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 08:07:33 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_hc2hc.native -fma -compact -variables 4 -pipeline-latency 4 -sign 1 -n 32 -dif -name hb_32 -include rdft/scalar/hb.h */
|
cannam@167
|
29
|
cannam@167
|
30 /*
|
cannam@167
|
31 * This function contains 434 FP additions, 260 FP multiplications,
|
cannam@167
|
32 * (or, 236 additions, 62 multiplications, 198 fused multiply/add),
|
cannam@167
|
33 * 102 stack variables, 7 constants, and 128 memory accesses
|
cannam@167
|
34 */
|
cannam@167
|
35 #include "rdft/scalar/hb.h"
|
cannam@167
|
36
|
cannam@167
|
37 static void hb_32(R *cr, R *ci, 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) * 62); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 62, MAKE_VOLATILE_STRIDE(64, rs)) {
|
cannam@167
|
49 E Tf, T5K, T7k, T8k, T7N, T8x, T1i, T3i, T2L, T3v, T4v, T5f, T6m, T6T, T42;
|
cannam@167
|
50 E T52, TZ, T6X, T1X, T3p, T8p, T8B, T26, T3o, T4n, T58, T7z, T7T, T4k, T59;
|
cannam@167
|
51 E T6a, T6p, TK, T6W, T2o, T3m, T8s, T8A, T2x, T3l, T4g, T55, T7G, T7S, T4d;
|
cannam@167
|
52 E T56, T61, T6o, Tu, T6f, T7r, T8y, T7Q, T8l, T1F, T3w, T2O, T3j, T4y, T53;
|
cannam@167
|
53 E T5R, T6U, T49, T5g;
|
cannam@167
|
54 {
|
cannam@167
|
55 E T3, T12, T6, T2D, T2G, T6g, T15, T6h, Td, T6k, T1g, T2J, Ta, T6j, T1b;
|
cannam@167
|
56 E T2I;
|
cannam@167
|
57 {
|
cannam@167
|
58 E T1, T2, T13, T14;
|
cannam@167
|
59 T1 = cr[0];
|
cannam@167
|
60 T2 = ci[WS(rs, 15)];
|
cannam@167
|
61 T3 = T1 + T2;
|
cannam@167
|
62 T12 = T1 - T2;
|
cannam@167
|
63 {
|
cannam@167
|
64 E T4, T5, T2E, T2F;
|
cannam@167
|
65 T4 = cr[WS(rs, 8)];
|
cannam@167
|
66 T5 = ci[WS(rs, 7)];
|
cannam@167
|
67 T6 = T4 + T5;
|
cannam@167
|
68 T2D = T4 - T5;
|
cannam@167
|
69 T2E = ci[WS(rs, 31)];
|
cannam@167
|
70 T2F = cr[WS(rs, 16)];
|
cannam@167
|
71 T2G = T2E + T2F;
|
cannam@167
|
72 T6g = T2E - T2F;
|
cannam@167
|
73 }
|
cannam@167
|
74 T13 = ci[WS(rs, 23)];
|
cannam@167
|
75 T14 = cr[WS(rs, 24)];
|
cannam@167
|
76 T15 = T13 + T14;
|
cannam@167
|
77 T6h = T13 - T14;
|
cannam@167
|
78 {
|
cannam@167
|
79 E Tb, Tc, T1c, T1d, T1e, T1f;
|
cannam@167
|
80 Tb = ci[WS(rs, 3)];
|
cannam@167
|
81 Tc = cr[WS(rs, 12)];
|
cannam@167
|
82 T1c = Tb - Tc;
|
cannam@167
|
83 T1d = ci[WS(rs, 19)];
|
cannam@167
|
84 T1e = cr[WS(rs, 28)];
|
cannam@167
|
85 T1f = T1d + T1e;
|
cannam@167
|
86 Td = Tb + Tc;
|
cannam@167
|
87 T6k = T1d - T1e;
|
cannam@167
|
88 T1g = T1c - T1f;
|
cannam@167
|
89 T2J = T1c + T1f;
|
cannam@167
|
90 }
|
cannam@167
|
91 {
|
cannam@167
|
92 E T8, T9, T17, T18, T19, T1a;
|
cannam@167
|
93 T8 = cr[WS(rs, 4)];
|
cannam@167
|
94 T9 = ci[WS(rs, 11)];
|
cannam@167
|
95 T17 = T8 - T9;
|
cannam@167
|
96 T18 = ci[WS(rs, 27)];
|
cannam@167
|
97 T19 = cr[WS(rs, 20)];
|
cannam@167
|
98 T1a = T18 + T19;
|
cannam@167
|
99 Ta = T8 + T9;
|
cannam@167
|
100 T6j = T18 - T19;
|
cannam@167
|
101 T1b = T17 - T1a;
|
cannam@167
|
102 T2I = T17 + T1a;
|
cannam@167
|
103 }
|
cannam@167
|
104 }
|
cannam@167
|
105 {
|
cannam@167
|
106 E T7, Te, T7i, T7j;
|
cannam@167
|
107 T7 = T3 + T6;
|
cannam@167
|
108 Te = Ta + Td;
|
cannam@167
|
109 Tf = T7 + Te;
|
cannam@167
|
110 T5K = T7 - Te;
|
cannam@167
|
111 T7i = T3 - T6;
|
cannam@167
|
112 T7j = T6k - T6j;
|
cannam@167
|
113 T7k = T7i - T7j;
|
cannam@167
|
114 T8k = T7i + T7j;
|
cannam@167
|
115 }
|
cannam@167
|
116 {
|
cannam@167
|
117 E T7L, T7M, T16, T1h;
|
cannam@167
|
118 T7L = T6g - T6h;
|
cannam@167
|
119 T7M = Ta - Td;
|
cannam@167
|
120 T7N = T7L - T7M;
|
cannam@167
|
121 T8x = T7M + T7L;
|
cannam@167
|
122 T16 = T12 - T15;
|
cannam@167
|
123 T1h = T1b + T1g;
|
cannam@167
|
124 T1i = FNMS(KP707106781, T1h, T16);
|
cannam@167
|
125 T3i = FMA(KP707106781, T1h, T16);
|
cannam@167
|
126 }
|
cannam@167
|
127 {
|
cannam@167
|
128 E T2H, T2K, T4t, T4u;
|
cannam@167
|
129 T2H = T2D + T2G;
|
cannam@167
|
130 T2K = T2I - T2J;
|
cannam@167
|
131 T2L = FNMS(KP707106781, T2K, T2H);
|
cannam@167
|
132 T3v = FMA(KP707106781, T2K, T2H);
|
cannam@167
|
133 T4t = T2G - T2D;
|
cannam@167
|
134 T4u = T1b - T1g;
|
cannam@167
|
135 T4v = FMA(KP707106781, T4u, T4t);
|
cannam@167
|
136 T5f = FNMS(KP707106781, T4u, T4t);
|
cannam@167
|
137 }
|
cannam@167
|
138 {
|
cannam@167
|
139 E T6i, T6l, T40, T41;
|
cannam@167
|
140 T6i = T6g + T6h;
|
cannam@167
|
141 T6l = T6j + T6k;
|
cannam@167
|
142 T6m = T6i - T6l;
|
cannam@167
|
143 T6T = T6i + T6l;
|
cannam@167
|
144 T40 = T12 + T15;
|
cannam@167
|
145 T41 = T2I + T2J;
|
cannam@167
|
146 T42 = FNMS(KP707106781, T41, T40);
|
cannam@167
|
147 T52 = FMA(KP707106781, T41, T40);
|
cannam@167
|
148 }
|
cannam@167
|
149 }
|
cannam@167
|
150 {
|
cannam@167
|
151 E TR, T7w, T1H, T1Y, T1K, T7t, T21, T65, TY, T7u, T7x, T1Q, T1V, T24, T68;
|
cannam@167
|
152 E T23, T7v, T7y;
|
cannam@167
|
153 {
|
cannam@167
|
154 E TL, TM, TN, TO, TP, TQ;
|
cannam@167
|
155 TL = ci[0];
|
cannam@167
|
156 TM = cr[WS(rs, 15)];
|
cannam@167
|
157 TN = TL + TM;
|
cannam@167
|
158 TO = cr[WS(rs, 7)];
|
cannam@167
|
159 TP = ci[WS(rs, 8)];
|
cannam@167
|
160 TQ = TO + TP;
|
cannam@167
|
161 TR = TN + TQ;
|
cannam@167
|
162 T7w = TN - TQ;
|
cannam@167
|
163 T1H = TO - TP;
|
cannam@167
|
164 T1Y = TL - TM;
|
cannam@167
|
165 }
|
cannam@167
|
166 {
|
cannam@167
|
167 E T1I, T1J, T63, T1Z, T20, T64;
|
cannam@167
|
168 T1I = ci[WS(rs, 16)];
|
cannam@167
|
169 T1J = cr[WS(rs, 31)];
|
cannam@167
|
170 T63 = T1I - T1J;
|
cannam@167
|
171 T1Z = ci[WS(rs, 24)];
|
cannam@167
|
172 T20 = cr[WS(rs, 23)];
|
cannam@167
|
173 T64 = T1Z - T20;
|
cannam@167
|
174 T1K = T1I + T1J;
|
cannam@167
|
175 T7t = T63 - T64;
|
cannam@167
|
176 T21 = T1Z + T20;
|
cannam@167
|
177 T65 = T63 + T64;
|
cannam@167
|
178 }
|
cannam@167
|
179 {
|
cannam@167
|
180 E TU, T1M, T1U, T67, TX, T1R, T1P, T66;
|
cannam@167
|
181 {
|
cannam@167
|
182 E TS, TT, T1S, T1T;
|
cannam@167
|
183 TS = cr[WS(rs, 3)];
|
cannam@167
|
184 TT = ci[WS(rs, 12)];
|
cannam@167
|
185 TU = TS + TT;
|
cannam@167
|
186 T1M = TS - TT;
|
cannam@167
|
187 T1S = ci[WS(rs, 20)];
|
cannam@167
|
188 T1T = cr[WS(rs, 27)];
|
cannam@167
|
189 T1U = T1S + T1T;
|
cannam@167
|
190 T67 = T1S - T1T;
|
cannam@167
|
191 }
|
cannam@167
|
192 {
|
cannam@167
|
193 E TV, TW, T1N, T1O;
|
cannam@167
|
194 TV = ci[WS(rs, 4)];
|
cannam@167
|
195 TW = cr[WS(rs, 11)];
|
cannam@167
|
196 TX = TV + TW;
|
cannam@167
|
197 T1R = TV - TW;
|
cannam@167
|
198 T1N = ci[WS(rs, 28)];
|
cannam@167
|
199 T1O = cr[WS(rs, 19)];
|
cannam@167
|
200 T1P = T1N + T1O;
|
cannam@167
|
201 T66 = T1N - T1O;
|
cannam@167
|
202 }
|
cannam@167
|
203 TY = TU + TX;
|
cannam@167
|
204 T7u = TU - TX;
|
cannam@167
|
205 T7x = T67 - T66;
|
cannam@167
|
206 T1Q = T1M + T1P;
|
cannam@167
|
207 T1V = T1R + T1U;
|
cannam@167
|
208 T24 = T1R - T1U;
|
cannam@167
|
209 T68 = T66 + T67;
|
cannam@167
|
210 T23 = T1M - T1P;
|
cannam@167
|
211 }
|
cannam@167
|
212 TZ = TR + TY;
|
cannam@167
|
213 T6X = T65 + T68;
|
cannam@167
|
214 {
|
cannam@167
|
215 E T1L, T1W, T8n, T8o;
|
cannam@167
|
216 T1L = T1H - T1K;
|
cannam@167
|
217 T1W = T1Q - T1V;
|
cannam@167
|
218 T1X = FNMS(KP707106781, T1W, T1L);
|
cannam@167
|
219 T3p = FMA(KP707106781, T1W, T1L);
|
cannam@167
|
220 T8n = T7u + T7t;
|
cannam@167
|
221 T8o = T7w + T7x;
|
cannam@167
|
222 T8p = FNMS(KP414213562, T8o, T8n);
|
cannam@167
|
223 T8B = FMA(KP414213562, T8n, T8o);
|
cannam@167
|
224 }
|
cannam@167
|
225 {
|
cannam@167
|
226 E T22, T25, T4l, T4m;
|
cannam@167
|
227 T22 = T1Y - T21;
|
cannam@167
|
228 T25 = T23 + T24;
|
cannam@167
|
229 T26 = FNMS(KP707106781, T25, T22);
|
cannam@167
|
230 T3o = FMA(KP707106781, T25, T22);
|
cannam@167
|
231 T4l = T1H + T1K;
|
cannam@167
|
232 T4m = T23 - T24;
|
cannam@167
|
233 T4n = FNMS(KP707106781, T4m, T4l);
|
cannam@167
|
234 T58 = FMA(KP707106781, T4m, T4l);
|
cannam@167
|
235 }
|
cannam@167
|
236 T7v = T7t - T7u;
|
cannam@167
|
237 T7y = T7w - T7x;
|
cannam@167
|
238 T7z = FMA(KP414213562, T7y, T7v);
|
cannam@167
|
239 T7T = FNMS(KP414213562, T7v, T7y);
|
cannam@167
|
240 {
|
cannam@167
|
241 E T4i, T4j, T62, T69;
|
cannam@167
|
242 T4i = T1Y + T21;
|
cannam@167
|
243 T4j = T1Q + T1V;
|
cannam@167
|
244 T4k = FNMS(KP707106781, T4j, T4i);
|
cannam@167
|
245 T59 = FMA(KP707106781, T4j, T4i);
|
cannam@167
|
246 T62 = TR - TY;
|
cannam@167
|
247 T69 = T65 - T68;
|
cannam@167
|
248 T6a = T62 + T69;
|
cannam@167
|
249 T6p = T69 - T62;
|
cannam@167
|
250 }
|
cannam@167
|
251 }
|
cannam@167
|
252 {
|
cannam@167
|
253 E TC, T7D, T28, T2p, T2b, T7A, T2s, T5W, TJ, T7B, T7E, T2h, T2m, T2v, T5Z;
|
cannam@167
|
254 E T2u, T7C, T7F;
|
cannam@167
|
255 {
|
cannam@167
|
256 E Tw, Tx, Ty, Tz, TA, TB;
|
cannam@167
|
257 Tw = cr[WS(rs, 1)];
|
cannam@167
|
258 Tx = ci[WS(rs, 14)];
|
cannam@167
|
259 Ty = Tw + Tx;
|
cannam@167
