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 08:07:52 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_hc2c.native -fma -compact -variables 4 -pipeline-latency 4 -sign 1 -n 16 -dif -name hc2cb_16 -include rdft/scalar/hc2cb.h */
|
cannam@167
|
29
|
cannam@167
|
30 /*
|
cannam@167
|
31 * This function contains 174 FP additions, 100 FP multiplications,
|
cannam@167
|
32 * (or, 104 additions, 30 multiplications, 70 fused multiply/add),
|
cannam@167
|
33 * 63 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 hc2cb_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(KP707106781, +0.707106781186547524400844362104849039284835938);
|
cannam@167
|
41 DK(KP414213562, +0.414213562373095048801688724209698078569671875);
|
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 TA, T1O, T21, T1h, T2P, T2S, T3b, T3p, T3q, T3D, T1k, T1P, Tf, T3y, T2A;
|
cannam@167
|
46 E T36, TL, T22, T3s, T3t, T3z, T2F, T2U, T2K, T2V, Tu, T3E, TX, T1n, T1T;
|
cannam@167
|
47 E T24, T1W, T25, T18, T1m;
|
cannam@167
|
48 {
|
cannam@167
|
49 E T3, Tw, T1g, T2Q, T6, T1d, Tz, T2R, Ta, TB, TE, T2y, Td, TG, TJ;
|
cannam@167
|
50 E T2x;
|
cannam@167
|
51 {
|
cannam@167
|
52 E T1, T2, T1e, T1f;
|
cannam@167
|
53 T1 = Rp[0];
|
cannam@167
|
54 T2 = Rm[WS(rs, 7)];
|
cannam@167
|
55 T3 = T1 + T2;
|
cannam@167
|
56 Tw = T1 - T2;
|
cannam@167
|
57 T1e = Ip[0];
|
cannam@167
|
58 T1f = Im[WS(rs, 7)];
|
cannam@167
|
59 T1g = T1e + T1f;
|
cannam@167
|
60 T2Q = T1e - T1f;
|
cannam@167
|
61 }
|
cannam@167
|
62 {
|
cannam@167
|
63 E T4, T5, Tx, Ty;
|
cannam@167
|
64 T4 = Rp[WS(rs, 4)];
|
cannam@167
|
65 T5 = Rm[WS(rs, 3)];
|
cannam@167
|
66 T6 = T4 + T5;
|
cannam@167
|
67 T1d = T4 - T5;
|
cannam@167
|
68 Tx = Ip[WS(rs, 4)];
|
cannam@167
|
69 Ty = Im[WS(rs, 3)];
|
cannam@167
|
70 Tz = Tx + Ty;
|
cannam@167
|
71 T2R = Tx - Ty;
|
cannam@167
|
72 }
|
cannam@167
|
73 {
|
cannam@167
|
74 E T8, T9, TC, TD;
|
cannam@167
|
75 T8 = Rp[WS(rs, 2)];
|
cannam@167
|
76 T9 = Rm[WS(rs, 5)];
|
cannam@167
|
77 Ta = T8 + T9;
|
cannam@167
|
78 TB = T8 - T9;
|
cannam@167
|
79 TC = Ip[WS(rs, 2)];
|
cannam@167
|
80 TD = Im[WS(rs, 5)];
|
cannam@167
|
81 TE = TC + TD;
|
cannam@167
|
82 T2y = TC - TD;
|
cannam@167
|
83 }
|
cannam@167
|
84 {
|
cannam@167
|
85 E Tb, Tc, TH, TI;
|
cannam@167
|
86 Tb = Rm[WS(rs, 1)];
|
cannam@167
|
87 Tc = Rp[WS(rs, 6)];
|
cannam@167
|
88 Td = Tb + Tc;
|
cannam@167
|
89 TG = Tb - Tc;
|
cannam@167
|
90 TH = Ip[WS(rs, 6)];
|
cannam@167
|
91 TI = Im[WS(rs, 1)];
|
cannam@167
|
92 TJ = TH + TI;
|
cannam@167
|
93 T2x = TH - TI;
|
cannam@167
|
94 }
|
cannam@167
|
95 TA = Tw - Tz;
|
cannam@167
|
96 T1O = Tw + Tz;
|
cannam@167
|
97 T21 = T1g - T1d;
|
cannam@167
|
98 T1h = T1d + T1g;
|
cannam@167
|
99 T2P = Ta - Td;
|
cannam@167
|
100 T2S = T2Q - T2R;
|
cannam@167
|
101 T3b = T2S - T2P;
|
cannam@167
|
102 {
|
cannam@167
|
103 E T1i, T1j, T7, Te;
|
cannam@167
|
104 T3p = T2Q + T2R;
|
cannam@167
|
105 T3q = T2y + T2x;
|
cannam@167
|
106 T3D = T3p - T3q;
|
cannam@167
|
107 T1i = TB + TE;
|
cannam@167
|
108 T1j = TG + TJ;
|
cannam@167
|
109 T1k = T1i - T1j;
|
cannam@167
|
110 T1P = T1i + T1j;
|
cannam@167
|
111 T7 = T3 + T6;
|
cannam@167
|
112 Te = Ta + Td;
|
cannam@167
|
113 Tf = T7 + Te;
|
cannam@167
|
114 T3y = T7 - Te;
|
cannam@167
|
115 {
|
cannam@167
|
116 E T2w, T2z, TF, TK;
|
cannam@167
|
117 T2w = T3 - T6;
|
cannam@167
|
118 T2z = T2x - T2y;
|
cannam@167
|
119 T2A = T2w + T2z;
|
cannam@167
|
120 T36 = T2w - T2z;
|
cannam@167
|
121 TF = TB - TE;
|
cannam@167
|
122 TK = TG - TJ;
|
cannam@167
|
123 TL = TF + TK;
|
cannam@167
|
124 T22 = TF - TK;
|
cannam@167
|
125 }
|
cannam@167
|
126 }
|
cannam@167
|
127 }
|
cannam@167
|
128 {
|
cannam@167
|
129 E Ti, T13, T11, T2C, Tl, TY, T16, T2D, Tp, TS, TQ, T2H, Ts, TN, TV;
|
