t2sv_32.c

快速fft变换

					     T4M = VSUB(T4J, T4L);
					     T28 = VFMA(T26, T27, T25);
					     T4R = VFNMS(T26, T24, T4Q);
					     T3l = VFMA(T3j, T3k, T3h);
					     T5z = VFNMS(T3j, T3g, T5y);
					     T3E = VFMA(T3C, T3D, T3B);
					     T5Z = VFNMS(T3C, T3A, T5Y);
					}
					{
					     V T3o, T3s, T2j, T4N, T7q, T4U, T3p, T5A;
					     T3o = LD(&(ri[WS(rs, 15)]), ms, &(ri[WS(rs, 1)]));
					     T3s = LD(&(ii[WS(rs, 15)]), ms, &(ii[WS(rs, 1)]));
					     T2j = VADD(T28, T2i);
					     T4N = VSUB(T28, T2i);
					     T7q = VADD(T4R, T4T);
					     T4U = VSUB(T4R, T4T);
					     T3v = LD(&(ri[WS(rs, 7)]), ms, &(ri[WS(rs, 1)]));
					     T3x = LD(&(ii[WS(rs, 7)]), ms, &(ii[WS(rs, 1)]));
					     T3p = VMUL(T3n, T3o);
					     T5A = VMUL(T3n, T3s);
					     T4O = VSUB(T4M, T4N);
					     T6A = VADD(T4M, T4N);
					     T2k = VADD(T22, T2j);
					     T7o = VSUB(T22, T2j);
					     T6B = VSUB(T4P, T4U);
					     T4V = VADD(T4P, T4U);
					     T7r = VSUB(T7p, T7q);
					     T8e = VADD(T7p, T7q);
					     T3w = VMUL(TP, T3v);
					     T3t = VFMA(T3r, T3s, T3p);
					     T5B = VFNMS(T3r, T3o, T5A);
					     T5W = VMUL(TP, T3x);
					}
				   }
			      }
			      {
				   V T2t, T2q, T50, T2L, T5q, T2u, T2x, T2A, T2C;
				   {
					V T2n, T2p, T2G, T2K, T5V, T3u, T5C, T7G, T5X, T2o, T4Z, T2H, T5D, T3F, T5p;
					V T3y, T60, T7H;
					T2n = LD(&(ri[WS(rs, 1)]), ms, &(ri[WS(rs, 1)]));
					T2p = LD(&(ii[WS(rs, 1)]), ms, &(ii[WS(rs, 1)]));
					T2G = LD(&(ri[WS(rs, 25)]), ms, &(ri[WS(rs, 1)]));
					T2K = LD(&(ii[WS(rs, 25)]), ms, &(ii[WS(rs, 1)]));
					T3y = VFMA(TR, T3x, T3w);
					T5V = VSUB(T3l, T3t);
					T3u = VADD(T3l, T3t);
					T5C = VSUB(T5z, T5B);
					T7G = VADD(T5z, T5B);
					T5X = VFNMS(TR, T3v, T5W);
					T2o = VMUL(T2, T2n);
					T4Z = VMUL(T2, T2p);
					T2H = VMUL(T2F, T2G);
					T5D = VSUB(T3y, T3E);
					T3F = VADD(T3y, T3E);
					T5p = VMUL(T2F, T2K);
					T2t = LD(&(ri[WS(rs, 17)]), ms, &(ri[WS(rs, 1)]));
					T60 = VSUB(T5X, T5Z);
					T7H = VADD(T5X, T5Z);
					T2q = VFMA(T5, T2p, T2o);
					T50 = VFNMS(T5, T2n, T4Z);
					T2L = VFMA(T2J, T2K, T2H);
					T5E = VSUB(T5C, T5D);
					T6P = VADD(T5C, T5D);
					T3G = VADD(T3u, T3F);
					T7L = VSUB(T3u, T3F);
					T5q = VFNMS(T2J, T2G, T5p);
					T6M = VSUB(T5V, T60);
					T61 = VADD(T5V, T60);
					T8n = VADD(T7G, T7H);
					T7I = VSUB(T7G, T7H);
					T2u = VMUL(T2s, T2t);
					T2x = LD(&(ii[WS(rs, 17)]), ms, &(ii[WS(rs, 1)]));
					T2A = LD(&(ri[WS(rs, 9)]), ms, &(ri[WS(rs, 1)]));
					T2C = LD(&(ii[WS(rs, 9)]), ms, &(ii[WS(rs, 1)]));
				   }
				   {
