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 07:39:52 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_hc2hc.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -n 15 -dit -name hf_15 -include hf.h */
|
Chris@10
|
29
|
Chris@10
|
30 /*
|
Chris@10
|
31 * This function contains 184 FP additions, 140 FP multiplications,
|
Chris@10
|
32 * (or, 72 additions, 28 multiplications, 112 fused multiply/add),
|
Chris@10
|
33 * 97 stack variables, 6 constants, and 60 memory accesses
|
Chris@10
|
34 */
|
Chris@10
|
35 #include "hf.h"
|
Chris@10
|
36
|
Chris@10
|
37 static void hf_15(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
|
Chris@10
|
38 {
|
Chris@10
|
39 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
|
Chris@10
|
40 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
|
Chris@10
|
41 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
|
Chris@10
|
42 DK(KP618033988, +0.618033988749894848204586834365638117720309180);
|
Chris@10
|
43 DK(KP866025403, +0.866025403784438646763723170752936183471402627);
|
Chris@10
|
44 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
|
Chris@10
|
45 {
|
Chris@10
|
46 INT m;
|
Chris@10
|
47 for (m = mb, W = W + ((mb - 1) * 28); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 28, MAKE_VOLATILE_STRIDE(30, rs)) {
|
Chris@10
|
48 E T3v, T3E, T3G, T3A, T3y, T3z, T3F, T3B;
|
Chris@10
|
49 {
|
Chris@10
|
50 E T1G, T3l, T3H, T3k, T1B, Tf, T37, T1y, T2Y, T2M, T2a, T2i, T39, Tz, T2U;
|
Chris@10
|
51 E T2t, T1O, T2e, T3a, TT, T10, T2V, T2z, T1V, T2f, T2C, T12, T15, T14, T21;
|
Chris@10
|
52 E T1c, T1Y, T13;
|
Chris@10
|
53 {
|
Chris@10
|
54 E T2I, T1k, T1m, T1p, T1o, T28, T1w, T25, T1n;
|
Chris@10
|
55 {
|
Chris@10
|
56 E T1, T3i, T9, Tc, Tb, T1D, T7, T1E, Ta, T1j, T1i, T1h;
|
Chris@10
|
57 T1 = cr[0];
|
Chris@10
|
58 T3i = ci[0];
|
Chris@10
|
59 {
|
Chris@10
|
60 E T3, T6, T2, T5, T1C, T4, T8;
|
Chris@10
|
61 T3 = cr[WS(rs, 5)];
|
Chris@10
|
62 T6 = ci[WS(rs, 5)];
|
Chris@10
|
63 T2 = W[8];
|
Chris@10
|
64 T5 = W[9];
|
Chris@10
|
65 T9 = cr[WS(rs, 10)];
|
Chris@10
|
66 Tc = ci[WS(rs, 10)];
|
Chris@10
|
67 T1C = T2 * T6;
|
Chris@10
|
68 T4 = T2 * T3;
|
Chris@10
|
69 T8 = W[18];
|
Chris@10
|
70 Tb = W[19];
|
Chris@10
|
71 T1D = FNMS(T5, T3, T1C);
|
Chris@10
|
72 T7 = FMA(T5, T6, T4);
|
Chris@10
|
73 T1E = T8 * Tc;
|
Chris@10
|
74 Ta = T8 * T9;
|
Chris@10
|
75 }
|
Chris@10
|
76 {
|
Chris@10
|
77 E T1g, T1F, Td, T1f, T3j, Te, T2H;
|
Chris@10
|
78 T1g = cr[WS(rs, 9)];
|
Chris@10
|
79 T1j = ci[WS(rs, 9)];
|
Chris@10
|
80 T1F = FNMS(Tb, T9, T1E);
|
Chris@10
|
81 Td = FMA(Tb, Tc, Ta);
|
Chris@10
|
82 T1f = W[16];
|
Chris@10
|
83 T1i = W[17];
|
Chris@10
|
84 T1G = T1D - T1F;
|
Chris@10
|
85 T3j = T1D + T1F;
|
Chris@10
|
86 T3l = Td - T7;
|
Chris@10
|
87 Te = T7 + Td;
|
Chris@10
|
88 T2H = T1f * T1j;
|
Chris@10
|
89 T1h = T1f * T1g;
|
Chris@10
|
90 T3H = T3j + T3i;
|
Chris@10
|
91 T3k = FNMS(KP500000000, T3j, T3i);
|
Chris@10
|
92 T1B = FNMS(KP500000000, Te, T1);
|
Chris@10
|
93 Tf = T1 + Te;
|
Chris@10
|
94 T2I = FNMS(T1i, T1g, T2H);
|
Chris@10
|
95 }
|
Chris@10
|
96 T1k = FMA(T1i, T1j, T1h);
|
Chris@10
|
97 {
|
Chris@10
|
98 E T1s, T1v, T1r, T1u, T27, T1t, T1l;
|
Chris@10
|
99 T1s = cr[WS(rs, 4)];
|
Chris@10
|
100 T1v = ci[WS(rs, 4)];
|
Chris@10
|
101 T1r = W[6];
|
Chris@10
|
102 T1u = W[7];
|
Chris@10
|
103 T1m = cr[WS(rs, 14)];
|
Chris@10
|
104 T1p = ci[WS(rs, 14)];
|
Chris@10
|
105 T27 = T1r * T1v;
|
Chris@10
|
106 T1t = T1r * T1s;
|
Chris@10
|
107 T1l = W[26];
|
Chris@10
|
108 T1o = W[27];
|
Chris@10
|
109 T28 = FNMS(T1u, T1s, T27);
|
Chris@10
|
110 T1w = FMA(T1u, T1v, T1t);
|
Chris@10
|
111 T25 = T1l * T1p;
|
Chris@10
|
112 T1n = T1l * T1m;
|
Chris@10
|
113 }
|
Chris@10
|
114 }
|
Chris@10
|
115 {
|
Chris@10
|
116 E Tl, T2p, Tn, Tq, Tp, T1M, Tx, T1J, To;
