cannam@167
|
1 /*
|
cannam@167
|
2 * Copyright (c) 2003, 2007-14 Matteo Frigo
|
cannam@167
|
3 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
|
cannam@167
|
4 *
|
cannam@167
|
5 * This program is free software; you can redistribute it and/or modify
|
cannam@167
|
6 * it under the terms of the GNU General Public License as published by
|
cannam@167
|
7 * the Free Software Foundation; either version 2 of the License, or
|
cannam@167
|
8 * (at your option) any later version.
|
cannam@167
|
9 *
|
cannam@167
|
10 * This program is distributed in the hope that it will be useful,
|
cannam@167
|
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
cannam@167
|
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
cannam@167
|
13 * GNU General Public License for more details.
|
cannam@167
|
14 *
|
cannam@167
|
15 * You should have received a copy of the GNU General Public License
|
cannam@167
|
16 * along with this program; if not, write to the Free Software
|
cannam@167
|
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
cannam@167
|
18 *
|
cannam@167
|
19 */
|
cannam@167
|
20
|
cannam@167
|
21 /* This file was automatically generated --- DO NOT EDIT */
|
cannam@167
|
22 /* Generated on Thu May 24 08:06:59 EDT 2018 */
|
cannam@167
|
23
|
cannam@167
|
24 #include "rdft/codelet-rdft.h"
|
cannam@167
|
25
|
cannam@167
|
26 #if defined(ARCH_PREFERS_FMA) || defined(ISA_EXTENSION_PREFERS_FMA)
|
cannam@167
|
27
|
cannam@167
|
28 /* Generated by: ../../../genfft/gen_hc2c.native -fma -compact -variables 4 -pipeline-latency 4 -n 32 -dit -name hc2cf_32 -include rdft/scalar/hc2cf.h */
|
cannam@167
|
29
|
cannam@167
|
30 /*
|
cannam@167
|
31 * This function contains 434 FP additions, 260 FP multiplications,
|
cannam@167
|
32 * (or, 236 additions, 62 multiplications, 198 fused multiply/add),
|
cannam@167
|
33 * 102 stack variables, 7 constants, and 128 memory accesses
|
cannam@167
|
34 */
|
cannam@167
|
35 #include "rdft/scalar/hc2cf.h"
|
cannam@167
|
36
|
cannam@167
|
37 static void hc2cf_32(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms)
|
cannam@167
|
38 {
|
cannam@167
|
39 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
cannam@167
|
40 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
cannam@167
|
41 DK(KP668178637, +0.668178637919298919997757686523080761552472251);
|
cannam@167
|
42 DK(KP198912367, +0.198912367379658006911597622644676228597850501);
|
cannam@167
|
43 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
cannam@167
|
44 DK(KP414213562, +0.414213562373095048801688724209698078569671875);
|
cannam@167
|
45 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
cannam@167
|
46 {
|
cannam@167
|
47 INT m;
|
cannam@167
|
48 for (m = mb, W = W + ((mb - 1) * 62); m < me; m = m + 1, Rp = Rp + ms, Ip = Ip + ms, Rm = Rm - ms, Im = Im - ms, W = W + 62, MAKE_VOLATILE_STRIDE(128, rs)) {
|
cannam@167
|
49 E T8, T8x, T3w, T87, Tl, T8y, T3B, T83, Tz, T6F, T3J, T5T, TM, T6G, T3Q;
|
cannam@167
|
50 E T5U, T11, T1e, T6M, T6J, T6K, T6L, T3Z, T5Y, T46, T5X, T1s, T1F, T6O, T6P;
|
cannam@167
|
51 E T6Q, T6R, T4e, T61, T4l, T60, T32, T7b, T79, T7N, T54, T6c, T5r, T6f, T29;
|
cannam@167
|
52 E T70, T6X, T7I, T4v, T65, T4S, T68, T3t, T76, T7e, T7O, T5b, T5s, T5i, T5t;
|
cannam@167
|
53 E T2A, T6Y, T73, T7J, T4C, T4T, T4J, T4U;
|
cannam@167
|
54 {
|
cannam@167
|
55 E T1, T86, T3, T6, T4, T84, T2, T7, T85, T5;
|
cannam@167
|
56 T1 = Rp[0];
|
cannam@167
|
57 T86 = Rm[0];
|
cannam@167
|
58 T3 = Rp[WS(rs, 8)];
|
cannam@167
|
59 T6 = Rm[WS(rs, 8)];
|
cannam@167
|
60 T2 = W[30];
|
cannam@167
|
61 T4 = T2 * T3;
|
cannam@167
|
62 T84 = T2 * T6;
|
cannam@167
|
63 T5 = W[31];
|
cannam@167
|
64 T7 = FMA(T5, T6, T4);
|
cannam@167
|
65 T85 = FNMS(T5, T3, T84);
|
cannam@167
|
66 T8 = T1 + T7;
|
cannam@167
|
67 T8x = T86 - T85;
|
cannam@167
|
68 T3w = T1 - T7;
|
cannam@167
|
69 T87 = T85 + T86;
|
cannam@167
|
70 }
|
cannam@167
|
71 {
|
cannam@167
|
72 E Ta, Td, Tb, T3x, Tg, Tj, Th, T3z, T9, Tf;
|
cannam@167
|
73 Ta = Rp[WS(rs, 4)];
|
cannam@167
|
74 Td = Rm[WS(rs, 4)];
|
cannam@167
|
75 T9 = W[14];
|
cannam@167
|
76 Tb = T9 * Ta;
|
cannam@167
|
77 T3x = T9 * Td;
|
cannam@167
|
78 Tg = Rp[WS(rs, 12)];
|
cannam@167
|
79 Tj = Rm[WS(rs, 12)];
|
cannam@167
|
80 Tf = W[46];
|
cannam@167
|
81 Th = Tf * Tg;
|
cannam@167
|
82 T3z = Tf * Tj;
|
cannam@167
|
83 {
|
cannam@167
|
84 E Te, T3y, Tk, T3A, Tc, Ti;
|
cannam@167
|
85 Tc = W[15];
|
cannam@167
|
86 Te = FMA(Tc, Td, Tb);
|
cannam@167
|
87 T3y = FNMS(Tc, Ta, T3x);
|
cannam@167
|
88 Ti = W[47];
|
cannam@167
|
89 Tk = FMA(Ti, Tj, Th);
|
cannam@167
|
90 T3A = FNMS(Ti, Tg, T3z);
|
cannam@167
|
91 Tl = Te + Tk;
|
cannam@167
|
92 T8y = Te - Tk;
|
cannam@167
|
93 T3B = T3y - T3A;
|
cannam@167
|
94 T83 = T3y + T3A;
|
cannam@167
|
95 }
|
cannam@167
|
96 }
|
cannam@167
|
97 {
|
cannam@167
|
98 E Ts, T3F, Ty, T3H, T3D, T3I;
|
cannam@167
|
99 {
|
cannam@167
|
100 E To, Tr, Tp, T3E, Tn, Tq;
|
cannam@167
|
101 To = Rp[WS(rs, 2)];
|
cannam@167
|
102 Tr = Rm[WS(rs, 2)];
|
cannam@167
|
103 Tn = W[6];
|
cannam@167
|
104 Tp = Tn * To;
|
cannam@167
|
105 T3E = Tn * Tr;
|
cannam@167
|
106 Tq = W[7];
|
cannam@167
|
107 Ts = FMA(Tq, Tr, Tp);
|
cannam@167
|
108 T3F = FNMS(Tq, To, T3E);
|
cannam@167
|
109 }
|
cannam@167
|
110 {
|
cannam@167
|
111 E Tu, Tx, Tv, T3G, Tt, Tw;
|
cannam@167
|
112 Tu = Rp[WS(rs, 10)];
|
cannam@167
|
113 Tx = Rm[WS(rs, 10)];
|
cannam@167
|
114 Tt = W[38];
|
cannam@167
|
115 Tv = Tt * Tu;
|
cannam@167
|
116 T3G = Tt * Tx;
|
cannam@167
|
117 Tw = W[39];
|
cannam@167
|
118 Ty = FMA(Tw, Tx, Tv);
|
cannam@167
|
119 T3H = FNMS(Tw, Tu, T3G);
|
cannam@167
|
120 }
|
cannam@167
|
121 Tz = Ts + Ty;
|
cannam@167
|
122 T6F = T3F + T3H;
|
cannam@167
|
123 T3D = Ts - Ty;
|
cannam@167
|
124 T3I = T3F - T3H;
|
cannam@167
|
125 T3J = T3D + T3I;
|
cannam@167
|
126 T5T = T3I - T3D;
|
cannam@167
|
127 }
|
cannam@167
|
128 {
|
cannam@167
|
129 E TF, T3M, TL, T3O, T3K, T3P;
|
cannam@167
|
130 {
|
cannam@167
|
131 E TB, TE, TC, T3L, TA, TD;
|
cannam@167
|
132 TB = Rp[WS(rs, 14)];
|
cannam@167
|
133 TE = Rm[WS(rs, 14)];
|
cannam@167
|
134 TA = W[54];
|
cannam@167
|
135 TC = TA * TB;
|
cannam@167
|
136 T3L = TA * TE;
|
cannam@167
|
137 TD = W[55];
|
cannam@167
|
138 TF = FMA(TD, TE, TC);
|
cannam@167
|
139 T3M = FNMS(TD, TB, T3L);
|
cannam@167
|
140 }
|
cannam@167
|
141 {
|
cannam@167
|
142 E TH, TK, TI, T3N, TG, TJ;
|
cannam@167
|
143 TH = Rp[WS(rs, 6)];
|
cannam@167
|
144 TK = Rm[WS(rs, 6)];
|
cannam@167
|
145 TG = W[22];
|
cannam@167
|
146 TI = TG * TH;
|
cannam@167
|
147 T3N = TG * TK;
|
cannam@167
|
148 TJ = W[23];
|
cannam@167
|
149 TL = FMA(TJ, TK, TI);
|
cannam@167
|
150 T3O = FNMS(TJ, TH, T3N);
|
cannam@167
|
151 }
|
cannam@167
|
152 TM = TF + TL;
|
cannam@167
|
153 T6G = T3M + T3O;
|
cannam@167
|
154 T3K = TF - TL;
|
cannam@167
|
155 T3P = T3M - T3O;
|
cannam@167
|
156 T3Q = T3K - T3P;
|
cannam@167
|
157 T5U = T3K + T3P;
|
cannam@167
|
158 }
|
cannam@167
|
159 {
|
cannam@167
|
160 E TU, T41, T1d, T3X, T10, T43, T17, T3V;
|
cannam@167
|
161 {
|
cannam@167
|
162 E TQ, TT, TR, T40, TP, TS;
|
cannam@167
|
163 TQ = Rp[WS(rs, 1)];
|
cannam@167
|
164 TT = Rm[WS(rs, 1)];
|
cannam@167
|
165 TP = W[2];
|
cannam@167
|
166 TR = TP * TQ;
|
cannam@167
|
167 T40 = TP * TT;
|
cannam@167
|
168 TS = W[3];
|
cannam@167
|
169 TU = FMA(TS, TT, TR);
|
cannam@167
|
170 T41 = FNMS(TS, TQ, T40);
|
cannam@167
|
171 }
|
cannam@167
|
172 {
|
cannam@167
|
173 E T19, T1c, T1a, T3W, T18, T1b;
|
cannam@167
|
174 T19 = Rp[WS(rs, 13)];
|
cannam@167
|
175 T1c = Rm[WS(rs, 13)];
|
cannam@167
|
176 T18 = W[50];
|
cannam@167
|
177 T1a = T18 * T19;
|
cannam@167
|
178 T3W = T18 * T1c;
|
cannam@167
|
179 T1b = W[51];
|
cannam@167
|
180 T1d = FMA(T1b, T1c, T1a);
|
cannam@167
|
181 T3X = FNMS(T1b, T19, T3W);
|
cannam@167
|
182 }
|
cannam@167
|
183 {
|
cannam@167
|
184 E TW, TZ, TX, T42, TV, TY;
|
cannam@167
|
185 TW = Rp[WS(rs, 9)];
|
cannam@167
|
186 TZ = Rm[WS(rs, 9)];
|
cannam@167
|
187 TV = W[34];
|
cannam@167
|
188 TX = TV * TW;
|
cannam@167
|
189 T42 = TV * TZ;
|
cannam@167
|
190 TY = W[35];
|
cannam@167
|
191 T10 = FMA(TY, TZ, TX);
|
cannam@167
|
192 T43 = FNMS(TY, TW, T42);
|
cannam@167
|
193 }
|
cannam@167
|
194 {
|
cannam@167
|
195 E T13, T16, T14, T3U, T12, T15;
|
cannam@167
|
196 T13 = Rp[WS(rs, 5)];
|
cannam@167
|
197 T16 = Rm[WS(rs, 5)];
|
cannam@167
|
198 T12 = W[18];
|
cannam@167
|
199 T14 = T12 * T13;
|
cannam@167
|
200 T3U = T12 * T16;
|
cannam@167
|
201 T15 = W[19];
|
cannam@167
|
202 T17 = FMA(T15, T16, T14);
|
cannam@167
|
203 T3V = FNMS(T15, T13, T3U);
|
cannam@167
|
204 }
|
cannam@167
|
205 T11 = TU + T10;
|
cannam@167
|
206 T1e = T17 + T1d;
|
cannam@167
|
207 T6M = T11 - T1e;
|
cannam@167
|
208 T6J = T41 + T43;
|
cannam@167
|
209 T6K = T3V + T3X;
|
cannam@167
|
210 T6L = T6J - T6K;
|
cannam@167
|
211 {
|
cannam@167
|
212 E T3T, T3Y, T44, T45;
|
cannam@167
|
213 T3T = TU - T10;
|
cannam@167
|
214 T3Y = T3V - T3X;
|
cannam@167
