cannam@167
|
1 /*
|
cannam@167
|
2 * Copyright (c) 2003, 2007-14 Matteo Frigo
|
cannam@167
|
3 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
|
cannam@167
|
4 *
|
cannam@167
|
5 * This program is free software; you can redistribute it and/or modify
|
cannam@167
|
6 * it under the terms of the GNU General Public License as published by
|
cannam@167
|
7 * the Free Software Foundation; either version 2 of the License, or
|
cannam@167
|
8 * (at your option) any later version.
|
cannam@167
|
9 *
|
cannam@167
|
10 * This program is distributed in the hope that it will be useful,
|
cannam@167
|
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
cannam@167
|
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
cannam@167
|
13 * GNU General Public License for more details.
|
cannam@167
|
14 *
|
cannam@167
|
15 * You should have received a copy of the GNU General Public License
|
cannam@167
|
16 * along with this program; if not, write to the Free Software
|
cannam@167
|
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
cannam@167
|
18 *
|
cannam@167
|
19 */
|
cannam@167
|
20
|
cannam@167
|
21 /* This file was automatically generated --- DO NOT EDIT */
|
cannam@167
|
22 /* Generated on Thu May 24 08:07:32 EDT 2018 */
|
cannam@167
|
23
|
cannam@167
|
24 #include "rdft/codelet-rdft.h"
|
cannam@167
|
25
|
cannam@167
|
26 #if defined(ARCH_PREFERS_FMA) || defined(ISA_EXTENSION_PREFERS_FMA)
|
cannam@167
|
27
|
cannam@167
|
28 /* Generated by: ../../../genfft/gen_hc2hc.native -fma -compact -variables 4 -pipeline-latency 4 -sign 1 -n 15 -dif -name hb_15 -include rdft/scalar/hb.h */
|
cannam@167
|
29
|
cannam@167
|
30 /*
|
cannam@167
|
31 * This function contains 184 FP additions, 140 FP multiplications,
|
cannam@167
|
32 * (or, 72 additions, 28 multiplications, 112 fused multiply/add),
|
cannam@167
|
33 * 78 stack variables, 6 constants, and 60 memory accesses
|
cannam@167
|
34 */
|
cannam@167
|
35 #include "rdft/scalar/hb.h"
|
cannam@167
|
36
|
cannam@167
|
37 static void hb_15(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
|
cannam@167
|
38 {
|
cannam@167
|
39 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
|
cannam@167
|
40 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
|
cannam@167
|
41 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
|
cannam@167
|
42 DK(KP618033988, +0.618033988749894848204586834365638117720309180);
|
cannam@167
|
43 DK(KP866025403, +0.866025403784438646763723170752936183471402627);
|
cannam@167
|
44 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
|
cannam@167
|
45 {
|
cannam@167
|
46 INT m;
|
cannam@167
|
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)) {
|
cannam@167
|
48 E T5, T11, T1C, T2U, T2f, T3f, TH, T19, T18, TS, T12, T13, T14, T3a, T3g;
|
cannam@167
|
49 E Ts, Tv, T37, T3h, T28, T2h, T21, T2g, T2V, T2W, T2X, T2Y, T2Z, T30, T31;
