Mercurial > hg > sv-dependency-builds
comparison src/fftw-3.3.3/dft/simd/common/n2fv_64.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 07:37:25 EST 2012 */ | |
23 | |
24 #include "codelet-dft.h" | |
25 | |
26 #ifdef HAVE_FMA | |
27 | |
28 /* Generated by: ../../../genfft/gen_notw_c.native -fma -reorder-insns -schedule-for-pipeline -simd -compact -variables 4 -pipeline-latency 8 -n 64 -name n2fv_64 -with-ostride 2 -include n2f.h -store-multiple 2 */ | |
29 | |
30 /* | |
31 * This function contains 456 FP additions, 258 FP multiplications, | |
32 * (or, 198 additions, 0 multiplications, 258 fused multiply/add), | |
33 * 178 stack variables, 15 constants, and 160 memory accesses | |
34 */ | |
35 #include "n2f.h" | |
36 | |
37 static void n2fv_64(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs) | |
38 { | |
39 DVK(KP773010453, +0.773010453362736960810906609758469800971041293); | |
40 DVK(KP820678790, +0.820678790828660330972281985331011598767386482); | |
41 DVK(KP881921264, +0.881921264348355029712756863660388349508442621); | |
42 DVK(KP534511135, +0.534511135950791641089685961295362908582039528); | |
43 DVK(KP995184726, +0.995184726672196886244836953109479921575474869); | |
44 DVK(KP098491403, +0.098491403357164253077197521291327432293052451); | |
45 DVK(KP980785280, +0.980785280403230449126182236134239036973933731); | |
46 DVK(KP956940335, +0.956940335732208864935797886980269969482849206); | |
47 DVK(KP303346683, +0.303346683607342391675883946941299872384187453); | |
48 DVK(KP831469612, +0.831469612302545237078788377617905756738560812); | |
49 DVK(KP923879532, +0.923879532511286756128183189396788286822416626); | |
50 DVK(KP668178637, +0.668178637919298919997757686523080761552472251); | |
51 DVK(KP198912367, +0.198912367379658006911597622644676228597850501); | |
52 DVK(KP414213562, +0.414213562373095048801688724209698078569671875); | |
53 DVK(KP707106781, +0.707106781186547524400844362104849039284835938); | |
54 { | |
55 INT i; | |
56 const R *xi; | |
57 R *xo; | |
58 xi = ri; | |
59 xo = ro; | |
60 for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(128, is), MAKE_VOLATILE_STRIDE(128, os)) { | |
61 V T7r, T7s, T7t, T7u, T5T, T5S, T5X, T65, T8a, T8b, T8e, T8g, T5Z, T5R, T67; | |
62 V T63, T5U, T64; | |
63 { | |
64 V T7, T26, T5k, T6A, T47, T69, T2V, T3z, T6B, T4e, T6a, T5n, T3M, T2Y, T27; | |
65 V Tm, T3A, T3l, T2a, TC, T5p, T4o, T6E, T6e, T3i, T3B, TR, T29, T4x, T5q; | |
66 V T6h, T6D, T39, T3H, T3I, T3c, T5N, T57, T72, T6w, T5O, T5e, T71, T6t, T2y; | |
67 V T1W, T2x, T1N, T33, T34, T3E, T32, T1p, T2v, T1g, T2u, T4M, T5K, T6p, T6Z; | |
68 V T6m, T6Y, T5L, T4T; | |
69 { | |
70 V T4g, T4l, T3j, Tu, Tx, T4h, TA, T4i; | |
71 { | |
72 V T1, T2, T23, T24, T4, T5, T20, T21; | |
73 T1 = LD(&(xi[0]), ivs, &(xi[0])); | |
74 T2 = LD(&(xi[WS(is, 32)]), ivs, &(xi[0])); | |
75 T23 = LD(&(xi[WS(is, 8)]), ivs, &(xi[0])); | |
76 T24 = LD(&(xi[WS(is, 40)]), ivs, &(xi[0])); | |
77 T4 = LD(&(xi[WS(is, 16)]), ivs, &(xi[0])); | |
78 T5 = LD(&(xi[WS(is, 48)]), ivs, &(xi[0])); | |
79 T20 = LD(&(xi[WS(is, 56)]), ivs, &(xi[0])); | |
80 T21 = LD(&(xi[WS(is, 24)]), ivs, &(xi[0])); | |
81 { | |
82 V Ta, T48, Tk, T4c, T49, Td, Tf, Tg; | |
83 { | |
84 V T8, T43, T3, T44, T25, T5i, T6, T45, T22, T9, Ti, Tj, Tb, Tc; | |
85 T8 = LD(&(xi[WS(is, 4)]), ivs, &(xi[0])); | |
86 T43 = VSUB(T1, T2); | |
87 T3 = VADD(T1, T2); | |
88 T44 = VSUB(T23, T24); | |
89 T25 = VADD(T23, T24); | |
90 T5i = VSUB(T4, T5); | |
91 T6 = VADD(T4, T5); | |
92 T45 = VSUB(T20, T21); | |
93 T22 = VADD(T20, T21); | |
94 T9 = LD(&(xi[WS(is, 36)]), ivs, &(xi[0])); | |
95 Ti = LD(&(xi[WS(is, 12)]), ivs, &(xi[0])); | |
96 Tj = LD(&(xi[WS(is, 44)]), ivs, &(xi[0])); | |
97 Tb = LD(&(xi[WS(is, 20)]), ivs, &(xi[0])); | |
98 Tc = LD(&(xi[WS(is, 52)]), ivs, &(xi[0])); | |
99 { | |
100 V T2T, T46, T5j, T2U; | |
101 T7 = VSUB(T3, T6); | |
102 T2T = VADD(T3, T6); | |
103 T46 = VADD(T44, T45); | |
104 T5j = VSUB(T45, T44); | |
105 T26 = VSUB(T22, T25); | |
106 T2U = VADD(T25, T22); | |
107 Ta = VADD(T8, T9); | |
108 T48 = VSUB(T8, T9); | |
109 Tk = VADD(Ti, Tj); | |
110 T4c = VSUB(Tj, Ti); | |
111 T5k = VFNMS(LDK(KP707106781), T5j, T5i); | |
112 T6A = VFMA(LDK(KP707106781), T5j, T5i); | |
113 T47 = VFMA(LDK(KP707106781), T46, T43); | |
114 T69 = VFNMS(LDK(KP707106781), T46, T43); | |
115 T2V = VADD(T2T, T2U); | |
116 T3z = VSUB(T2T, T2U); | |
117 T49 = VSUB(Tb, Tc); | |
118 Td = VADD(Tb, Tc); | |
119 } | |
120 Tf = LD(&(xi[WS(is, 60)]), ivs, &(xi[0])); | |
121 Tg = LD(&(xi[WS(is, 28)]), ivs, &(xi[0])); | |
122 } | |
123 { | |
124 V Te, T2W, T5l, T4a, Tq, Tt, Tv, Tw, T5m, T4d, Tl, T2X, Ty, Tz, To; | |
125 V Tp; | |
126 To = LD(&(xi[WS(is, 2)]), ivs, &(xi[0])); | |
127 Tp = LD(&(xi[WS(is, 34)]), ivs, &(xi[0])); | |
128 { | |
129 V Th, T4b, Tr, Ts; | |
130 Tr = LD(&(xi[WS(is, 18)]), ivs, &(xi[0])); | |
131 Ts = LD(&(xi[WS(is, 50)]), ivs, &(xi[0])); | |
132 Te = VSUB(Ta, Td); | |
133 T2W = VADD(Ta, Td); | |
134 T5l = VFMA(LDK(KP414213562), T48, T49); | |
135 T4a = VFNMS(LDK(KP414213562), T49, T48); | |
136 Th = VADD(Tf, Tg); | |
137 T4b = VSUB(Tf, Tg); | |
138 Tq = VADD(To, Tp); | |
139 T4g = VSUB(To, Tp); | |
140 T4l = VSUB(Tr, Ts); | |
141 Tt = VADD(Tr, Ts); | |
142 Tv = LD(&(xi[WS(is, 10)]), ivs, &(xi[0])); | |
143 Tw = LD(&(xi[WS(is, 42)]), ivs, &(xi[0])); | |
144 T5m = VFMA(LDK(KP414213562), T4b, T4c); | |
145 T4d = VFNMS(LDK(KP414213562), T4c, T4b); | |
146 Tl = VSUB(Th, Tk); | |
147 T2X = VADD(Th, Tk); | |
148 Ty = LD(&(xi[WS(is, 58)]), ivs, &(xi[0])); | |
149 Tz = LD(&(xi[WS(is, 26)]), ivs, &(xi[0])); | |
150 } | |
151 T3j = VADD(Tq, Tt); | |
152 Tu = VSUB(Tq, Tt); | |
153 Tx = VADD(Tv, Tw); | |
154 T4h = VSUB(Tv, Tw); | |
155 T6B = VSUB(T4d, T4a); | |
156 T4e = VADD(T4a, T4d); | |
157 T6a = VADD(T5l, T5m); | |
158 T5n = VSUB(T5l, T5m); | |
159 T3M = VSUB(T2X, T2W); | |
160 T2Y = VADD(T2W, T2X); | |
161 T27 = VSUB(Tl, Te); | |
162 Tm = VADD(Te, Tl); | |
163 TA = VADD(Ty, Tz); | |
164 T4i = VSUB(Ty, Tz); | |
165 } | |
166 } | |
167 } | |
168 { | |
169 V TK, T4p, T4u, T4k, T6d, T4n, T6c, TL, TN, TO, T3g, TJ, TF, TI; | |
170 { | |
171 V TD, TE, TG, TH; | |
172 TD = LD(&(xi[WS(is, 62)]), ivs, &(xi[0])); | |
173 TE = LD(&(xi[WS(is, 30)]), ivs, &(xi[0])); | |
174 TG = LD(&(xi[WS(is, 14)]), ivs, &(xi[0])); | |
175 TH = LD(&(xi[WS(is, 46)]), ivs, &(xi[0])); | |
176 TK = LD(&(xi[WS(is, 54)]), ivs, &(xi[0])); | |
177 { | |
178 V T3k, TB, T4j, T4m; | |
179 T3k = VADD(Tx, TA); | |
180 TB = VSUB(Tx, TA); | |
181 T4j = VADD(T4h, T4i); | |
182 T4m = VSUB(T4h, T4i); | |
183 T4p = VSUB(TD, TE); | |
184 TF = VADD(TD, TE); | |
185 T4u = VSUB(TH, TG); | |
186 TI = VADD(TG, TH); | |
187 T3A = VSUB(T3j, T3k); | |
188 T3l = VADD(T3j, T3k); | |
189 T2a = VFMA(LDK(KP414213562), Tu, TB); | |
190 TC = VFNMS(LDK(KP414213562), TB, Tu); | |
191 T4k = VFMA(LDK(KP707106781), T4j, T4g); | |
192 T6d = VFNMS(LDK(KP707106781), T4j, T4g); | |
193 T4n = VFMA(LDK(KP707106781), T4m, T4l); | |
194 T6c = VFNMS(LDK(KP707106781), T4m, T4l); | |
195 TL = LD(&(xi[WS(is, 22)]), ivs, &(xi[0])); | |
196 } | |
197 TN = LD(&(xi[WS(is, 6)]), ivs, &(xi[0])); | |
198 TO = LD(&(xi[WS(is, 38)]), ivs, &(xi[0])); | |
199 } | |
200 T3g = VADD(TF, TI); | |
201 TJ = VSUB(TF, TI); | |
202 { | |
203 V T3a, T1E, T52, T5b, T1x, T4Z, T6r, T6u, T5a, T1U, T55, T5c, T1L, T3b; | |
204 { | |
205 V T4V, T1t, T58, T1w, T1Q, T1T, T1I, T4Y, T59, T1J, T53, T1H; | |
206 { | |
207 V T1r, TM, T4r, TP, T4q, T1s, T1u, T1v; | |
208 T1r = LD(&(xi[WS(is, 63)]), ivs, &(xi[WS(is, 1)])); | |
209 T5p = VFMA(LDK(KP198912367), T4k, T4n); | |
210 T4o = VFNMS(LDK(KP198912367), T4n, T4k); | |
211 T6E = VFMA(LDK(KP668178637), T6c, T6d); | |
212 T6e = VFNMS(LDK(KP668178637), T6d, T6c); | |
213 TM = VADD(TK, TL); | |
214 T4r = VSUB(TK, TL); | |
215 TP = VADD(TN, TO); | |
216 T4q = VSUB(TN, TO); | |
217 T1s = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)])); | |
218 T1u = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)])); | |
219 T1v = LD(&(xi[WS(is, 47)]), ivs, &(xi[WS(is, 1)])); | |
220 { | |
221 V T1R, T4X, T6g, T4t, T6f, T4w, T1S, T1O, T1P; | |
