cannam@167
|
1 /*
|
cannam@167
|
2 * Copyright (c) 2003, 2007-14 Matteo Frigo
|
cannam@167
|
3 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
|
cannam@167
|
4 *
|
cannam@167
|
5 * This program is free software; you can redistribute it and/or modify
|
cannam@167
|
6 * it under the terms of the GNU General Public License as published by
|
cannam@167
|
7 * the Free Software Foundation; either version 2 of the License, or
|
cannam@167
|
8 * (at your option) any later version.
|
cannam@167
|
9 *
|
cannam@167
|
10 * This program is distributed in the hope that it will be useful,
|
cannam@167
|
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
cannam@167
|
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
cannam@167
|
13 * GNU General Public License for more details.
|
cannam@167
|
14 *
|
cannam@167
|
15 * You should have received a copy of the GNU General Public License
|
cannam@167
|
16 * along with this program; if not, write to the Free Software
|
cannam@167
|
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
cannam@167
|
18 *
|
cannam@167
|
19 */
|
cannam@167
|
20
|
cannam@167
|
21 /* This file was automatically generated --- DO NOT EDIT */
|
cannam@167
|
22 /* Generated on Thu May 24 08:07:13 EDT 2018 */
|
cannam@167
|
23
|
cannam@167
|
24 #include "rdft/codelet-rdft.h"
|
cannam@167
|
25
|
cannam@167
|
26 #if defined(ARCH_PREFERS_FMA) || defined(ISA_EXTENSION_PREFERS_FMA)
|
cannam@167
|
27
|
cannam@167
|
28 /* Generated by: ../../../genfft/gen_hc2cdft.native -fma -compact -variables 4 -pipeline-latency 4 -n 20 -dit -name hc2cfdft_20 -include rdft/scalar/hc2cf.h */
|
cannam@167
|
29
|
cannam@167
|
30 /*
|
cannam@167
|
31 * This function contains 286 FP additions, 188 FP multiplications,
|
cannam@167
|
32 * (or, 176 additions, 78 multiplications, 110 fused multiply/add),
|
cannam@167
|
33 * 153 stack variables, 5 constants, and 80 memory accesses
|
cannam@167
|
34 */
|
cannam@167
|
35 #include "rdft/scalar/hc2cf.h"
|
cannam@167
|
36
|
cannam@167
|
37 static void hc2cfdft_20(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms)
|
cannam@167
|
38 {
|
cannam@167
|
39 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
|
cannam@167
|
40 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
|
cannam@167
|
41 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
|
cannam@167
|
42 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
|
cannam@167
|
43 DK(KP618033988, +0.618033988749894848204586834365638117720309180);
|
cannam@167
|
44 {
|
cannam@167
|
45 INT m;
|
cannam@167
|
46 for (m = mb, W = W + ((mb - 1) * 38); m < me; m = m + 1, Rp = Rp + ms, Ip = Ip + ms, Rm = Rm - ms, Im = Im - ms, W = W + 38, MAKE_VOLATILE_STRIDE(80, rs)) {
|
cannam@167
|
47 E T2E, T4W, T3v, T4k, T2M, T4V, T3w, T4j, T2p, T2T, T5a, T5A, T3o, T3D, T4b;
|
cannam@167
|
48 E T4B, T1Y, T2S, T57, T5z, T3h, T3C, T44, T4A, TH, T2P, T50, T5x, T32, T3z;
|
cannam@167
|
49 E T3P, T4D, T1o, T2Q, T53, T5w, T39, T3A, T3W, T4E;
|
cannam@167
|
50 {
|
cannam@167
|
51 E T9, T1V, Tu, T2w, T1, T5, T6, T2Y, T1R, T1T, T1U, T40, T10, T2F, TE;
|
cannam@167
|
52 E T2C, T1y, T2m, T4g, TX, T33, TS, TW, Tw, TA, TB, T3L, T2y, T2A, T2B;
|
cannam@167
|
53 E T3t, T1q, T1u, T1v, T3d, T2i, T2k, T2l, T48, Tm, Tq, Tr, T3J, T2s, T2u;
|
cannam@167
|
54 E T2v, T3r, T1g, T20, T1l, T23, T1h, T3S, T21, T3k, TL, T2H, TQ, T2K, TM;
|
cannam@167
|
55 E T35, T2I, T4h, T1I, T1D, T2g, T2f, T46, T2c, T2e, T1E, T3b, T16, T1b, T29;
|
cannam@167
|
56 E T28, T3i, T25, T27, T17, T3Q, Tj, Te, T1P, T1O, T3Y, T1L, T1N, Tf, T2W;
|
cannam@167
|
57 E T2x, T2D;
|
cannam@167
|
58 {
|
cannam@167
|
59 E T7, T8, Ts, Tt;
|
cannam@167
|
60 T7 = Rp[WS(rs, 9)];
|
cannam@167
|
61 T8 = Rm[WS(rs, 9)];
|
cannam@167
|
62 T9 = T7 - T8;
|
cannam@167
|
63 T1V = T7 + T8;
|
cannam@167
|
64 Ts = Rp[WS(rs, 2)];
|
cannam@167
|
65 Tt = Rm[WS(rs, 2)];
|
cannam@167
|
66 Tu = Ts + Tt;
|
cannam@167
|
67 T2w = Ts - Tt;
|
cannam@167
|
68 {
|
cannam@167
|
69 E T4, T1S, T2, T3;
|
cannam@167
|
70 T2 = Ip[WS(rs, 9)];
|
cannam@167
|
71 T3 = Im[WS(rs, 9)];
|
cannam@167
|
72 T4 = T2 + T3;
|
cannam@167
|
73 T1S = T2 - T3;
|
cannam@167
|
74 T1 = W[36];
|
cannam@167
|
75 T5 = T1 * T4;
|
cannam@167
|
76 T6 = W[37];
|
cannam@167
|
77 T2Y = T6 * T4;
|
cannam@167
|
78 T1R = W[34];
|
cannam@167
|
79 T1T = T1R * T1S;
|
cannam@167
|
80 T1U = W[35];
|
cannam@167
|
81 T40 = T1U * T1S;
|
cannam@167
|
82 }
|
cannam@167
|
83 }
|
cannam@167
|
84 {
|
cannam@167
|
85 E TY, TZ, TC, TD;
|
cannam@167
|
86 TY = Ip[0];
|
cannam@167
|
87 TZ = Im[0];
|
cannam@167
|
88 T10 = TY + TZ;
|
cannam@167
|
89 T2F = TY - TZ;
|
cannam@167
|
90 TC = Rp[WS(rs, 7)];
|
cannam@167
|
91 TD = Rm[WS(rs, 7)];
|
cannam@167
|
92 TE = TC + TD;
|
cannam@167
|
93 T2C = TC - TD;
|
cannam@167
|
94 }
|
cannam@167
|
95 {
|
cannam@167
|
96 E T1w, T1x, TT, TU, TV;
|
cannam@167
|
97 T1w = Rp[WS(rs, 1)];
|
cannam@167
|
98 T1x = Rm[WS(rs, 1)];
|
cannam@167
|
99 T1y = T1w - T1x;
|
cannam@167
|
100 T2m = T1w + T1x;
|
cannam@167
|
101 TT = Rm[0];
|
cannam@167
|
102 TU = Rp[0];
|
cannam@167
|
103 TV = TT - TU;
|
cannam@167
|
104 T4g = TU + TT;
|
cannam@167
|
105 TX = W[0];
|
cannam@167
|
106 T33 = TX * TV;
|
cannam@167
|
107 TS = W[1];
|
cannam@167
|
108 TW = TS * TV;
|
cannam@167
|
109 }
|
cannam@167
|
110 {
|
cannam@167
|
111 E T1d, T1Z, TI, T2G;
|
cannam@167
|
112 {
|
cannam@167
|
113 E Tz, T2z, Tx, Ty;
|
cannam@167
|
114 Tx = Ip[WS(rs, 7)];
|
cannam@167
|
115 Ty = Im[WS(rs, 7)];
|
cannam@167
|
116 Tz = Tx - Ty;
|
cannam@167
|
117 T2z = Tx + Ty;
|
cannam@167
|
118 Tw = W[26];
|
cannam@167
|
119 TA = Tw * Tz;
|
cannam@167
|
120 TB = W[27];
|
cannam@167
|
121 T3L = TB * Tz;
|
cannam@167
|
122 T2y = W[28];
|
cannam@167
|
123 T2A = T2y * T2z;
|
cannam@167
|
124 T2B = W[29];
|
cannam@167
|
125 T3t = T2B * T2z;
|
cannam@167
|
126 }
|
cannam@167
|
127 {
|
cannam@167
|
