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 08:07:12 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 16 -dit -name hc2cfdft_16 -include rdft/scalar/hc2cf.h */
|
cannam@167
|
29
|
cannam@167
|
30 /*
|
cannam@167
|
31 * This function contains 206 FP additions, 132 FP multiplications,
|
cannam@167
|
32 * (or, 136 additions, 62 multiplications, 70 fused multiply/add),
|
cannam@167
|
33 * 67 stack variables, 4 constants, and 64 memory accesses
|
cannam@167
|
34 */
|
cannam@167
|
35 #include "rdft/scalar/hc2cf.h"
|
cannam@167
|
36
|
cannam@167
|
37 static void hc2cfdft_16(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(KP923879532, +0.923879532511286756128183189396788286822416626);
|
cannam@167
|
40 DK(KP414213562, +0.414213562373095048801688724209698078569671875);
|
cannam@167
|
41 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
cannam@167
|
42 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
|
cannam@167
|
43 {
|
cannam@167
|
44 INT m;
|
cannam@167
|
45 for (m = mb, W = W + ((mb - 1) * 30); m < me; m = m + 1, Rp = Rp + ms, Ip = Ip + ms, Rm = Rm - ms, Im = Im - ms, W = W + 30, MAKE_VOLATILE_STRIDE(64, rs)) {
|
cannam@167
|
46 E T1f, T2e, T1c, T2g, T1K, T3D, T2W, T3H, TR, T2j, T2R, T3E, T11, T2l, T1v;
|
cannam@167
|
47 E T3G, Ta, T2p, Tk, T2r, T3o, T3p, T1Y, T3z, T2G, T3w, Tv, T2u, TF, T2w;
|
cannam@167
|
48 E T3r, T3s, T2b, T3A, T2L, T3x;
|
cannam@167
|
49 {
|
cannam@167
|
50 E T1d, T1e, T1I, T16, T1A, T1D, T1E, T1C, T1G, T1H, T2U, T1b, T1z, T2S, T1w;
|
cannam@167
|
51 E T1y, T14, T15;
|
cannam@167
|
52 T1d = Ip[0];
|
cannam@167
|
53 T1e = Im[0];
|
cannam@167
|
54 T1I = T1d + T1e;
|
cannam@167
|
55 T14 = Ip[WS(rs, 4)];
|
cannam@167
|
56 T15 = Im[WS(rs, 4)];
|
cannam@167
|
57 T16 = T14 - T15;
|
cannam@167
|
58 T1A = T14 + T15;
|
cannam@167
|
59 {
|
cannam@167
|
60 E T1F, T19, T1a, T1x;
|
cannam@167
|
61 T1D = Rm[0];
|
cannam@167
|
62 T1E = Rp[0];
|
cannam@167
|
63 T1F = T1D - T1E;
|
cannam@167
|
64 T1C = W[0];
|
cannam@167
|
65 T1G = T1C * T1F;
|
cannam@167
|
66 T1H = W[1];
|
cannam@167
|
67 T2U = T1H * T1F;
|
cannam@167
|
68 T19 = Rp[WS(rs, 4)];
|
cannam@167
|
69 T1a = Rm[WS(rs, 4)];
|
cannam@167
|
70 T1x = T1a - T19;
|
cannam@167
|
71 T1b = T19 + T1a;
|
cannam@167
|
72 T1z = W[17];
|
cannam@167
|
73 T2S = T1z * T1x;
|
cannam@167
|
74 T1w = W[16];
|
cannam@167
|
75 T1y = T1w * T1x;
|
cannam@167
|
76 }
|
cannam@167
|
77 T1f = T1d - T1e;
|
cannam@167
|
78 T2e = T1E + T1D;
|
cannam@167
|
79 {
|
cannam@167
|
80 E T17, T2f, T13, T18;
|
cannam@167
|
81 T13 = W[14];
|
cannam@167
|
82 T17 = T13 * T16;
|
cannam@167
|
83 T2f = T13 * T1b;
|
cannam@167
|
84 T18 = W[15];
|
cannam@167
|
85 T1c = FNMS(T18, T1b, T17);
|
cannam@167
|
86 T2g = FMA(T18, T16, T2f);
|
cannam@167
|
87 }
|
cannam@167
|
88 {
|
cannam@167
|
89 E T1B, T1J, T2T, T2V;
|
cannam@167
|
90 T1B = FNMS(T1z, T1A, T1y);
|
cannam@167
|
91 T1J = FNMS(T1H, T1I, T1G);
|
cannam@167
|
92 T1K = T1B + T1J;
|
cannam@167
|
93 T3D = T1J - T1B;
|
cannam@167
|
94 T2T = FMA(T1w, T1A, T2S);
|
cannam@167
|
95 T2V = FMA(T1C, T1I, T2U);
|
cannam@167
|
96 T2W = T2T + T2V;
|
cannam@167
|
97 T3H = T2V - T2T;
|
cannam@167
|
98 }
|
cannam@167
|
99 }
|
cannam@167
|
100 {
|
cannam@167
|
101 E TL, T1n, TQ, T1m, T2N, T1j, T1l, TV, T1t, T10, T1s, T2P, T1p, T1r;
|
cannam@167
|
102 {
|
cannam@167
|
103 E TJ, TK, TO, TP, T1k;
|
cannam@167
|
104 TJ = Ip[WS(rs, 2)];
|
cannam@167
|
105 TK = Im[WS(rs, 2)];
|
cannam@167
|
106 TL = TJ - TK;
|
cannam@167
|
107 T1n = TJ + TK;
|
cannam@167
|
108 TO = Rp[WS(rs, 2)];
|
cannam@167
|
109 TP = Rm[WS(rs, 2)];
|
cannam@167
|
110 T1k = TP - TO;
|
cannam@167
|
111 TQ = TO + TP;
|
cannam@167
|
112 T1m = W[9];
|
cannam@167
|
113 T2N = T1m * T1k;
|
cannam@167
|
114 T1j = W[8];
|
cannam@167
|
115 T1l = T1j * T1k;
|
cannam@167
|
116 }
|
cannam@167
|
117 {
|
cannam@167
|
118 E TT, TU, TY, TZ, T1q;
|
cannam@167
|
119 TT = Ip[WS(rs, 6)];
|
cannam@167
|
120 TU = Im[WS(rs, 6)];
|
cannam@167
|
121 TV = TT - TU;
|
cannam@167
|
122 T1t = TT + TU;
|
cannam@167
|
123 TY = Rp[WS(rs, 6)];
|
cannam@167
|
124 TZ = Rm[WS(rs, 6)];
|
cannam@167
|
125 T1q = TZ - TY;
|
cannam@167
|
126 T10 = TY + TZ;
|
cannam@167
|
127 T1s = W[25];
|
cannam@167
|
128 T2P = T1s * T1q;
|
cannam@167
|
129 T1p = W[24];
|
cannam@167
|
130 T1r = T1p * T1q;
|
cannam@167
