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 08:07:55 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_hc2c.native -fma -compact -variables 4 -pipeline-latency 4 -sign 1 -twiddle-log3 -precompute-twiddles -n 32 -dif -name hc2cb2_32 -include rdft/scalar/hc2cb.h */
|
cannam@167
|
29
|
cannam@167
|
30 /*
|
cannam@167
|
31 * This function contains 488 FP additions, 350 FP multiplications,
|
cannam@167
|
32 * (or, 236 additions, 98 multiplications, 252 fused multiply/add),
|
cannam@167
|
33 * 164 stack variables, 7 constants, and 128 memory accesses
|
cannam@167
|
34 */
|
cannam@167
|
35 #include "rdft/scalar/hc2cb.h"
|
cannam@167
|
36
|
cannam@167
|
37 static void hc2cb2_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(KP980785280, +0.980785280403230449126182236134239036973933731);
|
cannam@167
|
40 DK(KP198912367, +0.198912367379658006911597622644676228597850501);
|
cannam@167
|
41 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
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 {
|
cannam@167
|
47 INT m;
|
cannam@167
|
48 for (m = mb, W = W + ((mb - 1) * 8); m < me; m = m + 1, Rp = Rp + ms, Ip = Ip + ms, Rm = Rm - ms, Im = Im - ms, W = W + 8, MAKE_VOLATILE_STRIDE(128, rs)) {
|
cannam@167
|
49 E T11, T14, T12, T37, T17, T1b, T39, T3a, T3v, T3d, T3x, T15, T16, T5X, T4p;
|
cannam@167
|
50 E T3G, T78, T7e, T8S, T9s, T8P, T8V, T98, T9m, T7I, T7C, T3y, T4b, T3C, T4g;
|
cannam@167
|
51 E T5u, T6b, T5I, T6e, T1a, T1c, T2O, T4r, T4s, T4W, T3J, T3K, T3Y, T5Z, T60;
|
cannam@167
|
52 E T66, T3i, T3q, T3l, T3e, T7S, T8K, T8m, T8E, T5k, T5U, T5R, T5e, T6i, T7s;
|
cannam@167
|
53 E T6O, T7o;
|
cannam@167
|
54 {
|
cannam@167
|
55 E T77, T9l, T7B, T7d, T9r, T7H, T3b, T5d, T19, T3I;
|
cannam@167
|
56 {
|
cannam@167
|
57 E T13, T3F, T38, T3c;
|
cannam@167
|
58 T11 = W[2];
|
cannam@167
|
59 T14 = W[3];
|
cannam@167
|
60 T12 = W[4];
|
cannam@167
|
61 T37 = W[0];
|
cannam@167
|
62 T13 = T11 * T12;
|
cannam@167
|
63 T3F = T37 * T12;
|
cannam@167
|
64 T38 = T37 * T11;
|
cannam@167
|
65 T3c = T37 * T14;
|
cannam@167
|
66 T17 = W[6];
|
cannam@167
|
67 T77 = T37 * T17;
|
cannam@167
|
68 T9l = T12 * T17;
|
cannam@167
|
69 T7B = T11 * T17;
|
cannam@167
|
70 T1b = W[7];
|
cannam@167
|
71 T7d = T37 * T1b;
|
cannam@167
|
72 T9r = T12 * T1b;
|
cannam@167
|
73 T7H = T11 * T1b;
|
cannam@167
|
74 T39 = W[1];
|
cannam@167
|
75 T3a = FNMS(T39, T14, T38);
|
cannam@167
|
76 T3v = FMA(T39, T14, T38);
|
cannam@167
|
77 T3b = T3a * T12;
|
cannam@167
|
78 T5d = T3v * T12;
|
cannam@167
|
79 T3d = FMA(T39, T11, T3c);
|
cannam@167
|
80 T3x = FNMS(T39, T11, T3c);
|
cannam@167
|
81 T15 = W[5];
|
cannam@167
|
82 T19 = T11 * T15;
|
cannam@167
|
83 T3I = T37 * T15;
|
cannam@167
|
84 T16 = FMA(T14, T15, T13);
|
cannam@167
|
85 T5X = FNMS(T14, T15, T13);
|
cannam@167
|
86 T4p = FMA(T39, T15, T3F);
|
cannam@167
|
87 T3G = FNMS(T39, T15, T3F);
|
cannam@167
|
88 }
|
cannam@167
|
89 T78 = FNMS(T39, T1b, T77);
|
cannam@167
|
90 T7e = FMA(T39, T17, T7d);
|
cannam@167
|
91 T8S = FMA(T14, T17, T7H);
|
cannam@167
|
92 T9s = FNMS(T15, T17, T9r);
|
cannam@167
|
93 T8P = FNMS(T14, T1b, T7B);
|
cannam@167
|
94 T8V = FMA(T39, T1b, T77);
|
cannam@167
|
95 T98 = FNMS(T39, T17, T7d);
|
cannam@167
|
96 T9m = FMA(T15, T1b, T9l);
|
cannam@167
|
97 T7I = FNMS(T14, T17, T7H);
|
cannam@167
|
98 T7C = FMA(T14, T1b, T7B);
|
cannam@167
|
99 {
|
cannam@167
|
100 E T3w, T3B, T5Y, T65;
|
cannam@167
|
101 T3w = T3v * T17;
|
cannam@167
|
102 T3y = FNMS(T3x, T1b, T3w);
|
cannam@167
|
103 T4b = FMA(T3x, T1b, T3w);
|
cannam@167
|
104 T3B = T3v * T1b;
|
cannam@167
|
105 T3C = FMA(T3x, T17, T3B);
|
cannam@167
|
106 T4g = FNMS(T3x, T17, T3B);
|
cannam@167
|
107 {
|
cannam@167
|
108 E T5t, T5H, T18, T2N;
|
cannam@167
|
109 T5t = T3a * T17;
|
cannam@167
|
110 T5u = FMA(T3d, T1b, T5t);
|
cannam@167
|
111 T6b = FNMS(T3d, T1b, T5t);
|
cannam@167
|
112 T5H = T3a * T1b;
|
cannam@167
|
113 T5I = FNMS(T3d, T17, T5H);
|
cannam@167
|
114 T6e = FMA(T3d, T17, T5H);
|
cannam@167
|
115 T18 = T16 * T17;
|
cannam@167
|
116 T2N = T16 * T1b;
|
cannam@167
|
117 T1a = FNMS(T14, T12, T19);
|
cannam@167
|
118 T1c = FMA(T1a, T1b, T18);
|
cannam@167
|
119 T2O = FNMS(T1a, T17, T2N);
|
cannam@167
|
120 }
|
cannam@167
|
121 {
|
cannam@167
|
122 E T4q, T4V, T3H, T3X;
|
cannam@167
|
123 T4q = T4p * T17;
|
cannam@167
|
124 T4V = T4p * T1b;
|
cannam@167
|
125 T4r = FNMS(T39, T12, T3I);
|
cannam@167
|
126 T4s = FMA(T4r, T1b, T4q);
|
cannam@167
|
127 T4W = FNMS(T4r, T17, T4V);
|
cannam@167
|
128 T3H = T3G * T17;
|
cannam@167
|
129 T3X = T3G * T1b;
|
cannam@167
|
130 T3J = FMA(T39, T12, T3I);
|
cannam@167
|
131 T3K = FMA(T3J, T1b, T3H);
|
cannam@167
|
132 T3Y = FNMS(T3J, T17, T3X);
|
cannam@167
|
133 }
|
cannam@167
|
134 T5Y = T5X * T17;
|
cannam@167
|
135 T65 = T5X * T1b;
|
cannam@167
|
136 T5Z = FMA(T14, T12, T19);
|
cannam@167
|
137 T60 = FMA(T5Z, T1b, T5Y);
|
cannam@167
|
138 T66 = FNMS(T5Z, T17, T65);
|
cannam@167
|
139 {
|
cannam@167
|
140 E T8D, T8J, T7R, T8l, T3h;
|
cannam@167
|
141 T3h = T3a * T15;
|
cannam@167
|
142 T3i = FNMS(T3d, T12, T3h);
|
cannam@167
|
143 T3q = FMA(T3d, T12, T3h);
|
cannam@167
|
144 T3l = FNMS(T3d, T15, T3b);
|
cannam@167
|
145 T8D = T3l * T17;
|
cannam@167
|
146 T8J = T3l * T1b;
|
cannam@167
|
147 T3e = FMA(T3d, T15, T3b);
|
cannam@167
|
148 T7R = T3e * T17;
|
cannam@167
|
149 T8l = T3e * T1b;
|
cannam@167
|
150 T7S = FMA(T3i, T1b, T7R);
|
cannam@167
|
151 T8K = FNMS(T3q, T17, T8J);
|
cannam@167
|
152 T8m = FNMS(T3i, T17, T8l);
|
cannam@167
|
153 T8E = FMA(T3q, T1b, T8D);
|
cannam@167
|
154 }
|
cannam@167
|
155 {
|
cannam@167
|
156 E T6h, T6N, T7n, T7r, T5j;
|
cannam@167
|
157 T5j = T3v * T15;
|
cannam@167
|
158 T5k = FMA(T3x, T12, T5j);
|
cannam@167
|
159 T5U = FNMS(T3x, T12, T5j);
|
cannam@167
|
160 T5R = FMA(T3x, T15, T5d);
|
cannam@167
|
161 T6h = T5R * T17;
|
cannam@167
|
162 T6N = T5R * T1b;
|
cannam@167
|
163 T5e = FNMS(T3x, T15, T5d);
|
cannam@167
|
164 T7n = T5e * T17;
|
cannam@167
|
165 T7r = T5e * T1b;
|
cannam@167
|
166 T6i = FMA(T5U, T1b, T6h);
|
cannam@167
|
167 T7s = FNMS(T5k, T17, T7r);
|
cannam@167
|
168 T6O = FNMS(T5U, T17, T6N);
|
cannam@167
|
169 T7o = FMA(T5k, T1b, T7n);
|
cannam@167
|
170 }
|
cannam@167
|
171 }
|
cannam@167
|
172 }
|
cannam@167
|
173 {
|
cannam@167
|
174 E Tf, T6j, T7V, T8W, T8p, T99, T1t, T3L, T2X, T3Z, T4Z, T5J, T6W, T7t, T4v;
|
cannam@167
|
175 E T5v, TZ, T7x, T28, T3S, T91, T9d, T2h, T3R, T4Q, T5B, T8a, T8v, T4N, T5C;
|
cannam@167
|
176 E T6J, T6Z, TK, T7w, T2z, T3P, T94, T9c, T2I, T3O, T4J, T5y, T8h, T8u, T4G;
|
cannam@167
|
177 E T5z, T6A, T6Y, Tu, T6P, T82, T9a, T8s, T8X, T1Q, T40, T30, T3M, T52, T5w;
|
cannam@167
|
178 E T6q, T7u, T4C, T5K;
|
cannam@167
|
179 {
|
cannam@167
|
180 E T3, T1d, T2S, T6Q, T6, T2P, T1g, T6R, Td, T6U, T1r, T2V, Ta, T6T, T1m;
|
cannam@167
|
181 E T2U;
|
cannam@167
|
182 {
|
cannam@167
|
183 E T1, T2, T1e, T1f;
|
cannam@167
|
184 T1 = Rp[0];
|
cannam@167
|
185 T2 = Rm[WS(rs, 15)];
|
cannam@167
|
186 T3 = T1 + T2;
|
cannam@167
|
187 T1d = T1 - T2;
|
cannam@167
|
188 {
|
cannam@167
|
189 E T2Q, T2R, T4, T5;
|
cannam@167
|
190 T2Q = Ip[0];
|
cannam@167
|
191 T2R = Im[WS(rs, 15)];
|
cannam@167
|
192 T2S = T2Q + T2R;
|
cannam@167
|
193 T6Q = T2Q - T2R;
|
cannam@167
|
194 T4 = Rp[WS(rs, 8)];
|
cannam@167
|
195 T5 = Rm[WS(rs, 7)];
|
cannam@167
|
196 T6 = T4 + T5;
|
cannam@167
|
197 T2P = T4 - T5;
|
cannam@167
|
198 }
|
cannam@167
|
199 T1e = Ip[WS(rs, 8)];
|
cannam@167
|
200 T1f = Im[WS(rs, 7)];
|
cannam@167
|
201 T1g = T1e + T1f;
|
cannam@167
|
202 T6R = T1e - T1f;
|
cannam@167
|
203 {
|
cannam@167
|
204 E Tb, Tc, T1n, T1o, T1p, T1q;
|
cannam@167
|
205 Tb = Rm[WS(rs, 3)];
|
cannam@167
|
206 Tc = Rp[WS(rs, 12)];
|
cannam@167
|
207 T1n = Tb - Tc;
|
cannam@167
|
208 T1o = Ip[WS(rs, 12)];
|
cannam@167
|
209 T1p = Im[WS(rs, 3)];
|
cannam@167
|
210 T1q = T1o + T1p;
|
cannam@167
|
211 Td = Tb + Tc;
|
cannam@167
|
212 T6U = T1o - T1p;
|
cannam@167
|
213 T1r = T1n - T1q;
|
cannam@167
|
214 T2V = T1n + T1q;
|
cannam@167
|
215 }
|
cannam@167
|
216 {
|
cannam@167
|
217 E T8, T9, T1i, T1j, T1k, T1l;
|
cannam@167
|
218 T8 = Rp[WS(rs, 4)];
|
cannam@167
|
219 T9 = Rm[WS(rs, 11)];
|
cannam@167
|
220 T1i = T8 - T9;
|
cannam@167
|
221 T1j = Ip[WS(rs, 4)];
|
cannam@167
|
222 T1k = Im[WS(rs, 11)];
|
cannam@167
|
223 T1l = T1j + T1k;
|
cannam@167
|
224 Ta = T8 + T9;
|
cannam@167
|
225 T6T = T1j - T1k;
|
cannam@167
|
226 T1m = T1i - T1l;
|
cannam@167
|
227 T2U = T1i + T1l;
|
cannam@167
|
228 }
|
cannam@167
|
229 }
|
cannam@167
|
230 {
|
cannam@167
|
231 E T7, Te, T7T, T7U;
|
cannam@167
|
232 T7 = T3 + T6;
|
cannam@167
|
233 Te = Ta + Td;
|
cannam@167
|
234 Tf = T7 + Te;
|
cannam@167
|
235 T6j = T7 - Te;
|
cannam@167
|
236 T7T = T3 - T6;
|
cannam@167
|
237 T7U = T6U - T6T;
|
cannam@167
|
238 T7V = T7T - T7U;
|
cannam@167
|
239 T8W = T7T + T7U;
|
cannam@167
|
240 }
|
cannam@167
|
241 {
|
cannam@167
|
242 E T8n, T8o, T1h, T1s;
|
cannam@167
|
243 T8n = T6Q - T6R;
|
cannam@167
|
244 T8o = Ta - Td;
|
cannam@167
|
245 T8p = T8n - T8o;
|
cannam@167
|
246 T99 = T8o + T8n;
|
cannam@167
|
247 T1h = T1d - T1g;
|
cannam@167
|
248 T1s = T1m + T1r;
|
cannam@167
|
249 T1t = FNMS(KP707106781, T1s, T1h);
|
cannam@167
|
250 T3L = FMA(KP707106781, T1s, T1h);
|
cannam@167
|
251 }
|
cannam@167
|
252 {
