Chris@42
|
1 /*
|
Chris@42
|
2 * Copyright (c) 2003, 2007-14 Matteo Frigo
|
Chris@42
|
3 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
|
Chris@42
|
4 *
|
Chris@42
|
5 * This program is free software; you can redistribute it and/or modify
|
Chris@42
|
6 * it under the terms of the GNU General Public License as published by
|
Chris@42
|
7 * the Free Software Foundation; either version 2 of the License, or
|
Chris@42
|
8 * (at your option) any later version.
|
Chris@42
|
9 *
|
Chris@42
|
10 * This program is distributed in the hope that it will be useful,
|
Chris@42
|
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
Chris@42
|
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
Chris@42
|
13 * GNU General Public License for more details.
|
Chris@42
|
14 *
|
Chris@42
|
15 * You should have received a copy of the GNU General Public License
|
Chris@42
|
16 * along with this program; if not, write to the Free Software
|
Chris@42
|
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
Chris@42
|
18 *
|
Chris@42
|
19 */
|
Chris@42
|
20
|
Chris@42
|
21 /* This file was automatically generated --- DO NOT EDIT */
|
Chris@42
|
22 /* Generated on Sat Jul 30 16:46:27 EDT 2016 */
|
Chris@42
|
23
|
Chris@42
|
24 #include "codelet-rdft.h"
|
Chris@42
|
25
|
Chris@42
|
26 #ifdef HAVE_FMA
|
Chris@42
|
27
|
Chris@42
|
28 /* Generated by: ../../../genfft/gen_hc2hc.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -n 32 -dit -name hf_32 -include hf.h */
|
Chris@42
|
29
|
Chris@42
|
30 /*
|
Chris@42
|
31 * This function contains 434 FP additions, 260 FP multiplications,
|
Chris@42
|
32 * (or, 236 additions, 62 multiplications, 198 fused multiply/add),
|
Chris@42
|
33 * 135 stack variables, 7 constants, and 128 memory accesses
|
Chris@42
|
34 */
|
Chris@42
|
35 #include "hf.h"
|
Chris@42
|
36
|
Chris@42
|
37 static void hf_32(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
|
Chris@42
|
38 {
|
Chris@42
|
39 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
Chris@42
|
40 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
Chris@42
|
41 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
Chris@42
|
42 DK(KP668178637, +0.668178637919298919997757686523080761552472251);
|
Chris@42
|
43 DK(KP198912367, +0.198912367379658006911597622644676228597850501);
|
Chris@42
|
44 DK(KP414213562, +0.414213562373095048801688724209698078569671875);
|
Chris@42
|
45 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
Chris@42
|
46 {
|
Chris@42
|
47 INT m;
|
Chris@42
|
48 for (m = mb, W = W + ((mb - 1) * 62); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 62, MAKE_VOLATILE_STRIDE(64, rs)) {
|
Chris@42
|
49 E T6D, T6A;
|
Chris@42
|
50 {
|
Chris@42
|
51 E T8y, T87, T8, T3w, T83, T3B, T8x, Tl, T6G, Tz, T3J, T5T, T6F, TM, T3Q;
|
Chris@42
|
52 E T5U, T46, T5X, T7E, T6M, T5Y, T3Z, T6J, T1f, T7D, T6R, T61, T4e, T6O, T1G;
|
Chris@42
|
53 E T60, T4l, T54, T6c, T7d, T7N, T32, T76, T6f, T5r, T4v, T65, T72, T7I, T29;
|
Chris@42
|
54 E T6V, T68, T4S, T5t, T5b, T7O, T79, T7e, T3t, T5s, T5i, T4H, T2y, T4B, T6X;
|
Chris@42
|
55 E T2m, T4w, T4F, T2s;
|
Chris@42
|
56 {
|
Chris@42
|
57 E T44, T1d, T3X, T6K, T11, T40, T42, T17, T5h, T5c;
|
Chris@42
|
58 {
|
Chris@42
|
59 E Ta, Td, Tg, T3x, Tb, Tj, Tf, Tc, Ti;
|
Chris@42
|
60 {
|
Chris@42
|
61 E T1, T86, T3, T6, T2, T5;
|
Chris@42
|
62 T1 = cr[0];
|
Chris@42
|
63 T86 = ci[0];
|
Chris@42
|
64 T3 = cr[WS(rs, 16)];
|
Chris@42
|
65 T6 = ci[WS(rs, 16)];
|
Chris@42
|
66 T2 = W[30];
|
Chris@42
|
67 T5 = W[31];
|
Chris@42
|
68 {
|
Chris@42
|
69 E T84, T4, T9, T85, T7;
|
Chris@42
|
70 Ta = cr[WS(rs, 8)];
|
Chris@42
|
71 Td = ci[WS(rs, 8)];
|
Chris@42
|
72 T84 = T2 * T6;
|
Chris@42
|
73 T4 = T2 * T3;
|
Chris@42
|
74 T9 = W[14];
|
Chris@42
|
75 Tg = cr[WS(rs, 24)];
|
Chris@42
|
76 T85 = FNMS(T5, T3, T84);
|
Chris@42
|
77 T7 = FMA(T5, T6, T4);
|
Chris@42
|
78 T3x = T9 * Td;
|
Chris@42
|
79 Tb = T9 * Ta;
|
Chris@42
|
80 T8y = T86 - T85;
|
Chris@42
|
81 T87 = T85 + T86;
|
Chris@42
|
82 T8 = T1 + T7;
|
Chris@42
|
83 T3w = T1 - T7;
|
Chris@42
|
84 Tj = ci[WS(rs, 24)];
|
Chris@42
|
85 Tf = W[46];
|
Chris@42
|
86 }
|
Chris@42
|
87 Tc = W[15];
|
Chris@42
|
88 Ti = W[47];
|
Chris@42
|
89 }
|
Chris@42
|
90 {
|
Chris@42
|
91 E Tu, Tx, T3F, Ts, Tw, T3G, Tv;
|
Chris@42
|
92 {
|
Chris@42
|
93 E To, Tr, Tp, T3E, Tq, Tt;
|
Chris@42
|
94 {
|
Chris@42
|
95 E T3y, Te, T3A, Tk, T3z, Th, Tn;
|
Chris@42
|
96 To = cr[WS(rs, 4)];
|
Chris@42
|
97 T3z = Tf * Tj;
|
Chris@42
|
98 Th = Tf * Tg;
|
Chris@42
|
99 T3y = FNMS(Tc, Ta, T3x);
|
Chris@42
|
100 Te = FMA(Tc, Td, Tb);
|
Chris@42
|
101 T3A = FNMS(Ti, Tg, T3z);
|
Chris@42
|
102 Tk = FMA(Ti, Tj, Th);
|
Chris@42
|
103 Tr = ci[WS(rs, 4)];
|
Chris@42
|
104 Tn = W[6];
|
Chris@42
|
105 T83 = T3y + T3A;
|
Chris@42
|
106 T3B = T3y - T3A;
|
Chris@42
|
107 T8x = Te - Tk;
|
Chris@42
|
108 Tl = Te + Tk;
|
Chris@42
|
109 Tp = Tn * To;
|
Chris@42
|
110 T3E = Tn * Tr;
|
Chris@42
|
111 }
|
Chris@42
|
112 Tq = W[7];
|
Chris@42
|
113 Tu = cr[WS(rs, 20)];
|
Chris@42
|
114 Tx = ci[WS(rs, 20)];
|
Chris@42
|
115 Tt = W[38];
|
Chris@42
|
116 T3F = FNMS(Tq, To, T3E);
|
Chris@42
|
117 Ts = FMA(Tq, Tr, Tp);
|
Chris@42
|
118 Tw = W[39];
|
Chris@42
|
119 T3G = Tt * Tx;
|
Chris@42
|
120 Tv = Tt * Tu;
|
Chris@42
|
121 }
|
Chris@42
|
122 {
|
Chris@42
|
123 E T3M, TF, TH, TK, TG, TJ, TE, TD, TC;
|
Chris@42
|
124 {
|
Chris@42
|
125 E TB, T3H, Ty, TA, T3I, T3D, T3L;
|
Chris@42
|
126 TB = cr[WS(rs, 28)];
|
Chris@42
|
127 TE = ci[WS(rs, 28)];
|
Chris@42
|
128 T3H = FNMS(Tw, Tu, T3G);
|
Chris@42
|
129 Ty = FMA(Tw, Tx, Tv);
|
Chris@42
|
130 TA = W[54];
|
Chris@42
|
131 TD = W[55];
|
Chris@42
|
132 T6G = T3F + T3H;
|
Chris@42
|
133 T3I = T3F - T3H;
|
Chris@42
|
134 Tz = Ts + Ty;
|
Chris@42
|
135 T3D = Ts - Ty;
|
Chris@42
|
136 T3L = TA * TE;
|
Chris@42
|
137 TC = TA * TB;
|
Chris@42
|
138 T3J = T3D - T3I;
|
Chris@42
|
139 T5T = T3D + T3I;
|
Chris@42
|
140 T3M = FNMS(TD, TB, T3L);
|
Chris@42
|
141 }
|
Chris@42
|
142 TF = FMA(TD, TE, TC);
|
Chris@42
|
143 TH = cr[WS(rs, 12)];
|
Chris@42
|
144 TK = ci[WS(rs, 12)];
|
Chris@42
|
145 TG = W[22];
|
Chris@42
|
146 TJ = W[23];
|
Chris@42
|
147 {
|
Chris@42
|
148 E TU, T3U, T13, T16, T3W, T10, T12, T15, T41, T14;
|
Chris@42
|
149 {
|
Chris@42
|
150 E T19, T1c, T18, T1b, T3P, T3K;
|
Chris@42
|
151 {
|
Chris@42
|
152 E TQ, TT, T3N, TI, TP, TS;
|
Chris@42
|
153 TQ = cr[WS(rs, 2)];
|
Chris@42
|
154 TT = ci[WS(rs, 2)];
|
Chris@42
|
155 T3N = TG * TK;
|
Chris@42
|
156 TI = TG * TH;
|
Chris@42
|
157 TP = W[2];
|
Chris@42
|
158 TS = W[3];
|
Chris@42
|
159 {
|
Chris@42
|
160 E T3O, TL, T3T, TR;
|
Chris@42
|
161 T3O = FNMS(TJ, TH, T3N);
|
Chris@42
|
162 TL = FMA(TJ, TK, TI);
|
Chris@42
|
163 T3T = TP * TT;
|
Chris@42
|
164 TR = TP * TQ;
|
Chris@42
|
165 T6F = T3M + T3O;
|
Chris@42
|
166 T3P = T3M - T3O;
|
Chris@42
|
167 TM = TF + TL;
|
Chris@42
|
168 T3K = TF - TL;
|
Chris@42
|
169 TU = FMA(TS, TT, TR);
|
Chris@42
|
170 T3U = FNMS(TS, TQ, T3T);
|
Chris@42
|
171 }
|
Chris@42
|
172 }
|
Chris@42
|
173 T3Q = T3K + T3P;
|
Chris@42
|
174 T5U = T3K - T3P;
|
Chris@42
|
175 T19 = cr[WS(rs, 26)];
|
Chris@42
|
176 T1c = ci[WS(rs, 26)];
|
Chris@42
|
177 T18 = W[50];
|
Chris@42
|
178 T1b = W[51];
|
Chris@42
|
179 {
|
Chris@42
|
180 E TW, TZ, TY, T3V, TX, T43, T1a, TV;
|
Chris@42
|
181 TW = cr[WS(rs, 18)];
|
Chris@42
|
182 TZ = ci[WS(rs, 18)];
|
Chris@42
|
183 T43 = T18 * T1c;
|
Chris@42
|
184 T1a = T18 * T19;
|
Chris@42
|
185 TV = W[34];
|
Chris@42
|
186 TY = W[35];
|
Chris@42
|
187 T44 = FNMS(T1b, T19, T43);
|
Chris@42
|
188 T1d = FMA(T1b, T1c, T1a);
|
Chris@42
|
189 T3V = TV * TZ;
|
Chris@42
|
190 TX = TV * TW;
|
Chris@42
|
191 T13 = cr[WS(rs, 10)];
|
Chris@42
|
192 T16 = ci[WS(rs, 10)];
|
Chris@42
|
193 T3W = FNMS(TY, TW, T3V);
|
Chris@42
|
194 T10 = FMA(TY, TZ, TX);
|
Chris@42
|
195 T12 = W[18];
|
Chris@42
|
196 T15 = W[19];
|
Chris@42
|
197 }
|
Chris@42
|
198 }
|
Chris@42
|
199 T3X = T3U - T3W;
|
Chris@42
|
200 T6K = T3U + T3W;
|
Chris@42
|
201 T11 = TU + T10;
|
Chris@42
|
202 T40 = TU - T10;
|
Chris@42
|
203 T41 = T12 * T16;
|
Chris@42
|
204 T14 = T12 * T13;
|
Chris@42
|
205 T42 = FNMS(T15, T13, T41);
|
Chris@42
|
206 T17 = FMA(T15, T16, T14);
|
Chris@42
|
207 }
|
Chris@42
|
208 }
|
Chris@42
|
209 }
|
Chris@42
|
210 }
|
Chris@42
|
211 {
|
Chris@42
|
212 E T49, T1l, T4j, T1E, T1u, T1x, T1w, T4b, T1r, T4g, T1v;
|
Chris@42
|
213 {
|
Chris@42
|