|
260 Tz = cr[WS(rs, 9)];
|
cannam@167
|
261 TA = ci[WS(rs, 6)];
|
cannam@167
|
262 TB = Tz + TA;
|
cannam@167
|
263 TC = Ty + TB;
|
cannam@167
|
264 T7D = Ty - TB;
|
cannam@167
|
265 T28 = Tz - TA;
|
cannam@167
|
266 T2p = Tw - Tx;
|
cannam@167
|
267 }
|
cannam@167
|
268 {
|
cannam@167
|
269 E T29, T2a, T5U, T2q, T2r, T5V;
|
cannam@167
|
270 T29 = ci[WS(rs, 30)];
|
cannam@167
|
271 T2a = cr[WS(rs, 17)];
|
cannam@167
|
272 T5U = T29 - T2a;
|
cannam@167
|
273 T2q = ci[WS(rs, 22)];
|
cannam@167
|
274 T2r = cr[WS(rs, 25)];
|
cannam@167
|
275 T5V = T2q - T2r;
|
cannam@167
|
276 T2b = T29 + T2a;
|
cannam@167
|
277 T7A = T5U - T5V;
|
cannam@167
|
278 T2s = T2q + T2r;
|
cannam@167
|
279 T5W = T5U + T5V;
|
cannam@167
|
280 }
|
cannam@167
|
281 {
|
cannam@167
|
282 E TF, T2d, T2l, T5Y, TI, T2i, T2g, T5X;
|
cannam@167
|
283 {
|
cannam@167
|
284 E TD, TE, T2j, T2k;
|
cannam@167
|
285 TD = cr[WS(rs, 5)];
|
cannam@167
|
286 TE = ci[WS(rs, 10)];
|
cannam@167
|
287 TF = TD + TE;
|
cannam@167
|
288 T2d = TD - TE;
|
cannam@167
|
289 T2j = ci[WS(rs, 18)];
|
cannam@167
|
290 T2k = cr[WS(rs, 29)];
|
cannam@167
|
291 T2l = T2j + T2k;
|
cannam@167
|
292 T5Y = T2j - T2k;
|
cannam@167
|
293 }
|
cannam@167
|
294 {
|
cannam@167
|
295 E TG, TH, T2e, T2f;
|
cannam@167
|
296 TG = ci[WS(rs, 2)];
|
cannam@167
|
297 TH = cr[WS(rs, 13)];
|
cannam@167
|
298 TI = TG + TH;
|
cannam@167
|
299 T2i = TG - TH;
|
cannam@167
|
300 T2e = ci[WS(rs, 26)];
|
cannam@167
|
301 T2f = cr[WS(rs, 21)];
|
cannam@167
|
302 T2g = T2e + T2f;
|
cannam@167
|
303 T5X = T2e - T2f;
|
cannam@167
|
304 }
|
cannam@167
|
305 TJ = TF + TI;
|
cannam@167
|
306 T7B = TF - TI;
|
cannam@167
|
307 T7E = T5Y - T5X;
|
cannam@167
|
308 T2h = T2d + T2g;
|
cannam@167
|
309 T2m = T2i + T2l;
|
cannam@167
|
310 T2v = T2i - T2l;
|
cannam@167
|
311 T5Z = T5X + T5Y;
|
cannam@167
|
312 T2u = T2d - T2g;
|
cannam@167
|
313 }
|
cannam@167
|
314 TK = TC + TJ;
|
cannam@167
|
315 T6W = T5W + T5Z;
|
cannam@167
|
316 {
|
cannam@167
|
317 E T2c, T2n, T8q, T8r;
|
cannam@167
|
318 T2c = T28 + T2b;
|
cannam@167
|
319 T2n = T2h - T2m;
|
cannam@167
|
320 T2o = FNMS(KP707106781, T2n, T2c);
|
cannam@167
|
321 T3m = FMA(KP707106781, T2n, T2c);
|
cannam@167
|
322 T8q = T7B + T7A;
|
cannam@167
|
323 T8r = T7D + T7E;
|
cannam@167
|
324 T8s = FMA(KP414213562, T8r, T8q);
|
cannam@167
|
325 T8A = FNMS(KP414213562, T8q, T8r);
|
cannam@167
|
326 }
|
cannam@167
|
327 {
|
cannam@167
|
328 E T2t, T2w, T4e, T4f;
|
cannam@167
|
329 T2t = T2p - T2s;
|
cannam@167
|
330 T2w = T2u + T2v;
|
cannam@167
|
331 T2x = FNMS(KP707106781, T2w, T2t);
|
cannam@167
|
332 T3l = FMA(KP707106781, T2w, T2t);
|
cannam@167
|
333 T4e = T2b - T28;
|
cannam@167
|
334 T4f = T2v - T2u;
|
cannam@167
|
335 T4g = FNMS(KP707106781, T4f, T4e);
|
cannam@167
|
336 T55 = FMA(KP707106781, T4f, T4e);
|
cannam@167
|
337 }
|
cannam@167
|
338 T7C = T7A - T7B;
|
cannam@167
|
339 T7F = T7D - T7E;
|
cannam@167
|
340 T7G = FNMS(KP414213562, T7F, T7C);
|
cannam@167
|
341 T7S = FMA(KP414213562, T7C, T7F);
|
cannam@167
|
342 {
|
cannam@167
|
343 E T4b, T4c, T5T, T60;
|
cannam@167
|
344 T4b = T2p + T2s;
|
cannam@167
|
345 T4c = T2h + T2m;
|
cannam@167
|
346 T4d = FNMS(KP707106781, T4c, T4b);
|
cannam@167
|
347 T56 = FMA(KP707106781, T4c, T4b);
|
cannam@167
|
348 T5T = TC - TJ;
|
cannam@167
|
349 T60 = T5W - T5Z;
|
cannam@167
|
350 T61 = T5T - T60;
|
cannam@167
|
351 T6o = T5T + T60;
|
cannam@167
|
352 }
|
cannam@167
|
353 }
|
cannam@167
|
354 {
|
cannam@167
|
355 E Ti, T5P, Tl, T5O, T1y, T1D, T7p, T7o, T44, T43, Tp, T5M, Ts, T5L, T1n;
|
cannam@167
|
356 E T1s, T7m, T7l, T47, T46;
|
cannam@167
|
357 {
|
cannam@167
|
358 E T1z, T1C, T1u, T1x;
|
cannam@167
|
359 {
|
cannam@167
|
360 E Tg, Th, T1A, T1B;
|
cannam@167
|
361 Tg = cr[WS(rs, 2)];
|
cannam@167
|
362 Th = ci[WS(rs, 13)];
|
cannam@167
|
363 Ti = Tg + Th;
|
cannam@167
|
364 T1z = Tg - Th;
|
cannam@167
|
365 T1A = ci[WS(rs, 21)];
|
cannam@167
|
366 T1B = cr[WS(rs, 26)];
|
cannam@167
|
367 T1C = T1A + T1B;
|
cannam@167
|
368 T5P = T1A - T1B;
|
cannam@167
|
369 }
|
cannam@167
|
370 {
|
cannam@167
|
371 E Tj, Tk, T1v, T1w;
|
cannam@167
|
372 Tj = cr[WS(rs, 10)];
|
cannam@167
|
373 Tk = ci[WS(rs, 5)];
|
cannam@167
|
374 Tl = Tj + Tk;
|
cannam@167
|
375 T1u = Tj - Tk;
|
cannam@167
|
376 T1v = ci[WS(rs, 29)];
|
cannam@167
|
377 T1w = cr[WS(rs, 18)];
|
cannam@167
|
378 T1x = T1v + T1w;
|
cannam@167
|
379 T5O = T1v - T1w;
|
cannam@167
|
380 }
|
cannam@167
|
381 T1y = T1u + T1x;
|
cannam@167
|
382 T1D = T1z - T1C;
|
cannam@167
|
383 T7p = T5O - T5P;
|
cannam@167
|
384 T7o = Ti - Tl;
|
cannam@167
|
385 T44 = T1z + T1C;
|
cannam@167
|
386 T43 = T1x - T1u;
|
cannam@167
|
387 }
|
cannam@167
|
388 {
|
cannam@167
|
389 E T1o, T1r, T1j, T1m;
|
cannam@167
|
390 {
|
cannam@167
|
391 E Tn, To, T1p, T1q;
|
cannam@167
|
392 Tn = ci[WS(rs, 1)];
|
cannam@167
|
393 To = cr[WS(rs, 14)];
|
cannam@167
|
394 Tp = Tn + To;
|
cannam@167
|
395 T1o = Tn - To;
|
cannam@167
|
396 T1p = ci[WS(rs, 25)];
|
cannam@167
|
397 T1q = cr[WS(rs, 22)];
|
cannam@167
|
398 T1r = T1p + T1q;
|
cannam@167
|
399 T5M = T1p - T1q;
|
cannam@167
|
400 }
|
cannam@167
|
401 {
|
cannam@167
|
402 E Tq, Tr, T1k, T1l;
|
cannam@167
|
403 Tq = cr[WS(rs, 6)];
|
cannam@167
|
404 Tr = ci[WS(rs, 9)];
|
cannam@167
|
405 Ts = Tq + Tr;
|
cannam@167
|
406 T1j = Tq - Tr;
|
cannam@167
|
407 T1k = ci[WS(rs, 17)];
|
cannam@167
|
408 T1l = cr[WS(rs, 30)];
|
cannam@167
|
409 T1m = T1k + T1l;
|
cannam@167
|
410 T5L = T1k - T1l;
|
cannam@167
|
411 }
|
cannam@167
|
412 T1n = T1j - T1m;
|
cannam@167
|
413 T1s = T1o - T1r;
|
cannam@167
|
414 T7m = Tp - Ts;
|
cannam@167
|
415 T7l = T5L - T5M;
|
cannam@167
|
416 T47 = T1o + T1r;
|
cannam@167
|
417 T46 = T1j + T1m;
|
cannam@167
|
418 }
|
cannam@167
|
419 {
|
cannam@167
|
420 E Tm, Tt, T7n, T7q;
|
cannam@167
|
421 Tm = Ti + Tl;
|
cannam@167
|
422 Tt = Tp + Ts;
|
cannam@167
|
423 Tu = Tm + Tt;
|
cannam@167
|
424 T6f = Tm - Tt;
|
cannam@167
|
425 T7n = T7l - T7m;
|
cannam@167
|
426 T7q = T7o + T7p;
|
cannam@167
|
427 T7r = T7n - T7q;
|
cannam@167
|
428 T8y = T7q + T7n;
|
cannam@167
|
429 }
|
cannam@167
|
430 {
|
cannam@167
|
431 E T7O, T7P, T1t, T1E;
|
cannam@167
|
432 T7O = T7o - T7p;
|
cannam@167
|
433 T7P = T7m + T7l;
|
cannam@167
|
434 T7Q = T7O - T7P;
|
cannam@167
|
435 T8l = T7O + T7P;
|
cannam@167
|
436 T1t = FNMS(KP414213562, T1s, T1n);
|
cannam@167
|
437 T1E = FMA(KP414213562, T1D, T1y);
|
cannam@167
|
438 T1F = T1t - T1E;
|
cannam@167
|
439 T3w = T1E + T1t;
|
cannam@167
|
440 }
|
cannam@167
|
441 {
|
cannam@167
|
442 E T2M, T2N, T4w, T4x;
|
cannam@167
|
443 T2M = FNMS(KP414213562, T1y, T1D);
|
cannam@167
|
444 T2N = FMA(KP414213562, T1n, T1s);
|
cannam@167
|
445 T2O = T2M - T2N;
|
cannam@167
|
446 T3j = T2M + T2N;
|
cannam@167
|
447 T4w = FMA(KP414213562, T43, T44);
|
cannam@167
|
448 T4x = FMA(KP414213562, T46, T47);
|
cannam@167
|
449 T4y = T4w - T4x;
|
cannam@167
|
450 T53 = T4w + T4x;
|
cannam@167
|
451 }
|
cannam@167
|
452 {
|
cannam@167
|
453 E T5N, T5Q, T45, T48;
|
cannam@167
|
454 T5N = T5L + T5M;
|
cannam@167
|
455 T5Q = T5O + T5P;
|
cannam@167
|
456 T5R = T5N - T5Q;
|
cannam@167
|
457 T6U = T5Q + T5N;
|
cannam@167
|
458 T45 = FNMS(KP414213562, T44, T43);
|
cannam@167
|
459 T48 = FNMS(KP414213562, T47, T46);
|
cannam@167
|
460 T49 = T45 + T48;
|
cannam@167
|
461 T5g = T48 - T45;
|
cannam@167
|
462 }
|
cannam@167
|
463 }
|
cannam@167
|
464 {
|
cannam@167
|
465 E Tv, T10, T6Q, T6V, T6Y, T6Z;
|
cannam@167
|
466 Tv = Tf + Tu;
|
cannam@167
|
467 T10 = TK + TZ;
|
cannam@167
|
468 T6Q = Tv - T10;
|
cannam@167
|
469 T6V = T6T + T6U;
|
cannam@167
|
470 T6Y = T6W + T6X;
|
cannam@167
|
471 T6Z = T6V - T6Y;
|
cannam@167
|
472 cr[0] = Tv + T10;
|
cannam@167
|
473 ci[0] = T6V + T6Y;
|
cannam@167
|
474 {
|
cannam@167
|
475 E T6P, T6R, T6S, T70;
|
cannam@167
|
476 T6P = W[30];
|
cannam@167
|
477 T6R = T6P * T6Q;
|
cannam@167
|
478 T6S = W[31];
|
cannam@167
|
479 T70 = T6S * T6Q;
|
cannam@167
|
480 cr[WS(rs, 16)] = FNMS(T6S, T6Z, T6R);
|
cannam@167
|
481 ci[WS(rs, 16)] = FMA(T6P, T6Z, T70);
|
cannam@167
|
482 }
|
cannam@167
|
483 }
|
cannam@167
|
484 {
|
cannam@167
|
485 E T8O, T8W, T8T, T8Z;
|
cannam@167
|
486 {
|
cannam@167
|
487 E T8M, T8N, T8R, T8S;
|
cannam@167
|
488 T8M = FMA(KP707106781, T8l, T8k);
|
cannam@167
|
489 T8N = T8A + T8B;
|
cannam@167
|
490 T8O = FNMS(KP923879532, T8N, T8M);
|
cannam@167
|
491 T8W = FMA(KP923879532, T8N, T8M);
|
cannam@167
|
492 T8R = FMA(KP707106781, T8y, T8x);
|
cannam@167
|
493 T8S = T8s + T8p;
|
cannam@167
|
494 T8T = FNMS(KP923879532, T8S, T8R);
|
cannam@167
|
495 T8Z = FMA(KP923879532, T8S, T8R);
|
cannam@167
|
496 }
|
cannam@167
|
497 {
|
cannam@167
|
498 E T8P, T8U, T8L, T8Q;
|
cannam@167
|
499 T8L = W[34];
|
cannam@167
|
500 T8P = T8L * T8O;
|
cannam@167
|
