Chris@10
|
1 /*
|
Chris@10
|
2 * Copyright (c) 2003, 2007-11 Matteo Frigo
|
Chris@10
|
3 * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology
|
Chris@10
|
4 *
|
Chris@10
|
5 * This program is free software; you can redistribute it and/or modify
|
Chris@10
|
6 * it under the terms of the GNU General Public License as published by
|
Chris@10
|
7 * the Free Software Foundation; either version 2 of the License, or
|
Chris@10
|
8 * (at your option) any later version.
|
Chris@10
|
9 *
|
Chris@10
|
10 * This program is distributed in the hope that it will be useful,
|
Chris@10
|
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
Chris@10
|
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
Chris@10
|
13 * GNU General Public License for more details.
|
Chris@10
|
14 *
|
Chris@10
|
15 * You should have received a copy of the GNU General Public License
|
Chris@10
|
16 * along with this program; if not, write to the Free Software
|
Chris@10
|
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
Chris@10
|
18 *
|
Chris@10
|
19 */
|
Chris@10
|
20
|
Chris@10
|
21 /* This file was automatically generated --- DO NOT EDIT */
|
Chris@10
|
22 /* Generated on Sun Nov 25 07:42:07 EST 2012 */
|
Chris@10
|
23
|
Chris@10
|
24 #include "codelet-rdft.h"
|
Chris@10
|
25
|
Chris@10
|
26 #ifdef HAVE_FMA
|
Chris@10
|
27
|
Chris@10
|
28 /* Generated by: ../../../genfft/gen_hc2cdft.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -sign 1 -n 16 -dif -name hc2cbdft2_16 -include hc2cb.h */
|
Chris@10
|
29
|
Chris@10
|
30 /*
|
Chris@10
|
31 * This function contains 206 FP additions, 100 FP multiplications,
|
Chris@10
|
32 * (or, 136 additions, 30 multiplications, 70 fused multiply/add),
|
Chris@10
|
33 * 97 stack variables, 3 constants, and 64 memory accesses
|
Chris@10
|
34 */
|
Chris@10
|
35 #include "hc2cb.h"
|
Chris@10
|
36
|
Chris@10
|
37 static void hc2cbdft2_16(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms)
|
Chris@10
|
38 {
|
Chris@10
|
39 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
Chris@10
|
40 DK(KP414213562, +0.414213562373095048801688724209698078569671875);
|
Chris@10
|
41 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
Chris@10
|
42 {
|
Chris@10
|
43 INT m;
|
Chris@10
|
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)) {
|
Chris@10
|
45 E T3w, T3z, T2Y, T3D, T3x, T3m, T3u, T3C, T3y, T3o, T3k, T3E, T3A;
|
Chris@10
|
46 {
|
Chris@10
|
47 E T20, Tf, T3Q, T32, T3V, T3f, T2a, TN, T2f, T1m, T3G, T2G, T3L, T2T, T26;
|
Chris@10
|
48 E T1F, T3M, T2N, T3H, T2W, T25, Tu, T1n, T1o, T3R, T3i, T2g, T1a, T21, T1y;
|
Chris@10
|
49 E T3W, T39;
|
Chris@10
|
50 {
|
Chris@10
|
51 E T2R, T1B, T2S, T1E;
|
Chris@10
|
52 {
|
Chris@10
|
53 E T1e, T3, T1C, TA, Tx, T6, T1D, T1h, Td, T1A, TL, T1k, Ta, TC, TF;
|
Chris@10
|
54 E T1z;
|
Chris@10
|
55 {
|
Chris@10
|
56 E T4, T5, T1f, T1g;
|
Chris@10
|
57 {
|
Chris@10
|
58 E T1, T2, Ty, Tz;
|
Chris@10
|
59 T1 = Rp[0];
|
Chris@10
|
60 T2 = Rm[WS(rs, 7)];
|
Chris@10
|
61 Ty = Ip[0];
|
Chris@10
|
62 Tz = Im[WS(rs, 7)];
|
Chris@10
|
63 T4 = Rp[WS(rs, 4)];
|
Chris@10
|
64 T1e = T1 - T2;
|
Chris@10
|
65 T3 = T1 + T2;
|
Chris@10
|
66 T1C = Ty - Tz;
|
Chris@10
|
67 TA = Ty + Tz;
|
Chris@10
|
68 T5 = Rm[WS(rs, 3)];
|
Chris@10
|
69 }
|
Chris@10
|
70 T1f = Ip[WS(rs, 4)];
|
Chris@10
|
71 T1g = Im[WS(rs, 3)];
|
Chris@10
|
72 {
|
Chris@10
|
73 E Tb, Tc, TI, TJ;
|
Chris@10
|
74 Tb = Rm[WS(rs, 1)];
|
Chris@10
|
75 Tx = T4 - T5;
|
Chris@10
|
76 T6 = T4 + T5;
|
Chris@10
|
77 T1D = T1f - T1g;
|
Chris@10
|
78 T1h = T1f + T1g;
|
Chris@10
|
79 Tc = Rp[WS(rs, 6)];
|
Chris@10
|
80 TI = Im[WS(rs, 1)];
|
Chris@10
|
81 TJ = Ip[WS(rs, 6)];
|
Chris@10
|
82 {
|
Chris@10
|
83 E T8, TH, TK, T9, TD, TE;
|
Chris@10
|
84 T8 = Rp[WS(rs, 2)];
|
Chris@10
|
85 Td = Tb + Tc;
|
Chris@10
|
86 TH = Tb - Tc;
|
Chris@10
|
87 T1A = TJ - TI;
|
Chris@10
|
88 TK = TI + TJ;
|
Chris@10
|
89 T9 = Rm[WS(rs, 5)];
|
Chris@10
|
90 TD = Ip[WS(rs, 2)];
|
Chris@10
|
91 TE = Im[WS(rs, 5)];
|
Chris@10
|
92 TL = TH + TK;
|
Chris@10
|
93 T1k = TH - TK;
|
Chris@10