128 E T1t, T2j, T1r, T1s;
|
cannam@167
|
129 T1r = Ip[WS(rs, 1)];
|
cannam@167
|
130 T1s = Im[WS(rs, 1)];
|
cannam@167
|
131 T1t = T1r + T1s;
|
cannam@167
|
132 T2j = T1r - T1s;
|
cannam@167
|
133 T1q = W[4];
|
cannam@167
|
134 T1u = T1q * T1t;
|
cannam@167
|
135 T1v = W[5];
|
cannam@167
|
136 T3d = T1v * T1t;
|
cannam@167
|
137 T2i = W[2];
|
cannam@167
|
138 T2k = T2i * T2j;
|
cannam@167
|
139 T2l = W[3];
|
cannam@167
|
140 T48 = T2l * T2j;
|
cannam@167
|
141 }
|
cannam@167
|
142 {
|
cannam@167
|
143 E Tp, T2t, Tn, To;
|
cannam@167
|
144 Tn = Ip[WS(rs, 2)];
|
cannam@167
|
145 To = Im[WS(rs, 2)];
|
cannam@167
|
146 Tp = Tn - To;
|
cannam@167
|
147 T2t = Tn + To;
|
cannam@167
|
148 Tm = W[6];
|
cannam@167
|
149 Tq = Tm * Tp;
|
cannam@167
|
150 Tr = W[7];
|
cannam@167
|
151 T3J = Tr * Tp;
|
cannam@167
|
152 T2s = W[8];
|
cannam@167
|
153 T2u = T2s * T2t;
|
cannam@167
|
154 T2v = W[9];
|
cannam@167
|
155 T3r = T2v * T2t;
|
cannam@167
|
156 }
|
cannam@167
|
157 {
|
cannam@167
|
158 E T1e, T1f, T1j, T1k;
|
cannam@167
|
159 T1e = Ip[WS(rs, 3)];
|
cannam@167
|
160 T1f = Im[WS(rs, 3)];
|
cannam@167
|
161 T1g = T1e - T1f;
|
cannam@167
|
162 T20 = T1e + T1f;
|
cannam@167
|
163 T1j = Rp[WS(rs, 3)];
|
cannam@167
|
164 T1k = Rm[WS(rs, 3)];
|
cannam@167
|
165 T1l = T1j + T1k;
|
cannam@167
|
166 T23 = T1j - T1k;
|
cannam@167
|
167 }
|
cannam@167
|
168 T1d = W[10];
|
cannam@167
|
169 T1h = T1d * T1g;
|
cannam@167
|
170 T3S = T1d * T1l;
|
cannam@167
|
171 T1Z = W[12];
|
cannam@167
|
172 T21 = T1Z * T20;
|
cannam@167
|
173 T3k = T1Z * T23;
|
cannam@167
|
174 {
|
cannam@167
|
175 E TJ, TK, TO, TP;
|
cannam@167
|
176 TJ = Ip[WS(rs, 5)];
|
cannam@167
|
177 TK = Im[WS(rs, 5)];
|
cannam@167
|
178 TL = TJ + TK;
|
cannam@167
|
179 T2H = TJ - TK;
|
cannam@167
|
180 TO = Rp[WS(rs, 5)];
|
cannam@167
|
181 TP = Rm[WS(rs, 5)];
|
cannam@167
|
182 TQ = TO - TP;
|
cannam@167
|
183 T2K = TO + TP;
|
cannam@167
|
184 }
|
cannam@167
|
185 TI = W[20];
|
cannam@167
|
186 TM = TI * TL;
|
cannam@167
|
187 T35 = TI * TQ;
|
cannam@167
|
188 T2G = W[18];
|
cannam@167
|
189 T2I = T2G * T2H;
|
cannam@167
|
190 T4h = T2G * T2K;
|
cannam@167
|
191 {
|
cannam@167
|
192 E T1G, T1H, T2d, T1B, T1C, T1A;
|
cannam@167
|
193 T1G = Rm[WS(rs, 6)];
|
cannam@167
|
194 T1H = Rp[WS(rs, 6)];
|
cannam@167
|
195 T1I = T1G - T1H;
|
cannam@167
|
196 T1B = Ip[WS(rs, 6)];
|
cannam@167
|
197 T1C = Im[WS(rs, 6)];
|
cannam@167
|
198 T1D = T1B + T1C;
|
cannam@167
|
199 T2d = T1B - T1C;
|
cannam@167
|
200 T2g = T1H + T1G;
|
cannam@167
|
201 T2f = W[23];
|
cannam@167
|
202 T46 = T2f * T2d;
|
cannam@167
|
203 T2c = W[22];
|
cannam@167
|
204 T2e = T2c * T2d;
|
cannam@167
|
205 T1A = W[24];
|
cannam@167
|
206 T1E = T1A * T1D;
|
cannam@167
|
207 T3b = T1A * T1I;
|
cannam@167
|
208 }
|
cannam@167
|
209 {
|
cannam@167
|
210 E T14, T15, T26, T19, T1a, T13;
|
cannam@167
|
211 T14 = Ip[WS(rs, 8)];
|
cannam@167
|
212 T15 = Im[WS(rs, 8)];
|
cannam@167
|
213 T16 = T14 - T15;
|
cannam@167
|
214 T19 = Rp[WS(rs, 8)];
|
cannam@167
|
215 T1a = Rm[WS(rs, 8)];
|
cannam@167
|
216 T1b = T19 + T1a;
|
cannam@167
|
217 T26 = T1a - T19;
|
cannam@167
|
218 T29 = T14 + T15;
|
cannam@167
|
219 T28 = W[32];
|
cannam@167
|
220 T3i = T28 * T26;
|
cannam@167
|
221 T25 = W[33];
|
cannam@167
|
222 T27 = T25 * T26;
|
cannam@167
|
223 T13 = W[30];
|
cannam@167
|
224 T17 = T13 * T16;
|
cannam@167
|
225 T3Q = T13 * T1b;
|
cannam@167
|
226 }
|
cannam@167
|
227 {
|
cannam@167
|
228 E Th, Ti, T1M, Tc, Td, Tb;
|
cannam@167
|
229 Th = Rm[WS(rs, 4)];
|
cannam@167
|
230 Ti = Rp[WS(rs, 4)];
|
cannam@167
|
231 Tj = Th - Ti;
|
cannam@167
|
232 Tc = Ip[WS(rs, 4)];
|
cannam@167
|
233 Td = Im[WS(rs, 4)];
|
cannam@167
|
234 Te = Tc + Td;
|
cannam@167
|
235 T1M = Tc - Td;
|
cannam@167
|
236 T1P = Ti + Th;
|
cannam@167
|
237 T1O = W[15];
|
cannam@167
|
238 T3Y = T1O * T1M;
|
cannam@167
|
239 T1L = W[14];
|
cannam@167
|
240 T1N = T1L * T1M;
|
cannam@167
|
241 Tb = W[16];
|
cannam@167
|
242 Tf = Tb * Te;
|
cannam@167
|
243 T2W = Tb * Tj;
|
cannam@167
|
244 }
|
cannam@167
|
245 }
|
cannam@167
|
246 T2x = FNMS(T2v, T2w, T2u);
|
cannam@167
|
247 T2D = FNMS(T2B, T2C, T2A);
|
cannam@167
|
248 T2E = T2x - T2D;
|
cannam@167
|
249 T4W = T2x + T2D;
|
cannam@167
|
250 {
|
cannam@167
|
251 E T3s, T3u, T2L, T4i, T2J;
|
cannam@167
|
252 T3s = FMA(T2s, T2w, T3r);
|
cannam@167
|
253 T3u = FMA(T2y, T2C, T3t);
|
cannam@167
|
254 T3v = T3s + T3u;
|
cannam@167
|
255 T4k = T3u - T3s;
|
cannam@167
|
256 T2J = W[19];
|
cannam@167
|
257 T2L = FNMS(T2J, T2K, T2I);
|
cannam@167
|
258 T4i = FMA(T2J, T2H, T4h);
|
cannam@167
|
259 T2M = T2F - T2L;
|
cannam@167
|
260 T4V = T4g + T4i;
|
cannam@167
|
261 T3w = T2L + T2F;
|
cannam@167
|
262 T4j = T4g - T4i;
|
cannam@167
|
263 }
|
cannam@167
|
264 {
|
cannam@167
|
265 E T2a, T3j, T24, T3l, T2o, T3n, T4a, T59, T22;
|
cannam@167
|
266 T2a = FMA(T28, T29, T27);
|
cannam@167
|
267 T3j = FNMS(T25, T29, T3i);
|
cannam@167
|
268 T22 = W[13];
|
cannam@167
|
269 T24 = FNMS(T22, T23, T21);
|
cannam@167
|
270 T3l = FMA(T22, T20, T3k);
|
cannam@167
|
271 {
|
cannam@167
|
272 E T2h, T2n, T47, T49;
|
cannam@167
|
273 T2h = FNMS(T2f, T2g, T2e);
|
cannam@167
|
274 T2n = FNMS(T2l, T2m, T2k);
|
cannam@167
|
275 T2o = T2h - T2n;
|
cannam@167
|
276 T3n = T2h + T2n;
|
cannam@167
|
277 T47 = FMA(T2c, T2g, T46);
|
cannam@167
|
278 T49 = FMA(T2i, T2m, T48);
|
cannam@167
|
279 T4a = T47 - T49;
|
cannam@167
|
280 T59 = T47 + T49;
|
cannam@167
|
281 }
|
cannam@167
|
282 {
|
cannam@167
|
283 E T2b, T58, T3m, T45;
|
cannam@167
|
284 T2b = T24 - T2a;
|
cannam@167
|
285 T2p = T2b - T2o;
|
cannam@167
|
286 T2T = T2b + T2o;
|
cannam@167
|
287 T58 = T2a + T24;
|
cannam@167
|
288 T5a = T58 + T59;
|
cannam@167
|
289 T5A = T59 - T58;
|
cannam@167
|
290 T3m = T3j - T3l;
|
cannam@167
|
291 T3o = T3m - T3n;
|
cannam@167
|
292 T3D = T3m + T3n;
|
cannam@167
|
