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 08:07:58 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 16 -dif -name hc2cbdft_16 -include rdft/scalar/hc2cb.h */
|
cannam@167
|
29
|
cannam@167
|
30 /*
|
cannam@167
|
31 * This function contains 206 FP additions, 100 FP multiplications,
|
cannam@167
|
32 * (or, 136 additions, 30 multiplications, 70 fused multiply/add),
|
cannam@167
|
33 * 66 stack variables, 3 constants, and 64 memory accesses
|
cannam@167
|
34 */
|
cannam@167
|
35 #include "rdft/scalar/hc2cb.h"
|
cannam@167
|
36
|
cannam@167
|
37 static void hc2cbdft_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 {
|
cannam@167
|
43 INT m;
|
cannam@167
|
44 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
|
45 E Tf, T20, T32, T3Q, T3f, T3V, TN, T2a, T1m, T2f, T2G, T3G, T2T, T3L, T1F;
|
cannam@167
|
46 E T26, T2J, T2M, T2N, T2U, T2V, T3H, Tu, T25, T3i, T3R, T1a, T2g, T1y, T21;
|
cannam@167
|
47 E T39, T3W, T1p, T2b;
|
cannam@167
|
48 {
|
cannam@167
|
49 E T3, T1e, TA, T1C, T6, Tx, T1h, T1D, Td, T1A, TL, T1k, Ta, T1z, TG;
|
cannam@167
|
50 E T1j;
|
cannam@167
|
51 {
|
cannam@167
|
52 E T1, T2, T1f, T1g;
|
cannam@167
|
53 T1 = Rp[0];
|
cannam@167
|
54 T2 = Rm[WS(rs, 7)];
|
cannam@167
|
55 T3 = T1 + T2;
|
cannam@167
|
56 T1e = T1 - T2;
|
cannam@167
|
57 {
|
cannam@167
|
58 E Ty, Tz, T4, T5;
|
cannam@167
|
59 Ty = Ip[0];
|
cannam@167
|
60 Tz = Im[WS(rs, 7)];
|
cannam@167
|
61 TA = Ty + Tz;
|
cannam@167
|
62 T1C = Ty - Tz;
|
cannam@167
|
63 T4 = Rp[WS(rs, 4)];
|
cannam@167
|
64 T5 = Rm[WS(rs, 3)];
|
cannam@167
|
65 T6 = T4 + T5;
|
cannam@167
|
66 Tx = T4 - T5;
|
cannam@167
|
67 }
|
cannam@167
|
68 T1f = Ip[WS(rs, 4)];
|
cannam@167
|
69 T1g = Im[WS(rs, 3)];
|
cannam@167
|
70 T1h = T1f + T1g;
|
cannam@167
|
71 T1D = T1f - T1g;
|
cannam@167
|
72 {
|
cannam@167
|
73 E Tb, Tc, TH, TI, TJ, TK;
|
cannam@167
|
74 Tb = Rm[WS(rs, 1)];
|
cannam@167
|
75 Tc = Rp[WS(rs, 6)];
|
cannam@167
|
76 TH = Tb - Tc;
|
cannam@167
|
77 TI = Im[WS(rs, 1)];
|
cannam@167
|
78 TJ = Ip[WS(rs, 6)];
|
cannam@167
|
79 TK = TI + TJ;
|
cannam@167
|
80 Td = Tb + Tc;
|
cannam@167
|
81 T1A = TJ - TI;
|
cannam@167
|
82 TL = TH + TK;
|
cannam@167
|
83 T1k = TH - TK;
|
cannam@167
|
84 }
|
cannam@167
|
85 {
|
cannam@167
|
86 E T8, T9, TC, TD, TE, TF;
|
cannam@167
|
87 T8 = Rp[WS(rs, 2)];
|
cannam@167
|
88 T9 = Rm[WS(rs, 5)];
|
cannam@167
|
89 TC = T8 - T9;
|
cannam@167
|
90 TD = Ip[WS(rs, 2)];
|
cannam@167
|
91 TE = Im[WS(rs, 5)];
|
cannam@167
|
92 TF = TD + TE;
|
cannam@167
|
93 Ta = T8 + T9;
|
cannam@167
|
94 T1z = TD - TE;
|
cannam@167
|
95 TG = TC + TF;
|
cannam@167
|
96 T1j = TC - TF;
|
cannam@167
|
97 }
|
cannam@167
|
98 }
|
cannam@167
|
99 {
|
cannam@167
|
100 E T7, Te, T30, T31;
|
cannam@167
|
101 T7 = T3 + T6;
|
cannam@167
|
102 Te = Ta + Td;
|
cannam@167
|
103 Tf = T7 + Te;
|
cannam@167
|
104 T20 = T7 - Te;
|
cannam@167
|
105 T30 = TA - Tx;
|
cannam@167
|
106 T31 = T1j - T1k;
|
cannam@167
|
107 T32 = FMA(KP707106781, T31, T30);
|
cannam@167
|
108 T3Q = FNMS(KP707106781, T31, T30);
|
cannam@167
|
109 }
|
cannam@167
|
110 {
|
cannam@167
|
111 E T3d, T3e, TB, TM;
|
cannam@167
|
112 T3d = T1e + T1h;
|
cannam@167
|
113 T3e = TG + TL;
|
cannam@167
|
114 T3f = FNMS(KP707106781, T3e, T3d);
|
cannam@167
|
115 T3V = FMA(KP707106781, T3e, T3d);
|
cannam@167
|
116 TB = Tx + TA;
|
cannam@167
|
117 TM = TG - TL;
|
cannam@167
|
118 TN = FMA(KP707106781, TM, TB);
|
cannam@167
|
119 T2a = FNMS(KP707106781, TM, TB);
|
cannam@167
|
120 }
|
cannam@167
|
121 {
|
cannam@167
|
122 E T1i, T1l, T2E, T2F;
|
cannam@167
|
123 T1i = T1e - T1h;
|
cannam@167
|
124 T1l = T1j + T1k;
|
cannam@167
|
125 T1m = FMA(KP707106781, T1l, T1i);
|
cannam@167
|
126 T2f = FNMS(KP707106781, T1l, T1i);
|
cannam@167
|
127 T2E = T3 - T6;
|
cannam@167
|
128 T2F = T1A - T1z;
|
cannam@167
|
129 T2G = T2E + T2F;
|
cannam@167
|
130 T3G = T2E - T2F;
|
cannam@167
|
131 }
|
cannam@167
|
132 {
|
cannam@167
|
133 E T2R, T2S, T1B, T1E;
|
cannam@167
|
134 T2R = Ta - Td;
|
cannam@167
|
135 T2S = T1C - T1D;
|
cannam@167
|
136 T2T = T2R + T2S;
|
cannam@167
|
