comparison src/fftw-3.3.3/dft/simd/common/t2sv_32.c @ 10:37bf6b4a2645

Add FFTW3
author Chris Cannam
date Wed, 20 Mar 2013 15:35:50 +0000
parents
children
comparison
equal deleted inserted replaced
9:c0fb53affa76 10:37bf6b4a2645
1 /*
2 * Copyright (c) 2003, 2007-11 Matteo Frigo
3 * Copyright (c) 2003, 2007-11 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 Sun Nov 25 07:39:28 EST 2012 */
23
24 #include "codelet-dft.h"
25
26 #ifdef HAVE_FMA
27
28 /* Generated by: ../../../genfft/gen_twiddle.native -fma -reorder-insns -schedule-for-pipeline -simd -compact -variables 4 -pipeline-latency 8 -twiddle-log3 -precompute-twiddles -n 32 -name t2sv_32 -include 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 * 204 stack variables, 7 constants, and 128 memory accesses
34 */
35 #include "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(KP831469612, +0.831469612302545237078788377617905756738560812);
40 DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
41 DVK(KP668178637, +0.668178637919298919997757686523080761552472251);
42 DVK(KP198912367, +0.198912367379658006911597622644676228597850501);
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 T6H, T74, T6U, T6E, T9r, T9t, T78, T7c, T6W, T6S, T73, T6K, T7a, T72, T9x;
50 V T9z;
51 {
52 V T2, T8, T3, T6, Te, Ti, T5, Tc;
53 T2 = LDW(&(W[0]));
54 T8 = LDW(&(W[TWVL * 4]));
55 T3 = LDW(&(W[TWVL * 2]));
56 T6 = LDW(&(W[TWVL * 3]));
57 Te = LDW(&(W[TWVL * 6]));
58 Ti = LDW(&(W[TWVL * 7]));
59 T5 = LDW(&(W[TWVL * 1]));
60 Tc = LDW(&(W[TWVL * 5]));
61 {
62 V T2X, T2T, T34, T31, Tq, T46, T97, T8H, TH, T98, T4b, T8D, TZ, T7f, T1g;
63 V T7g, T4j, T6t, T4q, T6u, T6x, T4z, T7m, T1J, T4G, T6y, T8d, T7l, T4O, T6A;
64 V T2k, T7o, T6B, T4V, T7r, T8e, T5E, T6P, T3G, T7L, T6M, T61, T8n, T7I, T55;
65 V T6I, T2N, T7A, T5s, T6F, T7x, T8i, T2R, T2U, T57, T3a, T5h, T62, T5L, T7J;
66 V T43, T63, T5S, T8o, T7O, T2V, T2Y, T32, T35;
67 {
68 V T1w, T23, T1K, T1F, T1s, T1N, T26, T1z, T2w, T2s, T3Q, T3M, T3r, T3n, T2b;
69 V T1U, T3C, T3j, T3z, T3f, T1R, T29, TR, Th, T2J, T2F, Td, TP, T1Z, T1V;
70 V T2g, T2c, T1m, T4u, T1D, T1G, T1p, T1t, T1E, T4D, T1x, T1A, T1q, T4v;
71 {
72 V T1, Ts, T19, TJ, T7, TM, Tb, T11, T1C, T1o, TA, T15, TE, T1d, Tw;
73 V T8G, Tk, Tn, Tj, TW, TS, To, Tt, Tx, TB, TF, Tl;
74 {
75 V T1Y, T1S, T2f, T2a;
76 T1 = LD(&(ri[0]), ms, &(ri[0]));
77 {
78 V Tr, T18, T4, Ta;
79 Tr = VMUL(T2, T8);
80 T18 = VMUL(T3, T8);
81 T4 = VMUL(T2, T3);
82 Ta = VMUL(T2, T6);
83 {
84 V T10, T1n, Tz, T14;
85 T10 = VMUL(T2, Te);
86 T1n = VMUL(T8, Te);
87 Tz = VMUL(T3, Te);
88 T14 = VMUL(T2, Ti);
89 {
90 V T1r, TD, T1c, Tv;
91 T1r = VMUL(T8, Ti);
92 TD = VMUL(T3, Ti);
93 T1c = VMUL(T3, Tc);
94 Tv = VMUL(T2, Tc);
95 T1w = VFNMS(T5, Tc, Tr);
96 Ts = VFMA(T5, Tc, Tr);
97 T19 = VFNMS(T6, Tc, T18);
98 T23 = VFMA(T6, Tc, T18);
99 TJ = VFNMS(T5, T6, T4);
100 T7 = VFMA(T5, T6, T4);
101 TM = VFMA(T5, T3, Ta);
102 Tb = VFNMS(T5, T3, Ta);
103 T11 = VFNMS(T5, Ti, T10);
104 T1C = VFMA(T5, Ti, T10);
105 T1o = VFMA(Tc, Ti, T1n);
106 TA = VFMA(T6, Ti, Tz);
107 T1K = VFNMS(T6, Ti, Tz);
108 T1F = VFNMS(T5, Te, T14);
109 T15 = VFMA(T5, Te, T14);
110 T1s = VFNMS(Tc, Te, T1r);
111 T1N = VFMA(T6, Te, TD);
112 TE = VFNMS(T6, Te, TD);
113 T26 = VFNMS(T6, T8, T1c);
114 T1d = VFMA(T6, T8, T1c);
115 T1z = VFMA(T5, T8, Tv);
116 Tw = VFNMS(T5, T8, Tv);
117 {
118 V T2v, T2r, T3P, T3L;
119 T2v = VMUL(T1w, Ti);
120 T2r = VMUL(T1w, Te);
121 T3P = VMUL(Ts, Ti);
122 T3L = VMUL(Ts, Te);
123 {
124 V T3q, T3m, T2W, T2S;
125 T3q = VMUL(T19, Ti);
126 T3m = VMUL(T19, Te);
127 T2W = VMUL(T23, Ti);
128 T2S = VMUL(T23, Te);
129 {
130 V T1T, T3i, T3e, T1Q;
131 T1T = VMUL(TJ, Tc);
132 T3i = VMUL(TJ, Ti);
133 T3e = VMUL(TJ, Te);
134 T1Q = VMUL(TJ, T8);
135 {
136 V Tg, T2I, T2E, T9;
137 Tg = VMUL(T7, Tc);
138 T2I = VMUL(T7, Ti);
139 T2E = VMUL(T7, Te);
140 T9 = VMUL(T7, T8);
141 T2w = VFNMS(T1z, Te, T2v);
142 T2s = VFMA(T1z, Ti, T2r);
143 T3Q = VFNMS(Tw, Te, T3P);
144 T3M = VFMA(Tw, Ti, T3L);
145 T3r = VFNMS(T1d, Te, T3q);
146 T3n = VFMA(T1d, Ti, T3m);
147 T2X = VFNMS(T26, Te, T2W);
148 T2T = VFMA(T26, Ti, T2S);
149 T2b = VFNMS(TM, T8, T1T);
150 T1U = VFMA(TM, T8, T1T);
151 T3C = VFNMS(TM, Te, T3i);
152 T3j = VFMA(TM, Te, T3i);
153 T3z = VFMA(TM, Ti, T3e);
154 T3f = VFNMS(TM, Ti, T3e);
155 T1R = VFNMS(TM, Tc, T1Q);
156 T29 = VFMA(TM, Tc, T1Q);
157 TR = VFNMS(Tb, T8, Tg);
158 Th = VFMA(Tb, T8, Tg);
159 T34 = VFMA(Tb, Te, T2I);
160 T2J = VFNMS(Tb, Te, T2I);
161 T31 = VFNMS(Tb, Ti, T2E);
162 T2F = VFMA(Tb, Ti, T2E);
163 Td = VFNMS(Tb, Tc, T9);
164 TP = VFMA(Tb, Tc, T9);
165 T1Y = VMUL(T1R, Ti);
166 T1S = VMUL(T1R, Te);
167 T2f = VMUL(T29, Ti);
168 T2a = VMUL(T29, Te);
169 T8G = LD(&(ii[0]), ms, &(ii[0]));
170 }
171 }
172 }
173 }
174 }
175 }
176 }
177 Tk = LD(&(ri[WS(rs, 16)]), ms, &(ri[0]));
178 {
179 V Tm, Tf, TV, TQ;
180 Tm = VMUL(Td, Ti);
181 Tf = VMUL(Td, Te);
182 TV = VMUL(TP, Ti);
183 TQ = VMUL(TP, Te);
184 T1Z = VFNMS(T1U, Te, T1Y);
185 T1V = VFMA(T1U, Ti, T1S);
186 T2g = VFNMS(T2b, Te, T2f);
187 T2c = VFMA(T2b, Ti, T2a);
188 Tn = VFNMS(Th, Te, Tm);
189 Tj = VFMA(Th, Ti, Tf);
190 TW = VFNMS(TR, Te, TV);
191 TS = VFMA(TR, Ti, TQ);
192 }
193 To = LD(&(ii[WS(rs, 16)]), ms, &(ii[0]));
194 }
195 Tt = LD(&(ri[WS(rs, 8)]), ms, &(ri[0]));
196 Tx = LD(&(ii[WS(rs, 8)]), ms, &(ii[0]));
197 TB = LD(&(ri[WS(rs, 24)]), ms, &(ri[0]));
198 TF = LD(&(ii[WS(rs, 24)]), ms, &(ii[0]));
199 Tl = VMUL(Tj, Tk);
200 {
201 V TO, T4f, TT, TX;
202 {
203 V Ty, T48, TG, T4a;
204 {
205 V TK, TN, T8E, Tu, T47, TC, T49, Tp, TL, T4e, T8F;
206 TK = LD(&(ri[WS(rs, 4)]), ms, &(ri[0]));
207 TN = LD(&(ii[WS(rs, 4)]), ms, &(ii[0]));
208 T8E = VMUL(Tj, To);
209 Tu = VMUL(Ts, Tt);
210 T47 = VMUL(Ts, Tx);
211 TC = VMUL(TA, TB);
212 T49 = VMUL(TA, TF);
213 Tp = VFMA(Tn, To, Tl);
214 TL = VMUL(TJ, TK);
215 T4e = VMUL(TJ, TN);
216 T8F = VFNMS(Tn, Tk, T8E);
217 Ty = VFMA(Tw, Tx, Tu);
218 T48 = VFNMS(Tw, Tt, T47);
219 TG = VFMA(TE, TF, TC);
220 T4a = VFNMS(TE, TB, T49);
221 Tq = VADD(T1, Tp);
222 T46 = VSUB(T1, Tp);
223 TO = VFMA(TM, TN, TL);
224 T97 = VSUB(T8G, T8F);
225 T8H = VADD(T8F, T8G);
226 T4f = VFNMS(TM, TK, T4e);
227 }
228 TH = VADD(Ty, TG);
229 T98 = VSUB(Ty, TG);
230 T4b = VSUB(T48, T4a);
231 T8D = VADD(T48, T4a);
232 TT = LD(&(ri[WS(rs, 20)]), ms, &(ri[0]));
233 TX = LD(&(ii[WS(rs, 20)]), ms, &(ii[0]));
234 }
235 {
236 V T12, T16, T1a, T1e, T4k, T4p;
237 T12 = LD(&(ri[WS(rs, 28)]), ms, &(ri[0]));
238 T16 = LD(&(ii[WS(rs, 28)]), ms, &(ii[0]));
239 T1a = LD(&(ri[WS(rs, 12)]), ms, &(ri[0]));
240 T1e = LD(&(ii[WS(rs, 12)]), ms, &(ii[0]));
241 {
242 V TY, T4h, T17, T4m, T1f, T4o, T4d, T4i;
243 {
244 V T1j, T1l, TU, T4g, T13, T4l, T1b, T4n, T1k, T4t;
245 T1j = LD(&(ri[WS(rs, 2)]), ms, &(ri[0]));
246 T1l = LD(&(ii[WS(rs, 2)]), ms, &(ii[0]));
247 TU = VMUL(TS, TT);
248 T4g = VMUL(TS, TX);
249 T13 = VMUL(T11, T12);
250 T4l = VMUL(T11, T16);
251 T1b = VMUL(T19, T1a);
252 T4n = VMUL(T19, T1e);
253 T1k = VMUL(T7, T1j);
254 T4t = VMUL(T7, T1l);
255 TY = VFMA(TW, TX, TU);
256 T4h = VFNMS(TW, TT, T4g);
257 T17 = VFMA(T15, T16, T13);
258 T4m = VFNMS(T15, T12, T4l);
259 T1f = VFMA(T1d, T1e, T1b);
260 T4o = VFNMS(T1d, T1a, T4n);
261 T1m = VFMA(Tb, T1l, T1k);
262 T4u = VFNMS(Tb, T1j, T4t);
263 }
264 TZ = VADD(TO, TY);
265 T4d = VSUB(TO, TY);
266 T7f = VADD(T4f, T4h);