|
cannam@167
|
50 E T1F, T1I, T1J, T1M, T1P, T1Q, T1R;
|
cannam@167
|
51 {
|
cannam@167
|
52 E T1, TX, T4, T2e, T10, T1B, T1A, T2d;
|
cannam@167
|
53 T1 = cr[0];
|
cannam@167
|
54 TX = ci[WS(rs, 14)];
|
cannam@167
|
55 {
|
cannam@167
|
56 E T2, T3, TY, TZ;
|
cannam@167
|
57 T2 = cr[WS(rs, 5)];
|
cannam@167
|
58 T3 = ci[WS(rs, 4)];
|
cannam@167
|
59 T4 = T2 + T3;
|
cannam@167
|
60 T2e = T2 - T3;
|
cannam@167
|
61 TY = ci[WS(rs, 9)];
|
cannam@167
|
62 TZ = cr[WS(rs, 10)];
|
cannam@167
|
63 T10 = TY - TZ;
|
cannam@167
|
64 T1B = TY + TZ;
|
cannam@167
|
65 }
|
cannam@167
|
66 T5 = T1 + T4;
|
cannam@167
|
67 T11 = TX + T10;
|
cannam@167
|
68 T1A = FNMS(KP500000000, T4, T1);
|
cannam@167
|
69 T1C = FNMS(KP866025403, T1B, T1A);
|
cannam@167
|
70 T2U = FMA(KP866025403, T1B, T1A);
|
cannam@167
|
71 T2d = FNMS(KP500000000, T10, TX);
|
cannam@167
|
72 T2f = FMA(KP866025403, T2e, T2d);
|
cannam@167
|
73 T3f = FNMS(KP866025403, T2e, T2d);
|
cannam@167
|
74 }
|
cannam@167
|
75 {
|
cannam@167
|
76 E Ta, T1W, T1D, Tl, T23, T1K, Tf, T1Z, T1G, TR, T1Y, T1H, Tq, T26, T1N;
|
cannam@167
|
77 E TG, T25, T1O, TM, T1V, T1E, TB, T22, T1L, T38, T39;
|
cannam@167
|
78 {
|
cannam@167
|
79 E T6, T7, T8, T9;
|
cannam@167
|
80 T6 = cr[WS(rs, 3)];
|
cannam@167
|
81 T7 = ci[WS(rs, 6)];
|
cannam@167
|
82 T8 = ci[WS(rs, 1)];
|
cannam@167
|
83 T9 = T7 + T8;
|
cannam@167
|
84 Ta = T6 + T9;
|
cannam@167
|
85 T1W = T7 - T8;
|
cannam@167
|
86 T1D = FNMS(KP500000000, T9, T6);
|
cannam@167
|
87 }
|
cannam@167
|
88 {
|
cannam@167
|
89 E Th, Ti, Tj, Tk;
|
cannam@167
|
90 Th = cr[WS(rs, 6)];
|
cannam@167
|
91 Ti = ci[WS(rs, 3)];
|
cannam@167
|
92 Tj = cr[WS(rs, 1)];
|
cannam@167
|
93 Tk = Ti + Tj;
|
cannam@167
|
94 Tl = Th + Tk;
|
cannam@167
|
95 T23 = Ti - Tj;
|
cannam@167
|
96 T1K = FNMS(KP500000000, Tk, Th);
|
cannam@167
|
97 }
|
cannam@167
|
98 {
|
cannam@167
|
99 E Tb, Tc, Td, Te;
|
cannam@167
|
100 Tb = ci[WS(rs, 2)];
|
cannam@167
|
101 Tc = cr[WS(rs, 2)];
|
cannam@167
|
102 Td = cr[WS(rs, 7)];
|
cannam@167
|
103 Te = Tc + Td;
|
cannam@167
|
104 Tf = Tb + Te;
|
cannam@167
|
105 T1Z = Tc - Td;
|
cannam@167
|
106 T1G = FNMS(KP500000000, Te, Tb);
|
cannam@167
|
107 }
|
cannam@167
|
108 {
|
cannam@167
|
109 E TQ, TN, TO, TP;
|
cannam@167
|
110 TQ = cr[WS(rs, 12)];
|
cannam@167
|
111 TN = ci[WS(rs, 12)];
|
cannam@167
|
112 TO = ci[WS(rs, 7)];
|
cannam@167
|
113 TP = TN + TO;
|
cannam@167
|
114 TR = TP - TQ;
|
cannam@167
|
115 T1Y = FMA(KP500000000, TP, TQ);
|
cannam@167
|
116 T1H = TO - TN;
|
cannam@167
|
117 }
|
cannam@167
|
118 {
|
cannam@167
|
119 E Tm, Tn, To, Tp;
|
cannam@167
|
120 Tm = ci[WS(rs, 5)];
|
cannam@167
|
121 Tn = ci[0];
|
cannam@167
|
122 To = cr[WS(rs, 4)];
|
cannam@167
|
123 Tp = Tn + To;
|
cannam@167
|
124 Tq = Tm + Tp;
|
cannam@167
|
125 T26 = Tn - To;
|
cannam@167
|
