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 08:07:12 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_hc2cdft.native -fma -compact -variables 4 -pipeline-latency 4 -n 32 -dit -name hc2cfdft_32 -include rdft/scalar/hc2cf.h */
|
cannam@167
|
29
|
cannam@167
|
30 /*
|
cannam@167
|
31 * This function contains 498 FP additions, 324 FP multiplications,
|
cannam@167
|
32 * (or, 300 additions, 126 multiplications, 198 fused multiply/add),
|
cannam@167
|
33 * 113 stack variables, 8 constants, and 128 memory accesses
|
cannam@167
|
34 */
|
cannam@167
|
35 #include "rdft/scalar/hc2cf.h"
|
cannam@167
|
36
|
cannam@167
|
37 static void hc2cfdft_32(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms)
|
cannam@167
|
38 {
|
cannam@167
|
39 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
cannam@167
|
40 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
cannam@167
|
41 DK(KP198912367, +0.198912367379658006911597622644676228597850501);
|
cannam@167
|
42 DK(KP668178637, +0.668178637919298919997757686523080761552472251);
|
cannam@167
|
43 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
cannam@167
|
44 DK(KP414213562, +0.414213562373095048801688724209698078569671875);
|
cannam@167
|
45 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
cannam@167
|
46 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
|
cannam@167
|
47 {
|
cannam@167
|
48 INT m;
|
cannam@167
|
49 for (m = mb, W = W + ((mb - 1) * 62); m < me; m = m + 1, Rp = Rp + ms, Ip = Ip + ms, Rm = Rm - ms, Im = Im - ms, W = W + 62, MAKE_VOLATILE_STRIDE(128, rs)) {
|
cannam@167
|
50 E T3B, T89, T61, T8l, T2F, T8t, T4B, T7p, T1n, T7L, T5e, T7I, T4u, T82, T5E;
|
cannam@167
|
51 E T7R, T3m, T8k, T5W, T8a, T2r, T8u, T4G, T7q, T12, T7K, T59, T7H, T4h, T81;
|
cannam@167
|
52 E T5z, T7Q, Tl, T7D, T4Y, T7A, T3Q, T5o, T7V, T84, T1K, T7t, T4M, T7s, T2V;
|
cannam@167
|
53 E T8n, T5L, T8e, T25, T7w, T4R, T7v, T38, T8o, T5Q, T8h, TG, T7E, T53, T7B;
|
cannam@167
|
54 E T43, T5t, T7Y, T85;
|
cannam@167
|
55 {
|
cannam@167
|
56 E T2E, T3z, T4y, T3y, T5Z, T3t, T3x, T2v, T2A, T3r, T3q, T5X, T3n, T3p, T2w;
|
cannam@167
|
57 E T4z, T3s, T3A;
|
cannam@167
|
58 {
|
cannam@167
|
59 E T2C, T2D, T3u, T3v, T3w;
|
cannam@167
|
60 T2C = Ip[0];
|
cannam@167
|
61 T2D = Im[0];
|
cannam@167
|
62 T2E = T2C - T2D;
|
cannam@167
|
63 T3z = T2C + T2D;
|
cannam@167
|
64 T3u = Rm[0];
|
cannam@167
|
65 T3v = Rp[0];
|
cannam@167
|
66 T3w = T3u - T3v;
|
cannam@167
|
67 T4y = T3v + T3u;
|
cannam@167
|
68 T3y = W[1];
|
cannam@167
|
69 T5Z = T3y * T3w;
|
cannam@167
|
70 T3t = W[0];
|
cannam@167
|
71 T3x = T3t * T3w;
|
cannam@167
|
72 {
|
cannam@167
|
73 E T2t, T2u, T3o, T2y, T2z, T2s;
|
cannam@167
|
74 T2t = Ip[WS(rs, 8)];
|
cannam@167
|
75 T2u = Im[WS(rs, 8)];
|
cannam@167
|
76 T2v = T2t - T2u;
|
cannam@167
|
77 T2y = Rp[WS(rs, 8)];
|
cannam@167
|
78 T2z = Rm[WS(rs, 8)];
|
cannam@167
|
79 T2A = T2y + T2z;
|
cannam@167
|
80 T3o = T2z - T2y;
|
cannam@167
|
81 T3r = T2t + T2u;
|
cannam@167
|
82 T3q = W[33];
|
cannam@167
|
83 T5X = T3q * T3o;
|
cannam@167
|
84 T3n = W[32];
|
cannam@167
|
85 T3p = T3n * T3o;
|
cannam@167
|
86 T2s = W[30];
|
cannam@167
|
87 T2w = T2s * T2v;
|
cannam@167
|
88 T4z = T2s * T2A;
|
cannam@167
|
89 }
|
cannam@167
|
90 }
|
cannam@167
|
91 T3s = FNMS(T3q, T3r, T3p);
|
cannam@167
|
92 T3A = FNMS(T3y, T3z, T3x);
|
cannam@167
|
93 T3B = T3s + T3A;
|
cannam@167
|
94 T89 = T3A - T3s;
|
cannam@167
|
95 {
|
cannam@167
|
96 E T5Y, T60, T2B, T4A, T2x;
|
cannam@167
|
97 T5Y = FMA(T3n, T3r, T5X);
|
cannam@167
|
98 T60 = FMA(T3t, T3z, T5Z);
|
cannam@167
|
99 T61 = T5Y + T60;
|
cannam@167
|
100 T8l = T60 - T5Y;
|
cannam@167
|
101 T2x = W[31];
|
cannam@167
|
102 T2B = FNMS(T2x, T2A, T2w);
|
cannam@167
|
103 T4A = FMA(T2x, T2v, T4z);
|
cannam@167
|
104 T2F = T2B + T2E;
|
cannam@167
|
105 T8t = T4y - T4A;
|
cannam@167
|
106 T4B = T4y + T4A;
|
cannam@167
|
107 T7p = T2E - T2B;
|
cannam@167
|
108 }
|
cannam@167
|
109 }
|
cannam@167
|
110 {
|
cannam@167
|
111 E T16, T4m, T1b, T4j, T17, T5a, T4k, T5A, T1g, T4s, T1l, T4p, T1h, T5c, T4q;
|
cannam@167
|
112 E T5C;
|
cannam@167
|
113 {
|
cannam@167
|
114 E T13, T4i, T1d, T4o;
|
cannam@167
|
115 {
|
cannam@167
|
116 E T14, T15, T19, T1a;
|
cannam@167
|
117 T14 = Ip[WS(rs, 3)];
|
cannam@167
|
118 T15 = Im[WS(rs, 3)];
|
cannam@167
|
119 T16 = T14 - T15;
|
cannam@167
|
120 T4m = T14 + T15;
|
cannam@167
|
121 T19 = Rp[WS(rs, 3)];
|
cannam@167
|
122 T1a = Rm[WS(rs, 3)];
|
cannam@167
|
123 T1b = T19 + T1a;
|
cannam@167
|
124 T4j = T19 - T1a;
|
cannam@167
|
125 }
|
cannam@167
|
126 T13 = W[10];
|
cannam@167
|
127 T17 = T13 * T16;
|
cannam@167
|
128 T5a = T13 * T1b;
|
cannam@167
|
129 T4i = W[12];
|
cannam@167
|
130 T4k = T4i * T4j;
|
cannam@167
|
131 T5A = T4i * T4m;
|
cannam@167
|
132 {
|
cannam@167
|
133 E T1e, T1f, T1j, T1k;
|
cannam@167
|
134 T1e = Ip[WS(rs, 11)];
|
cannam@167
|
135 T1f = Im[WS(rs, 11)];
|
cannam@167
|
136 T1g = T1e - T1f;
|
cannam@167
|
137 T4s = T1e + T1f;
|
cannam@167
|
138 T1j = Rp[WS(rs, 11)];
|
cannam@167
|
139 T1k = Rm[WS(rs, 11)];
|
cannam@167
|
140 T1l = T1j + T1k;
|
cannam@167
|
141 T4p = T1j - T1k;
|
cannam@167
|
142 }
|
cannam@167
|
143 T1d = W[42];
|
cannam@167
|
144 T1h = T1d * T1g;
|
cannam@167
|
145 T5c = T1d * T1l;
|
cannam@167
|
146 T4o = W[44];
|
cannam@167
|
147 T4q = T4o * T4p;
|
cannam@167
|
148 T5C = T4o * T4s;
|
cannam@167
|
149 }
|
cannam@167
|
150 {
|
cannam@167
|
151 E T1c, T5b, T1m, T5d, T18, T1i;
|
cannam@167
|
152 T18 = W[11];
|
cannam@167
|
153 T1c = FNMS(T18, T1b, T17);
|
cannam@167
|
154 T5b = FMA(T18, T16, T5a);
|
cannam@167
|
155 T1i = W[43];
|
cannam@167
|
156 T1m = FNMS(T1i, T1l, T1h);
|
cannam@167
|
157 T5d = FMA(T1i, T1g, T5c);
|
cannam@167
|
158 T1n = T1c + T1m;
|
cannam@167
|
159 T7L = T1c - T1m;
|
cannam@167
|
160 T5e = T5b + T5d;
|
cannam@167
|
161 T7I = T5b - T5d;
|
cannam@167
|
162 }
|
cannam@167
|
163 {
|
cannam@167
|
164 E T4n, T5B, T4t, T5D, T4l, T4r;
|
cannam@167
|
165 T4l = W[13];
|
cannam@167
|
166 T4n = FMA(T4l, T4m, T4k);
|
cannam@167
|
167 T5B = FNMS(T4l, T4j, T5A);
|
cannam@167
|
168 T4r = W[45];
|
cannam@167
|
169 T4t = FMA(T4r, T4s, T4q);
|
cannam@167
|
170 T5D = FNMS(T4r, T4p, T5C);
|
cannam@167
|
171 T4u = T4n + T4t;
|
cannam@167
|
172 T82 = T4t - T4n;
|
cannam@167
|
173 T5E = T5B + T5D;
|
cannam@167
|
174 T7R = T5D - T5B;
|
cannam@167
|
175 }
|
cannam@167
|
176 }
|
cannam@167
|
177 {
|
cannam@167
|
178 E T2a, T2f, T3e, T3d, T5S, T3a, T3c, T2b, T4C, T2k, T2p, T3k, T3j, T5U, T3g;
|
cannam@167
|
179 E T3i, T2l, T4E;
|
cannam@167
|
180 {
|
cannam@167
|
181 E T28, T29, T3b, T2d, T2e, T27;
|
cannam@167
|
182 T28 = Ip[WS(rs, 4)];
|
cannam@167
|
183 T29 = Im[WS(rs, 4)];
|
cannam@167
|
184 T2a = T28 - T29;
|
cannam@167
|
185 T2d = Rp[WS(rs, 4)];
|
cannam@167
|
186 T2e = Rm[WS(rs, 4)];
|
cannam@167
|
187 T2f = T2d + T2e;
|
cannam@167
|
188 T3b = T2e - T2d;
|
cannam@167
|
189 T3e = T28 + T29;
|
cannam@167
|
190 T3d = W[17];
|
cannam@167
|
191 T5S = T3d * T3b;
|
cannam@167
|
192 T3a = W[16];
|
cannam@167
|
193 T3c = T3a * T3b;
|
cannam@167
|
194 T27 = W[14];
|
cannam@167
|
195 T2b = T27 * T2a;
|
cannam@167
|
196 T4C = T27 * T2f;
|
cannam@167
|
197 }
|
cannam@167
|
198 {
|
cannam@167
|
199 E T2i, T2j, T3h, T2n, T2o, T2h;
|
cannam@167
|
200 T2i = Ip[WS(rs, 12)];
|
cannam@167
|
201 T2j = Im[WS(rs, 12)];
|
cannam@167
|
202 T2k = T2i - T2j;
|
cannam@167
|
203 T2n = Rp[WS(rs, 12)];
|
cannam@167
|
204 T2o = Rm[WS(rs, 12)];
|
cannam@167
|
205 T2p = T2n + T2o;
|
cannam@167
|
206 T3h = T2o - T2n;
|
cannam@167
|
207 T3k = T2i + T2j;
|
cannam@167
|
208 T3j = W[49];
|
cannam@167
|
209 T5U = T3j * T3h;
|
cannam@167
|
210 T3g = W[48];
|
cannam@167
|
211 T3i = T3g * T3h;
|
cannam@167
|
212 T2h = W[46];
|
cannam@167
|
213 T2l = T2h * T2k;
|
cannam@167
|
214 T4E = T2h * T2p;
|
cannam@167
|
215 }
|
cannam@167
|
216 {
|
cannam@167
|
217 E T3f, T3l, T5T, T5V;
|
cannam@167
|
218 T3f = FNMS(T3d, T3e, T3c);
|
cannam@167
|
219 T3l = FNMS(T3j, T3k, T3i);
|
cannam@167
|
220 T3m = T3f + T3l;
|
cannam@167
|
221 T8k = T3f - T3l;
|
cannam@167
|
222 T5T = FMA(T3a, T3e, T5S);
|
cannam@167
|
223 T5V = FMA(T3g, T3k, T5U);
|
cannam@167
|
224 T5W = T5T + T5V;
|
cannam@167
|
225 T8a = T5T - T5V;
|
cannam@167
|
226 {
|
cannam@167
|
227 E T2g, T4D, T2q, T4F, T2c, T2m;
|
cannam@167
|
228 T2c = W[15];
|
cannam@167
|
229 T2g = FNMS(T2c, T2f, T2b);
|
cannam@167
|
230 T4D = FMA(T2c, T2a, T4C);
|
cannam@167
|
231 T2m = W[47];
|
cannam@167
|
232 T2q = FNMS(T2m, T2p, T2l);
|
cannam@167
|
233 T4F = FMA(T2m, T2k, T4E);
|
cannam@167
|
234 T2r = T2g + T2q;
|
cannam@167
|
235 T8u = T2g - T2q;
|
cannam@167
|
236 T4G = T4D + T4F;
|
cannam@167
|
237 T7q = T4D - T4F;
|
cannam@167
|
238 }
|
cannam@167
|
239 }
|
cannam@167
|
240 }
|
cannam@167
|
241 {
|
cannam@167
|
242 E TL, T49, TQ, T46, TM, T55, T47, T5v, TV, T4f, T10, T4c, TW, T57, T4d;
|
cannam@167
|
243 E T5x;
|
cannam@167
|
244 {
|
cannam@167
|
245 E TI, T45, TS, T4b;
|
cannam@167
|
246 {
|
cannam@167
|
247 E TJ, TK, TO, TP;
|
cannam@167
|
248 TJ = Ip[WS(rs, 15)];
|
cannam@167
|
249 TK = Im[WS(rs, 15)];
|
cannam@167
|
250 TL = TJ - TK;
|
cannam@167
|
251 T49 = TJ + TK;
|
cannam@167
|
252 TO = Rp[WS(rs, 15)];
|
cannam@167
|
253 TP = Rm[WS(rs, 15)];
|
cannam@167
|
254 TQ = TO + TP;
|
cannam@167
|
255 T46 = TO - TP;
|
cannam@167
|
256 }
|
cannam@167
|
257 TI = W[58];
|
cannam@167
|
258 TM = TI * TL;
|
cannam@167
|
259 T55 = TI * TQ;
|
cannam@167
|
260 T45 = W[60];
|
cannam@167
|
261 T47 = T45 * T46;
|
cannam@167
|
262 T5v = T45 * T49;
|
cannam@167
|
263 {
|
cannam@167
|
264 E TT, TU, TY, TZ;
|
cannam@167
|
265 TT = Ip[WS(rs, 7)];
|
cannam@167
|
266 TU = Im[WS(rs, 7)];
|
cannam@167
|
267 TV = TT - TU;
|
cannam@167
|
268 T4f = TT + TU;
|
cannam@167
|
269 TY = Rp[WS(rs, 7)];
|
cannam@167
|
270 TZ = Rm[WS(rs, 7)];