267 T4i = VSUB(T4f, T4h);
268 T1g = VADD(T17, T1f);
269 T4k = VSUB(T17, T1f);
270 T7g = VADD(T4m, T4o);
271 T4p = VSUB(T4m, T4o);
272 T1D = LD(&(ri[WS(rs, 26)]), ms, &(ri[0]));
273 T1G = LD(&(ii[WS(rs, 26)]), ms, &(ii[0]));
274 T4j = VADD(T4d, T4i);
275 T6t = VSUB(T4i, T4d);
276 }
277 T1p = LD(&(ri[WS(rs, 18)]), ms, &(ri[0]));
278 T1t = LD(&(ii[WS(rs, 18)]), ms, &(ii[0]));
279 T4q = VSUB(T4k, T4p);
280 T6u = VADD(T4k, T4p);
281 T1E = VMUL(T1C, T1D);
282 T4D = VMUL(T1C, T1G);
283 T1x = LD(&(ri[WS(rs, 10)]), ms, &(ri[0]));
284 T1A = LD(&(ii[WS(rs, 10)]), ms, &(ii[0]));
285 T1q = VMUL(T1o, T1p);
286 T4v = VMUL(T1o, T1t);
287 }
288 }
289 }
290 {
291 V T3l, T5z, T3E, T5Z, T3v, T3x, T3w, T3t, T5B, T5W;
292 {
293 V T1P, T4J, T1W, T20, T2i, T4T, T1X, T4K, T24, T27;
294 {
295 V T2d, T2h, T1v, T4A, T7j, T4x, T2e, T4y, T1I, T4F, T7k, T4S;
296 {
297 V T1L, T1O, T1H, T4E, T1y, T4B, T1u, T4w, T1M, T4I, T1B, T4C;
298 T1L = LD(&(ri[WS(rs, 30)]), ms, &(ri[0]));
299 T1O = LD(&(ii[WS(rs, 30)]), ms, &(ii[0]));
300 T1H = VFMA(T1F, T1G, T1E);
301 T4E = VFNMS(T1F, T1D, T4D);
302 T1y = VMUL(T1w, T1x);
303 T4B = VMUL(T1w, T1A);
304 T1u = VFMA(T1s, T1t, T1q);
305 T4w = VFNMS(T1s, T1p, T4v);
306 T1M = VMUL(T1K, T1L);
307 T4I = VMUL(T1K, T1O);
308 T2d = LD(&(ri[WS(rs, 22)]), ms, &(ri[0]));
309 T2h = LD(&(ii[WS(rs, 22)]), ms, &(ii[0]));
310 T1B = VFMA(T1z, T1A, T1y);
311 T4C = VFNMS(T1z, T1x, T4B);
312 T1v = VADD(T1m, T1u);
313 T4A = VSUB(T1m, T1u);
314 T7j = VADD(T4u, T4w);
315 T4x = VSUB(T4u, T4w);
316 T1P = VFMA(T1N, T1O, T1M);
317 T4J = VFNMS(T1N, T1L, T4I);
318 T2e = VMUL(T2c, T2d);
319 T4y = VSUB(T1B, T1H);
320 T1I = VADD(T1B, T1H);
321 T4F = VSUB(T4C, T4E);
322 T7k = VADD(T4C, T4E);
323 T4S = VMUL(T2c, T2h);
324 }
325 T1W = LD(&(ri[WS(rs, 14)]), ms, &(ri[0]));
326 T20 = LD(&(ii[WS(rs, 14)]), ms, &(ii[0]));
327 T2i = VFMA(T2g, T2h, T2e);
328 T6x = VADD(T4x, T4y);
329 T4z = VSUB(T4x, T4y);
330 T7m = VSUB(T1v, T1I);
331 T1J = VADD(T1v, T1I);
332 T4G = VADD(T4A, T4F);
333 T6y = VSUB(T4A, T4F);
334 T8d = VADD(T7j, T7k);
335 T7l = VSUB(T7j, T7k);
336 T4T = VFNMS(T2g, T2d, T4S);
337 T1X = VMUL(T1V, T1W);
338 T4K = VMUL(T1V, T20);
339 T24 = LD(&(ri[WS(rs, 6)]), ms, &(ri[0]));
340 T27 = LD(&(ii[WS(rs, 6)]), ms, &(ii[0]));
341 }
342 {
343 V T22, T4P, T7p, T4M, T28, T4R, T3g, T3k;
344 T3g = LD(&(ri[WS(rs, 31)]), ms, &(ri[WS(rs, 1)]));
345 T3k = LD(&(ii[WS(rs, 31)]), ms, &(ii[WS(rs, 1)]));
346 {
347 V T3A, T3D, T21, T4L, T25, T4Q, T3h, T5y, T3B, T5Y;
348 T3A = LD(&(ri[WS(rs, 23)]), ms, &(ri[WS(rs, 1)]));
349 T3D = LD(&(ii[WS(rs, 23)]), ms, &(ii[WS(rs, 1)]));
350 T21 = VFMA(T1Z, T20, T1X);
351 T4L = VFNMS(T1Z, T1W, T4K);
352 T25 = VMUL(T23, T24);
353 T4Q = VMUL(T23, T27);
354 T3h = VMUL(T3f, T3g);
355 T5y = VMUL(T3f, T3k);
356 T3B = VMUL(T3z, T3A);
357 T5Y = VMUL(T3z, T3D);
358 T22 = VADD(T1P, T21);
359 T4P = VSUB(T1P, T21);
360 T7p = VADD(T4J, T4L);
361 T4M = VSUB(T4J, T4L);
362 T28 = VFMA(T26, T27, T25);
363 T4R = VFNMS(T26, T24, T4Q);
364 T3l = VFMA(T3j, T3k, T3h);
365 T5z = VFNMS(T3j, T3g, T5y);
366 T3E = VFMA(T3C, T3D, T3B);
367 T5Z = VFNMS(T3C, T3A, T5Y);
368 }
369 {
370 V T3o, T3s, T2j, T4N, T7q, T4U, T3p, T5A;
371 T3o = LD(&(ri[WS(rs, 15)]), ms, &(ri[WS(rs, 1)]));
372 T3s = LD(&(ii[WS(rs, 15)]), ms, &(ii[WS(rs, 1)]));
373 T2j = VADD(T28, T2i);
374 T4N = VSUB(T28, T2i);
375 T7q = VADD(T4R, T4T);
376 T4U = VSUB(T4R, T4T);
377 T3v = LD(&(ri[WS(rs, 7)]), ms, &(ri[WS(rs, 1)]));
378 T3x = LD(&(ii[WS(rs, 7)]), ms, &(ii[WS(rs, 1)]));
379 T3p = VMUL(T3n, T3o);
380 T5A = VMUL(T3n, T3s);
381 T4O = VSUB(T4M, T4N);
382 T6A = VADD(T4M, T4N);
383 T2k = VADD(T22, T2j);
384 T7o = VSUB(T22, T2j);
385 T6B = VSUB(T4P, T4U);
386 T4V = VADD(T4P, T4U);
387 T7r = VSUB(T7p, T7q);
388 T8e = VADD(T7p, T7q);
389 T3w = VMUL(TP, T3v);
390 T3t = VFMA(T3r, T3s, T3p);
391 T5B = VFNMS(T3r, T3o, T5A);
392 T5W = VMUL(TP, T3x);
393 }
394 }
395 }
396 {
397 V T2t, T2q, T50, T2L, T5q, T2u, T2x, T2A, T2C;
398 {
399 V T2n, T2p, T2G, T2K, T5V, T3u, T5C, T7G, T5X, T2o, T4Z, T2H, T5D, T3F, T5p;
400 V T3y, T60, T7H;
401 T2n = LD(&(ri[WS(rs, 1)]), ms, &(ri[WS(rs, 1)]));
402 T2p = LD(&(ii[WS(rs, 1)]), ms, &(ii[WS(rs, 1)]));
403 T2G = LD(&(ri[WS(rs, 25)]), ms, &(ri[WS(rs, 1)]));
404 T2K = LD(&(ii[WS(rs, 25)]), ms, &(ii[WS(rs, 1)]));
405 T3y = VFMA(TR, T3x, T3w);
406 T5V = VSUB(T3l, T3t);
407 T3u = VADD(T3l, T3t);
408 T5C = VSUB(T5z, T5B);
409 T7G = VADD(T5z, T5B);
410 T5X = VFNMS(TR, T3v, T5W);
411 T2o = VMUL(T2, T2n);
412 T4Z = VMUL(T2, T2p);
413 T2H = VMUL(T2F, T2G);
414 T5D = VSUB(T3y, T3E);
415 T3F = VADD(T3y, T3E);
416 T5p = VMUL(T2F, T2K);
417 T2t = LD(&(ri[WS(rs, 17)]), ms, &(ri[WS(rs, 1)]));
418 T60 = VSUB(T5X, T5Z);
419 T7H = VADD(T5X, T5Z);
420 T2q = VFMA(T5, T2p, T2o);
421 T50 = VFNMS(T5, T2n, T4Z);
422 T2L = VFMA(T2J, T2K, T2H);
423 T5E = VSUB(T5C, T5D);
424 T6P = VADD(T5C, T5D);
425 T3G = VADD(T3u, T3F);
426 T7L = VSUB(T3u, T3F);
427 T5q = VFNMS(T2J, T2G, T5p);
428 T6M = VSUB(T5V, T60);
429 T61 = VADD(T5V, T60);
430 T8n = VADD(T7G, T7H);
431 T7I = VSUB(T7G, T7H);
432 T2u = VMUL(T2s, T2t);
433 T2x = LD(&(ii[WS(rs, 17)]), ms, &(ii[WS(rs, 1)]));
434 T2A = LD(&(ri[WS(rs, 9)]), ms, &(ri[WS(rs, 1)]));
435 T2C = LD(&(ii[WS(rs, 9)]), ms, &(ii[WS(rs, 1)]));
436 }
437 {
438 V T3N, T2z, T5m, T3K, T5G, T41, T5Q, T3O, T7v, T53, T2M, T54, T7w, T5r, T3R;
439 V T3U, T3W;
440 {
441 V T3H, T3J, T3Y, T40, T52, T2D, T5o;
442 T3H = LD(&(ri[WS(rs, 3)]), ms, &(ri[WS(rs, 1)]));
443 T3J = LD(&(ii[WS(rs, 3)]), ms, &(ii[WS(rs, 1)]));
444 T3Y = LD(&(ri[WS(rs, 11)]), ms, &(ri[WS(rs, 1)]));
445 T40 = LD(&(ii[WS(rs, 11)]), ms, &(ii[WS(rs, 1)]));
446 T3N = LD(&(ri[WS(rs, 19)]), ms, &(ri[WS(rs, 1)]));
447 {
448 V T2y, T51, T2B, T5n;
449 T2y = VFMA(T2w, T2x, T2u);
450 T51 = VMUL(T2s, T2x);
451 T2B = VMUL(T8, T2A);
452 T5n = VMUL(T8, T2C);
453 {
454 V T3I, T5F, T3Z, T5P;
455 T3I = VMUL(T3, T3H);
456 T5F = VMUL(T3, T3J);
457 T3Z = VMUL(Td, T3Y);
458 T5P = VMUL(Td, T40);
459 T2z = VADD(T2q, T2y);
460 T5m = VSUB(T2q, T2y);
461 T52 = VFNMS(T2w, T2t, T51);
462 T2D = VFMA(Tc, T2C, T2B);
463 T5o = VFNMS(Tc, T2A, T5n);
464 T3K = VFMA(T6, T3J, T3I);
465 T5G = VFNMS(T6, T3H, T5F);
466 T41 = VFMA(Th, T40, T3Z);
467 T5Q = VFNMS(Th, T3Y, T5P);
468 T3O = VMUL(T3M, T3N);
469 }
470 }
471 T7v = VADD(T50, T52);
472 T53 = VSUB(T50, T52);
473 T2M = VADD(T2D, T2L);
474 T54 = VSUB(T2D, T2L);
475 T7w = VADD(T5o, T5q);
476 T5r = VSUB(T5o, T5q);
477 T3R = LD(&(ii[WS(rs, 19)]), ms, &(ii[WS(rs, 1)]));
478 T3U = LD(&(ri[WS(rs, 27)]), ms, &(ri[WS(rs, 1)]));
479 T3W = LD(&(ii[WS(rs, 27)]), ms, &(ii[WS(rs, 1)]));
480 }
481 {
482 V T2O, T37, T39, T3T, T5K, T5I, T3X, T5O, T56, T38, T5g, T7M, T5J;
483 {
484 V T3S, T5H, T3V, T5N, T2P, T2Q;
485 T2O = LD(&(ri[WS(rs, 5)]), ms, &(ri[WS(rs, 1)]));
486 T55 = VSUB(T53, T54);
487 T6I = VADD(T53, T54);
488 T2N = VADD(T2z, T2M);
489 T7A = VSUB(T2z, T2M);
490 T5s = VADD(T5m, T5r);
491 T6F = VSUB(T5m, T5r);
492 T7x = VSUB(T7v, T7w);
493 T8i = VADD(T7v, T7w);
494 T3S = VFMA(T3Q, T3R, T3O);
495 T5H = VMUL(T3M, T3R);
496 T3V = VMUL(Te, T3U);
497 T5N = VMUL(Te, T3W);
498 T2P = VMUL(T29, T2O);
499 T2Q = LD(&(ii[WS(rs, 5)]), ms, &(ii[WS(rs, 1)]));
500 T37 = LD(&(ri[WS(rs, 13)]), ms, &(ri[WS(rs, 1)]));
501 T39 = LD(&(ii[WS(rs, 13)]), ms, &(ii[WS(rs, 1)]));
502 T3T = VADD(T3K, T3S);