|
94 Ta = T8 + T9;
|
Chris@10
|
95 TC = T8 - T9;
|
Chris@10
|
96 TF = TD + TE;
|
Chris@10
|
97 T1z = TD - TE;
|
Chris@10
|
98 }
|
Chris@10
|
99 }
|
Chris@10
|
100 }
|
Chris@10
|
101 {
|
Chris@10
|
102 E T2E, TB, T1l, T1i, T3d, T3e, TM, T2F;
|
Chris@10
|
103 {
|
Chris@10
|
104 E T7, TG, Te, T30, T31, T1j;
|
Chris@10
|
105 T2E = T3 - T6;
|
Chris@10
|
106 T7 = T3 + T6;
|
Chris@10
|
107 T1j = TC - TF;
|
Chris@10
|
108 TG = TC + TF;
|
Chris@10
|
109 Te = Ta + Td;
|
Chris@10
|
110 T2R = Ta - Td;
|
Chris@10
|
111 TB = Tx + TA;
|
Chris@10
|
112 T30 = TA - Tx;
|
Chris@10
|
113 T31 = T1j - T1k;
|
Chris@10
|
114 T1l = T1j + T1k;
|
Chris@10
|
115 T1i = T1e - T1h;
|
Chris@10
|
116 T3d = T1e + T1h;
|
Chris@10
|
117 T20 = T7 - Te;
|
Chris@10
|
118 Tf = T7 + Te;
|
Chris@10
|
119 T3Q = FNMS(KP707106781, T31, T30);
|
Chris@10
|
120 T32 = FMA(KP707106781, T31, T30);
|
Chris@10
|
121 T3e = TG + TL;
|
Chris@10
|
122 TM = TG - TL;
|
Chris@10
|
123 }
|
Chris@10
|
124 T3V = FMA(KP707106781, T3e, T3d);
|
Chris@10
|
125 T3f = FNMS(KP707106781, T3e, T3d);
|
Chris@10
|
126 T2a = FNMS(KP707106781, TM, TB);
|
Chris@10
|
127 TN = FMA(KP707106781, TM, TB);
|
Chris@10
|
128 T2F = T1A - T1z;
|
Chris@10
|
129 T1B = T1z + T1A;
|
Chris@10
|
130 T2f = FNMS(KP707106781, T1l, T1i);
|
Chris@10
|
131 T1m = FMA(KP707106781, T1l, T1i);
|
Chris@10
|
132 T3G = T2E - T2F;
|
Chris@10
|
133 T2G = T2E + T2F;
|
Chris@10
|
134 T2S = T1C - T1D;
|
Chris@10
|
135 T1E = T1C + T1D;
|
Chris@10
|
136 }
|
Chris@10
|
137 }
|
Chris@10
|
138 {
|
Chris@10
|
139 E T34, TS, T2H, Tm, T1u, T2I, T33, TX, Tq, T14, Tp, T1v, T12, Tr, T15;
|
Chris@10
|
140 E T16;
|
Chris@10
|
141 {
|
Chris@10
|
142 E Tj, TT, Ti, T1s, TR, Tk, TU, TV;
|
Chris@10
|
143 {
|
Chris@10
|
144 E Tg, Th, TP, TQ;
|
Chris@10
|
145 Tg = Rp[WS(rs, 1)];
|
Chris@10
|
146 T3L = T2S - T2R;
|
Chris@10
|
147 T2T = T2R + T2S;
|
Chris@10
|
148 T26 = T1E - T1B;
|
Chris@10
|
149 T1F = T1B + T1E;
|
Chris@10
|
150 Th = Rm[WS(rs, 6)];
|
Chris@10
|
151 TP = Ip[WS(rs, 1)];
|
Chris@10
|
152 TQ = Im[WS(rs, 6)];
|
Chris@10
|
153 Tj = Rp[WS(rs, 5)];
|
Chris@10
|
154 TT = Tg - Th;
|
Chris@10
|
155 Ti = Tg + Th;
|
Chris@10
|
156 T1s = TP - TQ;
|
Chris@10
|
157 TR = TP + TQ;
|
Chris@10
|
158 Tk = Rm[WS(rs, 2)];
|
Chris@10
|
159 TU = Ip[WS(rs, 5)];
|
Chris@10
|
160 TV = Im[WS(rs, 2)];
|
Chris@10
|
161 }
|
Chris@10
|
162 {
|
Chris@10
|
163 E Tn, To, T10, T11;
|
Chris@10
|
164 Tn = Rm[0];
|
Chris@10
|
165 {
|
Chris@10
|
166 E TO, Tl, T1t, TW;
|
Chris@10
|
167 TO = Tj - Tk;
|
Chris@10
|
168 Tl = Tj + Tk;
|
Chris@10
|
169 T1t = TU - TV;
|
Chris@10
|
170 TW = TU + TV;
|
Chris@10
|
171 T34 = TR - TO;
|
Chris@10
|
172 TS = TO + TR;
|
Chris@10
|
173 T2H = Ti - Tl;
|
Chris@10
|
174 Tm = Ti + Tl;
|
Chris@10
|
175 T1u = T1s + T1t;
|
Chris@10
|
176 T2I = T1s - T1t;
|
Chris@10
|
177 T33 = TT + TW;
|
Chris@10
|
178 TX = TT - TW;
|
Chris@10
|
179 To = Rp[WS(rs, 7)];
|
Chris@10
|
180 }
|
Chris@10
|
181 T10 = Im[0];
|
Chris@10
|
182 T11 = Ip[WS(rs, 7)];
|
Chris@10
|
183 Tq = Rp[WS(rs, 3)];
|
Chris@10
|
184 T14 = Tn - To;
|
Chris@10
|
185 Tp = Tn + To;
|
Chris@10
|
186 T1v = T11 - T10;
|
Chris@10
|
187 T12 = T10 + T11;
|
Chris@10
|
188 Tr = Rm[WS(rs, 4)];
|
Chris@10
|
189 T15 = Ip[WS(rs, 3)];
|
Chris@10
|
190 T16 = Im[WS(rs, 4)];
|
Chris@10
|
191 }
|
Chris@10
|
192 }
|
Chris@10
|
193 {
|
Chris@10
|
194 E T13, T1x, T18, T35, T3g, T3h, T38, TY, T19;
|
Chris@10
|
195 {
|
Chris@10
|
196 E T2U, T2J, T37, Tt, T36, T2V, T2M, T2K, T2L;
|
Chris@10
|
197 T2U = T2H + T2I;
|
Chris@10
|
198 T2J = T2H - T2I;
|
Chris@10
|
199 {
|
Chris@10
|
200 E TZ, Ts, T1w, T17;
|
Chris@10
|
201 TZ = Tq - Tr;
|
Chris@10
|
202 Ts = Tq + Tr;
|
Chris@10
|
203 T1w = T15 - T16;
|
Chris@10
|
204 T17 = T15 + T16;
|
Chris@10
|
205 T37 = TZ + T12;
|
Chris@10
|
206 T13 = TZ - T12;
|
Chris@10
|
207 T2K = Tp - Ts;
|
Chris@10
|
208 Tt = Tp + Ts;
|
Chris@10
|
209 T1x = T1v + T1w;
|
Chris@10
|
210 T2L = T1v - T1w;
|
Chris@10
|
211 T36 = T14 + T17;
|
Chris@10
|
212 T18 = T14 - T17;
|
Chris@10
|
213 }
|
Chris@10
|
214 T2V = T2L - T2K;