|
131 }
|
cannam@167
|
132 {
|
cannam@167
|
133 E T2O, T2Q, T1o, T1u;
|
cannam@167
|
134 {
|
cannam@167
|
135 E TM, T2i, TI, TN;
|
cannam@167
|
136 TI = W[6];
|
cannam@167
|
137 TM = TI * TL;
|
cannam@167
|
138 T2i = TI * TQ;
|
cannam@167
|
139 TN = W[7];
|
cannam@167
|
140 TR = FNMS(TN, TQ, TM);
|
cannam@167
|
141 T2j = FMA(TN, TL, T2i);
|
cannam@167
|
142 }
|
cannam@167
|
143 T2O = FMA(T1j, T1n, T2N);
|
cannam@167
|
144 T2Q = FMA(T1p, T1t, T2P);
|
cannam@167
|
145 T2R = T2O + T2Q;
|
cannam@167
|
146 T3E = T2O - T2Q;
|
cannam@167
|
147 {
|
cannam@167
|
148 E TW, T2k, TS, TX;
|
cannam@167
|
149 TS = W[22];
|
cannam@167
|
150 TW = TS * TV;
|
cannam@167
|
151 T2k = TS * T10;
|
cannam@167
|
152 TX = W[23];
|
cannam@167
|
153 T11 = FNMS(TX, T10, TW);
|
cannam@167
|
154 T2l = FMA(TX, TV, T2k);
|
cannam@167
|
155 }
|
cannam@167
|
156 T1o = FNMS(T1m, T1n, T1l);
|
cannam@167
|
157 T1u = FNMS(T1s, T1t, T1r);
|
cannam@167
|
158 T1v = T1o + T1u;
|
cannam@167
|
159 T3G = T1o - T1u;
|
cannam@167
|
160 }
|
cannam@167
|
161 }
|
cannam@167
|
162 {
|
cannam@167
|
163 E T4, T1Q, T9, T1N, T5, T2o, T1O, T2C, Te, T1W, Tj, T1T, Tf, T2q, T1U;
|
cannam@167
|
164 E T2E, T6, Tg;
|
cannam@167
|
165 {
|
cannam@167
|
166 E T1, T1M, Tb, T1S;
|
cannam@167
|
167 {
|
cannam@167
|
168 E T2, T3, T7, T8;
|
cannam@167
|
169 T2 = Ip[WS(rs, 1)];
|
cannam@167
|
170 T3 = Im[WS(rs, 1)];
|
cannam@167
|
171 T4 = T2 - T3;
|
cannam@167
|
172 T1Q = T2 + T3;
|
cannam@167
|
173 T7 = Rp[WS(rs, 1)];
|
cannam@167
|
174 T8 = Rm[WS(rs, 1)];
|
cannam@167
|
175 T9 = T7 + T8;
|
cannam@167
|
176 T1N = T7 - T8;
|
cannam@167
|
177 }
|
cannam@167
|
178 T1 = W[2];
|
cannam@167
|
179 T5 = T1 * T4;
|
cannam@167
|
180 T2o = T1 * T9;
|
cannam@167
|
181 T1M = W[4];
|
cannam@167
|
182 T1O = T1M * T1N;
|
cannam@167
|
183 T2C = T1M * T1Q;
|
cannam@167
|
184 {
|
cannam@167
|
185 E Tc, Td, Th, Ti;
|
cannam@167
|
186 Tc = Ip[WS(rs, 5)];
|
cannam@167
|
187 Td = Im[WS(rs, 5)];
|
cannam@167
|
188 Te = Tc - Td;
|
cannam@167
|
189 T1W = Tc + Td;
|
cannam@167
|
190 Th = Rp[WS(rs, 5)];
|
cannam@167
|
191 Ti = Rm[WS(rs, 5)];
|
cannam@167
|
192 Tj = Th + Ti;
|
cannam@167
|
193 T1T = Th - Ti;
|
cannam@167
|
194 }
|
cannam@167
|
195 Tb = W[18];
|
cannam@167
|
196 Tf = Tb * Te;
|
cannam@167
|
197 T2q = Tb * Tj;
|
cannam@167
|
198 T1S = W[20];
|
cannam@167
|
199 T1U = T1S * T1T;
|
cannam@167
|
200 T2E = T1S * T1W;
|
cannam@167
|
201 }
|
cannam@167
|
202 T6 = W[3];
|
cannam@167
|
203 Ta = FNMS(T6, T9, T5);
|
cannam@167
|
204 T2p = FMA(T6, T4, T2o);
|
cannam@167
|
205 Tg = W[19];
|
cannam@167
|
206 Tk = FNMS(Tg, Tj, Tf);
|
cannam@167
|
207 T2r = FMA(Tg, Te, T2q);
|
cannam@167
|
208 T3o = Ta - Tk;
|
cannam@167
|
209 T3p = T2p - T2r;
|
cannam@167
|
210 {
|
cannam@167
|
211 E T1R, T2D, T1X, T2F, T1P, T1V;
|
cannam@167
|
212 T1P = W[5];
|
cannam@167
|
213 T1R = FMA(T1P, T1Q, T1O);
|
cannam@167
|
214 T2D = FNMS(T1P, T1N, T2C);
|
cannam@167
|
215 T1V = W[21];
|
cannam@167
|
216 T1X = FMA(T1V, T1W, T1U);
|
cannam@167
|
217 T2F = FNMS(T1V, T1T, T2E);
|
cannam@167
|
218 T1Y = T1R + T1X;
|
cannam@167
|
219 T3z = T1X - T1R;
|
cannam@167
|
220 T2G = T2D + T2F;
|
cannam@167
|
221 T3w = T2F - T2D;
|
cannam@167
|
222 }
|
cannam@167
|
223 }
|
cannam@167
|
224 {
|
cannam@167
|
225 E Tp, T23, Tu, T20, Tq, T2t, T21, T2H, Tz, T29, TE, T26, TA, T2v, T27;
|
cannam@167
|
226 E T2J, Tr, TB;
|
cannam@167
|
227 {
|
cannam@167
|
228 E Tm, T1Z, Tw, T25;
|
cannam@167
|
229 {
|
cannam@167
|
230 E Tn, To, Ts, Tt;
|
cannam@167
|
231 Tn = Ip[WS(rs, 7)];
|
cannam@167
|
232 To = Im[WS(rs, 7)];
|
cannam@167
|
233 Tp = Tn - To;
|
cannam@167
|
234 T23 = Tn + To;
|
cannam@167
|
235 Ts = Rp[WS(rs, 7)];
|
cannam@167
|
236 Tt = Rm[WS(rs, 7)];
|
cannam@167
|
237 Tu = Ts + Tt;
|
cannam@167
|
238 T20 = Ts - Tt;
|
cannam@167
|
239 }
|
cannam@167
|
240 Tm = W[26];
|
cannam@167
|
241 Tq = Tm * Tp;
|
cannam@167
|
242 T2t = Tm * Tu;
|
cannam@167
|
243 T1Z = W[28];
|
cannam@167
|
244 T21 = T1Z * T20;
|
cannam@167
|
245 T2H = T1Z * T23;
|
cannam@167
|
246 {
|
cannam@167
|
247 E Tx, Ty, TC, TD;
|
cannam@167
|
248 Tx = Ip[WS(rs, 3)];
|
cannam@167
|
249 Ty = Im[WS(rs, 3)];
|
cannam@167
|
250 Tz = Tx - Ty;
|
cannam@167
|
251 T29 = Tx + Ty;
|
cannam@167
|
252 TC = Rp[WS(rs, 3)];
|
cannam@167
|
253 TD = Rm[WS(rs, 3)];
|
cannam@167
|
