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 08:06:44 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 64 -name r2cfII_64 -dft-II -include rdft/scalar/r2cfII.h */
|
cannam@167
|
29
|
cannam@167
|
30 /*
|
cannam@167
|
31 * This function contains 434 FP additions, 320 FP multiplications,
|
cannam@167
|
32 * (or, 114 additions, 0 multiplications, 320 fused multiply/add),
|
cannam@167
|
33 * 118 stack variables, 31 constants, and 128 memory accesses
|
cannam@167
|
34 */
|
cannam@167
|
35 #include "rdft/scalar/r2cfII.h"
|
cannam@167
|
36
|
cannam@167
|
37 static void r2cfII_64(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(KP941544065, +0.941544065183020778412509402599502357185589796);
|
cannam@167
|
40 DK(KP903989293, +0.903989293123443331586200297230537048710132025);
|
cannam@167
|
41 DK(KP773010453, +0.773010453362736960810906609758469800971041293);
|
cannam@167
|
42 DK(KP472964775, +0.472964775891319928124438237972992463904131113);
|
cannam@167
|
43 DK(KP357805721, +0.357805721314524104672487743774474392487532769);
|
cannam@167
|
44 DK(KP820678790, +0.820678790828660330972281985331011598767386482);
|
cannam@167
|
45 DK(KP989176509, +0.989176509964780973451673738016243063983689533);
|
cannam@167
|
46 DK(KP803207531, +0.803207531480644909806676512963141923879569427);
|
cannam@167
|
47 DK(KP956940335, +0.956940335732208864935797886980269969482849206);
|
cannam@167
|
48 DK(KP741650546, +0.741650546272035369581266691172079863842265220);
|
cannam@167
|
49 DK(KP148335987, +0.148335987538347428753676511486911367000625355);
|
cannam@167
|
50 DK(KP303346683, +0.303346683607342391675883946941299872384187453);
|
cannam@167
|
51 DK(KP998795456, +0.998795456205172392714771604759100694443203615);
|
cannam@167
|
52 DK(KP740951125, +0.740951125354959091175616897495162729728955309);
|
cannam@167
|
53 DK(KP995184726, +0.995184726672196886244836953109479921575474869);
|
cannam@167
|
54 DK(KP906347169, +0.906347169019147157946142717268914412664134293);
|
cannam@167
|
55 DK(KP049126849, +0.049126849769467254105343321271313617079695752);
|
cannam@167
|
56 DK(KP098491403, +0.098491403357164253077197521291327432293052451);
|
cannam@167
|
57 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
cannam@167
|
58 DK(KP970031253, +0.970031253194543992603984207286100251456865962);
|
cannam@167
|
59 DK(KP857728610, +0.857728610000272069902269984284770137042490799);
|
cannam@167
|
60 DK(KP881921264, +0.881921264348355029712756863660388349508442621);
|
cannam@167
|
61 DK(KP599376933, +0.599376933681923766271389869014404232837890546);
|
cannam@167
|
62 DK(KP250486960, +0.250486960191305461595702160124721208578685568);
|
cannam@167
|
63 DK(KP534511135, +0.534511135950791641089685961295362908582039528);
|
cannam@167
|
64 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
cannam@167
|
65 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
cannam@167
|
66 DK(KP414213562, +0.414213562373095048801688724209698078569671875);
|
cannam@167
|
67 DK(KP198912367, +0.198912367379658006911597622644676228597850501);
|
cannam@167
|
68 DK(KP668178637, +0.668178637919298919997757686523080761552472251);
|
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(256, rs), MAKE_VOLATILE_STRIDE(256, csr), MAKE_VOLATILE_STRIDE(256, csi)) {
|
cannam@167
|
73 E Tm, T35, T3Z, T5h, Tv, T34, T3W, T5g, Td, T33, T6z, T6N, T3T, T5f, T65;
|
cannam@167
|
74 E T6j, T2b, T3n, T4O, T5C, T2C, T3q, T4D, T5z, TK, T3b, T4e, T5l, TR, T3c;
|
cannam@167
|
75 E T4b, T5k, T15, T38, T47, T5o, T1c, T39, T44, T5n, T1s, T3g, T4v, T5v, T1T;
|
cannam@167
|
76 E T3j, T4k, T5s, T2u, T3r, T4R, T5A, T2F, T3o, T4K, T5D, T1L, T3k, T4y, T5t;
|
cannam@167
|
77 E T1W, T3h, T4r, T5w;
|
cannam@167
|
78 {
|
cannam@167
|
79 E Te, Tj, Th, Tk, Tf, Tg;
|
cannam@167
|
80 Te = R0[WS(rs, 14)];
|
cannam@167
|
81 Tj = R0[WS(rs, 30)];
|
cannam@167
|
82 Tf = R0[WS(rs, 6)];
|
cannam@167
|
83 Tg = R0[WS(rs, 22)];
|
cannam@167
|
84 Th = Tf + Tg;
|
cannam@167
|
85 Tk = Tg - Tf;
|
cannam@167
|
86 {
|
cannam@167
|
87 E Ti, Tl, T3X, T3Y;
|
cannam@167
|
88 Ti = FNMS(KP707106781, Th, Te);
|
cannam@167
|
89 Tl = FNMS(KP707106781, Tk, Tj);
|
cannam@167
|
90 Tm = FNMS(KP668178637, Tl, Ti);
|
cannam@167
|
91 T35 = FMA(KP668178637, Ti, Tl);
|
cannam@167
|
92 T3X = FMA(KP707106781, Th, Te);
|
cannam@167
|
93 T3Y = FMA(KP707106781, Tk, Tj);
|
cannam@167
|
94 T3Z = FMA(KP198912367, T3Y, T3X);
|
cannam@167
|
95 T5h = FNMS(KP198912367, T3X, T3Y);
|
cannam@167
|
96 }
|
cannam@167
|
97 }
|
cannam@167
|
98 {
|
cannam@167
|
99 E Tn, Ts, Tq, Tt, To, Tp;
|
cannam@167
|
100 Tn = R0[WS(rs, 18)];
|
cannam@167
|
101 Ts = R0[WS(rs, 2)];
|
cannam@167
|
102 To = R0[WS(rs, 10)];
|
cannam@167
|
103 Tp = R0[WS(rs, 26)];
|
cannam@167
|
104 Tq = To + Tp;
|
cannam@167
|
105 Tt = To - Tp;
|
cannam@167
|
106 {
|
cannam@167
|
107 E Tr, Tu, T3U, T3V;
|
cannam@167
|
108 Tr = FNMS(KP707106781, Tq, Tn);
|
cannam@167
|
109 Tu = FNMS(KP707106781, Tt, Ts);
|
cannam@167
|
110 Tv = FNMS(KP668178637, Tu, Tr);
|
cannam@167
|
111 T34 = FMA(KP668178637, Tr, Tu);
|
cannam@167
|
112 T3U = FMA(KP707106781, Tq, Tn);
|
cannam@167
|
113 T3V = FMA(KP707106781, Tt, Ts);
|
cannam@167
|
114 T3W = FMA(KP198912367, T3V, T3U);
|
cannam@167
|
115 T5g = FNMS(KP198912367, T3U, T3V);
|
cannam@167
|
116 }
|
cannam@167
|
117 }
|
cannam@167
|
118 {
|
cannam@167
|
119 E T1, T61, T4, T62, T8, T3Q, Tb, T3R, T2, T3;
|
cannam@167
|
120 T1 = R0[0];
|
cannam@167
|
121 T61 = R0[WS(rs, 16)];
|
cannam@167
|
122 T2 = R0[WS(rs, 8)];
|
cannam@167
|
123 T3 = R0[WS(rs, 24)];
|
cannam@167
|
124 T4 = T2 - T3;
|
cannam@167
|
125 T62 = T2 + T3;
|
cannam@167
|
126 {
|
cannam@167
|
127 E T6, T7, T9, Ta;
|
cannam@167
|
128 T6 = R0[WS(rs, 20)];
|
cannam@167
|
129 T7 = R0[WS(rs, 4)];
|
cannam@167
|
130 T8 = FMA(KP414213562, T7, T6);
|
cannam@167
|
131 T3Q = FNMS(KP414213562, T6, T7);
|
cannam@167
|
132 T9 = R0[WS(rs, 12)];
|
cannam@167
|
133 Ta = R0[WS(rs, 28)];
|
cannam@167
|
134 Tb = FMA(KP414213562, Ta, T9);
|
cannam@167
|
135 T3R = FMS(KP414213562, T9, Ta);
|
cannam@167
|
136 }
|
cannam@167
|
137 {
|
cannam@167
|
138 E T5, Tc, T6x, T6y;
|
cannam@167
|
139 T5 = FNMS(KP707106781, T4, T1);
|
cannam@167
|
140 Tc = T8 - Tb;
|
cannam@167
|
141 Td = FNMS(KP923879532, Tc, T5);
|
cannam@167
|
142 T33 = FMA(KP923879532, Tc, T5);
|
cannam@167
|
143 T6x = FNMS(KP707106781, T62, T61);
|
cannam@167
|
144 T6y = T3R - T3Q;
|
cannam@167
|
145 T6z = FMA(KP923879532, T6y, T6x);
|
cannam@167
|
146 T6N = FNMS(KP923879532, T6y, T6x);
|
cannam@167
|
147 }
|
cannam@167
|
148 {
|
cannam@167
|
149 E T3P, T3S, T63, T64;
|
cannam@167
|
150 T3P = FMA(KP707106781, T4, T1);
|
cannam@167
|
151 T3S = T3Q + T3R;
|
cannam@167
|
152 T3T = FNMS(KP923879532, T3S, T3P);
|
cannam@167
|
153 T5f = FMA(KP923879532, T3S, T3P);
|
cannam@167
|
154 T63 = FMA(KP707106781, T62, T61);
|
cannam@167
|
155 T64 = T8 + Tb;
|
cannam@167
|
156 T65 = FMA(KP923879532, T64, T63);
|
cannam@167
|
157 T6j = FNMS(KP923879532, T64, T63);
|
cannam@167
|
158 }
|
cannam@167
|
159 }
|
cannam@167
|
160 {
|
cannam@167
|
161 E T1Z, T2w, T22, T2x, T26, T2A, T29, T2z, T20, T21;
|
cannam@167
|
162 T1Z = R1[WS(rs, 31)];
|
cannam@167
|
163 T2w = R1[WS(rs, 15)];
|
cannam@167
|
164 T20 = R1[WS(rs, 7)];
|
cannam@167
|
165 T21 = R1[WS(rs, 23)];
|
cannam@167
|
166 T22 = T20 - T21;
|
cannam@167
|
167 T2x = T20 + T21;
|
cannam@167
|
168 {
|
cannam@167
|
169 E T24, T25, T27, T28;
|
cannam@167
|
170 T24 = R1[WS(rs, 19)];
|
cannam@167
|
171 T25 = R1[WS(rs, 3)];
|
cannam@167
|
172 T26 = FMA(KP414213562, T25, T24);
|
cannam@167
|
173 T2A = FNMS(KP414213562, T24, T25);
|
cannam@167
|
174 T27 = R1[WS(rs, 11)];
|
cannam@167
|
175 T28 = R1[WS(rs, 27)];
|
cannam@167
|
176 T29 = FMA(KP414213562, T28, T27);
|
cannam@167
|
177 T2z = FMS(KP414213562, T27, T28);
|
cannam@167
|
178 }
|
cannam@167
|
179 {
|
cannam@167
|
180 E T23, T2a, T4M, T4N;
|
cannam@167
|
181 T23 = FMA(KP707106781, T22, T1Z);
|
cannam@167
|
182 T2a = T26 - T29;
|
cannam@167
|
183 T2b = FMA(KP923879532, T2a, T23);
|
cannam@167
|
184 T3n = FNMS(KP923879532, T2a, T23);
|
cannam@167
|
185 T4M = FMA(KP707106781, T2x, T2w);
|
cannam@167
|
186 T4N = T26 + T29;
|
cannam@167
|
187 T4O = FNMS(KP923879532, T4N, T4M);
|
cannam@167
|
188 T5C = FMA(KP923879532, T4N, T4M);
|
cannam@167
|
189 }
|
cannam@167
|
190 {
|
cannam@167
|
191 E T2y, T2B, T4B, T4C;
|
cannam@167
|
192 T2y = FNMS(KP707106781, T2x, T2w);
|
cannam@167
|
193 T2B = T2z - T2A;
|
cannam@167
|
194 T2C = FNMS(KP923879532, T2B, T2y);
