📄 n1bv_64.c
字号:
T4X = VSUB(T4V, T4W); T4Y = VSUB(T4U, T4X); T5O = VADD(T4X, T4U); } { V T51, T52, T3c, T3d; T51 = VSUB(T4Z, T50); T52 = VMUL(LDK(KP707106781), VADD(T4Q, T4T)); T53 = VSUB(T51, T52); T5P = VADD(T51, T52); T3c = VADD(T1q, T1n); T3d = VADD(T1y, T1z); T3e = VADD(T3c, T3d); T3W = VSUB(T3d, T3c); } } { V T7h, T7l, T7k, T7m; { V T7f, T7g, T7i, T7j; T7f = VADD(T78, T79); T7g = VADD(T74, T75); T7h = VSUB(T7f, T7g); T7l = VADD(T7f, T7g); T7i = VADD(T6X, T6Y); T7j = VADD(T70, T71); T7k = VBYI(VSUB(T7i, T7j)); T7m = VADD(T7i, T7j); } ST(&(xo[WS(os, 48)]), VSUB(T7h, T7k), ovs, &(xo[0])); ST(&(xo[0]), VADD(T7l, T7m), ovs, &(xo[0])); ST(&(xo[WS(os, 16)]), VADD(T7h, T7k), ovs, &(xo[0])); ST(&(xo[WS(os, 32)]), VSUB(T7l, T7m), ovs, &(xo[0])); } { V T76, T7a, T73, T7b, T6Z, T72; T76 = VSUB(T74, T75); T7a = VSUB(T78, T79); T6Z = VSUB(T6X, T6Y); T72 = VSUB(T70, T71); T73 = VMUL(LDK(KP707106781), VSUB(T6Z, T72)); T7b = VMUL(LDK(KP707106781), VADD(T6Z, T72)); { V T77, T7c, T7d, T7e; T77 = VBYI(VSUB(T73, T76)); T7c = VSUB(T7a, T7b); ST(&(xo[WS(os, 24)]), VADD(T77, T7c), ovs, &(xo[0])); ST(&(xo[WS(os, 40)]), VSUB(T7c, T77), ovs, &(xo[0])); T7d = VBYI(VADD(T76, T73)); T7e = VADD(T7a, T7b); ST(&(xo[WS(os, 8)]), VADD(T7d, T7e), ovs, &(xo[0])); ST(&(xo[WS(os, 56)]), VSUB(T7e, T7d), ovs, &(xo[0])); } } { V T6C, T6S, T6I, T6P, T6r, T6Q, T6L, T6T, T6y, T6H; T6y = VMUL(LDK(KP707106781), VSUB(T6u, T6x)); T6C = VSUB(T6y, T6B); T6S = VADD(T6B, T6y); T6H = VMUL(LDK(KP707106781), VADD(T6u, T6x)); T6I = VSUB(T6G, T6H); T6P = VADD(T6G, T6H); { V T6j, T6q, T6J, T6K; T6j = VFNMS(LDK(KP382683432), T6i, VMUL(LDK(KP923879532), T6f)); T6q = VFMA(LDK(KP923879532), T6m, VMUL(LDK(KP382683432), T6p)); T6r = VSUB(T6j, T6q); T6Q = VADD(T6j, T6q); T6J = VFMA(LDK(KP382683432), T6f, VMUL(LDK(KP923879532), T6i)); T6K = VFNMS(LDK(KP382683432), T6m, VMUL(LDK(KP923879532), T6p)); T6L = VSUB(T6J, T6K); T6T = VADD(T6J, T6K); } { V T6D, T6M, T6V, T6W; T6D = VBYI(VSUB(T6r, T6C)); T6M = VSUB(T6I, T6L); ST(&(xo[WS(os, 20)]), VADD(T6D, T6M), ovs, &(xo[0])); ST(&(xo[WS(os, 44)]), VSUB(T6M, T6D), ovs, &(xo[0])); T6V = VSUB(T6P, T6Q); T6W = VBYI(VSUB(T6T, T6S)); ST(&(xo[WS(os, 