📄 r2cf_25.c
字号:
} Cr[WS(csr, 3)] = FMA(KP992114701, T2V, T2P); { E T30, T34, T2S, T31, T35; T30 = FMA(KP855719849, T2Z, T2W); T34 = FNMS(KP923225144, T2R, T2Q); T2S = FMA(KP923225144, T2R, T2Q); Cr[WS(csr, 8)] = FNMS(KP897376177, T30, T2P); T31 = FNMS(KP237294955, T2S, T2P); Cr[WS(csr, 2)] = FMA(KP949179823, T2S, T2P); T35 = FNMS(KP997675361, T34, T33); { E T37, T36, T38, T2L; T37 = FNMS(KP904508497, T34, T32); T36 = FMA(KP560319534, T35, T32); T38 = FNMS(KP681693190, T37, T33); Cr[WS(csr, 12)] = FNMS(KP949179823, T36, T31); Cr[WS(csr, 7)] = FNMS(KP860541664, T38, T31); T2O = FNMS(KP809385824, T2K, T2I); T2L = FNMS(KP447417479, T2K, T2J); T2M = FNMS(KP690983005, T2L, T2I); } } Cr[WS(csr, 4)] = FNMS(KP992114701, T2O, T2v); } } } } T2y = FNMS(KP904730450, T2x, T2w); T2B = FMA(KP904730450, T2x, T2w); T2N = FNMS(KP999544308, T2M, T2H); { E T2z, T2D, T2F, T2E, T2G; T2z = FNMS(KP242145790, T2y, T2v); Cr[WS(csr, 1)] = FMA(KP968583161, T2y, T2v); T2D = FMA(KP904730450, T2C, T2B); T2F = T2A + T2B; Cr[WS(csr, 9)] = FNMS(KP803003575, T2N, T2v); T2E = FNMS(KP618033988, T2D, T2A); T2G = FMA(KP683113946, T2F, T2C); Cr[WS(csr, 6)] = FNMS(KP876091699, T2E, T2z); Cr[WS(csr, 11)] = FNMS(KP792626838, T2G, T2z); } }}static const kr2c_desc desc = { 25, "r2cf_25", {44, 12, 156, 0}, &GENUS };void X(codelet_r2cf_25) (planner *p) { X(kr2c_register) (p, r2cf_25, &desc);}#else /* HAVE_FMA *//* Generated by: ../../../genfft/gen_r2cf -compact -variables 4 -pipeline-latency 4 -n 25 -name r2cf_25 -include r2cf.h *//* * This function contains 200 FP additions, 140 FP multiplications, * (or, 117 additions, 57 multiplications, 83 fused multiply/add), * 101 stack variables, 40 constants, and 50 memory accesses */#include "r2cf.h"static void r2cf_25(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs){ DK(KP998026728, +0.998026728428271561952336806863450553336905220); DK(KP125581039, +0.125581039058626752152356449131262266244969664); DK(KP1_996053456, +1.996053456856543123904673613726901106673810439); DK(KP062790519, +0.062790519529313376076178224565631133122484832); DK(KP809016994, +0.809016994374947424102293417182819058860154590); DK(KP309016994, +0.309016994374947424102293417182819058860154590); DK(KP1_369094211, +1.369094211857377347464566715242418539779038465); DK(KP728968627, +0.728968627421411523146730319055259111372571664); DK(KP963507348, +0.963507348203430549974383005744259307057084020); DK(KP876306680, +0.876306680043863587308115903922062583399064238); DK(KP497379774, +0.497379774329709576484567492012895936835134813); DK(KP968583161, +0.968583161128631119490168375464735813836012403); DK(KP684547105, +0.684547105928688673732283357621209269889519233); DK(KP1_457937254, +1.457937254842823046293460638110518222745143328); DK(KP481753674, +0.481753674101715274987191502872129653528542010); DK(KP1_752613360, +1.752