📄 hc2cf_20.c
字号:
Rm[WS(rs, 2)] = FNMS(KP951056516, T3N, T3M); Rm[WS(rs, 6)] = FMA(KP951056516, T3N, T3M); T4U = T4Q - T4R; T4S = T4Q + T4R; } } } } Ip[WS(rs, 5)] = T4S + T4P; T4T = FNMS(KP250000000, T4S, T4P); T4Z = FMA(KP559016994, T4U, T4T); T4V = FNMS(KP559016994, T4U, T4T); Im[WS(rs, 2)] = -(FMA(KP951056516, T4Y, T4V)); Im[WS(rs, 6)] = FMS(KP951056516, T4Y, T4V); Ip[WS(rs, 9)] = FMA(KP951056516, T50, T4Z); Ip[WS(rs, 1)] = FNMS(KP951056516, T50, T4Z); }}static const tw_instr twinstr[] = { {TW_FULL, 1, 20}, {TW_NEXT, 1, 0}};static const hc2c_desc desc = { 20, "hc2cf_20", twinstr, &GENUS, {136, 38, 110, 0} };void X(codelet_hc2cf_20) (planner *p) { X(khc2c_register) (p, hc2cf_20, &desc, HC2C_VIA_RDFT);}#else /* HAVE_FMA *//* Generated by: ../../../genfft/gen_hc2c -compact -variables 4 -pipeline-latency 4 -n 20 -dit -name hc2cf_20 -include hc2cf.h *//* * This function contains 246 FP additions, 124 FP multiplications, * (or, 184 additions, 62 multiplications, 62 fused multiply/add), * 85 stack variables, 4 constants, and 80 memory accesses */#include "hc2cf.h"static void hc2cf_20(R *Rp, R *Ip, R *Rm, R *Im, 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); INT m; for (m = mb, W = W + ((mb - 1) * 38); m < me; m = m + 1, Rp = Rp + ms, Ip = Ip + ms, Rm = Rm - ms, Im = Im - ms, W = W + 38, MAKE_VOLATILE_STRIDE(rs)) { E Tj, T1R, T4j, T4s, T2q, T37, T3Q, T42, T1r, T1O, T1P, T3p, T3s, T3K, T3A; E T3B, T3Z, T1V, T1W, T1X, T23, T28, T4q, T2W, T2X, T4f, T33, T34, T35, T2G; E T2L, T2M, TG, T13, T14, T3i, T3l, T3J, T3D, T3E, T40, T1S, T1T, T1U, T2e; E T2j, T4p, T2T, T2U, T4e, T30, T31, T32, T2v, T2A, T2B; { E T1, T3O, T6, T3N, Tc, T2n, Th, T2o; T1 = Rp[0]; T3O = Rm[0]; { E T3, T5, T2, T4; T3 = Rp[WS(rs, 5)]; T5 = Rm[WS(rs, 5)]; T2 = W[18]; T4 = W[19]; T6 = FMA(T2, T3, T4 * T5); T3N = FNMS(T4, T3, T2 * T5); } { E T9, Tb, T8, Ta; T9 = Ip[WS(rs, 2)]; Tb = Im[WS(rs, 2)]; T8 = W[8]; Ta = W[9]; Tc = FMA(T8, T9, Ta * Tb); T2n = FNMS(Ta, T9, T8 * Tb); } { E Te, Tg, Td, Tf; Te = Ip[WS(rs, 7)]; Tg = Im[WS(rs, 7)]; Td = W[28]; Tf = W[29]; Th = FMA(Td, Te, Tf * Tg); T2o = FNMS(Tf, Te, Td * Tg); } { E T7, Ti, T4h, T4i; T7 = T1 + T6; Ti = Tc + Th; Tj = T7 - Ti; T1R = T7 + Ti; T4h = T3O - T3N; T4i = Tc - Th; T4j = T4h - T4i; T4s = T4i + T4h; } { E T2m, T2p, T3M, T3P; T2m = T1 - T6; T2p = T2n - T2o; T2q = T2m - T2p; T37 = T2m + T2p; T3M = T2n + T2o; T3P = T3N + T3O; T3Q = T3M + T3P; T42 = T3P - T3M; } } { E T1f, T3n, T21, T2C, T1N, T3r, T27, T2K, T1q, T3o, T22, T2F, T1C, T3q, T26; E T2H; { E T19, T1Z, T1e, T20; { E T16, T18, T15, T17; T16 = Rp[WS(rs, 