503 T5K = VSUB(T3K, T3S);
504 T5I = VFNMS(T3Q, T3N, T5H);
505 T3X = VFMA(Ti, T3W, T3V);
506 T5O = VFNMS(Ti, T3U, T5N);
507 T2R = VFMA(T2b, T2Q, T2P);
508 T56 = VMUL(T29, T2Q);
509 T38 = VMUL(T1R, T37);
510 T5g = VMUL(T1R, T39);
511 }
512 T2U = LD(&(ri[WS(rs, 21)]), ms, &(ri[WS(rs, 1)]));
513 T7M = VADD(T5G, T5I);
514 T5J = VSUB(T5G, T5I);
515 {
516 V T42, T5M, T7N, T5R;
517 T42 = VADD(T3X, T41);
518 T5M = VSUB(T3X, T41);
519 T7N = VADD(T5O, T5Q);
520 T5R = VSUB(T5O, T5Q);
521 T57 = VFNMS(T2b, T2O, T56);
522 T3a = VFMA(T1U, T39, T38);
523 T5h = VFNMS(T1U, T37, T5g);
524 T62 = VADD(T5K, T5J);
525 T5L = VSUB(T5J, T5K);
526 T7J = VSUB(T42, T3T);
527 T43 = VADD(T3T, T42);
528 T63 = VSUB(T5M, T5R);
529 T5S = VADD(T5M, T5R);
530 T8o = VADD(T7M, T7N);
531 T7O = VSUB(T7M, T7N);
532 T2V = VMUL(T2T, T2U);
533 }
534 T2Y = LD(&(ii[WS(rs, 21)]), ms, &(ii[WS(rs, 1)]));
535 T32 = LD(&(ri[WS(rs, 29)]), ms, &(ri[WS(rs, 1)]));
536 T35 = LD(&(ii[WS(rs, 29)]), ms, &(ii[WS(rs, 1)]));
537 }
538 }
539 }
540 }
541 }
542 {
543 V T5t, T5c, T5u, T5j, T8Z, T90;
544 {
545 V T7e, T8T, T8y, T7h, T8U, T8c, T8J, T44, T8u, T8q, T7y, T7D, T8w, T2m, T3d;
546 V T8h, T8R, T8P, T8k, T8x, T8B, T8f;
547 {
548 V T1i, T8O, T8N, T2l, T3c, T8j;
549 {
550 V T8p, T5b, T30, T59, T36, T5f, TI, T1h, T8m, T5a, T7B;
551 TI = VADD(Tq, TH);
552 T7e = VSUB(Tq, TH);
553 T8T = VSUB(T1g, TZ);
554 T1h = VADD(TZ, T1g);
555 T8y = VADD(T8n, T8o);
556 T8p = VSUB(T8n, T8o);
557 {
558 V T8C, T8I, T2Z, T58, T33, T5e;
559 T7h = VSUB(T7f, T7g);
560 T8C = VADD(T7f, T7g);
561 T8I = VADD(T8D, T8H);
562 T8U = VSUB(T8H, T8D);
563 T2Z = VFMA(T2X, T2Y, T2V);
564 T58 = VMUL(T2T, T2Y);
565 T33 = VMUL(T31, T32);
566 T5e = VMUL(T31, T35);
567 T1i = VADD(TI, T1h);
568 T8c = VSUB(TI, T1h);
569 T8O = VSUB(T8I, T8C);
570 T8J = VADD(T8C, T8I);
571 T5b = VSUB(T2R, T2Z);
572 T30 = VADD(T2R, T2Z);
573 T59 = VFNMS(T2X, T2U, T58);
574 T36 = VFMA(T34, T35, T33);
575 T5f = VFNMS(T34, T32, T5e);
576 }
577 T44 = VADD(T3G, T43);
578 T8m = VSUB(T3G, T43);
579 T5a = VSUB(T57, T59);
580 T7B = VADD(T57, T59);
581 {
582 V T5d, T3b, T5i, T7C;
583 T5d = VSUB(T36, T3a);
584 T3b = VADD(T36, T3a);
585 T5i = VSUB(T5f, T5h);
586 T7C = VADD(T5f, T5h);
587 T8N = VSUB(T2k, T1J);
588 T2l = VADD(T1J, T2k);
589 T8u = VADD(T8m, T8p);
590 T8q = VSUB(T8m, T8p);
591 T5t = VADD(T5b, T5a);
592 T5c = VSUB(T5a, T5b);
593 T7y = VSUB(T3b, T30);
594 T3c = VADD(T30, T3b);
595 T5u = VSUB(T5d, T5i);
596 T5j = VADD(T5d, T5i);
597 T8j = VADD(T7B, T7C);
598 T7D = VSUB(T7B, T7C);
599 }
600 }
601 T8w = VSUB(T1i, T2l);
602 T2m = VADD(T1i, T2l);
603 T3d = VADD(T2N, T3c);
604 T8h = VSUB(T2N, T3c);
605 T8R = VSUB(T8O, T8N);
606 T8P = VADD(T8N, T8O);
607 T8k = VSUB(T8i, T8j);
608 T8x = VADD(T8i, T8j);
609 T8B = VADD(T8d, T8e);
610 T8f = VSUB(T8d, T8e);
611 }
612 {
613 V T7P, T7K, T7X, T7Y, T82, T7z, T7W, T7i, T8a, T86, T91, T8V, T8W, T7t, T7E;
614 V T81;
615 {
616 V T84, T85, T7n, T7s, T8L, T45;
617 T8L = VSUB(T44, T3d);
618 T45 = VADD(T3d, T44);
619 {
620 V T8t, T8l, T8A, T8z;
621 T8t = VSUB(T8k, T8h);
622 T8l = VADD(T8h, T8k);
623 T8A = VADD(T8x, T8y);
624 T8z = VSUB(T8x, T8y);
625 {
626 V T8M, T8K, T8s, T8g;
627 T8M = VSUB(T8J, T8B);
628 T8K = VADD(T8B, T8J);
629 T8s = VSUB(T8c, T8f);
630 T8g = VADD(T8c, T8f);
631 ST(&(ri[0]), VADD(T2m, T45), ms, &(ri[0]));
632 ST(&(ri[WS(rs, 16)]), VSUB(T2m, T45), ms, &(ri[0]));
633 {
634 V T8v, T8Q, T8S, T8r;
635 T8v = VSUB(T8t, T8u);
636 T8Q = VADD(T8t, T8u);
637 T8S = VSUB(T8q, T8l);
638 T8r = VADD(T8l, T8q);
639 ST(&(ri[WS(rs, 8)]), VADD(T8w, T8z), ms, &(ri[0]));
640 ST(&(ri[WS(rs, 24)]), VSUB(T8w, T8z), ms, &(ri[0]));
641 ST(&(ii[WS(rs, 24)]), VSUB(T8M, T8L), ms, &(ii[0]));
642 ST(&(ii[WS(rs, 8)]), VADD(T8L, T8M), ms, &(ii[0]));
643 ST(&(ii[WS(rs, 16)]), VSUB(T8K, T8A), ms, &(ii[0]));
644 ST(&(ii[0]), VADD(T8A, T8K), ms, &(ii[0]));
645 ST(&(ri[WS(rs, 12)]), VFMA(LDK(KP707106781), T8v, T8s), ms, &(ri[0]));
646 ST(&(ri[WS(rs, 28)]), VFNMS(LDK(KP707106781), T8v, T8s), ms, &(ri[0]));
647 ST(&(ii[WS(rs, 20)]), VFNMS(LDK(KP707106781), T8Q, T8P), ms, &(ii[0]));
648 ST(&(ii[WS(rs, 4)]), VFMA(LDK(KP707106781), T8Q, T8P), ms, &(ii[0]));
649 ST(&(ii[WS(rs, 28)]), VFNMS(LDK(KP707106781), T8S, T8R), ms, &(ii[0]));
650 ST(&(ii[WS(rs, 12)]), VFMA(LDK(KP707106781), T8S, T8R), ms, &(ii[0]));
651 ST(&(ri[WS(rs, 4)]), VFMA(LDK(KP707106781), T8r, T8g), ms, &(ri[0]));
652 ST(&(ri[WS(rs, 20)]), VFNMS(LDK(KP707106781), T8r, T8g), ms, &(ri[0]));
653 }
654 }
655 }
656 T7P = VSUB(T7L, T7O);
657 T84 = VADD(T7L, T7O);
658 T85 = VADD(T7I, T7J);
659 T7K = VSUB(T7I, T7J);
660 T7X = VADD(T7m, T7l);
661 T7n = VSUB(T7l, T7m);
662 T7s = VADD(T7o, T7r);
663 T7Y = VSUB(T7o, T7r);
664 T82 = VADD(T7x, T7y);
665 T7z = VSUB(T7x, T7y);
666 T7W = VADD(T7e, T7h);
667 T7i = VSUB(T7e, T7h);
668 T8a = VFMA(LDK(KP414213562), T84, T85);
669 T86 = VFNMS(LDK(KP414213562), T85, T84);
670 T91 = VSUB(T8U, T8T);
671 T8V = VADD(T8T, T8U);
672 T8W = VADD(T7n, T7s);
673 T7t = VSUB(T7n, T7s);
674 T7E = VSUB(T7A, T7D);
675 T81 = VADD(T7A, T7D);
676 }
677 {
678 V T7S, T7u, T7T, T7F, T92, T7Z, T89, T83, T7U, T7Q;
679 T7S = VFNMS(LDK(KP707106781), T7t, T7i);
680 T7u = VFMA(LDK(KP707106781), T7t, T7i);
681 T7T = VFNMS(LDK(KP414213562), T7z, T7E);
682 T7F = VFMA(LDK(KP414213562), T7E, T7z);
683 T92 = VSUB(T7Y, T7X);
684 T7Z = VADD(T7X, T7Y);
685 T89 = VFNMS(LDK(KP414213562), T81, T82);
686 T83 = VFMA(LDK(KP414213562), T82, T81);
687 T7U = VFMA(LDK(KP414213562), T7K, T7P);
688 T7Q = VFNMS(LDK(KP414213562), T7P, T7K);
689 {
690 V T8X, T95, T93, T80, T88, T87, T7V, T94, T96, T7R, T8Y, T8b;
691 T8Z = VFNMS(LDK(KP707106781), T8W, T8V);
692 T8X = VFMA(LDK(KP707106781), T8W, T8V);
693 T95 = VFNMS(LDK(KP707106781), T92, T91);
694 T93 = VFMA(LDK(KP707106781), T92, T91);
695 T80 = VFMA(LDK(KP707106781), T7Z, T7W);
696 T88 = VFNMS(LDK(KP707106781), T7Z, T7W);
697 T90 = VSUB(T86, T83);
698 T87 = VADD(T83, T86);
699 T7V = VADD(T7T, T7U);
700 T94 = VSUB(T7U, T7T);
701 T96 = VADD(T7F, T7Q);
702 T7R = VSUB(T7F, T7Q);
703 T8Y = VADD(T89, T8a);
704 T8b = VSUB(T89, T8a);
705 ST(&(ri[WS(rs, 2)]), VFMA(LDK(KP923879532), T87, T80), ms, &(ri[0]));
706 ST(&(ri[WS(rs, 18)]), VFNMS(LDK(KP923879532), T87, T80), ms, &(ri[0]));
707 ST(&(ri[WS(rs, 30)]), VFMA(LDK(KP923879532), T7V, T7S), ms, &(ri[0]));
708 ST(&(ri[WS(rs, 14)]), VFNMS(LDK(KP923879532), T7V, T7S), ms, &(ri[0]));
709 ST(&(ii[WS(rs, 22)]), VFNMS(LDK(KP923879532), T94, T93), ms, &(ii[0]));
710 ST(&(ii[WS(rs, 6)]), VFMA(LDK(KP923879532), T94, T93), ms, &(ii[0]));
711 ST(&(ii[WS(rs, 30)]), VFMA(LDK(KP923879532), T96, T95), ms, &(ii[0]));
712 ST(&(ii[WS(rs, 14)]), VFNMS(LDK(KP923879532), T96, T95), ms, &(ii[0]));
713 ST(&(ri[WS(rs, 6)]), VFMA(LDK(KP923879532), T7R, T7u), ms, &(ri[0]));
714 ST(&(ri[WS(rs, 22)]), VFNMS(LDK(KP923879532), T7R, T7u), ms, &(ri[0]));
715 ST(&(ii[WS(rs, 18)]), VFNMS(LDK(KP923879532), T8Y, T8X), ms, &(ii[0]));
716 ST(&(ii[WS(rs, 2)]), VFMA(LDK(KP923879532), T8Y, T8X), ms, &(ii[0]));
717 ST(&(ri[WS(rs, 26)]), VFNMS(LDK(KP923879532), T8b, T88), ms, &(ri[0]));
718 ST(&(ri[WS(rs, 10)]), VFMA(LDK(KP923879532), T8b, T88), ms, &(ri[0]));
719 }
720 }
721 }
722 }
723 {
724 V T6s, T9o, T9n, T6v, T6N, T6Q, T6G, T6J, T68, T4Y, T9f, T9d, T9l, T9j, T6g;
725 V T6o, T6q, T6m, T66, T6a, T6p, T6j, T5x, T69;
726 {