|
cannam@167
|
195 T3q = FMA(KP923879532, T2B, T2y);
|
cannam@167
|
196 T4B = FMS(KP707106781, T22, T1Z);
|
cannam@167
|
197 T4C = T2A + T2z;
|
cannam@167
|
198 T4D = FNMS(KP923879532, T4C, T4B);
|
cannam@167
|
199 T5z = FMA(KP923879532, T4C, T4B);
|
cannam@167
|
200 }
|
cannam@167
|
201 }
|
cannam@167
|
202 {
|
cannam@167
|
203 E Ty, TL, TB, TM, TF, TP, TI, TO, Tz, TA;
|
cannam@167
|
204 Ty = R0[WS(rs, 17)];
|
cannam@167
|
205 TL = R0[WS(rs, 1)];
|
cannam@167
|
206 Tz = R0[WS(rs, 9)];
|
cannam@167
|
207 TA = R0[WS(rs, 25)];
|
cannam@167
|
208 TB = Tz + TA;
|
cannam@167
|
209 TM = Tz - TA;
|
cannam@167
|
210 {
|
cannam@167
|
211 E TD, TE, TG, TH;
|
cannam@167
|
212 TD = R0[WS(rs, 29)];
|
cannam@167
|
213 TE = R0[WS(rs, 13)];
|
cannam@167
|
214 TF = FMS(KP414213562, TE, TD);
|
cannam@167
|
215 TP = FMA(KP414213562, TD, TE);
|
cannam@167
|
216 TG = R0[WS(rs, 5)];
|
cannam@167
|
217 TH = R0[WS(rs, 21)];
|
cannam@167
|
218 TI = FNMS(KP414213562, TH, TG);
|
cannam@167
|
219 TO = FMA(KP414213562, TG, TH);
|
cannam@167
|
220 }
|
cannam@167
|
221 {
|
cannam@167
|
222 E TC, TJ, T4c, T4d;
|
cannam@167
|
223 TC = FNMS(KP707106781, TB, Ty);
|
cannam@167
|
224 TJ = TF - TI;
|
cannam@167
|
225 TK = FNMS(KP923879532, TJ, TC);
|
cannam@167
|
226 T3b = FMA(KP923879532, TJ, TC);
|
cannam@167
|
227 T4c = FMA(KP707106781, TM, TL);
|
cannam@167
|
228 T4d = TI + TF;
|
cannam@167
|
229 T4e = FNMS(KP923879532, T4d, T4c);
|
cannam@167
|
230 T5l = FMA(KP923879532, T4d, T4c);
|
cannam@167
|
231 }
|
cannam@167
|
232 {
|
cannam@167
|
233 E TN, TQ, T49, T4a;
|
cannam@167
|
234 TN = FNMS(KP707106781, TM, TL);
|
cannam@167
|
235 TQ = TO - TP;
|
cannam@167
|
236 TR = FNMS(KP923879532, TQ, TN);
|
cannam@167
|
237 T3c = FMA(KP923879532, TQ, TN);
|
cannam@167
|
238 T49 = FMA(KP707106781, TB, Ty);
|
cannam@167
|
239 T4a = TO + TP;
|
cannam@167
|
240 T4b = FNMS(KP923879532, T4a, T49);
|
cannam@167
|
241 T5k = FMA(KP923879532, T4a, T49);
|
cannam@167
|
242 }
|
cannam@167
|
243 }
|
cannam@167
|
244 {
|
cannam@167
|
245 E TT, T16, TW, T17, T10, T1a, T13, T19, TU, TV;
|
cannam@167
|
246 TT = R0[WS(rs, 15)];
|
cannam@167
|
247 T16 = R0[WS(rs, 31)];
|
cannam@167
|
248 TU = R0[WS(rs, 7)];
|
cannam@167
|
249 TV = R0[WS(rs, 23)];
|
cannam@167
|
250 TW = TU + TV;
|
cannam@167
|
251 T17 = TV - TU;
|
cannam@167
|
252 {
|
cannam@167
|
253 E TY, TZ, T11, T12;
|
cannam@167
|
254 TY = R0[WS(rs, 3)];
|
cannam@167
|
255 TZ = R0[WS(rs, 19)];
|
cannam@167
|
256 T10 = FMS(KP414213562, TZ, TY);
|
cannam@167
|
257 T1a = FMA(KP414213562, TY, TZ);
|
cannam@167
|
258 T11 = R0[WS(rs, 27)];
|
cannam@167
|
259 T12 = R0[WS(rs, 11)];
|
cannam@167
|
260 T13 = FNMS(KP414213562, T12, T11);
|
cannam@167
|
261 T19 = FMA(KP414213562, T11, T12);
|
cannam@167
|
262 }
|
cannam@167
|
263 {
|
cannam@167
|
264 E TX, T14, T45, T46;
|
cannam@167
|
265 TX = FNMS(KP707106781, TW, TT);
|
cannam@167
|
266 T14 = T10 - T13;
|
cannam@167
|
267 T15 = FNMS(KP923879532, T14, TX);
|
cannam@167
|
268 T38 = FMA(KP923879532, T14, TX);
|
cannam@167
|
269 T45 = FMA(KP707106781, T17, T16);
|
cannam@167
|
270 T46 = T10 + T13;
|
cannam@167
|
271 T47 = FNMS(KP923879532, T46, T45);
|
cannam@167
|
272 T5o = FMA(KP923879532, T46, T45);
|
cannam@167
|
273 }
|
cannam@167
|
274 {
|
cannam@167
|
275 E T18, T1b, T42, T43;
|
cannam@167
|
276 T18 = FNMS(KP707106781, T17, T16);
|
cannam@167
|
277 T1b = T19 - T1a;
|
cannam@167
|
278 T1c = FNMS(KP923879532, T1b, T18);
|
cannam@167
|
279 T39 = FMA(KP923879532, T1b, T18);
|
cannam@167
|
280 T42 = FMA(KP707106781, TW, TT);
|
cannam@167
|
281 T43 = T1a + T19;
|
cannam@167
|
282 T44 = FNMS(KP923879532, T43, T42);
|
cannam@167
|
283 T5n = FMA(KP923879532, T43, T42);
|
cannam@167
|
284 }
|
cannam@167
|
285 }
|
cannam@167
|
286 {
|
cannam@167
|
287 E T1g, T1N, T1j, T1O, T1n, T1R, T1q, T1Q, T1h, T1i;
|
cannam@167
|
288 T1g = R1[0];
|
cannam@167
|
289 T1N = R1[WS(rs, 16)];
|
cannam@167
|
290 T1h = R1[WS(rs, 8)];
|
cannam@167
|
291 T1i = R1[WS(rs, 24)];
|
cannam@167
|
292 T1j = T1h - T1i;
|
cannam@167
|
293 T1O = T1h + T1i;
|
cannam@167
|
294 {
|
cannam@167
|
295 E T1l, T1m, T1o, T1p;
|
cannam@167
|
296 T1l = R1[WS(rs, 20)];
|
cannam@167
|
297 T1m = R1[WS(rs, 4)];
|
cannam@167
|
298 T1n = FMA(KP414213562, T1m, T1l);
|
cannam@167
|
299 T1R = FNMS(KP414213562, T1l, T1m);
|
cannam@167
|
300 T1o = R1[WS(rs, 12)];
|
cannam@167
|
301 T1p = R1[WS(rs, 28)];
|
cannam@167
|
302 T1q = FMA(KP414213562, T1p, T1o);
|
cannam@167
|
303 T1Q = FMS(KP414213562, T1o, T1p);
|
cannam@167
|
304 }
|
cannam@167
|
305 {
|
cannam@167
|
306 E T1k, T1r, T4t, T4u;
|
cannam@167
|
307 T1k = FNMS(KP707106781, T1j, T1g);
|
cannam@167
|
308 T1r = T1n - T1q;
|
cannam@167
|
309 T1s = FNMS(KP923879532, T1r, T1k);
|
cannam@167
|
310 T3g = FMA(KP923879532, T1r, T1k);
|
cannam@167
|
311 T4t = FMA(KP707106781, T1O, T1N);
|
cannam@167
|
312 T4u = T1n + T1q;
|
cannam@167
|
313 T4v = FNMS(KP923879532, T4u, T4t);
|
cannam@167
|
314 T5v = FMA(KP923879532, T4u, T4t);
|
cannam@167
|
315 }
|
cannam@167
|
316 {
|
cannam@167
|
317 E T1P, T1S, T4i, T4j;
|
cannam@167
|
318 T1P = FNMS(KP707106781, T1O, T1N);
|
cannam@167
|
319 T1S = T1Q - T1R;
|
cannam@167
|
320 T1T = FNMS(KP923879532, T1S, T1P);
|
cannam@167
|
321 T3j = FMA(KP923879532, T1S, T1P);
|
cannam@167
|
322 T4i = FMA(KP707106781, T1j, T1g);
|
cannam@167
|
323 T4j = T1R + T1Q;
|
cannam@167
|
324 T4k = FNMS(KP923879532, T4j, T4i);
|
cannam@167
|
325 T5s = FMA(KP923879532, T4j, T4i);
|
cannam@167
|
326 }
|
cannam@167
|
327 }
|
cannam@167
|
328 {
|
cannam@167
|
329 E T2g, T4I, T2j, T4H, T2p, T4F, T2s, T4E;
|
cannam@167
|
330 {
|
cannam@167
|
331 E T2c, T2h, T2f, T2i, T2d, T2e;
|
cannam@167
|
332 T2c = R1[WS(rs, 13)];
|
cannam@167
|
333 T2h = R1[WS(rs, 29)];
|
cannam@167
|
334 T2d = R1[WS(rs, 5)];
|
cannam@167
|
335 T2e = R1[WS(rs, 21)];
|
cannam@167
|
336 T2f = T2d + T2e;
|
cannam@167
|
337 T2i = T2d - T2e;
|
cannam@167
|
338 T2g = FNMS(KP707106781, T2f, T2c);
|
cannam@167
|
339 T4I = FMS(KP707106781, T2i, T2h);
|
cannam@167
|
340 T2j = FMA(KP707106781, T2i, T2h);
|
cannam@167
|
341 T4H = FMA(KP707106781, T2f, T2c);
|
cannam@167
|
342 }
|
cannam@167
|
343 {
|
cannam@167
|
344 E T2l, T2q, T2o, T2r, T2m, T2n;
|
cannam@167
|
345 T2l = R1[WS(rs, 17)];
|
cannam@167
|
346 T2q = R1[WS(rs, 1)];
|
cannam@167
|
347 T2m = R1[WS(rs, 9)];
|
cannam@167
|
348 T2n = R1[WS(rs, 25)];
|
cannam@167
|
349 T2o = T2m + T2n;
|
cannam@167
|
350 T2r = T2m - T2n;
|
cannam@167
|
351 T2p = FNMS(KP707106781, T2o, T2l);
|
cannam@167
|
352 T4F = FMA(KP707106781, T2r, T2q);
|
cannam@167
|
353 T2s = FNMS(KP707106781, T2r, T2q);
|
cannam@167
|
354 T4E = FMA(KP707106781, T2o, T2l);
|
cannam@167
|
355 }
|
cannam@167
|
356 {
|
cannam@167
|
357 E T2k, T2t, T4P, T4Q;
|
cannam@167
|
358 T2k = FNMS(KP668178637, T2j, T2g);
|
cannam@167
|
359 T2t = FNMS(KP668178637, T2s, T2p);
|
cannam@167
|
360 T2u = T2k - T2t;
|
cannam@167
|
361 T3r = T2t + T2k;
|
cannam@167
|
362 T4P = FMA(KP198912367, T4H, T4I);
|
cannam@167
|
363 T4Q = FNMS(KP198912367, T4E, T4F);
|
cannam@167
|
364 T4R = T4P - T4Q;
|
cannam@167
|
365 T5A = T4Q + T4P;
|
cannam@167
|
366 }
|
cannam@167
|
367 {
|
cannam@167
|
368 E T2D, T2E, T4G, T4J;
|
cannam@167
|
369 T2D = FMA(KP668178637, T2p, T2s);
|
cannam@167
|
370 T2E = FMA(KP668178637, T2g, T2j);
|
cannam@167
|
371 T2F = T2D + T2E;
|
cannam@167
|
372 T3o = T2D - T2E;
|
cannam@167
|
373 T4G = FMA(KP198912367, T4F, T4E);
|
cannam@167
|
374 T4J = FNMS(KP198912367, T4I, T4H);
|
cannam@167
|
375 T4K = T4G - T4J;
|
cannam@167
|
376 T5D = T4G + T4J;
|
cannam@167
|
377 }
|
cannam@167
|
378 }
|
cannam@167
|
379 {
|
cannam@167
|
380 E T1x, T4p, T1A, T4o, T1G, T4m, T1J, T4l;
|
cannam@167
|
381 {
|
cannam@167
|
382 E T1t, T1y, T1w, T1z, T1u, T1v;
|
cannam@167
|
383 T1t = R1[WS(rs, 14)];
|
cannam@167
|
384 T1y = R1[WS(rs, 30)];
|
cannam@167
|
385 T1u = R1[WS(rs, 6)];
|
cannam@167
|
386 T1v = R1[WS(rs, 22)];
|
cannam@167
|
387 T1w = T1u + T1v;
|
cannam@167
|
388 T1z = T1u - T1v;
|
cannam@167
|
389 T1x = FNMS(KP707106781, T1w, T1t);
|
cannam@167
|
390 T4p = FMS(KP707106781, T1z, T1y);
|
cannam@167
|
391 T1A = FMA(KP707106781, T1z, T1y);
|
cannam@167
|
392 T4o = FMA(KP707106781, T1w, T1t);
|
cannam@167
|
393 }
|
cannam@167
|
394 {
|
cannam@167
|
395 E T1C, T1H, T1F, T1I, T1D, T1E;
|
cannam@167
|
396 T1C = R1[WS(rs, 18)];
|
cannam@167
|
397 T1H = R1[WS(rs, 2)];
|
cannam@167
|
398 T1D = R1[WS(rs, 10)];
|
cannam@167
|
399 T1E = R1[WS(rs, 26)];
|
cannam@167
|
