Chris@10
|
1 /*
|
Chris@10
|
2 * Copyright (c) 2003, 2007-11 Matteo Frigo
|
Chris@10
|
3 * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology
|
Chris@10
|
4 *
|
Chris@10
|
5 * This program is free software; you can redistribute it and/or modify
|
Chris@10
|
6 * it under the terms of the GNU General Public License as published by
|
Chris@10
|
7 * the Free Software Foundation; either version 2 of the License, or
|
Chris@10
|
8 * (at your option) any later version.
|
Chris@10
|
9 *
|
Chris@10
|
10 * This program is distributed in the hope that it will be useful,
|
Chris@10
|
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
Chris@10
|
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
Chris@10
|
13 * GNU General Public License for more details.
|
Chris@10
|
14 *
|
Chris@10
|
15 * You should have received a copy of the GNU General Public License
|
Chris@10
|
16 * along with this program; if not, write to the Free Software
|
Chris@10
|
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
Chris@10
|
18 *
|
Chris@10
|
19 */
|
Chris@10
|
20
|
Chris@10
|
21 /* This file was automatically generated --- DO NOT EDIT */
|
Chris@10
|
22 /* Generated on Sun Nov 25 07:35:45 EST 2012 */
|
Chris@10
|
23
|
Chris@10
|
24 #include "codelet-dft.h"
|
Chris@10
|
25
|
Chris@10
|
26 #ifdef HAVE_FMA
|
Chris@10
|
27
|
Chris@10
|
28 /* Generated by: ../../../genfft/gen_notw.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -n 32 -name n1_32 -include n.h */
|
Chris@10
|
29
|
Chris@10
|
30 /*
|
Chris@10
|
31 * This function contains 372 FP additions, 136 FP multiplications,
|
Chris@10
|
32 * (or, 236 additions, 0 multiplications, 136 fused multiply/add),
|
Chris@10
|
33 * 136 stack variables, 7 constants, and 128 memory accesses
|
Chris@10
|
34 */
|
Chris@10
|
35 #include "n.h"
|
Chris@10
|
36
|
Chris@10
|
37 static void n1_32(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
|
Chris@10
|
38 {
|
Chris@10
|
39 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
Chris@10
|
40 DK(KP198912367, +0.198912367379658006911597622644676228597850501);
|
Chris@10
|
41 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
Chris@10
|
42 DK(KP668178637, +0.668178637919298919997757686523080761552472251);
|
Chris@10
|
43 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
Chris@10
|
44 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
Chris@10
|
45 DK(KP414213562, +0.414213562373095048801688724209698078569671875);
|
Chris@10
|
46 {
|
Chris@10
|
47 INT i;
|
Chris@10
|
48 for (i = v; i > 0; i = i - 1, ri = ri + ivs, ii = ii + ivs, ro = ro + ovs, io = io + ovs, MAKE_VOLATILE_STRIDE(128, is), MAKE_VOLATILE_STRIDE(128, os)) {
|
Chris@10
|
49 E T3g, T3f, T3n, T3b, T3r, T3l, T3o, T3e, T3h, T3p;
|
Chris@10
|
50 {
|
Chris@10
|
51 E T2T, T3T, T4r, T7, T3t, T1z, T18, T4Z, Te, T50, T4s, T1f, T2W, T3u, T3U;
|
Chris@10
|
52 E T1G, Tm, T1n, T3X, T3y, T2Z, T1O, T53, T4w, Tt, T1u, T3W, T3B, T2Y, T1V;
|
Chris@10
|
53 E T52, T4z, T3O, T2t, T3L, T2K, T5F, TZ, T5I, T5X, T4R, T5k, T3M, T2E, T5j;
|
Chris@10
|
54 E T4W, T3P, T2N, T3H, T22, T3E, T2j, T4H, T4K, T5A, TK, T5D, T5W, T2k, T2l;
|
Chris@10
|
55 E T4G, T5h, T3F, T2d;
|
Chris@10
|
56 {
|
Chris@10
|
57 E Tj, T1L, Ti, T1I, T1j, Tk, T1k, T1l;
|
Chris@10
|
58 {
|
Chris@10
|
59 E T4, T1x, T3, T2R, T14, T5, T15, T16, T1C, T1F;
|
Chris@10
|
60 {
|
Chris@10
|
61 E T1, T2, T12, T13;
|
Chris@10
|
62 T1 = ri[0];
|
Chris@10
|
63 T2 = ri[WS(is, 16)];
|
Chris@10
|
64 T12 = ii[0];
|
Chris@10
|
65 T13 = ii[WS(is, 16)];
|
Chris@10
|
66 T4 = ri[WS(is, 8)];
|
Chris@10
|
67 T1x = T1 - T2;
|
Chris@10
|
68 T3 = T1 + T2;
|
Chris@10
|
69 T2R = T12 - T13;
|
Chris@10
|
70 T14 = T12 + T13;
|
Chris@10
|
71 T5 = ri[WS(is, 24)];
|
Chris@10
|
72 T15 = ii[WS(is, 8)];
|
Chris@10
|
73 T16 = ii[WS(is, 24)];
|
Chris@10
|
74 }
|
Chris@10
|
75 {
|
Chris@10
|
76 E Tb, T1A, Ta, T1B, T1b, Tc, T1c, T1d;
|
Chris@10
|
77 {
|
Chris@10
|
78 E T8, T9, T19, T1a;
|
Chris@10
|
79 T8 = ri[WS(is, 4)];
|
Chris@10
|
80 {
|
Chris@10
|
81 E T2S, T6, T1y, T17;
|
Chris@10
|
82 T2S = T4 - T5;
|
Chris@10
|
83 T6 = T4 + T5;
|
Chris@10
|
84 T1y = T15 - T16;
|
Chris@10
|
85 T17 = T15 + T16;
|
Chris@10
|
86 T2T = T2R - T2S;
|
Chris@10
|
87 T3T = T2S + T2R;
|
Chris@10
|
88 T4r = T3 - T6;
|
Chris@10
|
89 T7 = T3 + T6;
|
Chris@10
|
90 T3t = T1x - T1y;
|
Chris@10
|
91 T1z = T1x + T1y;
|
Chris@10
|
92 T18 = T14 + T17;
|
Chris@10
|
93 T4Z = T14 - T17;
|
Chris@10
|
94 T9 = ri[WS(is, 20)];
|
Chris@10
|
95 }
|
Chris@10
|
96 T19 = ii[WS(is, 4)];
|
Chris@10
|
97 T1a = ii[WS(is, 20)];
|
Chris@10
|
98 Tb = ri[WS(is, 28)];
|
Chris@10
|
99 T1A = T8 - T9;
|
Chris@10
|
100 Ta = T8 + T9;
|
Chris@10
|
101 T1B = T19 - T1a;
|
Chris@10
|
102 T1b = T19 + T1a;
|
Chris@10
|
103 Tc = ri[WS(is, 12)];
|
Chris@10
|
104 T1c = ii[WS(is, 28)];
|
Chris@10
|
105 T1d = ii[WS(is, 12)];
|
Chris@10
|
106 }
|
Chris@10
|
107 {
|
Chris@10
|
108 E T2U, T1D, Td, T1E, T1e, T2V;
|
Chris@10
|
109 T1C = T1A + T1B;
|
Chris@10
|
110 T2U = T1B - T1A;
|
Chris@10
|
111 T1D = Tb - Tc;
|
Chris@10
|
112 Td = Tb + Tc;
|
Chris@10
|
113 T1E = T1c - T1d;
|
Chris@10
|
114 T1e = T1c + T1d;
|
Chris@10
|
115 Te = Ta + Td;
|
Chris@10
|
116 T50 = Td - Ta;
|
Chris@10
|
117 T1F = T1D - T1E;
|
Chris@10
|
118 T2V = T1D + T1E;
|
Chris@10
|
119 T4s = T1b - T1e;
|
Chris@10
|
120 T1f = T1b + T1e;
|
Chris@10
|
121 T2W = T2U + T2V;
|
Chris@10
|
122 T3u = T2U - T2V;
|
Chris@10
|
123 }
|
Chris@10
|
124 }
|
Chris@10
|
125 {
|
Chris@10
|
126 E Tg, Th, T1h, T1i;
|
Chris@10
|
127 Tg = ri[WS(is, 2)];
|
Chris@10
|
128 T3U = T1F - T1C;
|
Chris@10
|
129 T1G = T1C + T1F;
|
Chris@10
|
130 Th = ri[WS(is, 18)];
|
Chris@10
|
131 T1h = ii[WS(is, 2)];
|
Chris@10
|
132 T1i = ii[WS(is, 18)];
|
Chris@10
|
133 Tj = ri[WS(is, 10)];
|
Chris@10
|
134 T1L = Tg - Th;
|
Chris@10
|
135 Ti = Tg + Th;
|
Chris@10
|
136 T1I = T1h - T1i;
|
Chris@10
|
137 T1j = T1h + T1i;
|
Chris@10
|
138 Tk = ri[WS(is, 26)];
|
Chris@10
|
139 T1k = ii[WS(is, 10)];
|
Chris@10
|
140 T1l = ii[WS(is, 26)];
|
Chris@10
|
141 }
|
Chris@10
|
142 }
|
Chris@10
|
143 {
|
Chris@10
|
144 E Tq, T1S, Tp, T1P, T1q, Tr, T1r, T1s;
|
Chris@10
|
145 {
|
Chris@10
|
146 E Tn, To, T1o, T1p, T1J, Tl;
|
Chris@10
|
147 Tn = ri[WS(is, 30)];
|
Chris@10
|
148 T1J = Tj - Tk;
|
Chris@10
|
