📄 hb2_25.c
字号:
T44 = FNMS(KP904730450, T43, T42); } } } } } } { E T45, T4L, T3G, T3I; T45 = FNMS(KP242145790, T44, T41); T4L = FMA(KP968583161, T44, T41); T3G = FMA(KP904730450, T3F, T3a); T3I = FNMS(KP904730450, T3F, T3a); { E T4v, T47, T4J, T3H; T4v = FNMS(KP541454447, T46, T45); T47 = FMA(KP541454447, T46, T45); T4J = FMA(KP968583161, T3G, T2F); T3H = FNMS(KP242145790, T3G, T2F); { E T4b, T4j, T4x, T4H; T4b = FMA(KP921177326, T4a, T47); T4j = FNMS(KP921177326, T4a, T47); T4x = FNMS(KP833417178, T4w, T4v); T4H = FMA(KP833417178, T4w, T4v); { E T3J, T4p, T4M, T4K; T3J = FMA(KP541454447, T3I, T3H); T4p = FNMS(KP541454447, T3I, T3H); T4M = TQ * T4J; T4K = TN * T4J; { E T4D, T4r, T4f, T3R; T4D = FMA(KP833417178, T4q, T4p); T4r = FNMS(KP833417178, T4q, T4p); T4f = FMA(KP921177326, T3Q, T3J); T3R = FNMS(KP921177326, T3Q, T3J); ci[WS(rs, 1)] = FMA(TN, T4L, T4M); cr[WS(rs, 1)] = FNMS(TQ, T4L, T4K); { E T4y, T4s, T4I, T4E; T4y = T4u * T4r; T4s = T4o * T4r; T4I = T4G * T4D; T4E = T4C * T4D; { E T4c, T3S, T4k, T4g; T4c = T3U * T3R; T3S = T2y * T3R; T4k = T4i * T4f; T4g = T4e * T4f; ci[WS(rs, 11)] = FMA(T4o, T4x, T4y); cr[WS(rs, 11)] = FNMS(T4u, T4x, T4s); ci[WS(rs, 16)] = FMA(T4C, T4H, T4I); cr[WS(rs, 16)] = FNMS(T4G, T4H, T4E); ci[WS(rs, 6)] = FMA(T2y, T4b, T4c); cr[WS(rs, 6)] = FNMS(T3U, T4b, T3S); ci[WS(rs, 21)] = FMA(T4e, T4j, T4k); cr[WS(rs, 21)] = FNMS(T4i, T4j, T4g); } } } } } } } }}static const tw_instr twinstr[] = { {TW_CEXP, 1, 1}, {TW_CEXP, 1, 3}, {TW_CEXP, 1, 9}, {TW_CEXP, 1, 24}, {TW_NEXT, 1, 0}};static const hc2hc_desc desc = { 25, "hb2_25", twinstr, &GENUS, {84, 78, 356, 0} };void X(codelet_hb2_25) (planner *p) { X(khc2hc_register) (p, hb2_25, &desc);}#else /* HAVE_FMA *//* Generated by: ../../../genfft/gen_hc2hc -compact -variables 4 -pipeline-latency 4 -sign 1 -twiddle-log3 -precompute-twiddles -n 25 -dif -name hb2_25 -include hb.h *//* * This function contains 440 FP additions, 340 FP multiplications, * (or, 280 additions, 180 multiplications, 160 fused multiply/add), * 155 stack variables, 20 constants, and 100 memory accesses */#include "hb.h"static void hb2_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) * 8); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 8, MAKE_VOLATILE_STRIDE(rs)) { E TN, TQ, TO, TR, TT, TY, T2t, T2r, TZ, TU, T4f, T4l, T2d, T4v, T5m; E T2j, T5l, T4X, T2v, T11, T3R, T1L, T5d, T6x, T5h, T6t, T25, T26, T27, T29; E T6D, T7v, T49, T7l, T7p, T7t, T2p, T2n, T4b, T4p, T5n, T6B, T5b, T5p, T6p; E T6r, T59, T4r; { E T2c, T4j, T2h, T4e, T2b, T4k, T2i, T4d; { E TP, TX, TS, TW; TN = W[0]; TQ = W[1]; TO = W[2]; TR = W[3]; TP = TN * TO; TX = TQ * TO; TS = TQ * TR; TW = TN * TR; TT = TP - TS; TY = TW + TX; T2t = TW - TX; T2r = TP + TS; TZ = W[5]; T2c = TQ * TZ; T4j = TO * TZ; T2h = TN * TZ; T4e = TR * TZ; TU = W[4]; T2b = TN * TU; T4k = TR * TU; T2i = TQ * TU; T4d = TO * TU; } T4f = T4d - T4e; T4l = T4j + T4k; { E T2s, T2u, TV, T10, T3P, T3Q, T1J, T1K; T2d = T2b - T2c; T4v = T2b + T2c; T5m = T4j - T4k; T2j = T2h + T2i; T5l = T4d + T4e; T4X = T2h - T2i; T2s = T2r * TU; T2u = T2t * TZ; T2v = T2s + T2u; TV = TT * TU; T10 = TY * TZ; T11 = TV + T10; T3P = T2r * TZ; T3Q = T2t * TU; T3R = T3P - T3Q; T1J = TT * TZ; T1K = TY * TU; T1L = T1J - T1K; T5d = TV - T10; T6x = T3P + T3Q; T5h = T1J + T1K; T6t = T2s - T2u; T25 = W[6]; T26 = W[7]; T27 = FMA(TT, T25, TY * T26); T29 = FNMS(TY, T25, TT * T26); T6D = FNMS(T4X, T25, T4v * T26); T7v = FNMS(T1L, T25, T11 * T26); T49 = FMA(T2r, T25, T2t * T26); T7l = FMA(T2d, T25, T2j * T26); T7p = FNMS(T2j, T25, T2d * T26); T7t = FMA(T11, T25, T1L * T26); T2p = FNMS(TZ, T25, TU * T26); T2n = FMA(TU, T25, TZ * T26); T4b = FNMS(T2t, T25, T2r * T26); T4p = FMA(T2v, T25, T3R * T26); T5n = FMA(T5l, T25, T5m * T26); T6B = FMA(T4v, T25, T4X * T26); T5b = FNMS(TQ, T25, TN * T26); T5p = FNMS(T5m, T25, T5l * T26); T6p = FMA(TO, T25, TR * T26); T6r = FNMS(TR, T25, TO * T26); T59 = FMA(TN, T25, TQ * T26); T4r = FNMS(T3R, T25, T2v * T26); } } { E T9, T6i, T40, T3z, T5Y, Ti, Tr, Ts, T1d, T1m, T1P, T2K, T4P, T3H, T4y; E T5G, T71, T65, T6N, T5z, T70, T64, T6K, T2Z, T4Q, T3I, T4B, T20, T5Z, T3C; E T43, T6j, TB, TK, TL, T1w, T1F, T1Q, T3f, T4S, T3K, T4F, T5V, T74, T68; E T6U, T5O, T73, T67, T6R, T3u, T4T, T3L, T4I; { E T1, T4, T7, T8, T3Z, T3Y, T3x, T3y; 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; T3Z = T5 - T6; T3Y = T2 - T3; } T9 = T1 + T8; T6i = FMA(KP951056516, T3Y, KP587785252 * T3Z); T40 = FNMS(KP951056516, T3Z, KP587785252 * T3Y); T3x = FNMS(KP250000000, T8, T1); T3y = KP559016994 * (T4 - T7); T3z = T3x - T3y; T5Y = T3y + T3x; } { E Ta, T2x, T5w, T2F, Th, T2w, T1e, T2P, T5B, T2X, T1l, T2O, Tj, T2N, T5D; E T2T, Tq, T2S, T15, T2B, T5u, T2H, T1c, 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; } T2x = KP559016994 * (Td - Tg); T5w = FMA(KP951056516, T2D, KP587785252 * T2E); T2F = FNMS(KP951056516, T2E, KP587785252 * T2D); Th = Td + Tg; T2w = FNMS(KP250000000, Th, Ta); } { E T1k, T2W, T1h, T2V; T1e = ci[WS(rs, 20)]; { E T1i, T1j, T1f, T1g; T1i = cr[WS(rs, 14)]; T1j = cr[WS(rs, 19)]; T1k = T1i + T1j; T2W = T1j - T1i; T1f = ci[WS(rs, 15)]; T1g = cr[WS(rs, 24)]; T1h = T1f - T1g; T2V = T1f + T1g; } T2P = KP559016994 * (T1h + T1k); T5B = FMA(KP951056516, T2V, KP587785252 * T2W); T2X = FNMS(KP951056516, T2W, KP587785252 * T2V); T1l = T1h - T1k; T2O = FNMS(KP250000000, T1l, T1e); } { 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); T5D = FMA(KP951056516, T2L, KP587785252 * T2M); T2T = KP559016994 * (Tm - Tp); Tq = Tm + Tp; T2S = FNMS(KP250000000, Tq, Tj); } { E T1b, T2A, T18, T2z; T15 = ci[WS(rs, 23)]; { E T19, T1a, T16, T17; T19 = ci[WS(rs, 13)]; T1a = cr[WS(rs, 16)]; T1b = T19 - T1a; T2A = T19 + T1a; T16 = ci[WS(rs, 18)]; T17 = cr[WS(rs, 21)]; T18 = T16 - T17; T2z = T16 + T17; } T2B = FNMS(KP951056516, T2A, KP587785252 * T2z); T5u = FMA(KP951056516, T2z, KP587785252 * T2A); T2H = KP559016994 * (T18 - T1b); T1c = T18 + T1b; T2G = FNMS(KP250000000, T1c, T15); } Ti = Ta + Th; Tr = Tj + Tq; Ts = Ti + Tr; T1d = T15 + T1c; T1m = T1e + T1l; T1P = T1d + T1m; { E T2C, T4w, T2J, T4x, T2y, T2I; T2y = T2w - T2x; T2C = T2y - T2B; T4w = T2y + T2B; T2I = T2G - T2H; T2J = T2F + T2I; T4x = T2I - T2F; T2K = FNMS(KP481753674, T2J, KP876306680 * T2C); T4P = FMA(KP728968627, T4x, KP684547105 * T4w); T3H = FMA(KP876306680, T2J, KP481753674 * T2C); T4y = FNMS(KP684547105, T4x, KP728968627 * T4w); } { E T5C, T6M, T5F, T6L, T5A, T5E; T5A = T2T + T2S; T5C = T5A - T5B; T6M = T5A + T5B; T5E = T2O + T2P; T5F = T5D + T5E; T6L = T5E - T5D; T5G = FNMS(KP844327925, T5F, KP535826794 * T5C); T71 = FMA(KP637423989, T6L, KP770513242 * T6M); T65 = FMA(KP535826794, T5F, KP844327925 * T5C); T6N = FNMS(KP637423989, T6M, KP770513242 * T6L); } { E T5v, T6I, T5y, T6J, T5t, T5x; T5t = T2x + T2w; T5v = T5t - T5u; T6I = T5t + T5u; T5x = T2H + T2G; T5y = T5w + T5x; T6J = T5x - T5w; T5z = FNMS(KP248689887, T5y, KP968583161 * T5v); T70 = FMA(KP535826794, T6J, KP844327925 * T6I); T64 = FMA(KP968583161, T5y, KP248689887 * T5v); T6K = FNMS(KP844327925, T6J, KP535826794 * T6I); } { E T2R, T4z, T2Y, T4A, T2Q, T2U; T2Q = T2O - T2P; T2R = T2N + T2Q; T4z = T2Q - T2N; T2U = T2S - T2T; T2Y = T2U - T2X; T4A = T2U + T2X; T2Z = FMA(KP904827052, T2R, KP425779291 * T2Y); T4Q = FNMS(KP992114701, T4z, KP125333233 * T4A); T3I = FNMS(KP425779291, T2R, KP904827052 * T2Y); T4B = FMA(KP125333233, T4z, KP992114701 * T4A); } } { E T1S, T1V, T1Y, T1Z, T3B, T3A, T41, T42; T1S = ci[WS(rs, 24)]; { E T1T, T1U, T1W, T1X; T1T = ci[WS(rs, 19)]; T1U = cr[WS(rs, 20)]; T1V = T1T - T1U; T1W = ci[WS(rs, 14)]; T1X = cr[WS(rs, 15)]; T1Y = T1W - T1X; T1Z = T1V + T1Y; T3B = T1W + T1X; T3A = T1T + T1U; } T20 = T1S + T1Z; T5Z = FMA(KP951056516, T3A, KP587785252 * T3B); T3C = FNMS(KP951056516, T3B, KP587785252 * T3A); T41 = FNMS(KP250000000, T1Z, T1S); T42 = KP559016994 * (T1V - T1Y); T43 = T41 - T42; T6j = T42 + T41; } { E Tt, T32, T5L, T3a, TA, T31, T1o, T36, T5J, T3c, T1v, T3b, TC, T3h, T5S; E T3p, TJ, T3g, T1x, T3l, T5Q, T3r, T1E, T3q; { E Tw, T38, Tz, T39; Tt = cr[WS(rs, 2)]; { E Tu, Tv, Tx, Ty; Tu = cr[WS(rs, 7)]; Tv = ci[WS(rs, 2)]; Tw = Tu + Tv; T38 = Tu - Tv; Tx = cr[WS(rs, 12)]; Ty = ci[WS(rs, 7)]; Tz = Tx + Ty; T39 = Tx - Ty; } T32 = KP559016994 * (Tw - Tz); T5L = FMA(KP951056516, T38, KP587785252 * T39); T3a = FNMS(KP951056516, T39, KP587785252 * T38); TA = Tw + Tz; T31 = FNMS(KP250000000, TA, Tt); } { E T1r, T34, T1u, T35; T1o = ci[WS(rs, 22)]; { E T1p, T1q, T1s, T1t; T1p = ci[WS(rs, 17)];
⌨️ 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 + -