cannam@167
|
1 /*
|
cannam@167
|
2 * Copyright (c) 2003, 2007-14 Matteo Frigo
|
cannam@167
|
3 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
|
cannam@167
|
4 *
|
cannam@167
|
5 * This program is free software; you can redistribute it and/or modify
|
cannam@167
|
6 * it under the terms of the GNU General Public License as published by
|
cannam@167
|
7 * the Free Software Foundation; either version 2 of the License, or
|
cannam@167
|
8 * (at your option) any later version.
|
cannam@167
|
9 *
|
cannam@167
|
10 * This program is distributed in the hope that it will be useful,
|
cannam@167
|
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
cannam@167
|
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
cannam@167
|
13 * GNU General Public License for more details.
|
cannam@167
|
14 *
|
cannam@167
|
15 * You should have received a copy of the GNU General Public License
|
cannam@167
|
16 * along with this program; if not, write to the Free Software
|
cannam@167
|
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
cannam@167
|
18 *
|
cannam@167
|
19 */
|
cannam@167
|
20
|
cannam@167
|
21 /* This file was automatically generated --- DO NOT EDIT */
|
cannam@167
|
22 /* Generated on Thu May 24 08:06:27 EDT 2018 */
|
cannam@167
|
23
|
cannam@167
|
24 #include "rdft/codelet-rdft.h"
|
cannam@167
|
25
|
cannam@167
|
26 #if defined(ARCH_PREFERS_FMA) || defined(ISA_EXTENSION_PREFERS_FMA)
|
cannam@167
|
27
|
cannam@167
|
28 /* Generated by: ../../../genfft/gen_r2cf.native -fma -compact -variables 4 -pipeline-latency 4 -n 128 -name r2cf_128 -include rdft/scalar/r2cf.h */
|
cannam@167
|
29
|
cannam@167
|
30 /*
|
cannam@167
|
31 * This function contains 956 FP additions, 516 FP multiplications,
|
cannam@167
|
32 * (or, 440 additions, 0 multiplications, 516 fused multiply/add),
|
cannam@167
|
33 * 186 stack variables, 31 constants, and 256 memory accesses
|
cannam@167
|
34 */
|
cannam@167
|
35 #include "rdft/scalar/r2cf.h"
|
cannam@167
|
36
|
cannam@167
|
37 static void r2cf_128(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
|
cannam@167
|
38 {
|
cannam@167
|
39 DK(KP803207531, +0.803207531480644909806676512963141923879569427);
|
cannam@167
|
40 DK(KP989176509, +0.989176509964780973451673738016243063983689533);
|
cannam@167
|
41 DK(KP741650546, +0.741650546272035369581266691172079863842265220);
|
cannam@167
|
42 DK(KP148335987, +0.148335987538347428753676511486911367000625355);
|
cannam@167
|
43 DK(KP740951125, +0.740951125354959091175616897495162729728955309);
|
cannam@167
|
44 DK(KP998795456, +0.998795456205172392714771604759100694443203615);
|
cannam@167
|
45 DK(KP906347169, +0.906347169019147157946142717268914412664134293);
|
cannam@167
|
46 DK(KP049126849, +0.049126849769467254105343321271313617079695752);
|
cannam@167
|
47 DK(KP970031253, +0.970031253194543992603984207286100251456865962);
|
cannam@167
|
48 DK(KP857728610, +0.857728610000272069902269984284770137042490799);
|
cannam@167
|
49 DK(KP250486960, +0.250486960191305461595702160124721208578685568);
|
cannam@167
|
50 DK(KP599376933, +0.599376933681923766271389869014404232837890546);
|
cannam@167
|
51 DK(KP941544065, +0.941544065183020778412509402599502357185589796);
|
cannam@167
|
52 DK(KP903989293, +0.903989293123443331586200297230537048710132025);
|
cannam@167
|
53 DK(KP357805721, +0.357805721314524104672487743774474392487532769);
|
cannam@167
|
54 DK(KP472964775, +0.472964775891319928124438237972992463904131113);
|
cannam@167
|
55 DK(KP773010453, +0.773010453362736960810906609758469800971041293);
|
cannam@167
|
56 DK(KP995184726, +0.995184726672196886244836953109479921575474869);
|
cannam@167
|
57 DK(KP098491403, +0.098491403357164253077197521291327432293052451);
|
cannam@167
|
58 DK(KP820678790, +0.820678790828660330972281985331011598767386482);
|
cannam@167
|
59 DK(KP956940335, +0.956940335732208864935797886980269969482849206);
|
cannam@167
|
60 DK(KP881921264, +0.881921264348355029712756863660388349508442621);
|
cannam@167
|
61 DK(KP534511135, +0.534511135950791641089685961295362908582039528);
|
cannam@167
|
62 DK(KP303346683, +0.303346683607342391675883946941299872384187453);
|
cannam@167
|
63 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
cannam@167
|
64 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
cannam@167
|
65 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
cannam@167
|
66 DK(KP668178637, +0.668178637919298919997757686523080761552472251);
|
cannam@167
|
67 DK(KP198912367, +0.198912367379658006911597622644676228597850501);
|
cannam@167
|
68 DK(KP414213562, +0.414213562373095048801688724209698078569671875);
|
cannam@167
|
69 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
cannam@167
|
70 {
|
cannam@167
|
71 INT i;
|
cannam@167
|
72 for (i = v; i > 0; i = i - 1, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(512, rs), MAKE_VOLATILE_STRIDE(512, csr), MAKE_VOLATILE_STRIDE(512, csi)) {
|
cannam@167
|
73 E TcD, TdR, T27, T7r, T5P, T8v, Tf, Ta5, Tu, Tbn, TcG, TdS, T2e, T8w, T5S;
|
cannam@167
|
74 E T7s, TK, Ta6, TcK, TdU, T2o, T5U, T7w, T8y, TZ, Ta7, TcN, TdV, T2x, T5V;
|
cannam@167
|
75 E T7z, T8z, T1g, Taa, TcU, Tex, TcX, Tew, T1v, Tab, T2M, T6A, T7E, T9b, T7H;
|
cannam@167
|
76 E T9a, T2T, T6z, T4X, T6L, Tdz, TeL, TdK, TeO, T5G, T6O, T8d, T9p, TaV, Tc3;
|
cannam@167
|
77 E Tbi, Tc4, T8o, T9s, T3I, T6E, Tde, TeE, Tdp, TeH, T4r, T6H, T7U, T9i, Tao;
|
cannam@167
|
78 E TbW, TaL, TbX, T85, T9l, T1L, Tad, Td3, TeA, Td6, Tez, T20, Tae, T37, T6x;
|
cannam@167
|
79 E T7L, T9e, T7O, T9d, T3e, T6w, TbZ, Tc0, T3Z, T4s, Tds, TeF, T4g, T4t, T80;
|
cannam@167
|
80 E T87, Tdl, TeI, T7X, T86, TaD, TaM, Tc6, Tc7, T5e, T5H, TdN, TeM, T5v, T5I;
|
cannam@167
|
81 E T8j, T8q, TdG, TeP, T8g, T8p, Tba, Tbj;
|
cannam@167
|
82 {
|
cannam@167
|
83 E T3, T23, Td, T25, T6, T5N, Ta, T24;
|
cannam@167
|
84 {
|
cannam@167
|
85 E T1, T2, Tb, Tc;
|
cannam@167
|
86 T1 = R0[0];
|
cannam@167
|
87 T2 = R0[WS(rs, 32)];
|
cannam@167
|
88 T3 = T1 + T2;
|
cannam@167
|
89 T23 = T1 - T2;
|
cannam@167
|
90 Tb = R0[WS(rs, 56)];
|
cannam@167
|
91 Tc = R0[WS(rs, 24)];
|
cannam@167
|
92 Td = Tb + Tc;
|
cannam@167
|
93 T25 = Tb - Tc;
|
cannam@167
|
94 }
|
cannam@167
|
95 {
|
cannam@167
|
96 E T4, T5, T8, T9;
|
cannam@167
|
97 T4 = R0[WS(rs, 16)];
|
cannam@167
|
98 T5 = R0[WS(rs, 48)];
|
cannam@167
|
99 T6 = T4 + T5;
|
cannam@167
|
100 T5N = T4 - T5;
|
cannam@167
|
101 T8 = R0[WS(rs, 8)];
|
cannam@167
|
102 T9 = R0[WS(rs, 40)];
|
cannam@167
|
103 Ta = T8 + T9;
|
cannam@167
|
104 T24 = T8 - T9;
|
cannam@167
|
105 }
|
cannam@167
|
106 TcD = T3 - T6;
|
cannam@167
|
107 TdR = Td - Ta;
|
cannam@167
|
108 {
|
cannam@167
|
109 E T26, T5O, T7, Te;
|
cannam@167
|
110 T26 = T24 + T25;
|
cannam@167
|
111 T27 = FMA(KP707106781, T26, T23);
|
cannam@167
|
112 T7r = FNMS(KP707106781, T26, T23);
|
cannam@167
|
113 T5O = T25 - T24;
|
cannam@167
|
114 T5P = FNMS(KP707106781, T5O, T5N);
|
cannam@167
|
115 T8v = FMA(KP707106781, T5O, T5N);
|
cannam@167
|
116 T7 = T3 + T6;
|
cannam@167
|
117 Te = Ta + Td;
|
cannam@167
|
118 Tf = T7 + Te;
|
cannam@167
|
119 Ta5 = T7 - Te;
|
cannam@167
|
120 }
|
cannam@167
|
121 }
|
cannam@167
|
122 {
|
cannam@167
|
123 E Ti, T28, Ts, T2c, Tl, T29, Tp, T2b;
|
cannam@167
|
124 {
|
cannam@167
|
125 E Tg, Th, Tq, Tr;
|
cannam@167
|
126 Tg = R0[WS(rs, 4)];
|
cannam@167
|
127 Th = R0[WS(rs, 36)];
|
cannam@167
|
128 Ti = Tg + Th;
|
cannam@167
|
129 T28 = Tg - Th;
|
cannam@167
|
130 Tq = R0[WS(rs, 12)];
|
cannam@167
|
131 Tr = R0[WS(rs, 44)];
|
cannam@167
|
132 Ts = Tq + Tr;
|
cannam@167
|
133 T2c = Tq - Tr;
|
cannam@167
|
134 }
|
cannam@167
|
135 {
|
cannam@167
|
136 E Tj, Tk, Tn, To;
|
cannam@167
|
137 Tj = R0[WS(rs, 20)];
|
cannam@167
|
138 Tk = R0[WS(rs, 52)];
|
cannam@167
|
139 Tl = Tj + Tk;
|
cannam@167
|
140 T29 = Tj - Tk;
|
cannam@167
|
141 Tn = R0[WS(rs, 60)];
|
cannam@167
|
142 To = R0[WS(rs, 28)];
|
cannam@167
|
143 Tp = Tn + To;
|
cannam@167
|
144 T2b = Tn - To;
|
cannam@167
|
145 }
|
cannam@167
|
146 {
|
cannam@167
|
147 E Tm, Tt, TcE, TcF;
|
cannam@167
|
148 Tm = Ti + Tl;
|
cannam@167
|
149 Tt = Tp + Ts;
|
cannam@167
|
150 Tu = Tm + Tt;
|
cannam@167
|
151 Tbn = Tt - Tm;
|
cannam@167
|
152 TcE = Ti - Tl;
|
cannam@167
|
153 TcF = Tp - Ts;
|
cannam@167
|
154 TcG = TcE + TcF;
|
cannam@167
|
155 TdS = TcF - TcE;
|
cannam@167
|
156 }
|
cannam@167
|
157 {
|
cannam@167
|
158 E T2a, T2d, T5Q, T5R;
|
cannam@167
|
159 T2a = FNMS(KP414213562, T29, T28);
|
cannam@167
|
160 T2d = FMA(KP414213562, T2c, T2b);
|
cannam@167
|
161 T2e = T2a + T2d;
|
cannam@167
|
162 T8w = T2d - T2a;
|
cannam@167
|
163 T5Q = FMA(KP414213562, T28, T29);
|
cannam@167
|
164 T5R = FNMS(KP414213562, T2b, T2c);
|
cannam@167
|
165 T5S = T5Q + T5R;
|
cannam@167
|
166 T7s = T5Q - T5R;
|
cannam@167
|
167 }
|
cannam@167
|
168 }
|
cannam@167
|
169 {
|
cannam@167
|
170 E Ty, T2g, TB, T2l, TF, T2m, TI, T2j;
|
cannam@167
|
171 {
|
cannam@167
|
172 E Tw, Tx, Tz, TA;
|
cannam@167
|
173 Tw = R0[WS(rs, 2)];
|
cannam@167
|
174 Tx = R0[WS(rs, 34)];
|
cannam@167
|
175 Ty = Tw + Tx;
|
cannam@167
|
176 T2g = Tw - Tx;
|
cannam@167
|
177 Tz = R0[WS(rs, 18)];
|
cannam@167
|
178 TA = R0[WS(rs, 50)];
|
cannam@167
|
179 TB = Tz + TA;
|
cannam@167
|
180 T2l = Tz - TA;
|
cannam@167
|
181 {
|
cannam@167
|
182 E TD, TE, T2h, TG, TH, T2i;
|
cannam@167
|
183 TD = R0[WS(rs, 10)];
|
cannam@167
|
184 TE = R0[WS(rs, 42)];
|
cannam@167
|
185 T2h = TD - TE;
|
cannam@167
|
186 TG = R0[WS(rs, 58)];
|
cannam@167
|
187 TH = R0[WS(rs, 26)];
|
cannam@167
|
188 T2i = TG - TH;
|
cannam@167
|
189 TF = TD + TE;
|
cannam@167
|
190 T2m = T2h - T2i;
|
cannam@167
|
191 TI = TG + TH;
|
cannam@167
|
192 T2j = T2h + T2i;
|
cannam@167
|
193 }
|
cannam@167
|
194 }
|
cannam@167
|
195 {
|
cannam@167
|
196 E TC, TJ, TcI, TcJ;
|
cannam@167
|
197 TC = Ty + TB;
|
cannam@167
|
198 TJ = TF + TI;
|
cannam@167
|
199 TK = TC + TJ;
|
cannam@167
|
200 Ta6 = TC - TJ;
|
cannam@167
|
201 TcI = Ty - TB;
|
cannam@167
|
202 TcJ = TI - TF;
|
cannam@167
|
203 TcK = FMA(KP414213562, TcJ, TcI);
|
cannam@167
|
204 TdU = FNMS(KP414213562, TcI, TcJ);
|
cannam@167
|
205 }
|
cannam@167
|
206 {
|
cannam@167
|
207 E T2k, T2n, T7u, T7v;
|
cannam@167
|
208 T2k = FMA(KP707106781, T2j, T2g);
|
cannam@167
|
209 T2n = FMA(KP707106781, T2m, T2l);
|
cannam@167
|
210 T2o = FNMS(KP198912367, T2n, T2k);
|
cannam@167
|
211 T5U = FMA(KP198912367, T2k, T2n);
|
cannam@167
|
212 T7u = FNMS(KP707106781, T2j, T2g);
|
cannam@167
|
213 T7v = FNMS(KP707106781, T2m, T2l);
|
cannam@167
|
214 T7w = FMA(KP668178637, T7v, T7u);
|
cannam@167
|
215 T8y = FNMS(KP668178637, T7u, T7v);
|
cannam@167
|
216 }
|
cannam@167
|
217 }
|
cannam@167
|
218 {
|
cannam@167
|
219 E TN, T2p, TQ, T2u, TU, T2v, TX, T2s;
|
cannam@167
|
220 {
|
cannam@167
|
221 E TL, TM, TO, TP;
|
cannam@167
|
222 TL = R0[WS(rs, 62)];
|
cannam@167
|
223 TM = R0[WS(rs, 30)];
|
cannam@167
|
224 TN = TL + TM;
|
cannam@167
|
225 T2p = TL - TM;
|
cannam@167
|
226 TO = R0[WS(rs, 14)];
|
cannam@167
|
227 TP = R0[WS(rs, 46)];
|
cannam@167
|
228 TQ = TO + TP;
|
cannam@167
|
229 T2u = TO - TP;
|
cannam@167
|
230 {
|
cannam@167
|
231 E TS, TT, T2q, TV, TW, T2r;
|
cannam@167
|
232 TS = R0[WS(rs, 6)];
|
cannam@167
|
233 TT = R0[WS(rs, 38)];
|
cannam@167
|
234 T2q = TS - TT;
|
cannam@167
|
235 TV = R0[WS(rs, 54)];
|
cannam@167
|
236 TW = R0[WS(rs, 22)];
|
cannam@167
|
237 T2r = TV - TW;
|
cannam@167
|
238 TU = TS + TT;
|
cannam@167
|
239 T2v = T2q - T2r;
|
cannam@167
|
240 TX = TV + TW;
|
cannam@167
|
241 T2s = T2q + T2r;
|
cannam@167
|
242 }
|
cannam@167
|
243 }
|
cannam@167
|
244 {
|
cannam@167
|
245 E TR, TY, TcL, TcM;
|
cannam@167
|
246 TR = TN + TQ;
|
cannam@167
|
247 TY = TU + TX;
|
cannam@167
|
248 TZ = TR + TY;
|
cannam@167
|
249 Ta7 = TR - TY;
|
cannam@167
|
250 TcL = TN - TQ;
|
cannam@167
|
251 TcM = TX - TU;
|
cannam@167
|
252 TcN = FNMS(KP414213562, TcM, TcL);
|
cannam@167
|
253 TdV = FMA(KP414213562, TcL, TcM);
|
cannam@167
|
254 }
|
cannam@167
|
255 {
|
cannam@167
|
256 E T2t, T2w, T7x, T7y;
|
cannam@167
|
257 T2t = FMA(KP707106781, T2s, T2p);
|
cannam@167
|
258 T2w = FMA(KP707106781, T2v, T2u);
|
cannam@167
|
259 T2x = FMA(KP198912367, T2w, T2t);
|
cannam@167
|
260 T5V = FNMS(KP198912367, T2t, T2w);
|
cannam@167
|
261 T7x = FNMS(KP707106781, T2s, T2p);
|
cannam@167
|
262 T7y = FNMS(KP707106781, T2v, T2u);
|
cannam@167
|
263 T7z = FNMS(KP668178637, T7y, T7x);
|
cannam@167
|
264 T8z = FMA(KP668178637, T7x, T7y);
|
cannam@167
|
265 }
|
cannam@167
|
266 }
|
cannam@167
|
267 {
|
cannam@167
|
268 E T14, T2A, T17, T2N, T1b, T2O, T1e, T2D, T1j, T1m, T2H, TcR, T2Q, T1q, T1t;
|
cannam@167
|
269 E T2K, TcS, T2R;
|
cannam@167
|
270 {
|
cannam@167
|
271 E T12, T13, T15, T16;
|
cannam@167
|
272 T12 = R0[WS(rs, 1)];
|
cannam@167
|
273 T13 = R0[WS(rs, 33)];
|
cannam@167
|
274 T14 = T12 + T13;
|
cannam@167
|
275 T2A = T12 - T13;
|
cannam@167
|
276 T15 = R0[WS(rs, 17)];
|
cannam@167
|
277 T16 = R0[WS(rs, 49)];
|
cannam@167
|
278 T17 = T15 + T16;
|
cannam@167
|
279 T2N = T15 - T16;
|
cannam@167
|
280 }
|
cannam@167
|
281 {
|
cannam@167
|
282 E T19, T1a, T2B, T1c, T1d, T2C;
|
cannam@167
|
283 T19 = R0[WS(rs, 9)];
|
cannam@167
|
284 T1a = R0[WS(rs, 41)];
|
cannam@167
|
285 T2B = T19 - T1a;
|
cannam@167
|
286 T1c = R0[WS(rs, 57)];
|
cannam@167
|
287 T1d = R0[WS(rs, 25)];
|
cannam@167
|
288 T2C = T1c - T1d;
|
cannam@167
|
289 T1b = T19 + T1a;
|
cannam@167
|
290 T2O = T2B - T2C;
|
cannam@167
|
291 T1e = T1c + T1d;
|
cannam@167
|
292 T2D = T2B + T2C;
|
cannam@167
|
293 }
|
cannam@167
|
294 {
|
cannam@167
|
295 E T2F, T2G, T2I, T2J;
|
cannam@167
|
296 {
|
cannam@167
|
297 E T1h, T1i, T1k, T1l;
|
cannam@167
|
298 T1h = R0[WS(rs, 5)];
|
cannam@167
|
299 T1i = R0[WS(rs, 37)];
|
cannam@167
|
300 T1j = T1h + T1i;
|
cannam@167
|
301 T2F = T1h - T1i;
|
cannam@167
|
302 T1k = R0[WS(rs, 21)];
|
cannam@167
|
303 T1l = R0[WS(rs, 53)];
|
cannam@167
|
304 T1m = T1k + T1l;
|
cannam@167
|
305 T2G = T1k - T1l;
|
cannam@167
|
306 }
|
cannam@167
|
307 T2H = FNMS(KP414213562, T2G, T2F);
|
cannam@167
|
308 TcR = T1j - T1m;
|
cannam@167
|
309 T2Q = FMA(KP414213562, T2F, T2G);
|
cannam@167
|
310 {
|
cannam@167
|
311 E T1o, T1p, T1r, T1s;
|
cannam@167
|
312 T1o = R0[WS(rs, 61)];
|
cannam@167
|
313 T1p = R0[WS(rs, 29)];
|
cannam@167
|
314 T1q = T1o + T1p;
|
cannam@167
|
315 T2I = T1o - T1p;
|
cannam@167
|
316 T1r = R0[WS(rs, 13)];
|
cannam@167
|
317 T1s = R0[WS(rs, 45)];
|
cannam@167
|
318 T1t = T1r + T1s;
|
cannam@167
|
319 T2J = T1r - T1s;
|
cannam@167
|
320 }
|
cannam@167
|
321 T2K = FMA(KP414213562, T2J, T2I);
|
cannam@167
|
322 TcS = T1q - T1t;
|
cannam@167
|
323 T2R = FNMS(KP414213562, T2I, T2J);
|
cannam@167
|
324 }
|
cannam@167
|
325 {
|
cannam@167
|
326 E T18, T1f, TcQ, TcT;
|
cannam@167
|
327 T18 = T14 + T17;
|
cannam@167
|
328 T1f = T1b + T1e;
|
cannam@167
|
329 T1g = T18 + T1f;
|
cannam@167
|
330 Taa = T18 - T1f;
|
cannam@167
|
331 TcQ = T14 - T17;
|
cannam@167
|
332 TcT = TcR + TcS;
|
cannam@167
|
333 TcU = FMA(KP707106781, TcT, TcQ);
|
cannam@167
|
334 Tex = FNMS(KP707106781, TcT, TcQ);
|
cannam@167
|
335 }
|
cannam@167
|
336 {
|
cannam@167
|
337 E TcV, TcW, T1n, T1u;
|
cannam@167
|
338 TcV = T1e - T1b;
|
cannam@167
|
339 TcW = TcS - TcR;
|
cannam@167
|
340 TcX = FMA(KP707106781, TcW, TcV);
|
cannam@167
|
341 Tew = FNMS(KP707106781, TcW, TcV);
|
cannam@167
|
342 T1n = T1j + T1m;
|
cannam@167
|
343 T1u = T1q + T1t;
|
cannam@167
|
344 T1v = T1n + T1u;
|
cannam@167
|
345 Tab = T1u - T1n;
|
cannam@167
|
346 }
|
cannam@167
|
347 {
|
cannam@167
|
348 E T2E, T2L, T7C, T7D;
|
cannam@167
|
349 T2E = FMA(KP707106781, T2D, T2A);
|
cannam@167
|
350 T2L = T2H + T2K;
|
cannam@167
|
351 T2M = FMA(KP923879532, T2L, T2E);
|
cannam@167
|
352 T6A = FNMS(KP923879532, T2L, T2E);
|
cannam@167
|
353 T7C = FNMS(KP707106781, T2D, T2A);
|
cannam@167
|
354 T7D = T2Q - T2R;
|
cannam@167
|
355 T7E = FMA(KP923879532, T7D, T7C);
|
cannam@167
|
356 T9b = FNMS(KP923879532, T7D, T7C);
|
cannam@167
|
357 }
|
cannam@167
|
358 {
|
cannam@167
|
359 E T7F, T7G, T2P, T2S;
|
cannam@167
|
360 T7F = FNMS(KP707106781, T2O, T2N);
|
cannam@167
|
361 T7G = T2K - T2H;
|
cannam@167
|
362 T7H = FMA(KP923879532, T7G, T7F);
|
cannam@167
|
363 T9a = FNMS(KP923879532, T7G, T7F);
|
cannam@167
|
364 T2P = FMA(KP707106781, T2O, T2N);
|
cannam@167
|
365 T2S = T2Q + T2R;
|
cannam@167
|
366 T2T = FMA(KP923879532, T2S, T2P);
|
cannam@167
|
367 T6z = FNMS(KP923879532, T2S, T2P);
|
cannam@167
|
368 }
|
cannam@167
|
369 }
|
cannam@167
|
370 {
|
cannam@167
|
371 E T4z, TaP, T5A, TaQ, T4G, TaT, T5B, TaS, Tbf, Tbg, T4O, Tdw, T5D, Tbc, Tbd;
|
cannam@167
|
372 E T4V, Tdx, T5E;
|
cannam@167
|
373 {
|
cannam@167
|
374 E T4x, T4y, T5y, T5z;
|
cannam@167
|
375 T4x = R1[WS(rs, 63)];
|
cannam@167
|
376 T4y = R1[WS(rs, 31)];
|
cannam@167
|
377 T4z = T4x - T4y;
|
cannam@167
|
378 TaP = T4x + T4y;
|
cannam@167
|
379 T5y = R1[WS(rs, 47)];
|
cannam@167
|
380 T5z = R1[WS(rs, 15)];
|
cannam@167
|
381 T5A = T5y - T5z;
|
cannam@167
|
382 TaQ = T5z + T5y;
|
cannam@167
|
383 }
|
cannam@167
|
384 {
|
cannam@167
|
385 E T4A, T4B, T4C, T4D, T4E, T4F;
|
cannam@167
|
386 T4A = R1[WS(rs, 7)];
|
cannam@167
|
387 T4B = R1[WS(rs, 39)];
|
cannam@167
|
388 T4C = T4A - T4B;
|
cannam@167
|
