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 07:41:14 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 -sign 1 -n 15 -dif -name hb_15 -include hb.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 * 93 stack variables, 6 constants, and 60 memory accesses
|
Chris@10
|
34 */
|
Chris@10
|
35 #include "hb.h"
|
Chris@10
|
36
|
Chris@10
|
37 static void hb_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, T3u, T3r, T3w, T3t;
|
Chris@10
|
49 {
|
Chris@10
|
50 E T5, T11, T1C, T2U, T2f, T3f, T19, T18, TS, TH, T14, T16, T3g, T3a, Ts;
|
Chris@10
|
51 E Tv, T37, T3h, T28, T2h, T1M, T21, T2g, T3n, T2X, T1P, T30, T3m, T1J, T2m;
|
Chris@10
|
52 {
|
Chris@10
|
53 E T1, TX, T2, T3, TY, TZ;
|
Chris@10
|
54 T1 = cr[0];
|
Chris@10
|
55 TX = ci[WS(rs, 14)];
|
Chris@10
|
56 T2 = cr[WS(rs, 5)];
|
Chris@10
|
57 T3 = ci[WS(rs, 4)];
|
Chris@10
|
58 TY = ci[WS(rs, 9)];
|
Chris@10
|
59 TZ = cr[WS(rs, 10)];
|
Chris@10
|
60 {
|
Chris@10
|
61 E T1W, T23, T1D, Ta, Tl, T1K, T1Z, T1H, T1G, Tf, TR, T1Y, T26, TI, T1O;
|
Chris@10
|
62 E T1N, Tq, TG, T25, Tx, Ty, Tz, TL, T1E;
|
Chris@10
|
63 {
|
Chris@10
|
64 E Tb, TQ, TN, TO, Te;
|
Chris@10
|
65 {
|
Chris@10
|
66 E T6, Th, Ti, Tj, T9, Tc, Td, Tk;
|
Chris@10
|
67 {
|
Chris@10
|
68 E T7, T8, T2e, T4;
|
Chris@10
|
69 T6 = cr[WS(rs, 3)];
|
Chris@10
|
70 T2e = T2 - T3;
|
Chris@10
|
71 T4 = T2 + T3;
|
Chris@10
|
72 {
|
Chris@10
|
73 E T1B, T10, T1A, T2d;
|
Chris@10
|
74 T1B = TY + TZ;
|
Chris@10
|
75 T10 = TY - TZ;
|
Chris@10
|
76 T7 = ci[WS(rs, 6)];
|
Chris@10
|
77 T5 = T1 + T4;
|
Chris@10
|
78 T1A = FNMS(KP500000000, T4, T1);
|
Chris@10
|
79 T11 = TX + T10;
|
Chris@10
|
80 T2d = FNMS(KP500000000, T10, TX);
|
Chris@10
|
81 T1C = FNMS(KP866025403, T1B, T1A);
|
Chris@10
|
82 T2U = FMA(KP866025403, T1B, T1A);
|
Chris@10
|
83 T2f = FMA(KP866025403, T2e, T2d);
|
Chris@10
|
84 T3f = FNMS(KP866025403, T2e, T2d);
|
Chris@10
|
85 T8 = ci[WS(rs, 1)];
|
Chris@10
|
86 }
|
Chris@10
|
87 Th = cr[WS(rs, 6)];
|
Chris@10
|
88 Ti = ci[WS(rs, 3)];
|
Chris@10
|
89 Tj = cr[WS(rs, 1)];
|
Chris@10
|
90 T9 = T7 + T8;
|
Chris@10
|
91 T1W = T7 - T8;
|
Chris@10
|
92 }
|
Chris@10
|
93 Tb = ci[WS(rs, 2)];
|
Chris@10
|
94 T23 = Ti - Tj;
|
Chris@10
|
95 Tk = Ti + Tj;
|
Chris@10
|
96 T1D = FNMS(KP500000000, T9, T6);
|
Chris@10
|
97 Ta = T6 + T9;
|
Chris@10
|
98 Tc = cr[WS(rs, 2)];
|
Chris@10
|
99 Tl = Th + Tk;
|
Chris@10
|
100 T1K = FNMS(KP500000000, Tk, Th);
|
Chris@10
|
101 Td = cr[WS(rs, 7)];
|
Chris@10
|
102 TQ = cr[WS(rs, 12)];
|
Chris@10
|
103 TN = ci[WS(rs, 12)];
|
Chris@10
|
104 TO = ci[WS(rs, 7)];
|
Chris@10
|
105 Te = Tc + Td;
|
Chris@10
|
106 T1Z = Tc - Td;
|
Chris@10
|
107 }
|
Chris@10
|
108 {
|
Chris@10
|
109 E Tm, TF, TC, TD, Tp, Tn, To, TP, TJ, TK, TE;
|
Chris@10
|
110 Tm = ci[WS(rs, 5)];
|
Chris@10
|
111 T1H = TO - TN;
|
Chris@10
|
112 TP = TN + TO;
|
Chris@10
|
113 T1G = FNMS(KP500000000, Te, Tb);
|
Chris@10
|
114 Tf = Tb + Te;
|
Chris@10
|
115 Tn = ci[0];
|
Chris@10
|
116 TR = TP - TQ;
|
Chris@10
|
117 T1Y = FMA(KP500000000, TP, TQ);
|
Chris@10
|
118 To = cr[WS(rs, 4)];
|
Chris@10
|
119 TF = cr[WS(rs, 9)];
|
Chris@10
|
120 TC = ci[WS(rs, 10)];
|
Chris@10
|
121 TD = cr[WS(rs, 14)];
|
Chris@10
|
122 Tp = Tn + To;
|
Chris@10
|
123 T26 = Tn - To;
|
Chris@10
|
124 TI = ci[WS(rs, 11)];
|
Chris@10
|
125 T1O = TC + TD;
|
Chris@10
|
126 TE = TC - TD;
|
Chris@10
|
127 T1N = FNMS(KP500000000, Tp, Tm);
|
Chris@10
|
128 Tq = Tm + Tp;
|
Chris@10
|
129 TJ = cr[WS(rs, 8)];
|
Chris@10
|
130 TG = TE - TF;
|
Chris@10
|
131 T25 = FMA(KP500000000, TE, TF);
|
Chris@10
|
132 TK = cr[WS(rs, 13)];
|
Chris@10
|
133 Tx = ci[WS(rs, 8)];
|
Chris@10
|
134 Ty = ci[WS(rs, 13)];
