📄 n2fv_64.c
字号:
T7F = VADD(T6b, T6c); STM2(&(xo[20]), T7F, ovs, &(xo[0])); T7G = VSUB(T6c, T6b); STM2(&(xo[108]), T7G, ovs, &(xo[0])); } { V T63, T64, T67, T6a; T63 = VSUB(T5N, T5U); T64 = VBYI(VSUB(T61, T5Y)); T7H = VSUB(T63, T64); STM2(&(xo[76]), T7H, ovs, &(xo[0])); T7I = VADD(T63, T64); STM2(&(xo[52]), T7I, ovs, &(xo[0])); T67 = VBYI(VSUB(T65, T66)); T6a = VSUB(T68, T69); T7J = VADD(T67, T6a); STM2(&(xo[44]), T7J, ovs, &(xo[0])); T7K = VSUB(T6a, T67); STM2(&(xo[84]), T7K, ovs, &(xo[0])); } } { V T7U, T7W, T7X, T7Z; { V T11, T2C, T2v, T2D, T2e, T2z, T2s, T2A; { V Tr, T10, T2t, T2u; Tr = VSUB(Tb, Tq); T10 = VSUB(TI, TZ); T11 = VADD(Tr, T10); T2C = VSUB(Tr, T10); T2t = VFNMS(LDK(KP634393284), T1B, VMUL(LDK(KP773010453), T1s)); T2u = VFMA(LDK(KP773010453), T2c, VMUL(LDK(KP634393284), T23)); T2v = VADD(T2t, T2u); T2D = VSUB(T2u, T2t); } { V T1C, T2d, T2o, T2r; T1C = VFMA(LDK(KP634393284), T1s, VMUL(LDK(KP773010453), T1B)); T2d = VFNMS(LDK(KP634393284), T2c, VMUL(LDK(KP773010453), T23)); T2e = VADD(T1C, T2d); T2z = VSUB(T2d, T1C); T2o = VSUB(T2i, T2n); T2r = VSUB(T2p, T2q); T2s = VADD(T2o, T2r); T2A = VSUB(T2r, T2o); } { V T2f, T2w, T7L, T7M; T2f = VADD(T11, T2e); T2w = VBYI(VADD(T2s, T2v)); T7L = VSUB(T2f, T2w); STM2(&(xo[114]), T7L, ovs, &(xo[2])); STN2(&(xo[112]), T7r, T7L, ovs); T7M = VADD(T2f, T2w); STM2(&(xo[14]), T7M, ovs, &(xo[2])); STN2(&(xo[12]), T7E, T7M, ovs); } { V T2F, T2G, T7N, T7O; T2F = VBYI(VADD(T2A, T2z)); T2G = VADD(T2C, T2D); T7N = VADD(T2F, T2G); STM2(&(xo[18]), T7N, ovs, &(xo[2])); STN2(&(xo[16]), T7s, T7N, ovs); T7O = VSUB(T2G, T2F); STM2(&(xo[110]), T7O, ovs, &(xo[2])); STN2(&(xo[108]), T7G, T7O, ovs); } { V T2x, T2y, T7P, T7Q; T2x = VSUB(T11, T2e); T2y = VBYI(VSUB(T2v, T2s)); T7P = VSUB(T2x, T2y); STM2(&(xo[78]), T7P, ovs, &(xo[2])); STN2(&(xo[76]), T7H, T7P, ovs); T7Q = VADD(T2x, T2y); STM2(&(xo[50]), T7Q, ovs, &(xo[2])); STN2(&(xo[48]), T7u, T7Q, ovs); } { V T2B, T2E, T7R, T7S; T2B = VBYI(VSUB(T2z, T2A)); T2E = VSUB(T2C, T2D); T7R = VADD(T2B, T2E); STM2(&(xo[46]), T7R, ovs, &(xo[2])); STN2(&(xo[44]), T7J, T7R, ovs); T7S = VSUB(T2E, T2B); STM2(&(xo[82]), T7S, ovs, &(xo[2])); STN2(&(xo[80]), T7t, T7S, ovs); } } { V