?? hb_25.c
字號:
T49 = T47 * T48; T3X = W[20]; T4c = T4a * T48; T42 = W[21]; cr[WS(rs, 16)] = FNMS(T4a, T4b, T49); T41 = T3X * T40; ci[WS(rs, 16)] = FMA(T47, T4b, T4c); } T46 = T42 * T40; { E T3V, T3R, T3U, T3T, T3W, T3S; cr[WS(rs, 11)] = FNMS(T42, T45, T41); T3S = FMA(KP921177326, T3v, T3o); T3w = FNMS(KP921177326, T3v, T3o); ci[WS(rs, 11)] = FMA(T3X, T45, T46); T3V = FNMS(KP921177326, T3O, T3L); T3P = FMA(KP921177326, T3O, T3L); T3R = W[40]; T3U = W[41]; T2d = W[10]; T3T = T3R * T3S; T3W = T3U * T3S; T3y = W[11]; T3x = T2d * T3w; cr[WS(rs, 21)] = FNMS(T3U, T3V, T3T); ci[WS(rs, 21)] = FMA(T3R, T3V, T3W); } } } } } } } } cr[WS(rs, 6)] = FNMS(T3y, T3P, T3x); T3Q = T3y * T3w; ci[WS(rs, 6)] = FMA(T2d, T3P, T3Q); }}static const tw_instr twinstr[] = { {TW_FULL, 1, 25}, {TW_NEXT, 1, 0}};static const hc2hc_desc desc = { 25, "hb_25", twinstr, &GENUS, {84, 48, 316, 0} };void X(codelet_hb_25) (planner *p) { X(khc2hc_register) (p, hb_25, &desc);}#else /* HAVE_FMA *//* Generated by: ../../../genfft/gen_hc2hc -compact -variables 4 -pipeline-latency 4 -sign 1 -n 25 -dif -name hb_25 -include hb.h *//* * This function contains 400 FP additions, 280 FP multiplications, * (or, 260 additions, 140 multiplications, 140 fused multiply/add), * 107 stack variables, 20 constants, and 100 memory accesses */#include "hb.h"static void hb_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(KP992114701, +0.992114701314477831049793042785778521453036709); DK(KP125333233, +0.125333233564304245373118759816508793942918247); DK(KP425779291, +0.425779291565072648862502445744251703979973042); DK(KP904827052, +0.904827052466019527713668647932697593970413911); DK(KP248689887, +0.248689887164854788242283746006447968417567406); DK(KP968583161, +0.968583161128631119490168375464735813836012403); DK(KP770513242, +0.770513242775789230803009636396177847271667672); DK(KP637423989, +0.637423989748689710176712811676016195434917298); DK(KP844327925, +0.844327925502015078548558063966681505381659241); DK(KP535826794, +0.535826794978996618271308767867639978063575346); DK(KP684547105, +0.684547105928688673732283357621209269889519233); DK(KP728968627, +0.728968627421411523146730319055259111372571664); DK(KP481753674, +0.481753674101715274987191502872129653528542010); DK(KP876306680, +0.876306680043863587308115903922062583399064238); DK(KP559016994, +0.559016994374947424102293417182819058860154590); DK(KP250000000, +0.250000000000000000000000000000000000000000000); DK(KP587785252, +0.587785252292473129168705954639072768597652438); DK(KP951056516, +0.951056516295153572116439333379382143405698634); 