293 T45 = T3j + T3l;
|
cannam@167
|
294 T4b = T45 + T4a;
|
cannam@167
|
295 T4B = T4a - T45;
|
cannam@167
|
296 }
|
cannam@167
|
297 }
|
cannam@167
|
298 {
|
cannam@167
|
299 E T1z, T3e, T1J, T3c, T1X, T3g, T42, T55, T1F;
|
cannam@167
|
300 T1z = FNMS(T1v, T1y, T1u);
|
cannam@167
|
301 T3e = FMA(T1q, T1y, T3d);
|
cannam@167
|
302 T1F = W[25];
|
cannam@167
|
303 T1J = FMA(T1F, T1I, T1E);
|
cannam@167
|
304 T3c = FNMS(T1F, T1D, T3b);
|
cannam@167
|
305 {
|
cannam@167
|
306 E T1Q, T1W, T3Z, T41;
|
cannam@167
|
307 T1Q = FNMS(T1O, T1P, T1N);
|
cannam@167
|
308 T1W = FNMS(T1U, T1V, T1T);
|
cannam@167
|
309 T1X = T1Q - T1W;
|
cannam@167
|
310 T3g = T1Q + T1W;
|
cannam@167
|
311 T3Z = FMA(T1L, T1P, T3Y);
|
cannam@167
|
312 T41 = FMA(T1R, T1V, T40);
|
cannam@167
|
313 T42 = T3Z - T41;
|
cannam@167
|
314 T55 = T3Z + T41;
|
cannam@167
|
315 }
|
cannam@167
|
316 {
|
cannam@167
|
317 E T1K, T56, T3f, T43;
|
cannam@167
|
318 T1K = T1z - T1J;
|
cannam@167
|
319 T1Y = T1K - T1X;
|
cannam@167
|
320 T2S = T1X + T1K;
|
cannam@167
|
321 T56 = T1J + T1z;
|
cannam@167
|
322 T57 = T55 + T56;
|
cannam@167
|
323 T5z = T55 - T56;
|
cannam@167
|
324 T3f = T3c - T3e;
|
cannam@167
|
325 T3h = T3f - T3g;
|
cannam@167
|
326 T3C = T3g + T3f;
|
cannam@167
|
327 T43 = T3c + T3e;
|
cannam@167
|
328 T44 = T42 + T43;
|
cannam@167
|
329 T4A = T42 - T43;
|
cannam@167
|
330 }
|
cannam@167
|
331 }
|
cannam@167
|
332 {
|
cannam@167
|
333 E Ta, T2Z, Tk, T2X, TG, T31, T3N, T4Y, Tg;
|
cannam@167
|
334 Ta = FNMS(T6, T9, T5);
|
cannam@167
|
335 T2Z = FMA(T1, T9, T2Y);
|
cannam@167
|
336 Tg = W[17];
|
cannam@167
|
337 Tk = FMA(Tg, Tj, Tf);
|
cannam@167
|
338 T2X = FNMS(Tg, Te, T2W);
|
cannam@167
|
339 {
|
cannam@167
|
340 E Tv, TF, T3K, T3M;
|
cannam@167
|
341 Tv = FNMS(Tr, Tu, Tq);
|
cannam@167
|
342 TF = FNMS(TB, TE, TA);
|
cannam@167
|
343 TG = Tv - TF;
|
cannam@167
|
344 T31 = Tv + TF;
|
cannam@167
|
345 T3K = FMA(Tm, Tu, T3J);
|
cannam@167
|
346 T3M = FMA(Tw, TE, T3L);
|
cannam@167
|
347 T3N = T3K - T3M;
|
cannam@167
|
348 T4Y = T3K + T3M;
|
cannam@167
|
349 }
|
cannam@167
|
350 {
|
cannam@167
|
351 E Tl, T4Z, T30, T3O;
|
cannam@167
|
352 Tl = Ta - Tk;
|
cannam@167
|
353 TH = Tl - TG;
|
cannam@167
|
354 T2P = TG + Tl;
|
cannam@167
|
355 T4Z = Tk + Ta;
|
cannam@167
|
356 T50 = T4Y + T4Z;
|
cannam@167
|
357 T5x = T4Y - T4Z;
|
cannam@167
|
358 T30 = T2X - T2Z;
|
cannam@167
|
359 T32 = T30 - T31;
|
cannam@167
|
360 T3z = T31 + T30;
|
cannam@167
|
361 T3O = T2X + T2Z;
|
cannam@167
|
362 T3P = T3N + T3O;
|
cannam@167
|
363 T4D = T3N - T3O;
|
cannam@167
|
364 }
|
cannam@167
|
365 }
|
cannam@167
|
366 {
|
cannam@167
|
367 E T11, T34, TR, T36, T1c, T3R, T1m, T3T, TN, T18, T1i;
|
cannam@167
|
368 T11 = FMA(TX, T10, TW);
|
cannam@167
|
369 T34 = FNMS(TS, T10, T33);
|
cannam@167
|
370 TN = W[21];
|
cannam@167
|
371 TR = FNMS(TN, TQ, TM);
|
cannam@167
|
372 T36 = FMA(TN, TL, T35);
|
cannam@167
|
373 T18 = W[31];
|
cannam@167
|
374 T1c = FNMS(T18, T1b, T17);
|
cannam@167
|
375 T3R = FMA(T18, T16, T3Q);
|
cannam@167
|
376 T1i = W[11];
|
cannam@167
|
377 T1m = FNMS(T1i, T1l, T1h);
|
cannam@167
|
378 T3T = FMA(T1i, T1g, T3S);
|
cannam@167
|
379 {
|
cannam@167
|
380 E T12, T1n, T51, T52;
|
cannam@167
|
381 T12 = TR - T11;
|
cannam@167
|
382 T1n = T1c - T1m;
|
cannam@167
|
383 T1o = T12 - T1n;
|
cannam@167
|
384 T2Q = T1n + T12;
|
cannam@167
|
385 T51 = T3R + T3T;
|
cannam@167
|
386 T52 = TR + T11;
|
cannam@167
|
387 T53 = T51 + T52;
|
cannam@167
|
388 T5w = T51 - T52;
|
cannam@167
|
389 }
|
cannam@167
|
390 {
|
cannam@167
|
391 E T37, T38, T3U, T3V;
|
cannam@167
|
392 T37 = T34 - T36;
|
cannam@167
|
393 T38 = T1c + T1m;
|
cannam@167
|
394 T39 = T37 - T38;
|
cannam@167
|
395 T3A = T38 + T37;
|
cannam@167
|
396 T3U = T3R - T3T;
|
cannam@167
|
397 T3V = T36 + T34;
|
cannam@167
|
398 T3W = T3U + T3V;
|
cannam@167
|
399 T4E = T3U - T3V;
|
cannam@167
|
400 }
|
cannam@167
|
401 }
|
cannam@167
|
402 }
|
cannam@167
|
403 {
|
cannam@167
|
404 E T4G, T4I, T2N, T2r, T4x, T4y, T4H, T4z;
|
cannam@167
|
405 {
|
cannam@167
|
406 E T4C, T4F, T1p, T2q;
|
cannam@167
|
407 T4C = T4A - T4B;
|
cannam@167
|
408 T4F = T4D - T4E;
|
cannam@167
|
409 T4G = FNMS(KP618033988, T4F, T4C);
|
cannam@167
|
410 T4I = FMA(KP618033988, T4C, T4F);
|
cannam@167
|
411 T2N = T2E + T2M;
|
cannam@167
|
412 T1p = TH + T1o;
|
cannam@167
|
413 T2q = T1Y + T2p;
|
cannam@167
|
414 T2r = T1p + T2q;
|
cannam@167
|
415 T4x = FMA(KP250000000, T2r, T2N);
|
cannam@167
|
416 T4y = T1p - T2q;
|
cannam@167
|
417 }
|
cannam@167
|
418 Im[WS(rs, 4)] = KP500000000 * (T2r - T2N);
|
cannam@167
|
419 T4H = FNMS(KP559016994, T4y, T4x);
|
cannam@167
|
420 Im[0] = -(KP500000000 * (FMA(KP951056516, T4I, T4H)));
|
cannam@167
|
421 Im[WS(rs, 8)] = -(KP500000000 * (FNMS(KP951056516, T4I, T4H)));
|
cannam@167
|
422 T4z = FMA(KP559016994, T4y, T4x);
|
cannam@167
|
423 Ip[WS(rs, 3)] = KP500000000 * (FNMS(KP951056516, T4G, T4z));
|
cannam@167
|
424 Ip[WS(rs, 7)] = KP500000000 * (FMA(KP951056516, T4G, T4z));
|
cannam@167
|
425 }
|
cannam@167
|
426 {
|
cannam@167
|
427 E T4S, T4U, T4J, T4M, T4N, T4O, T4T, T4P;
|
cannam@167
|
428 {
|
cannam@167
|
429 E T4Q, T4R, T4K, T4L;
|
cannam@167
|
430 T4Q = T2p - T1Y;
|
cannam@167
|
431 T4R = T1o - TH;
|
cannam@167
|
432 T4S = FNMS(KP618033988, T4R, T4Q);
|
cannam@167
|
433 T4U = FMA(KP618033988, T4Q, T4R);
|
cannam@167
|
434 T4J = T4j - T4k;
|
cannam@167
|
435 T4K = T4D + T4E;
|
cannam@167
|
436 T4L = T4A + T4B;
|
cannam@167
|
437 T4M = T4K + T4L;
|
cannam@167
|
438 T4N = FNMS(KP250000000, T4M, T4J);
|
cannam@167
|
439 T4O = T4K - T4L;
|
cannam@167
|
440 }
|
cannam@167
|
441 Rm[WS(rs, 4)] = KP500000000 * (T4J + T4M);
|
cannam@167
|
442 T4T = FMA(KP559016994, T4O, T4N);
|
cannam@167
|
443 Rm[WS(rs, 8)] = KP500000000 * (FMA(KP951056516, T4U, T4T));
|
cannam@167
|
444 Rm[0] = KP500000000 * (FNMS(KP951056516, T4U, T4T));