149 Tl = Tj + Tk;
|
Chris@10
|
150 {
|
Chris@10
|
151 E T1M, T1m, T3w, T1K;
|
Chris@10
|
152 T1M = T1k - T1l;
|
Chris@10
|
153 T1m = T1k + T1l;
|
Chris@10
|
154 T3w = T1J + T1I;
|
Chris@10
|
155 T1K = T1I - T1J;
|
Chris@10
|
156 {
|
Chris@10
|
157 E T4v, T3x, T1N, T4u;
|
Chris@10
|
158 T4v = Ti - Tl;
|
Chris@10
|
159 Tm = Ti + Tl;
|
Chris@10
|
160 T3x = T1L - T1M;
|
Chris@10
|
161 T1N = T1L + T1M;
|
Chris@10
|
162 T4u = T1j - T1m;
|
Chris@10
|
163 T1n = T1j + T1m;
|
Chris@10
|
164 T3X = FNMS(KP414213562, T3w, T3x);
|
Chris@10
|
165 T3y = FMA(KP414213562, T3x, T3w);
|
Chris@10
|
166 T2Z = FMA(KP414213562, T1K, T1N);
|
Chris@10
|
167 T1O = FNMS(KP414213562, T1N, T1K);
|
Chris@10
|
168 T53 = T4v + T4u;
|
Chris@10
|
169 T4w = T4u - T4v;
|
Chris@10
|
170 To = ri[WS(is, 14)];
|
Chris@10
|
171 }
|
Chris@10
|
172 }
|
Chris@10
|
173 T1o = ii[WS(is, 30)];
|
Chris@10
|
174 T1p = ii[WS(is, 14)];
|
Chris@10
|
175 Tq = ri[WS(is, 6)];
|
Chris@10
|
176 T1S = Tn - To;
|
Chris@10
|
177 Tp = Tn + To;
|
Chris@10
|
178 T1P = T1o - T1p;
|
Chris@10
|
179 T1q = T1o + T1p;
|
Chris@10
|
180 Tr = ri[WS(is, 22)];
|
Chris@10
|
181 T1r = ii[WS(is, 6)];
|
Chris@10
|
182 T1s = ii[WS(is, 22)];
|
Chris@10
|
183 }
|
Chris@10
|
184 {
|
Chris@10
|
185 E T4S, T4V, T2L, T2M;
|
Chris@10
|
186 {
|
Chris@10
|
187 E T2G, TN, T4N, T2r, T2s, TQ, T4O, T2J, TV, T2x, TU, T4T, T2w, TW, T2A;
|
Chris@10
|
188 E T2B;
|
Chris@10
|
189 {
|
Chris@10
|
190 E TO, TP, T2H, T2I;
|
Chris@10
|
191 {
|
Chris@10
|
192 E TL, TM, T2p, T2q, T1Q, Ts;
|
Chris@10
|
193 TL = ri[WS(is, 31)];
|
Chris@10
|
194 T1Q = Tq - Tr;
|
Chris@10
|
195 Ts = Tq + Tr;
|
Chris@10
|
196 {
|
Chris@10
|
197 E T1T, T1t, T3z, T1R;
|
Chris@10
|
198 T1T = T1r - T1s;
|
Chris@10
|
199 T1t = T1r + T1s;
|
Chris@10
|
200 T3z = T1Q + T1P;
|
Chris@10
|
201 T1R = T1P - T1Q;
|
Chris@10
|
202 {
|
Chris@10
|
203 E T4x, T3A, T1U, T4y;
|
Chris@10
|
204 T4x = Tp - Ts;
|
Chris@10
|
205 Tt = Tp + Ts;
|
Chris@10
|
206 T3A = T1S - T1T;
|
Chris@10
|
207 T1U = T1S + T1T;
|
Chris@10
|
208 T4y = T1q - T1t;
|
Chris@10
|
209 T1u = T1q + T1t;
|
Chris@10
|
210 T3W = FMA(KP414213562, T3z, T3A);
|
Chris@10
|
211 T3B = FNMS(KP414213562, T3A, T3z);
|
Chris@10
|
212 T2Y = FNMS(KP414213562, T1R, T1U);
|
Chris@10
|
213 T1V = FMA(KP414213562, T1U, T1R);
|
Chris@10
|
214 T52 = T4x - T4y;
|
Chris@10
|
215 T4z = T4x + T4y;
|
Chris@10
|
216 TM = ri[WS(is, 15)];
|
Chris@10
|
217 }
|
Chris@10
|
218 }
|
Chris@10
|
219 T2p = ii[WS(is, 31)];
|
Chris@10
|
220 T2q = ii[WS(is, 15)];
|
Chris@10
|
221 TO = ri[WS(is, 7)];
|
Chris@10
|
222 T2G = TL - TM;
|
Chris@10
|
223 TN = TL + TM;
|
Chris@10
|
224 T4N = T2p + T2q;
|
Chris@10
|
225 T2r = T2p - T2q;
|
Chris@10
|
226 TP = ri[WS(is, 23)];
|
Chris@10
|
227 T2H = ii[WS(is, 7)];
|
Chris@10
|
228 T2I = ii[WS(is, 23)];
|
Chris@10
|
229 }
|
Chris@10
|
230 {
|
Chris@10
|
231 E TS, TT, T2u, T2v;
|
Chris@10
|
232 TS = ri[WS(is, 3)];
|
Chris@10
|
233 T2s = TO - TP;
|
Chris@10
|
234 TQ = TO + TP;
|
Chris@10
|
235 T4O = T2H + T2I;
|
Chris@10
|
236 T2J = T2H - T2I;
|
Chris@10
|
237 TT = ri[WS(is, 19)];
|
Chris@10
|
238 T2u = ii[WS(is, 3)];
|
Chris@10
|
239 T2v = ii[WS(is, 19)];
|
Chris@10
|
240 TV = ri[WS(is, 27)];
|
Chris@10
|
241 T2x = TS - TT;
|
Chris@10
|
242 TU = TS + TT;
|
Chris@10
|
243 T4T = T2u + T2v;
|
Chris@10
|
244 T2w = T2u - T2v;
|
Chris@10
|
245 TW = ri[WS(is, 11)];
|
Chris@10
|
246 T2A = ii[WS(is, 27)];
|
Chris@10
|
247 T2B = ii[WS(is, 11)];
|
Chris@10
|
248 }
|
Chris@10
|
249 }
|
Chris@10
|
250 {
|
Chris@10
|
251 E T2z, T4U, T2C, TR, TY, T4Q, TX;
|
Chris@10
|
252 T3O = T2s + T2r;
|
Chris@10
|
253 T2t = T2r - T2s;
|
Chris@10
|
254 T2z = TV - TW;
|
Chris@10
|
255 TX = TV + TW;
|
Chris@10
|
256 T4U = T2A + T2B;
|
Chris@10
|
257 T2C = T2A - T2B;
|
Chris@10
|
258 T3L = T2G - T2J;
|
Chris@10
|
259 T2K = T2G + T2J;
|
Chris@10
|
260 T4S = TN - TQ;
|
Chris@10
|
261 TR = TN + TQ;
|
Chris@10
|
262 TY = TU + TX;
|
Chris@10
|
263 T4Q = TX - TU;
|
Chris@10
|
264 {
|
Chris@10
|
265 E T4P, T5G, T5H, T2y, T2D;
|
Chris@10
|
266 T4P = T4N - T4O;
|
Chris@10
|
267 T5G = T4N + T4O;
|
Chris@10
|
268 T5H = T4T + T4U;
|
Chris@10
|
269 T4V = T4T - T4U;
|
Chris@10
|
270 T5F = TR - TY;
|
Chris@10
|
271 TZ = TR + TY;
|
Chris@10
|
272 T5I = T5G - T5H;
|
Chris@10
|
273 T5X = T5G + T5H;
|
Chris@10
|
274 T2L = T2x + T2w;
|
Chris@10
|
275 T2y = T2w - T2x;
|
Chris@10
|
276 T2D = T2z + T2C;
|
Chris@10
|
277 T2M = T2z - T2C;
|
Chris@10
|
278 T4R = T4P - T4Q;
|
Chris@10
|
279 T5k = T4Q + T4P;
|
Chris@10
|
280 T3M = T2D - T2y;
|
Chris@10
|
281 T2E = T2y + T2D;
|
Chris@10
|
282 }
|
Chris@10
|
283 }
|
Chris@10
|
284 }
|
Chris@10
|
285 {
|
Chris@10
|
286 E T2f, Ty, T4C, T20, T21, TB, T4D, T2i, TG, T26, TF, T4I, T25, TH, T29;
|
Chris@10
|
287 E T2a;
|
Chris@10
|
288 {
|
Chris@10
|
289 E Tz, TA, T2g, T2h;
|
Chris@10
|
290 {
|
Chris@10
|
291 E Tw, Tx, T1Y, T1Z;
|
Chris@10
|
292 Tw = ri[WS(is, 1)];
|
Chris@10
|
293 T5j = T4S + T4V;
|
Chris@10
|
294 T4W = T4S - T4V;
|
Chris@10
|
295 T3P = T2L - T2M;
|
Chris@10
|
296 T2N = T2L + T2M;
|
Chris@10
|
297 Tx = ri[WS(is, 17)];
|
Chris@10
|
298 T1Y = ii[WS(is, 1)];
|
Chris@10
|
299 T1Z = ii[WS(is, 17)];
|
Chris@10
|
300 Tz = ri[WS(is, 9)];
|
Chris@10
|
301 T2f = Tw - Tx;
|
Chris@10
|
302 Ty = Tw + Tx;
|
Chris@10
|
303 T4C = T1Y + T1Z;
|
Chris@10
|
304 T20 = T1Y - T1Z;
|
Chris@10
|
305 TA = ri[WS(is, 25)];
|
Chris@10
|
306 T2g = ii[WS(is, 9)];
|
Chris@10
|
307 T2h = ii[WS(is, 25)];
|
Chris@10
|
308 }
|
Chris@10
|
309 {
|
Chris@10
|
310 E TD, TE, T23, T24;
|
Chris@10
|
311 TD = ri[WS(is, 5)];
|
Chris@10
|
312 T21 = Tz - TA;
|
Chris@10
|
313 TB = Tz + TA;
|
Chris@10
|
314 T4D = T2g + T2h;
|
Chris@10
|
315 T2i = T2g - T2h;
|
Chris@10
|
316 TE = ri[WS(is, 21)];
|
Chris@10
|
317 T23 = ii[WS(is, 5)];
|
Chris@10
|
318 T24 = ii[WS(is, 21)];
|
Chris@10
|
319 TG = ri[WS(is, 29)];
|
Chris@10
|
320 T26 = TD - TE;
|
Chris@10
|
321 TF = TD + TE;
|
Chris@10
|
322 T4I = T23 + T24;
|
Chris@10
|
323 T25 = T23 - T24;
|
Chris@10
|
324 TH = ri[WS(is, 13)];
|
Chris@10
|
325 T29 = ii[WS(is, 29)];
|
Chris@10
|
326 T2a = ii[WS(is, 13)];
|
Chris@10
|
327 }
|
Chris@10
|
328 }
|
Chris@10
|
329 {
|
Chris@10
|
330 E T28, T4J, T2b, TC, TJ, T4F, TI;
|
Chris@10
|
331 T3H = T21 + T20;
|
Chris@10
|
332 T22 = T20 - T21;
|
Chris@10