501 T8U = T8L * T8T;
|
cannam@167
|
502 T8Q = W[35];
|
cannam@167
|
503 cr[WS(rs, 18)] = FNMS(T8Q, T8T, T8P);
|
cannam@167
|
504 ci[WS(rs, 18)] = FMA(T8Q, T8O, T8U);
|
cannam@167
|
505 }
|
cannam@167
|
506 {
|
cannam@167
|
507 E T8X, T90, T8V, T8Y;
|
cannam@167
|
508 T8V = W[2];
|
cannam@167
|
509 T8X = T8V * T8W;
|
cannam@167
|
510 T90 = T8V * T8Z;
|
cannam@167
|
511 T8Y = W[3];
|
cannam@167
|
512 cr[WS(rs, 2)] = FNMS(T8Y, T8Z, T8X);
|
cannam@167
|
513 ci[WS(rs, 2)] = FMA(T8Y, T8W, T90);
|
cannam@167
|
514 }
|
cannam@167
|
515 }
|
cannam@167
|
516 {
|
cannam@167
|
517 E T86, T8e, T8b, T8h;
|
cannam@167
|
518 {
|
cannam@167
|
519 E T84, T85, T89, T8a;
|
cannam@167
|
520 T84 = FNMS(KP707106781, T7r, T7k);
|
cannam@167
|
521 T85 = T7S + T7T;
|
cannam@167
|
522 T86 = FNMS(KP923879532, T85, T84);
|
cannam@167
|
523 T8e = FMA(KP923879532, T85, T84);
|
cannam@167
|
524 T89 = FNMS(KP707106781, T7Q, T7N);
|
cannam@167
|
525 T8a = T7G + T7z;
|
cannam@167
|
526 T8b = FNMS(KP923879532, T8a, T89);
|
cannam@167
|
527 T8h = FMA(KP923879532, T8a, T89);
|
cannam@167
|
528 }
|
cannam@167
|
529 {
|
cannam@167
|
530 E T87, T8c, T83, T88;
|
cannam@167
|
531 T83 = W[26];
|
cannam@167
|
532 T87 = T83 * T86;
|
cannam@167
|
533 T8c = T83 * T8b;
|
cannam@167
|
534 T88 = W[27];
|
cannam@167
|
535 cr[WS(rs, 14)] = FNMS(T88, T8b, T87);
|
cannam@167
|
536 ci[WS(rs, 14)] = FMA(T88, T86, T8c);
|
cannam@167
|
537 }
|
cannam@167
|
538 {
|
cannam@167
|
539 E T8f, T8i, T8d, T8g;
|
cannam@167
|
540 T8d = W[58];
|
cannam@167
|
541 T8f = T8d * T8e;
|
cannam@167
|
542 T8i = T8d * T8h;
|
cannam@167
|
543 T8g = W[59];
|
cannam@167
|
544 cr[WS(rs, 30)] = FNMS(T8g, T8h, T8f);
|
cannam@167
|
545 ci[WS(rs, 30)] = FMA(T8g, T8e, T8i);
|
cannam@167
|
546 }
|
cannam@167
|
547 }
|
cannam@167
|
548 {
|
cannam@167
|
549 E T6C, T6K, T6H, T6N;
|
cannam@167
|
550 {
|
cannam@167
|
551 E T6A, T6B, T6F, T6G;
|
cannam@167
|
552 T6A = T5K - T5R;
|
cannam@167
|
553 T6B = T6p - T6o;
|
cannam@167
|
554 T6C = FNMS(KP707106781, T6B, T6A);
|
cannam@167
|
555 T6K = FMA(KP707106781, T6B, T6A);
|
cannam@167
|
556 T6F = T6m - T6f;
|
cannam@167
|
557 T6G = T61 - T6a;
|
cannam@167
|
558 T6H = FNMS(KP707106781, T6G, T6F);
|
cannam@167
|
559 T6N = FMA(KP707106781, T6G, T6F);
|
cannam@167
|
560 }
|
cannam@167
|
561 {
|
cannam@167
|
562 E T6D, T6I, T6z, T6E;
|
cannam@167
|
563 T6z = W[54];
|
cannam@167
|
564 T6D = T6z * T6C;
|
cannam@167
|
565 T6I = T6z * T6H;
|
cannam@167
|
566 T6E = W[55];
|
cannam@167
|
567 cr[WS(rs, 28)] = FNMS(T6E, T6H, T6D);
|
cannam@167
|
568 ci[WS(rs, 28)] = FMA(T6E, T6C, T6I);
|
cannam@167
|
569 }
|
cannam@167
|
570 {
|
cannam@167
|
571 E T6L, T6O, T6J, T6M;
|
cannam@167
|
572 T6J = W[22];
|
cannam@167
|
573 T6L = T6J * T6K;
|
cannam@167
|
574 T6O = T6J * T6N;
|
cannam@167
|
575 T6M = W[23];
|
cannam@167
|
576 cr[WS(rs, 12)] = FNMS(T6M, T6N, T6L);
|
cannam@167
|
577 ci[WS(rs, 12)] = FMA(T6M, T6K, T6O);
|
cannam@167
|
578 }
|
cannam@167
|
579 }
|
cannam@167
|
580 {
|
cannam@167
|
581 E T8u, T8G, T8D, T8J;
|
cannam@167
|
582 {
|
cannam@167
|
583 E T8m, T8t, T8z, T8C;
|
cannam@167
|
584 T8m = FNMS(KP707106781, T8l, T8k);
|
cannam@167
|
585 T8t = T8p - T8s;
|
cannam@167
|
586 T8u = FNMS(KP923879532, T8t, T8m);
|
cannam@167
|
587 T8G = FMA(KP923879532, T8t, T8m);
|
cannam@167
|
588 T8z = FNMS(KP707106781, T8y, T8x);
|
cannam@167
|
589 T8C = T8A - T8B;
|
cannam@167
|
590 T8D = FNMS(KP923879532, T8C, T8z);
|
cannam@167
|
591 T8J = FMA(KP923879532, T8C, T8z);
|
cannam@167
|
592 }
|
cannam@167
|
593 {
|
cannam@167
|
594 E T8j, T8v, T8w, T8E;
|
cannam@167
|
595 T8j = W[50];
|
cannam@167
|
596 T8v = T8j * T8u;
|
cannam@167
|
597 T8w = W[51];
|
cannam@167
|
598 T8E = T8w * T8u;
|
cannam@167
|
599 cr[WS(rs, 26)] = FNMS(T8w, T8D, T8v);
|
cannam@167
|
600 ci[WS(rs, 26)] = FMA(T8j, T8D, T8E);
|
cannam@167
|
601 }
|
cannam@167
|
602 {
|
cannam@167
|
603 E T8F, T8H, T8I, T8K;
|
cannam@167
|
604 T8F = W[18];
|
cannam@167
|
605 T8H = T8F * T8G;
|
cannam@167
|
606 T8I = W[19];
|
cannam@167
|
607 T8K = T8I * T8G;
|
cannam@167
|
608 cr[WS(rs, 10)] = FNMS(T8I, T8J, T8H);
|
cannam@167
|
609 ci[WS(rs, 10)] = FMA(T8F, T8J, T8K);
|
cannam@167
|
610 }
|
cannam@167
|
611 }
|
cannam@167
|
612 {
|
cannam@167
|
613 E T6c, T6u, T6r, T6x;
|
cannam@167
|
614 {
|
cannam@167
|
615 E T5S, T6b, T6n, T6q;
|
cannam@167
|
616 T5S = T5K + T5R;
|
cannam@167
|
617 T6b = T61 + T6a;
|
cannam@167
|
618 T6c = FNMS(KP707106781, T6b, T5S);
|
cannam@167
|
619 T6u = FMA(KP707106781, T6b, T5S);
|
cannam@167
|
620 T6n = T6f + T6m;
|
cannam@167
|
621 T6q = T6o + T6p;
|
cannam@167
|
622 T6r = FNMS(KP707106781, T6q, T6n);
|
cannam@167
|
623 T6x = FMA(KP707106781, T6q, T6n);
|
cannam@167
|
624 }
|
cannam@167
|
625 {
|
cannam@167
|
626 E T5J, T6d, T6e, T6s;
|
cannam@167
|
627 T5J = W[38];
|
cannam@167
|
628 T6d = T5J * T6c;
|
cannam@167
|
629 T6e = W[39];
|
cannam@167
|
630 T6s = T6e * T6c;
|
cannam@167
|
631 cr[WS(rs, 20)] = FNMS(T6e, T6r, T6d);
|
cannam@167
|
632 ci[WS(rs, 20)] = FMA(T5J, T6r, T6s);
|
cannam@167
|
633 }
|
cannam@167
|
634 {
|
cannam@167
|
635 E T6t, T6v, T6w, T6y;
|
cannam@167
|
636 T6t = W[6];
|
cannam@167
|
637 T6v = T6t * T6u;
|
cannam@167
|
638 T6w = W[7];
|
cannam@167
|
639 T6y = T6w * T6u;
|
cannam@167
|
640 cr[WS(rs, 4)] = FNMS(T6w, T6x, T6v);
|
cannam@167
|
641 ci[WS(rs, 4)] = FMA(T6t, T6x, T6y);
|
cannam@167
|
642 }
|
cannam@167
|
643 }
|
cannam@167
|
644 {
|
cannam@167
|
645 E T74, T7c, T79, T7f;
|
cannam@167
|
646 {
|
cannam@167
|
647 E T72, T73, T77, T78;
|
cannam@167
|
648 T72 = Tf - Tu;
|
cannam@167
|
649 T73 = T6X - T6W;
|
cannam@167
|
650 T74 = T72 - T73;
|
cannam@167
|
651 T7c = T72 + T73;
|
cannam@167
|
652 T77 = T6T - T6U;
|
cannam@167
|
653 T78 = TK - TZ;
|
cannam@167
|
654 T79 = T77 - T78;
|
cannam@167
|
655 T7f = T78 + T77;
|
cannam@167
|
656 }
|
cannam@167
|
657 {
|
cannam@167
|
658 E T75, T7a, T71, T76;
|
cannam@167
|
659 T71 = W[46];
|
cannam@167
|
660 T75 = T71 * T74;
|
cannam@167
|
661 T7a = T71 * T79;
|
cannam@167
|
662 T76 = W[47];
|
cannam@167
|
663 cr[WS(rs, 24)] = FNMS(T76, T79, T75);
|
cannam@167
|
664 ci[WS(rs, 24)] = FMA(T76, T74, T7a);
|
cannam@167
|
665 }
|
cannam@167
|
666 {
|
cannam@167
|
667 E T7d, T7g, T7b, T7e;
|
cannam@167
|
668 T7b = W[14];
|
cannam@167
|
669 T7d = T7b * T7c;
|
cannam@167
|
670 T7g = T7b * T7f;
|
cannam@167
|
671 T7e = W[15];
|
cannam@167
|
672 cr[WS(rs, 8)] = FNMS(T7e, T7f, T7d);
|
cannam@167
|
673 ci[WS(rs, 8)] = FMA(T7e, T7c, T7g);
|
cannam@167
|
674 }
|
cannam@167
|
675 }
|
cannam@167
|
676 {
|
cannam@167
|
677 E T7I, T7Y, T7V, T81;
|
cannam@167
|
678 {
|
cannam@167
|
679 E T7s, T7H, T7R, T7U;
|
cannam@167
|
680 T7s = FMA(KP707106781, T7r, T7k);
|
cannam@167
|
681 T7H = T7z - T7G;
|
cannam@167
|
682 T7I = FNMS(KP923879532, T7H, T7s);
|
cannam@167
|
683 T7Y = FMA(KP923879532, T7H, T7s);
|
cannam@167
|
684 T7R = FMA(KP707106781, T7Q, T7N);
|
cannam@167
|
685 T7U = T7S - T7T;
|
cannam@167
|
686 T7V = FNMS(KP923879532, T7U, T7R);
|
cannam@167
|
687 T81 = FMA(KP923879532, T7U, T7R);
|
cannam@167
|
688 }
|
cannam@167
|
689 {
|
cannam@167
|
690 E T7h, T7J, T7K, T7W;
|
cannam@167
|
691 T7h = W[42];
|
cannam@167
|
692 T7J = T7h * T7I;
|
cannam@167
|
693 T7K = W[43];
|
cannam@167
|
694 T7W = T7K * T7I;
|
cannam@167
|
695 cr[WS(rs, 22)] = FNMS(T7K, T7V, T7J);
|
cannam@167
|
696 ci[WS(rs, 22)] = FMA(T7h, T7V, T7W);
|
cannam@167
|
697 }
|
cannam@167
|
698 {
|
cannam@167
|
699 E T7X, T7Z, T80, T82;
|
cannam@167
|
700 T7X = W[10];
|
cannam@167
|
701 T7Z = T7X * T7Y;
|
cannam@167
|
702 T80 = W[11];
|
cannam@167
|
703 T82 = T80 * T7Y;
|
cannam@167
|
704 cr[WS(rs, 6)] = FNMS(T80, T81, T7Z);
|
cannam@167
|
705 ci[WS(rs, 6)] = FMA(T7X, T81, T82);
|
cannam@167
|
706 }
|
cannam@167
|
707 }
|
cannam@167
|
708 {
|
cannam@167
|
709 E T37, T2A, T38, T2W, T2T, T3c, T2Z, T34;
|
cannam@167
|
710 T37 = FNMS(KP923879532, T2O, T2L);
|
cannam@167
|
711 {
|
cannam@167
|
712 E T1G, T27, T2y, T2z;
|
cannam@167
|
713 T1G = FMA(KP923879532, T1F, T1i);
|
cannam@167
|
714 T27 = FMA(KP668178637, T26, T1X);
|
cannam@167
|
715 T2y = FNMS(KP668178637, T2x, T2o);
|
cannam@167
|
716 T2z = T27 - T2y;
|
cannam@167
|
717 T2A = FNMS(KP831469612, T2z, T1G);
|
cannam@167
|
718 T38 = T2y + T27;
|
cannam@167
|
719 T2W = FMA(KP831469612, T2z, T1G);
|
cannam@167
|
720 }
|
cannam@167
|
721 {
|
cannam@167
|
722 E T2P, T32, T2S, T33, T2Q, T2R;
|
cannam@167
|
723 T2P = FMA(KP923879532, T2O, T2L);
|
cannam@167
|
724 T32 = FNMS(KP923879532, T1F, T1i);
|
cannam@167
|
725 T2Q = FMA(KP668178637, T2o, T2x);
|
cannam@167
|
726 T2R = FNMS(KP668178637, T1X, T26);
|
cannam@167
|
727 T2S = T2Q - T2R;
|
cannam@167
|
728 T33 = T2Q + T2R;
|
cannam@167
|
729 T2T = FNMS(KP831469612, T2S, T2P);
|
cannam@167
|
730 T3c = FMA(KP831469612, T33, T32);
|
cannam@167
|