137 T3L = T2S - T2R;
|
cannam@167
|
138 T1B = T1z + T1A;
|
cannam@167
|
139 T1E = T1C + T1D;
|
cannam@167
|
140 T1F = T1B + T1E;
|
cannam@167
|
141 T26 = T1E - T1B;
|
cannam@167
|
142 }
|
cannam@167
|
143 }
|
cannam@167
|
144 {
|
cannam@167
|
145 E Ti, T1s, Tl, T1t, TS, TX, T34, T33, T2I, T2H, Tp, T1v, Ts, T1w, T13;
|
cannam@167
|
146 E T18, T37, T36, T2L, T2K;
|
cannam@167
|
147 {
|
cannam@167
|
148 E TT, TR, TO, TW;
|
cannam@167
|
149 {
|
cannam@167
|
150 E Tg, Th, TP, TQ;
|
cannam@167
|
151 Tg = Rp[WS(rs, 1)];
|
cannam@167
|
152 Th = Rm[WS(rs, 6)];
|
cannam@167
|
153 Ti = Tg + Th;
|
cannam@167
|
154 TT = Tg - Th;
|
cannam@167
|
155 TP = Ip[WS(rs, 1)];
|
cannam@167
|
156 TQ = Im[WS(rs, 6)];
|
cannam@167
|
157 TR = TP + TQ;
|
cannam@167
|
158 T1s = TP - TQ;
|
cannam@167
|
159 }
|
cannam@167
|
160 {
|
cannam@167
|
161 E Tj, Tk, TU, TV;
|
cannam@167
|
162 Tj = Rp[WS(rs, 5)];
|
cannam@167
|
163 Tk = Rm[WS(rs, 2)];
|
cannam@167
|
164 Tl = Tj + Tk;
|
cannam@167
|
165 TO = Tj - Tk;
|
cannam@167
|
166 TU = Ip[WS(rs, 5)];
|
cannam@167
|
167 TV = Im[WS(rs, 2)];
|
cannam@167
|
168 TW = TU + TV;
|
cannam@167
|
169 T1t = TU - TV;
|
cannam@167
|
170 }
|
cannam@167
|
171 TS = TO + TR;
|
cannam@167
|
172 TX = TT - TW;
|
cannam@167
|
173 T34 = TR - TO;
|
cannam@167
|
174 T33 = TT + TW;
|
cannam@167
|
175 T2I = T1s - T1t;
|
cannam@167
|
176 T2H = Ti - Tl;
|
cannam@167
|
177 }
|
cannam@167
|
178 {
|
cannam@167
|
179 E T14, T12, TZ, T17;
|
cannam@167
|
180 {
|
cannam@167
|
181 E Tn, To, T10, T11;
|
cannam@167
|
182 Tn = Rm[0];
|
cannam@167
|
183 To = Rp[WS(rs, 7)];
|
cannam@167
|
184 Tp = Tn + To;
|
cannam@167
|
185 T14 = Tn - To;
|
cannam@167
|
186 T10 = Im[0];
|
cannam@167
|
187 T11 = Ip[WS(rs, 7)];
|
cannam@167
|
188 T12 = T10 + T11;
|
cannam@167
|
189 T1v = T11 - T10;
|
cannam@167
|
190 }
|
cannam@167
|
191 {
|
cannam@167
|
192 E Tq, Tr, T15, T16;
|
cannam@167
|
193 Tq = Rp[WS(rs, 3)];
|
cannam@167
|
194 Tr = Rm[WS(rs, 4)];
|
cannam@167
|
195 Ts = Tq + Tr;
|
cannam@167
|
196 TZ = Tq - Tr;
|
cannam@167
|
197 T15 = Ip[WS(rs, 3)];
|
cannam@167
|
198 T16 = Im[WS(rs, 4)];
|
cannam@167
|
199 T17 = T15 + T16;
|
cannam@167
|
200 T1w = T15 - T16;
|
cannam@167
|
201 }
|
cannam@167
|
202 T13 = TZ - T12;
|
cannam@167
|
203 T18 = T14 - T17;
|
cannam@167
|
204 T37 = TZ + T12;
|
cannam@167
|
205 T36 = T14 + T17;
|
cannam@167
|
206 T2L = T1v - T1w;
|
cannam@167
|
207 T2K = Tp - Ts;
|
cannam@167
|
208 }
|
cannam@167
|
209 T2J = T2H - T2I;
|
cannam@167
|
210 T2M = T2K + T2L;
|
cannam@167
|
211 T2N = T2J + T2M;
|
cannam@167
|
212 T2U = T2H + T2I;
|
cannam@167
|
213 T2V = T2L - T2K;
|
cannam@167
|
214 T3H = T2V - T2U;
|
cannam@167
|
215 {
|
cannam@167
|
216 E Tm, Tt, T3g, T3h;
|
cannam@167
|
217 Tm = Ti + Tl;
|
cannam@167
|
218 Tt = Tp + Ts;
|
cannam@167
|
219 Tu = Tm + Tt;
|
cannam@167
|
220 T25 = Tm - Tt;
|
cannam@167
|
221 T3g = FNMS(KP414213562, T33, T34);
|
cannam@167
|
222 T3h = FNMS(KP414213562, T36, T37);
|
cannam@167
|
223 T3i = T3g + T3h;
|
cannam@167
|
224 T3R = T3h - T3g;
|
cannam@167
|
225 }
|
cannam@167
|
226 {
|
cannam@167
|
227 E TY, T19, T1u, T1x;
|
cannam@167
|
228 TY = FMA(KP414213562, TX, TS);
|
cannam@167
|
229 T19 = FNMS(KP414213562, T18, T13);
|
cannam@167
|
230 T1a = TY + T19;
|
cannam@167
|
231 T2g = T19 - TY;
|
cannam@167
|
232 T1u = T1s + T1t;
|
cannam@167
|
233 T1x = T1v + T1w;
|
cannam@167
|
234 T1y = T1u + T1x;
|
cannam@167
|
235 T21 = T1x - T1u;
|
cannam@167
|
236 }
|
cannam@167
|
237 {
|
cannam@167
|
238 E T35, T38, T1n, T1o;
|
cannam@167
|
239 T35 = FMA(KP414213562, T34, T33);
|
cannam@167
|
240 T38 = FMA(KP414213562, T37, T36);
|
cannam@167
|
241 T39 = T35 - T38;
|
cannam@167
|
242 T3W = T35 + T38;
|
cannam@167
|
243 T1n = FNMS(KP414213562, TS, TX);
|
cannam@167
|
244 T1o = FMA(KP414213562, T13, T18);
|
cannam@167
|
245 T1p = T1n + T1o;
|
cannam@167
|
246 T2b = T1n - T1o;
|
cannam@167
|
247 }
|
cannam@167
|
248 }
|
cannam@167
|
249 {
|
cannam@167
|
250 E Tv, T1G, T1b, T1q, T1c, T1H, Tw, T1r, T1I, T1d;
|
cannam@167
|
251 Tv = Tf + Tu;
|
cannam@167
|
252 T1G = T1y + T1F;
|
cannam@167
|