126 T1N = FNMS(KP500000000, Tp, Tm);
|
cannam@167
|
127 }
|
cannam@167
|
128 {
|
cannam@167
|
129 E TF, TC, TD, TE;
|
cannam@167
|
130 TF = cr[WS(rs, 9)];
|
cannam@167
|
131 TC = ci[WS(rs, 10)];
|
cannam@167
|
132 TD = cr[WS(rs, 14)];
|
cannam@167
|
133 TE = TC - TD;
|
cannam@167
|
134 TG = TE - TF;
|
cannam@167
|
135 T25 = FMA(KP500000000, TE, TF);
|
cannam@167
|
136 T1O = TC + TD;
|
cannam@167
|
137 }
|
cannam@167
|
138 {
|
cannam@167
|
139 E TI, TJ, TK, TL;
|
cannam@167
|
140 TI = ci[WS(rs, 11)];
|
cannam@167
|
141 TJ = cr[WS(rs, 8)];
|
cannam@167
|
142 TK = cr[WS(rs, 13)];
|
cannam@167
|
143 TL = TJ + TK;
|
cannam@167
|
144 TM = TI - TL;
|
cannam@167
|
145 T1V = FMA(KP500000000, TL, TI);
|
cannam@167
|
146 T1E = TJ - TK;
|
cannam@167
|
147 }
|
cannam@167
|
148 {
|
cannam@167
|
149 E Tx, Ty, Tz, TA;
|
cannam@167
|
150 Tx = ci[WS(rs, 8)];
|
cannam@167
|
151 Ty = ci[WS(rs, 13)];
|
cannam@167
|
152 Tz = cr[WS(rs, 11)];
|
cannam@167
|
153 TA = Ty - Tz;
|
cannam@167
|
154 TB = Tx + TA;
|
cannam@167
|
155 T22 = FNMS(KP500000000, TA, Tx);
|
cannam@167
|
156 T1L = Ty + Tz;
|
cannam@167
|
157 }
|
cannam@167
|
158 TH = TB - TG;
|
cannam@167
|
159 T19 = Ta - Tf;
|
cannam@167
|
160 T18 = Tl - Tq;
|
cannam@167
|
161 TS = TM - TR;
|
cannam@167
|
162 T12 = TM + TR;
|
cannam@167
|
163 T13 = TB + TG;
|
cannam@167
|
164 T14 = T12 + T13;
|
cannam@167
|
165 T38 = FNMS(KP866025403, T1W, T1V);
|
cannam@167
|
166 T39 = FMA(KP866025403, T1Z, T1Y);
|
cannam@167
|
167 T3a = T38 + T39;
|
cannam@167
|
168 T3g = T38 - T39;
|
cannam@167
|
169 {
|
cannam@167
|
170 E Tg, Tr, T1X, T20;
|
cannam@167
|
171 Tg = Ta + Tf;
|
cannam@167
|
172 Tr = Tl + Tq;
|
cannam@167
|
173 Ts = Tg + Tr;
|
cannam@167
|
174 Tv = Tg - Tr;
|
cannam@167
|
175 {
|
cannam@167
|
176 E T35, T36, T24, T27;
|
cannam@167
|
177 T35 = FNMS(KP866025403, T23, T22);
|
cannam@167
|
178 T36 = FMA(KP866025403, T26, T25);
|
cannam@167
|
179 T37 = T35 + T36;
|
cannam@167
|
180 T3h = T35 - T36;
|
cannam@167
|
181 T24 = FMA(KP866025403, T23, T22);
|
cannam@167
|
182 T27 = FNMS(KP866025403, T26, T25);
|
cannam@167
|
183 T28 = T24 + T27;
|
cannam@167
|
184 T2h = T24 - T27;
|
cannam@167
|
185 }
|
cannam@167
|
186 T1X = FMA(KP866025403, T1W, T1V);
|
cannam@167
|
187 T20 = FNMS(KP866025403, T1Z, T1Y);
|
cannam@167
|
188 T21 = T1X + T20;
|
cannam@167
|
189 T2g = T1X - T20;
|
cannam@167
|
190 T2V = FNMS(KP866025403, T1E, T1D);
|
cannam@167
|
191 T2W = FNMS(KP866025403, T1H, T1G);
|
cannam@167
|
192 T2X = T2V + T2W;
|
cannam@167
|
193 T2Y = FNMS(KP866025403, T1L, T1K);
|
cannam@167
|
194 T2Z = FNMS(KP866025403, T1O, T1N);
|
cannam@167
|
195 T30 = T2Y + T2Z;
|
cannam@167
|
196 T31 = T2X + T30;
|
cannam@167
|
197 T1F = FMA(KP866025403, T1E, T1D);
|
cannam@167
|
198 T1I = FMA(KP866025403, T1H, T1G);
|
cannam@167
|
199 T1J = T1F + T1I;
|
cannam@167
|
