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 16:52:04 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_hc2cdft.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -sign 1 -n 32 -dif -name hc2cbdft_32 -include hc2cb.h */
|
Chris@42
|
29
|
Chris@42
|
30 /*
|
Chris@42
|
31 * This function contains 498 FP additions, 260 FP multiplications,
|
Chris@42
|
32 * (or, 300 additions, 62 multiplications, 198 fused multiply/add),
|
Chris@42
|
33 * 165 stack variables, 7 constants, and 128 memory accesses
|
Chris@42
|
34 */
|
Chris@42
|
35 #include "hc2cb.h"
|
Chris@42
|
36
|
Chris@42
|
37 static void hc2cbdft_32(R *Rp, R *Ip, R *Rm, R *Im, 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(KP198912367, +0.198912367379658006911597622644676228597850501);
|
Chris@42
|
43 DK(KP668178637, +0.668178637919298919997757686523080761552472251);
|
Chris@42
|
44 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
Chris@42
|
45 DK(KP414213562, +0.414213562373095048801688724209698078569671875);
|
Chris@42
|
46 {
|
Chris@42
|
47 INT m;
|
Chris@42
|
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)) {
|
Chris@42
|
49 E T8e, T8h, T7S, T8l, T8f, T84, T8c, T8k, T8g, T86, T82, T8m, T8i;
|
Chris@42
|
50 {
|
Chris@42
|
51 E T4B, T3h, T3K, Tv, T8Y, T6T, T8L, T7i, T8X, T7f, T4Y, T1G, T4K, T1j, T4X;
|
Chris@42
|
52 E T2M, T8C, T6d, T8o, T66, T8K, T6M, T4L, T2P, T4C, T3o, T5q, T4q, T8p, T6C;
|
Chris@42
|
53 E T8B, T6z, T72, T2u, T75, T10, T3P, T3a, T3L, T4t, T4E, T8F, T8t, T4F, T4w;
|
Chris@42
|
54 E T8E, T8w, T6E, T6l, T6F, T6s, T76, T4P, T51, T2R, T28, T8P, T90, T7k, T71;
|
Chris@42
|
55 E T2p, T4R, T2x, T73, T6x, T6y;
|
Chris@42
|
56 {
|
Chris@42
|
57 E T3l, T16, T3m, T2H, T2E, T13, T64, T7, T3i, T2J, T1c, T3j, T1h, T2K, Te;
|
Chris@42
|
58 E T1z, T6R, T6a, Tt, T3g, T6b, T1E, T6Q, Tj, T1p, Ti, T3b, T1n, Tk, T1q;
|
Chris@42
|
59 E T1r;
|
Chris@42
|
60 {
|
Chris@42
|
61 E T1, T2, T4, T5;
|
Chris@42
|
62 {
|
Chris@42
|
63 E T14, T15, T2F, T2G;
|
Chris@42
|
64 T14 = Ip[0];
|
Chris@42
|
65 T15 = Im[WS(rs, 15)];
|
Chris@42
|
66 T2F = Ip[WS(rs, 8)];
|
Chris@42
|
67 T2G = Im[WS(rs, 7)];
|
Chris@42
|
68 T1 = Rp[0];
|
Chris@42
|
69 T3l = T14 - T15;
|
Chris@42
|
70 T16 = T14 + T15;
|
Chris@42
|
71 T3m = T2F - T2G;
|
Chris@42
|
72 T2H = T2F + T2G;
|
Chris@42
|
73 T2 = Rm[WS(rs, 15)];
|
Chris@42
|
74 T4 = Rp[WS(rs, 8)];
|
Chris@42
|
75 T5 = Rm[WS(rs, 7)];
|
Chris@42
|
76 }
|
Chris@42
|
77 {
|
Chris@42
|
78 E T1b, T1e, T18, Ta, T1f, Tb, Tc, T8, T9, T1g, T1d, Td;
|
Chris@42
|
79 {
|
Chris@42
|
80 E T19, T3, T6, T1a;
|
Chris@42
|
81 T19 = Ip[WS(rs, 4)];
|
Chris@42
|
82 T2E = T1 - T2;
|
Chris@42
|
83 T3 = T1 + T2;
|
Chris@42
|
84 T13 = T4 - T5;
|
Chris@42
|
85 T6 = T4 + T5;
|
Chris@42
|
86 T1a = Im[WS(rs, 11)];
|
Chris@42
|
87 T8 = Rp[WS(rs, 4)];
|
Chris@42
|
88 T9 = Rm[WS(rs, 11)];
|
Chris@42
|
89 T64 = T3 - T6;
|
Chris@42
|
90 T7 = T3 + T6;
|
Chris@42
|
91 T1b = T19 + T1a;
|
Chris@42
|
92 T3i = T19 - T1a;
|
Chris@42
|
93 }
|
Chris@42
|
94 T1e = Im[WS(rs, 3)];
|
Chris@42
|
95 T18 = T8 - T9;
|
Chris@42
|
96 Ta = T8 + T9;
|
Chris@42
|
97 T1f = Ip[WS(rs, 12)];
|
Chris@42
|
98 Tb = Rm[WS(rs, 3)];
|
Chris@42
|
99 Tc = Rp[WS(rs, 12)];
|
Chris@42
|
100 T2J = T18 - T1b;
|
Chris@42
|
101 T1c = T18 + T1b;
|
Chris@42
|
102 T1g = T1e + T1f;
|
Chris@42
|
103 T3j = T1f - T1e;
|
Chris@42
|
104 T1d = Tb - Tc;
|
Chris@42
|
105 Td = Tb + Tc;
|
Chris@42
|
106 T1h = T1d + T1g;
|
Chris@42
|
107 T2K = T1d - T1g;
|
Chris@42
|
108 T6x = Ta - Td;
|
Chris@42
|
109 Te = Ta + Td;
|
Chris@42
|
110 }
|
Chris@42
|
111 {
|
Chris@42
|
112 E Tq, T1A, Tp, T3e, T1y, Tr, T1B, T1C;
|
Chris@42
|
113 {
|
Chris@42
|
114 E Tn, To, T1w, T1x;
|
Chris@42
|
115 Tn = Rm[WS(rs, 1)];
|
Chris@42
|
116 To = Rp[WS(rs, 14)];
|
Chris@42
|
117 T1w = Im[WS(rs, 1)];
|
Chris@42
|
118 T1x = Ip[WS(rs, 14)];
|
Chris@42
|
119 Tq = Rp[WS(rs, 6)];
|
Chris@42
|
120 T1A = Tn - To;
|
Chris@42
|
121 Tp = Tn + To;
|
Chris@42
|
122 T3e = T1x - T1w;
|
Chris@42
|
123 T1y = T1w + T1x;
|
Chris@42
|
124 Tr = Rm[WS(rs, 9)];
|
Chris@42
|
125 T1B = Ip[WS(rs, 6)];
|
Chris@42
|
126 T1C = Im[WS(rs, 9)];
|
Chris@42
|
127 }
|
Chris@42
|
128 {
|
Chris@42
|
129 E Tg, Th, T1l, T1m;
|
Chris@42
|
130 Tg = Rp[WS(rs, 2)];
|
Chris@42
|
131 {
|
Chris@42
|
132 E T1v, Ts, T3f, T1D;
|
Chris@42
|
133 T1v = Tq - Tr;
|
Chris@42
|
134 Ts = Tq + Tr;
|
Chris@42
|
135 T3f = T1B - T1C;
|
Chris@42
|
136 T1D = T1B + T1C;
|
Chris@42
|
137 T1z = T1v - T1y;
|
Chris@42
|
138 T6R = T1v + T1y;
|
Chris@42
|
139 T6a = Tp - Ts;
|
Chris@42
|
140 Tt = Tp + Ts;
|
Chris@42
|
141 T3g = T3e + T3f;
|
Chris@42
|
142 T6b = T3e - T3f;
|
Chris@42
|
143 T1E = T1A - T1D;
|
Chris@42
|
144 T6Q = T1A + T1D;
|
Chris@42
|
145 Th = Rm[WS(rs, 13)];
|
Chris@42
|
146 }
|
Chris@42
|
147 T1l = Ip[WS(rs, 2)];
|
Chris@42
|
148 T1m = Im[WS(rs, 13)];
|
Chris@42
|
149 Tj = Rp[WS(rs, 10)];
|
Chris@42
|
150 T1p = Tg - Th;
|
Chris@42
|
151 Ti = Tg + Th;
|
Chris@42
|
152 T3b = T1l - T1m;
|
Chris@42
|
153 T1n = T1l + T1m;
|
Chris@42
|
154 Tk = Rm[WS(rs, 5)];
|
Chris@42
|
155 T1q = Ip[WS(rs, 10)];
|
Chris@42
|
156 T1r = Im[WS(rs, 5)];
|
Chris@42
|
157 }
|
Chris@42
|
158 }
|
Chris@42
|
159 }
|
Chris@42
|
160 {
|
Chris@42
|
161 E T4o, T67, T68, T4p, T2I, T1i, T2N, T1u, T1F, T2O, T6K, T17;
|
Chris@42
|
162 {
|
Chris@42
|
163 E Tf, T1o, T1t, Tu, T7g, T6P, T6S, T7h, T7d, T7e;
|
Chris@42
|
164 {
|
Chris@42
|
165 E T6O, T6N, T1k, Tl;
|
Chris@42
|
166 T4o = T7 - Te;
|
Chris@42
|
167 Tf = T7 + Te;
|
Chris@42
|
168 T1k = Tj - Tk;
|
Chris@42
|
169 Tl = Tj + Tk;
|
Chris@42
|
170 {
|
Chris@42
|
171 E T3c, T1s, Tm, T3d;
|
Chris@42
|
172 T3c = T1q - T1r;
|
Chris@42
|
173 T1s = T1q + T1r;
|
Chris@42
|
174 T1o = T1k + T1n;
|
Chris@42
|
175 T6O = T1n - T1k;
|
Chris@42
|
176 T67 = Ti - Tl;
|
Chris@42
|
177 Tm = Ti + Tl;
|
Chris@42
|
178 T3d = T3b + T3c;
|
Chris@42
|
179 T68 = T3b - T3c;
|
Chris@42
|
180 T1t = T1p - T1s;
|
Chris@42
|
181 T6N = T1p + T1s;
|
Chris@42
|
182 T4B = Tm - Tt;
|
Chris@42
|
183 Tu = Tm + Tt;
|
Chris@42
|
184 T4p = T3g - T3d;
|
Chris@42
|
185 T3h = T3d + T3g;
|
Chris@42
|
186 }
|
Chris@42
|
187 T7g = FNMS(KP414213562, T6N, T6O);
|
Chris@42
|
188 T6P = FMA(KP414213562, T6O, T6N);
|
Chris@42
|
189 T6S = FMA(KP414213562, T6R, T6Q);
|
Chris@42
|
190 T7h = FNMS(KP414213562, T6Q, T6R);
|
Chris@42
|
191 }
|
Chris@42
|
192 T3K = Tf - Tu;
|
Chris@42
|
193 Tv = Tf + Tu;
|
Chris@42
|
194 T8Y = T6P + T6S;
|
Chris@42
|
195 T6T = T6P - T6S;
|
Chris@42
|
196 T2I = T2E - T2H;
|
Chris@42
|
197 T7d = T2E + T2H;
|
Chris@42
|
198 T7e = T1c + T1h;
|
Chris@42
|
199 T1i = T1c - T1h;
|
Chris@42
|
200 T2N = FNMS(KP414213562, T1o, T1t);
|
Chris@42
|
201 T1u = FMA(KP414213562, T1t, T1o);
|
Chris@42
|
202 T8L = T7h - T7g;
|
Chris@42
|
203 T7i = T7g + T7h;
|
Chris@42
|
204 T8X = FMA(KP707106781, T7e, T7d);
|
Chris@42
|
205 T7f = FNMS(KP707106781, T7e, T7d);
|
Chris@42
|
206 T1F = FNMS(KP414213562, T1E, T1z);
|
Chris@42
|
207 T2O = FMA(KP414213562, T1z, T1E);
|
Chris@42
|
208 T6K = T16 - T13;
|
Chris@42
|
209 T17 = T13 + T16;
|
Chris@42
|
210 }
|
Chris@42
|
211 {
|
Chris@42
|
212 E T6L, T6A, T6B, T65, T3k, T2L, T69, T6c, T3n;
|
Chris@42
|
213 T4Y = T1F - T1u;
|
Chris@42
|
214 T1G = T1u + T1F;
|
Chris@42
|
215 T4K = FNMS(KP707106781, T1i, T17);
|
Chris@42
|
216 T1j = FMA(KP707106781, T1i, T17);
|
Chris@42
|
217 T2L = T2J + T2K;
|
Chris@42
|
218 T6L = T2J - T2K;
|
Chris@42
|
219 T6A = T67 + T68;
|
Chris@42
|
220 T69 = T67 - T68;
|
Chris@42
|
221 T6c = T6a + T6b;
|
Chris@42
|
222 T6B = T6b - T6a;
|
Chris@42
|
223 T4X = FNMS(KP707106781, T2L, T2I);
|
Chris@42
|
224 T2M = FMA(KP707106781, T2L, T2I);
|
Chris@42
|
225 T8C = T69 - T6c;
|
Chris@42
|
226 T6d = T69 + T6c;
|
Chris@42
|
227 T65 = T3j - T3i;
|
Chris@42
|
228 T3k = T3i + T3j;
|
Chris@42
|
229 T8o = T64 - T65;
|
Chris@42
|
230 T66 = T64 + T65;
|
Chris@42
|
231 T8K = FNMS(KP707106781, T6L, T6K);
|
Chris@42
|
232 T6M = FMA(KP707106781, T6L, T6K);
|
Chris@42
|
233 T3n = T3l + T3m;
|
Chris@42
|
234 T6y = T3l - T3m;
|
Chris@42
|
235 T4L = T2N - T2O;
|
Chris@42
|
236 T2P = T2N + T2O;
|
Chris@42
|
237 T4C = T3n - T3k;
|
Chris@42
|
238 T3o = T3k + T3n;
|
Chris@42
|
239 T5q = T4o - T4p;
|
Chris@42
|
240 T4q = T4o + T4p;
|
Chris@42
|
241 T8p = T6B - T6A;
|
Chris@42
|
242 T6C = T6A + T6B;
|
Chris@42
|
243 }
|
Chris@42
|
244 }
|
Chris@42
|
245 }
|
Chris@42
|
246 {
|
Chris@42
|
247 E T1M, T6V, T6f, TC, T31, T6j, T23, T6Y, T2v, T2i, TY, T6p, T6n, T35, T2n;
|
Chris@42
|
248 E T2w, T24, T1R, TJ, T6i, T6g, T2Y, T1W, T25, T2q, TN, T2r, T36, T2c, T29;
|
Chris@42
|
249 E TQ, T2s;
|
Chris@42
|
250 {
|
Chris@42
|
251 E TU, T2k, T33, T2j, TX, T2l, T2m, T34;
|
Chris@42
|
252 {
|
Chris@42
|
253 E T1Z, Ty, T20, T2Z, T1L, T1I, TB, T21, T2e, T2h;
|
Chris@42
|
254 {
|
Chris@42
|
255 E T1J, T1K, Tw, Tx, Tz, TA;
|
Chris@42
|
256 Tw = Rp[WS(rs, 1)];
