📄 hf_25.c
字号:
ci[WS(rs, 7)] = FMA(KP949179823, T6Q, T6P); cr[WS(rs, 12)] = FNMS(KP949179823, T6Q, T6P); cr[WS(rs, 7)] = FMA(KP860541664, T6O, T6H); ci[WS(rs, 2)] = FNMS(KP860541664, T6O, T6H); T7o = FMA(KP621716863, T7k, T7l); T7m = FNMS(KP614372930, T7l, T7k); T7h = FMA(KP249506682, T7g, T7d); T6n = FMA(KP557913902, T6e, T6d); T6f = FNMS(KP557913902, T6e, T6d); } } } } } } ci[WS(rs, 6)] = FNMS(KP949179823, T6o, T6n); ci[WS(rs, 11)] = FMA(KP949179823, T6o, T6n); cr[WS(rs, 8)] = FMA(KP943557151, T6m, T6f); ci[WS(rs, 1)] = FNMS(KP943557151, T6m, T6f); T7j = FNMS(KP557913902, T7i, T7h); T7n = FMA(KP557913902, T7i, T7h); cr[WS(rs, 23)] = -(FMA(KP943557151, T7o, T7n)); ci[WS(rs, 16)] = FNMS(KP943557151, T7o, T7n); cr[WS(rs, 18)] = FMS(KP949179823, T7m, T7j); cr[WS(rs, 13)] = -(FMA(KP949179823, T7m, T7j)); }}static const tw_instr twinstr[] = { {TW_FULL, 1, 25}, {TW_NEXT, 1, 0}};static const hc2hc_desc desc = { 25, "hf_25", twinstr, &GENUS, {84, 48, 316, 0} };void X(codelet_hf_25) (planner *p) { X(khc2hc_register) (p, hf_25, &desc);}#else /* HAVE_FMA *//* Generated by: ../../../genfft/gen_hc2hc -compact -variables 4 -pipeline-latency 4 -n 25 -dit -name hf_25 -include hf.h *//* * This function contains 400 FP additions, 280 FP multiplications, * (or, 260 additions, 140 multiplications, 140 fused multiply/add), * 101 stack variables, 20 constants, and 100 memory accesses */#include "hf.h"static void hf_25(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms){ DK(KP998026728, +0.998026728428271561952336806863450553336905220); DK(KP062790519, +0.062790519529313376076178224565631133122484832); DK(KP684547105, +0.684547105928688673732283357621209269889519233); DK(KP728968627, +0.728968627421411523146730319055259111372571664); DK(KP481753674, +0.481753674101715274987191502872129653528542010); DK(KP876306680, +0.876306680043863587308115903922062583399064238); DK(KP248689887, +0.248689887164854788242283746006447968417567406); DK(KP968583161, +0.968583161128631119490168375464735813836012403); DK(KP992114701, +0.992114701314477831049793042785778521453036709); DK(KP125333233, +0.125333233564304245373118759816508793942918247); DK(KP425779291, +0.425779291565072648862502445744251703979973042); DK(KP904827052, +0.904827052466019527713668647932697593970413911); DK(KP637423989, +0.637423989748689710176712811676016195434917298); DK(KP770513242, +0.770513242775789230803009636396177847271667672); DK(KP844327925, +0.844327925502015078548558063966681505381659241); DK(KP535826794, +0.535826794978996618271308767867639978063575346); DK(KP587785252, +0.587785252292473129168705954639072768597652438); DK(KP951056516, +0.951056516295153572116439333379382143405698634); DK(KP250000000, +0.250000000000000000000000000000000000000000000); DK(KP559016994, +0.559016994374947424102293417182819058860154590); INT m; for (m = mb, W = W + ((mb - 1) * 48); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 48, MAKE_VOLATILE_STRIDE(rs)) { E T1, T6b, T2l, T6g, To, T2m, T6e, T6f, T6a, T6H, T2u, T4I, T2i, T60, T3S; E