727 V T6d, T6e, T6c, T4s, T9c, T4X, T9h, T9b, T5T, T64, T5k, T5v, T9i, T6f;
728 {
729 V T4c, T4r, T4H, T4W, T99, T9a;
730 T6s = VSUB(T46, T4b);
731 T4c = VADD(T46, T4b);
732 T4r = VADD(T4j, T4q);
733 T9o = VSUB(T4q, T4j);
734 T6d = VFMA(LDK(KP414213562), T4z, T4G);
735 T4H = VFNMS(LDK(KP414213562), T4G, T4z);
736 T4W = VFMA(LDK(KP414213562), T4V, T4O);
737 T6e = VFNMS(LDK(KP414213562), T4O, T4V);
738 T9n = VADD(T98, T97);
739 T99 = VSUB(T97, T98);
740 T9a = VADD(T6t, T6u);
741 T6v = VSUB(T6t, T6u);
742 ST(&(ii[WS(rs, 26)]), VFNMS(LDK(KP923879532), T90, T8Z), ms, &(ii[0]));
743 ST(&(ii[WS(rs, 10)]), VFMA(LDK(KP923879532), T90, T8Z), ms, &(ii[0]));
744 T6c = VFMA(LDK(KP707106781), T4r, T4c);
745 T4s = VFNMS(LDK(KP707106781), T4r, T4c);
746 T9c = VADD(T4H, T4W);
747 T4X = VSUB(T4H, T4W);
748 T9h = VFNMS(LDK(KP707106781), T9a, T99);
749 T9b = VFMA(LDK(KP707106781), T9a, T99);
750 T6N = VSUB(T5S, T5L);
751 T5T = VADD(T5L, T5S);
752 T64 = VADD(T62, T63);
753 T6Q = VSUB(T62, T63);
754 T6G = VSUB(T5j, T5c);
755 T5k = VADD(T5c, T5j);
756 T5v = VADD(T5t, T5u);
757 T6J = VSUB(T5t, T5u);
758 }
759 T68 = VFNMS(LDK(KP923879532), T4X, T4s);
760 T4Y = VFMA(LDK(KP923879532), T4X, T4s);
761 T9f = VFNMS(LDK(KP923879532), T9c, T9b);
762 T9d = VFMA(LDK(KP923879532), T9c, T9b);
763 T9i = VSUB(T6e, T6d);
764 T6f = VADD(T6d, T6e);
765 {
766 V T6l, T5U, T6k, T65;
767 T6l = VFMA(LDK(KP707106781), T5T, T5E);
768 T5U = VFNMS(LDK(KP707106781), T5T, T5E);
769 T6k = VFMA(LDK(KP707106781), T64, T61);
770 T65 = VFNMS(LDK(KP707106781), T64, T61);
771 {
772 V T6i, T5l, T6h, T5w;
773 T6i = VFMA(LDK(KP707106781), T5k, T55);
774 T5l = VFNMS(LDK(KP707106781), T5k, T55);
775 T6h = VFMA(LDK(KP707106781), T5v, T5s);
776 T5w = VFNMS(LDK(KP707106781), T5v, T5s);
777 T9l = VFNMS(LDK(KP923879532), T9i, T9h);
778 T9j = VFMA(LDK(KP923879532), T9i, T9h);
779 T6g = VFMA(LDK(KP923879532), T6f, T6c);
780 T6o = VFNMS(LDK(KP923879532), T6f, T6c);
781 T6q = VFMA(LDK(KP198912367), T6k, T6l);
782 T6m = VFNMS(LDK(KP198912367), T6l, T6k);
783 T66 = VFNMS(LDK(KP668178637), T65, T5U);
784 T6a = VFMA(LDK(KP668178637), T5U, T65);
785 T6p = VFNMS(LDK(KP198912367), T6h, T6i);
786 T6j = VFMA(LDK(KP198912367), T6i, T6h);
787 T5x = VFMA(LDK(KP668178637), T5w, T5l);
788 T69 = VFNMS(LDK(KP668178637), T5l, T5w);
789 }
790 }
791 }
792 {
793 V T6Y, T6w, T9w, T6D, T9v, T9p, T9q, T71, T77, T6O, T76, T6R;
794 {
795 V T6Z, T6z, T6C, T70;
796 {
797 V T6n, T9g, T9e, T6r;
798 T6n = VADD(T6j, T6m);
799 T9g = VSUB(T6m, T6j);
800 T9e = VADD(T6p, T6q);
801 T6r = VSUB(T6p, T6q);
802 {
803 V T9k, T6b, T67, T9m;
804 T9k = VSUB(T6a, T69);
805 T6b = VADD(T69, T6a);
806 T67 = VSUB(T5x, T66);
807 T9m = VADD(T5x, T66);
808 ST(&(ii[WS(rs, 25)]), VFNMS(LDK(KP980785280), T9g, T9f), ms, &(ii[WS(rs, 1)]));
809 ST(&(ii[WS(rs, 9)]), VFMA(LDK(KP980785280), T9g, T9f), ms, &(ii[WS(rs, 1)]));
810 ST(&(ri[WS(rs, 1)]), VFMA(LDK(KP980785280), T6n, T6g), ms, &(ri[WS(rs, 1)]));
811 ST(&(ri[WS(rs, 17)]), VFNMS(LDK(KP980785280), T6n, T6g), ms, &(ri[WS(rs, 1)]));
812 ST(&(ri[WS(rs, 9)]), VFMA(LDK(KP980785280), T6r, T6o), ms, &(ri[WS(rs, 1)]));
813 ST(&(ri[WS(rs, 25)]), VFNMS(LDK(KP980785280), T6r, T6o), ms, &(ri[WS(rs, 1)]));
814 ST(&(ii[WS(rs, 17)]), VFNMS(LDK(KP980785280), T9e, T9d), ms, &(ii[WS(rs, 1)]));
815 ST(&(ii[WS(rs, 1)]), VFMA(LDK(KP980785280), T9e, T9d), ms, &(ii[WS(rs, 1)]));
816 ST(&(ri[WS(rs, 29)]), VFMA(LDK(KP831469612), T6b, T68), ms, &(ri[WS(rs, 1)]));
817 ST(&(ri[WS(rs, 13)]), VFNMS(LDK(KP831469612), T6b, T68), ms, &(ri[WS(rs, 1)]));
818 ST(&(ii[WS(rs, 21)]), VFNMS(LDK(KP831469612), T9k, T9j), ms, &(ii[WS(rs, 1)]));
819 ST(&(ii[WS(rs, 5)]), VFMA(LDK(KP831469612), T9k, T9j), ms, &(ii[WS(rs, 1)]));
820 ST(&(ii[WS(rs, 29)]), VFMA(LDK(KP831469612), T9m, T9l), ms, &(ii[WS(rs, 1)]));
821 ST(&(ii[WS(rs, 13)]), VFNMS(LDK(KP831469612), T9m, T9l), ms, &(ii[WS(rs, 1)]));
822 ST(&(ri[WS(rs, 5)]), VFMA(LDK(KP831469612), T67, T4Y), ms, &(ri[WS(rs, 1)]));
823 ST(&(ri[WS(rs, 21)]), VFNMS(LDK(KP831469612), T67, T4Y), ms, &(ri[WS(rs, 1)]));
824 T6Y = VFNMS(LDK(KP707106781), T6v, T6s);
825 T6w = VFMA(LDK(KP707106781), T6v, T6s);
826 }
827 }
828 T6Z = VFNMS(LDK(KP414213562), T6x, T6y);
829 T6z = VFMA(LDK(KP414213562), T6y, T6x);
830 T6C = VFNMS(LDK(KP414213562), T6B, T6A);
831 T70 = VFMA(LDK(KP414213562), T6A, T6B);
832 T9w = VADD(T6z, T6C);
833 T6D = VSUB(T6z, T6C);
834 T9v = VFNMS(LDK(KP707106781), T9o, T9n);
835 T9p = VFMA(LDK(KP707106781), T9o, T9n);
836 T9q = VSUB(T70, T6Z);
837 T71 = VADD(T6Z, T70);
838 T77 = VFMA(LDK(KP707106781), T6N, T6M);
839 T6O = VFNMS(LDK(KP707106781), T6N, T6M);
840 T76 = VFMA(LDK(KP707106781), T6Q, T6P);
841 T6R = VFNMS(LDK(KP707106781), T6Q, T6P);
842 T6H = VFNMS(LDK(KP707106781), T6G, T6F);
843 T74 = VFMA(LDK(KP707106781), T6G, T6F);
844 }
845 T6U = VFNMS(LDK(KP923879532), T6D, T6w);
846 T6E = VFMA(LDK(KP923879532), T6D, T6w);
847 T9r = VFMA(LDK(KP923879532), T9q, T9p);
848 T9t = VFNMS(LDK(KP923879532), T9q, T9p);
849 T78 = VFNMS(LDK(KP198912367), T77, T76);
850 T7c = VFMA(LDK(KP198912367), T76, T77);
851 T6W = VFMA(LDK(KP668178637), T6O, T6R);
852 T6S = VFNMS(LDK(KP668178637), T6R, T6O);
853 T73 = VFMA(LDK(KP707106781), T6J, T6I);
854 T6K = VFNMS(LDK(KP707106781), T6J, T6I);
855 T7a = VFMA(LDK(KP923879532), T71, T6Y);
856 T72 = VFNMS(LDK(KP923879532), T71, T6Y);
857 T9x = VFNMS(LDK(KP923879532), T9w, T9v);
858 T9z = VFMA(LDK(KP923879532), T9w, T9v);
859 }
860 }
861 }
862 }
863 }
864 {
865 V T7b, T75, T6L, T6V;
866 T7b = VFNMS(LDK(KP198912367), T73, T74);
867 T75 = VFMA(LDK(KP198912367), T74, T73);
868 T6L = VFMA(LDK(KP668178637), T6K, T6H);
869 T6V = VFNMS(LDK(KP668178637), T6H, T6K);
870 {
871 V T79, T9A, T9y, T7d;
872 T79 = VSUB(T75, T78);
873 T9A = VADD(T75, T78);
874 T9y = VSUB(T7c, T7b);
875 T7d = VADD(T7b, T7c);
876 {
877 V T9s, T6X, T6T, T9u;
878 T9s = VADD(T6V, T6W);
879 T6X = VSUB(T6V, T6W);
880 T6T = VADD(T6L, T6S);
881 T9u = VSUB(T6S, T6L);
882 ST(&(ii[WS(rs, 31)]), VFMA(LDK(KP980785280), T9A, T9z), ms, &(ii[WS(rs, 1)]));
883 ST(&(ii[WS(rs, 15)]), VFNMS(LDK(KP980785280), T9A, T9z), ms, &(ii[WS(rs, 1)]));
884 ST(&(ri[WS(rs, 7)]), VFMA(LDK(KP980785280), T79, T72), ms, &(ri[WS(rs, 1)]));
885 ST(&(ri[WS(rs, 23)]), VFNMS(LDK(KP980785280), T79, T72), ms, &(ri[WS(rs, 1)]));
886 ST(&(ri[WS(rs, 31)]), VFMA(LDK(KP980785280), T7d, T7a), ms, &(ri[WS(rs, 1)]));
887 ST(&(ri[WS(rs, 15)]), VFNMS(LDK(KP980785280), T7d, T7a), ms, &(ri[WS(rs, 1)]));
888 ST(&(ii[WS(rs, 23)]), VFNMS(LDK(KP980785280), T9y, T9x), ms, &(ii[WS(rs, 1)]));
889 ST(&(ii[WS(rs, 7)]), VFMA(LDK(KP980785280), T9y, T9x), ms, &(ii[WS(rs, 1)]));
890 ST(&(ri[WS(rs, 11)]), VFMA(LDK(KP831469612), T6X, T6U), ms, &(ri[WS(rs, 1)]));
891 ST(&(ri[WS(rs, 27)]), VFNMS(LDK(KP831469612), T6X, T6U), ms, &(ri[WS(rs, 1)]));
892 ST(&(ii[WS(rs, 19)]), VFNMS(LDK(KP831469612), T9s, T9r), ms, &(ii[WS(rs, 1)]));
893 ST(&(ii[WS(rs, 3)]), VFMA(LDK(KP831469612), T9s, T9r), ms, &(ii[WS(rs, 1)]));
894 ST(&(ii[WS(rs, 27)]), VFNMS(LDK(KP831469612), T9u, T9t), ms, &(ii[WS(rs, 1)]));
895 ST(&(ii[WS(rs, 11)]), VFMA(LDK(KP831469612), T9u, T9t), ms, &(ii[WS(rs, 1)]));
896 ST(&(ri[WS(rs, 3)]), VFMA(LDK(KP831469612), T6T, T6E), ms, &(ri[WS(rs, 1)]));
897 ST(&(ri[WS(rs, 19)]), VFNMS(LDK(KP831469612), T6T, T6E), ms, &(ri[WS(rs, 1)]));
898 }
