📄 n1_25.c
字号:
T1M = FMA(KP618033988, T1L, T1K); T1O = FNMS(KP618033988, T1K, T1L); { E T1V, T3f, T3m, T3n, T2W, T2Y, T32, T3g, T3h, T35, T3i, T2X; T3H = FMA(KP951056516, T1U, T1R); T1V = FNMS(KP951056516, T1U, T1R); T3f = FMA(KP951056516, T3e, T3b); T3r = FNMS(KP951056516, T3e, T3b); io[WS(os, 15)] = FNMS(KP951056516, T1O, T1N); io[WS(os, 10)] = FMA(KP951056516, T1O, T1N); io[WS(os, 20)] = FMA(KP951056516, T1M, T1J); io[WS(os, 5)] = FNMS(KP951056516, T1M, T1J); { E T30, T2a, T2p, T31, T33, T2F, T2U, T34, T2q, T2V; T30 = FMA(KP256756360, T22, T29); T2a = FNMS(KP256756360, T29, T22); T2p = FMA(KP634619297, T2o, T2h); T31 = FNMS(KP634619297, T2h, T2o); T33 = FMA(KP549754652, T2x, T2E); T2F = FNMS(KP549754652, T2E, T2x); T2U = FNMS(KP939062505, T2T, T2M); T34 = FMA(KP939062505, T2M, T2T); T3m = FNMS(KP871714437, T2p, T2a); T2q = FMA(KP871714437, T2p, T2a); T3n = FNMS(KP831864738, T2U, T2F); T2V = FMA(KP831864738, T2U, T2F); T2W = FMA(KP904730450, T2V, T2q); T2Y = FNMS(KP904730450, T2V, T2q); T32 = FNMS(KP871714437, T31, T30); T3g = FMA(KP871714437, T31, T30); T3h = FMA(KP831864738, T34, T33); T35 = FNMS(KP831864738, T34, T33); } io[WS(os, 1)] = FMA(KP968583161, T2W, T1V); T3i = FMA(KP904730450, T3h, T3g); T3k = FNMS(KP904730450, T3h, T3g); T36 = FMA(KP559154169, T35, T32); T38 = FNMS(KP683113946, T32, T35); ro[WS(os, 1)] = FMA(KP968583161, T3i, T3f); T2X = FNMS(KP242145790, T2W, T1V); T3o = FMA(KP559154169, T3n, T3m); T3q = FNMS(KP683113946, T3m, T3n); T3j = FNMS(KP242145790, T3i, T3f); T2Z = FMA(KP541454447, T2Y, T2X); T37 = FNMS(KP541454447, T2Y, T2X); } } } { E T47, T4R, T5A, T5w, T5y, T5E, T5G, T5z, T5t, T5x; { E T53, T5j, T5u, T5v, T5i, T5D, T5m, T5p, T5C, T3p, T3l, T5s, T5q, T5r; T47 = FMA(KP951056516, T46, T45); T53 = FNMS(KP951056516, T46, T45); T3p = FNMS(KP541454447, T3k, T3j); T3l = FMA(KP541454447, T3k, T3j); io[WS(os, 16)] = FNMS(KP833417178, T38, T37); io[WS(os, 11)] = FMA(KP833417178, T38, T37); io[WS(os, 21)] = FMA(KP921177326, T36, T2Z); io[WS(os, 6)] = FNMS(KP921177326, T36, T2Z); ro[WS(os, 11)] = FNMS(KP833417178, T3q, T3p); ro[WS(os, 16)] = FMA(KP833417178, T3q, T3p); ro[WS(os, 21)] = FNMS(KP921177326, T3o, T3l); ro[WS(os, 6)] = FMA(KP921177326, T3o, T3l); T5j = FMA(KP951056516, T4Q, T4P); T4R = FNMS(KP951056516, T4Q, T4P); { E T5k, T56, T59, T5l, T5n, T5d, T5g, T5o, T5a, T5h; T5k = FNMS(KP062914667, T54, T55); T56 = FMA(KP062914667, T55, T54); T59 = FMA(KP634619297, T58, T57); T5l = FNMS(KP634619297, T57, T58); T5n = FNMS(KP470564281, T5b, T5c); T5d = FMA(KP470564281, T5c, T5b); T5g = FMA(KP549754652, T5f, T5e); T5o = FNMS(KP549754652, T5e, T5f); T5u = FNMS(KP845997307, T59, T56); T5a = FMA(KP845997307, T59, T56); T5v = FNMS(KP968479752, T5g, T5d); T5h = FMA(KP968479752, T5g, T5d); T5i = FMA(KP906616052, T5h, T5a); T5A = FNMS(KP906616052, T5h, T5a); T5D = FNMS(KP845997307, T5l, T5k); T5m = FMA(KP845997307, T5l, T5k); T5p = FMA(KP968479752, T5o, T5n); T5C = FNMS(KP968479752, T5o, T5n); } ro[WS(os, 2)] = FMA(KP998026728, T5i, T53); T5s = FMA(KP906616052, T5p, T5m); T5q = FNMS(KP906616052, T5p, T5m); T5w = FNMS(KP560319534, T5v, T5u); T5y = FMA(KP681693190, T5u, T5v); T5E = FNMS(KP681693190, T5D, T5C); T5G = FMA(KP560319534, T5C, T5D); T5r = FMA(KP249506682, T5q, T5j); io[WS(os, 2)] = FNMS(KP998026728, T5q, T5j); T5z = FNMS(KP249506682, T5i, T53); T5t = FNMS(KP557913902, T5s, T5r); T5x = FMA(KP557913902, T5s, T5r); } { E T4W, T4M, T4O, T50, T52, T4V, T4F, T4N; { E T4Y, T4Z, T4C, T4E, T4I, T4T, T4S, T4L, T5F, T5B, T4U, T4D; T5F = FMA(KP557913902, T5A, T5z); T5B = FNMS(KP557913902, T5A, T5z); io[WS(os, 7)] = FMA(KP860541664, T5y, T5x); io[WS(os, 22)] = FNMS(KP860541664, T5y, T5x); io[WS(os, 17)] = FMA(KP949179823, T5w, T5t); io[WS(os, 12)] = FNMS(KP949179823, T5w, T5t); ro[WS(os, 12)] = FNMS(KP949179823, T5G, T5F); ro[WS(os, 17)] = FMA(KP949179823, T5G, T5F); ro[WS(os, 7)] = FNMS(KP860541664, T5E, T5B); ro[WS(os, 22)] = FMA(KP860541664, T5E, T5B); { E T4J, T4e, T4l, T4K, T4G, T4t, T4A, T4H, T4m, T4B; T4J = FNMS(KP062914667, T4a, T4d); T4e = FMA(KP062914667, T4d, T4a); T4l = FNMS(KP827271945, T4k, T4h); T4K = FMA(KP827271945, T4h, T4k); T4G = FNMS(KP126329378, T4p, T4s); T4t = FMA(KP126329378, T4s, T4p); T4A = FMA(KP939062505, T4z, T4w); T4H = FNMS(KP939062505, T4w, T4z); T4Y = FNMS(KP772036680, T4l, T4e); T4m = FMA(KP772036680, T4l, T4e); T4Z = FNMS(KP734762448, T4A, T4t); T4B = FMA(KP734762448, T4A, T4t); T4C = FMA(KP994076283, T4B, T4m); T4E = FNMS(KP994076283, T4B, T4m); T4I = FMA(KP734762448, T4H, T4G); T4T = FNMS(KP734762448, T4H, T4G); T4S = FMA(KP772036680, T4K, T4J); T4L = FNMS(KP772036680, T4K, T4J); } ro[WS(os, 3)] = FMA(KP998026728, T4C, T47); T4U = FMA(KP994076283, T4T, T4S); T4W = FNMS(KP994076283, T4T, T4S); T4M = FNMS(KP621716863, T4L, T4I); T4O = FMA(KP614372930, T4I, T4L); io[WS(os, 3)] = FNMS(KP998026728, T4U, T4R); T4D = FNMS(KP249506682, T4C, T47); T50 = FMA(KP614372930, T4Z, T4Y); T52 = FNMS(KP621716863, T4Y, T4Z); T4V = FMA(KP249506682, T4U, T4R); T4F = FNMS(KP557913902, T4E, T4D); T4N = FMA(KP557913902, T4E, T4D); } { E T3S, T3T, T3G, T41, T3K, T3N, T40, T51, T4X, T3Q, T3O, T3P; T51 = FMA(KP557913902, T4W, T4V); T4X = FNMS(KP557913902, T4W, T4V); ro[WS(os, 18)] = FNMS(KP949179823, T4O, T4N); ro[WS(os, 13)] = FMA(KP949179823, T4O, T4N); ro[WS(os, 8)] = FMA(KP943557151, T4M, T4F); ro[WS(os, 23)] = FNMS(KP943557151, T4M, T4F); io[WS(os, 8)] = FMA(KP943557151, T52, T51); io[WS(os, 23)] = FNMS(KP943557151, T52, T51); io[WS(os, 18)] = FNMS(KP949179823, T50, T4X); io[WS(os, 13)] = FMA(KP949179823, T50, T4X); { E T3I, T3u, T3x, T3J, T3L, T3B, T3E, T3M, T3y, T3F; T3I = FMA(KP126329378, T3s, T3t); T3u = FNMS(KP126329378, T3t, T3s); T3x = FNMS(KP470564281, T3w, T3v); T3J = FMA(KP470564281, T3v, T3w); T3L = FNMS(KP634619297, T3z, T3A); T3B = FMA(KP634619297, T3A, T3z); T3E = FNMS(KP827271945, T3D, T3C); T3M = FMA(KP827271945, T3C, T3D); T3S = FMA(KP912018591, T3x, T3u); T3y = FNMS(KP912018591, T3x, T3u); T3T = FMA(KP912575812, T3E, T3B); T3F = FNMS(KP912575812, T3E, T3B); T3G = FNMS(KP851038619, T3F, T3y); T3Y = FMA(KP851038619, T3F, T3y); T41 = FNMS(KP912018591, T3J, T3I); T3K = FMA(KP912018591, T3J, T3I); T3N = FMA(KP912575812, T3M, T3L); T40 = FNMS(KP912575812, T3M, T3L); } ro[WS(os, 4)] = FNMS(KP992114701, T3G, T3r); T3Q = FNMS(KP851038619, T3N, T3K); T3O = FMA(KP851038619, T3N, T3K); T3U = FNMS(KP525970792, T3T, T3S); T3W = FMA(KP726211448, T3S, T3T); T42 = FNMS(KP726211448, T41, T40); T44 = FMA(KP525970792, T40, T41); T3P = FMA(KP248028675, T3O, T3H); io[WS(os, 4)] = FNMS(KP992114701, T3O, T3H); T3X = FMA(KP248028675, T3G, T3r); T3R = FNMS(KP554608978, T3Q, T3P); T3V = FMA(KP554608978, T3Q, T3P); } } } } } T3Z = FMA(KP554608978, T3Y, T3X); T43 = FNMS(KP554608978, T3Y, T3X); io[WS(os, 9)] = FNMS(KP803003575, T3W, T3V); io[WS(os, 24)] = FMA(KP803003575, T3W, T3V); io[WS(os, 19)] = FNMS(KP943557151, T3U, T3R); io[WS(os, 14)] = FMA(KP943557151, T3U, T3R); ro[WS(os, 14)] = FNMS(KP943557151, T44, T43); ro[WS(os, 19)] = FMA(KP943557151, T44, T43); ro[WS(os, 24)] = FMA(KP803003575, T42, T3Z); ro[WS(os, 9)] = FNMS(KP803003575, T42, T3Z); }}static const kdft_desc desc = { 25, "n1_25", {84, 0, 268, 0}, &GENUS, 0, 0, 0, 0 };void X(codelet_n1_25) (planner *p) { X(kdft_register) (p, n1_25, &desc);}#else /* HAVE_FMA *//* Generated by: ../../../genfft/gen_notw -compact -variables 4 -pipeline-latency 4 -n 25 -name n1_25 -include n.h *//* * This function