253 T1b = FMA(KP923879532, T1a, TN);
|
cannam@167
|
254 T1q = FMA(KP923879532, T1p, T1m);
|
cannam@167
|
255 Tw = W[0];
|
cannam@167
|
256 T1c = Tw * T1b;
|
cannam@167
|
257 T1H = Tw * T1q;
|
cannam@167
|
258 T1d = W[1];
|
cannam@167
|
259 T1r = FMA(T1d, T1q, T1c);
|
cannam@167
|
260 T1I = FNMS(T1d, T1b, T1H);
|
cannam@167
|
261 Rp[0] = Tv - T1r;
|
cannam@167
|
262 Ip[0] = T1G + T1I;
|
cannam@167
|
263 Rm[0] = Tv + T1r;
|
cannam@167
|
264 Im[0] = T1I - T1G;
|
cannam@167
|
265 }
|
cannam@167
|
266 {
|
cannam@167
|
267 E T1N, T1J, T1L, T1M, T1V, T1Q, T1T, T1R, T1X, T1K, T1P;
|
cannam@167
|
268 T1N = T1F - T1y;
|
cannam@167
|
269 T1K = Tf - Tu;
|
cannam@167
|
270 T1J = W[14];
|
cannam@167
|
271 T1L = T1J * T1K;
|
cannam@167
|
272 T1M = W[15];
|
cannam@167
|
273 T1V = T1M * T1K;
|
cannam@167
|
274 T1Q = FNMS(KP923879532, T1a, TN);
|
cannam@167
|
275 T1T = FNMS(KP923879532, T1p, T1m);
|
cannam@167
|
276 T1P = W[16];
|
cannam@167
|
277 T1R = T1P * T1Q;
|
cannam@167
|
278 T1X = T1P * T1T;
|
cannam@167
|
279 {
|
cannam@167
|
280 E T1O, T1W, T1U, T1Y, T1S;
|
cannam@167
|
281 T1O = FNMS(T1M, T1N, T1L);
|
cannam@167
|
282 T1W = FMA(T1J, T1N, T1V);
|
cannam@167
|
283 T1S = W[17];
|
cannam@167
|
284 T1U = FMA(T1S, T1T, T1R);
|
cannam@167
|
285 T1Y = FNMS(T1S, T1Q, T1X);
|
cannam@167
|
286 Rp[WS(rs, 4)] = T1O - T1U;
|
cannam@167
|
287 Ip[WS(rs, 4)] = T1W + T1Y;
|
cannam@167
|
288 Rm[WS(rs, 4)] = T1O + T1U;
|
cannam@167
|
289 Im[WS(rs, 4)] = T1Y - T1W;
|
cannam@167
|
290 }
|
cannam@167
|
291 }
|
cannam@167
|
292 {
|
cannam@167
|
293 E T2r, T2n, T2p, T2q, T2z, T2u, T2x, T2v, T2B, T2o, T2t;
|
cannam@167
|
294 T2r = T26 - T25;
|
cannam@167
|
295 T2o = T20 - T21;
|
cannam@167
|
296 T2n = W[22];
|
cannam@167
|
297 T2p = T2n * T2o;
|
cannam@167
|
298 T2q = W[23];
|
cannam@167
|
299 T2z = T2q * T2o;
|
cannam@167
|
300 T2u = FNMS(KP923879532, T2b, T2a);
|
cannam@167
|
301 T2x = FNMS(KP923879532, T2g, T2f);
|
cannam@167
|
302 T2t = W[24];
|
cannam@167
|
303 T2v = T2t * T2u;
|
cannam@167
|
304 T2B = T2t * T2x;
|
cannam@167
|
305 {
|
cannam@167
|
306 E T2s, T2A, T2y, T2C, T2w;
|
cannam@167
|
307 T2s = FNMS(T2q, T2r, T2p);
|
cannam@167
|
308 T2A = FMA(T2n, T2r, T2z);
|
cannam@167
|
309 T2w = W[25];
|
cannam@167
|
310 T2y = FMA(T2w, T2x, T2v);
|
cannam@167
|
311 T2C = FNMS(T2w, T2u, T2B);
|
cannam@167
|
312 Rp[WS(rs, 6)] = T2s - T2y;
|
cannam@167
|
313 Ip[WS(rs, 6)] = T2A + T2C;
|
cannam@167
|
314 Rm[WS(rs, 6)] = T2s + T2y;
|
cannam@167
|
315 Im[WS(rs, 6)] = T2C - T2A;
|
cannam@167
|
316 }
|
cannam@167
|
317 }
|
cannam@167
|
318 {
|
cannam@167
|
319 E T27, T1Z, T23, T24, T2j, T2c, T2h, T2d, T2l, T22, T29;
|
cannam@167
|
320 T27 = T25 + T26;
|
cannam@167
|
321 T22 = T20 + T21;
|
cannam@167
|
322 T1Z = W[6];
|
cannam@167
|
323 T23 = T1Z * T22;
|
cannam@167
|
324 T24 = W[7];
|
cannam@167
|
325 T2j = T24 * T22;
|
cannam@167
|
326 T2c = FMA(KP923879532, T2b, T2a);
|
cannam@167
|
327 T2h = FMA(KP923879532, T2g, T2f);
|
cannam@167
|
328 T29 = W[8];
|
cannam@167
|
329 T2d = T29 * T2c;
|
cannam@167
|
330 T2l = T29 * T2h;
|
cannam@167
|
331 {
|
cannam@167
|
332 E T28, T2k, T2i, T2m, T2e;
|
cannam@167
|
333 T28 = FNMS(T24, T27, T23);
|
cannam@167
|
334 T2k = FMA(T1Z, T27, T2j);
|
cannam@167
|
335 T2e = W[9];
|
cannam@167
|
336 T2i = FMA(T2e, T2h, T2d);
|
cannam@167
|
337 T2m = FNMS(T2e, T2c, T2l);
|
cannam@167
|
338 Rp[WS(rs, 2)] = T28 - T2i;
|
cannam@167
|
339 Ip[WS(rs, 2)] = T2k + T2m;
|
cannam@167
|
340 Rm[WS(rs, 2)] = T28 + T2i;
|
cannam@167
|
341 Im[WS(rs, 2)] = T2m - T2k;
|
cannam@167
|
342 }
|
cannam@167
|
343 }
|
cannam@167
|
344 {
|
cannam@167
|
345 E T3N, T47, T43, T45, T46, T4f, T3F, T3J, T3K, T3Z, T3S, T3X, T3T, T41, T4a;
|
cannam@167
|
346 E T4d, T4b, T4h;
|
cannam@167
|
347 {
|
cannam@167
|
348 E T3M, T44, T3I, T3P, T49;
|
cannam@167
|
349 T3M = T2J - T2M;
|
cannam@167
|
350 T3N = FMA(KP707106781, T3M, T3L);
|
cannam@167
|
351 T47 = FNMS(KP707106781, T3M, T3L);
|
cannam@167
|
352 T44 = FNMS(KP707106781, T3H, T3G);
|
cannam@167
|
353 T43 = W[26];
|
cannam@167
|
354 T45 = T43 * T44;
|
cannam@167
|
355 T46 = W[27];
|
