📄 r2cfii_25.c
字号:
T3k = FNMS(KP560319534, T3j, T3f); T3i = FMA(KP681693190, T3h, T3e); } } Cr[WS(csr, 8)] = FMA(KP949179823, T3k, T3d); Cr[WS(csr, 3)] = FMA(KP860541664, T3i, T3d); T2P = FNMS(KP809385824, T2O, T2N); T2S = FMA(KP809385824, T2O, T2N); } } } { E T2F, T2K, T2M, T2Q; T2Q = FMA(KP248028675, T2P, T2B); { E T2U, T2W, T2E, T2I; T2U = FNMS(KP894834959, T2T, T2S); T2W = T2R + T2S; T2E = FMA(KP904730450, T2D, T2C); T2I = FNMS(KP904730450, T2D, T2C); { E T2V, T2X, T2J, T2L; T2V = FNMS(KP618033988, T2U, T2R); T2X = FNMS(KP690668130, T2W, T2T); T2F = FNMS(KP242145790, T2E, T2B); Cr[0] = FMA(KP968583161, T2E, T2B); T2J = T2H + T2I; T2L = FMA(KP904730450, T2G, T2I); Cr[WS(csr, 9)] = FMA(KP897376177, T2V, T2Q); Cr[WS(csr, 4)] = FNMS(KP803003575, T2X, T2Q); T2K = FNMS(KP683113946, T2J, T2G); T2M = FMA(KP618033988, T2L, T2H); } } Cr[WS(csr, 5)] = FMA(KP792626838, T2K, T2F); Cr[WS(csr, 10)] = FMA(KP876091699, T2M, T2F); } }}static const kr2c_desc desc = { 25, "r2cfII_25", {47, 12, 165, 0}, &GENUS };void X(codelet_r2cfII_25) (planner *p) { X(kr2c_register) (p, r2cfII_25, &desc);}#else /* HAVE_FMA *//* Generated by: ../../../genfft/gen_r2cf -compact -variables 4 -pipeline-latency 4 -n 25 -name r2cfII_25 -dft-II -include r2cfII.h *//* * This function contains 213 FP additions, 148 FP multiplications, * (or, 126 additions, 61 multiplications, 87 fused multiply/add), * 94 stack variables, 38 constants, and 50 memory accesses */#include "r2cfII.h"static void r2cfII_25(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs){ DK(KP1_996053456, +1.996053456856543123904673613726901106673810439); DK(KP062790519, +0.062790519529313376076178224565631133122484832); DK(KP125581039, +0.125581039058626752152356449131262266244969664); DK(KP998026728, +0.998026728428271561952336806863450553336905220); 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(KP1_457937254, +1.457937254842823046293460638110518222745143328); DK(KP684547105, +0.684547105928688673732283357621209269889519233); DK(KP1_752613360, +1.752613360087727174616231807844125166798128477); DK(KP481753674, +0.481753674101715274987191502872129653528542010); DK(KP1_937166322, +1.937166322257262238980336750929471627672024806); DK(KP248689887, +0.248689887164854788242283746006447968417567406); DK(KP992114701, +0.992114701314477831049793042785778521453036709); DK(KP250666467, +0.250666467128608490746237519633017587885836494); DK(KP1_809654104, +1.809654104932039055427337295865395187940827822); DK(KP425779291, +0.425779291565072648862502445744251703979973042); DK(KP1_541026485, +1.541026485551578461606019272792355694543335344); DK(KP637423989, +0.637423989748689710176712811676016195434917298); DK(KP1_688655851, +1.688655851004030157097116127933363010763318483); DK(KP535826794, +0.535826794978996618271308767867639978063575346