contains 352 FP additions, 184 FP multiplications, * (or, 260 additions, 92 multiplications, 92 fused multiply/add), * 101 stack variables, 20 constants, and 100 memory accesses */#include "n.h"static void n1_25(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs){ DK(KP425779291, +0.425779291565072648862502445744251703979973042); DK(KP904827052, +0.904827052466019527713668647932697593970413911); DK(KP637423989, +0.637423989748689710176712811676016195434917298); DK(KP770513242, +0.770513242775789230803009636396177847271667672); DK(KP998026728, +0.998026728428271561952336806863450553336905220); DK(KP062790519, +0.062790519529313376076178224565631133122484832); DK(KP992114701, +0.992114701314477831049793042785778521453036709); DK(KP125333233, +0.125333233564304245373118759816508793942918247); DK(KP684547105, +0.684547105928688673732283357621209269889519233); DK(KP728968627, +0.728968627421411523146730319055259111372571664); DK(KP481753674, +0.481753674101715274987191502872129653528542010); DK(KP876306680, +0.876306680043863587308115903922062583399064238); DK(KP844327925, +0.844327925502015078548558063966681505381659241); DK(KP535826794, +0.535826794978996618271308767867639978063575346); DK(KP248689887, +0.248689887164854788242283746006447968417567406); DK(KP968583161, +0.968583161128631119490168375464735813836012403); DK(KP250000000, +0.250000000000000000000000000000000000000000000); DK(KP559016994, +0.559016994374947424102293417182819058860154590); DK(KP587785252, +0.587785252292473129168705954639072768597652438); DK(KP951056516, +0.951056516295153572116439333379382143405698634); INT i; for (i = v; i > 0; i = i - 1, ri = ri + ivs, ii = ii + ivs, ro = ro + ovs, io = io + ovs, MAKE_VOLATILE_STRIDE(is), MAKE_VOLATILE_STRIDE(os)) { E T9, T4u, T2T, TP, T3H, TW, T5y, T3I, T2Q, T4v, Ti, Tr, Ts, T5m, T5n; E T5v, T18, T4G, T34, T3M, T1G, T4J, T38, T3T, T1v, T4K, T37, T3W, T1j, T4H; E T35, T3P, TB, TK, TL, T5p, T5q, T5w, T1T, T4N, T3c, T41, T2r, T4Q, T3e; E T4b, T2g, T4R, T3f, T48, T24, T4O, T3b, T44; { E T1, T4, T7, T8, T2S, T2R, TN, TO; T1 = ri[0]; { E T2, T3, T5, T6; T2 = ri[WS(is, 5)]; T3 = ri[WS(is, 20)]; T4 = T2 + T3; T5 = ri[WS(is, 10)]; T6 = ri[WS(is, 15)]; T7 = T5 + T6; T8 = T4 + T7; T2S = T5 - T6; T2R = T2 - T3; } T9 = T1 + T8; T4u = FNMS(KP587785252, T2R, KP951056516 * T2S); T2T = FMA(KP951056516, T2R, KP587785252 * T2S); TN = KP559016994 * (T4 - T7); TO = FNMS(KP250000000, T8, T1); TP = TN + TO; T3H = TO - TN; } { E