					V T3N, T2z, T5m, T3K, T5G, T41, T5Q, T3O, T7v, T53, T2M, T54, T7w, T5r, T3R;
					V T3U, T3W;
					{
					     V T3H, T3J, T3Y, T40, T52, T2D, T5o;
					     T3H = LD(&(ri[WS(rs, 3)]), ms, &(ri[WS(rs, 1)]));
					     T3J = LD(&(ii[WS(rs, 3)]), ms, &(ii[WS(rs, 1)]));
					     T3Y = LD(&(ri[WS(rs, 11)]), ms, &(ri[WS(rs, 1)]));
					     T40 = LD(&(ii[WS(rs, 11)]), ms, &(ii[WS(rs, 1)]));
					     T3N = LD(&(ri[WS(rs, 19)]), ms, &(ri[WS(rs, 1)]));
					     {
						  V T2y, T51, T2B, T5n;
						  T2y = VFMA(T2w, T2x, T2u);
						  T51 = VMUL(T2s, T2x);
						  T2B = VMUL(T8, T2A);
						  T5n = VMUL(T8, T2C);
						  {
						       V T3I, T5F, T3Z, T5P;
						       T3I = VMUL(T3, T3H);
						       T5F = VMUL(T3, T3J);
						       T3Z = VMUL(Td, T3Y);
						       T5P = VMUL(Td, T40);
						       T2z = VADD(T2q, T2y);
						       T5m = VSUB(T2q, T2y);
						       T52 = VFNMS(T2w, T2t, T51);
						       T2D = VFMA(Tc, T2C, T2B);
						       T5o = VFNMS(Tc, T2A, T5n);
						       T3K = VFMA(T6, T3J, T3I);
						       T5G = VFNMS(T6, T3H, T5F);
						       T41 = VFMA(Th, T40, T3Z);
						       T5Q = VFNMS(Th, T3Y, T5P);
						       T3O = VMUL(T3M, T3N);
						  }
					     }
					     T7v = VADD(T50, T52);
					     T53 = VSUB(T50, T52);
					     T2M = VADD(T2D, T2L);
					     T54 = VSUB(T2D, T2L);
					     T7w = VADD(T5o, T5q);
					     T5r = VSUB(T5o, T5q);
					     T3R = LD(&(ii[WS(rs, 19)]), ms, &(ii[WS(rs, 1)]));
					     T3U = LD(&(ri[WS(rs, 27)]), ms, &(ri[WS(rs, 1)]));
					     T3W = LD(&(ii[WS(rs, 27)]), ms, &(ii[WS(rs, 1)]));
					}
					{
					     V T2O, T37, T39, T3T, T5K, T5I, T3X, T5O, T56, T38, T5g, T7M, T5J;
					     {
						  V T3S, T5H, T3V, T5N, T2P, T2Q;
						  T2O = LD(&(ri[WS(rs, 5)]), ms, &(ri[WS(rs, 1)]));
						  T55 = VSUB(T53, T54);
						  T6I = VADD(T53, T54);
						  T2N = VADD(T2z, T2M);
						  T7A = VSUB(T2z, T2M);
						  T5s = VADD(T5m, T5r);
						  T6F = VSUB(T5m, T5r);
						  T7x = VSUB(T7v, T7w);
						  T8i = VADD(T7v, T7w);
						  T3S = VFMA(T3Q, T3R, T3O);
						  T5H = VMUL(T3M, T3R);
						  T3V = VMUL(Te, T3U);
						  T5N = VMUL(Te, T3W);
						  T2P = VMUL(T29, T2O);
						  T2Q = LD(&(ii[WS(rs, 5)]), ms, &(ii[WS(rs, 1)]));
						  T37 = LD(&(ri[WS(rs, 13)]), ms, &(ri[WS(rs, 1)]));
						  T39 = LD(&(ii[WS(rs, 13)]), ms, &(ii[WS(rs, 1)]));
						  T3T = VADD(T3K, T3S);
						  T5K = VSUB(T3K, T3S);
						  T5I = VFNMS(T3Q, T3N, T5H);
						  T3X = VFMA(Ti, T3W, T3V);
						  T5O = VFNMS(Ti, T3U, T5N);
						  T2R = VFMA(T2b, T2Q, T2P);
						  T56 = VMUL(T29, T2Q);