254 TE = TC + TD;
|
cannam@167
|
255 T26 = TC - TD;
|
cannam@167
|
256 }
|
cannam@167
|
257 Tw = W[10];
|
cannam@167
|
258 TA = Tw * Tz;
|
cannam@167
|
259 T2v = Tw * TE;
|
cannam@167
|
260 T25 = W[12];
|
cannam@167
|
261 T27 = T25 * T26;
|
cannam@167
|
262 T2J = T25 * T29;
|
cannam@167
|
263 }
|
cannam@167
|
264 Tr = W[27];
|
cannam@167
|
265 Tv = FNMS(Tr, Tu, Tq);
|
cannam@167
|
266 T2u = FMA(Tr, Tp, T2t);
|
cannam@167
|
267 TB = W[11];
|
cannam@167
|
268 TF = FNMS(TB, TE, TA);
|
cannam@167
|
269 T2w = FMA(TB, Tz, T2v);
|
cannam@167
|
270 T3r = T2u - T2w;
|
cannam@167
|
271 T3s = Tv - TF;
|
cannam@167
|
272 {
|
cannam@167
|
273 E T24, T2I, T2a, T2K, T22, T28;
|
cannam@167
|
274 T22 = W[29];
|
cannam@167
|
275 T24 = FMA(T22, T23, T21);
|
cannam@167
|
276 T2I = FNMS(T22, T20, T2H);
|
cannam@167
|
277 T28 = W[13];
|
cannam@167
|
278 T2a = FMA(T28, T29, T27);
|
cannam@167
|
279 T2K = FNMS(T28, T26, T2J);
|
cannam@167
|
280 T2b = T24 + T2a;
|
cannam@167
|
281 T3A = T2I - T2K;
|
cannam@167
|
282 T2L = T2I + T2K;
|
cannam@167
|
283 T3x = T2a - T24;
|
cannam@167
|
284 }
|
cannam@167
|
285 }
|
cannam@167
|
286 {
|
cannam@167
|
287 E TH, T3c, T36, T3g, T39, T3h, T1h, T32, T2d, T2A, T2y, T31, T2Y, T30, T2n;
|
cannam@167
|
288 E T3b;
|
cannam@167
|
289 {
|
cannam@167
|
290 E Tl, TG, T34, T35;
|
cannam@167
|
291 Tl = Ta + Tk;
|
cannam@167
|
292 TG = Tv + TF;
|
cannam@167
|
293 TH = Tl + TG;
|
cannam@167
|
294 T3c = Tl - TG;
|
cannam@167
|
295 T34 = T2L - T2G;
|
cannam@167
|
296 T35 = T1Y - T2b;
|
cannam@167
|
297 T36 = T34 + T35;
|
cannam@167
|
298 T3g = T34 - T35;
|
cannam@167
|
299 }
|
cannam@167
|
300 {
|
cannam@167
|
301 E T37, T38, T12, T1g;
|
cannam@167
|
302 T37 = T1K - T1v;
|
cannam@167
|
303 T38 = T2W - T2R;
|
cannam@167
|
304 T39 = T37 - T38;
|
cannam@167
|
305 T3h = T37 + T38;
|
cannam@167
|
306 T12 = TR + T11;
|
cannam@167
|
307 T1g = T1c + T1f;
|
cannam@167
|
308 T1h = T12 + T1g;
|
cannam@167
|
309 T32 = T1g - T12;
|
cannam@167
|
310 }
|
cannam@167
|
311 {
|
cannam@167
|
312 E T1L, T2c, T2s, T2x;
|
cannam@167
|
313 T1L = T1v + T1K;
|
cannam@167
|
314 T2c = T1Y + T2b;
|
cannam@167
|
315 T2d = T1L - T2c;
|
cannam@167
|
316 T2A = T2c + T1L;
|
cannam@167
|
317 T2s = T2p + T2r;
|
cannam@167
|
318 T2x = T2u + T2w;
|
cannam@167
|
319 T2y = T2s + T2x;
|
cannam@167
|
320 T31 = T2x - T2s;
|
cannam@167
|
321 }
|
cannam@167
|
322 {
|
cannam@167
|
323 E T2M, T2X, T2h, T2m;
|
cannam@167
|
324 T2M = T2G + T2L;
|
cannam@167
|
325 T2X = T2R + T2W;
|
cannam@167
|
326 T2Y = T2M - T2X;
|
cannam@167
|
327 T30 = T2M + T2X;
|
cannam@167
|
328 T2h = T2e + T2g;
|
cannam@167
|
329 T2m = T2j + T2l;
|
cannam@167
|
330 T2n = T2h + T2m;
|
cannam@167
|
331 T3b = T2h - T2m;
|
cannam@167
|
332 }
|
cannam@167
|
333 {
|
cannam@167
|
334 E T1i, T2Z, T2z, T2B;
|
cannam@167
|
335 T1i = TH + T1h;
|
cannam@167
|
336 Ip[0] = KP500000000 * (T1i + T2d);
|
cannam@167
|
337 Im[WS(rs, 7)] = KP500000000 * (T2d - T1i);
|
cannam@167
|
338 T2Z = T2n + T2y;
|
cannam@167
|
339 Rm[WS(rs, 7)] = KP500000000 * (T2Z - T30);
|
cannam@167
|
340 Rp[0] = KP500000000 * (T2Z + T30);
|
cannam@167
|
341 T2z = T2n - T2y;
|
cannam@167
|
342 Rm[WS(rs, 3)] = KP500000000 * (T2z - T2A);
|
cannam@167
|
343 Rp[WS(rs, 4)] = KP500000000 * (T2z + T2A);
|
cannam@167
|
344 T2B = T1h - TH;
|
cannam@167
|
345 Ip[WS(rs, 4)] = KP500000000 * (T2B + T2Y);
|
cannam@167
|
346 Im[WS(rs, 3)] = KP500000000 * (T2Y - T2B);
|
cannam@167
|
347 }
|
cannam@167
|
348 {
|
cannam@167
|
349 E T33, T3a, T3j, T3k;
|
cannam@167
|
350 T33 = T31 + T32;
|
cannam@167
|
351 T3a = T36 + T39;
|
cannam@167
|
352 Ip[WS(rs, 2)] = KP500000000 * (FMA(KP707106781, T3a, T33));
|
cannam@167
|
353 Im[WS(rs, 5)] = -(KP500000000 * (FNMS(KP707106781, T3a, T33)));
|
cannam@167
|
354 T3j = T3b + T3c;
|
cannam@167
|
355 T3k = T3g + T3h;
|
cannam@167
|
356 Rm[WS(rs, 5)] = KP500000000 * (FNMS(KP707106781, T3k, T3j));
|
cannam@167
|
357 Rp[WS(rs, 2)] = KP500000000 * (FMA(KP707106781, T3k, T3j));
|
cannam@167
|
358 }
|
cannam@167
|
359 {
|
cannam@167
|
360 E T3d, T3e, T3f, T3i;
|
cannam@167
|
361 T3d = T3b - T3c;
|
cannam@167
|
362 T3e = T39 - T36;
|
cannam@167
|
363 Rm[WS(rs, 1)] = KP500000000 * (FNMS(KP707106781, T3e, T3d));
|
cannam@167
|
364 Rp[WS(rs, 6)] = KP500000000 * (FMA(KP707106781, T3e, T3d));
|
cannam@167
|
365 T3f = T32 - T31;
