs, 13)]), ms, &(ri[WS(rs, 1)]));
					     T2x = LD(&(ii[WS(rs, 13)]), ms, &(ii[WS(rs, 1)]));
					     {
						  V T5e, T3l, T4w, T2c;
						  T5e = VFNMS(T3j, T3h, T5d);
						  T3l = VFMA(T3j, T3k, T3i);
						  T4w = VMUL(T2a, T2e);
						  T2c = VMUL(T2a, T2b);
						  {
						       V T7c, T59, T3f, T5a;
						       T7c = VADD(T56, T58);
						       T59 = VSUB(T56, T58);
						       T3f = VADD(T38, T3e);
						       T5a = VSUB(T38, T3e);
						       {
							    V T7d, T5h, T3s, T5c;
							    T7d = VADD(T5e, T5g);
							    T5h = VSUB(T5e, T5g);
							    T3s = VADD(T3l, T3r);
							    T5c = VSUB(T3l, T3r);
							    T4x = VFNMS(T2d, T2b, T4w);
							    T2f = VFMA(T2d, T2e, T2c);
							    T5b = VSUB(T59, T5a);
							    T5s = VADD(T5a, T59);
							    T2t = LDW(&(W[TWVL * 24]));
							    T7O = VADD(T7c, T7d);
							    T7e = VSUB(T7c, T7d);
							    T5t = VSUB(T5c, T5h);
							    T5i = VADD(T5c, T5h);
							    T79 = VSUB(T3s, T3f);
							    T3t = VADD(T3f, T3s);
						       }
						  }
					     }
					     T2w = LDW(&(W[TWVL * 25]));
					     T4G = VMUL(T2t, T2x);
					     T2v = VMUL(T2t, T2u);
					}
					T2h = LD(&(ri[WS(rs, 21)]), ms, &(ri[WS(rs, 1)]));
					T2k = LD(&(ii[WS(rs, 21)]), ms, &(ii[WS(rs, 1)]));
					T2g = LDW(&(W[TWVL * 40]));
					T2j = LDW(&(W[TWVL * 41]));
					T2o = LD(&(ri[WS(rs, 29)]), ms, &(ri[WS(rs, 1)]));
					T2r = LD(&(ii[WS(rs, 29)]), ms, &(ii[WS(rs, 1)]));
					T4H = VFNMS(T2w, T2u, T4G);
					T2y = VFMA(T2w, T2x, T2v);
					T2n = LDW(&(W[TWVL * 56]));
					T2q = LDW(&(W[TWVL * 57]));
					T4y = VMUL(T2g, T2k);
					T2i = VMUL(T2g, T2h);
				   }
			      }
			 }
		    }
	       }
	       {
		    V T4C, T4T, T4U, T4J, T7A, T7w, T7j, T75, T7i, T6U, T8p, T8n, T8v, T8t, T7q;
		    V T7y, T7t, T7z, T7g, T7k;
		    {
			 V T6E, T8j, T6H, T8k, T73, T6Y, T7S, T8i, T8h, T7V;
			 {
			      V T7P, T7Y, T7C, TO, T89, T8e, T3u, T7M, T8d, T1H, T7K, T7X, T2B, T7H;
			      {
				   V T71, T2m, T72, T4I, T2z, T4D, Tm, TN, T2A, T7J;
				   T6E = VSUB(T8, Tl);
				   Tm = VADD(T8, Tl);
				   TN = VADD(Tz, TM);
				   T8j = VSUB(TM, Tz);
				   T7P = VSUB(T7N, T7O);
				   T7Y = VADD(T7N, T7O);
				   {
					V T82, T4E, T2p, T4z, T2l, T88;
					T82 = VADD(T6F, T6G);
					T6H = VSUB(T6F, T6G);
					T4E = VMUL(T2n, T2r);
					T2p = VMUL(T2n, T2o);
					T4z = VFNMS(T2j, T2h, T4y);
					T2l = VFMA(T2j, T2k, T2i);
					T8k = VSUB(T87, T83);
					T88 = VADD(T83, T87);
					T7C = VSUB(Tm, TN);