); DK(KP851558583, +0.851558583130145297725004891488503407959946084); DK(KP904827052, +0.904827052466019527713668647932697593970413911); DK(KP1_984229402, +1.984229402628955662099586085571557042906073418); DK(KP125333233, +0.125333233564304245373118759816508793942918247); DK(KP1_274847979, +1.274847979497379420353425623352032390869834596); DK(KP770513242, +0.770513242775789230803009636396177847271667672); DK(KP844327925, +0.844327925502015078548558063966681505381659241); DK(KP1_071653589, +1.071653589957993236542617535735279956127150691); DK(KP293892626, +0.293892626146236564584352977319536384298826219); DK(KP475528258, +0.475528258147576786058219666689691071702849317); DK(KP250000000, +0.250000000000000000000000000000000000000000000); DK(KP587785252, +0.587785252292473129168705954639072768597652438); DK(KP951056516, +0.951056516295153572116439333379382143405698634); DK(KP559016994, +0.559016994374947424102293417182819058860154590); 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 TE, TR, T2i, T1z, TL, TS, TB, T2d, T1l, T1i, T2c, T9, T23, TZ, TW; E T22, Ti, T26, T16, T13, T25, Ts, T2a, T1e, T1b, T29, TP, TQ; { E TK, T1y, TH, T1x; TE = R0[0]; { E TI, TJ, TF, TG; TI = R0[WS(rs, 10)]; TJ = R1[WS(rs, 2)]; TK = TI - TJ; T1y = TI + TJ; TF = R0[WS(rs, 5)]; TG = R1[WS(rs, 7)]; TH = TF - TG; T1x = TF + TG; } TR = KP559016994 * (TH - TK); T2i = FNMS(KP587785252, T1x, KP951056516 * T1y); T1z = FMA(KP951056516, T1x, KP587785252 * T1y); TL = TH + TK; TS = FNMS(KP250000000, TL, TE); } { E Tt, Tw, Tz, TA, T1k, T1j, T1g, T1h; Tt = R0[WS(rs, 3)]; { E Tu, Tv, Tx, Ty; Tu = R0[WS(rs, 8)]; Tv = R1[WS(rs, 10)]; Tw = Tu - Tv; Tx = R1[0]; Ty = R1[WS(rs, 5)]; Tz = Tx + Ty; TA = Tw - Tz; T1k = Ty - Tx; T1j = Tu + Tv; } TB = Tt + TA; T2d = FNMS(KP293892626, T1j, KP475528258 * T1k); T1l = FMA(KP475528258, T1j, KP293892626 * T1k); T1g = FNMS(KP250000000, TA, Tt); T1h = KP559016994 * (Tw + Tz); T1i = T1g + T1h; T2c = T1g - T1h; } { E T1, T4, T7, T8, TY, TX, TU, TV; T1 = R0[WS(rs, 1)]; { E T2, T3, T5, T6; T2 = R0[WS(rs, 6)]; T3 = R1[WS(rs, 8)]; T4 = T2 - T3; T5 = R0[WS(rs, 11)]; T6 = R1[WS(rs, 3)]; T7 = T5 - T6; T8 = T4 + T7; TY = T5 + T6; TX = T2 + T3; } T9 = T1 + T8; T23 = FNMS(KP293892626, TX, KP475528258 * TY); TZ = FMA(KP475528258, TX, KP293892626 * TY); TU = KP559016994 * (T4 - T7); TV = FNMS(KP250000000, T8, T1); TW = TU + TV; T22 = TV - TU; } { E Ta, Td, Tg, Th, T15, T14, T11, T12; Ta = R0[WS(rs, 4)]; { E Tb, Tc, Te, Tf; Tb = R0[WS(rs, 9)]; Tc = R1[WS(rs, 11)]; Td = Tb - Tc; Te = R1[WS(rs, 1)]; Tf = R1[WS(rs, 6)]; Tg = Te + Tf; Th = Td - Tg; T15 = Tf - Te; T14 = Tb + Tc; } Ti = Ta + Th; T26 = FNMS(KP293892626, T14, KP475528258 * T15); T16 = FMA(KP475528258, T14, KP293892626 * T15); T11 = FNMS(KP250000000, Th, Ta); T12 = KP559016994 * (Td + Tg); T13 = T11 + T12; T25 = T11 - T12; } { E Tk, Tn, Tq, Tr, T1d, T1c, T19, T1a; Tk = R0[WS(rs, 2)]; { E Tl, Tm, To, Tp; Tl = R0[WS(rs, 7)]; Tm = R1[WS(rs, 9)]; Tn = Tl - Tm; To = R0[WS(rs, 12)]; Tp = R1[WS(rs, 4)]; Tq = To - Tp; Tr = Tn + Tq; T1d = To + Tp; T1c = Tl + Tm; } Ts = Tk + Tr; T2a = FNMS(KP293892626, T1c, KP475528258 * T1d); T1e = FMA(KP475528258, T1c, KP293892626 * T1d); T19 = KP559016994 * (Tn - Tq); T1a = FNMS(KP250000000, Tr, Tk); T1b = T19 + T1a; T29 = T1a - T19; } TP = TB - Ts; TQ = T9 - Ti; Ci[WS(csi, 2)] = FNMS(KP951056516, TQ, KP587785252 * TP); Ci[WS(csi, 7)] = FMA(KP587785252, TQ, KP951056516 * TP); { E TM, TD, TN, Tj, TC, TO; TM = TE + TL; Tj = T9 + Ti; TC = Ts + TB; TD = KP559016994 * (Tj - TC); TN = Tj + TC; Cr[WS(csr, 12)] = TM + TN; TO = FNMS(KP250000000, TN, TM); Cr[WS(csr, 2)] = TD + TO; Cr[WS(csr, 7)] = TO - TD; } { E TT, T1J, T1Y, T1U, T1X, T1P, T1V, T1M, T1W, T1A, T1B, T1r, T1C, T1v, T18; E T1n, T1o, T1G, T1D; TT = TR + TS; { E T1H, T1I, T1S, T1T; T1H = FNMS(KP844327925, TW, KP1_071653589 * TZ); T1I = FNMS(KP1_274847979, T16, KP770513242 * T13); T1J = T1H - T1I; T1Y = T1H + T1I; T1S = FMA(KP125333233, T1i, KP1_984229402 * T1l); T1T = FMA(KP904827052, T1b, KP851558583 * T1e); T1U = T1S - T1T; T1X = T1T + T1S; } { E T1N, T1O, T1K, T1L; T1N = FMA(KP535826794, TW, KP1_688655851 * TZ); T1O = FMA(KP637423989, T13, KP1_541026485 * T16); T1P = T1N - T1O; T1V = T1N + T1O; T1K = FNMS(KP1_809654104, T1e, KP425779291 * T1b); T1L = FNMS(KP992114701, T1i, KP250666467 * T1l); T1M = T1K - T1L; T1W = T1K + T1L; } { E T1p, T1q, T1t, T1u; T1p = FMA(KP844327925, T13, KP1_071653589 * T16); T1q = FMA(KP248689887, TW, KP1_937166322 * TZ); T1A = T1q + T1p; T1t = FMA(KP481753674, T1b, KP1_752613360 * T1e); T1u = FMA(KP684547105, T1i, KP1_457937254 * T1l); T1B = T1t + T1u; T1r = T1p - T1q; T1C = T1A + T1B; T1v = T1t - T1u; } { E T10, T17, T1f, T1m; T10 = FNMS(KP497379774, TZ, KP968583161 * TW); T17 = FNMS(KP1_688655851, T16, KP535826794 * T13); T18 = T10 + T17; T1f = FNMS(KP963507348, T1e, KP876306680 * T1b); T1m = FNMS(KP1_369094211, T1l, KP728968627 * T1i); T1n = T1f + T1m; T1o = T18 + T1n; T1G = T10 - T17; T1D = T1f - T1m; } { E T1R, T1Q, T20, T1Z; Cr[0] = TT + T1o; Ci[0] = -(T1z + T1C); T1R = KP559016994 * (T1P + T1M); T1Q = FMA(KP250000000, T1M - T1P, TT); Cr[WS(csr, 4)] = FMA(KP951056516, T1J, T1Q) + FMA(KP587785252, T1U, T1R); Cr[WS(csr, 9)] = FMA(KP951056516, T1U, T1Q) + FNMA(KP587785252, T1J, T1R); T20 = KP559016994 * (T1Y + T1X); T1Z = FMA(KP250000000, T1X - T1Y, T1z); Ci[WS(csi, 9)] = FMA(KP587785252, T1V, KP951056516 * T1W) + T1Z - T20; Ci[WS(csi, 4)] = FMA(KP587785252, T1W, T1Z) + FNMS(KP951056516, T1V, T20); { E T1E, T1F, T1s, T1w; T1E = FMS(KP250000000, T1C, T1z); T1F = KP559016994 * (T1B - T1A); Ci[WS(csi, 5)] = FMA(KP951056516, T1D, T1E) + FNMA(KP587785252, T1G, T1F); Ci[WS(csi, 