📄 hb_64.c
字号:
T6t = T6l + T6s; T6W = T6s - T6l; T66 = FMA(KP707106781, T65, T64); T76 = FNMS(KP707106781, T65, T64); T7k = T6c - T69; T6d = T69 + T6c; T6R = T6E - T6F; T6G = T6E + T6F; T6e = FMA(KP923879532, T6d, T66); T6Q = FNMS(KP923879532, T6d, T66); } { E T6N, T6S, T6K, T73, T72, T71; { E T6H, T6w, T70, T6u, T63; T6N = FNMS(KP980785280, T6G, T6D); T6H = FMA(KP980785280, T6G, T6D); T6w = W[3]; T70 = FNMS(KP980785280, T6R, T6Q); T6S = FMA(KP980785280, T6R, T6Q); T6u = FMA(KP980785280, T6t, T6e); T6K = FNMS(KP980785280, T6t, T6e); T63 = W[2]; { E T6Z, T74, T6I, T6v; T73 = FNMS(KP980785280, T6W, T6V); T6X = FMA(KP980785280, T6W, T6V); T72 = W[35]; T6I = T6w * T6u; T6v = T63 * T6u; T6Z = W[34]; T74 = T72 * T70; iio[-WS(ios, 61)] = FMA(T63, T6H, T6I); rio[WS(ios, 2)] = FNMS(T6w, T6H, T6v); T71 = T6Z * T70; iio[-WS(ios, 45)] = FMA(T6Z, T73, T74); } } { E T6M, T6J, T6O, T6L; T6M = W[67]; rio[WS(ios, 18)] = FNMS(T72, T73, T71); T6J = W[66]; T6O = T6M * T6K; T6U = W[99]; T6L = T6J * T6K; T6P = W[98]; iio[-WS(ios, 29)] = FMA(T6J, T6N, T6O); T6Y = T6U * T6S; rio[WS(ios, 34)] = FNMS(T6M, T6N, T6L); T6T = T6P * T6S; } } } { E TfQ, TfD, Tfk, TfI, TfT, TfA, Tfx; { E Tfb, Tfi, Tf4, TfL, TfM, Tfw, Tfp, Tfs; { E Tfu, Tfv, TeW, Tf3; Tg9 = FNMS(KP923879532, TeV, TeU); TeW = FMA(KP923879532, TeV, TeU); iio[-WS(ios, 13)] = FMA(T6P, T6X, T6Y); TfX = Tf2 - TeZ; Tf3 = TeZ + Tf2; rio[WS(ios, 50)] = FNMS(T6U, T6X, T6T); Tfb = FNMS(KP303346683, Tfa, Tf7); Tfu = FMA(KP303346683, Tf7, Tfa); Tfv = FNMS(KP303346683, Tfe, Tfh); Tfi = FMA(KP303346683, Tfh, Tfe); Tf4 = FMA(KP831469612, Tf3, TeW); TfL = FNMS(KP831469612, Tf3, TeW); TfM = Tfu - Tfv; Tfw = Tfu + Tfv; TfW = FNMS(KP923879532, Tfo, Tfn); Tfp = FMA(KP923879532, Tfo, Tfn); Tfs = Tfq + Tfr; Tga = Tfq - Tfr; } { E TfH, Tfj, TfG, Tft; TfN = FMA(KP956940335, TfM, TfL); TfQ = FNMS(KP956940335, TfM, TfL); TfH = Tfi - Tfb; Tfj = Tfb + Tfi; TfG = FNMS(KP831469612, Tfs, Tfp); Tft = FMA(KP831469612, Tfs, Tfp); TfD = FNMS(KP956940335, Tfj, Tf4); Tfk = FMA(KP956940335, Tfj, Tf4); TfI = FMA(KP956940335, TfH, TfG); TfT = FNMS(KP956940335, TfH, TfG); TfA = FNMS(KP956940335, Tfw, Tft); Tfx = FMA(KP956940335, Tfw, Tft); } } { E TfS, TfR, TeT, Tfm, Tfz, TfC; TeT = W[120]; Tfm = W[121]; { E TfP, Tfy, Tfl, TfU; TfP = W[88]; TfS = W[89]; Tfy = TeT * Tfx; Tfl = TeT * Tfk; TfU = TfP * TfT; TfR = TfP * TfQ; rio[WS(ios, 61)] = FNMS(Tfm, Tfk, Tfy); iio[-WS(ios, 2)] = FMA(Tfm, Tfx, Tfl); rio[WS(ios, 45)] = FNMS(TfS, TfQ, TfU); } iio[-WS(ios, 18)] = FMA(TfS, TfT, TfR); Tfz = W[56]; TfC = W[57]; { E TfF, TfE, TfB, TfO; TfF = W[24]; TfK = W[25]; TfE = Tfz * TfD; TfB = Tfz * TfA; TfO = TfF * TfN; TfJ = TfF * TfI; iio[-WS(ios, 34)] = FMA(TfC, TfA, TfE); rio[WS(ios, 29)] = FNMS(TfC, TfD, TfB); iio[-WS(ios, 50)] = FMA(TfK, TfI, TfO); } } } } rio[WS(ios, 13)] = FNMS(TfK, TfN, TfJ); { E T7F, T7C, T7x, T7G, T7B; { E T7D, T7l, T7f, T7E, T7z, T7o, T78, T7y; { E T7m, T7n, T7b, T7e; T7m = FNMS(KP668178637, T79, T7a); T7b = FMA(KP668178637, T7a, T79); T7e = FNMS(KP668178637, T7d, T7c); T7n = FMA(KP668178637, T7c, T7d); T7D = FNMS(KP923879532, T7k, T7j); T7l = FMA(KP923879532, T7k, T7j); T7f = T7b + T7e; T7E = T7b - T7e; T7z = T7n - T7m; T7o = T7m + T7n; T78 = FMA(KP923879532, T77, T76); T7y = FNMS(KP923879532, T77, T76); } { E T7v, T7A, T7s, T7L, T7K, T7J; { E T7p, T7i, T7I, T7g, T75; T7v = FNMS(KP831469612, T7o, T7l); T7p = FMA(KP831469612, T7o, T7l); T7i = W[115]; T7I = FNMS(KP831469612, T7z, T7y); T7A = FMA(KP831469612, T7z, T7y); T7g = FMA(KP831469612, T7f, T78); T7s = FNMS(KP831469612, T7f, T78); T75 = W[114]; { E T7H, T7M, T7q, T7h; T7L = FNMS(KP831469612, T7E, T7D); T7F = FMA(KP831469612, T7E, T7D); T7K = W[83]; T7q = T7i * T7g; T7h = T75 * T7g; T7H = W[82]; T7M = T7K * T7I; iio[-WS(ios, 5)] = FMA(T75, T7p, T7q); rio[WS(ios, 58)] = FNMS(T7i, T7p, T7h); T7J = T7H * T7I; iio[-WS(ios, 21)] = FMA(T7H, T7L, T7M); } } { E T7u, T7r, T7w, T7t; T7u = W[51]; rio[WS(ios, 42)] = FNMS(T7K, T7L, T7J); T7r = W[50]; T7w = T7u * T7s; T7C = W[19]; T7t = T7r * T7s; T7x = W[18]; iio[-WS(ios, 37)] = FMA(T7r, T7v, T7w); T7G = T7C * T7A; rio[WS(ios, 26)] = FNMS(T7u, T7v, T7t); T7B = T7x * T7A; } } } { E Tgy, Tgl, Tg6, Tgq, TgB, Tgi, Tgf; { E Tg1, Tg4, TfY, Tgt, Tgu, Tge, Tgc, Tgd; iio[-WS(ios, 53)] = FMA(T7x, T7F, T7G); rio[WS(ios, 10)] = FNMS(T7C, T7F, T7B); Tg1 = FNMS(KP534511135, Tg0, TfZ); Tgc = FMA(KP534511135, TfZ, Tg0); Tgd = FNMS(KP534511135, Tg2, Tg3); Tg4 = FMA(KP534511135, Tg3, Tg2); TfY = FMA(KP831469612, TfX, TfW); Tgt = FNMS(KP831469612, TfX, TfW); Tgu = Tgc - Tgd; Tge = Tgc + Tgd; { E Tgp, Tg5, Tgo, Tgb; Tgv = FMA(KP881921264, Tgu, Tgt); Tgy = FNMS(KP881921264, Tgu, Tgt); Tgp = Tg4 - Tg1; Tg5 = Tg1 + Tg4; Tgo = FNMS(KP831469612, Tga, Tg9); Tgb = FMA(KP831469612, Tga, Tg9); Tgl = FNMS(KP881921264, Tg5, TfY); Tg6 = FMA(KP881921264, Tg5, TfY); Tgq = FMA(KP881921264, Tgp, Tgo); TgB = FNMS(KP881921264, Tgp, Tgo); Tgi = FNMS(KP881921264, Tge, Tgb); Tgf = FMA(KP881921264, Tge, Tgb); } } { E TgA, Tgz, TfV, Tg8, Tgh, Tgk; TfV = W[8]; Tg8 = W[9]; { E Tgx, Tgg, Tg7, TgC; Tgx = W[40]; TgA = W[41]; Tgg = TfV * Tgf; Tg7 = TfV * Tg6; TgC = Tgx * TgB; Tgz = Tgx * Tgy; iio[-WS(ios, 58)] = FMA(Tg8, Tg6, Tgg); rio[WS(ios, 