cannam@167
|
130 E T2I, T2B, T2E;
|
cannam@167
|
131 {
|
cannam@167
|
132 E Tg, Th, TZ, T10;
|
cannam@167
|
133 Tg = Rp[WS(rs, 1)];
|
cannam@167
|
134 Th = Rm[WS(rs, 6)];
|
cannam@167
|
135 Ti = Tg + Th;
|
cannam@167
|
136 T13 = Tg - Th;
|
cannam@167
|
137 TZ = Ip[WS(rs, 1)];
|
cannam@167
|
138 T10 = Im[WS(rs, 6)];
|
cannam@167
|
139 T11 = TZ + T10;
|
cannam@167
|
140 T2C = TZ - T10;
|
cannam@167
|
141 }
|
cannam@167
|
142 {
|
cannam@167
|
143 E Tj, Tk, T14, T15;
|
cannam@167
|
144 Tj = Rp[WS(rs, 5)];
|
cannam@167
|
145 Tk = Rm[WS(rs, 2)];
|
cannam@167
|
146 Tl = Tj + Tk;
|
cannam@167
|
147 TY = Tj - Tk;
|
cannam@167
|
148 T14 = Ip[WS(rs, 5)];
|
cannam@167
|
149 T15 = Im[WS(rs, 2)];
|
cannam@167
|
150 T16 = T14 + T15;
|
cannam@167
|
151 T2D = T14 - T15;
|
cannam@167
|
152 }
|
cannam@167
|
153 {
|
cannam@167
|
154 E Tn, To, TO, TP;
|
cannam@167
|
155 Tn = Rm[0];
|
cannam@167
|
156 To = Rp[WS(rs, 7)];
|
cannam@167
|
157 Tp = Tn + To;
|
cannam@167
|
158 TS = Tn - To;
|
cannam@167
|
159 TO = Ip[WS(rs, 7)];
|
cannam@167
|
160 TP = Im[0];
|
cannam@167
|
161 TQ = TO + TP;
|
cannam@167
|
162 T2H = TO - TP;
|
cannam@167
|
163 }
|
cannam@167
|
164 {
|
cannam@167
|
165 E Tq, Tr, TT, TU;
|
cannam@167
|
166 Tq = Rp[WS(rs, 3)];
|
cannam@167
|
167 Tr = Rm[WS(rs, 4)];
|
cannam@167
|
168 Ts = Tq + Tr;
|
cannam@167
|
169 TN = Tq - Tr;
|
cannam@167
|
170 TT = Ip[WS(rs, 3)];
|
cannam@167
|
171 TU = Im[WS(rs, 4)];
|
cannam@167
|
172 TV = TT + TU;
|
cannam@167
|
173 T2I = TT - TU;
|
cannam@167
|
174 }
|
cannam@167
|
175 T3s = T2C + T2D;
|
cannam@167
|
176 T3t = T2H + T2I;
|
cannam@167
|
177 T3z = T3t - T3s;
|
cannam@167
|
178 T2B = Ti - Tl;
|
cannam@167
|
179 T2E = T2C - T2D;
|
cannam@167
|
180 T2F = T2B - T2E;
|
cannam@167
|
181 T2U = T2B + T2E;
|
cannam@167
|
182 {
|
cannam@167
|
183 E T2G, T2J, Tm, Tt;
|
cannam@167
|
184 T2G = Tp - Ts;
|
cannam@167
|
185 T2J = T2H - T2I;
|
cannam@167
|
186 T2K = T2G + T2J;
|
cannam@167
|
187 T2V = T2J - T2G;
|
cannam@167
|
188 Tm = Ti + Tl;
|
cannam@167
|
189 Tt = Tp + Ts;
|
cannam@167
|
190 Tu = Tm + Tt;
|
cannam@167
|
191 T3E = Tm - Tt;
|
cannam@167
|
192 }
|
cannam@167
|
193 {
|
cannam@167
|
194 E TR, TW, T1R, T1S;
|
cannam@167
|
195 TR = TN - TQ;
|
cannam@167
|
196 TW = TS - TV;
|
cannam@167
|
197 TX = FNMS(KP414213562, TW, TR);
|
cannam@167
|
198 T1n = FMA(KP414213562, TR, TW);
|
cannam@167
|
199 T1R = T11 - TY;
|
cannam@167
|
200 T1S = T13 + T16;
|
cannam@167
|
201 T1T = FNMS(KP414213562, T1S, T1R);
|
cannam@167
|
202 T24 = FMA(KP414213562, T1R, T1S);
|
cannam@167
|
203 }
|
cannam@167
|
204 {
|
cannam@167
|
205 E T1U, T1V, T12, T17;
|
cannam@167
|
206 T1U = TN + TQ;
|
cannam@167
|
207 T1V = TS + TV;
|
cannam@167
|
208 T1W = FNMS(KP414213562, T1V, T1U);
|
cannam@167
|
209 T25 = FMA(KP414213562, T1U, T1V);
|
cannam@167
|
210 T12 = TY + T11;
|
cannam@167
|
211 T17 = T13 - T16;
|
cannam@167
|
212 T18 = FMA(KP414213562, T17, T12);
|
cannam@167
|
213 T1m = FNMS(KP414213562, T12, T17);
|
cannam@167
|
214 }
|
cannam@167
|
215 }
|
cannam@167
|
216 Rp[0] = Tf + Tu;
|
cannam@167
|
217 {
|
cannam@167
|
218 E T3r, T3u, T3v, T3l, T3n, T3o, T3w, T3m;
|
cannam@167
|
219 T3r = T3p + T3q;
|
cannam@167
|
220 T3u = T3s + T3t;
|
cannam@167
|
221 T3v = T3r - T3u;
|
cannam@167
|
222 T3m = Tf - Tu;
|
cannam@167
|
223 T3l = W[14];
|
cannam@167
|
224 T3n = T3l * T3m;
|
cannam@167
|
225 T3o = W[15];
|
cannam@167
|
226 T3w = T3o * T3m;
|
cannam@167
|
227 Rm[0] = T3r + T3u;
|
cannam@167
|
228 Rm[WS(rs, 4)] = FMA(T3l, T3v, T3w);
|
cannam@167
|
229 Rp[WS(rs, 4)] = FNMS(T3o, T3v, T3n);
|
cannam@167
|
230 }
|
cannam@167
|
231 {
|
cannam@167
|
232 E T3A, T3F, T3B, T3G, T3x, T3C;
|
cannam@167
|
233 T3A = T3y - T3z;
|
cannam@167
|
234 T3F = T3D - T3E;
|
cannam@167
|
235 T3x = W[22];
|
cannam@167
|
