Chris@10: /* Chris@10: * Copyright (c) 2003, 2007-11 Matteo Frigo Chris@10: * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology Chris@10: * Chris@10: * This program is free software; you can redistribute it and/or modify Chris@10: * it under the terms of the GNU General Public License as published by Chris@10: * the Free Software Foundation; either version 2 of the License, or Chris@10: * (at your option) any later version. Chris@10: * Chris@10: * This program is distributed in the hope that it will be useful, Chris@10: * but WITHOUT ANY WARRANTY; without even the implied warranty of Chris@10: * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the Chris@10: * GNU General Public License for more details. Chris@10: * Chris@10: * You should have received a copy of the GNU General Public License Chris@10: * along with this program; if not, write to the Free Software Chris@10: * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA Chris@10: * Chris@10: */ Chris@10: Chris@10: /* This file was automatically generated --- DO NOT EDIT */ Chris@10: /* Generated on Sun Nov 25 07:37:08 EST 2012 */ Chris@10: Chris@10: #include "codelet-dft.h" Chris@10: Chris@10: #ifdef HAVE_FMA Chris@10: Chris@10: /* Generated by: ../../../genfft/gen_notw_c.native -fma -reorder-insns -schedule-for-pipeline -simd -compact -variables 4 -pipeline-latency 8 -sign 1 -n 128 -name n1bv_128 -include n1b.h */ Chris@10: Chris@10: /* Chris@10: * This function contains 1082 FP additions, 642 FP multiplications, Chris@10: * (or, 440 additions, 0 multiplications, 642 fused multiply/add), Chris@10: * 295 stack variables, 31 constants, and 256 memory accesses Chris@10: */ Chris@10: #include "n1b.h" Chris@10: Chris@10: static void n1bv_128(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs) Chris@10: { Chris@10: DVK(KP903989293, +0.903989293123443331586200297230537048710132025); Chris@10: DVK(KP941544065, +0.941544065183020778412509402599502357185589796); Chris@10: DVK(KP357805721, +0.357805721314524104672487743774474392487532769); Chris@10: DVK(KP472964775, +0.472964775891319928124438237972992463904131113); Chris@10: DVK(KP857728610, +0.857728610000272069902269984284770137042490799); Chris@10: DVK(KP970031253, +0.970031253194543992603984207286100251456865962); Chris@10: DVK(KP250486960, +0.250486960191305461595702160124721208578685568); Chris@10: DVK(KP998795456, +0.998795456205172392714771604759100694443203615); Chris@10: DVK(KP740951125, +0.740951125354959091175616897495162729728955309); Chris@10: DVK(KP599376933, +0.599376933681923766271389869014404232837890546); Chris@10: DVK(KP906347169, +0.906347169019147157946142717268914412664134293); Chris@10: DVK(KP049126849, +0.049126849769467254105343321271313617079695752); Chris@10: DVK(KP989176509, +0.989176509964780973451673738016243063983689533); Chris@10: DVK(KP803207531, +0.803207531480644909806676512963141923879569427); Chris@10: DVK(KP741650546, +0.741650546272035369581266691172079863842265220); Chris@10: DVK(KP956940335, +0.956940335732208864935797886980269969482849206); Chris@10: DVK(KP148335987, +0.148335987538347428753676511486911367000625355); Chris@10: DVK(KP881921264, +0.881921264348355029712756863660388349508442621); Chris@10: DVK(KP303346683, +0.303346683607342391675883946941299872384187453); Chris@10: DVK(KP534511135, +0.534511135950791641089685961295362908582039528); Chris@10: DVK(KP773010453, +0.773010453362736960810906609758469800971041293); Chris@10: DVK(KP995184726, +0.995184726672196886244836953109479921575474869); Chris@10: DVK(KP820678790, +0.820678790828660330972281985331011598767386482); Chris@10: DVK(KP098491403, +0.098491403357164253077197521291327432293052451); Chris@10: DVK(KP831469612, +0.831469612302545237078788377617905756738560812); Chris@10: DVK(KP980785280, +0.980785280403230449126182236134239036973933731); Chris@10: DVK(KP198912367, +0.198912367379658006911597622644676228597850501); Chris@10: DVK(KP668178637, +0.668178637919298919997757686523080761552472251); Chris@10: DVK(KP923879532, +0.923879532511286756128183189396788286822416626); Chris@10: DVK(KP414213562, +0.414213562373095048801688724209698078569671875); Chris@10: DVK(KP707106781, +0.707106781186547524400844362104849039284835938); Chris@10: { Chris@10: INT i; Chris@10: const R *xi; Chris@10: R *xo; Chris@10: xi = ii; Chris@10: xo = io; Chris@10: for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(256, is), MAKE_VOLATILE_STRIDE(256, os)) { Chris@10: V T6a, T5J, T6b, T5K, T6B, T6C, T6J, T6A, T6o, T6j, T6r, T68, T6d, T5O, T5R; Chris@10: V T6e, T6D, T6K; Chris@10: { Chris@10: V Tad, TcZ, T6Z, T8T, T4U, Tr, Tfq, TgG, Ted, Tgf, Td0, Tcc, T9k, T84, Tb6; Chris@10: V Tbt, Td8, TdK, TeK, Tgq, TeV, Tgt, T7q, T94, T3p, T5X, T7B, T97, T2G, T5U; Chris@10: V TbD, Tc0, Tdf, TdN, Tf5, Tgx, Tfg, TgA, T7J, T9b, T4E, T64, T7U, T9e, T3V; Chris@10: V T61, Td2, Td3, T85, T72, T4V, TI, Tcd, Tas, TgH, Tek, Tgg, Tft, T86, T75; Chris@10: V T4W, TZ, TaI, Tcf, Tdo, TdG, Tgi, Tet, Tgj, Teq, T8X, T7a, T5M, T1B, T8W; Chris@10: V T7d, T5N, T1s, TaX, Tcg, Tdr, TdH, Tgl, TeC, Tgm, Tez, T90, T7h, T5P, T2c; Chris@10: V T8Z, T7k, T5Q, T23, T3Y, T49, TdL, Tdb, Tbu, Tbl, Tgu, TeR, Tgr, TeY, Tf6; Chris@10: V TbG, T5V, T3s, T5Y, T3f, T95, T7E, T98, T7x, T4g, T4f, T4q, TbH, T41, TbI; Chris@10: V T44, T4h, T4j, T4k, Tf9, TbN; Chris@10: { Chris@10: V Tu, TF, Ty, TL, TW, Tah, Tx, Tag, Tee, Tz, TM, TN, Teh, Tan, TP; Chris@10: V TQ; Chris@10: { Chris@10: V TeG, T2A, Tbq, TeT, Tbp, TeH, T3m, T2x, Td6, T7o, T2q, T3l, T7z, Tbr, T2D; Chris@10: V T82, T83; Chris@10: { Chris@10: V Ta7, T3, Ta8, T4O, Taa, Tab, Ta, T4P, Te, Tc6, Th, Tc7, Tl, Tc9, Tca; Chris@10: V To; Chris@10: { Chris@10: V T1, T2, T4M, T4N; Chris@10: T1 = LD(&(xi[0]), ivs, &(xi[0])); Chris@10: T2 = LD(&(xi[WS(is, 64)]), ivs, &(xi[0])); Chris@10: T4M = LD(&(xi[WS(is, 32)]), ivs, &(xi[0])); Chris@10: T4N = LD(&(xi[WS(is, 96)]), ivs, &(xi[0])); Chris@10: { Chris@10: V T4, T5, T7, T8; Chris@10: T4 = LD(&(xi[WS(is, 16)]), ivs, &(xi[0])); Chris@10: T5 = LD(&(xi[WS(is, 80)]), ivs, &(xi[0])); Chris@10: T7 = LD(&(xi[WS(is, 112)]), ivs, &(xi[0])); Chris@10: T8 = LD(&(xi[WS(is, 48)]), ivs, &(xi[0])); Chris@10: { Chris@10: V Tc, T6, T9, Td, Tf, Tg; Chris@10: Tc = LD(&(xi[WS(is, 8)]), ivs, &(xi[0])); Chris@10: Ta7 = VADD(T1, T2); Chris@10: T3 = VSUB(T1, T2); Chris@10: Ta8 = VADD(T4M, T4N); Chris@10: T4O = VSUB(T4M, T4N); Chris@10: Taa = VADD(T4, T5); Chris@10: T6 = VSUB(T4, T5); Chris@10: Tab = VADD(T7, T8); Chris@10: T9 = VSUB(T7, T8); Chris@10: Td = LD(&(xi[WS(is, 72)]), ivs, &(xi[0])); Chris@10: Tf = LD(&(xi[WS(is, 40)]), ivs, &(xi[0])); Chris@10: Tg = LD(&(xi[WS(is, 104)]), ivs, &(xi[0])); Chris@10: { Chris@10: V Tj, Tk, Tm, Tn; Chris@10: Tj = LD(&(xi[WS(is, 120)]), ivs, &(xi[0])); Chris@10: Tk = LD(&(xi[WS(is, 56)]), ivs, &(xi[0])); Chris@10: Tm = LD(&(xi[WS(is, 88)]), ivs, &(xi[0])); Chris@10: Tn = LD(&(xi[WS(is, 24)]), ivs, &(xi[0])); Chris@10: Ta = VADD(T6, T9); Chris@10: T4P = VSUB(T6, T9); Chris@10: Te = VSUB(Tc, Td); Chris@10: Tc6 = VADD(Tc, Td); Chris@10: Th = VSUB(Tf, Tg); Chris@10: Tc7 = VADD(Tf, Tg); Chris@10: Tl = VSUB(Tj, Tk); Chris@10: Tc9 = VADD(Tj, Tk); Chris@10: Tca = VADD(Tn, Tm); Chris@10: To = VSUB(Tm, Tn); Chris@10: } Chris@10: } Chris@10: } Chris@10: } Chris@10: { Chris@10: V T6X, Tb, Te9, Ta9, Tc8, Tea, T4R, Ti, Tfo, Tac, Tp, T4S, Tcb, Teb, T4Q; Chris@10: T6X = VFNMS(LDK(KP707106781), Ta, T3); Chris@10: Tb = VFMA(LDK(KP707106781), Ta, T3); Chris@10: Te9 = VSUB(Ta7, Ta8); Chris@10: Ta9 = VADD(Ta7, Ta8); Chris@10: Tc8 = VADD(Tc6, Tc7); Chris@10: Tea = VSUB(Tc6, Tc7); Chris@10: T4R = VFMA(LDK(KP414213562), Te, Th); Chris@10: Ti = VFNMS(LDK(KP414213562), Th, Te); Chris@10: Tfo = VSUB(Taa, Tab); Chris@10: Tac = VADD(Taa, Tab); Chris@10: Tp = VFNMS(LDK(KP414213562), To, Tl); Chris@10: T4S = VFMA(LDK(KP414213562), Tl, To); Chris@10: Tcb = VADD(Tc9, Tca); Chris@10: Teb = VSUB(Tc9, Tca); Chris@10: T4Q = VFMA(LDK(KP707106781), T4P, T4O); Chris@10: T82 = VFNMS(LDK(KP707106781), T4P, T4O); Chris@10: { Chris@10: V T4T, T6Y, Tq, Tfp, Tec; Chris@10: T4T = VSUB(T4R, T4S); Chris@10: T6Y = VADD(T4R, T4S); Chris@10: T83 = VSUB(Ti, Tp); Chris@10: Tq = VADD(Ti, Tp); Chris@10: Tfp = VSUB(Tea, Teb); Chris@10: Tec = VADD(Tea, Teb); Chris@10: Tad = VSUB(Ta9, Tac); Chris@10: TcZ = VADD(Ta9, Tac); Chris@10: T6Z = VFMA(LDK(KP923879532), T6Y, T6X); Chris@10: T8T = VFNMS(LDK(KP923879532), T6Y, T6X); Chris@10: T4U = VFMA(LDK(KP923879532), T4T, T4Q); Chris@10: T6a = VFNMS(LDK(KP923879532), T4T, T4Q); Chris@10: Tr = VFMA(LDK(KP923879532), Tq, Tb); Chris@10: T5J = VFNMS(LDK(KP923879532), Tq, Tb); Chris@10: Tfq = VFMA(LDK(KP707106781), Tfp, Tfo); Chris@10: TgG = VFNMS(LDK(KP707106781), Tfp, Tfo); Chris@10: Ted = VFMA(LDK(KP707106781), Tec, Te9); Chris@10: Tgf = VFNMS(LDK(KP707106781), Tec, Te9); Chris@10: Td0 = VADD(Tc8, Tcb); Chris@10: Tcc = VSUB(Tc8, Tcb); Chris@10: } Chris@10: } Chris@10: } Chris@10: { Chris@10: V T2i, T3j, Tb2, T2B, Tb5, T3k, T2p, T2C; Chris@10: { Chris@10: V T2m, Tb0, Tb1, Tb3, T2l, T2n; Chris@10: { Chris@10: V T2g, T2h, T3h, T3i, T2j, T2k; Chris@10: T2g = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)])); Chris@10: T2h = LD(&(xi[WS(is, 65)]), ivs, &(xi[WS(is, 1)])); Chris@10: T3h = LD(&(xi[WS(is, 33)]), ivs, &(xi[WS(is, 1)])); Chris@10: T3i = LD(&(xi[WS(is, 97)]), ivs, &(xi[WS(is, 1)])); Chris@10: T2j = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)])); Chris@10: T2k = LD(&(xi[WS(is, 81)]), ivs, &(xi[WS(is, 1)])); Chris@10: T2m = LD(&(xi[WS(is, 113)]), ivs, &(xi[WS(is, 1)])); Chris@10: T9k = VFMA(LDK(KP923879532), T83, T82); Chris@10: T84 = VFNMS(LDK(KP923879532), T83, T82); Chris@10: T2i = VSUB(T2g, T2h); Chris@10: Tb0 = VADD(T2g, T2h); Chris@10: T3j = VSUB(T3h, T3i); Chris@10: Tb1 = VADD(T3h, T3i); Chris@10: Tb3 = VADD(T2j, T2k); Chris@10: T2l = VSUB(T2j, T2k); Chris@10: T2n = LD(&(xi[WS(is, 49)]), ivs, &(xi[WS(is, 1)])); Chris@10: } Chris@10: { Chris@10: V T2r, T2s, T2u, T2v; Chris@10: T2r = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)])); Chris@10: T2s = LD(&(xi[WS(is, 73)]), ivs, &(xi[WS(is, 1)])); Chris@10: T2u = LD(&(xi[WS(is, 41)]), ivs, &(xi[WS(is, 1)])); Chris@10: T2v = LD(&(xi[WS(is, 105)]), ivs, &(xi[WS(is, 1)])); Chris@10: TeG = VSUB(Tb0, Tb1); Chris@10: Tb2 = VADD(Tb0, Tb1); Chris@10: { Chris@10: V T2y, T2z, Tb4, T2o, Tbn, T2t, Tbo, T2w; Chris@10: T2y = LD(&(xi[WS(is, 121)]), ivs, &(xi[WS(is, 1)])); Chris@10: T2z = LD(&(xi[WS(is, 57)]), ivs, &(xi[WS(is, 1)])); Chris@10: Tb4 = VADD(T2m, T2n); Chris@10: T2o = VSUB(T2m, T2n); Chris@10: Tbn = VADD(T2r, T2s); Chris@10: T2t = VSUB(T2r, T2s); Chris@10: Tbo = VADD(T2u, T2v); Chris@10: T2w = VSUB(T2u, T2v); Chris@10: T2B = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)])); Chris@10: T2A = VSUB(T2y, T2z); Chris@10: Tbq = VADD(T2y, T2z); Chris@10: TeT = VSUB(Tb3, Tb4); Chris@10: Tb5 = VADD(Tb3, Tb4); Chris@10: T3k = VSUB(T2l, T2o); Chris@10: T2p = VADD(T2l, T2o); Chris@10: Tbp = VADD(Tbn, Tbo); Chris@10: TeH = VSUB(Tbn, Tbo); Chris@10: T3m = VFMA(LDK(KP414213562), T2t, T2w); Chris@10: T2x = VFNMS(LDK(KP414213562), T2w, T2t); Chris@10: T2C = LD(&(xi[WS(is, 89)]), ivs, &(xi[WS(is, 1)])); Chris@10: } Chris@10: } Chris@10: } Chris@10: Td6 = VADD(Tb2, Tb5); Chris@10: Tb6 = VSUB(Tb2, Tb5); Chris@10: T7o = VFNMS(LDK(KP707106781), T2p, T2i); Chris@10: T2q = VFMA(LDK(KP707106781), T2p, T2i); Chris@10: T3l = VFMA(LDK(KP707106781), T3k, T3j); Chris@10: T7z = VFNMS(LDK(KP707106781), T3k, T3j); Chris@10: Tbr = VADD(T2B, T2C); Chris@10: T2D = VSUB(T2B, T2C); Chris@10: } Chris@10: { Chris@10: V Tf1, Tfe, Tf2, TbZ, T3M, T4B, Tdd, T3F, T7H, T4A, T7S, TbW, Tf3, T4C, T3T; Chris@10: { Chris@10: V T3x, T4y, Tbz, T3Q, TbC, T4z, T3E, T3R, T3P, TbU, TbV, T3S; Chris@10: { Chris@10: V T3y, T3z, T3B, T3C; Chris@10: { Chris@10: V T3v, T3w, T4w, T4x; Chris@10: T3v = LD(&(xi[WS(is, 127)]), ivs, &(xi[WS(is, 1)])); Chris@10: T3w = LD(&(xi[WS(is, 63)]), ivs, &(xi[WS(is, 1)])); Chris@10: T4w = LD(&(xi[WS(is, 95)]), ivs, &(xi[WS(is, 1)])); Chris@10: T4x = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)])); Chris@10: T3y = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)])); Chris@10: { Chris@10: V Tbs, TeI, T3n, T2E, Tbx; Chris@10: Tbs = VADD(Tbq, Tbr); Chris@10: TeI = VSUB(Tbq, Tbr); Chris@10: T3n = VFNMS(LDK(KP414213562), T2A, T2D); Chris@10: T2E = VFMA(LDK(KP414213562), T2D, T2A); Chris@10: T3x = VSUB(T3v, T3w); Chris@10: Tbx = VADD(T3v, T3w); Chris@10: { Chris@10: V Tby, Td7, TeJ, TeU; Chris@10: T4y = VSUB(T4w, T4x); Chris@10: Tby = VADD(T4x, T4w); Chris@10: Td7 = VADD(Tbp, Tbs); Chris@10: Tbt = VSUB(Tbp, Tbs); Chris@10: TeJ = VADD(TeH, TeI); Chris@10: TeU = VSUB(TeH, TeI); Chris@10: { Chris@10: V T7p, T3o, T7A, T2F; Chris@10: T7p = VSUB(T3m, T3n); Chris@10: T3o = VADD(T3m, T3n); Chris@10: T7A = VSUB(T2x, T2E); Chris@10: T2F = VADD(T2x, T2E); Chris@10: Tbz = VADD(Tbx, Tby); Chris@10: Tf1 = VSUB(Tbx, Tby); Chris@10: Td8 = VADD(Td6, Td7); Chris@10: TdK = VSUB(Td6, Td7); Chris@10: TeK = VFMA(LDK(KP707106781), TeJ, TeG); Chris@10: Tgq = VFNMS(LDK(KP707106781), TeJ, TeG); Chris@10: TeV = VFMA(LDK(KP707106781), TeU, TeT); Chris@10: Tgt = VFNMS(LDK(KP707106781), TeU, TeT); Chris@10: T7q = VFMA(LDK(KP923879532), T7p, T7o); Chris@10: T94 = VFNMS(LDK(KP923879532), T7p, T7o); Chris@10: T3p = VFMA(LDK(KP923879532), T3o, T3l); Chris@10: T5X = VFNMS(LDK(KP923879532), T3o, T3l); Chris@10: T7B = VFNMS(LDK(KP923879532), T7A, T7z); Chris@10: T97 = VFMA(LDK(KP923879532), T7A, T7z); Chris@10: T2G = VFMA(LDK(KP923879532), T2F, T2q); Chris@10: T5U = VFNMS(LDK(KP923879532), T2F, T2q); Chris@10: T3z = LD(&(xi[WS(is, 79)]), ivs, &(xi[WS(is, 1)])); Chris@10: } Chris@10: } Chris@10: } Chris@10: T3B = LD(&(xi[WS(is, 111)]), ivs, &(xi[WS(is, 1)])); Chris@10: T3C = LD(&(xi[WS(is, 47)]), ivs, &(xi[WS(is, 1)])); Chris@10: } Chris@10: { Chris@10: V T3G, T3H, T3J, T3K; Chris@10: T3G = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)])); Chris@10: T3H = LD(&(xi[WS(is, 71)]), ivs, &(xi[WS(is, 1)])); Chris@10: T3J = LD(&(xi[WS(is, 103)]), ivs, &(xi[WS(is, 1)])); Chris@10: T3K = LD(&(xi[WS(is, 39)]), ivs, &(xi[WS(is, 1)])); Chris@10: { Chris@10: V T3N, T3A, TbA, T3D, TbB, T3I, TbX, T3L, TbY, T3O; Chris@10: T3N = LD(&(xi[WS(is, 119)]), ivs, &(xi[WS(is, 1)])); Chris@10: T3A = VSUB(T3y, T3z); Chris@10: TbA = VADD(T3y, T3z); Chris@10: T3D = VSUB(T3B, T3C); Chris@10: TbB = VADD(T3B, T3C); Chris@10: T3I = VSUB(T3G, T3H); Chris@10: TbX = VADD(T3G, T3H); Chris@10: T3L = VSUB(T3J, T3K); Chris@10: TbY = VADD(T3K, T3J); Chris@10: T3O = LD(&(xi[WS(is, 55)]), ivs, &(xi[WS(is, 1)])); Chris@10: T3Q = LD(&(xi[WS(is, 87)]), ivs, &(xi[WS(is, 1)])); Chris@10: Tfe = VSUB(TbB, TbA); Chris@10: TbC = VADD(TbA, TbB); Chris@10: T4z = VSUB(T3D, T3A); Chris@10: T3E = VADD(T3A, T3D); Chris@10: T3R = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)])); Chris@10: Tf2 = VSUB(TbX, TbY); Chris@10: TbZ = VADD(TbX, TbY); Chris@10: T3M = VFMA(LDK(KP414213562), T3L, T3I); Chris@10: T4B = VFNMS(LDK(KP414213562), T3I, T3L); Chris@10: T3P = VSUB(T3N, T3O); Chris@10: TbU = VADD(T3N, T3O); Chris@10: } Chris@10: } Chris@10: } Chris@10: Tdd = VADD(Tbz, TbC); Chris@10: TbD = VSUB(Tbz, TbC); Chris@10: TbV = VADD(T3R, T3Q); Chris@10: T3S = VSUB(T3Q, T3R); Chris@10: T3F = VFMA(LDK(KP707106781), T3E, T3x); Chris@10: T7H = VFNMS(LDK(KP707106781), T3E, T3x); Chris@10: T4A = VFMA(LDK(KP707106781), T4z, T4y); Chris@10: T7S = VFNMS(LDK(KP707106781), T4z, T4y); Chris@10: TbW = VADD(TbU, TbV); Chris@10: Tf3 = VSUB(TbU, TbV); Chris@10: T4C = VFMA(LDK(KP414213562), T3P, T3S); Chris@10: T3T = VFNMS(LDK(KP414213562), T3S, T3P); Chris@10: } Chris@10: { Chris@10: V TD, Tae, TE, TJ, TK, TU, TV; Chris@10: { Chris@10: V Ts, Tt, Tde, Tf4, Tff; Chris@10: Ts = LD(&(xi[WS(is, 4)]), ivs, &(xi[0])); Chris@10: Tt = LD(&(xi[WS(is, 68)]), ivs, &(xi[0])); Chris@10: TD = LD(&(xi[WS(is, 36)]), ivs, &(xi[0])); Chris@10: Tde = VADD(TbZ, TbW); Chris@10: Tc0 = VSUB(TbW, TbZ); Chris@10: Tf4 = VADD(Tf2, Tf3); Chris@10: Tff = VSUB(Tf3, Tf2); Chris@10: { Chris@10: V T7I, T4D, T7T, T3U; Chris@10: T7I = VSUB(T4C, T4B); Chris@10: T4D = VADD(T4B, T4C); Chris@10: T7T = VSUB(T3T, T3M); Chris@10: T3U = VADD(T3M, T3T); Chris@10: Tae = VADD(Ts, Tt); Chris@10: Tu = VSUB(Ts, Tt); Chris@10: Tdf = VADD(Tdd, Tde); Chris@10: TdN = VSUB(Tdd, Tde); Chris@10: Tf5 = VFMA(LDK(KP707106781), Tf4, Tf1); Chris@10: Tgx = VFNMS(LDK(KP707106781), Tf4, Tf1); Chris@10: Tfg = VFMA(LDK(KP707106781), Tff, Tfe); Chris@10: TgA = VFNMS(LDK(KP707106781), Tff, Tfe); Chris@10: T7J = VFMA(LDK(KP923879532), T7I, T7H); Chris@10: T9b = VFNMS(LDK(KP923879532), T7I, T7H); Chris@10: T4E = VFMA(LDK(KP923879532), T4D, T4A); Chris@10: T64 = VFNMS(LDK(KP923879532), T4D, T4A); Chris@10: T7U = VFNMS(LDK(KP923879532), T7T, T7S); Chris@10: T9e = VFMA(LDK(KP923879532), T7T, T7S); Chris@10: T3V = VFMA(LDK(KP923879532), T3U, T3F); Chris@10: T61 = VFNMS(LDK(KP923879532), T3U, T3F); Chris@10: TE = LD(&(xi[WS(is, 100)]), ivs, &(xi[0])); Chris@10: } Chris@10: } Chris@10: TJ = LD(&(xi[WS(is, 124)]), ivs, &(xi[0])); Chris@10: TK = LD(&(xi[WS(is, 60)]), ivs, &(xi[0])); Chris@10: TU = LD(&(xi[WS(is, 92)]), ivs, &(xi[0])); Chris@10: TV = LD(&(xi[WS(is, 28)]), ivs, &(xi[0])); Chris@10: { Chris@10: V Tal, Tam, Tv, Tw, Taf; Chris@10: Tv = LD(&(xi[WS(is, 20)]), ivs, &(xi[0])); Chris@10: Tw = LD(&(xi[WS(is, 84)]), ivs, &(xi[0])); Chris@10: Taf = VADD(TD, TE); Chris@10: TF = VSUB(TD, TE); Chris@10: Ty = LD(&(xi[WS(is, 116)]), ivs, &(xi[0])); Chris@10: TL = VSUB(TJ, TK); Chris@10: Tal = VADD(TJ, TK); Chris@10: TW = VSUB(TU, TV); Chris@10: Tam = VADD(TV, TU); Chris@10: Tah = VADD(Tv, Tw); Chris@10: Tx = VSUB(Tv, Tw); Chris@10: Tag = VADD(Tae, Taf); Chris@10: Tee = VSUB(Tae, Taf); Chris@10: Tz = LD(&(xi[WS(is, 52)]), ivs, &(xi[0])); Chris@10: TM = LD(&(xi[WS(is, 12)]), ivs, &(xi[0])); Chris@10: TN = LD(&(xi[WS(is, 76)]), ivs, &(xi[0])); Chris@10: Teh = VSUB(Tal, Tam); Chris@10: Tan = VADD(Tal, Tam); Chris@10: TP = LD(&(xi[WS(is, 108)]), ivs, &(xi[0])); Chris@10: TQ = LD(&(xi[WS(is, 44)]), ivs, &(xi[0])); Chris@10: } Chris@10: } Chris@10: } Chris@10: } Chris@10: { Chris@10: V Tev, TeA, Tdp, TaP, Tew, TaV, T1U, T29, T7f, T1N, T28, T7i, Tex, TaS, T21; Chris@10: V T2a; Chris@10: { Chris@10: V Tem, Ter, Ten, TaD, T1j, T1y, TaA, Tdm, T1c, T78, T7b, T1x, TaG, Teo, T1z; Chris@10: V T1q; Chris@10: { Chris@10: V T14, T1v, Taw, Taz, T1b, T1w, T1n, T1o, T1m, TaE, TaF, T1p; Chris@10: { Chris@10: V Tau, Tav, T15, T16, T18, T19; Chris@10: { Chris@10: V T12, Tai, TA, Tao, TO, T13; Chris@10: T12 = LD(&(xi[WS(is, 2)]), ivs, &(xi[0])); Chris@10: Tai = VADD(Ty, Tz); Chris@10: TA = VSUB(Ty, Tz); Chris@10: Tao = VADD(TM, TN); Chris@10: TO = VSUB(TM, TN); Chris@10: T13 = LD(&(xi[WS(is, 66)]), ivs, &(xi[0])); Chris@10: { Chris@10: V T1t, Tap, TR, Taj, Tef, TG, TB, T1u; Chris@10: T1t = LD(&(xi[WS(is, 34)]), ivs, &(xi[0])); Chris@10: Tap = VADD(TP, TQ); Chris@10: TR = VSUB(TP, TQ); Chris@10: Taj = VADD(Tah, Tai); Chris@10: Tef = VSUB(Tah, Tai); Chris@10: TG = VSUB(Tx, TA); Chris@10: TB = VADD(Tx, TA); Chris@10: Tau = VADD(T12, T13); Chris@10: T14 = VSUB(T12, T13); Chris@10: T1u = LD(&(xi[WS(is, 98)]), ivs, &(xi[0])); Chris@10: { Chris@10: V Taq, Tei, TX, TS, Tak; Chris@10: Taq = VADD(Tao, Tap); Chris@10: Tei = VSUB(Tap, Tao); Chris@10: TX = VSUB(TR, TO); Chris@10: TS = VADD(TO, TR); Chris@10: Tak = VSUB(Tag, Taj); Chris@10: Td2 = VADD(Tag, Taj); Chris@10: { Chris@10: V Teg, Tfr, T71, TH; Chris@10: Teg = VFNMS(LDK(KP414213562), Tef, Tee); Chris@10: Tfr = VFMA(LDK(KP414213562), Tee, Tef); Chris@10: T71 = VFNMS(LDK(KP707106781), TG, TF); Chris@10: TH = VFMA(LDK(KP707106781), TG, TF); Chris@10: { Chris@10: V T70, TC, Tar, Tej, Tfs; Chris@10: T70 = VFNMS(LDK(KP707106781), TB, Tu); Chris@10: TC = VFMA(LDK(KP707106781), TB, Tu); Chris@10: Tar = VSUB(Tan, Taq); Chris@10: Td3 = VADD(Tan, Taq); Chris@10: Tej = VFNMS(LDK(KP414213562), Tei, Teh); Chris@10: Tfs = VFMA(LDK(KP414213562), Teh, Tei); Chris@10: { Chris@10: V T74, TY, T73, TT; Chris@10: T74 = VFNMS(LDK(KP707106781), TX, TW); Chris@10: TY = VFMA(LDK(KP707106781), TX, TW); Chris@10: T73 = VFNMS(LDK(KP707106781), TS, TL); Chris@10: TT = VFMA(LDK(KP707106781), TS, TL); Chris@10: T85 = VFNMS(LDK(KP668178637), T70, T71); Chris@10: T72 = VFMA(LDK(KP668178637), T71, T70); Chris@10: T4V = VFMA(LDK(KP198912367), TC, TH); Chris@10: TI = VFNMS(LDK(KP198912367), TH, TC); Chris@10: Tcd = VSUB(Tak, Tar); Chris@10: Tas = VADD(Tak, Tar); Chris@10: TgH = VSUB(Teg, Tej); Chris@10: Tek = VADD(Teg, Tej); Chris@10: Tgg = VADD(Tfr, Tfs); Chris@10: Tft = VSUB(Tfr, Tfs); Chris@10: T86 = VFNMS(LDK(KP668178637), T73, T74); Chris@10: T75 = VFMA(LDK(KP668178637), T74, T73); Chris@10: T4W = VFMA(LDK(KP198912367), TT, TY); Chris@10: TZ = VFNMS(LDK(KP198912367), TY, TT); Chris@10: Tav = VADD(T1t, T1u); Chris@10: T1v = VSUB(T1t, T1u); Chris@10: } Chris@10: } Chris@10: } Chris@10: } Chris@10: } Chris@10: } Chris@10: T15 = LD(&(xi[WS(is, 18)]), ivs, &(xi[0])); Chris@10: T16 = LD(&(xi[WS(is, 82)]), ivs, &(xi[0])); Chris@10: T18 = LD(&(xi[WS(is, 114)]), ivs, &(xi[0])); Chris@10: T19 = LD(&(xi[WS(is, 50)]), ivs, &(xi[0])); Chris@10: { Chris@10: V T1d, T1e, T1g, T1h, Tax, T17, Tay, T1a; Chris@10: T1d = LD(&(xi[WS(is, 10)]), ivs, &(xi[0])); Chris@10: Taw = VADD(Tau, Tav); Chris@10: Tem = VSUB(Tau, Tav); Chris@10: T1e = LD(&(xi[WS(is, 74)]), ivs, &(xi[0])); Chris@10: T1g = LD(&(xi[WS(is, 42)]), ivs, &(xi[0])); Chris@10: T1h = LD(&(xi[WS(is, 106)]), ivs, &(xi[0])); Chris@10: Tax = VADD(T15, T16); Chris@10: T17 = VSUB(T15, T16); Chris@10: Tay = VADD(T18, T19); Chris@10: T1a = VSUB(T18, T19); Chris@10: { Chris@10: V T1k, T1f, TaB, T1i, TaC, T1l; Chris@10: T1k = LD(&(xi[WS(is, 122)]), ivs, &(xi[0])); Chris@10: T1f = VSUB(T1d, T1e); Chris@10: TaB = VADD(T1d, T1e); Chris@10: T1i = VSUB(T1g, T1h); Chris@10: TaC = VADD(T1g, T1h); Chris@10: T1l = LD(&(xi[WS(is, 58)]), ivs, &(xi[0])); Chris@10: Taz = VADD(Tax, Tay); Chris@10: Ter = VSUB(Tax, Tay); Chris@10: T1b = VADD(T17, T1a); Chris@10: T1w = VSUB(T17, T1a); Chris@10: T1n = LD(&(xi[WS(is, 26)]), ivs, &(xi[0])); Chris@10: T1o = LD(&(xi[WS(is, 90)]), ivs, &(xi[0])); Chris@10: Ten = VSUB(TaB, TaC); Chris@10: TaD = VADD(TaB, TaC); Chris@10: T1j = VFNMS(LDK(KP414213562), T1i, T1f); Chris@10: T1y = VFMA(LDK(KP414213562), T1f, T1i); Chris@10: T1m = VSUB(T1k, T1l); Chris@10: TaE = VADD(T1k, T1l); Chris@10: } Chris@10: } Chris@10: } Chris@10: TaA = VSUB(Taw, Taz); Chris@10: Tdm = VADD(Taw, Taz); Chris@10: TaF = VADD(T1n, T1o); Chris@10: T1p = VSUB(T1n, T1o); Chris@10: T1c = VFMA(LDK(KP707106781), T1b, T14); Chris@10: T78 = VFNMS(LDK(KP707106781), T1b, T14); Chris@10: T7b = VFNMS(LDK(KP707106781), T1w, T1v); Chris@10: T1x = VFMA(LDK(KP707106781), T1w, T1v); Chris@10: TaG = VADD(TaE, TaF); Chris@10: Teo = VSUB(TaE, TaF); Chris@10: T1z = VFNMS(LDK(KP414213562), T1m, T1p); Chris@10: T1q = VFMA(LDK(KP414213562), T1p, T1m); Chris@10: } Chris@10: { Chris@10: V T1F, T26, T1Q, TaT, TaL, TaO, T27, T1M, T1Y, T1Z, TaU, T1T, TaQ, T1X, T20; Chris@10: V TaR; Chris@10: { Chris@10: V T24, TaJ, T25, T1G, T1H, T1J, T1K, T1D, T1E; Chris@10: T1D = LD(&(xi[WS(is, 126)]), ivs, &(xi[0])); Chris@10: T1E = LD(&(xi[WS(is, 62)]), ivs, &(xi[0])); Chris@10: T24 = LD(&(xi[WS(is, 94)]), ivs, &(xi[0])); Chris@10: { Chris@10: V TaH, Tdn, Tes, Tep; Chris@10: TaH = VSUB(TaD, TaG); Chris@10: Tdn = VADD(TaD, TaG); Chris@10: Tes = VSUB(Ten, Teo); Chris@10: Tep = VADD(Ten, Teo); Chris@10: { Chris@10: V T79, T1A, T7c, T1r; Chris@10: T79 = VSUB(T1y, T1z); Chris@10: T1A = VADD(T1y, T1z); Chris@10: T7c = VSUB(T1j, T1q); Chris@10: T1r = VADD(T1j, T1q); Chris@10: TaJ = VADD(T1D, T1E); Chris@10: T1F = VSUB(T1D, T1E); Chris@10: TaI = VFNMS(LDK(KP414213562), TaH, TaA); Chris@10: Tcf = VFMA(LDK(KP414213562), TaA, TaH); Chris@10: Tdo = VADD(Tdm, Tdn); Chris@10: TdG = VSUB(Tdm, Tdn); Chris@10: Tgi = VFNMS(LDK(KP707106781), Tes, Ter); Chris@10: Tet = VFMA(LDK(KP707106781), Tes, Ter); Chris@10: Tgj = VFNMS(LDK(KP707106781), Tep, Tem); Chris@10: Teq = VFMA(LDK(KP707106781), Tep, Tem); Chris@10: T8X = VFNMS(LDK(KP923879532), T79, T78); Chris@10: T7a = VFMA(LDK(KP923879532), T79, T78); Chris@10: T5M = VFNMS(LDK(KP923879532), T1A, T1x); Chris@10: T1B = VFMA(LDK(KP923879532), T1A, T1x); Chris@10: T8W = VFMA(LDK(KP923879532), T7c, T7b); Chris@10: T7d = VFNMS(LDK(KP923879532), T7c, T7b); Chris@10: T5N = VFNMS(LDK(KP923879532), T1r, T1c); Chris@10: T1s = VFMA(LDK(KP923879532), T1r, T1c); Chris@10: T25 = LD(&(xi[WS(is, 30)]), ivs, &(xi[0])); Chris@10: } Chris@10: } Chris@10: T1G = LD(&(xi[WS(is, 14)]), ivs, &(xi[0])); Chris@10: T1H = LD(&(xi[WS(is, 78)]), ivs, &(xi[0])); Chris@10: T1J = LD(&(xi[WS(is, 110)]), ivs, &(xi[0])); Chris@10: T1K = LD(&(xi[WS(is, 46)]), ivs, &(xi[0])); Chris@10: { Chris@10: V T1R, T1I, TaM, T1L, TaN, T1S, T1O, T1P, TaK, T1V, T1W; Chris@10: T1O = LD(&(xi[WS(is, 6)]), ivs, &(xi[0])); Chris@10: T1P = LD(&(xi[WS(is, 70)]), ivs, &(xi[0])); Chris@10: T26 = VSUB(T24, T25); Chris@10: TaK = VADD(T25, T24); Chris@10: T1R = LD(&(xi[WS(is, 102)]), ivs, &(xi[0])); Chris@10: T1I = VSUB(T1G, T1H); Chris@10: TaM = VADD(T1G, T1H); Chris@10: T1L = VSUB(T1J, T1K); Chris@10: TaN = VADD(T1J, T1K); Chris@10: T1Q = VSUB(T1O, T1P); Chris@10: TaT = VADD(T1O, T1P); Chris@10: Tev = VSUB(TaJ, TaK); Chris@10: TaL = VADD(TaJ, TaK); Chris@10: T1S = LD(&(xi[WS(is, 38)]), ivs, &(xi[0])); Chris@10: T1V = LD(&(xi[WS(is, 118)]), ivs, &(xi[0])); Chris@10: T1W = LD(&(xi[WS(is, 54)]), ivs, &(xi[0])); Chris@10: TeA = VSUB(TaN, TaM); Chris@10: TaO = VADD(TaM, TaN); Chris@10: T27 = VSUB(T1L, T1I); Chris@10: T1M = VADD(T1I, T1L); Chris@10: T1Y = LD(&(xi[WS(is, 86)]), ivs, &(xi[0])); Chris@10: T1Z = LD(&(xi[WS(is, 22)]), ivs, &(xi[0])); Chris@10: TaU = VADD(T1S, T1R); Chris@10: T1T = VSUB(T1R, T1S); Chris@10: TaQ = VADD(T1V, T1W); Chris@10: T1X = VSUB(T1V, T1W); Chris@10: } Chris@10: } Chris@10: Tdp = VADD(TaL, TaO); Chris@10: TaP = VSUB(TaL, TaO); Chris@10: T20 = VSUB(T1Y, T1Z); Chris@10: TaR = VADD(T1Z, T1Y); Chris@10: Tew = VSUB(TaT, TaU); Chris@10: TaV = VADD(TaT, TaU); Chris@10: T1U = VFMA(LDK(KP414213562), T1T, T1Q); Chris@10: T29 = VFNMS(LDK(KP414213562), T1Q, T1T); Chris@10: T7f = VFNMS(LDK(KP707106781), T1M, T1F); Chris@10: T1N = VFMA(LDK(KP707106781), T1M, T1F); Chris@10: T28 = VFMA(LDK(KP707106781), T27, T26); Chris@10: T7i = VFNMS(LDK(KP707106781), T27, T26); Chris@10: Tex = VSUB(TaQ, TaR); Chris@10: TaS = VADD(TaQ, TaR); Chris@10: T21 = VFNMS(LDK(KP414213562), T20, T1X); Chris@10: T2a = VFMA(LDK(KP414213562), T1X, T20); Chris@10: } Chris@10: } Chris@10: { Chris@10: V T2J, T2U, T30, T3b, TeL, Tb9, TeO, Tbg, T2M, Tba, T2P, Tbb, T34, Tbh, T33; Chris@10: V T35; Chris@10: { Chris@10: V T2H, T2I, T2S, T2T, T2Y, T2Z, T39, T3a; Chris@10: T2H = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)])); Chris@10: { Chris@10: V Tdq, TaW, Tey, TeB; Chris@10: Tdq = VADD(TaV, TaS); Chris@10: TaW = VSUB(TaS, TaV); Chris@10: Tey = VADD(Tew, Tex); Chris@10: TeB = VSUB(Tex, Tew); Chris@10: { Chris@10: V T2b, T7g, T22, T7j; Chris@10: T2b = VADD(T29, T2a); Chris@10: T7g = VSUB(T2a, T29); Chris@10: T22 = VADD(T1U, T21); Chris@10: T7j = VSUB(T21, T1U); Chris@10: TaX = VFNMS(LDK(KP414213562), TaW, TaP); Chris@10: Tcg = VFMA(LDK(KP414213562), TaP, TaW); Chris@10: Tdr = VADD(Tdp, Tdq); Chris@10: TdH = VSUB(Tdp, Tdq); Chris@10: Tgl = VFNMS(LDK(KP707106781), TeB, TeA); Chris@10: TeC = VFMA(LDK(KP707106781), TeB, TeA); Chris@10: Tgm = VFNMS(LDK(KP707106781), Tey, Tev); Chris@10: Tez = VFMA(LDK(KP707106781), Tey, Tev); Chris@10: T90 = VFNMS(LDK(KP923879532), T7g, T7f); Chris@10: T7h = VFMA(LDK(KP923879532), T7g, T7f); Chris@10: T5P = VFNMS(LDK(KP923879532), T2b, T28); Chris@10: T2c = VFMA(LDK(KP923879532), T2b, T28); Chris@10: T8Z = VFMA(LDK(KP923879532), T7j, T7i); Chris@10: T7k = VFNMS(LDK(KP923879532), T7j, T7i); Chris@10: T5Q = VFNMS(LDK(KP923879532), T22, T1N); Chris@10: T23 = VFMA(LDK(KP923879532), T22, T1N); Chris@10: T2I = LD(&(xi[WS(is, 69)]), ivs, &(xi[WS(is, 1)])); Chris@10: } Chris@10: } Chris@10: T2S = LD(&(xi[WS(is, 37)]), ivs, &(xi[WS(is, 1)])); Chris@10: T2T = LD(&(xi[WS(is, 101)]), ivs, &(xi[WS(is, 1)])); Chris@10: T2Y = LD(&(xi[WS(is, 125)]), ivs, &(xi[WS(is, 1)])); Chris@10: T2Z = LD(&(xi[WS(is, 61)]), ivs, &(xi[WS(is, 1)])); Chris@10: T39 = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)])); Chris@10: T3a = LD(&(xi[WS(is, 93)]), ivs, &(xi[WS(is, 1)])); Chris@10: { Chris@10: V T2K, Tbe, Tbf, T2L, T2N, T2O, Tb7, Tb8, T31, T32; Chris@10: T2K = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)])); Chris@10: T2J = VSUB(T2H, T2I); Chris@10: Tb7 = VADD(T2H, T2I); Chris@10: T2U = VSUB(T2S, T2T); Chris@10: Tb8 = VADD(T2S, T2T); Chris@10: T30 = VSUB(T2Y, T2Z); Chris@10: Tbe = VADD(T2Y, T2Z); Chris@10: T3b = VSUB(T39, T3a); Chris@10: Tbf = VADD(T39, T3a); Chris@10: T2L = LD(&(xi[WS(is, 85)]), ivs, &(xi[WS(is, 1)])); Chris@10: T2N = LD(&(xi[WS(is, 117)]), ivs, &(xi[WS(is, 1)])); Chris@10: T2O = LD(&(xi[WS(is, 53)]), ivs, &(xi[WS(is, 1)])); Chris@10: TeL = VSUB(Tb7, Tb8); Chris@10: Tb9 = VADD(Tb7, Tb8); Chris@10: T31 = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)])); Chris@10: T32 = LD(&(xi[WS(is, 77)]), ivs, &(xi[WS(is, 1)])); Chris@10: TeO = VSUB(Tbe, Tbf); Chris@10: Tbg = VADD(Tbe, Tbf); Chris@10: T2M = VSUB(T2K, T2L); Chris@10: Tba = VADD(T2K, T2L); Chris@10: T2P = VSUB(T2N, T2O); Chris@10: Tbb = VADD(T2N, T2O); Chris@10: T34 = LD(&(xi[WS(is, 109)]), ivs, &(xi[WS(is, 1)])); Chris@10: Tbh = VADD(T31, T32); Chris@10: T33 = VSUB(T31, T32); Chris@10: T35 = LD(&(xi[WS(is, 45)]), ivs, &(xi[WS(is, 1)])); Chris@10: } Chris@10: } Chris@10: { Chris@10: V T4d, T4e, T4o, T4p; Chris@10: { Chris@10: V T2X, T3q, T7t, T7C, T3r, T3e, T7D, T7w; Chris@10: { Chris@10: V T47, TbE, Tbd, Td9, TeW, TeN, T7s, T2W, T7r, T2R, TeP, Tbj, T37, T3c, T48; Chris@10: { Chris@10: V T3W, T3X, TeM, Tbc, T2Q, T2V, Tbi, T36; Chris@10: T3W = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)])); Chris@10: T3X = LD(&(xi[WS(is, 67)]), ivs, &(xi[WS(is, 1)])); Chris@10: TeM = VSUB(Tba, Tbb); Chris@10: Tbc = VADD(Tba, Tbb); Chris@10: T2Q = VADD(T2M, T2P); Chris@10: T2V = VSUB(T2M, T2P); Chris@10: T47 = LD(&(xi[WS(is, 99)]), ivs, &(xi[WS(is, 1)])); Chris@10: Tbi = VADD(T34, T35); Chris@10: T36 = VSUB(T34, T35); Chris@10: TbE = VADD(T3W, T3X); Chris@10: T3Y = VSUB(T3W, T3X); Chris@10: Tbd = VSUB(Tb9, Tbc); Chris@10: Td9 = VADD(Tb9, Tbc); Chris@10: TeW = VFMA(LDK(KP414213562), TeL, TeM); Chris@10: TeN = VFNMS(LDK(KP414213562), TeM, TeL); Chris@10: T7s = VFNMS(LDK(KP707106781), T2V, T2U); Chris@10: T2W = VFMA(LDK(KP707106781), T2V, T2U); Chris@10: T7r = VFNMS(LDK(KP707106781), T2Q, T2J); Chris@10: T2R = VFMA(LDK(KP707106781), T2Q, T2J); Chris@10: TeP = VSUB(Tbh, Tbi); Chris@10: Tbj = VADD(Tbh, Tbi); Chris@10: T37 = VADD(T33, T36); Chris@10: T3c = VSUB(T33, T36); Chris@10: T48 = LD(&(xi[WS(is, 35)]), ivs, &(xi[WS(is, 1)])); Chris@10: } Chris@10: T2X = VFNMS(LDK(KP198912367), T2W, T2R); Chris@10: T3q = VFMA(LDK(KP198912367), T2R, T2W); Chris@10: T7t = VFMA(LDK(KP668178637), T7s, T7r); Chris@10: T7C = VFNMS(LDK(KP668178637), T7r, T7s); Chris@10: { Chris@10: V Tbk, Tda, TeX, TeQ; Chris@10: Tbk = VSUB(Tbg, Tbj); Chris@10: Tda = VADD(Tbg, Tbj); Chris@10: TeX = VFNMS(LDK(KP414213562), TeO, TeP); Chris@10: TeQ = VFMA(LDK(KP414213562), TeP, TeO); Chris@10: { Chris@10: V T7v, T3d, T7u, T38, TbF; Chris@10: T7v = VFNMS(LDK(KP707106781), T3c, T3b); Chris@10: T3d = VFMA(LDK(KP707106781), T3c, T3b); Chris@10: T7u = VFNMS(LDK(KP707106781), T37, T30); Chris@10: T38 = VFMA(LDK(KP707106781), T37, T30); Chris@10: T49 = VSUB(T47, T48); Chris@10: TbF = VADD(T48, T47); Chris@10: TdL = VSUB(Td9, Tda); Chris@10: Tdb = VADD(Td9, Tda); Chris@10: Tbu = VSUB(Tbd, Tbk); Chris@10: Tbl = VADD(Tbd, Tbk); Chris@10: Tgu = VSUB(TeN, TeQ); Chris@10: TeR = VADD(TeN, TeQ); Chris@10: Tgr = VSUB(TeW, TeX); Chris@10: TeY = VADD(TeW, TeX); Chris@10: T3r = VFNMS(LDK(KP198912367), T38, T3d); Chris@10: T3e = VFMA(LDK(KP198912367), T3d, T38); Chris@10: T7D = VFMA(LDK(KP668178637), T7u, T7v); Chris@10: T7w = VFNMS(LDK(KP668178637), T7v, T7u); Chris@10: Tf6 = VSUB(TbE, TbF); Chris@10: TbG = VADD(TbE, TbF); Chris@10: } Chris@10: } Chris@10: } Chris@10: T4d = LD(&(xi[WS(is, 123)]), ivs, &(xi[WS(is, 1)])); Chris@10: T5V = VSUB(T3q, T3r); Chris@10: T3s = VADD(T3q, T3r); Chris@10: T5Y = VSUB(T2X, T3e); Chris@10: T3f = VADD(T2X, T3e); Chris@10: T95 = VSUB(T7D, T7C); Chris@10: T7E = VADD(T7C, T7D); Chris@10: T98 = VSUB(T7t, T7w); Chris@10: T7x = VADD(T7t, T7w); Chris@10: T4e = LD(&(xi[WS(is, 59)]), ivs, &(xi[WS(is, 1)])); Chris@10: T4o = LD(&(xi[WS(is, 91)]), ivs, &(xi[WS(is, 1)])); Chris@10: T4p = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)])); Chris@10: } Chris@10: { Chris@10: V T3Z, T40, T42, T43, TbL, TbM; Chris@10: T3Z = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)])); Chris@10: T40 = LD(&(xi[WS(is, 83)]), ivs, &(xi[WS(is, 1)])); Chris@10: T42 = LD(&(xi[WS(is, 115)]), ivs, &(xi[WS(is, 1)])); Chris@10: T43 = LD(&(xi[WS(is, 51)]), ivs, &(xi[WS(is, 1)])); Chris@10: T4g = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)])); Chris@10: T4f = VSUB(T4d, T4e); Chris@10: TbL = VADD(T4d, T4e); Chris@10: T4q = VSUB(T4o, T4p); Chris@10: TbM = VADD(T4p, T4o); Chris@10: TbH = VADD(T3Z, T40); Chris@10: T41 = VSUB(T3Z, T40); Chris@10: TbI = VADD(T42, T43); Chris@10: T44 = VSUB(T42, T43); Chris@10: T4h = LD(&(xi[WS(is, 75)]), ivs, &(xi[WS(is, 1)])); Chris@10: T4j = LD(&(xi[WS(is, 107)]), ivs, &(xi[WS(is, 1)])); Chris@10: T4k = LD(&(xi[WS(is, 43)]), ivs, &(xi[WS(is, 1)])); Chris@10: Tf9 = VSUB(TbL, TbM); Chris@10: TbN = VADD(TbL, TbM); Chris@10: } Chris@10: } Chris@10: } Chris@10: } Chris@10: } Chris@10: { Chris@10: V TgB, Tgy, T62, T4H, T65, T4u, T9c, T7X, T9f, T7Q, Tg0, Tga, TfF, TeF, TfT; Chris@10: V TfU, TfP, Tg7, TfI, Tfy, Tfz, Tf0, TfA, Tfl, Tg1, TfS; Chris@10: { Chris@10: V Tc1, TbS, Tfc, Tfj, TdX, Te5, TdZ, TdR, Te7, Te3, TdU, Te4; Chris@10: { Chris@10: V TdF, TdS, Tdx, Td5, TdO, TdE, TdC, Tdt, Tdk; Chris@10: { Chris@10: V Tdc, TdA, T4F, T4c, T7V, T7M, T4G, T4t, T7W, T7P, TdB, Tdj; Chris@10: { Chris@10: V Td1, Tdg, TbK, Tf8, Tfh, T4b, T7L, T46, T7K, TbQ, Tfa, T4r, T4m, Td4; Chris@10: TdF = VSUB(TcZ, Td0); Chris@10: Td1 = VADD(TcZ, Td0); Chris@10: { Chris@10: V TbJ, Tf7, T4a, T45; Chris@10: TbJ = VADD(TbH, TbI); Chris@10: Tf7 = VSUB(TbI, TbH); Chris@10: T4a = VSUB(T44, T41); Chris@10: T45 = VADD(T41, T44); Chris@10: { Chris@10: V TbO, T4i, TbP, T4l; Chris@10: TbO = VADD(T4g, T4h); Chris@10: T4i = VSUB(T4g, T4h); Chris@10: TbP = VADD(T4j, T4k); Chris@10: T4l = VSUB(T4j, T4k); Chris@10: Tdg = VADD(TbG, TbJ); Chris@10: TbK = VSUB(TbG, TbJ); Chris@10: Tf8 = VFMA(LDK(KP414213562), Tf7, Tf6); Chris@10: Tfh = VFNMS(LDK(KP414213562), Tf6, Tf7); Chris@10: T4b = VFMA(LDK(KP707106781), T4a, T49); Chris@10: T7L = VFNMS(LDK(KP707106781), T4a, T49); Chris@10: T46 = VFMA(LDK(KP707106781), T45, T3Y); Chris@10: T7K = VFNMS(LDK(KP707106781), T45, T3Y); Chris@10: TbQ = VADD(TbO, TbP); Chris@10: Tfa = VSUB(TbP, TbO); Chris@10: T4r = VSUB(T4l, T4i); Chris@10: T4m = VADD(T4i, T4l); Chris@10: Td4 = VADD(Td2, Td3); Chris@10: TdS = VSUB(Td2, Td3); Chris@10: } Chris@10: } Chris@10: Tdc = VSUB(Td8, Tdb); Chris@10: TdA = VADD(Td8, Tdb); Chris@10: T4F = VFNMS(LDK(KP198912367), T46, T4b); Chris@10: T4c = VFMA(LDK(KP198912367), T4b, T46); Chris@10: T7V = VFMA(LDK(KP668178637), T7K, T7L); Chris@10: T7M = VFNMS(LDK(KP668178637), T7L, T7K); Chris@10: { Chris@10: V Tdh, TbR, Tfb, Tfi; Chris@10: Tdh = VADD(TbN, TbQ); Chris@10: TbR = VSUB(TbN, TbQ); Chris@10: Tfb = VFNMS(LDK(KP414213562), Tfa, Tf9); Chris@10: Tfi = VFMA(LDK(KP414213562), Tf9, Tfa); Chris@10: { Chris@10: V T4s, T7O, T4n, T7N, Tdi; Chris@10: T4s = VFMA(LDK(KP707106781), T4r, T4q); Chris@10: T7O = VFNMS(LDK(KP707106781), T4r, T4q); Chris@10: T4n = VFMA(LDK(KP707106781), T4m, T4f); Chris@10: T7N = VFNMS(LDK(KP707106781), T4m, T4f); Chris@10: Tdx = VADD(Td1, Td4); Chris@10: Td5 = VSUB(Td1, Td4); Chris@10: TdO = VSUB(Tdh, Tdg); Chris@10: Tdi = VADD(Tdg, Tdh); Chris@10: Tc1 = VSUB(TbR, TbK); Chris@10: TbS = VADD(TbK, TbR); Chris@10: TgB = VSUB(Tfb, Tf8); Chris@10: Tfc = VADD(Tf8, Tfb); Chris@10: Tgy = VSUB(Tfi, Tfh); Chris@10: Tfj = VADD(Tfh, Tfi); Chris@10: T4G = VFMA(LDK(KP198912367), T4n, T4s); Chris@10: T4t = VFNMS(LDK(KP198912367), T4s, T4n); Chris@10: T7W = VFNMS(LDK(KP668178637), T7N, T7O); Chris@10: T7P = VFMA(LDK(KP668178637), T7O, T7N); Chris@10: TdB = VADD(Tdf, Tdi); Chris@10: Tdj = VSUB(Tdf, Tdi); Chris@10: } Chris@10: } Chris@10: } Chris@10: T62 = VSUB(T4G, T4F); Chris@10: T4H = VADD(T4F, T4G); Chris@10: T65 = VSUB(T4t, T4c); Chris@10: T4u = VADD(T4c, T4t); Chris@10: T9c = VSUB(T7V, T7W); Chris@10: T7X = VADD(T7V, T7W); Chris@10: T9f = VSUB(T7P, T7M); Chris@10: T7Q = VADD(T7M, T7P); Chris@10: TdE = VADD(TdA, TdB); Chris@10: TdC = VSUB(TdA, TdB); Chris@10: Tdt = VSUB(Tdc, Tdj); Chris@10: Tdk = VADD(Tdc, Tdj); Chris@10: } Chris@10: { Chris@10: V TdT, Tdl, Tdv, TdJ, Te1, Te2, TdQ, Tdz, TdD, Tdu, Tdw; Chris@10: { Chris@10: V TdI, TdM, TdV, TdW, TdP, Tds, Tdy; Chris@10: TdI = VADD(TdG, TdH); Chris@10: TdT = VSUB(TdG, TdH); Chris@10: TdM = VFNMS(LDK(KP414213562), TdL, TdK); Chris@10: TdV = VFMA(LDK(KP414213562), TdK, TdL); Chris@10: TdW = VFMA(LDK(KP414213562), TdN, TdO); Chris@10: TdP = VFNMS(LDK(KP414213562), TdO, TdN); Chris@10: Tdl = VFNMS(LDK(KP707106781), Tdk, Td5); Chris@10: Tdv = VFMA(LDK(KP707106781), Tdk, Td5); Chris@10: Tds = VSUB(Tdo, Tdr); Chris@10: Tdy = VADD(Tdo, Tdr); Chris@10: TdJ = VFMA(LDK(KP707106781), TdI, TdF); Chris@10: Te1 = VFNMS(LDK(KP707106781), TdI, TdF); Chris@10: TdX = VSUB(TdV, TdW); Chris@10: Te2 = VADD(TdV, TdW); Chris@10: Te5 = VSUB(TdM, TdP); Chris@10: TdQ = VADD(TdM, TdP); Chris@10: Tdz = VSUB(Tdx, Tdy); Chris@10: TdD = VADD(Tdx, Tdy); Chris@10: Tdu = VFNMS(LDK(KP707106781), Tdt, Tds); Chris@10: Tdw = VFMA(LDK(KP707106781), Tdt, Tds); Chris@10: } Chris@10: TdZ = VFMA(LDK(KP923879532), TdQ, TdJ); Chris@10: TdR = VFNMS(LDK(KP923879532), TdQ, TdJ); Chris@10: Te7 = VFMA(LDK(KP923879532), Te2, Te1); Chris@10: Te3 = VFNMS(LDK(KP923879532), Te2, Te1); Chris@10: ST(&(xo[0]), VADD(TdD, TdE), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 64)]), VSUB(TdD, TdE), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 32)]), VFMAI(TdC, Tdz), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 96)]), VFNMSI(TdC, Tdz), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 112)]), VFNMSI(Tdw, Tdv), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 16)]), VFMAI(Tdw, Tdv), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 80)]), VFMAI(Tdu, Tdl), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 48)]), VFNMSI(Tdu, Tdl), ovs, &(xo[0])); Chris@10: TdU = VFMA(LDK(KP707106781), TdT, TdS); Chris@10: Te4 = VFNMS(LDK(KP707106781), TdT, TdS); Chris@10: } Chris@10: } Chris@10: { Chris@10: V Tcx, TcJ, TcI, Tcy, TcA, Tbm, Tcp, TaZ, Tcs, Tci, Tbv, TcB, TcD, TbT, Tc2; Chris@10: V TcE, Tat, TaY; Chris@10: Tcx = VFNMS(LDK(KP707106781), Tas, Tad); Chris@10: Tat = VFMA(LDK(KP707106781), Tas, Tad); Chris@10: TaY = VADD(TaI, TaX); Chris@10: TcJ = VSUB(TaI, TaX); Chris@10: { Chris@10: V Tce, Tch, Te8, Te6, TdY, Te0; Chris@10: TcI = VFNMS(LDK(KP707106781), Tcd, Tcc); Chris@10: Tce = VFMA(LDK(KP707106781), Tcd, Tcc); Chris@10: Tch = VSUB(Tcf, Tcg); Chris@10: Tcy = VADD(Tcf, Tcg); Chris@10: Te8 = VFNMS(LDK(KP923879532), Te5, Te4); Chris@10: Te6 = VFMA(LDK(KP923879532), Te5, Te4); Chris@10: TdY = VFNMS(LDK(KP923879532), TdX, TdU); Chris@10: Te0 = VFMA(LDK(KP923879532), TdX, TdU); Chris@10: TcA = VFNMS(LDK(KP707106781), Tbl, Tb6); Chris@10: Tbm = VFMA(LDK(KP707106781), Tbl, Tb6); Chris@10: Tcp = VFNMS(LDK(KP923879532), TaY, Tat); Chris@10: TaZ = VFMA(LDK(KP923879532), TaY, Tat); Chris@10: Tcs = VFNMS(LDK(KP923879532), Tch, Tce); Chris@10: Tci = VFMA(LDK(KP923879532), Tch, Tce); Chris@10: ST(&(xo[WS(os, 88)]), VFNMSI(Te6, Te3), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 40)]), VFMAI(Te6, Te3), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 104)]), VFMAI(Te8, Te7), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 24)]), VFNMSI(Te8, Te7), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 8)]), VFMAI(Te0, TdZ), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 120)]), VFNMSI(Te0, TdZ), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 72)]), VFMAI(TdY, TdR), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 56)]), VFNMSI(TdY, TdR), ovs, &(xo[0])); Chris@10: Tbv = VFMA(LDK(KP707106781), Tbu, Tbt); Chris@10: TcB = VFNMS(LDK(KP707106781), Tbu, Tbt); Chris@10: TcD = VFNMS(LDK(KP707106781), TbS, TbD); Chris@10: TbT = VFMA(LDK(KP707106781), TbS, TbD); Chris@10: Tc2 = VFMA(LDK(KP707106781), Tc1, Tc0); Chris@10: TcE = VFNMS(LDK(KP707106781), Tc1, Tc0); Chris@10: } Chris@10: { Chris@10: V TcR, Tcz, TcU, TcK, Tcq, Tcl, Tct, Tc4; Chris@10: { Chris@10: V Tcj, Tbw, Tck, Tc3; Chris@10: Tcj = VFMA(LDK(KP198912367), Tbm, Tbv); Chris@10: Tbw = VFNMS(LDK(KP198912367), Tbv, Tbm); Chris@10: Tck = VFMA(LDK(KP198912367), TbT, Tc2); Chris@10: Tc3 = VFNMS(LDK(KP198912367), Tc2, TbT); Chris@10: TcR = VFNMS(LDK(KP923879532), Tcy, Tcx); Chris@10: Tcz = VFMA(LDK(KP923879532), Tcy, Tcx); Chris@10: TcU = VFMA(LDK(KP923879532), TcJ, TcI); Chris@10: TcK = VFNMS(LDK(KP923879532), TcJ, TcI); Chris@10: Tcq = VADD(Tcj, Tck); Chris@10: Tcl = VSUB(Tcj, Tck); Chris@10: Tct = VSUB(Tbw, Tc3); Chris@10: Tc4 = VADD(Tbw, Tc3); Chris@10: } Chris@10: { Chris@10: V TfN, Tel, TfY, Tfu, Tfv, Tfw, TcT, TcX, TcQ, TcO, TcW, TcY, TcP, TcH, TfZ; Chris@10: V TeE; Chris@10: { Chris@10: V Teu, TcS, TcN, TcV, TcG, TeD; Chris@10: TfN = VFNMS(LDK(KP923879532), Tek, Ted); Chris@10: Tel = VFMA(LDK(KP923879532), Tek, Ted); Chris@10: { Chris@10: V TcL, TcC, Tcr, Tcv; Chris@10: TcL = VFNMS(LDK(KP668178637), TcA, TcB); Chris@10: TcC = VFMA(LDK(KP668178637), TcB, TcA); Chris@10: Tcr = VFNMS(LDK(KP980785280), Tcq, Tcp); Chris@10: Tcv = VFMA(LDK(KP980785280), Tcq, Tcp); Chris@10: { Chris@10: V Tco, Tcm, Tcu, Tcw; Chris@10: Tco = VFMA(LDK(KP980785280), Tcl, Tci); Chris@10: Tcm = VFNMS(LDK(KP980785280), Tcl, Tci); Chris@10: Tcu = VFMA(LDK(KP980785280), Tct, Tcs); Chris@10: Tcw = VFNMS(LDK(KP980785280), Tct, Tcs); Chris@10: { Chris@10: V Tcn, Tc5, TcM, TcF; Chris@10: Tcn = VFMA(LDK(KP980785280), Tc4, TaZ); Chris@10: Tc5 = VFNMS(LDK(KP980785280), Tc4, TaZ); Chris@10: TcM = VFNMS(LDK(KP668178637), TcD, TcE); Chris@10: TcF = VFMA(LDK(KP668178637), TcE, TcD); Chris@10: TfY = VFNMS(LDK(KP923879532), Tft, Tfq); Chris@10: Tfu = VFMA(LDK(KP923879532), Tft, Tfq); Chris@10: Tfv = VFMA(LDK(KP198912367), Teq, Tet); Chris@10: Teu = VFNMS(LDK(KP198912367), Tet, Teq); Chris@10: ST(&(xo[WS(os, 92)]), VFNMSI(Tcu, Tcr), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 36)]), VFMAI(Tcu, Tcr), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 100)]), VFMAI(Tcw, Tcv), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 28)]), VFNMSI(Tcw, Tcv), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 4)]), VFMAI(Tco, Tcn), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 124)]), VFNMSI(Tco, Tcn), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 68)]), VFMAI(Tcm, Tc5), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 60)]), VFNMSI(Tcm, Tc5), ovs, &(xo[0])); Chris@10: TcS = VADD(TcL, TcM); Chris@10: TcN = VSUB(TcL, TcM); Chris@10: TcV = VSUB(TcC, TcF); Chris@10: TcG = VADD(TcC, TcF); Chris@10: TeD = VFNMS(LDK(KP198912367), TeC, Tez); Chris@10: Tfw = VFMA(LDK(KP198912367), Tez, TeC); Chris@10: } Chris@10: } Chris@10: } Chris@10: TcT = VFMA(LDK(KP831469612), TcS, TcR); Chris@10: TcX = VFNMS(LDK(KP831469612), TcS, TcR); Chris@10: TcQ = VFMA(LDK(KP831469612), TcN, TcK); Chris@10: TcO = VFNMS(LDK(KP831469612), TcN, TcK); Chris@10: TcW = VFNMS(LDK(KP831469612), TcV, TcU); Chris@10: TcY = VFMA(LDK(KP831469612), TcV, TcU); Chris@10: TcP = VFMA(LDK(KP831469612), TcG, Tcz); Chris@10: TcH = VFNMS(LDK(KP831469612), TcG, Tcz); Chris@10: TfZ = VSUB(Teu, TeD); Chris@10: TeE = VADD(Teu, TeD); Chris@10: } Chris@10: { Chris@10: V TfQ, TeS, TfO, Tfx, TeZ, TfR, Tfd, Tfk; Chris@10: TfQ = VFNMS(LDK(KP923879532), TeR, TeK); Chris@10: TeS = VFMA(LDK(KP923879532), TeR, TeK); Chris@10: ST(&(xo[WS(os, 84)]), VFMAI(TcW, TcT), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 44)]), VFNMSI(TcW, TcT), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 108)]), VFNMSI(TcY, TcX), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 20)]), VFMAI(TcY, TcX), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 116)]), VFMAI(TcQ, TcP), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 12)]), VFNMSI(TcQ, TcP), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 52)]), VFMAI(TcO, TcH), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 76)]), VFNMSI(TcO, TcH), ovs, &(xo[0])); Chris@10: Tg0 = VFNMS(LDK(KP980785280), TfZ, TfY); Chris@10: Tga = VFMA(LDK(KP980785280), TfZ, TfY); Chris@10: TfF = VFNMS(LDK(KP980785280), TeE, Tel); Chris@10: TeF = VFMA(LDK(KP980785280), TeE, Tel); Chris@10: TfO = VADD(Tfv, Tfw); Chris@10: Tfx = VSUB(Tfv, Tfw); Chris@10: TeZ = VFMA(LDK(KP923879532), TeY, TeV); Chris@10: TfR = VFNMS(LDK(KP923879532), TeY, TeV); Chris@10: TfT = VFNMS(LDK(KP923879532), Tfc, Tf5); Chris@10: Tfd = VFMA(LDK(KP923879532), Tfc, Tf5); Chris@10: Tfk = VFMA(LDK(KP923879532), Tfj, Tfg); Chris@10: TfU = VFNMS(LDK(KP923879532), Tfj, Tfg); Chris@10: TfP = VFMA(LDK(KP980785280), TfO, TfN); Chris@10: Tg7 = VFNMS(LDK(KP980785280), TfO, TfN); Chris@10: TfI = VFNMS(LDK(KP980785280), Tfx, Tfu); Chris@10: Tfy = VFMA(LDK(KP980785280), Tfx, Tfu); Chris@10: Tfz = VFMA(LDK(KP098491403), TeS, TeZ); Chris@10: Tf0 = VFNMS(LDK(KP098491403), TeZ, TeS); Chris@10: TfA = VFMA(LDK(KP098491403), Tfd, Tfk); Chris@10: Tfl = VFNMS(LDK(KP098491403), Tfk, Tfd); Chris@10: Tg1 = VFNMS(LDK(KP820678790), TfQ, TfR); Chris@10: TfS = VFMA(LDK(KP820678790), TfR, TfQ); Chris@10: } Chris@10: } Chris@10: } Chris@10: } Chris@10: } Chris@10: { Chris@10: V T8x, T8y, T8F, T8w, T8k, T8f, T8n, T80, T9l, T76, T87, T8U, T89, T7e, T7l; Chris@10: V T8a; Chris@10: { Chris@10: V The, Tho, TgT, Tgp, Th7, Th8, Thf, Th6, Th3, Thl, TgW, TgM, TgU, TgP, TgX; Chris@10: V TgE; Chris@10: { Chris@10: V Th1, TgI, TgJ, TgK; Chris@10: { Chris@10: V Tgh, Thc, Tgk, TfG, TfB, TfJ, Tfm, Tg2, TfV, Tgn, TfL, TfH; Chris@10: Th1 = VFMA(LDK(KP923879532), Tgg, Tgf); Chris@10: Tgh = VFNMS(LDK(KP923879532), Tgg, Tgf); Chris@10: Thc = VFNMS(LDK(KP923879532), TgH, TgG); Chris@10: TgI = VFMA(LDK(KP923879532), TgH, TgG); Chris@10: TgJ = VFMA(LDK(KP668178637), Tgi, Tgj); Chris@10: Tgk = VFNMS(LDK(KP668178637), Tgj, Tgi); Chris@10: TfG = VADD(Tfz, TfA); Chris@10: TfB = VSUB(Tfz, TfA); Chris@10: TfJ = VSUB(Tf0, Tfl); Chris@10: Tfm = VADD(Tf0, Tfl); Chris@10: Tg2 = VFNMS(LDK(KP820678790), TfT, TfU); Chris@10: TfV = VFMA(LDK(KP820678790), TfU, TfT); Chris@10: Tgn = VFNMS(LDK(KP668178637), Tgm, Tgl); Chris@10: TgK = VFMA(LDK(KP668178637), Tgl, Tgm); Chris@10: TfL = VFMA(LDK(KP995184726), TfG, TfF); Chris@10: TfH = VFNMS(LDK(KP995184726), TfG, TfF); Chris@10: { Chris@10: V TfE, TfC, TfM, TfK; Chris@10: TfE = VFMA(LDK(KP995184726), TfB, Tfy); Chris@10: TfC = VFNMS(LDK(KP995184726), TfB, Tfy); Chris@10: TfM = VFNMS(LDK(KP995184726), TfJ, TfI); Chris@10: TfK = VFMA(LDK(KP995184726), TfJ, TfI); Chris@10: { Chris@10: V TfD, Tfn, Tg8, Tg3; Chris@10: TfD = VFMA(LDK(KP995184726), Tfm, TeF); Chris@10: Tfn = VFNMS(LDK(KP995184726), Tfm, TeF); Chris@10: Tg8 = VADD(Tg1, Tg2); Chris@10: Tg3 = VSUB(Tg1, Tg2); Chris@10: { Chris@10: V Tgb, TfW, Thd, Tgo; Chris@10: Tgb = VSUB(TfS, TfV); Chris@10: TfW = VADD(TfS, TfV); Chris@10: Thd = VSUB(Tgk, Tgn); Chris@10: Tgo = VADD(Tgk, Tgn); Chris@10: ST(&(xo[WS(os, 98)]), VFMAI(TfM, TfL), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 30)]), VFNMSI(TfM, TfL), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 94)]), VFNMSI(TfK, TfH), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 34)]), VFMAI(TfK, TfH), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 2)]), VFMAI(TfE, TfD), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 126)]), VFNMSI(TfE, TfD), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 66)]), VFMAI(TfC, Tfn), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 62)]), VFNMSI(TfC, Tfn), ovs, &(xo[0])); Chris@10: { Chris@10: V Tgd, Tg9, Tg6, Tg4; Chris@10: Tgd = VFNMS(LDK(KP773010453), Tg8, Tg7); Chris@10: Tg9 = VFMA(LDK(KP773010453), Tg8, Tg7); Chris@10: Tg6 = VFMA(LDK(KP773010453), Tg3, Tg0); Chris@10: Tg4 = VFNMS(LDK(KP773010453), Tg3, Tg0); Chris@10: { Chris@10: V Tge, Tgc, Tg5, TfX; Chris@10: Tge = VFMA(LDK(KP773010453), Tgb, Tga); Chris@10: Tgc = VFNMS(LDK(KP773010453), Tgb, Tga); Chris@10: Tg5 = VFMA(LDK(KP773010453), TfW, TfP); Chris@10: TfX = VFNMS(LDK(KP773010453), TfW, TfP); Chris@10: The = VFMA(LDK(KP831469612), Thd, Thc); Chris@10: Tho = VFNMS(LDK(KP831469612), Thd, Thc); Chris@10: TgT = VFMA(LDK(KP831469612), Tgo, Tgh); Chris@10: Tgp = VFNMS(LDK(KP831469612), Tgo, Tgh); Chris@10: ST(&(xo[WS(os, 110)]), VFNMSI(Tge, Tgd), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 18)]), VFMAI(Tge, Tgd), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 82)]), VFMAI(Tgc, Tg9), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 46)]), VFNMSI(Tgc, Tg9), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 114)]), VFMAI(Tg6, Tg5), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 14)]), VFNMSI(Tg6, Tg5), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 50)]), VFMAI(Tg4, TfX), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 78)]), VFNMSI(Tg4, TfX), ovs, &(xo[0])); Chris@10: } Chris@10: } Chris@10: } Chris@10: } Chris@10: } Chris@10: } Chris@10: { Chris@10: V Th4, Tgs, Tgv, Th5, Tgz, TgC, Th2, TgL; Chris@10: Th4 = VFMA(LDK(KP923879532), Tgr, Tgq); Chris@10: Tgs = VFNMS(LDK(KP923879532), Tgr, Tgq); Chris@10: Tgv = VFMA(LDK(KP923879532), Tgu, Tgt); Chris@10: Th5 = VFNMS(LDK(KP923879532), Tgu, Tgt); Chris@10: Th7 = VFMA(LDK(KP923879532), Tgy, Tgx); Chris@10: Tgz = VFNMS(LDK(KP923879532), Tgy, Tgx); Chris@10: TgC = VFMA(LDK(KP923879532), TgB, TgA); Chris@10: Th8 = VFNMS(LDK(KP923879532), TgB, TgA); Chris@10: Th2 = VADD(TgJ, TgK); Chris@10: TgL = VSUB(TgJ, TgK); Chris@10: { Chris@10: V TgN, Tgw, TgO, TgD; Chris@10: TgN = VFMA(LDK(KP534511135), Tgs, Tgv); Chris@10: Tgw = VFNMS(LDK(KP534511135), Tgv, Tgs); Chris@10: TgO = VFMA(LDK(KP534511135), Tgz, TgC); Chris@10: TgD = VFNMS(LDK(KP534511135), TgC, Tgz); Chris@10: Thf = VFNMS(LDK(KP303346683), Th4, Th5); Chris@10: Th6 = VFMA(LDK(KP303346683), Th5, Th4); Chris@10: Th3 = VFMA(LDK(KP831469612), Th2, Th1); Chris@10: Thl = VFNMS(LDK(KP831469612), Th2, Th1); Chris@10: TgW = VFNMS(LDK(KP831469612), TgL, TgI); Chris@10: TgM = VFMA(LDK(KP831469612), TgL, TgI); Chris@10: TgU = VADD(TgN, TgO); Chris@10: TgP = VSUB(TgN, TgO); Chris@10: TgX = VSUB(Tgw, TgD); Chris@10: TgE = VADD(Tgw, TgD); Chris@10: } Chris@10: } Chris@10: } Chris@10: { Chris@10: V T8u, T8v, T7R, T8d, T7G, Thm, Thh, Thp, Tha, T7Y, Thr, Thn; Chris@10: { Chris@10: V T7y, T7F, TgZ, TgV; Chris@10: T8u = VFNMS(LDK(KP831469612), T7x, T7q); Chris@10: T7y = VFMA(LDK(KP831469612), T7x, T7q); Chris@10: T7F = VFMA(LDK(KP831469612), T7E, T7B); Chris@10: T8v = VFNMS(LDK(KP831469612), T7E, T7B); Chris@10: T8x = VFNMS(LDK(KP831469612), T7Q, T7J); Chris@10: T7R = VFMA(LDK(KP831469612), T7Q, T7J); Chris@10: TgZ = VFMA(LDK(KP881921264), TgU, TgT); Chris@10: TgV = VFNMS(LDK(KP881921264), TgU, TgT); Chris@10: { Chris@10: V TgS, TgQ, Th0, TgY; Chris@10: TgS = VFMA(LDK(KP881921264), TgP, TgM); Chris@10: TgQ = VFNMS(LDK(KP881921264), TgP, TgM); Chris@10: Th0 = VFNMS(LDK(KP881921264), TgX, TgW); Chris@10: TgY = VFMA(LDK(KP881921264), TgX, TgW); Chris@10: { Chris@10: V TgR, TgF, Thg, Th9; Chris@10: TgR = VFMA(LDK(KP881921264), TgE, Tgp); Chris@10: TgF = VFNMS(LDK(KP881921264), TgE, Tgp); Chris@10: Thg = VFNMS(LDK(KP303346683), Th7, Th8); Chris@10: Th9 = VFMA(LDK(KP303346683), Th8, Th7); Chris@10: T8d = VFNMS(LDK(KP148335987), T7y, T7F); Chris@10: T7G = VFMA(LDK(KP148335987), T7F, T7y); Chris@10: ST(&(xo[WS(os, 106)]), VFMAI(Th0, TgZ), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 22)]), VFNMSI(Th0, TgZ), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 86)]), VFNMSI(TgY, TgV), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 42)]), VFMAI(TgY, TgV), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 10)]), VFMAI(TgS, TgR), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 118)]), VFNMSI(TgS, TgR), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 74)]), VFMAI(TgQ, TgF), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 54)]), VFNMSI(TgQ, TgF), ovs, &(xo[0])); Chris@10: Thm = VADD(Thf, Thg); Chris@10: Thh = VSUB(Thf, Thg); Chris@10: Thp = VSUB(Th6, Th9); Chris@10: Tha = VADD(Th6, Th9); Chris@10: T7Y = VFMA(LDK(KP831469612), T7X, T7U); Chris@10: T8y = VFNMS(LDK(KP831469612), T7X, T7U); Chris@10: } Chris@10: } Chris@10: } Chris@10: Thr = VFNMS(LDK(KP956940335), Thm, Thl); Chris@10: Thn = VFMA(LDK(KP956940335), Thm, Thl); Chris@10: { Chris@10: V Thk, Thi, Ths, Thq; Chris@10: Thk = VFMA(LDK(KP956940335), Thh, The); Chris@10: Thi = VFNMS(LDK(KP956940335), Thh, The); Chris@10: Ths = VFMA(LDK(KP956940335), Thp, Tho); Chris@10: Thq = VFNMS(LDK(KP956940335), Thp, Tho); Chris@10: { Chris@10: V Thj, Thb, T8e, T7Z; Chris@10: Thj = VFMA(LDK(KP956940335), Tha, Th3); Chris@10: Thb = VFNMS(LDK(KP956940335), Tha, Th3); Chris@10: T8e = VFNMS(LDK(KP148335987), T7R, T7Y); Chris@10: T7Z = VFMA(LDK(KP148335987), T7Y, T7R); Chris@10: T8F = VFMA(LDK(KP741650546), T8u, T8v); Chris@10: T8w = VFNMS(LDK(KP741650546), T8v, T8u); Chris@10: ST(&(xo[WS(os, 102)]), VFNMSI(Ths, Thr), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 