389 T4D = R1[WS(rs, 55)];
|
cannam@167
|
390 T4E = R1[WS(rs, 23)];
|
cannam@167
|
391 T4F = T4D - T4E;
|
cannam@167
|
392 T4G = T4C + T4F;
|
cannam@167
|
393 TaT = T4D + T4E;
|
cannam@167
|
394 T5B = T4F - T4C;
|
cannam@167
|
395 TaS = T4A + T4B;
|
cannam@167
|
396 }
|
cannam@167
|
397 {
|
cannam@167
|
398 E T4K, T4N, T4R, T4U;
|
cannam@167
|
399 {
|
cannam@167
|
400 E T4I, T4J, T4L, T4M;
|
cannam@167
|
401 T4I = R1[WS(rs, 3)];
|
cannam@167
|
402 T4J = R1[WS(rs, 35)];
|
cannam@167
|
403 T4K = T4I - T4J;
|
cannam@167
|
404 Tbf = T4I + T4J;
|
cannam@167
|
405 T4L = R1[WS(rs, 51)];
|
cannam@167
|
406 T4M = R1[WS(rs, 19)];
|
cannam@167
|
407 T4N = T4L - T4M;
|
cannam@167
|
408 Tbg = T4M + T4L;
|
cannam@167
|
409 }
|
cannam@167
|
410 T4O = FMA(KP414213562, T4N, T4K);
|
cannam@167
|
411 Tdw = Tbf - Tbg;
|
cannam@167
|
412 T5D = FNMS(KP414213562, T4K, T4N);
|
cannam@167
|
413 {
|
cannam@167
|
414 E T4P, T4Q, T4S, T4T;
|
cannam@167
|
415 T4P = R1[WS(rs, 59)];
|
cannam@167
|
416 T4Q = R1[WS(rs, 27)];
|
cannam@167
|
417 T4R = T4P - T4Q;
|
cannam@167
|
418 Tbc = T4P + T4Q;
|
cannam@167
|
419 T4S = R1[WS(rs, 43)];
|
cannam@167
|
420 T4T = R1[WS(rs, 11)];
|
cannam@167
|
421 T4U = T4S - T4T;
|
cannam@167
|
422 Tbd = T4T + T4S;
|
cannam@167
|
423 }
|
cannam@167
|
424 T4V = FNMS(KP414213562, T4U, T4R);
|
cannam@167
|
425 Tdx = Tbc - Tbd;
|
cannam@167
|
426 T5E = FMA(KP414213562, T4R, T4U);
|
cannam@167
|
427 }
|
cannam@167
|
428 {
|
cannam@167
|
429 E T4H, T4W, Tdv, Tdy;
|
cannam@167
|
430 T4H = FMA(KP707106781, T4G, T4z);
|
cannam@167
|
431 T4W = T4O + T4V;
|
cannam@167
|
432 T4X = FMA(KP923879532, T4W, T4H);
|
cannam@167
|
433 T6L = FNMS(KP923879532, T4W, T4H);
|
cannam@167
|
434 Tdv = TaP - TaQ;
|
cannam@167
|
435 Tdy = Tdw + Tdx;
|
cannam@167
|
436 Tdz = FMA(KP707106781, Tdy, Tdv);
|
cannam@167
|
437 TeL = FNMS(KP707106781, Tdy, Tdv);
|
cannam@167
|
438 }
|
cannam@167
|
439 {
|
cannam@167
|
440 E TdI, TdJ, T5C, T5F;
|
cannam@167
|
441 TdI = TaT - TaS;
|
cannam@167
|
442 TdJ = Tdx - Tdw;
|
cannam@167
|
443 TdK = FMA(KP707106781, TdJ, TdI);
|
cannam@167
|
444 TeO = FNMS(KP707106781, TdJ, TdI);
|
cannam@167
|
445 T5C = FMA(KP707106781, T5B, T5A);
|
cannam@167
|
446 T5F = T5D + T5E;
|
cannam@167
|
447 T5G = FMA(KP923879532, T5F, T5C);
|
cannam@167
|
448 T6O = FNMS(KP923879532, T5F, T5C);
|
cannam@167
|
449 }
|
cannam@167
|
450 {
|
cannam@167
|
451 E T8b, T8c, TaR, TaU;
|
cannam@167
|
452 T8b = FNMS(KP707106781, T4G, T4z);
|
cannam@167
|
453 T8c = T5E - T5D;
|
cannam@167
|
454 T8d = FMA(KP923879532, T8c, T8b);
|
cannam@167
|
455 T9p = FNMS(KP923879532, T8c, T8b);
|
cannam@167
|
456 TaR = TaP + TaQ;
|
cannam@167
|
457 TaU = TaS + TaT;
|
cannam@167
|
458 TaV = TaR - TaU;
|
cannam@167
|
459 Tc3 = TaR + TaU;
|
cannam@167
|
460 }
|
cannam@167
|
461 {
|
cannam@167
|
462 E Tbe, Tbh, T8m, T8n;
|
cannam@167
|
463 Tbe = Tbc + Tbd;
|
cannam@167
|
464 Tbh = Tbf + Tbg;
|
cannam@167
|
465 Tbi = Tbe - Tbh;
|
cannam@167
|
466 Tc4 = Tbh + Tbe;
|
cannam@167
|
467 T8m = FNMS(KP707106781, T5B, T5A);
|
cannam@167
|
468 T8n = T4V - T4O;
|
cannam@167
|
469 T8o = FNMS(KP923879532, T8n, T8m);
|
cannam@167
|
470 T9s = FMA(KP923879532, T8n, T8m);
|
cannam@167
|
471 }
|
cannam@167
|
472 }
|
cannam@167
|
473 {
|
cannam@167
|
474 E T3k, Tai, T4l, Taj, T3r, Tam, T4m, Tal, TaI, TaJ, T3z, Tdb, T4o, TaF, TaG;
|
cannam@167
|
475 E T3G, Tdc, T4p;
|
cannam@167
|
476 {
|
cannam@167
|
477 E T3i, T3j, T4j, T4k;
|
cannam@167
|
478 T3i = R1[0];
|
cannam@167
|
479 T3j = R1[WS(rs, 32)];
|
cannam@167
|
480 T3k = T3i - T3j;
|
cannam@167
|
481 Tai = T3i + T3j;
|
cannam@167
|
482 T4j = R1[WS(rs, 16)];
|
cannam@167
|
483 T4k = R1[WS(rs, 48)];
|
cannam@167
|
484 T4l = T4j - T4k;
|
cannam@167
|
485 Taj = T4j + T4k;
|
cannam@167
|
486 }
|
cannam@167
|
487 {
|
cannam@167
|
488 E T3l, T3m, T3n, T3o, T3p, T3q;
|
cannam@167
|
489 T3l = R1[WS(rs, 8)];
|
cannam@167
|
490 T3m = R1[WS(rs, 40)];
|
cannam@167
|
491 T3n = T3l - T3m;
|
cannam@167
|
492 T3o = R1[WS(rs, 56)];
|
cannam@167
|
493 T3p = R1[WS(rs, 24)];
|
cannam@167
|
494 T3q = T3o - T3p;
|
cannam@167
|
495 T3r = T3n + T3q;
|
cannam@167
|
496 Tam = T3o + T3p;
|
cannam@167
|
497 T4m = T3n - T3q;
|
cannam@167
|
498 Tal = T3l + T3m;
|
cannam@167
|
499 }
|
cannam@167
|
500 {
|
cannam@167
|
501 E T3v, T3y, T3C, T3F;
|
cannam@167
|
502 {
|
cannam@167
|
503 E T3t, T3u, T3w, T3x;
|
cannam@167
|
504 T3t = R1[WS(rs, 4)];
|
cannam@167
|
505 T3u = R1[WS(rs, 36)];
|
cannam@167
|
506 T3v = T3t - T3u;
|
cannam@167
|
507 TaI = T3t + T3u;
|
cannam@167
|
508 T3w = R1[WS(rs, 20)];
|
cannam@167
|
509 T3x = R1[WS(rs, 52)];
|
cannam@167
|
510 T3y = T3w - T3x;
|
cannam@167
|
511 TaJ = T3w + T3x;
|
cannam@167
|
512 }
|
cannam@167
|
513 T3z = FNMS(KP414213562, T3y, T3v);
|
cannam@167
|
514 Tdb = TaI - TaJ;
|
cannam@167
|
515 T4o = FMA(KP414213562, T3v, T3y);
|
cannam@167
|
516 {
|
cannam@167
|
517 E T3A, T3B, T3D, T3E;
|
cannam@167
|
518 T3A = R1[WS(rs, 60)];
|
cannam@167
|
519 T3B = R1[WS(rs, 28)];
|
cannam@167
|
520 T3C = T3A - T3B;
|
cannam@167
|
521 TaF = T3A + T3B;
|
cannam@167
|
522 T3D = R1[WS(rs, 12)];
|
cannam@167
|
523 T3E = R1[WS(rs, 44)];
|
cannam@167
|
524 T3F = T3D - T3E;
|
cannam@167
|
525 TaG = T3D + T3E;
|
cannam@167
|
526 }
|
cannam@167
|
527 T3G = FMA(KP414213562, T3F, T3C);
|
cannam@167
|
528 Tdc = TaF - TaG;
|
cannam@167
|
529 T4p = FNMS(KP414213562, T3C, T3F);
|
cannam@167
|
530 }
|
cannam@167
|
531 {
|
cannam@167
|
532 E T3s, T3H, Tda, Tdd;
|
cannam@167
|
533 T3s = FMA(KP707106781, T3r, T3k);
|
cannam@167
|
534 T3H = T3z + T3G;
|
cannam@167
|
535 T3I = FMA(KP923879532, T3H, T3s);
|
cannam@167
|
536 T6E = FNMS(KP923879532, T3H, T3s);
|
cannam@167
|
537 Tda = Tai - Taj;
|
cannam@167
|
538 Tdd = Tdb + Tdc;
|
cannam@167
|
539 Tde = FMA(KP707106781, Tdd, Tda);
|
cannam@167
|
540 TeE = FNMS(KP707106781, Tdd, Tda);
|
cannam@167
|
541 }
|
cannam@167
|
542 {
|
cannam@167
|
543 E Tdn, Tdo, T4n, T4q;
|
cannam@167
|
544 Tdn = Tam - Tal;
|
cannam@167
|
545 Tdo = Tdc - Tdb;
|
cannam@167
|
546 Tdp = FMA(KP707106781, Tdo, Tdn);
|
cannam@167
|
547 TeH = FNMS(KP707106781, Tdo, Tdn);
|
cannam@167
|
548 T4n = FMA(KP707106781, T4m, T4l);
|
cannam@167
|
549 T4q = T4o + T4p;
|
cannam@167
|
550 T4r = FMA(KP923879532, T4q, T4n);
|
cannam@167
|
551 T6H = FNMS(KP923879532, T4q, T4n);
|
cannam@167
|
552 }
|
cannam@167
|
553 {
|
cannam@167
|
554 E T7S, T7T, Tak, Tan;
|
cannam@167
|
555 T7S = FNMS(KP707106781, T3r, T3k);
|
cannam@167
|
556 T7T = T4o - T4p;
|
cannam@167
|
557 T7U = FMA(KP923879532, T7T, T7S);
|
cannam@167
|
558 T9i = FNMS(KP923879532, T7T, T7S);
|
cannam@167
|
559 Tak = Tai + Taj;
|
cannam@167
|
560 Tan = Tal + Tam;
|
cannam@167
|
561 Tao = Tak - Tan;
|
cannam@167
|
562 TbW = Tak + Tan;
|
cannam@167
|
563 }
|
cannam@167
|
564 {
|
cannam@167
|
565 E TaH, TaK, T83, T84;
|
cannam@167
|
566 TaH = TaF + TaG;
|
cannam@167
|
567 TaK = TaI + TaJ;
|
cannam@167
|
568 TaL = TaH - TaK;
|
cannam@167
|
569 TbX = TaK + TaH;
|
cannam@167
|
570 T83 = FNMS(KP707106781, T4m, T4l);
|
cannam@167
|
571 T84 = T3G - T3z;
|
cannam@167
|
572 T85 = FMA(KP923879532, T84, T83);
|
cannam@167
|
573 T9l = FNMS(KP923879532, T84, T83);
|
cannam@167
|
574 }
|
cannam@167
|
575 }
|
cannam@167
|
576 {
|
cannam@167
|
577 E T1z, T2V, T1C, T38, T1G, T39, T1J, T2Y, T1O, T1R, T32, Td0, T3b, T1V, T1Y;
|
cannam@167
|
578 E T35, Td1, T3c;
|
cannam@167
|
579 {
|
cannam@167
|
580 E T1x, T1y, T1A, T1B;
|
cannam@167
|
581 T1x = R0[WS(rs, 63)];
|
cannam@167
|
582 T1y = R0[WS(rs, 31)];
|
cannam@167
|
583 T1z = T1x + T1y;
|
cannam@167
|
584 T2V = T1x - T1y;
|
cannam@167
|
585 T1A = R0[WS(rs, 15)];
|
cannam@167
|
586 T1B = R0[WS(rs, 47)];
|
cannam@167
|
587 T1C = T1A + T1B;
|
cannam@167
|
588 T38 = T1A - T1B;
|
cannam@167
|
589 }
|
cannam@167
|
590 {
|
cannam@167
|
591 E T1E, T1F, T2W, T1H, T1I, T2X;
|
cannam@167
|
592 T1E = R0[WS(rs, 7)];
|
cannam@167
|
593 T1F = R0[WS(rs, 39)];
|
cannam@167
|
594 T2W = T1E - T1F;
|
cannam@167
|
595 T1H = R0[WS(rs, 55)];
|
cannam@167
|
596 T1I = R0[WS(rs, 23)];
|
cannam@167
|
597 T2X = T1H - T1I;
|
cannam@167
|
598 T1G = T1E + T1F;
|
cannam@167
|
599 T39 = T2W - T2X;
|
cannam@167
|
600 T1J = T1H + T1I;
|
cannam@167
|
601 T2Y = T2W + T2X;
|
cannam@167
|
602 }
|
cannam@167
|
603 {
|
cannam@167
|
604 E T30, T31, T33, T34;
|
cannam@167
|
605 {
|
cannam@167
|
606 E T1M, T1N, T1P, T1Q;
|
cannam@167
|
607 T1M = R0[WS(rs, 3)];
|
cannam@167
|
608 T1N = R0[WS(rs, 35)];
|
cannam@167
|
609 T1O = T1M + T1N;
|
cannam@167
|
610 T30 = T1M - T1N;
|
cannam@167
|
611 T1P = R0[WS(rs, 19)];
|
cannam@167
|
612 T1Q = R0[WS(rs, 51)];
|
cannam@167
|
613 T1R = T1P + T1Q;
|
cannam@167
|
614 T31 = T1P - T1Q;
|
cannam@167
|
615 }
|
cannam@167
|
616 T32 = FNMS(KP414213562, T31, T30);
|
cannam@167
|
617 Td0 = T1O - T1R;
|
cannam@167
|
618 T3b = FMA(KP414213562, T30, T31);
|
cannam@167
|
619 {
|
cannam@167
|
620 E T1T, T1U, T1W, T1X;
|
cannam@167
|
621 T1T = R0[WS(rs, 59)];
|
cannam@167
|
622 T1U = R0[WS(rs, 27)];
|
cannam@167
|
623 T1V = T1T + T1U;
|
cannam@167
|
624 T33 = T1T - T1U;
|
cannam@167
|
625 T1W = R0[WS(rs, 11)];
|
cannam@167
|
626 T1X = R0[WS(rs, 43)];
|
cannam@167
|
627 T1Y = T1W + T1X;
|
cannam@167
|
628 T34 = T1W - T1X;
|
cannam@167
|
629 }
|
cannam@167
|
630 T35 = FMA(KP414213562, T34, T33);
|
cannam@167
|
631 Td1 = T1V - T1Y;
|
cannam@167
|
632 T3c = FNMS(KP414213562, T33, T34);
|
cannam@167
|
633 }
|
cannam@167
|
634 {
|
cannam@167
|
635 E T1D, T1K, TcZ, Td2;
|
cannam@167
|
636 T1D = T1z + T1C;
|
cannam@167
|
637 T1K = T1G + T1J;
|
cannam@167
|
638 T1L = T1D + T1K;
|
cannam@167
|
639 Tad = T1D - T1K;
|
cannam@167
|
640 TcZ = T1z - T1C;
|
cannam@167
|
641 Td2 = Td0 + Td1;
|
cannam@167
|
642 Td3 = FMA(KP707106781, Td2, TcZ);
|
cannam@167
|
643 TeA = FNMS(KP707106781, Td2, TcZ);
|
cannam@167
|
644 }
|
cannam@167
|
645 {
|
cannam@167
|
646 E Td4, Td5, T1S, T1Z;
|
cannam@167
|
647 Td4 = T1J - T1G;
|
cannam@167
|
648 Td5 = Td1 - Td0;
|
cannam@167
|
649 Td6 = FMA(KP707106781, Td5, Td4);
|
cannam@167
|
650 Tez = FNMS(KP707106781, Td5, Td4);
|
cannam@167
|
651 T1S = T1O + T1R;
|
cannam@167
|
652 T1Z = T1V + T1Y;
|
cannam@167
|
653 T20 = T1S + T1Z;
|
cannam@167
|
654 Tae = T1Z - T1S;
|
cannam@167
|
655 }
|
cannam@167
|
656 {
|
cannam@167
|
657 E T2Z, T36, T7J, T7K;
|
cannam@167
|
658 T2Z = FMA(KP707106781, T2Y, T2V);
|
cannam@167
|
659 T36 = T32 + T35;
|
cannam@167
|
660 T37 = FMA(KP923879532, T36, T2Z);
|
cannam@167
|
661 T6x = FNMS(KP923879532, T36, T2Z);
|
cannam@167
|
662 T7J = FNMS(KP707106781, T2Y, T2V);
|
cannam@167
|
663 T7K = T3b - T3c;
|
cannam@167
|
664 T7L = FMA(KP923879532, T7K, T7J);
|
cannam@167
|
665 T9e = FNMS(KP923879532, T7K, T7J);
|
cannam@167
|
666 }
|
cannam@167
|
667 {
|
cannam@167
|
668 E T7M, T7N, T3a, T3d;
|
cannam@167
|
669 T7M = FNMS(KP707106781, T39, T38);
|
cannam@167
|
670 T7N = T35 - T32;
|
cannam@167
|
671 T7O = FMA(KP923879532, T7N, T7M);
|
cannam@167
|
672 T9d = FNMS(KP923879532, T7N, T7M);
|
cannam@167
|
673 T3a = FMA(KP707106781, T39, T38);
|
cannam@167
|
674 T3d = T3b + T3c;
|
cannam@167
|
675 T3e = FMA(KP923879532, T3d, T3a);
|
cannam@167
|
676 T6w = FNMS(KP923879532, T3d, T3a);
|
cannam@167
|
677 }
|
cannam@167
|
678 }
|
cannam@167
|
679 {
|
cannam@167
|
680 E T3L, Tdf, T3W, Tar, T42, Tdi, T4d, Tay, T3S, Tdg, T3X, Tau, T49, Tdj, T4e;
|
cannam@167
|
681 E TaB, Tdh, Tdk;
|
cannam@167
|
682 {
|
cannam@167
|
683 E T3J, T3K, Tap, T3U, T3V, Taq;
|
cannam@167
|
684 T3J = R1[WS(rs, 2)];
|
cannam@167
|
685 T3K = R1[WS(rs, 34)];
|
cannam@167
|
686 Tap = T3J + T3K;
|
cannam@167
|
687 T3U = R1[WS(rs, 18)];
|
cannam@167
|
688 T3V = R1[WS(rs, 50)];
|
cannam@167
|
689 Taq = T3U + T3V;
|
cannam@167
|
690 T3L = T3J - T3K;
|
cannam@167
|
691 Tdf = Tap - Taq;
|
cannam@167
|
692 T3W = T3U - T3V;
|
cannam@167
|
693 Tar = Tap + Taq;
|
cannam@167
|
694 }
|
cannam@167
|
695 {
|
cannam@167
|
696 E T40, T41, Taw, T4b, T4c, Tax;
|
cannam@167
|
697 T40 = R1[WS(rs, 62)];
|
cannam@167
|
698 T41 = R1[WS(rs, 30)];
|
cannam@167
|
699 Taw = T40 + T41;
|
cannam@167
|
700 T4b = R1[WS(rs, 14)];
|
cannam@167
|
701 T4c = R1[WS(rs, 46)];
|
cannam@167
|
702 Tax = T4b + T4c;
|
cannam@167
|
703 T42 = T40 - T41;
|
cannam@167
|
704 Tdi = Taw - Tax;
|
cannam@167
|
705 T4d = T4b - T4c;
|
cannam@167
|
706 Tay = Taw + Tax;
|
cannam@167
|
707 }
|
cannam@167
|
708 {
|
cannam@167
|
709 E T3O, Tas, T3R, Tat;
|
cannam@167
|
710 {
|
cannam@167
|
711 E T3M, T3N, T3P, T3Q;
|
cannam@167
|
712 T3M = R1[WS(rs, 10)];
|
cannam@167
|
713 T3N = R1[WS(rs, 42)];
|
cannam@167
|
714 T3O = T3M - T3N;
|
cannam@167
|
715 Tas = T3M + T3N;
|
cannam@167
|
716 T3P = R1[WS(rs, 58)];
|
cannam@167
|
717 T3Q = R1[WS(rs, 26)];
|
cannam@167
|
718 T3R = T3P - T3Q;
|
cannam@167
|
719 Tat = T3P + T3Q;
|
cannam@167
|
720 }
|
cannam@167
|
721 T3S = T3O + T3R;
|
cannam@167
|
722 Tdg = Tat - Tas;
|
cannam@167
|
723 T3X = T3O - T3R;
|
cannam@167
|
724 Tau = Tas + Tat;
|
cannam@167
|
725 }
|
cannam@167
|
726 {
|
cannam@167
|
727 E T45, Taz, T48, TaA;
|
cannam@167
|
728 {
|
cannam@167
|
729 E T43, T44, T46, T47;
|
cannam@167
|
730 T43 = R1[WS(rs, 6)];
|
cannam@167
|
731 T44 = R1[WS(rs, 38)];
|
cannam@167
|
732 T45 = T43 - T44;
|
cannam@167
|
733 Taz = T43 + T44;
|
cannam@167
|
734 T46 = R1[WS(rs, 54)];
|
cannam@167
|
735 T47 = R1[WS(rs, 22)];
|
cannam@167
|
736 T48 = T46 - T47;
|
cannam@167
|
737 TaA = T46 + T47;
|
cannam@167
|
738 }
|
cannam@167
|
739 T49 = T45 + T48;
|
cannam@167
|
740 Tdj = TaA - Taz;
|
cannam@167
|
741 T4e = T45 - T48;
|
cannam@167
|
742 TaB = Taz + TaA;
|
cannam@167
|
743 }
|
cannam@167
|
744 TbZ = Tar + Tau;
|
cannam@167
|
745 Tc0 = Tay + TaB;
|
cannam@167
|
746 {
|
cannam@167
|
747 E T3T, T3Y, Tdq, Tdr;
|
cannam@167
|
748 T3T = FMA(KP707106781, T3S, T3L);
|
cannam@167
|
749 T3Y = FMA(KP707106781, T3X, T3W);
|
cannam@167
|
750 T3Z = FNMS(KP198912367, T3Y, T3T);
|
cannam@167
|
751 T4s = FMA(KP198912367, T3T, T3Y);
|
cannam@167
|
752 Tdq = FNMS(KP414213562, Tdf, Tdg);
|
cannam@167
|
753 Tdr = FMA(KP414213562, Tdi, Tdj);
|
cannam@167
|
754 Tds = Tdq + Tdr;
|
cannam@167
|
755 TeF = Tdr - Tdq;
|
cannam@167
|
756 }
|
cannam@167
|
757 {
|
cannam@167
|
758 E T4a, T4f, T7Y, T7Z;
|
cannam@167
|
759 T4a = FMA(KP707106781, T49, T42);
|
cannam@167
|
760 T4f = FMA(KP707106781, T4e, T4d);
|
cannam@167
|
761 T4g = FMA(KP198912367, T4f, T4a);
|
cannam@167
|
762 T4t = FNMS(KP198912367, T4a, T4f);
|
cannam@167
|
763 T7Y = FNMS(KP707106781, T49, T42);
|
cannam@167
|
764 T7Z = FNMS(KP707106781, T4e, T4d);
|
cannam@167
|
765 T80 = FNMS(KP668178637, T7Z, T7Y);
|
cannam@167
|
766 T87 = FMA(KP668178637, T7Y, T7Z);
|
cannam@167
|
767 }
|
cannam@167
|
768 Tdh = FMA(KP414213562, Tdg, Tdf);
|
cannam@167
|
769 Tdk = FNMS(KP414213562, Tdj, Tdi);
|
cannam@167
|
770 Tdl = Tdh + Tdk;
|
cannam@167
|
771 TeI = Tdh - Tdk;
|
cannam@167
|
772 {
|
cannam@167
|
773 E T7V, T7W, Tav, TaC;
|
cannam@167
|
774 T7V = FNMS(KP707106781, T3S, T3L);
|
cannam@167
|
775 T7W = FNMS(KP707106781, T3X, T3W);
|
cannam@167
|
776 T7X = FMA(KP668178637, T7W, T7V);
|
cannam@167
|
777 T86 = FNMS(KP668178637, T7V, T7W);
|
cannam@167
|
778 Tav = Tar - Tau;
|
cannam@167
|
779 TaC = Tay - TaB;
|
cannam@167
|
780 TaD = Tav + TaC;
|
cannam@167
|
781 TaM = TaC - Tav;
|
cannam@167
|
782 }
|
cannam@167
|
783 }
|
cannam@167
|
784 {
|
cannam@167
|
785 E T50, TdA, T5b, TaY, T5h, TdD, T5s, Tb5, T57, TdB, T5c, Tb1, T5o, TdE, T5t;
|
cannam@167
|
786 E Tb8, TdC, TdF;
|
cannam@167
|
787 {
|
cannam@167
|
788 E T4Y, T4Z, TaW, T59, T5a, TaX;
|
cannam@167
|
789 T4Y = R1[WS(rs, 1)];
|
cannam@167
|
790 T4Z = R1[WS(rs, 33)];
|
cannam@167
|
791 TaW = T4Y + T4Z;
|
cannam@167
|
792 T59 = R1[WS(rs, 49)];
|
cannam@167
|
793 T5a = R1[WS(rs, 17)];
|
cannam@167
|
794 TaX = T5a + T59;
|
cannam@167
|
795 T50 = T4Y - T4Z;
|
cannam@167
|
796 TdA = TaW - TaX;
|
cannam@167
|
797 T5b = T59 - T5a;
|
cannam@167
|
798 TaY = TaW + TaX;
|
cannam@167
|
799 }
|
cannam@167
|
800 {
|
cannam@167
|
801 E T5f, T5g, Tb3, T5q, T5r, Tb4;
|
cannam@167
|
802 T5f = R1[WS(rs, 61)];
|
cannam@167
|
803 T5g = R1[WS(rs, 29)];
|
cannam@167
|
804 Tb3 = T5f + T5g;
|
cannam@167
|
805 T5q = R1[WS(rs, 45)];
|
cannam@167
|
806 T5r = R1[WS(rs, 13)];
|
cannam@167
|
807 Tb4 = T5r + T5q;
|
cannam@167
|
808 T5h = T5f - T5g;
|
cannam@167
|
809 TdD = Tb3 - Tb4;
|
cannam@167
|
810 T5s = T5q - T5r;
|
cannam@167
|
811 Tb5 = Tb3 + Tb4;
|
cannam@167
|
812 }
|
cannam@167
|
813 {
|
cannam@167
|
814 E T53, TaZ, T56, Tb0;
|
cannam@167
|
815 {
|
cannam@167
|
816 E T51, T52, T54, T55;
|
cannam@167
|
817 T51 = R1[WS(rs, 9)];
|
cannam@167
|
818 T52 = R1[WS(rs, 41)];
|
cannam@167
|
819 T53 = T51 - T52;
|
cannam@167
|
820 TaZ = T51 + T52;
|
cannam@167
|
821 T54 = R1[WS(rs, 57)];
|
cannam@167
|
822 T55 = R1[WS(rs, 25)];
|
cannam@167
|
823 T56 = T54 - T55;
|
cannam@167
|
824 Tb0 = T54 + T55;
|
cannam@167
|
825 }
|
cannam@167
|
826 T57 = T53 + T56;
|
cannam@167
|
827 TdB = Tb0 - TaZ;
|
cannam@167
|
828 T5c = T56 - T53;
|
cannam@167
|
829 Tb1 = TaZ + Tb0;
|
cannam@167
|
830 }
|
cannam@167
|
831 {
|
cannam@167
|
832 E T5k, Tb6, T5n, Tb7;
|
cannam@167
|
833 {
|
cannam@167
|
834 E T5i, T5j, T5l, T5m;
|
cannam@167
|
835 T5i = R1[WS(rs, 5)];
|
cannam@167
|
836 T5j = R1[WS(rs, 37)];
|
cannam@167
|
837 T5k = T5i - T5j;
|
