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 08:08:01 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 -sign 1 -n 20 -dif -name hc2cbdft2_20 -include rdft/scalar/hc2cb.h */
|
cannam@167
|
29
|
cannam@167
|
30 /*
|
cannam@167
|
31 * This function contains 286 FP additions, 148 FP multiplications,
|
cannam@167
|
32 * (or, 176 additions, 38 multiplications, 110 fused multiply/add),
|
cannam@167
|
33 * 104 stack variables, 4 constants, and 80 memory accesses
|
cannam@167
|
34 */
|
cannam@167
|
35 #include "rdft/scalar/hc2cb.h"
|
cannam@167
|
36
|
cannam@167
|
37 static void hc2cbdft2_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(KP250000000, +0.250000000000000000000000000000000000000000000);
|
cannam@167
|
42 DK(KP618033988, +0.618033988749894848204586834365638117720309180);
|
cannam@167
|
43 {
|
cannam@167
|
44 INT m;
|
cannam@167
|
45 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
|
46 E T27, T2o, T3T, T41, T2p, T40, T1N, T2Q, T1w, T2L, T4n, T59, T4A, T5e, T24;
|
cannam@167
|
47 E T2m, T2h, T2Z, T3P, T4J, T3W, T3Y, T7, TC, T2c, T2d, T3y, T3F, T3G, T3H;
|
cannam@167
|
48 E T46, T4d, T4e, T4f, T4r, T4u, T4v, T4w, T1E, T1H, T1I, T1J, TJ, T16, T17;
|
cannam@167
|
49 E T18;
|
cannam@167
|
50 {
|
cannam@167
|
51 E T3, T1A, TI, T25, T6, TF, T1D, T26, Te, T47, T4k, TO, T1e, T3z, T3M;
|
cannam@167
|
52 E T1S, Tt, T4a, T4h, TZ, T1p, T3C, T3J, T1Z, TA, T4b, T4i, T14, T1u, T3D;
|
cannam@167
|
53 E T3K, T22, Tl, T48, T4l, TT, T1j, T3A, T3N, T1V;
|
cannam@167
|
54 {
|
cannam@167
|
55 E T1, T2, TG, TH;
|
cannam@167
|
56 T1 = Rp[0];
|
cannam@167
|
57 T2 = Rm[WS(rs, 9)];
|
cannam@167
|
58 T3 = T1 + T2;
|
cannam@167
|
59 T1A = T1 - T2;
|
cannam@167
|
60 TG = Ip[0];
|
cannam@167
|
61 TH = Im[WS(rs, 9)];
|
cannam@167
|
62 TI = TG + TH;
|
cannam@167
|
63 T25 = TG - TH;
|
cannam@167
|
64 }
|
cannam@167
|
65 {
|
cannam@167
|
66 E T4, T5, T1B, T1C;
|
cannam@167
|
67 T4 = Rp[WS(rs, 5)];
|
cannam@167
|
68 T5 = Rm[WS(rs, 4)];
|
cannam@167
|
69 T6 = T4 + T5;
|
cannam@167
|
70 TF = T4 - T5;
|
cannam@167
|
71 T1B = Ip[WS(rs, 5)];
|
cannam@167
|
72 T1C = Im[WS(rs, 4)];
|
cannam@167
|
73 T1D = T1B + T1C;
|
cannam@167
|
74 T26 = T1B - T1C;
|
cannam@167
|
75 }
|
cannam@167
|
76 {
|
cannam@167
|
77 E Ta, T1a, TN, T1Q, Td, TK, T1d, T1R;
|
cannam@167
|
78 {
|
cannam@167
|
79 E T8, T9, TL, TM;
|
cannam@167
|
80 T8 = Rp[WS(rs, 4)];
|
cannam@167
|
81 T9 = Rm[WS(rs, 5)];
|
cannam@167
|
82 Ta = T8 + T9;
|
cannam@167
|
83 T1a = T8 - T9;
|
cannam@167
|
84 TL = Ip[WS(rs, 4)];
|
cannam@167
|
85 TM = Im[WS(rs, 5)];
|
cannam@167
|
86 TN = TL + TM;
|
cannam@167
|
87 T1Q = TL - TM;
|
cannam@167
|
88 }
|
cannam@167
|
89 {
|
cannam@167
|
90 E Tb, Tc, T1b, T1c;
|
cannam@167
|
91 Tb = Rp[WS(rs, 9)];
|
cannam@167
|
92 Tc = Rm[0];
|
cannam@167
|
93 Td = Tb + Tc;
|
cannam@167
|
94 TK = Tb - Tc;
|
cannam@167
|
95 T1b = Ip[WS(rs, 9)];
|
cannam@167
|
96 T1c = Im[0];
|
cannam@167
|
97 T1d = T1b + T1c;
|
cannam@167
|
98 T1R = T1b - T1c;
|
cannam@167
|
99 }
|
cannam@167
|
100 Te = Ta + Td;
|
cannam@167
|
101 T47 = TN - TK;
|
cannam@167
|
102 T4k = T1a + T1d;
|
cannam@167
|
103 TO = TK + TN;
|
cannam@167
|
104 T1e = T1a - T1d;
|
cannam@167
|
105 T3z = Ta - Td;
|
cannam@167
|
106 T3M = T1Q - T1R;
|
cannam@167
|
107 T1S = T1Q + T1R;
|
cannam@167
|
108 }
|
cannam@167
|
109 {
|
cannam@167
|
110 E Tp, T1l, TY, T1X, Ts, TV, T1o, T1Y;
|
cannam@167
|
111 {
|
cannam@167
|
112 E Tn, To, TW, TX;
|
cannam@167
|
113 Tn = Rp[WS(rs, 8)];
|
cannam@167
|
114 To = Rm[WS(rs, 1)];
|
cannam@167
|
115 Tp = Tn + To;
|
cannam@167
|
116 T1l = Tn - To;
|
cannam@167
|
117 TW = Ip[WS(rs, 8)];
|
cannam@167
|
118 TX = Im[WS(rs, 1)];
|
cannam@167
|
119 TY = TW + TX;
|
cannam@167
|
120 T1X = TW - TX;
|
cannam@167
|
121 }
|
cannam@167
|
122 {
|
cannam@167
|
123 E Tq, Tr, T1m, T1n;
|
cannam@167
|
124 Tq = Rm[WS(rs, 6)];
|
cannam@167
|
125 Tr = Rp[WS(rs, 3)];
|
cannam@167
|
126 Ts = Tq + Tr;
|
cannam@167
|
127 TV = Tq - Tr;
|
cannam@167
|
128 T1m = Im[WS(rs, 6)];
|
cannam@167
|
129 T1n = Ip[WS(rs, 3)];
|
cannam@167
|
130 T1o = T1m + T1n;
|
cannam@167
|
131 T1Y = T1n - T1m;
|
cannam@167
|
132 }
|
cannam@167
|
133 Tt = Tp + Ts;
|
cannam@167
|
134 T4a = TY - TV;
|
cannam@167
|
135 T4h = T1l - T1o;
|
cannam@167
|
136 TZ = TV + TY;
|
cannam@167
|
137 T1p = T1l + T1o;
|
cannam@167
|
138 T3C = Tp - Ts;
|
cannam@167
|
139 T3J = T1X - T1Y;
|
cannam@167
|
140 T1Z = T1X + T1Y;
|
cannam@167
|
141 }
|
cannam@167
|
142 {
|
cannam@167
|
143 E Tw, T1q, T13, T20, Tz, T10, T1t, T21;
|
cannam@167
|
144 {
|
cannam@167
|
145 E Tu, Tv, T11, T12;
|
cannam@167
|
146 Tu = Rm[WS(rs, 7)];
|
cannam@167
|
147 Tv = Rp[WS(rs, 2)];
|
cannam@167
|
148 Tw = Tu + Tv;
|
cannam@167
|
149 T1q = Tu - Tv;
|
cannam@167
|
150 T11 = Im[WS(rs, 7)];
|
cannam@167
|
151 T12 = Ip[WS(rs, 2)];
|
cannam@167
|
152 T13 = T11 + T12;
|
cannam@167
|
153 T20 = T12 - T11;
|
cannam@167
|
154 }
|
cannam@167
|
155 {
|
cannam@167
|
156 E Tx, Ty, T1r, T1s;
|
cannam@167
|
157 Tx = Rm[WS(rs, 2)];
|
cannam@167
|
158 Ty = Rp[WS(rs, 7)];
|
cannam@167
|
159 Tz = Tx + Ty;
|
cannam@167
|
160 T10 = Tx - Ty;
|
cannam@167
|
161 T1r = Im[WS(rs, 2)];
|
cannam@167
|
162 T1s = Ip[WS(rs, 7)];
|
cannam@167
|
163 T1t = T1r + T1s;
|
cannam@167
|
164 T21 = T1s - T1r;
|
cannam@167
|
165 }
|
cannam@167
|
166 TA = Tw + Tz;
|
cannam@167
|
167 T4b = T10 + T13;
|
cannam@167
|
168 T4i = T1q - T1t;
|
cannam@167
|
169 T14 = T10 - T13;
|
cannam@167
|
170 T1u = T1q + T1t;
|
cannam@167
|
171 T3D = Tw - Tz;