cannam@167
|
356 T4f = T46 * T44;
|
cannam@167
|
357 T3I = FMA(KP707106781, T3H, T3G);
|
cannam@167
|
358 T3F = W[10];
|
cannam@167
|
359 T3J = T3F * T3I;
|
cannam@167
|
360 T3K = W[11];
|
cannam@167
|
361 T3Z = T3K * T3I;
|
cannam@167
|
362 T3S = FMA(KP923879532, T3R, T3Q);
|
cannam@167
|
363 T3X = FNMS(KP923879532, T3W, T3V);
|
cannam@167
|
364 T3P = W[12];
|
cannam@167
|
365 T3T = T3P * T3S;
|
cannam@167
|
366 T41 = T3P * T3X;
|
cannam@167
|
367 T4a = FNMS(KP923879532, T3R, T3Q);
|
cannam@167
|
368 T4d = FMA(KP923879532, T3W, T3V);
|
cannam@167
|
369 T49 = W[28];
|
cannam@167
|
370 T4b = T49 * T4a;
|
cannam@167
|
371 T4h = T49 * T4d;
|
cannam@167
|
372 }
|
cannam@167
|
373 {
|
cannam@167
|
374 E T3O, T40, T3Y, T42, T3U;
|
cannam@167
|
375 T3O = FNMS(T3K, T3N, T3J);
|
cannam@167
|
376 T40 = FMA(T3F, T3N, T3Z);
|
cannam@167
|
377 T3U = W[13];
|
cannam@167
|
378 T3Y = FMA(T3U, T3X, T3T);
|
cannam@167
|
379 T42 = FNMS(T3U, T3S, T41);
|
cannam@167
|
380 Rp[WS(rs, 3)] = T3O - T3Y;
|
cannam@167
|
381 Ip[WS(rs, 3)] = T40 + T42;
|
cannam@167
|
382 Rm[WS(rs, 3)] = T3O + T3Y;
|
cannam@167
|
383 Im[WS(rs, 3)] = T42 - T40;
|
cannam@167
|
384 }
|
cannam@167
|
385 {
|
cannam@167
|
386 E T48, T4g, T4e, T4i, T4c;
|
cannam@167
|
387 T48 = FNMS(T46, T47, T45);
|
cannam@167
|
388 T4g = FMA(T43, T47, T4f);
|
cannam@167
|
389 T4c = W[29];
|
cannam@167
|
390 T4e = FMA(T4c, T4d, T4b);
|
cannam@167
|
391 T4i = FNMS(T4c, T4a, T4h);
|
cannam@167
|
392 Rp[WS(rs, 7)] = T48 - T4e;
|
cannam@167
|
393 Ip[WS(rs, 7)] = T4g + T4i;
|
cannam@167
|
394 Rm[WS(rs, 7)] = T48 + T4e;
|
cannam@167
|
395 Im[WS(rs, 7)] = T4i - T4g;
|
cannam@167
|
396 }
|
cannam@167
|
397 }
|
cannam@167
|
398 {
|
cannam@167
|
399 E T2X, T3t, T3p, T3r, T3s, T3B, T2D, T2P, T2Q, T3l, T3a, T3j, T3b, T3n, T3w;
|
cannam@167
|
400 E T3z, T3x, T3D;
|
cannam@167
|
401 {
|
cannam@167
|
402 E T2W, T3q, T2O, T2Z, T3v;
|
cannam@167
|
403 T2W = T2U + T2V;
|
cannam@167
|
404 T2X = FMA(KP707106781, T2W, T2T);
|
cannam@167
|
405 T3t = FNMS(KP707106781, T2W, T2T);
|
cannam@167
|
406 T3q = FNMS(KP707106781, T2N, T2G);
|
cannam@167
|
407 T3p = W[18];
|
cannam@167
|
408 T3r = T3p * T3q;
|
cannam@167
|
409 T3s = W[19];
|
cannam@167
|
410 T3B = T3s * T3q;
|
cannam@167
|
411 T2O = FMA(KP707106781, T2N, T2G);
|
cannam@167
|
412 T2D = W[2];
|
cannam@167
|
413 T2P = T2D * T2O;
|
cannam@167
|
414 T2Q = W[3];
|
cannam@167
|
415 T3l = T2Q * T2O;
|
cannam@167
|
416 T3a = FMA(KP923879532, T39, T32);
|
cannam@167
|
417 T3j = FNMS(KP923879532, T3i, T3f);
|
cannam@167
|
418 T2Z = W[4];
|
cannam@167
|
419 T3b = T2Z * T3a;
|
cannam@167
|
420 T3n = T2Z * T3j;
|
cannam@167
|
421 T3w = FNMS(KP923879532, T39, T32);
|
cannam@167
|
422 T3z = FMA(KP923879532, T3i, T3f);
|
cannam@167
|
423 T3v = W[20];
|
cannam@167
|
424 T3x = T3v * T3w;
|
cannam@167
|
425 T3D = T3v * T3z;
|
cannam@167
|
426 }
|
cannam@167
|
427 {
|
cannam@167
|
428 E T2Y, T3m, T3k, T3o, T3c;
|
cannam@167
|
429 T2Y = FNMS(T2Q, T2X, T2P);
|
cannam@167
|
430 T3m = FMA(T2D, T2X, T3l);
|
cannam@167
|
431 T3c = W[5];
|
cannam@167
|
432 T3k = FMA(T3c, T3j, T3b);
|
cannam@167
|
433 T3o = FNMS(T3c, T3a, T3n);
|
cannam@167
|
434 Rp[WS(rs, 1)] = T2Y - T3k;
|
cannam@167
|
435 Ip[WS(rs, 1)] = T3m + T3o;
|
cannam@167
|
436 Rm[WS(rs, 1)] = T2Y + T3k;
|
cannam@167
|
437 Im[WS(rs, 1)] = T3o - T3m;
|
cannam@167
|
438 }
|
cannam@167
|
439 {
|
cannam@167
|
440 E T3u, T3C, T3A, T3E, T3y;
|
cannam@167
|
441 T3u = FNMS(T3s, T3t, T3r);
|
cannam@167
|
442 T3C = FMA(T3p, T3t, T3B);
|
cannam@167
|
443 T3y = W[21];
|
cannam@167
|
444 T3A = FMA(T3y, T3z, T3x);
|
cannam@167
|
445 T3E = FNMS(T3y, T3w, T3D);
|
cannam@167
|
446 Rp[WS(rs, 5)] = T3u - T3A;
|
cannam@167
|
447 Ip[WS(rs, 5)] = T3C + T3E;
|
cannam@167
|
448 Rm[WS(rs, 5)] = T3u + T3A;
|
cannam@167
|
449 Im[WS(rs, 5)] = T3E - T3C;
|
cannam@167
|
450 }
|
cannam@167
|
451 }
|
cannam@167
|
452 }
|
cannam@167
|
453 }
|
cannam@167
|
454 }
|
cannam@167
|
455
|
cannam@167
|
456 static const tw_instr twinstr[] = {
|
cannam@167
|
457 {TW_FULL, 1, 16},
|
cannam@167
|
458 {TW_NEXT, 1, 0}
|
cannam@167
|