|
Chris@10
|
117 {
|
Chris@10
|
118 E Th, Tk, T26, T1q, Tg, Tj;
|
Chris@10
|
119 Th = cr[WS(rs, 3)];
|
Chris@10
|
120 Tk = ci[WS(rs, 3)];
|
Chris@10
|
121 T26 = FNMS(T1o, T1m, T25);
|
Chris@10
|
122 T1q = FMA(T1o, T1p, T1n);
|
Chris@10
|
123 Tg = W[4];
|
Chris@10
|
124 Tj = W[5];
|
Chris@10
|
125 {
|
Chris@10
|
126 E T29, T2J, T1x, T2L;
|
Chris@10
|
127 T29 = T26 - T28;
|
Chris@10
|
128 T2J = T26 + T28;
|
Chris@10
|
129 T1x = T1q + T1w;
|
Chris@10
|
130 T2L = T1q - T1w;
|
Chris@10
|
131 {
|
Chris@10
|
132 E T2o, Ti, T2K, T24;
|
Chris@10
|
133 T2o = Tg * Tk;
|
Chris@10
|
134 Ti = Tg * Th;
|
Chris@10
|
135 T2K = FNMS(KP500000000, T2J, T2I);
|
Chris@10
|
136 T37 = T2I + T2J;
|
Chris@10
|
137 T24 = FNMS(KP500000000, T1x, T1k);
|
Chris@10
|
138 T1y = T1k + T1x;
|
Chris@10
|
139 Tl = FMA(Tj, Tk, Ti);
|
Chris@10
|
140 T2Y = FMA(KP866025403, T2L, T2K);
|
Chris@10
|
141 T2M = FNMS(KP866025403, T2L, T2K);
|
Chris@10
|
142 T2a = FNMS(KP866025403, T29, T24);
|
Chris@10
|
143 T2i = FMA(KP866025403, T29, T24);
|
Chris@10
|
144 T2p = FNMS(Tj, Th, T2o);
|
Chris@10
|
145 }
|
Chris@10
|
146 }
|
Chris@10
|
147 }
|
Chris@10
|
148 {
|
Chris@10
|
149 E Tt, Tw, Ts, Tv, T1L, Tu, Tm;
|
Chris@10
|
150 Tt = cr[WS(rs, 13)];
|
Chris@10
|
151 Tw = ci[WS(rs, 13)];
|
Chris@10
|
152 Ts = W[24];
|
Chris@10
|
153 Tv = W[25];
|
Chris@10
|
154 Tn = cr[WS(rs, 8)];
|
Chris@10
|
155 Tq = ci[WS(rs, 8)];
|
Chris@10
|
156 T1L = Ts * Tw;
|
Chris@10
|
157 Tu = Ts * Tt;
|
Chris@10
|
158 Tm = W[14];
|
Chris@10
|
159 Tp = W[15];
|
Chris@10
|
160 T1M = FNMS(Tv, Tt, T1L);
|
Chris@10
|
161 Tx = FMA(Tv, Tw, Tu);
|
Chris@10
|
162 T1J = Tm * Tq;
|
Chris@10
|
163 To = Tm * Tn;
|
Chris@10
|
164 }
|
Chris@10
|
165 {
|
Chris@10
|
166 E TF, T2v, TH, TK, TJ, T1T, TR, T1Q, TI;
|
Chris@10
|
167 {
|
Chris@10
|
168 E TB, TE, T1K, Tr, TA, TD;
|
Chris@10
|
169 TB = cr[WS(rs, 12)];
|
Chris@10
|
170 TE = ci[WS(rs, 12)];
|
Chris@10
|
171 T1K = FNMS(Tp, Tn, T1J);
|
Chris@10
|
172 Tr = FMA(Tp, Tq, To);
|
Chris@10
|
173 TA = W[22];
|
Chris@10
|
174 TD = W[23];
|
Chris@10
|
175 {
|
Chris@10
|
176 E T1N, T2q, Ty, T2s;
|
Chris@10
|
177 T1N = T1K - T1M;
|
Chris@10
|
178 T2q = T1K + T1M;
|
Chris@10
|
179 Ty = Tr + Tx;
|
Chris@10
|
180 T2s = Tr - Tx;
|
Chris@10
|
181 {
|
Chris@10
|
182 E T2u, TC, T2r, T1I;
|
Chris@10
|
183 T2u = TA * TE;
|
Chris@10
|
184 TC = TA * TB;
|
Chris@10
|
185 T2r = FNMS(KP500000000, T2q, T2p);
|
Chris@10
|
186 T39 = T2p + T2q;
|
Chris@10
|
187 T1I = FNMS(KP500000000, Ty, Tl);
|
Chris@10
|
188 Tz = Tl + Ty;
|
Chris@10
|
189 TF = FMA(TD, TE, TC);
|
Chris@10
|
190 T2U = FMA(KP866025403, T2s, T2r);
|
Chris@10
|
191 T2t = FNMS(KP866025403, T2s, T2r);
|
Chris@10
|
192 T1O = FNMS(KP866025403, T1N, T1I);
|
Chris@10
|
193 T2e = FMA(KP866025403, T1N, T1I);
|
Chris@10
|
194 T2v = FNMS(TD, TB, T2u);
|
Chris@10
|
195 }
|
Chris@10
|
196 }
|
Chris@10
|
197 }
|
Chris@10
|
198 {
|
Chris@10
|
199 E TN, TQ, TM, TP, T1S, TO, TG;
|
Chris@10
|
200 TN = cr[WS(rs, 7)];
|
Chris@10
|
201 TQ = ci[WS(rs, 7)];
|
Chris@10
|
202 TM = W[12];
|
Chris@10
|
203 TP = W[13];
|
Chris@10
|
204 TH = cr[WS(rs, 2)];
|
Chris@10
|
205 TK = ci[WS(rs, 2)];
|
Chris@10
|
206 T1S = TM * TQ;
|
Chris@10
|
207 TO = TM * TN;
|
Chris@10
|
208 TG = W[2];
|
Chris@10
|
209 TJ = W[3];
|
Chris@10
|
210 T1T = FNMS(TP, TN, T1S);
|
Chris@10
|
211 TR = FMA(TP, TQ, TO);
|
Chris@10
|
212 T1Q = TG * TK;
|
Chris@10
|
213 TI = TG * TH;
|
Chris@10
|
214 }
|
Chris@10
|
215 {
|
Chris@10
|
216 E TW, TZ, T1R, TL, TV, TY;
|
Chris@10
|
217 TW = cr[WS(rs, 6)];
|
Chris@10
|
218 TZ = ci[WS(rs, 6)];
|
Chris@10
|
219 T1R = FNMS(TJ, TH, T1Q);
|
Chris@10
|
220 TL = FMA(TJ, TK, TI);
|
Chris@10
|
221 TV = W[10];
|
Chris@10
|
222 TY = W[11];
|
Chris@10
|
223 {
|
Chris@10
|
224 E T1U, T2w, TS, T2y;
|
Chris@10