|
cannam@167
|
271 T10 = TY + TZ;
|
cannam@167
|
272 T4c = TY - TZ;
|
cannam@167
|
273 }
|
cannam@167
|
274 TS = W[26];
|
cannam@167
|
275 TW = TS * TV;
|
cannam@167
|
276 T57 = TS * T10;
|
cannam@167
|
277 T4b = W[28];
|
cannam@167
|
278 T4d = T4b * T4c;
|
cannam@167
|
279 T5x = T4b * T4f;
|
cannam@167
|
280 }
|
cannam@167
|
281 {
|
cannam@167
|
282 E TR, T56, T11, T58, TN, TX;
|
cannam@167
|
283 TN = W[59];
|
cannam@167
|
284 TR = FNMS(TN, TQ, TM);
|
cannam@167
|
285 T56 = FMA(TN, TL, T55);
|
cannam@167
|
286 TX = W[27];
|
cannam@167
|
287 T11 = FNMS(TX, T10, TW);
|
cannam@167
|
288 T58 = FMA(TX, TV, T57);
|
cannam@167
|
289 T12 = TR + T11;
|
cannam@167
|
290 T7K = T56 - T58;
|
cannam@167
|
291 T59 = T56 + T58;
|
cannam@167
|
292 T7H = TR - T11;
|
cannam@167
|
293 }
|
cannam@167
|
294 {
|
cannam@167
|
295 E T4a, T5w, T4g, T5y, T48, T4e;
|
cannam@167
|
296 T48 = W[61];
|
cannam@167
|
297 T4a = FMA(T48, T49, T47);
|
cannam@167
|
298 T5w = FNMS(T48, T46, T5v);
|
cannam@167
|
299 T4e = W[29];
|
cannam@167
|
300 T4g = FMA(T4e, T4f, T4d);
|
cannam@167
|
301 T5y = FNMS(T4e, T4c, T5x);
|
cannam@167
|
302 T4h = T4a + T4g;
|
cannam@167
|
303 T81 = T5w - T5y;
|
cannam@167
|
304 T5z = T5w + T5y;
|
cannam@167
|
305 T7Q = T4g - T4a;
|
cannam@167
|
306 }
|
cannam@167
|
307 }
|
cannam@167
|
308 {
|
cannam@167
|
309 E T4, T3I, T9, T3F, T5, T4U, T3G, T5k, Te, T3O, Tj, T3L, Tf, T4W, T3M;
|
cannam@167
|
310 E T5m;
|
cannam@167
|
311 {
|
cannam@167
|
312 E T1, T3E, Tb, T3K;
|
cannam@167
|
313 {
|
cannam@167
|
314 E T2, T3, T7, T8;
|
cannam@167
|
315 T2 = Ip[WS(rs, 1)];
|
cannam@167
|
316 T3 = Im[WS(rs, 1)];
|
cannam@167
|
317 T4 = T2 - T3;
|
cannam@167
|
318 T3I = T2 + T3;
|
cannam@167
|
319 T7 = Rp[WS(rs, 1)];
|
cannam@167
|
320 T8 = Rm[WS(rs, 1)];
|
cannam@167
|
321 T9 = T7 + T8;
|
cannam@167
|
322 T3F = T7 - T8;
|
cannam@167
|
323 }
|
cannam@167
|
324 T1 = W[2];
|
cannam@167
|
325 T5 = T1 * T4;
|
cannam@167
|
326 T4U = T1 * T9;
|
cannam@167
|
327 T3E = W[4];
|
cannam@167
|
328 T3G = T3E * T3F;
|
cannam@167
|
329 T5k = T3E * T3I;
|
cannam@167
|
330 {
|
cannam@167
|
331 E Tc, Td, Th, Ti;
|
cannam@167
|
332 Tc = Ip[WS(rs, 9)];
|
cannam@167
|
333 Td = Im[WS(rs, 9)];
|
cannam@167
|
334 Te = Tc - Td;
|
cannam@167
|
335 T3O = Tc + Td;
|
cannam@167
|
336 Th = Rp[WS(rs, 9)];
|
cannam@167
|
337 Ti = Rm[WS(rs, 9)];
|
cannam@167
|
338 Tj = Th + Ti;
|
cannam@167
|
339 T3L = Th - Ti;
|
cannam@167
|
340 }
|
cannam@167
|
341 Tb = W[34];
|
cannam@167
|
342 Tf = Tb * Te;
|
cannam@167
|
343 T4W = Tb * Tj;
|
cannam@167
|
344 T3K = W[36];
|
cannam@167
|
345 T3M = T3K * T3L;
|
cannam@167
|
346 T5m = T3K * T3O;
|
cannam@167
|
347 }
|
cannam@167
|
348 {
|
cannam@167
|
349 E Ta, T4V, Tk, T4X, T6, Tg;
|
cannam@167
|
350 T6 = W[3];
|
cannam@167
|
351 Ta = FNMS(T6, T9, T5);
|
cannam@167
|
352 T4V = FMA(T6, T4, T4U);
|
cannam@167
|
353 Tg = W[35];
|
cannam@167
|
354 Tk = FNMS(Tg, Tj, Tf);
|
cannam@167
|
355 T4X = FMA(Tg, Te, T4W);
|
cannam@167
|
356 Tl = Ta + Tk;
|
cannam@167
|
357 T7D = T4V - T4X;
|
cannam@167
|
358 T4Y = T4V + T4X;
|
cannam@167
|
359 T7A = Ta - Tk;
|
cannam@167
|
360 }
|
cannam@167
|
361 {
|
cannam@167
|
362 E T3J, T5l, T3P, T5n, T3H, T3N, T7T, T7U;
|
cannam@167
|
363 T3H = W[5];
|
cannam@167
|
364 T3J = FMA(T3H, T3I, T3G);
|
cannam@167
|
365 T5l = FNMS(T3H, T3F, T5k);
|
cannam@167
|
366 T3N = W[37];
|
cannam@167
|
367 T3P = FMA(T3N, T3O, T3M);
|
cannam@167
|
368 T5n = FNMS(T3N, T3L, T5m);
|
cannam@167
|
369 T3Q = T3J + T3P;
|
cannam@167
|
370 T5o = T5l + T5n;
|
cannam@167
|
371 T7T = T3P - T3J;
|
cannam@167
|
372 T7U = T5l - T5n;
|
cannam@167
|
373 T7V = T7T - T7U;
|
cannam@167
|
374 T84 = T7U + T7T;
|
cannam@167
|
375 }
|
cannam@167
|
376 }
|
cannam@167
|
377 {
|
cannam@167
|
378 E T1t, T1y, T2N, T2M, T5H, T2J, T2L, T1u, T4I, T1D, T1I, T2T, T2S, T5J, T2P;
|
cannam@167
|
379 E T2R, T1E, T4K;
|
cannam@167
|
380 {
|
cannam@167
|
381 E T1r, T1s, T2K, T1w, T1x, T1q;
|
cannam@167
|
382 T1r = Ip[WS(rs, 2)];
|
cannam@167
|
383 T1s = Im[WS(rs, 2)];
|
cannam@167
|
384 T1t = T1r - T1s;
|
cannam@167
|
385 T1w = Rp[WS(rs, 2)];
|
cannam@167
|
386 T1x = Rm[WS(rs, 2)];
|
cannam@167
|
387 T1y = T1w + T1x;
|
cannam@167
|
388 T2K = T1x - T1w;
|
cannam@167
|
389 T2N = T1r + T1s;
|
cannam@167
|
390 T2M = W[9];
|
cannam@167
|
391 T5H = T2M * T2K;
|
cannam@167
|
392 T2J = W[8];
|
cannam@167
|
393 T2L = T2J * T2K;
|
cannam@167
|
394 T1q = W[6];
|
cannam@167
|
395 T1u = T1q * T1t;
|
cannam@167
|
396 T4I = T1q * T1y;
|
cannam@167
|
397 }
|
cannam@167
|
398 {
|
cannam@167
|
399 E T1B, T1C, T2Q, T1G, T1H, T1A;
|
cannam@167
|
400 T1B = Ip[WS(rs, 10)];
|
cannam@167
|
401 T1C = Im[WS(rs, 10)];
|
cannam@167
|
402 T1D = T1B - T1C;
|
cannam@167
|
403 T1G = Rp[WS(rs, 10)];
|
cannam@167
|
404 T1H = Rm[WS(rs, 10)];
|
cannam@167
|
405 T1I = T1G + T1H;
|
cannam@167
|
406 T2Q = T1H - T1G;
|
cannam@167
|
407 T2T = T1B + T1C;
|
cannam@167
|
408 T2S = W[41];
|
cannam@167
|
409 T5J = T2S * T2Q;
|
cannam@167
|
410 T2P = W[40];
|
cannam@167
|
411 T2R = T2P * T2Q;
|
cannam@167
|
412 T1A = W[38];
|
cannam@167
|
413 T1E = T1A * T1D;
|
cannam@167
|
414 T4K = T1A * T1I;
|
cannam@167
|
415 }
|
cannam@167
|
416 {
|
cannam@167
|
417 E T1z, T4J, T1J, T4L, T1v, T1F;
|
cannam@167
|
418 T1v = W[7];
|
cannam@167
|
419 T1z = FNMS(T1v, T1y, T1u);
|
cannam@167
|
420 T4J = FMA(T1v, T1t, T4I);
|
cannam@167
|
421 T1F = W[39];
|
cannam@167
|
422 T1J = FNMS(T1F, T1I, T1E);
|
cannam@167
|
423 T4L = FMA(T1F, T1D, T4K);
|
cannam@167
|
424 T1K = T1z + T1J;
|
cannam@167
|
425 T7t = T4J - T4L;
|
cannam@167
|
426 T4M = T4J + T4L;
|
cannam@167
|
427 T7s = T1z - T1J;
|
cannam@167
|
428 }
|
cannam@167
|
429 {
|
cannam@167
|
430 E T2O, T2U, T8c, T5I, T5K, T8d;
|
cannam@167
|
431 T2O = FNMS(T2M, T2N, T2L);
|
cannam@167
|
432 T2U = FNMS(T2S, T2T, T2R);
|
cannam@167
|
433 T8c = T2O - T2U;
|
cannam@167
|
434 T5I = FMA(T2J, T2N, T5H);
|
cannam@167
|
435 T5K = FMA(T2P, T2T, T5J);
|
cannam@167
|
436 T8d = T5I - T5K;
|
cannam@167
|
437 T2V = T2O + T2U;
|
cannam@167
|
438 T8n = T8c + T8d;
|
cannam@167
|
439 T5L = T5I + T5K;
|
cannam@167
|
440 T8e = T8c - T8d;
|
cannam@167
|
441 }
|
cannam@167
|
442 }
|
cannam@167
|
443 {
|
cannam@167
|
444 E T1O, T1T, T30, T2Z, T5M, T2W, T2Y, T1P, T4N, T1Y, T23, T36, T35, T5O, T32;
|
cannam@167
|
445 E T34, T1Z, T4P;
|
cannam@167
|
446 {
|
cannam@167
|
447 E T1M, T1N, T2X, T1R, T1S, T1L;
|
cannam@167
|
448 T1M = Ip[WS(rs, 14)];
|
cannam@167
|
449 T1N = Im[WS(rs, 14)];
|
cannam@167
|
450 T1O = T1M - T1N;
|
cannam@167
|
451 T1R = Rp[WS(rs, 14)];
|
cannam@167
|
452 T1S = Rm[WS(rs, 14)];
|
cannam@167
|
453 T1T = T1R + T1S;
|
cannam@167
|
454 T2X = T1S - T1R;
|
cannam@167
|
455 T30 = T1M + T1N;
|
cannam@167
|
456 T2Z = W[57];
|
cannam@167
|
457 T5M = T2Z * T2X;
|
cannam@167
|
458 T2W = W[56];
|
cannam@167
|
459 T2Y = T2W * T2X;
|
cannam@167
|
460 T1L = W[54];
|
cannam@167
|
461 T1P = T1L * T1O;
|
cannam@167
|
462 T4N = T1L * T1T;
|
cannam@167
|
463 }
|
cannam@167
|
464 {
|
cannam@167
|
465 E T1W, T1X, T33, T21, T22, T1V;
|
cannam@167
|
466 T1W = Ip[WS(rs, 6)];
|
cannam@167
|
467 T1X = Im[WS(rs, 6)];
|
cannam@167
|
468 T1Y = T1W - T1X;
|
cannam@167
|
469 T21 = Rp[WS(rs, 6)];
|
cannam@167
|
470 T22 = Rm[WS(rs, 6)];
|
cannam@167
|
471 T23 = T21 + T22;
|
cannam@167
|
472 T33 = T22 - T21;
|
cannam@167
|
473 T36 = T1W + T1X;
|
cannam@167
|
474 T35 = W[25];
|
cannam@167
|
475 T5O = T35 * T33;
|
cannam@167
|
476 T32 = W[24];
|
cannam@167
|
477 T34 = T32 * T33;
|
cannam@167
|
478 T1V = W[22];
|
cannam@167
|
479 T1Z = T1V * T1Y;
|
cannam@167
|
480 T4P = T1V * T23;
|
cannam@167
|
481 }
|
cannam@167
|
482 {
|
cannam@167
|
483 E T1U, T4O, T24, T4Q, T1Q, T20;
|
cannam@167
|
484 T1Q = W[55];
|
cannam@167
|
485 T1U = FNMS(T1Q, T1T, T1P);
|
cannam@167
|
486 T4O = FMA(T1Q, T1O, T4N);
|
cannam@167
|
487 T20 = W[23];
|
cannam@167
|
488 T24 = FNMS(T20, T23, T1Z);
|
cannam@167
|
489 T4Q = FMA(T20, T1Y, T4P);
|
cannam@167
|
490 T25 = T1U + T24;
|
cannam@167
|
491 T7w = T1U - T24;
|
cannam@167
|
492 T4R = T4O + T4Q;
|
cannam@167
|
493 T7v = T4O - T4Q;
|
cannam@167
|
494 }
|
cannam@167
|
495 {
|
cannam@167
|
496 E T31, T37, T8f, T5N, T5P, T8g;
|
cannam@167
|
497 T31 = FNMS(T2Z, T30, T2Y);
|
cannam@167
|
498 T37 = FNMS(T35, T36, T34);
|
cannam@167
|
499 T8f = T31 - T37;
|
cannam@167
|
500 T5N = FMA(T2W, T30, T5M);
|
cannam@167
|
501 T5P = FMA(T32, T36, T5O);
|
cannam@167
|
502 T8g = T5N - T5P;
|
cannam@167
|
503 T38 = T31 + T37;
|
cannam@167
|
504 T8o = T8g - T8f;
|
cannam@167
|
505 T5Q = T5N + T5P;
|
cannam@167
|
506 T8h = T8f + T8g;
|
cannam@167
|
507 }
|
cannam@167
|
508 }
|
cannam@167
|
509 {
|
cannam@167
|
510 E Tp, T3V, Tu, T3S, Tq, T4Z, T3T, T5p, Tz, T41, TE, T3Y, TA, T51, T3Z;
|
cannam@167
|
511 E T5r;
|
cannam@167
|
512 {
|
cannam@167
|
513 E Tm, T3R, Tw, T3X;
|
cannam@167
|
514 {
|
cannam@167
|
515 E Tn, To, Ts, Tt;
|
cannam@167
|
516 Tn = Ip[WS(rs, 5)];
|
cannam@167
|
517 To = Im[WS(rs, 5)];
|
cannam@167
|
518 Tp = Tn - To;
|
cannam@167
|
519 T3V = Tn + To;
|
cannam@167
|
520 Ts = Rp[WS(rs, 5)];
|
cannam@167
|
521 Tt = Rm[WS(rs, 5)];
|
cannam@167
|
522 Tu = Ts + Tt;
|
cannam@167
|
523 T3S = Ts - Tt;
|
cannam@167
|
524 }
|
cannam@167
|
525 Tm = W[18];
|
cannam@167
|
526 Tq = Tm * Tp;
|
cannam@167
|
527 T4Z = Tm * Tu;
|
cannam@167
|
528 T3R = W[20];
|
cannam@167
|
529 T3T = T3R * T3S;
|
cannam@167
|
530 T5p = T3R * T3V;
|
cannam@167
|
531 {
|
cannam@167
|
532 E Tx, Ty, TC, TD;
|
cannam@167
|
533 Tx = Ip[WS(rs, 13)];
|
cannam@167
|
534 Ty = Im[WS(rs, 13)];
|
cannam@167
|
535 Tz = Tx - Ty;
|
cannam@167
|
536 T41 = Tx + Ty;
|
cannam@167
|
537 TC = Rp[WS(rs, 13)];
|
cannam@167
|
538 TD = Rm[WS(rs, 13)];
|
cannam@167
|
539 TE = TC + TD;
|
cannam@167
|
540 T3Y = TC - TD;
|
cannam@167
|
541 }
|
cannam@167
|
542 Tw = W[50];
|
