Chris@82: /* Chris@82: * Copyright (c) 2003, 2007-14 Matteo Frigo Chris@82: * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology Chris@82: * Chris@82: * This program is free software; you can redistribute it and/or modify Chris@82: * it under the terms of the GNU General Public License as published by Chris@82: * the Free Software Foundation; either version 2 of the License, or Chris@82: * (at your option) any later version. Chris@82: * Chris@82: * This program is distributed in the hope that it will be useful, Chris@82: * but WITHOUT ANY WARRANTY; without even the implied warranty of Chris@82: * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the Chris@82: * GNU General Public License for more details. Chris@82: * Chris@82: * You should have received a copy of the GNU General Public License Chris@82: * along with this program; if not, write to the Free Software Chris@82: * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA Chris@82: * Chris@82: */ Chris@82: Chris@82: /* This file was automatically generated --- DO NOT EDIT */ Chris@82: /* Generated on Thu May 24 08:06:31 EDT 2018 */ Chris@82: Chris@82: #include "rdft/codelet-rdft.h" Chris@82: Chris@82: #if defined(ARCH_PREFERS_FMA) || defined(ISA_EXTENSION_PREFERS_FMA) Chris@82: Chris@82: /* Generated by: ../../../genfft/gen_hc2hc.native -fma -compact -variables 4 -pipeline-latency 4 -n 64 -dit -name hf_64 -include rdft/scalar/hf.h */ Chris@82: Chris@82: /* Chris@82: * This function contains 1038 FP additions, 644 FP multiplications, Chris@82: * (or, 520 additions, 126 multiplications, 518 fused multiply/add), Chris@82: * 190 stack variables, 15 constants, and 256 memory accesses Chris@82: */ Chris@82: #include "rdft/scalar/hf.h" Chris@82: Chris@82: static void hf_64(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms) Chris@82: { Chris@82: DK(KP881921264, +0.881921264348355029712756863660388349508442621); Chris@82: DK(KP956940335, +0.956940335732208864935797886980269969482849206); Chris@82: DK(KP831469612, +0.831469612302545237078788377617905756738560812); Chris@82: DK(KP773010453, +0.773010453362736960810906609758469800971041293); Chris@82: DK(KP995184726, +0.995184726672196886244836953109479921575474869); Chris@82: DK(KP980785280, +0.980785280403230449126182236134239036973933731); Chris@82: DK(KP668178637, +0.668178637919298919997757686523080761552472251); Chris@82: DK(KP303346683, +0.303346683607342391675883946941299872384187453); Chris@82: DK(KP534511135, +0.534511135950791641089685961295362908582039528); Chris@82: DK(KP820678790, +0.820678790828660330972281985331011598767386482); Chris@82: DK(KP098491403, +0.098491403357164253077197521291327432293052451); Chris@82: DK(KP923879532, +0.923879532511286756128183189396788286822416626); Chris@82: DK(KP198912367, +0.198912367379658006911597622644676228597850501); Chris@82: DK(KP707106781, +0.707106781186547524400844362104849039284835938); Chris@82: DK(KP414213562, +0.414213562373095048801688724209698078569671875); Chris@82: { Chris@82: INT m; Chris@82: for (m = mb, W = W + ((mb - 1) * 126); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 126, MAKE_VOLATILE_STRIDE(128, rs)) { Chris@82: E Tm, TeM, TjR, Tkl, T7e, TcA, TiV, Tjm, T1G, TeW, TeZ, Thr, T7Q, TcI, T7X; Chris@82: E TcJ, T29, Tf8, Tf5, Thw, T87, TcN, T8u, TcQ, T5K, TfS, Tgc, ThX, Taq, Tdm; Chris@82: E Tbj, Tdx, TN, Tjl, TeP, TiP, T7l, TcB, T7s, TcC, T1f, TeR, TeU, Ths, T7B; Chris@82: E TcF, T7I, TcG, T32, Tfj, Tfg, ThB, T8G, TcU, T93, TcX, T3X, Tfr, TfK, ThM; Chris@82: E T9h, Td3, Taa, Tde, T2A, Tf6, Tfb, Thx, T8m, TcR, T8x, TcO, T3t, Tfh, Tfm; Chris@82: E ThC, T8V, TcY, T96, TcV, T4o, TfL, Tfu, ThN, T9w, Tdf, Tad, Td4, T6b, Tg9; Chris@82: E TfV, ThY, TaF, Tdy, Tbm, Tdn, T4Q, ThJ, TfA, TfN, T9M, Tdh, Taf, Td8, T5h; Chris@82: E ThI, TfF, TfO, Ta1, Tdi, Tag, Tdb, T6D, ThU, Tg1, Tgf, TaV, TdA, Tbo, Tdr; Chris@82: E T74, ThT, Tg6, Tge, Tba, TdB, Tbp, Tdu; Chris@82: { Chris@82: E T1, TiT, T7, TiS, Te, T7a, Tk, T7c; Chris@82: T1 = cr[0]; Chris@82: TiT = ci[0]; Chris@82: { Chris@82: E T3, T6, T4, TiR, T2, T5; Chris@82: T3 = cr[WS(rs, 32)]; Chris@82: T6 = ci[WS(rs, 32)]; Chris@82: T2 = W[62]; Chris@82: T4 = T2 * T3; Chris@82: TiR = T2 * T6; Chris@82: T5 = W[63]; Chris@82: T7 = FMA(T5, T6, T4); Chris@82: TiS = FNMS(T5, T3, TiR); Chris@82: } Chris@82: { Chris@82: E Ta, Td, Tb, T79, T9, Tc; Chris@82: Ta = cr[WS(rs, 16)]; Chris@82: Td = ci[WS(rs, 16)]; Chris@82: T9 = W[30]; Chris@82: Tb = T9 * Ta; Chris@82: T79 = T9 * Td; Chris@82: Tc = W[31]; Chris@82: Te = FMA(Tc, Td, Tb); Chris@82: T7a = FNMS(Tc, Ta, T79); Chris@82: } Chris@82: { Chris@82: E Tg, Tj, Th, T7b, Tf, Ti; Chris@82: Tg = cr[WS(rs, 48)]; Chris@82: Tj = ci[WS(rs, 48)]; Chris@82: Tf = W[94]; Chris@82: Th = Tf * Tg; Chris@82: T7b = Tf * Tj; Chris@82: Ti = W[95]; Chris@82: Tk = FMA(Ti, Tj, Th); Chris@82: T7c = FNMS(Ti, Tg, T7b); Chris@82: } Chris@82: { Chris@82: E T8, Tl, TjP, TjQ; Chris@82: T8 = T1 + T7; Chris@82: Tl = Te + Tk; Chris@82: Tm = T8 + Tl; Chris@82: TeM = T8 - Tl; Chris@82: TjP = Te - Tk; Chris@82: TjQ = TiT - TiS; Chris@82: TjR = TjP + TjQ; Chris@82: Tkl = TjQ - TjP; Chris@82: } Chris@82: { Chris@82: E T78, T7d, TiQ, TiU; Chris@82: T78 = T1 - T7; Chris@82: T7d = T7a - T7c; Chris@82: T7e = T78 - T7d; Chris@82: TcA = T78 + T7d; Chris@82: TiQ = T7a + T7c; Chris@82: TiU = TiS + TiT; Chris@82: TiV = TiQ + TiU; Chris@82: Tjm = TiU - TiQ; Chris@82: } Chris@82: } Chris@82: { Chris@82: E T1l, T7S, T1E, T7O, T1r, T7U, T1y, T7M; Chris@82: { Chris@82: E T1h, T1k, T1i, T7R, T1g, T1j; Chris@82: T1h = cr[WS(rs, 60)]; Chris@82: T1k = ci[WS(rs, 60)]; Chris@82: T1g = W[118]; Chris@82: T1i = T1g * T1h; Chris@82: T7R = T1g * T1k; Chris@82: T1j = W[119]; Chris@82: T1l = FMA(T1j, T1k, T1i); Chris@82: T7S = FNMS(T1j, T1h, T7R); Chris@82: } Chris@82: { Chris@82: E T1A, T1D, T1B, T7N, T1z, T1C; Chris@82: T1A = cr[WS(rs, 44)]; Chris@82: T1D = ci[WS(rs, 44)]; Chris@82: T1z = W[86]; Chris@82: T1B = T1z * T1A; Chris@82: T7N = T1z * T1D; Chris@82: T1C = W[87]; Chris@82: T1E = FMA(T1C, T1D, T1B); Chris@82: T7O = FNMS(T1C, T1A, T7N); Chris@82: } Chris@82: { Chris@82: E T1n, T1q, T1o, T7T, T1m, T1p; Chris@82: T1n = cr[WS(rs, 28)]; Chris@82: T1q = ci[WS(rs, 28)]; Chris@82: T1m = W[54]; Chris@82: T1o = T1m * T1n; Chris@82: T7T = T1m * T1q; Chris@82: T1p = W[55]; Chris@82: T1r = FMA(T1p, T1q, T1o); Chris@82: T7U = FNMS(T1p, T1n, T7T); Chris@82: } Chris@82: { Chris@82: E T1u, T1x, T1v, T7L, T1t, T1w; Chris@82: T1u = cr[WS(rs, 12)]; Chris@82: T1x = ci[WS(rs, 12)]; Chris@82: T1t = W[22]; Chris@82: T1v = T1t * T1u; Chris@82: T7L = T1t * T1x; Chris@82: T1w = W[23]; Chris@82: T1y = FMA(T1w, T1x, T1v); Chris@82: T7M = FNMS(T1w, T1u, T7L); Chris@82: } Chris@82: { Chris@82: E T1s, T1F, TeX, TeY; Chris@82: T1s = T1l + T1r; Chris@82: T1F = T1y + T1E; Chris@82: T1G = T1s + T1F; Chris@82: TeW = T1s - T1F; Chris@82: TeX = T7S + T7U; Chris@82: TeY = T7M + T7O; Chris@82: TeZ = TeX - TeY; Chris@82: Thr = TeX + TeY; Chris@82: } Chris@82: { Chris@82: E T7K, T7P, T7V, T7W; Chris@82: T7K = T1l - T1r; Chris@82: T7P = T7M - T7O; Chris@82: T7Q = T7K - T7P; Chris@82: TcI = T7K + T7P; Chris@82: T7V = T7S - T7U; Chris@82: T7W = T1y - T1E; Chris@82: T7X = T7V + T7W; Chris@82: TcJ = T7V - T7W; Chris@82: } Chris@82: } Chris@82: { Chris@82: E T1O, T8p, T27, T85, T1U, T8r, T21, T83; Chris@82: { Chris@82: E T1K, T1N, T1L, T8o, T1J, T1M; Chris@82: T1K = cr[WS(rs, 2)]; Chris@82: T1N = ci[WS(rs, 2)]; Chris@82: T1J = W[2]; Chris@82: T1L = T1J * T1K; Chris@82: T8o = T1J * T1N; Chris@82: T1M = W[3]; Chris@82: T1O = FMA(T1M, T1N, T1L); Chris@82: T8p = FNMS(T1M, T1K, T8o); Chris@82: } Chris@82: { Chris@82: E T23, T26, T24, T84, T22, T25; Chris@82: T23 = cr[WS(rs, 50)]; Chris@82: T26 = ci[WS(rs, 50)]; Chris@82: T22 = W[98]; Chris@82: T24 = T22 * T23; Chris@82: T84 = T22 * T26; Chris@82: T25 = W[99]; Chris@82: T27 = FMA(T25, T26, T24); Chris@82: T85 = FNMS(T25, T23, T84); Chris@82: } Chris@82: { Chris@82: E T1Q, T1T, T1R, T8q, T1P, T1S; Chris@82: T1Q = cr[WS(rs, 34)]; Chris@82: T1T = ci[WS(rs, 34)]; Chris@82: T1P = W[66]; Chris@82: T1R = T1P * T1Q; Chris@82: T8q = T1P * T1T; Chris@82: T1S = W[67]; Chris@82: T1U = FMA(T1S, T1T, T1R); Chris@82: T8r = FNMS(T1S, T1Q, T8q); Chris@82: } Chris@82: { Chris@82: E T1X, T20, T1Y, T82, T1W, T1Z; Chris@82: T1X = cr[WS(rs, 18)]; Chris@82: T20 = ci[WS(rs, 18)]; Chris@82: T1W = W[34]; Chris@82: T1Y = T1W * T1X; Chris@82: T82 = T1W * T20; Chris@82: T1Z = W[35]; Chris@82: T21 = FMA(T1Z, T20, T1Y); Chris@82: T83 = FNMS(T1Z, T1X, T82); Chris@82: } Chris@82: { Chris@82: E T1V, T28, Tf3, Tf4; Chris@82: T1V = T1O + T1U; Chris@82: T28 = T21 + T27; Chris@82: T29 = T1V + T28; Chris@82: Tf8 = T1V - T28; Chris@82: Tf3 = T8p + T8r; Chris@82: Tf4 = T83 + T85; Chris@82: Tf5 = Tf3 - Tf4; Chris@82: Thw = Tf3 + Tf4; Chris@82: } Chris@82: { Chris@82: E T81, T86, T8s, T8t; Chris@82: T81 = T1O - T1U; Chris@82: T86 = T83 - T85; Chris@82: T87 = T81 - T86; Chris@82: TcN = T81 + T86; Chris@82: T8s = T8p - T8r; Chris@82: T8t = T21 - T27; Chris@82: T8u = T8s + T8t; Chris@82: TcQ = T8s - T8t; Chris@82: } Chris@82: } Chris@82: { Chris@82: E T5p, Tbf, T5I, Tao, T5v, Tbh, T5C, Tam; Chris@82: { Chris@82: E T5l, T5o, T5m, Tbe, T5k, T5n; Chris@82: T5l = cr[WS(rs, 63)]; Chris@82: T5o = ci[WS(rs, 63)]; Chris@82: T5k = W[124]; Chris@82: T5m = T5k * T5l; Chris@82: Tbe = T5k * T5o; Chris@82: T5n = W[125]; Chris@82: T5p = FMA(T5n, T5o, T5m); Chris@82: Tbf = FNMS(T5n, T5l, Tbe); Chris@82: } Chris@82: { Chris@82: E T5E, T5H, T5F, Tan, T5D, T5G; Chris@82: T5E = cr[WS(rs, 47)]; Chris@82: T5H = ci[WS(rs, 47)]; Chris@82: T5D = W[92]; Chris@82: T5F = T5D * T5E; Chris@82: Tan = T5D * T5H; Chris@82: T5G = W[93]; Chris@82: T5I = FMA(T5G, T5H, T5F); Chris@82: Tao = FNMS(T5G, T5E, Tan); Chris@82: } Chris@82: { Chris@82: E T5r, T5u, T5s, Tbg, T5q, T5t; Chris@82: T5r = cr[WS(rs, 31)]; Chris@82: T5u = ci[WS(rs, 31)]; Chris@82: T5q = W[60]; Chris@82: T5s = T5q * T5r; Chris@82: Tbg = T5q * T5u; Chris@82: T5t = W[61]; Chris@82: T5v = FMA(T5t, T5u, T5s); Chris@82: Tbh = FNMS(T5t, T5r, Tbg); Chris@82: } Chris@82: { Chris@82: E T5y, T5B, T5z, Tal, T5x, T5A; Chris@82: T5y = cr[WS(rs, 15)]; Chris@82: T5B = ci[WS(rs, 15)]; Chris@82: T5x = W[28]; Chris@82: T5z = T5x * T5y; Chris@82: Tal = T5x * T5B; Chris@82: T5A = W[29]; Chris@82: T5C = FMA(T5A, T5B, T5z); Chris@82: Tam = FNMS(T5A, T5y, Tal); Chris@82: } Chris@82: { Chris@82: E T5w, T5J, Tga, Tgb; Chris@82: T5w = T5p + T5v; Chris@82: T5J = T5C + T5I; Chris@82: T5K = T5w + T5J; Chris@82: TfS = T5w - T5J; Chris@82: Tga = Tbf + Tbh; Chris@82: Tgb = Tam + Tao; Chris@82: Tgc = Tga - Tgb; Chris@82: ThX = Tga + Tgb; Chris@82: } Chris@82: { Chris@82: E Tak, Tap, Tbd, Tbi; Chris@82: Tak = T5p - T5v; Chris@82: Tap = Tam - Tao; Chris@82: Taq = Tak - Tap; Chris@82: Tdm = Tak + Tap; Chris@82: Tbd = T5I - T5C; Chris@82: Tbi = Tbf - Tbh; Chris@82: Tbj = Tbd - Tbi; Chris@82: Tdx = Tbi + Tbd; Chris@82: } Chris@82: } Chris@82: { Chris@82: E Ts, T7h, TL, T7q, Ty, T7j, TF, T7o; Chris@82: { Chris@82: E To, Tr, Tp, T7g, Tn, Tq; Chris@82: To = cr[WS(rs, 8)]; Chris@82: Tr = ci[WS(rs, 8)]; Chris@82: Tn = W[14]; Chris@82: Tp = Tn * To; Chris@82: T7g = Tn * Tr; Chris@82: Tq = W[15]; Chris@82: Ts = FMA(Tq, Tr, Tp); Chris@82: T7h = FNMS(Tq, To, T7g); Chris@82: } Chris@82: { Chris@82: E TH, TK, TI, T7p, TG, TJ; Chris@82: TH = cr[WS(rs, 24)]; Chris@82: TK = ci[WS(rs, 24)]; Chris@82: TG = W[46]; Chris@82: TI = TG * TH; Chris@82: T7p = TG * TK; Chris@82: TJ = W[47]; Chris@82: TL = FMA(TJ, TK, TI); Chris@82: T7q = FNMS(TJ, TH, T7p); Chris@82: } Chris@82: { Chris@82: E Tu, Tx, Tv, T7i, Tt, Tw; Chris@82: Tu = cr[WS(rs, 40)]; Chris@82: Tx = ci[WS(rs, 40)]; Chris@82: Tt = W[78]; Chris@82: Tv = Tt * Tu; Chris@82: T7i = Tt * Tx; Chris@82: Tw = W[79]; Chris@82: Ty = FMA(Tw, Tx, Tv); Chris@82: T7j = FNMS(Tw, Tu, T7i); Chris@82: } Chris@82: { Chris@82: E TB, TE, TC, T7n, TA, TD; Chris@82: TB = cr[WS(rs, 56)]; Chris@82: TE = ci[WS(rs, 56)]; Chris@82: TA = W[110]; Chris@82: TC = TA * TB; Chris@82: T7n = TA * TE; Chris@82: TD = W[111]; Chris@82: TF = FMA(TD, TE, TC); Chris@82: T7o = FNMS(TD, TB, T7n); Chris@82: } Chris@82: { Chris@82: E Tz, TM, TeN, TeO; Chris@82: Tz = Ts + Ty; Chris@82: TM = TF + TL; Chris@82: TN = Tz + TM; Chris@82: Tjl = Tz - TM; Chris@82: TeN = T7o + T7q; Chris@82: TeO = T7h + T7j; Chris@82: TeP = TeN - TeO; Chris@82: TiP = TeO + TeN; Chris@82: } Chris@82: { Chris@82: E T7f, T7k, T7m, T7r; Chris@82: T7f = Ts - Ty; Chris@82: T7k = T7h - T7j; Chris@82: T7l = T7f - T7k; Chris@82: TcB = T7f + T7k; Chris@82: T7m = TF - TL; Chris@82: T7r = T7o - T7q; Chris@82: T7s = T7m + T7r; Chris@82: TcC = T7m - T7r; Chris@82: } Chris@82: } Chris@82: { Chris@82: E TU, T7D, T1d, T7z, T10, T7F, T17, T7x; Chris@82: { Chris@82: E TQ, TT, TR, T7C, TP, TS; Chris@82: TQ = cr[WS(rs, 4)]; Chris@82: TT = ci[WS(rs, 4)]; Chris@82: TP = W[6]; Chris@82: TR = TP * TQ; Chris@82: T7C = TP * TT; Chris@82: TS = W[7]; Chris@82: TU = FMA(TS, TT, TR); Chris@82: T7D = FNMS(TS, TQ, T7C); Chris@82: } Chris@82: { Chris@82: E T19, T1c, T1a, T7y, T18, T1b; Chris@82: T19 = cr[WS(rs, 52)]; Chris@82: T1c = ci[WS(rs, 52)]; Chris@82: T18 = W[102]; Chris@82: T1a = T18 * T19; Chris@82: T7y = T18 * T1c; Chris@82: T1b = W[103]; Chris@82: T1d = FMA(T1b, T1c, T1a); Chris@82: T7z = FNMS(T1b, T19, T7y); Chris@82: } Chris@82: { Chris@82: E TW, TZ, TX, T7E, TV, TY; Chris@82: TW = cr[WS(rs, 36)]; Chris@82: TZ = ci[WS(rs, 36)]; Chris@82: TV = W[70]; Chris@82: TX = TV * TW; Chris@82: T7E = TV * TZ; Chris@82: TY = W[71]; Chris@82: T10 = FMA(TY, TZ, TX); Chris@82: T7F = FNMS(TY, TW, T7E); Chris@82: } Chris@82: { Chris@82: E T13, T16, T14, T7w, T12, T15; Chris@82: T13 = cr[WS(rs, 20)]; Chris@82: T16 = ci[WS(rs, 20)]; Chris@82: T12 = W[38]; Chris@82: T14 = T12 * T13; Chris@82: T7w = T12 * T16; Chris@82: T15 = W[39]; Chris@82: T17 = FMA(T15, T16, T14); Chris@82: T7x = FNMS(T15, T13, T7w); Chris@82: } Chris@82: { Chris@82: E T11, T1e, TeS, TeT; Chris@82: T11 = TU + T10; Chris@82: T1e = T17 + T1d; Chris@82: T1f = T11 + T1e; Chris@82: TeR = T11 - T1e; Chris@82: TeS = T7D + T7F; Chris@82: TeT = T7x + T7z; Chris@82: TeU = TeS - TeT; Chris@82: Ths = TeS + TeT; Chris@82: } Chris@82: { Chris@82: E T7v, T7A, T7G, T7H; Chris@82: T7v = TU - T10; Chris@82: T7A = T7x - T7z; Chris@82: T7B = T7v - T7A; Chris@82: TcF = T7v + T7A; Chris@82: T7G = T7D - T7F; Chris@82: T7H = T17 - T1d; Chris@82: T7I = T7G + T7H; Chris@82: TcG = T7G - T7H; Chris@82: } Chris@82: } Chris@82: { Chris@82: E T2H, T8Y, T30, T8E, T2N, T90, T2U, T8C; Chris@82: { Chris@82: E T2D, T2G, T2E, T8X, T2C, T2F; Chris@82: T2D = cr[WS(rs, 62)]; Chris@82: T2G = ci[WS(rs, 62)]; Chris@82: T2C = W[122]; Chris@82: T2E = T2C * T2D; Chris@82: T8X = T2C * T2G; Chris@82: T2F = W[123]; Chris@82: T2H = FMA(T2F, T2G, T2E); Chris@82: T8Y = FNMS(T2F, T2D, T8X); Chris@82: } Chris@82: { Chris@82: E T2W, T2Z, T2X, T8D, T2V, T2Y; Chris@82: T2W = cr[WS(rs, 46)]; Chris@82: T2Z = ci[WS(rs, 46)]; Chris@82: T2V = W[90]; Chris@82: T2X = T2V * T2W; Chris@82: T8D = T2V * T2Z; Chris@82: T2Y = W[91]; Chris@82: T30 = FMA(T2Y, T2Z, T2X); Chris@82: T8E = FNMS(T2Y, T2W, T8D); Chris@82: } Chris@82: { Chris@82: E T2J, T2M, T2K, T8Z, T2I, T2L; Chris@82: T2J = cr[WS(rs, 30)]; Chris@82: T2M = ci[WS(rs, 30)]; Chris@82: T2I = W[58]; Chris@82: T2K = T2I * T2J; Chris@82: T8Z = T2I * T2M; Chris@82: T2L = W[59]; Chris@82: T2N = FMA(T2L, T2M, T2K); Chris@82: T90 = FNMS(T2L, T2J, T8Z); Chris@82: } Chris@82: { Chris@82: E T2Q, T2T, T2R, T8B, T2P, T2S; Chris@82: T2Q = cr[WS(rs, 14)]; Chris@82: T2T = ci[WS(rs, 14)]; Chris@82: T2P = W[26]; Chris@82: T2R = T2P * T2Q; Chris@82: T8B = T2P * T2T; Chris@82: T2S = W[27]; Chris@82: T2U = FMA(T2S, T2T, T2R); Chris@82: T8C = FNMS(T2S, T2Q, T8B); Chris@82: } Chris@82: { Chris@82: E T2O, T31, Tfe, Tff; Chris@82: T2O = T2H + T2N; Chris@82: T31 = T2U + T30; Chris@82: T32 = T2O + T31; Chris@82: Tfj = T2O - T31; Chris@82: Tfe = T8Y + T90; Chris@82: Tff = T8C + T8E; Chris@82: Tfg = Tfe - Tff; Chris@82: ThB = Tfe + Tff; Chris@82: } Chris@82: { Chris@82: E T8A, T8F, T91, T92; Chris@82: T8A = T2H - T2N; Chris@82: T8F = T8C - T8E; Chris@82: T8G = T8A - T8F; Chris@82: TcU = T8A + T8F; Chris@82: T91 = T8Y - T90; Chris@82: T92 = T2U - T30; Chris@82: T93 = T91 + T92; Chris@82: TcX = T91 - T92; Chris@82: } Chris@82: } Chris@82: { Chris@82: E T3C, Ta5, T3V, T9f, T3I, Ta7, T3P, T9d; Chris@82: { Chris@82: E T3y, T3B, T3z, Ta4, T3x, T3A; Chris@82: T3y = cr[WS(rs, 1)]; Chris@82: T3B = ci[WS(rs, 1)]; Chris@82: T3x = W[0]; Chris@82: T3z = T3x * T3y; Chris@82: Ta4 = T3x * T3B; Chris@82: T3A = W[1]; Chris@82: T3C = FMA(T3A, T3B, T3z); Chris@82: Ta5 = FNMS(T3A, T3y, Ta4); Chris@82: } Chris@82: { Chris@82: E T3R, T3U, T3S, T9e, T3Q, T3T; Chris@82: T3R = cr[WS(rs, 49)]; Chris@82: T3U = ci[WS(rs, 49)]; Chris@82: T3Q = W[96]; Chris@82: T3S = T3Q * T3R; Chris@82: T9e = T3Q * T3U; Chris@82: T3T = W[97]; Chris@82: T3V = FMA(T3T, T3U, T3S); Chris@82: T9f = FNMS(T3T, T3R, T9e); Chris@82: } Chris@82: { Chris@82: E T3E, T3H, T3F, Ta6, T3D, T3G; Chris@82: T3E = cr[WS(rs, 33)]; Chris@82: T3H = ci[WS(rs, 33)]; Chris@82: T3D = W[64]; Chris@82: T3F = T3D * T3E; Chris@82: Ta6 = T3D * T3H; Chris@82: T3G = W[65]; Chris@82: T3I = FMA(T3G, T3H, T3F); Chris@82: Ta7 = FNMS(T3G, T3E, Ta6); Chris@82: } Chris@82: { Chris@82: E T3L, T3O, T3M, T9c, T3K, T3N; Chris@82: T3L = cr[WS(rs, 17)]; Chris@82: T3O = ci[WS(rs, 17)]; Chris@82: T3K = W[32]; Chris@82: T3M = T3K * T3L; Chris@82: T9c = T3K * T3O; Chris@82: T3N = W[33]; Chris@82: T3P = FMA(T3N, T3O, T3M); Chris@82: T9d = FNMS(T3N, T3L, T9c); Chris@82: } Chris@82: { Chris@82: E T3J, T3W, TfI, TfJ; Chris@82: T3J = T3C + T3I; Chris@82: T3W = T3P + T3V; Chris@82: T3X = T3J + T3W; Chris@82: Tfr = T3J - T3W; Chris@82: TfI = Ta5 + Ta7; Chris@82: TfJ = T9d + T9f; Chris@82: TfK = TfI - TfJ; Chris@82: ThM = TfI + TfJ; Chris@82: } Chris@82: { Chris@82: E T9b, T9g, Ta8, Ta9; Chris@82: T9b = T3C - T3I; Chris@82: T9g = T9d - T9f; Chris@82: T9h = T9b - T9g; Chris@82: Td3 = T9b + T9g; Chris@82: Ta8 = Ta5 - Ta7; Chris@82: Ta9 = T3P - T3V; Chris@82: Taa = Ta8 + Ta9; Chris@82: Tde = Ta8 - Ta9; Chris@82: } Chris@82: } Chris@82: { Chris@82: E T2f, T8a, T2y, T8j, T2l, T8c, T2s, T8h; Chris@82: { Chris@82: E T2b, T2e, T2c, T89, T2a, T2d; Chris@82: T2b = cr[WS(rs, 10)]; Chris@82: T2e = ci[WS(rs, 10)]; Chris@82: T2a = W[18]; Chris@82: T2c = T2a * T2b; Chris@82: T89 = T2a * T2e; Chris@82: T2d = W[19]; Chris@82: T2f = FMA(T2d, T2e, T2c); Chris@82: T8a = FNMS(T2d, T2b, T89); Chris@82: } Chris@82: { Chris@82: E T2u, T2x, T2v, T8i, T2t, T2w; Chris@82: T2u = cr[WS(rs, 26)]; Chris@82: T2x = ci[WS(rs, 26)]; Chris@82: T2t = W[50]; Chris@82: T2v = T2t * T2u; Chris@82: T8i = T2t * T2x; Chris@82: T2w = W[51]; Chris@82: T2y = FMA(T2w, T2x, T2v); Chris@82: T8j = FNMS(T2w, T2u, T8i); Chris@82: } Chris@82: { Chris@82: E T2h, T2k, T2i, T8b, T2g, T2j; Chris@82: T2h = cr[WS(rs, 42)]; Chris@82: T2k = ci[WS(rs, 42)]; Chris@82: T2g = W[82]; Chris@82: T2i = T2g * T2h; Chris@82: T8b = T2g * T2k; Chris@82: T2j = W[83]; Chris@82: T2l = FMA(T2j, T2k, T2i); Chris@82: T8c = FNMS(T2j, T2h, T8b); Chris@82: } Chris@82: { Chris@82: E T2o, T2r, T2p, T8g, T2n, T2q; Chris@82: T2o = cr[WS(rs, 58)]; Chris@82: T2r = ci[WS(rs, 58)]; Chris@82: T2n = W[114]; Chris@82: T2p = T2n * T2o; Chris@82: T8g = T2n * T2r; Chris@82: T2q = W[115]; Chris@82: T2s = FMA(T2q, T2r, T2p); Chris@82: T8h = FNMS(T2q, T2o, T8g); Chris@82: } Chris@82: { Chris@82: E T2m, T2z, Tf9, Tfa; Chris@82: T2m = T2f + T2l; Chris@82: T2z = T2s + T2y; Chris@82: T2A = T2m + T2z; Chris@82: Tf6 = T2m - T2z; Chris@82: Tf9 = T8h + T8j; Chris@82: Tfa = T8a + T8c; Chris@82: Tfb = Tf9 - Tfa; Chris@82: Thx = Tfa + Tf9; Chris@82: { Chris@82: E T8e, T8v, T8l, T8w; Chris@82: { Chris@82: E T88, T8d, T8f, T8k; Chris@82: T88 = T2f - T2l; Chris@82: T8d = T8a - T8c; Chris@82: T8e = T88 - T8d; Chris@82: T8v = T88 + T8d; Chris@82: T8f = T2s - T2y; Chris@82: T8k = T8h - T8j; Chris@82: T8l = T8f + T8k; Chris@82: T8w = T8k - T8f; Chris@82: } Chris@82: T8m = T8e + T8l; Chris@82: TcR = T8l - T8e; Chris@82: T8x = T8v + T8w; Chris@82: TcO = T8v - T8w; Chris@82: } Chris@82: } Chris@82: } Chris@82: { Chris@82: E T38, T8J, T3r, T8S, T3e, T8L, T3l, T8Q; Chris@82: { Chris@82: E T34, T37, T35, T8I, T33, T36; Chris@82: T34 = cr[WS(rs, 6)]; Chris@82: T37 = ci[WS(rs, 6)]; Chris@82: T33 = W[10]; Chris@82: T35 = T33 * T34; Chris@82: T8I = T33 * T37; Chris@82: T36 = W[11]; Chris@82: T38 = FMA(T36, T37, T35); Chris@82: T8J = FNMS(T36, T34, T8I); Chris@82: } Chris@82: { Chris@82: E T3n, T3q, T3o, T8R, T3m, T3p; Chris@82: T3n = cr[WS(rs, 22)]; Chris@82: T3q = ci[WS(rs, 22)]; Chris@82: T3m = W[42]; Chris@82: T3o = T3m * T3n; Chris@82: T8R = T3m * T3q; Chris@82: T3p = W[43]; Chris@82: T3r = FMA(T3p, T3q, T3o); Chris@82: T8S = FNMS(T3p, T3n, T8R); Chris@82: } Chris@82: { Chris@82: E T3a, T3d, T3b, T8K, T39, T3c; Chris@82: T3a = cr[WS(rs, 38)]; Chris@82: T3d = ci[WS(rs, 38)]; Chris@82: T39 = W[74]; Chris@82: T3b = T39 * T3a; Chris@82: T8K = T39 * T3d; Chris@82: T3c = W[75]; Chris@82: T3e = FMA(T3c, T3d, T3b); Chris@82: T8L = FNMS(T3c, T3a, T8K); Chris@82: } Chris@82: { Chris@82: E T3h, T3k, T3i, T8P, T3g, T3j; Chris@82: T3h = cr[WS(rs, 54)]; Chris@82: T3k = ci[WS(rs, 54)]; Chris@82: T3g = W[106]; Chris@82: T3i = T3g * T3h; Chris@82: T8P = T3g * T3k; Chris@82: T3j = W[107]; Chris@82: T3l = FMA(T3j, T3k, T3i); Chris@82: T8Q = FNMS(T3j, T3h, T8P); Chris@82: } Chris@82: { Chris@82: E T3f, T3s, Tfk, Tfl; Chris@82: T3f = T38 + T3e; Chris@82: T3s = T3l + T3r; Chris@82: T3t = T3f + T3s; Chris@82: Tfh = T3f - T3s; Chris@82: Tfk = T8Q + T8S; Chris@82: Tfl = T8J + T8L; Chris@82: Tfm = Tfk - Tfl; Chris@82: ThC = Tfl + Tfk; Chris@82: { Chris@82: E T8N, T94, T8U, T95; Chris@82: { Chris@82: E T8H, T8M, T8O, T8T; Chris@82: T8H = T38 - T3e; Chris@82: T8M = T8J - T8L; Chris@82: T8N = T8H - T8M; Chris@82: T94 = T8H + T8M; Chris@82: T8O = T3l - T3r; Chris@82: T8T = T8Q - T8S; Chris@82: T8U = T8O + T8T; Chris@82: T95 = T8T - T8O; Chris@82: } Chris@82: T8V = T8N + T8U; Chris@82: TcY = T8U - T8N; Chris@82: T96 = T94 + T95; Chris@82: TcV = T94 - T95; Chris@82: } Chris@82: } Chris@82: } Chris@82: { Chris@82: E T43, T9k, T4m, T9t, T49, T9m, T4g, T9r; Chris@82: { Chris@82: E T3Z, T42, T40, T9j, T3Y, T41; Chris@82: T3Z = cr[WS(rs, 9)]; Chris@82: T42 = ci[WS(rs, 9)]; Chris@82: T3Y = W[16]; Chris@82: T40 = T3Y * T3Z; Chris@82: T9j = T3Y * T42; Chris@82: T41 = W[17]; Chris@82: T43 = FMA(T41, T42, T40); Chris@82: T9k = FNMS(T41, T3Z, T9j); Chris@82: } Chris@82: { Chris@82: E T4i, T4l, T4j, T9s, T4h, T4k; Chris@82: T4i = cr[WS(rs, 25)]; Chris@82: T4l = ci[WS(rs, 25)]; Chris@82: T4h = W[48]; Chris@82: T4j = T4h * T4i; Chris@82: T9s = T4h * T4l; Chris@82: T4k = W[49]; Chris@82: T4m = FMA(T4k, T4l, T4j); Chris@82: T9t = FNMS(T4k, T4i, T9s); Chris@82: } Chris@82: { Chris@82: E T45, T48, T46, T9l, T44, T47; Chris@82: T45 = cr[WS(rs, 41)]; Chris@82: T48 = ci[WS(rs, 41)]; Chris@82: T44 = W[80]; Chris@82: T46 = T44 * T45; Chris@82: T9l = T44 * T48; Chris@82: T47 = W[81]; Chris@82: T49 = FMA(T47, T48, T46); Chris@82: T9m = FNMS(T47, T45, T9l); Chris@82: } Chris@82: { Chris@82: E T4c, T4f, T4d, T9q, T4b, T4e; Chris@82: T4c = cr[WS(rs, 57)]; Chris@82: T4f = ci[WS(rs, 57)]; Chris@82: T4b = W[112]; Chris@82: T4d = T4b * T4c; Chris@82: T9q = T4b * T4f; Chris@82: T4e = W[113]; Chris@82: T4g = FMA(T4e, T4f, T4d); Chris@82: T9r = FNMS(T4e, T4c, T9q); Chris@82: } Chris@82: { Chris@82: E T4a, T4n, Tfs, Tft; Chris@82: T4a = T43 + T49; Chris@82: T4n = T4g + T4m; Chris@82: T4o = T4a + T4n; Chris@82: TfL = T4a - T4n; Chris@82: Tfs = T9r + T9t; Chris@82: Tft = T9k + T9m; Chris@82: Tfu = Tfs - Tft; Chris@82: ThN = Tft + Tfs; Chris@82: { Chris@82: E T9o, Tab, T9v, Tac; Chris@82: { Chris@82: E T9i, T9n, T9p, T9u; Chris@82: T9i = T43 - T49; Chris@82: T9n = T9k - T9m; Chris@82: T9o = T9i - T9n; Chris@82: Tab = T9i + T9n; Chris@82: T9p = T4g - T4m; Chris@82: T9u = T9r - T9t; Chris@82: T9v = T9p + T9u; Chris@82: Tac = T9u - T9p; Chris@82: } Chris@82: T9w = T9o + T9v; Chris@82: Tdf = T9v - T9o; Chris@82: Tad = Tab + Tac; Chris@82: Td4 = Tab - Tac; Chris@82: } Chris@82: } Chris@82: } Chris@82: { Chris@82: E T5Q, Tat, T69, TaC, T5W, Tav, T63, TaA; Chris@82: { Chris@82: E T5M, T5P, T5N, Tas, T5L, T5O; Chris@82: T5M = cr[WS(rs, 7)]; Chris@82: T5P = ci[WS(rs, 7)]; Chris@82: T5L = W[12]; Chris@82: T5N = T5L * T5M; Chris@82: Tas = T5L * T5P; Chris@82: T5O = W[13]; Chris@82: T5Q = FMA(T5O, T5P, T5N); Chris@82: Tat = FNMS(T5O, T5M, Tas); Chris@82: } Chris@82: { Chris@82: E T65, T68, T66, TaB, T64, T67; Chris@82: T65 = cr[WS(rs, 23)]; Chris@82: T68 = ci[WS(rs, 23)]; Chris@82: T64 = W[44]; Chris@82: T66 = T64 * T65; Chris@82: TaB = T64 * T68; Chris@82: T67 = W[45]; Chris@82: T69 = FMA(T67, T68, T66); Chris@82: TaC = FNMS(T67, T65, TaB); Chris@82: } Chris@82: { Chris@82: E T5S, T5V, T5T, Tau, T5R, T5U; Chris@82: T5S = cr[WS(rs, 39)]; Chris@82: T5V = ci[WS(rs, 39)]; Chris@82: T5R = W[76]; Chris@82: T5T = T5R * T5S; Chris@82: Tau = T5R * T5V; Chris@82: T5U = W[77]; Chris@82: T5W = FMA(T5U, T5V, T5T); Chris@82: Tav = FNMS(T5U, T5S, Tau); Chris@82: } Chris@82: { Chris@82: E T5Z, T62, T60, Taz, T5Y, T61; Chris@82: T5Z = cr[WS(rs, 55)]; Chris@82: T62 = ci[WS(rs, 55)]; Chris@82: T5Y = W[108]; Chris@82: T60 = T5Y * T5Z; Chris@82: Taz = T5Y * T62; Chris@82: T61 = W[109]; Chris@82: T63 = FMA(T61, T62, T60); Chris@82: TaA = FNMS(T61, T5Z, Taz); Chris@82: } Chris@82: { Chris@82: E T5X, T6a, TfT, TfU; Chris@82: T5X = T5Q + T5W; Chris@82: T6a = T63 + T69; Chris@82: T6b = T5X + T6a; Chris@82: Tg9 = T6a - T5X; Chris@82: TfT = TaA + TaC; Chris@82: TfU = Tat + Tav; Chris@82: TfV = TfT - TfU; Chris@82: ThY = TfU + TfT; Chris@82: { Chris@82: E Tax, Tbl, TaE, Tbk; Chris@82: { Chris@82: E Tar, Taw, Tay, TaD; Chris@82: Tar = T5Q - T5W; Chris@82: Taw = Tat - Tav; Chris@82: Tax = Tar - Taw; Chris@82: Tbl = Tar + Taw; Chris@82: Tay = T63 - T69; Chris@82: TaD = TaA - TaC; Chris@82: TaE = Tay + TaD; Chris@82: Tbk = Tay - TaD; Chris@82: } Chris@82: TaF = Tax + TaE; Chris@82: Tdy = TaE - Tax; Chris@82: Tbm = Tbk - Tbl; Chris@82: Tdn = Tbl + Tbk; Chris@82: } Chris@82: } Chris@82: } Chris@82: { Chris@82: E T4v, T9G, T4O, T9C, T4B, T9I, T4I, T9A; Chris@82: { Chris@82: E T4r, T4u, T4s, T9F, T4q, T4t; Chris@82: T4r = cr[WS(rs, 5)]; Chris@82: T4u = ci[WS(rs, 5)]; Chris@82: T4q = W[8]; Chris@82: T4s = T4q * T4r; Chris@82: T9F = T4q * T4u; Chris@82: T4t = W[9]; Chris@82: T4v = FMA(T4t, T4u, T4s); Chris@82: T9G = FNMS(T4t, T4r, T9F); Chris@82: } Chris@82: { Chris@82: E T4K, T4N, T4L, T9B, T4J, T4M; Chris@82: T4K = cr[WS(rs, 53)]; Chris@82: T4N = ci[WS(rs, 53)]; Chris@82: T4J = W[104]; Chris@82: T4L = T4J * T4K; Chris@82: T9B = T4J * T4N; Chris@82: T4M = W[105]; Chris@82: T4O = FMA(T4M, T4N, T4L); Chris@82: T9C = FNMS(T4M, T4K, T9B); Chris@82: } Chris@82: { Chris@82: E T4x, T4A, T4y, T9H, T4w, T4z; Chris@82: T4x = cr[WS(rs, 37)]; Chris@82: T4A = ci[WS(rs, 37)]; Chris@82: T4w = W[72]; Chris@82: T4y = T4w * T4x; Chris@82: T9H = T4w * T4A; Chris@82: T4z = W[73]; Chris@82: T4B = FMA(T4z, T4A, T4y); Chris@82: T9I = FNMS(T4z, T4x, T9H); Chris@82: } Chris@82: { Chris@82: E T4E, T4H, T4F, T9z, T4D, T4G; Chris@82: T4E = cr[WS(rs, 21)]; Chris@82: T4H = ci[WS(rs, 21)]; Chris@82: T4D = W[40]; Chris@82: T4F = T4D * T4E; Chris@82: T9z = T4D * T4H; Chris@82: T4G = W[41]; Chris@82: T4I = FMA(T4G, T4H, T4F); Chris@82: T9A = FNMS(T4G, T4E, T9z); Chris@82: } Chris@82: { Chris@82: E T4C, T4P, Tfw, Tfx, Tfy, Tfz; Chris@82: T4C = T4v + T4B; Chris@82: T4P = T4I + T4O; Chris@82: Tfw = T4C - T4P; Chris@82: Tfx = T9G + T9I; Chris@82: Tfy = T9A + T9C; Chris@82: Tfz = Tfx - Tfy; Chris@82: T4Q = T4C + T4P; Chris@82: ThJ = Tfx + Tfy; Chris@82: TfA = Tfw - Tfz; Chris@82: TfN = Tfw + Tfz; Chris@82: } Chris@82: { Chris@82: E T9E, Td6, T9L, Td7; Chris@82: { Chris@82: E T9y, T9D, T9J, T9K; Chris@82: T9y = T4v - T4B; Chris@82: T9D = T9A - T9C; Chris@82: T9E = T9y - T9D; Chris@82: Td6 = T9y + T9D; Chris@82: T9J = T9G - T9I; Chris@82: T9K = T4I - T4O; Chris@82: T9L = T9J + T9K; Chris@82: Td7 = T9J - T9K; Chris@82: } Chris@82: T9M = FNMS(KP414213562, T9L, T9E); Chris@82: Tdh = FNMS(KP414213562, Td6, Td7); Chris@82: Taf = FMA(KP414213562, T9E, T9L); Chris@82: Td8 = FMA(KP414213562, Td7, Td6); Chris@82: } Chris@82: } Chris@82: { Chris@82: E T4W, T9V, T5f, T9R, T52, T9X, T59, T9P; Chris@82: { Chris@82: E T4S, T4V, T4T, T9U, T4R, T4U; Chris@82: T4S = cr[WS(rs, 61)]; Chris@82: T4V = ci[WS(rs, 61)]; Chris@82: T4R = W[120]; Chris@82: T4T = T4R * T4S; Chris@82: T9U = T4R * T4V; Chris@82: T4U = W[121]; Chris@82: T4W = FMA(T4U, T4V, T4T); Chris@82: T9V = FNMS(T4U, T4S, T9U); Chris@82: } Chris@82: { Chris@82: E T5b, T5e, T5c, T9Q, T5a, T5d; Chris@82: T5b = cr[WS(rs, 45)]; Chris@82: T5e = ci[WS(rs, 45)]; Chris@82: T5a = W[88]; Chris@82: T5c = T5a * T5b; Chris@82: T9Q = T5a * T5e; Chris@82: T5d = W[89]; Chris@82: T5f = FMA(T5d, T5e, T5c); Chris@82: T9R = FNMS(T5d, T5b, T9Q); Chris@82: } Chris@82: { Chris@82: E T4Y, T51, T4Z, T9W, T4X, T50; Chris@82: T4Y = cr[WS(rs, 29)]; Chris@82: T51 = ci[WS(rs, 29)]; Chris@82: T4X = W[56]; Chris@82: T4Z = T4X * T4Y; Chris@82: T9W = T4X * T51; Chris@82: T50 = W[57]; Chris@82: T52 = FMA(T50, T51, T4Z); Chris@82: T9X = FNMS(T50, T4Y, T9W); Chris@82: } Chris@82: { Chris@82: E T55, T58, T56, T9O, T54, T57; Chris@82: T55 = cr[WS(rs, 13)]; Chris@82: T58 = ci[WS(rs, 13)]; Chris@82: T54 = W[24]; Chris@82: T56 = T54 * T55; Chris@82: T9O = T54 * T58; Chris@82: T57 = W[25]; Chris@82: T59 = FMA(T57, T58, T56); Chris@82: T9P = FNMS(T57, T55, T9O); Chris@82: } Chris@82: { Chris@82: E T53, T5g, TfB, TfC, TfD, TfE; Chris@82: T53 = T4W + T52; Chris@82: T5g = T59 + T5f; Chris@82: TfB = T53 - T5g; Chris@82: TfC = T9V + T9X; Chris@82: TfD = T9P + T9R; Chris@82: TfE = TfC - TfD; Chris@82: T5h = T53 + T5g; Chris@82: ThI = TfC + TfD; Chris@82: TfF = TfB + TfE; Chris@82: TfO = TfE - TfB; Chris@82: } Chris@82: { Chris@82: E T9T, Td9, Ta0, Tda; Chris@82: { Chris@82: E T9N, T9S, T9Y, T9Z; Chris@82: T9N = T4W - T52; Chris@82: T9S = T9P - T9R; Chris@82: T9T = T9N - T9S; Chris@82: Td9 = T9N + T9S; Chris@82: T9Y = T9V - T9X; Chris@82: T9Z = T59 - T5f; Chris@82: Ta0 = T9Y + T9Z; Chris@82: Tda = T9Y - T9Z; Chris@82: } Chris@82: Ta1 = FMA(KP414213562, Ta0, T9T); Chris@82: Tdi = FMA(KP414213562, Td9, Tda); Chris@82: Tag = FNMS(KP414213562, T9T, Ta0); Chris@82: Tdb = FNMS(KP414213562, Tda, Td9); Chris@82: } Chris@82: } Chris@82: { Chris@82: E T6i, TaQ, T6B, TaL, T6o, TaS, T6v, TaJ; Chris@82: { Chris@82: E T6e, T6h, T6f, TaP, T6d, T6g; Chris@82: T6e = cr[WS(rs, 3)]; Chris@82: T6h = ci[WS(rs, 3)]; Chris@82: T6d = W[4]; Chris@82: T6f = T6d * T6e; Chris@82: TaP = T6d * T6h; Chris@82: T6g = W[5]; Chris@82: T6i = FMA(T6g, T6h, T6f); Chris@82: TaQ = FNMS(T6g, T6e, TaP); Chris@82: } Chris@82: { Chris@82: E T6x, T6A, T6y, TaK, T6w, T6z; Chris@82: T6x = cr[WS(rs, 