36)]), VSUB(T6V, T6W), ovs, &(xo[0])); ST(&(xo[WS(os, 28)]), VADD(T6V, T6W), ovs, &(xo[0])); } { V T6N, T6O, T6R, T6U; T6N = VBYI(VADD(T6C, T6r)); T6O = VADD(T6I, T6L); ST(&(xo[WS(os, 12)]), VADD(T6N, T6O), ovs, &(xo[0])); ST(&(xo[WS(os, 52)]), VSUB(T6O, T6N), ovs, &(xo[0])); T6R = VADD(T6P, T6Q); T6U = VBYI(VADD(T6S, T6T)); ST(&(xo[WS(os, 60)]), VSUB(T6R, T6U), ovs, &(xo[0])); ST(&(xo[WS(os, 4)]), VADD(T6R, T6U), ovs, &(xo[0])); } } { V T5N, T68, T61, T69, T5U, T65, T5Y, T66; { V T5L, T5M, T5Z, T60; T5L = VADD(T4p, T4w); T5M = VADD(T5o, T5p); T5N = VSUB(T5L, T5M); T68 = VADD(T5L, T5M); T5Z = VFNMS(LDK(KP195090322), T5O, VMUL(LDK(KP980785280), T5P)); T60 = VFMA(LDK(KP195090322), T5R, VMUL(LDK(KP980785280), T5S)); T61 = VSUB(T5Z, T60); T69 = VADD(T5Z, T60); } { V T5Q, T5T, T5W, T5X; T5Q = VFMA(LDK(KP980785280), T5O, VMUL(LDK(KP195090322), T5P)); T5T = VFNMS(LDK(KP195090322), T5S, VMUL(LDK(KP980785280), T5R)); T5U = VSUB(T5Q, T5T); T65 = VADD(T5Q, T5T); T5W = VADD(T4E, T4L); T5X = VADD(T5u, T5r); T5Y = VSUB(T5W, T5X); T66 = VADD(T5X, T5W); } { V T5V, T62, T6b, T6c; T5V = VADD(T5N, T5U); T62 = VBYI(VADD(T5Y, T61)); ST(&(xo[WS(os, 50)]), VSUB(T5V, T62), ovs, &(xo[0])); ST(&(xo[WS(os, 14)]), VADD(T5V, T62), ovs, &(xo[0])); T6b = VBYI(VADD(T66, T65)); T6c = VADD(T68, T69); ST(&(xo[WS(os, 2)]), VADD(T6b, T6c), ovs, &(xo[0])); ST(&(xo[WS(os, 62)]), VSUB(T6c, T6b), ovs, &(xo[0])); } { V T63, T64, T67, T6a; T63 = VSUB(T5N, T5U); T64 = VBYI(VSUB(T61, T5Y)); ST(&(xo[WS(os, 46)]), VSUB(T63, T64), ovs, &(xo[0])); ST(&(xo[WS(os, 18)]), VADD(T63, T64), ovs, &(xo[0])); T67 = VBYI(VSUB(T65, T66)); T6a = VSUB(T68, T69); ST(&(xo[WS(os, 30)]), VADD(T67, T6a), ovs, &(xo[0])); ST(&(xo[WS(os, 34)]), VSUB(T6a, T67), ovs, &(xo[0])); } } { V T11, T2C, T2v, T2D, T2e, T2z, T2s, T2A; { V Tr, T10, T2t, T2u; Tr = VSUB(Tb, Tq); T10 = VSUB(TI, TZ); T11 = VSUB(Tr, T10); T2C = VADD(Tr, T10); T2t = VFNMS(LDK(KP471396736), T1s, VMUL(LDK(KP881921264), T1B)); T2u = VFMA(LDK(KP471396736), T23, VMUL(LDK(KP881921264), T2c)); T2v = VSUB(T2t, T2u); T2D = VADD(T2t, T2u); } { V T1C, T2d, T2i, T2r; T1C = VFMA(LDK(KP881921264), T1s, VMUL(LDK(KP471396736), T1B)); T2d = VFNMS(LDK(KP471396736), T2c, VMUL(LDK(KP881921264), T23)); T2e = VSUB(T1C, T2d); T2z = VADD(T1C, T2d); T2i = VSUB(T2g, T2h); T2r = VSUB(T2l, T2q); T2s = VSUB(T2i, T2r); T2A = VADD(T2r, T2i); } { V T2f, T2w, T2F, T2G; T2f = VADD(T11, T2e); T2w = VBYI(VADD(T2s, T2v)); ST(&(xo[WS(os, 53)]), VSUB(T2f, T2w), ovs, &(xo[WS(os, 1)])); ST(&(xo[WS(os, 11)]), VADD(T2f, T2w), ovs, &(xo[WS(os, 1)])); T2F = VBYI(VADD(T2A, T2z)); T2G = VADD(T2C, T2D); ST(&(xo[WS(os, 5)]), VADD(T2F, T2G), ovs, &(xo[WS(os, 1)])); ST(&(xo[WS(os, 59)]), VSUB(T2G, T2F), ovs, &(xo[WS(os, 1)])); } { V T2x, T2y, T2B, T2E; T2x = VSUB(T11, T2e); T2y = VBYI(VSUB(T2v, T2s)); ST(&(xo[WS(os, 43)]), VSUB(T2x, T2y), ovs, &(xo[WS(os, 1)])); ST(&(xo[WS(os, 21)]), VADD(T2x, T2y), ovs, &(xo[WS(os, 1)])); T2B = VBYI(VSUB(T2z, T2A)); T2E = VSUB(T2C, T2D); ST(&(xo[WS(os, 27)]), VADD(T2B, T2E), ovs, &(xo[WS(os, 1)])); ST(&(xo[WS(os, 37)]), VSUB(T2E, T2B), ovs, &(xo[WS(os, 1)])); } } { V T3n, T3O, T3J, T3R, T3y, T3Q, T3G, T3N; { V T3f, T3m, T3H, T3I; T3f = VFNMS(LDK(KP098017140), T3e, VMUL(LDK(KP995184726), T3b)); T3m = VFMA(LDK(KP995184726), T3i, VMUL(LDK(KP098017140), T3l)); T3n = VSUB(T3f, T3m); T3O = VADD(T3f, T3m); T3H = VFMA(LDK(KP098017140), T3b, VMUL(LDK(KP995184726), T3e)); T3I = VFNMS(LDK(KP098017140), T3i, VMUL(LDK(KP995184726), T3l)); T3J = VSUB(T3H, T3I); T3R = VADD(T3H, T3I); } { V T3u, T3x, T3C, T3F; T3u = VADD(T3q, T3t); T3x = VADD(T3v, T3w); T3y = VSUB(T3u, T3x); T3Q = VADD(T3x, T3u); T3C = VADD(T3A, T3B); T3F = VADD(T3D, T3E); T3G = VSUB(T3C, T3F); T3N = VADD(T3C, T3F); } { V T3z, T3K, T3T, T3U; T3z = VBYI(VSUB(T3n, T3y)); T3K = VSUB(T3G, T3J); ST(&(xo[WS(os, 17)]), VADD(T3z, T3K), ovs, &(xo[WS(os, 1)])); ST(&(xo[WS(os, 47)]), VSUB(T3K, T3z), ovs, &(xo[WS(os, 1)])); T3T = VSUB(T3N, T3O); T3U = VBYI(VSUB(T3R, T3Q)); ST(&(xo[WS(os, 33)]), VSUB(T3T, T3U), ovs, &(xo[WS(os, 1)])); ST(&(xo[WS(os, 31)]), VADD(T3T, T3U), ovs, &(xo[WS(os, 1)])); } { V T3L, T3M, T3P, T3S; T3L = VBYI(VADD(T3y, T3n)); T3M = VADD(T3G, T3J); ST(&(xo[WS(os, 15)]), VADD(T3L, T3M), ovs, &(xo[WS(os, 1)])); ST(&(xo[WS(os, 49)]), VSUB(T3M, T3L), ovs, &(xo[WS(os, 1)])); T3P = VADD(T3N, T3O); T3S = VBYI(VADD(T3Q, T3R)); ST(&(xo[WS(os, 