|
333 T28 = TG - TH;
|
Chris@10
|
334 TI = TG + TH;
|
Chris@10
|
335 T4J = T29 + T2a;
|
Chris@10
|
336 T2b = T29 - T2a;
|
Chris@10
|
337 T3E = T2f - T2i;
|
Chris@10
|
338 T2j = T2f + T2i;
|
Chris@10
|
339 T4H = Ty - TB;
|
Chris@10
|
340 TC = Ty + TB;
|
Chris@10
|
341 TJ = TF + TI;
|
Chris@10
|
342 T4F = TI - TF;
|
Chris@10
|
343 {
|
Chris@10
|
344 E T4E, T5B, T5C, T27, T2c;
|
Chris@10
|
345 T4E = T4C - T4D;
|
Chris@10
|
346 T5B = T4C + T4D;
|
Chris@10
|
347 T5C = T4I + T4J;
|
Chris@10
|
348 T4K = T4I - T4J;
|
Chris@10
|
349 T5A = TC - TJ;
|
Chris@10
|
350 TK = TC + TJ;
|
Chris@10
|
351 T5D = T5B - T5C;
|
Chris@10
|
352 T5W = T5B + T5C;
|
Chris@10
|
353 T2k = T26 + T25;
|
Chris@10
|
354 T27 = T25 - T26;
|
Chris@10
|
355 T2c = T28 + T2b;
|
Chris@10
|
356 T2l = T28 - T2b;
|
Chris@10
|
357 T4G = T4E - T4F;
|
Chris@10
|
358 T5h = T4F + T4E;
|
Chris@10
|
359 T3F = T2c - T27;
|
Chris@10
|
360 T2d = T27 + T2c;
|
Chris@10
|
361 }
|
Chris@10
|
362 }
|
Chris@10
|
363 }
|
Chris@10
|
364 }
|
Chris@10
|
365 }
|
Chris@10
|
366 }
|
Chris@10
|
367 {
|
Chris@10
|
368 E T3I, T2m, Tv, T60, T11, T10, T5Z, T1w;
|
Chris@10
|
369 {
|
Chris@10
|
370 E T5f, T5w, T5q, T5m, T5v, T5p;
|
Chris@10
|
371 {
|
Chris@10
|
372 E T5d, T5g, T5o, T4B, T5a, T5n, T5e, T56, T4Y, T57, T55;
|
Chris@10
|
373 {
|
Chris@10
|
374 E T4X, T4M, T5b, T5c, T51, T54;
|
Chris@10
|
375 {
|
Chris@10
|
376 E T4t, T4A, T58, T59, T4L;
|
Chris@10
|
377 T5d = T4r + T4s;
|
Chris@10
|
378 T4t = T4r - T4s;
|
Chris@10
|
379 T5g = T4H + T4K;
|
Chris@10
|
380 T4L = T4H - T4K;
|
Chris@10
|
381 T3I = T2k - T2l;
|
Chris@10
|
382 T2m = T2k + T2l;
|
Chris@10
|
383 T4A = T4w - T4z;
|
Chris@10
|
384 T5o = T4w + T4z;
|
Chris@10
|
385 T4X = FNMS(KP414213562, T4W, T4R);
|
Chris@10
|
386 T58 = FMA(KP414213562, T4R, T4W);
|
Chris@10
|
387 T59 = FNMS(KP414213562, T4G, T4L);
|
Chris@10
|
388 T4M = FMA(KP414213562, T4L, T4G);
|
Chris@10
|
389 T5b = FNMS(KP707106781, T4A, T4t);
|
Chris@10
|
390 T4B = FMA(KP707106781, T4A, T4t);
|
Chris@10
|
391 T5c = T59 + T58;
|
Chris@10
|
392 T5a = T58 - T59;
|
Chris@10
|
393 T5n = T50 + T4Z;
|
Chris@10
|
394 T51 = T4Z - T50;
|
Chris@10
|
395 T54 = T52 - T53;
|
Chris@10
|
396 T5e = T53 + T52;
|
Chris@10
|
397 }
|
Chris@10
|
398 ro[WS(os, 14)] = FNMS(KP923879532, T5c, T5b);
|
Chris@10
|
399 T56 = T4M + T4X;
|
Chris@10
|
400 T4Y = T4M - T4X;
|
Chris@10
|
401 T57 = FMA(KP707106781, T54, T51);
|
Chris@10
|
402 T55 = FNMS(KP707106781, T54, T51);
|
Chris@10
|
403 ro[WS(os, 30)] = FMA(KP923879532, T5c, T5b);
|
Chris@10
|
404 }
|
Chris@10
|
405 ro[WS(os, 6)] = FMA(KP923879532, T4Y, T4B);
|
Chris@10
|
406 ro[WS(os, 22)] = FNMS(KP923879532, T4Y, T4B);
|
Chris@10
|
407 io[WS(os, 6)] = FMA(KP923879532, T5a, T57);
|
Chris@10
|
408 io[WS(os, 22)] = FNMS(KP923879532, T5a, T57);
|
Chris@10
|
409 io[WS(os, 30)] = FMA(KP923879532, T56, T55);
|
Chris@10
|
410 io[WS(os, 14)] = FNMS(KP923879532, T56, T55);
|
Chris@10
|
411 {
|
Chris@10
|
412 E T5i, T5l, T5r, T5u, T5s, T5t;
|
Chris@10
|
413 T5i = FMA(KP414213562, T5h, T5g);
|
Chris@10
|
414 T5s = FNMS(KP414213562, T5g, T5h);
|
Chris@10
|
415 T5t = FMA(KP414213562, T5j, T5k);
|
Chris@10
|
416 T5l = FNMS(KP414213562, T5k, T5j);
|
Chris@10
|
417 T5r = FNMS(KP707106781, T5e, T5d);
|
Chris@10
|
418 T5f = FMA(KP707106781, T5e, T5d);
|
Chris@10
|
419 T5w = T5s + T5t;
|
Chris@10
|
420 T5u = T5s - T5t;
|
Chris@10
|
421 ro[WS(os, 26)] = FNMS(KP923879532, T5u, T5r);
|
Chris@10
|
422 T5q = T5l - T5i;
|
Chris@10
|
423 T5m = T5i + T5l;
|
Chris@10
|
424 T5v = FMA(KP707106781, T5o, T5n);
|
Chris@10
|
425 T5p = FNMS(KP707106781, T5o, T5n);
|
Chris@10
|
426 ro[WS(os, 10)] = FMA(KP923879532, T5u, T5r);
|
Chris@10
|
427 }
|
Chris@10
|
428 }
|
Chris@10
|
429 ro[WS(os, 2)] = FMA(KP923879532, T5m, T5f);
|
Chris@10
|
430 ro[WS(os, 18)] = FNMS(KP923879532, T5m, T5f);
|
Chris@10
|
431 io[WS(os, 2)] = FMA(KP923879532, T5w, T5v);
|
Chris@10
|
432 io[WS(os, 18)] = FNMS(KP923879532, T5w, T5v);
|
Chris@10
|
433 io[WS(os, 10)] = FMA(KP923879532, T5q, T5p);
|
Chris@10
|
434 io[WS(os, 26)] = FNMS(KP923879532, T5q, T5p);
|
Chris@10
|
435 {
|
Chris@10
|
436 E Tf, T1v, T5z, T5U, T1g, Tu, T5O, T5K, T5T, T5N, T5V, T5Y;
|
Chris@10
|
437 {
|
Chris@10
|
438 E T5E, T5J, T5P, T5S, T5L, T5M;
|
Chris@10
|
439 {
|
Chris@10
|
440 E T5x, T5y, T5Q, T5R;
|
Chris@10
|
441 Tf = T7 + Te;
|
Chris@10
|
442 T5x = T7 - Te;
|
Chris@10
|
443 T5y = T1n - T1u;
|
Chris@10
|
444 T1v = T1n + T1u;
|
Chris@10
|
445 T5E = T5A + T5D;
|
Chris@10
|
446 T5Q = T5D - T5A;
|
Chris@10
|
447 T5R = T5F + T5I;
|
Chris@10
|
448 T5J = T5F - T5I;
|
Chris@10
|
449 T5P = T5x - T5y;
|
Chris@10
|
450 T5z = T5x + T5y;
|
Chris@10
|
451 T5U = T5Q + T5R;
|
Chris@10
|
452 T5S = T5Q - T5R;
|
Chris@10
|
453 T1g = T18 + T1f;
|
Chris@10
|
454 T5L = T18 - T1f;
|
Chris@10
|
455 T5M = Tt - Tm;
|
Chris@10
|
456 Tu = Tm + Tt;
|
Chris@10
|
457 }
|
Chris@10
|
458 ro[WS(os, 28)] = FNMS(KP707106781, T5S, T5P);
|
Chris@10
|
459 T5O = T5J - T5E;
|
Chris@10
|
460 T5K = T5E + T5J;
|
Chris@10
|
461 T5T = T5M + T5L;
|
Chris@10
|
462 T5N = T5L - T5M;
|
Chris@10
|
463 ro[WS(os, 12)] = FMA(KP707106781, T5S, T5P);
|
Chris@10
|
464 }
|
Chris@10
|
465 ro[WS(os, 4)] = FMA(KP707106781, T5K, T5z);
|
Chris@10
|
466 ro[WS(os, 20)] = FNMS(KP707106781, T5K, T5z);
|
Chris@10
|
467 io[WS(os, 4)] = FMA(KP707106781, T5U, T5T);
|
Chris@10
|
468 io[WS(os, 20)] = FNMS(KP707106781, T5U, T5T);
|
Chris@10
|
469 io[WS(os, 12)] = FMA(KP707106781, T5O, T5N);
|
Chris@10
|
470 io[WS(os, 28)] = FNMS(KP707106781, T5O, T5N);
|
Chris@10
|
471 T5V = Tf - Tu;
|
Chris@10
|
472 Tv = Tf + Tu;
|
Chris@10
|
473 T60 = T5W + T5X;
|
Chris@10
|
474 T5Y = T5W - T5X;
|
Chris@10
|
475 ro[WS(os, 8)] = T5V + T5Y;
|
Chris@10
|
476 T11 = TZ - TK;
|
Chris@10
|
477 T10 = TK + TZ;
|
Chris@10
|
478 T5Z = T1g + T1v;
|
Chris@10
|
479 T1w = T1g - T1v;
|
Chris@10
|
480 ro[WS(os, 24)] = T5V - T5Y;
|
Chris@10
|
481 }
|
Chris@10
|
482 }
|
Chris@10
|
483 ro[0] = Tv + T10;
|
Chris@10
|
484 ro[WS(os, 16)] = Tv - T10;
|
Chris@10
|
485 io[0] = T5Z + T60;
|
Chris@10
|
486 io[WS(os, 16)] = T5Z - T60;
|
Chris@10
|
487 io[WS(os, 24)] = T1w - T11;
|
Chris@10
|
488 io[WS(os, 8)] = T11 + T1w;
|
Chris@10
|
489 {
|
Chris@10
|
490 E T39, T3k, T3j, T3a, T3d, T3c, T47, T4i, T4h, T41, T3D, T48, T4b, T4a, T4e;