|
cannam@167
|
253 E T2T, T2W, T4X, T4Y;
|
cannam@167
|
254 T2T = T2P + T2S;
|
cannam@167
|
255 T2W = T2U - T2V;
|
cannam@167
|
256 T2X = FNMS(KP707106781, T2W, T2T);
|
cannam@167
|
257 T3Z = FMA(KP707106781, T2W, T2T);
|
cannam@167
|
258 T4X = T2S - T2P;
|
cannam@167
|
259 T4Y = T1m - T1r;
|
cannam@167
|
260 T4Z = FMA(KP707106781, T4Y, T4X);
|
cannam@167
|
261 T5J = FNMS(KP707106781, T4Y, T4X);
|
cannam@167
|
262 }
|
cannam@167
|
263 {
|
cannam@167
|
264 E T6S, T6V, T4t, T4u;
|
cannam@167
|
265 T6S = T6Q + T6R;
|
cannam@167
|
266 T6V = T6T + T6U;
|
cannam@167
|
267 T6W = T6S - T6V;
|
cannam@167
|
268 T7t = T6S + T6V;
|
cannam@167
|
269 T4t = T1d + T1g;
|
cannam@167
|
270 T4u = T2U + T2V;
|
cannam@167
|
271 T4v = FNMS(KP707106781, T4u, T4t);
|
cannam@167
|
272 T5v = FMA(KP707106781, T4u, T4t);
|
cannam@167
|
273 }
|
cannam@167
|
274 }
|
cannam@167
|
275 {
|
cannam@167
|
276 E TR, T87, T1S, T29, T1V, T84, T2c, T6E, TY, T85, T88, T21, T26, T2f, T6H;
|
cannam@167
|
277 E T2e, T86, T89;
|
cannam@167
|
278 {
|
cannam@167
|
279 E TL, TM, TN, TO, TP, TQ;
|
cannam@167
|
280 TL = Rm[0];
|
cannam@167
|
281 TM = Rp[WS(rs, 15)];
|
cannam@167
|
282 TN = TL + TM;
|
cannam@167
|
283 TO = Rp[WS(rs, 7)];
|
cannam@167
|
284 TP = Rm[WS(rs, 8)];
|
cannam@167
|
285 TQ = TO + TP;
|
cannam@167
|
286 TR = TN + TQ;
|
cannam@167
|
287 T87 = TN - TQ;
|
cannam@167
|
288 T1S = TO - TP;
|
cannam@167
|
289 T29 = TL - TM;
|
cannam@167
|
290 }
|
cannam@167
|
291 {
|
cannam@167
|
292 E T1T, T1U, T6C, T2a, T2b, T6D;
|
cannam@167
|
293 T1T = Ip[WS(rs, 15)];
|
cannam@167
|
294 T1U = Im[0];
|
cannam@167
|
295 T6C = T1T - T1U;
|
cannam@167
|
296 T2a = Ip[WS(rs, 7)];
|
cannam@167
|
297 T2b = Im[WS(rs, 8)];
|
cannam@167
|
298 T6D = T2a - T2b;
|
cannam@167
|
299 T1V = T1T + T1U;
|
cannam@167
|
300 T84 = T6C - T6D;
|
cannam@167
|
301 T2c = T2a + T2b;
|
cannam@167
|
302 T6E = T6C + T6D;
|
cannam@167
|
303 }
|
cannam@167
|
304 {
|
cannam@167
|
305 E TU, T1X, T20, T6F, TX, T22, T25, T6G;
|
cannam@167
|
306 {
|
cannam@167
|
307 E TS, TT, T1Y, T1Z;
|
cannam@167
|
308 TS = Rp[WS(rs, 3)];
|
cannam@167
|
309 TT = Rm[WS(rs, 12)];
|
cannam@167
|
310 TU = TS + TT;
|
cannam@167
|
311 T1X = TS - TT;
|
cannam@167
|
312 T1Y = Ip[WS(rs, 3)];
|
cannam@167
|
313 T1Z = Im[WS(rs, 12)];
|
cannam@167
|
314 T20 = T1Y + T1Z;
|
cannam@167
|
315 T6F = T1Y - T1Z;
|
cannam@167
|
316 }
|
cannam@167
|
317 {
|
cannam@167
|
318 E TV, TW, T23, T24;
|
cannam@167
|
319 TV = Rm[WS(rs, 4)];
|
cannam@167
|
320 TW = Rp[WS(rs, 11)];
|
cannam@167
|
321 TX = TV + TW;
|
cannam@167
|
322 T22 = TV - TW;
|
cannam@167
|
323 T23 = Ip[WS(rs, 11)];
|
cannam@167
|
324 T24 = Im[WS(rs, 4)];
|
cannam@167
|
325 T25 = T23 + T24;
|
cannam@167
|
326 T6G = T23 - T24;
|
cannam@167
|
327 }
|
cannam@167
|
328 TY = TU + TX;
|
cannam@167
|
329 T85 = TU - TX;
|
cannam@167
|
330 T88 = T6G - T6F;
|
cannam@167
|
331 T21 = T1X + T20;
|
cannam@167
|
332 T26 = T22 + T25;
|
cannam@167
|
333 T2f = T22 - T25;
|
cannam@167
|
334 T6H = T6F + T6G;
|
cannam@167
|
335 T2e = T1X - T20;
|
cannam@167
|
336 }
|
cannam@167
|
337 TZ = TR + TY;
|
cannam@167
|
338 T7x = T6E + T6H;
|
cannam@167
|
339 {
|
cannam@167
|
340 E T1W, T27, T8Z, T90;
|
cannam@167
|
341 T1W = T1S - T1V;
|
cannam@167
|
342 T27 = T21 - T26;
|
cannam@167
|
343 T28 = FNMS(KP707106781, T27, T1W);
|
cannam@167
|
344 T3S = FMA(KP707106781, T27, T1W);
|
cannam@167
|
345 T8Z = T85 + T84;
|
cannam@167
|
346 T90 = T87 + T88;
|
cannam@167
|
347 T91 = FNMS(KP414213562, T90, T8Z);
|
cannam@167
|
348 T9d = FMA(KP414213562, T8Z, T90);
|
cannam@167
|
349 }
|
cannam@167
|
350 {
|
cannam@167
|
351 E T2d, T2g, T4O, T4P;
|
cannam@167
|
352 T2d = T29 - T2c;
|
cannam@167
|
353 T2g = T2e + T2f;
|
cannam@167
|
354 T2h = FNMS(KP707106781, T2g, T2d);
|
cannam@167
|
355 T3R = FMA(KP707106781, T2g, T2d);
|
cannam@167
|
356 T4O = T1S + T1V;
|
cannam@167
|
357 T4P = T2e - T2f;
|
cannam@167
|
358 T4Q = FNMS(KP707106781, T4P, T4O);
|
cannam@167
|
359 T5B = FMA(KP707106781, T4P, T4O);
|
cannam@167
|
360 }
|
cannam@167
|
361 T86 = T84 - T85;
|
cannam@167
|
362 T89 = T87 - T88;
|
cannam@167
|
363 T8a = FMA(KP414213562, T89, T86);
|
cannam@167
|
364 T8v = FNMS(KP414213562, T86, T89);
|
cannam@167
|
365 {
|
cannam@167
|
366 E T4L, T4M, T6B, T6I;
|
cannam@167
|
367 T4L = T29 + T2c;
|
cannam@167
|
368 T4M = T21 + T26;
|
cannam@167
|
369 T4N = FNMS(KP707106781, T4M, T4L);
|
cannam@167
|
370 T5C = FMA(KP707106781, T4M, T4L);
|
cannam@167
|
371 T6B = TR - TY;
|
cannam@167
|
372 T6I = T6E - T6H;
|
cannam@167
|
373 T6J = T6B + T6I;
|
cannam@167
|
374 T6Z = T6I - T6B;
|
cannam@167
|
375 }
|
cannam@167
|
376 }
|
cannam@167
|
377 {
|
cannam@167
|
378 E TC, T8e, T2j, T2A, T2m, T8b, T2D, T6v, TJ, T8c, T8f, T2s, T2x, T2G, T6y;
|
cannam@167
|
379 E T2F, T8d, T8g;
|
cannam@167
|
380 {
|
cannam@167
|
381 E Tw, Tx, Ty, Tz, TA, TB;
|
cannam@167
|
382 Tw = Rp[WS(rs, 1)];
|
cannam@167
|
383 Tx = Rm[WS(rs, 14)];
|
cannam@167
|
384 Ty = Tw + Tx;
|
cannam@167
|
385 Tz = Rp[WS(rs, 9)];
|
cannam@167
|
386 TA = Rm[WS(rs, 6)];
|
cannam@167
|
387 TB = Tz + TA;
|
cannam@167
|
388 TC = Ty + TB;
|
cannam@167
|
389 T8e = Ty - TB;
|
cannam@167
|
390 T2j = Tz - TA;
|
cannam@167
|
391 T2A = Tw - Tx;
|
cannam@167
|
392 }
|
cannam@167
|
393 {
|
cannam@167
|
394 E T2k, T2l, T6t, T2B, T2C, T6u;
|
cannam@167
|
395 T2k = Ip[WS(rs, 1)];
|
cannam@167
|
396 T2l = Im[WS(rs, 14)];
|
cannam@167
|
397 T6t = T2k - T2l;
|
cannam@167
|
398 T2B = Ip[WS(rs, 9)];
|
cannam@167
|
399 T2C = Im[WS(rs, 6)];
|
cannam@167
|
400 T6u = T2B - T2C;
|
cannam@167
|
401 T2m = T2k + T2l;
|
cannam@167
|
402 T8b = T6t - T6u;
|
cannam@167
|
403 T2D = T2B + T2C;
|
cannam@167
|
404 T6v = T6t + T6u;
|
cannam@167
|
405 }
|
cannam@167
|
406 {
|
cannam@167
|
407 E TF, T2o, T2r, T6w, TI, T2t, T2w, T6x;
|
cannam@167
|
408 {
|
cannam@167
|
409 E TD, TE, T2p, T2q;
|
cannam@167
|
410 TD = Rp[WS(rs, 5)];
|
cannam@167
|
411 TE = Rm[WS(rs, 10)];
|
cannam@167
|
412 TF = TD + TE;
|
cannam@167
|
413 T2o = TD - TE;
|
cannam@167
|
414 T2p = Ip[WS(rs, 5)];
|
cannam@167
|
415 T2q = Im[WS(rs, 10)];
|
cannam@167
|
416 T2r = T2p + T2q;
|
cannam@167
|
417 T6w = T2p - T2q;
|
cannam@167
|
418 }
|
cannam@167
|
419 {
|
cannam@167
|
420 E TG, TH, T2u, T2v;
|
cannam@167
|
421 TG = Rm[WS(rs, 2)];
|
cannam@167
|
422 TH = Rp[WS(rs, 13)];
|
cannam@167
|
423 TI = TG + TH;
|
cannam@167
|
424 T2t = TG - TH;
|
cannam@167
|
425 T2u = Ip[WS(rs, 13)];
|
cannam@167
|
426 T2v = Im[WS(rs, 2)];
|
cannam@167
|
427 T2w = T2u + T2v;
|
cannam@167
|
428 T6x = T2u - T2v;
|
cannam@167
|
429 }
|
cannam@167
|
430 TJ = TF + TI;
|
cannam@167
|
431 T8c = TF - TI;
|
cannam@167
|
432 T8f = T6x - T6w;
|
cannam@167
|
433 T2s = T2o + T2r;
|
cannam@167
|
434 T2x = T2t + T2w;
|
cannam@167
|
435 T2G = T2t - T2w;
|
cannam@167
|
436 T6y = T6w + T6x;
|
cannam@167
|
437 T2F = T2o - T2r;
|
cannam@167
|
438 }
|
cannam@167
|
439 TK = TC + TJ;
|
cannam@167
|
440 T7w = T6v + T6y;
|
cannam@167
|
441 {
|
cannam@167
|
442 E T2n, T2y, T92, T93;
|
cannam@167
|
443 T2n = T2j + T2m;
|
cannam@167
|
444 T2y = T2s - T2x;
|
cannam@167
|
445 T2z = FNMS(KP707106781, T2y, T2n);
|
cannam@167
|
446 T3P = FMA(KP707106781, T2y, T2n);
|
cannam@167
|
447 T92 = T8c + T8b;
|
cannam@167
|
448 T93 = T8e + T8f;
|
cannam@167
|
449 T94 = FMA(KP414213562, T93, T92);
|
cannam@167
|
450 T9c = FNMS(KP414213562, T92, T93);
|
cannam@167
|
451 }
|
cannam@167
|
452 {
|
cannam@167
|
453 E T2E, T2H, T4H, T4I;
|
cannam@167
|
454 T2E = T2A - T2D;
|
cannam@167
|
455 T2H = T2F + T2G;
|
cannam@167
|
456 T2I = FNMS(KP707106781, T2H, T2E);
|
cannam@167
|
457 T3O = FMA(KP707106781, T2H, T2E);
|
cannam@167
|
458 T4H = T2m - T2j;
|
cannam@167
|
459 T4I = T2G - T2F;
|
cannam@167
|
460 T4J = FNMS(KP707106781, T4I, T4H);
|
cannam@167
|
461 T5y = FMA(KP707106781, T4I, T4H);
|
cannam@167
|
462 }
|
cannam@167
|
463 T8d = T8b - T8c;
|
cannam@167
|
464 T8g = T8e - T8f;
|
cannam@167
|
465 T8h = FNMS(KP414213562, T8g, T8d);
|
cannam@167
|
466 T8u = FMA(KP414213562, T8d, T8g);
|
cannam@167
|
467 {
|
cannam@167
|
468 E T4E, T4F, T6s, T6z;
|
cannam@167
|
469 T4E = T2A + T2D;
|
cannam@167
|
470 T4F = T2s + T2x;
|
cannam@167
|
471 T4G = FNMS(KP707106781, T4F, T4E);
|
cannam@167
|
472 T5z = FMA(KP707106781, T4F, T4E);
|
cannam@167
|
473 T6s = TC - TJ;
|
cannam@167
|
474 T6z = T6v - T6y;
|
cannam@167
|
475 T6A = T6s - T6z;
|
cannam@167
|
476 T6Y = T6s + T6z;
|
cannam@167
|
477 }
|
cannam@167
|
478 }
|
cannam@167
|
479 {
|
cannam@167
|
480 E Ti, T6n, Tl, T6o, T1J, T1O, T80, T7Z, T4x, T4w, Tp, T6k, Ts, T6l, T1y;
|
cannam@167
|
481 E T1D, T7X, T7W, T4A, T4z;
|
cannam@167
|
482 {
|
cannam@167
|
483 E T1K, T1I, T1F, T1N;
|
cannam@167
|
484 {
|
cannam@167
|
485 E Tg, Th, T1G, T1H;
|
cannam@167
|
486 Tg = Rp[WS(rs, 2)];
|
cannam@167
|
487 Th = Rm[WS(rs, 13)];
|
cannam@167
|
488 Ti = Tg + Th;
|
cannam@167
|
489 T1K = Tg - Th;
|
cannam@167
|
490 T1G = Ip[WS(rs, 2)];
|
cannam@167
|
491 T1H = Im[WS(rs, 13)];
|
cannam@167
|
492 T1I = T1G + T1H;
|
cannam@167
|
493 T6n = T1G - T1H;
|
cannam@167
|
494 }
|
cannam@167
|
495 {
|
cannam@167
|
496 E Tj, Tk, T1L, T1M;
|
cannam@167
|
497 Tj = Rp[WS(rs, 10)];
|
cannam@167
|
498 Tk = Rm[WS(rs, 5)];
|
cannam@167
|
499 Tl = Tj + Tk;
|
cannam@167
|
500 T1F = Tj - Tk;
|
cannam@167
|
501 T1L = Ip[WS(rs, 10)];
|
cannam@167
|