|
225 T1U = T1R - T1T;
|
Chris@10
|
226 T2w = T1R + T1T;
|
Chris@10
|
227 TS = TL + TR;
|
Chris@10
|
228 T2y = TL - TR;
|
Chris@10
|
229 {
|
Chris@10
|
230 E T2B, TX, T2x, T1P;
|
Chris@10
|
231 T2B = TV * TZ;
|
Chris@10
|
232 TX = TV * TW;
|
Chris@10
|
233 T2x = FNMS(KP500000000, T2w, T2v);
|
Chris@10
|
234 T3a = T2v + T2w;
|
Chris@10
|
235 T1P = FNMS(KP500000000, TS, TF);
|
Chris@10
|
236 TT = TF + TS;
|
Chris@10
|
237 T10 = FMA(TY, TZ, TX);
|
Chris@10
|
238 T2V = FMA(KP866025403, T2y, T2x);
|
Chris@10
|
239 T2z = FNMS(KP866025403, T2y, T2x);
|
Chris@10
|
240 T1V = FNMS(KP866025403, T1U, T1P);
|
Chris@10
|
241 T2f = FMA(KP866025403, T1U, T1P);
|
Chris@10
|
242 T2C = FNMS(TY, TW, T2B);
|
Chris@10
|
243 }
|
Chris@10
|
244 }
|
Chris@10
|
245 }
|
Chris@10
|
246 {
|
Chris@10
|
247 E T18, T1b, T17, T1a, T20, T19, T11;
|
Chris@10
|
248 T18 = cr[WS(rs, 1)];
|
Chris@10
|
249 T1b = ci[WS(rs, 1)];
|
Chris@10
|
250 T17 = W[0];
|
Chris@10
|
251 T1a = W[1];
|
Chris@10
|
252 T12 = cr[WS(rs, 11)];
|
Chris@10
|
253 T15 = ci[WS(rs, 11)];
|
Chris@10
|
254 T20 = T17 * T1b;
|
Chris@10
|
255 T19 = T17 * T18;
|
Chris@10
|
256 T11 = W[20];
|
Chris@10
|
257 T14 = W[21];
|
Chris@10
|
258 T21 = FNMS(T1a, T18, T20);
|
Chris@10
|
259 T1c = FMA(T1a, T1b, T19);
|
Chris@10
|
260 T1Y = T11 * T15;
|
Chris@10
|
261 T13 = T11 * T12;
|
Chris@10
|
262 }
|
Chris@10
|
263 }
|
Chris@10
|
264 }
|
Chris@10
|
265 }
|
Chris@10
|
266 {
|
Chris@10
|
267 E T3I, T3O, T3w, T2d, T3J, T3P, T3x, T3C, T3D, T3f, T3g, T2Q, T2O, T3r, T3q;
|
Chris@10
|
268 E T2k, T2m;
|
Chris@10
|
269 {
|
Chris@10
|
270 E T3b, T1Z, T16, TU;
|
Chris@10
|
271 T3I = T39 + T3a;
|
Chris@10
|
272 T3b = T39 - T3a;
|
Chris@10
|
273 T1Z = FNMS(T14, T12, T1Y);
|
Chris@10
|
274 T16 = FMA(T14, T15, T13);
|
Chris@10
|
275 T3O = TT - Tz;
|
Chris@10
|
276 TU = Tz + TT;
|
Chris@10
|
277 {
|
Chris@10
|
278 E T1H, T2G, T2h, T3e, T3c, T34, T1W, T32, T30, T33, T2b, T2S, T2R;
|
Chris@10
|
279 {
|
Chris@10
|
280 E T2W, T22, T1d, T2F, T2E, T36, T2D;
|
Chris@10
|
281 T2W = T2U - T2V;
|
Chris@10
|
282 T3w = T2U + T2V;
|
Chris@10
|
283 T22 = T1Z - T21;
|
Chris@10
|
284 T2D = T1Z + T21;
|
Chris@10
|
285 T1d = T16 + T1c;
|
Chris@10
|
286 T2F = T16 - T1c;
|
Chris@10
|
287 T2E = FNMS(KP500000000, T2D, T2C);
|
Chris@10
|
288 T36 = T2C + T2D;
|
Chris@10
|
289 T2d = FMA(KP866025403, T1G, T1B);
|
Chris@10
|
290 T1H = FNMS(KP866025403, T1G, T1B);
|
Chris@10
|
291 {
|
Chris@10
|
292 E T1e, T1X, T38, T2X;
|
Chris@10
|
293 T1e = T10 + T1d;
|
Chris@10
|
294 T1X = FNMS(KP500000000, T1d, T10);
|
Chris@10
|
295 T38 = T36 - T37;
|
Chris@10
|
296 T3J = T36 + T37;
|
Chris@10
|
297 T2G = FNMS(KP866025403, T2F, T2E);
|
Chris@10
|
298 T2X = FMA(KP866025403, T2F, T2E);
|
Chris@10
|
299 {
|
Chris@10
|
300 E T1z, T23, T2Z, T1A;
|
Chris@10
|
301 T3P = T1y - T1e;
|
Chris@10
|
302 T1z = T1e + T1y;
|
Chris@10
|
303 T23 = FNMS(KP866025403, T22, T1X);
|
Chris@10
|
304 T2h = FMA(KP866025403, T22, T1X);
|
Chris@10
|
305 T3e = FMA(KP618033988, T38, T3b);
|
Chris@10
|
306 T3c = FNMS(KP618033988, T3b, T38);
|
Chris@10
|
307 T2Z = T2X - T2Y;
|
Chris@10
|
308 T3x = T2X + T2Y;
|
Chris@10
|
309 T1A = TU + T1z;
|
Chris@10
|
310 T34 = TU - T1z;
|
Chris@10
|
311 T3C = T1O - T1V;
|
Chris@10
|
312 T1W = T1O + T1V;
|
Chris@10
|
313 T32 = FNMS(KP618033988, T2W, T2Z);
|
Chris@10
|
314 T30 = FMA(KP618033988, T2Z, T2W);
|
Chris@10
|
315 cr[0] = Tf + T1A;
|
Chris@10
|
316 T33 = FNMS(KP250000000, T1A, Tf);
|
Chris@10
|
317 T2b = T23 + T2a;
|
Chris@10
|
318 T3D = T23 - T2a;
|
Chris@10
|
319 }
|
Chris@10
|
320 }
|
Chris@10
|
321 }
|
Chris@10
|
322 {
|
Chris@10
|
323 E T2A, T2N, T3d, T35, T2c;
|
Chris@10
|
324 T3f = T2t + T2z;
|
Chris@10
|
325 T2A = T2t - T2z;
|
Chris@10
|
326 T2N = T2G - T2M;
|
Chris@10
|
327 T3g = T2G + T2M;
|
Chris@10