|
Chris@42
|
257 Tx = Rm[WS(rs, 14)];
|
Chris@42
|
258 T1J = Ip[WS(rs, 1)];
|
Chris@42
|
259 T8B = T6y - T6x;
|
Chris@42
|
260 T6z = T6x + T6y;
|
Chris@42
|
261 T1Z = Tw - Tx;
|
Chris@42
|
262 Ty = Tw + Tx;
|
Chris@42
|
263 T1K = Im[WS(rs, 14)];
|
Chris@42
|
264 Tz = Rp[WS(rs, 9)];
|
Chris@42
|
265 TA = Rm[WS(rs, 6)];
|
Chris@42
|
266 T20 = Ip[WS(rs, 9)];
|
Chris@42
|
267 T2Z = T1J - T1K;
|
Chris@42
|
268 T1L = T1J + T1K;
|
Chris@42
|
269 T1I = Tz - TA;
|
Chris@42
|
270 TB = Tz + TA;
|
Chris@42
|
271 T21 = Im[WS(rs, 6)];
|
Chris@42
|
272 }
|
Chris@42
|
273 {
|
Chris@42
|
274 E T2f, T2g, TV, TW;
|
Chris@42
|
275 {
|
Chris@42
|
276 E TS, T30, T22, TT;
|
Chris@42
|
277 TS = Rp[WS(rs, 3)];
|
Chris@42
|
278 T1M = T1I + T1L;
|
Chris@42
|
279 T6V = T1L - T1I;
|
Chris@42
|
280 T6f = Ty - TB;
|
Chris@42
|
281 TC = Ty + TB;
|
Chris@42
|
282 T30 = T20 - T21;
|
Chris@42
|
283 T22 = T20 + T21;
|
Chris@42
|
284 TT = Rm[WS(rs, 12)];
|
Chris@42
|
285 T2f = Ip[WS(rs, 3)];
|
Chris@42
|
286 T31 = T2Z + T30;
|
Chris@42
|
287 T6j = T2Z - T30;
|
Chris@42
|
288 T23 = T1Z - T22;
|
Chris@42
|
289 T6Y = T1Z + T22;
|
Chris@42
|
290 T2e = TS - TT;
|
Chris@42
|
291 TU = TS + TT;
|
Chris@42
|
292 T2g = Im[WS(rs, 12)];
|
Chris@42
|
293 }
|
Chris@42
|
294 TV = Rm[WS(rs, 4)];
|
Chris@42
|
295 TW = Rp[WS(rs, 11)];
|
Chris@42
|
296 T2k = Im[WS(rs, 4)];
|
Chris@42
|
297 T33 = T2f - T2g;
|
Chris@42
|
298 T2h = T2f + T2g;
|
Chris@42
|
299 T2j = TV - TW;
|
Chris@42
|
300 TX = TV + TW;
|
Chris@42
|
301 T2l = Ip[WS(rs, 11)];
|
Chris@42
|
302 }
|
Chris@42
|
303 T2v = T2e - T2h;
|
Chris@42
|
304 T2i = T2e + T2h;
|
Chris@42
|
305 }
|
Chris@42
|
306 TY = TU + TX;
|
Chris@42
|
307 T6p = TU - TX;
|
Chris@42
|
308 T2m = T2k + T2l;
|
Chris@42
|
309 T34 = T2l - T2k;
|
Chris@42
|
310 {
|
Chris@42
|
311 E TF, T1T, T2W, T1S, TI, T1U, T1N, T1Q, T1V, T2X;
|
Chris@42
|
312 {
|
Chris@42
|
313 E T1O, T1P, TD, TE, TG, TH;
|
Chris@42
|
314 TD = Rp[WS(rs, 5)];
|
Chris@42
|
315 TE = Rm[WS(rs, 10)];
|
Chris@42
|
316 T6n = T34 - T33;
|
Chris@42
|
317 T35 = T33 + T34;
|
Chris@42
|
318 T2n = T2j + T2m;
|
Chris@42
|
319 T2w = T2j - T2m;
|
Chris@42
|
320 T1N = TD - TE;
|
Chris@42
|
321 TF = TD + TE;
|
Chris@42
|
322 T1O = Ip[WS(rs, 5)];
|
Chris@42
|
323 T1P = Im[WS(rs, 10)];
|
Chris@42
|
324 TG = Rm[WS(rs, 2)];
|
Chris@42
|
325 TH = Rp[WS(rs, 13)];
|
Chris@42
|
326 T1T = Im[WS(rs, 2)];
|
Chris@42
|
327 T2W = T1O - T1P;
|
Chris@42
|
328 T1Q = T1O + T1P;
|
Chris@42
|
329 T1S = TG - TH;
|
Chris@42
|
330 TI = TG + TH;
|
Chris@42
|
331 T1U = Ip[WS(rs, 13)];
|
Chris@42
|
332 }
|
Chris@42
|
333 T24 = T1N - T1Q;
|
Chris@42
|
334 T1R = T1N + T1Q;
|
Chris@42
|
335 TJ = TF + TI;
|
Chris@42
|
336 T6i = TF - TI;
|
Chris@42
|
337 T1V = T1T + T1U;
|
Chris@42
|
338 T2X = T1U - T1T;
|
Chris@42
|
339 {
|
Chris@42
|
340 E T2a, T2b, TL, TM, TO, TP;
|
Chris@42
|
341 TL = Rm[0];
|
Chris@42
|
342 TM = Rp[WS(rs, 15)];
|
Chris@42
|
343 T6g = T2X - T2W;
|
Chris@42
|
344 T2Y = T2W + T2X;
|
Chris@42
|
345 T1W = T1S + T1V;
|
Chris@42
|
346 T25 = T1S - T1V;
|
Chris@42
|
347 T2q = TL - TM;
|
Chris@42
|
348 TN = TL + TM;
|
Chris@42
|
349 T2a = Im[0];
|
Chris@42
|
350 T2b = Ip[WS(rs, 15)];
|
Chris@42
|
351 TO = Rp[WS(rs, 7)];
|
Chris@42
|
352 TP = Rm[WS(rs, 8)];
|
Chris@42
|
353 T2r = Ip[WS(rs, 7)];
|
Chris@42
|
354 T36 = T2b - T2a;
|
Chris@42
|
355 T2c = T2a + T2b;
|
Chris@42
|
356 T29 = TO - TP;
|
Chris@42
|
357 TQ = TO + TP;
|
Chris@42
|
358 T2s = Im[WS(rs, 8)];
|
Chris@42
|
359 }
|
Chris@42
|
360 }
|
Chris@42
|
361 }
|
Chris@42
|
362 {
|
Chris@42
|
363 E T2d, T4u, T4v, T6r, T6o, T6k, T8u, T8v, T6h;
|
Chris@42
|
364 {
|
Chris@42
|
365 E T4r, T6m, T32, T4s, T6q, T39, T8r, T8s;
|
Chris@42
|
366 {
|
Chris@42
|
367 E TK, TR, T37, T2t, TZ, T38;
|
Chris@42
|
368 T4r = TC - TJ;
|
Chris@42
|
369 TK = TC + TJ;
|
Chris@42
|
370 T2d = T29 - T2c;
|
Chris@42
|
371 T72 = T29 + T2c;
|
Chris@42
|
372 T6m = TN - TQ;
|
Chris@42
|
373 TR = TN + TQ;
|
Chris@42
|
374 T37 = T2r - T2s;
|
Chris@42
|
375 T2t = T2r + T2s;
|
Chris@42
|
376 T32 = T2Y + T31;
|
Chris@42
|
377 T4s = T31 - T2Y;
|
Chris@42
|
378 T4u = TR - TY;
|
Chris@42
|
379 TZ = TR + TY;
|
Chris@42
|
380 T38 = T36 + T37;
|
Chris@42
|
381 T6q = T36 - T37;
|
Chris@42
|
382 T2u = T2q - T2t;
|
Chris@42
|
383 T75 = T2q + T2t;
|
Chris@42
|
384 T10 = TK + TZ;
|
Chris@42
|
385 T3P = TK - TZ;
|
Chris@42
|
386 T4v = T38 - T35;
|
Chris@42
|
387 T39 = T35 + T38;
|
Chris@42
|
388 }
|
Chris@42
|
389 T8r = T6q - T6p;
|
Chris@42
|
390 T6r = T6p + T6q;
|
Chris@42
|
391 T3a = T32 + T39;
|
Chris@42
|
392 T3L = T39 - T32;
|
Chris@42
|
393 T8s = T6m - T6n;
|
Chris@42
|
394 T6o = T6m + T6n;
|
Chris@42
|
395 T4t = T4r - T4s;
|
Chris@42
|
396 T4E = T4r + T4s;
|
Chris@42
|
397 T8F = FNMS(KP414213562, T8r, T8s);
|
Chris@42
|
398 T8t = FMA(KP414213562, T8s, T8r);
|
Chris@42
|
399 T6k = T6i + T6j;
|
Chris@42
|
400 T8u = T6j - T6i;
|
Chris@42
|
401 T8v = T6f - T6g;
|
Chris@42
|
402 T6h = T6f + T6g;
|
Chris@42
|
403 }
|
Chris@42
|
404 {
|
Chris@42
|
405 E T6Z, T1Y, T4O, T26, T6W, T1X, T2o, T4N, T27;
|
Chris@42
|
406 T4F = T4v - T4u;
|
Chris@42
|
407 T4w = T4u + T4v;
|
Chris@42
|
408 T8E = FMA(KP414213562, T8u, T8v);
|
Chris@42
|
409 T8w = FNMS(KP414213562, T8v, T8u);
|
Chris@42
|
410 T6Z = T1R + T1W;
|
Chris@42
|
411 T1X = T1R - T1W;
|
Chris@42
|
412 T6E = FMA(KP414213562, T6h, T6k);
|
Chris@42
|
413 T6l = FNMS(KP414213562, T6k, T6h);
|
Chris@42
|
414 T6F = FNMS(KP414213562, T6o, T6r);
|
Chris@42
|
415 T6s = FMA(KP414213562, T6r, T6o);
|
Chris@42
|
416 T1Y = FMA(KP707106781, T1X, T1M);
|
Chris@42
|
417 T4O = FNMS(KP707106781, T1X, T1M);
|
Chris@42
|
418 T26 = T24 + T25;
|
Chris@42
|
419 T6W = T25 - T24;
|
Chris@42
|
420 T76 = T2i + T2n;
|
Chris@42
|
421 T2o = T2i - T2n;
|
Chris@42
|
422 T4N = FNMS(KP707106781, T26, T23);
|
Chris@42
|
423 T27 = FMA(KP707106781, T26, T23);
|
Chris@42
|
424 {
|
Chris@42
|
425 E T8O, T6X, T8N, T70;
|
Chris@42
|
426 T8O = FMA(KP707106781, T6W, T6V);
|
Chris@42
|
427 T6X = FNMS(KP707106781, T6W, T6V);
|
Chris@42
|
428 T8N = FMA(KP707106781, T6Z, T6Y);
|
Chris@42
|
429 T70 = FNMS(KP707106781, T6Z, T6Y);
|
Chris@42
|
430 T4P = FMA(KP668178637, T4O, T4N);
|
Chris@42
|
431 T51 = FNMS(KP668178637, T4N, T4O);
|
Chris@42
|
432 T2R = FNMS(KP198912367, T1Y, T27);
|
Chris@42
|
433 T28 = FMA(KP198912367, T27, T1Y);
|
Chris@42
|
434 T8P = FMA(KP198912367, T8O, T8N);
|
Chris@42
|
435 T90 = FNMS(KP198912367, T8N, T8O);
|
Chris@42
|
436 T7k = FNMS(KP668178637, T6X, T70);
|
Chris@42
|
437 T71 = FMA(KP668178637, T70, T6X);
|
Chris@42
|
438 T2p = FMA(KP707106781, T2o, T2d);
|
Chris@42
|
439 T4R = FNMS(KP707106781, T2o, T2d);
|
Chris@42
|
440 }
|
Chris@42
|
441 T2x = T2v + T2w;
|
Chris@42
|
442 T73 = T2v - T2w;
|
Chris@42
|
443 }
|
Chris@42
|
444 }
|
Chris@42
|
445 }
|
Chris@42
|
446 {
|
Chris@42
|
447 E T8S, T91, T7l, T78, T5U, T5X, T5y, T61, T5V, T5K, T5S, T60, T5W, T5M, T5I;
|
Chris@42
|
448 {
|
Chris@42
|
449 E T4S, T50, T4e, T4h, T3S, T4l, T4f, T44, T4c, T4k, T4g, T46, T42;
|
Chris@42
|
450 {
|
Chris@42
|
451 E T3Q, T3U, T40, T3Z, T3V, T3A, T3D, T3H, T3B, T3y, T3G, T3C;
|
Chris@42
|
452 {
|
Chris@42
|
453 E T11, T3t, T3w, T3q, T3x, T3v, T3F, T12, T2B, T2U, T3z, T2C;
|
Chris@42
|
454 {
|
Chris@42
|
455 E T3u, T2S, T2z, T3p, T4Q, T2y;
|
Chris@42
|
456 T3u = Tv - T10;
|
Chris@42
|
457 T11 = Tv + T10;
|
Chris@42
|
458 T4Q = FNMS(KP707106781, T2x, T2u);
|
Chris@42
|
459 T2y = FMA(KP707106781, T2x, T2u);
|
Chris@42
|
460 {
|
Chris@42
|
461 E T8R, T74, T8Q, T77;
|
Chris@42
|
462 T8R = FMA(KP707106781, T73, T72);
|
Chris@42
|
463 T74 = FNMS(KP707106781, T73, T72);
|
Chris@42
|
464 T8Q = FMA(KP707106781, T76, T75);
|
Chris@42
|
465 T77 = FNMS(KP707106781, T76, T75);
|
Chris@42
|
466 T4S = FNMS(KP668178637, T4R, T4Q);
|
Chris@42
|
467 T50 = FMA(KP668178637, T4Q, T4R);
|
Chris@42
|
468 T2S = FMA(KP198912367, T2p, T2y);
|
Chris@42
|
469 T2z = FNMS(KP198912367, T2y, T2p);
|
Chris@42
|
470 T8S = FMA(KP198912367, T8R, T8Q);
|
Chris@42
|
471 T91 = FNMS(KP198912367, T8Q, T8R);
|
Chris@42
|
472 T7l = FNMS(KP668178637, T74, T77);
|
Chris@42
|
473 T78 = FMA(KP668178637, T77, T74);
|
Chris@42
|
474 T3Q = T3o - T3h;
|
Chris@42
|
475 T3p = T3h + T3o;
|
Chris@42
|
476 }
|
Chris@42
|
477 T3t = W[30];
|
Chris@42
|
478 T3w = W[31];
|
Chris@42
|
479 T3q = T3a + T3p;
|
Chris@42
|
480 T3x = T3p - T3a;
|
Chris@42
|
481 T3v = T3t * T3u;
|
Chris@42
|
482 T3F = T3w * T3u;
|
Chris@42
|
483 {
|
Chris@42
|
484 E T1H, T2A, T2Q, T2T;
|
Chris@42
|
485 T3U = FNMS(KP923879532, T1G, T1j);
|
Chris@42
|
486 T1H = FMA(KP923879532, T1G, T1j);
|
Chris@42
|
487 T2A = T28 + T2z;
|
Chris@42
|
488 T40 = T2z - T28;
|
Chris@42
|
489 T3Z = FNMS(KP923879532, T2P, T2M);
|
Chris@42
|
490 T2Q = FMA(KP923879532, T2P, T2M);
|
Chris@42
|
491 T2T = T2R + T2S;
|
Chris@42
|
492 T3V = T2R - T2S;
|
Chris@42
|