|
215 T3Z = T3T + T3Y;
|
cannam@167
|
216 T5Y = T3T - T3Y;
|
cannam@167
|
217 T44 = T41 - T43;
|
cannam@167
|
218 T45 = T17 - T1d;
|
cannam@167
|
219 T46 = T44 - T45;
|
cannam@167
|
220 T5X = T44 + T45;
|
cannam@167
|
221 }
|
cannam@167
|
222 }
|
cannam@167
|
223 {
|
cannam@167
|
224 E T1l, T4g, T1E, T4c, T1r, T4i, T1y, T4a;
|
cannam@167
|
225 {
|
cannam@167
|
226 E T1h, T1k, T1i, T4f, T1g, T1j;
|
cannam@167
|
227 T1h = Rp[WS(rs, 15)];
|
cannam@167
|
228 T1k = Rm[WS(rs, 15)];
|
cannam@167
|
229 T1g = W[58];
|
cannam@167
|
230 T1i = T1g * T1h;
|
cannam@167
|
231 T4f = T1g * T1k;
|
cannam@167
|
232 T1j = W[59];
|
cannam@167
|
233 T1l = FMA(T1j, T1k, T1i);
|
cannam@167
|
234 T4g = FNMS(T1j, T1h, T4f);
|
cannam@167
|
235 }
|
cannam@167
|
236 {
|
cannam@167
|
237 E T1A, T1D, T1B, T4b, T1z, T1C;
|
cannam@167
|
238 T1A = Rp[WS(rs, 11)];
|
cannam@167
|
239 T1D = Rm[WS(rs, 11)];
|
cannam@167
|
240 T1z = W[42];
|
cannam@167
|
241 T1B = T1z * T1A;
|
cannam@167
|
242 T4b = T1z * T1D;
|
cannam@167
|
243 T1C = W[43];
|
cannam@167
|
244 T1E = FMA(T1C, T1D, T1B);
|
cannam@167
|
245 T4c = FNMS(T1C, T1A, T4b);
|
cannam@167
|
246 }
|
cannam@167
|
247 {
|
cannam@167
|
248 E T1n, T1q, T1o, T4h, T1m, T1p;
|
cannam@167
|
249 T1n = Rp[WS(rs, 7)];
|
cannam@167
|
250 T1q = Rm[WS(rs, 7)];
|
cannam@167
|
251 T1m = W[26];
|
cannam@167
|
252 T1o = T1m * T1n;
|
cannam@167
|
253 T4h = T1m * T1q;
|
cannam@167
|
254 T1p = W[27];
|
cannam@167
|
255 T1r = FMA(T1p, T1q, T1o);
|
cannam@167
|
256 T4i = FNMS(T1p, T1n, T4h);
|
cannam@167
|
257 }
|
cannam@167
|
258 {
|
cannam@167
|
259 E T1u, T1x, T1v, T49, T1t, T1w;
|
cannam@167
|
260 T1u = Rp[WS(rs, 3)];
|
cannam@167
|
261 T1x = Rm[WS(rs, 3)];
|
cannam@167
|
262 T1t = W[10];
|
cannam@167
|
263 T1v = T1t * T1u;
|
cannam@167
|
264 T49 = T1t * T1x;
|
cannam@167
|
265 T1w = W[11];
|
cannam@167
|
266 T1y = FMA(T1w, T1x, T1v);
|
cannam@167
|
267 T4a = FNMS(T1w, T1u, T49);
|
cannam@167
|
268 }
|
cannam@167
|
269 T1s = T1l + T1r;
|
cannam@167
|
270 T1F = T1y + T1E;
|
cannam@167
|
271 T6O = T1s - T1F;
|
cannam@167
|
272 T6P = T4g + T4i;
|
cannam@167
|
273 T6Q = T4a + T4c;
|
cannam@167
|
274 T6R = T6P - T6Q;
|
cannam@167
|
275 {
|
cannam@167
|
276 E T48, T4d, T4j, T4k;
|
cannam@167
|
277 T48 = T1l - T1r;
|
cannam@167
|
278 T4d = T4a - T4c;
|
cannam@167
|
279 T4e = T48 + T4d;
|
cannam@167
|
280 T61 = T48 - T4d;
|
cannam@167
|
281 T4j = T4g - T4i;
|
cannam@167
|
282 T4k = T1y - T1E;
|
cannam@167
|
283 T4l = T4j - T4k;
|
cannam@167
|
284 T60 = T4j + T4k;
|
cannam@167
|
285 }
|
cannam@167
|
286 }
|
cannam@167
|
287 {
|
cannam@167
|
288 E T2H, T5m, T30, T52, T2N, T5o, T2U, T50;
|
cannam@167
|
289 {
|
cannam@167
|
290 E T2D, T2G, T2E, T5l, T2C, T2F;
|
cannam@167
|
291 T2D = Ip[WS(rs, 15)];
|
cannam@167
|
292 T2G = Im[WS(rs, 15)];
|
cannam@167
|
293 T2C = W[60];
|
cannam@167
|
294 T2E = T2C * T2D;
|
cannam@167
|
295 T5l = T2C * T2G;
|
cannam@167
|
296 T2F = W[61];
|
cannam@167
|
297 T2H = FMA(T2F, T2G, T2E);
|
cannam@167
|
298 T5m = FNMS(T2F, T2D, T5l);
|
cannam@167
|
299 }
|
cannam@167
|
300 {
|
cannam@167
|
301 E T2W, T2Z, T2X, T51, T2V, T2Y;
|
cannam@167
|
302 T2W = Ip[WS(rs, 11)];
|
cannam@167
|
303 T2Z = Im[WS(rs, 11)];
|
cannam@167
|
304 T2V = W[44];
|
cannam@167
|
305 T2X = T2V * T2W;
|
cannam@167
|
306 T51 = T2V * T2Z;
|
cannam@167
|
307 T2Y = W[45];
|
cannam@167
|
308 T30 = FMA(T2Y, T2Z, T2X);
|
cannam@167
|
309 T52 = FNMS(T2Y, T2W, T51);
|
cannam@167
|
310 }
|
cannam@167
|
311 {
|
cannam@167
|
312 E T2J, T2M, T2K, T5n, T2I, T2L;
|
cannam@167
|
313 T2J = Ip[WS(rs, 7)];
|
cannam@167
|
314 T2M = Im[WS(rs, 7)];
|
cannam@167
|
315 T2I = W[28];
|
cannam@167
|
316 T2K = T2I * T2J;
|
cannam@167
|
317 T5n = T2I * T2M;
|
cannam@167
|
318 T2L = W[29];
|
cannam@167
|
319 T2N = FMA(T2L, T2M, T2K);
|
cannam@167
|
320 T5o = FNMS(T2L, T2J, T5n);
|
cannam@167
|
321 }
|
cannam@167
|
322 {
|
cannam@167
|
323 E T2Q, T2T, T2R, T4Z, T2P, T2S;
|
cannam@167
|
324 T2Q = Ip[WS(rs, 3)];
|
cannam@167
|
325 T2T = Im[WS(rs, 3)];
|
cannam@167
|
326 T2P = W[12];
|
cannam@167
|
327 T2R = T2P * T2Q;
|
cannam@167
|
328 T4Z = T2P * T2T;
|
cannam@167
|
329 T2S = W[13];
|
cannam@167
|
330 T2U = FMA(T2S, T2T, T2R);
|
cannam@167
|
331 T50 = FNMS(T2S, T2Q, T4Z);
|
cannam@167
|
332 }
|
cannam@167
|
333 {
|
cannam@167
|
334 E T2O, T31, T77, T78;
|
cannam@167
|
335 T2O = T2H + T2N;
|
cannam@167
|
336 T31 = T2U + T30;
|
cannam@167
|
337 T32 = T2O + T31;
|
cannam@167
|
338 T7b = T2O - T31;
|
cannam@167
|
339 T77 = T5m + T5o;
|
cannam@167
|
340 T78 = T50 + T52;
|
cannam@167
|
341 T79 = T77 - T78;
|
cannam@167
|
342 T7N = T77 + T78;
|
cannam@167
|
343 }
|
cannam@167
|
344 {
|
cannam@167
|
345 E T4Y, T53, T5p, T5q;
|
cannam@167
|
346 T4Y = T2H - T2N;
|
cannam@167
|
347 T53 = T50 - T52;
|
cannam@167
|
348 T54 = T4Y + T53;
|
cannam@167
|
349 T6c = T4Y - T53;
|
cannam@167
|
350 T5p = T5m - T5o;
|
cannam@167
|
351 T5q = T30 - T2U;
|
cannam@167
|
352 T5r = T5p + T5q;
|
cannam@167
|
353 T6f = T5q - T5p;
|
cannam@167
|
354 }
|
cannam@167
|
355 }
|
cannam@167
|
356 {
|
cannam@167
|
357 E T1O, T4N, T27, T4t, T1U, T4P, T21, T4r;
|
cannam@167
|
358 {
|
cannam@167
|
359 E T1K, T1N, T1L, T4M, T1J, T1M;
|
cannam@167
|
360 T1K = Ip[0];
|
cannam@167
|
361 T1N = Im[0];
|
cannam@167
|
362 T1J = W[0];
|
cannam@167
|
363 T1L = T1J * T1K;
|
cannam@167
|
364 T4M = T1J * T1N;
|
cannam@167
|
365 T1M = W[1];
|
cannam@167
|
366 T1O = FMA(T1M, T1N, T1L);
|
cannam@167
|
367 T4N = FNMS(T1M, T1K, T4M);
|
cannam@167
|
368 }
|
cannam@167
|
369 {
|
cannam@167
|
370 E T23, T26, T24, T4s, T22, T25;
|
cannam@167
|
371 T23 = Ip[WS(rs, 12)];
|
cannam@167
|
372 T26 = Im[WS(rs, 12)];
|
cannam@167
|
373 T22 = W[48];
|
cannam@167
|
374 T24 = T22 * T23;
|
cannam@167
|
375 T4s = T22 * T26;
|
cannam@167
|
376 T25 = W[49];
|
cannam@167
|
377 T27 = FMA(T25, T26, T24);
|
cannam@167
|
378 T4t = FNMS(T25, T23, T4s);
|
cannam@167
|
379 }
|
cannam@167
|
380 {
|
cannam@167
|
381 E T1Q, T1T, T1R, T4O, T1P, T1S;
|
cannam@167
|
382 T1Q = Ip[WS(rs, 8)];
|
cannam@167
|
383 T1T = Im[WS(rs, 8)];
|
cannam@167
|
384 T1P = W[32];
|
cannam@167
|
385 T1R = T1P * T1Q;
|
cannam@167
|
386 T4O = T1P * T1T;
|
cannam@167
|
387 T1S = W[33];
|
cannam@167
|
388 T1U = FMA(T1S, T1T, T1R);
|
cannam@167
|
389 T4P = FNMS(T1S, T1Q, T4O);
|
cannam@167
|
390 }
|
cannam@167
|
391 {
|
cannam@167
|
392 E T1X, T20, T1Y, T4q, T1W, T1Z;
|
cannam@167
|
393 T1X = Ip[WS(rs, 4)];
|
cannam@167
|
394 T20 = Im[WS(rs, 4)];
|
cannam@167
|
395 T1W = W[16];
|
cannam@167
|
396 T1Y = T1W * T1X;
|
cannam@167
|
397 T4q = T1W * T20;
|
cannam@167
|
398 T1Z = W[17];
|
cannam@167
|
399 T21 = FMA(T1Z, T20, T1Y);
|
cannam@167
|
400 T4r = FNMS(T1Z, T1X, T4q);
|
cannam@167
|
401 }
|
cannam@167
|
402 {
|
cannam@167
|
403 E T1V, T28, T6V, T6W;
|
cannam@167
|
404 T1V = T1O + T1U;
|
cannam@167
|
405 T28 = T21 + T27;
|
cannam@167
|
406 T29 = T1V + T28;
|
cannam@167
|
407 T70 = T1V - T28;
|
cannam@167
|
408 T6V = T4N + T4P;
|
cannam@167
|
409 T6W = T4r + T4t;
|
cannam@167
|
410 T6X = T6V - T6W;
|
cannam@167
|
411 T7I = T6V + T6W;
|
cannam@167
|
412 }
|
cannam@167
|
413 {
|
cannam@167
|
414 E T4p, T4u, T4Q, T4R;
|
cannam@167
|
415 T4p = T1O - T1U;
|
cannam@167
|
416 T4u = T4r - T4t;
|
cannam@167
|
417 T4v = T4p + T4u;
|
cannam@167
|
418 T65 = T4p - T4u;
|
cannam@167
|
419 T4Q = T4N - T4P;
|
cannam@167
|
420 T4R = T21 - T27;
|
cannam@167
|
421 T4S = T4Q - T4R;
|
cannam@167
|
422 T68 = T4Q + T4R;
|
cannam@167
|
423 }
|
cannam@167
|
424 }
|
cannam@167
|
425 {
|
cannam@167
|
426 E T38, T57, T3r, T5g, T3e, T59, T3l, T5e;
|
cannam@167
|
427 {
|
cannam@167
|
428 E T34, T37, T35, T56, T33, T36;
|
cannam@167
|
429 T34 = Ip[WS(rs, 1)];
|
cannam@167
|
430 T37 = Im[WS(rs, 1)];
|
cannam@167
|
431 T33 = W[4];
|
cannam@167
|
432 T35 = T33 * T34;
|
cannam@167
|
433 T56 = T33 * T37;
|
cannam@167
|
434 T36 = W[5];
|
cannam@167
|
435 T38 = FMA(T36, T37, T35);
|
cannam@167
|
436 T57 = FNMS(T36, T34, T56);
|
cannam@167
|
437 }
|
cannam@167
|
438 {
|
cannam@167
|
439 E T3n, T3q, T3o, T5f, T3m, T3p;
|
cannam@167
|
440 T3n = Ip[WS(rs, 5)];
|
cannam@167
|
441 T3q = Im[WS(rs, 5)];
|
cannam@167
|
442 T3m = W[20];
|
cannam@167
|
443 T3o = T3m * T3n;
|
cannam@167
|
444 T5f = T3m * T3q;
|
cannam@167
|
445 T3p = W[21];
|
cannam@167
|
446 T3r = FMA(T3p, T3q, T3o);
|
cannam@167
|
447 T5g = FNMS(T3p, T3n, T5f);
|
cannam@167
|
448 }
|
cannam@167
|
449 {
|
cannam@167
|
450 E T3a, T3d, T3b, T58, T39, T3c;
|
cannam@167
|
451 T3a = Ip[WS(rs, 9)];
|
cannam@167
|
452 T3d = Im[WS(rs, 9)];
|
cannam@167
|
453 T39 = W[36];
|
cannam@167
|
454 T3b = T39 * T3a;
|
cannam@167
|
455 T58 = T39 * T3d;
|
cannam@167
|
456 T3c = W[37];
|
cannam@167
|
457 T3e = FMA(T3c, T3d, T3b);
|
cannam@167
|
458 T59 = FNMS(T3c, T3a, T58);
|
cannam@167
|
459 }
|
cannam@167
|
460 {
|
cannam@167
|
461 E T3h, T3k, T3i, T5d, T3g, T3j;
|
cannam@167
|
462 T3h = Ip[WS(rs, 13)];
|
cannam@167
|
463 T3k = Im[WS(rs, 13)];
|
cannam@167
|
464 T3g = W[52];