214 E T1A, T1D, T1C, T4i, T1B;
|
Chris@42
|
215 {
|
Chris@42
|
216 E T1h, T1k, T1g, T1j, T48, T1i, T1z;
|
Chris@42
|
217 T1h = cr[WS(rs, 30)];
|
Chris@42
|
218 T1k = ci[WS(rs, 30)];
|
Chris@42
|
219 {
|
Chris@42
|
220 E T6L, T45, T1e, T3Y;
|
Chris@42
|
221 T6L = T42 + T44;
|
Chris@42
|
222 T45 = T42 - T44;
|
Chris@42
|
223 T1e = T17 + T1d;
|
Chris@42
|
224 T3Y = T17 - T1d;
|
Chris@42
|
225 T46 = T40 - T45;
|
Chris@42
|
226 T5X = T40 + T45;
|
Chris@42
|
227 T7E = T6K + T6L;
|
Chris@42
|
228 T6M = T6K - T6L;
|
Chris@42
|
229 T5Y = T3X - T3Y;
|
Chris@42
|
230 T3Z = T3X + T3Y;
|
Chris@42
|
231 T6J = T11 - T1e;
|
Chris@42
|
232 T1f = T11 + T1e;
|
Chris@42
|
233 T1g = W[58];
|
Chris@42
|
234 }
|
Chris@42
|
235 T1j = W[59];
|
Chris@42
|
236 T1A = cr[WS(rs, 22)];
|
Chris@42
|
237 T1D = ci[WS(rs, 22)];
|
Chris@42
|
238 T48 = T1g * T1k;
|
Chris@42
|
239 T1i = T1g * T1h;
|
Chris@42
|
240 T1z = W[42];
|
Chris@42
|
241 T1C = W[43];
|
Chris@42
|
242 T49 = FNMS(T1j, T1h, T48);
|
Chris@42
|
243 T1l = FMA(T1j, T1k, T1i);
|
Chris@42
|
244 T4i = T1z * T1D;
|
Chris@42
|
245 T1B = T1z * T1A;
|
Chris@42
|
246 }
|
Chris@42
|
247 {
|
Chris@42
|
248 E T1n, T1q, T1m, T1p, T4a, T1o, T1t;
|
Chris@42
|
249 T1n = cr[WS(rs, 14)];
|
Chris@42
|
250 T1q = ci[WS(rs, 14)];
|
Chris@42
|
251 T4j = FNMS(T1C, T1A, T4i);
|
Chris@42
|
252 T1E = FMA(T1C, T1D, T1B);
|
Chris@42
|
253 T1m = W[26];
|
Chris@42
|
254 T1p = W[27];
|
Chris@42
|
255 T1u = cr[WS(rs, 6)];
|
Chris@42
|
256 T1x = ci[WS(rs, 6)];
|
Chris@42
|
257 T4a = T1m * T1q;
|
Chris@42
|
258 T1o = T1m * T1n;
|
Chris@42
|
259 T1t = W[10];
|
Chris@42
|
260 T1w = W[11];
|
Chris@42
|
261 T4b = FNMS(T1p, T1n, T4a);
|
Chris@42
|
262 T1r = FMA(T1p, T1q, T1o);
|
Chris@42
|
263 T4g = T1t * T1x;
|
Chris@42
|
264 T1v = T1t * T1u;
|
Chris@42
|
265 }
|
Chris@42
|
266 }
|
Chris@42
|
267 {
|
Chris@42
|
268 E T4c, T6P, T1s, T4f, T4h, T1y;
|
Chris@42
|
269 T4c = T49 - T4b;
|
Chris@42
|
270 T6P = T49 + T4b;
|
Chris@42
|
271 T1s = T1l + T1r;
|
Chris@42
|
272 T4f = T1l - T1r;
|
Chris@42
|
273 T4h = FNMS(T1w, T1u, T4g);
|
Chris@42
|
274 T1y = FMA(T1w, T1x, T1v);
|
Chris@42
|
275 {
|
Chris@42
|
276 E T4k, T6Q, T4d, T1F;
|
Chris@42
|
277 T4k = T4h - T4j;
|
Chris@42
|
278 T6Q = T4h + T4j;
|
Chris@42
|
279 T4d = T1y - T1E;
|
Chris@42
|
280 T1F = T1y + T1E;
|
Chris@42
|
281 T7D = T6P + T6Q;
|
Chris@42
|
282 T6R = T6P - T6Q;
|
Chris@42
|
283 T61 = T4c - T4d;
|
Chris@42
|
284 T4e = T4c + T4d;
|
Chris@42
|
285 T6O = T1s - T1F;
|
Chris@42
|
286 T1G = T1s + T1F;
|
Chris@42
|
287 T60 = T4f + T4k;
|
Chris@42
|
288 T4l = T4f - T4k;
|
Chris@42
|
289 }
|
Chris@42
|
290 }
|
Chris@42
|
291 }
|
Chris@42
|
292 {
|
Chris@42
|
293 E T5n, T2H, T52, T30, T2Q, T2T, T2S, T5p, T2N, T4Z, T2R;
|
Chris@42
|
294 {
|
Chris@42
|
295 E T2W, T2Z, T2Y, T51, T2X;
|
Chris@42
|
296 {
|
Chris@42
|
297 E T2D, T2G, T2C, T2F, T5m, T2E, T2V;
|
Chris@42
|
298 T2D = cr[WS(rs, 31)];
|
Chris@42
|
299 T2G = ci[WS(rs, 31)];
|
Chris@42
|
300 T2C = W[60];
|
Chris@42
|
301 T2F = W[61];
|
Chris@42
|
302 T2W = cr[WS(rs, 23)];
|
Chris@42
|
303 T2Z = ci[WS(rs, 23)];
|
Chris@42
|
304 T5m = T2C * T2G;
|
Chris@42
|
305 T2E = T2C * T2D;
|
Chris@42
|
306 T2V = W[44];
|
Chris@42
|
307 T2Y = W[45];
|
Chris@42
|
308 T5n = FNMS(T2F, T2D, T5m);
|
Chris@42
|
309 T2H = FMA(T2F, T2G, T2E);
|
Chris@42
|
310 T51 = T2V * T2Z;
|
Chris@42
|
311 T2X = T2V * T2W;
|
Chris@42
|
312 }
|
Chris@42
|
313 {
|
Chris@42
|
314 E T2J, T2M, T2I, T2L, T5o, T2K, T2P;
|
Chris@42
|
315 T2J = cr[WS(rs, 15)];
|
Chris@42
|
316 T2M = ci[WS(rs, 15)];
|
Chris@42
|
317 T52 = FNMS(T2Y, T2W, T51);
|
Chris@42
|
318 T30 = FMA(T2Y, T2Z, T2X);
|
Chris@42
|
319 T2I = W[28];
|
Chris@42
|
320 T2L = W[29];
|
Chris@42
|
321 T2Q = cr[WS(rs, 7)];
|
Chris@42
|
322 T2T = ci[WS(rs, 7)];
|
Chris@42
|
323 T5o = T2I * T2M;
|
Chris@42
|
324 T2K = T2I * T2J;
|
Chris@42
|
325 T2P = W[12];
|
Chris@42
|
326 T2S = W[13];
|
Chris@42
|
327 T5p = FNMS(T2L, T2J, T5o);
|
Chris@42
|
328 T2N = FMA(T2L, T2M, T2K);
|
Chris@42
|
329 T4Z = T2P * T2T;
|
Chris@42
|
330 T2R = T2P * T2Q;
|
Chris@42
|
331 }
|
Chris@42
|
332 }
|
Chris@42
|
333 {
|
Chris@42
|
334 E T5q, T7b, T2O, T4Y, T50, T2U;
|
Chris@42
|
335 T5q = T5n - T5p;
|
Chris@42
|
336 T7b = T5n + T5p;
|
Chris@42
|
337 T2O = T2H + T2N;
|
Chris@42
|
338 T4Y = T2H - T2N;
|
Chris@42
|
339 T50 = FNMS(T2S, T2Q, T4Z);
|
Chris@42
|
340 T2U = FMA(T2S, T2T, T2R);
|
Chris@42
|
341 {
|
Chris@42
|
342 E T7c, T53, T31, T5l;
|
Chris@42
|
343 T7c = T50 + T52;
|
Chris@42
|
344 T53 = T50 - T52;
|
Chris@42
|
345 T31 = T2U + T30;
|
Chris@42
|
346 T5l = T30 - T2U;
|
Chris@42
|
347 T54 = T4Y - T53;
|
Chris@42
|
348 T6c = T4Y + T53;
|
Chris@42
|
349 T7d = T7b - T7c;
|
Chris@42
|
350 T7N = T7b + T7c;
|
Chris@42
|
351 T32 = T2O + T31;
|
Chris@42
|
352 T76 = T2O - T31;
|
Chris@42
|
353 T6f = T5q + T5l;
|
Chris@42
|
354 T5r = T5l - T5q;
|
Chris@42
|
355 }
|
Chris@42
|
356 }
|
Chris@42
|
357 }
|
Chris@42
|
358 {
|
Chris@42
|
359 E T4N, T1O, T4t, T27, T1X, T20, T1Z, T4P, T1U, T4q, T1Y;
|
Chris@42
|
360 {
|
Chris@42
|
361 E T23, T26, T25, T4s, T24;
|
Chris@42
|
362 {
|
Chris@42
|
363 E T1K, T1N, T1J, T1M, T4M, T1L, T22;
|
Chris@42
|
364 T1K = cr[WS(rs, 1)];
|
Chris@42
|
365 T1N = ci[WS(rs, 1)];
|
Chris@42
|
366 T1J = W[0];
|
Chris@42
|
367 T1M = W[1];
|
Chris@42
|
368 T23 = cr[WS(rs, 25)];
|
Chris@42
|
369 T26 = ci[WS(rs, 25)];
|
Chris@42
|
370 T4M = T1J * T1N;
|
Chris@42
|
371 T1L = T1J * T1K;
|
Chris@42
|
372 T22 = W[48];
|
Chris@42
|
373 T25 = W[49];
|
Chris@42
|
374 T4N = FNMS(T1M, T1K, T4M);
|
Chris@42
|
375 T1O = FMA(T1M, T1N, T1L);
|
Chris@42
|
376 T4s = T22 * T26;
|
Chris@42
|
377 T24 = T22 * T23;
|
Chris@42
|
378 }
|
Chris@42
|
379 {
|
Chris@42
|
380 E T1Q, T1T, T1P, T1S, T4O, T1R, T1W;
|
Chris@42
|
381 T1Q = cr[WS(rs, 17)];
|
Chris@42
|
382 T1T = ci[WS(rs, 17)];
|
Chris@42
|
383 T4t = FNMS(T25, T23, T4s);
|
Chris@42
|
384 T27 = FMA(T25, T26, T24);
|
Chris@42
|
385 T1P = W[32];
|
Chris@42
|
386 T1S = W[33];
|
Chris@42
|
387 T1X = cr[WS(rs, 9)];
|
Chris@42
|
388 T20 = ci[WS(rs, 9)];
|
Chris@42
|
389 T4O = T1P * T1T;
|
Chris@42
|
390 T1R = T1P * T1Q;
|
Chris@42
|
391 T1W = W[16];
|
Chris@42
|
392 T1Z = W[17];
|
Chris@42
|
393 T4P = FNMS(T1S, T1Q, T4O);
|
Chris@42
|
394 T1U = FMA(T1S, T1T, T1R);
|
Chris@42
|
395 T4q = T1W * T20;
|
Chris@42
|
396 T1Y = T1W * T1X;
|
Chris@42
|
397 }
|
Chris@42
|
398 }
|
Chris@42
|
399 {
|
Chris@42
|
400 E T4Q, T70, T1V, T4p, T4r, T21;
|
Chris@42
|
401 T4Q = T4N - T4P;
|
Chris@42
|
402 T70 = T4N + T4P;
|
Chris@42
|
403 T1V = T1O + T1U;
|
Chris@42
|
404 T4p = T1O - T1U;
|
Chris@42
|
405 T4r = FNMS(T1Z, T1X, T4q);
|
Chris@42
|
406 T21 = FMA(T1Z, T20, T1Y);
|
Chris@42
|
407 {
|
Chris@42
|
408 E T71, T4u, T4R, T28;
|
Chris@42
|
409 T71 = T4r + T4t;
|
Chris@42
|
410 T4u = T4r - T4t;
|
Chris@42
|
411 T4R = T21 - T27;
|
Chris@42
|
412 T28 = T21 + T27;
|
Chris@42
|
413 T4v = T4p - T4u;
|
Chris@42
|
414 T65 = T4p + T4u;
|
Chris@42
|
415 T72 = T70 - T71;
|
Chris@42
|
416 T7I = T70 + T71;
|
Chris@42
|
417 T29 = T1V + T28;
|
Chris@42
|
418 T6V = T1V - T28;
|
Chris@42
|
419 T68 = T4Q - T4R;
|
Chris@42
|
420 T4S = T4Q + T4R;
|
Chris@42
|
421 }
|
Chris@42
|
422 }
|
Chris@42
|
423 }
|
Chris@42
|
424 {
|
Chris@42
|
425 E T57, T38, T5g, T3r, T3h, T3k, T3j, T59, T3e, T5d, T3i;
|
Chris@42
|
426 {
|
Chris@42
|
427 E T3n, T3q, T3p, T5f, T3o;
|
Chris@42
|
428 {
|
Chris@42
|
429 E T34, T37, T33, T36, T56, T35, T3m;
|
Chris@42
|
430 T34 = cr[WS(rs, 3)];
|
Chris@42
|
431 T37 = ci[WS(rs, 3)];
|
Chris@42
|
432 T33 = W[4];
|
Chris@42
|
433 T36 = W[5];
|
Chris@42
|
434 T3n = cr[WS(rs, 11)];
|
Chris@42
|
435 T3q = ci[WS(rs, 11)];
|
Chris@42
|
436 T56 = T33 * T37;
|
Chris@42
|
437 T35 = T33 * T34;
|
Chris@42
|
438 T3m = W[20];
|
Chris@42
|
439 T3p = W[21];
|
Chris@42
|
440 T57 = FNMS(T36, T34, T56);
|
Chris@42
|
441 T38 = FMA(T36, T37, T35);
|
Chris@42
|
442 T5f = T3m * T3q;
|
Chris@42
|
443 T3o = T3m * T3n;
|
Chris@42
|
444 }
|
Chris@42
|
445 {
|
Chris@42
|
446 E T3a, T3d, T39, T3c, T58, T3b, T3g;
|
Chris@42
|
447 T3a = cr[WS(rs, 19)];
|
Chris@42
|
448 T3d = ci[WS(rs, 19)];
|
Chris@42
|
449 T5g = FNMS(T3p, T3n, T5f);
|
Chris@42
|
450 T3r = FMA(T3p, T3q, T3o);
|
Chris@42
|
451 T39 = W[36];
|
Chris@42
|
452 T3c = W[37];
|
Chris@42
|
453 T3h = cr[WS(rs, 27)];
|
Chris@42
|
454 T3k = ci[WS(rs, 27)];
|
Chris@42
|
455 T58 = T39 * T3d;
|
Chris@42
|
456 T3b = T39 * T3a;
|
Chris@42
|