|
cannam@167
|
172 T3K = T20 - T21;
|
cannam@167
|
173 T22 = T20 + T21;
|
cannam@167
|
174 }
|
cannam@167
|
175 {
|
cannam@167
|
176 E Th, T1f, TS, T1T, Tk, TP, T1i, T1U;
|
cannam@167
|
177 {
|
cannam@167
|
178 E Tf, Tg, TQ, TR;
|
cannam@167
|
179 Tf = Rm[WS(rs, 3)];
|
cannam@167
|
180 Tg = Rp[WS(rs, 6)];
|
cannam@167
|
181 Th = Tf + Tg;
|
cannam@167
|
182 T1f = Tf - Tg;
|
cannam@167
|
183 TQ = Im[WS(rs, 3)];
|
cannam@167
|
184 TR = Ip[WS(rs, 6)];
|
cannam@167
|
185 TS = TQ + TR;
|
cannam@167
|
186 T1T = TR - TQ;
|
cannam@167
|
187 }
|
cannam@167
|
188 {
|
cannam@167
|
189 E Ti, Tj, T1g, T1h;
|
cannam@167
|
190 Ti = Rp[WS(rs, 1)];
|
cannam@167
|
191 Tj = Rm[WS(rs, 8)];
|
cannam@167
|
192 Tk = Ti + Tj;
|
cannam@167
|
193 TP = Ti - Tj;
|
cannam@167
|
194 T1g = Ip[WS(rs, 1)];
|
cannam@167
|
195 T1h = Im[WS(rs, 8)];
|
cannam@167
|
196 T1i = T1g + T1h;
|
cannam@167
|
197 T1U = T1g - T1h;
|
cannam@167
|
198 }
|
cannam@167
|
199 Tl = Th + Tk;
|
cannam@167
|
200 T48 = TP + TS;
|
cannam@167
|
201 T4l = T1f + T1i;
|
cannam@167
|
202 TT = TP - TS;
|
cannam@167
|
203 T1j = T1f - T1i;
|
cannam@167
|
204 T3A = Th - Tk;
|
cannam@167
|
205 T3N = T1T - T1U;
|
cannam@167
|
206 T1V = T1T + T1U;
|
cannam@167
|
207 }
|
cannam@167
|
208 T27 = T25 + T26;
|
cannam@167
|
209 T2o = Tt - TA;
|
cannam@167
|
210 T3T = T25 - T26;
|
cannam@167
|
211 T41 = T3z - T3A;
|
cannam@167
|
212 T2p = Te - Tl;
|
cannam@167
|
213 {
|
cannam@167
|
214 E T1L, T1M, T1k, T1v;
|
cannam@167
|
215 T40 = T3C - T3D;
|
cannam@167
|
216 T1L = TO - TT;
|
cannam@167
|
217 T1M = TZ - T14;
|
cannam@167
|
218 T1N = FMA(KP618033988, T1M, T1L);
|
cannam@167
|
219 T2Q = FNMS(KP618033988, T1L, T1M);
|
cannam@167
|
220 T1k = T1e - T1j;
|
cannam@167
|
221 T1v = T1p - T1u;
|
cannam@167
|
222 T1w = FMA(KP618033988, T1v, T1k);
|
cannam@167
|
223 T2L = FNMS(KP618033988, T1k, T1v);
|
cannam@167
|
224 {
|
cannam@167
|
225 E T4j, T4m, T4y, T4z;
|
cannam@167
|
226 T4j = T4h - T4i;
|
cannam@167
|
227 T4m = T4k - T4l;
|
cannam@167
|
228 T4n = FNMS(KP618033988, T4m, T4j);
|
cannam@167
|
229 T59 = FMA(KP618033988, T4j, T4m);
|
cannam@167
|
230 T4y = T4a + T4b;
|
cannam@167
|
231 T4z = T47 + T48;
|
cannam@167
|
232 T4A = FNMS(KP618033988, T4z, T4y);
|
cannam@167
|
233 T5e = FMA(KP618033988, T4y, T4z);
|
cannam@167
|
234 }
|
cannam@167
|
235 }
|
cannam@167
|
236 {
|
cannam@167
|
237 E T3L, T3O, T4s, T4t;
|
cannam@167
|
238 {
|
cannam@167
|
239 E T1W, T23, T2f, T2g;
|
cannam@167
|
240 T1W = T1S + T1V;
|
cannam@167
|
241 T23 = T1Z + T22;
|
cannam@167
|
242 T24 = T1W + T23;
|
cannam@167
|
243 T2m = T1W - T23;
|
cannam@167
|
244 T2f = T1Z - T22;
|
cannam@167
|
245 T2g = T1S - T1V;
|
cannam@167
|
246 T2h = FNMS(KP618033988, T2g, T2f);
|
cannam@167
|
247 T2Z = FMA(KP618033988, T2f, T2g);
|
cannam@167
|
248 }
|
cannam@167
|
249 T3L = T3J - T3K;
|
cannam@167
|
250 T3O = T3M - T3N;
|
cannam@167
|
251 T3P = FNMS(KP618033988, T3O, T3L);
|
cannam@167
|
252 T4J = FMA(KP618033988, T3L, T3O);
|
cannam@167
|
253 {
|
cannam@167
|
254 E T3U, T3V, Tm, TB;
|
cannam@167
|
255 T3U = T3M + T3N;
|
cannam@167
|
256 T3V = T3J + T3K;
|
cannam@167
|
257 T3W = T3U + T3V;
|
cannam@167
|
258 T3Y = T3U - T3V;
|
cannam@167
|
259 T7 = T3 + T6;
|
cannam@167
|
260 Tm = Te + Tl;
|
cannam@167
|
261 TB = Tt + TA;
|
cannam@167
|
262 TC = Tm + TB;
|
cannam@167
|
263 T2c = FNMS(KP250000000, TC, T7);
|
cannam@167
|
264 T2d = Tm - TB;
|
cannam@167
|
265 }
|
cannam@167
|
266 {
|
cannam@167
|
267 E T3B, T3E, T49, T4c;
|
cannam@167
|
268 T3y = T3 - T6;
|
cannam@167
|
269 T3B = T3z + T3A;
|
cannam@167
|
270 T3E = T3C + T3D;
|
cannam@167
|
271 T3F = T3B + T3E;
|
cannam@167
|
272 T3G = FNMS(KP250000000, T3F, T3y);
|
cannam@167
|
273 T3H = T3B - T3E;
|
cannam@167
|
274 T46 = TI - TF;
|
cannam@167
|
275 T49 = T47 - T48;
|
cannam@167
|
276 T4c = T4a - T4b;
|
cannam@167
|
277 T4d = T49 + T4c;
|
cannam@167
|
278 T4e = FNMS(KP250000000, T4d, T46);
|
cannam@167
|
279 T4f = T49 - T4c;
|
cannam@167
|
280 }
|
cannam@167
|
281 T4r = T1A + T1D;
|
cannam@167
|
282 T4s = T4k + T4l;
|
cannam@167
|
283 T4t = T4h + T4i;
|
cannam@167
|
284 T4u = T4s + T4t;
|
cannam@167
|
285 T4v = FNMS(KP250000000, T4u, T4r);
|
cannam@167
|
286 T4w = T4s - T4t;
|
cannam@167
|
287 {
|
cannam@167
|
288 E T1F, T1G, TU, T15;
|
cannam@167
|
289 T1E = T1A - T1D;
|
cannam@167
|
290 T1F = T1e + T1j;
|
cannam@167
|
291 T1G = T1p + T1u;
|
cannam@167
|
292 T1H = T1F + T1G;
|
cannam@167
|
293 T1I = FNMS(KP250000000, T1H, T1E);
|
cannam@167
|
294 T1J = T1F - T1G;
|
cannam@167
|
295 TJ = TF + TI;
|
cannam@167
|
296 TU = TO + TT;
|
cannam@167
|
297 T15 = TZ + T14;
|
cannam@167
|
298 T16 = TU + T15;
|
cannam@167
|
299 T17 = FNMS(KP250000000, T16, TJ);
|
cannam@167
|
300 T18 = TU - T15;
|
cannam@167
|
301 }
|
cannam@167
|
302 }
|
cannam@167
|
303 }
|
cannam@167
|
304 {
|
cannam@167
|
305 E TD, T28, T3o, T3r, T3p, T3v, T2r, T3l, T2H, T35, T2b, T2j, T2k, T2z, T2D;
|
cannam@167
|
306 E T2F, T2G, T2T, T2X, T31, T32, T3d, T3h, T3j, T3k, T3t, T1x, T2u, T1O, T2x;
|
cannam@167
|
307 E T1y, T29, T2v, T2B, T2M, T38, T2R, T3b, T2N, T2V, T39, T3f, T3n, T1P, T2a;
|
cannam@167
|
308 E T1z;
|
cannam@167
|
309 TD = T7 + TC;
|
cannam@167
|
310 T28 = T24 + T27;
|
cannam@167
|
311 T3o = TJ + T16;
|
cannam@167
|
312 T3r = T1H + T1E;
|
cannam@167
|
313 T3n = W[8];
|
cannam@167
|
314 T3p = T3n * T3o;
|
cannam@167
|
315 T3v = T3n * T3r;
|
cannam@167
|
316 {
|
cannam@167
|
317 E T2q, T34, T2n, T33, T2l;
|
cannam@167
|
