📄 hf_64.c
字号:
E Tek, Tel, TdQ, Tdk; Tdo = FMA(KP707106781, Tdn, Tdm); Tek = FNMS(KP707106781, Tdn, Tdm); Tel = TdB - TdA; TdC = TdA + TdB; Tdz = FMA(KP707106781, Tdy, Tdx); Ten = FNMS(KP707106781, Tdy, Tdx); TdQ = FNMS(KP923879532, Tdj, Tdg); Tdk = FMA(KP923879532, Tdj, Tdg); TeE = FMA(KP923879532, Tel, Tek); Tem = FNMS(KP923879532, Tel, Tek); TdR = FNMS(KP820678790, TdQ, TdP); TdX = FMA(KP820678790, TdP, TdQ); TdI = FNMS(KP098491403, Tdd, Tdk); Tdl = FMA(KP098491403, Tdk, Tdd); Teo = Tdu - Tdr; Tdv = Tdr + Tdu; } } } { E TeB, Tef, TeF, Tep, TeC, Tei; TeF = FNMS(KP923879532, Teo, Ten); Tep = FMA(KP923879532, Teo, Ten); TeB = FMA(KP923879532, Tee, Ted); Tef = FNMS(KP923879532, Tee, Ted); TeG = FMA(KP303346683, TeF, TeE); TeK = FNMS(KP303346683, TeE, TeF); Tet = FMA(KP534511135, Tem, Tep); Teq = FNMS(KP534511135, Tep, Tem); TeC = FMA(KP923879532, Teh, Teg); Tei = FNMS(KP923879532, Teh, Teg); Tex = FNMS(KP668178637, Te5, Te6); Te7 = FMA(KP668178637, Te6, Te5); TeD = FNMS(KP303346683, TeC, TeB); TeJ = FMA(KP303346683, TeB, TeC); Teu = FNMS(KP534511135, Tef, Tei); Tej = FMA(KP534511135, Tei, Tef); } { E TdS, Tdw, Tey, Tea, TdT, TdD; Tey = FMA(KP668178637, Te8, Te9); Tea = FNMS(KP668178637, Te9, Te8); TdS = FNMS(KP923879532, Tdv, Tdo); Tdw = FMA(KP923879532, Tdv, Tdo); Tez = Tex + Tey; TkK = Tey - Tex; TkE = Te7 + Tea; Teb = Te7 - Tea; Te2 = FNMS(KP414213562, TcF, TcG); TcH = FMA(KP414213562, TcG, TcF); TdT = FNMS(KP923879532, TdC, Tdz); TdD = FMA(KP923879532, TdC, Tdz); Te0 = FNMS(KP707106781, TcD, TcA); TcE = FMA(KP707106781, TcD, TcA); TkB = FNMS(KP707106781, Tkm, Tkl); Tkn = FMA(KP707106781, Tkm, Tkl); TdU = FMA(KP820678790, TdT, TdS); TdY = FNMS(KP820678790, TdS, TdT); TdH = FMA(KP098491403, Tdw, TdD); TdE = FNMS(KP098491403, TdD, Tdw); TcK = FNMS(KP414213562, TcJ, TcI); Te1 = FMA(KP414213562, TcI, TcJ); } } { E Tkv, Tkp, Tew, Te4, TkH, TeL, TkL, Ter, Tks, TkA, TkI, TeH, TkM, Tev; { E TdG, TdO, TkF, TkG, TkN, TkO, TdJ, TdV; { E TdF, TkJ, TkD, Td2, TdW, TdZ, TdK; Tku = TdE - Tdl; TdF = Tdl + TdE; { E TkC, TcL, Tko, Te3, TcM; TkC = TcH - TcK; TcL = TcH + TcK; Tko = Te2 + Te1; Te3 = Te1 - Te2; TkJ = FNMS(KP923879532, TkC, TkB); TkD = FMA(KP923879532, TkC, TkB); TdK = FNMS(KP923879532, TcL, TcE); TcM = FMA(KP923879532, TcL, TcE); Tkv = FNMS(KP923879532, Tko, Tkn); Tkp = FMA(KP923879532, Tko, Tkn); Tew = FMA(KP923879532, Te3, Te0); Te4 = FNMS(KP923879532, Te3, Te0); Td2 = FMA(KP980785280, Td1, TcM); TdG = FNMS(KP980785280, Td1, TcM); } TdO = FMA(KP980785280, TdN, TdK); TdW = FNMS(KP980785280, TdN, TdK); TdZ = TdX - TdY; Tky = TdX + TdY; TkH = FNMS(KP831469612, TkE, TkD); TkF = FMA(KP831469612, TkE, TkD); cr[WS(rs, 1)] = FMA(KP995184726, TdF, Td2); ci[WS(rs, 30)] = FNMS(KP995184726, TdF, Td2); cr[WS(rs, 9)] = FMA(KP773010453, TdZ, TdW); ci[WS(rs, 