						  T38 = VMUL(T1R, T37);
						  T5g = VMUL(T1R, T39);
					     }
					     T2U = LD(&(ri[WS(rs, 21)]), ms, &(ri[WS(rs, 1)]));
					     T7M = VADD(T5G, T5I);
					     T5J = VSUB(T5G, T5I);
					     {
						  V T42, T5M, T7N, T5R;
						  T42 = VADD(T3X, T41);
						  T5M = VSUB(T3X, T41);
						  T7N = VADD(T5O, T5Q);
						  T5R = VSUB(T5O, T5Q);
						  T57 = VFNMS(T2b, T2O, T56);
						  T3a = VFMA(T1U, T39, T38);
						  T5h = VFNMS(T1U, T37, T5g);
						  T62 = VADD(T5K, T5J);
						  T5L = VSUB(T5J, T5K);
						  T7J = VSUB(T42, T3T);
						  T43 = VADD(T3T, T42);
						  T63 = VSUB(T5M, T5R);
						  T5S = VADD(T5M, T5R);
						  T8o = VADD(T7M, T7N);
						  T7O = VSUB(T7M, T7N);
						  T2V = VMUL(T2T, T2U);
					     }
					     T2Y = LD(&(ii[WS(rs, 21)]), ms, &(ii[WS(rs, 1)]));
					     T32 = LD(&(ri[WS(rs, 29)]), ms, &(ri[WS(rs, 1)]));
					     T35 = LD(&(ii[WS(rs, 29)]), ms, &(ii[WS(rs, 1)]));
					}
				   }
			      }
			 }
		    }
		    {
			 V T5t, T5c, T5u, T5j, T8Z, T90;
			 {
			      V T7e, T8T, T8y, T7h, T8U, T8c, T8J, T44, T8u, T8q, T7y, T7D, T8w, T2m, T3d;
			      V T8h, T8R, T8P, T8k, T8x, T8B, T8f;
			      {
				   V T1i, T8O, T8N, T2l, T3c, T8j;
				   {
					V T8p, T5b, T30, T59, T36, T5f, TI, T1h, T8m, T5a, T7B;
					TI = VADD(Tq, TH);
					T7e = VSUB(Tq, TH);
					T8T = VSUB(T1g, TZ);
					T1h = VADD(TZ, T1g);
					T8y = VADD(T8n, T8o);
					T8p = VSUB(T8n, T8o);
					{
					     V T8C, T8I, T2Z, T58, T33, T5e;
					     T7h = VSUB(T7f, T7g);
					     T8C = VADD(T7f, T7g);
					     T8I = VADD(T8D, T8H);
					     T8U = VSUB(T8H, T8D);
					     T2Z = VFMA(T2X, T2Y, T2V);
					     T58 = VMUL(T2T, T2Y);
					     T33 = VMUL(T31, T32);
					     T5e = VMUL(T31, T35);
					     T1i = VADD(TI, T1h);
					     T8c = VSUB(TI, T1h);
					     T8O = VSUB(T8I, T8C);
					     T8J = VADD(T8C, T8I);
					     T5b = VSUB(T2R, T2Z);
					     T30 = VADD(T2R, T2Z);
					     T59 = VFNMS(T2X, T2U, T58);
					     T36 = VFMA(T34, T35, T33);
					     T5f = VFNMS(T34, T32, T5e);
					}
					T44 = VADD(T3G, T43);
					T8m = VSUB(T3G, T43);
					T5a = VSUB(T57, T59);
					T7B = VADD(T57, T59);
					{
					     V T5d, T3b, T5i, T7C;
					     T5d = VSUB(T36, T3a);
					     T3b = VADD(T36, T3a);
					     T5i = VSUB(T5f, T5h);
					     T7C = VADD(T5f, T5h);
					     T8N = VSUB(T2k, T1J);
					     T2l = VADD(T1J, T2k);
					     T8u = VADD(T8m, T8p);
					     T8q = VSUB(T8m, T8p);
					     T5t = VADD(T5b, T5a);
					     T5c = VSUB(T5a, T5b);
					     T7y = VSUB(T3b, T30);
					     T3c = VADD(T30, T3b);
					     T5u = VSUB(T5d, T5i);