|
Chris@10
|
491 E T3N, T45, T3Z, T42, T3K, T3Q, T4d;
|
Chris@10
|
492 {
|
Chris@10
|
493 E T2e, T37, T1X, T33, T31, T2n, T2F, T2O;
|
Chris@10
|
494 {
|
Chris@10
|
495 E T1H, T1W, T2X, T30;
|
Chris@10
|
496 T39 = FMA(KP707106781, T1G, T1z);
|
Chris@10
|
497 T1H = FNMS(KP707106781, T1G, T1z);
|
Chris@10
|
498 T1W = T1O - T1V;
|
Chris@10
|
499 T3k = T1O + T1V;
|
Chris@10
|
500 T3j = FMA(KP707106781, T2W, T2T);
|
Chris@10
|
501 T2X = FNMS(KP707106781, T2W, T2T);
|
Chris@10
|
502 T30 = T2Y - T2Z;
|
Chris@10
|
503 T3a = T2Z + T2Y;
|
Chris@10
|
504 T3d = FMA(KP707106781, T2d, T22);
|
Chris@10
|
505 T2e = FNMS(KP707106781, T2d, T22);
|
Chris@10
|
506 T37 = FNMS(KP923879532, T1W, T1H);
|
Chris@10
|
507 T1X = FMA(KP923879532, T1W, T1H);
|
Chris@10
|
508 T33 = FMA(KP923879532, T30, T2X);
|
Chris@10
|
509 T31 = FNMS(KP923879532, T30, T2X);
|
Chris@10
|
510 T2n = FNMS(KP707106781, T2m, T2j);
|
Chris@10
|
511 T3c = FMA(KP707106781, T2m, T2j);
|
Chris@10
|
512 T3g = FMA(KP707106781, T2E, T2t);
|
Chris@10
|
513 T2F = FNMS(KP707106781, T2E, T2t);
|
Chris@10
|
514 T2O = FNMS(KP707106781, T2N, T2K);
|
Chris@10
|
515 T3f = FMA(KP707106781, T2N, T2K);
|
Chris@10
|
516 }
|
Chris@10
|
517 {
|
Chris@10
|
518 E T3V, T3Y, T3G, T3J;
|
Chris@10
|
519 {
|
Chris@10
|
520 E T3v, T35, T2o, T34, T2P, T3C;
|
Chris@10
|
521 T47 = FNMS(KP707106781, T3u, T3t);
|
Chris@10
|
522 T3v = FMA(KP707106781, T3u, T3t);
|
Chris@10
|
523 T35 = FNMS(KP668178637, T2e, T2n);
|
Chris@10
|
524 T2o = FMA(KP668178637, T2n, T2e);
|
Chris@10
|
525 T34 = FMA(KP668178637, T2F, T2O);
|
Chris@10
|
526 T2P = FNMS(KP668178637, T2O, T2F);
|
Chris@10
|
527 T3C = T3y - T3B;
|
Chris@10
|
528 T4i = T3y + T3B;
|
Chris@10
|
529 T4h = FNMS(KP707106781, T3U, T3T);
|
Chris@10
|
530 T3V = FMA(KP707106781, T3U, T3T);
|
Chris@10
|
531 {
|
Chris@10
|
532 E T38, T36, T32, T2Q;
|
Chris@10
|
533 T38 = T35 + T34;
|
Chris@10
|
534 T36 = T34 - T35;
|
Chris@10
|
535 T32 = T2o + T2P;
|
Chris@10
|
536 T2Q = T2o - T2P;
|
Chris@10
|
537 T41 = FNMS(KP923879532, T3C, T3v);
|
Chris@10
|
538 T3D = FMA(KP923879532, T3C, T3v);
|
Chris@10
|
539 ro[WS(os, 29)] = FMA(KP831469612, T38, T37);
|
Chris@10
|
540 ro[WS(os, 13)] = FNMS(KP831469612, T38, T37);
|
Chris@10
|
541 io[WS(os, 5)] = FMA(KP831469612, T36, T33);
|
Chris@10
|
542 io[WS(os, 21)] = FNMS(KP831469612, T36, T33);
|
Chris@10
|
543 io[WS(os, 29)] = FMA(KP831469612, T32, T31);
|
Chris@10
|
544 io[WS(os, 13)] = FNMS(KP831469612, T32, T31);
|
Chris@10
|
545 ro[WS(os, 5)] = FMA(KP831469612, T2Q, T1X);
|
Chris@10
|
546 ro[WS(os, 21)] = FNMS(KP831469612, T2Q, T1X);
|
Chris@10
|
547 T3Y = T3W - T3X;
|
Chris@10
|
548 T48 = T3X + T3W;
|
Chris@10
|
549 }
|
Chris@10
|
550 }
|
Chris@10
|
551 T4b = FMA(KP707106781, T3F, T3E);
|
Chris@10
|
552 T3G = FNMS(KP707106781, T3F, T3E);
|
Chris@10
|
553 T3J = FNMS(KP707106781, T3I, T3H);
|
Chris@10
|
554 T4a = FMA(KP707106781, T3I, T3H);
|
Chris@10
|
555 T4e = FMA(KP707106781, T3M, T3L);
|
Chris@10
|
556 T3N = FNMS(KP707106781, T3M, T3L);
|
Chris@10
|
557 T45 = FMA(KP923879532, T3Y, T3V);
|
Chris@10
|
558 T3Z = FNMS(KP923879532, T3Y, T3V);
|
Chris@10
|
559 T42 = FNMS(KP668178637, T3G, T3J);
|
Chris@10
|
560 T3K = FMA(KP668178637, T3J, T3G);
|
Chris@10
|
561 T3Q = FNMS(KP707106781, T3P, T3O);
|
Chris@10
|
562 T4d = FMA(KP707106781, T3P, T3O);
|
Chris@10
|
563 }
|
Chris@10
|
564 }
|
Chris@10
|
565 {
|
Chris@10
|
566 E T4p, T49, T4l, T4j, T4n, T4c, T43, T3R, T4m, T4f;
|
Chris@10
|
567 T43 = FMA(KP668178637, T3N, T3Q);
|
Chris@10
|
568 T3R = FNMS(KP668178637, T3Q, T3N);
|
Chris@10
|
569 T4p = FMA(KP923879532, T48, T47);
|
Chris@10
|
570 T49 = FNMS(KP923879532, T48, T47);
|
Chris@10
|
571 {
|
Chris@10
|
572 E T44, T46, T40, T3S;
|
Chris@10
|
573 T44 = T42 - T43;
|
Chris@10
|
574 T46 = T42 + T43;
|
Chris@10
|
575 T40 = T3R - T3K;
|
Chris@10
|
576 T3S = T3K + T3R;
|
Chris@10
|
577 ro[WS(os, 11)] = FMA(KP831469612, T44, T41);
|
Chris@10
|
578 ro[WS(os, 27)] = FNMS(KP831469612, T44, T41);
|
Chris@10
|
579 io[WS(os, 3)] = FMA(KP831469612, T46, T45);
|
Chris@10
|
580 io[WS(os, 19)] = FNMS(KP831469612, T46, T45);
|
Chris@10
|
581 io[WS(os, 11)] = FMA(KP831469612, T40, T3Z);
|
Chris@10
|
582 io[WS(os, 27)] = FNMS(KP831469612, T40, T3Z);
|
Chris@10
|
583 ro[WS(os, 3)] = FMA(KP831469612, T3S, T3D);
|
Chris@10
|
584 ro[WS(os, 19)] = FNMS(KP831469612, T3S, T3D);
|
Chris@10
|
585 }
|
Chris@10
|
586 T4l = FNMS(KP923879532, T4i, T4h);
|
Chris@10
|
587 T4j = FMA(KP923879532, T4i, T4h);
|
Chris@10
|
588 T4n = FNMS(KP198912367, T4a, T4b);
|
Chris@10
|
589 T4c = FMA(KP198912367, T4b, T4a);
|
Chris@10
|
590 T4m = FMA(KP198912367, T4d, T4e);
|
Chris@10
|
591 T4f = FNMS(KP198912367, T4e, T4d);
|
Chris@10
|
592 T3n = FNMS(KP923879532, T3a, T39);
|
Chris@10
|
593 T3b = FMA(KP923879532, T3a, T39);
|
Chris@10
|
594 {
|
Chris@10
|
595 E T4q, T4o, T4k, T4g;
|
Chris@10
|
596 T4q = T4n + T4m;
|
Chris@10
|
597 T4o = T4m - T4n;
|
Chris@10
|
598 T4k = T4c + T4f;
|
Chris@10
|
599 T4g = T4c - T4f;
|
Chris@10
|
600 ro[WS(os, 31)] = FMA(KP980785280, T4q, T4p);
|
Chris@10
|
601 ro[WS(os, 15)] = FNMS(KP980785280, T4q, T4p);
|
Chris@10
|
602 io[WS(os, 7)] = FMA(KP980785280, T4o, T4l);
|
Chris@10
|
603 io[WS(os, 23)] = FNMS(KP980785280, T4o, T4l);
|
Chris@10
|
604 io[WS(os, 31)] = FMA(KP980785280, T4k, T4j);
|
Chris@10
|
605 io[WS(os, 15)] = FNMS(KP980785280, T4k, T4j);
|
Chris@10
|
606 ro[WS(os, 7)] = FMA(KP980785280, T4g, T49);
|
Chris@10
|
607 ro[WS(os, 23)] = FNMS(KP980785280, T4g, T49);
|
Chris@10
|
608 }
|
Chris@10
|
609 T3r = FMA(KP923879532, T3k, T3j);
|
Chris@10
|
610 T3l = FNMS(KP923879532, T3k, T3j);
|
Chris@10
|
611 T3o = FNMS(KP198912367, T3c, T3d);
|
Chris@10
|
612 T3e = FMA(KP198912367, T3d, T3c);
|
Chris@10
|
613 }
|
Chris@10
|
614 }
|
Chris@10
|
615 }
|
Chris@10
|
616 }
|
Chris@10
|
617 T3h = FNMS(KP198912367, T3g, T3f);
|
Chris@10
|
618 T3p = FMA(KP198912367, T3f, T3g);
|
Chris@10
|
619 {
|
Chris@10
|
620 E T3s, T3q, T3i, T3m;
|
Chris@10
|
621 T3s = T3o + T3p;
|
Chris@10
|
622 T3q = T3o - T3p;
|
Chris@10
|
623 T3i = T3e + T3h;
|
Chris@10
|
624 T3m = T3h - T3e;
|
Chris@10
|
625 ro[WS(os, 9)] = FMA(KP980785280, T3q, T3n);
|
Chris@10
|
626 ro[WS(os, 25)] = FNMS(KP980785280, T3q, T3n);
|
Chris@10
|
627 io[WS(os, 1)] = FMA(KP980785280, T3s, T3r);