493 T12 = W[0];
|
Chris@42
|
494 T3A = FNMS(KP980785280, T2A, T1H);
|
Chris@42
|
495 T2B = FMA(KP980785280, T2A, T1H);
|
Chris@42
|
496 T3D = FNMS(KP980785280, T2T, T2Q);
|
Chris@42
|
497 T2U = FMA(KP980785280, T2T, T2Q);
|
Chris@42
|
498 T3z = W[32];
|
Chris@42
|
499 T2C = T12 * T2B;
|
Chris@42
|
500 }
|
Chris@42
|
501 }
|
Chris@42
|
502 {
|
Chris@42
|
503 E T2V, T3s, T2D, T3r;
|
Chris@42
|
504 T2D = W[1];
|
Chris@42
|
505 T3r = T12 * T2U;
|
Chris@42
|
506 T3H = T3z * T3D;
|
Chris@42
|
507 T3B = T3z * T3A;
|
Chris@42
|
508 T2V = FMA(T2D, T2U, T2C);
|
Chris@42
|
509 T3s = FNMS(T2D, T2B, T3r);
|
Chris@42
|
510 T3y = FNMS(T3w, T3x, T3v);
|
Chris@42
|
511 T3G = FMA(T3t, T3x, T3F);
|
Chris@42
|
512 Rm[0] = T11 + T2V;
|
Chris@42
|
513 Rp[0] = T11 - T2V;
|
Chris@42
|
514 Im[0] = T3s - T3q;
|
Chris@42
|
515 Ip[0] = T3q + T3s;
|
Chris@42
|
516 T3C = W[33];
|
Chris@42
|
517 }
|
Chris@42
|
518 }
|
Chris@42
|
519 {
|
Chris@42
|
520 E T4b, T3R, T47, T4a, T3J, T49, T4j, T3O, T3N, T43, T3W, T3T, T41, T4d, T3X;
|
Chris@42
|
521 E T45, T3Y;
|
Chris@42
|
522 {
|
Chris@42
|
523 E T3M, T48, T3I, T3E;
|
Chris@42
|
524 T3M = T3K + T3L;
|
Chris@42
|
525 T48 = T3K - T3L;
|
Chris@42
|
526 T3I = FNMS(T3C, T3A, T3H);
|
Chris@42
|
527 T3E = FMA(T3C, T3D, T3B);
|
Chris@42
|
528 T4b = T3Q - T3P;
|
Chris@42
|
529 T3R = T3P + T3Q;
|
Chris@42
|
530 Im[WS(rs, 8)] = T3I - T3G;
|
Chris@42
|
531 Ip[WS(rs, 8)] = T3G + T3I;
|
Chris@42
|
532 Rm[WS(rs, 8)] = T3y + T3E;
|
Chris@42
|
533 Rp[WS(rs, 8)] = T3y - T3E;
|
Chris@42
|
534 T47 = W[46];
|
Chris@42
|
535 T4a = W[47];
|
Chris@42
|
536 T3J = W[14];
|
Chris@42
|
537 T49 = T47 * T48;
|
Chris@42
|
538 T4j = T4a * T48;
|
Chris@42
|
539 T3O = W[15];
|
Chris@42
|
540 T3N = T3J * T3M;
|
Chris@42
|
541 T43 = T3O * T3M;
|
Chris@42
|
542 T3W = FMA(KP980785280, T3V, T3U);
|
Chris@42
|
543 T4e = FNMS(KP980785280, T3V, T3U);
|
Chris@42
|
544 T3T = W[16];
|
Chris@42
|
545 T4h = FNMS(KP980785280, T40, T3Z);
|
Chris@42
|
546 T41 = FMA(KP980785280, T40, T3Z);
|
Chris@42
|
547 T4d = W[48];
|
Chris@42
|
548 T3X = T3T * T3W;
|
Chris@42
|
549 }
|
Chris@42
|
550 T3S = FNMS(T3O, T3R, T3N);
|
Chris@42
|
551 T45 = T3T * T41;
|
Chris@42
|
552 T4l = T4d * T4h;
|
Chris@42
|
553 T4f = T4d * T4e;
|
Chris@42
|
554 T44 = FMA(T3J, T3R, T43);
|
Chris@42
|
555 T3Y = W[17];
|
Chris@42
|
556 T4c = FNMS(T4a, T4b, T49);
|
Chris@42
|
557 T4k = FMA(T47, T4b, T4j);
|
Chris@42
|
558 T4g = W[49];
|
Chris@42
|
559 T46 = FNMS(T3Y, T3W, T45);
|
Chris@42
|
560 T42 = FMA(T3Y, T41, T3X);
|
Chris@42
|
561 }
|
Chris@42
|
562 }
|
Chris@42
|
563 {
|
Chris@42
|
564 E T5v, T5r, T5w, T5A, T5G, T5F, T5B, T5g, T5j, T4I, T5n, T5h, T56, T5e, T5m;
|
Chris@42
|
565 E T5i, T58, T54;
|
Chris@42
|
566 {
|
Chris@42
|
567 E T4n, T4A, T5d, T4H, T59, T5c, T55, T4z, T5b, T5l, T4J, T4U, T53, T5f, T4V;
|
Chris@42
|
568 E T57, T4W;
|
Chris@42
|
569 {
|
Chris@42
|
570 E T4D, T4G, T4m, T4i, T5a, T4y, T4x;
|
Chris@42
|
571 T5v = T4C - T4B;
|
Chris@42
|
572 T4D = T4B + T4C;
|
Chris@42
|
573 T4m = FNMS(T4g, T4e, T4l);
|
Chris@42
|
574 T4i = FMA(T4g, T4h, T4f);
|
Chris@42
|
575 Im[WS(rs, 4)] = T46 - T44;
|
Chris@42
|
576 Ip[WS(rs, 4)] = T44 + T46;
|
Chris@42
|
577 Rm[WS(rs, 4)] = T3S + T42;
|
Chris@42
|
578 Rp[WS(rs, 4)] = T3S - T42;
|
Chris@42
|
579 Im[WS(rs, 12)] = T4m - T4k;
|
Chris@42
|
580 Ip[WS(rs, 12)] = T4k + T4m;
|
Chris@42
|
581 Rm[WS(rs, 12)] = T4c + T4i;
|
Chris@42
|
582 Rp[WS(rs, 12)] = T4c - T4i;
|
Chris@42
|
583 T4G = T4E + T4F;
|
Chris@42
|
584 T5r = T4F - T4E;
|
Chris@42
|
585 T5w = T4t - T4w;
|
Chris@42
|
586 T4x = T4t + T4w;
|
Chris@42
|
587 T4n = W[6];
|
Chris@42
|
588 T4A = W[7];
|
Chris@42
|
589 T5d = FNMS(KP707106781, T4G, T4D);
|
Chris@42
|
590 T4H = FMA(KP707106781, T4G, T4D);
|
Chris@42
|
591 T5a = FNMS(KP707106781, T4x, T4q);
|
Chris@42
|
592 T4y = FMA(KP707106781, T4x, T4q);
|
Chris@42
|
593 T59 = W[38];
|
Chris@42
|
594 T5c = W[39];
|
Chris@42
|
595 {
|
Chris@42
|
596 E T4M, T4T, T4Z, T52;
|
Chris@42
|
597 T4M = FMA(KP923879532, T4L, T4K);
|
Chris@42
|
598 T5A = FNMS(KP923879532, T4L, T4K);
|
Chris@42
|
599 T55 = T4A * T4y;
|
Chris@42
|
600 T4z = T4n * T4y;
|
Chris@42
|
601 T5b = T59 * T5a;
|
Chris@42
|
602 T5l = T5c * T5a;
|
Chris@42
|
603 T5G = T4P + T4S;
|
Chris@42
|
604 T4T = T4P - T4S;
|
Chris@42
|
605 T4Z = FMA(KP923879532, T4Y, T4X);
|
Chris@42
|
606 T5F = FNMS(KP923879532, T4Y, T4X);
|
Chris@42
|
607 T5B = T51 + T50;
|
Chris@42
|
608 T52 = T50 - T51;
|
Chris@42
|
609 T4J = W[8];
|
Chris@42
|
610 T4U = FMA(KP831469612, T4T, T4M);
|
Chris@42
|
611 T5g = FNMS(KP831469612, T4T, T4M);
|
Chris@42
|
612 T53 = FMA(KP831469612, T52, T4Z);
|
Chris@42
|
613 T5j = FNMS(KP831469612, T52, T4Z);
|
Chris@42
|
614 T5f = W[40];
|
Chris@42
|
615 T4V = T4J * T4U;
|
Chris@42
|
616 }
|
Chris@42
|
617 }
|
Chris@42
|
618 T4I = FNMS(T4A, T4H, T4z);
|
Chris@42
|
619 T57 = T4J * T53;
|
Chris@42
|
620 T5n = T5f * T5j;
|
Chris@42
|
621 T5h = T5f * T5g;
|
Chris@42
|
622 T56 = FMA(T4n, T4H, T55);
|
Chris@42
|
623 T4W = W[9];
|
Chris@42
|
624 T5e = FNMS(T5c, T5d, T5b);
|
Chris@42
|
625 T5m = FMA(T59, T5d, T5l);
|
Chris@42
|
626 T5i = W[41];
|
Chris@42
|
627 T58 = FNMS(T4W, T4U, T57);
|
Chris@42
|
628 T54 = FMA(T4W, T53, T4V);
|
Chris@42
|
629 }
|
Chris@42
|
630 {
|
Chris@42
|
631 E T5p, T5u, T5x, T5R, T5N, T5Q, T5J, T5t, T5P, T5Z, T5z, T5C, T5H, T5T, T5D;
|
Chris@42
|
632 E T5L, T5E;
|
Chris@42
|
633 {
|
Chris@42
|
634 E T5o, T5k, T5s, T5O;
|
Chris@42
|
635 T5o = FNMS(T5i, T5g, T5n);
|
Chris@42
|
636 T5k = FMA(T5i, T5j, T5h);
|
Chris@42
|
637 Im[WS(rs, 2)] = T58 - T56;
|
Chris@42
|
638 Ip[WS(rs, 2)] = T56 + T58;
|
Chris@42
|
639 Rm[WS(rs, 2)] = T4I + T54;
|
Chris@42
|
640 Rp[WS(rs, 2)] = T4I - T54;
|
Chris@42
|
641 Im[WS(rs, 10)] = T5o - T5m;
|
Chris@42
|
642 Ip[WS(rs, 10)] = T5m + T5o;
|
Chris@42
|
643 Rm[WS(rs, 10)] = T5e + T5k;
|
Chris@42
|
644 Rp[WS(rs, 10)] = T5e - T5k;
|
Chris@42
|
645 T5p = W[22];
|
Chris@42
|
646 T5u = W[23];
|
Chris@42
|
647 T5x = FMA(KP707106781, T5w, T5v);
|
Chris@42
|
648 T5R = FNMS(KP707106781, T5w, T5v);
|
Chris@42
|
649 T5s = FMA(KP707106781, T5r, T5q);
|
Chris@42
|
650 T5O = FNMS(KP707106781, T5r, T5q);
|
Chris@42
|
651 T5N = W[54];
|
Chris@42
|
652 T5Q = W[55];
|
Chris@42
|
653 T5J = T5u * T5s;
|
Chris@42
|
654 T5t = T5p * T5s;
|
Chris@42
|
655 T5P = T5N * T5O;
|
Chris@42
|
656 T5Z = T5Q * T5O;
|
Chris@42
|
657 T5z = W[24];
|
Chris@42
|
658 T5U = FMA(KP831469612, T5B, T5A);
|
Chris@42
|
659 T5C = FNMS(KP831469612, T5B, T5A);
|
Chris@42
|
660 T5X = FMA(KP831469612, T5G, T5F);
|
Chris@42
|
661 T5H = FNMS(KP831469612, T5G, T5F);
|
Chris@42
|
662 T5T = W[56];
|
Chris@42
|
663 T5D = T5z * T5C;
|
Chris@42
|
664 }
|
Chris@42
|
665 T5y = FNMS(T5u, T5x, T5t);
|
Chris@42
|
666 T5L = T5z * T5H;
|
Chris@42
|
667 T61 = T5T * T5X;
|
Chris@42
|
668 T5V = T5T * T5U;
|
Chris@42
|
669 T5K = FMA(T5p, T5x, T5J);
|
Chris@42
|
670 T5E = W[25];
|
Chris@42
|
671 T5S = FNMS(T5Q, T5R, T5P);
|
Chris@42
|
672 T60 = FMA(T5N, T5R, T5Z);
|
Chris@42
|
673 T5W = W[57];
|
Chris@42
|
674 T5M = FNMS(T5E, T5C, T5L);
|
Chris@42
|
675 T5I = FMA(T5E, T5H, T5D);
|
Chris@42
|
676 }
|
Chris@42
|
677 }
|
Chris@42
|
678 }
|
Chris@42
|
679 {
|
Chris@42
|
680 E T7P, T7L, T7K, T7Q, T7U, T80, T7Z, T7V, T9v, T9r, T9q, T9w, T9A, T9G, T9F;
|
Chris@42
|
681 E T9B, T9g, T9j, T8I, T9n, T9h, T96, T9e, T9m, T9i, T98, T94;
|
Chris@42
|
682 {
|
Chris@42
|
683 E T7A, T7D, T6I, T7H, T7B, T7q, T7y, T7G, T7C, T7s, T7o;
|
Chris@42
|
684 {
|
Chris@42
|
685 E T63, T7x, T6H, T6w, T7t, T7w, T6v, T7p, T7v, T7F, T6J, T7a, T7n, T7z, T7b;
|
Chris@42
|
686 E T7r, T7c;
|
Chris@42
|
687 {
|
Chris@42
|
688 E T6D, T6G, T62, T5Y;
|
Chris@42
|
689 T7P = FNMS(KP707106781, T6C, T6z);
|
Chris@42
|
690 T6D = FMA(KP707106781, T6C, T6z);
|
Chris@42
|
691 T62 = FNMS(T5W, T5U, T61);
|
Chris@42
|
692 T5Y = FMA(T5W, T5X, T5V);
|
Chris@42
|
693 Im[WS(rs, 6)] = T5M - T5K;
|
Chris@42
|
694 Ip[WS(rs, 6)] = T5K + T5M;
|
Chris@42
|
695 Rm[WS(rs, 6)] = T5y + T5I;
|
Chris@42
|
696 Rp[WS(rs, 6)] = T5y - T5I;
|
Chris@42
|
697 Im[WS(rs, 14)] = T62 - T60;
|
Chris@42
|
698 Ip[WS(rs, 14)] = T60 + T62;
|
Chris@42
|
699 Rm[WS(rs, 14)] = T5S + T5Y;
|
Chris@42
|
700 Rp[WS(rs, 14)] = T5S - T5Y;
|
Chris@42
|
701 T6G = T6E + T6F;
|
Chris@42
|
702 T7L = T6F - T6E;
|
Chris@42
|
703 {
|
Chris@42
|
704 E T6e, T6t, T7u, T6u;
|
Chris@42
|
705 T7K = FNMS(KP707106781, T6d, T66);
|
Chris@42
|
706 T6e = FMA(KP707106781, T6d, T66);
|
Chris@42
|
707 T6t = T6l + T6s;
|
Chris@42
|
708 T7Q = T6l - T6s;
|
Chris@42
|
709 T63 = W[2];
|
Chris@42
|
710 T7x = FNMS(KP923879532, T6G, T6D);
|
Chris@42
|
711 T6H = FMA(KP923879532, T6G, T6D);
|
Chris@42
|
712 T7u = FNMS(KP923879532, T6t, T6e);
|
Chris@42
|
713 T6u = FMA(KP923879532, T6t, T6e);
|
Chris@42
|
714 T6w = W[3];
|
Chris@42
|
715 T7t = W[34];
|
Chris@42
|
716 T7w = W[35];