cannam@167
|
838 Tb6 = T5i + T5j;
|
cannam@167
|
839 T5l = R1[WS(rs, 53)];
|
cannam@167
|
840 T5m = R1[WS(rs, 21)];
|
cannam@167
|
841 T5n = T5l - T5m;
|
cannam@167
|
842 Tb7 = T5l + T5m;
|
cannam@167
|
843 }
|
cannam@167
|
844 T5o = T5k + T5n;
|
cannam@167
|
845 TdE = Tb7 - Tb6;
|
cannam@167
|
846 T5t = T5n - T5k;
|
cannam@167
|
847 Tb8 = Tb6 + Tb7;
|
cannam@167
|
848 }
|
cannam@167
|
849 Tc6 = TaY + Tb1;
|
cannam@167
|
850 Tc7 = Tb5 + Tb8;
|
cannam@167
|
851 {
|
cannam@167
|
852 E T58, T5d, TdL, TdM;
|
cannam@167
|
853 T58 = FMA(KP707106781, T57, T50);
|
cannam@167
|
854 T5d = FMA(KP707106781, T5c, T5b);
|
cannam@167
|
855 T5e = FMA(KP198912367, T5d, T58);
|
cannam@167
|
856 T5H = FNMS(KP198912367, T58, T5d);
|
cannam@167
|
857 TdL = FNMS(KP414213562, TdA, TdB);
|
cannam@167
|
858 TdM = FMA(KP414213562, TdD, TdE);
|
cannam@167
|
859 TdN = TdL + TdM;
|
cannam@167
|
860 TeM = TdM - TdL;
|
cannam@167
|
861 }
|
cannam@167
|
862 {
|
cannam@167
|
863 E T5p, T5u, T8h, T8i;
|
cannam@167
|
864 T5p = FMA(KP707106781, T5o, T5h);
|
cannam@167
|
865 T5u = FMA(KP707106781, T5t, T5s);
|
cannam@167
|
866 T5v = FNMS(KP198912367, T5u, T5p);
|
cannam@167
|
867 T5I = FMA(KP198912367, T5p, T5u);
|
cannam@167
|
868 T8h = FNMS(KP707106781, T5o, T5h);
|
cannam@167
|
869 T8i = FNMS(KP707106781, T5t, T5s);
|
cannam@167
|
870 T8j = FMA(KP668178637, T8i, T8h);
|
cannam@167
|
871 T8q = FNMS(KP668178637, T8h, T8i);
|
cannam@167
|
872 }
|
cannam@167
|
873 TdC = FMA(KP414213562, TdB, TdA);
|
cannam@167
|
874 TdF = FNMS(KP414213562, TdE, TdD);
|
cannam@167
|
875 TdG = TdC + TdF;
|
cannam@167
|
876 TeP = TdF - TdC;
|
cannam@167
|
877 {
|
cannam@167
|
878 E T8e, T8f, Tb2, Tb9;
|
cannam@167
|
879 T8e = FNMS(KP707106781, T57, T50);
|
cannam@167
|
880 T8f = FNMS(KP707106781, T5c, T5b);
|
cannam@167
|
881 T8g = FNMS(KP668178637, T8f, T8e);
|
cannam@167
|
882 T8p = FMA(KP668178637, T8e, T8f);
|
cannam@167
|
883 Tb2 = TaY - Tb1;
|
cannam@167
|
884 Tb9 = Tb5 - Tb8;
|
cannam@167
|
885 Tba = Tb2 + Tb9;
|
cannam@167
|
886 Tbj = Tb9 - Tb2;
|
cannam@167
|
887 }
|
cannam@167
|
888 }
|
cannam@167
|
889 {
|
cannam@167
|
890 E T11, TbV, Tc9, Tcf, T22, Tcb, Tc2, Tce;
|
cannam@167
|
891 {
|
cannam@167
|
892 E Tv, T10, Tc5, Tc8;
|
cannam@167
|
893 Tv = Tf + Tu;
|
cannam@167
|
894 T10 = TK + TZ;
|
cannam@167
|
895 T11 = Tv + T10;
|
cannam@167
|
896 TbV = Tv - T10;
|
cannam@167
|
897 Tc5 = Tc3 + Tc4;
|
cannam@167
|
898 Tc8 = Tc6 + Tc7;
|
cannam@167
|
899 Tc9 = Tc5 - Tc8;
|
cannam@167
|
900 Tcf = Tc5 + Tc8;
|
cannam@167
|
901 }
|
cannam@167
|
902 {
|
cannam@167
|
903 E T1w, T21, TbY, Tc1;
|
cannam@167
|
904 T1w = T1g + T1v;
|
cannam@167
|
905 T21 = T1L + T20;
|
cannam@167
|
906 T22 = T1w + T21;
|
cannam@167
|
907 Tcb = T21 - T1w;
|
cannam@167
|
908 TbY = TbW + TbX;
|
cannam@167
|
909 Tc1 = TbZ + Tc0;
|
cannam@167
|
910 Tc2 = TbY - Tc1;
|
cannam@167
|
911 Tce = TbY + Tc1;
|
cannam@167
|
912 }
|
cannam@167
|
913 Cr[WS(csr, 32)] = T11 - T22;
|
cannam@167
|
914 Ci[WS(csi, 32)] = Tcf - Tce;
|
cannam@167
|
915 {
|
cannam@167
|
916 E Tca, Tcc, Tcd, Tcg;
|
cannam@167
|
917 Tca = Tc2 + Tc9;
|
cannam@167
|
918 Cr[WS(csr, 48)] = FNMS(KP707106781, Tca, TbV);
|
cannam@167
|
919 Cr[WS(csr, 16)] = FMA(KP707106781, Tca, TbV);
|
cannam@167
|
920 Tcc = Tc9 - Tc2;
|
cannam@167
|
921 Ci[WS(csi, 16)] = FMA(KP707106781, Tcc, Tcb);
|
cannam@167
|
922 Ci[WS(csi, 48)] = FMS(KP707106781, Tcc, Tcb);
|
cannam@167
|
923 Tcd = T11 + T22;
|
cannam@167
|
924 Tcg = Tce + Tcf;
|
cannam@167
|
925 Cr[WS(csr, 64)] = Tcd - Tcg;
|
cannam@167
|
926 Cr[0] = Tcd + Tcg;
|
cannam@167
|
927 }
|
cannam@167
|
928 }
|
cannam@167
|
929 {
|
cannam@167
|
930 E Tch, Tct, Tck, Tcu, Tco, Tcy, Tcr, Tcz, Tci, Tcj;
|
cannam@167
|
931 Tch = Tf - Tu;
|
cannam@167
|
932 Tct = TZ - TK;
|
cannam@167
|
933 Tci = T1g - T1v;
|
cannam@167
|
934 Tcj = T1L - T20;
|
cannam@167
|
935 Tck = Tci + Tcj;
|
cannam@167
|
936 Tcu = Tcj - Tci;
|
cannam@167
|
937 {
|
cannam@167
|
938 E Tcm, Tcn, Tcp, Tcq;
|
cannam@167
|
939 Tcm = TbW - TbX;
|
cannam@167
|
940 Tcn = Tc0 - TbZ;
|
cannam@167
|
941 Tco = FMA(KP414213562, Tcn, Tcm);
|
cannam@167
|
942 Tcy = FNMS(KP414213562, Tcm, Tcn);
|
cannam@167
|
943 Tcp = Tc3 - Tc4;
|
cannam@167
|
944 Tcq = Tc7 - Tc6;
|
cannam@167
|
945 Tcr = FNMS(KP414213562, Tcq, Tcp);
|
cannam@167
|
946 Tcz = FMA(KP414213562, Tcp, Tcq);
|
cannam@167
|
947 }
|
cannam@167
|
948 {
|
cannam@167
|
949 E Tcl, Tcs, Tcx, TcA;
|
cannam@167
|
950 Tcl = FMA(KP707106781, Tck, Tch);
|
cannam@167
|
951 Tcs = Tco + Tcr;
|
cannam@167
|
952 Cr[WS(csr, 56)] = FNMS(KP923879532, Tcs, Tcl);
|
cannam@167
|
953 Cr[WS(csr, 8)] = FMA(KP923879532, Tcs, Tcl);
|
cannam@167
|
954 Tcx = FMA(KP707106781, Tcu, Tct);
|
cannam@167
|
955 TcA = Tcy + Tcz;
|
cannam@167
|
956 Ci[WS(csi, 8)] = FMA(KP923879532, TcA, Tcx);
|
cannam@167
|
957 Ci[WS(csi, 56)] = FMS(KP923879532, TcA, Tcx);
|
cannam@167
|
958 }
|
cannam@167
|
959 {
|
cannam@167
|
960 E Tcv, Tcw, TcB, TcC;
|
cannam@167
|
961 Tcv = FNMS(KP707106781, Tcu, Tct);
|
cannam@167
|
962 Tcw = Tcr - Tco;
|
cannam@167
|
963 Ci[WS(csi, 24)] = FMS(KP923879532, Tcw, Tcv);
|
cannam@167
|
964 Ci[WS(csi, 40)] = FMA(KP923879532, Tcw, Tcv);
|
cannam@167
|
965 TcB = FNMS(KP707106781, Tck, Tch);
|
cannam@167
|
966 TcC = Tcz - Tcy;
|
cannam@167
|
967 Cr[WS(csr, 40)] = FNMS(KP923879532, TcC, TcB);
|
cannam@167
|
968 Cr[WS(csr, 24)] = FMA(KP923879532, TcC, TcB);
|
cannam@167
|
969 }
|
cannam@167
|
970 }
|
cannam@167
|
971 {
|
cannam@167
|
972 E Ta9, TbB, Tbp, TbL, Tag, TbM, TbJ, TbR, TaO, Tbw, Tbs, TbC, TbG, TbQ, Tbl;
|
cannam@167
|
973 E Tbx, Ta8, Tbo;
|
cannam@167
|
974 Ta8 = Ta6 + Ta7;
|
cannam@167
|
975 Ta9 = FMA(KP707106781, Ta8, Ta5);
|
cannam@167
|
976 TbB = FNMS(KP707106781, Ta8, Ta5);
|
cannam@167
|
977 Tbo = Ta7 - Ta6;
|
cannam@167
|
978 Tbp = FMA(KP707106781, Tbo, Tbn);
|
cannam@167
|
979 TbL = FNMS(KP707106781, Tbo, Tbn);
|
cannam@167
|
980 {
|
cannam@167
|
981 E Tac, Taf, TbH, TbI;
|
cannam@167
|
982 Tac = FMA(KP414213562, Tab, Taa);
|
cannam@167
|
983 Taf = FNMS(KP414213562, Tae, Tad);
|
cannam@167
|
984 Tag = Tac + Taf;
|
cannam@167
|
985 TbM = Taf - Tac;
|
cannam@167
|
986 TbH = FNMS(KP707106781, Tba, TaV);
|
cannam@167
|
987 TbI = FNMS(KP707106781, Tbj, Tbi);
|
cannam@167
|
988 TbJ = FMA(KP668178637, TbI, TbH);
|
cannam@167
|
989 TbR = FNMS(KP668178637, TbH, TbI);
|
cannam@167
|
990 }
|
cannam@167
|
991 {
|
cannam@167
|
992 E TaE, TaN, Tbq, Tbr;
|
cannam@167
|
993 TaE = FMA(KP707106781, TaD, Tao);
|
cannam@167
|
994 TaN = FMA(KP707106781, TaM, TaL);
|
cannam@167
|
995 TaO = FMA(KP198912367, TaN, TaE);
|
cannam@167
|
996 Tbw = FNMS(KP198912367, TaE, TaN);
|
cannam@167
|
997 Tbq = FNMS(KP414213562, Taa, Tab);
|
cannam@167
|
998 Tbr = FMA(KP414213562, Tad, Tae);
|
cannam@167
|
999 Tbs = Tbq + Tbr;
|
cannam@167
|
1000 TbC = Tbr - Tbq;
|
cannam@167
|
1001 }
|
cannam@167
|
1002 {
|
cannam@167
|
1003 E TbE, TbF, Tbb, Tbk;
|
cannam@167
|
1004 TbE = FNMS(KP707106781, TaD, Tao);
|
cannam@167
|
1005 TbF = FNMS(KP707106781, TaM, TaL);
|
cannam@167
|
1006 TbG = FNMS(KP668178637, TbF, TbE);
|
cannam@167
|
1007 TbQ = FMA(KP668178637, TbE, TbF);
|
cannam@167
|
1008 Tbb = FMA(KP707106781, Tba, TaV);
|
cannam@167
|
1009 Tbk = FMA(KP707106781, Tbj, Tbi);
|
cannam@167
|
1010 Tbl = FNMS(KP198912367, Tbk, Tbb);
|
cannam@167
|
1011 Tbx = FMA(KP198912367, Tbb, Tbk);
|
cannam@167
|
1012 }
|
cannam@167
|
1013 {
|
cannam@167
|
1014 E Tah, Tbm, Tbv, Tby;
|
cannam@167
|
1015 Tah = FMA(KP923879532, Tag, Ta9);
|
cannam@167
|
1016 Tbm = TaO + Tbl;
|
cannam@167
|
1017 Cr[WS(csr, 60)] = FNMS(KP980785280, Tbm, Tah);
|
cannam@167
|
1018 Cr[WS(csr, 4)] = FMA(KP980785280, Tbm, Tah);
|
cannam@167
|
1019 Tbv = FMA(KP923879532, Tbs, Tbp);
|
cannam@167
|
1020 Tby = Tbw + Tbx;
|
cannam@167
|
1021 Ci[WS(csi, 4)] = FMA(KP980785280, Tby, Tbv);
|
cannam@167
|
1022 Ci[WS(csi, 60)] = FMS(KP980785280, Tby, Tbv);
|
cannam@167
|
1023 }
|
cannam@167
|
1024 {
|
cannam@167
|
1025 E Tbt, Tbu, Tbz, TbA;
|
cannam@167
|
1026 Tbt = FNMS(KP923879532, Tbs, Tbp);
|
cannam@167
|
1027 Tbu = Tbl - TaO;
|
cannam@167
|
1028 Ci[WS(csi, 28)] = FMS(KP980785280, Tbu, Tbt);
|
cannam@167
|
1029 Ci[WS(csi, 36)] = FMA(KP980785280, Tbu, Tbt);
|
cannam@167
|
1030 Tbz = FNMS(KP923879532, Tag, Ta9);
|
cannam@167
|
1031 TbA = Tbx - Tbw;
|
cannam@167
|
1032 Cr[WS(csr, 36)] = FNMS(KP980785280, TbA, Tbz);
|
cannam@167
|
1033 Cr[WS(csr, 28)] = FMA(KP980785280, TbA, Tbz);
|
cannam@167
|
1034 }
|
cannam@167
|
1035 {
|
cannam@167
|
1036 E TbD, TbK, TbP, TbS;
|
cannam@167
|
1037 TbD = FMA(KP923879532, TbC, TbB);
|
cannam@167
|
1038 TbK = TbG + TbJ;
|
cannam@167
|
1039 Cr[WS(csr, 52)] = FNMS(KP831469612, TbK, TbD);
|
cannam@167
|
1040 Cr[WS(csr, 12)] = FMA(KP831469612, TbK, TbD);
|
cannam@167
|
1041 TbP = FNMS(KP923879532, TbM, TbL);
|
cannam@167
|
1042 TbS = TbQ + TbR;
|
cannam@167
|
1043 Ci[WS(csi, 12)] = -(FMA(KP831469612, TbS, TbP));
|
cannam@167
|
1044 Ci[WS(csi, 52)] = FNMS(KP831469612, TbS, TbP);
|
cannam@167
|
1045 }
|
cannam@167
|
1046 {
|
cannam@167
|
1047 E TbN, TbO, TbT, TbU;
|
cannam@167
|
1048 TbN = FMA(KP923879532, TbM, TbL);
|
cannam@167
|
1049 TbO = TbJ - TbG;
|
cannam@167
|
1050 Ci[WS(csi, 20)] = FMA(KP831469612, TbO, TbN);
|
cannam@167
|
1051 Ci[WS(csi, 44)] = FMS(KP831469612, TbO, TbN);
|
cannam@167
|
1052 TbT = FNMS(KP923879532, TbC, TbB);
|
cannam@167
|
1053 TbU = TbQ - TbR;
|
cannam@167
|
1054 Cr[WS(csr, 44)] = FNMS(KP831469612, TbU, TbT);
|
cannam@167
|
1055 Cr[WS(csr, 20)] = FMA(KP831469612, TbU, TbT);
|
cannam@167
|
1056 }
|
cannam@167
|
1057 }
|
cannam@167
|
1058 {
|
cannam@167
|
1059 E Tev, Tf7, Tfc, Tfm, Tff, Tfn, TeC, Tfi, TeK, Tf2, TeV, Tfh, TeY, Tf8, TeR;
|
cannam@167
|
1060 E Tf3;
|
cannam@167
|
1061 {
|
cannam@167
|
1062 E Tet, Teu, Tfa, Tfb;
|
cannam@167
|
1063 Tet = FNMS(KP707106781, TcG, TcD);
|
cannam@167
|
1064 Teu = TdV - TdU;
|
cannam@167
|
1065 Tev = FNMS(KP923879532, Teu, Tet);
|
cannam@167
|
1066 Tf7 = FMA(KP923879532, Teu, Tet);
|
cannam@167
|
1067 Tfa = FMA(KP923879532, TeF, TeE);
|
cannam@167
|
1068 Tfb = FMA(KP923879532, TeI, TeH);
|
cannam@167
|
1069 Tfc = FNMS(KP303346683, Tfb, Tfa);
|
cannam@167
|
1070 Tfm = FMA(KP303346683, Tfa, Tfb);
|
cannam@167
|
1071 }
|
cannam@167
|
1072 {
|
cannam@167
|
1073 E Tfd, Tfe, Tey, TeB;
|
cannam@167
|
1074 Tfd = FMA(KP923879532, TeM, TeL);
|
cannam@167
|
1075 Tfe = FNMS(KP923879532, TeP, TeO);
|
cannam@167
|
1076 Tff = FMA(KP303346683, Tfe, Tfd);
|
cannam@167
|
1077 Tfn = FNMS(KP303346683, Tfd, Tfe);
|
cannam@167
|
1078 Tey = FMA(KP668178637, Tex, Tew);
|
cannam@167
|
1079 TeB = FNMS(KP668178637, TeA, Tez);
|
cannam@167
|
1080 TeC = Tey - TeB;
|
cannam@167
|
1081 Tfi = Tey + TeB;
|
cannam@167
|
1082 }
|
cannam@167
|
1083 {
|
cannam@167
|
1084 E TeG, TeJ, TeT, TeU;
|
cannam@167
|
1085 TeG = FNMS(KP923879532, TeF, TeE);
|
cannam@167
|
1086 TeJ = FNMS(KP923879532, TeI, TeH);
|
cannam@167
|
1087 TeK = FMA(KP534511135, TeJ, TeG);
|
cannam@167
|
1088 Tf2 = FNMS(KP534511135, TeG, TeJ);
|
cannam@167
|
1089 TeT = FNMS(KP707106781, TdS, TdR);
|
cannam@167
|
1090 TeU = TcN - TcK;
|
cannam@167
|
1091 TeV = FMA(KP923879532, TeU, TeT);
|
cannam@167
|
1092 Tfh = FNMS(KP923879532, TeU, TeT);
|
cannam@167
|
1093 }
|
cannam@167
|
1094 {
|
cannam@167
|
1095 E TeW, TeX, TeN, TeQ;
|
cannam@167
|
1096 TeW = FMA(KP668178637, Tez, TeA);
|
cannam@167
|
1097 TeX = FNMS(KP668178637, Tew, Tex);
|
cannam@167
|
1098 TeY = TeW - TeX;
|
cannam@167
|
1099 Tf8 = TeX + TeW;
|
cannam@167
|
1100 TeN = FNMS(KP923879532, TeM, TeL);
|
cannam@167
|
1101 TeQ = FMA(KP923879532, TeP, TeO);
|
cannam@167
|
1102 TeR = FNMS(KP534511135, TeQ, TeN);
|
cannam@167
|
1103 Tf3 = FMA(KP534511135, TeN, TeQ);
|
cannam@167
|
1104 }
|
cannam@167
|
1105 {
|
cannam@167
|
1106 E TeD, TeS, Tf1, Tf4;
|
cannam@167
|
1107 TeD = FMA(KP831469612, TeC, Tev);
|
cannam@167
|
1108 TeS = TeK + TeR;
|
cannam@167
|
1109 Cr[WS(csr, 54)] = FNMS(KP881921264, TeS, TeD);
|
cannam@167
|
1110 Cr[WS(csr, 10)] = FMA(KP881921264, TeS, TeD);
|
cannam@167
|
1111 Tf1 = FMA(KP831469612, TeY, TeV);
|
cannam@167
|
1112 Tf4 = Tf2 + Tf3;
|
cannam@167
|
1113 Ci[WS(csi, 10)] = FMA(KP881921264, Tf4, Tf1);
|
cannam@167
|
1114 Ci[WS(csi, 54)] = FMS(KP881921264, Tf4, Tf1);
|
cannam@167
|
1115 }
|
cannam@167
|
1116 {
|
cannam@167
|
1117 E TeZ, Tf0, Tf5, Tf6;
|
cannam@167
|
1118 TeZ = FNMS(KP831469612, TeY, TeV);
|
cannam@167
|
1119 Tf0 = TeR - TeK;
|
cannam@167
|
1120 Ci[WS(csi, 22)] = FMS(KP881921264, Tf0, TeZ);
|
cannam@167
|
1121 Ci[WS(csi, 42)] = FMA(KP881921264, Tf0, TeZ);
|
cannam@167
|
1122 Tf5 = FNMS(KP831469612, TeC, Tev);
|
cannam@167
|
1123 Tf6 = Tf3 - Tf2;
|
cannam@167
|
1124 Cr[WS(csr, 42)] = FNMS(KP881921264, Tf6, Tf5);
|
cannam@167
|
1125 Cr[WS(csr, 22)] = FMA(KP881921264, Tf6, Tf5);
|
cannam@167
|
1126 }
|
cannam@167
|
1127 {
|
cannam@167
|
1128 E Tf9, Tfg, Tfl, Tfo;
|
cannam@167
|
1129 Tf9 = FMA(KP831469612, Tf8, Tf7);
|
cannam@167
|
1130 Tfg = Tfc + Tff;
|
cannam@167
|
1131 Cr[WS(csr, 58)] = FNMS(KP956940335, Tfg, Tf9);
|
cannam@167
|
1132 Cr[WS(csr, 6)] = FMA(KP956940335, Tfg, Tf9);
|
cannam@167
|
1133 Tfl = FMA(KP831469612, Tfi, Tfh);
|
cannam@167
|
1134 Tfo = Tfm + Tfn;
|
cannam@167
|
1135 Ci[WS(csi, 6)] = -(FMA(KP956940335, Tfo, Tfl));
|
cannam@167
|
1136 Ci[WS(csi, 58)] = FNMS(KP956940335, Tfo, Tfl);
|
cannam@167
|
1137 }
|
cannam@167
|
1138 {
|
cannam@167
|
1139 E Tfj, Tfk, Tfp, Tfq;
|
cannam@167
|
1140 Tfj = FNMS(KP831469612, Tfi, Tfh);
|
cannam@167
|
1141 Tfk = Tff - Tfc;
|
cannam@167
|
1142 Ci[WS(csi, 26)] = FMA(KP956940335, Tfk, Tfj);
|
cannam@167
|
1143 Ci[WS(csi, 38)] = FMS(KP956940335, Tfk, Tfj);
|
cannam@167
|
1144 Tfp = FNMS(KP831469612, Tf8, Tf7);
|
cannam@167
|
1145 Tfq = Tfm - Tfn;
|
cannam@167
|
1146 Cr[WS(csr, 38)] = FNMS(KP956940335, Tfq, Tfp);
|
cannam@167
|
1147 Cr[WS(csr, 26)] = FMA(KP956940335, Tfq, Tfp);
|
cannam@167
|
1148 }
|
cannam@167
|
1149 }
|
cannam@167
|
1150 {
|
cannam@167
|
1151 E TcP, Te9, Tee, Teo, Teh, Tep, Td8, Tek, Tdu, Te4, TdX, Tej, Te0, Tea, TdP;
|
cannam@167
|
1152 E Te5;
|
cannam@167
|
1153 {
|
cannam@167
|
1154 E TcH, TcO, Tec, Ted;
|
cannam@167
|
1155 TcH = FMA(KP707106781, TcG, TcD);
|
cannam@167
|
1156 TcO = TcK + TcN;
|
cannam@167
|
1157 TcP = FMA(KP923879532, TcO, TcH);
|
cannam@167
|
1158 Te9 = FNMS(KP923879532, TcO, TcH);
|
cannam@167
|
1159 Tec = FNMS(KP923879532, Tdl, Tde);
|
cannam@167
|
1160 Ted = FNMS(KP923879532, Tds, Tdp);
|
cannam@167
|
1161 Tee = FNMS(KP820678790, Ted, Tec);
|
cannam@167
|
1162 Teo = FMA(KP820678790, Tec, Ted);
|
cannam@167
|
1163 }
|
cannam@167
|
1164 {
|
cannam@167
|
1165 E Tef, Teg, TcY, Td7;
|
cannam@167
|
1166 Tef = FNMS(KP923879532, TdG, Tdz);
|
cannam@167
|
1167 Teg = FNMS(KP923879532, TdN, TdK);
|
cannam@167
|
1168 Teh = FMA(KP820678790, Teg, Tef);
|
cannam@167
|
1169 Tep = FNMS(KP820678790, Tef, Teg);
|
cannam@167
|
1170 TcY = FMA(KP198912367, TcX, TcU);
|
cannam@167
|
1171 Td7 = FNMS(KP198912367, Td6, Td3);
|
cannam@167
|
1172 Td8 = TcY + Td7;
|
cannam@167
|
1173 Tek = Td7 - TcY;
|
cannam@167
|
1174 }
|
cannam@167
|
1175 {
|
cannam@167
|
1176 E Tdm, Tdt, TdT, TdW;
|
cannam@167
|
1177 Tdm = FMA(KP923879532, Tdl, Tde);
|
cannam@167
|
1178 Tdt = FMA(KP923879532, Tds, Tdp);
|
cannam@167
|
1179 Tdu = FMA(KP098491403, Tdt, Tdm);
|
cannam@167
|
1180 Te4 = FNMS(KP098491403, Tdm, Tdt);
|
cannam@167
|
1181 TdT = FMA(KP707106781, TdS, TdR);
|
cannam@167
|
1182 TdW = TdU + TdV;
|
cannam@167
|
1183 TdX = FMA(KP923879532, TdW, TdT);
|
cannam@167
|
1184 Tej = FNMS(KP923879532, TdW, TdT);
|
cannam@167
|
1185 }
|
cannam@167
|
1186 {
|
cannam@167
|
1187 E TdY, TdZ, TdH, TdO;
|
cannam@167
|
1188 TdY = FNMS(KP198912367, TcU, TcX);
|
cannam@167
|
1189 TdZ = FMA(KP198912367, Td3, Td6);
|
cannam@167
|
1190 Te0 = TdY + TdZ;
|
cannam@167
|
1191 Tea = TdZ - TdY;
|
cannam@167
|
1192 TdH = FMA(KP923879532, TdG, Tdz);
|
cannam@167
|
1193 TdO = FMA(KP923879532, TdN, TdK);
|
cannam@167
|
1194 TdP = FNMS(KP098491403, TdO, TdH);
|
cannam@167
|
1195 Te5 = FMA(KP098491403, TdH, TdO);
|
cannam@167
|
1196 }
|
cannam@167
|
1197 {
|
cannam@167
|
1198 E Td9, TdQ, Te3, Te6;
|
cannam@167
|
1199 Td9 = FMA(KP980785280, Td8, TcP);
|
cannam@167
|
1200 TdQ = Tdu + TdP;
|
cannam@167
|
1201 Cr[WS(csr, 62)] = FNMS(KP995184726, TdQ, Td9);
|
cannam@167
|
1202 Cr[WS(csr, 2)] = FMA(KP995184726, TdQ, Td9);
|
cannam@167
|
1203 Te3 = FMA(KP980785280, Te0, TdX);
|
cannam@167
|
1204 Te6 = Te4 + Te5;
|
cannam@167
|
1205 Ci[WS(csi, 2)] = FMA(KP995184726, Te6, Te3);
|
cannam@167
|
1206 Ci[WS(csi, 62)] = FMS(KP995184726, Te6, Te3);
|
cannam@167
|
1207 }
|
cannam@167
|
1208 {
|
cannam@167
|
1209 E Te1, Te2, Te7, Te8;
|
cannam@167
|
1210 Te1 = FNMS(KP980785280, Te0, TdX);
|
cannam@167
|
1211 Te2 = TdP - Tdu;
|
cannam@167
|
1212 Ci[WS(csi, 30)] = FMS(KP995184726, Te2, Te1);
|
cannam@167
|
1213 Ci[WS(csi, 34)] = FMA(KP995184726, Te2, Te1);
|
cannam@167
|
1214 Te7 = FNMS(KP980785280, Td8, TcP);
|
cannam@167
|
1215 Te8 = Te5 - Te4;
|
cannam@167
|
1216 Cr[WS(csr, 34)] = FNMS(KP995184726, Te8, Te7);
|
cannam@167
|
1217 Cr[WS(csr, 30)] = FMA(KP995184726, Te8, Te7);
|
cannam@167
|
1218 }
|
cannam@167
|
1219 {
|
cannam@167
|
1220 E Teb, Tei, Ten, Teq;
|
cannam@167
|
1221 Teb = FMA(KP980785280, Tea, Te9);
|
cannam@167
