📄 hb_32.c
字号:
rio[WS(ios, 15)] = FNMS(T80, T81, T7Z); T8c = T83 * T8b; T87 = T83 * T86; iio[-WS(ios, 24)] = FMA(T88, T86, T8c); rio[WS(ios, 7)] = FNMS(T88, T8b, T87); } return W;}static const tw_instr twinstr[] = { {TW_FULL, 0, 32}, {TW_NEXT, 1, 0}};static const hc2hc_desc desc = { 32, "hb_32", twinstr, &GENUS, {236, 62, 198, 0}, 0, 0, 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, and 128 memory accesses *//* * Generator Id's : * $Id: algsimp.ml,v 1.8 2006-01-05 03:04:27 stevenj Exp $ * $Id: fft.ml,v 1.4 2006-01-05 03:04:27 stevenj Exp $ * $Id: gen_hc2hc.ml,v 1.15 2006-01-05 03:04:27 stevenj Exp $ */#include "hb.h"static const R *hb_32(R *rio, R *iio, const R *W, stride ios, INT m, INT dist){ DK(KP555570233, +0.555570233019602224742830813948532874374937191); DK(KP831469612, +0.831469612302545237078788377617905756738560812); DK(KP195090322, +0.195090322016128267848284868477022240927691618); DK(KP980785280, +0.980785280403230449126182236134239036973933731); DK(KP382683432, +0.382683432365089771728459984030398866761344562); DK(KP923879532, +0.923879532511286756128183189396788286822416626); DK(KP707106781, +0.707106781186547524400844362104849039284835938); INT i; for (i = m - 2; i > 0; i = i - 2, rio = rio + dist, iio = iio - dist, W = W + 62, MAKE_VOLATILE_STRIDE(ios)) { E T5n, T6y, T77, T5u, Tf, T3i, T5x, T76, T3G, T47, T1a, T2I, T5k, T6z, T2o; E T2Y, Tu, T3D, T6D, T73, T6G, T74, T1j, T2d, T1s, T2e, T55, T5z, T5c, T5A; E T3l, T48, TK, T3n, T6L, T7t, T6O, T7s, T1D, T2L, T1M, T2M, T4w, T62, T4D; E T61, T3q, T41, TZ, T3s, T6S, T7w, T6V, T7v, T1W, T2O, T25, T2P, T4P, T64; E T4W, T65, T3v, T42; { E T3, T5l, T2j, T5t, T6, T5s, T2m, T5m, Ta, T5i, T15, T5h, Td, T5e, T18; E T5f; { E T1, T2, T2h, T2i; T1 = rio[0]; T2 = iio[-WS(ios, 16)]; T3 = T1 + T2; T5l = T1 - T2; T2h = iio[0]; T2i = rio[WS(ios, 16)]; T2j = T2h - T2i; T5t = T2h + T2i; } { E T4, T5, T2k, T2l; T4 = rio[WS(ios, 8)]; T5 = iio[-WS(ios, 24)]; T6 = T4 + T5; T5s = T4 - T5; T2k = iio[-WS(ios, 8)]; T2l = rio[WS(ios, 24)]; T2m = T2k - T2l; T5m = T2k + T2l; } { E T8, T9, T13, T14; T8 = rio[WS(ios, 4)]; T9 = iio[-WS(ios, 20)]; Ta = T8 + T9; T5i = T8 - T9; T13 = iio[-WS(ios, 4)]; T14 = rio[WS(ios, 20)]; T15 = T13 - T14; T5h = T13 + T14; } { E Tb, Tc, T16, T17; Tb = iio[-WS(ios, 28)]; Tc = rio[WS(ios, 12)]; Td = Tb + Tc; T5e = Tb - Tc; T16 = iio[-WS(ios, 12)]; T17 = rio[WS(ios, 28)]; T18 = T16 - T17; T5f = T17 + T16; } { E T7, Te, T12, T19; T5n = T5l - T5m; T6y = T5t - T5s; T77 = T5l + T5m; T5u = T5s + T5t; T7 = T3 + T6; Te = Ta + Td; Tf = T7 + Te; T3i = T7 - Te; { E T5v, T5w, T3E, T3F; T5v = T5i + T5h; T5w = T5e + T5f; T5x = KP707106781 * (T5v - T5w); T76 = KP707106781 * (T5v + T5w); T3E = T2j + T2m; T3F = T15 + T18; T3G = T3E - T3F; T47 = T3F + T3E; } T12 = T3 - T6; T19 = T15 - T18; T1a = T12 + T19; T2I = T12 - T19; { E T5g, T5j, T2g, T2n; T5g = T5e - T5f; T5j = T5h - T5i; T5k = KP707106781 * (T5g - T5j); T6z = KP707106781 * (T5j + T5g); T2g = Td - Ta; T2n = T2j - T2m; T2o = T2g + T2n; T2Y = T2n - T2g; } } } { E Ti, T4Z, T1e, T53, Tl, T52, T1h, T50, Tp, T56, T1n, T5a, Ts, T59, T1q; E T57; { E Tg, Th, T1c, T1d; Tg = rio[WS(ios, 2)]; Th = iio[-WS(ios, 18)]; Ti = Tg + Th; T4Z = Tg - Th; T1c = iio[-WS(ios, 2)]; T1d = rio[WS(ios, 18)]; T1e = T1c - T1d; T53 = T1c + T1d; } { E Tj, Tk, T1f, T1g; Tj = rio[WS(ios, 10)]; Tk = iio[-WS(ios, 26)]; Tl = Tj + Tk; T52 = Tj - Tk; T1f = iio[-WS(ios, 10)]; T1g = rio[WS(ios, 26)]; T1h = T1f - T1g; T50 = T1f + T1g; } { E Tn, To, T1l, T1m; Tn = iio[-WS(ios, 30)]; To = rio[WS(ios, 14)]; Tp = Tn + To; T56 = Tn - To; T1l = iio[-WS(ios, 14)]; T1m = rio[WS(ios, 30)]; T1n = T1l - T1m; T5a = T1m + T1l; } { E Tq, Tr, T1o, T1p; Tq = rio[WS(ios, 6)]; Tr = iio[-WS(ios, 22)]; Ts = Tq + Tr; T59 = Tq - Tr; T1o = iio[-WS(ios, 6)]; T1p = rio[WS(ios, 22)]; T1q = T1o - T1p; T57 = T1o + T1p; } { E Tm, Tt, T6B, T6C; Tm = Ti + Tl; Tt = Tp + Ts; Tu = Tm + Tt; T3D = Tt - Tm; T6B = T53 - T52; T6C = T4Z + T50; T6D = FNMS(KP382683432, T6C, KP923879532 * T6B); T73 = FMA(KP382683432, T6B, KP923879532 * T6C); } { E T6E, T6F, T1b, T1i; T6E = T56 + T57; T6F = T59 + T5a; T6G = FNMS(KP923879532, T6F, KP382683432 * T6E); T74 = FMA(KP923879532, T6E, KP382683432 * T6F); T1b = Ti - Tl; T1i = T1e - T1h; T1j = T1b + T1i; T2d = T1i - T1b; } { E T1k, T1r, T51, T54; T1k = Tp - Ts; T1r = T1n - T1q; T1s = T1k - T1r; T2e = T1k + T1r; T51 = T4Z - T50; T54 = T52 + T53; T55 = FNMS(KP382683432, T54, KP923879532 * T51); T5z = FMA(KP923879532, T54, KP382683432 * T51); } { E T58, T5b, T3j, T3k; T58 = T56 - T57; T5b = T59 - T5a; T5c = FMA(KP923879532, T58, KP382683432 * T5b); T5A = FNMS(KP382683432, T58, KP923879532 * T5b); T3j = T1e + T1h; T3k = T1q + T1n; T3l = T3j - T3k; T48 = T3j + T3k; } } { E Ty, T4t, T1H, T4y, TB, T4x, T1K, T4u, TI, T4B, T1B, T4o, TF, T4A, T1y; E T4r; { E Tw, Tx, T1I, T1J; Tw = rio[WS(ios, 1)]; Tx = iio[-WS(ios, 17)]; Ty = Tw + Tx; T4t = Tw - Tx; { E T1F, T1G, Tz, TA; T1F = iio[-WS(ios, 1)]; T1G = rio[WS(ios, 17)]; T1H = T1F - T1G; T4y = T1F + T1G; Tz = rio[WS(ios, 9)]; TA = iio[-WS(ios, 25)]; TB = Tz + TA; T4x = Tz - TA; } T1I = iio[-WS(ios, 9)]; T1J = rio[WS(ios, 25)]; T1K = T1I - T1J; T4u = T1I + T1J; { E TG, TH, T4m, T1z, T1A, T4n; TG = iio[-WS(ios, 29)]; TH = rio[WS(ios, 13)]; T4m = TG - TH; T1z = iio[-WS(ios, 13)]; T1A = rio[WS(ios, 29)]; T4n = T1A + T1z; TI = TG + TH; T4B = T4m + T4n; T1B = T1z - T1A; T4o = T4m - T4n; } { E TD, TE, T4q, T1w, T1x, T4p; TD = rio[WS(ios, 5)]; TE = iio[-WS(ios, 21)]; T4q = TD - TE; T1w = iio[-WS(ios, 5)]; T1x = rio[WS(ios, 21)]; T4p = T1w + T1x; TF = TD + TE; T4A = T4q + T4p; T1y = T1w - T1x; T4r = T4p - T4q; } } { E TC, TJ, T6J, T6K; TC = Ty + TB; TJ = TF + TI; TK = TC + TJ; T3n = TC - TJ; T6J = T4y - T4x; T6K = KP707106781 * (T4r + T4o); T6L = T6J + T6K; T7t = T6J - T6K; } { E T6M, T6N, T1v, T1C; T6M = KP707106781 * (T4A + T4B); T6N = T4t + T4u; T6O = T6M + T6N; T7s = T6N - T6M; T1v = Ty - TB; T1C = T1y - T1B; T1D = T1v + T1C; T2L = T1v - T1C; } { E T1E, T1L, T4s, T4v; T1E = TI - TF; T1L = T1H - T1K; T1M = T1E + T1L; T2M = T1L - T1E; T4s = KP707106781 * (T4o - T4r); T4v = T4t - T4u; T4w = T4s + T4v; T62 = T4v - T4s; } { E T4z, T4C, T3o, T3p; T4z = T4x + T4y; T4C = KP707106781 * (T4A - T4B); T4D = T4z + T4C; T61 = T4z - T4C; T3o = T1H + T1K; T3p = T1y + T1B; T3q = T3o - T3p; T41 = T3p + T3o; } } { E TN, T4T, T20, T4N, TQ, T4M, T23, T4U, TX, T4Q, T1U, T4K, TU, T4R, T1R; E T4H; { E TL, TM, T21, T22; TL = iio[-WS(ios, 31)]; TM = rio[WS(ios, 15)]; TN = TL + TM; T4T = TL - TM; { E T1Y, T1Z, TO, TP; T1Y = iio[-WS(ios, 15)]; T1Z = rio[WS(ios, 31)]; T20 = T1Y - T1Z; T4N = T1Z + T1Y; TO = rio[WS(ios, 7)]; TP = iio[-WS(ios, 23)]; TQ = TO + TP; T4M = TO - TP; } T21 = iio[-WS(ios, 7)]; T22 = rio[WS(ios, 23)]; T23 = T21 - T22; T4U = T21 + T22; { E TV, TW, T4I, T1S, T1T, T4J; TV = iio[-WS(ios, 27)]; TW = rio[WS(ios, 11)]; T4I = TV - TW; T1S = iio[-WS(ios, 11)]; T1T = rio[WS(ios, 27)]; T4J = T1T + T1S; TX = TV + TW; T4Q = T4I - T4J; T1U = T1S - T1T; T4K = T4I + T4J; } { E TS, TT, T4F, T1P, T1Q, T4G; TS = rio[WS(ios, 3)]; TT = iio[-WS(ios, 19)]; T4F = TS - TT; T1P = iio[-WS(ios, 3)]; T1Q = rio[WS(ios, 19)]; T4G = T1P + T1Q; TU = TS + TT; T4R = T4G - T4F; T1R = T1P - T1Q; T4H = T4F + T4G; } } { E TR, TY, T6Q, T6R; TR = TN + TQ; TY = TU + TX; TZ = TR + TY; T3s = TR - TY; T6Q = KP707106781 * (T4R + T4Q); T6R = T4M + T4N; T6S = T6Q - T6R; T7w = T6Q + T6R; } { E T6T, T6U, T1O, T1V; T6T = KP707106781 * (T4H + T4K); T6U = T4T + T4U; T6V = T6T + T6U; T7v = T6U - T6T; T1O = TN - TQ; T1V = T1R - T1U; T1W = T1O + T1V; T2O = T1O - T1V; } { E T1X, T24, T4L, T4O; T1X = TX - TU; T24 = T20 - T23; T25 = T1X + T24; T2P = T24 - T1X; T4L = KP707106781 * (T4H - T4K); T4O = T4M - T4N; T4P = T4L + T4O; T64 = T4O - T4L; } { E T4S, T4V, T3t, T3u; T4S = KP707106781 * (T4Q - T4R); T4V = T4T - T4U; T4W = T4S + T4V; T65 = T4V - T4S; T3t = T20 + T23; T3u = T1R + T1U; T3v = T3t - T3u; T42 = T3u + T3t; } } { E Tv, T10, T4g, T4i, T4j, T4k, T4f, T4h; Tv = Tf + Tu; T10 = TK + TZ; T4g = Tv - T10; T4i = T48 + T47; T4j = T41 + T42; T4k = T4i - T4j; rio[0] = Tv + T10; iio[-WS(ios, 31)] = T4j + T4i; T4f = W[30]; T4h = W[31]; rio[WS(ios, 16)] = FNMS(T4h, T4k, T4f * T4g); iio[-WS(ios, 15)] = FMA(T4h, T4g, T4f * T4k); } { E T44, T4c, T4a, T4e; { E T40, T43, T46, T49; T40 = Tf - Tu; T43 = T41 - T42; T44 = T40 + T43; T4c = T40 - T43;
⌨️ 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 + -