457 T3g = W[52];
|
Chris@42
|
458 T3j = W[53];
|
Chris@42
|
459 T59 = FNMS(T3c, T3a, T58);
|
Chris@42
|
460 T3e = FMA(T3c, T3d, T3b);
|
Chris@42
|
461 T5d = T3g * T3k;
|
Chris@42
|
462 T3i = T3g * T3h;
|
Chris@42
|
463 }
|
Chris@42
|
464 }
|
Chris@42
|
465 {
|
Chris@42
|
466 E T5a, T78, T3f, T55, T5e, T3l, T77, T3s;
|
Chris@42
|
467 T5a = T57 - T59;
|
Chris@42
|
468 T78 = T57 + T59;
|
Chris@42
|
469 T3f = T38 + T3e;
|
Chris@42
|
470 T55 = T38 - T3e;
|
Chris@42
|
471 T5e = FNMS(T3j, T3h, T5d);
|
Chris@42
|
472 T3l = FMA(T3j, T3k, T3i);
|
Chris@42
|
473 T5h = T5e - T5g;
|
Chris@42
|
474 T77 = T5e + T5g;
|
Chris@42
|
475 T3s = T3l + T3r;
|
Chris@42
|
476 T5c = T3l - T3r;
|
Chris@42
|
477 T5t = T55 + T5a;
|
Chris@42
|
478 T5b = T55 - T5a;
|
Chris@42
|
479 T7O = T78 + T77;
|
Chris@42
|
480 T79 = T77 - T78;
|
Chris@42
|
481 T7e = T3s - T3f;
|
Chris@42
|
482 T3t = T3f + T3s;
|
Chris@42
|
483 }
|
Chris@42
|
484 }
|
Chris@42
|
485 {
|
Chris@42
|
486 E T4y, T2f, T2o, T2r, T4A, T2l, T2n, T2q, T4E, T2p;
|
Chris@42
|
487 {
|
Chris@42
|
488 E T2u, T2x, T2t, T2w;
|
Chris@42
|
489 {
|
Chris@42
|
490 E T2b, T2e, T2d, T4x, T2c, T2a;
|
Chris@42
|
491 T2b = cr[WS(rs, 5)];
|
Chris@42
|
492 T2e = ci[WS(rs, 5)];
|
Chris@42
|
493 T2a = W[8];
|
Chris@42
|
494 T5s = T5c - T5h;
|
Chris@42
|
495 T5i = T5c + T5h;
|
Chris@42
|
496 T2d = W[9];
|
Chris@42
|
497 T4x = T2a * T2e;
|
Chris@42
|
498 T2c = T2a * T2b;
|
Chris@42
|
499 T2u = cr[WS(rs, 13)];
|
Chris@42
|
500 T2x = ci[WS(rs, 13)];
|
Chris@42
|
501 T4y = FNMS(T2d, T2b, T4x);
|
Chris@42
|
502 T2f = FMA(T2d, T2e, T2c);
|
Chris@42
|
503 T2t = W[24];
|
Chris@42
|
504 T2w = W[25];
|
Chris@42
|
505 }
|
Chris@42
|
506 {
|
Chris@42
|
507 E T2h, T2k, T2j, T4z, T2i, T4G, T2v, T2g;
|
Chris@42
|
508 T2h = cr[WS(rs, 21)];
|
Chris@42
|
509 T2k = ci[WS(rs, 21)];
|
Chris@42
|
510 T4G = T2t * T2x;
|
Chris@42
|
511 T2v = T2t * T2u;
|
Chris@42
|
512 T2g = W[40];
|
Chris@42
|
513 T2j = W[41];
|
Chris@42
|
514 T4H = FNMS(T2w, T2u, T4G);
|
Chris@42
|
515 T2y = FMA(T2w, T2x, T2v);
|
Chris@42
|
516 T4z = T2g * T2k;
|
Chris@42
|
517 T2i = T2g * T2h;
|
Chris@42
|
518 T2o = cr[WS(rs, 29)];
|
Chris@42
|
519 T2r = ci[WS(rs, 29)];
|
Chris@42
|
520 T4A = FNMS(T2j, T2h, T4z);
|
Chris@42
|
521 T2l = FMA(T2j, T2k, T2i);
|
Chris@42
|
522 T2n = W[56];
|
Chris@42
|
523 T2q = W[57];
|
Chris@42
|
524 }
|
Chris@42
|
525 }
|
Chris@42
|
526 T4B = T4y - T4A;
|
Chris@42
|
527 T6X = T4y + T4A;
|
Chris@42
|
528 T2m = T2f + T2l;
|
Chris@42
|
529 T4w = T2f - T2l;
|
Chris@42
|
530 T4E = T2n * T2r;
|
Chris@42
|
531 T2p = T2n * T2o;
|
Chris@42
|
532 T4F = FNMS(T2q, T2o, T4E);
|
Chris@42
|
533 T2s = FMA(T2q, T2r, T2p);
|
Chris@42
|
534 }
|
Chris@42
|
535 }
|
Chris@42
|
536 {
|
Chris@42
|
537 E T6E, T8j, T6Y, T73, T6H, T8k, T5S, T8O, T8N, T5V, T6g, T6d, T69, T66, T5O;
|
Chris@42
|
538 E T5R;
|
Chris@42
|
539 {
|
Chris@42
|
540 E T4T, T4C, T4J, T4U, T7S, T7V;
|
Chris@42
|
541 {
|
Chris@42
|
542 E T7C, TO, T80, T7Z, T8e, T89, T8d, T1H, T8b, T3v, T7T, T7L, T7U, T7Q, T2A;
|
Chris@42
|
543 E T7P, T7K, T7W, T1I;
|
Chris@42
|
544 {
|
Chris@42
|
545 E T7X, T7Y, T7J, T82, T88;
|
Chris@42
|
546 {
|
Chris@42
|
547 E Tm, T4I, T6W, T4D, T2z, TN;
|
Chris@42
|
548 T6E = T8 - Tl;
|
Chris@42
|
549 Tm = T8 + Tl;
|
Chris@42
|
550 T4T = T4w + T4B;
|
Chris@42
|
551 T4C = T4w - T4B;
|
Chris@42
|
552 T4I = T4F - T4H;
|
Chris@42
|
553 T6W = T4F + T4H;
|
Chris@42
|
554 T4D = T2s - T2y;
|
Chris@42
|
555 T2z = T2s + T2y;
|
Chris@42
|
556 TN = Tz + TM;
|
Chris@42
|
557 T8j = Tz - TM;
|
Chris@42
|
558 T6Y = T6W - T6X;
|
Chris@42
|
559 T7J = T6X + T6W;
|
Chris@42
|
560 T4J = T4D + T4I;
|
Chris@42
|
561 T4U = T4I - T4D;
|
Chris@42
|
562 T2A = T2m + T2z;
|
Chris@42
|
563 T73 = T2m - T2z;
|
Chris@42
|
564 T7C = Tm - TN;
|
Chris@42
|
565 TO = Tm + TN;
|
Chris@42
|
566 }
|
Chris@42
|
567 T7P = T7N - T7O;
|
Chris@42
|
568 T7X = T7N + T7O;
|
Chris@42
|
569 T7Y = T7I + T7J;
|
Chris@42
|
570 T7K = T7I - T7J;
|
Chris@42
|
571 T6H = T6F - T6G;
|
Chris@42
|
572 T82 = T6G + T6F;
|
Chris@42
|
573 T88 = T83 + T87;
|
Chris@42
|
574 T8k = T87 - T83;
|
Chris@42
|
575 T80 = T7Y + T7X;
|
Chris@42
|
576 T7Z = T7X - T7Y;
|
Chris@42
|
577 T8e = T88 - T82;
|
Chris@42
|
578 T89 = T82 + T88;
|
Chris@42
|
579 }
|
Chris@42
|
580 {
|
Chris@42
|
581 E T7H, T7M, T2B, T3u;
|
Chris@42
|
582 T7H = T29 - T2A;
|
Chris@42
|
583 T2B = T29 + T2A;
|
Chris@42
|
584 T3u = T32 + T3t;
|
Chris@42
|
585 T7M = T32 - T3t;
|
Chris@42
|
586 T8d = T1f - T1G;
|
Chris@42
|
587 T1H = T1f + T1G;
|
Chris@42
|
588 T8b = T3u - T2B;
|
Chris@42
|
589 T3v = T2B + T3u;
|
Chris@42
|
590 T7T = T7H - T7K;
|
Chris@42
|
591 T7L = T7H + T7K;
|
Chris@42
|
592 T7U = T7M + T7P;
|
Chris@42
|
593 T7Q = T7M - T7P;
|
Chris@42
|
594 }
|
Chris@42
|
595 T7W = TO - T1H;
|
Chris@42
|
596 T1I = TO + T1H;
|
Chris@42
|
597 {
|
Chris@42
|
598 E T8g, T8h, T8f, T8i;
|
Chris@42
|
599 {
|
Chris@42
|
600 E T7R, T8c, T8a, T7G, T81, T7F;
|
Chris@42
|
601 T8g = T7Q - T7L;
|
Chris@42
|
602 T7R = T7L + T7Q;
|
Chris@42
|
603 T81 = T7E + T7D;
|
Chris@42
|
604 T7F = T7D - T7E;
|
Chris@42
|
605 cr[0] = T1I + T3v;
|
Chris@42
|
606 ci[WS(rs, 15)] = T1I - T3v;
|
Chris@42
|
607 ci[WS(rs, 7)] = T7W + T7Z;
|
Chris@42
|
608 cr[WS(rs, 8)] = T7W - T7Z;
|
Chris@42
|
609 T8c = T89 - T81;
|
Chris@42
|
610 T8a = T81 + T89;
|
Chris@42
|
611 T7G = T7C - T7F;
|
Chris@42
|
612 T7S = T7C + T7F;
|
Chris@42
|
613 T8h = T8e - T8d;
|
Chris@42
|
614 T8f = T8d + T8e;
|
Chris@42
|
615 ci[WS(rs, 23)] = T8b + T8c;
|
Chris@42
|
616 cr[WS(rs, 24)] = T8b - T8c;
|
Chris@42
|
617 ci[WS(rs, 31)] = T80 + T8a;
|
Chris@42
|
618 cr[WS(rs, 16)] = T80 - T8a;
|
Chris@42
|
619 cr[WS(rs, 4)] = FMA(KP707106781, T7R, T7G);
|
Chris@42
|
620 ci[WS(rs, 11)] = FNMS(KP707106781, T7R, T7G);
|
Chris@42
|
621 }
|
Chris@42
|
622 T8i = T7U - T7T;
|
Chris@42
|
623 T7V = T7T + T7U;
|
Chris@42
|
624 ci[WS(rs, 19)] = FMA(KP707106781, T8g, T8f);
|
Chris@42
|
625 cr[WS(rs, 28)] = FMS(KP707106781, T8g, T8f);
|
Chris@42
|
626 ci[WS(rs, 27)] = FMA(KP707106781, T8i, T8h);
|
Chris@42
|
627 cr[WS(rs, 20)] = FMS(KP707106781, T8i, T8h);
|
Chris@42
|
628 }
|
Chris@42
|
629 }
|
Chris@42
|
630 {
|
Chris@42
|
631 E T5C, T3S, T8C, T4n, T8H, T8B, T8I, T5F, T4L, T5H, T5M, T5Q, T5A, T5w, T4V;
|
Chris@42
|
632 {
|
Chris@42
|
633 E T5D, T47, T4m, T5E, T8z, T8A, T3C, T3R, T5j, T5u;
|
Chris@42
|
634 T5S = T3w + T3B;
|
Chris@42
|
635 T3C = T3w - T3B;
|
Chris@42
|
636 T3R = T3J + T3Q;
|
Chris@42
|
637 T8O = T3Q - T3J;
|
Chris@42
|
638 T5D = FNMS(KP414213562, T3Z, T46);
|
Chris@42
|
639 T47 = FMA(KP414213562, T46, T3Z);
|
Chris@42
|
640 ci[WS(rs, 3)] = FMA(KP707106781, T7V, T7S);
|
Chris@42
|
641 cr[WS(rs, 12)] = FNMS(KP707106781, T7V, T7S);
|
Chris@42
|
642 T5C = FMA(KP707106781, T3R, T3C);
|
Chris@42
|
643 T3S = FNMS(KP707106781, T3R, T3C);
|
Chris@42
|
644 T4m = FNMS(KP414213562, T4l, T4e);
|
Chris@42
|
645 T5E = FMA(KP414213562, T4e, T4l);
|
Chris@42
|
646 T8N = T8y - T8x;
|
Chris@42
|
647 T8z = T8x + T8y;
|
Chris@42
|
648 T8A = T5T - T5U;
|
Chris@42
|
649 T5V = T5T + T5U;
|
Chris@42
|
650 T8C = T47 + T4m;
|
Chris@42
|
651 T4n = T47 - T4m;
|
Chris@42
|
652 T8H = FNMS(KP707106781, T8A, T8z);
|
Chris@42
|
653 T8B = FMA(KP707106781, T8A, T8z);
|
Chris@42
|
654 T6g = T5i - T5b;
|
Chris@42
|
655 T5j = T5b + T5i;
|
Chris@42
|
656 T5u = T5s - T5t;
|
Chris@42
|
657 T6d = T5t + T5s;
|
Chris@42
|
658 {
|
Chris@42
|
659 E T5K, T5k, T5L, T5v, T4K;
|
Chris@42
|
660 T69 = T4J - T4C;
|
Chris@42
|
661 T4K = T4C + T4J;
|
Chris@42
|
662 T8I = T5E - T5D;
|
Chris@42
|
663 T5F = T5D + T5E;
|
Chris@42
|
664 T5K = FMA(KP707106781, T5j, T54);
|
Chris@42
|
665 T5k = FNMS(KP707106781, T5j, T54);
|
Chris@42
|
666 T5L = FMA(KP707106781, T5u, T5r);
|
Chris@42
|
667 T5v = FNMS(KP707106781, T5u, T5r);
|
Chris@42
|
668 T4L = FNMS(KP707106781, T4K, T4v);
|
Chris@42
|
669 T5H = FMA(KP707106781, T4K, T4v);
|
Chris@42
|
670 T5M = FNMS(KP198912367, T5L, T5K);
|
Chris@42
|
671 T5Q = FMA(KP198912367, T5K, T5L);
|
Chris@42
|
672 T5A = FNMS(KP668178637, T5k, T5v);
|
Chris@42
|
673 T5w = FMA(KP668178637, T5v, T5k);
|
Chris@42
|
674 T4V = T4T + T4U;
|
Chris@42
|
675 T66 = T4T - T4U;
|
Chris@42
|
676 }
|
Chris@42
|
677 }
|
Chris@42
|
678 {
|
Chris@42
|
679 E T5y, T4o, T8J, T8L, T5I, T4W;
|
Chris@42
|
680 T5y = FNMS(KP923879532, T4n, T3S);
|
Chris@42
|
681 T4o = FMA(KP923879532, T4n, T3S);
|
Chris@42
|
682 T8J = FMA(KP923879532, T8I, T8H);
|