|
cannam@167
|
366 T3i = T3g - T3h;
|
cannam@167
|
367 Ip[WS(rs, 6)] = KP500000000 * (FMA(KP707106781, T3i, T3f));
|
cannam@167
|
368 Im[WS(rs, 1)] = -(KP500000000 * (FNMS(KP707106781, T3i, T3f)));
|
cannam@167
|
369 }
|
cannam@167
|
370 }
|
cannam@167
|
371 {
|
cannam@167
|
372 E T3n, T3Z, T44, T4e, T47, T4f, T3u, T4a, T3C, T3U, T3N, T49, T3Q, T40, T3J;
|
cannam@167
|
373 E T3V;
|
cannam@167
|
374 {
|
cannam@167
|
375 E T3l, T3m, T42, T43;
|
cannam@167
|
376 T3l = T1f - T1c;
|
cannam@167
|
377 T3m = T2j - T2l;
|
cannam@167
|
378 T3n = T3l - T3m;
|
cannam@167
|
379 T3Z = T3m + T3l;
|
cannam@167
|
380 T42 = T3w - T3x;
|
cannam@167
|
381 T43 = T3A - T3z;
|
cannam@167
|
382 T44 = FMA(KP414213562, T43, T42);
|
cannam@167
|
383 T4e = FNMS(KP414213562, T42, T43);
|
cannam@167
|
384 }
|
cannam@167
|
385 {
|
cannam@167
|
386 E T45, T46, T3q, T3t;
|
cannam@167
|
387 T45 = T3E + T3D;
|
cannam@167
|
388 T46 = T3H - T3G;
|
cannam@167
|
389 T47 = FMA(KP414213562, T46, T45);
|
cannam@167
|
390 T4f = FNMS(KP414213562, T45, T46);
|
cannam@167
|
391 T3q = T3o - T3p;
|
cannam@167
|
392 T3t = T3r + T3s;
|
cannam@167
|
393 T3u = T3q + T3t;
|
cannam@167
|
394 T4a = T3q - T3t;
|
cannam@167
|
395 }
|
cannam@167
|
396 {
|
cannam@167
|
397 E T3y, T3B, T3L, T3M;
|
cannam@167
|
398 T3y = T3w + T3x;
|
cannam@167
|
399 T3B = T3z + T3A;
|
cannam@167
|
400 T3C = FMA(KP414213562, T3B, T3y);
|
cannam@167
|
401 T3U = FNMS(KP414213562, T3y, T3B);
|
cannam@167
|
402 T3L = T2e - T2g;
|
cannam@167
|
403 T3M = TR - T11;
|
cannam@167
|
404 T3N = T3L + T3M;
|
cannam@167
|
405 T49 = T3L - T3M;
|
cannam@167
|
406 }
|
cannam@167
|
407 {
|
cannam@167
|
408 E T3O, T3P, T3F, T3I;
|
cannam@167
|
409 T3O = T3p + T3o;
|
cannam@167
|
410 T3P = T3r - T3s;
|
cannam@167
|
411 T3Q = T3O + T3P;
|
cannam@167
|
412 T40 = T3P - T3O;
|
cannam@167
|
413 T3F = T3D - T3E;
|
cannam@167
|
414 T3I = T3G + T3H;
|
cannam@167
|
415 T3J = FNMS(KP414213562, T3I, T3F);
|
cannam@167
|
416 T3V = FMA(KP414213562, T3F, T3I);
|
cannam@167
|
417 }
|
cannam@167
|
418 {
|
cannam@167
|
419 E T3v, T3K, T3X, T3Y;
|
cannam@167
|
420 T3v = FMA(KP707106781, T3u, T3n);
|
cannam@167
|
421 T3K = T3C + T3J;
|
cannam@167
|
422 Ip[WS(rs, 1)] = KP500000000 * (FMA(KP923879532, T3K, T3v));
|
cannam@167
|
423 Im[WS(rs, 6)] = -(KP500000000 * (FNMS(KP923879532, T3K, T3v)));
|
cannam@167
|
424 T3X = FMA(KP707106781, T3Q, T3N);
|
cannam@167
|
425 T3Y = T3U + T3V;
|
cannam@167
|
426 Rm[WS(rs, 6)] = KP500000000 * (FNMS(KP923879532, T3Y, T3X));
|
cannam@167
|
427 Rp[WS(rs, 1)] = KP500000000 * (FMA(KP923879532, T3Y, T3X));
|
cannam@167
|
428 }
|
cannam@167
|
429 {
|
cannam@167
|
430 E T3R, T3S, T3T, T3W;
|
cannam@167
|
431 T3R = FNMS(KP707106781, T3Q, T3N);
|
cannam@167
|
432 T3S = T3J - T3C;
|
cannam@167
|
433 Rm[WS(rs, 2)] = KP500000000 * (FNMS(KP923879532, T3S, T3R));
|
cannam@167
|
434 Rp[WS(rs, 5)] = KP500000000 * (FMA(KP923879532, T3S, T3R));
|
cannam@167
|
435 T3T = FNMS(KP707106781, T3u, T3n);
|
cannam@167
|
436 T3W = T3U - T3V;
|
cannam@167
|
437 Ip[WS(rs, 5)] = KP500000000 * (FMA(KP923879532, T3W, T3T));
|
cannam@167
|
438 Im[WS(rs, 2)] = -(KP500000000 * (FNMS(KP923879532, T3W, T3T)));
|
cannam@167
|
439 }
|
cannam@167
|
440 {
|
cannam@167
|
441 E T41, T48, T4h, T4i;
|
cannam@167
|
442 T41 = FNMS(KP707106781, T40, T3Z);
|
cannam@167
|
443 T48 = T44 - T47;
|
cannam@167
|
444 Ip[WS(rs, 7)] = KP500000000 * (FMA(KP923879532, T48, T41));
|
cannam@167
|
445 Im[0] = -(KP500000000 * (FNMS(KP923879532, T48, T41)));
|
cannam@167
|
446 T4h = FNMS(KP707106781, T4a, T49);
|
cannam@167
|
447 T4i = T4e + T4f;
|
cannam@167
|
448 Rp[WS(rs, 7)] = KP500000000 * (FNMS(KP923879532, T4i, T4h));
|
cannam@167
|
449 Rm[0] = KP500000000 * (FMA(KP923879532, T4i, T4h));
|
cannam@167
|
450 }
|
cannam@167
|
451 {
|
cannam@167
|
452 E T4b, T4c, T4d, T4g;
|
cannam@167
|
453 T4b = FMA(KP707106781, T4a, T49);
|
cannam@167
|
454 T4c = T44 + T47;
|
cannam@167
|
455 Rm[WS(rs, 4)] = KP500000000 * (FNMS(KP923879532, T4c, T4b));
|
cannam@167
|
456 Rp[WS(rs, 3)] = KP500000000 * (FMA(KP923879532, T4c, T4b));
|
cannam@167
|
457 T4d = FMA(KP707106781, T40, T3Z);
|
cannam@167
|
458 T4g = T4e - T4f;
|
cannam@167
|
459 Ip[WS(rs, 3)] = KP500000000 * (FMA(KP923879532, T4g, T4d));
|
cannam@167
|
460 Im[WS(rs, 4)] = -(KP500000000 * (FNMS(KP923879532, T4g, T4d)));