|
Chris@10
|
135 Tz = cr[WS(rs, 11)];
|
Chris@10
|
136 TL = TJ + TK;
|
Chris@10
|
137 T1E = TJ - TK;
|
Chris@10
|
138 }
|
Chris@10
|
139 }
|
Chris@10
|
140 {
|
Chris@10
|
141 E Tg, T1L, Tr, T22, T12, T1X, T38, T13, T39, T20;
|
Chris@10
|
142 {
|
Chris@10
|
143 E TA, T1V, TM, TB;
|
Chris@10
|
144 Tg = Ta + Tf;
|
Chris@10
|
145 T19 = Ta - Tf;
|
Chris@10
|
146 T1L = Ty + Tz;
|
Chris@10
|
147 TA = Ty - Tz;
|
Chris@10
|
148 T1V = FMA(KP500000000, TL, TI);
|
Chris@10
|
149 TM = TI - TL;
|
Chris@10
|
150 T18 = Tl - Tq;
|
Chris@10
|
151 Tr = Tl + Tq;
|
Chris@10
|
152 TB = Tx + TA;
|
Chris@10
|
153 T22 = FNMS(KP500000000, TA, Tx);
|
Chris@10
|
154 T12 = TM + TR;
|
Chris@10
|
155 TS = TM - TR;
|
Chris@10
|
156 T1X = FMA(KP866025403, T1W, T1V);
|
Chris@10
|
157 T38 = FNMS(KP866025403, T1W, T1V);
|
Chris@10
|
158 T13 = TB + TG;
|
Chris@10
|
159 TH = TB - TG;
|
Chris@10
|
160 T39 = FMA(KP866025403, T1Z, T1Y);
|
Chris@10
|
161 T20 = FNMS(KP866025403, T1Z, T1Y);
|
Chris@10
|
162 }
|
Chris@10
|
163 {
|
Chris@10
|
164 E T35, T24, T27, T36;
|
Chris@10
|
165 T14 = T12 + T13;
|
Chris@10
|
166 T16 = T12 - T13;
|
Chris@10
|
167 T3g = T38 - T39;
|
Chris@10
|
168 T3a = T38 + T39;
|
Chris@10
|
169 T35 = FNMS(KP866025403, T23, T22);
|
Chris@10
|
170 T24 = FMA(KP866025403, T23, T22);
|
Chris@10
|
171 Ts = Tg + Tr;
|
Chris@10
|
172 Tv = Tg - Tr;
|
Chris@10
|
173 T27 = FNMS(KP866025403, T26, T25);
|
Chris@10
|
174 T36 = FMA(KP866025403, T26, T25);
|
Chris@10
|
175 T37 = T35 + T36;
|
Chris@10
|
176 T3h = T35 - T36;
|
Chris@10
|
177 T28 = T24 + T27;
|
Chris@10
|
178 T2h = T24 - T27;
|
Chris@10
|
179 {
|
Chris@10
|
180 E T1F, T1I, T2Y, T2Z, T2V, T2W;
|
Chris@10
|
181 T2V = FNMS(KP866025403, T1E, T1D);
|
Chris@10
|
182 T1F = FMA(KP866025403, T1E, T1D);
|
Chris@10
|
183 T1I = FMA(KP866025403, T1H, T1G);
|
Chris@10
|
184 T2W = FNMS(KP866025403, T1H, T1G);
|
Chris@10
|
185 T2Y = FNMS(KP866025403, T1L, T1K);
|
Chris@10
|
186 T1M = FMA(KP866025403, T1L, T1K);
|
Chris@10
|
187 T21 = T1X + T20;
|
Chris@10
|
188 T2g = T1X - T20;
|
Chris@10
|
189 T3n = T2V - T2W;
|
Chris@10
|
190 T2X = T2V + T2W;
|
Chris@10
|
191 T2Z = FNMS(KP866025403, T1O, T1N);
|
Chris@10
|
192 T1P = FMA(KP866025403, T1O, T1N);
|
Chris@10
|
193 T30 = T2Y + T2Z;
|
Chris@10
|
194 T3m = T2Y - T2Z;
|
Chris@10
|
195 T1J = T1F + T1I;
|
Chris@10
|
196 T2m = T1F - T1I;
|
Chris@10
|
197 }
|
Chris@10
|
198 }
|
Chris@10
|
199 }
|
Chris@10
|
200 }
|
Chris@10
|
201 }
|
Chris@10
|
202 {
|
Chris@10
|
203 E T31, T33, T2n, T1Q;
|
Chris@10
|
204 cr[0] = T5 + Ts;
|
Chris@10
|
205 T31 = T2X + T30;
|
Chris@10
|
206 T33 = T2X - T30;
|
Chris@10
|
207 T2n = T1M - T1P;
|
Chris@10
|
208 T1Q = T1M + T1P;
|
Chris@10
|
209 ci[0] = T11 + T14;
|
Chris@10
|
210 {
|
Chris@10
|
211 E T1T, T1R, T1r, T1o, T1n;
|
Chris@10
|
212 {
|
Chris@10
|
213 E T1q, T1a, TT, T1l, Tu, T17, T1p, T15;
|
Chris@10
|
214 T1q = FMA(KP618033988, T18, T19);
|
Chris@10
|
215 T1a = FNMS(KP618033988, T19, T18);
|
Chris@10
|
216 T1T = T1J - T1Q;
|
Chris@10
|
217 T1R = T1J + T1Q;
|
Chris@10
|
218 T15 = FNMS(KP250000000, T14, T11);
|
Chris@10
|
219 TT = FNMS(KP618033988, TS, TH);
|
Chris@10
|
220 T1l = FMA(KP618033988, TH, TS);
|
Chris@10
|
221 Tu = FNMS(KP250000000, Ts, T5);
|
Chris@10
|
222 T17 = FNMS(KP559016994, T16, T15);
|
Chris@10
|
223 T1p = FMA(KP559016994, T16, T15);
|
Chris@10
|
224 {
|
Chris@10
|
225 E T1h, T1m, T1e, T1x, T1w, T1v, T1g, T1d;
|
Chris@10
|
226 {
|
Chris@10
|
227 E TW, T1b, Tt, T1u, TU, T1k, Tw;
|
Chris@10
|
228 TW = W[5];
|
Chris@10
|
229 T1k = FMA(KP559016994, Tv, Tu);
|
Chris@10
|
230 Tw = FNMS(KP559016994, Tv, Tu);
|
Chris@10
|
231 T1b = FMA(KP951056516, T1a, T17);
