comparison src/fftw-3.3.8/dft/simd/common/t2sv_32.c @ 167:bd3cc4d1df30

Add FFTW 3.3.8 source, and a Linux build
author Chris Cannam <cannam@all-day-breakfast.com>
date Tue, 19 Nov 2019 14:52:55 +0000
parents
children
comparison
equal deleted inserted replaced
166:cbd6d7e562c7 167:bd3cc4d1df30
1 /*
2 * Copyright (c) 2003, 2007-14 Matteo Frigo
3 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
4 *
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 2 of the License, or
8 * (at your option) any later version.
9 *
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
14 *
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write to the Free Software
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
18 *
19 */
20
21 /* This file was automatically generated --- DO NOT EDIT */
22 /* Generated on Thu May 24 08:06:12 EDT 2018 */
23
24 #include "dft/codelet-dft.h"
25
26 #if defined(ARCH_PREFERS_FMA) || defined(ISA_EXTENSION_PREFERS_FMA)
27
28 /* Generated by: ../../../genfft/gen_twiddle.native -fma -simd -compact -variables 4 -pipeline-latency 8 -twiddle-log3 -precompute-twiddles -n 32 -name t2sv_32 -include dft/simd/ts.h */
29
30 /*
31 * This function contains 488 FP additions, 350 FP multiplications,
32 * (or, 236 additions, 98 multiplications, 252 fused multiply/add),
33 * 164 stack variables, 7 constants, and 128 memory accesses
34 */
35 #include "dft/simd/ts.h"
36
37 static void t2sv_32(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
38 {
39 DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
40 DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
41 DVK(KP198912367, +0.198912367379658006911597622644676228597850501);
42 DVK(KP668178637, +0.668178637919298919997757686523080761552472251);
43 DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
44 DVK(KP414213562, +0.414213562373095048801688724209698078569671875);
45 DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
46 {
47 INT m;
48 for (m = mb, W = W + (mb * 8); m < me; m = m + (2 * VL), ri = ri + ((2 * VL) * ms), ii = ii + ((2 * VL) * ms), W = W + ((2 * VL) * 8), MAKE_VOLATILE_STRIDE(64, rs)) {
49 V T2, T8, T3, T6, Te, Ti, T5, T7, TJ, Tb, TM, Tc, Ts, T23, T1w;
50 V T19, TA, TE, T1s, T1N, T1o, T1C, T1F, T1K, T15, T11, T2F, T31, T2J, T34;
51 V T3f, T3z, T3j, T3C, Tw, T3M, T3Q, T1z, T2s, T2w, T1d, T3n, T3r, T26, T2T;
52 V T2X, Th, TR, TP, Td, Tj, TW, Tn, TS, T1U, T2b, T29, T1R, T1V, T2g;
53 V T1Z, T2c;
54 {
55 V Tz, T1n, T10, TD, T1r, T14, T9, T1Q, Tv, T1c;
56 {
57 V T4, T18, Ta, Tr;
58 T2 = LDW(&(W[0]));
59 T8 = LDW(&(W[TWVL * 4]));
60 T3 = LDW(&(W[TWVL * 2]));
61 T6 = LDW(&(W[TWVL * 3]));
62 T4 = VMUL(T2, T3);
63 T18 = VMUL(T3, T8);
64 Ta = VMUL(T2, T6);
65 Tr = VMUL(T2, T8);
66 Te = LDW(&(W[TWVL * 6]));
67 Tz = VMUL(T3, Te);
68 T1n = VMUL(T8, Te);
69 T10 = VMUL(T2, Te);
70 Ti = LDW(&(W[TWVL * 7]));
71 TD = VMUL(T3, Ti);
72 T1r = VMUL(T8, Ti);
73 T14 = VMUL(T2, Ti);
74 T5 = LDW(&(W[TWVL * 1]));
75 T7 = VFMA(T5, T6, T4);
76 TJ = VFNMS(T5, T6, T4);
77 T9 = VMUL(T7, T8);
78 T1Q = VMUL(TJ, T8);
79 Tb = VFNMS(T5, T3, Ta);
80 TM = VFMA(T5, T3, Ta);
81 Tc = LDW(&(W[TWVL * 5]));
82 Tv = VMUL(T2, Tc);
83 T1c = VMUL(T3, Tc);
84 Ts = VFMA(T5, Tc, Tr);
85 T23 = VFMA(T6, Tc, T18);
86 T1w = VFNMS(T5, Tc, Tr);
87 T19 = VFNMS(T6, Tc, T18);
88 }
89 TA = VFMA(T6, Ti, Tz);
90 TE = VFNMS(T6, Te, TD);
91 T1s = VFNMS(Tc, Te, T1r);
92 T1N = VFMA(T6, Te, TD);
93 T1o = VFMA(Tc, Ti, T1n);
94 T1C = VFMA(T5, Ti, T10);
95 T1F = VFNMS(T5, Te, T14);
96 T1K = VFNMS(T6, Ti, Tz);
97 T15 = VFMA(T5, Te, T14);
98 T11 = VFNMS(T5, Ti, T10);
99 {
100 V T2E, T2I, T2S, T2W;
101 T2E = VMUL(T7, Te);
102 T2F = VFMA(Tb, Ti, T2E);
103 T31 = VFNMS(Tb, Ti, T2E);
104 T2I = VMUL(T7, Ti);
105 T2J = VFNMS(Tb, Te, T2I);
106 T34 = VFMA(Tb, Te, T2I);
107 {
108 V T3e, T3i, T3L, T3P;
109 T3e = VMUL(TJ, Te);
110 T3f = VFNMS(TM, Ti, T3e);
111 T3z = VFMA(TM, Ti, T3e);
112 T3i = VMUL(TJ, Ti);
113 T3j = VFMA(TM, Te, T3i);
114 T3C = VFNMS(TM, Te, T3i);
115 T3L = VMUL(Ts, Te);
116 T3P = VMUL(Ts, Ti);
117 Tw = VFNMS(T5, T8, Tv);
118 T3M = VFMA(Tw, Ti, T3L);
119 T3Q = VFNMS(Tw, Te, T3P);
120 }
121 {
122 V T2r, T2v, T3m, T3q;
123 T2r = VMUL(T1w, Te);
124 T2v = VMUL(T1w, Ti);
125 T1z = VFMA(T5, T8, Tv);
126 T2s = VFMA(T1z, Ti, T2r);
127 T2w = VFNMS(T1z, Te, T2v);
128 T3m = VMUL(T19, Te);
129 T3q = VMUL(T19, Ti);
130 T1d = VFMA(T6, T8, T1c);
131 T3n = VFMA(T1d, Ti, T3m);
132 T3r = VFNMS(T1d, Te, T3q);
133 }
134 T2S = VMUL(T23, Te);
135 T2W = VMUL(T23, Ti);
136 T26 = VFNMS(T6, T8, T1c);
137 T2T = VFMA(T26, Ti, T2S);
138 T2X = VFNMS(T26, Te, T2W);
139 {
140 V TQ, TV, Tf, Tm, Tg;
141 Tg = VMUL(T7, Tc);
142 Th = VFMA(Tb, T8, Tg);
143 TR = VFNMS(Tb, T8, Tg);
144 TP = VFMA(Tb, Tc, T9);
145 TQ = VMUL(TP, Te);
146 TV = VMUL(TP, Ti);
147 Td = VFNMS(Tb, Tc, T9);
148 Tf = VMUL(Td, Te);
149 Tm = VMUL(Td, Ti);
150 Tj = VFMA(Th, Ti, Tf);
151 TW = VFNMS(TR, Te, TV);
152 Tn = VFNMS(Th, Te, Tm);
153 TS = VFMA(TR, Ti, TQ);
154 }
155 {
156 V T2a, T2f, T1S, T1Y, T1T;
157 T1T = VMUL(TJ, Tc);
158 T1U = VFMA(TM, T8, T1T);
159 T2b = VFNMS(TM, T8, T1T);
160 T29 = VFMA(TM, Tc, T1Q);
161 T2a = VMUL(T29, Te);
162 T2f = VMUL(T29, Ti);
163 T1R = VFNMS(TM, Tc, T1Q);
164 T1S = VMUL(T1R, Te);
165 T1Y = VMUL(T1R, Ti);
166 T1V = VFMA(T1U, Ti, T1S);
167 T2g = VFNMS(T2b, Te, T2f);
168 T1Z = VFNMS(T1U, Te, T1Y);
169 T2c = VFMA(T2b, Ti, T2a);
170 }
171 }
172 }
173 {
174 V Tq, T46, T8H, T97, TH, T98, T4b, T8D, TZ, T7f, T4j, T6t, T1g, T7g, T4q;
175 V T6u, T1v, T1I, T7m, T7j, T7k, T7l, T4z, T6x, T4G, T6y, T22, T2j, T7o, T7p;
176 V T7q, T7r, T4O, T6A, T4V, T6B, T3G, T7L, T7I, T8n, T5E, T6P, T61, T6M, T2N;
177 V T7A, T7x, T8i, T55, T6I, T5s, T6F, T43, T7J, T7O, T8o, T5L, T62, T5S, T63;
178 V T3c, T7y, T7D, T8j, T5c, T5t, T5j, T5u;
179 {
180 V T1, T8G, Tk, Tl, To, T8E, Tp, T8F;
181 T1 = LD(&(ri[0]), ms, &(ri[0]));
182 T8G = LD(&(ii[0]), ms, &(ii[0]));
183 Tk = LD(&(ri[WS(rs, 16)]), ms, &(ri[0]));
184 Tl = VMUL(Tj, Tk);
185 To = LD(&(ii[WS(rs, 16)]), ms, &(ii[0]));
186 T8E = VMUL(Tj, To);
187 Tp = VFMA(Tn, To, Tl);
188 Tq = VADD(T1, Tp);
189 T46 = VSUB(T1, Tp);
190 T8F = VFNMS(Tn, Tk, T8E);
191 T8H = VADD(T8F, T8G);
192 T97 = VSUB(T8G, T8F);
193 }
194 {
195 V Tt, Tu, Tx, T47, TB, TC, TF, T49;
196 Tt = LD(&(ri[WS(rs, 8)]), ms, &(ri[0]));
197 Tu = VMUL(Ts, Tt);
198 Tx = LD(&(ii[WS(rs, 8)]), ms, &(ii[0]));
199 T47 = VMUL(Ts, Tx);
200 TB = LD(&(ri[WS(rs, 24)]), ms, &(ri[0]));
201 TC = VMUL(TA, TB);
202 TF = LD(&(ii[WS(rs, 24)]), ms, &(ii[0]));
203 T49 = VMUL(TA, TF);
204 {
205 V Ty, TG, T48, T4a;
206 Ty = VFMA(Tw, Tx, Tu);
207 TG = VFMA(TE, TF, TC);
208 TH = VADD(Ty, TG);
209 T98 = VSUB(Ty, TG);
210 T48 = VFNMS(Tw, Tt, T47);
211 T4a = VFNMS(TE, TB, T49);
212 T4b = VSUB(T48, T4a);
213 T8D = VADD(T48, T4a);
214 }
215 }
216 {
217 V TO, T4f, TY, T4h, T4d, T4i;
218 {
219 V TK, TL, TN, T4e;
220 TK = LD(&(ri[WS(rs, 4)]), ms, &(ri[0]));
221 TL = VMUL(TJ, TK);
222 TN = LD(&(ii[WS(rs, 4)]), ms, &(ii[0]));
223 T4e = VMUL(TJ, TN);
224 TO = VFMA(TM, TN, TL);
225 T4f = VFNMS(TM, TK, T4e);
226 }
227 {
228 V TT, TU, TX, T4g;
229 TT = LD(&(ri[WS(rs, 20)]), ms, &(ri[0]));
230 TU = VMUL(TS, TT);
231 TX = LD(&(ii[WS(rs, 20)]), ms, &(ii[0]));
232 T4g = VMUL(TS, TX);
233 TY = VFMA(TW, TX, TU);
234 T4h = VFNMS(TW, TT, T4g);
235 }
236 TZ = VADD(TO, TY);
237 T7f = VADD(T4f, T4h);
238 T4d = VSUB(TO, TY);
239 T4i = VSUB(T4f, T4h);
240 T4j = VADD(T4d, T4i);
241 T6t = VSUB(T4i, T4d);
242 }
243 {
244 V T17, T4m, T1f, T4o, T4k, T4p;
245 {
246 V T12, T13, T16, T4l;