|
cannam@167
|
445 T4P = FNMS(KP559016994, T4O, T4N);
|
cannam@167
|
446 Rp[WS(rs, 3)] = KP500000000 * (FMA(KP951056516, T4S, T4P));
|
cannam@167
|
447 Rp[WS(rs, 7)] = KP500000000 * (FNMS(KP951056516, T4S, T4P));
|
cannam@167
|
448 }
|
cannam@167
|
449 {
|
cannam@167
|
450 E T4d, T4f, T2O, T2V, T3G, T3H, T4e, T3I;
|
cannam@167
|
451 {
|
cannam@167
|
452 E T3X, T4c, T2R, T2U;
|
cannam@167
|
453 T3X = T3P - T3W;
|
cannam@167
|
454 T4c = T44 - T4b;
|
cannam@167
|
455 T4d = FMA(KP618033988, T4c, T3X);
|
cannam@167
|
456 T4f = FNMS(KP618033988, T3X, T4c);
|
cannam@167
|
457 T2O = T2M - T2E;
|
cannam@167
|
458 T2R = T2P + T2Q;
|
cannam@167
|
459 T2U = T2S + T2T;
|
cannam@167
|
460 T2V = T2R + T2U;
|
cannam@167
|
461 T3G = FNMS(KP250000000, T2V, T2O);
|
cannam@167
|
462 T3H = T2R - T2U;
|
cannam@167
|
463 }
|
cannam@167
|
464 Ip[WS(rs, 5)] = KP500000000 * (T2O + T2V);
|
cannam@167
|
465 T4e = FNMS(KP559016994, T3H, T3G);
|
cannam@167
|
466 Im[WS(rs, 2)] = -(KP500000000 * (FMA(KP951056516, T4f, T4e)));
|
cannam@167
|
467 Im[WS(rs, 6)] = -(KP500000000 * (FNMS(KP951056516, T4f, T4e)));
|
cannam@167
|
468 T3I = FMA(KP559016994, T3H, T3G);
|
cannam@167
|
469 Ip[WS(rs, 1)] = KP500000000 * (FNMS(KP951056516, T4d, T3I));
|
cannam@167
|
470 Ip[WS(rs, 9)] = KP500000000 * (FMA(KP951056516, T4d, T3I));
|
cannam@167
|
471 }
|
cannam@167
|
472 {
|
cannam@167
|
473 E T4u, T4w, T4l, T4o, T4p, T4q, T4v, T4r;
|
cannam@167
|
474 {
|
cannam@167
|
475 E T4s, T4t, T4m, T4n;
|
cannam@167
|
476 T4s = T2P - T2Q;
|
cannam@167
|
477 T4t = T2S - T2T;
|
cannam@167
|
478 T4u = FMA(KP618033988, T4t, T4s);
|
cannam@167
|
479 T4w = FNMS(KP618033988, T4s, T4t);
|
cannam@167
|
480 T4l = T4j + T4k;
|
cannam@167
|
481 T4m = T3P + T3W;
|
cannam@167
|
482 T4n = T44 + T4b;
|
cannam@167
|
483 T4o = T4m + T4n;
|
cannam@167
|
484 T4p = FNMS(KP250000000, T4o, T4l);
|
cannam@167
|
485 T4q = T4m - T4n;
|
cannam@167
|
486 }
|
cannam@167
|
487 Rp[WS(rs, 5)] = KP500000000 * (T4l + T4o);
|
cannam@167
|
488 T4v = FNMS(KP559016994, T4q, T4p);
|
cannam@167
|
489 Rm[WS(rs, 6)] = KP500000000 * (FMA(KP951056516, T4w, T4v));
|
cannam@167
|
490 Rm[WS(rs, 2)] = KP500000000 * (FNMS(KP951056516, T4w, T4v));
|
cannam@167
|
491 T4r = FMA(KP559016994, T4q, T4p);
|
cannam@167
|
492 Rp[WS(rs, 1)] = KP500000000 * (FMA(KP951056516, T4u, T4r));
|
cannam@167
|
493 Rp[WS(rs, 9)] = KP500000000 * (FNMS(KP951056516, T4u, T4r));
|
cannam@167
|
494 }
|
cannam@167
|
495 {
|
cannam@167
|
496 E T5C, T5E, T3x, T3q, T5t, T5u, T5D, T5v;
|
cannam@167
|
497 {
|
cannam@167
|
498 E T5y, T5B, T3a, T3p;
|
cannam@167
|
499 T5y = T5w - T5x;
|
cannam@167
|
500 T5B = T5z - T5A;
|
cannam@167
|
501 T5C = FNMS(KP618033988, T5B, T5y);
|
cannam@167
|
502 T5E = FMA(KP618033988, T5y, T5B);
|
cannam@167
|
503 T3x = T3v + T3w;
|
cannam@167
|
504 T3a = T32 + T39;
|
cannam@167
|
505 T3p = T3h + T3o;
|
cannam@167
|
506 T3q = T3a + T3p;
|
cannam@167
|
507 T5t = FMA(KP250000000, T3q, T3x);
|
cannam@167
|
508 T5u = T3p - T3a;
|
cannam@167
|
509 }
|
cannam@167
|
510 Im[WS(rs, 9)] = KP500000000 * (T3q - T3x);
|
cannam@167
|
511 T5D = FNMS(KP559016994, T5u, T5t);
|
cannam@167
|
512 Ip[WS(rs, 2)] = KP500000000 * (FMA(KP951056516, T5E, T5D));
|
cannam@167
|
513 Im[WS(rs, 1)] = -(KP500000000 * (FNMS(KP951056516, T5E, T5D)));
|
cannam@167
|
514 T5v = FMA(KP559016994, T5u, T5t);
|
cannam@167
|
515 Ip[WS(rs, 6)] = KP500000000 * (FMA(KP951056516, T5C, T5v));
|
cannam@167
|
516 Im[WS(rs, 5)] = -(KP500000000 * (FNMS(KP951056516, T5C, T5v)));
|
cannam@167
|
517 }
|
cannam@167
|
518 {
|
cannam@167
|
519 E T5O, T5Q, T5F, T5I, T5J, T5K, T5P, T5L;
|
cannam@167
|
520 {
|
cannam@167
|
521 E T5M, T5N, T5G, T5H;
|
cannam@167
|
522 T5M = T3o - T3h;
|
cannam@167
|
523 T5N = T39 - T32;
|
cannam@167
|
524 T5O = FNMS(KP618033988, T5N, T5M);
|
cannam@167
|
525 T5Q = FMA(KP618033988, T5M, T5N);
|
cannam@167
|
526 T5F = T4V - T4W;
|
cannam@167
|
527 T5G = T5x + T5w;
|
cannam@167
|
528 T5H = T5z + T5A;
|
cannam@167
|
529 T5I = T5G + T5H;
|
cannam@167
|
530 T5J = FNMS(KP250000000, T5I, T5F);
|
cannam@167
|
531 T5K = T5G - T5H;
|
cannam@167
|
532 }
|
cannam@167
|
533 Rm[WS(rs, 9)] = KP500000000 * (T5F + T5I);
|
cannam@167
|
534 T5P = FMA(KP559016994, T5K, T5J);
|
cannam@167
|
535 Rp[WS(rs, 6)] = KP500000000 * (FMA(KP951056516, T5Q, T5P));
|
cannam@167
|
536 Rm[WS(rs, 5)] = KP500000000 * (FNMS(KP951056516, T5Q, T5P));
|
cannam@167
|
537 T5L = FNMS(KP559016994, T5K, T5J);
|
cannam@167
|
538 Rp[WS(rs, 2)] = KP500000000 * (FNMS(KP951056516, T5O, T5L));
|
cannam@167
|
539 Rm[WS(rs, 1)] = KP500000000 * (FMA(KP951056516, T5O, T5L));
|
cannam@167
|
540 }
|
cannam@167
|
541 {
|
cannam@167
|
542 E T5q, T5s, T3y, T3F, T5l, T5m, T5r, T5n;
|
cannam@167
|
543 {
|
cannam@167
|
544 E T5o, T5p, T3B, T3E;
|
cannam@167
|
545 T5o = T50 - T53;
|
cannam@167
|
546 T5p = T5a - T57;
|
cannam@167
|
547 T5q = FNMS(KP618033988, T5p, T5o);
|
cannam@167
|
548 T5s = FMA(KP618033988, T5o, T5p);
|
cannam@167
|
549 T3y = T3w - T3v;
|
cannam@167
|
550 T3B = T3z + T3A;
|
cannam@167
|
551 T3E = T3C + T3D;
|
cannam@167
|
552 T3F = T3B + T3E;
|
cannam@167
|
553 T5l = FNMS(KP250000000, T3F, T3y);
|
cannam@167
|
554 T5m = T3B - T3E;
|
cannam@167
|
555 }
|
cannam@167
|
556 Ip[0] = KP500000000 * (T3y + T3F);
|
cannam@167
|
557 T5r = FNMS(KP559016994, T5m, T5l);
|
cannam@167
|
558 Ip[WS(rs, 8)] = KP500000000 * (FMA(KP951056516, T5s, T5r));
|
cannam@167
|
559 Im[WS(rs, 7)] = -(KP500000000 * (FNMS(KP951056516, T5s, T5r)));
|
cannam@167
|
560 T5n = FMA(KP559016994, T5m, T5l);
|
cannam@167
|
561 Ip[WS(rs, 4)] = KP500000000 * (FMA(KP951056516, T5q, T5n));
|
cannam@167
|
562 Im[WS(rs, 3)] = -(KP500000000 * (FNMS(KP951056516, T5q, T5n)));
|
cannam@167
|
563 }
|
cannam@167
|
564 {
|
cannam@167
|
565 E T5i, T5k, T4X, T5c, T5d, T5e, T5j, T5f;
|
cannam@167
|
566 {
|
cannam@167
|