|
Chris@10
|
232 T1h = FNMS(KP951056516, T1a, T17);
|
Chris@10
|
233 Tt = W[4];
|
Chris@10
|
234 T1m = FNMS(KP951056516, T1l, T1k);
|
Chris@10
|
235 T1u = FMA(KP951056516, T1l, T1k);
|
Chris@10
|
236 T1e = FMA(KP951056516, TT, Tw);
|
Chris@10
|
237 TU = FNMS(KP951056516, TT, Tw);
|
Chris@10
|
238 {
|
Chris@10
|
239 E T1t, TV, T1c, T1y;
|
Chris@10
|
240 T1x = FNMS(KP951056516, T1q, T1p);
|
Chris@10
|
241 T1r = FMA(KP951056516, T1q, T1p);
|
Chris@10
|
242 T1w = W[17];
|
Chris@10
|
243 T1t = W[16];
|
Chris@10
|
244 TV = Tt * TU;
|
Chris@10
|
245 T1c = TW * TU;
|
Chris@10
|
246 T1y = T1w * T1u;
|
Chris@10
|
247 T1v = T1t * T1u;
|
Chris@10
|
248 cr[WS(rs, 3)] = FNMS(TW, T1b, TV);
|
Chris@10
|
249 ci[WS(rs, 3)] = FMA(Tt, T1b, T1c);
|
Chris@10
|
250 ci[WS(rs, 9)] = FMA(T1t, T1x, T1y);
|
Chris@10
|
251 }
|
Chris@10
|
252 }
|
Chris@10
|
253 cr[WS(rs, 9)] = FNMS(T1w, T1x, T1v);
|
Chris@10
|
254 T1g = W[23];
|
Chris@10
|
255 T1d = W[22];
|
Chris@10
|
256 {
|
Chris@10
|
257 E T1j, T1s, T1i, T1f;
|
Chris@10
|
258 T1o = W[11];
|
Chris@10
|
259 T1i = T1g * T1e;
|
Chris@10
|
260 T1f = T1d * T1e;
|
Chris@10
|
261 T1j = W[10];
|
Chris@10
|
262 T1s = T1o * T1m;
|
Chris@10
|
263 ci[WS(rs, 12)] = FMA(T1d, T1h, T1i);
|
Chris@10
|
264 cr[WS(rs, 12)] = FNMS(T1g, T1h, T1f);
|
Chris@10
|
265 T1n = T1j * T1m;
|
Chris@10
|
266 ci[WS(rs, 6)] = FMA(T1j, T1r, T1s);
|
Chris@10
|
267 }
|
Chris@10
|
268 }
|
Chris@10
|
269 }
|
Chris@10
|
270 {
|
Chris@10
|
271 E T2v, T2u, T2r, T2w, T2t;
|
Chris@10
|
272 {
|
Chris@10
|
273 E T1S, T2N, T2o, T2E, T2Q, T2P, T2k, T2S, T29, T2z, T2R, T2j, T2O, T2i;
|
Chris@10
|
274 cr[WS(rs, 6)] = FNMS(T1o, T1r, T1n);
|
Chris@10
|
275 T1S = FNMS(KP250000000, T1R, T1C);
|
Chris@10
|
276 T2O = T1C + T1R;
|
Chris@10
|
277 T2N = W[18];
|
Chris@10
|
278 T2o = FMA(KP618033988, T2n, T2m);
|
Chris@10
|
279 T2E = FNMS(KP618033988, T2m, T2n);
|
Chris@10
|
280 T2Q = W[19];
|
Chris@10
|
281 T2P = T2N * T2O;
|
Chris@10
|
282 T2i = T2g + T2h;
|
Chris@10
|
283 T2k = T2g - T2h;
|
Chris@10
|
284 T2S = T2Q * T2O;
|
Chris@10
|
285 T29 = FMA(KP618033988, T28, T21);
|
Chris@10
|
286 T2z = FNMS(KP618033988, T21, T28);
|
Chris@10
|
287 T2R = T2f + T2i;
|
Chris@10
|
288 T2j = FNMS(KP250000000, T2i, T2f);
|
Chris@10
|
289 {
|
Chris@10
|
290 E T2D, T2p, T2I, T2A, T2a, T2s, T2c, T1z, T2l, T1U, T2y;
|
Chris@10
|
291 cr[WS(rs, 10)] = FNMS(T2Q, T2R, T2P);
|
Chris@10
|
292 T2l = FMA(KP559016994, T2k, T2j);
|
Chris@10
|
293 T2D = FNMS(KP559016994, T2k, T2j);
|
Chris@10
|
294 T1U = FMA(KP559016994, T1T, T1S);
|
Chris@10
|
295 T2y = FNMS(KP559016994, T1T, T1S);
|
Chris@10
|
296 ci[WS(rs, 10)] = FMA(T2N, T2R, T2S);
|
Chris@10
|
297 T2p = FMA(KP951056516, T2o, T2l);
|
Chris@10
|
298 T2v = FNMS(KP951056516, T2o, T2l);
|
Chris@10
|
299 T2I = FNMS(KP951056516, T2z, T2y);
|
Chris@10
|
300 T2A = FMA(KP951056516, T2z, T2y);
|
Chris@10
|
301 T2a = FNMS(KP951056516, T29, T1U);
|
Chris@10
|
302 T2s = FMA(KP951056516, T29, T1U);
|
Chris@10
|
303 T2c = W[1];
|
Chris@10
|
304 T1z = W[0];
|
Chris@10
|
305 {
|
Chris@10
|
306 E T2F, T2L, T2K, T2J;
|
Chris@10
|
307 {
|
Chris@10
|
308 E T2H, T2M, T2q, T2b;
|
Chris@10
|
309 T2F = FNMS(KP951056516, T2E, T2D);
|
Chris@10
|
310 T2L = FMA(KP951056516, T2E, T2D);
|
Chris@10
|
311 T2K = W[25];
|
Chris@10
|
312 T2q = T2c * T2a;
|
Chris@10
|
313 T2b = T1z * T2a;
|
Chris@10
|
314 T2H = W[24];
|
Chris@10
|
315 T2M = T2K * T2I;
|
Chris@10
|
316 ci[WS(rs, 1)] = FMA(T1z, T2p, T2q);
|
Chris@10
|
317 cr[WS(rs, 1)] = FNMS(T2c, T2p, T2b);
|
Chris@10
|
318 T2J = T2H * T2I;
|
Chris@10
|
319 ci[WS(rs, 13)] = FMA(T2H, T2L, T2M);
|
Chris@10
|
320 }
|
Chris@10
|
321 {
|
Chris@10
|
322 E T2C, T2x, T2G, T2B;