Chris@42
|
683 T8L = FNMS(KP923879532, T8I, T8H);
|
Chris@42
|
684 T5I = FMA(KP707106781, T4V, T4S);
|
Chris@42
|
685 T4W = FNMS(KP707106781, T4V, T4S);
|
Chris@42
|
686 {
|
Chris@42
|
687 E T8G, T8F, T8D, T8E;
|
Chris@42
|
688 {
|
Chris@42
|
689 E T5G, T5P, T5z, T4X, T5N, T5J;
|
Chris@42
|
690 T5O = FNMS(KP923879532, T5F, T5C);
|
Chris@42
|
691 T5G = FMA(KP923879532, T5F, T5C);
|
Chris@42
|
692 T5J = FNMS(KP198912367, T5I, T5H);
|
Chris@42
|
693 T5P = FMA(KP198912367, T5H, T5I);
|
Chris@42
|
694 T5z = FNMS(KP668178637, T4L, T4W);
|
Chris@42
|
695 T4X = FMA(KP668178637, T4W, T4L);
|
Chris@42
|
696 T5N = T5J + T5M;
|
Chris@42
|
697 T8G = T5M - T5J;
|
Chris@42
|
698 T8F = FNMS(KP923879532, T8C, T8B);
|
Chris@42
|
699 T8D = FMA(KP923879532, T8C, T8B);
|
Chris@42
|
700 {
|
Chris@42
|
701 E T5B, T8K, T8M, T5x;
|
Chris@42
|
702 T5B = T5z + T5A;
|
Chris@42
|
703 T8K = T5z - T5A;
|
Chris@42
|
704 T8M = T5w - T4X;
|
Chris@42
|
705 T5x = T4X + T5w;
|
Chris@42
|
706 ci[0] = FMA(KP980785280, T5N, T5G);
|
Chris@42
|
707 cr[WS(rs, 15)] = FNMS(KP980785280, T5N, T5G);
|
Chris@42
|
708 ci[WS(rs, 4)] = FNMS(KP831469612, T5B, T5y);
|
Chris@42
|
709 cr[WS(rs, 11)] = FMA(KP831469612, T5B, T5y);
|
Chris@42
|
710 ci[WS(rs, 28)] = FMA(KP831469612, T8K, T8J);
|
Chris@42
|
711 cr[WS(rs, 19)] = FMS(KP831469612, T8K, T8J);
|
Chris@42
|
712 ci[WS(rs, 20)] = FMA(KP831469612, T8M, T8L);
|
Chris@42
|
713 cr[WS(rs, 27)] = FMS(KP831469612, T8M, T8L);
|
Chris@42
|
714 cr[WS(rs, 3)] = FMA(KP831469612, T5x, T4o);
|
Chris@42
|
715 ci[WS(rs, 12)] = FNMS(KP831469612, T5x, T4o);
|
Chris@42
|
716 T8E = T5Q - T5P;
|
Chris@42
|
717 T5R = T5P + T5Q;
|
Chris@42
|
718 }
|
Chris@42
|
719 }
|
Chris@42
|
720 ci[WS(rs, 16)] = FMA(KP980785280, T8E, T8D);
|
Chris@42
|
721 cr[WS(rs, 31)] = FMS(KP980785280, T8E, T8D);
|
Chris@42
|
722 ci[WS(rs, 24)] = FMA(KP980785280, T8G, T8F);
|
Chris@42
|
723 cr[WS(rs, 23)] = FMS(KP980785280, T8G, T8F);
|
Chris@42
|
724 }
|
Chris@42
|
725 }
|
Chris@42
|
726 }
|
Chris@42
|
727 }
|
Chris@42
|
728 {
|
Chris@42
|
729 E T7y, T8q, T8p, T7B;
|
Chris@42
|
730 {
|
Chris@42
|
731 E T7a, T7m, T6I, T7f, T7A, T7w, T8r, T8l, T8m, T6T, T7k, T75, T8s, T7p, T7z;
|
Chris@42
|
732 E T7t;
|
Chris@42
|
733 {
|
Chris@42
|
734 E T7n, T6N, T6S, T7o, T7u, T7v;
|
Chris@42
|
735 T7a = T76 - T79;
|
Chris@42
|
736 T7u = T76 + T79;
|
Chris@42
|
737 cr[WS(rs, 7)] = FMA(KP980785280, T5R, T5O);
|
Chris@42
|
738 ci[WS(rs, 8)] = FNMS(KP980785280, T5R, T5O);
|
Chris@42
|
739 T7m = T6E + T6H;
|
Chris@42
|
740 T6I = T6E - T6H;
|
Chris@42
|
741 T7v = T7e - T7d;
|
Chris@42
|
742 T7f = T7d + T7e;
|
Chris@42
|
743 T7n = T6J - T6M;
|
Chris@42
|
744 T6N = T6J + T6M;
|
Chris@42
|
745 T7A = FMA(KP414213562, T7u, T7v);
|
Chris@42
|
746 T7w = FNMS(KP414213562, T7v, T7u);
|
Chris@42
|
747 T8r = T8k - T8j;
|
Chris@42
|
748 T8l = T8j + T8k;
|
Chris@42
|
749 T6S = T6O - T6R;
|
Chris@42
|
750 T7o = T6O + T6R;
|
Chris@42
|
751 {
|
Chris@42
|
752 E T7r, T7s, T6Z, T74;
|
Chris@42
|
753 T7r = T6V + T6Y;
|
Chris@42
|
754 T6Z = T6V - T6Y;
|
Chris@42
|
755 T74 = T72 - T73;
|
Chris@42
|
756 T7s = T72 + T73;
|
Chris@42
|
757 T8m = T6N - T6S;
|
Chris@42
|
758 T6T = T6N + T6S;
|
Chris@42
|
759 T7k = FNMS(KP414213562, T6Z, T74);
|
Chris@42
|
760 T75 = FMA(KP414213562, T74, T6Z);
|
Chris@42
|
761 T8s = T7o - T7n;
|
Chris@42
|
762 T7p = T7n + T7o;
|
Chris@42
|
763 T7z = FMA(KP414213562, T7r, T7s);
|
Chris@42
|
764 T7t = FNMS(KP414213562, T7s, T7r);
|
Chris@42
|
765 }
|
Chris@42
|
766 }
|
Chris@42
|
767 {
|
Chris@42
|
768 E T7i, T6U, T8t, T8v, T7j, T7g;
|
Chris@42
|
769 T7i = FNMS(KP707106781, T6T, T6I);
|
Chris@42
|
770 T6U = FMA(KP707106781, T6T, T6I);
|
Chris@42
|
771 T8t = FMA(KP707106781, T8s, T8r);
|
Chris@42
|
772 T8v = FNMS(KP707106781, T8s, T8r);
|
Chris@42
|
773 T7j = FMA(KP414213562, T7a, T7f);
|
Chris@42
|
774 T7g = FNMS(KP414213562, T7f, T7a);
|
Chris@42
|
775 {
|
Chris@42
|
776 E T7q, T7x, T8n, T8o;
|
Chris@42
|
777 T7y = FNMS(KP707106781, T7p, T7m);
|
Chris@42
|
778 T7q = FMA(KP707106781, T7p, T7m);
|
Chris@42
|
779 {
|
Chris@42
|
780 E T7l, T8u, T8w, T7h;
|
Chris@42
|
781 T7l = T7j - T7k;
|
Chris@42
|
782 T8u = T7k + T7j;
|
Chris@42
|
783 T8w = T7g - T75;
|
Chris@42
|
784 T7h = T75 + T7g;
|
Chris@42
|
785 ci[WS(rs, 5)] = FMA(KP923879532, T7l, T7i);
|
Chris@42
|
786 cr[WS(rs, 10)] = FNMS(KP923879532, T7l, T7i);
|
Chris@42
|
787 ci[WS(rs, 29)] = FMA(KP923879532, T8u, T8t);
|
Chris@42
|
788 cr[WS(rs, 18)] = FMS(KP923879532, T8u, T8t);
|
Chris@42
|
789 ci[WS(rs, 21)] = FMA(KP923879532, T8w, T8v);
|
Chris@42
|
790 cr[WS(rs, 26)] = FMS(KP923879532, T8w, T8v);
|
Chris@42
|
791 cr[WS(rs, 2)] = FMA(KP923879532, T7h, T6U);
|
Chris@42
|
792 ci[WS(rs, 13)] = FNMS(KP923879532, T7h, T6U);
|
Chris@42
|
793 T7x = T7t + T7w;
|
Chris@42
|
794 T8q = T7w - T7t;
|
Chris@42
|
795 }
|
Chris@42
|
796 T8p = FNMS(KP707106781, T8m, T8l);
|
Chris@42
|
797 T8n = FMA(KP707106781, T8m, T8l);
|
Chris@42
|
798 T8o = T7A - T7z;
|
Chris@42
|
799 T7B = T7z + T7A;
|
Chris@42
|
800 ci[WS(rs, 1)] = FMA(KP923879532, T7x, T7q);
|
Chris@42
|
801 cr[WS(rs, 14)] = FNMS(KP923879532, T7x, T7q);
|
Chris@42
|
802 ci[WS(rs, 17)] = FMA(KP923879532, T8o, T8n);
|
Chris@42
|
803 cr[WS(rs, 30)] = FMS(KP923879532, T8o, T8n);
|
Chris@42
|
804 }
|
Chris@42
|
805 }
|
Chris@42
|
806 }
|
Chris@42
|
807 {
|
Chris@42
|
808 E T6o, T5W, T8W, T63, T8V, T8P, T8Q, T6r, T6e, T6w;
|
Chris@42
|
809 {
|
Chris@42
|
810 E T6q, T6p, T5Z, T62;
|
Chris@42
|
811 ci[WS(rs, 25)] = FMA(KP923879532, T8q, T8p);
|
Chris@42
|
812 cr[WS(rs, 22)] = FMS(KP923879532, T8q, T8p);
|
Chris@42
|
813 cr[WS(rs, 6)] = FMA(KP923879532, T7B, T7y);
|
Chris@42
|
814 ci[WS(rs, 9)] = FNMS(KP923879532, T7B, T7y);
|
Chris@42
|
815 T6q = FNMS(KP414213562, T5X, T5Y);
|
Chris@42
|
816 T5Z = FMA(KP414213562, T5Y, T5X);
|
Chris@42
|
817 T62 = FNMS(KP414213562, T61, T60);
|
Chris@42
|
818 T6p = FMA(KP414213562, T60, T61);
|
Chris@42
|
819 T6o = FNMS(KP707106781, T5V, T5S);
|
Chris@42
|
820 T5W = FMA(KP707106781, T5V, T5S);
|
Chris@42
|
821 T8W = T5Z - T62;
|
Chris@42
|
822 T63 = T5Z + T62;
|
Chris@42
|
823 T8V = FNMS(KP707106781, T8O, T8N);
|
Chris@42
|
824 T8P = FMA(KP707106781, T8O, T8N);
|
Chris@42
|
825 T8Q = T6q + T6p;
|
Chris@42
|
826 T6r = T6p - T6q;
|
Chris@42
|
827 T6e = FMA(KP707106781, T6d, T6c);
|
Chris@42
|
828 T6w = FNMS(KP707106781, T6d, T6c);
|
Chris@42
|
829 }
|
Chris@42
|
830 {
|
Chris@42
|
831 E T6k, T8U, T6z, T6n, T8S, T8T, T8R, T6s;
|
Chris@42
|
832 {
|
Chris@42
|
833 E T64, T6y, T6l, T6i, T6v, T6m, T6b, T8X, T8Z, T8Y, T6j, T90;
|
Chris@42
|
834 {
|
Chris@42
|
835 E T6C, T6B, T6x, T6h;
|
Chris@42
|
836 T6k = FNMS(KP923879532, T63, T5W);
|
Chris@42
|
837 T64 = FMA(KP923879532, T63, T5W);
|
Chris@42
|
838 T6x = FNMS(KP707106781, T6g, T6f);
|
Chris@42
|
839 T6h = FMA(KP707106781, T6g, T6f);
|
Chris@42
|
840 {
|
Chris@42
|
841 E T6t, T67, T6u, T6a;
|
Chris@42
|
842 T6t = FNMS(KP707106781, T66, T65);
|
Chris@42
|
843 T67 = FMA(KP707106781, T66, T65);
|
Chris@42
|
844 T6u = FNMS(KP707106781, T69, T68);
|
Chris@42
|
845 T6a = FMA(KP707106781, T69, T68);
|
Chris@42
|
846 T6y = FMA(KP668178637, T6x, T6w);
|
Chris@42
|
847 T6C = FNMS(KP668178637, T6w, T6x);
|
Chris@42
|
848 T6l = FMA(KP198912367, T6e, T6h);
|
Chris@42
|
849 T6i = FNMS(KP198912367, T6h, T6e);
|
Chris@42
|
850 T6v = FNMS(KP668178637, T6u, T6t);
|
Chris@42
|
851 T6B = FMA(KP668178637, T6t, T6u);
|
Chris@42
|
852 T6m = FNMS(KP198912367, T67, T6a);
|
Chris@42
|
853 T6b = FMA(KP198912367, T6a, T67);
|
Chris@42
|
854 }
|
Chris@42
|
855 T8X = FMA(KP923879532, T8W, T8V);
|
Chris@42
|
856 T8Z = FNMS(KP923879532, T8W, T8V);
|
Chris@42
|
857 T6D = T6B - T6C;
|
Chris@42
|
858 T8Y = T6B + T6C;
|
Chris@42
|
859 }
|
Chris@42
|
860 T8U = T6i - T6b;
|
Chris@42
|
861 T6j = T6b + T6i;
|
Chris@42
|
862 T90 = T6y - T6v;
|
Chris@42
|
863 T6z = T6v + T6y;
|
Chris@42
|
864 ci[WS(rs, 18)] = FNMS(KP831469612, T8Y, T8X);
|
Chris@42
|
865 cr[WS(rs, 29)] = -(FMA(KP831469612, T8Y, T8X));
|
Chris@42
|
866 cr[WS(rs, 1)] = FMA(KP980785280, T6j, T64);
|
Chris@42
|
867 ci[WS(rs, 14)] = FNMS(KP980785280, T6j, T64);
|
Chris@42
|
868 cr[WS(rs, 21)] = FMS(KP831469612, T90, T8Z);
|
Chris@42
|
869 ci[WS(rs, 26)] = FMA(KP831469612, T90, T8Z);
|
Chris@42
|
870 T6n = T6l - T6m;
|
Chris@42
|
871 T8S = T6m + T6l;
|
Chris@42
|
872 }
|
Chris@42
|
873 T6A = FNMS(KP923879532, T6r, T6o);
|
Chris@42
|