|
cannam@167
|
465 T3i = T3g * T3h;
|
cannam@167
|
466 T5d = T3g * T3k;
|
cannam@167
|
467 T3j = W[53];
|
cannam@167
|
468 T3l = FMA(T3j, T3k, T3i);
|
cannam@167
|
469 T5e = FNMS(T3j, T3h, T5d);
|
cannam@167
|
470 }
|
cannam@167
|
471 {
|
cannam@167
|
472 E T3f, T3s, T7c, T7d;
|
cannam@167
|
473 T3f = T38 + T3e;
|
cannam@167
|
474 T3s = T3l + T3r;
|
cannam@167
|
475 T3t = T3f + T3s;
|
cannam@167
|
476 T76 = T3s - T3f;
|
cannam@167
|
477 T7c = T57 + T59;
|
cannam@167
|
478 T7d = T5e + T5g;
|
cannam@167
|
479 T7e = T7c - T7d;
|
cannam@167
|
480 T7O = T7c + T7d;
|
cannam@167
|
481 }
|
cannam@167
|
482 {
|
cannam@167
|
483 E T55, T5a, T5c, T5h;
|
cannam@167
|
484 T55 = T38 - T3e;
|
cannam@167
|
485 T5a = T57 - T59;
|
cannam@167
|
486 T5b = T55 + T5a;
|
cannam@167
|
487 T5s = T5a - T55;
|
cannam@167
|
488 T5c = T3l - T3r;
|
cannam@167
|
489 T5h = T5e - T5g;
|
cannam@167
|
490 T5i = T5c - T5h;
|
cannam@167
|
491 T5t = T5c + T5h;
|
cannam@167
|
492 }
|
cannam@167
|
493 }
|
cannam@167
|
494 {
|
cannam@167
|
495 E T2f, T4y, T2y, T4H, T2l, T4A, T2s, T4F;
|
cannam@167
|
496 {
|
cannam@167
|
497 E T2b, T2e, T2c, T4x, T2a, T2d;
|
cannam@167
|
498 T2b = Ip[WS(rs, 2)];
|
cannam@167
|
499 T2e = Im[WS(rs, 2)];
|
cannam@167
|
500 T2a = W[8];
|
cannam@167
|
501 T2c = T2a * T2b;
|
cannam@167
|
502 T4x = T2a * T2e;
|
cannam@167
|
503 T2d = W[9];
|
cannam@167
|
504 T2f = FMA(T2d, T2e, T2c);
|
cannam@167
|
505 T4y = FNMS(T2d, T2b, T4x);
|
cannam@167
|
506 }
|
cannam@167
|
507 {
|
cannam@167
|
508 E T2u, T2x, T2v, T4G, T2t, T2w;
|
cannam@167
|
509 T2u = Ip[WS(rs, 6)];
|
cannam@167
|
510 T2x = Im[WS(rs, 6)];
|
cannam@167
|
511 T2t = W[24];
|
cannam@167
|
512 T2v = T2t * T2u;
|
cannam@167
|
513 T4G = T2t * T2x;
|
cannam@167
|
514 T2w = W[25];
|
cannam@167
|
515 T2y = FMA(T2w, T2x, T2v);
|
cannam@167
|
516 T4H = FNMS(T2w, T2u, T4G);
|
cannam@167
|
517 }
|
cannam@167
|
518 {
|
cannam@167
|
519 E T2h, T2k, T2i, T4z, T2g, T2j;
|
cannam@167
|
520 T2h = Ip[WS(rs, 10)];
|
cannam@167
|
521 T2k = Im[WS(rs, 10)];
|
cannam@167
|
522 T2g = W[40];
|
cannam@167
|
523 T2i = T2g * T2h;
|
cannam@167
|
524 T4z = T2g * T2k;
|
cannam@167
|
525 T2j = W[41];
|
cannam@167
|
526 T2l = FMA(T2j, T2k, T2i);
|
cannam@167
|
527 T4A = FNMS(T2j, T2h, T4z);
|
cannam@167
|
528 }
|
cannam@167
|
529 {
|
cannam@167
|
530 E T2o, T2r, T2p, T4E, T2n, T2q;
|
cannam@167
|
531 T2o = Ip[WS(rs, 14)];
|
cannam@167
|
532 T2r = Im[WS(rs, 14)];
|
cannam@167
|
533 T2n = W[56];
|
cannam@167
|
534 T2p = T2n * T2o;
|
cannam@167
|
535 T4E = T2n * T2r;
|
cannam@167
|
536 T2q = W[57];
|
cannam@167
|
537 T2s = FMA(T2q, T2r, T2p);
|
cannam@167
|
538 T4F = FNMS(T2q, T2o, T4E);
|
cannam@167
|
539 }
|
cannam@167
|
540 {
|
cannam@167
|
541 E T2m, T2z, T71, T72;
|
cannam@167
|
542 T2m = T2f + T2l;
|
cannam@167
|
543 T2z = T2s + T2y;
|
cannam@167
|
544 T2A = T2m + T2z;
|
cannam@167
|
545 T6Y = T2z - T2m;
|
cannam@167
|
546 T71 = T4y + T4A;
|
cannam@167
|
547 T72 = T4F + T4H;
|
cannam@167
|
548 T73 = T71 - T72;
|
cannam@167
|
549 T7J = T71 + T72;
|
cannam@167
|
550 }
|
cannam@167
|
551 {
|
cannam@167
|
552 E T4w, T4B, T4D, T4I;
|
cannam@167
|
553 T4w = T2f - T2l;
|
cannam@167
|
554 T4B = T4y - T4A;
|
cannam@167
|
555 T4C = T4w + T4B;
|
cannam@167
|
556 T4T = T4B - T4w;
|
cannam@167
|
557 T4D = T2s - T2y;
|
cannam@167
|
558 T4I = T4F - T4H;
|
cannam@167
|
559 T4J = T4D - T4I;
|
cannam@167
|
560 T4U = T4D + T4I;
|
cannam@167
|
561 }
|
cannam@167
|
562 }
|
cannam@167
|
563 {
|
cannam@167
|
564 E TO, T7C, T7Z, T80, T89, T8e, T1H, T8d, T3v, T8b, T7L, T7T, T7Q, T7U, T7F;
|
cannam@167
|
565 E T81;
|
cannam@167
|
566 {
|
cannam@167
|
567 E Tm, TN, T7X, T7Y;
|
cannam@167
|
568 Tm = T8 + Tl;
|
cannam@167
|
569 TN = Tz + TM;
|
cannam@167
|
570 TO = Tm + TN;
|
cannam@167
|
571 T7C = Tm - TN;
|
cannam@167
|
572 T7X = T7I + T7J;
|
cannam@167
|
573 T7Y = T7N + T7O;
|
cannam@167
|
574 T7Z = T7X - T7Y;
|
cannam@167
|
575 T80 = T7X + T7Y;
|
cannam@167
|
576 }
|
cannam@167
|
577 {
|
cannam@167
|
578 E T82, T88, T1f, T1G;
|
cannam@167
|
579 T82 = T6F + T6G;
|
cannam@167
|
580 T88 = T83 + T87;
|
cannam@167
|
581 T89 = T82 + T88;
|
cannam@167
|
582 T8e = T88 - T82;
|
cannam@167
|
583 T1f = T11 + T1e;
|
cannam@167
|
584 T1G = T1s + T1F;
|
cannam@167
|
585 T1H = T1f + T1G;
|
cannam@167
|
586 T8d = T1G - T1f;
|
cannam@167
|
587 }
|
cannam@167
|
588 {
|
cannam@167
|
589 E T2B, T3u, T7H, T7K;
|
cannam@167
|
590 T2B = T29 + T2A;
|
cannam@167
|
591 T3u = T32 + T3t;
|
cannam@167
|
592 T3v = T2B + T3u;
|
cannam@167
|
593 T8b = T3u - T2B;
|
cannam@167
|
594 T7H = T29 - T2A;
|
cannam@167
|
595 T7K = T7I - T7J;
|
cannam@167
|
596 T7L = T7H + T7K;
|
cannam@167
|
597 T7T = T7K - T7H;
|
cannam@167
|
598 }
|
cannam@167
|
599 {
|
cannam@167
|
600 E T7M, T7P, T7D, T7E;
|
cannam@167
|
601 T7M = T32 - T3t;
|
cannam@167
|
602 T7P = T7N - T7O;
|
cannam@167
|
603 T7Q = T7M - T7P;
|
cannam@167
|
604 T7U = T7M + T7P;
|
cannam@167
|
605 T7D = T6J + T6K;
|
cannam@167
|
606 T7E = T6P + T6Q;
|
cannam@167
|
607 T7F = T7D - T7E;
|
cannam@167
|
608 T81 = T7D + T7E;
|
cannam@167
|
609 }
|
cannam@167
|
610 {
|
cannam@167
|
611 E T1I, T8a, T7W, T8c;
|
cannam@167
|
612 T1I = TO + T1H;
|
cannam@167
|
613 Rm[WS(rs, 15)] = T1I - T3v;
|
cannam@167
|
614 Rp[0] = T1I + T3v;
|
cannam@167
|
615 T8a = T81 + T89;
|
cannam@167
|
616 Im[WS(rs, 15)] = T80 - T8a;
|
cannam@167
|
617 Ip[0] = T80 + T8a;
|
cannam@167
|
618 T7W = TO - T1H;
|
cannam@167
|
619 Rm[WS(rs, 7)] = T7W - T7Z;
|
cannam@167
|
620 Rp[WS(rs, 8)] = T7W + T7Z;
|
cannam@167
|
621 T8c = T89 - T81;
|
cannam@167
|
622 Im[WS(rs, 7)] = T8b - T8c;
|
cannam@167
|
623 Ip[WS(rs, 8)] = T8b + T8c;
|
cannam@167
|
624 }
|
cannam@167
|
625 {
|
cannam@167
|
626 E T7G, T7R, T8f, T8g;
|
cannam@167
|
627 T7G = T7C + T7F;
|
cannam@167
|
628 T7R = T7L + T7Q;
|
cannam@167
|
629 Rm[WS(rs, 11)] = FNMS(KP707106781, T7R, T7G);
|
cannam@167
|
630 Rp[WS(rs, 4)] = FMA(KP707106781, T7R, T7G);
|
cannam@167
|
631 T8f = T8d + T8e;
|
cannam@167
|
632 T8g = T7T + T7U;
|
cannam@167
|
633 Im[WS(rs, 11)] = FMS(KP707106781, T8g, T8f);
|
cannam@167
|
634 Ip[WS(rs, 4)] = FMA(KP707106781, T8g, T8f);
|
cannam@167
|
635 }
|
cannam@167
|
636 {
|
cannam@167
|
637 E T7S, T7V, T8h, T8i;
|
cannam@167
|
638 T7S = T7C - T7F;
|
cannam@167
|
639 T7V = T7T - T7U;
|
cannam@167
|
640 Rm[WS(rs, 3)] = FNMS(KP707106781, T7V, T7S);
|
cannam@167
|
641 Rp[WS(rs, 12)] = FMA(KP707106781, T7V, T7S);
|
cannam@167
|
642 T8h = T8e - T8d;
|
cannam@167
|
643 T8i = T7Q - T7L;
|
cannam@167
|
644 Im[WS(rs, 3)] = FMS(KP707106781, T8i, T8h);
|
cannam@167
|
645 Ip[WS(rs, 12)] = FMA(KP707106781, T8i, T8h);
|
cannam@167
|
646 }
|
cannam@167
|
647 }
|
cannam@167
|
648 {
|
cannam@167
|
649 E T6I, T7m, T7w, T7A, T8l, T8r, T6T, T8m, T75, T7j, T7p, T8s, T7t, T7z, T7g;
|
cannam@167
|
650 E T7k;
|
cannam@167
|
651 {
|
cannam@167
|
652 E T6E, T6H, T7u, T7v;
|
cannam@167
|
653 T6E = T8 - Tl;
|
cannam@167
|
654 T6H = T6F - T6G;
|
cannam@167
|
655 T6I = T6E - T6H;
|
cannam@167
|
656 T7m = T6E + T6H;
|
cannam@167
|
657 T7u = T7b + T7e;
|
cannam@167
|
658 T7v = T79 + T76;
|
cannam@167
|
659 T7w = FNMS(KP414213562, T7v, T7u);
|
cannam@167
|
660 T7A = FMA(KP414213562, T7u, T7v);
|
cannam@167
|
661 }
|
cannam@167
|
662 {
|
cannam@167
|
663 E T8j, T8k, T6N, T6S;
|
cannam@167
|
664 T8j = TM - Tz;
|
cannam@167
|
665 T8k = T87 - T83;
|
cannam@167
|
666 T8l = T8j + T8k;
|
cannam@167
|
667 T8r = T8k - T8j;
|
cannam@167
|
668 T6N = T6L - T6M;
|
cannam@167
|
669 T6S = T6O + T6R;
|
cannam@167
|
670 T6T = T6N - T6S;
|
cannam@167
|
671 T8m = T6N + T6S;
|
cannam@167
|
672 }
|
cannam@167
|
673 {
|
cannam@167
|
674 E T6Z, T74, T7n, T7o;
|
cannam@167
|
675 T6Z = T6X - T6Y;
|
cannam@167
|
676 T74 = T70 - T73;
|
cannam@167
|
677 T75 = FMA(KP414213562, T74, T6Z);
|
cannam@167
|
678 T7j = FNMS(KP414213562, T6Z, T74);
|
cannam@167
|
679 T7n = T6M + T6L;
|
cannam@167
|
680 T7o = T6O - T6R;
|
cannam@167
|
681 T7p = T7n + T7o;
|
cannam@167
|
682 T8s = T7o - T7n;
|
cannam@167
|
683 }
|
cannam@167
|
684 {
|
cannam@167
|
685 E T7r, T7s, T7a, T7f;
|
cannam@167
|
686 T7r = T70 + T73;
|
cannam@167
|
687 T7s = T6X + T6Y;
|
cannam@167
|
688 T7t = FMA(KP414213562, T7s, T7r);
|
cannam@167
|
689 T7z = FNMS(KP414213562, T7r, T7s);
|
cannam@167
|
690 T7a = T76 - T79;
|
cannam@167
|
691 T7f = T7b - T7e;
|
cannam@167
|
692 T7g = FMA(KP414213562, T7f, T7a);
|
cannam@167
|
693 T7k = FNMS(KP414213562, T7a, T7f);
|
cannam@167
|
694 }
|
cannam@167
|
695 {
|
cannam@167
|
696 E T6U, T7h, T8t, T8u;
|
cannam@167
|
697 T6U = FMA(KP707106781, T6T, T6I);
|
cannam@167
|
698 T7h = T75 + T7g;
|
cannam@167
|
699 Rm[WS(rs, 9)] = FNMS(KP923879532, T7h, T6U);
|
cannam@167
|
700 Rp[WS(rs, 6)] = FMA(KP923879532, T7h, T6U);
|
cannam@167
|
701 T8t = FMA(KP707106781, T8s, T8r);
|
cannam@167
|
702 T8u = T7k - T7j;
|
cannam@167
|
703 Im[WS(rs, 9)] = FMS(KP923879532, T8u, T8t);
|
cannam@167
|
704 Ip[WS(rs, 6)] = FMA(KP923879532, T8u, T8t);
|
cannam@167
|
705 }