247 T12 = LD(&(ri[WS(rs, 28)]), ms, &(ri[0]));
248 T13 = VMUL(T11, T12);
249 T16 = LD(&(ii[WS(rs, 28)]), ms, &(ii[0]));
250 T4l = VMUL(T11, T16);
251 T17 = VFMA(T15, T16, T13);
252 T4m = VFNMS(T15, T12, T4l);
253 }
254 {
255 V T1a, T1b, T1e, T4n;
256 T1a = LD(&(ri[WS(rs, 12)]), ms, &(ri[0]));
257 T1b = VMUL(T19, T1a);
258 T1e = LD(&(ii[WS(rs, 12)]), ms, &(ii[0]));
259 T4n = VMUL(T19, T1e);
260 T1f = VFMA(T1d, T1e, T1b);
261 T4o = VFNMS(T1d, T1a, T4n);
262 }
263 T1g = VADD(T17, T1f);
264 T7g = VADD(T4m, T4o);
265 T4k = VSUB(T17, T1f);
266 T4p = VSUB(T4m, T4o);
267 T4q = VSUB(T4k, T4p);
268 T6u = VADD(T4k, T4p);
269 }
270 {
271 V T1m, T4u, T1H, T4E, T1u, T4w, T1B, T4C;
272 {
273 V T1j, T1k, T1l, T4t;
274 T1j = LD(&(ri[WS(rs, 2)]), ms, &(ri[0]));
275 T1k = VMUL(T7, T1j);
276 T1l = LD(&(ii[WS(rs, 2)]), ms, &(ii[0]));
277 T4t = VMUL(T7, T1l);
278 T1m = VFMA(Tb, T1l, T1k);
279 T4u = VFNMS(Tb, T1j, T4t);
280 }
281 {
282 V T1D, T1E, T1G, T4D;
283 T1D = LD(&(ri[WS(rs, 26)]), ms, &(ri[0]));
284 T1E = VMUL(T1C, T1D);
285 T1G = LD(&(ii[WS(rs, 26)]), ms, &(ii[0]));
286 T4D = VMUL(T1C, T1G);
287 T1H = VFMA(T1F, T1G, T1E);
288 T4E = VFNMS(T1F, T1D, T4D);
289 }
290 {
291 V T1p, T1q, T1t, T4v;
292 T1p = LD(&(ri[WS(rs, 18)]), ms, &(ri[0]));
293 T1q = VMUL(T1o, T1p);
294 T1t = LD(&(ii[WS(rs, 18)]), ms, &(ii[0]));
295 T4v = VMUL(T1o, T1t);
296 T1u = VFMA(T1s, T1t, T1q);
297 T4w = VFNMS(T1s, T1p, T4v);
298 }
299 {
300 V T1x, T1y, T1A, T4B;
301 T1x = LD(&(ri[WS(rs, 10)]), ms, &(ri[0]));
302 T1y = VMUL(T1w, T1x);
303 T1A = LD(&(ii[WS(rs, 10)]), ms, &(ii[0]));
304 T4B = VMUL(T1w, T1A);
305 T1B = VFMA(T1z, T1A, T1y);
306 T4C = VFNMS(T1z, T1x, T4B);
307 }
308 T1v = VADD(T1m, T1u);
309 T1I = VADD(T1B, T1H);
310 T7m = VSUB(T1v, T1I);
311 T7j = VADD(T4u, T4w);
312 T7k = VADD(T4C, T4E);
313 T7l = VSUB(T7j, T7k);
314 {
315 V T4x, T4y, T4A, T4F;
316 T4x = VSUB(T4u, T4w);
317 T4y = VSUB(T1B, T1H);
318 T4z = VSUB(T4x, T4y);
319 T6x = VADD(T4x, T4y);
320 T4A = VSUB(T1m, T1u);
321 T4F = VSUB(T4C, T4E);
322 T4G = VADD(T4A, T4F);
323 T6y = VSUB(T4A, T4F);
324 }
325 }
326 {
327 V T1P, T4J, T2i, T4T, T21, T4L, T28, T4R;
328 {
329 V T1L, T1M, T1O, T4I;
330 T1L = LD(&(ri[WS(rs, 30)]), ms, &(ri[0]));
331 T1M = VMUL(T1K, T1L);
332 T1O = LD(&(ii[WS(rs, 30)]), ms, &(ii[0]));
333 T4I = VMUL(T1K, T1O);
334 T1P = VFMA(T1N, T1O, T1M);
335 T4J = VFNMS(T1N, T1L, T4I);
336 }
337 {
338 V T2d, T2e, T2h, T4S;
339 T2d = LD(&(ri[WS(rs, 22)]), ms, &(ri[0]));
340 T2e = VMUL(T2c, T2d);
341 T2h = LD(&(ii[WS(rs, 22)]), ms, &(ii[0]));
342 T4S = VMUL(T2c, T2h);
343 T2i = VFMA(T2g, T2h, T2e);
344 T4T = VFNMS(T2g, T2d, T4S);
345 }
346 {
347 V T1W, T1X, T20, T4K;
348 T1W = LD(&(ri[WS(rs, 14)]), ms, &(ri[0]));
349 T1X = VMUL(T1V, T1W);
350 T20 = LD(&(ii[WS(rs, 14)]), ms, &(ii[0]));
351 T4K = VMUL(T1V, T20);
352 T21 = VFMA(T1Z, T20, T1X);
353 T4L = VFNMS(T1Z, T1W, T4K);
354 }
355 {
356 V T24, T25, T27, T4Q;
357 T24 = LD(&(ri[WS(rs, 6)]), ms, &(ri[0]));
358 T25 = VMUL(T23, T24);
359 T27 = LD(&(ii[WS(rs, 6)]), ms, &(ii[0]));
360 T4Q = VMUL(T23, T27);
361 T28 = VFMA(T26, T27, T25);
362 T4R = VFNMS(T26, T24, T4Q);
363 }
364 T22 = VADD(T1P, T21);
365 T2j = VADD(T28, T2i);
366 T7o = VSUB(T22, T2j);
367 T7p = VADD(T4J, T4L);
368 T7q = VADD(T4R, T4T);
369 T7r = VSUB(T7p, T7q);
370 {
371 V T4M, T4N, T4P, T4U;
372 T4M = VSUB(T4J, T4L);
373 T4N = VSUB(T28, T2i);
374 T4O = VSUB(T4M, T4N);
375 T6A = VADD(T4M, T4N);
376 T4P = VSUB(T1P, T21);
377 T4U = VSUB(T4R, T4T);
378 T4V = VADD(T4P, T4U);
379 T6B = VSUB(T4P, T4U);
380 }
381 }
382 {
383 V T3l, T5z, T3E, T5Z, T3t, T5B, T3y, T5X;
384 {
385 V T3g, T3h, T3k, T5y;
386 T3g = LD(&(ri[WS(rs, 31)]), ms, &(ri[WS(rs, 1)]));
387 T3h = VMUL(T3f, T3g);
388 T3k = LD(&(ii[WS(rs, 31)]), ms, &(ii[WS(rs, 1)]));
389 T5y = VMUL(T3f, T3k);
390 T3l = VFMA(T3j, T3k, T3h);
391 T5z = VFNMS(T3j, T3g, T5y);
392 }
393 {
394 V T3A, T3B, T3D, T5Y;
395 T3A = LD(&(ri[WS(rs, 23)]), ms, &(ri[WS(rs, 1)]));
396 T3B = VMUL(T3z, T3A);
397 T3D = LD(&(ii[WS(rs, 23)]), ms, &(ii[WS(rs, 1)]));
398 T5Y = VMUL(T3z, T3D);
399 T3E = VFMA(T3C, T3D, T3B);
400 T5Z = VFNMS(T3C, T3A, T5Y);
401 }
402 {
403 V T3o, T3p, T3s, T5A;
404 T3o = LD(&(ri[WS(rs, 15)]), ms, &(ri[WS(rs, 1)]));
405 T3p = VMUL(T3n, T3o);
406 T3s = LD(&(ii[WS(rs, 15)]), ms, &(ii[WS(rs, 1)]));
407 T5A = VMUL(T3n, T3s);
408 T3t = VFMA(T3r, T3s, T3p);
409 T5B = VFNMS(T3r, T3o, T5A);
410 }
411 {
412 V T3v, T3w, T3x, T5W;
413 T3v = LD(&(ri[WS(rs, 7)]), ms, &(ri[WS(rs, 1)]));
414 T3w = VMUL(TP, T3v);
415 T3x = LD(&(ii[WS(rs, 7)]), ms, &(ii[WS(rs, 1)]));
416 T5W = VMUL(TP, T3x);
417 T3y = VFMA(TR, T3x, T3w);
418 T5X = VFNMS(TR, T3v, T5W);
419 }
420 {
421 V T3u, T3F, T7G, T7H;
422 T3u = VADD(T3l, T3t);
423 T3F = VADD(T3y, T3E);
424 T3G = VADD(T3u, T3F);
425 T7L = VSUB(T3u, T3F);
426 T7G = VADD(T5z, T5B);
427 T7H = VADD(T5X, T5Z);
428 T7I = VSUB(T7G, T7H);
429 T8n = VADD(T7G, T7H);
430 }
431 {
432 V T5C, T5D, T5V, T60;
433 T5C = VSUB(T5z, T5B);
434 T5D = VSUB(T3y, T3E);
435 T5E = VSUB(T5C, T5D);
436 T6P = VADD(T5C, T5D);
437 T5V = VSUB(T3l, T3t);
438 T60 = VSUB(T5X, T5Z);
439 T61 = VADD(T5V, T60);
440 T6M = VSUB(T5V, T60);
441 }
442 }
443 {
444 V T2q, T50, T2L, T5q, T2y, T52, T2D, T5o;
445 {
446 V T2n, T2o, T2p, T4Z;
447 T2n = LD(&(ri[WS(rs, 1)]), ms, &(ri[WS(rs, 1)]));
448 T2o = VMUL(T2, T2n);
449 T2p = LD(&(ii[WS(rs, 1)]), ms, &(ii[WS(rs, 1)]));
450 T4Z = VMUL(T2, T2p);
451 T2q = VFMA(T5, T2p, T2o);
452 T50 = VFNMS(T5, T2n, T4Z);
453 }
454 {
455 V T2G, T2H, T2K, T5p;
456 T2G = LD(&(ri[WS(rs, 25)]), ms, &(ri[WS(rs, 1)]));
457 T2H = VMUL(T2F, T2G);
458 T2K = LD(&(ii[WS(rs, 25)]), ms, &(ii[WS(rs, 1)]));
459 T5p = VMUL(T2F, T2K);
460 T2L = VFMA(T2J, T2K, T2H);
461 T5q = VFNMS(T2J, T2G, T5p);
462 }
463 {
464 V T2t, T2u, T2x, T51;
465 T2t = LD(&(ri[WS(rs, 17)]), ms, &(ri[WS(rs, 1)]));
466 T2u = VMUL(T2s, T2t);
467 T2x = LD(&(ii[WS(rs, 17)]), ms, &(ii[WS(rs, 1)]));
468 T51 = VMUL(T2s, T2x);
469 T2y = VFMA(T2w, T2x, T2u);
470 T52 = VFNMS(T2w, T2t, T51);
471 }
472 {
473 V T2A, T2B, T2C, T5n;
474 T2A = LD(&(ri[WS(rs, 9)]), ms, &(ri[WS(rs, 1)]));
475 T2B = VMUL(T8, T2A);
476 T2C = LD(&(ii[WS(rs, 9)]), ms, &(ii[WS(rs, 1)]));
477 T5n = VMUL(T8, T2C);
478 T2D = VFMA(Tc, T2C, T2B);
479 T5o = VFNMS(Tc, T2A, T5n);
480 }
481 {
482 V T2z, T2M, T7v, T7w;
483 T2z = VADD(T2q, T2y);
484 T2M = VADD(T2D, T2L);
485 T2N = VADD(T2z, T2M);
486 T7A = VSUB(T2z, T2M);
487 T7v = VADD(T50, T52);
488 T7w = VADD(T5o, T5q);
489 T7x = VSUB(T7v, T7w);
490 T8i = VADD(T7v, T7w);
491 }
492 {
493 V T53, T54, T5m, T5r;
494 T53 = VSUB(T50, T52);
495 T54 = VSUB(T2D, T2L);
496 T55 = VSUB(T53, T54);
497 T6I = VADD(T53, T54);
498 T5m = VSUB(T2q, T2y);
499 T5r = VSUB(T5o, T5q);
500 T5s = VADD(T5m, T5r);
501 T6F = VSUB(T5m, T5r);
502 }
503 }
504 {
505 V T3K, T5G, T41, T5Q, T3S, T5I, T3X, T5O;
506 {
507 V T3H, T3I, T3J, T5F;
508 T3H = LD(&(ri[WS(rs, 3)]), ms, &(ri[WS(rs, 1)]));
509 T3I = VMUL(T3, T3H);
510 T3J = LD(&(ii[WS(rs, 3)]), ms, &(ii[WS(rs, 1)]));
511 T5F = VMUL(T3, T3J);
512 T3K = VFMA(T6, T3J, T3I);
513 T5G = VFNMS(T6, T3H, T5F);
514 }
515 {
516 V T3Y, T3Z, T40, T5P;
517 T3Y = LD(&(ri[WS(rs, 11)]), ms, &(ri[WS(rs, 1)]));
518 T3Z = VMUL(Td, T3Y);
519 T40 = LD(&(ii[WS(rs, 11)]), ms, &(ii[WS(rs, 1)]));
520 T5P = VMUL(Td, T40);
521 T41 = VFMA(Th, T40, T3Z);
522 T5Q = VFNMS(Th, T3Y, T5P);
523 }
524 {
525 V T3N, T3O, T3R, T5H;
526 T3N = LD(&(ri[WS(rs, 19)]), ms, &(ri[WS(rs, 1)]));
527 T3O = VMUL(T3M, T3N);
528 T3R = LD(&(ii[WS(rs, 19)]), ms, &(ii[WS(rs, 1)]));
