📄 n2sv_32.c
字号:
STM4(&(ro[30]), T66, ovs, &(ro[0])); T67 = VADD(T55, T56); STM4(&(io[14]), T67, ovs, &(io[0])); T68 = VADD(T57, T5a); STM4(&(ro[14]), T68, ovs, &(ro[0])); } { V T5f, T5r, T5u, T5w, T5m, T5q, T5p, T5v; { V T5d, T5e, T5s, T5t; T5d = VADD(T4r, T4s); T5e = VMUL(LDK(KP707106781), VADD(T53, T52)); T5f = VADD(T5d, T5e); T5r = VSUB(T5d, T5e); T5s = VFNMS(LDK(KP382683432), T5h, VMUL(LDK(KP923879532), T5g)); T5t = VFMA(LDK(KP923879532), T5k, VMUL(LDK(KP382683432), T5j)); T5u = VSUB(T5s, T5t); T5w = VADD(T5s, T5t); } { V T5i, T5l, T5n, T5o; T5i = VFMA(LDK(KP382683432), T5g, VMUL(LDK(KP923879532), T5h)); T5l = VFNMS(LDK(KP382683432), T5k, VMUL(LDK(KP923879532), T5j)); T5m = VADD(T5i, T5l); T5q = VSUB(T5l, T5i); T5n = VADD(T50, T4Z); T5o = VMUL(LDK(KP707106781), VADD(T4w, T4z)); T5p = VSUB(T5n, T5o); T5v = VADD(T5n, T5o); } T69 = VSUB(T5f, T5m); STM4(&(ro[18]), T69, ovs, &(ro[0])); T6a = VSUB(T5v, T5w); STM4(&(io[18]), T6a, ovs, &(io[0])); T6b = VADD(T5f, T5m); STM4(&(ro[2]), T6b, ovs, &(ro[0])); T6c = VADD(T5v, T5w); STM4(&(io[2]), T6c, ovs, &(io[0])); T6d = VSUB(T5p, T5q); STM4(&(io[26]), T6d, ovs, &(io[0])); T6e = VSUB(T5r, T5u); STM4(&(ro[26]), T6e, ovs, &(ro[0])); T6f = VADD(T5p, T5q); STM4(&(io[10]), T6f, ovs, &(io[0])); T6g = VADD(T5r, T5u); STM4(&(ro[10]), T6g, ovs, &(ro[0])); } { V T5z, T5P, T5S, T5U, T5K, T5O, T5N, T5T; { V T5x, T5y, T5Q, T5R; T5x = VSUB(T7, Te); T5y = VSUB(T1n, T1u); T5z = VADD(T5x, T5y); T5P = VSUB(T5x, T5y); T5Q = VSUB(T5D, T5A); T5R = VADD(T5F, T5I); T5S = VMUL(LDK(KP707106781), VSUB(T5Q, T5R)); T5U = VMUL(LDK(KP707106781), VADD(T5Q, T5R)); } { V T5E, T5J, T5L, T5M; T5E = VADD(T5A, T5D); T5J = VSUB(T5F, T5I); T5K = VMUL(LDK(KP707106781), VADD(T5E, T5J)); T5O = VMUL(LDK(KP707106781), VSUB(T5J, T5E)); T5L = VSUB(T18, T1f); T5M = VSUB(Tt, Tm); T5N = VSUB(T5L, T5M); T5T = VADD(T5M, T5L); } T6h = VSUB(T5z, T5K); STM4(&(ro[20]), T6h, ovs, &(ro[0])); T6i = VSUB(T5T, T5U); STM4(&(io[20]), T6i, ovs, &(io[0])); T6j = VADD(T5z, T5K); STM4(&(ro[4]), T6j, ovs, &(ro[0])); T6k = VADD(T5T, T5U); STM4(&(io[4]), T6k, ovs, &(io[0])); T6l = VSUB(T5N, T5O); STM4(&(io[28]), T6l, ovs, &(io[0])); T6m = VSUB(T5P, T5S); STM4(&(ro[28]), T6m, ovs, &(ro[0])); T6n = VADD(T5N, T5O); STM4(&(io[12]), T6n, ovs, &(io[0])); T6o = VADD(T5P, T5S); STM4(&(ro[12]), T6o, ovs, &(ro[0])); } { V Tv, T5V, T5Y, T60, T10, T11, T1w, T5Z; { V Tf, Tu, T5W, T5X; Tf = VADD(T7, Te); Tu = VADD(Tm, Tt); Tv = VADD(Tf, Tu); T5V = VSUB(Tf, Tu); T5W = VADD(T5B, T5C); T5X = VADD(T5G, T5H); T5Y = VSUB(T5W, T5X); T60 = VADD(T5W, T5X); } { V TK, TZ, T1g, T1v; TK = VADD(TC, TJ); TZ = VADD(TR, TY); T10 = VADD(TK, TZ); T11 = VSUB(TZ, TK); T1g = VADD(T18, T1f); T1v = VADD(T1n, T1u); T1w = VSUB(T1g, T1v); T5Z = VADD(T1g, T1v); } T6p = VSUB(Tv, T10); STM4(&(ro[16]), T6p, ovs, &(ro[0])); T6q = VSUB(T5Z, T60); STM4(&(io[16]), T6q, ovs, &(io[0])); T6r = VADD(Tv, T10); STM4(&(ro[0]), T6r, ovs, &(ro[0])); T6s = VADD(T5Z, T60); STM4(&(io[0]), T6s, ovs, &(io[0])); T6t = VADD(T11, T1w); STM4(&(io[8]), T6t, ovs, &(io[0])); T6u = VADD(T5V, T5Y); STM4(&(ro[8]), T6u, ovs, &(ro[0])); T6v = VSUB(T1w, T11); STM4(&(io[24]), T6v, ovs, &(io[0])); T6w = VSUB(T5V, T5Y); STM4(&(ro[24]), T6w, ovs, &(ro[0])); } { V T6x, T6y, T6z, T6A, T6B, T6C, T6D, T6E; { V T1X, T33, T31, T37, T2o, T34, T2P, T35; { V T1H, T1W, T2X, T30; T1H = VSUB(T1z, T1G); T1W = VSUB(T1O, T1V); T1X = VADD(T1H, T1W); T33 = VSUB(T1H, T1W); T2X = VSUB(T2T, T2W); T30 = VSUB(T2Y, T2Z); T31 = VSUB(T2X, T30); T37 = VADD(T2X, T30); } { V T2e, T2n, T2F, T2O; T2e = VSUB(T22, T2d); T2n = VSUB(T2j, T2m); T2o = VFMA(LDK(KP980785280), T2e, VMUL(LDK(KP195090322), T2n)); T34 = VFNMS(LDK(KP980785280), T2n, VMUL(LDK(KP195090322), T2e)); T2F = VSUB(T2t, T2E); T2O = VSUB(T2K, T2N); T2P = VFNMS(LDK(KP980785280), T2O, VMUL(LDK(KP195090322), T2F)); T35 = VFMA(LDK(KP195090322), T2O, VMUL(LDK(KP980785280), T2F)); } { V T2Q, T38, T32, T36; T2Q = VADD(T2o, T2P); T6x = VSUB(T1X, T2Q); STM4(&(ro[23]), T6x, ovs, &(ro[1])); T6y = VADD(T1X, T2Q); STM4(&(ro[7]), T6y, ovs, &(ro[1])); T38 = VADD(T34, T35); T6z = VSUB(T37, T38); STM4(&(io[23]), T6z, ovs, &(io[1])); T6A = VADD(T37, T38); STM4(&(io[7]), T6A, ovs, &(io[1])); T32 = VSUB(T2P, T2o); T6B = VSUB(T31, T32); STM4(&(io[31]), T6B, ovs, &(io[1])); T6C = VADD(T31, T32); STM4(&(io[15]), T6C, ovs, &(io[1])); T36 = VSUB(T34, T35); T6D = VSUB(T33, T36); STM4(&(ro[31]), T6D, ovs, &(ro[1])); T6E = VADD(T33, T36); STM4(&(ro[15]), T6E, ovs, &(ro[1])); } } { V T3D, T41, T3Z, T45, T3K, T42, T3R, T43; { V T3v, T3C, T3V, T3Y; T3v = VSUB(T3t, T3u); T3C = VSUB(T3y, T3B); T3D = VADD(T3v, T3C); T41 = VSUB(T3v, T3C); T3V = VSUB(T3T, T3U); T3Y = VSUB(T3W, T3X); T3Z = VSUB(T3V, T3Y); T45 = VADD(T3V, T3Y); } { V T3G, T3J, T3N, T3Q; T3G = VSUB(T3E, T3F); T3J = VSUB(T3H, T3I); T3K = VFMA(LDK(KP555570233), T3G, VMUL(LDK(KP831469612), T3J)); T42 = VFNMS(LDK(KP831469612), T3G, VMUL(LDK(KP555570233), T3J)); T3N = VSUB(T3L, T3M); T3Q = VSUB(T3O, T3P); T3R = VFNMS(LDK(KP831469612), T3Q, VMUL(LDK(KP555570233), T3N)); T43 = VFMA(LDK(KP831469612), T3N, VMUL(LDK(KP555570233), T3Q)); } { V T3S, T6F, T6G, T46, T6H, T6I; T3S = VADD(T3K, T3R); T6F = VSUB(T3D, T3S); STM4(&(ro[21]), T6F, ovs, &(ro[1])); STN4(&(ro[20]), T6h, T6F, T61, T6x, ovs); T6G = VADD(T3D, T3S); STM4(&(ro[5]), T6G, ovs, &(ro[1])); STN4(&(ro[4]), T6j, T6G, T63, T6y, ovs); T46 = VADD(T42, T43); T6H = VSUB(T45, T46); STM4(&(io[21]), T6H, ovs, &(io[1])); STN4(&(io[20]), T6i, T6H, T62, T6z, ovs); T6I = VADD(T45, T46); STM4(&(io[5]), T6I, ovs, &(io[1])); STN4(&(io[4]), T6k, T6I, T64, T6A, ovs); } { V T40, T6J, T6K, T44, T6L, T6M; T40 = VSUB(T3R, T3K); T6J = VSUB(T3Z, T40); STM4(&(io[29]), T6J, ovs, &(io[1])); STN4(&(io[28]), T6l, T6J, T65, T6B, ovs); T6K = VADD(T3Z, T40); STM4(&(io[13]), T6K, ovs, &(io[1])); STN4(&(io[12]), T6n, T6K, T67, T6C, ovs); T44 = VSUB(T42, T43); T6L = VSUB(T41, T44); STM4(&(ro[29]), T6L, ovs, &(ro[1])); STN4(&(ro[28]), T6m, T6L, T66, T6D, ovs); T6M = VADD(T41, T44); STM4(&(ro[13]), T6M, ovs, &(ro[1])); STN4(&(ro[12]), T6o, T6M, T68, T6E, ovs); } } } { V T6N, T6O, T6P, T6Q, T6R, T6S, T6T, T6U; { V T49, T4l, T4j, T4p, T4c, T4m, T4f, T4n; { V T47, T48, T4h, T4i; T47 = VADD(T3t, T3u); T48 = VADD(T3X, T3W); T49 = VADD(T47, T48); T4l = VSUB(T47, T48); T4h = VADD(T3T, T3U); T4i = VADD(T3y, T3B); T4j = VSUB(T4h, T4i); T4p = VADD(T4h, T4i); } { V T4a, T4b, T4d, T4e; T4a = VADD(T3E, T3F); T4b = VADD(T3H, T3I); T4c = VFMA(LDK(KP980785280), T4a, VMUL(LDK(KP195090322), T4b)); T4m = VFNMS(LDK(KP195090322), T4a, VMUL(LDK(KP980785280), T4b)); T4d = VADD(T3L, T3M); T4e = VADD(T3O, T3P); T4f = VFNMS(LDK(KP195090322), T4e, VMUL(LDK(KP980785280), T4d)); T4n = VFMA(LDK(KP195090322), T4d, VMUL(LDK(KP980785280), T4e)); } { V T4g, T4q, T4k, T4o; T4g = VADD(T4c, T4f); T6N = VSUB(T49, T4g); STM4(&(ro[17]), T6N, ovs, &(ro[1])); T6O = VADD(T49, T4g); STM4(&(ro[1]), T6O, ovs, &(ro[1])); T4q = VADD(T4m, T4n); T6P = VSUB(T4p, T4q); STM4(&(io[17]), T6P, ovs, &(io[1])); T6Q = VADD(T4p, T4q); STM4(&(io[1]), T6Q, ovs, &(io[1])); T4k = VSUB(T4f, T4c); T6R = VSUB(T4j, T4k); STM4(&(io[25]), T6R, ovs, &(io[1])); T6S = VADD(T4j, T4k); STM4(&(io[9]), T6S, ovs, &(io[1])); T4o = VSUB(T4m, T4n); T6T = VSUB(T4l, T4o); STM4(&(ro[25]), T6T, ovs, &(ro[1])); T6U = VADD(T4l, T4o); STM4(&(ro[9]), T6U, ovs, &(ro[1])); } } { V T3b, T3n, T3l, T3r, T3e, T3o, T3h, T3p; { V T39, T3a, T3j, T3k; T39 = VADD(T1z, T1G); T3a = VADD(T2Z, T2Y); T3b = VADD(T39, T3a); T3n = VSUB(T39, T3a); T3j = VADD(T2T, T2W); T3k = VADD(T1O, T1V); T3l = VSUB(T3j, T3k); T3r = VADD(T3j, T3k); } { V T3c, T3d, T3f, T3g; T3c = VADD(T22, T2d); T3d = VADD(T2j, T2m); T3e = VFMA(LDK(KP555570233), T3c, VMUL(LDK(KP831469612), T3d)); T3o = VFNMS(LDK(KP555570233), T3d, VMUL(LDK(KP831469612), T3c)); T3f = VADD(T2t, T2E); T3g = VADD(T2K, T2N); T3h = VFNMS(LDK(KP555570233), T3g, VMUL(LDK(KP831469612), T3f)); T3p = VFMA(LDK(KP831469612), T3g, VMUL(LDK(KP555570233), T3f)); } { V T3i, T6V, T6W, T3s, T6X, T6Y; T3i = VADD(T3e, T3h); T6V = VSUB(T3b, T3i); STM4(&(ro[19]), T6V, ovs, &(ro[1])); STN4(&(ro[16]), T6p, T6N, T69, T6V, ovs); T6W = VADD(T3b, T3i); STM4(&(ro[3]), T6W, ovs, &(ro[1])); STN4(&(ro[0]), T6r, T6O, T6b, T6W, ovs); T3s = VADD(T3o, T3p); T6X = VSUB(T3r, T3s); STM4(&(io[19]), T6X, ovs, &(io[1])); STN4(&(io[16]), T6q, T6P, T6a, T6X, ovs); T6Y = VADD(T3r, T3s); STM4(&(io[3]), T6Y, ovs, &(io[1])); STN4(&(io[0]), T6s, T6Q, T6c, T6Y, ovs); } { V T3m, T6Z, T70, T3q, T71, T72; T3m = VSUB(T3h, T3e); T6Z = VSUB(T3l, T3m); STM4(&(io[27]), T6Z, ovs, &(io[1])); STN4(&(io[24]), T6v, T6R, T6d, T6Z, ovs); T70 = VADD(T3l, T3m); STM4(&(io[11]), T70, ovs, &(io[1])); STN4(&(io[8]), T6t, T6S, T6f, T70, ovs); T3q = VSUB(T3o, T3p); T71 = VSUB(T3n, T3q); STM4(&(ro[27]), T71, ovs, &(ro[1])); STN4(&(ro[24]), T6w, T6T, T6e, T71, ovs); T72 = VADD(T3n, T3q); STM4(&(ro[11]), T72, ovs, &(ro[1])); STN4(&(ro[8]), T6u, T6U, T6g, T72, ovs); } } } } }}static const kdft_desc desc = { 32, "n2sv_32", {340, 52, 32, 0}, &GENUS, 0, 1, 0, 0 };void X(codelet_n2sv_32) (planner *p) { X(kdft_register) (p, n2sv_32, &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 + -