|
328 T3d = FMA(KP559016994, T34, T33);
|
Chris@10
|
329 T35 = FNMS(KP559016994, T34, T33);
|
Chris@10
|
330 T2c = T1W + T2b;
|
Chris@10
|
331 T2S = T1W - T2b;
|
Chris@10
|
332 cr[WS(rs, 3)] = FMA(KP951056516, T3c, T35);
|
Chris@10
|
333 ci[WS(rs, 2)] = FNMS(KP951056516, T3c, T35);
|
Chris@10
|
334 cr[WS(rs, 6)] = FMA(KP951056516, T3e, T3d);
|
Chris@10
|
335 ci[WS(rs, 5)] = FNMS(KP951056516, T3e, T3d);
|
Chris@10
|
336 cr[WS(rs, 5)] = T1H + T2c;
|
Chris@10
|
337 T2R = FNMS(KP250000000, T2c, T1H);
|
Chris@10
|
338 T2Q = FNMS(KP618033988, T2A, T2N);
|
Chris@10
|
339 T2O = FMA(KP618033988, T2N, T2A);
|
Chris@10
|
340 }
|
Chris@10
|
341 {
|
Chris@10
|
342 E T2T, T31, T2g, T2j;
|
Chris@10
|
343 T2T = FMA(KP559016994, T2S, T2R);
|
Chris@10
|
344 T31 = FNMS(KP559016994, T2S, T2R);
|
Chris@10
|
345 T2g = T2e + T2f;
|
Chris@10
|
346 T3r = T2e - T2f;
|
Chris@10
|
347 T3q = T2h - T2i;
|
Chris@10
|
348 T2j = T2h + T2i;
|
Chris@10
|
349 ci[WS(rs, 3)] = FMA(KP951056516, T30, T2T);
|
Chris@10
|
350 ci[0] = FNMS(KP951056516, T30, T2T);
|
Chris@10
|
351 ci[WS(rs, 6)] = FMA(KP951056516, T32, T31);
|
Chris@10
|
352 cr[WS(rs, 2)] = FNMS(KP951056516, T32, T31);
|
Chris@10
|
353 T2k = T2g + T2j;
|
Chris@10
|
354 T2m = T2g - T2j;
|
Chris@10
|
355 }
|
Chris@10
|
356 }
|
Chris@10
|
357 }
|
Chris@10
|
358 {
|
Chris@10
|
359 E T3m, T3s, T3u, T3o, T3h, T2l, T2n, T2P;
|
Chris@10
|
360 ci[WS(rs, 4)] = T2d + T2k;
|
Chris@10
|
361 T2l = FNMS(KP250000000, T2k, T2d);
|
Chris@10
|
362 T3m = FMA(KP866025403, T3l, T3k);
|
Chris@10
|
363 T3v = FNMS(KP866025403, T3l, T3k);
|
Chris@10
|
364 T3s = FNMS(KP618033988, T3r, T3q);
|
Chris@10
|
365 T3u = FMA(KP618033988, T3q, T3r);
|
Chris@10
|
366 T2n = FMA(KP559016994, T2m, T2l);
|
Chris@10
|
367 T2P = FNMS(KP559016994, T2m, T2l);
|
Chris@10
|
368 ci[WS(rs, 1)] = FMA(KP951056516, T2Q, T2P);
|
Chris@10
|
369 cr[WS(rs, 7)] = FNMS(KP951056516, T2Q, T2P);
|
Chris@10
|
370 cr[WS(rs, 1)] = FMA(KP951056516, T2O, T2n);
|
Chris@10
|
371 cr[WS(rs, 4)] = FNMS(KP951056516, T2O, T2n);
|
Chris@10
|
372 T3o = T3f - T3g;
|
Chris@10
|
373 T3h = T3f + T3g;
|
Chris@10
|
374 {
|
Chris@10
|
375 E T3S, T3Q, T3K, T3M, T3n, T3p, T3t, T3L, T3R, T3N;
|
Chris@10
|
376 cr[WS(rs, 10)] = -(T3h + T3m);
|
Chris@10
|
377 T3n = FNMS(KP250000000, T3h, T3m);
|
Chris@10
|
378 T3S = FNMS(KP618033988, T3O, T3P);
|
Chris@10
|
379 T3Q = FMA(KP618033988, T3P, T3O);
|
Chris@10
|
380 T3p = FNMS(KP559016994, T3o, T3n);
|
Chris@10
|
381 T3t = FMA(KP559016994, T3o, T3n);
|
Chris@10
|
382 ci[WS(rs, 7)] = FMA(KP951056516, T3s, T3p);
|
Chris@10
|
383 cr[WS(rs, 13)] = FMS(KP951056516, T3s, T3p);
|
Chris@10
|
384 ci[WS(rs, 13)] = FNMS(KP951056516, T3u, T3t);
|
Chris@10
|
385 ci[WS(rs, 10)] = FMA(KP951056516, T3u, T3t);
|
Chris@10
|
386 T3K = T3I + T3J;
|
Chris@10
|
387 T3M = T3I - T3J;
|
Chris@10
|
388 ci[WS(rs, 14)] = T3K + T3H;
|
Chris@10
|
389 T3L = FNMS(KP250000000, T3K, T3H);
|
Chris@10
|
390 T3E = FMA(KP618033988, T3D, T3C);
|
Chris@10
|
391 T3G = FNMS(KP618033988, T3C, T3D);
|
Chris@10
|
392 T3R = FNMS(KP559016994, T3M, T3L);
|
Chris@10
|
393 T3N = FMA(KP559016994, T3M, T3L);
|
Chris@10
|
394 ci[WS(rs, 8)] = FMA(KP951056516, T3Q, T3N);
|
Chris@10
|
395 cr[WS(rs, 9)] = FMS(KP951056516, T3Q, T3N);
|
Chris@10
|
396 ci[WS(rs, 11)] = FMA(KP951056516, T3S, T3R);
|
Chris@10
|
397 cr[WS(rs, 12)] = FMS(KP951056516, T3S, T3R);
|
Chris@10
|
398 T3A = T3x - T3w;
|
Chris@10
|
399 T3y = T3w + T3x;
|
Chris@10
|
400 }
|
Chris@10
|
401 }
|
Chris@10
|
402 }
|
Chris@10
|
403 }
|
Chris@10
|
404 ci[WS(rs, 9)] = T3y + T3v;
|
Chris@10
|
405 T3z = FNMS(KP250000000, T3y, T3v);
|
Chris@10
|
406 T3F = FMA(KP559016994, T3A, T3z);
|
Chris@10
|
407 T3B = FNMS(KP559016994, T3A, T3z);
|
Chris@10
|
408 cr[WS(rs, 14)] = -(FMA(KP951056516, T3E, T3B));
|