					TO = VADD(Tm, TN);
					{
					     V T4F, T2s, T4A, T4B;
					     T4F = VFNMS(T2q, T2o, T4E);
					     T2s = VFMA(T2q, T2r, T2p);
					     T71 = VADD(T4x, T4z);
					     T4A = VSUB(T4x, T4z);
					     T2m = VADD(T2f, T2l);
					     T4B = VSUB(T2f, T2l);
					     T89 = VADD(T82, T88);
					     T8e = VSUB(T88, T82);
					     T72 = VADD(T4F, T4H);
					     T4I = VSUB(T4F, T4H);
					     T2z = VADD(T2s, T2y);
					     T4D = VSUB(T2s, T2y);
					     T4C = VSUB(T4A, T4B);
					     T4T = VADD(T4B, T4A);
					}
				   }
				   T3u = VADD(T32, T3t);
				   T7M = VSUB(T32, T3t);
				   T7J = VADD(T71, T72);
				   T73 = VSUB(T71, T72);
				   T4U = VSUB(T4D, T4I);
				   T4J = VADD(T4D, T4I);
				   T6Y = VSUB(T2z, T2m);
				   T2A = VADD(T2m, T2z);
				   T8d = VSUB(T1G, T1f);
				   T1H = VADD(T1f, T1G);
				   T7K = VSUB(T7I, T7J);
				   T7X = VADD(T7I, T7J);
				   T2B = VADD(T29, T2A);
				   T7H = VSUB(T29, T2A);
			      }
			      {
				   V T1I, T80, T7Q, T7U, T7F, T7L, T7T, T3v, T8b, T8c, T8a, T7W, T81, T7Z;
				   T7W = VSUB(TO, T1H);
				   T1I = VADD(TO, T1H);
				   T7Z = VSUB(T7X, T7Y);
				   T80 = VADD(T7X, T7Y);
				   T7Q = VSUB(T7M, T7P);
				   T7U = VADD(T7M, T7P);
				   T7F = VSUB(T7D, T7E);
				   T81 = VADD(T7D, T7E);
				   T7L = VADD(T7H, T7K);
				   T7T = VSUB(T7K, T7H);
				   T3v = VADD(T2B, T3u);
				   T8b = VSUB(T3u, T2B);
				   ST(&(ri[WS(rs, 24)]), VSUB(T7W, T7Z), ms, &(ri[0]));
				   ST(&(ri[WS(rs, 8)]), VADD(T7W, T7Z), ms, &(ri[0]));
				   T8c = VSUB(T89, T81);
				   T8a = VADD(T81, T89);
				   {
					V T8f, T8g, T7G, T7R;
					T7S = VSUB(T7C, T7F);
					T7G = VADD(T7C, T7F);
					T7R = VADD(T7L, T7Q);
					T8i = VSUB(T7Q, T7L);
					T8h = VSUB(T8e, T8d);
					T8f = VADD(T8d, T8e);
					ST(&(ri[0]), VADD(T1I, T3v), ms, &(ri[0]));
					ST(&(ri[WS(rs, 16)]), VSUB(T1I, T3v), ms, &(ri[0]));
					T8g = VADD(T7T, T7U);
					T7V = VSUB(T7T, T7U);
					ST(&(ii[WS(rs, 16)]), VSUB(T8a, T80), ms, &(ii[0]));
					ST(&(ii[0]), VADD(T80, T8a), ms, &(ii[0]));
					ST(&(ii[WS(rs, 24)]), VSUB(T8c, T8b), ms, &(ii[0]));
					ST(&(ii[WS(rs, 8)]), VADD(T8b, T8c), ms, &(ii[0]));
					ST(&(ri[WS(rs, 4)]), VFMA(LDK(KP707106781), T7R, T7G), ms, &(ri[0]));
					ST(&(ri[WS(rs, 20)]), VFNMS(LDK(KP707106781), T7R, T7G), ms, &(ri[0]));
					ST(&(ii[WS(rs, 20)]), VFNMS(LDK(KP707106781), T8g, T8f), ms, &(ii[0]));
					ST(&(ii[WS(rs, 4)]), VFMA(LDK(KP707106781), T8g, T8f), ms, &(ii[0]));
				   }
			      }