cannam@167
|
543 TA = Tw * Tz;
|
cannam@167
|
544 T51 = Tw * TE;
|
cannam@167
|
545 T3X = W[52];
|
cannam@167
|
546 T3Z = T3X * T3Y;
|
cannam@167
|
547 T5r = T3X * T41;
|
cannam@167
|
548 }
|
cannam@167
|
549 {
|
cannam@167
|
550 E Tv, T50, TF, T52, Tr, TB;
|
cannam@167
|
551 Tr = W[19];
|
cannam@167
|
552 Tv = FNMS(Tr, Tu, Tq);
|
cannam@167
|
553 T50 = FMA(Tr, Tp, T4Z);
|
cannam@167
|
554 TB = W[51];
|
cannam@167
|
555 TF = FNMS(TB, TE, TA);
|
cannam@167
|
556 T52 = FMA(TB, Tz, T51);
|
cannam@167
|
557 TG = Tv + TF;
|
cannam@167
|
558 T7E = Tv - TF;
|
cannam@167
|
559 T53 = T50 + T52;
|
cannam@167
|
560 T7B = T50 - T52;
|
cannam@167
|
561 }
|
cannam@167
|
562 {
|
cannam@167
|
563 E T3W, T5q, T42, T5s, T3U, T40, T7W, T7X;
|
cannam@167
|
564 T3U = W[21];
|
cannam@167
|
565 T3W = FMA(T3U, T3V, T3T);
|
cannam@167
|
566 T5q = FNMS(T3U, T3S, T5p);
|
cannam@167
|
567 T40 = W[53];
|
cannam@167
|
568 T42 = FMA(T40, T41, T3Z);
|
cannam@167
|
569 T5s = FNMS(T40, T3Y, T5r);
|
cannam@167
|
570 T43 = T3W + T42;
|
cannam@167
|
571 T5t = T5q + T5s;
|
cannam@167
|
572 T7W = T5s - T5q;
|
cannam@167
|
573 T7X = T3W - T42;
|
cannam@167
|
574 T7Y = T7W + T7X;
|
cannam@167
|
575 T85 = T7W - T7X;
|
cannam@167
|
576 }
|
cannam@167
|
577 }
|
cannam@167
|
578 {
|
cannam@167
|
579 E T1p, T6i, T2H, T68, T5g, T67, T4T, T6h, T4w, T6m, T5G, T6c, T3D, T6n, T63;
|
cannam@167
|
580 E T6f;
|
cannam@167
|
581 {
|
cannam@167
|
582 E TH, T1o, T4H, T4S;
|
cannam@167
|
583 TH = Tl + TG;
|
cannam@167
|
584 T1o = T12 + T1n;
|
cannam@167
|
585 T1p = TH + T1o;
|
cannam@167
|
586 T6i = TH - T1o;
|
cannam@167
|
587 {
|
cannam@167
|
588 E T26, T2G, T54, T5f;
|
cannam@167
|
589 T26 = T1K + T25;
|
cannam@167
|
590 T2G = T2r + T2F;
|
cannam@167
|
591 T2H = T26 + T2G;
|
cannam@167
|
592 T68 = T2G - T26;
|
cannam@167
|
593 T54 = T4Y + T53;
|
cannam@167
|
594 T5f = T59 + T5e;
|
cannam@167
|
595 T5g = T54 + T5f;
|
cannam@167
|
596 T67 = T5f - T54;
|
cannam@167
|
597 }
|
cannam@167
|
598 T4H = T4B + T4G;
|
cannam@167
|
599 T4S = T4M + T4R;
|
cannam@167
|
600 T4T = T4H + T4S;
|
cannam@167
|
601 T6h = T4H - T4S;
|
cannam@167
|
602 {
|
cannam@167
|
603 E T44, T4v, T6b, T5u, T5F, T6a;
|
cannam@167
|
604 T44 = T3Q + T43;
|
cannam@167
|
605 T4v = T4h + T4u;
|
cannam@167
|
606 T6b = T44 - T4v;
|
cannam@167
|
607 T5u = T5o + T5t;
|
cannam@167
|
608 T5F = T5z + T5E;
|
cannam@167
|
609 T6a = T5F - T5u;
|
cannam@167
|
610 T4w = T44 + T4v;
|
cannam@167
|
611 T6m = T6a - T6b;
|
cannam@167
|
612 T5G = T5u + T5F;
|
cannam@167
|
613 T6c = T6a + T6b;
|
cannam@167
|
614 }
|
cannam@167
|
615 {
|
cannam@167
|
616 E T39, T3C, T6d, T5R, T62, T6e;
|
cannam@167
|
617 T39 = T2V + T38;
|
cannam@167
|
618 T3C = T3m + T3B;
|
cannam@167
|
619 T6d = T3C - T39;
|
cannam@167
|
620 T5R = T5L + T5Q;
|
cannam@167
|
621 T62 = T5W + T61;
|
cannam@167
|
622 T6e = T62 - T5R;
|
cannam@167
|
623 T3D = T39 + T3C;
|
cannam@167
|
624 T6n = T6d + T6e;
|
cannam@167
|
625 T63 = T5R + T62;
|
cannam@167
|
626 T6f = T6d - T6e;
|
cannam@167
|
627 }
|
cannam@167
|
628 }
|
cannam@167
|
629 {
|
cannam@167
|
630 E T2I, T4x, T65, T66;
|
cannam@167
|
631 T2I = T1p + T2H;
|
cannam@167
|
632 T4x = T3D - T4w;
|
cannam@167
|
633 Ip[0] = KP500000000 * (T2I + T4x);
|
cannam@167
|
634 Im[WS(rs, 15)] = KP500000000 * (T4x - T2I);
|
cannam@167
|
635 T65 = T4T + T5g;
|
cannam@167
|
636 T66 = T5G + T63;
|
cannam@167
|
637 Rm[WS(rs, 15)] = KP500000000 * (T65 - T66);
|
cannam@167
|
638 Rp[0] = KP500000000 * (T65 + T66);
|
cannam@167
|
639 }
|
cannam@167
|
640 {
|
cannam@167
|
641 E T5h, T5i, T5j, T64;
|
cannam@167
|
642 T5h = T4T - T5g;
|
cannam@167
|
643 T5i = T4w + T3D;
|
cannam@167
|
644 Rm[WS(rs, 7)] = KP500000000 * (T5h - T5i);
|
cannam@167
|
645 Rp[WS(rs, 8)] = KP500000000 * (T5h + T5i);
|
cannam@167
|
646 T5j = T2H - T1p;
|
cannam@167
|
647 T64 = T5G - T63;
|
cannam@167
|
648 Ip[WS(rs, 8)] = KP500000000 * (T5j + T64);
|
cannam@167
|
649 Im[WS(rs, 7)] = KP500000000 * (T64 - T5j);
|
cannam@167
|
650 }
|
cannam@167
|
651 {
|
cannam@167
|
652 E T69, T6g, T6p, T6q;
|
cannam@167
|
653 T69 = T67 + T68;
|
cannam@167
|
654 T6g = T6c + T6f;
|
cannam@167
|
655 Ip[WS(rs, 4)] = KP500000000 * (FMA(KP707106781, T6g, T69));
|
cannam@167
|
656 Im[WS(rs, 11)] = -(KP500000000 * (FNMS(KP707106781, T6g, T69)));
|
cannam@167
|
657 T6p = T6h + T6i;
|
cannam@167
|
658 T6q = T6m + T6n;
|
cannam@167
|
659 Rm[WS(rs, 11)] = KP500000000 * (FNMS(KP707106781, T6q, T6p));
|
cannam@167
|
660 Rp[WS(rs, 4)] = KP500000000 * (FMA(KP707106781, T6q, T6p));
|
cannam@167
|
661 }
|
cannam@167
|
662 {
|
cannam@167
|
663 E T6j, T6k, T6l, T6o;
|
cannam@167
|
664 T6j = T6h - T6i;
|
cannam@167
|
665 T6k = T6f - T6c;
|
cannam@167
|
666 Rm[WS(rs, 3)] = KP500000000 * (FNMS(KP707106781, T6k, T6j));
|
cannam@167
|
667 Rp[WS(rs, 12)] = KP500000000 * (FMA(KP707106781, T6k, T6j));
|
cannam@167
|
668 T6l = T68 - T67;
|
cannam@167
|
669 T6o = T6m - T6n;
|
cannam@167
|
670 Ip[WS(rs, 12)] = KP500000000 * (FMA(KP707106781, T6o, T6l));
|
cannam@167
|
671 Im[WS(rs, 3)] = -(KP500000000 * (FNMS(KP707106781, T6o, T6l)));
|
cannam@167
|
672 }
|
cannam@167
|
673 }
|
cannam@167
|
674 {
|
cannam@167
|
675 E T6t, T75, T6T, T7f, T6A, T7g, T6W, T76, T6I, T7k, T70, T7a, T6P, T7l, T71;
|
cannam@167
|
676 E T7d;
|
cannam@167
|
677 {
|
cannam@167
|
678 E T6r, T6s, T6R, T6S;
|
cannam@167
|
679 T6r = T4R - T4M;
|
cannam@167
|
680 T6s = T2F - T2r;
|
cannam@167
|
681 T6t = T6r + T6s;
|
cannam@167
|
682 T75 = T6s - T6r;
|
cannam@167
|
683 T6R = T4B - T4G;
|
cannam@167
|
684 T6S = T1K - T25;
|
cannam@167
|
685 T6T = T6R + T6S;
|
cannam@167
|
686 T7f = T6R - T6S;
|
cannam@167
|
687 }
|
cannam@167
|
688 {
|
cannam@167
|
689 E T6w, T6U, T6z, T6V;
|
cannam@167
|
690 {
|
cannam@167
|
691 E T6u, T6v, T6x, T6y;
|
cannam@167
|
692 T6u = Tl - TG;
|
cannam@167
|
693 T6v = T4Y - T53;
|
cannam@167
|
694 T6w = T6u - T6v;
|
cannam@167
|
695 T6U = T6v + T6u;
|
cannam@167
|
696 T6x = T59 - T5e;
|
cannam@167
|
697 T6y = T12 - T1n;
|
cannam@167
|
698 T6z = T6x + T6y;
|
cannam@167
|
699 T6V = T6x - T6y;
|
cannam@167
|
700 }
|
cannam@167
|
701 T6A = T6w + T6z;
|
cannam@167
|
702 T7g = T6w - T6z;
|
cannam@167
|
703 T6W = T6U + T6V;
|
cannam@167
|
704 T76 = T6V - T6U;
|
cannam@167
|
705 }
|
cannam@167
|
706 {
|
cannam@167
|
707 E T6E, T78, T6H, T79;
|
cannam@167
|
708 {
|
cannam@167
|
709 E T6C, T6D, T6F, T6G;
|
cannam@167
|
710 T6C = T5t - T5o;
|
cannam@167
|
711 T6D = T4u - T4h;
|
cannam@167
|
712 T6E = T6C + T6D;
|
cannam@167
|
713 T78 = T6C - T6D;
|
cannam@167
|
714 T6F = T43 - T3Q;
|
cannam@167
|
715 T6G = T5z - T5E;
|
cannam@167
|
716 T6H = T6F + T6G;
|
cannam@167
|
717 T79 = T6G - T6F;
|
cannam@167
|
718 }
|
cannam@167
|
719 T6I = FMA(KP414213562, T6H, T6E);
|
cannam@167
|
720 T7k = FNMS(KP414213562, T78, T79);
|
cannam@167
|
721 T70 = FNMS(KP414213562, T6E, T6H);
|
cannam@167
|
722 T7a = FMA(KP414213562, T79, T78);
|
cannam@167
|
723 }
|
cannam@167
|
724 {
|
cannam@167
|
725 E T6L, T7b, T6O, T7c;
|
cannam@167
|
726 {
|
cannam@167
|
727 E T6J, T6K, T6M, T6N;
|
cannam@167
|
728 T6J = T5Q - T5L;
|
cannam@167
|
729 T6K = T3B - T3m;
|
cannam@167
|
730 T6L = T6J + T6K;
|
cannam@167
|
731 T7b = T6K - T6J;
|
cannam@167
|
732 T6M = T2V - T38;
|
cannam@167
|
733 T6N = T61 - T5W;
|
cannam@167
|
734 T6O = T6M + T6N;
|
cannam@167
|
735 T7c = T6N - T6M;
|
cannam@167
|
736 }
|
cannam@167
|
737 T6P = FNMS(KP414213562, T6O, T6L);
|
cannam@167
|
738 T7l = FNMS(KP414213562, T7b, T7c);
|
cannam@167
|
739 T71 = FMA(KP414213562, T6L, T6O);
|
cannam@167
|
740 T7d = FMA(KP414213562, T7c, T7b);
|
cannam@167
|
741 }
|
cannam@167
|
742 {
|
cannam@167
|
743 E T6B, T6Q, T73, T74;
|
cannam@167
|
744 T6B = FMA(KP707106781, T6A, T6t);
|
cannam@167
|
745 T6Q = T6I + T6P;
|
cannam@167
|
746 Ip[WS(rs, 2)] = KP500000000 * (FMA(KP923879532, T6Q, T6B));
|
cannam@167
|
747 Im[WS(rs, 13)] = -(KP500000000 * (FNMS(KP923879532, T6Q, T6B)));
|
cannam@167
|
748 T73 = FMA(KP707106781, T6W, T6T);
|
cannam@167
|
749 T74 = T70 + T71;
|
cannam@167
|
750 Rm[WS(rs, 13)] = KP500000000 * (FNMS(KP923879532, T74, T73));
|
cannam@167
|
751 Rp[WS(rs, 2)] = KP500000000 * (FMA(KP923879532, T74, T73));
|
cannam@167
|
752 }
|
cannam@167
|
753 {
|
cannam@167
|
754 E T6X, T6Y, T6Z, T72;
|
cannam@167
|
755 T6X = FNMS(KP707106781, T6W, T6T);
|
cannam@167
|
756 T6Y = T6P - T6I;
|
cannam@167
|
757 Rm[WS(rs, 5)] = KP500000000 * (FNMS(KP923879532, T6Y, T6X));
|
cannam@167
|
758 Rp[WS(rs, 10)] = KP500000000 * (FMA(KP923879532, T6Y, T6X));
|
cannam@167
|
759 T6Z = FNMS(KP707106781, T6A, T6t);
|
cannam@167
|
760 T72 = T70 - T71;
|
cannam@167
|
761 Ip[WS(rs, 10)] = KP500000000 * (FMA(KP923879532, T72, T6Z));
|
cannam@167
|
762 Im[WS(rs, 5)] = -(KP500000000 * (FNMS(KP923879532, T72, T6Z)));
|
cannam@167
|
763 }
|
cannam@167
|
764 {
|
cannam@167
|
765 E T77, T7e, T7n, T7o;
|
cannam@167
|
766 T77 = FNMS(KP707106781, T76, T75);
|
cannam@167
|
767 T7e = T7a - T7d;
|
cannam@167
|
768 Ip[WS(rs, 14)] = KP500000000 * (FMA(KP923879532, T7e, T77));
|
cannam@167
|
769 Im[WS(rs, 1)] = -(KP500000000 * (FNMS(KP923879532, T7e, T77)));
|
cannam@167
|
770 T7n = FNMS(KP707106781, T7g, T7f);
|
cannam@167
|
771 T7o = T7k + T7l;
|
cannam@167
|
772 Rp[WS(rs, 14)] = KP500000000 * (FNMS(KP923879532, T7o, T7n));
|
cannam@167
|
773 Rm[WS(rs, 1)] = KP500000000 * (FMA(KP923879532, T7o, T7n));
|
cannam@167
|
774 }
|
cannam@167
|
775 {
|
cannam@167
|
776 E T7h, T7i, T7j, T7m;
|
cannam@167
|
777 T7h = FMA(KP707106781, T7g, T7f);
|
cannam@167
|
778 T7i = T7a + T7d;
|
cannam@167
|
779 Rm[WS(rs, 9)] = KP500000000 * (FNMS(KP923879532, T7i, T7h));
|
cannam@167
|
780 Rp[WS(rs, 6)] = KP500000000 * (FMA(KP923879532, T7i, T7h));
|
cannam@167
|
781 T7j = FMA(KP707106781, T76, T75);
|
cannam@167
|
782 T7m = T7k - T7l;
|
cannam@167
|