459 };
|
cannam@167
|
460
|
cannam@167
|
461 static const hc2c_desc desc = { 16, "hc2cbdft_16", twinstr, &GENUS, {136, 30, 70, 0} };
|
cannam@167
|
462
|
cannam@167
|
463 void X(codelet_hc2cbdft_16) (planner *p) {
|
cannam@167
|
464 X(khc2c_register) (p, hc2cbdft_16, &desc, HC2C_VIA_DFT);
|
cannam@167
|
465 }
|
cannam@167
|
466 #else
|
cannam@167
|
467
|
cannam@167
|
468 /* Generated by: ../../../genfft/gen_hc2cdft.native -compact -variables 4 -pipeline-latency 4 -sign 1 -n 16 -dif -name hc2cbdft_16 -include rdft/scalar/hc2cb.h */
|
cannam@167
|
469
|
cannam@167
|
470 /*
|
cannam@167
|
471 * This function contains 206 FP additions, 84 FP multiplications,
|
cannam@167
|
472 * (or, 168 additions, 46 multiplications, 38 fused multiply/add),
|
cannam@167
|
473 * 60 stack variables, 3 constants, and 64 memory accesses
|
cannam@167
|
474 */
|
cannam@167
|
475 #include "rdft/scalar/hc2cb.h"
|
cannam@167
|
476
|
cannam@167
|
477 static void hc2cbdft_16(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms)
|
cannam@167
|
478 {
|
cannam@167
|
479 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
cannam@167
|
480 DK(KP382683432, +0.382683432365089771728459984030398866761344562);
|
cannam@167
|
481 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
cannam@167
|
482 {
|
cannam@167
|
483 INT m;
|
cannam@167
|
484 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
|
485 E TB, T2L, T30, T1n, Tf, T1U, T2H, T3p, T1E, T1Z, TM, T31, T2s, T3k, T1i;
|
cannam@167
|
486 E T2M, Tu, T1Y, T2Q, T2X, T2T, T2Y, TY, T1d, T19, T1e, T2v, T2C, T2y, T2D;
|
cannam@167
|
487 E T1x, T1V;
|
cannam@167
|
488 {
|
cannam@167
|
489 E T3, T1j, TA, T1B, T6, Tx, T1m, T1C, Ta, TC, TF, T1y, Td, TH, TK;
|
cannam@167
|
490 E T1z;
|
cannam@167
|
491 {
|
cannam@167
|
492 E T1, T2, Ty, Tz;
|
cannam@167
|
493 T1 = Rp[0];
|
cannam@167
|
494 T2 = Rm[WS(rs, 7)];
|
cannam@167
|
495 T3 = T1 + T2;
|
cannam@167
|
496 T1j = T1 - T2;
|
cannam@167
|
497 Ty = Ip[0];
|
cannam@167
|
498 Tz = Im[WS(rs, 7)];
|
cannam@167
|
499 TA = Ty + Tz;
|
cannam@167
|
500 T1B = Ty - Tz;
|
cannam@167
|
501 }
|
cannam@167
|
502 {
|
cannam@167
|
503 E T4, T5, T1k, T1l;
|
cannam@167
|
504 T4 = Rp[WS(rs, 4)];
|
cannam@167
|
505 T5 = Rm[WS(rs, 3)];
|
cannam@167
|
506 T6 = T4 + T5;
|
cannam@167
|
507 Tx = T4 - T5;
|
cannam@167
|
508 T1k = Ip[WS(rs, 4)];
|
cannam@167
|
509 T1l = Im[WS(rs, 3)];
|
cannam@167
|
510 T1m = T1k + T1l;
|
cannam@167
|
511 T1C = T1k - T1l;
|
cannam@167
|
512 }
|
cannam@167
|
513 {
|
cannam@167
|
514 E T8, T9, TD, TE;
|
cannam@167
|
515 T8 = Rp[WS(rs, 2)];
|
cannam@167
|
516 T9 = Rm[WS(rs, 5)];
|
cannam@167
|
517 Ta = T8 + T9;
|
cannam@167
|
518 TC = T8 - T9;
|
cannam@167
|
519 TD = Ip[WS(rs, 2)];
|
cannam@167
|
520 TE = Im[WS(rs, 5)];
|
cannam@167
|
521 TF = TD + TE;
|
cannam@167
|
522 T1y = TD - TE;
|
cannam@167
|
523 }
|
cannam@167
|
524 {
|
cannam@167
|
525 E Tb, Tc, TI, TJ;
|
cannam@167
|
526 Tb = Rm[WS(rs, 1)];
|
cannam@167
|
527 Tc = Rp[WS(rs, 6)];
|
cannam@167
|
528 Td = Tb + Tc;
|
cannam@167
|
529 TH = Tb - Tc;
|
cannam@167
|
530 TI = Im[WS(rs, 1)];
|
cannam@167
|
531 TJ = Ip[WS(rs, 6)];
|
cannam@167
|
532 TK = TI + TJ;
|
cannam@167
|
533 T1z = TJ - TI;
|
cannam@167
|
534 }
|
cannam@167
|
535 {
|
cannam@167
|
536 E T7, Te, TG, TL;
|
cannam@167
|
537 TB = Tx + TA;
|
cannam@167
|
538 T2L = TA - Tx;
|
cannam@167
|
539 T30 = T1j + T1m;
|
cannam@167
|
540 T1n = T1j - T1m;
|
cannam@167
|
541 T7 = T3 + T6;
|
cannam@167
|
542 Te = Ta + Td;
|
cannam@167
|
543 Tf = T7 + Te;
|
cannam@167
|
544 T1U = T7 - Te;
|
cannam@167
|
545 {
|
cannam@167
|
546 E T2F, T2G, T1A, T1D;
|
cannam@167
|
547 T2F = Ta - Td;
|
cannam@167
|
548 T2G = T1B - T1C;
|
cannam@167
|
549 T2H = T2F + T2G;
|
cannam@167
|
550 T3p = T2G - T2F;
|
cannam@167
|
551 T1A = T1y + T1z;
|
cannam@167
|
552 T1D = T1B + T1C;
|
cannam@167
|
553 T1E = T1A + T1D;
|
cannam@167
|
554 T1Z = T1D - T1A;
|
cannam@167
|
555 }
|
cannam@167
|
556 TG = TC + TF;
|
cannam@167
|
557 TL = TH + TK;
|
cannam@167
|
558 TM = KP707106781 * (TG - TL);
|
cannam@167
|
559 T31 = KP707106781 * (TG + TL);
|
cannam@167