400 T1F = T1D + T1E;
|
cannam@167
|
401 T1I = T1D - T1E;
|
cannam@167
|
402 T1G = FNMS(KP707106781, T1F, T1C);
|
cannam@167
|
403 T4m = FMA(KP707106781, T1I, T1H);
|
cannam@167
|
404 T1J = FNMS(KP707106781, T1I, T1H);
|
cannam@167
|
405 T4l = FMA(KP707106781, T1F, T1C);
|
cannam@167
|
406 }
|
cannam@167
|
407 {
|
cannam@167
|
408 E T1B, T1K, T4w, T4x;
|
cannam@167
|
409 T1B = FNMS(KP668178637, T1A, T1x);
|
cannam@167
|
410 T1K = FNMS(KP668178637, T1J, T1G);
|
cannam@167
|
411 T1L = T1B - T1K;
|
cannam@167
|
412 T3k = T1K + T1B;
|
cannam@167
|
413 T4w = FMA(KP198912367, T4o, T4p);
|
cannam@167
|
414 T4x = FNMS(KP198912367, T4l, T4m);
|
cannam@167
|
415 T4y = T4w - T4x;
|
cannam@167
|
416 T5t = T4x + T4w;
|
cannam@167
|
417 }
|
cannam@167
|
418 {
|
cannam@167
|
419 E T1U, T1V, T4n, T4q;
|
cannam@167
|
420 T1U = FMA(KP668178637, T1G, T1J);
|
cannam@167
|
421 T1V = FMA(KP668178637, T1x, T1A);
|
cannam@167
|
422 T1W = T1U + T1V;
|
cannam@167
|
423 T3h = T1U - T1V;
|
cannam@167
|
424 T4n = FMA(KP198912367, T4m, T4l);
|
cannam@167
|
425 T4q = FNMS(KP198912367, T4p, T4o);
|
cannam@167
|
426 T4r = T4n - T4q;
|
cannam@167
|
427 T5w = T4n + T4q;
|
cannam@167
|
428 }
|
cannam@167
|
429 }
|
cannam@167
|
430 {
|
cannam@167
|
431 E Tx, T2N, T6P, T6V, T1e, T6Q, T2X, T31, T1Y, T2L, T2Q, T6W, T2U, T30, T2H;
|
cannam@167
|
432 E T2K, Tw, T6O;
|
cannam@167
|
433 Tw = Tm - Tv;
|
cannam@167
|
434 Tx = FNMS(KP831469612, Tw, Td);
|
cannam@167
|
435 T2N = FMA(KP831469612, Tw, Td);
|
cannam@167
|
436 T6O = T34 + T35;
|
cannam@167
|
437 T6P = FMA(KP831469612, T6O, T6N);
|
cannam@167
|
438 T6V = FNMS(KP831469612, T6O, T6N);
|
cannam@167
|
439 {
|
cannam@167
|
440 E TS, T1d, T2V, T2W;
|
cannam@167
|
441 TS = FMA(KP534511135, TR, TK);
|
cannam@167
|
442 T1d = FMA(KP534511135, T1c, T15);
|
cannam@167
|
443 T1e = TS - T1d;
|
cannam@167
|
444 T6Q = TS + T1d;
|
cannam@167
|
445 T2V = FNMS(KP831469612, T2u, T2b);
|
cannam@167
|
446 T2W = FMA(KP831469612, T2F, T2C);
|
cannam@167
|
447 T2X = FNMS(KP250486960, T2W, T2V);
|
cannam@167
|
448 T31 = FMA(KP250486960, T2V, T2W);
|
cannam@167
|
449 }
|
cannam@167
|
450 {
|
cannam@167
|
451 E T1M, T1X, T2O, T2P;
|
cannam@167
|
452 T1M = FNMS(KP831469612, T1L, T1s);
|
cannam@167
|
453 T1X = FNMS(KP831469612, T1W, T1T);
|
cannam@167
|
454 T1Y = FMA(KP599376933, T1X, T1M);
|
cannam@167
|
455 T2L = FNMS(KP599376933, T1M, T1X);
|
cannam@167
|
456 T2O = FNMS(KP534511135, TK, TR);
|
cannam@167
|
457 T2P = FNMS(KP534511135, T15, T1c);
|
cannam@167
|
458 T2Q = T2O - T2P;
|
cannam@167
|
459 T6W = T2O + T2P;
|
cannam@167
|
460 }
|
cannam@167
|
461 {
|
cannam@167
|
462 E T2S, T2T, T2v, T2G;
|
cannam@167
|
463 T2S = FMA(KP831469612, T1L, T1s);
|
cannam@167
|
464 T2T = FMA(KP831469612, T1W, T1T);
|
cannam@167
|
465 T2U = FNMS(KP250486960, T2T, T2S);
|
cannam@167
|
466 T30 = FMA(KP250486960, T2S, T2T);
|
cannam@167
|
467 T2v = FMA(KP831469612, T2u, T2b);
|
cannam@167
|
468 T2G = FNMS(KP831469612, T2F, T2C);
|
cannam@167
|
469 T2H = FMA(KP599376933, T2G, T2v);
|
cannam@167
|
470 T2K = FNMS(KP599376933, T2v, T2G);
|
cannam@167
|
471 }
|
cannam@167
|
472 {
|
cannam@167
|
473 E T1f, T2I, T6X, T6Y;
|
cannam@167
|
474 T1f = FMA(KP881921264, T1e, Tx);
|
cannam@167
|
475 T2I = T1Y - T2H;
|
cannam@167
|
476 Cr[WS(csr, 26)] = FNMS(KP857728610, T2I, T1f);
|
cannam@167
|
477 Cr[WS(csr, 5)] = FMA(KP857728610, T2I, T1f);
|
cannam@167
|
478 T6X = FNMS(KP881921264, T6W, T6V);
|
cannam@167
|
479 T6Y = T2L + T2K;
|
cannam@167
|
480 Ci[WS(csi, 26)] = FMS(KP857728610, T6Y, T6X);
|
cannam@167
|
481 Ci[WS(csi, 5)] = FMA(KP857728610, T6Y, T6X);
|
cannam@167
|
482 }
|
cannam@167
|
483 {
|
cannam@167
|
484 E T2J, T2M, T6Z, T70;
|
cannam@167
|
485 T2J = FNMS(KP881921264, T1e, Tx);
|
cannam@167
|
486 T2M = T2K - T2L;
|
cannam@167
|
487 Cr[WS(csr, 21)] = FNMS(KP857728610, T2M, T2J);
|
cannam@167
|
488 Cr[WS(csr, 10)] = FMA(KP857728610, T2M, T2J);
|
cannam@167
|
489 T6Z = FMA(KP881921264, T6W, T6V);
|
cannam@167
|
490 T70 = T1Y + T2H;
|
cannam@167
|
491 Ci[WS(csi, 10)] = -(FMA(KP857728610, T70, T6Z));
|
cannam@167
|
492 Ci[WS(csi, 21)] = FNMS(KP857728610, T70, T6Z);
|
cannam@167
|
493 }
|
cannam@167
|
494 {
|
cannam@167
|
495 E T2R, T2Y, T6R, T6S;
|
cannam@167
|
496 T2R = FMA(KP881921264, T2Q, T2N);
|
cannam@167
|
497 T2Y = T2U - T2X;
|
cannam@167
|
498 Cr[WS(csr, 29)] = FNMS(KP970031253, T2Y, T2R);
|
cannam@167
|
499 Cr[WS(csr, 2)] = FMA(KP970031253, T2Y, T2R);
|
cannam@167
|
500 T6R = FMA(KP881921264, T6Q, T6P);
|
cannam@167
|
501 T6S = T30 + T31;
|
cannam@167
|
502 Ci[WS(csi, 2)] = -(FMA(KP970031253, T6S, T6R));
|
cannam@167
|
503 Ci[WS(csi, 29)] = FNMS(KP970031253, T6S, T6R);
|
cannam@167
|
504 }
|
cannam@167
|
505 {
|
cannam@167
|
506 E T2Z, T32, T6T, T6U;
|
cannam@167
|
507 T2Z = FNMS(KP881921264, T2Q, T2N);
|
cannam@167
|
508 T32 = T30 - T31;
|
cannam@167
|
509 Cr[WS(csr, 18)] = FNMS(KP970031253, T32, T2Z);
|
cannam@167
|
510 Cr[WS(csr, 13)] = FMA(KP970031253, T32, T2Z);
|
cannam@167
|
511 T6T = FNMS(KP881921264, T6Q, T6P);
|
cannam@167
|
512 T6U = T2U + T2X;
|
cannam@167
|
513 Ci[WS(csi, 18)] = -(FMA(KP970031253, T6U, T6T));
|
cannam@167
|
514 Ci[WS(csi, 13)] = FNMS(KP970031253, T6U, T6T);
|
cannam@167
|
515 }
|
cannam@167
|
516 }
|
cannam@167
|
517 {
|
cannam@167
|
518 E T5j, T5L, T67, T6d, T5q, T68, T5V, T5Z, T5y, T5J, T5O, T6e, T5S, T5Y, T5F;
|
cannam@167
|
519 E T5I, T5i, T66;
|
cannam@167
|
520 T5i = T5g - T5h;
|
cannam@167
|
521 T5j = FNMS(KP980785280, T5i, T5f);
|
cannam@167
|
522 T5L = FMA(KP980785280, T5i, T5f);
|
cannam@167
|
523 T66 = T3W + T3Z;
|
cannam@167
|
524 T67 = FMA(KP980785280, T66, T65);
|
cannam@167
|
525 T6d = FNMS(KP980785280, T66, T65);
|
cannam@167
|
526 {
|
cannam@167
|
527 E T5m, T5p, T5T, T5U;
|
cannam@167
|
528 T5m = FMA(KP098491403, T5l, T5k);
|
cannam@167
|
529 T5p = FMA(KP098491403, T5o, T5n);
|
cannam@167
|
530 T5q = T5m - T5p;
|
cannam@167
|
531 T68 = T5m + T5p;
|
cannam@167
|
532 T5T = FMA(KP980785280, T5A, T5z);
|
cannam@167
|
533 T5U = FMA(KP980785280, T5D, T5C);
|
cannam@167
|
534 T5V = FMA(KP049126849, T5U, T5T);
|
cannam@167
|
535 T5Z = FNMS(KP049126849, T5T, T5U);
|
cannam@167
|
536 }
|
cannam@167
|
537 {
|
cannam@167
|
538 E T5u, T5x, T5M, T5N;
|
cannam@167
|
539 T5u = FNMS(KP980785280, T5t, T5s);
|
cannam@167
|
540 T5x = FNMS(KP980785280, T5w, T5v);
|
cannam@167
|
541 T5y = FMA(KP906347169, T5x, T5u);
|
cannam@167
|
542 T5J = FNMS(KP906347169, T5u, T5x);
|
cannam@167
|
543 T5M = FNMS(KP098491403, T5k, T5l);
|
cannam@167
|
544 T5N = FNMS(KP098491403, T5n, T5o);
|
cannam@167
|
545 T5O = T5M - T5N;
|
cannam@167
|
546 T6e = T5M + T5N;
|
cannam@167
|
547 }
|
cannam@167
|
548 {
|
cannam@167
|
549 E T5Q, T5R, T5B, T5E;
|
cannam@167
|
550 T5Q = FMA(KP980785280, T5t, T5s);
|
cannam@167
|
551 T5R = FMA(KP980785280, T5w, T5v);
|
cannam@167
|
552 T5S = FNMS(KP049126849, T5R, T5Q);
|
cannam@167
|
553 T5Y = FMA(KP049126849, T5Q, T5R);
|
cannam@167
|
554 T5B = FNMS(KP980785280, T5A, T5z);
|
cannam@167
|
555 T5E = FNMS(KP980785280, T5D, T5C);
|
cannam@167
|
556 T5F = FNMS(KP906347169, T5E, T5B);
|
cannam@167
|
557 T5I = FMA(KP906347169, T5B, T5E);
|
cannam@167
|
558 }
|
cannam@167
|
559 {
|
cannam@167
|
560 E T5r, T5G, T6f, T6g;
|
cannam@167
|
561 T5r = FMA(KP995184726, T5q, T5j);
|
cannam@167
|
562 T5G = T5y + T5F;
|
cannam@167
|
563 Cr[WS(csr, 24)] = FNMS(KP740951125, T5G, T5r);
|
cannam@167
|
564 Cr[WS(csr, 7)] = FMA(KP740951125, T5G, T5r);
|
cannam@167
|
565 T6f = FNMS(KP995184726, T6e, T6d);
|
cannam@167
|
566 T6g = T5J + T5I;
|
cannam@167
|
567 Ci[WS(csi, 24)] = FMS(KP740951125, T6g, T6f);
|
cannam@167
|
568 Ci[WS(csi, 7)] = FMA(KP740951125, T6g, T6f);
|
cannam@167
|
569 }
|
cannam@167
|
570 {
|
cannam@167
|
571 E T5H, T5K, T6h, T6i;
|
cannam@167
|
572 T5H = FNMS(KP995184726, T5q, T5j);
|
cannam@167
|
573 T5K = T5I - T5J;
|
cannam@167
|
574 Cr[WS(csr, 23)] = FNMS(KP740951125, T5K, T5H);
|
cannam@167
|
575 Cr[WS(csr, 8)] = FMA(KP740951125, T5K, T5H);
|
cannam@167
|
576 T6h = FMA(KP995184726, T6e, T6d);
|
cannam@167
|
577 T6i = T5F - T5y;
|
cannam@167
|
578 Ci[WS(csi, 8)] = FMS(KP740951125, T6i, T6h);
|
cannam@167
|
579 Ci[WS(csi, 23)] = FMA(KP740951125, T6i, T6h);
|
cannam@167
|
580 }
|
cannam@167
|
581 {
|
cannam@167
|
582 E T5P, T5W, T69, T6a;
|
cannam@167
|
583 T5P = FMA(KP995184726, T5O, T5L);
|
cannam@167
|
584 T5W = T5S + T5V;
|
cannam@167
|
585 Cr[WS(csr, 31)] = FNMS(KP998795456, T5W, T5P);
|
cannam@167
|
586 Cr[0] = FMA(KP998795456, T5W, T5P);
|
cannam@167
|