|
cannam@167
|
461 }
|
cannam@167
|
462 }
|
cannam@167
|
463 }
|
cannam@167
|
464 }
|
cannam@167
|
465 }
|
cannam@167
|
466
|
cannam@167
|
467 static const tw_instr twinstr[] = {
|
cannam@167
|
468 {TW_FULL, 1, 16},
|
cannam@167
|
469 {TW_NEXT, 1, 0}
|
cannam@167
|
470 };
|
cannam@167
|
471
|
cannam@167
|
472 static const hc2c_desc desc = { 16, "hc2cfdft_16", twinstr, &GENUS, {136, 62, 70, 0} };
|
cannam@167
|
473
|
cannam@167
|
474 void X(codelet_hc2cfdft_16) (planner *p) {
|
cannam@167
|
475 X(khc2c_register) (p, hc2cfdft_16, &desc, HC2C_VIA_DFT);
|
cannam@167
|
476 }
|
cannam@167
|
477 #else
|
cannam@167
|
478
|
cannam@167
|
479 /* Generated by: ../../../genfft/gen_hc2cdft.native -compact -variables 4 -pipeline-latency 4 -n 16 -dit -name hc2cfdft_16 -include rdft/scalar/hc2cf.h */
|
cannam@167
|
480
|
cannam@167
|
481 /*
|
cannam@167
|
482 * This function contains 206 FP additions, 100 FP multiplications,
|
cannam@167
|
483 * (or, 168 additions, 62 multiplications, 38 fused multiply/add),
|
cannam@167
|
484 * 61 stack variables, 4 constants, and 64 memory accesses
|
cannam@167
|
485 */
|
cannam@167
|
486 #include "rdft/scalar/hc2cf.h"
|
cannam@167
|
487
|
cannam@167
|
488 static void hc2cfdft_16(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms)
|
cannam@167
|
489 {
|
cannam@167
|
490 DK(KP461939766, +0.461939766255643378064091594698394143411208313);
|
cannam@167
|
491 DK(KP191341716, +0.191341716182544885864229992015199433380672281);
|
cannam@167
|
492 DK(KP353553390, +0.353553390593273762200422181052424519642417969);
|
cannam@167
|
493 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
|
cannam@167
|
494 {
|
cannam@167
|
495 INT m;
|
cannam@167
|
496 for (m = mb, W = W + ((mb - 1) * 30); m < me; m = m + 1, Rp = Rp + ms, Ip = Ip + ms, Rm = Rm - ms, Im = Im - ms, W = W + 30, MAKE_VOLATILE_STRIDE(64, rs)) {
|
cannam@167
|
497 E T19, T3h, T21, T2Y, T1o, T3d, T2s, T39, TW, T3i, T24, T2Z, T1z, T3c, T2p;
|
cannam@167
|
498 E T3a, Tj, T2S, T28, T2R, T1L, T36, T2i, T32, TC, T2V, T2b, T2U, T1W, T35;
|
cannam@167
|
499 E T2l, T33;
|
cannam@167
|
500 {
|
cannam@167
|
501 E T10, T1m, T14, T1k, T18, T1h, T1f, T1Z;
|
cannam@167
|
502 {
|
cannam@167
|
503 E TY, TZ, T12, T13;
|
cannam@167
|
504 TY = Ip[WS(rs, 4)];
|
cannam@167
|
505 TZ = Im[WS(rs, 4)];
|
cannam@167
|
506 T10 = TY - TZ;
|
cannam@167
|
507 T1m = TY + TZ;
|
cannam@167
|
508 T12 = Rp[WS(rs, 4)];
|
cannam@167
|
509 T13 = Rm[WS(rs, 4)];
|
cannam@167
|
510 T14 = T12 + T13;
|
cannam@167
|
511 T1k = T12 - T13;
|
cannam@167
|
512 }
|
cannam@167
|
513 {
|
cannam@167
|
514 E T16, T17, T1d, T1e;
|
cannam@167
|
515 T16 = Ip[0];
|
cannam@167
|
516 T17 = Im[0];
|
cannam@167
|
517 T18 = T16 - T17;
|
cannam@167
|
518 T1h = T16 + T17;
|
cannam@167
|
519 T1d = Rm[0];
|
cannam@167
|
520 T1e = Rp[0];
|
cannam@167
|
521 T1f = T1d - T1e;
|
cannam@167
|
522 T1Z = T1e + T1d;
|
cannam@167
|
523 }
|
cannam@167
|
524 {
|
cannam@167
|
525 E T15, T20, TX, T11;
|
cannam@167
|
526 TX = W[14];
|
cannam@167
|
527 T11 = W[15];
|
cannam@167
|
528 T15 = FNMS(T11, T14, TX * T10);
|
cannam@167
|
529 T20 = FMA(TX, T14, T11 * T10);
|
cannam@167
|
530 T19 = T15 + T18;
|
cannam@167
|
531 T3h = T1Z - T20;
|
cannam@167
|
532 T21 = T1Z + T20;
|
cannam@167
|
533 T2Y = T18 - T15;
|
cannam@167
|
534 }
|
cannam@167
|
535 {
|
cannam@167
|
536 E T1i, T2r, T1n, T2q;
|
cannam@167
|
537 {
|
cannam@167
|
538 E T1c, T1g, T1j, T1l;
|
cannam@167
|
539 T1c = W[0];
|
cannam@167
|
540 T1g = W[1];
|
cannam@167
|
541 T1i = FNMS(T1g, T1h, T1c * T1f);
|
cannam@167
|
542 T2r = FMA(T1g, T1f, T1c * T1h);
|
cannam@167
|
543 T1j = W[16];
|
cannam@167
|
544 T1l = W[17];
|
cannam@167
|
545 T1n = FMA(T1j, T1k, T1l * T1m);
|
cannam@167
|
546 T2q = FNMS(T1l, T1k, T1j * T1m);
|
cannam@167
|
547 }
|
cannam@167
|
548 T1o = T1i - T1n;
|
cannam@167
|
549 T3d = T2r - T2q;
|
cannam@167
|
550 T2s = T2q + T2r;
|
cannam@167
|
551 T39 = T1n + T1i;
|
cannam@167
|
552 }
|
cannam@167
|
553 }
|
cannam@167
|
554 {
|
cannam@167
|
555 E TH, T1s, TL, T1q, TQ, T1x, TU, T1v;
|
cannam@167
|
556 {
|
cannam@167
|
557 E TF, TG, TJ, TK;
|
cannam@167
|
558 TF = Ip[WS(rs, 2)];
|
cannam@167
|
559 TG = Im[WS(rs, 2)];
|
cannam@167
|
560 TH = TF - TG;
|
cannam@167
|
561 T1s = TF + TG;
|
cannam@167
|
