📄 hc2cfdft2_32.c
字号:
Rm[WS(rs, 8)] = KP500000000 * (FNMS(KP980785280, Taw, Tav)); } } } } } } Im[WS(rs, 8)] = -(KP500000000 * (FNMS(KP980785280, TaA, Tax))); Ip[WS(rs, 7)] = KP500000000 * (FMA(KP980785280, TaA, Tax)); }}static const tw_instr twinstr[] = { {TW_CEXP, 1, 1}, {TW_CEXP, 1, 3}, {TW_CEXP, 1, 9}, {TW_CEXP, 1, 27}, {TW_NEXT, 1, 0}};static const hc2c_desc desc = { 32, "hc2cfdft2_32", twinstr, &GENUS, {300, 162, 252, 0} };void X(codelet_hc2cfdft2_32) (planner *p) { X(khc2c_register) (p, hc2cfdft2_32, &desc, HC2C_VIA_DFT);}#else /* HAVE_FMA *//* Generated by: ../../../genfft/gen_hc2cdft -compact -variables 4 -pipeline-latency 4 -twiddle-log3 -precompute-twiddles -n 32 -dit -name hc2cfdft2_32 -include hc2cf.h *//* * This function contains 552 FP additions, 300 FP multiplications, * (or, 440 additions, 188 multiplications, 112 fused multiply/add), * 166 stack variables, 9 constants, and 128 memory accesses */#include "hc2cf.h"static void hc2cfdft2_32(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms){ DK(KP277785116, +0.277785116509801112371415406974266437187468595); DK(KP415734806, +0.415734806151272618539394188808952878369280406); DK(KP097545161, +0.097545161008064133924142434238511120463845809); DK(KP490392640, +0.490392640201615224563091118067119518486966865); DK(KP707106781, +0.707106781186547524400844362104849039284835938); DK(KP191341716, +0.191341716182544885864229992015199433380672281); DK(KP461939766, +0.461939766255643378064091594698394143411208313); DK(KP353553390, +0.353553390593273762200422181052424519642417969); DK(KP500000000, +0.500000000000000000000000000000000000000000000); INT m; for (m = mb, W = W + ((mb - 1) * 8); m < me; m = m + 1, Rp = Rp + ms, Ip = Ip + ms, Rm = Rm - ms, Im = Im - ms, W = W + 8, MAKE_VOLATILE_STRIDE(rs)) { E T1, T4, T2, T5, T7, T1b, T1d, Td, Ti, Tk, Tj, Tl, TL, TR, T2h; E T2O, T16, T2l, T10, T2K, Tm, Tq, T3s, T3K, T3w, T3M, T4e, T4u, T4i, T4w; E Ty, TE, T3h, T3j, T2q, T2u, T4l, T4n, T1v, T1B, T3E, T3G, T2B, T2F, T3Y; E T40, T1f, T1G, T1i, T1H, T1j, T1M, T1n, T1I, T23, T2U, T26, T2V, T27, T30; E T2b, T2W; { E Tw, T1A, TD, T1t, Tx, T1z, TC, T1u, TJ, T15, TQ, TY, TK, T14, TP; E TZ; { E T3, Tc, T6, Tb; T1 = W[0]; T4 = W[1]; T2 = W[2]; T5 = W[3]; T3 = T1 * T2; Tc = T4 * T2; T6 = T4 * T5; Tb = T1 * T5; T7 = T3 + T6; T1b = T3 - T6; T1d = Tb + Tc; Td = Tb - Tc; Ti = W[4]; Tw = T1 * Ti; T1A = T5 * Ti; TD = T4 * Ti; T1t = T2 * Ti; Tk = W[5]; Tx = T4 * Tk; T1z = T2 * Tk; TC = T1 * Tk; T1u = T5 * Tk; Tj = W[6]; TJ = T1 * Tj; T15 = T5 * Tj; TQ = T4 * Tj; TY = T2 * Tj; Tl = W[7]; TK = T4 * Tl; T14 = T2 * Tl; TP = T1 * Tl; TZ = T5 * Tl; } TL = TJ + TK; TR = TP - TQ; T2h = TJ - TK; T2O = T14 - T15; T16 = T14 + T15; T2l = TP + TQ; T10 = TY - TZ; T2K = TY + TZ; Tm = FMA(Ti, Tj, Tk * Tl); Tq = FNMS(Tk, Tj, Ti * Tl); { E T3q, T3r, T3u, T3v; T3q = T7 * Tj; T3r = Td * Tl; T3s = T3q + T3r; T3K = T3q - T3r; T3u = T7 * Tl; T3v = Td * Tj; T3w = T3u - T3v; T3M = T3u + T3v; } { E T4c, T4d, T4g, T4h; T4c = T1b * Tj; T4d = T1d * Tl; T4e = T4c - T4d; T4u = T4c + T4d; T4g = T1b * Tl; T4h = T1d * Tj; T4i = T4g + T4h; T4w = T4g - T4h; Ty = Tw - Tx; TE = TC + TD; T3h = FMA(Ty, Tj, TE * Tl); T3j = FNMS(TE, Tj, Ty * Tl); } T2q = T1t - T1u; T2u = T1z + T1A; T4l = FMA(T2q, Tj, T2u * Tl); T4n = FNMS(T2u, Tj, T2q * Tl); T1v = T1t + T1u; T1B = T1z - T1A; T3E = FMA(T1v, Tj, T1B * Tl); T3G = FNMS(T1B, Tj, T1v * Tl); T2B = Tw + Tx; T2F = TC - TD; T3Y = FMA(T2B, Tj, T2F * Tl); T40 = FNMS(T2F, Tj, T2B * Tl); { E T1c, T1e, T1g, T1h; T1c = T1b * Ti; T1e = T1d * Tk; T1f = T1c - T1e; T1G = T1c + T1e; T1g = T1b * Tk; T1h = T1d * Ti; T1i = T1g + T1h; T1H = T1g - T1h; } T1j = FMA(T1f, Tj, T1i * Tl); T1M = FNMS(T1H, Tj, T1G * Tl); T1n = FNMS(T1i, Tj, T1f * Tl); T1I = FMA(T1G, Tj, T1H * Tl); { E T21, T22, T24, T25; T21 = T7 * Ti; T22 = Td * Tk; T23 = T21 + T22; T2U = T21 - T22; T24 = T7 * Tk; T25 = Td * Ti; T26 = T24 - T25; T2V = T24 + T25; } T27 = FMA(T23, Tj, T26 * Tl); T30 = FNMS(T2V, Tj, T2U * Tl); T2b = FNMS(T26, Tj, T23 * Tl); T2W = FMA(T2U, Tj, T2V * Tl); } { E T38, T7l, T7S, T8Y, T7Z, T91, T3A, T6k, T4F, T83, T5C, T6n, T2T, T84, T4I; E T7m, T2g, T4M, T4P, T2z, T3T, T6m, T7O, T7V, T7j, T87, T5v, T6j, T7L, T7U; E T7g, T86, Tv, TW, T61, T4U, T4X, T62, T4b, T6c, T7v, T7C, T5g, T6f, T74; E T8G, T7s, T7B, T71, T8F, T1s, T1R, T65, T51, T54, T64, T4A, T6g, T7G, T8U; E T5n, T6d, T7b, T8J, T7z, T8R, T78, T8I; { E T2E, T2I, T3p, T5w, T37, T4D, T3g, T5A, T2N, T2R, T3y, T5x, T2Z, T33, T3l; E T5z; { E T2C, T2D, T3o, T2G, T2H, T3n; T2C = Ip[WS(rs, 4)]; T2D = Im[WS(rs, 4)]; T3o = T2C + T2D; T2G = Rp[WS(rs, 4)]; T2H = Rm[WS(rs, 4)]; T3n = T2G - T2H; T2E = T2C - T2D; T2I = T2G + T2H; T3p = FMA(Ti, T3n, Tk * T3o); T5w = FNMS(Tk, T3n, Ti * T3o); } { E T35, T36, T3f, T3c, T3d, T3e; T35 = Ip[0]; T36 = Im[0]; T3f = T35 + T36; T3c = Rm[0]; T3d = Rp[0]; T3e = T3c - T3d; T37 = T35 - T36; T4D = T3d + T3c; T3g = FNMS(T4, T3f, T1 * T3e); T5A = FMA(T4, T3e, T1 * T3f); } { E T2L, T2M, T3x, T2P, T2Q, T3t; T2L = Ip[WS(rs, 12)]; T2M = Im[WS(rs, 12)]; T3x = T2L + T2M; T2P = Rp[WS(rs, 12)]; T2Q = Rm[WS(rs, 