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 07:41:57 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_hc2c.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -sign 1 -twiddle-log3 -precompute-twiddles -n 16 -dif -name hc2cb2_16 -include hc2cb.h */
|
Chris@10
|
29
|
Chris@10
|
30 /*
|
Chris@10
|
31 * This function contains 196 FP additions, 134 FP multiplications,
|
Chris@10
|
32 * (or, 104 additions, 42 multiplications, 92 fused multiply/add),
|
Chris@10
|
33 * 112 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 hc2cb2_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(KP707106781, +0.707106781186547524400844362104849039284835938);
|
Chris@10
|
41 DK(KP414213562, +0.414213562373095048801688724209698078569671875);
|
Chris@10
|
42 {
|
Chris@10
|
43 INT m;
|
Chris@10
|
44 for (m = mb, W = W + ((mb - 1) * 8); m < me; m = m + 1, Rp = Rp + ms, Ip = Ip + ms, Rm = Rm - ms, Im = Im - ms, W = W + 8, MAKE_VOLATILE_STRIDE(64, rs)) {
|
Chris@10
|
45 E Tv, TB, TF, Ty, T1J, T1O, T1N, T1K;
|
Chris@10
|
46 {
|
Chris@10
|
47 E Tw, T2z, T2C, Tx, T3f, T3l, T2F, T3r, Tz;
|
Chris@10
|
48 Tv = W[0];
|
Chris@10
|
49 Tw = W[2];
|
Chris@10
|
50 T2z = W[6];
|
Chris@10
|
51 T2C = W[7];
|
Chris@10
|
52 TB = W[4];
|
Chris@10
|
53 Tx = Tv * Tw;
|
Chris@10
|
54 T3f = Tv * T2z;
|
Chris@10
|
55 T3l = Tv * T2C;
|
Chris@10
|
56 T2F = Tv * TB;
|
Chris@10
|
57 T3r = Tw * TB;
|
Chris@10
|
58 TF = W[5];
|
Chris@10
|
59 Ty = W[1];
|
Chris@10
|
60 Tz = W[3];
|
Chris@10
|
61 {
|
Chris@10
|
62 E T2G, T3z, T3m, T3g, T3L, T3s, T1V, TA, T3w, T3Q, T30, T3C, TE, T1X, T1D;
|
Chris@10
|
63 E TG, T1G, T1o, T2p, T1Y, T2u, T2c, T1Z, TL, T1t, T2d, T3n, T35, T3R, T3F;
|
Chris@10
|
64 E T1w, T20, T3M, Tf, T3h, T2L, T2e, TW, T2Q, T36, T3I, T3N, T2V, T37, T1d;
|
Chris@10
|
65 E Tu, T3S, T18, T1z, T1i, T24, T2g, T27, T2h;
|
Chris@10
|
66 {
|
Chris@10
|
67 E T2K, TQ, TV, T2H;
|
Chris@10
|
68 {
|
Chris@10
|
69 E TH, T3, T32, T1s, T1p, T6, T33, TK, TM, Ta, TS, T2J, TP, TR, Td;
|
Chris@10
|
70 E TT, TI, TJ;
|
Chris@10
|
71 {
|
Chris@10
|
72 E T1q, T1r, T4, T5;
|
Chris@10
|
73 {
|
Chris@10
|
74 E T1, T1n, TC, T2b, T1W, T2, T3v, T2Z, TD;
|
Chris@10
|
75 T1 = Rp[0];
|
Chris@10
|
76 T3v = Tw * TF;
|
Chris@10
|
77 T2Z = Tv * TF;
|
Chris@10
|
78 T2G = FNMS(Ty, TF, T2F);
|
Chris@10
|
79 T3z = FMA(Ty, TF, T2F);
|
Chris@10
|
80 T3m = FNMS(Ty, T2z, T3l);
|
Chris@10
|
81 T3g = FMA(Ty, T2C, T3f);
|
Chris@10
|
82 T3L = FNMS(Tz, TF, T3r);
|
Chris@10
|
83 T3s = FMA(Tz, TF, T3r);
|
Chris@10
|
84 T1V = FMA(Ty, Tz, Tx);
|
Chris@10
|
85 TA = FNMS(Ty, Tz, Tx);
|
Chris@10
|
86 TD = Tv * Tz;
|
Chris@10
|
87 T3w = FNMS(Tz, TB, T3v);
|
Chris@10
|
88 T3Q = FMA(Tz, TB, T3v);
|
Chris@10
|
89 T30 = FMA(Ty, TB, T2Z);
|
Chris@10
|
90 T3C = FNMS(Ty, TB, T2Z);
|
Chris@10
|
91 T1n = TA * TF;
|
Chris@10
|
92 TC = TA * TB;
|
Chris@10
|
93 T2b = T1V * TF;
|
Chris@10
|
94 T1W = T1V * TB;
|
Chris@10
|
95 TE = FMA(Ty, Tw, TD);
|
Chris@10
|
96 T1X = FNMS(Ty, Tw, TD);
|
Chris@10
|
97 T2 = Rm[WS(rs, 7)];
|
Chris@10
|
98 T1q = Ip[0];
|
Chris@10
|
99 T1D = FMA(TE, TF, TC);
|
Chris@10
|
100 TG = FNMS(TE, TF, TC);
|
Chris@10
|
101 T1G = FNMS(TE, TB, T1n);
|
Chris@10
|
102 T1o = FMA(TE, TB, T1n);
|
Chris@10
|
103 T2p = FMA(T1X, TF, T1W);
|
Chris@10
|
104 T1Y = FNMS(T1X, TF, T1W);
|
Chris@10
|
105 T2u = FNMS(T1X, TB, T2b);
|
Chris@10
|
106 T2c = FMA(T1X, TB, T2b);
|
Chris@10
|
107 TH = T1 - T2;
|
Chris@10
|
108 T3 = T1 + T2;
|
Chris@10
|
109 T1r = Im[WS(rs, 7)];
|
Chris@10
|
110 }
|
Chris@10
|
111 T4 = Rp[WS(rs, 4)];
|
Chris@10
|
112 T5 = Rm[WS(rs, 3)];
|
Chris@10
|
113 TI = Ip[WS(rs, 4)];
|
Chris@10
|
114 T32 = T1q - T1r;
|
Chris@10
|
115 T1s = T1q + T1r;
|
Chris@10
|
116 T1p = T4 - T5;
|
Chris@10
|
117 T6 = T4 + T5;
|
Chris@10
|
118 TJ = Im[WS(rs, 3)];
|