731 T2Z = FMA(KP831469612, T2S, T2P);
|
cannam@167
|
732 T34 = FNMS(KP831469612, T33, T32);
|
cannam@167
|
733 }
|
cannam@167
|
734 {
|
cannam@167
|
735 E T2B, T2U, T11, T2C;
|
cannam@167
|
736 T11 = W[40];
|
cannam@167
|
737 T2B = T11 * T2A;
|
cannam@167
|
738 T2U = T11 * T2T;
|
cannam@167
|
739 T2C = W[41];
|
cannam@167
|
740 cr[WS(rs, 21)] = FNMS(T2C, T2T, T2B);
|
cannam@167
|
741 ci[WS(rs, 21)] = FMA(T2C, T2A, T2U);
|
cannam@167
|
742 }
|
cannam@167
|
743 {
|
cannam@167
|
744 E T2X, T30, T2V, T2Y;
|
cannam@167
|
745 T2V = W[8];
|
cannam@167
|
746 T2X = T2V * T2W;
|
cannam@167
|
747 T30 = T2V * T2Z;
|
cannam@167
|
748 T2Y = W[9];
|
cannam@167
|
749 cr[WS(rs, 5)] = FNMS(T2Y, T2Z, T2X);
|
cannam@167
|
750 ci[WS(rs, 5)] = FMA(T2Y, T2W, T30);
|
cannam@167
|
751 }
|
cannam@167
|
752 {
|
cannam@167
|
753 E T39, T36, T3a, T31, T35;
|
cannam@167
|
754 T39 = FNMS(KP831469612, T38, T37);
|
cannam@167
|
755 T36 = W[25];
|
cannam@167
|
756 T3a = T36 * T34;
|
cannam@167
|
757 T31 = W[24];
|
cannam@167
|
758 T35 = T31 * T34;
|
cannam@167
|
759 cr[WS(rs, 13)] = FNMS(T36, T39, T35);
|
cannam@167
|
760 ci[WS(rs, 13)] = FMA(T31, T39, T3a);
|
cannam@167
|
761 }
|
cannam@167
|
762 {
|
cannam@167
|
763 E T3f, T3e, T3g, T3b, T3d;
|
cannam@167
|
764 T3f = FMA(KP831469612, T38, T37);
|
cannam@167
|
765 T3e = W[57];
|
cannam@167
|
766 T3g = T3e * T3c;
|
cannam@167
|
767 T3b = W[56];
|
cannam@167
|
768 T3d = T3b * T3c;
|
cannam@167
|
769 cr[WS(rs, 29)] = FNMS(T3e, T3f, T3d);
|
cannam@167
|
770 ci[WS(rs, 29)] = FMA(T3b, T3f, T3g);
|
cannam@167
|
771 }
|
cannam@167
|
772 }
|
cannam@167
|
773 {
|
cannam@167
|
774 E T4z, T4C, T4W, T4O, T4q, T4Z, T4G, T4T;
|
cannam@167
|
775 T4z = FMA(KP923879532, T4y, T4v);
|
cannam@167
|
776 {
|
cannam@167
|
777 E T4M, T4A, T4B, T4N;
|
cannam@167
|
778 T4M = FMA(KP923879532, T49, T42);
|
cannam@167
|
779 T4A = FMA(KP668178637, T4d, T4g);
|
cannam@167
|
780 T4B = FMA(KP668178637, T4k, T4n);
|
cannam@167
|
781 T4N = T4A + T4B;
|
cannam@167
|
782 T4C = T4A - T4B;
|
cannam@167
|
783 T4W = FMA(KP831469612, T4N, T4M);
|
cannam@167
|
784 T4O = FNMS(KP831469612, T4N, T4M);
|
cannam@167
|
785 }
|
cannam@167
|
786 {
|
cannam@167
|
787 E T4a, T4R, T4p, T4S, T4h, T4o;
|
cannam@167
|
788 T4a = FNMS(KP923879532, T49, T42);
|
cannam@167
|
789 T4R = FNMS(KP923879532, T4y, T4v);
|
cannam@167
|
790 T4h = FNMS(KP668178637, T4g, T4d);
|
cannam@167
|
791 T4o = FNMS(KP668178637, T4n, T4k);
|
cannam@167
|
792 T4p = T4h + T4o;
|
cannam@167
|
793 T4S = T4h - T4o;
|
cannam@167
|
794 T4q = FNMS(KP831469612, T4p, T4a);
|
cannam@167
|
795 T4Z = FNMS(KP831469612, T4S, T4R);
|
cannam@167
|
796 T4G = FMA(KP831469612, T4p, T4a);
|
cannam@167
|
797 T4T = FMA(KP831469612, T4S, T4R);
|
cannam@167
|
798 }
|
cannam@167
|
799 {
|
cannam@167
|
800 E T4P, T4U, T4L, T4Q;
|
cannam@167
|
801 T4L = W[20];
|
cannam@167
|
802 T4P = T4L * T4O;
|
cannam@167
|
803 T4U = T4L * T4T;
|
cannam@167
|
804 T4Q = W[21];
|
cannam@167
|
805 cr[WS(rs, 11)] = FNMS(T4Q, T4T, T4P);
|
cannam@167
|
806 ci[WS(rs, 11)] = FMA(T4Q, T4O, T4U);
|
cannam@167
|
807 }
|
cannam@167
|
808 {
|
cannam@167
|
809 E T4X, T50, T4V, T4Y;
|
cannam@167
|
810 T4V = W[52];
|
cannam@167
|
811 T4X = T4V * T4W;
|
cannam@167
|
812 T50 = T4V * T4Z;
|
cannam@167
|
813 T4Y = W[53];
|
cannam@167
|
814 cr[WS(rs, 27)] = FNMS(T4Y, T4Z, T4X);
|
cannam@167
|
815 ci[WS(rs, 27)] = FMA(T4Y, T4W, T50);
|
cannam@167
|
816 }
|
cannam@167
|
817 {
|
cannam@167
|
818 E T4D, T4s, T4E, T3Z, T4r;
|
cannam@167
|
819 T4D = FNMS(KP831469612, T4C, T4z);
|
cannam@167
|
820 T4s = W[37];
|
cannam@167
|
821 T4E = T4s * T4q;
|
cannam@167
|
822 T3Z = W[36];
|
cannam@167
|
823 T4r = T3Z * T4q;
|
cannam@167
|
824 cr[WS(rs, 19)] = FNMS(T4s, T4D, T4r);
|
cannam@167
|
825 ci[WS(rs, 19)] = FMA(T3Z, T4D, T4E);
|
cannam@167
|
826 }
|
cannam@167
|
827 {
|
cannam@167
|
828 E T4J, T4I, T4K, T4F, T4H;
|
cannam@167
|
829 T4J = FMA(KP831469612, T4C, T4z);
|
cannam@167
|
830 T4I = W[5];
|
cannam@167
|
831 T4K = T4I * T4G;
|
cannam@167
|
832 T4F = W[4];
|
cannam@167
|
833 T4H = T4F * T4G;
|
cannam@167
|
834 cr[WS(rs, 3)] = FNMS(T4I, T4J, T4H);
|
cannam@167
|
835 ci[WS(rs, 3)] = FMA(T4F, T4J, T4K);
|
cannam@167
|
836 }
|
cannam@167
|
837 }
|
cannam@167
|
838 {
|
cannam@167
|
839 E T3x, T3A, T3U, T3M, T3s, T3X, T3E, T3R;
|
cannam@167
|
840 T3x = FMA(KP923879532, T3w, T3v);
|
cannam@167
|
841 {
|
cannam@167
|
842 E T3K, T3y, T3z, T3L;
|
cannam@167
|
843 T3K = FNMS(KP923879532, T3j, T3i);
|
cannam@167
|
844 T3y = FMA(KP198912367, T3l, T3m);
|
cannam@167
|
845 T3z = FNMS(KP198912367, T3o, T3p);
|
cannam@167
|
846 T3L = T3z - T3y;
|
cannam@167
|
847 T3A = T3y + T3z;
|
cannam@167
|
848 T3U = FMA(KP980785280, T3L, T3K);
|
cannam@167
|
849 T3M = FNMS(KP980785280, T3L, T3K);
|
cannam@167
|
850 }
|
cannam@167
|
851 {
|
cannam@167
|
852 E T3k, T3P, T3r, T3Q, T3n, T3q;
|
cannam@167
|
853 T3k = FMA(KP923879532, T3j, T3i);
|
cannam@167
|
854 T3P = FNMS(KP923879532, T3w, T3v);
|
cannam@167
|
855 T3n = FNMS(KP198912367, T3m, T3l);
|
cannam@167
|
856 T3q = FMA(KP198912367, T3p, T3o);
|
cannam@167
|
857 T3r = T3n + T3q;
|
cannam@167
|
858 T3Q = T3n - T3q;
|
cannam@167
|
859 T3s = FNMS(KP980785280, T3r, T3k);
|
cannam@167
|
860 T3X = FMA(KP980785280, T3Q, T3P);
|
cannam@167
|
861 T3E = FMA(KP980785280, T3r, T3k);
|
cannam@167
|
862 T3R = FNMS(KP980785280, T3Q, T3P);
|
cannam@167
|
863 }
|
cannam@167
|
864 {
|
cannam@167
|
865 E T3N, T3S, T3J, T3O;
|
cannam@167
|
866 T3J = W[48];
|
cannam@167
|
867 T3N = T3J * T3M;
|
cannam@167
|
868 T3S = T3J * T3R;
|
cannam@167
|
869 T3O = W[49];
|
cannam@167
|
870 cr[WS(rs, 25)] = FNMS(T3O, T3R, T3N);
|
cannam@167
|
871 ci[WS(rs, 25)] = FMA(T3O, T3M, T3S);
|
cannam@167
|
872 }
|
cannam@167
|
873 {
|
cannam@167
|
874 E T3V, T3Y, T3T, T3W;
|
cannam@167
|
875 T3T = W[16];
|
cannam@167
|
876 T3V = T3T * T3U;
|
cannam@167
|
877 T3Y = T3T * T3X;
|
cannam@167
|
878 T3W = W[17];
|
cannam@167
|
879 cr[WS(rs, 9)] = FNMS(T3W, T3X, T3V);
|
cannam@167
|
880 ci[WS(rs, 9)] = FMA(T3W, T3U, T3Y);
|
cannam@167
|
881 }
|
cannam@167
|
882 {
|
cannam@167
|
883 E T3B, T3u, T3C, T3h, T3t;
|
cannam@167
|
884 T3B = FNMS(KP980785280, T3A, T3x);
|
cannam@167
|
885 T3u = W[33];
|
cannam@167
|
886 T3C = T3u * T3s;
|
cannam@167
|
887 T3h = W[32];
|
cannam@167
|
888 T3t = T3h * T3s;
|
cannam@167
|
889 cr[WS(rs, 17)] = FNMS(T3u, T3B, T3t);
|
cannam@167
|
890 ci[WS(rs, 17)] = FMA(T3h, T3B, T3C);
|
cannam@167
|
891 }
|
cannam@167
|
892 {
|
cannam@167
|
893 E T3H, T3G, T3I, T3D, T3F;
|
cannam@167
|
894 T3H = FMA(KP980785280, T3A, T3x);
|
cannam@167
|
895 T3G = W[1];
|
cannam@167
|
896 T3I = T3G * T3E;
|
cannam@167
|
897 T3D = W[0];
|
cannam@167
|
898 T3F = T3D * T3E;
|
cannam@167
|
899 cr[WS(rs, 1)] = FNMS(T3G, T3H, T3F);
|
cannam@167
|
900 ci[WS(rs, 1)] = FMA(T3D, T3H, T3I);
|
cannam@167
|
901 }
|
cannam@167
|
902 }
|
cannam@167
|
903 {
|
cannam@167
|
904 E T5h, T5k, T5E, T5w, T5c, T5H, T5o, T5B;
|
cannam@167
|
905 T5h = FMA(KP923879532, T5g, T5f);
|
cannam@167
|
906 {
|
cannam@167
|
907 E T5u, T5i, T5j, T5v;
|
cannam@167
|
908 T5u = FMA(KP923879532, T53, T52);
|
cannam@167
|
909 T5i = FMA(KP198912367, T55, T56);
|
cannam@167
|
910 T5j = FMA(KP198912367, T58, T59);
|
cannam@167
|
911 T5v = T5i + T5j;
|
cannam@167
|
912 T5k = T5i - T5j;
|
cannam@167
|
913 T5E = FMA(KP980785280, T5v, T5u);
|
cannam@167
|
914 T5w = FNMS(KP980785280, T5v, T5u);
|
cannam@167
|
915 }
|
cannam@167
|
916 {
|
cannam@167
|
917 E T54, T5z, T5b, T5A, T57, T5a;
|
cannam@167
|
918 T54 = FNMS(KP923879532, T53, T52);
|
cannam@167
|
919 T5z = FNMS(KP923879532, T5g, T5f);
|
cannam@167
|
920 T57 = FNMS(KP198912367, T56, T55);
|
cannam@167
|
921 T5a = FNMS(KP198912367, T59, T58);
|
cannam@167
|
922 T5b = T57 + T5a;
|
cannam@167
|
923 T5A = T5a - T57;
|
cannam@167
|
924 T5c = FMA(KP980785280, T5b, T54);
|
cannam@167
|
925 T5H = FNMS(KP980785280, T5A, T5z);
|
cannam@167
|
926 T5o = FNMS(KP980785280, T5b, T54);
|
cannam@167
|
927 T5B = FMA(KP980785280, T5A, T5z);
|
cannam@167
|
928 }
|
cannam@167
|
929 {
|
cannam@167
|
930 E T5x, T5C, T5t, T5y;
|
cannam@167
|
931 T5t = W[28];
|
cannam@167
|
932 T5x = T5t * T5w;
|
cannam@167
|
933 T5C = T5t * T5B;
|
cannam@167
|
934 T5y = W[29];
|
cannam@167
|
935 cr[WS(rs, 15)] = FNMS(T5y, T5B, T5x);
|
cannam@167
|
936 ci[WS(rs, 15)] = FMA(T5y, T5w, T5C);
|
cannam@167
|
937 }
|
cannam@167
|
938 {
|
cannam@167
|
939 E T5F, T5I, T5D, T5G;
|
cannam@167
|
940 T5D = W[60];
|
cannam@167
|
941 T5F = T5D * T5E;
|
cannam@167
|
942 T5I = T5D * T5H;
|
cannam@167
|
943 T5G = W[61];
|
cannam@167
|
944 cr[WS(rs, 31)] = FNMS(T5G, T5H, T5F);
|
cannam@167
|
945 ci[WS(rs, 31)] = FMA(T5G, T5E, T5I);
|
cannam@167
|
946 }
|
cannam@167
|
947 {
|
cannam@167
|
948 E T5l, T5e, T5m, T51, T5d;
|
cannam@167
|
949 T5l = FNMS(KP980785280, T5k, T5h);
|