|
Chris@10
|
215 T2M = T2K + T2L;
|
Chris@10
|
216 T3M = T2J - T2M;
|
Chris@10
|
217 T2N = T2J + T2M;
|
Chris@10
|
218 T3H = T2V - T2U;
|
Chris@10
|
219 T2W = T2U + T2V;
|
Chris@10
|
220 T35 = FMA(KP414213562, T34, T33);
|
Chris@10
|
221 T3g = FNMS(KP414213562, T33, T34);
|
Chris@10
|
222 T25 = Tm - Tt;
|
Chris@10
|
223 Tu = Tm + Tt;
|
Chris@10
|
224 T3h = FNMS(KP414213562, T36, T37);
|
Chris@10
|
225 T38 = FMA(KP414213562, T37, T36);
|
Chris@10
|
226 }
|
Chris@10
|
227 T1n = FNMS(KP414213562, TS, TX);
|
Chris@10
|
228 TY = FMA(KP414213562, TX, TS);
|
Chris@10
|
229 T19 = FNMS(KP414213562, T18, T13);
|
Chris@10
|
230 T1o = FMA(KP414213562, T13, T18);
|
Chris@10
|
231 T3R = T3h - T3g;
|
Chris@10
|
232 T3i = T3g + T3h;
|
Chris@10
|
233 T2g = T19 - TY;
|
Chris@10
|
234 T1a = TY + T19;
|
Chris@10
|
235 T21 = T1x - T1u;
|
Chris@10
|
236 T1y = T1u + T1x;
|
Chris@10
|
237 T3W = T35 + T38;
|
Chris@10
|
238 T39 = T35 - T38;
|
Chris@10
|
239 }
|
Chris@10
|
240 }
|
Chris@10
|
241 }
|
Chris@10
|
242 {
|
Chris@10
|
243 E T27, T22, T2c, T2u, T2x, T2h, T2s, T2A, T2w, T2B, T2v;
|
Chris@10
|
244 {
|
Chris@10
|
245 E T1K, Tv, T1G, T1N, T1Q, T1b, T2b, T1p, Tw, T1d;
|
Chris@10
|
246 T1K = Tf - Tu;
|
Chris@10
|
247 Tv = Tf + Tu;
|
Chris@10
|
248 T1G = T1y + T1F;
|
Chris@10
|
249 T1N = T1F - T1y;
|
Chris@10
|
250 T1Q = FNMS(KP923879532, T1a, TN);
|
Chris@10
|
251 T1b = FMA(KP923879532, T1a, TN);
|
Chris@10
|
252 T2b = T1n - T1o;
|
Chris@10
|
253 T1p = T1n + T1o;
|
Chris@10
|
254 Tw = W[0];
|
Chris@10
|
255 T1d = W[1];
|
Chris@10
|
256 {
|
Chris@10
|
257 E T1T, T1O, T1W, T1S, T1X, T1R;
|
Chris@10
|
258 {
|
Chris@10
|
259 E T1J, T1M, T1L, T1V, T1P, T1q;
|
Chris@10
|
260 T1T = FNMS(KP923879532, T1p, T1m);
|
Chris@10
|
261 T1q = FMA(KP923879532, T1p, T1m);
|
Chris@10
|
262 {
|
Chris@10
|
263 E T1c, T1I, T1H, T1r;
|
Chris@10
|
264 T1c = Tw * T1b;
|
Chris@10
|
265 T1J = W[14];
|
Chris@10
|
266 T1H = Tw * T1q;
|
Chris@10
|
267 T1r = FMA(T1d, T1q, T1c);
|
Chris@10
|
268 T1M = W[15];
|
Chris@10
|
269 T1L = T1J * T1K;
|
Chris@10
|
270 T1I = FNMS(T1d, T1b, T1H);
|
Chris@10
|
271 Rm[0] = Tv + T1r;
|
Chris@10
|
272 Rp[0] = Tv - T1r;
|
Chris@10
|
273 T1V = T1M * T1K;
|
Chris@10
|
274 Im[0] = T1I - T1G;
|
Chris@10
|
275 Ip[0] = T1G + T1I;
|
Chris@10
|
276 T1P = W[16];
|
Chris@10
|
277 }
|
Chris@10
|
278 T1O = FNMS(T1M, T1N, T1L);
|
Chris@10
|
279 T1W = FMA(T1J, T1N, T1V);
|
Chris@10
|
280 T1S = W[17];
|
Chris@10
|
281 T1X = T1P * T1T;
|
Chris@10
|
282 T1R = T1P * T1Q;
|
Chris@10
|
283 }
|
Chris@10
|
284 {
|
Chris@10
|
285 E T2r, T2n, T2q, T2p, T2z, T2t, T2o, T1Y, T1U;
|
Chris@10
|
286 T27 = T25 + T26;
|
Chris@10
|
287 T2r = T26 - T25;
|
Chris@10
|
288 T2o = T20 - T21;
|
Chris@10
|
289 T22 = T20 + T21;
|
Chris@10
|
290 T1Y = FNMS(T1S, T1Q, T1X);
|
Chris@10
|
291 T1U = FMA(T1S, T1T, T1R);
|
Chris@10
|
292 T2n = W[22];
|
Chris@10
|
293 T2q = W[23];
|
Chris@10
|
294 Im[WS(rs, 4)] = T1Y - T1W;
|
Chris@10
|
295 Ip[WS(rs, 4)] = T1W + T1Y;
|
Chris@10
|
296 Rm[WS(rs, 4)] = T1O + T1U;
|
Chris@10
|
297 Rp[WS(rs, 4)] = T1O - T1U;
|
Chris@10
|
298 T2p = T2n * T2o;
|
Chris@10
|
299 T2z = T2q * T2o;
|
Chris@10
|
300 T2c = FMA(KP923879532, T2b, T2a);
|
Chris@10
|
301 T2u = FNMS(KP923879532, T2b, T2a);
|
Chris@10
|
302 T2x = FNMS(KP923879532, T2g, T2f);
|
Chris@10
|
303 T2h = FMA(KP923879532, T2g, T2f);
|
Chris@10
|
304 T2t = W[24];
|
Chris@10
|
305 T2s = FNMS(T2q, T2r, T2p);
|
Chris@10
|
306 T2A = FMA(T2n, T2r, T2z);
|
Chris@10
|
307 T2w = W[25];
|
Chris@10
|
308 T2B = T2t * T2x;
|
Chris@10
|
309 T2v = T2t * T2u;
|
Chris@10
|
310 }
|
Chris@10
|
311 }
|
Chris@10
|
312 }
|
Chris@10
|
313 {
|
Chris@10
|
314 E T28, T2k, T2e, T2l, T2d;
|
Chris@10
|
315 {
|
Chris@10
|
316 E T1Z, T24, T23, T2j, T29, T2C, T2y;
|
Chris@10
|
317 T2C = FNMS(T2w, T2u, T2B);
|
Chris@10
|
318 T2y = FMA(T2w, T2x, T2v);
|
Chris@10
|
319 T1Z = W[6];
|
Chris@10
|
320 T24 = W[7];
|
Chris@10
|
321 Im[WS(rs, 6)] = T2C - T2A;
|
Chris@10
|
322 Ip[WS(rs, 6)] = T2A + T2C;
|
Chris@10
|
323 Rm[WS(rs, 6)] = T2s + T2y;
|
Chris@10
|