|
Chris@10
|
628 io[WS(os, 17)] = FNMS(KP980785280, T3s, T3r);
|
Chris@10
|
629 io[WS(os, 9)] = FMA(KP980785280, T3m, T3l);
|
Chris@10
|
630 io[WS(os, 25)] = FNMS(KP980785280, T3m, T3l);
|
Chris@10
|
631 ro[WS(os, 1)] = FMA(KP980785280, T3i, T3b);
|
Chris@10
|
632 ro[WS(os, 17)] = FNMS(KP980785280, T3i, T3b);
|
Chris@10
|
633 }
|
Chris@10
|
634 }
|
Chris@10
|
635 }
|
Chris@10
|
636 }
|
Chris@10
|
637
|
Chris@10
|
638 static const kdft_desc desc = { 32, "n1_32", {236, 0, 136, 0}, &GENUS, 0, 0, 0, 0 };
|
Chris@10
|
639
|
Chris@10
|
640 void X(codelet_n1_32) (planner *p) {
|
Chris@10
|
641 X(kdft_register) (p, n1_32, &desc);
|
Chris@10
|
642 }
|
Chris@10
|
643
|
Chris@10
|
644 #else /* HAVE_FMA */
|
Chris@10
|
645
|
Chris@10
|
646 /* Generated by: ../../../genfft/gen_notw.native -compact -variables 4 -pipeline-latency 4 -n 32 -name n1_32 -include n.h */
|
Chris@10
|
647
|
Chris@10
|
648 /*
|
Chris@10
|
649 * This function contains 372 FP additions, 84 FP multiplications,
|
Chris@10
|
650 * (or, 340 additions, 52 multiplications, 32 fused multiply/add),
|
Chris@10
|
651 * 100 stack variables, 7 constants, and 128 memory accesses
|
Chris@10
|
652 */
|
Chris@10
|
653 #include "n.h"
|
Chris@10
|
654
|
Chris@10
|
655 static void n1_32(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
|
Chris@10
|
656 {
|
Chris@10
|
657 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
Chris@10
|
658 DK(KP555570233, +0.555570233019602224742830813948532874374937191);
|
Chris@10
|
659 DK(KP195090322, +0.195090322016128267848284868477022240927691618);
|
Chris@10
|
660 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
Chris@10
|
661 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
Chris@10
|
662 DK(KP382683432, +0.382683432365089771728459984030398866761344562);
|
Chris@10
|
663 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
Chris@10
|
664 {
|
Chris@10
|
665 INT i;
|
Chris@10
|
666 for (i = v; i > 0; i = i - 1, ri = ri + ivs, ii = ii + ivs, ro = ro + ovs, io = io + ovs, MAKE_VOLATILE_STRIDE(128, is), MAKE_VOLATILE_STRIDE(128, os)) {
|
Chris@10
|
667 E T7, T4r, T4Z, T18, T1z, T3t, T3T, T2T, Te, T1f, T50, T4s, T2W, T3u, T1G;
|
Chris@10
|
668 E T3U, Tm, T1n, T1O, T2Z, T3y, T3X, T4w, T53, Tt, T1u, T1V, T2Y, T3B, T3W;
|
Chris@10
|
669 E T4z, T52, T2t, T3L, T3O, T2K, TR, TY, T5F, T5G, T5H, T5I, T4R, T5j, T2E;
|
Chris@10
|
670 E T3P, T4W, T5k, T2N, T3M, T22, T3E, T3H, T2j, TC, TJ, T5A, T5B, T5C, T5D;
|
Chris@10
|
671 E T4G, T5g, T2d, T3F, T4L, T5h, T2m, T3I;
|
Chris@10
|
672 {
|
Chris@10
|
673 E T3, T1x, T14, T2S, T6, T2R, T17, T1y;
|
Chris@10
|
674 {
|
Chris@10
|
675 E T1, T2, T12, T13;
|
Chris@10
|
676 T1 = ri[0];
|
Chris@10
|
677 T2 = ri[WS(is, 16)];
|
Chris@10
|
678 T3 = T1 + T2;
|
Chris@10
|
679 T1x = T1 - T2;
|
Chris@10
|
680 T12 = ii[0];
|
Chris@10
|
681 T13 = ii[WS(is, 16)];
|
Chris@10
|
682 T14 = T12 + T13;
|
Chris@10
|
683 T2S = T12 - T13;
|
Chris@10
|
684 }
|
Chris@10
|
685 {
|
Chris@10
|
686 E T4, T5, T15, T16;
|
Chris@10
|
687 T4 = ri[WS(is, 8)];
|
Chris@10
|
688 T5 = ri[WS(is, 24)];
|
Chris@10
|
689 T6 = T4 + T5;
|
Chris@10
|
690 T2R = T4 - T5;
|
Chris@10
|
691 T15 = ii[WS(is, 8)];
|
Chris@10
|
692 T16 = ii[WS(is, 24)];
|
Chris@10
|
693 T17 = T15 + T16;
|
Chris@10
|
694 T1y = T15 - T16;
|
Chris@10
|
695 }
|
Chris@10
|
696 T7 = T3 + T6;
|
Chris@10
|
697 T4r = T3 - T6;
|
Chris@10
|
698 T4Z = T14 - T17;
|
Chris@10
|
699 T18 = T14 + T17;
|
Chris@10
|
700 T1z = T1x - T1y;
|
Chris@10
|
701 T3t = T1x + T1y;
|
Chris@10
|
702 T3T = T2S - T2R;
|
Chris@10
|
703 T2T = T2R + T2S;
|
Chris@10
|
704 }
|
Chris@10
|
705 {
|
Chris@10
|
706 E Ta, T1B, T1b, T1A, Td, T1D, T1e, T1E;
|
Chris@10
|
707 {
|
Chris@10
|
708 E T8, T9, T19, T1a;
|
Chris@10
|
709 T8 = ri[WS(is, 4)];
|
Chris@10
|
710 T9 = ri[WS(is, 20)];
|
Chris@10
|
711 Ta = T8 + T9;
|
Chris@10
|
712 T1B = T8 - T9;
|
Chris@10
|
713 T19 = ii[WS(is, 4)];
|
Chris@10
|
714 T1a = ii[WS(is, 20)];
|
Chris@10
|
715 T1b = T19 + T1a;
|
Chris@10
|
716 T1A = T19 - T1a;
|
Chris@10
|
717 }
|
Chris@10
|
718 {
|
Chris@10
|
719 E Tb, Tc, T1c, T1d;
|
Chris@10
|
720 Tb = ri[WS(is, 28)];
|
Chris@10
|
721 Tc = ri[WS(is, 12)];
|
Chris@10
|
722 Td = Tb + Tc;
|
Chris@10
|
723 T1D = Tb - Tc;
|
Chris@10
|
724 T1c = ii[WS(is, 28)];
|
Chris@10
|
725 T1d = ii[WS(is, 12)];
|
Chris@10
|
726 T1e = T1c + T1d;
|
Chris@10
|
727 T1E = T1c - T1d;
|
Chris@10
|
728 }
|
Chris@10
|
729 Te = Ta + Td;
|
Chris@10
|
730 T1f = T1b + T1e;
|
Chris@10
|
731 T50 = Td - Ta;
|
Chris@10
|
732 T4s = T1b - T1e;
|
Chris@10
|
733 {
|
Chris@10
|
734 E T2U, T2V, T1C, T1F;
|
Chris@10
|
735 T2U = T1D - T1E;
|
Chris@10
|
736 T2V = T1B + T1A;
|
Chris@10
|
737 T2W = KP707106781 * (T2U - T2V);
|
Chris@10
|
738 T3u = KP707106781 * (T2V + T2U);
|
Chris@10
|
739 T1C = T1A - T1B;
|
Chris@10
|
740 T1F = T1D + T1E;
|
Chris@10
|
741 T1G = KP707106781 * (T1C - T1F);
|
Chris@10
|
742 T3U = KP707106781 * (T1C + T1F);
|
Chris@10
|
743 }
|
Chris@10
|
744 }
|
Chris@10
|
745 {
|
Chris@10
|
746 E Ti, T1L, T1j, T1J, Tl, T1I, T1m, T1M, T1K, T1N;
|
Chris@10
|
747 {
|
Chris@10
|
748 E Tg, Th, T1h, T1i;
|
Chris@10
|
749 Tg = ri[WS(is, 2)];
|
Chris@10
|
750 Th = ri[WS(is, 18)];
|
Chris@10
|
751 Ti = Tg + Th;
|
Chris@10
|
752 T1L = Tg - Th;
|
Chris@10
|
753 T1h = ii[WS(is, 2)];
|
Chris@10
|
754 T1i = ii[WS(is, 18)];
|
Chris@10
|
755 T1j = T1h + T1i;
|
Chris@10
|
756 T1J = T1h - T1i;
|
Chris@10
|
757 }
|
Chris@10
|
758 {
|
Chris@10
|
759 E Tj, Tk, T1k, T1l;
|
Chris@10
|
760 Tj = ri[WS(is, 10)];
|
Chris@10
|
761 Tk = ri[WS(is, 26)];
|
Chris@10
|
762 Tl = Tj + Tk;
|
Chris@10
|
763 T1I = Tj - Tk;
|
Chris@10
|
764 T1k = ii[WS(is, 10)];
|
Chris@10
|
765 T1l = ii[WS(is, 26)];
|
Chris@10
|
766 T1m = T1k + T1l;
|
Chris@10
|
767 T1M = T1k - T1l;
|
Chris@10
|
768 }
|
Chris@10
|
769 Tm = Ti + Tl;
|
Chris@10
|
770 T1n = T1j + T1m;
|
Chris@10
|
771 T1K = T1I + T1J;
|
Chris@10
|
772 T1N = T1L - T1M;
|
Chris@10
|
773 T1O = FNMS(KP923879532, T1N, KP382683432 * T1K);
|
Chris@10
|
774 T2Z = FMA(KP923879532, T1K, KP382683432 * T1N);
|
Chris@10
|
775 {
|
Chris@10
|
776 E T3w, T3x, T4u, T4v;
|
Chris@10
|
777 T3w = T1J - T1I;
|
Chris@10
|
778 T3x = T1L + T1M;
|
Chris@10
|
779 T3y = FNMS(KP382683432, T3x, KP923879532 * T3w);
|
Chris@10
|