|
cannam@167
|
706 {
|
cannam@167
|
707 E T7i, T7l, T8v, T8w;
|
cannam@167
|
708 T7i = FNMS(KP707106781, T6T, T6I);
|
cannam@167
|
709 T7l = T7j + T7k;
|
cannam@167
|
710 Rp[WS(rs, 14)] = FNMS(KP923879532, T7l, T7i);
|
cannam@167
|
711 Rm[WS(rs, 1)] = FMA(KP923879532, T7l, T7i);
|
cannam@167
|
712 T8v = FNMS(KP707106781, T8s, T8r);
|
cannam@167
|
713 T8w = T7g - T75;
|
cannam@167
|
714 Im[WS(rs, 1)] = FMS(KP923879532, T8w, T8v);
|
cannam@167
|
715 Ip[WS(rs, 14)] = FMA(KP923879532, T8w, T8v);
|
cannam@167
|
716 }
|
cannam@167
|
717 {
|
cannam@167
|
718 E T7q, T7x, T8n, T8o;
|
cannam@167
|
719 T7q = FMA(KP707106781, T7p, T7m);
|
cannam@167
|
720 T7x = T7t + T7w;
|
cannam@167
|
721 Rm[WS(rs, 13)] = FNMS(KP923879532, T7x, T7q);
|
cannam@167
|
722 Rp[WS(rs, 2)] = FMA(KP923879532, T7x, T7q);
|
cannam@167
|
723 T8n = FMA(KP707106781, T8m, T8l);
|
cannam@167
|
724 T8o = T7z + T7A;
|
cannam@167
|
725 Im[WS(rs, 13)] = FMS(KP923879532, T8o, T8n);
|
cannam@167
|
726 Ip[WS(rs, 2)] = FMA(KP923879532, T8o, T8n);
|
cannam@167
|
727 }
|
cannam@167
|
728 {
|
cannam@167
|
729 E T7y, T7B, T8p, T8q;
|
cannam@167
|
730 T7y = FNMS(KP707106781, T7p, T7m);
|
cannam@167
|
731 T7B = T7z - T7A;
|
cannam@167
|
732 Rm[WS(rs, 5)] = FNMS(KP923879532, T7B, T7y);
|
cannam@167
|
733 Rp[WS(rs, 10)] = FMA(KP923879532, T7B, T7y);
|
cannam@167
|
734 T8p = FNMS(KP707106781, T8m, T8l);
|
cannam@167
|
735 T8q = T7w - T7t;
|
cannam@167
|
736 Im[WS(rs, 5)] = FMS(KP923879532, T8q, T8p);
|
cannam@167
|
737 Ip[WS(rs, 10)] = FMA(KP923879532, T8q, T8p);
|
cannam@167
|
738 }
|
cannam@167
|
739 }
|
cannam@167
|
740 {
|
cannam@167
|
741 E T3S, T5C, T4n, T8I, T8B, T8H, T5F, T8C, T5w, T5Q, T5A, T5M, T4X, T5P, T5z;
|
cannam@167
|
742 E T5J;
|
cannam@167
|
743 {
|
cannam@167
|
744 E T3C, T3R, T5D, T5E;
|
cannam@167
|
745 T3C = T3w + T3B;
|
cannam@167
|
746 T3R = T3J + T3Q;
|
cannam@167
|
747 T3S = FMA(KP707106781, T3R, T3C);
|
cannam@167
|
748 T5C = FNMS(KP707106781, T3R, T3C);
|
cannam@167
|
749 {
|
cannam@167
|
750 E T47, T4m, T8z, T8A;
|
cannam@167
|
751 T47 = FMA(KP414213562, T46, T3Z);
|
cannam@167
|
752 T4m = FNMS(KP414213562, T4l, T4e);
|
cannam@167
|
753 T4n = T47 + T4m;
|
cannam@167
|
754 T8I = T4m - T47;
|
cannam@167
|
755 T8z = T8x - T8y;
|
cannam@167
|
756 T8A = T5T + T5U;
|
cannam@167
|
757 T8B = FMA(KP707106781, T8A, T8z);
|
cannam@167
|
758 T8H = FNMS(KP707106781, T8A, T8z);
|
cannam@167
|
759 }
|
cannam@167
|
760 T5D = FNMS(KP414213562, T3Z, T46);
|
cannam@167
|
761 T5E = FMA(KP414213562, T4e, T4l);
|
cannam@167
|
762 T5F = T5D - T5E;
|
cannam@167
|
763 T8C = T5D + T5E;
|
cannam@167
|
764 {
|
cannam@167
|
765 E T5k, T5K, T5v, T5L, T5j, T5u;
|
cannam@167
|
766 T5j = T5b + T5i;
|
cannam@167
|
767 T5k = FMA(KP707106781, T5j, T54);
|
cannam@167
|
768 T5K = FNMS(KP707106781, T5j, T54);
|
cannam@167
|
769 T5u = T5s + T5t;
|
cannam@167
|
770 T5v = FMA(KP707106781, T5u, T5r);
|
cannam@167
|
771 T5L = FNMS(KP707106781, T5u, T5r);
|
cannam@167
|
772 T5w = FNMS(KP198912367, T5v, T5k);
|
cannam@167
|
773 T5Q = FNMS(KP668178637, T5K, T5L);
|
cannam@167
|
774 T5A = FMA(KP198912367, T5k, T5v);
|
cannam@167
|
775 T5M = FMA(KP668178637, T5L, T5K);
|
cannam@167
|
776 }
|
cannam@167
|
777 {
|
cannam@167
|
778 E T4L, T5H, T4W, T5I, T4K, T4V;
|
cannam@167
|
779 T4K = T4C + T4J;
|
cannam@167
|
780 T4L = FMA(KP707106781, T4K, T4v);
|
cannam@167
|
781 T5H = FNMS(KP707106781, T4K, T4v);
|
cannam@167
|
782 T4V = T4T + T4U;
|
cannam@167
|
783 T4W = FMA(KP707106781, T4V, T4S);
|
cannam@167
|
784 T5I = FNMS(KP707106781, T4V, T4S);
|
cannam@167
|
785 T4X = FMA(KP198912367, T4W, T4L);
|
cannam@167
|
786 T5P = FMA(KP668178637, T5H, T5I);
|
cannam@167
|
787 T5z = FNMS(KP198912367, T4L, T4W);
|
cannam@167
|
788 T5J = FNMS(KP668178637, T5I, T5H);
|
cannam@167
|
789 }
|
cannam@167
|
790 }
|
cannam@167
|
791 {
|
cannam@167
|
792 E T4o, T5x, T8D, T8E;
|
cannam@167
|
793 T4o = FMA(KP923879532, T4n, T3S);
|
cannam@167
|
794 T5x = T4X + T5w;
|
cannam@167
|
795 Rm[WS(rs, 14)] = FNMS(KP980785280, T5x, T4o);
|
cannam@167
|
796 Rp[WS(rs, 1)] = FMA(KP980785280, T5x, T4o);
|
cannam@167
|
797 T8D = FMA(KP923879532, T8C, T8B);
|
cannam@167
|
798 T8E = T5z + T5A;
|
cannam@167
|
799 Im[WS(rs, 14)] = FMS(KP980785280, T8E, T8D);
|
cannam@167
|
800 Ip[WS(rs, 1)] = FMA(KP980785280, T8E, T8D);
|
cannam@167
|
801 }
|
cannam@167
|
802 {
|
cannam@167
|
803 E T5y, T5B, T8F, T8G;
|
cannam@167
|
804 T5y = FNMS(KP923879532, T4n, T3S);
|
cannam@167
|
805 T5B = T5z - T5A;
|
cannam@167
|
806 Rm[WS(rs, 6)] = FNMS(KP980785280, T5B, T5y);
|
cannam@167
|
807 Rp[WS(rs, 9)] = FMA(KP980785280, T5B, T5y);
|
cannam@167
|
808 T8F = FNMS(KP923879532, T8C, T8B);
|
cannam@167
|
809 T8G = T5w - T4X;
|
cannam@167
|
810 Im[WS(rs, 6)] = FMS(KP980785280, T8G, T8F);
|
cannam@167
|
811 Ip[WS(rs, 9)] = FMA(KP980785280, T8G, T8F);
|
cannam@167
|
812 }
|
cannam@167
|
813 {
|
cannam@167
|
814 E T5G, T5N, T8L, T8M;
|
cannam@167
|
815 T5G = FNMS(KP923879532, T5F, T5C);
|
cannam@167
|
816 T5N = T5J + T5M;
|
cannam@167
|
817 Rp[WS(rs, 13)] = FNMS(KP831469612, T5N, T5G);
|
cannam@167
|
818 Rm[WS(rs, 2)] = FMA(KP831469612, T5N, T5G);
|
cannam@167
|
819 T8L = FNMS(KP923879532, T8I, T8H);
|
cannam@167
|
820 T8M = T5P + T5Q;
|
cannam@167
|
821 Im[WS(rs, 2)] = -(FMA(KP831469612, T8M, T8L));
|
cannam@167
|
822 Ip[WS(rs, 13)] = FNMS(KP831469612, T8M, T8L);
|
cannam@167
|
823 }
|
cannam@167
|
824 {
|
cannam@167
|
825 E T5O, T5R, T8J, T8K;
|
cannam@167
|
826 T5O = FMA(KP923879532, T5F, T5C);
|
cannam@167
|
827 T5R = T5P - T5Q;
|
cannam@167
|
828 Rm[WS(rs, 10)] = FNMS(KP831469612, T5R, T5O);
|
cannam@167
|
829 Rp[WS(rs, 5)] = FMA(KP831469612, T5R, T5O);
|
cannam@167
|
830 T8J = FMA(KP923879532, T8I, T8H);
|
cannam@167
|
831 T8K = T5M - T5J;
|
cannam@167
|
832 Im[WS(rs, 10)] = FMS(KP831469612, T8K, T8J);
|
cannam@167
|
833 Ip[WS(rs, 5)] = FMA(KP831469612, T8K, T8J);
|
cannam@167
|
834 }
|
cannam@167
|
835 }
|
cannam@167
|
836 {
|
cannam@167
|
837 E T5W, T6o, T63, T8W, T8P, T8V, T6r, T8Q, T6i, T6C, T6m, T6y, T6b, T6B, T6l;
|
cannam@167
|
838 E T6v;
|
cannam@167
|
839 {
|
cannam@167
|
840 E T5S, T5V, T6p, T6q;
|
cannam@167
|
841 T5S = T3w - T3B;
|
cannam@167
|
842 T5V = T5T - T5U;
|
cannam@167
|
843 T5W = FMA(KP707106781, T5V, T5S);
|
cannam@167
|
844 T6o = FNMS(KP707106781, T5V, T5S);
|
cannam@167
|
845 {
|
cannam@167
|
846 E T5Z, T62, T8N, T8O;
|
cannam@167
|
847 T5Z = FMA(KP414213562, T5Y, T5X);
|
cannam@167
|
848 T62 = FNMS(KP414213562, T61, T60);
|
cannam@167
|
849 T63 = T5Z - T62;
|
cannam@167
|
850 T8W = T5Z + T62;
|
cannam@167
|
851 T8N = T8y + T8x;
|
cannam@167
|
852 T8O = T3Q - T3J;
|
cannam@167
|
853 T8P = FMA(KP707106781, T8O, T8N);
|
cannam@167
|
854 T8V = FNMS(KP707106781, T8O, T8N);
|
cannam@167
|
855 }
|
cannam@167
|
856 T6p = FNMS(KP414213562, T5X, T5Y);
|
cannam@167
|
857 T6q = FMA(KP414213562, T60, T61);
|
cannam@167
|
858 T6r = T6p + T6q;
|
cannam@167
|
859 T8Q = T6q - T6p;
|
cannam@167
|
860 {
|
cannam@167
|
861 E T6e, T6x, T6h, T6w, T6d, T6g;
|
cannam@167
|
862 T6d = T5t - T5s;
|
cannam@167
|
863 T6e = FNMS(KP707106781, T6d, T6c);
|
cannam@167
|
864 T6x = FMA(KP707106781, T6d, T6c);
|
cannam@167
|
865 T6g = T5i - T5b;
|
cannam@167
|
866 T6h = FNMS(KP707106781, T6g, T6f);
|
cannam@167
|
867 T6w = FMA(KP707106781, T6g, T6f);
|
cannam@167
|
868 T6i = FMA(KP668178637, T6h, T6e);
|
cannam@167
|
869 T6C = FNMS(KP198912367, T6w, T6x);
|
cannam@167
|
870 T6m = FNMS(KP668178637, T6e, T6h);
|
cannam@167
|
871 T6y = FMA(KP198912367, T6x, T6w);
|
cannam@167
|
872 }
|
cannam@167
|
873 {
|
cannam@167
|
874 E T67, T6u, T6a, T6t, T66, T69;
|
cannam@167
|
875 T66 = T4U - T4T;
|
cannam@167
|
876 T67 = FNMS(KP707106781, T66, T65);
|
cannam@167
|
877 T6u = FMA(KP707106781, T66, T65);
|
cannam@167
|
878 T69 = T4C - T4J;
|
cannam@167
|
879 T6a = FNMS(KP707106781, T69, T68);
|
cannam@167
|
880 T6t = FMA(KP707106781, T69, T68);
|
cannam@167
|
881 T6b = FMA(KP668178637, T6a, T67);
|
cannam@167
|
882 T6B = FNMS(KP198912367, T6t, T6u);
|
cannam@167
|
883 T6l = FNMS(KP668178637, T67, T6a);
|
cannam@167
|
884 T6v = FMA(KP198912367, T6u, T6t);
|
cannam@167
|
885 }
|
cannam@167
|
886 }
|
cannam@167
|
887 {
|
cannam@167
|
888 E T64, T6j, T8R, T8S;
|
cannam@167
|
889 T64 = FMA(KP923879532, T63, T5W);
|
cannam@167
|
890 T6j = T6b + T6i;
|
cannam@167
|
891 Rm[WS(rs, 12)] = FNMS(KP831469612, T6j, T64);
|
cannam@167
|
892 Rp[WS(rs, 3)] = FMA(KP831469612, T6j, T64);
|
cannam@167
|
893 T8R = FMA(KP923879532, T8Q, T8P);
|
cannam@167
|
894 T8S = T6l - T6m;
|
cannam@167
|
895 Im[WS(rs, 12)] = FMS(KP831469612, T8S, T8R);
|
cannam@167
|
896 Ip[WS(rs, 3)] = FMA(KP831469612, T8S, T8R);
|
cannam@167
|
897 }
|
cannam@167
|
898 {
|
cannam@167
|
899 E T6k, T6n, T8T, T8U;
|
cannam@167
|
900 T6k = FNMS(KP923879532, T63, T5W);
|
cannam@167
|
901 T6n = T6l + T6m;
|
cannam@167
|
902 Rm[WS(rs, 4)] = FNMS(KP831469612, T6n, T6k);
|
cannam@167
|
903 Rp[WS(rs, 11)] = FMA(KP831469612, T6n, T6k);
|