502 T1M = Im[WS(rs, 5)];
|
cannam@167
|
503 T1N = T1L + T1M;
|
cannam@167
|
504 T6o = T1L - T1M;
|
cannam@167
|
505 }
|
cannam@167
|
506 T1J = T1F + T1I;
|
cannam@167
|
507 T1O = T1K - T1N;
|
cannam@167
|
508 T80 = T6n - T6o;
|
cannam@167
|
509 T7Z = Ti - Tl;
|
cannam@167
|
510 T4x = T1K + T1N;
|
cannam@167
|
511 T4w = T1I - T1F;
|
cannam@167
|
512 }
|
cannam@167
|
513 {
|
cannam@167
|
514 E T1z, T1x, T1u, T1C;
|
cannam@167
|
515 {
|
cannam@167
|
516 E Tn, To, T1v, T1w;
|
cannam@167
|
517 Tn = Rm[WS(rs, 1)];
|
cannam@167
|
518 To = Rp[WS(rs, 14)];
|
cannam@167
|
519 Tp = Tn + To;
|
cannam@167
|
520 T1z = Tn - To;
|
cannam@167
|
521 T1v = Ip[WS(rs, 14)];
|
cannam@167
|
522 T1w = Im[WS(rs, 1)];
|
cannam@167
|
523 T1x = T1v + T1w;
|
cannam@167
|
524 T6k = T1v - T1w;
|
cannam@167
|
525 }
|
cannam@167
|
526 {
|
cannam@167
|
527 E Tq, Tr, T1A, T1B;
|
cannam@167
|
528 Tq = Rp[WS(rs, 6)];
|
cannam@167
|
529 Tr = Rm[WS(rs, 9)];
|
cannam@167
|
530 Ts = Tq + Tr;
|
cannam@167
|
531 T1u = Tq - Tr;
|
cannam@167
|
532 T1A = Ip[WS(rs, 6)];
|
cannam@167
|
533 T1B = Im[WS(rs, 9)];
|
cannam@167
|
534 T1C = T1A + T1B;
|
cannam@167
|
535 T6l = T1A - T1B;
|
cannam@167
|
536 }
|
cannam@167
|
537 T1y = T1u - T1x;
|
cannam@167
|
538 T1D = T1z - T1C;
|
cannam@167
|
539 T7X = Tp - Ts;
|
cannam@167
|
540 T7W = T6k - T6l;
|
cannam@167
|
541 T4A = T1z + T1C;
|
cannam@167
|
542 T4z = T1u + T1x;
|
cannam@167
|
543 }
|
cannam@167
|
544 {
|
cannam@167
|
545 E Tm, Tt, T7Y, T81;
|
cannam@167
|
546 Tm = Ti + Tl;
|
cannam@167
|
547 Tt = Tp + Ts;
|
cannam@167
|
548 Tu = Tm + Tt;
|
cannam@167
|
549 T6P = Tm - Tt;
|
cannam@167
|
550 T7Y = T7W - T7X;
|
cannam@167
|
551 T81 = T7Z + T80;
|
cannam@167
|
552 T82 = T7Y - T81;
|
cannam@167
|
553 T9a = T81 + T7Y;
|
cannam@167
|
554 }
|
cannam@167
|
555 {
|
cannam@167
|
556 E T8q, T8r, T1E, T1P;
|
cannam@167
|
557 T8q = T7Z - T80;
|
cannam@167
|
558 T8r = T7X + T7W;
|
cannam@167
|
559 T8s = T8q - T8r;
|
cannam@167
|
560 T8X = T8q + T8r;
|
cannam@167
|
561 T1E = FNMS(KP414213562, T1D, T1y);
|
cannam@167
|
562 T1P = FMA(KP414213562, T1O, T1J);
|
cannam@167
|
563 T1Q = T1E - T1P;
|
cannam@167
|
564 T40 = T1P + T1E;
|
cannam@167
|
565 }
|
cannam@167
|
566 {
|
cannam@167
|
567 E T2Y, T2Z, T50, T51;
|
cannam@167
|
568 T2Y = FNMS(KP414213562, T1J, T1O);
|
cannam@167
|
569 T2Z = FMA(KP414213562, T1y, T1D);
|
cannam@167
|
570 T30 = T2Y - T2Z;
|
cannam@167
|
571 T3M = T2Y + T2Z;
|
cannam@167
|
572 T50 = FMA(KP414213562, T4w, T4x);
|
cannam@167
|
573 T51 = FMA(KP414213562, T4z, T4A);
|
cannam@167
|
574 T52 = T50 - T51;
|
cannam@167
|
575 T5w = T50 + T51;
|
cannam@167
|
576 }
|
cannam@167
|
577 {
|
cannam@167
|
578 E T6m, T6p, T4y, T4B;
|
cannam@167
|
579 T6m = T6k + T6l;
|
cannam@167
|
580 T6p = T6n + T6o;
|
cannam@167
|
581 T6q = T6m - T6p;
|
cannam@167
|
582 T7u = T6p + T6m;
|
cannam@167
|
583 T4y = FNMS(KP414213562, T4x, T4w);
|
cannam@167
|
584 T4B = FNMS(KP414213562, T4A, T4z);
|
cannam@167
|
585 T4C = T4y + T4B;
|
cannam@167
|
586 T5K = T4B - T4y;
|
cannam@167
|
587 }
|
cannam@167
|
588 }
|
cannam@167
|
589 {
|
cannam@167
|
590 E Tv, T10, T7p, T7v, T7y, T7z, T7q, T7A;
|
cannam@167
|
591 Tv = Tf + Tu;
|
cannam@167
|
592 T10 = TK + TZ;
|
cannam@167
|
593 T7p = Tv - T10;
|
cannam@167
|
594 T7v = T7t + T7u;
|
cannam@167
|
595 T7y = T7w + T7x;
|
cannam@167
|
596 T7z = T7v - T7y;
|
cannam@167
|
597 Rp[0] = Tv + T10;
|
cannam@167
|
598 Rm[0] = T7v + T7y;
|
cannam@167
|
599 T7q = T7o * T7p;
|
cannam@167
|
600 Rp[WS(rs, 8)] = FNMS(T7s, T7z, T7q);
|
cannam@167
|
601 T7A = T7s * T7p;
|
cannam@167
|
602 Rm[WS(rs, 8)] = FMA(T7o, T7z, T7A);
|
cannam@167
|
603 }
|
cannam@167
|
604 {
|
cannam@167
|
605 E T9p, T9x, T9v, T9z;
|
cannam@167
|
606 {
|
cannam@167
|
607 E T9n, T9o, T9t, T9u;
|
cannam@167
|
608 T9n = FMA(KP707106781, T8X, T8W);
|
cannam@167
|
609 T9o = T9c + T9d;
|
cannam@167
|
610 T9p = FNMS(KP923879532, T9o, T9n);
|
cannam@167
|
611 T9x = FMA(KP923879532, T9o, T9n);
|
cannam@167
|
612 T9t = FMA(KP707106781, T9a, T99);
|
cannam@167
|
613 T9u = T94 + T91;
|
cannam@167
|
614 T9v = FNMS(KP923879532, T9u, T9t);
|
cannam@167
|
615 T9z = FMA(KP923879532, T9u, T9t);
|
cannam@167
|
616 }
|
cannam@167
|
617 {
|
cannam@167
|
618 E T9q, T9w, T9y, T9A;
|
cannam@167
|
619 T9q = T9m * T9p;
|
cannam@167
|
620 Rp[WS(rs, 9)] = FNMS(T9s, T9v, T9q);
|
cannam@167
|
621 T9w = T9m * T9v;
|
cannam@167
|
622 Rm[WS(rs, 9)] = FMA(T9s, T9p, T9w);
|
cannam@167
|
623 T9y = T3v * T9x;
|
cannam@167
|
624 Rp[WS(rs, 1)] = FNMS(T3x, T9z, T9y);
|
cannam@167
|
625 T9A = T3v * T9z;
|
cannam@167
|
626 Rm[WS(rs, 1)] = FMA(T3x, T9x, T9A);
|
cannam@167
|
627 }
|
cannam@167
|
628 }
|
cannam@167
|
629 {
|
cannam@167
|
630 E T8H, T8Q, T8N, T8T;
|
cannam@167
|
631 {
|
cannam@167
|
632 E T8F, T8G, T8L, T8M;
|
cannam@167
|
633 T8F = FNMS(KP707106781, T82, T7V);
|
cannam@167
|
634 T8G = T8u + T8v;
|
cannam@167
|
635 T8H = FNMS(KP923879532, T8G, T8F);
|
cannam@167
|
636 T8Q = FMA(KP923879532, T8G, T8F);
|
cannam@167
|
637 T8L = FNMS(KP707106781, T8s, T8p);
|
cannam@167
|
638 T8M = T8h + T8a;
|
cannam@167
|
639 T8N = FNMS(KP923879532, T8M, T8L);
|
cannam@167
|
640 T8T = FMA(KP923879532, T8M, T8L);
|
cannam@167
|
641 }
|
cannam@167
|
642 {
|
cannam@167
|
643 E T8I, T8O, T8R, T8U;
|
cannam@167
|
644 T8I = T8E * T8H;
|
cannam@167
|
645 Rp[WS(rs, 7)] = FNMS(T8K, T8N, T8I);
|
cannam@167
|
646 T8O = T8E * T8N;
|
cannam@167
|
647 Rm[WS(rs, 7)] = FMA(T8K, T8H, T8O);
|
cannam@167
|
648 T8R = T8P * T8Q;
|
cannam@167
|
649 Rp[WS(rs, 15)] = FNMS(T8S, T8T, T8R);
|
cannam@167
|
650 T8U = T8P * T8T;
|
cannam@167
|
651 Rm[WS(rs, 15)] = FMA(T8S, T8Q, T8U);
|
cannam@167
|
652 }
|
cannam@167
|
653 }
|
cannam@167
|
654 {
|
cannam@167
|
655 E T7b, T7j, T7h, T7l;
|
cannam@167
|
656 {
|
cannam@167
|
657 E T79, T7a, T7f, T7g;
|
cannam@167
|
658 T79 = T6j - T6q;
|
cannam@167
|
659 T7a = T6Z - T6Y;
|
cannam@167
|
660 T7b = FNMS(KP707106781, T7a, T79);
|
cannam@167
|
661 T7j = FMA(KP707106781, T7a, T79);
|
cannam@167
|
662 T7f = T6W - T6P;
|
cannam@167
|
663 T7g = T6A - T6J;
|
cannam@167
|
664 T7h = FNMS(KP707106781, T7g, T7f);
|
cannam@167
|
665 T7l = FMA(KP707106781, T7g, T7f);
|
cannam@167
|
666 }
|
cannam@167
|
667 {
|
cannam@167
|
668 E T7c, T7i, T7k, T7m;
|
cannam@167
|
669 T7c = T78 * T7b;
|
cannam@167
|
670 Rp[WS(rs, 14)] = FNMS(T7e, T7h, T7c);
|
cannam@167
|
671 T7i = T78 * T7h;
|
cannam@167
|
672 Rm[WS(rs, 14)] = FMA(T7e, T7b, T7i);
|
cannam@167
|
673 T7k = T5X * T7j;
|
cannam@167
|
674 Rp[WS(rs, 6)] = FNMS(T5Z, T7l, T7k);
|
cannam@167
|
675 T7m = T5X * T7l;
|
cannam@167
|
676 Rm[WS(rs, 6)] = FMA(T5Z, T7j, T7m);
|
cannam@167
|
677 }
|
cannam@167
|
678 }
|
cannam@167
|
679 {
|
cannam@167
|
680 E T96, T9h, T9f, T9j;
|
cannam@167
|
681 {
|
cannam@167
|
682 E T8Y, T95, T9b, T9e;
|
cannam@167
|
683 T8Y = FNMS(KP707106781, T8X, T8W);
|
cannam@167
|
684 T95 = T91 - T94;
|
cannam@167
|
685 T96 = FNMS(KP923879532, T95, T8Y);
|
cannam@167
|
686 T9h = FMA(KP923879532, T95, T8Y);
|
cannam@167
|
687 T9b = FNMS(KP707106781, T9a, T99);
|
cannam@167
|
688 T9e = T9c - T9d;
|
cannam@167
|
689 T9f = FNMS(KP923879532, T9e, T9b);
|
cannam@167
|
690 T9j = FMA(KP923879532, T9e, T9b);
|
cannam@167
|
691 }
|
cannam@167
|
692 {
|
cannam@167
|
693 E T97, T9g, T9i, T9k;
|
cannam@167
|
694 T97 = T8V * T96;
|
cannam@167
|
695 Rp[WS(rs, 13)] = FNMS(T98, T9f, T97);
|
cannam@167
|
696 T9g = T98 * T96;
|
cannam@167
|
697 Rm[WS(rs, 13)] = FMA(T8V, T9f, T9g);
|
cannam@167
|
698 T9i = T3G * T9h;
|
cannam@167
|
699 Rp[WS(rs, 5)] = FNMS(T3J, T9j, T9i);
|
cannam@167
|
700 T9k = T3J * T9h;
|
cannam@167
|
701 Rm[WS(rs, 5)] = FMA(T3G, T9j, T9k);
|
cannam@167
|
702 }
|
cannam@167
|
703 }
|
cannam@167
|
704 {
|
cannam@167
|
705 E T6L, T73, T71, T75;
|
cannam@167
|
706 {
|
cannam@167
|
707 E T6r, T6K, T6X, T70;
|
cannam@167
|
708 T6r = T6j + T6q;
|
cannam@167
|
709 T6K = T6A + T6J;
|
cannam@167
|
710 T6L = FNMS(KP707106781, T6K, T6r);
|
cannam@167
|
711 T73 = FMA(KP707106781, T6K, T6r);
|
cannam@167
|
712 T6X = T6P + T6W;
|
cannam@167
|
713 T70 = T6Y + T6Z;
|
cannam@167
|
714 T71 = FNMS(KP707106781, T70, T6X);
|
cannam@167
|
715 T75 = FMA(KP707106781, T70, T6X);
|
cannam@167
|
716 }
|
cannam@167
|
717 {
|
cannam@167
|
718 E T6M, T72, T74, T76;
|
cannam@167
|
719 T6M = T6i * T6L;
|
cannam@167
|
720 Rp[WS(rs, 10)] = FNMS(T6O, T71, T6M);
|
cannam@167
|
721 T72 = T6O * T6L;
|
cannam@167
|
722 Rm[WS(rs, 10)] = FMA(T6i, T71, T72);
|
cannam@167
|
723 T74 = T3a * T73;
|
cannam@167
|
724 Rp[WS(rs, 2)] = FNMS(T3d, T75, T74);
|
cannam@167
|
725 T76 = T3d * T73;
|
cannam@167
|
726 Rm[WS(rs, 2)] = FMA(T3a, T75, T76);
|
cannam@167
|
727 }
|
cannam@167
|
728 }
|
cannam@167
|
729 {
|
cannam@167
|
730 E T7F, T7N, T7L, T7P;
|
cannam@167
|
731 {
|
cannam@167
|
732 E T7D, T7E, T7J, T7K;
|
cannam@167
|
733 T7D = Tf - Tu;
|
cannam@167
|
734 T7E = T7x - T7w;
|
cannam@167
|
735 T7F = T7D - T7E;
|
cannam@167
|
736 T7N = T7D + T7E;
|
cannam@167
|
737 T7J = T7t - T7u;
|
cannam@167
|
738 T7K = TK - TZ;
|
cannam@167
|
739 T7L = T7J - T7K;
|
cannam@167
|
740 T7P = T7K + T7J;
|
cannam@167
|
741 }
|
cannam@167
|
742 {
|
cannam@167
|
743 E T7G, T7M, T7O, T7Q;
|
cannam@167
|
744 T7G = T7C * T7F;
|
cannam@167
|
745 Rp[WS(rs, 12)] = FNMS(T7I, T7L, T7G);
|
cannam@167
|
746 T7M = T7C * T7L;
|
cannam@167
|