Chris@10
|
119 }
|
Chris@10
|
120 {
|
Chris@10
|
121 E TN, TO, T8, T9, Tb, Tc;
|
Chris@10
|
122 T8 = Rp[WS(rs, 2)];
|
Chris@10
|
123 T9 = Rm[WS(rs, 5)];
|
Chris@10
|
124 TN = Ip[WS(rs, 2)];
|
Chris@10
|
125 T33 = TI - TJ;
|
Chris@10
|
126 TK = TI + TJ;
|
Chris@10
|
127 TM = T8 - T9;
|
Chris@10
|
128 Ta = T8 + T9;
|
Chris@10
|
129 TO = Im[WS(rs, 5)];
|
Chris@10
|
130 Tb = Rm[WS(rs, 1)];
|
Chris@10
|
131 Tc = Rp[WS(rs, 6)];
|
Chris@10
|
132 TS = Ip[WS(rs, 6)];
|
Chris@10
|
133 T2J = TN - TO;
|
Chris@10
|
134 TP = TN + TO;
|
Chris@10
|
135 TR = Tb - Tc;
|
Chris@10
|
136 Td = Tb + Tc;
|
Chris@10
|
137 TT = Im[WS(rs, 1)];
|
Chris@10
|
138 }
|
Chris@10
|
139 {
|
Chris@10
|
140 E T2I, TU, Te, T31, T34, T3D;
|
Chris@10
|
141 T1Z = TH + TK;
|
Chris@10
|
142 TL = TH - TK;
|
Chris@10
|
143 T1t = T1p + T1s;
|
Chris@10
|
144 T2d = T1s - T1p;
|
Chris@10
|
145 T2I = TS - TT;
|
Chris@10
|
146 TU = TS + TT;
|
Chris@10
|
147 Te = Ta + Td;
|
Chris@10
|
148 T31 = Ta - Td;
|
Chris@10
|
149 T34 = T32 - T33;
|
Chris@10
|
150 T3D = T32 + T33;
|
Chris@10
|
151 {
|
Chris@10
|
152 E T1u, T1v, T3E, T7;
|
Chris@10
|
153 T3E = T2J + T2I;
|
Chris@10
|
154 T2K = T2I - T2J;
|
Chris@10
|
155 TQ = TM - TP;
|
Chris@10
|
156 T1u = TM + TP;
|
Chris@10
|
157 T3n = T34 - T31;
|
Chris@10
|
158 T35 = T31 + T34;
|
Chris@10
|
159 TV = TR - TU;
|
Chris@10
|
160 T1v = TR + TU;
|
Chris@10
|
161 T3R = T3D - T3E;
|
Chris@10
|
162 T3F = T3D + T3E;
|
Chris@10
|
163 T2H = T3 - T6;
|
Chris@10
|
164 T7 = T3 + T6;
|
Chris@10
|
165 T1w = T1u - T1v;
|
Chris@10
|
166 T20 = T1u + T1v;
|
Chris@10
|
167 T3M = T7 - Te;
|
Chris@10
|
168 Tf = T7 + Te;
|
Chris@10
|
169 }
|
Chris@10
|
170 }
|
Chris@10
|
171 }
|
Chris@10
|
172 {
|
Chris@10
|
173 E T1e, Ti, T2N, T1c, T19, Tl, T2O, T1h, Tq, T13, Tp, T2S, T11, Tr, T14;
|
Chris@10
|
174 E T15;
|
Chris@10
|
175 {
|
Chris@10
|
176 E Tj, Tk, T1f, T1g;
|
Chris@10
|
177 {
|
Chris@10
|
178 E Tg, Th, T1a, T1b;
|
Chris@10
|
179 Tg = Rp[WS(rs, 1)];
|
Chris@10
|
180 T3h = T2H - T2K;
|
Chris@10
|
181 T2L = T2H + T2K;
|
Chris@10
|
182 T2e = TQ - TV;
|
Chris@10
|
183 TW = TQ + TV;
|
Chris@10
|
184 Th = Rm[WS(rs, 6)];
|
Chris@10
|
185 T1a = Ip[WS(rs, 1)];
|
Chris@10
|
186 T1b = Im[WS(rs, 6)];
|
Chris@10
|
187 Tj = Rp[WS(rs, 5)];
|
Chris@10
|
188 T1e = Tg - Th;
|
Chris@10
|
189 Ti = Tg + Th;
|
Chris@10
|
190 T2N = T1a - T1b;
|
Chris@10
|
191 T1c = T1a + T1b;
|
Chris@10
|
192 Tk = Rm[WS(rs, 2)];
|
Chris@10
|
193 T1f = Ip[WS(rs, 5)];
|
Chris@10
|
194 T1g = Im[WS(rs, 2)];
|
Chris@10
|
195 }
|
Chris@10
|
196 {
|
Chris@10
|
197 E Tn, To, TZ, T10;
|
Chris@10
|
198 Tn = Rm[0];
|
Chris@10
|
199 T19 = Tj - Tk;
|
Chris@10
|
200 Tl = Tj + Tk;
|
Chris@10
|
201 T2O = T1f - T1g;
|
Chris@10
|
202 T1h = T1f + T1g;
|
Chris@10
|
203 To = Rp[WS(rs, 7)];
|
Chris@10
|
204 TZ = Ip[WS(rs, 7)];
|
Chris@10
|
205 T10 = Im[0];
|
Chris@10
|
206 Tq = Rp[WS(rs, 3)];
|
Chris@10
|
207 T13 = Tn - To;
|
Chris@10
|
208 Tp = Tn + To;
|
Chris@10
|
209 T2S = TZ - T10;
|
Chris@10
|
210 T11 = TZ + T10;
|
Chris@10
|
211 Tr = Rm[WS(rs, 4)];
|
Chris@10
|
212 T14 = Ip[WS(rs, 3)];
|
Chris@10
|
213 T15 = Im[WS(rs, 4)];
|
Chris@10
|
214 }
|
Chris@10
|
215 }
|
Chris@10
|
216 {
|
Chris@10
|
217 E TY, T16, Tm, Tt;
|
Chris@10
|
218 {
|
Chris@10
|
219 E T2P, T3G, Ts, T2M, T3H, T2U, T2T, T2R;
|
Chris@10
|
220 T2P = T2N - T2O;
|
Chris@10
|
221 T3G = T2N + T2O;
|
Chris@10
|
222 TY = Tq - Tr;
|
Chris@10
|
223 Ts = Tq + Tr;
|
Chris@10
|
224 T2T = T14 - T15;
|
Chris@10
|
225 T16 = T14 + T15;
|
Chris@10
|
226 T2M = Ti - Tl;
|
Chris@10
|
227 Tm = Ti + Tl;
|
Chris@10
|
228 T3H = T2S + T2T;
|
Chris@10
|
229 T2U = T2S - T2T;
|
Chris@10
|
230 Tt = Tp + Ts;
|
Chris@10
|
231 T2R = Tp - Ts;
|
Chris@10
|
232 T2Q = T2M - T2P;
|
Chris@10
|
233 T36 = T2M + T2P;
|
Chris@10
|