324 Rp[WS(rs, 6)] = T2s - T2y;
|
Chris@10
|
325 T23 = T1Z * T22;
|
Chris@10
|
326 T2j = T24 * T22;
|
Chris@10
|
327 T29 = W[8];
|
Chris@10
|
328 T28 = FNMS(T24, T27, T23);
|
Chris@10
|
329 T2k = FMA(T1Z, T27, T2j);
|
Chris@10
|
330 T2e = W[9];
|
Chris@10
|
331 T2l = T29 * T2h;
|
Chris@10
|
332 T2d = T29 * T2c;
|
Chris@10
|
333 }
|
Chris@10
|
334 {
|
Chris@10
|
335 E T4a, T4d, T3O, T4h, T4b, T40, T48, T4g, T4c, T42, T3Y;
|
Chris@10
|
336 {
|
Chris@10
|
337 E T3N, T47, T43, T46, T3F, T45, T4f, T3K, T3J, T3S, T3X, T3Z, T49, T41, T3T;
|
Chris@10
|
338 E T3U;
|
Chris@10
|
339 {
|
Chris@10
|
340 E T44, T3I, T2m, T2i, T3P;
|
Chris@10
|
341 T44 = FNMS(KP707106781, T3H, T3G);
|
Chris@10
|
342 T3I = FMA(KP707106781, T3H, T3G);
|
Chris@10
|
343 T2m = FNMS(T2e, T2c, T2l);
|
Chris@10
|
344 T2i = FMA(T2e, T2h, T2d);
|
Chris@10
|
345 T3N = FMA(KP707106781, T3M, T3L);
|
Chris@10
|
346 T47 = FNMS(KP707106781, T3M, T3L);
|
Chris@10
|
347 Im[WS(rs, 2)] = T2m - T2k;
|
Chris@10
|
348 Ip[WS(rs, 2)] = T2k + T2m;
|
Chris@10
|
349 Rm[WS(rs, 2)] = T28 + T2i;
|
Chris@10
|
350 Rp[WS(rs, 2)] = T28 - T2i;
|
Chris@10
|
351 T43 = W[26];
|
Chris@10
|
352 T46 = W[27];
|
Chris@10
|
353 T3F = W[10];
|
Chris@10
|
354 T45 = T43 * T44;
|
Chris@10
|
355 T4f = T46 * T44;
|
Chris@10
|
356 T3K = W[11];
|
Chris@10
|
357 T3J = T3F * T3I;
|
Chris@10
|
358 T4a = FNMS(KP923879532, T3R, T3Q);
|
Chris@10
|
359 T3S = FMA(KP923879532, T3R, T3Q);
|
Chris@10
|
360 T3X = FNMS(KP923879532, T3W, T3V);
|
Chris@10
|
361 T4d = FMA(KP923879532, T3W, T3V);
|
Chris@10
|
362 T3Z = T3K * T3I;
|
Chris@10
|
363 T3P = W[12];
|
Chris@10
|
364 T49 = W[28];
|
Chris@10
|
365 T41 = T3P * T3X;
|
Chris@10
|
366 T3T = T3P * T3S;
|
Chris@10
|
367 }
|
Chris@10
|
368 T3O = FNMS(T3K, T3N, T3J);
|
Chris@10
|
369 T4h = T49 * T4d;
|
Chris@10
|
370 T4b = T49 * T4a;
|
Chris@10
|
371 T40 = FMA(T3F, T3N, T3Z);
|
Chris@10
|
372 T3U = W[13];
|
Chris@10
|
373 T48 = FNMS(T46, T47, T45);
|
Chris@10
|
374 T4g = FMA(T43, T47, T4f);
|
Chris@10
|
375 T4c = W[29];
|
Chris@10
|
376 T42 = FNMS(T3U, T3S, T41);
|
Chris@10
|
377 T3Y = FMA(T3U, T3X, T3T);
|
Chris@10
|
378 }
|
Chris@10
|
379 {
|
Chris@10
|
380 E T3t, T2X, T3p, T3s, T2D, T3r, T3B, T2Q, T2P, T3a, T3j, T3l, T3v, T3n, T3b;
|
Chris@10
|
381 E T3c;
|
Chris@10
|
382 {
|
Chris@10
|
383 E T2O, T3q, T4i, T4e, T2Z;
|
Chris@10
|
384 T4i = FNMS(T4c, T4a, T4h);
|
Chris@10
|
385 T4e = FMA(T4c, T4d, T4b);
|
Chris@10
|
386 Im[WS(rs, 3)] = T42 - T40;
|
Chris@10
|
387 Ip[WS(rs, 3)] = T40 + T42;
|
Chris@10
|
388 Rm[WS(rs, 3)] = T3O + T3Y;
|
Chris@10
|
389 Rp[WS(rs, 3)] = T3O - T3Y;
|
Chris@10
|
390 Im[WS(rs, 7)] = T4i - T4g;
|
Chris@10
|
391 Ip[WS(rs, 7)] = T4g + T4i;
|
Chris@10
|
392 Rm[WS(rs, 7)] = T48 + T4e;
|
Chris@10
|
393 Rp[WS(rs, 7)] = T48 - T4e;
|
Chris@10
|
394 T3t = FNMS(KP707106781, T2W, T2T);
|
Chris@10
|
395 T2X = FMA(KP707106781, T2W, T2T);
|
Chris@10
|
396 T2O = FMA(KP707106781, T2N, T2G);
|
Chris@10
|
397 T3q = FNMS(KP707106781, T2N, T2G);
|
Chris@10
|
398 T3p = W[18];
|
Chris@10
|
399 T3s = W[19];
|
Chris@10
|
400 T2D = W[2];
|
Chris@10
|
401 T3r = T3p * T3q;
|
Chris@10
|
402 T3B = T3s * T3q;
|
Chris@10
|
403 T2Q = W[3];
|
Chris@10
|
404 T2P = T2D * T2O;
|
Chris@10
|
405 T3a = FMA(KP923879532, T39, T32);
|
Chris@10
|
406 T3w = FNMS(KP923879532, T39, T32);
|
Chris@10
|
407 T3z = FMA(KP923879532, T3i, T3f);
|
Chris@10
|
408 T3j = FNMS(KP923879532, T3i, T3f);
|
Chris@10
|
409 T3l = T2Q * T2O;
|
Chris@10
|
410 T2Z = W[4];
|
Chris@10
|
411 T3v = W[20];
|
Chris@10
|
412 T3n = T2Z * T3j;
|
Chris@10
|
413 T3b = T2Z * T3a;
|
Chris@10
|
414 }
|
Chris@10
|
415 T2Y = FNMS(T2Q, T2X, T2P);
|
Chris@10
|
416 T3D = T3v * T3z;
|
Chris@10
|
417 T3x = T3v * T3w;
|
Chris@10
|
418 T3m = FMA(T2D, T2X, T3l);
|
Chris@10
|
419 T3c = W[5];
|
Chris@10
|
420 T3u = FNMS(T3s, T3t, T3r);
|
Chris@10
|
421 T3C = FMA(T3p, T3t, T3B);
|
Chris@10
|
422 T3y = W[21];
|
Chris@10
|
423 T3o = FNMS(T3c, T3a, T3n);
|
Chris@10
|
424 T3k = FMA(T3c, T3j, T3b);
|
Chris@10
|
425 }
|
Chris@10
|
426 }
|
Chris@10
|
427 }
|
Chris@10
|
428 }