747 Rm[WS(rs, 12)] = FMA(T7I, T7F, T7M);
|
cannam@167
|
748 T7O = T4p * T7N;
|
cannam@167
|
749 Rp[WS(rs, 4)] = FNMS(T4r, T7P, T7O);
|
cannam@167
|
750 T7Q = T4p * T7P;
|
cannam@167
|
751 Rm[WS(rs, 4)] = FMA(T4r, T7N, T7Q);
|
cannam@167
|
752 }
|
cannam@167
|
753 }
|
cannam@167
|
754 {
|
cannam@167
|
755 E T8j, T8z, T8x, T8B;
|
cannam@167
|
756 {
|
cannam@167
|
757 E T83, T8i, T8t, T8w;
|
cannam@167
|
758 T83 = FMA(KP707106781, T82, T7V);
|
cannam@167
|
759 T8i = T8a - T8h;
|
cannam@167
|
760 T8j = FNMS(KP923879532, T8i, T83);
|
cannam@167
|
761 T8z = FMA(KP923879532, T8i, T83);
|
cannam@167
|
762 T8t = FMA(KP707106781, T8s, T8p);
|
cannam@167
|
763 T8w = T8u - T8v;
|
cannam@167
|
764 T8x = FNMS(KP923879532, T8w, T8t);
|
cannam@167
|
765 T8B = FMA(KP923879532, T8w, T8t);
|
cannam@167
|
766 }
|
cannam@167
|
767 {
|
cannam@167
|
768 E T8k, T8y, T8A, T8C;
|
cannam@167
|
769 T8k = T7S * T8j;
|
cannam@167
|
770 Rp[WS(rs, 11)] = FNMS(T8m, T8x, T8k);
|
cannam@167
|
771 T8y = T8m * T8j;
|
cannam@167
|
772 Rm[WS(rs, 11)] = FMA(T7S, T8x, T8y);
|
cannam@167
|
773 T8A = T16 * T8z;
|
cannam@167
|
774 Rp[WS(rs, 3)] = FNMS(T1a, T8B, T8A);
|
cannam@167
|
775 T8C = T1a * T8z;
|
cannam@167
|
776 Rm[WS(rs, 3)] = FMA(T16, T8B, T8C);
|
cannam@167
|
777 }
|
cannam@167
|
778 }
|
cannam@167
|
779 {
|
cannam@167
|
780 E T3r, T2L, T3s, T3f, T35, T3z, T3j, T3o;
|
cannam@167
|
781 T3r = FNMS(KP923879532, T30, T2X);
|
cannam@167
|
782 {
|
cannam@167
|
783 E T1R, T2i, T2J, T2K;
|
cannam@167
|
784 T1R = FMA(KP923879532, T1Q, T1t);
|
cannam@167
|
785 T2i = FMA(KP668178637, T2h, T28);
|
cannam@167
|
786 T2J = FNMS(KP668178637, T2I, T2z);
|
cannam@167
|
787 T2K = T2i - T2J;
|
cannam@167
|
788 T2L = FNMS(KP831469612, T2K, T1R);
|
cannam@167
|
789 T3s = T2J + T2i;
|
cannam@167
|
790 T3f = FMA(KP831469612, T2K, T1R);
|
cannam@167
|
791 }
|
cannam@167
|
792 {
|
cannam@167
|
793 E T31, T3m, T34, T3n, T32, T33;
|
cannam@167
|
794 T31 = FMA(KP923879532, T30, T2X);
|
cannam@167
|
795 T3m = FNMS(KP923879532, T1Q, T1t);
|
cannam@167
|
796 T32 = FMA(KP668178637, T2z, T2I);
|
cannam@167
|
797 T33 = FNMS(KP668178637, T28, T2h);
|
cannam@167
|
798 T34 = T32 - T33;
|
cannam@167
|
799 T3n = T32 + T33;
|
cannam@167
|
800 T35 = FNMS(KP831469612, T34, T31);
|
cannam@167
|
801 T3z = FMA(KP831469612, T3n, T3m);
|
cannam@167
|
802 T3j = FMA(KP831469612, T34, T31);
|
cannam@167
|
803 T3o = FNMS(KP831469612, T3n, T3m);
|
cannam@167
|
804 }
|
cannam@167
|
805 {
|
cannam@167
|
806 E T2M, T36, T3g, T3k;
|
cannam@167
|
807 T2M = T1c * T2L;
|
cannam@167
|
808 Ip[WS(rs, 10)] = FNMS(T2O, T35, T2M);
|
cannam@167
|
809 T36 = T1c * T35;
|
cannam@167
|
810 Im[WS(rs, 10)] = FMA(T2O, T2L, T36);
|
cannam@167
|
811 T3g = T3e * T3f;
|
cannam@167
|
812 Ip[WS(rs, 2)] = FNMS(T3i, T3j, T3g);
|
cannam@167
|
813 T3k = T3e * T3j;
|
cannam@167
|
814 Im[WS(rs, 2)] = FMA(T3i, T3f, T3k);
|
cannam@167
|
815 {
|
cannam@167
|
816 E T3A, T3E, T3D, T3p, T3u, T3t;
|
cannam@167
|
817 T3A = T3y * T3z;
|
cannam@167
|
818 T3E = T3C * T3z;
|
cannam@167
|
819 T3D = FMA(KP831469612, T3s, T3r);
|
cannam@167
|
820 Ip[WS(rs, 14)] = FNMS(T3C, T3D, T3A);
|
cannam@167
|
821 Im[WS(rs, 14)] = FMA(T3y, T3D, T3E);
|
cannam@167
|
822 T3p = T3l * T3o;
|
cannam@167
|
823 T3u = T3q * T3o;
|
cannam@167
|
824 T3t = FNMS(KP831469612, T3s, T3r);
|
cannam@167
|
825 Ip[WS(rs, 6)] = FNMS(T3q, T3t, T3p);
|
cannam@167
|
826 Im[WS(rs, 6)] = FMA(T3l, T3t, T3u);
|
cannam@167
|
827 }
|
cannam@167
|
828 }
|
cannam@167
|
829 }
|
cannam@167
|
830 {
|
cannam@167
|
831 E T53, T56, T5p, T5h, T4T, T5r, T59, T5n;
|
cannam@167
|
832 T53 = FMA(KP923879532, T52, T4Z);
|
cannam@167
|
833 {
|
cannam@167
|
834 E T5f, T54, T55, T5g;
|
cannam@167
|
835 T5f = FMA(KP923879532, T4C, T4v);
|
cannam@167
|
836 T54 = FMA(KP668178637, T4G, T4J);
|
cannam@167
|
837 T55 = FMA(KP668178637, T4N, T4Q);
|
cannam@167
|
838 T5g = T54 + T55;
|
cannam@167
|
839 T56 = T54 - T55;
|
cannam@167
|
840 T5p = FMA(KP831469612, T5g, T5f);
|
cannam@167
|
841 T5h = FNMS(KP831469612, T5g, T5f);
|
cannam@167
|
842 }
|
cannam@167
|
843 {
|
cannam@167
|
844 E T4D, T5l, T4S, T5m, T4K, T4R;
|
cannam@167
|
845 T4D = FNMS(KP923879532, T4C, T4v);
|
cannam@167
|
846 T5l = FNMS(KP923879532, T52, T4Z);
|
cannam@167
|
847 T4K = FNMS(KP668178637, T4J, T4G);
|
cannam@167
|
848 T4R = FNMS(KP668178637, T4Q, T4N);
|
cannam@167
|
849 T4S = T4K + T4R;
|
cannam@167
|
850 T5m = T4K - T4R;
|
cannam@167
|
851 T4T = FNMS(KP831469612, T4S, T4D);
|
cannam@167
|
852 T5r = FNMS(KP831469612, T5m, T5l);
|
cannam@167
|
853 T59 = FMA(KP831469612, T4S, T4D);
|
cannam@167
|
854 T5n = FMA(KP831469612, T5m, T5l);
|
cannam@167
|
855 }
|
cannam@167
|
856 {
|
cannam@167
|
857 E T5i, T5o, T5q, T5s;
|
cannam@167
|
858 T5i = T5e * T5h;
|
cannam@167
|
859 Ip[WS(rs, 5)] = FNMS(T5k, T5n, T5i);
|
cannam@167
|
860 T5o = T5e * T5n;
|
cannam@167
|
861 Im[WS(rs, 5)] = FMA(T5k, T5h, T5o);
|
cannam@167
|
862 T5q = T17 * T5p;
|
cannam@167
|
863 Ip[WS(rs, 13)] = FNMS(T1b, T5r, T5q);
|
cannam@167
|
864 T5s = T17 * T5r;
|
cannam@167
|
865 Im[WS(rs, 13)] = FMA(T1b, T5p, T5s);
|
cannam@167
|
866 {
|
cannam@167
|
867 E T5a, T5c, T5b, T4U, T58, T57;
|
cannam@167
|
868 T5a = T11 * T59;
|
cannam@167
|
869 T5c = T14 * T59;
|
cannam@167
|
870 T5b = FMA(KP831469612, T56, T53);
|
cannam@167
|
871 Ip[WS(rs, 1)] = FNMS(T14, T5b, T5a);
|
cannam@167
|
872 Im[WS(rs, 1)] = FMA(T11, T5b, T5c);
|
cannam@167
|
873 T4U = T4s * T4T;
|
cannam@167
|
874 T58 = T4W * T4T;
|
cannam@167
|
875 T57 = FNMS(KP831469612, T56, T53);
|
cannam@167
|
876 Ip[WS(rs, 9)] = FNMS(T4W, T57, T4U);
|
cannam@167
|
877 Im[WS(rs, 9)] = FMA(T4s, T57, T58);
|
cannam@167
|
878 }
|
cannam@167
|
879 }
|
cannam@167
|
880 }
|
cannam@167
|
881 {
|
cannam@167
|
882 E T41, T44, T4l, T4e, T3V, T4n, T47, T4j;
|
cannam@167
|
883 T41 = FMA(KP923879532, T40, T3Z);
|
cannam@167
|
884 {
|
cannam@167
|
885 E T4c, T42, T43, T4d;
|
cannam@167
|
886 T4c = FNMS(KP923879532, T3M, T3L);
|
cannam@167
|
887 T42 = FMA(KP198912367, T3O, T3P);
|
cannam@167
|
888 T43 = FNMS(KP198912367, T3R, T3S);
|
cannam@167
|
889 T4d = T43 - T42;
|
cannam@167
|
890 T44 = T42 + T43;
|
cannam@167
|
891 T4l = FMA(KP980785280, T4d, T4c);
|
cannam@167
|
892 T4e = FNMS(KP980785280, T4d, T4c);
|
cannam@167
|
893 }
|
cannam@167
|
894 {
|
cannam@167
|
895 E T3N, T4h, T3U, T4i, T3Q, T3T;
|
cannam@167
|
896 T3N = FMA(KP923879532, T3M, T3L);
|
cannam@167
|
897 T4h = FNMS(KP923879532, T40, T3Z);
|
cannam@167
|
898 T3Q = FNMS(KP198912367, T3P, T3O);
|
cannam@167
|
899 T3T = FMA(KP198912367, T3S, T3R);
|
cannam@167
|
900 T3U = T3Q + T3T;
|
cannam@167
|
901 T4i = T3Q - T3T;
|
cannam@167
|
902 T3V = FNMS(KP980785280, T3U, T3N);
|
cannam@167
|
903 T4n = FMA(KP980785280, T4i, T4h);
|
cannam@167
|
904 T47 = FMA(KP980785280, T3U, T3N);
|
cannam@167
|
905 T4j = FNMS(KP980785280, T4i, T4h);
|
cannam@167
|
906 }
|
cannam@167
|
907 {
|
cannam@167
|
908 E T4f, T4k, T4m, T4o;
|
cannam@167
|
909 T4f = T4b * T4e;
|
cannam@167
|
910 Ip[WS(rs, 12)] = FNMS(T4g, T4j, T4f);
|
cannam@167
|
911 T4k = T4b * T4j;
|
cannam@167
|
912 Im[WS(rs, 12)] = FMA(T4g, T4e, T4k);
|
cannam@167
|
913 T4m = T12 * T4l;
|
cannam@167
|
914 Ip[WS(rs, 4)] = FNMS(T15, T4n, T4m);
|
cannam@167
|
915 T4o = T12 * T4n;
|
cannam@167
|
916 Im[WS(rs, 4)] = FMA(T15, T4l, T4o);
|
cannam@167
|
917 {
|
cannam@167
|
918 E T48, T4a, T49, T3W, T46, T45;
|
cannam@167
|
919 T48 = T37 * T47;
|
cannam@167
|
920 T4a = T39 * T47;
|
cannam@167
|
921 T49 = FMA(KP980785280, T44, T41);
|
cannam@167
|
922 Ip[0] = FNMS(T39, T49, T48);
|
cannam@167
|
923 Im[0] = FMA(T37, T49, T4a);
|
cannam@167
|
924 T3W = T3K * T3V;
|
cannam@167
|
925 T46 = T3Y * T3V;
|
cannam@167
|
926 T45 = FNMS(KP980785280, T44, T41);
|
cannam@167
|
927 Ip[WS(rs, 8)] = FNMS(T3Y, T45, T3W);
|
cannam@167
|
928 Im[WS(rs, 8)] = FMA(T3K, T45, T46);
|
cannam@167
|
929 }
|
cannam@167
|
930 }
|
cannam@167
|
931 }
|
cannam@167
|
932 {
|
cannam@167
|
933 E T5L, T5O, T6c, T63, T5F, T6f, T5S, T69;
|
cannam@167
|
934 T5L = FMA(KP923879532, T5K, T5J);
|
cannam@167
|
935 {
|
cannam@167
|
936 E T61, T5M, T5N, T62;
|
cannam@167
|
937 T61 = FMA(KP923879532, T5w, T5v);
|
cannam@167
|
938 T5M = FMA(KP198912367, T5y, T5z);
|
cannam@167
|
939 T5N = FMA(KP198912367, T5B, T5C);
|
cannam@167
|
940 T62 = T5M + T5N;
|
cannam@167
|
941 T5O = T5M - T5N;
|
cannam@167
|
942 T6c = FMA(KP980785280, T62, T61);
|
cannam@167
|
943 T63 = FNMS(KP980785280, T62, T61);
|
cannam@167
|
944 }
|
cannam@167
|
945 {
|
cannam@167
|
946 E T5x, T67, T5E, T68, T5A, T5D;
|
cannam@167
|
947 T5x = FNMS(KP923879532, T5w, T5v);
|
cannam@167
|
948 T67 = FNMS(KP923879532, T5K, T5J);
|
cannam@167
|
949 T5A = FNMS(KP198912367, T5z, T5y);
|
cannam@167
|
950 T5D = FNMS(KP198912367, T5C, T5B);
|
cannam@167
|
951 T5E = T5A + T5D;
|
cannam@167
|
952 T68 = T5D - T5A;
|
cannam@167
|
953 T5F = FMA(KP980785280, T5E, T5x);
|
cannam@167
|
954 T6f = FNMS(KP980785280, T68, T67);
|
cannam@167
|
955 T5S = FNMS(KP980785280, T5E, T5x);
|
cannam@167
|
956 T69 = FMA(KP980785280, T68, T67);
|
cannam@167
|
957 }
|
cannam@167
|
958 {
|
cannam@167
|
959 E T64, T6a, T6d, T6g;
|
cannam@167
|
960 T64 = T60 * T63;
|
cannam@167
|
961 Ip[WS(rs, 7)] = FNMS(T66, T69, T64);
|
cannam@167
|
962 T6a = T60 * T69;
|
cannam@167
|
963 Im[WS(rs, 7)] = FMA(T66, T63, T6a);
|
cannam@167
|