|
Chris@10
|
323 T2C = W[13];
|
Chris@10
|
324 cr[WS(rs, 13)] = FNMS(T2K, T2L, T2J);
|
Chris@10
|
325 T2x = W[12];
|
Chris@10
|
326 T2G = T2C * T2A;
|
Chris@10
|
327 T2u = W[7];
|
Chris@10
|
328 T2B = T2x * T2A;
|
Chris@10
|
329 T2r = W[6];
|
Chris@10
|
330 ci[WS(rs, 7)] = FMA(T2x, T2F, T2G);
|
Chris@10
|
331 T2w = T2u * T2s;
|
Chris@10
|
332 cr[WS(rs, 7)] = FNMS(T2C, T2F, T2B);
|
Chris@10
|
333 T2t = T2r * T2s;
|
Chris@10
|
334 }
|
Chris@10
|
335 }
|
Chris@10
|
336 }
|
Chris@10
|
337 }
|
Chris@10
|
338 {
|
Chris@10
|
339 E T32, T3N, T3E, T3o, T3Q, T3P, T3k, T3S, T3z, T3b, T3j, T3R, T3O, T3i;
|
Chris@10
|
340 ci[WS(rs, 4)] = FMA(T2r, T2v, T2w);
|
Chris@10
|
341 cr[WS(rs, 4)] = FNMS(T2u, T2v, T2t);
|
Chris@10
|
342 T3O = T2U + T31;
|
Chris@10
|
343 T32 = FNMS(KP250000000, T31, T2U);
|
Chris@10
|
344 T3N = W[8];
|
Chris@10
|
345 T3E = FMA(KP618033988, T3m, T3n);
|
Chris@10
|
346 T3o = FNMS(KP618033988, T3n, T3m);
|
Chris@10
|
347 T3Q = W[9];
|
Chris@10
|
348 T3P = T3N * T3O;
|
Chris@10
|
349 T3k = T3g - T3h;
|
Chris@10
|
350 T3i = T3g + T3h;
|
Chris@10
|
351 T3S = T3Q * T3O;
|
Chris@10
|
352 T3z = FMA(KP618033988, T37, T3a);
|
Chris@10
|
353 T3b = FNMS(KP618033988, T3a, T37);
|
Chris@10
|
354 T3j = FNMS(KP250000000, T3i, T3f);
|
Chris@10
|
355 T3R = T3f + T3i;
|
Chris@10
|
356 {
|
Chris@10
|
357 E T3D, T3p, T3A, T3I, T3s, T3c, T3e, T2T, T3l, T3y, T34;
|
Chris@10
|
358 cr[WS(rs, 5)] = FNMS(T3Q, T3R, T3P);
|
Chris@10
|
359 T3D = FMA(KP559016994, T3k, T3j);
|
Chris@10
|
360 T3l = FNMS(KP559016994, T3k, T3j);
|
Chris@10
|
361 T3y = FMA(KP559016994, T33, T32);
|
Chris@10
|
362 T34 = FNMS(KP559016994, T33, T32);
|
Chris@10
|
363 ci[WS(rs, 5)] = FMA(T3N, T3R, T3S);
|
Chris@10
|
364 T3v = FMA(KP951056516, T3o, T3l);
|
Chris@10
|
365 T3p = FNMS(KP951056516, T3o, T3l);
|
Chris@10
|
366 T3A = FNMS(KP951056516, T3z, T3y);
|
Chris@10
|
367 T3I = FMA(KP951056516, T3z, T3y);
|
Chris@10
|
368 T3s = FNMS(KP951056516, T3b, T34);
|
Chris@10
|
369 T3c = FMA(KP951056516, T3b, T34);
|
Chris@10
|
370 T3e = W[3];
|
Chris@10
|
371 T2T = W[2];
|
Chris@10
|
372 {
|
Chris@10
|
373 E T3L, T3F, T3K, T3J;
|
Chris@10
|
374 {
|
Chris@10
|
375 E T3H, T3M, T3q, T3d;
|
Chris@10
|
376 T3L = FNMS(KP951056516, T3E, T3D);
|
Chris@10
|
377 T3F = FMA(KP951056516, T3E, T3D);
|
Chris@10
|
378 T3K = W[27];
|
Chris@10
|
379 T3q = T3e * T3c;
|
Chris@10
|
380 T3d = T2T * T3c;
|
Chris@10
|
381 T3H = W[26];
|
Chris@10
|
382 T3M = T3K * T3I;
|
Chris@10
|
383 ci[WS(rs, 2)] = FMA(T2T, T3p, T3q);
|
Chris@10
|
384 cr[WS(rs, 2)] = FNMS(T3e, T3p, T3d);
|
Chris@10
|
385 T3J = T3H * T3I;
|
Chris@10
|
386 ci[WS(rs, 14)] = FMA(T3H, T3L, T3M);
|
Chris@10
|
387 }
|
Chris@10
|
388 {
|
Chris@10
|
389 E T3C, T3x, T3G, T3B;
|
Chris@10
|
390 T3C = W[21];
|
Chris@10
|
391 cr[WS(rs, 14)] = FNMS(T3K, T3L, T3J);
|
Chris@10
|
392 T3x = W[20];
|
Chris@10
|
393 T3G = T3C * T3A;
|
Chris@10
|
394 T3u = W[15];
|
Chris@10
|
395 T3B = T3x * T3A;
|
Chris@10
|
396 T3r = W[14];
|
Chris@10
|
397 ci[WS(rs, 11)] = FMA(T3x, T3F, T3G);
|
Chris@10
|
398 T3w = T3u * T3s;
|
Chris@10
|
399 cr[WS(rs, 11)] = FNMS(T3C, T3F, T3B);
|
Chris@10
|
400 T3t = T3r * T3s;
|
Chris@10
|
401 }
|
Chris@10
|
402 }
|
Chris@10
|
403 }
|
Chris@10
|
404 }
|
Chris@10
|
405 }
|
Chris@10
|
406 }
|
Chris@10
|
407 }
|
Chris@10
|
408 }
|
Chris@10
|
409 ci[WS(rs, 8)] = FMA(T3r, T3v, T3w);
|
Chris@10
|
410 cr[WS(rs, 8)] = FNMS(T3u, T3v, T3t);
|
Chris@10
|
411 }
|
Chris@10
|
412 }
|
Chris@10
|
413 }
|
Chris@10
|
414
|
Chris@10
|
415 static const tw_instr twinstr[] = {
|
Chris@10
|
416 {TW_FULL, 1, 15},
|
Chris@10
|
417 {TW_NEXT, 1, 0}
|
Chris@10
|
418 };
|
Chris@10
|
419