26)]), VFMAI(Ths, Thr), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 90)]), VFMAI(Thq, Thn), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 38)]), VFNMSI(Thq, Thn), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 122)]), VFMAI(Thk, Thj), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 6)]), VFNMSI(Thk, Thj), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 58)]), VFMAI(Thi, Thb), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 70)]), VFNMSI(Thi, Thb), ovs, &(xo[0])); Chris@10: T8k = VADD(T8d, T8e); Chris@10: T8f = VSUB(T8d, T8e); Chris@10: T8n = VSUB(T7G, T7Z); Chris@10: T80 = VADD(T7G, T7Z); Chris@10: } Chris@10: } Chris@10: T9l = VSUB(T72, T75); Chris@10: T76 = VADD(T72, T75); Chris@10: T87 = VSUB(T85, T86); Chris@10: T8U = VADD(T85, T86); Chris@10: T89 = VFNMS(LDK(KP303346683), T7a, T7d); Chris@10: T7e = VFMA(LDK(KP303346683), T7d, T7a); Chris@10: T7l = VFMA(LDK(KP303346683), T7k, T7h); Chris@10: T8a = VFNMS(LDK(KP303346683), T7h, T7k); Chris@10: } Chris@10: } Chris@10: { Chris@10: V T11, T5h, T5a, T55, T5d, T4K, T5C, T5x, T5F, T5q, T4X, T4Z, T1C, T2d, T50; Chris@10: { Chris@10: V T5k, T3g, T3t, T5l, T5n, T4v, T4I, T5o, T8G, T8z; Chris@10: T5k = VFNMS(LDK(KP980785280), T3f, T2G); Chris@10: T3g = VFMA(LDK(KP980785280), T3f, T2G); Chris@10: T8G = VFMA(LDK(KP741650546), T8x, T8y); Chris@10: T8z = VFNMS(LDK(KP741650546), T8y, T8x); Chris@10: { Chris@10: V T8r, T77, T8C, T88; Chris@10: T8r = VFNMS(LDK(KP831469612), T76, T6Z); Chris@10: T77 = VFMA(LDK(KP831469612), T76, T6Z); Chris@10: T8C = VFNMS(LDK(KP831469612), T87, T84); Chris@10: T88 = VFMA(LDK(KP831469612), T87, T84); Chris@10: { Chris@10: V T8D, T7m, T8s, T8b; Chris@10: T8D = VSUB(T7e, T7l); Chris@10: T7m = VADD(T7e, T7l); Chris@10: T8s = VADD(T89, T8a); Chris@10: T8b = VSUB(T89, T8a); Chris@10: { Chris@10: V T8M, T8H, T8P, T8A; Chris@10: T8M = VADD(T8F, T8G); Chris@10: T8H = VSUB(T8F, T8G); Chris@10: T8P = VSUB(T8w, T8z); Chris@10: T8A = VADD(T8w, T8z); Chris@10: { Chris@10: V T8E, T8O, T8j, T7n; Chris@10: T8E = VFMA(LDK(KP956940335), T8D, T8C); Chris@10: T8O = VFNMS(LDK(KP956940335), T8D, T8C); Chris@10: T8j = VFNMS(LDK(KP956940335), T7m, T77); Chris@10: T7n = VFMA(LDK(KP956940335), T7m, T77); Chris@10: { Chris@10: V T8t, T8L, T8m, T8c; Chris@10: T8t = VFNMS(LDK(KP956940335), T8s, T8r); Chris@10: T8L = VFMA(LDK(KP956940335), T8s, T8r); Chris@10: T8m = VFNMS(LDK(KP956940335), T8b, T88); Chris@10: T8c = VFMA(LDK(KP956940335), T8b, T88); Chris@10: { Chris@10: V T8K, T8I, T8S, T8Q; Chris@10: T8K = VFMA(LDK(KP803207531), T8H, T8E); Chris@10: T8I = VFNMS(LDK(KP803207531), T8H, T8E); Chris@10: T8S = VFNMS(LDK(KP803207531), T8P, T8O); Chris@10: T8Q = VFMA(LDK(KP803207531), T8P, T8O); Chris@10: { Chris@10: V T8p, T8l, T8h, T81; Chris@10: T8p = VFNMS(LDK(KP989176509), T8k, T8j); Chris@10: T8l = VFMA(LDK(KP989176509), T8k, T8j); Chris@10: T8h = VFMA(LDK(KP989176509), T80, T7n); Chris@10: T81 = VFNMS(LDK(KP989176509), T80, T7n); Chris@10: { Chris@10: V T8J, T8B, T8R, T8N; Chris@10: T8J = VFMA(LDK(KP803207531), T8A, T8t); Chris@10: T8B = VFNMS(LDK(KP803207531), T8A, T8t); Chris@10: T8R = VFMA(LDK(KP803207531), T8M, T8L); Chris@10: T8N = VFNMS(LDK(KP803207531), T8M, T8L); Chris@10: { Chris@10: V T8q, T8o, T8i, T8g; Chris@10: T8q = VFMA(LDK(KP989176509), T8n, T8m); Chris@10: T8o = VFNMS(LDK(KP989176509), T8n, T8m); Chris@10: T8i = VFMA(LDK(KP989176509), T8f, T8c); Chris@10: T8g = VFNMS(LDK(KP989176509), T8f, T8c); Chris@10: ST(&(xo[WS(os, 13)]), VFMAI(T8K, T8J), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 115)]), VFNMSI(T8K, T8J), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 77)]), VFMAI(T8I, T8B), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 51)]), VFNMSI(T8I, T8B), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 109)]), VFMAI(T8S, T8R), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 19)]), VFNMSI(T8S, T8R), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 83)]), VFNMSI(T8Q, T8N), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 45)]), VFMAI(T8Q, T8N), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 99)]), VFNMSI(T8q, T8p), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 29)]), VFMAI(T8q, T8p), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 93)]), VFMAI(T8o, T8l), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 35)]), VFNMSI(T8o, T8l), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 125)]), VFMAI(T8i, T8h), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 3)]), VFNMSI(T8i, T8h), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 61)]), VFMAI(T8g, T81), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 67)]), VFNMSI(T8g, T81), ovs, &(xo[WS(os, 1)])); Chris@10: T3t = VFMA(LDK(KP980785280), T3s, T3p); Chris@10: T5l = VFNMS(LDK(KP980785280), T3s, T3p); Chris@10: } Chris@10: } Chris@10: } Chris@10: } Chris@10: } Chris@10: } Chris@10: } Chris@10: } Chris@10: } Chris@10: T5n = VFNMS(LDK(KP980785280), T4u, T3V); Chris@10: T4v = VFMA(LDK(KP980785280), T4u, T3V); Chris@10: T4I = VFMA(LDK(KP980785280), T4H, T4E); Chris@10: T5o = VFNMS(LDK(KP980785280), T4H, T4E); Chris@10: { Chris@10: V T53, T3u, T54, T4J, T5v, T5m, T5w, T5p, T10; Chris@10: T6b = VSUB(TI, TZ); Chris@10: T10 = VADD(TI, TZ); Chris@10: T53 = VFMA(LDK(KP049126849), T3g, T3t); Chris@10: T3u = VFNMS(LDK(KP049126849), T3t, T3g); Chris@10: T54 = VFMA(LDK(KP049126849), T4v, T4I); Chris@10: T4J = VFNMS(LDK(KP049126849), T4I, T4v); Chris@10: T5v = VFNMS(LDK(KP906347169), T5k, T5l); Chris@10: T5m = VFMA(LDK(KP906347169), T5l, T5k); Chris@10: T5w = VFNMS(LDK(KP906347169), T5n, T5o); Chris@10: T5p = VFMA(LDK(KP906347169), T5o, T5n); Chris@10: T11 = VFMA(LDK(KP980785280), T10, Tr); Chris@10: T5h = VFNMS(LDK(KP980785280), T10, Tr); Chris@10: T5a = VADD(T53, T54); Chris@10: T55 = VSUB(T53, T54); Chris@10: T5d = VSUB(T3u, T4J); Chris@10: T4K = VADD(T3u, T4J); Chris@10: T5C = VADD(T5v, T5w); Chris@10: T5x = VSUB(T5v, T5w); Chris@10: T5F = VSUB(T5m, T5p); Chris@10: T5q = VADD(T5m, T5p); Chris@10: T4X = VSUB(T4V, T4W); Chris@10: T5K = VADD(T4V, T4W); Chris@10: } Chris@10: T4Z = VFMA(LDK(KP098491403), T1s, T1B); Chris@10: T1C = VFNMS(LDK(KP098491403), T1B, T1s); Chris@10: T2d = VFNMS(LDK(KP098491403), T2c, T23); Chris@10: T50 = VFMA(LDK(KP098491403), T23, T2c); Chris@10: } Chris@10: { Chris@10: V T9y, T9t, T9B, T9i, T9n, T9o, T9F, T8V, T9Q, T9m, T9R, T92, Ta0, T9V, Ta3; Chris@10: V T9O; Chris@10: { Chris@10: V T9I, T9J, T9L, T9d, T5s, T4Y, T5t, T2e, T5i, T51, T9r, T9a, T9g, T9M, T96; Chris@10: V T99; Chris@10: T9I = VFMA(LDK(KP831469612), T95, T94); Chris@10: T96 = VFNMS(LDK(KP831469612), T95, T94); Chris@10: T99 = VFNMS(LDK(KP831469612), T98, T97); Chris@10: T9J = VFMA(LDK(KP831469612), T98, T97); Chris@10: T9L = VFMA(LDK(KP831469612), T9c, T9b); Chris@10: T9d = VFNMS(LDK(KP831469612), T9c, T9b); Chris@10: T5s = VFNMS(LDK(KP980785280), T4X, T4U); Chris@10: T4Y = VFMA(LDK(KP980785280), T4X, T4U); Chris@10: T5t = VSUB(T1C, T2d); Chris@10: T2e = VADD(T1C, T2d); Chris@10: T5i = VADD(T4Z, T50); Chris@10: T51 = VSUB(T4Z, T50); Chris@10: T9r = VFNMS(LDK(KP599376933), T96, T99); Chris@10: T9a = VFMA(LDK(KP599376933), T99, T96); Chris@10: T9g = VFNMS(LDK(KP831469612), T9f, T9e); Chris@10: T9M = VFMA(LDK(KP831469612), T9f, T9e); Chris@10: { Chris@10: V T5u, T5E, T8Y, T91; Chris@10: T5u = VFNMS(LDK(KP995184726), T5t, T5s); Chris@10: T5E = VFMA(LDK(KP995184726), T5t, T5s); Chris@10: { Chris@10: V T59, T2f, T5j, T5B; Chris@10: T59 = VFNMS(LDK(KP995184726), T2e, T11); Chris@10: T2f = VFMA(LDK(KP995184726), T2e, T11); Chris@10: T5j = VFMA(LDK(KP995184726), T5i, T5h); Chris@10: T5B = VFNMS(LDK(KP995184726), T5i, T5h); Chris@10: { Chris@10: V T5c, T52, T9s, T9h; Chris@10: T5c = VFNMS(LDK(KP995184726), T51, T4Y); Chris@10: T52 = VFMA(LDK(KP995184726), T51, T4Y); Chris@10: T9s = VFNMS(LDK(KP599376933), T9d, T9g); Chris@10: T9h = VFMA(LDK(KP599376933), T9g, T9d); Chris@10: { Chris@10: V T5A, T5y, T5I, T5G; Chris@10: T5A = VFMA(LDK(KP740951125), T5x, T5u); Chris@10: T5y = VFNMS(LDK(KP740951125), T5x, T5u); Chris@10: T5I = VFMA(LDK(KP740951125), T5F, T5E); Chris@10: T5G = VFNMS(LDK(KP740951125), T5F, T5E); Chris@10: { Chris@10: V T5f, T5b, T57, T4L; Chris@10: T5f = VFMA(LDK(KP998795456), T5a, T59); Chris@10: T5b = VFNMS(LDK(KP998795456), T5a, T59); Chris@10: T57 = VFMA(LDK(KP998795456), T4K, T2f); Chris@10: T4L = VFNMS(LDK(KP998795456), T4K, T2f); Chris@10: { Chris@10: V T5z, T5r, T5H, T5D; Chris@10: T5z = VFMA(LDK(KP740951125), T5q, T5j); Chris@10: T5r = VFNMS(LDK(KP740951125), T5q, T5j); Chris@10: T5H = VFNMS(LDK(KP740951125), T5C, T5B); Chris@10: T5D = VFMA(LDK(KP740951125), T5C, T5B); Chris@10: { Chris@10: V T5g, T5e, T58, T56; Chris@10: T5g = VFNMS(LDK(KP998795456), T5d, T5c); Chris@10: T5e = VFMA(LDK(KP998795456), T5d, T5c); Chris@10: T58 = VFMA(LDK(KP998795456), T55, T52); Chris@10: T56 = VFNMS(LDK(KP998795456), T55, T52); Chris@10: T9y = VADD(T9r, T9s); Chris@10: T9t = VSUB(T9r, T9s); Chris@10: T9B = VSUB(T9a, T9h); Chris@10: T9i = VADD(T9a, T9h); Chris@10: ST(&(xo[WS(os, 113)]), VFMAI(T5A, T5z), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 15)]), VFNMSI(T5A, T5z), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 49)]), VFMAI(T5y, T5r), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 79)]), VFNMSI(T5y, T5r), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 111)]), VFNMSI(T5I, T5H), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 17)]), VFMAI(T5I, T5H), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 81)]), VFMAI(T5G, T5D), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 47)]), VFNMSI(T5G, T5D), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 97)]), VFMAI(T5g, T5f), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 31)]), VFNMSI(T5g, T5f), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 95)]), VFNMSI(T5e, T5b), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 33)]), VFMAI(T5e, T5b), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 1)]), VFMAI(T58, T57), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 127)]), VFNMSI(T58, T57), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 65)]), VFMAI(T56, T4L), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 63)]), VFNMSI(T56, T4L), ovs, &(xo[WS(os, 1)])); Chris@10: } Chris@10: } Chris@10: } Chris@10: } Chris@10: } Chris@10: } Chris@10: T9n = VFNMS(LDK(KP534511135), T8W, T8X); Chris@10: T8Y = VFMA(LDK(KP534511135), T8X, T8W); Chris@10: T91 = VFMA(LDK(KP534511135), T90, T8Z); Chris@10: T9o = VFNMS(LDK(KP534511135), T8Z, T90); Chris@10: { Chris@10: V T9T, T9K, T9U, T9N; Chris@10: T9T = VFMA(LDK(KP250486960), T9I, T9J); Chris@10: T9K = VFNMS(LDK(KP250486960), T9J, T9I); Chris@10: T9U = VFMA(LDK(KP250486960), T9L, T9M); Chris@10: T9N = VFNMS(LDK(KP250486960), T9M, T9L); Chris@10: T9F = VFNMS(LDK(KP831469612), T8U, T8T); Chris@10: T8V = VFMA(LDK(KP831469612), T8U, T8T); Chris@10: T9Q = VFMA(LDK(KP831469612), T9l, T9k); Chris@10: T9m = VFNMS(LDK(KP831469612), T9l, T9k); Chris@10: T9R = VSUB(T8Y, T91); Chris@10: T92 = VADD(T8Y, T91); Chris@10: Ta0 = VADD(T9T, T9U); Chris@10: T9V = VSUB(T9T, T9U); Chris@10: Ta3 = VSUB(T9K, T9N); Chris@10: T9O = VADD(T9K, T9N); Chris@10: } Chris@10: } Chris@10: } Chris@10: { Chris@10: V T6y, T6z, T63, T9Y, T9W, Ta6, Ta4, T9D, T9z, T9v, T9j, T6h, T60, T9H, T9Z; Chris@10: V T9A, T9q, T66, T9X, T9P; Chris@10: { Chris@10: V T5W, T9S, Ta2, T9x, T93, T5Z, T9G, T9p; Chris@10: T6y = VFMA(LDK(KP980785280), T5V, T5U); Chris@10: T5W = VFNMS(LDK(KP980785280), T5V, T5U); Chris@10: T9S = VFMA(LDK(KP881921264), T9R, T9Q); Chris@10: Ta2 = VFNMS(LDK(KP881921264), T9R, T9Q); Chris@10: T9x = VFNMS(LDK(KP881921264), T92, T8V); Chris@10: T93 = VFMA(LDK(KP881921264), T92, T8V); Chris@10: T5Z = VFMA(LDK(KP980785280), T5Y, T5X); Chris@10: T6z = VFNMS(LDK(KP980785280), T5Y, T5X); Chris@10: T6B = VFMA(LDK(KP980785280), T62, T61); Chris@10: T63 = VFNMS(LDK(KP980785280), T62, T61); Chris@10: T9G = VADD(T9n, T9o); Chris@10: T9p = VSUB(T9n, T9o); Chris@10: T9Y = VFMA(LDK(KP970031253), T9V, T9S); Chris@10: T9W = VFNMS(LDK(KP970031253), T9V, T9S); Chris@10: Ta6 = VFNMS(LDK(KP970031253), Ta3, Ta2); Chris@10: Ta4 = VFMA(LDK(KP970031253), Ta3, Ta2); Chris@10: T9D = VFNMS(LDK(KP857728610), T9y, T9x); Chris@10: T9z = VFMA(LDK(KP857728610), T9y, T9x); Chris@10: T9v = VFMA(LDK(KP857728610), T9i, T93); Chris@10: T9j = VFNMS(LDK(KP857728610), T9i, T93); Chris@10: T6h = VFMA(LDK(KP472964775), T5W, T5Z); Chris@10: T60 = VFNMS(LDK(KP472964775), T5Z, T5W); Chris@10: T9H = VFMA(LDK(KP881921264), T9G, T9F); Chris@10: T9Z = VFNMS(LDK(KP881921264), T9G, T9F); Chris@10: T9A = VFMA(LDK(KP881921264), T9p, T9m); Chris@10: T9q = VFNMS(LDK(KP881921264), T9p, T9m); Chris@10: T66 = VFMA(LDK(KP980785280), T65, T64); Chris@10: T6C = VFNMS(LDK(KP980785280), T65, T64); Chris@10: } Chris@10: T9X = VFMA(LDK(KP970031253), T9O, T9H); Chris@10: T9P = VFNMS(LDK(KP970031253), T9O, T9H); Chris@10: { Chris@10: V Ta5, Ta1, T9E, T9C; Chris@10: Ta5 = VFMA(LDK(KP970031253), Ta0, T9Z); Chris@10: Ta1 = VFNMS(LDK(KP970031253), Ta0, T9Z); Chris@10: T9E = VFMA(LDK(KP857728610), T9B, T9A); Chris@10: T9C = VFNMS(LDK(KP857728610), T9B, T9A); Chris@10: { Chris@10: V T9w, T9u, T6i, T67; Chris@10: T9w = VFMA(LDK(KP857728610), T9t, T9q); Chris@10: T9u = VFNMS(LDK(KP857728610), T9t, T9q); Chris@10: T6i = VFMA(LDK(KP472964775), T63, T66); Chris@10: T67 = VFNMS(LDK(KP472964775), T66, T63); Chris@10: T6J = VFNMS(LDK(KP357805721), T6y, T6z); Chris@10: T6A = VFMA(LDK(KP357805721), T6z, T6y); Chris@10: ST(&(xo[WS(os, 5)]), VFMAI(T9Y, T9X), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 123)]), VFNMSI(T9Y, T9X), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 69)]), VFMAI(T9W, T9P), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 59)]), VFNMSI(T9W, T9P), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 101)]), VFMAI(Ta6, Ta5), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 27)]), VFNMSI(Ta6, Ta5), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 91)]), VFNMSI(Ta4, Ta1), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 37)]), VFMAI(Ta4, Ta1), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 107)]), VFNMSI(T9E, T9D), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 21)]), VFMAI(T9E, T9D), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 85)]), VFMAI(T9C, T9z), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 43)]), VFNMSI(T9C, T9z), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 117)]), VFMAI(T9w, T9v), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 11)]), VFNMSI(T9w, T9v), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 53)]), VFMAI(T9u, T9j), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 75)]), VFNMSI(T9u, T9j), ovs, &(xo[WS(os, 1)])); Chris@10: T6o = VADD(T6h, T6i); Chris@10: T6j = VSUB(T6h, T6i); Chris@10: T6r = VSUB(T60, T67); Chris@10: T68 = VADD(T60, T67); Chris@10: } Chris@10: } Chris@10: T6d = VFMA(LDK(KP820678790), T5M, T5N); Chris@10: T5O = VFNMS(LDK(KP820678790), T5N, T5M); Chris@10: T5R = VFNMS(LDK(KP820678790), T5Q, T5P); Chris@10: T6e = VFMA(LDK(KP820678790), T5P, T5Q); Chris@10: } Chris@10: } Chris@10: } Chris@10: } Chris@10: } Chris@10: } Chris@10: T6D = VFMA(LDK(KP357805721), T6C, T6B); Chris@10: T6K = VFNMS(LDK(KP357805721), T6B, T6C); Chris@10: { Chris@10: V T5L, T6v, T6c, T6G; Chris@10: T5L = VFNMS(LDK(KP980785280), T5K, T5J); Chris@10: T6v = VFMA(LDK(KP980785280), T5K, T5J); Chris@10: T6c = VFMA(LDK(KP980785280), T6b, T6a); Chris@10: T6G = VFNMS(LDK(KP980785280), T6b, T6a); Chris@10: { Chris@10: V T5S, T6H, T6f, T6w; Chris@10: T5S = VADD(T5O, T5R); Chris@10: T6H = VSUB(T5O, T5R); Chris@10: T6f = VSUB(T6d, T6e); Chris@10: T6w = VADD(T6d, T6e); Chris@10: { Chris@10: V T6L, T6Q, T6E, T6T; Chris@10: T6L = VSUB(T6J, T6K); Chris@10: T6Q = VADD(T6J, T6K); Chris@10: T6E = VADD(T6A, T6D); Chris@10: T6T = VSUB(T6A, T6D); Chris@10: { Chris@10: V T6S, T6I, T5T, T6n; Chris@10: T6S = VFNMS(LDK(KP773010453), T6H, T6G); Chris@10: T6I = VFMA(LDK(KP773010453), T6H, T6G); Chris@10: T5T = VFNMS(LDK(KP773010453), T5S, T5L); Chris@10: T6n = VFMA(LDK(KP773010453), T5S, T5L); Chris@10: { Chris@10: V T6P, T6x, T6g, T6q; Chris@10: T6P = VFNMS(LDK(KP773010453), T6w, T6v); Chris@10: T6x = VFMA(LDK(KP773010453), T6w, T6v); Chris@10: T6g = VFMA(LDK(KP773010453), T6f, T6c); Chris@10: T6q = VFNMS(LDK(KP773010453), T6f, T6c); Chris@10: { Chris@10: V T6M, T6O, T6U, T6W; Chris@10: T6M = VFNMS(LDK(KP941544065), T6L, T6I); Chris@10: T6O = VFMA(LDK(KP941544065), T6L, T6I); Chris@10: T6U = VFNMS(LDK(KP941544065), T6T, T6S); Chris@10: T6W = VFMA(LDK(KP941544065), T6T, T6S); Chris@10: { Chris@10: V T6p, T6t, T69, T6l; Chris@10: T6p = VFNMS(LDK(KP903989293), T6o, T6n); Chris@10: T6t = VFMA(LDK(KP903989293), T6o, T6n); Chris@10: T69 = VFNMS(LDK(KP903989293), T68, T5T); Chris@10: T6l = VFMA(LDK(KP903989293), T68, T5T); Chris@10: { Chris@10: V T6F, T6N, T6R, T6V; Chris@10: T6F = VFNMS(LDK(KP941544065), T6E, T6x); Chris@10: T6N = VFMA(LDK(KP941544065), T6E, T6x); Chris@10: T6R = VFMA(LDK(KP941544065), T6Q, T6P); Chris@10: T6V = VFNMS(LDK(KP941544065), T6Q, T6P); Chris@10: { Chris@10: V T6s, T6u, T6k, T6m; Chris@10: T6s = VFMA(LDK(KP903989293), T6r, T6q); Chris@10: T6u = VFNMS(LDK(KP903989293), T6r, T6q); Chris@10: T6k = VFNMS(LDK(KP903989293), T6j, T6g); Chris@10: T6m = VFMA(LDK(KP903989293), T6j, T6g); Chris@10: ST(&(xo[WS(os, 121)]), VFMAI(T6O, T6N), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 7)]), VFNMSI(T6O, T6N), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 57)]), VFMAI(T6M, T6F), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 71)]), VFNMSI(T6M, T6F), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 103)]), VFNMSI(T6W, T6V), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 25)]), VFMAI(T6W, T6V), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 