📄 hc2cb_20.c
字号:
T1Q = W[11]; T1K = T1I * T1G; T1H = T1F * T1G; T1L = W[10]; T1U = T1Q * T1O; Rm[WS(rs, 9)] = FMA(T1F, T1J, T1K); Rp[WS(rs, 9)] = FNMS(T1I, T1J, T1H); T1P = T1L * T1O; Rm[WS(rs, 3)] = FMA(T1L, T1T, T1U); } } } } } } } Rp[WS(rs, 3)] = FNMS(T1Q, T1T, T1P); }}static const tw_instr twinstr[] = { {TW_FULL, 1, 20}, {TW_NEXT, 1, 0}};static const hc2c_desc desc = { 20, "hc2cb_20", twinstr, &GENUS, {136, 38, 110, 0} };void X(codelet_hc2cb_20) (planner *p) { X(khc2c_register) (p, hc2cb_20, &desc, HC2C_VIA_RDFT);}#else /* HAVE_FMA *//* Generated by: ../../../genfft/gen_hc2c -compact -variables 4 -pipeline-latency 4 -sign 1 -n 20 -dif -name hc2cb_20 -include hc2cb.h *//* * This function contains 246 FP additions, 124 FP multiplications, * (or, 184 additions, 62 multiplications, 62 fused multiply/add), * 97 stack variables, 4 constants, and 80 memory accesses */#include "hc2cb.h"static void hc2cb_20(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms){ DK(KP250000000, +0.250000000000000000000000000000000000000000000); DK(KP559016994, +0.559016994374947424102293417182819058860154590); DK(KP587785252, +0.587785252292473129168705954639072768597652438); DK(KP951056516, +0.951056516295153572116439333379382143405698634); 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 T7, T3T, T49, TE, T1v, T2T, T3g, T2d, T13, T3n, T3o, T1i, T26, T4e, T4d; E T23, T1n, T42, T3Z, T1m, T2h, T2I, T2i, T2P, T30, T37, T38, Tm, TB, TC; E T46, T47, T4a, T2a, T2b, T2e, T1w, T1x, T1y, T3O, T3R, T3U, T3h, T3i, T3j; E TH, TK, TL; { E T3, T2R, T1r, T3e, T6, T3f, T1u, T2S; { E T1, T2, T1p, T1q; T1 = Rp[0]; T2 = Rm[WS(rs, 9)]; T3 = T1 + T2; T2R = T1 - T2; T1p = Ip[0]; T1q = Im[WS(rs, 9)]; T1r = T1p - T1q; T3e = T1p + T1q; } { E T4, T5, T1s, T1t; T4 = Rp[WS(rs, 5)]; T5 = Rm[WS(rs, 4)]; T6 = T4 + T5; T3f = T4 - T5; T1s = Ip[WS(rs, 5)]; T1t = Im[WS(rs, 4)]; T1u = T1s - T1t; T2S = T1s + T1t; } T7 = T3 + T6; T3T = T2R - T2S; T49 = T3f + T3e; TE = T3 - T6; T1v = T1r - T1u; T2T = T2R + T2S; T3g = T3e - T3f; T2d = T1r + T1u; } { E Te, T3M, T3X, TF, TV, T2E, T2W, T21, TA, T3Q, T41, TJ, T1h, T2O, T36; E T25, Tl, T3N, T3Y, TG, T12, T2H, T2Z, T22, Tt, T3P, T40, TI, T1a, T2L; E T33, T24; { E Ta, T2U, TR, T2C, Td, T2D, TU, T2V; { E T8, T9, TP, TQ; T8 = Rp[WS(rs, 4)]; T9 = Rm[WS(rs, 5)]; Ta = T8 + T9; T2U = T8 - T9; TP = Ip[WS(rs, 4)]; TQ = Im[WS(rs, 5)]; TR = TP - TQ; T2C = TP + TQ; } { E Tb, Tc, TS, TT; Tb = Rp[WS(rs, 9)]; Tc = Rm[0]; Td = Tb + Tc; T2D = Tb - Tc; TS = Ip[WS(rs, 9)]; TT = Im[0]; TU = TS - TT; T2V = TS + TT; } Te = Ta + Td; T3M = T2U - T2V; T3X = T2D + T2C; TF = Ta - Td; TV = TR - TU; T2E = T2C - T2D; T2W = T2U + T2V; T21 = TR + TU; } { E Tw, T34, T1d, T2N, Tz, T2M, T1g, T35; { E Tu, Tv, T1b, T1c; Tu = Rm[WS(rs, 7)]; Tv = Rp[WS(rs, 2)]; Tw = Tu + Tv; T34 = Tu - Tv; T1b = Ip[WS(rs, 2)]; T1c = Im[WS(rs, 7)]; T1d = T1b - T1c; T2N = T1b + T1c; } { E Tx, Ty, T1e, T1f; Tx = Rm[WS(rs, 2)]; Ty = Rp[WS(rs, 7)]; Tz = Tx + Ty; T2M = Tx - Ty; T1e = Ip[WS(rs, 7)]; T1f = Im[WS(rs, 2)]; T1g = T1e - T1f; T35 = T1e + T1f; } TA = Tw + Tz; T3Q = T34 + T35; T41 = T2M - T2N; TJ = Tw - Tz; T1h = T1d - T1g; T2O = T2M + T2N; T36 = T34 - T35; T25 = T1d + T1g; } { E Th, T2X, TY, T2G, Tk, T2F, T11, T2Y; { E Tf, Tg, TW, TX; Tf = Rm[WS(rs, 3)]; Tg = Rp[WS(rs, 6)]; Th = Tf + Tg; T2X = Tf - Tg; TW = Ip[WS(rs, 6)]; TX = Im[WS(rs, 3)]; TY = TW - TX; T2G = TW + TX; } { E Ti, Tj, TZ, T10; Ti = Rp[WS(rs, 1)]; Tj = Rm[WS(rs, 8)]; Tk = Ti + Tj; T2F = Ti - Tj; TZ = Ip[WS(rs, 1)]; T10 = Im[WS(rs, 8)]; T11 = TZ - T10; T2Y = TZ + T10; } Tl = Th + Tk; T3N = T2X - T2Y; T3Y = T2F - T2G; TG = Th - Tk; T12 = TY - T11; T2H = T2F + T2G; T2Z = T2X + T2Y; T22 = TY + T11; } { E Tp, T31, T16, T2J, Ts, T2K, T19, T32; { E Tn, To, T14, T15; Tn = Rp[WS(rs, 8)]; To = Rm[WS(rs, 1)]; Tp = Tn + To; T31 = Tn - To; T14 = Ip[WS(rs, 8)]; T15 = Im[WS(rs, 1)]; T16 = T14 - T15; T2J = T14 + T15; } { E Tq, Tr, T17, T18; Tq = Rm[WS(rs, 6)]; Tr = Rp[WS(rs, 3)]; Ts = Tq + Tr; T2K = Tq - Tr; T17 = Ip[WS(rs, 3)]; T18 = Im[WS(rs, 6)]; T19 = T17 - T18; T32 = T17 + T18; } Tt = Tp + Ts; T3P = T31 + T32; T40 = T2K + T2J; TI = Tp - Ts; T1a = T16 - T19; T2L = T2J - T2K; T33 = T31 - T32; T24 = T16 + T19; } T13 = TV - T12; T3n = T2W - T2Z; T3o = T33 - T36; T1i = T1a - T1h; T26 = T24 - T25; T4e = T3P - T3Q; T4d = T3M - T3N; T23 = T21 - T22; T1n = TI - TJ; T42 = T40 - T41; T3Z = T3X - T3Y; T1m = TF - TG; T2h = Te - Tl; T2I = T2E + T2H; T2i = Tt - TA; T2P = T2L + T2O; T30 = T2W + T2Z; T37 = T33 + T36; T38 = T30 + T37; Tm = Te + Tl; TB = Tt + TA; TC = Tm + TB; T46 = T3X + T3Y; T47 = T40 + T41; T4a = T46 + T47; T2a = T21 + T22; T2b = T24 + T25; T2e = T2a + T2b; T1w = TV + T12; T1x = T1a + T1h; T1y = T1w + T1x; T3O = T3M + T3N; T3R = T3P + T3Q; T3U = T3O + T3R; T3h = T2E - T2H; T3i = T2L - T2O; T3j = T3h + T3i; TH = TF + TG; TK = TI + TJ; TL = TH + TK; } Rp[0] = T7 + TC; Rm[0] = T2d + T2e; { E T1U, T1W, T1T, T1V; T1U = TE + TL; T1W = T1v + T1y; T1T = W[18]; T1V = W[19]; Rp[WS(rs, 5)] = FNMS(T1V, T1W, T1T * T1U); Rm[WS(rs, 5)] = FMA(T1V, T1U, T1T * T1W); } { E T4y, T4A, T4x, T4z; T4y = T3T + T3U; T4A = T49 + T4a; T4x = W[8]; T4z = W[9]; Ip[WS(rs, 2)] = FNMS(T4z, T4A, T4x * T4y); Im[WS(rs, 2)] = FMA(T4x, T4A, T4z * T4y); } { E T3I, T3K, T3H, T3J; T3I = T2T + T38; T3K = T3g + T3j; T3H = W[28]; T3J = W[29]; Ip[WS(rs, 7)] = FNMS(T3J, T3K, T3H * T3I); Im[WS(rs, 7)] = FMA(T3H, T3K, T3J * T3I); } { E T27, T2j, T2v, T2r, T2g, T2u, T20, T2q; T27 = FMA(KP951056516, T23, KP587785252 * T26); T2j = FMA(KP951056516, T2h, KP587785252 * T2i); T2v = FNMS(KP951056516, T2i, KP587785252 * T2h); T2r = FNMS(KP951056516, T26, KP587785252 * T23); { E T2c, T2f, T1Y, T1Z; T2c = KP559016994 * (T2a - T2b); T2f = FNMS(KP250000000, T2e, T2d); T2g = T2c + T2f; T2u = T2f - T2c; T1Y = KP559016994 * (Tm - TB); T1Z = FNMS(KP250000000, TC, T7); T20 = T1Y + T1Z; T2q = T1Z - T1Y; } { E T28, T2k, T1X, T29; T28 = T20 + T27; T2k = T2g - T2j; T1X = W[6]; T29 = W[7]; Rp[WS(rs, 2)] = FNMS(T29, T2k, T1X * T28); Rm[WS(rs, 2)] = FMA(T29, T28, T1X * T2k); } { E T2y, T2A, T2x, T2z; T2y = T2q - T2r; T2A = T2v + T2u; T2x = W[22]; T2z = W[23]; Rp[WS(rs, 6)] = FNMS(T2z, T2A, T2x * T2y); Rm[WS(rs, 6)] = FMA(T2z, T2y, T2x * T2A); } { E T2m, T2o, T2l, T2n; T2m = T20 - T27; T2o = T2j + T2g; T2l = W[30]; T2n = W[31]; Rp[WS(rs, 8)] = FNMS(T2n, T2o, T2l * T2m); Rm[WS(rs, 8)] = FMA(T2n, T2m, T2l * T2o); } { E T2s, T2w, T2p, T2t; T2s = T2q + T2r; T2w = T2u - T2v; T2p = W[14]; T2t = W[15]; Rp[WS(rs, 4)] = FNMS(T2t, T2w, T2p * T2s); Rm[WS(rs, 4)] = FMA(T2t, T2s, T2p * T2w); } } { E T43, T4f, T4r, T4m, T4c, T4q, T3W, T4n; T43 = FMA(KP951056516, T3Z, KP587785252 * T42); T4f = FMA(KP951056516, T4d, KP587785252 * T4e); T4r = FNMS(KP951056516, T4e, KP587785252 * T4d); T4m = FNMS(KP951056516, T42, KP587785252 * T3Z); { E T48, T4b, T3S, T3V; T48 = KP559016994 * (T46 - T47); T4b = FNMS(KP250000000, T4a, T49); T4c = T48 + T4b; T4q = T4b - T48; T3S = KP559016994 * (T3O - T3R); T3V = FNMS(KP250000000, T3U, T3T); T3W = T3S + T3V; T4n = T3V - T3S; } { E T44, T4g, T3L, T45; T44 = T3W - T43; T4g = T4c + T4f; T3L = W[0]; T45 = W[1]; Ip[0] = FNMS(T45, T4g, T3L * T44); Im[0] = FMA(T3L, T4g, T45 * T44); } { E T4u, T4w, T4t, T4v; T4u = T4n - T4m; T4w = T4q + T4r; T4t = W[32]; T4v = W[33]; Ip[WS(rs, 8)] = FNMS(T4v, T4w, T4t * T4u); Im[WS(rs, 8)] = FMA(T4t, T4w, T4v * T4u); } { E T4i, T4k, T4h, T4j; T4i = T43 + T3W; T4k = T4c - T4f; T4h = W[16]; T4j = W[17]; Ip[WS(rs, 4)] = FNMS(T4j, T4k, T4h * T4i); Im[WS(rs, 4)] = FMA(T4h, T4k, T4j * T4i); } { E T4o, T4s, T4l, T4p; T4o = T4m + T4n; T4s = T4q - T4r; T4l = W[24]; T4p = W[25]; Ip[WS(rs, 6)] = FNMS(T4p, T4s, T4l * T4o); Im[WS(rs, 6)] = FMA(T4l, T4s, T4p * T4o); } } { E