cannam@167
|
950 T5e = W[45];
|
cannam@167
|
951 T5m = T5e * T5c;
|
cannam@167
|
952 T51 = W[44];
|
cannam@167
|
953 T5d = T51 * T5c;
|
cannam@167
|
954 cr[WS(rs, 23)] = FNMS(T5e, T5l, T5d);
|
cannam@167
|
955 ci[WS(rs, 23)] = FMA(T51, T5l, T5m);
|
cannam@167
|
956 }
|
cannam@167
|
957 {
|
cannam@167
|
958 E T5r, T5q, T5s, T5n, T5p;
|
cannam@167
|
959 T5r = FMA(KP980785280, T5k, T5h);
|
cannam@167
|
960 T5q = W[13];
|
cannam@167
|
961 T5s = T5q * T5o;
|
cannam@167
|
962 T5n = W[12];
|
cannam@167
|
963 T5p = T5n * T5o;
|
cannam@167
|
964 cr[WS(rs, 7)] = FNMS(T5q, T5r, T5p);
|
cannam@167
|
965 ci[WS(rs, 7)] = FMA(T5n, T5r, T5s);
|
cannam@167
|
966 }
|
cannam@167
|
967 }
|
cannam@167
|
968 }
|
cannam@167
|
969 }
|
cannam@167
|
970 }
|
cannam@167
|
971
|
cannam@167
|
972 static const tw_instr twinstr[] = {
|
cannam@167
|
973 {TW_FULL, 1, 32},
|
cannam@167
|
974 {TW_NEXT, 1, 0}
|
cannam@167
|
975 };
|
cannam@167
|
976
|
cannam@167
|
977 static const hc2hc_desc desc = { 32, "hb_32", twinstr, &GENUS, {236, 62, 198, 0} };
|
cannam@167
|
978
|
cannam@167
|
979 void X(codelet_hb_32) (planner *p) {
|
cannam@167
|
980 X(khc2hc_register) (p, hb_32, &desc);
|
cannam@167
|
981 }
|
cannam@167
|
982 #else
|
cannam@167
|
983
|
cannam@167
|
984 /* Generated by: ../../../genfft/gen_hc2hc.native -compact -variables 4 -pipeline-latency 4 -sign 1 -n 32 -dif -name hb_32 -include rdft/scalar/hb.h */
|
cannam@167
|
985
|
cannam@167
|
986 /*
|
cannam@167
|
987 * This function contains 434 FP additions, 208 FP multiplications,
|
cannam@167
|
988 * (or, 340 additions, 114 multiplications, 94 fused multiply/add),
|
cannam@167
|
989 * 98 stack variables, 7 constants, and 128 memory accesses
|
cannam@167
|
990 */
|
cannam@167
|
991 #include "rdft/scalar/hb.h"
|
cannam@167
|
992
|
cannam@167
|
993 static void hb_32(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
|
cannam@167
|
994 {
|
cannam@167
|
995 DK(KP555570233, +0.555570233019602224742830813948532874374937191);
|
cannam@167
|
996 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
cannam@167
|
997 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
cannam@167
|
998 DK(KP195090322, +0.195090322016128267848284868477022240927691618);
|
cannam@167
|
999 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
cannam@167
|
1000 DK(KP382683432, +0.382683432365089771728459984030398866761344562);
|
cannam@167
|
1001 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
cannam@167
|
1002 {
|
cannam@167
|
1003 INT m;
|
cannam@167
|
1004 for (m = mb, W = W + ((mb - 1) * 62); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 62, MAKE_VOLATILE_STRIDE(64, rs)) {
|
cannam@167
|
1005 E T4o, T6y, T70, T5u, Tf, T12, T5x, T6z, T3m, T3Y, T29, T2y, T4v, T71, T2U;
|
cannam@167
|
1006 E T3M, Tu, T1U, T6D, T73, T6G, T74, T1h, T2z, T2X, T3o, T4D, T5A, T4K, T5z;
|
cannam@167
|
1007 E T30, T3n, TK, T1j, T6S, T7w, T6V, T7v, T1y, T2B, T3c, T3S, T4X, T61, T54;
|
cannam@167
|
1008 E T62, T3f, T3T, TZ, T1A, T6L, T7z, T6O, T7y, T1P, T2C, T35, T3P, T5g, T64;
|
cannam@167
|
1009 E T5n, T65, T38, T3Q;
|
cannam@167
|
1010 {
|
cannam@167
|
1011 E T3, T4m, T24, T4q, T27, T4t, T6, T5s, Ta, T4p, T1X, T5t, T20, T4n, Td;
|
cannam@167
|
1012 E T4s;
|
cannam@167
|
1013 {
|
cannam@167
|
1014 E T1, T2, T22, T23;
|
cannam@167
|
1015 T1 = cr[0];
|
cannam@167
|
1016 T2 = ci[WS(rs, 15)];
|
cannam@167
|
1017 T3 = T1 + T2;
|
cannam@167
|
1018 T4m = T1 - T2;
|
cannam@167
|
1019 T22 = ci[WS(rs, 27)];
|
cannam@167
|
1020 T23 = cr[WS(rs, 20)];
|
cannam@167
|
1021 T24 = T22 - T23;
|
cannam@167
|
1022 T4q = T22 + T23;
|
cannam@167
|
1023 }
|
cannam@167
|
1024 {
|
cannam@167
|
1025 E T25, T26, T4, T5;
|
cannam@167
|
1026 T25 = ci[WS(rs, 19)];
|
cannam@167
|
1027 T26 = cr[WS(rs, 28)];
|
cannam@167
|
1028 T27 = T25 - T26;
|
cannam@167
|
1029 T4t = T25 + T26;
|
cannam@167
|
1030 T4 = cr[WS(rs, 8)];
|
cannam@167
|
1031 T5 = ci[WS(rs, 7)];
|
cannam@167
|
1032 T6 = T4 + T5;
|
cannam@167
|
1033 T5s = T4 - T5;
|
cannam@167
|
1034 }
|
cannam@167
|
1035 {
|
cannam@167
|
1036 E T8, T9, T1V, T1W;
|
cannam@167
|
1037 T8 = cr[WS(rs, 4)];
|
cannam@167
|
1038 T9 = ci[WS(rs, 11)];
|
cannam@167
|
1039 Ta = T8 + T9;
|
cannam@167
|
1040 T4p = T8 - T9;
|
cannam@167
|
1041 T1V = ci[WS(rs, 31)];
|
cannam@167
|
1042 T1W = cr[WS(rs, 16)];
|
cannam@167
|
1043 T1X = T1V - T1W;
|
cannam@167
|
1044 T5t = T1V + T1W;
|
cannam@167
|
1045 }
|
cannam@167
|
1046 {
|
cannam@167
|
1047 E T1Y, T1Z, Tb, Tc;
|
cannam@167
|
1048 T1Y = ci[WS(rs, 23)];
|
cannam@167
|
1049 T1Z = cr[WS(rs, 24)];
|
cannam@167
|
1050 T20 = T1Y - T1Z;
|
cannam@167
|
1051 T4n = T1Y + T1Z;
|
cannam@167
|
1052 Tb = ci[WS(rs, 3)];
|
cannam@167
|
1053 Tc = cr[WS(rs, 12)];
|
cannam@167
|
1054 Td = Tb + Tc;
|
cannam@167
|
1055 T4s = Tb - Tc;
|
cannam@167
|
1056 }
|
cannam@167
|
1057 {
|
cannam@167
|
1058 E T7, Te, T21, T28;
|
cannam@167
|
1059 T4o = T4m - T4n;
|
cannam@167
|
1060 T6y = T4m + T4n;
|
cannam@167
|
1061 T70 = T5t - T5s;
|
cannam@167
|
1062 T5u = T5s + T5t;
|
cannam@167
|
1063 T7 = T3 + T6;
|
cannam@167
|
1064 Te = Ta + Td;
|
cannam@167
|
1065 Tf = T7 + Te;
|
cannam@167
|
1066 T12 = T7 - Te;
|
cannam@167
|
1067 {
|
cannam@167
|
1068 E T5v, T5w, T3k, T3l;
|
cannam@167
|
1069 T5v = T4p + T4q;
|
cannam@167
|
1070 T5w = T4s + T4t;
|
cannam@167
|
1071 T5x = KP707106781 * (T5v - T5w);
|
cannam@167
|
1072 T6z = KP707106781 * (T5v + T5w);
|
cannam@167
|
1073 T3k = T1X - T20;
|
cannam@167
|
1074 T3l = Ta - Td;
|
cannam@167
|
1075 T3m = T3k - T3l;
|
cannam@167
|
1076 T3Y = T3l + T3k;
|
cannam@167
|
1077 }
|
cannam@167
|
1078 T21 = T1X + T20;
|
cannam@167
|
1079 T28 = T24 + T27;
|
cannam@167
|
1080 T29 = T21 - T28;
|
cannam@167
|
1081 T2y = T21 + T28;
|
cannam@167
|
1082 {
|
cannam@167
|
1083 E T4r, T4u, T2S, T2T;
|
cannam@167
|
1084 T4r = T4p - T4q;
|
cannam@167
|
1085 T4u = T4s - T4t;
|
cannam@167
|
1086 T4v = KP707106781 * (T4r + T4u);
|
cannam@167
|
1087 T71 = KP707106781 * (T4r - T4u);
|
cannam@167
|
1088 T2S = T3 - T6;
|
cannam@167
|
1089 T2T = T27 - T24;
|
cannam@167
|
1090 T2U = T2S - T2T;
|
cannam@167
|
1091 T3M = T2S + T2T;
|
cannam@167
|
1092 }
|
cannam@167
|
1093 }
|
cannam@167
|
1094 }
|
cannam@167
|
1095 {
|
cannam@167
|
1096 E Ti, T4H, T1c, T4F, T1f, T4I, Tl, T4E, Tp, T4A, T15, T4y, T18, T4B, Ts;
|
cannam@167
|
1097 E T4x;
|
cannam@167
|
1098 {
|
cannam@167
|
1099 E Tg, Th, T1a, T1b;
|
cannam@167
|
1100 Tg = cr[WS(rs, 2)];
|
cannam@167
|
1101 Th = ci[WS(rs, 13)];
|
cannam@167
|
1102 Ti = Tg + Th;
|
cannam@167
|
1103 T4H = Tg - Th;
|
cannam@167
|
1104 T1a = ci[WS(rs, 29)];
|
cannam@167
|
1105 T1b = cr[WS(rs, 18)];
|
cannam@167
|
1106 T1c = T1a - T1b;
|
cannam@167
|
1107 T4F = T1a + T1b;
|
cannam@167
|
1108 }
|
cannam@167
|
1109 {
|
cannam@167
|
1110 E T1d, T1e, Tj, Tk;
|
cannam@167
|
1111 T1d = ci[WS(rs, 21)];
|
cannam@167
|
1112 T1e = cr[WS(rs, 26)];
|
cannam@167
|
1113 T1f = T1d - T1e;
|
cannam@167
|
1114 T4I = T1d + T1e;
|
cannam@167
|
1115 Tj = cr[WS(rs, 10)];
|
cannam@167
|
1116 Tk = ci[WS(rs, 5)];
|
cannam@167
|
1117 Tl = Tj + Tk;
|
cannam@167
|
1118 T4E = Tj - Tk;
|
cannam@167
|
1119 }
|
cannam@167
|
1120 {
|
cannam@167
|
1121 E Tn, To, T13, T14;
|
cannam@167
|
1122 Tn = ci[WS(rs, 1)];
|
cannam@167
|
1123 To = cr[WS(rs, 14)];
|
cannam@167
|
1124 Tp = Tn + To;
|
cannam@167
|
1125 T4A = Tn - To;
|
cannam@167
|
1126 T13 = ci[WS(rs, 17)];
|
cannam@167
|
1127 T14 = cr[WS(rs, 30)];
|
cannam@167
|
1128 T15 = T13 - T14;
|
cannam@167
|
1129 T4y = T13 + T14;
|
cannam@167
|
1130 }
|
cannam@167
|
1131 {
|
cannam@167
|
1132 E T16, T17, Tq, Tr;
|
cannam@167
|
1133 T16 = ci[WS(rs, 25)];
|
cannam@167
|
1134 T17 = cr[WS(rs, 22)];
|
cannam@167
|
1135 T18 = T16 - T17;
|
cannam@167
|
1136 T4B = T16 + T17;
|
cannam@167
|
1137 Tq = cr[WS(rs, 6)];
|
cannam@167
|
1138 Tr = ci[WS(rs, 9)];
|
cannam@167
|
1139 Ts = Tq + Tr;
|
cannam@167
|
1140 T4x = Tq - Tr;
|
cannam@167
|
1141 }
|
cannam@167
|
1142 {
|
cannam@167
|
1143 E Tm, Tt, T6B, T6C;
|
cannam@167
|
1144 Tm = Ti + Tl;
|
cannam@167
|
1145 Tt = Tp + Ts;
|
cannam@167
|
1146 Tu = Tm + Tt;
|
cannam@167
|
1147 T1U = Tm - Tt;
|
cannam@167
|
1148 T6B = T4H + T4I;
|
cannam@167
|
1149 T6C = T4F - T4E;
|
cannam@167
|
1150 T6D = FNMS(KP923879532, T6C, KP382683432 * T6B);
|
cannam@167
|
1151 T73 = FMA(KP382683432, T6C, KP923879532 * T6B);
|
cannam@167
|
1152 }
|
cannam@167
|
1153 {
|
cannam@167
|
1154 E T6E, T6F, T19, T1g;
|
cannam@167
|
1155 T6E = T4A + T4B;
|
cannam@167
|
1156 T6F = T4x + T4y;
|
cannam@167
|
1157 T6G = FNMS(KP923879532, T6F, KP382683432 * T6E);
|
cannam@167
|
1158 T74 = FMA(KP382683432, T6F, KP923879532 * T6E);
|
cannam@167
|
1159 T19 = T15 + T18;
|
cannam@167
|
1160 T1g = T1c + T1f;
|
cannam@167
|
1161 T1h = T19 - T1g;
|
cannam@167
|
1162 T2z = T1g + T19;
|
cannam@167
|
1163 }
|
cannam@167
|
1164 {
|