22)] = FNMS(KP773010453, TdZ, TdW); TkG = TeJ + TeK; TeL = TeJ - TeK; TkL = FMA(KP831469612, TkK, TkJ); TkN = FNMS(KP831469612, TkK, TkJ); TkO = Teq - Tej; Ter = Tej + Teq; } ci[WS(rs, 34)] = FNMS(KP956940335, TkG, TkF); cr[WS(rs, 61)] = -(FMA(KP956940335, TkG, TkF)); ci[WS(rs, 42)] = FMA(KP881921264, TkO, TkN); cr[WS(rs, 53)] = FMS(KP881921264, TkO, TkN); TdJ = TdH - TdI; Tks = TdI + TdH; TdV = TdR + TdU; TkA = TdU - TdR; ci[WS(rs, 14)] = FMA(KP995184726, TdJ, TdG); cr[WS(rs, 17)] = FNMS(KP995184726, TdJ, TdG); ci[WS(rs, 6)] = FMA(KP773010453, TdV, TdO); cr[WS(rs, 25)] = FNMS(KP773010453, TdV, TdO); TkI = TeG - TeD; TeH = TeD + TeG; TkM = Teu + Tet; Tev = Tet - Teu; } { E Tes, TeA, Tkr, Tkz, Tec, TeI; Tes = FNMS(KP831469612, Teb, Te4); Tec = FMA(KP831469612, Teb, Te4); ci[WS(rs, 50)] = FMA(KP956940335, TkI, TkH); cr[WS(rs, 45)] = FMS(KP956940335, TkI, TkH); ci[WS(rs, 58)] = FMA(KP881921264, TkM, TkL); cr[WS(rs, 37)] = FMS(KP881921264, TkM, TkL); TeA = FMA(KP831469612, Tez, Tew); TeI = FNMS(KP831469612, Tez, Tew); Tkt = FNMS(KP980785280, Tkq, Tkp); Tkr = FMA(KP980785280, Tkq, Tkp); cr[WS(rs, 5)] = FMA(KP881921264, Ter, Tec); ci[WS(rs, 26)] = FNMS(KP881921264, Ter, Tec); cr[WS(rs, 13)] = FMA(KP956940335, TeL, TeI); ci[WS(rs, 18)] = FNMS(KP956940335, TeL, TeI); Tkx = FMA(KP980785280, Tkw, Tkv); Tkz = FNMS(KP980785280, Tkw, Tkv); ci[WS(rs, 62)] = FMA(KP995184726, Tks, Tkr); cr[WS(rs, 33)] = FMS(KP995184726, Tks, Tkr); ci[WS(rs, 54)] = FMA(KP773010453, TkA, Tkz); cr[WS(rs, 41)] = FMS(KP773010453, TkA, Tkz); ci[WS(rs, 10)] = FMA(KP881921264, Tev, Tes); cr[WS(rs, 21)] = FNMS(KP881921264, Tev, Tes); ci[WS(rs, 2)] = FMA(KP956940335, TeH, TeA); cr[WS(rs, 29)] = FNMS(KP956940335, TeH, TeA); } } } } } } ci[WS(rs, 46)] = FMA(KP995184726, Tku, Tkt); cr[WS(rs, 49)] = FMS(KP995184726, Tku, Tkt); ci[WS(rs, 38)] = FNMS(KP773010453, Tky, Tkx); cr[WS(rs, 57)] = -(FMA(KP773010453, Tky, Tkx)); }}static const tw_instr twinstr[] = { {TW_FULL, 1, 64}, {TW_NEXT, 1, 0}};static const hc2hc_desc desc = { 64, "hf_64", twinstr, &GENUS, {520, 126, 518, 0} };void X(codelet_hf_64) (planner *p) { X(khc2hc_register) (p, hf_64, &desc);}#else /* HAVE_FMA *//* Generated by: ../../../genfft/gen_hc2hc -compact -variables 4 -pipeline-latency 4 -n 64 -dit -name hf_64 -include hf.h *//* * This function contains 1038 FP additions, 500 FP multiplications, * (or, 808 additions, 270 multiplications, 230 fused multiply/add), * 176 stack variables, 15 constants, and 256 memory accesses */#include "hf.h"static void hf_64(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms){ DK(KP290284677, +0.290284677254462367636192375817395274691476278); DK(KP956940335, +0.956940335732208864935797886980269969482849206); DK(KP881921264, +0.881921264348355029712756863660388349508442621); DK(KP471396736, +0.471396736825997648556387625905254377657460319); DK(KP555570233, +0.555570233019602224742830813948532874374937191); DK(KP831469612, +0.831469612302545237078788377617905756738560812); DK(KP098017140, +0.098017140329560601994195563888641845861136673); DK(KP995184726, +0.995184726672196886244836953109479921575474869); DK(KP773010453, +0.773010453362736960810906609758469800971041293); DK(KP634393284, +0.634393284163645498215171613225493370675687095); DK(KP980785280, +0.980785280403230449126182236134239036973933731); DK(KP195090322, +0.195090322016128267848284868477022240927691618); DK(KP382683432, +0.382683432365089771728459984030398866761344562); DK(KP923879532, +0.923879532511286756128183189396788286822416626); DK(KP707106781, +0.707106781186547524400844362104849039284835938); INT m; for (m = mb, W = W + ((mb - 1) * 126); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 126, MAKE_VOLATILE_STRIDE(rs)) { E Tj, TcL, ThT, Tin, T6b, Taz, TgT, Thn, TG, Thm, TcO, TgO, T6m, Tim, TaC; E ThQ, T14, Tfr, T6y, T9O, TaG, Tc0, TcU, TeE, T1r, Tfq, T6J, T9P, TaJ, Tc1; E TcZ, TeF, T1Q, T2d, Tfu, Tfv, Tfw, Tfx, T6Q, TaM, Tdb, TeI, T71, TaQ, T7a; E TaN, Td6, TeJ, T77, TaP, T2B, T2Y, Tfz, TfA, TfB, TfC, T7h, TaW, Tdm, TeL; E T7s, TaU, T7B, TaX, Tdh, TeM, T7y, TaT, T5j, TfR, Tec, TeX, TfY, Tgy, T8D; E Tbl, T8O, Tbx, T9l, Tbm, TdV, Tf0, T9i, Tbw, T3M, TfL, TdL, TeT, TfI, Tgt; E T7K, Tbd, T7V, Tb3, T8s, Tbe, Tdu, TeQ, T8p, Tb2, T4x, TfJ, TdE, TdM, TfO; E Tgu, T87, T8u, T8i, T8v, Tba, Tbh, Tdz, TdN, Tb7, Tbg, T64, TfZ, Te5, Ted; E TfU, Tgz, T90, T9n, T9b, T9o, Tbt, TbA, Te0, Tee, Tbq, Tbz; { E T1, TgR, T6, TgQ, Tc, T68, Th, T69; T1 = cr[0]; TgR = ci[0]; { E T3, T5, T2, T4; T3 = cr[WS(rs, 32)]; T5 = ci[WS(rs, 32)]; T2 = W[62]; T4 = W[63]; T6 = FMA(T2, T3, T4 * T5); TgQ = FNMS(T4, T3, T2 * T5); } { E T9, Tb, T8, Ta; T9 = cr[WS(rs, 16)]; Tb = ci[WS(rs, 16)]; T8 = W[30]; Ta = W[31]; Tc = FMA(T8, T9, Ta * Tb); T68 = FNMS(Ta, T9, T8 * Tb); } { E Te, Tg, Td, Tf; Te = cr[WS(rs, 48)]; Tg = ci[WS(rs, 48)]; Td = W[94]; Tf = W[95]; Th = FMA(Td, Te, Tf * Tg); T69 = FNMS(Tf, Te, Td * Tg); } { E T7, Ti, ThR, ThS; T7 = T1 + T6; Ti = Tc + Th; Tj = T7 + Ti; TcL = T7 - Ti; ThR = Tc - Th; ThS = TgR - TgQ; ThT = ThR + ThS; Tin = ThS - ThR; } { E T67, T6a, TgP, TgS; T67 = T1 - T6; T6a = T68 - T69; T6b = T67 - T6a; Taz = T67 + T6a; TgP = T68 + T69; TgS = TgQ + TgR; TgT = TgP + TgS; Thn = TgS - TgP; } } { E To, T6d, Tt, T6e, T6c, T6f, Tz, T6i, TE, T6j, T6h, T6k; { E Tl, Tn, Tk, Tm; Tl = cr[WS(rs, 8)]; Tn = ci[WS(rs, 8)]; Tk = W[14]; Tm = W[15]; To = FMA(Tk, Tl, Tm * Tn); T6d = FNMS(Tm, Tl, Tk * Tn); } { E Tq, Ts, Tp, Tr; Tq = cr[WS(rs, 40)]; Ts = ci[WS(rs, 40)]; Tp = W[78]; Tr = W[79]; Tt = FMA(Tp, Tq, Tr * Ts); T6e = FNMS(Tr, Tq, Tp * Ts); } T6c = To - Tt; T6f = T6d - T6e; { E Tw, Ty, Tv, Tx; Tw = cr[WS(rs, 56)]; Ty = ci[WS(rs, 56)]; Tv = W[110]; Tx = W[111]; Tz = FMA(Tv, Tw, Tx * Ty); T6i = FNMS(Tx, Tw, Tv * Ty); } { E TB, TD, TA, TC; TB = cr[WS(rs, 24)]; TD = ci[WS(rs, 24)]; TA = W[46]; TC = W[47]; TE = FMA(TA, TB, TC * TD); T6j = FNMS(TC, TB, TA * TD); } T6h = Tz - TE; T6k = T6i - T6j; { E Tu, TF, TcM, TcN; Tu = To + Tt; TF = Tz + TE; TG = Tu + TF; Thm = Tu - TF; TcM = T6i + T6j; TcN = T6d + T6e; TcO = TcM - TcN; TgO = TcN + TcM; } { E T6g, T6l, TaA, TaB; T6g = T6c - T6f; T6l = T6h + T6k; T6m = KP707106781 * (T6g + T6l); Tim = KP707106781 * (T6l - T6g); TaA = T6c + T6f; TaB = T6h - T6k; TaC = KP707106781 * (TaA + TaB); ThQ = KP707106781 * (TaA - TaB); } } { E TS, TcR, T6o, T6v, T13, TcS, T6r, T6w, T6s, T6x; { E TM, T6t, TR, T6u; { E TJ, TL, TI, TK; TJ = cr[WS(rs, 4)]; TL = ci[WS(rs, 4)]; TI = W[6]; TK = W[7]; TM = FMA(TI, TJ, TK * TL); T6t = FNMS(TK, TJ, TI * TL); } { E TO, TQ, TN, TP; TO = cr[WS(rs, 36)]; TQ = ci[WS(rs, 36)]; TN = W[70]; TP = W[71]; TR = FMA(TN, TO, TP * TQ); T6u = FNMS(TP, TO, TN * TQ); } TS = TM + TR; TcR = T6t + T6u; T6o = TM - TR; T6v = T6t - T6u; } { E TX, T6p, T12, T6q; { E TU, TW, TT, TV; TU = cr[WS(rs, 20)]; TW = ci[WS(rs, 20)]; TT = W[38]; TV = W[39]; TX = FMA(TT, TU, TV * TW); T6p = FNMS(TV, TU, TT * TW); } { E TZ, T11, TY, T10; TZ = cr[WS(rs, 52)]; T11 = ci[WS(rs, 52)]; TY = W[102]; T10 = W[103]; T12 = FMA(TY, TZ, T10 * T11); T6q = FNMS(T10, TZ, TY * T11); } T13 = TX + T12; TcS = T6p + T6q; T6r = T6p - T6q; T6w = TX - T12; } T14 = TS + T13; Tfr = TcR + TcS; T6s = T6o - T6r; T6x = T6v + T6w; T6y = FNMS(KP382683432, T6x, KP923879532 * T6s); T9O = FMA(KP923879532, T6x, KP382683432 * T6s); { E TaE, TaF, TcQ, TcT; TaE = T6v - T6w; TaF = T6o + T6r; TaG = FMA(KP382683432, TaE, KP923879532 * TaF); Tc0 = FNMS(KP923879532, TaE, KP382683432 * TaF); TcQ = TS - T13; TcT = TcR - TcS; TcU = TcQ + TcT; TeE = TcQ - TcT; } } { E T1f, TcW, T6B, T6E, T1q, TcX, T6C, T6H, T6D, T6I; { E T19, T6z, T1e, T6A; { E T16, T18, T15, T17; T16 = cr[WS(rs, 60)]; T18 = ci[WS(rs, 60)]; T15 = W[118]; T17 = W[119]; T19 = FMA(T15, T16, T17 * T18); T6z = FNMS(T17, T16, T15 * T18); } { E T1b, T1d, T1a, T1c; T1b = cr[WS(rs, 28)]; T1d = ci[WS(rs, 28)]; T1a = W[54]; T1c = W[55]; T1e = FMA(T1a, T1b, T1c * T1d); T6A = FNMS(T1c, T1b, T1a * T1d); } T1f = T19 + T1e; TcW = T6z + T6A; T6B = T6z - T6A; T6E = T19 - T1e; } { E T1k, T6F, T1p, T6G; { E T1h, T1j, T1g, T1i; T1h = cr[WS(rs, 12)]; T1j = ci[WS(rs, 12)]; T1g = W[22]; T1i = W[23]; T1k = FMA(T1g, T1h, T1i * T1j); T6F = FNMS(T1i, T1h, T1g * T1j); } { E T1m, T1o, T1l, T1n; T1m = cr[WS(rs, 44)]; T1o = ci[WS(rs, 44)]; T1l = W[86]; T1n = W[87]; T1p = FMA(T1l, T1m, T1n * T1o)
⌨️ 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 + -