|
Chris@10
|
420 static const hc2hc_desc desc = { 15, "hb_15", twinstr, &GENUS, {72, 28, 112, 0} };
|
Chris@10
|
421
|
Chris@10
|
422 void X(codelet_hb_15) (planner *p) {
|
Chris@10
|
423 X(khc2hc_register) (p, hb_15, &desc);
|
Chris@10
|
424 }
|
Chris@10
|
425 #else /* HAVE_FMA */
|
Chris@10
|
426
|
Chris@10
|
427 /* Generated by: ../../../genfft/gen_hc2hc.native -compact -variables 4 -pipeline-latency 4 -sign 1 -n 15 -dif -name hb_15 -include hb.h */
|
Chris@10
|
428
|
Chris@10
|
429 /*
|
Chris@10
|
430 * This function contains 184 FP additions, 112 FP multiplications,
|
Chris@10
|
431 * (or, 128 additions, 56 multiplications, 56 fused multiply/add),
|
Chris@10
|
432 * 75 stack variables, 6 constants, and 60 memory accesses
|
Chris@10
|
433 */
|
Chris@10
|
434 #include "hb.h"
|
Chris@10
|
435
|
Chris@10
|
436 static void hb_15(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
|
Chris@10
|
437 {
|
Chris@10
|
438 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
|
Chris@10
|
439 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
|
Chris@10
|
440 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
|
Chris@10
|
441 DK(KP587785252, +0.587785252292473129168705954639072768597652438);
|
Chris@10
|
442 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
|
Chris@10
|
443 DK(KP866025403, +0.866025403784438646763723170752936183471402627);
|
Chris@10
|
444 {
|
Chris@10
|
445 INT m;
|
Chris@10
|
446 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
|
447 E T5, T10, T1J, T2C, T2c, T2M, TH, T18, T17, TS, T2Q, T2R, T2S, Tg, Tr;
|
Chris@10
|
448 E Ts, T11, T12, T13, T2N, T2O, T2P, T1u, T1x, T1y, T1W, T1Z, T28, T1P, T1S;
|
Chris@10
|
449 E T27, T1B, T1E, T1F, T2G, T2H, T2I, T2D, T2E, T2F;
|
Chris@10
|
450 {
|
Chris@10
|
451 E T1, TW, T4, T2a, TZ, T1I, T1H, T2b;
|
Chris@10
|
452 T1 = cr[0];
|
Chris@10
|
453 TW = ci[WS(rs, 14)];
|
Chris@10
|
454 {
|
Chris@10
|
455 E T2, T3, TX, TY;
|
Chris@10
|
456 T2 = cr[WS(rs, 5)];
|
Chris@10
|
457 T3 = ci[WS(rs, 4)];
|
Chris@10
|
458 T4 = T2 + T3;
|
Chris@10
|
459 T2a = KP866025403 * (T2 - T3);
|
Chris@10
|
460 TX = ci[WS(rs, 9)];
|
Chris@10
|
461 TY = cr[WS(rs, 10)];
|
Chris@10
|
462 TZ = TX - TY;
|
Chris@10
|
463 T1I = KP866025403 * (TX + TY);
|
Chris@10
|
464 }
|
Chris@10
|
465 T5 = T1 + T4;
|
Chris@10
|
466 T10 = TW + TZ;
|
Chris@10
|
467 T1H = FNMS(KP500000000, T4, T1);
|
Chris@10
|
468 T1J = T1H - T1I;
|
Chris@10
|
469 T2C = T1H + T1I;
|
Chris@10
|
470 T2b = FNMS(KP500000000, TZ, TW);
|
Chris@10
|
471 T2c = T2a + T2b;
|
Chris@10
|
472 T2M = T2b - T2a;
|
Chris@10
|
473 }
|
Chris@10
|
474 {
|
Chris@10
|
475 E Ta, T1N, T1s, Tl, T1U, T1z, Tf, T1Q, T1v, TG, T1R, T1w, Tq, T1X, T1C;
|
Chris@10
|
476 E TM, T1V, T1A, TB, T1O, T1t, TR, T1Y, T1D;
|
Chris@10
|
477 {
|
Chris@10
|
478 E T6, T7, T8, T9;
|
Chris@10
|
479 T6 = cr[WS(rs, 3)];
|
Chris@10
|
480 T7 = ci[WS(rs, 6)];
|
Chris@10
|
481 T8 = ci[WS(rs, 1)];
|
Chris@10
|
482 T9 = T7 + T8;
|
Chris@10
|
483 Ta = T6 + T9;
|
Chris@10
|
484 T1N = KP866025403 * (T7 - T8);
|
Chris@10
|
485 T1s = FNMS(KP500000000, T9, T6);
|
Chris@10
|
486 }
|
Chris@10
|
487 {
|
Chris@10
|
488 E Th, Ti, Tj, Tk;
|
Chris@10
|
489 Th = cr[WS(rs, 6)];
|
Chris@10
|
490 Ti = ci[WS(rs, 3)];
|
Chris@10
|
491 Tj = cr[WS(rs, 1)];
|
Chris@10
|
492 Tk = Ti + Tj;
|
Chris@10
|
493 Tl = Th + Tk;
|
Chris@10
|
494 T1U = KP866025403 * (Ti - Tj);
|
Chris@10
|
495 T1z = FNMS(KP500000000, Tk, Th);
|
Chris@10
|
496 }
|
Chris@10
|
497 {
|
Chris@10
|
498 E Tb, Tc, Td, Te;
|
Chris@10
|
499 Tb = ci[WS(rs, 2)];
|
Chris@10
|
500 Tc = cr[WS(rs, 2)];
|
Chris@10
|
501 Td = cr[WS(rs, 7)];
|
Chris@10
|
502 Te = Tc + Td;
|
Chris@10
|