567 E T5g, T5h, T54, T5b;
|
cannam@167
|
568 T5g = T3z - T3A;
|
cannam@167
|
569 T5h = T3C - T3D;
|
cannam@167
|
570 T5i = FMA(KP618033988, T5h, T5g);
|
cannam@167
|
571 T5k = FNMS(KP618033988, T5g, T5h);
|
cannam@167
|
572 T4X = T4V + T4W;
|
cannam@167
|
573 T54 = T50 + T53;
|
cannam@167
|
574 T5b = T57 + T5a;
|
cannam@167
|
575 T5c = T54 + T5b;
|
cannam@167
|
576 T5d = FNMS(KP250000000, T5c, T4X);
|
cannam@167
|
577 T5e = T54 - T5b;
|
cannam@167
|
578 }
|
cannam@167
|
579 Rp[0] = KP500000000 * (T4X + T5c);
|
cannam@167
|
580 T5j = FNMS(KP559016994, T5e, T5d);
|
cannam@167
|
581 Rp[WS(rs, 8)] = KP500000000 * (FMA(KP951056516, T5k, T5j));
|
cannam@167
|
582 Rm[WS(rs, 7)] = KP500000000 * (FNMS(KP951056516, T5k, T5j));
|
cannam@167
|
583 T5f = FMA(KP559016994, T5e, T5d);
|
cannam@167
|
584 Rp[WS(rs, 4)] = KP500000000 * (FNMS(KP951056516, T5i, T5f));
|
cannam@167
|
585 Rm[WS(rs, 3)] = KP500000000 * (FMA(KP951056516, T5i, T5f));
|
cannam@167
|
586 }
|
cannam@167
|
587 }
|
cannam@167
|
588 }
|
cannam@167
|
589 }
|
cannam@167
|
590
|
cannam@167
|
591 static const tw_instr twinstr[] = {
|
cannam@167
|
592 {TW_FULL, 1, 20},
|
cannam@167
|
593 {TW_NEXT, 1, 0}
|
cannam@167
|
594 };
|
cannam@167
|
595
|
cannam@167
|
596 static const hc2c_desc desc = { 20, "hc2cfdft_20", twinstr, &GENUS, {176, 78, 110, 0} };
|
cannam@167
|
597
|
cannam@167
|
598 void X(codelet_hc2cfdft_20) (planner *p) {
|
cannam@167
|
599 X(khc2c_register) (p, hc2cfdft_20, &desc, HC2C_VIA_DFT);
|
cannam@167
|
600 }
|
cannam@167
|
601 #else
|
cannam@167
|
602
|
cannam@167
|
603 /* Generated by: ../../../genfft/gen_hc2cdft.native -compact -variables 4 -pipeline-latency 4 -n 20 -dit -name hc2cfdft_20 -include rdft/scalar/hc2cf.h */
|
cannam@167
|
604
|
cannam@167
|
605 /*
|
cannam@167
|
606 * This function contains 286 FP additions, 140 FP multiplications,
|
cannam@167
|
607 * (or, 224 additions, 78 multiplications, 62 fused multiply/add),
|
cannam@167
|
608 * 98 stack variables, 5 constants, and 80 memory accesses
|
cannam@167
|
609 */
|
cannam@167
|
610 #include "rdft/scalar/hc2cf.h"
|
cannam@167
|
611
|
cannam@167
|
612 static void hc2cfdft_20(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms)
|
cannam@167
|
613 {
|
cannam@167
|
614 DK(KP125000000, +0.125000000000000000000000000000000000000000000);
|
cannam@167
|
615 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
|
cannam@167
|
616 DK(KP279508497, +0.279508497187473712051146708591409529430077295);
|
cannam@167
|
617 DK(KP293892626, +0.293892626146236564584352977319536384298826219);
|
cannam@167
|
618 DK(KP475528258, +0.475528258147576786058219666689691071702849317);
|
cannam@167
|
619 {
|
cannam@167
|
620 INT m;
|
cannam@167
|
621 for (m = mb, W = W + ((mb - 1) * 38); m < me; m = m + 1, Rp = Rp + ms, Ip = Ip + ms, Rm = Rm - ms, Im = Im - ms, W = W + 38, MAKE_VOLATILE_STRIDE(80, rs)) {
|
cannam@167
|
622 E T12, T2w, T4o, T4V, T2H, T3a, T4y, T4Y, T1z, T2v, T25, T2y, T2s, T2z, T4v;
|
cannam@167
|
623 E T4X, T4r, T4U, T3A, T3Z, T2X, T37, T3k, T41, T2M, T39, T3v, T3Y, T2S, T36;
|
cannam@167
|
624 E T3p, T42, Td, T4G, T33, T3N, Tw, T4H, T32, T3O;
|
cannam@167
|
625 {
|
cannam@167
|
626 E T3, T3L, T1x, T2V, Th, Tl, TC, T3g, Tq, Tu, TH, T3h, T7, Tb, T1q;
|
cannam@167
|
627 E T2U, TR, T2P, T1F, T3r, T23, T2K, T2f, T3y, T1k, T3m, T2q, T2E, T10, T2Q;
|
cannam@167
|
628 E T1K, T3s, T1U, T2J, T2a, T3x, T1b, T3l, T2l, T2D;
|
cannam@167
|
629 {
|
cannam@167
|
630 E T1, T2, T1s, T1u, T1v, T1w, T1r, T1t;
|
cannam@167
|
631 T1 = Ip[0];
|
cannam@167
|
632 T2 = Im[0];
|
cannam@167
|
633 T1s = T1 + T2;
|
cannam@167
|
634 T1u = Rp[0];
|
cannam@167
|
635 T1v = Rm[0];
|
cannam@167
|
636 T1w = T1u - T1v;
|
cannam@167
|
637 T3 = T1 - T2;
|
cannam@167
|
638 T3L = T1u + T1v;
|
cannam@167
|
639 T1r = W[0];
|
cannam@167
|
640 T1t = W[1];
|
cannam@167
|
641 T1x = FNMS(T1t, T1w, T1r * T1s);
|
cannam@167
|
642 T2V = FMA(T1r, T1w, T1t * T1s);
|
cannam@167
|
643 }
|
cannam@167
|
644 {
|
cannam@167
|
645 E Tf, Tg, Tz, Tj, Tk, TB, Ty, TA;
|
cannam@167
|
646 Tf = Ip[WS(rs, 2)];
|
cannam@167
|
647 Tg = Im[WS(rs, 2)];
|
cannam@167
|
648 Tz = Tf - Tg;
|
cannam@167
|
649 Tj = Rp[WS(rs, 2)];
|
cannam@167
|
650 Tk = Rm[WS(rs, 2)];
|
cannam@167
|
651 TB = Tj + Tk;
|
cannam@167
|
652 Th = Tf + Tg;
|
cannam@167
|
653 Tl = Tj - Tk;
|
cannam@167
|
654 Ty = W[6];
|
cannam@167
|
655 TA = W[7];
|
cannam@167
|
656 TC = FNMS(TA, TB, Ty * Tz);
|
cannam@167
|
657 T3g = FMA(TA, Tz, Ty * TB);
|
cannam@167
|
658 }
|
cannam@167
|
659 {
|
cannam@167
|
660 E To, Tp, TE, Ts, Tt, TG, TD, TF;
|
cannam@167
|
661 To = Ip[WS(rs, 7)];
|
cannam@167
|
662 Tp = Im[WS(rs, 7)];
|
cannam@167
|
663 TE = To - Tp;
|
cannam@167
|
664 Ts = Rp[WS(rs, 7)];
|
cannam@167
|
665 Tt = Rm[WS(rs, 7)];
|
cannam@167
|
666 TG = Ts + Tt;
|
cannam@167
|
667 Tq = To + Tp;
|
cannam@167
|
668 Tu = Ts - Tt;
|
cannam@167
|
669 TD = W[26];
|
cannam@167
|
670 TF = W[27];
|
cannam@167
|
671 TH = FNMS(TF, TG, TD * TE);
|
cannam@167
|
672 T3h = FMA(TF, TE, TD * TG);
|
cannam@167
|
673 }
|
cannam@167
|
674 {
|
cannam@167
|
675 E T5, T6, T1n, T9, Ta, T1p, T1m, T1o;
|
cannam@167
|
676 T5 = Ip[WS(rs, 5)];
|
cannam@167
|
677 T6 = Im[WS(rs, 5)];
|
cannam@167
|
678 T1n = T5 + T6;
|
cannam@167
|
679 T9 = Rp[WS(rs, 5)];
|
cannam@167
|
680 Ta = Rm[WS(rs, 5)];
|
cannam@167
|
681 T1p = T9 - Ta;
|
cannam@167
|
682 T7 = T5 - T6;
|
cannam@167
|
683 Tb = T9 + Ta;
|
cannam@167
|
684 T1m = W[20];
|
cannam@167
|
685 T1o = W[21];
|
cannam@167
|
686 T1q = FNMS(T1o, T1p, T1m * T1n);
|
cannam@167
|
687 T2U = FMA(T1m, T1p, T1o * T1n);
|
cannam@167
|
688 }
|
cannam@167
|
689 {
|
cannam@167
|
690 E TM, T1C, TQ, T1E;
|
cannam@167
|
691 {
|
cannam@167
|
692 E TK, TL, TO, TP;
|
cannam@167
|
693 TK = Ip[WS(rs, 4)];
|
cannam@167
|
694 TL = Im[WS(rs, 4)];
|
cannam@167
|
695 TM = TK + TL;
|
cannam@167
|
696 T1C = TK - TL;