562 TJ = Rp[WS(rs, 2)];
|
cannam@167
|
563 TK = Rm[WS(rs, 2)];
|
cannam@167
|
564 TL = TJ + TK;
|
cannam@167
|
565 T1q = TJ - TK;
|
cannam@167
|
566 }
|
cannam@167
|
567 {
|
cannam@167
|
568 E TO, TP, TS, TT;
|
cannam@167
|
569 TO = Ip[WS(rs, 6)];
|
cannam@167
|
570 TP = Im[WS(rs, 6)];
|
cannam@167
|
571 TQ = TO - TP;
|
cannam@167
|
572 T1x = TO + TP;
|
cannam@167
|
573 TS = Rp[WS(rs, 6)];
|
cannam@167
|
574 TT = Rm[WS(rs, 6)];
|
cannam@167
|
575 TU = TS + TT;
|
cannam@167
|
576 T1v = TS - TT;
|
cannam@167
|
577 }
|
cannam@167
|
578 {
|
cannam@167
|
579 E TM, T22, TV, T23;
|
cannam@167
|
580 {
|
cannam@167
|
581 E TE, TI, TN, TR;
|
cannam@167
|
582 TE = W[6];
|
cannam@167
|
583 TI = W[7];
|
cannam@167
|
584 TM = FNMS(TI, TL, TE * TH);
|
cannam@167
|
585 T22 = FMA(TE, TL, TI * TH);
|
cannam@167
|
586 TN = W[22];
|
cannam@167
|
587 TR = W[23];
|
cannam@167
|
588 TV = FNMS(TR, TU, TN * TQ);
|
cannam@167
|
589 T23 = FMA(TN, TU, TR * TQ);
|
cannam@167
|
590 }
|
cannam@167
|
591 TW = TM + TV;
|
cannam@167
|
592 T3i = TM - TV;
|
cannam@167
|
593 T24 = T22 + T23;
|
cannam@167
|
594 T2Z = T22 - T23;
|
cannam@167
|
595 }
|
cannam@167
|
596 {
|
cannam@167
|
597 E T1t, T2n, T1y, T2o;
|
cannam@167
|
598 {
|
cannam@167
|
599 E T1p, T1r, T1u, T1w;
|
cannam@167
|
600 T1p = W[8];
|
cannam@167
|
601 T1r = W[9];
|
cannam@167
|
602 T1t = FMA(T1p, T1q, T1r * T1s);
|
cannam@167
|
603 T2n = FNMS(T1r, T1q, T1p * T1s);
|
cannam@167
|
604 T1u = W[24];
|
cannam@167
|
605 T1w = W[25];
|
cannam@167
|
606 T1y = FMA(T1u, T1v, T1w * T1x);
|
cannam@167
|
607 T2o = FNMS(T1w, T1v, T1u * T1x);
|
cannam@167
|
608 }
|
cannam@167
|
609 T1z = T1t + T1y;
|
cannam@167
|
610 T3c = T1y - T1t;
|
cannam@167
|
611 T2p = T2n + T2o;
|
cannam@167
|
612 T3a = T2n - T2o;
|
cannam@167
|
613 }
|
cannam@167
|
614 }
|
cannam@167
|
615 {
|
cannam@167
|
616 E T4, T1E, T8, T1C, Td, T1J, Th, T1H;
|
cannam@167
|
617 {
|
cannam@167
|
618 E T2, T3, T6, T7;
|
cannam@167
|
619 T2 = Ip[WS(rs, 1)];
|
cannam@167
|
620 T3 = Im[WS(rs, 1)];
|
cannam@167
|
621 T4 = T2 - T3;
|
cannam@167
|
622 T1E = T2 + T3;
|
cannam@167
|
623 T6 = Rp[WS(rs, 1)];
|
cannam@167
|
624 T7 = Rm[WS(rs, 1)];
|
cannam@167
|
625 T8 = T6 + T7;
|
cannam@167
|
626 T1C = T6 - T7;
|
cannam@167
|
627 }
|
cannam@167
|
628 {
|
cannam@167
|
629 E Tb, Tc, Tf, Tg;
|
cannam@167
|
630 Tb = Ip[WS(rs, 5)];
|
cannam@167
|
631 Tc = Im[WS(rs, 5)];
|
cannam@167
|
632 Td = Tb - Tc;
|
cannam@167
|
633 T1J = Tb + Tc;
|
cannam@167
|
634 Tf = Rp[WS(rs, 5)];
|
cannam@167
|
635 Tg = Rm[WS(rs, 5)];
|
cannam@167
|
636 Th = Tf + Tg;
|
cannam@167
|
637 T1H = Tf - Tg;
|
cannam@167
|
638 }
|
cannam@167
|
639 {
|
cannam@167
|
640 E T9, T26, Ti, T27;
|
cannam@167
|
641 {
|
cannam@167
|
642 E T1, T5, Ta, Te;
|
cannam@167
|
643 T1 = W[2];
|
cannam@167
|
644 T5 = W[3];
|
cannam@167
|
645 T9 = FNMS(T5, T8, T1 * T4);
|
cannam@167
|
646 T26 = FMA(T1, T8, T5 * T4);
|
cannam@167
|
647 Ta = W[18];
|
cannam@167
|
648 Te = W[19];
|
cannam@167
|
649 Ti = FNMS(Te, Th, Ta * Td);
|
cannam@167
|
650 T27 = FMA(Ta, Th, Te * Td);
|
cannam@167
|
651 }
|
cannam@167
|
652 Tj = T9 + Ti;
|
cannam@167
|
653 T2S = T26 - T27;
|
cannam@167
|
654 T28 = T26 + T27;
|
cannam@167
|
655 T2R = T9 - Ti;
|
cannam@167
|
656 }
|
cannam@167
|
657 {
|
cannam@167
|
658 E T1F, T2g, T1K, T2h;
|
cannam@167
|
659 {
|
cannam@167
|
660 E T1B, T1D, T1G, T1I;
|
cannam@167
|
661 T1B = W[4];
|
cannam@167
|
662 T1D = W[5];
|
cannam@167
|
663 T1F = FMA(T1B, T1C, T1D * T1E);
|
cannam@167
|
664 T2g = FNMS(T1D, T1C, T1B * T1E);
|
cannam@167
|
665 T1G = W[20];
|
cannam@167
|
666 T1I = W[21];
|
cannam@167
|
667 T1K = FMA(T1G, T1H, T1I * T1J);
|
cannam@167
|
668 T2h = FNMS(T1I, T1H, T1G * T1J);
|
cannam@167
|
669 }
|
cannam@167
|
670 T1L = T1F + T1K;
|
cannam@167
|
671 T36 = T2g - T2h;
|
cannam@167
|
672 T2i = T2g + T2h;
|
cannam@167
|
673 T32 = T1K - T1F;
|
cannam@167
|
674 }
|
cannam@167
|
675 }
|
cannam@167
|
676 {
|
cannam@167
|
677 E Tn, T1P, Tr, T1N, Tw, T1U, TA, T1S;
|
cannam@167
|
678 {
|
cannam@167
|
679 E Tl, Tm, Tp, Tq;
|
cannam@167
|
680 Tl = Ip[WS(rs, 7)];
|
cannam@167
|
681 Tm = Im[WS(rs, 7)];
|
cannam@167
|
682 Tn = Tl - Tm;
|
cannam@167
|
683 T1P = Tl + Tm;
|
cannam@167
|
684 Tp = Rp[WS(rs, 7)];
|
cannam@167
|
685 Tq = Rm[WS(rs, 7)];
|
cannam@167
|
686 Tr = Tp + Tq;
|
cannam@167
|