cannam@167
|
1165 E T2V, T2W, T4z, T4C;
|
cannam@167
|
1166 T2V = T15 - T18;
|
cannam@167
|
1167 T2W = Tp - Ts;
|
cannam@167
|
1168 T2X = T2V - T2W;
|
cannam@167
|
1169 T3o = T2W + T2V;
|
cannam@167
|
1170 T4z = T4x - T4y;
|
cannam@167
|
1171 T4C = T4A - T4B;
|
cannam@167
|
1172 T4D = FNMS(KP382683432, T4C, KP923879532 * T4z);
|
cannam@167
|
1173 T5A = FMA(KP382683432, T4z, KP923879532 * T4C);
|
cannam@167
|
1174 }
|
cannam@167
|
1175 {
|
cannam@167
|
1176 E T4G, T4J, T2Y, T2Z;
|
cannam@167
|
1177 T4G = T4E + T4F;
|
cannam@167
|
1178 T4J = T4H - T4I;
|
cannam@167
|
1179 T4K = FMA(KP923879532, T4G, KP382683432 * T4J);
|
cannam@167
|
1180 T5z = FNMS(KP382683432, T4G, KP923879532 * T4J);
|
cannam@167
|
1181 T2Y = Ti - Tl;
|
cannam@167
|
1182 T2Z = T1c - T1f;
|
cannam@167
|
1183 T30 = T2Y + T2Z;
|
cannam@167
|
1184 T3n = T2Y - T2Z;
|
cannam@167
|
1185 }
|
cannam@167
|
1186 }
|
cannam@167
|
1187 {
|
cannam@167
|
1188 E Ty, T4N, TB, T4Y, T1p, T4O, T1m, T4Z, TI, T52, T1w, T4V, TF, T51, T1t;
|
cannam@167
|
1189 E T4S;
|
cannam@167
|
1190 {
|
cannam@167
|
1191 E Tw, Tx, T1k, T1l;
|
cannam@167
|
1192 Tw = cr[WS(rs, 1)];
|
cannam@167
|
1193 Tx = ci[WS(rs, 14)];
|
cannam@167
|
1194 Ty = Tw + Tx;
|
cannam@167
|
1195 T4N = Tw - Tx;
|
cannam@167
|
1196 {
|
cannam@167
|
1197 E Tz, TA, T1n, T1o;
|
cannam@167
|
1198 Tz = cr[WS(rs, 9)];
|
cannam@167
|
1199 TA = ci[WS(rs, 6)];
|
cannam@167
|
1200 TB = Tz + TA;
|
cannam@167
|
1201 T4Y = Tz - TA;
|
cannam@167
|
1202 T1n = ci[WS(rs, 22)];
|
cannam@167
|
1203 T1o = cr[WS(rs, 25)];
|
cannam@167
|
1204 T1p = T1n - T1o;
|
cannam@167
|
1205 T4O = T1n + T1o;
|
cannam@167
|
1206 }
|
cannam@167
|
1207 T1k = ci[WS(rs, 30)];
|
cannam@167
|
1208 T1l = cr[WS(rs, 17)];
|
cannam@167
|
1209 T1m = T1k - T1l;
|
cannam@167
|
1210 T4Z = T1k + T1l;
|
cannam@167
|
1211 {
|
cannam@167
|
1212 E TG, TH, T4T, T1u, T1v, T4U;
|
cannam@167
|
1213 TG = ci[WS(rs, 2)];
|
cannam@167
|
1214 TH = cr[WS(rs, 13)];
|
cannam@167
|
1215 T4T = TG - TH;
|
cannam@167
|
1216 T1u = ci[WS(rs, 18)];
|
cannam@167
|
1217 T1v = cr[WS(rs, 29)];
|
cannam@167
|
1218 T4U = T1u + T1v;
|
cannam@167
|
1219 TI = TG + TH;
|
cannam@167
|
1220 T52 = T4T + T4U;
|
cannam@167
|
1221 T1w = T1u - T1v;
|
cannam@167
|
1222 T4V = T4T - T4U;
|
cannam@167
|
1223 }
|
cannam@167
|
1224 {
|
cannam@167
|
1225 E TD, TE, T4Q, T1r, T1s, T4R;
|
cannam@167
|
1226 TD = cr[WS(rs, 5)];
|
cannam@167
|
1227 TE = ci[WS(rs, 10)];
|
cannam@167
|
1228 T4Q = TD - TE;
|
cannam@167
|
1229 T1r = ci[WS(rs, 26)];
|
cannam@167
|
1230 T1s = cr[WS(rs, 21)];
|
cannam@167
|
1231 T4R = T1r + T1s;
|
cannam@167
|
1232 TF = TD + TE;
|
cannam@167
|
1233 T51 = T4Q + T4R;
|
cannam@167
|
1234 T1t = T1r - T1s;
|
cannam@167
|
1235 T4S = T4Q - T4R;
|
cannam@167
|
1236 }
|
cannam@167
|
1237 }
|
cannam@167
|
1238 {
|
cannam@167
|
1239 E TC, TJ, T6Q, T6R;
|
cannam@167
|
1240 TC = Ty + TB;
|
cannam@167
|
1241 TJ = TF + TI;
|
cannam@167
|
1242 TK = TC + TJ;
|
cannam@167
|
1243 T1j = TC - TJ;
|
cannam@167
|
1244 T6Q = T4Z - T4Y;
|
cannam@167
|
1245 T6R = KP707106781 * (T4S - T4V);
|
cannam@167
|
1246 T6S = T6Q + T6R;
|
cannam@167
|
1247 T7w = T6Q - T6R;
|
cannam@167
|
1248 }
|
cannam@167
|
1249 {
|
cannam@167
|
1250 E T6T, T6U, T1q, T1x;
|
cannam@167
|
1251 T6T = T4N + T4O;
|
cannam@167
|
1252 T6U = KP707106781 * (T51 + T52);
|
cannam@167
|
1253 T6V = T6T - T6U;
|
cannam@167
|
1254 T7v = T6T + T6U;
|
cannam@167
|
1255 T1q = T1m + T1p;
|
cannam@167
|
1256 T1x = T1t + T1w;
|
cannam@167
|
1257 T1y = T1q - T1x;
|
cannam@167
|
1258 T2B = T1q + T1x;
|
cannam@167
|
1259 }
|
cannam@167
|
1260 {
|
cannam@167
|
1261 E T3a, T3b, T4P, T4W;
|
cannam@167
|
1262 T3a = T1m - T1p;
|
cannam@167
|
1263 T3b = TF - TI;
|
cannam@167
|
1264 T3c = T3a - T3b;
|
cannam@167
|
1265 T3S = T3b + T3a;
|
cannam@167
|
1266 T4P = T4N - T4O;
|
cannam@167
|
1267 T4W = KP707106781 * (T4S + T4V);
|
cannam@167
|
1268 T4X = T4P - T4W;
|
cannam@167
|
1269 T61 = T4P + T4W;
|
cannam@167
|
1270 }
|
cannam@167
|
1271 {
|
cannam@167
|
1272 E T50, T53, T3d, T3e;
|
cannam@167
|
1273 T50 = T4Y + T4Z;
|
cannam@167
|
1274 T53 = KP707106781 * (T51 - T52);
|
cannam@167
|
1275 T54 = T50 - T53;
|
cannam@167
|
1276 T62 = T50 + T53;
|
cannam@167
|
1277 T3d = Ty - TB;
|
cannam@167
|
1278 T3e = T1w - T1t;
|
cannam@167
|
1279 T3f = T3d - T3e;
|
cannam@167
|
1280 T3T = T3d + T3e;
|
cannam@167
|
1281 }
|
cannam@167
|
1282 }
|
cannam@167
|
1283 {
|
cannam@167
|
1284 E TN, T56, TQ, T5h, T1G, T57, T1D, T5i, TX, T5l, T1N, T5e, TU, T5k, T1K;
|
cannam@167
|
1285 E T5b;
|
cannam@167
|
1286 {
|
cannam@167
|
1287 E TL, TM, T1B, T1C;
|
cannam@167
|
1288 TL = ci[0];
|
cannam@167
|
1289 TM = cr[WS(rs, 15)];
|
cannam@167
|
1290 TN = TL + TM;
|
cannam@167
|
1291 T56 = TL - TM;
|
cannam@167
|
1292 {
|
cannam@167
|
1293 E TO, TP, T1E, T1F;
|
cannam@167
|
1294 TO = cr[WS(rs, 7)];
|
cannam@167
|
1295 TP = ci[WS(rs, 8)];
|
cannam@167
|
1296 TQ = TO + TP;
|
cannam@167
|
1297 T5h = TO - TP;
|
cannam@167
|
1298 T1E = ci[WS(rs, 24)];
|
cannam@167
|
1299 T1F = cr[WS(rs, 23)];
|
cannam@167
|
1300 T1G = T1E - T1F;
|
cannam@167
|
1301 T57 = T1E + T1F;
|
cannam@167
|
1302 }
|
cannam@167
|
1303 T1B = ci[WS(rs, 16)];
|
cannam@167
|
1304 T1C = cr[WS(rs, 31)];
|
cannam@167
|
1305 T1D = T1B - T1C;
|
cannam@167
|
1306 T5i = T1B + T1C;
|
cannam@167
|
1307 {
|
cannam@167
|
1308 E TV, TW, T5c, T1L, T1M, T5d;
|
cannam@167
|
1309 TV = ci[WS(rs, 4)];
|
cannam@167
|
1310 TW = cr[WS(rs, 11)];
|
cannam@167
|
1311 T5c = TV - TW;
|
cannam@167
|
1312 T1L = ci[WS(rs, 20)];
|
cannam@167
|
1313 T1M = cr[WS(rs, 27)];
|
cannam@167
|
1314 T5d = T1L + T1M;
|
cannam@167
|
1315 TX = TV + TW;
|
cannam@167
|
1316 T5l = T5c + T5d;
|
cannam@167
|
1317 T1N = T1L - T1M;
|
cannam@167
|
1318 T5e = T5c - T5d;
|
cannam@167
|
1319 }
|
cannam@167
|
1320 {
|
cannam@167
|
1321 E TS, TT, T59, T1I, T1J, T5a;
|
cannam@167
|
1322 TS = cr[WS(rs, 3)];
|
cannam@167
|
1323 TT = ci[WS(rs, 12)];
|
cannam@167
|
1324 T59 = TS - TT;
|
cannam@167
|
1325 T1I = ci[WS(rs, 28)];
|
cannam@167
|
1326 T1J = cr[WS(rs, 19)];
|
cannam@167
|
1327 T5a = T1I + T1J;
|
cannam@167
|
1328 TU = TS + TT;
|
cannam@167
|
1329 T5k = T59 + T5a;
|
cannam@167
|
1330 T1K = T1I - T1J;
|
cannam@167
|
1331 T5b = T59 - T5a;
|
cannam@167
|
1332 }
|
cannam@167
|
1333 }
|
cannam@167
|
1334 {
|
cannam@167
|
1335 E TR, TY, T6J, T6K;
|
cannam@167
|
1336 TR = TN + TQ;
|
cannam@167
|
1337 TY = TU + TX;
|
cannam@167
|
1338 TZ = TR + TY;
|
cannam@167
|
1339 T1A = TR - TY;
|
cannam@167
|
1340 T6J = KP707106781 * (T5b - T5e);
|
cannam@167
|
1341 T6K = T5h + T5i;
|
cannam@167
|
1342 T6L = T6J - T6K;
|
cannam@167
|
1343 T7z = T6K + T6J;
|
cannam@167
|
1344 }
|
cannam@167
|
1345 {
|
cannam@167
|
1346 E T6M, T6N, T1H, T1O;
|
cannam@167
|
1347 T6M = T56 + T57;
|
cannam@167
|
1348 T6N = KP707106781 * (T5k + T5l);
|
cannam@167
|
1349 T6O = T6M - T6N;
|
cannam@167
|
1350 T7y = T6M + T6N;
|
cannam@167
|
1351 T1H = T1D + T1G;
|
cannam@167
|
1352 T1O = T1K + T1N;
|
cannam@167
|
1353 T1P = T1H - T1O;
|
cannam@167
|
1354 T2C = T1H + T1O;
|
cannam@167
|
1355 }
|
cannam@167
|
1356 {
|
cannam@167
|
1357 E T33, T34, T58, T5f;
|
cannam@167
|
1358 T33 = T1D - T1G;
|
cannam@167
|
1359 T34 = TU - TX;
|
cannam@167
|
1360 T35 = T33 - T34;
|
cannam@167
|
1361 T3P = T34 + T33;
|
cannam@167
|
1362 T58 = T56 - T57;
|
cannam@167
|
1363 T5f = KP707106781 * (T5b + T5e);
|
cannam@167
|
1364 T5g = T58 - T5f;
|
cannam@167
|
1365 T64 = T58 + T5f;
|
cannam@167
|
1366 }
|
cannam@167
|
1367 {
|
cannam@167
|
1368 E T5j, T5m, T36, T37;
|
cannam@167
|
1369 T5j = T5h - T5i;
|
cannam@167
|
1370 T5m = KP707106781 * (T5k - T5l);
|
cannam@167
|
1371 T5n = T5j - T5m;
|
cannam@167
|
1372 T65 = T5j + T5m;
|
cannam@167
|
1373 T36 = TN - TQ;
|
cannam@167
|
1374 T37 = T1N - T1K;
|
cannam@167
|
1375 T38 = T36 - T37;
|
cannam@167
|
1376 T3Q = T36 + T37;
|
cannam@167
|
1377 }
|
cannam@167
|
1378 }
|
cannam@167
|
1379 {
|
cannam@167
|
1380 E Tv, T10, T2w, T2A, T2D, T2E, T2v, T2x;
|
cannam@167
|
1381 Tv = Tf + Tu;
|
cannam@167
|
1382 T10 = TK + TZ;
|
cannam@167
|
1383 T2w = Tv - T10;
|
cannam@167
|
1384 T2A = T2y + T2z;
|
cannam@167
|
1385 T2D = T2B + T2C;
|
cannam@167
|
1386 T2E = T2A - T2D;
|
cannam@167
|
1387 cr[0] = Tv + T10;
|
cannam@167
|
1388 ci[0] = T2A + T2D;
|
cannam@167
|
1389 T2v = W[30];
|
cannam@167
|
1390 T2x = W[31];
|
cannam@167
|
1391 cr[WS(rs, 16)] = FNMS(T2x, T2E, T2v * T2w);
|
cannam@167
|
1392 ci[WS(rs, 16)] = FMA(T2x, T2w, T2v * T2E);
|
cannam@167
|
1393 }
|
cannam@167
|
1394 {
|
cannam@167
|
1395 E T2I, T2O, T2M, T2Q;
|
cannam@167
|
1396 {
|
cannam@167
|
1397 E T2G, T2H, T2K, T2L;
|
cannam@167
|
1398 T2G = Tf - Tu;
|
cannam@167
|
1399 T2H = T2C - T2B;
|
cannam@167
|
1400 T2I = T2G - T2H;
|
cannam@167
|
1401 T2O = T2G + T2H;
|
cannam@167
|
1402 T2K = T2y - T2z;
|
cannam@167