234 T3I = T3G + T3H;
|
Chris@10
|
235 T3N = T3H - T3G;
|
Chris@10
|
236 T2V = T2R + T2U;
|
Chris@10
|
237 T37 = T2U - T2R;
|
Chris@10
|
238 }
|
Chris@10
|
239 {
|
Chris@10
|
240 E T25, T26, T22, T23, T12, T17;
|
Chris@10
|
241 T12 = TY - T11;
|
Chris@10
|
242 T25 = TY + T11;
|
Chris@10
|
243 T26 = T13 + T16;
|
Chris@10
|
244 T17 = T13 - T16;
|
Chris@10
|
245 T22 = T1c - T19;
|
Chris@10
|
246 T1d = T19 + T1c;
|
Chris@10
|
247 Tu = Tm + Tt;
|
Chris@10
|
248 T3S = Tm - Tt;
|
Chris@10
|
249 T18 = FNMS(KP414213562, T17, T12);
|
Chris@10
|
250 T1z = FMA(KP414213562, T12, T17);
|
Chris@10
|
251 T1i = T1e - T1h;
|
Chris@10
|
252 T23 = T1e + T1h;
|
Chris@10
|
253 T24 = FNMS(KP414213562, T23, T22);
|
Chris@10
|
254 T2g = FMA(KP414213562, T22, T23);
|
Chris@10
|
255 T27 = FNMS(KP414213562, T26, T25);
|
Chris@10
|
256 T2h = FMA(KP414213562, T25, T26);
|
Chris@10
|
257 }
|
Chris@10
|
258 }
|
Chris@10
|
259 }
|
Chris@10
|
260 }
|
Chris@10
|
261 {
|
Chris@10
|
262 E T1j, T1y, T3V, T3X, T3W, T38, T3i, T3o, T2W, T3K, T3B, T3A;
|
Chris@10
|
263 Rp[0] = Tf + Tu;
|
Chris@10
|
264 T3A = Tf - Tu;
|
Chris@10
|
265 T1j = FMA(KP414213562, T1i, T1d);
|
Chris@10
|
266 T1y = FNMS(KP414213562, T1d, T1i);
|
Chris@10
|
267 T3K = T3C * T3A;
|
Chris@10
|
268 T3B = T3z * T3A;
|
Chris@10
|
269 {
|
Chris@10
|
270 E T3O, T3T, T3J, T3P, T3U;
|
Chris@10
|
271 T3O = T3M - T3N;
|
Chris@10
|
272 T3V = T3M + T3N;
|
Chris@10
|
273 T3X = T3S + T3R;
|
Chris@10
|
274 T3T = T3R - T3S;
|
Chris@10
|
275 Rm[0] = T3F + T3I;
|
Chris@10
|
276 T3J = T3F - T3I;
|
Chris@10
|
277 T3P = T3L * T3O;
|
Chris@10
|
278 T3U = T3L * T3T;
|
Chris@10
|
279 T3W = TA * T3V;
|
Chris@10
|
280 Rp[WS(rs, 4)] = FNMS(T3C, T3J, T3B);
|
Chris@10
|
281 Rm[WS(rs, 4)] = FMA(T3z, T3J, T3K);
|
Chris@10
|
282 Rp[WS(rs, 6)] = FNMS(T3Q, T3T, T3P);
|
Chris@10
|
283 Rm[WS(rs, 6)] = FMA(T3Q, T3O, T3U);
|
Chris@10
|
284 T38 = T36 + T37;
|
Chris@10
|
285 T3i = T37 - T36;
|
Chris@10
|
286 T3o = T2Q - T2V;
|
Chris@10
|
287 T2W = T2Q + T2V;
|
Chris@10
|
288 }
|
Chris@10
|
289 {
|
Chris@10
|
290 E T2q, T21, T28, T2w, T2v, T2f, T2i, T2r;
|
Chris@10
|
291 {
|
Chris@10
|
292 E T2Y, T3a, T3c, T3d, T39, T3e, T3b, T2X, T3Y;
|
Chris@10
|
293 Rp[WS(rs, 2)] = FNMS(TE, T3X, T3W);
|
Chris@10
|
294 T3Y = TA * T3X;
|
Chris@10
|
295 {
|
Chris@10
|
296 E T3t, T3j, T3x, T3p;
|
Chris@10
|
297 T3t = FMA(KP707106781, T3i, T3h);
|
Chris@10
|
298 T3j = FNMS(KP707106781, T3i, T3h);
|
Chris@10
|
299 T3x = FMA(KP707106781, T3o, T3n);
|
Chris@10
|
300 T3p = FNMS(KP707106781, T3o, T3n);
|
Chris@10
|
301 Rm[WS(rs, 2)] = FMA(TE, T3V, T3Y);
|
Chris@10
|
302 {
|
Chris@10
|
303 E T3u, T3k, T3y, T3q;
|
Chris@10
|
304 T3u = T3s * T3t;
|
Chris@10
|
305 T3k = T3g * T3j;
|
Chris@10
|
306 T3y = T3s * T3x;
|
Chris@10
|
307 T3q = T3g * T3p;
|
Chris@10
|
308 Rp[WS(rs, 3)] = FNMS(T3w, T3x, T3u);
|
Chris@10
|
309 Rp[WS(rs, 7)] = FNMS(T3m, T3p, T3k);
|
Chris@10
|
310 Rm[WS(rs, 3)] = FMA(T3w, T3t, T3y);
|
Chris@10
|
311 Rm[WS(rs, 7)] = FMA(T3m, T3j, T3q);
|
Chris@10
|
312 T3b = FMA(KP707106781, T2W, T2L);
|
Chris@10
|
313 T2X = FNMS(KP707106781, T2W, T2L);
|
Chris@10
|
314 }
|
Chris@10
|
315 }
|
Chris@10
|
316 T2Y = T2G * T2X;
|
Chris@10
|
317 T3a = T30 * T2X;
|
Chris@10
|
318 T3c = T1V * T3b;
|
Chris@10
|
319 T3d = FMA(KP707106781, T38, T35);
|
Chris@10
|
320 T39 = FNMS(KP707106781, T38, T35);
|
Chris@10
|
321 T3e = T1X * T3b;
|
Chris@10
|
322 T2q = FMA(KP707106781, T20, T1Z);
|
Chris@10
|
323 T21 = FNMS(KP707106781, T20, T1Z);
|
Chris@10
|
324 Rp[WS(rs, 1)] = FNMS(T1X, T3d, T3c);
|
Chris@10
|
325 Rm[WS(rs, 5)] = FMA(T2G, T39, T3a);
|
Chris@10
|
326 Rp[WS(rs, 5)] = FNMS(T30, T39, T2Y);
|
Chris@10
|
327 Rm[WS(rs, 1)] = FMA(T1V, T3d, T3e);
|
Chris@10
|
328 T28 = T24 + T27;
|
Chris@10
|
329 T2w = T27 - T24;
|
Chris@10
|