|
Chris@10
|
429 }
|
Chris@10
|
430 T3E = FNMS(T3y, T3w, T3D);
|
Chris@10
|
431 T3A = FMA(T3y, T3z, T3x);
|
Chris@10
|
432 Im[WS(rs, 1)] = T3o - T3m;
|
Chris@10
|
433 Ip[WS(rs, 1)] = T3m + T3o;
|
Chris@10
|
434 Rm[WS(rs, 1)] = T2Y + T3k;
|
Chris@10
|
435 Rp[WS(rs, 1)] = T2Y - T3k;
|
Chris@10
|
436 Im[WS(rs, 5)] = T3E - T3C;
|
Chris@10
|
437 Ip[WS(rs, 5)] = T3C + T3E;
|
Chris@10
|
438 Rm[WS(rs, 5)] = T3u + T3A;
|
Chris@10
|
439 Rp[WS(rs, 5)] = T3u - T3A;
|
Chris@10
|
440 }
|
Chris@10
|
441 }
|
Chris@10
|
442 }
|
Chris@10
|
443
|
Chris@10
|
444 static const tw_instr twinstr[] = {
|
Chris@10
|
445 {TW_FULL, 1, 16},
|
Chris@10
|
446 {TW_NEXT, 1, 0}
|
Chris@10
|
447 };
|
Chris@10
|
448
|
Chris@10
|
449 static const hc2c_desc desc = { 16, "hc2cbdft2_16", twinstr, &GENUS, {136, 30, 70, 0} };
|
Chris@10
|
450
|
Chris@10
|
451 void X(codelet_hc2cbdft2_16) (planner *p) {
|
Chris@10
|
452 X(khc2c_register) (p, hc2cbdft2_16, &desc, HC2C_VIA_DFT);
|
Chris@10
|
453 }
|
Chris@10
|
454 #else /* HAVE_FMA */
|
Chris@10
|
455
|
Chris@10
|
456 /* Generated by: ../../../genfft/gen_hc2cdft.native -compact -variables 4 -pipeline-latency 4 -sign 1 -n 16 -dif -name hc2cbdft2_16 -include hc2cb.h */
|
Chris@10
|
457
|
Chris@10
|
458 /*
|
Chris@10
|
459 * This function contains 206 FP additions, 84 FP multiplications,
|
Chris@10
|
460 * (or, 168 additions, 46 multiplications, 38 fused multiply/add),
|
Chris@10
|
461 * 60 stack variables, 3 constants, and 64 memory accesses
|
Chris@10
|
462 */
|
Chris@10
|
463 #include "hc2cb.h"
|
Chris@10
|
464
|
Chris@10
|
465 static void hc2cbdft2_16(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms)
|
Chris@10
|
466 {
|
Chris@10
|
467 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
Chris@10
|
468 DK(KP382683432, +0.382683432365089771728459984030398866761344562);
|
Chris@10
|
469 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
Chris@10
|
470 {
|
Chris@10
|
471 INT m;
|
Chris@10
|
472 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)) {
|
Chris@10
|
473 E TB, T2L, T30, T1n, Tf, T1U, T2H, T3p, T1E, T1Z, TM, T31, T2s, T3k, T1i;
|
Chris@10
|
474 E T2M, Tu, T1Y, T2Q, T2X, T2T, T2Y, TY, T1d, T19, T1e, T2v, T2C, T2y, T2D;
|
Chris@10
|
475 E T1x, T1V;
|
Chris@10
|
476 {
|
Chris@10
|
477 E T3, T1j, TA, T1B, T6, Tx, T1m, T1C, Ta, TC, TF, T1y, Td, TH, TK;
|
Chris@10
|
478 E T1z;
|
Chris@10
|
479 {
|
Chris@10
|
480 E T1, T2, Ty, Tz;
|
Chris@10
|
481 T1 = Rp[0];
|
Chris@10
|
482 T2 = Rm[WS(rs, 7)];
|
Chris@10
|
483 T3 = T1 + T2;
|
Chris@10
|
484 T1j = T1 - T2;
|
Chris@10
|
485 Ty = Ip[0];
|
Chris@10
|
486 Tz = Im[WS(rs, 7)];
|
Chris@10
|
487 TA = Ty + Tz;
|
Chris@10
|
488 T1B = Ty - Tz;
|
Chris@10
|
489 }
|
Chris@10
|
490 {
|
Chris@10
|
491 E T4, T5, T1k, T1l;
|
Chris@10
|
492 T4 = Rp[WS(rs, 4)];
|
Chris@10
|
493 T5 = Rm[WS(rs, 3)];
|
Chris@10
|
494 T6 = T4 + T5;
|
Chris@10
|
495 Tx = T4 - T5;
|
Chris@10
|
496 T1k = Ip[WS(rs, 4)];
|
Chris@10
|
497 T1l = Im[WS(rs, 3)];
|
Chris@10
|
498 T1m = T1k + T1l;
|
Chris@10
|
499 T1C = T1k - T1l;
|
Chris@10
|
500 }
|
Chris@10
|
501 {
|
Chris@10
|
502 E T8, T9, TD, TE;
|
Chris@10
|
503 T8 = Rp[WS(rs, 2)];
|
Chris@10
|
504 T9 = Rm[WS(rs, 5)];
|
Chris@10
|
505 Ta = T8 + T9;
|
Chris@10
|
506 TC = T8 - T9;
|
Chris@10
|
507 TD = Ip[WS(rs, 2)];
|
Chris@10
|
508 TE = Im[WS(rs, 5)];
|
Chris@10
|
509 TF = TD + TE;
|
Chris@10
|
510 T1y = TD - TE;
|
Chris@10
|
511 }
|
Chris@10
|
512 {
|
Chris@10
|
513 E Tb, Tc, TI, TJ;
|
Chris@10
|
514 Tb = Rm[WS(rs, 1)];
|
Chris@10
|
515 Tc = Rp[WS(rs, 6)];
|
Chris@10
|
516 Td = Tb + Tc;
|
Chris@10
|
517 TH = Tb - Tc;
|
Chris@10
|
518 TI = Im[WS(rs, 1)];
|
Chris@10
|
519 TJ = Ip[WS(rs, 6)];
|
Chris@10
|
520 TK = TI + TJ;
|
Chris@10
|
521 T1z = TJ - TI;
|
Chris@10
|
522 }
|
Chris@10
|
523 {
|
Chris@10
|
524 E T7, Te, TG, TL;
|
Chris@10
|
525 TB = Tx + TA;
|
Chris@10
|
526 T2L = TA - Tx;
|
Chris@10
|
527 T30 = T1j + T1m;
|
Chris@10
|
528 T1n = T1j - T1m;
|
Chris@10
|
529 T7 = T3 + T6;
|
Chris@10
|
530 Te = Ta + Td;
|
Chris@10
|
531 Tf = T7 + Te;
|
Chris@10
|