780 T3X = FMA(KP382683432, T3w, KP923879532 * T3x);
|
Chris@10
|
781 T4u = T1j - T1m;
|
Chris@10
|
782 T4v = Ti - Tl;
|
Chris@10
|
783 T4w = T4u - T4v;
|
Chris@10
|
784 T53 = T4v + T4u;
|
Chris@10
|
785 }
|
Chris@10
|
786 }
|
Chris@10
|
787 {
|
Chris@10
|
788 E Tp, T1S, T1q, T1Q, Ts, T1P, T1t, T1T, T1R, T1U;
|
Chris@10
|
789 {
|
Chris@10
|
790 E Tn, To, T1o, T1p;
|
Chris@10
|
791 Tn = ri[WS(is, 30)];
|
Chris@10
|
792 To = ri[WS(is, 14)];
|
Chris@10
|
793 Tp = Tn + To;
|
Chris@10
|
794 T1S = Tn - To;
|
Chris@10
|
795 T1o = ii[WS(is, 30)];
|
Chris@10
|
796 T1p = ii[WS(is, 14)];
|
Chris@10
|
797 T1q = T1o + T1p;
|
Chris@10
|
798 T1Q = T1o - T1p;
|
Chris@10
|
799 }
|
Chris@10
|
800 {
|
Chris@10
|
801 E Tq, Tr, T1r, T1s;
|
Chris@10
|
802 Tq = ri[WS(is, 6)];
|
Chris@10
|
803 Tr = ri[WS(is, 22)];
|
Chris@10
|
804 Ts = Tq + Tr;
|
Chris@10
|
805 T1P = Tq - Tr;
|
Chris@10
|
806 T1r = ii[WS(is, 6)];
|
Chris@10
|
807 T1s = ii[WS(is, 22)];
|
Chris@10
|
808 T1t = T1r + T1s;
|
Chris@10
|
809 T1T = T1r - T1s;
|
Chris@10
|
810 }
|
Chris@10
|
811 Tt = Tp + Ts;
|
Chris@10
|
812 T1u = T1q + T1t;
|
Chris@10
|
813 T1R = T1P + T1Q;
|
Chris@10
|
814 T1U = T1S - T1T;
|
Chris@10
|
815 T1V = FMA(KP382683432, T1R, KP923879532 * T1U);
|
Chris@10
|
816 T2Y = FNMS(KP923879532, T1R, KP382683432 * T1U);
|
Chris@10
|
817 {
|
Chris@10
|
818 E T3z, T3A, T4x, T4y;
|
Chris@10
|
819 T3z = T1Q - T1P;
|
Chris@10
|
820 T3A = T1S + T1T;
|
Chris@10
|
821 T3B = FMA(KP923879532, T3z, KP382683432 * T3A);
|
Chris@10
|
822 T3W = FNMS(KP382683432, T3z, KP923879532 * T3A);
|
Chris@10
|
823 T4x = Tp - Ts;
|
Chris@10
|
824 T4y = T1q - T1t;
|
Chris@10
|
825 T4z = T4x + T4y;
|
Chris@10
|
826 T52 = T4x - T4y;
|
Chris@10
|
827 }
|
Chris@10
|
828 }
|
Chris@10
|
829 {
|
Chris@10
|
830 E TN, T2p, T2J, T4S, TQ, T2G, T2s, T4T, TU, T2x, T2w, T4O, TX, T2z, T2C;
|
Chris@10
|
831 E T4P;
|
Chris@10
|
832 {
|
Chris@10
|
833 E TL, TM, T2H, T2I;
|
Chris@10
|
834 TL = ri[WS(is, 31)];
|
Chris@10
|
835 TM = ri[WS(is, 15)];
|
Chris@10
|
836 TN = TL + TM;
|
Chris@10
|
837 T2p = TL - TM;
|
Chris@10
|
838 T2H = ii[WS(is, 31)];
|
Chris@10
|
839 T2I = ii[WS(is, 15)];
|
Chris@10
|
840 T2J = T2H - T2I;
|
Chris@10
|
841 T4S = T2H + T2I;
|
Chris@10
|
842 }
|
Chris@10
|
843 {
|
Chris@10
|
844 E TO, TP, T2q, T2r;
|
Chris@10
|
845 TO = ri[WS(is, 7)];
|
Chris@10
|
846 TP = ri[WS(is, 23)];
|
Chris@10
|
847 TQ = TO + TP;
|
Chris@10
|
848 T2G = TO - TP;
|
Chris@10
|
849 T2q = ii[WS(is, 7)];
|
Chris@10
|
850 T2r = ii[WS(is, 23)];
|
Chris@10
|
851 T2s = T2q - T2r;
|
Chris@10
|
852 T4T = T2q + T2r;
|
Chris@10
|
853 }
|
Chris@10
|
854 {
|
Chris@10
|
855 E TS, TT, T2u, T2v;
|
Chris@10
|
856 TS = ri[WS(is, 3)];
|
Chris@10
|
857 TT = ri[WS(is, 19)];
|
Chris@10
|
858 TU = TS + TT;
|
Chris@10
|
859 T2x = TS - TT;
|
Chris@10
|
860 T2u = ii[WS(is, 3)];
|
Chris@10
|
861 T2v = ii[WS(is, 19)];
|
Chris@10
|
862 T2w = T2u - T2v;
|
Chris@10
|
863 T4O = T2u + T2v;
|
Chris@10
|
864 }
|
Chris@10
|
865 {
|
Chris@10
|
866 E TV, TW, T2A, T2B;
|
Chris@10
|
867 TV = ri[WS(is, 27)];
|
Chris@10
|
868 TW = ri[WS(is, 11)];
|
Chris@10
|
869 TX = TV + TW;
|
Chris@10
|
870 T2z = TV - TW;
|
Chris@10
|
871 T2A = ii[WS(is, 27)];
|
Chris@10
|
872 T2B = ii[WS(is, 11)];
|
Chris@10
|
873 T2C = T2A - T2B;
|
Chris@10
|
874 T4P = T2A + T2B;
|
Chris@10
|
875 }
|
Chris@10
|
876 T2t = T2p - T2s;
|
Chris@10
|
877 T3L = T2p + T2s;
|
Chris@10
|
878 T3O = T2J - T2G;
|
Chris@10
|
879 T2K = T2G + T2J;
|
Chris@10
|
880 TR = TN + TQ;
|
Chris@10
|
881 TY = TU + TX;
|
Chris@10
|
882 T5F = TR - TY;
|
Chris@10
|
883 {
|
Chris@10
|
884 E T4N, T4Q, T2y, T2D;
|
Chris@10
|
885 T5G = T4S + T4T;
|
Chris@10
|
886 T5H = T4O + T4P;
|
Chris@10
|
887 T5I = T5G - T5H;
|
Chris@10
|
888 T4N = TN - TQ;
|
Chris@10
|
889 T4Q = T4O - T4P;
|
Chris@10
|
890 T4R = T4N - T4Q;
|
Chris@10
|
891 T5j = T4N + T4Q;
|
Chris@10
|
892 T2y = T2w - T2x;
|
Chris@10
|
893 T2D = T2z + T2C;
|
Chris@10
|
894 T2E = KP707106781 * (T2y - T2D);
|
Chris@10
|
895 T3P = KP707106781 * (T2y + T2D);
|
Chris@10
|
896 {
|
Chris@10
|
897 E T4U, T4V, T2L, T2M;
|
Chris@10
|
898 T4U = T4S - T4T;
|
Chris@10
|
899 T4V = TX - TU;
|
Chris@10
|
900 T4W = T4U - T4V;
|
Chris@10
|
901 T5k = T4V + T4U;
|
Chris@10
|
902 T2L = T2z - T2C;
|
Chris@10
|
903 T2M = T2x + T2w;
|
Chris@10
|
904 T2N = KP707106781 * (T2L - T2M);
|
Chris@10
|
905 T3M = KP707106781 * (T2M + T2L);
|
Chris@10
|
906 }
|
Chris@10
|
907 }
|
Chris@10
|
908 }
|
Chris@10
|
909 {
|
Chris@10
|
910 E Ty, T2f, T21, T4C, TB, T1Y, T2i, T4D, TF, T28, T2b, T4I, TI, T23, T26;
|
Chris@10
|
911 E T4J;
|
Chris@10
|
912 {
|
Chris@10
|
913 E Tw, Tx, T1Z, T20;
|
Chris@10
|
914 Tw = ri[WS(is, 1)];
|
Chris@10
|
915 Tx = ri[WS(is, 17)];
|
Chris@10
|
916 Ty = Tw + Tx;
|
Chris@10
|
917 T2f = Tw - Tx;
|
Chris@10
|
918 T1Z = ii[WS(is, 1)];
|
Chris@10
|
919 T20 = ii[WS(is, 17)];
|
Chris@10
|
920 T21 = T1Z - T20;
|
Chris@10
|
921 T4C = T1Z + T20;
|
Chris@10
|
922 }
|
Chris@10
|
923 {
|
Chris@10
|
924 E Tz, TA, T2g, T2h;
|
Chris@10
|
925 Tz = ri[WS(is, 9)];
|
Chris@10
|
926 TA = ri[WS(is, 25)];
|
Chris@10
|
927 TB = Tz + TA;
|
Chris@10
|
928 T1Y = Tz - TA;
|
Chris@10
|
929 T2g = ii[WS(is, 9)];
|
Chris@10
|
930 T2h = ii[WS(is, 25)];
|
Chris@10
|
931 T2i = T2g - T2h;
|
Chris@10
|
932 T4D = T2g + T2h;
|
Chris@10
|
933 }
|
Chris@10
|
934 {
|
Chris@10
|
935 E TD, TE, T29, T2a;
|
Chris@10
|
936 TD = ri[WS(is, 5)];
|
Chris@10
|
937 TE = ri[WS(is, 21)];
|
Chris@10
|
938 TF = TD + TE;
|
Chris@10
|
939 T28 = TD - TE;
|
Chris@10
|
940 T29 = ii[WS(is, 5)];
|
Chris@10
|
941 T2a = ii[WS(is, 21)];
|
Chris@10
|
942 T2b = T29 - T2a;
|
Chris@10
|
943 T4I = T29 + T2a;
|
Chris@10
|
944 }
|
Chris@10
|
945 {
|
Chris@10
|
946 E TG, TH, T24, T25;
|
Chris@10
|
947 TG = ri[WS(is, 29)];
|
Chris@10
|
948 TH = ri[WS(is, 13)];
|
Chris@10
|
949 TI = TG + TH;
|
Chris@10
|
950 T23 = TG - TH;
|
Chris@10
|
951 T24 = ii[WS(is, 29)];
|
Chris@10
|
952 T25 = ii[WS(is, 13)];
|
Chris@10
|
953 T26 = T24 - T25;
|
Chris@10
|
954 T4J = T24 + T25;
|
Chris@10
|
955 }
|
Chris@10
|
956 T22 = T1Y + T21;
|
Chris@10
|
957 T3E = T2f + T2i;
|
Chris@10
|
958 T3H = T21 - T1Y;
|
Chris@10
|
959 T2j = T2f - T2i;
|
Chris@10
|
960 TC = Ty + TB;
|
Chris@10
|
961 TJ = TF + TI;
|
Chris@10
|
962 T5A = TC - TJ;
|
Chris@10
|
963 {
|
Chris@10