|
1403 T2L = TK - TZ;
|
cannam@167
|
1404 T2M = T2K - T2L;
|
cannam@167
|
1405 T2Q = T2L + T2K;
|
cannam@167
|
1406 }
|
cannam@167
|
1407 {
|
cannam@167
|
1408 E T2F, T2J, T2N, T2P;
|
cannam@167
|
1409 T2F = W[46];
|
cannam@167
|
1410 T2J = W[47];
|
cannam@167
|
1411 cr[WS(rs, 24)] = FNMS(T2J, T2M, T2F * T2I);
|
cannam@167
|
1412 ci[WS(rs, 24)] = FMA(T2F, T2M, T2J * T2I);
|
cannam@167
|
1413 T2N = W[14];
|
cannam@167
|
1414 T2P = W[15];
|
cannam@167
|
1415 cr[WS(rs, 8)] = FNMS(T2P, T2Q, T2N * T2O);
|
cannam@167
|
1416 ci[WS(rs, 8)] = FMA(T2N, T2Q, T2P * T2O);
|
cannam@167
|
1417 }
|
cannam@167
|
1418 }
|
cannam@167
|
1419 {
|
cannam@167
|
1420 E T1i, T2a, T2o, T2k, T2d, T2l, T1R, T2p;
|
cannam@167
|
1421 T1i = T12 + T1h;
|
cannam@167
|
1422 T2a = T1U + T29;
|
cannam@167
|
1423 T2o = T29 - T1U;
|
cannam@167
|
1424 T2k = T12 - T1h;
|
cannam@167
|
1425 {
|
cannam@167
|
1426 E T2b, T2c, T1z, T1Q;
|
cannam@167
|
1427 T2b = T1j + T1y;
|
cannam@167
|
1428 T2c = T1P - T1A;
|
cannam@167
|
1429 T2d = KP707106781 * (T2b + T2c);
|
cannam@167
|
1430 T2l = KP707106781 * (T2c - T2b);
|
cannam@167
|
1431 T1z = T1j - T1y;
|
cannam@167
|
1432 T1Q = T1A + T1P;
|
cannam@167
|
1433 T1R = KP707106781 * (T1z + T1Q);
|
cannam@167
|
1434 T2p = KP707106781 * (T1z - T1Q);
|
cannam@167
|
1435 }
|
cannam@167
|
1436 {
|
cannam@167
|
1437 E T1S, T2e, T11, T1T;
|
cannam@167
|
1438 T1S = T1i - T1R;
|
cannam@167
|
1439 T2e = T2a - T2d;
|
cannam@167
|
1440 T11 = W[38];
|
cannam@167
|
1441 T1T = W[39];
|
cannam@167
|
1442 cr[WS(rs, 20)] = FNMS(T1T, T2e, T11 * T1S);
|
cannam@167
|
1443 ci[WS(rs, 20)] = FMA(T1T, T1S, T11 * T2e);
|
cannam@167
|
1444 }
|
cannam@167
|
1445 {
|
cannam@167
|
1446 E T2s, T2u, T2r, T2t;
|
cannam@167
|
1447 T2s = T2k + T2l;
|
cannam@167
|
1448 T2u = T2o + T2p;
|
cannam@167
|
1449 T2r = W[22];
|
cannam@167
|
1450 T2t = W[23];
|
cannam@167
|
1451 cr[WS(rs, 12)] = FNMS(T2t, T2u, T2r * T2s);
|
cannam@167
|
1452 ci[WS(rs, 12)] = FMA(T2r, T2u, T2t * T2s);
|
cannam@167
|
1453 }
|
cannam@167
|
1454 {
|
cannam@167
|
1455 E T2g, T2i, T2f, T2h;
|
cannam@167
|
1456 T2g = T1i + T1R;
|
cannam@167
|
1457 T2i = T2a + T2d;
|
cannam@167
|
1458 T2f = W[6];
|
cannam@167
|
1459 T2h = W[7];
|
cannam@167
|
1460 cr[WS(rs, 4)] = FNMS(T2h, T2i, T2f * T2g);
|
cannam@167
|
1461 ci[WS(rs, 4)] = FMA(T2h, T2g, T2f * T2i);
|
cannam@167
|
1462 }
|
cannam@167
|
1463 {
|
cannam@167
|
1464 E T2m, T2q, T2j, T2n;
|
cannam@167
|
1465 T2m = T2k - T2l;
|
cannam@167
|
1466 T2q = T2o - T2p;
|
cannam@167
|
1467 T2j = W[54];
|
cannam@167
|
1468 T2n = W[55];
|
cannam@167
|
1469 cr[WS(rs, 28)] = FNMS(T2n, T2q, T2j * T2m);
|
cannam@167
|
1470 ci[WS(rs, 28)] = FMA(T2j, T2q, T2n * T2m);
|
cannam@167
|
1471 }
|
cannam@167
|
1472 }
|
cannam@167
|
1473 {
|
cannam@167
|
1474 E T3O, T4a, T40, T4e, T3V, T4f, T43, T4b, T3N, T3Z;
|
cannam@167
|
1475 T3N = KP707106781 * (T3n + T3o);
|
cannam@167
|
1476 T3O = T3M - T3N;
|
cannam@167
|
1477 T4a = T3M + T3N;
|
cannam@167
|
1478 T3Z = KP707106781 * (T30 + T2X);
|
cannam@167
|
1479 T40 = T3Y - T3Z;
|
cannam@167
|
1480 T4e = T3Y + T3Z;
|
cannam@167
|
1481 {
|
cannam@167
|
1482 E T3R, T3U, T41, T42;
|
cannam@167
|
1483 T3R = FNMS(KP382683432, T3Q, KP923879532 * T3P);
|
cannam@167
|
1484 T3U = FMA(KP923879532, T3S, KP382683432 * T3T);
|
cannam@167
|
1485 T3V = T3R - T3U;
|
cannam@167
|
1486 T4f = T3U + T3R;
|
cannam@167
|
1487 T41 = FNMS(KP382683432, T3S, KP923879532 * T3T);
|
cannam@167
|
1488 T42 = FMA(KP382683432, T3P, KP923879532 * T3Q);
|
cannam@167
|
1489 T43 = T41 - T42;
|
cannam@167
|
1490 T4b = T41 + T42;
|
cannam@167
|
1491 }
|
cannam@167
|
1492 {
|
cannam@167
|
1493 E T3W, T44, T3L, T3X;
|
cannam@167
|
1494 T3W = T3O - T3V;
|
cannam@167
|
1495 T44 = T40 - T43;
|
cannam@167
|
1496 T3L = W[50];
|
cannam@167
|
1497 T3X = W[51];
|
cannam@167
|
1498 cr[WS(rs, 26)] = FNMS(T3X, T44, T3L * T3W);
|
cannam@167
|
1499 ci[WS(rs, 26)] = FMA(T3X, T3W, T3L * T44);
|
cannam@167
|
1500 }
|
cannam@167
|
1501 {
|
cannam@167
|
1502 E T4i, T4k, T4h, T4j;
|
cannam@167
|
1503 T4i = T4a + T4b;
|
cannam@167
|
1504 T4k = T4e + T4f;
|
cannam@167
|
1505 T4h = W[2];
|
cannam@167
|
1506 T4j = W[3];
|
cannam@167
|
1507 cr[WS(rs, 2)] = FNMS(T4j, T4k, T4h * T4i);
|
cannam@167
|
1508 ci[WS(rs, 2)] = FMA(T4h, T4k, T4j * T4i);
|
cannam@167
|
1509 }
|
cannam@167
|
1510 {
|
cannam@167
|
1511 E T46, T48, T45, T47;
|
cannam@167
|
1512 T46 = T3O + T3V;
|
cannam@167
|
1513 T48 = T40 + T43;
|
cannam@167
|
1514 T45 = W[18];
|
cannam@167
|
1515 T47 = W[19];
|
cannam@167
|
1516 cr[WS(rs, 10)] = FNMS(T47, T48, T45 * T46);
|
cannam@167
|
1517 ci[WS(rs, 10)] = FMA(T47, T46, T45 * T48);
|
cannam@167
|
1518 }
|
cannam@167
|
1519 {
|
cannam@167
|
1520 E T4c, T4g, T49, T4d;
|
cannam@167
|
1521 T4c = T4a - T4b;
|
cannam@167
|
1522 T4g = T4e - T4f;
|
cannam@167
|
1523 T49 = W[34];
|
cannam@167
|
1524 T4d = W[35];
|
cannam@167
|
1525 cr[WS(rs, 18)] = FNMS(T4d, T4g, T49 * T4c);
|
cannam@167
|
1526 ci[WS(rs, 18)] = FMA(T49, T4g, T4d * T4c);
|
cannam@167
|
1527 }
|
cannam@167
|
1528 }
|
cannam@167
|
1529 {
|
cannam@167
|
1530 E T32, T3A, T3q, T3E, T3h, T3F, T3t, T3B, T31, T3p;
|
cannam@167
|
1531 T31 = KP707106781 * (T2X - T30);
|
cannam@167
|
1532 T32 = T2U - T31;
|
cannam@167
|
1533 T3A = T2U + T31;
|
cannam@167
|
1534 T3p = KP707106781 * (T3n - T3o);
|
cannam@167
|
1535 T3q = T3m - T3p;
|
cannam@167
|
1536 T3E = T3m + T3p;
|
cannam@167
|
1537 {
|
cannam@167
|
1538 E T39, T3g, T3r, T3s;
|
cannam@167
|
1539 T39 = FNMS(KP923879532, T38, KP382683432 * T35);
|
cannam@167
|
1540 T3g = FMA(KP382683432, T3c, KP923879532 * T3f);
|
cannam@167
|
1541 T3h = T39 - T3g;
|
cannam@167
|
1542 T3F = T3g + T39;
|
cannam@167
|
1543 T3r = FNMS(KP923879532, T3c, KP382683432 * T3f);
|
cannam@167
|
1544 T3s = FMA(KP923879532, T35, KP382683432 * T38);
|
cannam@167
|
1545 T3t = T3r - T3s;
|
cannam@167
|
1546 T3B = T3r + T3s;
|
cannam@167
|
1547 }
|
cannam@167
|
1548 {
|
cannam@167
|
1549 E T3i, T3u, T2R, T3j;
|
cannam@167
|
1550 T3i = T32 - T3h;
|
cannam@167
|
1551 T3u = T3q - T3t;
|
cannam@167
|
1552 T2R = W[58];
|
cannam@167
|
1553 T3j = W[59];
|
cannam@167
|
1554 cr[WS(rs, 30)] = FNMS(T3j, T3u, T2R * T3i);
|
cannam@167
|
1555 ci[WS(rs, 30)] = FMA(T3j, T3i, T2R * T3u);
|
cannam@167
|
1556 }
|
cannam@167
|
1557 {
|
cannam@167
|
1558 E T3I, T3K, T3H, T3J;
|
cannam@167
|
1559 T3I = T3A + T3B;
|
cannam@167
|
1560 T3K = T3E + T3F;
|
cannam@167
|
1561 T3H = W[10];
|
cannam@167
|
1562 T3J = W[11];
|
cannam@167
|
1563 cr[WS(rs, 6)] = FNMS(T3J, T3K, T3H * T3I);
|
cannam@167
|
1564 ci[WS(rs, 6)] = FMA(T3H, T3K, T3J * T3I);
|
cannam@167
|
1565 }
|
cannam@167
|
1566 {
|
cannam@167
|
1567 E T3w, T3y, T3v, T3x;
|
cannam@167
|
1568 T3w = T32 + T3h;
|
cannam@167
|
1569 T3y = T3q + T3t;
|
cannam@167
|
1570 T3v = W[26];
|
cannam@167
|
1571 T3x = W[27];
|
cannam@167
|
1572 cr[WS(rs, 14)] = FNMS(T3x, T3y, T3v * T3w);
|
cannam@167
|
1573 ci[WS(rs, 14)] = FMA(T3x, T3w, T3v * T3y);
|
cannam@167
|
1574 }
|
cannam@167
|
1575 {
|
cannam@167
|
1576 E T3C, T3G, T3z, T3D;
|
cannam@167
|
1577 T3C = T3A - T3B;
|
cannam@167
|
1578 T3G = T3E - T3F;
|
cannam@167
|
1579 T3z = W[42];
|
cannam@167
|
1580 T3D = W[43];
|
cannam@167
|
1581 cr[WS(rs, 22)] = FNMS(T3D, T3G, T3z * T3C);
|
cannam@167
|
1582 ci[WS(rs, 22)] = FMA(T3z, T3G, T3D * T3C);
|
cannam@167
|
1583 }
|
cannam@167
|
1584 }
|
cannam@167
|
1585 {
|
cannam@167
|
1586 E T60, T6m, T6f, T6n, T67, T6r, T6c, T6q;
|
cannam@167
|
1587 {
|
cannam@167
|
1588 E T5Y, T5Z, T6d, T6e;
|
cannam@167
|
1589 T5Y = T4o + T4v;
|
cannam@167
|
1590 T5Z = T5z + T5A;
|
cannam@167
|
1591 T60 = T5Y + T5Z;
|
cannam@167
|
1592 T6m = T5Y - T5Z;
|
cannam@167
|
1593 T6d = FMA(KP195090322, T61, KP980785280 * T62);
|
cannam@167
|
1594 T6e = FNMS(KP195090322, T64, KP980785280 * T65);
|
cannam@167
|
1595 T6f = T6d + T6e;
|
cannam@167
|
1596 T6n = T6e - T6d;
|
cannam@167
|
1597 }
|
cannam@167
|
1598 {
|
cannam@167
|
1599 E T63, T66, T6a, T6b;
|
cannam@167
|
1600 T63 = FNMS(KP195090322, T62, KP980785280 * T61);
|
cannam@167
|
1601 T66 = FMA(KP980785280, T64, KP195090322 * T65);
|
cannam@167
|
1602 T67 = T63 + T66;
|
cannam@167
|
1603 T6r = T63 - T66;
|
cannam@167
|
1604 T6a = T5u + T5x;
|
cannam@167
|
1605 T6b = T4K + T4D;
|
cannam@167
|
1606 T6c = T6a + T6b;
|
cannam@167
|
1607 T6q = T6a - T6b;
|
cannam@167
|
1608 }
|
cannam@167
|
1609 {
|
cannam@167
|
1610 E T68, T6g, T5X, T69;
|
cannam@167
|
1611 T68 = T60 - T67;
|
cannam@167
|
1612 T6g = T6c - T6f;
|
cannam@167
|
1613 T5X = W[32];
|
cannam@167
|
1614 T69 = W[33];
|
cannam@167
|
1615 cr[WS(rs, 17)] = FNMS(T69, T6g, T5X * T68);
|
cannam@167
|
1616 ci[WS(rs, 17)] = FMA(T69, T68, T5X * T6g);
|
cannam@167
|
1617 }
|
cannam@167
|
1618 {
|
cannam@167
|
1619 E T6u, T6w, T6t, T6v;
|
cannam@167
|
1620 T6u = T6m + T6n;
|