783 Ip[WS(rs, 6)] = KP500000000 * (FMA(KP923879532, T7m, T7j));
|
cannam@167
|
784 Im[WS(rs, 9)] = -(KP500000000 * (FNMS(KP923879532, T7m, T7j)));
|
cannam@167
|
785 }
|
cannam@167
|
786 }
|
cannam@167
|
787 {
|
cannam@167
|
788 E T7z, T9T, T8L, T9x, T8z, T9J, T8V, T97, T7O, T8W, T8C, T8M, T9t, T9Y, T9E;
|
cannam@167
|
789 E T9O, T88, T90, T8G, T8Q, T9e, T9U, T9A, T9K, T9m, T9Z, T9F, T9R, T8r, T91;
|
cannam@167
|
790 E T8H, T8T;
|
cannam@167
|
791 {
|
cannam@167
|
792 E T7r, T9v, T7y, T9w, T7u, T7x;
|
cannam@167
|
793 T7r = T7p - T7q;
|
cannam@167
|
794 T9v = T8t - T8u;
|
cannam@167
|
795 T7u = T7s - T7t;
|
cannam@167
|
796 T7x = T7v + T7w;
|
cannam@167
|
797 T7y = T7u + T7x;
|
cannam@167
|
798 T9w = T7u - T7x;
|
cannam@167
|
799 T7z = FMA(KP707106781, T7y, T7r);
|
cannam@167
|
800 T9T = FNMS(KP707106781, T9w, T9v);
|
cannam@167
|
801 T8L = FNMS(KP707106781, T7y, T7r);
|
cannam@167
|
802 T9x = FMA(KP707106781, T9w, T9v);
|
cannam@167
|
803 }
|
cannam@167
|
804 {
|
cannam@167
|
805 E T8v, T95, T8y, T96, T8w, T8x;
|
cannam@167
|
806 T8v = T8t + T8u;
|
cannam@167
|
807 T95 = T7q + T7p;
|
cannam@167
|
808 T8w = T7t + T7s;
|
cannam@167
|
809 T8x = T7v - T7w;
|
cannam@167
|
810 T8y = T8w + T8x;
|
cannam@167
|
811 T96 = T8x - T8w;
|
cannam@167
|
812 T8z = FMA(KP707106781, T8y, T8v);
|
cannam@167
|
813 T9J = FNMS(KP707106781, T96, T95);
|
cannam@167
|
814 T8V = FNMS(KP707106781, T8y, T8v);
|
cannam@167
|
815 T97 = FMA(KP707106781, T96, T95);
|
cannam@167
|
816 }
|
cannam@167
|
817 {
|
cannam@167
|
818 E T7G, T8A, T7N, T8B;
|
cannam@167
|
819 {
|
cannam@167
|
820 E T7C, T7F, T7J, T7M;
|
cannam@167
|
821 T7C = T7A - T7B;
|
cannam@167
|
822 T7F = T7D + T7E;
|
cannam@167
|
823 T7G = FNMS(KP414213562, T7F, T7C);
|
cannam@167
|
824 T8A = FMA(KP414213562, T7C, T7F);
|
cannam@167
|
825 T7J = T7H - T7I;
|
cannam@167
|
826 T7M = T7K + T7L;
|
cannam@167
|
827 T7N = FMA(KP414213562, T7M, T7J);
|
cannam@167
|
828 T8B = FNMS(KP414213562, T7J, T7M);
|
cannam@167
|
829 }
|
cannam@167
|
830 T7O = T7G + T7N;
|
cannam@167
|
831 T8W = T7G - T7N;
|
cannam@167
|
832 T8C = T8A + T8B;
|
cannam@167
|
833 T8M = T8B - T8A;
|
cannam@167
|
834 }
|
cannam@167
|
835 {
|
cannam@167
|
836 E T9p, T9M, T9s, T9N;
|
cannam@167
|
837 {
|
cannam@167
|
838 E T9n, T9o, T9q, T9r;
|
cannam@167
|
839 T9n = T7R - T7Q;
|
cannam@167
|
840 T9o = T85 - T84;
|
cannam@167
|
841 T9p = FNMS(KP707106781, T9o, T9n);
|
cannam@167
|
842 T9M = FMA(KP707106781, T9o, T9n);
|
cannam@167
|
843 T9q = T81 - T82;
|
cannam@167
|
844 T9r = T7Y - T7V;
|
cannam@167
|
845 T9s = FNMS(KP707106781, T9r, T9q);
|
cannam@167
|
846 T9N = FMA(KP707106781, T9r, T9q);
|
cannam@167
|
847 }
|
cannam@167
|
848 T9t = FNMS(KP668178637, T9s, T9p);
|
cannam@167
|
849 T9Y = FNMS(KP198912367, T9M, T9N);
|
cannam@167
|
850 T9E = FMA(KP668178637, T9p, T9s);
|
cannam@167
|
851 T9O = FMA(KP198912367, T9N, T9M);
|
cannam@167
|
852 }
|
cannam@167
|
853 {
|
cannam@167
|
854 E T80, T8O, T87, T8P;
|
cannam@167
|
855 {
|
cannam@167
|
856 E T7S, T7Z, T83, T86;
|
cannam@167
|
857 T7S = T7Q + T7R;
|
cannam@167
|
858 T7Z = T7V + T7Y;
|
cannam@167
|
859 T80 = FMA(KP707106781, T7Z, T7S);
|
cannam@167
|
860 T8O = FNMS(KP707106781, T7Z, T7S);
|
cannam@167
|
861 T83 = T81 + T82;
|
cannam@167
|
862 T86 = T84 + T85;
|
cannam@167
|
863 T87 = FMA(KP707106781, T86, T83);
|
cannam@167
|
864 T8P = FNMS(KP707106781, T86, T83);
|
cannam@167
|
865 }
|
cannam@167
|
866 T88 = FMA(KP198912367, T87, T80);
|
cannam@167
|
867 T90 = FMA(KP668178637, T8O, T8P);
|
cannam@167
|
868 T8G = FNMS(KP198912367, T80, T87);
|
cannam@167
|
869 T8Q = FNMS(KP668178637, T8P, T8O);
|
cannam@167
|
870 }
|
cannam@167
|
871 {
|
cannam@167
|
872 E T9a, T9z, T9d, T9y;
|
cannam@167
|
873 {
|
cannam@167
|
874 E T98, T99, T9b, T9c;
|
cannam@167
|
875 T98 = T7K - T7L;
|
cannam@167
|
876 T99 = T7H + T7I;
|
cannam@167
|
877 T9a = FMA(KP414213562, T99, T98);
|
cannam@167
|
878 T9z = FNMS(KP414213562, T98, T99);
|
cannam@167
|
879 T9b = T7D - T7E;
|
cannam@167
|
880 T9c = T7A + T7B;
|
cannam@167
|
881 T9d = FNMS(KP414213562, T9c, T9b);
|
cannam@167
|
882 T9y = FMA(KP414213562, T9b, T9c);
|
cannam@167
|
883 }
|
cannam@167
|
884 T9e = T9a - T9d;
|
cannam@167
|
885 T9U = T9d + T9a;
|
cannam@167
|
886 T9A = T9y - T9z;
|
cannam@167
|
887 T9K = T9y + T9z;
|
cannam@167
|
888 }
|
cannam@167
|
889 {
|
cannam@167
|
890 E T9i, T9P, T9l, T9Q;
|
cannam@167
|
891 {
|
cannam@167
|
892 E T9g, T9h, T9j, T9k;
|
cannam@167
|
893 T9g = T8a + T89;
|
cannam@167
|
894 T9h = T8n - T8o;
|
cannam@167
|
895 T9i = FNMS(KP707106781, T9h, T9g);
|
cannam@167
|
896 T9P = FMA(KP707106781, T9h, T9g);
|
cannam@167
|
897 T9j = T8l - T8k;
|
cannam@167
|
898 T9k = T8h - T8e;
|
cannam@167
|
899 T9l = FNMS(KP707106781, T9k, T9j);
|
cannam@167
|
900 T9Q = FMA(KP707106781, T9k, T9j);
|
cannam@167
|
901 }
|
cannam@167
|
902 T9m = FNMS(KP668178637, T9l, T9i);
|
cannam@167
|
903 T9Z = FNMS(KP198912367, T9P, T9Q);
|
cannam@167
|
904 T9F = FMA(KP668178637, T9i, T9l);
|
cannam@167
|
905 T9R = FMA(KP198912367, T9Q, T9P);
|
cannam@167
|
906 }
|
cannam@167
|
907 {
|
cannam@167
|
908 E T8j, T8R, T8q, T8S;
|
cannam@167
|
909 {
|
cannam@167
|
910 E T8b, T8i, T8m, T8p;
|
cannam@167
|
911 T8b = T89 - T8a;
|
cannam@167
|
912 T8i = T8e + T8h;
|
cannam@167
|
913 T8j = FMA(KP707106781, T8i, T8b);
|
cannam@167
|
914 T8R = FNMS(KP707106781, T8i, T8b);
|
cannam@167
|
915 T8m = T8k + T8l;
|
cannam@167
|
916 T8p = T8n + T8o;
|
cannam@167
|
917 T8q = FMA(KP707106781, T8p, T8m);
|
cannam@167
|
918 T8S = FNMS(KP707106781, T8p, T8m);
|
cannam@167
|
919 }
|
cannam@167
|
920 T8r = FNMS(KP198912367, T8q, T8j);
|
cannam@167
|
921 T91 = FNMS(KP668178637, T8R, T8S);
|
cannam@167
|
922 T8H = FMA(KP198912367, T8j, T8q);
|
cannam@167
|
923 T8T = FMA(KP668178637, T8S, T8R);
|
cannam@167
|
924 }
|
cannam@167
|
925 {
|
cannam@167
|
926 E T7P, T8s, T8J, T8K;
|
cannam@167
|
927 T7P = FMA(KP923879532, T7O, T7z);
|
cannam@167
|
928 T8s = T88 + T8r;
|
cannam@167
|
929 Ip[WS(rs, 1)] = KP500000000 * (FMA(KP980785280, T8s, T7P));
|
cannam@167
|
930 Im[WS(rs, 14)] = -(KP500000000 * (FNMS(KP980785280, T8s, T7P)));
|
cannam@167
|
931 T8J = FMA(KP923879532, T8C, T8z);
|
cannam@167
|
932 T8K = T8G + T8H;
|
cannam@167
|
933 Rm[WS(rs, 14)] = KP500000000 * (FNMS(KP980785280, T8K, T8J));
|
cannam@167
|
934 Rp[WS(rs, 1)] = KP500000000 * (FMA(KP980785280, T8K, T8J));
|
cannam@167
|
935 }
|
cannam@167
|
936 {
|
cannam@167
|
937 E T8D, T8E, T8F, T8I;
|
cannam@167
|
938 T8D = FNMS(KP923879532, T8C, T8z);
|
cannam@167
|
939 T8E = T8r - T88;
|
cannam@167
|
940 Rm[WS(rs, 6)] = KP500000000 * (FNMS(KP980785280, T8E, T8D));
|
cannam@167
|
941 Rp[WS(rs, 9)] = KP500000000 * (FMA(KP980785280, T8E, T8D));
|
cannam@167
|
942 T8F = FNMS(KP923879532, T7O, T7z);
|
cannam@167
|
943 T8I = T8G - T8H;
|
cannam@167
|
944 Ip[WS(rs, 9)] = KP500000000 * (FMA(KP980785280, T8I, T8F));
|
cannam@167
|
945 Im[WS(rs, 6)] = -(KP500000000 * (FNMS(KP980785280, T8I, T8F)));
|
cannam@167
|
946 }
|
cannam@167
|
947 {
|
cannam@167
|
948 E T8N, T8U, T93, T94;
|
cannam@167
|
949 T8N = FNMS(KP923879532, T8M, T8L);
|
cannam@167
|
950 T8U = T8Q + T8T;
|
cannam@167
|
951 Ip[WS(rs, 13)] = KP500000000 * (FNMS(KP831469612, T8U, T8N));
|
cannam@167
|
952 Im[WS(rs, 2)] = -(KP500000000 * (FMA(KP831469612, T8U, T8N)));
|
cannam@167
|
953 T93 = FNMS(KP923879532, T8W, T8V);
|
cannam@167
|
954 T94 = T90 + T91;
|
cannam@167
|
955 Rp[WS(rs, 13)] = KP500000000 * (FNMS(KP831469612, T94, T93));
|
cannam@167
|
956 Rm[WS(rs, 2)] = KP500000000 * (FMA(KP831469612, T94, T93));
|
cannam@167
|
957 }
|
cannam@167
|
958 {
|
cannam@167
|
959 E T8X, T8Y, T8Z, T92;
|
cannam@167
|
960 T8X = FMA(KP923879532, T8W, T8V);
|
cannam@167
|
961 T8Y = T8T - T8Q;
|
cannam@167
|
962 Rm[WS(rs, 10)] = KP500000000 * (FNMS(KP831469612, T8Y, T8X));
|
cannam@167
|
963 Rp[WS(rs, 5)] = KP500000000 * (FMA(KP831469612, T8Y, T8X));
|
cannam@167
|
964 T8Z = FMA(KP923879532, T8M, T8L);
|
cannam@167
|
965 T92 = T90 - T91;
|
cannam@167
|
966 Ip[WS(rs, 5)] = KP500000000 * (FMA(KP831469612, T92, T8Z));
|
cannam@167
|
967 Im[WS(rs, 10)] = -(KP500000000 * (FNMS(KP831469612, T92, T8Z)));
|
cannam@167
|
968 }
|
cannam@167
|
969 {
|
cannam@167
|
970 E T9f, T9u, T9H, T9I;
|
cannam@167
|
971 T9f = FMA(KP923879532, T9e, T97);
|
cannam@167
|
972 T9u = T9m - T9t;
|
cannam@167
|
973 Ip[WS(rs, 3)] = KP500000000 * (FMA(KP831469612, T9u, T9f));
|
cannam@167
|
974 Im[WS(rs, 12)] = -(KP500000000 * (FNMS(KP831469612, T9u, T9f)));
|
cannam@167
|
975 T9H = FMA(KP923879532, T9A, T9x);
|
cannam@167
|
976 T9I = T9E + T9F;
|
cannam@167
|
977 Rm[WS(rs, 12)] = KP500000000 * (FNMS(KP831469612, T9I, T9H));
|
cannam@167
|
978 Rp[WS(rs, 3)] = KP500000000 * (FMA(KP831469612, T9I, T9H));
|
cannam@167
|
979 }
|
cannam@167
|
980 {
|
cannam@167
|
981 E T9B, T9C, T9D, T9G;
|
cannam@167
|
982 T9B = FNMS(KP923879532, T9A, T9x);
|
cannam@167
|
983 T9C = T9t + T9m;
|
cannam@167
|
984 Rm[WS(rs, 4)] = KP500000000 * (FNMS(KP831469612, T9C, T9B));
|
cannam@167
|
985 Rp[WS(rs, 11)] = KP500000000 * (FMA(KP831469612, T9C, T9B));
|
cannam@167
|
986 T9D = FNMS(KP923879532, T9e, T97);
|
cannam@167
|
987 T9G = T9E - T9F;
|
cannam@167
|
988 Ip[WS(rs, 11)] = KP500000000 * (FMA(KP831469612, T9G, T9D));
|
cannam@167
|
989 Im[WS(rs, 4)] = -(KP500000000 * (FNMS(KP831469612, T9G, T9D)));
|
cannam@167
|
990 }
|
cannam@167
|
991 {
|
cannam@167
|
992 E T9L, T9S, Ta1, Ta2;
|
cannam@167
|
993 T9L = FMA(KP923879532, T9K, T9J);
|
cannam@167
|
994 T9S = T9O - T9R;
|
cannam@167
|
995 Ip[WS(rs, 15)] = KP500000000 * (FMA(KP980785280, T9S, T9L));
|
cannam@167
|
996 Im[0] = -(KP500000000 * (FNMS(KP980785280, T9S, T9L)));
|
cannam@167
|
997 Ta1 = FMA(KP923879532, T9U, T9T);
|
cannam@167
|
998 Ta2 = T9Y + T9Z;
|
cannam@167
|
999 Rp[WS(rs, 15)] = KP500000000 * (FNMS(KP980785280, Ta2, Ta1));
