Chris@10
|
1 /*
|
Chris@10
|
2 * Copyright (c) 2003, 2007-11 Matteo Frigo
|
Chris@10
|
3 * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology
|
Chris@10
|
4 *
|
Chris@10
|
5 * This program is free software; you can redistribute it and/or modify
|
Chris@10
|
6 * it under the terms of the GNU General Public License as published by
|
Chris@10
|
7 * the Free Software Foundation; either version 2 of the License, or
|
Chris@10
|
8 * (at your option) any later version.
|
Chris@10
|
9 *
|
Chris@10
|
10 * This program is distributed in the hope that it will be useful,
|
Chris@10
|
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
Chris@10
|
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
Chris@10
|
13 * GNU General Public License for more details.
|
Chris@10
|
14 *
|
Chris@10
|
15 * You should have received a copy of the GNU General Public License
|
Chris@10
|
16 * along with this program; if not, write to the Free Software
|
Chris@10
|
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
Chris@10
|
18 *
|
Chris@10
|
19 */
|
Chris@10
|
20
|
Chris@10
|
21 /* This file was automatically generated --- DO NOT EDIT */
|
Chris@10
|
22 /* Generated on Sun Nov 25 07:41:09 EST 2012 */
|
Chris@10
|
23
|
Chris@10
|
24 #include "codelet-rdft.h"
|
Chris@10
|
25
|
Chris@10
|
26 #ifdef HAVE_FMA
|
Chris@10
|
27
|
Chris@10
|
28 /* Generated by: ../../../genfft/gen_r2cb.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -sign 1 -n 128 -name r2cb_128 -include r2cb.h */
|
Chris@10
|
29
|
Chris@10
|
30 /*
|
Chris@10
|
31 * This function contains 956 FP additions, 540 FP multiplications,
|
Chris@10
|
32 * (or, 416 additions, 0 multiplications, 540 fused multiply/add),
|
Chris@10
|
33 * 242 stack variables, 36 constants, and 256 memory accesses
|
Chris@10
|
34 */
|
Chris@10
|
35 #include "r2cb.h"
|
Chris@10
|
36
|
Chris@10
|
37 static void r2cb_128(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
|
Chris@10
|
38 {
|
Chris@10
|
39 DK(KP1_715457220, +1.715457220000544139804539968569540274084981599);
|
Chris@10
|
40 DK(KP1_606415062, +1.606415062961289819613353025926283847759138854);
|
Chris@10
|
41 DK(KP599376933, +0.599376933681923766271389869014404232837890546);
|
Chris@10
|
42 DK(KP741650546, +0.741650546272035369581266691172079863842265220);
|
Chris@10
|
43 DK(KP1_978353019, +1.978353019929561946903347476032486127967379067);
|
Chris@10
|
44 DK(KP1_940062506, +1.940062506389087985207968414572200502913731924);
|
Chris@10
|
45 DK(KP148335987, +0.148335987538347428753676511486911367000625355);
|
Chris@10
|
46 DK(KP250486960, +0.250486960191305461595702160124721208578685568);
|
Chris@10
|
47 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
Chris@10
|
48 DK(KP1_807978586, +1.807978586246886663172400594461074097420264050);
|
Chris@10
|
49 DK(KP1_481902250, +1.481902250709918182351233794990325459457910619);
|
Chris@10
|
50 DK(KP472964775, +0.472964775891319928124438237972992463904131113);
|
Chris@10
|
51 DK(KP906347169, +0.906347169019147157946142717268914412664134293);
|
Chris@10
|
52 DK(KP1_997590912, +1.997590912410344785429543209518201388886407229);
|
Chris@10
|
53 DK(KP1_883088130, +1.883088130366041556825018805199004714371179592);
|
Chris@10
|
54 DK(KP049126849, +0.049126849769467254105343321271313617079695752);
|
Chris@10
|
55 DK(KP357805721, +0.357805721314524104672487743774474392487532769);
|
Chris@10
|
56 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
Chris@10
|
57 DK(KP1_763842528, +1.763842528696710059425513727320776699016885241);
|
Chris@10
|
58 DK(KP1_913880671, +1.913880671464417729871595773960539938965698411);
|
Chris@10
|
59 DK(KP534511135, +0.534511135950791641089685961295362908582039528);
|
Chris@10
|
60 DK(KP303346683, +0.303346683607342391675883946941299872384187453);
|
Chris@10
|
61 DK(KP1_662939224, +1.662939224605090474157576755235811513477121624);
|
Chris@10
|
62 DK(KP668178637, +0.668178637919298919997757686523080761552472251);
|
Chris@10
|
63 DK(KP1_990369453, +1.990369453344393772489673906218959843150949737);
|
Chris@10
|
64 DK(KP1_546020906, +1.546020906725473921621813219516939601942082586);
|
Chris@10
|
65 DK(KP098491403, +0.098491403357164253077197521291327432293052451);
|
Chris@10
|
66 DK(KP820678790, +0.820678790828660330972281985331011598767386482);
|
Chris@10
|
67 DK(KP1_961570560, +1.961570560806460898252364472268478073947867462);
|
Chris@10
|
68 DK(KP198912367, +0.198912367379658006911597622644676228597850501);
|
Chris@10
|
69 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
Chris@10
|
70 DK(KP1_847759065, +1.847759065022573512256366378793576573644833252);
|
Chris@10
|
71 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
Chris@10
|
72 DK(KP414213562, +0.414213562373095048801688724209698078569671875);
|
Chris@10
|
73 DK(KP1_414213562, +1.414213562373095048801688724209698078569671875);
|
Chris@10
|
74 DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
|
Chris@10
|
75 {
|
Chris@10
|
76 INT i;
|
Chris@10
|
77 for (i = v; i > 0; i = i - 1, R0 = R0 + ovs, R1 = R1 + ovs, Cr = Cr + ivs, Ci = Ci + ivs, MAKE_VOLATILE_STRIDE(512, rs), MAKE_VOLATILE_STRIDE(512, csr), MAKE_VOLATILE_STRIDE(512, csi)) {
|
Chris@10
|
78 E T9H, T9I, T9X, T9Y;
|
Chris@10
|
79 {
|
Chris@10
|
80 E Tdr, T9, Tcl, Ta9, T6b, T2d, T91, T7j, Tg, Tds, Tcm, Tae, T92, T7m, T6c;
|
Chris@10
|
81 E T2o, Tdu, Tw, Tco, Tap, TeM, Tdx, T6f, T2G, T6e, T2P, T94, T7t, Tcp, Tak;
|
Chris@10
|
82 E T95, T7q, TdM, T1i, TcL, TbD, Tf0, Te6, T6q, T42, T6B, T5t, T9r, T8j, TcA;
|
Chris@10
|
83 E TaY, T9g, T7S, TdA, TM, Tcv, TaN, TeP, TdI, T6i, T38, T6l, T3F, T9b, T7J;
|
Chris@10
|
84 E Tcs, Taw, T98, T7y, T1N, TeW, T6x, T4H, Te8, TdV, T6w, T4Q, T9j, T86, TcO;
|
Chris@10
|
85 E TcI, T9k, T83, TbI, Tbl, T22, TeV, Te0, Te9, T58, T6u, T6t, T5h, T9m, T8d;
|
Chris@10
|
86 E TcP, TcF, T9n, T8a, TbJ, Tbw, Te3, T1x, TcB, TbG, Tf1, TdP, T6C, T4p, T6r;
|
Chris@10
|
87 E T5w, T9h, T8m, TcM, Tb9, T9s, T7Z, TaB, TaG, TdF, T11, Tct, TaQ, TeQ, TdD;
|
Chris@10
|
88 E T6m, T3v, T7B, T7E, T6j, T3I, T99, T7M;
|
Chris@10
|
89 {
|
Chris@10
|
90 E TaU, TaX, T7Q, T7R, Tbk, Tbf;
|
Chris@10
|
91 {
|
Chris@10
|
92 E Td, T2e, Tc, Tab, T2m, Te, T2f, T2g;
|
Chris@10
|
93 {
|
Chris@10
|
94 E T7h, T27, T2c, T7i;
|
Chris@10
|
95 {
|
Chris@10
|
96 E T4, T26, T29, T25, T3, T28, T8, T2a;
|
Chris@10
|
97 T4 = Cr[WS(csr, 32)];
|
Chris@10
|
98 T26 = Ci[WS(csi, 32)];
|
Chris@10
|
99 {
|
Chris@10
|
100 E T1, T2, T6, T7;
|
Chris@10
|
101 T1 = Cr[0];
|
Chris@10
|
102 T2 = Cr[WS(csr, 64)];
|
Chris@10
|
103 T6 = Cr[WS(csr, 16)];
|
Chris@10
|
104 T7 = Cr[WS(csr, 48)];
|
Chris@10
|
105 T29 = Ci[WS(csi, 16)];
|
Chris@10
|
106 T25 = T1 - T2;
|
Chris@10
|
107 T3 = T1 + T2;
|
Chris@10
|
108 T28 = T6 - T7;
|
Chris@10
|
109 T8 = T6 + T7;
|
Chris@10
|
110 T2a = Ci[WS(csi, 48)];
|
Chris@10
|
111 }
|
Chris@10
|
112 {
|
Chris@10
|
113 E Ta7, T5, Ta8, T2b;
|
Chris@10
|
114 Ta7 = FNMS(KP2_000000000, T4, T3);
|
Chris@10
|
115 T5 = FMA(KP2_000000000, T4, T3);
|
Chris@10
|
116 T7h = FMA(KP2_000000000, T26, T25);
|
Chris@10
|
117 T27 = FNMS(KP2_000000000, T26, T25);
|
Chris@10
|
118 Ta8 = T29 - T2a;
|
Chris@10
|
119 T2b = T29 + T2a;
|
Chris@10
|
120 Tdr = FNMS(KP2_000000000, T8, T5);
|
Chris@10
|
121 T9 = FMA(KP2_000000000, T8, T5);
|
Chris@10
|
122 Tcl = FMA(KP2_000000000, Ta8, Ta7);
|
Chris@10
|
123 Ta9 = FNMS(KP2_000000000, Ta8, Ta7);
|
Chris@10
|
124 T2c = T28 - T2b;
|
Chris@10
|
125 T7i = T28 + T2b;
|
Chris@10
|
126 }
|
Chris@10
|
127 }
|
Chris@10
|
128 {
|
Chris@10
|
129 E Ta, Tb, T2k, T2l;
|
Chris@10
|
130 Ta = Cr[WS(csr, 8)];
|
Chris@10
|
131 T6b = FNMS(KP1_414213562, T2c, T27);
|
Chris@10
|
132 T2d = FMA(KP1_414213562, T2c, T27);
|
Chris@10
|
133 T91 = FMA(KP1_414213562, T7i, T7h);
|
Chris@10
|
134 T7j = FNMS(KP1_414213562, T7i, T7h);
|
Chris@10
|
135 Tb = Cr[WS(csr, 56)];
|
Chris@10
|
136 T2k = Ci[WS(csi, 8)];
|
Chris@10
|
137 T2l = Ci[WS(csi, 56)];
|
Chris@10
|
138 Td = Cr[WS(csr, 40)];
|
Chris@10
|
139 T2e = Ta - Tb;
|
Chris@10
|
140 Tc = Ta + Tb;
|
Chris@10
|
141 Tab = T2k - T2l;
|
Chris@10
|
142 T2m = T2k + T2l;
|
Chris@10
|
143 Te = Cr[WS(csr, 24)];
|
Chris@10
|
144 T2f = Ci[WS(csi, 40)];
|
Chris@10
|
145 T2g = Ci[WS(csi, 24)];
|
Chris@10
|
146 }
|
Chris@10
|
147 }
|
Chris@10
|
148 {
|
Chris@10
|
149 E Tag, Taj, T7o, T7p;
|
Chris@10
|
150 {
|
Chris@10
|
151 E T2q, Tk, Tam, T2K, T2H, Tn, Tan, T2t, Tu, Tah, T2E, T2N, Tr, T2v, T2y;
|
Chris@10
|
152 E Tai;
|
Chris@10
|
153 {
|
Chris@10
|
154 E Tl, Tm, T2r, T2s;
|
Chris@10
|
155 {
|
Chris@10
|
156 E Ti, Tj, T2j, Tf, T2I, T2J;
|
Chris@10
|
157 Ti = Cr[WS(csr, 4)];
|
Chris@10
|
158 T2j = Td - Te;
|
Chris@10
|
159 Tf = Td + Te;
|
Chris@10
|
160 {
|
Chris@10
|
161 E Tac, T2h, T7k, T2n;
|
Chris@10
|
162 Tac = T2f - T2g;
|
Chris@10
|
163 T2h = T2f + T2g;
|
Chris@10
|
164 T7k = T2m - T2j;
|
Chris@10
|
165 T2n = T2j + T2m;
|
Chris@10
|
166 {
|
Chris@10
|
167 E Taa, Tad, T7l, T2i;
|
Chris@10
|
168 Taa = Tc - Tf;
|
Chris@10
|
169 Tg = Tc + Tf;
|
Chris@10
|
170 Tad = Tab - Tac;
|
Chris@10
|
171 Tds = Tac + Tab;
|
Chris@10
|
172 T7l = T2e + T2h;
|
Chris@10
|
173 T2i = T2e - T2h;
|
Chris@10
|
174 Tcm = Taa + Tad;
|
Chris@10
|
175 Tae = Taa - Tad;
|
Chris@10
|
176 T92 = FMA(KP414213562, T7k, T7l);
|
Chris@10
|
177 T7m = FNMS(KP414213562, T7l, T7k);
|
Chris@10
|
178 T6c = FMA(KP414213562, T2i, T2n);
|
Chris@10
|
179 T2o = FNMS(KP414213562, T2n, T2i);
|
Chris@10
|
180 Tj = Cr[WS(csr, 60)];
|
Chris@10
|
181 }
|
Chris@10
|
182 }
|
Chris@10
|
183 T2I = Ci[WS(csi, 4)];
|
Chris@10
|
184 T2J = Ci[WS(csi, 60)];
|
Chris@10
|
185 Tl = Cr[WS(csr, 36)];
|
Chris@10
|
186 T2q = Ti - Tj;
|
Chris@10
|
187 Tk = Ti + Tj;
|
Chris@10
|
188 Tam = T2I - T2J;
|
Chris@10
|
189 T2K = T2I + T2J;
|
Chris@10
|
190 Tm = Cr[WS(csr, 28)];
|
Chris@10
|
191 }
|
Chris@10
|
192 T2r = Ci[WS(csi, 36)];
|
Chris@10
|
193 T2s = Ci[WS(csi, 28)];
|
Chris@10
|
194 {
|
Chris@10
|
195 E Ts, Tt, T2B, T2C;
|
Chris@10
|
196 Ts = Cr[WS(csr, 12)];
|
Chris@10
|
197 T2H = Tl - Tm;
|
Chris@10
|
198 Tn = Tl + Tm;
|
Chris@10
|
199 Tan = T2r - T2s;
|
Chris@10
|
200 T2t = T2r + T2s;
|
Chris@10
|
201 Tt = Cr[WS(csr, 52)];
|
Chris@10
|
202 T2B = Ci[WS(csi, 12)];
|
Chris@10
|
203 T2C = Ci[WS(csi, 52)];
|
Chris@10
|
204 {
|
Chris@10
|
205 E Tp, T2A, T2D, Tq, T2w, T2x;
|
Chris@10
|
206 Tp = Cr[WS(csr, 20)];
|
Chris@10
|
207 Tu = Ts + Tt;
|
Chris@10
|
208 T2A = Ts - Tt;
|
Chris@10
|
209 Tah = T2C - T2B;
|
Chris@10
|
210 T2D = T2B + T2C;
|
Chris@10
|
211 Tq = Cr[WS(csr, 44)];
|
Chris@10
|
212 T2w = Ci[WS(csi, 20)];
|
Chris@10
|
213 T2x = Ci[WS(csi, 44)];
|
Chris@10
|
214 T2E = T2A - T2D;
|
Chris@10
|
215 T2N = T2A + T2D;
|
Chris@10
|
216 Tr = Tp + Tq;
|
Chris@10
|
217 T2v = Tp - Tq;
|
Chris@10
|
218 T2y = T2w + T2x;
|
Chris@10
|
219 Tai = T2w - T2x;
|
Chris@10
|
220 }
|
Chris@10
|
221 }
|
Chris@10
|
222 }
|
Chris@10
|
223 {
|
Chris@10
|
224 E T2M, Tdv, Tdw, T2u, T2F, T7s, T7r, T2L, T2O;
|
Chris@10
|
225 {
|
Chris@10
|
226 E To, T2z, Tv, Tal, Tao;
|
Chris@10
|
227 Tag = Tk - Tn;
|
Chris@10
|
228 To = Tk + Tn;
|
Chris@10
|
229 T2M = T2v + T2y;
|
Chris@10
|
230 T2z = T2v - T2y;
|
Chris@10
|
231 Tv = Tr + Tu;
|
Chris@10
|
232 Tal = Tr - Tu;
|
Chris@10
|
233 Tao = Tam - Tan;
|
Chris@10
|
234 Tdv = Tan + Tam;
|
Chris@10
|
235 Tdu = To - Tv;
|
Chris@10
|
236 Tw = To + Tv;
|
Chris@10
|
237 Tco = Tao - Tal;
|
Chris@10
|
238 Tap = Tal + Tao;
|
Chris@10
|
239 Tdw = Tai + Tah;
|
Chris@10
|
240 Taj = Tah - Tai;
|
Chris@10
|
241 T7o = T2q + T2t;
|
Chris@10
|
242 T2u = T2q - T2t;
|
Chris@10
|
243 T2F = T2z + T2E;
|
Chris@10
|
244 T7s = T2E - T2z;
|
Chris@10
|
245 }
|
Chris@10
|
246 T7r = T2K - T2H;
|
Chris@10
|
247 T2L = T2H + T2K;
|
Chris@10
|
248 TeM = Tdw + Tdv;
|
Chris@10
|
249 Tdx = Tdv - Tdw;
|
Chris@10
|
250 T6f = FNMS(KP707106781, T2F, T2u);
|
Chris@10
|
251 T2G = FMA(KP707106781, T2F, T2u);
|
Chris@10
|
252 T2O = T2M - T2N;
|
Chris@10
|
253 T7p = T2M + T2N;
|
Chris@10
|
254 T6e = FNMS(KP707106781, T2O, T2L);
|
Chris@10
|
255 T2P = FMA(KP707106781, T2O, T2L);
|
Chris@10
|
256 T94 = FMA(KP707106781, T7s, T7r);
|
Chris@10
|
257 T7t = FNMS(KP707106781, T7s, T7r);
|
Chris@10
|
258 }
|
Chris@10
|
259 }
|
Chris@10
|
260 {
|
Chris@10
|
261 E T3M, T16, TbA, T5o, T5l, T19, TbB, T3P, T1g, TaV, T40, T5r, T1d, T3R, T3U;
|
Chris@10
|
262 E TaW;
|
Chris@10
|
263 {
|
Chris@10
|
264 E T17, T18, T3N, T3O;
|
Chris@10
|
265 {
|
Chris@10
|
266 E T14, T15, T5m, T5n;
|
Chris@10
|
267 T14 = Cr[WS(csr, 1)];
|
Chris@10
|
268 Tcp = Tag - Taj;
|
Chris@10
|
269 Tak = Tag + Taj;
|
Chris@10
|
270 T95 = FMA(KP707106781, T7p, T7o);
|
Chris@10
|
271 T7q = FNMS(KP707106781, T7p, T7o);
|
Chris@10
|
272 T15 = Cr[WS(csr, 63)];
|
Chris@10
|
273 T5m = Ci[WS(csi, 1)];
|
Chris@10
|
274 T5n = Ci[WS(csi, 63)];
|
Chris@10
|
275 T17 = Cr[WS(csr, 33)];
|
Chris@10
|
276 T3M = T14 - T15;
|
Chris@10
|
277 T16 = T14 + T15;
|
Chris@10
|
278 TbA = T5m - T5n;
|
Chris@10
|
279 T5o = T5m + T5n;
|
Chris@10
|
280 T18 = Cr[WS(csr, 31)];
|
Chris@10
|
281 }
|
Chris@10
|
282 T3N = Ci[WS(csi, 33)];
|
Chris@10
|
283 T3O = Ci[WS(csi, 31)];
|
Chris@10
|
284 {
|
Chris@10
|
285 E T1e, T1f, T3X, T3Y;
|
Chris@10
|
286 T1e = Cr[WS(csr, 15)];
|
Chris@10
|
287 T5l = T17 - T18;
|
Chris@10
|
288 T19 = T17 + T18;
|
Chris@10
|
289 TbB = T3N - T3O;
|
Chris@10
|
290 T3P = T3N + T3O;
|
Chris@10
|
291 T1f = Cr[WS(csr, 49)];
|
Chris@10
|
292 T3X = Ci[WS(csi, 15)];
|
Chris@10
|
293 T3Y = Ci[WS(csi, 49)];
|
Chris@10
|
294 {
|
Chris@10
|
295 E T1b, T3W, T3Z, T1c, T3S, T3T;
|
Chris@10
|
296 T1b = Cr[WS(csr, 17)];
|
Chris@10
|
297 T1g = T1e + T1f;
|
Chris@10
|
298 T3W = T1e - T1f;
|
Chris@10
|
299 TaV = T3Y - T3X;
|
Chris@10
|
300 T3Z = T3X + T3Y;
|
Chris@10
|
301 T1c = Cr[WS(csr, 47)];
|
Chris@10
|
302 T3S = Ci[WS(csi, 17)];
|
Chris@10
|
303 T3T = Ci[WS(csi, 47)];
|
Chris@10
|
304 T40 = T3W - T3Z;
|
Chris@10
|
305 T5r = T3W + T3Z;
|
Chris@10
|
306 T1d = T1b + T1c;
|
Chris@10
|
307 T3R = T1b - T1c;
|
Chris@10
|
308 T3U = T3S + T3T;
|
Chris@10
|
309 TaW = T3S - T3T;
|
Chris@10
|
310 }
|
Chris@10
|
311 }
|
Chris@10
|
312 }
|
Chris@10
|
313 {
|
Chris@10
|
314 E T5q, Te4, Te5, T3Q, T41, T8i, T8h, T5p, T5s;
|
Chris@10
|
315 {
|
Chris@10
|
316 E T1a, T3V, T1h, Tbz, TbC;
|
Chris@10
|
317 TaU = T16 - T19;
|
Chris@10
|
318 T1a = T16 + T19;
|
Chris@10
|
319 T5q = T3R + T3U;
|
Chris@10
|
320 T3V = T3R - T3U;
|
Chris@10
|
321 T1h = T1d + T1g;
|
Chris@10
|
322 Tbz = T1d - T1g;
|
Chris@10
|
323 TbC = TbA - TbB;
|
Chris@10
|
324 Te4 = TbB + TbA;
|
Chris@10
|
325 TdM = T1a - T1h;
|
Chris@10
|
326 T1i = T1a + T1h;
|
Chris@10
|
327 TcL = TbC - Tbz;
|
Chris@10
|
328 TbD = Tbz + TbC;
|
Chris@10
|
329 Te5 = TaW + TaV;
|
Chris@10
|
330 TaX = TaV - TaW;
|
Chris@10
|
331 T7Q = T3M + T3P;
|
Chris@10
|
332 T3Q = T3M - T3P;
|
Chris@10
|
333 T41 = T3V + T40;
|
Chris@10
|
334 T8i = T40 - T3V;
|
Chris@10
|
335 }
|
Chris@10
|
336 T8h = T5o - T5l;
|
Chris@10
|
337 T5p = T5l + T5o;
|
Chris@10
|
338 Tf0 = Te5 + Te4;
|
Chris@10
|
339 Te6 = Te4 - Te5;
|
Chris@10
|
340 T6q = FNMS(KP707106781, T41, T3Q);
|
Chris@10
|
341 T42 = FMA(KP707106781, T41, T3Q);
|
Chris@10
|
342 T5s = T5q - T5r;
|
Chris@10
|
343 T7R = T5q + T5r;
|
Chris@10
|
344 T6B = FNMS(KP707106781, T5s, T5p);
|
Chris@10
|
345 T5t = FMA(KP707106781, T5s, T5p);
|
Chris@10
|
346 T9r = FMA(KP707106781, T8i, T8h);
|
Chris@10
|
347 T8j = FNMS(KP707106781, T8i, T8h);
|
Chris@10
|
348 }
|
Chris@10
|
349 }
|
Chris@10
|
350 }
|
Chris@10
|
351 }
|
Chris@10
|
352 {
|
Chris@10
|
353 E Tas, Tav, T7w, T7x;
|
Chris@10
|
354 {
|
Chris@10
|
355 E T2S, TA, TaK, T3A, T3x, TD, TaL, T2V, TK, Tat, T36, T3D, TH, T2X, T30;
|
Chris@10
|
356 E Tau;
|
Chris@10
|
357 {
|
Chris@10
|
358 E TB, TC, T2T, T2U;
|
Chris@10
|
359 {
|
Chris@10
|
360 E Ty, Tz, T3y, T3z;
|
Chris@10
|
361 Ty = Cr[WS(csr, 2)];
|
Chris@10
|
362 TcA = TaU - TaX;
|
Chris@10
|
363 TaY = TaU + TaX;
|
Chris@10
|
364 T9g = FMA(KP707106781, T7R, T7Q);
|
Chris@10
|
365 T7S = FNMS(KP707106781, T7R, T7Q);
|
Chris@10
|
366 Tz = Cr[WS(csr, 62)];
|
Chris@10
|
367 T3y = Ci[WS(csi, 2)];
|
Chris@10
|
368 T3z = Ci[WS(csi, 62)];
|
Chris@10
|
369 TB = Cr[WS(csr, 34)];
|
Chris@10
|
370 T2S = Ty - Tz;
|
Chris@10
|
371 TA = Ty + Tz;
|
Chris@10
|
372 TaK = T3y - T3z;
|
Chris@10
|
373 T3A = T3y + T3z;
|
Chris@10
|
374 TC = Cr[WS(csr, 30)];
|
Chris@10
|
375 }
|
Chris@10
|
376 T2T = Ci[WS(csi, 34)];
|
Chris@10
|
377 T2U = Ci[WS(csi, 30)];
|
Chris@10
|
378 {
|
Chris@10
|
379 E TI, TJ, T33, T34;
|
Chris@10
|
380 TI = Cr[WS(csr, 14)];
|
Chris@10
|
381 T3x = TB - TC;
|
Chris@10
|
382 TD = TB + TC;
|
Chris@10
|
383 TaL = T2T - T2U;
|