63)]), VSUB(T3P, T3S), ovs, &(xo[WS(os, 1)])); ST(&(xo[WS(os, 1)]), VADD(T3P, T3S), ovs, &(xo[WS(os, 1)])); } } { V T4N, T5G, T5z, T5H, T5m, T5D, T5w, T5E; { V T4x, T4M, T5x, T5y; T4x = VSUB(T4p, T4w); T4M = VSUB(T4E, T4L); T4N = VSUB(T4x, T4M); T5G = VADD(T4x, T4M); T5x = VFNMS(LDK(KP555570233), T4Y, VMUL(LDK(KP831469612), T53)); T5y = VFMA(LDK(KP555570233), T5f, VMUL(LDK(KP831469612), T5k)); T5z = VSUB(T5x, T5y); T5H = VADD(T5x, T5y); } { V T54, T5l, T5q, T5v; T54 = VFMA(LDK(KP831469612), T4Y, VMUL(LDK(KP555570233), T53)); T5l = VFNMS(LDK(KP555570233), T5k, VMUL(LDK(KP831469612), T5f)); T5m = VSUB(T54, T5l); T5D = VADD(T54, T5l); T5q = VSUB(T5o, T5p); T5v = VSUB(T5r, T5u); T5w = VSUB(T5q, T5v); T5E = VADD(T5v, T5q); } { V T5n, T5A, T5J, T5K; T5n = VADD(T4N, T5m); T5A = VBYI(VADD(T5w, T5z)); ST(&(xo[WS(os, 54)]), VSUB(T5n, T5A), ovs, &(xo[0])); ST(&(xo[WS(os, 10)]), VADD(T5n, T5A), ovs, &(xo[0])); T5J = VBYI(VADD(T5E, T5D)); T5K = VADD(T5G, T5H); ST(&(xo[WS(os, 6)]), VADD(T5J, T5K), ovs, &(xo[0])); ST(&(xo[WS(os, 58)]), VSUB(T5K, T5J), ovs, &(xo[0])); } { V T5B, T5C, T5F, T5I; T5B = VSUB(T4N, T5m); T5C = VBYI(VSUB(T5z, T5w)); ST(&(xo[WS(os, 42)]), VSUB(T5B, T5C), ovs, &(xo[0])); ST(&(xo[WS(os, 22)]), VADD(T5B, T5C), ovs, &(xo[0])); T5F = VBYI(VSUB(T5D, T5E)); T5I = VSUB(T5G, T5H); ST(&(xo[WS(os, 26)]), VADD(T5F, T5I), ovs, &(xo[0])); ST(&(xo[WS(os, 38)]), VSUB(T5I, T5F), ovs, &(xo[0])); } } { V T2J, T34, T2X, T35, T2Q, T31, T2U, T32; { V T2H, T2I, T2V, T2W; T2H = VADD(Tb, Tq); T2I = VADD(T2g, T2h); T2J = VSUB(T2H, T2I); T34 = VADD(T2H, T2I); T2V = VFNMS(LDK(KP290284677), T2K, VMUL(LDK(KP956940335), T2L)); T2W = VFMA(LDK(KP290284677), T2N, VMUL(LDK(KP956940335), T2O)); T2X = VSUB(T2V, T2W); T35 = VADD(T2V, T2W); } { V T2M, T2P, T2S, T2T; T2M = VFMA(LDK(KP956940335), T2K, VMUL(LDK(KP290284677), T2L)); T2P = VFNMS(LDK(KP290284677), T2O, VMUL(LDK(KP956940335), T2N)); T2Q = VSUB(T2M, T2P); T31 = VADD(T2M, T2P); T2S = VADD(TI, TZ); T2T = VADD(T2q, T2l); T2U = VSUB(T2S, T2T); T32 = VADD(T2T, T2S); } { V T2R, T2Y, T37, T38; T2R = VADD(T2J, T2Q); T2Y = VBYI(VADD(T2U, T2X)); ST(&(xo[WS(os, 51)]), VSUB(T2R, T2Y), ovs, &(xo[WS(os, 1)])); ST(&(xo[WS(os, 