503 Tf = Tb + Te;
|
Chris@10
|
504 T1Q = KP866025403 * (Tc - Td);
|
Chris@10
|
505 T1v = FNMS(KP500000000, Te, Tb);
|
Chris@10
|
506 }
|
Chris@10
|
507 {
|
Chris@10
|
508 E TF, TC, TD, TE;
|
Chris@10
|
509 TF = cr[WS(rs, 12)];
|
Chris@10
|
510 TC = ci[WS(rs, 12)];
|
Chris@10
|
511 TD = ci[WS(rs, 7)];
|
Chris@10
|
512 TE = TC + TD;
|
Chris@10
|
513 TG = TE - TF;
|
Chris@10
|
514 T1R = FMA(KP500000000, TE, TF);
|
Chris@10
|
515 T1w = KP866025403 * (TD - TC);
|
Chris@10
|
516 }
|
Chris@10
|
517 {
|
Chris@10
|
518 E Tm, Tn, To, Tp;
|
Chris@10
|
519 Tm = ci[WS(rs, 5)];
|
Chris@10
|
520 Tn = ci[0];
|
Chris@10
|
521 To = cr[WS(rs, 4)];
|
Chris@10
|
522 Tp = Tn + To;
|
Chris@10
|
523 Tq = Tm + Tp;
|
Chris@10
|
524 T1X = KP866025403 * (Tn - To);
|
Chris@10
|
525 T1C = FNMS(KP500000000, Tp, Tm);
|
Chris@10
|
526 }
|
Chris@10
|
527 {
|
Chris@10
|
528 E TI, TJ, TK, TL;
|
Chris@10
|
529 TI = ci[WS(rs, 8)];
|
Chris@10
|
530 TJ = ci[WS(rs, 13)];
|
Chris@10
|
531 TK = cr[WS(rs, 11)];
|
Chris@10
|
532 TL = TJ - TK;
|
Chris@10
|
533 TM = TI + TL;
|
Chris@10
|
534 T1V = FNMS(KP500000000, TL, TI);
|
Chris@10
|
535 T1A = KP866025403 * (TJ + TK);
|
Chris@10
|
536 }
|
Chris@10
|
537 {
|
Chris@10
|
538 E Tx, Ty, Tz, TA;
|
Chris@10
|
539 Tx = ci[WS(rs, 11)];
|
Chris@10
|
540 Ty = cr[WS(rs, 8)];
|
Chris@10
|
541 Tz = cr[WS(rs, 13)];
|
Chris@10
|
542 TA = Ty + Tz;
|
Chris@10
|
543 TB = Tx - TA;
|
Chris@10
|
544 T1O = FMA(KP500000000, TA, Tx);
|
Chris@10
|
545 T1t = KP866025403 * (Ty - Tz);
|
Chris@10
|
546 }
|
Chris@10
|
547 {
|
Chris@10
|
548 E TQ, TN, TO, TP;
|
Chris@10
|
549 TQ = cr[WS(rs, 9)];
|
Chris@10
|
550 TN = ci[WS(rs, 10)];
|
Chris@10
|
551 TO = cr[WS(rs, 14)];
|
Chris@10
|
552 TP = TN - TO;
|
Chris@10
|
553 TR = TP - TQ;
|
Chris@10
|
554 T1Y = FMA(KP500000000, TP, TQ);
|
Chris@10
|
555 T1D = KP866025403 * (TN + TO);
|
Chris@10
|
556 }
|
Chris@10
|
557 TH = TB - TG;
|
Chris@10
|
558 T18 = Tl - Tq;
|
Chris@10
|
559 T17 = Ta - Tf;
|
Chris@10
|
560 TS = TM - TR;
|
Chris@10
|
561 T2Q = T1V - T1U;
|
Chris@10
|
562 T2R = T1X + T1Y;
|
Chris@10
|
563 T2S = T2Q - T2R;
|
Chris@10
|
564 Tg = Ta + Tf;
|
Chris@10
|
565 Tr = Tl + Tq;
|
Chris@10
|
566 Ts = Tg + Tr;
|
Chris@10
|
567 T11 = TB + TG;
|
Chris@10
|
568 T12 = TM + TR;
|
Chris@10
|
569 T13 = T11 + T12;
|
Chris@10
|
570 T2N = T1O - T1N;
|
Chris@10
|
571 T2O = T1Q + T1R;
|
Chris@10
|
572 T2P = T2N - T2O;
|
Chris@10
|
573 T1u = T1s + T1t;
|
Chris@10
|
574 T1x = T1v + T1w;
|
Chris@10
|
575 T1y = T1u + T1x;
|
Chris@10
|
576 T1W = T1U + T1V;
|
Chris@10
|
577 T1Z = T1X - T1Y;
|
Chris@10
|
578 T28 = T1W + T1Z;
|
Chris@10
|
579 T1P = T1N + T1O;
|
Chris@10
|
580 T1S = T1Q - T1R;
|
Chris@10
|
581 T27 = T1P + T1S;
|
Chris@10
|
582 T1B = T1z + T1A;
|
Chris@10
|
583 T1E = T1C + T1D;
|
Chris@10
|
584 T1F = T1B + T1E;
|
Chris@10
|
585 T2G = T1z - T1A;
|
Chris@10
|
586 T2H = T1C - T1D;
|
Chris@10
|
587 T2I = T2G + T2H;
|
Chris@10
|
588 T2D = T1s - T1t;
|
Chris@10
|
589 T2E = T1v - T1w;
|
Chris@10
|
590 T2F = T2D + T2E;
|
Chris@10
|
591 }
|
Chris@10
|
592 cr[0] = T5 + Ts;
|
Chris@10
|
593 ci[0] = T10 + T13;
|
Chris@10
|
594 {
|
Chris@10
|
595 E TT, T19, T1k, T1h, T16, T1l, Tw, T1g;
|
Chris@10
|
596 TT = FNMS(KP951056516, TS, KP587785252 * TH);
|
Chris@10
|
597 T19 = FNMS(KP951056516, T18, KP587785252 * T17);
|
Chris@10
|
598 T1k = FMA(KP951056516, T17, KP587785252 * T18);
|
Chris@10
|
599 T1h = FMA(KP951056516, TH, KP587785252 * TS);
|
Chris@10
|
600 {
|
Chris@10
|
601 E T14, T15, Tu, Tv;
|
Chris@10
|
602 T14 = FNMS(KP250000000, T13, T10);
|
Chris@10
|
603 T15 = KP559016994 * (T11 - T12);
|
Chris@10
|
604 T16 = T14 - T15;
|
Chris@10
|
605 T1l = T15 + T14;