899 }
900 }
901 }
902 }
903 VLEAVE();
904 }
905
906 static const tw_instr twinstr[] = {
907 VTW(0, 1),
908 VTW(0, 3),
909 VTW(0, 9),
910 VTW(0, 27),
911 {TW_NEXT, (2 * VL), 0}
912 };
913
914 static const ct_desc desc = { 32, XSIMD_STRING("t2sv_32"), twinstr, &GENUS, {236, 98, 252, 0}, 0, 0, 0 };
915
916 void XSIMD(codelet_t2sv_32) (planner *p) {
917 X(kdft_dit_register) (p, t2sv_32, &desc);
918 }
919 #else /* HAVE_FMA */
920
921 /* Generated by: ../../../genfft/gen_twiddle.native -simd -compact -variables 4 -pipeline-latency 8 -twiddle-log3 -precompute-twiddles -n 32 -name t2sv_32 -include ts.h */
922
923 /*
924 * This function contains 488 FP additions, 280 FP multiplications,
925 * (or, 376 additions, 168 multiplications, 112 fused multiply/add),
926 * 158 stack variables, 7 constants, and 128 memory accesses
927 */
928 #include "ts.h"
929
930 static void t2sv_32(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
931 {
932 DVK(KP195090322, +0.195090322016128267848284868477022240927691618);
933 DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
934 DVK(KP555570233, +0.555570233019602224742830813948532874374937191);
935 DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
936 DVK(KP382683432, +0.382683432365089771728459984030398866761344562);
937 DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
938 DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
939 {
940 INT m;
941 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)) {
942 V T2, T5, T3, T6, T8, TM, TO, Td, T9, Te, Th, Tl, TD, TH, T1y;
943 V T1H, T15, T1A, T11, T1F, T1n, T1p, T2q, T2I, T2u, T2K, T2V, T3b, T2Z, T3d;
944 V Tu, Ty, T3l, T3n, T1t, T1v, T2f, T2h, T1a, T1e, T32, T34, T1W, T1Y, T2C;
945 V T2E, Tg, TR, Tk, TS, Tm, TV, To, TT, T1M, T21, T1P, T22, T1Q, T25;
946 V T1S, T23;
947 {
948 V Ts, T1d, Tx, T18, Tt, T1c, Tw, T19, TB, T14, TG, TZ, TC, T13, TF;
949 V T10;
950 {
951 V T4, Tc, T7, Tb;
952 T2 = LDW(&(W[0]));
953 T5 = LDW(&(W[TWVL * 1]));
954 T3 = LDW(&(W[TWVL * 2]));
955 T6 = LDW(&(W[TWVL * 3]));
956 T4 = VMUL(T2, T3);
957 Tc = VMUL(T5, T3);
958 T7 = VMUL(T5, T6);
959 Tb = VMUL(T2, T6);
960 T8 = VADD(T4, T7);
961 TM = VSUB(T4, T7);
962 TO = VADD(Tb, Tc);
963 Td = VSUB(Tb, Tc);
964 T9 = LDW(&(W[TWVL * 4]));
965 Ts = VMUL(T2, T9);
966 T1d = VMUL(T6, T9);
967 Tx = VMUL(T5, T9);
968 T18 = VMUL(T3, T9);
969 Te = LDW(&(W[TWVL * 5]));
970 Tt = VMUL(T5, Te);
971 T1c = VMUL(T3, Te);
972 Tw = VMUL(T2, Te);
973 T19 = VMUL(T6, Te);
974 Th = LDW(&(W[TWVL * 6]));
975 TB = VMUL(T3, Th);
976 T14 = VMUL(T5, Th);
977 TG = VMUL(T6, Th);
978 TZ = VMUL(T2, Th);
979 Tl = LDW(&(W[TWVL * 7]));
980 TC = VMUL(T6, Tl);
981 T13 = VMUL(T2, Tl);
982 TF = VMUL(T3, Tl);
983 T10 = VMUL(T5, Tl);
984 }
985 TD = VADD(TB, TC);
986 TH = VSUB(TF, TG);
987 T1y = VADD(TZ, T10);
988 T1H = VADD(TF, TG);
989 T15 = VADD(T13, T14);
990 T1A = VSUB(T13, T14);
991 T11 = VSUB(TZ, T10);
992 T1F = VSUB(TB, TC);
993 T1n = VFMA(T9, Th, VMUL(Te, Tl));
994 T1p = VFNMS(Te, Th, VMUL(T9, Tl));
995 {
996 V T2o, T2p, T2s, T2t;
997 T2o = VMUL(T8, Th);
998 T2p = VMUL(Td, Tl);
999 T2q = VADD(T2o, T2p);
1000 T2I = VSUB(T2o, T2p);
1001 T2s = VMUL(T8, Tl);
1002 T2t = VMUL(Td, Th);
1003 T2u = VSUB(T2s, T2t);
1004 T2K = VADD(T2s, T2t);
1005 }
1006 {
1007 V T2T, T2U, T2X, T2Y;
1008 T2T = VMUL(TM, Th);
1009 T2U = VMUL(TO, Tl);
1010 T2V = VSUB(T2T, T2U);
1011 T3b = VADD(T2T, T2U);
1012 T2X = VMUL(TM, Tl);
1013 T2Y = VMUL(TO, Th);
1014 T2Z = VADD(T2X, T2Y);
1015 T3d = VSUB(T2X, T2Y);
1016 Tu = VADD(Ts, Tt);
1017 Ty = VSUB(Tw, Tx);
1018 T3l = VFMA(Tu, Th, VMUL(Ty, Tl));
1019 T3n = VFNMS(Ty, Th, VMUL(Tu, Tl));
1020 }
1021 T1t = VSUB(Ts, Tt);
1022 T1v = VADD(Tw, Tx);
1023 T2f = VFMA(T1t, Th, VMUL(T1v, Tl));
1024 T2h = VFNMS(T1v, Th, VMUL(T1t, Tl));
1025 T1a = VSUB(T18, T19);
1026 T1e = VADD(T1c, T1d);
1027 T32 = VFMA(T1a, Th, VMUL(T1e, Tl));
1028 T34 = VFNMS(T1e, Th, VMUL(T1a, Tl));
1029 T1W = VADD(T18, T19);
1030 T1Y = VSUB(T1c, T1d);
1031 T2C = VFMA(T1W, Th, VMUL(T1Y, Tl));
1032 T2E = VFNMS(T1Y, Th, VMUL(T1W, Tl));
1033 {
1034 V Ta, Tf, Ti, Tj;
1035 Ta = VMUL(T8, T9);
1036 Tf = VMUL(Td, Te);
1037 Tg = VSUB(Ta, Tf);
1038 TR = VADD(Ta, Tf);
1039 Ti = VMUL(T8, Te);
1040 Tj = VMUL(Td, T9);
1041 Tk = VADD(Ti, Tj);
1042 TS = VSUB(Ti, Tj);
1043 }
1044 Tm = VFMA(Tg, Th, VMUL(Tk, Tl));
1045 TV = VFNMS(TS, Th, VMUL(TR, Tl));
1046 To = VFNMS(Tk, Th, VMUL(Tg, Tl));
1047 TT = VFMA(TR, Th, VMUL(TS, Tl));
1048 {
1049 V T1K, T1L, T1N, T1O;
1050 T1K = VMUL(TM, T9);
1051 T1L = VMUL(TO, Te);
1052 T1M = VSUB(T1K, T1L);
1053 T21 = VADD(T1K, T1L);
1054 T1N = VMUL(TM, Te);
1055 T1O = VMUL(TO, T9);
1056 T1P = VADD(T1N, T1O);
1057 T22 = VSUB(T1N, T1O);
1058 }
1059 T1Q = VFMA(T1M, Th, VMUL(T1P, Tl));
1060 T25 = VFNMS(T22, Th, VMUL(T21, Tl));
1061 T1S = VFNMS(T1P, Th, VMUL(T1M, Tl));
1062 T23 = VFMA(T21, Th, VMUL(T22, Tl));
1063 }
1064 {
1065 V TL, T6f, T8c, T8q, T3F, T5t, T7I, T7W, T2y, T6B, T6y, T7j, T4k, T5J, T4B;
1066 V T5G, T3h, T6H, T6O, T7o, T4L, T5N, T52, T5Q, T1i, T7V, T6i, T7D, T3K, T5u;
1067 V T3P, T5v, T1E, T6n, T6m, T7e, T3W, T5y, T41, T5z, T29, T6p, T6s, T7f, T47;
1068 V T5B, T4c, T5C, T2R, T6z, T6E, T7k, T4v, T5H, T4E, T5K, T3y, T6P, T6K, T7p;
1069 V T4W, T5R, T55, T5O;
1070 {
1071 V T1, T7G, Tq, T7F, TA, T3C, TJ, T3D, Tn, Tp;
1072 T1 = LD(&(ri[0]), ms, &(ri[0]));
1073 T7G = LD(&(ii[0]), ms, &(ii[0]));
1074 Tn = LD(&(ri[WS(rs, 16)]), ms, &(ri[0]));
1075 Tp = LD(&(ii[WS(rs, 16)]), ms, &(ii[0]));
1076 Tq = VFMA(Tm, Tn, VMUL(To, Tp));
1077 T7F = VFNMS(To, Tn, VMUL(Tm, Tp));
1078 {
1079 V Tv, Tz, TE, TI;
1080 Tv = LD(&(ri[WS(rs, 8)]), ms, &(ri[0]));
1081 Tz = LD(&(ii[WS(rs, 8)]), ms, &(ii[0]));
1082 TA = VFMA(Tu, Tv, VMUL(Ty, Tz));
1083 T3C = VFNMS(Ty, Tv, VMUL(Tu, Tz));
1084 TE = LD(&(ri[WS(rs, 24)]), ms, &(ri[0]));
1085 TI = LD(&(ii[WS(rs, 24)]), ms, &(ii[0]));
1086 TJ = VFMA(TD, TE, VMUL(TH, TI));
1087 T3D = VFNMS(TH, TE, VMUL(TD, TI));
1088 }
1089 {
1090 V Tr, TK, T8a, T8b;
1091 Tr = VADD(T1, Tq);
1092 TK = VADD(TA, TJ);
1093 TL = VADD(Tr, TK);
1094 T6f = VSUB(Tr, TK);
1095 T8a = VSUB(T7G, T7F);
1096 T8b = VSUB(TA, TJ);
1097 T8c = VSUB(T8a, T8b);
1098 T8q = VADD(T8b, T8a);
1099 }
1100 {
1101 V T3B, T3E, T7E, T7H;
1102 T3B = VSUB(T1, Tq);
1103 T3E = VSUB(T3C, T3D);
1104 T3F = VSUB(T3B, T3E);
1105 T5t = VADD(T3B, T3E);
1106 T7E = VADD(T3C, T3D);
1107 T7H = VADD(T7F, T7G);
1108 T7I = VADD(T7E, T7H);
1109 T7W = VSUB(T7H, T7E);
1110 }
1111 }
1112 {
1113 V T2e, T4g, T2w, T4z, T2j, T4h, T2n, T4y;
1114 {
1115 V T2c, T2d, T2r, T2v;
1116 T2c = LD(&(ri[WS(rs, 1)]), ms, &(ri[WS(rs, 1)]));
1117 T2d = LD(&(ii[WS(rs, 1)]), ms, &(ii[WS(rs, 1)]));
1118 T2e = VFMA(T2, T2c, VMUL(T5, T2d));
1119 T4g = VFNMS(T5, T2c, VMUL(T2, T2d));
1120 T2r = LD(&(ri[WS(rs, 25)]), ms, &(ri[WS(rs, 1)]));
1121 T2v = LD(&(ii[WS(rs, 25)]), ms, &(ii[WS(rs, 1)]));
1122 T2w = VFMA(T2q, T2r, VMUL(T2u, T2v));
1123 T4z = VFNMS(T2u, T2r, VMUL(T2q, T2v));
1124 }
1125 {
1126 V T2g, T2i, T2l, T2m;
1127 T2g = LD(&(ri[WS(rs, 17)]), ms, &(ri[WS(rs, 1)]));
1128 T2i = LD(&(ii[WS(rs, 17)]), ms, &(ii[WS(rs, 1)]));
1129 T2j = VFMA(T2f, T2g, VMUL(T2h, T2i));
1130 T4h = VFNMS(T2h, T2g, VMUL(T2f, T2i));
1131 T2l = LD(&(ri[WS(rs, 9)]), ms, &(ri[WS(rs, 1)]));
1132 T2m = LD(&(ii[WS(rs, 9)]), ms, &(ii[WS(rs, 1)]));
1133 T2n = VFMA(T9, T2l, VMUL(Te, T2m));