318 T2q = FNMS(KP618033988, T2p, T2o);
|
cannam@167
|
319 T34 = FMA(KP618033988, T2o, T2p);
|
cannam@167
|
320 T2l = FNMS(KP250000000, T24, T27);
|
cannam@167
|
321 T2n = FNMS(KP559016994, T2m, T2l);
|
cannam@167
|
322 T33 = FMA(KP559016994, T2m, T2l);
|
cannam@167
|
323 T2r = FMA(KP951056516, T2q, T2n);
|
cannam@167
|
324 T3l = FNMS(KP951056516, T34, T33);
|
cannam@167
|
325 T2H = FNMS(KP951056516, T2q, T2n);
|
cannam@167
|
326 T35 = FMA(KP951056516, T34, T33);
|
cannam@167
|
327 }
|
cannam@167
|
328 {
|
cannam@167
|
329 E T2i, T2E, T2e, T30, T3i, T2Y;
|
cannam@167
|
330 T2e = FNMS(KP559016994, T2d, T2c);
|
cannam@167
|
331 T2i = FNMS(KP951056516, T2h, T2e);
|
cannam@167
|
332 T2E = FMA(KP951056516, T2h, T2e);
|
cannam@167
|
333 T2b = W[14];
|
cannam@167
|
334 T2j = T2b * T2i;
|
cannam@167
|
335 T2k = W[15];
|
cannam@167
|
336 T2z = T2k * T2i;
|
cannam@167
|
337 T2D = W[22];
|
cannam@167
|
338 T2F = T2D * T2E;
|
cannam@167
|
339 T2G = W[23];
|
cannam@167
|
340 T2T = T2G * T2E;
|
cannam@167
|
341 T2Y = FMA(KP559016994, T2d, T2c);
|
cannam@167
|
342 T30 = FNMS(KP951056516, T2Z, T2Y);
|
cannam@167
|
343 T3i = FMA(KP951056516, T2Z, T2Y);
|
cannam@167
|
344 T2X = W[30];
|
cannam@167
|
345 T31 = T2X * T30;
|
cannam@167
|
346 T32 = W[31];
|
cannam@167
|
347 T3d = T32 * T30;
|
cannam@167
|
348 T3h = W[6];
|
cannam@167
|
349 T3j = T3h * T3i;
|
cannam@167
|
350 T3k = W[7];
|
cannam@167
|
351 T3t = T3k * T3i;
|
cannam@167
|
352 }
|
cannam@167
|
353 {
|
cannam@167
|
354 E T19, T1K, TE, T2t;
|
cannam@167
|
355 T19 = FMA(KP559016994, T18, T17);
|
cannam@167
|
356 T1x = FMA(KP951056516, T1w, T19);
|
cannam@167
|
357 T2u = FNMS(KP951056516, T1w, T19);
|
cannam@167
|
358 T1K = FMA(KP559016994, T1J, T1I);
|
cannam@167
|
359 T1O = FNMS(KP951056516, T1N, T1K);
|
cannam@167
|
360 T2x = FMA(KP951056516, T1N, T1K);
|
cannam@167
|
361 TE = W[0];
|
cannam@167
|
362 T1y = TE * T1x;
|
cannam@167
|
363 T29 = TE * T1O;
|
cannam@167
|
364 T2t = W[16];
|
cannam@167
|
365 T2v = T2t * T2u;
|
cannam@167
|
366 T2B = T2t * T2x;
|
cannam@167
|
367 }
|
cannam@167
|
368 {
|
cannam@167
|
369 E T2K, T2P, T2J, T37;
|
cannam@167
|
370 T2K = FNMS(KP559016994, T18, T17);
|
cannam@167
|
371 T2M = FMA(KP951056516, T2L, T2K);
|
cannam@167
|
372 T38 = FNMS(KP951056516, T2L, T2K);
|
cannam@167
|
373 T2P = FNMS(KP559016994, T1J, T1I);
|
cannam@167
|
374 T2R = FNMS(KP951056516, T2Q, T2P);
|
cannam@167
|
375 T3b = FMA(KP951056516, T2Q, T2P);
|
cannam@167
|
376 T2J = W[24];
|
cannam@167
|
377 T2N = T2J * T2M;
|
cannam@167
|
378 T2V = T2J * T2R;
|
cannam@167
|
379 T37 = W[32];
|
cannam@167
|
380 T39 = T37 * T38;
|
cannam@167
|
381 T3f = T37 * T3b;
|
cannam@167
|
382 }
|
cannam@167
|
383 T1z = W[1];
|
cannam@167
|
384 T1P = FMA(T1z, T1O, T1y);
|
cannam@167
|
385 T2a = FNMS(T1z, T1x, T29);
|
cannam@167
|
386 Rp[0] = TD - T1P;
|
cannam@167
|
387 Ip[0] = T28 + T2a;
|
cannam@167
|
388 Rm[0] = TD + T1P;
|
cannam@167
|
389 Im[0] = T2a - T28;
|
cannam@167
|
390 {
|
cannam@167
|
391 E T3m, T3u, T3s, T3w, T3q;
|
cannam@167
|
392 T3m = FNMS(T3k, T3l, T3j);
|
cannam@167
|
393 T3u = FMA(T3h, T3l, T3t);
|
cannam@167
|
394 T3q = W[9];
|
cannam@167
|
395 T3s = FMA(T3q, T3r, T3p);
|
cannam@167
|
396 T3w = FNMS(T3q, T3o, T3v);
|
cannam@167
|
397 Rp[WS(rs, 2)] = T3m - T3s;
|
cannam@167
|
398 Ip[WS(rs, 2)] = T3u + T3w;
|
cannam@167
|
399 Rm[WS(rs, 2)] = T3m + T3s;
|
cannam@167
|
400 Im[WS(rs, 2)] = T3w - T3u;
|
cannam@167
|
401 }
|
cannam@167
|
402 {
|
cannam@167
|
403 E T2s, T2A, T2y, T2C, T2w;
|
cannam@167
|
404 T2s = FNMS(T2k, T2r, T2j);
|
cannam@167
|
405 T2A = FMA(T2b, T2r, T2z);
|
cannam@167
|
406 T2w = W[17];
|
cannam@167
|
407 T2y = FMA(T2w, T2x, T2v);
|
cannam@167
|
408 T2C = FNMS(T2w, T2u, T2B);
|
cannam@167
|
409 Rp[WS(rs, 4)] = T2s - T2y;
|
cannam@167
|
410 Ip[WS(rs, 4)] = T2A + T2C;
|
cannam@167
|
411 Rm[WS(rs, 4)] = T2s + T2y;
|
cannam@167
|
412 Im[WS(rs, 4)] = T2C - T2A;
|
cannam@167
|
413 }
|
cannam@167
|
414 {
|
cannam@167
|
415 E T2I, T2U, T2S, T2W, T2O;
|
cannam@167
|
416 T2I = FNMS(T2G, T2H, T2F);
|
cannam@167
|
417 T2U = FMA(T2D, T2H, T2T);
|
cannam@167
|
418 T2O = W[25];
|
cannam@167
|
419 T2S = FMA(T2O, T2R, T2N);
|
cannam@167
|
420 T2W = FNMS(T2O, T2M, T2V);
|
cannam@167
|
421 Rp[WS(rs, 6)] = T2I - T2S;
|
cannam@167
|
422 Ip[WS(rs, 6)] = T2U + T2W;
|
cannam@167
|
423 Rm[WS(rs, 6)] = T2I + T2S;
|
cannam@167
|
424 Im[WS(rs, 6)] = T2W - T2U;
|
cannam@167
|
425 }
|
cannam@167
|
426 {
|
cannam@167
|
427 E T36, T3e, T3c, T3g, T3a;
|
cannam@167
|
428 T36 = FNMS(T32, T35, T31);
|
cannam@167
|
429 T3e = FMA(T2X, T35, T3d);
|
cannam@167
|
430 T3a = W[33];
|
cannam@167
|
431 T3c = FMA(T3a, T3b, T39);
|
cannam@167
|
432 T3g = FNMS(T3a, T38, T3f);
|
cannam@167
|
433 Rp[WS(rs, 8)] = T36 - T3c;
|
cannam@167
|
434 Ip[WS(rs, 8)] = T3e + T3g;
|
cannam@167
|
435 Rm[WS(rs, 8)] = T36 + T3c;
|
cannam@167
|
436 Im[WS(rs, 8)] = T3g - T3e;
|
cannam@167
|
437 }
|
cannam@167
|
438 }
|
cannam@167
|
439 {
|
cannam@167
|
440 E T55, T51, T53, T54, T5h, T5I, T5L, T5J, T5P, T43, T5F, T4P, T5p, T3x, T3R;
|
cannam@167
|
441 E T3S, T4D, T5l, T5n, T5o, T5x, T4H, T4L, T4M, T4X, T5B, T5D, T5E, T5N, T4o;
|
cannam@167
|
442 E T4S, T4B, T4V, T4p, T4F, T4T, T4Z, T5a, T5s, T5f, T5v, T5b, T5j, T5t, T5z;
|
cannam@167
|
443 E T52, T5H;
|
cannam@167
|
444 T55 = T3W + T3T;
|
cannam@167
|
445 T52 = T3y + T3F;
|
cannam@167
|
446 T51 = W[18];
|
cannam@167
|
447 T53 = T51 * T52;
|
cannam@167
|
448 T54 = W[19];
|
cannam@167
|
449 T5h = T54 * T52;