|
Chris@10
|
606 Tu = FNMS(KP250000000, Ts, T5);
|
Chris@10
|
607 Tv = KP559016994 * (Tg - Tr);
|
Chris@10
|
608 Tw = Tu - Tv;
|
Chris@10
|
609 T1g = Tv + Tu;
|
Chris@10
|
610 }
|
Chris@10
|
611 {
|
Chris@10
|
612 E TU, T1a, Tt, TV;
|
Chris@10
|
613 TU = Tw + TT;
|
Chris@10
|
614 T1a = T16 - T19;
|
Chris@10
|
615 Tt = W[4];
|
Chris@10
|
616 TV = W[5];
|
Chris@10
|
617 cr[WS(rs, 3)] = FNMS(TV, T1a, Tt * TU);
|
Chris@10
|
618 ci[WS(rs, 3)] = FMA(TV, TU, Tt * T1a);
|
Chris@10
|
619 }
|
Chris@10
|
620 {
|
Chris@10
|
621 E T1o, T1q, T1n, T1p;
|
Chris@10
|
622 T1o = T1g + T1h;
|
Chris@10
|
623 T1q = T1l - T1k;
|
Chris@10
|
624 T1n = W[16];
|
Chris@10
|
625 T1p = W[17];
|
Chris@10
|
626 cr[WS(rs, 9)] = FNMS(T1p, T1q, T1n * T1o);
|
Chris@10
|
627 ci[WS(rs, 9)] = FMA(T1p, T1o, T1n * T1q);
|
Chris@10
|
628 }
|
Chris@10
|
629 {
|
Chris@10
|
630 E T1c, T1e, T1b, T1d;
|
Chris@10
|
631 T1c = Tw - TT;
|
Chris@10
|
632 T1e = T19 + T16;
|
Chris@10
|
633 T1b = W[22];
|
Chris@10
|
634 T1d = W[23];
|
Chris@10
|
635 cr[WS(rs, 12)] = FNMS(T1d, T1e, T1b * T1c);
|
Chris@10
|
636 ci[WS(rs, 12)] = FMA(T1d, T1c, T1b * T1e);
|
Chris@10
|
637 }
|
Chris@10
|
638 {
|
Chris@10
|
639 E T1i, T1m, T1f, T1j;
|
Chris@10
|
640 T1i = T1g - T1h;
|
Chris@10
|
641 T1m = T1k + T1l;
|
Chris@10
|
642 T1f = W[10];
|
Chris@10
|
643 T1j = W[11];
|
Chris@10
|
644 cr[WS(rs, 6)] = FNMS(T1j, T1m, T1f * T1i);
|
Chris@10
|
645 ci[WS(rs, 6)] = FMA(T1j, T1i, T1f * T1m);
|
Chris@10
|
646 }
|
Chris@10
|
647 }
|
Chris@10
|
648 {
|
Chris@10
|
649 E T21, T2n, T26, T2q, T1M, T2y, T2m, T2f, T2A, T2r, T2x, T2z;
|
Chris@10
|
650 {
|
Chris@10
|
651 E T1T, T20, T24, T25;
|
Chris@10
|
652 T1T = T1P - T1S;
|
Chris@10
|
653 T20 = T1W - T1Z;
|
Chris@10
|
654 T21 = FMA(KP951056516, T1T, KP587785252 * T20);
|
Chris@10
|
655 T2n = FNMS(KP951056516, T20, KP587785252 * T1T);
|
Chris@10
|
656 T24 = T1u - T1x;
|
Chris@10
|
657 T25 = T1B - T1E;
|
Chris@10
|
658 T26 = FMA(KP951056516, T24, KP587785252 * T25);
|
Chris@10
|
659 T2q = FNMS(KP951056516, T25, KP587785252 * T24);
|
Chris@10
|
660 }
|
Chris@10
|
661 {
|
Chris@10
|
662 E T1G, T1K, T1L, T29, T2d, T2e;
|
Chris@10
|
663 T1G = KP559016994 * (T1y - T1F);
|
Chris@10
|
664 T1K = T1y + T1F;
|
Chris@10
|
665 T1L = FNMS(KP250000000, T1K, T1J);
|
Chris@10
|
666 T1M = T1G + T1L;
|
Chris@10
|
667 T2y = T1J + T1K;
|
Chris@10
|
668 T2m = T1L - T1G;
|
Chris@10
|
669 T29 = KP559016994 * (T27 - T28);
|
Chris@10
|
670 T2d = T27 + T28;
|
Chris@10
|
671 T2e = FNMS(KP250000000, T2d, T2c);
|
Chris@10
|
672 T2f = T29 + T2e;
|
Chris@10
|
673 T2A = T2c + T2d;
|
Chris@10
|
674 T2r = T2e - T29;
|
Chris@10
|
675 }
|
Chris@10
|
676 T2x = W[18];
|
Chris@10
|
677 T2z = W[19];
|
Chris@10
|
678 cr[WS(rs, 10)] = FNMS(T2z, T2A, T2x * T2y);
|
Chris@10
|
679 ci[WS(rs, 10)] = FMA(T2z, T2y, T2x * T2A);
|
Chris@10
|
680 {
|
Chris@10
|
681 E T2u, T2w, T2t, T2v;
|
Chris@10
|
682 T2u = T2m + T2n;
|
Chris@10
|
683 T2w = T2r - T2q;
|
Chris@10
|
684 T2t = W[24];
|
Chris@10
|
685 T2v = W[25];
|
Chris@10
|
686 cr[WS(rs, 13)] = FNMS(T2v, T2w, T2t * T2u);
|
Chris@10
|
687 ci[WS(rs, 13)] = FMA(T2v, T2u, T2t * T2w);
|
Chris@10
|
688 }
|
Chris@10
|
689 {
|
Chris@10
|
690 E T22, T2g, T1r, T23;
|
Chris@10
|
691 T22 = T1M - T21;
|
Chris@10
|
692 T2g = T26 + T2f;
|
Chris@10
|
693 T1r = W[0];
|
Chris@10
|
694 T23 = W[1];
|
Chris@10
|
695 cr[WS(rs, 1)] = FNMS(T23, T2g, T1r * T22);
|
Chris@10
|
696 ci[WS(rs, 1)] = FMA(T23, T22, T1r * T2g);
|
Chris@10
|
697 }
|
Chris@10
|
698 {
|
Chris@10
|
699 E T2i, T2k, T2h, T2j;
|
Chris@10
|
700 T2i = T1M + T21;
|
Chris@10
|
701 T2k = T2f - T26;
|
Chris@10
|
702 T2h = W[6];
|
Chris@10
|
703 T2j = W[7];