Chris@10
|
384 T2V = T2T + T2U;
|
Chris@10
|
385 TJ = Cr[WS(csr, 50)];
|
Chris@10
|
386 T33 = Ci[WS(csi, 14)];
|
Chris@10
|
387 T34 = Ci[WS(csi, 50)];
|
Chris@10
|
388 {
|
Chris@10
|
389 E TF, T32, T35, TG, T2Y, T2Z;
|
Chris@10
|
390 TF = Cr[WS(csr, 18)];
|
Chris@10
|
391 TK = TI + TJ;
|
Chris@10
|
392 T32 = TI - TJ;
|
Chris@10
|
393 Tat = T34 - T33;
|
Chris@10
|
394 T35 = T33 + T34;
|
Chris@10
|
395 TG = Cr[WS(csr, 46)];
|
Chris@10
|
396 T2Y = Ci[WS(csi, 18)];
|
Chris@10
|
397 T2Z = Ci[WS(csi, 46)];
|
Chris@10
|
398 T36 = T32 - T35;
|
Chris@10
|
399 T3D = T32 + T35;
|
Chris@10
|
400 TH = TF + TG;
|
Chris@10
|
401 T2X = TF - TG;
|
Chris@10
|
402 T30 = T2Y + T2Z;
|
Chris@10
|
403 Tau = T2Y - T2Z;
|
Chris@10
|
404 }
|
Chris@10
|
405 }
|
Chris@10
|
406 }
|
Chris@10
|
407 {
|
Chris@10
|
408 E T3C, TdG, TdH, T2W, T37, T7I, T7H, T3B, T3E;
|
Chris@10
|
409 {
|
Chris@10
|
410 E TE, T31, TL, TaJ, TaM;
|
Chris@10
|
411 Tas = TA - TD;
|
Chris@10
|
412 TE = TA + TD;
|
Chris@10
|
413 T3C = T2X + T30;
|
Chris@10
|
414 T31 = T2X - T30;
|
Chris@10
|
415 TL = TH + TK;
|
Chris@10
|
416 TaJ = TH - TK;
|
Chris@10
|
417 TaM = TaK - TaL;
|
Chris@10
|
418 TdG = TaL + TaK;
|
Chris@10
|
419 TdA = TE - TL;
|
Chris@10
|
420 TM = TE + TL;
|
Chris@10
|
421 Tcv = TaM - TaJ;
|
Chris@10
|
422 TaN = TaJ + TaM;
|
Chris@10
|
423 TdH = Tau + Tat;
|
Chris@10
|
424 Tav = Tat - Tau;
|
Chris@10
|
425 T7w = T2S + T2V;
|
Chris@10
|
426 T2W = T2S - T2V;
|
Chris@10
|
427 T37 = T31 + T36;
|
Chris@10
|
428 T7I = T36 - T31;
|
Chris@10
|
429 }
|
Chris@10
|
430 T7H = T3A - T3x;
|
Chris@10
|
431 T3B = T3x + T3A;
|
Chris@10
|
432 TeP = TdH + TdG;
|
Chris@10
|
433 TdI = TdG - TdH;
|
Chris@10
|
434 T6i = FNMS(KP707106781, T37, T2W);
|
Chris@10
|
435 T38 = FMA(KP707106781, T37, T2W);
|
Chris@10
|
436 T3E = T3C - T3D;
|
Chris@10
|
437 T7x = T3C + T3D;
|
Chris@10
|
438 T6l = FNMS(KP707106781, T3E, T3B);
|
Chris@10
|
439 T3F = FMA(KP707106781, T3E, T3B);
|
Chris@10
|
440 T9b = FMA(KP707106781, T7I, T7H);
|
Chris@10
|
441 T7J = FNMS(KP707106781, T7I, T7H);
|
Chris@10
|
442 }
|
Chris@10
|
443 }
|
Chris@10
|
444 {
|
Chris@10
|
445 E T4r, T4I, T1F, Tbb, T4u, T4L, Tbj, TdS, T1I, Tbd, T4N, T4A, T4B, T1L, Tbc;
|
Chris@10
|
446 E T4E, T1M, Tbg;
|
Chris@10
|
447 {
|
Chris@10
|
448 E T1z, T1A, T1C, T1D, Tbi, Tbh;
|
Chris@10
|
449 T1z = Cr[WS(csr, 5)];
|
Chris@10
|
450 Tcs = Tas - Tav;
|
Chris@10
|
451 Taw = Tas + Tav;
|
Chris@10
|
452 T98 = FMA(KP707106781, T7x, T7w);
|
Chris@10
|
453 T7y = FNMS(KP707106781, T7x, T7w);
|
Chris@10
|
454 T1A = Cr[WS(csr, 59)];
|
Chris@10
|
455 T1C = Cr[WS(csr, 37)];
|
Chris@10
|
456 T1D = Cr[WS(csr, 27)];
|
Chris@10
|
457 {
|
Chris@10
|
458 E T4s, T1B, T1E, T4t, T4J, T4K;
|
Chris@10
|
459 T4s = Ci[WS(csi, 37)];
|
Chris@10
|
460 T4r = T1z - T1A;
|
Chris@10
|
461 T1B = T1z + T1A;
|
Chris@10
|
462 T4I = T1C - T1D;
|
Chris@10
|
463 T1E = T1C + T1D;
|
Chris@10
|
464 T4t = Ci[WS(csi, 27)];
|
Chris@10
|
465 T4J = Ci[WS(csi, 5)];
|
Chris@10
|
466 T4K = Ci[WS(csi, 59)];
|
Chris@10
|
467 T1F = T1B + T1E;
|
Chris@10
|
468 Tbb = T1B - T1E;
|
Chris@10
|
469 T4u = T4s + T4t;
|
Chris@10
|
470 Tbi = T4s - T4t;
|
Chris@10
|
471 Tbh = T4J - T4K;
|
Chris@10
|
472 T4L = T4J + T4K;
|
Chris@10
|
473 }
|
Chris@10
|
474 {
|
Chris@10
|
475 E T1J, T4w, T4z, T1K, T4C, T4D;
|
Chris@10
|
476 {
|
Chris@10
|
477 E T1G, T1H, T4x, T4y;
|
Chris@10
|
478 T1G = Cr[WS(csr, 21)];
|
Chris@10
|
479 Tbj = Tbh - Tbi;
|
Chris@10
|
480 TdS = Tbi + Tbh;
|
Chris@10
|
481 T1H = Cr[WS(csr, 43)];
|
Chris@10
|
482 T4x = Ci[WS(csi, 21)];
|
Chris@10
|
483 T4y = Ci[WS(csi, 43)];
|
Chris@10
|
484 T1J = Cr[WS(csr, 11)];
|
Chris@10
|
485 T4w = T1G - T1H;
|
Chris@10
|
486 T1I = T1G + T1H;
|
Chris@10
|
487 Tbd = T4x - T4y;
|
Chris@10
|
488 T4z = T4x + T4y;
|
Chris@10
|
489 T1K = Cr[WS(csr, 53)];
|
Chris@10
|
490 T4C = Ci[WS(csi, 11)];
|
Chris@10
|
491 T4D = Ci[WS(csi, 53)];
|
Chris@10
|
492 }
|
Chris@10
|
493 T4N = T4w + T4z;
|
Chris@10
|
494 T4A = T4w - T4z;
|
Chris@10
|
495 T4B = T1J - T1K;
|
Chris@10
|
496 T1L = T1J + T1K;
|
Chris@10
|
497 Tbc = T4D - T4C;
|
Chris@10
|
498 T4E = T4C + T4D;
|
Chris@10
|
499 }
|
Chris@10
|
500 }
|
Chris@10
|
501 T1M = T1I + T1L;
|
Chris@10
|
502 Tbg = T1I - T1L;
|
Chris@10
|
503 {
|
Chris@10
|
504 E TdT, Tbe, T4F, T4O;
|
Chris@10
|
505 TdT = Tbd + Tbc;
|
Chris@10
|
506 Tbe = Tbc - Tbd;
|
Chris@10
|
507 T4F = T4B - T4E;
|
Chris@10
|
508 T4O = T4B + T4E;
|
Chris@10
|
509 {
|
Chris@10
|
510 E TdR, TdU, T81, T4v, T4G, T85;
|
Chris@10
|
511 TdR = T1F - T1M;
|
Chris@10
|
512 T1N = T1F + T1M;
|
Chris@10
|
513 TeW = TdT + TdS;
|
Chris@10
|
514 TdU = TdS - TdT;
|
Chris@10
|
515 T81 = T4r + T4u;
|
Chris@10
|
516 T4v = T4r - T4u;
|
Chris@10
|
517 T4G = T4A + T4F;
|
Chris@10
|
518 T85 = T4F - T4A;
|
Chris@10
|
519 {
|
Chris@10
|
520 E T84, T4M, T4P, T82, TcG, TcH;
|
Chris@10
|
521 T84 = T4L - T4I;
|
Chris@10
|
522 T4M = T4I + T4L;
|
Chris@10
|
523 T6x = FNMS(KP707106781, T4G, T4v);
|
Chris@10
|
524 T4H = FMA(KP707106781, T4G, T4v);
|
Chris@10
|
525 Te8 = TdR + TdU;
|
Chris@10
|
526 TdV = TdR - TdU;
|
Chris@10
|
527 T4P = T4N - T4O;
|
Chris@10
|
528 T82 = T4N + T4O;
|
Chris@10
|
529 Tbk = Tbg + Tbj;
|
Chris@10
|
530 TcG = Tbj - Tbg;
|
Chris@10
|
531 T6w = FNMS(KP707106781, T4P, T4M);
|
Chris@10
|
532 T4Q = FMA(KP707106781, T4P, T4M);
|
Chris@10
|
533 T9j = FMA(KP707106781, T85, T84);
|
Chris@10
|
534 T86 = FNMS(KP707106781, T85, T84);
|
Chris@10
|
535 TcH = Tbb - Tbe;
|
Chris@10
|
536 Tbf = Tbb + Tbe;
|
Chris@10
|
537 TcO = FMA(KP414213562, TcG, TcH);
|
Chris@10
|
538 TcI = FNMS(KP414213562, TcH, TcG);
|
Chris@10
|
539 T9k = FMA(KP707106781, T82, T81);
|
Chris@10
|
540 T83 = FNMS(KP707106781, T82, T81);
|
Chris@10
|
541 }
|
Chris@10
|
542 }
|
Chris@10
|
543 }
|
Chris@10
|
544 }
|
Chris@10
|
545 }
|
Chris@10
|
546 {
|
Chris@10
|
547 E T88, T89, Tbv, Tbq;
|
Chris@10
|
548 {
|
Chris@10
|
549 E T4S, T59, T4V, Tbm, T1U, T5c, TdX, Tbu, T1X, T53, Tbo, T52, T20, T54, T5e;
|
Chris@10
|
550 E T51;
|
Chris@10
|
551 {
|
Chris@10
|
552 E T1R, T1Q, T1S, T1O, T1P;
|
Chris@10
|
553 T1O = Cr[WS(csr, 3)];
|
Chris@10
|
554 T1P = Cr[WS(csr, 61)];
|
Chris@10
|
555 T1R = Cr[WS(csr, 29)];
|
Chris@10
|
556 TbI = FMA(KP414213562, Tbf, Tbk);
|
Chris@10
|
557 Tbl = FNMS(KP414213562, Tbk, Tbf);
|
Chris@10
|
558 T1Q = T1O + T1P;
|
Chris@10
|
559 T4S = T1O - T1P;
|
Chris@10
|
560 T1S = Cr[WS(csr, 35)];
|
Chris@10
|
561 {
|
Chris@10
|
562 E Tbt, Tbs, T4X, T50;
|
Chris@10
|
563 {
|
Chris@10
|
564 E T5a, T5b, T4T, T4U, T1T;
|
Chris@10
|
565 T4T = Ci[WS(csi, 29)];
|
Chris@10
|
566 T4U = Ci[WS(csi, 35)];
|
Chris@10
|
567 T1T = T1R + T1S;
|
Chris@10
|
568 T59 = T1R - T1S;
|
Chris@10
|
569 T5a = Ci[WS(csi, 3)];
|
Chris@10
|
570 Tbt = T4T - T4U;
|
Chris@10
|
571 T4V = T4T + T4U;
|
Chris@10
|
572 T5b = Ci[WS(csi, 61)];
|
Chris@10
|
573 Tbm = T1Q - T1T;
|
Chris@10
|
574 T1U = T1Q + T1T;
|
Chris@10
|
575 T5c = T5a + T5b;
|
Chris@10
|
576 Tbs = T5b - T5a;
|
Chris@10
|
577 }
|
Chris@10
|
578 {
|
Chris@10
|
579 E T4Y, T4Z, T1V, T1W, T1Y, T1Z;
|
Chris@10
|
580 T1V = Cr[WS(csr, 13)];
|
Chris@10
|
581 T1W = Cr[WS(csr, 51)];
|
Chris@10
|
582 TdX = Tbt + Tbs;
|
Chris@10
|
583 Tbu = Tbs - Tbt;
|
Chris@10
|
584 T4Y = Ci[WS(csi, 13)];
|
Chris@10
|
585 T4X = T1V - T1W;
|
Chris@10
|
586 T1X = T1V + T1W;
|
Chris@10
|
587 T4Z = Ci[WS(csi, 51)];
|
Chris@10
|
588 T1Y = Cr[WS(csr, 19)];
|
Chris@10
|
589 T1Z = Cr[WS(csr, 45)];
|
Chris@10
|
590 T53 = Ci[WS(csi, 19)];
|
Chris@10
|
591 Tbo = T4Y - T4Z;
|
Chris@10
|
592 T50 = T4Y + T4Z;
|
Chris@10
|
593 T52 = T1Y - T1Z;
|
Chris@10
|
594 T20 = T1Y + T1Z;
|
Chris@10
|
595 T54 = Ci[WS(csi, 45)];
|
Chris@10
|
596 }
|
Chris@10
|
597 T5e = T4X + T50;
|
Chris@10
|
598 T51 = T4X - T50;
|
Chris@10
|
599 }
|
Chris@10
|
600 }
|
Chris@10
|
601 {
|
Chris@10
|
602 E T21, Tbr, T55, Tbn;
|
Chris@10
|
603 T21 = T1X + T20;
|
Chris@10
|
604 Tbr = T1X - T20;
|
Chris@10
|
605 T55 = T53 + T54;
|
Chris@10
|
606 Tbn = T54 - T53;
|
Chris@10
|
607 {
|
Chris@10
|
608 E T4W, TdW, Tbp, T5f, TdZ, T57, T8c, TdY, T56;
|
Chris@10
|
609 T88 = T4S + T4V;
|
Chris@10
|
610 T4W = T4S - T4V;
|
Chris@10
|
611 T22 = T1U + T21;
|
Chris@10
|
612 TdW = T1U - T21;
|
Chris@10
|
613 TdY = Tbo + Tbn;
|
Chris@10
|
614 Tbp = Tbn - Tbo;
|
Chris@10
|
615 T56 = T52 - T55;
|
Chris@10
|
616 T5f = T52 + T55;
|
Chris@10
|
617 TeV = TdY + TdX;
|
Chris@10
|
618 TdZ = TdX - TdY;
|
Chris@10
|
619 T57 = T51 + T56;
|
Chris@10
|
620 T8c = T56 - T51;
|
Chris@10
|
621 {
|
Chris@10
|
622 E T8b, T5d, T5g, TcD, TcE;
|
Chris@10
|
623 T8b = T59 + T5c;
|
Chris@10
|
624 T5d = T59 - T5c;
|
Chris@10
|
625 T5g = T5e - T5f;
|
Chris@10
|
626 T89 = T5e + T5f;
|
Chris@10
|
627 Te0 = TdW + TdZ;
|
Chris@10
|
628 Te9 = TdZ - TdW;
|
Chris@10
|
629 T58 = FMA(KP707106781, T57, T4W);
|
Chris@10
|
630 T6u = FNMS(KP707106781, T57, T4W);
|
Chris@10
|
631 T6t = FNMS(KP707106781, T5g, T5d);
|
Chris@10
|
632 T5h = FMA(KP707106781, T5g, T5d);
|
Chris@10
|
633 Tbv = Tbr + Tbu;
|
Chris@10
|
634 TcD = Tbu - Tbr;
|
Chris@10
|
635 TcE = Tbm - Tbp;
|
Chris@10
|
636 Tbq = Tbm + Tbp;
|
Chris@10
|
637 T9m = FNMS(KP707106781, T8c, T8b);
|
Chris@10
|
638 T8d = FMA(KP707106781, T8c, T8b);
|
Chris@10
|
639 TcP = FNMS(KP414213562, TcD, TcE);
|
Chris@10
|
640 TcF = FMA(KP414213562, TcE, TcD);
|
Chris@10
|
641 }
|
Chris@10
|
642 }
|
Chris@10
|
643 }
|
Chris@10
|
644 }
|
Chris@10
|
645 {
|
Chris@10
|
646 E Tb3, Tb8, T7V, T7Y;
|
Chris@10
|
647 {
|
Chris@10
|
648 E T7T, T4c, TaZ, T1p, TdO, Tb2, T7U, T47, T1t, T4e, T1s, Tb5, T4m, T1u, T4f;
|
Chris@10
|
649 E T4g;
|
Chris@10
|
650 {
|
Chris@10
|
651 E T1m, T43, T1l, Tb0, T4b, T1n, T44, T45;
|
Chris@10
|
652 {
|
Chris@10
|
653 E T1j, T1k, T49, T4a;
|
Chris@10
|
654 T1j = Cr[WS(csr, 9)];
|
Chris@10
|
655 T9n = FMA(KP707106781, T89, T88);
|
Chris@10
|
656 T8a = FNMS(KP707106781, T89, T88);
|
Chris@10
|
657 TbJ = FNMS(KP414213562, Tbq, Tbv);
|
Chris@10
|
658 Tbw = FMA(KP414213562, Tbv, Tbq);
|
Chris@10
|
659 T1k = Cr[WS(csr, 55)];
|
Chris@10
|
660 T49 = Ci[WS(csi, 9)];
|
Chris@10
|
661 T4a = Ci[WS(csi, 55)];
|
Chris@10
|
662 T1m = Cr[WS(csr, 41)];
|
Chris@10
|
663 T43 = T1j - T1k;
|
Chris@10
|
664 T1l = T1j + T1k;
|
Chris@10
|
665 Tb0 = T49 - T4a;
|
Chris@10
|
666 T4b = T49 + T4a;
|
Chris@10
|
667 T1n = Cr[WS(csr, 23)];
|
Chris@10
|
668 T44 = Ci[WS(csi, 41)];
|
Chris@10
|
669 T45 = Ci[WS(csi, 23)];
|
Chris@10
|
670 }
|
Chris@10
|
671 {
|
Chris@10
|
672 E T1q, T1r, T4k, T4l;
|
Chris@10
|
673 T1q = Cr[WS(csr, 7)];
|
Chris@10
|
674 {
|
Chris@10
|
675 E T48, T1o, Tb1, T46;
|
Chris@10
|
676 T48 = T1m - T1n;
|
Chris@10
|
677 T1o = T1m + T1n;
|
Chris@10
|
678 Tb1 = T44 - T45;
|
Chris@10
|
679 T46 = T44 + T45;
|
Chris@10
|
680 T7T = T4b - T48;
|
Chris@10
|
681 T4c = T48 + T4b;
|
Chris@10
|
682 TaZ = T1l - T1o;
|
Chris@10
|
683 T1p = T1l + T1o;
|
Chris@10
|
684 TdO = Tb1 + Tb0;
|
Chris@10
|
685 Tb2 = Tb0 - Tb1;
|
Chris@10
|
686 T7U = T43 + T46;
|
Chris@10
|
687 T47 = T43 - T46;
|
Chris@10
|
688 T1r = Cr[WS(csr, 57)];
|
Chris@10
|
689 }
|
Chris@10
|
690 T4k = Ci[WS(csi, 7)];
|
Chris@10
|
691 T4l = Ci[WS(csi, 57)];
|
Chris@10
|
692 T1t = Cr[WS(csr, 25)];
|
Chris@10
|
693 T4e = T1q - T1r;
|
Chris@10
|
694 T1s = T1q + T1r;
|
Chris@10
|
695 Tb5 = T4l - T4k;
|
Chris@10
|
696 T4m = T4k + T4l;
|
Chris@10
|
697 T1u = Cr[WS(csr, 39)];
|
Chris@10
|
698 T4f = Ci[WS(csi, 25)];
|
Chris@10
|
699 T4g = Ci[WS(csi, 39)];
|
Chris@10
|
700 }
|
Chris@10
|
701 }
|
Chris@10
|
702 {
|
Chris@10
|
703 E T7W, TdN, T7X, T5u, T4d, T4o, T5v, T8k, T8l;
|
Chris@10
|
704 {
|
Chris@10
|
705 E T4n, T1w, T4i, TbE, TbF, Tb4, Tb7;
|
Chris@10
|
706 {
|
Chris@10
|
707 E T4j, T1v, Tb6, T4h;
|
Chris@10
|
708 T4j = T1t - T1u;
|
Chris@10
|
709 T1v = T1t + T1u;
|
Chris@10
|
710 Tb6 = T4f - T4g;
|
Chris@10
|
711 T4h = T4f + T4g;
|
Chris@10
|
712 T7W = T4j + T4m;
|
Chris@10
|
713 T4n = T4j - T4m;
|
Chris@10
|
714 Tb4 = T1s - T1v;
|
Chris@10
|
715 T1w = T1s + T1v;
|
Chris@10
|
716 TdN = Tb6 + Tb5;
|
Chris@10
|
717 Tb7 = Tb5 - Tb6;
|
Chris@10
|
718 T7X = T4e + T4h;
|
Chris@10
|
719 T4i = T4e - T4h;
|
Chris@10
|
720 }
|
Chris@10
|
721 Tb3 = TaZ - Tb2;
|
Chris@10
|
722 TbE = TaZ + Tb2;
|
Chris@10
|
723 TbF = Tb7 - Tb4;
|
Chris@10
|
724 Tb8 = Tb4 + Tb7;
|
Chris@10
|
725 Te3 = T1p - T1w;
|
Chris@10
|
726 T1x = T1p + T1w;
|
Chris@10
|
727 TcB = TbE - TbF;
|
Chris@10
|
728 TbG = TbE + TbF;
|
Chris@10
|
729 T5u = FMA(KP414213562, T47, T4c);
|
Chris@10
|
730 T4d = FNMS(KP414213562, T4c, T47);
|
Chris@10
|
731 T4o = FMA(KP414213562, T4n, T4i);
|
Chris@10
|
732 T5v = FNMS(KP414213562, T4i, T4n);
|
Chris@10
|
733 }
|
Chris@10
|
734 Tf1 = TdO + TdN;
|
Chris@10
|
735 TdP = TdN - TdO;
|
Chris@10
|
736 T6C = T4o - T4d;
|
Chris@10
|
737 T4p = T4d + T4o;
|
Chris@10
|
738 T7V = FNMS(KP414213562, T7U, T7T);
|
Chris@10
|
739 T8k = FMA(KP414213562, T7T, T7U);
|
Chris@10
|
740 T8l = FMA(KP414213562, T7W, T7X);
|
Chris@10
|
741 T7Y = FNMS(KP414213562, T7X, T7W);
|
Chris@10
|
742 T6r = T5u - T5v;
|
Chris@10
|
743 T5w = T5u + T5v;
|
Chris@10
|
744 T9h = T8k + T8l;
|
Chris@10
|
745 T8m = T8k - T8l;
|
Chris@10
|
746 }
|
Chris@10
|
747 }
|
Chris@10
|
748 {
|
Chris@10
|
749 E T7z, T3i, Tax, TT, TdC, TaA, T7A, T3d, TX, T3k, TW, TaD, T3s, TY, T3l;
|
Chris@10
|
750 E T3m;
|
Chris@10
|
751 {
|
Chris@10
|
752 E TQ, T39, TP, Tay, T3h, TR, T3a, T3b;
|
Chris@10
|
753 {
|
Chris@10
|
754 E TN, TO, T3f, T3g;
|
Chris@10
|
755 TN = Cr[WS(csr, 10)];
|
Chris@10
|
756 TcM = Tb8 - Tb3;
|
Chris@10
|
757 Tb9 = Tb3 + Tb8;
|
Chris@10
|
758 T9s = T7V - T7Y;
|
Chris@10
|
759 T7Z = T7V + T7Y;
|
Chris@10
|
760 TO = Cr[WS(csr, 54)];
|
Chris@10
|
761 T3f = Ci[WS(csi, 10)];
|
Chris@10
|
762 T3g = Ci[WS(csi, 54)];
|
Chris@10
|
763 TQ = Cr[WS(csr, 42)];
|
Chris@10
|
764 T39 = TN - TO;
|
Chris@10
|
765 TP = TN + TO;
|
Chris@10
|
766 Tay = T3f - T3g;
|
Chris@10
|
767 T3h = T3f + T3g;
|
Chris@10
|
768 TR = Cr[WS(csr, 22)];
|
Chris@10
|
769 T3a = Ci[WS(csi, 42)];
|
Chris@10
|
770 T3b = Ci[WS(csi, 22)];
|
Chris@10
|
771 }
|
Chris@10
|
772 {
|
Chris@10
|
773 E TU, TV, T3q, T3r;
|
Chris@10
|
774 TU = Cr[WS(csr, 6)];
|
Chris@10
|
775 {
|
Chris@10
|
776 E T3e, TS, Taz, T3c;
|
Chris@10
|
777 T3e = TQ - TR;
|
Chris@10
|
778 TS = TQ + TR;
|
Chris@10
|
779 Taz = T3a - T3b;
|
Chris@10
|
780 T3c = T3a + T3b;
|
Chris@10
|
781 T7z = T3h - T3e;
|
Chris@10
|
782 T3i = T3e + T3h;
|
Chris@10
|
783 Tax = TP - TS;
|
Chris@10
|
784 TT = TP + TS;
|
Chris@10
|
785 TdC = Taz + Tay;
|
Chris@10
|
786 TaA = Tay - Taz;
|
Chris@10
|
787 T7A = T39 + T3c;
|
Chris@10
|
788 T3d = T39 - T3c;
|
Chris@10
|
789 TV = Cr[WS(csr, 58)];
|
Chris@10
|
790 }
|
Chris@10
|
791 T3q = Ci[WS(csi, 6)];
|
Chris@10
|
792 T3r = Ci[WS(csi, 58)];
|
Chris@10
|
793 TX = Cr[WS(csr, 26)];
|
Chris@10
|
794 T3k = TU - TV;
|
Chris@10
|
795 TW = TU + TV;
|
Chris@10
|
796 TaD = T3r - T3q;
|
Chris@10
|
797 T3s = T3q + T3r;
|
Chris@10
|
798 TY = Cr[WS(csr, 38)];
|
Chris@10
|
799 T3l = Ci[WS(csi, 26)];
|
Chris@10
|
800 T3m = Ci[WS(csi, 38)];
|
Chris@10
|
801 }
|
Chris@10
|
802 }
|
Chris@10
|
803 {
|
Chris@10
|
804 E T7C, TdB, T7D, T3G, T3j, T3u, T3H, T7K, T7L;
|
Chris@10
|
805 {
|
Chris@10
|
806 E T3t, T10, T3o, TaO, TaP, TaC, TaF;
|
Chris@10
|
807 {
|
Chris@10
|
808 E T3p, TZ, TaE, T3n;
|
Chris@10
|
809 T3p = TX - TY;
|
Chris@10
|
810 TZ = TX + TY;
|
Chris@10
|
811 TaE = T3l - T3m;
|
Chris@10
|
812 T3n = T3l + T3m;
|
Chris@10
|
813 T7C = T3p + T3s;
|
Chris@10
|
814 T3t = T3p - T3s;
|
Chris@10
|
815 TaC = TW - TZ;
|
Chris@10
|
816 T10 = TW + TZ;
|
Chris@10
|
817 TdB = TaE + TaD;
|
Chris@10
|
818 TaF = TaD - TaE;
|
Chris@10
|
819 T7D = T3k + T3n;
|
Chris@10
|
820 T3o = T3k - T3n;
|
Chris@10
|
821 }
|
Chris@10
|
822 TaB = Tax - TaA;
|
Chris@10
|
823 TaO = Tax + TaA;
|
Chris@10
|
824 TaP = TaF - TaC;
|
Chris@10
|
825 TaG = TaC + TaF;
|
Chris@10
|
826 TdF = TT - T10;
|
Chris@10
|
827 T11 = TT + T10;
|
Chris@10
|
828 Tct = TaO - TaP;
|
Chris@10
|
829 TaQ = TaO + TaP;
|
Chris@10
|
830 T3G = FMA(KP414213562, T3d, T3i);
|
Chris@10
|
831 T3j = FNMS(KP414213562, T3i, T3d);
|
Chris@10
|
832 T3u = FMA(KP414213562, T3t, T3o);
|
Chris@10
|
833 T3H = FNMS(KP414213562, T3o, T3t);
|
Chris@10
|
834 }
|
Chris@10
|
835 TeQ = TdC + TdB;
|
Chris@10
|
836 TdD = TdB - TdC;
|
Chris@10
|
837 T6m = T3u - T3j;
|
Chris@10
|
838 T3v = T3j + T3u;
|
Chris@10
|
839 T7B = FNMS(KP414213562, T7A, T7z);
|
Chris@10
|
840 T7K = FMA(KP414213562, T7z, T7A);
|
Chris@10
|
841 T7L = FMA(KP414213562, T7C, T7D);
|
Chris@10
|
842 T7E = FNMS(KP414213562, T7D, T7C);
|
Chris@10
|
843 T6j = T3G - T3H;
|
Chris@10
|
844 T3I = T3G + T3H;
|
Chris@10
|
845 T99 = T7K + T7L;