687 T1N = Tp - Tq;
|
cannam@167
|
688 }
|
cannam@167
|
689 {
|
cannam@167
|
690 E Tu, Tv, Ty, Tz;
|
cannam@167
|
691 Tu = Ip[WS(rs, 3)];
|
cannam@167
|
692 Tv = Im[WS(rs, 3)];
|
cannam@167
|
693 Tw = Tu - Tv;
|
cannam@167
|
694 T1U = Tu + Tv;
|
cannam@167
|
695 Ty = Rp[WS(rs, 3)];
|
cannam@167
|
696 Tz = Rm[WS(rs, 3)];
|
cannam@167
|
697 TA = Ty + Tz;
|
cannam@167
|
698 T1S = Ty - Tz;
|
cannam@167
|
699 }
|
cannam@167
|
700 {
|
cannam@167
|
701 E Ts, T29, TB, T2a;
|
cannam@167
|
702 {
|
cannam@167
|
703 E Tk, To, Tt, Tx;
|
cannam@167
|
704 Tk = W[26];
|
cannam@167
|
705 To = W[27];
|
cannam@167
|
706 Ts = FNMS(To, Tr, Tk * Tn);
|
cannam@167
|
707 T29 = FMA(Tk, Tr, To * Tn);
|
cannam@167
|
708 Tt = W[10];
|
cannam@167
|
709 Tx = W[11];
|
cannam@167
|
710 TB = FNMS(Tx, TA, Tt * Tw);
|
cannam@167
|
711 T2a = FMA(Tt, TA, Tx * Tw);
|
cannam@167
|
712 }
|
cannam@167
|
713 TC = Ts + TB;
|
cannam@167
|
714 T2V = Ts - TB;
|
cannam@167
|
715 T2b = T29 + T2a;
|
cannam@167
|
716 T2U = T29 - T2a;
|
cannam@167
|
717 }
|
cannam@167
|
718 {
|
cannam@167
|
719 E T1Q, T2j, T1V, T2k;
|
cannam@167
|
720 {
|
cannam@167
|
721 E T1M, T1O, T1R, T1T;
|
cannam@167
|
722 T1M = W[28];
|
cannam@167
|
723 T1O = W[29];
|
cannam@167
|
724 T1Q = FMA(T1M, T1N, T1O * T1P);
|
cannam@167
|
725 T2j = FNMS(T1O, T1N, T1M * T1P);
|
cannam@167
|
726 T1R = W[12];
|
cannam@167
|
727 T1T = W[13];
|
cannam@167
|
728 T1V = FMA(T1R, T1S, T1T * T1U);
|
cannam@167
|
729 T2k = FNMS(T1T, T1S, T1R * T1U);
|
cannam@167
|
730 }
|
cannam@167
|
731 T1W = T1Q + T1V;
|
cannam@167
|
732 T35 = T1V - T1Q;
|
cannam@167
|
733 T2l = T2j + T2k;
|
cannam@167
|
734 T33 = T2j - T2k;
|
cannam@167
|
735 }
|
cannam@167
|
736 }
|
cannam@167
|
737 {
|
cannam@167
|
738 E T1b, T2f, T2u, T2w, T1Y, T2e, T2d, T2v;
|
cannam@167
|
739 {
|
cannam@167
|
740 E TD, T1a, T2m, T2t;
|
cannam@167
|
741 TD = Tj + TC;
|
cannam@167
|
742 T1a = TW + T19;
|
cannam@167
|
743 T1b = TD + T1a;
|
cannam@167
|
744 T2f = T1a - TD;
|
cannam@167
|
745 T2m = T2i + T2l;
|
cannam@167
|
746 T2t = T2p + T2s;
|
cannam@167
|
747 T2u = T2m - T2t;
|
cannam@167
|
748 T2w = T2m + T2t;
|
cannam@167
|
749 }
|
cannam@167
|
750 {
|
cannam@167
|
751 E T1A, T1X, T25, T2c;
|
cannam@167
|
752 T1A = T1o - T1z;
|
cannam@167
|
753 T1X = T1L + T1W;
|
cannam@167
|
754 T1Y = T1A - T1X;
|
cannam@167
|
755 T2e = T1X + T1A;
|
cannam@167
|
756 T25 = T21 + T24;
|
cannam@167
|
757 T2c = T28 + T2b;
|
cannam@167
|
758 T2d = T25 - T2c;
|
cannam@167
|
759 T2v = T25 + T2c;
|
cannam@167
|
760 }
|
cannam@167
|
761 Ip[0] = KP500000000 * (T1b + T1Y);
|
cannam@167
|
762 Rp[0] = KP500000000 * (T2v + T2w);
|
cannam@167
|
763 Im[WS(rs, 7)] = KP500000000 * (T1Y - T1b);
|
cannam@167
|
764 Rm[WS(rs, 7)] = KP500000000 * (T2v - T2w);
|
cannam@167
|
765 Rm[WS(rs, 3)] = KP500000000 * (T2d - T2e);
|
cannam@167
|
766 Im[WS(rs, 3)] = KP500000000 * (T2u - T2f);
|
cannam@167
|
767 Rp[WS(rs, 4)] = KP500000000 * (T2d + T2e);
|
cannam@167
|
768 Ip[WS(rs, 4)] = KP500000000 * (T2f + T2u);
|
cannam@167
|
769 }
|
cannam@167
|
770 {
|
cannam@167
|
771 E T2z, T2L, T2J, T2P, T2C, T2M, T2F, T2N;
|
cannam@167
|
772 {
|
cannam@167
|
773 E T2x, T2y, T2H, T2I;
|
cannam@167
|
774 T2x = T2b - T28;
|
cannam@167
|
775 T2y = T19 - TW;
|
cannam@167
|
776 T2z = KP500000000 * (T2x + T2y);
|
cannam@167
|
777 T2L = KP500000000 * (T2y - T2x);
|
cannam@167
|
778 T2H = T21 - T24;
|
cannam@167
|
779 T2I = Tj - TC;
|
cannam@167
|
780 T2J = KP500000000 * (T2H - T2I);
|
cannam@167
|
781 T2P = KP500000000 * (T2H + T2I);
|
cannam@167
|
782 }
|
cannam@167
|
783 {
|
cannam@167
|
784 E T2A, T2B, T2D, T2E;
|
cannam@167
|
785 T2A = T2l - T2i;
|
cannam@167
|
786 T2B = T1L - T1W;
|
cannam@167
|
787 T2C = T2A + T2B;
|
cannam@167
|
788 T2M = T2A - T2B;
|
cannam@167
|
789 T2D = T1z + T1o;
|
cannam@167
|
790 T2E = T2s - T2p;
|
cannam@167
|
791 T2F = T2D - T2E;
|
cannam@167
|
792 T2N = T2D + T2E;
|
cannam@167
|
793 }
|
cannam@167
|
794 {
|
cannam@167
|
795 E T2G, T2Q, T2K, T2O;
|
cannam@167
|
796 T2G = KP353553390 * (T2C + T2F);
|
cannam@167
|
797 Ip[WS(rs, 2)] = T2z + T2G;
|
cannam@167
|
798 Im[WS(rs, 5)] = T2G - T2z;
|
cannam@167
|
799 T2Q = KP353553390 * (T2M + T2N);
|
cannam@167
|
800 Rm[WS(rs, 5)] = T2P - T2Q;
|
cannam@167
|
801 Rp[WS(rs, 2)] = T2P + T2Q;
|
cannam@167
|
802 T2K = KP353553390 * (T2F - T2C);