|
cannam@167
|
697 TO = Rp[WS(rs, 4)];
|
cannam@167
|
698 TP = Rm[WS(rs, 4)];
|
cannam@167
|
699 TQ = TO - TP;
|
cannam@167
|
700 T1E = TO + TP;
|
cannam@167
|
701 }
|
cannam@167
|
702 {
|
cannam@167
|
703 E TJ, TN, T1B, T1D;
|
cannam@167
|
704 TJ = W[16];
|
cannam@167
|
705 TN = W[17];
|
cannam@167
|
706 TR = FNMS(TN, TQ, TJ * TM);
|
cannam@167
|
707 T2P = FMA(TN, TM, TJ * TQ);
|
cannam@167
|
708 T1B = W[14];
|
cannam@167
|
709 T1D = W[15];
|
cannam@167
|
710 T1F = FNMS(T1D, T1E, T1B * T1C);
|
cannam@167
|
711 T3r = FMA(T1D, T1C, T1B * T1E);
|
cannam@167
|
712 }
|
cannam@167
|
713 }
|
cannam@167
|
714 {
|
cannam@167
|
715 E T1Y, T2c, T22, T2e;
|
cannam@167
|
716 {
|
cannam@167
|
717 E T1W, T1X, T20, T21;
|
cannam@167
|
718 T1W = Ip[WS(rs, 1)];
|
cannam@167
|
719 T1X = Im[WS(rs, 1)];
|
cannam@167
|
720 T1Y = T1W + T1X;
|
cannam@167
|
721 T2c = T1W - T1X;
|
cannam@167
|
722 T20 = Rp[WS(rs, 1)];
|
cannam@167
|
723 T21 = Rm[WS(rs, 1)];
|
cannam@167
|
724 T22 = T20 - T21;
|
cannam@167
|
725 T2e = T20 + T21;
|
cannam@167
|
726 }
|
cannam@167
|
727 {
|
cannam@167
|
728 E T1V, T1Z, T2b, T2d;
|
cannam@167
|
729 T1V = W[4];
|
cannam@167
|
730 T1Z = W[5];
|
cannam@167
|
731 T23 = FNMS(T1Z, T22, T1V * T1Y);
|
cannam@167
|
732 T2K = FMA(T1Z, T1Y, T1V * T22);
|
cannam@167
|
733 T2b = W[2];
|
cannam@167
|
734 T2d = W[3];
|
cannam@167
|
735 T2f = FNMS(T2d, T2e, T2b * T2c);
|
cannam@167
|
736 T3y = FMA(T2d, T2c, T2b * T2e);
|
cannam@167
|
737 }
|
cannam@167
|
738 }
|
cannam@167
|
739 {
|
cannam@167
|
740 E T1f, T2n, T1j, T2p;
|
cannam@167
|
741 {
|
cannam@167
|
742 E T1d, T1e, T1h, T1i;
|
cannam@167
|
743 T1d = Ip[WS(rs, 3)];
|
cannam@167
|
744 T1e = Im[WS(rs, 3)];
|
cannam@167
|
745 T1f = T1d - T1e;
|
cannam@167
|
746 T2n = T1d + T1e;
|
cannam@167
|
747 T1h = Rp[WS(rs, 3)];
|
cannam@167
|
748 T1i = Rm[WS(rs, 3)];
|
cannam@167
|
749 T1j = T1h + T1i;
|
cannam@167
|
750 T2p = T1h - T1i;
|
cannam@167
|
751 }
|
cannam@167
|
752 {
|
cannam@167
|
753 E T1c, T1g, T2m, T2o;
|
cannam@167
|
754 T1c = W[10];
|
cannam@167
|
755 T1g = W[11];
|
cannam@167
|
756 T1k = FNMS(T1g, T1j, T1c * T1f);
|
cannam@167
|
757 T3m = FMA(T1c, T1j, T1g * T1f);
|
cannam@167
|
758 T2m = W[12];
|
cannam@167
|
759 T2o = W[13];
|
cannam@167
|
760 T2q = FNMS(T2o, T2p, T2m * T2n);
|
cannam@167
|
761 T2E = FMA(T2m, T2p, T2o * T2n);
|
cannam@167
|
762 }
|
cannam@167
|
763 }
|
cannam@167
|
764 {
|
cannam@167
|
765 E TV, T1H, TZ, T1J;
|
cannam@167
|
766 {
|
cannam@167
|
767 E TT, TU, TX, TY;
|
cannam@167
|
768 TT = Ip[WS(rs, 9)];
|
cannam@167
|
769 TU = Im[WS(rs, 9)];
|
cannam@167
|
770 TV = TT + TU;
|
cannam@167
|
771 T1H = TT - TU;
|
cannam@167
|
772 TX = Rp[WS(rs, 9)];
|
cannam@167
|
773 TY = Rm[WS(rs, 9)];
|
cannam@167
|
774 TZ = TX - TY;
|
cannam@167
|
775 T1J = TX + TY;
|
cannam@167
|
776 }
|
cannam@167
|
777 {
|
cannam@167
|
778 E TS, TW, T1G, T1I;
|
cannam@167
|
779 TS = W[36];
|
cannam@167
|
780 TW = W[37];
|
cannam@167
|
781 T10 = FNMS(TW, TZ, TS * TV);
|
cannam@167
|
782 T2Q = FMA(TW, TV, TS * TZ);
|
cannam@167
|
783 T1G = W[34];
|
cannam@167
|
784 T1I = W[35];
|
cannam@167
|
785 T1K = FNMS(T1I, T1J, T1G * T1H);
|
cannam@167
|
786 T3s = FMA(T1I, T1H, T1G * T1J);
|
cannam@167
|
787 }
|
cannam@167
|
788 }
|
cannam@167
|
789 {
|
cannam@167
|
790 E T1P, T27, T1T, T29;
|
cannam@167
|
791 {
|
cannam@167
|
792 E T1N, T1O, T1R, T1S;
|
cannam@167
|
793 T1N = Ip[WS(rs, 6)];
|
cannam@167
|
794 T1O = Im[WS(rs, 6)];
|
cannam@167
|
795 T1P = T1N + T1O;
|
cannam@167
|
796 T27 = T1N - T1O;
|
cannam@167
|
797 T1R = Rp[WS(rs, 6)];
|
cannam@167
|
798 T1S = Rm[WS(rs, 6)];
|
cannam@167
|
799 T1T = T1R - T1S;
|
cannam@167
|
800 T29 = T1R + T1S;
|
cannam@167
|
801 }
|
cannam@167
|
802 {
|
cannam@167
|
803 E T1M, T1Q, T26, T28;
|
cannam@167
|
804 T1M = W[24];
|
cannam@167
|
805 T1Q = W[25];
|
cannam@167
|
806 T1U = FNMS(T1Q, T1T, T1M * T1P);
|
cannam@167
|
807 T2J = FMA(T1Q, T1P, T1M * T1T);
|
cannam@167
|
808 T26 = W[22];
|
cannam@167
|
809 T28 = W[23];
|
cannam@167
|
810 T2a = FNMS(T28, T29, T26 * T27);
|
cannam@167
|
811 T3x = FMA(T28, T27, T26 * T29);
|
cannam@167
|
812 }
|
cannam@167
|
813 }
|
cannam@167
|
814 {
|
cannam@167
|
815 E T16, T2k, T1a, T2i;
|
cannam@167
|
816 {
|
cannam@167
|
817 E T14, T15, T18, T19;
|
cannam@167
|
818 T14 = Ip[WS(rs, 8)];
|
cannam@167
|
819 T15 = Im[WS(rs, 8)];
|
cannam@167
|
820 T16 = T14 - T15;
|
cannam@167
|
821 T2k = T14 + T15;
|
cannam@167
|
822 T18 = Rp[WS(rs, 8)];
|
cannam@167
|
823 T19 = Rm[WS(rs, 8)];
|
cannam@167
|
824 T1a = T18 + T19;
|
cannam@167
|
825 T2i = T19 - T18;
|
cannam@167
|
826 }
|
cannam@167
|
827 {
|
cannam@167
|
828 E T13, T17, T2h, T2j;
|
cannam@167
|
829 T13 = W[30];
|
cannam@167
|
830 T17 = W[31];
|
cannam@167
|
831 T1b = FNMS(T17, T1a, T13 * T16);
|
cannam@167
|
832 T3l = FMA(T13, T1a, T17 * T16);
|
cannam@167
|
833 T2h = W[33];
|
cannam@167
|
834 T2j = W[32];
|
cannam@167
|
835 T2l = FMA(T2h, T2i, T2j * T2k);
|
cannam@167
|
836 T2D = FNMS(T2h, T2k, T2j * T2i);
|
cannam@167
|
837 }
|
cannam@167
|
838 }
|
cannam@167
|
839 {
|
cannam@167
|
840 E T2g, T2r, T3n, T3o;
|
cannam@167
|
841 {
|
cannam@167
|
842 E TI, T11, T4m, T4n;
|
cannam@167
|
843 TI = TC - TH;
|
cannam@167
|
844 T11 = TR - T10;
|
cannam@167
|
845 T12 = TI - T11;
|
cannam@167
|
846 T2w = TI + T11;
|
cannam@167
|
847 T4m = T3g + T3h;
|
cannam@167
|
848 T4n = TR + T10;
|
cannam@167
|
849 T4o = T4m + T4n;
|
cannam@167
|
850 T4V = T4m - T4n;
|
cannam@167
|
851 }
|
cannam@167
|
852 {
|
cannam@167
|
853 E T2F, T2G, T4w, T4x;
|
cannam@167
|
854 T2F = T2D - T2E;
|
cannam@167
|
855 T2G = T2a + T2f;
|
cannam@167
|
856 T2H = T2F - T2G;
|
cannam@167
|
857 T3a = T2F + T2G;
|
cannam@167
|
858 T4w = T2l + T2q;
|
cannam@167
|
859 T4x = T3x + T3y;
|
cannam@167
|
860 T4y = T4w + T4x;
|
cannam@167
|
861 T4Y = T4x - T4w;
|
cannam@167
|
862 }
|
cannam@167
|
863 {