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 T9, T5Q, T3y, T39, T5v, Ti, Tr, Ts, TZ, T18, T1z, T2k, T4l, T3h, T44; E T5d, T6C, T5C, T6o, T56, T6B, T5B, T6l, T2z, T4m, T3i, T47, T1K, T5w, T3c; E T3B, T5R, TB, TK, TL, T1i, T1r, T1A, T2P, T4o, T3k, T4b, T5s, T6F, T5F; E T6v, T5l, T6E, T5E, T6s, T34, T4p, T3l, T4e; { E T1, T4, T7, T8, T3x, T3w, T37, T38; T1 = cr[0]; { E T2, T3, T5, T6; T2 = cr[WS(rs, 5)]; T3 = ci[WS(rs, 4)]; T4 = T2 + T3; T5 = cr[WS(rs, 10)]; T6 = ci[WS(rs, 9)]; T7 = T5 + T6; T8 = T4 + T7; T3x = T5 - T6; T3w = T2 - T3; } T9 = T1 + T8; T5Q = FMA(KP951056516, T3w, KP587785252 * T3x); T3y = FNMS(KP951056516, T3x, KP587785252 * T3w); T37 = FNMS(KP250000000, T8, T1); T38 = KP559016994 * (T4 - T7); T39 = T37 - T38; T5v = T38 + T37; } { E Ta, T27, T53, T2f, Th, T26, T10, T2p, T58, T2x, T17, T2o, Tj, T2n, T5a; E T2t, Tq, T2s, TR, T2b, T51, T2h, TY, T2g; { E Tg, T2e, Td, T2d; Ta = cr[WS(rs, 1)]; { E Te, Tf, Tb, Tc; Te = cr[WS(rs, 11)]; Tf = ci[WS(rs, 8)]; Tg = Te + Tf; T2e = Te - Tf; Tb = cr[WS(rs, 6)]; Tc = ci[WS(rs, 3)]; Td = Tb + Tc; T2d = Tb - Tc; } T27 = KP559016994 * (Td - Tg); T53 = FMA(KP951056516, T2d, KP587785252 * T2e); T2f = FNMS(KP951056516, T2e, KP587785252 * T2d); Th = Td + Tg; T26 = FNMS(KP250000000, Th, Ta); } { E T16, T2w, T13, T2v; T10 = ci[WS(rs, 20)]; { E T14, T15, T11, T12; T14 = cr[WS(rs, 14)]; T15 = cr[WS(rs, 19)]; T16 = T14 + T15; T2w = T15 - T14; T11 = ci[WS(rs, 15)]; T12 = cr[WS(rs, 24)]; T13 = T11 - T12; T2v = T11 + T12; } T2p = KP559016994 * (T13 + T16); T58 = FMA(KP951056516, T2v, KP587785252 * T2w); T2x = FNMS(KP951056516, T2w, KP587785252 * T2v); T17 = T13 - T16; T2o = FNMS(KP250000000, T17, T10); } { E Tp, T2m, Tm, T2l; Tj = cr[WS(rs, 4)]; { E Tn, To, Tk, Tl; Tn = ci[WS(rs, 10)]; To = ci[WS(rs, 5)]; Tp = Tn + To; T2m = Tn - To; Tk = cr[WS(rs, 9)]; Tl = ci[0]; Tm = Tk + Tl; T2l = Tk - Tl; } T2n = FNMS(KP951056516, T2m, KP587785252 * T2l); T5a = FMA(KP951056516, T2l, KP587785252 * T2m); T2t = KP559016994 * (Tm - Tp); Tq = Tm + Tp; T2s = FNMS(KP250000000, Tq, Tj); } { E TX, T2a, TU, T29; TR = ci[WS(rs, 23)]; { E TV, TW, TS, TT; TV = ci[WS(rs, 13)]; TW = cr[WS(rs, 16)]; TX = TV - TW; T2a = TV + TW; TS = ci[WS(rs, 18)]; TT = cr[WS(rs, 21)]; TU = TS - TT; T29 = TS + TT; } T2b = FNMS(KP951056516, T2a, KP587785252 * T29); T51 = FMA(KP951056516, T29, KP587785252 * T2a); T2h = KP559016994 * (TU - TX); TY = TU + TX; T2g = FNMS(KP250000000, TY, TR); } Ti = Ta + Th; Tr = Tj + Tq; Ts = Ti + Tr; TZ = TR + TY; T18 = T10 + T17; T1z = TZ + T18; { E T2c, T42, T2j, T43, T28, T2i; T28 = T26 - T27; T2c = T28 - T2b; T42 = T28 + T2b; T2i = T2g - T2h; T2j = T2f + T2i; T43 = T2i - T2f; T2k = FNMS(KP481753674, T2j, KP876306680 * T2c); T4l = FMA(KP728968627, T43, KP684547105 * T42); T3h = FMA(KP876306680, T2j, KP481753674 * T2c); T44 = FNMS(KP684547105, T43, KP728968627 * T42); } { E T59, T6n, T5c, T6m, T57, T5b; T57 = T2t + T2s; T59 = T57 - T58; T6n = T57 + T58; T5b = T2o + T2p; T5c = T5a + T5b; T6m = T5b - T5a; T5d = FNMS(KP844327925, T5c, KP535826794 * T59); T6C = FMA(KP637423989, T6m, KP770513242 * T6n); T5C = FMA(KP535826794, T5c, KP844327925 * T59); T6o = FNMS(KP637423989, T6n, KP770513242 * T6m); } { E T52, T6j, T55, T6k, T50, T54; T50 = T27 + T26; T52 = T50 - T51; T6j = T50 + T51; T54 = T2h + T2g; T55 = T53 + T54; T6k = T54 - T53; T56 = FNMS(KP248689887, T55, KP968583161 * T52); T6B = FMA(KP535826794, T6k, KP844327925 * T6j); T5B = FMA(KP968583161, T55, KP248689887 * T52); T6l = FNMS(KP844327925, T6k, KP535826794 * T6j); } { E T2r, T45, T2y, T46, T2q, T2u; T2q = T2o - T2p; T2r = T2n + T2q; T45 = T2q - T2n; T2u = T2s - T2t; T2y = T2u - T2x; T46 = T2u + T2x; T2z = FMA(KP904827052, T2r, KP425779291 * T2y); T4m = FNMS(KP992114701, T45, KP125333233 * T46); T3i = FNMS(KP425779291, T2r, KP904827052 * T2y); T47 = FMA(KP125333233, T45, KP992114701 * T46); } } { E T1C, T1F, T1I, T1J, T3b, T3a, T3z, T3A; T1C = ci[WS(rs, 24)]; { E T1D, T1E, T1G, T1H; T1D = ci[WS(rs, 19)]; T1E = cr[WS(rs, 20)]; T1F = T1D - T1E; T1G = ci[WS(rs, 14)]; T1H = cr[WS(rs, 15)]; T1I = T1G - T1H; T1J = T1F + T1I; T3b = T1G + T1H; T3a = T1D + T1E; } T1K = T1C + T1J; T5w = FMA(KP951056516, T3a, KP587785252 * T3b); T3c = FNMS(KP951056516, T3b, KP587785252 * T3a); T3z = FNMS(KP250000000, T1J, T1C); T3A = KP559016994 * (T1F - T1I); T3B = T3z - T3A; T5R = T3A + T3z; } { E Tt, T2C, T5i, T2K, TA, T2B, T1a, T2G, T5g, T2M, T1h, T2L, TC, T2R, T5p; E T2Z, TJ, T2Q, T1j, T2V, T5n, T31, T1q, T30; { E Tw, T2I, Tz, T2J; Tt = cr[WS(rs, 2)]; { E Tu, Tv, Tx, Ty; Tu = cr[WS(rs, 7)]; Tv = ci[WS(rs, 2)]; Tw = Tu + Tv; T2I = Tu - Tv; Tx = cr[WS(rs, 12)]; Ty = ci[WS(rs, 7)]; Tz = Tx + Ty; T2J = Tx - Ty; } T2C = KP559016994 * (Tw - Tz); T5i = FMA(KP951056516, T2I, KP587785252 * T2J); T2K = FNMS(KP951056516, T2J, KP587785252 * T2I); TA = Tw + Tz; T2B = FNMS(KP250000000, TA, Tt); } { E T1d, T2E, T1g, T2F; T1a = ci[WS(rs, 22)]; { E T1b, T1c, T1e, T1f; T1b = ci[WS(rs, 17)]; T1c = cr[WS(rs, 22)]; T1d = T1b - T1c; T2E = T1b + T1c; T1e = ci[WS(rs, 12)]; T1f = cr[WS(rs, 17)]; T1g = T1e - T1f; T2F = T1e + T1f; } T2G = FNMS(KP951056516, T2F, KP587785252 * T2E); T5g = FMA(KP951056516, T2E, KP587785252 * T2F); T2M = KP559016994 * (T1d - T1g); T1h = T1d + T1g; T2L = FNMS(KP250000000, T1h, T1a); } { E TI, T2Y, TF, T2X; TC = cr[WS(rs, 3)]; { E TG, TH, TD, TE; TG = ci[WS(rs, 11)]; TH = ci[WS(rs, 6)]; TI = TG + TH; T2Y = TG - TH; TD = cr[WS(rs, 8)]; TE = ci[WS(rs, 1)]; TF = TD + TE; T2X = TD - TE; } T2R = KP559016994 * (TF - TI); T5p = FMA(KP951056516, T2X, KP587785252 * T2Y); T2Z = FNMS(KP951056516, T2Y, KP587785252 * T2X); TJ = TF + TI; T2Q = FNMS(KP250000000, TJ, TC); } { E T1p, T2U, T1m, T2T; T1j = ci[WS(rs, 21)]; { E T1n, T1o, T1k, T1l; T1n = cr[WS(rs, 13)]; T1o = cr[WS(rs, 18)]; T1p = T1n + T1o; T2U = T1o - T1n; T1k = ci[WS(rs, 16)]; T1l = cr[WS(rs, 23)]; T1m = T1k - T1l; T2T = T1k + T1l; } T2V = FNMS(KP951056516, T2U, KP587785252 * T2T); T5n = FMA(KP951056516, T2T, KP587785252 * T2U); T31 = KP559016994 * (T1m + T1p); T1q = T1m - T1p; T30 = FNMS(KP250000000, T1q, T1j); } TB = Tt + TA; TK = TC + TJ; TL = TB + TK; T1i = T1a + T1h; T1r = T1j + T1q; T1A = T1i + T1r; { E T2H, T49, T2O, T4a, T2D, T2N; T2D = T2B - T2C; T2H = T2D - T2G; T49 = T2D + T2G; T2N = T2L - T2M; T2O = T2K + T2N; T4a = T2N - T2K; T2P = FNMS(KP844327925, T2O, KP535826794 * T2H); T4o = FMA(KP062790519, T4a, KP998026728 * T49); T3k = FMA(KP535826794, T2O, KP844327925 * T2H); T4b = FNMS(KP998026728, T4a, KP062790519 * T49); } { E T5o, T6u, T5r, T6t, T5m, T5q; T5m = T2R + T2Q; T5o = T5m - T5n; T6u = T5m + T5n; T5q = T30 + T31; T5r = T5p + T5q; T6t = T5q - T5p; T5s = FNMS(KP684547105, T5r, KP728968627 * T5o); T6F = FNMS(KP992114701, T6t, KP125333233 * T6u); T5F = FMA(KP728968627, T5r, KP684547105 * T5o); T6v = FMA(KP125333233, T6t, KP992114701 * T6u); } { E T5h, T6r, T5k, T6q, T5f, T5j; T5f = T2C + T2B; T5h = T5f - T5g; T6r = T5f + T5g; T5j = T2M + T2L; T5k = T5i + T5j; T6q = T5j - T5i; T5l = FNMS(KP481753674, T5k, KP876306680 * T5h); T6E = FNMS(KP425779291, T6q, KP904827052 * T6r); T5E = FMA(KP876306680, T5k, KP481753674 * T5h); T6s = FMA(KP904827052, T6q, KP425779291 * T6r); } {
?? 快捷鍵說明
復制代碼
Ctrl + C
搜索代碼
Ctrl + F
全屏模式
F11
切換主題
Ctrl + Shift + D
顯示快捷鍵
?
增大字號
Ctrl + =
減小字號
Ctrl + -