|
Chris@10
|
846 T7M = T7K - T7L;
|
Chris@10
|
847 }
|
Chris@10
|
848 }
|
Chris@10
|
849 }
|
Chris@10
|
850 }
|
Chris@10
|
851 }
|
Chris@10
|
852 {
|
Chris@10
|
853 E Tcw, T9c, T7F, Tev, Teu, TeD, Tep, TeG, Tez, TeE, Tes;
|
Chris@10
|
854 {
|
Chris@10
|
855 E TbX, TbY, Tc7, TbP, Tar, Tc5, Tc1, Tc0, Tc4, Tba, TbS, TbL, TbQ, TaS, Tbx;
|
Chris@10
|
856 E Tc8;
|
Chris@10
|
857 {
|
Chris@10
|
858 E TeO, TaH, TeR, TeL, TeU, TeZ, Tf2, TeX, Tfh, Tfn, Tfo, Tfm;
|
Chris@10
|
859 {
|
Chris@10
|
860 E T12, Tfg, Tfj, Tx, Tff, T24, Tfi, Tfk, Th, T1y, T23;
|
Chris@10
|
861 TeO = TM - T11;
|
Chris@10
|
862 T12 = TM + T11;
|
Chris@10
|
863 Tcw = TaG - TaB;
|
Chris@10
|
864 TaH = TaB + TaG;
|
Chris@10
|
865 T9c = T7B - T7E;
|
Chris@10
|
866 T7F = T7B + T7E;
|
Chris@10
|
867 Tfg = TeQ + TeP;
|
Chris@10
|
868 TeR = TeP - TeQ;
|
Chris@10
|
869 TeL = FNMS(KP2_000000000, Tg, T9);
|
Chris@10
|
870 Th = FMA(KP2_000000000, Tg, T9);
|
Chris@10
|
871 T1y = T1i + T1x;
|
Chris@10
|
872 TeU = T1i - T1x;
|
Chris@10
|
873 TeZ = T1N - T22;
|
Chris@10
|
874 T23 = T1N + T22;
|
Chris@10
|
875 Tfj = Tf1 + Tf0;
|
Chris@10
|
876 Tf2 = Tf0 - Tf1;
|
Chris@10
|
877 Tx = FMA(KP2_000000000, Tw, Th);
|
Chris@10
|
878 Tff = FNMS(KP2_000000000, Tw, Th);
|
Chris@10
|
879 T24 = T1y + T23;
|
Chris@10
|
880 Tfi = T1y - T23;
|
Chris@10
|
881 TeX = TeV - TeW;
|
Chris@10
|
882 Tfk = TeW + TeV;
|
Chris@10
|
883 {
|
Chris@10
|
884 E T13, Tfp, Tfl, Tfq;
|
Chris@10
|
885 T13 = FMA(KP2_000000000, T12, Tx);
|
Chris@10
|
886 Tfp = FNMS(KP2_000000000, T12, Tx);
|
Chris@10
|
887 Tfh = FNMS(KP2_000000000, Tfg, Tff);
|
Chris@10
|
888 Tfn = FMA(KP2_000000000, Tfg, Tff);
|
Chris@10
|
889 Tfl = Tfj - Tfk;
|
Chris@10
|
890 Tfq = Tfk + Tfj;
|
Chris@10
|
891 R0[0] = FMA(KP2_000000000, T24, T13);
|
Chris@10
|
892 R0[WS(rs, 32)] = FNMS(KP2_000000000, T24, T13);
|
Chris@10
|
893 R0[WS(rs, 48)] = FMA(KP2_000000000, Tfq, Tfp);
|
Chris@10
|
894 R0[WS(rs, 16)] = FNMS(KP2_000000000, Tfq, Tfp);
|
Chris@10
|
895 Tfo = Tfi + Tfl;
|
Chris@10
|
896 Tfm = Tfi - Tfl;
|
Chris@10
|
897 }
|
Chris@10
|
898 }
|
Chris@10
|
899 {
|
Chris@10
|
900 E Tf7, TeN, Tfa, Tf3, Tf8, TeS;
|
Chris@10
|
901 R0[WS(rs, 8)] = FMA(KP1_414213562, Tfm, Tfh);
|
Chris@10
|
902 R0[WS(rs, 40)] = FNMS(KP1_414213562, Tfm, Tfh);
|
Chris@10
|
903 R0[WS(rs, 56)] = FMA(KP1_414213562, Tfo, Tfn);
|
Chris@10
|
904 R0[WS(rs, 24)] = FNMS(KP1_414213562, Tfo, Tfn);
|
Chris@10
|
905 Tf7 = FMA(KP2_000000000, TeM, TeL);
|
Chris@10
|
906 TeN = FNMS(KP2_000000000, TeM, TeL);
|
Chris@10
|
907 Tfa = Tf2 - TeZ;
|
Chris@10
|
908 Tf3 = TeZ + Tf2;
|
Chris@10
|
909 Tf8 = TeO + TeR;
|
Chris@10
|
910 TeS = TeO - TeR;
|
Chris@10
|
911 {
|
Chris@10
|
912 E TbH, TbK, TaI, TaR;
|
Chris@10
|
913 {
|
Chris@10
|
914 E Taf, Tf9, Tfd, Tf5, TeT, Tfb, TeY, Taq;
|
Chris@10
|
915 TbX = FNMS(KP1_414213562, Tae, Ta9);
|
Chris@10
|
916 Taf = FMA(KP1_414213562, Tae, Ta9);
|
Chris@10
|
917 Tf9 = FNMS(KP1_414213562, Tf8, Tf7);
|
Chris@10
|
918 Tfd = FMA(KP1_414213562, Tf8, Tf7);
|
Chris@10
|
919 Tf5 = FNMS(KP1_414213562, TeS, TeN);
|
Chris@10
|
920 TeT = FMA(KP1_414213562, TeS, TeN);
|
Chris@10
|
921 Tfb = TeU - TeX;
|
Chris@10
|
922 TeY = TeU + TeX;
|
Chris@10
|
923 Taq = FNMS(KP414213562, Tap, Tak);
|
Chris@10
|
924 TbY = FMA(KP414213562, Tak, Tap);
|
Chris@10
|
925 Tc7 = FNMS(KP707106781, TbG, TbD);
|
Chris@10
|
926 TbH = FMA(KP707106781, TbG, TbD);
|
Chris@10
|
927 {
|
Chris@10
|
928 E Tfc, Tfe, Tf6, Tf4;
|
Chris@10
|
929 Tfc = FNMS(KP414213562, Tfb, Tfa);
|
Chris@10
|
930 Tfe = FMA(KP414213562, Tfa, Tfb);
|
Chris@10
|
931 Tf6 = FMA(KP414213562, TeY, Tf3);
|
Chris@10
|
932 Tf4 = FNMS(KP414213562, Tf3, TeY);
|
Chris@10
|
933 TbP = FNMS(KP1_847759065, Taq, Taf);
|
Chris@10
|
934 Tar = FMA(KP1_847759065, Taq, Taf);
|
Chris@10
|
935 R0[WS(rs, 44)] = FMA(KP1_847759065, Tfc, Tf9);
|
Chris@10
|
936 R0[WS(rs, 12)] = FNMS(KP1_847759065, Tfc, Tf9);
|
Chris@10
|
937 R0[WS(rs, 60)] = FMA(KP1_847759065, Tfe, Tfd);
|
Chris@10
|
938 R0[WS(rs, 28)] = FNMS(KP1_847759065, Tfe, Tfd);
|
Chris@10
|
939 R0[WS(rs, 52)] = FMA(KP1_847759065, Tf6, Tf5);
|
Chris@10
|
940 R0[WS(rs, 20)] = FNMS(KP1_847759065, Tf6, Tf5);
|
Chris@10
|
941 R0[WS(rs, 4)] = FMA(KP1_847759065, Tf4, TeT);
|
Chris@10
|
942 R0[WS(rs, 36)] = FNMS(KP1_847759065, Tf4, TeT);
|
Chris@10
|
943 TbK = TbI + TbJ;
|
Chris@10
|
944 Tc5 = TbI - TbJ;
|
Chris@10
|
945 }
|
Chris@10
|
946 }
|
Chris@10
|
947 Tc1 = FNMS(KP707106781, TaH, Taw);
|
Chris@10
|
948 TaI = FMA(KP707106781, TaH, Taw);
|
Chris@10
|
949 TaR = FMA(KP707106781, TaQ, TaN);
|
Chris@10
|
950 Tc0 = FNMS(KP707106781, TaQ, TaN);
|
Chris@10
|
951 Tc4 = FNMS(KP707106781, Tb9, TaY);
|
Chris@10
|
952 Tba = FMA(KP707106781, Tb9, TaY);
|
Chris@10
|
953 TbS = FNMS(KP923879532, TbK, TbH);
|
Chris@10
|
954 TbL = FMA(KP923879532, TbK, TbH);
|
Chris@10
|
955 TbQ = FMA(KP198912367, TaI, TaR);
|
Chris@10
|
956 TaS = FNMS(KP198912367, TaR, TaI);
|
Chris@10
|
957 Tbx = Tbl + Tbw;
|
Chris@10
|
958 Tc8 = Tbw - Tbl;
|
Chris@10
|
959 }
|
Chris@10
|
960 }
|
Chris@10
|
961 }
|
Chris@10
|
962 {
|
Chris@10
|
963 E Ten, Teo, Tex, Tef, Tdz, Ter, Teq, TdQ, Tei, Teb, Teg, TdK, Te1, Tey;
|
Chris@10
|
964 {
|
Chris@10
|
965 E Te7, Tea, TdE, TdJ;
|
Chris@10
|
966 {
|
Chris@10
|
967 E Tdt, TbR, TbV, TbN, TaT, TbT, Tby, Tdy;
|
Chris@10
|
968 Ten = FMA(KP2_000000000, Tds, Tdr);
|
Chris@10
|
969 Tdt = FNMS(KP2_000000000, Tds, Tdr);
|
Chris@10
|
970 TbR = FNMS(KP1_961570560, TbQ, TbP);
|
Chris@10
|
971 TbV = FMA(KP1_961570560, TbQ, TbP);
|
Chris@10
|
972 TbN = FNMS(KP1_961570560, TaS, Tar);
|
Chris@10
|
973 TaT = FMA(KP1_961570560, TaS, Tar);
|
Chris@10
|
974 TbT = FNMS(KP923879532, Tbx, Tba);
|
Chris@10
|
975 Tby = FMA(KP923879532, Tbx, Tba);
|
Chris@10
|
976 Tdy = Tdu - Tdx;
|
Chris@10
|
977 Teo = Tdu + Tdx;
|
Chris@10
|
978 Tex = Te6 - Te3;
|
Chris@10
|
979 Te7 = Te3 + Te6;
|
Chris@10
|
980 {
|
Chris@10
|
981 E TbU, TbW, TbO, TbM;
|
Chris@10
|
982 TbU = FNMS(KP820678790, TbT, TbS);
|
Chris@10
|
983 TbW = FMA(KP820678790, TbS, TbT);
|
Chris@10
|
984 TbO = FMA(KP098491403, Tby, TbL);
|
Chris@10
|
985 TbM = FNMS(KP098491403, TbL, Tby);
|
Chris@10
|
986 Tef = FNMS(KP1_414213562, Tdy, Tdt);
|
Chris@10
|
987 Tdz = FMA(KP1_414213562, Tdy, Tdt);
|
Chris@10
|
988 R0[WS(rs, 41)] = FMA(KP1_546020906, TbU, TbR);
|
Chris@10
|
989 R0[WS(rs, 9)] = FNMS(KP1_546020906, TbU, TbR);
|
Chris@10
|
990 R0[WS(rs, 57)] = FMA(KP1_546020906, TbW, TbV);
|
Chris@10
|
991 R0[WS(rs, 25)] = FNMS(KP1_546020906, TbW, TbV);
|
Chris@10
|
992 R0[WS(rs, 49)] = FMA(KP1_990369453, TbO, TbN);
|
Chris@10
|
993 R0[WS(rs, 17)] = FNMS(KP1_990369453, TbO, TbN);
|
Chris@10
|
994 R0[WS(rs, 1)] = FMA(KP1_990369453, TbM, TaT);
|
Chris@10
|
995 R0[WS(rs, 33)] = FNMS(KP1_990369453, TbM, TaT);
|
Chris@10
|
996 Tea = Te8 + Te9;
|
Chris@10
|
997 Tev = Te8 - Te9;
|
Chris@10
|
998 }
|
Chris@10
|
999 }
|
Chris@10
|
1000 Ter = TdA - TdD;
|
Chris@10
|
1001 TdE = TdA + TdD;
|
Chris@10
|
1002 TdJ = TdF + TdI;
|
Chris@10
|
1003 Teq = TdI - TdF;
|
Chris@10
|
1004 Teu = TdM - TdP;
|
Chris@10
|
1005 TdQ = TdM + TdP;
|
Chris@10
|
1006 Tei = FNMS(KP707106781, Tea, Te7);
|
Chris@10
|
1007 Teb = FMA(KP707106781, Tea, Te7);
|
Chris@10
|
1008 Teg = FMA(KP414213562, TdE, TdJ);
|
Chris@10
|
1009 TdK = FNMS(KP414213562, TdJ, TdE);
|
Chris@10
|
1010 Te1 = TdV + Te0;
|
Chris@10
|
1011 Tey = Te0 - TdV;
|
Chris@10
|
1012 }
|
Chris@10
|
1013 {
|
Chris@10
|
1014 E Tcd, TbZ, Tcg, Tc9, Tce, Tc2;
|
Chris@10
|
1015 {
|
Chris@10
|
1016 E Teh, Tel, Ted, TdL, Tej, Te2;
|
Chris@10
|
1017 Teh = FNMS(KP1_847759065, Teg, Tef);
|
Chris@10
|
1018 Tel = FMA(KP1_847759065, Teg, Tef);
|
Chris@10
|
1019 Ted = FNMS(KP1_847759065, TdK, Tdz);
|
Chris@10
|
1020 TdL = FMA(KP1_847759065, TdK, Tdz);
|
Chris@10
|
1021 Tej = FNMS(KP707106781, Te1, TdQ);
|
Chris@10
|
1022 Te2 = FMA(KP707106781, Te1, TdQ);
|
Chris@10
|
1023 {
|
Chris@10
|
1024 E Tek, Tem, Tee, Tec;
|
Chris@10
|
1025 Tek = FNMS(KP668178637, Tej, Tei);
|
Chris@10
|
1026 Tem = FMA(KP668178637, Tei, Tej);
|
Chris@10
|
1027 Tee = FMA(KP198912367, Te2, Teb);
|
Chris@10
|
1028 Tec = FNMS(KP198912367, Teb, Te2);
|
Chris@10
|
1029 Tcd = FMA(KP1_847759065, TbY, TbX);
|
Chris@10
|
1030 TbZ = FNMS(KP1_847759065, TbY, TbX);
|
Chris@10
|
1031 R0[WS(rs, 42)] = FMA(KP1_662939224, Tek, Teh);
|
Chris@10
|
1032 R0[WS(rs, 10)] = FNMS(KP1_662939224, Tek, Teh);
|
Chris@10
|
1033 R0[WS(rs, 58)] = FMA(KP1_662939224, Tem, Tel);
|
Chris@10
|
1034 R0[WS(rs, 26)] = FNMS(KP1_662939224, Tem, Tel);
|
Chris@10
|
1035 R0[WS(rs, 50)] = FMA(KP1_961570560, Tee, Ted);
|
Chris@10
|
1036 R0[WS(rs, 18)] = FNMS(KP1_961570560, Tee, Ted);
|
Chris@10
|
1037 R0[WS(rs, 2)] = FMA(KP1_961570560, Tec, TdL);
|
Chris@10
|
1038 R0[WS(rs, 34)] = FNMS(KP1_961570560, Tec, TdL);
|
Chris@10
|
1039 }
|
Chris@10
|
1040 }
|
Chris@10
|
1041 Tcg = FMA(KP923879532, Tc8, Tc7);
|
Chris@10
|
1042 Tc9 = FNMS(KP923879532, Tc8, Tc7);
|
Chris@10
|
1043 Tce = FMA(KP668178637, Tc0, Tc1);
|
Chris@10
|
1044 Tc2 = FNMS(KP668178637, Tc1, Tc0);
|
Chris@10
|
1045 {
|
Chris@10
|
1046 E Tcf, Tcj, Tcb, Tc3, Tch, Tc6;
|
Chris@10
|
1047 Tcf = FNMS(KP1_662939224, Tce, Tcd);
|
Chris@10
|
1048 Tcj = FMA(KP1_662939224, Tce, Tcd);
|
Chris@10
|
1049 Tcb = FMA(KP1_662939224, Tc2, TbZ);
|
Chris@10
|
1050 Tc3 = FNMS(KP1_662939224, Tc2, TbZ);
|
Chris@10
|
1051 Tch = FMA(KP923879532, Tc5, Tc4);
|
Chris@10
|
1052 Tc6 = FNMS(KP923879532, Tc5, Tc4);
|
Chris@10
|
1053 {
|
Chris@10
|
1054 E Tci, Tck, Tcc, Tca;
|
Chris@10
|
1055 Tci = FNMS(KP303346683, Tch, Tcg);
|
Chris@10
|
1056 Tck = FMA(KP303346683, Tcg, Tch);
|
Chris@10
|
1057 Tcc = FMA(KP534511135, Tc6, Tc9);
|
Chris@10
|
1058 Tca = FNMS(KP534511135, Tc9, Tc6);
|
Chris@10
|
1059 TeD = FMA(KP1_414213562, Teo, Ten);
|
Chris@10
|
1060 Tep = FNMS(KP1_414213562, Teo, Ten);
|
Chris@10
|
1061 R0[WS(rs, 45)] = FMA(KP1_913880671, Tci, Tcf);
|
Chris@10
|
1062 R0[WS(rs, 13)] = FNMS(KP1_913880671, Tci, Tcf);
|
Chris@10
|
1063 R0[WS(rs, 61)] = FMA(KP1_913880671, Tck, Tcj);
|
Chris@10
|
1064 R0[WS(rs, 29)] = FNMS(KP1_913880671, Tck, Tcj);
|
Chris@10
|
1065 R0[WS(rs, 53)] = FMA(KP1_763842528, Tcc, Tcb);
|
Chris@10
|
1066 R0[WS(rs, 21)] = FNMS(KP1_763842528, Tcc, Tcb);
|
Chris@10
|
1067 R0[WS(rs, 5)] = FMA(KP1_763842528, Tca, Tc3);
|
Chris@10
|
1068 R0[WS(rs, 37)] = FNMS(KP1_763842528, Tca, Tc3);
|
Chris@10
|
1069 }
|
Chris@10
|
1070 }
|
Chris@10
|
1071 TeG = FMA(KP707106781, Tey, Tex);
|
Chris@10
|
1072 Tez = FNMS(KP707106781, Tey, Tex);
|
Chris@10
|
1073 TeE = FMA(KP414213562, Teq, Ter);
|
Chris@10
|
1074 Tes = FNMS(KP414213562, Ter, Teq);
|
Chris@10
|
1075 }
|
Chris@10
|
1076 }
|
Chris@10
|
1077 }
|
Chris@10
|
1078 {
|
Chris@10
|
1079 E T5L, T5M, T61, T62;
|
Chris@10
|
1080 {
|
Chris@10
|
1081 E Td3, Td4, Tdd, TcV, Tcr, Tdb, Td7, Td6, Tda, TcC, TcY, TcR, TcW, Tcy, TcJ;
|
Chris@10
|
1082 E Tde;
|
Chris@10
|
1083 {
|
Chris@10
|
1084 E TcN, TcQ, Tcu, Tcx;
|
Chris@10
|
1085 {
|
Chris@10
|
1086 E Tcn, TeF, TeJ, TeB, Tet, TeH, Tew, Tcq;
|
Chris@10
|
1087 Td3 = FMA(KP1_414213562, Tcm, Tcl);
|
Chris@10
|
1088 Tcn = FNMS(KP1_414213562, Tcm, Tcl);
|
Chris@10
|
1089 TeF = FNMS(KP1_847759065, TeE, TeD);
|
Chris@10
|
1090 TeJ = FMA(KP1_847759065, TeE, TeD);
|
Chris@10
|
1091 TeB = FMA(KP1_847759065, Tes, Tep);
|
Chris@10
|
1092 Tet = FNMS(KP1_847759065, Tes, Tep);
|
Chris@10
|
1093 TeH = FMA(KP707106781, Tev, Teu);
|
Chris@10
|
1094 Tew = FNMS(KP707106781, Tev, Teu);
|
Chris@10
|
1095 Tcq = FNMS(KP414213562, Tcp, Tco);
|
Chris@10
|
1096 Td4 = FMA(KP414213562, Tco, Tcp);
|
Chris@10
|
1097 Tdd = FMA(KP707106781, TcM, TcL);
|
Chris@10
|
1098 TcN = FNMS(KP707106781, TcM, TcL);
|
Chris@10
|
1099 {
|
Chris@10
|
1100 E TeI, TeK, TeC, TeA;
|
Chris@10
|
1101 TeI = FNMS(KP198912367, TeH, TeG);
|
Chris@10
|
1102 TeK = FMA(KP198912367, TeG, TeH);
|
Chris@10
|
1103 TeC = FMA(KP668178637, Tew, Tez);
|
Chris@10
|
1104 TeA = FNMS(KP668178637, Tez, Tew);
|
Chris@10
|
1105 TcV = FMA(KP1_847759065, Tcq, Tcn);
|
Chris@10
|
1106 Tcr = FNMS(KP1_847759065, Tcq, Tcn);
|
Chris@10
|
1107 R0[WS(rs, 46)] = FMA(KP1_961570560, TeI, TeF);
|
Chris@10
|
1108 R0[WS(rs, 14)] = FNMS(KP1_961570560, TeI, TeF);
|
Chris@10
|
1109 R0[WS(rs, 62)] = FMA(KP1_961570560, TeK, TeJ);
|
Chris@10
|
1110 R0[WS(rs, 30)] = FNMS(KP1_961570560, TeK, TeJ);
|
Chris@10
|
1111 R0[WS(rs, 54)] = FMA(KP1_662939224, TeC, TeB);
|
Chris@10
|
1112 R0[WS(rs, 22)] = FNMS(KP1_662939224, TeC, TeB);
|
Chris@10
|
1113 R0[WS(rs, 6)] = FMA(KP1_662939224, TeA, Tet);
|
Chris@10
|
1114 R0[WS(rs, 38)] = FNMS(KP1_662939224, TeA, Tet);
|
Chris@10
|
1115 TcQ = TcO - TcP;
|
Chris@10
|
1116 Tdb = TcO + TcP;
|
Chris@10
|
1117 }
|
Chris@10
|
1118 }
|
Chris@10
|
1119 Td7 = FMA(KP707106781, Tct, Tcs);
|
Chris@10
|
1120 Tcu = FNMS(KP707106781, Tct, Tcs);
|
Chris@10
|
1121 Tcx = FNMS(KP707106781, Tcw, Tcv);
|
Chris@10
|
1122 Td6 = FMA(KP707106781, Tcw, Tcv);
|
Chris@10
|
1123 Tda = FMA(KP707106781, TcB, TcA);
|
Chris@10
|
1124 TcC = FNMS(KP707106781, TcB, TcA);
|
Chris@10
|
1125 TcY = FNMS(KP923879532, TcQ, TcN);
|
Chris@10
|
1126 TcR = FMA(KP923879532, TcQ, TcN);
|
Chris@10
|
1127 TcW = FMA(KP668178637, Tcu, Tcx);
|
Chris@10
|
1128 Tcy = FNMS(KP668178637, Tcx, Tcu);
|
Chris@10
|
1129 TcJ = TcF - TcI;
|
Chris@10
|
1130 Tde = TcI + TcF;
|
Chris@10
|
1131 }
|
Chris@10
|
1132 {
|
Chris@10
|
1133 E Tdj, Td5, Tdm, Tdf, Tdk, Td8;
|
Chris@10
|
1134 {
|
Chris@10
|
1135 E TcX, Td1, TcT, Tcz, TcZ, TcK;
|
Chris@10
|
1136 TcX = FNMS(KP1_662939224, TcW, TcV);
|
Chris@10
|
1137 Td1 = FMA(KP1_662939224, TcW, TcV);
|
Chris@10
|
1138 TcT = FNMS(KP1_662939224, Tcy, Tcr);
|
Chris@10
|
1139 Tcz = FMA(KP1_662939224, Tcy, Tcr);
|
Chris@10
|
1140 TcZ = FNMS(KP923879532, TcJ, TcC);
|
Chris@10
|
1141 TcK = FMA(KP923879532, TcJ, TcC);
|
Chris@10
|
1142 {
|
Chris@10
|
1143 E Td0, Td2, TcU, TcS;
|
Chris@10
|
1144 Td0 = FNMS(KP534511135, TcZ, TcY);
|
Chris@10
|
1145 Td2 = FMA(KP534511135, TcY, TcZ);
|
Chris@10
|
1146 TcU = FMA(KP303346683, TcK, TcR);
|
Chris@10
|
1147 TcS = FNMS(KP303346683, TcR, TcK);
|
Chris@10
|
1148 Tdj = FMA(KP1_847759065, Td4, Td3);
|
Chris@10
|
1149 Td5 = FNMS(KP1_847759065, Td4, Td3);
|
Chris@10
|
1150 R0[WS(rs, 43)] = FMA(KP1_763842528, Td0, TcX);
|
Chris@10
|
1151 R0[WS(rs, 11)] = FNMS(KP1_763842528, Td0, TcX);
|
Chris@10
|
1152 R0[WS(rs, 59)] = FMA(KP1_763842528, Td2, Td1);
|
Chris@10
|
1153 R0[WS(rs, 27)] = FNMS(KP1_763842528, Td2, Td1);
|
Chris@10
|
1154 R0[WS(rs, 51)] = FMA(KP1_913880671, TcU, TcT);
|
Chris@10
|
1155 R0[WS(rs, 19)] = FNMS(KP1_913880671, TcU, TcT);
|
Chris@10
|
1156 R0[WS(rs, 3)] = FMA(KP1_913880671, TcS, Tcz);
|
Chris@10
|
1157 R0[WS(rs, 35)] = FNMS(KP1_913880671, TcS, Tcz);
|
Chris@10
|
1158 }
|
Chris@10
|
1159 }
|
Chris@10
|
1160 Tdm = FMA(KP923879532, Tde, Tdd);
|
Chris@10
|
1161 Tdf = FNMS(KP923879532, Tde, Tdd);
|
Chris@10
|
1162 Tdk = FMA(KP198912367, Td6, Td7);
|
Chris@10
|
1163 Td8 = FNMS(KP198912367, Td7, Td6);
|
Chris@10
|
1164 {
|
Chris@10
|
1165 E T5F, T2R, T5G, T3K, T64, T5S, T5X, T5x, T5U, T4q, T4R, T63, T5P, T5i, T5V;
|
Chris@10
|
1166 E T5A;
|
Chris@10
|
1167 {
|
Chris@10
|
1168 E T5N, T5O, T5R, T3w, T3J, T5Q, T5y, T5z;
|
Chris@10
|
1169 {
|
Chris@10
|
1170 E T2p, Tdl, Tdp, Tdh, Td9, Tdn, Tdc, T2Q;
|
Chris@10
|
1171 T5N = FNMS(KP1_847759065, T2o, T2d);
|
Chris@10
|
1172 T2p = FMA(KP1_847759065, T2o, T2d);
|
Chris@10
|
1173 Tdl = FNMS(KP1_961570560, Tdk, Tdj);
|
Chris@10
|
1174 Tdp = FMA(KP1_961570560, Tdk, Tdj);
|
Chris@10
|
1175 Tdh = FMA(KP1_961570560, Td8, Td5);
|
Chris@10
|
1176 Td9 = FNMS(KP1_961570560, Td8, Td5);
|
Chris@10
|
1177 Tdn = FMA(KP923879532, Tdb, Tda);
|
Chris@10
|
1178 Tdc = FNMS(KP923879532, Tdb, Tda);
|
Chris@10
|
1179 T2Q = FNMS(KP198912367, T2P, T2G);
|
Chris@10
|
1180 T5O = FMA(KP198912367, T2G, T2P);
|
Chris@10
|
1181 T5R = FNMS(KP923879532, T3v, T38);
|
Chris@10
|
1182 T3w = FMA(KP923879532, T3v, T38);
|
Chris@10
|
1183 {
|
Chris@10
|
1184 E Tdo, Tdq, Tdi, Tdg;
|
Chris@10
|
1185 Tdo = FNMS(KP098491403, Tdn, Tdm);
|
Chris@10
|
1186 Tdq = FMA(KP098491403, Tdm, Tdn);
|
Chris@10
|
1187 Tdi = FMA(KP820678790, Tdc, Tdf);
|
Chris@10
|
1188 Tdg = FNMS(KP820678790, Tdf, Tdc);
|
Chris@10
|
1189 T5F = FNMS(KP1_961570560, T2Q, T2p);
|
Chris@10
|
1190 T2R = FMA(KP1_961570560, T2Q, T2p);
|
Chris@10
|
1191 R0[WS(rs, 47)] = FMA(KP1_990369453, Tdo, Tdl);
|
Chris@10
|
1192 R0[WS(rs, 15)] = FNMS(KP1_990369453, Tdo, Tdl);
|
Chris@10
|
1193 R0[WS(rs, 63)] = FMA(KP1_990369453, Tdq, Tdp);
|
Chris@10
|
1194 R0[WS(rs, 31)] = FNMS(KP1_990369453, Tdq, Tdp);
|
Chris@10
|
1195 R0[WS(rs, 55)] = FMA(KP1_546020906, Tdi, Tdh);
|
Chris@10
|