Chris@10
|
409 cr[WS(rs, 11)] = FMS(KP951056516, T3E, T3B);
|
Chris@10
|
410 ci[WS(rs, 12)] = FMA(KP951056516, T3G, T3F);
|
Chris@10
|
411 cr[WS(rs, 8)] = FMS(KP951056516, T3G, T3F);
|
Chris@10
|
412 }
|
Chris@10
|
413 }
|
Chris@10
|
414 }
|
Chris@10
|
415
|
Chris@10
|
416 static const tw_instr twinstr[] = {
|
Chris@10
|
417 {TW_FULL, 1, 15},
|
Chris@10
|
418 {TW_NEXT, 1, 0}
|
Chris@10
|
419 };
|
Chris@10
|
420
|
Chris@10
|
421 static const hc2hc_desc desc = { 15, "hf_15", twinstr, &GENUS, {72, 28, 112, 0} };
|
Chris@10
|
422
|
Chris@10
|
423 void X(codelet_hf_15) (planner *p) {
|
Chris@10
|
424 X(khc2hc_register) (p, hf_15, &desc);
|
Chris@10
|
425 }
|
Chris@10
|
426 #else /* HAVE_FMA */
|
Chris@10
|
427
|
Chris@10
|
428 /* Generated by: ../../../genfft/gen_hc2hc.native -compact -variables 4 -pipeline-latency 4 -n 15 -dit -name hf_15 -include hf.h */
|
Chris@10
|
429
|
Chris@10
|
430 /*
|
Chris@10
|
431 * This function contains 184 FP additions, 112 FP multiplications,
|
Chris@10
|
432 * (or, 128 additions, 56 multiplications, 56 fused multiply/add),
|
Chris@10
|
433 * 65 stack variables, 6 constants, and 60 memory accesses
|
Chris@10
|
434 */
|
Chris@10
|
435 #include "hf.h"
|
Chris@10
|
436
|
Chris@10
|
437 static void hf_15(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
|
Chris@10
|
438 {
|
Chris@10
|
439 DK(KP587785252, +0.587785252292473129168705954639072768597652438);
|
Chris@10
|
440 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
|
Chris@10
|
441 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
|
Chris@10
|
442 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
|
Chris@10
|
443 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
|
Chris@10
|
444 DK(KP866025403, +0.866025403784438646763723170752936183471402627);
|
Chris@10
|
445 {
|
Chris@10
|
446 INT m;
|
Chris@10
|
447 for (m = mb, W = W + ((mb - 1) * 28); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 28, MAKE_VOLATILE_STRIDE(30, rs)) {
|
Chris@10
|
448 E T1q, T2Q, Td, T1n, T2T, T3l, T13, T1k, T1l, T2E, T2F, T3j, T1H, T1T, T2k;
|
Chris@10
|
449 E T2w, T2f, T2v, T1M, T1U, Tu, TL, TM, T2H, T2I, T3i, T1w, T1Q, T29, T2t;
|
Chris@10
|
450 E T24, T2s, T1B, T1R;
|
Chris@10
|
451 {
|
Chris@10
|
452 E T1, T2R, T6, T1o, Tb, T1p, Tc, T2S;
|
Chris@10
|
453 T1 = cr[0];
|
Chris@10
|
454 T2R = ci[0];
|
Chris@10
|
455 {
|
Chris@10
|
456 E T3, T5, T2, T4;
|
Chris@10
|
457 T3 = cr[WS(rs, 5)];
|
Chris@10
|
458 T5 = ci[WS(rs, 5)];
|
Chris@10
|
459 T2 = W[8];
|
Chris@10
|
460 T4 = W[9];
|
Chris@10
|
461 T6 = FMA(T2, T3, T4 * T5);
|
Chris@10
|
462 T1o = FNMS(T4, T3, T2 * T5);
|
Chris@10
|
463 }
|
Chris@10
|
464 {
|
Chris@10
|
465 E T8, Ta, T7, T9;
|
Chris@10
|
466 T8 = cr[WS(rs, 10)];
|
Chris@10
|
467 Ta = ci[WS(rs, 10)];
|
Chris@10
|
468 T7 = W[18];
|
Chris@10
|
469 T9 = W[19];
|
Chris@10
|
470 Tb = FMA(T7, T8, T9 * Ta);
|
Chris@10
|
471 T1p = FNMS(T9, T8, T7 * Ta);
|
Chris@10
|
472 }
|
Chris@10
|
473 T1q = KP866025403 * (T1o - T1p);
|
Chris@10
|
474 T2Q = KP866025403 * (Tb - T6);
|
Chris@10
|
475 Tc = T6 + Tb;
|
Chris@10
|
476 Td = T1 + Tc;
|
Chris@10
|
477 T1n = FNMS(KP500000000, Tc, T1);
|
Chris@10
|
478 T2S = T1o + T1p;
|
Chris@10
|
479 T2T = FNMS(KP500000000, T2S, T2R);
|
Chris@10
|
480 T3l = T2S + T2R;
|
Chris@10
|
481 }
|
Chris@10
|
482 {
|
Chris@10
|
483 E TR, T2c, T18, T2h, TW, T1E, T11, T1F, T12, T2d, T1d, T1J, T1i, T1K, T1j;
|
Chris@10
|
484 E T2i;
|
Chris@10
|
485 {
|
Chris@10
|
486 E TO, TQ, TN, TP;
|
Chris@10
|
487 TO = cr[WS(rs, 6)];
|
Chris@10
|
488 TQ = ci[WS(rs, 6)];
|
Chris@10
|
489 TN = W[10];
|
Chris@10
|
490 TP = W[11];
|
Chris@10
|
491 TR = FMA(TN, TO, TP * TQ);
|
Chris@10
|
492 T2c = FNMS(TP, TO, TN * TQ);
|
Chris@10
|
493 }
|
Chris@10
|
494 {
|
Chris@10
|
495 E T15, T17, T14, T16;
|
Chris@10
|