330 T2v = FNMS(KP707106781, T2e, T2d);
|
Chris@10
|
331 T2f = FMA(KP707106781, T2e, T2d);
|
Chris@10
|
332 T2i = T2g - T2h;
|
Chris@10
|
333 T2r = T2g + T2h;
|
Chris@10
|
334 }
|
Chris@10
|
335 {
|
Chris@10
|
336 E TX, T1k, T1x, T1A;
|
Chris@10
|
337 T1J = FMA(KP707106781, TW, TL);
|
Chris@10
|
338 TX = FNMS(KP707106781, TW, TL);
|
Chris@10
|
339 {
|
Chris@10
|
340 E T2l, T29, T2n, T2j;
|
Chris@10
|
341 T2l = FNMS(KP923879532, T28, T21);
|
Chris@10
|
342 T29 = FMA(KP923879532, T28, T21);
|
Chris@10
|
343 T2n = FMA(KP923879532, T2i, T2f);
|
Chris@10
|
344 T2j = FNMS(KP923879532, T2i, T2f);
|
Chris@10
|
345 {
|
Chris@10
|
346 E T2o, T2m, T2k, T2a;
|
Chris@10
|
347 T2o = Tz * T2l;
|
Chris@10
|
348 T2m = Tw * T2l;
|
Chris@10
|
349 T2k = T2c * T29;
|
Chris@10
|
350 T2a = T1Y * T29;
|
Chris@10
|
351 Im[WS(rs, 1)] = FMA(Tw, T2n, T2o);
|
Chris@10
|
352 Ip[WS(rs, 1)] = FNMS(Tz, T2n, T2m);
|
Chris@10
|
353 Im[WS(rs, 5)] = FMA(T1Y, T2j, T2k);
|
Chris@10
|
354 Ip[WS(rs, 5)] = FNMS(T2c, T2j, T2a);
|
Chris@10
|
355 T1k = T18 - T1j;
|
Chris@10
|
356 T1O = T1j + T18;
|
Chris@10
|
357 }
|
Chris@10
|
358 }
|
Chris@10
|
359 T1N = FMA(KP707106781, T1w, T1t);
|
Chris@10
|
360 T1x = FNMS(KP707106781, T1w, T1t);
|
Chris@10
|
361 T1A = T1y - T1z;
|
Chris@10
|
362 T1K = T1y + T1z;
|
Chris@10
|
363 {
|
Chris@10
|
364 E T1E, T1l, T1H, T1B;
|
Chris@10
|
365 T1E = FMA(KP923879532, T1k, TX);
|
Chris@10
|
366 T1l = FNMS(KP923879532, T1k, TX);
|
Chris@10
|
367 T1H = FMA(KP923879532, T1A, T1x);
|
Chris@10
|
368 T1B = FNMS(KP923879532, T1A, T1x);
|
Chris@10
|
369 {
|
Chris@10
|
370 E T1I, T1F, T1C, T1m;
|
Chris@10
|
371 T1I = T1G * T1E;
|
Chris@10
|
372 T1F = T1D * T1E;
|
Chris@10
|
373 T1C = T1o * T1l;
|
Chris@10
|
374 T1m = TG * T1l;
|
Chris@10
|
375 Im[WS(rs, 2)] = FMA(T1D, T1H, T1I);
|
Chris@10
|
376 Ip[WS(rs, 2)] = FNMS(T1G, T1H, T1F);
|
Chris@10
|
377 Im[WS(rs, 6)] = FMA(TG, T1B, T1C);
|
Chris@10
|
378 Ip[WS(rs, 6)] = FNMS(T1o, T1B, T1m);
|
Chris@10
|
379 }
|
Chris@10
|
380 }
|
Chris@10
|
381 {
|
Chris@10
|
382 E T2A, T2s, T2D, T2x;
|
Chris@10
|
383 T2A = FMA(KP923879532, T2r, T2q);
|
Chris@10
|
384 T2s = FNMS(KP923879532, T2r, T2q);
|
Chris@10
|
385 T2D = FNMS(KP923879532, T2w, T2v);
|
Chris@10
|
386 T2x = FMA(KP923879532, T2w, T2v);
|
Chris@10
|
387 {
|
Chris@10
|
388 E T2B, T2t, T2E, T2y;
|
Chris@10
|
389 T2B = T2z * T2A;
|
Chris@10
|
390 T2t = T2p * T2s;
|
Chris@10
|
391 T2E = T2z * T2D;
|
Chris@10
|
392 T2y = T2p * T2x;
|
Chris@10
|
393 Ip[WS(rs, 7)] = FNMS(T2C, T2D, T2B);
|
Chris@10
|
394 Ip[WS(rs, 3)] = FNMS(T2u, T2x, T2t);
|
Chris@10
|
395 Im[WS(rs, 7)] = FMA(T2C, T2A, T2E);
|
Chris@10
|
396 Im[WS(rs, 3)] = FMA(T2u, T2s, T2y);
|
Chris@10
|
397 }
|
Chris@10
|
398 }
|
Chris@10
|
399 }
|
Chris@10
|
400 }
|
Chris@10
|
401 }
|
Chris@10
|
402 }
|
Chris@10
|
403 }
|
Chris@10
|
404 {
|
Chris@10
|
405 E T1L, T1R, T1P, T1T;
|
Chris@10
|
406 T1L = FNMS(KP923879532, T1K, T1J);
|
Chris@10
|
407 T1R = FMA(KP923879532, T1K, T1J);
|
Chris@10
|
408 T1P = FNMS(KP923879532, T1O, T1N);
|
Chris@10
|
409 T1T = FMA(KP923879532, T1O, T1N);
|
Chris@10
|
410 {
|
Chris@10
|
411 E T1S, T1M, T1U, T1Q;
|
Chris@10
|
412 T1S = Tv * T1R;
|
Chris@10
|
413 T1M = TB * T1L;
|
Chris@10
|
414 T1U = Tv * T1T;
|
Chris@10
|
415 T1Q = TB * T1P;
|
Chris@10
|
416 Ip[0] = FNMS(Ty, T1T, T1S);
|
Chris@10
|
417 Ip[WS(rs, 4)] = FNMS(TF, T1P, T1M);
|
Chris@10
|
418 Im[0] = FMA(Ty, T1R, T1U);
|
Chris@10
|
419 Im[WS(rs, 4)] = FMA(TF, T1L, T1Q);
|
Chris@10
|
420 }
|
Chris@10
|
421 }
|
Chris@10
|
422 }
|
Chris@10
|
423 }
|
Chris@10
|
424 }
|
Chris@10
|
425
|
Chris@10
|
426 static const tw_instr twinstr[] = {
|
Chris@10
|
427 {TW_CEXP, 1, 1},
|
Chris@10
|
428 {TW_CEXP, 1, 3},
|
Chris@10
|
429 {TW_CEXP, 1, 9},
|
Chris@10
|
430 {TW_CEXP, 1, 15},
|
Chris@10