89)]), VFMAI(T6U, T6R), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 39)]), VFNMSI(T6U, T6R), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 105)]), VFMAI(T6u, T6t), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 23)]), VFNMSI(T6u, T6t), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 87)]), VFNMSI(T6s, T6p), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 41)]), VFMAI(T6s, T6p), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 9)]), VFMAI(T6m, T6l), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 119)]), VFNMSI(T6m, T6l), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 73)]), VFMAI(T6k, T69), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 55)]), VFNMSI(T6k, T69), ovs, &(xo[WS(os, 1)])); Chris@10: } Chris@10: } Chris@10: } Chris@10: } Chris@10: } Chris@10: } Chris@10: } Chris@10: } Chris@10: } Chris@10: } Chris@10: } Chris@10: VLEAVE(); Chris@10: } Chris@10: Chris@10: static const kdft_desc desc = { 128, XSIMD_STRING("n1bv_128"), {440, 0, 642, 0}, &GENUS, 0, 0, 0, 0 }; Chris@10: Chris@10: void XSIMD(codelet_n1bv_128) (planner *p) { Chris@10: X(kdft_register) (p, n1bv_128, &desc); Chris@10: } Chris@10: Chris@10: #else /* HAVE_FMA */ Chris@10: Chris@10: /* Generated by: ../../../genfft/gen_notw_c.native -simd -compact -variables 4 -pipeline-latency 8 -sign 1 -n 128 -name n1bv_128 -include n1b.h */ Chris@10: Chris@10: /* Chris@10: * This function contains 1082 FP additions, 330 FP multiplications, Chris@10: * (or, 938 additions, 186 multiplications, 144 fused multiply/add), Chris@10: * 194 stack variables, 31 constants, and 256 memory accesses Chris@10: */ Chris@10: #include "n1b.h" Chris@10: Chris@10: static void n1bv_128(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs) Chris@10: { Chris@10: DVK(KP146730474, +0.146730474455361751658850129646717819706215317); Chris@10: DVK(KP989176509, +0.989176509964780973451673738016243063983689533); Chris@10: DVK(KP595699304, +0.595699304492433343467036528829969889511926338); Chris@10: DVK(KP803207531, +0.803207531480644909806676512963141923879569427); Chris@10: DVK(KP049067674, +0.049067674327418014254954976942682658314745363); Chris@10: DVK(KP998795456, +0.998795456205172392714771604759100694443203615); Chris@10: DVK(KP671558954, +0.671558954847018400625376850427421803228750632); Chris@10: DVK(KP740951125, +0.740951125354959091175616897495162729728955309); Chris@10: DVK(KP514102744, +0.514102744193221726593693838968815772608049120); Chris@10: DVK(KP857728610, +0.857728610000272069902269984284770137042490799); Chris@10: DVK(KP242980179, +0.242980179903263889948274162077471118320990783); Chris@10: DVK(KP970031253, +0.970031253194543992603984207286100251456865962); Chris@10: DVK(KP427555093, +0.427555093430282094320966856888798534304578629); Chris@10: DVK(KP903989293, +0.903989293123443331586200297230537048710132025); Chris@10: DVK(KP336889853, +0.336889853392220050689253212619147570477766780); Chris@10: DVK(KP941544065, +0.941544065183020778412509402599502357185589796); Chris@10: DVK(KP634393284, +0.634393284163645498215171613225493370675687095); Chris@10: DVK(KP773010453, +0.773010453362736960810906609758469800971041293); Chris@10: DVK(KP098017140, +0.098017140329560601994195563888641845861136673); Chris@10: DVK(KP995184726, +0.995184726672196886244836953109479921575474869); Chris@10: DVK(KP290284677, +0.290284677254462367636192375817395274691476278); Chris@10: DVK(KP956940335, +0.956940335732208864935797886980269969482849206); Chris@10: DVK(KP471396736, +0.471396736825997648556387625905254377657460319); Chris@10: DVK(KP881921264, +0.881921264348355029712756863660388349508442621); Chris@10: DVK(KP980785280, +0.980785280403230449126182236134239036973933731); Chris@10: DVK(KP195090322, +0.195090322016128267848284868477022240927691618); Chris@10: DVK(KP555570233, +0.555570233019602224742830813948532874374937191); Chris@10: DVK(KP831469612, +0.831469612302545237078788377617905756738560812); Chris@10: DVK(KP382683432, +0.382683432365089771728459984030398866761344562); Chris@10: DVK(KP923879532, +0.923879532511286756128183189396788286822416626); Chris@10: DVK(KP707106781, +0.707106781186547524400844362104849039284835938); Chris@10: { Chris@10: INT i; Chris@10: const R *xi; Chris@10: R *xo; Chris@10: xi = ii; Chris@10: xo = io; Chris@10: for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(256, is), MAKE_VOLATILE_STRIDE(256, os)) { Chris@10: V T49, T6e, Tev, TgK, TfA, TgL, T4U, T5J, T7R, T9o, Tah, TdG, Tcw, TdB, T84; Chris@10: V T8T, Tfk, Tfo, T1G, T64, Tgs, Th6, T2p, T62, T7t, T9c, Tce, Tdm, T7i, T9e; Chris@10: V Tc8, Tdp, TgF, TgG, T4q, T4V, TeC, Tfx, T4H, T4W, T7X, T86, Tcr, TdH, T7U; Chris@10: V T85, Taw, TdC, Tf3, Tf7, Tr, T5X, Tgl, Th3, T1a, T5V, T7a, T95, TbD, Tdf; Chris@10: V T6Z, T97, Tbx, Tdi, Tgy, Tgz, TgA, TaN, Tdv, TeK, Tfu, T2W, T5M, T35, T5N; Chris@10: V T7F, T8X, TaI, Tdu, T7C, T8W, TgB, TgC, TgD, Tb4, Tdy, TeR, Tfv, T3x, T5P; Chris@10: V T3G, T5Q, T7M, T90, TaZ, Tdx, T7J, T8Z, Tbm, Tdg, TbG, Tdj, Tgo, Th4, Tf0; Chris@10: V Tf8, T76, T98, T7d, T94, T10, T5Y, T1d, T5U, TbX, Tdn, Tch, Tdq, Tgv, Th7; Chris@10: V Tfh, Tfp, T7p, T9f, T7w, T9b, T2f, T65, T2s, T61; Chris@10: { Chris@10: V T47, Ta8, T4O, Ta7, T44, Tcu, T4P, Tct, Taa, Tab, T3P, Tac, T4R, Tad, Tae; Chris@10: V T3W, Taf, T4S; Chris@10: { Chris@10: V T45, T46, T4M, T4N; Chris@10: T45 = LD(&(xi[WS(is, 32)]), ivs, &(xi[0])); Chris@10: T46 = LD(&(xi[WS(is, 96)]), ivs, &(xi[0])); Chris@10: T47 = VSUB(T45, T46); Chris@10: Ta8 = VADD(T45, T46); Chris@10: T4M = LD(&(xi[0]), ivs, &(xi[0])); Chris@10: T4N = LD(&(xi[WS(is, 64)]), ivs, &(xi[0])); Chris@10: T4O = VSUB(T4M, T4N); Chris@10: Ta7 = VADD(T4M, T4N); Chris@10: } Chris@10: { Chris@10: V T3Y, T3Z, T40, T41, T42, T43; Chris@10: T3Y = LD(&(xi[WS(is, 16)]), ivs, &(xi[0])); Chris@10: T3Z = LD(&(xi[WS(is, 80)]), ivs, &(xi[0])); Chris@10: T40 = VSUB(T3Y, T3Z); Chris@10: T41 = LD(&(xi[WS(is, 112)]), ivs, &(xi[0])); Chris@10: T42 = LD(&(xi[WS(is, 48)]), ivs, &(xi[0])); Chris@10: T43 = VSUB(T41, T42); Chris@10: T44 = VMUL(LDK(KP707106781), VSUB(T40, T43)); Chris@10: Tcu = VADD(T41, T42); Chris@10: T4P = VMUL(LDK(KP707106781), VADD(T40, T43)); Chris@10: Tct = VADD(T3Y, T3Z); Chris@10: } Chris@10: { Chris@10: V T3L, T3O, T3S, T3V; Chris@10: { Chris@10: V T3J, T3K, T3M, T3N; Chris@10: T3J = LD(&(xi[WS(is, 8)]), ivs, &(xi[0])); Chris@10: T3K = LD(&(xi[WS(is, 72)]), ivs, &(xi[0])); Chris@10: T3L = VSUB(T3J, T3K); Chris@10: Taa = VADD(T3J, T3K); Chris@10: T3M = LD(&(xi[WS(is, 40)]), ivs, &(xi[0])); Chris@10: T3N = LD(&(xi[WS(is, 104)]), ivs, &(xi[0])); Chris@10: T3O = VSUB(T3M, T3N); Chris@10: Tab = VADD(T3M, T3N); Chris@10: } Chris@10: T3P = VFNMS(LDK(KP382683432), T3O, VMUL(LDK(KP923879532), T3L)); Chris@10: Tac = VSUB(Taa, Tab); Chris@10: T4R = VFMA(LDK(KP382683432), T3L, VMUL(LDK(KP923879532), T3O)); Chris@10: { Chris@10: V T3Q, T3R, T3T, T3U; Chris@10: T3Q = LD(&(xi[WS(is, 120)]), ivs, &(xi[0])); Chris@10: T3R = LD(&(xi[WS(is, 56)]), ivs, &(xi[0])); Chris@10: T3S = VSUB(T3Q, T3R); Chris@10: Tad = VADD(T3Q, T3R); Chris@10: T3T = LD(&(xi[WS(is, 24)]), ivs, &(xi[0])); Chris@10: T3U = LD(&(xi[WS(is, 88)]), ivs, &(xi[0])); Chris@10: T3V = VSUB(T3T, T3U); Chris@10: Tae = VADD(T3T, T3U); Chris@10: } Chris@10: T3W = VFMA(LDK(KP923879532), T3S, VMUL(LDK(KP382683432), T3V)); Chris@10: Taf = VSUB(Tad, Tae); Chris@10: T4S = VFNMS(LDK(KP382683432), T3S, VMUL(LDK(KP923879532), T3V)); Chris@10: } Chris@10: { Chris@10: V T3X, T48, Tet, Teu; Chris@10: T3X = VSUB(T3P, T3W); Chris@10: T48 = VSUB(T44, T47); Chris@10: T49 = VSUB(T3X, T48); Chris@10: T6e = VADD(T48, T3X); Chris@10: Tet = VADD(Ta7, Ta8); Chris@10: Teu = VADD(Tct, Tcu); Chris@10: Tev = VSUB(Tet, Teu); Chris@10: TgK = VADD(Tet, Teu); Chris@10: } Chris@10: { Chris@10: V Tfy, Tfz, T4Q, T4T; Chris@10: Tfy = VADD(Taa, Tab); Chris@10: Tfz = VADD(Tad, Tae); Chris@10: TfA = VSUB(Tfy, Tfz); Chris@10: TgL = VADD(Tfy, Tfz); Chris@10: T4Q = VSUB(T4O, T4P); Chris@10: T4T = VSUB(T4R, T4S); Chris@10: T4U = VSUB(T4Q, T4T); Chris@10: T5J = VADD(T4Q, T4T); Chris@10: } Chris@10: { Chris@10: V T7P, T7Q, Ta9, Tag; Chris@10: T7P = VADD(T4R, T4S); Chris@10: T7Q = VADD(T47, T44); Chris@10: T7R = VSUB(T7P, T7Q); Chris@10: T9o = VADD(T7Q, T7P); Chris@10: Ta9 = VSUB(Ta7, Ta8); Chris@10: Tag = VMUL(LDK(KP707106781), VADD(Tac, Taf)); Chris@10: Tah = VSUB(Ta9, Tag); Chris@10: TdG = VADD(Ta9, Tag); Chris@10: } Chris@10: { Chris@10: V Tcs, Tcv, T82, T83; Chris@10: Tcs = VMUL(LDK(KP707106781), VSUB(Tac, Taf)); Chris@10: Tcv = VSUB(Tct, Tcu); Chris@10: Tcw = VSUB(Tcs, Tcv); Chris@10: TdB = VADD(Tcv, Tcs); Chris@10: T82 = VADD(T4O, T4P); Chris@10: T83 = VADD(T3P, T3W); Chris@10: T84 = VSUB(T82, T83); Chris@10: T8T = VADD(T82, T83); Chris@10: } Chris@10: } Chris@10: { Chris@10: V Tca, Tcb, T1i, Tfm, T2n, Tc5, Tc6, T1p, Tfn, T2k, T1x, Tfi, T2h, Tc0, T1E; Chris@10: V Tfj, T2i, Tc3, T1l, T1o, Tcc, Tcd; Chris@10: { Chris@10: V T1g, T1h, T2l, T2m; Chris@10: T1g = LD(&(xi[WS(is, 127)]), ivs, &(xi[WS(is, 1)])); Chris@10: T1h = LD(&(xi[WS(is, 63)]), ivs, &(xi[WS(is, 1)])); Chris@10: Tca = VADD(T1g, T1h); Chris@10: T2l = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)])); Chris@10: T2m = LD(&(xi[WS(is, 95)]), ivs, &(xi[WS(is, 1)])); Chris@10: Tcb = VADD(T2l, T2m); Chris@10: T1i = VSUB(T1g, T1h); Chris@10: Tfm = VADD(Tca, Tcb); Chris@10: T2n = VSUB(T2l, T2m); Chris@10: } Chris@10: { Chris@10: V T1j, T1k, T1m, T1n; Chris@10: T1j = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)])); Chris@10: T1k = LD(&(xi[WS(is, 79)]), ivs, &(xi[WS(is, 1)])); Chris@10: T1l = VSUB(T1j, T1k); Chris@10: Tc5 = VADD(T1j, T1k); Chris@10: T1m = LD(&(xi[WS(is, 111)]), ivs, &(xi[WS(is, 1)])); Chris@10: T1n = LD(&(xi[WS(is, 47)]), ivs, &(xi[WS(is, 1)])); Chris@10: T1o = VSUB(T1m, T1n); Chris@10: Tc6 = VADD(T1m, T1n); Chris@10: } Chris@10: T1p = VMUL(LDK(KP707106781), VADD(T1l, T1o)); Chris@10: Tfn = VADD(Tc5, Tc6); Chris@10: T2k = VMUL(LDK(KP707106781), VSUB(T1l, T1o)); Chris@10: { Chris@10: V T1t, TbY, T1w, TbZ; Chris@10: { Chris@10: V T1r, T1s, T1u, T1v; Chris@10: T1r = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)])); Chris@10: T1s = LD(&(xi[WS(is, 71)]), ivs, &(xi[WS(is, 1)])); Chris@10: T1t = VSUB(T1r, T1s); Chris@10: TbY = VADD(T1r, T1s); Chris@10: T1u = LD(&(xi[WS(is, 39)]), ivs, &(xi[WS(is, 1)])); Chris@10: T1v = LD(&(xi[WS(is, 103)]), ivs, &(xi[WS(is, 1)])); Chris@10: T1w = VSUB(T1u, T1v); Chris@10: TbZ = VADD(T1u, T1v); Chris@10: } Chris@10: T1x = VFMA(LDK(KP382683432), T1t, VMUL(LDK(KP923879532), T1w)); Chris@10: Tfi = VADD(TbY, TbZ); Chris@10: T2h = VFNMS(LDK(KP382683432), T1w, VMUL(LDK(KP923879532), T1t)); Chris@10: Tc0 = VSUB(TbY, TbZ); Chris@10: } Chris@10: { Chris@10: V T1A, Tc2, T1D, Tc1; Chris@10: { Chris@10: V T1y, T1z, T1B, T1C; Chris@10: T1y = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)])); Chris@10: T1z = LD(&(xi[WS(is, 87)]), ivs, &(xi[WS(is, 1)])); Chris@10: T1A = VSUB(T1y, T1z); Chris@10: Tc2 = VADD(T1y, T1z); Chris@10: T1B = LD(&(xi[WS(is, 119)]), ivs, &(xi[WS(is, 1)])); Chris@10: T1C = LD(&(xi[WS(is, 55)]), ivs, &(xi[WS(is, 1)])); Chris@10: T1D = VSUB(T1B, T1C); Chris@10: Tc1 = VADD(T1B, T1C); Chris@10: } Chris@10: T1E = VFNMS(LDK(KP382683432), T1D, VMUL(LDK(KP923879532), T1A)); Chris@10: Tfj = VADD(Tc1, Tc2); Chris@10: T2i = VFMA(LDK(KP923879532), T1D, VMUL(LDK(KP382683432), T1A)); Chris@10: Tc3 = VSUB(Tc1, Tc2); Chris@10: } Chris@10: Tfk = VSUB(Tfi, Tfj); Chris@10: Tfo = VSUB(Tfm, Tfn); Chris@10: { Chris@10: V T1q, T1F, Tgq, Tgr; Chris@10: T1q = VSUB(T1i, T1p); Chris@10: T1F = VSUB(T1x, T1E); Chris@10: T1G = VSUB(T1q, T1F); Chris@10: T64 = VADD(T1q, T1F); Chris@10: Tgq = VADD(Tfm, Tfn); Chris@10: Tgr = VADD(Tfi, Tfj); Chris@10: Tgs = VSUB(Tgq, Tgr); Chris@10: Th6 = VADD(Tgq, Tgr); Chris@10: } Chris@10: { Chris@10: V T2j, T2o, T7r, T7s; Chris@10: T2j = VSUB(T2h, T2i); Chris@10: T2o = VSUB(T2k, T2n); Chris@10: T2p = VSUB(T2j, T2o); Chris@10: T62 = VADD(T2o, T2j); Chris@10: T7r = VADD(T1x, T1E); Chris@10: T7s = VADD(T2n, T2k); Chris@10: T7t = VSUB(T7r, T7s); Chris@10: T9c = VADD(T7s, T7r); Chris@10: } Chris@10: Tcc = VSUB(Tca, Tcb); Chris@10: Tcd = VMUL(LDK(KP707106781), VADD(Tc0, Tc3)); Chris@10: Tce = VSUB(Tcc, Tcd); Chris@10: Tdm = VADD(Tcc, Tcd); Chris@10: { Chris@10: V T7g, T7h, Tc4, Tc7; Chris@10: T7g = VADD(T1i, T1p); Chris@10: T7h = VADD(T2h, T2i); Chris@10: T7i = VSUB(T7g, T7h); Chris@10: T9e = VADD(T7g, T7h); Chris@10: Tc4 = VMUL(LDK(KP707106781), VSUB(Tc0, Tc3)); Chris@10: Tc7 = VSUB(Tc5, Tc6); Chris@10: Tc8 = VSUB(Tc4, Tc7); Chris@10: Tdp = VADD(Tc7, Tc4); Chris@10: } Chris@10: } Chris@10: { Chris@10: V T4c, Tew, T4o, Tak, T4A, Tez, T4E, Tau, T4j, Tex, T4l, Tan, T4x, TeA, T4F; Chris@10: V Tar, Tcp, Tcq; Chris@10: { Chris@10: V T4a, T4b, Tai, T4m, T4n, Taj; Chris@10: T4a = LD(&(xi[WS(is, 4)]), ivs, &(xi[0])); Chris@10: T4b = LD(&(xi[WS(is, 68)]), ivs, &(xi[0])); Chris@10: Tai = VADD(T4a, T4b); Chris@10: T4m = LD(&(xi[WS(is, 36)]), ivs, &(xi[0])); Chris@10: T4n = LD(&(xi[WS(is, 100)]), ivs, &(xi[0])); Chris@10: Taj = VADD(T4m, T4n); Chris@10: T4c = VSUB(T4a, T4b); Chris@10: Tew = VADD(Tai, Taj); Chris@10: T4o = VSUB(T4m, T4n); Chris@10: Tak = VSUB(Tai, Taj); Chris@10: } Chris@10: { Chris@10: V T4y, T4z, Tat, T4C, T4D, Tas; Chris@10: T4y = LD(&(xi[WS(is, 28)]), ivs, &(xi[0])); Chris@10: T4z = LD(&(xi[WS(is, 92)]), ivs, &(xi[0])); Chris@10: Tat = VADD(T4y, T4z); Chris@10: T4C = LD(&(xi[WS(is, 124)]), ivs, &(xi[0])); Chris@10: T4D = LD(&(xi[WS(is, 60)]), ivs, &(xi[0])); Chris@10: Tas = VADD(T4C, T4D); Chris@10: T4A = VSUB(T4y, T4z); Chris@10: Tez = VADD(Tas, Tat); Chris@10: T4E = VSUB(T4C, T4D); Chris@10: Tau = VSUB(Tas, Tat); Chris@10: } Chris@10: { Chris@10: V T4f, Tal, T4i, Tam; Chris@10: { Chris@10: V T4d, T4e, T4g, T4h; Chris@10: T4d = LD(&(xi[WS(is, 20)]), ivs, &(xi[0])); Chris@10: T4e = LD(&(xi[WS(is, 84)]), ivs, &(xi[0])); Chris@10: T4f = VSUB(T4d, T4e); Chris@10: Tal = VADD(T4d, T4e); Chris@10: T4g = LD(&(xi[WS(is, 116)]), ivs, &(xi[0])); Chris@10: T4h = LD(&(xi[WS(is, 52)]), ivs, &(xi[0])); Chris@10: T4i = VSUB(T4g, T4h); Chris@10: Tam = VADD(T4g, T4h); Chris@10: } Chris@10: T4j = VMUL(LDK(KP707106781), VADD(T4f, T4i)); Chris@10: Tex = VADD(Tal, Tam); Chris@10: T4l = VMUL(LDK(KP707106781), VSUB(T4f, T4i)); Chris@10: Tan = VSUB(Tal, Tam); Chris@10: } Chris@10: { Chris@10: V T4t, Tap, T4w, Taq; Chris@10: { Chris@10: V T4r, T4s, T4u, T4v; Chris@10: T4r = LD(&(xi[WS(is, 12)]), ivs, &(xi[0])); Chris@10: T4s = LD(&(xi[WS(is, 76)]), ivs, &(xi[0])); Chris@10: T4t = VSUB(T4r, T4s); Chris@10: Tap = VADD(T4r, T4s); Chris@10: T4u = LD(&(xi[WS(is, 108)]), ivs, &(xi[0])); Chris@10: T4v = LD(&(xi[WS(is, 44)]), ivs, &(xi[0])); Chris@10: T4w = VSUB(T4u, T4v); Chris@10: Taq = VADD(T4u, T4v); Chris@10: } Chris@10: T4x = VMUL(LDK(KP707106781), VSUB(T4t, T4w)); Chris@10: TeA = VADD(Tap, Taq); Chris@10: T4F = VMUL(LDK(KP707106781), VADD(T4t, T4w)); Chris@10: Tar = VSUB(Tap, Taq); Chris@10: } Chris@10: TgF = VADD(Tew, Tex); Chris@10: TgG = VADD(Tez, TeA); Chris@10: { Chris@10: V T4k, T4p, Tey, TeB; Chris@10: T4k = VSUB(T4c, T4j); Chris@10: T4p = VSUB(T4l, T4o); Chris@10: T4q = VFNMS(LDK(KP555570233), T4p, VMUL(LDK(KP831469612), T4k)); Chris@10: T4V = VFMA(LDK(KP831469612), T4p, VMUL(LDK(KP555570233), T4k)); Chris@10: Tey = VSUB(Tew, Tex); Chris@10: TeB = VSUB(Tez, TeA); Chris@10: TeC = VMUL(LDK(KP707106781), VADD(Tey, TeB)); Chris@10: Tfx = VMUL(LDK(KP707106781), VSUB(Tey, TeB)); Chris@10: } Chris@10: { Chris@10: V T4B, T4G, T7V, T7W; Chris@10: T4B = VSUB(T4x, T4A); Chris@10: T4G = VSUB(T4E, T4F); Chris@10: T4H = VFMA(LDK(KP555570233), T4B, VMUL(LDK(KP831469612), T4G)); Chris@10: T4W = VFNMS(LDK(KP555570233), T4G, VMUL(LDK(KP831469612), T4B)); Chris@10: T7V = VADD(T4A, T4x); Chris@10: T7W = VADD(T4E, T4F); Chris@10: T7X = VFMA(LDK(KP195090322), T7V, VMUL(LDK(KP980785280), T7W)); Chris@10: T86 = VFNMS(LDK(KP195090322), T7W, VMUL(LDK(KP980785280), T7V)); Chris@10: } Chris@10: Tcp = VFNMS(LDK(KP382683432), Tan, VMUL(LDK(KP923879532), Tak)); Chris@10: Tcq = VFMA(LDK(KP923879532), Tau, VMUL(LDK(KP382683432), Tar)); Chris@10: Tcr = VSUB(Tcp, Tcq); Chris@10: TdH = VADD(Tcp, Tcq); Chris@10: { Chris@10: V T7S, T7T, Tao, Tav; Chris@10: T7S = VADD(T4c, T4j); Chris@10: T7T = VADD(T4o, T4l); Chris@10: T7U = VFNMS(LDK(KP195090322), T7T, VMUL(LDK(KP980785280), T7S)); Chris@10: T85 = VFMA(LDK(KP980785280), T7T, VMUL(LDK(KP195090322), T7S)); Chris@10: Tao = VFMA(LDK(KP382683432), Tak, VMUL(LDK(KP923879532), Tan)); Chris@10: Tav = VFNMS(LDK(KP382683432), Tau, VMUL(LDK(KP923879532), Tar)); Chris@10: Taw = VSUB(Tao, Tav); Chris@10: TdC = VADD(Tao, Tav); Chris@10: } Chris@10: } Chris@10: { Chris@10: V Tbz, TbA, T3, Tf5, T18, Tbu, Tbv, Ta, Tf6, T15, Ti, Tf1, T12, Tbp, Tp; Chris@10: V Tf2, T13, Tbs, T6, T9, TbB, TbC; Chris@10: { Chris@10: V T1, T2, T16, T17; Chris@10: T1 = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)])); Chris@10: T2 = LD(&(xi[WS(is, 65)]), ivs, &(xi[WS(is, 1)])); Chris@10: Tbz = VADD(T1, T2); Chris@10: T16 = LD(&(xi[WS(is, 33)]), ivs, &(xi[WS(is, 1)])); Chris@10: T17 = LD(&(xi[WS(is, 97)]), ivs, &(xi[WS(is, 1)])); Chris@10: TbA = VADD(T16, T17); Chris@10: T3 = VSUB(T1, T2); Chris@10: Tf5 = VADD(Tbz, TbA); Chris@10: T18 = VSUB(T16, T17); Chris@10: } Chris@10: { Chris@10: V T4, T5, T7, T8; Chris@10: T4 = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)])); Chris@10: T5 = LD(&(xi[WS(is, 81)]), ivs, &(xi[WS(is, 1)])); Chris@10: T6 = VSUB(T4, T5); Chris@10: Tbu = VADD(T4, T5); Chris@10: T7 = LD(&(xi[WS(is, 113)]), ivs, &(xi[WS(is, 1)])); Chris@10: T8 = LD(&(xi[WS(is, 49)]), ivs, &(xi[WS(is, 1)])); Chris@10: T9 = VSUB(T7, T8); Chris@10: Tbv = VADD(T7, T8); Chris@10: } Chris@10: Ta = VMUL(LDK(KP707106781), VADD(T6, T9)); Chris@10: Tf6 = VADD(Tbu, Tbv); Chris@10: T15 = VMUL(LDK(KP707106781), VSUB(T6, T9)); Chris@10: { Chris@10: V Te, Tbn, Th, Tbo; Chris@10: { Chris@10: V