			 }
			 {
			      V T7f, T7a, T7m, T6I, T7s, T7r, T8r, T8l, T8m, T6T, T8s, T7p;
			      {
				   V T7n, T6N, T6S, T7o, T7u, T7v, T6Z, T74;
				   T7f = VSUB(T7b, T7e);
				   T7u = VADD(T7b, T7e);
				   T7v = VADD(T78, T79);
				   T7a = VSUB(T78, T79);
				   ST(&(ri[WS(rs, 12)]), VFMA(LDK(KP707106781), T7V, T7S), ms, &(ri[0]));
				   ST(&(ri[WS(rs, 28)]), VFNMS(LDK(KP707106781), T7V, T7S), ms, &(ri[0]));
				   ST(&(ii[WS(rs, 28)]), VFNMS(LDK(KP707106781), T8i, T8h), ms, &(ii[0]));
				   ST(&(ii[WS(rs, 12)]), VFMA(LDK(KP707106781), T8i, T8h), ms, &(ii[0]));
				   T7m = VADD(T6E, T6H);
				   T6I = VSUB(T6E, T6H);
				   T7A = VFMA(LDK(KP414213562), T7u, T7v);
				   T7w = VFNMS(LDK(KP414213562), T7v, T7u);
				   T7n = VADD(T6M, T6L);
				   T6N = VSUB(T6L, T6M);
				   T6S = VADD(T6O, T6R);
				   T7o = VSUB(T6O, T6R);
				   T7s = VADD(T6X, T6Y);
				   T6Z = VSUB(T6X, T6Y);
				   T74 = VSUB(T70, T73);
				   T7r = VADD(T70, T73);
				   T8r = VSUB(T8k, T8j);
				   T8l = VADD(T8j, T8k);
				   T8m = VADD(T6N, T6S);
				   T6T = VSUB(T6N, T6S);
				   T7j = VFNMS(LDK(KP414213562), T6Z, T74);
				   T75 = VFMA(LDK(KP414213562), T74, T6Z);
				   T8s = VSUB(T7o, T7n);
				   T7p = VADD(T7n, T7o);
			      }
			      T7i = VFNMS(LDK(KP707106781), T6T, T6I);
			      T6U = VFMA(LDK(KP707106781), T6T, T6I);
			      T8p = VFNMS(LDK(KP707106781), T8m, T8l);
			      T8n = VFMA(LDK(KP707106781), T8m, T8l);
			      T8v = VFNMS(LDK(KP707106781), T8s, T8r);
			      T8t = VFMA(LDK(KP707106781), T8s, T8r);
			      T7q = VFMA(LDK(KP707106781), T7p, T7m);
			      T7y = VFNMS(LDK(KP707106781), T7p, T7m);
			      T7t = VFMA(LDK(KP414213562), T7s, T7r);
			      T7z = VFNMS(LDK(KP414213562), T7r, T7s);
			      T7g = VFNMS(LDK(KP414213562), T7f, T7a);
			      T7k = VFMA(LDK(KP414213562), T7a, T7f);
			 }
		    }
		    {
			 V T5S, T8O, T8N, T5V, T6d, T6g, T66, T4L, T5I, T69, T5y, T4o, T8J, T8L, T5M;
			 V T5Q, T5A, T5w, T5H, T4W, T5O, T5G, T8D, T8F;
			 {
			      V T5C, T3S, T8C, T4n, T8H, T8B, T8I, T5F, T5L, T5k, T5K, T5v, T4V;
			      {
				   V T5D, T47, T4m, T5E, T8z, T8A, T3C, T3R, T5j, T5u, T4K;
				   T5S = VSUB(T3w, T3B);
				   T3C = VADD(T3w, T3B);
				   T3R = VADD(T3J, T3Q);
				   T8O = VSUB(T3Q, T3J);
				   {
					V T8o, T7B, T7x, T8q;
					T8o = VADD(T7z, T7A);
					T7B = VSUB(T7z, T7A);
					T7x = VADD(T7t, T7w);
					T8q = VSUB(T7w, T7t);
					{
					     V T8u, T7l, T7h, T8w;
					     T8u = VSUB(T7k, T7j);