874 T6s = FMA(KP923879532, T6r, T6o);
|
Chris@42
|
875 T8T = FNMS(KP923879532, T8Q, T8P);
|
Chris@42
|
876 T8R = FMA(KP923879532, T8Q, T8P);
|
Chris@42
|
877 ci[WS(rs, 6)] = FMA(KP980785280, T6n, T6k);
|
Chris@42
|
878 cr[WS(rs, 9)] = FNMS(KP980785280, T6n, T6k);
|
Chris@42
|
879 ci[WS(rs, 2)] = FMA(KP831469612, T6z, T6s);
|
Chris@42
|
880 cr[WS(rs, 13)] = FNMS(KP831469612, T6z, T6s);
|
Chris@42
|
881 ci[WS(rs, 30)] = FMA(KP980785280, T8S, T8R);
|
Chris@42
|
882 cr[WS(rs, 17)] = FMS(KP980785280, T8S, T8R);
|
Chris@42
|
883 ci[WS(rs, 22)] = FMA(KP980785280, T8U, T8T);
|
Chris@42
|
884 cr[WS(rs, 25)] = FMS(KP980785280, T8U, T8T);
|
Chris@42
|
885 }
|
Chris@42
|
886 }
|
Chris@42
|
887 }
|
Chris@42
|
888 }
|
Chris@42
|
889 }
|
Chris@42
|
890 cr[WS(rs, 5)] = FMA(KP831469612, T6D, T6A);
|
Chris@42
|
891 ci[WS(rs, 10)] = FNMS(KP831469612, T6D, T6A);
|
Chris@42
|
892 }
|
Chris@42
|
893 }
|
Chris@42
|
894 }
|
Chris@42
|
895
|
Chris@42
|
896 static const tw_instr twinstr[] = {
|
Chris@42
|
897 {TW_FULL, 1, 32},
|
Chris@42
|
898 {TW_NEXT, 1, 0}
|
Chris@42
|
899 };
|
Chris@42
|
900
|
Chris@42
|
901 static const hc2hc_desc desc = { 32, "hf_32", twinstr, &GENUS, {236, 62, 198, 0} };
|
Chris@42
|
902
|
Chris@42
|
903 void X(codelet_hf_32) (planner *p) {
|
Chris@42
|
904 X(khc2hc_register) (p, hf_32, &desc);
|
Chris@42
|
905 }
|
Chris@42
|
906 #else /* HAVE_FMA */
|
Chris@42
|
907
|
Chris@42
|
908 /* Generated by: ../../../genfft/gen_hc2hc.native -compact -variables 4 -pipeline-latency 4 -n 32 -dit -name hf_32 -include hf.h */
|
Chris@42
|
909
|
Chris@42
|
910 /*
|
Chris@42
|
911 * This function contains 434 FP additions, 208 FP multiplications,
|
Chris@42
|
912 * (or, 340 additions, 114 multiplications, 94 fused multiply/add),
|
Chris@42
|
913 * 96 stack variables, 7 constants, and 128 memory accesses
|
Chris@42
|
914 */
|
Chris@42
|
915 #include "hf.h"
|
Chris@42
|
916
|
Chris@42
|
917 static void hf_32(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
|
Chris@42
|
918 {
|
Chris@42
|
919 DK(KP555570233, +0.555570233019602224742830813948532874374937191);
|
Chris@42
|
920 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
Chris@42
|
921 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
Chris@42
|
922 DK(KP195090322, +0.195090322016128267848284868477022240927691618);
|
Chris@42
|
923 DK(KP382683432, +0.382683432365089771728459984030398866761344562);
|
Chris@42
|
924 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
Chris@42
|
925 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
Chris@42
|
926 {
|
Chris@42
|
927 INT m;
|
Chris@42
|
928 for (m = mb, W = W + ((mb - 1) * 62); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 62, MAKE_VOLATILE_STRIDE(64, rs)) {
|
Chris@42
|
929 E Tj, T5F, T7C, T7Q, T35, T4T, T78, T7m, T1Q, T61, T5Y, T6J, T3K, T56, T41;
|
Chris@42
|
930 E T59, T2B, T67, T6e, T6O, T4b, T5g, T4s, T5d, TG, T7l, T5I, T73, T3a, T4U;
|
Chris@42
|
931 E T3f, T4V, T14, T5K, T5N, T6F, T3m, T4Z, T3r, T4Y, T1r, T5P, T5S, T6E, T3x;
|
Chris@42
|
932 E T52, T3C, T51, T2d, T5Z, T64, T6K, T3V, T5a, T44, T57, T2Y, T6f, T6a, T6P;
|
Chris@42
|
933 E T4m, T5e, T4v, T5h;
|
Chris@42
|
934 {
|
Chris@42
|
935 E T1, T76, T6, T75, Tc, T32, Th, T33;
|
Chris@42
|
936 T1 = cr[0];
|
Chris@42
|
937 T76 = ci[0];
|
Chris@42
|
938 {
|
Chris@42
|
939 E T3, T5, T2, T4;
|
Chris@42
|
940 T3 = cr[WS(rs, 16)];
|
Chris@42
|
941 T5 = ci[WS(rs, 16)];
|
Chris@42
|
942 T2 = W[30];
|
Chris@42
|
943 T4 = W[31];
|
Chris@42
|
944 T6 = FMA(T2, T3, T4 * T5);
|
Chris@42
|
945 T75 = FNMS(T4, T3, T2 * T5);
|
Chris@42
|
946 }
|
Chris@42
|
947 {
|
Chris@42
|
948 E T9, Tb, T8, Ta;
|
Chris@42
|
949 T9 = cr[WS(rs, 8)];
|
Chris@42
|
950 Tb = ci[WS(rs, 8)];
|
Chris@42
|
951 T8 = W[14];
|
Chris@42
|
952 Ta = W[15];
|
Chris@42
|
953 Tc = FMA(T8, T9, Ta * Tb);
|
Chris@42
|
954 T32 = FNMS(Ta, T9, T8 * Tb);
|
Chris@42
|
955 }
|
Chris@42
|
956 {
|
Chris@42
|
957 E Te, Tg, Td, Tf;
|
Chris@42
|
958 Te = cr[WS(rs, 24)];
|
Chris@42
|
959 Tg = ci[WS(rs, 24)];
|
Chris@42
|
960 Td = W[46];
|
Chris@42
|
961 Tf = W[47];
|
Chris@42
|
962 Th = FMA(Td, Te, Tf * Tg);
|
Chris@42
|
963 T33 = FNMS(Tf, Te, Td * Tg);
|
Chris@42
|
964 }
|
Chris@42
|
965 {
|
Chris@42
|
966 E T7, Ti, T7A, T7B;
|
Chris@42
|
967 T7 = T1 + T6;
|
Chris@42
|
968 Ti = Tc + Th;
|
Chris@42
|
969 Tj = T7 + Ti;
|
Chris@42
|
970 T5F = T7 - Ti;
|
Chris@42
|
971 T7A = Tc - Th;
|
Chris@42
|
972 T7B = T76 - T75;
|
Chris@42
|
973 T7C = T7A + T7B;
|
Chris@42
|
974 T7Q = T7B - T7A;
|
Chris@42
|
975 }
|
Chris@42
|
976 {
|
Chris@42
|
977 E T31, T34, T74, T77;
|
Chris@42
|
978 T31 = T1 - T6;
|
Chris@42
|
979 T34 = T32 - T33;
|
Chris@42
|
980 T35 = T31 + T34;
|
Chris@42
|
981 T4T = T31 - T34;
|
Chris@42
|
982 T74 = T32 + T33;
|
Chris@42
|
983 T77 = T75 + T76;
|
Chris@42
|
984 T78 = T74 + T77;
|
Chris@42
|
985 T7m = T77 - T74;
|
Chris@42
|
986 }
|
Chris@42
|
987 }
|
Chris@42
|
988 {
|
Chris@42
|
989 E T1y, T3X, T1O, T3I, T1D, T3Y, T1J, T3H;
|
Chris@42
|
990 {
|
Chris@42
|
991 E T1v, T1x, T1u, T1w;
|
Chris@42
|
992 T1v = cr[WS(rs, 1)];
|
Chris@42
|
993 T1x = ci[WS(rs, 1)];
|
Chris@42
|
994 T1u = W[0];
|
Chris@42
|
995 T1w = W[1];
|
Chris@42
|
996 T1y = FMA(T1u, T1v, T1w * T1x);
|
Chris@42
|
997 T3X = FNMS(T1w, T1v, T1u * T1x);
|
Chris@42
|
998 }
|
Chris@42
|
999 {
|
Chris@42
|
1000 E T1L, T1N, T1K, T1M;
|
Chris@42
|
1001 T1L = cr[WS(rs, 25)];
|
Chris@42
|
1002 T1N = ci[WS(rs, 25)];
|
Chris@42
|
1003 T1K = W[48];
|
Chris@42
|
1004 T1M = W[49];
|
Chris@42
|
1005 T1O = FMA(T1K, T1L, T1M * T1N);
|
Chris@42
|
1006 T3I = FNMS(T1M, T1L, T1K * T1N);
|
Chris@42
|
1007 }
|
Chris@42
|
1008 {
|
Chris@42
|
1009 E T1A, T1C, T1z, T1B;
|
Chris@42
|
1010 T1A = cr[WS(rs, 17)];
|
Chris@42
|
1011 T1C = ci[WS(rs, 17)];
|
Chris@42
|
1012 T1z = W[32];
|
Chris@42
|
1013 T1B = W[33];
|
Chris@42
|
1014 T1D = FMA(T1z, T1A, T1B * T1C);
|
Chris@42
|
1015 T3Y = FNMS(T1B, T1A, T1z * T1C);
|
Chris@42
|
1016 }
|
Chris@42
|
1017 {
|
Chris@42
|
1018 E T1G, T1I, T1F, T1H;
|
Chris@42
|
1019 T1G = cr[WS(rs, 9)];
|
Chris@42
|
1020 T1I = ci[WS(rs, 9)];
|
Chris@42
|
1021 T1F = W[16];
|
Chris@42
|
1022 T1H = W[17];
|
Chris@42
|
1023 T1J = FMA(T1F, T1G, T1H * T1I);
|
Chris@42
|
1024 T3H = FNMS(T1H, T1G, T1F * T1I);
|
Chris@42
|
1025 }
|
Chris@42
|
1026 {
|
Chris@42
|
1027 E T1E, T1P, T5W, T5X;
|
Chris@42
|
1028 T1E = T1y + T1D;
|
Chris@42
|
1029 T1P = T1J + T1O;
|
Chris@42
|
1030 T1Q = T1E + T1P;
|
Chris@42
|
1031 T61 = T1E - T1P;
|
Chris@42
|
1032 T5W = T3X + T3Y;
|
Chris@42
|
1033 T5X = T3H + T3I;
|
Chris@42
|
1034 T5Y = T5W - T5X;
|
Chris@42
|
1035 T6J = T5W + T5X;
|
Chris@42
|
1036 }
|
Chris@42
|
1037 {
|
Chris@42
|
1038 E T3G, T3J, T3Z, T40;
|
Chris@42
|
1039 T3G = T1y - T1D;
|
Chris@42
|
1040 T3J = T3H - T3I;
|
Chris@42
|
1041 T3K = T3G + T3J;
|
Chris@42
|
1042 T56 = T3G - T3J;
|
Chris@42
|
1043 T3Z = T3X - T3Y;
|
Chris@42
|
1044 T40 = T1J - T1O;
|
Chris@42
|
1045 T41 = T3Z - T40;
|
Chris@42
|
1046 T59 = T3Z + T40;
|
Chris@42
|
1047 }
|
Chris@42
|
1048 }
|
Chris@42
|
1049 {
|
Chris@42
|
1050 E T2j, T47, T2z, T4q, T2o, T48, T2u, T4p;
|
Chris@42
|
1051 {
|
Chris@42
|
1052 E T2g, T2i, T2f, T2h;
|
Chris@42
|
1053 T2g = cr[WS(rs, 31)];
|
Chris@42
|
1054 T2i = ci[WS(rs, 31)];
|
Chris@42
|
1055 T2f = W[60];
|
Chris@42
|
1056 T2h = W[61];
|
Chris@42
|
1057 T2j = FMA(T2f, T2g, T2h * T2i);
|
Chris@42
|
1058 T47 = FNMS(T2h, T2g, T2f * T2i);
|
Chris@42
|
1059 }
|
Chris@42
|
1060 {
|
Chris@42
|
1061 E T2w, T2y, T2v, T2x;
|
Chris@42
|
1062 T2w = cr[WS(rs, 23)];
|
Chris@42
|
1063 T2y = ci[WS(rs, 23)];
|
Chris@42
|
1064 T2v = W[44];
|
Chris@42
|
1065 T2x = W[45];
|
Chris@42
|
1066 T2z = FMA(T2v, T2w, T2x * T2y);
|
Chris@42
|
1067 T4q = FNMS(T2x, T2w, T2v * T2y);
|
Chris@42
|
1068 }
|
Chris@42
|
1069 {
|
Chris@42
|
1070 E T2l, T2n, T2k, T2m;
|
Chris@42
|
1071 T2l = cr[WS(rs, 15)];
|
Chris@42
|
1072 T2n = ci[WS(rs, 15)];
|
Chris@42
|
1073 T2k = W[28];
|
Chris@42
|
1074 T2m = W[29];
|
Chris@42
|
1075 T2o = FMA(T2k, T2l, T2m * T2n);
|
Chris@42
|
1076 T48 = FNMS(T2m, T2l, T2k * T2n);
|
Chris@42
|
1077 }
|
Chris@42
|
1078 {
|
Chris@42
|
1079 E T2r, T2t, T2q, T2s;
|
Chris@42
|
1080 T2r = cr[WS(rs, 7)];
|
Chris@42
|
1081 T2t = ci[WS(rs, 7)];
|
Chris@42
|
1082 T2q = W[12];
|
Chris@42
|
1083 T2s = W[13];
|
Chris@42
|
1084 T2u = FMA(T2q, T2r, T2s * T2t);
|
Chris@42
|
1085 T4p = FNMS(T2s, T2r, T2q * T2t);
|
Chris@42
|
1086 }
|
Chris@42
|
1087 {
|
Chris@42