496 T15 = cr[WS(rs, 9)];
|
Chris@10
|
497 T17 = ci[WS(rs, 9)];
|
Chris@10
|
498 T14 = W[16];
|
Chris@10
|
499 T16 = W[17];
|
Chris@10
|
500 T18 = FMA(T14, T15, T16 * T17);
|
Chris@10
|
501 T2h = FNMS(T16, T15, T14 * T17);
|
Chris@10
|
502 }
|
Chris@10
|
503 {
|
Chris@10
|
504 E TT, TV, TS, TU;
|
Chris@10
|
505 TT = cr[WS(rs, 11)];
|
Chris@10
|
506 TV = ci[WS(rs, 11)];
|
Chris@10
|
507 TS = W[20];
|
Chris@10
|
508 TU = W[21];
|
Chris@10
|
509 TW = FMA(TS, TT, TU * TV);
|
Chris@10
|
510 T1E = FNMS(TU, TT, TS * TV);
|
Chris@10
|
511 }
|
Chris@10
|
512 {
|
Chris@10
|
513 E TY, T10, TX, TZ;
|
Chris@10
|
514 TY = cr[WS(rs, 1)];
|
Chris@10
|
515 T10 = ci[WS(rs, 1)];
|
Chris@10
|
516 TX = W[0];
|
Chris@10
|
517 TZ = W[1];
|
Chris@10
|
518 T11 = FMA(TX, TY, TZ * T10);
|
Chris@10
|
519 T1F = FNMS(TZ, TY, TX * T10);
|
Chris@10
|
520 }
|
Chris@10
|
521 T12 = TW + T11;
|
Chris@10
|
522 T2d = T1E + T1F;
|
Chris@10
|
523 {
|
Chris@10
|
524 E T1a, T1c, T19, T1b;
|
Chris@10
|
525 T1a = cr[WS(rs, 14)];
|
Chris@10
|
526 T1c = ci[WS(rs, 14)];
|
Chris@10
|
527 T19 = W[26];
|
Chris@10
|
528 T1b = W[27];
|
Chris@10
|
529 T1d = FMA(T19, T1a, T1b * T1c);
|
Chris@10
|
530 T1J = FNMS(T1b, T1a, T19 * T1c);
|
Chris@10
|
531 }
|
Chris@10
|
532 {
|
Chris@10
|
533 E T1f, T1h, T1e, T1g;
|
Chris@10
|
534 T1f = cr[WS(rs, 4)];
|
Chris@10
|
535 T1h = ci[WS(rs, 4)];
|
Chris@10
|
536 T1e = W[6];
|
Chris@10
|
537 T1g = W[7];
|
Chris@10
|
538 T1i = FMA(T1e, T1f, T1g * T1h);
|
Chris@10
|
539 T1K = FNMS(T1g, T1f, T1e * T1h);
|
Chris@10
|
540 }
|
Chris@10
|
541 T1j = T1d + T1i;
|
Chris@10
|
542 T2i = T1J + T1K;
|
Chris@10
|
543 {
|
Chris@10
|
544 E T1D, T1G, T2g, T2j;
|
Chris@10
|
545 T13 = TR + T12;
|
Chris@10
|
546 T1k = T18 + T1j;
|
Chris@10
|
547 T1l = T13 + T1k;
|
Chris@10
|
548 T2E = T2c + T2d;
|
Chris@10
|
549 T2F = T2h + T2i;
|
Chris@10
|
550 T3j = T2E + T2F;
|
Chris@10
|
551 T1D = FNMS(KP500000000, T12, TR);
|
Chris@10
|
552 T1G = KP866025403 * (T1E - T1F);
|
Chris@10
|
553 T1H = T1D - T1G;
|
Chris@10
|
554 T1T = T1D + T1G;
|
Chris@10
|
555 T2g = KP866025403 * (T1d - T1i);
|
Chris@10
|
556 T2j = FNMS(KP500000000, T2i, T2h);
|
Chris@10
|
557 T2k = T2g - T2j;
|
Chris@10
|
558 T2w = T2g + T2j;
|
Chris@10
|
559 {
|
Chris@10
|
560 E T2b, T2e, T1I, T1L;
|
Chris@10
|
561 T2b = KP866025403 * (T11 - TW);
|
Chris@10
|
562 T2e = FNMS(KP500000000, T2d, T2c);
|
Chris@10
|
563 T2f = T2b + T2e;
|
Chris@10
|
564 T2v = T2e - T2b;
|
Chris@10
|
565 T1I = FNMS(KP500000000, T1j, T18);
|
Chris@10
|
566 T1L = KP866025403 * (T1J - T1K);
|
Chris@10
|
567 T1M = T1I - T1L;
|
Chris@10
|
568 T1U = T1I + T1L;
|
Chris@10
|
569 }
|
Chris@10
|
570 }
|
Chris@10
|
571 }
|
Chris@10
|
572 {
|
Chris@10
|
573 E Ti, T21, Tz, T26, Tn, T1t, Ts, T1u, Tt, T22, TE, T1y, TJ, T1z, TK;
|
Chris@10
|
574 E T27;
|
Chris@10
|
575 {
|
Chris@10
|
576 E Tf, Th, Te, Tg;
|
Chris@10
|
577 Tf = cr[WS(rs, 3)];
|
Chris@10
|
578 Th = ci[WS(rs, 3)];
|
Chris@10
|
579 Te = W[4];
|
Chris@10
|
580 Tg = W[5];
|
Chris@10
|
581 Ti = FMA(Te, Tf, Tg * Th);
|
Chris@10
|
582 T21 = FNMS(Tg, Tf, Te * Th);
|
Chris@10
|
583 }
|
Chris@10
|
584 {
|
Chris@10
|
585 E Tw, Ty, Tv, Tx;
|
Chris@10
|
586 Tw = cr[WS(rs, 12)];
|
Chris@10
|
587 Ty = ci[WS(rs, 12)];
|
Chris@10
|
588 Tv = W[22];
|
Chris@10
|
589 Tx = W[23];
|
Chris@10
|
590 Tz = FMA(Tv, Tw, Tx * Ty);
|
Chris@10
|
591 T26 = FNMS(Tx, Tw, Tv * Ty);
|
Chris@10
|
592 }
|
Chris@10
|
593 {
|
Chris@10
|
594 E Tk, Tm, Tj, Tl;
|
Chris@10
|
595 Tk = cr[WS(rs, 8)];
|
Chris@10
|
596 Tm = ci[WS(rs, 8)];
|
Chris@10
|
597 Tj = W[14];
|
Chris@10
|
598 Tl = W[15];
|
Chris@10
|
599 Tn = FMA(Tj, Tk, Tl * Tm);
|
Chris@10
|
600 T1t = FNMS(Tl, Tk, Tj * Tm);
|
Chris@10
|
601 }
|
Chris@10
|
602 {
|
Chris@10
|
603 E Tp, Tr, To, Tq;
|
Chris@10
|
604 Tp = cr[WS(rs, 13)];
|