					     T7l = VADD(T7j, T7k);
					     T7h = VSUB(T75, T7g);
					     T8w = VADD(T75, T7g);
					     ST(&(ri[WS(rs, 10)]), VFMA(LDK(KP923879532), T7B, T7y), ms, &(ri[0]));
					     ST(&(ri[WS(rs, 26)]), VFNMS(LDK(KP923879532), T7B, T7y), ms, &(ri[0]));
					     ST(&(ii[WS(rs, 18)]), VFNMS(LDK(KP923879532), T8o, T8n), ms, &(ii[0]));
					     ST(&(ii[WS(rs, 2)]), VFMA(LDK(KP923879532), T8o, T8n), ms, &(ii[0]));
					     ST(&(ii[WS(rs, 26)]), VFNMS(LDK(KP923879532), T8q, T8p), ms, &(ii[0]));
					     ST(&(ii[WS(rs, 10)]), VFMA(LDK(KP923879532), T8q, T8p), ms, &(ii[0]));
					     ST(&(ri[WS(rs, 2)]), VFMA(LDK(KP923879532), T7x, T7q), ms, &(ri[0]));
					     ST(&(ri[WS(rs, 18)]), VFNMS(LDK(KP923879532), T7x, T7q), ms, &(ri[0]));
					     ST(&(ri[WS(rs, 30)]), VFMA(LDK(KP923879532), T7l, T7i), ms, &(ri[0]));
					     ST(&(ri[WS(rs, 14)]), VFNMS(LDK(KP923879532), T7l, T7i), ms, &(ri[0]));
					     ST(&(ii[WS(rs, 22)]), VFNMS(LDK(KP923879532), T8u, T8t), ms, &(ii[0]));
					     ST(&(ii[WS(rs, 6)]), VFMA(LDK(KP923879532), T8u, T8t), ms, &(ii[0]));
					     ST(&(ii[WS(rs, 30)]), VFMA(LDK(KP923879532), T8w, T8v), ms, &(ii[0]));
					     ST(&(ii[WS(rs, 14)]), VFNMS(LDK(KP923879532), T8w, T8v), ms, &(ii[0]));
					     ST(&(ri[WS(rs, 6)]), VFMA(LDK(KP923879532), T7h, T6U), ms, &(ri[0]));
					     ST(&(ri[WS(rs, 22)]), VFNMS(LDK(KP923879532), T7h, T6U), ms, &(ri[0]));
					     T5C = VFMA(LDK(KP707106781), T3R, T3C);
					     T3S = VFNMS(LDK(KP707106781), T3R, T3C);
					}
				   }
				   T5D = VFMA(LDK(KP414213562), T3Z, T46);
				   T47 = VFNMS(LDK(KP414213562), T46, T3Z);
				   T4m = VFMA(LDK(KP414213562), T4l, T4e);
				   T5E = VFNMS(LDK(KP414213562), T4e, T4l);
				   T8N = VADD(T8y, T8x);
				   T8z = VSUB(T8x, T8y);
				   T8A = VADD(T5T, T5U);
				   T5V = VSUB(T5T, T5U);
				   T6d = VSUB(T5i, T5b);
				   T5j = VADD(T5b, T5i);
				   T5u = VADD(T5s, T5t);
				   T6g = VSUB(T5s, T5t);
				   T66 = VSUB(T4J, T4C);
				   T4K = VADD(T4C, T4J);
				   T8C = VADD(T47, T4m);
				   T4n = VSUB(T47, T4m);
				   T8H = VFNMS(LDK(KP707106781), T8A, T8z);
				   T8B = VFMA(LDK(KP707106781), T8A, T8z);
				   T8I = VSUB(T5E, T5D);
				   T5F = VADD(T5D, T5E);
				   T5L = VFMA(LDK(KP707106781), T5j, T54);
				   T5k = VFNMS(LDK(KP707106781), T5j, T54);
				   T5K = VFMA(LDK(KP707106781), T5u, T5r);
				   T5v = VFNMS(LDK(KP707106781), T5u, T5r);
				   T4L = VFNMS(LDK(KP707106781), T4K, T4v);
				   T5I = VFMA(LDK(KP707106781), T4K, T4v);