|
Chris@42
|
717 T6v = T63 * T6u;
|
Chris@42
|
718 T7p = T6w * T6u;
|
Chris@42
|
719 T7v = T7t * T7u;
|
Chris@42
|
720 T7F = T7w * T7u;
|
Chris@42
|
721 }
|
Chris@42
|
722 {
|
Chris@42
|
723 E T6U, T79, T7j, T7m;
|
Chris@42
|
724 T7U = FNMS(KP923879532, T6T, T6M);
|
Chris@42
|
725 T6U = FMA(KP923879532, T6T, T6M);
|
Chris@42
|
726 T79 = T71 - T78;
|
Chris@42
|
727 T80 = T71 + T78;
|
Chris@42
|
728 T7Z = FMA(KP923879532, T7i, T7f);
|
Chris@42
|
729 T7j = FNMS(KP923879532, T7i, T7f);
|
Chris@42
|
730 T7m = T7k + T7l;
|
Chris@42
|
731 T7V = T7k - T7l;
|
Chris@42
|
732 T6J = W[4];
|
Chris@42
|
733 T7A = FNMS(KP831469612, T79, T6U);
|
Chris@42
|
734 T7a = FMA(KP831469612, T79, T6U);
|
Chris@42
|
735 T7D = FNMS(KP831469612, T7m, T7j);
|
Chris@42
|
736 T7n = FMA(KP831469612, T7m, T7j);
|
Chris@42
|
737 T7z = W[36];
|
Chris@42
|
738 T7b = T6J * T7a;
|
Chris@42
|
739 }
|
Chris@42
|
740 }
|
Chris@42
|
741 T6I = FNMS(T6w, T6H, T6v);
|
Chris@42
|
742 T7r = T6J * T7n;
|
Chris@42
|
743 T7H = T7z * T7D;
|
Chris@42
|
744 T7B = T7z * T7A;
|
Chris@42
|
745 T7q = FMA(T63, T6H, T7p);
|
Chris@42
|
746 T7c = W[5];
|
Chris@42
|
747 T7y = FNMS(T7w, T7x, T7v);
|
Chris@42
|
748 T7G = FMA(T7t, T7x, T7F);
|
Chris@42
|
749 T7C = W[37];
|
Chris@42
|
750 T7s = FNMS(T7c, T7a, T7r);
|
Chris@42
|
751 T7o = FMA(T7c, T7n, T7b);
|
Chris@42
|
752 }
|
Chris@42
|
753 {
|
Chris@42
|
754 E T8n, T9d, T8H, T8A, T99, T9c, T8z, T95, T9b, T9l, T8J, T8U, T93, T9f, T8V;
|
Chris@42
|
755 E T97, T8W;
|
Chris@42
|
756 {
|
Chris@42
|
757 E T8D, T8G, T7I, T7E;
|
Chris@42
|
758 T9v = FNMS(KP707106781, T8C, T8B);
|
Chris@42
|
759 T8D = FMA(KP707106781, T8C, T8B);
|
Chris@42
|
760 T7I = FNMS(T7C, T7A, T7H);
|
Chris@42
|
761 T7E = FMA(T7C, T7D, T7B);
|
Chris@42
|
762 Im[WS(rs, 1)] = T7s - T7q;
|
Chris@42
|
763 Ip[WS(rs, 1)] = T7q + T7s;
|
Chris@42
|
764 Rm[WS(rs, 1)] = T6I + T7o;
|
Chris@42
|
765 Rp[WS(rs, 1)] = T6I - T7o;
|
Chris@42
|
766 Im[WS(rs, 9)] = T7I - T7G;
|
Chris@42
|
767 Ip[WS(rs, 9)] = T7G + T7I;
|
Chris@42
|
768 Rm[WS(rs, 9)] = T7y + T7E;
|
Chris@42
|
769 Rp[WS(rs, 9)] = T7y - T7E;
|
Chris@42
|
770 T8G = T8E - T8F;
|
Chris@42
|
771 T9r = T8E + T8F;
|
Chris@42
|
772 {
|
Chris@42
|
773 E T8q, T8x, T9a, T8y;
|
Chris@42
|
774 T9q = FNMS(KP707106781, T8p, T8o);
|
Chris@42
|
775 T8q = FMA(KP707106781, T8p, T8o);
|
Chris@42
|
776 T8x = T8t - T8w;
|
Chris@42
|
777 T9w = T8w + T8t;
|
Chris@42
|
778 T8n = W[10];
|
Chris@42
|
779 T9d = FNMS(KP923879532, T8G, T8D);
|
Chris@42
|
780 T8H = FMA(KP923879532, T8G, T8D);
|
Chris@42
|
781 T9a = FNMS(KP923879532, T8x, T8q);
|
Chris@42
|
782 T8y = FMA(KP923879532, T8x, T8q);
|
Chris@42
|
783 T8A = W[11];
|
Chris@42
|
784 T99 = W[42];
|
Chris@42
|
785 T9c = W[43];
|
Chris@42
|
786 T8z = T8n * T8y;
|
Chris@42
|
787 T95 = T8A * T8y;
|
Chris@42
|
788 T9b = T99 * T9a;
|
Chris@42
|
789 T9l = T9c * T9a;
|
Chris@42
|
790 }
|
Chris@42
|
791 {
|
Chris@42
|
792 E T8M, T8T, T8Z, T92;
|
Chris@42
|
793 T9A = FNMS(KP923879532, T8L, T8K);
|
Chris@42
|
794 T8M = FMA(KP923879532, T8L, T8K);
|
Chris@42
|
795 T8T = T8P - T8S;
|
Chris@42
|
796 T9G = T8P + T8S;
|
Chris@42
|
797 T9F = FMA(KP923879532, T8Y, T8X);
|
Chris@42
|
798 T8Z = FNMS(KP923879532, T8Y, T8X);
|
Chris@42
|
799 T92 = T90 + T91;
|
Chris@42
|
800 T9B = T91 - T90;
|
Chris@42
|
801 T8J = W[12];
|
Chris@42
|
802 T9g = FNMS(KP980785280, T8T, T8M);
|
Chris@42
|
803 T8U = FMA(KP980785280, T8T, T8M);
|
Chris@42
|
804 T9j = FMA(KP980785280, T92, T8Z);
|
Chris@42
|
805 T93 = FNMS(KP980785280, T92, T8Z);
|
Chris@42
|
806 T9f = W[44];
|
Chris@42
|
807 T8V = T8J * T8U;
|
Chris@42
|
808 }
|
Chris@42
|
809 }
|
Chris@42
|
810 T8I = FNMS(T8A, T8H, T8z);
|
Chris@42
|
811 T97 = T8J * T93;
|
Chris@42
|
812 T9n = T9f * T9j;
|
Chris@42
|
813 T9h = T9f * T9g;
|
Chris@42
|
814 T96 = FMA(T8n, T8H, T95);
|
Chris@42
|
815 T8W = W[13];
|
Chris@42
|
816 T9e = FNMS(T9c, T9d, T9b);
|
Chris@42
|
817 T9m = FMA(T99, T9d, T9l);
|
Chris@42
|
818 T9i = W[45];
|
Chris@42
|
819 T98 = FNMS(T8W, T8U, T97);
|
Chris@42
|
820 T94 = FMA(T8W, T93, T8V);
|
Chris@42
|
821 }
|
Chris@42
|
822 }
|
Chris@42
|
823 {
|
Chris@42
|
824 E T9U, T9X, T9y, Ta1, T9V, T9K, T9S, Ta0, T9W, T9M, T9I;
|
Chris@42
|
825 {
|
Chris@42
|
826 E T9p, T9R, T9x, T9u, T9N, T9Q, T9t, T9J, T9P, T9Z, T9z, T9C, T9H, T9T, T9D;
|
Chris@42
|
827 E T9L, T9E;
|
Chris@42
|
828 {
|
Chris@42
|
829 E T9o, T9k, T9O, T9s;
|
Chris@42
|
830 T9o = FNMS(T9i, T9g, T9n);
|
Chris@42
|
831 T9k = FMA(T9i, T9j, T9h);
|
Chris@42
|
832 Im[WS(rs, 3)] = T98 - T96;
|
Chris@42
|
833 Ip[WS(rs, 3)] = T96 + T98;
|
Chris@42
|
834 Rm[WS(rs, 3)] = T8I + T94;
|
Chris@42
|
835 Rp[WS(rs, 3)] = T8I - T94;
|
Chris@42
|
836 Im[WS(rs, 11)] = T9o - T9m;
|
Chris@42
|
837 Ip[WS(rs, 11)] = T9m + T9o;
|
Chris@42
|
838 Rm[WS(rs, 11)] = T9e + T9k;
|
Chris@42
|
839 Rp[WS(rs, 11)] = T9e - T9k;
|
Chris@42
|
840 T9p = W[26];
|
Chris@42
|
841 T9R = FMA(KP923879532, T9w, T9v);
|
Chris@42
|
842 T9x = FNMS(KP923879532, T9w, T9v);
|
Chris@42
|
843 T9O = FMA(KP923879532, T9r, T9q);
|
Chris@42
|
844 T9s = FNMS(KP923879532, T9r, T9q);
|
Chris@42
|
845 T9u = W[27];
|
Chris@42
|
846 T9N = W[58];
|
Chris@42
|
847 T9Q = W[59];
|
Chris@42
|
848 T9t = T9p * T9s;
|
Chris@42
|
849 T9J = T9u * T9s;
|
Chris@42
|
850 T9P = T9N * T9O;
|
Chris@42
|
851 T9Z = T9Q * T9O;
|
Chris@42
|
852 T9z = W[28];
|
Chris@42
|
853 T9U = FNMS(KP980785280, T9B, T9A);
|
Chris@42
|
854 T9C = FMA(KP980785280, T9B, T9A);
|
Chris@42
|
855 T9X = FMA(KP980785280, T9G, T9F);
|
Chris@42
|
856 T9H = FNMS(KP980785280, T9G, T9F);
|
Chris@42
|
857 T9T = W[60];
|
Chris@42
|
858 T9D = T9z * T9C;
|
Chris@42
|
859 }
|
Chris@42
|
860 T9y = FNMS(T9u, T9x, T9t);
|
Chris@42
|
861 T9L = T9z * T9H;
|
Chris@42
|
862 Ta1 = T9T * T9X;
|
Chris@42
|
863 T9V = T9T * T9U;
|
Chris@42
|
864 T9K = FMA(T9p, T9x, T9J);
|
Chris@42
|
865 T9E = W[29];
|
Chris@42
|
866 T9S = FNMS(T9Q, T9R, T9P);
|
Chris@42
|
867 Ta0 = FMA(T9N, T9R, T9Z);
|
Chris@42
|
868 T9W = W[61];
|
Chris@42
|
869 T9M = FNMS(T9E, T9C, T9L);
|
Chris@42
|
870 T9I = FMA(T9E, T9H, T9D);
|
Chris@42
|
871 }
|
Chris@42
|
872 {
|
Chris@42
|
873 E T7J, T8b, T7R, T7O, T87, T8a, T7N, T83, T89, T8j, T7T, T7W, T81, T8d, T7X;
|
Chris@42
|
874 E T85, T7Y;
|
Chris@42
|
875 {
|
Chris@42
|
876 E Ta2, T9Y, T88, T7M;
|
Chris@42
|
877 Ta2 = FNMS(T9W, T9U, Ta1);
|
Chris@42
|
878 T9Y = FMA(T9W, T9X, T9V);
|
Chris@42
|
879 Im[WS(rs, 7)] = T9M - T9K;
|
Chris@42
|
880 Ip[WS(rs, 7)] = T9K + T9M;
|
Chris@42
|
881 Rm[WS(rs, 7)] = T9y + T9I;
|
Chris@42
|
882 Rp[WS(rs, 7)] = T9y - T9I;
|
Chris@42
|
883 Im[WS(rs, 15)] = Ta2 - Ta0;
|
Chris@42
|
884 Ip[WS(rs, 15)] = Ta0 + Ta2;
|
Chris@42
|
885 Rm[WS(rs, 15)] = T9S + T9Y;
|
Chris@42
|
886 Rp[WS(rs, 15)] = T9S - T9Y;
|
Chris@42
|
887 T7J = W[18];
|
Chris@42
|
888 T8b = FNMS(KP923879532, T7Q, T7P);
|
Chris@42
|
889 T7R = FMA(KP923879532, T7Q, T7P);
|
Chris@42
|
890 T88 = FNMS(KP923879532, T7L, T7K);
|
Chris@42
|
891 T7M = FMA(KP923879532, T7L, T7K);
|
Chris@42
|
892 T7O = W[19];
|
Chris@42
|
893 T87 = W[50];
|
Chris@42
|
894 T8a = W[51];
|
Chris@42
|
895 T7N = T7J * T7M;
|
Chris@42
|
896 T83 = T7O * T7M;
|
Chris@42
|
897 T89 = T87 * T88;
|
Chris@42
|
898 T8j = T8a * T88;
|
Chris@42
|
899 T7T = W[20];
|
Chris@42
|
900 T8e = FNMS(KP831469612, T7V, T7U);
|
Chris@42
|
901 T7W = FMA(KP831469612, T7V, T7U);
|
Chris@42
|
902 T8h = FMA(KP831469612, T80, T7Z);
|
Chris@42
|
903 T81 = FNMS(KP831469612, T80, T7Z);
|
Chris@42
|
904 T8d = W[52];
|
Chris@42
|
905 T7X = T7T * T7W;
|
Chris@42
|
906 }
|
Chris@42
|
907 T7S = FNMS(T7O, T7R, T7N);
|
Chris@42
|
908 T85 = T7T * T81;
|
Chris@42
|
909 T8l = T8d * T8h;
|
Chris@42
|
910 T8f = T8d * T8e;
|
Chris@42
|
911 T84 = FMA(T7J, T7R, T83);
|
Chris@42
|
912 T7Y = W[21];
|
Chris@42
|
913 T8c = FNMS(T8a, T8b, T89);
|
Chris@42
|
914 T8k = FMA(T87, T8b, T8j);
|
Chris@42
|
915 T8g = W[53];
|
Chris@42
|
916 T86 = FNMS(T7Y, T7W, T85);
|
Chris@42
|
917 T82 = FMA(T7Y, T81, T7X);
|
Chris@42
|
918 }
|
Chris@42
|
919 }
|
Chris@42
|
920 }
|
Chris@42
|
921 }
|
Chris@42
|
922 }
|
Chris@42
|
923 T8m = FNMS(T8g, T8e, T8l);
|
Chris@42
|
924 T8i = FMA(T8g, T8h, T8f);
|
Chris@42
|
925 Im[WS(rs, 5)] = T86 - T84;
|
Chris@42
|
926 Ip[WS(rs, 5)] = T84 + T86;
|
Chris@42
|
927 Rm[WS(rs, 5)] = T7S + T82;
|
Chris@42
|
928 Rp[WS(rs, 5)] = T7S - T82;
|
Chris@42
|
929 Im[WS(rs, 13)] = T8m - T8k;
|
Chris@42
|
930 Ip[WS(rs, 13)] = T8k + T8m;
|
Chris@42
|
931 Rm[WS(rs, 13)] = T8c + T8i;
|
Chris@42
|
932 Rp[WS(rs, 13)] = T8c - T8i;
|
Chris@42
|
933 }
|
Chris@42
|
934 }
|
Chris@42
|
935 }
|
Chris@42
|
936
|
Chris@42
|
937 static const tw_instr twinstr[] = {
|
Chris@42
|
938 {TW_FULL, 1, 32},
|
Chris@42
|
939 {TW_NEXT, 1, 0}
|
Chris@42
|
940 };
|
Chris@42
|
941
|
Chris@42
|
942 static const hc2c_desc desc = { 32, "hc2cbdft_32", twinstr, &GENUS, {300, 62, 198, 0} };
|