613360087727174616231807844125166798128477); DK(KP248689887, +0.248689887164854788242283746006447968417567406); DK(KP1_937166322, +1.937166322257262238980336750929471627672024806); DK(KP992114701, +0.992114701314477831049793042785778521453036709); DK(KP250666467, +0.250666467128608490746237519633017587885836494); DK(KP425779291, +0.425779291565072648862502445744251703979973042); DK(KP1_809654104, +1.809654104932039055427337295865395187940827822); DK(KP1_274847979, +1.274847979497379420353425623352032390869834596); DK(KP770513242, +0.770513242775789230803009636396177847271667672); DK(KP844327925, +0.844327925502015078548558063966681505381659241); DK(KP1_071653589, +1.071653589957993236542617535735279956127150691); DK(KP125333233, +0.125333233564304245373118759816508793942918247); DK(KP1_984229402, +1.984229402628955662099586085571557042906073418); DK(KP904827052, +0.904827052466019527713668647932697593970413911); DK(KP851558583, +0.851558583130145297725004891488503407959946084); DK(KP637423989, +0.637423989748689710176712811676016195434917298); DK(KP1_541026485, +1.541026485551578461606019272792355694543335344); DK(KP535826794, +0.535826794978996618271308767867639978063575346); DK(KP1_688655851, +1.688655851004030157097116127933363010763318483); DK(KP293892626, +0.293892626146236564584352977319536384298826219); DK(KP475528258, +0.475528258147576786058219666689691071702849317); 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, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(rs), MAKE_VOLATILE_STRIDE(csr), MAKE_VOLATILE_STRIDE(csi)) { E T8, T1j, T1V, T1l, T7, T9, Ta, T12, T2u, T1O, T19, T1P, Ti, T2r, T1K; E Tp, T1L, Tx, T2q, T1H, TE, T1I, TN, T2t, T1R, TU, T1S, T6, T1k, T3; E T2s, T2v; T8 = R0[0]; { E T4, T5, T1, T2; T4 = R0[WS(rs, 5)]; T5 = R1[WS(rs, 7)]; T6 = T4 + T5; T1k = T4 - T5; T1 = R1[WS(rs, 2)]; T2 = R0[WS(rs, 10)]; T3 = T1 + T2; T1j = T1 - T2; } T1V = KP951056516 * T1k; T1l = FMA(KP951056516, T1j, KP587785252 * T1k); T7 = KP559016994 * (T3 - T6); T9 = T3 + T6; Ta = FNMS(KP250000000, T9, T8); { E T16, T13, T14, TY, T17, T11, T15, T18; T16 = R1[WS(rs, 1)]; { E TW, TX, TZ, T10; TW = R0[WS(rs, 4)]; TX = R1[WS(rs, 11)]; T13 = TW + TX; TZ = R1[WS(rs, 6)]; T10 = R0[WS(rs, 9)]; T14 = TZ + T10; TY = TW - TX; T17 = T13 + T14; T11 = TZ - T10; } T12 = FMA(KP475528258, TY, KP293892626 * T11); T2u = T16 + T17; T1O = FNMS(KP293892626, TY, KP475528258 * T11); T15 = KP559016994 * (T13 - T14); T18 = FNMS(KP250000000, T17, T16); T19 = T15 + T18; T1P = T18 - T15; } { E Tm, Tj, Tk, Te, Tn, Th, Tl, To; Tm = R1[0]; { E Tc, Td, Tf, Tg; Tc = R0[WS(rs, 3)]; Td = R1[WS(rs, 10)]; Tj = Tc + Td; Tf = R1[WS(rs, 5)]; Tg = R0[WS(rs, 8)]; Tk = Tf + Tg; Te = Tc - Td; Tn = Tj + Tk; Th = Tf - Tg; } Ti = FMA(KP475528258, Te, KP293892626 * Th); T2r = Tm + Tn; T1K = FNMS(KP293892626, Te, KP475528258 * Th); Tl = KP559016994 * (Tj - Tk); To = FNMS(KP250000000, Tn, Tm); Tp = Tl + To; T1L = To - Tl; } { E TB, Ty, Tz, Tt, TC, Tw, TA, TD; TB = R0[WS(rs, 2)]; { E Tr, Ts, Tu, Tv; Tr = R1[WS(rs, 4)]; Ts = R0[WS(rs, 12)]; Ty = Tr + Ts; Tu = R0[WS(rs, 7)]; Tv = R1[WS(rs, 9)]; Tz = Tu + Tv; Tt = Tr - Ts; TC = Ty + Tz; Tw = Tu - Tv; } Tx = FMA(KP475528258, Tt, KP293892626 * Tw); T2q = TB + TC; T1H = FNMS(KP293892626, Tt, KP475528258 * Tw); TA = KP559016994 * (Ty - Tz); TD = FNMS(KP250000000, TC, TB); TE = TA + TD; T1I = TD - TA; } { E TR, TO, TP, TJ, TS, TM, TQ, TT; TR = R0[WS(rs, 1)]; { E TH, TI, TK, TL; TH = R1[WS(rs, 3)]; TI = R0[WS(rs, 11)]; TO = TH + TI; TK = R0[WS(rs, 6)]; TL = R1[WS(rs, 8)]; TP = TK + TL; TJ = TH - TI; TS = TO + TP; TM = TK - TL; } TN = FMA(KP475528258, TJ, KP293892626 * TM); T2t = TR + TS; T1R = FNMS(KP293892626, TJ, KP475528258 * TM); TQ = KP559016994 * (TO - TP); TT = FNMS(KP250000000, TS, TR); TU = TQ + TT; T1S = TT - TQ; } T2s = T2q - T2r; T2v = T2t - T2u; Ci[WS(csi, 5)] = FNMS(KP587785252, T2v, KP951056516 * T2s); Ci[WS(csi, 10)] = FMA(KP587785252, T2s, KP951056516 * T2v); { E T2z, T2y, T2A, T2w, T2x, T2B; T2z = T8 + T9; T2w = T2r + T2q; T2x = T2t + T2u; T2y = KP559016994 * (T2w - T2x); T2A = T2w + T2x; Cr[0] = T2z + T2A; T2B = FNMS(KP250000000, T2A, T2z); Cr[WS(csr, 5)] = T2y + T2B; Cr[WS(csr, 10)] = T2B - T2y; } { E Tb, Tq, TF, TG, T1E, T1F, T1G, T1B, T1C, T1D, TV, T1a, T1b, T1o, T1r; E T1s, T1z, T1x, T1e, T1h, T1i, T1u, T1t; Tb = T7 + Ta; Tq = FMA(KP1_688655851, Ti, KP535826794 * Tp); TF = FMA(KP1_541026485, Tx, KP637423989 * TE); TG = Tq - TF; T1E = FMA(KP851558583, TN, KP904827052 * TU); T1F = FMA(KP1_984229402, T12, KP125333233 * T19); T1G = T1E + T1F; T1B = FNMS(KP844327925, Tp, KP1_071653589 * Ti); T1C = FNMS(KP1_274847979, Tx, KP770513242 * TE); T1D = T1B + T1C; TV = FNMS(KP425779291, TU, KP1_809654104 * TN); T1a = FNMS(KP992114701, T19, KP250666467 * T12); T1b = TV + T1a; { E T1m, T1n, T1p, T1q; T1m = FMA(KP1_937166322, Ti, KP248689887 * Tp); T1n = FMA(KP1_071653589, Tx, KP844327925 * TE); T1o = T1m + T1n; T1p = FMA(KP1_752613360, TN, KP481753674 * TU); T1q = FMA(KP1_457937254, T12, KP684547105 * T19); T1r = T1p + T1q; T1s = T1o + T1r; T1z = T1q - T1p; T1x = T1n - T1m; } { E T1c, T1d, T1f, T1g; T1c = FNMS(KP497379774, Ti, KP968583161 * Tp); T1d = FNMS(KP1_688655851, Tx, KP535826794 * TE); T1e = T1c + T1d; T1f = FNMS(KP963507348, TN, KP876306680 * TU); T1g = FNMS(KP1_369094211, T12, KP728968627 * T19); T1h = T1f + T1g; T1i = T1e + T1h; T1u = T1f - T1g; T1t = T1d - T1c; } Cr[WS(csr, 1)] = Tb + T1i; Ci[WS(csi, 1)] = -(T1l + T1s); Cr[WS(csr, 4)] = Tb + TG + T1b; Ci[WS(csi, 4)] = T1l + T1D - T1G; Ci[WS(csi, 9)] = FMA(KP309016994, T1D, T1l) + FMA(KP587785252, T1a - TV, KP809016994 * T1G) - (KP951056516 * (Tq + TF)); Cr[WS(csr, 9)] = FMA(KP309016994, TG, Tb) + FMA(KP951056516, T1B - T1C, KP587785252 * (T1F - T1E)) - (KP809016994 * T1b); { E T1v, T1w, T1y, T1A; T1v = FMS(KP250000000, T1s, T1l); T1w = KP559016994 * (T1r - T1o); Ci[WS(csi, 11)] = FMA(KP587785252, T1t, KP951056516 * T1u) + T1v - T1w; Ci[WS(csi, 6)] = FMA(KP951056516, T1t, T1v) + FNMS(KP587785252, T1u, T1w); T1y = FNMS(KP250000000, T1i, Tb); T1A = KP559016994 * (T1e - T1h); Cr[WS(csr, 11)] = FMA(KP587785252, T1x, T1y) + FNMA(KP951056516, T1z, T1A); Cr[WS(csr, 6)] = FMA(KP951056516, T1x, T1A) + FMA(KP587785252, T1z, T1y); } } { E T1W, T1X, T1J, T1M, T1N, T21, T22, T23, T1Q, T1T, T1U, T1Y, T1Z, T20, T26; E T29, T2a, T2k, T2j, T2l, T2m, T2d, T2o, T2i; T1W = FNMS(KP587785252, T1j, T1V); T1X = Ta - T7; T1J = FNMS(KP125333233, T1I, KP1_984229402 * T1H); T1M = FMA(KP1_457937254, T1K, KP684547105 * T1L); T1N = T1J - T1M; T21 = FNMS(KP1_996053456, T1R, KP062790519 * T1S); T22 = FMA(KP1_541026485, T1O, KP637423989 * T1P); T23 = T21 - T22; T1Q = FNMS(KP770513242, T1P, KP1_274847979 * T1O); T1T = FMA(KP125581039, T1R, KP998026728 * T1S); T1U = T1Q - T1T; T1Y = FNMS(KP1_369094211, T1K, KP728968627 * T1L); T1Z = FMA(KP250666467, T1H, KP992114701 * T1I); T20 = T1Y - T1Z; { E T24, T25, T27, T28; T24 = FNMS(KP481753674, T1L, KP1_752613360 * T1K); T25 = FMA(KP851558583, T1H, KP904827052 * T1I); T26 = T24 - T25; T27 = FNMS(KP844327925, T1S, KP1_071653589 * T1R); T28 = FNMS(KP998026728, T1P, KP125581039 * T1O); T29 = T27 + T28; T2a = T26 + T29; T2k = T27 - T28; T2j = T24 + T25; } { E T2b, T2c, T2g, T2h; T2b = FNMS(KP425779291, T1I, KP1_809654104 * T1H); T2c = FMA(KP963507348, T1K, KP876306680 * T1L); T2l = T2c + T2b; T2g = FMA(KP1_688655851, T1R, KP535826794 * T1S); T2h = FMA(KP1_996053456, T1O, KP062790519 * T1P); T2m = T2g + T2h; T2d = T2b - T2c; T2o = T2l + T2m; T2i = T2g - T2h; } Ci[WS(csi, 2)] = T1W + T2a; Cr[WS(csr, 2)] = T1X + T2o; Ci[WS(csi, 3)] = T1N + T1U - T1W; Cr[WS(csr, 3)] = T1X + T20 + T23; Cr[WS(csr, 8)] = FMA(KP309016994, T20, T1X) + FNMA(KP809016994, T23, KP587785252 * (T1T + T1Q)) - (KP951056516 * (T1M + T1J)); Ci[WS(csi, 8)] = FNMS(KP587785252, T21 + T22, KP309016994 * T1N) + FNMA(KP809016994, T1U, KP951056516 * (T1Y + T1Z)) - T1W; { E T2e, T2f, T2n, T2p; T2e = KP559016994 * (T26 - T29); T2f = FNMS(KP250000000, T2a, T1W); Ci[WS(csi, 7)] = FMA(KP951056516, T2d, T2e) + FNMS(KP587785252, T2i, T2f); Ci[WS(csi, 12)] = FMA(KP587785252, T2d, T2f) + FMS(KP951056516, T2i, T2e); T2n = KP559016994 * (T2l - T2m); T2p = FNMS(KP250000000, T2o, T1X); Cr[WS(csr, 7)] = FMA(KP951056516, T2j, KP587785252 * T2k) + T2n + T2p; Cr[WS(csr, 12)] = FMA(KP587785252, T2j, T2p) + FNMA(KP951056516, T2k, T2n); } } }}static const kr2c_desc desc = { 25, "r2cf_25", {117, 57, 83, 0}, &GENUS };void X(codelet_r2cf_25) (planner *p) { X(kr2c_register) (p, r2cf_25, &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 + -