|
431 {TW_NEXT, 1, 0}
|
Chris@10
|
432 };
|
Chris@10
|
433
|
Chris@10
|
434 static const hc2c_desc desc = { 16, "hc2cb2_16", twinstr, &GENUS, {104, 42, 92, 0} };
|
Chris@10
|
435
|
Chris@10
|
436 void X(codelet_hc2cb2_16) (planner *p) {
|
Chris@10
|
437 X(khc2c_register) (p, hc2cb2_16, &desc, HC2C_VIA_RDFT);
|
Chris@10
|
438 }
|
Chris@10
|
439 #else /* HAVE_FMA */
|
Chris@10
|
440
|
Chris@10
|
441 /* Generated by: ../../../genfft/gen_hc2c.native -compact -variables 4 -pipeline-latency 4 -sign 1 -twiddle-log3 -precompute-twiddles -n 16 -dif -name hc2cb2_16 -include hc2cb.h */
|
Chris@10
|
442
|
Chris@10
|
443 /*
|
Chris@10
|
444 * This function contains 196 FP additions, 108 FP multiplications,
|
Chris@10
|
445 * (or, 156 additions, 68 multiplications, 40 fused multiply/add),
|
Chris@10
|
446 * 80 stack variables, 3 constants, and 64 memory accesses
|
Chris@10
|
447 */
|
Chris@10
|
448 #include "hc2cb.h"
|
Chris@10
|
449
|
Chris@10
|
450 static void hc2cb2_16(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms)
|
Chris@10
|
451 {
|
Chris@10
|
452 DK(KP382683432, +0.382683432365089771728459984030398866761344562);
|
Chris@10
|
453 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
Chris@10
|
454 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
Chris@10
|
455 {
|
Chris@10
|
456 INT m;
|
Chris@10
|
457 for (m = mb, W = W + ((mb - 1) * 8); m < me; m = m + 1, Rp = Rp + ms, Ip = Ip + ms, Rm = Rm - ms, Im = Im - ms, W = W + 8, MAKE_VOLATILE_STRIDE(64, rs)) {
|
Chris@10
|
458 E Tv, Ty, T1l, T1n, T1p, T1t, T27, T25, Tz, Tw, TB, T21, T1P, T1H, T1X;
|
Chris@10
|
459 E T17, T1L, T1N, T1v, T1w, T1x, T1B, T2F, T2T, T2b, T2R, T3j, T3x, T35, T3t;
|
Chris@10
|
460 {
|
Chris@10
|
461 E TA, T1J, T15, T1G, Tx, T1K, T16, T1F;
|
Chris@10
|
462 {
|
Chris@10
|
463 E T1m, T1s, T1o, T1r;
|
Chris@10
|
464 Tv = W[0];
|
Chris@10
|
465 Ty = W[1];
|
Chris@10
|
466 T1l = W[2];
|
Chris@10
|
467 T1n = W[3];
|
Chris@10
|
468 T1m = Tv * T1l;
|
Chris@10
|
469 T1s = Ty * T1l;
|
Chris@10
|
470 T1o = Ty * T1n;
|
Chris@10
|
471 T1r = Tv * T1n;
|
Chris@10
|
472 T1p = T1m + T1o;
|
Chris@10
|
473 T1t = T1r - T1s;
|
Chris@10
|
474 T27 = T1r + T1s;
|
Chris@10
|
475 T25 = T1m - T1o;
|
Chris@10
|
476 Tz = W[5];
|
Chris@10
|
477 TA = Ty * Tz;
|
Chris@10
|
478 T1J = T1l * Tz;
|
Chris@10
|
479 T15 = Tv * Tz;
|
Chris@10
|
480 T1G = T1n * Tz;
|
Chris@10
|
481 Tw = W[4];
|
Chris@10
|
482 Tx = Tv * Tw;
|
Chris@10
|
483 T1K = T1n * Tw;
|
Chris@10
|
484 T16 = Ty * Tw;
|
Chris@10
|
485 T1F = T1l * Tw;
|
Chris@10
|
486 }
|
Chris@10
|
487 TB = Tx - TA;
|
Chris@10
|
488 T21 = T1J + T1K;
|
Chris@10
|
489 T1P = T15 - T16;
|
Chris@10
|
490 T1H = T1F + T1G;
|
Chris@10
|
491 T1X = T1F - T1G;
|
Chris@10
|
492 T17 = T15 + T16;
|
Chris@10
|
493 T1L = T1J - T1K;
|
Chris@10
|
494 T1N = Tx + TA;
|
Chris@10
|
495 T1v = W[6];
|
Chris@10
|
496 T1w = W[7];
|
Chris@10
|
497 T1x = FMA(Tv, T1v, Ty * T1w);
|
Chris@10
|
498 T1B = FNMS(Ty, T1v, Tv * T1w);
|
Chris@10
|
499 {
|
Chris@10
|
500 E T2D, T2E, T29, T2a;
|
Chris@10
|
501 T2D = T25 * Tz;
|
Chris@10
|
502 T2E = T27 * Tw;
|
Chris@10
|
503 T2F = T2D + T2E;
|
Chris@10
|
504 T2T = T2D - T2E;
|
Chris@10
|
505 T29 = T25 * Tw;
|
Chris@10
|
506 T2a = T27 * Tz;
|
Chris@10
|
507 T2b = T29 - T2a;
|
Chris@10
|
508 T2R = T29 + T2a;
|
Chris@10
|
509 }
|
Chris@10
|
510 {
|
Chris@10
|
511 E T3h, T3i, T33, T34;
|
Chris@10
|
512 T3h = T1p * Tz;
|
Chris@10
|
513 T3i = T1t * Tw;
|
Chris@10
|
514 T3j = T3h + T3i;
|
Chris@10
|
515 T3x = T3h - T3i;
|
Chris@10
|
516 T33 = T1p * Tw;
|
Chris@10
|
517 T34 = T1t * Tz;
|
Chris@10
|
518 T35 = T33 - T34;
|
Chris@10
|
519 T3t = T33 + T34;
|
Chris@10
|
520 }
|
Chris@10
|
521 }
|
Chris@10
|
522 {
|
Chris@10
|
523 E T7, T36, T3k, TC, T1f, T2e, T2I, T1Q, Te, TJ, T1R, T18, T2L, T37, T2l;