529 T5H = VMUL(T3M, T3R);
530 T3S = VFMA(T3Q, T3R, T3O);
531 T5I = VFNMS(T3Q, T3N, T5H);
532 }
533 {
534 V T3U, T3V, T3W, T5N;
535 T3U = LD(&(ri[WS(rs, 27)]), ms, &(ri[WS(rs, 1)]));
536 T3V = VMUL(Te, T3U);
537 T3W = LD(&(ii[WS(rs, 27)]), ms, &(ii[WS(rs, 1)]));
538 T5N = VMUL(Te, T3W);
539 T3X = VFMA(Ti, T3W, T3V);
540 T5O = VFNMS(Ti, T3U, T5N);
541 }
542 {
543 V T3T, T42, T7M, T7N;
544 T3T = VADD(T3K, T3S);
545 T42 = VADD(T3X, T41);
546 T43 = VADD(T3T, T42);
547 T7J = VSUB(T42, T3T);
548 T7M = VADD(T5G, T5I);
549 T7N = VADD(T5O, T5Q);
550 T7O = VSUB(T7M, T7N);
551 T8o = VADD(T7M, T7N);
552 }
553 {
554 V T5J, T5K, T5M, T5R;
555 T5J = VSUB(T5G, T5I);
556 T5K = VSUB(T3K, T3S);
557 T5L = VSUB(T5J, T5K);
558 T62 = VADD(T5K, T5J);
559 T5M = VSUB(T3X, T41);
560 T5R = VSUB(T5O, T5Q);
561 T5S = VADD(T5M, T5R);
562 T63 = VSUB(T5M, T5R);
563 }
564 }
565 {
566 V T2R, T57, T3a, T5h, T2Z, T59, T36, T5f;
567 {
568 V T2O, T2P, T2Q, T56;
569 T2O = LD(&(ri[WS(rs, 5)]), ms, &(ri[WS(rs, 1)]));
570 T2P = VMUL(T29, T2O);
571 T2Q = LD(&(ii[WS(rs, 5)]), ms, &(ii[WS(rs, 1)]));
572 T56 = VMUL(T29, T2Q);
573 T2R = VFMA(T2b, T2Q, T2P);
574 T57 = VFNMS(T2b, T2O, T56);
575 }
576 {
577 V T37, T38, T39, T5g;
578 T37 = LD(&(ri[WS(rs, 13)]), ms, &(ri[WS(rs, 1)]));
579 T38 = VMUL(T1R, T37);
580 T39 = LD(&(ii[WS(rs, 13)]), ms, &(ii[WS(rs, 1)]));
581 T5g = VMUL(T1R, T39);
582 T3a = VFMA(T1U, T39, T38);
583 T5h = VFNMS(T1U, T37, T5g);
584 }
585 {
586 V T2U, T2V, T2Y, T58;
587 T2U = LD(&(ri[WS(rs, 21)]), ms, &(ri[WS(rs, 1)]));
588 T2V = VMUL(T2T, T2U);
589 T2Y = LD(&(ii[WS(rs, 21)]), ms, &(ii[WS(rs, 1)]));
590 T58 = VMUL(T2T, T2Y);
591 T2Z = VFMA(T2X, T2Y, T2V);
592 T59 = VFNMS(T2X, T2U, T58);
593 }
594 {
595 V T32, T33, T35, T5e;
596 T32 = LD(&(ri[WS(rs, 29)]), ms, &(ri[WS(rs, 1)]));
597 T33 = VMUL(T31, T32);
598 T35 = LD(&(ii[WS(rs, 29)]), ms, &(ii[WS(rs, 1)]));
599 T5e = VMUL(T31, T35);
600 T36 = VFMA(T34, T35, T33);
601 T5f = VFNMS(T34, T32, T5e);
602 }
603 {
604 V T30, T3b, T7B, T7C;
605 T30 = VADD(T2R, T2Z);
606 T3b = VADD(T36, T3a);
607 T3c = VADD(T30, T3b);
608 T7y = VSUB(T3b, T30);
609 T7B = VADD(T57, T59);
610 T7C = VADD(T5f, T5h);
611 T7D = VSUB(T7B, T7C);
612 T8j = VADD(T7B, T7C);
613 }
614 {
615 V T5a, T5b, T5d, T5i;
616 T5a = VSUB(T57, T59);
617 T5b = VSUB(T2R, T2Z);
618 T5c = VSUB(T5a, T5b);
619 T5t = VADD(T5b, T5a);
620 T5d = VSUB(T36, T3a);
621 T5i = VSUB(T5f, T5h);
622 T5j = VADD(T5d, T5i);
623 T5u = VSUB(T5d, T5i);
624 }
625 }
626 {
627 V T1i, T8c, T8z, T8A, T8J, T8O, T2l, T8N, T45, T8L, T8l, T8t, T8q, T8u, T8f;
628 V T8B;
629 {
630 V TI, T1h, T8x, T8y;
631 TI = VADD(Tq, TH);
632 T1h = VADD(TZ, T1g);
633 T1i = VADD(TI, T1h);
634 T8c = VSUB(TI, T1h);
635 T8x = VADD(T8i, T8j);
636 T8y = VADD(T8n, T8o);
637 T8z = VSUB(T8x, T8y);
638 T8A = VADD(T8x, T8y);
639 }
640 {
641 V T8C, T8I, T1J, T2k;
642 T8C = VADD(T7f, T7g);
643 T8I = VADD(T8D, T8H);
644 T8J = VADD(T8C, T8I);
645 T8O = VSUB(T8I, T8C);
646 T1J = VADD(T1v, T1I);
647 T2k = VADD(T22, T2j);
648 T2l = VADD(T1J, T2k);
649 T8N = VSUB(T2k, T1J);
650 }
651 {
652 V T3d, T44, T8h, T8k;
653 T3d = VADD(T2N, T3c);
654 T44 = VADD(T3G, T43);
655 T45 = VADD(T3d, T44);
656 T8L = VSUB(T44, T3d);
657 T8h = VSUB(T2N, T3c);
658 T8k = VSUB(T8i, T8j);
659 T8l = VADD(T8h, T8k);
660 T8t = VSUB(T8k, T8h);
661 }
662 {
663 V T8m, T8p, T8d, T8e;
664 T8m = VSUB(T3G, T43);
665 T8p = VSUB(T8n, T8o);
666 T8q = VSUB(T8m, T8p);
667 T8u = VADD(T8m, T8p);
668 T8d = VADD(T7j, T7k);
669 T8e = VADD(T7p, T7q);
670 T8f = VSUB(T8d, T8e);
671 T8B = VADD(T8d, T8e);
672 }
673 {
674 V T2m, T8K, T8w, T8M;
675 T2m = VADD(T1i, T2l);
676 ST(&(ri[WS(rs, 16)]), VSUB(T2m, T45), ms, &(ri[0]));
677 ST(&(ri[0]), VADD(T2m, T45), ms, &(ri[0]));
678 T8K = VADD(T8B, T8J);
679 ST(&(ii[0]), VADD(T8A, T8K), ms, &(ii[0]));
680 ST(&(ii[WS(rs, 16)]), VSUB(T8K, T8A), ms, &(ii[0]));
681 T8w = VSUB(T1i, T2l);
682 ST(&(ri[WS(rs, 24)]), VSUB(T8w, T8z), ms, &(ri[0]));
683 ST(&(ri[WS(rs, 8)]), VADD(T8w, T8z), ms, &(ri[0]));
684 T8M = VSUB(T8J, T8B);
685 ST(&(ii[WS(rs, 8)]), VADD(T8L, T8M), ms, &(ii[0]));
686 ST(&(ii[WS(rs, 24)]), VSUB(T8M, T8L), ms, &(ii[0]));
687 }
688 {
689 V T8g, T8r, T8P, T8Q;
690 T8g = VADD(T8c, T8f);
691 T8r = VADD(T8l, T8q);
692 ST(&(ri[WS(rs, 20)]), VFNMS(LDK(KP707106781), T8r, T8g), ms, &(ri[0]));
693 ST(&(ri[WS(rs, 4)]), VFMA(LDK(KP707106781), T8r, T8g), ms, &(ri[0]));
694 T8P = VADD(T8N, T8O);
695 T8Q = VADD(T8t, T8u);
696 ST(&(ii[WS(rs, 4)]), VFMA(LDK(KP707106781), T8Q, T8P), ms, &(ii[0]));
697 ST(&(ii[WS(rs, 20)]), VFNMS(LDK(KP707106781), T8Q, T8P), ms, &(ii[0]));
698 }
699 {
700 V T8s, T8v, T8R, T8S;
701 T8s = VSUB(T8c, T8f);
702 T8v = VSUB(T8t, T8u);
703 ST(&(ri[WS(rs, 28)]), VFNMS(LDK(KP707106781), T8v, T8s), ms, &(ri[0]));
704 ST(&(ri[WS(rs, 12)]), VFMA(LDK(KP707106781), T8v, T8s), ms, &(ri[0]));
705 T8R = VSUB(T8O, T8N);
706 T8S = VSUB(T8q, T8l);
707 ST(&(ii[WS(rs, 12)]), VFMA(LDK(KP707106781), T8S, T8R), ms, &(ii[0]));
708 ST(&(ii[WS(rs, 28)]), VFNMS(LDK(KP707106781), T8S, T8R), ms, &(ii[0]));
709 }
710 }
711 {
712 V T7i, T7W, T86, T8a, T8V, T91, T7t, T8W, T7F, T7T, T7Z, T92, T83, T89, T7Q;
713 V T7U;
714 {
715 V T7e, T7h, T84, T85;
716 T7e = VSUB(Tq, TH);
717 T7h = VSUB(T7f, T7g);
718 T7i = VSUB(T7e, T7h);
719 T7W = VADD(T7e, T7h);
720 T84 = VADD(T7L, T7O);
721 T85 = VADD(T7I, T7J);
722 T86 = VFNMS(LDK(KP414213562), T85, T84);
723 T8a = VFMA(LDK(KP414213562), T84, T85);
724 }
725 {
726 V T8T, T8U, T7n, T7s;
727 T8T = VSUB(T1g, TZ);
728 T8U = VSUB(T8H, T8D);
729 T8V = VADD(T8T, T8U);
730 T91 = VSUB(T8U, T8T);
731 T7n = VSUB(T7l, T7m);
732 T7s = VADD(T7o, T7r);
733 T7t = VSUB(T7n, T7s);
734 T8W = VADD(T7n, T7s);
735 }
736 {
737 V T7z, T7E, T7X, T7Y;
738 T7z = VSUB(T7x, T7y);
739 T7E = VSUB(T7A, T7D);
740 T7F = VFMA(LDK(KP414213562), T7E, T7z);
741 T7T = VFNMS(LDK(KP414213562), T7z, T7E);
742 T7X = VADD(T7m, T7l);
743 T7Y = VSUB(T7o, T7r);
744 T7Z = VADD(T7X, T7Y);
745 T92 = VSUB(T7Y, T7X);
746 }
747 {
748 V T81, T82, T7K, T7P;
749 T81 = VADD(T7A, T7D);
750 T82 = VADD(T7x, T7y);
751 T83 = VFMA(LDK(KP414213562), T82, T81);
752 T89 = VFNMS(LDK(KP414213562), T81, T82);
753 T7K = VSUB(T7I, T7J);
754 T7P = VSUB(T7L, T7O);
755 T7Q = VFNMS(LDK(KP414213562), T7P, T7K);
756 T7U = VFMA(LDK(KP414213562), T7K, T7P);
757 }
758 {
759 V T7u, T7R, T93, T94;
760 T7u = VFMA(LDK(KP707106781), T7t, T7i);
761 T7R = VSUB(T7F, T7Q);
762 ST(&(ri[WS(rs, 22)]), VFNMS(LDK(KP923879532), T7R, T7u), ms, &(ri[0]));
763 ST(&(ri[WS(rs, 6)]), VFMA(LDK(KP923879532), T7R, T7u), ms, &(ri[0]));
764 T93 = VFMA(LDK(KP707106781), T92, T91);
765 T94 = VSUB(T7U, T7T);
766 ST(&(ii[WS(rs, 6)]), VFMA(LDK(KP923879532), T94, T93), ms, &(ii[0]));
767 ST(&(ii[WS(rs, 22)]), VFNMS(LDK(KP923879532), T94, T93), ms, &(ii[0]));
768 }
769 {
770 V T7S, T7V, T95, T96;
771 T7S = VFNMS(LDK(KP707106781), T7t, T7i);
772 T7V = VADD(T7T, T7U);
773 ST(&(ri[WS(rs, 14)]), VFNMS(LDK(KP923879532), T7V, T7S), ms, &(ri[0]));
774 ST(&(ri[WS(rs, 30)]), VFMA(LDK(KP923879532), T7V, T7S), ms, &(ri[0]));
775 T95 = VFNMS(LDK(KP707106781), T92, T91);
776 T96 = VADD(T7F, T7Q);
777 ST(&(ii[WS(rs, 14)]), VFNMS(LDK(KP923879532), T96, T95), ms, &(ii[0]));
778 ST(&(ii[WS(rs, 30)]), VFMA(LDK(KP923879532), T96, T95), ms, &(ii[0]));
779 }
780 {
781 V T80, T87, T8X, T8Y;
782 T80 = VFMA(LDK(KP707106781), T7Z, T7W);
783 T87 = VADD(T83, T86);
784 ST(&(ri[WS(rs, 18)]), VFNMS(LDK(KP923879532), T87, T80), ms, &(ri[0]));
785 ST(&(ri[WS(rs, 2)]), VFMA(LDK(KP923879532), T87, T80), ms, &(ri[0]));