Chris@42
|
943
|
Chris@42
|
944 void X(codelet_hc2cbdft_32) (planner *p) {
|
Chris@42
|
945 X(khc2c_register) (p, hc2cbdft_32, &desc, HC2C_VIA_DFT);
|
Chris@42
|
946 }
|
Chris@42
|
947 #else /* HAVE_FMA */
|
Chris@42
|
948
|
Chris@42
|
949 /* Generated by: ../../../genfft/gen_hc2cdft.native -compact -variables 4 -pipeline-latency 4 -sign 1 -n 32 -dif -name hc2cbdft_32 -include hc2cb.h */
|
Chris@42
|
950
|
Chris@42
|
951 /*
|
Chris@42
|
952 * This function contains 498 FP additions, 208 FP multiplications,
|
Chris@42
|
953 * (or, 404 additions, 114 multiplications, 94 fused multiply/add),
|
Chris@42
|
954 * 102 stack variables, 7 constants, and 128 memory accesses
|
Chris@42
|
955 */
|
Chris@42
|
956 #include "hc2cb.h"
|
Chris@42
|
957
|
Chris@42
|
958 static void hc2cbdft_32(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms)
|
Chris@42
|
959 {
|
Chris@42
|
960 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
Chris@42
|
961 DK(KP555570233, +0.555570233019602224742830813948532874374937191);
|
Chris@42
|
962 DK(KP195090322, +0.195090322016128267848284868477022240927691618);
|
Chris@42
|
963 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
Chris@42
|
964 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
Chris@42
|
965 DK(KP382683432, +0.382683432365089771728459984030398866761344562);
|
Chris@42
|
966 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
Chris@42
|
967 {
|
Chris@42
|
968 INT m;
|
Chris@42
|
969 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)) {
|
Chris@42
|
970 E Tf, T4a, T6h, T7Z, T6P, T8e, T1j, T4v, T2R, T4L, T5C, T7E, T6a, T7U, T3n;
|
Chris@42
|
971 E T4q, TZ, T38, T2p, T4B, T7M, T7R, T2y, T4C, T5Y, T63, T6C, T86, T4i, T4n;
|
Chris@42
|
972 E T6z, T85, TK, T31, T1Y, T4y, T7J, T7Q, T27, T4z, T5R, T62, T6v, T83, T4f;
|
Chris@42
|
973 E T4m, T6s, T82, Tu, T4p, T6o, T8f, T6M, T80, T1G, T4K, T2I, T4w, T5J, T7T;
|
Chris@42
|
974 E T67, T7F, T3g, T4b;
|
Chris@42
|
975 {
|
Chris@42
|
976 E T3, T2M, T16, T3k, T6, T13, T2P, T3l, Td, T3i, T1h, T2K, Ta, T3h, T1c;
|
Chris@42
|
977 E T2J;
|
Chris@42
|
978 {
|
Chris@42
|
979 E T1, T2, T2N, T2O;
|
Chris@42
|
980 T1 = Rp[0];
|
Chris@42
|
981 T2 = Rm[WS(rs, 15)];
|
Chris@42
|
982 T3 = T1 + T2;
|
Chris@42
|
983 T2M = T1 - T2;
|
Chris@42
|
984 {
|
Chris@42
|
985 E T14, T15, T4, T5;
|
Chris@42
|
986 T14 = Ip[0];
|
Chris@42
|
987 T15 = Im[WS(rs, 15)];
|
Chris@42
|
988 T16 = T14 + T15;
|
Chris@42
|
989 T3k = T14 - T15;
|
Chris@42
|
990 T4 = Rp[WS(rs, 8)];
|
Chris@42
|
991 T5 = Rm[WS(rs, 7)];
|
Chris@42
|
992 T6 = T4 + T5;
|
Chris@42
|
993 T13 = T4 - T5;
|
Chris@42
|
994 }
|
Chris@42
|
995 T2N = Ip[WS(rs, 8)];
|
Chris@42
|
996 T2O = Im[WS(rs, 7)];
|
Chris@42
|
997 T2P = T2N + T2O;
|
Chris@42
|
998 T3l = T2N - T2O;
|
Chris@42
|
999 {
|
Chris@42
|
1000 E Tb, Tc, T1d, T1e, T1f, T1g;
|
Chris@42
|
1001 Tb = Rm[WS(rs, 3)];
|
Chris@42
|
1002 Tc = Rp[WS(rs, 12)];
|
Chris@42
|
1003 T1d = Tb - Tc;
|
Chris@42
|
1004 T1e = Im[WS(rs, 3)];
|
Chris@42
|
1005 T1f = Ip[WS(rs, 12)];
|
Chris@42
|
1006 T1g = T1e + T1f;
|
Chris@42
|
1007 Td = Tb + Tc;
|
Chris@42
|
1008 T3i = T1f - T1e;
|
Chris@42
|
1009 T1h = T1d + T1g;
|
Chris@42
|
1010 T2K = T1d - T1g;
|
Chris@42
|
1011 }
|
Chris@42
|
1012 {
|
Chris@42
|
1013 E T8, T9, T18, T19, T1a, T1b;
|
Chris@42
|
1014 T8 = Rp[WS(rs, 4)];
|
Chris@42
|
1015 T9 = Rm[WS(rs, 11)];
|
Chris@42
|
1016 T18 = T8 - T9;
|
Chris@42
|
1017 T19 = Ip[WS(rs, 4)];
|
Chris@42
|
1018 T1a = Im[WS(rs, 11)];
|
Chris@42
|
1019 T1b = T19 + T1a;
|
Chris@42
|
1020 Ta = T8 + T9;
|
Chris@42
|
1021 T3h = T19 - T1a;
|
Chris@42
|
1022 T1c = T18 + T1b;
|
Chris@42
|
1023 T2J = T18 - T1b;
|
Chris@42
|
1024 }
|
Chris@42
|
1025 }
|
Chris@42
|
1026 {
|
Chris@42
|
1027 E T7, Te, T6f, T6g;
|
Chris@42
|
1028 T7 = T3 + T6;
|
Chris@42
|
1029 Te = Ta + Td;
|
Chris@42
|
1030 Tf = T7 + Te;
|
Chris@42
|
1031 T4a = T7 - Te;
|
Chris@42
|
1032 T6f = T16 - T13;
|
Chris@42
|
1033 T6g = KP707106781 * (T2J - T2K);
|
Chris@42
|
1034 T6h = T6f + T6g;
|
Chris@42
|
1035 T7Z = T6f - T6g;
|
Chris@42
|
1036 }
|
Chris@42
|
1037 {
|
Chris@42
|
1038 E T6N, T6O, T17, T1i;
|
Chris@42
|
1039 T6N = T2M + T2P;
|
Chris@42
|
1040 T6O = KP707106781 * (T1c + T1h);
|
Chris@42
|
1041 T6P = T6N - T6O;
|
Chris@42
|
1042 T8e = T6O + T6N;
|
Chris@42
|
1043 T17 = T13 + T16;
|
Chris@42
|
1044 T1i = KP707106781 * (T1c - T1h);
|
Chris@42
|
1045 T1j = T17 + T1i;
|
Chris@42
|
1046 T4v = T17 - T1i;
|
Chris@42
|
1047 }
|
Chris@42
|
1048 {
|
Chris@42
|
1049 E T2L, T2Q, T5A, T5B;
|
Chris@42
|
1050 T2L = KP707106781 * (T2J + T2K);
|
Chris@42
|
1051 T2Q = T2M - T2P;
|
Chris@42
|
1052 T2R = T2L + T2Q;
|
Chris@42
|
1053 T4L = T2Q - T2L;
|
Chris@42
|
1054 T5A = T3 - T6;
|
Chris@42
|
1055 T5B = T3i - T3h;
|
Chris@42
|
1056 T5C = T5A + T5B;
|
Chris@42
|
1057 T7E = T5A - T5B;
|
Chris@42
|
1058 }
|
Chris@42
|
1059 {
|
Chris@42
|
1060 E T68, T69, T3j, T3m;
|
Chris@42
|
1061 T68 = Ta - Td;
|
Chris@42
|
1062 T69 = T3k - T3l;
|
Chris@42
|
1063 T6a = T68 + T69;
|
Chris@42
|
1064 T7U = T69 - T68;
|
Chris@42
|
1065 T3j = T3h + T3i;
|
Chris@42
|
1066 T3m = T3k + T3l;
|
Chris@42
|
1067 T3n = T3j + T3m;
|
Chris@42
|
1068 T4q = T3m - T3j;
|
Chris@42
|
1069 }
|
Chris@42
|
1070 }
|
Chris@42
|
1071 {
|
Chris@42
|
1072 E TR, T5S, T29, T2t, T2c, T5W, T2w, T37, TY, T5T, T5V, T2i, T2n, T2r, T34;
|
Chris@42
|
1073 E T2q, T6A, T6B;
|
Chris@42
|
1074 {
|
Chris@42
|
1075 E TL, TM, TN, TO, TP, TQ;
|
Chris@42
|
1076 TL = Rm[0];
|
Chris@42
|
1077 TM = Rp[WS(rs, 15)];
|
Chris@42
|
1078 TN = TL + TM;
|
Chris@42
|
1079 TO = Rp[WS(rs, 7)];
|
Chris@42
|
1080 TP = Rm[WS(rs, 8)];
|
Chris@42
|
1081 TQ = TO + TP;
|
Chris@42
|
1082 TR = TN + TQ;
|
Chris@42
|
1083 T5S = TN - TQ;
|
Chris@42
|
1084 T29 = TO - TP;
|
Chris@42
|
1085 T2t = TL - TM;
|
Chris@42
|
1086 }
|
Chris@42
|
1087 {
|
Chris@42
|
1088 E T2a, T2b, T35, T2u, T2v, T36;
|
Chris@42
|
1089 T2a = Im[0];
|
Chris@42
|
1090 T2b = Ip[WS(rs, 15)];
|
Chris@42
|
1091 T35 = T2b - T2a;
|
Chris@42
|
1092 T2u = Ip[WS(rs, 7)];
|
Chris@42
|
1093 T2v = Im[WS(rs, 8)];
|
Chris@42
|
1094 T36 = T2u - T2v;
|
Chris@42
|
1095 T2c = T2a + T2b;
|
Chris@42
|
1096 T5W = T35 - T36;
|
Chris@42
|
1097 T2w = T2u + T2v;
|
Chris@42
|
1098 T37 = T35 + T36;
|
Chris@42
|
1099 }
|
Chris@42
|
1100 {
|
Chris@42
|
1101 E TU, T2e, T2h, T32, TX, T2j, T2m, T33;
|
Chris@42
|
1102 {
|
Chris@42
|
1103 E TS, TT, T2f, T2g;
|
Chris@42
|
1104 TS = Rp[WS(rs, 3)];
|
Chris@42
|
1105 TT = Rm[WS(rs, 12)];
|
Chris@42
|
1106 TU = TS + TT;
|
Chris@42
|
1107 T2e = TS - TT;
|
Chris@42
|
1108 T2f = Ip[WS(rs, 3)];
|
Chris@42
|
1109 T2g = Im[WS(rs, 12)];
|
Chris@42
|
1110 T2h = T2f + T2g;
|
Chris@42
|
1111 T32 = T2f - T2g;
|
Chris@42
|
1112 }
|
Chris@42
|
1113 {
|
Chris@42
|
1114 E TV, TW, T2k, T2l;
|
Chris@42
|
1115 TV = Rm[WS(rs, 4)];
|
Chris@42
|
1116 TW = Rp[WS(rs, 11)];
|
Chris@42
|
1117 TX = TV + TW;
|
Chris@42
|
1118 T2j = TV - TW;
|
Chris@42
|
1119 T2k = Im[WS(rs, 4)];
|
Chris@42
|
1120 T2l = Ip[WS(rs, 11)];
|
Chris@42
|
1121 T2m = T2k + T2l;
|
Chris@42
|
1122 T33 = T2l - T2k;
|
Chris@42
|
1123 }
|
Chris@42
|
1124 TY = TU + TX;
|
Chris@42
|
1125 T5T = T33 - T32;
|
Chris@42
|
1126 T5V = TU - TX;
|
Chris@42
|
1127 T2i = T2e + T2h;
|
Chris@42
|
1128 T2n = T2j + T2m;
|
Chris@42
|
1129 T2r = T2j - T2m;
|
Chris@42
|
1130 T34 = T32 + T33;
|
Chris@42
|
1131 T2q = T2e - T2h;
|
Chris@42
|
1132 }
|
Chris@42
|
1133 TZ = TR + TY;
|
Chris@42
|
1134 T38 = T34 + T37;
|
Chris@42
|
1135 {
|
Chris@42
|
1136 E T2d, T2o, T7K, T7L;
|
Chris@42
|
1137 T2d = T29 - T2c;
|
Chris@42
|
1138 T2o = KP707106781 * (T2i - T2n);
|
Chris@42
|
1139 T2p = T2d + T2o;
|
Chris@42
|
1140 T4B = T2d - T2o;
|
Chris@42
|
1141 T7K = T5S - T5T;
|
Chris@42
|
1142 T7L = T5W - T5V;
|
Chris@42
|
1143 T7M = FMA(KP382683432, T7K, KP923879532 * T7L);
|
Chris@42
|
1144 T7R = FNMS(KP923879532, T7K, KP382683432 * T7L);
|
Chris@42
|
1145 }
|
Chris@42
|
1146 {
|
Chris@42
|
1147 E T2s, T2x, T5U, T5X;
|
Chris@42
|
1148 T2s = KP707106781 * (T2q + T2r);
|
Chris@42
|
1149 T2x = T2t - T2w;
|
Chris@42
|
1150 T2y = T2s + T2x;
|
Chris@42
|
1151 T4C = T2x - T2s;
|
Chris@42
|
1152 T5U = T5S + T5T;
|
Chris@42
|
1153 T5X = T5V + T5W;
|
Chris@42
|
1154 T5Y = FMA(KP923879532, T5U, KP382683432 * T5X);
|
Chris@42
|
1155 T63 = FNMS(KP382683432, T5U, KP923879532 * T5X);
|
Chris@42
|
1156 }
|
Chris@42
|
1157 T6A = T2t + T2w;
|
Chris@42
|
1158 T6B = KP707106781 * (T2i + T2n);
|
Chris@42
|
1159 T6C = T6A - T6B;
|
Chris@42
|
1160 T86 = T6B + T6A;
|
Chris@42
|
1161 {
|
Chris@42
|
1162 E T4g, T4h, T6x, T6y;
|
Chris@42
|
1163 T4g = TR - TY;
|
Chris@42
|
1164 T4h = T37 - T34;
|
Chris@42
|
1165 T4i = T4g + T4h;
|
Chris@42
|
1166 T4n = T4h - T4g;
|
Chris@42
|
1167 T6x = KP707106781 * (T2q - T2r);
|
Chris@42
|
1168 T6y = T29 + T2c;
|
Chris@42
|
1169 T6z = T6x - T6y;
|
Chris@42
|
1170 T85 = T6y + T6x;