|
cannam@167
|
450 T5I = T46 + T4d;
|
cannam@167
|
451 T5L = T4u + T4r;
|
cannam@167
|
452 T5H = W[28];
|
cannam@167
|
453 T5J = T5H * T5I;
|
cannam@167
|
454 T5P = T5H * T5L;
|
cannam@167
|
455 {
|
cannam@167
|
456 E T42, T4O, T3Z, T4N, T3X;
|
cannam@167
|
457 T42 = FNMS(KP618033988, T41, T40);
|
cannam@167
|
458 T4O = FMA(KP618033988, T40, T41);
|
cannam@167
|
459 T3X = FNMS(KP250000000, T3W, T3T);
|
cannam@167
|
460 T3Z = FNMS(KP559016994, T3Y, T3X);
|
cannam@167
|
461 T4N = FMA(KP559016994, T3Y, T3X);
|
cannam@167
|
462 T43 = FNMS(KP951056516, T42, T3Z);
|
cannam@167
|
463 T5F = FNMS(KP951056516, T4O, T4N);
|
cannam@167
|
464 T4P = FMA(KP951056516, T4O, T4N);
|
cannam@167
|
465 T5p = FMA(KP951056516, T42, T3Z);
|
cannam@167
|
466 }
|
cannam@167
|
467 {
|
cannam@167
|
468 E T3Q, T5m, T3I, T4K, T5C, T4I;
|
cannam@167
|
469 T3I = FNMS(KP559016994, T3H, T3G);
|
cannam@167
|
470 T3Q = FMA(KP951056516, T3P, T3I);
|
cannam@167
|
471 T5m = FNMS(KP951056516, T3P, T3I);
|
cannam@167
|
472 T3x = W[2];
|
cannam@167
|
473 T3R = T3x * T3Q;
|
cannam@167
|
474 T3S = W[3];
|
cannam@167
|
475 T4D = T3S * T3Q;
|
cannam@167
|
476 T5l = W[34];
|
cannam@167
|
477 T5n = T5l * T5m;
|
cannam@167
|
478 T5o = W[35];
|
cannam@167
|
479 T5x = T5o * T5m;
|
cannam@167
|
480 T4I = FMA(KP559016994, T3H, T3G);
|
cannam@167
|
481 T4K = FNMS(KP951056516, T4J, T4I);
|
cannam@167
|
482 T5C = FMA(KP951056516, T4J, T4I);
|
cannam@167
|
483 T4H = W[10];
|
cannam@167
|
484 T4L = T4H * T4K;
|
cannam@167
|
485 T4M = W[11];
|
cannam@167
|
486 T4X = T4M * T4K;
|
cannam@167
|
487 T5B = W[26];
|
cannam@167
|
488 T5D = T5B * T5C;
|
cannam@167
|
489 T5E = W[27];
|
cannam@167
|
490 T5N = T5E * T5C;
|
cannam@167
|
491 }
|
cannam@167
|
492 {
|
cannam@167
|
493 E T4g, T4x, T45, T4R;
|
cannam@167
|
494 T4g = FNMS(KP559016994, T4f, T4e);
|
cannam@167
|
495 T4o = FMA(KP951056516, T4n, T4g);
|
cannam@167
|
496 T4S = FNMS(KP951056516, T4n, T4g);
|
cannam@167
|
497 T4x = FNMS(KP559016994, T4w, T4v);
|
cannam@167
|
498 T4B = FNMS(KP951056516, T4A, T4x);
|
cannam@167
|
499 T4V = FMA(KP951056516, T4A, T4x);
|
cannam@167
|
500 T45 = W[4];
|
cannam@167
|
501 T4p = T45 * T4o;
|
cannam@167
|
502 T4F = T45 * T4B;
|
cannam@167
|
503 T4R = W[12];
|
cannam@167
|
504 T4T = T4R * T4S;
|
cannam@167
|
505 T4Z = T4R * T4V;
|
cannam@167
|
506 }
|
cannam@167
|
507 {
|
cannam@167
|
508 E T58, T5d, T57, T5r;
|
cannam@167
|
509 T58 = FMA(KP559016994, T4f, T4e);
|
cannam@167
|
510 T5a = FMA(KP951056516, T59, T58);
|
cannam@167
|
511 T5s = FNMS(KP951056516, T59, T58);
|
cannam@167
|
512 T5d = FMA(KP559016994, T4w, T4v);
|
cannam@167
|
513 T5f = FNMS(KP951056516, T5e, T5d);
|
cannam@167
|
514 T5v = FMA(KP951056516, T5e, T5d);
|
cannam@167
|
515 T57 = W[20];
|
cannam@167
|
516 T5b = T57 * T5a;
|
cannam@167
|
517 T5j = T57 * T5f;
|
cannam@167
|
518 T5r = W[36];
|
cannam@167
|
519 T5t = T5r * T5s;
|
cannam@167
|
520 T5z = T5r * T5v;
|
cannam@167
|
521 }
|
cannam@167
|
522 {
|
cannam@167
|
523 E T44, T4E, T4C, T4G, T4q;
|
cannam@167
|
524 T44 = FNMS(T3S, T43, T3R);
|
cannam@167
|
525 T4E = FMA(T3x, T43, T4D);
|
cannam@167
|
526 T4q = W[5];
|
cannam@167
|
527 T4C = FMA(T4q, T4B, T4p);
|
cannam@167
|
528 T4G = FNMS(T4q, T4o, T4F);
|
cannam@167
|
529 Rp[WS(rs, 1)] = T44 - T4C;
|
cannam@167
|
530 Ip[WS(rs, 1)] = T4E + T4G;
|
cannam@167
|
531 Rm[WS(rs, 1)] = T44 + T4C;
|
cannam@167
|
532 Im[WS(rs, 1)] = T4G - T4E;
|
cannam@167
|
533 }
|
cannam@167
|
534 {
|
cannam@167
|
535 E T5G, T5O, T5M, T5Q, T5K;
|
cannam@167
|
536 T5G = FNMS(T5E, T5F, T5D);
|
cannam@167
|
537 T5O = FMA(T5B, T5F, T5N);
|
cannam@167
|
538 T5K = W[29];
|
cannam@167
|
539 T5M = FMA(T5K, T5L, T5J);
|
cannam@167
|
540 T5Q = FNMS(T5K, T5I, T5P);
|
cannam@167
|
541 Rp[WS(rs, 7)] = T5G - T5M;
|
cannam@167
|
542 Ip[WS(rs, 7)] = T5O + T5Q;
|
cannam@167
|
543 Rm[WS(rs, 7)] = T5G + T5M;
|
cannam@167
|
544 Im[WS(rs, 7)] = T5Q - T5O;
|
cannam@167
|
545 }
|
cannam@167
|
546 {
|
cannam@167
|
547 E T4Q, T4Y, T4W, T50, T4U;
|
cannam@167
|
548 T4Q = FNMS(T4M, T4P, T4L);
|
cannam@167
|
549 T4Y = FMA(T4H, T4P, T4X);
|
cannam@167
|
550 T4U = W[13];
|
cannam@167
|
551 T4W = FMA(T4U, T4V, T4T);
|
cannam@167
|
552 T50 = FNMS(T4U, T4S, T4Z);
|
cannam@167
|
553 Rp[WS(rs, 3)] = T4Q - T4W;
|
cannam@167
|
554 Ip[WS(rs, 3)] = T4Y + T50;
|
cannam@167
|
555 Rm[WS(rs, 3)] = T4Q + T4W;
|
cannam@167
|
556 Im[WS(rs, 3)] = T50 - T4Y;
|
cannam@167
|
557 }
|
cannam@167
|
558 {
|
cannam@167
|
559 E T56, T5i, T5g, T5k, T5c;
|
cannam@167
|
560 T56 = FNMS(T54, T55, T53);
|
cannam@167
|
561 T5i = FMA(T51, T55, T5h);
|
cannam@167
|
562 T5c = W[21];
|
cannam@167
|
563 T5g = FMA(T5c, T5f, T5b);
|
cannam@167
|
564 T5k = FNMS(T5c, T5a, T5j);
|
cannam@167
|
565 Rp[WS(rs, 5)] = T56 - T5g;
|
cannam@167
|
566 Ip[WS(rs, 5)] = T5i + T5k;
|
cannam@167
|
567 Rm[WS(rs, 5)] = T56 + T5g;
|
cannam@167
|
568 Im[WS(rs, 5)] = T5k - T5i;
|
cannam@167
|
569 }
|
cannam@167
|
570 {
|
cannam@167
|
571 E T5q, T5y, T5w, T5A, T5u;
|
cannam@167
|
572 T5q = FNMS(T5o, T5p, T5n);
|
cannam@167
|
573 T5y = FMA(T5l, T5p, T5x);
|
cannam@167
|
574 T5u = W[37];
|
cannam@167
|
575 T5w = FMA(T5u, T5v, T5t);
|
cannam@167
|
576 T5A = FNMS(T5u, T5s, T5z);
|
cannam@167
|
577 Rp[WS(rs, 9)] = T5q - T5w;
|
cannam@167
|
578 Ip[WS(rs, 9)] = T5y + T5A;
|
cannam@167
|
579 Rm[WS(rs, 9)] = T5q + T5w;
|
cannam@167
|
580 Im[WS(rs, 9)] = T5A - T5y;
|
cannam@167
|
581 }
|
cannam@167
|
582 }
|
cannam@167
|
583 }
|
cannam@167
|
584 }
|
cannam@167
|
585 }