T3j, T3Q, T3J, T3R, T3y, T3N, T3G, T3O; { V T3b, T3i, T3H, T3I; T3b = VADD(T39, T3a); T3i = VADD(T3e, T3h); T3j = VADD(T3b, T3i); T3Q = VSUB(T3b, T3i); T3H = VFNMS(LDK(KP290284677), T3m, VMUL(LDK(KP956940335), T3p)); T3I = VFMA(LDK(KP290284677), T3t, VMUL(LDK(KP956940335), T3w)); T3J = VADD(T3H, T3I); T3R = VSUB(T3I, T3H); } { V T3q, T3x, T3C, T3F; T3q = VFMA(LDK(KP956940335), T3m, VMUL(LDK(KP290284677), T3p)); T3x = VFNMS(LDK(KP290284677), T3w, VMUL(LDK(KP956940335), T3t)); T3y = VADD(T3q, T3x); T3N = VSUB(T3x, T3q); T3C = VADD(T3A, T3B); T3F = VADD(T3D, T3E); T3G = VADD(T3C, T3F); T3O = VSUB(T3F, T3C); } { V T3z, T3K, T7T, T3T, T3U, T7V; T3z = VADD(T3j, T3y); T3K = VBYI(VADD(T3G, T3J)); T7T = VSUB(T3z, T3K); STM2(&(xo[122]), T7T, ovs, &(xo[2])); STN2(&(xo[120]), T7v, T7T, ovs); T7U = VADD(T3z, T3K); STM2(&(xo[6]), T7U, ovs, &(xo[2])); T3T = VBYI(VADD(T3O, T3N)); T3U = VADD(T3Q, T3R); T7V = VADD(T3T, T3U); STM2(&(xo[26]), T7V, ovs, &(xo[2])); STN2(&(xo[24]), T7x, T7V, ovs); T7W = VSUB(T3U, T3T); STM2(&(xo[102]), T7W, ovs, &(xo[2])); } { V T3L, T3M, T7Y, T3P, T3S, T80; T3L = VSUB(T3j, T3y); T3M = VBYI(VSUB(T3J, T3G)); T7X = VSUB(T3L, T3M); STM2(&(xo[70]), T7X, ovs, &(xo[2])); T7Y = VADD(T3L, T3M); STM2(&(xo[58]), T7Y, ovs, &(xo[2])); STN2(&(xo[56]), T7A, T7Y, ovs); T3P = VBYI(VSUB(T3N, T3O)); T3S = VSUB(T3Q, T3R); T7Z = VADD(T3P, T3S); STM2(&(xo[38]), T7Z, ovs, &(xo[2])); T80 = VSUB(T3S, T3P); STM2(&(xo[90]), T80, ovs, &(xo[2])); STN2(&(xo[88]), T7C, T80, ovs); } } { V T81, T83, T86, T88; { V T4N, T5G, T5z, T5H, T5m, T5D, T5w, T5E; { V T4x, T4M, T5x, T5y; T4x = VADD(T4p, T4w); T4M = VADD(T4E, T4L); T4N = VADD(T4x, T4M); T5G = VSUB(T4x, T4M); T5x = VFNMS(LDK(KP195090322), T4Y, VMUL(LDK(KP980785280), T53)); T5y = VFMA(LDK(KP195090322), T5f, VMUL(LDK(KP980785280), T5k)); T5z = VADD(T5x, T5y); T5H = VSUB(T5y, T5x); } { V T54, T5l, T5s, T5v; T54 = VFMA(LDK(KP980785280), T4Y, VMUL(LDK(KP195090322), T53)); T5l = VFNMS(LDK(KP195090322), T5k, VMUL(LDK(KP980785280), T5f)); T5m = VADD(T54, T5l); T5D = VSUB(T5l, T54); T5s = VADD(T5q, T5r); T5v = VADD(T5t, T5u); T5w = VADD(T5s, T5v); T5E = VSUB(T5v, T5s); } { V T5n, T5A, T82, T5J, T5K, T84; T5n = VADD(T4N, T5m); T5A = VBYI(VADD(T5w, T5z)); T81 = VSUB(T5n, T5A); STM2(&(xo[124]), T81, ovs, &(xo[0])); T82 = VADD(T5n, T5A); STM2(&(xo[4]), T82, ovs, &(xo[0])); STN2(&(xo[4]), T82, T7U, ovs); T5J = VBYI(VADD(T5E, T5D)); T5K = VADD(T5G, T5H); T83 = VADD(T5J, T5K); STM2(&(xo[28]), T83, ovs, &(xo[0])); T84 = VSUB(T5K, T5J); STM2(&(xo[100]), T84, ovs, &(xo[0])); STN2(&(xo[100]), T84, T7W, ovs); } { V T5B, T5C, T85, T5F, T5I, T87; T5B = VSUB(T4N, T5m); T5C = VBYI(VSUB(T5z, T5w)); T85 = VSUB(T5B, T5C); STM2(&(xo[68]), T85, ovs, &(xo[0])); STN2(&(xo[68]), T85, T7X, ovs); T86 = VADD(T5B, T5C); STM2(&(xo[60]), T86, ovs, &(xo[0])); T5F = VBYI(VSUB(T5D, T5E)); T5I = VSUB(T5G, T5H); T87 = VADD(T5F, T5I); STM2(&(xo[36]), T87, ovs, &(xo[0])); STN2(&(xo[36]), T87, T7Z, ovs); T88 = VSUB(T5I, T5F); STM2(&(xo[92]), T88, ovs, &(xo[0])); } } { V T2J, T34, T2X, T35, T2Q, T31, T2U, T32; { V T2H, T2I, T2V, T2W; T2H = VADD(Tb, Tq); T2I = VADD(T2q, T2p); T2J = VADD(T2H, T2I); T34 = VSUB(T2H, T2I); T2V = VFNMS(LDK(KP098017140), T2L, VMUL(LDK(KP995184726), T2K)); T2W = VFMA(LDK(KP995184726), T2O, VMUL(LDK(KP098017140), T2N)); T2X = VADD(T2V, T2W); T35 = VSUB(T2W, T2V); } { V T2M, T2P, T2S, T2T; T2M = VFMA(LDK(KP098017140), T2K, VMUL(LDK(KP995184726), T2L)); T2P = VFNMS(LDK(KP098017140), T2O, VMUL(LDK(KP995184726), T2N)); T2Q = VADD(T2M, T2P); T31 = VSUB(T2P, T2M); T2S = VADD(T2n, T2i); T2T = VADD(TZ, TI); T2U = VADD(T2S, T2T); T32 = VSUB(T2T, T2S); } { V T2R, T2Y, T89, T8a; T2R = VADD(T2J, T2Q); T2Y = VBYI(VADD(T2U, T2X)); T89 = VSUB(T2R, T2Y); STM2(&(xo[126]), T89, ovs, &(xo[2])); STN2(&(xo[124]), T81, T89, ovs); T8a = VADD(T2R, T2Y); STM2(&(xo[2]), T8a, ovs, &(xo[2])); STN2(&(xo[0]), T7p, T8a, ovs); } { V T37, T38, T8b, T8c; T37 = VBYI(VADD(T32, T31)); T38 = VADD(T34, T35); T8b = VADD(T37, T38); STM2(&(xo[30]), T8b, ovs, &(xo[2])); STN2(&(xo[28]), T83, T8b, ovs); T8c = VSUB(T38, T37); STM2(&(xo[98]), T8c, ovs, &(xo[2])); STN2(&(xo[96]), T7q, T8c, ovs); } { V T2Z, T30, T8d, T8e; T2Z = VSUB(T2J, T2Q); T30 = VBYI(VSUB(T2X, T2U)); T8d = VSUB(T2Z, T30); STM2(&(xo[66]), T8d, ovs, &(xo[2])); STN2(&(xo[64]), T7n, T8d, ovs); T8e = VADD(T2Z, T30); STM2(&(xo[62]), T8e, ovs, &(xo[2])); STN2(&(xo[60]), T86, T8e, ovs); } { V