|
Chris@42
|
1171 }
|
Chris@42
|
1172 }
|
Chris@42
|
1173 {
|
Chris@42
|
1174 E TC, T5L, T1I, T22, T1L, T5P, T25, T30, TJ, T5M, T5O, T1R, T1W, T20, T2X;
|
Chris@42
|
1175 E T1Z, T6t, T6u;
|
Chris@42
|
1176 {
|
Chris@42
|
1177 E Tw, Tx, Ty, Tz, TA, TB;
|
Chris@42
|
1178 Tw = Rp[WS(rs, 1)];
|
Chris@42
|
1179 Tx = Rm[WS(rs, 14)];
|
Chris@42
|
1180 Ty = Tw + Tx;
|
Chris@42
|
1181 Tz = Rp[WS(rs, 9)];
|
Chris@42
|
1182 TA = Rm[WS(rs, 6)];
|
Chris@42
|
1183 TB = Tz + TA;
|
Chris@42
|
1184 TC = Ty + TB;
|
Chris@42
|
1185 T5L = Ty - TB;
|
Chris@42
|
1186 T1I = Tz - TA;
|
Chris@42
|
1187 T22 = Tw - Tx;
|
Chris@42
|
1188 }
|
Chris@42
|
1189 {
|
Chris@42
|
1190 E T1J, T1K, T2Y, T23, T24, T2Z;
|
Chris@42
|
1191 T1J = Ip[WS(rs, 1)];
|
Chris@42
|
1192 T1K = Im[WS(rs, 14)];
|
Chris@42
|
1193 T2Y = T1J - T1K;
|
Chris@42
|
1194 T23 = Ip[WS(rs, 9)];
|
Chris@42
|
1195 T24 = Im[WS(rs, 6)];
|
Chris@42
|
1196 T2Z = T23 - T24;
|
Chris@42
|
1197 T1L = T1J + T1K;
|
Chris@42
|
1198 T5P = T2Y - T2Z;
|
Chris@42
|
1199 T25 = T23 + T24;
|
Chris@42
|
1200 T30 = T2Y + T2Z;
|
Chris@42
|
1201 }
|
Chris@42
|
1202 {
|
Chris@42
|
1203 E TF, T1N, T1Q, T2V, TI, T1S, T1V, T2W;
|
Chris@42
|
1204 {
|
Chris@42
|
1205 E TD, TE, T1O, T1P;
|
Chris@42
|
1206 TD = Rp[WS(rs, 5)];
|
Chris@42
|
1207 TE = Rm[WS(rs, 10)];
|
Chris@42
|
1208 TF = TD + TE;
|
Chris@42
|
1209 T1N = TD - TE;
|
Chris@42
|
1210 T1O = Ip[WS(rs, 5)];
|
Chris@42
|
1211 T1P = Im[WS(rs, 10)];
|
Chris@42
|
1212 T1Q = T1O + T1P;
|
Chris@42
|
1213 T2V = T1O - T1P;
|
Chris@42
|
1214 }
|
Chris@42
|
1215 {
|
Chris@42
|
1216 E TG, TH, T1T, T1U;
|
Chris@42
|
1217 TG = Rm[WS(rs, 2)];
|
Chris@42
|
1218 TH = Rp[WS(rs, 13)];
|
Chris@42
|
1219 TI = TG + TH;
|
Chris@42
|
1220 T1S = TG - TH;
|
Chris@42
|
1221 T1T = Im[WS(rs, 2)];
|
Chris@42
|
1222 T1U = Ip[WS(rs, 13)];
|
Chris@42
|
1223 T1V = T1T + T1U;
|
Chris@42
|
1224 T2W = T1U - T1T;
|
Chris@42
|
1225 }
|
Chris@42
|
1226 TJ = TF + TI;
|
Chris@42
|
1227 T5M = T2W - T2V;
|
Chris@42
|
1228 T5O = TF - TI;
|
Chris@42
|
1229 T1R = T1N + T1Q;
|
Chris@42
|
1230 T1W = T1S + T1V;
|
Chris@42
|
1231 T20 = T1S - T1V;
|
Chris@42
|
1232 T2X = T2V + T2W;
|
Chris@42
|
1233 T1Z = T1N - T1Q;
|
Chris@42
|
1234 }
|
Chris@42
|
1235 TK = TC + TJ;
|
Chris@42
|
1236 T31 = T2X + T30;
|
Chris@42
|
1237 {
|
Chris@42
|
1238 E T1M, T1X, T7H, T7I;
|
Chris@42
|
1239 T1M = T1I + T1L;
|
Chris@42
|
1240 T1X = KP707106781 * (T1R - T1W);
|
Chris@42
|
1241 T1Y = T1M + T1X;
|
Chris@42
|
1242 T4y = T1M - T1X;
|
Chris@42
|
1243 T7H = T5L - T5M;
|
Chris@42
|
1244 T7I = T5P - T5O;
|
Chris@42
|
1245 T7J = FNMS(KP923879532, T7I, KP382683432 * T7H);
|
Chris@42
|
1246 T7Q = FMA(KP923879532, T7H, KP382683432 * T7I);
|
Chris@42
|
1247 }
|
Chris@42
|
1248 {
|
Chris@42
|
1249 E T21, T26, T5N, T5Q;
|
Chris@42
|
1250 T21 = KP707106781 * (T1Z + T20);
|
Chris@42
|
1251 T26 = T22 - T25;
|
Chris@42
|
1252 T27 = T21 + T26;
|
Chris@42
|
1253 T4z = T26 - T21;
|
Chris@42
|
1254 T5N = T5L + T5M;
|
Chris@42
|
1255 T5Q = T5O + T5P;
|
Chris@42
|
1256 T5R = FNMS(KP382683432, T5Q, KP923879532 * T5N);
|
Chris@42
|
1257 T62 = FMA(KP382683432, T5N, KP923879532 * T5Q);
|
Chris@42
|
1258 }
|
Chris@42
|
1259 T6t = T22 + T25;
|
Chris@42
|
1260 T6u = KP707106781 * (T1R + T1W);
|
Chris@42
|
1261 T6v = T6t - T6u;
|
Chris@42
|
1262 T83 = T6u + T6t;
|
Chris@42
|
1263 {
|
Chris@42
|
1264 E T4d, T4e, T6q, T6r;
|
Chris@42
|
1265 T4d = TC - TJ;
|
Chris@42
|
1266 T4e = T30 - T2X;
|
Chris@42
|
1267 T4f = T4d - T4e;
|
Chris@42
|
1268 T4m = T4d + T4e;
|
Chris@42
|
1269 T6q = T1L - T1I;
|
Chris@42
|
1270 T6r = KP707106781 * (T1Z - T20);
|
Chris@42
|
1271 T6s = T6q + T6r;
|
Chris@42
|
1272 T82 = T6q - T6r;
|
Chris@42
|
1273 }
|
Chris@42
|
1274 }
|
Chris@42
|
1275 {
|
Chris@42
|
1276 E Ti, T3a, Tl, T3b, T1o, T1t, T6j, T6i, T5E, T5D, Tp, T3d, Ts, T3e, T1z;
|
Chris@42
|
1277 E T1E, T6m, T6l, T5H, T5G;
|
Chris@42
|
1278 {
|
Chris@42
|
1279 E T1p, T1n, T1k, T1s;
|
Chris@42
|
1280 {
|
Chris@42
|
1281 E Tg, Th, T1l, T1m;
|
Chris@42
|
1282 Tg = Rp[WS(rs, 2)];
|
Chris@42
|
1283 Th = Rm[WS(rs, 13)];
|
Chris@42
|
1284 Ti = Tg + Th;
|
Chris@42
|
1285 T1p = Tg - Th;
|
Chris@42
|
1286 T1l = Ip[WS(rs, 2)];
|
Chris@42
|
1287 T1m = Im[WS(rs, 13)];
|
Chris@42
|
1288 T1n = T1l + T1m;
|
Chris@42
|
1289 T3a = T1l - T1m;
|
Chris@42
|
1290 }
|
Chris@42
|
1291 {
|
Chris@42
|
1292 E Tj, Tk, T1q, T1r;
|
Chris@42
|
1293 Tj = Rp[WS(rs, 10)];
|
Chris@42
|
1294 Tk = Rm[WS(rs, 5)];
|
Chris@42
|
1295 Tl = Tj + Tk;
|
Chris@42
|
1296 T1k = Tj - Tk;
|
Chris@42
|
1297 T1q = Ip[WS(rs, 10)];
|
Chris@42
|
1298 T1r = Im[WS(rs, 5)];
|
Chris@42
|
1299 T1s = T1q + T1r;
|
Chris@42
|
1300 T3b = T1q - T1r;
|
Chris@42
|
1301 }
|
Chris@42
|
1302 T1o = T1k + T1n;
|
Chris@42
|
1303 T1t = T1p - T1s;
|
Chris@42
|
1304 T6j = T1p + T1s;
|
Chris@42
|
1305 T6i = T1n - T1k;
|
Chris@42
|
1306 T5E = T3a - T3b;
|
Chris@42
|
1307 T5D = Ti - Tl;
|
Chris@42
|
1308 }
|
Chris@42
|
1309 {
|
Chris@42
|
1310 E T1A, T1y, T1v, T1D;
|
Chris@42
|
1311 {
|
Chris@42
|
1312 E Tn, To, T1w, T1x;
|
Chris@42
|
1313 Tn = Rm[WS(rs, 1)];
|
Chris@42
|
1314 To = Rp[WS(rs, 14)];
|
Chris@42
|
1315 Tp = Tn + To;
|
Chris@42
|
1316 T1A = Tn - To;
|
Chris@42
|
1317 T1w = Im[WS(rs, 1)];
|
Chris@42
|
1318 T1x = Ip[WS(rs, 14)];
|
Chris@42
|
1319 T1y = T1w + T1x;
|
Chris@42
|
1320 T3d = T1x - T1w;
|
Chris@42
|
1321 }
|
Chris@42
|
1322 {
|
Chris@42
|
1323 E Tq, Tr, T1B, T1C;
|
Chris@42
|
1324 Tq = Rp[WS(rs, 6)];
|
Chris@42
|
1325 Tr = Rm[WS(rs, 9)];
|
Chris@42
|
1326 Ts = Tq + Tr;
|
Chris@42
|
1327 T1v = Tq - Tr;
|
Chris@42
|
1328 T1B = Ip[WS(rs, 6)];
|
Chris@42
|
1329 T1C = Im[WS(rs, 9)];
|
Chris@42
|
1330 T1D = T1B + T1C;
|
Chris@42
|
1331 T3e = T1B - T1C;
|
Chris@42
|
1332 }
|
Chris@42
|
1333 T1z = T1v - T1y;
|
Chris@42
|
1334 T1E = T1A - T1D;
|
Chris@42
|
1335 T6m = T1A + T1D;
|
Chris@42
|
1336 T6l = T1v + T1y;
|
Chris@42
|
1337 T5H = T3d - T3e;
|
Chris@42
|
1338 T5G = Tp - Ts;
|
Chris@42
|
1339 }
|
Chris@42
|
1340 {
|
Chris@42
|
1341 E Tm, Tt, T6k, T6n;
|
Chris@42
|
1342 Tm = Ti + Tl;
|
Chris@42
|
1343 Tt = Tp + Ts;
|
Chris@42
|
1344 Tu = Tm + Tt;
|
Chris@42
|
1345 T4p = Tm - Tt;
|
Chris@42
|
1346 T6k = FMA(KP382683432, T6i, KP923879532 * T6j);
|
Chris@42
|
1347 T6n = FMA(KP382683432, T6l, KP923879532 * T6m);
|
Chris@42
|
1348 T6o = T6k - T6n;
|
Chris@42
|
1349 T8f = T6k + T6n;
|
Chris@42
|
1350 }
|
Chris@42
|
1351 {
|
Chris@42
|
1352 E T6K, T6L, T1u, T1F;
|
Chris@42
|
1353 T6K = FNMS(KP923879532, T6i, KP382683432 * T6j);
|
Chris@42
|
1354 T6L = FNMS(KP923879532, T6l, KP382683432 * T6m);
|
Chris@42
|
1355 T6M = T6K + T6L;
|
Chris@42
|
1356 T80 = T6K - T6L;
|
Chris@42
|
1357 T1u = FMA(KP923879532, T1o, KP382683432 * T1t);
|
Chris@42
|
1358 T1F = FNMS(KP382683432, T1E, KP923879532 * T1z);
|
Chris@42
|
1359 T1G = T1u + T1F;
|
Chris@42
|
1360 T4K = T1F - T1u;
|
Chris@42
|
1361 }
|
Chris@42
|
1362 {
|
Chris@42
|
1363 E T2G, T2H, T5F, T5I;
|
Chris@42
|
1364 T2G = FNMS(KP382683432, T1o, KP923879532 * T1t);
|
Chris@42
|
1365 T2H = FMA(KP382683432, T1z, KP923879532 * T1E);
|
Chris@42
|
1366 T2I = T2G + T2H;
|
Chris@42
|
1367 T4w = T2G - T2H;
|
Chris@42
|
1368 T5F = T5D - T5E;
|
Chris@42
|
1369 T5I = T5G + T5H;
|
Chris@42
|
1370 T5J = KP707106781 * (T5F + T5I);
|
Chris@42
|
1371 T7T = KP707106781 * (T5F - T5I);
|
Chris@42
|
1372 }
|
Chris@42
|
1373 {
|
Chris@42
|
1374 E T65, T66, T3c, T3f;
|
Chris@42
|
1375 T65 = T5D + T5E;
|
Chris@42
|
1376 T66 = T5H - T5G;
|
Chris@42
|
1377 T67 = KP707106781 * (T65 + T66);
|
Chris@42
|
1378 T7F = KP707106781 * (T66 - T65);
|
Chris@42
|
1379 T3c = T3a + T3b;
|
Chris@42
|
1380 T3f = T3d + T3e;
|
Chris@42
|
1381 T3g = T3c + T3f;
|
Chris@42
|
1382 T4b = T3f - T3c;
|
Chris@42
|
1383 }
|
Chris@42
|
1384 }
|
Chris@42
|
1385 {
|
Chris@42
|
1386 E T11, T3s, T3p, T3u, T3K, T40, T3G, T3Y, T2T, T43, T3z, T3P, T2B, T45, T3x;
|
Chris@42
|
1387 E T3T;
|
Chris@42
|
1388 {
|
Chris@42
|
1389 E Tv, T10, T3E, T3F;
|
Chris@42
|
1390 Tv = Tf + Tu;
|
Chris@42
|
1391 T10 = TK + TZ;
|
Chris@42
|
1392 T11 = Tv + T10;
|
Chris@42
|
1393 T3s = Tv - T10;
|
Chris@42
|
1394 {
|
Chris@42
|
1395 E T39, T3o, T3I, T3J;
|
Chris@42
|
1396 T39 = T31 + T38;
|
Chris@42
|
1397 T3o = T3g + T3n;
|
Chris@42
|
1398 T3p = T39 + T3o;
|
Chris@42
|
1399 T3u = T3o - T39;
|
Chris@42
|
1400 T3I = TK - TZ;
|
Chris@42
|
1401 T3J = T3n - T3g;
|
Chris@42
|
1402 T3K = T3I + T3J;
|
Chris@42
|
1403 T40 = T3J - T3I;
|
Chris@42
|
1404 }
|
Chris@42
|
1405 T3E = Tf - Tu;
|
Chris@42
|
1406 T3F = T38 - T31;
|
Chris@42
|
1407 T3G = T3E + T3F;
|
Chris@42
|
1408 T3Y = T3E - T3F;
|
Chris@42
|
1409 {
|
Chris@42
|
1410 E T2S, T3N, T2F, T3O, T2D, T2E;
|
Chris@42
|
1411 T2S = T2I + T2R;
|
Chris@42
|
1412 T3N = T1j - T1G;
|
Chris@42
|
1413 T2D = FNMS(KP195090322, T1Y, KP980785280 * T27);
|
Chris@42
|
1414 T2E = FMA(KP195090322, T2p, KP980785280 * T2y);