|
Chris@10
|
524 E T3l, Tm, T1T, TT, T1h, T2A, T2N, T3b, T3n, Tt, T1U, T12, T1i, T2t, T2O;
|
Chris@10
|
525 E T3e, T3o;
|
Chris@10
|
526 {
|
Chris@10
|
527 E T3, T2c, T1b, T2H, T6, T2G, T1e, T2d;
|
Chris@10
|
528 {
|
Chris@10
|
529 E T1, T2, T19, T1a;
|
Chris@10
|
530 T1 = Rp[0];
|
Chris@10
|
531 T2 = Rm[WS(rs, 7)];
|
Chris@10
|
532 T3 = T1 + T2;
|
Chris@10
|
533 T2c = T1 - T2;
|
Chris@10
|
534 T19 = Ip[0];
|
Chris@10
|
535 T1a = Im[WS(rs, 7)];
|
Chris@10
|
536 T1b = T19 - T1a;
|
Chris@10
|
537 T2H = T19 + T1a;
|
Chris@10
|
538 }
|
Chris@10
|
539 {
|
Chris@10
|
540 E T4, T5, T1c, T1d;
|
Chris@10
|
541 T4 = Rp[WS(rs, 4)];
|
Chris@10
|
542 T5 = Rm[WS(rs, 3)];
|
Chris@10
|
543 T6 = T4 + T5;
|
Chris@10
|
544 T2G = T4 - T5;
|
Chris@10
|
545 T1c = Ip[WS(rs, 4)];
|
Chris@10
|
546 T1d = Im[WS(rs, 3)];
|
Chris@10
|
547 T1e = T1c - T1d;
|
Chris@10
|
548 T2d = T1c + T1d;
|
Chris@10
|
549 }
|
Chris@10
|
550 T7 = T3 + T6;
|
Chris@10
|
551 T36 = T2c + T2d;
|
Chris@10
|
552 T3k = T2H - T2G;
|
Chris@10
|
553 TC = T3 - T6;
|
Chris@10
|
554 T1f = T1b - T1e;
|
Chris@10
|
555 T2e = T2c - T2d;
|
Chris@10
|
556 T2I = T2G + T2H;
|
Chris@10
|
557 T1Q = T1b + T1e;
|
Chris@10
|
558 }
|
Chris@10
|
559 {
|
Chris@10
|
560 E Ta, T2f, TI, T2g, Td, T2i, TF, T2j;
|
Chris@10
|
561 {
|
Chris@10
|
562 E T8, T9, TG, TH;
|
Chris@10
|
563 T8 = Rp[WS(rs, 2)];
|
Chris@10
|
564 T9 = Rm[WS(rs, 5)];
|
Chris@10
|
565 Ta = T8 + T9;
|
Chris@10
|
566 T2f = T8 - T9;
|
Chris@10
|
567 TG = Ip[WS(rs, 2)];
|
Chris@10
|
568 TH = Im[WS(rs, 5)];
|
Chris@10
|
569 TI = TG - TH;
|
Chris@10
|
570 T2g = TG + TH;
|
Chris@10
|
571 }
|
Chris@10
|
572 {
|
Chris@10
|
573 E Tb, Tc, TD, TE;
|
Chris@10
|
574 Tb = Rm[WS(rs, 1)];
|
Chris@10
|
575 Tc = Rp[WS(rs, 6)];
|
Chris@10
|
576 Td = Tb + Tc;
|
Chris@10
|
577 T2i = Tb - Tc;
|
Chris@10
|
578 TD = Ip[WS(rs, 6)];
|
Chris@10
|
579 TE = Im[WS(rs, 1)];
|
Chris@10
|
580 TF = TD - TE;
|
Chris@10
|
581 T2j = TD + TE;
|
Chris@10
|
582 }
|
Chris@10
|
583 Te = Ta + Td;
|
Chris@10
|
584 TJ = TF - TI;
|
Chris@10
|
585 T1R = TI + TF;
|
Chris@10
|
586 T18 = Ta - Td;
|
Chris@10
|
587 {
|
Chris@10
|
588 E T2J, T2K, T2h, T2k;
|
Chris@10
|
589 T2J = T2f + T2g;
|
Chris@10
|
590 T2K = T2i + T2j;
|
Chris@10
|
591 T2L = KP707106781 * (T2J - T2K);
|
Chris@10
|
592 T37 = KP707106781 * (T2J + T2K);
|
Chris@10
|
593 T2h = T2f - T2g;
|
Chris@10
|
594 T2k = T2i - T2j;
|
Chris@10
|
595 T2l = KP707106781 * (T2h + T2k);
|
Chris@10
|
596 T3l = KP707106781 * (T2h - T2k);
|
Chris@10
|
597 }
|
Chris@10
|
598 }
|
Chris@10
|
599 {
|
Chris@10
|
600 E Ti, T2x, TO, T2v, Tl, T2u, TR, T2y, TL, TS;
|
Chris@10
|
601 {
|
Chris@10
|
602 E Tg, Th, TM, TN;
|
Chris@10
|
603 Tg = Rp[WS(rs, 1)];
|
Chris@10
|
604 Th = Rm[WS(rs, 6)];
|
Chris@10
|
605 Ti = Tg + Th;
|
Chris@10
|
606 T2x = Tg - Th;
|
Chris@10
|
607 TM = Ip[WS(rs, 1)];
|
Chris@10
|
608 TN = Im[WS(rs, 6)];
|
Chris@10
|
609 TO = TM - TN;
|
Chris@10
|
610 T2v = TM + TN;
|
Chris@10
|
611 }
|
Chris@10
|
612 {
|
Chris@10
|
613 E Tj, Tk, TP, TQ;
|
Chris@10
|
614 Tj = Rp[WS(rs, 5)];
|
Chris@10
|
615 Tk = Rm[WS(rs, 2)];
|
Chris@10
|
616 Tl = Tj + Tk;
|
Chris@10
|
617 T2u = Tj - Tk;
|
Chris@10
|
618 TP = Ip[WS(rs, 5)];
|
Chris@10
|
619 TQ = Im[WS(rs, 2)];
|
Chris@10
|
620 TR = TP - TQ;
|
Chris@10
|
621 T2y = TP + TQ;
|
Chris@10
|
622 }
|
Chris@10
|
623 Tm = Ti + Tl;
|
Chris@10
|
624 T1T = TO + TR;
|
Chris@10
|
625 TL = Ti - Tl;
|
Chris@10
|
626 TS = TO - TR;
|
Chris@10
|
627 TT = TL - TS;
|
Chris@10
|
628 T1h = TL + TS;
|
Chris@10
|
629 {
|
Chris@10
|
630 E T2w, T2z, T39, T3a;
|
Chris@10
|
631 T2w = T2u + T2v;
|
Chris@10
|
632 T2z = T2x - T2y;
|
Chris@10
|
633 T2A = FMA(KP923879532, T2w, KP382683432 * T2z);
|
Chris@10
|
634 T2N = FNMS(KP382683432, T2w, KP923879532 * T2z);
|