222 T1O = LD(&(xi[WS(is, 55)]), ivs, &(xi[WS(is, 1)])); | |
223 T1P = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)])); | |
224 T1R = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)])); | |
225 { | |
226 V T3h, TQ, T4s, T4v; | |
227 T3h = VADD(TP, TM); | |
228 TQ = VSUB(TM, TP); | |
229 T4s = VADD(T4q, T4r); | |
230 T4v = VSUB(T4r, T4q); | |
231 T4V = VSUB(T1r, T1s); | |
232 T1t = VADD(T1r, T1s); | |
233 T58 = VSUB(T1v, T1u); | |
234 T1w = VADD(T1u, T1v); | |
235 T4X = VSUB(T1O, T1P); | |
236 T1Q = VADD(T1O, T1P); | |
237 T3i = VADD(T3g, T3h); | |
238 T3B = VSUB(T3g, T3h); | |
239 TR = VFNMS(LDK(KP414213562), TQ, TJ); | |
240 T29 = VFMA(LDK(KP414213562), TJ, TQ); | |
241 T6g = VFNMS(LDK(KP707106781), T4s, T4p); | |
242 T4t = VFMA(LDK(KP707106781), T4s, T4p); | |
243 T6f = VFNMS(LDK(KP707106781), T4v, T4u); | |
244 T4w = VFMA(LDK(KP707106781), T4v, T4u); | |
245 T1S = LD(&(xi[WS(is, 39)]), ivs, &(xi[WS(is, 1)])); | |
246 } | |
247 { | |
248 V T4W, T1A, T50, T51, T1D, T1F, T1G; | |
249 { | |
250 V T1y, T1z, T1B, T1C; | |
251 T1y = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)])); | |
252 T1z = LD(&(xi[WS(is, 35)]), ivs, &(xi[WS(is, 1)])); | |
253 T1B = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)])); | |
254 T1C = LD(&(xi[WS(is, 51)]), ivs, &(xi[WS(is, 1)])); | |
255 T4x = VFNMS(LDK(KP198912367), T4w, T4t); | |
256 T5q = VFMA(LDK(KP198912367), T4t, T4w); | |
257 T6h = VFNMS(LDK(KP668178637), T6g, T6f); | |
258 T6D = VFMA(LDK(KP668178637), T6f, T6g); | |
259 T4W = VSUB(T1R, T1S); | |
260 T1T = VADD(T1R, T1S); | |
261 T1A = VADD(T1y, T1z); | |
262 T50 = VSUB(T1y, T1z); | |
263 T51 = VSUB(T1C, T1B); | |
264 T1D = VADD(T1B, T1C); | |
265 } | |
266 T1F = LD(&(xi[WS(is, 59)]), ivs, &(xi[WS(is, 1)])); | |
267 T1G = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)])); | |
268 T1I = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)])); | |
269 T4Y = VADD(T4W, T4X); | |
270 T59 = VSUB(T4X, T4W); | |
271 T1J = LD(&(xi[WS(is, 43)]), ivs, &(xi[WS(is, 1)])); | |
272 T3a = VADD(T1A, T1D); | |
273 T1E = VSUB(T1A, T1D); | |
274 T52 = VFMA(LDK(KP414213562), T51, T50); | |
275 T5b = VFNMS(LDK(KP414213562), T50, T51); | |
276 T53 = VSUB(T1F, T1G); | |
277 T1H = VADD(T1F, T1G); | |
278 } | |
279 } | |
280 } | |
281 { | |
282 V T37, T54, T1K, T38; | |
283 T1x = VSUB(T1t, T1w); | |
284 T37 = VADD(T1t, T1w); | |
285 T4Z = VFMA(LDK(KP707106781), T4Y, T4V); | |
286 T6r = VFNMS(LDK(KP707106781), T4Y, T4V); | |
287 T54 = VSUB(T1J, T1I); | |
288 T1K = VADD(T1I, T1J); | |
289 T6u = VFNMS(LDK(KP707106781), T59, T58); | |
290 T5a = VFMA(LDK(KP707106781), T59, T58); | |
291 T38 = VADD(T1T, T1Q); | |
292 T1U = VSUB(T1Q, T1T); | |
293 T55 = VFNMS(LDK(KP414213562), T54, T53); | |
294 T5c = VFMA(LDK(KP414213562), T53, T54); | |
295 T1L = VSUB(T1H, T1K); | |
296 T3b = VADD(T1H, T1K); | |
297 T39 = VADD(T37, T38); | |
298 T3H = VSUB(T37, T38); | |
299 } | |
300 } | |
301 { | |
302 V T4A, TW, T4N, TZ, T1j, T1m, T4O, T4D, T13, T4F, T16, T4G, T1a, T4I, T4J; | |
303 V T1d; | |
304 { | |
305 V TU, TV, TX, TY, T56, T6v; | |
306 TU = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)])); | |
307 T56 = VADD(T52, T55); | |
308 T6v = VSUB(T55, T52); | |
309 { | |
310 V T5d, T6s, T1V, T1M; | |
311 T5d = VADD(T5b, T5c); | |
312 T6s = VSUB(T5c, T5b); | |
313 T1V = VSUB(T1L, T1E); | |
314 T1M = VADD(T1E, T1L); | |
315 T3I = VSUB(T3b, T3a); | |
316 T3c = VADD(T3a, T3b); | |
317 T5N = VFNMS(LDK(KP923879532), T56, T4Z); | |
318 T57 = VFMA(LDK(KP923879532), T56, T4Z); | |
319 T72 = VFNMS(LDK(KP923879532), T6v, T6u); | |
320 T6w = VFMA(LDK(KP923879532), T6v, T6u); | |
321 T5O = VFNMS(LDK(KP923879532), T5d, T5a); | |
322 T5e = VFMA(LDK(KP923879532), T5d, T5a); | |
323 T71 = VFMA(LDK(KP923879532), T6s, T6r); | |
324 T6t = VFNMS(LDK(KP923879532), T6s, T6r); | |
325 T2y = VFNMS(LDK(KP707106781), T1V, T1U); | |
326 T1W = VFMA(LDK(KP707106781), T1V, T1U); | |
327 T2x = VFNMS(LDK(KP707106781), T1M, T1x); | |
328 T1N = VFMA(LDK(KP707106781), T1M, T1x); | |
329 TV = LD(&(xi[WS(is, 33)]), ivs, &(xi[WS(is, 1)])); | |
330 } | |
331 TX = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)])); | |
332 TY = LD(&(xi[WS(is, 49)]), ivs, &(xi[WS(is, 1)])); | |
333 { | |
334 V T1h, T1i, T1k, T1l; | |
335 T1h = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)])); | |
336 T1i = LD(&(xi[WS(is, 41)]), ivs, &(xi[WS(is, 1)])); | |
337 T1k = LD(&(xi[WS(is, 57)]), ivs, &(xi[WS(is, 1)])); | |
338 T1l = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)])); | |
339 { | |
340 V T11, T4B, T4C, T12, T14, T15; | |
341 T11 = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)])); | |
342 T4A = VSUB(TU, TV); | |
343 TW = VADD(TU, TV); | |
344 T4N = VSUB(TX, TY); | |
345 TZ = VADD(TX, TY); | |
346 T1j = VADD(T1h, T1i); | |
347 T4B = VSUB(T1h, T1i); | |
348 T1m = VADD(T1k, T1l); | |
349 T4C = VSUB(T1k, T1l); | |
350 T12 = LD(&(xi[WS(is, 37)]), ivs, &(xi[WS(is, 1)])); | |
351 T14 = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)])); | |
352 T15 = LD(&(xi[WS(is, 53)]), ivs, &(xi[WS(is, 1)])); | |
353 { | |
354 V T18, T19, T1b, T1c; | |
355 T18 = LD(&(xi[WS(is, 61)]), ivs, &(xi[WS(is, 1)])); | |
356 T19 = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)])); | |
357 T1b = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)])); | |
358 T1c = LD(&(xi[WS(is, 45)]), ivs, &(xi[WS(is, 1)])); | |
359 T4O = VSUB(T4B, T4C); | |
360 T4D = VADD(T4B, T4C); | |
361 T13 = VADD(T11, T12); | |
362 T4F = VSUB(T11, T12); | |
363 T16 = VADD(T14, T15); | |
364 T4G = VSUB(T14, T15); | |
365 T1a = VADD(T18, T19); | |
366 T4I = VSUB(T18, T19); | |
367 T4J = VSUB(T1b, T1c); | |
368 T1d = VADD(T1b, T1c); | |
369 } | |
370 } | |
371 } | |
372 } | |
373 { | |
374 V T30, T10, T6k, T4E, T4Q, T4H, T17, T6n, T4P, T1e, T4K, T4R, T1n, T31; | |
375 T30 = VADD(TW, TZ); | |
376 T10 = VSUB(TW, TZ); | |
377 T6k = VFNMS(LDK(KP707106781), T4D, T4A); | |
378 T4E = VFMA(LDK(KP707106781), T4D, T4A); | |
379 T4Q = VFMA(LDK(KP414213562), T4F, T4G); | |
380 T4H = VFNMS(LDK(KP414213562), T4G, T4F); | |
381 T33 = VADD(T13, T16); | |
382 T17 = VSUB(T13, T16); | |
383 T6n = VFNMS(LDK(KP707106781), T4O, T4N); | |
384 T4P = VFMA(LDK(KP707106781), T4O, T4N); | |
385 T34 = VADD(T1a, T1d); | |
386 T1e = VSUB(T1a, T1d); | |
387 T4K = VFMA(LDK(KP414213562), T4J, T4I); | |
388 T4R = VFNMS(LDK(KP414213562), T4I, T4J); | |
389 T1n = VSUB(T1j, T1m); | |
390 T31 = VADD(T1j, T1m); | |
391 { | |
392 V T1f, T1o, T6o, T4L, T4S, T6l; | |
393 T1f = VADD(T17, T1e); | |
394 T1o = VSUB(T17, T1e); | |
395 T6o = VSUB(T4H, T4K); | |
396 T4L = VADD(T4H, T4K); | |
397 T4S = VADD(T4Q, T4R); | |
398 T6l = VSUB(T4Q, T4R); | |
399 T3E = VSUB(T30, T31); | |
400 T32 = VADD(T30, T31); | |
401 T1p = VFMA(LDK(KP707106781), T1o, T1n); | |
402 T2v = VFNMS(LDK(KP707106781), T1o, T1n); | |
403 T1g = VFMA(LDK(KP707106781), T1f, T10); | |
404 T2u = VFNMS(LDK(KP707106781), T1f, T10); | |
405 T4M = VFMA(LDK(KP923879532), T4L, T4E); | |
406 T5K = VFNMS(LDK(KP923879532), T4L, T4E); | |
407 T6p = VFMA(LDK(KP923879532), T6o, T6n); | |
408 T6Z = VFNMS(LDK(KP923879532), T6o, T6n); | |
409 T6m = VFNMS(LDK(KP923879532), T6l, T6k); | |
410 T6Y = VFMA(LDK(KP923879532), T6l, T6k); | |
411 T5L = VFNMS(LDK(KP923879532), T4S, T4P); | |
412 T4T = VFMA(LDK(KP923879532), T4S, T4P); | |
413 } | |
414 } | |
415 } | |
416 } | |
417 } | |
418 } | |
419 { | |
420 V T6b, T6F, T7n, T7o, T7p, T7q, T7v, T7w, T7x, T7y, T7z, T7A, T7B, T7C, T7f; | |
421 V T6X, T70, T79, T7a, T73, T6C, T76, T77, T6i; | |
422 { | |
423 V T2Z, T3r, T3s, T3m, T3d, T3v; | |
424 T2Z = VSUB(T2V, T2Y); | |
425 T3r = VADD(T2V, T2Y); | |
426 T3s = VADD(T3l, T3i); | |
427 T3m = VSUB(T3i, T3l); | |
428 T3d = VSUB(T39, T3c); | |
429 T3v = VADD(T39, T3c); | |
430 { | |
431 V T3x, T3t, T3P, T3J, T3D, T3V, T3Q, T3G, T36, T3u, T3Y, T3O, T6V, T6W; | |
432 { | |
433 V T3N, T3C, T3F, T35; | |
434 T3N = VSUB(T3B, T3A); | |
435 T3C = VADD(T3A, T3B); | |
436 T3F = VSUB(T33, T34); | |
437 T35 = VADD(T33, T34); | |
438 T3x = VSUB(T3r, T3s); | |
439 T3t = VADD(T3r, T3s); | |
440 T3P = VFMA(LDK(KP414213562), T3H, T3I); | |