1196 R0[WS(rs, 23)] = FNMS(KP1_546020906, Tdi, Tdh);
|
Chris@10
|
1197 R0[WS(rs, 7)] = FMA(KP1_546020906, Tdg, Td9);
|
Chris@10
|
1198 R0[WS(rs, 39)] = FNMS(KP1_546020906, Tdg, Td9);
|
Chris@10
|
1199 T3J = FMA(KP923879532, T3I, T3F);
|
Chris@10
|
1200 T5Q = FNMS(KP923879532, T3I, T3F);
|
Chris@10
|
1201 }
|
Chris@10
|
1202 }
|
Chris@10
|
1203 T5G = FMA(KP098491403, T3w, T3J);
|
Chris@10
|
1204 T3K = FNMS(KP098491403, T3J, T3w);
|
Chris@10
|
1205 T64 = FMA(KP820678790, T5Q, T5R);
|
Chris@10
|
1206 T5S = FNMS(KP820678790, T5R, T5Q);
|
Chris@10
|
1207 T5X = FNMS(KP923879532, T5w, T5t);
|
Chris@10
|
1208 T5x = FMA(KP923879532, T5w, T5t);
|
Chris@10
|
1209 T5U = FNMS(KP923879532, T4p, T42);
|
Chris@10
|
1210 T4q = FMA(KP923879532, T4p, T42);
|
Chris@10
|
1211 T4R = FNMS(KP198912367, T4Q, T4H);
|
Chris@10
|
1212 T5y = FMA(KP198912367, T4H, T4Q);
|
Chris@10
|
1213 T63 = FMA(KP1_961570560, T5O, T5N);
|
Chris@10
|
1214 T5P = FNMS(KP1_961570560, T5O, T5N);
|
Chris@10
|
1215 T5z = FNMS(KP198912367, T58, T5h);
|
Chris@10
|
1216 T5i = FMA(KP198912367, T5h, T58);
|
Chris@10
|
1217 T5V = T5y - T5z;
|
Chris@10
|
1218 T5A = T5y + T5z;
|
Chris@10
|
1219 }
|
Chris@10
|
1220 {
|
Chris@10
|
1221 E T5W, T5I, T5Z, T5J;
|
Chris@10
|
1222 {
|
Chris@10
|
1223 E T5D, T3L, T67, T5B, T5Y, T5j, T65, T69, T66, T5k;
|
Chris@10
|
1224 T5D = FNMS(KP1_990369453, T3K, T2R);
|
Chris@10
|
1225 T3L = FMA(KP1_990369453, T3K, T2R);
|
Chris@10
|
1226 T5W = FNMS(KP980785280, T5V, T5U);
|
Chris@10
|
1227 T67 = FMA(KP980785280, T5V, T5U);
|
Chris@10
|
1228 T5I = FNMS(KP980785280, T5A, T5x);
|
Chris@10
|
1229 T5B = FMA(KP980785280, T5A, T5x);
|
Chris@10
|
1230 T5Y = T5i - T4R;
|
Chris@10
|
1231 T5j = T4R + T5i;
|
Chris@10
|
1232 T65 = FNMS(KP1_546020906, T64, T63);
|
Chris@10
|
1233 T69 = FMA(KP1_546020906, T64, T63);
|
Chris@10
|
1234 T5Z = FNMS(KP980785280, T5Y, T5X);
|
Chris@10
|
1235 T66 = FMA(KP980785280, T5Y, T5X);
|
Chris@10
|
1236 T5J = FNMS(KP980785280, T5j, T4q);
|
Chris@10
|
1237 T5k = FMA(KP980785280, T5j, T4q);
|
Chris@10
|
1238 {
|
Chris@10
|
1239 E T68, T6a, T5E, T5C;
|
Chris@10
|
1240 T68 = FNMS(KP357805721, T67, T66);
|
Chris@10
|
1241 T6a = FMA(KP357805721, T66, T67);
|
Chris@10
|
1242 T5E = FMA(KP049126849, T5k, T5B);
|
Chris@10
|
1243 T5C = FNMS(KP049126849, T5B, T5k);
|
Chris@10
|
1244 R1[WS(rs, 60)] = FMA(KP1_883088130, T6a, T69);
|
Chris@10
|
1245 R1[WS(rs, 28)] = FNMS(KP1_883088130, T6a, T69);
|
Chris@10
|
1246 R1[WS(rs, 44)] = FMA(KP1_883088130, T68, T65);
|
Chris@10
|
1247 R1[WS(rs, 12)] = FNMS(KP1_883088130, T68, T65);
|
Chris@10
|
1248 R1[0] = FMA(KP1_997590912, T5C, T3L);
|
Chris@10
|
1249 R1[WS(rs, 32)] = FNMS(KP1_997590912, T5C, T3L);
|
Chris@10
|
1250 R1[WS(rs, 16)] = FNMS(KP1_997590912, T5E, T5D);
|
Chris@10
|
1251 R1[WS(rs, 48)] = FMA(KP1_997590912, T5E, T5D);
|
Chris@10
|
1252 }
|
Chris@10
|
1253 }
|
Chris@10
|
1254 {
|
Chris@10
|
1255 E T5H, T5K, T5T, T60;
|
Chris@10
|
1256 T5L = FMA(KP1_990369453, T5G, T5F);
|
Chris@10
|
1257 T5H = FNMS(KP1_990369453, T5G, T5F);
|
Chris@10
|
1258 T5K = FNMS(KP906347169, T5J, T5I);
|
Chris@10
|
1259 T5M = FMA(KP906347169, T5I, T5J);
|
Chris@10
|
1260 T61 = FMA(KP1_546020906, T5S, T5P);
|
Chris@10
|
1261 T5T = FNMS(KP1_546020906, T5S, T5P);
|
Chris@10
|
1262 T60 = FNMS(KP472964775, T5Z, T5W);
|
Chris@10
|
1263 T62 = FMA(KP472964775, T5W, T5Z);
|
Chris@10
|
1264 R1[WS(rs, 40)] = FMA(KP1_481902250, T5K, T5H);
|
Chris@10
|
1265 R1[WS(rs, 8)] = FNMS(KP1_481902250, T5K, T5H);
|
Chris@10
|
1266 R1[WS(rs, 4)] = FMA(KP1_807978586, T60, T5T);
|
Chris@10
|
1267 R1[WS(rs, 36)] = FNMS(KP1_807978586, T60, T5T);
|
Chris@10
|
1268 }
|
Chris@10
|
1269 }
|
Chris@10
|
1270 }
|
Chris@10
|
1271 }
|
Chris@10
|
1272 }
|
Chris@10
|
1273 {
|
Chris@10
|
1274 E T8B, T8C, T8R, T8S;
|
Chris@10
|
1275 {
|
Chris@10
|
1276 E T8v, T7v, T8w, T7O, T8N, T8n, T8U, T8I, T8T, T8F, T8K, T80, T87, T8e, T8L;
|
Chris@10
|
1277 E T8q;
|
Chris@10
|
1278 {
|
Chris@10
|
1279 E T8D, T8E, T8H, T8G, T8o, T8p;
|
Chris@10
|
1280 {
|
Chris@10
|
1281 E T7n, T7u, T7G, T7N;
|
Chris@10
|
1282 T8D = FMA(KP1_847759065, T7m, T7j);
|
Chris@10
|
1283 T7n = FNMS(KP1_847759065, T7m, T7j);
|
Chris@10
|
1284 R1[WS(rs, 52)] = FMA(KP1_807978586, T62, T61);
|
Chris@10
|
1285 R1[WS(rs, 20)] = FNMS(KP1_807978586, T62, T61);
|
Chris@10
|
1286 R1[WS(rs, 56)] = FMA(KP1_481902250, T5M, T5L);
|
Chris@10
|
1287 R1[WS(rs, 24)] = FNMS(KP1_481902250, T5M, T5L);
|
Chris@10
|
1288 T7u = FNMS(KP668178637, T7t, T7q);
|
Chris@10
|
1289 T8E = FMA(KP668178637, T7q, T7t);
|
Chris@10
|
1290 T8H = FMA(KP923879532, T7F, T7y);
|
Chris@10
|
1291 T7G = FNMS(KP923879532, T7F, T7y);
|
Chris@10
|
1292 T7N = FMA(KP923879532, T7M, T7J);
|
Chris@10
|
1293 T8G = FNMS(KP923879532, T7M, T7J);
|
Chris@10
|
1294 T8v = FNMS(KP1_662939224, T7u, T7n);
|
Chris@10
|
1295 T7v = FMA(KP1_662939224, T7u, T7n);
|
Chris@10
|
1296 T8w = FMA(KP303346683, T7G, T7N);
|
Chris@10
|
1297 T7O = FNMS(KP303346683, T7N, T7G);
|
Chris@10
|
1298 }
|
Chris@10
|
1299 T8N = FNMS(KP923879532, T8m, T8j);
|
Chris@10
|
1300 T8n = FMA(KP923879532, T8m, T8j);
|
Chris@10
|
1301 T8U = FMA(KP534511135, T8G, T8H);
|
Chris@10
|
1302 T8I = FNMS(KP534511135, T8H, T8G);
|
Chris@10
|
1303 T8T = FMA(KP1_662939224, T8E, T8D);
|
Chris@10
|
1304 T8F = FNMS(KP1_662939224, T8E, T8D);
|
Chris@10
|
1305 T8K = FMA(KP923879532, T7Z, T7S);
|
Chris@10
|
1306 T80 = FNMS(KP923879532, T7Z, T7S);
|
Chris@10
|
1307 T87 = FNMS(KP668178637, T86, T83);
|
Chris@10
|
1308 T8o = FMA(KP668178637, T83, T86);
|
Chris@10
|
1309 T8p = FMA(KP668178637, T8a, T8d);
|
Chris@10
|
1310 T8e = FNMS(KP668178637, T8d, T8a);
|
Chris@10
|
1311 T8L = T8o + T8p;
|
Chris@10
|
1312 T8q = T8o - T8p;
|
Chris@10
|
1313 }
|
Chris@10
|
1314 {
|
Chris@10
|
1315 E T8M, T8y, T8P, T8z;
|
Chris@10
|
1316 {
|
Chris@10
|
1317 E T8t, T7P, T8X, T8r, T8O, T8f, T8V, T8Z, T8W, T8g;
|
Chris@10
|
1318 T8t = FNMS(KP1_913880671, T7O, T7v);
|
Chris@10
|
1319 T7P = FMA(KP1_913880671, T7O, T7v);
|
Chris@10
|
1320 T8M = FNMS(KP831469612, T8L, T8K);
|
Chris@10
|
1321 T8X = FMA(KP831469612, T8L, T8K);
|
Chris@10
|
1322 T8y = FNMS(KP831469612, T8q, T8n);
|
Chris@10
|
1323 T8r = FMA(KP831469612, T8q, T8n);
|
Chris@10
|
1324 T8O = T8e - T87;
|
Chris@10
|
1325 T8f = T87 + T8e;
|
Chris@10
|
1326 T8V = FNMS(KP1_763842528, T8U, T8T);
|
Chris@10
|
1327 T8Z = FMA(KP1_763842528, T8U, T8T);
|
Chris@10
|
1328 T8P = FNMS(KP831469612, T8O, T8N);
|
Chris@10
|
1329 T8W = FMA(KP831469612, T8O, T8N);
|
Chris@10
|
1330 T8z = FNMS(KP831469612, T8f, T80);
|
Chris@10
|
1331 T8g = FMA(KP831469612, T8f, T80);
|
Chris@10
|
1332 {
|
Chris@10
|
1333 E T8Y, T90, T8u, T8s;
|
Chris@10
|
1334 T8Y = FNMS(KP250486960, T8X, T8W);
|
Chris@10
|
1335 T90 = FMA(KP250486960, T8W, T8X);
|
Chris@10
|
1336 T8u = FMA(KP148335987, T8g, T8r);
|
Chris@10
|
1337 T8s = FNMS(KP148335987, T8r, T8g);
|
Chris@10
|
1338 R1[WS(rs, 61)] = FMA(KP1_940062506, T90, T8Z);
|
Chris@10
|
1339 R1[WS(rs, 29)] = FNMS(KP1_940062506, T90, T8Z);
|
Chris@10
|
1340 R1[WS(rs, 45)] = FMA(KP1_940062506, T8Y, T8V);
|
Chris@10
|
1341 R1[WS(rs, 13)] = FNMS(KP1_940062506, T8Y, T8V);
|
Chris@10
|
1342 R1[WS(rs, 1)] = FMA(KP1_978353019, T8s, T7P);
|
Chris@10
|
1343 R1[WS(rs, 33)] = FNMS(KP1_978353019, T8s, T7P);
|
Chris@10
|
1344 R1[WS(rs, 17)] = FNMS(KP1_978353019, T8u, T8t);
|
Chris@10
|
1345 R1[WS(rs, 49)] = FMA(KP1_978353019, T8u, T8t);
|
Chris@10
|
1346 }
|
Chris@10
|
1347 }
|
Chris@10
|
1348 {
|
Chris@10
|
1349 E T8x, T8A, T8J, T8Q;
|
Chris@10
|
1350 T8B = FMA(KP1_913880671, T8w, T8v);
|
Chris@10
|
1351 T8x = FNMS(KP1_913880671, T8w, T8v);
|
Chris@10
|
1352 T8A = FNMS(KP741650546, T8z, T8y);
|
Chris@10
|
1353 T8C = FMA(KP741650546, T8y, T8z);
|
Chris@10
|
1354 T8R = FMA(KP1_763842528, T8I, T8F);
|
Chris@10
|
1355 T8J = FNMS(KP1_763842528, T8I, T8F);
|
Chris@10
|
1356 T8Q = FNMS(KP599376933, T8P, T8M);
|
Chris@10
|
1357 T8S = FMA(KP599376933, T8M, T8P);
|
Chris@10
|
1358 R1[WS(rs, 41)] = FMA(KP1_606415062, T8A, T8x);
|
Chris@10
|
1359 R1[WS(rs, 9)] = FNMS(KP1_606415062, T8A, T8x);
|
Chris@10
|
1360 R1[WS(rs, 5)] = FMA(KP1_715457220, T8Q, T8J);
|
Chris@10
|
1361 R1[WS(rs, 37)] = FNMS(KP1_715457220, T8Q, T8J);
|
Chris@10
|
1362 }
|
Chris@10
|
1363 }
|
Chris@10
|
1364 }
|
Chris@10
|
1365 {
|
Chris@10
|
1366 E T6R, T6S, T77, T78;
|
Chris@10
|
1367 {
|
Chris@10
|
1368 E T6L, T6h, T6M, T6o, T73, T6D, T7a, T6Y, T79, T6V, T70, T6s, T6y, T6v, T71;
|
Chris@10
|
1369 E T6G;
|
Chris@10
|
1370 {
|
Chris@10
|
1371 E T6T, T6U, T6X, T6W, T6E, T6F;
|
Chris@10
|
1372 {
|
Chris@10
|
1373 E T6d, T6g, T6k, T6n;
|
Chris@10
|
1374 T6T = FMA(KP1_847759065, T6c, T6b);
|
Chris@10
|
1375 T6d = FNMS(KP1_847759065, T6c, T6b);
|
Chris@10
|
1376 R1[WS(rs, 53)] = FMA(KP1_715457220, T8S, T8R);
|
Chris@10
|
1377 R1[WS(rs, 21)] = FNMS(KP1_715457220, T8S, T8R);
|
Chris@10
|
1378 R1[WS(rs, 57)] = FMA(KP1_606415062, T8C, T8B);
|
Chris@10
|
1379 R1[WS(rs, 25)] = FNMS(KP1_606415062, T8C, T8B);
|
Chris@10
|
1380 T6g = FNMS(KP668178637, T6f, T6e);
|
Chris@10
|
1381 T6U = FMA(KP668178637, T6e, T6f);
|
Chris@10
|
1382 T6X = FMA(KP923879532, T6j, T6i);
|
Chris@10
|
1383 T6k = FNMS(KP923879532, T6j, T6i);
|
Chris@10
|
1384 T6n = FNMS(KP923879532, T6m, T6l);
|
Chris@10
|
1385 T6W = FMA(KP923879532, T6m, T6l);
|
Chris@10
|
1386 T6L = FMA(KP1_662939224, T6g, T6d);
|
Chris@10
|
1387 T6h = FNMS(KP1_662939224, T6g, T6d);
|
Chris@10
|
1388 T6M = FMA(KP534511135, T6k, T6n);
|
Chris@10
|
1389 T6o = FNMS(KP534511135, T6n, T6k);
|
Chris@10
|
1390 }
|
Chris@10
|
1391 T73 = FMA(KP923879532, T6C, T6B);
|
Chris@10
|
1392 T6D = FNMS(KP923879532, T6C, T6B);
|
Chris@10
|
1393 T7a = FMA(KP303346683, T6W, T6X);
|
Chris@10
|
1394 T6Y = FNMS(KP303346683, T6X, T6W);
|
Chris@10
|
1395 T79 = FMA(KP1_662939224, T6U, T6T);
|
Chris@10
|
1396 T6V = FNMS(KP1_662939224, T6U, T6T);
|
Chris@10
|
1397 T70 = FMA(KP923879532, T6r, T6q);
|
Chris@10
|
1398 T6s = FNMS(KP923879532, T6r, T6q);
|
Chris@10
|
1399 T6y = FNMS(KP668178637, T6x, T6w);
|
Chris@10
|
1400 T6E = FMA(KP668178637, T6w, T6x);
|
Chris@10
|
1401 T6F = FNMS(KP668178637, T6t, T6u);
|
Chris@10
|
1402 T6v = FMA(KP668178637, T6u, T6t);
|
Chris@10
|
1403 T71 = T6E + T6F;
|
Chris@10
|
1404 T6G = T6E - T6F;
|
Chris@10
|
1405 }
|
Chris@10
|
1406 {
|
Chris@10
|
1407 E T72, T6O, T75, T6P;
|
Chris@10
|
1408 {
|
Chris@10
|
1409 E T6J, T6p, T7d, T6H, T74, T6z, T7b, T7f, T7c, T6A;
|
Chris@10
|
1410 T6J = FNMS(KP1_763842528, T6o, T6h);
|
Chris@10
|
1411 T6p = FMA(KP1_763842528, T6o, T6h);
|
Chris@10
|
1412 T72 = FNMS(KP831469612, T71, T70);
|
Chris@10
|
1413 T7d = FMA(KP831469612, T71, T70);
|
Chris@10
|
1414 T6O = FNMS(KP831469612, T6G, T6D);
|
Chris@10
|
1415 T6H = FMA(KP831469612, T6G, T6D);
|
Chris@10
|
1416 T74 = T6y + T6v;
|
Chris@10
|
1417 T6z = T6v - T6y;
|
Chris@10
|
1418 T7b = FNMS(KP1_913880671, T7a, T79);
|
Chris@10
|
1419 T7f = FMA(KP1_913880671, T7a, T79);
|
Chris@10
|
1420 T75 = FNMS(KP831469612, T74, T73);
|
Chris@10
|
1421 T7c = FMA(KP831469612, T74, T73);
|
Chris@10
|
1422 T6P = FNMS(KP831469612, T6z, T6s);
|
Chris@10
|
1423 T6A = FMA(KP831469612, T6z, T6s);
|
Chris@10
|
1424 {
|
Chris@10
|
1425 E T7e, T7g, T6K, T6I;
|
Chris@10
|
1426 T7e = FNMS(KP148335987, T7d, T7c);
|
Chris@10
|
1427 T7g = FMA(KP148335987, T7c, T7d);
|
Chris@10
|
1428 T6K = FMA(KP250486960, T6A, T6H);
|
Chris@10
|
1429 T6I = FNMS(KP250486960, T6H, T6A);
|
Chris@10
|
1430 R1[WS(rs, 62)] = FMA(KP1_978353019, T7g, T7f);
|
Chris@10
|
1431 R1[WS(rs, 30)] = FNMS(KP1_978353019, T7g, T7f);
|
Chris@10
|
1432 R1[WS(rs, 46)] = FMA(KP1_978353019, T7e, T7b);
|
Chris@10
|
1433 R1[WS(rs, 14)] = FNMS(KP1_978353019, T7e, T7b);
|
Chris@10
|
1434 R1[WS(rs, 2)] = FMA(KP1_940062506, T6I, T6p);
|
Chris@10
|
1435 R1[WS(rs, 34)] = FNMS(KP1_940062506, T6I, T6p);
|
Chris@10
|
1436 R1[WS(rs, 18)] = FNMS(KP1_940062506, T6K, T6J);
|
Chris@10
|
1437 R1[WS(rs, 50)] = FMA(KP1_940062506, T6K, T6J);
|
Chris@10
|
1438 }
|
Chris@10
|
1439 }
|
Chris@10
|
1440 {
|
Chris@10
|
1441 E T6N, T6Q, T6Z, T76;
|
Chris@10
|
1442 T6R = FMA(KP1_763842528, T6M, T6L);
|
Chris@10
|
1443 T6N = FNMS(KP1_763842528, T6M, T6L);
|
Chris@10
|
1444 T6Q = FNMS(KP599376933, T6P, T6O);
|
Chris@10
|
1445 T6S = FMA(KP599376933, T6O, T6P);
|
Chris@10
|
1446 T77 = FMA(KP1_913880671, T6Y, T6V);
|
Chris@10
|
1447 T6Z = FNMS(KP1_913880671, T6Y, T6V);
|
Chris@10
|
1448 T76 = FNMS(KP741650546, T75, T72);
|
Chris@10
|
1449 T78 = FMA(KP741650546, T72, T75);
|
Chris@10
|
1450 R1[WS(rs, 42)] = FMA(KP1_715457220, T6Q, T6N);
|
Chris@10
|
1451 R1[WS(rs, 10)] = FNMS(KP1_715457220, T6Q, T6N);
|
Chris@10
|
1452 R1[WS(rs, 6)] = FMA(KP1_606415062, T76, T6Z);
|
Chris@10
|
1453 R1[WS(rs, 38)] = FNMS(KP1_606415062, T76, T6Z);
|
Chris@10
|
1454 }
|
Chris@10
|
1455 }
|
Chris@10
|
1456 }
|
Chris@10
|
1457 {
|
Chris@10
|
1458 E T9B, T97, T9C, T9e, T9T, T9t, Ta0, T9O, T9Z, T9L, T9Q, T9i, T9l, T9o, T9R;
|
Chris@10
|
1459 E T9w;
|
Chris@10
|
1460 {
|
Chris@10
|
1461 E T9J, T9K, T9N, T9M, T9u, T9v;
|
Chris@10
|
1462 {
|
Chris@10
|
1463 E T93, T96, T9a, T9d;
|
Chris@10
|
1464 T9J = FMA(KP1_847759065, T92, T91);
|
Chris@10
|
1465 T93 = FNMS(KP1_847759065, T92, T91);
|
Chris@10
|
1466 R1[WS(rs, 54)] = FMA(KP1_606415062, T78, T77);
|
Chris@10
|
1467 R1[WS(rs, 22)] = FNMS(KP1_606415062, T78, T77);
|
Chris@10
|
1468 R1[WS(rs, 58)] = FMA(KP1_715457220, T6S, T6R);
|
Chris@10
|
1469 R1[WS(rs, 26)] = FNMS(KP1_715457220, T6S, T6R);
|
Chris@10
|
1470 T96 = FNMS(KP198912367, T95, T94);
|
Chris@10
|
1471 T9K = FMA(KP198912367, T94, T95);
|
Chris@10
|
1472 T9N = FMA(KP923879532, T99, T98);
|
Chris@10
|
1473 T9a = FNMS(KP923879532, T99, T98);
|
Chris@10
|
1474 T9d = FNMS(KP923879532, T9c, T9b);
|
Chris@10
|
1475 T9M = FMA(KP923879532, T9c, T9b);
|
Chris@10
|
1476 T9B = FMA(KP1_961570560, T96, T93);
|
Chris@10
|
1477 T97 = FNMS(KP1_961570560, T96, T93);
|
Chris@10
|
1478 T9C = FMA(KP820678790, T9a, T9d);
|
Chris@10
|
1479 T9e = FNMS(KP820678790, T9d, T9a);
|
Chris@10
|
1480 }
|
Chris@10
|
1481 T9T = FMA(KP923879532, T9s, T9r);
|
Chris@10
|
1482 T9t = FNMS(KP923879532, T9s, T9r);
|
Chris@10
|
1483 Ta0 = FMA(KP098491403, T9M, T9N);
|
Chris@10
|
1484 T9O = FNMS(KP098491403, T9N, T9M);
|
Chris@10
|
1485 T9Z = FMA(KP1_961570560, T9K, T9J);
|
Chris@10
|
1486 T9L = FNMS(KP1_961570560, T9K, T9J);
|
Chris@10
|
1487 T9Q = FMA(KP923879532, T9h, T9g);
|
Chris@10
|
1488 T9i = FNMS(KP923879532, T9h, T9g);
|
Chris@10
|
1489 T9l = FNMS(KP198912367, T9k, T9j);
|
Chris@10
|
1490 T9u = FMA(KP198912367, T9j, T9k);
|
Chris@10
|
1491 T9v = FMA(KP198912367, T9m, T9n);
|
Chris@10
|
1492 T9o = FNMS(KP198912367, T9n, T9m);
|
Chris@10
|
1493 T9R = T9u + T9v;
|
Chris@10
|
1494 T9w = T9u - T9v;
|
Chris@10
|
1495 }
|
Chris@10
|
1496 {
|
Chris@10
|
1497 E T9S, T9E, T9V, T9F;
|
Chris@10
|
1498 {
|
Chris@10
|
1499 E T9z, T9f, Ta3, T9x, T9U, T9p, Ta1, Ta5, Ta2, T9q;
|
Chris@10
|
1500 T9z = FNMS(KP1_546020906, T9e, T97);
|
Chris@10
|
1501 T9f = FMA(KP1_546020906, T9e, T97);
|
Chris@10
|
1502 T9S = FNMS(KP980785280, T9R, T9Q);
|
Chris@10
|
1503 Ta3 = FMA(KP980785280, T9R, T9Q);
|
Chris@10
|
1504 T9E = FNMS(KP980785280, T9w, T9t);
|
Chris@10
|
1505 T9x = FMA(KP980785280, T9w, T9t);
|
Chris@10
|
1506 T9U = T9l - T9o;
|
Chris@10
|
1507 T9p = T9l + T9o;
|
Chris@10
|
1508 Ta1 = FNMS(KP1_990369453, Ta0, T9Z);
|
Chris@10
|
1509 Ta5 = FMA(KP1_990369453, Ta0, T9Z);
|
Chris@10
|
1510 T9V = FNMS(KP980785280, T9U, T9T);
|
Chris@10
|
1511 Ta2 = FMA(KP980785280, T9U, T9T);
|
Chris@10
|
1512 T9F = FMA(KP980785280, T9p, T9i);
|
Chris@10
|
1513 T9q = FNMS(KP980785280, T9p, T9i);
|
Chris@10
|
1514 {
|
Chris@10
|
1515 E Ta4, Ta6, T9A, T9y;
|
Chris@10
|
1516 Ta4 = FNMS(KP049126849, Ta3, Ta2);
|
Chris@10
|
1517 Ta6 = FMA(KP049126849, Ta2, Ta3);
|
Chris@10
|
1518 T9A = FMA(KP357805721, T9q, T9x);
|
Chris@10
|
1519 T9y = FNMS(KP357805721, T9x, T9q);
|
Chris@10
|
1520 R1[WS(rs, 63)] = FMA(KP1_997590912, Ta6, Ta5);
|
Chris@10
|
1521 R1[WS(rs, 31)] = FNMS(KP1_997590912, Ta6, Ta5);
|
Chris@10
|
1522 R1[WS(rs, 47)] = FMA(KP1_997590912, Ta4, Ta1);
|
Chris@10
|
1523 R1[WS(rs, 15)] = FNMS(KP1_997590912, Ta4, Ta1);
|
Chris@10
|
1524 R1[WS(rs, 3)] = FMA(KP1_883088130, T9y, T9f);
|
Chris@10
|
1525 R1[WS(rs, 35)] = FNMS(KP1_883088130, T9y, T9f);
|
Chris@10
|
1526 R1[WS(rs, 19)] = FNMS(KP1_883088130, T9A, T9z);
|
Chris@10
|
1527 R1[WS(rs, 51)] = FMA(KP1_883088130, T9A, T9z);
|
Chris@10
|
1528 }
|
Chris@10
|
1529 }
|
Chris@10
|
1530 {
|
Chris@10
|
1531 E T9D, T9G, T9P, T9W;
|
Chris@10
|
1532 T9H = FMA(KP1_546020906, T9C, T9B);
|
Chris@10
|
1533 T9D = FNMS(KP1_546020906, T9C, T9B);