					     T5j = VADD(T5d, T5i);
					     T8j = VADD(T7B, T7C);
					     T7D = VSUB(T7B, T7C);
					}
				   }
				   T8w = VSUB(T1i, T2l);
				   T2m = VADD(T1i, T2l);
				   T3d = VADD(T2N, T3c);
				   T8h = VSUB(T2N, T3c);
				   T8R = VSUB(T8O, T8N);
				   T8P = VADD(T8N, T8O);
				   T8k = VSUB(T8i, T8j);
				   T8x = VADD(T8i, T8j);
				   T8B = VADD(T8d, T8e);
				   T8f = VSUB(T8d, T8e);
			      }
			      {
				   V T7P, T7K, T7X, T7Y, T82, T7z, T7W, T7i, T8a, T86, T91, T8V, T8W, T7t, T7E;
				   V T81;
				   {
					V T84, T85, T7n, T7s, T8L, T45;
					T8L = VSUB(T44, T3d);
					T45 = VADD(T3d, T44);
					{
					     V T8t, T8l, T8A, T8z;
					     T8t = VSUB(T8k, T8h);
					     T8l = VADD(T8h, T8k);
					     T8A = VADD(T8x, T8y);
					     T8z = VSUB(T8x, T8y);
					     {
						  V T8M, T8K, T8s, T8g;
						  T8M = VSUB(T8J, T8B);
						  T8K = VADD(T8B, T8J);
						  T8s = VSUB(T8c, T8f);
						  T8g = VADD(T8c, T8f);
						  ST(&(ri[0]), VADD(T2m, T45), ms, &(ri[0]));
						  ST(&(ri[WS(rs, 16)]), VSUB(T2m, T45), ms, &(ri[0]));
						  {
						       V T8v, T8Q, T8S, T8r;
						       T8v = VSUB(T8t, T8u);
						       T8Q = VADD(T8t, T8u);
						       T8S = VSUB(T8q, T8l);
						       T8r = VADD(T8l, T8q);
						       ST(&(ri[WS(rs, 8)]), VADD(T8w, T8z), ms, &(ri[0]));
						       ST(&(ri[WS(rs, 24)]), VSUB(T8w, T8z), ms, &(ri[0]));
						       ST(&(ii[WS(rs, 24)]), VSUB(T8M, T8L), ms, &(ii[0]));
						       ST(&(ii[WS(rs, 8)]), VADD(T8L, T8M), ms, &(ii[0]));
						       ST(&(ii[WS(rs, 16)]), VSUB(T8K, T8A), ms, &(ii[0]));
						       ST(&(ii[0]), VADD(T8A, T8K), ms, &(ii[0]));
						       ST(&(ri[WS(rs, 12)]), VFMA(LDK(KP707106781), T8v, T8s), ms, &(ri[0]));
						       ST(&(ri[WS(rs, 28)]), VFNMS(LDK(KP707106781), T8v, T8s), ms, &(ri[0]));
						       ST(&(ii[WS(rs, 20)]), VFNMS(LDK(KP707106781), T8Q, T8P), ms, &(ii[0]));
						       ST(&(ii[WS(rs, 4)]), VFMA(LDK(KP707106781), T8Q, T8P), ms, &(ii[0]));
						       ST(&(ii[WS(rs, 28)]), VFNMS(LDK(KP707106781), T8S, T8R), ms, &(ii[0]));
						       ST(&(ii[WS(rs, 12)]), VFMA(LDK(KP707106781), T8S, T8R), ms, &(ii[0]));
						       ST(&(ri[WS(rs, 4)]), VFMA(LDK(KP707106781), T8r, T8g), ms, &(ri[0]));
						       ST(&(ri[WS(rs, 20)]), VFNMS(LDK(KP707106781), T8r, T8g), ms, &(ri[0]));
						  }
					     }
					}
					T7P = VSUB(T7L, T7O);
					T84 = VADD(T7L, T7O);
					T85 = VADD(T7I, T7J);
					T7K = VSUB(T7I, T7J);
					T7X = VADD(T7m, T7l);
					T7n = VSUB(T7l, T7m);
					T7s = VADD(T7o, T7r);
					T7Y = VSUB(T7o, T7r);
					T82 = VADD(T7x, T7y);