|
1222 Tei = Tee + Teh;
|
cannam@167
|
1223 Cr[WS(csr, 50)] = FNMS(KP773010453, Tei, Teb);
|
cannam@167
|
1224 Cr[WS(csr, 14)] = FMA(KP773010453, Tei, Teb);
|
cannam@167
|
1225 Ten = FNMS(KP980785280, Tek, Tej);
|
cannam@167
|
1226 Teq = Teo + Tep;
|
cannam@167
|
1227 Ci[WS(csi, 14)] = -(FMA(KP773010453, Teq, Ten));
|
cannam@167
|
1228 Ci[WS(csi, 50)] = FNMS(KP773010453, Teq, Ten);
|
cannam@167
|
1229 }
|
cannam@167
|
1230 {
|
cannam@167
|
1231 E Tel, Tem, Ter, Tes;
|
cannam@167
|
1232 Tel = FMA(KP980785280, Tek, Tej);
|
cannam@167
|
1233 Tem = Teh - Tee;
|
cannam@167
|
1234 Ci[WS(csi, 18)] = FMA(KP773010453, Tem, Tel);
|
cannam@167
|
1235 Ci[WS(csi, 46)] = FMS(KP773010453, Tem, Tel);
|
cannam@167
|
1236 Ter = FNMS(KP980785280, Tea, Te9);
|
cannam@167
|
1237 Tes = Teo - Tep;
|
cannam@167
|
1238 Cr[WS(csr, 46)] = FNMS(KP773010453, Tes, Ter);
|
cannam@167
|
1239 Cr[WS(csr, 18)] = FMA(KP773010453, Tes, Ter);
|
cannam@167
|
1240 }
|
cannam@167
|
1241 }
|
cannam@167
|
1242 {
|
cannam@167
|
1243 E T6v, T77, T6C, T7i, T6Y, T78, T6V, T7h, T6R, T7n, T72, T7f, T6K, T7m, T73;
|
cannam@167
|
1244 E T7c;
|
cannam@167
|
1245 {
|
cannam@167
|
1246 E T6t, T6u, T6T, T6U;
|
cannam@167
|
1247 T6t = FNMS(KP923879532, T2e, T27);
|
cannam@167
|
1248 T6u = T5U - T5V;
|
cannam@167
|
1249 T6v = FNMS(KP980785280, T6u, T6t);
|
cannam@167
|
1250 T77 = FMA(KP980785280, T6u, T6t);
|
cannam@167
|
1251 {
|
cannam@167
|
1252 E T6y, T6B, T6W, T6X;
|
cannam@167
|
1253 T6y = FMA(KP820678790, T6x, T6w);
|
cannam@167
|
1254 T6B = FNMS(KP820678790, T6A, T6z);
|
cannam@167
|
1255 T6C = T6y - T6B;
|
cannam@167
|
1256 T7i = T6B + T6y;
|
cannam@167
|
1257 T6W = FNMS(KP820678790, T6w, T6x);
|
cannam@167
|
1258 T6X = FMA(KP820678790, T6z, T6A);
|
cannam@167
|
1259 T6Y = T6W - T6X;
|
cannam@167
|
1260 T78 = T6X + T6W;
|
cannam@167
|
1261 }
|
cannam@167
|
1262 T6T = FNMS(KP923879532, T5S, T5P);
|
cannam@167
|
1263 T6U = T2x - T2o;
|
cannam@167
|
1264 T6V = FNMS(KP980785280, T6U, T6T);
|
cannam@167
|
1265 T7h = FMA(KP980785280, T6U, T6T);
|
cannam@167
|
1266 {
|
cannam@167
|
1267 E T6N, T7d, T6Q, T7e, T6M, T6P;
|
cannam@167
|
1268 T6M = T5I - T5H;
|
cannam@167
|
1269 T6N = FNMS(KP980785280, T6M, T6L);
|
cannam@167
|
1270 T7d = FMA(KP980785280, T6M, T6L);
|
cannam@167
|
1271 T6P = T5v - T5e;
|
cannam@167
|
1272 T6Q = FMA(KP980785280, T6P, T6O);
|
cannam@167
|
1273 T7e = FNMS(KP980785280, T6P, T6O);
|
cannam@167
|
1274 T6R = FNMS(KP472964775, T6Q, T6N);
|
cannam@167
|
1275 T7n = FNMS(KP357805721, T7d, T7e);
|
cannam@167
|
1276 T72 = FMA(KP472964775, T6N, T6Q);
|
cannam@167
|
1277 T7f = FMA(KP357805721, T7e, T7d);
|
cannam@167
|
1278 }
|
cannam@167
|
1279 {
|
cannam@167
|
1280 E T6G, T7a, T6J, T7b, T6F, T6I;
|
cannam@167
|
1281 T6F = T4s - T4t;
|
cannam@167
|
1282 T6G = FNMS(KP980785280, T6F, T6E);
|
cannam@167
|
1283 T7a = FMA(KP980785280, T6F, T6E);
|
cannam@167
|
1284 T6I = T4g - T3Z;
|
cannam@167
|
1285 T6J = FNMS(KP980785280, T6I, T6H);
|
cannam@167
|
1286 T7b = FMA(KP980785280, T6I, T6H);
|
cannam@167
|
1287 T6K = FNMS(KP472964775, T6J, T6G);
|
cannam@167
|
1288 T7m = FNMS(KP357805721, T7a, T7b);
|
cannam@167
|
1289 T73 = FMA(KP472964775, T6G, T6J);
|
cannam@167
|
1290 T7c = FMA(KP357805721, T7b, T7a);
|
cannam@167
|
1291 }
|
cannam@167
|
1292 }
|
cannam@167
|
1293 {
|
cannam@167
|
1294 E T6D, T6S, T71, T74;
|
cannam@167
|
1295 T6D = FMA(KP773010453, T6C, T6v);
|
cannam@167
|
1296 T6S = T6K + T6R;
|
cannam@167
|
1297 Cr[WS(csr, 55)] = FNMS(KP903989293, T6S, T6D);
|
cannam@167
|
1298 Cr[WS(csr, 9)] = FMA(KP903989293, T6S, T6D);
|
cannam@167
|
1299 T71 = FNMS(KP773010453, T6Y, T6V);
|
cannam@167
|
1300 T74 = T72 - T73;
|
cannam@167
|
1301 Ci[WS(csi, 9)] = FMS(KP903989293, T74, T71);
|
cannam@167
|
1302 Ci[WS(csi, 55)] = FMA(KP903989293, T74, T71);
|
cannam@167
|
1303 }
|
cannam@167
|
1304 {
|
cannam@167
|
1305 E T6Z, T70, T75, T76;
|
cannam@167
|
1306 T6Z = FMA(KP773010453, T6Y, T6V);
|
cannam@167
|
1307 T70 = T6R - T6K;
|
cannam@167
|
1308 Ci[WS(csi, 23)] = FMA(KP903989293, T70, T6Z);
|
cannam@167
|
1309 Ci[WS(csi, 41)] = FMS(KP903989293, T70, T6Z);
|
cannam@167
|
1310 T75 = FNMS(KP773010453, T6C, T6v);
|
cannam@167
|
1311 T76 = T73 + T72;
|
cannam@167
|
1312 Cr[WS(csr, 41)] = FNMS(KP903989293, T76, T75);
|
cannam@167
|
1313 Cr[WS(csr, 23)] = FMA(KP903989293, T76, T75);
|
cannam@167
|
1314 }
|
cannam@167
|
1315 {
|
cannam@167
|
1316 E T79, T7g, T7l, T7o;
|
cannam@167
|
1317 T79 = FMA(KP773010453, T78, T77);
|
cannam@167
|
1318 T7g = T7c + T7f;
|
cannam@167
|
1319 Cr[WS(csr, 57)] = FNMS(KP941544065, T7g, T79);
|
cannam@167
|
1320 Cr[WS(csr, 7)] = FMA(KP941544065, T7g, T79);
|
cannam@167
|
1321 T7l = FMA(KP773010453, T7i, T7h);
|
cannam@167
|
1322 T7o = T7m - T7n;
|
cannam@167
|
1323 Ci[WS(csi, 7)] = FMA(KP941544065, T7o, T7l);
|
cannam@167
|
1324 Ci[WS(csi, 57)] = FMS(KP941544065, T7o, T7l);
|
cannam@167
|
1325 }
|
cannam@167
|
1326 {
|
cannam@167
|
1327 E T7j, T7k, T7p, T7q;
|
cannam@167
|
1328 T7j = FNMS(KP773010453, T7i, T7h);
|
cannam@167
|
1329 T7k = T7f - T7c;
|
cannam@167
|
1330 Ci[WS(csi, 25)] = FMS(KP941544065, T7k, T7j);
|
cannam@167
|
1331 Ci[WS(csi, 39)] = FMA(KP941544065, T7k, T7j);
|
cannam@167
|
1332 T7p = FNMS(KP773010453, T78, T77);
|
cannam@167
|
1333 T7q = T7m + T7n;
|
cannam@167
|
1334 Cr[WS(csr, 39)] = FMA(KP941544065, T7q, T7p);
|
cannam@167
|
1335 Cr[WS(csr, 25)] = FNMS(KP941544065, T7q, T7p);
|
cannam@167
|
1336 }
|
cannam@167
|
1337 }
|
cannam@167
|
1338 {
|
cannam@167
|
1339 E T99, T9L, T9g, T9W, T9C, T9M, T9z, T9V, T9v, Ta0, T9H, T9T, T9o, Ta1, T9G;
|
cannam@167
|
1340 E T9Q;
|
cannam@167
|
1341 {
|
cannam@167
|
1342 E T97, T98, T9x, T9y;
|
cannam@167
|
1343 T97 = FNMS(KP923879532, T7s, T7r);
|
cannam@167
|
1344 T98 = T8z - T8y;
|
cannam@167
|
1345 T99 = FNMS(KP831469612, T98, T97);
|
cannam@167
|
1346 T9L = FMA(KP831469612, T98, T97);
|
cannam@167
|
1347 {
|
cannam@167
|
1348 E T9c, T9f, T9A, T9B;
|
cannam@167
|
1349 T9c = FMA(KP534511135, T9b, T9a);
|
cannam@167
|
1350 T9f = FNMS(KP534511135, T9e, T9d);
|
cannam@167
|
1351 T9g = T9c - T9f;
|
cannam@167
|
1352 T9W = T9c + T9f;
|
cannam@167
|
1353 T9A = FMA(KP534511135, T9d, T9e);
|
cannam@167
|
1354 T9B = FNMS(KP534511135, T9a, T9b);
|
cannam@167
|
1355 T9C = T9A - T9B;
|
cannam@167
|
1356 T9M = T9B + T9A;
|
cannam@167
|
1357 }
|
cannam@167
|
1358 T9x = FNMS(KP923879532, T8w, T8v);
|
cannam@167
|
1359 T9y = T7z - T7w;
|
cannam@167
|
1360 T9z = FMA(KP831469612, T9y, T9x);
|
cannam@167
|
1361 T9V = FNMS(KP831469612, T9y, T9x);
|
cannam@167
|
1362 {
|
cannam@167
|
1363 E T9r, T9R, T9u, T9S, T9q, T9t;
|
cannam@167
|
1364 T9q = T8p - T8q;
|
cannam@167
|
1365 T9r = FNMS(KP831469612, T9q, T9p);
|
cannam@167
|
1366 T9R = FMA(KP831469612, T9q, T9p);
|
cannam@167
|
1367 T9t = T8j - T8g;
|
cannam@167
|
1368 T9u = FNMS(KP831469612, T9t, T9s);
|
cannam@167
|
1369 T9S = FMA(KP831469612, T9t, T9s);
|
cannam@167
|
1370 T9v = FMA(KP599376933, T9u, T9r);
|
cannam@167
|
1371 Ta0 = FMA(KP250486960, T9R, T9S);
|
cannam@167
|
1372 T9H = FNMS(KP599376933, T9r, T9u);
|
cannam@167
|
1373 T9T = FNMS(KP250486960, T9S, T9R);
|
cannam@167
|
1374 }
|
cannam@167
|
1375 {
|
cannam@167
|
1376 E T9k, T9O, T9n, T9P, T9j, T9m;
|
cannam@167
|
1377 T9j = T87 - T86;
|
cannam@167
|
1378 T9k = FNMS(KP831469612, T9j, T9i);
|
cannam@167
|
1379 T9O = FMA(KP831469612, T9j, T9i);
|
cannam@167
|
1380 T9m = T7X - T80;
|
cannam@167
|
1381 T9n = FNMS(KP831469612, T9m, T9l);
|
cannam@167
|
1382 T9P = FMA(KP831469612, T9m, T9l);
|
cannam@167
|
1383 T9o = FMA(KP599376933, T9n, T9k);
|
cannam@167
|
1384 Ta1 = FMA(KP250486960, T9O, T9P);
|
cannam@167
|
1385 T9G = FNMS(KP599376933, T9k, T9n);
|
cannam@167
|
1386 T9Q = FNMS(KP250486960, T9P, T9O);
|
cannam@167
|
1387 }
|
cannam@167
|
1388 }
|
cannam@167
|
1389 {
|
cannam@167
|
1390 E T9h, T9w, T9F, T9I;
|
cannam@167
|
1391 T9h = FMA(KP881921264, T9g, T99);
|
cannam@167
|
1392 T9w = T9o + T9v;
|
cannam@167
|
1393 Cr[WS(csr, 53)] = FNMS(KP857728610, T9w, T9h);
|
cannam@167
|
1394 Cr[WS(csr, 11)] = FMA(KP857728610, T9w, T9h);
|
cannam@167
|
1395 T9F = FMA(KP881921264, T9C, T9z);
|
cannam@167
|
1396 T9I = T9G - T9H;
|
cannam@167
|
1397 Ci[WS(csi, 11)] = FMA(KP857728610, T9I, T9F);
|
cannam@167
|
1398 Ci[WS(csi, 53)] = FMS(KP857728610, T9I, T9F);
|
cannam@167
|
1399 }
|
cannam@167
|
1400 {
|
cannam@167
|
1401 E T9D, T9E, T9J, T9K;
|
cannam@167
|
1402 T9D = FNMS(KP881921264, T9C, T9z);
|
cannam@167
|
1403 T9E = T9v - T9o;
|
cannam@167
|
1404 Ci[WS(csi, 21)] = FMS(KP857728610, T9E, T9D);
|
cannam@167
|
1405 Ci[WS(csi, 43)] = FMA(KP857728610, T9E, T9D);
|
cannam@167
|
1406 T9J = FNMS(KP881921264, T9g, T99);
|
cannam@167
|
1407 T9K = T9G + T9H;
|
cannam@167
|
1408 Cr[WS(csr, 43)] = FMA(KP857728610, T9K, T9J);
|
cannam@167
|
1409 Cr[WS(csr, 21)] = FNMS(KP857728610, T9K, T9J);
|
cannam@167
|
1410 }
|
cannam@167
|
1411 {
|
cannam@167
|
1412 E T9N, T9U, T9Z, Ta2;
|
cannam@167
|
1413 T9N = FMA(KP881921264, T9M, T9L);
|
cannam@167
|
1414 T9U = T9Q + T9T;
|
cannam@167
|
1415 Cr[WS(csr, 59)] = FNMS(KP970031253, T9U, T9N);
|
cannam@167
|
1416 Cr[WS(csr, 5)] = FMA(KP970031253, T9U, T9N);
|
cannam@167
|
1417 T9Z = FMA(KP881921264, T9W, T9V);
|
cannam@167
|
1418 Ta2 = Ta0 - Ta1;
|
cannam@167
|
1419 Ci[WS(csi, 5)] = FMS(KP970031253, Ta2, T9Z);
|
cannam@167
|
1420 Ci[WS(csi, 59)] = FMA(KP970031253, Ta2, T9Z);
|
cannam@167
|
1421 }
|
cannam@167
|
1422 {
|
cannam@167
|
1423 E T9X, T9Y, Ta3, Ta4;
|
cannam@167
|
1424 T9X = FNMS(KP881921264, T9W, T9V);
|
cannam@167
|
1425 T9Y = T9T - T9Q;
|
cannam@167
|
1426 Ci[WS(csi, 27)] = FMA(KP970031253, T9Y, T9X);
|
cannam@167
|
1427 Ci[WS(csi, 37)] = FMS(KP970031253, T9Y, T9X);
|
cannam@167
|
1428 Ta3 = FNMS(KP881921264, T9M, T9L);
|
cannam@167
|
1429 Ta4 = Ta1 + Ta0;
|
cannam@167
|
1430 Cr[WS(csr, 37)] = FNMS(KP970031253, Ta4, Ta3);
|
cannam@167
|
1431 Cr[WS(csr, 27)] = FMA(KP970031253, Ta4, Ta3);
|
cannam@167
|
1432 }
|
cannam@167
|
1433 }
|
cannam@167
|
1434 {
|
cannam@167
|
1435 E T2z, T69, T3g, T6k, T60, T6a, T5X, T6j, T5L, T6p, T64, T6h, T4w, T6o, T65;
|
cannam@167
|
1436 E T6e;
|
cannam@167
|
1437 {
|
cannam@167
|
1438 E T2f, T2y, T5T, T5W;
|
cannam@167
|
1439 T2f = FMA(KP923879532, T2e, T27);
|
cannam@167
|
1440 T2y = T2o + T2x;
|
cannam@167
|
1441 T2z = FMA(KP980785280, T2y, T2f);
|
cannam@167
|
1442 T69 = FNMS(KP980785280, T2y, T2f);
|
cannam@167
|
1443 {
|
cannam@167
|
1444 E T2U, T3f, T5Y, T5Z;
|
cannam@167
|
1445 T2U = FNMS(KP098491403, T2T, T2M);
|
cannam@167
|
1446 T3f = FMA(KP098491403, T3e, T37);
|
cannam@167
|
1447 T3g = T2U + T3f;
|
cannam@167
|
1448 T6k = T3f - T2U;
|
cannam@167
|
1449 T5Y = FMA(KP098491403, T2M, T2T);
|
cannam@167
|
1450 T5Z = FNMS(KP098491403, T37, T3e);
|
cannam@167
|
1451 T60 = T5Y + T5Z;
|
cannam@167
|
1452 T6a = T5Y - T5Z;
|
cannam@167
|
1453 }
|
cannam@167
|
1454 T5T = FMA(KP923879532, T5S, T5P);
|
cannam@167
|
1455 T5W = T5U + T5V;
|
cannam@167
|
1456 T5X = FMA(KP980785280, T5W, T5T);
|
cannam@167
|
1457 T6j = FNMS(KP980785280, T5W, T5T);
|
cannam@167
|
1458 {
|
cannam@167
|
1459 E T5x, T6f, T5K, T6g, T5w, T5J;
|
cannam@167
|
1460 T5w = T5e + T5v;
|
cannam@167
|
1461 T5x = FMA(KP980785280, T5w, T4X);
|
cannam@167
|
1462 T6f = FNMS(KP980785280, T5w, T4X);
|
cannam@167
|
1463 T5J = T5H + T5I;
|
cannam@167
|
1464 T5K = FMA(KP980785280, T5J, T5G);
|
cannam@167
|
1465 T6g = FNMS(KP980785280, T5J, T5G);
|
cannam@167
|
1466 T5L = FNMS(KP049126849, T5K, T5x);
|
cannam@167
|
1467 T6p = FNMS(KP906347169, T6f, T6g);
|
cannam@167
|
1468 T64 = FMA(KP049126849, T5x, T5K);
|
cannam@167
|
1469 T6h = FMA(KP906347169, T6g, T6f);
|
cannam@167
|
1470 }
|
cannam@167
|
1471 {
|
cannam@167
|
1472 E T4i, T6c, T4v, T6d, T4h, T4u;
|
cannam@167
|
1473 T4h = T3Z + T4g;
|
cannam@167
|
1474 T4i = FMA(KP980785280, T4h, T3I);
|
cannam@167
|
1475 T6c = FNMS(KP980785280, T4h, T3I);
|
cannam@167
|
1476 T4u = T4s + T4t;
|
cannam@167
|
1477 T4v = FMA(KP980785280, T4u, T4r);
|
cannam@167
|
1478 T6d = FNMS(KP980785280, T4u, T4r);
|
cannam@167
|
1479 T4w = FNMS(KP049126849, T4v, T4i);
|
cannam@167
|
1480 T6o = FNMS(KP906347169, T6c, T6d);
|
cannam@167
|
1481 T65 = FMA(KP049126849, T4i, T4v);
|
cannam@167
|
1482 T6e = FMA(KP906347169, T6d, T6c);
|
cannam@167
|
1483 }
|
cannam@167
|
1484 }
|
cannam@167
|
1485 {
|
cannam@167
|
1486 E T3h, T5M, T63, T66;
|
cannam@167
|
1487 T3h = FMA(KP995184726, T3g, T2z);
|
cannam@167
|
1488 T5M = T4w + T5L;
|
cannam@167
|
1489 Cr[WS(csr, 63)] = FNMS(KP998795456, T5M, T3h);
|
cannam@167
|
1490 Cr[WS(csr, 1)] = FMA(KP998795456, T5M, T3h);
|
cannam@167
|
1491 T63 = FMA(KP995184726, T60, T5X);
|
cannam@167
|
1492 T66 = T64 - T65;
|
cannam@167
|
1493 Ci[WS(csi, 1)] = FMS(KP998795456, T66, T63);
|
cannam@167
|
1494 Ci[WS(csi, 63)] = FMA(KP998795456, T66, T63);
|
cannam@167
|
1495 }
|
cannam@167
|
1496 {
|
cannam@167
|
1497 E T61, T62, T67, T68;
|
cannam@167
|
1498 T61 = FNMS(KP995184726, T60, T5X);
|
cannam@167
|
1499 T62 = T5L - T4w;
|
cannam@167
|
1500 Ci[WS(csi, 31)] = FMA(KP998795456, T62, T61);
|
cannam@167
|
1501 Ci[WS(csi, 33)] = FMS(KP998795456, T62, T61);
|
cannam@167
|
1502 T67 = FNMS(KP995184726, T3g, T2z);
|
cannam@167
|
1503 T68 = T65 + T64;
|
cannam@167
|
1504 Cr[WS(csr, 33)] = FNMS(KP998795456, T68, T67);
|
cannam@167
|
1505 Cr[WS(csr, 31)] = FMA(KP998795456, T68, T67);
|
cannam@167
|
1506 }
|
cannam@167
|
1507 {
|
cannam@167
|
1508 E T6b, T6i, T6n, T6q;
|
cannam@167
|
1509 T6b = FMA(KP995184726, T6a, T69);
|
cannam@167
|
1510 T6i = T6e + T6h;
|
cannam@167
|
1511 Cr[WS(csr, 49)] = FNMS(KP740951125, T6i, T6b);
|
cannam@167
|
1512 Cr[WS(csr, 15)] = FMA(KP740951125, T6i, T6b);
|
cannam@167
|
1513 T6n = FMA(KP995184726, T6k, T6j);
|
cannam@167
|
1514 T6q = T6o - T6p;
|
cannam@167
|
1515 Ci[WS(csi, 15)] = FMA(KP740951125, T6q, T6n);
|
cannam@167
|
1516 Ci[WS(csi, 49)] = FMS(KP740951125, T6q, T6n);
|
cannam@167
|
1517 }
|
cannam@167
|
1518 {
|
cannam@167
|
1519 E T6l, T6m, T6r, T6s;
|
cannam@167
|
1520 T6l = FNMS(KP995184726, T6k, T6j);
|
cannam@167
|
1521 T6m = T6h - T6e;
|
cannam@167
|
1522 Ci[WS(csi, 17)] = FMS(KP740951125, T6m, T6l);
|
cannam@167
|
1523 Ci[WS(csi, 47)] = FMA(KP740951125, T6m, T6l);
|
cannam@167
|
1524 T6r = FNMS(KP995184726, T6a, T69);
|
cannam@167
|
1525 T6s = T6o + T6p;
|
cannam@167
|
1526 Cr[WS(csr, 47)] = FMA(KP740951125, T6s, T6r);
|
cannam@167
|
1527 Cr[WS(csr, 17)] = FNMS(KP740951125, T6s, T6r);
|
cannam@167
|
1528 }
|
cannam@167
|
1529 }
|
cannam@167
|
1530 {
|
cannam@167
|
1531 E T7B, T8N, T7Q, T8Y, T8E, T8O, T8B, T8X, T8t, T92, T8J, T8V, T8a, T93, T8I;
|
cannam@167
|
1532 E T8S;
|
cannam@167
|
1533 {
|
cannam@167
|
1534 E T7t, T7A, T8x, T8A;
|
cannam@167
|
1535 T7t = FMA(KP923879532, T7s, T7r);
|
cannam@167
|
1536 T7A = T7w + T7z;
|
cannam@167
|
1537 T7B = FMA(KP831469612, T7A, T7t);
|
cannam@167
|
1538 T8N = FNMS(KP831469612, T7A, T7t);
|
cannam@167
|
1539 {
|
cannam@167
|
1540 E T7I, T7P, T8C, T8D;
|
cannam@167
|
1541 T7I = FMA(KP303346683, T7H, T7E);
|
cannam@167
|
1542 T7P = FNMS(KP303346683, T7O, T7L);
|
cannam@167
|
1543 T7Q = T7I + T7P;
|
cannam@167
|
1544 T8Y = T7P - T7I;
|
cannam@167
|
1545 T8C = FNMS(KP303346683, T7E, T7H);
|
cannam@167
|
1546 T8D = FMA(KP303346683, T7L, T7O);
|
cannam@167
|
1547 T8E = T8C + T8D;
|
cannam@167
|
1548 T8O = T8D - T8C;
|
cannam@167
|
1549 }
|
cannam@167
|
1550 T8x = FMA(KP923879532, T8w, T8v);
|
cannam@167
|
1551 T8A = T8y + T8z;
|
cannam@167
|
1552 T8B = FMA(KP831469612, T8A, T8x);
|
cannam@167
|
1553 T8X = FNMS(KP831469612, T8A, T8x);
|
cannam@167
|
1554 {
|
cannam@167
|
1555 E T8l, T8T, T8s, T8U, T8k, T8r;
|
cannam@167
|
1556 T8k = T8g + T8j;
|
cannam@167
|
1557 T8l = FMA(KP831469612, T8k, T8d);
|
cannam@167
|
1558 T8T = FNMS(KP831469612, T8k, T8d);
|
cannam@167
|
1559 T8r = T8p + T8q;
|
cannam@167
|
1560 T8s = FMA(KP831469612, T8r, T8o);
|
cannam@167
|
1561 T8U = FNMS(KP831469612, T8r, T8o);
|
cannam@167
|
1562 T8t = FMA(KP148335987, T8s, T8l);
|
cannam@167
|
1563 T92 = FMA(KP741650546, T8T, T8U);
|
cannam@167
|
1564 T8J = FNMS(KP148335987, T8l, T8s);
|
cannam@167
|
1565 T8V = FNMS(KP741650546, T8U, T8T);
|
cannam@167
|
1566 }
|
cannam@167
|
1567 {
|
cannam@167
|
1568 E T82, T8Q, T89, T8R, T81, T88;
|
cannam@167
|
1569 T81 = T7X + T80;
|
cannam@167
|
1570 T82 = FMA(KP831469612, T81, T7U);
|
cannam@167
|
1571 T8Q = FNMS(KP831469612, T81, T7U);
|
cannam@167
|
1572 T88 = T86 + T87;
|
cannam@167
|
1573 T89 = FMA(KP831469612, T88, T85);
|
cannam@167
|
1574 T8R = FNMS(KP831469612, T88, T85);
|
cannam@167
|
1575 T8a = FMA(KP148335987, T89, T82);
|
cannam@167
|
1576 T93 = FMA(KP741650546, T8Q, T8R);
|
cannam@167
|
1577 T8I = FNMS(KP148335987, T82, T89);
|
cannam@167
|
1578 T8S = FNMS(KP741650546, T8R, T8Q);
|
cannam@167
|
1579 }
|
cannam@167
|
1580 }
|
cannam@167
|
1581 {
|
cannam@167
|
1582 E T7R, T8u, T8H, T8K;
|
cannam@167
|
1583 T7R = FMA(KP956940335, T7Q, T7B);
|
cannam@167
|
1584 T8u = T8a + T8t;
|
cannam@167
|
1585 Cr[WS(csr, 61)] = FNMS(KP989176509, T8u, T7R);
|
cannam@167
|
1586 Cr[WS(csr, 3)] = FMA(KP989176509, T8u, T7R);
|
cannam@167
|
1587 T8H = FMA(KP956940335, T8E, T8B);
|
cannam@167
|
1588 T8K = T8I - T8J;
|
cannam@167
|
1589 Ci[WS(csi, 3)] = FMA(KP989176509, T8K, T8H);
|
cannam@167
|