|
1088 E T2p, T2A, T6c, T6d;
|
Chris@42
|
1089 T2p = T2j + T2o;
|
Chris@42
|
1090 T2A = T2u + T2z;
|
Chris@42
|
1091 T2B = T2p + T2A;
|
Chris@42
|
1092 T67 = T2p - T2A;
|
Chris@42
|
1093 T6c = T47 + T48;
|
Chris@42
|
1094 T6d = T4p + T4q;
|
Chris@42
|
1095 T6e = T6c - T6d;
|
Chris@42
|
1096 T6O = T6c + T6d;
|
Chris@42
|
1097 }
|
Chris@42
|
1098 {
|
Chris@42
|
1099 E T49, T4a, T4o, T4r;
|
Chris@42
|
1100 T49 = T47 - T48;
|
Chris@42
|
1101 T4a = T2u - T2z;
|
Chris@42
|
1102 T4b = T49 - T4a;
|
Chris@42
|
1103 T5g = T49 + T4a;
|
Chris@42
|
1104 T4o = T2j - T2o;
|
Chris@42
|
1105 T4r = T4p - T4q;
|
Chris@42
|
1106 T4s = T4o + T4r;
|
Chris@42
|
1107 T5d = T4o - T4r;
|
Chris@42
|
1108 }
|
Chris@42
|
1109 }
|
Chris@42
|
1110 {
|
Chris@42
|
1111 E To, T37, TE, T3d, Tt, T38, Tz, T3c;
|
Chris@42
|
1112 {
|
Chris@42
|
1113 E Tl, Tn, Tk, Tm;
|
Chris@42
|
1114 Tl = cr[WS(rs, 4)];
|
Chris@42
|
1115 Tn = ci[WS(rs, 4)];
|
Chris@42
|
1116 Tk = W[6];
|
Chris@42
|
1117 Tm = W[7];
|
Chris@42
|
1118 To = FMA(Tk, Tl, Tm * Tn);
|
Chris@42
|
1119 T37 = FNMS(Tm, Tl, Tk * Tn);
|
Chris@42
|
1120 }
|
Chris@42
|
1121 {
|
Chris@42
|
1122 E TB, TD, TA, TC;
|
Chris@42
|
1123 TB = cr[WS(rs, 12)];
|
Chris@42
|
1124 TD = ci[WS(rs, 12)];
|
Chris@42
|
1125 TA = W[22];
|
Chris@42
|
1126 TC = W[23];
|
Chris@42
|
1127 TE = FMA(TA, TB, TC * TD);
|
Chris@42
|
1128 T3d = FNMS(TC, TB, TA * TD);
|
Chris@42
|
1129 }
|
Chris@42
|
1130 {
|
Chris@42
|
1131 E Tq, Ts, Tp, Tr;
|
Chris@42
|
1132 Tq = cr[WS(rs, 20)];
|
Chris@42
|
1133 Ts = ci[WS(rs, 20)];
|
Chris@42
|
1134 Tp = W[38];
|
Chris@42
|
1135 Tr = W[39];
|
Chris@42
|
1136 Tt = FMA(Tp, Tq, Tr * Ts);
|
Chris@42
|
1137 T38 = FNMS(Tr, Tq, Tp * Ts);
|
Chris@42
|
1138 }
|
Chris@42
|
1139 {
|
Chris@42
|
1140 E Tw, Ty, Tv, Tx;
|
Chris@42
|
1141 Tw = cr[WS(rs, 28)];
|
Chris@42
|
1142 Ty = ci[WS(rs, 28)];
|
Chris@42
|
1143 Tv = W[54];
|
Chris@42
|
1144 Tx = W[55];
|
Chris@42
|
1145 Tz = FMA(Tv, Tw, Tx * Ty);
|
Chris@42
|
1146 T3c = FNMS(Tx, Tw, Tv * Ty);
|
Chris@42
|
1147 }
|
Chris@42
|
1148 {
|
Chris@42
|
1149 E Tu, TF, T5G, T5H;
|
Chris@42
|
1150 Tu = To + Tt;
|
Chris@42
|
1151 TF = Tz + TE;
|
Chris@42
|
1152 TG = Tu + TF;
|
Chris@42
|
1153 T7l = Tu - TF;
|
Chris@42
|
1154 T5G = T3c + T3d;
|
Chris@42
|
1155 T5H = T37 + T38;
|
Chris@42
|
1156 T5I = T5G - T5H;
|
Chris@42
|
1157 T73 = T5H + T5G;
|
Chris@42
|
1158 }
|
Chris@42
|
1159 {
|
Chris@42
|
1160 E T36, T39, T3b, T3e;
|
Chris@42
|
1161 T36 = To - Tt;
|
Chris@42
|
1162 T39 = T37 - T38;
|
Chris@42
|
1163 T3a = T36 + T39;
|
Chris@42
|
1164 T4U = T36 - T39;
|
Chris@42
|
1165 T3b = Tz - TE;
|
Chris@42
|
1166 T3e = T3c - T3d;
|
Chris@42
|
1167 T3f = T3b - T3e;
|
Chris@42
|
1168 T4V = T3b + T3e;
|
Chris@42
|
1169 }
|
Chris@42
|
1170 }
|
Chris@42
|
1171 {
|
Chris@42
|
1172 E TM, T3n, T12, T3k, TR, T3o, TX, T3j;
|
Chris@42
|
1173 {
|
Chris@42
|
1174 E TJ, TL, TI, TK;
|
Chris@42
|
1175 TJ = cr[WS(rs, 2)];
|
Chris@42
|
1176 TL = ci[WS(rs, 2)];
|
Chris@42
|
1177 TI = W[2];
|
Chris@42
|
1178 TK = W[3];
|
Chris@42
|
1179 TM = FMA(TI, TJ, TK * TL);
|
Chris@42
|
1180 T3n = FNMS(TK, TJ, TI * TL);
|
Chris@42
|
1181 }
|
Chris@42
|
1182 {
|
Chris@42
|
1183 E TZ, T11, TY, T10;
|
Chris@42
|
1184 TZ = cr[WS(rs, 26)];
|
Chris@42
|
1185 T11 = ci[WS(rs, 26)];
|
Chris@42
|
1186 TY = W[50];
|
Chris@42
|
1187 T10 = W[51];
|
Chris@42
|
1188 T12 = FMA(TY, TZ, T10 * T11);
|
Chris@42
|
1189 T3k = FNMS(T10, TZ, TY * T11);
|
Chris@42
|
1190 }
|
Chris@42
|
1191 {
|
Chris@42
|
1192 E TO, TQ, TN, TP;
|
Chris@42
|
1193 TO = cr[WS(rs, 18)];
|
Chris@42
|
1194 TQ = ci[WS(rs, 18)];
|
Chris@42
|
1195 TN = W[34];
|
Chris@42
|
1196 TP = W[35];
|
Chris@42
|
1197 TR = FMA(TN, TO, TP * TQ);
|
Chris@42
|
1198 T3o = FNMS(TP, TO, TN * TQ);
|
Chris@42
|
1199 }
|
Chris@42
|
1200 {
|
Chris@42
|
1201 E TU, TW, TT, TV;
|
Chris@42
|
1202 TU = cr[WS(rs, 10)];
|
Chris@42
|
1203 TW = ci[WS(rs, 10)];
|
Chris@42
|
1204 TT = W[18];
|
Chris@42
|
1205 TV = W[19];
|
Chris@42
|
1206 TX = FMA(TT, TU, TV * TW);
|
Chris@42
|
1207 T3j = FNMS(TV, TU, TT * TW);
|
Chris@42
|
1208 }
|
Chris@42
|
1209 {
|
Chris@42
|
1210 E TS, T13, T5L, T5M;
|
Chris@42
|
1211 TS = TM + TR;
|
Chris@42
|
1212 T13 = TX + T12;
|
Chris@42
|
1213 T14 = TS + T13;
|
Chris@42
|
1214 T5K = TS - T13;
|
Chris@42
|
1215 T5L = T3n + T3o;
|
Chris@42
|
1216 T5M = T3j + T3k;
|
Chris@42
|
1217 T5N = T5L - T5M;
|
Chris@42
|
1218 T6F = T5L + T5M;
|
Chris@42
|
1219 }
|
Chris@42
|
1220 {
|
Chris@42
|
1221 E T3i, T3l, T3p, T3q;
|
Chris@42
|
1222 T3i = TM - TR;
|
Chris@42
|
1223 T3l = T3j - T3k;
|
Chris@42
|
1224 T3m = T3i + T3l;
|
Chris@42
|
1225 T4Z = T3i - T3l;
|
Chris@42
|
1226 T3p = T3n - T3o;
|
Chris@42
|
1227 T3q = TX - T12;
|
Chris@42
|
1228 T3r = T3p - T3q;
|
Chris@42
|
1229 T4Y = T3p + T3q;
|
Chris@42
|
1230 }
|
Chris@42
|
1231 }
|
Chris@42
|
1232 {
|
Chris@42
|
1233 E T19, T3t, T1p, T3A, T1e, T3u, T1k, T3z;
|
Chris@42
|
1234 {
|
Chris@42
|
1235 E T16, T18, T15, T17;
|
Chris@42
|
1236 T16 = cr[WS(rs, 30)];
|
Chris@42
|
1237 T18 = ci[WS(rs, 30)];
|
Chris@42
|
1238 T15 = W[58];
|
Chris@42
|
1239 T17 = W[59];
|
Chris@42
|
1240 T19 = FMA(T15, T16, T17 * T18);
|
Chris@42
|
1241 T3t = FNMS(T17, T16, T15 * T18);
|
Chris@42
|
1242 }
|
Chris@42
|
1243 {
|
Chris@42
|
1244 E T1m, T1o, T1l, T1n;
|
Chris@42
|
1245 T1m = cr[WS(rs, 22)];
|
Chris@42
|
1246 T1o = ci[WS(rs, 22)];
|
Chris@42
|
1247 T1l = W[42];
|
Chris@42
|
1248 T1n = W[43];
|
Chris@42
|
1249 T1p = FMA(T1l, T1m, T1n * T1o);
|
Chris@42
|
1250 T3A = FNMS(T1n, T1m, T1l * T1o);
|
Chris@42
|
1251 }
|
Chris@42
|
1252 {
|
Chris@42
|
1253 E T1b, T1d, T1a, T1c;
|
Chris@42
|
1254 T1b = cr[WS(rs, 14)];
|
Chris@42
|
1255 T1d = ci[WS(rs, 14)];
|
Chris@42
|
1256 T1a = W[26];
|
Chris@42
|
1257 T1c = W[27];
|
Chris@42
|
1258 T1e = FMA(T1a, T1b, T1c * T1d);
|
Chris@42
|
1259 T3u = FNMS(T1c, T1b, T1a * T1d);
|
Chris@42
|
1260 }
|
Chris@42
|
1261 {
|
Chris@42
|
1262 E T1h, T1j, T1g, T1i;
|
Chris@42
|
1263 T1h = cr[WS(rs, 6)];
|
Chris@42
|
1264 T1j = ci[WS(rs, 6)];
|
Chris@42
|
1265 T1g = W[10];
|
Chris@42
|
1266 T1i = W[11];
|
Chris@42
|
1267 T1k = FMA(T1g, T1h, T1i * T1j);
|
Chris@42
|
1268 T3z = FNMS(T1i, T1h, T1g * T1j);
|
Chris@42
|
1269 }
|
Chris@42
|
1270 {
|
Chris@42
|
1271 E T1f, T1q, T5Q, T5R;
|
Chris@42
|
1272 T1f = T19 + T1e;
|
Chris@42
|
1273 T1q = T1k + T1p;
|
Chris@42
|
1274 T1r = T1f + T1q;
|
Chris@42
|
1275 T5P = T1f - T1q;
|
Chris@42
|
1276 T5Q = T3t + T3u;
|
Chris@42
|
1277 T5R = T3z + T3A;
|
Chris@42
|
1278 T5S = T5Q - T5R;
|
Chris@42
|
1279 T6E = T5Q + T5R;
|
Chris@42
|
1280 }
|
Chris@42
|
1281 {
|
Chris@42
|
1282 E T3v, T3w, T3y, T3B;
|
Chris@42
|
1283 T3v = T3t - T3u;
|
Chris@42
|
1284 T3w = T1k - T1p;
|
Chris@42
|
1285 T3x = T3v - T3w;
|
Chris@42
|
1286 T52 = T3v + T3w;
|
Chris@42
|
1287 T3y = T19 - T1e;
|
Chris@42
|
1288 T3B = T3z - T3A;
|
Chris@42
|
1289 T3C = T3y + T3B;
|
Chris@42
|
1290 T51 = T3y - T3B;
|
Chris@42
|
1291 }
|
Chris@42
|
1292 }
|
Chris@42
|
1293 {
|
Chris@42
|
1294 E T1V, T3M, T20, T3N, T3L, T3O, T26, T3Q, T2b, T3R, T3S, T3T;
|
Chris@42
|
1295 {
|
Chris@42
|
1296 E T1S, T1U, T1R, T1T;
|
Chris@42
|
1297 T1S = cr[WS(rs, 5)];
|
Chris@42
|
1298 T1U = ci[WS(rs, 5)];
|
Chris@42
|
1299 T1R = W[8];
|
Chris@42
|
1300 T1T = W[9];
|
Chris@42
|
1301 T1V = FMA(T1R, T1S, T1T * T1U);
|
Chris@42
|
1302 T3M = FNMS(T1T, T1S, T1R * T1U);
|
Chris@42
|
1303 }
|
Chris@42
|
1304 {
|
Chris@42
|
1305 E T1X, T1Z, T1W, T1Y;
|
Chris@42
|
1306 T1X = cr[WS(rs, 21)];
|
Chris@42
|
1307 T1Z = ci[WS(rs, 21)];
|
Chris@42
|
1308 T1W = W[40];
|
Chris@42
|
1309 T1Y = W[41];
|
Chris@42
|
1310 T20 = FMA(T1W, T1X, T1Y * T1Z);
|
Chris@42
|
1311 T3N = FNMS(T1Y, T1X, T1W * T1Z);
|
Chris@42
|
1312 }
|
Chris@42
|
1313 T3L = T1V - T20;
|
Chris@42
|
1314 T3O = T3M - T3N;
|
Chris@42
|
1315 {
|
Chris@42
|
1316 E T23, T25, T22, T24;
|
Chris@42
|
1317 T23 = cr[WS(rs, 29)];
|
Chris@42
|
1318 T25 = ci[WS(rs, 29)];
|
Chris@42
|
1319 T22 = W[56];
|
Chris@42
|
1320 T24 = W[57];
|
Chris@42
|
1321 T26 = FMA(T22, T23, T24 * T25);
|
Chris@42
|
1322 T3Q = FNMS(T24, T23, T22 * T25);
|
Chris@42
|
1323 }
|
Chris@42
|
1324 {
|
Chris@42
|
1325 E T28, T2a, T27, T29;
|
Chris@42
|
1326 T28 = cr[WS(rs, 13)];
|
Chris@42
|
1327 T2a = ci[WS(rs, 13)];
|
Chris@42
|
1328 T27 = W[24];
|
Chris@42
|
1329 T29 = W[25];
|
Chris@42
|
1330 T2b = FMA(T27, T28, T29 * T2a);
|
Chris@42
|
1331 T3R = FNMS(T29, T28, T27 * T2a);
|
Chris@42