Chris@10
|
605 Tr = ci[WS(rs, 13)];
|
Chris@10
|
606 To = W[24];
|
Chris@10
|
607 Tq = W[25];
|
Chris@10
|
608 Ts = FMA(To, Tp, Tq * Tr);
|
Chris@10
|
609 T1u = FNMS(Tq, Tp, To * Tr);
|
Chris@10
|
610 }
|
Chris@10
|
611 Tt = Tn + Ts;
|
Chris@10
|
612 T22 = T1t + T1u;
|
Chris@10
|
613 {
|
Chris@10
|
614 E TB, TD, TA, TC;
|
Chris@10
|
615 TB = cr[WS(rs, 2)];
|
Chris@10
|
616 TD = ci[WS(rs, 2)];
|
Chris@10
|
617 TA = W[2];
|
Chris@10
|
618 TC = W[3];
|
Chris@10
|
619 TE = FMA(TA, TB, TC * TD);
|
Chris@10
|
620 T1y = FNMS(TC, TB, TA * TD);
|
Chris@10
|
621 }
|
Chris@10
|
622 {
|
Chris@10
|
623 E TG, TI, TF, TH;
|
Chris@10
|
624 TG = cr[WS(rs, 7)];
|
Chris@10
|
625 TI = ci[WS(rs, 7)];
|
Chris@10
|
626 TF = W[12];
|
Chris@10
|
627 TH = W[13];
|
Chris@10
|
628 TJ = FMA(TF, TG, TH * TI);
|
Chris@10
|
629 T1z = FNMS(TH, TG, TF * TI);
|
Chris@10
|
630 }
|
Chris@10
|
631 TK = TE + TJ;
|
Chris@10
|
632 T27 = T1y + T1z;
|
Chris@10
|
633 {
|
Chris@10
|
634 E T1s, T1v, T25, T28;
|
Chris@10
|
635 Tu = Ti + Tt;
|
Chris@10
|
636 TL = Tz + TK;
|
Chris@10
|
637 TM = Tu + TL;
|
Chris@10
|
638 T2H = T21 + T22;
|
Chris@10
|
639 T2I = T26 + T27;
|
Chris@10
|
640 T3i = T2H + T2I;
|
Chris@10
|
641 T1s = FNMS(KP500000000, Tt, Ti);
|
Chris@10
|
642 T1v = KP866025403 * (T1t - T1u);
|
Chris@10
|
643 T1w = T1s - T1v;
|
Chris@10
|
644 T1Q = T1s + T1v;
|
Chris@10
|
645 T25 = KP866025403 * (TJ - TE);
|
Chris@10
|
646 T28 = FNMS(KP500000000, T27, T26);
|
Chris@10
|
647 T29 = T25 + T28;
|
Chris@10
|
648 T2t = T28 - T25;
|
Chris@10
|
649 {
|
Chris@10
|
650 E T20, T23, T1x, T1A;
|
Chris@10
|
651 T20 = KP866025403 * (Ts - Tn);
|
Chris@10
|
652 T23 = FNMS(KP500000000, T22, T21);
|
Chris@10
|
653 T24 = T20 + T23;
|
Chris@10
|
654 T2s = T23 - T20;
|
Chris@10
|
655 T1x = FNMS(KP500000000, TK, Tz);
|
Chris@10
|
656 T1A = KP866025403 * (T1y - T1z);
|
Chris@10
|
657 T1B = T1x - T1A;
|
Chris@10
|
658 T1R = T1x + T1A;
|
Chris@10
|
659 }
|
Chris@10
|
660 }
|
Chris@10
|
661 }
|
Chris@10
|
662 {
|
Chris@10
|
663 E T2C, T1m, T2B, T2K, T2M, T2G, T2J, T2L, T2D;
|
Chris@10
|
664 T2C = KP559016994 * (TM - T1l);
|
Chris@10
|
665 T1m = TM + T1l;
|
Chris@10
|
666 T2B = FNMS(KP250000000, T1m, Td);
|
Chris@10
|
667 T2G = T2E - T2F;
|
Chris@10
|
668 T2J = T2H - T2I;
|
Chris@10
|
669 T2K = FNMS(KP587785252, T2J, KP951056516 * T2G);
|
Chris@10
|
670 T2M = FMA(KP951056516, T2J, KP587785252 * T2G);
|
Chris@10
|
671 cr[0] = Td + T1m;
|
Chris@10
|
672 T2L = T2C + T2B;
|
Chris@10
|
673 ci[WS(rs, 5)] = T2L - T2M;
|
Chris@10
|
674 cr[WS(rs, 6)] = T2L + T2M;
|
Chris@10
|
675 T2D = T2B - T2C;
|
Chris@10
|
676 ci[WS(rs, 2)] = T2D - T2K;
|
Chris@10
|
677 cr[WS(rs, 3)] = T2D + T2K;
|
Chris@10
|
678 }
|
Chris@10
|
679 {
|
Chris@10
|
680 E T3k, T3m, T3n, T3h, T3p, T3f, T3g, T3q, T3o;
|
Chris@10
|
681 T3k = KP559016994 * (T3i - T3j);
|
Chris@10
|
682 T3m = T3i + T3j;
|
Chris@10
|
683 T3n = FNMS(KP250000000, T3m, T3l);
|
Chris@10
|
684 T3f = T1k - T13;
|
Chris@10
|
685 T3g = Tu - TL;
|
Chris@10
|
686 T3h = FNMS(KP951056516, T3g, KP587785252 * T3f);
|
Chris@10
|
687 T3p = FMA(KP587785252, T3g, KP951056516 * T3f);
|
Chris@10
|
688 ci[WS(rs, 14)] = T3m + T3l;
|
Chris@10
|
689 T3q = T3n - T3k;
|
Chris@10
|
690 cr[WS(rs, 12)] = T3p - T3q;
|
Chris@10
|
691 ci[WS(rs, 11)] = T3p + T3q;
|
Chris@10
|
692 T3o = T3k + T3n;
|
Chris@10
|
693 cr[WS(rs, 9)] = T3h - T3o;
|
Chris@10
|
694 ci[WS(rs, 8)] = T3h + T3o;
|
Chris@10
|
695 }
|
Chris@10
|
696 {
|
Chris@10
|
697 E T2y, T2A, T1r, T1O, T2p, T2q, T2z, T2r;
|
Chris@10
|
698 {
|
Chris@10
|
699 E T2u, T2x, T1C, T1N;
|
Chris@10
|
700 T2u = T2s - T2t;
|
Chris@10
|
701 T2x = T2v - T2w;
|
Chris@10
|
702 T2y = FMA(KP951056516, T2u, KP587785252 * T2x);
|
Chris@10
|
703 T2A = FNMS(KP587785252, T2u, KP951056516 * T2x);
|
Chris@10
|
