📄 r2cb_64.c
字号:
R0[WS(rs, 26)] = FMA(KP1_847759065, T62, T61); R0[WS(rs, 2)] = FMA(KP1_847759065, T60, T5P); R0[WS(rs, 18)] = FNMS(KP1_847759065, T60, T5P); T4G = FNMS(KP414213562, T4F, T4A); T5c = FMA(KP414213562, T4A, T4F); } { E T5a, T59, T5h, T5i, T58, T4H; T5a = FMA(KP198912367, T4Y, T57); T58 = FNMS(KP198912367, T57, T4Y); T59 = FNMS(KP1_847759065, T4G, T4v); T4H = FMA(KP1_847759065, T4G, T4v); T5h = FMA(KP1_847759065, T5c, T5b); T5d = FNMS(KP1_847759065, T5c, T5b); T5i = FMA(KP668178637, T5e, T5f); T5g = FNMS(KP668178637, T5f, T5e); R0[WS(rs, 1)] = FMA(KP1_961570560, T58, T4H); R0[WS(rs, 17)] = FNMS(KP1_961570560, T58, T4H); R0[WS(rs, 29)] = FMA(KP1_662939224, T5i, T5h); R0[WS(rs, 13)] = FNMS(KP1_662939224, T5i, T5h); R0[WS(rs, 25)] = FMA(KP1_961570560, T5a, T59); R0[WS(rs, 9)] = FNMS(KP1_961570560, T5a, T59); } } } } } { E T43, T42, T46, T4a, T49, T3V, T3G, T47, T3P, T3v, T3X, T3T, T3U, T3N, T5B; E T5E; { E T5s, T5D, T5z, T5l, T5C, T5v, T5o, T5A; R0[WS(rs, 21)] = FMA(KP1_662939224, T5g, T5d); R0[WS(rs, 5)] = FNMS(KP1_662939224, T5g, T5d); T5s = FNMS(KP707106781, T5r, T5q); T5D = FMA(KP707106781, T5r, T5q); T5z = FMA(KP1_414213562, T5k, T5j); T5l = FNMS(KP1_414213562, T5k, T5j); T5C = FMA(KP707106781, T5u, T5t); T5v = FNMS(KP707106781, T5u, T5t); T5o = FNMS(KP414213562, T5n, T5m); T5A = FMA(KP414213562, T5m, T5n); { E T5y, T5x, T5F, T5G, T5w, T5p; T5y = FMA(KP668178637, T5s, T5v); T5w = FNMS(KP668178637, T5v, T5s); T5x = FMA(KP1_847759065, T5o, T5l); T5p = FNMS(KP1_847759065, T5o, T5l); T5F = FMA(KP1_847759065, T5A, T5z); T5B = FNMS(KP1_847759065, T5A, T5z); T5G = FMA(KP198912367, T5C, T5D); T5E = FNMS(KP198912367, T5D, T5C); R0[WS(rs, 3)] = FMA(KP1_662939224, T5w, T5p); R0[WS(rs, 19)] = FNMS(KP1_662939224, T5w, T5p); R0[WS(rs, 31)] = FMA(KP1_961570560, T5G, T5F); R0[WS(rs, 15)] = FNMS(KP1_961570560, T5G, T5F); R0[WS(rs, 27)] = FMA(KP1_662939224, T5y, T5x); R0[WS(rs, 11)] = FNMS(KP1_662939224, T5y, T5x); } } { E T3R, T3n, T3J, T3S, T3u, T3M; T3R = FMA(KP1_847759065, T3m, T3j); T3n = FNMS(KP1_847759065, T3m, T3j); R0[WS(rs, 23)] = FMA(KP1_961570560, T5E, T5B); R0[WS(rs, 7)] = FNMS(KP1_961570560, T5E, T5B); { E T3q, T3t, T3y, T3F; T43 = FMA(KP707106781, T3p, T3o); T3q = FNMS(KP707106781, T3p, T3o); T3t = FNMS(KP707106781, T3s, T3r); T42 = FMA(KP707106781, T3s, T3r); T46 = FMA(KP707106781, T3x, T3w); T3y = FNMS(KP707106781, T3x, T3w); T3F = T3B + T3E; T4a = T3B - T3E; T49 = FMA(KP707106781, T3I, T3H); T3J = FNMS(KP707106781, T3I, T3H); T3S = FMA(KP668178637, T3q, T3t); T3u = FNMS(KP668178637, T3t, T3q); T3V = FMA(KP923879532, T3F, T3y); T3G = FNMS(KP923879532, T3F, T3y); T3M = T3K - T3L; T47 = T3K + T3L; } T3P = FNMS(KP1_662939224, T3u, T3n); T3v = FMA(KP1_662939224, T3u, T3n); T3X = FMA(KP1_662939224, T3S, T3R); T3T = FNMS(KP1_662939224, T3S, T3R); T3U = FNMS(KP923879532, T3M, T3J); T3N = FMA(KP923879532, T3M, T3J); } { E T2X, T2W, T30, T34, T33, T2P, T2u, T31, T2J, T1P, T2R, T2N, T2O, T2H; { E T2L, T1n, T2D, T2M, T1O, T2G; T2L = FNMS(KP1_847759065, T1m, T1b); T1n = FMA(KP1_847759065, T1m, T1b); { E T3W, T3Y, T3Q, T3O; T3W = FNMS(KP534511135, T3V, T3U); T3Y = FMA(KP534511135, T3U, T3V); T3Q = FMA(KP303346683, T3G, T3N); T3O = FNMS(KP303346683, T3N, T3G); R1[WS(rs, 21)] = FMA(KP1_763842528, T3W, T3T); R1[WS(rs, 5)] = FNMS(KP1_763842528, T3W, T3T); R1[WS(rs, 29)] = FMA(KP1_763842528, T3Y, T3X); R1[WS(rs, 13)] = FNMS(KP1_763842528, T3Y, T3X); R1[WS(rs, 25)] = FMA(KP1_913880671, T3Q, T3P); R1[WS(rs, 9)] = FNMS(KP1_913880671, T3Q, T3P); R1[WS(rs, 1)] = FMA(KP1_913880671, T3O, T3v); R1[WS(rs, 17)] = FNMS(KP1_913880671, T3O, T3v); } { E T1E, T1N, T26, T2t; T2X = FNMS(KP707106781, T1D, T1s); T1E = FMA(KP707106781, T1D, T1s); T1N = FMA(KP707106781, T1M, T1J); T2W = FNMS(KP707106781, T1M, T1J); T30 = FNMS(KP707106781, T25, T1U); T26 = FMA(KP707106781, T25, T1U); T2t = T2h + T2s; T34 = T2s - T2h; T33 = FNMS(KP707106781, T2C, T2z); T2D = FMA(KP707106781, T2C, T2z); T2M = FMA(KP198912367, T1E, T1N); T1O = FNMS(KP198912367, T1N, T1E); T2P = FNMS(KP923879532, T2t, T26); T2u = FMA(KP923879532, T2t, T26); T2G = T2E + T2F; T31 = T2E - T2F; } T2J = FNMS(KP1_961570560, T1O, T1n); T1P = FMA(KP1_961570560, T1O, T1n); T2R = FMA(KP1_961570560, T2M, T2L); T2N = FNMS(KP1_961570560, T2M, T2L); T2O = FNMS(KP923879532, T2G, T2D); T2H = FMA(KP923879532, T2G, T2D); } { E T4j, T48, T4d, T45, T4l, T4h, T4i, T4b; { E T4f, T41, T4g, T44; T4f = FMA(KP1_847759065, T40, T3Z); T41 = FNMS(KP1_847759065, T40, T3Z); { E T2Q, T2S, T2K, T2I; T2Q = FNMS(KP820678790, T2P, T2O); T2S = FMA(KP820678790, T2O, T2P); T2K = FMA(KP098491403, T2u, T2H); T2I = FNMS(KP098491403, T2H, T2u); R1[WS(rs, 20)] = FMA(KP1_546020906, T2Q, T2N); R1[WS(rs, 4)] = FNMS(KP1_546020906, T2Q, T2N); R1[WS(rs, 28)] = FMA(KP1_546020906, T2S, T2R); R1[WS(rs, 12)] = FNMS(KP1_546020906, T2S, T2R); R1[WS(rs, 24)] = FMA(KP1_990369453, T2K, T2J); R1[WS(rs, 8)] = FNMS(KP1_990369453, T2K, T2J); R1[0] = FMA(KP1_990369453, T2I, T1P); R1[WS(rs, 16)] = FNMS(KP1_990369453, T2I, T1P); } T4g = FMA(KP198912367, T42, T43); T44 = FNMS(KP198912367, T43, T42); T4j = FMA(KP923879532, T47, T46); T48 = FNMS(KP923879532, T47, T46); T4d = FMA(KP1_961570560, T44, T41); T45 = FNMS(KP1_961570560, T44, T41); T4l = FMA(KP1_961570560, T4g, T4f); T4h = FNMS(KP1_961570560, T4g, T4f); T4i = FMA(KP923879532, T4a, T49); T4b = FNMS(KP923879532, T4a, T49); } { E T39, T2V, T3a, T2Y; T39 = FMA(KP1_847759065, T2U, T2T); T2V = FNMS(KP1_847759065, T2U, T2T); { E T4k, T4m, T4e, T4c; T4k = FNMS(KP098491403, T4j, T4i); T4m = FMA(KP098491403, T4i, T4j); T4e = FMA(KP820678790, T48, T4b); T4c = FNMS(KP820678790, T4b, T48); R1[WS(rs, 23)] = FMA(KP1_990369453, T4k, T4h); R1[WS(rs, 7)] = FNMS(KP1_990369453, T4k, T4h); R1[WS(rs, 31)] = FMA(KP1_990369453, T4m, T4l); R1[WS(rs, 15)] = FNMS(KP1_990369453, T4m, T4l); R1[WS(rs, 27)] = FMA(KP1_546020906, T4e, T4d); R1[WS(rs, 11)] = FNMS(KP1_546020906, T4e, T4d); R1[WS(rs, 3)] = FMA(KP1_546020906, T4c, T45); R1[WS(rs, 19)] = FNMS(KP1_546020906, T4c, T45); } T3a = FMA(KP668178637, T2W, T2X); T2Y = FNMS(KP668178637, T2X, T2W); T3d = FMA(KP923879532, T31, T30); T32 = FNMS(KP923879532, T31, T30); T37 = FMA(KP1_662939224, T2Y, T2V); T2Z = FNMS(KP1_662939224, T2Y, T2V); T3f = FMA(KP1_662939224, T3a, T39); T3b = FNMS(KP1_662939224, T3a, T39); T3c = FMA(KP923879532, T34, T33); T35 = FNMS(KP923879532, T34, T33); } } } } } } { E T3g, T3e, T36, T38; T3g = FMA(KP303346683, T3c, T3d); T3e = FNMS(KP303346683, T3d, T3c); T36 = FNMS(KP534511135, T35, T32); T38 = FMA(KP534511135, T32, T35); R1[WS(rs, 22)] = FMA(KP1_913880671, T3e, T3b); R1[WS(rs, 6)] = FNMS(KP1_913880671, T3e, T3b); R1[WS(rs, 30)] = FMA(KP1_913880671, T3g, T3f); R1[WS(rs, 14)] = FNMS(KP1_913880671, T3g, T3f); R1[WS(rs, 26)] = FMA(KP1_763842528, T38, T37); R1[WS(rs, 10)] = FNMS(KP1_763842528, T38, T37); R1[WS(rs, 2)] = FMA(KP1_763842528, T36, T2Z); R1[WS(rs, 18)] = FNMS(KP1_763842528, T36, T2Z); } }}static const kr2c_desc desc = { 64, "r2cb_64", {178, 0, 216, 0}, &GENUS };void X(codelet_r2cb_64) (planner *p) { X(kr2c_register) (p, r2cb_64, &desc);}#else /* HAVE_FMA *//* Generated by: ../../../genfft/gen_r2cb -compact -variables 4 -pipeline-latency 4 -sign 1 -n 64 -name r2cb_64 -include r2cb.h *//* * This function contains 394 FP additions, 134 FP multiplications, * (or, 342 additions, 82 multiplications, 52 fused multiply/add), * 110 stack variables, 19 constants, and 128 memory accesses */#include "r2cb.h"static void r2cb_64(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs){ DK(KP1_268786568, +1.268786568327290996430343226450986741351374190); DK(KP1_546020906, +1.546020906725473921621813219516939601942082586); DK(KP196034280, +0.196034280659121203988391127777283691722273346); DK(KP1_990369453, +1.990369453344393772489673906218959843150949737); DK(KP942793473, +0.942793473651995297112775251810508755314920638); DK(KP1_763842528, +1.763842528696710059425513727320776699016885241); DK(KP580569354, +0.580569354508924735272384751634790549382952557); DK(KP1_913880671, +1.913880671464417729871595773960539938965698411); DK(KP1_111140466, +1.111140466039204449485661627897065748749874382); DK(KP1_662939224, +1.662939224605090474157576755235811513477121624); DK(KP390180644, +0.390180644032256535696569736954044481855383236); DK(KP1_961570560, +1.961570560806460898252364472268478073947867462); DK(KP923879532, +0.923879532511286756128183189396788286822416626); DK(KP382683432, +0.382683432365089771728459984030398866761344562); DK(KP707106781, +0.707106781186547524400844362104849039284835938); DK(KP765366864, +0.765366864730179543456919968060797733522689125); DK(KP1_847759065, +1.847759065022573512256366378793576573644833252); DK(KP1_414213562, +1.414213562373095048801688724209698078569671875); DK(KP2_000000000, +2.000000000000000000000000000000000000000000000); INT i; for (i = v; i > 0; i = i - 1, R0 = R0 + ovs, R1 = R1 + ovs, Cr = Cr + ivs, Ci = Ci + ivs, MAKE_VOLATILE_STRIDE(rs), MAKE_VOLATILE_STRIDE(csr), MAKE_VOLATILE_STRIDE(csi)) { E Ta, T2S, T18, T2u, T3F, T4V, T5l, T61, Th, T2T, T1h, T2v, T3M, T4W, T5o; E T62, T3Q, T5q, T5u, T44, Tp, Tw, T2V, T2W, T2X, T2Y, T3X, T5t, T1r, T2x; E T41, T5r, T1A, T2y, T4a, T5y, T5N, T4H, TN, T31, T4E, T5z, T39, T3q, T1L; E T2B, T4h, T5M, T2h, T2F, T12, T36, T5D, T5J, T5G, T5K, T1U, T26, T23, T27; E T4p, T4z, T4w, T4A, T34, T3r; { E T5, T3A, T3, T3y, T9, T3C, T17, T3D, T6, T14; { E T4, T3z, T1, T2; T4 = Cr[WS(csr, 16)]; T5 = KP2_000000000 * T4; T3z = Ci[WS(csi, 16)]; T3A = KP2_000000000 * T3z; T1 = Cr[0]; T2 = Cr[WS(csr, 32)]; T3 = T1 + T2; T3y = T1 - T2; { E T7, T8, T15, T16; T7 = Cr[WS(csr, 8)]; T8 = Cr[WS(csr, 24)]; T9 = KP2_000000000 * (T7 + T8); T3C = T7 - T8; T15 = Ci[WS(csi, 8)]; T16 = Ci[WS(csi, 24)]; T17 = KP2_000000000 * (T15 - T16); T3D = T15 + T16; } } T6 = T3 + T5; Ta = T6 + T9; T2S = T6 - T9; T14 = T3 - T5; T18 = T14 - T17; T2u = T14 + T17; { E T3B, T3E, T5j, T5k; T3B = T3y - T3A; T3E = KP1_414213562 * (T3C - T3D); T3F = T3B + T3E; T4V = T3B - T3E; T5j = T3y + T3A; T5k = KP1_414213562 * (T3C + T3D); T5l = T5j - T5k; T61 = T5j + T5k; } } { E Td, T3G, T1c, T3K, Tg, T3J, T1f, T3H, T19, T1g; { E Tb, Tc, T1a, T1b; Tb = Cr[WS(csr, 4)]; Tc = Cr[WS(csr, 28)]; Td = Tb + Tc; T3G = Tb - Tc; T1a = Ci[WS(csi, 4)]; T1b = Ci[WS(csi, 28)]; T1c = T1a - T1b; T3K = T1a + T1b; } { E Te, Tf, T1d, T1e; Te = Cr[WS(csr, 20)]; Tf = Cr[WS(csr, 12)]; Tg = Te + Tf; T3J = Te - Tf; T1d = Ci[WS(csi, 20)]; T1e = Ci[WS(csi, 12)]; T1f = T1d - T1e; T3H = T1d + T1e; } Th = KP2_000000000 * (Td + Tg); T2T = KP2_000000000 * (T1f + T1c); T19 = Td - Tg; T1g = T1c - T1f; T1h = KP1_414213562 * (T19 - T1g); T2v = KP1_414213562 * (T19 + T1g); { E T3I, T3L, T5m, T5n; T3I = T3G - T3H; T3L = T3J + T3K; T3M = FNMS(KP765366864, T3L, KP1_847759065 * T3I); T4W = FMA(KP765366864, T3I, KP1_847759065 * T3L); T5m = T3G + T3H; T5n = T3K - T3J; T5o = FNMS(KP1_847759065, T5n, KP765366864 * T5m); T62 = FMA(KP1_847759065, T5m, KP765366864 * T5n); } } { E Tl, T3O, T1v, T43, To, T42, T1y, T3P, Ts, T3R, T1p, T3S, Tv, T3U, T1m; E T3V; { E Tj, Tk, T1t, T1u; Tj = Cr[WS(csr, 2)]; Tk = Cr[WS(csr, 30)]; Tl = Tj + Tk; T3O = Tj - Tk; T1t = Ci[WS(csi, 2)]; T1u = Ci[WS(csi, 30)]; T1v = T1t - T1u; T43 = T1t + T1u; } { E Tm, Tn, T1w, T1x; Tm = Cr[WS(csr, 18)]; Tn = Cr[WS(csr, 14)]; To = Tm + Tn; T42 = Tm - Tn; T1w = Ci[WS(csi, 18)]; T1x = Ci[WS(csi, 14)]; T1y = T1w - T1x; T3P = T1w + T1x; } { E Tq, Tr, T1n, T1o; Tq = Cr[WS(csr, 10)]; Tr = Cr[WS(csr, 22)]; Ts = Tq + Tr; T3R = Tq - Tr; T1n = Ci[WS(csi, 10)]; T1o = Ci[WS(csi, 22)]; T1p = T1n - T1o; T3S = T1n + T1o; } { E Tt, Tu, T1k, T1l; Tt = Cr[WS(csr, 6)]; Tu = Cr[WS(csr, 26)]; Tv = Tt + Tu; T3U = Tt - Tu; T1k = Ci[WS(csi, 26)]; T1l = Ci[WS(csi, 6)]; T1m = T1k - T1l; T3V = T1l + T1k; } T3Q = T3O - T3P; T5q = T3O + T3P; T5u = T43 - T42; T44 = T42 + T43; Tp = Tl + To; Tw = Ts + Tv; T2V = Tp - Tw; { E T3T, T3W, T1j, T1q; T2W = T1y + T1v; T2X = T1p + T1m; T2Y = T2W - T2X; T3T = T3R - T3S; T3W = T3U - T3V; T3X = KP707106781 * (T3T + T3W); T5t = KP707106781 * (T3T - T3W); T1j = Tl - To; T1q = T1m - T1p; T1r = T1j + T1q; T2x = T1j - T1q; { E T3Z, T40, T1s, T1z; T3Z = T3R + T3S; T40 = T3U + T3V; T41 = KP707106781 * (T3Z - T40); T5r = KP707106781 * (T3Z + T40); T1s = Ts - Tv; T1z = T1v - T1y; T1A = T1s + T1z; T2y = T1z - T1s; } } } { E TB, T48, T2c, T4G, TE, T4F, T2f, T49, TI, T4b, T1J, T4c, TL, T4e, T1G; E T4f; { E Tz, TA, T2a, T2b; Tz = Cr[WS(csr, 1)]; TA = Cr[WS(csr, 31)]; TB = Tz + TA; T48 = Tz - TA; T2a = Ci[WS(csi, 1)]; T2b = Ci[WS(csi, 31)]; T2c = T2a - T2b; T4G = T2a + T2b; } { E TC, TD, T2d, T2e; TC = Cr[WS(csr, 17)]; TD = Cr[WS(csr, 15)]; TE = TC + TD; T4F = TC - TD; T2d = Ci[WS(csi, 17)]; T2e = Ci[WS(csi, 15)]; T2f = T2d - T2e; T49 = T2d + T2e; } { E TG, TH, T1H, T1I; TG = Cr[WS(csr, 9)]; TH = Cr[WS(csr, 23)]; TI = TG + TH;
⌨️ 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 + -