|
Chris@10
|
1534 T9G = FNMS(KP472964775, T9F, T9E);
|
Chris@10
|
1535 T9I = FMA(KP472964775, T9E, T9F);
|
Chris@10
|
1536 T9X = FMA(KP1_990369453, T9O, T9L);
|
Chris@10
|
1537 T9P = FNMS(KP1_990369453, T9O, T9L);
|
Chris@10
|
1538 T9W = FNMS(KP906347169, T9V, T9S);
|
Chris@10
|
1539 T9Y = FMA(KP906347169, T9S, T9V);
|
Chris@10
|
1540 R1[WS(rs, 43)] = FMA(KP1_807978586, T9G, T9D);
|
Chris@10
|
1541 R1[WS(rs, 11)] = FNMS(KP1_807978586, T9G, T9D);
|
Chris@10
|
1542 R1[WS(rs, 7)] = FMA(KP1_481902250, T9W, T9P);
|
Chris@10
|
1543 R1[WS(rs, 39)] = FNMS(KP1_481902250, T9W, T9P);
|
Chris@10
|
1544 }
|
Chris@10
|
1545 }
|
Chris@10
|
1546 }
|
Chris@10
|
1547 }
|
Chris@10
|
1548 }
|
Chris@10
|
1549 }
|
Chris@10
|
1550 }
|
Chris@10
|
1551 }
|
Chris@10
|
1552 R1[WS(rs, 55)] = FMA(KP1_481902250, T9Y, T9X);
|
Chris@10
|
1553 R1[WS(rs, 23)] = FNMS(KP1_481902250, T9Y, T9X);
|
Chris@10
|
1554 R1[WS(rs, 59)] = FMA(KP1_807978586, T9I, T9H);
|
Chris@10
|
1555 R1[WS(rs, 27)] = FNMS(KP1_807978586, T9I, T9H);
|
Chris@10
|
1556 }
|
Chris@10
|
1557 }
|
Chris@10
|
1558 }
|
Chris@10
|
1559
|
Chris@10
|
1560 static const kr2c_desc desc = { 128, "r2cb_128", {416, 0, 540, 0}, &GENUS };
|
Chris@10
|
1561
|
Chris@10
|
1562 void X(codelet_r2cb_128) (planner *p) {
|
Chris@10
|
1563 X(kr2c_register) (p, r2cb_128, &desc);
|
Chris@10
|
1564 }
|
Chris@10
|
1565
|
Chris@10
|
1566 #else /* HAVE_FMA */
|
Chris@10
|
1567
|
Chris@10
|
1568 /* Generated by: ../../../genfft/gen_r2cb.native -compact -variables 4 -pipeline-latency 4 -sign 1 -n 128 -name r2cb_128 -include r2cb.h */
|
Chris@10
|
1569
|
Chris@10
|
1570 /*
|
Chris@10
|
1571 * This function contains 956 FP additions, 342 FP multiplications,
|
Chris@10
|
1572 * (or, 812 additions, 198 multiplications, 144 fused multiply/add),
|
Chris@10
|
1573 * 198 stack variables, 39 constants, and 256 memory accesses
|
Chris@10
|
1574 */
|
Chris@10
|
1575 #include "r2cb.h"
|
Chris@10
|
1576
|
Chris@10
|
1577 static void r2cb_128(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
|
Chris@10
|
1578 {
|
Chris@10
|
1579 DK(KP1_028205488, +1.028205488386443453187387677937631545216098241);
|
Chris@10
|
1580 DK(KP1_715457220, +1.715457220000544139804539968569540274084981599);
|
Chris@10
|
1581 DK(KP1_606415062, +1.606415062961289819613353025926283847759138854);
|
Chris@10
|
1582 DK(KP1_191398608, +1.191398608984866686934073057659939779023852677);
|
Chris@10
|
1583 DK(KP1_940062506, +1.940062506389087985207968414572200502913731924);
|
Chris@10
|
1584 DK(KP485960359, +0.485960359806527779896548324154942236641981567);
|
Chris@10
|
1585 DK(KP293460948, +0.293460948910723503317700259293435639412430633);
|
Chris@10
|
1586 DK(KP1_978353019, +1.978353019929561946903347476032486127967379067);
|
Chris@10
|
1587 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
Chris@10
|
1588 DK(KP555570233, +0.555570233019602224742830813948532874374937191);
|
Chris@10
|
1589 DK(KP855110186, +0.855110186860564188641933713777597068609157259);
|
Chris@10
|
1590 DK(KP1_807978586, +1.807978586246886663172400594461074097420264050);
|
Chris@10
|
1591 DK(KP1_481902250, +1.481902250709918182351233794990325459457910619);
|
Chris@10
|
1592 DK(KP1_343117909, +1.343117909694036801250753700854843606457501264);
|
Chris@10
|
1593 DK(KP1_883088130, +1.883088130366041556825018805199004714371179592);
|
Chris@10
|
1594 DK(KP673779706, +0.673779706784440101378506425238295140955533559);
|
Chris@10
|
1595 DK(KP098135348, +0.098135348654836028509909953885365316629490726);
|
Chris@10
|
1596 DK(KP1_997590912, +1.997590912410344785429543209518201388886407229);
|
Chris@10
|
1597 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
Chris@10
|
1598 DK(KP195090322, +0.195090322016128267848284868477022240927691618);
|
Chris@10
|
1599 DK(KP580569354, +0.580569354508924735272384751634790549382952557);
|
Chris@10
|
1600 DK(KP1_913880671, +1.913880671464417729871595773960539938965698411);
|
Chris@10
|
1601 DK(KP942793473, +0.942793473651995297112775251810508755314920638);
|
Chris@10
|
1602 DK(KP1_763842528, +1.763842528696710059425513727320776699016885241);
|
Chris@10
|
1603 DK(KP1_111140466, +1.111140466039204449485661627897065748749874382);
|
Chris@10
|
1604 DK(KP1_662939224, +1.662939224605090474157576755235811513477121624);
|
Chris@10
|
1605 DK(KP1_268786568, +1.268786568327290996430343226450986741351374190);
|
Chris@10
|
1606 DK(KP1_546020906, +1.546020906725473921621813219516939601942082586);
|
Chris@10
|
1607 DK(KP196034280, +0.196034280659121203988391127777283691722273346);
|
Chris@10
|
1608 DK(KP1_990369453, +1.990369453344393772489673906218959843150949737);
|
Chris@10
|
1609 DK(KP390180644, +0.390180644032256535696569736954044481855383236);
|
Chris@10
|
1610 DK(KP1_961570560, +1.961570560806460898252364472268478073947867462);
|
Chris@10
|
1611 DK(KP382683432, +0.382683432365089771728459984030398866761344562);
|
Chris@10
|
1612 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
Chris@10
|
1613 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
Chris@10
|
1614 DK(KP765366864, +0.765366864730179543456919968060797733522689125);
|
Chris@10
|
1615 DK(KP1_847759065, +1.847759065022573512256366378793576573644833252);
|
Chris@10
|
1616 DK(KP1_414213562, +1.414213562373095048801688724209698078569671875);
|
Chris@10
|
1617 DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
|
Chris@10
|
1618 {
|
Chris@10
|
1619 INT i;
|
Chris@10
|
1620 for (i = v; i > 0; i = i - 1, R0 = R0 + ovs, R1 = R1 + ovs, Cr = Cr + ivs, Ci = Ci + ivs, MAKE_VOLATILE_STRIDE(512, rs), MAKE_VOLATILE_STRIDE(512, csr), MAKE_VOLATILE_STRIDE(512, csi)) {
|
Chris@10
|
1621 E Ta, T6q, T2a, T5k, T8x, Tbx, TcF, Ten, Th, T6r, T2j, T5l, T8E, Tby, TcI;
|
Chris@10
|
1622 E Teo, Tx, T6t, TcM, Teq, TcP, Ter, T2t, T5n, T2C, T5o, T8Q, TbA, T8X, TbB;
|
Chris@10
|
1623 E T6w, T7L, T1j, T6L, Tde, TeC, TdL, TeR, T3v, T5z, T4I, T5O, T9O, TbM, TaV;
|
Chris@10
|
1624 E Tc1, T78, T7Z, TN, T6z, TcU, Teu, Td8, Tey, T2N, T5r, T3j, T5v, T9a, TbE;
|
Chris@10
|
1625 E T9A, TbI, T6H, T7O, T1O, T7V, T48, T4u, Tds, TeG, T5E, T5K, Taf, TbP, Tdp;
|
Chris@10
|
1626 E TeF, T6U, T72, Tam, TbQ, T23, T7U, T4r, T4v, Tdz, TeJ, T5H, T5L, Tay, TbS;
|
Chris@10
|
1627 E Tdw, TeI, T6Z, T73, TaF, TbT, T1y, T75, Tdl, TeQ, TdI, TeD, T3O, T5N, T4z;
|
Chris@10
|
1628 E T5A, Ta3, Tc0, TaO, TbN, T6O, T80, T12, T6E, Td1, Tex, Td5, Tev, T36, T5u;
|
Chris@10
|
1629 E T3a, T5s, T9p, TbH, T9t, TbF, T6C, T7P;
|
Chris@10
|
1630 {
|
Chris@10
|
1631 E T5, T8s, T3, T8q, T9, T8u, T29, T8v, T6, T26;
|
Chris@10
|
1632 {
|
Chris@10
|
1633 E T4, T8r, T1, T2;
|
Chris@10
|
1634 T4 = Cr[WS(csr, 32)];
|
Chris@10
|
1635 T5 = KP2_000000000 * T4;
|
Chris@10
|
1636 T8r = Ci[WS(csi, 32)];
|
Chris@10
|
1637 T8s = KP2_000000000 * T8r;
|
Chris@10
|
1638 T1 = Cr[0];
|
Chris@10
|
1639 T2 = Cr[WS(csr, 64)];
|
Chris@10
|
1640 T3 = T1 + T2;
|
Chris@10
|
1641 T8q = T1 - T2;
|
Chris@10
|
1642 {
|
Chris@10
|
1643 E T7, T8, T27, T28;
|
Chris@10
|
1644 T7 = Cr[WS(csr, 16)];
|
Chris@10
|
1645 T8 = Cr[WS(csr, 48)];
|
Chris@10
|
1646 T9 = KP2_000000000 * (T7 + T8);
|
Chris@10
|
1647 T8u = T7 - T8;
|
Chris@10
|
1648 T27 = Ci[WS(csi, 16)];
|
Chris@10
|
1649 T28 = Ci[WS(csi, 48)];
|
Chris@10
|
1650 T29 = KP2_000000000 * (T27 - T28);
|
Chris@10
|
1651 T8v = T27 + T28;
|
Chris@10
|
1652 }
|
Chris@10
|
1653 }
|
Chris@10
|
1654 T6 = T3 + T5;
|
Chris@10
|
1655 Ta = T6 + T9;
|
Chris@10
|
1656 T6q = T6 - T9;
|
Chris@10
|
1657 T26 = T3 - T5;
|
Chris@10
|
1658 T2a = T26 - T29;
|
Chris@10
|
1659 T5k = T26 + T29;
|
Chris@10
|
1660 {
|
Chris@10
|
1661 E T8t, T8w, TcD, TcE;
|
Chris@10
|
1662 T8t = T8q - T8s;
|
Chris@10
|
1663 T8w = KP1_414213562 * (T8u - T8v);
|
Chris@10
|
1664 T8x = T8t + T8w;
|
Chris@10
|
1665 Tbx = T8t - T8w;
|
Chris@10
|
1666 TcD = T8q + T8s;
|
Chris@10
|
1667 TcE = KP1_414213562 * (T8u + T8v);
|
Chris@10
|
1668 TcF = TcD - TcE;
|
Chris@10
|
1669 Ten = TcD + TcE;
|
Chris@10
|
1670 }
|
Chris@10
|
1671 }
|
Chris@10
|
1672 {
|
Chris@10
|
1673 E Td, T8y, T2e, T8C, Tg, T8B, T2h, T8z, T2b, T2i;
|
Chris@10
|
1674 {
|
Chris@10
|
1675 E Tb, Tc, T2c, T2d;
|
Chris@10
|
1676 Tb = Cr[WS(csr, 8)];
|
Chris@10
|
1677 Tc = Cr[WS(csr, 56)];
|
Chris@10
|
1678 Td = Tb + Tc;
|
Chris@10
|
1679 T8y = Tb - Tc;
|
Chris@10
|
1680 T2c = Ci[WS(csi, 8)];
|
Chris@10
|
1681 T2d = Ci[WS(csi, 56)];
|
Chris@10
|
1682 T2e = T2c - T2d;
|
Chris@10
|
1683 T8C = T2c + T2d;
|
Chris@10
|
1684 }
|
Chris@10
|
1685 {
|
Chris@10
|
1686 E Te, Tf, T2f, T2g;
|
Chris@10
|
1687 Te = Cr[WS(csr, 40)];
|
Chris@10
|
1688 Tf = Cr[WS(csr, 24)];
|
Chris@10
|
1689 Tg = Te + Tf;
|
Chris@10
|
1690 T8B = Te - Tf;
|
Chris@10
|
1691 T2f = Ci[WS(csi, 40)];
|
Chris@10
|
1692 T2g = Ci[WS(csi, 24)];
|
Chris@10
|
1693 T2h = T2f - T2g;
|
Chris@10
|
1694 T8z = T2f + T2g;
|
Chris@10
|
1695 }
|
Chris@10
|
1696 Th = KP2_000000000 * (Td + Tg);
|
Chris@10
|
1697 T6r = KP2_000000000 * (T2h + T2e);
|
Chris@10
|
1698 T2b = Td - Tg;
|
Chris@10
|
1699 T2i = T2e - T2h;
|
Chris@10
|
1700 T2j = KP1_414213562 * (T2b - T2i);
|
Chris@10
|
1701 T5l = KP1_414213562 * (T2b + T2i);
|
Chris@10
|
1702 {
|
Chris@10
|
1703 E T8A, T8D, TcG, TcH;
|
Chris@10
|
1704 T8A = T8y - T8z;
|
Chris@10
|
1705 T8D = T8B + T8C;
|
Chris@10
|
1706 T8E = FNMS(KP765366864, T8D, KP1_847759065 * T8A);
|
Chris@10
|
1707 Tby = FMA(KP765366864, T8A, KP1_847759065 * T8D);
|
Chris@10
|
1708 TcG = T8y + T8z;
|
Chris@10
|
1709 TcH = T8C - T8B;
|
Chris@10
|
1710 TcI = FNMS(KP1_847759065, TcH, KP765366864 * TcG);
|
Chris@10
|
1711 Teo = FMA(KP1_847759065, TcG, KP765366864 * TcH);
|
Chris@10
|
1712 }
|
Chris@10
|
1713 }
|
Chris@10
|
1714 {
|
Chris@10
|
1715 E Tl, T8G, T2x, T8V, To, T8U, T2A, T8H, Tv, T8S, T2o, T8O, Ts, T8R, T2r;
|
Chris@10
|
1716 E T8L;
|
Chris@10
|
1717 {
|
Chris@10
|
1718 E Tj, Tk, T2y, T2z;
|
Chris@10
|
1719 Tj = Cr[WS(csr, 4)];
|
Chris@10
|
1720 Tk = Cr[WS(csr, 60)];
|
Chris@10
|
1721 Tl = Tj + Tk;
|
Chris@10
|
1722 T8G = Tj - Tk;
|
Chris@10
|
1723 {
|
Chris@10
|
1724 E T2v, T2w, Tm, Tn;
|
Chris@10
|
1725 T2v = Ci[WS(csi, 4)];
|
Chris@10
|
1726 T2w = Ci[WS(csi, 60)];
|
Chris@10
|
1727 T2x = T2v - T2w;
|
Chris@10
|
1728 T8V = T2v + T2w;
|
Chris@10
|
1729 Tm = Cr[WS(csr, 36)];
|
Chris@10
|
1730 Tn = Cr[WS(csr, 28)];
|
Chris@10
|
1731 To = Tm + Tn;
|
Chris@10
|
1732 T8U = Tm - Tn;
|
Chris@10
|
1733 }
|
Chris@10
|
1734 T2y = Ci[WS(csi, 36)];
|
Chris@10
|
1735 T2z = Ci[WS(csi, 28)];
|
Chris@10
|
1736 T2A = T2y - T2z;
|
Chris@10
|
1737 T8H = T2y + T2z;
|
Chris@10
|
1738 {
|
Chris@10
|
1739 E Tt, Tu, T8M, T2m, T2n, T8N;
|
Chris@10
|
1740 Tt = Cr[WS(csr, 12)];
|
Chris@10
|
1741 Tu = Cr[WS(csr, 52)];
|
Chris@10
|
1742 T8M = Tt - Tu;
|
Chris@10
|
1743 T2m = Ci[WS(csi, 52)];
|
Chris@10
|
1744 T2n = Ci[WS(csi, 12)];
|
Chris@10
|
1745 T8N = T2n + T2m;
|
Chris@10
|
1746 Tv = Tt + Tu;
|
Chris@10
|
1747 T8S = T8M + T8N;
|
Chris@10
|
1748 T2o = T2m - T2n;
|
Chris@10
|
1749 T8O = T8M - T8N;
|
Chris@10
|
1750 }
|
Chris@10
|
1751 {
|
Chris@10
|
1752 E Tq, Tr, T8J, T2p, T2q, T8K;
|
Chris@10
|
1753 Tq = Cr[WS(csr, 20)];
|
Chris@10
|
1754 Tr = Cr[WS(csr, 44)];
|
Chris@10
|
1755 T8J = Tq - Tr;
|
Chris@10
|
1756 T2p = Ci[WS(csi, 20)];
|
Chris@10
|
1757 T2q = Ci[WS(csi, 44)];
|
Chris@10
|
1758 T8K = T2p + T2q;
|
Chris@10
|
1759 Ts = Tq + Tr;
|
Chris@10
|
1760 T8R = T8J + T8K;
|
Chris@10
|
1761 T2r = T2p - T2q;
|
Chris@10
|
1762 T8L = T8J - T8K;
|
Chris@10
|
1763 }
|
Chris@10
|
1764 }
|
Chris@10
|
1765 {
|
Chris@10
|
1766 E Tp, Tw, TcK, TcL;
|
Chris@10
|
1767 Tp = Tl + To;
|
Chris@10
|
1768 Tw = Ts + Tv;
|
Chris@10
|
1769 Tx = KP2_000000000 * (Tp + Tw);
|
Chris@10
|
1770 T6t = Tp - Tw;
|
Chris@10
|
1771 TcK = T8G + T8H;
|
Chris@10
|
1772 TcL = KP707106781 * (T8R + T8S);
|
Chris@10
|
1773 TcM = TcK - TcL;
|
Chris@10
|
1774 Teq = TcK + TcL;
|
Chris@10
|
1775 }
|
Chris@10
|
1776 {
|
Chris@10
|
1777 E TcN, TcO, T2l, T2s;
|
Chris@10
|
1778 TcN = KP707106781 * (T8L - T8O);
|
Chris@10
|
1779 TcO = T8V - T8U;
|
Chris@10
|
1780 TcP = TcN + TcO;
|
Chris@10
|
1781 Ter = TcO - TcN;
|
Chris@10
|
1782 T2l = Tl - To;
|
Chris@10
|
1783 T2s = T2o - T2r;
|
Chris@10
|
1784 T2t = T2l + T2s;
|
Chris@10
|
1785 T5n = T2l - T2s;
|
Chris@10
|
1786 }
|
Chris@10
|
1787 {
|
Chris@10
|
1788 E T2u, T2B, T8I, T8P;
|
Chris@10
|
1789 T2u = Ts - Tv;
|
Chris@10
|
1790 T2B = T2x - T2A;
|
Chris@10
|
1791 T2C = T2u + T2B;
|
Chris@10
|
1792 T5o = T2B - T2u;
|
Chris@10
|
1793 T8I = T8G - T8H;
|
Chris@10
|
1794 T8P = KP707106781 * (T8L + T8O);
|
Chris@10
|
1795 T8Q = T8I + T8P;
|
Chris@10
|
1796 TbA = T8I - T8P;
|
Chris@10
|
1797 }
|
Chris@10
|
1798 {
|
Chris@10
|
1799 E T8T, T8W, T6u, T6v;
|
Chris@10
|
1800 T8T = KP707106781 * (T8R - T8S);
|
Chris@10
|
1801 T8W = T8U + T8V;
|
Chris@10
|
1802 T8X = T8T + T8W;
|
Chris@10
|
1803 TbB = T8W - T8T;
|
Chris@10
|
1804 T6u = T2A + T2x;
|
Chris@10
|
1805 T6v = T2r + T2o;
|
Chris@10
|
1806 T6w = T6u - T6v;
|
Chris@10
|
1807 T7L = KP2_000000000 * (T6v + T6u);
|
Chris@10
|
1808 }
|
Chris@10
|
1809 }
|
Chris@10
|
1810 {
|
Chris@10
|
1811 E T17, T9E, T4D, TaT, T1a, TaS, T4G, T9F, T1h, TaQ, T3q, T9M, T1e, TaP, T3t;
|
Chris@10
|
1812 E T9J;
|
Chris@10
|
1813 {
|
Chris@10
|
1814 E T15, T16, T4E, T4F;
|
Chris@10
|
1815 T15 = Cr[WS(csr, 1)];
|
Chris@10
|
1816 T16 = Cr[WS(csr, 63)];
|
Chris@10
|
1817 T17 = T15 + T16;
|
Chris@10
|
1818 T9E = T15 - T16;
|
Chris@10
|
1819 {
|
Chris@10
|
1820 E T4B, T4C, T18, T19;
|
Chris@10
|
1821 T4B = Ci[WS(csi, 1)];
|
Chris@10
|
1822 T4C = Ci[WS(csi, 63)];
|
Chris@10
|
1823 T4D = T4B - T4C;
|
Chris@10
|
1824 TaT = T4B + T4C;
|
Chris@10
|
1825 T18 = Cr[WS(csr, 33)];
|
Chris@10
|
1826 T19 = Cr[WS(csr, 31)];
|
Chris@10
|
1827 T1a = T18 + T19;
|
Chris@10
|
1828 TaS = T18 - T19;
|
Chris@10
|
1829 }
|
Chris@10
|
1830 T4E = Ci[WS(csi, 33)];
|
Chris@10
|
1831 T4F = Ci[WS(csi, 31)];
|
Chris@10
|
1832 T4G = T4E - T4F;
|
Chris@10
|
1833 T9F = T4E + T4F;
|
Chris@10
|
1834 {
|
Chris@10
|
1835 E T1f, T1g, T9K, T3o, T3p, T9L;
|
Chris@10
|
1836 T1f = Cr[WS(csr, 15)];
|
Chris@10
|
1837 T1g = Cr[WS(csr, 49)];
|
Chris@10
|
1838 T9K = T1f - T1g;
|
Chris@10
|
1839 T3o = Ci[WS(csi, 49)];
|
Chris@10
|
1840 T3p = Ci[WS(csi, 15)];
|
Chris@10
|
1841 T9L = T3p + T3o;
|
Chris@10
|
1842 T1h = T1f + T1g;
|
Chris@10
|
1843 TaQ = T9K + T9L;
|
Chris@10
|
1844 T3q = T3o - T3p;
|
Chris@10
|
1845 T9M = T9K - T9L;
|
Chris@10
|
1846 }
|
Chris@10
|
1847 {
|
Chris@10
|
1848 E T1c, T1d, T9H, T3r, T3s, T9I;
|
Chris@10
|
1849 T1c = Cr[WS(csr, 17)];
|
Chris@10
|
1850 T1d = Cr[WS(csr, 47)];
|
Chris@10
|
1851 T9H = T1c - T1d;
|
Chris@10
|
1852 T3r = Ci[WS(csi, 17)];
|
Chris@10
|
1853 T3s = Ci[WS(csi, 47)];
|
Chris@10
|
1854 T9I = T3r + T3s;
|
Chris@10
|
1855 T1e = T1c + T1d;
|
Chris@10
|
1856 TaP = T9H + T9I;
|
Chris@10
|
1857 T3t = T3r - T3s;
|
Chris@10
|
1858 T9J = T9H - T9I;
|
Chris@10
|
1859 }
|
Chris@10
|
1860 }
|
Chris@10
|
1861 {
|
Chris@10
|
1862 E T1b, T1i, Tdc, Tdd;
|
Chris@10
|
1863 T1b = T17 + T1a;
|
Chris@10
|
1864 T1i = T1e + T1h;
|
Chris@10
|
1865 T1j = T1b + T1i;
|
Chris@10
|
1866 T6L = T1b - T1i;
|
Chris@10
|
1867 Tdc = T9E + T9F;
|
Chris@10
|
1868 Tdd = KP707106781 * (TaP + TaQ);
|
Chris@10
|
1869 Tde = Tdc - Tdd;
|
Chris@10
|
1870 TeC = Tdc + Tdd;
|
Chris@10
|
1871 }
|
Chris@10
|
1872 {
|
Chris@10
|
1873 E TdJ, TdK, T3n, T3u;
|
Chris@10
|
1874 TdJ = KP707106781 * (T9J - T9M);
|
Chris@10
|
1875 TdK = TaT - TaS;
|
Chris@10
|
1876 TdL = TdJ + TdK;
|
Chris@10
|
1877 TeR = TdK - TdJ;
|
Chris@10
|
1878 T3n = T17 - T1a;
|
Chris@10
|
1879 T3u = T3q - T3t;
|
Chris@10
|
1880 T3v = T3n + T3u;
|
Chris@10
|
1881 T5z = T3n - T3u;
|
Chris@10
|
1882 }
|
Chris@10
|
1883 {
|
Chris@10
|
1884 E T4A, T4H, T9G, T9N;
|
Chris@10
|
1885 T4A = T1e - T1h;
|
Chris@10
|
1886 T4H = T4D - T4G;
|
Chris@10
|
1887 T4I = T4A + T4H;
|
Chris@10
|
1888 T5O = T4H - T4A;
|
Chris@10
|
1889 T9G = T9E - T9F;
|
Chris@10
|
1890 T9N = KP707106781 * (T9J + T9M);
|
Chris@10
|
1891 T9O = T9G + T9N;
|
Chris@10
|
1892 TbM = T9G - T9N;
|
Chris@10
|
1893 }
|
Chris@10
|
1894 {
|
Chris@10
|
1895 E TaR, TaU, T76, T77;
|
Chris@10
|
1896 TaR = KP707106781 * (TaP - TaQ);
|
Chris@10
|
1897 TaU = TaS + TaT;
|
Chris@10
|
1898 TaV = TaR + TaU;
|
Chris@10
|
1899 Tc1 = TaU - TaR;
|
Chris@10
|
1900 T76 = T4G + T4D;
|
Chris@10
|
1901 T77 = T3t + T3q;
|
Chris@10
|
1902 T78 = T76 - T77;
|
Chris@10
|
1903 T7Z = T77 + T76;
|
Chris@10
|
1904 }
|
Chris@10
|
1905 }
|
Chris@10
|
1906 {
|
Chris@10
|
1907 E TB, T90, T3e, T9y, TE, T9x, T3h, T91, TL, T9v, T2I, T98, TI, T9u, T2L;
|
Chris@10
|
1908 E T95;
|
Chris@10
|
1909 {
|
Chris@10
|
1910 E Tz, TA, T3f, T3g;
|