|
1332 }
|
Chris@42
|
1333 T3S = T3Q - T3R;
|
Chris@42
|
1334 T3T = T26 - T2b;
|
Chris@42
|
1335 {
|
Chris@42
|
1336 E T21, T2c, T62, T63;
|
Chris@42
|
1337 T21 = T1V + T20;
|
Chris@42
|
1338 T2c = T26 + T2b;
|
Chris@42
|
1339 T2d = T21 + T2c;
|
Chris@42
|
1340 T5Z = T21 - T2c;
|
Chris@42
|
1341 T62 = T3Q + T3R;
|
Chris@42
|
1342 T63 = T3M + T3N;
|
Chris@42
|
1343 T64 = T62 - T63;
|
Chris@42
|
1344 T6K = T63 + T62;
|
Chris@42
|
1345 }
|
Chris@42
|
1346 {
|
Chris@42
|
1347 E T3P, T3U, T42, T43;
|
Chris@42
|
1348 T3P = T3L + T3O;
|
Chris@42
|
1349 T3U = T3S - T3T;
|
Chris@42
|
1350 T3V = KP707106781 * (T3P - T3U);
|
Chris@42
|
1351 T5a = KP707106781 * (T3P + T3U);
|
Chris@42
|
1352 T42 = T3T + T3S;
|
Chris@42
|
1353 T43 = T3L - T3O;
|
Chris@42
|
1354 T44 = KP707106781 * (T42 - T43);
|
Chris@42
|
1355 T57 = KP707106781 * (T43 + T42);
|
Chris@42
|
1356 }
|
Chris@42
|
1357 }
|
Chris@42
|
1358 {
|
Chris@42
|
1359 E T2G, T4i, T2L, T4j, T4h, T4k, T2R, T4d, T2W, T4e, T4c, T4f;
|
Chris@42
|
1360 {
|
Chris@42
|
1361 E T2D, T2F, T2C, T2E;
|
Chris@42
|
1362 T2D = cr[WS(rs, 3)];
|
Chris@42
|
1363 T2F = ci[WS(rs, 3)];
|
Chris@42
|
1364 T2C = W[4];
|
Chris@42
|
1365 T2E = W[5];
|
Chris@42
|
1366 T2G = FMA(T2C, T2D, T2E * T2F);
|
Chris@42
|
1367 T4i = FNMS(T2E, T2D, T2C * T2F);
|
Chris@42
|
1368 }
|
Chris@42
|
1369 {
|
Chris@42
|
1370 E T2I, T2K, T2H, T2J;
|
Chris@42
|
1371 T2I = cr[WS(rs, 19)];
|
Chris@42
|
1372 T2K = ci[WS(rs, 19)];
|
Chris@42
|
1373 T2H = W[36];
|
Chris@42
|
1374 T2J = W[37];
|
Chris@42
|
1375 T2L = FMA(T2H, T2I, T2J * T2K);
|
Chris@42
|
1376 T4j = FNMS(T2J, T2I, T2H * T2K);
|
Chris@42
|
1377 }
|
Chris@42
|
1378 T4h = T2G - T2L;
|
Chris@42
|
1379 T4k = T4i - T4j;
|
Chris@42
|
1380 {
|
Chris@42
|
1381 E T2O, T2Q, T2N, T2P;
|
Chris@42
|
1382 T2O = cr[WS(rs, 27)];
|
Chris@42
|
1383 T2Q = ci[WS(rs, 27)];
|
Chris@42
|
1384 T2N = W[52];
|
Chris@42
|
1385 T2P = W[53];
|
Chris@42
|
1386 T2R = FMA(T2N, T2O, T2P * T2Q);
|
Chris@42
|
1387 T4d = FNMS(T2P, T2O, T2N * T2Q);
|
Chris@42
|
1388 }
|
Chris@42
|
1389 {
|
Chris@42
|
1390 E T2T, T2V, T2S, T2U;
|
Chris@42
|
1391 T2T = cr[WS(rs, 11)];
|
Chris@42
|
1392 T2V = ci[WS(rs, 11)];
|
Chris@42
|
1393 T2S = W[20];
|
Chris@42
|
1394 T2U = W[21];
|
Chris@42
|
1395 T2W = FMA(T2S, T2T, T2U * T2V);
|
Chris@42
|
1396 T4e = FNMS(T2U, T2T, T2S * T2V);
|
Chris@42
|
1397 }
|
Chris@42
|
1398 T4c = T2R - T2W;
|
Chris@42
|
1399 T4f = T4d - T4e;
|
Chris@42
|
1400 {
|
Chris@42
|
1401 E T2M, T2X, T68, T69;
|
Chris@42
|
1402 T2M = T2G + T2L;
|
Chris@42
|
1403 T2X = T2R + T2W;
|
Chris@42
|
1404 T2Y = T2M + T2X;
|
Chris@42
|
1405 T6f = T2M - T2X;
|
Chris@42
|
1406 T68 = T4d + T4e;
|
Chris@42
|
1407 T69 = T4i + T4j;
|
Chris@42
|
1408 T6a = T68 - T69;
|
Chris@42
|
1409 T6P = T69 + T68;
|
Chris@42
|
1410 }
|
Chris@42
|
1411 {
|
Chris@42
|
1412 E T4g, T4l, T4t, T4u;
|
Chris@42
|
1413 T4g = T4c + T4f;
|
Chris@42
|
1414 T4l = T4h - T4k;
|
Chris@42
|
1415 T4m = KP707106781 * (T4g - T4l);
|
Chris@42
|
1416 T5e = KP707106781 * (T4l + T4g);
|
Chris@42
|
1417 T4t = T4h + T4k;
|
Chris@42
|
1418 T4u = T4f - T4c;
|
Chris@42
|
1419 T4v = KP707106781 * (T4t - T4u);
|
Chris@42
|
1420 T5h = KP707106781 * (T4t + T4u);
|
Chris@42
|
1421 }
|
Chris@42
|
1422 }
|
Chris@42
|
1423 {
|
Chris@42
|
1424 E T1t, T6X, T7a, T7c, T30, T7b, T70, T71;
|
Chris@42
|
1425 {
|
Chris@42
|
1426 E TH, T1s, T72, T79;
|
Chris@42
|
1427 TH = Tj + TG;
|
Chris@42
|
1428 T1s = T14 + T1r;
|
Chris@42
|
1429 T1t = TH + T1s;
|
Chris@42
|
1430 T6X = TH - T1s;
|
Chris@42
|
1431 T72 = T6F + T6E;
|
Chris@42
|
1432 T79 = T73 + T78;
|
Chris@42
|
1433 T7a = T72 + T79;
|
Chris@42
|
1434 T7c = T79 - T72;
|
Chris@42
|
1435 }
|
Chris@42
|
1436 {
|
Chris@42
|
1437 E T2e, T2Z, T6Y, T6Z;
|
Chris@42
|
1438 T2e = T1Q + T2d;
|
Chris@42
|
1439 T2Z = T2B + T2Y;
|
Chris@42
|
1440 T30 = T2e + T2Z;
|
Chris@42
|
1441 T7b = T2Z - T2e;
|
Chris@42
|
1442 T6Y = T6O + T6P;
|
Chris@42
|
1443 T6Z = T6J + T6K;
|
Chris@42
|
1444 T70 = T6Y - T6Z;
|
Chris@42
|
1445 T71 = T6Z + T6Y;
|
Chris@42
|
1446 }
|
Chris@42
|
1447 ci[WS(rs, 15)] = T1t - T30;
|
Chris@42
|
1448 cr[WS(rs, 24)] = T7b - T7c;
|
Chris@42
|
1449 ci[WS(rs, 23)] = T7b + T7c;
|
Chris@42
|
1450 cr[0] = T1t + T30;
|
Chris@42
|
1451 cr[WS(rs, 8)] = T6X - T70;
|
Chris@42
|
1452 cr[WS(rs, 16)] = T71 - T7a;
|
Chris@42
|
1453 ci[WS(rs, 31)] = T71 + T7a;
|
Chris@42
|
1454 ci[WS(rs, 7)] = T6X + T70;
|
Chris@42
|
1455 }
|
Chris@42
|
1456 {
|
Chris@42
|
1457 E T4X, T5p, T7D, T7J, T54, T7y, T5z, T5D, T5c, T5m, T5s, T7I, T5w, T5C, T5j;
|
Chris@42
|
1458 E T5n, T4W, T7z;
|
Chris@42
|
1459 T4W = KP707106781 * (T4U + T4V);
|
Chris@42
|
1460 T4X = T4T - T4W;
|
Chris@42
|
1461 T5p = T4T + T4W;
|
Chris@42
|
1462 T7z = KP707106781 * (T3a - T3f);
|
Chris@42
|
1463 T7D = T7z + T7C;
|
Chris@42
|
1464 T7J = T7C - T7z;
|
Chris@42
|
1465 {
|
Chris@42
|
1466 E T50, T53, T5x, T5y;
|
Chris@42
|
1467 T50 = FMA(KP923879532, T4Y, KP382683432 * T4Z);
|
Chris@42
|
1468 T53 = FNMS(KP923879532, T52, KP382683432 * T51);
|
Chris@42
|
1469 T54 = T50 + T53;
|
Chris@42
|
1470 T7y = T50 - T53;
|
Chris@42
|
1471 T5x = T5d + T5e;
|
Chris@42
|
1472 T5y = T5g + T5h;
|
Chris@42
|
1473 T5z = FNMS(KP980785280, T5y, KP195090322 * T5x);
|
Chris@42
|
1474 T5D = FMA(KP980785280, T5x, KP195090322 * T5y);
|
Chris@42
|
1475 }
|
Chris@42
|
1476 {
|
Chris@42
|
1477 E T58, T5b, T5q, T5r;
|
Chris@42
|
1478 T58 = T56 - T57;
|
Chris@42
|
1479 T5b = T59 - T5a;
|
Chris@42
|
1480 T5c = FMA(KP831469612, T58, KP555570233 * T5b);
|
Chris@42
|
1481 T5m = FNMS(KP831469612, T5b, KP555570233 * T58);
|
Chris@42
|
1482 T5q = FNMS(KP382683432, T4Y, KP923879532 * T4Z);
|
Chris@42
|
1483 T5r = FMA(KP382683432, T52, KP923879532 * T51);
|
Chris@42
|
1484 T5s = T5q + T5r;
|
Chris@42
|
1485 T7I = T5r - T5q;
|
Chris@42
|
1486 }
|
Chris@42
|
1487 {
|
Chris@42
|
1488 E T5u, T5v, T5f, T5i;
|
Chris@42
|
1489 T5u = T56 + T57;
|
Chris@42
|
1490 T5v = T59 + T5a;
|
Chris@42
|
1491 T5w = FMA(KP195090322, T5u, KP980785280 * T5v);
|
Chris@42
|
1492 T5C = FNMS(KP195090322, T5v, KP980785280 * T5u);
|
Chris@42
|
1493 T5f = T5d - T5e;
|
Chris@42
|
1494 T5i = T5g - T5h;
|
Chris@42
|
1495 T5j = FNMS(KP555570233, T5i, KP831469612 * T5f);
|
Chris@42
|
1496 T5n = FMA(KP555570233, T5f, KP831469612 * T5i);
|
Chris@42
|
1497 }
|
Chris@42
|
1498 {
|
Chris@42
|
1499 E T55, T5k, T7H, T7K;
|
Chris@42
|
1500 T55 = T4X + T54;
|
Chris@42
|
1501 T5k = T5c + T5j;
|
Chris@42
|
1502 ci[WS(rs, 12)] = T55 - T5k;
|
Chris@42
|
1503 cr[WS(rs, 3)] = T55 + T5k;
|
Chris@42
|
1504 T7H = T5n - T5m;
|
Chris@42
|
1505 T7K = T7I + T7J;
|
Chris@42
|
1506 cr[WS(rs, 19)] = T7H - T7K;
|
Chris@42
|
1507 ci[WS(rs, 28)] = T7H + T7K;
|
Chris@42
|
1508 }
|
Chris@42
|
1509 {
|
Chris@42
|
1510 E T7L, T7M, T5l, T5o;
|
Chris@42
|
1511 T7L = T5j - T5c;
|
Chris@42
|
1512 T7M = T7J - T7I;
|
Chris@42
|
1513 cr[WS(rs, 27)] = T7L - T7M;
|
Chris@42
|
1514 ci[WS(rs, 20)] = T7L + T7M;
|
Chris@42
|
1515 T5l = T4X - T54;
|
Chris@42
|
1516 T5o = T5m + T5n;
|
Chris@42
|
1517 cr[WS(rs, 11)] = T5l - T5o;
|
Chris@42
|
1518 ci[WS(rs, 4)] = T5l + T5o;
|
Chris@42
|
1519 }
|
Chris@42
|
1520 {
|
Chris@42
|
1521 E T5t, T5A, T7x, T7E;
|
Chris@42
|
1522 T5t = T5p - T5s;
|
Chris@42
|
1523 T5A = T5w + T5z;
|
Chris@42
|
1524 ci[WS(rs, 8)] = T5t - T5A;
|
Chris@42
|
1525 cr[WS(rs, 7)] = T5t + T5A;
|
Chris@42
|
1526 T7x = T5z - T5w;
|
Chris@42
|
1527 T7E = T7y + T7D;
|
Chris@42
|
1528 cr[WS(rs, 31)] = T7x - T7E;
|
Chris@42
|
1529 ci[WS(rs, 16)] = T7x + T7E;
|
Chris@42
|
1530 }
|
Chris@42
|
1531 {
|
Chris@42
|
1532 E T7F, T7G, T5B, T5E;
|
Chris@42
|
1533 T7F = T5D - T5C;
|
Chris@42
|
1534 T7G = T7D - T7y;
|
Chris@42
|
1535 cr[WS(rs, 23)] = T7F - T7G;
|
Chris@42
|
1536 ci[WS(rs, 24)] = T7F + T7G;
|
Chris@42
|
1537 T5B = T5p + T5s;
|
Chris@42
|
1538 T5E = T5C + T5D;
|
Chris@42
|
1539 cr[WS(rs, 15)] = T5B - T5E;
|
Chris@42
|
1540 ci[0] = T5B + T5E;
|
Chris@42
|
1541 }
|
Chris@42
|
1542 }
|
Chris@42
|
1543 {
|
Chris@42
|
1544 E T6H, T6T, T7g, T7i, T6M, T6U, T6R, T6V;
|
Chris@42
|
1545 {
|
Chris@42
|
1546 E T6D, T6G, T7e, T7f;
|
Chris@42
|
1547 T6D = Tj - TG;
|
Chris@42
|
1548 T6G = T6E - T6F;
|
Chris@42
|
1549 T6H = T6D - T6G;
|
Chris@42
|
1550 T6T = T6D + T6G;
|
Chris@42
|
1551 T7e = T14 - T1r;
|
Chris@42
|
1552 T7f = T78 - T73;
|
Chris@42
|
1553 T7g = T7e + T7f;
|
Chris@42
|
1554 T7i = T7f - T7e;
|
Chris@42
|
1555 }
|
Chris@42
|
1556 {
|
Chris@42