|
cannam@167
|
1000 Rm[0] = KP500000000 * (FMA(KP980785280, Ta2, Ta1));
|
cannam@167
|
1001 }
|
cannam@167
|
1002 {
|
cannam@167
|
1003 E T9V, T9W, T9X, Ta0;
|
cannam@167
|
1004 T9V = FNMS(KP923879532, T9U, T9T);
|
cannam@167
|
1005 T9W = T9O + T9R;
|
cannam@167
|
1006 Rm[WS(rs, 8)] = KP500000000 * (FNMS(KP980785280, T9W, T9V));
|
cannam@167
|
1007 Rp[WS(rs, 7)] = KP500000000 * (FMA(KP980785280, T9W, T9V));
|
cannam@167
|
1008 T9X = FNMS(KP923879532, T9K, T9J);
|
cannam@167
|
1009 Ta0 = T9Y - T9Z;
|
cannam@167
|
1010 Ip[WS(rs, 7)] = KP500000000 * (FMA(KP980785280, Ta0, T9X));
|
cannam@167
|
1011 Im[WS(rs, 8)] = -(KP500000000 * (FNMS(KP980785280, Ta0, T9X)));
|
cannam@167
|
1012 }
|
cannam@167
|
1013 }
|
cannam@167
|
1014 }
|
cannam@167
|
1015 }
|
cannam@167
|
1016 }
|
cannam@167
|
1017
|
cannam@167
|
1018 static const tw_instr twinstr[] = {
|
cannam@167
|
1019 {TW_FULL, 1, 32},
|
cannam@167
|
1020 {TW_NEXT, 1, 0}
|
cannam@167
|
1021 };
|
cannam@167
|
1022
|
cannam@167
|
1023 static const hc2c_desc desc = { 32, "hc2cfdft_32", twinstr, &GENUS, {300, 126, 198, 0} };
|
cannam@167
|
1024
|
cannam@167
|
1025 void X(codelet_hc2cfdft_32) (planner *p) {
|
cannam@167
|
1026 X(khc2c_register) (p, hc2cfdft_32, &desc, HC2C_VIA_DFT);
|
cannam@167
|
1027 }
|
cannam@167
|
1028 #else
|
cannam@167
|
1029
|
cannam@167
|
1030 /* Generated by: ../../../genfft/gen_hc2cdft.native -compact -variables 4 -pipeline-latency 4 -n 32 -dit -name hc2cfdft_32 -include rdft/scalar/hc2cf.h */
|
cannam@167
|
1031
|
cannam@167
|
1032 /*
|
cannam@167
|
1033 * This function contains 498 FP additions, 228 FP multiplications,
|
cannam@167
|
1034 * (or, 404 additions, 134 multiplications, 94 fused multiply/add),
|
cannam@167
|
1035 * 106 stack variables, 9 constants, and 128 memory accesses
|
cannam@167
|
1036 */
|
cannam@167
|
1037 #include "rdft/scalar/hc2cf.h"
|
cannam@167
|
1038
|
cannam@167
|
1039 static void hc2cfdft_32(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms)
|
cannam@167
|
1040 {
|
cannam@167
|
1041 DK(KP277785116, +0.277785116509801112371415406974266437187468595);
|
cannam@167
|
1042 DK(KP415734806, +0.415734806151272618539394188808952878369280406);
|
cannam@167
|
1043 DK(KP097545161, +0.097545161008064133924142434238511120463845809);
|
cannam@167
|
1044 DK(KP490392640, +0.490392640201615224563091118067119518486966865);
|
cannam@167
|
1045 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
cannam@167
|
1046 DK(KP191341716, +0.191341716182544885864229992015199433380672281);
|
cannam@167
|
1047 DK(KP461939766, +0.461939766255643378064091594698394143411208313);
|
cannam@167
|
1048 DK(KP353553390, +0.353553390593273762200422181052424519642417969);
|
cannam@167
|
1049 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
|
cannam@167
|
1050 {
|
cannam@167
|
1051 INT m;
|
cannam@167
|
1052 for (m = mb, W = W + ((mb - 1) * 62); m < me; m = m + 1, Rp = Rp + ms, Ip = Ip + ms, Rm = Rm - ms, Im = Im - ms, W = W + 62, MAKE_VOLATILE_STRIDE(128, rs)) {
|
cannam@167
|
1053 E T2S, T5K, T52, T5N, T7p, T8r, T7i, T8o, T2q, T7t, T45, T6L, T2d, T7u, T48;
|
cannam@167
|
1054 E T6M, T1A, T4c, T4f, T1T, T3f, T5M, T7e, T7l, T6J, T7x, T4V, T5J, T7b, T7k;
|
cannam@167
|
1055 E T6G, T7w, Tj, TC, T5r, T4k, T4n, T5s, T3D, T5C, T6V, T72, T4G, T5F, T6u;
|
cannam@167
|
1056 E T86, T6S, T71, T6r, T85, TW, T1f, T5v, T4r, T4u, T5u, T40, T5G, T76, T8k;
|
cannam@167
|
1057 E T4N, T5D, T6B, T89, T6Z, T8h, T6y, T88;
|
cannam@167
|
1058 {
|
cannam@167
|
1059 E T1Y, T22, T2L, T4W, T2p, T43, T2A, T50, T27, T2b, T2Q, T4X, T2h, T2l, T2F;
|
cannam@167
|
1060 E T4Z;
|
cannam@167
|
1061 {
|
cannam@167
|
1062 E T1W, T1X, T2K, T20, T21, T2I, T2H, T2J;
|
cannam@167
|
1063 T1W = Ip[WS(rs, 4)];
|
cannam@167
|
1064 T1X = Im[WS(rs, 4)];
|
cannam@167
|
1065 T2K = T1W + T1X;
|
cannam@167
|
1066 T20 = Rp[WS(rs, 4)];
|
cannam@167
|
1067 T21 = Rm[WS(rs, 4)];
|
cannam@167
|
1068 T2I = T20 - T21;
|
cannam@167
|
1069 T1Y = T1W - T1X;
|
cannam@167
|
1070 T22 = T20 + T21;
|
cannam@167
|
1071 T2H = W[16];
|
cannam@167
|
1072 T2J = W[17];
|
cannam@167
|
1073 T2L = FMA(T2H, T2I, T2J * T2K);
|
cannam@167
|
1074 T4W = FNMS(T2J, T2I, T2H * T2K);
|
cannam@167
|
1075 }
|
cannam@167
|
1076 {
|
cannam@167
|
1077 E T2n, T2o, T2z, T2v, T2w, T2x, T2u, T2y;
|
cannam@167
|
1078 T2n = Ip[0];
|
cannam@167
|
1079 T2o = Im[0];
|
cannam@167
|
1080 T2z = T2n + T2o;
|
cannam@167
|
1081 T2v = Rm[0];
|
cannam@167
|
1082 T2w = Rp[0];
|
cannam@167
|
1083 T2x = T2v - T2w;
|
cannam@167
|
1084 T2p = T2n - T2o;
|
cannam@167
|
1085 T43 = T2w + T2v;
|
cannam@167
|
1086 T2u = W[0];
|
cannam@167
|
1087 T2y = W[1];
|
cannam@167
|
1088 T2A = FNMS(T2y, T2z, T2u * T2x);
|
cannam@167
|
1089 T50 = FMA(T2y, T2x, T2u * T2z);
|
cannam@167
|
1090 }
|
cannam@167
|
1091 {
|
cannam@167
|
1092 E T25, T26, T2P, T29, T2a, T2N, T2M, T2O;
|
cannam@167
|
1093 T25 = Ip[WS(rs, 12)];
|
cannam@167
|
1094 T26 = Im[WS(rs, 12)];
|
cannam@167
|
1095 T2P = T25 + T26;
|
cannam@167
|
1096 T29 = Rp[WS(rs, 12)];
|
cannam@167
|
1097 T2a = Rm[WS(rs, 12)];
|
cannam@167
|
1098 T2N = T29 - T2a;
|
cannam@167
|
1099 T27 = T25 - T26;
|
cannam@167
|
1100 T2b = T29 + T2a;
|
cannam@167
|
1101 T2M = W[48];
|
cannam@167
|
1102 T2O = W[49];
|
cannam@167
|
1103 T2Q = FMA(T2M, T2N, T2O * T2P);
|
cannam@167
|
1104 T4X = FNMS(T2O, T2N, T2M * T2P);
|
cannam@167
|
1105 }
|
cannam@167
|
1106 {
|
cannam@167
|
1107 E T2f, T2g, T2E, T2j, T2k, T2C, T2B, T2D;
|
cannam@167
|
1108 T2f = Ip[WS(rs, 8)];
|
cannam@167
|
1109 T2g = Im[WS(rs, 8)];
|
cannam@167
|
1110 T2E = T2f + T2g;
|
cannam@167
|
1111 T2j = Rp[WS(rs, 8)];
|
cannam@167
|
1112 T2k = Rm[WS(rs, 8)];
|
cannam@167
|
1113 T2C = T2j - T2k;
|
cannam@167
|
1114 T2h = T2f - T2g;
|
cannam@167
|
1115 T2l = T2j + T2k;
|
cannam@167
|
1116 T2B = W[32];
|
cannam@167
|
1117 T2D = W[33];
|
cannam@167
|
1118 T2F = FMA(T2B, T2C, T2D * T2E);
|
cannam@167
|
1119 T4Z = FNMS(T2D, T2C, T2B * T2E);
|
cannam@167
|
1120 }
|
cannam@167
|
1121 {
|
cannam@167
|
1122 E T2G, T2R, T7g, T7h;
|
cannam@167
|
1123 T2G = T2A - T2F;
|
cannam@167
|
1124 T2R = T2L + T2Q;
|
cannam@167
|
1125 T2S = T2G - T2R;
|
cannam@167
|
1126 T5K = T2R + T2G;
|
cannam@167
|
1127 {
|
cannam@167
|
1128 E T4Y, T51, T7n, T7o;
|
cannam@167
|
1129 T4Y = T4W + T4X;
|
cannam@167
|
1130 T51 = T4Z + T50;
|
cannam@167
|
1131 T52 = T4Y + T51;
|
cannam@167
|
1132 T5N = T51 - T4Y;
|
cannam@167
|
1133 T7n = T2Q - T2L;
|
cannam@167
|
1134 T7o = T50 - T4Z;
|
cannam@167
|
1135 T7p = T7n + T7o;
|
cannam@167
|
1136 T8r = T7o - T7n;
|
cannam@167
|
1137 }
|
cannam@167
|
1138 T7g = T2F + T2A;
|
cannam@167
|
1139 T7h = T4W - T4X;
|
cannam@167
|
1140 T7i = T7g - T7h;
|
cannam@167
|
1141 T8o = T7h + T7g;
|
cannam@167
|
1142 {
|
cannam@167
|
1143 E T2m, T44, T2e, T2i;
|
cannam@167
|
1144 T2e = W[30];
|
cannam@167
|
1145 T2i = W[31];
|
cannam@167
|
1146 T2m = FNMS(T2i, T2l, T2e * T2h);
|
cannam@167
|
1147 T44 = FMA(T2e, T2l, T2i * T2h);
|
cannam@167
|
1148 T2q = T2m + T2p;
|
cannam@167
|
1149 T7t = T43 - T44;
|
cannam@167
|
1150 T45 = T43 + T44;
|
cannam@167
|
1151 T6L = T2p - T2m;
|
cannam@167
|
1152 }
|
cannam@167
|
1153 {
|
cannam@167
|
1154 E T23, T46, T2c, T47;
|
cannam@167
|
1155 {
|
cannam@167
|
1156 E T1V, T1Z, T24, T28;
|
cannam@167
|
1157 T1V = W[14];
|
cannam@167
|
1158 T1Z = W[15];
|
cannam@167
|
1159 T23 = FNMS(T1Z, T22, T1V * T1Y);
|
cannam@167
|
1160 T46 = FMA(T1V, T22, T1Z * T1Y);
|
cannam@167
|
1161 T24 = W[46];
|
cannam@167
|
1162 T28 = W[47];
|
cannam@167
|
1163 T2c = FNMS(T28, T2b, T24 * T27);
|
cannam@167
|
1164 T47 = FMA(T24, T2b, T28 * T27);
|
cannam@167
|
1165 }
|
cannam@167
|
1166 T2d = T23 + T2c;
|
cannam@167
|
1167 T7u = T23 - T2c;
|
cannam@167
|
1168 T48 = T46 + T47;
|
cannam@167
|
1169 T6M = T46 - T47;
|
cannam@167
|
1170 }
|
cannam@167
|
1171 }
|
cannam@167
|
1172 }
|
cannam@167
|
1173 {
|
cannam@167
|
1174 E T1q, T4a, T2X, T4P, T1S, T4e, T3d, T4T, T1z, T4b, T32, T4Q, T1J, T4d, T38;
|
cannam@167
|
1175 E T4S;
|
cannam@167
|
1176 {
|
cannam@167
|
1177 E T1l, T2W, T1p, T2U;
|
cannam@167
|
1178 {
|
cannam@167
|
1179 E T1j, T1k, T1n, T1o;
|
cannam@167
|
1180 T1j = Ip[WS(rs, 2)];
|
cannam@167
|
1181 T1k = Im[WS(rs, 2)];
|
cannam@167
|
1182 T1l = T1j - T1k;
|
cannam@167
|
1183 T2W = T1j + T1k;
|
cannam@167
|
1184 T1n = Rp[WS(rs, 2)];
|
cannam@167
|
1185 T1o = Rm[WS(rs, 2)];
|
cannam@167
|
1186 T1p = T1n + T1o;
|
cannam@167
|
1187 T2U = T1n - T1o;
|
cannam@167
|
1188 }
|
cannam@167
|
1189 {
|
cannam@167
|
1190 E T1i, T1m, T2T, T2V;
|
cannam@167
|
1191 T1i = W[6];
|
cannam@167
|
1192 T1m = W[7];
|
cannam@167
|
1193 T1q = FNMS(T1m, T1p, T1i * T1l);
|
cannam@167
|
1194 T4a = FMA(T1i, T1p, T1m * T1l);
|
cannam@167
|
1195 T2T = W[8];
|
cannam@167
|
1196 T2V = W[9];
|
cannam@167
|
1197 T2X = FMA(T2T, T2U, T2V * T2W);
|
cannam@167
|
1198 T4P = FNMS(T2V, T2U, T2T * T2W);
|
cannam@167
|
1199 }
|
cannam@167
|
1200 }
|
cannam@167
|
1201 {
|
cannam@167
|
1202 E T1N, T3c, T1R, T3a;
|
cannam@167
|
1203 {
|
cannam@167
|
1204 E T1L, T1M, T1P, T1Q;
|
cannam@167
|
1205 T1L = Ip[WS(rs, 6)];
|
cannam@167
|
1206 T1M = Im[WS(rs, 6)];
|
cannam@167
|
1207 T1N = T1L - T1M;
|
cannam@167
|
1208 T3c = T1L + T1M;
|
cannam@167
|
1209 T1P = Rp[WS(rs, 6)];
|
cannam@167
|
1210 T1Q = Rm[WS(rs, 6)];
|
cannam@167
|
1211 T1R = T1P + T1Q;
|
cannam@167
|
1212 T3a = T1P - T1Q;
|
cannam@167
|
1213 }
|
cannam@167
|
1214 {
|
cannam@167
|
1215 E T1K, T1O, T39, T3b;
|
cannam@167
|
1216 T1K = W[22];
|
cannam@167
|
1217 T1O = W[23];
|
cannam@167
|
1218 T1S = FNMS(T1O, T1R, T1K * T1N);
|
cannam@167
|
1219 T4e = FMA(T1K, T1R, T1O * T1N);
|
cannam@167
|
1220 T39 = W[24];
|
cannam@167
|
1221 T3b = W[25];
|
cannam@167
|
1222 T3d = FMA(T39, T3a, T3b * T3c);
|
cannam@167
|
1223 T4T = FNMS(T3b, T3a, T39 * T3c);
|
cannam@167
|
1224 }
|
cannam@167