441 T3J = VFNMS(LDK(KP414213562), T3I, T3H); | |
442 T3D = VFMA(LDK(KP707106781), T3C, T3z); | |
443 T3V = VFNMS(LDK(KP707106781), T3C, T3z); | |
444 T3Q = VFMA(LDK(KP414213562), T3E, T3F); | |
445 T3G = VFNMS(LDK(KP414213562), T3F, T3E); | |
446 T36 = VSUB(T32, T35); | |
447 T3u = VADD(T32, T35); | |
448 T3Y = VFNMS(LDK(KP707106781), T3N, T3M); | |
449 T3O = VFMA(LDK(KP707106781), T3N, T3M); | |
450 } | |
451 T6b = VFNMS(LDK(KP923879532), T6a, T69); | |
452 T6V = VFMA(LDK(KP923879532), T6a, T69); | |
453 T6W = VADD(T6E, T6D); | |
454 T6F = VSUB(T6D, T6E); | |
455 { | |
456 V T3K, T3Z, T3e, T3n; | |
457 T3K = VADD(T3G, T3J); | |
458 T3Z = VSUB(T3J, T3G); | |
459 T3e = VADD(T36, T3d); | |
460 T3n = VSUB(T3d, T36); | |
461 { | |
462 V T3w, T3y, T3R, T3W; | |
463 T3w = VADD(T3u, T3v); | |
464 T3y = VSUB(T3v, T3u); | |
465 T3R = VSUB(T3P, T3Q); | |
466 T3W = VADD(T3Q, T3P); | |
467 { | |
468 V T42, T40, T3L, T3T; | |
469 T42 = VFNMS(LDK(KP923879532), T3Z, T3Y); | |
470 T40 = VFMA(LDK(KP923879532), T3Z, T3Y); | |
471 T3L = VFNMS(LDK(KP923879532), T3K, T3D); | |
472 T3T = VFMA(LDK(KP923879532), T3K, T3D); | |
473 { | |
474 V T3o, T3q, T3f, T3p; | |
475 T3o = VFNMS(LDK(KP707106781), T3n, T3m); | |
476 T3q = VFMA(LDK(KP707106781), T3n, T3m); | |
477 T3f = VFNMS(LDK(KP707106781), T3e, T2Z); | |
478 T3p = VFMA(LDK(KP707106781), T3e, T2Z); | |
479 T7n = VFNMSI(T3y, T3x); | |
480 STM2(&(xo[96]), T7n, ovs, &(xo[0])); | |
481 T7o = VFMAI(T3y, T3x); | |
482 STM2(&(xo[32]), T7o, ovs, &(xo[0])); | |
483 T7p = VADD(T3t, T3w); | |
484 STM2(&(xo[0]), T7p, ovs, &(xo[0])); | |
485 T7q = VSUB(T3t, T3w); | |
486 STM2(&(xo[64]), T7q, ovs, &(xo[0])); | |
487 { | |
488 V T41, T3X, T3S, T3U; | |
489 T41 = VFMA(LDK(KP923879532), T3W, T3V); | |
490 T3X = VFNMS(LDK(KP923879532), T3W, T3V); | |
491 T3S = VFNMS(LDK(KP923879532), T3R, T3O); | |
492 T3U = VFMA(LDK(KP923879532), T3R, T3O); | |
493 T7r = VFMAI(T3q, T3p); | |
494 STM2(&(xo[16]), T7r, ovs, &(xo[0])); | |
495 T7s = VFNMSI(T3q, T3p); | |
496 STM2(&(xo[112]), T7s, ovs, &(xo[0])); | |
497 T7t = VFMAI(T3o, T3f); | |
498 STM2(&(xo[80]), T7t, ovs, &(xo[0])); | |
499 T7u = VFNMSI(T3o, T3f); | |
500 STM2(&(xo[48]), T7u, ovs, &(xo[0])); | |
501 T7v = VFNMSI(T40, T3X); | |
502 STM2(&(xo[88]), T7v, ovs, &(xo[0])); | |
503 T7w = VFMAI(T40, T3X); | |
504 STM2(&(xo[40]), T7w, ovs, &(xo[0])); | |
505 T7x = VFMAI(T42, T41); | |
506 STM2(&(xo[104]), T7x, ovs, &(xo[0])); | |
507 T7y = VFNMSI(T42, T41); | |
508 STM2(&(xo[24]), T7y, ovs, &(xo[0])); | |
509 T7z = VFMAI(T3U, T3T); | |
510 STM2(&(xo[8]), T7z, ovs, &(xo[0])); | |
511 T7A = VFNMSI(T3U, T3T); | |
512 STM2(&(xo[120]), T7A, ovs, &(xo[0])); | |
513 T7B = VFMAI(T3S, T3L); | |
514 STM2(&(xo[72]), T7B, ovs, &(xo[0])); | |
515 T7C = VFNMSI(T3S, T3L); | |
516 STM2(&(xo[56]), T7C, ovs, &(xo[0])); | |
517 T7f = VFNMS(LDK(KP831469612), T6W, T6V); | |
518 T6X = VFMA(LDK(KP831469612), T6W, T6V); | |
519 } | |
520 } | |
521 } | |
522 } | |
523 } | |
524 T70 = VFMA(LDK(KP303346683), T6Z, T6Y); | |
525 T79 = VFNMS(LDK(KP303346683), T6Y, T6Z); | |
526 T7a = VFNMS(LDK(KP303346683), T71, T72); | |
527 T73 = VFMA(LDK(KP303346683), T72, T71); | |
528 T6C = VFNMS(LDK(KP923879532), T6B, T6A); | |
529 T76 = VFMA(LDK(KP923879532), T6B, T6A); | |
530 T77 = VSUB(T6e, T6h); | |
531 T6i = VADD(T6e, T6h); | |
532 } | |
533 } | |
534 { | |
535 V T2r, T2D, T2C, T2s, T5H, T5o, T5v, T5D, T7L, T7O, T7Q, T7S, T5r, T5I, T5x; | |
536 V T5h, T5F, T5B; | |
537 { | |
538 V TT, T2f, T7E, T7F, T7H, T7J, T2n, T1Y, T28, T2b, T2l, T2p, T2j, T2k; | |
539 { | |
540 V T1X, T2d, T7h, T7l, T2e, T1q, T75, T7d, T7m, T7k, T7c, T7e, Tn, TS; | |
541 T2r = VFNMS(LDK(KP707106781), Tm, T7); | |
542 Tn = VFMA(LDK(KP707106781), Tm, T7); | |
543 TS = VADD(TC, TR); | |
544 T2D = VSUB(TR, TC); | |
545 { | |
546 V T7b, T7j, T74, T7i, T78, T7g; | |
547 T1X = VFNMS(LDK(KP198912367), T1W, T1N); | |
548 T2d = VFMA(LDK(KP198912367), T1N, T1W); | |
549 T7g = VADD(T79, T7a); | |
550 T7b = VSUB(T79, T7a); | |
551 T7j = VSUB(T73, T70); | |
552 T74 = VADD(T70, T73); | |
553 T7i = VFNMS(LDK(KP831469612), T77, T76); | |
554 T78 = VFMA(LDK(KP831469612), T77, T76); | |
555 T2j = VFNMS(LDK(KP923879532), TS, Tn); | |
556 TT = VFMA(LDK(KP923879532), TS, Tn); | |
557 T7h = VFMA(LDK(KP956940335), T7g, T7f); | |
558 T7l = VFNMS(LDK(KP956940335), T7g, T7f); | |
559 T2e = VFMA(LDK(KP198912367), T1g, T1p); | |
560 T1q = VFNMS(LDK(KP198912367), T1p, T1g); | |
561 T75 = VFNMS(LDK(KP956940335), T74, T6X); | |
562 T7d = VFMA(LDK(KP956940335), T74, T6X); | |
563 T7m = VFNMS(LDK(KP956940335), T7j, T7i); | |
564 T7k = VFMA(LDK(KP956940335), T7j, T7i); | |
565 T7c = VFNMS(LDK(KP956940335), T7b, T78); | |
566 T7e = VFMA(LDK(KP956940335), T7b, T78); | |
567 } | |
568 T2k = VADD(T2e, T2d); | |
569 T2f = VSUB(T2d, T2e); | |
570 { | |
571 V T7D, T7G, T7I, T7K; | |
572 T7D = VFNMSI(T7k, T7h); | |
573 STM2(&(xo[90]), T7D, ovs, &(xo[2])); | |
574 STN2(&(xo[88]), T7v, T7D, ovs); | |
575 T7E = VFMAI(T7k, T7h); | |
576 STM2(&(xo[38]), T7E, ovs, &(xo[2])); | |
577 T7F = VFMAI(T7m, T7l); | |
578 STM2(&(xo[102]), T7F, ovs, &(xo[2])); | |
579 T7G = VFNMSI(T7m, T7l); | |
580 STM2(&(xo[26]), T7G, ovs, &(xo[2])); | |
581 STN2(&(xo[24]), T7y, T7G, ovs); | |
582 T7H = VFMAI(T7e, T7d); | |
583 STM2(&(xo[6]), T7H, ovs, &(xo[2])); | |
584 T7I = VFNMSI(T7e, T7d); | |
585 STM2(&(xo[122]), T7I, ovs, &(xo[2])); | |
586 STN2(&(xo[120]), T7A, T7I, ovs); | |
587 T7J = VFMAI(T7c, T75); | |
588 STM2(&(xo[70]), T7J, ovs, &(xo[2])); | |
589 T7K = VFNMSI(T7c, T75); | |
590 STM2(&(xo[58]), T7K, ovs, &(xo[2])); | |
591 STN2(&(xo[56]), T7C, T7K, ovs); | |
592 T2n = VSUB(T1X, T1q); | |
593 T1Y = VADD(T1q, T1X); | |
594 } | |
595 T2C = VFNMS(LDK(KP707106781), T27, T26); | |
596 T28 = VFMA(LDK(KP707106781), T27, T26); | |
597 T2b = VSUB(T29, T2a); | |
598 T2s = VADD(T2a, T29); | |
599 } | |
600 T2l = VFNMS(LDK(KP980785280), T2k, T2j); | |
601 T2p = VFMA(LDK(KP980785280), T2k, T2j); | |
602 { | |
603 V T5z, T4z, T5A, T5g; | |
604 { | |
605 V T4f, T4y, T1Z, T2h, T4U, T5t, T2m, T2c, T5u, T5f; | |
606 T5H = VFNMS(LDK(KP923879532), T4e, T47); | |
607 T4f = VFMA(LDK(KP923879532), T4e, T47); | |
608 T4y = VADD(T4o, T4x); | |
609 T5T = VSUB(T4x, T4o); | |
610 T1Z = VFNMS(LDK(KP980785280), T1Y, TT); | |
611 T2h = VFMA(LDK(KP980785280), T1Y, TT); | |
612 T4U = VFNMS(LDK(KP098491403), T4T, T4M); | |
613 T5t = VFMA(LDK(KP098491403), T4M, T4T); | |
614 T2m = VFNMS(LDK(KP923879532), T2b, T28); | |
615 T2c = VFMA(LDK(KP923879532), T2b, T28); | |
616 T5u = VFMA(LDK(KP098491403), T57, T5e); | |
617 T5f = VFNMS(LDK(KP098491403), T5e, T57); | |
618 T5z = VFNMS(LDK(KP980785280), T4y, T4f); | |
619 T4z = VFMA(LDK(KP980785280), T4y, T4f); | |
620 T5S = VFNMS(LDK(KP923879532), T5n, T5k); | |
621 T5o = VFMA(LDK(KP923879532), T5n, T5k); | |
622 { | |
623 V T2o, T2q, T2i, T2g; | |
624 T2o = VFMA(LDK(KP980785280), T2n, T2m); | |
625 T2q = VFNMS(LDK(KP980785280), T2n, T2m); | |
626 T2i = VFMA(LDK(KP980785280), T2f, T2c); | |
627 T2g = VFNMS(LDK(KP980785280), T2f, T2c); | |
628 T5A = VADD(T5t, T5u); | |
629 T5v = VSUB(T5t, T5u); | |
630 T5D = VSUB(T5f, T4U); | |
631 T5g = VADD(T4U, T5f); | |
632 T7L = VFNMSI(T2o, T2l); | |
633 STM2(&(xo[92]), T7L, ovs, &(xo[0])); | |
634 { | |
635 V T7M, T7N, T7P, T7R; | |
636 T7M = VFMAI(T2o, T2l); | |
637 STM2(&(xo[36]), T7M, ovs, &(xo[0])); | |
638 STN2(&(xo[36]), T7M, T7E, ovs); | |
639 T7N = VFMAI(T2q, T2p); | |
640 STM2(&(xo[100]), T7N, ovs, &(xo[0])); | |
641 STN2(&(xo[100]), T7N, T7F, ovs); | |
642 T7O = VFNMSI(T2q, T2p); | |
643 STM2(&(xo[28]), T7O, ovs, &(xo[0])); | |
644 T7P = VFMAI(T2i, T2h); | |
645 STM2(&(xo[4]), T7P, ovs, &(xo[0])); | |
646 STN2(&(xo[4]), T7P, T7H, ovs); | |
647 T7Q = VFNMSI(T2i, T2h); | |
648 STM2(&(xo[124]), T7Q, ovs, &(xo[0])); | |
649 T7R = VFMAI(T2g, T1Z); | |
650 STM2(&(xo[68]), T7R, ovs, &(xo[0])); | |
651 STN2(&(xo[68]), T7R, T7J, ovs); | |
652 T7S = VFNMSI(T2g, T1Z); | |
653 STM2(&(xo[60]), T7S, ovs, &(xo[0])); | |
654 T5r = VSUB(T5p, T5q); | |
655 T5I = VADD(T5p, T5q); | |
656 } | |
657 } | |
658 } | |
659 T5x = VFMA(LDK(KP995184726), T5g, T4z); | |