|
Chris@42
|
1415 T2F = T2D + T2E;
|
Chris@42
|
1416 T3O = T2D - T2E;
|
Chris@42
|
1417 T2T = T2F + T2S;
|
Chris@42
|
1418 T43 = T3N - T3O;
|
Chris@42
|
1419 T3z = T2S - T2F;
|
Chris@42
|
1420 T3P = T3N + T3O;
|
Chris@42
|
1421 }
|
Chris@42
|
1422 {
|
Chris@42
|
1423 E T1H, T3S, T2A, T3R, T28, T2z;
|
Chris@42
|
1424 T1H = T1j + T1G;
|
Chris@42
|
1425 T3S = T2R - T2I;
|
Chris@42
|
1426 T28 = FMA(KP980785280, T1Y, KP195090322 * T27);
|
Chris@42
|
1427 T2z = FNMS(KP195090322, T2y, KP980785280 * T2p);
|
Chris@42
|
1428 T2A = T28 + T2z;
|
Chris@42
|
1429 T3R = T2z - T28;
|
Chris@42
|
1430 T2B = T1H + T2A;
|
Chris@42
|
1431 T45 = T3S - T3R;
|
Chris@42
|
1432 T3x = T1H - T2A;
|
Chris@42
|
1433 T3T = T3R + T3S;
|
Chris@42
|
1434 }
|
Chris@42
|
1435 }
|
Chris@42
|
1436 {
|
Chris@42
|
1437 E T2U, T3q, T12, T2C;
|
Chris@42
|
1438 T12 = W[0];
|
Chris@42
|
1439 T2C = W[1];
|
Chris@42
|
1440 T2U = FMA(T12, T2B, T2C * T2T);
|
Chris@42
|
1441 T3q = FNMS(T2C, T2B, T12 * T2T);
|
Chris@42
|
1442 Rp[0] = T11 - T2U;
|
Chris@42
|
1443 Ip[0] = T3p + T3q;
|
Chris@42
|
1444 Rm[0] = T11 + T2U;
|
Chris@42
|
1445 Im[0] = T3q - T3p;
|
Chris@42
|
1446 }
|
Chris@42
|
1447 {
|
Chris@42
|
1448 E T41, T47, T46, T48;
|
Chris@42
|
1449 {
|
Chris@42
|
1450 E T3X, T3Z, T42, T44;
|
Chris@42
|
1451 T3X = W[46];
|
Chris@42
|
1452 T3Z = W[47];
|
Chris@42
|
1453 T41 = FNMS(T3Z, T40, T3X * T3Y);
|
Chris@42
|
1454 T47 = FMA(T3Z, T3Y, T3X * T40);
|
Chris@42
|
1455 T42 = W[48];
|
Chris@42
|
1456 T44 = W[49];
|
Chris@42
|
1457 T46 = FMA(T42, T43, T44 * T45);
|
Chris@42
|
1458 T48 = FNMS(T44, T43, T42 * T45);
|
Chris@42
|
1459 }
|
Chris@42
|
1460 Rp[WS(rs, 12)] = T41 - T46;
|
Chris@42
|
1461 Ip[WS(rs, 12)] = T47 + T48;
|
Chris@42
|
1462 Rm[WS(rs, 12)] = T41 + T46;
|
Chris@42
|
1463 Im[WS(rs, 12)] = T48 - T47;
|
Chris@42
|
1464 }
|
Chris@42
|
1465 {
|
Chris@42
|
1466 E T3v, T3B, T3A, T3C;
|
Chris@42
|
1467 {
|
Chris@42
|
1468 E T3r, T3t, T3w, T3y;
|
Chris@42
|
1469 T3r = W[30];
|
Chris@42
|
1470 T3t = W[31];
|
Chris@42
|
1471 T3v = FNMS(T3t, T3u, T3r * T3s);
|
Chris@42
|
1472 T3B = FMA(T3t, T3s, T3r * T3u);
|
Chris@42
|
1473 T3w = W[32];
|
Chris@42
|
1474 T3y = W[33];
|
Chris@42
|
1475 T3A = FMA(T3w, T3x, T3y * T3z);
|
Chris@42
|
1476 T3C = FNMS(T3y, T3x, T3w * T3z);
|
Chris@42
|
1477 }
|
Chris@42
|
1478 Rp[WS(rs, 8)] = T3v - T3A;
|
Chris@42
|
1479 Ip[WS(rs, 8)] = T3B + T3C;
|
Chris@42
|
1480 Rm[WS(rs, 8)] = T3v + T3A;
|
Chris@42
|
1481 Im[WS(rs, 8)] = T3C - T3B;
|
Chris@42
|
1482 }
|
Chris@42
|
1483 {
|
Chris@42
|
1484 E T3L, T3V, T3U, T3W;
|
Chris@42
|
1485 {
|
Chris@42
|
1486 E T3D, T3H, T3M, T3Q;
|
Chris@42
|
1487 T3D = W[14];
|
Chris@42
|
1488 T3H = W[15];
|
Chris@42
|
1489 T3L = FNMS(T3H, T3K, T3D * T3G);
|
Chris@42
|
1490 T3V = FMA(T3H, T3G, T3D * T3K);
|
Chris@42
|
1491 T3M = W[16];
|
Chris@42
|
1492 T3Q = W[17];
|
Chris@42
|
1493 T3U = FMA(T3M, T3P, T3Q * T3T);
|
Chris@42
|
1494 T3W = FNMS(T3Q, T3P, T3M * T3T);
|
Chris@42
|
1495 }
|
Chris@42
|
1496 Rp[WS(rs, 4)] = T3L - T3U;
|
Chris@42
|
1497 Ip[WS(rs, 4)] = T3V + T3W;
|
Chris@42
|
1498 Rm[WS(rs, 4)] = T3L + T3U;
|
Chris@42
|
1499 Im[WS(rs, 4)] = T3W - T3V;
|
Chris@42
|
1500 }
|
Chris@42
|
1501 }
|
Chris@42
|
1502 {
|
Chris@42
|
1503 E T7O, T8m, T7W, T8o, T8E, T8U, T8A, T8S, T8h, T8X, T8t, T8J, T89, T8Z, T8r;
|
Chris@42
|
1504 E T8N;
|
Chris@42
|
1505 {
|
Chris@42
|
1506 E T7G, T7N, T8y, T8z;
|
Chris@42
|
1507 T7G = T7E + T7F;
|
Chris@42
|
1508 T7N = T7J + T7M;
|
Chris@42
|
1509 T7O = T7G + T7N;
|
Chris@42
|
1510 T8m = T7G - T7N;
|
Chris@42
|
1511 {
|
Chris@42
|
1512 E T7S, T7V, T8C, T8D;
|
Chris@42
|
1513 T7S = T7Q + T7R;
|
Chris@42
|
1514 T7V = T7T + T7U;
|
Chris@42
|
1515 T7W = T7S + T7V;
|
Chris@42
|
1516 T8o = T7V - T7S;
|
Chris@42
|
1517 T8C = T7J - T7M;
|
Chris@42
|
1518 T8D = T7U - T7T;
|
Chris@42
|
1519 T8E = T8C + T8D;
|
Chris@42
|
1520 T8U = T8D - T8C;
|
Chris@42
|
1521 }
|
Chris@42
|
1522 T8y = T7E - T7F;
|
Chris@42
|
1523 T8z = T7R - T7Q;
|
Chris@42
|
1524 T8A = T8y + T8z;
|
Chris@42
|
1525 T8S = T8y - T8z;
|
Chris@42
|
1526 {
|
Chris@42
|
1527 E T8g, T8H, T8d, T8I, T8b, T8c;
|
Chris@42
|
1528 T8g = T8e - T8f;
|
Chris@42
|
1529 T8H = T7Z - T80;
|
Chris@42
|
1530 T8b = FNMS(KP980785280, T82, KP195090322 * T83);
|
Chris@42
|
1531 T8c = FNMS(KP980785280, T85, KP195090322 * T86);
|
Chris@42
|
1532 T8d = T8b + T8c;
|
Chris@42
|
1533 T8I = T8b - T8c;
|
Chris@42
|
1534 T8h = T8d + T8g;
|
Chris@42
|
1535 T8X = T8H - T8I;
|
Chris@42
|
1536 T8t = T8g - T8d;
|
Chris@42
|
1537 T8J = T8H + T8I;
|
Chris@42
|
1538 }
|
Chris@42
|
1539 {
|
Chris@42
|
1540 E T81, T8L, T88, T8M, T84, T87;
|
Chris@42
|
1541 T81 = T7Z + T80;
|
Chris@42
|
1542 T8L = T8f + T8e;
|
Chris@42
|
1543 T84 = FMA(KP195090322, T82, KP980785280 * T83);
|
Chris@42
|
1544 T87 = FMA(KP195090322, T85, KP980785280 * T86);
|
Chris@42
|
1545 T88 = T84 - T87;
|
Chris@42
|
1546 T8M = T84 + T87;
|
Chris@42
|
1547 T89 = T81 + T88;
|
Chris@42
|
1548 T8Z = T8M + T8L;
|
Chris@42
|
1549 T8r = T81 - T88;
|
Chris@42
|
1550 T8N = T8L - T8M;
|
Chris@42
|
1551 }
|
Chris@42
|
1552 }
|
Chris@42
|
1553 {
|
Chris@42
|
1554 E T7X, T8j, T8i, T8k;
|
Chris@42
|
1555 {
|
Chris@42
|
1556 E T7D, T7P, T7Y, T8a;
|
Chris@42
|
1557 T7D = W[10];
|
Chris@42
|
1558 T7P = W[11];
|
Chris@42
|
1559 T7X = FNMS(T7P, T7W, T7D * T7O);
|
Chris@42
|
1560 T8j = FMA(T7P, T7O, T7D * T7W);
|
Chris@42
|
1561 T7Y = W[12];
|
Chris@42
|
1562 T8a = W[13];
|
Chris@42
|
1563 T8i = FMA(T7Y, T89, T8a * T8h);
|
Chris@42
|
1564 T8k = FNMS(T8a, T89, T7Y * T8h);
|
Chris@42
|
1565 }
|
Chris@42
|
1566 Rp[WS(rs, 3)] = T7X - T8i;
|
Chris@42
|
1567 Ip[WS(rs, 3)] = T8j + T8k;
|
Chris@42
|
1568 Rm[WS(rs, 3)] = T7X + T8i;
|
Chris@42
|
1569 Im[WS(rs, 3)] = T8k - T8j;
|
Chris@42
|
1570 }
|
Chris@42
|
1571 {
|
Chris@42
|
1572 E T8V, T91, T90, T92;
|
Chris@42
|
1573 {
|
Chris@42
|
1574 E T8R, T8T, T8W, T8Y;
|
Chris@42
|
1575 T8R = W[58];
|
Chris@42
|
1576 T8T = W[59];
|
Chris@42
|
1577 T8V = FNMS(T8T, T8U, T8R * T8S);
|
Chris@42
|
1578 T91 = FMA(T8T, T8S, T8R * T8U);
|
Chris@42
|
1579 T8W = W[60];
|
Chris@42
|
1580 T8Y = W[61];
|
Chris@42
|
1581 T90 = FMA(T8W, T8X, T8Y * T8Z);
|
Chris@42
|
1582 T92 = FNMS(T8Y, T8X, T8W * T8Z);
|
Chris@42
|
1583 }
|
Chris@42
|
1584 Rp[WS(rs, 15)] = T8V - T90;
|
Chris@42
|
1585 Ip[WS(rs, 15)] = T91 + T92;
|
Chris@42
|
1586 Rm[WS(rs, 15)] = T8V + T90;
|
Chris@42
|
1587 Im[WS(rs, 15)] = T92 - T91;
|
Chris@42
|
1588 }
|
Chris@42
|
1589 {
|
Chris@42
|
1590 E T8p, T8v, T8u, T8w;
|
Chris@42
|
1591 {
|
Chris@42
|
1592 E T8l, T8n, T8q, T8s;
|
Chris@42
|
1593 T8l = W[42];
|
Chris@42
|
1594 T8n = W[43];
|
Chris@42
|
1595 T8p = FNMS(T8n, T8o, T8l * T8m);
|
Chris@42
|
1596 T8v = FMA(T8n, T8m, T8l * T8o);
|
Chris@42
|
1597 T8q = W[44];
|
Chris@42
|
1598 T8s = W[45];
|
Chris@42
|
1599 T8u = FMA(T8q, T8r, T8s * T8t);
|
Chris@42
|
1600 T8w = FNMS(T8s, T8r, T8q * T8t);
|
Chris@42
|
1601 }
|
Chris@42
|
1602 Rp[WS(rs, 11)] = T8p - T8u;
|
Chris@42
|
1603 Ip[WS(rs, 11)] = T8v + T8w;
|
Chris@42
|
1604 Rm[WS(rs, 11)] = T8p + T8u;
|
Chris@42
|
1605 Im[WS(rs, 11)] = T8w - T8v;
|
Chris@42
|
1606 }
|
Chris@42
|
1607 {
|
Chris@42
|
1608 E T8F, T8P, T8O, T8Q;
|
Chris@42
|
1609 {
|
Chris@42
|
1610 E T8x, T8B, T8G, T8K;
|
Chris@42
|
1611 T8x = W[26];
|
Chris@42
|
1612 T8B = W[27];
|
Chris@42
|
1613 T8F = FNMS(T8B, T8E, T8x * T8A);
|
Chris@42
|
1614 T8P = FMA(T8B, T8A, T8x * T8E);
|
Chris@42
|
1615 T8G = W[28];
|
Chris@42
|
1616 T8K = W[29];
|
Chris@42
|
1617 T8O = FMA(T8G, T8J, T8K * T8N);
|
Chris@42
|
1618 T8Q = FNMS(T8K, T8J, T8G * T8N);
|
Chris@42
|
1619 }
|
Chris@42
|
1620 Rp[WS(rs, 7)] = T8F - T8O;
|
Chris@42
|
1621 Ip[WS(rs, 7)] = T8P + T8Q;
|
Chris@42
|
1622 Rm[WS(rs, 7)] = T8F + T8O;
|
Chris@42
|
1623 Im[WS(rs, 7)] = T8Q - T8P;
|
Chris@42
|
1624 }
|
Chris@42
|
1625 }
|
Chris@42
|
1626 {
|
Chris@42
|
1627 E T4k, T4S, T4s, T4U, T5a, T5q, T56, T5o, T4N, T5t, T4Z, T5f, T4F, T5v, T4X;
|
Chris@42
|
1628 E T5j;
|
Chris@42
|
1629 {
|
Chris@42
|
1630 E T4c, T4j, T54, T55;
|
Chris@42
|
1631 T4c = T4a + T4b;
|
Chris@42
|
1632 T4j = KP707106781 * (T4f + T4i);
|
Chris@42
|
1633 T4k = T4c + T4j;
|
Chris@42
|
1634 T4S = T4c - T4j;
|
Chris@42
|
1635 {
|
Chris@42
|
1636 E T4o, T4r, T58, T59;
|
Chris@42
|
1637 T4o = KP707106781 * (T4m + T4n);
|
Chris@42
|
1638 T4r = T4p + T4q;
|
Chris@42
|
1639 T4s = T4o + T4r;
|
Chris@42
|
1640 T4U = T4r - T4o;
|
Chris@42
|
1641 T58 = KP707106781 * (T4f - T4i);
|
Chris@42
|
1642 T59 = T4q - T4p;
|
Chris@42
|
1643 T5a = T58 + T59;
|
Chris@42
|
1644 T5q = T59 - T58;
|
Chris@42
|
1645 }
|
Chris@42
|
1646 T54 = T4a - T4b;
|
Chris@42
|
1647 T55 = KP707106781 * (T4n - T4m);
|
Chris@42
|
1648 T56 = T54 + T55;
|
Chris@42
|
1649 T5o = T54 - T55;
|
Chris@42
|
1650 {
|
Chris@42
|
1651 E T4M, T5d, T4J, T5e, T4H, T4I;
|
Chris@42
|