|
cannam@167
|
864 E T1l, T1y, T1L, T24;
|
cannam@167
|
865 T1l = T1b - T1k;
|
cannam@167
|
866 T1y = T1q - T1x;
|
cannam@167
|
867 T1z = T1l + T1y;
|
cannam@167
|
868 T2v = T1y - T1l;
|
cannam@167
|
869 T1L = T1F - T1K;
|
cannam@167
|
870 T24 = T1U - T23;
|
cannam@167
|
871 T25 = T1L - T24;
|
cannam@167
|
872 T2y = T1L + T24;
|
cannam@167
|
873 }
|
cannam@167
|
874 T2g = T2a - T2f;
|
cannam@167
|
875 T2r = T2l - T2q;
|
cannam@167
|
876 T2s = T2g - T2r;
|
cannam@167
|
877 T2z = T2r + T2g;
|
cannam@167
|
878 {
|
cannam@167
|
879 E T4t, T4u, T4p, T4q;
|
cannam@167
|
880 T4t = T3r + T3s;
|
cannam@167
|
881 T4u = T1U + T23;
|
cannam@167
|
882 T4v = T4t + T4u;
|
cannam@167
|
883 T4X = T4t - T4u;
|
cannam@167
|
884 T4p = T3l + T3m;
|
cannam@167
|
885 T4q = T1q + T1x;
|
cannam@167
|
886 T4r = T4p + T4q;
|
cannam@167
|
887 T4U = T4p - T4q;
|
cannam@167
|
888 }
|
cannam@167
|
889 {
|
cannam@167
|
890 E T3w, T3z, T2T, T2W;
|
cannam@167
|
891 T3w = T2D + T2E;
|
cannam@167
|
892 T3z = T3x - T3y;
|
cannam@167
|
893 T3A = T3w + T3z;
|
cannam@167
|
894 T3Z = T3z - T3w;
|
cannam@167
|
895 T2T = T1b + T1k;
|
cannam@167
|
896 T2W = T2U + T2V;
|
cannam@167
|
897 T2X = T2T + T2W;
|
cannam@167
|
898 T37 = T2T - T2W;
|
cannam@167
|
899 }
|
cannam@167
|
900 {
|
cannam@167
|
901 E T3i, T3j, T2I, T2L;
|
cannam@167
|
902 T3i = T3g - T3h;
|
cannam@167
|
903 T3j = T2Q - T2P;
|
cannam@167
|
904 T3k = T3i + T3j;
|
cannam@167
|
905 T41 = T3i - T3j;
|
cannam@167
|
906 T2I = T1F + T1K;
|
cannam@167
|
907 T2L = T2J + T2K;
|
cannam@167
|
908 T2M = T2I + T2L;
|
cannam@167
|
909 T39 = T2I - T2L;
|
cannam@167
|
910 }
|
cannam@167
|
911 {
|
cannam@167
|
912 E T3t, T3u, T2O, T2R;
|
cannam@167
|
913 T3t = T3r - T3s;
|
cannam@167
|
914 T3u = T2K - T2J;
|
cannam@167
|
915 T3v = T3t + T3u;
|
cannam@167
|
916 T3Y = T3t - T3u;
|
cannam@167
|
917 T2O = TC + TH;
|
cannam@167
|
918 T2R = T2P + T2Q;
|
cannam@167
|
919 T2S = T2O + T2R;
|
cannam@167
|
920 T36 = T2O - T2R;
|
cannam@167
|
921 }
|
cannam@167
|
922 T3n = T3l - T3m;
|
cannam@167
|
923 T3o = T2U - T2V;
|
cannam@167
|
924 T3p = T3n + T3o;
|
cannam@167
|
925 T42 = T3n - T3o;
|
cannam@167
|
926 {
|
cannam@167
|
927 E Tc, T3M, T4, T8;
|
cannam@167
|
928 T4 = W[18];
|
cannam@167
|
929 T8 = W[19];
|
cannam@167
|
930 Tc = FNMS(T8, Tb, T4 * T7);
|
cannam@167
|
931 T3M = FMA(T4, Tb, T8 * T7);
|
cannam@167
|
932 Td = T3 - Tc;
|
cannam@167
|
933 T4G = T3L + T3M;
|
cannam@167
|
934 T33 = Tc + T3;
|
cannam@167
|
935 T3N = T3L - T3M;
|
cannam@167
|
936 }
|
cannam@167
|
937 {
|
cannam@167
|
938 E Tm, T30, Tv, T31;
|
cannam@167
|
939 {
|
cannam@167
|
940 E Te, Ti, Tn, Tr;
|
cannam@167
|
941 Te = W[8];
|
cannam@167
|
942 Ti = W[9];
|
cannam@167
|
943 Tm = FNMS(Ti, Tl, Te * Th);
|
cannam@167
|
944 T30 = FMA(Ti, Th, Te * Tl);
|
cannam@167
|
945 Tn = W[28];
|
cannam@167
|
946 Tr = W[29];
|
cannam@167
|
947 Tv = FNMS(Tr, Tu, Tn * Tq);
|
cannam@167
|
948 T31 = FMA(Tr, Tq, Tn * Tu);
|
cannam@167
|
949 }
|
cannam@167
|
950 Tw = Tm - Tv;
|
cannam@167
|
951 T4H = Tm + Tv;
|
cannam@167
|
952 T32 = T30 + T31;
|
cannam@167
|
953 T3O = T31 - T30;
|
cannam@167
|
954 }
|
cannam@167
|
955 }
|
cannam@167
|
956 }
|
cannam@167
|
957 {
|
cannam@167
|
958 E T3C, T3E, Tx, T2u, T3d, T3e, T3D, T3f;
|
cannam@167
|
959 {
|
cannam@167
|
960 E T3q, T3B, T1A, T2t;
|
cannam@167
|
961 T3q = T3k - T3p;
|
cannam@167
|
962 T3B = T3v - T3A;
|
cannam@167
|
963 T3C = FMA(KP475528258, T3q, KP293892626 * T3B);
|
cannam@167
|
964 T3E = FNMS(KP293892626, T3q, KP475528258 * T3B);
|
cannam@167
|
965 Tx = Td - Tw;
|
cannam@167
|
966 T1A = T12 + T1z;
|
cannam@167
|
967 T2t = T25 + T2s;
|
cannam@167
|
968 T2u = T1A + T2t;
|
cannam@167
|
969 T3d = KP279508497 * (T1A - T2t);
|
cannam@167
|
970 T3e = FNMS(KP125000000, T2u, KP500000000 * Tx);
|
cannam@167
|
971 }
|
cannam@167
|
972 Ip[WS(rs, 5)] = KP500000000 * (Tx + T2u);
|
cannam@167
|
973 T3D = T3d - T3e;
|
cannam@167
|
974 Im[WS(rs, 2)] = T3D - T3E;
|
cannam@167
|
975 Im[WS(rs, 6)] = T3D + T3E;
|
cannam@167
|
976 T3f = T3d + T3e;
|
cannam@167
|
977 Ip[WS(rs, 1)] = T3f - T3C;
|
cannam@167
|
978 Ip[WS(rs, 9)] = T3f + T3C;
|
cannam@167
|
979 }
|
cannam@167
|
980 {
|
cannam@167
|
981 E T3H, T3T, T3P, T3Q, T3K, T3R, T3U, T3S;
|
cannam@167
|
982 {
|
cannam@167
|
983 E T3F, T3G, T3I, T3J;
|
cannam@167
|
984 T3F = T12 - T1z;
|
cannam@167
|
985 T3G = T25 - T2s;
|
cannam@167
|
986 T3H = FMA(KP475528258, T3F, KP293892626 * T3G);
|
cannam@167
|
987 T3T = FNMS(KP293892626, T3F, KP475528258 * T3G);
|
cannam@167
|
988 T3P = T3N + T3O;
|
cannam@167
|
989 T3I = T3k + T3p;
|
cannam@167
|
990 T3J = T3v + T3A;
|
cannam@167
|
991 T3Q = T3I + T3J;
|
cannam@167
|
992 T3K = KP279508497 * (T3I - T3J);
|
cannam@167
|
993 T3R = FNMS(KP125000000, T3Q, KP500000000 * T3P);
|
cannam@167
|
994 }
|
cannam@167
|
995 Rp[WS(rs, 5)] = KP500000000 * (T3P + T3Q);
|
cannam@167
|
996 T3U = T3R - T3K;
|
cannam@167
|
997 Rm[WS(rs, 6)] = T3T + T3U;
|
cannam@167
|
998 Rm[WS(rs, 2)] = T3U - T3T;
|
cannam@167
|
999 T3S = T3K + T3R;
|
cannam@167
|
1000 Rp[WS(rs, 1)] = T3H + T3S;
|
cannam@167
|
1001 Rp[WS(rs, 9)] = T3S - T3H;
|
cannam@167
|
1002 }
|
cannam@167
|
1003 {
|
cannam@167
|
1004 E T44, T46, T2C, T2B, T3V, T3W, T45, T3X;
|
cannam@167
|
1005 {
|
cannam@167
|
1006 E T40, T43, T2x, T2A;
|
cannam@167
|
1007 T40 = T3Y - T3Z;
|
cannam@167
|
1008 T43 = T41 - T42;
|
cannam@167
|
1009 T44 = FNMS(KP293892626, T43, KP475528258 * T40);
|
cannam@167
|
1010 T46 = FMA(KP475528258, T43, KP293892626 * T40);
|
cannam@167
|
1011 T2C = Tw + Td;
|
cannam@167
|
1012 T2x = T2v - T2w;
|
cannam@167
|
1013 T2A = T2y + T2z;
|
cannam@167
|
1014 T2B = T2x - T2A;
|
cannam@167
|
1015 T3V = FMA(KP500000000, T2C, KP125000000 * T2B);
|
cannam@167
|
1016 T3W = KP279508497 * (T2x + T2A);
|
cannam@167
|
1017 }
|
cannam@167
|
1018 Im[WS(rs, 4)] = KP500000000 * (T2B - T2C);
|