660 T5h = VFNMS(LDK(KP995184726), T5g, T4z); | |
661 T5F = VFMA(LDK(KP995184726), T5A, T5z); | |
662 T5B = VFNMS(LDK(KP995184726), T5A, T5z); | |
663 } | |
664 } | |
665 { | |
666 V T6J, T6R, T6L, T6z, T6T, T6P; | |
667 { | |
668 V T6N, T6j, T6O, T6y; | |
669 { | |
670 V T6q, T6H, T5C, T5s, T6I, T6x; | |
671 T6q = VFNMS(LDK(KP534511135), T6p, T6m); | |
672 T6H = VFMA(LDK(KP534511135), T6m, T6p); | |
673 T5C = VFNMS(LDK(KP980785280), T5r, T5o); | |
674 T5s = VFMA(LDK(KP980785280), T5r, T5o); | |
675 T6I = VFMA(LDK(KP534511135), T6t, T6w); | |
676 T6x = VFNMS(LDK(KP534511135), T6w, T6t); | |
677 T6N = VFMA(LDK(KP831469612), T6i, T6b); | |
678 T6j = VFNMS(LDK(KP831469612), T6i, T6b); | |
679 { | |
680 V T5E, T5G, T5y, T5w; | |
681 T5E = VFNMS(LDK(KP995184726), T5D, T5C); | |
682 T5G = VFMA(LDK(KP995184726), T5D, T5C); | |
683 T5y = VFMA(LDK(KP995184726), T5v, T5s); | |
684 T5w = VFNMS(LDK(KP995184726), T5v, T5s); | |
685 T6O = VADD(T6H, T6I); | |
686 T6J = VSUB(T6H, T6I); | |
687 T6R = VSUB(T6x, T6q); | |
688 T6y = VADD(T6q, T6x); | |
689 { | |
690 V T7T, T7U, T7V, T7W; | |
691 T7T = VFMAI(T5E, T5B); | |
692 STM2(&(xo[94]), T7T, ovs, &(xo[2])); | |
693 STN2(&(xo[92]), T7L, T7T, ovs); | |
694 T7U = VFNMSI(T5E, T5B); | |
695 STM2(&(xo[34]), T7U, ovs, &(xo[2])); | |
696 STN2(&(xo[32]), T7o, T7U, ovs); | |
697 T7V = VFNMSI(T5G, T5F); | |
698 STM2(&(xo[98]), T7V, ovs, &(xo[2])); | |
699 STN2(&(xo[96]), T7n, T7V, ovs); | |
700 T7W = VFMAI(T5G, T5F); | |
701 STM2(&(xo[30]), T7W, ovs, &(xo[2])); | |
702 STN2(&(xo[28]), T7O, T7W, ovs); | |
703 { | |
704 V T7X, T7Y, T7Z, T80; | |
705 T7X = VFMAI(T5y, T5x); | |
706 STM2(&(xo[126]), T7X, ovs, &(xo[2])); | |
707 STN2(&(xo[124]), T7Q, T7X, ovs); | |
708 T7Y = VFNMSI(T5y, T5x); | |
709 STM2(&(xo[2]), T7Y, ovs, &(xo[2])); | |
710 STN2(&(xo[0]), T7p, T7Y, ovs); | |
711 T7Z = VFMAI(T5w, T5h); | |
712 STM2(&(xo[62]), T7Z, ovs, &(xo[2])); | |
713 STN2(&(xo[60]), T7S, T7Z, ovs); | |
714 T80 = VFNMSI(T5w, T5h); | |
715 STM2(&(xo[66]), T80, ovs, &(xo[2])); | |
716 STN2(&(xo[64]), T7q, T80, ovs); | |
717 } | |
718 } | |
719 } | |
720 } | |
721 T6L = VFMA(LDK(KP881921264), T6y, T6j); | |
722 T6z = VFNMS(LDK(KP881921264), T6y, T6j); | |
723 T6T = VFMA(LDK(KP881921264), T6O, T6N); | |
724 T6P = VFNMS(LDK(KP881921264), T6O, T6N); | |
725 } | |
726 { | |
727 V T2H, T2P, T81, T84, T85, T87, T2J, T2B, T2R, T2N; | |
728 { | |
729 V T2L, T2t, T2M, T2A; | |
730 { | |
731 V T2z, T2F, T6Q, T6G, T2G, T2w; | |
732 T2z = VFMA(LDK(KP668178637), T2y, T2x); | |
733 T2F = VFNMS(LDK(KP668178637), T2x, T2y); | |
734 T6Q = VFMA(LDK(KP831469612), T6F, T6C); | |
735 T6G = VFNMS(LDK(KP831469612), T6F, T6C); | |
736 T2G = VFNMS(LDK(KP668178637), T2u, T2v); | |
737 T2w = VFMA(LDK(KP668178637), T2v, T2u); | |
738 T2L = VFNMS(LDK(KP923879532), T2s, T2r); | |
739 T2t = VFMA(LDK(KP923879532), T2s, T2r); | |
740 { | |
741 V T6S, T6U, T6M, T6K; | |
742 T6S = VFNMS(LDK(KP881921264), T6R, T6Q); | |
743 T6U = VFMA(LDK(KP881921264), T6R, T6Q); | |
744 T6M = VFMA(LDK(KP881921264), T6J, T6G); | |
745 T6K = VFNMS(LDK(KP881921264), T6J, T6G); | |
746 T2M = VADD(T2G, T2F); | |
747 T2H = VSUB(T2F, T2G); | |
748 T2P = VSUB(T2z, T2w); | |
749 T2A = VADD(T2w, T2z); | |
750 T81 = VFMAI(T6S, T6P); | |
751 STM2(&(xo[86]), T81, ovs, &(xo[2])); | |
752 { | |
753 V T82, T83, T86, T88; | |
754 T82 = VFNMSI(T6S, T6P); | |
755 STM2(&(xo[42]), T82, ovs, &(xo[2])); | |
756 STN2(&(xo[40]), T7w, T82, ovs); | |
757 T83 = VFNMSI(T6U, T6T); | |
758 STM2(&(xo[106]), T83, ovs, &(xo[2])); | |
759 STN2(&(xo[104]), T7x, T83, ovs); | |
760 T84 = VFMAI(T6U, T6T); | |
761 STM2(&(xo[22]), T84, ovs, &(xo[2])); | |
762 T85 = VFMAI(T6M, T6L); | |
763 STM2(&(xo[118]), T85, ovs, &(xo[2])); | |
764 T86 = VFNMSI(T6M, T6L); | |
765 STM2(&(xo[10]), T86, ovs, &(xo[2])); | |
766 STN2(&(xo[8]), T7z, T86, ovs); | |
767 T87 = VFMAI(T6K, T6z); | |
768 STM2(&(xo[54]), T87, ovs, &(xo[2])); | |
769 T88 = VFNMSI(T6K, T6z); | |
770 STM2(&(xo[74]), T88, ovs, &(xo[2])); | |
771 STN2(&(xo[72]), T7B, T88, ovs); | |
772 } | |
773 } | |
774 } | |
775 T2J = VFMA(LDK(KP831469612), T2A, T2t); | |
776 T2B = VFNMS(LDK(KP831469612), T2A, T2t); | |
777 T2R = VFNMS(LDK(KP831469612), T2M, T2L); | |
778 T2N = VFMA(LDK(KP831469612), T2M, T2L); | |
779 } | |
780 { | |
781 V T61, T5J, T62, T5Q; | |
782 { | |
783 V T5M, T5V, T2O, T2E, T5W, T5P; | |
784 T5M = VFMA(LDK(KP820678790), T5L, T5K); | |
785 T5V = VFNMS(LDK(KP820678790), T5K, T5L); | |
786 T2O = VFMA(LDK(KP923879532), T2D, T2C); | |
787 T2E = VFNMS(LDK(KP923879532), T2D, T2C); | |
788 T5W = VFNMS(LDK(KP820678790), T5N, T5O); | |
789 T5P = VFMA(LDK(KP820678790), T5O, T5N); | |
790 T61 = VFNMS(LDK(KP980785280), T5I, T5H); | |
791 T5J = VFMA(LDK(KP980785280), T5I, T5H); | |
792 { | |
793 V T2Q, T2S, T2K, T2I; | |
794 T2Q = VFNMS(LDK(KP831469612), T2P, T2O); | |
795 T2S = VFMA(LDK(KP831469612), T2P, T2O); | |
796 T2K = VFMA(LDK(KP831469612), T2H, T2E); | |
797 T2I = VFNMS(LDK(KP831469612), T2H, T2E); | |
798 T62 = VADD(T5V, T5W); | |
799 T5X = VSUB(T5V, T5W); | |
800 T65 = VSUB(T5P, T5M); | |
801 T5Q = VADD(T5M, T5P); | |
802 { | |
803 V T89, T8c, T8d, T8f; | |
804 T89 = VFMAI(T2Q, T2N); | |
805 STM2(&(xo[84]), T89, ovs, &(xo[0])); | |
806 STN2(&(xo[84]), T89, T81, ovs); | |
807 T8a = VFNMSI(T2Q, T2N); | |
808 STM2(&(xo[44]), T8a, ovs, &(xo[0])); | |
809 T8b = VFNMSI(T2S, T2R); | |
810 STM2(&(xo[108]), T8b, ovs, &(xo[0])); | |
811 T8c = VFMAI(T2S, T2R); | |
812 STM2(&(xo[20]), T8c, ovs, &(xo[0])); | |
813 STN2(&(xo[20]), T8c, T84, ovs); | |
814 T8d = VFMAI(T2K, T2J); | |
815 STM2(&(xo[116]), T8d, ovs, &(xo[0])); | |
816 STN2(&(xo[116]), T8d, T85, ovs); | |
817 T8e = VFNMSI(T2K, T2J); | |
818 STM2(&(xo[12]), T8e, ovs, &(xo[0])); | |
819 T8f = VFMAI(T2I, T2B); | |
820 STM2(&(xo[52]), T8f, ovs, &(xo[0])); | |
821 STN2(&(xo[52]), T8f, T87, ovs); | |
822 T8g = VFNMSI(T2I, T2B); | |
823 STM2(&(xo[76]), T8g, ovs, &(xo[0])); | |
824 } | |
825 } | |
826 } | |
827 T5Z = VFMA(LDK(KP773010453), T5Q, T5J); | |
828 T5R = VFNMS(LDK(KP773010453), T5Q, T5J); | |
829 T67 = VFNMS(LDK(KP773010453), T62, T61); | |
830 T63 = VFMA(LDK(KP773010453), T62, T61); | |
831 } | |
832 } | |
833 } | |
834 } | |
835 } | |
836 } | |
837 T5U = VFMA(LDK(KP980785280), T5T, T5S); | |
838 T64 = VFNMS(LDK(KP980785280), T5T, T5S); | |
839 { | |
840 V T68, T66, T5Y, T60; | |
841 T68 = VFNMS(LDK(KP773010453), T65, T64); | |
842 T66 = VFMA(LDK(KP773010453), T65, T64); | |
843 T5Y = VFNMS(LDK(KP773010453), T5X, T5U); | |
844 T60 = VFMA(LDK(KP773010453), T5X, T5U); | |
845 { | |
846 V T8h, T8i, T8j, T8k; | |
847 T8h = VFNMSI(T66, T63); | |
848 STM2(&(xo[82]), T8h, ovs, &(xo[2])); | |
849 STN2(&(xo[80]), T7t, T8h, ovs); | |
850 T8i = VFMAI(T66, T63); | |
851 STM2(&(xo[46]), T8i, ovs, &(xo[2])); | |
852 STN2(&(xo[44]), T8a, T8i, ovs); | |
853 T8j = VFMAI(T68, T67); | |
854 STM2(&(xo[110]), T8j, ovs, &(xo[2])); | |
855 STN2(&(xo[108]), T8b, T8j, ovs); | |
856 T8k = VFNMSI(T68, T67); | |
857 STM2(&(xo[18]), T8k, ovs, &(xo[2])); | |
858 STN2(&(xo[16]), T7r, T8k, ovs); | |
859 { | |
860 V T8l, T8m, T8n, T8o; | |
861 T8l = VFMAI(T60, T5Z); | |
862 STM2(&(xo[14]), T8l, ovs, &(xo[2])); | |
863 STN2(&(xo[12]), T8e, T8l, ovs); | |
864 T8m = VFNMSI(T60, T5Z); | |
865 STM2(&(xo[114]), T8m, ovs, &(xo[2])); | |
866 STN2(&(xo[112]), T7s, T8m, ovs); | |
867 T8n = VFMAI(T5Y, T5R); | |
868 STM2(&(xo[78]), T8n, ovs, &(xo[2])); | |
869 STN2(&(xo[76]), T8g, T8n, ovs); | |
870 T8o = VFNMSI(T5Y, T5R); | |
871 STM2(&(xo[50]), T8o, ovs, &(xo[2])); | |
872 STN2(&(xo[48]), T7u, T8o, ovs); | |
873 } | |
874 } | |
875 } | |
876 } | |
877 } | |
878 VLEAVE(); | |
879 } | |
880 | |
881 static const kdft_desc desc = { 64, XSIMD_STRING("n2fv_64"), {198, 0, 258, 0}, &GENUS, 0, 2, 0, 0 }; | |
882 | |
883 void XSIMD(codelet_n2fv_64) (planner *p) { | |
884 X(kdft_register) (p, n2fv_64, &desc); | |
885 } | |
886 | |
887 #else /* HAVE_FMA */ | |
888 | |
889 /* Generated by: ../../../genfft/gen_notw_c.native -simd -compact -variables 4 -pipeline-latency 8 -n 64 -name n2fv_64 -with-ostride 2 -include n2f.h -store-multiple 2 */ | |
890 | |
891 /* | |