10)] = FMA(KP951056516, T1G, KP587785252 * T1D) + T1E + T1F; T1s = FNMS(KP250000000, T1o, TT); T1w = KP559016994 * (T18 - T1n); Cr[WS(csr, 5)] = FMA(KP587785252, T1r, T1s) + FMS(KP951056516, T1v, T1w); Cr[WS(csr, 10)] = T1w + FMA(KP587785252, T1v, T1s) - (KP951056516 * T1r); } } } { E T21, T2z, T2L, T2K, T2M, T2F, T2P, T2C, T2Q, T2l, T2o, T2p, T2w, T2u, T28; E T2f, T2g, T2s, T2h; T21 = TS - TR; { E T2x, T2y, T2I, T2J; T2x = FNMS(KP844327925, T29, KP1_071653589 * T2a); T2y = FNMS(KP125581039, T2d, KP998026728 * T2c); T2z = T2x + T2y; T2L = T2y - T2x; T2I = FNMS(KP481753674, T22, KP1_752613360 * T23); T2J = FMA(KP904827052, T25, KP851558583 * T26); T2K = T2I + T2J; T2M = T2I - T2J; } { E T2D, T2E, T2A, T2B; T2D = FMA(KP535826794, T29, KP1_688655851 * T2a); T2E = FMA(KP062790519, T2c, KP1_996053456 * T2d); T2F = T2D + T2E; T2P = T2E - T2D; T2A = FMA(KP876306680, T22, KP963507348 * T23); T2B = FNMS(KP425779291, T25, KP1_809654104 * T26); T2C = T2A + T2B; T2Q = T2A - T2B; } { E T2j, T2k, T2m, T2n; T2j = FNMS(KP125333233, T25, KP1_984229402 * T26); T2k = FMA(KP684547105, T22, KP1_457937254 * T23); T2l = T2j - T2k; T2m = FNMS(KP770513242, T2c, KP1_274847979 * T2d); T2n = FMA(KP998026728, T29, KP125581039 * T2a); T2o = T2m - T2n; T2p = T2l + T2o; T2w = T2k + T2j; T2u = T2n + T2m; } { E T24, T27, T2b, T2e; T24 = FNMS(KP1_369094211, T23, KP728968627 * T22); T27 = FMA(KP992114701, T25, KP250666467 * T26); T28 = T24 - T27; T2b = FNMS(KP1_996053456, T2a, KP062790519 * T29); T2e = FMA(KP637423989, T2c, KP1_541026485 * T2d); T2f = T2b - T2e; T2g = T28 + T2f; T2s = T24 + T27; T2h = T2b + T2e; } { E T2H, T2G, T2O, T2N; Cr[WS(csr, 1)] = T21 + T2g; Ci[WS(csi, 1)] = T2p - T2i; T2H = KP559016994 * (T2C - T2F); T2G = FNMS(KP250000000, T2C + T2F, T21); Cr[WS(csr, 8)] = FMA(KP951056516, T2z, T2G) + FNMA(KP587785252, T2K, T2H); Cr[WS(csr, 3)] = FMA(KP951056516, T2K, KP587785252 * T2z) + T2G + T2H; T2O = KP559016994 * (T2M + T2L); T2N = FMA(KP250000000, T2L - T2M, T2i); Ci[WS(csi, 3)] = T2N + FMA(KP587785252, T2P, T2O) - (KP951056516 * T2Q); Ci[WS(csi, 8)] = FMA(KP587785252, T2Q, T2N) + FMS(KP951056516, T2P, T2O); { E T2t, T2v, T2q, T2r; T2t = FNMS(KP250000000, T2g, T21); T2v = KP559016994 * (T28 - T2f); Cr[WS(csr, 6)] = FMA(KP951056516, T2u, T2t) + FNMA(KP587785252, T2w, T2v); Cr[WS(csr, 11)] = FMA(KP951056516, T2w, T2v) + FMA(KP587785252, T2u, T2t); T2q = KP250000000 * T2p; T2r = KP559016994 * (T2l - T2o); Ci[WS(csi, 6)] = FMS(KP951056516, T2h, T2i + T2q) + FNMA(KP587785252, T2s, T2r); Ci[WS(csi, 11)] = FMA(KP951056516, T2s, KP587785252 * T2h) + T2r - (T2i + T2q); } } } }}static const kr2c_desc desc = { 25, "r2cfII_25", {126, 61, 87, 0}, &GENUS };void X(codelet_r2cfII_25) (planner *p) { X(kr2c_register) (p, r2cfII_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 + -