📄 hf_15.c
字号:
} } ci[WS(rs, 9)] = T3y + T3v; T3z = FNMS(KP250000000, T3y, T3v); T3F = FMA(KP559016994, T3A, T3z); T3B = FNMS(KP559016994, T3A, T3z); cr[WS(rs, 14)] = -(FMA(KP951056516, T3E, T3B)); cr[WS(rs, 11)] = FMS(KP951056516, T3E, T3B); ci[WS(rs, 12)] = FMA(KP951056516, T3G, T3F); cr[WS(rs, 8)] = FMS(KP951056516, T3G, T3F); }}static const tw_instr twinstr[] = { {TW_FULL, 1, 15}, {TW_NEXT, 1, 0}};static const hc2hc_desc desc = { 15, "hf_15", twinstr, &GENUS, {72, 28, 112, 0} };void X(codelet_hf_15) (planner *p) { X(khc2hc_register) (p, hf_15, &desc);}#else /* HAVE_FMA *//* Generated by: ../../../genfft/gen_hc2hc -compact -variables 4 -pipeline-latency 4 -n 15 -dit -name hf_15 -include hf.h *//* * This function contains 184 FP additions, 112 FP multiplications, * (or, 128 additions, 56 multiplications, 56 fused multiply/add), * 65 stack variables, 6 constants, and 60 memory accesses */#include "hf.h"static void hf_15(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms){ DK(KP587785252, +0.587785252292473129168705954639072768597652438); DK(KP951056516, +0.951056516295153572116439333379382143405698634); DK(KP250000000, +0.250000000000000000000000000000000000000000000); DK(KP559016994, +0.559016994374947424102293417182819058860154590); DK(KP500000000, +0.500000000000000000000000000000000000000000000); DK(KP866025403, +0.866025403784438646763723170752936183471402627); INT m; for (m = mb, W = W + ((mb - 1) * 28); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 28, MAKE_VOLATILE_STRIDE(rs)) { E T1q, T2Q, Td, T1n, T2T, T3l, T13, T1k, T1l, T2E, T2F, T3j, T1H, T1T, T2k; E T2w, T2f, T2v, T1M, T1U, Tu, TL, TM, T2H, T2I, T3i, T1w, T1Q, T29, T2t; E T24, T2s, T1B, T1R; { E T1, T2R, T6, T1o, Tb, T1p, Tc, T2S; T1 = cr[0]; T2R = 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); T1o = FNMS(T4, T3, T2 * T5); } { E T8, Ta, T7, T9; T8 = cr[WS(rs, 10)]; Ta = ci[WS(rs, 10)]; T7 = W[18]; T9 = W[19]; Tb = FMA(T7, T8, T9 * Ta); T1p = FNMS(T9, T8, T7 * Ta); } T1q = KP866025403 * (T1o - T1p); T2Q = KP866025403 * (Tb - T6); Tc = T6 + Tb; Td = T1 + Tc; T1n = FNMS(KP500000000, Tc, T1); T2S = T1o + T1p; T2T = FNMS(KP500000000, T2S, T2R); T3l = T2S + T2R; } { E TR, T2c, T18, T2h, TW, T1E, T11, T1F, T12, T2d, T1d, T1J, T1i, T1K, T1j; E T2i; { E TO, TQ, TN, TP; TO = cr[WS(rs, 6)]; TQ = ci[WS(rs, 6)]; TN = W[10]; TP = W[11]; TR = FMA(TN, TO, TP * TQ); T2c = FNMS(TP, TO, TN * TQ); } { E T15, T17, T14, T16; T15 = cr[WS(rs, 9)]; T17 = ci[WS(rs, 9)]; T14 = W[16]; T16 = W[17]; T18 = FMA(T14, T15, T16 * T17); T2h = FNMS(T16, T15, T14 * T17); } { E TT, TV, TS, TU; TT = cr[WS(rs, 11)]; TV = ci[WS(rs, 11)]; TS = W[20]; TU = W[21]; TW = FMA(TS, TT, TU * TV); T1E = FNMS(TU, TT, TS * TV); } { E TY, T10, TX, TZ; TY = cr[WS(rs, 1)]; T10 = ci[WS(rs, 1)]; TX = W[0]; TZ = W[1]; T11 = FMA(TX, TY, TZ * T10); T1F = FNMS(TZ, TY, TX * T10); } T12 = TW + T11; T2d = T1E + T1F; { 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); T1J = FNMS(T1b, T1a, T19 * T1c); } { E T1f, T1h, T1e, T1g; T1f = cr[WS(rs, 4)]; T1h = ci[WS(rs, 4)]; T1e = W[6]; T1g = W[7]; T1i = FMA(T1e, T1f, T1g * T1h); T1K = FNMS(T1g, T1f, T1e * T1h); } T1j = T1d + T1i; T2i = T1J + T1K; { E T1D, T1G, T2g, T2j; T13 = TR + T12; T1k = T18 + T1j; T1l = T13 + T1k; T2E = T2c + T2d; T2F = T2h + T2i; T3j = T2E + T2F; T1D = FNMS(KP500000000, T12, TR); T1G = KP866025403 * (T1E - T1F); T1H = T1D - T1G; T1T = T1D + T1G; T2g = KP866025403 * (T1d - T1i); T2j = FNMS(KP500000000, T2i, T2h); T2k = T2g - T2j; T2w = T2g + T2j; { E T2b, T2e, T1I, T1L; T2b = KP866025403 * (T11 - TW); T2e = FNMS(KP500000000, T2d, T2c); T2f = T2b + T2e; T2v = T2e - T2b; T1I = FNMS(KP500000000, T1j, T18); T1L = KP866025403 * (T1J - T1K); T1M = T1I - T1L; T1U = T1I + T1L; } } } { E Ti, T21, Tz, T26, Tn, T1t, Ts, T1u, Tt, T22, TE, T1y, TJ, T1z, TK; E T27; { E Tf, Th, Te, Tg; Tf = cr[WS(rs, 3)]; Th = ci[WS(rs, 3)]; Te = W[4]; Tg = W[5]; Ti = FMA(Te, Tf, Tg * Th); T21 = FNMS(Tg, Tf, Te * Th); } { E Tw, Ty, Tv, Tx; Tw = cr[WS(rs, 12)]; Ty = ci[WS(rs, 12)]; Tv = W[22]; Tx = W[23]; Tz = FMA(Tv, Tw, Tx * Ty); T26 = FNMS(Tx, Tw, Tv * Ty); } { E Tk, Tm, Tj, Tl; Tk = cr[WS(rs, 8)]; Tm = ci[WS(rs, 8)]; Tj = W[14]; Tl = W[15]; Tn = FMA(Tj, Tk, Tl * Tm); T1t = FNMS(Tl, Tk, Tj * Tm); } { E Tp, Tr, To, Tq; Tp = cr[WS(rs, 13)]; Tr = ci[WS(rs, 13)]; To = W[24]; Tq = W[25]; Ts = FMA(To, Tp, Tq * Tr); T1u = FNMS(Tq, Tp, To * Tr); } Tt = Tn + Ts; T22 = T1t + T1u; { E TB, TD, TA, TC; TB = cr[WS(rs, 2)]; TD = ci[WS(rs, 2)]; TA = W[2]; TC = W[3]; TE = FMA(TA, TB, TC * TD); T1y = FNMS(TC, TB, TA * TD); } { E TG, TI, TF, TH; TG = cr[WS(rs, 7)]; TI = ci[WS(rs, 7)]; TF = W[12]; TH = W[13]; TJ = FMA(TF, TG, TH * TI); T1z = FNMS(TH, TG, TF * TI); } TK = TE + TJ; T27 = T1y + T1z; { E T1s, T1v, T25, T28; Tu = Ti + Tt; TL = Tz + TK; TM = Tu + TL; T2H = T21 + T22; T2I = T26 + T27; T3i = T2H + T2I; T1s = FNMS(KP500000000, Tt, Ti); T1v = KP866025403 * (T1t - T1u); T1w = T1s - T1v; T1Q = T1s + T1v; T25 = KP866025403 * (TJ - TE); T28 = FNMS(KP500000000, T27, T26); T29 = T25 + T28; T2t = T28 - T25; { E T20, T23, T1x, T1A; T20 = KP866025403 * (Ts - Tn); T23 = FNMS(KP500000000, T22, T21); T24 = T20 + T23; T2s = T23 - T20; T1x = FNMS(KP500000000, TK, Tz); T1A = KP866025403 * (T1y - T1z); T1B = T1x - T1A; T1R = T1x + T1A; } } } { E T2C, T1m, T2B, T2K, T2M, T2G, T2J, T2L, T2D; T2C = KP559016994 * (TM - T1l); T1m = TM + T1l; T2B = FNMS(KP250000000, T1m, Td); T2G = T2E - T2F; T2J = T2H - T2I; T2K = FNMS(KP587785252, T2J, KP951056516 * T2G); T2M = FMA(KP951056516, T2J, KP587785252 * T2G); cr[0] = Td + T1m; T2L = T2C + T2B; ci[WS(rs, 5)] = T2L - T2M; cr[WS(rs, 6)] = T2L + T2M; T2D = T2B - T2C; ci[WS(rs, 