T2N, T2K, T2L, TS, T2O, TV, T2M, T2P; T2N = ii[0]; { E TQ, TR, TT, TU; TQ = ii[WS(is, 5)]; TR = ii[WS(is, 20)]; T2K = TQ + TR; TT = ii[WS(is, 10)]; TU = ii[WS(is, 15)]; T2L = TT + TU; TS = TQ - TR; T2O = T2K + T2L; TV = TT - TU; } TW = FMA(KP951056516, TS, KP587785252 * TV); T5y = T2N + T2O; T3I = FNMS(KP587785252, TS, KP951056516 * TV); T2M = KP559016994 * (T2K - T2L); T2P = FNMS(KP250000000, T2O, T2N); T2Q = T2M + T2P; T4v = T2P - T2M; } { E Ta, T1c, Tj, T1z, Th, T1h, TY, T1g, T13, T1d, T16, T1b, Tq, T1E, T1l; E T1D, T1q, T1A, T1t, T1y; Ta = ri[WS(is, 1)]; T1c = ii[WS(is, 1)]; Tj = ri[WS(is, 4)]; T1z = ii[WS(is, 4)]; { E Tb, Tc, Td, Te, Tf, Tg; Tb = ri[WS(is, 6)]; Tc = ri[WS(is, 21)]; Td = Tb + Tc; Te = ri[WS(is, 11)]; Tf = ri[WS(is, 16)]; Tg = Te + Tf; Th = Td + Tg; T1h = Te - Tf; TY = KP559016994 * (Td - Tg); T1g = Tb - Tc; } { E T11, T12, T19, T14, T15, T1a; T11 = ii[WS(is, 6)]; T12 = ii[WS(is, 21)]; T19 = T11 + T12; T14 = ii[WS(is, 11)]; T15 = ii[WS(is, 16)]; T1a = T14 + T15; T13 = T11 - T12; T1d = T19 + T1a; T16 = T14 - T15; T1b = KP559016994 * (T19 - T1a); } { E Tk, Tl, Tm, Tn, To, Tp; Tk = ri[WS(is, 9)]; Tl = ri[WS(is, 24)]; Tm = Tk + Tl; Tn = ri[WS(is, 14)]; To = ri[WS(is, 19)]; Tp = Tn + To; Tq = Tm + Tp; T1E = Tn - To; T1l = KP559016994 * (Tm - Tp); T1D = Tk - Tl; } { E T1o, T1p, T1w, T1r, T1s, T1x; T1o = ii[WS(is, 9)]; T1p = ii[WS(is, 24)]; T1w = T1o + T1p; T1r = ii[WS(is, 14)]; T1s = ii[WS(is, 19)]; T1x = T1r + T1s; T1q = T1o - T1p; T1A = T1w + T1x; T1t = T1r - T1s; T1y = KP559016994 * (T1w - T1x); } Ti = Ta + Th; Tr = Tj + Tq; Ts = Ti + Tr; T5m = T1c + T1d; T5n = T1z + T1A; T5v = T5m + T5n; { E T17, T3L, T10, T3K, TZ; T17 = FMA(KP951056516, T13, KP587785252 * T16); T3L = FNMS(KP587785252, T13, KP951056516 * T16); TZ = FNMS(KP250000000, Th, Ta); T10 = TY + TZ; T3K = TZ - TY; T18 = T10 + T17; T4G = T3K + T3L; T34 = T10 - T17; T3M = T3K - T3L; } { E T1F, T3R, T1C, T3S, T1B; T1F = FMA(KP951056516, T1D, KP587785252 * T1E); T3R = FNMS(KP587785252, T1D, KP951056516 * T1E); T1B = FNMS(KP250000000, T1A, T1z); T1C = T1y + T1B; T3S = T1B - T1y; T1G = T1C - T1F; T4J = T3S - T3R; T38 = T1F + T1C; T3T = T3R + T3S; } { E T1u, T3V, T1n, T3U, T1m; T1u = FMA(KP951056516, T1q, KP587785252 * T1t); T3V = FNMS(KP587785252, T1q, KP951056516 * T1t); T1m = FNMS(KP250000000, Tq, Tj); T1n = T1l + T1m; T3U = T1m - T1l; T1v = T1n + T1u; T4K = T3U + T3V; T37 = T1n - T1u; T3W = T3U - T3V;
⌨️ 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 + -