|
560 {
|
cannam@167
|
561 E T2q, T2r, T1g, T1h;
|
cannam@167
|
562 T2q = T3 - T6;
|
cannam@167
|
563 T2r = T1z - T1y;
|
cannam@167
|
564 T2s = T2q + T2r;
|
cannam@167
|
565 T3k = T2q - T2r;
|
cannam@167
|
566 T1g = TC - TF;
|
cannam@167
|
567 T1h = TH - TK;
|
cannam@167
|
568 T1i = KP707106781 * (T1g + T1h);
|
cannam@167
|
569 T2M = KP707106781 * (T1g - T1h);
|
cannam@167
|
570 }
|
cannam@167
|
571 }
|
cannam@167
|
572 }
|
cannam@167
|
573 {
|
cannam@167
|
574 E Ti, TT, TR, T1r, Tl, TO, TW, T1s, Tp, T14, T12, T1u, Ts, TZ, T17;
|
cannam@167
|
575 E T1v;
|
cannam@167
|
576 {
|
cannam@167
|
577 E Tg, Th, TP, TQ;
|
cannam@167
|
578 Tg = Rp[WS(rs, 1)];
|
cannam@167
|
579 Th = Rm[WS(rs, 6)];
|
cannam@167
|
580 Ti = Tg + Th;
|
cannam@167
|
581 TT = Tg - Th;
|
cannam@167
|
582 TP = Ip[WS(rs, 1)];
|
cannam@167
|
583 TQ = Im[WS(rs, 6)];
|
cannam@167
|
584 TR = TP + TQ;
|
cannam@167
|
585 T1r = TP - TQ;
|
cannam@167
|
586 }
|
cannam@167
|
587 {
|
cannam@167
|
588 E Tj, Tk, TU, TV;
|
cannam@167
|
589 Tj = Rp[WS(rs, 5)];
|
cannam@167
|
590 Tk = Rm[WS(rs, 2)];
|
cannam@167
|
591 Tl = Tj + Tk;
|
cannam@167
|
592 TO = Tj - Tk;
|
cannam@167
|
593 TU = Ip[WS(rs, 5)];
|
cannam@167
|
594 TV = Im[WS(rs, 2)];
|
cannam@167
|
595 TW = TU + TV;
|
cannam@167
|
596 T1s = TU - TV;
|
cannam@167
|
597 }
|
cannam@167
|
598 {
|
cannam@167
|
599 E Tn, To, T10, T11;
|
cannam@167
|
600 Tn = Rm[0];
|
cannam@167
|
601 To = Rp[WS(rs, 7)];
|
cannam@167
|
602 Tp = Tn + To;
|
cannam@167
|
603 T14 = Tn - To;
|
cannam@167
|
604 T10 = Im[0];
|
cannam@167
|
605 T11 = Ip[WS(rs, 7)];
|
cannam@167
|
606 T12 = T10 + T11;
|
cannam@167
|
607 T1u = T11 - T10;
|
cannam@167
|
608 }
|
cannam@167
|
609 {
|
cannam@167
|
610 E Tq, Tr, T15, T16;
|
cannam@167
|
611 Tq = Rp[WS(rs, 3)];
|
cannam@167
|
612 Tr = Rm[WS(rs, 4)];
|
cannam@167
|
613 Ts = Tq + Tr;
|
cannam@167
|
614 TZ = Tq - Tr;
|
cannam@167
|
615 T15 = Ip[WS(rs, 3)];
|
cannam@167
|
616 T16 = Im[WS(rs, 4)];
|
cannam@167
|
617 T17 = T15 + T16;
|
cannam@167
|
618 T1v = T15 - T16;
|
cannam@167
|
619 }
|
cannam@167
|
620 {
|
cannam@167
|
621 E Tm, Tt, T2O, T2P;
|
cannam@167
|
622 Tm = Ti + Tl;
|
cannam@167
|
623 Tt = Tp + Ts;
|
cannam@167
|
624 Tu = Tm + Tt;
|
cannam@167
|
625 T1Y = Tm - Tt;
|
cannam@167
|
626 T2O = TR - TO;
|
cannam@167
|
627 T2P = TT + TW;
|
cannam@167
|
628 T2Q = FMA(KP382683432, T2O, KP923879532 * T2P);
|
cannam@167
|
629 T2X = FNMS(KP923879532, T2O, KP382683432 * T2P);
|
cannam@167
|
630 }
|
cannam@167
|
631 {
|
cannam@167
|
632 E T2R, T2S, TS, TX;
|
cannam@167
|
633 T2R = TZ + T12;
|
cannam@167
|
634 T2S = T14 + T17;
|
cannam@167
|
635 T2T = FMA(KP382683432, T2R, KP923879532 * T2S);
|
cannam@167
|
636 T2Y = FNMS(KP923879532, T2R, KP382683432 * T2S);
|
cannam@167
|
637 TS = TO + TR;
|
cannam@167
|
638 TX = TT - TW;
|
cannam@167
|
639 TY = FMA(KP923879532, TS, KP382683432 * TX);
|
cannam@167
|
640 T1d = FNMS(KP382683432, TS, KP923879532 * TX);
|
cannam@167
|
641 }
|
cannam@167
|
642 {
|
cannam@167
|
643 E T13, T18, T2t, T2u;
|
cannam@167
|
644 T13 = TZ - T12;
|
cannam@167
|
645 T18 = T14 - T17;
|
cannam@167
|
646 T19 = FNMS(KP382683432, T18, KP923879532 * T13);
|
cannam@167
|
647 T1e = FMA(KP382683432, T13, KP923879532 * T18);
|
cannam@167
|
648 T2t = Ti - Tl;
|
cannam@167
|
649 T2u = T1r - T1s;
|
cannam@167
|
650 T2v = T2t - T2u;
|
cannam@167
|
651 T2C = T2t + T2u;
|
cannam@167
|
652 }
|
cannam@167
|
653 {
|
cannam@167
|
654 E T2w, T2x, T1t, T1w;
|
cannam@167
|
655 T2w = Tp - Ts;
|
cannam@167
|
656 T2x = T1u - T1v;
|
cannam@167
|
657 T2y = T2w + T2x;
|
cannam@167
|
658 T2D = T2x - T2w;
|
cannam@167
|
659 T1t = T1r + T1s;
|
cannam@167
|
660 T1w = T1u + T1v;
|
cannam@167
|
661 T1x = T1t + T1w;
|
cannam@167
|
662 T1V = T1w - T1t;
|
cannam@167
|
663 }
|
cannam@167
|
664 }
|
cannam@167
|
665 {
|
cannam@167
|
666 E Tv, T1F, T1b, T1N, T1p, T1P, T1L, T1R;
|
cannam@167
|
667 Tv = Tf + Tu;
|
cannam@167
|
668 T1F = T1x + T1E;
|
cannam@167
|
669 {
|
cannam@167
|
670 E TN, T1a, T1f, T1o;
|
cannam@167
|
671 TN = TB + TM;
|
cannam@167
|
672 T1a = TY + T19;
|
cannam@167
|