964 T6d = T6b * T6c;
|
cannam@167
|
965 Ip[WS(rs, 15)] = FNMS(T6e, T6f, T6d);
|
cannam@167
|
966 T6g = T6b * T6f;
|
cannam@167
|
967 Im[WS(rs, 15)] = FMA(T6e, T6c, T6g);
|
cannam@167
|
968 {
|
cannam@167
|
969 E T5T, T5W, T5V, T5G, T5Q, T5P;
|
cannam@167
|
970 T5T = T5R * T5S;
|
cannam@167
|
971 T5W = T5U * T5S;
|
cannam@167
|
972 T5V = FMA(KP980785280, T5O, T5L);
|
cannam@167
|
973 Ip[WS(rs, 3)] = FNMS(T5U, T5V, T5T);
|
cannam@167
|
974 Im[WS(rs, 3)] = FMA(T5R, T5V, T5W);
|
cannam@167
|
975 T5G = T5u * T5F;
|
cannam@167
|
976 T5Q = T5I * T5F;
|
cannam@167
|
977 T5P = FNMS(KP980785280, T5O, T5L);
|
cannam@167
|
978 Ip[WS(rs, 11)] = FNMS(T5I, T5P, T5G);
|
cannam@167
|
979 Im[WS(rs, 11)] = FMA(T5u, T5P, T5Q);
|
cannam@167
|
980 }
|
cannam@167
|
981 }
|
cannam@167
|
982 }
|
cannam@167
|
983 }
|
cannam@167
|
984 }
|
cannam@167
|
985 }
|
cannam@167
|
986 }
|
cannam@167
|
987
|
cannam@167
|
988 static const tw_instr twinstr[] = {
|
cannam@167
|
989 {TW_CEXP, 1, 1},
|
cannam@167
|
990 {TW_CEXP, 1, 3},
|
cannam@167
|
991 {TW_CEXP, 1, 9},
|
cannam@167
|
992 {TW_CEXP, 1, 27},
|
cannam@167
|
993 {TW_NEXT, 1, 0}
|
cannam@167
|
994 };
|
cannam@167
|
995
|
cannam@167
|
996 static const hc2c_desc desc = { 32, "hc2cb2_32", twinstr, &GENUS, {236, 98, 252, 0} };
|
cannam@167
|
997
|
cannam@167
|
998 void X(codelet_hc2cb2_32) (planner *p) {
|
cannam@167
|
999 X(khc2c_register) (p, hc2cb2_32, &desc, HC2C_VIA_RDFT);
|
cannam@167
|
1000 }
|
cannam@167
|
1001 #else
|
cannam@167
|
1002
|
cannam@167
|
1003 /* Generated by: ../../../genfft/gen_hc2c.native -compact -variables 4 -pipeline-latency 4 -sign 1 -twiddle-log3 -precompute-twiddles -n 32 -dif -name hc2cb2_32 -include rdft/scalar/hc2cb.h */
|
cannam@167
|
1004
|
cannam@167
|
1005 /*
|
cannam@167
|
1006 * This function contains 488 FP additions, 280 FP multiplications,
|
cannam@167
|
1007 * (or, 376 additions, 168 multiplications, 112 fused multiply/add),
|
cannam@167
|
1008 * 160 stack variables, 7 constants, and 128 memory accesses
|
cannam@167
|
1009 */
|
cannam@167
|
1010 #include "rdft/scalar/hc2cb.h"
|
cannam@167
|
1011
|
cannam@167
|
1012 static void hc2cb2_32(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms)
|
cannam@167
|
1013 {
|
cannam@167
|
1014 DK(KP555570233, +0.555570233019602224742830813948532874374937191);
|
cannam@167
|
1015 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
cannam@167
|
1016 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
cannam@167
|
1017 DK(KP195090322, +0.195090322016128267848284868477022240927691618);
|
cannam@167
|
1018 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
cannam@167
|
1019 DK(KP382683432, +0.382683432365089771728459984030398866761344562);
|
cannam@167
|
1020 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
cannam@167
|
1021 {
|
cannam@167
|
1022 INT m;
|
cannam@167
|
1023 for (m = mb, W = W + ((mb - 1) * 8); m < me; m = m + 1, Rp = Rp + ms, Ip = Ip + ms, Rm = Rm - ms, Im = Im - ms, W = W + 8, MAKE_VOLATILE_STRIDE(128, rs)) {
|
cannam@167
|
1024 E T11, T14, T12, T15, T17, T2z, T2B, T1c, T18, T1d, T1g, T1k, T2F, T2L, T3t;
|
cannam@167
|
1025 E T4H, T3h, T3V, T3b, T4v, T4T, T4X, T6t, T71, T6z, T75, T81, T8x, T8f, T8z;
|
cannam@167
|
1026 E T2R, T2V, T8p, T8t, T4r, T4t, T53, T69, T3n, T3r, T7P, T7T, T4P, T4R, T6F;
|
cannam@167
|
1027 E T6R, T1f, T2X, T1j, T2Y, T1l, T31, T2d, T2Z, T49, T4h, T4c, T4i, T4d, T4n;
|
cannam@167
|
1028 E T4f, T4j;
|
cannam@167
|
1029 {
|
cannam@167
|
1030 E T2P, T3q, T2U, T3l, T2Q, T3p, T2T, T3m, T2D, T3g, T2K, T39, T2E, T3f, T2J;
|
cannam@167
|
1031 E T3a;
|
cannam@167
|
1032 {
|
cannam@167
|
1033 E T13, T1b, T16, T1a;
|
cannam@167
|
1034 T11 = W[0];
|
cannam@167
|
1035 T14 = W[1];
|
cannam@167
|
1036 T12 = W[2];
|
cannam@167
|
1037 T15 = W[3];
|
cannam@167
|
1038 T13 = T11 * T12;
|
cannam@167
|
1039 T1b = T14 * T12;
|
cannam@167
|
1040 T16 = T14 * T15;
|
cannam@167
|
1041 T1a = T11 * T15;
|
cannam@167
|
1042 T17 = T13 + T16;
|
cannam@167
|
1043 T2z = T13 - T16;
|
cannam@167
|
1044 T2B = T1a + T1b;
|
cannam@167
|
1045 T1c = T1a - T1b;
|
cannam@167
|
1046 T18 = W[4];
|
cannam@167
|
1047 T2P = T12 * T18;
|
cannam@167
|
1048 T3q = T14 * T18;
|
cannam@167
|
1049 T2U = T15 * T18;
|
cannam@167
|
1050 T3l = T11 * T18;
|
cannam@167
|
1051 T1d = W[5];
|
cannam@167
|
1052 T2Q = T15 * T1d;
|
cannam@167
|
1053 T3p = T11 * T1d;
|
cannam@167
|
1054 T2T = T12 * T1d;
|
cannam@167
|
1055 T3m = T14 * T1d;
|
cannam@167
|
1056 T1g = W[6];
|
cannam@167
|
1057 T2D = T11 * T1g;
|
cannam@167
|
1058 T3g = T15 * T1g;
|
cannam@167
|
1059 T2K = T14 * T1g;
|
cannam@167
|
1060 T39 = T12 * T1g;
|
cannam@167
|
1061 T1k = W[7];
|
cannam@167
|
1062 T2E = T14 * T1k;
|
cannam@167
|
1063 T3f = T12 * T1k;
|
cannam@167
|
1064 T2J = T11 * T1k;
|
cannam@167
|
1065 T3a = T15 * T1k;
|
cannam@167
|
1066 }
|
cannam@167
|
1067 T2F = T2D - T2E;
|
cannam@167
|
1068 T2L = T2J + T2K;
|
cannam@167
|
1069 T3t = T39 - T3a;
|
cannam@167
|
1070 T4H = T2J - T2K;
|
cannam@167
|
1071 T3h = T3f - T3g;
|
cannam@167
|
1072 T3V = T3f + T3g;
|
cannam@167
|
1073 T3b = T39 + T3a;
|
cannam@167
|
1074 T4v = T2D + T2E;
|
cannam@167
|
1075 T4T = FMA(T18, T1g, T1d * T1k);
|
cannam@167
|
1076 T4X = FNMS(T1d, T1g, T18 * T1k);
|
cannam@167
|
1077 {
|
cannam@167
|
1078 E T6r, T6s, T6x, T6y;
|
cannam@167
|
1079 T6r = T17 * T1g;
|
cannam@167
|
1080 T6s = T1c * T1k;
|
cannam@167
|
1081 T6t = T6r - T6s;
|
cannam@167
|
1082 T71 = T6r + T6s;
|
cannam@167
|
1083 T6x = T17 * T1k;
|
cannam@167
|
1084 T6y = T1c * T1g;
|
cannam@167
|
1085 T6z = T6x + T6y;
|
cannam@167
|
1086 T75 = T6x - T6y;
|
cannam@167
|
1087 }
|
cannam@167
|
1088 {
|
cannam@167
|
1089 E T7Z, T80, T8d, T8e;
|
cannam@167
|
1090 T7Z = T2z * T1g;
|
cannam@167
|
1091 T80 = T2B * T1k;
|
cannam@167
|
1092 T81 = T7Z + T80;
|
cannam@167
|
1093 T8x = T7Z - T80;
|
cannam@167
|
1094 T8d = T2z * T1k;
|
cannam@167
|
1095 T8e = T2B * T1g;
|
cannam@167
|
1096 T8f = T8d - T8e;
|
cannam@167
|
1097 T8z = T8d + T8e;
|
cannam@167
|
1098 T2R = T2P - T2Q;
|
cannam@167
|
1099 T2V = T2T + T2U;
|
cannam@167
|
1100 T8p = FMA(T2R, T1g, T2V * T1k);
|
cannam@167
|
1101 T8t = FNMS(T2V, T1g, T2R * T1k);
|
cannam@167
|
1102 }
|
cannam@167
|
1103 T4r = T2P + T2Q;
|
cannam@167
|
1104 T4t = T2T - T2U;
|
cannam@167
|
1105 T53 = FMA(T4r, T1g, T4t * T1k);
|
cannam@167
|
1106 T69 = FNMS(T4t, T1g, T4r * T1k);
|
cannam@167
|
1107 T3n = T3l + T3m;
|
cannam@167
|
1108 T3r = T3p - T3q;
|
cannam@167
|
1109 T7P = FMA(T3n, T1g, T3r * T1k);
|
cannam@167
|
1110 T7T = FNMS(T3r, T1g, T3n * T1k);
|
cannam@167
|
1111 T4P = T3l - T3m;
|
cannam@167
|
1112 T4R = T3p + T3q;
|
cannam@167
|
1113 T6F = FMA(T4P, T1g, T4R * T1k);
|
cannam@167
|
1114 T6R = FNMS(T4R, T1g, T4P * T1k);
|
cannam@167
|
1115 {
|
cannam@167
|
1116 E T19, T1e, T1h, T1i;
|
cannam@167
|
1117 T19 = T17 * T18;
|
cannam@167
|
1118 T1e = T1c * T1d;
|
cannam@167
|
1119 T1f = T19 + T1e;
|
cannam@167
|
1120 T2X = T19 - T1e;
|
cannam@167
|
1121 T1h = T17 * T1d;
|
cannam@167
|
1122 T1i = T1c * T18;
|
cannam@167
|
1123 T1j = T1h - T1i;
|
cannam@167
|
1124 T2Y = T1h + T1i;
|
cannam@167
|
1125 }
|
cannam@167
|
1126 T1l = FMA(T1f, T1g, T1j * T1k);
|
cannam@167
|
1127 T31 = FNMS(T2Y, T1g, T2X * T1k);
|
cannam@167
|
1128 T2d = FNMS(T1j, T1g, T1f * T1k);
|
cannam@167
|
1129 T2Z = FMA(T2X, T1g, T2Y * T1k);
|
cannam@167
|
1130 {
|
cannam@167
|
1131 E T47, T48, T4a, T4b;
|
cannam@167
|
1132 T47 = T2z * T18;
|
cannam@167
|
1133 T48 = T2B * T1d;
|
cannam@167
|
1134 T49 = T47 - T48;
|
cannam@167
|
1135 T4h = T47 + T48;
|
cannam@167
|
1136 T4a = T2z * T1d;
|
cannam@167
|
1137 T4b = T2B * T18;
|
cannam@167
|
1138 T4c = T4a + T4b;
|
cannam@167
|
1139 T4i = T4a - T4b;
|
cannam@167
|
1140 }
|
cannam@167
|
1141 T4d = FMA(T49, T1g, T4c * T1k);
|
cannam@167
|
1142 T4n = FNMS(T4i, T1g, T4h * T1k);
|
cannam@167
|
1143 T4f = FNMS(T4c, T1g, T49 * T1k);
|
cannam@167
|
1144 T4j = FMA(T4h, T1g, T4i * T1k);
|
cannam@167
|
1145 }
|
cannam@167
|
1146 {
|
cannam@167
|
1147 E T56, T7b, T7C, T6c, Tf, T1m, T6f, T7c, T3Y, T4I, T2t, T32, T5d, T7D, T3w;
|
cannam@167
|
1148 E T4w, Tu, T2e, T7g, T7F, T7j, T7G, T1B, T33, T3z, T40, T5l, T6i, T5s, T6h;
|
cannam@167
|
1149 E T3C, T3Z, TK, T1D, T7v, T86, T7y, T85, T1S, T35, T3O, T4C, T5F, T6J, T5M;
|
cannam@167
|
1150 E T6K, T3R, T4D, TZ, T1U, T7o, T89, T7r, T88, T29, T36, T3H, T4z, T5Y, T6M;
|
cannam@167
|
1151 E T65, T6N, T3K, T4A;
|
cannam@167
|
1152 {
|
cannam@167
|
1153 E T3, T54, T2h, T6b, T6, T6a, T2k, T55, Ta, T57, T2o, T58, Td, T5a, T2r;
|
cannam@167
|
1154 E T5b;
|
cannam@167
|
1155 {
|
cannam@167
|
1156 E T1, T2, T2f, T2g;
|
cannam@167
|
1157 T1 = Rp[0];
|
cannam@167
|
1158 T2 = Rm[WS(rs, 15)];
|
cannam@167
|
1159 T3 = T1 + T2;
|
cannam@167
|
1160 T54 = T1 - T2;
|
cannam@167
|
1161 T2f = Ip[0];
|
cannam@167
|
1162 T2g = Im[WS(rs, 15)];
|
cannam@167
|
1163 T2h = T2f - T2g;
|
cannam@167
|
1164 T6b = T2f + T2g;
|
cannam@167
|
1165 }
|
cannam@167
|
1166 {
|
cannam@167
|
1167 E T4, T5, T2i, T2j;
|
cannam@167
|
1168 T4 = Rp[WS(rs, 8)];
|
cannam@167
|
1169 T5 = Rm[WS(rs, 7)];
|
cannam@167
|
1170 T6 = T4 + T5;
|
cannam@167
|
1171 T6a = T4 - T5;
|
cannam@167
|
1172 T2i = Ip[WS(rs, 8)];
|
cannam@167
|
1173 T2j = Im[WS(rs, 7)];
|
cannam@167
|
1174 T2k = T2i - T2j;
|
cannam@167
|
1175 T55 = T2i + T2j;
|
cannam@167
|
1176 }
|
cannam@167
|