1134 T4y = VFNMS(Te, T2l, VMUL(T9, T2m));
1135 }
1136 {
1137 V T2k, T2x, T6w, T6x;
1138 T2k = VADD(T2e, T2j);
1139 T2x = VADD(T2n, T2w);
1140 T2y = VADD(T2k, T2x);
1141 T6B = VSUB(T2k, T2x);
1142 T6w = VADD(T4g, T4h);
1143 T6x = VADD(T4y, T4z);
1144 T6y = VSUB(T6w, T6x);
1145 T7j = VADD(T6w, T6x);
1146 }
1147 {
1148 V T4i, T4j, T4x, T4A;
1149 T4i = VSUB(T4g, T4h);
1150 T4j = VSUB(T2n, T2w);
1151 T4k = VADD(T4i, T4j);
1152 T5J = VSUB(T4i, T4j);
1153 T4x = VSUB(T2e, T2j);
1154 T4A = VSUB(T4y, T4z);
1155 T4B = VSUB(T4x, T4A);
1156 T5G = VADD(T4x, T4A);
1157 }
1158 }
1159 {
1160 V T31, T4Y, T3f, T4J, T36, T4Z, T3a, T4I;
1161 {
1162 V T2W, T30, T3c, T3e;
1163 T2W = LD(&(ri[WS(rs, 31)]), ms, &(ri[WS(rs, 1)]));
1164 T30 = LD(&(ii[WS(rs, 31)]), ms, &(ii[WS(rs, 1)]));
1165 T31 = VFMA(T2V, T2W, VMUL(T2Z, T30));
1166 T4Y = VFNMS(T2Z, T2W, VMUL(T2V, T30));
1167 T3c = LD(&(ri[WS(rs, 23)]), ms, &(ri[WS(rs, 1)]));
1168 T3e = LD(&(ii[WS(rs, 23)]), ms, &(ii[WS(rs, 1)]));
1169 T3f = VFMA(T3b, T3c, VMUL(T3d, T3e));
1170 T4J = VFNMS(T3d, T3c, VMUL(T3b, T3e));
1171 }
1172 {
1173 V T33, T35, T38, T39;
1174 T33 = LD(&(ri[WS(rs, 15)]), ms, &(ri[WS(rs, 1)]));
1175 T35 = LD(&(ii[WS(rs, 15)]), ms, &(ii[WS(rs, 1)]));
1176 T36 = VFMA(T32, T33, VMUL(T34, T35));
1177 T4Z = VFNMS(T34, T33, VMUL(T32, T35));
1178 T38 = LD(&(ri[WS(rs, 7)]), ms, &(ri[WS(rs, 1)]));
1179 T39 = LD(&(ii[WS(rs, 7)]), ms, &(ii[WS(rs, 1)]));
1180 T3a = VFMA(TR, T38, VMUL(TS, T39));
1181 T4I = VFNMS(TS, T38, VMUL(TR, T39));
1182 }
1183 {
1184 V T37, T3g, T6M, T6N;
1185 T37 = VADD(T31, T36);
1186 T3g = VADD(T3a, T3f);
1187 T3h = VADD(T37, T3g);
1188 T6H = VSUB(T37, T3g);
1189 T6M = VADD(T4Y, T4Z);
1190 T6N = VADD(T4I, T4J);
1191 T6O = VSUB(T6M, T6N);
1192 T7o = VADD(T6M, T6N);
1193 }
1194 {
1195 V T4H, T4K, T50, T51;
1196 T4H = VSUB(T31, T36);
1197 T4K = VSUB(T4I, T4J);
1198 T4L = VSUB(T4H, T4K);
1199 T5N = VADD(T4H, T4K);
1200 T50 = VSUB(T4Y, T4Z);
1201 T51 = VSUB(T3a, T3f);
1202 T52 = VADD(T50, T51);
1203 T5Q = VSUB(T50, T51);
1204 }
1205 }
1206 {
1207 V TQ, T3G, T1g, T3N, TX, T3H, T17, T3M;
1208 {
1209 V TN, TP, T1b, T1f;
1210 TN = LD(&(ri[WS(rs, 4)]), ms, &(ri[0]));
1211 TP = LD(&(ii[WS(rs, 4)]), ms, &(ii[0]));
1212 TQ = VFMA(TM, TN, VMUL(TO, TP));
1213 T3G = VFNMS(TO, TN, VMUL(TM, TP));
1214 T1b = LD(&(ri[WS(rs, 12)]), ms, &(ri[0]));
1215 T1f = LD(&(ii[WS(rs, 12)]), ms, &(ii[0]));
1216 T1g = VFMA(T1a, T1b, VMUL(T1e, T1f));
1217 T3N = VFNMS(T1e, T1b, VMUL(T1a, T1f));
1218 }
1219 {
1220 V TU, TW, T12, T16;
1221 TU = LD(&(ri[WS(rs, 20)]), ms, &(ri[0]));
1222 TW = LD(&(ii[WS(rs, 20)]), ms, &(ii[0]));
1223 TX = VFMA(TT, TU, VMUL(TV, TW));
1224 T3H = VFNMS(TV, TU, VMUL(TT, TW));
1225 T12 = LD(&(ri[WS(rs, 28)]), ms, &(ri[0]));
1226 T16 = LD(&(ii[WS(rs, 28)]), ms, &(ii[0]));
1227 T17 = VFMA(T11, T12, VMUL(T15, T16));
1228 T3M = VFNMS(T15, T12, VMUL(T11, T16));
1229 }
1230 {
1231 V TY, T1h, T6g, T6h;
1232 TY = VADD(TQ, TX);
1233 T1h = VADD(T17, T1g);
1234 T1i = VADD(TY, T1h);
1235 T7V = VSUB(T1h, TY);
1236 T6g = VADD(T3G, T3H);
1237 T6h = VADD(T3M, T3N);
1238 T6i = VSUB(T6g, T6h);
1239 T7D = VADD(T6g, T6h);
1240 }
1241 {
1242 V T3I, T3J, T3L, T3O;
1243 T3I = VSUB(T3G, T3H);
1244 T3J = VSUB(TQ, TX);
1245 T3K = VSUB(T3I, T3J);
1246 T5u = VADD(T3J, T3I);
1247 T3L = VSUB(T17, T1g);
1248 T3O = VSUB(T3M, T3N);
1249 T3P = VADD(T3L, T3O);
1250 T5v = VSUB(T3L, T3O);
1251 }
1252 }
1253 {
1254 V T1m, T3S, T1C, T3Z, T1r, T3T, T1x, T3Y;
1255 {
1256 V T1k, T1l, T1z, T1B;
1257 T1k = LD(&(ri[WS(rs, 2)]), ms, &(ri[0]));
1258 T1l = LD(&(ii[WS(rs, 2)]), ms, &(ii[0]));
1259 T1m = VFMA(T8, T1k, VMUL(Td, T1l));
1260 T3S = VFNMS(Td, T1k, VMUL(T8, T1l));
1261 T1z = LD(&(ri[WS(rs, 26)]), ms, &(ri[0]));
1262 T1B = LD(&(ii[WS(rs, 26)]), ms, &(ii[0]));
1263 T1C = VFMA(T1y, T1z, VMUL(T1A, T1B));
1264 T3Z = VFNMS(T1A, T1z, VMUL(T1y, T1B));
1265 }
1266 {
1267 V T1o, T1q, T1u, T1w;
1268 T1o = LD(&(ri[WS(rs, 18)]), ms, &(ri[0]));
1269 T1q = LD(&(ii[WS(rs, 18)]), ms, &(ii[0]));
1270 T1r = VFMA(T1n, T1o, VMUL(T1p, T1q));
1271 T3T = VFNMS(T1p, T1o, VMUL(T1n, T1q));
1272 T1u = LD(&(ri[WS(rs, 10)]), ms, &(ri[0]));
1273 T1w = LD(&(ii[WS(rs, 10)]), ms, &(ii[0]));
1274 T1x = VFMA(T1t, T1u, VMUL(T1v, T1w));
1275 T3Y = VFNMS(T1v, T1u, VMUL(T1t, T1w));
1276 }
1277 {
1278 V T1s, T1D, T6k, T6l;
1279 T1s = VADD(T1m, T1r);
1280 T1D = VADD(T1x, T1C);
1281 T1E = VADD(T1s, T1D);
1282 T6n = VSUB(T1s, T1D);
1283 T6k = VADD(T3S, T3T);
1284 T6l = VADD(T3Y, T3Z);
1285 T6m = VSUB(T6k, T6l);
1286 T7e = VADD(T6k, T6l);
1287 }
1288 {
1289 V T3U, T3V, T3X, T40;
1290 T3U = VSUB(T3S, T3T);
1291 T3V = VSUB(T1x, T1C);
1292 T3W = VADD(T3U, T3V);
1293 T5y = VSUB(T3U, T3V);
1294 T3X = VSUB(T1m, T1r);
1295 T40 = VSUB(T3Y, T3Z);
1296 T41 = VSUB(T3X, T40);
1297 T5z = VADD(T3X, T40);
1298 }
1299 }
1300 {
1301 V T1J, T43, T27, T4a, T1U, T44, T20, T49;
1302 {
1303 V T1G, T1I, T24, T26;
1304 T1G = LD(&(ri[WS(rs, 30)]), ms, &(ri[0]));
1305 T1I = LD(&(ii[WS(rs, 30)]), ms, &(ii[0]));
1306 T1J = VFMA(T1F, T1G, VMUL(T1H, T1I));
1307 T43 = VFNMS(T1H, T1G, VMUL(T1F, T1I));
1308 T24 = LD(&(ri[WS(rs, 22)]), ms, &(ri[0]));
1309 T26 = LD(&(ii[WS(rs, 22)]), ms, &(ii[0]));
1310 T27 = VFMA(T23, T24, VMUL(T25, T26));
1311 T4a = VFNMS(T25, T24, VMUL(T23, T26));
1312 }
1313 {
1314 V T1R, T1T, T1X, T1Z;
1315 T1R = LD(&(ri[WS(rs, 14)]), ms, &(ri[0]));
1316 T1T = LD(&(ii[WS(rs, 14)]), ms, &(ii[0]));
1317 T1U = VFMA(T1Q, T1R, VMUL(T1S, T1T));
1318 T44 = VFNMS(T1S, T1R, VMUL(T1Q, T1T));
1319 T1X = LD(&(ri[WS(rs, 6)]), ms, &(ri[0]));
1320 T1Z = LD(&(ii[WS(rs, 6)]), ms, &(ii[0]));
1321 T20 = VFMA(T1W, T1X, VMUL(T1Y, T1Z));
1322 T49 = VFNMS(T1Y, T1X, VMUL(T1W, T1Z));
1323 }
1324 {
1325 V T1V, T28, T6q, T6r;
1326 T1V = VADD(T1J, T1U);
1327 T28 = VADD(T20, T27);
1328 T29 = VADD(T1V, T28);
1329 T6p = VSUB(T1V, T28);
1330 T6q = VADD(T43, T44);
1331 T6r = VADD(T49, T4a);
1332 T6s = VSUB(T6q, T6r);
1333 T7f = VADD(T6q, T6r);
1334 }
1335 {
1336 V T45, T46, T48, T4b;
1337 T45 = VSUB(T43, T44);
1338 T46 = VSUB(T20, T27);
1339 T47 = VADD(T45, T46);
1340 T5B = VSUB(T45, T46);
1341 T48 = VSUB(T1J, T1U);
1342 T4b = VSUB(T49, T4a);
1343 T4c = VSUB(T48, T4b);
1344 T5C = VADD(T48, T4b);
1345 }
1346 }
1347 {
1348 V T2B, T4r, T2G, T4s, T4q, T4t, T2M, T4m, T2P, T4n, T4l, T4o;
1349 {
1350 V T2z, T2A, T2D, T2F;
1351 T2z = LD(&(ri[WS(rs, 5)]), ms, &(ri[WS(rs, 1)]));
1352 T2A = LD(&(ii[WS(rs, 5)]), ms, &(ii[WS(rs, 1)]));
1353 T2B = VFMA(T21, T2z, VMUL(T22, T2A));
1354 T4r = VFNMS(T22, T2z, VMUL(T21, T2A));
1355 T2D = LD(&(ri[WS(rs, 21)]), ms, &(ri[WS(rs, 1)]));
1356 T2F = LD(&(ii[WS(rs, 21)]), ms, &(ii[WS(rs, 1)]));
1357 T2G = VFMA(T2C, T2D, VMUL(T2E, T2F));
1358 T4s = VFNMS(T2E, T2D, VMUL(T2C, T2F));
1359 }
1360 T4q = VSUB(T2B, T2G);
1361 T4t = VSUB(T4r, T4s);
1362 {
1363 V T2J, T2L, T2N, T2O;
1364 T2J = LD(&(ri[WS(rs, 29)]), ms, &(ri[WS(rs, 1)]));
1365 T2L = LD(&(ii[WS(rs, 29)]), ms, &(ii[WS(rs, 1)]));
1366 T2M = VFMA(T2I, T2J, VMUL(T2K, T2L));
1367 T4m = VFNMS(T2K, T2J, VMUL(T2I, T2L));
1368 T2N = LD(&(ri[WS(rs, 13)]), ms, &(ri[WS(rs, 1)]));
1369 T2O = LD(&(ii[WS(rs, 13)]), ms, &(ii[WS(rs, 1)]));
1370 T2P = VFMA(T1M, T2N, VMUL(T1P, T2O));
1371 T4n = VFNMS(T1P, T2N, VMUL(T1M, T2O));
1372 }
1373 T4l = VSUB(T2M, T2P);
1374 T4o = VSUB(T4m, T4n);
1375 {
1376 V T2H, T2Q, T6C, T6D;