					T7z = VSUB(T7x, T7y);
					T7W = VADD(T7e, T7h);
					T7i = VSUB(T7e, T7h);
					T8a = VFMA(LDK(KP414213562), T84, T85);
					T86 = VFNMS(LDK(KP414213562), T85, T84);
					T91 = VSUB(T8U, T8T);
					T8V = VADD(T8T, T8U);
					T8W = VADD(T7n, T7s);
					T7t = VSUB(T7n, T7s);
					T7E = VSUB(T7A, T7D);
					T81 = VADD(T7A, T7D);
				   }
				   {
					V T7S, T7u, T7T, T7F, T92, T7Z, T89, T83, T7U, T7Q;
					T7S = VFNMS(LDK(KP707106781), T7t, T7i);
					T7u = VFMA(LDK(KP707106781), T7t, T7i);
					T7T = VFNMS(LDK(KP414213562), T7z, T7E);
					T7F = VFMA(LDK(KP414213562), T7E, T7z);
					T92 = VSUB(T7Y, T7X);
					T7Z = VADD(T7X, T7Y);
					T89 = VFNMS(LDK(KP414213562), T81, T82);
					T83 = VFMA(LDK(KP414213562), T82, T81);
					T7U = VFMA(LDK(KP414213562), T7K, T7P);
					T7Q = VFNMS(LDK(KP414213562), T7P, T7K);
					{
					     V T8X, T95, T93, T80, T88, T87, T7V, T94, T96, T7R, T8Y, T8b;
					     T8Z = VFNMS(LDK(KP707106781), T8W, T8V);
					     T8X = VFMA(LDK(KP707106781), T8W, T8V);
					     T95 = VFNMS(LDK(KP707106781), T92, T91);
					     T93 = VFMA(LDK(KP707106781), T92, T91);
					     T80 = VFMA(LDK(KP707106781), T7Z, T7W);
					     T88 = VFNMS(LDK(KP707106781), T7Z, T7W);
					     T90 = VSUB(T86, T83);
					     T87 = VADD(T83, T86);
					     T7V = VADD(T7T, T7U);
					     T94 = VSUB(T7U, T7T);
					     T96 = VADD(T7F, T7Q);
					     T7R = VSUB(T7F, T7Q);
					     T8Y = VADD(T89, T8a);
					     T8b = VSUB(T89, T8a);
					     ST(&(ri[WS(rs, 2)]), VFMA(LDK(KP923879532), T87, T80), ms, &(ri[0]));
					     ST(&(ri[WS(rs, 18)]), VFNMS(LDK(KP923879532), T87, T80), ms, &(ri[0]));
					     ST(&(ri[WS(rs, 30)]), VFMA(LDK(KP923879532), T7V, T7S), ms, &(ri[0]));
					     ST(&(ri[WS(rs, 14)]), VFNMS(LDK(KP923879532), T7V, T7S), ms, &(ri[0]));
					     ST(&(ii[WS(rs, 22)]), VFNMS(LDK(KP923879532), T94, T93), ms, &(ii[0]));
					     ST(&(ii[WS(rs, 6)]), VFMA(LDK(KP923879532), T94, T93), ms, &(ii[0]));
					     ST(&(ii[WS(rs, 30)]), VFMA(LDK(KP923879532), T96, T95), ms, &(ii[0]));
					     ST(&(ii[WS(rs, 14)]), VFNMS(LDK(KP923879532), T96, T95), ms, &(ii[0]));
					     ST(&(ri[WS(rs, 6)]), VFMA(LDK(KP923879532), T7R, T7u), ms, &(ri[0]));
					     ST(&(ri[WS(rs, 22)]), VFNMS(LDK(KP923879532), T7R, T7u), ms, &(ri[0]));
					     ST(&(ii[WS(rs, 18)]), VFNMS(LDK(KP923879532), T8Y, T8X), ms, &(ii[0]));
					     ST(&(ii[WS(rs, 2)]), VFMA(LDK(KP923879532), T8Y, T8X), ms, &(ii[0]));
					     ST(&(ri[WS(rs, 26)]), VFNMS(LDK(KP923879532), T8b, T88), ms, &(ri[0]));
					     ST(&(ri[WS(rs, 10)]), VFMA(LDK(KP923879532), T8b, T88), ms, &(ri[0]));
					}