587 T69 = FMA(KP995184726, T68, T67);
|
cannam@167
|
588 T6a = T5Y + T5Z;
|
cannam@167
|
589 Ci[0] = -(FMA(KP998795456, T6a, T69));
|
cannam@167
|
590 Ci[WS(csi, 31)] = FNMS(KP998795456, T6a, T69);
|
cannam@167
|
591 }
|
cannam@167
|
592 {
|
cannam@167
|
593 E T5X, T60, T6b, T6c;
|
cannam@167
|
594 T5X = FNMS(KP995184726, T5O, T5L);
|
cannam@167
|
595 T60 = T5Y - T5Z;
|
cannam@167
|
596 Cr[WS(csr, 16)] = FNMS(KP998795456, T60, T5X);
|
cannam@167
|
597 Cr[WS(csr, 15)] = FMA(KP998795456, T60, T5X);
|
cannam@167
|
598 T6b = FNMS(KP995184726, T68, T67);
|
cannam@167
|
599 T6c = T5V - T5S;
|
cannam@167
|
600 Ci[WS(csi, 16)] = FMS(KP998795456, T6c, T6b);
|
cannam@167
|
601 Ci[WS(csi, 15)] = FMA(KP998795456, T6c, T6b);
|
cannam@167
|
602 }
|
cannam@167
|
603 }
|
cannam@167
|
604 {
|
cannam@167
|
605 E T37, T3z, T6B, T6H, T3e, T6C, T3J, T3M, T3m, T3w, T3C, T6I, T3G, T3N, T3t;
|
cannam@167
|
606 E T3x, T36, T6A;
|
cannam@167
|
607 T36 = T34 - T35;
|
cannam@167
|
608 T37 = FNMS(KP831469612, T36, T33);
|
cannam@167
|
609 T3z = FMA(KP831469612, T36, T33);
|
cannam@167
|
610 T6A = Tv + Tm;
|
cannam@167
|
611 T6B = FMA(KP831469612, T6A, T6z);
|
cannam@167
|
612 T6H = FNMS(KP831469612, T6A, T6z);
|
cannam@167
|
613 {
|
cannam@167
|
614 E T3a, T3d, T3H, T3I;
|
cannam@167
|
615 T3a = FNMS(KP303346683, T39, T38);
|
cannam@167
|
616 T3d = FNMS(KP303346683, T3c, T3b);
|
cannam@167
|
617 T3e = T3a - T3d;
|
cannam@167
|
618 T6C = T3d + T3a;
|
cannam@167
|
619 T3H = FNMS(KP831469612, T3o, T3n);
|
cannam@167
|
620 T3I = FMA(KP831469612, T3r, T3q);
|
cannam@167
|
621 T3J = FMA(KP148335987, T3I, T3H);
|
cannam@167
|
622 T3M = FNMS(KP148335987, T3H, T3I);
|
cannam@167
|
623 }
|
cannam@167
|
624 {
|
cannam@167
|
625 E T3i, T3l, T3A, T3B;
|
cannam@167
|
626 T3i = FNMS(KP831469612, T3h, T3g);
|
cannam@167
|
627 T3l = FNMS(KP831469612, T3k, T3j);
|
cannam@167
|
628 T3m = FNMS(KP741650546, T3l, T3i);
|
cannam@167
|
629 T3w = FMA(KP741650546, T3i, T3l);
|
cannam@167
|
630 T3A = FMA(KP303346683, T3b, T3c);
|
cannam@167
|
631 T3B = FMA(KP303346683, T38, T39);
|
cannam@167
|
632 T3C = T3A - T3B;
|
cannam@167
|
633 T6I = T3A + T3B;
|
cannam@167
|
634 }
|
cannam@167
|
635 {
|
cannam@167
|
636 E T3E, T3F, T3p, T3s;
|
cannam@167
|
637 T3E = FMA(KP831469612, T3h, T3g);
|
cannam@167
|
638 T3F = FMA(KP831469612, T3k, T3j);
|
cannam@167
|
639 T3G = FMA(KP148335987, T3F, T3E);
|
cannam@167
|
640 T3N = FNMS(KP148335987, T3E, T3F);
|
cannam@167
|
641 T3p = FMA(KP831469612, T3o, T3n);
|
cannam@167
|
642 T3s = FNMS(KP831469612, T3r, T3q);
|
cannam@167
|
643 T3t = FNMS(KP741650546, T3s, T3p);
|
cannam@167
|
644 T3x = FMA(KP741650546, T3p, T3s);
|
cannam@167
|
645 }
|
cannam@167
|
646 {
|
cannam@167
|
647 E T3f, T3u, T6J, T6K;
|
cannam@167
|
648 T3f = FMA(KP956940335, T3e, T37);
|
cannam@167
|
649 T3u = T3m - T3t;
|
cannam@167
|
650 Cr[WS(csr, 25)] = FNMS(KP803207531, T3u, T3f);
|
cannam@167
|
651 Cr[WS(csr, 6)] = FMA(KP803207531, T3u, T3f);
|
cannam@167
|
652 T6J = FMA(KP956940335, T6I, T6H);
|
cannam@167
|
653 T6K = T3w + T3x;
|
cannam@167
|
654 Ci[WS(csi, 6)] = -(FMA(KP803207531, T6K, T6J));
|
cannam@167
|
655 Ci[WS(csi, 25)] = FNMS(KP803207531, T6K, T6J);
|
cannam@167
|
656 }
|
cannam@167
|
657 {
|
cannam@167
|
658 E T3v, T3y, T6L, T6M;
|
cannam@167
|
659 T3v = FNMS(KP956940335, T3e, T37);
|
cannam@167
|
660 T3y = T3w - T3x;
|
cannam@167
|
661 Cr[WS(csr, 22)] = FNMS(KP803207531, T3y, T3v);
|
cannam@167
|
662 Cr[WS(csr, 9)] = FMA(KP803207531, T3y, T3v);
|
cannam@167
|
663 T6L = FNMS(KP956940335, T6I, T6H);
|
cannam@167
|
664 T6M = T3m + T3t;
|
cannam@167
|
665 Ci[WS(csi, 22)] = -(FMA(KP803207531, T6M, T6L));
|
cannam@167
|
666 Ci[WS(csi, 9)] = FNMS(KP803207531, T6M, T6L);
|
cannam@167
|
667 }
|
cannam@167
|
668 {
|
cannam@167
|
669 E T3D, T3K, T6D, T6E;
|
cannam@167
|
670 T3D = FMA(KP956940335, T3C, T3z);
|
cannam@167
|
671 T3K = T3G - T3J;
|
cannam@167
|
672 Cr[WS(csr, 30)] = FNMS(KP989176509, T3K, T3D);
|
cannam@167
|
673 Cr[WS(csr, 1)] = FMA(KP989176509, T3K, T3D);
|
cannam@167
|
674 T6D = FMA(KP956940335, T6C, T6B);
|
cannam@167
|
675 T6E = T3N + T3M;
|
cannam@167
|
676 Ci[WS(csi, 30)] = FMS(KP989176509, T6E, T6D);
|
cannam@167
|
677 Ci[WS(csi, 1)] = FMA(KP989176509, T6E, T6D);
|
cannam@167
|
678 }
|
cannam@167
|
679 {
|
cannam@167
|
680 E T3L, T3O, T6F, T6G;
|
cannam@167
|
681 T3L = FNMS(KP956940335, T3C, T3z);
|
cannam@167
|
682 T3O = T3M - T3N;
|
cannam@167
|
683 Cr[WS(csr, 17)] = FNMS(KP989176509, T3O, T3L);
|
cannam@167
|
684 Cr[WS(csr, 14)] = FMA(KP989176509, T3O, T3L);
|
cannam@167
|
685 T6F = FNMS(KP956940335, T6C, T6B);
|
cannam@167
|
686 T6G = T3G + T3J;
|
cannam@167
|
687 Ci[WS(csi, 14)] = -(FMA(KP989176509, T6G, T6F));
|
cannam@167
|
688 Ci[WS(csi, 17)] = FNMS(KP989176509, T6G, T6F);
|
cannam@167
|
689 }
|
cannam@167
|
690 }
|
cannam@167
|
691 {
|
cannam@167
|
692 E T41, T4Z, T6l, T6r, T4g, T6m, T59, T5c, T4A, T4W, T52, T6s, T56, T5d, T4T;
|
cannam@167
|
693 E T4X, T40, T6k;
|
cannam@167
|
694 T40 = T3W - T3Z;
|
cannam@167
|
695 T41 = FNMS(KP980785280, T40, T3T);
|
cannam@167
|
696 T4Z = FMA(KP980785280, T40, T3T);
|
cannam@167
|
697 T6k = T5g + T5h;
|
cannam@167
|
698 T6l = FNMS(KP980785280, T6k, T6j);
|
cannam@167
|
699 T6r = FMA(KP980785280, T6k, T6j);
|
cannam@167
|
700 {
|
cannam@167
|
701 E T48, T4f, T57, T58;
|
cannam@167
|
702 T48 = FNMS(KP820678790, T47, T44);
|
cannam@167
|
703 T4f = FNMS(KP820678790, T4e, T4b);
|
cannam@167
|
704 T4g = T48 - T4f;
|
cannam@167
|
705 T6m = T4f + T48;
|
cannam@167
|
706 T57 = FMA(KP980785280, T4K, T4D);
|
cannam@167
|
707 T58 = FMA(KP980785280, T4R, T4O);
|
cannam@167
|
708 T59 = FNMS(KP357805721, T58, T57);
|
cannam@167
|
709 T5c = FMA(KP357805721, T57, T58);
|
cannam@167
|
710 }
|
cannam@167
|
711 {
|
cannam@167
|
712 E T4s, T4z, T50, T51;
|
cannam@167
|
713 T4s = FNMS(KP980785280, T4r, T4k);
|
cannam@167
|
714 T4z = FNMS(KP980785280, T4y, T4v);
|
cannam@167
|
715 T4A = FNMS(KP472964775, T4z, T4s);
|
cannam@167
|
716 T4W = FMA(KP472964775, T4s, T4z);
|
cannam@167
|
717 T50 = FMA(KP820678790, T4b, T4e);
|
cannam@167
|
718 T51 = FMA(KP820678790, T44, T47);
|
cannam@167
|
719 T52 = T50 - T51;
|
cannam@167
|
720 T6s = T50 + T51;
|
cannam@167
|
721 }
|
cannam@167
|
722 {
|
cannam@167
|
723 E T54, T55, T4L, T4S;
|
cannam@167
|
724 T54 = FMA(KP980785280, T4r, T4k);
|
cannam@167
|
725 T55 = FMA(KP980785280, T4y, T4v);
|
cannam@167
|
726 T56 = FMA(KP357805721, T55, T54);
|
cannam@167
|
727 T5d = FNMS(KP357805721, T54, T55);
|
cannam@167
|
728 T4L = FNMS(KP980785280, T4K, T4D);
|
cannam@167
|
729 T4S = FNMS(KP980785280, T4R, T4O);
|
cannam@167
|
730 T4T = FMA(KP472964775, T4S, T4L);
|
cannam@167
|
731 T4X = FNMS(KP472964775, T4L, T4S);
|
cannam@167
|
732 }
|
cannam@167
|
733 {
|
cannam@167
|
734 E T4h, T4U, T6t, T6u;
|
cannam@167
|
735 T4h = FMA(KP773010453, T4g, T41);
|
cannam@167
|
736 T4U = T4A + T4T;
|
cannam@167
|
737 Cr[WS(csr, 27)] = FNMS(KP903989293, T4U, T4h);
|
cannam@167
|
738 Cr[WS(csr, 4)] = FMA(KP903989293, T4U, T4h);
|
cannam@167
|
739 T6t = FMA(KP773010453, T6s, T6r);
|
cannam@167
|
740 T6u = T4W + T4X;
|
cannam@167
|
741 Ci[WS(csi, 4)] = -(FMA(KP903989293, T6u, T6t));
|
cannam@167
|
742 Ci[WS(csi, 27)] = FNMS(KP903989293, T6u, T6t);
|
cannam@167
|
743 }
|
cannam@167
|
744 {
|
cannam@167
|
745 E T4V, T4Y, T6v, T6w;
|
cannam@167
|
746 T4V = FNMS(KP773010453, T4g, T41);
|
cannam@167
|
747 T4Y = T4W - T4X;
|
cannam@167
|
748 Cr[WS(csr, 20)] = FNMS(KP903989293, T4Y, T4V);
|
cannam@167
|
749 Cr[WS(csr, 11)] = FMA(KP903989293, T4Y, T4V);
|
cannam@167
|
750 T6v = FNMS(KP773010453, T6s, T6r);
|
cannam@167
|
751 T6w = T4T - T4A;
|
cannam@167
|
752 Ci[WS(csi, 20)] = FMS(KP903989293, T6w, T6v);
|
cannam@167
|
753 Ci[WS(csi, 11)] = FMA(KP903989293, T6w, T6v);
|
cannam@167
|
754 }
|
cannam@167
|
755 {
|
cannam@167
|
756 E T53, T5a, T6n, T6o;
|
cannam@167
|
757 T53 = FMA(KP773010453, T52, T4Z);
|
cannam@167
|
758 T5a = T56 + T59;
|
cannam@167
|
759 Cr[WS(csr, 28)] = FNMS(KP941544065, T5a, T53);
|
cannam@167
|
760 Cr[WS(csr, 3)] = FMA(KP941544065, T5a, T53);
|
cannam@167
|
761 T6n = FMA(KP773010453, T6m, T6l);
|
cannam@167
|
762 T6o = T5d + T5c;
|
cannam@167
|
763 Ci[WS(csi, 28)] = FMS(KP941544065, T6o, T6n);
|
cannam@167
|
764 Ci[WS(csi, 3)] = FMA(KP941544065, T6o, T6n);
|
cannam@167
|
765 }
|
cannam@167
|
766 {
|
cannam@167
|
767 E T5b, T5e, T6p, T6q;
|
cannam@167
|
768 T5b = FNMS(KP773010453, T52, T4Z);
|
cannam@167
|
769 T5e = T5c - T5d;
|
cannam@167
|