236 T3B = T3x * T3A;
|
cannam@167
|
237 T3G = T3x * T3F;
|
cannam@167
|
238 T3C = W[23];
|
cannam@167
|
239 Rp[WS(rs, 6)] = FNMS(T3C, T3F, T3B);
|
cannam@167
|
240 Rm[WS(rs, 6)] = FMA(T3C, T3A, T3G);
|
cannam@167
|
241 }
|
cannam@167
|
242 {
|
cannam@167
|
243 E T3I, T3L, T3J, T3M, T3H, T3K;
|
cannam@167
|
244 T3I = T3y + T3z;
|
cannam@167
|
245 T3L = T3E + T3D;
|
cannam@167
|
246 T3H = W[6];
|
cannam@167
|
247 T3J = T3H * T3I;
|
cannam@167
|
248 T3M = T3H * T3L;
|
cannam@167
|
249 T3K = W[7];
|
cannam@167
|
250 Rp[WS(rs, 2)] = FNMS(T3K, T3L, T3J);
|
cannam@167
|
251 Rm[WS(rs, 2)] = FMA(T3K, T3I, T3M);
|
cannam@167
|
252 }
|
cannam@167
|
253 {
|
cannam@167
|
254 E T38, T3g, T3d, T3j, T37, T3c;
|
cannam@167
|
255 T37 = T2V - T2U;
|
cannam@167
|
256 T38 = FNMS(KP707106781, T37, T36);
|
cannam@167
|
257 T3g = FMA(KP707106781, T37, T36);
|
cannam@167
|
258 T3c = T2F - T2K;
|
cannam@167
|
259 T3d = FNMS(KP707106781, T3c, T3b);
|
cannam@167
|
260 T3j = FMA(KP707106781, T3c, T3b);
|
cannam@167
|
261 {
|
cannam@167
|
262 E T39, T3e, T35, T3a;
|
cannam@167
|
263 T35 = W[26];
|
cannam@167
|
264 T39 = T35 * T38;
|
cannam@167
|
265 T3e = T35 * T3d;
|
cannam@167
|
266 T3a = W[27];
|
cannam@167
|
267 Rp[WS(rs, 7)] = FNMS(T3a, T3d, T39);
|
cannam@167
|
268 Rm[WS(rs, 7)] = FMA(T3a, T38, T3e);
|
cannam@167
|
269 }
|
cannam@167
|
270 {
|
cannam@167
|
271 E T3h, T3k, T3f, T3i;
|
cannam@167
|
272 T3f = W[10];
|
cannam@167
|
273 T3h = T3f * T3g;
|
cannam@167
|
274 T3k = T3f * T3j;
|
cannam@167
|
275 T3i = W[11];
|
cannam@167
|
276 Rp[WS(rs, 3)] = FNMS(T3i, T3j, T3h);
|
cannam@167
|
277 Rm[WS(rs, 3)] = FMA(T3i, T3g, T3k);
|
cannam@167
|
278 }
|
cannam@167
|
279 }
|
cannam@167
|
280 {
|
cannam@167
|
281 E T2M, T30, T2X, T33, T2L, T2T, T2W;
|
cannam@167
|
282 T2L = T2F + T2K;
|
cannam@167
|
283 T2M = FNMS(KP707106781, T2L, T2A);
|
cannam@167
|
284 T30 = FMA(KP707106781, T2L, T2A);
|
cannam@167
|
285 T2T = T2P + T2S;
|
cannam@167
|
286 T2W = T2U + T2V;
|
cannam@167
|
287 T2X = FNMS(KP707106781, T2W, T2T);
|
cannam@167
|
288 T33 = FMA(KP707106781, T2W, T2T);
|
cannam@167
|
289 {
|
cannam@167
|
290 E T2v, T2N, T2O, T2Y;
|
cannam@167
|
291 T2v = W[18];
|
cannam@167
|
292 T2N = T2v * T2M;
|
cannam@167
|
293 T2O = W[19];
|
cannam@167
|
294 T2Y = T2O * T2M;
|
cannam@167
|
295 Rp[WS(rs, 5)] = FNMS(T2O, T2X, T2N);
|
cannam@167
|
296 Rm[WS(rs, 5)] = FMA(T2v, T2X, T2Y);
|
cannam@167
|
297 }
|
cannam@167
|
298 {
|
cannam@167
|
299 E T2Z, T31, T32, T34;
|
cannam@167
|
300 T2Z = W[2];
|
cannam@167
|
301 T31 = T2Z * T30;
|
cannam@167
|
302 T32 = W[3];
|
cannam@167
|
303 T34 = T32 * T30;
|
cannam@167
|
304 Rp[WS(rs, 1)] = FNMS(T32, T33, T31);
|
cannam@167
|
305 Rm[WS(rs, 1)] = FMA(T2Z, T33, T34);
|
cannam@167
|
306 }
|
cannam@167
|
307 }
|
cannam@167
|
308 {
|
cannam@167
|
309 E T1Y, T2a, T27, T2d;
|
cannam@167
|
310 {
|
cannam@167
|
311 E T1Q, T1X, T23, T26;
|
cannam@167
|
312 T1Q = FNMS(KP707106781, T1P, T1O);
|
cannam@167
|
313 T1X = T1T + T1W;
|
cannam@167
|
314 T1Y = FMA(KP923879532, T1X, T1Q);
|
cannam@167
|
315 T2a = FNMS(KP923879532, T1X, T1Q);
|
cannam@167
|
316 T23 = FMA(KP707106781, T22, T21);
|
cannam@167
|
317 T26 = T24 - T25;
|
cannam@167
|
318 T27 = FNMS(KP923879532, T26, T23);
|
cannam@167
|
319 T2d = FMA(KP923879532, T26, T23);
|
cannam@167
|
320 }
|
cannam@167
|
321 {
|
cannam@167
|
322 E T1N, T1Z, T20, T28;
|
cannam@167
|
323 T1N = W[20];
|
cannam@167
|
324 T1Z = T1N * T1Y;
|
cannam@167
|
325 T20 = W[21];
|
cannam@167
|
326 T28 = T20 * T1Y;
|
cannam@167
|
327 Ip[WS(rs, 5)] = FNMS(T20, T27, T1Z);
|
cannam@167
|
328 Im[WS(rs, 5)] = FMA(T1N, T27, T28);
|
cannam@167
|
329 }
|
cannam@167
|
330 {
|
cannam@167
|
331 E T29, T2b, T2c, T2e;
|
cannam@167
|
332 T29 = W[4];
|
cannam@167
|
333 T2b = T29 * T2a;
|