13)]), VADD(T2R, T2Y), ovs, &(xo[WS(os, 1)])); T37 = VBYI(VADD(T32, T31)); T38 = VADD(T34, T35); ST(&(xo[WS(os, 3)]), VADD(T37, T38), ovs, &(xo[WS(os, 1)])); ST(&(xo[WS(os, 61)]), VSUB(T38, T37), ovs, &(xo[WS(os, 1)])); } { V T2Z, T30, T33, T36; T2Z = VSUB(T2J, T2Q); T30 = VBYI(VSUB(T2X, T2U)); ST(&(xo[WS(os, 45)]), VSUB(T2Z, T30), ovs, &(xo[WS(os, 1)])); ST(&(xo[WS(os, 19)]), VADD(T2Z, T30), ovs, &(xo[WS(os, 1)])); T33 = VBYI(VSUB(T31, T32)); T36 = VSUB(T34, T35); ST(&(xo[WS(os, 29)]), VADD(T33, T36), ovs, &(xo[WS(os, 1)])); ST(&(xo[WS(os, 35)]), VSUB(T36, T33), ovs, &(xo[WS(os, 1)])); } } { V T41, T4g, T4b, T4j, T44, T4i, T48, T4f; { V T3X, T40, T49, T4a; T3X = VFNMS(LDK(KP634393284), T3W, VMUL(LDK(KP773010453), T3V)); T40 = VFMA(LDK(KP773010453), T3Y, VMUL(LDK(KP634393284), T3Z)); T41 = VSUB(T3X, T40); T4g = VADD(T3X, T40); T49 = VFMA(LDK(KP634393284), T3V, VMUL(LDK(KP773010453), T3W)); T4a = VFNMS(LDK(KP634393284), T3Y, VMUL(LDK(KP773010453), T3Z)); T4b = VSUB(T49, T4a); T4j = VADD(T49, T4a); } { V T42, T43, T46, T47; T42 = VSUB(T3D, T3E); T43 = VSUB(T3w, T3v); T44 = VSUB(T42, T43); T4i = VADD(T43, T42); T46 = VSUB(T3A, T3B); T47 = VSUB(T3q, T3t); T48 = VSUB(T46, T47); T4f = VADD(T46, T47); } { V T45, T4c, T4l, T4m; T45 = VBYI(VSUB(T41, T44)); T4c = VSUB(T48, T4b); ST(&(xo[WS(os, 23)]), VADD(T45, T4c), ovs, &(xo[WS(os, 1)])); ST(&(xo[WS(os, 41)]), VSUB(T4c, T45), ovs, &(xo[WS(os, 1)])); T4l = VSUB(T4f, T4g); T4m = VBYI(VSUB(T4j, T4i)); ST(&(xo[WS(os, 39)]), VSUB(T4l, T4m), ovs, &(xo[WS(os, 1)])); ST(&(xo[WS(os, 25)]), VADD(T4l, T4m), ovs, &(xo[WS(os, 1)])); } { V T4d, T4e, T4h, T4k; T4d = VBYI(VADD(T44, T41)); T4e = VADD(T48, T4b); ST(&(xo[WS(os, 9)]), VADD(T4d, T4e), ovs, &(xo[WS(os, 1)])); ST(&(xo[WS(os, 55)]), VSUB(T4e, T4d), ovs, &(xo[WS(os, 1)])); T4h = VADD(T4f, T4g); T4k = VBYI(VADD(T4i, T4j)); ST(&(xo[WS(os, 57)]), VSUB(T4h, T4k), ovs, &(xo[WS(os, 1)])); ST(&(xo[WS(os, 7)]), VADD(T4h, T4k), ovs, &(xo[WS(os, 1)])); } } }}static const kdft_desc desc = { 64, "n1bv_64", {404, 72, 52, 0}, &GENUS, 0, 0, 0, 0 };void X(codelet_n1bv_64) (planner *p) { X(kdft_register) (p, n1bv_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 + -