5)] = FNMS(Tg8, Tgf, Tg7); iio[-WS(ios, 42)] = FMA(TgA, Tgy, TgC); } rio[WS(ios, 21)] = FNMS(TgA, TgB, Tgz); Tgh = W[72]; Tgk = W[73]; { E Tgn, Tgm, Tgj, Tgw; Tgn = W[104]; Tgs = W[105]; Tgm = Tgh * Tgl; Tgj = Tgh * Tgi; Tgw = Tgn * Tgv; Tgr = Tgn * Tgq; rio[WS(ios, 37)] = FNMS(Tgk, Tgi, Tgm); iio[-WS(ios, 26)] = FMA(Tgk, Tgl, Tgj); rio[WS(ios, 53)] = FNMS(Tgs, Tgq, Tgw); } } } } } } } iio[-WS(ios, 10)] = FMA(Tgs, Tgv, Tgr); } return W;}static const tw_instr twinstr[] = { {TW_FULL, 0, 64}, {TW_NEXT, 1, 0}};static const hc2hc_desc desc = { 64, "hb_64", twinstr, &GENUS, {520, 126, 518, 0}, 0, 0, 0 };void X(codelet_hb_64) (planner *p) { X(khc2hc_register) (p, hb_64, &desc);}#else /* HAVE_FMA *//* Generated by: ../../../genfft/gen_hc2hc -compact -variables 4 -pipeline-latency 4 -sign 1 -n 64 -dif -name hb_64 -include hb.h *//* * This function contains 1038 FP additions, 500 FP multiplications, * (or, 808 additions, 270 multiplications, 230 fused multiply/add), * 196 stack variables, and 256 memory accesses *//* * Generator Id's : * $Id: algsimp.ml,v 1.8 2006-01-05 03:04:27 stevenj Exp $ * $Id: fft.ml,v 1.4 2006-01-05 03:04:27 stevenj Exp $ * $Id: gen_hc2hc.ml,v 1.15 2006-01-05 03:04:27 stevenj Exp $ */#include "hb.h"static const R *hb_64(R *rio, R *iio, const R *W, stride ios, INT m, INT dist){ DK(KP634393284, +0.634393284163645498215171613225493370675687095); DK(KP773010453, +0.773010453362736960810906609758469800971041293); DK(KP098017140, +0.098017140329560601994195563888641845861136673); DK(KP995184726, +0.995184726672196886244836953109479921575474869); DK(KP471396736, +0.471396736825997648556387625905254377657460319); DK(KP881921264, +0.881921264348355029712756863660388349508442621); DK(KP290284677, +0.290284677254462367636192375817395274691476278); DK(KP956940335, +0.956940335732208864935797886980269969482849206); DK(KP195090322, +0.195090322016128267848284868477022240927691618); DK(KP980785280, +0.980785280403230449126182236134239036973933731); DK(KP555570233, +0.555570233019602224742830813948532874374937191); DK(KP831469612, +0.831469612302545237078788377617905756738560812); DK(KP382683432, +0.382683432365089771728459984030398866761344562); DK(KP923879532, +0.923879532511286756128183189396788286822416626); DK(KP707106781, +0.707106781186547524400844362104849039284835938); INT i; for (i = m - 2; i > 0; i = i - 2, rio = rio + dist, iio = iio - dist, W = W + 126, MAKE_VOLATILE_STRIDE(ios)) { E Tf, T7i, Tfa, ThM, Tgp, ThH, T2c, T5O, T4T, T6n, Tcp, Ted, TcA, TdE, T87; E T9o, TK, T93, T2P, T4F, Tfo, Thz, T5T, T6j, Tbx, TdI, Tfl, ThA, T7r, T81; E