1590 Ci[WS(csi, 61)] = FMS(KP989176509, T8K, T8H);
|
cannam@167
|
1591 }
|
cannam@167
|
1592 {
|
cannam@167
|
1593 E T8F, T8G, T8L, T8M;
|
cannam@167
|
1594 T8F = FNMS(KP956940335, T8E, T8B);
|
cannam@167
|
1595 T8G = T8t - T8a;
|
cannam@167
|
1596 Ci[WS(csi, 29)] = FMS(KP989176509, T8G, T8F);
|
cannam@167
|
1597 Ci[WS(csi, 35)] = FMA(KP989176509, T8G, T8F);
|
cannam@167
|
1598 T8L = FNMS(KP956940335, T7Q, T7B);
|
cannam@167
|
1599 T8M = T8I + T8J;
|
cannam@167
|
1600 Cr[WS(csr, 35)] = FMA(KP989176509, T8M, T8L);
|
cannam@167
|
1601 Cr[WS(csr, 29)] = FNMS(KP989176509, T8M, T8L);
|
cannam@167
|
1602 }
|
cannam@167
|
1603 {
|
cannam@167
|
1604 E T8P, T8W, T91, T94;
|
cannam@167
|
1605 T8P = FMA(KP956940335, T8O, T8N);
|
cannam@167
|
1606 T8W = T8S + T8V;
|
cannam@167
|
1607 Cr[WS(csr, 51)] = FNMS(KP803207531, T8W, T8P);
|
cannam@167
|
1608 Cr[WS(csr, 13)] = FMA(KP803207531, T8W, T8P);
|
cannam@167
|
1609 T91 = FNMS(KP956940335, T8Y, T8X);
|
cannam@167
|
1610 T94 = T92 - T93;
|
cannam@167
|
1611 Ci[WS(csi, 13)] = FMS(KP803207531, T94, T91);
|
cannam@167
|
1612 Ci[WS(csi, 51)] = FMA(KP803207531, T94, T91);
|
cannam@167
|
1613 }
|
cannam@167
|
1614 {
|
cannam@167
|
1615 E T8Z, T90, T95, T96;
|
cannam@167
|
1616 T8Z = FMA(KP956940335, T8Y, T8X);
|
cannam@167
|
1617 T90 = T8V - T8S;
|
cannam@167
|
1618 Ci[WS(csi, 19)] = FMA(KP803207531, T90, T8Z);
|
cannam@167
|
1619 Ci[WS(csi, 45)] = FMS(KP803207531, T90, T8Z);
|
cannam@167
|
1620 T95 = FNMS(KP956940335, T8O, T8N);
|
cannam@167
|
1621 T96 = T93 + T92;
|
cannam@167
|
1622 Cr[WS(csr, 45)] = FNMS(KP803207531, T96, T95);
|
cannam@167
|
1623 Cr[WS(csr, 19)] = FMA(KP803207531, T96, T95);
|
cannam@167
|
1624 }
|
cannam@167
|
1625 }
|
cannam@167
|
1626 }
|
cannam@167
|
1627 }
|
cannam@167
|
1628 }
|
cannam@167
|
1629
|
cannam@167
|
1630 static const kr2c_desc desc = { 128, "r2cf_128", {440, 0, 516, 0}, &GENUS };
|
cannam@167
|
1631
|
cannam@167
|
1632 void X(codelet_r2cf_128) (planner *p) {
|
cannam@167
|
1633 X(kr2c_register) (p, r2cf_128, &desc);
|
cannam@167
|
1634 }
|
cannam@167
|
1635
|
cannam@167
|
1636 #else
|
cannam@167
|
1637
|
cannam@167
|
1638 /* Generated by: ../../../genfft/gen_r2cf.native -compact -variables 4 -pipeline-latency 4 -n 128 -name r2cf_128 -include rdft/scalar/r2cf.h */
|
cannam@167
|
1639
|
cannam@167
|
1640 /*
|
cannam@167
|
1641 * This function contains 956 FP additions, 330 FP multiplications,
|
cannam@167
|
1642 * (or, 812 additions, 186 multiplications, 144 fused multiply/add),
|
cannam@167
|
1643 * 186 stack variables, 31 constants, and 256 memory accesses
|
cannam@167
|
1644 */
|
cannam@167
|
1645 #include "rdft/scalar/r2cf.h"
|
cannam@167
|
1646
|
cannam@167
|
1647 static void r2cf_128(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
|
cannam@167
|
1648 {
|
cannam@167
|
1649 DK(KP803207531, +0.803207531480644909806676512963141923879569427);
|
cannam@167
|
1650 DK(KP595699304, +0.595699304492433343467036528829969889511926338);
|
cannam@167
|
1651 DK(KP146730474, +0.146730474455361751658850129646717819706215317);
|
cannam@167
|
1652 DK(KP989176509, +0.989176509964780973451673738016243063983689533);
|
cannam@167
|
1653 DK(KP740951125, +0.740951125354959091175616897495162729728955309);
|
cannam@167
|
1654 DK(KP671558954, +0.671558954847018400625376850427421803228750632);
|
cannam@167
|
1655 DK(KP049067674, +0.049067674327418014254954976942682658314745363);
|
cannam@167
|
1656 DK(KP998795456, +0.998795456205172392714771604759100694443203615);
|
cannam@167
|
1657 DK(KP242980179, +0.242980179903263889948274162077471118320990783);
|
cannam@167
|
1658 DK(KP970031253, +0.970031253194543992603984207286100251456865962);
|
cannam@167
|
1659 DK(KP514102744, +0.514102744193221726593693838968815772608049120);
|
cannam@167
|
1660 DK(KP857728610, +0.857728610000272069902269984284770137042490799);
|
cannam@167
|
1661 DK(KP336889853, +0.336889853392220050689253212619147570477766780);
|
cannam@167
|
1662 DK(KP941544065, +0.941544065183020778412509402599502357185589796);
|
cannam@167
|
1663 DK(KP427555093, +0.427555093430282094320966856888798534304578629);
|
cannam@167
|
1664 DK(KP903989293, +0.903989293123443331586200297230537048710132025);
|
cannam@167
|
1665 DK(KP098017140, +0.098017140329560601994195563888641845861136673);
|
cannam@167
|
1666 DK(KP995184726, +0.995184726672196886244836953109479921575474869);
|
cannam@167
|
1667 DK(KP634393284, +0.634393284163645498215171613225493370675687095);
|
cannam@167
|
1668 DK(KP773010453, +0.773010453362736960810906609758469800971041293);
|
cannam@167
|
1669 DK(KP881921264, +0.881921264348355029712756863660388349508442621);
|
cannam@167
|
1670 DK(KP471396736, +0.471396736825997648556387625905254377657460319);
|
cannam@167
|
1671 DK(KP956940335, +0.956940335732208864935797886980269969482849206);
|
cannam@167
|
1672 DK(KP290284677, +0.290284677254462367636192375817395274691476278);
|
cannam@167
|
1673 DK(KP555570233, +0.555570233019602224742830813948532874374937191);
|
cannam@167
|
1674 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
cannam@167
|
1675 DK(KP195090322, +0.195090322016128267848284868477022240927691618);
|
cannam@167
|
1676 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
cannam@167
|
1677 DK(KP382683432, +0.382683432365089771728459984030398866761344562);
|
cannam@167
|
1678 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
cannam@167
|
1679 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
cannam@167
|
1680 {
|
cannam@167
|
1681 INT i;
|
cannam@167
|
1682 for (i = v; i > 0; i = i - 1, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(512, rs), MAKE_VOLATILE_STRIDE(512, csr), MAKE_VOLATILE_STRIDE(512, csi)) {
|
cannam@167
|
1683 E TcD, TdU, T27, T7r, T5S, T8y, Tf, Ta5, Tu, Tbq, TcG, TdV, T2e, T8z, T5V;
|
cannam@167
|
1684 E T7s, TK, Ta6, TcK, TdX, T2o, T5X, T7w, T8B, TZ, Ta7, TcN, TdY, T2x, T5Y;
|
cannam@167
|
1685 E T7z, T8C, T1g, Taa, TcU, TeA, TcX, Tez, T1v, Tab, T2M, T6z, T7E, T9e, T7H;
|
cannam@167
|
1686 E T9d, T2T, T6A, T4X, T6L, Tdz, TeL, TdK, TeP, T5G, T6P, T8d, T9p, TaV, Tc3;
|
cannam@167
|
1687 E Tbi, Tc4, T8o, T9t, T3I, T6H, Tde, TeH, Tdp, TeF, T4r, T6F, T7U, T9l, Tao;
|
cannam@167
|
1688 E TbW, TaL, TbX, T85, T9j, T1L, Tad, Td3, Tew, Td6, Tex, T20, Tae, T37, T6x;
|
cannam@167
|
1689 E T7L, T9a, T7O, T9b, T3e, T6w, TbZ, Tc0, T3Z, T4s, Tds, TeI, T4g, T4t, T80;
|
cannam@167
|
1690 E T87, Tdl, TeE, T7X, T86, TaD, TaM, Tc6, Tc7, T5e, T5H, TdN, TeM, T5v, T5I;
|
cannam@167
|
1691 E T8j, T8q, TdG, TeO, T8g, T8p, Tba, Tbj;
|
cannam@167
|
1692 {
|
cannam@167
|
1693 E T3, T23, Td, T25, T6, T5R, Ta, T24;
|
cannam@167
|
1694 {
|
cannam@167
|
1695 E T1, T2, Tb, Tc;
|
cannam@167
|
1696 T1 = R0[0];
|
cannam@167
|
1697 T2 = R0[WS(rs, 32)];
|
cannam@167
|
1698 T3 = T1 + T2;
|
cannam@167
|
1699 T23 = T1 - T2;
|
cannam@167
|
1700 Tb = R0[WS(rs, 56)];
|
cannam@167
|
1701 Tc = R0[WS(rs, 24)];
|
cannam@167
|
1702 Td = Tb + Tc;
|
cannam@167
|
1703 T25 = Tb - Tc;
|
cannam@167
|
1704 }
|
cannam@167
|
1705 {
|
cannam@167
|
1706 E T4, T5, T8, T9;
|
cannam@167
|
1707 T4 = R0[WS(rs, 16)];
|
cannam@167
|
1708 T5 = R0[WS(rs, 48)];
|
cannam@167
|
1709 T6 = T4 + T5;
|
cannam@167
|
1710 T5R = T4 - T5;
|
cannam@167
|
1711 T8 = R0[WS(rs, 8)];
|
cannam@167
|
1712 T9 = R0[WS(rs, 40)];
|
cannam@167
|
1713 Ta = T8 + T9;
|
cannam@167
|
1714 T24 = T8 - T9;
|
cannam@167
|
1715 }
|
cannam@167
|
1716 TcD = T3 - T6;
|
cannam@167
|
1717 TdU = Td - Ta;
|
cannam@167
|
1718 {
|
cannam@167
|
1719 E T26, T5Q, T7, Te;
|
cannam@167
|
1720 T26 = KP707106781 * (T24 + T25);
|
cannam@167
|
1721 T27 = T23 + T26;
|
cannam@167
|
1722 T7r = T23 - T26;
|
cannam@167
|
1723 T5Q = KP707106781 * (T25 - T24);
|
cannam@167
|
1724 T5S = T5Q - T5R;
|
cannam@167
|
1725 T8y = T5R + T5Q;
|
cannam@167
|
1726 T7 = T3 + T6;
|
cannam@167
|
1727 Te = Ta + Td;
|
cannam@167
|
1728 Tf = T7 + Te;
|
cannam@167
|
1729 Ta5 = T7 - Te;
|
cannam@167
|
1730 }
|
cannam@167
|
1731 }
|
cannam@167
|
1732 {
|
cannam@167
|
1733 E Ti, T28, Ts, T2c, Tl, T29, Tp, T2b;
|
cannam@167
|
1734 {
|
cannam@167
|
1735 E Tg, Th, Tq, Tr;
|
cannam@167
|
1736 Tg = R0[WS(rs, 4)];
|
cannam@167
|
1737 Th = R0[WS(rs, 36)];
|
cannam@167
|
1738 Ti = Tg + Th;
|
cannam@167
|
1739 T28 = Tg - Th;
|
cannam@167
|
1740 Tq = R0[WS(rs, 12)];
|
cannam@167
|
1741 Tr = R0[WS(rs, 44)];
|
cannam@167
|
1742 Ts = Tq + Tr;
|
cannam@167
|
1743 T2c = Tq - Tr;
|
cannam@167
|
1744 }
|
cannam@167
|
1745 {
|
cannam@167
|
1746 E Tj, Tk, Tn, To;
|
cannam@167
|
1747 Tj = R0[WS(rs, 20)];
|
cannam@167
|
1748 Tk = R0[WS(rs, 52)];
|
cannam@167
|
1749 Tl = Tj + Tk;
|
cannam@167
|
1750 T29 = Tj - Tk;
|
cannam@167
|
1751 Tn = R0[WS(rs, 60)];
|
cannam@167
|
1752 To = R0[WS(rs, 28)];
|
cannam@167
|
1753 Tp = Tn + To;
|
cannam@167
|
1754 T2b = Tn - To;
|
cannam@167
|
1755 }
|
cannam@167
|
1756 {
|
cannam@167
|
1757 E Tm, Tt, TcE, TcF;
|
cannam@167
|
1758 Tm = Ti + Tl;
|
cannam@167
|
1759 Tt = Tp + Ts;
|
cannam@167
|
1760 Tu = Tm + Tt;
|
cannam@167
|
1761 Tbq = Tt - Tm;
|
cannam@167
|
1762 TcE = Ti - Tl;
|
cannam@167
|
1763 TcF = Tp - Ts;
|
cannam@167
|
1764 TcG = KP707106781 * (TcE + TcF);
|
cannam@167
|
1765 TdV = KP707106781 * (TcF - TcE);
|
cannam@167
|
1766 }
|
cannam@167
|
1767 {
|
cannam@167
|
1768 E T2a, T2d, T5T, T5U;
|
cannam@167
|
1769 T2a = FNMS(KP382683432, T29, KP923879532 * T28);
|
cannam@167
|
1770 T2d = FMA(KP923879532, T2b, KP382683432 * T2c);
|
cannam@167
|
1771 T2e = T2a + T2d;
|
cannam@167
|
1772 T8z = T2d - T2a;
|
cannam@167
|
1773 T5T = FNMS(KP923879532, T2c, KP382683432 * T2b);
|
cannam@167
|
1774 T5U = FMA(KP382683432, T28, KP923879532 * T29);
|
cannam@167
|
1775 T5V = T5T - T5U;
|
cannam@167
|
1776 T7s = T5U + T5T;
|
cannam@167
|
1777 }
|
cannam@167
|
1778 }
|
cannam@167
|
1779 {
|
cannam@167
|
1780 E Ty, T2g, TB, T2m, TF, T2l, TI, T2j;
|
cannam@167
|
1781 {
|
cannam@167
|
1782 E Tw, Tx, Tz, TA;
|
cannam@167
|
1783 Tw = R0[WS(rs, 2)];
|
cannam@167
|
1784 Tx = R0[WS(rs, 34)];
|
cannam@167
|
1785 Ty = Tw + Tx;
|
cannam@167
|
1786 T2g = Tw - Tx;
|
cannam@167
|
1787 Tz = R0[WS(rs, 18)];
|
cannam@167
|
1788 TA = R0[WS(rs, 50)];
|
cannam@167
|
1789 TB = Tz + TA;
|
cannam@167
|
1790 T2m = Tz - TA;
|
cannam@167
|
1791 {
|
cannam@167
|
1792 E TD, TE, T2h, TG, TH, T2i;
|
cannam@167
|
1793 TD = R0[WS(rs, 10)];
|
cannam@167
|
1794 TE = R0[WS(rs, 42)];
|
cannam@167
|
1795 T2h = TD - TE;
|
cannam@167
|
1796 TG = R0[WS(rs, 58)];
|
cannam@167
|
1797 TH = R0[WS(rs, 26)];
|
cannam@167
|
1798 T2i = TG - TH;
|
cannam@167
|
1799 TF = TD + TE;
|
cannam@167
|
1800 T2l = KP707106781 * (T2i - T2h);
|
cannam@167
|
1801 TI = TG + TH;
|
cannam@167
|
1802 T2j = KP707106781 * (T2h + T2i);
|
cannam@167
|
1803 }
|
cannam@167
|
1804 }
|
cannam@167
|
1805 {
|
cannam@167
|
1806 E TC, TJ, TcI, TcJ;
|
cannam@167
|
1807 TC = Ty + TB;
|
cannam@167
|
1808 TJ = TF + TI;
|
cannam@167
|
1809 TK = TC + TJ;
|
cannam@167
|
1810 Ta6 = TC - TJ;
|
cannam@167
|
1811 TcI = Ty - TB;
|
cannam@167
|
1812 TcJ = TI - TF;
|
cannam@167
|
1813 TcK = FMA(KP923879532, TcI, KP382683432 * TcJ);
|
cannam@167
|
1814 TdX = FNMS(KP382683432, TcI, KP923879532 * TcJ);
|
cannam@167
|
1815 }
|
cannam@167
|
1816 {
|
cannam@167
|
1817 E T2k, T2n, T7u, T7v;
|
cannam@167
|
1818 T2k = T2g + T2j;
|
cannam@167
|
1819 T2n = T2l - T2m;
|
cannam@167
|
1820 T2o = FMA(KP980785280, T2k, KP195090322 * T2n);
|
cannam@167
|
1821 T5X = FNMS(KP195090322, T2k, KP980785280 * T2n);
|
cannam@167
|
1822 T7u = T2g - T2j;
|
cannam@167
|
1823 T7v = T2m + T2l;
|
cannam@167
|
1824 T7w = FMA(KP831469612, T7u, KP555570233 * T7v);
|
cannam@167
|
1825 T8B = FNMS(KP555570233, T7u, KP831469612 * T7v);
|
cannam@167
|
1826 }
|
cannam@167
|
1827 }
|
cannam@167
|
1828 {
|
cannam@167
|
1829 E TN, T2p, TQ, T2v, TU, T2u, TX, T2s;
|
cannam@167
|
1830 {
|
cannam@167
|
1831 E TL, TM, TO, TP;
|
cannam@167
|
1832 TL = R0[WS(rs, 62)];
|
cannam@167
|
1833 TM = R0[WS(rs, 30)];
|
cannam@167
|
1834 TN = TL + TM;
|
cannam@167
|
1835 T2p = TL - TM;
|
cannam@167
|
1836 TO = R0[WS(rs, 14)];
|
cannam@167
|
1837 TP = R0[WS(rs, 46)];
|
cannam@167
|
1838 TQ = TO + TP;
|
cannam@167
|
1839 T2v = TO - TP;
|
cannam@167
|
1840 {
|
cannam@167
|
1841 E TS, TT, T2q, TV, TW, T2r;
|
cannam@167
|
1842 TS = R0[WS(rs, 6)];
|
cannam@167
|
1843 TT = R0[WS(rs, 38)];
|
cannam@167
|
1844 T2q = TS - TT;
|
cannam@167
|
1845 TV = R0[WS(rs, 54)];
|
cannam@167
|
1846 TW = R0[WS(rs, 22)];
|
cannam@167
|
1847 T2r = TV - TW;
|
cannam@167
|
1848 TU = TS + TT;
|
cannam@167
|
1849 T2u = KP707106781 * (T2r - T2q);
|
cannam@167
|
1850 TX = TV + TW;
|
cannam@167
|
1851 T2s = KP707106781 * (T2q + T2r);
|
cannam@167
|
1852 }
|
cannam@167
|
1853 }
|
cannam@167
|
1854 {
|
cannam@167
|
1855 E TR, TY, TcL, TcM;
|
cannam@167
|
1856 TR = TN + TQ;
|
cannam@167
|
1857 TY = TU + TX;
|
cannam@167
|
1858 TZ = TR + TY;
|
cannam@167
|
1859 Ta7 = TR - TY;
|
cannam@167
|
1860 TcL = TN - TQ;
|
cannam@167
|
1861 TcM = TX - TU;
|
cannam@167
|
1862 TcN = FNMS(KP382683432, TcM, KP923879532 * TcL);
|
cannam@167
|
1863 TdY = FMA(KP382683432, TcL, KP923879532 * TcM);
|
cannam@167
|
1864 }
|
cannam@167
|
1865 {
|
cannam@167
|
1866 E T2t, T2w, T7x, T7y;
|
cannam@167
|
1867 T2t = T2p + T2s;
|
cannam@167
|
1868 T2w = T2u - T2v;
|
cannam@167
|
1869 T2x = FNMS(KP195090322, T2w, KP980785280 * T2t);
|
cannam@167
|
1870 T5Y = FMA(KP195090322, T2t, KP980785280 * T2w);
|
cannam@167
|
1871 T7x = T2p - T2s;
|
cannam@167
|
1872 T7y = T2v + T2u;
|
cannam@167
|
1873 T7z = FNMS(KP555570233, T7y, KP831469612 * T7x);
|
cannam@167
|
1874 T8C = FMA(KP555570233, T7x, KP831469612 * T7y);
|
cannam@167
|
1875 }
|
cannam@167
|
1876 }
|
cannam@167
|
1877 {
|
cannam@167
|
1878 E T14, T2N, T17, T2D, T1b, T2O, T1e, T2C, T1j, T1m, T2K, TcR, T2Q, T1q, T1t;
|
cannam@167
|
1879 E T2H, TcS, T2R;
|
cannam@167
|
1880 {
|
cannam@167
|
1881 E T12, T13, T15, T16;
|
cannam@167
|
1882 T12 = R0[WS(rs, 1)];
|
cannam@167
|
1883 T13 = R0[WS(rs, 33)];
|
cannam@167
|
1884 T14 = T12 + T13;
|
cannam@167
|
1885 T2N = T12 - T13;
|
cannam@167
|
1886 T15 = R0[WS(rs, 17)];
|
cannam@167
|
1887 T16 = R0[WS(rs, 49)];
|
cannam@167
|
1888 T17 = T15 + T16;
|
cannam@167
|
1889 T2D = T15 - T16;
|
cannam@167
|
1890 }
|
cannam@167
|
1891 {
|
cannam@167
|
1892 E T19, T1a, T2B, T1c, T1d, T2A;
|
cannam@167
|
1893 T19 = R0[WS(rs, 9)];
|
cannam@167
|
1894 T1a = R0[WS(rs, 41)];
|
cannam@167
|
1895 T2B = T19 - T1a;
|
cannam@167
|
1896 T1c = R0[WS(rs, 57)];
|
cannam@167
|
1897 T1d = R0[WS(rs, 25)];
|
cannam@167
|
1898 T2A = T1c - T1d;
|
cannam@167
|
1899 T1b = T19 + T1a;
|
cannam@167
|
1900 T2O = KP707106781 * (T2B + T2A);
|
cannam@167
|
1901 T1e = T1c + T1d;
|
cannam@167
|
1902 T2C = KP707106781 * (T2A - T2B);
|
cannam@167
|
1903 }
|
cannam@167
|
1904 {
|
cannam@167
|
1905 E T2I, T2J, T2F, T2G;
|
cannam@167
|
1906 {
|
cannam@167
|
1907 E T1h, T1i, T1k, T1l;
|
cannam@167
|
1908 T1h = R0[WS(rs, 5)];
|
cannam@167
|
1909 T1i = R0[WS(rs, 37)];
|
cannam@167
|
1910 T1j = T1h + T1i;
|
cannam@167
|
1911 T2I = T1h - T1i;
|
cannam@167
|
1912 T1k = R0[WS(rs, 21)];
|
cannam@167
|
1913 T1l = R0[WS(rs, 53)];
|
cannam@167
|
1914 T1m = T1k + T1l;
|
cannam@167
|
1915 T2J = T1k - T1l;
|
cannam@167
|
1916 }
|
cannam@167
|
1917 T2K = FMA(KP382683432, T2I, KP923879532 * T2J);
|
cannam@167
|
1918 TcR = T1j - T1m;
|
cannam@167
|
1919 T2Q = FNMS(KP382683432, T2J, KP923879532 * T2I);
|
cannam@167
|
1920 {
|
cannam@167
|
1921 E T1o, T1p, T1r, T1s;
|
cannam@167
|
1922 T1o = R0[WS(rs, 61)];
|
cannam@167
|
1923 T1p = R0[WS(rs, 29)];
|
cannam@167
|
1924 T1q = T1o + T1p;
|
cannam@167
|
1925 T2F = T1o - T1p;
|
cannam@167
|
1926 T1r = R0[WS(rs, 13)];
|
cannam@167
|
1927 T1s = R0[WS(rs, 45)];
|
cannam@167
|
1928 T1t = T1r + T1s;
|
cannam@167
|
1929 T2G = T1r - T1s;
|
cannam@167
|
1930 }
|
cannam@167
|
1931 T2H = FNMS(KP923879532, T2G, KP382683432 * T2F);
|
cannam@167
|
1932 TcS = T1q - T1t;
|
cannam@167
|
1933 T2R = FMA(KP923879532, T2F, KP382683432 * T2G);
|
cannam@167
|
1934 }
|
cannam@167
|
1935 {
|
cannam@167
|
1936 E T18, T1f, TcQ, TcT;
|
cannam@167
|
1937 T18 = T14 + T17;
|
cannam@167
|
1938 T1f = T1b + T1e;
|
cannam@167
|
1939 T1g = T18 + T1f;
|
cannam@167
|
1940 Taa = T18 - T1f;
|
cannam@167
|
1941 TcQ = T14 - T17;
|
cannam@167
|
1942 TcT = KP707106781 * (TcR + TcS);
|
cannam@167
|
1943 TcU = TcQ + TcT;
|
cannam@167
|
1944 TeA = TcQ - TcT;
|
cannam@167
|
1945 }
|
cannam@167
|
1946 {
|
cannam@167
|
1947 E TcV, TcW, T1n, T1u;
|
cannam@167
|
1948 TcV = T1e - T1b;
|
cannam@167
|
1949 TcW = KP707106781 * (TcS - TcR);
|
cannam@167
|
1950 TcX = TcV + TcW;
|
cannam@167
|
1951 Tez = TcW - TcV;
|
cannam@167
|
1952 T1n = T1j + T1m;
|
cannam@167
|
1953 T1u = T1q + T1t;
|
cannam@167
|
1954 T1v = T1n + T1u;
|
cannam@167
|
1955 Tab = T1u - T1n;
|
cannam@167
|
1956 }
|
cannam@167
|
1957 {
|
cannam@167
|
1958 E T2E, T2L, T7C, T7D;
|
cannam@167
|
1959 T2E = T2C - T2D;
|
cannam@167
|
1960 T2L = T2H - T2K;
|
cannam@167
|
1961 T2M = T2E + T2L;
|
cannam@167
|
1962 T6z = T2L - T2E;
|
cannam@167
|
1963 T7C = T2N - T2O;
|
cannam@167
|
1964 T7D = T2K + T2H;
|
cannam@167
|
1965 T7E = T7C + T7D;
|
cannam@167
|
1966 T9e = T7C - T7D;
|
cannam@167
|
1967 }
|
cannam@167
|
1968 {
|
cannam@167
|
1969 E T7F, T7G, T2P, T2S;
|
cannam@167
|
1970 T7F = T2D + T2C;