|
cannam@167
|
586
|
cannam@167
|
587 static const tw_instr twinstr[] = {
|
cannam@167
|
588 {TW_FULL, 1, 20},
|
cannam@167
|
589 {TW_NEXT, 1, 0}
|
cannam@167
|
590 };
|
cannam@167
|
591
|
cannam@167
|
592 static const hc2c_desc desc = { 20, "hc2cbdft2_20", twinstr, &GENUS, {176, 38, 110, 0} };
|
cannam@167
|
593
|
cannam@167
|
594 void X(codelet_hc2cbdft2_20) (planner *p) {
|
cannam@167
|
595 X(khc2c_register) (p, hc2cbdft2_20, &desc, HC2C_VIA_DFT);
|
cannam@167
|
596 }
|
cannam@167
|
597 #else
|
cannam@167
|
598
|
cannam@167
|
599 /* Generated by: ../../../genfft/gen_hc2cdft.native -compact -variables 4 -pipeline-latency 4 -sign 1 -n 20 -dif -name hc2cbdft2_20 -include rdft/scalar/hc2cb.h */
|
cannam@167
|
600
|
cannam@167
|
601 /*
|
cannam@167
|
602 * This function contains 286 FP additions, 124 FP multiplications,
|
cannam@167
|
603 * (or, 224 additions, 62 multiplications, 62 fused multiply/add),
|
cannam@167
|
604 * 89 stack variables, 4 constants, and 80 memory accesses
|
cannam@167
|
605 */
|
cannam@167
|
606 #include "rdft/scalar/hc2cb.h"
|
cannam@167
|
607
|
cannam@167
|
608 static void hc2cbdft2_20(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms)
|
cannam@167
|
609 {
|
cannam@167
|
610 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
|
cannam@167
|
611 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
|
cannam@167
|
612 DK(KP587785252, +0.587785252292473129168705954639072768597652438);
|
cannam@167
|
613 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
|
cannam@167
|
614 {
|
cannam@167
|
615 INT m;
|
cannam@167
|
616 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
|
617 E T7, T3N, T4a, T16, T1G, T3g, T3D, T26, T1k, T3A, T3B, T1v, T2e, T48, T47;
|
cannam@167
|
618 E T2d, T1L, T43, T40, T1K, T2l, T3t, T2m, T3w, T3n, T3p, TC, T2b, T4d, T4f;
|
cannam@167
|
619 E T23, T2j, T1B, T1H, T3U, T3W, T3G, T3I, T11, T17;
|
cannam@167
|
620 {
|
cannam@167
|
621 E T3, T1C, T15, T24, T6, T12, T1F, T25;
|
cannam@167
|
622 {
|
cannam@167
|
623 E T1, T2, T13, T14;
|
cannam@167
|
624 T1 = Rp[0];
|
cannam@167
|
625 T2 = Rm[WS(rs, 9)];
|
cannam@167
|
626 T3 = T1 + T2;
|
cannam@167
|
627 T1C = T1 - T2;
|
cannam@167
|
628 T13 = Ip[0];
|
cannam@167
|
629 T14 = Im[WS(rs, 9)];
|
cannam@167
|
630 T15 = T13 + T14;
|
cannam@167
|
631 T24 = T13 - T14;
|
cannam@167
|
632 }
|
cannam@167
|
633 {
|
cannam@167
|
634 E T4, T5, T1D, T1E;
|
cannam@167
|
635 T4 = Rp[WS(rs, 5)];
|
cannam@167
|
636 T5 = Rm[WS(rs, 4)];
|
cannam@167
|
637 T6 = T4 + T5;
|
cannam@167
|
638 T12 = T4 - T5;
|
cannam@167
|
639 T1D = Ip[WS(rs, 5)];
|
cannam@167
|
640 T1E = Im[WS(rs, 4)];
|
cannam@167
|
641 T1F = T1D + T1E;
|
cannam@167
|
642 T25 = T1D - T1E;
|
cannam@167
|
643 }
|
cannam@167
|
644 T7 = T3 + T6;
|
cannam@167
|
645 T3N = T15 - T12;
|
cannam@167
|
646 T4a = T1C + T1F;
|
cannam@167
|
647 T16 = T12 + T15;
|
cannam@167
|
648 T1G = T1C - T1F;
|
cannam@167
|
649 T3g = T3 - T6;
|
cannam@167
|
650 T3D = T24 - T25;
|
cannam@167
|
651 T26 = T24 + T25;
|
cannam@167
|
652 }
|
cannam@167
|
653 {
|
cannam@167
|
654 E Te, T3O, T3Y, TJ, T1e, T3h, T3r, T1R, TA, T3S, T42, TZ, T1u, T3l, T3v;
|
cannam@167
|
655 E T21, Tl, T3P, T3Z, TO, T1j, T3i, T3s, T1U, Tt, T3R, T41, TU, T1p, T3k;
|
cannam@167
|
656 E T3u, T1Y;
|
cannam@167
|
657 {
|
cannam@167
|
658 E Ta, T1a, TI, T1P, Td, TF, T1d, T1Q;
|
cannam@167
|
659 {
|
cannam@167
|
660 E T8, T9, TG, TH;
|
cannam@167
|
661 T8 = Rp[WS(rs, 4)];
|
cannam@167
|
662 T9 = Rm[WS(rs, 5)];
|
cannam@167
|
663 Ta = T8 + T9;
|
cannam@167
|
664 T1a = T8 - T9;
|
cannam@167
|
665 TG = Ip[WS(rs, 4)];
|
cannam@167
|
666 TH = Im[WS(rs, 5)];
|
cannam@167
|
667 TI = TG + TH;
|
cannam@167
|
668 T1P = TG - TH;
|
cannam@167
|
669 }
|
cannam@167
|
670 {
|
cannam@167
|
671 E Tb, Tc, T1b, T1c;
|
cannam@167
|
672 Tb = Rp[WS(rs, 9)];
|
cannam@167
|
673 Tc = Rm[0];
|
cannam@167
|
674 Td = Tb + Tc;
|
cannam@167
|
675 TF = Tb - Tc;
|
cannam@167
|
676 T1b = Ip[WS(rs, 9)];
|
cannam@167
|
677 T1c = Im[0];
|
cannam@167
|
678 T1d = T1b + T1c;
|
cannam@167
|
679 T1Q = T1b - T1c;
|
cannam@167
|
680 }
|
cannam@167
|
681 Te = Ta + Td;
|
cannam@167
|
682 T3O = TI - TF;
|
cannam@167
|
683 T3Y = T1a + T1d;
|
cannam@167
|
684 TJ = TF + TI;
|
cannam@167
|
685 T1e = T1a - T1d;
|
cannam@167
|
686 T3h = Ta - Td;
|
cannam@167
|
687 T3r = T1P - T1Q;
|
cannam@167
|
688 T1R = T1P + T1Q;
|
cannam@167
|
689 }
|
cannam@167
|
690 {
|
cannam@167
|
691 E Tw, T1q, TY, T1Z, Tz, TV, T1t, T20;
|
cannam@167
|
692 {
|
cannam@167
|
693 E Tu, Tv, TW, TX;
|
cannam@167
|
694 Tu = Rm[WS(rs, 7)];
|
cannam@167
|
695 Tv = Rp[WS(rs, 2)];
|
cannam@167
|
696 Tw = Tu + Tv;
|
cannam@167
|
697 T1q = Tu - Tv;
|
cannam@167
|
698 TW = Im[WS(rs, 7)];
|
cannam@167
|
699 TX = Ip[WS(rs, 2)];
|
cannam@167
|
700 TY = TW + TX;
|
cannam@167
|
701 T1Z = TX - TW;
|
cannam@167
|
702 }
|
cannam@167
|
703 {
|
cannam@167
|
704 E Tx, Ty, T1r, T1s;
|
cannam@167
|
705 Tx = Rm[WS(rs, 2)];
|
cannam@167
|
706 Ty = Rp[WS(rs, 7)];
|
cannam@167
|
707 Tz = Tx + Ty;
|
cannam@167
|
708 TV = Tx - Ty;
|
cannam@167
|
709 T1r = Im[WS(rs, 2)];
|
cannam@167
|
710 T1s = Ip[WS(rs, 7)];
|
cannam@167
|
711 T1t = T1r + T1s;
|
cannam@167
|
712 T20 = T1s - T1r;
|
cannam@167
|
713 }
|
cannam@167
|
714 TA = Tw + Tz;
|
cannam@167
|
715 T3S = TV + TY;
|
cannam@167
|
716 T42 = T1q - T1t;
|
cannam@167
|
717 TZ = TV - TY;
|
cannam@167
|
718 T1u = T1q + T1t;
|
cannam@167
|
719 T3l = Tw - Tz;
|
cannam@167
|
720 T3v = T1Z - T20;
|
cannam@167
|