1177 {
|
cannam@167
|
1178 E T8, T9, T2m, T2n;
|
cannam@167
|
1179 T8 = Rp[WS(rs, 4)];
|
cannam@167
|
1180 T9 = Rm[WS(rs, 11)];
|
cannam@167
|
1181 Ta = T8 + T9;
|
cannam@167
|
1182 T57 = T8 - T9;
|
cannam@167
|
1183 T2m = Ip[WS(rs, 4)];
|
cannam@167
|
1184 T2n = Im[WS(rs, 11)];
|
cannam@167
|
1185 T2o = T2m - T2n;
|
cannam@167
|
1186 T58 = T2m + T2n;
|
cannam@167
|
1187 }
|
cannam@167
|
1188 {
|
cannam@167
|
1189 E Tb, Tc, T2p, T2q;
|
cannam@167
|
1190 Tb = Rm[WS(rs, 3)];
|
cannam@167
|
1191 Tc = Rp[WS(rs, 12)];
|
cannam@167
|
1192 Td = Tb + Tc;
|
cannam@167
|
1193 T5a = Tb - Tc;
|
cannam@167
|
1194 T2p = Ip[WS(rs, 12)];
|
cannam@167
|
1195 T2q = Im[WS(rs, 3)];
|
cannam@167
|
1196 T2r = T2p - T2q;
|
cannam@167
|
1197 T5b = T2p + T2q;
|
cannam@167
|
1198 }
|
cannam@167
|
1199 {
|
cannam@167
|
1200 E T7, Te, T2l, T2s;
|
cannam@167
|
1201 T56 = T54 - T55;
|
cannam@167
|
1202 T7b = T54 + T55;
|
cannam@167
|
1203 T7C = T6b - T6a;
|
cannam@167
|
1204 T6c = T6a + T6b;
|
cannam@167
|
1205 T7 = T3 + T6;
|
cannam@167
|
1206 Te = Ta + Td;
|
cannam@167
|
1207 Tf = T7 + Te;
|
cannam@167
|
1208 T1m = T7 - Te;
|
cannam@167
|
1209 {
|
cannam@167
|
1210 E T6d, T6e, T3W, T3X;
|
cannam@167
|
1211 T6d = T57 + T58;
|
cannam@167
|
1212 T6e = T5a + T5b;
|
cannam@167
|
1213 T6f = KP707106781 * (T6d - T6e);
|
cannam@167
|
1214 T7c = KP707106781 * (T6d + T6e);
|
cannam@167
|
1215 T3W = T2h - T2k;
|
cannam@167
|
1216 T3X = Ta - Td;
|
cannam@167
|
1217 T3Y = T3W - T3X;
|
cannam@167
|
1218 T4I = T3X + T3W;
|
cannam@167
|
1219 }
|
cannam@167
|
1220 T2l = T2h + T2k;
|
cannam@167
|
1221 T2s = T2o + T2r;
|
cannam@167
|
1222 T2t = T2l - T2s;
|
cannam@167
|
1223 T32 = T2l + T2s;
|
cannam@167
|
1224 {
|
cannam@167
|
1225 E T59, T5c, T3u, T3v;
|
cannam@167
|
1226 T59 = T57 - T58;
|
cannam@167
|
1227 T5c = T5a - T5b;
|
cannam@167
|
1228 T5d = KP707106781 * (T59 + T5c);
|
cannam@167
|
1229 T7D = KP707106781 * (T59 - T5c);
|
cannam@167
|
1230 T3u = T3 - T6;
|
cannam@167
|
1231 T3v = T2r - T2o;
|
cannam@167
|
1232 T3w = T3u - T3v;
|
cannam@167
|
1233 T4w = T3u + T3v;
|
cannam@167
|
1234 }
|
cannam@167
|
1235 }
|
cannam@167
|
1236 }
|
cannam@167
|
1237 {
|
cannam@167
|
1238 E Ti, T5p, T1w, T5n, Tl, T5m, T1z, T5q, Tp, T5i, T1p, T5g, Ts, T5f, T1s;
|
cannam@167
|
1239 E T5j;
|
cannam@167
|
1240 {
|
cannam@167
|
1241 E Tg, Th, T1u, T1v;
|
cannam@167
|
1242 Tg = Rp[WS(rs, 2)];
|
cannam@167
|
1243 Th = Rm[WS(rs, 13)];
|
cannam@167
|
1244 Ti = Tg + Th;
|
cannam@167
|
1245 T5p = Tg - Th;
|
cannam@167
|
1246 T1u = Ip[WS(rs, 2)];
|
cannam@167
|
1247 T1v = Im[WS(rs, 13)];
|
cannam@167
|
1248 T1w = T1u - T1v;
|
cannam@167
|
1249 T5n = T1u + T1v;
|
cannam@167
|
1250 }
|
cannam@167
|
1251 {
|
cannam@167
|
1252 E Tj, Tk, T1x, T1y;
|
cannam@167
|
1253 Tj = Rp[WS(rs, 10)];
|
cannam@167
|
1254 Tk = Rm[WS(rs, 5)];
|
cannam@167
|
1255 Tl = Tj + Tk;
|
cannam@167
|
1256 T5m = Tj - Tk;
|
cannam@167
|
1257 T1x = Ip[WS(rs, 10)];
|
cannam@167
|
1258 T1y = Im[WS(rs, 5)];
|
cannam@167
|
1259 T1z = T1x - T1y;
|
cannam@167
|
1260 T5q = T1x + T1y;
|
cannam@167
|
1261 }
|
cannam@167
|
1262 {
|
cannam@167
|
1263 E Tn, To, T1n, T1o;
|
cannam@167
|
1264 Tn = Rm[WS(rs, 1)];
|
cannam@167
|
1265 To = Rp[WS(rs, 14)];
|
cannam@167
|
1266 Tp = Tn + To;
|
cannam@167
|
1267 T5i = Tn - To;
|
cannam@167
|
1268 T1n = Ip[WS(rs, 14)];
|
cannam@167
|
1269 T1o = Im[WS(rs, 1)];
|
cannam@167
|
1270 T1p = T1n - T1o;
|
cannam@167
|
1271 T5g = T1n + T1o;
|
cannam@167
|
1272 }
|
cannam@167
|
1273 {
|
cannam@167
|
1274 E Tq, Tr, T1q, T1r;
|
cannam@167
|
1275 Tq = Rp[WS(rs, 6)];
|
cannam@167
|
1276 Tr = Rm[WS(rs, 9)];
|
cannam@167
|
1277 Ts = Tq + Tr;
|
cannam@167
|
1278 T5f = Tq - Tr;
|
cannam@167
|
1279 T1q = Ip[WS(rs, 6)];
|
cannam@167
|
1280 T1r = Im[WS(rs, 9)];
|
cannam@167
|
1281 T1s = T1q - T1r;
|
cannam@167
|
1282 T5j = T1q + T1r;
|
cannam@167
|
1283 }
|
cannam@167
|
1284 {
|
cannam@167
|
1285 E Tm, Tt, T7e, T7f;
|
cannam@167
|
1286 Tm = Ti + Tl;
|
cannam@167
|
1287 Tt = Tp + Ts;
|
cannam@167
|
1288 Tu = Tm + Tt;
|
cannam@167
|
1289 T2e = Tm - Tt;
|
cannam@167
|
1290 T7e = T5p + T5q;
|
cannam@167
|
1291 T7f = T5n - T5m;
|
cannam@167
|
1292 T7g = FNMS(KP923879532, T7f, KP382683432 * T7e);
|
cannam@167
|
1293 T7F = FMA(KP382683432, T7f, KP923879532 * T7e);
|
cannam@167
|
1294 }
|
cannam@167
|
1295 {
|
cannam@167
|
1296 E T7h, T7i, T1t, T1A;
|
cannam@167
|
1297 T7h = T5i + T5j;
|
cannam@167
|
1298 T7i = T5f + T5g;
|
cannam@167
|
1299 T7j = FNMS(KP923879532, T7i, KP382683432 * T7h);
|
cannam@167
|
1300 T7G = FMA(KP382683432, T7i, KP923879532 * T7h);
|
cannam@167
|
1301 T1t = T1p + T1s;
|
cannam@167
|
1302 T1A = T1w + T1z;
|
cannam@167
|
1303 T1B = T1t - T1A;
|
cannam@167
|
1304 T33 = T1A + T1t;
|
cannam@167
|
1305 }
|
cannam@167
|
1306 {
|
cannam@167
|
1307 E T3x, T3y, T5h, T5k;
|
cannam@167
|
1308 T3x = T1p - T1s;
|
cannam@167
|
1309 T3y = Tp - Ts;
|
cannam@167
|
1310 T3z = T3x - T3y;
|
cannam@167
|
1311 T40 = T3y + T3x;
|
cannam@167
|
1312 T5h = T5f - T5g;
|
cannam@167
|
1313 T5k = T5i - T5j;
|
cannam@167
|
1314 T5l = FNMS(KP382683432, T5k, KP923879532 * T5h);
|
cannam@167
|
1315 T6i = FMA(KP382683432, T5h, KP923879532 * T5k);
|
cannam@167
|
1316 }
|
cannam@167
|
1317 {
|
cannam@167
|
1318 E T5o, T5r, T3A, T3B;
|
cannam@167
|
1319 T5o = T5m + T5n;
|
cannam@167
|
1320 T5r = T5p - T5q;
|
cannam@167
|
1321 T5s = FMA(KP923879532, T5o, KP382683432 * T5r);
|
cannam@167
|
1322 T6h = FNMS(KP382683432, T5o, KP923879532 * T5r);
|
cannam@167
|
1323 T3A = Ti - Tl;
|
cannam@167
|
1324 T3B = T1w - T1z;
|
cannam@167
|
1325 T3C = T3A + T3B;
|
cannam@167
|
1326 T3Z = T3A - T3B;
|
cannam@167
|
1327 }
|
cannam@167
|
1328 }
|
cannam@167
|
1329 {
|
cannam@167
|
1330 E Ty, T5v, T1G, T5H, TB, T5G, T1J, T5w, TI, T5K, T1Q, T5D, TF, T5J, T1N;
|
cannam@167
|
1331 E T5A;
|
cannam@167
|
1332 {
|
cannam@167
|
1333 E Tw, Tx, T1H, T1I;
|
cannam@167
|
1334 Tw = Rp[WS(rs, 1)];
|
cannam@167
|
1335 Tx = Rm[WS(rs, 14)];
|
cannam@167
|
1336 Ty = Tw + Tx;
|
cannam@167
|
1337 T5v = Tw - Tx;
|
cannam@167
|
1338 {
|
cannam@167
|
1339 E T1E, T1F, Tz, TA;
|
cannam@167
|
1340 T1E = Ip[WS(rs, 1)];
|
cannam@167
|
1341 T1F = Im[WS(rs, 14)];
|
cannam@167
|
1342 T1G = T1E - T1F;
|
cannam@167
|
1343 T5H = T1E + T1F;
|
cannam@167
|
1344 Tz = Rp[WS(rs, 9)];
|
cannam@167
|
1345 TA = Rm[WS(rs, 6)];
|
cannam@167
|
1346 TB = Tz + TA;
|
cannam@167
|
1347 T5G = Tz - TA;
|
cannam@167
|
1348 }
|
cannam@167
|
1349 T1H = Ip[WS(rs, 9)];
|
cannam@167
|
1350 T1I = Im[WS(rs, 6)];
|
cannam@167
|
1351 T1J = T1H - T1I;
|
cannam@167
|
1352 T5w = T1H + T1I;
|
cannam@167
|
1353 {
|
cannam@167
|
1354 E TG, TH, T5B, T1O, T1P, T5C;
|
cannam@167
|
1355 TG = Rm[WS(rs, 2)];
|
cannam@167
|
1356 TH = Rp[WS(rs, 13)];
|
cannam@167
|
1357 T5B = TG - TH;
|
cannam@167
|
1358 T1O = Ip[WS(rs, 13)];
|
cannam@167
|
1359 T1P = Im[WS(rs, 2)];
|
cannam@167
|
1360 T5C = T1O + T1P;
|
cannam@167
|
1361 TI = TG + TH;
|
cannam@167
|
1362 T5K = T5B + T5C;
|
cannam@167
|
1363 T1Q = T1O - T1P;
|
cannam@167
|
1364 T5D = T5B - T5C;
|
cannam@167
|
1365 }
|
cannam@167
|
1366 {
|
cannam@167
|
1367 E TD, TE, T5y, T1L, T1M, T5z;
|
cannam@167
|
1368 TD = Rp[WS(rs, 5)];
|
cannam@167
|
1369 TE = Rm[WS(rs, 10)];
|
cannam@167
|
1370 T5y = TD - TE;
|
cannam@167
|
1371 T1L = Ip[WS(rs, 5)];
|
cannam@167
|
1372 T1M = Im[WS(rs, 10)];
|
cannam@167
|
1373 T5z = T1L + T1M;
|
cannam@167
|
1374 TF = TD + TE;
|
cannam@167
|
1375 T5J = T5y + T5z;
|
cannam@167
|
1376 T1N = T1L - T1M;
|
cannam@167
|
1377 T5A = T5y - T5z;
|
cannam@167
|
1378 }
|
cannam@167
|
1379 }
|
cannam@167
|
1380 {
|
cannam@167
|
1381 E TC, TJ, T7t, T7u;
|
cannam@167
|
1382 TC = Ty + TB;
|
cannam@167
|
1383 TJ = TF + TI;
|
cannam@167
|
1384 TK = TC + TJ;
|
cannam@167
|
1385 T1D = TC - TJ;
|
cannam@167
|
1386 T7t = T5H - T5G;
|
cannam@167
|
1387 T7u = KP707106781 * (T5A - T5D);
|
cannam@167
|
1388 T7v = T7t + T7u;
|
cannam@167
|
1389 T86 = T7t - T7u;
|
cannam@167
|
1390 }
|
cannam@167
|
1391 {
|
cannam@167
|
1392 E T7w, T7x, T1K, T1R;
|
cannam@167
|
1393 T7w = T5v + T5w;
|
cannam@167
|
1394 T7x = KP707106781 * (T5J + T5K);
|
cannam@167
|
1395 T7y = T7w - T7x;
|
cannam@167
|
1396 T85 = T7w + T7x;
|
cannam@167
|
1397 T1K = T1G + T1J;
|
cannam@167
|
1398 T1R = T1N + T1Q;
|
cannam@167
|
1399 T1S = T1K - T1R;
|
cannam@167
|
1400 T35 = T1K + T1R;
|
cannam@167
|
1401 }
|
cannam@167
|
1402 {
|
cannam@167
|
1403 E T3M, T3N, T5x, T5E;
|
cannam@167
|
1404 T3M = T1G - T1J;
|
cannam@167
|
1405 T3N = TF - TI;
|
cannam@167
|
1406 T3O = T3M - T3N;
|
cannam@167
|
1407 T4C = T3N + T3M;
|
cannam@167
|
1408 T5x = T5v - T5w;
|
cannam@167
|
1409 T5E = KP707106781 * (T5A + T5D);
|
cannam@167
|
1410 T5F = T5x - T5E;
|
cannam@167
|
1411 T6J = T5x + T5E;
|
cannam@167
|
1412 }
|
cannam@167
|
1413 {
|
cannam@167
|
1414 E T5I, T5L, T3P, T3Q;
|
cannam@167
|
1415 T5I = T5G + T5H;
|
cannam@167
|
1416 T5L = KP707106781 * (T5J - T5K);
|
cannam@167
|
1417 T5M = T5I - T5L;
|
cannam@167
|
1418 T6K = T5I + T5L;
|
cannam@167
|
1419 T3P = Ty - TB;
|
cannam@167
|
1420 T3Q = T1Q - T1N;
|
cannam@167
|
1421 T3R = T3P - T3Q;
|
cannam@167
|
1422 T4D = T3P + T3Q;
|
cannam@167
|
1423 }
|
cannam@167
|
1424 }
|
cannam@167
|
1425 {
|
cannam@167
|
1426 E TN, T5O, T1X, T60, TQ, T5Z, T20, T5P, TX, T63, T27, T5W, TU, T62, T24;
|
cannam@167
|
1427 E T5T;
|