673 T1b = TN + T1a;
|
cannam@167
|
674 T1N = TN - T1a;
|
cannam@167
|
675 T1f = T1d + T1e;
|
cannam@167
|
676 T1o = T1i + T1n;
|
cannam@167
|
677 T1p = T1f + T1o;
|
cannam@167
|
678 T1P = T1o - T1f;
|
cannam@167
|
679 {
|
cannam@167
|
680 E T1I, T1K, T1H, T1J;
|
cannam@167
|
681 T1I = Tf - Tu;
|
cannam@167
|
682 T1K = T1E - T1x;
|
cannam@167
|
683 T1H = W[14];
|
cannam@167
|
684 T1J = W[15];
|
cannam@167
|
685 T1L = FNMS(T1J, T1K, T1H * T1I);
|
cannam@167
|
686 T1R = FMA(T1J, T1I, T1H * T1K);
|
cannam@167
|
687 }
|
cannam@167
|
688 }
|
cannam@167
|
689 {
|
cannam@167
|
690 E T1q, T1G, Tw, T1c;
|
cannam@167
|
691 Tw = W[0];
|
cannam@167
|
692 T1c = W[1];
|
cannam@167
|
693 T1q = FMA(Tw, T1b, T1c * T1p);
|
cannam@167
|
694 T1G = FNMS(T1c, T1b, Tw * T1p);
|
cannam@167
|
695 Rp[0] = Tv - T1q;
|
cannam@167
|
696 Ip[0] = T1F + T1G;
|
cannam@167
|
697 Rm[0] = Tv + T1q;
|
cannam@167
|
698 Im[0] = T1G - T1F;
|
cannam@167
|
699 }
|
cannam@167
|
700 {
|
cannam@167
|
701 E T1Q, T1S, T1M, T1O;
|
cannam@167
|
702 T1M = W[16];
|
cannam@167
|
703 T1O = W[17];
|
cannam@167
|
704 T1Q = FMA(T1M, T1N, T1O * T1P);
|
cannam@167
|
705 T1S = FNMS(T1O, T1N, T1M * T1P);
|
cannam@167
|
706 Rp[WS(rs, 4)] = T1L - T1Q;
|
cannam@167
|
707 Ip[WS(rs, 4)] = T1R + T1S;
|
cannam@167
|
708 Rm[WS(rs, 4)] = T1L + T1Q;
|
cannam@167
|
709 Im[WS(rs, 4)] = T1S - T1R;
|
cannam@167
|
710 }
|
cannam@167
|
711 }
|
cannam@167
|
712 {
|
cannam@167
|
713 E T25, T2j, T29, T2l, T21, T2b, T2h, T2n;
|
cannam@167
|
714 {
|
cannam@167
|
715 E T23, T24, T27, T28;
|
cannam@167
|
716 T23 = TB - TM;
|
cannam@167
|
717 T24 = T1d - T1e;
|
cannam@167
|
718 T25 = T23 + T24;
|
cannam@167
|
719 T2j = T23 - T24;
|
cannam@167
|
720 T27 = T19 - TY;
|
cannam@167
|
721 T28 = T1n - T1i;
|
cannam@167
|
722 T29 = T27 + T28;
|
cannam@167
|
723 T2l = T28 - T27;
|
cannam@167
|
724 }
|
cannam@167
|
725 {
|
cannam@167
|
726 E T1W, T20, T1T, T1X;
|
cannam@167
|
727 T1W = T1U + T1V;
|
cannam@167
|
728 T20 = T1Y + T1Z;
|
cannam@167
|
729 T1T = W[6];
|
cannam@167
|
730 T1X = W[7];
|
cannam@167
|
731 T21 = FNMS(T1X, T20, T1T * T1W);
|
cannam@167
|
732 T2b = FMA(T1X, T1W, T1T * T20);
|
cannam@167
|
733 }
|
cannam@167
|
734 {
|
cannam@167
|
735 E T2e, T2g, T2d, T2f;
|
cannam@167
|
736 T2e = T1U - T1V;
|
cannam@167
|
737 T2g = T1Z - T1Y;
|
cannam@167
|
738 T2d = W[22];
|
cannam@167
|
739 T2f = W[23];
|
cannam@167
|
740 T2h = FNMS(T2f, T2g, T2d * T2e);
|
cannam@167
|
741 T2n = FMA(T2f, T2e, T2d * T2g);
|
cannam@167
|
742 }
|
cannam@167
|
743 {
|
cannam@167
|
744 E T2a, T2c, T22, T26;
|
cannam@167
|
745 T22 = W[8];
|
cannam@167
|
746 T26 = W[9];
|
cannam@167
|
747 T2a = FMA(T22, T25, T26 * T29);
|
cannam@167
|
748 T2c = FNMS(T26, T25, T22 * T29);
|
cannam@167
|
749 Rp[WS(rs, 2)] = T21 - T2a;
|
cannam@167
|
750 Ip[WS(rs, 2)] = T2b + T2c;
|
cannam@167
|
751 Rm[WS(rs, 2)] = T21 + T2a;
|
cannam@167
|
752 Im[WS(rs, 2)] = T2c - T2b;
|
cannam@167
|
753 }
|
cannam@167
|
754 {
|
cannam@167
|
755 E T2m, T2o, T2i, T2k;
|
cannam@167
|
756 T2i = W[24];
|
cannam@167
|
757 T2k = W[25];
|
cannam@167
|
758 T2m = FMA(T2i, T2j, T2k * T2l);
|
cannam@167
|
759 T2o = FNMS(T2k, T2j, T2i * T2l);
|
cannam@167
|
760 Rp[WS(rs, 6)] = T2h - T2m;
|
cannam@167
|
761 Ip[WS(rs, 6)] = T2n + T2o;
|
cannam@167
|
762 Rm[WS(rs, 6)] = T2h + T2m;
|
cannam@167
|
763 Im[WS(rs, 6)] = T2o - T2n;
|
cannam@167
|
764 }
|
cannam@167
|
765 }
|
cannam@167
|
766 {
|
cannam@167
|
767 E T2A, T38, T2I, T3a, T2V, T3d, T33, T3f, T2z, T2E;
|
cannam@167
|
768 T2z = KP707106781 * (T2v + T2y);
|
cannam@167
|
769 T2A = T2s + T2z;
|
cannam@167
|
770 T38 = T2s - T2z;
|
cannam@167
|
771 T2E = KP707106781 * (T2C + T2D);
|
cannam@167
|
772 T2I = T2E + T2H;
|
cannam@167
|
773 T3a = T2H - T2E;
|
cannam@167
|
774 {
|
cannam@167
|
775 E T2N, T2U, T2Z, T32;
|
cannam@167
|
776 T2N = T2L + T2M;
|
cannam@167
|
777 T2U = T2Q - T2T;
|
cannam@167
|
778 T2V = T2N + T2U;
|
cannam@167
|
779 T3d = T2N - T2U;
|
cannam@167
|
780 T2Z = T2X + T2Y;
|
cannam@167
|
781 T32 = T30 - T31;
|
cannam@167
|
782 T33 = T2Z + T32;
|
cannam@167
|
783 T3f = T32 - T2Z;
|
cannam@167