|
Chris@10
|
704 cr[WS(rs, 4)] = FNMS(T2j, T2k, T2h * T2i);
|
Chris@10
|
705 ci[WS(rs, 4)] = FMA(T2j, T2i, T2h * T2k);
|
Chris@10
|
706 }
|
Chris@10
|
707 {
|
Chris@10
|
708 E T2o, T2s, T2l, T2p;
|
Chris@10
|
709 T2o = T2m - T2n;
|
Chris@10
|
710 T2s = T2q + T2r;
|
Chris@10
|
711 T2l = W[12];
|
Chris@10
|
712 T2p = W[13];
|
Chris@10
|
713 cr[WS(rs, 7)] = FNMS(T2p, T2s, T2l * T2o);
|
Chris@10
|
714 ci[WS(rs, 7)] = FMA(T2p, T2o, T2l * T2s);
|
Chris@10
|
715 }
|
Chris@10
|
716 }
|
Chris@10
|
717 {
|
Chris@10
|
718 E T31, T3h, T36, T3k, T2K, T3g, T2Y, T2U, T3l, T39, T2B, T2L;
|
Chris@10
|
719 {
|
Chris@10
|
720 E T2Z, T30, T34, T35;
|
Chris@10
|
721 T2Z = T2N + T2O;
|
Chris@10
|
722 T30 = T2Q + T2R;
|
Chris@10
|
723 T31 = FNMS(KP951056516, T30, KP587785252 * T2Z);
|
Chris@10
|
724 T3h = FMA(KP951056516, T2Z, KP587785252 * T30);
|
Chris@10
|
725 T34 = T2D - T2E;
|
Chris@10
|
726 T35 = T2G - T2H;
|
Chris@10
|
727 T36 = FNMS(KP951056516, T35, KP587785252 * T34);
|
Chris@10
|
728 T3k = FMA(KP951056516, T34, KP587785252 * T35);
|
Chris@10
|
729 }
|
Chris@10
|
730 {
|
Chris@10
|
731 E T2X, T2J, T2W, T38, T2T, T37;
|
Chris@10
|
732 T2X = KP559016994 * (T2F - T2I);
|
Chris@10
|
733 T2J = T2F + T2I;
|
Chris@10
|
734 T2W = FNMS(KP250000000, T2J, T2C);
|
Chris@10
|
735 T2K = T2C + T2J;
|
Chris@10
|
736 T3g = T2X + T2W;
|
Chris@10
|
737 T2Y = T2W - T2X;
|
Chris@10
|
738 T38 = KP559016994 * (T2P - T2S);
|
Chris@10
|
739 T2T = T2P + T2S;
|
Chris@10
|
740 T37 = FNMS(KP250000000, T2T, T2M);
|
Chris@10
|
741 T2U = T2M + T2T;
|
Chris@10
|
742 T3l = T38 + T37;
|
Chris@10
|
743 T39 = T37 - T38;
|
Chris@10
|
744 }
|
Chris@10
|
745 T2B = W[8];
|
Chris@10
|
746 T2L = W[9];
|
Chris@10
|
747 cr[WS(rs, 5)] = FNMS(T2L, T2U, T2B * T2K);
|
Chris@10
|
748 ci[WS(rs, 5)] = FMA(T2L, T2K, T2B * T2U);
|
Chris@10
|
749 {
|
Chris@10
|
750 E T3o, T3q, T3n, T3p;
|
Chris@10
|
751 T3o = T3g + T3h;
|
Chris@10
|
752 T3q = T3l - T3k;
|
Chris@10
|
753 T3n = W[26];
|
Chris@10
|
754 T3p = W[27];
|
Chris@10
|
755 cr[WS(rs, 14)] = FNMS(T3p, T3q, T3n * T3o);
|
Chris@10
|
756 ci[WS(rs, 14)] = FMA(T3n, T3q, T3p * T3o);
|
Chris@10
|
757 }
|
Chris@10
|
758 {
|
Chris@10
|
759 E T32, T3a, T2V, T33;
|
Chris@10
|
760 T32 = T2Y - T31;
|
Chris@10
|
761 T3a = T36 + T39;
|
Chris@10
|
762 T2V = W[2];
|
Chris@10
|
763 T33 = W[3];
|
Chris@10
|
764 cr[WS(rs, 2)] = FNMS(T33, T3a, T2V * T32);
|
Chris@10
|
765 ci[WS(rs, 2)] = FMA(T2V, T3a, T33 * T32);
|
Chris@10
|
766 }
|
Chris@10
|
767 {
|
Chris@10
|
768 E T3c, T3e, T3b, T3d;
|
Chris@10
|
769 T3c = T2Y + T31;
|
Chris@10
|
770 T3e = T39 - T36;
|
Chris@10
|
771 T3b = W[14];
|
Chris@10
|
772 T3d = W[15];
|
Chris@10
|
773 cr[WS(rs, 8)] = FNMS(T3d, T3e, T3b * T3c);
|
Chris@10
|
774 ci[WS(rs, 8)] = FMA(T3b, T3e, T3d * T3c);
|
Chris@10
|
775 }
|
Chris@10
|
776 {
|
Chris@10
|
777 E T3i, T3m, T3f, T3j;
|
Chris@10
|
778 T3i = T3g - T3h;
|
Chris@10
|
779 T3m = T3k + T3l;
|
Chris@10
|
780 T3f = W[20];
|
Chris@10
|
781 T3j = W[21];
|
Chris@10
|
782 cr[WS(rs, 11)] = FNMS(T3j, T3m, T3f * T3i);
|
Chris@10
|
783 ci[WS(rs, 11)] = FMA(T3f, T3m, T3j * T3i);
|
Chris@10
|
784 }
|
Chris@10
|
785 }
|
Chris@10
|
786 }
|
Chris@10
|
787 }
|
Chris@10
|
788 }
|
Chris@10
|
789
|
Chris@10
|
790 static const tw_instr twinstr[] = {
|
Chris@10
|
791 {TW_FULL, 1, 15},
|
Chris@10
|
792 {TW_NEXT, 1, 0}
|
Chris@10
|
793 };
|
Chris@10
|
794
|
Chris@10
|
795 static const hc2hc_desc desc = { 15, "hb_15", twinstr, &GENUS, {128, 56, 56, 0} };
|
Chris@10
|
796
|
Chris@10
|
797 void X(codelet_hb_15) (planner *p) {
|
Chris@10
|
798 X(khc2hc_register) (p, hb_15, &desc);
|
Chris@10
|
799 }
|
Chris@10
|
800 #endif /* HAVE_FMA */
|