📄 hb_32.c
字号:
} } } } } } T5p = T5n * T5o; ci[WS(rs, 7)] = FMA(T5n, T5r, T5s); cr[WS(rs, 7)] = FNMS(T5q, T5r, T5p); }}static const tw_instr twinstr[] = { {TW_FULL, 1, 32}, {TW_NEXT, 1, 0}};static const hc2hc_desc desc = { 32, "hb_32", twinstr, &GENUS, {236, 62, 198, 0} };void X(codelet_hb_32) (planner *p) { X(khc2hc_register) (p, hb_32, &desc);}#else /* HAVE_FMA *//* Generated by: ../../../genfft/gen_hc2hc -compact -variables 4 -pipeline-latency 4 -sign 1 -n 32 -dif -name hb_32 -include hb.h *//* * This function contains 434 FP additions, 208 FP multiplications, * (or, 340 additions, 114 multiplications, 94 fused multiply/add), * 98 stack variables, 7 constants, and 128 memory accesses */#include "hb.h"static void hb_32(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms){ DK(KP555570233, +0.555570233019602224742830813948532874374937191); DK(KP831469612, +0.831469612302545237078788377617905756738560812); DK(KP980785280, +0.980785280403230449126182236134239036973933731); DK(KP195090322, +0.195090322016128267848284868477022240927691618); DK(KP923879532, +0.923879532511286756128183189396788286822416626); DK(KP382683432, +0.382683432365089771728459984030398866761344562); DK(KP707106781, +0.707106781186547524400844362104849039284835938); INT m; for (m = mb, W = W + ((mb - 1) * 62); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 62, MAKE_VOLATILE_STRIDE(rs)) { E T4o, T6y, T70, T5u, Tf, T12, T5x, T6z, T3m, T3Y, T29, T2y, T4v, T71, T2U; E T3M, Tu, T1U, T6D, T73, T6G, T74, T1h, T2z, T2X, T3o, T4D, T5A, T4K, T5z; E T30, T3n, TK, T1j, T6S, T7w, T6V, T7v, T1y, T2B, T3c, T3S, T4X, T61, T54; E T62, T3f, T3T, TZ, T1A, T6L, T7z, T6O, T7y, T1P, T2C, T35, T3P, T5g, T64; E T5n, T65, T38, T3Q; { E T3, T4m, T24, T4q, T27, T4t, T6, T5s, Ta, T4p, T1X, T5t, T20, T4n, Td; E T4s; { E T1, T2, T22, T23; T1 = cr[0]; T2 = ci[WS(rs, 15)]; T3 = T1 + T2; T4m = T1 - T2; T22 = ci[WS(rs, 27)]; T23 = cr[WS(rs, 20)]; T24 = T22 - T23; T4q = T22 + T23; } { E T25, T26, T4, T5; T25 = ci[WS(rs, 19)]; T26 = cr[WS(rs, 28)]; T27 = T25 - T26; T4t = T25 + T26; T4 = cr[WS(rs, 8)]; T5 = ci[WS(rs, 7)]; T6 = T4 + T5; T5s = T4 - T5; } { E T8, T9, T1V, T1W; T8 = cr[WS(rs, 4)]; T9 = ci[WS(rs, 11)]; Ta = T8 + T9; T4p = T8 - T9; T1V = ci[WS(rs, 31)]; T1W = cr[WS(rs, 16)]; T1X = T1V - T1W; T5t = T1V + T1W; } { E T1Y, T1Z, Tb, Tc; T1Y = ci[WS(rs, 23)]; T1Z = cr[WS(rs, 24)]; T20 = T1Y - T1Z; T4n = T1Y + T1Z; Tb = ci[WS(rs, 3)]; Tc = cr[WS(rs, 12)]; Td = Tb + Tc; T4s = Tb - Tc; } { E T7, Te, T21, T28; T4o = T4m - T4n; T6y = T4m + T4n; T70 = T5t - T5s; T5u = T5s + T5t; T7 = T3 + T6; Te = Ta + Td; Tf = T7 + Te; T12 = T7 - Te; { E T5v, T5w, T3k, T3l; T5v = T4p + T4q; T5w = T4s + T4t; T5x = KP707106781 * (T5v - T5w); T6z = KP707106781 * (T5v + T5w); T3k = T1X - T20; T3l = Ta - Td; T3m = T3k - T3l; T3Y = T3l + T3k; } T21 = T1X + T20; T28 = T24 + T27; T29 = T21 - T28; T2y = T21 + T28; { E T4r, T4u, T2S, T2T; T4r = T4p - T4q; T4u = T4s - T4t; T4v = KP707106781 * (T4r + T4u); T71 = KP707106781 * (T4r - T4u); T2S = T3 - T6; T2T = T27 - T24; T2U = T2S - T2T; T3M = T2S + T2T; } } } { E Ti, T4H, T1c, T4F, T1f, T4I, Tl, T4E, Tp, T4A, T15, T4y, T18, T4B, Ts; E T4x; { E Tg, Th, T1a, T1b; Tg = cr[WS(rs, 2)]; Th = ci[WS(rs, 13)]; Ti = Tg + Th; T4H = Tg - Th; T1a = ci[WS(rs, 29)]; T1b = cr[WS(rs, 18)]; T1c = T1a - T1b; T4F = T1a + T1b; } { E T1d, T1e, Tj, Tk; T1d = ci[WS(rs, 21)]; T1e = cr[WS(rs, 26)]; T1f = T1d - T1e; T4I = T1d + T1e; Tj = cr[WS(rs, 10)]; Tk = ci[WS(rs, 5)]; Tl = Tj + Tk; T4E = Tj - Tk; } { E Tn, To, T13, T14; Tn = ci[WS(rs, 1)]; To = cr[WS(rs, 14)]; Tp = Tn + To; T4A = Tn - To; T13 = ci[WS(rs, 17)]; T14 = cr[WS(rs, 30)]; T15 = T13 - T14; T4y = T13 + T14; } { E T16, T17, Tq, Tr; T16 = ci[WS(rs, 25)]; T17 = cr[WS(rs, 22)]; T18 = T16 - T17; T4B = T16 + T17; Tq = cr[WS(rs, 6)]; Tr = ci[WS(rs, 9)]; Ts = Tq + Tr; T4x = Tq - Tr; } { E Tm, Tt, T6B, T6C; Tm = Ti + Tl; Tt = Tp + Ts; Tu = Tm + Tt; T1U = Tm - Tt; T6B = T4H + T4I; T6C = T4F - T4E; T6D = FNMS(KP923879532, T6C, KP382683432 * T6B); T73 = FMA(KP382683432, T6C, KP923879532 * T6B); } { E T6E, T6F, T19, T1g; T6E = T4A + T4B; T6F = T4x + T4y; T6G = FNMS(KP923879532, T6F, KP382683432 * T6E); T74 = FMA(KP382683432, T6F, KP923879532 * T6E); T19 = T15 + T18; T1g = T1c + T1f; T1h = T19 - T1g; T2z = T1g + T19; } { E T2V, T2W, T4z, T4C; T2V = T15 - T18; T2W = Tp - Ts; T2X = T2V - T2W; T3o = T2W + T2V; T4z = T4x - T4y; T4C = T4A - T4B; T4D = FNMS(KP382683432, T4C, KP923879532 * T4z); T5A = FMA(KP382683432, T4z, KP923879532 * T4C); } { E T4G, T4J, T2Y, T2Z; T4G = T4E + T4F; T4J = T4H - T4I; T4K = FMA(KP923879532, T4G, KP382683432 * T4J); T5z = FNMS(KP382683432, T4G, KP923879532 * T4J); T2Y = Ti - Tl; T2Z = T1c - T1f; T30 = T2Y + T2Z; T3n = T2Y - T2Z; } } { E Ty, T4N, TB, T4Y, T1p, T4O, T1m, T4Z, TI, T52, T1w, T4V, TF, T51, T1t; E T4S; { E Tw, Tx, T1k, T1l; Tw = cr[WS(rs, 1)]; Tx = ci[WS(rs, 14)]; Ty = Tw + Tx; T4N = Tw - Tx; { E Tz, TA, T1n, T1o; Tz = cr[WS(rs, 9)]; TA = ci[WS(rs, 6)]; TB = Tz + TA; T4Y = Tz - TA; T1n = ci[WS(rs, 22)]; T1o = cr[WS(rs, 25)]; T1p = T1n - T1o; T4O = T1n + T1o; } T1k = ci[WS(rs, 30)]; T1l = cr[WS(rs, 17)]; T1m = T1k - T1l; T4Z = T1k + T1l; { E TG, TH, T4T, T1u, T1v, T4U; TG = ci[WS(rs, 2)]; TH = cr[WS(rs, 13)]; T4T = TG - TH; T1u = ci[WS(rs, 18)]; T1v = cr[WS(rs, 29)]; T4U = T1u + T1v; TI = TG + TH; T52 = T4T + T4U; T1w = T1u - T1v; T4V = T4T - T4U; } { E TD, TE, T4Q, T1r, T1s, T4R; TD = cr[WS(rs, 5)]; TE = ci[WS(rs, 10)]; T4Q = TD - TE; T1r = ci[WS(rs, 26)]; T1s = cr[WS(rs, 21)]; T4R = T1r + T1s; TF = TD + TE; T51 = T4Q + T4R; T1t = T1r - T1s; T4S = T4Q - T4R; } } { E TC, TJ, T6Q, T6R; TC = Ty + TB; TJ = TF + TI; TK = TC + TJ; T1j = TC - TJ; T6Q = T4Z - T4Y; T6R = KP707106781 * (T4S - T4V); T6S = T6Q + T6R; T7w = T6Q - T6R; } { E T6T, T6U, T1q, T1x; T6T = T4N + T4O; T6U = KP707106781 * (T51 + T52); T6V = T6T - T6U; T7v = T6T + T6U; T1q = T1m + T1p; T1x = T1t + T1w; T1y = T1q - T1x; T2B = T1q + T1x; } { E T3a, T3b, T4P, T4W; T3a = T1m - T1p; T3b = TF - TI; T3c = T3a - T3b; T3S = T3b + T3a; T4P = T4N - T4O; T4W = KP707106781 * (T4S + T4V); T4X = T4P - T4W; T61 = T4P + T4W; } { E T50, T53, T3d, T3e; T50 = T4Y + T4Z; T53 = KP707106781 * (T51 - T52); T54 = T50 - T53; T62 = T50 + T53; T3d = Ty - TB; T3e = T1w - T1t; T3f = T3d - T3e; T3T = T3d + T3e; } } { E TN, T56, TQ, T5h, T1G, T57, T1D, T5i, TX, T5l, T1N, T5e, TU, T5k, T1K; E T5b; { E TL, TM, T1B, T1C; TL = ci[0]; TM = cr[WS(rs, 15)]; TN = TL + TM; T56 = TL - TM; { E TO, TP, T1E, T1F; TO = cr[WS(rs, 7)]; TP = ci[WS(rs, 8)]; TQ = TO + TP; T5h = TO - TP; T1E = ci[WS(rs, 24)]; T1F = cr[WS(rs, 23)]; T1G = T1E - T1F; T57 = T1E + T1F; } T1B = ci[WS(rs, 16)]; T1C = cr[WS(rs, 31)]; T1D = T1B - T1C; T5i = T1B + T1C; { E TV, TW, T5c, T1L, T1M, T5d; TV = ci[WS(rs, 4)]; TW = cr[WS(rs, 11)]; T5c = TV - TW; T1L = ci[WS(rs, 20)]; T1M = cr[WS(rs, 27)]; T5d = T1L + T1M; TX = TV + TW; T5l = T5c + T5d; T1N = T1L - T1M; T5e = T5c - T5d; } { E TS, TT, T59, T1I, T1J, T5a; TS = cr[WS(rs, 3)]; TT = ci[WS(rs, 12)]; T59 = TS - TT; T1I = ci[WS(rs, 28)]; T1J = cr[WS(rs, 19)]; T5a = T1I + T1J; TU = TS + TT; T5k = T59 + T5a; T1K = T1I - T1J; T5b = T59 - T5a; } } { E TR, TY, T6J, T6K; TR = TN + TQ; TY = TU + TX; TZ = TR + TY; T1A = TR - TY; T6J = KP707106781 * (T5b - T5e); T6K = T5h + T5i; T6L = T6J - T6K; T7z = T6K + T6J; } { E T6M, T6N, T1H, T1O; T6M = T56 + T57; T6N = KP707106781 * (T5k + T5l); T6O = T6M - T6N; T7y = T6M + T6N; T1H = T1D + T1G; T1O = T1K + T1N; T1P = T1H - T1O; T2C = T1H + T1O; } { E T33, T34, T58, T5f; T33 = T1D - T1G; T34 = TU - TX; T35 = T33 - T34; T3P = T34 + T33; T58 = T56 - T57; T5f = KP707106781 * (T5b + T5e); T5g = T58 - T5f; T64 = T58 + T5f; } { E T5j, T5m, T36, T37; T5j = T5h - T5i; T5m = KP707106781 * (T5k - T5l); T5n = T5j - T5m; T65 = T5j + T5m; T36 = TN - TQ; T37 = T1N - T1K; T38 = T36 - T37; T3Q = T36 + T37; } } { E Tv, T10, T2w, T2A, T2D, T2E, T2v, T2x; Tv = Tf + Tu; T10 = TK + TZ; T2w = Tv - T10; T2A = T2y + T2z; T2D = T2B + T2C; T2E = T2A - T2D; cr[0] = Tv + T10; ci[0] = T2A + T2D; T2v = W[30]; T2x = W[31]; cr[WS(rs, 16)] = FNMS(T2x, T2E, T2v * T2w); ci[WS(rs, 16)] = FMA(T2x, T2w, T2v * T2E); } { E T2I, T2O, T2M, T2Q; { E T2G, T2H, T2K, T2L; T2G = Tf - Tu; T2H = T2C - T2B; T2I = T2G - T2H; T2O = T2G + T2H; T2K = T2y - T2z;
⌨️ 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 + -