892 * This function contains 456 FP additions, 124 FP multiplications, | |
893 * (or, 404 additions, 72 multiplications, 52 fused multiply/add), | |
894 * 128 stack variables, 15 constants, and 160 memory accesses | |
895 */ | |
896 #include "n2f.h" | |
897 | |
898 static void n2fv_64(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs) | |
899 { | |
900 DVK(KP471396736, +0.471396736825997648556387625905254377657460319); | |
901 DVK(KP881921264, +0.881921264348355029712756863660388349508442621); | |
902 DVK(KP098017140, +0.098017140329560601994195563888641845861136673); | |
903 DVK(KP995184726, +0.995184726672196886244836953109479921575474869); | |
904 DVK(KP290284677, +0.290284677254462367636192375817395274691476278); | |
905 DVK(KP956940335, +0.956940335732208864935797886980269969482849206); | |
906 DVK(KP634393284, +0.634393284163645498215171613225493370675687095); | |
907 DVK(KP773010453, +0.773010453362736960810906609758469800971041293); | |
908 DVK(KP555570233, +0.555570233019602224742830813948532874374937191); | |
909 DVK(KP831469612, +0.831469612302545237078788377617905756738560812); | |
910 DVK(KP980785280, +0.980785280403230449126182236134239036973933731); | |
911 DVK(KP195090322, +0.195090322016128267848284868477022240927691618); | |
912 DVK(KP382683432, +0.382683432365089771728459984030398866761344562); | |
913 DVK(KP923879532, +0.923879532511286756128183189396788286822416626); | |
914 DVK(KP707106781, +0.707106781186547524400844362104849039284835938); | |
915 { | |
916 INT i; | |
917 const R *xi; | |
918 R *xo; | |
919 xi = ri; | |
920 xo = ro; | |
921 for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(128, is), MAKE_VOLATILE_STRIDE(128, os)) { | |
922 V T4p, T5q, Tb, T39, T2n, T3A, T6f, T6T, Tq, T3B, T6i, T76, T2i, T3a, T4w; | |
923 V T5r, TI, T2p, T6C, T6V, T3h, T3E, T4L, T5u, TZ, T2q, T6F, T6U, T3e, T3D; | |
924 V T4E, T5t, T23, T2N, T6t, T71, T6w, T72, T2c, T2O, T3t, T41, T5f, T5R, T5k; | |
925 V T5S, T3w, T42, T1s, T2K, T6m, T6Y, T6p, T6Z, T1B, T2L, T3m, T3Y, T4Y, T5O; | |
926 V T53, T5P, T3p, T3Z; | |
927 { | |
928 V T3, T4n, T2m, T4o, T6, T5p, T9, T5o; | |
929 { | |
930 V T1, T2, T2k, T2l; | |
931 T1 = LD(&(xi[0]), ivs, &(xi[0])); | |
932 T2 = LD(&(xi[WS(is, 32)]), ivs, &(xi[0])); | |
933 T3 = VSUB(T1, T2); | |
934 T4n = VADD(T1, T2); | |
935 T2k = LD(&(xi[WS(is, 16)]), ivs, &(xi[0])); | |
936 T2l = LD(&(xi[WS(is, 48)]), ivs, &(xi[0])); | |
937 T2m = VSUB(T2k, T2l); | |
938 T4o = VADD(T2k, T2l); | |
939 } | |
940 { | |
941 V T4, T5, T7, T8; | |
942 T4 = LD(&(xi[WS(is, 8)]), ivs, &(xi[0])); | |
943 T5 = LD(&(xi[WS(is, 40)]), ivs, &(xi[0])); | |
944 T6 = VSUB(T4, T5); | |
945 T5p = VADD(T4, T5); | |
946 T7 = LD(&(xi[WS(is, 56)]), ivs, &(xi[0])); | |
947 T8 = LD(&(xi[WS(is, 24)]), ivs, &(xi[0])); | |
948 T9 = VSUB(T7, T8); | |
949 T5o = VADD(T7, T8); | |
950 } | |
951 T4p = VSUB(T4n, T4o); | |
952 T5q = VSUB(T5o, T5p); | |
953 { | |
954 V Ta, T2j, T6d, T6e; | |
955 Ta = VMUL(LDK(KP707106781), VADD(T6, T9)); | |
956 Tb = VADD(T3, Ta); | |
957 T39 = VSUB(T3, Ta); | |
958 T2j = VMUL(LDK(KP707106781), VSUB(T9, T6)); | |
959 T2n = VSUB(T2j, T2m); | |
960 T3A = VADD(T2m, T2j); | |
961 T6d = VADD(T4n, T4o); | |
962 T6e = VADD(T5p, T5o); | |
963 T6f = VADD(T6d, T6e); | |
964 T6T = VSUB(T6d, T6e); | |
965 } | |
966 } | |
967 { | |
968 V Te, T4q, To, T4u, Th, T4r, Tl, T4t; | |
969 { | |
970 V Tc, Td, Tm, Tn; | |
971 Tc = LD(&(xi[WS(is, 4)]), ivs, &(xi[0])); | |
972 Td = LD(&(xi[WS(is, 36)]), ivs, &(xi[0])); | |
973 Te = VSUB(Tc, Td); | |
974 T4q = VADD(Tc, Td); | |
975 Tm = LD(&(xi[WS(is, 12)]), ivs, &(xi[0])); | |
976 Tn = LD(&(xi[WS(is, 44)]), ivs, &(xi[0])); | |
977 To = VSUB(Tm, Tn); | |
978 T4u = VADD(Tm, Tn); | |
979 } | |
980 { | |
981 V Tf, Tg, Tj, Tk; | |
982 Tf = LD(&(xi[WS(is, 20)]), ivs, &(xi[0])); | |
983 Tg = LD(&(xi[WS(is, 52)]), ivs, &(xi[0])); | |
984 Th = VSUB(Tf, Tg); | |
985 T4r = VADD(Tf, Tg); | |
986 Tj = LD(&(xi[WS(is, 60)]), ivs, &(xi[0])); | |
987 Tk = LD(&(xi[WS(is, 28)]), ivs, &(xi[0])); | |
988 Tl = VSUB(Tj, Tk); | |
989 T4t = VADD(Tj, Tk); | |
990 } | |
991 { | |
992 V Ti, Tp, T6g, T6h; | |
993 Ti = VFNMS(LDK(KP382683432), Th, VMUL(LDK(KP923879532), Te)); | |
994 Tp = VFMA(LDK(KP923879532), Tl, VMUL(LDK(KP382683432), To)); | |
995 Tq = VADD(Ti, Tp); | |
996 T3B = VSUB(Tp, Ti); | |
997 T6g = VADD(T4q, T4r); | |
998 T6h = VADD(T4t, T4u); | |
999 T6i = VADD(T6g, T6h); | |
1000 T76 = VSUB(T6h, T6g); | |
1001 } | |
1002 { | |
1003 V T2g, T2h, T4s, T4v; | |
1004 T2g = VFNMS(LDK(KP923879532), To, VMUL(LDK(KP382683432), Tl)); | |
1005 T2h = VFMA(LDK(KP382683432), Te, VMUL(LDK(KP923879532), Th)); | |
1006 T2i = VSUB(T2g, T2h); | |
1007 T3a = VADD(T2h, T2g); | |
1008 T4s = VSUB(T4q, T4r); | |
1009 T4v = VSUB(T4t, T4u); | |
1010 T4w = VMUL(LDK(KP707106781), VADD(T4s, T4v)); | |
1011 T5r = VMUL(LDK(KP707106781), VSUB(T4v, T4s)); | |
1012 } | |
1013 } | |
1014 { | |
1015 V Tu, T4F, TG, T4G, TB, T4J, TD, T4I; | |
1016 { | |
1017 V Ts, Tt, TE, TF; | |
1018 Ts = LD(&(xi[WS(is, 62)]), ivs, &(xi[0])); | |
1019 Tt = LD(&(xi[WS(is, 30)]), ivs, &(xi[0])); | |
1020 Tu = VSUB(Ts, Tt); | |
1021 T4F = VADD(Ts, Tt); | |
1022 TE = LD(&(xi[WS(is, 14)]), ivs, &(xi[0])); | |
1023 TF = LD(&(xi[WS(is, 46)]), ivs, &(xi[0])); | |
1024 TG = VSUB(TE, TF); | |
1025 T4G = VADD(TE, TF); | |
1026 { | |
1027 V Tv, Tw, Tx, Ty, Tz, TA; | |
1028 Tv = LD(&(xi[WS(is, 6)]), ivs, &(xi[0])); | |
1029 Tw = LD(&(xi[WS(is, 38)]), ivs, &(xi[0])); | |
1030 Tx = VSUB(Tv, Tw); | |
1031 Ty = LD(&(xi[WS(is, 54)]), ivs, &(xi[0])); | |
1032 Tz = LD(&(xi[WS(is, 22)]), ivs, &(xi[0])); | |
1033 TA = VSUB(Ty, Tz); | |
1034 TB = VMUL(LDK(KP707106781), VADD(Tx, TA)); | |
1035 T4J = VADD(Tv, Tw); | |
1036 TD = VMUL(LDK(KP707106781), VSUB(TA, Tx)); | |
1037 T4I = VADD(Ty, Tz); | |
1038 } | |
1039 } | |
1040 { | |
1041 V TC, TH, T6A, T6B; | |
1042 TC = VADD(Tu, TB); | |
1043 TH = VSUB(TD, TG); | |
1044 TI = VFMA(LDK(KP195090322), TC, VMUL(LDK(KP980785280), TH)); | |
1045 T2p = VFNMS(LDK(KP195090322), TH, VMUL(LDK(KP980785280), TC)); | |
1046 T6A = VADD(T4F, T4G); | |
1047 T6B = VADD(T4J, T4I); | |
1048 T6C = VADD(T6A, T6B); | |
1049 T6V = VSUB(T6A, T6B); | |
1050 } | |
1051 { | |
1052 V T3f, T3g, T4H, T4K; | |
1053 T3f = VSUB(Tu, TB); | |
1054 T3g = VADD(TG, TD); | |
1055 T3h = VFNMS(LDK(KP555570233), T3g, VMUL(LDK(KP831469612), T3f)); | |
1056 T3E = VFMA(LDK(KP555570233), T3f, VMUL(LDK(KP831469612), T3g)); | |
1057 T4H = VSUB(T4F, T4G); | |
1058 T4K = VSUB(T4I, T4J); | |
1059 T4L = VFNMS(LDK(KP382683432), T4K, VMUL(LDK(KP923879532), T4H)); | |
1060 T5u = VFMA(LDK(KP382683432), T4H, VMUL(LDK(KP923879532), T4K)); | |
1061 } | |
1062 } | |
1063 { | |
1064 V TS, T4z, TW, T4y, TP, T4C, TX, T4B; | |
1065 { | |
1066 V TQ, TR, TU, TV; | |
1067 TQ = LD(&(xi[WS(is, 18)]), ivs, &(xi[0])); | |
1068 TR = LD(&(xi[WS(is, 50)]), ivs, &(xi[0])); | |
1069 TS = VSUB(TQ, TR); | |
1070 T4z = VADD(TQ, TR); | |
1071 TU = LD(&(xi[WS(is, 2)]), ivs, &(xi[0])); | |
1072 TV = LD(&(xi[WS(is, 34)]), ivs, &(xi[0])); | |
1073 TW = VSUB(TU, TV); | |
1074 T4y = VADD(TU, TV); | |
1075 { | |
1076 V TJ, TK, TL, TM, TN, TO; | |
1077 TJ = LD(&(xi[WS(is, 58)]), ivs, &(xi[0])); | |
1078 TK = LD(&(xi[WS(is, 26)]), ivs, &(xi[0])); | |
1079 TL = VSUB(TJ, TK); | |
1080 TM = LD(&(xi[WS(is, 10)]), ivs, &(xi[0])); | |
1081 TN = LD(&(xi[WS(is, 42)]), ivs, &(xi[0])); | |
1082 TO = VSUB(TM, TN); | |
1083 TP = VMUL(LDK(KP707106781), VSUB(TL, TO)); | |
1084 T4C = VADD(TM, TN); | |
1085 TX = VMUL(LDK(KP707106781), VADD(TO, TL)); | |
1086 T4B = VADD(TJ, TK); | |
1087 } | |
1088 } | |
1089 { | |
1090 V TT, TY, T6D, T6E; | |
1091 TT = VSUB(TP, TS); | |
1092 TY = VADD(TW, TX); | |
1093 TZ = VFNMS(LDK(KP195090322), TY, VMUL(LDK(KP980785280), TT)); | |
1094 T2q = VFMA(LDK(KP980785280), TY, VMUL(LDK(KP195090322), TT)); | |
1095 T6D = VADD(T4y, T4z); | |
1096 T6E = VADD(T4C, T4B); | |
1097 T6F = VADD(T6D, T6E); | |
1098 T6U = VSUB(T6D, T6E); | |
1099 } | |
1100 { | |
1101 V T3c, T3d, T4A, T4D; | |
1102 T3c = VSUB(TW, TX); | |
1103 T3d = VADD(TS, TP); | |
1104 T3e = VFMA(LDK(KP831469612), T3c, VMUL(LDK(KP555570233), T3d)); | |