Chris@10
|
1911 Tz = Cr[WS(csr, 2)];
|
Chris@10
|
1912 TA = Cr[WS(csr, 62)];
|
Chris@10
|
1913 TB = Tz + TA;
|
Chris@10
|
1914 T90 = Tz - TA;
|
Chris@10
|
1915 {
|
Chris@10
|
1916 E T3c, T3d, TC, TD;
|
Chris@10
|
1917 T3c = Ci[WS(csi, 2)];
|
Chris@10
|
1918 T3d = Ci[WS(csi, 62)];
|
Chris@10
|
1919 T3e = T3c - T3d;
|
Chris@10
|
1920 T9y = T3c + T3d;
|
Chris@10
|
1921 TC = Cr[WS(csr, 34)];
|
Chris@10
|
1922 TD = Cr[WS(csr, 30)];
|
Chris@10
|
1923 TE = TC + TD;
|
Chris@10
|
1924 T9x = TC - TD;
|
Chris@10
|
1925 }
|
Chris@10
|
1926 T3f = Ci[WS(csi, 34)];
|
Chris@10
|
1927 T3g = Ci[WS(csi, 30)];
|
Chris@10
|
1928 T3h = T3f - T3g;
|
Chris@10
|
1929 T91 = T3f + T3g;
|
Chris@10
|
1930 {
|
Chris@10
|
1931 E TJ, TK, T96, T2G, T2H, T97;
|
Chris@10
|
1932 TJ = Cr[WS(csr, 14)];
|
Chris@10
|
1933 TK = Cr[WS(csr, 50)];
|
Chris@10
|
1934 T96 = TJ - TK;
|
Chris@10
|
1935 T2G = Ci[WS(csi, 50)];
|
Chris@10
|
1936 T2H = Ci[WS(csi, 14)];
|
Chris@10
|
1937 T97 = T2H + T2G;
|
Chris@10
|
1938 TL = TJ + TK;
|
Chris@10
|
1939 T9v = T96 + T97;
|
Chris@10
|
1940 T2I = T2G - T2H;
|
Chris@10
|
1941 T98 = T96 - T97;
|
Chris@10
|
1942 }
|
Chris@10
|
1943 {
|
Chris@10
|
1944 E TG, TH, T93, T2J, T2K, T94;
|
Chris@10
|
1945 TG = Cr[WS(csr, 18)];
|
Chris@10
|
1946 TH = Cr[WS(csr, 46)];
|
Chris@10
|
1947 T93 = TG - TH;
|
Chris@10
|
1948 T2J = Ci[WS(csi, 18)];
|
Chris@10
|
1949 T2K = Ci[WS(csi, 46)];
|
Chris@10
|
1950 T94 = T2J + T2K;
|
Chris@10
|
1951 TI = TG + TH;
|
Chris@10
|
1952 T9u = T93 + T94;
|
Chris@10
|
1953 T2L = T2J - T2K;
|
Chris@10
|
1954 T95 = T93 - T94;
|
Chris@10
|
1955 }
|
Chris@10
|
1956 }
|
Chris@10
|
1957 {
|
Chris@10
|
1958 E TF, TM, TcS, TcT;
|
Chris@10
|
1959 TF = TB + TE;
|
Chris@10
|
1960 TM = TI + TL;
|
Chris@10
|
1961 TN = TF + TM;
|
Chris@10
|
1962 T6z = TF - TM;
|
Chris@10
|
1963 TcS = T90 + T91;
|
Chris@10
|
1964 TcT = KP707106781 * (T9u + T9v);
|
Chris@10
|
1965 TcU = TcS - TcT;
|
Chris@10
|
1966 Teu = TcS + TcT;
|
Chris@10
|
1967 }
|
Chris@10
|
1968 {
|
Chris@10
|
1969 E Td6, Td7, T2F, T2M;
|
Chris@10
|
1970 Td6 = KP707106781 * (T95 - T98);
|
Chris@10
|
1971 Td7 = T9y - T9x;
|
Chris@10
|
1972 Td8 = Td6 + Td7;
|
Chris@10
|
1973 Tey = Td7 - Td6;
|
Chris@10
|
1974 T2F = TB - TE;
|
Chris@10
|
1975 T2M = T2I - T2L;
|
Chris@10
|
1976 T2N = T2F + T2M;
|
Chris@10
|
1977 T5r = T2F - T2M;
|
Chris@10
|
1978 }
|
Chris@10
|
1979 {
|
Chris@10
|
1980 E T3b, T3i, T92, T99;
|
Chris@10
|
1981 T3b = TI - TL;
|
Chris@10
|
1982 T3i = T3e - T3h;
|
Chris@10
|
1983 T3j = T3b + T3i;
|
Chris@10
|
1984 T5v = T3i - T3b;
|
Chris@10
|
1985 T92 = T90 - T91;
|
Chris@10
|
1986 T99 = KP707106781 * (T95 + T98);
|
Chris@10
|
1987 T9a = T92 + T99;
|
Chris@10
|
1988 TbE = T92 - T99;
|
Chris@10
|
1989 }
|
Chris@10
|
1990 {
|
Chris@10
|
1991 E T9w, T9z, T6F, T6G;
|
Chris@10
|
1992 T9w = KP707106781 * (T9u - T9v);
|
Chris@10
|
1993 T9z = T9x + T9y;
|
Chris@10
|
1994 T9A = T9w + T9z;
|
Chris@10
|
1995 TbI = T9z - T9w;
|
Chris@10
|
1996 T6F = T3h + T3e;
|
Chris@10
|
1997 T6G = T2L + T2I;
|
Chris@10
|
1998 T6H = T6F - T6G;
|
Chris@10
|
1999 T7O = T6G + T6F;
|
Chris@10
|
2000 }
|
Chris@10
|
2001 }
|
Chris@10
|
2002 {
|
Chris@10
|
2003 E T1G, Taj, T3Q, Ta5, T46, Tak, T6R, Ta6, T1N, Tag, Tah, T3X, T3Z, Taa, Tad;
|
Chris@10
|
2004 E T6S, Tdn, Tdo;
|
Chris@10
|
2005 {
|
Chris@10
|
2006 E T1A, T1B, T1C, T1D, T1E, T1F;
|
Chris@10
|
2007 T1A = Cr[WS(csr, 5)];
|
Chris@10
|
2008 T1B = Cr[WS(csr, 59)];
|
Chris@10
|
2009 T1C = T1A + T1B;
|
Chris@10
|
2010 T1D = Cr[WS(csr, 37)];
|
Chris@10
|
2011 T1E = Cr[WS(csr, 27)];
|
Chris@10
|
2012 T1F = T1D + T1E;
|
Chris@10
|
2013 T1G = T1C + T1F;
|
Chris@10
|
2014 Taj = T1D - T1E;
|
Chris@10
|
2015 T3Q = T1C - T1F;
|
Chris@10
|
2016 Ta5 = T1A - T1B;
|
Chris@10
|
2017 }
|
Chris@10
|
2018 {
|
Chris@10
|
2019 E T40, T41, T42, T43, T44, T45;
|
Chris@10
|
2020 T40 = Ci[WS(csi, 5)];
|
Chris@10
|
2021 T41 = Ci[WS(csi, 59)];
|
Chris@10
|
2022 T42 = T40 - T41;
|
Chris@10
|
2023 T43 = Ci[WS(csi, 37)];
|
Chris@10
|
2024 T44 = Ci[WS(csi, 27)];
|
Chris@10
|
2025 T45 = T43 - T44;
|
Chris@10
|
2026 T46 = T42 - T45;
|
Chris@10
|
2027 Tak = T40 + T41;
|
Chris@10
|
2028 T6R = T45 + T42;
|
Chris@10
|
2029 Ta6 = T43 + T44;
|
Chris@10
|
2030 }
|
Chris@10
|
2031 {
|
Chris@10
|
2032 E T1J, Ta8, T3W, Ta9, T1M, Tab, T3T, Tac;
|
Chris@10
|
2033 {
|
Chris@10
|
2034 E T1H, T1I, T3U, T3V;
|
Chris@10
|
2035 T1H = Cr[WS(csr, 21)];
|
Chris@10
|
2036 T1I = Cr[WS(csr, 43)];
|
Chris@10
|
2037 T1J = T1H + T1I;
|
Chris@10
|
2038 Ta8 = T1H - T1I;
|
Chris@10
|
2039 T3U = Ci[WS(csi, 21)];
|
Chris@10
|
2040 T3V = Ci[WS(csi, 43)];
|
Chris@10
|
2041 T3W = T3U - T3V;
|
Chris@10
|
2042 Ta9 = T3U + T3V;
|
Chris@10
|
2043 }
|
Chris@10
|
2044 {
|
Chris@10
|
2045 E T1K, T1L, T3R, T3S;
|
Chris@10
|
2046 T1K = Cr[WS(csr, 11)];
|
Chris@10
|
2047 T1L = Cr[WS(csr, 53)];
|
Chris@10
|
2048 T1M = T1K + T1L;
|
Chris@10
|
2049 Tab = T1K - T1L;
|
Chris@10
|
2050 T3R = Ci[WS(csi, 53)];
|
Chris@10
|
2051 T3S = Ci[WS(csi, 11)];
|
Chris@10
|
2052 T3T = T3R - T3S;
|
Chris@10
|
2053 Tac = T3S + T3R;
|
Chris@10
|
2054 }
|
Chris@10
|
2055 T1N = T1J + T1M;
|
Chris@10
|
2056 Tag = Ta8 + Ta9;
|
Chris@10
|
2057 Tah = Tab + Tac;
|
Chris@10
|
2058 T3X = T3T - T3W;
|
Chris@10
|
2059 T3Z = T1J - T1M;
|
Chris@10
|
2060 Taa = Ta8 - Ta9;
|
Chris@10
|
2061 Tad = Tab - Tac;
|
Chris@10
|
2062 T6S = T3W + T3T;
|
Chris@10
|
2063 }
|
Chris@10
|
2064 T1O = T1G + T1N;
|
Chris@10
|
2065 T7V = T6S + T6R;
|
Chris@10
|
2066 {
|
Chris@10
|
2067 E T3Y, T47, Tdq, Tdr;
|
Chris@10
|
2068 T3Y = T3Q + T3X;
|
Chris@10
|
2069 T47 = T3Z + T46;
|
Chris@10
|
2070 T48 = FNMS(KP382683432, T47, KP923879532 * T3Y);
|
Chris@10
|
2071 T4u = FMA(KP382683432, T3Y, KP923879532 * T47);
|
Chris@10
|
2072 Tdq = KP707106781 * (Taa - Tad);
|
Chris@10
|
2073 Tdr = Tak - Taj;
|
Chris@10
|
2074 Tds = Tdq + Tdr;
|
Chris@10
|
2075 TeG = Tdr - Tdq;
|
Chris@10
|
2076 }
|
Chris@10
|
2077 {
|
Chris@10
|
2078 E T5C, T5D, Ta7, Tae;
|
Chris@10
|
2079 T5C = T3Q - T3X;
|
Chris@10
|
2080 T5D = T46 - T3Z;
|
Chris@10
|
2081 T5E = FNMS(KP923879532, T5D, KP382683432 * T5C);
|
Chris@10
|
2082 T5K = FMA(KP923879532, T5C, KP382683432 * T5D);
|
Chris@10
|
2083 Ta7 = Ta5 - Ta6;
|
Chris@10
|
2084 Tae = KP707106781 * (Taa + Tad);
|
Chris@10
|
2085 Taf = Ta7 + Tae;
|
Chris@10
|
2086 TbP = Ta7 - Tae;
|
Chris@10
|
2087 }
|
Chris@10
|
2088 Tdn = Ta5 + Ta6;
|
Chris@10
|
2089 Tdo = KP707106781 * (Tag + Tah);
|
Chris@10
|
2090 Tdp = Tdn - Tdo;
|
Chris@10
|
2091 TeF = Tdn + Tdo;
|
Chris@10
|
2092 {
|
Chris@10
|
2093 E T6Q, T6T, Tai, Tal;
|
Chris@10
|
2094 T6Q = T1G - T1N;
|
Chris@10
|
2095 T6T = T6R - T6S;
|
Chris@10
|
2096 T6U = T6Q - T6T;
|
Chris@10
|
2097 T72 = T6Q + T6T;
|
Chris@10
|
2098 Tai = KP707106781 * (Tag - Tah);
|
Chris@10
|
2099 Tal = Taj + Tak;
|
Chris@10
|
2100 Tam = Tai + Tal;
|
Chris@10
|
2101 TbQ = Tal - Tai;
|
Chris@10
|
2102 }
|
Chris@10
|
2103 }
|
Chris@10
|
2104 {
|
Chris@10
|
2105 E T1V, TaC, T49, Tao, T4p, TaD, T6W, Tap, T22, Taz, TaA, T4g, T4i, Tat, Taw;
|
Chris@10
|
2106 E T6X, Tdu, Tdv;
|
Chris@10
|
2107 {
|
Chris@10
|
2108 E T1P, T1Q, T1R, T1S, T1T, T1U;
|
Chris@10
|
2109 T1P = Cr[WS(csr, 3)];
|
Chris@10
|
2110 T1Q = Cr[WS(csr, 61)];
|
Chris@10
|
2111 T1R = T1P + T1Q;
|
Chris@10
|
2112 T1S = Cr[WS(csr, 29)];
|
Chris@10
|
2113 T1T = Cr[WS(csr, 35)];
|
Chris@10
|
2114 T1U = T1S + T1T;
|
Chris@10
|
2115 T1V = T1R + T1U;
|
Chris@10
|
2116 TaC = T1S - T1T;
|
Chris@10
|
2117 T49 = T1R - T1U;
|
Chris@10
|
2118 Tao = T1P - T1Q;
|
Chris@10
|
2119 }
|
Chris@10
|
2120 {
|
Chris@10
|
2121 E T4j, T4k, T4l, T4m, T4n, T4o;
|
Chris@10
|
2122 T4j = Ci[WS(csi, 61)];
|
Chris@10
|
2123 T4k = Ci[WS(csi, 3)];
|
Chris@10
|
2124 T4l = T4j - T4k;
|
Chris@10
|
2125 T4m = Ci[WS(csi, 29)];
|
Chris@10
|
2126 T4n = Ci[WS(csi, 35)];
|
Chris@10
|
2127 T4o = T4m - T4n;
|
Chris@10
|
2128 T4p = T4l - T4o;
|
Chris@10
|
2129 TaD = T4k + T4j;
|
Chris@10
|
2130 T6W = T4o + T4l;
|
Chris@10
|
2131 Tap = T4m + T4n;
|
Chris@10
|
2132 }
|
Chris@10
|
2133 {
|
Chris@10
|
2134 E T1Y, Tar, T4f, Tas, T21, Tau, T4c, Tav;
|
Chris@10
|
2135 {
|
Chris@10
|
2136 E T1W, T1X, T4d, T4e;
|
Chris@10
|
2137 T1W = Cr[WS(csr, 13)];
|
Chris@10
|
2138 T1X = Cr[WS(csr, 51)];
|
Chris@10
|
2139 T1Y = T1W + T1X;
|
Chris@10
|
2140 Tar = T1W - T1X;
|
Chris@10
|
2141 T4d = Ci[WS(csi, 13)];
|
Chris@10
|
2142 T4e = Ci[WS(csi, 51)];
|
Chris@10
|
2143 T4f = T4d - T4e;
|
Chris@10
|
2144 Tas = T4d + T4e;
|
Chris@10
|
2145 }
|
Chris@10
|
2146 {
|
Chris@10
|
2147 E T1Z, T20, T4a, T4b;
|
Chris@10
|
2148 T1Z = Cr[WS(csr, 19)];
|
Chris@10
|
2149 T20 = Cr[WS(csr, 45)];
|
Chris@10
|
2150 T21 = T1Z + T20;
|
Chris@10
|
2151 Tau = T1Z - T20;
|
Chris@10
|
2152 T4a = Ci[WS(csi, 45)];
|
Chris@10
|
2153 T4b = Ci[WS(csi, 19)];
|
Chris@10
|
2154 T4c = T4a - T4b;
|
Chris@10
|
2155 Tav = T4b + T4a;
|
Chris@10
|
2156 }
|
Chris@10
|
2157 T22 = T1Y + T21;
|
Chris@10
|
2158 Taz = Tar + Tas;
|
Chris@10
|
2159 TaA = Tau + Tav;
|
Chris@10
|
2160 T4g = T4c - T4f;
|
Chris@10
|
2161 T4i = T1Y - T21;
|
Chris@10
|
2162 Tat = Tar - Tas;
|
Chris@10
|
2163 Taw = Tau - Tav;
|
Chris@10
|
2164 T6X = T4f + T4c;
|
Chris@10
|
2165 }
|
Chris@10
|
2166 T23 = T1V + T22;
|
Chris@10
|
2167 T7U = T6X + T6W;
|
Chris@10
|
2168 {
|
Chris@10
|
2169 E T4h, T4q, Tdx, Tdy;
|
Chris@10
|
2170 T4h = T49 + T4g;
|
Chris@10
|
2171 T4q = T4i + T4p;
|
Chris@10
|
2172 T4r = FMA(KP923879532, T4h, KP382683432 * T4q);
|
Chris@10
|
2173 T4v = FNMS(KP382683432, T4h, KP923879532 * T4q);
|
Chris@10
|
2174 Tdx = KP707106781 * (Tat - Taw);
|
Chris@10
|
2175 Tdy = TaC + TaD;
|
Chris@10
|
2176 Tdz = Tdx - Tdy;
|
Chris@10
|
2177 TeJ = Tdx + Tdy;
|
Chris@10
|
2178 }
|
Chris@10
|
2179 {
|
Chris@10
|
2180 E T5F, T5G, Taq, Tax;
|
Chris@10
|
2181 T5F = T49 - T4g;
|
Chris@10
|
2182 T5G = T4p - T4i;
|
Chris@10
|
2183 T5H = FMA(KP382683432, T5F, KP923879532 * T5G);
|
Chris@10
|
2184 T5L = FNMS(KP923879532, T5F, KP382683432 * T5G);
|
Chris@10
|
2185 Taq = Tao - Tap;
|
Chris@10
|
2186 Tax = KP707106781 * (Tat + Taw);
|
Chris@10
|
2187 Tay = Taq + Tax;
|
Chris@10
|
2188 TbS = Taq - Tax;
|
Chris@10
|
2189 }
|
Chris@10
|
2190 Tdu = Tao + Tap;
|
Chris@10
|
2191 Tdv = KP707106781 * (Taz + TaA);
|
Chris@10
|
2192 Tdw = Tdu - Tdv;
|
Chris@10
|
2193 TeI = Tdu + Tdv;
|
Chris@10
|
2194 {
|
Chris@10
|
2195 E T6V, T6Y, TaB, TaE;
|
Chris@10
|
2196 T6V = T1V - T22;
|
Chris@10
|
2197 T6Y = T6W - T6X;
|
Chris@10
|
2198 T6Z = T6V + T6Y;
|
Chris@10
|
2199 T73 = T6Y - T6V;
|
Chris@10
|
2200 TaB = KP707106781 * (Taz - TaA);
|
Chris@10
|
2201 TaE = TaC - TaD;
|
Chris@10
|
2202 TaF = TaB + TaE;
|
Chris@10
|
2203 TbT = TaE - TaB;
|
Chris@10
|
2204 }
|
Chris@10
|
2205 }
|
Chris@10
|
2206 {
|
Chris@10
|
2207 E T1m, T3z, T1p, T3C, T3w, T3D, Tdg, Tdf, T9U, T9R, T1t, T3I, T1w, T3L, T3F;
|
Chris@10
|
2208 E T3M, Tdj, Tdi, Ta1, T9Y;
|
Chris@10
|
2209 {
|
Chris@10
|
2210 E T9P, T9T, T9S, T9Q;
|
Chris@10
|
2211 {
|
Chris@10
|
2212 E T1k, T1l, T3x, T3y;
|
Chris@10
|
2213 T1k = Cr[WS(csr, 9)];
|
Chris@10
|
2214 T1l = Cr[WS(csr, 55)];
|
Chris@10
|
2215 T1m = T1k + T1l;
|
Chris@10
|
2216 T9P = T1k - T1l;
|
Chris@10
|
2217 T3x = Ci[WS(csi, 9)];
|
Chris@10
|
2218 T3y = Ci[WS(csi, 55)];
|
Chris@10
|
2219 T3z = T3x - T3y;
|
Chris@10
|
2220 T9T = T3x + T3y;
|
Chris@10
|
2221 }
|
Chris@10
|
2222 {
|
Chris@10
|
2223 E T1n, T1o, T3A, T3B;
|
Chris@10
|
2224 T1n = Cr[WS(csr, 41)];
|
Chris@10
|
2225 T1o = Cr[WS(csr, 23)];
|
Chris@10
|
2226 T1p = T1n + T1o;
|
Chris@10
|
2227 T9S = T1n - T1o;
|
Chris@10
|
2228 T3A = Ci[WS(csi, 41)];
|
Chris@10
|
2229 T3B = Ci[WS(csi, 23)];
|
Chris@10
|
2230 T3C = T3A - T3B;
|
Chris@10
|
2231 T9Q = T3A + T3B;
|
Chris@10
|
2232 }
|
Chris@10
|
2233 T3w = T1m - T1p;
|
Chris@10
|
2234 T3D = T3z - T3C;
|
Chris@10
|
2235 Tdg = T9T - T9S;
|
Chris@10
|
2236 Tdf = T9P + T9Q;
|
Chris@10
|
2237 T9U = T9S + T9T;
|
Chris@10
|
2238 T9R = T9P - T9Q;
|
Chris@10
|
2239 }
|
Chris@10
|
2240 {
|
Chris@10
|
2241 E T9W, Ta0, T9Z, T9X;
|
Chris@10
|
2242 {
|
Chris@10
|
2243 E T1r, T1s, T3G, T3H;
|
Chris@10
|
2244 T1r = Cr[WS(csr, 7)];
|
Chris@10
|
2245 T1s = Cr[WS(csr, 57)];
|
Chris@10
|
2246 T1t = T1r + T1s;
|
Chris@10
|
2247 T9W = T1r - T1s;
|
Chris@10
|
2248 T3G = Ci[WS(csi, 57)];
|
Chris@10
|
2249 T3H = Ci[WS(csi, 7)];
|
Chris@10
|
2250 T3I = T3G - T3H;
|
Chris@10
|
2251 Ta0 = T3H + T3G;
|
Chris@10
|
2252 }
|
Chris@10
|
2253 {
|
Chris@10
|
2254 E T1u, T1v, T3J, T3K;
|
Chris@10
|
2255 T1u = Cr[WS(csr, 25)];
|
Chris@10
|
2256 T1v = Cr[WS(csr, 39)];
|
Chris@10
|
2257 T1w = T1u + T1v;
|
Chris@10
|
2258 T9Z = T1u - T1v;
|
Chris@10
|
2259 T3J = Ci[WS(csi, 25)];
|
Chris@10
|
2260 T3K = Ci[WS(csi, 39)];
|
Chris@10
|
2261 T3L = T3J - T3K;
|
Chris@10
|
2262 T9X = T3J + T3K;
|
Chris@10
|
2263 }
|
Chris@10
|
2264 T3F = T1t - T1w;
|
Chris@10
|
2265 T3M = T3I - T3L;
|
Chris@10
|
2266 Tdj = T9Z + Ta0;
|
Chris@10
|
2267 Tdi = T9W + T9X;
|
Chris@10
|
2268 Ta1 = T9Z - Ta0;
|
Chris@10
|
2269 T9Y = T9W - T9X;
|
Chris@10
|
2270 }
|
Chris@10
|
2271 {
|
Chris@10
|
2272 E T1q, T1x, Tdh, Tdk;
|
Chris@10
|
2273 T1q = T1m + T1p;
|
Chris@10
|
2274 T1x = T1t + T1w;
|
Chris@10
|
2275 T1y = T1q + T1x;
|
Chris@10
|
2276 T75 = T1q - T1x;
|
Chris@10
|
2277 Tdh = FNMS(KP923879532, Tdg, KP382683432 * Tdf);
|
Chris@10
|
2278 Tdk = FNMS(KP923879532, Tdj, KP382683432 * Tdi);
|
Chris@10
|
2279 Tdl = Tdh + Tdk;
|
Chris@10
|
2280 TeQ = Tdh - Tdk;
|
Chris@10
|
2281 }
|
Chris@10
|
2282 {
|
Chris@10
|
2283 E TdG, TdH, T3E, T3N;
|
Chris@10
|
2284 TdG = FMA(KP923879532, Tdf, KP382683432 * Tdg);
|
Chris@10
|
2285 TdH = FMA(KP923879532, Tdi, KP382683432 * Tdj);
|
Chris@10
|
2286 TdI = TdG - TdH;
|
Chris@10
|
2287 TeD = TdG + TdH;
|
Chris@10
|
2288 T3E = T3w - T3D;
|
Chris@10
|
2289 T3N = T3F + T3M;
|
Chris@10
|
2290 T3O = KP707106781 * (T3E + T3N);
|
Chris@10
|
2291 T5N = KP707106781 * (T3E - T3N);
|
Chris@10
|
2292 }
|
Chris@10
|
2293 {
|
Chris@10
|
2294 E T4x, T4y, T9V, Ta2;
|
Chris@10
|
2295 T4x = T3w + T3D;
|
Chris@10
|
2296 T4y = T3M - T3F;
|
Chris@10
|
2297 T4z = KP707106781 * (T4x + T4y);
|
Chris@10
|
2298 T5A = KP707106781 * (T4y - T4x);
|
Chris@10
|
2299 T9V = FNMS(KP382683432, T9U, KP923879532 * T9R);
|
Chris@10
|
2300 Ta2 = FMA(KP923879532, T9Y, KP382683432 * Ta1);
|
Chris@10
|
2301 Ta3 = T9V + Ta2;
|
Chris@10
|
2302 Tc0 = T9V - Ta2;
|
Chris@10
|
2303 }
|
Chris@10
|
2304 {
|
Chris@10
|
2305 E TaM, TaN, T6M, T6N;
|
Chris@10
|
2306 TaM = FMA(KP382683432, T9R, KP923879532 * T9U);
|
Chris@10
|
2307 TaN = FNMS(KP382683432, T9Y, KP923879532 * Ta1);
|
Chris@10
|
2308 TaO = TaM + TaN;
|
Chris@10
|
2309 TbN = TaN - TaM;
|
Chris@10
|
2310 T6M = T3L + T3I;
|
Chris@10
|
2311 T6N = T3C + T3z;
|
Chris@10
|
2312 T6O = T6M - T6N;
|
Chris@10
|
2313 T80 = T6N + T6M;
|
Chris@10
|
2314 }
|
Chris@10
|
2315 }
|
Chris@10
|
2316 {
|
Chris@10
|
2317 E TQ, T2R, TT, T2U, T2O, T2V, TcW, TcV, T9g, T9d, TX, T30, T10, T33, T2X;
|
Chris@10
|
2318 E T34, TcZ, TcY, T9n, T9k;
|
Chris@10
|
2319 {
|
Chris@10
|
2320 E T9b, T9f, T9e, T9c;
|
Chris@10
|
2321 {
|
Chris@10
|
2322 E TO, TP, T2P, T2Q;
|
Chris@10
|
2323 TO = Cr[WS(csr, 10)];
|
Chris@10
|
2324 TP = Cr[WS(csr, 54)];
|
Chris@10
|
2325 TQ = TO + TP;
|
Chris@10
|
2326 T9b = TO - TP;
|
Chris@10
|
2327 T2P = Ci[WS(csi, 10)];
|
Chris@10
|
2328 T2Q = Ci[WS(csi, 54)];
|
Chris@10
|
2329 T2R = T2P - T2Q;
|
Chris@10
|
2330 T9f = T2P + T2Q;
|
Chris@10
|
2331 }
|
Chris@10
|
2332 {
|
Chris@10
|
2333 E TR, TS, T2S, T2T;
|
Chris@10
|
2334 TR = Cr[WS(csr, 42)];
|
Chris@10
|
2335 TS = Cr[WS(csr, 22)];
|
Chris@10
|
2336 TT = TR + TS;
|
Chris@10
|
2337 T9e = TR - TS;
|
Chris@10
|
2338 T2S = Ci[WS(csi, 42)];
|
Chris@10
|
2339 T2T = Ci[WS(csi, 22)];
|
Chris@10
|
2340 T2U = T2S - T2T;
|
Chris@10
|
2341 T9c = T2S + T2T;
|
Chris@10
|
2342 }
|
Chris@10
|
2343 T2O = TQ - TT;
|
Chris@10
|
2344 T2V = T2R - T2U;
|
Chris@10
|
2345 TcW = T9f - T9e;
|
Chris@10
|
2346 TcV = T9b + T9c;
|
Chris@10
|
2347 T9g = T9e + T9f;
|
Chris@10
|
2348 T9d = T9b - T9c;
|
Chris@10
|
2349 }
|
Chris@10
|
2350 {
|
Chris@10
|
2351 E T9i, T9m, T9l, T9j;
|
Chris@10
|
2352 {
|
Chris@10
|
2353 E TV, TW, T2Y, T2Z;
|
Chris@10
|
2354 TV = Cr[WS(csr, 6)];
|
Chris@10
|
2355 TW = Cr[WS(csr, 58)];
|
Chris@10
|
2356 TX = TV + TW;
|
Chris@10
|
2357 T9i = TV - TW;
|
Chris@10
|
2358 T2Y = Ci[WS(csi, 58)];
|
Chris@10
|
2359 T2Z = Ci[WS(csi, 6)];
|
Chris@10
|
2360 T30 = T2Y - T2Z;
|
Chris@10
|
2361 T9m = T2Z + T2Y;
|
Chris@10
|
2362 }
|
Chris@10
|
2363 {
|
Chris@10
|
2364 E TY, TZ, T31, T32;
|
Chris@10
|
2365 TY = Cr[WS(csr, 26)];
|
Chris@10
|
2366 TZ = Cr[WS(csr, 38)];
|
Chris@10
|
2367 T10 = TY + TZ;
|