12)]; T3t = T2P - T2Q; T2N = T2L - T2M; T2R = T2P + T2Q; T3y = FMA(T3s, T3t, T3w * T3x); T5x = FNMS(T3w, T3t, T3s * T3x); } { E T2X, T2Y, T3k, T31, T32, T3i; T2X = Ip[WS(rs, 8)]; T2Y = Im[WS(rs, 8)]; T3k = T2X + T2Y; T31 = Rp[WS(rs, 8)]; T32 = Rm[WS(rs, 8)]; T3i = T31 - T32; T2Z = T2X - T2Y; T33 = T31 + T32; T3l = FMA(T3h, T3i, T3j * T3k); T5z = FNMS(T3j, T3i, T3h * T3k); } { E T34, T7Q, T7R, T4E, T5y, T5B; T34 = FNMS(T30, T33, T2W * T2Z); T38 = T34 + T37; T7l = T37 - T34; T7Q = T3l + T3g; T7R = T5w - T5x; T7S = T7Q - T7R; T8Y = T7R + T7Q; { E T7X, T7Y, T3m, T3z; T7X = T3y - T3p; T7Y = T5A - T5z; T7Z = T7X + T7Y; T91 = T7Y - T7X; T3m = T3g - T3l; T3z = T3p + T3y; T3A = T3m - T3z; T6k = T3z + T3m; } T4E = FMA(T2W, T33, T30 * T2Z); T4F = T4D + T4E; T83 = T4D - T4E; T5y = T5w + T5x; T5B = T5z + T5A; T5C = T5y + T5B; T6n = T5B - T5y; { E T2J, T2S, T4G, T4H; T2J = FNMS(T2F, T2I, T2B * T2E); T2S = FNMS(T2O, T2R, T2K * T2N); T2T = T2J + T2S; T84 = T2J - T2S; T4G = FMA(T2B, T2I, T2F * T2E); T4H = FMA(T2K, T2R, T2O * T2N); T4I = T4G + T4H; T7m = T4G - T4H; } } } { E T20, T5p, T3D, T4K, T2y, T5t, T3R, T4O, T2f, T5q, T3I, T4L, T2p, T5s, T3O; E T4N; { E T1W, T3C, T1Z, T3B; { E T1U, T1V, T1X, T1Y; T1U = Ip[WS(rs, 2)]; T1V = Im[WS(rs, 2)]; T1W = T1U - T1V; T3C = T1U + T1V; T1X = Rp[WS(rs, 2)]; T1Y = Rm[WS(rs, 2)]; T1Z = T1X + T1Y; T3B = T1X - T1Y; } T20 = FNMS(T1d, T1Z, T1b * T1W); T5p = FNMS(T1H, T3B, T1G * T3C); T3D = FMA(T1G, T3B, T1H * T3C); T4K = FMA(T1b, T1Z, T1d * T1W); } { E T2t, T3Q, T2x, T3P; { E T2r, T2s, T2v, T2w; T2r = Ip[WS(rs, 6)]; T2s = Im[WS(rs, 6)]; T2t = T2r - T2s; T3Q = T2r + T2s; T2v = Rp[WS(rs, 6)]; T2w = Rm[WS(rs, 6)]; T2x = T2v + T2w; T3P = T2v - T2w; } T2y = FNMS(T2u, T2x, T2q * T2t); T5t = FNMS(T1i, T3P, T1f * T3Q); T3R = FMA(T1f, T3P, T1i * T3Q); T4O = FMA(T2q, T2x, T2u * T2t); } { E T2a, T3H, T2e, T3F; { E T28, T29, T2c, T2d; T28 = Ip[WS(rs, 10)]; T29 = Im[WS(rs, 10)]; T2a = T28 - T29; T3H = T28 + T29; T2c = Rp[WS(rs, 10)]; T2d = Rm[WS(rs, 10)]; T2e = T2c + T2d; T3F = T2c - T2d; } T2f = FNMS(T2b, T2e, T27 * T2a); T5q = FNMS(T3G, T3F, T3E * T3H); T3I = FMA(T3E, T3F, T3G * T3H); T4L = FMA(T27, T2e, T2b * T2a); } { E T2k, T3N, T2o, T3L; { E T2i, T2j, T2m, T2n; T2i = Ip[WS(rs, 14)]; T2j = Im[WS(rs, 14)]; T2k = T2i - T2j; T3N = T2i + T2j; T2m = Rp[WS(rs, 14)]; T2n = Rm[WS(rs, 14)]; T2o = T2m + T2n; T3L = T2m - T2n; } T2p = FNMS(T2l, T2o, T2h * T2k); T5s = FNMS(T3M, T3L, T3K * T3N); T3O = FMA(T3K, T3L, T3M * T3N); T4N = FMA(T2h, T2o, T2l * T2k); } { E T3J, T3S, T5r, T5u; T2g = T20 + T2f; T4M = T4K + T4L; T4P = T4N + T4O; T2z = T2p + T2y; T3J = T3D + T3I; T3S = T3O + T3R; T3T = T3J + T3S; T6m = T3S - T3J; { E T7M, T7N, T7h, T7i; T7M = T5s - T5t; T7N = T3R - T3O; T7O = T7M + T7N; T7V = T7M - T7N; T7h = T4N - T4O; T7i = T2p - T2y; T7j = T7h + T7i; T87 = T7h - T7i; } T5r = T5p + T5q; T5u = T5s + T5t; T5v = T5r + T5u; T6j = T5u - T5r; { E T7J, T7K, T7e, T7f; T7J = T3I - T3D; T7K = T5p - T5q; T7L = T7J - T7K; T7U = T7K + T7J; T7e = T20 - T2f; T7f = T4K - T4L; T7g = T7e - T7f; T86 = T7f + T7e; } } } { E Th, T5a, T3X, T4S, TV, T5e, T49, T4W, Tu, T5b, T42, T4T, TI, T5d, T46; E T4V; { E Ta, T3W, Tg, T3V; { E T8, T9, Te, Tf; T8 = Ip[WS(rs, 1)]; T9 = Im[WS(rs, 1)]; Ta = T8 - T9; T3W = T8 + T9; Te = Rp[WS(rs, 1)]; Tf = Rm[WS(rs, 1)]; Tg = Te + Tf; T3V = Te - Tf; } Th = FNMS(Td, Tg, T7 * Ta); T5a = FNMS(T5, T3V, T2 * T3W); T3X = FMA(T2, T3V, T5 * T3W); T4S = FMA(T7, Tg, Td * Ta); } { E TO, T48, TU, T47; { E TM, TN, TS, TT; TM = Ip[WS(rs, 13)]; TN = Im[WS(rs, 13)]; TO = TM - TN; T48 = TM + TN; TS = Rp[WS(rs, 13)]; TT = Rm[WS(rs, 13)]; TU = TS + TT; T47 = TS - TT; } TV = FNMS(TR, TU, TL * TO); T5e = FNMS(Tl, T47, Tj * T48); T49 = FMA(Tj, T47, Tl * T48); T4W = FMA(TL, TU, TR * TO); } { E Tp, T41, Tt, T3Z; { E Tn, To, Tr, Ts; Tn = Ip[WS(rs, 9)]; To = Im[WS(rs, 9)]; Tp = Tn - To; T41 = Tn + To; Tr = Rp[WS(rs, 9)]; Ts = Rm[WS(rs, 9)]; Tt = Tr + Ts; T3Z = Tr - Ts; } Tu = FNMS(Tq, Tt, Tm * Tp); T5b = FNMS(T40, T3Z, T3Y * T41); T42 = FMA(T3Y, T3Z, T40 * T41); T4T = FMA(Tm, Tt, Tq * Tp); } { E TB, T45, TH, T44; { E Tz, TA, TF, TG; Tz = Ip[WS(rs, 5)]; TA = Im[WS(rs, 5)]; TB = Tz - TA; T45 = Tz + TA; TF = Rp[WS(rs, 5)]; TG = Rm[WS(rs, 5)]; TH = TF + TG; T44 = TF - TG; } TI = FNMS(TE, TH, Ty * TB); T5d = FNMS(T2V, T44, T2U * T45); T46 = FMA(T2U, T44, T2V * T45); T4V = FMA(Ty, TH, TE * TB); } Tv = Th + Tu; TW = TI + TV; T61 = Tv - TW; T4U = T4S + T4T; T4X = T4V + T4W; T62 = T4U - T4X; { E T43, T4a, T7t, T7u; T43 = T3X + T42; T4a = T46 + T49; T4b = T43 + T4a; T6c = T4a - T43; T7t = T5e - T5d; T7u = T46 - T49; T7v = T7t + T7u; T7C = T7t - T7u; } { E T5c, T5f, T72, T73; T5c = T5a + T5b; T5f = T5d + T5e; T5g = T5c + T5f; T6f = T5f - T5c; T72 = T4S - T4T; T73 = TI - TV; T74 = T72 + T73; T8G = T72 - T73; } { E T7q, T7r, T6Z, T70; T7q = T42 - T3X; T7r = T5a - T5b; T7s = T7q - T7r; T7B = T7r + T7q; T6Z = Th - Tu; T70 = T4V - T4W; T71 = T6Z - T70; T8F = T6Z + T70; } } { E T1a, T5h, T4k, T4Z, T1Q, T5l, T4y, T53, T1r, T5i, T4p, T50, T1F, T5k, T4t; E T52;
⌨️ 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 + -