1105 T3D = VFNMS(LDK(KP555570233), T3c, VMUL(LDK(KP831469612), T3d)); | |
1106 T4A = VSUB(T4y, T4z); | |
1107 T4D = VSUB(T4B, T4C); | |
1108 T4E = VFMA(LDK(KP923879532), T4A, VMUL(LDK(KP382683432), T4D)); | |
1109 T5t = VFNMS(LDK(KP382683432), T4A, VMUL(LDK(KP923879532), T4D)); | |
1110 } | |
1111 } | |
1112 { | |
1113 V T1F, T55, T2a, T56, T1M, T5h, T27, T5g, T58, T59, T1U, T5a, T25, T5b, T5c; | |
1114 V T21, T5d, T24; | |
1115 { | |
1116 V T1D, T1E, T28, T29; | |
1117 T1D = LD(&(xi[WS(is, 63)]), ivs, &(xi[WS(is, 1)])); | |
1118 T1E = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)])); | |
1119 T1F = VSUB(T1D, T1E); | |
1120 T55 = VADD(T1D, T1E); | |
1121 T28 = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)])); | |
1122 T29 = LD(&(xi[WS(is, 47)]), ivs, &(xi[WS(is, 1)])); | |
1123 T2a = VSUB(T28, T29); | |
1124 T56 = VADD(T28, T29); | |
1125 } | |
1126 { | |
1127 V T1G, T1H, T1I, T1J, T1K, T1L; | |
1128 T1G = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)])); | |
1129 T1H = LD(&(xi[WS(is, 39)]), ivs, &(xi[WS(is, 1)])); | |
1130 T1I = VSUB(T1G, T1H); | |
1131 T1J = LD(&(xi[WS(is, 55)]), ivs, &(xi[WS(is, 1)])); | |
1132 T1K = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)])); | |
1133 T1L = VSUB(T1J, T1K); | |
1134 T1M = VMUL(LDK(KP707106781), VADD(T1I, T1L)); | |
1135 T5h = VADD(T1G, T1H); | |
1136 T27 = VMUL(LDK(KP707106781), VSUB(T1L, T1I)); | |
1137 T5g = VADD(T1J, T1K); | |
1138 } | |
1139 { | |
1140 V T1Q, T1T, T1X, T20; | |
1141 { | |
1142 V T1O, T1P, T1R, T1S; | |
1143 T1O = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)])); | |
1144 T1P = LD(&(xi[WS(is, 35)]), ivs, &(xi[WS(is, 1)])); | |
1145 T1Q = VSUB(T1O, T1P); | |
1146 T58 = VADD(T1O, T1P); | |
1147 T1R = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)])); | |
1148 T1S = LD(&(xi[WS(is, 51)]), ivs, &(xi[WS(is, 1)])); | |
1149 T1T = VSUB(T1R, T1S); | |
1150 T59 = VADD(T1R, T1S); | |
1151 } | |
1152 T1U = VFNMS(LDK(KP382683432), T1T, VMUL(LDK(KP923879532), T1Q)); | |
1153 T5a = VSUB(T58, T59); | |
1154 T25 = VFMA(LDK(KP382683432), T1Q, VMUL(LDK(KP923879532), T1T)); | |
1155 { | |
1156 V T1V, T1W, T1Y, T1Z; | |
1157 T1V = LD(&(xi[WS(is, 59)]), ivs, &(xi[WS(is, 1)])); | |
1158 T1W = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)])); | |
1159 T1X = VSUB(T1V, T1W); | |
1160 T5b = VADD(T1V, T1W); | |
1161 T1Y = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)])); | |
1162 T1Z = LD(&(xi[WS(is, 43)]), ivs, &(xi[WS(is, 1)])); | |
1163 T20 = VSUB(T1Y, T1Z); | |
1164 T5c = VADD(T1Y, T1Z); | |
1165 } | |
1166 T21 = VFMA(LDK(KP923879532), T1X, VMUL(LDK(KP382683432), T20)); | |
1167 T5d = VSUB(T5b, T5c); | |
1168 T24 = VFNMS(LDK(KP923879532), T20, VMUL(LDK(KP382683432), T1X)); | |
1169 } | |
1170 { | |
1171 V T1N, T22, T6r, T6s; | |
1172 T1N = VADD(T1F, T1M); | |
1173 T22 = VADD(T1U, T21); | |
1174 T23 = VSUB(T1N, T22); | |
1175 T2N = VADD(T1N, T22); | |
1176 T6r = VADD(T55, T56); | |
1177 T6s = VADD(T5h, T5g); | |
1178 T6t = VADD(T6r, T6s); | |
1179 T71 = VSUB(T6r, T6s); | |
1180 } | |
1181 { | |
1182 V T6u, T6v, T26, T2b; | |
1183 T6u = VADD(T58, T59); | |
1184 T6v = VADD(T5b, T5c); | |
1185 T6w = VADD(T6u, T6v); | |
1186 T72 = VSUB(T6v, T6u); | |
1187 T26 = VSUB(T24, T25); | |
1188 T2b = VSUB(T27, T2a); | |
1189 T2c = VSUB(T26, T2b); | |
1190 T2O = VADD(T2b, T26); | |
1191 } | |
1192 { | |
1193 V T3r, T3s, T57, T5e; | |
1194 T3r = VSUB(T1F, T1M); | |
1195 T3s = VADD(T25, T24); | |
1196 T3t = VADD(T3r, T3s); | |
1197 T41 = VSUB(T3r, T3s); | |
1198 T57 = VSUB(T55, T56); | |
1199 T5e = VMUL(LDK(KP707106781), VADD(T5a, T5d)); | |
1200 T5f = VADD(T57, T5e); | |
1201 T5R = VSUB(T57, T5e); | |
1202 } | |
1203 { | |
1204 V T5i, T5j, T3u, T3v; | |
1205 T5i = VSUB(T5g, T5h); | |
1206 T5j = VMUL(LDK(KP707106781), VSUB(T5d, T5a)); | |
1207 T5k = VADD(T5i, T5j); | |
1208 T5S = VSUB(T5j, T5i); | |
1209 T3u = VADD(T2a, T27); | |
1210 T3v = VSUB(T21, T1U); | |
1211 T3w = VADD(T3u, T3v); | |
1212 T42 = VSUB(T3v, T3u); | |
1213 } | |
1214 } | |
1215 { | |
1216 V T1q, T4P, T1v, T4O, T1n, T50, T1w, T4Z, T4U, T4V, T18, T4W, T1z, T4R, T4S; | |
1217 V T1f, T4T, T1y; | |
1218 { | |
1219 V T1o, T1p, T1t, T1u; | |
1220 T1o = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)])); | |
1221 T1p = LD(&(xi[WS(is, 49)]), ivs, &(xi[WS(is, 1)])); | |
1222 T1q = VSUB(T1o, T1p); | |
1223 T4P = VADD(T1o, T1p); | |
1224 T1t = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)])); | |
1225 T1u = LD(&(xi[WS(is, 33)]), ivs, &(xi[WS(is, 1)])); | |
1226 T1v = VSUB(T1t, T1u); | |
1227 T4O = VADD(T1t, T1u); | |
1228 } | |
1229 { | |
1230 V T1h, T1i, T1j, T1k, T1l, T1m; | |
1231 T1h = LD(&(xi[WS(is, 57)]), ivs, &(xi[WS(is, 1)])); | |
1232 T1i = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)])); | |
1233 T1j = VSUB(T1h, T1i); | |
1234 T1k = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)])); | |
1235 T1l = LD(&(xi[WS(is, 41)]), ivs, &(xi[WS(is, 1)])); | |
1236 T1m = VSUB(T1k, T1l); | |
1237 T1n = VMUL(LDK(KP707106781), VSUB(T1j, T1m)); | |
1238 T50 = VADD(T1k, T1l); | |
1239 T1w = VMUL(LDK(KP707106781), VADD(T1m, T1j)); | |
1240 T4Z = VADD(T1h, T1i); | |
1241 } | |
1242 { | |
1243 V T14, T17, T1b, T1e; | |
1244 { | |
1245 V T12, T13, T15, T16; | |
1246 T12 = LD(&(xi[WS(is, 61)]), ivs, &(xi[WS(is, 1)])); | |
1247 T13 = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)])); | |
1248 T14 = VSUB(T12, T13); | |
1249 T4U = VADD(T12, T13); | |
1250 T15 = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)])); | |
1251 T16 = LD(&(xi[WS(is, 45)]), ivs, &(xi[WS(is, 1)])); | |
1252 T17 = VSUB(T15, T16); | |
1253 T4V = VADD(T15, T16); | |
1254 } | |
1255 T18 = VFNMS(LDK(KP923879532), T17, VMUL(LDK(KP382683432), T14)); | |
1256 T4W = VSUB(T4U, T4V); | |
1257 T1z = VFMA(LDK(KP923879532), T14, VMUL(LDK(KP382683432), T17)); | |
1258 { | |
1259 V T19, T1a, T1c, T1d; | |
1260 T19 = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)])); | |
1261 T1a = LD(&(xi[WS(is, 37)]), ivs, &(xi[WS(is, 1)])); | |
1262 T1b = VSUB(T19, T1a); | |
1263 T4R = VADD(T19, T1a); | |
1264 T1c = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)])); | |
1265 T1d = LD(&(xi[WS(is, 53)]), ivs, &(xi[WS(is, 1)])); | |
1266 T1e = VSUB(T1c, T1d); | |
1267 T4S = VADD(T1c, T1d); | |
1268 } | |
1269 T1f = VFMA(LDK(KP382683432), T1b, VMUL(LDK(KP923879532), T1e)); | |
1270 T4T = VSUB(T4R, T4S); | |
1271 T1y = VFNMS(LDK(KP382683432), T1e, VMUL(LDK(KP923879532), T1b)); | |
1272 } | |
1273 { | |
1274 V T1g, T1r, T6k, T6l; | |
1275 T1g = VSUB(T18, T1f); | |
1276 T1r = VSUB(T1n, T1q); | |
1277 T1s = VSUB(T1g, T1r); | |
1278 T2K = VADD(T1r, T1g); | |
1279 T6k = VADD(T4O, T4P); | |
1280 T6l = VADD(T50, T4Z); | |
1281 T6m = VADD(T6k, T6l); | |
1282 T6Y = VSUB(T6k, T6l); | |
1283 } | |
1284 { | |
1285 V T6n, T6o, T1x, T1A; | |
1286 T6n = VADD(T4R, T4S); | |
1287 T6o = VADD(T4U, T4V); | |
1288 T6p = VADD(T6n, T6o); | |
1289 T6Z = VSUB(T6o, T6n); | |
1290 T1x = VADD(T1v, T1w); | |
1291 T1A = VADD(T1y, T1z); | |
1292 T1B = VSUB(T1x, T1A); | |
1293 T2L = VADD(T1x, T1A); | |
1294 } | |
1295 { | |
1296 V T3k, T3l, T4Q, T4X; | |
1297 T3k = VSUB(T1v, T1w); | |
1298 T3l = VADD(T1f, T18); | |
1299 T3m = VADD(T3k, T3l); | |
1300 T3Y = VSUB(T3k, T3l); | |
1301 T4Q = VSUB(T4O, T4P); | |
1302 T4X = VMUL(LDK(KP707106781), VADD(T4T, T4W)); | |
1303 T4Y = VADD(T4Q, T4X); | |
1304 T5O = VSUB(T4Q, T4X); | |
1305 } | |
1306 { | |
1307 V T51, T52, T3n, T3o; | |
1308 T51 = VSUB(T4Z, T50); | |
1309 T52 = VMUL(LDK(KP707106781), VSUB(T4W, T4T)); | |
1310 T53 = VADD(T51, T52); | |
1311 T5P = VSUB(T52, T51); | |
1312 T3n = VADD(T1q, T1n); | |
1313 T3o = VSUB(T1z, T1y); | |
1314 T3p = VADD(T3n, T3o); | |
1315 T3Z = VSUB(T3o, T3n); | |
1316 } | |
1317 } | |
1318 { | |
1319 V T7n, T7o, T7p, T7q, T7r, T7s, T7t, T7u, T7v, T7w, T7x, T7y, T7z, T7A, T7B; | |
1320 V T7C, T7D, T7E, T7F, T7G, T7H, T7I, T7J, T7K; | |
1321 { | |
1322 V T6N, T6R, T6Q, T6S; | |
1323 { | |
1324 V T6L, T6M, T6O, T6P; | |
1325 T6L = VADD(T6f, T6i); | |
1326 T6M = VADD(T6F, T6C); | |
1327 T6N = VADD(T6L, T6M); | |
1328 T6R = VSUB(T6L, T6M); | |
1329 T6O = VADD(T6m, T6p); | |