721 T21 = T1Z + T20;
|
cannam@167
|
722 }
|
cannam@167
|
723 {
|
cannam@167
|
724 E Th, T1f, TN, T1S, Tk, TK, T1i, T1T;
|
cannam@167
|
725 {
|
cannam@167
|
726 E Tf, Tg, TL, TM;
|
cannam@167
|
727 Tf = Rm[WS(rs, 3)];
|
cannam@167
|
728 Tg = Rp[WS(rs, 6)];
|
cannam@167
|
729 Th = Tf + Tg;
|
cannam@167
|
730 T1f = Tf - Tg;
|
cannam@167
|
731 TL = Im[WS(rs, 3)];
|
cannam@167
|
732 TM = Ip[WS(rs, 6)];
|
cannam@167
|
733 TN = TL + TM;
|
cannam@167
|
734 T1S = TM - TL;
|
cannam@167
|
735 }
|
cannam@167
|
736 {
|
cannam@167
|
737 E Ti, Tj, T1g, T1h;
|
cannam@167
|
738 Ti = Rp[WS(rs, 1)];
|
cannam@167
|
739 Tj = Rm[WS(rs, 8)];
|
cannam@167
|
740 Tk = Ti + Tj;
|
cannam@167
|
741 TK = Ti - Tj;
|
cannam@167
|
742 T1g = Ip[WS(rs, 1)];
|
cannam@167
|
743 T1h = Im[WS(rs, 8)];
|
cannam@167
|
744 T1i = T1g + T1h;
|
cannam@167
|
745 T1T = T1g - T1h;
|
cannam@167
|
746 }
|
cannam@167
|
747 Tl = Th + Tk;
|
cannam@167
|
748 T3P = TK + TN;
|
cannam@167
|
749 T3Z = T1f + T1i;
|
cannam@167
|
750 TO = TK - TN;
|
cannam@167
|
751 T1j = T1f - T1i;
|
cannam@167
|
752 T3i = Th - Tk;
|
cannam@167
|
753 T3s = T1S - T1T;
|
cannam@167
|
754 T1U = T1S + T1T;
|
cannam@167
|
755 }
|
cannam@167
|
756 {
|
cannam@167
|
757 E Tp, T1l, TT, T1W, Ts, TQ, T1o, T1X;
|
cannam@167
|
758 {
|
cannam@167
|
759 E Tn, To, TR, TS;
|
cannam@167
|
760 Tn = Rp[WS(rs, 8)];
|
cannam@167
|
761 To = Rm[WS(rs, 1)];
|
cannam@167
|
762 Tp = Tn + To;
|
cannam@167
|
763 T1l = Tn - To;
|
cannam@167
|
764 TR = Ip[WS(rs, 8)];
|
cannam@167
|
765 TS = Im[WS(rs, 1)];
|
cannam@167
|
766 TT = TR + TS;
|
cannam@167
|
767 T1W = TR - TS;
|
cannam@167
|
768 }
|
cannam@167
|
769 {
|
cannam@167
|
770 E Tq, Tr, T1m, T1n;
|
cannam@167
|
771 Tq = Rm[WS(rs, 6)];
|
cannam@167
|
772 Tr = Rp[WS(rs, 3)];
|
cannam@167
|
773 Ts = Tq + Tr;
|
cannam@167
|
774 TQ = Tq - Tr;
|
cannam@167
|
775 T1m = Im[WS(rs, 6)];
|
cannam@167
|
776 T1n = Ip[WS(rs, 3)];
|
cannam@167
|
777 T1o = T1m + T1n;
|
cannam@167
|
778 T1X = T1n - T1m;
|
cannam@167
|
779 }
|
cannam@167
|
780 Tt = Tp + Ts;
|
cannam@167
|
781 T3R = TT - TQ;
|
cannam@167
|
782 T41 = T1l - T1o;
|
cannam@167
|
783 TU = TQ + TT;
|
cannam@167
|
784 T1p = T1l + T1o;
|
cannam@167
|
785 T3k = Tp - Ts;
|
cannam@167
|
786 T3u = T1W - T1X;
|
cannam@167
|
787 T1Y = T1W + T1X;
|
cannam@167
|
788 }
|
cannam@167
|
789 T1k = T1e - T1j;
|
cannam@167
|
790 T3A = T3h - T3i;
|
cannam@167
|
791 T3B = T3k - T3l;
|
cannam@167
|
792 T1v = T1p - T1u;
|
cannam@167
|
793 T2e = T1Y - T21;
|
cannam@167
|
794 T48 = T3R + T3S;
|
cannam@167
|
795 T47 = T3O + T3P;
|
cannam@167
|
796 T2d = T1R - T1U;
|
cannam@167
|
797 T1L = TU - TZ;
|
cannam@167
|
798 T43 = T41 - T42;
|
cannam@167
|
799 T40 = T3Y - T3Z;
|
cannam@167
|
800 T1K = TJ - TO;
|
cannam@167
|
801 T2l = Te - Tl;
|
cannam@167
|
802 T3t = T3r - T3s;
|
cannam@167
|
803 T2m = Tt - TA;
|
cannam@167
|
804 T3w = T3u - T3v;
|
cannam@167
|
805 {
|
cannam@167
|
806 E T3j, T3m, Tm, TB;
|
cannam@167
|
807 T3j = T3h + T3i;
|
cannam@167
|
808 T3m = T3k + T3l;
|
cannam@167
|
809 T3n = T3j + T3m;
|
cannam@167
|
810 T3p = KP559016994 * (T3j - T3m);
|
cannam@167
|
811 Tm = Te + Tl;
|
cannam@167
|
812 TB = Tt + TA;
|
cannam@167
|
813 TC = Tm + TB;
|
cannam@167
|
814 T2b = KP559016994 * (Tm - TB);
|
cannam@167
|
815 }
|
cannam@167
|
816 {
|
cannam@167
|
817 E T4b, T4c, T3Q, T3T;
|
cannam@167
|
818 T4b = T3Y + T3Z;
|
cannam@167
|
819 T4c = T41 + T42;
|
cannam@167
|
820 T4d = T4b + T4c;
|
cannam@167
|
821 T4f = KP559016994 * (T4b - T4c);
|
cannam@167
|
822 {
|
cannam@167
|
823 E T1V, T22, T1z, T1A;
|
cannam@167
|
824 T1V = T1R + T1U;
|
cannam@167
|
825 T22 = T1Y + T21;
|
cannam@167
|
826 T23 = T1V + T22;
|
cannam@167
|
827 T2j = KP559016994 * (T1V - T22);
|
cannam@167
|
828 T1z = T1e + T1j;
|
cannam@167
|
829 T1A = T1p + T1u;
|
cannam@167
|
830 T1B = KP559016994 * (T1z - T1A);
|
cannam@167
|
831 T1H = T1z + T1A;
|
cannam@167
|
832 }
|
cannam@167
|
833 T3Q = T3O - T3P;
|
cannam@167
|
834 T3T = T3R - T3S;
|
cannam@167
|
835 T3U = T3Q + T3T;
|
cannam@167
|
836 T3W = KP559016994 * (T3Q - T3T);
|
cannam@167
|
837 {
|
cannam@167
|
838 E T3E, T3F, TP, T10;
|
cannam@167
|
839 T3E = T3r + T3s;
|
cannam@167
|
840 T3F = T3u + T3v;
|
cannam@167
|
841 T3G = T3E + T3F;
|
cannam@167
|
842 T3I = KP559016994 * (T3E - T3F);
|
cannam@167
|
843 TP = TJ + TO;
|
cannam@167
|
844 T10 = TU + TZ;
|
cannam@167
|
845 T11 = KP559016994 * (TP - T10);
|
cannam@167
|
846 T17 = TP + T10;
|
cannam@167
|
847 }
|
cannam@167
|
848 }
|
cannam@167
|
849 }
|
cannam@167
|
850 {
|
cannam@167
|
851 E TD, T27, T3c, T3e, T2o, T36, T2A, T2U, T1N, T2Z, T2t, T2J, T1x, T2X, T2r;
|
cannam@167
|
852 E T2F, T2g, T34, T2y, T2Q;
|
cannam@167
|
853 TD = T7 + TC;
|
cannam@167
|
854 T27 = T23 + T26;
|
cannam@167
|
855 {
|
cannam@167
|
856 E T39, T3b, T38, T3a;
|
cannam@167
|
857 T39 = T16 + T17;
|
cannam@167
|
858 T3b = T1H + T1G;
|
cannam@167
|
859 T38 = W[8];
|
cannam@167
|
860 T3a = W[9];
|
cannam@167
|
861 T3c = FMA(T38, T39, T3a * T3b);
|
cannam@167
|
862 T3e = FNMS(T3a, T39, T38 * T3b);
|
cannam@167
|
863 }
|
cannam@167
|
864 {
|
cannam@167
|
865 E T2n, T2S, T2k, T2T, T2i;
|
cannam@167
|
866 T2n = FNMS(KP951056516, T2m, KP587785252 * T2l);
|
cannam@167
|
867 T2S = FMA(KP951056516, T2l, KP587785252 * T2m);
|
cannam@167
|
868 T2i = FNMS(KP250000000, T23, T26);
|
cannam@167
|
869 T2k = T2i - T2j;
|
cannam@167
|
870 T2T = T2j + T2i;
|
cannam@167
|
871 T2o = T2k - T2n;
|
cannam@167
|
872 T36 = T2T - T2S;
|
cannam@167
|
873 T2A = T2n + T2k;