704 T1r = T1n - T1q;
|
Chris@10
|
705 T1C = T1w + T1B;
|
Chris@10
|
706 T1N = T1H + T1M;
|
Chris@10
|
707 T1O = T1C + T1N;
|
Chris@10
|
708 T2p = KP559016994 * (T1C - T1N);
|
Chris@10
|
709 T2q = FNMS(KP250000000, T1O, T1r);
|
Chris@10
|
710 }
|
Chris@10
|
711 cr[WS(rs, 5)] = T1r + T1O;
|
Chris@10
|
712 T2z = T2q - T2p;
|
Chris@10
|
713 cr[WS(rs, 2)] = T2z - T2A;
|
Chris@10
|
714 ci[WS(rs, 6)] = T2z + T2A;
|
Chris@10
|
715 T2r = T2p + T2q;
|
Chris@10
|
716 ci[0] = T2r - T2y;
|
Chris@10
|
717 ci[WS(rs, 3)] = T2r + T2y;
|
Chris@10
|
718 }
|
Chris@10
|
719 {
|
Chris@10
|
720 E T35, T3d, T39, T3a, T38, T3b, T3e, T3c;
|
Chris@10
|
721 {
|
Chris@10
|
722 E T33, T34, T36, T37;
|
Chris@10
|
723 T33 = T1w - T1B;
|
Chris@10
|
724 T34 = T1H - T1M;
|
Chris@10
|
725 T35 = FMA(KP951056516, T33, KP587785252 * T34);
|
Chris@10
|
726 T3d = FNMS(KP587785252, T33, KP951056516 * T34);
|
Chris@10
|
727 T39 = T2T - T2Q;
|
Chris@10
|
728 T36 = T2v + T2w;
|
Chris@10
|
729 T37 = T2s + T2t;
|
Chris@10
|
730 T3a = T37 + T36;
|
Chris@10
|
731 T38 = KP559016994 * (T36 - T37);
|
Chris@10
|
732 T3b = FNMS(KP250000000, T3a, T39);
|
Chris@10
|
733 }
|
Chris@10
|
734 ci[WS(rs, 9)] = T3a + T39;
|
Chris@10
|
735 T3e = T38 + T3b;
|
Chris@10
|
736 cr[WS(rs, 8)] = T3d - T3e;
|
Chris@10
|
737 ci[WS(rs, 12)] = T3d + T3e;
|
Chris@10
|
738 T3c = T38 - T3b;
|
Chris@10
|
739 cr[WS(rs, 11)] = T35 + T3c;
|
Chris@10
|
740 cr[WS(rs, 14)] = T3c - T35;
|
Chris@10
|
741 }
|
Chris@10
|
742 {
|
Chris@10
|
743 E T2X, T31, T2U, T2P, T2Y, T2Z, T32, T30;
|
Chris@10
|
744 {
|
Chris@10
|
745 E T2V, T2W, T2N, T2O;
|
Chris@10
|
746 T2V = T1T - T1U;
|
Chris@10
|
747 T2W = T1Q - T1R;
|
Chris@10
|
748 T2X = FNMS(KP587785252, T2W, KP951056516 * T2V);
|
Chris@10
|
749 T31 = FMA(KP951056516, T2W, KP587785252 * T2V);
|
Chris@10
|
750 T2U = T2Q + T2T;
|
Chris@10
|
751 T2N = T2k - T2f;
|
Chris@10
|
752 T2O = T24 + T29;
|
Chris@10
|
753 T2P = T2N - T2O;
|
Chris@10
|
754 T2Y = FMA(KP250000000, T2P, T2U);
|
Chris@10
|
755 T2Z = KP559016994 * (T2O + T2N);
|
Chris@10
|
756 }
|
Chris@10
|
757 cr[WS(rs, 10)] = T2P - T2U;
|
Chris@10
|
758 T32 = T2Z + T2Y;
|
Chris@10
|
759 ci[WS(rs, 10)] = T31 + T32;
|
Chris@10
|
760 ci[WS(rs, 13)] = T32 - T31;
|
Chris@10
|
761 T30 = T2Y - T2Z;
|
Chris@10
|
762 cr[WS(rs, 13)] = T2X - T30;
|
Chris@10
|
763 ci[WS(rs, 7)] = T2X + T30;
|
Chris@10
|
764 }
|
Chris@10
|
765 {
|
Chris@10
|
766 E T2m, T2o, T1P, T1W, T1X, T1Y, T1Z, T2n;
|
Chris@10
|
767 {
|
Chris@10
|
768 E T2a, T2l, T1S, T1V;
|
Chris@10
|
769 T2a = T24 - T29;
|
Chris@10
|
770 T2l = T2f + T2k;
|
Chris@10
|
771 T2m = FMA(KP951056516, T2a, KP587785252 * T2l);
|
Chris@10
|
772 T2o = FNMS(KP587785252, T2a, KP951056516 * T2l);
|
Chris@10
|
773 T1P = T1n + T1q;
|
Chris@10
|
774 T1S = T1Q + T1R;
|
Chris@10
|
775 T1V = T1T + T1U;
|
Chris@10
|
776 T1W = T1S + T1V;
|
Chris@10
|
777 T1X = KP559016994 * (T1S - T1V);
|
Chris@10
|
778 T1Y = FNMS(KP250000000, T1W, T1P);
|
Chris@10
|
779 }
|
Chris@10
|
780 ci[WS(rs, 4)] = T1P + T1W;
|
Chris@10
|
781 T1Z = T1X + T1Y;
|
Chris@10
|
782 cr[WS(rs, 4)] = T1Z - T2m;
|
Chris@10
|
783 cr[WS(rs, 1)] = T1Z + T2m;
|
Chris@10
|
784 T2n = T1Y - T1X;
|
Chris@10
|
785 cr[WS(rs, 7)] = T2n - T2o;
|
Chris@10
|
786 ci[WS(rs, 1)] = T2n + T2o;
|
Chris@10
|
787 }
|
Chris@10
|
788 }
|
Chris@10
|
789 }
|
Chris@10
|
790 }
|
Chris@10
|
791
|
Chris@10
|
792 static const tw_instr twinstr[] = {
|
Chris@10
|
793 {TW_FULL, 1, 15},
|
Chris@10
|
794 {TW_NEXT, 1, 0}
|
Chris@10
|
795 };
|
Chris@10
|
796
|
Chris@10
|
797 static const hc2hc_desc desc = { 15, "hf_15", twinstr, &GENUS, {128, 56, 56, 0} };
|
Chris@10
|
798
|
Chris@10
|
799 void X(codelet_hf_15) (planner *p) {
|
Chris@10
|
800 X(khc2hc_register) (p, hf_15, &desc);
|
Chris@10
|
801 }
|
Chris@10
|
802 #endif /* HAVE_FMA */
|