786 T8X = VFMA(LDK(KP707106781), T8W, T8V);
787 T8Y = VADD(T89, T8a);
788 ST(&(ii[WS(rs, 2)]), VFMA(LDK(KP923879532), T8Y, T8X), ms, &(ii[0]));
789 ST(&(ii[WS(rs, 18)]), VFNMS(LDK(KP923879532), T8Y, T8X), ms, &(ii[0]));
790 }
791 {
792 V T88, T8b, T8Z, T90;
793 T88 = VFNMS(LDK(KP707106781), T7Z, T7W);
794 T8b = VSUB(T89, T8a);
795 ST(&(ri[WS(rs, 26)]), VFNMS(LDK(KP923879532), T8b, T88), ms, &(ri[0]));
796 ST(&(ri[WS(rs, 10)]), VFMA(LDK(KP923879532), T8b, T88), ms, &(ri[0]));
797 T8Z = VFNMS(LDK(KP707106781), T8W, T8V);
798 T90 = VSUB(T86, T83);
799 ST(&(ii[WS(rs, 10)]), VFMA(LDK(KP923879532), T90, T8Z), ms, &(ii[0]));
800 ST(&(ii[WS(rs, 26)]), VFNMS(LDK(KP923879532), T90, T8Z), ms, &(ii[0]));
801 }
802 }
803 {
804 V T4s, T6c, T4X, T9c, T9b, T9h, T6f, T9i, T66, T6q, T6a, T6m, T5x, T6p, T69;
805 V T6j;
806 {
807 V T4c, T4r, T6d, T6e;
808 T4c = VADD(T46, T4b);
809 T4r = VADD(T4j, T4q);
810 T4s = VFNMS(LDK(KP707106781), T4r, T4c);
811 T6c = VFMA(LDK(KP707106781), T4r, T4c);
812 {
813 V T4H, T4W, T99, T9a;
814 T4H = VFNMS(LDK(KP414213562), T4G, T4z);
815 T4W = VFMA(LDK(KP414213562), T4V, T4O);
816 T4X = VSUB(T4H, T4W);
817 T9c = VADD(T4H, T4W);
818 T99 = VSUB(T97, T98);
819 T9a = VADD(T6t, T6u);
820 T9b = VFMA(LDK(KP707106781), T9a, T99);
821 T9h = VFNMS(LDK(KP707106781), T9a, T99);
822 }
823 T6d = VFMA(LDK(KP414213562), T4z, T4G);
824 T6e = VFNMS(LDK(KP414213562), T4O, T4V);
825 T6f = VADD(T6d, T6e);
826 T9i = VSUB(T6e, T6d);
827 {
828 V T5U, T6l, T65, T6k, T5T, T64;
829 T5T = VADD(T5L, T5S);
830 T5U = VFNMS(LDK(KP707106781), T5T, T5E);
831 T6l = VFMA(LDK(KP707106781), T5T, T5E);
832 T64 = VADD(T62, T63);
833 T65 = VFNMS(LDK(KP707106781), T64, T61);
834 T6k = VFMA(LDK(KP707106781), T64, T61);
835 T66 = VFNMS(LDK(KP668178637), T65, T5U);
836 T6q = VFMA(LDK(KP198912367), T6k, T6l);
837 T6a = VFMA(LDK(KP668178637), T5U, T65);
838 T6m = VFNMS(LDK(KP198912367), T6l, T6k);
839 }
840 {
841 V T5l, T6i, T5w, T6h, T5k, T5v;
842 T5k = VADD(T5c, T5j);
843 T5l = VFNMS(LDK(KP707106781), T5k, T55);
844 T6i = VFMA(LDK(KP707106781), T5k, T55);
845 T5v = VADD(T5t, T5u);
846 T5w = VFNMS(LDK(KP707106781), T5v, T5s);
847 T6h = VFMA(LDK(KP707106781), T5v, T5s);
848 T5x = VFMA(LDK(KP668178637), T5w, T5l);
849 T6p = VFNMS(LDK(KP198912367), T6h, T6i);
850 T69 = VFNMS(LDK(KP668178637), T5l, T5w);
851 T6j = VFMA(LDK(KP198912367), T6i, T6h);
852 }
853 }
854 {
855 V T4Y, T67, T9j, T9k;
856 T4Y = VFMA(LDK(KP923879532), T4X, T4s);
857 T67 = VSUB(T5x, T66);
858 ST(&(ri[WS(rs, 21)]), VFNMS(LDK(KP831469612), T67, T4Y), ms, &(ri[WS(rs, 1)]));
859 ST(&(ri[WS(rs, 5)]), VFMA(LDK(KP831469612), T67, T4Y), ms, &(ri[WS(rs, 1)]));
860 T9j = VFMA(LDK(KP923879532), T9i, T9h);
861 T9k = VSUB(T6a, T69);
862 ST(&(ii[WS(rs, 5)]), VFMA(LDK(KP831469612), T9k, T9j), ms, &(ii[WS(rs, 1)]));
863 ST(&(ii[WS(rs, 21)]), VFNMS(LDK(KP831469612), T9k, T9j), ms, &(ii[WS(rs, 1)]));
864 }
865 {
866 V T68, T6b, T9l, T9m;
867 T68 = VFNMS(LDK(KP923879532), T4X, T4s);
868 T6b = VADD(T69, T6a);
869 ST(&(ri[WS(rs, 13)]), VFNMS(LDK(KP831469612), T6b, T68), ms, &(ri[WS(rs, 1)]));
870 ST(&(ri[WS(rs, 29)]), VFMA(LDK(KP831469612), T6b, T68), ms, &(ri[WS(rs, 1)]));
871 T9l = VFNMS(LDK(KP923879532), T9i, T9h);
872 T9m = VADD(T5x, T66);
873 ST(&(ii[WS(rs, 13)]), VFNMS(LDK(KP831469612), T9m, T9l), ms, &(ii[WS(rs, 1)]));
874 ST(&(ii[WS(rs, 29)]), VFMA(LDK(KP831469612), T9m, T9l), ms, &(ii[WS(rs, 1)]));
875 }
876 {
877 V T6g, T6n, T9d, T9e;
878 T6g = VFMA(LDK(KP923879532), T6f, T6c);
879 T6n = VADD(T6j, T6m);
880 ST(&(ri[WS(rs, 17)]), VFNMS(LDK(KP980785280), T6n, T6g), ms, &(ri[WS(rs, 1)]));
881 ST(&(ri[WS(rs, 1)]), VFMA(LDK(KP980785280), T6n, T6g), ms, &(ri[WS(rs, 1)]));
882 T9d = VFMA(LDK(KP923879532), T9c, T9b);
883 T9e = VADD(T6p, T6q);
884 ST(&(ii[WS(rs, 1)]), VFMA(LDK(KP980785280), T9e, T9d), ms, &(ii[WS(rs, 1)]));
885 ST(&(ii[WS(rs, 17)]), VFNMS(LDK(KP980785280), T9e, T9d), ms, &(ii[WS(rs, 1)]));
886 }
887 {
888 V T6o, T6r, T9f, T9g;
889 T6o = VFNMS(LDK(KP923879532), T6f, T6c);
890 T6r = VSUB(T6p, T6q);
891 ST(&(ri[WS(rs, 25)]), VFNMS(LDK(KP980785280), T6r, T6o), ms, &(ri[WS(rs, 1)]));
892 ST(&(ri[WS(rs, 9)]), VFMA(LDK(KP980785280), T6r, T6o), ms, &(ri[WS(rs, 1)]));
893 T9f = VFNMS(LDK(KP923879532), T9c, T9b);
894 T9g = VSUB(T6m, T6j);
895 ST(&(ii[WS(rs, 9)]), VFMA(LDK(KP980785280), T9g, T9f), ms, &(ii[WS(rs, 1)]));
896 ST(&(ii[WS(rs, 25)]), VFNMS(LDK(KP980785280), T9g, T9f), ms, &(ii[WS(rs, 1)]));
897 }
898 }
899 {
900 V T6w, T6Y, T6D, T9w, T9p, T9v, T71, T9q, T6S, T7c, T6W, T78, T6L, T7b, T6V;
901 V T75;
902 {
903 V T6s, T6v, T6Z, T70;
904 T6s = VSUB(T46, T4b);
905 T6v = VSUB(T6t, T6u);
906 T6w = VFMA(LDK(KP707106781), T6v, T6s);
907 T6Y = VFNMS(LDK(KP707106781), T6v, T6s);
908 {
909 V T6z, T6C, T9n, T9o;
910 T6z = VFMA(LDK(KP414213562), T6y, T6x);
911 T6C = VFNMS(LDK(KP414213562), T6B, T6A);
912 T6D = VSUB(T6z, T6C);
913 T9w = VADD(T6z, T6C);
914 T9n = VADD(T98, T97);
915 T9o = VSUB(T4q, T4j);
916 T9p = VFMA(LDK(KP707106781), T9o, T9n);
917 T9v = VFNMS(LDK(KP707106781), T9o, T9n);
918 }
919 T6Z = VFNMS(LDK(KP414213562), T6x, T6y);
920 T70 = VFMA(LDK(KP414213562), T6A, T6B);
921 T71 = VADD(T6Z, T70);
922 T9q = VSUB(T70, T6Z);
923 {
924 V T6O, T77, T6R, T76, T6N, T6Q;
925 T6N = VSUB(T5S, T5L);
926 T6O = VFNMS(LDK(KP707106781), T6N, T6M);
927 T77 = VFMA(LDK(KP707106781), T6N, T6M);
928 T6Q = VSUB(T62, T63);
929 T6R = VFNMS(LDK(KP707106781), T6Q, T6P);
930 T76 = VFMA(LDK(KP707106781), T6Q, T6P);
931 T6S = VFNMS(LDK(KP668178637), T6R, T6O);
932 T7c = VFMA(LDK(KP198912367), T76, T77);
933 T6W = VFMA(LDK(KP668178637), T6O, T6R);
934 T78 = VFNMS(LDK(KP198912367), T77, T76);
935 }
936 {
937 V T6H, T74, T6K, T73, T6G, T6J;
938 T6G = VSUB(T5j, T5c);
939 T6H = VFNMS(LDK(KP707106781), T6G, T6F);
940 T74 = VFMA(LDK(KP707106781), T6G, T6F);
941 T6J = VSUB(T5t, T5u);
942 T6K = VFNMS(LDK(KP707106781), T6J, T6I);
943 T73 = VFMA(LDK(KP707106781), T6J, T6I);
944 T6L = VFMA(LDK(KP668178637), T6K, T6H);
945 T7b = VFNMS(LDK(KP198912367), T73, T74);
946 T6V = VFNMS(LDK(KP668178637), T6H, T6K);
947 T75 = VFMA(LDK(KP198912367), T74, T73);
948 }
949 }
950 {
951 V T6E, T6T, T9r, T9s;
952 T6E = VFMA(LDK(KP923879532), T6D, T6w);
953 T6T = VADD(T6L, T6S);
954 ST(&(ri[WS(rs, 19)]), VFNMS(LDK(KP831469612), T6T, T6E), ms, &(ri[WS(rs, 1)]));
955 ST(&(ri[WS(rs, 3)]), VFMA(LDK(KP831469612), T6T, T6E), ms, &(ri[WS(rs, 1)]));
956 T9r = VFMA(LDK(KP923879532), T9q, T9p);
957 T9s = VADD(T6V, T6W);
958 ST(&(ii[WS(rs, 3)]), VFMA(LDK(KP831469612), T9s, T9r), ms, &(ii[WS(rs, 1)]));
959 ST(&(ii[WS(rs, 19)]), VFNMS(LDK(KP831469612), T9s, T9r), ms, &(ii[WS(rs, 1)]));
960 }
961 {
962 V T6U, T6X, T9t, T9u;
963 T6U = VFNMS(LDK(KP923879532), T6D, T6w);
964 T6X = VSUB(T6V, T6W);
965 ST(&(ri[WS(rs, 27)]), VFNMS(LDK(KP831469612), T6X, T6U), ms, &(ri[WS(rs, 1)]));
966 ST(&(ri[WS(rs, 11)]), VFMA(LDK(KP831469612), T6X, T6U), ms, &(ri[WS(rs, 1)]));
967 T9t = VFNMS(LDK(KP923879532), T9q, T9p);
968 T9u = VSUB(T6S, T6L);
969 ST(&(ii[WS(rs, 11)]), VFMA(LDK(KP831469612), T9u, T9t), ms, &(ii[WS(rs, 1)]));
970 ST(&(ii[WS(rs, 27)]), VFNMS(LDK(KP831469612), T9u, T9t), ms, &(ii[WS(rs, 1)]));
971 }
972 {
973 V T72, T79, T9x, T9y;
974 T72 = VFNMS(LDK(KP923879532), T71, T6Y);
975 T79 = VSUB(T75, T78);
976 ST(&(ri[WS(rs, 23)]), VFNMS(LDK(KP980785280), T79, T72), ms, &(ri[WS(rs, 1)]));
977 ST(&(ri[WS(rs, 7)]), VFMA(LDK(KP980785280), T79, T72), ms, &(ri[WS(rs, 1)]));
978 T9x = VFNMS(LDK(KP923879532), T9w, T9v);
979 T9y = VSUB(T7c, T7b);