|
1225 }
|
cannam@167
|
1226 {
|
cannam@167
|
1227 E T1u, T31, T1y, T2Z;
|
cannam@167
|
1228 {
|
cannam@167
|
1229 E T1s, T1t, T1w, T1x;
|
cannam@167
|
1230 T1s = Ip[WS(rs, 10)];
|
cannam@167
|
1231 T1t = Im[WS(rs, 10)];
|
cannam@167
|
1232 T1u = T1s - T1t;
|
cannam@167
|
1233 T31 = T1s + T1t;
|
cannam@167
|
1234 T1w = Rp[WS(rs, 10)];
|
cannam@167
|
1235 T1x = Rm[WS(rs, 10)];
|
cannam@167
|
1236 T1y = T1w + T1x;
|
cannam@167
|
1237 T2Z = T1w - T1x;
|
cannam@167
|
1238 }
|
cannam@167
|
1239 {
|
cannam@167
|
1240 E T1r, T1v, T2Y, T30;
|
cannam@167
|
1241 T1r = W[38];
|
cannam@167
|
1242 T1v = W[39];
|
cannam@167
|
1243 T1z = FNMS(T1v, T1y, T1r * T1u);
|
cannam@167
|
1244 T4b = FMA(T1r, T1y, T1v * T1u);
|
cannam@167
|
1245 T2Y = W[40];
|
cannam@167
|
1246 T30 = W[41];
|
cannam@167
|
1247 T32 = FMA(T2Y, T2Z, T30 * T31);
|
cannam@167
|
1248 T4Q = FNMS(T30, T2Z, T2Y * T31);
|
cannam@167
|
1249 }
|
cannam@167
|
1250 }
|
cannam@167
|
1251 {
|
cannam@167
|
1252 E T1E, T37, T1I, T35;
|
cannam@167
|
1253 {
|
cannam@167
|
1254 E T1C, T1D, T1G, T1H;
|
cannam@167
|
1255 T1C = Ip[WS(rs, 14)];
|
cannam@167
|
1256 T1D = Im[WS(rs, 14)];
|
cannam@167
|
1257 T1E = T1C - T1D;
|
cannam@167
|
1258 T37 = T1C + T1D;
|
cannam@167
|
1259 T1G = Rp[WS(rs, 14)];
|
cannam@167
|
1260 T1H = Rm[WS(rs, 14)];
|
cannam@167
|
1261 T1I = T1G + T1H;
|
cannam@167
|
1262 T35 = T1G - T1H;
|
cannam@167
|
1263 }
|
cannam@167
|
1264 {
|
cannam@167
|
1265 E T1B, T1F, T34, T36;
|
cannam@167
|
1266 T1B = W[54];
|
cannam@167
|
1267 T1F = W[55];
|
cannam@167
|
1268 T1J = FNMS(T1F, T1I, T1B * T1E);
|
cannam@167
|
1269 T4d = FMA(T1B, T1I, T1F * T1E);
|
cannam@167
|
1270 T34 = W[56];
|
cannam@167
|
1271 T36 = W[57];
|
cannam@167
|
1272 T38 = FMA(T34, T35, T36 * T37);
|
cannam@167
|
1273 T4S = FNMS(T36, T35, T34 * T37);
|
cannam@167
|
1274 }
|
cannam@167
|
1275 }
|
cannam@167
|
1276 {
|
cannam@167
|
1277 E T33, T3e, T4R, T4U;
|
cannam@167
|
1278 T1A = T1q + T1z;
|
cannam@167
|
1279 T4c = T4a + T4b;
|
cannam@167
|
1280 T4f = T4d + T4e;
|
cannam@167
|
1281 T1T = T1J + T1S;
|
cannam@167
|
1282 T33 = T2X + T32;
|
cannam@167
|
1283 T3e = T38 + T3d;
|
cannam@167
|
1284 T3f = T33 + T3e;
|
cannam@167
|
1285 T5M = T3e - T33;
|
cannam@167
|
1286 {
|
cannam@167
|
1287 E T7c, T7d, T6H, T6I;
|
cannam@167
|
1288 T7c = T4S - T4T;
|
cannam@167
|
1289 T7d = T3d - T38;
|
cannam@167
|
1290 T7e = T7c + T7d;
|
cannam@167
|
1291 T7l = T7c - T7d;
|
cannam@167
|
1292 T6H = T4d - T4e;
|
cannam@167
|
1293 T6I = T1J - T1S;
|
cannam@167
|
1294 T6J = T6H + T6I;
|
cannam@167
|
1295 T7x = T6H - T6I;
|
cannam@167
|
1296 }
|
cannam@167
|
1297 T4R = T4P + T4Q;
|
cannam@167
|
1298 T4U = T4S + T4T;
|
cannam@167
|
1299 T4V = T4R + T4U;
|
cannam@167
|
1300 T5J = T4U - T4R;
|
cannam@167
|
1301 {
|
cannam@167
|
1302 E T79, T7a, T6E, T6F;
|
cannam@167
|
1303 T79 = T32 - T2X;
|
cannam@167
|
1304 T7a = T4P - T4Q;
|
cannam@167
|
1305 T7b = T79 - T7a;
|
cannam@167
|
1306 T7k = T7a + T79;
|
cannam@167
|
1307 T6E = T1q - T1z;
|
cannam@167
|
1308 T6F = T4a - T4b;
|
cannam@167
|
1309 T6G = T6E - T6F;
|
cannam@167
|
1310 T7w = T6F + T6E;
|
cannam@167
|
1311 }
|
cannam@167
|
1312 }
|
cannam@167
|
1313 }
|
cannam@167
|
1314 {
|
cannam@167
|
1315 E T9, T4i, T3l, T4A, TB, T4m, T3B, T4E, Ti, T4j, T3q, T4B, Ts, T4l, T3w;
|
cannam@167
|
1316 E T4D;
|
cannam@167
|
1317 {
|
cannam@167
|
1318 E T4, T3k, T8, T3i;
|
cannam@167
|
1319 {
|
cannam@167
|
1320 E T2, T3, T6, T7;
|
cannam@167
|
1321 T2 = Ip[WS(rs, 1)];
|
cannam@167
|
1322 T3 = Im[WS(rs, 1)];
|
cannam@167
|
1323 T4 = T2 - T3;
|
cannam@167
|
1324 T3k = T2 + T3;
|
cannam@167
|
1325 T6 = Rp[WS(rs, 1)];
|
cannam@167
|
1326 T7 = Rm[WS(rs, 1)];
|
cannam@167
|
1327 T8 = T6 + T7;
|
cannam@167
|
1328 T3i = T6 - T7;
|
cannam@167
|
1329 }
|
cannam@167
|
1330 {
|
cannam@167
|
1331 E T1, T5, T3h, T3j;
|
cannam@167
|
1332 T1 = W[2];
|
cannam@167
|
1333 T5 = W[3];
|
cannam@167
|
1334 T9 = FNMS(T5, T8, T1 * T4);
|
cannam@167
|
1335 T4i = FMA(T1, T8, T5 * T4);
|
cannam@167
|
1336 T3h = W[4];
|
cannam@167
|
1337 T3j = W[5];
|
cannam@167
|
1338 T3l = FMA(T3h, T3i, T3j * T3k);
|
cannam@167
|
1339 T4A = FNMS(T3j, T3i, T3h * T3k);
|
cannam@167
|
1340 }
|
cannam@167
|
1341 }
|
cannam@167
|
1342 {
|
cannam@167
|
1343 E Tw, T3A, TA, T3y;
|
cannam@167
|
1344 {
|
cannam@167
|
1345 E Tu, Tv, Ty, Tz;
|
cannam@167
|
1346 Tu = Ip[WS(rs, 13)];
|
cannam@167
|
1347 Tv = Im[WS(rs, 13)];
|
cannam@167
|
1348 Tw = Tu - Tv;
|
cannam@167
|
1349 T3A = Tu + Tv;
|
cannam@167
|
1350 Ty = Rp[WS(rs, 13)];
|
cannam@167
|
1351 Tz = Rm[WS(rs, 13)];
|
cannam@167
|
1352 TA = Ty + Tz;
|
cannam@167
|
1353 T3y = Ty - Tz;
|
cannam@167
|
1354 }
|
cannam@167
|
1355 {
|
cannam@167
|
1356 E Tt, Tx, T3x, T3z;
|
cannam@167
|
1357 Tt = W[50];
|
cannam@167
|
1358 Tx = W[51];
|
cannam@167
|
1359 TB = FNMS(Tx, TA, Tt * Tw);
|
cannam@167
|
1360 T4m = FMA(Tt, TA, Tx * Tw);
|
cannam@167
|
1361 T3x = W[52];
|
cannam@167
|
1362 T3z = W[53];
|
cannam@167
|
1363 T3B = FMA(T3x, T3y, T3z * T3A);
|
cannam@167
|
1364 T4E = FNMS(T3z, T3y, T3x * T3A);
|
cannam@167
|
1365 }
|
cannam@167
|
1366 }
|
cannam@167
|
1367 {
|
cannam@167
|
1368 E Td, T3p, Th, T3n;
|
cannam@167
|
1369 {
|
cannam@167
|
1370 E Tb, Tc, Tf, Tg;
|
cannam@167
|
1371 Tb = Ip[WS(rs, 9)];
|
cannam@167
|
1372 Tc = Im[WS(rs, 9)];
|
cannam@167
|
1373 Td = Tb - Tc;
|
cannam@167
|
1374 T3p = Tb + Tc;
|
cannam@167
|
1375 Tf = Rp[WS(rs, 9)];
|
cannam@167
|
1376 Tg = Rm[WS(rs, 9)];
|
cannam@167
|
1377 Th = Tf + Tg;
|
cannam@167
|
1378 T3n = Tf - Tg;
|
cannam@167
|
1379 }
|
cannam@167
|
1380 {
|
cannam@167
|
1381 E Ta, Te, T3m, T3o;
|
cannam@167
|
1382 Ta = W[34];
|
cannam@167
|
1383 Te = W[35];
|
cannam@167
|
1384 Ti = FNMS(Te, Th, Ta * Td);
|
cannam@167
|
1385 T4j = FMA(Ta, Th, Te * Td);
|
cannam@167
|
1386 T3m = W[36];
|
cannam@167
|
1387 T3o = W[37];
|
cannam@167
|
1388 T3q = FMA(T3m, T3n, T3o * T3p);
|
cannam@167
|
1389 T4B = FNMS(T3o, T3n, T3m * T3p);
|
cannam@167
|
1390 }
|
cannam@167
|
1391 }
|
cannam@167
|
1392 {
|
cannam@167
|
1393 E Tn, T3v, Tr, T3t;
|
cannam@167
|
1394 {
|
cannam@167
|
1395 E Tl, Tm, Tp, Tq;
|
cannam@167
|
1396 Tl = Ip[WS(rs, 5)];
|
cannam@167
|
1397 Tm = Im[WS(rs, 5)];
|
cannam@167
|
1398 Tn = Tl - Tm;
|
cannam@167
|
1399 T3v = Tl + Tm;
|
cannam@167
|
1400 Tp = Rp[WS(rs, 5)];
|
cannam@167
|
1401 Tq = Rm[WS(rs, 5)];
|
cannam@167
|
1402 Tr = Tp + Tq;
|
cannam@167
|
1403 T3t = Tp - Tq;
|
cannam@167
|
1404 }
|
cannam@167
|
1405 {
|
cannam@167
|
1406 E Tk, To, T3s, T3u;
|
cannam@167
|
1407 Tk = W[18];
|
cannam@167
|
1408 To = W[19];
|
cannam@167
|
1409 Ts = FNMS(To, Tr, Tk * Tn);
|
cannam@167
|
1410 T4l = FMA(Tk, Tr, To * Tn);
|
cannam@167
|
1411 T3s = W[20];
|
cannam@167
|
1412 T3u = W[21];
|
cannam@167
|
1413 T3w = FMA(T3s, T3t, T3u * T3v);
|
cannam@167
|
1414 T4D = FNMS(T3u, T3t, T3s * T3v);
|
cannam@167
|
1415 }
|
cannam@167
|
1416 }
|
cannam@167
|
1417 Tj = T9 + Ti;
|
cannam@167
|
1418 TC = Ts + TB;
|
cannam@167
|
1419 T5r = Tj - TC;
|
cannam@167
|
1420 T4k = T4i + T4j;
|
cannam@167
|
1421 T4n = T4l + T4m;
|
cannam@167
|
1422 T5s = T4k - T4n;
|
cannam@167
|
1423 {
|
cannam@167
|
1424 E T3r, T3C, T6T, T6U;
|
cannam@167
|
1425 T3r = T3l + T3q;
|
cannam@167
|
1426 T3C = T3w + T3B;
|
cannam@167
|
1427 T3D = T3r + T3C;
|
cannam@167
|
1428 T5C = T3C - T3r;
|
cannam@167
|
1429 T6T = T4E - T4D;
|
cannam@167
|
1430 T6U = T3w - T3B;
|
cannam@167
|
1431 T6V = T6T + T6U;
|
cannam@167
|
1432 T72 = T6T - T6U;
|
cannam@167
|
1433 }
|
cannam@167
|
1434 {
|
cannam@167
|
1435 E T4C, T4F, T6s, T6t;
|
cannam@167
|
1436 T4C = T4A + T4B;
|
cannam@167
|
1437 T4F = T4D + T4E;
|
cannam@167
|
1438 T4G = T4C + T4F;
|
cannam@167
|
1439 T5F = T4F - T4C;
|
cannam@167
|
1440 T6s = T4i - T4j;
|
cannam@167
|
1441 T6t = Ts - TB;
|
cannam@167
|
1442 T6u = T6s + T6t;
|
cannam@167
|
1443 T86 = T6s - T6t;
|
cannam@167
|
1444 }
|
cannam@167
|
1445 {
|
cannam@167
|
1446 E T6Q, T6R, T6p, T6q;
|
cannam@167
|
1447 T6Q = T3q - T3l;
|
cannam@167
|
1448 T6R = T4A - T4B;
|
cannam@167
|
1449 T6S = T6Q - T6R;
|
cannam@167
|
1450 T71 = T6R + T6Q;
|
cannam@167
|
1451 T6p = T9 - Ti;
|
cannam@167
|
1452 T6q = T4l - T4m;
|
cannam@167
|
1453 T6r = T6p - T6q;
|
cannam@167
|
1454 T85 = T6p + T6q;
|
cannam@167
|
1455 }
|
cannam@167
|
1456 }
|
cannam@167
|
1457 {
|
cannam@167
|
1458 E TM, T4p, T3I, T4H, T1e, T4t, T3Y, T4L, TV, T4q, T3N, T4I, T15, T4s, T3T;
|
cannam@167
|
1459 E T4K;
|
cannam@167
|
1460 {
|
cannam@167
|
1461 E TH, T3H, TL, T3F;
|
cannam@167
|
1462 {
|
cannam@167
|
1463 E TF, TG, TJ, TK;
|
cannam@167
|
1464 TF = Ip[WS(rs, 15)];
|
cannam@167
|
1465 TG = Im[WS(rs, 15)];
|
cannam@167
|
1466 TH = TF - TG;
|
cannam@167
|
1467 T3H = TF + TG;
|
cannam@167
|
1468 TJ = Rp[WS(rs, 15)];
|
cannam@167
|
1469 TK = Rm[WS(rs, 15)];
|
cannam@167
|
1470 TL = TJ + TK;
|
cannam@167
|
1471 T3F = TJ - TK;
|
cannam@167
|
1472 }
|
cannam@167
|
1473 {
|
cannam@167
|
1474 E TE, TI, T3E, T3G;
|
cannam@167
|
1475 TE = W[58];
|
cannam@167
|
1476 TI = W[59];
|
cannam@167
|
1477 TM = FNMS(TI, TL, TE * TH);
|
cannam@167
|
1478 T4p = FMA(TE, TL, TI * TH);
|
cannam@167
|
1479 T3E = W[60];
|
cannam@167
|
1480 T3G = W[61];
|
cannam@167
|
1481 T3I = FMA(T3E, T3F, T3G * T3H);
|
cannam@167
|
1482 T4H = FNMS(T3G, T3F, T3E * T3H);
|
cannam@167
|
1483 }
|
cannam@167
|
1484 }
|
cannam@167
|
1485 {
|
cannam@167
|
1486 E T19, T3X, T1d, T3V;
|
cannam@167
|
1487 {
|
cannam@167
|
1488 E T17, T18, T1b, T1c;
|
cannam@167
|
1489 T17 = Ip[WS(rs, 11)];
|
cannam@167
|
1490 T18 = Im[WS(rs, 11)];
|
cannam@167
|
1491 T19 = T17 - T18;
|
cannam@167
|
1492 T3X = T17 + T18;
|
cannam@167
|
1493 T1b = Rp[WS(rs, 11)];
|
cannam@167
|
1494 T1c = Rm[WS(rs, 11)];
|
cannam@167
|
1495 T1d = T1b + T1c;
|
cannam@167
|
1496 T3V = T1b - T1c;
|
cannam@167
|
1497 }
|
cannam@167
|
1498 {
|