cannam@167
|
904 T8T = FNMS(KP923879532, T8Q, T8P);
|
cannam@167
|
905 T8U = T6i - T6b;
|
cannam@167
|
906 Im[WS(rs, 4)] = FMS(KP831469612, T8U, T8T);
|
cannam@167
|
907 Ip[WS(rs, 11)] = FMA(KP831469612, T8U, T8T);
|
cannam@167
|
908 }
|
cannam@167
|
909 {
|
cannam@167
|
910 E T6s, T6z, T8X, T8Y;
|
cannam@167
|
911 T6s = FNMS(KP923879532, T6r, T6o);
|
cannam@167
|
912 T6z = T6v + T6y;
|
cannam@167
|
913 Rm[WS(rs, 8)] = FNMS(KP980785280, T6z, T6s);
|
cannam@167
|
914 Rp[WS(rs, 7)] = FMA(KP980785280, T6z, T6s);
|
cannam@167
|
915 T8X = FNMS(KP923879532, T8W, T8V);
|
cannam@167
|
916 T8Y = T6C - T6B;
|
cannam@167
|
917 Im[WS(rs, 8)] = FMS(KP980785280, T8Y, T8X);
|
cannam@167
|
918 Ip[WS(rs, 7)] = FMA(KP980785280, T8Y, T8X);
|
cannam@167
|
919 }
|
cannam@167
|
920 {
|
cannam@167
|
921 E T6A, T6D, T8Z, T90;
|
cannam@167
|
922 T6A = FMA(KP923879532, T6r, T6o);
|
cannam@167
|
923 T6D = T6B + T6C;
|
cannam@167
|
924 Rp[WS(rs, 15)] = FNMS(KP980785280, T6D, T6A);
|
cannam@167
|
925 Rm[0] = FMA(KP980785280, T6D, T6A);
|
cannam@167
|
926 T8Z = FMA(KP923879532, T8W, T8V);
|
cannam@167
|
927 T90 = T6y - T6v;
|
cannam@167
|
928 Im[0] = FMS(KP980785280, T90, T8Z);
|
cannam@167
|
929 Ip[WS(rs, 15)] = FMA(KP980785280, T90, T8Z);
|
cannam@167
|
930 }
|
cannam@167
|
931 }
|
cannam@167
|
932 }
|
cannam@167
|
933 }
|
cannam@167
|
934 }
|
cannam@167
|
935
|
cannam@167
|
936 static const tw_instr twinstr[] = {
|
cannam@167
|
937 {TW_FULL, 1, 32},
|
cannam@167
|
938 {TW_NEXT, 1, 0}
|
cannam@167
|
939 };
|
cannam@167
|
940
|
cannam@167
|
941 static const hc2c_desc desc = { 32, "hc2cf_32", twinstr, &GENUS, {236, 62, 198, 0} };
|
cannam@167
|
942
|
cannam@167
|
943 void X(codelet_hc2cf_32) (planner *p) {
|
cannam@167
|
944 X(khc2c_register) (p, hc2cf_32, &desc, HC2C_VIA_RDFT);
|
cannam@167
|
945 }
|
cannam@167
|
946 #else
|
cannam@167
|
947
|
cannam@167
|
948 /* Generated by: ../../../genfft/gen_hc2c.native -compact -variables 4 -pipeline-latency 4 -n 32 -dit -name hc2cf_32 -include rdft/scalar/hc2cf.h */
|
cannam@167
|
949
|
cannam@167
|
950 /*
|
cannam@167
|
951 * This function contains 434 FP additions, 208 FP multiplications,
|
cannam@167
|
952 * (or, 340 additions, 114 multiplications, 94 fused multiply/add),
|
cannam@167
|
953 * 96 stack variables, 7 constants, and 128 memory accesses
|
cannam@167
|
954 */
|
cannam@167
|
955 #include "rdft/scalar/hc2cf.h"
|
cannam@167
|
956
|
cannam@167
|
957 static void hc2cf_32(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms)
|
cannam@167
|
958 {
|
cannam@167
|
959 DK(KP195090322, +0.195090322016128267848284868477022240927691618);
|
cannam@167
|
960 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
cannam@167
|
961 DK(KP555570233, +0.555570233019602224742830813948532874374937191);
|
cannam@167
|
962 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
cannam@167
|
963 DK(KP382683432, +0.382683432365089771728459984030398866761344562);
|
cannam@167
|
964 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
cannam@167
|
965 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
cannam@167
|
966 {
|
cannam@167
|
967 INT m;
|
cannam@167
|
968 for (m = mb, W = W + ((mb - 1) * 62); m < me; m = m + 1, Rp = Rp + ms, Ip = Ip + ms, Rm = Rm - ms, Im = Im - ms, W = W + 62, MAKE_VOLATILE_STRIDE(128, rs)) {
|
cannam@167
|
969 E Tj, T5F, T7C, T7Q, T35, T4T, T78, T7m, T1Q, T61, T5Y, T6J, T3K, T59, T41;
|
cannam@167
|
970 E T56, T2B, T67, T6e, T6O, T4b, T5d, T4s, T5g, TG, T7l, T5I, T73, T3a, T4U;
|
cannam@167
|
971 E T3f, T4V, T14, T5N, T5M, T6E, T3m, T4Y, T3r, T4Z, T1r, T5P, T5S, T6F, T3x;
|
cannam@167
|
972 E T51, T3C, T52, T2d, T5Z, T64, T6K, T3V, T57, T44, T5a, T2Y, T6f, T6a, T6P;
|
cannam@167
|
973 E T4m, T5h, T4v, T5e;
|
cannam@167
|
974 {
|
cannam@167
|
975 E T1, T76, T6, T75, Tc, T32, Th, T33;
|
cannam@167
|
976 T1 = Rp[0];
|
cannam@167
|
977 T76 = Rm[0];
|
cannam@167
|
978 {
|
cannam@167
|
979 E T3, T5, T2, T4;
|
cannam@167
|
980 T3 = Rp[WS(rs, 8)];
|
cannam@167
|
981 T5 = Rm[WS(rs, 8)];
|
cannam@167
|
982 T2 = W[30];
|
cannam@167
|
983 T4 = W[31];
|
cannam@167
|
984 T6 = FMA(T2, T3, T4 * T5);
|
cannam@167
|
985 T75 = FNMS(T4, T3, T2 * T5);
|
cannam@167
|
986 }
|
cannam@167
|
987 {
|
cannam@167
|
988 E T9, Tb, T8, Ta;
|
cannam@167
|
989 T9 = Rp[WS(rs, 4)];
|
cannam@167
|
990 Tb = Rm[WS(rs, 4)];
|
cannam@167
|
991 T8 = W[14];
|
cannam@167
|
992 Ta = W[15];
|
cannam@167
|
993 Tc = FMA(T8, T9, Ta * Tb);
|
cannam@167
|
994 T32 = FNMS(Ta, T9, T8 * Tb);
|
cannam@167
|
995 }
|
cannam@167
|
996 {
|
cannam@167
|
997 E Te, Tg, Td, Tf;
|
cannam@167
|
998 Te = Rp[WS(rs, 12)];
|
cannam@167
|
999 Tg = Rm[WS(rs, 12)];
|
cannam@167
|
1000 Td = W[46];
|
cannam@167
|
1001 Tf = W[47];
|
cannam@167
|
1002 Th = FMA(Td, Te, Tf * Tg);
|
cannam@167
|
1003 T33 = FNMS(Tf, Te, Td * Tg);
|
cannam@167
|
1004 }
|
cannam@167
|
1005 {
|
cannam@167
|
1006 E T7, Ti, T7A, T7B;
|
cannam@167
|
1007 T7 = T1 + T6;
|
cannam@167
|
1008 Ti = Tc + Th;
|
cannam@167
|
1009 Tj = T7 + Ti;
|
cannam@167
|
1010 T5F = T7 - Ti;
|
cannam@167
|
1011 T7A = T76 - T75;
|
cannam@167
|
1012 T7B = Tc - Th;
|
cannam@167
|
1013 T7C = T7A - T7B;
|
cannam@167
|
1014 T7Q = T7B + T7A;
|
cannam@167
|
1015 }
|
cannam@167
|
1016 {
|
cannam@167
|
1017 E T31, T34, T74, T77;
|
cannam@167
|
1018 T31 = T1 - T6;
|
cannam@167
|
1019 T34 = T32 - T33;
|
cannam@167
|
1020 T35 = T31 - T34;
|
cannam@167
|
1021 T4T = T31 + T34;
|
cannam@167
|
1022 T74 = T32 + T33;
|
cannam@167
|
1023 T77 = T75 + T76;
|
cannam@167
|
1024 T78 = T74 + T77;
|
cannam@167
|
1025 T7m = T77 - T74;
|
cannam@167
|
1026 }
|
cannam@167
|
1027 }
|
cannam@167
|
1028 {
|
cannam@167
|
1029 E T1y, T3G, T1O, T3Z, T1D, T3H, T1J, T3Y;
|
cannam@167
|
1030 {
|
cannam@167
|
1031 E T1v, T1x, T1u, T1w;
|
cannam@167
|
1032 T1v = Ip[0];
|
cannam@167
|
1033 T1x = Im[0];
|
cannam@167
|
1034 T1u = W[0];
|
cannam@167
|
1035 T1w = W[1];
|
cannam@167
|
1036 T1y = FMA(T1u, T1v, T1w * T1x);
|
cannam@167
|
1037 T3G = FNMS(T1w, T1v, T1u * T1x);
|
cannam@167
|
1038 }
|
cannam@167
|
1039 {
|
cannam@167
|
1040 E T1L, T1N, T1K, T1M;
|
cannam@167
|
1041 T1L = Ip[WS(rs, 12)];
|
cannam@167
|
1042 T1N = Im[WS(rs, 12)];
|
cannam@167
|
1043 T1K = W[48];
|
cannam@167
|
1044 T1M = W[49];
|
cannam@167
|
1045 T1O = FMA(T1K, T1L, T1M * T1N);
|
cannam@167
|
1046 T3Z = FNMS(T1M, T1L, T1K * T1N);
|
cannam@167
|
1047 }
|
cannam@167
|
1048 {
|
cannam@167
|
1049 E T1A, T1C, T1z, T1B;
|
cannam@167
|
1050 T1A = Ip[WS(rs, 8)];
|
cannam@167
|
1051 T1C = Im[WS(rs, 8)];
|
cannam@167
|
1052 T1z = W[32];
|
cannam@167
|
1053 T1B = W[33];
|
cannam@167
|
1054 T1D = FMA(T1z, T1A, T1B * T1C);
|
cannam@167
|
1055 T3H = FNMS(T1B, T1A, T1z * T1C);
|
cannam@167
|
1056 }
|
cannam@167
|
1057 {
|
cannam@167
|
1058 E T1G, T1I, T1F, T1H;
|
cannam@167
|
1059 T1G = Ip[WS(rs, 4)];
|
cannam@167
|
1060 T1I = Im[WS(rs, 4)];
|
cannam@167
|
1061 T1F = W[16];
|
cannam@167
|
1062 T1H = W[17];
|
cannam@167
|
1063 T1J = FMA(T1F, T1G, T1H * T1I);
|
cannam@167
|
1064 T3Y = FNMS(T1H, T1G, T1F * T1I);
|
cannam@167
|
1065 }
|
cannam@167
|
1066 {
|
cannam@167
|
1067 E T1E, T1P, T5W, T5X;
|
cannam@167
|
1068 T1E = T1y + T1D;
|
cannam@167
|
1069 T1P = T1J + T1O;
|
cannam@167
|
1070 T1Q = T1E + T1P;
|
cannam@167
|
1071 T61 = T1E - T1P;
|
cannam@167
|
1072 T5W = T3G + T3H;
|
cannam@167
|
1073 T5X = T3Y + T3Z;
|
cannam@167
|
1074 T5Y = T5W - T5X;
|
cannam@167
|
1075 T6J = T5W + T5X;
|
cannam@167
|
1076 }
|
cannam@167
|
1077 {
|
cannam@167
|
1078 E T3I, T3J, T3X, T40;
|
cannam@167
|
1079 T3I = T3G - T3H;
|
cannam@167
|
1080 T3J = T1J - T1O;
|
cannam@167
|
1081 T3K = T3I + T3J;
|
cannam@167
|
1082 T59 = T3I - T3J;
|
cannam@167
|
1083 T3X = T1y - T1D;
|
cannam@167
|
1084 T40 = T3Y - T3Z;
|
cannam@167
|
1085 T41 = T3X - T40;
|
cannam@167
|
1086 T56 = T3X + T40;
|
cannam@167
|
1087 }
|
cannam@167
|
1088 }
|
cannam@167
|
1089 {
|
cannam@167
|
1090 E T2j, T4o, T2z, T49, T2o, T4p, T2u, T48;
|
cannam@167
|
1091 {
|
cannam@167
|
1092 E T2g, T2i, T2f, T2h;
|
cannam@167
|
1093 T2g = Ip[WS(rs, 15)];
|
cannam@167
|
1094 T2i = Im[WS(rs, 15)];
|
cannam@167
|
1095 T2f = W[60];
|
cannam@167
|
1096 T2h = W[61];
|
cannam@167
|
1097 T2j = FMA(T2f, T2g, T2h * T2i);
|
cannam@167
|
1098 T4o = FNMS(T2h, T2g, T2f * T2i);
|
cannam@167
|
1099 }
|
cannam@167
|
1100 {
|
cannam@167
|
1101 E T2w, T2y, T2v, T2x;
|
cannam@167
|
1102 T2w = Ip[WS(rs, 11)];
|
cannam@167
|
1103 T2y = Im[WS(rs, 11)];
|
cannam@167
|
1104 T2v = W[44];
|
cannam@167
|
1105 T2x = W[45];
|
cannam@167
|
1106 T2z = FMA(T2v, T2w, T2x * T2y);
|
cannam@167
|
1107 T49 = FNMS(T2x, T2w, T2v * T2y);
|
cannam@167
|
1108 }
|
cannam@167
|
1109 {
|
cannam@167
|
1110 E T2l, T2n, T2k, T2m;
|
cannam@167
|
1111 T2l = Ip[WS(rs, 7)];
|
cannam@167
|
1112 T2n = Im[WS(rs, 7)];
|
cannam@167
|
1113 T2k = W[28];
|
cannam@167
|
1114 T2m = W[29];
|
cannam@167
|
1115 T2o = FMA(T2k, T2l, T2m * T2n);
|
cannam@167
|
1116 T4p = FNMS(T2m, T2l, T2k * T2n);
|
cannam@167
|
1117 }
|
cannam@167