|
cannam@167
|
1971 T7G = T2R - T2Q;
|
cannam@167
|
1972 T7H = T7F + T7G;
|
cannam@167
|
1973 T9d = T7G - T7F;
|
cannam@167
|
1974 T2P = T2N + T2O;
|
cannam@167
|
1975 T2S = T2Q + T2R;
|
cannam@167
|
1976 T2T = T2P + T2S;
|
cannam@167
|
1977 T6A = T2P - T2S;
|
cannam@167
|
1978 }
|
cannam@167
|
1979 }
|
cannam@167
|
1980 {
|
cannam@167
|
1981 E T4z, TaP, T5B, TaQ, T4G, TaT, T5y, TaS, Tbf, Tbg, T4O, Tdw, T5E, Tbc, Tbd;
|
cannam@167
|
1982 E T4V, Tdx, T5D;
|
cannam@167
|
1983 {
|
cannam@167
|
1984 E T4x, T4y, T5z, T5A;
|
cannam@167
|
1985 T4x = R1[WS(rs, 63)];
|
cannam@167
|
1986 T4y = R1[WS(rs, 31)];
|
cannam@167
|
1987 T4z = T4x - T4y;
|
cannam@167
|
1988 TaP = T4x + T4y;
|
cannam@167
|
1989 T5z = R1[WS(rs, 15)];
|
cannam@167
|
1990 T5A = R1[WS(rs, 47)];
|
cannam@167
|
1991 T5B = T5z - T5A;
|
cannam@167
|
1992 TaQ = T5z + T5A;
|
cannam@167
|
1993 }
|
cannam@167
|
1994 {
|
cannam@167
|
1995 E T4A, T4B, T4C, T4D, T4E, T4F;
|
cannam@167
|
1996 T4A = R1[WS(rs, 7)];
|
cannam@167
|
1997 T4B = R1[WS(rs, 39)];
|
cannam@167
|
1998 T4C = T4A - T4B;
|
cannam@167
|
1999 T4D = R1[WS(rs, 55)];
|
cannam@167
|
2000 T4E = R1[WS(rs, 23)];
|
cannam@167
|
2001 T4F = T4D - T4E;
|
cannam@167
|
2002 T4G = KP707106781 * (T4C + T4F);
|
cannam@167
|
2003 TaT = T4D + T4E;
|
cannam@167
|
2004 T5y = KP707106781 * (T4F - T4C);
|
cannam@167
|
2005 TaS = T4A + T4B;
|
cannam@167
|
2006 }
|
cannam@167
|
2007 {
|
cannam@167
|
2008 E T4K, T4N, T4R, T4U;
|
cannam@167
|
2009 {
|
cannam@167
|
2010 E T4I, T4J, T4L, T4M;
|
cannam@167
|
2011 T4I = R1[WS(rs, 3)];
|
cannam@167
|
2012 T4J = R1[WS(rs, 35)];
|
cannam@167
|
2013 T4K = T4I - T4J;
|
cannam@167
|
2014 Tbf = T4I + T4J;
|
cannam@167
|
2015 T4L = R1[WS(rs, 19)];
|
cannam@167
|
2016 T4M = R1[WS(rs, 51)];
|
cannam@167
|
2017 T4N = T4L - T4M;
|
cannam@167
|
2018 Tbg = T4L + T4M;
|
cannam@167
|
2019 }
|
cannam@167
|
2020 T4O = FNMS(KP382683432, T4N, KP923879532 * T4K);
|
cannam@167
|
2021 Tdw = Tbf - Tbg;
|
cannam@167
|
2022 T5E = FMA(KP382683432, T4K, KP923879532 * T4N);
|
cannam@167
|
2023 {
|
cannam@167
|
2024 E T4P, T4Q, T4S, T4T;
|
cannam@167
|
2025 T4P = R1[WS(rs, 59)];
|
cannam@167
|
2026 T4Q = R1[WS(rs, 27)];
|
cannam@167
|
2027 T4R = T4P - T4Q;
|
cannam@167
|
2028 Tbc = T4P + T4Q;
|
cannam@167
|
2029 T4S = R1[WS(rs, 11)];
|
cannam@167
|
2030 T4T = R1[WS(rs, 43)];
|
cannam@167
|
2031 T4U = T4S - T4T;
|
cannam@167
|
2032 Tbd = T4S + T4T;
|
cannam@167
|
2033 }
|
cannam@167
|
2034 T4V = FMA(KP923879532, T4R, KP382683432 * T4U);
|
cannam@167
|
2035 Tdx = Tbc - Tbd;
|
cannam@167
|
2036 T5D = FNMS(KP923879532, T4U, KP382683432 * T4R);
|
cannam@167
|
2037 }
|
cannam@167
|
2038 {
|
cannam@167
|
2039 E T4H, T4W, Tdv, Tdy;
|
cannam@167
|
2040 T4H = T4z + T4G;
|
cannam@167
|
2041 T4W = T4O + T4V;
|
cannam@167
|
2042 T4X = T4H + T4W;
|
cannam@167
|
2043 T6L = T4H - T4W;
|
cannam@167
|
2044 Tdv = TaP - TaQ;
|
cannam@167
|
2045 Tdy = KP707106781 * (Tdw + Tdx);
|
cannam@167
|
2046 Tdz = Tdv + Tdy;
|
cannam@167
|
2047 TeL = Tdv - Tdy;
|
cannam@167
|
2048 }
|
cannam@167
|
2049 {
|
cannam@167
|
2050 E TdI, TdJ, T5C, T5F;
|
cannam@167
|
2051 TdI = TaT - TaS;
|
cannam@167
|
2052 TdJ = KP707106781 * (Tdx - Tdw);
|
cannam@167
|
2053 TdK = TdI + TdJ;
|
cannam@167
|
2054 TeP = TdJ - TdI;
|
cannam@167
|
2055 T5C = T5y - T5B;
|
cannam@167
|
2056 T5F = T5D - T5E;
|
cannam@167
|
2057 T5G = T5C + T5F;
|
cannam@167
|
2058 T6P = T5F - T5C;
|
cannam@167
|
2059 }
|
cannam@167
|
2060 {
|
cannam@167
|
2061 E T8b, T8c, TaR, TaU;
|
cannam@167
|
2062 T8b = T4z - T4G;
|
cannam@167
|
2063 T8c = T5E + T5D;
|
cannam@167
|
2064 T8d = T8b + T8c;
|
cannam@167
|
2065 T9p = T8b - T8c;
|
cannam@167
|
2066 TaR = TaP + TaQ;
|
cannam@167
|
2067 TaU = TaS + TaT;
|
cannam@167
|
2068 TaV = TaR - TaU;
|
cannam@167
|
2069 Tc3 = TaR + TaU;
|
cannam@167
|
2070 }
|
cannam@167
|
2071 {
|
cannam@167
|
2072 E Tbe, Tbh, T8m, T8n;
|
cannam@167
|
2073 Tbe = Tbc + Tbd;
|
cannam@167
|
2074 Tbh = Tbf + Tbg;
|
cannam@167
|
2075 Tbi = Tbe - Tbh;
|
cannam@167
|
2076 Tc4 = Tbh + Tbe;
|
cannam@167
|
2077 T8m = T5B + T5y;
|
cannam@167
|
2078 T8n = T4V - T4O;
|
cannam@167
|
2079 T8o = T8m + T8n;
|
cannam@167
|
2080 T9t = T8n - T8m;
|
cannam@167
|
2081 }
|
cannam@167
|
2082 }
|
cannam@167
|
2083 {
|
cannam@167
|
2084 E T3k, Tai, T4m, Taj, T3r, Tam, T4j, Tal, TaI, TaJ, T3z, Tdb, T4p, TaF, TaG;
|
cannam@167
|
2085 E T3G, Tdc, T4o;
|
cannam@167
|
2086 {
|
cannam@167
|
2087 E T3i, T3j, T4k, T4l;
|
cannam@167
|
2088 T3i = R1[0];
|
cannam@167
|
2089 T3j = R1[WS(rs, 32)];
|
cannam@167
|
2090 T3k = T3i - T3j;
|
cannam@167
|
2091 Tai = T3i + T3j;
|
cannam@167
|
2092 T4k = R1[WS(rs, 16)];
|
cannam@167
|
2093 T4l = R1[WS(rs, 48)];
|
cannam@167
|
2094 T4m = T4k - T4l;
|
cannam@167
|
2095 Taj = T4k + T4l;
|
cannam@167
|
2096 }
|
cannam@167
|
2097 {
|
cannam@167
|
2098 E T3l, T3m, T3n, T3o, T3p, T3q;
|
cannam@167
|
2099 T3l = R1[WS(rs, 8)];
|
cannam@167
|
2100 T3m = R1[WS(rs, 40)];
|
cannam@167
|
2101 T3n = T3l - T3m;
|
cannam@167
|
2102 T3o = R1[WS(rs, 56)];
|
cannam@167
|
2103 T3p = R1[WS(rs, 24)];
|
cannam@167
|
2104 T3q = T3o - T3p;
|
cannam@167
|
2105 T3r = KP707106781 * (T3n + T3q);
|
cannam@167
|
2106 Tam = T3o + T3p;
|
cannam@167
|
2107 T4j = KP707106781 * (T3q - T3n);
|
cannam@167
|
2108 Tal = T3l + T3m;
|
cannam@167
|
2109 }
|
cannam@167
|
2110 {
|
cannam@167
|
2111 E T3v, T3y, T3C, T3F;
|
cannam@167
|
2112 {
|
cannam@167
|
2113 E T3t, T3u, T3w, T3x;
|
cannam@167
|
2114 T3t = R1[WS(rs, 4)];
|
cannam@167
|
2115 T3u = R1[WS(rs, 36)];
|
cannam@167
|
2116 T3v = T3t - T3u;
|
cannam@167
|
2117 TaI = T3t + T3u;
|
cannam@167
|
2118 T3w = R1[WS(rs, 20)];
|
cannam@167
|
2119 T3x = R1[WS(rs, 52)];
|
cannam@167
|
2120 T3y = T3w - T3x;
|
cannam@167
|
2121 TaJ = T3w + T3x;
|
cannam@167
|
2122 }
|
cannam@167
|
2123 T3z = FNMS(KP382683432, T3y, KP923879532 * T3v);
|
cannam@167
|
2124 Tdb = TaI - TaJ;
|
cannam@167
|
2125 T4p = FMA(KP382683432, T3v, KP923879532 * T3y);
|
cannam@167
|
2126 {
|
cannam@167
|
2127 E T3A, T3B, T3D, T3E;
|
cannam@167
|
2128 T3A = R1[WS(rs, 60)];
|
cannam@167
|
2129 T3B = R1[WS(rs, 28)];
|
cannam@167
|
2130 T3C = T3A - T3B;
|
cannam@167
|
2131 TaF = T3A + T3B;
|
cannam@167
|
2132 T3D = R1[WS(rs, 12)];
|
cannam@167
|
2133 T3E = R1[WS(rs, 44)];
|
cannam@167
|
2134 T3F = T3D - T3E;
|
cannam@167
|
2135 TaG = T3D + T3E;
|
cannam@167
|
2136 }
|
cannam@167
|
2137 T3G = FMA(KP923879532, T3C, KP382683432 * T3F);
|
cannam@167
|
2138 Tdc = TaF - TaG;
|
cannam@167
|
2139 T4o = FNMS(KP923879532, T3F, KP382683432 * T3C);
|
cannam@167
|
2140 }
|
cannam@167
|
2141 {
|
cannam@167
|
2142 E T3s, T3H, Tda, Tdd;
|
cannam@167
|
2143 T3s = T3k + T3r;
|
cannam@167
|
2144 T3H = T3z + T3G;
|
cannam@167
|
2145 T3I = T3s + T3H;
|
cannam@167
|
2146 T6H = T3s - T3H;
|
cannam@167
|
2147 Tda = Tai - Taj;
|
cannam@167
|
2148 Tdd = KP707106781 * (Tdb + Tdc);
|
cannam@167
|
2149 Tde = Tda + Tdd;
|
cannam@167
|
2150 TeH = Tda - Tdd;
|
cannam@167
|
2151 }
|
cannam@167
|
2152 {
|
cannam@167
|
2153 E Tdn, Tdo, T4n, T4q;
|
cannam@167
|
2154 Tdn = Tam - Tal;
|
cannam@167
|
2155 Tdo = KP707106781 * (Tdc - Tdb);
|
cannam@167
|
2156 Tdp = Tdn + Tdo;
|
cannam@167
|
2157 TeF = Tdo - Tdn;
|
cannam@167
|
2158 T4n = T4j - T4m;
|
cannam@167
|
2159 T4q = T4o - T4p;
|
cannam@167
|
2160 T4r = T4n + T4q;
|
cannam@167
|
2161 T6F = T4q - T4n;
|
cannam@167
|
2162 }
|
cannam@167
|
2163 {
|
cannam@167
|
2164 E T7S, T7T, Tak, Tan;
|
cannam@167
|
2165 T7S = T3k - T3r;
|
cannam@167
|
2166 T7T = T4p + T4o;
|
cannam@167
|
2167 T7U = T7S + T7T;
|
cannam@167
|
2168 T9l = T7S - T7T;
|
cannam@167
|
2169 Tak = Tai + Taj;
|
cannam@167
|
2170 Tan = Tal + Tam;
|
cannam@167
|
2171 Tao = Tak - Tan;
|
cannam@167
|
2172 TbW = Tak + Tan;
|
cannam@167
|
2173 }
|
cannam@167
|
2174 {
|
cannam@167
|
2175 E TaH, TaK, T83, T84;
|
cannam@167
|
2176 TaH = TaF + TaG;
|
cannam@167
|
2177 TaK = TaI + TaJ;
|
cannam@167
|
2178 TaL = TaH - TaK;
|
cannam@167
|
2179 TbX = TaK + TaH;
|
cannam@167
|
2180 T83 = T4m + T4j;
|
cannam@167
|
2181 T84 = T3G - T3z;
|
cannam@167
|
2182 T85 = T83 + T84;
|
cannam@167
|
2183 T9j = T84 - T83;
|
cannam@167
|
2184 }
|
cannam@167
|
2185 }
|
cannam@167
|
2186 {
|
cannam@167
|
2187 E T1z, T2V, T1C, T39, T1G, T38, T1J, T2Y, T1O, T1R, T32, Td0, T3c, T1V, T1Y;
|
cannam@167
|
2188 E T35, Td1, T3b;
|
cannam@167
|
2189 {
|
cannam@167
|
2190 E T1x, T1y, T1A, T1B;
|
cannam@167
|
2191 T1x = R0[WS(rs, 63)];
|
cannam@167
|
2192 T1y = R0[WS(rs, 31)];
|
cannam@167
|
2193 T1z = T1x + T1y;
|
cannam@167
|
2194 T2V = T1x - T1y;
|
cannam@167
|
2195 T1A = R0[WS(rs, 15)];
|
cannam@167
|
2196 T1B = R0[WS(rs, 47)];
|
cannam@167
|
2197 T1C = T1A + T1B;
|
cannam@167
|
2198 T39 = T1A - T1B;
|
cannam@167
|
2199 }
|
cannam@167
|
2200 {
|
cannam@167
|
2201 E T1E, T1F, T2W, T1H, T1I, T2X;
|
cannam@167
|
2202 T1E = R0[WS(rs, 7)];
|
cannam@167
|
2203 T1F = R0[WS(rs, 39)];
|
cannam@167
|
2204 T2W = T1E - T1F;
|
cannam@167
|
2205 T1H = R0[WS(rs, 55)];
|
cannam@167
|
2206 T1I = R0[WS(rs, 23)];
|
cannam@167
|
2207 T2X = T1H - T1I;
|
cannam@167
|
2208 T1G = T1E + T1F;
|
cannam@167
|
2209 T38 = KP707106781 * (T2X - T2W);
|
cannam@167
|
2210 T1J = T1H + T1I;
|
cannam@167
|
2211 T2Y = KP707106781 * (T2W + T2X);
|
cannam@167
|
2212 }
|
cannam@167
|
2213 {
|
cannam@167
|
2214 E T30, T31, T33, T34;
|
cannam@167
|
2215 {
|
cannam@167
|
2216 E T1M, T1N, T1P, T1Q;
|
cannam@167
|
2217 T1M = R0[WS(rs, 3)];
|
cannam@167
|
2218 T1N = R0[WS(rs, 35)];
|
cannam@167
|
2219 T1O = T1M + T1N;
|
cannam@167
|
2220 T30 = T1M - T1N;
|
cannam@167
|
2221 T1P = R0[WS(rs, 19)];
|
cannam@167
|
2222 T1Q = R0[WS(rs, 51)];
|
cannam@167
|
2223 T1R = T1P + T1Q;
|
cannam@167
|
2224 T31 = T1P - T1Q;
|
cannam@167
|
2225 }
|
cannam@167
|
2226 T32 = FNMS(KP382683432, T31, KP923879532 * T30);
|
cannam@167
|
2227 Td0 = T1O - T1R;
|
cannam@167
|
2228 T3c = FMA(KP382683432, T30, KP923879532 * T31);
|
cannam@167
|
2229 {
|
cannam@167
|
2230 E T1T, T1U, T1W, T1X;
|
cannam@167
|
2231 T1T = R0[WS(rs, 59)];
|
cannam@167
|
2232 T1U = R0[WS(rs, 27)];
|
cannam@167
|
2233 T1V = T1T + T1U;
|
cannam@167
|
2234 T33 = T1T - T1U;
|
cannam@167
|
2235 T1W = R0[WS(rs, 11)];
|
cannam@167
|
2236 T1X = R0[WS(rs, 43)];
|
cannam@167
|
2237 T1Y = T1W + T1X;
|
cannam@167
|
2238 T34 = T1W - T1X;
|
cannam@167
|
2239 }
|
cannam@167
|
2240 T35 = FMA(KP923879532, T33, KP382683432 * T34);
|
cannam@167
|
2241 Td1 = T1V - T1Y;
|
cannam@167
|
2242 T3b = FNMS(KP923879532, T34, KP382683432 * T33);
|
cannam@167
|
2243 }
|
cannam@167
|
2244 {
|
cannam@167
|
2245 E T1D, T1K, TcZ, Td2;
|
cannam@167
|
2246 T1D = T1z + T1C;
|
cannam@167
|
2247 T1K = T1G + T1J;
|
cannam@167
|
2248 T1L = T1D + T1K;
|
cannam@167
|
2249 Tad = T1D - T1K;
|
cannam@167
|
2250 TcZ = T1z - T1C;
|
cannam@167
|
2251 Td2 = KP707106781 * (Td0 + Td1);
|
cannam@167
|
2252 Td3 = TcZ + Td2;
|
cannam@167
|
2253 Tew = TcZ - Td2;
|
cannam@167
|
2254 }
|
cannam@167
|
2255 {
|
cannam@167
|
2256 E Td4, Td5, T1S, T1Z;
|
cannam@167
|
2257 Td4 = T1J - T1G;
|
cannam@167
|
2258 Td5 = KP707106781 * (Td1 - Td0);
|
cannam@167
|
2259 Td6 = Td4 + Td5;
|
cannam@167
|
2260 Tex = Td5 - Td4;
|
cannam@167
|
2261 T1S = T1O + T1R;
|
cannam@167
|
2262 T1Z = T1V + T1Y;
|
cannam@167
|
2263 T20 = T1S + T1Z;
|
cannam@167
|
2264 Tae = T1Z - T1S;
|
cannam@167
|
2265 }
|
cannam@167
|
2266 {
|
cannam@167
|
2267 E T2Z, T36, T7J, T7K;
|
cannam@167
|
2268 T2Z = T2V + T2Y;
|
cannam@167
|
2269 T36 = T32 + T35;
|
cannam@167
|
2270 T37 = T2Z + T36;
|
cannam@167
|
2271 T6x = T2Z - T36;
|
cannam@167
|
2272 T7J = T2V - T2Y;
|
cannam@167
|
2273 T7K = T3c + T3b;
|
cannam@167
|
2274 T7L = T7J + T7K;
|
cannam@167
|
2275 T9a = T7J - T7K;
|
cannam@167
|
2276 }
|
cannam@167
|
2277 {
|
cannam@167
|
2278 E T7M, T7N, T3a, T3d;
|
cannam@167
|
2279 T7M = T39 + T38;
|
cannam@167
|
2280 T7N = T35 - T32;
|
cannam@167
|
2281 T7O = T7M + T7N;
|
cannam@167
|
2282 T9b = T7N - T7M;
|
cannam@167
|
2283 T3a = T38 - T39;
|
cannam@167
|
2284 T3d = T3b - T3c;
|
cannam@167
|
2285 T3e = T3a + T3d;
|
cannam@167
|
2286 T6w = T3d - T3a;
|
cannam@167
|
2287 }
|
cannam@167
|
2288 }
|
cannam@167
|
2289 {
|
cannam@167
|
2290 E T3L, Tdf, T3X, Tar, T42, Tdi, T4e, Tay, T3S, Tdg, T3U, Tau, T49, Tdj, T4b;
|
cannam@167
|
2291 E TaB, Tdh, Tdk;
|
cannam@167
|
2292 {
|
cannam@167
|
2293 E T3J, T3K, Tap, T3V, T3W, Taq;
|
cannam@167
|
2294 T3J = R1[WS(rs, 2)];
|
cannam@167
|
2295 T3K = R1[WS(rs, 34)];
|
cannam@167
|
2296 Tap = T3J + T3K;
|
cannam@167
|
2297 T3V = R1[WS(rs, 18)];
|
cannam@167
|
2298 T3W = R1[WS(rs, 50)];
|
cannam@167
|
2299 Taq = T3V + T3W;
|
cannam@167
|
2300 T3L = T3J - T3K;
|
cannam@167
|
2301 Tdf = Tap - Taq;
|
cannam@167
|
2302 T3X = T3V - T3W;
|
cannam@167
|
2303 Tar = Tap + Taq;
|
cannam@167
|
2304 }
|
cannam@167
|
2305 {
|
cannam@167
|
2306 E T40, T41, Taw, T4c, T4d, Tax;
|
cannam@167
|
2307 T40 = R1[WS(rs, 62)];
|
cannam@167
|
2308 T41 = R1[WS(rs, 30)];
|
cannam@167
|
2309 Taw = T40 + T41;
|
cannam@167
|
2310 T4c = R1[WS(rs, 14)];
|
cannam@167
|
2311 T4d = R1[WS(rs, 46)];
|
cannam@167
|
2312 Tax = T4c + T4d;
|
cannam@167
|
2313 T42 = T40 - T41;
|
cannam@167
|
2314 Tdi = Taw - Tax;
|
cannam@167
|
2315 T4e = T4c - T4d;
|
cannam@167
|
2316 Tay = Taw + Tax;
|
cannam@167
|
2317 }
|
cannam@167
|
2318 {
|
cannam@167
|
2319 E T3O, Tas, T3R, Tat;
|
cannam@167
|
2320 {
|
cannam@167
|
2321 E T3M, T3N, T3P, T3Q;
|
cannam@167
|
2322 T3M = R1[WS(rs, 10)];
|
cannam@167
|
2323 T3N = R1[WS(rs, 42)];
|
cannam@167
|
2324 T3O = T3M - T3N;
|
cannam@167
|
2325 Tas = T3M + T3N;
|
cannam@167
|
2326 T3P = R1[WS(rs, 58)];
|
cannam@167
|
2327 T3Q = R1[WS(rs, 26)];
|
cannam@167
|
2328 T3R = T3P - T3Q;
|
cannam@167
|
2329 Tat = T3P + T3Q;
|
cannam@167
|
2330 }
|
cannam@167
|
2331 T3S = KP707106781 * (T3O + T3R);
|
cannam@167
|
2332 Tdg = Tat - Tas;
|
cannam@167
|
2333 T3U = KP707106781 * (T3R - T3O);
|
cannam@167
|
2334 Tau = Tas + Tat;
|
cannam@167
|
2335 }
|
cannam@167
|
2336 {
|
cannam@167
|
2337 E T45, Taz, T48, TaA;
|
cannam@167
|
2338 {
|
cannam@167
|
2339 E T43, T44, T46, T47;
|
cannam@167
|
2340 T43 = R1[WS(rs, 6)];
|
cannam@167
|
2341 T44 = R1[WS(rs, 38)];
|
cannam@167
|
2342 T45 = T43 - T44;
|
cannam@167
|
2343 Taz = T43 + T44;
|
cannam@167
|
2344 T46 = R1[WS(rs, 54)];
|
cannam@167
|
2345 T47 = R1[WS(rs, 22)];
|
cannam@167
|
2346 T48 = T46 - T47;
|
cannam@167
|
2347 TaA = T46 + T47;
|
cannam@167
|
2348 }
|
cannam@167
|
2349 T49 = KP707106781 * (T45 + T48);
|
cannam@167
|
2350 Tdj = TaA - Taz;
|
cannam@167
|
2351 T4b = KP707106781 * (T48 - T45);
|
cannam@167
|
2352 TaB = Taz + TaA;
|
cannam@167
|
2353 }
|
cannam@167
|
2354 TbZ = Tar + Tau;
|
cannam@167
|
2355 Tc0 = Tay + TaB;
|
cannam@167
|
2356 {
|
cannam@167
|
2357 E T3T, T3Y, Tdq, Tdr;
|
cannam@167
|
2358 T3T = T3L + T3S;
|
cannam@167
|
2359 T3Y = T3U - T3X;
|
cannam@167
|
2360 T3Z = FMA(KP980785280, T3T, KP195090322 * T3Y);
|
cannam@167
|
2361 T4s = FNMS(KP195090322, T3T, KP980785280 * T3Y);
|
cannam@167
|
2362 Tdq = FNMS(KP382683432, Tdf, KP923879532 * Tdg);
|
cannam@167
|
2363 Tdr = FMA(KP382683432, Tdi, KP923879532 * Tdj);
|
cannam@167
|
2364 Tds = Tdq + Tdr;
|
cannam@167
|
2365 TeI = Tdr - Tdq;
|
cannam@167
|
2366 }
|
cannam@167
|
2367 {
|
cannam@167
|
2368 E T4a, T4f, T7Y, T7Z;
|
cannam@167
|
2369 T4a = T42 + T49;
|
cannam@167
|
2370 T4f = T4b - T4e;
|
cannam@167
|
2371 T4g = FNMS(KP195090322, T4f, KP980785280 * T4a);
|
cannam@167
|
2372 T4t = FMA(KP195090322, T4a, KP980785280 * T4f);
|
cannam@167
|
2373 T7Y = T42 - T49;
|
cannam@167
|
2374 T7Z = T4e + T4b;
|
cannam@167
|
2375 T80 = FNMS(KP555570233, T7Z, KP831469612 * T7Y);
|
cannam@167
|
2376 T87 = FMA(KP555570233, T7Y, KP831469612 * T7Z);
|
cannam@167
|
2377 }
|
cannam@167
|
2378 Tdh = FMA(KP923879532, Tdf, KP382683432 * Tdg);
|
cannam@167
|
2379 Tdk = FNMS(KP382683432, Tdj, KP923879532 * Tdi);
|
cannam@167
|
2380 Tdl = Tdh + Tdk;
|
cannam@167
|
2381 TeE = Tdk - Tdh;
|
cannam@167
|
2382 {
|
cannam@167
|
2383 E T7V, T7W, Tav, TaC;
|
cannam@167
|
2384 T7V = T3L - T3S;
|
cannam@167
|
2385 T7W = T3X + T3U;
|
cannam@167
|
2386 T7X = FMA(KP831469612, T7V, KP555570233 * T7W);
|
cannam@167
|
2387 T86 = FNMS(KP555570233, T7V, KP831469612 * T7W);
|
cannam@167
|
2388 Tav = Tar - Tau;
|
cannam@167
|
2389 TaC = Tay - TaB;
|
cannam@167
|
2390 TaD = KP707106781 * (Tav + TaC);
|
cannam@167
|
2391 TaM = KP707106781 * (TaC - Tav);
|
cannam@167
|
2392 }
|
cannam@167
|
2393 }
|
cannam@167
|
2394 {
|
cannam@167
|
2395 E T50, TdA, T5c, TaY, T5h, TdD, T5t, Tb5, T57, TdB, T59, Tb1, T5o, TdE, T5q;
|
cannam@167
|
2396 E Tb8, TdC, TdF;
|
cannam@167
|
2397 {
|
cannam@167
|
2398 E T4Y, T4Z, TaW, T5a, T5b, TaX;
|
cannam@167
|
2399 T4Y = R1[WS(rs, 1)];
|
cannam@167
|
2400 T4Z = R1[WS(rs, 33)];
|
cannam@167
|
2401 TaW = T4Y + T4Z;
|
cannam@167
|
2402 T5a = R1[WS(rs, 17)];
|
cannam@167
|
2403 T5b = R1[WS(rs, 49)];
|
cannam@167
|
2404 TaX = T5a + T5b;
|
cannam@167
|
2405 T50 = T4Y - T4Z;
|
cannam@167
|
2406 TdA = TaW - TaX;
|
cannam@167
|
2407 T5c = T5a - T5b;
|
cannam@167
|
2408 TaY = TaW + TaX;
|
cannam@167
|
2409 }
|
cannam@167
|
2410 {
|
cannam@167
|
2411 E T5f, T5g, Tb3, T5r, T5s, Tb4;
|
cannam@167
|
2412 T5f = R1[WS(rs, 61)];
|
cannam@167
|
2413 T5g = R1[WS(rs, 29)];
|
cannam@167
|
2414 Tb3 = T5f + T5g;