980 ST(&(ii[WS(rs, 7)]), VFMA(LDK(KP980785280), T9y, T9x), ms, &(ii[WS(rs, 1)]));
981 ST(&(ii[WS(rs, 23)]), VFNMS(LDK(KP980785280), T9y, T9x), ms, &(ii[WS(rs, 1)]));
982 }
983 {
984 V T7a, T7d, T9z, T9A;
985 T7a = VFMA(LDK(KP923879532), T71, T6Y);
986 T7d = VADD(T7b, T7c);
987 ST(&(ri[WS(rs, 15)]), VFNMS(LDK(KP980785280), T7d, T7a), ms, &(ri[WS(rs, 1)]));
988 ST(&(ri[WS(rs, 31)]), VFMA(LDK(KP980785280), T7d, T7a), ms, &(ri[WS(rs, 1)]));
989 T9z = VFMA(LDK(KP923879532), T9w, T9v);
990 T9A = VADD(T75, T78);
991 ST(&(ii[WS(rs, 15)]), VFNMS(LDK(KP980785280), T9A, T9z), ms, &(ii[WS(rs, 1)]));
992 ST(&(ii[WS(rs, 31)]), VFMA(LDK(KP980785280), T9A, T9z), ms, &(ii[WS(rs, 1)]));
993 }
994 }
995 }
996 }
997 }
998 VLEAVE();
999 }
1000
1001 static const tw_instr twinstr[] = {
1002 VTW(0, 1),
1003 VTW(0, 3),
1004 VTW(0, 9),
1005 VTW(0, 27),
1006 {TW_NEXT, (2 * VL), 0}
1007 };
1008
1009 static const ct_desc desc = { 32, XSIMD_STRING("t2sv_32"), twinstr, &GENUS, {236, 98, 252, 0}, 0, 0, 0 };
1010
1011 void XSIMD(codelet_t2sv_32) (planner *p) {
1012 X(kdft_dit_register) (p, t2sv_32, &desc);
1013 }
1014 #else
1015
1016 /* Generated by: ../../../genfft/gen_twiddle.native -simd -compact -variables 4 -pipeline-latency 8 -twiddle-log3 -precompute-twiddles -n 32 -name t2sv_32 -include dft/simd/ts.h */
1017
1018 /*
1019 * This function contains 488 FP additions, 280 FP multiplications,
1020 * (or, 376 additions, 168 multiplications, 112 fused multiply/add),
1021 * 158 stack variables, 7 constants, and 128 memory accesses
1022 */
1023 #include "dft/simd/ts.h"
1024
1025 static void t2sv_32(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
1026 {
1027 DVK(KP195090322, +0.195090322016128267848284868477022240927691618);
1028 DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
1029 DVK(KP555570233, +0.555570233019602224742830813948532874374937191);
1030 DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
1031 DVK(KP382683432, +0.382683432365089771728459984030398866761344562);
1032 DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
1033 DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
1034 {
1035 INT m;
1036 for (m = mb, W = W + (mb * 8); m < me; m = m + (2 * VL), ri = ri + ((2 * VL) * ms), ii = ii + ((2 * VL) * ms), W = W + ((2 * VL) * 8), MAKE_VOLATILE_STRIDE(64, rs)) {
1037 V T2, T5, T3, T6, T8, TM, TO, Td, T9, Te, Th, Tl, TD, TH, T1y;
1038 V T1H, T15, T1A, T11, T1F, T1n, T1p, T2q, T2I, T2u, T2K, T2V, T3b, T2Z, T3d;
1039 V Tu, Ty, T3l, T3n, T1t, T1v, T2f, T2h, T1a, T1e, T32, T34, T1W, T1Y, T2C;
1040 V T2E, Tg, TR, Tk, TS, Tm, TV, To, TT, T1M, T21, T1P, T22, T1Q, T25;
1041 V T1S, T23;
1042 {
1043 V Ts, T1d, Tx, T18, Tt, T1c, Tw, T19, TB, T14, TG, TZ, TC, T13, TF;
1044 V T10;
1045 {
1046 V T4, Tc, T7, Tb;
1047 T2 = LDW(&(W[0]));
1048 T5 = LDW(&(W[TWVL * 1]));
1049 T3 = LDW(&(W[TWVL * 2]));
1050 T6 = LDW(&(W[TWVL * 3]));
1051 T4 = VMUL(T2, T3);
1052 Tc = VMUL(T5, T3);
1053 T7 = VMUL(T5, T6);
1054 Tb = VMUL(T2, T6);
1055 T8 = VADD(T4, T7);
1056 TM = VSUB(T4, T7);
1057 TO = VADD(Tb, Tc);
1058 Td = VSUB(Tb, Tc);
1059 T9 = LDW(&(W[TWVL * 4]));
1060 Ts = VMUL(T2, T9);
1061 T1d = VMUL(T6, T9);
1062 Tx = VMUL(T5, T9);
1063 T18 = VMUL(T3, T9);
1064 Te = LDW(&(W[TWVL * 5]));
1065 Tt = VMUL(T5, Te);
1066 T1c = VMUL(T3, Te);
1067 Tw = VMUL(T2, Te);
1068 T19 = VMUL(T6, Te);
1069 Th = LDW(&(W[TWVL * 6]));
1070 TB = VMUL(T3, Th);
1071 T14 = VMUL(T5, Th);
1072 TG = VMUL(T6, Th);
1073 TZ = VMUL(T2, Th);
1074 Tl = LDW(&(W[TWVL * 7]));
1075 TC = VMUL(T6, Tl);
1076 T13 = VMUL(T2, Tl);
1077 TF = VMUL(T3, Tl);
1078 T10 = VMUL(T5, Tl);
1079 }
1080 TD = VADD(TB, TC);
1081 TH = VSUB(TF, TG);
1082 T1y = VADD(TZ, T10);
1083 T1H = VADD(TF, TG);
1084 T15 = VADD(T13, T14);
1085 T1A = VSUB(T13, T14);
1086 T11 = VSUB(TZ, T10);
1087 T1F = VSUB(TB, TC);
1088 T1n = VFMA(T9, Th, VMUL(Te, Tl));
1089 T1p = VFNMS(Te, Th, VMUL(T9, Tl));
1090 {
1091 V T2o, T2p, T2s, T2t;
1092 T2o = VMUL(T8, Th);
1093 T2p = VMUL(Td, Tl);
1094 T2q = VADD(T2o, T2p);
1095 T2I = VSUB(T2o, T2p);
1096 T2s = VMUL(T8, Tl);
1097 T2t = VMUL(Td, Th);
1098 T2u = VSUB(T2s, T2t);
1099 T2K = VADD(T2s, T2t);
1100 }
1101 {
1102 V T2T, T2U, T2X, T2Y;
1103 T2T = VMUL(TM, Th);
1104 T2U = VMUL(TO, Tl);
1105 T2V = VSUB(T2T, T2U);
1106 T3b = VADD(T2T, T2U);
1107 T2X = VMUL(TM, Tl);
1108 T2Y = VMUL(TO, Th);
1109 T2Z = VADD(T2X, T2Y);
1110 T3d = VSUB(T2X, T2Y);
1111 Tu = VADD(Ts, Tt);
1112 Ty = VSUB(Tw, Tx);
1113 T3l = VFMA(Tu, Th, VMUL(Ty, Tl));
1114 T3n = VFNMS(Ty, Th, VMUL(Tu, Tl));
1115 }
1116 T1t = VSUB(Ts, Tt);
1117 T1v = VADD(Tw, Tx);
1118 T2f = VFMA(T1t, Th, VMUL(T1v, Tl));
1119 T2h = VFNMS(T1v, Th, VMUL(T1t, Tl));
1120 T1a = VSUB(T18, T19);
1121 T1e = VADD(T1c, T1d);
1122 T32 = VFMA(T1a, Th, VMUL(T1e, Tl));
1123 T34 = VFNMS(T1e, Th, VMUL(T1a, Tl));
1124 T1W = VADD(T18, T19);
1125 T1Y = VSUB(T1c, T1d);
1126 T2C = VFMA(T1W, Th, VMUL(T1Y, Tl));
1127 T2E = VFNMS(T1Y, Th, VMUL(T1W, Tl));
1128 {
1129 V Ta, Tf, Ti, Tj;
1130 Ta = VMUL(T8, T9);
1131 Tf = VMUL(Td, Te);
1132 Tg = VSUB(Ta, Tf);
1133 TR = VADD(Ta, Tf);
1134 Ti = VMUL(T8, Te);
1135 Tj = VMUL(Td, T9);
1136 Tk = VADD(Ti, Tj);
1137 TS = VSUB(Ti, Tj);
1138 }
1139 Tm = VFMA(Tg, Th, VMUL(Tk, Tl));
1140 TV = VFNMS(TS, Th, VMUL(TR, Tl));
1141 To = VFNMS(Tk, Th, VMUL(Tg, Tl));
1142 TT = VFMA(TR, Th, VMUL(TS, Tl));
1143 {
1144 V T1K, T1L, T1N, T1O;
1145 T1K = VMUL(TM, T9);
1146 T1L = VMUL(TO, Te);
1147 T1M = VSUB(T1K, T1L);
1148 T21 = VADD(T1K, T1L);
1149 T1N = VMUL(TM, Te);
1150 T1O = VMUL(TO, T9);
1151 T1P = VADD(T1N, T1O);
1152 T22 = VSUB(T1N, T1O);
1153 }
1154 T1Q = VFMA(T1M, Th, VMUL(T1P, Tl));
1155 T25 = VFNMS(T22, Th, VMUL(T21, Tl));
1156 T1S = VFNMS(T1P, Th, VMUL(T1M, Tl));
1157 T23 = VFMA(T21, Th, VMUL(T22, Tl));
1158 }
1159 {
1160 V TL, T6f, T8c, T8q, T3F, T5t, T7I, T7W, T2y, T6B, T6y, T7j, T4k, T5J, T4B;
1161 V T5G, T3h, T6H, T6O, T7o, T4L, T5N, T52, T5Q, T1i, T7V, T6i, T7D, T3K, T5u;
1162 V T3P, T5v, T1E, T6n, T6m, T7e, T3W, T5y, T41, T5z, T29, T6p, T6s, T7f, T47;
1163 V T5B, T4c, T5C, T2R, T6z, T6E, T7k, T4v, T5H, T4E, T5K, T3y, T6P, T6K, T7p;
1164 V T4W, T5R, T55, T5O;
1165 {
1166 V T1, T7G, Tq, T7F, TA, T3C, TJ, T3D, Tn, Tp;
1167 T1 = LD(&(ri[0]), ms, &(ri[0]));
1168 T7G = LD(&(ii[0]), ms, &(ii[0]));
1169 Tn = LD(&(ri[WS(rs, 16)]), ms, &(ri[0]));
1170 Tp = LD(&(ii[WS(rs, 16)]), ms, &(ii[0]));
1171 Tq = VFMA(Tm, Tn, VMUL(To, Tp));
1172 T7F = VFNMS(To, Tn, VMUL(Tm, Tp));
1173 {
1174 V Tv, Tz, TE, TI;
1175 Tv = LD(&(ri[WS(rs, 8)]), ms, &(ri[0]));
1176 Tz = LD(&(ii[WS(rs, 8)]), ms, &(ii[0]));
1177 TA = VFMA(Tu, Tv, VMUL(Ty, Tz));
1178 T3C = VFNMS(Ty, Tv, VMUL(Tu, Tz));
1179 TE = LD(&(ri[WS(rs, 24)]), ms, &(ri[0]));
1180 TI = LD(&(ii[WS(rs, 24)]), ms, &(ii[0]));
1181 TJ = VFMA(TD, TE, VMUL(TH, TI));
1182 T3D = VFNMS(TH, TE, VMUL(TD, TI));
1183 }
1184 {
1185 V Tr, TK, T8a, T8b;
1186 Tr = VADD(T1, Tq);
1187 TK = VADD(TA, TJ);
1188 TL = VADD(Tr, TK);
1189 T6f = VSUB(Tr, TK);
1190 T8a = VSUB(T7G, T7F);
1191 T8b = VSUB(TA, TJ);
1192 T8c = VSUB(T8a, T8b);
1193 T8q = VADD(T8b, T8a);
1194 }
1195 {
1196 V T3B, T3E, T7E, T7H;
1197 T3B = VSUB(T1, Tq);
1198 T3E = VSUB(T3C, T3D);
1199 T3F = VSUB(T3B, T3E);
1200 T5t = VADD(T3B, T3E);
1201 T7E = VADD(T3C, T3D);
1202 T7H = VADD(T7F, T7G);
1203 T7I = VADD(T7E, T7H);
1204 T7W = VSUB(T7H, T7E);
1205 }
1206 }
1207 {
1208 V T2e, T4g, T2w, T4z, T2j, T4h, T2n, T4y;
1209 {
1210 V T2c, T2d, T2r, T2v;
1211 T2c = LD(&(ri[WS(rs, 1)]), ms, &(ri[WS(rs, 1)]));
1212 T2d = LD(&(ii[WS(rs, 1)]), ms, &(ii[WS(rs, 1)]));