1377 T2H = VADD(T2B, T2G);
1378 T2Q = VADD(T2M, T2P);
1379 T2R = VADD(T2H, T2Q);
1380 T6z = VSUB(T2Q, T2H);
1381 T6C = VADD(T4r, T4s);
1382 T6D = VADD(T4m, T4n);
1383 T6E = VSUB(T6C, T6D);
1384 T7k = VADD(T6C, T6D);
1385 }
1386 {
1387 V T4p, T4u, T4C, T4D;
1388 T4p = VSUB(T4l, T4o);
1389 T4u = VADD(T4q, T4t);
1390 T4v = VMUL(LDK(KP707106781), VSUB(T4p, T4u));
1391 T5H = VMUL(LDK(KP707106781), VADD(T4u, T4p));
1392 T4C = VSUB(T4t, T4q);
1393 T4D = VADD(T4l, T4o);
1394 T4E = VMUL(LDK(KP707106781), VSUB(T4C, T4D));
1395 T5K = VMUL(LDK(KP707106781), VADD(T4C, T4D));
1396 }
1397 }
1398 {
1399 V T3k, T4M, T3p, T4N, T4O, T4P, T3t, T4S, T3w, T4T, T4R, T4U;
1400 {
1401 V T3i, T3j, T3m, T3o;
1402 T3i = LD(&(ri[WS(rs, 3)]), ms, &(ri[WS(rs, 1)]));
1403 T3j = LD(&(ii[WS(rs, 3)]), ms, &(ii[WS(rs, 1)]));
1404 T3k = VFMA(T3, T3i, VMUL(T6, T3j));
1405 T4M = VFNMS(T6, T3i, VMUL(T3, T3j));
1406 T3m = LD(&(ri[WS(rs, 19)]), ms, &(ri[WS(rs, 1)]));
1407 T3o = LD(&(ii[WS(rs, 19)]), ms, &(ii[WS(rs, 1)]));
1408 T3p = VFMA(T3l, T3m, VMUL(T3n, T3o));
1409 T4N = VFNMS(T3n, T3m, VMUL(T3l, T3o));
1410 }
1411 T4O = VSUB(T4M, T4N);
1412 T4P = VSUB(T3k, T3p);
1413 {
1414 V T3r, T3s, T3u, T3v;
1415 T3r = LD(&(ri[WS(rs, 27)]), ms, &(ri[WS(rs, 1)]));
1416 T3s = LD(&(ii[WS(rs, 27)]), ms, &(ii[WS(rs, 1)]));
1417 T3t = VFMA(Th, T3r, VMUL(Tl, T3s));
1418 T4S = VFNMS(Tl, T3r, VMUL(Th, T3s));
1419 T3u = LD(&(ri[WS(rs, 11)]), ms, &(ri[WS(rs, 1)]));
1420 T3v = LD(&(ii[WS(rs, 11)]), ms, &(ii[WS(rs, 1)]));
1421 T3w = VFMA(Tg, T3u, VMUL(Tk, T3v));
1422 T4T = VFNMS(Tk, T3u, VMUL(Tg, T3v));
1423 }
1424 T4R = VSUB(T3t, T3w);
1425 T4U = VSUB(T4S, T4T);
1426 {
1427 V T3q, T3x, T6I, T6J;
1428 T3q = VADD(T3k, T3p);
1429 T3x = VADD(T3t, T3w);
1430 T3y = VADD(T3q, T3x);
1431 T6P = VSUB(T3x, T3q);
1432 T6I = VADD(T4M, T4N);
1433 T6J = VADD(T4S, T4T);
1434 T6K = VSUB(T6I, T6J);
1435 T7p = VADD(T6I, T6J);
1436 }
1437 {
1438 V T4Q, T4V, T53, T54;
1439 T4Q = VSUB(T4O, T4P);
1440 T4V = VADD(T4R, T4U);
1441 T4W = VMUL(LDK(KP707106781), VSUB(T4Q, T4V));
1442 T5R = VMUL(LDK(KP707106781), VADD(T4Q, T4V));
1443 T53 = VSUB(T4R, T4U);
1444 T54 = VADD(T4P, T4O);
1445 T55 = VMUL(LDK(KP707106781), VSUB(T53, T54));
1446 T5O = VMUL(LDK(KP707106781), VADD(T54, T53));
1447 }
1448 }
1449 {
1450 V T2b, T7x, T7K, T7M, T3A, T7L, T7A, T7B;
1451 {
1452 V T1j, T2a, T7C, T7J;
1453 T1j = VADD(TL, T1i);
1454 T2a = VADD(T1E, T29);
1455 T2b = VADD(T1j, T2a);
1456 T7x = VSUB(T1j, T2a);
1457 T7C = VADD(T7e, T7f);
1458 T7J = VADD(T7D, T7I);
1459 T7K = VADD(T7C, T7J);
1460 T7M = VSUB(T7J, T7C);
1461 }
1462 {
1463 V T2S, T3z, T7y, T7z;
1464 T2S = VADD(T2y, T2R);
1465 T3z = VADD(T3h, T3y);
1466 T3A = VADD(T2S, T3z);
1467 T7L = VSUB(T3z, T2S);
1468 T7y = VADD(T7j, T7k);
1469 T7z = VADD(T7o, T7p);
1470 T7A = VSUB(T7y, T7z);
1471 T7B = VADD(T7y, T7z);
1472 }
1473 ST(&(ri[WS(rs, 16)]), VSUB(T2b, T3A), ms, &(ri[0]));
1474 ST(&(ii[WS(rs, 16)]), VSUB(T7K, T7B), ms, &(ii[0]));
1475 ST(&(ri[0]), VADD(T2b, T3A), ms, &(ri[0]));
1476 ST(&(ii[0]), VADD(T7B, T7K), ms, &(ii[0]));
1477 ST(&(ri[WS(rs, 24)]), VSUB(T7x, T7A), ms, &(ri[0]));
1478 ST(&(ii[WS(rs, 24)]), VSUB(T7M, T7L), ms, &(ii[0]));
1479 ST(&(ri[WS(rs, 8)]), VADD(T7x, T7A), ms, &(ri[0]));
1480 ST(&(ii[WS(rs, 8)]), VADD(T7L, T7M), ms, &(ii[0]));
1481 }
1482 {
1483 V T7h, T7t, T7Q, T7S, T7m, T7u, T7r, T7v;
1484 {
1485 V T7d, T7g, T7O, T7P;
1486 T7d = VSUB(TL, T1i);
1487 T7g = VSUB(T7e, T7f);
1488 T7h = VADD(T7d, T7g);
1489 T7t = VSUB(T7d, T7g);
1490 T7O = VSUB(T29, T1E);
1491 T7P = VSUB(T7I, T7D);
1492 T7Q = VADD(T7O, T7P);
1493 T7S = VSUB(T7P, T7O);
1494 }
1495 {
1496 V T7i, T7l, T7n, T7q;
1497 T7i = VSUB(T2y, T2R);
1498 T7l = VSUB(T7j, T7k);
1499 T7m = VADD(T7i, T7l);
1500 T7u = VSUB(T7l, T7i);
1501 T7n = VSUB(T3h, T3y);
1502 T7q = VSUB(T7o, T7p);
1503 T7r = VSUB(T7n, T7q);
1504 T7v = VADD(T7n, T7q);
1505 }
1506 {
1507 V T7s, T7N, T7w, T7R;
1508 T7s = VMUL(LDK(KP707106781), VADD(T7m, T7r));
1509 ST(&(ri[WS(rs, 20)]), VSUB(T7h, T7s), ms, &(ri[0]));
1510 ST(&(ri[WS(rs, 4)]), VADD(T7h, T7s), ms, &(ri[0]));
1511 T7N = VMUL(LDK(KP707106781), VADD(T7u, T7v));
1512 ST(&(ii[WS(rs, 4)]), VADD(T7N, T7Q), ms, &(ii[0]));
1513 ST(&(ii[WS(rs, 20)]), VSUB(T7Q, T7N), ms, &(ii[0]));
1514 T7w = VMUL(LDK(KP707106781), VSUB(T7u, T7v));
1515 ST(&(ri[WS(rs, 28)]), VSUB(T7t, T7w), ms, &(ri[0]));
1516 ST(&(ri[WS(rs, 12)]), VADD(T7t, T7w), ms, &(ri[0]));
1517 T7R = VMUL(LDK(KP707106781), VSUB(T7r, T7m));
1518 ST(&(ii[WS(rs, 12)]), VADD(T7R, T7S), ms, &(ii[0]));
1519 ST(&(ii[WS(rs, 28)]), VSUB(T7S, T7R), ms, &(ii[0]));
1520 }
1521 }
1522 {
1523 V T6j, T7X, T83, T6X, T6u, T7U, T77, T7b, T70, T82, T6G, T6U, T74, T7a, T6R;
1524 V T6V;
1525 {
1526 V T6o, T6t, T6A, T6F;
1527 T6j = VSUB(T6f, T6i);
1528 T7X = VADD(T7V, T7W);
1529 T83 = VSUB(T7W, T7V);
1530 T6X = VADD(T6f, T6i);
1531 T6o = VSUB(T6m, T6n);
1532 T6t = VADD(T6p, T6s);
1533 T6u = VMUL(LDK(KP707106781), VSUB(T6o, T6t));
1534 T7U = VMUL(LDK(KP707106781), VADD(T6o, T6t));
1535 {
1536 V T75, T76, T6Y, T6Z;
1537 T75 = VADD(T6H, T6K);
1538 T76 = VADD(T6O, T6P);
1539 T77 = VFNMS(LDK(KP382683432), T76, VMUL(LDK(KP923879532), T75));
1540 T7b = VFMA(LDK(KP923879532), T76, VMUL(LDK(KP382683432), T75));
1541 T6Y = VADD(T6n, T6m);
1542 T6Z = VSUB(T6p, T6s);
1543 T70 = VMUL(LDK(KP707106781), VADD(T6Y, T6Z));
1544 T82 = VMUL(LDK(KP707106781), VSUB(T6Z, T6Y));
1545 }
1546 T6A = VSUB(T6y, T6z);
1547 T6F = VSUB(T6B, T6E);
1548 T6G = VFMA(LDK(KP923879532), T6A, VMUL(LDK(KP382683432), T6F));
1549 T6U = VFNMS(LDK(KP923879532), T6F, VMUL(LDK(KP382683432), T6A));
1550 {
1551 V T72, T73, T6L, T6Q;
1552 T72 = VADD(T6y, T6z);
1553 T73 = VADD(T6B, T6E);
1554 T74 = VFMA(LDK(KP382683432), T72, VMUL(LDK(KP923879532), T73));
1555 T7a = VFNMS(LDK(KP382683432), T73, VMUL(LDK(KP923879532), T72));
1556 T6L = VSUB(T6H, T6K);
1557 T6Q = VSUB(T6O, T6P);
1558 T6R = VFNMS(LDK(KP923879532), T6Q, VMUL(LDK(KP382683432), T6L));
1559 T6V = VFMA(LDK(KP382683432), T6Q, VMUL(LDK(KP923879532), T6L));
1560 }
1561 }
1562 {
1563 V T6v, T6S, T81, T84;
1564 T6v = VADD(T6j, T6u);
1565 T6S = VADD(T6G, T6R);
1566 ST(&(ri[WS(rs, 22)]), VSUB(T6v, T6S), ms, &(ri[0]));
1567 ST(&(ri[WS(rs, 6)]), VADD(T6v, T6S), ms, &(ri[0]));
1568 T81 = VADD(T6U, T6V);
1569 T84 = VADD(T82, T83);
1570 ST(&(ii[WS(rs, 6)]), VADD(T81, T84), ms, &(ii[0]));
1571 ST(&(ii[WS(rs, 22)]), VSUB(T84, T81), ms, &(ii[0]));
1572 }
1573 {
1574 V T6T, T6W, T85, T86;
1575 T6T = VSUB(T6j, T6u);
1576 T6W = VSUB(T6U, T6V);
1577 ST(&(ri[WS(rs, 30)]), VSUB(T6T, T6W), ms, &(ri[0]));
1578 ST(&(ri[WS(rs, 14)]), VADD(T6T, T6W), ms, &(ri[0]));
1579 T85 = VSUB(T6R, T6G);
1580 T86 = VSUB(T83, T82);
1581 ST(&(ii[WS(rs, 14)]), VADD(T85, T86), ms, &(ii[0]));
1582 ST(&(ii[WS(rs, 30)]), VSUB(T86, T85), ms, &(ii[0]));
1583 }
1584 {
1585 V T71, T78, T7T, T7Y;
1586 T71 = VADD(T6X, T70);
1587 T78 = VADD(T74, T77);
1588 ST(&(ri[WS(rs, 18)]), VSUB(T71, T78), ms, &(ri[0]));
1589 ST(&(ri[WS(rs, 2)]), VADD(T71, T78), ms, &(ri[0]));
1590 T7T = VADD(T7a, T7b);
1591 T7Y = VADD(T7U, T7X);
1592 ST(&(ii[WS(rs, 2)]), VADD(T7T, T7Y), ms, &(ii[0]));
1593 ST(&(ii[WS(rs, 18)]), VSUB(T7Y, T7T), ms, &(ii[0]));
1594 }
1595 {
1596 V T79, T7c, T7Z, T80;
1597 T79 = VSUB(T6X, T70);
1598 T7c = VSUB(T7a, T7b);