|
cannam@167
|
2415 T5r = R1[WS(rs, 13)];
|
cannam@167
|
2416 T5s = R1[WS(rs, 45)];
|
cannam@167
|
2417 Tb4 = T5r + T5s;
|
cannam@167
|
2418 T5h = T5f - T5g;
|
cannam@167
|
2419 TdD = Tb3 - Tb4;
|
cannam@167
|
2420 T5t = T5r - T5s;
|
cannam@167
|
2421 Tb5 = Tb3 + Tb4;
|
cannam@167
|
2422 }
|
cannam@167
|
2423 {
|
cannam@167
|
2424 E T53, TaZ, T56, Tb0;
|
cannam@167
|
2425 {
|
cannam@167
|
2426 E T51, T52, T54, T55;
|
cannam@167
|
2427 T51 = R1[WS(rs, 9)];
|
cannam@167
|
2428 T52 = R1[WS(rs, 41)];
|
cannam@167
|
2429 T53 = T51 - T52;
|
cannam@167
|
2430 TaZ = T51 + T52;
|
cannam@167
|
2431 T54 = R1[WS(rs, 57)];
|
cannam@167
|
2432 T55 = R1[WS(rs, 25)];
|
cannam@167
|
2433 T56 = T54 - T55;
|
cannam@167
|
2434 Tb0 = T54 + T55;
|
cannam@167
|
2435 }
|
cannam@167
|
2436 T57 = KP707106781 * (T53 + T56);
|
cannam@167
|
2437 TdB = Tb0 - TaZ;
|
cannam@167
|
2438 T59 = KP707106781 * (T56 - T53);
|
cannam@167
|
2439 Tb1 = TaZ + Tb0;
|
cannam@167
|
2440 }
|
cannam@167
|
2441 {
|
cannam@167
|
2442 E T5k, Tb6, T5n, Tb7;
|
cannam@167
|
2443 {
|
cannam@167
|
2444 E T5i, T5j, T5l, T5m;
|
cannam@167
|
2445 T5i = R1[WS(rs, 5)];
|
cannam@167
|
2446 T5j = R1[WS(rs, 37)];
|
cannam@167
|
2447 T5k = T5i - T5j;
|
cannam@167
|
2448 Tb6 = T5i + T5j;
|
cannam@167
|
2449 T5l = R1[WS(rs, 53)];
|
cannam@167
|
2450 T5m = R1[WS(rs, 21)];
|
cannam@167
|
2451 T5n = T5l - T5m;
|
cannam@167
|
2452 Tb7 = T5l + T5m;
|
cannam@167
|
2453 }
|
cannam@167
|
2454 T5o = KP707106781 * (T5k + T5n);
|
cannam@167
|
2455 TdE = Tb7 - Tb6;
|
cannam@167
|
2456 T5q = KP707106781 * (T5n - T5k);
|
cannam@167
|
2457 Tb8 = Tb6 + Tb7;
|
cannam@167
|
2458 }
|
cannam@167
|
2459 Tc6 = TaY + Tb1;
|
cannam@167
|
2460 Tc7 = Tb5 + Tb8;
|
cannam@167
|
2461 {
|
cannam@167
|
2462 E T58, T5d, TdL, TdM;
|
cannam@167
|
2463 T58 = T50 + T57;
|
cannam@167
|
2464 T5d = T59 - T5c;
|
cannam@167
|
2465 T5e = FMA(KP980785280, T58, KP195090322 * T5d);
|
cannam@167
|
2466 T5H = FNMS(KP195090322, T58, KP980785280 * T5d);
|
cannam@167
|
2467 TdL = FNMS(KP382683432, TdA, KP923879532 * TdB);
|
cannam@167
|
2468 TdM = FMA(KP382683432, TdD, KP923879532 * TdE);
|
cannam@167
|
2469 TdN = TdL + TdM;
|
cannam@167
|
2470 TeM = TdM - TdL;
|
cannam@167
|
2471 }
|
cannam@167
|
2472 {
|
cannam@167
|
2473 E T5p, T5u, T8h, T8i;
|
cannam@167
|
2474 T5p = T5h + T5o;
|
cannam@167
|
2475 T5u = T5q - T5t;
|
cannam@167
|
2476 T5v = FNMS(KP195090322, T5u, KP980785280 * T5p);
|
cannam@167
|
2477 T5I = FMA(KP195090322, T5p, KP980785280 * T5u);
|
cannam@167
|
2478 T8h = T5h - T5o;
|
cannam@167
|
2479 T8i = T5t + T5q;
|
cannam@167
|
2480 T8j = FNMS(KP555570233, T8i, KP831469612 * T8h);
|
cannam@167
|
2481 T8q = FMA(KP555570233, T8h, KP831469612 * T8i);
|
cannam@167
|
2482 }
|
cannam@167
|
2483 TdC = FMA(KP923879532, TdA, KP382683432 * TdB);
|
cannam@167
|
2484 TdF = FNMS(KP382683432, TdE, KP923879532 * TdD);
|
cannam@167
|
2485 TdG = TdC + TdF;
|
cannam@167
|
2486 TeO = TdF - TdC;
|
cannam@167
|
2487 {
|
cannam@167
|
2488 E T8e, T8f, Tb2, Tb9;
|
cannam@167
|
2489 T8e = T50 - T57;
|
cannam@167
|
2490 T8f = T5c + T59;
|
cannam@167
|
2491 T8g = FMA(KP831469612, T8e, KP555570233 * T8f);
|
cannam@167
|
2492 T8p = FNMS(KP555570233, T8e, KP831469612 * T8f);
|
cannam@167
|
2493 Tb2 = TaY - Tb1;
|
cannam@167
|
2494 Tb9 = Tb5 - Tb8;
|
cannam@167
|
2495 Tba = KP707106781 * (Tb2 + Tb9);
|
cannam@167
|
2496 Tbj = KP707106781 * (Tb9 - Tb2);
|
cannam@167
|
2497 }
|
cannam@167
|
2498 }
|
cannam@167
|
2499 {
|
cannam@167
|
2500 E T11, TbV, Tc9, Tcf, T22, Tcb, Tc2, Tce;
|
cannam@167
|
2501 {
|
cannam@167
|
2502 E Tv, T10, Tc5, Tc8;
|
cannam@167
|
2503 Tv = Tf + Tu;
|
cannam@167
|
2504 T10 = TK + TZ;
|
cannam@167
|
2505 T11 = Tv + T10;
|
cannam@167
|
2506 TbV = Tv - T10;
|
cannam@167
|
2507 Tc5 = Tc3 + Tc4;
|
cannam@167
|
2508 Tc8 = Tc6 + Tc7;
|
cannam@167
|
2509 Tc9 = Tc5 - Tc8;
|
cannam@167
|
2510 Tcf = Tc5 + Tc8;
|
cannam@167
|
2511 }
|
cannam@167
|
2512 {
|
cannam@167
|
2513 E T1w, T21, TbY, Tc1;
|
cannam@167
|
2514 T1w = T1g + T1v;
|
cannam@167
|
2515 T21 = T1L + T20;
|
cannam@167
|
2516 T22 = T1w + T21;
|
cannam@167
|
2517 Tcb = T21 - T1w;
|
cannam@167
|
2518 TbY = TbW + TbX;
|
cannam@167
|
2519 Tc1 = TbZ + Tc0;
|
cannam@167
|
2520 Tc2 = TbY - Tc1;
|
cannam@167
|
2521 Tce = TbY + Tc1;
|
cannam@167
|
2522 }
|
cannam@167
|
2523 Cr[WS(csr, 32)] = T11 - T22;
|
cannam@167
|
2524 Ci[WS(csi, 32)] = Tcf - Tce;
|
cannam@167
|
2525 {
|
cannam@167
|
2526 E Tca, Tcc, Tcd, Tcg;
|
cannam@167
|
2527 Tca = KP707106781 * (Tc2 + Tc9);
|
cannam@167
|
2528 Cr[WS(csr, 48)] = TbV - Tca;
|
cannam@167
|
2529 Cr[WS(csr, 16)] = TbV + Tca;
|
cannam@167
|
2530 Tcc = KP707106781 * (Tc9 - Tc2);
|
cannam@167
|
2531 Ci[WS(csi, 16)] = Tcb + Tcc;
|
cannam@167
|
2532 Ci[WS(csi, 48)] = Tcc - Tcb;
|
cannam@167
|
2533 Tcd = T11 + T22;
|
cannam@167
|
2534 Tcg = Tce + Tcf;
|
cannam@167
|
2535 Cr[WS(csr, 64)] = Tcd - Tcg;
|
cannam@167
|
2536 Cr[0] = Tcd + Tcg;
|
cannam@167
|
2537 }
|
cannam@167
|
2538 }
|
cannam@167
|
2539 {
|
cannam@167
|
2540 E Tch, Tcu, Tck, Tct, Tco, Tcy, Tcr, Tcz, Tci, Tcj;
|
cannam@167
|
2541 Tch = Tf - Tu;
|
cannam@167
|
2542 Tcu = TZ - TK;
|
cannam@167
|
2543 Tci = T1g - T1v;
|
cannam@167
|
2544 Tcj = T1L - T20;
|
cannam@167
|
2545 Tck = KP707106781 * (Tci + Tcj);
|
cannam@167
|
2546 Tct = KP707106781 * (Tcj - Tci);
|
cannam@167
|
2547 {
|
cannam@167
|
2548 E Tcm, Tcn, Tcp, Tcq;
|
cannam@167
|
2549 Tcm = TbW - TbX;
|
cannam@167
|
2550 Tcn = Tc0 - TbZ;
|
cannam@167
|
2551 Tco = FMA(KP923879532, Tcm, KP382683432 * Tcn);
|
cannam@167
|
2552 Tcy = FNMS(KP382683432, Tcm, KP923879532 * Tcn);
|
cannam@167
|
2553 Tcp = Tc3 - Tc4;
|
cannam@167
|
2554 Tcq = Tc7 - Tc6;
|
cannam@167
|
2555 Tcr = FNMS(KP382683432, Tcq, KP923879532 * Tcp);
|
cannam@167
|
2556 Tcz = FMA(KP382683432, Tcp, KP923879532 * Tcq);
|
cannam@167
|
2557 }
|
cannam@167
|
2558 {
|
cannam@167
|
2559 E Tcl, Tcs, Tcx, TcA;
|
cannam@167
|
2560 Tcl = Tch + Tck;
|
cannam@167
|
2561 Tcs = Tco + Tcr;
|
cannam@167
|
2562 Cr[WS(csr, 56)] = Tcl - Tcs;
|
cannam@167
|
2563 Cr[WS(csr, 8)] = Tcl + Tcs;
|
cannam@167
|
2564 Tcx = Tcu + Tct;
|
cannam@167
|
2565 TcA = Tcy + Tcz;
|
cannam@167
|
2566 Ci[WS(csi, 8)] = Tcx + TcA;
|
cannam@167
|
2567 Ci[WS(csi, 56)] = TcA - Tcx;
|
cannam@167
|
2568 }
|
cannam@167
|
2569 {
|
cannam@167
|
2570 E Tcv, Tcw, TcB, TcC;
|
cannam@167
|
2571 Tcv = Tct - Tcu;
|
cannam@167
|
2572 Tcw = Tcr - Tco;
|
cannam@167
|
2573 Ci[WS(csi, 24)] = Tcv + Tcw;
|
cannam@167
|
2574 Ci[WS(csi, 40)] = Tcw - Tcv;
|
cannam@167
|
2575 TcB = Tch - Tck;
|
cannam@167
|
2576 TcC = Tcz - Tcy;
|
cannam@167
|
2577 Cr[WS(csr, 40)] = TcB - TcC;
|
cannam@167
|
2578 Cr[WS(csr, 24)] = TcB + TcC;
|
cannam@167
|
2579 }
|
cannam@167
|
2580 }
|
cannam@167
|
2581 {
|
cannam@167
|
2582 E Ta9, TbB, Tbs, TbM, Tag, TbL, TbJ, TbR, TaO, Tbw, Tbp, TbC, TbG, TbQ, Tbl;
|
cannam@167
|
2583 E Tbx, Ta8, Tbr;
|
cannam@167
|
2584 Ta8 = KP707106781 * (Ta6 + Ta7);
|
cannam@167
|
2585 Ta9 = Ta5 + Ta8;
|
cannam@167
|
2586 TbB = Ta5 - Ta8;
|
cannam@167
|
2587 Tbr = KP707106781 * (Ta7 - Ta6);
|
cannam@167
|
2588 Tbs = Tbq + Tbr;
|
cannam@167
|
2589 TbM = Tbr - Tbq;
|
cannam@167
|
2590 {
|
cannam@167
|
2591 E Tac, Taf, TbH, TbI;
|
cannam@167
|
2592 Tac = FMA(KP923879532, Taa, KP382683432 * Tab);
|
cannam@167
|
2593 Taf = FNMS(KP382683432, Tae, KP923879532 * Tad);
|
cannam@167
|
2594 Tag = Tac + Taf;
|
cannam@167
|
2595 TbL = Taf - Tac;
|
cannam@167
|
2596 TbH = TaV - Tba;
|
cannam@167
|
2597 TbI = Tbj - Tbi;
|
cannam@167
|
2598 TbJ = FNMS(KP555570233, TbI, KP831469612 * TbH);
|
cannam@167
|
2599 TbR = FMA(KP555570233, TbH, KP831469612 * TbI);
|
cannam@167
|
2600 }
|
cannam@167
|
2601 {
|
cannam@167
|
2602 E TaE, TaN, Tbn, Tbo;
|
cannam@167
|
2603 TaE = Tao + TaD;
|
cannam@167
|
2604 TaN = TaL + TaM;
|
cannam@167
|
2605 TaO = FMA(KP980785280, TaE, KP195090322 * TaN);
|
cannam@167
|
2606 Tbw = FNMS(KP195090322, TaE, KP980785280 * TaN);
|
cannam@167
|
2607 Tbn = FNMS(KP382683432, Taa, KP923879532 * Tab);
|
cannam@167
|
2608 Tbo = FMA(KP382683432, Tad, KP923879532 * Tae);
|
cannam@167
|
2609 Tbp = Tbn + Tbo;
|
cannam@167
|
2610 TbC = Tbo - Tbn;
|
cannam@167
|
2611 }
|
cannam@167
|
2612 {
|
cannam@167
|
2613 E TbE, TbF, Tbb, Tbk;
|
cannam@167
|
2614 TbE = Tao - TaD;
|
cannam@167
|
2615 TbF = TaM - TaL;
|
cannam@167
|
2616 TbG = FMA(KP831469612, TbE, KP555570233 * TbF);
|
cannam@167
|
2617 TbQ = FNMS(KP555570233, TbE, KP831469612 * TbF);
|
cannam@167
|
2618 Tbb = TaV + Tba;
|
cannam@167
|
2619 Tbk = Tbi + Tbj;
|
cannam@167
|
2620 Tbl = FNMS(KP195090322, Tbk, KP980785280 * Tbb);
|
cannam@167
|
2621 Tbx = FMA(KP195090322, Tbb, KP980785280 * Tbk);
|
cannam@167
|
2622 }
|
cannam@167
|
2623 {
|
cannam@167
|
2624 E Tah, Tbm, Tbv, Tby;
|
cannam@167
|
2625 Tah = Ta9 + Tag;
|
cannam@167
|
2626 Tbm = TaO + Tbl;
|
cannam@167
|
2627 Cr[WS(csr, 60)] = Tah - Tbm;
|
cannam@167
|
2628 Cr[WS(csr, 4)] = Tah + Tbm;
|
cannam@167
|
2629 Tbv = Tbs + Tbp;
|
cannam@167
|
2630 Tby = Tbw + Tbx;
|
cannam@167
|
2631 Ci[WS(csi, 4)] = Tbv + Tby;
|
cannam@167
|
2632 Ci[WS(csi, 60)] = Tby - Tbv;
|
cannam@167
|
2633 }
|
cannam@167
|
2634 {
|
cannam@167
|
2635 E Tbt, Tbu, Tbz, TbA;
|
cannam@167
|
2636 Tbt = Tbp - Tbs;
|
cannam@167
|
2637 Tbu = Tbl - TaO;
|
cannam@167
|
2638 Ci[WS(csi, 28)] = Tbt + Tbu;
|
cannam@167
|
2639 Ci[WS(csi, 36)] = Tbu - Tbt;
|
cannam@167
|
2640 Tbz = Ta9 - Tag;
|
cannam@167
|
2641 TbA = Tbx - Tbw;
|
cannam@167
|
2642 Cr[WS(csr, 36)] = Tbz - TbA;
|
cannam@167
|
2643 Cr[WS(csr, 28)] = Tbz + TbA;
|
cannam@167
|
2644 }
|
cannam@167
|
2645 {
|
cannam@167
|
2646 E TbD, TbK, TbP, TbS;
|
cannam@167
|
2647 TbD = TbB + TbC;
|
cannam@167
|
2648 TbK = TbG + TbJ;
|
cannam@167
|
2649 Cr[WS(csr, 52)] = TbD - TbK;
|
cannam@167
|
2650 Cr[WS(csr, 12)] = TbD + TbK;
|
cannam@167
|
2651 TbP = TbM + TbL;
|
cannam@167
|
2652 TbS = TbQ + TbR;
|
cannam@167
|
2653 Ci[WS(csi, 12)] = TbP + TbS;
|
cannam@167
|
2654 Ci[WS(csi, 52)] = TbS - TbP;
|
cannam@167
|
2655 }
|
cannam@167
|
2656 {
|
cannam@167
|
2657 E TbN, TbO, TbT, TbU;
|
cannam@167
|
2658 TbN = TbL - TbM;
|
cannam@167
|
2659 TbO = TbJ - TbG;
|
cannam@167
|
2660 Ci[WS(csi, 20)] = TbN + TbO;
|
cannam@167
|
2661 Ci[WS(csi, 44)] = TbO - TbN;
|
cannam@167
|
2662 TbT = TbB - TbC;
|
cannam@167
|
2663 TbU = TbR - TbQ;
|
cannam@167
|
2664 Cr[WS(csr, 44)] = TbT - TbU;
|
cannam@167
|
2665 Cr[WS(csr, 20)] = TbT + TbU;
|
cannam@167
|
2666 }
|
cannam@167
|
2667 }
|
cannam@167
|
2668 {
|
cannam@167
|
2669 E Tev, Tf7, Tfc, Tfm, Tff, Tfn, TeC, Tfh, TeK, Tf2, TeV, Tf8, TeY, Tfi, TeR;
|
cannam@167
|
2670 E Tf3;
|
cannam@167
|
2671 {
|
cannam@167
|
2672 E Tet, Teu, Tfa, Tfb;
|
cannam@167
|
2673 Tet = TcD - TcG;
|
cannam@167
|
2674 Teu = TdY - TdX;
|
cannam@167
|
2675 Tev = Tet - Teu;
|
cannam@167
|
2676 Tf7 = Tet + Teu;
|
cannam@167
|
2677 Tfa = TeF + TeE;
|
cannam@167
|
2678 Tfb = TeH + TeI;
|
cannam@167
|
2679 Tfc = FMA(KP290284677, Tfa, KP956940335 * Tfb);
|
cannam@167
|
2680 Tfm = FNMS(KP290284677, Tfb, KP956940335 * Tfa);
|
cannam@167
|
2681 }
|
cannam@167
|
2682 {
|
cannam@167
|
2683 E Tfd, Tfe, Tey, TeB;
|
cannam@167
|
2684 Tfd = TeL + TeM;
|
cannam@167
|
2685 Tfe = TeP + TeO;
|
cannam@167
|
2686 Tff = FNMS(KP290284677, Tfe, KP956940335 * Tfd);
|
cannam@167
|
2687 Tfn = FMA(KP956940335, Tfe, KP290284677 * Tfd);
|
cannam@167
|
2688 Tey = FMA(KP555570233, Tew, KP831469612 * Tex);
|
cannam@167
|
2689 TeB = FNMS(KP555570233, TeA, KP831469612 * Tez);
|
cannam@167
|
2690 TeC = Tey - TeB;
|
cannam@167
|
2691 Tfh = TeB + Tey;
|
cannam@167
|
2692 }
|
cannam@167
|
2693 {
|
cannam@167
|
2694 E TeG, TeJ, TeT, TeU;
|
cannam@167
|
2695 TeG = TeE - TeF;
|
cannam@167
|
2696 TeJ = TeH - TeI;
|
cannam@167
|
2697 TeK = FMA(KP471396736, TeG, KP881921264 * TeJ);
|
cannam@167
|
2698 Tf2 = FNMS(KP471396736, TeJ, KP881921264 * TeG);
|
cannam@167
|
2699 TeT = FNMS(KP555570233, Tex, KP831469612 * Tew);
|
cannam@167
|
2700 TeU = FMA(KP831469612, TeA, KP555570233 * Tez);
|
cannam@167
|
2701 TeV = TeT - TeU;
|
cannam@167
|
2702 Tf8 = TeU + TeT;
|
cannam@167
|
2703 }
|
cannam@167
|
2704 {
|
cannam@167
|
2705 E TeW, TeX, TeN, TeQ;
|
cannam@167
|
2706 TeW = TcN - TcK;
|
cannam@167
|
2707 TeX = TdV - TdU;
|
cannam@167
|
2708 TeY = TeW - TeX;
|
cannam@167
|
2709 Tfi = TeX + TeW;
|
cannam@167
|
2710 TeN = TeL - TeM;
|
cannam@167
|
2711 TeQ = TeO - TeP;
|
cannam@167
|
2712 TeR = FNMS(KP471396736, TeQ, KP881921264 * TeN);
|
cannam@167
|
2713 Tf3 = FMA(KP881921264, TeQ, KP471396736 * TeN);
|
cannam@167
|
2714 }
|
cannam@167
|
2715 {
|
cannam@167
|
2716 E TeD, TeS, Tf1, Tf4;
|
cannam@167
|
2717 TeD = Tev + TeC;
|
cannam@167
|
2718 TeS = TeK + TeR;
|
cannam@167
|
2719 Cr[WS(csr, 54)] = TeD - TeS;
|
cannam@167
|
2720 Cr[WS(csr, 10)] = TeD + TeS;
|
cannam@167
|
2721 Tf1 = TeY + TeV;
|
cannam@167
|
2722 Tf4 = Tf2 + Tf3;
|
cannam@167
|
2723 Ci[WS(csi, 10)] = Tf1 + Tf4;
|
cannam@167
|
2724 Ci[WS(csi, 54)] = Tf4 - Tf1;
|
cannam@167
|
2725 }
|
cannam@167
|
2726 {
|
cannam@167
|
2727 E TeZ, Tf0, Tf5, Tf6;
|
cannam@167
|
2728 TeZ = TeV - TeY;
|
cannam@167
|
2729 Tf0 = TeR - TeK;
|
cannam@167
|
2730 Ci[WS(csi, 22)] = TeZ + Tf0;
|
cannam@167
|
2731 Ci[WS(csi, 42)] = Tf0 - TeZ;
|
cannam@167
|
2732 Tf5 = Tev - TeC;
|
cannam@167
|
2733 Tf6 = Tf3 - Tf2;
|
cannam@167
|
2734 Cr[WS(csr, 42)] = Tf5 - Tf6;
|
cannam@167
|
2735 Cr[WS(csr, 22)] = Tf5 + Tf6;
|
cannam@167
|
2736 }
|
cannam@167
|
2737 {
|
cannam@167
|
2738 E Tf9, Tfg, Tfl, Tfo;
|
cannam@167
|
2739 Tf9 = Tf7 + Tf8;
|
cannam@167
|
2740 Tfg = Tfc + Tff;
|
cannam@167
|
2741 Cr[WS(csr, 58)] = Tf9 - Tfg;
|
cannam@167
|
2742 Cr[WS(csr, 6)] = Tf9 + Tfg;
|
cannam@167
|
2743 Tfl = Tfi + Tfh;
|
cannam@167
|
2744 Tfo = Tfm + Tfn;
|
cannam@167
|
2745 Ci[WS(csi, 6)] = Tfl + Tfo;
|
cannam@167
|
2746 Ci[WS(csi, 58)] = Tfo - Tfl;
|
cannam@167
|
2747 }
|
cannam@167
|
2748 {
|
cannam@167
|
2749 E Tfj, Tfk, Tfp, Tfq;
|
cannam@167
|
2750 Tfj = Tfh - Tfi;
|
cannam@167
|
2751 Tfk = Tff - Tfc;
|
cannam@167
|
2752 Ci[WS(csi, 26)] = Tfj + Tfk;
|
cannam@167
|
2753 Ci[WS(csi, 38)] = Tfk - Tfj;
|
cannam@167
|
2754 Tfp = Tf7 - Tf8;
|
cannam@167
|
2755 Tfq = Tfn - Tfm;
|
cannam@167
|
2756 Cr[WS(csr, 38)] = Tfp - Tfq;
|
cannam@167
|
2757 Cr[WS(csr, 26)] = Tfp + Tfq;
|
cannam@167
|
2758 }
|
cannam@167
|
2759 }
|
cannam@167
|
2760 {
|
cannam@167
|
2761 E TcP, Te9, Tee, Teo, Teh, Tep, Td8, Tej, Tdu, Te4, TdT, Tea, Te0, Tek, TdP;
|
cannam@167
|
2762 E Te5;
|
cannam@167
|
2763 {
|
cannam@167
|
2764 E TcH, TcO, Tec, Ted;
|
cannam@167
|
2765 TcH = TcD + TcG;
|
cannam@167
|
2766 TcO = TcK + TcN;
|
cannam@167
|
2767 TcP = TcH + TcO;
|
cannam@167
|
2768 Te9 = TcH - TcO;
|
cannam@167
|
2769 Tec = Tde - Tdl;
|
cannam@167
|
2770 Ted = Tds - Tdp;
|
cannam@167
|
2771 Tee = FMA(KP773010453, Tec, KP634393284 * Ted);
|
cannam@167
|
2772 Teo = FNMS(KP634393284, Tec, KP773010453 * Ted);
|
cannam@167
|
2773 }
|
cannam@167
|
2774 {
|
cannam@167
|
2775 E Tef, Teg, TcY, Td7;
|
cannam@167
|
2776 Tef = Tdz - TdG;
|
cannam@167
|
2777 Teg = TdN - TdK;
|
cannam@167
|
2778 Teh = FNMS(KP634393284, Teg, KP773010453 * Tef);
|
cannam@167
|
2779 Tep = FMA(KP634393284, Tef, KP773010453 * Teg);
|
cannam@167
|
2780 TcY = FMA(KP980785280, TcU, KP195090322 * TcX);
|
cannam@167
|
2781 Td7 = FNMS(KP195090322, Td6, KP980785280 * Td3);
|
cannam@167
|
2782 Td8 = TcY + Td7;
|
cannam@167
|
2783 Tej = Td7 - TcY;
|
cannam@167
|
2784 }
|
cannam@167
|
2785 {
|
cannam@167
|
2786 E Tdm, Tdt, TdR, TdS;
|
cannam@167
|
2787 Tdm = Tde + Tdl;
|
cannam@167
|
2788 Tdt = Tdp + Tds;
|
cannam@167
|
2789 Tdu = FMA(KP995184726, Tdm, KP098017140 * Tdt);
|
cannam@167
|
2790 Te4 = FNMS(KP098017140, Tdm, KP995184726 * Tdt);
|
cannam@167
|
2791 TdR = FNMS(KP195090322, TcU, KP980785280 * TcX);
|
cannam@167
|
2792 TdS = FMA(KP195090322, Td3, KP980785280 * Td6);
|
cannam@167
|
2793 TdT = TdR + TdS;
|
cannam@167
|
2794 Tea = TdS - TdR;
|
cannam@167
|
2795 }
|
cannam@167
|
2796 {
|
cannam@167
|
2797 E TdW, TdZ, TdH, TdO;
|
cannam@167
|
2798 TdW = TdU + TdV;
|
cannam@167
|
2799 TdZ = TdX + TdY;
|
cannam@167
|
2800 Te0 = TdW + TdZ;
|
cannam@167
|
2801 Tek = TdZ - TdW;
|
cannam@167
|
2802 TdH = Tdz + TdG;
|
cannam@167
|
2803 TdO = TdK + TdN;
|
cannam@167
|
2804 TdP = FNMS(KP098017140, TdO, KP995184726 * TdH);
|
cannam@167
|
2805 Te5 = FMA(KP098017140, TdH, KP995184726 * TdO);
|
cannam@167
|
2806 }
|
cannam@167
|
2807 {
|
cannam@167
|
2808 E Td9, TdQ, Te3, Te6;
|
cannam@167
|
2809 Td9 = TcP + Td8;
|
cannam@167
|
2810 TdQ = Tdu + TdP;
|
cannam@167
|
2811 Cr[WS(csr, 62)] = Td9 - TdQ;
|
cannam@167
|
2812 Cr[WS(csr, 2)] = Td9 + TdQ;
|
cannam@167
|
2813 Te3 = Te0 + TdT;
|
cannam@167
|
2814 Te6 = Te4 + Te5;
|
cannam@167
|
2815 Ci[WS(csi, 2)] = Te3 + Te6;
|
cannam@167
|
2816 Ci[WS(csi, 62)] = Te6 - Te3;
|
cannam@167
|
2817 }
|
cannam@167
|
2818 {
|
cannam@167
|
2819 E Te1, Te2, Te7, Te8;
|
cannam@167
|
2820 Te1 = TdT - Te0;
|
cannam@167
|
2821 Te2 = TdP - Tdu;
|
cannam@167
|
2822 Ci[WS(csi, 30)] = Te1 + Te2;
|
cannam@167
|
2823 Ci[WS(csi, 34)] = Te2 - Te1;
|
cannam@167
|
2824 Te7 = TcP - Td8;
|
cannam@167
|
2825 Te8 = Te5 - Te4;
|
cannam@167
|
2826 Cr[WS(csr, 34)] = Te7 - Te8;
|
cannam@167
|
2827 Cr[WS(csr, 30)] = Te7 + Te8;
|
cannam@167
|
2828 }
|
cannam@167
|
2829 {
|
cannam@167