|
1557 E T6I, T6L, T6N, T6Q;
|
Chris@42
|
1558 T6I = T1Q - T2d;
|
Chris@42
|
1559 T6L = T6J - T6K;
|
Chris@42
|
1560 T6M = T6I + T6L;
|
Chris@42
|
1561 T6U = T6I - T6L;
|
Chris@42
|
1562 T6N = T2B - T2Y;
|
Chris@42
|
1563 T6Q = T6O - T6P;
|
Chris@42
|
1564 T6R = T6N - T6Q;
|
Chris@42
|
1565 T6V = T6N + T6Q;
|
Chris@42
|
1566 }
|
Chris@42
|
1567 {
|
Chris@42
|
1568 E T6S, T7h, T6W, T7d;
|
Chris@42
|
1569 T6S = KP707106781 * (T6M + T6R);
|
Chris@42
|
1570 ci[WS(rs, 11)] = T6H - T6S;
|
Chris@42
|
1571 cr[WS(rs, 4)] = T6H + T6S;
|
Chris@42
|
1572 T7h = KP707106781 * (T6V - T6U);
|
Chris@42
|
1573 cr[WS(rs, 20)] = T7h - T7i;
|
Chris@42
|
1574 ci[WS(rs, 27)] = T7h + T7i;
|
Chris@42
|
1575 T6W = KP707106781 * (T6U + T6V);
|
Chris@42
|
1576 cr[WS(rs, 12)] = T6T - T6W;
|
Chris@42
|
1577 ci[WS(rs, 3)] = T6T + T6W;
|
Chris@42
|
1578 T7d = KP707106781 * (T6R - T6M);
|
Chris@42
|
1579 cr[WS(rs, 28)] = T7d - T7g;
|
Chris@42
|
1580 ci[WS(rs, 19)] = T7d + T7g;
|
Chris@42
|
1581 }
|
Chris@42
|
1582 }
|
Chris@42
|
1583 {
|
Chris@42
|
1584 E T5J, T7n, T7t, T6n, T5U, T7k, T6x, T6B, T6q, T7s, T66, T6k, T6u, T6A, T6h;
|
Chris@42
|
1585 E T6l;
|
Chris@42
|
1586 {
|
Chris@42
|
1587 E T5O, T5T, T60, T65;
|
Chris@42
|
1588 T5J = T5F - T5I;
|
Chris@42
|
1589 T7n = T7l + T7m;
|
Chris@42
|
1590 T7t = T7m - T7l;
|
Chris@42
|
1591 T6n = T5F + T5I;
|
Chris@42
|
1592 T5O = T5K + T5N;
|
Chris@42
|
1593 T5T = T5P - T5S;
|
Chris@42
|
1594 T5U = KP707106781 * (T5O + T5T);
|
Chris@42
|
1595 T7k = KP707106781 * (T5O - T5T);
|
Chris@42
|
1596 {
|
Chris@42
|
1597 E T6v, T6w, T6o, T6p;
|
Chris@42
|
1598 T6v = T6e + T6f;
|
Chris@42
|
1599 T6w = T67 + T6a;
|
Chris@42
|
1600 T6x = FMA(KP382683432, T6v, KP923879532 * T6w);
|
Chris@42
|
1601 T6B = FNMS(KP923879532, T6v, KP382683432 * T6w);
|
Chris@42
|
1602 T6o = T5K - T5N;
|
Chris@42
|
1603 T6p = T5P + T5S;
|
Chris@42
|
1604 T6q = KP707106781 * (T6o + T6p);
|
Chris@42
|
1605 T7s = KP707106781 * (T6p - T6o);
|
Chris@42
|
1606 }
|
Chris@42
|
1607 T60 = T5Y - T5Z;
|
Chris@42
|
1608 T65 = T61 - T64;
|
Chris@42
|
1609 T66 = FMA(KP382683432, T60, KP923879532 * T65);
|
Chris@42
|
1610 T6k = FNMS(KP923879532, T60, KP382683432 * T65);
|
Chris@42
|
1611 {
|
Chris@42
|
1612 E T6s, T6t, T6b, T6g;
|
Chris@42
|
1613 T6s = T61 + T64;
|
Chris@42
|
1614 T6t = T5Y + T5Z;
|
Chris@42
|
1615 T6u = FNMS(KP382683432, T6t, KP923879532 * T6s);
|
Chris@42
|
1616 T6A = FMA(KP923879532, T6t, KP382683432 * T6s);
|
Chris@42
|
1617 T6b = T67 - T6a;
|
Chris@42
|
1618 T6g = T6e - T6f;
|
Chris@42
|
1619 T6h = FNMS(KP382683432, T6g, KP923879532 * T6b);
|
Chris@42
|
1620 T6l = FMA(KP923879532, T6g, KP382683432 * T6b);
|
Chris@42
|
1621 }
|
Chris@42
|
1622 }
|
Chris@42
|
1623 {
|
Chris@42
|
1624 E T5V, T6i, T7r, T7u;
|
Chris@42
|
1625 T5V = T5J + T5U;
|
Chris@42
|
1626 T6i = T66 + T6h;
|
Chris@42
|
1627 ci[WS(rs, 13)] = T5V - T6i;
|
Chris@42
|
1628 cr[WS(rs, 2)] = T5V + T6i;
|
Chris@42
|
1629 T7r = T6l - T6k;
|
Chris@42
|
1630 T7u = T7s + T7t;
|
Chris@42
|
1631 cr[WS(rs, 18)] = T7r - T7u;
|
Chris@42
|
1632 ci[WS(rs, 29)] = T7r + T7u;
|
Chris@42
|
1633 }
|
Chris@42
|
1634 {
|
Chris@42
|
1635 E T7v, T7w, T6j, T6m;
|
Chris@42
|
1636 T7v = T6h - T66;
|
Chris@42
|
1637 T7w = T7t - T7s;
|
Chris@42
|
1638 cr[WS(rs, 26)] = T7v - T7w;
|
Chris@42
|
1639 ci[WS(rs, 21)] = T7v + T7w;
|
Chris@42
|
1640 T6j = T5J - T5U;
|
Chris@42
|
1641 T6m = T6k + T6l;
|
Chris@42
|
1642 cr[WS(rs, 10)] = T6j - T6m;
|
Chris@42
|
1643 ci[WS(rs, 5)] = T6j + T6m;
|
Chris@42
|
1644 }
|
Chris@42
|
1645 {
|
Chris@42
|
1646 E T6r, T6y, T7j, T7o;
|
Chris@42
|
1647 T6r = T6n + T6q;
|
Chris@42
|
1648 T6y = T6u + T6x;
|
Chris@42
|
1649 cr[WS(rs, 14)] = T6r - T6y;
|
Chris@42
|
1650 ci[WS(rs, 1)] = T6r + T6y;
|
Chris@42
|
1651 T7j = T6B - T6A;
|
Chris@42
|
1652 T7o = T7k + T7n;
|
Chris@42
|
1653 cr[WS(rs, 30)] = T7j - T7o;
|
Chris@42
|
1654 ci[WS(rs, 17)] = T7j + T7o;
|
Chris@42
|
1655 }
|
Chris@42
|
1656 {
|
Chris@42
|
1657 E T7p, T7q, T6z, T6C;
|
Chris@42
|
1658 T7p = T6x - T6u;
|
Chris@42
|
1659 T7q = T7n - T7k;
|
Chris@42
|
1660 cr[WS(rs, 22)] = T7p - T7q;
|
Chris@42
|
1661 ci[WS(rs, 25)] = T7p + T7q;
|
Chris@42
|
1662 T6z = T6n - T6q;
|
Chris@42
|
1663 T6C = T6A + T6B;
|
Chris@42
|
1664 ci[WS(rs, 9)] = T6z - T6C;
|
Chris@42
|
1665 cr[WS(rs, 6)] = T6z + T6C;
|
Chris@42
|
1666 }
|
Chris@42
|
1667 }
|
Chris@42
|
1668 {
|
Chris@42
|
1669 E T3h, T4D, T7R, T7X, T3E, T7O, T4N, T4R, T46, T4A, T4G, T7W, T4K, T4Q, T4x;
|
Chris@42
|
1670 E T4B, T3g, T7P;
|
Chris@42
|
1671 T3g = KP707106781 * (T3a + T3f);
|
Chris@42
|
1672 T3h = T35 - T3g;
|
Chris@42
|
1673 T4D = T35 + T3g;
|
Chris@42
|
1674 T7P = KP707106781 * (T4V - T4U);
|
Chris@42
|
1675 T7R = T7P + T7Q;
|
Chris@42
|
1676 T7X = T7Q - T7P;
|
Chris@42
|
1677 {
|
Chris@42
|
1678 E T3s, T3D, T4L, T4M;
|
Chris@42
|
1679 T3s = FNMS(KP923879532, T3r, KP382683432 * T3m);
|
Chris@42
|
1680 T3D = FMA(KP923879532, T3x, KP382683432 * T3C);
|
Chris@42
|
1681 T3E = T3s + T3D;
|
Chris@42
|
1682 T7O = T3D - T3s;
|
Chris@42
|
1683 T4L = T4s + T4v;
|
Chris@42
|
1684 T4M = T4b + T4m;
|
Chris@42
|
1685 T4N = FNMS(KP195090322, T4M, KP980785280 * T4L);
|
Chris@42
|
1686 T4R = FMA(KP980785280, T4M, KP195090322 * T4L);
|
Chris@42
|
1687 }
|
Chris@42
|
1688 {
|
Chris@42
|
1689 E T3W, T45, T4E, T4F;
|
Chris@42
|
1690 T3W = T3K - T3V;
|
Chris@42
|
1691 T45 = T41 - T44;
|
Chris@42
|
1692 T46 = FNMS(KP555570233, T45, KP831469612 * T3W);
|
Chris@42
|
1693 T4A = FMA(KP831469612, T45, KP555570233 * T3W);
|
Chris@42
|
1694 T4E = FMA(KP382683432, T3r, KP923879532 * T3m);
|
Chris@42
|
1695 T4F = FNMS(KP382683432, T3x, KP923879532 * T3C);
|
Chris@42
|
1696 T4G = T4E + T4F;
|
Chris@42
|
1697 T7W = T4E - T4F;
|
Chris@42
|
1698 }
|
Chris@42
|
1699 {
|
Chris@42
|
1700 E T4I, T4J, T4n, T4w;
|
Chris@42
|
1701 T4I = T41 + T44;
|
Chris@42
|
1702 T4J = T3K + T3V;
|
Chris@42
|
1703 T4K = FMA(KP195090322, T4I, KP980785280 * T4J);
|
Chris@42
|
1704 T4Q = FNMS(KP980785280, T4I, KP195090322 * T4J);
|
Chris@42
|
1705 T4n = T4b - T4m;
|
Chris@42
|
1706 T4w = T4s - T4v;
|
Chris@42
|
1707 T4x = FMA(KP555570233, T4n, KP831469612 * T4w);
|
Chris@42
|
1708 T4B = FNMS(KP831469612, T4n, KP555570233 * T4w);
|
Chris@42
|
1709 }
|
Chris@42
|
1710 {
|
Chris@42
|
1711 E T3F, T4y, T7V, T7Y;
|
Chris@42
|
1712 T3F = T3h + T3E;
|
Chris@42
|
1713 T4y = T46 + T4x;
|
Chris@42
|
1714 cr[WS(rs, 13)] = T3F - T4y;
|
Chris@42
|
1715 ci[WS(rs, 2)] = T3F + T4y;
|
Chris@42
|
1716 T7V = T4B - T4A;
|
Chris@42
|
1717 T7Y = T7W + T7X;
|
Chris@42
|
1718 cr[WS(rs, 29)] = T7V - T7Y;
|
Chris@42
|
1719 ci[WS(rs, 18)] = T7V + T7Y;
|
Chris@42
|
1720 }
|
Chris@42
|
1721 {
|
Chris@42
|
1722 E T7Z, T80, T4z, T4C;
|
Chris@42
|
1723 T7Z = T4x - T46;
|
Chris@42
|
1724 T80 = T7X - T7W;
|
Chris@42
|
1725 cr[WS(rs, 21)] = T7Z - T80;
|
Chris@42
|
1726 ci[WS(rs, 26)] = T7Z + T80;
|
Chris@42
|
1727 T4z = T3h - T3E;
|
Chris@42
|
1728 T4C = T4A + T4B;
|
Chris@42
|
1729 ci[WS(rs, 10)] = T4z - T4C;
|
Chris@42
|
1730 cr[WS(rs, 5)] = T4z + T4C;
|
Chris@42
|
1731 }
|
Chris@42
|
1732 {
|
Chris@42
|
1733 E T4H, T4O, T7N, T7S;
|
Chris@42
|
1734 T4H = T4D + T4G;
|
Chris@42
|
1735 T4O = T4K + T4N;
|
Chris@42
|
1736 ci[WS(rs, 14)] = T4H - T4O;
|
Chris@42
|
1737 cr[WS(rs, 1)] = T4H + T4O;
|
Chris@42
|
1738 T7N = T4R - T4Q;
|
Chris@42
|
1739 T7S = T7O + T7R;
|
Chris@42
|
1740 cr[WS(rs, 17)] = T7N - T7S;
|
Chris@42
|
1741 ci[WS(rs, 30)] = T7N + T7S;
|
Chris@42
|
1742 }
|
Chris@42
|
1743 {
|
Chris@42
|
1744 E T7T, T7U, T4P, T4S;
|
Chris@42
|
1745 T7T = T4N - T4K;
|
Chris@42
|
1746 T7U = T7R - T7O;
|
Chris@42
|
1747 cr[WS(rs, 25)] = T7T - T7U;
|
Chris@42
|
1748 ci[WS(rs, 22)] = T7T + T7U;
|
Chris@42
|
1749 T4P = T4D - T4G;
|
Chris@42
|
1750 T4S = T4Q + T4R;
|
Chris@42
|
1751 cr[WS(rs, 9)] = T4P - T4S;
|
Chris@42
|
1752 ci[WS(rs, 6)] = T4P + T4S;
|
Chris@42
|
1753 }
|
Chris@42
|
1754 }
|
Chris@42
|
1755 }
|
Chris@42
|
1756 }
|
Chris@42
|
1757 }
|
Chris@42
|
1758
|
Chris@42
|
1759 static const tw_instr twinstr[] = {
|
Chris@42
|
1760 {TW_FULL, 1, 32},
|
Chris@42
|
1761 {TW_NEXT, 1, 0}
|
Chris@42
|
1762 };
|
Chris@42
|
1763
|
Chris@42
|
1764 static const hc2hc_desc desc = { 32, "hf_32", twinstr, &GENUS, {340, 114, 94, 0} };
|
Chris@42
|
1765
|
Chris@42
|
1766 void X(codelet_hf_32) (planner *p) {
|
Chris@42
|
1767 X(khc2hc_register) (p, hf_32, &desc);
|
Chris@42
|
1768 }
|
Chris@42
|
1769 #endif /* HAVE_FMA */
|