T5D, T4r, T58, T3Z, T5C, T4q, T5b, TS, T5W, T2G, T5s, T4g, T4M, T2R, T5t; E T4h, T4P, T1l, T5X, T37, T5v, T4k, T4T, T3e, T5w, T4j, T4W, T1P, T5Z, T3v; E T5A, T4o, T54, T3C, T5z, T4n, T51; { E T6, T2o, Tb, T2p, Tc, T6c, Th, T2r, Tm, T2s, Tn, T6d; T1 = cr[0]; T6b = ci[0]; { E T3, T5, T2, T4; T3 = cr[WS(rs, 5)]; T5 = ci[WS(rs, 5)]; T2 = W[8]; T4 = W[9]; T6 = FMA(T2, T3, T4 * T5); T2o = FNMS(T4, T3, T2 * T5); } { E T8, Ta, T7, T9; T8 = cr[WS(rs, 20)]; Ta = ci[WS(rs, 20)]; T7 = W[38]; T9 = W[39]; Tb = FMA(T7, T8, T9 * Ta); T2p = FNMS(T9, T8, T7 * Ta); } Tc = T6 + Tb; T6c = T2o + T2p; { E Te, Tg, Td, Tf; Te = cr[WS(rs, 10)]; Tg = ci[WS(rs, 10)]; Td = W[18]; Tf = W[19]; Th = FMA(Td, Te, Tf * Tg); T2r = FNMS(Tf, Te, Td * Tg); } { E Tj, Tl, Ti, Tk; Tj = cr[WS(rs, 15)]; Tl = ci[WS(rs, 15)]; Ti = W[28]; Tk = W[29]; Tm = FMA(Ti, Tj, Tk * Tl); T2s = FNMS(Tk, Tj, Ti * Tl); } Tn = Th + Tm; T6d = T2r + T2s; T2l = KP559016994 * (Tc - Tn); T6g = KP559016994 * (T6c - T6d); To = Tc + Tn; T2m = FNMS(KP250000000, To, T1); T6e = T6c + T6d; T6f = FNMS(KP250000000, T6e, T6b); { E T68, T69, T2q, T2t; T68 = Th - Tm; T69 = T6 - Tb; T6a = FNMS(KP587785252, T69, KP951056516 * T68); T6H = FMA(KP951056516, T69, KP587785252 * T68); T2q = T2o - T2p; T2t = T2r - T2s; T2u = FMA(KP951056516, T2q, KP587785252 * T2t); T4I = FNMS(KP587785252, T2q, KP951056516 * T2t); } } { E T1U, T3O, T3E, T3F, T3X, T3W, T3J, T3M, T3P, T25, T2g, T2h; { E T1R, T1T, T1Q, T1S; T1R = cr[WS(rs, 3)]; T1T = ci[WS(rs, 3)]; T1Q = W[4]; T1S = W[5]; T1U = FMA(T1Q, T1R, T1S * T1T); T3O = FNMS(T1S, T1R, T1Q * T1T); } { E T1Z, T3H, T2f, T3L, T24, T3I, T2a, T3K; { E T1W, T1Y, T1V, T1X; T1W = cr[WS(rs, 8)]; T1Y = ci[WS(rs, 8)]; T1V = W[14]; T1X = W[15]; T1Z = FMA(T1V, T1W, T1X * T1Y); T3H = FNMS(T1X, T1W, T1V * T1Y); } { E T2c, T2e, T2b, T2d; T2c = cr[WS(rs, 18)]; T2e = ci[WS(rs, 18)]; T2b = W[34]; T2d = W[35]; T2f = FMA(T2b, T2c, T2d * T2e); T3L = FNMS(T2d, T2c, T2b * T2e); } { E T21, T23, T20, T22; T21 = cr[WS(rs, 23)]; T23 = ci[WS(rs, 23)]; T20 = W[44]; T22 = W[45]; T24 = FMA(T20, T21, T22 * T23); T3I = FNMS(T22, T21, T20 * T23); } { E T27, T29, T26, T28; T27 = cr[WS(rs, 13)]; T29 = ci[WS(rs, 13)]; T26 = W[24]; T28 = W[25]; T2a = FMA(T26, T27, T28 * T29); T3K = FNMS(T28, T27, T26 * T29); } T3E = T1Z - T24; T3F = T2a - T2f; T3X = T3K - T3L; T3W = T3H - T3I; T3J = T3H + T3I; T3M = T3K + T3L; T3P = T3J + T3M; T25 = T1Z + T24; T2g = T2a + T2f; T2h = T25 + T2g; } T2i = T1U + T2h; T60 = T3O + T3P; { E T3G, T57, T3R, T56, T3N, T3Q; T3G = FMA(KP951056516, T3E, KP587785252 * T3F); T57 = FNMS(KP587785252, T3E, KP951056516 * T3F); T3N = KP559016994 * (T3J - T3M); T3Q = FNMS(KP250000000, T3P, T3O); T3R = T3N + T3Q; T56 = T3Q - T3N; T3S = T3G + T3R; T5D = T57 + T56; T4r = T3R - T3G; T58 = T56 - T57; } { E T3Y, T5a, T3V, T59, T3T, T3U; T3Y = FMA(KP951056516, T3W, KP587785252 * T3X); T5a = FNMS(KP587785252, T3W, KP951056516 * T3X); T3T = KP559016994 * (T25 - T2g); T3U = FNMS(KP250000000, T2h, T1U); T3V = T3T + T3U; T59 = T3U - T3T; T3Z = T3V - T3Y; T5C = T59 - T5a; T4q = T3V + T3Y; T5b = T59 + T5a; } } { E