cannam@167
|
334 T2c = W[5];
|
cannam@167
|
335 T2e = T2c * T2a;
|
cannam@167
|
336 Ip[WS(rs, 1)] = FNMS(T2c, T2d, T2b);
|
cannam@167
|
337 Im[WS(rs, 1)] = FMA(T29, T2d, T2e);
|
cannam@167
|
338 }
|
cannam@167
|
339 }
|
cannam@167
|
340 {
|
cannam@167
|
341 E T1a, T1s, T1p, T1v;
|
cannam@167
|
342 {
|
cannam@167
|
343 E TM, T19, T1l, T1o;
|
cannam@167
|
344 TM = FNMS(KP707106781, TL, TA);
|
cannam@167
|
345 T19 = TX - T18;
|
cannam@167
|
346 T1a = FNMS(KP923879532, T19, TM);
|
cannam@167
|
347 T1s = FMA(KP923879532, T19, TM);
|
cannam@167
|
348 T1l = FNMS(KP707106781, T1k, T1h);
|
cannam@167
|
349 T1o = T1m - T1n;
|
cannam@167
|
350 T1p = FNMS(KP923879532, T1o, T1l);
|
cannam@167
|
351 T1v = FMA(KP923879532, T1o, T1l);
|
cannam@167
|
352 }
|
cannam@167
|
353 {
|
cannam@167
|
354 E Tv, T1b, T1c, T1q;
|
cannam@167
|
355 Tv = W[24];
|
cannam@167
|
356 T1b = Tv * T1a;
|
cannam@167
|
357 T1c = W[25];
|
cannam@167
|
358 T1q = T1c * T1a;
|
cannam@167
|
359 Ip[WS(rs, 6)] = FNMS(T1c, T1p, T1b);
|
cannam@167
|
360 Im[WS(rs, 6)] = FMA(Tv, T1p, T1q);
|
cannam@167
|
361 }
|
cannam@167
|
362 {
|
cannam@167
|
363 E T1r, T1t, T1u, T1w;
|
cannam@167
|
364 T1r = W[8];
|
cannam@167
|
365 T1t = T1r * T1s;
|
cannam@167
|
366 T1u = W[9];
|
cannam@167
|
367 T1w = T1u * T1s;
|
cannam@167
|
368 Ip[WS(rs, 2)] = FNMS(T1u, T1v, T1t);
|
cannam@167
|
369 Im[WS(rs, 2)] = FMA(T1r, T1v, T1w);
|
cannam@167
|
370 }
|
cannam@167
|
371 }
|
cannam@167
|
372 {
|
cannam@167
|
373 E T2i, T2q, T2n, T2t;
|
cannam@167
|
374 {
|
cannam@167
|
375 E T2g, T2h, T2l, T2m;
|
cannam@167
|
376 T2g = FMA(KP707106781, T1P, T1O);
|
cannam@167
|
377 T2h = T24 + T25;
|
cannam@167
|
378 T2i = FNMS(KP923879532, T2h, T2g);
|
cannam@167
|
379 T2q = FMA(KP923879532, T2h, T2g);
|
cannam@167
|
380 T2l = FNMS(KP707106781, T22, T21);
|
cannam@167
|
381 T2m = T1W - T1T;
|
cannam@167
|
382 T2n = FMA(KP923879532, T2m, T2l);
|
cannam@167
|
383 T2t = FNMS(KP923879532, T2m, T2l);
|
cannam@167
|
384 }
|
cannam@167
|
385 {
|
cannam@167
|
386 E T2j, T2o, T2f, T2k;
|
cannam@167
|
387 T2f = W[12];
|
cannam@167
|
388 T2j = T2f * T2i;
|
cannam@167
|
389 T2o = T2f * T2n;
|
cannam@167
|
390 T2k = W[13];
|
cannam@167
|
391 Ip[WS(rs, 3)] = FNMS(T2k, T2n, T2j);
|
cannam@167
|
392 Im[WS(rs, 3)] = FMA(T2k, T2i, T2o);
|
cannam@167
|
393 }
|
cannam@167
|
394 {
|
cannam@167
|
395 E T2r, T2u, T2p, T2s;
|
cannam@167
|
396 T2p = W[28];
|
cannam@167
|
397 T2r = T2p * T2q;
|
cannam@167
|
398 T2u = T2p * T2t;
|
cannam@167
|
399 T2s = W[29];
|
cannam@167
|
400 Ip[WS(rs, 7)] = FNMS(T2s, T2t, T2r);
|
cannam@167
|
401 Im[WS(rs, 7)] = FMA(T2s, T2q, T2u);
|
cannam@167
|
402 }
|
cannam@167
|
403 }
|
cannam@167
|
404 {
|
cannam@167
|
405 E T1A, T1I, T1F, T1L;
|
cannam@167
|
406 {
|
cannam@167
|
407 E T1y, T1z, T1D, T1E;
|
cannam@167
|
408 T1y = FMA(KP707106781, TL, TA);
|
cannam@167
|
409 T1z = T1m + T1n;
|
cannam@167
|
410 T1A = FNMS(KP923879532, T1z, T1y);
|
cannam@167
|
411 T1I = FMA(KP923879532, T1z, T1y);
|
cannam@167
|
412 T1D = FMA(KP707106781, T1k, T1h);
|
cannam@167
|
413 T1E = T18 + TX;
|
cannam@167
|
414 T1F = FNMS(KP923879532, T1E, T1D);
|
cannam@167
|
415 T1L = FMA(KP923879532, T1E, T1D);
|
cannam@167
|
416 }
|
cannam@167
|
417 {
|
cannam@167
|
418 E T1B, T1G, T1x, T1C;
|
cannam@167
|
419 T1x = W[16];
|
cannam@167
|
420 T1B = T1x * T1A;
|
cannam@167
|
421 T1G = T1x * T1F;
|
cannam@167
|
422 T1C = W[17];
|
cannam@167
|
423 Ip[WS(rs, 4)] = FNMS(T1C, T1F, T1B);
|
cannam@167
|
424 Im[WS(rs, 4)] = FMA(T1C, T1A, T1G);
|
cannam@167
|
425 }
|
cannam@167
|
426 {
|
cannam@167
|
427 E T1J, T1M, T1H, T1K;
|
cannam@167
|
428 T1H = W[0];
|
cannam@167
|
429 T1J = T1H * T1I;
|
cannam@167
|
430 T1M = T1H * T1L;
|
cannam@167
|
431 T1K = W[1];
|
cannam@167
|
432 Ip[0] = FNMS(T1K, T1L, T1J);