1213 T2e = VFMA(T2, T2c, VMUL(T5, T2d));
1214 T4g = VFNMS(T5, T2c, VMUL(T2, T2d));
1215 T2r = LD(&(ri[WS(rs, 25)]), ms, &(ri[WS(rs, 1)]));
1216 T2v = LD(&(ii[WS(rs, 25)]), ms, &(ii[WS(rs, 1)]));
1217 T2w = VFMA(T2q, T2r, VMUL(T2u, T2v));
1218 T4z = VFNMS(T2u, T2r, VMUL(T2q, T2v));
1219 }
1220 {
1221 V T2g, T2i, T2l, T2m;
1222 T2g = LD(&(ri[WS(rs, 17)]), ms, &(ri[WS(rs, 1)]));
1223 T2i = LD(&(ii[WS(rs, 17)]), ms, &(ii[WS(rs, 1)]));
1224 T2j = VFMA(T2f, T2g, VMUL(T2h, T2i));
1225 T4h = VFNMS(T2h, T2g, VMUL(T2f, T2i));
1226 T2l = LD(&(ri[WS(rs, 9)]), ms, &(ri[WS(rs, 1)]));
1227 T2m = LD(&(ii[WS(rs, 9)]), ms, &(ii[WS(rs, 1)]));
1228 T2n = VFMA(T9, T2l, VMUL(Te, T2m));
1229 T4y = VFNMS(Te, T2l, VMUL(T9, T2m));
1230 }
1231 {
1232 V T2k, T2x, T6w, T6x;
1233 T2k = VADD(T2e, T2j);
1234 T2x = VADD(T2n, T2w);
1235 T2y = VADD(T2k, T2x);
1236 T6B = VSUB(T2k, T2x);
1237 T6w = VADD(T4g, T4h);
1238 T6x = VADD(T4y, T4z);
1239 T6y = VSUB(T6w, T6x);
1240 T7j = VADD(T6w, T6x);
1241 }
1242 {
1243 V T4i, T4j, T4x, T4A;
1244 T4i = VSUB(T4g, T4h);
1245 T4j = VSUB(T2n, T2w);
1246 T4k = VADD(T4i, T4j);
1247 T5J = VSUB(T4i, T4j);
1248 T4x = VSUB(T2e, T2j);
1249 T4A = VSUB(T4y, T4z);
1250 T4B = VSUB(T4x, T4A);
1251 T5G = VADD(T4x, T4A);
1252 }
1253 }
1254 {
1255 V T31, T4Y, T3f, T4J, T36, T4Z, T3a, T4I;
1256 {
1257 V T2W, T30, T3c, T3e;
1258 T2W = LD(&(ri[WS(rs, 31)]), ms, &(ri[WS(rs, 1)]));
1259 T30 = LD(&(ii[WS(rs, 31)]), ms, &(ii[WS(rs, 1)]));
1260 T31 = VFMA(T2V, T2W, VMUL(T2Z, T30));
1261 T4Y = VFNMS(T2Z, T2W, VMUL(T2V, T30));
1262 T3c = LD(&(ri[WS(rs, 23)]), ms, &(ri[WS(rs, 1)]));
1263 T3e = LD(&(ii[WS(rs, 23)]), ms, &(ii[WS(rs, 1)]));
1264 T3f = VFMA(T3b, T3c, VMUL(T3d, T3e));
1265 T4J = VFNMS(T3d, T3c, VMUL(T3b, T3e));
1266 }
1267 {
1268 V T33, T35, T38, T39;
1269 T33 = LD(&(ri[WS(rs, 15)]), ms, &(ri[WS(rs, 1)]));
1270 T35 = LD(&(ii[WS(rs, 15)]), ms, &(ii[WS(rs, 1)]));
1271 T36 = VFMA(T32, T33, VMUL(T34, T35));
1272 T4Z = VFNMS(T34, T33, VMUL(T32, T35));
1273 T38 = LD(&(ri[WS(rs, 7)]), ms, &(ri[WS(rs, 1)]));
1274 T39 = LD(&(ii[WS(rs, 7)]), ms, &(ii[WS(rs, 1)]));
1275 T3a = VFMA(TR, T38, VMUL(TS, T39));
1276 T4I = VFNMS(TS, T38, VMUL(TR, T39));
1277 }
1278 {
1279 V T37, T3g, T6M, T6N;
1280 T37 = VADD(T31, T36);
1281 T3g = VADD(T3a, T3f);
1282 T3h = VADD(T37, T3g);
1283 T6H = VSUB(T37, T3g);
1284 T6M = VADD(T4Y, T4Z);
1285 T6N = VADD(T4I, T4J);
1286 T6O = VSUB(T6M, T6N);
1287 T7o = VADD(T6M, T6N);
1288 }
1289 {
1290 V T4H, T4K, T50, T51;
1291 T4H = VSUB(T31, T36);
1292 T4K = VSUB(T4I, T4J);
1293 T4L = VSUB(T4H, T4K);
1294 T5N = VADD(T4H, T4K);
1295 T50 = VSUB(T4Y, T4Z);
1296 T51 = VSUB(T3a, T3f);
1297 T52 = VADD(T50, T51);
1298 T5Q = VSUB(T50, T51);
1299 }
1300 }
1301 {
1302 V TQ, T3G, T1g, T3N, TX, T3H, T17, T3M;
1303 {
1304 V TN, TP, T1b, T1f;
1305 TN = LD(&(ri[WS(rs, 4)]), ms, &(ri[0]));
1306 TP = LD(&(ii[WS(rs, 4)]), ms, &(ii[0]));
1307 TQ = VFMA(TM, TN, VMUL(TO, TP));
1308 T3G = VFNMS(TO, TN, VMUL(TM, TP));
1309 T1b = LD(&(ri[WS(rs, 12)]), ms, &(ri[0]));
1310 T1f = LD(&(ii[WS(rs, 12)]), ms, &(ii[0]));
1311 T1g = VFMA(T1a, T1b, VMUL(T1e, T1f));
1312 T3N = VFNMS(T1e, T1b, VMUL(T1a, T1f));
1313 }
1314 {
1315 V TU, TW, T12, T16;
1316 TU = LD(&(ri[WS(rs, 20)]), ms, &(ri[0]));
1317 TW = LD(&(ii[WS(rs, 20)]), ms, &(ii[0]));
1318 TX = VFMA(TT, TU, VMUL(TV, TW));
1319 T3H = VFNMS(TV, TU, VMUL(TT, TW));
1320 T12 = LD(&(ri[WS(rs, 28)]), ms, &(ri[0]));
1321 T16 = LD(&(ii[WS(rs, 28)]), ms, &(ii[0]));
1322 T17 = VFMA(T11, T12, VMUL(T15, T16));
1323 T3M = VFNMS(T15, T12, VMUL(T11, T16));
1324 }
1325 {
1326 V TY, T1h, T6g, T6h;
1327 TY = VADD(TQ, TX);
1328 T1h = VADD(T17, T1g);
1329 T1i = VADD(TY, T1h);
1330 T7V = VSUB(T1h, TY);
1331 T6g = VADD(T3G, T3H);
1332 T6h = VADD(T3M, T3N);
1333 T6i = VSUB(T6g, T6h);
1334 T7D = VADD(T6g, T6h);
1335 }
1336 {
1337 V T3I, T3J, T3L, T3O;
1338 T3I = VSUB(T3G, T3H);
1339 T3J = VSUB(TQ, TX);
1340 T3K = VSUB(T3I, T3J);
1341 T5u = VADD(T3J, T3I);
1342 T3L = VSUB(T17, T1g);
1343 T3O = VSUB(T3M, T3N);
1344 T3P = VADD(T3L, T3O);
1345 T5v = VSUB(T3L, T3O);
1346 }
1347 }
1348 {
1349 V T1m, T3S, T1C, T3Z, T1r, T3T, T1x, T3Y;
1350 {
1351 V T1k, T1l, T1z, T1B;
1352 T1k = LD(&(ri[WS(rs, 2)]), ms, &(ri[0]));
1353 T1l = LD(&(ii[WS(rs, 2)]), ms, &(ii[0]));
1354 T1m = VFMA(T8, T1k, VMUL(Td, T1l));
1355 T3S = VFNMS(Td, T1k, VMUL(T8, T1l));
1356 T1z = LD(&(ri[WS(rs, 26)]), ms, &(ri[0]));
1357 T1B = LD(&(ii[WS(rs, 26)]), ms, &(ii[0]));
1358 T1C = VFMA(T1y, T1z, VMUL(T1A, T1B));
1359 T3Z = VFNMS(T1A, T1z, VMUL(T1y, T1B));
1360 }
1361 {
1362 V T1o, T1q, T1u, T1w;
1363 T1o = LD(&(ri[WS(rs, 18)]), ms, &(ri[0]));
1364 T1q = LD(&(ii[WS(rs, 18)]), ms, &(ii[0]));
1365 T1r = VFMA(T1n, T1o, VMUL(T1p, T1q));
1366 T3T = VFNMS(T1p, T1o, VMUL(T1n, T1q));
1367 T1u = LD(&(ri[WS(rs, 10)]), ms, &(ri[0]));
1368 T1w = LD(&(ii[WS(rs, 10)]), ms, &(ii[0]));
1369 T1x = VFMA(T1t, T1u, VMUL(T1v, T1w));
1370 T3Y = VFNMS(T1v, T1u, VMUL(T1t, T1w));
1371 }
1372 {
1373 V T1s, T1D, T6k, T6l;
1374 T1s = VADD(T1m, T1r);
1375 T1D = VADD(T1x, T1C);
1376 T1E = VADD(T1s, T1D);
1377 T6n = VSUB(T1s, T1D);
1378 T6k = VADD(T3S, T3T);
1379 T6l = VADD(T3Y, T3Z);
1380 T6m = VSUB(T6k, T6l);
1381 T7e = VADD(T6k, T6l);
1382 }
1383 {
1384 V T3U, T3V, T3X, T40;
1385 T3U = VSUB(T3S, T3T);
1386 T3V = VSUB(T1x, T1C);
1387 T3W = VADD(T3U, T3V);
1388 T5y = VSUB(T3U, T3V);
1389 T3X = VSUB(T1m, T1r);
1390 T40 = VSUB(T3Y, T3Z);
1391 T41 = VSUB(T3X, T40);
1392 T5z = VADD(T3X, T40);
1393 }
1394 }
1395 {
1396 V T1J, T43, T27, T4a, T1U, T44, T20, T49;
1397 {
1398 V T1G, T1I, T24, T26;
1399 T1G = LD(&(ri[WS(rs, 30)]), ms, &(ri[0]));
1400 T1I = LD(&(ii[WS(rs, 30)]), ms, &(ii[0]));
1401 T1J = VFMA(T1F, T1G, VMUL(T1H, T1I));
1402 T43 = VFNMS(T1H, T1G, VMUL(T1F, T1I));
1403 T24 = LD(&(ri[WS(rs, 22)]), ms, &(ri[0]));
1404 T26 = LD(&(ii[WS(rs, 22)]), ms, &(ii[0]));
1405 T27 = VFMA(T23, T24, VMUL(T25, T26));
1406 T4a = VFNMS(T25, T24, VMUL(T23, T26));
1407 }
1408 {
1409 V T1R, T1T, T1X, T1Z;
1410 T1R = LD(&(ri[WS(rs, 14)]), ms, &(ri[0]));
1411 T1T = LD(&(ii[WS(rs, 14)]), ms, &(ii[0]));
1412 T1U = VFMA(T1Q, T1R, VMUL(T1S, T1T));
1413 T44 = VFNMS(T1S, T1R, VMUL(T1Q, T1T));
1414 T1X = LD(&(ri[WS(rs, 6)]), ms, &(ri[0]));
1415 T1Z = LD(&(ii[WS(rs, 6)]), ms, &(ii[0]));
1416 T20 = VFMA(T1W, T1X, VMUL(T1Y, T1Z));
1417 T49 = VFNMS(T1Y, T1X, VMUL(T1W, T1Z));
1418 }
1419 {
1420 V T1V, T28, T6q, T6r;
1421 T1V = VADD(T1J, T1U);
1422 T28 = VADD(T20, T27);
1423 T29 = VADD(T1V, T28);
1424 T6p = VSUB(T1V, T28);
1425 T6q = VADD(T43, T44);
1426 T6r = VADD(T49, T4a);
1427 T6s = VSUB(T6q, T6r);
1428 T7f = VADD(T6q, T6r);
1429 }
1430 {
1431 V T45, T46, T48, T4b;
1432 T45 = VSUB(T43, T44);
1433 T46 = VSUB(T20, T27);
1434 T47 = VADD(T45, T46);
1435 T5B = VSUB(T45, T46);
1436 T48 = VSUB(T1J, T1U);
1437 T4b = VSUB(T49, T4a);
1438 T4c = VSUB(T48, T4b);
1439 T5C = VADD(T48, T4b);
1440 }
1441 }
1442 {
1443 V T2B, T4r, T2G, T4s, T4q, T4t, T2M, T4m, T2P, T4n, T4l, T4o;
1444 {
1445 V T2z, T2A, T2D, T2F;
1446 T2z = LD(&(ri[WS(rs, 5)]), ms, &(ri[WS(rs, 1)]));
1447 T2A = LD(&(ii[WS(rs, 5)]), ms, &(ii[WS(rs, 1)]));
1448 T2B = VFMA(T21, T2z, VMUL(T22, T2A));
1449 T4r = VFNMS(T22, T2z, VMUL(T21, T2A));
1450 T2D = LD(&(ri[WS(rs, 21)]), ms, &(ri[WS(rs, 1)]));
1451 T2F = LD(&(ii[WS(rs, 21)]), ms, &(ii[WS(rs, 1)]));
1452 T2G = VFMA(T2C, T2D, VMUL(T2E, T2F));
1453 T4s = VFNMS(T2E, T2D, VMUL(T2C, T2F));
1454 }
1455 T4q = VSUB(T2B, T2G);
1456 T4t = VSUB(T4r, T4s);
1457 {
1458 V T2J, T2L, T2N, T2O;
1459 T2J = LD(&(ri[WS(rs, 29)]), ms, &(ri[WS(rs, 1)]));