1330 T6P = VADD(T6t, T6w); | |
1331 T6Q = VADD(T6O, T6P); | |
1332 T6S = VBYI(VSUB(T6P, T6O)); | |
1333 } | |
1334 T7n = VSUB(T6N, T6Q); | |
1335 STM2(&(xo[64]), T7n, ovs, &(xo[0])); | |
1336 T7o = VADD(T6R, T6S); | |
1337 STM2(&(xo[32]), T7o, ovs, &(xo[0])); | |
1338 T7p = VADD(T6N, T6Q); | |
1339 STM2(&(xo[0]), T7p, ovs, &(xo[0])); | |
1340 T7q = VSUB(T6R, T6S); | |
1341 STM2(&(xo[96]), T7q, ovs, &(xo[0])); | |
1342 } | |
1343 { | |
1344 V T6j, T6G, T6y, T6H, T6q, T6x; | |
1345 T6j = VSUB(T6f, T6i); | |
1346 T6G = VSUB(T6C, T6F); | |
1347 T6q = VSUB(T6m, T6p); | |
1348 T6x = VSUB(T6t, T6w); | |
1349 T6y = VMUL(LDK(KP707106781), VADD(T6q, T6x)); | |
1350 T6H = VMUL(LDK(KP707106781), VSUB(T6x, T6q)); | |
1351 { | |
1352 V T6z, T6I, T6J, T6K; | |
1353 T6z = VADD(T6j, T6y); | |
1354 T6I = VBYI(VADD(T6G, T6H)); | |
1355 T7r = VSUB(T6z, T6I); | |
1356 STM2(&(xo[112]), T7r, ovs, &(xo[0])); | |
1357 T7s = VADD(T6z, T6I); | |
1358 STM2(&(xo[16]), T7s, ovs, &(xo[0])); | |
1359 T6J = VSUB(T6j, T6y); | |
1360 T6K = VBYI(VSUB(T6H, T6G)); | |
1361 T7t = VSUB(T6J, T6K); | |
1362 STM2(&(xo[80]), T7t, ovs, &(xo[0])); | |
1363 T7u = VADD(T6J, T6K); | |
1364 STM2(&(xo[48]), T7u, ovs, &(xo[0])); | |
1365 } | |
1366 } | |
1367 { | |
1368 V T6X, T7i, T78, T7g, T74, T7f, T7b, T7j, T6W, T77; | |
1369 T6W = VMUL(LDK(KP707106781), VADD(T6U, T6V)); | |
1370 T6X = VADD(T6T, T6W); | |
1371 T7i = VSUB(T6T, T6W); | |
1372 T77 = VMUL(LDK(KP707106781), VSUB(T6V, T6U)); | |
1373 T78 = VADD(T76, T77); | |
1374 T7g = VSUB(T77, T76); | |
1375 { | |
1376 V T70, T73, T79, T7a; | |
1377 T70 = VFMA(LDK(KP923879532), T6Y, VMUL(LDK(KP382683432), T6Z)); | |
1378 T73 = VFNMS(LDK(KP382683432), T72, VMUL(LDK(KP923879532), T71)); | |
1379 T74 = VADD(T70, T73); | |
1380 T7f = VSUB(T73, T70); | |
1381 T79 = VFNMS(LDK(KP382683432), T6Y, VMUL(LDK(KP923879532), T6Z)); | |
1382 T7a = VFMA(LDK(KP382683432), T71, VMUL(LDK(KP923879532), T72)); | |
1383 T7b = VADD(T79, T7a); | |
1384 T7j = VSUB(T7a, T79); | |
1385 } | |
1386 { | |
1387 V T75, T7c, T7l, T7m; | |
1388 T75 = VADD(T6X, T74); | |
1389 T7c = VBYI(VADD(T78, T7b)); | |
1390 T7v = VSUB(T75, T7c); | |
1391 STM2(&(xo[120]), T7v, ovs, &(xo[0])); | |
1392 T7w = VADD(T75, T7c); | |
1393 STM2(&(xo[8]), T7w, ovs, &(xo[0])); | |
1394 T7l = VBYI(VADD(T7g, T7f)); | |
1395 T7m = VADD(T7i, T7j); | |
1396 T7x = VADD(T7l, T7m); | |
1397 STM2(&(xo[24]), T7x, ovs, &(xo[0])); | |
1398 T7y = VSUB(T7m, T7l); | |
1399 STM2(&(xo[104]), T7y, ovs, &(xo[0])); | |
1400 } | |
1401 { | |
1402 V T7d, T7e, T7h, T7k; | |
1403 T7d = VSUB(T6X, T74); | |
1404 T7e = VBYI(VSUB(T7b, T78)); | |
1405 T7z = VSUB(T7d, T7e); | |
1406 STM2(&(xo[72]), T7z, ovs, &(xo[0])); | |
1407 T7A = VADD(T7d, T7e); | |
1408 STM2(&(xo[56]), T7A, ovs, &(xo[0])); | |
1409 T7h = VBYI(VSUB(T7f, T7g)); | |
1410 T7k = VSUB(T7i, T7j); | |
1411 T7B = VADD(T7h, T7k); | |
1412 STM2(&(xo[40]), T7B, ovs, &(xo[0])); | |
1413 T7C = VSUB(T7k, T7h); | |
1414 STM2(&(xo[88]), T7C, ovs, &(xo[0])); | |
1415 } | |
1416 } | |
1417 { | |
1418 V T5N, T68, T61, T69, T5U, T65, T5Y, T66; | |
1419 { | |
1420 V T5L, T5M, T5Z, T60; | |
1421 T5L = VSUB(T4p, T4w); | |
1422 T5M = VSUB(T5u, T5t); | |
1423 T5N = VADD(T5L, T5M); | |
1424 T68 = VSUB(T5L, T5M); | |
1425 T5Z = VFNMS(LDK(KP555570233), T5O, VMUL(LDK(KP831469612), T5P)); | |
1426 T60 = VFMA(LDK(KP555570233), T5R, VMUL(LDK(KP831469612), T5S)); | |
1427 T61 = VADD(T5Z, T60); | |
1428 T69 = VSUB(T60, T5Z); | |
1429 } | |
1430 { | |
1431 V T5Q, T5T, T5W, T5X; | |
1432 T5Q = VFMA(LDK(KP831469612), T5O, VMUL(LDK(KP555570233), T5P)); | |
1433 T5T = VFNMS(LDK(KP555570233), T5S, VMUL(LDK(KP831469612), T5R)); | |
1434 T5U = VADD(T5Q, T5T); | |
1435 T65 = VSUB(T5T, T5Q); | |
1436 T5W = VSUB(T5r, T5q); | |
1437 T5X = VSUB(T4L, T4E); | |
1438 T5Y = VADD(T5W, T5X); | |
1439 T66 = VSUB(T5X, T5W); | |
1440 } | |
1441 { | |
1442 V T5V, T62, T6b, T6c; | |
1443 T5V = VADD(T5N, T5U); | |
1444 T62 = VBYI(VADD(T5Y, T61)); | |
1445 T7D = VSUB(T5V, T62); | |
1446 STM2(&(xo[116]), T7D, ovs, &(xo[0])); | |
1447 T7E = VADD(T5V, T62); | |
1448 STM2(&(xo[12]), T7E, ovs, &(xo[0])); | |
1449 T6b = VBYI(VADD(T66, T65)); | |
1450 T6c = VADD(T68, T69); | |
1451 T7F = VADD(T6b, T6c); | |
1452 STM2(&(xo[20]), T7F, ovs, &(xo[0])); | |
1453 T7G = VSUB(T6c, T6b); | |
1454 STM2(&(xo[108]), T7G, ovs, &(xo[0])); | |
1455 } | |
1456 { | |
1457 V T63, T64, T67, T6a; | |
1458 T63 = VSUB(T5N, T5U); | |
1459 T64 = VBYI(VSUB(T61, T5Y)); | |
1460 T7H = VSUB(T63, T64); | |
1461 STM2(&(xo[76]), T7H, ovs, &(xo[0])); | |
1462 T7I = VADD(T63, T64); | |
1463 STM2(&(xo[52]), T7I, ovs, &(xo[0])); | |
1464 T67 = VBYI(VSUB(T65, T66)); | |
1465 T6a = VSUB(T68, T69); | |
1466 T7J = VADD(T67, T6a); | |
1467 STM2(&(xo[44]), T7J, ovs, &(xo[0])); | |
1468 T7K = VSUB(T6a, T67); | |
1469 STM2(&(xo[84]), T7K, ovs, &(xo[0])); | |
1470 } | |
1471 } | |
1472 { | |
1473 V T7U, T7W, T7X, T7Z; | |
1474 { | |
1475 V T11, T2C, T2v, T2D, T2e, T2z, T2s, T2A; | |
1476 { | |
1477 V Tr, T10, T2t, T2u; | |
1478 Tr = VSUB(Tb, Tq); | |
1479 T10 = VSUB(TI, TZ); | |
1480 T11 = VADD(Tr, T10); | |
1481 T2C = VSUB(Tr, T10); | |
1482 T2t = VFNMS(LDK(KP634393284), T1B, VMUL(LDK(KP773010453), T1s)); | |
1483 T2u = VFMA(LDK(KP773010453), T2c, VMUL(LDK(KP634393284), T23)); | |
1484 T2v = VADD(T2t, T2u); | |
1485 T2D = VSUB(T2u, T2t); | |
1486 } | |
1487 { | |
1488 V T1C, T2d, T2o, T2r; | |
1489 T1C = VFMA(LDK(KP634393284), T1s, VMUL(LDK(KP773010453), T1B)); | |
1490 T2d = VFNMS(LDK(KP634393284), T2c, VMUL(LDK(KP773010453), T23)); | |
1491 T2e = VADD(T1C, T2d); | |
1492 T2z = VSUB(T2d, T1C); | |
1493 T2o = VSUB(T2i, T2n); | |
1494 T2r = VSUB(T2p, T2q); | |
1495 T2s = VADD(T2o, T2r); | |
1496 T2A = VSUB(T2r, T2o); | |
1497 } | |
1498 { | |
1499 V T2f, T2w, T7L, T7M; | |
1500 T2f = VADD(T11, T2e); | |
1501 T2w = VBYI(VADD(T2s, T2v)); | |
1502 T7L = VSUB(T2f, T2w); | |
1503 STM2(&(xo[114]), T7L, ovs, &(xo[2])); | |
1504 STN2(&(xo[112]), T7r, T7L, ovs); | |
1505 T7M = VADD(T2f, T2w); | |
1506 STM2(&(xo[14]), T7M, ovs, &(xo[2])); | |
1507 STN2(&(xo[12]), T7E, T7M, ovs); | |
1508 } | |
1509 { | |
1510 V T2F, T2G, T7N, T7O; | |
1511 T2F = VBYI(VADD(T2A, T2z)); | |
1512 T2G = VADD(T2C, T2D); | |
1513 T7N = VADD(T2F, T2G); | |
1514 STM2(&(xo[18]), T7N, ovs, &(xo[2])); | |
1515 STN2(&(xo[16]), T7s, T7N, ovs); | |
1516 T7O = VSUB(T2G, T2F); | |
1517 STM2(&(xo[110]), T7O, ovs, &(xo[2])); | |
1518 STN2(&(xo[108]), T7G, T7O, ovs); | |
1519 } | |
1520 { | |
1521 V T2x, T2y, T7P, T7Q; | |
1522 T2x = VSUB(T11, T2e); | |
1523 T2y = VBYI(VSUB(T2v, T2s)); | |
1524 T7P = VSUB(T2x, T2y); | |
1525 STM2(&(xo[78]), T7P, ovs, &(xo[2])); | |
1526 STN2(&(xo[76]), T7H, T7P, ovs); | |
1527 T7Q = VADD(T2x, T2y); | |
1528 STM2(&(xo[50]), T7Q, ovs, &(xo[2])); | |
1529 STN2(&(xo[48]), T7u, T7Q, ovs); | |
1530 } | |
1531 { | |
1532 V T2B, T2E, T7R, T7S; | |
1533 T2B = VBYI(VSUB(T2z, T2A)); | |
1534 T2E = VSUB(T2C, T2D); | |
1535 T7R = VADD(T2B, T2E); | |
1536 STM2(&(xo[46]), T7R, ovs, &(xo[2])); | |
1537 STN2(&(xo[44]), T7J, T7R, ovs); | |
1538 T7S = VSUB(T2E, T2B); | |
1539 STM2(&(xo[82]), T7S, ovs, &(xo[2])); | |
1540 STN2(&(xo[80]), T7t, T7S, ovs); | |
1541 } | |
1542 } | |
1543 { | |
1544 V T3j, T3Q, T3J, T3R, T3y, T3N, T3G, T3O; | |
1545 { | |
1546 V T3b, T3i, T3H, T3I; | |
1547 T3b = VADD(T39, T3a); | |
1548 T3i = VADD(T3e, T3h); | |
1549 T3j = VADD(T3b, T3i); | |
1550 T3Q = VSUB(T3b, T3i); | |
1551 T3H = VFNMS(LDK(KP290284677), T3m, VMUL(LDK(KP956940335), T3p)); | |
1552 T3I = VFMA(LDK(KP290284677), T3t, VMUL(LDK(KP956940335), T3w)); | |
1553 T3J = VADD(T3H, T3I); | |
1554 T3R = VSUB(T3I, T3H); | |
1555 } | |
1556 { | |
1557 V T3q, T3x, T3C, T3F; | |
1558 T3q = VFMA(LDK(KP956940335), T3m, VMUL(LDK(KP290284677), T3p)); | |
1559 T3x = VFNMS(LDK(KP290284677), T3w, VMUL(LDK(KP956940335), T3t)); | |
1560 T3y = VADD(T3q, T3x); | |
1561 T3N = VSUB(T3x, T3q); | |
1562 T3C = VADD(T3A, T3B); | |
1563 T3F = VADD(T3D, T3E); | |
1564 T3G = VADD(T3C, T3F); | |
1565 T3O = VSUB(T3F, T3C); | |
1566 } | |
1567 { | |
1568 V T3z, T3K, T7T, T3T, T3U, T7V; | |
1569 T3z = VADD(T3j, T3y); | |