532 T1U = T7 - Te;
|
Chris@10
|
533 {
|
Chris@10
|
534 E T2F, T2G, T1A, T1D;
|
Chris@10
|
535 T2F = Ta - Td;
|
Chris@10
|
536 T2G = T1B - T1C;
|
Chris@10
|
537 T2H = T2F + T2G;
|
Chris@10
|
538 T3p = T2G - T2F;
|
Chris@10
|
539 T1A = T1y + T1z;
|
Chris@10
|
540 T1D = T1B + T1C;
|
Chris@10
|
541 T1E = T1A + T1D;
|
Chris@10
|
542 T1Z = T1D - T1A;
|
Chris@10
|
543 }
|
Chris@10
|
544 TG = TC + TF;
|
Chris@10
|
545 TL = TH + TK;
|
Chris@10
|
546 TM = KP707106781 * (TG - TL);
|
Chris@10
|
547 T31 = KP707106781 * (TG + TL);
|
Chris@10
|
548 {
|
Chris@10
|
549 E T2q, T2r, T1g, T1h;
|
Chris@10
|
550 T2q = T3 - T6;
|
Chris@10
|
551 T2r = T1z - T1y;
|
Chris@10
|
552 T2s = T2q + T2r;
|
Chris@10
|
553 T3k = T2q - T2r;
|
Chris@10
|
554 T1g = TC - TF;
|
Chris@10
|
555 T1h = TH - TK;
|
Chris@10
|
556 T1i = KP707106781 * (T1g + T1h);
|
Chris@10
|
557 T2M = KP707106781 * (T1g - T1h);
|
Chris@10
|
558 }
|
Chris@10
|
559 }
|
Chris@10
|
560 }
|
Chris@10
|
561 {
|
Chris@10
|
562 E Ti, TT, TR, T1r, Tl, TO, TW, T1s, Tp, T14, T12, T1u, Ts, TZ, T17;
|
Chris@10
|
563 E T1v;
|
Chris@10
|
564 {
|
Chris@10
|
565 E Tg, Th, TP, TQ;
|
Chris@10
|
566 Tg = Rp[WS(rs, 1)];
|
Chris@10
|
567 Th = Rm[WS(rs, 6)];
|
Chris@10
|
568 Ti = Tg + Th;
|
Chris@10
|
569 TT = Tg - Th;
|
Chris@10
|
570 TP = Ip[WS(rs, 1)];
|
Chris@10
|
571 TQ = Im[WS(rs, 6)];
|
Chris@10
|
572 TR = TP + TQ;
|
Chris@10
|
573 T1r = TP - TQ;
|
Chris@10
|
574 }
|
Chris@10
|
575 {
|
Chris@10
|
576 E Tj, Tk, TU, TV;
|
Chris@10
|
577 Tj = Rp[WS(rs, 5)];
|
Chris@10
|
578 Tk = Rm[WS(rs, 2)];
|
Chris@10
|
579 Tl = Tj + Tk;
|
Chris@10
|
580 TO = Tj - Tk;
|
Chris@10
|
581 TU = Ip[WS(rs, 5)];
|
Chris@10
|
582 TV = Im[WS(rs, 2)];
|
Chris@10
|
583 TW = TU + TV;
|
Chris@10
|
584 T1s = TU - TV;
|
Chris@10
|
585 }
|
Chris@10
|
586 {
|
Chris@10
|
587 E Tn, To, T10, T11;
|
Chris@10
|
588 Tn = Rm[0];
|
Chris@10
|
589 To = Rp[WS(rs, 7)];
|
Chris@10
|
590 Tp = Tn + To;
|
Chris@10
|
591 T14 = Tn - To;
|
Chris@10
|
592 T10 = Im[0];
|
Chris@10
|
593 T11 = Ip[WS(rs, 7)];
|
Chris@10
|
594 T12 = T10 + T11;
|
Chris@10
|
595 T1u = T11 - T10;
|
Chris@10
|
596 }
|
Chris@10
|
597 {
|
Chris@10
|
598 E Tq, Tr, T15, T16;
|
Chris@10
|
599 Tq = Rp[WS(rs, 3)];
|
Chris@10
|
600 Tr = Rm[WS(rs, 4)];
|
Chris@10
|
601 Ts = Tq + Tr;
|
Chris@10
|
602 TZ = Tq - Tr;
|
Chris@10
|
603 T15 = Ip[WS(rs, 3)];
|
Chris@10
|
604 T16 = Im[WS(rs, 4)];
|
Chris@10
|
605 T17 = T15 + T16;
|
Chris@10
|
606 T1v = T15 - T16;
|
Chris@10
|
607 }
|
Chris@10
|
608 {
|
Chris@10
|
609 E Tm, Tt, T2O, T2P;
|
Chris@10
|
610 Tm = Ti + Tl;
|
Chris@10
|
611 Tt = Tp + Ts;
|
Chris@10
|
612 Tu = Tm + Tt;
|
Chris@10
|
613 T1Y = Tm - Tt;
|
Chris@10
|
614 T2O = TR - TO;
|
Chris@10
|
615 T2P = TT + TW;
|
Chris@10
|
616 T2Q = FMA(KP382683432, T2O, KP923879532 * T2P);
|
Chris@10
|
617 T2X = FNMS(KP923879532, T2O, KP382683432 * T2P);
|
Chris@10
|
618 }
|
Chris@10
|
619 {
|
Chris@10
|
620 E T2R, T2S, TS, TX;
|
Chris@10
|
621 T2R = TZ + T12;
|
Chris@10
|
622 T2S = T14 + T17;
|
Chris@10
|
623 T2T = FMA(KP382683432, T2R, KP923879532 * T2S);
|
Chris@10
|
624 T2Y = FNMS(KP923879532, T2R, KP382683432 * T2S);
|
Chris@10
|
625 TS = TO + TR;
|
Chris@10
|
626 TX = TT - TW;
|
Chris@10
|
627 TY = FMA(KP923879532, TS, KP382683432 * TX);
|
Chris@10
|
628 T1d = FNMS(KP382683432, TS, KP923879532 * TX);
|
Chris@10
|
629 }
|
Chris@10
|
630 {
|
Chris@10
|
631 E T13, T18, T2t, T2u;
|
Chris@10
|
632 T13 = TZ - T12;
|
Chris@10
|
633 T18 = T14 - T17;
|
Chris@10
|
634 T19 = FNMS(KP382683432, T18, KP923879532 * T13);
|
Chris@10
|
635 T1e = FMA(KP382683432, T13, KP923879532 * T18);
|
Chris@10
|
636 T2t = Ti - Tl;
|
Chris@10
|
637 T2u = T1r - T1s;
|
Chris@10
|
638 T2v = T2t - T2u;
|
Chris@10
|
639 T2C = T2t + T2u;
|
Chris@10
|
640 }
|
Chris@10
|
641 {
|
Chris@10
|
642 E T2w, T2x, T1t, T1w;
|
Chris@10
|
643 T2w = Tp - Ts;
|
Chris@10
|
644 T2x = T1u - T1v;
|
Chris@10
|
645 T2y = T2w + T2x;
|
Chris@10
|
646 T2D = T2x - T2w;
|
Chris@10
|
647 T1t = T1r + T1s;
|
Chris@10
|
648 T1w = T1u + T1v;
|
Chris@10
|
649 T1x = T1t + T1w;
|
Chris@10
|