TbE, TdH, TZ, T94, T38, T4G, Tfv, ThC, T5W, T6k, TbQ, TdK, Tfs, ThD, T7w; E T82, TbX, TdL, Tu, T84, Tfh, ThG, Tgm, ThN, T2v, T6m, T4K, T5P, Tce, TdF; E TcD, Tec, T7l, T9p, T1L, T20, T9c, T9d, T9e, T9f, T40, T66, Tg1, Thu, Tg8; E Thv, Tg5, Thr, T4n, T67, T4j, T69, T4w, T6a, TaT, TdW, Tb8, TdZ, TfU, Ths; E T7O, T8y, T7T, T8z, Tbc, TdX, Tbj, Te0, T1g, T1v, T97, T98, T99, T9a, T3j; E T5Z, TfI, Thk, TfP, Thl, TfM, Tho, T3G, T60, T3C, T62, T3P, T63, Tak, TdQ; E Tav, TdT, TfB, Thn, T7D, T8v, T7I, T8w, TaD, TdP, TaG, TdS; { E T3, Tcm, T4O, Tcv, T6, Tcu, T4R, Tcn, Td, Tcy, T2a, Tch, Ta, Tcx, T27; E Tck; { E T1, T2, T4P, T4Q; T1 = rio[0]; T2 = iio[-WS(ios, 32)]; T3 = T1 + T2; Tcm = T1 - T2; { E T4M, T4N, T4, T5; T4M = iio[0]; T4N = rio[WS(ios, 32)]; T4O = T4M - T4N; Tcv = T4M + T4N; T4 = rio[WS(ios, 16)]; T5 = iio[-WS(ios, 48)]; T6 = T4 + T5; Tcu = T4 - T5; } T4P = iio[-WS(ios, 16)]; T4Q = rio[WS(ios, 48)]; T4R = T4P - T4Q; Tcn = T4P + T4Q; { E Tb, Tc, Tcf, T28, T29, Tcg; Tb = iio[-WS(ios, 56)]; Tc = rio[WS(ios, 24)]; Tcf = Tb - Tc; T28 = iio[-WS(ios, 24)]; T29 = rio[WS(ios, 56)]; Tcg = T29 + T28; Td = Tb + Tc; Tcy = Tcf + Tcg; T2a = T28 - T29; Tch = Tcf - Tcg; } { E T8, T9, Tcj, T25, T26, Tci; T8 = rio[WS(ios, 8)]; T9 = iio[-WS(ios, 40)]; Tcj = T8 - T9; T25 = iio[-WS(ios, 8)]; T26 = rio[WS(ios, 40)]; Tci = T25 + T26; Ta = T8 + T9; Tcx = Tcj + Tci; T27 = T25 - T26; Tck = Tci - Tcj; } } { E T7, Te, Tf8, Tf9; T7 = T3 + T6; Te = Ta + Td; Tf = T7 + Te; T7i = T7 - Te; Tf8 = Tcv - Tcu; Tf9 = KP707106781 * (Tck + Tch); Tfa = Tf8 + Tf9; ThM = Tf8 - Tf9; } { E Tgn, Tgo, T24, T2b; Tgn = KP707106781 * (Tcx + Tcy); Tgo = Tcm + Tcn; Tgp = Tgn + Tgo; ThH = Tgo - Tgn; T24 = T3 - T6; T2b = T27 - T2a; T2c = T24 + T2b; T5O = T24 - T2b; } { E T4L, T4S, Tcl, Tco; T4L = Td - Ta; T4S = T4O - T4R; T4T = T4L + T4S; T6n = T4S - T4L; Tcl = KP707106781 * (Tch - Tck); Tco = Tcm - Tcn; Tcp = Tcl + Tco; Ted = Tco - Tcl; } { E Tcw, Tcz, T85, T86; Tcw = Tcu + Tcv; Tcz = KP707106781 * (Tcx - Tcy); TcA = Tcw + Tcz; TdE = Tcw - Tcz; T85 = T4O + T4R; T86 = T27 + T2a; T87 = T85 - T86; T9o = T86 + T85; } } { E TC, Tby, T2x, Tbu, T2N, Tbz, T7o, Tbv, TJ, TbB, TbC, T2E, T2G, Tbp, Tbs; E T7p, Tfj, Tfk; { E Tw, Tx, Ty, Tz, TA, TB; Tw = rio[WS(ios, 2)]; Tx = iio[-WS(ios, 34)]; Ty = Tw + Tx; Tz = rio[WS(ios, 18)]; TA = iio[-WS(ios, 50)]; TB = Tz + TA; TC = Ty + TB; Tby = Tz - TA; T2x = Ty - TB; Tbu = Tw - Tx; } { E T2H, T2I, T2J, T2K, T2L, T2M; T2H = iio[-WS(ios, 2)]; T2I = rio[WS(ios, 34)]; T2J = T2H - T2I; T2K = iio[-WS(ios, 18)]; T2L =
⌨️ 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 + -