770 Cr[WS(csr, 19)] = FNMS(KP941544065, T5e, T5b);
|
cannam@167
|
771 Cr[WS(csr, 12)] = FMA(KP941544065, T5e, T5b);
|
cannam@167
|
772 T6p = FNMS(KP773010453, T6m, T6l);
|
cannam@167
|
773 T6q = T59 - T56;
|
cannam@167
|
774 Ci[WS(csi, 12)] = FMS(KP941544065, T6q, T6p);
|
cannam@167
|
775 Ci[WS(csi, 19)] = FMA(KP941544065, T6q, T6p);
|
cannam@167
|
776 }
|
cannam@167
|
777 }
|
cannam@167
|
778 }
|
cannam@167
|
779 }
|
cannam@167
|
780 }
|
cannam@167
|
781
|
cannam@167
|
782 static const kr2c_desc desc = { 64, "r2cfII_64", {114, 0, 320, 0}, &GENUS };
|
cannam@167
|
783
|
cannam@167
|
784 void X(codelet_r2cfII_64) (planner *p) {
|
cannam@167
|
785 X(kr2c_register) (p, r2cfII_64, &desc);
|
cannam@167
|
786 }
|
cannam@167
|
787
|
cannam@167
|
788 #else
|
cannam@167
|
789
|
cannam@167
|
790 /* Generated by: ../../../genfft/gen_r2cf.native -compact -variables 4 -pipeline-latency 4 -n 64 -name r2cfII_64 -dft-II -include rdft/scalar/r2cfII.h */
|
cannam@167
|
791
|
cannam@167
|
792 /*
|
cannam@167
|
793 * This function contains 434 FP additions, 206 FP multiplications,
|
cannam@167
|
794 * (or, 342 additions, 114 multiplications, 92 fused multiply/add),
|
cannam@167
|
795 * 118 stack variables, 31 constants, and 128 memory accesses
|
cannam@167
|
796 */
|
cannam@167
|
797 #include "rdft/scalar/r2cfII.h"
|
cannam@167
|
798
|
cannam@167
|
799 static void r2cfII_64(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
|
cannam@167
|
800 {
|
cannam@167
|
801 DK(KP242980179, +0.242980179903263889948274162077471118320990783);
|
cannam@167
|
802 DK(KP970031253, +0.970031253194543992603984207286100251456865962);
|
cannam@167
|
803 DK(KP857728610, +0.857728610000272069902269984284770137042490799);
|
cannam@167
|
804 DK(KP514102744, +0.514102744193221726593693838968815772608049120);
|
cannam@167
|
805 DK(KP471396736, +0.471396736825997648556387625905254377657460319);
|
cannam@167
|
806 DK(KP881921264, +0.881921264348355029712756863660388349508442621);
|
cannam@167
|
807 DK(KP427555093, +0.427555093430282094320966856888798534304578629);
|
cannam@167
|
808 DK(KP903989293, +0.903989293123443331586200297230537048710132025);
|
cannam@167
|
809 DK(KP336889853, +0.336889853392220050689253212619147570477766780);
|
cannam@167
|
810 DK(KP941544065, +0.941544065183020778412509402599502357185589796);
|
cannam@167
|
811 DK(KP773010453, +0.773010453362736960810906609758469800971041293);
|
cannam@167
|
812 DK(KP634393284, +0.634393284163645498215171613225493370675687095);
|
cannam@167
|
813 DK(KP595699304, +0.595699304492433343467036528829969889511926338);
|
cannam@167
|
814 DK(KP803207531, +0.803207531480644909806676512963141923879569427);
|
cannam@167
|
815 DK(KP146730474, +0.146730474455361751658850129646717819706215317);
|
cannam@167
|
816 DK(KP989176509, +0.989176509964780973451673738016243063983689533);
|
cannam@167
|
817 DK(KP956940335, +0.956940335732208864935797886980269969482849206);
|
cannam@167
|
818 DK(KP290284677, +0.290284677254462367636192375817395274691476278);
|
cannam@167
|
819 DK(KP049067674, +0.049067674327418014254954976942682658314745363);
|
cannam@167
|
820 DK(KP998795456, +0.998795456205172392714771604759100694443203615);
|
cannam@167
|
821 DK(KP671558954, +0.671558954847018400625376850427421803228750632);
|
cannam@167
|
822 DK(KP740951125, +0.740951125354959091175616897495162729728955309);
|
cannam@167
|
823 DK(KP098017140, +0.098017140329560601994195563888641845861136673);
|
cannam@167
|
824 DK(KP995184726, +0.995184726672196886244836953109479921575474869);
|
cannam@167
|
825 DK(KP382683432, +0.382683432365089771728459984030398866761344562);
|
cannam@167
|
826 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
cannam@167
|
827 DK(KP555570233, +0.555570233019602224742830813948532874374937191);
|
cannam@167
|
828 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
cannam@167
|
829 DK(KP195090322, +0.195090322016128267848284868477022240927691618);
|
cannam@167
|
830 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
cannam@167
|
831 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
cannam@167
|
832 {
|
cannam@167
|
833 INT i;
|
cannam@167
|
834 for (i = v; i > 0; i = i - 1, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(256, rs), MAKE_VOLATILE_STRIDE(256, csr), MAKE_VOLATILE_STRIDE(256, csi)) {
|
cannam@167
|
835 E Tm, T34, T3Z, T5g, Tv, T35, T3W, T5h, Td, T33, T6B, T6Q, T3T, T5f, T68;
|
cannam@167
|
836 E T6m, T2b, T3n, T4O, T5D, T2F, T3r, T4K, T5z, TK, T3c, T47, T5n, TR, T3b;
|
cannam@167
|
837 E T44, T5o, T15, T38, T4e, T5l, T1c, T39, T4b, T5k, T1s, T3g, T4v, T5w, T1W;
|
cannam@167
|
838 E T3k, T4k, T5s, T2u, T3q, T4R, T5A, T2y, T3o, T4H, T5C, T1L, T3j, T4y, T5t;
|
cannam@167
|
839 E T1P, T3h, T4r, T5v;
|
cannam@167
|
840 {
|
cannam@167
|
841 E Te, Tk, Th, Tj, Tf, Tg;
|
cannam@167
|
842 Te = R0[WS(rs, 2)];
|
cannam@167
|
843 Tk = R0[WS(rs, 18)];
|
cannam@167
|
844 Tf = R0[WS(rs, 10)];
|
cannam@167
|
845 Tg = R0[WS(rs, 26)];
|
cannam@167
|
846 Th = KP707106781 * (Tf - Tg);
|
cannam@167
|
847 Tj = KP707106781 * (Tf + Tg);
|
cannam@167
|
848 {
|
cannam@167
|
849 E Ti, Tl, T3X, T3Y;
|
cannam@167
|
850 Ti = Te + Th;
|
cannam@167
|
851 Tl = Tj + Tk;
|
cannam@167
|
852 Tm = FNMS(KP195090322, Tl, KP980785280 * Ti);
|
cannam@167
|
853 T34 = FMA(KP195090322, Ti, KP980785280 * Tl);
|
cannam@167
|
854 T3X = Tk - Tj;
|
cannam@167
|
855 T3Y = Te - Th;
|
cannam@167
|
856 T3Z = FNMS(KP555570233, T3Y, KP831469612 * T3X);
|
cannam@167
|
857 T5g = FMA(KP831469612, T3Y, KP555570233 * T3X);
|
cannam@167
|
858 }
|
cannam@167
|
859 }
|
cannam@167
|
860 {
|
cannam@167
|
861 E Tq, Tt, Tp, Ts, Tn, To;
|
cannam@167
|
862 Tq = R0[WS(rs, 30)];
|
cannam@167
|
863 Tt = R0[WS(rs, 14)];
|
cannam@167
|
864 Tn = R0[WS(rs, 6)];
|
cannam@167
|
865 To = R0[WS(rs, 22)];
|
cannam@167
|
866 Tp = KP707106781 * (Tn - To);
|
cannam@167
|
867 Ts = KP707106781 * (Tn + To);
|
cannam@167
|
868 {
|
cannam@167
|
869 E Tr, Tu, T3U, T3V;
|
cannam@167
|
870 Tr = Tp - Tq;
|
cannam@167
|
871 Tu = Ts + Tt;
|
cannam@167
|
872 Tv = FMA(KP980785280, Tr, KP195090322 * Tu);
|
cannam@167
|
873 T35 = FNMS(KP980785280, Tu, KP195090322 * Tr);
|
cannam@167
|
874 T3U = Tt - Ts;
|
cannam@167
|
875 T3V = Tp + Tq;
|
cannam@167
|
876 T3W = FNMS(KP555570233, T3V, KP831469612 * T3U);
|
cannam@167
|
877 T5h = FMA(KP831469612, T3V, KP555570233 * T3U);
|
cannam@167
|
878 }
|
cannam@167
|
879 }
|
cannam@167
|
880 {
|
cannam@167
|
881 E T1, T66, T4, T65, T8, T3Q, Tb, T3R, T2, T3;
|
cannam@167
|
882 T1 = R0[0];
|
cannam@167
|
883 T66 = R0[WS(rs, 16)];
|
cannam@167
|
884 T2 = R0[WS(rs, 8)];
|
cannam@167
|
885 T3 = R0[WS(rs, 24)];
|
cannam@167
|
886 T4 = KP707106781 * (T2 - T3);
|
cannam@167
|
887 T65 = KP707106781 * (T2 + T3);
|
cannam@167
|
888 {
|
cannam@167
|
889 E T6, T7, T9, Ta;
|
cannam@167
|
890 T6 = R0[WS(rs, 4)];
|
cannam@167
|
891 T7 = R0[WS(rs, 20)];
|
cannam@167
|
892 T8 = FNMS(KP382683432, T7, KP923879532 * T6);
|
cannam@167
|
893 T3Q = FMA(KP382683432, T6, KP923879532 * T7);
|
cannam@167
|
894 T9 = R0[WS(rs, 12)];
|
cannam@167
|
895 Ta = R0[WS(rs, 28)];
|
cannam@167
|
896 Tb = FNMS(KP923879532, Ta, KP382683432 * T9);
|
cannam@167
|
897 T3R = FMA(KP923879532, T9, KP382683432 * Ta);
|
cannam@167
|
898 }
|
cannam@167
|
899 {
|
cannam@167
|
900 E T5, Tc, T6z, T6A;
|
cannam@167
|
901 T5 = T1 + T4;
|
cannam@167
|
902 Tc = T8 + Tb;
|
cannam@167
|
903 Td = T5 + Tc;
|
cannam@167
|
904 T33 = T5 - Tc;
|
cannam@167
|
905 T6z = Tb - T8;
|
cannam@167
|
906 T6A = T66 - T65;
|
cannam@167
|
907 T6B = T6z - T6A;
|
cannam@167
|
908 T6Q = T6z + T6A;
|
cannam@167
|
909 }
|
cannam@167
|
910 {
|
cannam@167
|
911 E T3P, T3S, T64, T67;
|
cannam@167
|
912 T3P = T1 - T4;
|
cannam@167
|
913 T3S = T3Q - T3R;
|
cannam@167
|
914 T3T = T3P - T3S;
|
cannam@167
|
915 T5f = T3P + T3S;
|
cannam@167
|
916 T64 = T3Q + T3R;
|
cannam@167
|
917 T67 = T65 + T66;
|
cannam@167
|
918 T68 = T64 + T67;
|
cannam@167
|
919 T6m = T67 - T64;
|
cannam@167
|
920 }
|
cannam@167
|
921 }
|
cannam@167
|
922 {
|
cannam@167
|
923 E T22, T2D, T21, T2C, T26, T2z, T29, T2A, T1Z, T20;
|
cannam@167
|
924 T22 = R1[WS(rs, 31)];
|
cannam@167
|
925 T2D = R1[WS(rs, 15)];
|
cannam@167
|
926 T1Z = R1[WS(rs, 7)];
|
cannam@167
|
927 T20 = R1[WS(rs, 23)];
|
cannam@167
|
928 T21 = KP707106781 * (T1Z - T20);
|
cannam@167
|
929 T2C = KP707106781 * (T1Z + T20);
|
cannam@167
|
930 {
|
cannam@167
|
931 E T24, T25, T27, T28;
|
cannam@167
|
932 T24 = R1[WS(rs, 3)];
|
cannam@167
|
933 T25 = R1[WS(rs, 19)];