1460 T2L = LD(&(ii[WS(rs, 29)]), ms, &(ii[WS(rs, 1)]));
1461 T2M = VFMA(T2I, T2J, VMUL(T2K, T2L));
1462 T4m = VFNMS(T2K, T2J, VMUL(T2I, T2L));
1463 T2N = LD(&(ri[WS(rs, 13)]), ms, &(ri[WS(rs, 1)]));
1464 T2O = LD(&(ii[WS(rs, 13)]), ms, &(ii[WS(rs, 1)]));
1465 T2P = VFMA(T1M, T2N, VMUL(T1P, T2O));
1466 T4n = VFNMS(T1P, T2N, VMUL(T1M, T2O));
1467 }
1468 T4l = VSUB(T2M, T2P);
1469 T4o = VSUB(T4m, T4n);
1470 {
1471 V T2H, T2Q, T6C, T6D;
1472 T2H = VADD(T2B, T2G);
1473 T2Q = VADD(T2M, T2P);
1474 T2R = VADD(T2H, T2Q);
1475 T6z = VSUB(T2Q, T2H);
1476 T6C = VADD(T4r, T4s);
1477 T6D = VADD(T4m, T4n);
1478 T6E = VSUB(T6C, T6D);
1479 T7k = VADD(T6C, T6D);
1480 }
1481 {
1482 V T4p, T4u, T4C, T4D;
1483 T4p = VSUB(T4l, T4o);
1484 T4u = VADD(T4q, T4t);
1485 T4v = VMUL(LDK(KP707106781), VSUB(T4p, T4u));
1486 T5H = VMUL(LDK(KP707106781), VADD(T4u, T4p));
1487 T4C = VSUB(T4t, T4q);
1488 T4D = VADD(T4l, T4o);
1489 T4E = VMUL(LDK(KP707106781), VSUB(T4C, T4D));
1490 T5K = VMUL(LDK(KP707106781), VADD(T4C, T4D));
1491 }
1492 }
1493 {
1494 V T3k, T4M, T3p, T4N, T4O, T4P, T3t, T4S, T3w, T4T, T4R, T4U;
1495 {
1496 V T3i, T3j, T3m, T3o;
1497 T3i = LD(&(ri[WS(rs, 3)]), ms, &(ri[WS(rs, 1)]));
1498 T3j = LD(&(ii[WS(rs, 3)]), ms, &(ii[WS(rs, 1)]));
1499 T3k = VFMA(T3, T3i, VMUL(T6, T3j));
1500 T4M = VFNMS(T6, T3i, VMUL(T3, T3j));
1501 T3m = LD(&(ri[WS(rs, 19)]), ms, &(ri[WS(rs, 1)]));
1502 T3o = LD(&(ii[WS(rs, 19)]), ms, &(ii[WS(rs, 1)]));
1503 T3p = VFMA(T3l, T3m, VMUL(T3n, T3o));
1504 T4N = VFNMS(T3n, T3m, VMUL(T3l, T3o));
1505 }
1506 T4O = VSUB(T4M, T4N);
1507 T4P = VSUB(T3k, T3p);
1508 {
1509 V T3r, T3s, T3u, T3v;
1510 T3r = LD(&(ri[WS(rs, 27)]), ms, &(ri[WS(rs, 1)]));
1511 T3s = LD(&(ii[WS(rs, 27)]), ms, &(ii[WS(rs, 1)]));
1512 T3t = VFMA(Th, T3r, VMUL(Tl, T3s));
1513 T4S = VFNMS(Tl, T3r, VMUL(Th, T3s));
1514 T3u = LD(&(ri[WS(rs, 11)]), ms, &(ri[WS(rs, 1)]));
1515 T3v = LD(&(ii[WS(rs, 11)]), ms, &(ii[WS(rs, 1)]));
1516 T3w = VFMA(Tg, T3u, VMUL(Tk, T3v));
1517 T4T = VFNMS(Tk, T3u, VMUL(Tg, T3v));
1518 }
1519 T4R = VSUB(T3t, T3w);
1520 T4U = VSUB(T4S, T4T);
1521 {
1522 V T3q, T3x, T6I, T6J;
1523 T3q = VADD(T3k, T3p);
1524 T3x = VADD(T3t, T3w);
1525 T3y = VADD(T3q, T3x);
1526 T6P = VSUB(T3x, T3q);
1527 T6I = VADD(T4M, T4N);
1528 T6J = VADD(T4S, T4T);
1529 T6K = VSUB(T6I, T6J);
1530 T7p = VADD(T6I, T6J);
1531 }
1532 {
1533 V T4Q, T4V, T53, T54;
1534 T4Q = VSUB(T4O, T4P);
1535 T4V = VADD(T4R, T4U);
1536 T4W = VMUL(LDK(KP707106781), VSUB(T4Q, T4V));
1537 T5R = VMUL(LDK(KP707106781), VADD(T4Q, T4V));
1538 T53 = VSUB(T4R, T4U);
1539 T54 = VADD(T4P, T4O);
1540 T55 = VMUL(LDK(KP707106781), VSUB(T53, T54));
1541 T5O = VMUL(LDK(KP707106781), VADD(T54, T53));
1542 }
1543 }
1544 {
1545 V T2b, T7x, T7K, T7M, T3A, T7L, T7A, T7B;
1546 {
1547 V T1j, T2a, T7C, T7J;
1548 T1j = VADD(TL, T1i);
1549 T2a = VADD(T1E, T29);
1550 T2b = VADD(T1j, T2a);
1551 T7x = VSUB(T1j, T2a);
1552 T7C = VADD(T7e, T7f);
1553 T7J = VADD(T7D, T7I);
1554 T7K = VADD(T7C, T7J);
1555 T7M = VSUB(T7J, T7C);
1556 }
1557 {
1558 V T2S, T3z, T7y, T7z;
1559 T2S = VADD(T2y, T2R);
1560 T3z = VADD(T3h, T3y);
1561 T3A = VADD(T2S, T3z);
1562 T7L = VSUB(T3z, T2S);
1563 T7y = VADD(T7j, T7k);
1564 T7z = VADD(T7o, T7p);
1565 T7A = VSUB(T7y, T7z);
1566 T7B = VADD(T7y, T7z);
1567 }
1568 ST(&(ri[WS(rs, 16)]), VSUB(T2b, T3A), ms, &(ri[0]));
1569 ST(&(ii[WS(rs, 16)]), VSUB(T7K, T7B), ms, &(ii[0]));
1570 ST(&(ri[0]), VADD(T2b, T3A), ms, &(ri[0]));
1571 ST(&(ii[0]), VADD(T7B, T7K), ms, &(ii[0]));
1572 ST(&(ri[WS(rs, 24)]), VSUB(T7x, T7A), ms, &(ri[0]));
1573 ST(&(ii[WS(rs, 24)]), VSUB(T7M, T7L), ms, &(ii[0]));
1574 ST(&(ri[WS(rs, 8)]), VADD(T7x, T7A), ms, &(ri[0]));
1575 ST(&(ii[WS(rs, 8)]), VADD(T7L, T7M), ms, &(ii[0]));
1576 }
1577 {
1578 V T7h, T7t, T7Q, T7S, T7m, T7u, T7r, T7v;
1579 {
1580 V T7d, T7g, T7O, T7P;
1581 T7d = VSUB(TL, T1i);
1582 T7g = VSUB(T7e, T7f);
1583 T7h = VADD(T7d, T7g);
1584 T7t = VSUB(T7d, T7g);
1585 T7O = VSUB(T29, T1E);
1586 T7P = VSUB(T7I, T7D);
1587 T7Q = VADD(T7O, T7P);
1588 T7S = VSUB(T7P, T7O);
1589 }
1590 {
1591 V T7i, T7l, T7n, T7q;
1592 T7i = VSUB(T2y, T2R);
1593 T7l = VSUB(T7j, T7k);
1594 T7m = VADD(T7i, T7l);
1595 T7u = VSUB(T7l, T7i);
1596 T7n = VSUB(T3h, T3y);
1597 T7q = VSUB(T7o, T7p);
1598 T7r = VSUB(T7n, T7q);
1599 T7v = VADD(T7n, T7q);
1600 }
1601 {
1602 V T7s, T7N, T7w, T7R;
1603 T7s = VMUL(LDK(KP707106781), VADD(T7m, T7r));
1604 ST(&(ri[WS(rs, 20)]), VSUB(T7h, T7s), ms, &(ri[0]));
1605 ST(&(ri[WS(rs, 4)]), VADD(T7h, T7s), ms, &(ri[0]));
1606 T7N = VMUL(LDK(KP707106781), VADD(T7u, T7v));
1607 ST(&(ii[WS(rs, 4)]), VADD(T7N, T7Q), ms, &(ii[0]));
1608 ST(&(ii[WS(rs, 20)]), VSUB(T7Q, T7N), ms, &(ii[0]));
1609 T7w = VMUL(LDK(KP707106781), VSUB(T7u, T7v));
1610 ST(&(ri[WS(rs, 28)]), VSUB(T7t, T7w), ms, &(ri[0]));
1611 ST(&(ri[WS(rs, 12)]), VADD(T7t, T7w), ms, &(ri[0]));
1612 T7R = VMUL(LDK(KP707106781), VSUB(T7r, T7m));
1613 ST(&(ii[WS(rs, 12)]), VADD(T7R, T7S), ms, &(ii[0]));
1614 ST(&(ii[WS(rs, 28)]), VSUB(T7S, T7R), ms, &(ii[0]));
1615 }
1616 }
1617 {
1618 V T6j, T7X, T83, T6X, T6u, T7U, T77, T7b, T70, T82, T6G, T6U, T74, T7a, T6R;
1619 V T6V;
1620 {
1621 V T6o, T6t, T6A, T6F;
1622 T6j = VSUB(T6f, T6i);
1623 T7X = VADD(T7V, T7W);
1624 T83 = VSUB(T7W, T7V);
1625 T6X = VADD(T6f, T6i);
1626 T6o = VSUB(T6m, T6n);
1627 T6t = VADD(T6p, T6s);
1628 T6u = VMUL(LDK(KP707106781), VSUB(T6o, T6t));
1629 T7U = VMUL(LDK(KP707106781), VADD(T6o, T6t));
1630 {
1631 V T75, T76, T6Y, T6Z;
1632 T75 = VADD(T6H, T6K);
1633 T76 = VADD(T6O, T6P);
1634 T77 = VFNMS(LDK(KP382683432), T76, VMUL(LDK(KP923879532), T75));
1635 T7b = VFMA(LDK(KP923879532), T76, VMUL(LDK(KP382683432), T75));
1636 T6Y = VADD(T6n, T6m);
1637 T6Z = VSUB(T6p, T6s);
1638 T70 = VMUL(LDK(KP707106781), VADD(T6Y, T6Z));
1639 T82 = VMUL(LDK(KP707106781), VSUB(T6Z, T6Y));
1640 }
1641 T6A = VSUB(T6y, T6z);
1642 T6F = VSUB(T6B, T6E);
1643 T6G = VFMA(LDK(KP923879532), T6A, VMUL(LDK(KP382683432), T6F));
1644 T6U = VFNMS(LDK(KP923879532), T6F, VMUL(LDK(KP382683432), T6A));
1645 {
1646 V T72, T73, T6L, T6Q;
1647 T72 = VADD(T6y, T6z);
1648 T73 = VADD(T6B, T6E);
1649 T74 = VFMA(LDK(KP382683432), T72, VMUL(LDK(KP923879532), T73));
1650 T7a = VFNMS(LDK(KP382683432), T73, VMUL(LDK(KP923879532), T72));
1651 T6L = VSUB(T6H, T6K);
1652 T6Q = VSUB(T6O, T6P);
1653 T6R = VFNMS(LDK(KP923879532), T6Q, VMUL(LDK(KP382683432), T6L));
1654 T6V = VFMA(LDK(KP382683432), T6Q, VMUL(LDK(KP923879532), T6L));
1655 }
1656 }
1657 {
1658 V T6v, T6S, T81, T84;
1659 T6v = VADD(T6j, T6u);
1660 T6S = VADD(T6G, T6R);
1661 ST(&(ri[WS(rs, 22)]), VSUB(T6v, T6S), ms, &(ri[0]));
1662 ST(&(ri[WS(rs, 6)]), VADD(T6v, T6S), ms, &(ri[0]));
1663 T81 = VADD(T6U, T6V);
1664 T84 = VADD(T82, T83);
1665 ST(&(ii[WS(rs, 6)]), VADD(T81, T84), ms, &(ii[0]));
1666 ST(&(ii[WS(rs, 22)]), VSUB(T84, T81), ms, &(ii[0]));
1667 }
1668 {
1669 V T6T, T6W, T85, T86;
1670 T6T = VSUB(T6j, T6u);
1671 T6W = VSUB(T6U, T6V);
1672 ST(&(ri[WS(rs, 30)]), VSUB(T6T, T6W), ms, &(ri[0]));
1673 ST(&(ri[WS(rs, 14)]), VADD(T6T, T6W), ms, &(ri[0]));
1674 T85 = VSUB(T6R, T6G);
1675 T86 = VSUB(T83, T82);
1676 ST(&(ii[WS(rs, 14)]), VADD(T85, T86), ms, &(ii[0]));
1677 ST(&(ii[WS(rs, 30)]), VSUB(T86, T85), ms, &(ii[0]));
1678 }
1679 {
1680 V T71, T78, T7T, T7Y;
1681 T71 = VADD(T6X, T70);
1682 T78 = VADD(T74, T77);
1683 ST(&(ri[WS(rs, 18)]), VSUB(T71, T78), ms, &(ri[0]));
1684 ST(&(ri[WS(rs, 2)]), VADD(T71, T78), ms, &(ri[0]));
1685 T7T = VADD(T7a, T7b);
1686 T7Y = VADD(T7U, T7X);