1652 T4M = T4K + T4L;
|
Chris@42
|
1653 T5d = T4v - T4w;
|
Chris@42
|
1654 T4H = FNMS(KP831469612, T4y, KP555570233 * T4z);
|
Chris@42
|
1655 T4I = FMA(KP831469612, T4B, KP555570233 * T4C);
|
Chris@42
|
1656 T4J = T4H + T4I;
|
Chris@42
|
1657 T5e = T4H - T4I;
|
Chris@42
|
1658 T4N = T4J + T4M;
|
Chris@42
|
1659 T5t = T5d - T5e;
|
Chris@42
|
1660 T4Z = T4M - T4J;
|
Chris@42
|
1661 T5f = T5d + T5e;
|
Chris@42
|
1662 }
|
Chris@42
|
1663 {
|
Chris@42
|
1664 E T4x, T5i, T4E, T5h, T4A, T4D;
|
Chris@42
|
1665 T4x = T4v + T4w;
|
Chris@42
|
1666 T5i = T4L - T4K;
|
Chris@42
|
1667 T4A = FMA(KP555570233, T4y, KP831469612 * T4z);
|
Chris@42
|
1668 T4D = FNMS(KP831469612, T4C, KP555570233 * T4B);
|
Chris@42
|
1669 T4E = T4A + T4D;
|
Chris@42
|
1670 T5h = T4D - T4A;
|
Chris@42
|
1671 T4F = T4x + T4E;
|
Chris@42
|
1672 T5v = T5i - T5h;
|
Chris@42
|
1673 T4X = T4x - T4E;
|
Chris@42
|
1674 T5j = T5h + T5i;
|
Chris@42
|
1675 }
|
Chris@42
|
1676 }
|
Chris@42
|
1677 {
|
Chris@42
|
1678 E T4t, T4P, T4O, T4Q;
|
Chris@42
|
1679 {
|
Chris@42
|
1680 E T49, T4l, T4u, T4G;
|
Chris@42
|
1681 T49 = W[6];
|
Chris@42
|
1682 T4l = W[7];
|
Chris@42
|
1683 T4t = FNMS(T4l, T4s, T49 * T4k);
|
Chris@42
|
1684 T4P = FMA(T4l, T4k, T49 * T4s);
|
Chris@42
|
1685 T4u = W[8];
|
Chris@42
|
1686 T4G = W[9];
|
Chris@42
|
1687 T4O = FMA(T4u, T4F, T4G * T4N);
|
Chris@42
|
1688 T4Q = FNMS(T4G, T4F, T4u * T4N);
|
Chris@42
|
1689 }
|
Chris@42
|
1690 Rp[WS(rs, 2)] = T4t - T4O;
|
Chris@42
|
1691 Ip[WS(rs, 2)] = T4P + T4Q;
|
Chris@42
|
1692 Rm[WS(rs, 2)] = T4t + T4O;
|
Chris@42
|
1693 Im[WS(rs, 2)] = T4Q - T4P;
|
Chris@42
|
1694 }
|
Chris@42
|
1695 {
|
Chris@42
|
1696 E T5r, T5x, T5w, T5y;
|
Chris@42
|
1697 {
|
Chris@42
|
1698 E T5n, T5p, T5s, T5u;
|
Chris@42
|
1699 T5n = W[54];
|
Chris@42
|
1700 T5p = W[55];
|
Chris@42
|
1701 T5r = FNMS(T5p, T5q, T5n * T5o);
|
Chris@42
|
1702 T5x = FMA(T5p, T5o, T5n * T5q);
|
Chris@42
|
1703 T5s = W[56];
|
Chris@42
|
1704 T5u = W[57];
|
Chris@42
|
1705 T5w = FMA(T5s, T5t, T5u * T5v);
|
Chris@42
|
1706 T5y = FNMS(T5u, T5t, T5s * T5v);
|
Chris@42
|
1707 }
|
Chris@42
|
1708 Rp[WS(rs, 14)] = T5r - T5w;
|
Chris@42
|
1709 Ip[WS(rs, 14)] = T5x + T5y;
|
Chris@42
|
1710 Rm[WS(rs, 14)] = T5r + T5w;
|
Chris@42
|
1711 Im[WS(rs, 14)] = T5y - T5x;
|
Chris@42
|
1712 }
|
Chris@42
|
1713 {
|
Chris@42
|
1714 E T4V, T51, T50, T52;
|
Chris@42
|
1715 {
|
Chris@42
|
1716 E T4R, T4T, T4W, T4Y;
|
Chris@42
|
1717 T4R = W[38];
|
Chris@42
|
1718 T4T = W[39];
|
Chris@42
|
1719 T4V = FNMS(T4T, T4U, T4R * T4S);
|
Chris@42
|
1720 T51 = FMA(T4T, T4S, T4R * T4U);
|
Chris@42
|
1721 T4W = W[40];
|
Chris@42
|
1722 T4Y = W[41];
|
Chris@42
|
1723 T50 = FMA(T4W, T4X, T4Y * T4Z);
|
Chris@42
|
1724 T52 = FNMS(T4Y, T4X, T4W * T4Z);
|
Chris@42
|
1725 }
|
Chris@42
|
1726 Rp[WS(rs, 10)] = T4V - T50;
|
Chris@42
|
1727 Ip[WS(rs, 10)] = T51 + T52;
|
Chris@42
|
1728 Rm[WS(rs, 10)] = T4V + T50;
|
Chris@42
|
1729 Im[WS(rs, 10)] = T52 - T51;
|
Chris@42
|
1730 }
|
Chris@42
|
1731 {
|
Chris@42
|
1732 E T5b, T5l, T5k, T5m;
|
Chris@42
|
1733 {
|
Chris@42
|
1734 E T53, T57, T5c, T5g;
|
Chris@42
|
1735 T53 = W[22];
|
Chris@42
|
1736 T57 = W[23];
|
Chris@42
|
1737 T5b = FNMS(T57, T5a, T53 * T56);
|
Chris@42
|
1738 T5l = FMA(T57, T56, T53 * T5a);
|
Chris@42
|
1739 T5c = W[24];
|
Chris@42
|
1740 T5g = W[25];
|
Chris@42
|
1741 T5k = FMA(T5c, T5f, T5g * T5j);
|
Chris@42
|
1742 T5m = FNMS(T5g, T5f, T5c * T5j);
|
Chris@42
|
1743 }
|
Chris@42
|
1744 Rp[WS(rs, 6)] = T5b - T5k;
|
Chris@42
|
1745 Ip[WS(rs, 6)] = T5l + T5m;
|
Chris@42
|
1746 Rm[WS(rs, 6)] = T5b + T5k;
|
Chris@42
|
1747 Im[WS(rs, 6)] = T5m - T5l;
|
Chris@42
|
1748 }
|
Chris@42
|
1749 }
|
Chris@42
|
1750 {
|
Chris@42
|
1751 E T60, T6W, T6c, T6Y, T7e, T7u, T7a, T7s, T6R, T7x, T73, T7j, T6F, T7z, T71;
|
Chris@42
|
1752 E T7n;
|
Chris@42
|
1753 {
|
Chris@42
|
1754 E T5K, T5Z, T78, T79;
|
Chris@42
|
1755 T5K = T5C + T5J;
|
Chris@42
|
1756 T5Z = T5R + T5Y;
|
Chris@42
|
1757 T60 = T5K + T5Z;
|
Chris@42
|
1758 T6W = T5K - T5Z;
|
Chris@42
|
1759 {
|
Chris@42
|
1760 E T64, T6b, T7c, T7d;
|
Chris@42
|
1761 T64 = T62 + T63;
|
Chris@42
|
1762 T6b = T67 + T6a;
|
Chris@42
|
1763 T6c = T64 + T6b;
|
Chris@42
|
1764 T6Y = T6b - T64;
|
Chris@42
|
1765 T7c = T5R - T5Y;
|
Chris@42
|
1766 T7d = T6a - T67;
|
Chris@42
|
1767 T7e = T7c + T7d;
|
Chris@42
|
1768 T7u = T7d - T7c;
|
Chris@42
|
1769 }
|
Chris@42
|
1770 T78 = T5C - T5J;
|
Chris@42
|
1771 T79 = T63 - T62;
|
Chris@42
|
1772 T7a = T78 + T79;
|
Chris@42
|
1773 T7s = T78 - T79;
|
Chris@42
|
1774 {
|
Chris@42
|
1775 E T6Q, T7h, T6J, T7i, T6H, T6I;
|
Chris@42
|
1776 T6Q = T6M + T6P;
|
Chris@42
|
1777 T7h = T6h - T6o;
|
Chris@42
|
1778 T6H = FNMS(KP555570233, T6s, KP831469612 * T6v);
|
Chris@42
|
1779 T6I = FMA(KP555570233, T6z, KP831469612 * T6C);
|
Chris@42
|
1780 T6J = T6H + T6I;
|
Chris@42
|
1781 T7i = T6H - T6I;
|
Chris@42
|
1782 T6R = T6J + T6Q;
|
Chris@42
|
1783 T7x = T7h - T7i;
|
Chris@42
|
1784 T73 = T6Q - T6J;
|
Chris@42
|
1785 T7j = T7h + T7i;
|
Chris@42
|
1786 }
|
Chris@42
|
1787 {
|
Chris@42
|
1788 E T6p, T7m, T6E, T7l, T6w, T6D;
|
Chris@42
|
1789 T6p = T6h + T6o;
|
Chris@42
|
1790 T7m = T6P - T6M;
|
Chris@42
|
1791 T6w = FMA(KP831469612, T6s, KP555570233 * T6v);
|
Chris@42
|
1792 T6D = FNMS(KP555570233, T6C, KP831469612 * T6z);
|
Chris@42
|
1793 T6E = T6w + T6D;
|
Chris@42
|
1794 T7l = T6D - T6w;
|
Chris@42
|
1795 T6F = T6p + T6E;
|
Chris@42
|
1796 T7z = T7m - T7l;
|
Chris@42
|
1797 T71 = T6p - T6E;
|
Chris@42
|
1798 T7n = T7l + T7m;
|
Chris@42
|
1799 }
|
Chris@42
|
1800 }
|
Chris@42
|
1801 {
|
Chris@42
|
1802 E T6d, T6T, T6S, T6U;
|
Chris@42
|
1803 {
|
Chris@42
|
1804 E T5z, T61, T6e, T6G;
|
Chris@42
|
1805 T5z = W[2];
|
Chris@42
|
1806 T61 = W[3];
|
Chris@42
|
1807 T6d = FNMS(T61, T6c, T5z * T60);
|
Chris@42
|
1808 T6T = FMA(T61, T60, T5z * T6c);
|
Chris@42
|
1809 T6e = W[4];
|
Chris@42
|
1810 T6G = W[5];
|
Chris@42
|
1811 T6S = FMA(T6e, T6F, T6G * T6R);
|
Chris@42
|
1812 T6U = FNMS(T6G, T6F, T6e * T6R);
|
Chris@42
|
1813 }
|
Chris@42
|
1814 Rp[WS(rs, 1)] = T6d - T6S;
|
Chris@42
|
1815 Ip[WS(rs, 1)] = T6T + T6U;
|
Chris@42
|
1816 Rm[WS(rs, 1)] = T6d + T6S;
|
Chris@42
|
1817 Im[WS(rs, 1)] = T6U - T6T;
|
Chris@42
|
1818 }
|
Chris@42
|
1819 {
|
Chris@42
|
1820 E T7v, T7B, T7A, T7C;
|
Chris@42
|
1821 {
|
Chris@42
|
1822 E T7r, T7t, T7w, T7y;
|
Chris@42
|
1823 T7r = W[50];
|
Chris@42
|
1824 T7t = W[51];
|
Chris@42
|
1825 T7v = FNMS(T7t, T7u, T7r * T7s);
|
Chris@42
|
1826 T7B = FMA(T7t, T7s, T7r * T7u);
|
Chris@42
|
1827 T7w = W[52];
|
Chris@42
|
1828 T7y = W[53];
|
Chris@42
|
1829 T7A = FMA(T7w, T7x, T7y * T7z);
|
Chris@42
|
1830 T7C = FNMS(T7y, T7x, T7w * T7z);
|
Chris@42
|
1831 }
|
Chris@42
|
1832 Rp[WS(rs, 13)] = T7v - T7A;
|
Chris@42
|
1833 Ip[WS(rs, 13)] = T7B + T7C;
|
Chris@42
|
1834 Rm[WS(rs, 13)] = T7v + T7A;
|
Chris@42
|
1835 Im[WS(rs, 13)] = T7C - T7B;
|
Chris@42
|
1836 }
|
Chris@42
|
1837 {
|
Chris@42
|
1838 E T6Z, T75, T74, T76;
|
Chris@42
|
1839 {
|
Chris@42
|
1840 E T6V, T6X, T70, T72;
|
Chris@42
|
1841 T6V = W[34];
|
Chris@42
|
1842 T6X = W[35];
|
Chris@42
|
1843 T6Z = FNMS(T6X, T6Y, T6V * T6W);
|
Chris@42
|
1844 T75 = FMA(T6X, T6W, T6V * T6Y);
|
Chris@42
|
1845 T70 = W[36];
|
Chris@42
|
1846 T72 = W[37];
|
Chris@42
|
1847 T74 = FMA(T70, T71, T72 * T73);
|
Chris@42
|
1848 T76 = FNMS(T72, T71, T70 * T73);
|
Chris@42
|
1849 }
|
Chris@42
|
1850 Rp[WS(rs, 9)] = T6Z - T74;
|
Chris@42
|
1851 Ip[WS(rs, 9)] = T75 + T76;
|
Chris@42
|
1852 Rm[WS(rs, 9)] = T6Z + T74;
|
Chris@42
|
1853 Im[WS(rs, 9)] = T76 - T75;
|
Chris@42
|
1854 }
|
Chris@42
|
1855 {
|
Chris@42
|
1856 E T7f, T7p, T7o, T7q;
|
Chris@42
|
1857 {
|
Chris@42
|
1858 E T77, T7b, T7g, T7k;
|
Chris@42
|
1859 T77 = W[18];
|
Chris@42
|
1860 T7b = W[19];
|
Chris@42
|
1861 T7f = FNMS(T7b, T7e, T77 * T7a);
|
Chris@42
|
1862 T7p = FMA(T7b, T7a, T77 * T7e);
|
Chris@42
|
1863 T7g = W[20];
|
Chris@42
|
1864 T7k = W[21];
|
Chris@42
|
1865 T7o = FMA(T7g, T7j, T7k * T7n);
|
Chris@42
|
1866 T7q = FNMS(T7k, T7j, T7g * T7n);
|
Chris@42
|
1867 }
|
Chris@42
|
1868 Rp[WS(rs, 5)] = T7f - T7o;
|
Chris@42
|
1869 Ip[WS(rs, 5)] = T7p + T7q;
|
Chris@42
|
1870 Rm[WS(rs, 5)] = T7f + T7o;
|
Chris@42
|
1871 Im[WS(rs, 5)] = T7q - T7p;
|
Chris@42
|
1872 }
|
Chris@42
|
1873 }
|
Chris@42
|
1874 }
|
Chris@42
|
1875 }
|
Chris@42
|
1876 }
|
Chris@42
|
1877
|
Chris@42
|
1878 static const tw_instr twinstr[] = {
|
Chris@42
|
1879 {TW_FULL, 1, 32},
|
Chris@42
|
1880 {TW_NEXT, 1, 0}
|
Chris@42
|
1881 };
|
Chris@42
|
1882
|
Chris@42
|
1883 static const hc2c_desc desc = { 32, "hc2cbdft_32", twinstr, &GENUS, {404, 114, 94, 0} };
|
Chris@42
|
1884
|
Chris@42
|
1885 void X(codelet_hc2cbdft_32) (planner *p) {
|
Chris@42
|
1886 X(khc2c_register) (p, hc2cbdft_32, &desc, HC2C_VIA_DFT);
|
Chris@42
|
1887 }
|
Chris@42
|
1888 #endif /* HAVE_FMA */
|