cannam@167
|
1428 {
|
cannam@167
|
1429 E TL, TM, T1Y, T1Z;
|
cannam@167
|
1430 TL = Rm[0];
|
cannam@167
|
1431 TM = Rp[WS(rs, 15)];
|
cannam@167
|
1432 TN = TL + TM;
|
cannam@167
|
1433 T5O = TL - TM;
|
cannam@167
|
1434 {
|
cannam@167
|
1435 E T1V, T1W, TO, TP;
|
cannam@167
|
1436 T1V = Ip[WS(rs, 15)];
|
cannam@167
|
1437 T1W = Im[0];
|
cannam@167
|
1438 T1X = T1V - T1W;
|
cannam@167
|
1439 T60 = T1V + T1W;
|
cannam@167
|
1440 TO = Rp[WS(rs, 7)];
|
cannam@167
|
1441 TP = Rm[WS(rs, 8)];
|
cannam@167
|
1442 TQ = TO + TP;
|
cannam@167
|
1443 T5Z = TO - TP;
|
cannam@167
|
1444 }
|
cannam@167
|
1445 T1Y = Ip[WS(rs, 7)];
|
cannam@167
|
1446 T1Z = Im[WS(rs, 8)];
|
cannam@167
|
1447 T20 = T1Y - T1Z;
|
cannam@167
|
1448 T5P = T1Y + T1Z;
|
cannam@167
|
1449 {
|
cannam@167
|
1450 E TV, TW, T5U, T25, T26, T5V;
|
cannam@167
|
1451 TV = Rm[WS(rs, 4)];
|
cannam@167
|
1452 TW = Rp[WS(rs, 11)];
|
cannam@167
|
1453 T5U = TV - TW;
|
cannam@167
|
1454 T25 = Ip[WS(rs, 11)];
|
cannam@167
|
1455 T26 = Im[WS(rs, 4)];
|
cannam@167
|
1456 T5V = T25 + T26;
|
cannam@167
|
1457 TX = TV + TW;
|
cannam@167
|
1458 T63 = T5U + T5V;
|
cannam@167
|
1459 T27 = T25 - T26;
|
cannam@167
|
1460 T5W = T5U - T5V;
|
cannam@167
|
1461 }
|
cannam@167
|
1462 {
|
cannam@167
|
1463 E TS, TT, T5R, T22, T23, T5S;
|
cannam@167
|
1464 TS = Rp[WS(rs, 3)];
|
cannam@167
|
1465 TT = Rm[WS(rs, 12)];
|
cannam@167
|
1466 T5R = TS - TT;
|
cannam@167
|
1467 T22 = Ip[WS(rs, 3)];
|
cannam@167
|
1468 T23 = Im[WS(rs, 12)];
|
cannam@167
|
1469 T5S = T22 + T23;
|
cannam@167
|
1470 TU = TS + TT;
|
cannam@167
|
1471 T62 = T5R + T5S;
|
cannam@167
|
1472 T24 = T22 - T23;
|
cannam@167
|
1473 T5T = T5R - T5S;
|
cannam@167
|
1474 }
|
cannam@167
|
1475 }
|
cannam@167
|
1476 {
|
cannam@167
|
1477 E TR, TY, T7m, T7n;
|
cannam@167
|
1478 TR = TN + TQ;
|
cannam@167
|
1479 TY = TU + TX;
|
cannam@167
|
1480 TZ = TR + TY;
|
cannam@167
|
1481 T1U = TR - TY;
|
cannam@167
|
1482 T7m = KP707106781 * (T5T - T5W);
|
cannam@167
|
1483 T7n = T5Z + T60;
|
cannam@167
|
1484 T7o = T7m - T7n;
|
cannam@167
|
1485 T89 = T7n + T7m;
|
cannam@167
|
1486 }
|
cannam@167
|
1487 {
|
cannam@167
|
1488 E T7p, T7q, T21, T28;
|
cannam@167
|
1489 T7p = T5O + T5P;
|
cannam@167
|
1490 T7q = KP707106781 * (T62 + T63);
|
cannam@167
|
1491 T7r = T7p - T7q;
|
cannam@167
|
1492 T88 = T7p + T7q;
|
cannam@167
|
1493 T21 = T1X + T20;
|
cannam@167
|
1494 T28 = T24 + T27;
|
cannam@167
|
1495 T29 = T21 - T28;
|
cannam@167
|
1496 T36 = T21 + T28;
|
cannam@167
|
1497 }
|
cannam@167
|
1498 {
|
cannam@167
|
1499 E T3F, T3G, T5Q, T5X;
|
cannam@167
|
1500 T3F = T1X - T20;
|
cannam@167
|
1501 T3G = TU - TX;
|
cannam@167
|
1502 T3H = T3F - T3G;
|
cannam@167
|
1503 T4z = T3G + T3F;
|
cannam@167
|
1504 T5Q = T5O - T5P;
|
cannam@167
|
1505 T5X = KP707106781 * (T5T + T5W);
|
cannam@167
|
1506 T5Y = T5Q - T5X;
|
cannam@167
|
1507 T6M = T5Q + T5X;
|
cannam@167
|
1508 }
|
cannam@167
|
1509 {
|
cannam@167
|
1510 E T61, T64, T3I, T3J;
|
cannam@167
|
1511 T61 = T5Z - T60;
|
cannam@167
|
1512 T64 = KP707106781 * (T62 - T63);
|
cannam@167
|
1513 T65 = T61 - T64;
|
cannam@167
|
1514 T6N = T61 + T64;
|
cannam@167
|
1515 T3I = TN - TQ;
|
cannam@167
|
1516 T3J = T27 - T24;
|
cannam@167
|
1517 T3K = T3I - T3J;
|
cannam@167
|
1518 T4A = T3I + T3J;
|
cannam@167
|
1519 }
|
cannam@167
|
1520 }
|
cannam@167
|
1521 {
|
cannam@167
|
1522 E Tv, T10, T30, T34, T37, T38;
|
cannam@167
|
1523 Tv = Tf + Tu;
|
cannam@167
|
1524 T10 = TK + TZ;
|
cannam@167
|
1525 T30 = Tv - T10;
|
cannam@167
|
1526 T34 = T32 + T33;
|
cannam@167
|
1527 T37 = T35 + T36;
|
cannam@167
|
1528 T38 = T34 - T37;
|
cannam@167
|
1529 Rp[0] = Tv + T10;
|
cannam@167
|
1530 Rm[0] = T34 + T37;
|
cannam@167
|
1531 Rp[WS(rs, 8)] = FNMS(T31, T38, T2Z * T30);
|
cannam@167
|
1532 Rm[WS(rs, 8)] = FMA(T31, T30, T2Z * T38);
|
cannam@167
|
1533 }
|
cannam@167
|
1534 {
|
cannam@167
|
1535 E T3e, T3o, T3k, T3s;
|
cannam@167
|
1536 {
|
cannam@167
|
1537 E T3c, T3d, T3i, T3j;
|
cannam@167
|
1538 T3c = Tf - Tu;
|
cannam@167
|
1539 T3d = T36 - T35;
|
cannam@167
|
1540 T3e = T3c - T3d;
|
cannam@167
|
1541 T3o = T3c + T3d;
|
cannam@167
|
1542 T3i = T32 - T33;
|
cannam@167
|
1543 T3j = TK - TZ;
|
cannam@167
|
1544 T3k = T3i - T3j;
|
cannam@167
|
1545 T3s = T3j + T3i;
|
cannam@167
|
1546 }
|
cannam@167
|
1547 Rp[WS(rs, 12)] = FNMS(T3h, T3k, T3b * T3e);
|
cannam@167
|
1548 Rm[WS(rs, 12)] = FMA(T3b, T3k, T3h * T3e);
|
cannam@167
|
1549 Rp[WS(rs, 4)] = FNMS(T3r, T3s, T3n * T3o);
|
cannam@167
|
1550 Rm[WS(rs, 4)] = FMA(T3n, T3s, T3r * T3o);
|
cannam@167
|
1551 }
|
cannam@167
|
1552 {
|
cannam@167
|
1553 E T1C, T2u, T2M, T2G, T2x, T2H, T2b, T2N;
|
cannam@167
|
1554 T1C = T1m + T1B;
|
cannam@167
|
1555 T2u = T2e + T2t;
|
cannam@167
|
1556 T2M = T2t - T2e;
|
cannam@167
|
1557 T2G = T1m - T1B;
|
cannam@167
|
1558 {
|
cannam@167
|
1559 E T2v, T2w, T1T, T2a;
|
cannam@167
|
1560 T2v = T1D + T1S;
|
cannam@167
|
1561 T2w = T29 - T1U;
|
cannam@167
|
1562 T2x = KP707106781 * (T2v + T2w);
|
cannam@167
|
1563 T2H = KP707106781 * (T2w - T2v);
|
cannam@167
|
1564 T1T = T1D - T1S;
|
cannam@167
|
1565 T2a = T1U + T29;
|
cannam@167
|
1566 T2b = KP707106781 * (T1T + T2a);
|
cannam@167
|
1567 T2N = KP707106781 * (T1T - T2a);
|
cannam@167
|
1568 }
|
cannam@167
|
1569 {
|
cannam@167
|
1570 E T2c, T2y, T2S, T2W;
|
cannam@167
|
1571 T2c = T1C - T2b;
|
cannam@167
|
1572 T2y = T2u - T2x;
|
cannam@167
|
1573 Rp[WS(rs, 10)] = FNMS(T2d, T2y, T1l * T2c);
|
cannam@167
|
1574 Rm[WS(rs, 10)] = FMA(T2d, T2c, T1l * T2y);
|
cannam@167
|
1575 T2S = T2G + T2H;
|
cannam@167
|
1576 T2W = T2M + T2N;
|
cannam@167
|
1577 Rp[WS(rs, 6)] = FNMS(T2V, T2W, T2R * T2S);
|
cannam@167
|
1578 Rm[WS(rs, 6)] = FMA(T2R, T2W, T2V * T2S);
|
cannam@167
|
1579 }
|
cannam@167
|
1580 {
|
cannam@167
|
1581 E T2A, T2C, T2I, T2O;
|
cannam@167
|
1582 T2A = T1C + T2b;
|
cannam@167
|
1583 T2C = T2u + T2x;
|
cannam@167
|
1584 Rp[WS(rs, 2)] = FNMS(T2B, T2C, T2z * T2A);
|
cannam@167
|
1585 Rm[WS(rs, 2)] = FMA(T2B, T2A, T2z * T2C);
|
cannam@167
|
1586 T2I = T2G - T2H;
|
cannam@167
|
1587 T2O = T2M - T2N;
|
cannam@167
|
1588 Rp[WS(rs, 14)] = FNMS(T2L, T2O, T2F * T2I);
|
cannam@167
|
1589 Rm[WS(rs, 14)] = FMA(T2F, T2O, T2L * T2I);
|
cannam@167
|
1590 }
|
cannam@167
|
1591 }
|
cannam@167
|
1592 {
|
cannam@167
|
1593 E T4y, T4U, T4K, T4Y, T4F, T4Z, T4N, T4V, T4x, T4J;
|
cannam@167
|
1594 T4x = KP707106781 * (T3Z + T40);
|
cannam@167
|
1595 T4y = T4w - T4x;
|
cannam@167
|
1596 T4U = T4w + T4x;
|
cannam@167
|
1597 T4J = KP707106781 * (T3C + T3z);
|
cannam@167
|
1598 T4K = T4I - T4J;
|
cannam@167
|
1599 T4Y = T4I + T4J;
|
cannam@167
|
1600 {
|
cannam@167
|
1601 E T4B, T4E, T4L, T4M;
|
cannam@167
|
1602 T4B = FNMS(KP382683432, T4A, KP923879532 * T4z);
|
cannam@167
|
1603 T4E = FMA(KP923879532, T4C, KP382683432 * T4D);
|
cannam@167
|
1604 T4F = T4B - T4E;
|
cannam@167
|
1605 T4Z = T4E + T4B;
|
cannam@167
|
1606 T4L = FNMS(KP382683432, T4C, KP923879532 * T4D);
|
cannam@167
|
1607 T4M = FMA(KP382683432, T4z, KP923879532 * T4A);
|
cannam@167
|
1608 T4N = T4L - T4M;
|
cannam@167
|
1609 T4V = T4L + T4M;
|
cannam@167
|
1610 }
|
cannam@167
|
1611 {
|
cannam@167
|
1612 E T4G, T4O, T51, T52;
|
cannam@167
|
1613 T4G = T4y - T4F;
|
cannam@167
|
1614 T4O = T4K - T4N;
|
cannam@167
|
1615 Rp[WS(rs, 13)] = FNMS(T4H, T4O, T4v * T4G);
|
cannam@167
|
1616 Rm[WS(rs, 13)] = FMA(T4H, T4G, T4v * T4O);
|
cannam@167
|
1617 T51 = T4U + T4V;
|
cannam@167
|
1618 T52 = T4Y + T4Z;
|
cannam@167
|
1619 Rp[WS(rs, 1)] = FNMS(T1c, T52, T17 * T51);
|
cannam@167
|
1620 Rm[WS(rs, 1)] = FMA(T17, T52, T1c * T51);
|
cannam@167
|
1621 }
|
cannam@167
|
1622 {
|
cannam@167
|
1623 E T4Q, T4S, T4W, T50;
|
cannam@167
|
1624 T4Q = T4y + T4F;
|
cannam@167
|
1625 T4S = T4K + T4N;
|
cannam@167
|
1626 Rp[WS(rs, 5)] = FNMS(T4R, T4S, T4P * T4Q);
|
cannam@167
|
1627 Rm[WS(rs, 5)] = FMA(T4R, T4Q, T4P * T4S);
|
cannam@167
|
1628 T4W = T4U - T4V;
|
cannam@167
|
1629 T50 = T4Y - T4Z;
|
cannam@167
|
1630 Rp[WS(rs, 9)] = FNMS(T4X, T50, T4T * T4W);
|
cannam@167
|
1631 Rm[WS(rs, 9)] = FMA(T4T, T50, T4X * T4W);
|
cannam@167
|
1632 }
|
cannam@167
|
1633 }
|
cannam@167
|
1634 {
|
cannam@167
|
1635 E T3E, T4k, T42, T4o, T3T, T4p, T45, T4l, T3D, T41;
|
cannam@167
|
1636 T3D = KP707106781 * (T3z - T3C);
|
cannam@167
|
1637 T3E = T3w - T3D;
|
cannam@167
|
1638 T4k = T3w + T3D;
|
cannam@167
|
1639 T41 = KP707106781 * (T3Z - T40);
|
cannam@167
|
1640 T42 = T3Y - T41;
|
cannam@167
|
1641 T4o = T3Y + T41;
|
cannam@167
|
1642 {
|
cannam@167
|
1643 E T3L, T3S, T43, T44;
|
cannam@167
|
1644 T3L = FNMS(KP923879532, T3K, KP382683432 * T3H);
|
cannam@167
|
1645 T3S = FMA(KP382683432, T3O, KP923879532 * T3R);
|
cannam@167
|
1646 T3T = T3L - T3S;
|
cannam@167
|
1647 T4p = T3S + T3L;
|
cannam@167
|
1648 T43 = FNMS(KP923879532, T3O, KP382683432 * T3R);
|
cannam@167
|
1649 T44 = FMA(KP923879532, T3H, KP382683432 * T3K);
|
cannam@167
|
1650 T45 = T43 - T44;
|
cannam@167
|
1651 T4l = T43 + T44;
|
cannam@167
|
1652 }
|
cannam@167
|
1653 {
|
cannam@167
|
1654 E T3U, T46, T4s, T4u;
|
cannam@167
|
1655 T3U = T3E - T3T;
|
cannam@167
|
1656 T46 = T42 - T45;
|
cannam@167
|
1657 Rp[WS(rs, 15)] = FNMS(T3V, T46, T3t * T3U);
|
cannam@167
|
1658 Rm[WS(rs, 15)] = FMA(T3V, T3U, T3t * T46);
|
cannam@167
|
1659 T4s = T4k + T4l;
|
cannam@167
|