|
cannam@167
|
874 T2U = T2S + T2T;
|
cannam@167
|
875 }
|
cannam@167
|
876 {
|
cannam@167
|
877 E T1M, T2H, T1J, T2I, T1I;
|
cannam@167
|
878 T1M = FMA(KP951056516, T1K, KP587785252 * T1L);
|
cannam@167
|
879 T2H = FNMS(KP951056516, T1L, KP587785252 * T1K);
|
cannam@167
|
880 T1I = FNMS(KP250000000, T1H, T1G);
|
cannam@167
|
881 T1J = T1B + T1I;
|
cannam@167
|
882 T2I = T1I - T1B;
|
cannam@167
|
883 T1N = T1J - T1M;
|
cannam@167
|
884 T2Z = T2I - T2H;
|
cannam@167
|
885 T2t = T1M + T1J;
|
cannam@167
|
886 T2J = T2H + T2I;
|
cannam@167
|
887 }
|
cannam@167
|
888 {
|
cannam@167
|
889 E T1w, T2E, T19, T2D, T18;
|
cannam@167
|
890 T1w = FMA(KP951056516, T1k, KP587785252 * T1v);
|
cannam@167
|
891 T2E = FNMS(KP951056516, T1v, KP587785252 * T1k);
|
cannam@167
|
892 T18 = FNMS(KP250000000, T17, T16);
|
cannam@167
|
893 T19 = T11 + T18;
|
cannam@167
|
894 T2D = T18 - T11;
|
cannam@167
|
895 T1x = T19 + T1w;
|
cannam@167
|
896 T2X = T2D + T2E;
|
cannam@167
|
897 T2r = T19 - T1w;
|
cannam@167
|
898 T2F = T2D - T2E;
|
cannam@167
|
899 }
|
cannam@167
|
900 {
|
cannam@167
|
901 E T2f, T2P, T2c, T2O, T2a;
|
cannam@167
|
902 T2f = FNMS(KP951056516, T2e, KP587785252 * T2d);
|
cannam@167
|
903 T2P = FMA(KP951056516, T2d, KP587785252 * T2e);
|
cannam@167
|
904 T2a = FNMS(KP250000000, TC, T7);
|
cannam@167
|
905 T2c = T2a - T2b;
|
cannam@167
|
906 T2O = T2b + T2a;
|
cannam@167
|
907 T2g = T2c + T2f;
|
cannam@167
|
908 T34 = T2O + T2P;
|
cannam@167
|
909 T2y = T2c - T2f;
|
cannam@167
|
910 T2Q = T2O - T2P;
|
cannam@167
|
911 }
|
cannam@167
|
912 {
|
cannam@167
|
913 E T1O, T28, TE, T1y;
|
cannam@167
|
914 TE = W[0];
|
cannam@167
|
915 T1y = W[1];
|
cannam@167
|
916 T1O = FMA(TE, T1x, T1y * T1N);
|
cannam@167
|
917 T28 = FNMS(T1y, T1x, TE * T1N);
|
cannam@167
|
918 Rp[0] = TD - T1O;
|
cannam@167
|
919 Ip[0] = T27 + T28;
|
cannam@167
|
920 Rm[0] = TD + T1O;
|
cannam@167
|
921 Im[0] = T28 - T27;
|
cannam@167
|
922 }
|
cannam@167
|
923 {
|
cannam@167
|
924 E T37, T3d, T33, T35;
|
cannam@167
|
925 T33 = W[6];
|
cannam@167
|
926 T35 = W[7];
|
cannam@167
|
927 T37 = FNMS(T35, T36, T33 * T34);
|
cannam@167
|
928 T3d = FMA(T35, T34, T33 * T36);
|
cannam@167
|
929 Rp[WS(rs, 2)] = T37 - T3c;
|
cannam@167
|
930 Ip[WS(rs, 2)] = T3d + T3e;
|
cannam@167
|
931 Rm[WS(rs, 2)] = T37 + T3c;
|
cannam@167
|
932 Im[WS(rs, 2)] = T3e - T3d;
|
cannam@167
|
933 }
|
cannam@167
|
934 {
|
cannam@167
|
935 E T2p, T2v, T2u, T2w;
|
cannam@167
|
936 {
|
cannam@167
|
937 E T29, T2h, T2q, T2s;
|
cannam@167
|
938 T29 = W[14];
|
cannam@167
|
939 T2h = W[15];
|
cannam@167
|
940 T2p = FNMS(T2h, T2o, T29 * T2g);
|
cannam@167
|
941 T2v = FMA(T2h, T2g, T29 * T2o);
|
cannam@167
|
942 T2q = W[16];
|
cannam@167
|
943 T2s = W[17];
|
cannam@167
|
944 T2u = FMA(T2q, T2r, T2s * T2t);
|
cannam@167
|
945 T2w = FNMS(T2s, T2r, T2q * T2t);
|
cannam@167
|
946 }
|
cannam@167
|
947 Rp[WS(rs, 4)] = T2p - T2u;
|
cannam@167
|
948 Ip[WS(rs, 4)] = T2v + T2w;
|
cannam@167
|
949 Rm[WS(rs, 4)] = T2p + T2u;
|
cannam@167
|
950 Im[WS(rs, 4)] = T2w - T2v;
|
cannam@167
|
951 }
|
cannam@167
|
952 {
|
cannam@167
|
953 E T2B, T2L, T2K, T2M;
|
cannam@167
|
954 {
|
cannam@167
|
955 E T2x, T2z, T2C, T2G;
|
cannam@167
|
956 T2x = W[22];
|
cannam@167
|
957 T2z = W[23];
|
cannam@167
|
958 T2B = FNMS(T2z, T2A, T2x * T2y);
|
cannam@167
|
959 T2L = FMA(T2z, T2y, T2x * T2A);
|
cannam@167
|
960 T2C = W[24];
|
cannam@167
|
961 T2G = W[25];
|
cannam@167
|
962 T2K = FMA(T2C, T2F, T2G * T2J);
|
cannam@167
|
963 T2M = FNMS(T2G, T2F, T2C * T2J);
|
cannam@167
|
964 }
|
cannam@167
|
965 Rp[WS(rs, 6)] = T2B - T2K;
|
cannam@167
|
966 Ip[WS(rs, 6)] = T2L + T2M;
|
cannam@167
|
967 Rm[WS(rs, 6)] = T2B + T2K;
|
cannam@167
|
968 Im[WS(rs, 6)] = T2M - T2L;
|
cannam@167
|
969 }
|
cannam@167
|
970 {
|
cannam@167
|
971 E T2V, T31, T30, T32;
|
cannam@167
|
972 {
|
cannam@167
|
973 E T2N, T2R, T2W, T2Y;
|
cannam@167
|
974 T2N = W[30];
|
cannam@167
|
975 T2R = W[31];
|
cannam@167
|
976 T2V = FNMS(T2R, T2U, T2N * T2Q);
|
cannam@167
|
977 T31 = FMA(T2R, T2Q, T2N * T2U);
|
cannam@167
|
978 T2W = W[32];
|
cannam@167
|
979 T2Y = W[33];
|
cannam@167
|
980 T30 = FMA(T2W, T2X, T2Y * T2Z);
|
cannam@167
|
981 T32 = FNMS(T2Y, T2X, T2W * T2Z);
|
cannam@167
|
982 }
|
cannam@167
|
983 Rp[WS(rs, 8)] = T2V - T30;
|
cannam@167
|
984 Ip[WS(rs, 8)] = T31 + T32;
|
cannam@167
|
985 Rm[WS(rs, 8)] = T2V + T30;
|
cannam@167
|
986 Im[WS(rs, 8)] = T32 - T31;
|
cannam@167
|
987 }
|
cannam@167
|
988 }
|
cannam@167
|
989 {
|
cannam@167
|
990 E T4F, T4P, T5c, T5e, T3y, T54, T4o, T4S, T4h, T4Z, T4x, T4N, T45, T4X, T4v;
|
cannam@167
|
991 E T4J, T3K, T56, T4s, T4U;
|
cannam@167
|
992 {
|
cannam@167
|
993 E T4C, T4E, T4B, T4D;
|
cannam@167
|
994 T4C = T3g + T3n;
|
cannam@167
|
995 T4E = T3G + T3D;
|
cannam@167
|
996 T4B = W[18];
|
cannam@167
|
997 T4D = W[19];
|
cannam@167
|
998 T4F = FNMS(T4D, T4E, T4B * T4C);
|
cannam@167
|
999 T4P = FMA(T4D, T4C, T4B * T4E);
|
cannam@167
|
1000 }
|
cannam@167
|
1001 {
|
cannam@167
|
1002 E T59, T5b, T58, T5a;
|
cannam@167
|
1003 T59 = T3N + T3U;
|
cannam@167
|
1004 T5b = T4d + T4a;
|
cannam@167
|
1005 T58 = W[28];
|
cannam@167
|
1006 T5a = W[29];
|
cannam@167
|
1007 T5c = FMA(T58, T59, T5a * T5b);
|
cannam@167
|
1008 T5e = FNMS(T5a, T59, T58 * T5b);
|
cannam@167
|
1009 }
|
cannam@167
|
1010 {
|
cannam@167
|
1011 E T3x, T4n, T3q, T4m, T3o;
|
cannam@167
|
1012 T3x = FNMS(KP951056516, T3w, KP587785252 * T3t);
|
cannam@167
|