Chris@10
|
2368 T9l = TY - TZ;
|
Chris@10
|
2369 T31 = Ci[WS(csi, 26)];
|
Chris@10
|
2370 T32 = Ci[WS(csi, 38)];
|
Chris@10
|
2371 T33 = T31 - T32;
|
Chris@10
|
2372 T9j = T31 + T32;
|
Chris@10
|
2373 }
|
Chris@10
|
2374 T2X = TX - T10;
|
Chris@10
|
2375 T34 = T30 - T33;
|
Chris@10
|
2376 TcZ = T9l + T9m;
|
Chris@10
|
2377 TcY = T9i + T9j;
|
Chris@10
|
2378 T9n = T9l - T9m;
|
Chris@10
|
2379 T9k = T9i - T9j;
|
Chris@10
|
2380 }
|
Chris@10
|
2381 {
|
Chris@10
|
2382 E TU, T11, TcX, Td0;
|
Chris@10
|
2383 TU = TQ + TT;
|
Chris@10
|
2384 T11 = TX + T10;
|
Chris@10
|
2385 T12 = TU + T11;
|
Chris@10
|
2386 T6E = TU - T11;
|
Chris@10
|
2387 TcX = FNMS(KP923879532, TcW, KP382683432 * TcV);
|
Chris@10
|
2388 Td0 = FNMS(KP923879532, TcZ, KP382683432 * TcY);
|
Chris@10
|
2389 Td1 = TcX + Td0;
|
Chris@10
|
2390 Tex = TcX - Td0;
|
Chris@10
|
2391 }
|
Chris@10
|
2392 {
|
Chris@10
|
2393 E Td3, Td4, T2W, T35;
|
Chris@10
|
2394 Td3 = FMA(KP923879532, TcV, KP382683432 * TcW);
|
Chris@10
|
2395 Td4 = FMA(KP923879532, TcY, KP382683432 * TcZ);
|
Chris@10
|
2396 Td5 = Td3 - Td4;
|
Chris@10
|
2397 Tev = Td3 + Td4;
|
Chris@10
|
2398 T2W = T2O - T2V;
|
Chris@10
|
2399 T35 = T2X + T34;
|
Chris@10
|
2400 T36 = KP707106781 * (T2W + T35);
|
Chris@10
|
2401 T5u = KP707106781 * (T2W - T35);
|
Chris@10
|
2402 }
|
Chris@10
|
2403 {
|
Chris@10
|
2404 E T38, T39, T9h, T9o;
|
Chris@10
|
2405 T38 = T2O + T2V;
|
Chris@10
|
2406 T39 = T34 - T2X;
|
Chris@10
|
2407 T3a = KP707106781 * (T38 + T39);
|
Chris@10
|
2408 T5s = KP707106781 * (T39 - T38);
|
Chris@10
|
2409 T9h = FNMS(KP382683432, T9g, KP923879532 * T9d);
|
Chris@10
|
2410 T9o = FMA(KP923879532, T9k, KP382683432 * T9n);
|
Chris@10
|
2411 T9p = T9h + T9o;
|
Chris@10
|
2412 TbH = T9h - T9o;
|
Chris@10
|
2413 }
|
Chris@10
|
2414 {
|
Chris@10
|
2415 E T9r, T9s, T6A, T6B;
|
Chris@10
|
2416 T9r = FMA(KP382683432, T9d, KP923879532 * T9g);
|
Chris@10
|
2417 T9s = FNMS(KP382683432, T9k, KP923879532 * T9n);
|
Chris@10
|
2418 T9t = T9r + T9s;
|
Chris@10
|
2419 TbF = T9s - T9r;
|
Chris@10
|
2420 T6A = T33 + T30;
|
Chris@10
|
2421 T6B = T2U + T2R;
|
Chris@10
|
2422 T6C = T6A - T6B;
|
Chris@10
|
2423 T7P = T6B + T6A;
|
Chris@10
|
2424 }
|
Chris@10
|
2425 }
|
Chris@10
|
2426 {
|
Chris@10
|
2427 E T13, T8f, Ty, T8e, T25, T8h, T8k, T8p, Ti, T14, T8o;
|
Chris@10
|
2428 T13 = KP2_000000000 * (TN + T12);
|
Chris@10
|
2429 T8f = KP2_000000000 * (T7P + T7O);
|
Chris@10
|
2430 Ti = Ta + Th;
|
Chris@10
|
2431 Ty = Ti + Tx;
|
Chris@10
|
2432 T8e = Ti - Tx;
|
Chris@10
|
2433 {
|
Chris@10
|
2434 E T1z, T24, T8i, T8j;
|
Chris@10
|
2435 T1z = T1j + T1y;
|
Chris@10
|
2436 T24 = T1O + T23;
|
Chris@10
|
2437 T25 = KP2_000000000 * (T1z + T24);
|
Chris@10
|
2438 T8h = T1z - T24;
|
Chris@10
|
2439 T8i = T80 + T7Z;
|
Chris@10
|
2440 T8j = T7V + T7U;
|
Chris@10
|
2441 T8k = T8i - T8j;
|
Chris@10
|
2442 T8p = KP2_000000000 * (T8j + T8i);
|
Chris@10
|
2443 }
|
Chris@10
|
2444 T14 = Ty + T13;
|
Chris@10
|
2445 R0[WS(rs, 32)] = T14 - T25;
|
Chris@10
|
2446 R0[0] = T14 + T25;
|
Chris@10
|
2447 T8o = Ty - T13;
|
Chris@10
|
2448 R0[WS(rs, 16)] = T8o - T8p;
|
Chris@10
|
2449 R0[WS(rs, 48)] = T8o + T8p;
|
Chris@10
|
2450 {
|
Chris@10
|
2451 E T8g, T8l, T8m, T8n;
|
Chris@10
|
2452 T8g = T8e - T8f;
|
Chris@10
|
2453 T8l = KP1_414213562 * (T8h - T8k);
|
Chris@10
|
2454 R0[WS(rs, 40)] = T8g - T8l;
|
Chris@10
|
2455 R0[WS(rs, 8)] = T8g + T8l;
|
Chris@10
|
2456 T8m = T8e + T8f;
|
Chris@10
|
2457 T8n = KP1_414213562 * (T8h + T8k);
|
Chris@10
|
2458 R0[WS(rs, 24)] = T8m - T8n;
|
Chris@10
|
2459 R0[WS(rs, 56)] = T8m + T8n;
|
Chris@10
|
2460 }
|
Chris@10
|
2461 }
|
Chris@10
|
2462 {
|
Chris@10
|
2463 E T7M, T86, T82, T8a, T7R, T87, T7X, T89, T7K, T7Y, T81;
|
Chris@10
|
2464 T7K = Ta - Th;
|
Chris@10
|
2465 T7M = T7K - T7L;
|
Chris@10
|
2466 T86 = T7K + T7L;
|
Chris@10
|
2467 T7Y = T1O - T23;
|
Chris@10
|
2468 T81 = T7Z - T80;
|
Chris@10
|
2469 T82 = T7Y + T81;
|
Chris@10
|
2470 T8a = T81 - T7Y;
|
Chris@10
|
2471 {
|
Chris@10
|
2472 E T7N, T7Q, T7T, T7W;
|
Chris@10
|
2473 T7N = TN - T12;
|
Chris@10
|
2474 T7Q = T7O - T7P;
|
Chris@10
|
2475 T7R = KP1_414213562 * (T7N - T7Q);
|
Chris@10
|
2476 T87 = KP1_414213562 * (T7N + T7Q);
|
Chris@10
|
2477 T7T = T1j - T1y;
|
Chris@10
|
2478 T7W = T7U - T7V;
|
Chris@10
|
2479 T7X = T7T + T7W;
|
Chris@10
|
2480 T89 = T7T - T7W;
|
Chris@10
|
2481 }
|
Chris@10
|
2482 {
|
Chris@10
|
2483 E T7S, T83, T8c, T8d;
|
Chris@10
|
2484 T7S = T7M + T7R;
|
Chris@10
|
2485 T83 = FNMS(KP765366864, T82, KP1_847759065 * T7X);
|
Chris@10
|
2486 R0[WS(rs, 36)] = T7S - T83;
|
Chris@10
|
2487 R0[WS(rs, 4)] = T7S + T83;
|
Chris@10
|
2488 T8c = T86 + T87;
|
Chris@10
|
2489 T8d = FMA(KP1_847759065, T89, KP765366864 * T8a);
|
Chris@10
|
2490 R0[WS(rs, 28)] = T8c - T8d;
|
Chris@10
|
2491 R0[WS(rs, 60)] = T8c + T8d;
|
Chris@10
|
2492 }
|
Chris@10
|
2493 {
|
Chris@10
|
2494 E T84, T85, T88, T8b;
|
Chris@10
|
2495 T84 = T7M - T7R;
|
Chris@10
|
2496 T85 = FMA(KP765366864, T7X, KP1_847759065 * T82);
|
Chris@10
|
2497 R0[WS(rs, 20)] = T84 - T85;
|
Chris@10
|
2498 R0[WS(rs, 52)] = T84 + T85;
|
Chris@10
|
2499 T88 = T86 - T87;
|
Chris@10
|
2500 T8b = FNMS(KP1_847759065, T8a, KP765366864 * T89);
|
Chris@10
|
2501 R0[WS(rs, 44)] = T88 - T8b;
|
Chris@10
|
2502 R0[WS(rs, 12)] = T88 + T8b;
|
Chris@10
|
2503 }
|
Chris@10
|
2504 }
|
Chris@10
|
2505 {
|
Chris@10
|
2506 E T2E, T4O, T4K, T4S, T3l, T4P, T4t, T4R;
|
Chris@10
|
2507 {
|
Chris@10
|
2508 E T2k, T2D, T4w, T4J;
|
Chris@10
|
2509 T2k = T2a + T2j;
|
Chris@10
|
2510 T2D = FNMS(KP765366864, T2C, KP1_847759065 * T2t);
|
Chris@10
|
2511 T2E = T2k + T2D;
|
Chris@10
|
2512 T4O = T2k - T2D;
|
Chris@10
|
2513 T4w = T4u + T4v;
|
Chris@10
|
2514 T4J = T4z + T4I;
|
Chris@10
|
2515 T4K = T4w + T4J;
|
Chris@10
|
2516 T4S = T4J - T4w;
|
Chris@10
|
2517 }
|
Chris@10
|
2518 {
|
Chris@10
|
2519 E T37, T3k, T3P, T4s;
|
Chris@10
|
2520 T37 = T2N + T36;
|
Chris@10
|
2521 T3k = T3a + T3j;
|
Chris@10
|
2522 T3l = FNMS(KP390180644, T3k, KP1_961570560 * T37);
|
Chris@10
|
2523 T4P = FMA(KP390180644, T37, KP1_961570560 * T3k);
|
Chris@10
|
2524 T3P = T3v + T3O;
|
Chris@10
|
2525 T4s = T48 + T4r;
|
Chris@10
|
2526 T4t = T3P + T4s;
|
Chris@10
|
2527 T4R = T3P - T4s;
|
Chris@10
|
2528 }
|
Chris@10
|
2529 {
|
Chris@10
|
2530 E T3m, T4L, T4U, T4V;
|
Chris@10
|
2531 T3m = T2E + T3l;
|
Chris@10
|
2532 T4L = FNMS(KP196034280, T4K, KP1_990369453 * T4t);
|
Chris@10
|
2533 R0[WS(rs, 33)] = T3m - T4L;
|
Chris@10
|
2534 R0[WS(rs, 1)] = T3m + T4L;
|
Chris@10
|
2535 T4U = T4O + T4P;
|
Chris@10
|
2536 T4V = FMA(KP1_546020906, T4R, KP1_268786568 * T4S);
|
Chris@10
|
2537 R0[WS(rs, 25)] = T4U - T4V;
|
Chris@10
|
2538 R0[WS(rs, 57)] = T4U + T4V;
|
Chris@10
|
2539 }
|
Chris@10
|
2540 {
|
Chris@10
|
2541 E T4M, T4N, T4Q, T4T;
|
Chris@10
|
2542 T4M = T2E - T3l;
|
Chris@10
|
2543 T4N = FMA(KP196034280, T4t, KP1_990369453 * T4K);
|
Chris@10
|
2544 R0[WS(rs, 17)] = T4M - T4N;
|
Chris@10
|
2545 R0[WS(rs, 49)] = T4M + T4N;
|
Chris@10
|
2546 T4Q = T4O - T4P;
|
Chris@10
|
2547 T4T = FNMS(KP1_546020906, T4S, KP1_268786568 * T4R);
|
Chris@10
|
2548 R0[WS(rs, 41)] = T4Q - T4T;
|
Chris@10
|
2549 R0[WS(rs, 9)] = T4Q + T4T;
|
Chris@10
|
2550 }
|
Chris@10
|
2551 }
|
Chris@10
|
2552 {
|
Chris@10
|
2553 E T6y, T7e, T7a, T7i, T6J, T7f, T71, T7h;
|
Chris@10
|
2554 {
|
Chris@10
|
2555 E T6s, T6x, T74, T79;
|
Chris@10
|
2556 T6s = T6q - T6r;
|
Chris@10
|
2557 T6x = KP1_414213562 * (T6t - T6w);
|
Chris@10
|
2558 T6y = T6s + T6x;
|
Chris@10
|
2559 T7e = T6s - T6x;
|
Chris@10
|
2560 T74 = KP707106781 * (T72 + T73);
|
Chris@10
|
2561 T79 = T75 + T78;
|
Chris@10
|
2562 T7a = T74 + T79;
|
Chris@10
|
2563 T7i = T79 - T74;
|
Chris@10
|
2564 }
|
Chris@10
|
2565 {
|
Chris@10
|
2566 E T6D, T6I, T6P, T70;
|
Chris@10
|
2567 T6D = T6z + T6C;
|
Chris@10
|
2568 T6I = T6E + T6H;
|
Chris@10
|
2569 T6J = FNMS(KP765366864, T6I, KP1_847759065 * T6D);
|
Chris@10
|
2570 T7f = FMA(KP765366864, T6D, KP1_847759065 * T6I);
|
Chris@10
|
2571 T6P = T6L + T6O;
|
Chris@10
|
2572 T70 = KP707106781 * (T6U + T6Z);
|
Chris@10
|
2573 T71 = T6P + T70;
|
Chris@10
|
2574 T7h = T6P - T70;
|
Chris@10
|
2575 }
|
Chris@10
|
2576 {
|
Chris@10
|
2577 E T6K, T7b, T7k, T7l;
|
Chris@10
|
2578 T6K = T6y + T6J;
|
Chris@10
|
2579 T7b = FNMS(KP390180644, T7a, KP1_961570560 * T71);
|
Chris@10
|
2580 R0[WS(rs, 34)] = T6K - T7b;
|
Chris@10
|
2581 R0[WS(rs, 2)] = T6K + T7b;
|
Chris@10
|
2582 T7k = T7e + T7f;
|
Chris@10
|
2583 T7l = FMA(KP1_662939224, T7h, KP1_111140466 * T7i);
|
Chris@10
|
2584 R0[WS(rs, 26)] = T7k - T7l;
|
Chris@10
|
2585 R0[WS(rs, 58)] = T7k + T7l;
|
Chris@10
|
2586 }
|
Chris@10
|
2587 {
|
Chris@10
|
2588 E T7c, T7d, T7g, T7j;
|
Chris@10
|
2589 T7c = T6y - T6J;
|
Chris@10
|
2590 T7d = FMA(KP390180644, T71, KP1_961570560 * T7a);
|
Chris@10
|
2591 R0[WS(rs, 18)] = T7c - T7d;
|
Chris@10
|
2592 R0[WS(rs, 50)] = T7c + T7d;
|
Chris@10
|
2593 T7g = T7e - T7f;
|
Chris@10
|
2594 T7j = FNMS(KP1_662939224, T7i, KP1_111140466 * T7h);
|
Chris@10
|
2595 R0[WS(rs, 42)] = T7g - T7j;
|
Chris@10
|
2596 R0[WS(rs, 10)] = T7g + T7j;
|
Chris@10
|
2597 }
|
Chris@10
|
2598 }
|
Chris@10
|
2599 {
|
Chris@10
|
2600 E T4Y, T5c, T58, T5g, T51, T5d, T55, T5f;
|
Chris@10
|
2601 {
|
Chris@10
|
2602 E T4W, T4X, T56, T57;
|
Chris@10
|
2603 T4W = T2a - T2j;
|
Chris@10
|
2604 T4X = FMA(KP765366864, T2t, KP1_847759065 * T2C);
|
Chris@10
|
2605 T4Y = T4W - T4X;
|
Chris@10
|
2606 T5c = T4W + T4X;
|
Chris@10
|
2607 T56 = T48 - T4r;
|
Chris@10
|
2608 T57 = T4I - T4z;
|
Chris@10
|
2609 T58 = T56 + T57;
|
Chris@10
|
2610 T5g = T57 - T56;
|
Chris@10
|
2611 }
|
Chris@10
|
2612 {
|
Chris@10
|
2613 E T4Z, T50, T53, T54;
|
Chris@10
|
2614 T4Z = T2N - T36;
|
Chris@10
|
2615 T50 = T3j - T3a;
|
Chris@10
|
2616 T51 = FNMS(KP1_662939224, T50, KP1_111140466 * T4Z);
|
Chris@10
|
2617 T5d = FMA(KP1_662939224, T4Z, KP1_111140466 * T50);
|
Chris@10
|
2618 T53 = T3v - T3O;
|
Chris@10
|
2619 T54 = T4v - T4u;
|
Chris@10
|
2620 T55 = T53 + T54;
|
Chris@10
|
2621 T5f = T53 - T54;
|
Chris@10
|
2622 }
|
Chris@10
|
2623 {
|
Chris@10
|
2624 E T52, T59, T5i, T5j;
|
Chris@10
|
2625 T52 = T4Y + T51;
|
Chris@10
|
2626 T59 = FNMS(KP942793473, T58, KP1_763842528 * T55);
|
Chris@10
|
2627 R0[WS(rs, 37)] = T52 - T59;
|
Chris@10
|
2628 R0[WS(rs, 5)] = T52 + T59;
|
Chris@10
|
2629 T5i = T5c + T5d;
|
Chris@10
|
2630 T5j = FMA(KP1_913880671, T5f, KP580569354 * T5g);
|
Chris@10
|
2631 R0[WS(rs, 29)] = T5i - T5j;
|
Chris@10
|
2632 R0[WS(rs, 61)] = T5i + T5j;
|
Chris@10
|
2633 }
|
Chris@10
|
2634 {
|
Chris@10
|
2635 E T5a, T5b, T5e, T5h;
|
Chris@10
|
2636 T5a = T4Y - T51;
|
Chris@10
|
2637 T5b = FMA(KP942793473, T55, KP1_763842528 * T58);
|
Chris@10
|
2638 R0[WS(rs, 21)] = T5a - T5b;
|
Chris@10
|
2639 R0[WS(rs, 53)] = T5a + T5b;
|
Chris@10
|
2640 T5e = T5c - T5d;
|
Chris@10
|
2641 T5h = FNMS(KP1_913880671, T5g, KP580569354 * T5f);
|
Chris@10
|
2642 R0[WS(rs, 45)] = T5e - T5h;
|
Chris@10
|
2643 R0[WS(rs, 13)] = T5e + T5h;
|
Chris@10
|
2644 }
|
Chris@10
|
2645 }
|
Chris@10
|
2646 {
|
Chris@10
|
2647 E T7o, T7C, T7y, T7G, T7r, T7D, T7v, T7F;
|
Chris@10
|
2648 {
|
Chris@10
|
2649 E T7m, T7n, T7w, T7x;
|
Chris@10
|
2650 T7m = T6q + T6r;
|
Chris@10
|
2651 T7n = KP1_414213562 * (T6t + T6w);
|
Chris@10
|
2652 T7o = T7m - T7n;
|
Chris@10
|
2653 T7C = T7m + T7n;
|
Chris@10
|
2654 T7w = KP707106781 * (T6U - T6Z);
|
Chris@10
|
2655 T7x = T78 - T75;
|
Chris@10
|
2656 T7y = T7w + T7x;
|
Chris@10
|
2657 T7G = T7x - T7w;
|
Chris@10
|
2658 }
|
Chris@10
|
2659 {
|
Chris@10
|
2660 E T7p, T7q, T7t, T7u;
|
Chris@10
|
2661 T7p = T6z - T6C;
|
Chris@10
|
2662 T7q = T6H - T6E;
|
Chris@10
|
2663 T7r = FNMS(KP1_847759065, T7q, KP765366864 * T7p);
|
Chris@10
|
2664 T7D = FMA(KP1_847759065, T7p, KP765366864 * T7q);
|
Chris@10
|
2665 T7t = T6L - T6O;
|
Chris@10
|
2666 T7u = KP707106781 * (T73 - T72);
|
Chris@10
|
2667 T7v = T7t + T7u;
|
Chris@10
|
2668 T7F = T7t - T7u;
|
Chris@10
|
2669 }
|
Chris@10
|
2670 {
|
Chris@10
|
2671 E T7s, T7z, T7I, T7J;
|
Chris@10
|
2672 T7s = T7o + T7r;
|
Chris@10
|
2673 T7z = FNMS(KP1_111140466, T7y, KP1_662939224 * T7v);
|
Chris@10
|
2674 R0[WS(rs, 38)] = T7s - T7z;
|
Chris@10
|
2675 R0[WS(rs, 6)] = T7s + T7z;
|
Chris@10
|
2676 T7I = T7C + T7D;
|
Chris@10
|
2677 T7J = FMA(KP1_961570560, T7F, KP390180644 * T7G);
|
Chris@10
|
2678 R0[WS(rs, 30)] = T7I - T7J;
|
Chris@10
|
2679 R0[WS(rs, 62)] = T7I + T7J;
|
Chris@10
|
2680 }
|
Chris@10
|
2681 {
|
Chris@10
|
2682 E T7A, T7B, T7E, T7H;
|
Chris@10
|
2683 T7A = T7o - T7r;
|
Chris@10
|
2684 T7B = FMA(KP1_111140466, T7v, KP1_662939224 * T7y);
|
Chris@10
|
2685 R0[WS(rs, 22)] = T7A - T7B;
|
Chris@10
|
2686 R0[WS(rs, 54)] = T7A + T7B;
|
Chris@10
|
2687 T7E = T7C - T7D;
|
Chris@10
|
2688 T7H = FNMS(KP1_961570560, T7G, KP390180644 * T7F);
|
Chris@10
|
2689 R0[WS(rs, 46)] = T7E - T7H;
|
Chris@10
|
2690 R0[WS(rs, 14)] = T7E + T7H;
|
Chris@10
|
2691 }
|
Chris@10
|
2692 }
|
Chris@10
|
2693 {
|
Chris@10
|
2694 E T5q, T5U, T5Q, T5Y, T5x, T5V, T5J, T5X;
|
Chris@10
|
2695 {
|
Chris@10
|
2696 E T5m, T5p, T5M, T5P;
|
Chris@10
|
2697 T5m = T5k - T5l;
|
Chris@10
|
2698 T5p = FNMS(KP1_847759065, T5o, KP765366864 * T5n);
|
Chris@10
|
2699 T5q = T5m + T5p;
|
Chris@10
|
2700 T5U = T5m - T5p;
|
Chris@10
|
2701 T5M = T5K + T5L;
|
Chris@10
|
2702 T5P = T5N + T5O;
|
Chris@10
|
2703 T5Q = T5M + T5P;
|
Chris@10
|
2704 T5Y = T5P - T5M;
|
Chris@10
|
2705 }
|
Chris@10
|
2706 {
|
Chris@10
|
2707 E T5t, T5w, T5B, T5I;
|
Chris@10
|
2708 T5t = T5r + T5s;
|
Chris@10
|
2709 T5w = T5u + T5v;
|
Chris@10
|
2710 T5x = FNMS(KP1_111140466, T5w, KP1_662939224 * T5t);
|
Chris@10
|
2711 T5V = FMA(KP1_111140466, T5t, KP1_662939224 * T5w);
|
Chris@10
|
2712 T5B = T5z + T5A;
|
Chris@10
|
2713 T5I = T5E + T5H;
|
Chris@10
|
2714 T5J = T5B + T5I;
|
Chris@10
|
2715 T5X = T5B - T5I;
|
Chris@10
|
2716 }
|
Chris@10
|
2717 {
|
Chris@10
|
2718 E T5y, T5R, T60, T61;
|
Chris@10
|
2719 T5y = T5q + T5x;
|
Chris@10
|
2720 T5R = FNMS(KP580569354, T5Q, KP1_913880671 * T5J);
|
Chris@10
|
2721 R0[WS(rs, 35)] = T5y - T5R;
|
Chris@10
|
2722 R0[WS(rs, 3)] = T5y + T5R;
|
Chris@10
|
2723 T60 = T5U + T5V;
|
Chris@10
|
2724 T61 = FMA(KP1_763842528, T5X, KP942793473 * T5Y);
|
Chris@10
|
2725 R0[WS(rs, 27)] = T60 - T61;
|
Chris@10
|
2726 R0[WS(rs, 59)] = T60 + T61;
|
Chris@10
|
2727 }
|
Chris@10
|
2728 {
|
Chris@10
|
2729 E T5S, T5T, T5W, T5Z;
|
Chris@10
|
2730 T5S = T5q - T5x;
|
Chris@10
|
2731 T5T = FMA(KP580569354, T5J, KP1_913880671 * T5Q);
|
Chris@10
|
2732 R0[WS(rs, 19)] = T5S - T5T;
|
Chris@10
|
2733 R0[WS(rs, 51)] = T5S + T5T;
|
Chris@10
|
2734 T5W = T5U - T5V;
|
Chris@10
|
2735 T5Z = FNMS(KP1_763842528, T5Y, KP942793473 * T5X);
|
Chris@10
|
2736 R0[WS(rs, 43)] = T5W - T5Z;
|
Chris@10
|
2737 R0[WS(rs, 11)] = T5W + T5Z;
|
Chris@10
|
2738 }
|
Chris@10
|
2739 }
|
Chris@10
|
2740 {
|
Chris@10
|
2741 E T64, T6i, T6e, T6m, T67, T6j, T6b, T6l;
|
Chris@10
|
2742 {
|
Chris@10
|
2743 E T62, T63, T6c, T6d;
|
Chris@10
|
2744 T62 = T5k + T5l;
|
Chris@10
|
2745 T63 = FMA(KP1_847759065, T5n, KP765366864 * T5o);
|
Chris@10
|
2746 T64 = T62 - T63;
|
Chris@10
|
2747 T6i = T62 + T63;
|
Chris@10
|
2748 T6c = T5E - T5H;
|
Chris@10
|
2749 T6d = T5O - T5N;
|
Chris@10
|
2750 T6e = T6c + T6d;
|
Chris@10
|
2751 T6m = T6d - T6c;
|
Chris@10
|
2752 }
|
Chris@10
|
2753 {
|
Chris@10
|
2754 E T65, T66, T69, T6a;
|
Chris@10
|
2755 T65 = T5r - T5s;
|
Chris@10
|
2756 T66 = T5v - T5u;
|
Chris@10
|
2757 T67 = FNMS(KP1_961570560, T66, KP390180644 * T65);
|
Chris@10
|
2758 T6j = FMA(KP1_961570560, T65, KP390180644 * T66);
|
Chris@10
|
2759 T69 = T5z - T5A;
|
Chris@10
|
2760 T6a = T5L - T5K;
|
Chris@10
|
2761 T6b = T69 + T6a;
|
Chris@10
|
2762 T6l = T69 - T6a;
|
Chris@10
|
2763 }
|
Chris@10
|
2764 {
|
Chris@10
|
2765 E T68, T6f, T6o, T6p;
|
Chris@10
|
2766 T68 = T64 + T67;
|
Chris@10
|
2767 T6f = FNMS(KP1_268786568, T6e, KP1_546020906 * T6b);
|
Chris@10
|
2768 R0[WS(rs, 39)] = T68 - T6f;
|
Chris@10
|
2769 R0[WS(rs, 7)] = T68 + T6f;
|
Chris@10
|
2770 T6o = T6i + T6j;
|
Chris@10
|
2771 T6p = FMA(KP1_990369453, T6l, KP196034280 * T6m);
|
Chris@10
|
2772 R0[WS(rs, 31)] = T6o - T6p;
|
Chris@10
|
2773 R0[WS(rs, 63)] = T6o + T6p;
|
Chris@10
|
2774 }
|
Chris@10
|
2775 {
|
Chris@10
|
2776 E T6g, T6h, T6k, T6n;
|
Chris@10
|
2777 T6g = T64 - T67;
|
Chris@10
|
2778 T6h = FMA(KP1_268786568, T6b, KP1_546020906 * T6e);
|
Chris@10
|
2779 R0[WS(rs, 23)] = T6g - T6h;
|
Chris@10
|
2780 R0[WS(rs, 55)] = T6g + T6h;
|
Chris@10
|
2781 T6k = T6i - T6j;
|
Chris@10
|
2782 T6n = FNMS(KP1_990369453, T6m, KP196034280 * T6l);
|
Chris@10
|