cannam@167
|
1499 E T16, T1a, T3U, T3W;
|
cannam@167
|
1500 T16 = W[42];
|
cannam@167
|
1501 T1a = W[43];
|
cannam@167
|
1502 T1e = FNMS(T1a, T1d, T16 * T19);
|
cannam@167
|
1503 T4t = FMA(T16, T1d, T1a * T19);
|
cannam@167
|
1504 T3U = W[44];
|
cannam@167
|
1505 T3W = W[45];
|
cannam@167
|
1506 T3Y = FMA(T3U, T3V, T3W * T3X);
|
cannam@167
|
1507 T4L = FNMS(T3W, T3V, T3U * T3X);
|
cannam@167
|
1508 }
|
cannam@167
|
1509 }
|
cannam@167
|
1510 {
|
cannam@167
|
1511 E TQ, T3M, TU, T3K;
|
cannam@167
|
1512 {
|
cannam@167
|
1513 E TO, TP, TS, TT;
|
cannam@167
|
1514 TO = Ip[WS(rs, 7)];
|
cannam@167
|
1515 TP = Im[WS(rs, 7)];
|
cannam@167
|
1516 TQ = TO - TP;
|
cannam@167
|
1517 T3M = TO + TP;
|
cannam@167
|
1518 TS = Rp[WS(rs, 7)];
|
cannam@167
|
1519 TT = Rm[WS(rs, 7)];
|
cannam@167
|
1520 TU = TS + TT;
|
cannam@167
|
1521 T3K = TS - TT;
|
cannam@167
|
1522 }
|
cannam@167
|
1523 {
|
cannam@167
|
1524 E TN, TR, T3J, T3L;
|
cannam@167
|
1525 TN = W[26];
|
cannam@167
|
1526 TR = W[27];
|
cannam@167
|
1527 TV = FNMS(TR, TU, TN * TQ);
|
cannam@167
|
1528 T4q = FMA(TN, TU, TR * TQ);
|
cannam@167
|
1529 T3J = W[28];
|
cannam@167
|
1530 T3L = W[29];
|
cannam@167
|
1531 T3N = FMA(T3J, T3K, T3L * T3M);
|
cannam@167
|
1532 T4I = FNMS(T3L, T3K, T3J * T3M);
|
cannam@167
|
1533 }
|
cannam@167
|
1534 }
|
cannam@167
|
1535 {
|
cannam@167
|
1536 E T10, T3S, T14, T3Q;
|
cannam@167
|
1537 {
|
cannam@167
|
1538 E TY, TZ, T12, T13;
|
cannam@167
|
1539 TY = Ip[WS(rs, 3)];
|
cannam@167
|
1540 TZ = Im[WS(rs, 3)];
|
cannam@167
|
1541 T10 = TY - TZ;
|
cannam@167
|
1542 T3S = TY + TZ;
|
cannam@167
|
1543 T12 = Rp[WS(rs, 3)];
|
cannam@167
|
1544 T13 = Rm[WS(rs, 3)];
|
cannam@167
|
1545 T14 = T12 + T13;
|
cannam@167
|
1546 T3Q = T12 - T13;
|
cannam@167
|
1547 }
|
cannam@167
|
1548 {
|
cannam@167
|
1549 E TX, T11, T3P, T3R;
|
cannam@167
|
1550 TX = W[10];
|
cannam@167
|
1551 T11 = W[11];
|
cannam@167
|
1552 T15 = FNMS(T11, T14, TX * T10);
|
cannam@167
|
1553 T4s = FMA(TX, T14, T11 * T10);
|
cannam@167
|
1554 T3P = W[12];
|
cannam@167
|
1555 T3R = W[13];
|
cannam@167
|
1556 T3T = FMA(T3P, T3Q, T3R * T3S);
|
cannam@167
|
1557 T4K = FNMS(T3R, T3Q, T3P * T3S);
|
cannam@167
|
1558 }
|
cannam@167
|
1559 }
|
cannam@167
|
1560 TW = TM + TV;
|
cannam@167
|
1561 T1f = T15 + T1e;
|
cannam@167
|
1562 T5v = TW - T1f;
|
cannam@167
|
1563 T4r = T4p + T4q;
|
cannam@167
|
1564 T4u = T4s + T4t;
|
cannam@167
|
1565 T5u = T4r - T4u;
|
cannam@167
|
1566 {
|
cannam@167
|
1567 E T3O, T3Z, T74, T75;
|
cannam@167
|
1568 T3O = T3I + T3N;
|
cannam@167
|
1569 T3Z = T3T + T3Y;
|
cannam@167
|
1570 T40 = T3O + T3Z;
|
cannam@167
|
1571 T5G = T3Z - T3O;
|
cannam@167
|
1572 T74 = T4H - T4I;
|
cannam@167
|
1573 T75 = T3Y - T3T;
|
cannam@167
|
1574 T76 = T74 + T75;
|
cannam@167
|
1575 T8k = T74 - T75;
|
cannam@167
|
1576 }
|
cannam@167
|
1577 {
|
cannam@167
|
1578 E T4J, T4M, T6z, T6A;
|
cannam@167
|
1579 T4J = T4H + T4I;
|
cannam@167
|
1580 T4M = T4K + T4L;
|
cannam@167
|
1581 T4N = T4J + T4M;
|
cannam@167
|
1582 T5D = T4J - T4M;
|
cannam@167
|
1583 T6z = T4p - T4q;
|
cannam@167
|
1584 T6A = T15 - T1e;
|
cannam@167
|
1585 T6B = T6z + T6A;
|
cannam@167
|
1586 T89 = T6z - T6A;
|
cannam@167
|
1587 }
|
cannam@167
|
1588 {
|
cannam@167
|
1589 E T6X, T6Y, T6w, T6x;
|
cannam@167
|
1590 T6X = T3N - T3I;
|
cannam@167
|
1591 T6Y = T4K - T4L;
|
cannam@167
|
1592 T6Z = T6X - T6Y;
|
cannam@167
|
1593 T8h = T6X + T6Y;
|
cannam@167
|
1594 T6w = TM - TV;
|
cannam@167
|
1595 T6x = T4s - T4t;
|
cannam@167
|
1596 T6y = T6w - T6x;
|
cannam@167
|
1597 T88 = T6w + T6x;
|
cannam@167
|
1598 }
|
cannam@167
|
1599 }
|
cannam@167
|
1600 {
|
cannam@167
|
1601 E T1h, T5i, T5c, T5m, T5f, T5n, T2s, T58, T42, T4y, T4w, T57, T54, T56, T4h;
|
cannam@167
|
1602 E T5h;
|
cannam@167
|
1603 {
|
cannam@167
|
1604 E TD, T1g, T5a, T5b;
|
cannam@167
|
1605 TD = Tj + TC;
|
cannam@167
|
1606 T1g = TW + T1f;
|
cannam@167
|
1607 T1h = TD + T1g;
|
cannam@167
|
1608 T5i = TD - T1g;
|
cannam@167
|
1609 T5a = T4N - T4G;
|
cannam@167
|
1610 T5b = T3D - T40;
|
cannam@167
|
1611 T5c = T5a + T5b;
|
cannam@167
|
1612 T5m = T5a - T5b;
|
cannam@167
|
1613 }
|
cannam@167
|
1614 {
|
cannam@167
|
1615 E T5d, T5e, T1U, T2r;
|
cannam@167
|
1616 T5d = T3f + T2S;
|
cannam@167
|
1617 T5e = T52 - T4V;
|
cannam@167
|
1618 T5f = T5d - T5e;
|
cannam@167
|
1619 T5n = T5d + T5e;
|
cannam@167
|
1620 T1U = T1A + T1T;
|
cannam@167
|
1621 T2r = T2d + T2q;
|
cannam@167
|
1622 T2s = T1U + T2r;
|
cannam@167
|
1623 T58 = T2r - T1U;
|
cannam@167
|
1624 }
|
cannam@167
|
1625 {
|
cannam@167
|
1626 E T3g, T41, T4o, T4v;
|
cannam@167
|
1627 T3g = T2S - T3f;
|
cannam@167
|
1628 T41 = T3D + T40;
|
cannam@167
|
1629 T42 = T3g - T41;
|
cannam@167
|
1630 T4y = T41 + T3g;
|
cannam@167
|
1631 T4o = T4k + T4n;
|
cannam@167
|
1632 T4v = T4r + T4u;
|
cannam@167
|
1633 T4w = T4o + T4v;
|
cannam@167
|
1634 T57 = T4v - T4o;
|
cannam@167
|
1635 }
|
cannam@167
|
1636 {
|
cannam@167
|
1637 E T4O, T53, T49, T4g;
|
cannam@167
|
1638 T4O = T4G + T4N;
|
cannam@167
|
1639 T53 = T4V + T52;
|
cannam@167
|
1640 T54 = T4O - T53;
|
cannam@167
|
1641 T56 = T4O + T53;
|
cannam@167
|
1642 T49 = T45 + T48;
|
cannam@167
|
1643 T4g = T4c + T4f;
|
cannam@167
|
1644 T4h = T49 + T4g;
|
cannam@167
|
1645 T5h = T49 - T4g;
|
cannam@167
|
1646 }
|
cannam@167
|
1647 {
|
cannam@167
|
1648 E T2t, T55, T4x, T4z;
|
cannam@167
|
1649 T2t = T1h + T2s;
|
cannam@167
|
1650 Ip[0] = KP500000000 * (T2t + T42);
|
cannam@167
|
1651 Im[WS(rs, 15)] = KP500000000 * (T42 - T2t);
|
cannam@167
|
1652 T55 = T4h + T4w;
|
cannam@167
|
1653 Rm[WS(rs, 15)] = KP500000000 * (T55 - T56);
|
cannam@167
|
1654 Rp[0] = KP500000000 * (T55 + T56);
|
cannam@167
|
1655 T4x = T4h - T4w;
|
cannam@167
|
1656 Rm[WS(rs, 7)] = KP500000000 * (T4x - T4y);
|
cannam@167
|
1657 Rp[WS(rs, 8)] = KP500000000 * (T4x + T4y);
|
cannam@167
|
1658 T4z = T2s - T1h;
|
cannam@167
|
1659 Ip[WS(rs, 8)] = KP500000000 * (T4z + T54);
|
cannam@167
|
1660 Im[WS(rs, 7)] = KP500000000 * (T54 - T4z);
|
cannam@167
|
1661 }
|
cannam@167
|
1662 {
|
cannam@167
|
1663 E T59, T5g, T5p, T5q;
|
cannam@167
|
1664 T59 = KP500000000 * (T57 + T58);
|
cannam@167
|
1665 T5g = KP353553390 * (T5c + T5f);
|
cannam@167
|
1666 Ip[WS(rs, 4)] = T59 + T5g;
|
cannam@167
|
1667 Im[WS(rs, 11)] = T5g - T59;
|
cannam@167
|
1668 T5p = KP500000000 * (T5h + T5i);
|
cannam@167
|
1669 T5q = KP353553390 * (T5m + T5n);
|
cannam@167
|
1670 Rm[WS(rs, 11)] = T5p - T5q;
|
cannam@167
|
1671 Rp[WS(rs, 4)] = T5p + T5q;
|
cannam@167
|
1672 }
|
cannam@167
|
1673 {
|
cannam@167
|
1674 E T5j, T5k, T5l, T5o;
|
cannam@167
|
1675 T5j = KP500000000 * (T5h - T5i);
|
cannam@167
|
1676 T5k = KP353553390 * (T5f - T5c);
|
cannam@167
|
1677 Rm[WS(rs, 3)] = T5j - T5k;
|
cannam@167
|
1678 Rp[WS(rs, 12)] = T5j + T5k;
|
cannam@167
|
1679 T5l = KP500000000 * (T58 - T57);
|
cannam@167
|
1680 T5o = KP353553390 * (T5m - T5n);
|
cannam@167
|
1681 Ip[WS(rs, 12)] = T5l + T5o;
|
cannam@167
|
1682 Im[WS(rs, 3)] = T5o - T5l;
|
cannam@167
|
1683 }
|
cannam@167
|
1684 }
|
cannam@167
|
1685 {
|
cannam@167
|
1686 E T5x, T6g, T6a, T6k, T6d, T6l, T5A, T66, T5I, T60, T5T, T6f, T5W, T65, T5P;
|
cannam@167
|
1687 E T61;
|
cannam@167
|
1688 {
|
cannam@167
|
1689 E T5t, T5w, T68, T69;
|
cannam@167
|
1690 T5t = T5r - T5s;
|
cannam@167
|
1691 T5w = T5u + T5v;
|
cannam@167
|
1692 T5x = KP353553390 * (T5t + T5w);
|
cannam@167
|
1693 T6g = KP353553390 * (T5t - T5w);
|
cannam@167
|
1694 T68 = T5D - T5C;
|
cannam@167
|
1695 T69 = T5G - T5F;
|
cannam@167
|
1696 T6a = FMA(KP461939766, T68, KP191341716 * T69);
|
cannam@167
|
1697 T6k = FNMS(KP461939766, T69, KP191341716 * T68);
|
cannam@167
|
1698 }
|
cannam@167
|
1699 {
|
cannam@167
|
1700 E T6b, T6c, T5y, T5z;
|
cannam@167
|
1701 T6b = T5K - T5J;
|
cannam@167
|
1702 T6c = T5N - T5M;
|
cannam@167
|
1703 T6d = FNMS(KP461939766, T6c, KP191341716 * T6b);
|
cannam@167
|
1704 T6l = FMA(KP461939766, T6b, KP191341716 * T6c);
|
cannam@167
|
1705 T5y = T4f - T4c;
|
cannam@167
|
1706 T5z = T2q - T2d;
|
cannam@167
|
1707 T5A = KP500000000 * (T5y + T5z);
|
cannam@167
|
1708 T66 = KP500000000 * (T5z - T5y);
|
cannam@167
|
1709 }
|
cannam@167
|
1710 {
|
cannam@167
|
1711 E T5E, T5H, T5R, T5S;
|
cannam@167
|
1712 T5E = T5C + T5D;
|
cannam@167
|
1713 T5H = T5F + T5G;
|
cannam@167
|
1714 T5I = FMA(KP191341716, T5E, KP461939766 * T5H);
|
cannam@167
|
1715 T60 = FNMS(KP191341716, T5H, KP461939766 * T5E);
|
cannam@167
|
1716 T5R = T45 - T48;
|
cannam@167
|
1717 T5S = T1A - T1T;
|
cannam@167
|
1718 T5T = KP500000000 * (T5R + T5S);
|
cannam@167
|
1719 T6f = KP500000000 * (T5R - T5S);
|
cannam@167
|
1720 }
|
cannam@167
|
1721 {
|
cannam@167
|
1722 E T5U, T5V, T5L, T5O;
|
cannam@167
|
1723 T5U = T5s + T5r;
|
cannam@167
|
1724 T5V = T5u - T5v;
|
cannam@167
|
1725 T5W = KP353553390 * (T5U + T5V);
|
cannam@167
|
1726 T65 = KP353553390 * (T5V - T5U);
|
cannam@167
|
1727 T5L = T5J + T5K;
|
cannam@167
|
1728 T5O = T5M + T5N;
|
cannam@167
|
1729 T5P = FNMS(KP191341716, T5O, KP461939766 * T5L);
|
cannam@167
|
1730 T61 = FMA(KP191341716, T5L, KP461939766 * T5O);
|
cannam@167
|
1731 }
|
cannam@167
|
1732 {
|
cannam@167
|
1733 E T5B, T5Q, T63, T64;
|
cannam@167
|
1734 T5B = T5x + T5A;
|
cannam@167
|
1735 T5Q = T5I + T5P;
|
cannam@167
|
1736 Ip[WS(rs, 2)] = T5B + T5Q;
|
cannam@167
|
1737 Im[WS(rs, 13)] = T5Q - T5B;
|
cannam@167
|
1738 T63 = T5T + T5W;
|
cannam@167
|
1739 T64 = T60 + T61;
|
cannam@167
|
1740 Rm[WS(rs, 13)] = T63 - T64;
|
cannam@167
|
1741 Rp[WS(rs, 2)] = T63 + T64;
|
cannam@167
|
1742 }
|
cannam@167
|
1743 {
|
cannam@167
|
1744 E T5X, T5Y, T5Z, T62;
|
cannam@167
|
1745 T5X = T5T - T5W;
|
cannam@167
|
1746 T5Y = T5P - T5I;
|
cannam@167
|
1747 Rm[WS(rs, 5)] = T5X - T5Y;
|
cannam@167
|
1748 Rp[WS(rs, 10)] = T5X + T5Y;
|
cannam@167
|
1749 T5Z = T5A - T5x;