1687 ST(&(ii[WS(rs, 2)]), VADD(T7T, T7Y), ms, &(ii[0]));
1688 ST(&(ii[WS(rs, 18)]), VSUB(T7Y, T7T), ms, &(ii[0]));
1689 }
1690 {
1691 V T79, T7c, T7Z, T80;
1692 T79 = VSUB(T6X, T70);
1693 T7c = VSUB(T7a, T7b);
1694 ST(&(ri[WS(rs, 26)]), VSUB(T79, T7c), ms, &(ri[0]));
1695 ST(&(ri[WS(rs, 10)]), VADD(T79, T7c), ms, &(ri[0]));
1696 T7Z = VSUB(T77, T74);
1697 T80 = VSUB(T7X, T7U);
1698 ST(&(ii[WS(rs, 10)]), VADD(T7Z, T80), ms, &(ii[0]));
1699 ST(&(ii[WS(rs, 26)]), VSUB(T80, T7Z), ms, &(ii[0]));
1700 }
1701 }
1702 {
1703 V T3R, T5d, T8r, T8x, T4e, T8o, T5n, T5r, T4G, T5a, T5g, T8w, T5k, T5q, T57;
1704 V T5b, T3Q, T8p;
1705 T3Q = VMUL(LDK(KP707106781), VSUB(T3K, T3P));
1706 T3R = VSUB(T3F, T3Q);
1707 T5d = VADD(T3F, T3Q);
1708 T8p = VMUL(LDK(KP707106781), VSUB(T5v, T5u));
1709 T8r = VADD(T8p, T8q);
1710 T8x = VSUB(T8q, T8p);
1711 {
1712 V T42, T4d, T5l, T5m;
1713 T42 = VFNMS(LDK(KP923879532), T41, VMUL(LDK(KP382683432), T3W));
1714 T4d = VFMA(LDK(KP382683432), T47, VMUL(LDK(KP923879532), T4c));
1715 T4e = VSUB(T42, T4d);
1716 T8o = VADD(T42, T4d);
1717 T5l = VADD(T4L, T4W);
1718 T5m = VADD(T52, T55);
1719 T5n = VFNMS(LDK(KP555570233), T5m, VMUL(LDK(KP831469612), T5l));
1720 T5r = VFMA(LDK(KP831469612), T5m, VMUL(LDK(KP555570233), T5l));
1721 }
1722 {
1723 V T4w, T4F, T5e, T5f;
1724 T4w = VSUB(T4k, T4v);
1725 T4F = VSUB(T4B, T4E);
1726 T4G = VFMA(LDK(KP980785280), T4w, VMUL(LDK(KP195090322), T4F));
1727 T5a = VFNMS(LDK(KP980785280), T4F, VMUL(LDK(KP195090322), T4w));
1728 T5e = VFMA(LDK(KP923879532), T3W, VMUL(LDK(KP382683432), T41));
1729 T5f = VFNMS(LDK(KP923879532), T47, VMUL(LDK(KP382683432), T4c));
1730 T5g = VADD(T5e, T5f);
1731 T8w = VSUB(T5f, T5e);
1732 }
1733 {
1734 V T5i, T5j, T4X, T56;
1735 T5i = VADD(T4k, T4v);
1736 T5j = VADD(T4B, T4E);
1737 T5k = VFMA(LDK(KP555570233), T5i, VMUL(LDK(KP831469612), T5j));
1738 T5q = VFNMS(LDK(KP555570233), T5j, VMUL(LDK(KP831469612), T5i));
1739 T4X = VSUB(T4L, T4W);
1740 T56 = VSUB(T52, T55);
1741 T57 = VFNMS(LDK(KP980785280), T56, VMUL(LDK(KP195090322), T4X));
1742 T5b = VFMA(LDK(KP195090322), T56, VMUL(LDK(KP980785280), T4X));
1743 }
1744 {
1745 V T4f, T58, T8v, T8y;
1746 T4f = VADD(T3R, T4e);
1747 T58 = VADD(T4G, T57);
1748 ST(&(ri[WS(rs, 23)]), VSUB(T4f, T58), ms, &(ri[WS(rs, 1)]));
1749 ST(&(ri[WS(rs, 7)]), VADD(T4f, T58), ms, &(ri[WS(rs, 1)]));
1750 T8v = VADD(T5a, T5b);
1751 T8y = VADD(T8w, T8x);
1752 ST(&(ii[WS(rs, 7)]), VADD(T8v, T8y), ms, &(ii[WS(rs, 1)]));
1753 ST(&(ii[WS(rs, 23)]), VSUB(T8y, T8v), ms, &(ii[WS(rs, 1)]));
1754 }
1755 {
1756 V T59, T5c, T8z, T8A;
1757 T59 = VSUB(T3R, T4e);
1758 T5c = VSUB(T5a, T5b);
1759 ST(&(ri[WS(rs, 31)]), VSUB(T59, T5c), ms, &(ri[WS(rs, 1)]));
1760 ST(&(ri[WS(rs, 15)]), VADD(T59, T5c), ms, &(ri[WS(rs, 1)]));
1761 T8z = VSUB(T57, T4G);
1762 T8A = VSUB(T8x, T8w);
1763 ST(&(ii[WS(rs, 15)]), VADD(T8z, T8A), ms, &(ii[WS(rs, 1)]));
1764 ST(&(ii[WS(rs, 31)]), VSUB(T8A, T8z), ms, &(ii[WS(rs, 1)]));
1765 }
1766 {
1767 V T5h, T5o, T8n, T8s;
1768 T5h = VADD(T5d, T5g);
1769 T5o = VADD(T5k, T5n);
1770 ST(&(ri[WS(rs, 19)]), VSUB(T5h, T5o), ms, &(ri[WS(rs, 1)]));
1771 ST(&(ri[WS(rs, 3)]), VADD(T5h, T5o), ms, &(ri[WS(rs, 1)]));
1772 T8n = VADD(T5q, T5r);
1773 T8s = VADD(T8o, T8r);
1774 ST(&(ii[WS(rs, 3)]), VADD(T8n, T8s), ms, &(ii[WS(rs, 1)]));
1775 ST(&(ii[WS(rs, 19)]), VSUB(T8s, T8n), ms, &(ii[WS(rs, 1)]));
1776 }
1777 {
1778 V T5p, T5s, T8t, T8u;
1779 T5p = VSUB(T5d, T5g);
1780 T5s = VSUB(T5q, T5r);
1781 ST(&(ri[WS(rs, 27)]), VSUB(T5p, T5s), ms, &(ri[WS(rs, 1)]));
1782 ST(&(ri[WS(rs, 11)]), VADD(T5p, T5s), ms, &(ri[WS(rs, 1)]));
1783 T8t = VSUB(T5n, T5k);
1784 T8u = VSUB(T8r, T8o);
1785 ST(&(ii[WS(rs, 11)]), VADD(T8t, T8u), ms, &(ii[WS(rs, 1)]));
1786 ST(&(ii[WS(rs, 27)]), VSUB(T8u, T8t), ms, &(ii[WS(rs, 1)]));
1787 }
1788 }
1789 {
1790 V T5x, T5Z, T8d, T8j, T5E, T88, T69, T6d, T5M, T5W, T62, T8i, T66, T6c, T5T;
1791 V T5X, T5w, T89;
1792 T5w = VMUL(LDK(KP707106781), VADD(T5u, T5v));
1793 T5x = VSUB(T5t, T5w);
1794 T5Z = VADD(T5t, T5w);
1795 T89 = VMUL(LDK(KP707106781), VADD(T3K, T3P));
1796 T8d = VADD(T89, T8c);
1797 T8j = VSUB(T8c, T89);
1798 {
1799 V T5A, T5D, T67, T68;
1800 T5A = VFNMS(LDK(KP382683432), T5z, VMUL(LDK(KP923879532), T5y));
1801 T5D = VFMA(LDK(KP923879532), T5B, VMUL(LDK(KP382683432), T5C));
1802 T5E = VSUB(T5A, T5D);
1803 T88 = VADD(T5A, T5D);
1804 T67 = VADD(T5N, T5O);
1805 T68 = VADD(T5Q, T5R);
1806 T69 = VFNMS(LDK(KP195090322), T68, VMUL(LDK(KP980785280), T67));
1807 T6d = VFMA(LDK(KP195090322), T67, VMUL(LDK(KP980785280), T68));
1808 }
1809 {
1810 V T5I, T5L, T60, T61;
1811 T5I = VSUB(T5G, T5H);
1812 T5L = VSUB(T5J, T5K);
1813 T5M = VFMA(LDK(KP555570233), T5I, VMUL(LDK(KP831469612), T5L));
1814 T5W = VFNMS(LDK(KP831469612), T5I, VMUL(LDK(KP555570233), T5L));
1815 T60 = VFMA(LDK(KP382683432), T5y, VMUL(LDK(KP923879532), T5z));
1816 T61 = VFNMS(LDK(KP382683432), T5B, VMUL(LDK(KP923879532), T5C));
1817 T62 = VADD(T60, T61);
1818 T8i = VSUB(T61, T60);
1819 }
1820 {
1821 V T64, T65, T5P, T5S;
1822 T64 = VADD(T5G, T5H);
1823 T65 = VADD(T5J, T5K);
1824 T66 = VFMA(LDK(KP980785280), T64, VMUL(LDK(KP195090322), T65));
1825 T6c = VFNMS(LDK(KP195090322), T64, VMUL(LDK(KP980785280), T65));
1826 T5P = VSUB(T5N, T5O);
1827 T5S = VSUB(T5Q, T5R);
1828 T5T = VFNMS(LDK(KP831469612), T5S, VMUL(LDK(KP555570233), T5P));
1829 T5X = VFMA(LDK(KP831469612), T5P, VMUL(LDK(KP555570233), T5S));
1830 }
1831 {
1832 V T5F, T5U, T8h, T8k;
1833 T5F = VADD(T5x, T5E);
1834 T5U = VADD(T5M, T5T);
1835 ST(&(ri[WS(rs, 21)]), VSUB(T5F, T5U), ms, &(ri[WS(rs, 1)]));
1836 ST(&(ri[WS(rs, 5)]), VADD(T5F, T5U), ms, &(ri[WS(rs, 1)]));
1837 T8h = VADD(T5W, T5X);
1838 T8k = VADD(T8i, T8j);
1839 ST(&(ii[WS(rs, 5)]), VADD(T8h, T8k), ms, &(ii[WS(rs, 1)]));
1840 ST(&(ii[WS(rs, 21)]), VSUB(T8k, T8h), ms, &(ii[WS(rs, 1)]));
1841 }
1842 {
1843 V T5V, T5Y, T8l, T8m;
1844 T5V = VSUB(T5x, T5E);
1845 T5Y = VSUB(T5W, T5X);
1846 ST(&(ri[WS(rs, 29)]), VSUB(T5V, T5Y), ms, &(ri[WS(rs, 1)]));
1847 ST(&(ri[WS(rs, 13)]), VADD(T5V, T5Y), ms, &(ri[WS(rs, 1)]));
1848 T8l = VSUB(T5T, T5M);
1849 T8m = VSUB(T8j, T8i);
1850 ST(&(ii[WS(rs, 13)]), VADD(T8l, T8m), ms, &(ii[WS(rs, 1)]));
1851 ST(&(ii[WS(rs, 29)]), VSUB(T8m, T8l), ms, &(ii[WS(rs, 1)]));
1852 }
1853 {
1854 V T63, T6a, T87, T8e;
1855 T63 = VADD(T5Z, T62);
1856 T6a = VADD(T66, T69);
1857 ST(&(ri[WS(rs, 17)]), VSUB(T63, T6a), ms, &(ri[WS(rs, 1)]));
1858 ST(&(ri[WS(rs, 1)]), VADD(T63, T6a), ms, &(ri[WS(rs, 1)]));
1859 T87 = VADD(T6c, T6d);
1860 T8e = VADD(T88, T8d);
1861 ST(&(ii[WS(rs, 1)]), VADD(T87, T8e), ms, &(ii[WS(rs, 1)]));
1862 ST(&(ii[WS(rs, 17)]), VSUB(T8e, T87), ms, &(ii[WS(rs, 1)]));
1863 }
1864 {
1865 V T6b, T6e, T8f, T8g;
1866 T6b = VSUB(T5Z, T62);
1867 T6e = VSUB(T6c, T6d);
1868 ST(&(ri[WS(rs, 25)]), VSUB(T6b, T6e), ms, &(ri[WS(rs, 1)]));
1869 ST(&(ri[WS(rs, 9)]), VADD(T6b, T6e), ms, &(ri[WS(rs, 1)]));
1870 T8f = VSUB(T69, T66);
1871 T8g = VSUB(T8d, T88);
1872 ST(&(ii[WS(rs, 9)]), VADD(T8f, T8g), ms, &(ii[WS(rs, 1)]));
1873 ST(&(ii[WS(rs, 25)]), VSUB(T8g, T8f), ms, &(ii[WS(rs, 1)]));
1874 }
1875 }
1876 }
1877 }
1878 }
1879 VLEAVE();
1880 }
1881
1882 static const tw_instr twinstr[] = {
1883 VTW(0, 1),
1884 VTW(0, 3),
1885 VTW(0, 9),
1886 VTW(0, 27),
1887 {TW_NEXT, (2 * VL), 0}
1888 };
1889
1890 static const ct_desc desc = { 32, XSIMD_STRING("t2sv_32"), twinstr, &GENUS, {376, 168, 112, 0}, 0, 0, 0 };
1891
1892 void XSIMD(codelet_t2sv_32) (planner *p) {
1893 X(kdft_dit_register) (p, t2sv_32, &desc);
1894 }
1895 #endif