|
cannam@167
|
934 T26 = FNMS(KP382683432, T25, KP923879532 * T24);
|
cannam@167
|
935 T2z = FMA(KP382683432, T24, KP923879532 * T25);
|
cannam@167
|
936 T27 = R1[WS(rs, 11)];
|
cannam@167
|
937 T28 = R1[WS(rs, 27)];
|
cannam@167
|
938 T29 = FNMS(KP923879532, T28, KP382683432 * T27);
|
cannam@167
|
939 T2A = FMA(KP923879532, T27, KP382683432 * T28);
|
cannam@167
|
940 }
|
cannam@167
|
941 {
|
cannam@167
|
942 E T23, T2a, T4M, T4N;
|
cannam@167
|
943 T23 = T21 - T22;
|
cannam@167
|
944 T2a = T26 + T29;
|
cannam@167
|
945 T2b = T23 + T2a;
|
cannam@167
|
946 T3n = T23 - T2a;
|
cannam@167
|
947 T4M = T29 - T26;
|
cannam@167
|
948 T4N = T2D - T2C;
|
cannam@167
|
949 T4O = T4M - T4N;
|
cannam@167
|
950 T5D = T4M + T4N;
|
cannam@167
|
951 }
|
cannam@167
|
952 {
|
cannam@167
|
953 E T2B, T2E, T4I, T4J;
|
cannam@167
|
954 T2B = T2z + T2A;
|
cannam@167
|
955 T2E = T2C + T2D;
|
cannam@167
|
956 T2F = T2B + T2E;
|
cannam@167
|
957 T3r = T2E - T2B;
|
cannam@167
|
958 T4I = T21 + T22;
|
cannam@167
|
959 T4J = T2z - T2A;
|
cannam@167
|
960 T4K = T4I + T4J;
|
cannam@167
|
961 T5z = T4J - T4I;
|
cannam@167
|
962 }
|
cannam@167
|
963 }
|
cannam@167
|
964 {
|
cannam@167
|
965 E Ty, TP, TB, TO, TF, TL, TI, TM, Tz, TA;
|
cannam@167
|
966 Ty = R0[WS(rs, 1)];
|
cannam@167
|
967 TP = R0[WS(rs, 17)];
|
cannam@167
|
968 Tz = R0[WS(rs, 9)];
|
cannam@167
|
969 TA = R0[WS(rs, 25)];
|
cannam@167
|
970 TB = KP707106781 * (Tz - TA);
|
cannam@167
|
971 TO = KP707106781 * (Tz + TA);
|
cannam@167
|
972 {
|
cannam@167
|
973 E TD, TE, TG, TH;
|
cannam@167
|
974 TD = R0[WS(rs, 5)];
|
cannam@167
|
975 TE = R0[WS(rs, 21)];
|
cannam@167
|
976 TF = FNMS(KP382683432, TE, KP923879532 * TD);
|
cannam@167
|
977 TL = FMA(KP382683432, TD, KP923879532 * TE);
|
cannam@167
|
978 TG = R0[WS(rs, 13)];
|
cannam@167
|
979 TH = R0[WS(rs, 29)];
|
cannam@167
|
980 TI = FNMS(KP923879532, TH, KP382683432 * TG);
|
cannam@167
|
981 TM = FMA(KP923879532, TG, KP382683432 * TH);
|
cannam@167
|
982 }
|
cannam@167
|
983 {
|
cannam@167
|
984 E TC, TJ, T45, T46;
|
cannam@167
|
985 TC = Ty + TB;
|
cannam@167
|
986 TJ = TF + TI;
|
cannam@167
|
987 TK = TC + TJ;
|
cannam@167
|
988 T3c = TC - TJ;
|
cannam@167
|
989 T45 = TI - TF;
|
cannam@167
|
990 T46 = TP - TO;
|
cannam@167
|
991 T47 = T45 - T46;
|
cannam@167
|
992 T5n = T45 + T46;
|
cannam@167
|
993 }
|
cannam@167
|
994 {
|
cannam@167
|
995 E TN, TQ, T42, T43;
|
cannam@167
|
996 TN = TL + TM;
|
cannam@167
|
997 TQ = TO + TP;
|
cannam@167
|
998 TR = TN + TQ;
|
cannam@167
|
999 T3b = TQ - TN;
|
cannam@167
|
1000 T42 = Ty - TB;
|
cannam@167
|
1001 T43 = TL - TM;
|
cannam@167
|
1002 T44 = T42 - T43;
|
cannam@167
|
1003 T5o = T42 + T43;
|
cannam@167
|
1004 }
|
cannam@167
|
1005 }
|
cannam@167
|
1006 {
|
cannam@167
|
1007 E TW, T1a, TV, T19, T10, T16, T13, T17, TT, TU;
|
cannam@167
|
1008 TW = R0[WS(rs, 31)];
|
cannam@167
|
1009 T1a = R0[WS(rs, 15)];
|
cannam@167
|
1010 TT = R0[WS(rs, 7)];
|
cannam@167
|
1011 TU = R0[WS(rs, 23)];
|
cannam@167
|
1012 TV = KP707106781 * (TT - TU);
|
cannam@167
|
1013 T19 = KP707106781 * (TT + TU);
|
cannam@167
|
1014 {
|
cannam@167
|
1015 E TY, TZ, T11, T12;
|
cannam@167
|
1016 TY = R0[WS(rs, 3)];
|
cannam@167
|
1017 TZ = R0[WS(rs, 19)];
|
cannam@167
|
1018 T10 = FNMS(KP382683432, TZ, KP923879532 * TY);
|
cannam@167
|
1019 T16 = FMA(KP382683432, TY, KP923879532 * TZ);
|
cannam@167
|
1020 T11 = R0[WS(rs, 11)];
|
cannam@167
|
1021 T12 = R0[WS(rs, 27)];
|
cannam@167
|
1022 T13 = FNMS(KP923879532, T12, KP382683432 * T11);
|
cannam@167
|
1023 T17 = FMA(KP923879532, T11, KP382683432 * T12);
|
cannam@167
|
1024 }
|
cannam@167
|
1025 {
|
cannam@167
|
1026 E TX, T14, T4c, T4d;
|
cannam@167
|
1027 TX = TV - TW;
|
cannam@167
|
1028 T14 = T10 + T13;
|
cannam@167
|
1029 T15 = TX + T14;
|
cannam@167
|
1030 T38 = TX - T14;
|
cannam@167
|
1031 T4c = T13 - T10;
|
cannam@167
|
1032 T4d = T1a - T19;
|
cannam@167
|
1033 T4e = T4c - T4d;
|
cannam@167
|
1034 T5l = T4c + T4d;
|
cannam@167
|
1035 }
|
cannam@167
|
1036 {
|
cannam@167
|
1037 E T18, T1b, T49, T4a;
|
cannam@167
|
1038 T18 = T16 + T17;
|
cannam@167
|
1039 T1b = T19 + T1a;
|
cannam@167
|
1040 T1c = T18 + T1b;
|
cannam@167
|
1041 T39 = T1b - T18;
|
cannam@167
|
1042 T49 = TV + TW;
|
cannam@167
|
1043 T4a = T16 - T17;
|
cannam@167
|
1044 T4b = T49 + T4a;
|
cannam@167
|
1045 T5k = T4a - T49;
|
cannam@167
|
1046 }
|
cannam@167
|
1047 }
|
cannam@167
|
1048 {
|
cannam@167
|
1049 E T1g, T1U, T1j, T1T, T1n, T1Q, T1q, T1R, T1h, T1i;
|
cannam@167
|
1050 T1g = R1[0];
|
cannam@167
|
1051 T1U = R1[WS(rs, 16)];
|
cannam@167
|
1052 T1h = R1[WS(rs, 8)];
|
cannam@167
|
1053 T1i = R1[WS(rs, 24)];
|
cannam@167
|
1054 T1j = KP707106781 * (T1h - T1i);
|
cannam@167
|
1055 T1T = KP707106781 * (T1h + T1i);
|
cannam@167
|
1056 {
|
cannam@167
|
1057 E T1l, T1m, T1o, T1p;
|
cannam@167
|
1058 T1l = R1[WS(rs, 4)];
|
cannam@167
|
1059 T1m = R1[WS(rs, 20)];
|
cannam@167
|
1060 T1n = FNMS(KP382683432, T1m, KP923879532 * T1l);
|
cannam@167
|
1061 T1Q = FMA(KP382683432, T1l, KP923879532 * T1m);
|
cannam@167
|
1062 T1o = R1[WS(rs, 12)];
|
cannam@167
|
1063 T1p = R1[WS(rs, 28)];
|
cannam@167
|
1064 T1q = FNMS(KP923879532, T1p, KP382683432 * T1o);
|
cannam@167
|
1065 T1R = FMA(KP923879532, T1o, KP382683432 * T1p);
|
cannam@167
|
1066 }
|
cannam@167
|
1067 {
|
cannam@167
|
1068 E T1k, T1r, T4t, T4u;
|
cannam@167
|
1069 T1k = T1g + T1j;
|
cannam@167
|
1070 T1r = T1n + T1q;
|
cannam@167
|
1071 T1s = T1k + T1r;
|
cannam@167
|
1072 T3g = T1k - T1r;
|
cannam@167
|
1073 T4t = T1q - T1n;
|
cannam@167
|
1074 T4u = T1U - T1T;
|
cannam@167
|
1075 T4v = T4t - T4u;
|
cannam@167
|
1076 T5w = T4t + T4u;
|
cannam@167
|
1077 }
|
cannam@167
|
1078 {
|
cannam@167
|
1079 E T1S, T1V, T4i, T4j;
|
cannam@167
|
1080 T1S = T1Q + T1R;
|
cannam@167
|
1081 T1V = T1T + T1U;
|
cannam@167
|
1082 T1W = T1S + T1V;
|
cannam@167
|
1083 T3k = T1V - T1S;
|
cannam@167
|
1084 T4i = T1g - T1j;
|
cannam@167
|
1085 T4j = T1Q - T1R;
|
cannam@167
|
1086 T4k = T4i - T4j;
|
cannam@167
|
1087 T5s = T4i + T4j;
|
cannam@167
|
1088 }
|
cannam@167
|
1089 }
|
cannam@167
|
1090 {
|
cannam@167
|
1091 E T2g, T4F, T2j, T4E, T2p, T4C, T2s, T4B;
|
cannam@167
|
1092 {
|
cannam@167
|
1093 E T2c, T2i, T2f, T2h, T2d, T2e;
|
cannam@167
|
1094 T2c = R1[WS(rs, 1)];
|
cannam@167
|
1095 T2i = R1[WS(rs, 17)];
|
cannam@167
|
1096 T2d = R1[WS(rs, 9)];
|
cannam@167
|
1097 T2e = R1[WS(rs, 25)];
|
cannam@167
|
1098 T2f = KP707106781 * (T2d - T2e);
|
cannam@167
|
1099 T2h = KP707106781 * (T2d + T2e);
|
cannam@167
|
1100 T2g = T2c + T2f;
|
cannam@167
|
1101 T4F = T2c - T2f;
|
cannam@167
|
1102 T2j = T2h + T2i;
|
cannam@167
|
1103 T4E = T2i - T2h;
|
cannam@167
|
1104 }
|
cannam@167
|
1105 {
|
cannam@167
|
1106 E T2o, T2r, T2n, T2q, T2l, T2m;
|
cannam@167
|
1107 T2o = R1[WS(rs, 29)];
|
cannam@167
|
1108 T2r = R1[WS(rs, 13)];
|
cannam@167
|
1109 T2l = R1[WS(rs, 5)];
|
cannam@167
|
1110 T2m = R1[WS(rs, 21)];
|
cannam@167
|
1111 T2n = KP707106781 * (T2l - T2m);
|
cannam@167
|
1112 T2q = KP707106781 * (T2l + T2m);
|
cannam@167
|
1113 T2p = T2n - T2o;
|
cannam@167
|
1114 T4C = T2n + T2o;
|
cannam@167
|
1115 T2s = T2q + T2r;
|
cannam@167
|
1116 T4B = T2r - T2q;
|
cannam@167
|
1117 }
|
cannam@167
|
1118 {
|
cannam@167
|
1119 E T2k, T2t, T4P, T4Q;
|
cannam@167
|
1120 T2k = FNMS(KP195090322, T2j, KP980785280 * T2g);
|
cannam@167
|
1121 T2t = FMA(KP980785280, T2p, KP195090322 * T2s);
|
cannam@167
|
1122 T2u = T2k + T2t;
|
cannam@167
|
1123 T3q = T2t - T2k;
|
cannam@167
|
1124 T4P = FMA(KP831469612, T4F, KP555570233 * T4E);
|
cannam@167
|
1125 T4Q = FMA(KP831469612, T4C, KP555570233 * T4B);
|
cannam@167
|
1126 T4R = T4P + T4Q;
|
cannam@167
|
1127 T5A = T4P - T4Q;
|
cannam@167
|
1128 }
|
cannam@167
|
1129 {
|
cannam@167
|
1130 E T2w, T2x, T4D, T4G;
|
cannam@167
|
1131 T2w = FNMS(KP980785280, T2s, KP195090322 * T2p);
|
cannam@167
|
1132 T2x = FMA(KP195090322, T2g, KP980785280 * T2j);
|
cannam@167
|
1133 T2y = T2w - T2x;
|
cannam@167
|
1134 T3o = T2x + T2w;
|
cannam@167
|
1135 T4D = FNMS(KP555570233, T4C, KP831469612 * T4B);
|
cannam@167
|
1136 T4G = FNMS(KP555570233, T4F, KP831469612 * T4E);
|
cannam@167
|
1137 T4H = T4D - T4G;
|
cannam@167
|
1138 T5C = T4G + T4D;
|
cannam@167
|
1139 }
|
cannam@167
|
1140 }
|
cannam@167
|
1141 {
|
cannam@167
|
1142 E T1x, T4p, T1A, T4o, T1G, T4m, T1J, T4l;
|
cannam@167
|
1143 {
|
cannam@167
|
1144 E T1t, T1z, T1w, T1y, T1u, T1v;
|
cannam@167
|
1145 T1t = R1[WS(rs, 2)];
|
cannam@167
|
1146 T1z = R1[WS(rs, 18)];