cannam@167
|
1019 T45 = T3W - T3V;
|
cannam@167
|
1020 Im[0] = T45 - T46;
|
cannam@167
|
1021 Im[WS(rs, 8)] = T45 + T46;
|
cannam@167
|
1022 T3X = T3V + T3W;
|
cannam@167
|
1023 Ip[WS(rs, 3)] = T3X - T44;
|
cannam@167
|
1024 Ip[WS(rs, 7)] = T3X + T44;
|
cannam@167
|
1025 }
|
cannam@167
|
1026 {
|
cannam@167
|
1027 E T49, T4h, T4a, T4d, T4e, T4f, T4i, T4g;
|
cannam@167
|
1028 {
|
cannam@167
|
1029 E T47, T48, T4b, T4c;
|
cannam@167
|
1030 T47 = T2y - T2z;
|
cannam@167
|
1031 T48 = T2w + T2v;
|
cannam@167
|
1032 T49 = FNMS(KP293892626, T48, KP475528258 * T47);
|
cannam@167
|
1033 T4h = FMA(KP475528258, T48, KP293892626 * T47);
|
cannam@167
|
1034 T4a = T3N - T3O;
|
cannam@167
|
1035 T4b = T41 + T42;
|
cannam@167
|
1036 T4c = T3Y + T3Z;
|
cannam@167
|
1037 T4d = T4b + T4c;
|
cannam@167
|
1038 T4e = FNMS(KP125000000, T4d, KP500000000 * T4a);
|
cannam@167
|
1039 T4f = KP279508497 * (T4b - T4c);
|
cannam@167
|
1040 }
|
cannam@167
|
1041 Rm[WS(rs, 4)] = KP500000000 * (T4a + T4d);
|
cannam@167
|
1042 T4i = T4f + T4e;
|
cannam@167
|
1043 Rm[WS(rs, 8)] = T4h + T4i;
|
cannam@167
|
1044 Rm[0] = T4i - T4h;
|
cannam@167
|
1045 T4g = T4e - T4f;
|
cannam@167
|
1046 Rp[WS(rs, 3)] = T49 + T4g;
|
cannam@167
|
1047 Rp[WS(rs, 7)] = T4g - T49;
|
cannam@167
|
1048 }
|
cannam@167
|
1049 {
|
cannam@167
|
1050 E T50, T52, T34, T2Z, T4R, T4S, T51, T4T;
|
cannam@167
|
1051 {
|
cannam@167
|
1052 E T4W, T4Z, T2N, T2Y;
|
cannam@167
|
1053 T4W = T4U - T4V;
|
cannam@167
|
1054 T4Z = T4X - T4Y;
|
cannam@167
|
1055 T50 = FNMS(KP293892626, T4Z, KP475528258 * T4W);
|
cannam@167
|
1056 T52 = FMA(KP293892626, T4W, KP475528258 * T4Z);
|
cannam@167
|
1057 T34 = T32 + T33;
|
cannam@167
|
1058 T2N = T2H - T2M;
|
cannam@167
|
1059 T2Y = T2S + T2X;
|
cannam@167
|
1060 T2Z = T2N - T2Y;
|
cannam@167
|
1061 T4R = FMA(KP500000000, T34, KP125000000 * T2Z);
|
cannam@167
|
1062 T4S = KP279508497 * (T2Y + T2N);
|
cannam@167
|
1063 }
|
cannam@167
|
1064 Im[WS(rs, 9)] = KP500000000 * (T2Z - T34);
|
cannam@167
|
1065 T51 = T4R - T4S;
|
cannam@167
|
1066 Ip[WS(rs, 2)] = T51 + T52;
|
cannam@167
|
1067 Im[WS(rs, 1)] = T52 - T51;
|
cannam@167
|
1068 T4T = T4R + T4S;
|
cannam@167
|
1069 Ip[WS(rs, 6)] = T4T + T50;
|
cannam@167
|
1070 Im[WS(rs, 5)] = T50 - T4T;
|
cannam@167
|
1071 }
|
cannam@167
|
1072 {
|
cannam@167
|
1073 E T5c, T5d, T53, T56, T57, T58, T5e, T59;
|
cannam@167
|
1074 {
|
cannam@167
|
1075 E T5a, T5b, T54, T55;
|
cannam@167
|
1076 T5a = T2M + T2H;
|
cannam@167
|
1077 T5b = T2S - T2X;
|
cannam@167
|
1078 T5c = FNMS(KP293892626, T5b, KP475528258 * T5a);
|
cannam@167
|
1079 T5d = FMA(KP475528258, T5b, KP293892626 * T5a);
|
cannam@167
|
1080 T53 = T4G - T4H;
|
cannam@167
|
1081 T54 = T4V + T4U;
|
cannam@167
|
1082 T55 = T4X + T4Y;
|
cannam@167
|
1083 T56 = T54 + T55;
|
cannam@167
|
1084 T57 = FNMS(KP125000000, T56, KP500000000 * T53);
|
cannam@167
|
1085 T58 = KP279508497 * (T54 - T55);
|
cannam@167
|
1086 }
|
cannam@167
|
1087 Rm[WS(rs, 9)] = KP500000000 * (T53 + T56);
|
cannam@167
|
1088 T5e = T58 + T57;
|
cannam@167
|
1089 Rp[WS(rs, 6)] = T5d + T5e;
|
cannam@167
|
1090 Rm[WS(rs, 5)] = T5e - T5d;
|
cannam@167
|
1091 T59 = T57 - T58;
|
cannam@167
|
1092 Rp[WS(rs, 2)] = T59 - T5c;
|
cannam@167
|
1093 Rm[WS(rs, 1)] = T5c + T59;
|
cannam@167
|
1094 }
|
cannam@167
|
1095 {
|
cannam@167
|
1096 E T4A, T4C, T35, T3c, T4j, T4k, T4B, T4l;
|
cannam@167
|
1097 {
|
cannam@167
|
1098 E T4s, T4z, T38, T3b;
|
cannam@167
|
1099 T4s = T4o - T4r;
|
cannam@167
|
1100 T4z = T4v - T4y;
|
cannam@167
|
1101 T4A = FNMS(KP475528258, T4z, KP293892626 * T4s);
|
cannam@167
|
1102 T4C = FMA(KP475528258, T4s, KP293892626 * T4z);
|
cannam@167
|
1103 T35 = T33 - T32;
|
cannam@167
|
1104 T38 = T36 + T37;
|
cannam@167
|
1105 T3b = T39 + T3a;
|
cannam@167
|
1106 T3c = T38 + T3b;
|
cannam@167
|
1107 T4j = FNMS(KP125000000, T3c, KP500000000 * T35);
|
cannam@167
|
1108 T4k = KP279508497 * (T38 - T3b);
|
cannam@167
|
1109 }
|
cannam@167
|
1110 Ip[0] = KP500000000 * (T35 + T3c);
|
cannam@167
|
1111 T4B = T4k + T4j;
|
cannam@167
|
1112 Ip[WS(rs, 4)] = T4B + T4C;
|
cannam@167
|
1113 Im[WS(rs, 3)] = T4C - T4B;
|
cannam@167
|
1114 T4l = T4j - T4k;
|
cannam@167
|
1115 Ip[WS(rs, 8)] = T4l + T4A;
|
cannam@167
|
1116 Im[WS(rs, 7)] = T4A - T4l;
|
cannam@167
|
1117 }
|
cannam@167
|
1118 {
|
cannam@167
|
1119 E T4O, T4P, T4I, T4J, T4F, T4K, T4Q, T4L;
|
cannam@167
|
1120 {
|
cannam@167
|
1121 E T4M, T4N, T4D, T4E;
|
cannam@167
|
1122 T4M = T36 - T37;
|
cannam@167
|
1123 T4N = T39 - T3a;
|
cannam@167
|
1124 T4O = FMA(KP475528258, T4M, KP293892626 * T4N);
|
cannam@167
|
1125 T4P = FNMS(KP293892626, T4M, KP475528258 * T4N);
|
cannam@167
|
1126 T4I = T4G + T4H;
|
cannam@167
|
1127 T4D = T4o + T4r;
|
cannam@167
|
1128 T4E = T4v + T4y;
|
cannam@167
|
1129 T4J = T4D + T4E;
|
cannam@167
|
1130 T4F = KP279508497 * (T4D - T4E);
|
cannam@167
|
1131 T4K = FNMS(KP125000000, T4J, KP500000000 * T4I);
|
cannam@167
|
1132 }
|
cannam@167
|
1133 Rp[0] = KP500000000 * (T4I + T4J);
|
cannam@167
|
1134 T4Q = T4K - T4F;
|
cannam@167
|
1135 Rp[WS(rs, 8)] = T4P + T4Q;
|
cannam@167
|
1136 Rm[WS(rs, 7)] = T4Q - T4P;
|
cannam@167
|
1137 T4L = T4F + T4K;
|
cannam@167
|
1138 Rp[WS(rs, 4)] = T4L - T4O;
|
cannam@167
|
1139 Rm[WS(rs, 3)] = T4O + T4L;
|
cannam@167
|
1140 }
|
cannam@167
|
1141 }
|
cannam@167
|
1142 }
|
cannam@167
|
1143 }
|
cannam@167
|
1144
|
cannam@167
|
1145 static const tw_instr twinstr[] = {
|
cannam@167
|
1146 {TW_FULL, 1, 20},
|
cannam@167
|
1147 {TW_NEXT, 1, 0}
|
cannam@167
|
1148 };
|
cannam@167
|
1149
|
cannam@167
|
1150 static const hc2c_desc desc = { 20, "hc2cfdft_20", twinstr, &GENUS, {224, 78, 62, 0} };
|
cannam@167
|
1151
|
cannam@167
|
1152 void X(codelet_hc2cfdft_20) (planner *p) {
|
cannam@167
|
1153 X(khc2c_register) (p, hc2cfdft_20, &desc, HC2C_VIA_DFT);
|
cannam@167
|
1154 }
|
cannam@167
|
1155 #endif
|