|
784 }
|
cannam@167
|
785 {
|
cannam@167
|
786 E T2J, T35, T34, T36;
|
cannam@167
|
787 {
|
cannam@167
|
788 E T2p, T2B, T2K, T2W;
|
cannam@167
|
789 T2p = W[2];
|
cannam@167
|
790 T2B = W[3];
|
cannam@167
|
791 T2J = FNMS(T2B, T2I, T2p * T2A);
|
cannam@167
|
792 T35 = FMA(T2B, T2A, T2p * T2I);
|
cannam@167
|
793 T2K = W[4];
|
cannam@167
|
794 T2W = W[5];
|
cannam@167
|
795 T34 = FMA(T2K, T2V, T2W * T33);
|
cannam@167
|
796 T36 = FNMS(T2W, T2V, T2K * T33);
|
cannam@167
|
797 }
|
cannam@167
|
798 Rp[WS(rs, 1)] = T2J - T34;
|
cannam@167
|
799 Ip[WS(rs, 1)] = T35 + T36;
|
cannam@167
|
800 Rm[WS(rs, 1)] = T2J + T34;
|
cannam@167
|
801 Im[WS(rs, 1)] = T36 - T35;
|
cannam@167
|
802 }
|
cannam@167
|
803 {
|
cannam@167
|
804 E T3b, T3h, T3g, T3i;
|
cannam@167
|
805 {
|
cannam@167
|
806 E T37, T39, T3c, T3e;
|
cannam@167
|
807 T37 = W[18];
|
cannam@167
|
808 T39 = W[19];
|
cannam@167
|
809 T3b = FNMS(T39, T3a, T37 * T38);
|
cannam@167
|
810 T3h = FMA(T39, T38, T37 * T3a);
|
cannam@167
|
811 T3c = W[20];
|
cannam@167
|
812 T3e = W[21];
|
cannam@167
|
813 T3g = FMA(T3c, T3d, T3e * T3f);
|
cannam@167
|
814 T3i = FNMS(T3e, T3d, T3c * T3f);
|
cannam@167
|
815 }
|
cannam@167
|
816 Rp[WS(rs, 5)] = T3b - T3g;
|
cannam@167
|
817 Ip[WS(rs, 5)] = T3h + T3i;
|
cannam@167
|
818 Rm[WS(rs, 5)] = T3b + T3g;
|
cannam@167
|
819 Im[WS(rs, 5)] = T3i - T3h;
|
cannam@167
|
820 }
|
cannam@167
|
821 }
|
cannam@167
|
822 {
|
cannam@167
|
823 E T3m, T3E, T3q, T3G, T3v, T3J, T3z, T3L, T3l, T3o;
|
cannam@167
|
824 T3l = KP707106781 * (T2D - T2C);
|
cannam@167
|
825 T3m = T3k + T3l;
|
cannam@167
|
826 T3E = T3k - T3l;
|
cannam@167
|
827 T3o = KP707106781 * (T2v - T2y);
|
cannam@167
|
828 T3q = T3o + T3p;
|
cannam@167
|
829 T3G = T3p - T3o;
|
cannam@167
|
830 {
|
cannam@167
|
831 E T3t, T3u, T3x, T3y;
|
cannam@167
|
832 T3t = T2L - T2M;
|
cannam@167
|
833 T3u = T2X - T2Y;
|
cannam@167
|
834 T3v = T3t + T3u;
|
cannam@167
|
835 T3J = T3t - T3u;
|
cannam@167
|
836 T3x = T31 + T30;
|
cannam@167
|
837 T3y = T2Q + T2T;
|
cannam@167
|
838 T3z = T3x - T3y;
|
cannam@167
|
839 T3L = T3y + T3x;
|
cannam@167
|
840 }
|
cannam@167
|
841 {
|
cannam@167
|
842 E T3r, T3B, T3A, T3C;
|
cannam@167
|
843 {
|
cannam@167
|
844 E T3j, T3n, T3s, T3w;
|
cannam@167
|
845 T3j = W[10];
|
cannam@167
|
846 T3n = W[11];
|
cannam@167
|
847 T3r = FNMS(T3n, T3q, T3j * T3m);
|
cannam@167
|
848 T3B = FMA(T3n, T3m, T3j * T3q);
|
cannam@167
|
849 T3s = W[12];
|
cannam@167
|
850 T3w = W[13];
|
cannam@167
|
851 T3A = FMA(T3s, T3v, T3w * T3z);
|
cannam@167
|
852 T3C = FNMS(T3w, T3v, T3s * T3z);
|
cannam@167
|
853 }
|
cannam@167
|
854 Rp[WS(rs, 3)] = T3r - T3A;
|
cannam@167
|
855 Ip[WS(rs, 3)] = T3B + T3C;
|
cannam@167
|
856 Rm[WS(rs, 3)] = T3r + T3A;
|
cannam@167
|
857 Im[WS(rs, 3)] = T3C - T3B;
|
cannam@167
|
858 }
|
cannam@167
|
859 {
|
cannam@167
|
860 E T3H, T3N, T3M, T3O;
|
cannam@167
|
861 {
|
cannam@167
|
862 E T3D, T3F, T3I, T3K;
|
cannam@167
|
863 T3D = W[26];
|
cannam@167
|
864 T3F = W[27];
|
cannam@167
|
865 T3H = FNMS(T3F, T3G, T3D * T3E);
|
cannam@167
|
866 T3N = FMA(T3F, T3E, T3D * T3G);
|
cannam@167
|
867 T3I = W[28];
|
cannam@167
|
868 T3K = W[29];
|
cannam@167
|
869 T3M = FMA(T3I, T3J, T3K * T3L);
|
cannam@167
|
870 T3O = FNMS(T3K, T3J, T3I * T3L);
|
cannam@167
|
871 }
|
cannam@167
|
872 Rp[WS(rs, 7)] = T3H - T3M;
|
cannam@167
|
873 Ip[WS(rs, 7)] = T3N + T3O;
|
cannam@167
|
874 Rm[WS(rs, 7)] = T3H + T3M;
|
cannam@167
|
875 Im[WS(rs, 7)] = T3O - T3N;
|
cannam@167
|
876 }
|
cannam@167
|
877 }
|
cannam@167
|
878 }
|
cannam@167
|
879 }
|
cannam@167
|
880 }
|
cannam@167
|
881
|
cannam@167
|
882 static const tw_instr twinstr[] = {
|
cannam@167
|
883 {TW_FULL, 1, 16},
|
cannam@167
|
884 {TW_NEXT, 1, 0}
|
cannam@167
|
885 };
|
cannam@167
|
886
|
cannam@167
|
887 static const hc2c_desc desc = { 16, "hc2cbdft_16", twinstr, &GENUS, {168, 46, 38, 0} };
|
cannam@167
|
888
|
cannam@167
|
889 void X(codelet_hc2cbdft_16) (planner *p) {
|
cannam@167
|
890 X(khc2c_register) (p, hc2cbdft_16, &desc, HC2C_VIA_DFT);
|
cannam@167
|
891 }
|
cannam@167
|
892 #endif
|