650 T1V = T1w - T1t;
|
Chris@10
|
651 }
|
Chris@10
|
652 }
|
Chris@10
|
653 {
|
Chris@10
|
654 E Tv, T1F, T1b, T1N, T1p, T1P, T1L, T1R;
|
Chris@10
|
655 Tv = Tf + Tu;
|
Chris@10
|
656 T1F = T1x + T1E;
|
Chris@10
|
657 {
|
Chris@10
|
658 E TN, T1a, T1f, T1o;
|
Chris@10
|
659 TN = TB + TM;
|
Chris@10
|
660 T1a = TY + T19;
|
Chris@10
|
661 T1b = TN + T1a;
|
Chris@10
|
662 T1N = TN - T1a;
|
Chris@10
|
663 T1f = T1d + T1e;
|
Chris@10
|
664 T1o = T1i + T1n;
|
Chris@10
|
665 T1p = T1f + T1o;
|
Chris@10
|
666 T1P = T1o - T1f;
|
Chris@10
|
667 {
|
Chris@10
|
668 E T1I, T1K, T1H, T1J;
|
Chris@10
|
669 T1I = Tf - Tu;
|
Chris@10
|
670 T1K = T1E - T1x;
|
Chris@10
|
671 T1H = W[14];
|
Chris@10
|
672 T1J = W[15];
|
Chris@10
|
673 T1L = FNMS(T1J, T1K, T1H * T1I);
|
Chris@10
|
674 T1R = FMA(T1J, T1I, T1H * T1K);
|
Chris@10
|
675 }
|
Chris@10
|
676 }
|
Chris@10
|
677 {
|
Chris@10
|
678 E T1q, T1G, Tw, T1c;
|
Chris@10
|
679 Tw = W[0];
|
Chris@10
|
680 T1c = W[1];
|
Chris@10
|
681 T1q = FMA(Tw, T1b, T1c * T1p);
|
Chris@10
|
682 T1G = FNMS(T1c, T1b, Tw * T1p);
|
Chris@10
|
683 Rp[0] = Tv - T1q;
|
Chris@10
|
684 Ip[0] = T1F + T1G;
|
Chris@10
|
685 Rm[0] = Tv + T1q;
|
Chris@10
|
686 Im[0] = T1G - T1F;
|
Chris@10
|
687 }
|
Chris@10
|
688 {
|
Chris@10
|
689 E T1Q, T1S, T1M, T1O;
|
Chris@10
|
690 T1M = W[16];
|
Chris@10
|
691 T1O = W[17];
|
Chris@10
|
692 T1Q = FMA(T1M, T1N, T1O * T1P);
|
Chris@10
|
693 T1S = FNMS(T1O, T1N, T1M * T1P);
|
Chris@10
|
694 Rp[WS(rs, 4)] = T1L - T1Q;
|
Chris@10
|
695 Ip[WS(rs, 4)] = T1R + T1S;
|
Chris@10
|
696 Rm[WS(rs, 4)] = T1L + T1Q;
|
Chris@10
|
697 Im[WS(rs, 4)] = T1S - T1R;
|
Chris@10
|
698 }
|
Chris@10
|
699 }
|
Chris@10
|
700 {
|
Chris@10
|
701 E T25, T2j, T29, T2l, T21, T2b, T2h, T2n;
|
Chris@10
|
702 {
|
Chris@10
|
703 E T23, T24, T27, T28;
|
Chris@10
|
704 T23 = TB - TM;
|
Chris@10
|
705 T24 = T1d - T1e;
|
Chris@10
|
706 T25 = T23 + T24;
|
Chris@10
|
707 T2j = T23 - T24;
|
Chris@10
|
708 T27 = T19 - TY;
|
Chris@10
|
709 T28 = T1n - T1i;
|
Chris@10
|
710 T29 = T27 + T28;
|
Chris@10
|
711 T2l = T28 - T27;
|
Chris@10
|
712 }
|
Chris@10
|
713 {
|
Chris@10
|
714 E T1W, T20, T1T, T1X;
|
Chris@10
|
715 T1W = T1U + T1V;
|
Chris@10
|
716 T20 = T1Y + T1Z;
|
Chris@10
|
717 T1T = W[6];
|
Chris@10
|
718 T1X = W[7];
|
Chris@10
|
719 T21 = FNMS(T1X, T20, T1T * T1W);
|
Chris@10
|
720 T2b = FMA(T1X, T1W, T1T * T20);
|
Chris@10
|
721 }
|
Chris@10
|
722 {
|
Chris@10
|
723 E T2e, T2g, T2d, T2f;
|
Chris@10
|
724 T2e = T1U - T1V;
|
Chris@10
|
725 T2g = T1Z - T1Y;
|
Chris@10
|
726 T2d = W[22];
|
Chris@10
|
727 T2f = W[23];
|
Chris@10
|
728 T2h = FNMS(T2f, T2g, T2d * T2e);
|
Chris@10
|
729 T2n = FMA(T2f, T2e, T2d * T2g);
|
Chris@10
|
730 }
|
Chris@10
|
731 {
|
Chris@10
|
732 E T2a, T2c, T22, T26;
|
Chris@10
|
733 T22 = W[8];
|
Chris@10
|
734 T26 = W[9];
|
Chris@10
|
735 T2a = FMA(T22, T25, T26 * T29);
|
Chris@10
|
736 T2c = FNMS(T26, T25, T22 * T29);
|
Chris@10
|
737 Rp[WS(rs, 2)] = T21 - T2a;
|
Chris@10
|
738 Ip[WS(rs, 2)] = T2b + T2c;
|
Chris@10
|
739 Rm[WS(rs, 2)] = T21 + T2a;
|
Chris@10
|
740 Im[WS(rs, 2)] = T2c - T2b;
|
Chris@10
|
741 }
|
Chris@10
|
742 {
|
Chris@10
|
743 E T2m, T2o, T2i, T2k;
|
Chris@10
|
744 T2i = W[24];
|
Chris@10
|
745 T2k = W[25];
|
Chris@10
|
746 T2m = FMA(T2i, T2j, T2k * T2l);
|
Chris@10
|
747 T2o = FNMS(T2k, T2j, T2i * T2l);
|
Chris@10
|
748 Rp[WS(rs, 6)] = T2h - T2m;
|
Chris@10
|
749 Ip[WS(rs, 6)] = T2n + T2o;
|
Chris@10
|
750 Rm[WS(rs, 6)] = T2h + T2m;
|
Chris@10
|
751 Im[WS(rs, 6)] = T2o - T2n;
|
Chris@10
|
752 }
|
Chris@10
|
753 }
|
Chris@10
|
754 {
|
Chris@10
|
755 E T2A, T38, T2I, T3a, T2V, T3d, T33, T3f, T2z, T2E;
|
Chris@10
|
756 T2z = KP707106781 * (T2v + T2y);
|
Chris@10
|
757 T2A = T2s + T2z;
|
Chris@10
|
758 T38 = T2s - T2z;
|
Chris@10
|
759 T2E = KP707106781 * (T2C + T2D);
|
Chris@10
|
760 T2I = T2E + T2H;
|
Chris@10
|
761 T3a = T2H - T2E;
|
Chris@10
|
762 {
|
Chris@10
|
763 E T2N, T2U, T2Z, T32;
|
Chris@10
|
764 T2N = T2L + T2M;
|
Chris@10
|
765 T2U = T2Q - T2T;
|
Chris@10
|
766 T2V = T2N + T2U;
|
Chris@10
|