|
2830 E Teb, Tei, Ten, Teq;
|
cannam@167
|
2831 Teb = Te9 + Tea;
|
cannam@167
|
2832 Tei = Tee + Teh;
|
cannam@167
|
2833 Cr[WS(csr, 50)] = Teb - Tei;
|
cannam@167
|
2834 Cr[WS(csr, 14)] = Teb + Tei;
|
cannam@167
|
2835 Ten = Tek + Tej;
|
cannam@167
|
2836 Teq = Teo + Tep;
|
cannam@167
|
2837 Ci[WS(csi, 14)] = Ten + Teq;
|
cannam@167
|
2838 Ci[WS(csi, 50)] = Teq - Ten;
|
cannam@167
|
2839 }
|
cannam@167
|
2840 {
|
cannam@167
|
2841 E Tel, Tem, Ter, Tes;
|
cannam@167
|
2842 Tel = Tej - Tek;
|
cannam@167
|
2843 Tem = Teh - Tee;
|
cannam@167
|
2844 Ci[WS(csi, 18)] = Tel + Tem;
|
cannam@167
|
2845 Ci[WS(csi, 46)] = Tem - Tel;
|
cannam@167
|
2846 Ter = Te9 - Tea;
|
cannam@167
|
2847 Tes = Tep - Teo;
|
cannam@167
|
2848 Cr[WS(csr, 46)] = Ter - Tes;
|
cannam@167
|
2849 Cr[WS(csr, 18)] = Ter + Tes;
|
cannam@167
|
2850 }
|
cannam@167
|
2851 }
|
cannam@167
|
2852 {
|
cannam@167
|
2853 E T6v, T77, T6C, T7h, T6Y, T7i, T6V, T78, T6R, T7n, T73, T7f, T6K, T7m, T72;
|
cannam@167
|
2854 E T7c;
|
cannam@167
|
2855 {
|
cannam@167
|
2856 E T6t, T6u, T6T, T6U;
|
cannam@167
|
2857 T6t = T27 - T2e;
|
cannam@167
|
2858 T6u = T5Y - T5X;
|
cannam@167
|
2859 T6v = T6t - T6u;
|
cannam@167
|
2860 T77 = T6t + T6u;
|
cannam@167
|
2861 {
|
cannam@167
|
2862 E T6y, T6B, T6W, T6X;
|
cannam@167
|
2863 T6y = FMA(KP773010453, T6w, KP634393284 * T6x);
|
cannam@167
|
2864 T6B = FNMS(KP634393284, T6A, KP773010453 * T6z);
|
cannam@167
|
2865 T6C = T6y - T6B;
|
cannam@167
|
2866 T7h = T6B + T6y;
|
cannam@167
|
2867 T6W = T2x - T2o;
|
cannam@167
|
2868 T6X = T5V - T5S;
|
cannam@167
|
2869 T6Y = T6W - T6X;
|
cannam@167
|
2870 T7i = T6X + T6W;
|
cannam@167
|
2871 }
|
cannam@167
|
2872 T6T = FNMS(KP634393284, T6w, KP773010453 * T6x);
|
cannam@167
|
2873 T6U = FMA(KP634393284, T6z, KP773010453 * T6A);
|
cannam@167
|
2874 T6V = T6T - T6U;
|
cannam@167
|
2875 T78 = T6U + T6T;
|
cannam@167
|
2876 {
|
cannam@167
|
2877 E T6N, T7d, T6Q, T7e, T6M, T6O;
|
cannam@167
|
2878 T6M = T5I - T5H;
|
cannam@167
|
2879 T6N = T6L - T6M;
|
cannam@167
|
2880 T7d = T6L + T6M;
|
cannam@167
|
2881 T6O = T5v - T5e;
|
cannam@167
|
2882 T6Q = T6O - T6P;
|
cannam@167
|
2883 T7e = T6P + T6O;
|
cannam@167
|
2884 T6R = FNMS(KP427555093, T6Q, KP903989293 * T6N);
|
cannam@167
|
2885 T7n = FMA(KP941544065, T7e, KP336889853 * T7d);
|
cannam@167
|
2886 T73 = FMA(KP903989293, T6Q, KP427555093 * T6N);
|
cannam@167
|
2887 T7f = FNMS(KP336889853, T7e, KP941544065 * T7d);
|
cannam@167
|
2888 }
|
cannam@167
|
2889 {
|
cannam@167
|
2890 E T6G, T7a, T6J, T7b, T6E, T6I;
|
cannam@167
|
2891 T6E = T4g - T3Z;
|
cannam@167
|
2892 T6G = T6E - T6F;
|
cannam@167
|
2893 T7a = T6F + T6E;
|
cannam@167
|
2894 T6I = T4t - T4s;
|
cannam@167
|
2895 T6J = T6H - T6I;
|
cannam@167
|
2896 T7b = T6H + T6I;
|
cannam@167
|
2897 T6K = FMA(KP427555093, T6G, KP903989293 * T6J);
|
cannam@167
|
2898 T7m = FNMS(KP336889853, T7b, KP941544065 * T7a);
|
cannam@167
|
2899 T72 = FNMS(KP427555093, T6J, KP903989293 * T6G);
|
cannam@167
|
2900 T7c = FMA(KP336889853, T7a, KP941544065 * T7b);
|
cannam@167
|
2901 }
|
cannam@167
|
2902 }
|
cannam@167
|
2903 {
|
cannam@167
|
2904 E T6D, T6S, T71, T74;
|
cannam@167
|
2905 T6D = T6v + T6C;
|
cannam@167
|
2906 T6S = T6K + T6R;
|
cannam@167
|
2907 Cr[WS(csr, 55)] = T6D - T6S;
|
cannam@167
|
2908 Cr[WS(csr, 9)] = T6D + T6S;
|
cannam@167
|
2909 T71 = T6Y + T6V;
|
cannam@167
|
2910 T74 = T72 + T73;
|
cannam@167
|
2911 Ci[WS(csi, 9)] = T71 + T74;
|
cannam@167
|
2912 Ci[WS(csi, 55)] = T74 - T71;
|
cannam@167
|
2913 }
|
cannam@167
|
2914 {
|
cannam@167
|
2915 E T6Z, T70, T75, T76;
|
cannam@167
|
2916 T6Z = T6V - T6Y;
|
cannam@167
|
2917 T70 = T6R - T6K;
|
cannam@167
|
2918 Ci[WS(csi, 23)] = T6Z + T70;
|
cannam@167
|
2919 Ci[WS(csi, 41)] = T70 - T6Z;
|
cannam@167
|
2920 T75 = T6v - T6C;
|
cannam@167
|
2921 T76 = T73 - T72;
|
cannam@167
|
2922 Cr[WS(csr, 41)] = T75 - T76;
|
cannam@167
|
2923 Cr[WS(csr, 23)] = T75 + T76;
|
cannam@167
|
2924 }
|
cannam@167
|
2925 {
|
cannam@167
|
2926 E T79, T7g, T7l, T7o;
|
cannam@167
|
2927 T79 = T77 + T78;
|
cannam@167
|
2928 T7g = T7c + T7f;
|
cannam@167
|
2929 Cr[WS(csr, 57)] = T79 - T7g;
|
cannam@167
|
2930 Cr[WS(csr, 7)] = T79 + T7g;
|
cannam@167
|
2931 T7l = T7i + T7h;
|
cannam@167
|
2932 T7o = T7m + T7n;
|
cannam@167
|
2933 Ci[WS(csi, 7)] = T7l + T7o;
|
cannam@167
|
2934 Ci[WS(csi, 57)] = T7o - T7l;
|
cannam@167
|
2935 }
|
cannam@167
|
2936 {
|
cannam@167
|
2937 E T7j, T7k, T7p, T7q;
|
cannam@167
|
2938 T7j = T7h - T7i;
|
cannam@167
|
2939 T7k = T7f - T7c;
|
cannam@167
|
2940 Ci[WS(csi, 25)] = T7j + T7k;
|
cannam@167
|
2941 Ci[WS(csi, 39)] = T7k - T7j;
|
cannam@167
|
2942 T7p = T77 - T78;
|
cannam@167
|
2943 T7q = T7n - T7m;
|
cannam@167
|
2944 Cr[WS(csr, 39)] = T7p - T7q;
|
cannam@167
|
2945 Cr[WS(csr, 25)] = T7p + T7q;
|
cannam@167
|
2946 }
|
cannam@167
|
2947 }
|
cannam@167
|
2948 {
|
cannam@167
|
2949 E T99, T9L, T9g, T9V, T9C, T9W, T9z, T9M, T9v, Ta1, T9H, T9T, T9o, Ta0, T9G;
|
cannam@167
|
2950 E T9Q;
|
cannam@167
|
2951 {
|
cannam@167
|
2952 E T97, T98, T9x, T9y;
|
cannam@167
|
2953 T97 = T7r - T7s;
|
cannam@167
|
2954 T98 = T8C - T8B;
|
cannam@167
|
2955 T99 = T97 - T98;
|
cannam@167
|
2956 T9L = T97 + T98;
|
cannam@167
|
2957 {
|
cannam@167
|
2958 E T9c, T9f, T9A, T9B;
|
cannam@167
|
2959 T9c = FMA(KP471396736, T9a, KP881921264 * T9b);
|
cannam@167
|
2960 T9f = FNMS(KP471396736, T9e, KP881921264 * T9d);
|
cannam@167
|
2961 T9g = T9c - T9f;
|
cannam@167
|
2962 T9V = T9f + T9c;
|
cannam@167
|
2963 T9A = T7z - T7w;
|
cannam@167
|
2964 T9B = T8z - T8y;
|
cannam@167
|
2965 T9C = T9A - T9B;
|
cannam@167
|
2966 T9W = T9B + T9A;
|
cannam@167
|
2967 }
|
cannam@167
|
2968 T9x = FNMS(KP471396736, T9b, KP881921264 * T9a);
|
cannam@167
|
2969 T9y = FMA(KP881921264, T9e, KP471396736 * T9d);
|
cannam@167
|
2970 T9z = T9x - T9y;
|
cannam@167
|
2971 T9M = T9y + T9x;
|
cannam@167
|
2972 {
|
cannam@167
|
2973 E T9r, T9R, T9u, T9S, T9q, T9s;
|
cannam@167
|
2974 T9q = T8q - T8p;
|
cannam@167
|
2975 T9r = T9p - T9q;
|
cannam@167
|
2976 T9R = T9p + T9q;
|
cannam@167
|
2977 T9s = T8j - T8g;
|
cannam@167
|
2978 T9u = T9s - T9t;
|
cannam@167
|
2979 T9S = T9t + T9s;
|
cannam@167
|
2980 T9v = FNMS(KP514102744, T9u, KP857728610 * T9r);
|
cannam@167
|
2981 Ta1 = FMA(KP970031253, T9S, KP242980179 * T9R);
|
cannam@167
|
2982 T9H = FMA(KP857728610, T9u, KP514102744 * T9r);
|
cannam@167
|
2983 T9T = FNMS(KP242980179, T9S, KP970031253 * T9R);
|
cannam@167
|
2984 }
|
cannam@167
|
2985 {
|
cannam@167
|
2986 E T9k, T9O, T9n, T9P, T9i, T9m;
|
cannam@167
|
2987 T9i = T80 - T7X;
|
cannam@167
|
2988 T9k = T9i - T9j;
|
cannam@167
|
2989 T9O = T9j + T9i;
|
cannam@167
|
2990 T9m = T87 - T86;
|
cannam@167
|
2991 T9n = T9l - T9m;
|
cannam@167
|
2992 T9P = T9l + T9m;
|
cannam@167
|
2993 T9o = FMA(KP514102744, T9k, KP857728610 * T9n);
|
cannam@167
|
2994 Ta0 = FNMS(KP242980179, T9P, KP970031253 * T9O);
|
cannam@167
|
2995 T9G = FNMS(KP514102744, T9n, KP857728610 * T9k);
|
cannam@167
|
2996 T9Q = FMA(KP242980179, T9O, KP970031253 * T9P);
|
cannam@167
|
2997 }
|
cannam@167
|
2998 }
|
cannam@167
|
2999 {
|
cannam@167
|
3000 E T9h, T9w, T9F, T9I;
|
cannam@167
|
3001 T9h = T99 + T9g;
|
cannam@167
|
3002 T9w = T9o + T9v;
|
cannam@167
|
3003 Cr[WS(csr, 53)] = T9h - T9w;
|
cannam@167
|
3004 Cr[WS(csr, 11)] = T9h + T9w;
|
cannam@167
|
3005 T9F = T9C + T9z;
|
cannam@167
|
3006 T9I = T9G + T9H;
|
cannam@167
|
3007 Ci[WS(csi, 11)] = T9F + T9I;
|
cannam@167
|
3008 Ci[WS(csi, 53)] = T9I - T9F;
|
cannam@167
|
3009 }
|
cannam@167
|
3010 {
|
cannam@167
|
3011 E T9D, T9E, T9J, T9K;
|
cannam@167
|
3012 T9D = T9z - T9C;
|
cannam@167
|
3013 T9E = T9v - T9o;
|
cannam@167
|
3014 Ci[WS(csi, 21)] = T9D + T9E;
|
cannam@167
|
3015 Ci[WS(csi, 43)] = T9E - T9D;
|
cannam@167
|
3016 T9J = T99 - T9g;
|
cannam@167
|
3017 T9K = T9H - T9G;
|
cannam@167
|
3018 Cr[WS(csr, 43)] = T9J - T9K;
|
cannam@167
|
3019 Cr[WS(csr, 21)] = T9J + T9K;
|
cannam@167
|
3020 }
|
cannam@167
|
3021 {
|
cannam@167
|
3022 E T9N, T9U, T9Z, Ta2;
|
cannam@167
|
3023 T9N = T9L + T9M;
|
cannam@167
|
3024 T9U = T9Q + T9T;
|
cannam@167
|
3025 Cr[WS(csr, 59)] = T9N - T9U;
|
cannam@167
|
3026 Cr[WS(csr, 5)] = T9N + T9U;
|
cannam@167
|
3027 T9Z = T9W + T9V;
|
cannam@167
|
3028 Ta2 = Ta0 + Ta1;
|
cannam@167
|
3029 Ci[WS(csi, 5)] = T9Z + Ta2;
|
cannam@167
|
3030 Ci[WS(csi, 59)] = Ta2 - T9Z;
|
cannam@167
|
3031 }
|
cannam@167
|
3032 {
|
cannam@167
|
3033 E T9X, T9Y, Ta3, Ta4;
|
cannam@167
|
3034 T9X = T9V - T9W;
|
cannam@167
|
3035 T9Y = T9T - T9Q;
|
cannam@167
|
3036 Ci[WS(csi, 27)] = T9X + T9Y;
|
cannam@167
|
3037 Ci[WS(csi, 37)] = T9Y - T9X;
|
cannam@167
|
3038 Ta3 = T9L - T9M;
|
cannam@167
|
3039 Ta4 = Ta1 - Ta0;
|
cannam@167
|
3040 Cr[WS(csr, 37)] = Ta3 - Ta4;
|
cannam@167
|
3041 Cr[WS(csr, 27)] = Ta3 + Ta4;
|
cannam@167
|
3042 }
|
cannam@167
|
3043 }
|
cannam@167
|
3044 {
|
cannam@167
|
3045 E T2z, T69, T3g, T6j, T60, T6k, T5P, T6a, T5L, T6p, T65, T6h, T4w, T6o, T64;
|
cannam@167
|
3046 E T6e;
|
cannam@167
|
3047 {
|
cannam@167
|
3048 E T2f, T2y, T5N, T5O;
|
cannam@167
|
3049 T2f = T27 + T2e;
|
cannam@167
|
3050 T2y = T2o + T2x;
|
cannam@167
|
3051 T2z = T2f + T2y;
|
cannam@167
|
3052 T69 = T2f - T2y;
|
cannam@167
|
3053 {
|
cannam@167
|
3054 E T2U, T3f, T5W, T5Z;
|
cannam@167
|
3055 T2U = FMA(KP098017140, T2M, KP995184726 * T2T);
|
cannam@167
|
3056 T3f = FNMS(KP098017140, T3e, KP995184726 * T37);
|
cannam@167
|
3057 T3g = T2U + T3f;
|
cannam@167
|
3058 T6j = T3f - T2U;
|
cannam@167
|
3059 T5W = T5S + T5V;
|
cannam@167
|
3060 T5Z = T5X + T5Y;
|
cannam@167
|
3061 T60 = T5W + T5Z;
|
cannam@167
|
3062 T6k = T5Z - T5W;
|
cannam@167
|
3063 }
|
cannam@167
|
3064 T5N = FNMS(KP098017140, T2T, KP995184726 * T2M);
|
cannam@167
|
3065 T5O = FMA(KP995184726, T3e, KP098017140 * T37);
|
cannam@167
|
3066 T5P = T5N + T5O;
|
cannam@167
|
3067 T6a = T5O - T5N;
|
cannam@167
|
3068 {
|
cannam@167
|
3069 E T5x, T6f, T5K, T6g, T5w, T5J;
|
cannam@167
|
3070 T5w = T5e + T5v;
|
cannam@167
|
3071 T5x = T4X + T5w;
|
cannam@167
|
3072 T6f = T4X - T5w;
|
cannam@167
|
3073 T5J = T5H + T5I;
|
cannam@167
|
3074 T5K = T5G + T5J;
|
cannam@167
|
3075 T6g = T5J - T5G;
|
cannam@167
|
3076 T5L = FNMS(KP049067674, T5K, KP998795456 * T5x);
|
cannam@167
|
3077 T6p = FMA(KP671558954, T6f, KP740951125 * T6g);
|
cannam@167
|
3078 T65 = FMA(KP049067674, T5x, KP998795456 * T5K);
|
cannam@167
|
3079 T6h = FNMS(KP671558954, T6g, KP740951125 * T6f);
|
cannam@167
|
3080 }
|
cannam@167
|
3081 {
|
cannam@167
|
3082 E T4i, T6c, T4v, T6d, T4h, T4u;
|
cannam@167
|
3083 T4h = T3Z + T4g;
|
cannam@167
|
3084 T4i = T3I + T4h;
|
cannam@167
|
3085 T6c = T3I - T4h;
|
cannam@167
|
3086 T4u = T4s + T4t;
|
cannam@167
|
3087 T4v = T4r + T4u;
|
cannam@167
|
3088 T6d = T4u - T4r;
|
cannam@167
|
3089 T4w = FMA(KP998795456, T4i, KP049067674 * T4v);
|
cannam@167
|
3090 T6o = FNMS(KP671558954, T6c, KP740951125 * T6d);
|
cannam@167
|
3091 T64 = FNMS(KP049067674, T4i, KP998795456 * T4v);
|
cannam@167
|
3092 T6e = FMA(KP740951125, T6c, KP671558954 * T6d);
|
cannam@167
|
3093 }
|
cannam@167
|
3094 }
|
cannam@167
|
3095 {
|
cannam@167
|
3096 E T3h, T5M, T63, T66;
|
cannam@167
|
3097 T3h = T2z + T3g;
|
cannam@167
|
3098 T5M = T4w + T5L;
|
cannam@167
|
3099 Cr[WS(csr, 63)] = T3h - T5M;
|
cannam@167
|
3100 Cr[WS(csr, 1)] = T3h + T5M;
|
cannam@167
|
3101 T63 = T60 + T5P;
|
cannam@167
|
3102 T66 = T64 + T65;
|
cannam@167
|
3103 Ci[WS(csi, 1)] = T63 + T66;
|
cannam@167
|
3104 Ci[WS(csi, 63)] = T66 - T63;
|
cannam@167
|
3105 }
|
cannam@167
|
3106 {
|
cannam@167
|
3107 E T61, T62, T67, T68;
|
cannam@167
|
3108 T61 = T5P - T60;
|
cannam@167
|
3109 T62 = T5L - T4w;
|
cannam@167
|
3110 Ci[WS(csi, 31)] = T61 + T62;
|
cannam@167
|
3111 Ci[WS(csi, 33)] = T62 - T61;
|
cannam@167
|
3112 T67 = T2z - T3g;
|
cannam@167
|
3113 T68 = T65 - T64;
|
cannam@167
|
3114 Cr[WS(csr, 33)] = T67 - T68;
|
cannam@167
|
3115 Cr[WS(csr, 31)] = T67 + T68;
|
cannam@167
|
3116 }
|
cannam@167
|
3117 {
|
cannam@167
|
3118 E T6b, T6i, T6n, T6q;
|
cannam@167
|
3119 T6b = T69 + T6a;
|
cannam@167
|
3120 T6i = T6e + T6h;
|
cannam@167
|
3121 Cr[WS(csr, 49)] = T6b - T6i;
|
cannam@167
|
3122 Cr[WS(csr, 15)] = T6b + T6i;
|
cannam@167
|
3123 T6n = T6k + T6j;
|
cannam@167
|
3124 T6q = T6o + T6p;
|
cannam@167
|
3125 Ci[WS(csi, 15)] = T6n + T6q;
|
cannam@167
|
3126 Ci[WS(csi, 49)] = T6q - T6n;
|
cannam@167
|
3127 }
|
cannam@167
|
3128 {
|
cannam@167
|
3129 E T6l, T6m, T6r, T6s;
|
cannam@167
|
3130 T6l = T6j - T6k;
|
cannam@167
|
3131 T6m = T6h - T6e;
|
cannam@167
|
3132 Ci[WS(csi, 17)] = T6l + T6m;
|
cannam@167
|
3133 Ci[WS(csi, 47)] = T6m - T6l;
|
cannam@167
|
3134 T6r = T69 - T6a;
|
cannam@167
|
3135 T6s = T6p - T6o;
|
cannam@167
|
3136 Cr[WS(csr, 47)] = T6r - T6s;
|
cannam@167
|
3137 Cr[WS(csr, 17)] = T6r + T6s;
|
cannam@167
|
3138 }
|
cannam@167
|
3139 }
|
cannam@167
|
3140 {
|
cannam@167
|
3141 E T7B, T8N, T7Q, T8X, T8E, T8Y, T8x, T8O, T8t, T93, T8J, T8V, T8a, T92, T8I;
|
cannam@167
|
3142 E T8S;
|
cannam@167
|
3143 {
|
cannam@167
|
3144 E T7t, T7A, T8v, T8w;
|
cannam@167
|
3145 T7t = T7r + T7s;
|
cannam@167
|
3146 T7A = T7w + T7z;
|
cannam@167
|
3147 T7B = T7t + T7A;
|
cannam@167
|
3148 T8N = T7t - T7A;
|
cannam@167
|
3149 {
|
cannam@167
|
3150 E T7I, T7P, T8A, T8D;
|
cannam@167
|
3151 T7I = FMA(KP956940335, T7E, KP290284677 * T7H);
|
cannam@167
|
3152 T7P = FNMS(KP290284677, T7O, KP956940335 * T7L);
|
cannam@167
|
3153 T7Q = T7I + T7P;
|
cannam@167
|
3154 T8X = T7P - T7I;
|
cannam@167
|
3155 T8A = T8y + T8z;
|
cannam@167
|
3156 T8D = T8B + T8C;
|
cannam@167
|
3157 T8E = T8A + T8D;
|
cannam@167
|
3158 T8Y = T8D - T8A;
|
cannam@167
|
3159 }
|
cannam@167
|
3160 T8v = FNMS(KP290284677, T7E, KP956940335 * T7H);
|
cannam@167
|
3161 T8w = FMA(KP290284677, T7L, KP956940335 * T7O);
|
cannam@167
|
3162 T8x = T8v + T8w;
|
cannam@167
|
3163 T8O = T8w - T8v;
|
cannam@167
|
3164 {
|
cannam@167
|
3165 E T8l, T8T, T8s, T8U, T8k, T8r;
|
cannam@167
|
3166 T8k = T8g + T8j;
|
cannam@167
|
3167 T8l = T8d + T8k;
|
cannam@167
|
3168 T8T = T8d - T8k;
|
cannam@167
|
3169 T8r = T8p + T8q;
|
cannam@167
|
3170 T8s = T8o + T8r;
|
cannam@167
|
3171 T8U = T8r - T8o;
|
cannam@167
|
3172 T8t = FNMS(KP146730474, T8s, KP989176509 * T8l);
|
cannam@167
|
3173 T93 = FMA(KP595699304, T8T, KP803207531 * T8U);
|
cannam@167
|
3174 T8J = FMA(KP146730474, T8l, KP989176509 * T8s);
|
cannam@167
|
3175 T8V = FNMS(KP595699304, T8U, KP803207531 * T8T);
|
cannam@167
|
3176 }
|
cannam@167
|
3177 {
|
cannam@167
|
3178 E T82, T8Q, T89, T8R, T81, T88;
|
cannam@167
|
3179 T81 = T7X + T80;
|
cannam@167
|
3180 T82 = T7U + T81;
|
cannam@167
|
3181 T8Q = T7U - T81;
|
cannam@167
|
3182 T88 = T86 + T87;
|
cannam@167
|
3183 T89 = T85 + T88;
|
cannam@167
|
3184 T8R = T88 - T85;
|
cannam@167
|
3185 T8a = FMA(KP989176509, T82, KP146730474 * T89);
|
cannam@167
|
3186 T92 = FNMS(KP595699304, T8Q, KP803207531 * T8R);
|
cannam@167
|
3187 T8I = FNMS(KP146730474, T82, KP989176509 * T89);
|
cannam@167
|
3188 T8S = FMA(KP803207531, T8Q, KP595699304 * T8R);
|
cannam@167
|
3189 }
|
cannam@167
|
3190 }
|
cannam@167
|
3191 {
|
cannam@167
|
3192 E T7R, T8u, T8H, T8K;
|
cannam@167
|
3193 T7R = T7B + T7Q;
|
cannam@167
|
3194 T8u = T8a + T8t;
|
cannam@167
|
3195 Cr[WS(csr, 61)] = T7R - T8u;
|
cannam@167
|
3196 Cr[WS(csr, 3)] = T7R + T8u;
|
cannam@167
|
3197 T8H = T8E + T8x;
|
cannam@167
|
3198 T8K = T8I + T8J;
|
cannam@167
|
3199 Ci[WS(csi, 3)] = T8H + T8K;
|
cannam@167
|
3200 Ci[WS(csi, 61)] = T8K - T8H;
|
cannam@167
|
3201 }
|
cannam@167
|
3202 {
|
cannam@167
|
3203 E T8F, T8G, T8L, T8M;
|
cannam@167
|
3204 T8F = T8x - T8E;
|
cannam@167
|
3205 T8G = T8t - T8a;
|
cannam@167
|
3206 Ci[WS(csi, 29)] = T8F + T8G;
|
cannam@167
|
3207 Ci[WS(csi, 35)] = T8G - T8F;
|
cannam@167
|
3208 T8L = T7B - T7Q;
|
cannam@167
|
3209 T8M = T8J - T8I;
|
cannam@167
|
3210 Cr[WS(csr, 35)] = T8L - T8M;
|
cannam@167
|
3211 Cr[WS(csr, 29)] = T8L + T8M;
|
cannam@167
|
3212 }
|
cannam@167
|
3213 {
|
cannam@167
|
3214 E T8P, T8W, T91, T94;
|
cannam@167
|
3215 T8P = T8N + T8O;
|
cannam@167
|
3216 T8W = T8S + T8V;
|
cannam@167
|
3217 Cr[WS(csr, 51)] = T8P - T8W;
|
cannam@167
|
3218 Cr[WS(csr, 13)] = T8P + T8W;
|
cannam@167
|
3219 T91 = T8Y + T8X;
|
cannam@167
|
3220 T94 = T92 + T93;
|
cannam@167
|
3221 Ci[WS(csi, 13)] = T91 + T94;
|
cannam@167
|
3222 Ci[WS(csi, 51)] = T94 - T91;
|
cannam@167
|
3223 }
|
cannam@167
|
3224 {
|
cannam@167
|
3225 E T8Z, T90, T95, T96;
|
cannam@167
|
3226 T8Z = T8X - T8Y;
|
cannam@167
|
3227 T90 = T8V - T8S;
|
cannam@167
|
3228 Ci[WS(csi, 19)] = T8Z + T90;
|
cannam@167
|
3229 Ci[WS(csi, 45)] = T90 - T8Z;
|
cannam@167
|
3230 T95 = T8N - T8O;
|
cannam@167
|
3231 T96 = T93 - T92;
|
cannam@167
|
3232 Cr[WS(csr, 45)] = T95 - T96;
|
cannam@167
|
3233 Cr[WS(csr, 19)] = T95 + T96;
|
cannam@167
|
3234 }
|
cannam@167
|
3235 }
|
cannam@167
|
3236 }
|
cannam@167
|
3237 }
|
cannam@167
|
3238 }
|
cannam@167
|
3239
|
cannam@167
|
3240 static const kr2c_desc desc = { 128, "r2cf_128", {812, 186, 144, 0}, &GENUS };
|
cannam@167
|
3241
|
cannam@167
|
3242 void X(codelet_r2cf_128) (planner *p) {
|
cannam@167
|
3243 X(kr2c_register) (p, r2cf_128, &desc);
|
cannam@167
|
3244 }
|
cannam@167
|
3245
|
cannam@167
|
3246 #endif
|