|
cannam@167
|
433 Im[0] = FMA(T1K, T1I, T1M);
|
cannam@167
|
434 }
|
cannam@167
|
435 }
|
cannam@167
|
436 }
|
cannam@167
|
437 }
|
cannam@167
|
438 }
|
cannam@167
|
439
|
cannam@167
|
440 static const tw_instr twinstr[] = {
|
cannam@167
|
441 {TW_FULL, 1, 16},
|
cannam@167
|
442 {TW_NEXT, 1, 0}
|
cannam@167
|
443 };
|
cannam@167
|
444
|
cannam@167
|
445 static const hc2c_desc desc = { 16, "hc2cb_16", twinstr, &GENUS, {104, 30, 70, 0} };
|
cannam@167
|
446
|
cannam@167
|
447 void X(codelet_hc2cb_16) (planner *p) {
|
cannam@167
|
448 X(khc2c_register) (p, hc2cb_16, &desc, HC2C_VIA_RDFT);
|
cannam@167
|
449 }
|
cannam@167
|
450 #else
|
cannam@167
|
451
|
cannam@167
|
452 /* Generated by: ../../../genfft/gen_hc2c.native -compact -variables 4 -pipeline-latency 4 -sign 1 -n 16 -dif -name hc2cb_16 -include rdft/scalar/hc2cb.h */
|
cannam@167
|
453
|
cannam@167
|
454 /*
|
cannam@167
|
455 * This function contains 174 FP additions, 84 FP multiplications,
|
cannam@167
|
456 * (or, 136 additions, 46 multiplications, 38 fused multiply/add),
|
cannam@167
|
457 * 50 stack variables, 3 constants, and 64 memory accesses
|
cannam@167
|
458 */
|
cannam@167
|
459 #include "rdft/scalar/hc2cb.h"
|
cannam@167
|
460
|
cannam@167
|
461 static void hc2cb_16(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms)
|
cannam@167
|
462 {
|
cannam@167
|
463 DK(KP382683432, +0.382683432365089771728459984030398866761344562);
|
cannam@167
|
464 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
cannam@167
|
465 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
cannam@167
|
466 {
|
cannam@167
|
467 INT m;
|
cannam@167
|
468 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
|
469 E T7, T2K, T2W, Tw, T17, T1S, T2k, T1w, Te, TD, T1x, T10, T2n, T2L, T1Z;
|
cannam@167
|
470 E T2X, Tm, T1z, TN, T19, T2e, T2p, T2P, T2Z, Tt, T1A, TW, T1a, T27, T2q;
|
cannam@167
|
471 E T2S, T30;
|
cannam@167
|
472 {
|
cannam@167
|
473 E T3, T1Q, T13, T2j, T6, T2i, T16, T1R;
|
cannam@167
|
474 {
|
cannam@167
|
475 E T1, T2, T11, T12;
|
cannam@167
|
476 T1 = Rp[0];
|
cannam@167
|
477 T2 = Rm[WS(rs, 7)];
|
cannam@167
|
478 T3 = T1 + T2;
|
cannam@167
|
479 T1Q = T1 - T2;
|
cannam@167
|
480 T11 = Ip[0];
|
cannam@167
|
481 T12 = Im[WS(rs, 7)];
|
cannam@167
|
482 T13 = T11 - T12;
|
cannam@167
|
483 T2j = T11 + T12;
|
cannam@167
|
484 }
|
cannam@167
|
485 {
|
cannam@167
|
486 E T4, T5, T14, T15;
|
cannam@167
|
487 T4 = Rp[WS(rs, 4)];
|
cannam@167
|
488 T5 = Rm[WS(rs, 3)];
|
cannam@167
|
489 T6 = T4 + T5;
|
cannam@167
|
490 T2i = T4 - T5;
|
cannam@167
|
491 T14 = Ip[WS(rs, 4)];
|
cannam@167
|
492 T15 = Im[WS(rs, 3)];
|
cannam@167
|
493 T16 = T14 - T15;
|
cannam@167
|
494 T1R = T14 + T15;
|
cannam@167
|
495 }
|
cannam@167
|
496 T7 = T3 + T6;
|
cannam@167
|
497 T2K = T1Q + T1R;
|
cannam@167
|
498 T2W = T2j - T2i;
|
cannam@167
|
499 Tw = T3 - T6;
|
cannam@167
|
500 T17 = T13 - T16;
|
cannam@167
|
501 T1S = T1Q - T1R;
|
cannam@167
|
502 T2k = T2i + T2j;
|
cannam@167
|
503 T1w = T13 + T16;
|
cannam@167
|
504 }
|
cannam@167
|
505 {
|
cannam@167
|
506 E Ta, T1T, TC, T1U, Td, T1W, Tz, T1X;
|
cannam@167
|
507 {
|
cannam@167
|
508 E T8, T9, TA, TB;
|
cannam@167
|
509 T8 = Rp[WS(rs, 2)];
|
cannam@167
|
510 T9 = Rm[WS(rs, 5)];
|
cannam@167
|
511 Ta = T8 + T9;
|
cannam@167
|
512 T1T = T8 - T9;
|
cannam@167
|
513 TA = Ip[WS(rs, 2)];
|
cannam@167
|
514 TB = Im[WS(rs, 5)];
|
cannam@167
|
515 TC = TA - TB;
|
cannam@167
|
516 T1U = TA + TB;
|
cannam@167
|
517 }
|
cannam@167
|
518 {
|
cannam@167
|
519 E Tb, Tc, Tx, Ty;
|
cannam@167
|
520 Tb = Rm[WS(rs, 1)];
|
cannam@167
|
521 Tc = Rp[WS(rs, 6)];
|
cannam@167
|
522 Td = Tb + Tc;
|
cannam@167
|
523 T1W = Tb - Tc;
|
cannam@167
|
524 Tx = Ip[WS(rs, 6)];
|
cannam@167
|
525 Ty = Im[WS(rs, 1)];
|
cannam@167