4)]; T18 = Rm[WS(rs, 4)]; T15 = W[14]; T17 = W[15]; T19 = FMA(T15, T16, T17 * T18); T1Z = FNMS(T17, T16, T15 * T18); } { E T1b, T1d, T1a, T1c; T1b = Rp[WS(rs, 9)]; T1d = Rm[WS(rs, 9)]; T1a = W[34]; T1c = W[35]; T1e = FMA(T1a, T1b, T1c * T1d); T20 = FNMS(T1c, T1b, T1a * T1d); } T1f = T19 + T1e; T3n = T1Z + T20; T21 = T1Z - T20; T2C = T19 - T1e; } { E T1H, T2I, T1M, T2J; { E T1E, T1G, T1D, T1F; T1E = Ip[WS(rs, 8)]; T1G = Im[WS(rs, 8)]; T1D = W[32]; T1F = W[33]; T1H = FMA(T1D, T1E, T1F * T1G); T2I = FNMS(T1F, T1E, T1D * T1G); } { E T1J, T1L, T1I, T1K; T1J = Ip[WS(rs, 3)]; T1L = Im[WS(rs, 3)]; T1I = W[12]; T1K = W[13]; T1M = FMA(T1I, T1J, T1K * T1L); T2J = FNMS(T1K, T1J, T1I * T1L); } T1N = T1H + T1M; T3r = T2I + T2J; T27 = T1H - T1M; T2K = T2I - T2J; } { E T1k, T2D, T1p, T2E; { E T1h, T1j, T1g, T1i; T1h = Ip[WS(rs, 6)]; T1j = Im[WS(rs, 6)]; T1g = W[24]; T1i = W[25]; T1k = FMA(T1g, T1h, T1i * T1j); T2D = FNMS(T1i, T1h, T1g * T1j); } { E T1m, T1o, T1l, T1n; T1m = Ip[WS(rs, 1)]; T1o = Im[WS(rs, 1)]; T1l = W[4]; T1n = W[5]; T1p = FMA(T1l, T1m, T1n * T1o); T2E = FNMS(T1n, T1m, T1l * T1o); } T1q = T1k + T1p; T3o = T2D + T2E; T22 = T1k - T1p; T2F = T2D - T2E; } { E T1w, T24, T1B, T25; { E T1t, T1v, T1s, T1u; T1t = Rp[WS(rs, 6)]; T1v = Rm[WS(rs, 6)]; T1s = W[22]; T1u = W[23]; T1w = FMA(T1s, T1t, T1u * T1v); T24 = FNMS(T1u, T1t, T1s * T1v); } { E T1y, T1A, T1x, T1z; T1y = Rp[WS(rs, 1)]; T1A = Rm[WS(rs, 1)]; T1x = W[2]; T1z = W[3]; T1B = FMA(T1x, T1y, T1z * T1A); T25 = FNMS(T1z, T1y, T1x * T1A); } T1C = T1w + T1B; T3q = T24 + T25; T26 = T24 - T25; T2H = T1w - T1B; } T1r = T1f - T1q; T1O = T1C - T1N; T1P = T1r + T1O; T3p = T3n + T3o; T3s = T3q + T3r; T3K = T3p + T3s; T3A = T3n - T3o; T3B = T3r - T3q; T3Z = T3B - T3A; T1V = T1f + T1q; T1W = T1C + T1N; T1X = T1V + T1W; T23 = T21 + T22; T28 = T26 + T27; T4q = T23 + T28; T2W = T21 - T22; T2X = T26 - T27; T4f = T2W + T2X; T33 = T2C + T2F; T34 = T2H + T2K; T35 = T33 + T34; T2G = T2C - T2F; T2L = T2H - T2K; T2M = T2G + T2L; } { E Tu, T3g, T2c, T2r, T12, T3k, T2f, T2z, TF, T3h, T2d, T2u, TR, T3j, T2i; E T2w; { E To, T2a, Tt, T2b; { E Tl, Tn, Tk, Tm; Tl = Rp[WS(rs, 2)]; Tn = Rm[WS(rs, 2)]; Tk = W[6]; Tm = W[7]; To = FMA(Tk, Tl, Tm * Tn); T2a = FNMS(Tm, Tl, Tk * Tn); } { E Tq, Ts, Tp, Tr; Tq = Rp[WS(rs, 7)]; Ts = Rm[WS(rs, 7)]; Tp = W[26]; Tr = W[27]; Tt = FMA(Tp, Tq, Tr * Ts); T2b = FNMS(Tr, Tq, Tp * Ts); } Tu = To + Tt; T3g = T2a + T2b; T2c = T2a - T2b; T2r = To - Tt; } { E TW, T2x, T11, T2y; { E TT, TV, TS, TU; TT = Ip[0]; TV = Im[0]; TS = W[0]; TU = W[1]; TW = FMA(TS, TT, TU * TV); T2x = FNMS(TU, TT, TS * TV); } { E TY, T10, TX, TZ; TY = Ip[WS(rs, 5)]; T10 = Im[WS(rs, 5)]; TX = W[20]; TZ = W[21]; T11 = FMA(TX, TY, TZ * T10); T2y = FNMS(TZ, TY, TX * T10); } T12 = TW + T11; T3k = T2x + T2y; T2f = T11 - TW; T2z = T2x - T2y; } { E Tz, T2s, TE, T2t; { E Tw, Ty, Tv, Tx; Tw = Ip[WS(rs, 4)]; Ty = Im[WS(rs, 4)]; Tv = W[16]; Tx = W[17]; Tz = FMA(Tv, Tw, Tx * Ty); T2s = FNMS(Tx, Tw, Tv * Ty); } { E TB, TD, TA, TC; TB = Ip[WS(rs, 9)]; TD = Im[WS(rs, 9)]; TA = W[36]; TC = W[37]; TE = FMA(TA, TB, TC * TD); T2t = FNMS(TC, TB, TA * TD); } TF = Tz + TE; T3h = T2s + T2t; T2d = Tz - TE; T2u = T2s - T2t; } { E TL, T2g, TQ, T2h; { E TI, TK, TH, TJ; TI = Rp[WS(rs, 8)]; TK = Rm[WS(rs, 8)]; TH = W[30]; TJ = W[31]; TL = FMA(TH, TI, TJ * TK); T2g = FNMS(TJ, TI, TH * TK); } { E TN, TP, TM, TO; TN = Rp[WS(rs, 3)]; TP = Rm[WS(rs, 3)]; TM = W[10]; TO = W[11]; TQ = FMA(TM, TN, TO * TP); T2h = FNMS(TO, TN, TM * TP); } TR = TL + TQ; T3j = T2g + T2h; T2i = T2g - T2h; T2w = TL - TQ; } TG = Tu - TF; T13 = TR - T12; T14 = TG + T13; T3i = T3g + T3h; T3l = T3j + T3k; T3J = T3i + T3l; T3D = T3g - T3h; T3E = T3j - T3k; T40 = T3D + T3E; T1S = Tu + TF; T1T = TR + T12; T1U = T1S + T1T; T2e = T2c + T2d; T2j = T2f - T2i; T4p = T2j - T2e; T2T = T2c - T2d; T2U = T2i + T2f; T4e = T2T + T2U; T30 = T2r + T2u; T31 = T2w + T2z; T32 = T30 + T31; T2v = T2r - T2u; T2A = T2w - T2z; T2B = T2v + T2A; } { E T3y, T1Q, T3x, T3G, T3I, T3C, T3F, T3H, T3z; T3y = KP559016994 * (T14 - T1P); T1Q = T14 + T1P; T3x = FNMS(KP250000000, T1Q, Tj); T3C = T3A + T3B; T3F = T3D - T3E; T3G = FNMS(KP587785252, T3F, KP951056516 * T3C); T3I = FMA(KP951056516, T3F, KP587785252 * T3C); Rm[WS(rs, 9)] = Tj + T1Q; T3H = T3y + T3x; Rm[WS(rs, 5)] = T3H - T3I; Rp[WS(rs, 6)] = T3H + T3I; T3z = T3x - T3y; Rp[WS(rs, 2)] = T3z - T3G; Rm[WS(rs, 1)] = T3z + T3G; } { E T47, T41, T46, T45, T49, T43, T44, T4a, T48; T47 = KP559016994 * (T40 + T3Z); T41 = T3Z - T40; T46 = FMA(KP250000000, T41, T42); T43 = T13 - TG; T44 = T1r - T1O; T45 = FMA(KP587785252, T43, KP951056516 * T44); T49 = FNMS(KP587785252, T44, KP951056516 * T43); Im[WS(rs, 9)] = T41 - T42; T4a = T47 + T46; Im[WS(rs, 5)] = T49 - T4a; Ip[WS(rs, 6)] = T49 + T4a; T48 = T46 - T47; Im[WS(rs, 1)] = T45 - T48; Ip[WS(rs, 2)] = T45 + T48; } { E T3d, T1Y, T3e, T3u, T3w, T3m, T3t, T3v, T3f; T3d = KP559016994 * (T1U - T1X); T1Y = T1U + T1X; T3e = FNMS(KP250000000, T1Y, T1R); T3m = T3i - T3l; T3t = T3p - T3s; T3u = FMA(KP951056516, T3m, KP587785252 * T3t); T3w = FNMS(KP587785252, T3m, KP951056516 * T3t); Rp[0] = T1R + T1Y; T3v = T3e - T3d; Rm[WS(rs, 7)] = T3v - T3w; Rp[WS(rs, 8)] = T3v + T3w; T3f = T3d + T3e; Rp[WS(rs, 4)] = T3f - T3u; Rm[WS(rs, 3)] = T3f + T3u; } { E T3U, T3L, T3V, T3T, T3X, T3R, T3S, T3Y, T3W; T3U = KP559016994 * (T3J - T3K); T3L = T3J + T3K; T3V = FNMS(KP250000000, T3L, T3Q); T3R = T1S - T1T; T3S = T1V - T1W; T3T = FMA(KP951056516, T3R, KP587785252 * T3S); T3X = FNMS(KP951056516, T3S, KP587785252 * T3R); Ip[0] = T3L + T3Q; T3Y = T3V - T3U; Im[WS(rs, 7)] = T3X - T3Y; Ip[WS(rs, 8)] = T3X + T3Y; T3W = T3U + T3V; Im[WS(rs, 3)] = T3T - T3W; Ip[WS(rs, 4)] = T3T + T3W; } { E T2P, T2N, T2O, T2l, T2R, T29, T2k, T2S, T2Q; T2P = KP559016994 * (T2B - T2M); T2N = T2B + T2M; T2O = FNMS(KP250000000, T2N, T2q); T29 = T23 - T28; T2k = T2e + T2j; T2l = FNMS(KP587785252, T2k, KP951056516 * T29); T2R = FMA(KP951056516, T2k, KP587785252 * T29); Rm[WS(rs, 4)] = T2q + T2N; T2S = T2P + T2O; Rm[WS(rs, 8)] = T2R + T2S; Rm[0] = T2S - T2R; T2Q = T2O - T2P; Rp[WS(rs, 3)] = T2l + T2Q; Rp[WS(rs, 7)] = T2Q - T2l; } { E T4w, T4r, T4x, T4v, T4A, T4t, T4u, T4z, T4y; T4w = KP559016994 * (T4p + T4q); T4r = T4p - T4q; T4x = FMA(KP250000000, T4r, T4s); T4t = T2v - T2A; T4u = T2G - T2L; T4v = FMA(KP951056516, T4t, KP587785252 * T4u); T4A = FNMS(KP587785252, T4t, KP951056516 * T4u); Im[WS(rs, 4)] = T4r - T4s; T4z = T4w + T4x; Ip[WS(rs, 3)] = T4z - T4A; Ip[WS(rs, 7)] = T4A + T4z; T4y = T4w - T4x; Im[WS(rs, 8)] = T4v + T4y; Im[0] = T4y - T4v; } { E T36, T38, T39, T2Z, T3b, T2V, T2Y, T3c, T3a; T36 = KP559016994 * (T32 - T35); T38 = T32 + T35; T39 = FNMS(KP250000000, T38, T37); T2V = T2T - T2U; T2Y = T2W - T2X; T2Z = FMA(KP951056516, T2V, KP587785252 * T2Y); T3b = FNMS(KP587785252, T2V, KP951056516 * T2Y); Rp[WS(rs, 5)] = T37 + T38; T3c = T39 - T36; Rm[WS(rs, 6)] = T3b + T3c; Rm[WS(rs, 2)] = T3c - T3b; T3a = T36 + T39; Rp[WS(rs, 1)] = T2Z + T3a; Rp[WS(rs, 9)] = T3a - T2Z; } { E T4g, T4k, T4l, T4d, T4o, T4b, T4c, T4n, T4m; T4g = KP559016994 * (T4e - T4f); T4k = T4e + T4f; T4l = FNMS(KP250000000, T4k, T4j); T4b = T33 - T34; T4c = T30 - T31; T4d = FNMS(KP587785252, T4c, KP951056516 * T4b); T4o = FMA(KP951056516, T4c, KP587785252 * T4b); Ip[WS(rs, 5)] = T4k + T4j; T4n = T4g + T4l; Ip[WS(rs, 1)] = T4n - T4o; Ip[WS(rs, 9)] = T4o + T4n; T4m = T4g - T4l; Im[WS(rs, 6)] = T4d + T4m; Im[WS(rs, 2)] = T4m - T4d; } }}static const tw_instr twinstr[] = { {TW_FULL, 1, 20}, {TW_NEXT, 1, 0}};static const hc2c_desc desc = { 20, "hc2cf_20", twinstr, &GENUS, {184, 62, 62, 0} };void X(codelet_hc2cf_20) (planner *p) { X(khc2c_register) (p, hc2cf_20, &desc, HC2C_VIA_RDFT);}#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 + -