1013 T4n = FMA(KP951056516, T3t, KP587785252 * T3w);
|
cannam@167
|
1014 T3o = FNMS(KP250000000, T3n, T3g);
|
cannam@167
|
1015 T3q = T3o - T3p;
|
cannam@167
|
1016 T4m = T3p + T3o;
|
cannam@167
|
1017 T3y = T3q - T3x;
|
cannam@167
|
1018 T54 = T4m + T4n;
|
cannam@167
|
1019 T4o = T4m - T4n;
|
cannam@167
|
1020 T4S = T3q + T3x;
|
cannam@167
|
1021 }
|
cannam@167
|
1022 {
|
cannam@167
|
1023 E T49, T4M, T4g, T4L, T4e;
|
cannam@167
|
1024 T49 = FNMS(KP951056516, T48, KP587785252 * T47);
|
cannam@167
|
1025 T4M = FMA(KP951056516, T47, KP587785252 * T48);
|
cannam@167
|
1026 T4e = FNMS(KP250000000, T4d, T4a);
|
cannam@167
|
1027 T4g = T4e - T4f;
|
cannam@167
|
1028 T4L = T4f + T4e;
|
cannam@167
|
1029 T4h = T49 + T4g;
|
cannam@167
|
1030 T4Z = T4M + T4L;
|
cannam@167
|
1031 T4x = T4g - T49;
|
cannam@167
|
1032 T4N = T4L - T4M;
|
cannam@167
|
1033 }
|
cannam@167
|
1034 {
|
cannam@167
|
1035 E T44, T4I, T3X, T4H, T3V;
|
cannam@167
|
1036 T44 = FNMS(KP951056516, T43, KP587785252 * T40);
|
cannam@167
|
1037 T4I = FMA(KP951056516, T40, KP587785252 * T43);
|
cannam@167
|
1038 T3V = FNMS(KP250000000, T3U, T3N);
|
cannam@167
|
1039 T3X = T3V - T3W;
|
cannam@167
|
1040 T4H = T3W + T3V;
|
cannam@167
|
1041 T45 = T3X - T44;
|
cannam@167
|
1042 T4X = T4H - T4I;
|
cannam@167
|
1043 T4v = T3X + T44;
|
cannam@167
|
1044 T4J = T4H + T4I;
|
cannam@167
|
1045 }
|
cannam@167
|
1046 {
|
cannam@167
|
1047 E T3C, T4q, T3J, T4r, T3H;
|
cannam@167
|
1048 T3C = FNMS(KP951056516, T3B, KP587785252 * T3A);
|
cannam@167
|
1049 T4q = FMA(KP951056516, T3A, KP587785252 * T3B);
|
cannam@167
|
1050 T3H = FNMS(KP250000000, T3G, T3D);
|
cannam@167
|
1051 T3J = T3H - T3I;
|
cannam@167
|
1052 T4r = T3I + T3H;
|
cannam@167
|
1053 T3K = T3C + T3J;
|
cannam@167
|
1054 T56 = T4r - T4q;
|
cannam@167
|
1055 T4s = T4q + T4r;
|
cannam@167
|
1056 T4U = T3J - T3C;
|
cannam@167
|
1057 }
|
cannam@167
|
1058 {
|
cannam@167
|
1059 E T4O, T4Q, T4G, T4K;
|
cannam@167
|
1060 T4G = W[20];
|
cannam@167
|
1061 T4K = W[21];
|
cannam@167
|
1062 T4O = FMA(T4G, T4J, T4K * T4N);
|
cannam@167
|
1063 T4Q = FNMS(T4K, T4J, T4G * T4N);
|
cannam@167
|
1064 Rp[WS(rs, 5)] = T4F - T4O;
|
cannam@167
|
1065 Ip[WS(rs, 5)] = T4P + T4Q;
|
cannam@167
|
1066 Rm[WS(rs, 5)] = T4F + T4O;
|
cannam@167
|
1067 Im[WS(rs, 5)] = T4Q - T4P;
|
cannam@167
|
1068 }
|
cannam@167
|
1069 {
|
cannam@167
|
1070 E T57, T5d, T53, T55;
|
cannam@167
|
1071 T53 = W[26];
|
cannam@167
|
1072 T55 = W[27];
|
cannam@167
|
1073 T57 = FNMS(T55, T56, T53 * T54);
|
cannam@167
|
1074 T5d = FMA(T55, T54, T53 * T56);
|
cannam@167
|
1075 Rp[WS(rs, 7)] = T57 - T5c;
|
cannam@167
|
1076 Ip[WS(rs, 7)] = T5d + T5e;
|
cannam@167
|
1077 Rm[WS(rs, 7)] = T57 + T5c;
|
cannam@167
|
1078 Im[WS(rs, 7)] = T5e - T5d;
|
cannam@167
|
1079 }
|
cannam@167
|
1080 {
|
cannam@167
|
1081 E T3L, T4j, T4i, T4k;
|
cannam@167
|
1082 {
|
cannam@167
|
1083 E T3f, T3z, T3M, T46;
|
cannam@167
|
1084 T3f = W[2];
|
cannam@167
|
1085 T3z = W[3];
|
cannam@167
|
1086 T3L = FNMS(T3z, T3K, T3f * T3y);
|
cannam@167
|
1087 T4j = FMA(T3z, T3y, T3f * T3K);
|
cannam@167
|
1088 T3M = W[4];
|
cannam@167
|
1089 T46 = W[5];
|
cannam@167
|
1090 T4i = FMA(T3M, T45, T46 * T4h);
|
cannam@167
|
1091 T4k = FNMS(T46, T45, T3M * T4h);
|
cannam@167
|
1092 }
|
cannam@167
|
1093 Rp[WS(rs, 1)] = T3L - T4i;
|
cannam@167
|
1094 Ip[WS(rs, 1)] = T4j + T4k;
|
cannam@167
|
1095 Rm[WS(rs, 1)] = T3L + T4i;
|
cannam@167
|
1096 Im[WS(rs, 1)] = T4k - T4j;
|
cannam@167
|
1097 }
|
cannam@167
|
1098 {
|
cannam@167
|
1099 E T4t, T4z, T4y, T4A;
|
cannam@167
|
1100 {
|
cannam@167
|
1101 E T4l, T4p, T4u, T4w;
|
cannam@167
|
1102 T4l = W[10];
|
cannam@167
|
1103 T4p = W[11];
|
cannam@167
|
1104 T4t = FNMS(T4p, T4s, T4l * T4o);
|
cannam@167
|
1105 T4z = FMA(T4p, T4o, T4l * T4s);
|
cannam@167
|
1106 T4u = W[12];
|
cannam@167
|
1107 T4w = W[13];
|
cannam@167
|
1108 T4y = FMA(T4u, T4v, T4w * T4x);
|
cannam@167
|
1109 T4A = FNMS(T4w, T4v, T4u * T4x);
|
cannam@167
|
1110 }
|
cannam@167
|
1111 Rp[WS(rs, 3)] = T4t - T4y;
|
cannam@167
|
1112 Ip[WS(rs, 3)] = T4z + T4A;
|
cannam@167
|
1113 Rm[WS(rs, 3)] = T4t + T4y;
|
cannam@167
|
1114 Im[WS(rs, 3)] = T4A - T4z;
|
cannam@167
|
1115 }
|
cannam@167
|
1116 {
|
cannam@167
|
1117 E T4V, T51, T50, T52;
|
cannam@167
|
1118 {
|
cannam@167
|
1119 E T4R, T4T, T4W, T4Y;
|
cannam@167
|
1120 T4R = W[34];
|
cannam@167
|
1121 T4T = W[35];
|
cannam@167
|
1122 T4V = FNMS(T4T, T4U, T4R * T4S);
|
cannam@167
|
1123 T51 = FMA(T4T, T4S, T4R * T4U);
|
cannam@167
|
1124 T4W = W[36];
|
cannam@167
|
1125 T4Y = W[37];
|
cannam@167
|
1126 T50 = FMA(T4W, T4X, T4Y * T4Z);
|
cannam@167
|
1127 T52 = FNMS(T4Y, T4X, T4W * T4Z);
|
cannam@167
|
1128 }
|
cannam@167
|
1129 Rp[WS(rs, 9)] = T4V - T50;
|
cannam@167
|
1130 Ip[WS(rs, 9)] = T51 + T52;
|
cannam@167
|
1131 Rm[WS(rs, 9)] = T4V + T50;
|
cannam@167
|
1132 Im[WS(rs, 9)] = T52 - T51;
|
cannam@167
|
1133 }
|
cannam@167
|
1134 }
|
cannam@167
|
1135 }
|
cannam@167
|
1136 }
|
cannam@167
|
1137 }
|
cannam@167
|
1138
|
cannam@167
|
1139 static const tw_instr twinstr[] = {
|
cannam@167
|
1140 {TW_FULL, 1, 20},
|
cannam@167
|
1141 {TW_NEXT, 1, 0}
|
cannam@167
|
1142 };
|
cannam@167
|
1143
|
cannam@167
|
1144 static const hc2c_desc desc = { 20, "hc2cbdft2_20", twinstr, &GENUS, {224, 62, 62, 0} };
|
cannam@167
|
1145
|
cannam@167
|
1146 void X(codelet_hc2cbdft2_20) (planner *p) {
|
cannam@167
|
1147 X(khc2c_register) (p, hc2cbdft2_20, &desc, HC2C_VIA_DFT);
|
cannam@167
|
1148 }
|
cannam@167
|
1149 #endif
|