|
526 Tz = Tx - Ty;
|
cannam@167
|
527 T1X = Tx + Ty;
|
cannam@167
|
528 }
|
cannam@167
|
529 Te = Ta + Td;
|
cannam@167
|
530 TD = Tz - TC;
|
cannam@167
|
531 T1x = TC + Tz;
|
cannam@167
|
532 T10 = Ta - Td;
|
cannam@167
|
533 {
|
cannam@167
|
534 E T2l, T2m, T1V, T1Y;
|
cannam@167
|
535 T2l = T1T + T1U;
|
cannam@167
|
536 T2m = T1W + T1X;
|
cannam@167
|
537 T2n = KP707106781 * (T2l - T2m);
|
cannam@167
|
538 T2L = KP707106781 * (T2l + T2m);
|
cannam@167
|
539 T1V = T1T - T1U;
|
cannam@167
|
540 T1Y = T1W - T1X;
|
cannam@167
|
541 T1Z = KP707106781 * (T1V + T1Y);
|
cannam@167
|
542 T2X = KP707106781 * (T1V - T1Y);
|
cannam@167
|
543 }
|
cannam@167
|
544 }
|
cannam@167
|
545 {
|
cannam@167
|
546 E Ti, T2b, TI, T29, Tl, T28, TL, T2c, TF, TM;
|
cannam@167
|
547 {
|
cannam@167
|
548 E Tg, Th, TG, TH;
|
cannam@167
|
549 Tg = Rp[WS(rs, 1)];
|
cannam@167
|
550 Th = Rm[WS(rs, 6)];
|
cannam@167
|
551 Ti = Tg + Th;
|
cannam@167
|
552 T2b = Tg - Th;
|
cannam@167
|
553 TG = Ip[WS(rs, 1)];
|
cannam@167
|
554 TH = Im[WS(rs, 6)];
|
cannam@167
|
555 TI = TG - TH;
|
cannam@167
|
556 T29 = TG + TH;
|
cannam@167
|
557 }
|
cannam@167
|
558 {
|
cannam@167
|
559 E Tj, Tk, TJ, TK;
|
cannam@167
|
560 Tj = Rp[WS(rs, 5)];
|
cannam@167
|
561 Tk = Rm[WS(rs, 2)];
|
cannam@167
|
562 Tl = Tj + Tk;
|
cannam@167
|
563 T28 = Tj - Tk;
|
cannam@167
|
564 TJ = Ip[WS(rs, 5)];
|
cannam@167
|
565 TK = Im[WS(rs, 2)];
|
cannam@167
|
566 TL = TJ - TK;
|
cannam@167
|
567 T2c = TJ + TK;
|
cannam@167
|
568 }
|
cannam@167
|
569 Tm = Ti + Tl;
|
cannam@167
|
570 T1z = TI + TL;
|
cannam@167
|
571 TF = Ti - Tl;
|
cannam@167
|
572 TM = TI - TL;
|
cannam@167
|
573 TN = TF - TM;
|
cannam@167
|
574 T19 = TF + TM;
|
cannam@167
|
575 {
|
cannam@167
|
576 E T2a, T2d, T2N, T2O;
|
cannam@167
|
577 T2a = T28 + T29;
|
cannam@167
|
578 T2d = T2b - T2c;
|
cannam@167
|
579 T2e = FMA(KP923879532, T2a, KP382683432 * T2d);
|
cannam@167
|
580 T2p = FNMS(KP382683432, T2a, KP923879532 * T2d);
|
cannam@167
|
581 T2N = T2b + T2c;
|
cannam@167
|
582 T2O = T29 - T28;
|
cannam@167
|
583 T2P = FNMS(KP923879532, T2O, KP382683432 * T2N);
|
cannam@167
|
584 T2Z = FMA(KP382683432, T2O, KP923879532 * T2N);
|
cannam@167
|
585 }
|
cannam@167
|
586 }
|
cannam@167
|
587 {
|
cannam@167
|
588 E Tp, T24, TR, T22, Ts, T21, TU, T25, TO, TV;
|
cannam@167
|
589 {
|
cannam@167
|
590 E Tn, To, TP, TQ;
|
cannam@167
|
591 Tn = Rm[0];
|
cannam@167
|
592 To = Rp[WS(rs, 7)];
|
cannam@167
|
593 Tp = Tn + To;
|
cannam@167
|
594 T24 = Tn - To;
|
cannam@167
|
595 TP = Ip[WS(rs, 7)];
|
cannam@167
|
596 TQ = Im[0];
|
cannam@167
|
597 TR = TP - TQ;
|
cannam@167
|
598 T22 = TP + TQ;
|
cannam@167
|
599 }
|
cannam@167
|
600 {
|
cannam@167
|
601 E Tq, Tr, TS, TT;
|
cannam@167
|
602 Tq = Rp[WS(rs, 3)];
|
cannam@167
|
603 Tr = Rm[WS(rs, 4)];
|
cannam@167
|
604 Ts = Tq + Tr;
|
cannam@167
|
605 T21 = Tq - Tr;
|
cannam@167
|
606 TS = Ip[WS(rs, 3)];
|
cannam@167
|
607 TT = Im[WS(rs, 4)];
|
cannam@167
|
608 TU = TS - TT;
|
cannam@167
|
609 T25 = TS + TT;
|
cannam@167
|
610 }
|
cannam@167
|
611 Tt = Tp + Ts;
|
cannam@167
|
612 T1A = TR + TU;
|
cannam@167
|
613 TO = Tp - Ts;
|
cannam@167
|
614 TV = TR - TU;
|
cannam@167
|
615 TW = TO + TV;
|
cannam@167
|
616 T1a = TV - TO;
|
cannam@167
|
617 {
|
cannam@167
|
618 E T23, T26, T2Q, T2R;
|
cannam@167
|
619 T23 = T21 - T22;
|
cannam@167
|
620 T26 = T24 - T25;
|
cannam@167
|
621 T27 = FNMS(KP382683432, T26, KP923879532 * T23);
|
cannam@167
|
622 T2q = FMA(KP382683432, T23, KP923879532 * T26);
|
cannam@167
|
623 T2Q = T24 + T25;
|
cannam@167
|
624 T2R = T21 + T22;
|
cannam@167
|
625 T2S = FNMS(KP923879532, T2R, KP382683432 * T2Q);
|
cannam@167
|
626 T30 = FMA(KP382683432, T2R, KP923879532 * T2Q);
|
cannam@167
|
627 }
|
cannam@167
|
628 }
|
cannam@167