|
964 E T4E, T4F, T27, T2c;
|
Chris@10
|
965 T5B = T4C + T4D;
|
Chris@10
|
966 T5C = T4I + T4J;
|
Chris@10
|
967 T5D = T5B - T5C;
|
Chris@10
|
968 T4E = T4C - T4D;
|
Chris@10
|
969 T4F = TI - TF;
|
Chris@10
|
970 T4G = T4E - T4F;
|
Chris@10
|
971 T5g = T4F + T4E;
|
Chris@10
|
972 T27 = T23 - T26;
|
Chris@10
|
973 T2c = T28 + T2b;
|
Chris@10
|
974 T2d = KP707106781 * (T27 - T2c);
|
Chris@10
|
975 T3F = KP707106781 * (T2c + T27);
|
Chris@10
|
976 {
|
Chris@10
|
977 E T4H, T4K, T2k, T2l;
|
Chris@10
|
978 T4H = Ty - TB;
|
Chris@10
|
979 T4K = T4I - T4J;
|
Chris@10
|
980 T4L = T4H - T4K;
|
Chris@10
|
981 T5h = T4H + T4K;
|
Chris@10
|
982 T2k = T2b - T28;
|
Chris@10
|
983 T2l = T23 + T26;
|
Chris@10
|
984 T2m = KP707106781 * (T2k - T2l);
|
Chris@10
|
985 T3I = KP707106781 * (T2k + T2l);
|
Chris@10
|
986 }
|
Chris@10
|
987 }
|
Chris@10
|
988 }
|
Chris@10
|
989 {
|
Chris@10
|
990 E T4B, T57, T5a, T5c, T4Y, T56, T55, T5b;
|
Chris@10
|
991 {
|
Chris@10
|
992 E T4t, T4A, T58, T59;
|
Chris@10
|
993 T4t = T4r - T4s;
|
Chris@10
|
994 T4A = KP707106781 * (T4w - T4z);
|
Chris@10
|
995 T4B = T4t + T4A;
|
Chris@10
|
996 T57 = T4t - T4A;
|
Chris@10
|
997 T58 = FNMS(KP923879532, T4L, KP382683432 * T4G);
|
Chris@10
|
998 T59 = FMA(KP382683432, T4W, KP923879532 * T4R);
|
Chris@10
|
999 T5a = T58 - T59;
|
Chris@10
|
1000 T5c = T58 + T59;
|
Chris@10
|
1001 }
|
Chris@10
|
1002 {
|
Chris@10
|
1003 E T4M, T4X, T51, T54;
|
Chris@10
|
1004 T4M = FMA(KP923879532, T4G, KP382683432 * T4L);
|
Chris@10
|
1005 T4X = FNMS(KP923879532, T4W, KP382683432 * T4R);
|
Chris@10
|
1006 T4Y = T4M + T4X;
|
Chris@10
|
1007 T56 = T4X - T4M;
|
Chris@10
|
1008 T51 = T4Z - T50;
|
Chris@10
|
1009 T54 = KP707106781 * (T52 - T53);
|
Chris@10
|
1010 T55 = T51 - T54;
|
Chris@10
|
1011 T5b = T51 + T54;
|
Chris@10
|
1012 }
|
Chris@10
|
1013 ro[WS(os, 22)] = T4B - T4Y;
|
Chris@10
|
1014 io[WS(os, 22)] = T5b - T5c;
|
Chris@10
|
1015 ro[WS(os, 6)] = T4B + T4Y;
|
Chris@10
|
1016 io[WS(os, 6)] = T5b + T5c;
|
Chris@10
|
1017 io[WS(os, 30)] = T55 - T56;
|
Chris@10
|
1018 ro[WS(os, 30)] = T57 - T5a;
|
Chris@10
|
1019 io[WS(os, 14)] = T55 + T56;
|
Chris@10
|
1020 ro[WS(os, 14)] = T57 + T5a;
|
Chris@10
|
1021 }
|
Chris@10
|
1022 {
|
Chris@10
|
1023 E T5f, T5r, T5u, T5w, T5m, T5q, T5p, T5v;
|
Chris@10
|
1024 {
|
Chris@10
|
1025 E T5d, T5e, T5s, T5t;
|
Chris@10
|
1026 T5d = T4r + T4s;
|
Chris@10
|
1027 T5e = KP707106781 * (T53 + T52);
|
Chris@10
|
1028 T5f = T5d + T5e;
|
Chris@10
|
1029 T5r = T5d - T5e;
|
Chris@10
|
1030 T5s = FNMS(KP382683432, T5h, KP923879532 * T5g);
|
Chris@10
|
1031 T5t = FMA(KP923879532, T5k, KP382683432 * T5j);
|
Chris@10
|
1032 T5u = T5s - T5t;
|
Chris@10
|
1033 T5w = T5s + T5t;
|
Chris@10
|
1034 }
|
Chris@10
|
1035 {
|
Chris@10
|
1036 E T5i, T5l, T5n, T5o;
|
Chris@10
|
1037 T5i = FMA(KP382683432, T5g, KP923879532 * T5h);
|
Chris@10
|
1038 T5l = FNMS(KP382683432, T5k, KP923879532 * T5j);
|
Chris@10
|
1039 T5m = T5i + T5l;
|
Chris@10
|
1040 T5q = T5l - T5i;
|
Chris@10
|
1041 T5n = T50 + T4Z;
|
Chris@10
|
1042 T5o = KP707106781 * (T4w + T4z);
|
Chris@10
|
1043 T5p = T5n - T5o;
|
Chris@10
|
1044 T5v = T5n + T5o;
|
Chris@10
|
1045 }
|
Chris@10
|
1046 ro[WS(os, 18)] = T5f - T5m;
|
Chris@10
|
1047 io[WS(os, 18)] = T5v - T5w;
|
Chris@10
|
1048 ro[WS(os, 2)] = T5f + T5m;
|
Chris@10
|
1049 io[WS(os, 2)] = T5v + T5w;
|
Chris@10
|
1050 io[WS(os, 26)] = T5p - T5q;
|
Chris@10
|
1051 ro[WS(os, 26)] = T5r - T5u;
|
Chris@10
|
1052 io[WS(os, 10)] = T5p + T5q;
|
Chris@10
|
1053 ro[WS(os, 10)] = T5r + T5u;
|
Chris@10
|
1054 }
|
Chris@10
|
1055 {
|
Chris@10
|
1056 E T5z, T5P, T5S, T5U, T5K, T5O, T5N, T5T;
|
Chris@10
|
1057 {
|
Chris@10
|
1058 E T5x, T5y, T5Q, T5R;
|
Chris@10
|
1059 T5x = T7 - Te;
|
Chris@10
|
1060 T5y = T1n - T1u;
|
Chris@10
|
1061 T5z = T5x + T5y;
|
Chris@10
|
1062 T5P = T5x - T5y;
|
Chris@10
|
1063 T5Q = T5D - T5A;
|
Chris@10
|
1064 T5R = T5F + T5I;
|
Chris@10
|
1065 T5S = KP707106781 * (T5Q - T5R);
|
Chris@10
|
1066 T5U = KP707106781 * (T5Q + T5R);
|
Chris@10
|
1067 }
|
Chris@10
|
1068 {
|
Chris@10
|
1069 E T5E, T5J, T5L, T5M;
|
Chris@10
|
1070 T5E = T5A + T5D;
|
Chris@10
|
1071 T5J = T5F - T5I;
|
Chris@10
|
1072 T5K = KP707106781 * (T5E + T5J);
|
Chris@10
|
1073 T5O = KP707106781 * (T5J - T5E);
|
Chris@10
|
1074 T5L = T18 - T1f;
|
Chris@10
|
1075 T5M = Tt - Tm;
|
Chris@10
|
1076 T5N = T5L - T5M;
|
Chris@10
|
1077 T5T = T5M + T5L;
|
Chris@10
|
1078 }
|
Chris@10
|
1079 ro[WS(os, 20)] = T5z - T5K;
|
Chris@10
|
1080 io[WS(os, 20)] = T5T - T5U;
|
Chris@10
|
1081 ro[WS(os, 4)] = T5z + T5K;
|
Chris@10
|
1082 io[WS(os, 4)] = T5T + T5U;
|
Chris@10
|
1083 io[WS(os, 28)] = T5N - T5O;
|
Chris@10
|
1084 ro[WS(os, 28)] = T5P - T5S;
|
Chris@10
|
1085 io[WS(os, 12)] = T5N + T5O;
|
Chris@10
|
1086 ro[WS(os, 12)] = T5P + T5S;
|
Chris@10
|
1087 }
|
Chris@10
|
1088 {
|
Chris@10
|
1089 E Tv, T5V, T5Y, T60, T10, T11, T1w, T5Z;
|
Chris@10
|
1090 {
|
Chris@10
|
1091 E Tf, Tu, T5W, T5X;
|
Chris@10
|
1092 Tf = T7 + Te;
|
Chris@10
|
1093 Tu = Tm + Tt;
|
Chris@10
|
1094 Tv = Tf + Tu;
|
Chris@10
|
1095 T5V = Tf - Tu;
|
Chris@10
|
1096 T5W = T5B + T5C;
|
Chris@10
|
1097 T5X = T5G + T5H;
|
Chris@10
|
1098 T5Y = T5W - T5X;
|
Chris@10
|
1099 T60 = T5W + T5X;
|
Chris@10
|
1100 }
|
Chris@10
|
1101 {
|
Chris@10
|
1102 E TK, TZ, T1g, T1v;
|
Chris@10
|
1103 TK = TC + TJ;
|
Chris@10
|
1104 TZ = TR + TY;
|
Chris@10
|
1105 T10 = TK + TZ;
|
Chris@10
|
1106 T11 = TZ - TK;
|
Chris@10
|
1107 T1g = T18 + T1f;
|
Chris@10
|
1108 T1v = T1n + T1u;
|
Chris@10
|
1109 T1w = T1g - T1v;
|
Chris@10
|
1110 T5Z = T1g + T1v;
|
Chris@10
|
1111 }
|
Chris@10
|
1112 ro[WS(os, 16)] = Tv - T10;
|
Chris@10
|
1113 io[WS(os, 16)] = T5Z - T60;
|
Chris@10
|
1114 ro[0] = Tv + T10;
|
Chris@10
|
1115 io[0] = T5Z + T60;
|
Chris@10
|
1116 io[WS(os, 8)] = T11 + T1w;
|
Chris@10
|
1117 ro[WS(os, 8)] = T5V + T5Y;
|
Chris@10
|
1118 io[WS(os, 24)] = T1w - T11;
|
Chris@10
|
1119 ro[WS(os, 24)] = T5V - T5Y;
|
Chris@10
|
1120 }
|
Chris@10
|
1121 {
|
Chris@10
|
1122 E T1X, T33, T31, T37, T2o, T34, T2P, T35;
|
Chris@10
|
1123 {
|
Chris@10
|
1124 E T1H, T1W, T2X, T30;
|
Chris@10
|
1125 T1H = T1z - T1G;
|
Chris@10
|
1126 T1W = T1O - T1V;
|
Chris@10
|
1127 T1X = T1H + T1W;
|
Chris@10
|
1128 T33 = T1H - T1W;
|
Chris@10
|