T1j, T1o, T1M, T1J, T1B, T1N, TO, T1I; T1j = FNMS(KP951056516, T1i, KP587785252 * T13); T1o = FNMS(KP951056516, T1n, KP587785252 * T1m); T1M = FMA(KP951056516, T1m, KP587785252 * T1n); T1J = FMA(KP951056516, T13, KP587785252 * T1i); { E T1z, T1A, TM, TN; T1z = FNMS(KP250000000, T1y, T1v); T1A = KP559016994 * (T1w - T1x); T1B = T1z - T1A; T1N = T1A + T1z; TM = FNMS(KP250000000, TL, TE); TN = KP559016994 * (TH - TK); TO = TM - TN; T1I = TN + TM; } { E T1k, T1C, TD, T1l; T1k = TO - T1j; T1C = T1o + T1B; TD = W[2]; T1l = W[3]; Rp[WS(rs, 1)] = FNMS(T1l, T1C, TD * T1k); Rm[WS(rs, 1)] = FMA(T1l, T1k, TD * T1C); } { E T1Q, T1S, T1P, T1R; T1Q = T1I + T1J; T1S = T1N - T1M; T1P = W[26]; T1R = W[27]; Rp[WS(rs, 7)] = FNMS(T1R, T1S, T1P * T1Q); Rm[WS(rs, 7)] = FMA(T1R, T1Q, T1P * T1S); } { E T1E, T1G, T1D, T1F; T1E = TO + T1j; T1G = T1B - T1o; T1D = W[34]; T1F = W[35]; Rp[WS(rs, 9)] = FNMS(T1F, T1G, T1D * T1E); Rm[WS(rs, 9)] = FMA(T1F, T1E, T1D * T1G); } { E T1K, T1O, T1H, T1L; T1K = T1I - T1J; T1O = T1M + T1N; T1H = W[10]; T1L = W[11]; Rp[WS(rs, 3)] = FNMS(T1L, T1O, T1H * T1K); Rm[WS(rs, 3)] = FMA(T1L, T1K, T1H * T1O); } } { E T2Q, T3p, T3B, T3x, T3m, T3A, T3b, T3w; T2Q = FNMS(KP951056516, T2P, KP587785252 * T2I); T3p = FNMS(KP951056516, T3o, KP587785252 * T3n); T3B = FMA(KP951056516, T3n, KP587785252 * T3o); T3x = FMA(KP951056516, T2I, KP587785252 * T2P); { E T3k, T3l, T39, T3a; T3k = FNMS(KP250000000, T3j, T3g); T3l = KP559016994 * (T3h - T3i); T3m = T3k - T3l; T3A = T3l + T3k; T39 = FNMS(KP250000000, T38, T2T); T3a = KP559016994 * (T30 - T37); T3b = T39 - T3a; T3w = T3a + T39; } { E T3c, T3q, T2B, T3d; T3c = T2Q + T3b; T3q = T3m - T3p; T2B = W[4]; T3d = W[5]; Ip[WS(rs, 1)] = FNMS(T3d, T3q, T2B * T3c); Im[WS(rs, 1)] = FMA(T2B, T3q, T3d * T3c); } { E T3E, T3G, T3D, T3F; T3E = T3x + T3w; T3G = T3A - T3B; T3D = W[36]; T3F = W[37]; Ip[WS(rs, 9)] = FNMS(T3F, T3G, T3D * T3E); Im[WS(rs, 9)] = FMA(T3D, T3G, T3F * T3E); } { E T3s, T3u, T3r, T3t; T3s = T3b - T2Q; T3u = T3m + T3p; T3r = W[12]; T3t = W[13]; Ip[WS(rs, 3)] = FNMS(T3t, T3u, T3r * T3s); Im[WS(rs, 3)] = FMA(T3r, T3u, T3t * T3s); } { E T3y, T3C, T3v, T3z; T3y = T3w - T3x; T3C = T3A + T3B; T3v = W[20]; T3z = W[21]; Ip[WS(rs, 5)] = FNMS(T3z, T3C, T3v * T3y); Im[WS(rs, 5)] = FMA(T3v, T3C, T3z * T3y); } } }}static const tw_instr twinstr[] = { {TW_FULL, 1, 20}, {TW_NEXT, 1, 0}};static const hc2c_desc desc = { 20, "hc2cb_20", twinstr, &GENUS, {184, 62, 62, 0} };void X(codelet_hc2cb_20) (planner *p) { X(khc2c_register) (p, hc2cb_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 + -