767 T3d = T2N - T2U;
|
Chris@10
|
768 T2Z = T2X + T2Y;
|
Chris@10
|
769 T32 = T30 - T31;
|
Chris@10
|
770 T33 = T2Z + T32;
|
Chris@10
|
771 T3f = T32 - T2Z;
|
Chris@10
|
772 }
|
Chris@10
|
773 {
|
Chris@10
|
774 E T2J, T35, T34, T36;
|
Chris@10
|
775 {
|
Chris@10
|
776 E T2p, T2B, T2K, T2W;
|
Chris@10
|
777 T2p = W[2];
|
Chris@10
|
778 T2B = W[3];
|
Chris@10
|
779 T2J = FNMS(T2B, T2I, T2p * T2A);
|
Chris@10
|
780 T35 = FMA(T2B, T2A, T2p * T2I);
|
Chris@10
|
781 T2K = W[4];
|
Chris@10
|
782 T2W = W[5];
|
Chris@10
|
783 T34 = FMA(T2K, T2V, T2W * T33);
|
Chris@10
|
784 T36 = FNMS(T2W, T2V, T2K * T33);
|
Chris@10
|
785 }
|
Chris@10
|
786 Rp[WS(rs, 1)] = T2J - T34;
|
Chris@10
|
787 Ip[WS(rs, 1)] = T35 + T36;
|
Chris@10
|
788 Rm[WS(rs, 1)] = T2J + T34;
|
Chris@10
|
789 Im[WS(rs, 1)] = T36 - T35;
|
Chris@10
|
790 }
|
Chris@10
|
791 {
|
Chris@10
|
792 E T3b, T3h, T3g, T3i;
|
Chris@10
|
793 {
|
Chris@10
|
794 E T37, T39, T3c, T3e;
|
Chris@10
|
795 T37 = W[18];
|
Chris@10
|
796 T39 = W[19];
|
Chris@10
|
797 T3b = FNMS(T39, T3a, T37 * T38);
|
Chris@10
|
798 T3h = FMA(T39, T38, T37 * T3a);
|
Chris@10
|
799 T3c = W[20];
|
Chris@10
|
800 T3e = W[21];
|
Chris@10
|
801 T3g = FMA(T3c, T3d, T3e * T3f);
|
Chris@10
|
802 T3i = FNMS(T3e, T3d, T3c * T3f);
|
Chris@10
|
803 }
|
Chris@10
|
804 Rp[WS(rs, 5)] = T3b - T3g;
|
Chris@10
|
805 Ip[WS(rs, 5)] = T3h + T3i;
|
Chris@10
|
806 Rm[WS(rs, 5)] = T3b + T3g;
|
Chris@10
|
807 Im[WS(rs, 5)] = T3i - T3h;
|
Chris@10
|
808 }
|
Chris@10
|
809 }
|
Chris@10
|
810 {
|
Chris@10
|
811 E T3m, T3E, T3q, T3G, T3v, T3J, T3z, T3L, T3l, T3o;
|
Chris@10
|
812 T3l = KP707106781 * (T2D - T2C);
|
Chris@10
|
813 T3m = T3k + T3l;
|
Chris@10
|
814 T3E = T3k - T3l;
|
Chris@10
|
815 T3o = KP707106781 * (T2v - T2y);
|
Chris@10
|
816 T3q = T3o + T3p;
|
Chris@10
|
817 T3G = T3p - T3o;
|
Chris@10
|
818 {
|
Chris@10
|
819 E T3t, T3u, T3x, T3y;
|
Chris@10
|
820 T3t = T2L - T2M;
|
Chris@10
|
821 T3u = T2X - T2Y;
|
Chris@10
|
822 T3v = T3t + T3u;
|
Chris@10
|
823 T3J = T3t - T3u;
|
Chris@10
|
824 T3x = T31 + T30;
|
Chris@10
|
825 T3y = T2Q + T2T;
|
Chris@10
|
826 T3z = T3x - T3y;
|
Chris@10
|
827 T3L = T3y + T3x;
|
Chris@10
|
828 }
|
Chris@10
|
829 {
|
Chris@10
|
830 E T3r, T3B, T3A, T3C;
|
Chris@10
|
831 {
|
Chris@10
|
832 E T3j, T3n, T3s, T3w;
|
Chris@10
|
833 T3j = W[10];
|
Chris@10
|
834 T3n = W[11];
|
Chris@10
|
835 T3r = FNMS(T3n, T3q, T3j * T3m);
|
Chris@10
|
836 T3B = FMA(T3n, T3m, T3j * T3q);
|
Chris@10
|
837 T3s = W[12];
|
Chris@10
|
838 T3w = W[13];
|
Chris@10
|
839 T3A = FMA(T3s, T3v, T3w * T3z);
|
Chris@10
|
840 T3C = FNMS(T3w, T3v, T3s * T3z);
|
Chris@10
|
841 }
|
Chris@10
|
842 Rp[WS(rs, 3)] = T3r - T3A;
|
Chris@10
|
843 Ip[WS(rs, 3)] = T3B + T3C;
|
Chris@10
|
844 Rm[WS(rs, 3)] = T3r + T3A;
|
Chris@10
|
845 Im[WS(rs, 3)] = T3C - T3B;
|
Chris@10
|
846 }
|
Chris@10
|
847 {
|
Chris@10
|
848 E T3H, T3N, T3M, T3O;
|
Chris@10
|
849 {
|
Chris@10
|
850 E T3D, T3F, T3I, T3K;
|
Chris@10
|
851 T3D = W[26];
|
Chris@10
|
852 T3F = W[27];
|
Chris@10
|
853 T3H = FNMS(T3F, T3G, T3D * T3E);
|
Chris@10
|
854 T3N = FMA(T3F, T3E, T3D * T3G);
|
Chris@10
|
855 T3I = W[28];
|
Chris@10
|
856 T3K = W[29];
|
Chris@10
|
857 T3M = FMA(T3I, T3J, T3K * T3L);
|
Chris@10
|
858 T3O = FNMS(T3K, T3J, T3I * T3L);
|
Chris@10
|
859 }
|
Chris@10
|
860 Rp[WS(rs, 7)] = T3H - T3M;
|
Chris@10
|
861 Ip[WS(rs, 7)] = T3N + T3O;
|
Chris@10
|
862 Rm[WS(rs, 7)] = T3H + T3M;
|
Chris@10
|
863 Im[WS(rs, 7)] = T3O - T3N;
|
Chris@10
|
864 }
|
Chris@10
|
865 }
|
Chris@10
|
866 }
|
Chris@10
|
867 }
|
Chris@10
|
868 }
|
Chris@10
|
869
|
Chris@10
|
870 static const tw_instr twinstr[] = {
|
Chris@10
|
871 {TW_FULL, 1, 16},
|
Chris@10
|
872 {TW_NEXT, 1, 0}
|
Chris@10
|
873 };
|
Chris@10
|
874
|
Chris@10
|
875 static const hc2c_desc desc = { 16, "hc2cbdft2_16", twinstr, &GENUS, {168, 46, 38, 0} };
|
Chris@10
|
876
|
Chris@10
|
877 void X(codelet_hc2cbdft2_16) (planner *p) {
|
Chris@10
|
878 X(khc2c_register) (p, hc2cbdft2_16, &desc, HC2C_VIA_DFT);
|
Chris@10
|
879 }
|
Chris@10
|
880 #endif /* HAVE_FMA */
|