51)]; Chris@82: T6A = ci[WS(rs, 51)]; Chris@82: T6w = W[100]; Chris@82: T6y = T6w * T6x; Chris@82: TaK = T6w * T6A; Chris@82: T6z = W[101]; Chris@82: T6B = FMA(T6z, T6A, T6y); Chris@82: TaL = FNMS(T6z, T6x, TaK); Chris@82: } Chris@82: { Chris@82: E T6k, T6n, T6l, TaR, T6j, T6m; Chris@82: T6k = cr[WS(rs, 35)]; Chris@82: T6n = ci[WS(rs, 35)]; Chris@82: T6j = W[68]; Chris@82: T6l = T6j * T6k; Chris@82: TaR = T6j * T6n; Chris@82: T6m = W[69]; Chris@82: T6o = FMA(T6m, T6n, T6l); Chris@82: TaS = FNMS(T6m, T6k, TaR); Chris@82: } Chris@82: { Chris@82: E T6r, T6u, T6s, TaI, T6q, T6t; Chris@82: T6r = cr[WS(rs, 19)]; Chris@82: T6u = ci[WS(rs, 19)]; Chris@82: T6q = W[36]; Chris@82: T6s = T6q * T6r; Chris@82: TaI = T6q * T6u; Chris@82: T6t = W[37]; Chris@82: T6v = FMA(T6t, T6u, T6s); Chris@82: TaJ = FNMS(T6t, T6r, TaI); Chris@82: } Chris@82: { Chris@82: E T6p, T6C, TfX, TfY, TfZ, Tg0; Chris@82: T6p = T6i + T6o; Chris@82: T6C = T6v + T6B; Chris@82: TfX = T6p - T6C; Chris@82: TfY = TaQ + TaS; Chris@82: TfZ = TaJ + TaL; Chris@82: Tg0 = TfY - TfZ; Chris@82: T6D = T6p + T6C; Chris@82: ThU = TfY + TfZ; Chris@82: Tg1 = TfX - Tg0; Chris@82: Tgf = TfX + Tg0; Chris@82: } Chris@82: { Chris@82: E TaN, Tdp, TaU, Tdq; Chris@82: { Chris@82: E TaH, TaM, TaO, TaT; Chris@82: TaH = T6i - T6o; Chris@82: TaM = TaJ - TaL; Chris@82: TaN = TaH - TaM; Chris@82: Tdp = TaH + TaM; Chris@82: TaO = T6B - T6v; Chris@82: TaT = TaQ - TaS; Chris@82: TaU = TaO - TaT; Chris@82: Tdq = TaT + TaO; Chris@82: } Chris@82: TaV = FMA(KP414213562, TaU, TaN); Chris@82: TdA = FNMS(KP414213562, Tdp, Tdq); Chris@82: Tbo = FNMS(KP414213562, TaN, TaU); Chris@82: Tdr = FMA(KP414213562, Tdq, Tdp); Chris@82: } Chris@82: } Chris@82: { Chris@82: E T6J, Tb5, T72, Tb0, T6P, Tb7, T6W, TaY; Chris@82: { Chris@82: E T6F, T6I, T6G, Tb4, T6E, T6H; Chris@82: T6F = cr[WS(rs, 59)]; Chris@82: T6I = ci[WS(rs, 59)]; Chris@82: T6E = W[116]; Chris@82: T6G = T6E * T6F; Chris@82: Tb4 = T6E * T6I; Chris@82: T6H = W[117]; Chris@82: T6J = FMA(T6H, T6I, T6G); Chris@82: Tb5 = FNMS(T6H, T6F, Tb4); Chris@82: } Chris@82: { Chris@82: E T6Y, T71, T6Z, TaZ, T6X, T70; Chris@82: T6Y = cr[WS(rs, 43)]; Chris@82: T71 = ci[WS(rs, 43)]; Chris@82: T6X = W[84]; Chris@82: T6Z = T6X * T6Y; Chris@82: TaZ = T6X * T71; Chris@82: T70 = W[85]; Chris@82: T72 = FMA(T70, T71, T6Z); Chris@82: Tb0 = FNMS(T70, T6Y, TaZ); Chris@82: } Chris@82: { Chris@82: E T6L, T6O, T6M, Tb6, T6K, T6N; Chris@82: T6L = cr[WS(rs, 27)]; Chris@82: T6O = ci[WS(rs, 27)]; Chris@82: T6K = W[52]; Chris@82: T6M = T6K * T6L; Chris@82: Tb6 = T6K * T6O; Chris@82: T6N = W[53]; Chris@82: T6P = FMA(T6N, T6O, T6M); Chris@82: Tb7 = FNMS(T6N, T6L, Tb6); Chris@82: } Chris@82: { Chris@82: E T6S, T6V, T6T, TaX, T6R, T6U; Chris@82: T6S = cr[WS(rs, 11)]; Chris@82: T6V = ci[WS(rs, 11)]; Chris@82: T6R = W[20]; Chris@82: T6T = T6R * T6S; Chris@82: TaX = T6R * T6V; Chris@82: T6U = W[21]; Chris@82: T6W = FMA(T6U, T6V, T6T); Chris@82: TaY = FNMS(T6U, T6S, TaX); Chris@82: } Chris@82: { Chris@82: E T6Q, T73, Tg2, Tg3, Tg4, Tg5; Chris@82: T6Q = T6J + T6P; Chris@82: T73 = T6W + T72; Chris@82: Tg2 = T6Q - T73; Chris@82: Tg3 = Tb5 + Tb7; Chris@82: Tg4 = TaY + Tb0; Chris@82: Tg5 = Tg3 - Tg4; Chris@82: T74 = T6Q + T73; Chris@82: ThT = Tg3 + Tg4; Chris@82: Tg6 = Tg2 + Tg5; Chris@82: Tge = Tg2 - Tg5; Chris@82: } Chris@82: { Chris@82: E Tb2, Tds, Tb9, Tdt; Chris@82: { Chris@82: E TaW, Tb1, Tb3, Tb8; Chris@82: TaW = T6J - T6P; Chris@82: Tb1 = TaY - Tb0; Chris@82: Tb2 = TaW - Tb1; Chris@82: Tds = TaW + Tb1; Chris@82: Tb3 = T72 - T6W; Chris@82: Tb8 = Tb5 - Tb7; Chris@82: Tb9 = Tb3 - Tb8; Chris@82: Tdt = Tb8 + Tb3; Chris@82: } Chris@82: Tba = FNMS(KP414213562, Tb9, Tb2); Chris@82: TdB = FMA(KP414213562, Tds, Tdt); Chris@82: Tbp = FMA(KP414213562, Tb2, Tb9); Chris@82: Tdu = FNMS(KP414213562, Tdt, Tds); Chris@82: } Chris@82: } Chris@82: { Chris@82: E T1I, Tio, T3v, Tj1, TiX, Tj2, Tir, TiN, T76, TiJ, TiC, TiG, T5j, TiK, Tix; Chris@82: E TiF; Chris@82: { Chris@82: E TO, T1H, Tip, Tiq; Chris@82: TO = Tm + TN; Chris@82: T1H = T1f + T1G; Chris@82: T1I = TO + T1H; Chris@82: Tio = TO - T1H; Chris@82: { Chris@82: E T2B, T3u, TiO, TiW; Chris@82: T2B = T29 + T2A; Chris@82: T3u = T32 + T3t; Chris@82: T3v = T2B + T3u; Chris@82: Tj1 = T2B - T3u; Chris@82: TiO = Ths + Thr; Chris@82: TiW = TiP + TiV; Chris@82: TiX = TiO + TiW; Chris@82: Tj2 = TiW - TiO; Chris@82: } Chris@82: Tip = ThB + ThC; Chris@82: Tiq = Thw + Thx; Chris@82: Tir = Tip - Tiq; Chris@82: TiN = Tiq + Tip; Chris@82: { Chris@82: E T6c, T75, Tiy, Tiz, TiA, TiB; Chris@82: T6c = T5K + T6b; Chris@82: T75 = T6D + T74; Chris@82: Tiy = T6c - T75; Chris@82: Tiz = ThX + ThY; Chris@82: TiA = ThU + ThT; Chris@82: TiB = Tiz - TiA; Chris@82: T76 = T6c + T75; Chris@82: TiJ = Tiz + TiA; Chris@82: TiC = Tiy - TiB; Chris@82: TiG = Tiy + TiB; Chris@82: } Chris@82: { Chris@82: E T4p, T5i, Tit, Tiu, Tiv, Tiw; Chris@82: T4p = T3X + T4o; Chris@82: T5i = T4Q + T5h; Chris@82: Tit = T4p - T5i; Chris@82: Tiu = ThM + ThN; Chris@82: Tiv = ThJ + ThI; Chris@82: Tiw = Tiu - Tiv; Chris@82: T5j = T4p + T5i; Chris@82: TiK = Tiu + Tiv; Chris@82: Tix = Tit + Tiw; Chris@82: TiF = Tit - Tiw; Chris@82: } Chris@82: } Chris@82: { Chris@82: E T3w, T77, Tj3, Tj4; Chris@82: T3w = T1I + T3v; Chris@82: T77 = T5j + T76; Chris@82: ci[WS(rs, 31)] = T3w - T77; Chris@82: cr[0] = T3w + T77; Chris@82: Tj3 = Tj1 + Tj2; Chris@82: Tj4 = TiC - Tix; Chris@82: cr[WS(rs, 56)] = FMS(KP707106781, Tj4, Tj3); Chris@82: ci[WS(rs, 39)] = FMA(KP707106781, Tj4, Tj3); Chris@82: } Chris@82: { Chris@82: E Tj5, Tj6, Tis, TiD; Chris@82: Tj5 = Tj2 - Tj1; Chris@82: Tj6 = TiG - TiF; Chris@82: cr[WS(rs, 40)] = FMS(KP707106781, Tj6, Tj5); Chris@82: ci[WS(rs, 55)] = FMA(KP707106781, Tj6, Tj5); Chris@82: Tis = Tio - Tir; Chris@82: TiD = Tix + TiC; Chris@82: ci[WS(rs, 23)] = FNMS(KP707106781, TiD, Tis); Chris@82: cr[WS(rs, 8)] = FMA(KP707106781, TiD, Tis); Chris@82: } Chris@82: { Chris@82: E TiE, TiH, TiM, TiY; Chris@82: TiE = Tio + Tir; Chris@82: TiH = TiF + TiG; Chris@82: cr[WS(rs, 24)] = FNMS(KP707106781, TiH, TiE); Chris@82: ci[WS(rs, 7)] = FMA(KP707106781, TiH, TiE); Chris@82: TiM = TiK + TiJ; Chris@82: TiY = TiN + TiX; Chris@82: cr[WS(rs, 32)] = TiM - TiY; Chris@82: ci[WS(rs, 63)] = TiM + TiY; Chris@82: } Chris@82: { Chris@82: E TiZ, Tj0, TiI, TiL; Chris@82: TiZ = T76 - T5j; Chris@82: Tj0 = TiX - TiN; Chris@82: cr[WS(rs, 48)] = TiZ - Tj0; Chris@82: ci[WS(rs, 47)] = TiZ + Tj0; Chris@82: TiI = T1I - T3v; Chris@82: TiL = TiJ - TiK; Chris@82: cr[WS(rs, 16)] = TiI - TiL; Chris@82: ci[WS(rs, 15)] = TiI + TiL; Chris@82: } Chris@82: } Chris@82: { Chris@82: E T99, Tk2, TbB, TjW, Taj, TbL, Tbv, TbF, Tce, Tcy, Tci, Tcu, Tc7, Tcx, Tch; Chris@82: E Tcr, TbZ, Tkg, Tcn, Tka, Tbs, TbM, Tbw, TbI, T80, Tk9, Tkf, Tby, TbS, TjV; Chris@82: E Tk1, Tck; Chris@82: { Chris@82: E T8z, Tbz, T98, TbA; Chris@82: { Chris@82: E T8n, T8y, T8W, T97; Chris@82: T8n = FMA(KP707106781, T8m, T87); Chris@82: T8y = FMA(KP707106781, T8x, T8u); Chris@82: T8z = FNMS(KP198912367, T8y, T8n); Chris@82: Tbz = FMA(KP198912367, T8n, T8y); Chris@82: T8W = FMA(KP707106781, T8V, T8G); Chris@82: T97 = FMA(KP707106781, T96, T93); Chris@82: T98 = FMA(KP198912367, T97, T8W); Chris@82: TbA = FNMS(KP198912367, T8W, T97); Chris@82: } Chris@82: T99 = T8z + T98; Chris@82: Tk2 = T98 - T8z; Chris@82: TbB = Tbz - TbA; Chris@82: TjW = Tbz + TbA; Chris@82: } Chris@82: { Chris@82: E Ta3, TbD, Tai, TbE; Chris@82: { Chris@82: E T9x, Ta2, Tae, Tah; Chris@82: T9x = FMA(KP707106781, T9w, T9h); Chris@82: Ta2 = T9M + Ta1; Chris@82: Ta3 = FMA(KP923879532, Ta2, T9x); Chris@82: TbD = FNMS(KP923879532, Ta2, T9x); Chris@82: Tae = FMA(KP707106781, Tad, Taa); Chris@82: Tah = Taf + Tag; Chris@82: Tai = FMA(KP923879532, Tah, Tae); Chris@82: TbE = FNMS(KP923879532, Tah, Tae); Chris@82: } Chris@82: Taj = FNMS(KP098491403, Tai, Ta3); Chris@82: TbL = FNMS(KP820678790, TbD, TbE); Chris@82: Tbv = FMA(KP098491403, Ta3, Tai); Chris@82: TbF = FMA(KP820678790, TbE, TbD); Chris@82: } Chris@82: { Chris@82: E Tca, Tcs, Tcd, Tct; Chris@82: { Chris@82: E Tc8, Tc9, Tcb, Tcc; Chris@82: Tc8 = FNMS(KP707106781, TaF, Taq); Chris@82: Tc9 = Tbp - Tbo; Chris@82: Tca = FNMS(KP923879532, Tc9, Tc8); Chris@82: Tcs = FMA(KP923879532, Tc9, Tc8); Chris@82: Tcb = FNMS(KP707106781, Tbm, Tbj); Chris@82: Tcc = Tba - TaV; Chris@82: Tcd = FMA(KP923879532, Tcc, Tcb); Chris@82: Tct = FNMS(KP923879532, Tcc, Tcb); Chris@82: } Chris@82: Tce = FNMS(KP534511135, Tcd, Tca); Chris@82: Tcy = FNMS(KP303346683, Tcs, Tct); Chris@82: Tci = FMA(KP534511135, Tca, Tcd); Chris@82: Tcu = FMA(KP303346683, Tct, Tcs); Chris@82: } Chris@82: { Chris@82: E Tc3, Tcp, Tc6, Tcq; Chris@82: { Chris@82: E Tc1, Tc2, Tc4, Tc5; Chris@82: Tc1 = FNMS(KP707106781, T9w, T9h); Chris@82: Tc2 = Taf - Tag; Chris@82: Tc3 = FNMS(KP923879532, Tc2, Tc1); Chris@82: Tcp = FMA(KP923879532, Tc2, Tc1); Chris@82: Tc4 = FNMS(KP707106781, Tad, Taa); Chris@82: Tc5 = Ta1 - T9M; Chris@82: Tc6 = FNMS(KP923879532, Tc5, Tc4); Chris@82: Tcq = FMA(KP923879532, Tc5, Tc4); Chris@82: } Chris@82: Tc7 = FNMS(KP534511135, Tc6, Tc3); Chris@82: Tcx = FNMS(KP303346683, Tcp, Tcq); Chris@82: Tch = FMA(KP534511135, Tc3, Tc6); Chris@82: Tcr = FMA(KP303346683, Tcq, Tcp); Chris@82: } Chris@82: { Chris@82: E TbV, Tcm, TbY, Tcl; Chris@82: { Chris@82: E TbT, TbU, TbW, TbX; Chris@82: TbT = FNMS(KP707106781, T96, T93); Chris@82: TbU = FNMS(KP707106781, T8V, T8G); Chris@82: TbV = FMA(KP668178637, TbU, TbT); Chris@82: Tcm = FNMS(KP668178637, TbT, TbU); Chris@82: TbW = FNMS(KP707106781, T8x, T8u); Chris@82: TbX = FNMS(KP707106781, T8m, T87); Chris@82: TbY = FNMS(KP668178637, TbX, TbW); Chris@82: Tcl = FMA(KP668178637, TbW, TbX); Chris@82: } Chris@82: TbZ = TbV - TbY; Chris@82: Tkg = Tcl - Tcm; Chris@82: Tcn = Tcl + Tcm; Chris@82: Tka = TbY + TbV; Chris@82: } Chris@82: { Chris@82: E Tbc, TbG, Tbr, TbH; Chris@82: { Chris@82: E TaG, Tbb, Tbn, Tbq; Chris@82: TaG = FMA(KP707106781, TaF, Taq); Chris@82: Tbb = TaV + Tba; Chris@82: Tbc = FMA(KP923879532, Tbb, TaG); Chris@82: TbG = FNMS(KP923879532, Tbb, TaG); Chris@82: Tbn = FMA(KP707106781, Tbm, Tbj); Chris@82: Tbq = Tbo + Tbp; Chris@82: Tbr = FMA(KP923879532, Tbq, Tbn); Chris@82: TbH = FNMS(KP923879532, Tbq, Tbn); Chris@82: } Chris@82: Tbs = FNMS(KP098491403, Tbr, Tbc); Chris@82: TbM = FNMS(KP820678790, TbG, TbH); Chris@82: Tbw = FMA(KP098491403, Tbc, Tbr); Chris@82: TbI = FMA(KP820678790, TbH, TbG); Chris@82: } Chris@82: { Chris@82: E T7u, TbO, TjT, Tk7, T7Z, Tk8, TbR, TjU, T7t, TjS; Chris@82: T7t = T7l + T7s; Chris@82: T7u = FMA(KP707106781, T7t, T7e); Chris@82: TbO = FNMS(KP707106781, T7t, T7e); Chris@82: TjS = TcB - TcC; Chris@82: TjT = FMA(KP707106781, TjS, TjR); Chris@82: Tk7 = FNMS(KP707106781, TjS, TjR); Chris@82: { Chris@82: E T7J, T7Y, TbP, TbQ; Chris@82: T7J = FNMS(KP414213562, T7I, T7B); Chris@82: T7Y = FMA(KP414213562, T7X, T7Q); Chris@82: T7Z = T7J + T7Y; Chris@82: Tk8 = T7Y - T7J; Chris@82: TbP = FMA(KP414213562, T7B, T7I); Chris@82: TbQ = FNMS(KP414213562, T7Q, T7X); Chris@82: TbR = TbP - TbQ; Chris@82: TjU = TbP + TbQ; Chris@82: } Chris@82: T80 = FMA(KP923879532, T7Z, T7u); Chris@82: Tk9 = FMA(KP923879532, Tk8, Tk7); Chris@82: Tkf = FNMS(KP923879532, Tk8, Tk7); Chris@82: Tby = FNMS(KP923879532, T7Z, T7u); Chris@82: TbS = FNMS(KP923879532, TbR, TbO); Chris@82: TjV = FMA(KP923879532, TjU, TjT); Chris@82: Tk1 = FNMS(KP923879532, TjU, TjT); Chris@82: Tck = FMA(KP923879532, TbR, TbO); Chris@82: } Chris@82: { Chris@82: E T9a, Tbt, TbK, TbN; Chris@82: T9a = FMA(KP980785280, T99, T80); Chris@82: Tbt = Taj + Tbs; Chris@82: cr[WS(rs, 31)] = FNMS(KP995184726, Tbt, T9a); Chris@82: ci[0] = FMA(KP995184726, Tbt, T9a); Chris@82: TbK = FNMS(KP980785280, TbB, Tby); Chris@82: TbN = TbL + TbM; Chris@82: cr[WS(rs, 23)] = FMA(KP773010453, TbN, TbK); Chris@82: ci[WS(rs, 8)] = FNMS(KP773010453, TbN, TbK); Chris@82: } Chris@82: { Chris@82: E Tkb, Tkc, Tkj, Tkk; Chris@82: Tkb = FMA(KP831469612, Tka, Tk9); Chris@82: Tkc = Tcx - Tcy; Chris@82: cr[WS(rs, 35)] = FMS(KP956940335, Tkc, Tkb); Chris@82: ci[WS(rs, 60)] = FMA(KP956940335, Tkc, Tkb); Chris@82: Tkj = FNMS(KP831469612, Tkg, Tkf); Chris@82: Tkk = Tce - Tc7; Chris@82: cr[WS(rs, 43)] = FMS(KP881921264, Tkk, Tkj); Chris@82: ci[WS(rs, 52)] = FMA(KP881921264, Tkk, Tkj); Chris@82: } Chris@82: { Chris@82: E Tbu, Tbx, TbC, TbJ; Chris@82: Tbu = FNMS(KP980785280, T99, T80); Chris@82: Tbx = Tbv + Tbw; Chris@82: ci[WS(rs, 16)] = FNMS(KP995184726, Tbx, Tbu); Chris@82: cr[WS(rs, 15)] = FMA(KP995184726, Tbx, Tbu); Chris@82: TbC = FMA(KP980785280, TbB, Tby); Chris@82: TbJ = TbF + TbI; Chris@82: ci[WS(rs, 24)] = FNMS(KP773010453, TbJ, TbC); Chris@82: cr[WS(rs, 7)] = FMA(KP773010453, TbJ, TbC); Chris@82: } Chris@82: { Chris@82: E Tkd, Tke, Tkh, Tki; Chris@82: Tkd = FNMS(KP831469612, Tka, Tk9); Chris@82: Tke = Tcu - Tcr; Chris@82: cr[WS(rs, 51)] = FMS(KP956940335, Tke, Tkd); Chris@82: ci[WS(rs, 44)] = FMA(KP956940335, Tke, Tkd); Chris@82: Tkh = FMA(KP831469612, Tkg, Tkf); Chris@82: Tki = Tci - Tch; Chris@82: cr[WS(rs, 59)] = FMS(KP881921264, Tki, Tkh); Chris@82: ci[WS(rs, 36)] = FMA(KP881921264, Tki, Tkh); Chris@82: } Chris@82: { Chris@82: E Tc0, Tcf, Tcw, Tcz; Chris@82: Tc0 = FMA(KP831469612, TbZ, TbS); Chris@82: Tcf = Tc7 + Tce; Chris@82: cr[WS(rs, 27)] = FNMS(KP881921264, Tcf, Tc0); Chris@82: ci[WS(rs, 4)] = FMA(KP881921264, Tcf, Tc0); Chris@82: Tcw = FNMS(KP831469612, Tcn, Tck); Chris@82: Tcz = Tcx + Tcy; Chris@82: cr[WS(rs, 19)] = FMA(KP956940335, Tcz, Tcw); Chris@82: ci[WS(rs, 12)] = FNMS(KP956940335, Tcz, Tcw); Chris@82: } Chris@82: { Chris@82: E TjX, TjY, Tk5, Tk6; Chris@82: TjX = FMA(KP980785280, TjW, TjV); Chris@82: TjY = Tbw - Tbv; Chris@82: cr[WS(rs, 63)] = FMS(KP995184726, TjY, TjX); Chris@82: ci[WS(rs, 32)] = FMA(KP995184726, TjY, TjX); Chris@82: Tk5 = FNMS(KP980785280, Tk2, Tk1); Chris@82: Tk6 = TbI - TbF; Chris@82: cr[WS(rs, 55)] = FMS(KP773010453, Tk6, Tk5); Chris@82: ci[WS(rs, 40)] = FMA(KP773010453, Tk6, Tk5); Chris@82: } Chris@82: { Chris@82: E Tcg, Tcj, Tco, Tcv; Chris@82: Tcg = FNMS(KP831469612, TbZ, TbS); Chris@82: Tcj = Tch + Tci; Chris@82: ci[WS(rs, 20)] = FNMS(KP881921264, Tcj, Tcg); Chris@82: cr[WS(rs, 11)] = FMA(KP881921264, Tcj, Tcg); Chris@82: Tco = FMA(KP831469612, Tcn, Tck); Chris@82: Tcv = Tcr + Tcu; Chris@82: ci[WS(rs, 28)] = FNMS(KP956940335, Tcv, Tco); Chris@82: cr[WS(rs, 3)] = FMA(KP956940335, Tcv, Tco); Chris@82: } Chris@82: { Chris@82: E TjZ, Tk0, Tk3, Tk4; Chris@82: TjZ = FNMS(KP980785280, TjW, TjV); Chris@82: Tk0 = Tbs - Taj; Chris@82: cr[WS(rs, 47)] = FMS(KP995184726, Tk0, TjZ); Chris@82: ci[WS(rs, 48)] = FMA(KP995184726, Tk0, TjZ); Chris@82: Tk3 = FMA(KP980785280, Tk2, Tk1); Chris@82: Tk4 = TbL - TbM; Chris@82: cr[WS(rs, 39)] = FMS(KP773010453, Tk4, Tk3); Chris@82: ci[WS(rs, 56)] = FMA(KP773010453, Tk4, Tk3); Chris@82: } Chris@82: } Chris@82: { Chris@82: E Thu, Ti8, Tj9, Tjf, ThF, Tjg, Tib, Tja, ThR, Til, Ti6, Tif, Ti2, Tim, Ti5; Chris@82: E Tii; Chris@82: { Chris@82: E Thq, Tht, Tj7, Tj8; Chris@82: Thq = Tm - TN; Chris@82: Tht = Thr - Ths; Chris@82: Thu = Thq - Tht; Chris@82: Ti8 = Thq + Tht; Chris@82: Tj7 = T1f - T1G; Chris@82: Tj8 = TiV - TiP; Chris@82: Tj9 = Tj7 + Tj8; Chris@82: Tjf = Tj8 - Tj7; Chris@82: } Chris@82: { Chris@82: E Thz, Ti9, ThE, Tia; Chris@82: { Chris@82: E Thv, Thy, ThA, ThD; Chris@82: Thv = T29 - T2A; Chris@82: Thy = Thw - Thx; Chris@82: Thz = Thv + Thy; Chris@82: Ti9 = Thv - Thy; Chris@82: ThA = T32 - T3t; Chris@82: ThD = ThB - ThC; Chris@82: ThE = ThA - ThD; Chris@82: Tia = ThA + ThD; Chris@82: } Chris@82: ThF = Thz + ThE; Chris@82: Tjg = Tia - Ti9; Chris@82: Tib = Ti9 + Tia; Chris@82: Tja = Thz - ThE; Chris@82: } Chris@82: { Chris@82: E ThL, Tid, ThQ, Tie; Chris@82: { Chris@82: E ThH, ThK, ThO, ThP; Chris@82: ThH = T3X - T4o; Chris@82: ThK = ThI - ThJ; Chris@82: ThL = ThH - ThK; Chris@82: Tid = ThH + ThK; Chris@82: ThO = ThM - ThN; Chris@82: ThP = T4Q - T5h; Chris@82: ThQ = ThO - ThP; Chris@82: Tie = ThO + ThP; Chris@82: } Chris@82: ThR = FMA(KP414213562, ThQ, ThL); Chris@82: Til = FMA(KP414213562, Tid, Tie); Chris@82: Ti6 = FNMS(KP414213562, ThL, ThQ); Chris@82: Tif = FNMS(KP414213562, Tie, Tid); Chris@82: } Chris@82: { Chris@82: E ThW, Tig, Ti1, Tih; Chris@82: { Chris@82: E ThS, ThV, ThZ, Ti0; Chris@82: ThS = T5K - T6b; Chris@82: ThV = ThT - ThU; Chris@82: ThW = ThS - ThV; Chris@82: Tig = ThS + ThV; Chris@82: ThZ = ThX - ThY; Chris@82: Ti0 = T74 - T6D; Chris@82: Ti1 = ThZ + Ti0; Chris@82: Tih = Ti0 - ThZ; Chris@82: } Chris@82: Ti2 = FNMS(KP414213562, Ti1, ThW); Chris@82: Tim = FMA(KP414213562, Tig, Tih); Chris@82: Ti5 = FMA(KP414213562, ThW, Ti1); Chris@82: Tii = FNMS(KP414213562, Tih, Tig); Chris@82: } Chris@82: { Chris@82: E ThG, Ti3, Tjh, Tji; Chris@82: ThG = FMA(KP707106781, ThF, Thu); Chris@82: Ti3 = ThR + Ti2; Chris@82: ci[WS(rs, 27)] = FNMS(KP923879532, Ti3, ThG); Chris@82: cr[WS(rs, 4)] = FMA(KP923879532, Ti3, ThG); Chris@82: Tjh = FMA(KP707106781, Tjg, Tjf); Chris@82: Tji = Ti6 + Ti5; Chris@82: cr[WS(rs, 36)] = FMS(KP923879532, Tji, Tjh); Chris@82: ci[WS(rs, 59)] = FMA(KP923879532, Tji, Tjh); Chris@82: } Chris@82: { Chris@82: E Tjj, Tjk, Ti4, Ti7; Chris@82: Tjj = FNMS(KP707106781, Tjg, Tjf); Chris@82: Tjk = Ti2 - ThR; Chris@82: cr[WS(rs, 52)] = FMS(KP923879532, Tjk, Tjj); Chris@82: ci[WS(rs, 43)] = FMA(KP923879532, Tjk, Tjj); Chris@82: Ti4 = FNMS(KP707106781, ThF, Thu); Chris@82: Ti7 = Ti5 - Ti6; Chris@82: cr[WS(rs, 20)] = FNMS(KP923879532, Ti7, Ti4); Chris@82: ci[WS(rs, 11)] = FMA(KP923879532, Ti7, Ti4); Chris@82: } Chris@82: { Chris@82: E Tic, Tij, Tjb, Tjc; Chris@82: Tic = FMA(KP707106781, Tib, Ti8); Chris@82: Tij = Tif + Tii; Chris@82: cr[WS(rs, 28)] = FNMS(KP923879532, Tij, Tic); Chris@82: ci[WS(rs, 3)] = FMA(KP923879532, Tij, Tic); Chris@82: Tjb = FMA(KP707106781, Tja, Tj9); Chris@82: Tjc = Tim - Til; Chris@82: cr[WS(rs, 60)] = FMS(KP923879532, Tjc, Tjb); Chris@82: ci[WS(rs, 35)] = FMA(KP923879532, Tjc, Tjb); Chris@82: } Chris@82: { Chris@82: E Tjd, Tje, Tik, Tin; Chris@82: Tjd = FNMS(KP707106781, Tja, Tj9); Chris@82: Tje = Tii - Tif; Chris@82: cr[WS(rs, 44)] = FMS(KP923879532, Tje, Tjd); Chris@82: ci[WS(rs, 51)] = FMA(KP923879532, Tje, Tjd); Chris@82: Tik = FNMS(KP707106781, Tib, Ti8); Chris@82: Tin = Til + Tim; Chris@82: ci[WS(rs, 19)] = FNMS(KP923879532, Tin, Tik); Chris@82: cr[WS(rs, 12)] = FMA(KP923879532, Tin, Tik); Chris@82: } Chris@82: } Chris@82: { Chris@82: E Tf2, TjJ, Tgo, TjD, TgI, Tjv, Tha, Tjp, Tfp, Tjw, Tgr, Tjq, Th4, Tho, Th7; Chris@82: E Thk, TfR, TgB, Tgl, Tgv, TgP, TjK, Thd, TjE, TgX, Thn, Th8, Thh, Tgi, TgC; Chris@82: E Tgm, Tgy; Chris@82: { Chris@82: E TeQ, TjB, Tf1, TjC, TeV, Tf0; Chris@82: TeQ = TeM + TeP; Chris@82: TjB = Tjm - Tjl; Chris@82: TeV = TeR - TeU; Chris@82: Tf0 = TeW + TeZ; Chris@82: Tf1 = TeV + Tf0; Chris@82: TjC = Tf0 - TeV; Chris@82: Tf2 = FNMS(KP707106781, Tf1, TeQ); Chris@82: TjJ = FNMS(KP707106781, TjC, TjB); Chris@82: Tgo = FMA(KP707106781, Tf1, TeQ); Chris@82: TjD = FMA(KP707106781, TjC, TjB); Chris@82: } Chris@82: { Chris@82: E TgE, Tjn, TgH, Tjo, TgF, TgG; Chris@82: TgE = TeM - TeP; Chris@82: Tjn = Tjl + Tjm; Chris@82: TgF = TeR + TeU; Chris@82: TgG = TeW - TeZ; Chris@82: TgH = TgF + TgG; Chris@82: Tjo = TgF - TgG; Chris@82: TgI = FMA(KP707106781, TgH, TgE); Chris@82: Tjv = FNMS(KP707106781, Tjo, Tjn); Chris@82: Tha = FNMS(KP707106781, TgH, TgE); Chris@82: Tjp = FMA(KP707106781, Tjo, Tjn); Chris@82: } Chris@82: { Chris@82: E Tfd, Tgp, Tfo, Tgq; Chris@82: { Chris@82: E Tf7, Tfc, Tfi, Tfn; Chris@82: Tf7 = Tf5 + Tf6; Chris@82: Tfc = Tf8 + Tfb; Chris@82: Tfd = FMA(KP414213562, Tfc, Tf7); Chris@82: Tgp = FNMS(KP414213562, Tf7, Tfc); Chris@82: Tfi = Tfg + Tfh; Chris@82: Tfn = Tfj + Tfm; Chris@82: Tfo = FNMS(KP414213562, Tfn, Tfi); Chris@82: Tgq = FMA(KP414213562, Tfi, Tfn); Chris@82: } Chris@82: Tfp = Tfd - Tfo; Chris@82: Tjw = Tgq - Tgp; Chris@82: Tgr = Tgp + Tgq; Chris@82: Tjq = Tfd + Tfo; Chris@82: } Chris@82: { Chris@82: E Th0, Thi, Th3, Thj; Chris@82: { Chris@82: E TgY, TgZ, Th1, Th2; Chris@82: TgY = TfS - TfV; Chris@82: TgZ = Tgf + Tge; Chris@82: Th0 = FMA(KP707106781, TgZ, TgY); Chris@82: Thi = FNMS(KP707106781, TgZ, TgY); Chris@82: Th1 = Tgc + Tg9; Chris@82: Th2 = Tg6 - Tg1; Chris@82: Th3 = FMA(KP707106781, Th2, Th1); Chris@82: Thj = FNMS(KP707106781, Th2, Th1); Chris@82: } Chris@82: Th4 = FNMS(KP198912367, Th3, Th0); Chris@82: Tho = FNMS(KP668178637, Thi, Thj); Chris@82: Th7 = FMA(KP198912367, Th0, Th3); Chris@82: Thk = FMA(KP668178637, Thj, Thi); Chris@82: } Chris@82: { Chris@82: E TfH, Tgt, TfQ, Tgu; Chris@82: { Chris@82: E Tfv, TfG, TfM, TfP; Chris@82: Tfv = Tfr + Tfu; Chris@82: TfG = TfA + TfF; Chris@82: TfH = FNMS(KP707106781, TfG, Tfv); Chris@82: Tgt = FMA(KP707106781, TfG, Tfv); Chris@82: TfM = TfK + TfL; Chris@82: TfP = TfN + TfO; Chris@82: TfQ = FNMS(KP707106781, TfP, TfM); Chris@82: Tgu = FMA(KP707106781, TfP, TfM); Chris@82: } Chris@82: TfR = FMA(KP668178637, TfQ, TfH); Chris@82: TgB = FMA(KP198912367, Tgt, Tgu); Chris@82: Tgl = FNMS(KP668178637, TfH, TfQ); Chris@82: Tgv = FNMS(KP198912367, Tgu, Tgt); Chris@82: } Chris@82: { Chris@82: E TgL, Thc, TgO, Thb; Chris@82: { Chris@82: E TgJ, TgK, TgM, TgN; Chris@82: TgJ = Tf8 - Tfb; Chris@82: TgK = Tf5 - Tf6; Chris@82: TgL = FMA(KP414213562, TgK, TgJ); Chris@82: Thc = FNMS(KP414213562, TgJ, TgK); Chris@82: TgM = Tfj - Tfm; Chris@82: TgN = Tfg - Tfh; Chris@82: TgO = FNMS(KP414213562, TgN, TgM); Chris@82: Thb = FMA(KP414213562, TgM, TgN); Chris@82: } Chris@82: TgP = TgL + TgO; Chris@82: TjK = TgL - TgO; Chris@82: Thd = Thb - Thc; Chris@82: TjE = Thc + Thb; Chris@82: } Chris@82: { Chris@82: E TgT, Thf, TgW, Thg; Chris@82: { Chris@82: E TgR, TgS, TgU, TgV; Chris@82: TgR = Tfr - Tfu; Chris@82: TgS = TfN - TfO; Chris@82: TgT = FMA(KP707106781, TgS, TgR); Chris@82: Thf = FNMS(KP707106781, TgS, TgR); Chris@82: TgU = TfK - TfL; Chris@82: TgV = TfF - TfA; Chris@82: TgW = FMA(KP707106781, TgV, TgU); Chris@82: Thg = FNMS(KP707106781, TgV, TgU); Chris@82: } Chris@82: TgX = FMA(KP198912367, TgW, TgT); Chris@82: Thn = FMA(KP668178637, Thf, Thg); Chris@82: Th8 = FNMS(KP198912367, TgT, TgW); Chris@82: Thh = FNMS(KP668178637, Thg, Thf); Chris@82: } Chris@82: { Chris@82: E Tg8, Tgw, Tgh, Tgx; Chris@82: { Chris@82: E TfW, Tg7, Tgd, Tgg; Chris@82: TfW = TfS + TfV; Chris@82: Tg7 = Tg1 + Tg6; Chris@82: Tg8 = FNMS(KP707106781, Tg7, TfW); Chris@82: Tgw = FMA(KP707106781, Tg7, TfW); Chris@82: Tgd = Tg9 - Tgc; Chris@82: Tgg = Tge - Tgf; Chris@82: Tgh = FNMS(KP707106781, Tgg, Tgd); Chris@82: Tgx = FMA(KP707106781, Tgg, Tgd); Chris@82: } Chris@82: Tgi = FMA(KP668178637, Tgh, Tg8); Chris@82: TgC = FMA(KP198912367, Tgw, Tgx); Chris@82: Tgm = FNMS(KP668178637, Tg8, Tgh); Chris@82: Tgy = FNMS(KP198912367, Tgx, Tgw); Chris@82: } Chris@82: { Chris@82: E Tfq, Tgj, TgA, TgD; Chris@82: Tfq = FMA(KP923879532, Tfp, Tf2); Chris@82: Tgj = TfR + Tgi; Chris@82: ci[WS(rs, 25)] = FNMS(KP831469612, Tgj, Tfq); Chris@82: cr[WS(rs, 6)] = FMA(KP831469612, Tgj, Tfq); Chris@82: TgA = FNMS(KP923879532, Tgr, Tgo); Chris@82: TgD = TgB + TgC; Chris@82: ci[WS(rs, 17)] = FNMS(KP980785280, TgD, TgA); Chris@82: cr[WS(rs, 14)] = FMA(KP980785280, TgD, TgA); Chris@82: } Chris@82: { Chris@82: E TjF, TjG, TjN, TjO; Chris@82: TjF = FMA(KP923879532, TjE, TjD); Chris@82: TjG = Th8 + Th7; Chris@82: cr[WS(rs, 34)] = FMS(KP980785280, TjG, TjF); Chris@82: ci[WS(rs, 61)] = FMA(KP980785280, TjG, TjF); Chris@82: TjN = FNMS(KP923879532, TjK, TjJ); Chris@82: TjO = Thk - Thh; Chris@82: cr[WS(rs, 42)] = FMS(KP831469612, TjO, TjN); Chris@82: ci[WS(rs, 53)] = FMA(KP831469612, TjO, TjN); Chris@82: } Chris@82: { Chris@82: E Tgk, Tgn, Tgs, Tgz; Chris@82: Tgk = FNMS(KP923879532, Tfp, Tf2); Chris@82: Tgn = Tgl + Tgm; Chris@82: cr[WS(rs, 22)] = FMA(KP831469612, Tgn, Tgk); Chris@82: ci[WS(rs, 9)] = FNMS(KP831469612, Tgn, Tgk); Chris@82: Tgs = FMA(KP923879532, Tgr, Tgo); Chris@82: Tgz = Tgv + Tgy; Chris@82: cr[WS(rs, 30)] = FNMS(KP980785280, Tgz, Tgs); Chris@82: ci[WS(rs, 1)] = FMA(KP980785280, Tgz, Tgs); Chris@82: } Chris@82: { Chris@82: E TjH, TjI, TjL, TjM; Chris@82: TjH = FNMS(KP923879532, TjE, TjD); Chris@82: TjI = Th4 - TgX; Chris@82: cr[WS(rs, 50)] = FMS(KP980785280, TjI, TjH); Chris@82: ci[WS(rs, 45)] = FMA(KP980785280, TjI, TjH); Chris@82: TjL = FMA(KP923879532, TjK, TjJ); Chris@82: TjM = Thn + Tho; Chris@82: cr[WS(rs, 58)] = -(FMA(KP831469612, TjM, TjL)); Chris@82: ci[WS(rs, 37)] = FNMS(KP831469612, TjM, TjL); Chris@82: } Chris@82: { Chris@82: E TgQ, Th5, Thm, Thp; Chris@82: TgQ = FMA(KP923879532, TgP, TgI); Chris@82: Th5 = TgX + Th4; Chris@82: ci[WS(rs, 29)] = FNMS(KP980785280, Th5, TgQ); Chris@82: cr[WS(rs, 2)] = FMA(KP980785280, Th5, TgQ); Chris@82: Thm = FNMS(KP923879532, Thd, Tha); Chris@82: Thp = Thn - Tho; Chris@82: ci[WS(rs, 21)] = FNMS(KP831469612, Thp, Thm); Chris@82: cr[WS(rs, 10)] = FMA(KP831469612, Thp, Thm); Chris@82: } Chris@82: { Chris@82: E Tjr, Tjs, Tjz, TjA; Chris@82: Tjr = FMA(KP923879532, Tjq, Tjp); Chris@82: Tjs = TgC - TgB; Chris@82: cr[WS(rs, 62)] = FMS(KP980785280, Tjs, Tjr); Chris@82: ci[WS(rs, 33)] = FMA(KP980785280, Tjs, Tjr); Chris@82: Tjz = FNMS(KP923879532, Tjw, Tjv); Chris@82: TjA = Tgi - TfR; Chris@82: cr[WS(rs, 54)] = FMS(KP831469612, TjA, Tjz); Chris@82: ci[WS(rs, 41)] = FMA(KP831469612, TjA, Tjz); Chris@82: } Chris@82: { Chris@82: E Th6, Th9, The, Thl; Chris@82: Th6 = FNMS(KP923879532, TgP, TgI); Chris@82: Th9 = Th7 - Th8; Chris@82: cr[WS(rs, 18)] = FNMS(KP980785280, Th9, Th6); Chris@82: ci[WS(rs, 13)] = FMA(KP980785280, Th9, Th6); Chris@82: The = FMA(KP923879532, Thd, Tha); Chris@82: Thl = Thh + Thk; Chris@82: cr[WS(rs, 26)] = FNMS(KP831469612, Thl, The); Chris@82: ci[WS(rs, 5)] = FMA(KP831469612, Thl, The); Chris@82: } Chris@82: { Chris@82: E Tjt, Tju, Tjx, Tjy; Chris@82: Tjt = FNMS(KP923879532, Tjq, Tjp); Chris@82: Tju = Tgy - Tgv; Chris@82: cr[WS(rs, 46)] = FMS(KP980785280, Tju, Tjt); Chris@82: ci[WS(rs, 49)] = FMA(KP980785280, Tju, Tjt); Chris@82: Tjx = FMA(KP923879532, Tjw, Tjv); Chris@82: Tjy = Tgl - Tgm; Chris@82: cr[WS(rs, 38)] = FMS(KP831469612, Tjy, Tjx); Chris@82: ci[WS(rs, 57)] = FMA(KP831469612, Tjy, Tjx); Chris@82: } Chris@82: } Chris@82: { Chris@82: E Td1, Tkw, TdN, Tkq, Tdl, TdX, TdI, TdR, Teq, TeK, Tet, TeG, Tej, TeJ, Teu; Chris@82: E TeD, Teb, TkK, Tez, TkE, TdE, TdY, TdH, TdU, TcM, TkD, TkJ, TdK, Te4, Tkp; Chris@82: E Tkv, Tew; Chris@82: { Chris@82: E TcT, TdM, Td0, TdL; Chris@82: { Chris@82: E TcP, TcS, TcW, TcZ; Chris@82: TcP = FMA(KP707106781, TcO, TcN); Chris@82: TcS = FMA(KP707106781, TcR, TcQ); Chris@82: TcT = FMA(KP198912367, TcS, TcP); Chris@82: TdM = FNMS(KP198912367, TcP, TcS); Chris@82: TcW = FMA(KP707106781, TcV, TcU); Chris@82: TcZ = FMA(KP707106781, TcY, TcX); Chris@82: Td0 = FNMS(KP198912367, TcZ, TcW); Chris@82: TdL = FMA(KP198912367, TcW, TcZ); Chris@82: } Chris@82: Td1 = TcT + Td0; Chris@82: Tkw = TcT - Td0; Chris@82: TdN = TdL - TdM; Chris@82: Tkq = TdM + TdL; Chris@82: } Chris@82: { Chris@82: E Tdd, TdP, Tdk, TdQ; Chris@82: { Chris@82: E Td5, Tdc, Tdg, Tdj; Chris@82: Td5 = FMA(KP707106781, Td4, Td3); Chris@82: Tdc = Td8 + Tdb; Chris@82: Tdd = FMA(KP923879532, Tdc, Td5); Chris@82: TdP = FNMS(KP923879532, Tdc, Td5); Chris@82: Tdg = FMA(KP707106781, Tdf, Tde); Chris@82: Tdj = Tdh + Tdi; Chris@82: Tdk = FMA(KP923879532, Tdj, Tdg); Chris@82: TdQ = FNMS(KP923879532, Tdj, Tdg); Chris@82: } Chris@82: Tdl = FMA(KP098491403, Tdk, Tdd); Chris@82: TdX = FMA(KP820678790, TdP, TdQ); Chris@82: TdI = FNMS(KP098491403, Tdd, Tdk); Chris@82: TdR = FNMS(KP820678790, TdQ, TdP); Chris@82: } Chris@82: { Chris@82: E Tem, TeE, Tep, TeF; Chris@82: { Chris@82: E Tek, Tel, Ten, Teo; Chris@82: Tek = FNMS(KP707106781, Tdn, Tdm); Chris@82: Tel = TdB - TdA; Chris@82: Tem = FNMS(KP923879532, Tel, Tek); Chris@82: TeE = FMA(KP923879532, Tel, Tek); Chris@82: Ten = FNMS(KP707106781, Tdy, Tdx); Chris@82: Teo = Tdu - Tdr; Chris@82: Tep = FMA(KP923879532, Teo, Ten); Chris@82: TeF = FNMS(KP923879532, Teo, Ten); Chris@82: } Chris@82: Teq = FNMS(KP534511135, Tep, Tem); Chris@82: TeK = FNMS(KP303346683, TeE, TeF); Chris@82: Tet = FMA(KP534511135, Tem, Tep); Chris@82: TeG = FMA(KP303346683, TeF, TeE); Chris@82: } Chris@82: { Chris@82: E Tef, TeB, Tei, TeC; Chris@82: { Chris@82: E Ted, Tee, Teg, Teh; Chris@82: Ted = FNMS(KP707106781, Td4, Td3); Chris@82: Tee = Tdi - Tdh; Chris@82: Tef = FNMS(KP923879532, Tee, Ted); Chris@82: TeB = FMA(KP923879532, Tee, Ted); Chris@82: Teg = FNMS(KP707106781, Tdf, Tde); Chris@82: Teh = Td8 - Tdb; Chris@82: Tei = FNMS(KP923879532, Teh, Teg); Chris@82: TeC = FMA(KP923879532, Teh, Teg); Chris@82: } Chris@82: Tej = FMA(KP534511135, Tei, Tef); Chris@82: TeJ = FMA(KP303346683, TeB, TeC); Chris@82: Teu = FNMS(KP534511135, Tef, Tei); Chris@82: TeD = FNMS(KP303346683, TeC, TeB); Chris@82: } Chris@82: { Chris@82: E Te7, Tex, Tea, Tey; Chris@82: { Chris@82: E Te5, Te6, Te8, Te9; Chris@82: Te5 = FNMS(KP707106781, TcR, TcQ); Chris@82: Te6 = FNMS(KP707106781, TcO, TcN); Chris@82: Te7 = FMA(KP668178637, Te6, Te5); Chris@82: Tex = FNMS(KP668178637, Te5, Te6); Chris@82: Te8 = FNMS(KP707106781, TcY, TcX); Chris@82: Te9 = FNMS(KP707106781, TcV, TcU); Chris@82: Tea = FNMS(KP668178637, Te9, Te8); Chris@82: Tey = FMA(KP668178637, Te8, Te9); Chris@82: } Chris@82: Teb = Te7 - Tea; Chris@82: TkK = Tey - Tex; Chris@82: Tez = Tex + Tey; Chris@82: TkE = Te7 + Tea; Chris@82: } Chris@82: { Chris@82: E Tdw, TdS, TdD, TdT; Chris@82: { Chris@82: E Tdo, Tdv, Tdz, TdC; Chris@82: Tdo = FMA(KP707106781, Tdn, Tdm); Chris@82: Tdv = Tdr + Tdu; Chris@82: Tdw = FMA(KP923879532, Tdv, Tdo); Chris@82: TdS = FNMS(KP923879532, Tdv, Tdo); Chris@82: Tdz = FMA(KP707106781, Tdy, Tdx); Chris@82: TdC = TdA + TdB; Chris@82: TdD = FMA(KP923879532, TdC, Tdz); Chris@82: TdT = FNMS(KP923879532, TdC, Tdz); Chris@82: } Chris@82: TdE = FNMS(KP098491403, TdD, Tdw); Chris@82: TdY = FNMS(KP820678790, TdS, TdT); Chris@82: TdH = FMA(KP098491403, Tdw, TdD); Chris@82: TdU = FMA(KP820678790, TdT, TdS); Chris@82: } Chris@82: { Chris@82: E TcE, Te0, Tkn, TkB, TcL, TkC, Te3, Tko, TcD, Tkm; Chris@82: TcD = TcB + TcC; Chris@82: TcE = FMA(KP707106781, TcD, TcA); Chris@82: Te0 = FNMS(KP707106781, TcD, TcA); Chris@82: Tkm = T7s - T7l; Chris@82: Tkn = FMA(KP707106781, Tkm, Tkl); Chris@82: TkB = FNMS(KP707106781, Tkm, Tkl); Chris@82: { Chris@82: E TcH, TcK, Te1, Te2; Chris@82: TcH = FMA(KP414213562, TcG, TcF); Chris@82: TcK = FNMS(KP414213562, TcJ, TcI); Chris@82: TcL = TcH + TcK; Chris@82: TkC = TcH - TcK; Chris@82: Te1 = FMA(KP414213562, TcI, TcJ); Chris@82: Te2 = FNMS(KP414213562, TcF, TcG); Chris@82: Te3 = Te1 - Te2; Chris@82: Tko = Te2 + Te1; Chris@82: } Chris@82: TcM = FMA(KP923879532, TcL, TcE); Chris@82: TkD = FMA(KP923879532, TkC, TkB); Chris@82: TkJ = FNMS(KP923879532, TkC, TkB); Chris@82: TdK = FNMS(KP923879532, TcL, TcE); Chris@82: Te4 = FNMS(KP923879532, Te3, Te0); Chris@82: Tkp = FMA(KP923879532, Tko, Tkn); Chris@82: Tkv = FNMS(KP923879532, Tko, Tkn); Chris@82: Tew = FMA(KP923879532, Te3, Te0); Chris@82: } Chris@82: { Chris@82: E Td2, TdF, TdW, TdZ; Chris@82: Td2 = FMA(KP980785280, Td1, TcM); Chris@82: TdF = Tdl + TdE; Chris@82: ci[WS(rs, 30)] = FNMS(KP995184726, TdF, Td2); Chris@82: cr[WS(rs, 1)] = FMA(KP995184726, TdF, Td2); Chris@82: TdW = FNMS(KP980785280, TdN, TdK); Chris@82: TdZ = TdX - TdY; Chris@82: ci[WS(rs, 22)] = FNMS(KP773010453, TdZ, TdW); Chris@82: cr[WS(rs, 9)] = FMA(KP773010453, TdZ, TdW); Chris@82: } Chris@82: { Chris@82: E TkF, TkG, TkN, TkO; Chris@82: TkF = FMA(KP831469612, TkE, TkD); Chris@82: TkG = TeJ + TeK; Chris@82: cr[WS(rs, 61)] = -(FMA(KP956940335, TkG, TkF)); Chris@82: ci[WS(rs, 34)] = FNMS(KP956940335, TkG, TkF); Chris@82: TkN = FNMS(KP831469612, TkK, TkJ); Chris@82: TkO = Teq - Tej; Chris@82: cr[WS(rs, 53)] = FMS(KP881921264, TkO, TkN); Chris@82: ci[WS(rs, 42)] = FMA(KP881921264, TkO, TkN); Chris@82: } Chris@82: { Chris@82: E TdG, TdJ, TdO, TdV; Chris@82: TdG = FNMS(KP980785280, Td1, TcM); Chris@82: TdJ = TdH - TdI; Chris@82: cr[WS(rs, 17)] = FNMS(KP995184726, TdJ, TdG); Chris@82: ci[WS(rs, 14)] = FMA(KP995184726, TdJ, TdG); Chris@82: TdO = FMA(KP980785280, TdN, TdK); Chris@82: TdV = TdR + TdU; Chris@82: cr[WS(rs, 25)] = FNMS(KP773010453, TdV, TdO); Chris@82: ci[WS(rs, 6)] = FMA(KP773010453, TdV, TdO); Chris@82: } Chris@82: { Chris@82: E TkH, TkI, TkL, TkM; Chris@82: TkH = FNMS(KP831469612, TkE, TkD); Chris@82: TkI = TeG - TeD; Chris@82: cr[WS(rs, 45)] = FMS(KP956940335, TkI, TkH); Chris@82: ci[WS(rs, 50)] = FMA(KP956940335, TkI, TkH); Chris@82: TkL = FMA(KP831469612, TkK, TkJ); Chris@82: TkM = Teu + Tet; Chris@82: cr[WS(rs, 37)] = FMS(KP881921264, TkM, TkL); Chris@82: ci[WS(rs, 58)] = FMA(KP881921264, TkM, TkL); Chris@82: } Chris@82: { Chris@82: E Tec, Ter, TeI, TeL; Chris@82: Tec = FMA(KP831469612, Teb, Te4); Chris@82: Ter = Tej + Teq; Chris@82: ci[WS(rs, 26)] = FNMS(KP881921264, Ter, Tec); Chris@82: cr[WS(rs, 5)] = FMA(KP881921264, Ter, Tec); Chris@82: TeI = FNMS(KP831469612, Tez, Tew); Chris@82: TeL = TeJ - TeK; Chris@82: ci[WS(rs, 18)] = FNMS(KP956940335, TeL, TeI); Chris@82: cr[WS(rs, 13)] = FMA(KP956940335, TeL, TeI); Chris@82: } Chris@82: { Chris@82: E Tkr, Tks, Tkz, TkA; Chris@82: Tkr = FMA(KP980785280, Tkq, Tkp); Chris@82: Tks = TdI + TdH; Chris@82: cr[WS(rs, 33)] = FMS(KP995184726, Tks, Tkr); Chris@82: ci[WS(rs, 62)] = FMA(KP995184726, Tks, Tkr); Chris@82: Tkz = FNMS(KP980785280, Tkw, Tkv); Chris@82: TkA = TdU - TdR; Chris@82: cr[WS(rs, 41)] = FMS(KP773010453, TkA, Tkz); Chris@82: ci[WS(rs, 54)] = FMA(KP773010453, TkA, Tkz); Chris@82: } Chris@82: { Chris@82: E Tes, Tev, TeA, TeH; Chris@82: Tes = FNMS(KP831469612, Teb, Te4); Chris@82: Tev = Tet - Teu; Chris@82: cr[WS(rs, 21)] = FNMS(KP881921264, Tev, Tes); Chris@82: ci[WS(rs, 10)] = FMA(KP881921264, Tev, Tes); Chris@82: TeA = FMA(KP831469612, Tez, Tew); Chris@82: TeH = TeD + TeG; Chris@82: cr[WS(rs, 29)] = FNMS(KP956940335, TeH, TeA); Chris@82: ci[WS(rs, 2)] = FMA(KP956940335, TeH, TeA); Chris@82: } Chris@82: { Chris@82: E Tkt, Tku, Tkx, Tky; Chris@82: Tkt = FNMS(KP980785280, Tkq, Tkp); Chris@82: Tku = TdE - Tdl; Chris@82: cr[WS(rs, 49)] = FMS(KP995184726, Tku, Tkt); Chris@82: ci[WS(rs, 46)] = FMA(KP995184726, Tku, Tkt); Chris@82: Tkx = FMA(KP980785280, Tkw, Tkv); Chris@82: Tky = TdX + TdY; Chris@82: cr[WS(rs, 57)] = -(FMA(KP773010453, Tky, Tkx)); Chris@82: ci[WS(rs, 38)] = FNMS(KP773010453, Tky, Tkx); Chris@82: } Chris@82: } Chris@82: } Chris@82: } Chris@82: } Chris@82: Chris@82: static const tw_instr twinstr[] = { Chris@82: {TW_FULL, 1, 64}, Chris@82: {TW_NEXT, 1, 0} Chris@82: }; Chris@82: Chris@82: static const hc2hc_desc desc = { 64, "hf_64", twinstr, &GENUS, {520, 126, 518, 0} }; Chris@82: Chris@82: void X(codelet_hf_64) (planner *p) { Chris@82: X(khc2hc_register) (p, hf_64, &desc); Chris@82: } Chris@82: #else Chris@82: Chris@82: /* Generated by: ../../../genfft/gen_hc2hc.native -compact -variables 4 -pipeline-latency 4 -n 64 -dit -name hf_64 -include rdft/scalar/hf.h */ Chris@82: Chris@82: /* Chris@82: * This function contains 1038 FP additions, 500 FP multiplications, Chris@82: * (or, 808 additions, 270 multiplications, 230 fused multiply/add), Chris@82: * 176 stack variables, 15 constants, and 256 memory accesses Chris@82: */ Chris@82: #include "rdft/scalar/hf.h" Chris@82: Chris@82: static void hf_64(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms) Chris@82: { Chris@82: DK(KP290284677, +0.290284677254462367636192375817395274691476278); Chris@82: DK(KP956940335, +0.956940335732208864935797886980269969482849206); Chris@82: DK(KP881921264, +0.881921264348355029712756863660388349508442621); Chris@82: DK(KP471396736, +0.471396736825997648556387625905254377657460319); Chris@82: DK(KP555570233, +0.555570233019602224742830813948532874374937191); Chris@82: DK(KP831469612, +0.831469612302545237078788377617905756738560812); Chris@82: DK(KP098017140, +0.098017140329560601994195563888641845861136673); Chris@82: DK(KP995184726, +0.995184726672196886244836953109479921575474869); Chris@82: DK(KP773010453, +0.773010453362736960810906609758469800971041293); Chris@82: DK(KP634393284, +0.634393284163645498215171613225493370675687095); Chris@82: DK(KP980785280, +0.980785280403230449126182236134239036973933731); Chris@82: DK(KP195090322, +0.195090322016128267848284868477022240927691618); Chris@82: DK(KP382683432, +0.382683432365089771728459984030398866761344562); Chris@82: DK(KP923879532, +0.923879532511286756128183189396788286822416626); Chris@82: DK(KP707106781, +0.707106781186547524400844362104849039284835938); Chris@82: { Chris@82: INT m; Chris@82: for (m = mb, W = W + ((mb - 1) * 126); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 126, MAKE_VOLATILE_STRIDE(128, rs)) { Chris@82: E Tj, TcL, ThT, Tin, T6b, Taz, TgT, Thn, TG, Thm, TcO, TgO, T6m, Tim, TaC; Chris@82: E ThQ, T14, Tfr, T6y, T9O, TaG, Tc0, TcU, TeE, T1r, Tfq, T6J, T9P, TaJ, Tc1; Chris@82: E TcZ, TeF, T1Q, T2d, Tfu, Tfv, Tfw, Tfx, T6Q, TaM, Tdb, TeI, T71, TaQ, T7a; Chris@82: E TaN, Td6, TeJ, T77, TaP, T2B, T2Y, Tfz, TfA, TfB, TfC, T7h, TaW, Tdm, TeL; Chris@82: E T7s, TaU, T7B, TaX, Tdh, TeM, T7y, TaT, T5j, TfR, Tec, TeX, TfY, Tgy, T8D; Chris@82: E Tbl, T8O, Tbx, T9l, Tbm, TdV, Tf0, T9i, Tbw, T3M, TfL, TdL, TeT, TfI, Tgt; Chris@82: E T7K, Tbd, T7V, Tb3, T8s, Tbe, Tdu, TeQ, T8p, Tb2, T4x, TfJ, TdE, TdM, TfO; Chris@82: E Tgu, T87, T8u, T8i, T8v, Tba, Tbh, Tdz, TdN, Tb7, Tbg, T64, TfZ, Te5, Ted; Chris@82: E TfU, Tgz, T90, T9n, T9b, T9o, Tbt, TbA, Te0, Tee, Tbq, Tbz; Chris@82: { Chris@82: E T1, TgR, T6, TgQ, Tc, T68, Th, T69; Chris@82: T1 = cr[0]; Chris@82: TgR = ci[0]; Chris@82: { Chris@82: E T3, T5, T2, T4; Chris@82: T3 = cr[WS(rs, 32)]; Chris@82: T5 = ci[WS(rs, 32)]; Chris@82: T2 = W[62]; Chris@82: T4 = W[63]; Chris@82: T6 = FMA(T2, T3, T4 * T5); Chris@82: TgQ = FNMS(T4, T3, T2 * T5); Chris@82: } Chris@82: { Chris@82: E T9, Tb, T8, Ta; Chris@82: T9 = cr[WS(rs, 16)]; Chris@82: Tb = ci[WS(rs, 16)]; Chris@82: T8 = W[30]; Chris@82: Ta = W[31]; Chris@82: Tc = FMA(T8, T9, Ta * Tb); Chris@82: T68 = FNMS(Ta, T9, T8 * Tb); Chris@82: } Chris@82: { Chris@82: E Te, Tg, Td, Tf; Chris@82: Te = cr[WS(rs, 48)]; Chris@82: Tg = ci[WS(rs, 48)]; Chris@82: Td = W[94]; Chris@82: Tf = W[95]; Chris@82: Th = FMA(Td, Te, Tf * Tg); Chris@82: T69 = FNMS(Tf, Te, Td * Tg); Chris@82: } Chris@82: { Chris@82: E T7, Ti, ThR, ThS; Chris@82: T7 = T1 + T6; Chris@82: Ti = Tc + Th; Chris@82: Tj = T7 + Ti; Chris@82: TcL = T7 - Ti; Chris@82: ThR = Tc - Th; Chris@82: ThS = TgR - TgQ; Chris@82: ThT = ThR + ThS; Chris@82: Tin = ThS - ThR; Chris@82: } Chris@82: { Chris@82: E T67, T6a, TgP, TgS; Chris@82: T67 = T1 - T6; Chris@82: T6a = T68 - T69; Chris@82: T6b = T67 - T6a; Chris@82: Taz = T67 + T6a; Chris@82: TgP = T68 + T69; Chris@82: TgS = TgQ + TgR; Chris@82: TgT = TgP + TgS; Chris@82: Thn = TgS - TgP; Chris@82: } Chris@82: } Chris@82: { Chris@82: E To, T6d, Tt, T6e, T6c, T6f, Tz, T6i, TE, T6j, T6h, T6k; Chris@82: { Chris@82: E Tl, Tn, Tk, Tm; Chris@82: Tl = cr[WS(rs, 8)]; Chris@82: Tn = ci[WS(rs, 8)]; Chris@82: Tk = W[14]; Chris@82: Tm = W[15]; Chris@82: To = FMA(Tk, Tl, Tm * Tn); Chris@82: T6d = FNMS(Tm, Tl, Tk * Tn); Chris@82: } Chris@82: { Chris@82: E Tq, Ts, Tp, Tr; Chris@82: Tq = cr[WS(rs, 40)]; Chris@82: Ts = ci[WS(rs, 40)]; Chris@82: Tp = W[78]; Chris@82: Tr = W[79]; Chris@82: Tt = FMA(Tp, Tq, Tr * Ts); Chris@82: T6e = FNMS(Tr, Tq, Tp * Ts); Chris@82: } Chris@82: T6c = To - Tt; Chris@82: T6f = T6d - T6e; Chris@82: { Chris@82: E Tw, Ty, Tv, Tx; Chris@82: Tw = cr[WS(rs, 56)]; Chris@82: Ty = ci[WS(rs, 56)]; Chris@82: Tv = W[110]; Chris@82: Tx = W[111]; Chris@82: Tz = FMA(Tv, Tw, Tx * Ty); Chris@82: T6i = FNMS(Tx, Tw, Tv * Ty); Chris@82: } Chris@82: { Chris@82: E TB, TD, TA, TC; Chris@82: TB = cr[WS(rs, 24)]; Chris@82: TD = ci[WS(rs, 24)]; Chris@82: TA = W[46]; Chris@82: TC = W[47]; Chris@82: TE = FMA(TA, TB, TC * TD); Chris@82: T6j = FNMS(TC, TB, TA * TD); Chris@82: } Chris@82: T6h = Tz - TE; Chris@82: T6k = T6i - T6j; Chris@82: { Chris@82: E Tu, TF, TcM, TcN; Chris@82: Tu = To + Tt; Chris@82: TF = Tz + TE; Chris@82: TG = Tu + TF; Chris@82: Thm = Tu - TF; Chris@82: TcM = T6i + T6j; Chris@82: TcN = T6d + T6e; Chris@82: TcO = TcM - TcN; Chris@82: TgO = TcN + TcM; Chris@82: } Chris@82: { Chris@82: E T6g, T6l, TaA, TaB; Chris@82: T6g = T6c - T6f; Chris@82: T6l = T6h + T6k; Chris@82: T6m = KP707106781 * (T6g + T6l); Chris@82: Tim = KP707106781 * (T6l - T6g); Chris@82: TaA = T6c + T6f; Chris@82: TaB = T6h - T6k; Chris@82: TaC = KP707106781 * (TaA + TaB); Chris@82: ThQ = KP707106781 * (TaA - TaB); Chris@82: } Chris@82: } Chris@82: { Chris@82: E TS, TcR, T6o, T6v, T13, TcS, T6r, T6w, T6s, T6x; Chris@82: { Chris@82: E TM, T6t, TR, T6u; Chris@82: { Chris@82: E TJ, TL, TI, TK; Chris@82: TJ = cr[WS(rs, 4)]; Chris@82: TL = ci[WS(rs, 4)]; Chris@82: TI = W[6]; Chris@82: TK = W[7]; Chris@82: TM = FMA(TI, TJ, TK * TL); Chris@82: T6t = FNMS(TK, TJ, TI * TL); Chris@82: } Chris@82: { Chris@82: E TO, TQ, TN, TP; Chris@82: TO = cr[WS(rs, 36)]; Chris@82: TQ = ci[WS(rs, 36)]; Chris@82: TN = W[70]; Chris@82: TP = W[71]; Chris@82: TR = FMA(TN, TO, TP * TQ); Chris@82: T6u = FNMS(TP, TO, TN * TQ); Chris@82: } Chris@82: TS = TM + TR; Chris@82: TcR = T6t + T6u; Chris@82: T6o = TM - TR; Chris@82: T6v = T6t - T6u; Chris@82: } Chris@82: { Chris@82: E TX, T6p, T12, T6q; Chris@82: { Chris@82: E TU, TW, TT, TV; Chris@82: TU = cr[WS(rs, 20)]; Chris@82: TW = ci[WS(rs, 20)]; Chris@82: TT = W[38]; Chris@82: TV = W[39]; Chris@82: TX = FMA(TT, TU, TV * TW); Chris@82: T6p = FNMS(TV, TU, TT * TW); Chris@82: } Chris@82: { Chris@82: E TZ, T11, TY, T10; Chris@82: TZ = cr[WS(rs, 52)]; Chris@82: T11 = ci[WS(rs, 52)]; Chris@82: TY = W[102]; Chris@82: T10 = W[103]; Chris@82: T12 = FMA(TY, TZ, T10 * T11); Chris@82: T6q = FNMS(T10, TZ, TY * T11); Chris@82: } Chris@82: T13 = TX + T12; Chris@82: TcS = T6p + T6q; Chris@82: T6r = T6p - T6q; Chris@82: T6w = TX - T12; Chris@82: } Chris@82: T14 = TS + T13; Chris@82: Tfr = TcR + TcS; Chris@82: T6s = T6o - T6r; Chris@82: T6x = T6v + T6w; Chris@82: T6y = FNMS(KP382683432, T6x, KP923879532 * T6s); Chris@82: T9O = FMA(KP923879532, T6x, KP382683432 * T6s); Chris@82: { Chris@82: E TaE, TaF, TcQ, TcT; Chris@82: TaE = T6v - T6w; Chris@82: TaF = T6o + T6r; Chris@82: TaG = FMA(KP382683432, TaE, KP923879532 * TaF); Chris@82: Tc0 = FNMS(KP923879532, TaE, KP382683432 * TaF); Chris@82: TcQ = TS - T13; Chris@82: TcT = TcR - TcS; Chris@82: TcU = TcQ + TcT; Chris@82: TeE = TcQ - TcT; Chris@82: } Chris@82: } Chris@82: { Chris@82: E T1f, TcW, T6B, T6E, T1q, TcX, T6C, T6H, T6D, T6I; Chris@82: { Chris@82: E T19, T6z, T1e, T6A; Chris@82: { Chris@82: E T16, T18, T15, T17; Chris@82: T16 = cr[WS(rs, 60)]; Chris@82: T18 = ci[WS(rs, 60)]; Chris@82: T15 = W[118]; Chris@82: T17 = W[119]; Chris@82: T19 = FMA(T15, T16, T17 * T18); Chris@82: T6z = FNMS(T17, T16, T15 * T18); Chris@82: } Chris@82: { Chris@82: E T1b, T1d, T1a, T1c; Chris@82: T1b = cr[WS(rs, 28)]; Chris@82: T1d = ci[WS(rs, 28)]; Chris@82: T1a = W[54]; Chris@82: T1c = W[55]; Chris@82: T1e = FMA(T1a, T1b, T1c * T1d); Chris@82: T6A = FNMS(T1c, T1b, T1a * T1d); Chris@82: } Chris@82: T1f = T19 + T1e; Chris@82: TcW = T6z + T6A; Chris@82: T6B = T6z - T6A; Chris@82: T6E = T19 - T1e; Chris@82: } Chris@82: { Chris@82: E T1k, T6F, T1p, T6G; Chris@82: { Chris@82: E T1h, T1j, T1g, T1i; Chris@82: T1h = cr[WS(rs, 12)]; Chris@82: T1j = ci[WS(rs, 12)]; Chris@82: T1g = W[22]; Chris@82: T1i = W[23]; Chris@82: T1k = FMA(T1g, T1h, T1i * T1j); Chris@82: T6F = FNMS(T1i, T1h, T1g * T1j); Chris@82: } Chris@82: { Chris@82: E T1m, T1o, T1l, T1n; Chris@82: T1m = cr[WS(rs, 44)]; Chris@82: T1o = ci[WS(rs, 44)]; Chris@82: T1l = W[86]; Chris@82: T1n = W[87]; Chris@82: T1p = FMA(T1l, T1m, T1n * T1o); Chris@82: T6G = FNMS(T1n, T1m, T1l * T1o); Chris@82: } Chris@82: T1q = T1k + T1p; Chris@82: TcX = T6F + T6G; Chris@82: T6C = T1k - T1p; Chris@82: T6H = T6F - T6G; Chris@82: } Chris@82: T1r = T1f + T1q; Chris@82: Tfq = TcW + TcX; Chris@82: T6D = T6B + T6C; Chris@82: T6I = T6E - T6H; Chris@82: T6J = FMA(KP382683432, T6D, KP923879532 * T6I); Chris@82: T9P = FNMS(KP923879532, T6D, KP382683432 * T6I); Chris@82: { Chris@82: E TaH, TaI, TcV, TcY; Chris@82: TaH = T6E + T6H; Chris@82: TaI = T6B - T6C; Chris@82: TaJ = FNMS(KP382683432, TaI, KP923879532 * TaH); Chris@82: Tc1 = FMA(KP923879532, TaI, KP382683432 * TaH); Chris@82: TcV = T1f - T1q; Chris@82: TcY = TcW - TcX; Chris@82: TcZ = TcV - TcY; Chris@82: TeF = TcV + TcY; Chris@82: } Chris@82: } Chris@82: { Chris@82: E T1y, T73, T1D, T74, T1E, Td7, T1J, T6N, T1O, T6O, T1P, Td8, T21, Td4, T6R; Chris@82: E T6U, T2c, Td3, T6W, T6Z; Chris@82: { Chris@82: E T1v, T1x, T1u, T1w; Chris@82: T1v = cr[WS(rs, 2)]; Chris@82: T1x = ci[WS(rs, 2)]; Chris@82: T1u = W[2]; Chris@82: T1w = W[3]; Chris@82: T1y = FMA(T1u, T1v, T1w * T1x); Chris@82: T73 = FNMS(T1w, T1v, T1u * T1x); Chris@82: } Chris@82: { Chris@82: E T1A, T1C, T1z, T1B; Chris@82: T1A = cr[WS(rs, 34)]; Chris@82: T1C = ci[WS(rs, 34)]; Chris@82: T1z = W[66]; Chris@82: T1B = W[67]; Chris@82: T1D = FMA(T1z, T1A, T1B * T1C); Chris@82: T74 = FNMS(T1B, T1A, T1z * T1C); Chris@82: } Chris@82: T1E = T1y + T1D; Chris@82: Td7 = T73 + T74; Chris@82: { Chris@82: E T1G, T1I, T1F, T1H; Chris@82: T1G = cr[WS(rs, 18)]; Chris@82: T1I = ci[WS(rs, 18)]; Chris@82: T1F = W[34]; Chris@82: T1H = W[35]; Chris@82: T1J = FMA(T1F, T1G, T1H * T1I); Chris@82: T6N = FNMS(T1H, T1G, T1F * T1I); Chris@82: } Chris@82: { Chris@82: E T1L, T1N, T1K, T1M; Chris@82: T1L = cr[WS(rs, 50)]; Chris@82: T1N = ci[WS(rs, 50)]; Chris@82: T1K = W[98]; Chris@82: T1M = W[99]; Chris@82: T1O = FMA(T1K, T1L, T1M * T1N); Chris@82: T6O = FNMS(T1M, T1L, T1K * T1N); Chris@82: } Chris@82: T1P = T1J + T1O; Chris@82: Td8 = T6N + T6O; Chris@82: { Chris@82: E T1V, T6S, T20, T6T; Chris@82: { Chris@82: E T1S, T1U, T1R, T1T; Chris@82: T1S = cr[WS(rs, 10)]; Chris@82: T1U = ci[WS(rs, 10)]; Chris@82: T1R = W[18]; Chris@82: T1T = W[19]; Chris@82: T1V = FMA(T1R, T1S, T1T * T1U); Chris@82: T6S = FNMS(T1T, T1S, T1R * T1U); Chris@82: } Chris@82: { Chris@82: E T1X, T1Z, T1W, T1Y; Chris@82: T1X = cr[WS(rs, 42)]; Chris@82: T1Z = ci[WS(rs, 42)]; Chris@82: T1W = W[82]; Chris@82: T1Y = W[83]; Chris@82: T20 = FMA(T1W, T1X, T1Y * T1Z); Chris@82: T6T = FNMS(T1Y, T1X, T1W * T1Z); Chris@82: } Chris@82: T21 = T1V + T20; Chris@82: Td4 = T6S + T6T; Chris@82: T6R = T1V - T20; Chris@82: T6U = T6S - T6T; Chris@82: } Chris@82: { Chris@82: E T26, T6X, T2b, T6Y; Chris@82: { Chris@82: E T23, T25, T22, T24; Chris@82: T23 = cr[WS(rs, 58)]; Chris@82: T25 = ci[WS(rs, 58)]; Chris@82: T22 = W[114]; Chris@82: T24 = W[115]; Chris@82: T26 = FMA(T22, T23, T24 * T25); Chris@82: T6X = FNMS(T24, T23, T22 * T25); Chris@82: } Chris@82: { Chris@82: E T28, T2a, T27, T29; Chris@82: T28 = cr[WS(rs, 26)]; Chris@82: T2a = ci[WS(rs, 26)]; Chris@82: T27 = W[50]; Chris@82: T29 = W[51]; Chris@82: T2b = FMA(T27, T28, T29 * T2a); Chris@82: T6Y = FNMS(T29, T28, T27 * T2a); Chris@82: } Chris@82: T2c = T26 + T2b; Chris@82: Td3 = T6X + T6Y; Chris@82: T6W = T26 - T2b; Chris@82: T6Z = T6X - T6Y; Chris@82: } Chris@82: T1Q = T1E + T1P; Chris@82: T2d = T21 + T2c; Chris@82: Tfu = T1Q - T2d; Chris@82: Tfv = Td7 + Td8; Chris@82: Tfw = Td4 + Td3; Chris@82: Tfx = Tfv - Tfw; Chris@82: { Chris@82: E T6M, T6P, Td9, Tda; Chris@82: T6M = T1y - T1D; Chris@82: T6P = T6N - T6O; Chris@82: T6Q = T6M - T6P; Chris@82: TaM = T6M + T6P; Chris@82: Td9 = Td7 - Td8; Chris@82: Tda = T21 - T2c; Chris@82: Tdb = Td9 - Tda; Chris@82: TeI = Td9 + Tda; Chris@82: } Chris@82: { Chris@82: E T6V, T70, T78, T79; Chris@82: T6V = T6R - T6U; Chris@82: T70 = T6W + T6Z; Chris@82: T71 = KP707106781 * (T6V + T70); Chris@82: TaQ = KP707106781 * (T70 - T6V); Chris@82: T78 = T6R + T6U; Chris@82: T79 = T6Z - T6W; Chris@82: T7a = KP707106781 * (T78 + T79); Chris@82: TaN = KP707106781 * (T78 - T79); Chris@82: } Chris@82: { Chris@82: E Td2, Td5, T75, T76; Chris@82: Td2 = T1E - T1P; Chris@82: Td5 = Td3 - Td4; Chris@82: Td6 = Td2 - Td5; Chris@82: TeJ = Td2 + Td5; Chris@82: T75 = T73 - T74; Chris@82: T76 = T1J - T1O; Chris@82: T77 = T75 + T76; Chris@82: TaP = T75 - T76; Chris@82: } Chris@82: } Chris@82: { Chris@82: E T2j, T7u, T2o, T7v, T2p, Tdd, T2u, T7e, T2z, T7f, T2A, Tde, T2M, Tdk, T7i; Chris@82: E T7l, T2X, Tdj, T7n, T7q; Chris@82: { Chris@82: E T2g, T2i, T2f, T2h; Chris@82: T2g = cr[WS(rs, 62)]; Chris@82: T2i = ci[WS(rs, 62)]; Chris@82: T2f = W[122]; Chris@82: T2h = W[123]; Chris@82: T2j = FMA(T2f, T2g, T2h * T2i); Chris@82: T7u = FNMS(T2h, T2g, T2f * T2i); Chris@82: } Chris@82: { Chris@82: E T2l, T2n, T2k, T2m; Chris@82: T2l = cr[WS(rs, 30)]; Chris@82: T2n = ci[WS(rs, 30)]; Chris@82: T2k = W[58]; Chris@82: T2m = W[59]; Chris@82: T2o = FMA(T2k, T2l, T2m * T2n); Chris@82: T7v = FNMS(T2m, T2l, T2k * T2n); Chris@82: } Chris@82: T2p = T2j + T2o; Chris@82: Tdd = T7u + T7v; Chris@82: { Chris@82: E T2r, T2t, T2q, T2s; Chris@82: T2r = cr[WS(rs, 14)]; Chris@82: T2t = ci[WS(rs, 14)]; Chris@82: T2q = W[26]; Chris@82: T2s = W[27]; Chris@82: T2u = FMA(T2q, T2r, T2s * T2t); Chris@82: T7e = FNMS(T2s, T2r, T2q * T2t); Chris@82: } Chris@82: { Chris@82: E T2w, T2y, T2v, T2x; Chris@82: T2w = cr[WS(rs, 46)]; Chris@82: T2y = ci[WS(rs, 46)]; Chris@82: T2v = W[90]; Chris@82: T2x = W[91]; Chris@82: T2z = FMA(T2v, T2w, T2x * T2y); Chris@82: T7f = FNMS(T2x, T2w, T2v * T2y); Chris@82: } Chris@82: T2A = T2u + T2z; Chris@82: Tde = T7e + T7f; Chris@82: { Chris@82: E T2G, T7j, T2L, T7k; Chris@82: { Chris@82: E T2D, T2F, T2C, T2E; Chris@82: T2D = cr[WS(rs, 6)]; Chris@82: T2F = ci[WS(rs, 6)]; Chris@82: T2C = W[10]; Chris@82: T2E = W[11]; Chris@82: T2G = FMA(T2C, T2D, T2E * T2F); Chris@82: T7j = FNMS(T2E, T2D, T2C * T2F); Chris@82: } Chris@82: { Chris@82: E T2I, T2K, T2H, T2J; Chris@82: T2I = cr[WS(rs, 38)]; Chris@82: T2K = ci[WS(rs, 38)]; Chris@82: T2H = W[74]; Chris@82: T2J = W[75]; Chris@82: T2L = FMA(T2H, T2I, T2J * T2K); Chris@82: T7k = FNMS(T2J, T2I, T2H * T2K); Chris@82: } Chris@82: T2M = T2G + T2L; Chris@82: Tdk = T7j + T7k; Chris@82: T7i = T2G - T2L; Chris@82: T7l = T7j - T7k; Chris@82: } Chris@82: { Chris@82: E T2R, T7o, T2W, T7p; Chris@82: { Chris@82: E T2O, T2Q, T2N, T2P; Chris@82: T2O = cr[WS(rs, 54)]; Chris@82: T2Q = ci[WS(rs, 54)]; Chris@82: T2N = W[106]; Chris@82: T2P = W[107]; Chris@82: T2R = FMA(T2N, T2O, T2P * T2Q); Chris@82: T7o = FNMS(T2P, T2O, T2N * T2Q); Chris@82: } Chris@82: { Chris@82: E T2T, T2V, T2S, T2U; Chris@82: T2T = cr[WS(rs, 22)]; Chris@82: T2V = ci[WS(rs, 22)]; Chris@82: T2S = W[42]; Chris@82: T2U = W[43]; Chris@82: T2W = FMA(T2S, T2T, T2U * T2V); Chris@82: T7p = FNMS(T2U, T2T, T2S * T2V); Chris@82: } Chris@82: T2X = T2R + T2W; Chris@82: Tdj = T7o + T7p; Chris@82: T7n = T2R - T2W; Chris@82: T7q = T7o - T7p; Chris@82: } Chris@82: T2B = T2p + T2A; Chris@82: T2Y = T2M + T2X; Chris@82: Tfz = T2B - T2Y; Chris@82: TfA = Tdd + Tde; Chris@82: TfB = Tdk + Tdj; Chris@82: TfC = TfA - TfB; Chris@82: { Chris@82: E T7d, T7g, Tdi, Tdl; Chris@82: T7d = T2j - T2o; Chris@82: T7g = T7e - T7f; Chris@82: T7h = T7d - T7g; Chris@82: TaW = T7d + T7g; Chris@82: Tdi = T2p - T2A; Chris@82: Tdl = Tdj - Tdk; Chris@82: Tdm = Tdi - Tdl; Chris@82: TeL = Tdi + Tdl; Chris@82: } Chris@82: { Chris@82: E T7m, T7r, T7z, T7A; Chris@82: T7m = T7i - T7l; Chris@82: T7r = T7n + T7q; Chris@82: T7s = KP707106781 * (T7m + T7r); Chris@82: TaU = KP707106781 * (T7r - T7m); Chris@82: T7z = T7i + T7l; Chris@82: T7A = T7q - T7n; Chris@82: T7B = KP707106781 * (T7z + T7A); Chris@82: TaX = KP707106781 * (T7z - T7A); Chris@82: } Chris@82: { Chris@82: E Tdf, Tdg, T7w, T7x; Chris@82: Tdf = Tdd - Tde; Chris@82: Tdg = T2M - T2X; Chris@82: Tdh = Tdf - Tdg; Chris@82: TeM = Tdf + Tdg; Chris@82: T7w = T7u - T7v; Chris@82: T7x = T2u - T2z; Chris@82: T7y = T7w + T7x; Chris@82: TaT = T7w - T7x; Chris@82: } Chris@82: } Chris@82: { Chris@82: E T4D, T9e, T4I, T9f, T4J, TdR, T4O, T8A, T4T, T8B, T4U, TdS, T56, Tea, T8E; Chris@82: E T8H, T5h, Te9, T8J, T8M; Chris@82: { Chris@82: E T4A, T4C, T4z, T4B; Chris@82: T4A = cr[WS(rs, 63)]; Chris@82: T4C = ci[WS(rs, 63)]; Chris@82: T4z = W[124]; Chris@82: T4B = W[125]; Chris@82: T4D = FMA(T4z, T4A, T4B * T4C); Chris@82: T9e = FNMS(T4B, T4A, T4z * T4C); Chris@82: } Chris@82: { Chris@82: E T4F, T4H, T4E, T4G; Chris@82: T4F = cr[WS(rs, 31)]; Chris@82: T4H = ci[WS(rs, 31)]; Chris@82: T4E = W[60]; Chris@82: T4G = W[61]; Chris@82: T4I = FMA(T4E, T4F, T4G * T4H); Chris@82: T9f = FNMS(T4G, T4F, T4E * T4H); Chris@82: } Chris@82: T4J = T4D + T4I; Chris@82: TdR = T9e + T9f; Chris@82: { Chris@82: E T4L, T4N, T4K, T4M; Chris@82: T4L = cr[WS(rs, 15)]; Chris@82: T4N = ci[WS(rs, 15)]; Chris@82: T4K = W[28]; Chris@82: T4M = W[29]; Chris@82: T4O = FMA(T4K, T4L, T4M * T4N); Chris@82: T8A = FNMS(T4M, T4L, T4K * T4N); Chris@82: } Chris@82: { Chris@82: E T4Q, T4S, T4P, T4R; Chris@82: T4Q = cr[WS(rs, 47)]; Chris@82: T4S = ci[WS(rs, 47)]; Chris@82: T4P = W[92]; Chris@82: T4R = W[93]; Chris@82: T4T = FMA(T4P, T4Q, T4R * T4S); Chris@82: T8B = FNMS(T4R, T4Q, T4P * T4S); Chris@82: } Chris@82: T4U = T4O + T4T; Chris@82: TdS = T8A + T8B; Chris@82: { Chris@82: E T50, T8F, T55, T8G; Chris@82: { Chris@82: E T4X, T4Z, T4W, T4Y; Chris@82: T4X = cr[WS(rs, 7)]; Chris@82: T4Z = ci[WS(rs, 7)]; Chris@82: T4W = W[12]; Chris@82: T4Y = W[13]; Chris@82: T50 = FMA(T4W, T4X, T4Y * T4Z); Chris@82: T8F = FNMS(T4Y, T4X, T4W * T4Z); Chris@82: } Chris@82: { Chris@82: E T52, T54, T51, T53; Chris@82: T52 = cr[WS(rs, 39)]; Chris@82: T54 = ci[WS(rs, 39)]; Chris@82: T51 = W[76]; Chris@82: T53 = W[77]; Chris@82: T55 = FMA(T51, T52, T53 * T54); Chris@82: T8G = FNMS(T53, T52, T51 * T54); Chris@82: } Chris@82: T56 = T50 + T55; Chris@82: Tea = T8F + T8G; Chris@82: T8E = T50 - T55; Chris@82: T8H = T8F - T8G; Chris@82: } Chris@82: { Chris@82: E T5b, T8K, T5g, T8L; Chris@82: { Chris@82: E T58, T5a, T57, T59; Chris@82: T58 = cr[WS(rs, 55)]; Chris@82: T5a = ci[WS(rs, 55)]; Chris@82: T57 = W[108]; Chris@82: T59 = W[109]; Chris@82: T5b = FMA(T57, T58, T59 * T5a); Chris@82: T8K = FNMS(T59, T58, T57 * T5a); Chris@82: } Chris@82: { Chris@82: E T5d, T5f, T5c, T5e; Chris@82: T5d = cr[WS(rs, 23)]; Chris@82: T5f = ci[WS(rs, 23)]; Chris@82: T5c = W[44]; Chris@82: T5e = W[45]; Chris@82: T5g = FMA(T5c, T5d, T5e * T5f); Chris@82: T8L = FNMS(T5e, T5d, T5c * T5f); Chris@82: } Chris@82: T5h = T5b + T5g; Chris@82: Te9 = T8K + T8L; Chris@82: T8J = T5b - T5g; Chris@82: T8M = T8K - T8L; Chris@82: } Chris@82: { Chris@82: E T4V, T5i, Te8, Teb; Chris@82: T4V = T4J + T4U; Chris@82: T5i = T56 + T5h; Chris@82: T5j = T4V + T5i; Chris@82: TfR = T4V - T5i; Chris@82: Te8 = T4J - T4U; Chris@82: Teb = Te9 - Tea; Chris@82: Tec = Te8 - Teb; Chris@82: TeX = Te8 + Teb; Chris@82: } Chris@82: { Chris@82: E TfW, TfX, T8z, T8C; Chris@82: TfW = TdR + TdS; Chris@82: TfX = Tea + Te9; Chris@82: TfY = TfW - TfX; Chris@82: Tgy = TfW + TfX; Chris@82: T8z = T4D - T4I; Chris@82: T8C = T8A - T8B; Chris@82: T8D = T8z - T8C; Chris@82: Tbl = T8z + T8C; Chris@82: } Chris@82: { Chris@82: E T8I, T8N, T9j, T9k; Chris@82: T8I = T8E - T8H; Chris@82: T8N = T8J + T8M; Chris@82: T8O = KP707106781 * (T8I + T8N); Chris@82: Tbx = KP707106781 * (T8N - T8I); Chris@82: T9j = T8E + T8H; Chris@82: T9k = T8M - T8J; Chris@82: T9l = KP707106781 * (T9j + T9k); Chris@82: Tbm = KP707106781 * (T9j - T9k); Chris@82: } Chris@82: { Chris@82: E TdT, TdU, T9g, T9h; Chris@82: TdT = TdR - TdS; Chris@82: TdU = T56 - T5h; Chris@82: TdV = TdT - TdU; Chris@82: Tf0 = TdT + TdU; Chris@82: T9g = T9e - T9f; Chris@82: T9h = T4O - T4T; Chris@82: T9i = T9g + T9h; Chris@82: Tbw = T9g - T9h; Chris@82: } Chris@82: } Chris@82: { Chris@82: E T36, T7G, T3b, T7H, T3c, TdH, T3h, T8m, T3m, T8n, T3n, TdI, T3z, Tds, T7L; Chris@82: E T7O, T3K, Tdr, T7S, T7T; Chris@82: { Chris@82: E T33, T35, T32, T34; Chris@82: T33 = cr[WS(rs, 1)]; Chris@82: T35 = ci[WS(rs, 1)]; Chris@82: T32 = W[0]; Chris@82: T34 = W[1]; Chris@82: T36 = FMA(T32, T33, T34 * T35); Chris@82: T7G = FNMS(T34, T33, T32 * T35); Chris@82: } Chris@82: { Chris@82: E T38, T3a, T37, T39; Chris@82: T38 = cr[WS(rs, 33)]; Chris@82: T3a = ci[WS(rs, 33)]; Chris@82: T37 = W[64]; Chris@82: T39 = W[65]; Chris@82: T3b = FMA(T37, T38, T39 * T3a); Chris@82: T7H = FNMS(T39, T38, T37 * T3a); Chris@82: } Chris@82: T3c = T36 + T3b; Chris@82: TdH = T7G + T7H; Chris@82: { Chris@82: E T3e, T3g, T3d, T3f; Chris@82: T3e = cr[WS(rs, 17)]; Chris@82: T3g = ci[WS(rs, 17)]; Chris@82: T3d = W[32]; Chris@82: T3f = W[33]; Chris@82: T3h = FMA(T3d, T3e, T3f * T3g); Chris@82: T8m = FNMS(T3f, T3e, T3d * T3g); Chris@82: } Chris@82: { Chris@82: E T3j, T3l, T3i, T3k; Chris@82: T3j = cr[WS(rs, 49)]; Chris@82: T3l = ci[WS(rs, 49)]; Chris@82: T3i = W[96]; Chris@82: T3k = W[97]; Chris@82: T3m = FMA(T3i, T3j, T3k * T3l); Chris@82: T8n = FNMS(T3k, T3j, T3i * T3l); Chris@82: } Chris@82: T3n = T3h + T3m; Chris@82: TdI = T8m + T8n; Chris@82: { Chris@82: E T3t, T7M, T3y, T7N; Chris@82: { Chris@82: E T3q, T3s, T3p, T3r; Chris@82: T3q = cr[WS(rs, 9)]; Chris@82: T3s = ci[WS(rs, 9)]; Chris@82: T3p = W[16]; Chris@82: T3r = W[17]; Chris@82: T3t = FMA(T3p, T3q, T3r * T3s); Chris@82: T7M = FNMS(T3r, T3q, T3p * T3s); Chris@82: } Chris@82: { Chris@82: E T3v, T3x, T3u, T3w; Chris@82: T3v = cr[WS(rs, 41)]; Chris@82: T3x = ci[WS(rs, 41)]; Chris@82: T3u = W[80]; Chris@82: T3w = W[81]; Chris@82: T3y = FMA(T3u, T3v, T3w * T3x); Chris@82: T7N = FNMS(T3w, T3v, T3u * T3x); Chris@82: } Chris@82: T3z = T3t + T3y; Chris@82: Tds = T7M + T7N; Chris@82: T7L = T3t - T3y; Chris@82: T7O = T7M - T7N; Chris@82: } Chris@82: { Chris@82: E T3E, T7Q, T3J, T7R; Chris@82: { Chris@82: E T3B, T3D, T3A, T3C; Chris@82: T3B = cr[WS(rs, 57)]; Chris@82: T3D = ci[WS(rs, 57)]; Chris@82: T3A = W[112]; Chris@82: T3C = W[113]; Chris@82: T3E = FMA(T3A, T3B, T3C * T3D); Chris@82: T7Q = FNMS(T3C, T3B, T3A * T3D); Chris@82: } Chris@82: { Chris@82: E T3G, T3I, T3F, T3H; Chris@82: T3G = cr[WS(rs, 25)]; Chris@82: T3I = ci[WS(rs, 25)]; Chris@82: T3F = W[48]; Chris@82: T3H = W[49]; Chris@82: T3J = FMA(T3F, T3G, T3H * T3I); Chris@82: T7R = FNMS(T3H, T3G, T3F * T3I); Chris@82: } Chris@82: T3K = T3E + T3J; Chris@82: Tdr = T7Q + T7R; Chris@82: T7S = T7Q - T7R; Chris@82: T7T = T3E - T3J; Chris@82: } Chris@82: { Chris@82: E T3o, T3L, TdJ, TdK; Chris@82: T3o = T3c + T3n; Chris@82: T3L = T3z + T3K; Chris@82: T3M = T3o + T3L; Chris@82: TfL = T3o - T3L; Chris@82: TdJ = TdH - TdI; Chris@82: TdK = T3z - T3K; Chris@82: TdL = TdJ - TdK; Chris@82: TeT = TdJ + TdK; Chris@82: } Chris@82: { Chris@82: E TfG, TfH, T7I, T7J; Chris@82: TfG = TdH + TdI; Chris@82: TfH = Tds + Tdr; Chris@82: TfI = TfG - TfH; Chris@82: Tgt = TfG + TfH; Chris@82: T7I = T7G - T7H; Chris@82: T7J = T3h - T3m; Chris@82: T7K = T7I + T7J; Chris@82: Tbd = T7I - T7J; Chris@82: } Chris@82: { Chris@82: E T7P, T7U, T8q, T8r; Chris@82: T7P = T7L + T7O; Chris@82: T7U = T7S - T7T; Chris@82: T7V = KP707106781 * (T7P + T7U); Chris@82: Tb3 = KP707106781 * (T7P - T7U); Chris@82: T8q = T7L - T7O; Chris@82: T8r = T7T + T7S; Chris@82: T8s = KP707106781 * (T8q + T8r); Chris@82: Tbe = KP707106781 * (T8r - T8q); Chris@82: } Chris@82: { Chris@82: E Tdq, Tdt, T8l, T8o; Chris@82: Tdq = T3c - T3n; Chris@82: Tdt = Tdr - Tds; Chris@82: Tdu = Tdq - Tdt; Chris@82: TeQ = Tdq + Tdt; Chris@82: T8l = T36 - T3b; Chris@82: T8o = T8m - T8n; Chris@82: T8p = T8l - T8o; Chris@82: Tb2 = T8l + T8o; Chris@82: } Chris@82: } Chris@82: { Chris@82: E T3X, Tdw, T7Z, T82, T4v, TdB, T8b, T8g, T48, Tdx, T80, T85, T4k, TdA, T8a; Chris@82: E T8d; Chris@82: { Chris@82: E T3R, T7X, T3W, T7Y; Chris@82: { Chris@82: E T3O, T3Q, T3N, T3P; Chris@82: T3O = cr[WS(rs, 5)]; Chris@82: T3Q = ci[WS(rs, 5)]; Chris@82: T3N = W[8]; Chris@82: T3P = W[9]; Chris@82: T3R = FMA(T3N, T3O, T3P * T3Q); Chris@82: T7X = FNMS(T3P, T3O, T3N * T3Q); Chris@82: } Chris@82: { Chris@82: E T3T, T3V, T3S, T3U; Chris@82: T3T = cr[WS(rs, 37)]; Chris@82: T3V = ci[WS(rs, 37)]; Chris@82: T3S = W[72]; Chris@82: T3U = W[73]; Chris@82: T3W = FMA(T3S, T3T, T3U * T3V); Chris@82: T7Y = FNMS(T3U, T3T, T3S * T3V); Chris@82: } Chris@82: T3X = T3R + T3W; Chris@82: Tdw = T7X + T7Y; Chris@82: T7Z = T7X - T7Y; Chris@82: T82 = T3R - T3W; Chris@82: } Chris@82: { Chris@82: E T4p, T8e, T4u, T8f; Chris@82: { Chris@82: E T4m, T4o, T4l, T4n; Chris@82: T4m = cr[WS(rs, 13)]; Chris@82: T4o = ci[WS(rs, 13)]; Chris@82: T4l = W[24]; Chris@82: T4n = W[25]; Chris@82: T4p = FMA(T4l, T4m, T4n * T4o); Chris@82: T8e = FNMS(T4n, T4m, T4l * T4o); Chris@82: } Chris@82: { Chris@82: E T4r, T4t, T4q, T4s; Chris@82: T4r = cr[WS(rs, 45)]; Chris@82: T4t = ci[WS(rs, 45)]; Chris@82: T4q = W[88]; Chris@82: T4s = W[89]; Chris@82: T4u = FMA(T4q, T4r, T4s * T4t); Chris@82: T8f = FNMS(T4s, T4r, T4q * T4t); Chris@82: } Chris@82: T4v = T4p + T4u; Chris@82: TdB = T8e + T8f; Chris@82: T8b = T4p - T4u; Chris@82: T8g = T8e - T8f; Chris@82: } Chris@82: { Chris@82: E T42, T83, T47, T84; Chris@82: { Chris@82: E T3Z, T41, T3Y, T40; Chris@82: T3Z = cr[WS(rs, 21)]; Chris@82: T41 = ci[WS(rs, 21)]; Chris@82: T3Y = W[40]; Chris@82: T40 = W[41]; Chris@82: T42 = FMA(T3Y, T3Z, T40 * T41); Chris@82: T83 = FNMS(T40, T3Z, T3Y * T41); Chris@82: } Chris@82: { Chris@82: E T44, T46, T43, T45; Chris@82: T44 = cr[WS(rs, 53)]; Chris@82: T46 = ci[WS(rs, 53)]; Chris@82: T43 = W[104]; Chris@82: T45 = W[105]; Chris@82: T47 = FMA(T43, T44, T45 * T46); Chris@82: T84 = FNMS(T45, T44, T43 * T46); Chris@82: } Chris@82: T48 = T42 + T47; Chris@82: Tdx = T83 + T84; Chris@82: T80 = T42 - T47; Chris@82: T85 = T83 - T84; Chris@82: } Chris@82: { Chris@82: E T4e, T88, T4j, T89; Chris@82: { Chris@82: E T4b, T4d, T4a, T4c; Chris@82: T4b = cr[WS(rs, 61)]; Chris@82: T4d = ci[WS(rs, 61)]; Chris@82: T4a = W[120]; Chris@82: T4c = W[121]; Chris@82: T4e = FMA(T4a, T4b, T4c * T4d); Chris@82: T88 = FNMS(T4c, T4b, T4a * T4d); Chris@82: } Chris@82: { Chris@82: E T4g, T4i, T4f, T4h; Chris@82: T4g = cr[WS(rs, 29)]; Chris@82: T4i = ci[WS(rs, 29)]; Chris@82: T4f = W[56]; Chris@82: T4h = W[57]; Chris@82: T4j = FMA(T4f, T4g, T4h * T4i); Chris@82: T89 = FNMS(T4h, T4g, T4f * T4i); Chris@82: } Chris@82: T4k = T4e + T4j; Chris@82: TdA = T88 + T89; Chris@82: T8a = T88 - T89; Chris@82: T8d = T4e - T4j; Chris@82: } Chris@82: { Chris@82: E T49, T4w, TdC, TdD; Chris@82: T49 = T3X + T48; Chris@82: T4w = T4k + T4v; Chris@82: T4x = T49 + T4w; Chris@82: TfJ = T49 - T4w; Chris@82: TdC = TdA - TdB; Chris@82: TdD = T4k - T4v; Chris@82: TdE = TdC - TdD; Chris@82: TdM = TdD + TdC; Chris@82: } Chris@82: { Chris@82: E TfM, TfN, T81, T86; Chris@82: TfM = TdA + TdB; Chris@82: TfN = Tdw + Tdx; Chris@82: TfO = TfM - TfN; Chris@82: Tgu = TfN + TfM; Chris@82: T81 = T7Z + T80; Chris@82: T86 = T82 - T85; Chris@82: T87 = FMA(KP923879532, T81, KP382683432 * T86); Chris@82: T8u = FNMS(KP382683432, T81, KP923879532 * T86); Chris@82: } Chris@82: { Chris@82: E T8c, T8h, Tb8, Tb9; Chris@82: T8c = T8a + T8b; Chris@82: T8h = T8d - T8g; Chris@82: T8i = FNMS(KP382683432, T8h, KP923879532 * T8c); Chris@82: T8v = FMA(KP382683432, T8c, KP923879532 * T8h); Chris@82: Tb8 = T8d + T8g; Chris@82: Tb9 = T8a - T8b; Chris@82: Tba = FNMS(KP382683432, Tb9, KP923879532 * Tb8); Chris@82: Tbh = FMA(KP923879532, Tb9, KP382683432 * Tb8); Chris@82: } Chris@82: { Chris@82: E Tdv, Tdy, Tb5, Tb6; Chris@82: Tdv = T3X - T48; Chris@82: Tdy = Tdw - Tdx; Chris@82: Tdz = Tdv + Tdy; Chris@82: TdN = Tdv - Tdy; Chris@82: Tb5 = T7Z - T80; Chris@82: Tb6 = T82 + T85; Chris@82: Tb7 = FMA(KP382683432, Tb5, KP923879532 * Tb6); Chris@82: Tbg = FNMS(KP382683432, Tb6, KP923879532 * Tb5); Chris@82: } Chris@82: } Chris@82: { Chris@82: E T5u, Te2, T8Q, T8X, T62, TdY, T94, T99, T5F, Te3, T8T, T8Y, T5R, TdX, T93; Chris@82: E T96; Chris@82: { Chris@82: E T5o, T8V, T5t, T8W; Chris@82: { Chris@82: E T5l, T5n, T5k, T5m; Chris@82: T5l = cr[WS(rs, 3)]; Chris@82: T5n = ci[WS(rs, 3)]; Chris@82: T5k = W[4]; Chris@82: T5m = W[5]; Chris@82: T5o = FMA(T5k, T5l, T5m * T5n); Chris@82: T8V = FNMS(T5m, T5l, T5k * T5n); Chris@82: } Chris@82: { Chris@82: E T5q, T5s, T5p, T5r; Chris@82: T5q = cr[WS(rs, 35)]; Chris@82: T5s = ci[WS(rs, 35)]; Chris@82: T5p = W[68]; Chris@82: T5r = W[69]; Chris@82: T5t = FMA(T5p, T5q, T5r * T5s); Chris@82: T8W = FNMS(T5r, T5q, T5p * T5s); Chris@82: } Chris@82: T5u = T5o + T5t; Chris@82: Te2 = T8V + T8W; Chris@82: T8Q = T5o - T5t; Chris@82: T8X = T8V - T8W; Chris@82: } Chris@82: { Chris@82: E T5W, T97, T61, T98; Chris@82: { Chris@82: E T5T, T5V, T5S, T5U; Chris@82: T5T = cr[WS(rs, 11)]; Chris@82: T5V = ci[WS(rs, 11)]; Chris@82: T5S = W[20]; Chris@82: T5U = W[21]; Chris@82: T5W = FMA(T5S, T5T, T5U * T5V); Chris@82: T97 = FNMS(T5U, T5T, T5S * T5V); Chris@82: } Chris@82: { Chris@82: E T5Y, T60, T5X, T5Z; Chris@82: T5Y = cr[WS(rs, 43)]; Chris@82: T60 = ci[WS(rs, 43)]; Chris@82: T5X = W[84]; Chris@82: T5Z = W[85]; Chris@82: T61 = FMA(T5X, T5Y, T5Z * T60); Chris@82: T98 = FNMS(T5Z, T5Y, T5X * T60); Chris@82: } Chris@82: T62 = T5W + T61; Chris@82: TdY = T97 + T98; Chris@82: T94 = T5W - T61; Chris@82: T99 = T97 - T98; Chris@82: } Chris@82: { Chris@82: E T5z, T8R, T5E, T8S; Chris@82: { Chris@82: E T5w, T5y, T5v, T5x; Chris@82: T5w = cr[WS(rs, 19)]; Chris@82: T5y = ci[WS(rs, 19)]; Chris@82: T5v = W[36]; Chris@82: T5x = W[37]; Chris@82: T5z = FMA(T5v, T5w, T5x * T5y); Chris@82: T8R = FNMS(T5x, T5w, T5v * T5y); Chris@82: } Chris@82: { Chris@82: E T5B, T5D, T5A, T5C; Chris@82: T5B = cr[WS(rs, 51)]; Chris@82: T5D = ci[WS(rs, 51)]; Chris@82: T5A = W[100]; Chris@82: T5C = W[101]; Chris@82: T5E = FMA(T5A, T5B, T5C * T5D); Chris@82: T8S = FNMS(T5C, T5B, T5A * T5D); Chris@82: } Chris@82: T5F = T5z + T5E; Chris@82: Te3 = T8R + T8S; Chris@82: T8T = T8R - T8S; Chris@82: T8Y = T5z - T5E; Chris@82: } Chris@82: { Chris@82: E T5L, T91, T5Q, T92; Chris@82: { Chris@82: E T5I, T5K, T5H, T5J; Chris@82: T5I = cr[WS(rs, 59)]; Chris@82: T5K = ci[WS(rs, 59)]; Chris@82: T5H = W[116]; Chris@82: T5J = W[117]; Chris@82: T5L = FMA(T5H, T5I, T5J * T5K); Chris@82: T91 = FNMS(T5J, T5I, T5H * T5K); Chris@82: } Chris@82: { Chris@82: E T5N, T5P, T5M, T5O; Chris@82: T5N = cr[WS(rs, 27)]; Chris@82: T5P = ci[WS(rs, 27)]; Chris@82: T5M = W[52]; Chris@82: T5O = W[53]; Chris@82: T5Q = FMA(T5M, T5N, T5O * T5P); Chris@82: T92 = FNMS(T5O, T5N, T5M * T5P); Chris@82: } Chris@82: T5R = T5L + T5Q; Chris@82: TdX = T91 + T92; Chris@82: T93 = T91 - T92; Chris@82: T96 = T5L - T5Q; Chris@82: } Chris@82: { Chris@82: E T5G, T63, Te1, Te4; Chris@82: T5G = T5u + T5F; Chris@82: T63 = T5R + T62; Chris@82: T64 = T5G + T63; Chris@82: TfZ = T5G - T63; Chris@82: Te1 = T5u - T5F; Chris@82: Te4 = Te2 - Te3; Chris@82: Te5 = Te1 - Te4; Chris@82: Ted = Te1 + Te4; Chris@82: } Chris@82: { Chris@82: E TfS, TfT, T8U, T8Z; Chris@82: TfS = TdX + TdY; Chris@82: TfT = Te2 + Te3; Chris@82: TfU = TfS - TfT; Chris@82: Tgz = TfT + TfS; Chris@82: T8U = T8Q - T8T; Chris@82: T8Z = T8X + T8Y; Chris@82: T90 = FNMS(KP382683432, T8Z, KP923879532 * T8U); Chris@82: T9n = FMA(KP923879532, T8Z, KP382683432 * T8U); Chris@82: } Chris@82: { Chris@82: E T95, T9a, Tbr, Tbs; Chris@82: T95 = T93 + T94; Chris@82: T9a = T96 - T99; Chris@82: T9b = FMA(KP382683432, T95, KP923879532 * T9a); Chris@82: T9o = FNMS(KP382683432, T9a, KP923879532 * T95); Chris@82: Tbr = T96 + T99; Chris@82: Tbs = T93 - T94; Chris@82: Tbt = FNMS(KP382683432, Tbs, KP923879532 * Tbr); Chris@82: TbA = FMA(KP923879532, Tbs, KP382683432 * Tbr); Chris@82: } Chris@82: { Chris@82: E TdW, TdZ, Tbo, Tbp; Chris@82: TdW = T5R - T62; Chris@82: TdZ = TdX - TdY; Chris@82: Te0 = TdW + TdZ; Chris@82: Tee = TdZ - TdW; Chris@82: Tbo = T8X - T8Y; Chris@82: Tbp = T8Q + T8T; Chris@82: Tbq = FMA(KP382683432, Tbo, KP923879532 * Tbp); Chris@82: Tbz = FNMS(KP382683432, Tbp, KP923879532 * Tbo); Chris@82: } Chris@82: } Chris@82: { Chris@82: E T1t, Tgn, TgK, TgL, TgV, Th1, T30, Th0, T66, TgX, Tgw, TgE, TgB, TgF, Tgq; Chris@82: E TgM; Chris@82: { Chris@82: E TH, T1s, TgI, TgJ; Chris@82: TH = Tj + TG; Chris@82: T1s = T14 + T1r; Chris@82: T1t = TH + T1s; Chris@82: Tgn = TH - T1s; Chris@82: TgI = Tgy + Tgz; Chris@82: TgJ = Tgt + Tgu; Chris@82: TgK = TgI - TgJ; Chris@82: TgL = TgJ + TgI; Chris@82: } Chris@82: { Chris@82: E TgN, TgU, T2e, T2Z; Chris@82: TgN = Tfr + Tfq; Chris@82: TgU = TgO + TgT; Chris@82: TgV = TgN + TgU; Chris@82: Th1 = TgU - TgN; Chris@82: T2e = T1Q + T2d; Chris@82: T2Z = T2B + T2Y; Chris@82: T30 = T2e + T2Z; Chris@82: Th0 = T2e - T2Z; Chris@82: } Chris@82: { Chris@82: E T4y, T65, Tgs, Tgv; Chris@82: T4y = T3M + T4x; Chris@82: T65 = T5j + T64; Chris@82: T66 = T4y + T65; Chris@82: TgX = T65 - T4y; Chris@82: Tgs = T3M - T4x; Chris@82: Tgv = Tgt - Tgu; Chris@82: Tgw = Tgs + Tgv; Chris@82: TgE = Tgs - Tgv; Chris@82: } Chris@82: { Chris@82: E Tgx, TgA, Tgo, Tgp; Chris@82: Tgx = T5j - T64; Chris@82: TgA = Tgy - Tgz; Chris@82: TgB = Tgx - TgA; Chris@82: TgF = Tgx + TgA; Chris@82: Tgo = TfA + TfB; Chris@82: Tgp = Tfv + Tfw; Chris@82: Tgq = Tgo - Tgp; Chris@82: TgM = Tgp + Tgo; Chris@82: } Chris@82: { Chris@82: E T31, TgW, TgY, TgH; Chris@82: T31 = T1t + T30; Chris@82: ci[WS(rs, 31)] = T31 - T66; Chris@82: cr[0] = T31 + T66; Chris@82: TgW = TgM + TgV; Chris@82: cr[WS(rs, 32)] = TgL - TgW; Chris@82: ci[WS(rs, 63)] = TgL + TgW; Chris@82: TgY = TgV - TgM; Chris@82: cr[WS(rs, 48)] = TgX - TgY; Chris@82: ci[WS(rs, 47)] = TgX + TgY; Chris@82: TgH = T1t - T30; Chris@82: cr[WS(rs, 16)] = TgH - TgK; Chris@82: ci[WS(rs, 15)] = TgH + TgK; Chris@82: } Chris@82: { Chris@82: E Tgr, TgC, TgZ, Th2; Chris@82: Tgr = Tgn - Tgq; Chris@82: TgC = KP707106781 * (Tgw + TgB); Chris@82: ci[WS(rs, 23)] = Tgr - TgC; Chris@82: cr[WS(rs, 8)] = Tgr + TgC; Chris@82: TgZ = KP707106781 * (TgB - Tgw); Chris@82: Th2 = Th0 + Th1; Chris@82: cr[WS(rs, 56)] = TgZ - Th2; Chris@82: ci[WS(rs, 39)] = TgZ + Th2; Chris@82: } Chris@82: { Chris@82: E Th3, Th4, TgD, TgG; Chris@82: Th3 = KP707106781 * (TgF - TgE); Chris@82: Th4 = Th1 - Th0; Chris@82: cr[WS(rs, 40)] = Th3 - Th4; Chris@82: ci[WS(rs, 55)] = Th3 + Th4; Chris@82: TgD = Tgn + Tgq; Chris@82: TgG = KP707106781 * (TgE + TgF); Chris@82: cr[WS(rs, 24)] = TgD - TgG; Chris@82: ci[WS(rs, 7)] = TgD + TgG; Chris@82: } Chris@82: } Chris@82: { Chris@82: E T6L, T9x, ThV, Ti1, T7E, Ti0, T9A, ThO, T8y, T9K, T9u, T9E, T9r, T9L, T9v; Chris@82: E T9H; Chris@82: { Chris@82: E T6n, T6K, ThP, ThU; Chris@82: T6n = T6b + T6m; Chris@82: T6K = T6y + T6J; Chris@82: T6L = T6n - T6K; Chris@82: T9x = T6n + T6K; Chris@82: ThP = T9O - T9P; Chris@82: ThU = ThQ + ThT; Chris@82: ThV = ThP + ThU; Chris@82: Ti1 = ThU - ThP; Chris@82: } Chris@82: { Chris@82: E T7c, T9y, T7D, T9z; Chris@82: { Chris@82: E T72, T7b, T7t, T7C; Chris@82: T72 = T6Q + T71; Chris@82: T7b = T77 + T7a; Chris@82: T7c = FMA(KP195090322, T72, KP980785280 * T7b); Chris@82: T9y = FNMS(KP195090322, T7b, KP980785280 * T72); Chris@82: T7t = T7h + T7s; Chris@82: T7C = T7y + T7B; Chris@82: T7D = FNMS(KP980785280, T7C, KP195090322 * T7t); Chris@82: T9z = FMA(KP980785280, T7t, KP195090322 * T7C); Chris@82: } Chris@82: T7E = T7c + T7D; Chris@82: Ti0 = T9z - T9y; Chris@82: T9A = T9y + T9z; Chris@82: ThO = T7c - T7D; Chris@82: } Chris@82: { Chris@82: E T8k, T9D, T8x, T9C; Chris@82: { Chris@82: E T7W, T8j, T8t, T8w; Chris@82: T7W = T7K + T7V; Chris@82: T8j = T87 + T8i; Chris@82: T8k = T7W - T8j; Chris@82: T9D = T7W + T8j; Chris@82: T8t = T8p + T8s; Chris@82: T8w = T8u + T8v; Chris@82: T8x = T8t - T8w; Chris@82: T9C = T8t + T8w; Chris@82: } Chris@82: T8y = FMA(KP634393284, T8k, KP773010453 * T8x); Chris@82: T9K = FMA(KP995184726, T9D, KP098017140 * T9C); Chris@82: T9u = FNMS(KP773010453, T8k, KP634393284 * T8x); Chris@82: T9E = FNMS(KP098017140, T9D, KP995184726 * T9C); Chris@82: } Chris@82: { Chris@82: E T9d, T9G, T9q, T9F; Chris@82: { Chris@82: E T8P, T9c, T9m, T9p; Chris@82: T8P = T8D + T8O; Chris@82: T9c = T90 + T9b; Chris@82: T9d = T8P - T9c; Chris@82: T9G = T8P + T9c; Chris@82: T9m = T9i + T9l; Chris@82: T9p = T9n + T9o; Chris@82: T9q = T9m - T9p; Chris@82: T9F = T9m + T9p; Chris@82: } Chris@82: T9r = FNMS(KP634393284, T9q, KP773010453 * T9d); Chris@82: T9L = FNMS(KP995184726, T9F, KP098017140 * T9G); Chris@82: T9v = FMA(KP773010453, T9q, KP634393284 * T9d); Chris@82: T9H = FMA(KP098017140, T9F, KP995184726 * T9G); Chris@82: } Chris@82: { Chris@82: E T7F, T9s, ThZ, Ti2; Chris@82: T7F = T6L + T7E; Chris@82: T9s = T8y + T9r; Chris@82: ci[WS(rs, 24)] = T7F - T9s; Chris@82: cr[WS(rs, 7)] = T7F + T9s; Chris@82: ThZ = T9v - T9u; Chris@82: Ti2 = Ti0 + Ti1; Chris@82: cr[WS(rs, 39)] = ThZ - Ti2; Chris@82: ci[WS(rs, 56)] = ThZ + Ti2; Chris@82: } Chris@82: { Chris@82: E Ti3, Ti4, T9t, T9w; Chris@82: Ti3 = T9r - T8y; Chris@82: Ti4 = Ti1 - Ti0; Chris@82: cr[WS(rs, 55)] = Ti3 - Ti4; Chris@82: ci[WS(rs, 40)] = Ti3 + Ti4; Chris@82: T9t = T6L - T7E; Chris@82: T9w = T9u + T9v; Chris@82: cr[WS(rs, 23)] = T9t - T9w; Chris@82: ci[WS(rs, 8)] = T9t + T9w; Chris@82: } Chris@82: { Chris@82: E T9B, T9I, ThN, ThW; Chris@82: T9B = T9x + T9A; Chris@82: T9I = T9E + T9H; Chris@82: cr[WS(rs, 31)] = T9B - T9I; Chris@82: ci[0] = T9B + T9I; Chris@82: ThN = T9L - T9K; Chris@82: ThW = ThO + ThV; Chris@82: cr[WS(rs, 63)] = ThN - ThW; Chris@82: ci[WS(rs, 32)] = ThN + ThW; Chris@82: } Chris@82: { Chris@82: E ThX, ThY, T9J, T9M; Chris@82: ThX = T9H - T9E; Chris@82: ThY = ThV - ThO; Chris@82: cr[WS(rs, 47)] = ThX - ThY; Chris@82: ci[WS(rs, 48)] = ThX + ThY; Chris@82: T9J = T9x - T9A; Chris@82: T9M = T9K + T9L; Chris@82: ci[WS(rs, 16)] = T9J - T9M; Chris@82: cr[WS(rs, 15)] = T9J + T9M; Chris@82: } Chris@82: } Chris@82: { Chris@82: E Tft, Tg7, Tgh, Tgl, Th9, Thf, TfE, Th6, TfQ, Tg4, Tga, The, Tge, Tgk, Tg1; Chris@82: E Tg5; Chris@82: { Chris@82: E Tfp, Tfs, Tgf, Tgg; Chris@82: Tfp = Tj - TG; Chris@82: Tfs = Tfq - Tfr; Chris@82: Tft = Tfp - Tfs; Chris@82: Tg7 = Tfp + Tfs; Chris@82: Tgf = TfY + TfZ; Chris@82: Tgg = TfR + TfU; Chris@82: Tgh = FMA(KP382683432, Tgf, KP923879532 * Tgg); Chris@82: Tgl = FNMS(KP923879532, Tgf, KP382683432 * Tgg); Chris@82: } Chris@82: { Chris@82: E Th7, Th8, Tfy, TfD; Chris@82: Th7 = T14 - T1r; Chris@82: Th8 = TgT - TgO; Chris@82: Th9 = Th7 + Th8; Chris@82: Thf = Th8 - Th7; Chris@82: Tfy = Tfu + Tfx; Chris@82: TfD = Tfz - TfC; Chris@82: TfE = KP707106781 * (Tfy + TfD); Chris@82: Th6 = KP707106781 * (Tfy - TfD); Chris@82: } Chris@82: { Chris@82: E TfK, TfP, Tg8, Tg9; Chris@82: TfK = TfI - TfJ; Chris@82: TfP = TfL - TfO; Chris@82: TfQ = FMA(KP382683432, TfK, KP923879532 * TfP); Chris@82: Tg4 = FNMS(KP923879532, TfK, KP382683432 * TfP); Chris@82: Tg8 = Tfu - Tfx; Chris@82: Tg9 = Tfz + TfC; Chris@82: Tga = KP707106781 * (Tg8 + Tg9); Chris@82: The = KP707106781 * (Tg9 - Tg8); Chris@82: } Chris@82: { Chris@82: E Tgc, Tgd, TfV, Tg0; Chris@82: Tgc = TfL + TfO; Chris@82: Tgd = TfI + TfJ; Chris@82: Tge = FNMS(KP382683432, Tgd, KP923879532 * Tgc); Chris@82: Tgk = FMA(KP923879532, Tgd, KP382683432 * Tgc); Chris@82: TfV = TfR - TfU; Chris@82: Tg0 = TfY - TfZ; Chris@82: Tg1 = FNMS(KP382683432, Tg0, KP923879532 * TfV); Chris@82: Tg5 = FMA(KP923879532, Tg0, KP382683432 * TfV); Chris@82: } Chris@82: { Chris@82: E TfF, Tg2, Thd, Thg; Chris@82: TfF = Tft + TfE; Chris@82: Tg2 = TfQ + Tg1; Chris@82: ci[WS(rs, 27)] = TfF - Tg2; Chris@82: cr[WS(rs, 4)] = TfF + Tg2; Chris@82: Thd = Tg5 - Tg4; Chris@82: Thg = The + Thf; Chris@82: cr[WS(rs, 36)] = Thd - Thg; Chris@82: ci[WS(rs, 59)] = Thd + Thg; Chris@82: } Chris@82: { Chris@82: E Thh, Thi, Tg3, Tg6; Chris@82: Thh = Tg1 - TfQ; Chris@82: Thi = Thf - The; Chris@82: cr[WS(rs, 52)] = Thh - Thi; Chris@82: ci[WS(rs, 43)] = Thh + Thi; Chris@82: Tg3 = Tft - TfE; Chris@82: Tg6 = Tg4 + Tg5; Chris@82: cr[WS(rs, 20)] = Tg3 - Tg6; Chris@82: ci[WS(rs, 11)] = Tg3 + Tg6; Chris@82: } Chris@82: { Chris@82: E Tgb, Tgi, Th5, Tha; Chris@82: Tgb = Tg7 + Tga; Chris@82: Tgi = Tge + Tgh; Chris@82: cr[WS(rs, 28)] = Tgb - Tgi; Chris@82: ci[WS(rs, 3)] = Tgb + Tgi; Chris@82: Th5 = Tgl - Tgk; Chris@82: Tha = Th6 + Th9; Chris@82: cr[WS(rs, 60)] = Th5 - Tha; Chris@82: ci[WS(rs, 35)] = Th5 + Tha; Chris@82: } Chris@82: { Chris@82: E Thb, Thc, Tgj, Tgm; Chris@82: Thb = Tgh - Tge; Chris@82: Thc = Th9 - Th6; Chris@82: cr[WS(rs, 44)] = Thb - Thc; Chris@82: ci[WS(rs, 51)] = Thb + Thc; Chris@82: Tgj = Tg7 - Tga; Chris@82: Tgm = Tgk + Tgl; Chris@82: ci[WS(rs, 19)] = Tgj - Tgm; Chris@82: cr[WS(rs, 12)] = Tgj + Tgm; Chris@82: } Chris@82: } Chris@82: { Chris@82: E TeH, Tf9, TeO, Thk, Thp, Thv, Tfc, Thu, Tf3, Tfn, Tf7, Tfj, TeW, Tfm, Tf6; Chris@82: E Tfg; Chris@82: { Chris@82: E TeD, TeG, Tfa, Tfb; Chris@82: TeD = TcL + TcO; Chris@82: TeG = KP707106781 * (TeE + TeF); Chris@82: TeH = TeD - TeG; Chris@82: Tf9 = TeD + TeG; Chris@82: { Chris@82: E TeK, TeN, Thl, Tho; Chris@82: TeK = FMA(KP923879532, TeI, KP382683432 * TeJ); Chris@82: TeN = FNMS(KP923879532, TeM, KP382683432 * TeL); Chris@82: TeO = TeK + TeN; Chris@82: Thk = TeK - TeN; Chris@82: Thl = KP707106781 * (TcU - TcZ); Chris@82: Tho = Thm + Thn; Chris@82: Thp = Thl + Tho; Chris@82: Thv = Tho - Thl; Chris@82: } Chris@82: Tfa = FNMS(KP382683432, TeI, KP923879532 * TeJ); Chris@82: Tfb = FMA(KP382683432, TeM, KP923879532 * TeL); Chris@82: Tfc = Tfa + Tfb; Chris@82: Thu = Tfb - Tfa; Chris@82: { Chris@82: E TeZ, Tfh, Tf2, Tfi, TeY, Tf1; Chris@82: TeY = KP707106781 * (Te5 + Te0); Chris@82: TeZ = TeX - TeY; Chris@82: Tfh = TeX + TeY; Chris@82: Tf1 = KP707106781 * (Ted + Tee); Chris@82: Tf2 = Tf0 - Tf1; Chris@82: Tfi = Tf0 + Tf1; Chris@82: Tf3 = FNMS(KP555570233, Tf2, KP831469612 * TeZ); Chris@82: Tfn = FMA(KP980785280, Tfh, KP195090322 * Tfi); Chris@82: Tf7 = FMA(KP555570233, TeZ, KP831469612 * Tf2); Chris@82: Tfj = FNMS(KP980785280, Tfi, KP195090322 * Tfh); Chris@82: } Chris@82: { Chris@82: E TeS, Tfe, TeV, Tff, TeR, TeU; Chris@82: TeR = KP707106781 * (TdN + TdM); Chris@82: TeS = TeQ - TeR; Chris@82: Tfe = TeQ + TeR; Chris@82: TeU = KP707106781 * (Tdz + TdE); Chris@82: TeV = TeT - TeU; Chris@82: Tff = TeT + TeU; Chris@82: TeW = FMA(KP831469612, TeS, KP555570233 * TeV); Chris@82: Tfm = FNMS(KP195090322, Tff, KP980785280 * Tfe); Chris@82: Tf6 = FNMS(KP831469612, TeV, KP555570233 * TeS); Chris@82: Tfg = FMA(KP195090322, Tfe, KP980785280 * Tff); Chris@82: } Chris@82: } Chris@82: { Chris@82: E TeP, Tf4, Tht, Thw; Chris@82: TeP = TeH + TeO; Chris@82: Tf4 = TeW + Tf3; Chris@82: ci[WS(rs, 25)] = TeP - Tf4; Chris@82: cr[WS(rs, 6)] = TeP + Tf4; Chris@82: Tht = Tf7 - Tf6; Chris@82: Thw = Thu + Thv; Chris@82: cr[WS(rs, 38)] = Tht - Thw; Chris@82: ci[WS(rs, 57)] = Tht + Thw; Chris@82: } Chris@82: { Chris@82: E Thx, Thy, Tf5, Tf8; Chris@82: Thx = Tf3 - TeW; Chris@82: Thy = Thv - Thu; Chris@82: cr[WS(rs, 54)] = Thx - Thy; Chris@82: ci[WS(rs, 41)] = Thx + Thy; Chris@82: Tf5 = TeH - TeO; Chris@82: Tf8 = Tf6 + Tf7; Chris@82: cr[WS(rs, 22)] = Tf5 - Tf8; Chris@82: ci[WS(rs, 9)] = Tf5 + Tf8; Chris@82: } Chris@82: { Chris@82: E Tfd, Tfk, Thj, Thq; Chris@82: Tfd = Tf9 - Tfc; Chris@82: Tfk = Tfg + Tfj; Chris@82: ci[WS(rs, 17)] = Tfd - Tfk; Chris@82: cr[WS(rs, 14)] = Tfd + Tfk; Chris@82: Thj = Tfj - Tfg; Chris@82: Thq = Thk + Thp; Chris@82: cr[WS(rs, 62)] = Thj - Thq; Chris@82: ci[WS(rs, 33)] = Thj + Thq; Chris@82: } Chris@82: { Chris@82: E Thr, Ths, Tfl, Tfo; Chris@82: Thr = Tfn - Tfm; Chris@82: Ths = Thp - Thk; Chris@82: cr[WS(rs, 46)] = Thr - Ths; Chris@82: ci[WS(rs, 49)] = Thr + Ths; Chris@82: Tfl = Tf9 + Tfc; Chris@82: Tfo = Tfm + Tfn; Chris@82: cr[WS(rs, 30)] = Tfl - Tfo; Chris@82: ci[WS(rs, 1)] = Tfl + Tfo; Chris@82: } Chris@82: } Chris@82: { Chris@82: E Td1, Ten, Tdo, ThA, ThD, ThJ, Teq, ThI, Teh, TeB, Tel, Tex, TdQ, TeA, Tek; Chris@82: E Teu; Chris@82: { Chris@82: E TcP, Td0, Teo, Tep; Chris@82: TcP = TcL - TcO; Chris@82: Td0 = KP707106781 * (TcU + TcZ); Chris@82: Td1 = TcP - Td0; Chris@82: Ten = TcP + Td0; Chris@82: { Chris@82: E Tdc, Tdn, ThB, ThC; Chris@82: Tdc = FNMS(KP923879532, Tdb, KP382683432 * Td6); Chris@82: Tdn = FMA(KP923879532, Tdh, KP382683432 * Tdm); Chris@82: Tdo = Tdc + Tdn; Chris@82: ThA = Tdn - Tdc; Chris@82: ThB = KP707106781 * (TeF - TeE); Chris@82: ThC = Thn - Thm; Chris@82: ThD = ThB + ThC; Chris@82: ThJ = ThC - ThB; Chris@82: } Chris@82: Teo = FMA(KP382683432, Tdb, KP923879532 * Td6); Chris@82: Tep = FNMS(KP382683432, Tdh, KP923879532 * Tdm); Chris@82: Teq = Teo + Tep; Chris@82: ThI = Teo - Tep; Chris@82: { Chris@82: E Te7, Tew, Teg, Tev, Te6, Tef; Chris@82: Te6 = KP707106781 * (Te0 - Te5); Chris@82: Te7 = TdV - Te6; Chris@82: Tew = TdV + Te6; Chris@82: Tef = KP707106781 * (Ted - Tee); Chris@82: Teg = Tec - Tef; Chris@82: Tev = Tec + Tef; Chris@82: Teh = FMA(KP555570233, Te7, KP831469612 * Teg); Chris@82: TeB = FMA(KP980785280, Tew, KP195090322 * Tev); Chris@82: Tel = FNMS(KP831469612, Te7, KP555570233 * Teg); Chris@82: Tex = FNMS(KP195090322, Tew, KP980785280 * Tev); Chris@82: } Chris@82: { Chris@82: E TdG, Tet, TdP, Tes, TdF, TdO; Chris@82: TdF = KP707106781 * (Tdz - TdE); Chris@82: TdG = Tdu - TdF; Chris@82: Tet = Tdu + TdF; Chris@82: TdO = KP707106781 * (TdM - TdN); Chris@82: TdP = TdL - TdO; Chris@82: Tes = TdL + TdO; Chris@82: TdQ = FNMS(KP555570233, TdP, KP831469612 * TdG); Chris@82: TeA = FNMS(KP980785280, Tes, KP195090322 * Tet); Chris@82: Tek = FMA(KP831469612, TdP, KP555570233 * TdG); Chris@82: Teu = FMA(KP195090322, Tes, KP980785280 * Tet); Chris@82: } Chris@82: } Chris@82: { Chris@82: E Tdp, Tei, ThH, ThK; Chris@82: Tdp = Td1 + Tdo; Chris@82: Tei = TdQ + Teh; Chris@82: cr[WS(rs, 26)] = Tdp - Tei; Chris@82: ci[WS(rs, 5)] = Tdp + Tei; Chris@82: ThH = Tel - Tek; Chris@82: ThK = ThI + ThJ; Chris@82: cr[WS(rs, 58)] = ThH - ThK; Chris@82: ci[WS(rs, 37)] = ThH + ThK; Chris@82: } Chris@82: { Chris@82: E ThL, ThM, Tej, Tem; Chris@82: ThL = Teh - TdQ; Chris@82: ThM = ThJ - ThI; Chris@82: cr[WS(rs, 42)] = ThL - ThM; Chris@82: ci[WS(rs, 53)] = ThL + ThM; Chris@82: Tej = Td1 - Tdo; Chris@82: Tem = Tek + Tel; Chris@82: ci[WS(rs, 21)] = Tej - Tem; Chris@82: cr[WS(rs, 10)] = Tej + Tem; Chris@82: } Chris@82: { Chris@82: E Ter, Tey, Thz, ThE; Chris@82: Ter = Ten + Teq; Chris@82: Tey = Teu + Tex; Chris@82: ci[WS(rs, 29)] = Ter - Tey; Chris@82: cr[WS(rs, 2)] = Ter + Tey; Chris@82: Thz = TeB - TeA; Chris@82: ThE = ThA + ThD; Chris@82: cr[WS(rs, 34)] = Thz - ThE; Chris@82: ci[WS(rs, 61)] = Thz + ThE; Chris@82: } Chris@82: { Chris@82: E ThF, ThG, Tez, TeC; Chris@82: ThF = Tex - Teu; Chris@82: ThG = ThD - ThA; Chris@82: cr[WS(rs, 50)] = ThF - ThG; Chris@82: ci[WS(rs, 45)] = ThF + ThG; Chris@82: Tez = Ten - Teq; Chris@82: TeC = TeA + TeB; Chris@82: cr[WS(rs, 18)] = Tez - TeC; Chris@82: ci[WS(rs, 13)] = Tez + TeC; Chris@82: } Chris@82: } Chris@82: { Chris@82: E Tc3, Tcv, TiD, TiJ, Tca, TiI, Tcy, TiA, Tci, TcI, Tcs, TcC, Tcp, TcJ, Tct; Chris@82: E TcF; Chris@82: { Chris@82: E TbZ, Tc2, TiB, TiC; Chris@82: TbZ = Taz - TaC; Chris@82: Tc2 = Tc0 + Tc1; Chris@82: Tc3 = TbZ - Tc2; Chris@82: Tcv = TbZ + Tc2; Chris@82: TiB = TaG - TaJ; Chris@82: TiC = Tin - Tim; Chris@82: TiD = TiB + TiC; Chris@82: TiJ = TiC - TiB; Chris@82: } Chris@82: { Chris@82: E Tc6, Tcw, Tc9, Tcx; Chris@82: { Chris@82: E Tc4, Tc5, Tc7, Tc8; Chris@82: Tc4 = TaP - TaQ; Chris@82: Tc5 = TaM - TaN; Chris@82: Tc6 = FMA(KP831469612, Tc4, KP555570233 * Tc5); Chris@82: Tcw = FNMS(KP555570233, Tc4, KP831469612 * Tc5); Chris@82: Tc7 = TaW - TaX; Chris@82: Tc8 = TaT - TaU; Chris@82: Tc9 = FNMS(KP831469612, Tc8, KP555570233 * Tc7); Chris@82: Tcx = FMA(KP555570233, Tc8, KP831469612 * Tc7); Chris@82: } Chris@82: Tca = Tc6 + Tc9; Chris@82: TiI = Tcx - Tcw; Chris@82: Tcy = Tcw + Tcx; Chris@82: TiA = Tc6 - Tc9; Chris@82: } Chris@82: { Chris@82: E Tce, TcB, Tch, TcA; Chris@82: { Chris@82: E Tcc, Tcd, Tcf, Tcg; Chris@82: Tcc = Tbd - Tbe; Chris@82: Tcd = Tb7 - Tba; Chris@82: Tce = Tcc - Tcd; Chris@82: TcB = Tcc + Tcd; Chris@82: Tcf = Tb2 - Tb3; Chris@82: Tcg = Tbh - Tbg; Chris@82: Tch = Tcf - Tcg; Chris@82: TcA = Tcf + Tcg; Chris@82: } Chris@82: Tci = FMA(KP471396736, Tce, KP881921264 * Tch); Chris@82: TcI = FMA(KP956940335, TcB, KP290284677 * TcA); Chris@82: Tcs = FNMS(KP881921264, Tce, KP471396736 * Tch); Chris@82: TcC = FNMS(KP290284677, TcB, KP956940335 * TcA); Chris@82: } Chris@82: { Chris@82: E Tcl, TcE, Tco, TcD; Chris@82: { Chris@82: E Tcj, Tck, Tcm, Tcn; Chris@82: Tcj = Tbl - Tbm; Chris@82: Tck = TbA - Tbz; Chris@82: Tcl = Tcj - Tck; Chris@82: TcE = Tcj + Tck; Chris@82: Tcm = Tbw - Tbx; Chris@82: Tcn = Tbq - Tbt; Chris@82: Tco = Tcm - Tcn; Chris@82: TcD = Tcm + Tcn; Chris@82: } Chris@82: Tcp = FNMS(KP471396736, Tco, KP881921264 * Tcl); Chris@82: TcJ = FNMS(KP956940335, TcD, KP290284677 * TcE); Chris@82: Tct = FMA(KP881921264, Tco, KP471396736 * Tcl); Chris@82: TcF = FMA(KP290284677, TcD, KP956940335 * TcE); Chris@82: } Chris@82: { Chris@82: E Tcb, Tcq, TiH, TiK; Chris@82: Tcb = Tc3 + Tca; Chris@82: Tcq = Tci + Tcp; Chris@82: ci[WS(rs, 26)] = Tcb - Tcq; Chris@82: cr[WS(rs, 5)] = Tcb + Tcq; Chris@82: TiH = Tct - Tcs; Chris@82: TiK = TiI + TiJ; Chris@82: cr[WS(rs, 37)] = TiH - TiK; Chris@82: ci[WS(rs, 58)] = TiH + TiK; Chris@82: } Chris@82: { Chris@82: E TiL, TiM, Tcr, Tcu; Chris@82: TiL = Tcp - Tci; Chris@82: TiM = TiJ - TiI; Chris@82: cr[WS(rs, 53)] = TiL - TiM; Chris@82: ci[WS(rs, 42)] = TiL + TiM; Chris@82: Tcr = Tc3 - Tca; Chris@82: Tcu = Tcs + Tct; Chris@82: cr[WS(rs, 21)] = Tcr - Tcu; Chris@82: ci[WS(rs, 10)] = Tcr + Tcu; Chris@82: } Chris@82: { Chris@82: E Tcz, TcG, Tiz, TiE; Chris@82: Tcz = Tcv + Tcy; Chris@82: TcG = TcC + TcF; Chris@82: cr[WS(rs, 29)] = Tcz - TcG; Chris@82: ci[WS(rs, 2)] = Tcz + TcG; Chris@82: Tiz = TcJ - TcI; Chris@82: TiE = TiA + TiD; Chris@82: cr[WS(rs, 61)] = Tiz - TiE; Chris@82: ci[WS(rs, 34)] = Tiz + TiE; Chris@82: } Chris@82: { Chris@82: E TiF, TiG, TcH, TcK; Chris@82: TiF = TcF - TcC; Chris@82: TiG = TiD - TiA; Chris@82: cr[WS(rs, 45)] = TiF - TiG; Chris@82: ci[WS(rs, 50)] = TiF + TiG; Chris@82: TcH = Tcv - Tcy; Chris@82: TcK = TcI + TcJ; Chris@82: ci[WS(rs, 18)] = TcH - TcK; Chris@82: cr[WS(rs, 13)] = TcH + TcK; Chris@82: } Chris@82: } Chris@82: { Chris@82: E TaL, TbJ, Tip, Tiv, Tb0, Tiu, TbM, Tik, Tbk, TbW, TbG, TbQ, TbD, TbX, TbH; Chris@82: E TbT; Chris@82: { Chris@82: E TaD, TaK, Til, Tio; Chris@82: TaD = Taz + TaC; Chris@82: TaK = TaG + TaJ; Chris@82: TaL = TaD - TaK; Chris@82: TbJ = TaD + TaK; Chris@82: Til = Tc1 - Tc0; Chris@82: Tio = Tim + Tin; Chris@82: Tip = Til + Tio; Chris@82: Tiv = Tio - Til; Chris@82: } Chris@82: { Chris@82: E TaS, TbK, TaZ, TbL; Chris@82: { Chris@82: E TaO, TaR, TaV, TaY; Chris@82: TaO = TaM + TaN; Chris@82: TaR = TaP + TaQ; Chris@82: TaS = FNMS(KP980785280, TaR, KP195090322 * TaO); Chris@82: TbK = FMA(KP195090322, TaR, KP980785280 * TaO); Chris@82: TaV = TaT + TaU; Chris@82: TaY = TaW + TaX; Chris@82: TaZ = FMA(KP980785280, TaV, KP195090322 * TaY); Chris@82: TbL = FNMS(KP195090322, TaV, KP980785280 * TaY); Chris@82: } Chris@82: Tb0 = TaS + TaZ; Chris@82: Tiu = TbK - TbL; Chris@82: TbM = TbK + TbL; Chris@82: Tik = TaZ - TaS; Chris@82: } Chris@82: { Chris@82: E Tbc, TbO, Tbj, TbP; Chris@82: { Chris@82: E Tb4, Tbb, Tbf, Tbi; Chris@82: Tb4 = Tb2 + Tb3; Chris@82: Tbb = Tb7 + Tba; Chris@82: Tbc = Tb4 - Tbb; Chris@82: TbO = Tb4 + Tbb; Chris@82: Tbf = Tbd + Tbe; Chris@82: Tbi = Tbg + Tbh; Chris@82: Tbj = Tbf - Tbi; Chris@82: TbP = Tbf + Tbi; Chris@82: } Chris@82: Tbk = FMA(KP634393284, Tbc, KP773010453 * Tbj); Chris@82: TbW = FNMS(KP995184726, TbP, KP098017140 * TbO); Chris@82: TbG = FNMS(KP634393284, Tbj, KP773010453 * Tbc); Chris@82: TbQ = FMA(KP995184726, TbO, KP098017140 * TbP); Chris@82: } Chris@82: { Chris@82: E Tbv, TbR, TbC, TbS; Chris@82: { Chris@82: E Tbn, Tbu, Tby, TbB; Chris@82: Tbn = Tbl + Tbm; Chris@82: Tbu = Tbq + Tbt; Chris@82: Tbv = Tbn - Tbu; Chris@82: TbR = Tbn + Tbu; Chris@82: Tby = Tbw + Tbx; Chris@82: TbB = Tbz + TbA; Chris@82: TbC = Tby - TbB; Chris@82: TbS = Tby + TbB; Chris@82: } Chris@82: TbD = FNMS(KP773010453, TbC, KP634393284 * Tbv); Chris@82: TbX = FMA(KP098017140, TbR, KP995184726 * TbS); Chris@82: TbH = FMA(KP773010453, Tbv, KP634393284 * TbC); Chris@82: TbT = FNMS(KP098017140, TbS, KP995184726 * TbR); Chris@82: } Chris@82: { Chris@82: E Tb1, TbE, Tit, Tiw; Chris@82: Tb1 = TaL - Tb0; Chris@82: TbE = Tbk + TbD; Chris@82: ci[WS(rs, 22)] = Tb1 - TbE; Chris@82: cr[WS(rs, 9)] = Tb1 + TbE; Chris@82: Tit = TbD - Tbk; Chris@82: Tiw = Tiu + Tiv; Chris@82: cr[WS(rs, 57)] = Tit - Tiw; Chris@82: ci[WS(rs, 38)] = Tit + Tiw; Chris@82: } Chris@82: { Chris@82: E Tix, Tiy, TbF, TbI; Chris@82: Tix = TbH - TbG; Chris@82: Tiy = Tiv - Tiu; Chris@82: cr[WS(rs, 41)] = Tix - Tiy; Chris@82: ci[WS(rs, 54)] = Tix + Tiy; Chris@82: TbF = TaL + Tb0; Chris@82: TbI = TbG + TbH; Chris@82: cr[WS(rs, 25)] = TbF - TbI; Chris@82: ci[WS(rs, 6)] = TbF + TbI; Chris@82: } Chris@82: { Chris@82: E TbN, TbU, Tij, Tiq; Chris@82: TbN = TbJ + TbM; Chris@82: TbU = TbQ + TbT; Chris@82: ci[WS(rs, 30)] = TbN - TbU; Chris@82: cr[WS(rs, 1)] = TbN + TbU; Chris@82: Tij = TbX - TbW; Chris@82: Tiq = Tik + Tip; Chris@82: cr[WS(rs, 33)] = Tij - Tiq; Chris@82: ci[WS(rs, 62)] = Tij + Tiq; Chris@82: } Chris@82: { Chris@82: E Tir, Tis, TbV, TbY; Chris@82: Tir = TbT - TbQ; Chris@82: Tis = Tip - Tik; Chris@82: cr[WS(rs, 49)] = Tir - Tis; Chris@82: ci[WS(rs, 46)] = Tir + Tis; Chris@82: TbV = TbJ - TbM; Chris@82: TbY = TbW + TbX; Chris@82: cr[WS(rs, 17)] = TbV - TbY; Chris@82: ci[WS(rs, 14)] = TbV + TbY; Chris@82: } Chris@82: } Chris@82: { Chris@82: E T9R, Taj, Ti9, Tif, T9Y, Tie, Tam, Ti6, Ta6, Taw, Tag, Taq, Tad, Tax, Tah; Chris@82: E Tat; Chris@82: { Chris@82: E T9N, T9Q, Ti7, Ti8; Chris@82: T9N = T6b - T6m; Chris@82: T9Q = T9O + T9P; Chris@82: T9R = T9N - T9Q; Chris@82: Taj = T9N + T9Q; Chris@82: Ti7 = T6J - T6y; Chris@82: Ti8 = ThT - ThQ; Chris@82: Ti9 = Ti7 + Ti8; Chris@82: Tif = Ti8 - Ti7; Chris@82: } Chris@82: { Chris@82: E T9U, Tak, T9X, Tal; Chris@82: { Chris@82: E T9S, T9T, T9V, T9W; Chris@82: T9S = T6Q - T71; Chris@82: T9T = T77 - T7a; Chris@82: T9U = FNMS(KP831469612, T9T, KP555570233 * T9S); Chris@82: Tak = FMA(KP831469612, T9S, KP555570233 * T9T); Chris@82: T9V = T7h - T7s; Chris@82: T9W = T7y - T7B; Chris@82: T9X = FMA(KP555570233, T9V, KP831469612 * T9W); Chris@82: Tal = FNMS(KP555570233, T9W, KP831469612 * T9V); Chris@82: } Chris@82: T9Y = T9U + T9X; Chris@82: Tie = Tak - Tal; Chris@82: Tam = Tak + Tal; Chris@82: Ti6 = T9X - T9U; Chris@82: } Chris@82: { Chris@82: E Ta2, Tao, Ta5, Tap; Chris@82: { Chris@82: E Ta0, Ta1, Ta3, Ta4; Chris@82: Ta0 = T8p - T8s; Chris@82: Ta1 = T87 - T8i; Chris@82: Ta2 = Ta0 - Ta1; Chris@82: Tao = Ta0 + Ta1; Chris@82: Ta3 = T7K - T7V; Chris@82: Ta4 = T8v - T8u; Chris@82: Ta5 = Ta3 - Ta4; Chris@82: Tap = Ta3 + Ta4; Chris@82: } Chris@82: Ta6 = FMA(KP471396736, Ta2, KP881921264 * Ta5); Chris@82: Taw = FNMS(KP956940335, Tap, KP290284677 * Tao); Chris@82: Tag = FNMS(KP471396736, Ta5, KP881921264 * Ta2); Chris@82: Taq = FMA(KP956940335, Tao, KP290284677 * Tap); Chris@82: } Chris@82: { Chris@82: E Ta9, Tar, Tac, Tas; Chris@82: { Chris@82: E Ta7, Ta8, Taa, Tab; Chris@82: Ta7 = T8D - T8O; Chris@82: Ta8 = T9n - T9o; Chris@82: Ta9 = Ta7 - Ta8; Chris@82: Tar = Ta7 + Ta8; Chris@82: Taa = T9i - T9l; Chris@82: Tab = T9b - T90; Chris@82: Tac = Taa - Tab; Chris@82: Tas = Taa + Tab; Chris@82: } Chris@82: Tad = FNMS(KP881921264, Tac, KP471396736 * Ta9); Chris@82: Tax = FMA(KP290284677, Tar, KP956940335 * Tas); Chris@82: Tah = FMA(KP881921264, Ta9, KP471396736 * Tac); Chris@82: Tat = FNMS(KP290284677, Tas, KP956940335 * Tar); Chris@82: } Chris@82: { Chris@82: E T9Z, Tae, Tid, Tig; Chris@82: T9Z = T9R - T9Y; Chris@82: Tae = Ta6 + Tad; Chris@82: ci[WS(rs, 20)] = T9Z - Tae; Chris@82: cr[WS(rs, 11)] = T9Z + Tae; Chris@82: Tid = Tad - Ta6; Chris@82: Tig = Tie + Tif; Chris@82: cr[WS(rs, 59)] = Tid - Tig; Chris@82: ci[WS(rs, 36)] = Tid + Tig; Chris@82: } Chris@82: { Chris@82: E Tih, Tii, Taf, Tai; Chris@82: Tih = Tah - Tag; Chris@82: Tii = Tif - Tie; Chris@82: cr[WS(rs, 43)] = Tih - Tii; Chris@82: ci[WS(rs, 52)] = Tih + Tii; Chris@82: Taf = T9R + T9Y; Chris@82: Tai = Tag + Tah; Chris@82: cr[WS(rs, 27)] = Taf - Tai; Chris@82: ci[WS(rs, 4)] = Taf + Tai; Chris@82: } Chris@82: { Chris@82: E Tan, Tau, Ti5, Tia; Chris@82: Tan = Taj + Tam; Chris@82: Tau = Taq + Tat; Chris@82: ci[WS(rs, 28)] = Tan - Tau; Chris@82: cr[WS(rs, 3)] = Tan + Tau; Chris@82: Ti5 = Tax - Taw; Chris@82: Tia = Ti6 + Ti9; Chris@82: cr[WS(rs, 35)] = Ti5 - Tia; Chris@82: ci[WS(rs, 60)] = Ti5 + Tia; Chris@82: } Chris@82: { Chris@82: E Tib, Tic, Tav, Tay; Chris@82: Tib = Tat - Taq; Chris@82: Tic = Ti9 - Ti6; Chris@82: cr[WS(rs, 51)] = Tib - Tic; Chris@82: ci[WS(rs, 44)] = Tib + Tic; Chris@82: Tav = Taj - Tam; Chris@82: Tay = Taw + Tax; Chris@82: cr[WS(rs, 19)] = Tav - Tay; Chris@82: ci[WS(rs, 12)] = Tav + Tay; Chris@82: } Chris@82: } Chris@82: } Chris@82: } Chris@82: } Chris@82: Chris@82: static const tw_instr twinstr[] = { Chris@82: {TW_FULL, 1, 64}, Chris@82: {TW_NEXT, 1, 0} Chris@82: }; Chris@82: Chris@82: static const hc2hc_desc desc = { 64, "hf_64", twinstr, &GENUS, {808, 270, 230, 0} }; Chris@82: Chris@82: void X(codelet_hf_64) (planner *p) { Chris@82: X(khc2hc_register) (p, hf_64, &desc); Chris@82: } Chris@82: #endif