T33, T36, T8f, T8g; T33 = VBYI(VSUB(T31, T32)); T36 = VSUB(T34, T35); T8f = VADD(T33, T36); STM2(&(xo[34]), T8f, ovs, &(xo[2])); STN2(&(xo[32]), T7o, T8f, ovs); T8g = VSUB(T36, T33); STM2(&(xo[94]), T8g, ovs, &(xo[2])); STN2(&(xo[92]), T88, T8g, ovs); } } { V T3X, T4i, T4b, T4j, T44, T4f, T48, T4g; { V T3V, T3W, T49, T4a; T3V = VSUB(T39, T3a); T3W = VSUB(T3E, T3D); T3X = VADD(T3V, T3W); T4i = VSUB(T3V, T3W); T49 = VFNMS(LDK(KP471396736), T3Y, VMUL(LDK(KP881921264), T3Z)); T4a = VFMA(LDK(KP471396736), T41, VMUL(LDK(KP881921264), T42)); T4b = VADD(T49, T4a); T4j = VSUB(T4a, T49); } { V T40, T43, T46, T47; T40 = VFMA(LDK(KP881921264), T3Y, VMUL(LDK(KP471396736), T3Z)); T43 = VFNMS(LDK(KP471396736), T42, VMUL(LDK(KP881921264), T41)); T44 = VADD(T40, T43); T4f = VSUB(T43, T40); T46 = VSUB(T3B, T3A); T47 = VSUB(T3h, T3e); T48 = VADD(T46, T47); T4g = VSUB(T47, T46); } { V T45, T4c, T8h, T8i; T45 = VADD(T3X, T44); T4c = VBYI(VADD(T48, T4b)); T8h = VSUB(T45, T4c); STM2(&(xo[118]), T8h, ovs, &(xo[2])); STN2(&(xo[116]), T7D, T8h, ovs); T8i = VADD(T45, T4c); STM2(&(xo[10]), T8i, ovs, &(xo[2])); STN2(&(xo[8]), T7w, T8i, ovs); } { V T4l, T4m, T8j, T8k; T4l = VBYI(VADD(T4g, T4f)); T4m = VADD(T4i, T4j); T8j = VADD(T4l, T4m); STM2(&(xo[22]), T8j, ovs, &(xo[2])); STN2(&(xo[20]), T7F, T8j, ovs); T8k = VSUB(T4m, T4l); STM2(&(xo[106]), T8k, ovs, &(xo[2])); STN2(&(xo[104]), T7y, T8k, ovs); } { V T4d, T4e, T8l, T8m; T4d = VSUB(T3X, T44); T4e = VBYI(VSUB(T4b, T48)); T8l = VSUB(T4d, T4e); STM2(&(xo[74]), T8l, ovs, &(xo[2])); STN2(&(xo[72]), T7z, T8l, ovs); T8m = VADD(T4d, T4e); STM2(&(xo[54]), T8m, ovs, &(xo[2])); STN2(&(xo[52]), T7I, T8m, ovs); } { V T4h, T4k, T8n, T8o; T4h = VBYI(VSUB(T4f, T4g)); T4k = VSUB(T4i, T4j); T8n = VADD(T4h, T4k); STM2(&(xo[42]), T8n, ovs, &(xo[2])); STN2(&(xo[40]), T7B, T8n, ovs); T8o = VSUB(T4k, T4h); STM2(&(xo[86]), T8o, ovs, &(xo[2])); STN2(&(xo[84]), T7K, T8o, ovs); } } } } } }}static const kdft_desc desc = { 64, "n2fv_64", {404, 72, 52, 0}, &GENUS, 0, 2, 0, 0 };void X(codelet_n2fv_64) (planner *p) { X(kdft_register) (p, n2fv_64, &desc);}#endif /* HAVE_FMA */
⌨️ 快捷键说明
复制代码
Ctrl + C
搜索代码
Ctrl + F
全屏模式
F11
切换主题
Ctrl + Shift + D
显示快捷键
?
增大字号
Ctrl + =
减小字号
Ctrl + -