1660 T4u = T4o + T4p;
|
cannam@167
|
1661 Rp[WS(rs, 3)] = FNMS(T4t, T4u, T4r * T4s);
|
cannam@167
|
1662 Rm[WS(rs, 3)] = FMA(T4r, T4u, T4t * T4s);
|
cannam@167
|
1663 }
|
cannam@167
|
1664 {
|
cannam@167
|
1665 E T4e, T4g, T4m, T4q;
|
cannam@167
|
1666 T4e = T3E + T3T;
|
cannam@167
|
1667 T4g = T42 + T45;
|
cannam@167
|
1668 Rp[WS(rs, 7)] = FNMS(T4f, T4g, T4d * T4e);
|
cannam@167
|
1669 Rm[WS(rs, 7)] = FMA(T4f, T4e, T4d * T4g);
|
cannam@167
|
1670 T4m = T4k - T4l;
|
cannam@167
|
1671 T4q = T4o - T4p;
|
cannam@167
|
1672 Rp[WS(rs, 11)] = FNMS(T4n, T4q, T4j * T4m);
|
cannam@167
|
1673 Rm[WS(rs, 11)] = FMA(T4j, T4q, T4n * T4m);
|
cannam@167
|
1674 }
|
cannam@167
|
1675 }
|
cannam@167
|
1676 {
|
cannam@167
|
1677 E T6I, T72, T6X, T73, T6P, T77, T6U, T76;
|
cannam@167
|
1678 {
|
cannam@167
|
1679 E T6G, T6H, T6V, T6W;
|
cannam@167
|
1680 T6G = T56 + T5d;
|
cannam@167
|
1681 T6H = T6h + T6i;
|
cannam@167
|
1682 T6I = T6G + T6H;
|
cannam@167
|
1683 T72 = T6G - T6H;
|
cannam@167
|
1684 T6V = FMA(KP195090322, T6J, KP980785280 * T6K);
|
cannam@167
|
1685 T6W = FNMS(KP195090322, T6M, KP980785280 * T6N);
|
cannam@167
|
1686 T6X = T6V + T6W;
|
cannam@167
|
1687 T73 = T6W - T6V;
|
cannam@167
|
1688 }
|
cannam@167
|
1689 {
|
cannam@167
|
1690 E T6L, T6O, T6S, T6T;
|
cannam@167
|
1691 T6L = FNMS(KP195090322, T6K, KP980785280 * T6J);
|
cannam@167
|
1692 T6O = FMA(KP980785280, T6M, KP195090322 * T6N);
|
cannam@167
|
1693 T6P = T6L + T6O;
|
cannam@167
|
1694 T77 = T6L - T6O;
|
cannam@167
|
1695 T6S = T6c + T6f;
|
cannam@167
|
1696 T6T = T5s + T5l;
|
cannam@167
|
1697 T6U = T6S + T6T;
|
cannam@167
|
1698 T76 = T6S - T6T;
|
cannam@167
|
1699 }
|
cannam@167
|
1700 {
|
cannam@167
|
1701 E T6Q, T6Y, T79, T7a;
|
cannam@167
|
1702 T6Q = T6I - T6P;
|
cannam@167
|
1703 T6Y = T6U - T6X;
|
cannam@167
|
1704 Ip[WS(rs, 8)] = FNMS(T6R, T6Y, T6F * T6Q);
|
cannam@167
|
1705 Im[WS(rs, 8)] = FMA(T6R, T6Q, T6F * T6Y);
|
cannam@167
|
1706 T79 = T72 + T73;
|
cannam@167
|
1707 T7a = T76 + T77;
|
cannam@167
|
1708 Ip[WS(rs, 4)] = FNMS(T1d, T7a, T18 * T79);
|
cannam@167
|
1709 Im[WS(rs, 4)] = FMA(T18, T7a, T1d * T79);
|
cannam@167
|
1710 }
|
cannam@167
|
1711 {
|
cannam@167
|
1712 E T6Z, T70, T74, T78;
|
cannam@167
|
1713 T6Z = T6I + T6P;
|
cannam@167
|
1714 T70 = T6U + T6X;
|
cannam@167
|
1715 Ip[0] = FNMS(T14, T70, T11 * T6Z);
|
cannam@167
|
1716 Im[0] = FMA(T14, T6Z, T11 * T70);
|
cannam@167
|
1717 T74 = T72 - T73;
|
cannam@167
|
1718 T78 = T76 - T77;
|
cannam@167
|
1719 Ip[WS(rs, 12)] = FNMS(T75, T78, T71 * T74);
|
cannam@167
|
1720 Im[WS(rs, 12)] = FMA(T71, T78, T75 * T74);
|
cannam@167
|
1721 }
|
cannam@167
|
1722 }
|
cannam@167
|
1723 {
|
cannam@167
|
1724 E T84, T8q, T8l, T8r, T8b, T8v, T8i, T8u;
|
cannam@167
|
1725 {
|
cannam@167
|
1726 E T82, T83, T8j, T8k;
|
cannam@167
|
1727 T82 = T7b + T7c;
|
cannam@167
|
1728 T83 = T7F + T7G;
|
cannam@167
|
1729 T84 = T82 - T83;
|
cannam@167
|
1730 T8q = T82 + T83;
|
cannam@167
|
1731 T8j = FMA(KP195090322, T86, KP980785280 * T85);
|
cannam@167
|
1732 T8k = FMA(KP195090322, T89, KP980785280 * T88);
|
cannam@167
|
1733 T8l = T8j - T8k;
|
cannam@167
|
1734 T8r = T8j + T8k;
|
cannam@167
|
1735 }
|
cannam@167
|
1736 {
|
cannam@167
|
1737 E T87, T8a, T8g, T8h;
|
cannam@167
|
1738 T87 = FNMS(KP980785280, T86, KP195090322 * T85);
|
cannam@167
|
1739 T8a = FNMS(KP980785280, T89, KP195090322 * T88);
|
cannam@167
|
1740 T8b = T87 + T8a;
|
cannam@167
|
1741 T8v = T87 - T8a;
|
cannam@167
|
1742 T8g = T7C - T7D;
|
cannam@167
|
1743 T8h = T7g - T7j;
|
cannam@167
|
1744 T8i = T8g + T8h;
|
cannam@167
|
1745 T8u = T8g - T8h;
|
cannam@167
|
1746 }
|
cannam@167
|
1747 {
|
cannam@167
|
1748 E T8c, T8m, T8y, T8A;
|
cannam@167
|
1749 T8c = T84 - T8b;
|
cannam@167
|
1750 T8m = T8i - T8l;
|
cannam@167
|
1751 Ip[WS(rs, 11)] = FNMS(T8f, T8m, T81 * T8c);
|
cannam@167
|
1752 Im[WS(rs, 11)] = FMA(T8f, T8c, T81 * T8m);
|
cannam@167
|
1753 T8y = T8q + T8r;
|
cannam@167
|
1754 T8A = T8u - T8v;
|
cannam@167
|
1755 Ip[WS(rs, 15)] = FNMS(T8z, T8A, T8x * T8y);
|
cannam@167
|
1756 Im[WS(rs, 15)] = FMA(T8x, T8A, T8z * T8y);
|
cannam@167
|
1757 }
|
cannam@167
|
1758 {
|
cannam@167
|
1759 E T8n, T8o, T8s, T8w;
|
cannam@167
|
1760 T8n = T84 + T8b;
|
cannam@167
|
1761 T8o = T8i + T8l;
|
cannam@167
|
1762 Ip[WS(rs, 3)] = FNMS(T1j, T8o, T1f * T8n);
|
cannam@167
|
1763 Im[WS(rs, 3)] = FMA(T1j, T8n, T1f * T8o);
|
cannam@167
|
1764 T8s = T8q - T8r;
|
cannam@167
|
1765 T8w = T8u + T8v;
|
cannam@167
|
1766 Ip[WS(rs, 7)] = FNMS(T8t, T8w, T8p * T8s);
|
cannam@167
|
1767 Im[WS(rs, 7)] = FMA(T8p, T8w, T8t * T8s);
|
cannam@167
|
1768 }
|
cannam@167
|
1769 }
|
cannam@167
|
1770 {
|
cannam@167
|
1771 E T5u, T6u, T6n, T6v, T67, T6B, T6k, T6A;
|
cannam@167
|
1772 {
|
cannam@167
|
1773 E T5e, T5t, T6l, T6m;
|
cannam@167
|
1774 T5e = T56 - T5d;
|
cannam@167
|
1775 T5t = T5l - T5s;
|
cannam@167
|
1776 T5u = T5e + T5t;
|
cannam@167
|
1777 T6u = T5e - T5t;
|
cannam@167
|
1778 T6l = FMA(KP831469612, T5F, KP555570233 * T5M);
|
cannam@167
|
1779 T6m = FNMS(KP831469612, T5Y, KP555570233 * T65);
|
cannam@167
|
1780 T6n = T6l + T6m;
|
cannam@167
|
1781 T6v = T6m - T6l;
|
cannam@167
|
1782 }
|
cannam@167
|
1783 {
|
cannam@167
|
1784 E T5N, T66, T6g, T6j;
|
cannam@167
|
1785 T5N = FNMS(KP831469612, T5M, KP555570233 * T5F);
|
cannam@167
|
1786 T66 = FMA(KP555570233, T5Y, KP831469612 * T65);
|
cannam@167
|
1787 T67 = T5N + T66;
|
cannam@167
|
1788 T6B = T5N - T66;
|
cannam@167
|
1789 T6g = T6c - T6f;
|
cannam@167
|
1790 T6j = T6h - T6i;
|
cannam@167
|
1791 T6k = T6g + T6j;
|
cannam@167
|
1792 T6A = T6g - T6j;
|
cannam@167
|
1793 }
|
cannam@167
|
1794 {
|
cannam@167
|
1795 E T68, T6o, T6D, T6E;
|
cannam@167
|
1796 T68 = T5u - T67;
|
cannam@167
|
1797 T6o = T6k - T6n;
|
cannam@167
|
1798 Ip[WS(rs, 10)] = FNMS(T69, T6o, T53 * T68);
|
cannam@167
|
1799 Im[WS(rs, 10)] = FMA(T69, T68, T53 * T6o);
|
cannam@167
|
1800 T6D = T6u + T6v;
|
cannam@167
|
1801 T6E = T6A + T6B;
|
cannam@167
|
1802 Ip[WS(rs, 6)] = FNMS(T4c, T6E, T49 * T6D);
|
cannam@167
|
1803 Im[WS(rs, 6)] = FMA(T49, T6E, T4c * T6D);
|
cannam@167
|
1804 }
|
cannam@167
|
1805 {
|
cannam@167
|
1806 E T6p, T6q, T6w, T6C;
|
cannam@167
|
1807 T6p = T5u + T67;
|
cannam@167
|
1808 T6q = T6k + T6n;
|
cannam@167
|
1809 Ip[WS(rs, 2)] = FNMS(T4i, T6q, T4h * T6p);
|
cannam@167
|
1810 Im[WS(rs, 2)] = FMA(T4i, T6p, T4h * T6q);
|
cannam@167
|
1811 T6w = T6u - T6v;
|
cannam@167
|
1812 T6C = T6A - T6B;
|
cannam@167
|
1813 Ip[WS(rs, 14)] = FNMS(T6z, T6C, T6t * T6w);
|
cannam@167
|
1814 Im[WS(rs, 14)] = FMA(T6t, T6C, T6z * T6w);
|
cannam@167
|
1815 }
|
cannam@167
|
1816 }
|
cannam@167
|
1817 {
|
cannam@167
|
1818 E T7l, T7Q, T7L, T7R, T7A, T7V, T7I, T7U;
|
cannam@167
|
1819 {
|
cannam@167
|
1820 E T7d, T7k, T7J, T7K;
|
cannam@167
|
1821 T7d = T7b - T7c;
|
cannam@167
|
1822 T7k = T7g + T7j;
|
cannam@167
|
1823 T7l = T7d - T7k;
|
cannam@167
|
1824 T7Q = T7d + T7k;
|
cannam@167
|
1825 T7J = FNMS(KP555570233, T7v, KP831469612 * T7y);
|
cannam@167
|
1826 T7K = FMA(KP555570233, T7o, KP831469612 * T7r);
|
cannam@167
|
1827 T7L = T7J - T7K;
|
cannam@167
|
1828 T7R = T7J + T7K;
|
cannam@167
|
1829 }
|
cannam@167
|
1830 {
|
cannam@167
|
1831 E T7s, T7z, T7E, T7H;
|
cannam@167
|
1832 T7s = FNMS(KP555570233, T7r, KP831469612 * T7o);
|
cannam@167
|
1833 T7z = FMA(KP831469612, T7v, KP555570233 * T7y);
|
cannam@167
|
1834 T7A = T7s - T7z;
|
cannam@167
|
1835 T7V = T7z + T7s;
|
cannam@167
|
1836 T7E = T7C + T7D;
|
cannam@167
|
1837 T7H = T7F - T7G;
|
cannam@167
|
1838 T7I = T7E - T7H;
|
cannam@167
|
1839 T7U = T7E + T7H;
|
cannam@167
|
1840 }
|
cannam@167
|
1841 {
|
cannam@167
|
1842 E T7B, T7M, T7X, T7Y;
|
cannam@167
|
1843 T7B = T7l - T7A;
|
cannam@167
|
1844 T7M = T7I - T7L;
|
cannam@167
|
1845 Ip[WS(rs, 13)] = FNMS(T1k, T7M, T1g * T7B);
|
cannam@167
|
1846 Im[WS(rs, 13)] = FMA(T1k, T7B, T1g * T7M);
|
cannam@167
|
1847 T7X = T7Q + T7R;
|
cannam@167
|
1848 T7Y = T7U + T7V;
|
cannam@167
|
1849 Ip[WS(rs, 1)] = FNMS(T15, T7Y, T12 * T7X);
|
cannam@167
|
1850 Im[WS(rs, 1)] = FMA(T12, T7Y, T15 * T7X);
|
cannam@167
|
1851 }
|
cannam@167
|
1852 {
|
cannam@167
|
1853 E T7N, T7O, T7S, T7W;
|
cannam@167
|
1854 T7N = T7l + T7A;
|
cannam@167
|
1855 T7O = T7I + T7L;
|
cannam@167
|
1856 Ip[WS(rs, 5)] = FNMS(T2Y, T7O, T2X * T7N);
|
cannam@167
|
1857 Im[WS(rs, 5)] = FMA(T2Y, T7N, T2X * T7O);
|
cannam@167
|
1858 T7S = T7Q - T7R;
|
cannam@167
|
1859 T7W = T7U - T7V;
|
cannam@167
|
1860 Ip[WS(rs, 9)] = FNMS(T7T, T7W, T7P * T7S);
|
cannam@167
|
1861 Im[WS(rs, 9)] = FMA(T7P, T7W, T7T * T7S);
|
cannam@167
|
1862 }
|
cannam@167
|
1863 }
|
cannam@167
|
1864 }
|
cannam@167
|
1865 }
|
cannam@167
|
1866 }
|
cannam@167
|
1867 }
|
cannam@167
|
1868
|
cannam@167
|
1869 static const tw_instr twinstr[] = {
|
cannam@167
|
1870 {TW_CEXP, 1, 1},
|
cannam@167
|
1871 {TW_CEXP, 1, 3},
|
cannam@167
|
1872 {TW_CEXP, 1, 9},
|
cannam@167
|
1873 {TW_CEXP, 1, 27},
|
cannam@167
|
1874 {TW_NEXT, 1, 0}
|
cannam@167
|
1875 };
|
cannam@167
|
1876
|
cannam@167
|
1877 static const hc2c_desc desc = { 32, "hc2cb2_32", twinstr, &GENUS, {376, 168, 112, 0} };
|
cannam@167
|
1878
|
cannam@167
|
1879 void X(codelet_hc2cb2_32) (planner *p) {
|
cannam@167
|
1880 X(khc2c_register) (p, hc2cb2_32, &desc, HC2C_VIA_RDFT);
|
cannam@167
|
1881 }
|
cannam@167
|
1882 #endif
|