|
cannam@167
|
1147 T1u = R1[WS(rs, 10)];
|
cannam@167
|
1148 T1v = R1[WS(rs, 26)];
|
cannam@167
|
1149 T1w = KP707106781 * (T1u - T1v);
|
cannam@167
|
1150 T1y = KP707106781 * (T1u + T1v);
|
cannam@167
|
1151 T1x = T1t + T1w;
|
cannam@167
|
1152 T4p = T1t - T1w;
|
cannam@167
|
1153 T1A = T1y + T1z;
|
cannam@167
|
1154 T4o = T1z - T1y;
|
cannam@167
|
1155 }
|
cannam@167
|
1156 {
|
cannam@167
|
1157 E T1F, T1I, T1E, T1H, T1C, T1D;
|
cannam@167
|
1158 T1F = R1[WS(rs, 30)];
|
cannam@167
|
1159 T1I = R1[WS(rs, 14)];
|
cannam@167
|
1160 T1C = R1[WS(rs, 6)];
|
cannam@167
|
1161 T1D = R1[WS(rs, 22)];
|
cannam@167
|
1162 T1E = KP707106781 * (T1C - T1D);
|
cannam@167
|
1163 T1H = KP707106781 * (T1C + T1D);
|
cannam@167
|
1164 T1G = T1E - T1F;
|
cannam@167
|
1165 T4m = T1E + T1F;
|
cannam@167
|
1166 T1J = T1H + T1I;
|
cannam@167
|
1167 T4l = T1I - T1H;
|
cannam@167
|
1168 }
|
cannam@167
|
1169 {
|
cannam@167
|
1170 E T1B, T1K, T4w, T4x;
|
cannam@167
|
1171 T1B = FNMS(KP195090322, T1A, KP980785280 * T1x);
|
cannam@167
|
1172 T1K = FMA(KP980785280, T1G, KP195090322 * T1J);
|
cannam@167
|
1173 T1L = T1B + T1K;
|
cannam@167
|
1174 T3j = T1K - T1B;
|
cannam@167
|
1175 T4w = FMA(KP831469612, T4p, KP555570233 * T4o);
|
cannam@167
|
1176 T4x = FMA(KP831469612, T4m, KP555570233 * T4l);
|
cannam@167
|
1177 T4y = T4w + T4x;
|
cannam@167
|
1178 T5t = T4w - T4x;
|
cannam@167
|
1179 }
|
cannam@167
|
1180 {
|
cannam@167
|
1181 E T1N, T1O, T4n, T4q;
|
cannam@167
|
1182 T1N = FNMS(KP980785280, T1J, KP195090322 * T1G);
|
cannam@167
|
1183 T1O = FMA(KP195090322, T1x, KP980785280 * T1A);
|
cannam@167
|
1184 T1P = T1N - T1O;
|
cannam@167
|
1185 T3h = T1O + T1N;
|
cannam@167
|
1186 T4n = FNMS(KP555570233, T4m, KP831469612 * T4l);
|
cannam@167
|
1187 T4q = FNMS(KP555570233, T4p, KP831469612 * T4o);
|
cannam@167
|
1188 T4r = T4n - T4q;
|
cannam@167
|
1189 T5v = T4q + T4n;
|
cannam@167
|
1190 }
|
cannam@167
|
1191 }
|
cannam@167
|
1192 {
|
cannam@167
|
1193 E Tx, T2N, T69, T6f, T1e, T6e, T2X, T30, T1Y, T2L, T2Q, T62, T2U, T31, T2H;
|
cannam@167
|
1194 E T2K, Tw, T63;
|
cannam@167
|
1195 Tw = Tm + Tv;
|
cannam@167
|
1196 Tx = Td + Tw;
|
cannam@167
|
1197 T2N = Td - Tw;
|
cannam@167
|
1198 T63 = T35 - T34;
|
cannam@167
|
1199 T69 = T63 - T68;
|
cannam@167
|
1200 T6f = T63 + T68;
|
cannam@167
|
1201 {
|
cannam@167
|
1202 E TS, T1d, T2V, T2W;
|
cannam@167
|
1203 TS = FNMS(KP098017140, TR, KP995184726 * TK);
|
cannam@167
|
1204 T1d = FMA(KP995184726, T15, KP098017140 * T1c);
|
cannam@167
|
1205 T1e = TS + T1d;
|
cannam@167
|
1206 T6e = T1d - TS;
|
cannam@167
|
1207 T2V = T2b - T2u;
|
cannam@167
|
1208 T2W = T2y + T2F;
|
cannam@167
|
1209 T2X = FNMS(KP671558954, T2W, KP740951125 * T2V);
|
cannam@167
|
1210 T30 = FMA(KP671558954, T2V, KP740951125 * T2W);
|
cannam@167
|
1211 }
|
cannam@167
|
1212 {
|
cannam@167
|
1213 E T1M, T1X, T2O, T2P;
|
cannam@167
|
1214 T1M = T1s + T1L;
|
cannam@167
|
1215 T1X = T1P - T1W;
|
cannam@167
|
1216 T1Y = FMA(KP998795456, T1M, KP049067674 * T1X);
|
cannam@167
|
1217 T2L = FNMS(KP049067674, T1M, KP998795456 * T1X);
|
cannam@167
|
1218 T2O = FMA(KP098017140, TK, KP995184726 * TR);
|
cannam@167
|
1219 T2P = FNMS(KP995184726, T1c, KP098017140 * T15);
|
cannam@167
|
1220 T2Q = T2O + T2P;
|
cannam@167
|
1221 T62 = T2P - T2O;
|
cannam@167
|
1222 }
|
cannam@167
|
1223 {
|
cannam@167
|
1224 E T2S, T2T, T2v, T2G;
|
cannam@167
|
1225 T2S = T1s - T1L;
|
cannam@167
|
1226 T2T = T1P + T1W;
|
cannam@167
|
1227 T2U = FMA(KP740951125, T2S, KP671558954 * T2T);
|
cannam@167
|
1228 T31 = FNMS(KP671558954, T2S, KP740951125 * T2T);
|
cannam@167
|
1229 T2v = T2b + T2u;
|
cannam@167
|
1230 T2G = T2y - T2F;
|
cannam@167
|
1231 T2H = FNMS(KP049067674, T2G, KP998795456 * T2v);
|
cannam@167
|
1232 T2K = FMA(KP049067674, T2v, KP998795456 * T2G);
|
cannam@167
|
1233 }
|
cannam@167
|
1234 {
|
cannam@167
|
1235 E T1f, T2I, T6b, T6c;
|
cannam@167
|
1236 T1f = Tx + T1e;
|
cannam@167
|
1237 T2I = T1Y + T2H;
|
cannam@167
|
1238 Cr[WS(csr, 31)] = T1f - T2I;
|
cannam@167
|
1239 Cr[0] = T1f + T2I;
|
cannam@167
|
1240 T6b = T2L + T2K;
|
cannam@167
|
1241 T6c = T62 + T69;
|
cannam@167
|
1242 Ci[WS(csi, 31)] = T6b - T6c;
|
cannam@167
|
1243 Ci[0] = T6b + T6c;
|
cannam@167
|
1244 }
|
cannam@167
|
1245 {
|
cannam@167
|
1246 E T2J, T2M, T61, T6a;
|
cannam@167
|
1247 T2J = Tx - T1e;
|
cannam@167
|
1248 T2M = T2K - T2L;
|
cannam@167
|
1249 Cr[WS(csr, 16)] = T2J - T2M;
|
cannam@167
|
1250 Cr[WS(csr, 15)] = T2J + T2M;
|
cannam@167
|
1251 T61 = T2H - T1Y;
|
cannam@167
|
1252 T6a = T62 - T69;
|
cannam@167
|
1253 Ci[WS(csi, 16)] = T61 - T6a;
|
cannam@167
|
1254 Ci[WS(csi, 15)] = T61 + T6a;
|
cannam@167
|
1255 }
|
cannam@167
|
1256 {
|
cannam@167
|
1257 E T2R, T2Y, T6h, T6i;
|
cannam@167
|
1258 T2R = T2N + T2Q;
|
cannam@167
|
1259 T2Y = T2U + T2X;
|
cannam@167
|
1260 Cr[WS(csr, 24)] = T2R - T2Y;
|
cannam@167
|
1261 Cr[WS(csr, 7)] = T2R + T2Y;
|
cannam@167
|
1262 T6h = T31 + T30;
|
cannam@167
|
1263 T6i = T6e + T6f;
|
cannam@167
|
1264 Ci[WS(csi, 24)] = T6h - T6i;
|
cannam@167
|
1265 Ci[WS(csi, 7)] = T6h + T6i;
|
cannam@167
|
1266 }
|
cannam@167
|
1267 {
|
cannam@167
|
1268 E T2Z, T32, T6d, T6g;
|
cannam@167
|
1269 T2Z = T2N - T2Q;
|
cannam@167
|
1270 T32 = T30 - T31;
|
cannam@167
|
1271 Cr[WS(csr, 23)] = T2Z - T32;
|
cannam@167
|
1272 Cr[WS(csr, 8)] = T2Z + T32;
|
cannam@167
|
1273 T6d = T2X - T2U;
|
cannam@167
|
1274 T6g = T6e - T6f;
|
cannam@167
|
1275 Ci[WS(csi, 23)] = T6d - T6g;
|
cannam@167
|
1276 Ci[WS(csi, 8)] = T6d + T6g;
|
cannam@167
|
1277 }
|
cannam@167
|
1278 }
|
cannam@167
|
1279 {
|
cannam@167
|
1280 E T5j, T5L, T6R, T6X, T5q, T6W, T5V, T5Y, T5y, T5J, T5O, T6O, T5S, T5Z, T5F;
|
cannam@167
|
1281 E T5I, T5i, T6P;
|
cannam@167
|
1282 T5i = T5g - T5h;
|
cannam@167
|
1283 T5j = T5f - T5i;
|
cannam@167
|
1284 T5L = T5f + T5i;
|
cannam@167
|
1285 T6P = T3Z + T3W;
|
cannam@167
|
1286 T6R = T6P - T6Q;
|
cannam@167
|
1287 T6X = T6P + T6Q;
|
cannam@167
|
1288 {
|
cannam@167
|
1289 E T5m, T5p, T5T, T5U;
|
cannam@167
|
1290 T5m = FMA(KP290284677, T5k, KP956940335 * T5l);
|
cannam@167
|
1291 T5p = FNMS(KP290284677, T5o, KP956940335 * T5n);
|
cannam@167
|
1292 T5q = T5m - T5p;
|
cannam@167
|
1293 T6W = T5p + T5m;
|
cannam@167
|
1294 T5T = T5z + T5A;
|
cannam@167
|
1295 T5U = T5C + T5D;
|
cannam@167
|
1296 T5V = FNMS(KP146730474, T5U, KP989176509 * T5T);
|
cannam@167
|
1297 T5Y = FMA(KP146730474, T5T, KP989176509 * T5U);
|
cannam@167
|
1298 }
|
cannam@167
|
1299 {
|
cannam@167
|
1300 E T5u, T5x, T5M, T5N;
|
cannam@167
|
1301 T5u = T5s - T5t;
|
cannam@167
|
1302 T5x = T5v - T5w;
|
cannam@167
|
1303 T5y = FMA(KP803207531, T5u, KP595699304 * T5x);
|
cannam@167
|
1304 T5J = FNMS(KP595699304, T5u, KP803207531 * T5x);
|
cannam@167
|
1305 T5M = FMA(KP956940335, T5o, KP290284677 * T5n);
|
cannam@167
|
1306 T5N = FNMS(KP290284677, T5l, KP956940335 * T5k);
|
cannam@167
|
1307 T5O = T5M + T5N;
|
cannam@167
|
1308 T6O = T5N - T5M;
|
cannam@167
|
1309 }
|
cannam@167
|
1310 {
|
cannam@167
|
1311 E T5Q, T5R, T5B, T5E;
|
cannam@167
|
1312 T5Q = T5s + T5t;
|
cannam@167
|
1313 T5R = T5v + T5w;
|
cannam@167
|
1314 T5S = FMA(KP989176509, T5Q, KP146730474 * T5R);
|
cannam@167
|
1315 T5Z = FNMS(KP146730474, T5Q, KP989176509 * T5R);
|
cannam@167
|
1316 T5B = T5z - T5A;
|
cannam@167
|
1317 T5E = T5C - T5D;
|
cannam@167
|
1318 T5F = FNMS(KP595699304, T5E, KP803207531 * T5B);
|
cannam@167
|
1319 T5I = FMA(KP595699304, T5B, KP803207531 * T5E);
|
cannam@167
|
1320 }
|
cannam@167
|
1321 {
|
cannam@167
|
1322 E T5r, T5G, T6T, T6U;
|
cannam@167
|
1323 T5r = T5j + T5q;
|
cannam@167
|
1324 T5G = T5y + T5F;
|
cannam@167
|
1325 Cr[WS(csr, 25)] = T5r - T5G;
|
cannam@167
|
1326 Cr[WS(csr, 6)] = T5r + T5G;
|
cannam@167
|
1327 T6T = T5J + T5I;
|
cannam@167
|
1328 T6U = T6O + T6R;
|
cannam@167
|
1329 Ci[WS(csi, 25)] = T6T - T6U;
|
cannam@167
|
1330 Ci[WS(csi, 6)] = T6T + T6U;
|
cannam@167
|
1331 }
|
cannam@167
|
1332 {
|
cannam@167
|
1333 E T5H, T5K, T6N, T6S;
|
cannam@167
|
1334 T5H = T5j - T5q;
|
cannam@167
|
1335 T5K = T5I - T5J;
|
cannam@167
|
1336 Cr[WS(csr, 22)] = T5H - T5K;
|
cannam@167
|
1337 Cr[WS(csr, 9)] = T5H + T5K;
|
cannam@167
|
1338 T6N = T5F - T5y;
|
cannam@167
|
1339 T6S = T6O - T6R;
|
cannam@167
|
1340 Ci[WS(csi, 22)] = T6N - T6S;
|
cannam@167
|
1341 Ci[WS(csi, 9)] = T6N + T6S;
|
cannam@167
|
1342 }
|
cannam@167
|
1343 {
|
cannam@167
|
1344 E T5P, T5W, T6Z, T70;
|
cannam@167
|
1345 T5P = T5L + T5O;
|
cannam@167
|
1346 T5W = T5S + T5V;