1129 T2X = T2T - T2W;
|
Chris@10
|
1130 T30 = T2Y - T2Z;
|
Chris@10
|
1131 T31 = T2X - T30;
|
Chris@10
|
1132 T37 = T2X + T30;
|
Chris@10
|
1133 }
|
Chris@10
|
1134 {
|
Chris@10
|
1135 E T2e, T2n, T2F, T2O;
|
Chris@10
|
1136 T2e = T22 - T2d;
|
Chris@10
|
1137 T2n = T2j - T2m;
|
Chris@10
|
1138 T2o = FMA(KP980785280, T2e, KP195090322 * T2n);
|
Chris@10
|
1139 T34 = FNMS(KP980785280, T2n, KP195090322 * T2e);
|
Chris@10
|
1140 T2F = T2t - T2E;
|
Chris@10
|
1141 T2O = T2K - T2N;
|
Chris@10
|
1142 T2P = FNMS(KP980785280, T2O, KP195090322 * T2F);
|
Chris@10
|
1143 T35 = FMA(KP195090322, T2O, KP980785280 * T2F);
|
Chris@10
|
1144 }
|
Chris@10
|
1145 {
|
Chris@10
|
1146 E T2Q, T38, T32, T36;
|
Chris@10
|
1147 T2Q = T2o + T2P;
|
Chris@10
|
1148 ro[WS(os, 23)] = T1X - T2Q;
|
Chris@10
|
1149 ro[WS(os, 7)] = T1X + T2Q;
|
Chris@10
|
1150 T38 = T34 + T35;
|
Chris@10
|
1151 io[WS(os, 23)] = T37 - T38;
|
Chris@10
|
1152 io[WS(os, 7)] = T37 + T38;
|
Chris@10
|
1153 T32 = T2P - T2o;
|
Chris@10
|
1154 io[WS(os, 31)] = T31 - T32;
|
Chris@10
|
1155 io[WS(os, 15)] = T31 + T32;
|
Chris@10
|
1156 T36 = T34 - T35;
|
Chris@10
|
1157 ro[WS(os, 31)] = T33 - T36;
|
Chris@10
|
1158 ro[WS(os, 15)] = T33 + T36;
|
Chris@10
|
1159 }
|
Chris@10
|
1160 }
|
Chris@10
|
1161 {
|
Chris@10
|
1162 E T3D, T41, T3Z, T45, T3K, T42, T3R, T43;
|
Chris@10
|
1163 {
|
Chris@10
|
1164 E T3v, T3C, T3V, T3Y;
|
Chris@10
|
1165 T3v = T3t - T3u;
|
Chris@10
|
1166 T3C = T3y - T3B;
|
Chris@10
|
1167 T3D = T3v + T3C;
|
Chris@10
|
1168 T41 = T3v - T3C;
|
Chris@10
|
1169 T3V = T3T - T3U;
|
Chris@10
|
1170 T3Y = T3W - T3X;
|
Chris@10
|
1171 T3Z = T3V - T3Y;
|
Chris@10
|
1172 T45 = T3V + T3Y;
|
Chris@10
|
1173 }
|
Chris@10
|
1174 {
|
Chris@10
|
1175 E T3G, T3J, T3N, T3Q;
|
Chris@10
|
1176 T3G = T3E - T3F;
|
Chris@10
|
1177 T3J = T3H - T3I;
|
Chris@10
|
1178 T3K = FMA(KP555570233, T3G, KP831469612 * T3J);
|
Chris@10
|
1179 T42 = FNMS(KP831469612, T3G, KP555570233 * T3J);
|
Chris@10
|
1180 T3N = T3L - T3M;
|
Chris@10
|
1181 T3Q = T3O - T3P;
|
Chris@10
|
1182 T3R = FNMS(KP831469612, T3Q, KP555570233 * T3N);
|
Chris@10
|
1183 T43 = FMA(KP831469612, T3N, KP555570233 * T3Q);
|
Chris@10
|
1184 }
|
Chris@10
|
1185 {
|
Chris@10
|
1186 E T3S, T46, T40, T44;
|
Chris@10
|
1187 T3S = T3K + T3R;
|
Chris@10
|
1188 ro[WS(os, 21)] = T3D - T3S;
|
Chris@10
|
1189 ro[WS(os, 5)] = T3D + T3S;
|
Chris@10
|
1190 T46 = T42 + T43;
|
Chris@10
|
1191 io[WS(os, 21)] = T45 - T46;
|
Chris@10
|
1192 io[WS(os, 5)] = T45 + T46;
|
Chris@10
|
1193 T40 = T3R - T3K;
|
Chris@10
|
1194 io[WS(os, 29)] = T3Z - T40;
|
Chris@10
|
1195 io[WS(os, 13)] = T3Z + T40;
|
Chris@10
|
1196 T44 = T42 - T43;
|
Chris@10
|
1197 ro[WS(os, 29)] = T41 - T44;
|
Chris@10
|
1198 ro[WS(os, 13)] = T41 + T44;
|
Chris@10
|
1199 }
|
Chris@10
|
1200 }
|
Chris@10
|
1201 {
|
Chris@10
|
1202 E T49, T4l, T4j, T4p, T4c, T4m, T4f, T4n;
|
Chris@10
|
1203 {
|
Chris@10
|
1204 E T47, T48, T4h, T4i;
|
Chris@10
|
1205 T47 = T3t + T3u;
|
Chris@10
|
1206 T48 = T3X + T3W;
|
Chris@10
|
1207 T49 = T47 + T48;
|
Chris@10
|
1208 T4l = T47 - T48;
|
Chris@10
|
1209 T4h = T3T + T3U;
|
Chris@10
|
1210 T4i = T3y + T3B;
|
Chris@10
|
1211 T4j = T4h - T4i;
|
Chris@10
|
1212 T4p = T4h + T4i;
|
Chris@10
|
1213 }
|
Chris@10
|
1214 {
|
Chris@10
|
1215 E T4a, T4b, T4d, T4e;
|
Chris@10
|
1216 T4a = T3E + T3F;
|
Chris@10
|
1217 T4b = T3H + T3I;
|
Chris@10
|
1218 T4c = FMA(KP980785280, T4a, KP195090322 * T4b);
|
Chris@10
|
1219 T4m = FNMS(KP195090322, T4a, KP980785280 * T4b);
|
Chris@10
|
1220 T4d = T3L + T3M;
|
Chris@10
|
1221 T4e = T3O + T3P;
|
Chris@10
|
1222 T4f = FNMS(KP195090322, T4e, KP980785280 * T4d);
|
Chris@10
|
1223 T4n = FMA(KP195090322, T4d, KP980785280 * T4e);
|
Chris@10
|
1224 }
|
Chris@10
|
1225 {
|
Chris@10
|
1226 E T4g, T4q, T4k, T4o;
|
Chris@10
|
1227 T4g = T4c + T4f;
|
Chris@10
|
1228 ro[WS(os, 17)] = T49 - T4g;
|
Chris@10
|
1229 ro[WS(os, 1)] = T49 + T4g;
|
Chris@10
|
1230 T4q = T4m + T4n;
|
Chris@10
|
1231 io[WS(os, 17)] = T4p - T4q;
|
Chris@10
|
1232 io[WS(os, 1)] = T4p + T4q;
|
Chris@10
|
1233 T4k = T4f - T4c;
|
Chris@10
|
1234 io[WS(os, 25)] = T4j - T4k;
|
Chris@10
|
1235 io[WS(os, 9)] = T4j + T4k;
|
Chris@10
|
1236 T4o = T4m - T4n;
|
Chris@10
|
1237 ro[WS(os, 25)] = T4l - T4o;
|
Chris@10
|
1238 ro[WS(os, 9)] = T4l + T4o;
|
Chris@10
|
1239 }
|
Chris@10
|
1240 }
|
Chris@10
|
1241 {
|
Chris@10
|
1242 E T3b, T3n, T3l, T3r, T3e, T3o, T3h, T3p;
|
Chris@10
|
1243 {
|
Chris@10
|
1244 E T39, T3a, T3j, T3k;
|
Chris@10
|
1245 T39 = T1z + T1G;
|
Chris@10
|
1246 T3a = T2Z + T2Y;
|
Chris@10
|
1247 T3b = T39 + T3a;
|
Chris@10
|
1248 T3n = T39 - T3a;
|
Chris@10
|
1249 T3j = T2T + T2W;
|
Chris@10
|
1250 T3k = T1O + T1V;
|
Chris@10
|
1251 T3l = T3j - T3k;
|
Chris@10
|
1252 T3r = T3j + T3k;
|
Chris@10
|
1253 }
|
Chris@10
|
1254 {
|
Chris@10
|
1255 E T3c, T3d, T3f, T3g;
|
Chris@10
|
1256 T3c = T22 + T2d;
|
Chris@10
|
1257 T3d = T2j + T2m;
|
Chris@10
|
1258 T3e = FMA(KP555570233, T3c, KP831469612 * T3d);
|
Chris@10
|
1259 T3o = FNMS(KP555570233, T3d, KP831469612 * T3c);
|
Chris@10
|
1260 T3f = T2t + T2E;
|
Chris@10
|
1261 T3g = T2K + T2N;
|
Chris@10
|
1262 T3h = FNMS(KP555570233, T3g, KP831469612 * T3f);
|
Chris@10
|
1263 T3p = FMA(KP831469612, T3g, KP555570233 * T3f);
|
Chris@10
|
1264 }
|
Chris@10
|
1265 {
|
Chris@10
|
1266 E T3i, T3s, T3m, T3q;
|
Chris@10
|
1267 T3i = T3e + T3h;
|
Chris@10
|
1268 ro[WS(os, 19)] = T3b - T3i;
|
Chris@10
|
1269 ro[WS(os, 3)] = T3b + T3i;
|
Chris@10
|
1270 T3s = T3o + T3p;
|
Chris@10
|
1271 io[WS(os, 19)] = T3r - T3s;
|
Chris@10
|
1272 io[WS(os, 3)] = T3r + T3s;
|
Chris@10
|
1273 T3m = T3h - T3e;
|
Chris@10
|
1274 io[WS(os, 27)] = T3l - T3m;
|
Chris@10
|
1275 io[WS(os, 11)] = T3l + T3m;
|
Chris@10
|
1276 T3q = T3o - T3p;
|
Chris@10
|
1277 ro[WS(os, 27)] = T3n - T3q;
|
Chris@10
|
1278 ro[WS(os, 11)] = T3n + T3q;
|
Chris@10
|
1279 }
|
Chris@10
|
1280 }
|
Chris@10
|
1281 }
|
Chris@10
|
1282 }
|
Chris@10
|
1283 }
|
Chris@10
|
1284
|
Chris@10
|
1285 static const kdft_desc desc = { 32, "n1_32", {340, 52, 32, 0}, &GENUS, 0, 0, 0, 0 };
|
Chris@10
|
1286
|
Chris@10
|
1287 void X(codelet_n1_32) (planner *p) {
|
Chris@10
|
1288 X(kdft_register) (p, n1_32, &desc);
|
Chris@10
|
1289 }
|
Chris@10
|
1290
|
Chris@10
|
1291 #endif /* HAVE_FMA */
|