|
629 {
|
cannam@167
|
630 E Tf, Tu, T1u, T1y, T1B, T1C, T1t, T1v;
|
cannam@167
|
631 Tf = T7 + Te;
|
cannam@167
|
632 Tu = Tm + Tt;
|
cannam@167
|
633 T1u = Tf - Tu;
|
cannam@167
|
634 T1y = T1w + T1x;
|
cannam@167
|
635 T1B = T1z + T1A;
|
cannam@167
|
636 T1C = T1y - T1B;
|
cannam@167
|
637 Rp[0] = Tf + Tu;
|
cannam@167
|
638 Rm[0] = T1y + T1B;
|
cannam@167
|
639 T1t = W[14];
|
cannam@167
|
640 T1v = W[15];
|
cannam@167
|
641 Rp[WS(rs, 4)] = FNMS(T1v, T1C, T1t * T1u);
|
cannam@167
|
642 Rm[WS(rs, 4)] = FMA(T1v, T1u, T1t * T1C);
|
cannam@167
|
643 }
|
cannam@167
|
644 {
|
cannam@167
|
645 E T2U, T34, T32, T36;
|
cannam@167
|
646 {
|
cannam@167
|
647 E T2M, T2T, T2Y, T31;
|
cannam@167
|
648 T2M = T2K - T2L;
|
cannam@167
|
649 T2T = T2P + T2S;
|
cannam@167
|
650 T2U = T2M - T2T;
|
cannam@167
|
651 T34 = T2M + T2T;
|
cannam@167
|
652 T2Y = T2W + T2X;
|
cannam@167
|
653 T31 = T2Z - T30;
|
cannam@167
|
654 T32 = T2Y - T31;
|
cannam@167
|
655 T36 = T2Y + T31;
|
cannam@167
|
656 }
|
cannam@167
|
657 {
|
cannam@167
|
658 E T2J, T2V, T33, T35;
|
cannam@167
|
659 T2J = W[20];
|
cannam@167
|
660 T2V = W[21];
|
cannam@167
|
661 Ip[WS(rs, 5)] = FNMS(T2V, T32, T2J * T2U);
|
cannam@167
|
662 Im[WS(rs, 5)] = FMA(T2V, T2U, T2J * T32);
|
cannam@167
|
663 T33 = W[4];
|
cannam@167
|
664 T35 = W[5];
|
cannam@167
|
665 Ip[WS(rs, 1)] = FNMS(T35, T36, T33 * T34);
|
cannam@167
|
666 Im[WS(rs, 1)] = FMA(T35, T34, T33 * T36);
|
cannam@167
|
667 }
|
cannam@167
|
668 }
|
cannam@167
|
669 {
|
cannam@167
|
670 E T3a, T3g, T3e, T3i;
|
cannam@167
|
671 {
|
cannam@167
|
672 E T38, T39, T3c, T3d;
|
cannam@167
|
673 T38 = T2K + T2L;
|
cannam@167
|
674 T39 = T2Z + T30;
|
cannam@167
|
675 T3a = T38 - T39;
|
cannam@167
|
676 T3g = T38 + T39;
|
cannam@167
|
677 T3c = T2W - T2X;
|
cannam@167
|
678 T3d = T2P - T2S;
|
cannam@167
|
679 T3e = T3c + T3d;
|
cannam@167
|
680 T3i = T3c - T3d;
|
cannam@167
|
681 }
|
cannam@167
|
682 {
|
cannam@167
|
683 E T37, T3b, T3f, T3h;
|
cannam@167
|
684 T37 = W[12];
|
cannam@167
|
685 T3b = W[13];
|
cannam@167
|
686 Ip[WS(rs, 3)] = FNMS(T3b, T3e, T37 * T3a);
|
cannam@167
|
687 Im[WS(rs, 3)] = FMA(T37, T3e, T3b * T3a);
|
cannam@167
|
688 T3f = W[28];
|
cannam@167
|
689 T3h = W[29];
|
cannam@167
|
690 Ip[WS(rs, 7)] = FNMS(T3h, T3i, T3f * T3g);
|
cannam@167
|
691 Im[WS(rs, 7)] = FMA(T3f, T3i, T3h * T3g);
|
cannam@167
|
692 }
|
cannam@167
|
693 }
|
cannam@167
|
694 {
|
cannam@167
|
695 E TY, T1e, T1c, T1g;
|
cannam@167
|
696 {
|
cannam@167
|
697 E TE, TX, T18, T1b;
|
cannam@167
|
698 TE = Tw + TD;
|
cannam@167
|
699 TX = KP707106781 * (TN + TW);
|
cannam@167
|
700 TY = TE - TX;
|
cannam@167
|
701 T1e = TE + TX;
|
cannam@167
|
702 T18 = T10 + T17;
|
cannam@167
|
703 T1b = KP707106781 * (T19 + T1a);
|
cannam@167
|
704 T1c = T18 - T1b;
|
cannam@167
|
705 T1g = T18 + T1b;
|
cannam@167
|
706 }
|
cannam@167
|
707 {
|
cannam@167
|
708 E Tv, TZ, T1d, T1f;
|
cannam@167
|
709 Tv = W[18];
|
cannam@167
|
710 TZ = W[19];
|
cannam@167
|
711 Rp[WS(rs, 5)] = FNMS(TZ, T1c, Tv * TY);
|
cannam@167
|
712 Rm[WS(rs, 5)] = FMA(TZ, TY, Tv * T1c);
|
cannam@167
|
713 T1d = W[2];
|
cannam@167
|
714 T1f = W[3];
|
cannam@167
|
715 Rp[WS(rs, 1)] = FNMS(T1f, T1g, T1d * T1e);
|
cannam@167
|
716 Rm[WS(rs, 1)] = FMA(T1f, T1e, T1d * T1g);
|
cannam@167
|
717 }
|
cannam@167
|
718 }
|
cannam@167
|
719 {
|
cannam@167
|
720 E T1k, T1q, T1o, T1s;
|
cannam@167
|
721 {
|
cannam@167
|
722 E T1i, T1j, T1m, T1n;
|
cannam@167
|
723 T1i = Tw - TD;
|
cannam@167
|
724 T1j = KP707106781 * (T1a - T19);
|
cannam@167
|
725 T1k = T1i - T1j;
|
cannam@167
|
726 T1q = T1i + T1j;
|
cannam@167
|
727 T1m = T17 - T10;
|
cannam@167
|
728 T1n = KP707106781 * (TN - TW);
|
cannam@167
|
729 T1o = T1m - T1n;
|
cannam@167
|
730 T1s = T1m + T1n;
|
cannam@167