Chris@10
|
635 T39 = T2x + T2y;
|
Chris@10
|
636 T3a = T2v - T2u;
|
Chris@10
|
637 T3b = FNMS(KP923879532, T3a, KP382683432 * T39);
|
Chris@10
|
638 T3n = FMA(KP382683432, T3a, KP923879532 * T39);
|
Chris@10
|
639 }
|
Chris@10
|
640 }
|
Chris@10
|
641 {
|
Chris@10
|
642 E Tp, T2q, TX, T2o, Ts, T2n, T10, T2r, TU, T11;
|
Chris@10
|
643 {
|
Chris@10
|
644 E Tn, To, TV, TW;
|
Chris@10
|
645 Tn = Rm[0];
|
Chris@10
|
646 To = Rp[WS(rs, 7)];
|
Chris@10
|
647 Tp = Tn + To;
|
Chris@10
|
648 T2q = Tn - To;
|
Chris@10
|
649 TV = Ip[WS(rs, 7)];
|
Chris@10
|
650 TW = Im[0];
|
Chris@10
|
651 TX = TV - TW;
|
Chris@10
|
652 T2o = TV + TW;
|
Chris@10
|
653 }
|
Chris@10
|
654 {
|
Chris@10
|
655 E Tq, Tr, TY, TZ;
|
Chris@10
|
656 Tq = Rp[WS(rs, 3)];
|
Chris@10
|
657 Tr = Rm[WS(rs, 4)];
|
Chris@10
|
658 Ts = Tq + Tr;
|
Chris@10
|
659 T2n = Tq - Tr;
|
Chris@10
|
660 TY = Ip[WS(rs, 3)];
|
Chris@10
|
661 TZ = Im[WS(rs, 4)];
|
Chris@10
|
662 T10 = TY - TZ;
|
Chris@10
|
663 T2r = TY + TZ;
|
Chris@10
|
664 }
|
Chris@10
|
665 Tt = Tp + Ts;
|
Chris@10
|
666 T1U = TX + T10;
|
Chris@10
|
667 TU = Tp - Ts;
|
Chris@10
|
668 T11 = TX - T10;
|
Chris@10
|
669 T12 = TU + T11;
|
Chris@10
|
670 T1i = T11 - TU;
|
Chris@10
|
671 {
|
Chris@10
|
672 E T2p, T2s, T3c, T3d;
|
Chris@10
|
673 T2p = T2n - T2o;
|
Chris@10
|
674 T2s = T2q - T2r;
|
Chris@10
|
675 T2t = FNMS(KP382683432, T2s, KP923879532 * T2p);
|
Chris@10
|
676 T2O = FMA(KP382683432, T2p, KP923879532 * T2s);
|
Chris@10
|
677 T3c = T2q + T2r;
|
Chris@10
|
678 T3d = T2n + T2o;
|
Chris@10
|
679 T3e = FNMS(KP923879532, T3d, KP382683432 * T3c);
|
Chris@10
|
680 T3o = FMA(KP382683432, T3d, KP923879532 * T3c);
|
Chris@10
|
681 }
|
Chris@10
|
682 }
|
Chris@10
|
683 {
|
Chris@10
|
684 E Tf, Tu, T1O, T1S, T1V, T1W;
|
Chris@10
|
685 Tf = T7 + Te;
|
Chris@10
|
686 Tu = Tm + Tt;
|
Chris@10
|
687 T1O = Tf - Tu;
|
Chris@10
|
688 T1S = T1Q + T1R;
|
Chris@10
|
689 T1V = T1T + T1U;
|
Chris@10
|
690 T1W = T1S - T1V;
|
Chris@10
|
691 Rp[0] = Tf + Tu;
|
Chris@10
|
692 Rm[0] = T1S + T1V;
|
Chris@10
|
693 Rp[WS(rs, 4)] = FNMS(T1P, T1W, T1N * T1O);
|
Chris@10
|
694 Rm[WS(rs, 4)] = FMA(T1P, T1O, T1N * T1W);
|
Chris@10
|
695 }
|
Chris@10
|
696 {
|
Chris@10
|
697 E T3g, T3r, T3q, T3s;
|
Chris@10
|
698 {
|
Chris@10
|
699 E T38, T3f, T3m, T3p;
|
Chris@10
|
700 T38 = T36 - T37;
|
Chris@10
|
701 T3f = T3b + T3e;
|
Chris@10
|
702 T3g = T38 - T3f;
|
Chris@10
|
703 T3r = T38 + T3f;
|
Chris@10
|
704 T3m = T3k + T3l;
|
Chris@10
|
705 T3p = T3n - T3o;
|
Chris@10
|
706 T3q = T3m - T3p;
|
Chris@10
|
707 T3s = T3m + T3p;
|
Chris@10
|
708 }
|
Chris@10
|
709 Ip[WS(rs, 5)] = FNMS(T3j, T3q, T35 * T3g);
|
Chris@10
|
710 Im[WS(rs, 5)] = FMA(T3j, T3g, T35 * T3q);
|
Chris@10
|
711 Ip[WS(rs, 1)] = FNMS(T1n, T3s, T1l * T3r);
|
Chris@10
|
712 Im[WS(rs, 1)] = FMA(T1n, T3r, T1l * T3s);
|
Chris@10
|
713 }
|
Chris@10
|
714 {
|
Chris@10
|
715 E T3w, T3B, T3A, T3C;
|
Chris@10
|
716 {
|
Chris@10
|
717 E T3u, T3v, T3y, T3z;
|
Chris@10
|
718 T3u = T36 + T37;
|
Chris@10
|
719 T3v = T3n + T3o;
|
Chris@10
|
720 T3w = T3u - T3v;
|
Chris@10
|
721 T3B = T3u + T3v;
|
Chris@10
|
722 T3y = T3k - T3l;
|
Chris@10
|
723 T3z = T3b - T3e;
|
Chris@10
|
724 T3A = T3y + T3z;
|
Chris@10
|
725 T3C = T3y - T3z;
|
Chris@10
|
726 }
|
Chris@10
|
727 Ip[WS(rs, 3)] = FNMS(T3x, T3A, T3t * T3w);
|
Chris@10
|
728 Im[WS(rs, 3)] = FMA(T3t, T3A, T3x * T3w);
|
Chris@10
|
729 Ip[WS(rs, 7)] = FNMS(T1w, T3C, T1v * T3B);
|
Chris@10
|
730 Im[WS(rs, 7)] = FMA(T1v, T3C, T1w * T3B);
|
Chris@10
|
731 }
|
Chris@10
|
732 {
|
Chris@10
|
733 E T14, T1q, T1k, T1u;
|
Chris@10
|
734 {
|
Chris@10
|
735 E TK, T13, T1g, T1j;
|
Chris@10
|
736 TK = TC + TJ;
|
Chris@10
|
737 T13 = KP707106781 * (TT + T12);
|
Chris@10
|
738 T14 = TK - T13;
|
Chris@10
|
739 T1q = TK + T13;
|