|
1118 {
|
cannam@167
|
1119 E T2r, T2t, T2q, T2s;
|
cannam@167
|
1120 T2r = Ip[WS(rs, 3)];
|
cannam@167
|
1121 T2t = Im[WS(rs, 3)];
|
cannam@167
|
1122 T2q = W[12];
|
cannam@167
|
1123 T2s = W[13];
|
cannam@167
|
1124 T2u = FMA(T2q, T2r, T2s * T2t);
|
cannam@167
|
1125 T48 = FNMS(T2s, T2r, T2q * T2t);
|
cannam@167
|
1126 }
|
cannam@167
|
1127 {
|
cannam@167
|
1128 E T2p, T2A, T6c, T6d;
|
cannam@167
|
1129 T2p = T2j + T2o;
|
cannam@167
|
1130 T2A = T2u + T2z;
|
cannam@167
|
1131 T2B = T2p + T2A;
|
cannam@167
|
1132 T67 = T2p - T2A;
|
cannam@167
|
1133 T6c = T4o + T4p;
|
cannam@167
|
1134 T6d = T48 + T49;
|
cannam@167
|
1135 T6e = T6c - T6d;
|
cannam@167
|
1136 T6O = T6c + T6d;
|
cannam@167
|
1137 }
|
cannam@167
|
1138 {
|
cannam@167
|
1139 E T47, T4a, T4q, T4r;
|
cannam@167
|
1140 T47 = T2j - T2o;
|
cannam@167
|
1141 T4a = T48 - T49;
|
cannam@167
|
1142 T4b = T47 - T4a;
|
cannam@167
|
1143 T5d = T47 + T4a;
|
cannam@167
|
1144 T4q = T4o - T4p;
|
cannam@167
|
1145 T4r = T2u - T2z;
|
cannam@167
|
1146 T4s = T4q + T4r;
|
cannam@167
|
1147 T5g = T4q - T4r;
|
cannam@167
|
1148 }
|
cannam@167
|
1149 }
|
cannam@167
|
1150 {
|
cannam@167
|
1151 E To, T36, TE, T3d, Tt, T37, Tz, T3c;
|
cannam@167
|
1152 {
|
cannam@167
|
1153 E Tl, Tn, Tk, Tm;
|
cannam@167
|
1154 Tl = Rp[WS(rs, 2)];
|
cannam@167
|
1155 Tn = Rm[WS(rs, 2)];
|
cannam@167
|
1156 Tk = W[6];
|
cannam@167
|
1157 Tm = W[7];
|
cannam@167
|
1158 To = FMA(Tk, Tl, Tm * Tn);
|
cannam@167
|
1159 T36 = FNMS(Tm, Tl, Tk * Tn);
|
cannam@167
|
1160 }
|
cannam@167
|
1161 {
|
cannam@167
|
1162 E TB, TD, TA, TC;
|
cannam@167
|
1163 TB = Rp[WS(rs, 6)];
|
cannam@167
|
1164 TD = Rm[WS(rs, 6)];
|
cannam@167
|
1165 TA = W[22];
|
cannam@167
|
1166 TC = W[23];
|
cannam@167
|
1167 TE = FMA(TA, TB, TC * TD);
|
cannam@167
|
1168 T3d = FNMS(TC, TB, TA * TD);
|
cannam@167
|
1169 }
|
cannam@167
|
1170 {
|
cannam@167
|
1171 E Tq, Ts, Tp, Tr;
|
cannam@167
|
1172 Tq = Rp[WS(rs, 10)];
|
cannam@167
|
1173 Ts = Rm[WS(rs, 10)];
|
cannam@167
|
1174 Tp = W[38];
|
cannam@167
|
1175 Tr = W[39];
|
cannam@167
|
1176 Tt = FMA(Tp, Tq, Tr * Ts);
|
cannam@167
|
1177 T37 = FNMS(Tr, Tq, Tp * Ts);
|
cannam@167
|
1178 }
|
cannam@167
|
1179 {
|
cannam@167
|
1180 E Tw, Ty, Tv, Tx;
|
cannam@167
|
1181 Tw = Rp[WS(rs, 14)];
|
cannam@167
|
1182 Ty = Rm[WS(rs, 14)];
|
cannam@167
|
1183 Tv = W[54];
|
cannam@167
|
1184 Tx = W[55];
|
cannam@167
|
1185 Tz = FMA(Tv, Tw, Tx * Ty);
|
cannam@167
|
1186 T3c = FNMS(Tx, Tw, Tv * Ty);
|
cannam@167
|
1187 }
|
cannam@167
|
1188 {
|
cannam@167
|
1189 E Tu, TF, T5G, T5H;
|
cannam@167
|
1190 Tu = To + Tt;
|
cannam@167
|
1191 TF = Tz + TE;
|
cannam@167
|
1192 TG = Tu + TF;
|
cannam@167
|
1193 T7l = TF - Tu;
|
cannam@167
|
1194 T5G = T36 + T37;
|
cannam@167
|
1195 T5H = T3c + T3d;
|
cannam@167
|
1196 T5I = T5G - T5H;
|
cannam@167
|
1197 T73 = T5G + T5H;
|
cannam@167
|
1198 }
|
cannam@167
|
1199 {
|
cannam@167
|
1200 E T38, T39, T3b, T3e;
|
cannam@167
|
1201 T38 = T36 - T37;
|
cannam@167
|
1202 T39 = To - Tt;
|
cannam@167
|
1203 T3a = T38 - T39;
|
cannam@167
|
1204 T4U = T39 + T38;
|
cannam@167
|
1205 T3b = Tz - TE;
|
cannam@167
|
1206 T3e = T3c - T3d;
|
cannam@167
|
1207 T3f = T3b + T3e;
|
cannam@167
|
1208 T4V = T3b - T3e;
|
cannam@167
|
1209 }
|
cannam@167
|
1210 }
|
cannam@167
|
1211 {
|
cannam@167
|
1212 E TM, T3i, T12, T3p, TR, T3j, TX, T3o;
|
cannam@167
|
1213 {
|
cannam@167
|
1214 E TJ, TL, TI, TK;
|
cannam@167
|
1215 TJ = Rp[WS(rs, 1)];
|
cannam@167
|
1216 TL = Rm[WS(rs, 1)];
|
cannam@167
|
1217 TI = W[2];
|
cannam@167
|
1218 TK = W[3];
|
cannam@167
|
1219 TM = FMA(TI, TJ, TK * TL);
|
cannam@167
|
1220 T3i = FNMS(TK, TJ, TI * TL);
|
cannam@167
|
1221 }
|
cannam@167
|
1222 {
|
cannam@167
|
1223 E TZ, T11, TY, T10;
|
cannam@167
|
1224 TZ = Rp[WS(rs, 13)];
|
cannam@167
|
1225 T11 = Rm[WS(rs, 13)];
|
cannam@167
|
1226 TY = W[50];
|
cannam@167
|
1227 T10 = W[51];
|
cannam@167
|
1228 T12 = FMA(TY, TZ, T10 * T11);
|
cannam@167
|
1229 T3p = FNMS(T10, TZ, TY * T11);
|
cannam@167
|
1230 }
|
cannam@167
|
1231 {
|
cannam@167
|
1232 E TO, TQ, TN, TP;
|
cannam@167
|
1233 TO = Rp[WS(rs, 9)];
|
cannam@167
|
1234 TQ = Rm[WS(rs, 9)];
|
cannam@167
|
1235 TN = W[34];
|
cannam@167
|
1236 TP = W[35];
|
cannam@167
|
1237 TR = FMA(TN, TO, TP * TQ);
|
cannam@167
|
1238 T3j = FNMS(TP, TO, TN * TQ);
|
cannam@167
|
1239 }
|
cannam@167
|
1240 {
|
cannam@167
|
1241 E TU, TW, TT, TV;
|
cannam@167
|
1242 TU = Rp[WS(rs, 5)];
|
cannam@167
|
1243 TW = Rm[WS(rs, 5)];
|
cannam@167
|
1244 TT = W[18];
|
cannam@167
|
1245 TV = W[19];
|
cannam@167
|
1246 TX = FMA(TT, TU, TV * TW);
|
cannam@167
|
1247 T3o = FNMS(TV, TU, TT * TW);
|
cannam@167
|
1248 }
|
cannam@167
|
1249 {
|
cannam@167
|
1250 E TS, T13, T5K, T5L;
|
cannam@167
|
1251 TS = TM + TR;
|
cannam@167
|
1252 T13 = TX + T12;
|
cannam@167
|
1253 T14 = TS + T13;
|
cannam@167
|
1254 T5N = TS - T13;
|
cannam@167
|
1255 T5K = T3i + T3j;
|
cannam@167
|
1256 T5L = T3o + T3p;
|
cannam@167
|
1257 T5M = T5K - T5L;
|
cannam@167
|
1258 T6E = T5K + T5L;
|
cannam@167
|
1259 }
|
cannam@167
|
1260 {
|
cannam@167
|
1261 E T3k, T3l, T3n, T3q;
|
cannam@167
|
1262 T3k = T3i - T3j;
|
cannam@167
|
1263 T3l = TX - T12;
|
cannam@167
|
1264 T3m = T3k + T3l;
|
cannam@167
|
1265 T4Y = T3k - T3l;
|
cannam@167
|
1266 T3n = TM - TR;
|
cannam@167
|
1267 T3q = T3o - T3p;
|
cannam@167
|
1268 T3r = T3n - T3q;
|
cannam@167
|
1269 T4Z = T3n + T3q;
|
cannam@167
|
1270 }
|
cannam@167
|
1271 }
|
cannam@167
|
1272 {
|
cannam@167
|
1273 E T19, T3t, T1p, T3A, T1e, T3u, T1k, T3z;
|
cannam@167
|
1274 {
|
cannam@167
|
1275 E T16, T18, T15, T17;
|
cannam@167
|
1276 T16 = Rp[WS(rs, 15)];
|
cannam@167
|
1277 T18 = Rm[WS(rs, 15)];
|
cannam@167
|
1278 T15 = W[58];
|
cannam@167
|
1279 T17 = W[59];
|
cannam@167
|
1280 T19 = FMA(T15, T16, T17 * T18);
|
cannam@167
|
1281 T3t = FNMS(T17, T16, T15 * T18);
|
cannam@167
|
1282 }
|
cannam@167
|
1283 {
|
cannam@167
|
1284 E T1m, T1o, T1l, T1n;
|
cannam@167
|
1285 T1m = Rp[WS(rs, 11)];
|
cannam@167
|
1286 T1o = Rm[WS(rs, 11)];
|
cannam@167
|
1287 T1l = W[42];
|
cannam@167
|
1288 T1n = W[43];
|
cannam@167
|
1289 T1p = FMA(T1l, T1m, T1n * T1o);
|
cannam@167
|
1290 T3A = FNMS(T1n, T1m, T1l * T1o);
|
cannam@167
|
1291 }
|
cannam@167
|
1292 {
|
cannam@167
|
1293 E T1b, T1d, T1a, T1c;
|
cannam@167
|
1294 T1b = Rp[WS(rs, 7)];
|
cannam@167
|
1295 T1d = Rm[WS(rs, 7)];
|
cannam@167
|
1296 T1a = W[26];
|
cannam@167
|
1297 T1c = W[27];
|
cannam@167
|
1298 T1e = FMA(T1a, T1b, T1c * T1d);
|
cannam@167
|
1299 T3u = FNMS(T1c, T1b, T1a * T1d);
|
cannam@167
|
1300 }
|
cannam@167
|
1301 {
|
cannam@167
|
1302 E T1h, T1j, T1g, T1i;
|
cannam@167
|
1303 T1h = Rp[WS(rs, 3)];
|
cannam@167
|
1304 T1j = Rm[WS(rs, 3)];
|
cannam@167
|
1305 T1g = W[10];
|
cannam@167
|
1306 T1i = W[11];
|
cannam@167
|
1307 T1k = FMA(T1g, T1h, T1i * T1j);
|
cannam@167
|
1308 T3z = FNMS(T1i, T1h, T1g * T1j);
|
cannam@167
|
1309 }
|
cannam@167
|
1310 {
|
cannam@167
|
1311 E T1f, T1q, T5Q, T5R;
|
cannam@167
|
1312 T1f = T19 + T1e;
|
cannam@167
|
1313 T1q = T1k + T1p;
|
cannam@167
|
1314 T1r = T1f + T1q;
|
cannam@167
|
1315 T5P = T1f - T1q;
|
cannam@167
|
1316 T5Q = T3t + T3u;
|
cannam@167
|
1317 T5R = T3z + T3A;
|
cannam@167
|
1318 T5S = T5Q - T5R;
|
cannam@167
|
1319 T6F = T5Q + T5R;
|
cannam@167
|
1320 }
|
cannam@167
|
1321 {
|
cannam@167
|
1322 E T3v, T3w, T3y, T3B;
|
cannam@167
|
1323 T3v = T3t - T3u;
|
cannam@167
|
1324 T3w = T1k - T1p;
|
cannam@167
|
1325 T3x = T3v + T3w;
|
cannam@167
|
1326 T51 = T3v - T3w;
|
cannam@167
|
1327 T3y = T19 - T1e;
|
cannam@167
|
1328 T3B = T3z - T3A;
|
cannam@167
|
1329 T3C = T3y - T3B;
|
cannam@167
|
1330 T52 = T3y + T3B;
|
cannam@167
|
1331 }
|
cannam@167
|
1332 }
|
cannam@167
|
1333 {
|
cannam@167
|
1334 E T1V, T3R, T20, T3S, T3Q, T3T, T26, T3M, T2b, T3N, T3L, T3O;
|
cannam@167
|
1335 {
|
cannam@167
|
1336 E T1S, T1U, T1R, T1T;
|
cannam@167
|
1337 T1S = Ip[WS(rs, 2)];
|
cannam@167
|
1338 T1U = Im[WS(rs, 2)];
|
cannam@167
|
1339 T1R = W[8];
|
cannam@167
|
1340 T1T = W[9];
|
cannam@167
|
1341 T1V = FMA(T1R, T1S, T1T * T1U);
|
cannam@167
|
1342 T3R = FNMS(T1T, T1S, T1R * T1U);
|
cannam@167
|
1343 }
|
cannam@167
|
1344 {
|
cannam@167
|
1345 E T1X, T1Z, T1W, T1Y;
|
cannam@167
|
1346 T1X = Ip[WS(rs, 10)];
|
cannam@167
|
1347 T1Z = Im[WS(rs, 10)];
|
cannam@167
|
1348 T1W = W[40];
|
cannam@167
|
1349 T1Y = W[41];
|
cannam@167
|
1350 T20 = FMA(T1W, T1X, T1Y * T1Z);
|
cannam@167
|
1351 T3S = FNMS(T1Y, T1X, T1W * T1Z);
|
cannam@167
|
1352 }
|
cannam@167
|
1353 T3Q = T1V - T20;
|
cannam@167
|
1354 T3T = T3R - T3S;
|
cannam@167
|
1355 {
|
cannam@167
|
1356 E T23, T25, T22, T24;
|
cannam@167
|
1357 T23 = Ip[WS(rs, 14)];
|
cannam@167
|
1358 T25 = Im[WS(rs, 14)];
|
cannam@167
|
1359 T22 = W[56];
|
cannam@167
|
1360 T24 = W[57];
|
cannam@167
|
1361 T26 = FMA(T22, T23, T24 * T25);
|
cannam@167
|
1362 T3M = FNMS(T24, T23, T22 * T25);
|
cannam@167
|
1363 }
|
cannam@167
|
1364 {
|
cannam@167
|
1365 E T28, T2a, T27, T29;