|
cannam@167
|
803 Rm[WS(rs, 1)] = T2J - T2K;
|
cannam@167
|
804 Rp[WS(rs, 6)] = T2J + T2K;
|
cannam@167
|
805 T2O = KP353553390 * (T2M - T2N);
|
cannam@167
|
806 Ip[WS(rs, 6)] = T2L + T2O;
|
cannam@167
|
807 Im[WS(rs, 1)] = T2O - T2L;
|
cannam@167
|
808 }
|
cannam@167
|
809 }
|
cannam@167
|
810 {
|
cannam@167
|
811 E T30, T3w, T3F, T3j, T2X, T3G, T3D, T3L, T3m, T3v, T38, T3q, T3A, T3K, T3f;
|
cannam@167
|
812 E T3r;
|
cannam@167
|
813 {
|
cannam@167
|
814 E T2T, T2W, T34, T37;
|
cannam@167
|
815 T30 = KP500000000 * (T2Y - T2Z);
|
cannam@167
|
816 T3w = KP500000000 * (T2Z + T2Y);
|
cannam@167
|
817 T3F = KP500000000 * (T3h - T3i);
|
cannam@167
|
818 T3j = KP500000000 * (T3h + T3i);
|
cannam@167
|
819 T2T = T2R - T2S;
|
cannam@167
|
820 T2W = T2U + T2V;
|
cannam@167
|
821 T2X = KP353553390 * (T2T + T2W);
|
cannam@167
|
822 T3G = KP353553390 * (T2T - T2W);
|
cannam@167
|
823 {
|
cannam@167
|
824 E T3B, T3C, T3k, T3l;
|
cannam@167
|
825 T3B = T3a + T39;
|
cannam@167
|
826 T3C = T3d - T3c;
|
cannam@167
|
827 T3D = FNMS(KP461939766, T3C, KP191341716 * T3B);
|
cannam@167
|
828 T3L = FMA(KP461939766, T3B, KP191341716 * T3C);
|
cannam@167
|
829 T3k = T2S + T2R;
|
cannam@167
|
830 T3l = T2U - T2V;
|
cannam@167
|
831 T3m = KP353553390 * (T3k + T3l);
|
cannam@167
|
832 T3v = KP353553390 * (T3l - T3k);
|
cannam@167
|
833 }
|
cannam@167
|
834 T34 = T32 + T33;
|
cannam@167
|
835 T37 = T35 - T36;
|
cannam@167
|
836 T38 = FMA(KP191341716, T34, KP461939766 * T37);
|
cannam@167
|
837 T3q = FNMS(KP191341716, T37, KP461939766 * T34);
|
cannam@167
|
838 {
|
cannam@167
|
839 E T3y, T3z, T3b, T3e;
|
cannam@167
|
840 T3y = T33 - T32;
|
cannam@167
|
841 T3z = T36 + T35;
|
cannam@167
|
842 T3A = FMA(KP461939766, T3y, KP191341716 * T3z);
|
cannam@167
|
843 T3K = FNMS(KP461939766, T3z, KP191341716 * T3y);
|
cannam@167
|
844 T3b = T39 - T3a;
|
cannam@167
|
845 T3e = T3c + T3d;
|
cannam@167
|
846 T3f = FNMS(KP191341716, T3e, KP461939766 * T3b);
|
cannam@167
|
847 T3r = FMA(KP191341716, T3b, KP461939766 * T3e);
|
cannam@167
|
848 }
|
cannam@167
|
849 }
|
cannam@167
|
850 {
|
cannam@167
|
851 E T31, T3g, T3t, T3u;
|
cannam@167
|
852 T31 = T2X + T30;
|
cannam@167
|
853 T3g = T38 + T3f;
|
cannam@167
|
854 Ip[WS(rs, 1)] = T31 + T3g;
|
cannam@167
|
855 Im[WS(rs, 6)] = T3g - T31;
|
cannam@167
|
856 T3t = T3j + T3m;
|
cannam@167
|
857 T3u = T3q + T3r;
|
cannam@167
|
858 Rm[WS(rs, 6)] = T3t - T3u;
|
cannam@167
|
859 Rp[WS(rs, 1)] = T3t + T3u;
|
cannam@167
|
860 }
|
cannam@167
|
861 {
|
cannam@167
|
862 E T3n, T3o, T3p, T3s;
|
cannam@167
|
863 T3n = T3j - T3m;
|
cannam@167
|
864 T3o = T3f - T38;
|
cannam@167
|
865 Rm[WS(rs, 2)] = T3n - T3o;
|
cannam@167
|
866 Rp[WS(rs, 5)] = T3n + T3o;
|
cannam@167
|
867 T3p = T30 - T2X;
|
cannam@167
|
868 T3s = T3q - T3r;
|
cannam@167
|
869 Ip[WS(rs, 5)] = T3p + T3s;
|
cannam@167
|
870 Im[WS(rs, 2)] = T3s - T3p;
|
cannam@167
|
871 }
|
cannam@167
|
872 {
|
cannam@167
|
873 E T3x, T3E, T3N, T3O;
|
cannam@167
|
874 T3x = T3v + T3w;
|
cannam@167
|
875 T3E = T3A + T3D;
|
cannam@167
|
876 Ip[WS(rs, 3)] = T3x + T3E;
|
cannam@167
|
877 Im[WS(rs, 4)] = T3E - T3x;
|
cannam@167
|
878 T3N = T3F + T3G;
|
cannam@167
|
879 T3O = T3K + T3L;
|
cannam@167
|
880 Rm[WS(rs, 4)] = T3N - T3O;
|
cannam@167
|
881 Rp[WS(rs, 3)] = T3N + T3O;
|
cannam@167
|
882 }
|
cannam@167
|
883 {
|
cannam@167
|
884 E T3H, T3I, T3J, T3M;
|
cannam@167
|
885 T3H = T3F - T3G;
|
cannam@167
|
886 T3I = T3D - T3A;
|
cannam@167
|
887 Rm[0] = T3H - T3I;
|
cannam@167
|
888 Rp[WS(rs, 7)] = T3H + T3I;
|
cannam@167
|
889 T3J = T3w - T3v;
|
cannam@167
|
890 T3M = T3K - T3L;
|
cannam@167
|
891 Ip[WS(rs, 7)] = T3J + T3M;
|
cannam@167
|
892 Im[0] = T3M - T3J;
|
cannam@167
|
893 }
|
cannam@167
|
894 }
|
cannam@167
|
895 }
|
cannam@167
|
896 }
|
cannam@167
|
897 }
|
cannam@167
|
898
|
cannam@167
|
899 static const tw_instr twinstr[] = {
|
cannam@167
|
900 {TW_FULL, 1, 16},
|
cannam@167
|
901 {TW_NEXT, 1, 0}
|
cannam@167
|
902 };
|
cannam@167
|
903
|
cannam@167
|
904 static const hc2c_desc desc = { 16, "hc2cfdft_16", twinstr, &GENUS, {168, 62, 38, 0} };
|
cannam@167
|
905
|
cannam@167
|
906 void X(codelet_hc2cfdft_16) (planner *p) {
|
cannam@167
|
907 X(khc2c_register) (p, hc2cfdft_16, &desc, HC2C_VIA_DFT);
|
cannam@167
|
908 }
|
cannam@167
|
909 #endif
|