2783 R0[WS(rs, 47)] = T6k - T6n;
|
Chris@10
|
2784 R0[WS(rs, 15)] = T6k + T6n;
|
Chris@10
|
2785 }
|
Chris@10
|
2786 }
|
Chris@10
|
2787 {
|
Chris@10
|
2788 E T8Z, Tb1, T9C, Tb2, Tbe, Tbq, Tbb, Tbp, TaX, Tbs, Tb5, Tbi, TaI, Tbt, Tb4;
|
Chris@10
|
2789 E Tbl;
|
Chris@10
|
2790 {
|
Chris@10
|
2791 E T8F, T8Y, Tb9, Tba;
|
Chris@10
|
2792 T8F = T8x + T8E;
|
Chris@10
|
2793 T8Y = FNMS(KP390180644, T8X, KP1_961570560 * T8Q);
|
Chris@10
|
2794 T8Z = T8F + T8Y;
|
Chris@10
|
2795 Tb1 = T8F - T8Y;
|
Chris@10
|
2796 {
|
Chris@10
|
2797 E T9q, T9B, Tbc, Tbd;
|
Chris@10
|
2798 T9q = T9a + T9p;
|
Chris@10
|
2799 T9B = T9t + T9A;
|
Chris@10
|
2800 T9C = FNMS(KP196034280, T9B, KP1_990369453 * T9q);
|
Chris@10
|
2801 Tb2 = FMA(KP196034280, T9q, KP1_990369453 * T9B);
|
Chris@10
|
2802 Tbc = T9a - T9p;
|
Chris@10
|
2803 Tbd = T9A - T9t;
|
Chris@10
|
2804 Tbe = FNMS(KP1_546020906, Tbd, KP1_268786568 * Tbc);
|
Chris@10
|
2805 Tbq = FMA(KP1_546020906, Tbc, KP1_268786568 * Tbd);
|
Chris@10
|
2806 }
|
Chris@10
|
2807 Tb9 = T8x - T8E;
|
Chris@10
|
2808 Tba = FMA(KP390180644, T8Q, KP1_961570560 * T8X);
|
Chris@10
|
2809 Tbb = Tb9 - Tba;
|
Chris@10
|
2810 Tbp = Tb9 + Tba;
|
Chris@10
|
2811 {
|
Chris@10
|
2812 E TaW, Tbg, TaL, Tbh, TaJ, TaK;
|
Chris@10
|
2813 TaW = TaO + TaV;
|
Chris@10
|
2814 Tbg = T9O - Ta3;
|
Chris@10
|
2815 TaJ = FMA(KP195090322, Taf, KP980785280 * Tam);
|
Chris@10
|
2816 TaK = FNMS(KP195090322, Tay, KP980785280 * TaF);
|
Chris@10
|
2817 TaL = TaJ + TaK;
|
Chris@10
|
2818 Tbh = TaK - TaJ;
|
Chris@10
|
2819 TaX = TaL + TaW;
|
Chris@10
|
2820 Tbs = Tbg - Tbh;
|
Chris@10
|
2821 Tb5 = TaW - TaL;
|
Chris@10
|
2822 Tbi = Tbg + Tbh;
|
Chris@10
|
2823 }
|
Chris@10
|
2824 {
|
Chris@10
|
2825 E Ta4, Tbk, TaH, Tbj, Tan, TaG;
|
Chris@10
|
2826 Ta4 = T9O + Ta3;
|
Chris@10
|
2827 Tbk = TaV - TaO;
|
Chris@10
|
2828 Tan = FNMS(KP195090322, Tam, KP980785280 * Taf);
|
Chris@10
|
2829 TaG = FMA(KP980785280, Tay, KP195090322 * TaF);
|
Chris@10
|
2830 TaH = Tan + TaG;
|
Chris@10
|
2831 Tbj = Tan - TaG;
|
Chris@10
|
2832 TaI = Ta4 + TaH;
|
Chris@10
|
2833 Tbt = Tbk - Tbj;
|
Chris@10
|
2834 Tb4 = Ta4 - TaH;
|
Chris@10
|
2835 Tbl = Tbj + Tbk;
|
Chris@10
|
2836 }
|
Chris@10
|
2837 }
|
Chris@10
|
2838 {
|
Chris@10
|
2839 E T9D, TaY, Tbr, Tbu;
|
Chris@10
|
2840 T9D = T8Z + T9C;
|
Chris@10
|
2841 TaY = FNMS(KP098135348, TaX, KP1_997590912 * TaI);
|
Chris@10
|
2842 R1[WS(rs, 32)] = T9D - TaY;
|
Chris@10
|
2843 R1[0] = T9D + TaY;
|
Chris@10
|
2844 Tbr = Tbp - Tbq;
|
Chris@10
|
2845 Tbu = FNMS(KP1_883088130, Tbt, KP673779706 * Tbs);
|
Chris@10
|
2846 R1[WS(rs, 44)] = Tbr - Tbu;
|
Chris@10
|
2847 R1[WS(rs, 12)] = Tbr + Tbu;
|
Chris@10
|
2848 }
|
Chris@10
|
2849 {
|
Chris@10
|
2850 E Tbv, Tbw, TaZ, Tb0;
|
Chris@10
|
2851 Tbv = Tbp + Tbq;
|
Chris@10
|
2852 Tbw = FMA(KP1_883088130, Tbs, KP673779706 * Tbt);
|
Chris@10
|
2853 R1[WS(rs, 28)] = Tbv - Tbw;
|
Chris@10
|
2854 R1[WS(rs, 60)] = Tbv + Tbw;
|
Chris@10
|
2855 TaZ = T8Z - T9C;
|
Chris@10
|
2856 Tb0 = FMA(KP098135348, TaI, KP1_997590912 * TaX);
|
Chris@10
|
2857 R1[WS(rs, 16)] = TaZ - Tb0;
|
Chris@10
|
2858 R1[WS(rs, 48)] = TaZ + Tb0;
|
Chris@10
|
2859 }
|
Chris@10
|
2860 {
|
Chris@10
|
2861 E Tb3, Tb6, Tbf, Tbm;
|
Chris@10
|
2862 Tb3 = Tb1 - Tb2;
|
Chris@10
|
2863 Tb6 = FNMS(KP1_481902250, Tb5, KP1_343117909 * Tb4);
|
Chris@10
|
2864 R1[WS(rs, 40)] = Tb3 - Tb6;
|
Chris@10
|
2865 R1[WS(rs, 8)] = Tb3 + Tb6;
|
Chris@10
|
2866 Tbf = Tbb + Tbe;
|
Chris@10
|
2867 Tbm = FNMS(KP855110186, Tbl, KP1_807978586 * Tbi);
|
Chris@10
|
2868 R1[WS(rs, 36)] = Tbf - Tbm;
|
Chris@10
|
2869 R1[WS(rs, 4)] = Tbf + Tbm;
|
Chris@10
|
2870 }
|
Chris@10
|
2871 {
|
Chris@10
|
2872 E Tbn, Tbo, Tb7, Tb8;
|
Chris@10
|
2873 Tbn = Tbb - Tbe;
|
Chris@10
|
2874 Tbo = FMA(KP855110186, Tbi, KP1_807978586 * Tbl);
|
Chris@10
|
2875 R1[WS(rs, 20)] = Tbn - Tbo;
|
Chris@10
|
2876 R1[WS(rs, 52)] = Tbn + Tbo;
|
Chris@10
|
2877 Tb7 = Tb1 + Tb2;
|
Chris@10
|
2878 Tb8 = FMA(KP1_481902250, Tb4, KP1_343117909 * Tb5);
|
Chris@10
|
2879 R1[WS(rs, 24)] = Tb7 - Tb8;
|
Chris@10
|
2880 R1[WS(rs, 56)] = Tb7 + Tb8;
|
Chris@10
|
2881 }
|
Chris@10
|
2882 }
|
Chris@10
|
2883 {
|
Chris@10
|
2884 E TcR, TdR, Tda, TdS, Te4, Teg, Te1, Tef, TdN, Tei, TdV, Te8, TdC, Tej, TdU;
|
Chris@10
|
2885 E Teb;
|
Chris@10
|
2886 {
|
Chris@10
|
2887 E TcJ, TcQ, TdZ, Te0;
|
Chris@10
|
2888 TcJ = TcF + TcI;
|
Chris@10
|
2889 TcQ = FNMS(KP1_111140466, TcP, KP1_662939224 * TcM);
|
Chris@10
|
2890 TcR = TcJ + TcQ;
|
Chris@10
|
2891 TdR = TcJ - TcQ;
|
Chris@10
|
2892 {
|
Chris@10
|
2893 E Td2, Td9, Te2, Te3;
|
Chris@10
|
2894 Td2 = TcU + Td1;
|
Chris@10
|
2895 Td9 = Td5 + Td8;
|
Chris@10
|
2896 Tda = FNMS(KP580569354, Td9, KP1_913880671 * Td2);
|
Chris@10
|
2897 TdS = FMA(KP580569354, Td2, KP1_913880671 * Td9);
|
Chris@10
|
2898 Te2 = TcU - Td1;
|
Chris@10
|
2899 Te3 = Td8 - Td5;
|
Chris@10
|
2900 Te4 = FNMS(KP1_763842528, Te3, KP942793473 * Te2);
|
Chris@10
|
2901 Teg = FMA(KP1_763842528, Te2, KP942793473 * Te3);
|
Chris@10
|
2902 }
|
Chris@10
|
2903 TdZ = TcF - TcI;
|
Chris@10
|
2904 Te0 = FMA(KP1_111140466, TcM, KP1_662939224 * TcP);
|
Chris@10
|
2905 Te1 = TdZ - Te0;
|
Chris@10
|
2906 Tef = TdZ + Te0;
|
Chris@10
|
2907 {
|
Chris@10
|
2908 E TdM, Te6, TdF, Te7, TdD, TdE;
|
Chris@10
|
2909 TdM = TdI + TdL;
|
Chris@10
|
2910 Te6 = Tde - Tdl;
|
Chris@10
|
2911 TdD = FMA(KP555570233, Tdp, KP831469612 * Tds);
|
Chris@10
|
2912 TdE = FNMS(KP555570233, Tdw, KP831469612 * Tdz);
|
Chris@10
|
2913 TdF = TdD + TdE;
|
Chris@10
|
2914 Te7 = TdE - TdD;
|
Chris@10
|
2915 TdN = TdF + TdM;
|
Chris@10
|
2916 Tei = Te6 - Te7;
|
Chris@10
|
2917 TdV = TdM - TdF;
|
Chris@10
|
2918 Te8 = Te6 + Te7;
|
Chris@10
|
2919 }
|
Chris@10
|
2920 {
|
Chris@10
|
2921 E Tdm, Tea, TdB, Te9, Tdt, TdA;
|
Chris@10
|
2922 Tdm = Tde + Tdl;
|
Chris@10
|
2923 Tea = TdL - TdI;
|
Chris@10
|
2924 Tdt = FNMS(KP555570233, Tds, KP831469612 * Tdp);
|
Chris@10
|
2925 TdA = FMA(KP831469612, Tdw, KP555570233 * Tdz);
|
Chris@10
|
2926 TdB = Tdt + TdA;
|
Chris@10
|
2927 Te9 = Tdt - TdA;
|
Chris@10
|
2928 TdC = Tdm + TdB;
|
Chris@10
|
2929 Tej = Tea - Te9;
|
Chris@10
|
2930 TdU = Tdm - TdB;
|
Chris@10
|
2931 Teb = Te9 + Tea;
|
Chris@10
|
2932 }
|
Chris@10
|
2933 }
|
Chris@10
|
2934 {
|
Chris@10
|
2935 E Tdb, TdO, Teh, Tek;
|
Chris@10
|
2936 Tdb = TcR + Tda;
|
Chris@10
|
2937 TdO = FNMS(KP293460948, TdN, KP1_978353019 * TdC);
|
Chris@10
|
2938 R1[WS(rs, 33)] = Tdb - TdO;
|
Chris@10
|
2939 R1[WS(rs, 1)] = Tdb + TdO;
|
Chris@10
|
2940 Teh = Tef - Teg;
|
Chris@10
|
2941 Tek = FNMS(KP1_940062506, Tej, KP485960359 * Tei);
|
Chris@10
|
2942 R1[WS(rs, 45)] = Teh - Tek;
|
Chris@10
|
2943 R1[WS(rs, 13)] = Teh + Tek;
|
Chris@10
|
2944 }
|
Chris@10
|
2945 {
|
Chris@10
|
2946 E Tel, Tem, TdP, TdQ;
|
Chris@10
|
2947 Tel = Tef + Teg;
|
Chris@10
|
2948 Tem = FMA(KP1_940062506, Tei, KP485960359 * Tej);
|
Chris@10
|
2949 R1[WS(rs, 29)] = Tel - Tem;
|
Chris@10
|
2950 R1[WS(rs, 61)] = Tel + Tem;
|
Chris@10
|
2951 TdP = TcR - Tda;
|
Chris@10
|
2952 TdQ = FMA(KP293460948, TdC, KP1_978353019 * TdN);
|
Chris@10
|
2953 R1[WS(rs, 17)] = TdP - TdQ;
|
Chris@10
|
2954 R1[WS(rs, 49)] = TdP + TdQ;
|
Chris@10
|
2955 }
|
Chris@10
|
2956 {
|
Chris@10
|
2957 E TdT, TdW, Te5, Tec;
|
Chris@10
|
2958 TdT = TdR - TdS;
|
Chris@10
|
2959 TdW = FNMS(KP1_606415062, TdV, KP1_191398608 * TdU);
|
Chris@10
|
2960 R1[WS(rs, 41)] = TdT - TdW;
|
Chris@10
|
2961 R1[WS(rs, 9)] = TdT + TdW;
|
Chris@10
|
2962 Te5 = Te1 + Te4;
|
Chris@10
|
2963 Tec = FNMS(KP1_028205488, Teb, KP1_715457220 * Te8);
|
Chris@10
|
2964 R1[WS(rs, 37)] = Te5 - Tec;
|
Chris@10
|
2965 R1[WS(rs, 5)] = Te5 + Tec;
|
Chris@10
|
2966 }
|
Chris@10
|
2967 {
|
Chris@10
|
2968 E Ted, Tee, TdX, TdY;
|
Chris@10
|
2969 Ted = Te1 - Te4;
|
Chris@10
|
2970 Tee = FMA(KP1_028205488, Te8, KP1_715457220 * Teb);
|
Chris@10
|
2971 R1[WS(rs, 21)] = Ted - Tee;
|
Chris@10
|
2972 R1[WS(rs, 53)] = Ted + Tee;
|
Chris@10
|
2973 TdX = TdR + TdS;
|
Chris@10
|
2974 TdY = FMA(KP1_606415062, TdU, KP1_191398608 * TdV);
|
Chris@10
|
2975 R1[WS(rs, 25)] = TdX - TdY;
|
Chris@10
|
2976 R1[WS(rs, 57)] = TdX + TdY;
|
Chris@10
|
2977 }
|
Chris@10
|
2978 }
|
Chris@10
|
2979 {
|
Chris@10
|
2980 E TbD, Tc7, TbK, Tc8, Tck, Tcw, Tch, Tcv, Tc3, Tcy, Tcb, Tco, TbW, Tcz, Tca;
|
Chris@10
|
2981 E Tcr;
|
Chris@10
|
2982 {
|
Chris@10
|
2983 E Tbz, TbC, Tcf, Tcg;
|
Chris@10
|
2984 Tbz = Tbx - Tby;
|
Chris@10
|
2985 TbC = FNMS(KP1_662939224, TbB, KP1_111140466 * TbA);
|
Chris@10
|
2986 TbD = Tbz + TbC;
|
Chris@10
|
2987 Tc7 = Tbz - TbC;
|
Chris@10
|
2988 {
|
Chris@10
|
2989 E TbG, TbJ, Tci, Tcj;
|
Chris@10
|
2990 TbG = TbE + TbF;
|
Chris@10
|
2991 TbJ = TbH + TbI;
|
Chris@10
|
2992 TbK = FNMS(KP942793473, TbJ, KP1_763842528 * TbG);
|
Chris@10
|
2993 Tc8 = FMA(KP942793473, TbG, KP1_763842528 * TbJ);
|
Chris@10
|
2994 Tci = TbE - TbF;
|
Chris@10
|
2995 Tcj = TbI - TbH;
|
Chris@10
|
2996 Tck = FNMS(KP1_913880671, Tcj, KP580569354 * Tci);
|
Chris@10
|
2997 Tcw = FMA(KP1_913880671, Tci, KP580569354 * Tcj);
|
Chris@10
|
2998 }
|
Chris@10
|
2999 Tcf = Tbx + Tby;
|
Chris@10
|
3000 Tcg = FMA(KP1_662939224, TbA, KP1_111140466 * TbB);
|
Chris@10
|
3001 Tch = Tcf - Tcg;
|
Chris@10
|
3002 Tcv = Tcf + Tcg;
|
Chris@10
|
3003 {
|
Chris@10
|
3004 E Tc2, Tcm, TbZ, Tcn, TbX, TbY;
|
Chris@10
|
3005 Tc2 = Tc0 + Tc1;
|
Chris@10
|
3006 Tcm = TbM - TbN;
|
Chris@10
|
3007 TbX = FMA(KP831469612, TbP, KP555570233 * TbQ);
|
Chris@10
|
3008 TbY = FNMS(KP831469612, TbS, KP555570233 * TbT);
|
Chris@10
|
3009 TbZ = TbX + TbY;
|
Chris@10
|
3010 Tcn = TbY - TbX;
|
Chris@10
|
3011 Tc3 = TbZ + Tc2;
|
Chris@10
|
3012 Tcy = Tcm - Tcn;
|
Chris@10
|
3013 Tcb = Tc2 - TbZ;
|
Chris@10
|
3014 Tco = Tcm + Tcn;
|
Chris@10
|
3015 }
|
Chris@10
|
3016 {
|
Chris@10
|
3017 E TbO, Tcq, TbV, Tcp, TbR, TbU;
|
Chris@10
|
3018 TbO = TbM + TbN;
|
Chris@10
|
3019 Tcq = Tc1 - Tc0;
|
Chris@10
|
3020 TbR = FNMS(KP831469612, TbQ, KP555570233 * TbP);
|
Chris@10
|
3021 TbU = FMA(KP555570233, TbS, KP831469612 * TbT);
|
Chris@10
|
3022 TbV = TbR + TbU;
|
Chris@10
|
3023 Tcp = TbR - TbU;
|
Chris@10
|
3024 TbW = TbO + TbV;
|
Chris@10
|
3025 Tcz = Tcq - Tcp;
|
Chris@10
|
3026 Tca = TbO - TbV;
|
Chris@10
|
3027 Tcr = Tcp + Tcq;
|
Chris@10
|
3028 }
|
Chris@10
|
3029 }
|
Chris@10
|
3030 {
|
Chris@10
|
3031 E TbL, Tc4, Tcx, TcA;
|
Chris@10
|
3032 TbL = TbD + TbK;
|
Chris@10
|
3033 Tc4 = FNMS(KP485960359, Tc3, KP1_940062506 * TbW);
|
Chris@10
|
3034 R1[WS(rs, 34)] = TbL - Tc4;
|
Chris@10
|
3035 R1[WS(rs, 2)] = TbL + Tc4;
|
Chris@10
|
3036 Tcx = Tcv - Tcw;
|
Chris@10
|
3037 TcA = FNMS(KP1_978353019, Tcz, KP293460948 * Tcy);
|
Chris@10
|
3038 R1[WS(rs, 46)] = Tcx - TcA;
|
Chris@10
|
3039 R1[WS(rs, 14)] = Tcx + TcA;
|
Chris@10
|
3040 }
|
Chris@10
|
3041 {
|
Chris@10
|
3042 E TcB, TcC, Tc5, Tc6;
|
Chris@10
|
3043 TcB = Tcv + Tcw;
|
Chris@10
|
3044 TcC = FMA(KP1_978353019, Tcy, KP293460948 * Tcz);
|
Chris@10
|
3045 R1[WS(rs, 30)] = TcB - TcC;
|
Chris@10
|
3046 R1[WS(rs, 62)] = TcB + TcC;
|
Chris@10
|
3047 Tc5 = TbD - TbK;
|
Chris@10
|
3048 Tc6 = FMA(KP485960359, TbW, KP1_940062506 * Tc3);
|
Chris@10
|
3049 R1[WS(rs, 18)] = Tc5 - Tc6;
|
Chris@10
|
3050 R1[WS(rs, 50)] = Tc5 + Tc6;
|
Chris@10
|
3051 }
|
Chris@10
|
3052 {
|
Chris@10
|
3053 E Tc9, Tcc, Tcl, Tcs;
|
Chris@10
|
3054 Tc9 = Tc7 - Tc8;
|
Chris@10
|
3055 Tcc = FNMS(KP1_715457220, Tcb, KP1_028205488 * Tca);
|
Chris@10
|
3056 R1[WS(rs, 42)] = Tc9 - Tcc;
|
Chris@10
|
3057 R1[WS(rs, 10)] = Tc9 + Tcc;
|
Chris@10
|
3058 Tcl = Tch + Tck;
|
Chris@10
|
3059 Tcs = FNMS(KP1_191398608, Tcr, KP1_606415062 * Tco);
|
Chris@10
|
3060 R1[WS(rs, 38)] = Tcl - Tcs;
|
Chris@10
|
3061 R1[WS(rs, 6)] = Tcl + Tcs;
|
Chris@10
|
3062 }
|
Chris@10
|
3063 {
|
Chris@10
|
3064 E Tct, Tcu, Tcd, Tce;
|
Chris@10
|
3065 Tct = Tch - Tck;
|
Chris@10
|
3066 Tcu = FMA(KP1_191398608, Tco, KP1_606415062 * Tcr);
|
Chris@10
|
3067 R1[WS(rs, 22)] = Tct - Tcu;
|
Chris@10
|
3068 R1[WS(rs, 54)] = Tct + Tcu;
|
Chris@10
|
3069 Tcd = Tc7 + Tc8;
|
Chris@10
|
3070 Tce = FMA(KP1_715457220, Tca, KP1_028205488 * Tcb);
|
Chris@10
|
3071 R1[WS(rs, 26)] = Tcd - Tce;
|
Chris@10
|
3072 R1[WS(rs, 58)] = Tcd + Tce;
|
Chris@10
|
3073 }
|
Chris@10
|
3074 }
|
Chris@10
|
3075 {
|
Chris@10
|
3076 E Tet, TeX, TeA, TeY, Tfa, Tfm, Tf7, Tfl, TeT, Tfo, Tf1, Tfe, TeM, Tfp, Tf0;
|
Chris@10
|
3077 E Tfh;
|
Chris@10
|
3078 {
|
Chris@10
|
3079 E Tep, Tes, Tf5, Tf6;
|
Chris@10
|
3080 Tep = Ten - Teo;
|
Chris@10
|
3081 Tes = FNMS(KP1_961570560, Ter, KP390180644 * Teq);
|
Chris@10
|
3082 Tet = Tep + Tes;
|
Chris@10
|
3083 TeX = Tep - Tes;
|
Chris@10
|
3084 {
|
Chris@10
|
3085 E Tew, Tez, Tf8, Tf9;
|
Chris@10
|
3086 Tew = Teu - Tev;
|
Chris@10
|
3087 Tez = Tex + Tey;
|
Chris@10
|
3088 TeA = FNMS(KP1_268786568, Tez, KP1_546020906 * Tew);
|
Chris@10
|
3089 TeY = FMA(KP1_268786568, Tew, KP1_546020906 * Tez);
|
Chris@10
|
3090 Tf8 = Teu + Tev;
|
Chris@10
|
3091 Tf9 = Tey - Tex;
|
Chris@10
|
3092 Tfa = FNMS(KP1_990369453, Tf9, KP196034280 * Tf8);
|
Chris@10
|
3093 Tfm = FMA(KP1_990369453, Tf8, KP196034280 * Tf9);
|
Chris@10
|
3094 }
|
Chris@10
|
3095 Tf5 = Ten + Teo;
|
Chris@10
|
3096 Tf6 = FMA(KP1_961570560, Teq, KP390180644 * Ter);
|
Chris@10
|
3097 Tf7 = Tf5 - Tf6;
|
Chris@10
|
3098 Tfl = Tf5 + Tf6;
|
Chris@10
|
3099 {
|
Chris@10
|
3100 E TeS, Tfc, TeP, Tfd, TeN, TeO;
|
Chris@10
|
3101 TeS = TeQ + TeR;
|
Chris@10
|
3102 Tfc = TeC + TeD;
|
Chris@10
|
3103 TeN = FMA(KP980785280, TeF, KP195090322 * TeG);
|
Chris@10
|
3104 TeO = FMA(KP980785280, TeI, KP195090322 * TeJ);
|
Chris@10
|
3105 TeP = TeN - TeO;
|
Chris@10
|
3106 Tfd = TeN + TeO;
|
Chris@10
|
3107 TeT = TeP + TeS;
|
Chris@10
|
3108 Tfo = Tfc + Tfd;
|
Chris@10
|
3109 Tf1 = TeS - TeP;
|
Chris@10
|
3110 Tfe = Tfc - Tfd;
|
Chris@10
|
3111 }
|
Chris@10
|
3112 {
|
Chris@10
|
3113 E TeE, Tfg, TeL, Tff, TeH, TeK;
|
Chris@10
|
3114 TeE = TeC - TeD;
|
Chris@10
|
3115 Tfg = TeR - TeQ;
|
Chris@10
|
3116 TeH = FNMS(KP980785280, TeG, KP195090322 * TeF);
|
Chris@10
|
3117 TeK = FNMS(KP980785280, TeJ, KP195090322 * TeI);
|
Chris@10
|
3118 TeL = TeH + TeK;
|
Chris@10
|
3119 Tff = TeH - TeK;
|
Chris@10
|
3120 TeM = TeE + TeL;
|
Chris@10
|
3121 Tfp = Tfg - Tff;
|
Chris@10
|
3122 Tf0 = TeE - TeL;
|
Chris@10
|
3123 Tfh = Tff + Tfg;
|
Chris@10
|
3124 }
|
Chris@10
|
3125 }
|
Chris@10
|
3126 {
|
Chris@10
|
3127 E TeB, TeU, Tfn, Tfq;
|
Chris@10
|
3128 TeB = Tet + TeA;
|
Chris@10
|
3129 TeU = FNMS(KP673779706, TeT, KP1_883088130 * TeM);
|
Chris@10
|
3130 R1[WS(rs, 35)] = TeB - TeU;
|
Chris@10
|
3131 R1[WS(rs, 3)] = TeB + TeU;
|
Chris@10
|
3132 Tfn = Tfl - Tfm;
|
Chris@10
|
3133 Tfq = FNMS(KP1_997590912, Tfp, KP098135348 * Tfo);
|
Chris@10
|
3134 R1[WS(rs, 47)] = Tfn - Tfq;
|
Chris@10
|
3135 R1[WS(rs, 15)] = Tfn + Tfq;
|
Chris@10
|
3136 }
|
Chris@10
|
3137 {
|
Chris@10
|
3138 E Tfr, Tfs, TeV, TeW;
|
Chris@10
|
3139 Tfr = Tfl + Tfm;
|
Chris@10
|
3140 Tfs = FMA(KP1_997590912, Tfo, KP098135348 * Tfp);
|
Chris@10
|
3141 R1[WS(rs, 31)] = Tfr - Tfs;
|
Chris@10
|
3142 R1[WS(rs, 63)] = Tfr + Tfs;
|
Chris@10
|
3143 TeV = Tet - TeA;
|
Chris@10
|
3144 TeW = FMA(KP673779706, TeM, KP1_883088130 * TeT);
|
Chris@10
|
3145 R1[WS(rs, 19)] = TeV - TeW;
|
Chris@10
|
3146 R1[WS(rs, 51)] = TeV + TeW;
|
Chris@10
|
3147 }
|
Chris@10
|
3148 {
|
Chris@10
|
3149 E TeZ, Tf2, Tfb, Tfi;
|
Chris@10
|
3150 TeZ = TeX - TeY;
|
Chris@10
|
3151 Tf2 = FNMS(KP1_807978586, Tf1, KP855110186 * Tf0);
|
Chris@10
|
3152 R1[WS(rs, 43)] = TeZ - Tf2;
|
Chris@10
|
3153 R1[WS(rs, 11)] = TeZ + Tf2;
|
Chris@10
|
3154 Tfb = Tf7 + Tfa;
|
Chris@10
|
3155 Tfi = FNMS(KP1_343117909, Tfh, KP1_481902250 * Tfe);
|
Chris@10
|
3156 R1[WS(rs, 39)] = Tfb - Tfi;
|
Chris@10
|
3157 R1[WS(rs, 7)] = Tfb + Tfi;
|
Chris@10
|
3158 }
|
Chris@10
|
3159 {
|
Chris@10
|
3160 E Tfj, Tfk, Tf3, Tf4;
|
Chris@10
|
3161 Tfj = Tf7 - Tfa;
|
Chris@10
|
3162 Tfk = FMA(KP1_343117909, Tfe, KP1_481902250 * Tfh);
|
Chris@10
|
3163 R1[WS(rs, 23)] = Tfj - Tfk;
|
Chris@10
|
3164 R1[WS(rs, 55)] = Tfj + Tfk;
|
Chris@10
|
3165 Tf3 = TeX + TeY;
|
Chris@10
|
3166 Tf4 = FMA(KP1_807978586, Tf0, KP855110186 * Tf1);
|
Chris@10
|
3167 R1[WS(rs, 27)] = Tf3 - Tf4;
|
Chris@10
|
3168 R1[WS(rs, 59)] = Tf3 + Tf4;
|
Chris@10
|
3169 }
|
Chris@10
|
3170 }
|
Chris@10
|
3171 }
|
Chris@10
|
3172 }
|
Chris@10
|
3173 }
|
Chris@10
|
3174
|
Chris@10
|
3175 static const kr2c_desc desc = { 128, "r2cb_128", {812, 198, 144, 0}, &GENUS };
|
Chris@10
|
3176
|
Chris@10
|
3177 void X(codelet_r2cb_128) (planner *p) {
|
Chris@10
|
3178 X(kr2c_register) (p, r2cb_128, &desc);
|
Chris@10
|
3179 }
|
Chris@10
|
3180
|
Chris@10
|
3181 #endif /* HAVE_FMA */
|