1570 T3K = VBYI(VADD(T3G, T3J)); | |
1571 T7T = VSUB(T3z, T3K); | |
1572 STM2(&(xo[122]), T7T, ovs, &(xo[2])); | |
1573 STN2(&(xo[120]), T7v, T7T, ovs); | |
1574 T7U = VADD(T3z, T3K); | |
1575 STM2(&(xo[6]), T7U, ovs, &(xo[2])); | |
1576 T3T = VBYI(VADD(T3O, T3N)); | |
1577 T3U = VADD(T3Q, T3R); | |
1578 T7V = VADD(T3T, T3U); | |
1579 STM2(&(xo[26]), T7V, ovs, &(xo[2])); | |
1580 STN2(&(xo[24]), T7x, T7V, ovs); | |
1581 T7W = VSUB(T3U, T3T); | |
1582 STM2(&(xo[102]), T7W, ovs, &(xo[2])); | |
1583 } | |
1584 { | |
1585 V T3L, T3M, T7Y, T3P, T3S, T80; | |
1586 T3L = VSUB(T3j, T3y); | |
1587 T3M = VBYI(VSUB(T3J, T3G)); | |
1588 T7X = VSUB(T3L, T3M); | |
1589 STM2(&(xo[70]), T7X, ovs, &(xo[2])); | |
1590 T7Y = VADD(T3L, T3M); | |
1591 STM2(&(xo[58]), T7Y, ovs, &(xo[2])); | |
1592 STN2(&(xo[56]), T7A, T7Y, ovs); | |
1593 T3P = VBYI(VSUB(T3N, T3O)); | |
1594 T3S = VSUB(T3Q, T3R); | |
1595 T7Z = VADD(T3P, T3S); | |
1596 STM2(&(xo[38]), T7Z, ovs, &(xo[2])); | |
1597 T80 = VSUB(T3S, T3P); | |
1598 STM2(&(xo[90]), T80, ovs, &(xo[2])); | |
1599 STN2(&(xo[88]), T7C, T80, ovs); | |
1600 } | |
1601 } | |
1602 { | |
1603 V T81, T83, T86, T88; | |
1604 { | |
1605 V T4N, T5G, T5z, T5H, T5m, T5D, T5w, T5E; | |
1606 { | |
1607 V T4x, T4M, T5x, T5y; | |
1608 T4x = VADD(T4p, T4w); | |
1609 T4M = VADD(T4E, T4L); | |
1610 T4N = VADD(T4x, T4M); | |
1611 T5G = VSUB(T4x, T4M); | |
1612 T5x = VFNMS(LDK(KP195090322), T4Y, VMUL(LDK(KP980785280), T53)); | |
1613 T5y = VFMA(LDK(KP195090322), T5f, VMUL(LDK(KP980785280), T5k)); | |
1614 T5z = VADD(T5x, T5y); | |
1615 T5H = VSUB(T5y, T5x); | |
1616 } | |
1617 { | |
1618 V T54, T5l, T5s, T5v; | |
1619 T54 = VFMA(LDK(KP980785280), T4Y, VMUL(LDK(KP195090322), T53)); | |
1620 T5l = VFNMS(LDK(KP195090322), T5k, VMUL(LDK(KP980785280), T5f)); | |
1621 T5m = VADD(T54, T5l); | |
1622 T5D = VSUB(T5l, T54); | |
1623 T5s = VADD(T5q, T5r); | |
1624 T5v = VADD(T5t, T5u); | |
1625 T5w = VADD(T5s, T5v); | |
1626 T5E = VSUB(T5v, T5s); | |
1627 } | |
1628 { | |
1629 V T5n, T5A, T82, T5J, T5K, T84; | |
1630 T5n = VADD(T4N, T5m); | |
1631 T5A = VBYI(VADD(T5w, T5z)); | |
1632 T81 = VSUB(T5n, T5A); | |
1633 STM2(&(xo[124]), T81, ovs, &(xo[0])); | |
1634 T82 = VADD(T5n, T5A); | |
1635 STM2(&(xo[4]), T82, ovs, &(xo[0])); | |
1636 STN2(&(xo[4]), T82, T7U, ovs); | |
1637 T5J = VBYI(VADD(T5E, T5D)); | |
1638 T5K = VADD(T5G, T5H); | |
1639 T83 = VADD(T5J, T5K); | |
1640 STM2(&(xo[28]), T83, ovs, &(xo[0])); | |
1641 T84 = VSUB(T5K, T5J); | |
1642 STM2(&(xo[100]), T84, ovs, &(xo[0])); | |
1643 STN2(&(xo[100]), T84, T7W, ovs); | |
1644 } | |
1645 { | |
1646 V T5B, T5C, T85, T5F, T5I, T87; | |
1647 T5B = VSUB(T4N, T5m); | |
1648 T5C = VBYI(VSUB(T5z, T5w)); | |
1649 T85 = VSUB(T5B, T5C); | |
1650 STM2(&(xo[68]), T85, ovs, &(xo[0])); | |
1651 STN2(&(xo[68]), T85, T7X, ovs); | |
1652 T86 = VADD(T5B, T5C); | |
1653 STM2(&(xo[60]), T86, ovs, &(xo[0])); | |
1654 T5F = VBYI(VSUB(T5D, T5E)); | |
1655 T5I = VSUB(T5G, T5H); | |
1656 T87 = VADD(T5F, T5I); | |
1657 STM2(&(xo[36]), T87, ovs, &(xo[0])); | |
1658 STN2(&(xo[36]), T87, T7Z, ovs); | |
1659 T88 = VSUB(T5I, T5F); | |
1660 STM2(&(xo[92]), T88, ovs, &(xo[0])); | |
1661 } | |
1662 } | |
1663 { | |
1664 V T2J, T34, T2X, T35, T2Q, T31, T2U, T32; | |
1665 { | |
1666 V T2H, T2I, T2V, T2W; | |
1667 T2H = VADD(Tb, Tq); | |
1668 T2I = VADD(T2q, T2p); | |
1669 T2J = VADD(T2H, T2I); | |
1670 T34 = VSUB(T2H, T2I); | |
1671 T2V = VFNMS(LDK(KP098017140), T2L, VMUL(LDK(KP995184726), T2K)); | |
1672 T2W = VFMA(LDK(KP995184726), T2O, VMUL(LDK(KP098017140), T2N)); | |
1673 T2X = VADD(T2V, T2W); | |
1674 T35 = VSUB(T2W, T2V); | |
1675 } | |
1676 { | |
1677 V T2M, T2P, T2S, T2T; | |
1678 T2M = VFMA(LDK(KP098017140), T2K, VMUL(LDK(KP995184726), T2L)); | |
1679 T2P = VFNMS(LDK(KP098017140), T2O, VMUL(LDK(KP995184726), T2N)); | |
1680 T2Q = VADD(T2M, T2P); | |
1681 T31 = VSUB(T2P, T2M); | |
1682 T2S = VADD(T2n, T2i); | |
1683 T2T = VADD(TZ, TI); | |
1684 T2U = VADD(T2S, T2T); | |
1685 T32 = VSUB(T2T, T2S); | |
1686 } | |
1687 { | |
1688 V T2R, T2Y, T89, T8a; | |
1689 T2R = VADD(T2J, T2Q); | |
1690 T2Y = VBYI(VADD(T2U, T2X)); | |
1691 T89 = VSUB(T2R, T2Y); | |
1692 STM2(&(xo[126]), T89, ovs, &(xo[2])); | |
1693 STN2(&(xo[124]), T81, T89, ovs); | |
1694 T8a = VADD(T2R, T2Y); | |
1695 STM2(&(xo[2]), T8a, ovs, &(xo[2])); | |
1696 STN2(&(xo[0]), T7p, T8a, ovs); | |
1697 } | |
1698 { | |
1699 V T37, T38, T8b, T8c; | |
1700 T37 = VBYI(VADD(T32, T31)); | |
1701 T38 = VADD(T34, T35); | |
1702 T8b = VADD(T37, T38); | |
1703 STM2(&(xo[30]), T8b, ovs, &(xo[2])); | |
1704 STN2(&(xo[28]), T83, T8b, ovs); | |
1705 T8c = VSUB(T38, T37); | |
1706 STM2(&(xo[98]), T8c, ovs, &(xo[2])); | |
1707 STN2(&(xo[96]), T7q, T8c, ovs); | |
1708 } | |
1709 { | |
1710 V T2Z, T30, T8d, T8e; | |
1711 T2Z = VSUB(T2J, T2Q); | |
1712 T30 = VBYI(VSUB(T2X, T2U)); | |
1713 T8d = VSUB(T2Z, T30); | |
1714 STM2(&(xo[66]), T8d, ovs, &(xo[2])); | |
1715 STN2(&(xo[64]), T7n, T8d, ovs); | |
1716 T8e = VADD(T2Z, T30); | |
1717 STM2(&(xo[62]), T8e, ovs, &(xo[2])); | |
1718 STN2(&(xo[60]), T86, T8e, ovs); | |
1719 } | |
1720 { | |
1721 V T33, T36, T8f, T8g; | |
1722 T33 = VBYI(VSUB(T31, T32)); | |
1723 T36 = VSUB(T34, T35); | |
1724 T8f = VADD(T33, T36); | |
1725 STM2(&(xo[34]), T8f, ovs, &(xo[2])); | |
1726 STN2(&(xo[32]), T7o, T8f, ovs); | |
1727 T8g = VSUB(T36, T33); | |
1728 STM2(&(xo[94]), T8g, ovs, &(xo[2])); | |
1729 STN2(&(xo[92]), T88, T8g, ovs); | |
1730 } | |
1731 } | |
1732 { | |
1733 V T3X, T4i, T4b, T4j, T44, T4f, T48, T4g; | |
1734 { | |
1735 V T3V, T3W, T49, T4a; | |
1736 T3V = VSUB(T39, T3a); | |
1737 T3W = VSUB(T3E, T3D); | |
1738 T3X = VADD(T3V, T3W); | |
1739 T4i = VSUB(T3V, T3W); | |
1740 T49 = VFNMS(LDK(KP471396736), T3Y, VMUL(LDK(KP881921264), T3Z)); | |
1741 T4a = VFMA(LDK(KP471396736), T41, VMUL(LDK(KP881921264), T42)); | |
1742 T4b = VADD(T49, T4a); | |
1743 T4j = VSUB(T4a, T49); | |
1744 } | |
1745 { | |
1746 V T40, T43, T46, T47; | |
1747 T40 = VFMA(LDK(KP881921264), T3Y, VMUL(LDK(KP471396736), T3Z)); | |
1748 T43 = VFNMS(LDK(KP471396736), T42, VMUL(LDK(KP881921264), T41)); | |
1749 T44 = VADD(T40, T43); | |
1750 T4f = VSUB(T43, T40); | |
1751 T46 = VSUB(T3B, T3A); | |
1752 T47 = VSUB(T3h, T3e); | |
1753 T48 = VADD(T46, T47); | |
1754 T4g = VSUB(T47, T46); | |
1755 } | |
1756 { | |
1757 V T45, T4c, T8h, T8i; | |
1758 T45 = VADD(T3X, T44); | |
1759 T4c = VBYI(VADD(T48, T4b)); | |
1760 T8h = VSUB(T45, T4c); | |
1761 STM2(&(xo[118]), T8h, ovs, &(xo[2])); | |
1762 STN2(&(xo[116]), T7D, T8h, ovs); | |
1763 T8i = VADD(T45, T4c); | |
1764 STM2(&(xo[10]), T8i, ovs, &(xo[2])); | |
1765 STN2(&(xo[8]), T7w, T8i, ovs); | |
1766 } | |
1767 { | |
1768 V T4l, T4m, T8j, T8k; | |
1769 T4l = VBYI(VADD(T4g, T4f)); | |
1770 T4m = VADD(T4i, T4j); | |
1771 T8j = VADD(T4l, T4m); | |
1772 STM2(&(xo[22]), T8j, ovs, &(xo[2])); | |
1773 STN2(&(xo[20]), T7F, T8j, ovs); | |
1774 T8k = VSUB(T4m, T4l); | |
1775 STM2(&(xo[106]), T8k, ovs, &(xo[2])); | |
1776 STN2(&(xo[104]), T7y, T8k, ovs); | |
1777 } | |
1778 { | |
1779 V T4d, T4e, T8l, T8m; | |
1780 T4d = VSUB(T3X, T44); | |
1781 T4e = VBYI(VSUB(T4b, T48)); | |
1782 T8l = VSUB(T4d, T4e); | |
1783 STM2(&(xo[74]), T8l, ovs, &(xo[2])); | |
1784 STN2(&(xo[72]), T7z, T8l, ovs); | |
1785 T8m = VADD(T4d, T4e); | |
1786 STM2(&(xo[54]), T8m, ovs, &(xo[2])); | |
1787 STN2(&(xo[52]), T7I, T8m, ovs); | |
1788 } | |
1789 { | |
1790 V T4h, T4k, T8n, T8o; | |
1791 T4h = VBYI(VSUB(T4f, T4g)); | |
1792 T4k = VSUB(T4i, T4j); | |
1793 T8n = VADD(T4h, T4k); | |
1794 STM2(&(xo[42]), T8n, ovs, &(xo[2])); | |
1795 STN2(&(xo[40]), T7B, T8n, ovs); | |
1796 T8o = VSUB(T4k, T4h); | |
1797 STM2(&(xo[86]), T8o, ovs, &(xo[2])); | |
1798 STN2(&(xo[84]), T7K, T8o, ovs); | |
1799 } | |
1800 } | |
1801 } | |
1802 } | |
1803 } | |
1804 } | |
1805 } | |
1806 VLEAVE(); | |
1807 } | |
1808 | |
1809 static const kdft_desc desc = { 64, XSIMD_STRING("n2fv_64"), {404, 72, 52, 0}, &GENUS, 0, 2, 0, 0 }; | |
1810 | |
1811 void XSIMD(codelet_n2fv_64) (planner *p) { | |
1812 X(kdft_register) (p, n2fv_64, &desc); | |
1813 } | |
1814 | |
1815 #endif /* HAVE_FMA */ |