|
cannam@167
|
1347 Cr[WS(csr, 30)] = T5P - T5W;
|
cannam@167
|
1348 Cr[WS(csr, 1)] = T5P + T5W;
|
cannam@167
|
1349 T6Z = T5Z + T5Y;
|
cannam@167
|
1350 T70 = T6W + T6X;
|
cannam@167
|
1351 Ci[WS(csi, 30)] = T6Z - T70;
|
cannam@167
|
1352 Ci[WS(csi, 1)] = T6Z + T70;
|
cannam@167
|
1353 }
|
cannam@167
|
1354 {
|
cannam@167
|
1355 E T5X, T60, T6V, T6Y;
|
cannam@167
|
1356 T5X = T5L - T5O;
|
cannam@167
|
1357 T60 = T5Y - T5Z;
|
cannam@167
|
1358 Cr[WS(csr, 17)] = T5X - T60;
|
cannam@167
|
1359 Cr[WS(csr, 14)] = T5X + T60;
|
cannam@167
|
1360 T6V = T5V - T5S;
|
cannam@167
|
1361 T6Y = T6W - T6X;
|
cannam@167
|
1362 Ci[WS(csi, 17)] = T6V - T6Y;
|
cannam@167
|
1363 Ci[WS(csi, 14)] = T6V + T6Y;
|
cannam@167
|
1364 }
|
cannam@167
|
1365 }
|
cannam@167
|
1366 {
|
cannam@167
|
1367 E T37, T3z, T6n, T6t, T3e, T6s, T3J, T3M, T3m, T3x, T3C, T6k, T3G, T3N, T3t;
|
cannam@167
|
1368 E T3w, T36, T6l;
|
cannam@167
|
1369 T36 = T34 + T35;
|
cannam@167
|
1370 T37 = T33 - T36;
|
cannam@167
|
1371 T3z = T33 + T36;
|
cannam@167
|
1372 T6l = Tv - Tm;
|
cannam@167
|
1373 T6n = T6l - T6m;
|
cannam@167
|
1374 T6t = T6l + T6m;
|
cannam@167
|
1375 {
|
cannam@167
|
1376 E T3a, T3d, T3H, T3I;
|
cannam@167
|
1377 T3a = FMA(KP634393284, T38, KP773010453 * T39);
|
cannam@167
|
1378 T3d = FNMS(KP634393284, T3c, KP773010453 * T3b);
|
cannam@167
|
1379 T3e = T3a - T3d;
|
cannam@167
|
1380 T6s = T3d + T3a;
|
cannam@167
|
1381 T3H = T3n + T3o;
|
cannam@167
|
1382 T3I = T3q + T3r;
|
cannam@167
|
1383 T3J = FNMS(KP336889853, T3I, KP941544065 * T3H);
|
cannam@167
|
1384 T3M = FMA(KP336889853, T3H, KP941544065 * T3I);
|
cannam@167
|
1385 }
|
cannam@167
|
1386 {
|
cannam@167
|
1387 E T3i, T3l, T3A, T3B;
|
cannam@167
|
1388 T3i = T3g - T3h;
|
cannam@167
|
1389 T3l = T3j - T3k;
|
cannam@167
|
1390 T3m = FMA(KP903989293, T3i, KP427555093 * T3l);
|
cannam@167
|
1391 T3x = FNMS(KP427555093, T3i, KP903989293 * T3l);
|
cannam@167
|
1392 T3A = FMA(KP773010453, T3c, KP634393284 * T3b);
|
cannam@167
|
1393 T3B = FNMS(KP634393284, T39, KP773010453 * T38);
|
cannam@167
|
1394 T3C = T3A + T3B;
|
cannam@167
|
1395 T6k = T3B - T3A;
|
cannam@167
|
1396 }
|
cannam@167
|
1397 {
|
cannam@167
|
1398 E T3E, T3F, T3p, T3s;
|
cannam@167
|
1399 T3E = T3g + T3h;
|
cannam@167
|
1400 T3F = T3j + T3k;
|
cannam@167
|
1401 T3G = FMA(KP941544065, T3E, KP336889853 * T3F);
|
cannam@167
|
1402 T3N = FNMS(KP336889853, T3E, KP941544065 * T3F);
|
cannam@167
|
1403 T3p = T3n - T3o;
|
cannam@167
|
1404 T3s = T3q - T3r;
|
cannam@167
|
1405 T3t = FNMS(KP427555093, T3s, KP903989293 * T3p);
|
cannam@167
|
1406 T3w = FMA(KP427555093, T3p, KP903989293 * T3s);
|
cannam@167
|
1407 }
|
cannam@167
|
1408 {
|
cannam@167
|
1409 E T3f, T3u, T6p, T6q;
|
cannam@167
|
1410 T3f = T37 + T3e;
|
cannam@167
|
1411 T3u = T3m + T3t;
|
cannam@167
|
1412 Cr[WS(csr, 27)] = T3f - T3u;
|
cannam@167
|
1413 Cr[WS(csr, 4)] = T3f + T3u;
|
cannam@167
|
1414 T6p = T3x + T3w;
|
cannam@167
|
1415 T6q = T6k + T6n;
|
cannam@167
|
1416 Ci[WS(csi, 27)] = T6p - T6q;
|
cannam@167
|
1417 Ci[WS(csi, 4)] = T6p + T6q;
|
cannam@167
|
1418 }
|
cannam@167
|
1419 {
|
cannam@167
|
1420 E T3v, T3y, T6j, T6o;
|
cannam@167
|
1421 T3v = T37 - T3e;
|
cannam@167
|
1422 T3y = T3w - T3x;
|
cannam@167
|
1423 Cr[WS(csr, 20)] = T3v - T3y;
|
cannam@167
|
1424 Cr[WS(csr, 11)] = T3v + T3y;
|
cannam@167
|
1425 T6j = T3t - T3m;
|
cannam@167
|
1426 T6o = T6k - T6n;
|
cannam@167
|
1427 Ci[WS(csi, 20)] = T6j - T6o;
|
cannam@167
|
1428 Ci[WS(csi, 11)] = T6j + T6o;
|
cannam@167
|
1429 }
|
cannam@167
|
1430 {
|
cannam@167
|
1431 E T3D, T3K, T6v, T6w;
|
cannam@167
|
1432 T3D = T3z + T3C;
|
cannam@167
|
1433 T3K = T3G + T3J;
|
cannam@167
|
1434 Cr[WS(csr, 28)] = T3D - T3K;
|
cannam@167
|
1435 Cr[WS(csr, 3)] = T3D + T3K;
|
cannam@167
|
1436 T6v = T3N + T3M;
|
cannam@167
|
1437 T6w = T6s + T6t;
|
cannam@167
|
1438 Ci[WS(csi, 28)] = T6v - T6w;
|
cannam@167
|
1439 Ci[WS(csi, 3)] = T6v + T6w;
|
cannam@167
|
1440 }
|
cannam@167
|
1441 {
|
cannam@167
|
1442 E T3L, T3O, T6r, T6u;
|
cannam@167
|
1443 T3L = T3z - T3C;
|
cannam@167
|
1444 T3O = T3M - T3N;
|
cannam@167
|
1445 Cr[WS(csr, 19)] = T3L - T3O;
|
cannam@167
|
1446 Cr[WS(csr, 12)] = T3L + T3O;
|
cannam@167
|
1447 T6r = T3J - T3G;
|
cannam@167
|
1448 T6u = T6s - T6t;
|
cannam@167
|
1449 Ci[WS(csi, 19)] = T6r - T6u;
|
cannam@167
|
1450 Ci[WS(csi, 12)] = T6r + T6u;
|
cannam@167
|
1451 }
|
cannam@167
|
1452 }
|
cannam@167
|
1453 {
|
cannam@167
|
1454 E T41, T4Z, T6D, T6J, T4g, T6I, T59, T5d, T4A, T4X, T52, T6y, T56, T5c, T4T;
|
cannam@167
|
1455 E T4W, T40, T6C;
|
cannam@167
|
1456 T40 = T3W - T3Z;
|
cannam@167
|
1457 T41 = T3T + T40;
|
cannam@167
|
1458 T4Z = T3T - T40;
|
cannam@167
|
1459 T6C = T5g + T5h;
|
cannam@167
|
1460 T6D = T6B - T6C;
|
cannam@167
|
1461 T6J = T6C + T6B;
|
cannam@167
|
1462 {
|
cannam@167
|
1463 E T48, T4f, T57, T58;
|
cannam@167
|
1464 T48 = FMA(KP881921264, T44, KP471396736 * T47);
|
cannam@167
|
1465 T4f = FMA(KP881921264, T4b, KP471396736 * T4e);
|
cannam@167
|
1466 T4g = T48 - T4f;
|
cannam@167
|
1467 T6I = T48 + T4f;
|
cannam@167
|
1468 T57 = T4K + T4H;
|
cannam@167
|
1469 T58 = T4R + T4O;
|
cannam@167
|
1470 T59 = FMA(KP514102744, T57, KP857728610 * T58);
|
cannam@167
|
1471 T5d = FNMS(KP857728610, T57, KP514102744 * T58);
|
cannam@167
|
1472 }
|
cannam@167
|
1473 {
|
cannam@167
|
1474 E T4s, T4z, T50, T51;
|
cannam@167
|
1475 T4s = T4k + T4r;
|
cannam@167
|
1476 T4z = T4v - T4y;
|
cannam@167
|
1477 T4A = FMA(KP970031253, T4s, KP242980179 * T4z);
|
cannam@167
|
1478 T4X = FNMS(KP242980179, T4s, KP970031253 * T4z);
|
cannam@167
|
1479 T50 = FNMS(KP471396736, T4b, KP881921264 * T4e);
|
cannam@167
|
1480 T51 = FNMS(KP471396736, T44, KP881921264 * T47);
|
cannam@167
|
1481 T52 = T50 - T51;
|
cannam@167
|
1482 T6y = T51 + T50;
|
cannam@167
|
1483 }
|
cannam@167
|
1484 {
|
cannam@167
|
1485 E T54, T55, T4L, T4S;
|
cannam@167
|
1486 T54 = T4k - T4r;
|
cannam@167
|
1487 T55 = T4y + T4v;
|
cannam@167
|
1488 T56 = FMA(KP514102744, T54, KP857728610 * T55);
|
cannam@167
|
1489 T5c = FNMS(KP514102744, T55, KP857728610 * T54);
|
cannam@167
|
1490 T4L = T4H - T4K;
|
cannam@167
|
1491 T4S = T4O - T4R;
|
cannam@167
|
1492 T4T = FNMS(KP242980179, T4S, KP970031253 * T4L);
|
cannam@167
|
1493 T4W = FMA(KP242980179, T4L, KP970031253 * T4S);
|
cannam@167
|
1494 }
|
cannam@167
|
1495 {
|
cannam@167
|
1496 E T4h, T4U, T6F, T6G;
|
cannam@167
|
1497 T4h = T41 + T4g;
|
cannam@167
|
1498 T4U = T4A + T4T;
|
cannam@167
|
1499 Cr[WS(csr, 29)] = T4h - T4U;
|
cannam@167
|
1500 Cr[WS(csr, 2)] = T4h + T4U;
|
cannam@167
|
1501 T6F = T4X + T4W;
|
cannam@167
|
1502 T6G = T6y + T6D;
|
cannam@167
|
1503 Ci[WS(csi, 29)] = T6F - T6G;
|
cannam@167
|
1504 Ci[WS(csi, 2)] = T6F + T6G;
|
cannam@167
|
1505 }
|
cannam@167
|
1506 {
|
cannam@167
|
1507 E T4V, T4Y, T6x, T6E;
|
cannam@167
|
1508 T4V = T41 - T4g;
|
cannam@167
|
1509 T4Y = T4W - T4X;
|
cannam@167
|
1510 Cr[WS(csr, 18)] = T4V - T4Y;
|
cannam@167
|
1511 Cr[WS(csr, 13)] = T4V + T4Y;
|
cannam@167
|
1512 T6x = T4T - T4A;
|
cannam@167
|
1513 T6E = T6y - T6D;
|
cannam@167
|
1514 Ci[WS(csi, 18)] = T6x - T6E;
|
cannam@167
|
1515 Ci[WS(csi, 13)] = T6x + T6E;
|
cannam@167
|
1516 }
|
cannam@167
|
1517 {
|
cannam@167
|
1518 E T53, T5a, T6L, T6M;
|
cannam@167
|
1519 T53 = T4Z - T52;
|
cannam@167
|
1520 T5a = T56 - T59;
|
cannam@167
|
1521 Cr[WS(csr, 21)] = T53 - T5a;
|
cannam@167
|
1522 Cr[WS(csr, 10)] = T53 + T5a;
|
cannam@167
|
1523 T6L = T5d - T5c;
|
cannam@167
|
1524 T6M = T6J - T6I;
|
cannam@167
|
1525 Ci[WS(csi, 21)] = T6L - T6M;
|
cannam@167
|
1526 Ci[WS(csi, 10)] = T6L + T6M;
|
cannam@167
|
1527 }
|
cannam@167
|
1528 {
|
cannam@167
|
1529 E T5b, T5e, T6H, T6K;
|
cannam@167
|
1530 T5b = T4Z + T52;
|
cannam@167
|
1531 T5e = T5c + T5d;
|
cannam@167
|
1532 Cr[WS(csr, 26)] = T5b - T5e;
|
cannam@167
|
1533 Cr[WS(csr, 5)] = T5b + T5e;
|
cannam@167
|
1534 T6H = T56 + T59;
|
cannam@167
|
1535 T6K = T6I + T6J;
|
cannam@167
|
1536 Ci[WS(csi, 5)] = -(T6H + T6K);
|
cannam@167
|
1537 Ci[WS(csi, 26)] = T6K - T6H;
|
cannam@167
|
1538 }
|
cannam@167
|
1539 }
|
cannam@167
|
1540 }
|
cannam@167
|
1541 }
|
cannam@167
|
1542 }
|
cannam@167
|
1543
|
cannam@167
|
1544 static const kr2c_desc desc = { 64, "r2cfII_64", {342, 114, 92, 0}, &GENUS };
|
cannam@167
|
1545
|
cannam@167
|
1546 void X(codelet_r2cfII_64) (planner *p) {
|
cannam@167
|
1547 X(kr2c_register) (p, r2cfII_64, &desc);
|
cannam@167
|
1548 }
|
cannam@167
|
1549
|
cannam@167
|
1550 #endif
|