Tu, T2N, T2B, T2E, T2I, T2H, T2K, T2L, T2O, TF, TQ, TR; { E Tr, Tt, Tq, Ts; Tr = cr[WS(rs, 1)]; Tt = ci[WS(rs, 1)]; Tq = W[0]; Ts = W[1]; Tu = FMA(Tq, Tr, Ts * Tt); T2N = FNMS(Ts, Tr, Tq * Tt); } { E Tz, T2z, TP, T2D, TE, T2A, TK, T2C; { E Tw, Ty, Tv, Tx; Tw = cr[WS(rs, 6)]; Ty = ci[WS(rs, 6)]; Tv = W[10]; Tx = W[11]; Tz = FMA(Tv, Tw, Tx * Ty); T2z = FNMS(Tx, Tw, Tv * Ty); } { E TM, TO, TL, TN; TM = cr[WS(rs, 16)]; TO = ci[WS(rs, 16)]; TL = W[30]; TN = W[31]; TP = FMA(TL, TM, TN * TO); T2D = FNMS(TN, TM, TL * TO); } { E TB, TD, TA, TC; TB = cr[WS(rs, 21)]; TD = ci[WS(rs, 21)]; TA = W[40]; TC = W[41]; TE = FMA(TA, TB, TC * TD); T2A = FNMS(TC, TB, TA * TD); } { E TH, TJ, TG, TI; TH = cr[WS(rs, 11)]; TJ = ci[WS(rs, 11)]; TG = W[20]; TI = W[21]; TK = FMA(TG, TH, TI * TJ); T2C = FNMS(TI, TH, TG * TJ); } T2B = T2z - T2A; T2E = T2C - T2D; T2I = TK - TP; T2H = Tz - TE; T2K = T2z + T2A; T2L = T2C + T2D; T2O = T2K + T2L; TF = Tz + TE; TQ = TK + TP; TR = TF + TQ; } TS = Tu + TR; T5W = T2N + T2O; { E T2F, T4L, T2y, T4K, T2w, T2x; T2F = FMA(KP951056516, T2B, KP587785252 * T2E); T4L = FNMS(KP587785252, T2B, KP951056516 * T2E); T2w = KP559016994 * (TF - TQ); T2x = FNMS(KP250000000, TR, Tu); T2y = T2w + T2x; T4K = T2x - T2w; T2G = T2y - T2F; T5s = T4K - T4L; T4g = T2y + T2F; T4M = T4K + T4L; } { E T2J, T4O, T2Q, T4N, T2M, T2P; T2J = FMA(KP951056516, T2H, KP587785252 * T2I); T4O = FNMS(KP587785252, T2H, KP951056516 * T2I); T2M = KP559016994 * (T2K - T2L); T2P = FNMS(KP250000000, T2O, T2N); T2Q = T2M + T2P; T4N = T2P - T2M; T2R = T2J + T2Q; T5t = T4O + T4N; T4h = T2Q - T2J; T4P = T4N - T4O; } } { E TX, T33, T2T, T2U, T3c, T3b, T2Y, T31, T34, T18, T1j, T1k; { E TU, TW, TT, TV; TU = cr[WS(rs, 4)]; TW = ci[WS(rs, 4)]; TT = W[6]; TV = W[7]; TX = FMA(TT, TU, TV * TW); T33 = FNMS(TV, TU, TT * TW); } { E T12, T2W, T1i, T30, T17, T2X, T1d, T2Z; { E TZ, T11, TY, T10; TZ = cr[WS(rs, 9)]; T11 = ci[WS(rs, 9)]; TY = W[16]; T10 = W[17]; T12 = FMA(TY, TZ, T10 * T11); T2W = FNMS(T10, TZ, TY * T11); } { E T1f, T1h, T1e, T1g; T1f = cr[WS(rs, 19)]; T1h = ci[WS(rs, 19)]; T1e = W[36]; T1g = W[37]; T1i = FMA(T1e, T1f, T1g * T1h); T30 = FNMS(T1g, T1f, T1e * T1h); } { E T14, T16, T13, T15; T14 = cr[WS(rs, 24)]; T16 = ci[WS(rs, 24)]; T13 = W[46]; T15 = W[47]; T17 = FMA(T13, T14, T15 * T16); T2X = FNMS(T15, T14, T13 * T16); } { E T1a, T1c, T19, T1b; T1a = cr[WS(rs, 14)]; T1c = ci[WS(rs, 14)]; T19 = W[26]; T1b = W[27]; T1d = FMA(T19, T1a, T1b * T1c); T2Z = FNMS(T1b, T1a, T19 * T1c); } T2T = T17 - T12; T2U = T1d - T1i; T3c = T2Z - T30; T3b = T2W - T2X; T2Y = T2W + T2X; T31 = T2Z + T30; T34 = T2Y + T31; T18 = T12 + T17; T1j = T1d + T1i; T1k = T18 + T1j; } T1l = TX + T1k; T5X = T33 + T34; { E T2V, T4S, T36, T4R, T32, T35; T2V = FNMS(KP587785252, T2U, KP951056516 * T2T); T4S = FMA(KP587785252, T2T, KP951056516 * T2U); T32 = KP559016994 * (T2Y - T31); T35 = FNMS(KP250000000, T34, T33); T36 = T32 + T35; T4R = T35 - T32; T37 = T2V - T36; T5v = T4S + T4R; T4k = T2V + T36; T4T = T4R - T4S; } {
⌨️ 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 + -