2)] = T2D - T2K; cr[WS(rs, 3)] = T2D + T2K; } { E T3k, T3m, T3n, T3h, T3p, T3f, T3g, T3q, T3o; T3k = KP559016994 * (T3i - T3j); T3m = T3i + T3j; T3n = FNMS(KP250000000, T3m, T3l); T3f = T1k - T13; T3g = Tu - TL; T3h = FNMS(KP951056516, T3g, KP587785252 * T3f); T3p = FMA(KP587785252, T3g, KP951056516 * T3f); ci[WS(rs, 14)] = T3m + T3l; T3q = T3n - T3k; cr[WS(rs, 12)] = T3p - T3q; ci[WS(rs, 11)] = T3p + T3q; T3o = T3k + T3n; cr[WS(rs, 9)] = T3h - T3o; ci[WS(rs, 8)] = T3h + T3o; } { E T2y, T2A, T1r, T1O, T2p, T2q, T2z, T2r; { E T2u, T2x, T1C, T1N; T2u = T2s - T2t; T2x = T2v - T2w; T2y = FMA(KP951056516, T2u, KP587785252 * T2x); T2A = FNMS(KP587785252, T2u, KP951056516 * T2x); T1r = T1n - T1q; T1C = T1w + T1B; T1N = T1H + T1M; T1O = T1C + T1N; T2p = KP559016994 * (T1C - T1N); T2q = FNMS(KP250000000, T1O, T1r); } cr[WS(rs, 5)] = T1r + T1O; T2z = T2q - T2p; cr[WS(rs, 2)] = T2z - T2A; ci[WS(rs, 6)] = T2z + T2A; T2r = T2p + T2q; ci[0] = T2r - T2y; ci[WS(rs, 3)] = T2r + T2y; } { E T35, T3d, T39, T3a, T38, T3b, T3e, T3c; { E T33, T34, T36, T37; T33 = T1w - T1B; T34 = T1H - T1M; T35 = FMA(KP951056516, T33, KP587785252 * T34); T3d = FNMS(KP587785252, T33, KP951056516 * T34); T39 = T2T - T2Q; T36 = T2v + T2w; T37 = T2s + T2t; T3a = T37 + T36; T38 = KP559016994 * (T36 - T37); T3b = FNMS(KP250000000, T3a, T39); } ci[WS(rs, 9)] = T3a + T39; T3e = T38 + T3b; cr[WS(rs, 8)] = T3d - T3e; ci[WS(rs, 12)] = T3d + T3e; T3c = T38 - T3b; cr[WS(rs, 11)] = T35 + T3c; cr[WS(rs, 14)] = T3c - T35; } { E T2X, T31, T2U, T2P, T2Y, T2Z, T32, T30; { E T2V, T2W, T2N, T2O; T2V = T1T - T1U; T2W = T1Q - T1R; T2X = FNMS(KP587785252, T2W, KP951056516 * T2V); T31 = FMA(KP951056516, T2W, KP587785252 * T2V); T2U = T2Q + T2T; T2N = T2k - T2f; T2O = T24 + T29; T2P = T2N - T2O; T2Y = FMA(KP250000000, T2P, T2U); T2Z = KP559016994 * (T2O + T2N); } cr[WS(rs, 10)] = T2P - T2U; T32 = T2Z + T2Y; ci[WS(rs, 10)] = T31 + T32; ci[WS(rs, 13)] = T32 - T31; T30 = T2Y - T2Z; cr[WS(rs, 13)] = T2X - T30; ci[WS(rs, 7)] = T2X + T30; } { E T2m, T2o, T1P, T1W, T1X, T1Y, T1Z, T2n; { E T2a, T2l, T1S, T1V; T2a = T24 - T29; T2l = T2f + T2k; T2m = FMA(KP951056516, T2a, KP587785252 * T2l); T2o = FNMS(KP587785252, T2a, KP951056516 * T2l); T1P = T1n + T1q; T1S = T1Q + T1R; T1V = T1T + T1U; T1W = T1S + T1V; T1X = KP559016994 * (T1S - T1V); T1Y = FNMS(KP250000000, T1W, T1P); } ci[WS(rs, 4)] = T1P + T1W; T1Z = T1X + T1Y; cr[WS(rs, 4)] = T1Z - T2m; cr[WS(rs, 1)] = T1Z + T2m; T2n = T1Y - T1X; cr[WS(rs, 7)] = T2n - T2o; ci[WS(rs, 1)] = T2n + T2o; } }}static const tw_instr twinstr[] = { {TW_FULL, 1, 15}, {TW_NEXT, 1, 0}};static const hc2hc_desc desc = { 15, "hf_15", twinstr, &GENUS, {128, 56, 56, 0} };void X(codelet_hf_15) (planner *p) { X(khc2hc_register) (p, hf_15, &desc);}#endif /* HAVE_FMA */
⌨️ 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 + -