|
cannam@167
|
1366 T28 = Ip[WS(rs, 6)];
|
cannam@167
|
1367 T2a = Im[WS(rs, 6)];
|
cannam@167
|
1368 T27 = W[24];
|
cannam@167
|
1369 T29 = W[25];
|
cannam@167
|
1370 T2b = FMA(T27, T28, T29 * T2a);
|
cannam@167
|
1371 T3N = FNMS(T29, T28, T27 * T2a);
|
cannam@167
|
1372 }
|
cannam@167
|
1373 T3L = T26 - T2b;
|
cannam@167
|
1374 T3O = T3M - T3N;
|
cannam@167
|
1375 {
|
cannam@167
|
1376 E T21, T2c, T62, T63;
|
cannam@167
|
1377 T21 = T1V + T20;
|
cannam@167
|
1378 T2c = T26 + T2b;
|
cannam@167
|
1379 T2d = T21 + T2c;
|
cannam@167
|
1380 T5Z = T2c - T21;
|
cannam@167
|
1381 T62 = T3R + T3S;
|
cannam@167
|
1382 T63 = T3M + T3N;
|
cannam@167
|
1383 T64 = T62 - T63;
|
cannam@167
|
1384 T6K = T62 + T63;
|
cannam@167
|
1385 }
|
cannam@167
|
1386 {
|
cannam@167
|
1387 E T3P, T3U, T42, T43;
|
cannam@167
|
1388 T3P = T3L - T3O;
|
cannam@167
|
1389 T3U = T3Q + T3T;
|
cannam@167
|
1390 T3V = KP707106781 * (T3P - T3U);
|
cannam@167
|
1391 T57 = KP707106781 * (T3U + T3P);
|
cannam@167
|
1392 T42 = T3T - T3Q;
|
cannam@167
|
1393 T43 = T3L + T3O;
|
cannam@167
|
1394 T44 = KP707106781 * (T42 - T43);
|
cannam@167
|
1395 T5a = KP707106781 * (T42 + T43);
|
cannam@167
|
1396 }
|
cannam@167
|
1397 }
|
cannam@167
|
1398 {
|
cannam@167
|
1399 E T2G, T4c, T2L, T4d, T4e, T4f, T2R, T4i, T2W, T4j, T4h, T4k;
|
cannam@167
|
1400 {
|
cannam@167
|
1401 E T2D, T2F, T2C, T2E;
|
cannam@167
|
1402 T2D = Ip[WS(rs, 1)];
|
cannam@167
|
1403 T2F = Im[WS(rs, 1)];
|
cannam@167
|
1404 T2C = W[4];
|
cannam@167
|
1405 T2E = W[5];
|
cannam@167
|
1406 T2G = FMA(T2C, T2D, T2E * T2F);
|
cannam@167
|
1407 T4c = FNMS(T2E, T2D, T2C * T2F);
|
cannam@167
|
1408 }
|
cannam@167
|
1409 {
|
cannam@167
|
1410 E T2I, T2K, T2H, T2J;
|
cannam@167
|
1411 T2I = Ip[WS(rs, 9)];
|
cannam@167
|
1412 T2K = Im[WS(rs, 9)];
|
cannam@167
|
1413 T2H = W[36];
|
cannam@167
|
1414 T2J = W[37];
|
cannam@167
|
1415 T2L = FMA(T2H, T2I, T2J * T2K);
|
cannam@167
|
1416 T4d = FNMS(T2J, T2I, T2H * T2K);
|
cannam@167
|
1417 }
|
cannam@167
|
1418 T4e = T4c - T4d;
|
cannam@167
|
1419 T4f = T2G - T2L;
|
cannam@167
|
1420 {
|
cannam@167
|
1421 E T2O, T2Q, T2N, T2P;
|
cannam@167
|
1422 T2O = Ip[WS(rs, 13)];
|
cannam@167
|
1423 T2Q = Im[WS(rs, 13)];
|
cannam@167
|
1424 T2N = W[52];
|
cannam@167
|
1425 T2P = W[53];
|
cannam@167
|
1426 T2R = FMA(T2N, T2O, T2P * T2Q);
|
cannam@167
|
1427 T4i = FNMS(T2P, T2O, T2N * T2Q);
|
cannam@167
|
1428 }
|
cannam@167
|
1429 {
|
cannam@167
|
1430 E T2T, T2V, T2S, T2U;
|
cannam@167
|
1431 T2T = Ip[WS(rs, 5)];
|
cannam@167
|
1432 T2V = Im[WS(rs, 5)];
|
cannam@167
|
1433 T2S = W[20];
|
cannam@167
|
1434 T2U = W[21];
|
cannam@167
|
1435 T2W = FMA(T2S, T2T, T2U * T2V);
|
cannam@167
|
1436 T4j = FNMS(T2U, T2T, T2S * T2V);
|
cannam@167
|
1437 }
|
cannam@167
|
1438 T4h = T2R - T2W;
|
cannam@167
|
1439 T4k = T4i - T4j;
|
cannam@167
|
1440 {
|
cannam@167
|
1441 E T2M, T2X, T68, T69;
|
cannam@167
|
1442 T2M = T2G + T2L;
|
cannam@167
|
1443 T2X = T2R + T2W;
|
cannam@167
|
1444 T2Y = T2M + T2X;
|
cannam@167
|
1445 T6f = T2X - T2M;
|
cannam@167
|
1446 T68 = T4c + T4d;
|
cannam@167
|
1447 T69 = T4i + T4j;
|
cannam@167
|
1448 T6a = T68 - T69;
|
cannam@167
|
1449 T6P = T68 + T69;
|
cannam@167
|
1450 }
|
cannam@167
|
1451 {
|
cannam@167
|
1452 E T4g, T4l, T4t, T4u;
|
cannam@167
|
1453 T4g = T4e - T4f;
|
cannam@167
|
1454 T4l = T4h + T4k;
|
cannam@167
|
1455 T4m = KP707106781 * (T4g - T4l);
|
cannam@167
|
1456 T5h = KP707106781 * (T4g + T4l);
|
cannam@167
|
1457 T4t = T4h - T4k;
|
cannam@167
|
1458 T4u = T4f + T4e;
|
cannam@167
|
1459 T4v = KP707106781 * (T4t - T4u);
|
cannam@167
|
1460 T5e = KP707106781 * (T4u + T4t);
|
cannam@167
|
1461 }
|
cannam@167
|
1462 }
|
cannam@167
|
1463 {
|
cannam@167
|
1464 E T1t, T6X, T7a, T7c, T30, T7b, T70, T71;
|
cannam@167
|
1465 {
|
cannam@167
|
1466 E TH, T1s, T72, T79;
|
cannam@167
|
1467 TH = Tj + TG;
|
cannam@167
|
1468 T1s = T14 + T1r;
|
cannam@167
|
1469 T1t = TH + T1s;
|
cannam@167
|
1470 T6X = TH - T1s;
|
cannam@167
|
1471 T72 = T6E + T6F;
|
cannam@167
|
1472 T79 = T73 + T78;
|
cannam@167
|
1473 T7a = T72 + T79;
|
cannam@167
|
1474 T7c = T79 - T72;
|
cannam@167
|
1475 }
|
cannam@167
|
1476 {
|
cannam@167
|
1477 E T2e, T2Z, T6Y, T6Z;
|
cannam@167
|
1478 T2e = T1Q + T2d;
|
cannam@167
|
1479 T2Z = T2B + T2Y;
|
cannam@167
|
1480 T30 = T2e + T2Z;
|
cannam@167
|
1481 T7b = T2Z - T2e;
|
cannam@167
|
1482 T6Y = T6J + T6K;
|
cannam@167
|
1483 T6Z = T6O + T6P;
|
cannam@167
|
1484 T70 = T6Y - T6Z;
|
cannam@167
|
1485 T71 = T6Y + T6Z;
|
cannam@167
|
1486 }
|
cannam@167
|
1487 Rm[WS(rs, 15)] = T1t - T30;
|
cannam@167
|
1488 Im[WS(rs, 15)] = T71 - T7a;
|
cannam@167
|
1489 Rp[0] = T1t + T30;
|
cannam@167
|
1490 Ip[0] = T71 + T7a;
|
cannam@167
|
1491 Rm[WS(rs, 7)] = T6X - T70;
|
cannam@167
|
1492 Im[WS(rs, 7)] = T7b - T7c;
|
cannam@167
|
1493 Rp[WS(rs, 8)] = T6X + T70;
|
cannam@167
|
1494 Ip[WS(rs, 8)] = T7b + T7c;
|
cannam@167
|
1495 }
|
cannam@167
|
1496 {
|
cannam@167
|
1497 E T6H, T6T, T7g, T7i, T6M, T6U, T6R, T6V;
|
cannam@167
|
1498 {
|
cannam@167
|
1499 E T6D, T6G, T7e, T7f;
|
cannam@167
|
1500 T6D = Tj - TG;
|
cannam@167
|
1501 T6G = T6E - T6F;
|
cannam@167
|
1502 T6H = T6D + T6G;
|
cannam@167
|
1503 T6T = T6D - T6G;
|
cannam@167
|
1504 T7e = T1r - T14;
|
cannam@167
|
1505 T7f = T78 - T73;
|
cannam@167
|
1506 T7g = T7e + T7f;
|
cannam@167
|
1507 T7i = T7f - T7e;
|
cannam@167
|
1508 }
|
cannam@167
|
1509 {
|
cannam@167
|
1510 E T6I, T6L, T6N, T6Q;
|
cannam@167
|
1511 T6I = T1Q - T2d;
|
cannam@167
|
1512 T6L = T6J - T6K;
|
cannam@167
|
1513 T6M = T6I + T6L;
|
cannam@167
|
1514 T6U = T6L - T6I;
|
cannam@167
|
1515 T6N = T2B - T2Y;
|
cannam@167
|
1516 T6Q = T6O - T6P;
|
cannam@167
|
1517 T6R = T6N - T6Q;
|
cannam@167
|
1518 T6V = T6N + T6Q;
|
cannam@167
|
1519 }
|
cannam@167
|
1520 {
|
cannam@167
|
1521 E T6S, T7d, T6W, T7h;
|
cannam@167
|
1522 T6S = KP707106781 * (T6M + T6R);
|
cannam@167
|
1523 Rm[WS(rs, 11)] = T6H - T6S;
|
cannam@167
|
1524 Rp[WS(rs, 4)] = T6H + T6S;
|
cannam@167
|
1525 T7d = KP707106781 * (T6U + T6V);
|
cannam@167
|
1526 Im[WS(rs, 11)] = T7d - T7g;
|
cannam@167
|
1527 Ip[WS(rs, 4)] = T7d + T7g;
|
cannam@167
|
1528 T6W = KP707106781 * (T6U - T6V);
|
cannam@167
|
1529 Rm[WS(rs, 3)] = T6T - T6W;
|
cannam@167
|
1530 Rp[WS(rs, 12)] = T6T + T6W;
|
cannam@167
|
1531 T7h = KP707106781 * (T6R - T6M);
|
cannam@167
|
1532 Im[WS(rs, 3)] = T7h - T7i;
|
cannam@167
|
1533 Ip[WS(rs, 12)] = T7h + T7i;
|
cannam@167
|
1534 }
|
cannam@167
|
1535 }
|
cannam@167
|
1536 {
|
cannam@167
|
1537 E T5J, T7n, T7t, T6n, T5U, T7k, T6x, T6B, T6q, T7s, T66, T6k, T6u, T6A, T6h;
|
cannam@167
|
1538 E T6l;
|
cannam@167
|
1539 {
|
cannam@167
|
1540 E T5O, T5T, T60, T65;
|
cannam@167
|
1541 T5J = T5F - T5I;
|
cannam@167
|
1542 T7n = T7l + T7m;
|
cannam@167
|
1543 T7t = T7m - T7l;
|
cannam@167
|
1544 T6n = T5F + T5I;
|
cannam@167
|
1545 T5O = T5M - T5N;
|
cannam@167
|
1546 T5T = T5P + T5S;
|
cannam@167
|
1547 T5U = KP707106781 * (T5O - T5T);
|
cannam@167
|
1548 T7k = KP707106781 * (T5O + T5T);
|
cannam@167
|
1549 {
|
cannam@167
|
1550 E T6v, T6w, T6o, T6p;
|
cannam@167
|
1551 T6v = T67 + T6a;
|
cannam@167
|
1552 T6w = T6e + T6f;
|
cannam@167
|
1553 T6x = FNMS(KP382683432, T6w, KP923879532 * T6v);
|
cannam@167
|
1554 T6B = FMA(KP923879532, T6w, KP382683432 * T6v);
|
cannam@167
|
1555 T6o = T5N + T5M;
|
cannam@167
|
1556 T6p = T5P - T5S;
|
cannam@167
|
1557 T6q = KP707106781 * (T6o + T6p);
|
cannam@167
|
1558 T7s = KP707106781 * (T6p - T6o);
|
cannam@167
|
1559 }
|
cannam@167
|
1560 T60 = T5Y - T5Z;
|
cannam@167
|
1561 T65 = T61 - T64;
|
cannam@167
|
1562 T66 = FMA(KP923879532, T60, KP382683432 * T65);
|
cannam@167
|
1563 T6k = FNMS(KP923879532, T65, KP382683432 * T60);
|
cannam@167
|
1564 {
|
cannam@167
|
1565 E T6s, T6t, T6b, T6g;
|
cannam@167
|
1566 T6s = T5Y + T5Z;
|
cannam@167
|
1567 T6t = T61 + T64;
|
cannam@167
|
1568 T6u = FMA(KP382683432, T6s, KP923879532 * T6t);
|
cannam@167
|
1569 T6A = FNMS(KP382683432, T6t, KP923879532 * T6s);
|
cannam@167
|
1570 T6b = T67 - T6a;
|
cannam@167
|
1571 T6g = T6e - T6f;
|
cannam@167
|
1572 T6h = FNMS(KP923879532, T6g, KP382683432 * T6b);
|
cannam@167
|
1573 T6l = FMA(KP382683432, T6g, KP923879532 * T6b);
|
cannam@167
|
1574 }
|
cannam@167
|
1575 }
|
cannam@167
|
1576 {
|
cannam@167
|
1577 E T5V, T6i, T7r, T7u;
|
cannam@167
|
1578 T5V = T5J + T5U;
|
cannam@167
|
1579 T6i = T66 + T6h;
|
cannam@167
|
1580 Rm[WS(rs, 9)] = T5V - T6i;
|
cannam@167
|
1581 Rp[WS(rs, 6)] = T5V + T6i;
|
cannam@167
|
1582 T7r = T6k + T6l;
|
cannam@167
|
1583 T7u = T7s + T7t;
|
cannam@167
|
1584 Im[WS(rs, 9)] = T7r - T7u;
|
cannam@167
|
1585 Ip[WS(rs, 6)] = T7r + T7u;
|
cannam@167
|
1586 }
|
cannam@167
|
1587 {
|
cannam@167
|
1588 E T6j, T6m, T7v, T7w;
|
cannam@167
|
1589 T6j = T5J - T5U;
|
cannam@167
|
1590 T6m = T6k - T6l;
|
cannam@167
|
1591 Rm[WS(rs, 1)] = T6j - T6m;
|
cannam@167
|
1592 Rp[WS(rs, 14)] = T6j + T6m;
|
cannam@167
|
1593 T7v = T6h - T66;