Chris@10
|
740 T1g = T18 + T1f;
|
Chris@10
|
741 T1j = KP707106781 * (T1h + T1i);
|
Chris@10
|
742 T1k = T1g - T1j;
|
Chris@10
|
743 T1u = T1g + T1j;
|
Chris@10
|
744 }
|
Chris@10
|
745 Rp[WS(rs, 5)] = FNMS(T17, T1k, TB * T14);
|
Chris@10
|
746 Rm[WS(rs, 5)] = FMA(T17, T14, TB * T1k);
|
Chris@10
|
747 Rp[WS(rs, 1)] = FNMS(T1t, T1u, T1p * T1q);
|
Chris@10
|
748 Rm[WS(rs, 1)] = FMA(T1t, T1q, T1p * T1u);
|
Chris@10
|
749 }
|
Chris@10
|
750 {
|
Chris@10
|
751 E T1A, T1I, T1E, T1M;
|
Chris@10
|
752 {
|
Chris@10
|
753 E T1y, T1z, T1C, T1D;
|
Chris@10
|
754 T1y = TC - TJ;
|
Chris@10
|
755 T1z = KP707106781 * (T1i - T1h);
|
Chris@10
|
756 T1A = T1y - T1z;
|
Chris@10
|
757 T1I = T1y + T1z;
|
Chris@10
|
758 T1C = T1f - T18;
|
Chris@10
|
759 T1D = KP707106781 * (TT - T12);
|
Chris@10
|
760 T1E = T1C - T1D;
|
Chris@10
|
761 T1M = T1C + T1D;
|
Chris@10
|
762 }
|
Chris@10
|
763 Rp[WS(rs, 7)] = FNMS(T1B, T1E, T1x * T1A);
|
Chris@10
|
764 Rm[WS(rs, 7)] = FMA(T1x, T1E, T1B * T1A);
|
Chris@10
|
765 Rp[WS(rs, 3)] = FNMS(T1L, T1M, T1H * T1I);
|
Chris@10
|
766 Rm[WS(rs, 3)] = FMA(T1H, T1M, T1L * T1I);
|
Chris@10
|
767 }
|
Chris@10
|
768 {
|
Chris@10
|
769 E T2C, T2S, T2Q, T2U;
|
Chris@10
|
770 {
|
Chris@10
|
771 E T2m, T2B, T2M, T2P;
|
Chris@10
|
772 T2m = T2e - T2l;
|
Chris@10
|
773 T2B = T2t - T2A;
|
Chris@10
|
774 T2C = T2m - T2B;
|
Chris@10
|
775 T2S = T2m + T2B;
|
Chris@10
|
776 T2M = T2I - T2L;
|
Chris@10
|
777 T2P = T2N - T2O;
|
Chris@10
|
778 T2Q = T2M - T2P;
|
Chris@10
|
779 T2U = T2M + T2P;
|
Chris@10
|
780 }
|
Chris@10
|
781 Ip[WS(rs, 6)] = FNMS(T2F, T2Q, T2b * T2C);
|
Chris@10
|
782 Im[WS(rs, 6)] = FMA(T2F, T2C, T2b * T2Q);
|
Chris@10
|
783 Ip[WS(rs, 2)] = FNMS(T2T, T2U, T2R * T2S);
|
Chris@10
|
784 Im[WS(rs, 2)] = FMA(T2T, T2S, T2R * T2U);
|
Chris@10
|
785 }
|
Chris@10
|
786 {
|
Chris@10
|
787 E T2X, T31, T30, T32;
|
Chris@10
|
788 {
|
Chris@10
|
789 E T2V, T2W, T2Y, T2Z;
|
Chris@10
|
790 T2V = T2e + T2l;
|
Chris@10
|
791 T2W = T2N + T2O;
|
Chris@10
|
792 T2X = T2V - T2W;
|
Chris@10
|
793 T31 = T2V + T2W;
|
Chris@10
|
794 T2Y = T2I + T2L;
|
Chris@10
|
795 T2Z = T2A + T2t;
|
Chris@10
|
796 T30 = T2Y - T2Z;
|
Chris@10
|
797 T32 = T2Y + T2Z;
|
Chris@10
|
798 }
|
Chris@10
|
799 Ip[WS(rs, 4)] = FNMS(Tz, T30, Tw * T2X);
|
Chris@10
|
800 Im[WS(rs, 4)] = FMA(Tw, T30, Tz * T2X);
|
Chris@10
|
801 Ip[0] = FNMS(Ty, T32, Tv * T31);
|
Chris@10
|
802 Im[0] = FMA(Tv, T32, Ty * T31);
|
Chris@10
|
803 }
|
Chris@10
|
804 {
|
Chris@10
|
805 E T20, T26, T24, T28;
|
Chris@10
|
806 {
|
Chris@10
|
807 E T1Y, T1Z, T22, T23;
|
Chris@10
|
808 T1Y = T7 - Te;
|
Chris@10
|
809 T1Z = T1U - T1T;
|
Chris@10
|
810 T20 = T1Y - T1Z;
|
Chris@10
|
811 T26 = T1Y + T1Z;
|
Chris@10
|
812 T22 = T1Q - T1R;
|
Chris@10
|
813 T23 = Tm - Tt;
|
Chris@10
|
814 T24 = T22 - T23;
|
Chris@10
|
815 T28 = T23 + T22;
|
Chris@10
|
816 }
|
Chris@10
|
817 Rp[WS(rs, 6)] = FNMS(T21, T24, T1X * T20);
|
Chris@10
|
818 Rm[WS(rs, 6)] = FMA(T1X, T24, T21 * T20);
|
Chris@10
|
819 Rp[WS(rs, 2)] = FNMS(T27, T28, T25 * T26);
|
Chris@10
|
820 Rm[WS(rs, 2)] = FMA(T25, T28, T27 * T26);
|
Chris@10
|
821 }
|
Chris@10
|
822 }
|
Chris@10
|
823 }
|
Chris@10
|
824 }
|
Chris@10
|
825 }
|
Chris@10
|
826
|
Chris@10
|
827 static const tw_instr twinstr[] = {
|
Chris@10
|
828 {TW_CEXP, 1, 1},
|
Chris@10
|
829 {TW_CEXP, 1, 3},
|
Chris@10
|
830 {TW_CEXP, 1, 9},
|
Chris@10
|
831 {TW_CEXP, 1, 15},
|
Chris@10
|
832 {TW_NEXT, 1, 0}
|
Chris@10
|
833 };
|
Chris@10
|
834
|
Chris@10
|
835 static const hc2c_desc desc = { 16, "hc2cb2_16", twinstr, &GENUS, {156, 68, 40, 0} };
|
Chris@10
|
836
|
Chris@10
|
837 void X(codelet_hc2cb2_16) (planner *p) {
|
Chris@10
|
838 X(khc2c_register) (p, hc2cb2_16, &desc, HC2C_VIA_RDFT);
|
Chris@10
|
839 }
|
Chris@10
|
840 #endif /* HAVE_FMA */
|