cannam@167
|
1621 T6w = T6q + T6r;
|
cannam@167
|
1622 T6t = W[16];
|
cannam@167
|
1623 T6v = W[17];
|
cannam@167
|
1624 cr[WS(rs, 9)] = FNMS(T6v, T6w, T6t * T6u);
|
cannam@167
|
1625 ci[WS(rs, 9)] = FMA(T6t, T6w, T6v * T6u);
|
cannam@167
|
1626 }
|
cannam@167
|
1627 {
|
cannam@167
|
1628 E T6i, T6k, T6h, T6j;
|
cannam@167
|
1629 T6i = T60 + T67;
|
cannam@167
|
1630 T6k = T6c + T6f;
|
cannam@167
|
1631 T6h = W[0];
|
cannam@167
|
1632 T6j = W[1];
|
cannam@167
|
1633 cr[WS(rs, 1)] = FNMS(T6j, T6k, T6h * T6i);
|
cannam@167
|
1634 ci[WS(rs, 1)] = FMA(T6j, T6i, T6h * T6k);
|
cannam@167
|
1635 }
|
cannam@167
|
1636 {
|
cannam@167
|
1637 E T6o, T6s, T6l, T6p;
|
cannam@167
|
1638 T6o = T6m - T6n;
|
cannam@167
|
1639 T6s = T6q - T6r;
|
cannam@167
|
1640 T6l = W[48];
|
cannam@167
|
1641 T6p = W[49];
|
cannam@167
|
1642 cr[WS(rs, 25)] = FNMS(T6p, T6s, T6l * T6o);
|
cannam@167
|
1643 ci[WS(rs, 25)] = FMA(T6l, T6s, T6p * T6o);
|
cannam@167
|
1644 }
|
cannam@167
|
1645 }
|
cannam@167
|
1646 {
|
cannam@167
|
1647 E T7u, T7Q, T7J, T7R, T7B, T7V, T7G, T7U;
|
cannam@167
|
1648 {
|
cannam@167
|
1649 E T7s, T7t, T7H, T7I;
|
cannam@167
|
1650 T7s = T6y + T6z;
|
cannam@167
|
1651 T7t = T73 + T74;
|
cannam@167
|
1652 T7u = T7s - T7t;
|
cannam@167
|
1653 T7Q = T7s + T7t;
|
cannam@167
|
1654 T7H = FMA(KP195090322, T7w, KP980785280 * T7v);
|
cannam@167
|
1655 T7I = FMA(KP195090322, T7z, KP980785280 * T7y);
|
cannam@167
|
1656 T7J = T7H - T7I;
|
cannam@167
|
1657 T7R = T7H + T7I;
|
cannam@167
|
1658 }
|
cannam@167
|
1659 {
|
cannam@167
|
1660 E T7x, T7A, T7E, T7F;
|
cannam@167
|
1661 T7x = FNMS(KP980785280, T7w, KP195090322 * T7v);
|
cannam@167
|
1662 T7A = FNMS(KP980785280, T7z, KP195090322 * T7y);
|
cannam@167
|
1663 T7B = T7x + T7A;
|
cannam@167
|
1664 T7V = T7x - T7A;
|
cannam@167
|
1665 T7E = T70 - T71;
|
cannam@167
|
1666 T7F = T6D - T6G;
|
cannam@167
|
1667 T7G = T7E + T7F;
|
cannam@167
|
1668 T7U = T7E - T7F;
|
cannam@167
|
1669 }
|
cannam@167
|
1670 {
|
cannam@167
|
1671 E T7C, T7K, T7r, T7D;
|
cannam@167
|
1672 T7C = T7u - T7B;
|
cannam@167
|
1673 T7K = T7G - T7J;
|
cannam@167
|
1674 T7r = W[44];
|
cannam@167
|
1675 T7D = W[45];
|
cannam@167
|
1676 cr[WS(rs, 23)] = FNMS(T7D, T7K, T7r * T7C);
|
cannam@167
|
1677 ci[WS(rs, 23)] = FMA(T7D, T7C, T7r * T7K);
|
cannam@167
|
1678 }
|
cannam@167
|
1679 {
|
cannam@167
|
1680 E T7Y, T80, T7X, T7Z;
|
cannam@167
|
1681 T7Y = T7Q + T7R;
|
cannam@167
|
1682 T80 = T7U - T7V;
|
cannam@167
|
1683 T7X = W[60];
|
cannam@167
|
1684 T7Z = W[61];
|
cannam@167
|
1685 cr[WS(rs, 31)] = FNMS(T7Z, T80, T7X * T7Y);
|
cannam@167
|
1686 ci[WS(rs, 31)] = FMA(T7X, T80, T7Z * T7Y);
|
cannam@167
|
1687 }
|
cannam@167
|
1688 {
|
cannam@167
|
1689 E T7M, T7O, T7L, T7N;
|
cannam@167
|
1690 T7M = T7u + T7B;
|
cannam@167
|
1691 T7O = T7G + T7J;
|
cannam@167
|
1692 T7L = W[12];
|
cannam@167
|
1693 T7N = W[13];
|
cannam@167
|
1694 cr[WS(rs, 7)] = FNMS(T7N, T7O, T7L * T7M);
|
cannam@167
|
1695 ci[WS(rs, 7)] = FMA(T7N, T7M, T7L * T7O);
|
cannam@167
|
1696 }
|
cannam@167
|
1697 {
|
cannam@167
|
1698 E T7S, T7W, T7P, T7T;
|
cannam@167
|
1699 T7S = T7Q - T7R;
|
cannam@167
|
1700 T7W = T7U + T7V;
|
cannam@167
|
1701 T7P = W[28];
|
cannam@167
|
1702 T7T = W[29];
|
cannam@167
|
1703 cr[WS(rs, 15)] = FNMS(T7T, T7W, T7P * T7S);
|
cannam@167
|
1704 ci[WS(rs, 15)] = FMA(T7P, T7W, T7T * T7S);
|
cannam@167
|
1705 }
|
cannam@167
|
1706 }
|
cannam@167
|
1707 {
|
cannam@167
|
1708 E T4M, T5M, T5F, T5N, T5p, T5R, T5C, T5Q;
|
cannam@167
|
1709 {
|
cannam@167
|
1710 E T4w, T4L, T5D, T5E;
|
cannam@167
|
1711 T4w = T4o - T4v;
|
cannam@167
|
1712 T4L = T4D - T4K;
|
cannam@167
|
1713 T4M = T4w + T4L;
|
cannam@167
|
1714 T5M = T4w - T4L;
|
cannam@167
|
1715 T5D = FMA(KP831469612, T4X, KP555570233 * T54);
|
cannam@167
|
1716 T5E = FNMS(KP831469612, T5g, KP555570233 * T5n);
|
cannam@167
|
1717 T5F = T5D + T5E;
|
cannam@167
|
1718 T5N = T5E - T5D;
|
cannam@167
|
1719 }
|
cannam@167
|
1720 {
|
cannam@167
|
1721 E T55, T5o, T5y, T5B;
|
cannam@167
|
1722 T55 = FNMS(KP831469612, T54, KP555570233 * T4X);
|
cannam@167
|
1723 T5o = FMA(KP555570233, T5g, KP831469612 * T5n);
|
cannam@167
|
1724 T5p = T55 + T5o;
|
cannam@167
|
1725 T5R = T55 - T5o;
|
cannam@167
|
1726 T5y = T5u - T5x;
|
cannam@167
|
1727 T5B = T5z - T5A;
|
cannam@167
|
1728 T5C = T5y + T5B;
|
cannam@167
|
1729 T5Q = T5y - T5B;
|
cannam@167
|
1730 }
|
cannam@167
|
1731 {
|
cannam@167
|
1732 E T5q, T5G, T4l, T5r;
|
cannam@167
|
1733 T5q = T4M - T5p;
|
cannam@167
|
1734 T5G = T5C - T5F;
|
cannam@167
|
1735 T4l = W[40];
|
cannam@167
|
1736 T5r = W[41];
|
cannam@167
|
1737 cr[WS(rs, 21)] = FNMS(T5r, T5G, T4l * T5q);
|
cannam@167
|
1738 ci[WS(rs, 21)] = FMA(T5r, T5q, T4l * T5G);
|
cannam@167
|
1739 }
|
cannam@167
|
1740 {
|
cannam@167
|
1741 E T5U, T5W, T5T, T5V;
|
cannam@167
|
1742 T5U = T5M + T5N;
|
cannam@167
|
1743 T5W = T5Q + T5R;
|
cannam@167
|
1744 T5T = W[24];
|
cannam@167
|
1745 T5V = W[25];
|
cannam@167
|
1746 cr[WS(rs, 13)] = FNMS(T5V, T5W, T5T * T5U);
|
cannam@167
|
1747 ci[WS(rs, 13)] = FMA(T5T, T5W, T5V * T5U);
|
cannam@167
|
1748 }
|
cannam@167
|
1749 {
|
cannam@167
|
1750 E T5I, T5K, T5H, T5J;
|
cannam@167
|
1751 T5I = T4M + T5p;
|
cannam@167
|
1752 T5K = T5C + T5F;
|
cannam@167
|
1753 T5H = W[8];
|
cannam@167
|
1754 T5J = W[9];
|
cannam@167
|
1755 cr[WS(rs, 5)] = FNMS(T5J, T5K, T5H * T5I);
|
cannam@167
|
1756 ci[WS(rs, 5)] = FMA(T5J, T5I, T5H * T5K);
|
cannam@167
|
1757 }
|
cannam@167
|
1758 {
|
cannam@167
|
1759 E T5O, T5S, T5L, T5P;
|
cannam@167
|
1760 T5O = T5M - T5N;
|
cannam@167
|
1761 T5S = T5Q - T5R;
|
cannam@167
|
1762 T5L = W[56];
|
cannam@167
|
1763 T5P = W[57];
|
cannam@167
|
1764 cr[WS(rs, 29)] = FNMS(T5P, T5S, T5L * T5O);
|
cannam@167
|
1765 ci[WS(rs, 29)] = FMA(T5L, T5S, T5P * T5O);
|
cannam@167
|
1766 }
|
cannam@167
|
1767 }
|
cannam@167
|
1768 {
|
cannam@167
|
1769 E T6I, T7g, T79, T7h, T6X, T7l, T76, T7k;
|
cannam@167
|
1770 {
|
cannam@167
|
1771 E T6A, T6H, T77, T78;
|
cannam@167
|
1772 T6A = T6y - T6z;
|
cannam@167
|
1773 T6H = T6D + T6G;
|
cannam@167
|
1774 T6I = T6A - T6H;
|
cannam@167
|
1775 T7g = T6A + T6H;
|
cannam@167
|
1776 T77 = FNMS(KP555570233, T6S, KP831469612 * T6V);
|
cannam@167
|
1777 T78 = FMA(KP555570233, T6L, KP831469612 * T6O);
|
cannam@167
|
1778 T79 = T77 - T78;
|
cannam@167
|
1779 T7h = T77 + T78;
|
cannam@167
|
1780 }
|
cannam@167
|
1781 {
|
cannam@167
|
1782 E T6P, T6W, T72, T75;
|
cannam@167
|
1783 T6P = FNMS(KP555570233, T6O, KP831469612 * T6L);
|
cannam@167
|
1784 T6W = FMA(KP831469612, T6S, KP555570233 * T6V);
|
cannam@167
|
1785 T6X = T6P - T6W;
|
cannam@167
|
1786 T7l = T6W + T6P;
|
cannam@167
|
1787 T72 = T70 + T71;
|
cannam@167
|
1788 T75 = T73 - T74;
|
cannam@167
|
1789 T76 = T72 - T75;
|
cannam@167
|
1790 T7k = T72 + T75;
|
cannam@167
|
1791 }
|
cannam@167
|
1792 {
|
cannam@167
|
1793 E T6Y, T7a, T6x, T6Z;
|
cannam@167
|
1794 T6Y = T6I - T6X;
|
cannam@167
|
1795 T7a = T76 - T79;
|
cannam@167
|
1796 T6x = W[52];
|
cannam@167
|
1797 T6Z = W[53];
|
cannam@167
|
1798 cr[WS(rs, 27)] = FNMS(T6Z, T7a, T6x * T6Y);
|
cannam@167
|
1799 ci[WS(rs, 27)] = FMA(T6Z, T6Y, T6x * T7a);
|
cannam@167
|
1800 }
|
cannam@167
|
1801 {
|
cannam@167
|
1802 E T7o, T7q, T7n, T7p;
|
cannam@167
|
1803 T7o = T7g + T7h;
|
cannam@167
|
1804 T7q = T7k + T7l;
|
cannam@167
|
1805 T7n = W[4];
|
cannam@167
|
1806 T7p = W[5];
|
cannam@167
|
1807 cr[WS(rs, 3)] = FNMS(T7p, T7q, T7n * T7o);
|
cannam@167
|
1808 ci[WS(rs, 3)] = FMA(T7n, T7q, T7p * T7o);
|
cannam@167
|
1809 }
|
cannam@167
|
1810 {
|
cannam@167
|
1811 E T7c, T7e, T7b, T7d;
|
cannam@167
|
1812 T7c = T6I + T6X;
|
cannam@167
|
1813 T7e = T76 + T79;
|
cannam@167
|
1814 T7b = W[20];
|
cannam@167
|
1815 T7d = W[21];
|
cannam@167
|
1816 cr[WS(rs, 11)] = FNMS(T7d, T7e, T7b * T7c);
|
cannam@167
|
1817 ci[WS(rs, 11)] = FMA(T7d, T7c, T7b * T7e);
|
cannam@167
|
1818 }
|
cannam@167
|
1819 {
|
cannam@167
|
1820 E T7i, T7m, T7f, T7j;
|
cannam@167
|
1821 T7i = T7g - T7h;
|
cannam@167
|
1822 T7m = T7k - T7l;
|
cannam@167
|
1823 T7f = W[36];
|
cannam@167
|
1824 T7j = W[37];
|
cannam@167
|
1825 cr[WS(rs, 19)] = FNMS(T7j, T7m, T7f * T7i);
|
cannam@167
|
1826 ci[WS(rs, 19)] = FMA(T7f, T7m, T7j * T7i);
|
cannam@167
|
1827 }
|
cannam@167
|
1828 }
|
cannam@167
|
1829 }
|
cannam@167
|
1830 }
|
cannam@167
|
1831 }
|
cannam@167
|
1832
|
cannam@167
|
1833 static const tw_instr twinstr[] = {
|
cannam@167
|
1834 {TW_FULL, 1, 32},
|
cannam@167
|
1835 {TW_NEXT, 1, 0}
|
cannam@167
|
1836 };
|
cannam@167
|
1837
|
cannam@167
|
1838 static const hc2hc_desc desc = { 32, "hb_32", twinstr, &GENUS, {340, 114, 94, 0} };
|
cannam@167
|
1839
|
cannam@167
|
1840 void X(codelet_hb_32) (planner *p) {
|
cannam@167
|
1841 X(khc2hc_register) (p, hb_32, &desc);
|
cannam@167
|
1842 }
|
cannam@167
|
1843 #endif
|