|
cannam@167
|
1594 T7w = T7t - T7s;
|
cannam@167
|
1595 Im[WS(rs, 1)] = T7v - T7w;
|
cannam@167
|
1596 Ip[WS(rs, 14)] = T7v + T7w;
|
cannam@167
|
1597 }
|
cannam@167
|
1598 {
|
cannam@167
|
1599 E T6r, T6y, T7j, T7o;
|
cannam@167
|
1600 T6r = T6n + T6q;
|
cannam@167
|
1601 T6y = T6u + T6x;
|
cannam@167
|
1602 Rm[WS(rs, 13)] = T6r - T6y;
|
cannam@167
|
1603 Rp[WS(rs, 2)] = T6r + T6y;
|
cannam@167
|
1604 T7j = T6A + T6B;
|
cannam@167
|
1605 T7o = T7k + T7n;
|
cannam@167
|
1606 Im[WS(rs, 13)] = T7j - T7o;
|
cannam@167
|
1607 Ip[WS(rs, 2)] = T7j + T7o;
|
cannam@167
|
1608 }
|
cannam@167
|
1609 {
|
cannam@167
|
1610 E T6z, T6C, T7p, T7q;
|
cannam@167
|
1611 T6z = T6n - T6q;
|
cannam@167
|
1612 T6C = T6A - T6B;
|
cannam@167
|
1613 Rm[WS(rs, 5)] = T6z - T6C;
|
cannam@167
|
1614 Rp[WS(rs, 10)] = T6z + T6C;
|
cannam@167
|
1615 T7p = T6x - T6u;
|
cannam@167
|
1616 T7q = T7n - T7k;
|
cannam@167
|
1617 Im[WS(rs, 5)] = T7p - T7q;
|
cannam@167
|
1618 Ip[WS(rs, 10)] = T7p + T7q;
|
cannam@167
|
1619 }
|
cannam@167
|
1620 }
|
cannam@167
|
1621 {
|
cannam@167
|
1622 E T3h, T4D, T7R, T7X, T3E, T7O, T4N, T4R, T46, T4A, T4G, T7W, T4K, T4Q, T4x;
|
cannam@167
|
1623 E T4B, T3g, T7P;
|
cannam@167
|
1624 T3g = KP707106781 * (T3a - T3f);
|
cannam@167
|
1625 T3h = T35 - T3g;
|
cannam@167
|
1626 T4D = T35 + T3g;
|
cannam@167
|
1627 T7P = KP707106781 * (T4V - T4U);
|
cannam@167
|
1628 T7R = T7P + T7Q;
|
cannam@167
|
1629 T7X = T7Q - T7P;
|
cannam@167
|
1630 {
|
cannam@167
|
1631 E T3s, T3D, T4L, T4M;
|
cannam@167
|
1632 T3s = FNMS(KP923879532, T3r, KP382683432 * T3m);
|
cannam@167
|
1633 T3D = FMA(KP382683432, T3x, KP923879532 * T3C);
|
cannam@167
|
1634 T3E = T3s - T3D;
|
cannam@167
|
1635 T7O = T3s + T3D;
|
cannam@167
|
1636 T4L = T4b + T4m;
|
cannam@167
|
1637 T4M = T4s + T4v;
|
cannam@167
|
1638 T4N = FNMS(KP555570233, T4M, KP831469612 * T4L);
|
cannam@167
|
1639 T4R = FMA(KP831469612, T4M, KP555570233 * T4L);
|
cannam@167
|
1640 }
|
cannam@167
|
1641 {
|
cannam@167
|
1642 E T3W, T45, T4E, T4F;
|
cannam@167
|
1643 T3W = T3K - T3V;
|
cannam@167
|
1644 T45 = T41 - T44;
|
cannam@167
|
1645 T46 = FMA(KP980785280, T3W, KP195090322 * T45);
|
cannam@167
|
1646 T4A = FNMS(KP980785280, T45, KP195090322 * T3W);
|
cannam@167
|
1647 T4E = FMA(KP923879532, T3m, KP382683432 * T3r);
|
cannam@167
|
1648 T4F = FNMS(KP923879532, T3x, KP382683432 * T3C);
|
cannam@167
|
1649 T4G = T4E + T4F;
|
cannam@167
|
1650 T7W = T4F - T4E;
|
cannam@167
|
1651 }
|
cannam@167
|
1652 {
|
cannam@167
|
1653 E T4I, T4J, T4n, T4w;
|
cannam@167
|
1654 T4I = T3K + T3V;
|
cannam@167
|
1655 T4J = T41 + T44;
|
cannam@167
|
1656 T4K = FMA(KP555570233, T4I, KP831469612 * T4J);
|
cannam@167
|
1657 T4Q = FNMS(KP555570233, T4J, KP831469612 * T4I);
|
cannam@167
|
1658 T4n = T4b - T4m;
|
cannam@167
|
1659 T4w = T4s - T4v;
|
cannam@167
|
1660 T4x = FNMS(KP980785280, T4w, KP195090322 * T4n);
|
cannam@167
|
1661 T4B = FMA(KP195090322, T4w, KP980785280 * T4n);
|
cannam@167
|
1662 }
|
cannam@167
|
1663 {
|
cannam@167
|
1664 E T3F, T4y, T7V, T7Y;
|
cannam@167
|
1665 T3F = T3h + T3E;
|
cannam@167
|
1666 T4y = T46 + T4x;
|
cannam@167
|
1667 Rm[WS(rs, 8)] = T3F - T4y;
|
cannam@167
|
1668 Rp[WS(rs, 7)] = T3F + T4y;
|
cannam@167
|
1669 T7V = T4A + T4B;
|
cannam@167
|
1670 T7Y = T7W + T7X;
|
cannam@167
|
1671 Im[WS(rs, 8)] = T7V - T7Y;
|
cannam@167
|
1672 Ip[WS(rs, 7)] = T7V + T7Y;
|
cannam@167
|
1673 }
|
cannam@167
|
1674 {
|
cannam@167
|
1675 E T4z, T4C, T7Z, T80;
|
cannam@167
|
1676 T4z = T3h - T3E;
|
cannam@167
|
1677 T4C = T4A - T4B;
|
cannam@167
|
1678 Rm[0] = T4z - T4C;
|
cannam@167
|
1679 Rp[WS(rs, 15)] = T4z + T4C;
|
cannam@167
|
1680 T7Z = T4x - T46;
|
cannam@167
|
1681 T80 = T7X - T7W;
|
cannam@167
|
1682 Im[0] = T7Z - T80;
|
cannam@167
|
1683 Ip[WS(rs, 15)] = T7Z + T80;
|
cannam@167
|
1684 }
|
cannam@167
|
1685 {
|
cannam@167
|
1686 E T4H, T4O, T7N, T7S;
|
cannam@167
|
1687 T4H = T4D + T4G;
|
cannam@167
|
1688 T4O = T4K + T4N;
|
cannam@167
|
1689 Rm[WS(rs, 12)] = T4H - T4O;
|
cannam@167
|
1690 Rp[WS(rs, 3)] = T4H + T4O;
|
cannam@167
|
1691 T7N = T4Q + T4R;
|
cannam@167
|
1692 T7S = T7O + T7R;
|
cannam@167
|
1693 Im[WS(rs, 12)] = T7N - T7S;
|
cannam@167
|
1694 Ip[WS(rs, 3)] = T7N + T7S;
|
cannam@167
|
1695 }
|
cannam@167
|
1696 {
|
cannam@167
|
1697 E T4P, T4S, T7T, T7U;
|
cannam@167
|
1698 T4P = T4D - T4G;
|
cannam@167
|
1699 T4S = T4Q - T4R;
|
cannam@167
|
1700 Rm[WS(rs, 4)] = T4P - T4S;
|
cannam@167
|
1701 Rp[WS(rs, 11)] = T4P + T4S;
|
cannam@167
|
1702 T7T = T4N - T4K;
|
cannam@167
|
1703 T7U = T7R - T7O;
|
cannam@167
|
1704 Im[WS(rs, 4)] = T7T - T7U;
|
cannam@167
|
1705 Ip[WS(rs, 11)] = T7T + T7U;
|
cannam@167
|
1706 }
|
cannam@167
|
1707 }
|
cannam@167
|
1708 {
|
cannam@167
|
1709 E T4X, T5p, T7D, T7J, T54, T7y, T5z, T5D, T5c, T5m, T5s, T7I, T5w, T5C, T5j;
|
cannam@167
|
1710 E T5n, T4W, T7z;
|
cannam@167
|
1711 T4W = KP707106781 * (T4U + T4V);
|
cannam@167
|
1712 T4X = T4T - T4W;
|
cannam@167
|
1713 T5p = T4T + T4W;
|
cannam@167
|
1714 T7z = KP707106781 * (T3a + T3f);
|
cannam@167
|
1715 T7D = T7z + T7C;
|
cannam@167
|
1716 T7J = T7C - T7z;
|
cannam@167
|
1717 {
|
cannam@167
|
1718 E T50, T53, T5x, T5y;
|
cannam@167
|
1719 T50 = FNMS(KP382683432, T4Z, KP923879532 * T4Y);
|
cannam@167
|
1720 T53 = FMA(KP923879532, T51, KP382683432 * T52);
|
cannam@167
|
1721 T54 = T50 - T53;
|
cannam@167
|
1722 T7y = T50 + T53;
|
cannam@167
|
1723 T5x = T5d + T5e;
|
cannam@167
|
1724 T5y = T5g + T5h;
|
cannam@167
|
1725 T5z = FNMS(KP195090322, T5y, KP980785280 * T5x);
|
cannam@167
|
1726 T5D = FMA(KP195090322, T5x, KP980785280 * T5y);
|
cannam@167
|
1727 }
|
cannam@167
|
1728 {
|
cannam@167
|
1729 E T58, T5b, T5q, T5r;
|
cannam@167
|
1730 T58 = T56 - T57;
|
cannam@167
|
1731 T5b = T59 - T5a;
|
cannam@167
|
1732 T5c = FMA(KP555570233, T58, KP831469612 * T5b);
|
cannam@167
|
1733 T5m = FNMS(KP831469612, T58, KP555570233 * T5b);
|
cannam@167
|
1734 T5q = FMA(KP382683432, T4Y, KP923879532 * T4Z);
|
cannam@167
|
1735 T5r = FNMS(KP382683432, T51, KP923879532 * T52);
|
cannam@167
|
1736 T5s = T5q + T5r;
|
cannam@167
|
1737 T7I = T5r - T5q;
|
cannam@167
|
1738 }
|
cannam@167
|
1739 {
|
cannam@167
|
1740 E T5u, T5v, T5f, T5i;
|
cannam@167
|
1741 T5u = T56 + T57;
|
cannam@167
|
1742 T5v = T59 + T5a;
|
cannam@167
|
1743 T5w = FMA(KP980785280, T5u, KP195090322 * T5v);
|
cannam@167
|
1744 T5C = FNMS(KP195090322, T5u, KP980785280 * T5v);
|
cannam@167
|
1745 T5f = T5d - T5e;
|
cannam@167
|
1746 T5i = T5g - T5h;
|
cannam@167
|
1747 T5j = FNMS(KP831469612, T5i, KP555570233 * T5f);
|
cannam@167
|
1748 T5n = FMA(KP831469612, T5f, KP555570233 * T5i);
|
cannam@167
|
1749 }
|
cannam@167
|
1750 {
|
cannam@167
|
1751 E T55, T5k, T7H, T7K;
|
cannam@167
|
1752 T55 = T4X + T54;
|
cannam@167
|
1753 T5k = T5c + T5j;
|
cannam@167
|
1754 Rm[WS(rs, 10)] = T55 - T5k;
|
cannam@167
|
1755 Rp[WS(rs, 5)] = T55 + T5k;
|
cannam@167
|
1756 T7H = T5m + T5n;
|
cannam@167
|
1757 T7K = T7I + T7J;
|
cannam@167
|
1758 Im[WS(rs, 10)] = T7H - T7K;
|
cannam@167
|
1759 Ip[WS(rs, 5)] = T7H + T7K;
|
cannam@167
|
1760 }
|
cannam@167
|
1761 {
|
cannam@167
|
1762 E T5l, T5o, T7L, T7M;
|
cannam@167
|
1763 T5l = T4X - T54;
|
cannam@167
|
1764 T5o = T5m - T5n;
|
cannam@167
|
1765 Rm[WS(rs, 2)] = T5l - T5o;
|
cannam@167
|
1766 Rp[WS(rs, 13)] = T5l + T5o;
|
cannam@167
|
1767 T7L = T5j - T5c;
|
cannam@167
|
1768 T7M = T7J - T7I;
|
cannam@167
|
1769 Im[WS(rs, 2)] = T7L - T7M;
|
cannam@167
|
1770 Ip[WS(rs, 13)] = T7L + T7M;
|
cannam@167
|
1771 }
|
cannam@167
|
1772 {
|
cannam@167
|
1773 E T5t, T5A, T7x, T7E;
|
cannam@167
|
1774 T5t = T5p + T5s;
|
cannam@167
|
1775 T5A = T5w + T5z;
|
cannam@167
|
1776 Rm[WS(rs, 14)] = T5t - T5A;
|
cannam@167
|
1777 Rp[WS(rs, 1)] = T5t + T5A;
|
cannam@167
|
1778 T7x = T5C + T5D;
|
cannam@167
|
1779 T7E = T7y + T7D;
|
cannam@167
|
1780 Im[WS(rs, 14)] = T7x - T7E;
|
cannam@167
|
1781 Ip[WS(rs, 1)] = T7x + T7E;
|
cannam@167
|
1782 }
|
cannam@167
|
1783 {
|
cannam@167
|
1784 E T5B, T5E, T7F, T7G;
|
cannam@167
|
1785 T5B = T5p - T5s;
|
cannam@167
|
1786 T5E = T5C - T5D;
|
cannam@167
|
1787 Rm[WS(rs, 6)] = T5B - T5E;
|
cannam@167
|
1788 Rp[WS(rs, 9)] = T5B + T5E;
|
cannam@167
|
1789 T7F = T5z - T5w;
|
cannam@167
|
1790 T7G = T7D - T7y;
|
cannam@167
|
1791 Im[WS(rs, 6)] = T7F - T7G;
|
cannam@167
|
1792 Ip[WS(rs, 9)] = T7F + T7G;
|
cannam@167
|
1793 }
|
cannam@167
|
1794 }
|
cannam@167
|
1795 }
|
cannam@167
|
1796 }
|
cannam@167
|
1797 }
|
cannam@167
|
1798
|
cannam@167
|
1799 static const tw_instr twinstr[] = {
|
cannam@167
|
1800 {TW_FULL, 1, 32},
|
cannam@167
|
1801 {TW_NEXT, 1, 0}
|
cannam@167
|
1802 };
|
cannam@167
|
1803
|
cannam@167
|
1804 static const hc2c_desc desc = { 32, "hc2cf_32", twinstr, &GENUS, {340, 114, 94, 0} };
|
cannam@167
|
1805
|
cannam@167
|
1806 void X(codelet_hc2cf_32) (planner *p) {
|
cannam@167
|
1807 X(khc2c_register) (p, hc2cf_32, &desc, HC2C_VIA_RDFT);
|
cannam@167
|
1808 }
|
cannam@167
|
1809 #endif
|