|
cannam@167
|
1750 T62 = T60 - T61;
|
cannam@167
|
1751 Ip[WS(rs, 10)] = T5Z + T62;
|
cannam@167
|
1752 Im[WS(rs, 5)] = T62 - T5Z;
|
cannam@167
|
1753 }
|
cannam@167
|
1754 {
|
cannam@167
|
1755 E T67, T6e, T6n, T6o;
|
cannam@167
|
1756 T67 = T65 + T66;
|
cannam@167
|
1757 T6e = T6a + T6d;
|
cannam@167
|
1758 Ip[WS(rs, 6)] = T67 + T6e;
|
cannam@167
|
1759 Im[WS(rs, 9)] = T6e - T67;
|
cannam@167
|
1760 T6n = T6f + T6g;
|
cannam@167
|
1761 T6o = T6k + T6l;
|
cannam@167
|
1762 Rm[WS(rs, 9)] = T6n - T6o;
|
cannam@167
|
1763 Rp[WS(rs, 6)] = T6n + T6o;
|
cannam@167
|
1764 }
|
cannam@167
|
1765 {
|
cannam@167
|
1766 E T6h, T6i, T6j, T6m;
|
cannam@167
|
1767 T6h = T6f - T6g;
|
cannam@167
|
1768 T6i = T6d - T6a;
|
cannam@167
|
1769 Rm[WS(rs, 1)] = T6h - T6i;
|
cannam@167
|
1770 Rp[WS(rs, 14)] = T6h + T6i;
|
cannam@167
|
1771 T6j = T66 - T65;
|
cannam@167
|
1772 T6m = T6k - T6l;
|
cannam@167
|
1773 Ip[WS(rs, 14)] = T6j + T6m;
|
cannam@167
|
1774 Im[WS(rs, 1)] = T6m - T6j;
|
cannam@167
|
1775 }
|
cannam@167
|
1776 }
|
cannam@167
|
1777 {
|
cannam@167
|
1778 E T6D, T7W, T6O, T7M, T7C, T7L, T7z, T7V, T7r, T81, T7H, T7T, T78, T80, T7G;
|
cannam@167
|
1779 E T7Q;
|
cannam@167
|
1780 {
|
cannam@167
|
1781 E T6v, T6C, T7v, T7y;
|
cannam@167
|
1782 T6v = FNMS(KP191341716, T6u, KP461939766 * T6r);
|
cannam@167
|
1783 T6C = FMA(KP461939766, T6y, KP191341716 * T6B);
|
cannam@167
|
1784 T6D = T6v + T6C;
|
cannam@167
|
1785 T7W = T6v - T6C;
|
cannam@167
|
1786 {
|
cannam@167
|
1787 E T6K, T6N, T7A, T7B;
|
cannam@167
|
1788 T6K = KP353553390 * (T6G + T6J);
|
cannam@167
|
1789 T6N = KP500000000 * (T6L - T6M);
|
cannam@167
|
1790 T6O = T6K + T6N;
|
cannam@167
|
1791 T7M = T6N - T6K;
|
cannam@167
|
1792 T7A = FMA(KP191341716, T6r, KP461939766 * T6u);
|
cannam@167
|
1793 T7B = FNMS(KP191341716, T6y, KP461939766 * T6B);
|
cannam@167
|
1794 T7C = T7A + T7B;
|
cannam@167
|
1795 T7L = T7B - T7A;
|
cannam@167
|
1796 }
|
cannam@167
|
1797 T7v = KP500000000 * (T7t + T7u);
|
cannam@167
|
1798 T7y = KP353553390 * (T7w + T7x);
|
cannam@167
|
1799 T7z = T7v + T7y;
|
cannam@167
|
1800 T7V = T7v - T7y;
|
cannam@167
|
1801 {
|
cannam@167
|
1802 E T7j, T7R, T7q, T7S, T7f, T7m;
|
cannam@167
|
1803 T7f = KP707106781 * (T7b + T7e);
|
cannam@167
|
1804 T7j = T7f + T7i;
|
cannam@167
|
1805 T7R = T7i - T7f;
|
cannam@167
|
1806 T7m = KP707106781 * (T7k + T7l);
|
cannam@167
|
1807 T7q = T7m + T7p;
|
cannam@167
|
1808 T7S = T7p - T7m;
|
cannam@167
|
1809 T7r = FNMS(KP097545161, T7q, KP490392640 * T7j);
|
cannam@167
|
1810 T81 = FMA(KP415734806, T7R, KP277785116 * T7S);
|
cannam@167
|
1811 T7H = FMA(KP097545161, T7j, KP490392640 * T7q);
|
cannam@167
|
1812 T7T = FNMS(KP415734806, T7S, KP277785116 * T7R);
|
cannam@167
|
1813 }
|
cannam@167
|
1814 {
|
cannam@167
|
1815 E T70, T7O, T77, T7P, T6W, T73;
|
cannam@167
|
1816 T6W = KP707106781 * (T6S + T6V);
|
cannam@167
|
1817 T70 = T6W + T6Z;
|
cannam@167
|
1818 T7O = T6Z - T6W;
|
cannam@167
|
1819 T73 = KP707106781 * (T71 + T72);
|
cannam@167
|
1820 T77 = T73 + T76;
|
cannam@167
|
1821 T7P = T76 - T73;
|
cannam@167
|
1822 T78 = FMA(KP490392640, T70, KP097545161 * T77);
|
cannam@167
|
1823 T80 = FNMS(KP415734806, T7O, KP277785116 * T7P);
|
cannam@167
|
1824 T7G = FNMS(KP097545161, T70, KP490392640 * T77);
|
cannam@167
|
1825 T7Q = FMA(KP277785116, T7O, KP415734806 * T7P);
|
cannam@167
|
1826 }
|
cannam@167
|
1827 }
|
cannam@167
|
1828 {
|
cannam@167
|
1829 E T6P, T7s, T7J, T7K;
|
cannam@167
|
1830 T6P = T6D + T6O;
|
cannam@167
|
1831 T7s = T78 + T7r;
|
cannam@167
|
1832 Ip[WS(rs, 1)] = T6P + T7s;
|
cannam@167
|
1833 Im[WS(rs, 14)] = T7s - T6P;
|
cannam@167
|
1834 T7J = T7z + T7C;
|
cannam@167
|
1835 T7K = T7G + T7H;
|
cannam@167
|
1836 Rm[WS(rs, 14)] = T7J - T7K;
|
cannam@167
|
1837 Rp[WS(rs, 1)] = T7J + T7K;
|
cannam@167
|
1838 }
|
cannam@167
|
1839 {
|
cannam@167
|
1840 E T7D, T7E, T7F, T7I;
|
cannam@167
|
1841 T7D = T7z - T7C;
|
cannam@167
|
1842 T7E = T7r - T78;
|
cannam@167
|
1843 Rm[WS(rs, 6)] = T7D - T7E;
|
cannam@167
|
1844 Rp[WS(rs, 9)] = T7D + T7E;
|
cannam@167
|
1845 T7F = T6O - T6D;
|
cannam@167
|
1846 T7I = T7G - T7H;
|
cannam@167
|
1847 Ip[WS(rs, 9)] = T7F + T7I;
|
cannam@167
|
1848 Im[WS(rs, 6)] = T7I - T7F;
|
cannam@167
|
1849 }
|
cannam@167
|
1850 {
|
cannam@167
|
1851 E T7N, T7U, T83, T84;
|
cannam@167
|
1852 T7N = T7L + T7M;
|
cannam@167
|
1853 T7U = T7Q + T7T;
|
cannam@167
|
1854 Ip[WS(rs, 5)] = T7N + T7U;
|
cannam@167
|
1855 Im[WS(rs, 10)] = T7U - T7N;
|
cannam@167
|
1856 T83 = T7V + T7W;
|
cannam@167
|
1857 T84 = T80 + T81;
|
cannam@167
|
1858 Rm[WS(rs, 10)] = T83 - T84;
|
cannam@167
|
1859 Rp[WS(rs, 5)] = T83 + T84;
|
cannam@167
|
1860 }
|
cannam@167
|
1861 {
|
cannam@167
|
1862 E T7X, T7Y, T7Z, T82;
|
cannam@167
|
1863 T7X = T7V - T7W;
|
cannam@167
|
1864 T7Y = T7T - T7Q;
|
cannam@167
|
1865 Rm[WS(rs, 2)] = T7X - T7Y;
|
cannam@167
|
1866 Rp[WS(rs, 13)] = T7X + T7Y;
|
cannam@167
|
1867 T7Z = T7M - T7L;
|
cannam@167
|
1868 T82 = T80 - T81;
|
cannam@167
|
1869 Ip[WS(rs, 13)] = T7Z + T82;
|
cannam@167
|
1870 Im[WS(rs, 2)] = T82 - T7Z;
|
cannam@167
|
1871 }
|
cannam@167
|
1872 }
|
cannam@167
|
1873 {
|
cannam@167
|
1874 E T8b, T8U, T8e, T8K, T8A, T8J, T8x, T8T, T8t, T8Z, T8F, T8R, T8m, T8Y, T8E;
|
cannam@167
|
1875 E T8O;
|
cannam@167
|
1876 {
|
cannam@167
|
1877 E T87, T8a, T8v, T8w;
|
cannam@167
|
1878 T87 = FNMS(KP461939766, T86, KP191341716 * T85);
|
cannam@167
|
1879 T8a = FMA(KP191341716, T88, KP461939766 * T89);
|
cannam@167
|
1880 T8b = T87 + T8a;
|
cannam@167
|
1881 T8U = T87 - T8a;
|
cannam@167
|
1882 {
|
cannam@167
|
1883 E T8c, T8d, T8y, T8z;
|
cannam@167
|
1884 T8c = KP353553390 * (T7x - T7w);
|
cannam@167
|
1885 T8d = KP500000000 * (T6M + T6L);
|
cannam@167
|
1886 T8e = T8c + T8d;
|
cannam@167
|
1887 T8K = T8d - T8c;
|
cannam@167
|
1888 T8y = FMA(KP461939766, T85, KP191341716 * T86);
|
cannam@167
|
1889 T8z = FNMS(KP461939766, T88, KP191341716 * T89);
|
cannam@167
|
1890 T8A = T8y + T8z;
|
cannam@167
|
1891 T8J = T8z - T8y;
|
cannam@167
|
1892 }
|
cannam@167
|
1893 T8v = KP500000000 * (T7t - T7u);
|
cannam@167
|
1894 T8w = KP353553390 * (T6G - T6J);
|
cannam@167
|
1895 T8x = T8v + T8w;
|
cannam@167
|
1896 T8T = T8v - T8w;
|
cannam@167
|
1897 {
|
cannam@167
|
1898 E T8p, T8P, T8s, T8Q, T8n, T8q;
|
cannam@167
|
1899 T8n = KP707106781 * (T7l - T7k);
|
cannam@167
|
1900 T8p = T8n + T8o;
|
cannam@167
|
1901 T8P = T8o - T8n;
|
cannam@167
|
1902 T8q = KP707106781 * (T7b - T7e);
|
cannam@167
|
1903 T8s = T8q + T8r;
|
cannam@167
|
1904 T8Q = T8r - T8q;
|
cannam@167
|
1905 T8t = FNMS(KP277785116, T8s, KP415734806 * T8p);
|
cannam@167
|
1906 T8Z = FMA(KP490392640, T8P, KP097545161 * T8Q);
|
cannam@167
|
1907 T8F = FMA(KP277785116, T8p, KP415734806 * T8s);
|
cannam@167
|
1908 T8R = FNMS(KP490392640, T8Q, KP097545161 * T8P);
|
cannam@167
|
1909 }
|
cannam@167
|
1910 {
|
cannam@167
|
1911 E T8i, T8M, T8l, T8N, T8g, T8j;
|
cannam@167
|
1912 T8g = KP707106781 * (T72 - T71);
|
cannam@167
|
1913 T8i = T8g + T8h;
|
cannam@167
|
1914 T8M = T8h - T8g;
|
cannam@167
|
1915 T8j = KP707106781 * (T6S - T6V);
|
cannam@167
|
1916 T8l = T8j + T8k;
|
cannam@167
|
1917 T8N = T8k - T8j;
|
cannam@167
|
1918 T8m = FMA(KP415734806, T8i, KP277785116 * T8l);
|
cannam@167
|
1919 T8Y = FNMS(KP490392640, T8M, KP097545161 * T8N);
|
cannam@167
|
1920 T8E = FNMS(KP277785116, T8i, KP415734806 * T8l);
|
cannam@167
|
1921 T8O = FMA(KP097545161, T8M, KP490392640 * T8N);
|
cannam@167
|
1922 }
|
cannam@167
|
1923 }
|
cannam@167
|
1924 {
|
cannam@167
|
1925 E T8f, T8u, T8H, T8I;
|
cannam@167
|
1926 T8f = T8b + T8e;
|
cannam@167
|
1927 T8u = T8m + T8t;
|
cannam@167
|
1928 Ip[WS(rs, 3)] = T8f + T8u;
|
cannam@167
|
1929 Im[WS(rs, 12)] = T8u - T8f;
|
cannam@167
|
1930 T8H = T8x + T8A;
|
cannam@167
|
1931 T8I = T8E + T8F;
|
cannam@167
|
1932 Rm[WS(rs, 12)] = T8H - T8I;
|
cannam@167
|
1933 Rp[WS(rs, 3)] = T8H + T8I;
|
cannam@167
|
1934 }
|
cannam@167
|
1935 {
|
cannam@167
|
1936 E T8B, T8C, T8D, T8G;
|
cannam@167
|
1937 T8B = T8x - T8A;
|
cannam@167
|
1938 T8C = T8t - T8m;
|
cannam@167
|
1939 Rm[WS(rs, 4)] = T8B - T8C;
|
cannam@167
|
1940 Rp[WS(rs, 11)] = T8B + T8C;
|
cannam@167
|
1941 T8D = T8e - T8b;
|
cannam@167
|
1942 T8G = T8E - T8F;
|
cannam@167
|
1943 Ip[WS(rs, 11)] = T8D + T8G;
|
cannam@167
|
1944 Im[WS(rs, 4)] = T8G - T8D;
|
cannam@167
|
1945 }
|
cannam@167
|
1946 {
|
cannam@167
|
1947 E T8L, T8S, T91, T92;
|
cannam@167
|
1948 T8L = T8J + T8K;
|
cannam@167
|
1949 T8S = T8O + T8R;
|
cannam@167
|
1950 Ip[WS(rs, 7)] = T8L + T8S;
|
cannam@167
|
1951 Im[WS(rs, 8)] = T8S - T8L;
|
cannam@167
|
1952 T91 = T8T + T8U;
|
cannam@167
|
1953 T92 = T8Y + T8Z;
|
cannam@167
|
1954 Rm[WS(rs, 8)] = T91 - T92;
|
cannam@167
|
1955 Rp[WS(rs, 7)] = T91 + T92;
|
cannam@167
|
1956 }
|
cannam@167
|
1957 {
|
cannam@167
|
1958 E T8V, T8W, T8X, T90;
|
cannam@167
|
1959 T8V = T8T - T8U;
|
cannam@167
|
1960 T8W = T8R - T8O;
|
cannam@167
|
1961 Rm[0] = T8V - T8W;
|
cannam@167
|
1962 Rp[WS(rs, 15)] = T8V + T8W;
|
cannam@167
|
1963 T8X = T8K - T8J;
|
cannam@167
|
1964 T90 = T8Y - T8Z;
|
cannam@167
|
1965 Ip[WS(rs, 15)] = T8X + T90;
|
cannam@167
|
1966 Im[0] = T90 - T8X;
|
cannam@167
|
1967 }
|
cannam@167
|
1968 }
|
cannam@167
|
1969 }
|
cannam@167
|
1970 }
|
cannam@167
|
1971 }
|
cannam@167
|
1972
|
cannam@167
|
1973 static const tw_instr twinstr[] = {
|
cannam@167
|
1974 {TW_FULL, 1, 32},
|
cannam@167
|
1975 {TW_NEXT, 1, 0}
|
cannam@167
|
1976 };
|
cannam@167
|
1977
|
cannam@167
|
1978 static const hc2c_desc desc = { 32, "hc2cfdft_32", twinstr, &GENUS, {404, 134, 94, 0} };
|
cannam@167
|
1979
|
cannam@167
|
1980 void X(codelet_hc2cfdft_32) (planner *p) {
|
cannam@167
|
1981 X(khc2c_register) (p, hc2cfdft_32, &desc, HC2C_VIA_DFT);
|
cannam@167
|
1982 }
|
cannam@167
|
1983 #endif
|