Tc, Td, Tf, Tg; Chris@10: Tc = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)])); Chris@10: Td = LD(&(xi[WS(is, 73)]), ivs, &(xi[WS(is, 1)])); Chris@10: Te = VSUB(Tc, Td); Chris@10: Tbn = VADD(Tc, Td); Chris@10: Tf = LD(&(xi[WS(is, 41)]), ivs, &(xi[WS(is, 1)])); Chris@10: Tg = LD(&(xi[WS(is, 105)]), ivs, &(xi[WS(is, 1)])); Chris@10: Th = VSUB(Tf, Tg); Chris@10: Tbo = VADD(Tf, Tg); Chris@10: } Chris@10: Ti = VFMA(LDK(KP382683432), Te, VMUL(LDK(KP923879532), Th)); Chris@10: Tf1 = VADD(Tbn, Tbo); Chris@10: T12 = VFNMS(LDK(KP382683432), Th, VMUL(LDK(KP923879532), Te)); Chris@10: Tbp = VSUB(Tbn, Tbo); Chris@10: } Chris@10: { Chris@10: V Tl, Tbr, To, Tbq; Chris@10: { Chris@10: V Tj, Tk, Tm, Tn; Chris@10: Tj = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)])); Chris@10: Tk = LD(&(xi[WS(is, 89)]), ivs, &(xi[WS(is, 1)])); Chris@10: Tl = VSUB(Tj, Tk); Chris@10: Tbr = VADD(Tj, Tk); Chris@10: Tm = LD(&(xi[WS(is, 121)]), ivs, &(xi[WS(is, 1)])); Chris@10: Tn = LD(&(xi[WS(is, 57)]), ivs, &(xi[WS(is, 1)])); Chris@10: To = VSUB(Tm, Tn); Chris@10: Tbq = VADD(Tm, Tn); Chris@10: } Chris@10: Tp = VFNMS(LDK(KP382683432), To, VMUL(LDK(KP923879532), Tl)); Chris@10: Tf2 = VADD(Tbq, Tbr); Chris@10: T13 = VFMA(LDK(KP923879532), To, VMUL(LDK(KP382683432), Tl)); Chris@10: Tbs = VSUB(Tbq, Tbr); Chris@10: } Chris@10: Tf3 = VSUB(Tf1, Tf2); Chris@10: Tf7 = VSUB(Tf5, Tf6); Chris@10: { Chris@10: V Tb, Tq, Tgj, Tgk; Chris@10: Tb = VSUB(T3, Ta); Chris@10: Tq = VSUB(Ti, Tp); Chris@10: Tr = VSUB(Tb, Tq); Chris@10: T5X = VADD(Tb, Tq); Chris@10: Tgj = VADD(Tf5, Tf6); Chris@10: Tgk = VADD(Tf1, Tf2); Chris@10: Tgl = VSUB(Tgj, Tgk); Chris@10: Th3 = VADD(Tgj, Tgk); Chris@10: } Chris@10: { Chris@10: V T14, T19, T78, T79; Chris@10: T14 = VSUB(T12, T13); Chris@10: T19 = VSUB(T15, T18); Chris@10: T1a = VSUB(T14, T19); Chris@10: T5V = VADD(T19, T14); Chris@10: T78 = VADD(Ti, Tp); Chris@10: T79 = VADD(T18, T15); Chris@10: T7a = VSUB(T78, T79); Chris@10: T95 = VADD(T79, T78); Chris@10: } Chris@10: TbB = VSUB(Tbz, TbA); Chris@10: TbC = VMUL(LDK(KP707106781), VADD(Tbp, Tbs)); Chris@10: TbD = VSUB(TbB, TbC); Chris@10: Tdf = VADD(TbB, TbC); Chris@10: { Chris@10: V T6X, T6Y, Tbt, Tbw; Chris@10: T6X = VADD(T3, Ta); Chris@10: T6Y = VADD(T12, T13); Chris@10: T6Z = VSUB(T6X, T6Y); Chris@10: T97 = VADD(T6X, T6Y); Chris@10: Tbt = VMUL(LDK(KP707106781), VSUB(Tbp, Tbs)); Chris@10: Tbw = VSUB(Tbu, Tbv); Chris@10: Tbx = VSUB(Tbt, Tbw); Chris@10: Tdi = VADD(Tbw, Tbt); Chris@10: } Chris@10: } Chris@10: { Chris@10: V TaK, TaJ, T2U, TeE, T2Z, TaF, TaG, T2R, TeF, T30, T2C, TeH, T32, TaA, T2J; Chris@10: V TeI, T33, TaD, T2N, T2Q, TaL, TaM; Chris@10: { Chris@10: V T2S, T2T, T2X, T2Y; Chris@10: T2S = LD(&(xi[WS(is, 34)]), ivs, &(xi[0])); Chris@10: T2T = LD(&(xi[WS(is, 98)]), ivs, &(xi[0])); Chris@10: TaK = VADD(T2S, T2T); Chris@10: T2X = LD(&(xi[WS(is, 2)]), ivs, &(xi[0])); Chris@10: T2Y = LD(&(xi[WS(is, 66)]), ivs, &(xi[0])); Chris@10: TaJ = VADD(T2X, T2Y); Chris@10: T2U = VSUB(T2S, T2T); Chris@10: TeE = VADD(TaJ, TaK); Chris@10: T2Z = VSUB(T2X, T2Y); Chris@10: } Chris@10: { Chris@10: V T2L, T2M, T2O, T2P; Chris@10: T2L = LD(&(xi[WS(is, 18)]), ivs, &(xi[0])); Chris@10: T2M = LD(&(xi[WS(is, 82)]), ivs, &(xi[0])); Chris@10: T2N = VSUB(T2L, T2M); Chris@10: TaF = VADD(T2L, T2M); Chris@10: T2O = LD(&(xi[WS(is, 114)]), ivs, &(xi[0])); Chris@10: T2P = LD(&(xi[WS(is, 50)]), ivs, &(xi[0])); Chris@10: T2Q = VSUB(T2O, T2P); Chris@10: TaG = VADD(T2O, T2P); Chris@10: } Chris@10: T2R = VMUL(LDK(KP707106781), VSUB(T2N, T2Q)); Chris@10: TeF = VADD(TaF, TaG); Chris@10: T30 = VMUL(LDK(KP707106781), VADD(T2N, T2Q)); Chris@10: { Chris@10: V T2y, Tay, T2B, Taz; Chris@10: { Chris@10: V T2w, T2x, T2z, T2A; Chris@10: T2w = LD(&(xi[WS(is, 10)]), ivs, &(xi[0])); Chris@10: T2x = LD(&(xi[WS(is, 74)]), ivs, &(xi[0])); Chris@10: T2y = VSUB(T2w, T2x); Chris@10: Tay = VADD(T2w, T2x); Chris@10: T2z = LD(&(xi[WS(is, 42)]), ivs, &(xi[0])); Chris@10: T2A = LD(&(xi[WS(is, 106)]), ivs, &(xi[0])); Chris@10: T2B = VSUB(T2z, T2A); Chris@10: Taz = VADD(T2z, T2A); Chris@10: } Chris@10: T2C = VFNMS(LDK(KP382683432), T2B, VMUL(LDK(KP923879532), T2y)); Chris@10: TeH = VADD(Tay, Taz); Chris@10: T32 = VFMA(LDK(KP382683432), T2y, VMUL(LDK(KP923879532), T2B)); Chris@10: TaA = VSUB(Tay, Taz); Chris@10: } Chris@10: { Chris@10: V T2F, TaB, T2I, TaC; Chris@10: { Chris@10: V T2D, T2E, T2G, T2H; Chris@10: T2D = LD(&(xi[WS(is, 122)]), ivs, &(xi[0])); Chris@10: T2E = LD(&(xi[WS(is, 58)]), ivs, &(xi[0])); Chris@10: T2F = VSUB(T2D, T2E); Chris@10: TaB = VADD(T2D, T2E); Chris@10: T2G = LD(&(xi[WS(is, 26)]), ivs, &(xi[0])); Chris@10: T2H = LD(&(xi[WS(is, 90)]), ivs, &(xi[0])); Chris@10: T2I = VSUB(T2G, T2H); Chris@10: TaC = VADD(T2G, T2H); Chris@10: } Chris@10: T2J = VFMA(LDK(KP923879532), T2F, VMUL(LDK(KP382683432), T2I)); Chris@10: TeI = VADD(TaB, TaC); Chris@10: T33 = VFNMS(LDK(KP382683432), T2F, VMUL(LDK(KP923879532), T2I)); Chris@10: TaD = VSUB(TaB, TaC); Chris@10: } Chris@10: Tgy = VADD(TeE, TeF); Chris@10: Tgz = VADD(TeH, TeI); Chris@10: TgA = VSUB(Tgy, Tgz); Chris@10: TaL = VSUB(TaJ, TaK); Chris@10: TaM = VMUL(LDK(KP707106781), VADD(TaA, TaD)); Chris@10: TaN = VSUB(TaL, TaM); Chris@10: Tdv = VADD(TaL, TaM); Chris@10: { Chris@10: V TeG, TeJ, T2K, T2V; Chris@10: TeG = VSUB(TeE, TeF); Chris@10: TeJ = VSUB(TeH, TeI); Chris@10: TeK = VFMA(LDK(KP382683432), TeG, VMUL(LDK(KP923879532), TeJ)); Chris@10: Tfu = VFNMS(LDK(KP382683432), TeJ, VMUL(LDK(KP923879532), TeG)); Chris@10: T2K = VSUB(T2C, T2J); Chris@10: T2V = VSUB(T2R, T2U); Chris@10: T2W = VSUB(T2K, T2V); Chris@10: T5M = VADD(T2V, T2K); Chris@10: } Chris@10: { Chris@10: V T31, T34, T7D, T7E; Chris@10: T31 = VSUB(T2Z, T30); Chris@10: T34 = VSUB(T32, T33); Chris@10: T35 = VSUB(T31, T34); Chris@10: T5N = VADD(T31, T34); Chris@10: T7D = VADD(T32, T33); Chris@10: T7E = VADD(T2U, T2R); Chris@10: T7F = VSUB(T7D, T7E); Chris@10: T8X = VADD(T7E, T7D); Chris@10: } Chris@10: { Chris@10: V TaE, TaH, T7A, T7B; Chris@10: TaE = VMUL(LDK(KP707106781), VSUB(TaA, TaD)); Chris@10: TaH = VSUB(TaF, TaG); Chris@10: TaI = VSUB(TaE, TaH); Chris@10: Tdu = VADD(TaH, TaE); Chris@10: T7A = VADD(T2Z, T30); Chris@10: T7B = VADD(T2C, T2J); Chris@10: T7C = VSUB(T7A, T7B); Chris@10: T8W = VADD(T7A, T7B); Chris@10: } Chris@10: } Chris@10: { Chris@10: V Tb1, Tb0, T3v, TeO, T3A, TaW, TaX, T3s, TeP, T3B, T3d, TeL, T3D, TaR, T3k; Chris@10: V TeM, T3E, TaU, T3o, T3r, Tb2, Tb3; Chris@10: { Chris@10: V T3t, T3u, T3y, T3z; Chris@10: T3t = LD(&(xi[WS(is, 30)]), ivs, &(xi[0])); Chris@10: T3u = LD(&(xi[WS(is, 94)]), ivs, &(xi[0])); Chris@10: Tb1 = VADD(T3t, T3u); Chris@10: T3y = LD(&(xi[WS(is, 126)]), ivs, &(xi[0])); Chris@10: T3z = LD(&(xi[WS(is, 62)]), ivs, &(xi[0])); Chris@10: Tb0 = VADD(T3y, T3z); Chris@10: T3v = VSUB(T3t, T3u); Chris@10: TeO = VADD(Tb0, Tb1); Chris@10: T3A = VSUB(T3y, T3z); Chris@10: } Chris@10: { Chris@10: V T3m, T3n, T3p, T3q; Chris@10: T3m = LD(&(xi[WS(is, 14)]), ivs, &(xi[0])); Chris@10: T3n = LD(&(xi[WS(is, 78)]), ivs, &(xi[0])); Chris@10: T3o = VSUB(T3m, T3n); Chris@10: TaW = VADD(T3m, T3n); Chris@10: T3p = LD(&(xi[WS(is, 110)]), ivs, &(xi[0])); Chris@10: T3q = LD(&(xi[WS(is, 46)]), ivs, &(xi[0])); Chris@10: T3r = VSUB(T3p, T3q); Chris@10: TaX = VADD(T3p, T3q); Chris@10: } Chris@10: T3s = VMUL(LDK(KP707106781), VSUB(T3o, T3r)); Chris@10: TeP = VADD(TaW, TaX); Chris@10: T3B = VMUL(LDK(KP707106781), VADD(T3o, T3r)); Chris@10: { Chris@10: V T39, TaP, T3c, TaQ; Chris@10: { Chris@10: V T37, T38, T3a, T3b; Chris@10: T37 = LD(&(xi[WS(is, 6)]), ivs, &(xi[0])); Chris@10: T38 = LD(&(xi[WS(is, 70)]), ivs, &(xi[0])); Chris@10: T39 = VSUB(T37, T38); Chris@10: TaP = VADD(T37, T38); Chris@10: T3a = LD(&(xi[WS(is, 38)]), ivs, &(xi[0])); Chris@10: T3b = LD(&(xi[WS(is, 102)]), ivs, &(xi[0])); Chris@10: T3c = VSUB(T3a, T3b); Chris@10: TaQ = VADD(T3a, T3b); Chris@10: } Chris@10: T3d = VFNMS(LDK(KP382683432), T3c, VMUL(LDK(KP923879532), T39)); Chris@10: TeL = VADD(TaP, TaQ); Chris@10: T3D = VFMA(LDK(KP382683432), T39, VMUL(LDK(KP923879532), T3c)); Chris@10: TaR = VSUB(TaP, TaQ); Chris@10: } Chris@10: { Chris@10: V T3g, TaS, T3j, TaT; Chris@10: { Chris@10: V T3e, T3f, T3h, T3i; Chris@10: T3e = LD(&(xi[WS(is, 118)]), ivs, &(xi[0])); Chris@10: T3f = LD(&(xi[WS(is, 54)]), ivs, &(xi[0])); Chris@10: T3g = VSUB(T3e, T3f); Chris@10: TaS = VADD(T3e, T3f); Chris@10: T3h = LD(&(xi[WS(is, 22)]), ivs, &(xi[0])); Chris@10: T3i = LD(&(xi[WS(is, 86)]), ivs, &(xi[0])); Chris@10: T3j = VSUB(T3h, T3i); Chris@10: TaT = VADD(T3h, T3i); Chris@10: } Chris@10: T3k = VFMA(LDK(KP923879532), T3g, VMUL(LDK(KP382683432), T3j)); Chris@10: TeM = VADD(TaS, TaT); Chris@10: T3E = VFNMS(LDK(KP382683432), T3g, VMUL(LDK(KP923879532), T3j)); Chris@10: TaU = VSUB(TaS, TaT); Chris@10: } Chris@10: TgB = VADD(TeO, TeP); Chris@10: TgC = VADD(TeL, TeM); Chris@10: TgD = VSUB(TgB, TgC); Chris@10: Tb2 = VSUB(Tb0, Tb1); Chris@10: Tb3 = VMUL(LDK(KP707106781), VADD(TaR, TaU)); Chris@10: Tb4 = VSUB(Tb2, Tb3); Chris@10: Tdy = VADD(Tb2, Tb3); Chris@10: { Chris@10: V TeN, TeQ, T3l, T3w; Chris@10: TeN = VSUB(TeL, TeM); Chris@10: TeQ = VSUB(TeO, TeP); Chris@10: TeR = VFNMS(LDK(KP382683432), TeQ, VMUL(LDK(KP923879532), TeN)); Chris@10: Tfv = VFMA(LDK(KP923879532), TeQ, VMUL(LDK(KP382683432), TeN)); Chris@10: T3l = VSUB(T3d, T3k); Chris@10: T3w = VSUB(T3s, T3v); Chris@10: T3x = VSUB(T3l, T3w); Chris@10: T5P = VADD(T3w, T3l); Chris@10: } Chris@10: { Chris@10: V T3C, T3F, T7K, T7L; Chris@10: T3C = VSUB(T3A, T3B); Chris@10: T3F = VSUB(T3D, T3E); Chris@10: T3G = VSUB(T3C, T3F); Chris@10: T5Q = VADD(T3C, T3F); Chris@10: T7K = VADD(T3A, T3B); Chris@10: T7L = VADD(T3d, T3k); Chris@10: T7M = VSUB(T7K, T7L); Chris@10: T90 = VADD(T7K, T7L); Chris@10: } Chris@10: { Chris@10: V TaV, TaY, T7H, T7I; Chris@10: TaV = VMUL(LDK(KP707106781), VSUB(TaR, TaU)); Chris@10: TaY = VSUB(TaW, TaX); Chris@10: TaZ = VSUB(TaV, TaY); Chris@10: Tdx = VADD(TaY, TaV); Chris@10: T7H = VADD(T3D, T3E); Chris@10: T7I = VADD(T3v, T3s); Chris@10: T7J = VSUB(T7H, T7I); Chris@10: T8Z = VADD(T7I, T7H); Chris@10: } Chris@10: } Chris@10: { Chris@10: V TB, TeU, TF, Tba, TS, TeX, TW, Tbh, Ty, TeV, TG, Tbd, TP, TeY, TX; Chris@10: V Tbk; Chris@10: { Chris@10: V Tz, TA, Tb9, TD, TE, Tb8; Chris@10: Tz = LD(&(xi[WS(is, 37)]), ivs, &(xi[WS(is, 1)])); Chris@10: TA = LD(&(xi[WS(is, 101)]), ivs, &(xi[WS(is, 1)])); Chris@10: Tb9 = VADD(Tz, TA); Chris@10: TD = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)])); Chris@10: TE = LD(&(xi[WS(is, 69)]), ivs, &(xi[WS(is, 1)])); Chris@10: Tb8 = VADD(TD, TE); Chris@10: TB = VSUB(Tz, TA); Chris@10: TeU = VADD(Tb8, Tb9); Chris@10: TF = VSUB(TD, TE); Chris@10: Tba = VSUB(Tb8, Tb9); Chris@10: } Chris@10: { Chris@10: V TQ, TR, Tbg, TU, TV, Tbf; Chris@10: TQ = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)])); Chris@10: TR = LD(&(xi[WS(is, 93)]), ivs, &(xi[WS(is, 1)])); Chris@10: Tbg = VADD(TQ, TR); Chris@10: TU = LD(&(xi[WS(is, 125)]), ivs, &(xi[WS(is, 1)])); Chris@10: TV = LD(&(xi[WS(is, 61)]), ivs, &(xi[WS(is, 1)])); Chris@10: Tbf = VADD(TU, TV); Chris@10: TS = VSUB(TQ, TR); Chris@10: TeX = VADD(Tbf, Tbg); Chris@10: TW = VSUB(TU, TV); Chris@10: Tbh = VSUB(Tbf, Tbg); Chris@10: } Chris@10: { Chris@10: V Tu, Tbb, Tx, Tbc; Chris@10: { Chris@10: V Ts, Tt, Tv, Tw; Chris@10: Ts = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)])); Chris@10: Tt = LD(&(xi[WS(is, 85)]), ivs, &(xi[WS(is, 1)])); Chris@10: Tu = VSUB(Ts, Tt); Chris@10: Tbb = VADD(Ts, Tt); Chris@10: Tv = LD(&(xi[WS(is, 117)]), ivs, &(xi[WS(is, 1)])); Chris@10: Tw = LD(&(xi[WS(is, 53)]), ivs, &(xi[WS(is, 1)])); Chris@10: Tx = VSUB(Tv, Tw); Chris@10: Tbc = VADD(Tv, Tw); Chris@10: } Chris@10: Ty = VMUL(LDK(KP707106781), VSUB(Tu, Tx)); Chris@10: TeV = VADD(Tbb, Tbc); Chris@10: TG = VMUL(LDK(KP707106781), VADD(Tu, Tx)); Chris@10: Tbd = VSUB(Tbb, Tbc); Chris@10: } Chris@10: { Chris@10: V TL, Tbi, TO, Tbj; Chris@10: { Chris@10: V TJ, TK, TM, TN; Chris@10: TJ = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)])); Chris@10: TK = LD(&(xi[WS(is, 77)]), ivs, &(xi[WS(is, 1)])); Chris@10: TL = VSUB(TJ, TK); Chris@10: Tbi = VADD(TJ, TK); Chris@10: TM = LD(&(xi[WS(is, 109)]), ivs, &(xi[WS(is, 1)])); Chris@10: TN = LD(&(xi[WS(is, 45)]), ivs, &(xi[WS(is, 1)])); Chris@10: TO = VSUB(TM, TN); Chris@10: Tbj = VADD(TM, TN); Chris@10: } Chris@10: TP = VMUL(LDK(KP707106781), VSUB(TL, TO)); Chris@10: TeY = VADD(Tbi, Tbj); Chris@10: TX = VMUL(LDK(KP707106781), VADD(TL, TO)); Chris@10: Tbk = VSUB(Tbi, Tbj); Chris@10: } Chris@10: { Chris@10: V Tbe, Tbl, TeW, TeZ; Chris@10: Tbe = VFNMS(LDK(KP382683432), Tbd, VMUL(LDK(KP923879532), Tba)); Chris@10: Tbl = VFMA(LDK(KP923879532), Tbh, VMUL(LDK(KP382683432), Tbk)); Chris@10: Tbm = VSUB(Tbe, Tbl); Chris@10: Tdg = VADD(Tbe, Tbl); Chris@10: { Chris@10: V TbE, TbF, Tgm, Tgn; Chris@10: TbE = VFMA(LDK(KP382683432), Tba, VMUL(LDK(KP923879532), Tbd)); Chris@10: TbF = VFNMS(LDK(KP382683432), Tbh, VMUL(LDK(KP923879532), Tbk)); Chris@10: TbG = VSUB(TbE, TbF); Chris@10: Tdj = VADD(TbE, TbF); Chris@10: Tgm = VADD(TeU, TeV); Chris@10: Tgn = VADD(TeX, TeY); Chris@10: Tgo = VSUB(Tgm, Tgn); Chris@10: Th4 = VADD(Tgm, Tgn); Chris@10: } Chris@10: TeW = VSUB(TeU, TeV); Chris@10: TeZ = VSUB(TeX, TeY); Chris@10: Tf0 = VMUL(LDK(KP707106781), VSUB(TeW, TeZ)); Chris@10: Tf8 = VMUL(LDK(KP707106781), VADD(TeW, TeZ)); Chris@10: { Chris@10: V T72, T7b, T75, T7c; Chris@10: { Chris@10: V T70, T71, T73, T74; Chris@10: T70 = VADD(TB, Ty); Chris@10: T71 = VADD(TF, TG); Chris@10: T72 = VFMA(LDK(KP980785280), T70, VMUL(LDK(KP195090322), T71)); Chris@10: T7b = VFNMS(LDK(KP195090322), T70, VMUL(LDK(KP980785280), T71)); Chris@10: T73 = VADD(TS, TP); Chris@10: T74 = VADD(TW, TX); Chris@10: T75 = VFNMS(LDK(KP195090322), T74, VMUL(LDK(KP980785280), T73)); Chris@10: T7c = VFMA(LDK(KP195090322), T73, VMUL(LDK(KP980785280), T74)); Chris@10: } Chris@10: T76 = VSUB(T72, T75); Chris@10: T98 = VADD(T7b, T7c); Chris@10: T7d = VSUB(T7b, T7c); Chris@10: T94 = VADD(T72, T75); Chris@10: } Chris@10: { Chris@10: V TI, T1b, TZ, T1c; Chris@10: { Chris@10: V TC, TH, TT, TY; Chris@10: TC = VSUB(Ty, TB); Chris@10: TH = VSUB(TF, TG); Chris@10: TI = VFMA(LDK(KP831469612), TC, VMUL(LDK(KP555570233), TH)); Chris@10: T1b = VFNMS(LDK(KP555570233), TC, VMUL(LDK(KP831469612), TH)); Chris@10: TT = VSUB(TP, TS); Chris@10: TY = VSUB(TW, TX); Chris@10: TZ = VFNMS(LDK(KP555570233), TY, VMUL(LDK(KP831469612), TT)); Chris@10: T1c = VFMA(LDK(KP555570233), TT, VMUL(LDK(KP831469612), TY)); Chris@10: } Chris@10: T10 = VSUB(TI, TZ); Chris@10: T5Y = VADD(T1b, T1c); Chris@10: T1d = VSUB(T1b, T1c); Chris@10: T5U = VADD(TI, TZ); Chris@10: } Chris@10: } Chris@10: } Chris@10: { Chris@10: V T1Q, Tfb, T1U, TbL, T27, Tfe, T2b, TbS, T1N, Tfc, T1V, TbO, T24, Tff, T2c; Chris@10: V TbV; Chris@10: { Chris@10: V T1O, T1P, TbK, T1S, T1T, TbJ; Chris@10: T1O = LD(&(xi[WS(is, 35)]), ivs, &(xi[WS(is, 1)])); Chris@10: T1P = LD(&(xi[WS(is, 99)]), ivs, &(xi[WS(is, 1)])); Chris@10: TbK = VADD(T1O, T1P); Chris@10: T1S = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)])); Chris@10: T1T = LD(&(xi[WS(is, 67)]), ivs, &(xi[WS(is, 1)])); Chris@10: TbJ = VADD(T1S, T1T); Chris@10: T1Q = VSUB(T1O, T1P); Chris@10: Tfb = VADD(TbJ, TbK); Chris@10: T1U = VSUB(T1S, T1T); Chris@10: TbL = VSUB(TbJ, TbK); Chris@10: } Chris@10: { Chris@10: V T25, T26, TbR, T29, T2a, TbQ; Chris@10: T25 = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)])); Chris@10: T26 = LD(&(xi[WS(is, 91)]), ivs, &(xi[WS(is, 1)])); Chris@10: TbR = VADD(T25, T26); Chris@10: T29 = LD(&(xi[WS(is, 123)]), ivs, &(xi[WS(is, 1)])); Chris@10: T2a = LD(&(xi[WS(is, 59)]), ivs, &(xi[WS(is, 1)])); Chris@10: TbQ = VADD(T29, T2a); Chris@10: T27 = VSUB(T25, T26); Chris@10: Tfe = VADD(TbQ, TbR); Chris@10: T2b = VSUB(T29, T2a); Chris@10: TbS = VSUB(TbQ, TbR); Chris@10: } Chris@10: { Chris@10: V T1J, TbM, T1M, TbN; Chris@10: { Chris@10: V T1H, T1I, T1K, T1L; Chris@10: T1H = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)])); Chris@10: T1I = LD(&(xi[WS(is, 83)]), ivs, &(xi[WS(is, 1)])); Chris@10: T1J = VSUB(T1H, T1I); Chris@10: TbM = VADD(T1H, T1I); Chris@10: T1K = LD(&(xi[WS(is, 115)]), ivs, &(xi[WS(is, 1)])); Chris@10: T1L = LD(&(xi[WS(is, 51)]), ivs, &(xi[WS(is, 1)])); Chris@10: T1M = VSUB(T1K, T1L); Chris@10: TbN = VADD(T1K, T1L); Chris@10: } Chris@10: T1N = VMUL(LDK(KP707106781), VSUB(T1J, T1M)); Chris@10: Tfc = VADD(TbM, TbN); Chris@10: T1V = VMUL(LDK(KP707106781), VADD(T1J, T1M)); Chris@10: TbO = VSUB(TbM, TbN); Chris@10: } Chris@10: { Chris@10: V T20, TbT, T23, TbU; Chris@10: { Chris@10: V T1Y, T1Z, T21, T22; Chris@10: T1Y = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)])); Chris@10: T1Z = LD(&(xi[WS(is, 75)]), ivs, &(xi[WS(is, 1)])); Chris@10: T20 = VSUB(T1Y, T1Z); Chris@10: TbT = VADD(T1Y, T1Z); Chris@10: T21 = LD(&(xi[WS(is, 107)]), ivs, &(xi[WS(is, 1)])); Chris@10: T22 = LD(&(xi[WS(is, 43)]), ivs, &(xi[WS(is, 1)])); Chris@10: T23 = VSUB(T21, T22); Chris@10: TbU = VADD(T21, T22); Chris@10: } Chris@10: T24 = VMUL(LDK(KP707106781), VSUB(T20, T23)); Chris@10: Tff = VADD(TbT, TbU); Chris@10: T2c = VMUL(LDK(KP707106781), VADD(T20, T23)); Chris@10: TbV = VSUB(TbT, TbU); Chris@10: } Chris@10: { Chris@10: V TbP, TbW, Tfd, Tfg; Chris@10: TbP = VFNMS(LDK(KP382683432), TbO, VMUL(LDK(KP923879532), TbL)); Chris@10: TbW = VFMA(LDK(KP923879532), TbS, VMUL(LDK(KP382683432), TbV)); Chris@10: TbX = VSUB(TbP, TbW); Chris@10: Tdn = VADD(TbP, TbW); Chris@10: { Chris@10: V Tcf, Tcg, Tgt, Tgu; Chris@10: Tcf = VFMA(LDK(KP382683432), TbL, VMUL(LDK(KP923879532), TbO)); Chris@10: Tcg = VFNMS(LDK(KP382683432), TbS, VMUL(LDK(KP923879532), TbV)); Chris@10: Tch = VSUB(Tcf, Tcg); Chris@10: Tdq = VADD(Tcf, Tcg); Chris@10: Tgt = VADD(Tfb, Tfc); Chris@10: Tgu = VADD(Tfe, Tff); Chris@10: Tgv = VSUB(Tgt, Tgu); Chris@10: Th7 = VADD(Tgt, Tgu); Chris@10: } Chris@10: Tfd = VSUB(Tfb, Tfc); Chris@10: Tfg = VSUB(Tfe, Tff); Chris@10: Tfh = VMUL(LDK(KP707106781), VSUB(Tfd, Tfg)); Chris@10: Tfp = VMUL(LDK(KP707106781), VADD(Tfd, Tfg)); Chris@10: { Chris@10: V T7l, T7u, T7o, T7v; Chris@10: { Chris@10: V T7j, T7k, T7m, T7n; Chris@10: T7j = VADD(T1Q, T1N); Chris@10: T7k = VADD(T1U, T1V); Chris@10: T7l = VFMA(LDK(KP980785280), T7j, VMUL(LDK(KP195090322), T7k)); Chris@10: T7u = VFNMS(LDK(KP195090322), T7j, VMUL(LDK(KP980785280), T7k)); Chris@10: T7m = VADD(T27, T24); Chris@10: T7n = VADD(T2b, T2c); Chris@10: T7o = VFNMS(LDK(KP195090322), T7n, VMUL(LDK(KP980785280), T7m)); Chris@10: T7v = VFMA(LDK(KP195090322), T7m, VMUL(LDK(KP980785280), T7n)); Chris@10: } Chris@10: T7p = VSUB(T7l, T7o); Chris@10: T9f = VADD(T7u, T7v); Chris@10: T7w = VSUB(T7u, T7v); Chris@10: T9b = VADD(T7l, T7o); Chris@10: } Chris@10: { Chris@10: V T1X, T2q, T2e, T2r; Chris@10: { Chris@10: V T1R, T1W, T28, T2d; Chris@10: T1R = VSUB(T1N, T1Q); Chris@10: T1W = VSUB(T1U, T1V); Chris@10: T1X = VFMA(LDK(KP831469612), T1R, VMUL(LDK(KP555570233), T1W)); Chris@10: T2q = VFNMS(LDK(KP555570233), T1R, VMUL(LDK(KP831469612), T1W)); Chris@10: T28 = VSUB(T24, T27); Chris@10: T2d = VSUB(T2b, T2c); Chris@10: T2e = VFNMS(LDK(KP555570233), T2d, VMUL(LDK(KP831469612), T28)); Chris@10: T2r = VFMA(LDK(KP555570233), T28, VMUL(LDK(KP831469612), T2d)); Chris@10: } Chris@10: T2f = VSUB(T1X, T2e); Chris@10: T65 = VADD(T2q, T2r); Chris@10: T2s = VSUB(T2q, T2r); Chris@10: T61 = VADD(T1X, T2e); Chris@10: } Chris@10: } Chris@10: } Chris@10: { Chris@10: V Tgx, TgW, TgR, TgZ, TgI, TgY, TgO, TgV; Chris@10: { Chris@10: V Tgp, Tgw, TgP, TgQ; Chris@10: Tgp = VFNMS(LDK(KP382683432), Tgo, VMUL(LDK(KP923879532), Tgl)); Chris@10: Tgw = VFMA(LDK(KP923879532), Tgs, VMUL(LDK(KP382683432), Tgv)); Chris@10: Tgx = VSUB(Tgp, Tgw); Chris@10: TgW = VADD(Tgp, Tgw); Chris@10: TgP = VFMA(LDK(KP382683432), Tgl, VMUL(LDK(KP923879532), Tgo)); Chris@10: TgQ = VFNMS(LDK(KP382683432), Tgs, VMUL(LDK(KP923879532), Tgv)); Chris@10: TgR = VSUB(TgP, TgQ); Chris@10: TgZ = VADD(TgP, TgQ); Chris@10: } Chris@10: { Chris@10: V TgE, TgH, TgM, TgN; Chris@10: TgE = VMUL(LDK(KP707106781), VSUB(TgA, TgD)); Chris@10: TgH = VSUB(TgF, TgG); Chris@10: TgI = VSUB(TgE, TgH); Chris@10: TgY = VADD(TgH, TgE); Chris@10: TgM = VSUB(TgK, TgL); Chris@10: TgN = VMUL(LDK(KP707106781), VADD(TgA, TgD)); Chris@10: TgO = VSUB(TgM, TgN); Chris@10: TgV = VADD(TgM, TgN); Chris@10: } Chris@10: { Chris@10: V TgJ, TgS, Th1, Th2; Chris@10: TgJ = VBYI(VSUB(Tgx, TgI)); Chris@10: TgS = VSUB(TgO, TgR); Chris@10: ST(&(xo[WS(os, 40)]), VADD(TgJ, TgS), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 88)]), VSUB(TgS, TgJ), ovs, &(xo[0])); Chris@10: Th1 = VSUB(TgV, TgW); Chris@10: Th2 = VBYI(VSUB(TgZ, TgY)); Chris@10: ST(&(xo[WS(os, 72)]), VSUB(Th1, Th2), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 56)]), VADD(Th1, Th2), ovs, &(xo[0])); Chris@10: } Chris@10: { Chris@10: V TgT, TgU, TgX, Th0; Chris@10: TgT = VBYI(VADD(TgI, Tgx)); Chris@10: TgU = VADD(TgO, TgR); Chris@10: ST(&(xo[WS(os, 24)]), VADD(TgT, TgU), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 104)]), VSUB(TgU, TgT), ovs, &(xo[0])); Chris@10: TgX = VADD(TgV, TgW); Chris@10: Th0 = VBYI(VADD(TgY, TgZ)); Chris@10: ST(&(xo[WS(os, 120)]), VSUB(TgX, Th0), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 8)]), VADD(TgX, Th0), ovs, &(xo[0])); Chris@10: } Chris@10: } Chris@10: { Chris@10: V Th9, Thh, Thq, Ths, Thc, Thm, Thg, Thl, Thn, Thr; Chris@10: { Chris@10: V Th5, Th8, Tho, Thp; Chris@10: Th5 = VSUB(Th3, Th4); Chris@10: Th8 = VSUB(Th6, Th7); Chris@10: Th9 = VMUL(LDK(KP707106781), VSUB(Th5, Th8)); Chris@10: Thh = VMUL(LDK(KP707106781), VADD(Th5, Th8)); Chris@10: Tho = VADD(Th3, Th4); Chris@10: Thp = VADD(Th6, Th7); Chris@10: Thq = VBYI(VSUB(Tho, Thp)); Chris@10: Ths = VADD(Tho, Thp); Chris@10: } Chris@10: { Chris@10: V Tha, Thb, The, Thf; Chris@10: Tha = VADD(Tgy, Tgz); Chris@10: Thb = VADD(TgB, TgC); Chris@10: Thc = VSUB(Tha, Thb); Chris@10: Thm = VADD(Tha, Thb); Chris@10: The = VADD(TgK, TgL); Chris@10: Thf = VADD(TgF, TgG); Chris@10: Thg = VSUB(The, Thf); Chris@10: Thl = VADD(The, Thf); Chris@10: } Chris@10: Thn = VSUB(Thl, Thm); Chris@10: ST(&(xo[WS(os, 96)]), VSUB(Thn, Thq), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 32)]), VADD(Thn, Thq), ovs, &(xo[0])); Chris@10: Thr = VADD(Thl, Thm); Chris@10: ST(&(xo[WS(os, 64)]), VSUB(Thr, Ths), ovs, &(xo[0])); Chris@10: ST(&(xo[0]), VADD(Thr, Ths), ovs, &(xo[0])); Chris@10: { Chris@10: V Thd, Thi, Thj, Thk; Chris@10: Thd = VBYI(VSUB(Th9, Thc)); Chris@10: Thi = VSUB(Thg, Thh); Chris@10: ST(&(xo[WS(os, 48)]), VADD(Thd, Thi), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 80)]), VSUB(Thi, Thd), ovs, &(xo[0])); Chris@10: Thj = VBYI(VADD(Thc, Th9)); Chris@10: Thk = VADD(Thg, Thh); Chris@10: ST(&(xo[WS(os, 16)]), VADD(Thj, Thk), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 112)]), VSUB(Thk, Thj), ovs, &(xo[0])); Chris@10: } Chris@10: } Chris@10: { Chris@10: V TeT, TfM, TfC, TfK, Tfs, TfN, TfF, TfJ; Chris@10: { Chris@10: V TeD, TeS, Tfw, TfB; Chris@10: TeD = VSUB(Tev, TeC); Chris@10: TeS = VSUB(TeK, TeR); Chris@10: TeT = VSUB(TeD, TeS); Chris@10: TfM = VADD(TeD, TeS); Chris@10: Tfw = VSUB(Tfu, Tfv); Chris@10: TfB = VSUB(Tfx, TfA); Chris@10: TfC = VSUB(Tfw, TfB); Chris@10: TfK = VADD(TfB, Tfw); Chris@10: { Chris@10: V Tfa, TfD, Tfr, TfE; Chris@10: { Chris@10: V Tf4, Tf9, Tfl, Tfq; Chris@10: Tf4 = VSUB(Tf0, Tf3); Chris@10: Tf9 = VSUB(Tf7, Tf8); Chris@10: Tfa = VFMA(LDK(KP831469612), Tf4, VMUL(LDK(KP555570233), Tf9)); Chris@10: TfD = VFNMS(LDK(KP555570233), Tf4, VMUL(LDK(KP831469612), Tf9)); Chris@10: Tfl = VSUB(Tfh, Tfk); Chris@10: Tfq = VSUB(Tfo, Tfp); Chris@10: Tfr = VFNMS(LDK(KP555570233), Tfq, VMUL(LDK(KP831469612), Tfl)); Chris@10: TfE = VFMA(LDK(KP555570233), Tfl, VMUL(LDK(KP831469612), Tfq)); Chris@10: } Chris@10: Tfs = VSUB(Tfa, Tfr); Chris@10: TfN = VADD(TfD, TfE); Chris@10: TfF = VSUB(TfD, TfE); Chris@10: TfJ = VADD(Tfa, Tfr); Chris@10: } Chris@10: } Chris@10: { Chris@10: V Tft, TfG, TfP, TfQ; Chris@10: Tft = VADD(TeT, Tfs); Chris@10: TfG = VBYI(VADD(TfC, TfF)); Chris@10: ST(&(xo[WS(os, 108)]), VSUB(Tft, TfG), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 20)]), VADD(Tft, TfG), ovs, &(xo[0])); Chris@10: TfP = VBYI(VADD(TfK, TfJ)); Chris@10: TfQ = VADD(TfM, TfN); Chris@10: ST(&(xo[WS(os, 12)]), VADD(TfP, TfQ), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 116)]), VSUB(TfQ, TfP), ovs, &(xo[0])); Chris@10: } Chris@10: { Chris@10: V TfH, TfI, TfL, TfO; Chris@10: TfH = VSUB(TeT, Tfs); Chris@10: TfI = VBYI(VSUB(TfF, TfC)); Chris@10: ST(&(xo[WS(os, 84)]), VSUB(TfH, TfI), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 44)]), VADD(TfH, TfI), ovs, &(xo[0])); Chris@10: TfL = VBYI(VSUB(TfJ, TfK)); Chris@10: TfO = VSUB(TfM, TfN); Chris@10: ST(&(xo[WS(os, 52)]), VADD(TfL, TfO), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 76)]), VSUB(TfO, TfL), ovs, &(xo[0])); Chris@10: } Chris@10: } Chris@10: { Chris@10: V TfT, Tge, Tg4, Tgc, Tg0, Tgf, Tg7, Tgb; Chris@10: { Chris@10: V TfR, TfS, Tg2, Tg3; Chris@10: TfR = VADD(Tev, TeC); Chris@10: TfS = VADD(Tfu, Tfv); Chris@10: TfT = VSUB(TfR, TfS); Chris@10: Tge = VADD(TfR, TfS); Chris@10: Tg2 = VADD(TeK, TeR); Chris@10: Tg3 = VADD(TfA, Tfx); Chris@10: Tg4 = VSUB(Tg2, Tg3); Chris@10: Tgc = VADD(Tg3, Tg2); Chris@10: { Chris@10: V TfW, Tg5, TfZ, Tg6; Chris@10: { Chris@10: V TfU, TfV, TfX, TfY; Chris@10: TfU = VADD(Tf3, Tf0); Chris@10: TfV = VADD(Tf7, Tf8); Chris@10: TfW = VFMA(LDK(KP980785280), TfU, VMUL(LDK(KP195090322), TfV)); Chris@10: Tg5 = VFNMS(LDK(KP195090322), TfU, VMUL(LDK(KP980785280), TfV)); Chris@10: TfX = VADD(Tfk, Tfh); Chris@10: TfY = VADD(Tfo, Tfp); Chris@10: TfZ = VFNMS(LDK(KP195090322), TfY, VMUL(LDK(KP980785280), TfX)); Chris@10: Tg6 = VFMA(LDK(KP195090322), TfX, VMUL(LDK(KP980785280), TfY)); Chris@10: } Chris@10: Tg0 = VSUB(TfW, TfZ); Chris@10: Tgf = VADD(Tg5, Tg6); Chris@10: Tg7 = VSUB(Tg5, Tg6); Chris@10: Tgb = VADD(TfW, TfZ); Chris@10: } Chris@10: } Chris@10: { Chris@10: V Tg1, Tg8, Tgh, Tgi; Chris@10: Tg1 = VADD(TfT, Tg0); Chris@10: Tg8 = VBYI(VADD(Tg4, Tg7)); Chris@10: ST(&(xo[WS(os, 100)]), VSUB(Tg1, Tg8), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 28)]), VADD(Tg1, Tg8), ovs, &(xo[0])); Chris@10: Tgh = VBYI(VADD(Tgc, Tgb)); Chris@10: Tgi = VADD(Tge, Tgf); Chris@10: ST(&(xo[WS(os, 4)]), VADD(Tgh, Tgi), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 124)]), VSUB(Tgi, Tgh), ovs, &(xo[0])); Chris@10: } Chris@10: { Chris@10: V Tg9, Tga, Tgd, Tgg; Chris@10: Tg9 = VSUB(TfT, Tg0); Chris@10: Tga = VBYI(VSUB(Tg7, Tg4)); Chris@10: ST(&(xo[WS(os, 92)]), VSUB(Tg9, Tga), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 36)]), VADD(Tg9, Tga), ovs, &(xo[0])); Chris@10: Tgd = VBYI(VSUB(Tgb, Tgc)); Chris@10: Tgg = VSUB(Tge, Tgf); Chris@10: ST(&(xo[WS(os, 60)]), VADD(Tgd, Tgg), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 68)]), VSUB(Tgg, Tgd), ovs, &(xo[0])); Chris@10: } Chris@10: } Chris@10: { Chris@10: V Tb7, Td8, TcI, Td0, Tcy, Tda, TcG, TcP, Tck, TcJ, TcB, TcF, TcW, Tdb, Td3; Chris@10: V Td7; Chris@10: { Chris@10: V Tax, TcZ, Tb6, TcY, TaO, Tb5; Chris@10: Tax = VSUB(Tah, Taw); Chris@10: TcZ = VADD(Tcw, Tcr); Chris@10: TaO = VFMA(LDK(KP831469612), TaI, VMUL(LDK(KP555570233), TaN)); Chris@10: Tb5 = VFNMS(LDK(KP555570233), Tb4, VMUL(LDK(KP831469612), TaZ)); Chris@10: Tb6 = VSUB(TaO, Tb5); Chris@10: TcY = VADD(TaO, Tb5); Chris@10: Tb7 = VSUB(Tax, Tb6); Chris@10: Td8 = VADD(TcZ, TcY); Chris@10: TcI = VADD(Tax, Tb6); Chris@10: Td0 = VSUB(TcY, TcZ); Chris@10: } Chris@10: { Chris@10: V Tcx, TcN, Tco, TcO, Tcm, Tcn; Chris@10: Tcx = VSUB(Tcr, Tcw); Chris@10: TcN = VADD(Tah, Taw); Chris@10: Tcm = VFNMS(LDK(KP555570233), TaI, VMUL(LDK(KP831469612), TaN)); Chris@10: Tcn = VFMA(LDK(KP555570233), TaZ, VMUL(LDK(KP831469612), Tb4)); Chris@10: Tco = VSUB(Tcm, Tcn); Chris@10: TcO = VADD(Tcm, Tcn); Chris@10: Tcy = VSUB(Tco, Tcx); Chris@10: Tda = VADD(TcN, TcO); Chris@10: TcG = VADD(Tcx, Tco); Chris@10: TcP = VSUB(TcN, TcO); Chris@10: } Chris@10: { Chris@10: V TbI, Tcz, Tcj, TcA; Chris@10: { Chris@10: V Tby, TbH, Tc9, Tci; Chris@10: Tby = VSUB(Tbm, Tbx); Chris@10: TbH = VSUB(TbD, TbG); Chris@10: TbI = VFMA(LDK(KP881921264), Tby, VMUL(LDK(KP471396736), TbH)); Chris@10: Tcz = VFNMS(LDK(KP471396736), Tby, VMUL(LDK(KP881921264), TbH)); Chris@10: Tc9 = VSUB(TbX, Tc8); Chris@10: Tci = VSUB(Tce, Tch); Chris@10: Tcj = VFNMS(LDK(KP471396736), Tci, VMUL(LDK(KP881921264), Tc9)); Chris@10: TcA = VFMA(LDK(KP471396736), Tc9, VMUL(LDK(KP881921264), Tci)); Chris@10: } Chris@10: Tck = VSUB(TbI, Tcj); Chris@10: TcJ = VADD(Tcz, TcA); Chris@10: TcB = VSUB(Tcz, TcA); Chris@10: TcF = VADD(TbI, Tcj); Chris@10: } Chris@10: { Chris@10: V TcS, Td1, TcV, Td2; Chris@10: { Chris@10: V TcQ, TcR, TcT, TcU; Chris@10: TcQ = VADD(Tbx, Tbm); Chris@10: TcR = VADD(TbD, TbG); Chris@10: TcS = VFMA(LDK(KP956940335), TcQ, VMUL(LDK(KP290284677), TcR)); Chris@10: Td1 = VFNMS(LDK(KP290284677), TcQ, VMUL(LDK(KP956940335), TcR)); Chris@10: TcT = VADD(Tc8, TbX); Chris@10: TcU = VADD(Tce, Tch); Chris@10: TcV = VFNMS(LDK(KP290284677), TcU, VMUL(LDK(KP956940335), TcT)); Chris@10: Td2 = VFMA(LDK(KP290284677), TcT, VMUL(LDK(KP956940335), TcU)); Chris@10: } Chris@10: TcW = VSUB(TcS, TcV); Chris@10: Tdb = VADD(Td1, Td2); Chris@10: Td3 = VSUB(Td1, Td2); Chris@10: Td7 = VADD(TcS, TcV); Chris@10: } Chris@10: { Chris@10: V Tcl, TcC, Td9, Tdc; Chris@10: Tcl = VADD(Tb7, Tck); Chris@10: TcC = VBYI(VADD(Tcy, TcB)); Chris@10: ST(&(xo[WS(os, 106)]), VSUB(Tcl, TcC), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 22)]), VADD(Tcl, TcC), ovs, &(xo[0])); Chris@10: Td9 = VBYI(VSUB(Td7, Td8)); Chris@10: Tdc = VSUB(Tda, Tdb); Chris@10: ST(&(xo[WS(os, 58)]), VADD(Td9, Tdc), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 70)]), VSUB(Tdc, Td9), ovs, &(xo[0])); Chris@10: } Chris@10: { Chris@10: V Tdd, Tde, TcD, TcE; Chris@10: Tdd = VBYI(VADD(Td8, Td7)); Chris@10: Tde = VADD(Tda, Tdb); Chris@10: ST(&(xo[WS(os, 6)]), VADD(Tdd, Tde), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 122)]), VSUB(Tde, Tdd), ovs, &(xo[0])); Chris@10: TcD = VSUB(Tb7, Tck); Chris@10: TcE = VBYI(VSUB(TcB, Tcy)); Chris@10: ST(&(xo[WS(os, 86)]), VSUB(TcD, TcE), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 42)]), VADD(TcD, TcE), ovs, &(xo[0])); Chris@10: } Chris@10: { Chris@10: V TcH, TcK, TcX, Td4; Chris@10: TcH = VBYI(VSUB(TcF, TcG)); Chris@10: TcK = VSUB(TcI, TcJ); Chris@10: ST(&(xo[WS(os, 54)]), VADD(TcH, TcK), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 74)]), VSUB(TcK, TcH), ovs, &(xo[0])); Chris@10: TcX = VADD(TcP, TcW); Chris@10: Td4 = VBYI(VADD(Td0, Td3)); Chris@10: ST(&(xo[WS(os, 102)]), VSUB(TcX, Td4), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 26)]), VADD(TcX, Td4), ovs, &(xo[0])); Chris@10: } Chris@10: { Chris@10: V Td5, Td6, TcL, TcM; Chris@10: Td5 = VSUB(TcP, TcW); Chris@10: Td6 = VBYI(VSUB(Td3, Td0)); Chris@10: ST(&(xo[WS(os, 90)]), VSUB(Td5, Td6), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 38)]), VADD(Td5, Td6), ovs, &(xo[0])); Chris@10: TcL = VBYI(VADD(TcG, TcF)); Chris@10: TcM = VADD(TcI, TcJ); Chris@10: ST(&(xo[WS(os, 10)]), VADD(TcL, TcM), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 118)]), VSUB(TcM, TcL), ovs, &(xo[0])); Chris@10: } Chris@10: } Chris@10: { Chris@10: V TdE, Tel, TdW, Tee, TdM, Teo, TdT, Tea, Tdt, TdX, TdP, TdU, Te7, Tep, Teh; Chris@10: V Tem; Chris@10: { Chris@10: V TdD, Tec, TdA, Ted, Tdw, Tdz; Chris@10: TdD = VADD(TdB, TdC); Chris@10: Tec = VSUB(TdG, TdH); Chris@10: Tdw = VFMA(LDK(KP980785280), Tdu, VMUL(LDK(KP195090322), Tdv)); Chris@10: Tdz = VFNMS(LDK(KP195090322), Tdy, VMUL(LDK(KP980785280), Tdx)); Chris@10: TdA = VADD(Tdw, Tdz); Chris@10: Ted = VSUB(Tdw, Tdz); Chris@10: TdE = VSUB(TdA, TdD); Chris@10: Tel = VADD(Tec, Ted); Chris@10: TdW = VADD(TdD, TdA); Chris@10: Tee = VSUB(Tec, Ted); Chris@10: } Chris@10: { Chris@10: V TdI, Te9, TdL, Te8, TdJ, TdK; Chris@10: TdI = VADD(TdG, TdH); Chris@10: Te9 = VSUB(TdC, TdB); Chris@10: TdJ = VFNMS(LDK(KP195090322), Tdu, VMUL(LDK(KP980785280), Tdv)); Chris@10: TdK = VFMA(LDK(KP195090322), Tdx, VMUL(LDK(KP980785280), Tdy)); Chris@10: TdL = VADD(TdJ, TdK); Chris@10: Te8 = VSUB(TdJ, TdK); Chris@10: TdM = VSUB(TdI, TdL); Chris@10: Teo = VADD(Te9, Te8); Chris@10: TdT = VADD(TdI, TdL); Chris@10: Tea = VSUB(Te8, Te9); Chris@10: } Chris@10: { Chris@10: V Tdl, TdN, Tds, TdO; Chris@10: { Chris@10: V Tdh, Tdk, Tdo, Tdr; Chris@10: Tdh = VADD(Tdf, Tdg); Chris@10: Tdk = VADD(Tdi, Tdj); Chris@10: Tdl = VFNMS(LDK(KP098017140), Tdk, VMUL(LDK(KP995184726), Tdh)); Chris@10: TdN = VFMA(LDK(KP098017140), Tdh, VMUL(LDK(KP995184726), Tdk)); Chris@10: Tdo = VADD(Tdm, Tdn); Chris@10: Tdr = VADD(Tdp, Tdq); Chris@10: Tds = VFMA(LDK(KP995184726), Tdo, VMUL(LDK(KP098017140), Tdr)); Chris@10: TdO = VFNMS(LDK(KP098017140), Tdo, VMUL(LDK(KP995184726), Tdr)); Chris@10: } Chris@10: Tdt = VSUB(Tdl, Tds); Chris@10: TdX = VADD(TdN, TdO); Chris@10: TdP = VSUB(TdN, TdO); Chris@10: TdU = VADD(Tdl, Tds); Chris@10: } Chris@10: { Chris@10: V Te3, Tef, Te6, Teg; Chris@10: { Chris@10: V Te1, Te2, Te4, Te5; Chris@10: Te1 = VSUB(Tdf, Tdg); Chris@10: Te2 = VSUB(Tdj, Tdi); Chris@10: Te3 = VFNMS(LDK(KP634393284), Te2, VMUL(LDK(KP773010453), Te1)); Chris@10: Tef = VFMA(LDK(KP634393284), Te1, VMUL(LDK(KP773010453), Te2)); Chris@10: Te4 = VSUB(Tdm, Tdn); Chris@10: Te5 = VSUB(Tdq, Tdp); Chris@10: Te6 = VFMA(LDK(KP773010453), Te4, VMUL(LDK(KP634393284), Te5)); Chris@10: Teg = VFNMS(LDK(KP634393284), Te4, VMUL(LDK(KP773010453), Te5)); Chris@10: } Chris@10: Te7 = VSUB(Te3, Te6); Chris@10: Tep = VADD(Tef, Teg); Chris@10: Teh = VSUB(Tef, Teg); Chris@10: Tem = VADD(Te3, Te6); Chris@10: } Chris@10: { Chris@10: V TdF, TdQ, Ten, Teq; Chris@10: TdF = VBYI(VSUB(Tdt, TdE)); Chris@10: TdQ = VSUB(TdM, TdP); Chris@10: ST(&(xo[WS(os, 34)]), VADD(TdF, TdQ), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 94)]), VSUB(TdQ, TdF), ovs, &(xo[0])); Chris@10: Ten = VADD(Tel, Tem); Chris@10: Teq = VBYI(VADD(Teo, Tep)); Chris@10: ST(&(xo[WS(os, 114)]), VSUB(Ten, Teq), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 14)]), VADD(Ten, Teq), ovs, &(xo[0])); Chris@10: } Chris@10: { Chris@10: V Ter, Tes, TdR, TdS; Chris@10: Ter = VSUB(Tel, Tem); Chris@10: Tes = VBYI(VSUB(Tep, Teo)); Chris@10: ST(&(xo[WS(os, 78)]), VSUB(Ter, Tes), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 50)]), VADD(Ter, Tes), ovs, &(xo[0])); Chris@10: TdR = VBYI(VADD(TdE, Tdt)); Chris@10: TdS = VADD(TdM, TdP); Chris@10: ST(&(xo[WS(os, 30)]), VADD(TdR, TdS), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 98)]), VSUB(TdS, TdR), ovs, &(xo[0])); Chris@10: } Chris@10: { Chris@10: V TdV, TdY, Teb, Tei; Chris@10: TdV = VADD(TdT, TdU); Chris@10: TdY = VBYI(VADD(TdW, TdX)); Chris@10: ST(&(xo[WS(os, 126)]), VSUB(TdV, TdY), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 2)]), VADD(TdV, TdY), ovs, &(xo[0])); Chris@10: Teb = VBYI(VSUB(Te7, Tea)); Chris@10: Tei = VSUB(Tee, Teh); Chris@10: ST(&(xo[WS(os, 46)]), VADD(Teb, Tei), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 82)]), VSUB(Tei, Teb), ovs, &(xo[0])); Chris@10: } Chris@10: { Chris@10: V Tej, Tek, TdZ, Te0; Chris@10: Tej = VBYI(VADD(Tea, Te7)); Chris@10: Tek = VADD(Tee, Teh); Chris@10: ST(&(xo[WS(os, 18)]), VADD(Tej, Tek), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 110)]), VSUB(Tek, Tej), ovs, &(xo[0])); Chris@10: TdZ = VSUB(TdT, TdU); Chris@10: Te0 = VBYI(VSUB(TdX, TdW)); Chris@10: ST(&(xo[WS(os, 66)]), VSUB(TdZ, Te0), ovs, &(xo[0])); Chris@10: ST(&(xo[WS(os, 62)]), VADD(TdZ, Te0), ovs, &(xo[0])); Chris@10: } Chris@10: } Chris@10: { Chris@10: V T7z, T8n, T8f, T8k, T8x, T8P, T8H, T8M, T80, T8L, T8O, T8c, T8j, T8A, T8E; Chris@10: V T8m; Chris@10: { Chris@10: V T7f, T8d, T7y, T8e; Chris@10: { Chris@10: V T77, T7e, T7q, T7x; Chris@10: T77 = VADD(T6Z, T76); Chris@10: T7e = VADD(T7a, T7d); Chris@10: T7f = VFNMS(LDK(KP336889853), T7e, VMUL(LDK(KP941544065), T77)); Chris@10: T8d = VFMA(LDK(KP336889853), T77, VMUL(LDK(KP941544065), T7e)); Chris@10: T7q = VADD(T7i, T7p); Chris@10: T7x = VADD(T7t, T7w); Chris@10: T7y = VFMA(LDK(KP941544065), T7q, VMUL(LDK(KP336889853), T7x)); Chris@10: T8e = VFNMS(LDK(KP336889853), T7q, VMUL(LDK(KP941544065), T7x)); Chris@10: } Chris@10: T7z = VSUB(T7f, T7y); Chris@10: T8n = VADD(T8d, T8e); Chris@10: T8f = VSUB(T8d, T8e); Chris@10: T8k = VADD(T7f, T7y); Chris@10: } Chris@10: { Chris@10: V T8t, T8F, T8w, T8G; Chris@10: { Chris@10: V T8r, T8s, T8u, T8v; Chris@10: T8r = VSUB(T6Z, T76); Chris@10: T8s = VSUB(T7d, T7a); Chris@10: T8t = VFNMS(LDK(KP427555093), T8s, VMUL(LDK(KP903989293), T8r)); Chris@10: T8F = VFMA(LDK(KP427555093), T8r, VMUL(LDK(KP903989293), T8s)); Chris@10: T8u = VSUB(T7i, T7p); Chris@10: T8v = VSUB(T7w, T7t); Chris@10: T8w = VFMA(LDK(KP903989293), T8u, VMUL(LDK(KP427555093), T8v)); Chris@10: T8G = VFNMS(LDK(KP427555093), T8u, VMUL(LDK(KP903989293), T8v)); Chris@10: } Chris@10: T8x = VSUB(T8t, T8w); Chris@10: T8P = VADD(T8F, T8G); Chris@10: T8H = VSUB(T8F, T8G); Chris@10: T8M = VADD(T8t, T8w); Chris@10: } Chris@10: { Chris@10: V T7Z, T8z, T88, T8C, T7O, T8D, T8b, T8y, T7Y, T87; Chris@10: T7Y = VSUB(T7U, T7X); Chris@10: T7Z = VADD(T7R, T7Y); Chris@10: T8z = VSUB(T7Y, T7R); Chris@10: T87 = VSUB(T85, T86); Chris@10: T88 = VADD(T84, T87); Chris@10: T8C = VSUB(T84, T87); Chris@10: { Chris@10: V T7G, T7N, T89, T8a; Chris@10: T7G = VFMA(LDK(KP634393284), T7C, VMUL(LDK(KP773010453), T7F)); Chris@10: T7N = VFNMS(LDK(KP634393284), T7M, VMUL(LDK(KP773010453), T7J)); Chris@10: T7O = VADD(T7G, T7N); Chris@10: T8D = VSUB(T7G, T7N); Chris@10: T89 = VFNMS(LDK(KP634393284), T7F, VMUL(LDK(KP773010453), T7C)); Chris@10: T8a = VFMA(LDK(KP773010453), T7M, VMUL(LDK(KP634393284), T7J)); Chris@10: T8b = VADD(T89, T8a); Chris@10: T8y = VSUB(T89, T8a); Chris@10: } Chris@10: T80 = VSUB(T7O, T7Z); Chris@10: T8L = VADD(T8C, T8D); Chris@10: T8O = VADD(T8z, T8y); Chris@10: T8c = VSUB(T88, T8b); Chris@10: T8j = VADD(T88, T8b); Chris@10: T8A = VSUB(T8y, T8z); Chris@10: T8E = VSUB(T8C, T8D); Chris@10: T8m = VADD(T7Z, T7O); Chris@10: } Chris@10: { Chris@10: V T81, T8g, T8N, T8Q; Chris@10: T81 = VBYI(VSUB(T7z, T80)); Chris@10: T8g = VSUB(T8c, T8f); Chris@10: ST(&(xo[WS(os, 39)]), VADD(T81, T8g), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 89)]), VSUB(T8g, T81), ovs, &(xo[WS(os, 1)])); Chris@10: T8N = VADD(T8L, T8M); Chris@10: T8Q = VBYI(VADD(T8O, T8P)); Chris@10: ST(&(xo[WS(os, 119)]), VSUB(T8N, T8Q), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 9)]), VADD(T8N, T8Q), ovs, &(xo[WS(os, 1)])); Chris@10: } Chris@10: { Chris@10: V T8R, T8S, T8h, T8i; Chris@10: T8R = VSUB(T8L, T8M); Chris@10: T8S = VBYI(VSUB(T8P, T8O)); Chris@10: ST(&(xo[WS(os, 73)]), VSUB(T8R, T8S), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 55)]), VADD(T8R, T8S), ovs, &(xo[WS(os, 1)])); Chris@10: T8h = VBYI(VADD(T80, T7z)); Chris@10: T8i = VADD(T8c, T8f); Chris@10: ST(&(xo[WS(os, 25)]), VADD(T8h, T8i), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 103)]), VSUB(T8i, T8h), ovs, &(xo[WS(os, 1)])); Chris@10: } Chris@10: { Chris@10: V T8l, T8o, T8B, T8I; Chris@10: T8l = VADD(T8j, T8k); Chris@10: T8o = VBYI(VADD(T8m, T8n)); Chris@10: ST(&(xo[WS(os, 121)]), VSUB(T8l, T8o), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 7)]), VADD(T8l, T8o), ovs, &(xo[WS(os, 1)])); Chris@10: T8B = VBYI(VSUB(T8x, T8A)); Chris@10: T8I = VSUB(T8E, T8H); Chris@10: ST(&(xo[WS(os, 41)]), VADD(T8B, T8I), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 87)]), VSUB(T8I, T8B), ovs, &(xo[WS(os, 1)])); Chris@10: } Chris@10: { Chris@10: V T8J, T8K, T8p, T8q; Chris@10: T8J = VBYI(VADD(T8A, T8x)); Chris@10: T8K = VADD(T8E, T8H); Chris@10: ST(&(xo[WS(os, 23)]), VADD(T8J, T8K), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 105)]), VSUB(T8K, T8J), ovs, &(xo[WS(os, 1)])); Chris@10: T8p = VSUB(T8j, T8k); Chris@10: T8q = VBYI(VSUB(T8n, T8m)); Chris@10: ST(&(xo[WS(os, 71)]), VSUB(T8p, T8q), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 57)]), VADD(T8p, T8q), ovs, &(xo[WS(os, 1)])); Chris@10: } Chris@10: } Chris@10: { Chris@10: V T2v, T5d, T55, T5a, T5n, T5F, T5x, T5C, T4K, T5B, T5E, T52, T59, T5q, T5u; Chris@10: V T5c; Chris@10: { Chris@10: V T1f, T53, T2u, T54; Chris@10: { Chris@10: V T11, T1e, T2g, T2t; Chris@10: T11 = VADD(Tr, T10); Chris@10: T1e = VADD(T1a, T1d); Chris@10: T1f = VFNMS(LDK(KP242980179), T1e, VMUL(LDK(KP970031253), T11)); Chris@10: T53 = VFMA(LDK(KP242980179), T11, VMUL(LDK(KP970031253), T1e)); Chris@10: T2g = VADD(T1G, T2f); Chris@10: T2t = VADD(T2p, T2s); Chris@10: T2u = VFMA(LDK(KP970031253), T2g, VMUL(LDK(KP242980179), T2t)); Chris@10: T54 = VFNMS(LDK(KP242980179), T2g, VMUL(LDK(KP970031253), T2t)); Chris@10: } Chris@10: T2v = VSUB(T1f, T2u); Chris@10: T5d = VADD(T53, T54); Chris@10: T55 = VSUB(T53, T54); Chris@10: T5a = VADD(T1f, T2u); Chris@10: } Chris@10: { Chris@10: V T5j, T5v, T5m, T5w; Chris@10: { Chris@10: V T5h, T5i, T5k, T5l; Chris@10: T5h = VSUB(Tr, T10); Chris@10: T5i = VSUB(T1d, T1a); Chris@10: T5j = VFNMS(LDK(KP514102744), T5i, VMUL(LDK(KP857728610), T5h)); Chris@10: T5v = VFMA(LDK(KP514102744), T5h, VMUL(LDK(KP857728610), T5i)); Chris@10: T5k = VSUB(T1G, T2f); Chris@10: T5l = VSUB(T2s, T2p); Chris@10: T5m = VFMA(LDK(KP857728610), T5k, VMUL(LDK(KP514102744), T5l)); Chris@10: T5w = VFNMS(LDK(KP514102744), T5k, VMUL(LDK(KP857728610), T5l)); Chris@10: } Chris@10: T5n = VSUB(T5j, T5m); Chris@10: T5F = VADD(T5v, T5w); Chris@10: T5x = VSUB(T5v, T5w); Chris@10: T5C = VADD(T5j, T5m); Chris@10: } Chris@10: { Chris@10: V T4J, T5p, T4Y, T5s, T3I, T5t, T51, T5o, T4I, T4X; Chris@10: T4I = VSUB(T4q, T4H); Chris@10: T4J = VADD(T49, T4I); Chris@10: T5p = VSUB(T4I, T49); Chris@10: T4X = VSUB(T4V, T4W); Chris@10: T4Y = VADD(T4U, T4X); Chris@10: T5s = VSUB(T4U, T4X); Chris@10: { Chris@10: V T36, T3H, T4Z, T50; Chris@10: T36 = VFMA(LDK(KP881921264), T2W, VMUL(LDK(KP471396736), T35)); Chris@10: T3H = VFNMS(LDK(KP471396736), T3G, VMUL(LDK(KP881921264), T3x)); Chris@10: T3I = VADD(T36, T3H); Chris@10: T5t = VSUB(T36, T3H); Chris@10: T4Z = VFNMS(LDK(KP471396736), T2W, VMUL(LDK(KP881921264), T35)); Chris@10: T50 = VFMA(LDK(KP471396736), T3x, VMUL(LDK(KP881921264), T3G)); Chris@10: T51 = VADD(T4Z, T50); Chris@10: T5o = VSUB(T4Z, T50); Chris@10: } Chris@10: T4K = VSUB(T3I, T4J); Chris@10: T5B = VADD(T5s, T5t); Chris@10: T5E = VADD(T5p, T5o); Chris@10: T52 = VSUB(T4Y, T51); Chris@10: T59 = VADD(T4Y, T51); Chris@10: T5q = VSUB(T5o, T5p); Chris@10: T5u = VSUB(T5s, T5t); Chris@10: T5c = VADD(T4J, T3I); Chris@10: } Chris@10: { Chris@10: V T4L, T56, T5D, T5G; Chris@10: T4L = VBYI(VSUB(T2v, T4K)); Chris@10: T56 = VSUB(T52, T55); Chris@10: ST(&(xo[WS(os, 37)]), VADD(T4L, T56), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 91)]), VSUB(T56, T4L), ovs, &(xo[WS(os, 1)])); Chris@10: T5D = VADD(T5B, T5C); Chris@10: T5G = VBYI(VADD(T5E, T5F)); Chris@10: ST(&(xo[WS(os, 117)]), VSUB(T5D, T5G), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 11)]), VADD(T5D, T5G), ovs, &(xo[WS(os, 1)])); Chris@10: } Chris@10: { Chris@10: V T5H, T5I, T57, T58; Chris@10: T5H = VSUB(T5B, T5C); Chris@10: T5I = VBYI(VSUB(T5F, T5E)); Chris@10: ST(&(xo[WS(os, 75)]), VSUB(T5H, T5I), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 53)]), VADD(T5H, T5I), ovs, &(xo[WS(os, 1)])); Chris@10: T57 = VBYI(VADD(T4K, T2v)); Chris@10: T58 = VADD(T52, T55); Chris@10: ST(&(xo[WS(os, 27)]), VADD(T57, T58), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 101)]), VSUB(T58, T57), ovs, &(xo[WS(os, 1)])); Chris@10: } Chris@10: { Chris@10: V T5b, T5e, T5r, T5y; Chris@10: T5b = VADD(T59, T5a); Chris@10: T5e = VBYI(VADD(T5c, T5d)); Chris@10: ST(&(xo[WS(os, 123)]), VSUB(T5b, T5e), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 5)]), VADD(T5b, T5e), ovs, &(xo[WS(os, 1)])); Chris@10: T5r = VBYI(VSUB(T5n, T5q)); Chris@10: T5y = VSUB(T5u, T5x); Chris@10: ST(&(xo[WS(os, 43)]), VADD(T5r, T5y), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 85)]), VSUB(T5y, T5r), ovs, &(xo[WS(os, 1)])); Chris@10: } Chris@10: { Chris@10: V T5z, T5A, T5f, T5g; Chris@10: T5z = VBYI(VADD(T5q, T5n)); Chris@10: T5A = VADD(T5u, T5x); Chris@10: ST(&(xo[WS(os, 21)]), VADD(T5z, T5A), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 107)]), VSUB(T5A, T5z), ovs, &(xo[WS(os, 1)])); Chris@10: T5f = VSUB(T59, T5a); Chris@10: T5g = VBYI(VSUB(T5d, T5c)); Chris@10: ST(&(xo[WS(os, 69)]), VSUB(T5f, T5g), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 59)]), VADD(T5f, T5g), ovs, &(xo[WS(os, 1)])); Chris@10: } Chris@10: } Chris@10: { Chris@10: V T9i, T9B, T9t, T9x, T9O, Ta3, T9V, T9Z, T93, Ta0, Ta2, T9q, T9y, T9H, T9S; Chris@10: V T9A; Chris@10: { Chris@10: V T9a, T9r, T9h, T9s; Chris@10: { Chris@10: V T96, T99, T9d, T9g; Chris@10: T96 = VSUB(T94, T95); Chris@10: T99 = VSUB(T97, T98); Chris@10: T9a = VFMA(LDK(KP740951125), T96, VMUL(LDK(KP671558954), T99)); Chris@10: T9r = VFNMS(LDK(KP671558954), T96, VMUL(LDK(KP740951125), T99)); Chris@10: T9d = VSUB(T9b, T9c); Chris@10: T9g = VSUB(T9e, T9f); Chris@10: T9h = VFNMS(LDK(KP671558954), T9g, VMUL(LDK(KP740951125), T9d)); Chris@10: T9s = VFMA(LDK(KP671558954), T9d, VMUL(LDK(KP740951125), T9g)); Chris@10: } Chris@10: T9i = VSUB(T9a, T9h); Chris@10: T9B = VADD(T9r, T9s); Chris@10: T9t = VSUB(T9r, T9s); Chris@10: T9x = VADD(T9a, T9h); Chris@10: } Chris@10: { Chris@10: V T9K, T9T, T9N, T9U; Chris@10: { Chris@10: V T9I, T9J, T9L, T9M; Chris@10: T9I = VADD(T95, T94); Chris@10: T9J = VADD(T97, T98); Chris@10: T9K = VFMA(LDK(KP998795456), T9I, VMUL(LDK(KP049067674), T9J)); Chris@10: T9T = VFNMS(LDK(KP049067674), T9I, VMUL(LDK(KP998795456), T9J)); Chris@10: T9L = VADD(T9c, T9b); Chris@10: T9M = VADD(T9e, T9f); Chris@10: T9N = VFNMS(LDK(KP049067674), T9M, VMUL(LDK(KP998795456), T9L)); Chris@10: T9U = VFMA(LDK(KP049067674), T9L, VMUL(LDK(KP998795456), T9M)); Chris@10: } Chris@10: T9O = VSUB(T9K, T9N); Chris@10: Ta3 = VADD(T9T, T9U); Chris@10: T9V = VSUB(T9T, T9U); Chris@10: T9Z = VADD(T9K, T9N); Chris@10: } Chris@10: { Chris@10: V T8V, T9F, T9p, T9R, T92, T9Q, T9m, T9G, T8U, T9n; Chris@10: T8U = VADD(T7U, T7X); Chris@10: T8V = VSUB(T8T, T8U); Chris@10: T9F = VADD(T8T, T8U); Chris@10: T9n = VADD(T85, T86); Chris@10: T9p = VSUB(T9n, T9o); Chris@10: T9R = VADD(T9o, T9n); Chris@10: { Chris@10: V T8Y, T91, T9k, T9l; Chris@10: T8Y = VFMA(LDK(KP098017140), T8W, VMUL(LDK(KP995184726), T8X)); Chris@10: T91 = VFNMS(LDK(KP098017140), T90, VMUL(LDK(KP995184726), T8Z)); Chris@10: T92 = VSUB(T8Y, T91); Chris@10: T9Q = VADD(T8Y, T91); Chris@10: T9k = VFNMS(LDK(KP098017140), T8X, VMUL(LDK(KP995184726), T8W)); Chris@10: T9l = VFMA(LDK(KP995184726), T90, VMUL(LDK(KP098017140), T8Z)); Chris@10: T9m = VSUB(T9k, T9l); Chris@10: T9G = VADD(T9k, T9l); Chris@10: } Chris@10: T93 = VSUB(T8V, T92); Chris@10: Ta0 = VADD(T9R, T9Q); Chris@10: Ta2 = VADD(T9F, T9G); Chris@10: T9q = VSUB(T9m, T9p); Chris@10: T9y = VADD(T9p, T9m); Chris@10: T9H = VSUB(T9F, T9G); Chris@10: T9S = VSUB(T9Q, T9R); Chris@10: T9A = VADD(T8V, T92); Chris@10: } Chris@10: { Chris@10: V T9j, T9u, Ta1, Ta4; Chris@10: T9j = VADD(T93, T9i); Chris@10: T9u = VBYI(VADD(T9q, T9t)); Chris@10: ST(&(xo[WS(os, 111)]), VSUB(T9j, T9u), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 17)]), VADD(T9j, T9u), ovs, &(xo[WS(os, 1)])); Chris@10: Ta1 = VBYI(VSUB(T9Z, Ta0)); Chris@10: Ta4 = VSUB(Ta2, Ta3); Chris@10: ST(&(xo[WS(os, 63)]), VADD(Ta1, Ta4), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 65)]), VSUB(Ta4, Ta1), ovs, &(xo[WS(os, 1)])); Chris@10: } Chris@10: { Chris@10: V Ta5, Ta6, T9v, T9w; Chris@10: Ta5 = VBYI(VADD(Ta0, T9Z)); Chris@10: Ta6 = VADD(Ta2, Ta3); Chris@10: ST(&(xo[WS(os, 1)]), VADD(Ta5, Ta6), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 127)]), VSUB(Ta6, Ta5), ovs, &(xo[WS(os, 1)])); Chris@10: T9v = VSUB(T93, T9i); Chris@10: T9w = VBYI(VSUB(T9t, T9q)); Chris@10: ST(&(xo[WS(os, 81)]), VSUB(T9v, T9w), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 47)]), VADD(T9v, T9w), ovs, &(xo[WS(os, 1)])); Chris@10: } Chris@10: { Chris@10: V T9z, T9C, T9P, T9W; Chris@10: T9z = VBYI(VSUB(T9x, T9y)); Chris@10: T9C = VSUB(T9A, T9B); Chris@10: ST(&(xo[WS(os, 49)]), VADD(T9z, T9C), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 79)]), VSUB(T9C, T9z), ovs, &(xo[WS(os, 1)])); Chris@10: T9P = VADD(T9H, T9O); Chris@10: T9W = VBYI(VADD(T9S, T9V)); Chris@10: ST(&(xo[WS(os, 97)]), VSUB(T9P, T9W), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 31)]), VADD(T9P, T9W), ovs, &(xo[WS(os, 1)])); Chris@10: } Chris@10: { Chris@10: V T9X, T9Y, T9D, T9E; Chris@10: T9X = VSUB(T9H, T9O); Chris@10: T9Y = VBYI(VSUB(T9V, T9S)); Chris@10: ST(&(xo[WS(os, 95)]), VSUB(T9X, T9Y), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 33)]), VADD(T9X, T9Y), ovs, &(xo[WS(os, 1)])); Chris@10: T9D = VBYI(VADD(T9y, T9x)); Chris@10: T9E = VADD(T9A, T9B); Chris@10: ST(&(xo[WS(os, 15)]), VADD(T9D, T9E), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 113)]), VSUB(T9E, T9D), ovs, &(xo[WS(os, 1)])); Chris@10: } Chris@10: } Chris@10: { Chris@10: V T68, T6r, T6j, T6n, T6E, T6T, T6L, T6P, T5T, T6Q, T6S, T6g, T6o, T6x, T6I; Chris@10: V T6q; Chris@10: { Chris@10: V T60, T6h, T67, T6i; Chris@10: { Chris@10: V T5W, T5Z, T63, T66; Chris@10: T5W = VSUB(T5U, T5V); Chris@10: T5Z = VSUB(T5X, T5Y); Chris@10: T60 = VFMA(LDK(KP803207531), T5W, VMUL(LDK(KP595699304), T5Z)); Chris@10: T6h = VFNMS(LDK(KP595699304), T5W, VMUL(LDK(KP803207531), T5Z)); Chris@10: T63 = VSUB(T61, T62); Chris@10: T66 = VSUB(T64, T65); Chris@10: T67 = VFNMS(LDK(KP595699304), T66, VMUL(LDK(KP803207531), T63)); Chris@10: T6i = VFMA(LDK(KP595699304), T63, VMUL(LDK(KP803207531), T66)); Chris@10: } Chris@10: T68 = VSUB(T60, T67); Chris@10: T6r = VADD(T6h, T6i); Chris@10: T6j = VSUB(T6h, T6i); Chris@10: T6n = VADD(T60, T67); Chris@10: } Chris@10: { Chris@10: V T6A, T6J, T6D, T6K; Chris@10: { Chris@10: V T6y, T6z, T6B, T6C; Chris@10: T6y = VADD(T5V, T5U); Chris@10: T6z = VADD(T5X, T5Y); Chris@10: T6A = VFMA(LDK(KP989176509), T6y, VMUL(LDK(KP146730474), T6z)); Chris@10: T6J = VFNMS(LDK(KP146730474), T6y, VMUL(LDK(KP989176509), T6z)); Chris@10: T6B = VADD(T62, T61); Chris@10: T6C = VADD(T64, T65); Chris@10: T6D = VFNMS(LDK(KP146730474), T6C, VMUL(LDK(KP989176509), T6B)); Chris@10: T6K = VFMA(LDK(KP146730474), T6B, VMUL(LDK(KP989176509), T6C)); Chris@10: } Chris@10: T6E = VSUB(T6A, T6D); Chris@10: T6T = VADD(T6J, T6K); Chris@10: T6L = VSUB(T6J, T6K); Chris@10: T6P = VADD(T6A, T6D); Chris@10: } Chris@10: { Chris@10: V T5L, T6v, T6f, T6H, T5S, T6G, T6c, T6w, T5K, T6d; Chris@10: T5K = VADD(T4q, T4H); Chris@10: T5L = VSUB(T5J, T5K); Chris@10: T6v = VADD(T5J, T5K); Chris@10: T6d = VADD(T4V, T4W); Chris@10: T6f = VSUB(T6d, T6e); Chris@10: T6H = VADD(T6e, T6d); Chris@10: { Chris@10: V T5O, T5R, T6a, T6b; Chris@10: T5O = VFMA(LDK(KP956940335), T5M, VMUL(LDK(KP290284677), T5N)); Chris@10: T5R = VFNMS(LDK(KP290284677), T5Q, VMUL(LDK(KP956940335), T5P)); Chris@10: T5S = VSUB(T5O, T5R); Chris@10: T6G = VADD(T5O, T5R); Chris@10: T6a = VFNMS(LDK(KP290284677), T5M, VMUL(LDK(KP956940335), T5N)); Chris@10: T6b = VFMA(LDK(KP290284677), T5P, VMUL(LDK(KP956940335), T5Q)); Chris@10: T6c = VSUB(T6a, T6b); Chris@10: T6w = VADD(T6a, T6b); Chris@10: } Chris@10: T5T = VSUB(T5L, T5S); Chris@10: T6Q = VADD(T6H, T6G); Chris@10: T6S = VADD(T6v, T6w); Chris@10: T6g = VSUB(T6c, T6f); Chris@10: T6o = VADD(T6f, T6c); Chris@10: T6x = VSUB(T6v, T6w); Chris@10: T6I = VSUB(T6G, T6H); Chris@10: T6q = VADD(T5L, T5S); Chris@10: } Chris@10: { Chris@10: V T69, T6k, T6R, T6U; Chris@10: T69 = VADD(T5T, T68); Chris@10: T6k = VBYI(VADD(T6g, T6j)); Chris@10: ST(&(xo[WS(os, 109)]), VSUB(T69, T6k), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 19)]), VADD(T69, T6k), ovs, &(xo[WS(os, 1)])); Chris@10: T6R = VBYI(VSUB(T6P, T6Q)); Chris@10: T6U = VSUB(T6S, T6T); Chris@10: ST(&(xo[WS(os, 61)]), VADD(T6R, T6U), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 67)]), VSUB(T6U, T6R), ovs, &(xo[WS(os, 1)])); Chris@10: } Chris@10: { Chris@10: V T6V, T6W, T6l, T6m; Chris@10: T6V = VBYI(VADD(T6Q, T6P)); Chris@10: T6W = VADD(T6S, T6T); Chris@10: ST(&(xo[WS(os, 3)]), VADD(T6V, T6W), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 125)]), VSUB(T6W, T6V), ovs, &(xo[WS(os, 1)])); Chris@10: T6l = VSUB(T5T, T68); Chris@10: T6m = VBYI(VSUB(T6j, T6g)); Chris@10: ST(&(xo[WS(os, 83)]), VSUB(T6l, T6m), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 45)]), VADD(T6l, T6m), ovs, &(xo[WS(os, 1)])); Chris@10: } Chris@10: { Chris@10: V T6p, T6s, T6F, T6M; Chris@10: T6p = VBYI(VSUB(T6n, T6o)); Chris@10: T6s = VSUB(T6q, T6r); Chris@10: ST(&(xo[WS(os, 51)]), VADD(T6p, T6s), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 77)]), VSUB(T6s, T6p), ovs, &(xo[WS(os, 1)])); Chris@10: T6F = VADD(T6x, T6E); Chris@10: T6M = VBYI(VADD(T6I, T6L)); Chris@10: ST(&(xo[WS(os, 99)]), VSUB(T6F, T6M), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 29)]), VADD(T6F, T6M), ovs, &(xo[WS(os, 1)])); Chris@10: } Chris@10: { Chris@10: V T6N, T6O, T6t, T6u; Chris@10: T6N = VSUB(T6x, T6E); Chris@10: T6O = VBYI(VSUB(T6L, T6I)); Chris@10: ST(&(xo[WS(os, 93)]), VSUB(T6N, T6O), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 35)]), VADD(T6N, T6O), ovs, &(xo[WS(os, 1)])); Chris@10: T6t = VBYI(VADD(T6o, T6n)); Chris@10: T6u = VADD(T6q, T6r); Chris@10: ST(&(xo[WS(os, 13)]), VADD(T6t, T6u), ovs, &(xo[WS(os, 1)])); Chris@10: ST(&(xo[WS(os, 115)]), VSUB(T6u, T6t), ovs, &(xo[WS(os, 1)])); Chris@10: } Chris@10: } Chris@10: } Chris@10: } Chris@10: VLEAVE(); Chris@10: } Chris@10: Chris@10: static const kdft_desc desc = { 128, XSIMD_STRING("n1bv_128"), {938, 186, 144, 0}, &GENUS, 0, 0, 0, 0 }; Chris@10: Chris@10: void XSIMD(codelet_n1bv_128) (planner *p) { Chris@10: X(kdft_register) (p, n1bv_128, &desc); Chris@10: } Chris@10: Chris@10: #endif /* HAVE_FMA */