|
731 }
|
cannam@167
|
732 {
|
cannam@167
|
733 E T1h, T1l, T1p, T1r;
|
cannam@167
|
734 T1h = W[26];
|
cannam@167
|
735 T1l = W[27];
|
cannam@167
|
736 Rp[WS(rs, 7)] = FNMS(T1l, T1o, T1h * T1k);
|
cannam@167
|
737 Rm[WS(rs, 7)] = FMA(T1h, T1o, T1l * T1k);
|
cannam@167
|
738 T1p = W[10];
|
cannam@167
|
739 T1r = W[11];
|
cannam@167
|
740 Rp[WS(rs, 3)] = FNMS(T1r, T1s, T1p * T1q);
|
cannam@167
|
741 Rm[WS(rs, 3)] = FMA(T1p, T1s, T1r * T1q);
|
cannam@167
|
742 }
|
cannam@167
|
743 }
|
cannam@167
|
744 {
|
cannam@167
|
745 E T2g, T2u, T2s, T2w;
|
cannam@167
|
746 {
|
cannam@167
|
747 E T20, T2f, T2o, T2r;
|
cannam@167
|
748 T20 = T1S - T1Z;
|
cannam@167
|
749 T2f = T27 - T2e;
|
cannam@167
|
750 T2g = T20 - T2f;
|
cannam@167
|
751 T2u = T20 + T2f;
|
cannam@167
|
752 T2o = T2k - T2n;
|
cannam@167
|
753 T2r = T2p - T2q;
|
cannam@167
|
754 T2s = T2o - T2r;
|
cannam@167
|
755 T2w = T2o + T2r;
|
cannam@167
|
756 }
|
cannam@167
|
757 {
|
cannam@167
|
758 E T1P, T2h, T2t, T2v;
|
cannam@167
|
759 T1P = W[24];
|
cannam@167
|
760 T2h = W[25];
|
cannam@167
|
761 Ip[WS(rs, 6)] = FNMS(T2h, T2s, T1P * T2g);
|
cannam@167
|
762 Im[WS(rs, 6)] = FMA(T2h, T2g, T1P * T2s);
|
cannam@167
|
763 T2t = W[8];
|
cannam@167
|
764 T2v = W[9];
|
cannam@167
|
765 Ip[WS(rs, 2)] = FNMS(T2v, T2w, T2t * T2u);
|
cannam@167
|
766 Im[WS(rs, 2)] = FMA(T2v, T2u, T2t * T2w);
|
cannam@167
|
767 }
|
cannam@167
|
768 }
|
cannam@167
|
769 {
|
cannam@167
|
770 E T2A, T2G, T2E, T2I;
|
cannam@167
|
771 {
|
cannam@167
|
772 E T2y, T2z, T2C, T2D;
|
cannam@167
|
773 T2y = T1S + T1Z;
|
cannam@167
|
774 T2z = T2p + T2q;
|
cannam@167
|
775 T2A = T2y - T2z;
|
cannam@167
|
776 T2G = T2y + T2z;
|
cannam@167
|
777 T2C = T2k + T2n;
|
cannam@167
|
778 T2D = T2e + T27;
|
cannam@167
|
779 T2E = T2C - T2D;
|
cannam@167
|
780 T2I = T2C + T2D;
|
cannam@167
|
781 }
|
cannam@167
|
782 {
|
cannam@167
|
783 E T2x, T2B, T2F, T2H;
|
cannam@167
|
784 T2x = W[16];
|
cannam@167
|
785 T2B = W[17];
|
cannam@167
|
786 Ip[WS(rs, 4)] = FNMS(T2B, T2E, T2x * T2A);
|
cannam@167
|
787 Im[WS(rs, 4)] = FMA(T2x, T2E, T2B * T2A);
|
cannam@167
|
788 T2F = W[0];
|
cannam@167
|
789 T2H = W[1];
|
cannam@167
|
790 Ip[0] = FNMS(T2H, T2I, T2F * T2G);
|
cannam@167
|
791 Im[0] = FMA(T2F, T2I, T2H * T2G);
|
cannam@167
|
792 }
|
cannam@167
|
793 }
|
cannam@167
|
794 {
|
cannam@167
|
795 E T1G, T1M, T1K, T1O;
|
cannam@167
|
796 {
|
cannam@167
|
797 E T1E, T1F, T1I, T1J;
|
cannam@167
|
798 T1E = T7 - Te;
|
cannam@167
|
799 T1F = T1A - T1z;
|
cannam@167
|
800 T1G = T1E - T1F;
|
cannam@167
|
801 T1M = T1E + T1F;
|
cannam@167
|
802 T1I = T1w - T1x;
|
cannam@167
|
803 T1J = Tm - Tt;
|
cannam@167
|
804 T1K = T1I - T1J;
|
cannam@167
|
805 T1O = T1J + T1I;
|
cannam@167
|
806 }
|
cannam@167
|
807 {
|
cannam@167
|
808 E T1D, T1H, T1L, T1N;
|
cannam@167
|
809 T1D = W[22];
|
cannam@167
|
810 T1H = W[23];
|
cannam@167
|
811 Rp[WS(rs, 6)] = FNMS(T1H, T1K, T1D * T1G);
|
cannam@167
|
812 Rm[WS(rs, 6)] = FMA(T1D, T1K, T1H * T1G);
|
cannam@167
|
813 T1L = W[6];
|
cannam@167
|
814 T1N = W[7];
|
cannam@167
|
815 Rp[WS(rs, 2)] = FNMS(T1N, T1O, T1L * T1M);
|
cannam@167
|
816 Rm[WS(rs, 2)] = FMA(T1L, T1O, T1N * T1M);
|
cannam@167
|
817 }
|
cannam@167
|
818 }
|
cannam@167
|
819 }
|
cannam@167
|
820 }
|
cannam@167
|
821 }
|
cannam@167
|
822
|
cannam@167
|
823 static const tw_instr twinstr[] = {
|
cannam@167
|
824 {TW_FULL, 1, 16},
|
cannam@167
|
825 {TW_NEXT, 1, 0}
|
cannam@167
|
826 };
|
cannam@167
|
827
|
cannam@167
|
828 static const hc2c_desc desc = { 16, "hc2cb_16", twinstr, &GENUS, {136, 46, 38, 0} };
|
cannam@167
|
829
|
cannam@167
|
830 void X(codelet_hc2cb_16) (planner *p) {
|
cannam@167
|
831 X(khc2c_register) (p, hc2cb_16, &desc, HC2C_VIA_RDFT);
|
cannam@167
|
832 }
|
cannam@167
|
833 #endif
|