1599 ST(&(ri[WS(rs, 26)]), VSUB(T79, T7c), ms, &(ri[0]));
1600 ST(&(ri[WS(rs, 10)]), VADD(T79, T7c), ms, &(ri[0]));
1601 T7Z = VSUB(T77, T74);
1602 T80 = VSUB(T7X, T7U);
1603 ST(&(ii[WS(rs, 10)]), VADD(T7Z, T80), ms, &(ii[0]));
1604 ST(&(ii[WS(rs, 26)]), VSUB(T80, T7Z), ms, &(ii[0]));
1605 }
1606 }
1607 {
1608 V T3R, T5d, T8r, T8x, T4e, T8o, T5n, T5r, T4G, T5a, T5g, T8w, T5k, T5q, T57;
1609 V T5b, T3Q, T8p;
1610 T3Q = VMUL(LDK(KP707106781), VSUB(T3K, T3P));
1611 T3R = VSUB(T3F, T3Q);
1612 T5d = VADD(T3F, T3Q);
1613 T8p = VMUL(LDK(KP707106781), VSUB(T5v, T5u));
1614 T8r = VADD(T8p, T8q);
1615 T8x = VSUB(T8q, T8p);
1616 {
1617 V T42, T4d, T5l, T5m;
1618 T42 = VFNMS(LDK(KP923879532), T41, VMUL(LDK(KP382683432), T3W));
1619 T4d = VFMA(LDK(KP382683432), T47, VMUL(LDK(KP923879532), T4c));
1620 T4e = VSUB(T42, T4d);
1621 T8o = VADD(T42, T4d);
1622 T5l = VADD(T4L, T4W);
1623 T5m = VADD(T52, T55);
1624 T5n = VFNMS(LDK(KP555570233), T5m, VMUL(LDK(KP831469612), T5l));
1625 T5r = VFMA(LDK(KP831469612), T5m, VMUL(LDK(KP555570233), T5l));
1626 }
1627 {
1628 V T4w, T4F, T5e, T5f;
1629 T4w = VSUB(T4k, T4v);
1630 T4F = VSUB(T4B, T4E);
1631 T4G = VFMA(LDK(KP980785280), T4w, VMUL(LDK(KP195090322), T4F));
1632 T5a = VFNMS(LDK(KP980785280), T4F, VMUL(LDK(KP195090322), T4w));
1633 T5e = VFMA(LDK(KP923879532), T3W, VMUL(LDK(KP382683432), T41));
1634 T5f = VFNMS(LDK(KP923879532), T47, VMUL(LDK(KP382683432), T4c));
1635 T5g = VADD(T5e, T5f);
1636 T8w = VSUB(T5f, T5e);
1637 }
1638 {
1639 V T5i, T5j, T4X, T56;
1640 T5i = VADD(T4k, T4v);
1641 T5j = VADD(T4B, T4E);
1642 T5k = VFMA(LDK(KP555570233), T5i, VMUL(LDK(KP831469612), T5j));
1643 T5q = VFNMS(LDK(KP555570233), T5j, VMUL(LDK(KP831469612), T5i));
1644 T4X = VSUB(T4L, T4W);
1645 T56 = VSUB(T52, T55);
1646 T57 = VFNMS(LDK(KP980785280), T56, VMUL(LDK(KP195090322), T4X));
1647 T5b = VFMA(LDK(KP195090322), T56, VMUL(LDK(KP980785280), T4X));
1648 }
1649 {
1650 V T4f, T58, T8v, T8y;
1651 T4f = VADD(T3R, T4e);
1652 T58 = VADD(T4G, T57);
1653 ST(&(ri[WS(rs, 23)]), VSUB(T4f, T58), ms, &(ri[WS(rs, 1)]));
1654 ST(&(ri[WS(rs, 7)]), VADD(T4f, T58), ms, &(ri[WS(rs, 1)]));
1655 T8v = VADD(T5a, T5b);
1656 T8y = VADD(T8w, T8x);
1657 ST(&(ii[WS(rs, 7)]), VADD(T8v, T8y), ms, &(ii[WS(rs, 1)]));
1658 ST(&(ii[WS(rs, 23)]), VSUB(T8y, T8v), ms, &(ii[WS(rs, 1)]));
1659 }
1660 {
1661 V T59, T5c, T8z, T8A;
1662 T59 = VSUB(T3R, T4e);
1663 T5c = VSUB(T5a, T5b);
1664 ST(&(ri[WS(rs, 31)]), VSUB(T59, T5c), ms, &(ri[WS(rs, 1)]));
1665 ST(&(ri[WS(rs, 15)]), VADD(T59, T5c), ms, &(ri[WS(rs, 1)]));
1666 T8z = VSUB(T57, T4G);
1667 T8A = VSUB(T8x, T8w);
1668 ST(&(ii[WS(rs, 15)]), VADD(T8z, T8A), ms, &(ii[WS(rs, 1)]));
1669 ST(&(ii[WS(rs, 31)]), VSUB(T8A, T8z), ms, &(ii[WS(rs, 1)]));
1670 }
1671 {
1672 V T5h, T5o, T8n, T8s;
1673 T5h = VADD(T5d, T5g);
1674 T5o = VADD(T5k, T5n);
1675 ST(&(ri[WS(rs, 19)]), VSUB(T5h, T5o), ms, &(ri[WS(rs, 1)]));
1676 ST(&(ri[WS(rs, 3)]), VADD(T5h, T5o), ms, &(ri[WS(rs, 1)]));
1677 T8n = VADD(T5q, T5r);
1678 T8s = VADD(T8o, T8r);
1679 ST(&(ii[WS(rs, 3)]), VADD(T8n, T8s), ms, &(ii[WS(rs, 1)]));
1680 ST(&(ii[WS(rs, 19)]), VSUB(T8s, T8n), ms, &(ii[WS(rs, 1)]));
1681 }
1682 {
1683 V T5p, T5s, T8t, T8u;
1684 T5p = VSUB(T5d, T5g);
1685 T5s = VSUB(T5q, T5r);
1686 ST(&(ri[WS(rs, 27)]), VSUB(T5p, T5s), ms, &(ri[WS(rs, 1)]));
1687 ST(&(ri[WS(rs, 11)]), VADD(T5p, T5s), ms, &(ri[WS(rs, 1)]));
1688 T8t = VSUB(T5n, T5k);
1689 T8u = VSUB(T8r, T8o);
1690 ST(&(ii[WS(rs, 11)]), VADD(T8t, T8u), ms, &(ii[WS(rs, 1)]));
1691 ST(&(ii[WS(rs, 27)]), VSUB(T8u, T8t), ms, &(ii[WS(rs, 1)]));
1692 }
1693 }
1694 {
1695 V T5x, T5Z, T8d, T8j, T5E, T88, T69, T6d, T5M, T5W, T62, T8i, T66, T6c, T5T;
1696 V T5X, T5w, T89;
1697 T5w = VMUL(LDK(KP707106781), VADD(T5u, T5v));
1698 T5x = VSUB(T5t, T5w);
1699 T5Z = VADD(T5t, T5w);
1700 T89 = VMUL(LDK(KP707106781), VADD(T3K, T3P));
1701 T8d = VADD(T89, T8c);
1702 T8j = VSUB(T8c, T89);
1703 {
1704 V T5A, T5D, T67, T68;
1705 T5A = VFNMS(LDK(KP382683432), T5z, VMUL(LDK(KP923879532), T5y));
1706 T5D = VFMA(LDK(KP923879532), T5B, VMUL(LDK(KP382683432), T5C));
1707 T5E = VSUB(T5A, T5D);
1708 T88 = VADD(T5A, T5D);
1709 T67 = VADD(T5N, T5O);
1710 T68 = VADD(T5Q, T5R);
1711 T69 = VFNMS(LDK(KP195090322), T68, VMUL(LDK(KP980785280), T67));
1712 T6d = VFMA(LDK(KP195090322), T67, VMUL(LDK(KP980785280), T68));
1713 }
1714 {
1715 V T5I, T5L, T60, T61;
1716 T5I = VSUB(T5G, T5H);
1717 T5L = VSUB(T5J, T5K);
1718 T5M = VFMA(LDK(KP555570233), T5I, VMUL(LDK(KP831469612), T5L));
1719 T5W = VFNMS(LDK(KP831469612), T5I, VMUL(LDK(KP555570233), T5L));
1720 T60 = VFMA(LDK(KP382683432), T5y, VMUL(LDK(KP923879532), T5z));
1721 T61 = VFNMS(LDK(KP382683432), T5B, VMUL(LDK(KP923879532), T5C));
1722 T62 = VADD(T60, T61);
1723 T8i = VSUB(T61, T60);
1724 }
1725 {
1726 V T64, T65, T5P, T5S;
1727 T64 = VADD(T5G, T5H);
1728 T65 = VADD(T5J, T5K);
1729 T66 = VFMA(LDK(KP980785280), T64, VMUL(LDK(KP195090322), T65));
1730 T6c = VFNMS(LDK(KP195090322), T64, VMUL(LDK(KP980785280), T65));
1731 T5P = VSUB(T5N, T5O);
1732 T5S = VSUB(T5Q, T5R);
1733 T5T = VFNMS(LDK(KP831469612), T5S, VMUL(LDK(KP555570233), T5P));
1734 T5X = VFMA(LDK(KP831469612), T5P, VMUL(LDK(KP555570233), T5S));
1735 }
1736 {
1737 V T5F, T5U, T8h, T8k;
1738 T5F = VADD(T5x, T5E);
1739 T5U = VADD(T5M, T5T);
1740 ST(&(ri[WS(rs, 21)]), VSUB(T5F, T5U), ms, &(ri[WS(rs, 1)]));
1741 ST(&(ri[WS(rs, 5)]), VADD(T5F, T5U), ms, &(ri[WS(rs, 1)]));
1742 T8h = VADD(T5W, T5X);
1743 T8k = VADD(T8i, T8j);
1744 ST(&(ii[WS(rs, 5)]), VADD(T8h, T8k), ms, &(ii[WS(rs, 1)]));
1745 ST(&(ii[WS(rs, 21)]), VSUB(T8k, T8h), ms, &(ii[WS(rs, 1)]));
1746 }
1747 {
1748 V T5V, T5Y, T8l, T8m;
1749 T5V = VSUB(T5x, T5E);
1750 T5Y = VSUB(T5W, T5X);
1751 ST(&(ri[WS(rs, 29)]), VSUB(T5V, T5Y), ms, &(ri[WS(rs, 1)]));
1752 ST(&(ri[WS(rs, 13)]), VADD(T5V, T5Y), ms, &(ri[WS(rs, 1)]));
1753 T8l = VSUB(T5T, T5M);
1754 T8m = VSUB(T8j, T8i);
1755 ST(&(ii[WS(rs, 13)]), VADD(T8l, T8m), ms, &(ii[WS(rs, 1)]));
1756 ST(&(ii[WS(rs, 29)]), VSUB(T8m, T8l), ms, &(ii[WS(rs, 1)]));
1757 }
1758 {
1759 V T63, T6a, T87, T8e;
1760 T63 = VADD(T5Z, T62);
1761 T6a = VADD(T66, T69);
1762 ST(&(ri[WS(rs, 17)]), VSUB(T63, T6a), ms, &(ri[WS(rs, 1)]));
1763 ST(&(ri[WS(rs, 1)]), VADD(T63, T6a), ms, &(ri[WS(rs, 1)]));
1764 T87 = VADD(T6c, T6d);
1765 T8e = VADD(T88, T8d);
1766 ST(&(ii[WS(rs, 1)]), VADD(T87, T8e), ms, &(ii[WS(rs, 1)]));
1767 ST(&(ii[WS(rs, 17)]), VSUB(T8e, T87), ms, &(ii[WS(rs, 1)]));
1768 }
1769 {
1770 V T6b, T6e, T8f, T8g;
1771 T6b = VSUB(T5Z, T62);
1772 T6e = VSUB(T6c, T6d);
1773 ST(&(ri[WS(rs, 25)]), VSUB(T6b, T6e), ms, &(ri[WS(rs, 1)]));
1774 ST(&(ri[WS(rs, 9)]), VADD(T6b, T6e), ms, &(ri[WS(rs, 1)]));
1775 T8f = VSUB(T69, T66);
1776 T8g = VSUB(T8d, T88);
1777 ST(&(ii[WS(rs, 9)]), VADD(T8f, T8g), ms, &(ii[WS(rs, 1)]));
1778 ST(&(ii[WS(rs, 25)]), VSUB(T8g, T8f), ms, &(ii[WS(rs, 1)]));
1779 }
1780 }
1781 }
1782 }
1783 }
1784 VLEAVE();
1785 }
1786
1787 static const tw_instr twinstr[] = {
1788 VTW(0, 1),
1789 VTW(0, 3),
1790 VTW(0, 9),
1791 VTW(0, 27),
1792 {TW_NEXT, (2 * VL), 0}
1793 };
1794
1795 static const ct_desc desc = { 32, XSIMD_STRING("t2sv_32"), twinstr, &GENUS, {376, 168, 112, 0}, 0, 0, 0 };
1796
1797 void XSIMD(codelet_t2sv_32) (planner *p) {
1798 X(kdft_dit_register) (p, t2sv_32, &desc);
1799 }
1800 #endif /* HAVE_FMA */