200 T1M = FMA(KP866025403, T1L, T1K);
|
cannam@167
|
201 T1P = FMA(KP866025403, T1O, T1N);
|
cannam@167
|
202 T1Q = T1M + T1P;
|
cannam@167
|
203 T1R = T1J + T1Q;
|
cannam@167
|
204 }
|
cannam@167
|
205 }
|
cannam@167
|
206 cr[0] = T5 + Ts;
|
cannam@167
|
207 ci[0] = T11 + T14;
|
cannam@167
|
208 {
|
cannam@167
|
209 E T1a, T1q, T17, T1p, TU, T1u, T1e, T1m, T15, T16;
|
cannam@167
|
210 T1a = FNMS(KP618033988, T19, T18);
|
cannam@167
|
211 T1q = FMA(KP618033988, T18, T19);
|
cannam@167
|
212 T15 = FNMS(KP250000000, T14, T11);
|
cannam@167
|
213 T16 = T12 - T13;
|
cannam@167
|
214 T17 = FNMS(KP559016994, T16, T15);
|
cannam@167
|
215 T1p = FMA(KP559016994, T16, T15);
|
cannam@167
|
216 {
|
cannam@167
|
217 E TT, T1l, Tw, T1k, Tu;
|
cannam@167
|
218 TT = FNMS(KP618033988, TS, TH);
|
cannam@167
|
219 T1l = FMA(KP618033988, TH, TS);
|
cannam@167
|
220 Tu = FNMS(KP250000000, Ts, T5);
|
cannam@167
|
221 Tw = FNMS(KP559016994, Tv, Tu);
|
cannam@167
|
222 T1k = FMA(KP559016994, Tv, Tu);
|
cannam@167
|
223 TU = FNMS(KP951056516, TT, Tw);
|
cannam@167
|
224 T1u = FMA(KP951056516, T1l, T1k);
|
cannam@167
|
225 T1e = FMA(KP951056516, TT, Tw);
|
cannam@167
|
226 T1m = FNMS(KP951056516, T1l, T1k);
|
cannam@167
|
227 }
|
cannam@167
|
228 {
|
cannam@167
|
229 E T1b, TW, T1c, Tt, TV;
|
cannam@167
|
230 T1b = FMA(KP951056516, T1a, T17);
|
cannam@167
|
231 TW = W[5];
|
cannam@167
|
232 T1c = TW * TU;
|
cannam@167
|
233 Tt = W[4];
|
cannam@167
|
234 TV = Tt * TU;
|
cannam@167
|
235 cr[WS(rs, 3)] = FNMS(TW, T1b, TV);
|
cannam@167
|
236 ci[WS(rs, 3)] = FMA(Tt, T1b, T1c);
|
cannam@167
|
237 }
|
cannam@167
|
238 {
|
cannam@167
|
239 E T1x, T1w, T1y, T1t, T1v;
|
cannam@167
|
240 T1x = FNMS(KP951056516, T1q, T1p);
|
cannam@167
|
241 T1w = W[17];
|
cannam@167
|
242 T1y = T1w * T1u;
|
cannam@167
|
243 T1t = W[16];
|
cannam@167
|
244 T1v = T1t * T1u;
|
cannam@167
|
245 cr[WS(rs, 9)] = FNMS(T1w, T1x, T1v);
|
cannam@167
|
246 ci[WS(rs, 9)] = FMA(T1t, T1x, T1y);
|
cannam@167
|
247 }
|
cannam@167
|
248 {
|
cannam@167
|
249 E T1h, T1g, T1i, T1d, T1f;
|
cannam@167
|
250 T1h = FNMS(KP951056516, T1a, T17);
|
cannam@167
|
251 T1g = W[23];
|
cannam@167
|
252 T1i = T1g * T1e;
|
cannam@167
|
253 T1d = W[22];
|
cannam@167
|
254 T1f = T1d * T1e;
|
cannam@167
|
255 cr[WS(rs, 12)] = FNMS(T1g, T1h, T1f);
|
cannam@167
|
256 ci[WS(rs, 12)] = FMA(T1d, T1h, T1i);
|
cannam@167
|
257 }
|
cannam@167
|
258 {
|
cannam@167
|
259 E T1r, T1o, T1s, T1j, T1n;
|
cannam@167
|
260 T1r = FMA(KP951056516, T1q, T1p);
|
cannam@167
|
261 T1o = W[11];
|
cannam@167
|
262 T1s = T1o * T1m;
|
cannam@167
|
263 T1j = W[10];
|
cannam@167
|
264 T1n = T1j * T1m;
|
cannam@167
|
265 cr[WS(rs, 6)] = FNMS(T1o, T1r, T1n);
|
cannam@167
|
266 ci[WS(rs, 6)] = FMA(T1j, T1r, T1s);
|
cannam@167
|
267 }
|
cannam@167
|
268 }
|
cannam@167
|
269 {
|
cannam@167
|
270 E T2o, T2E, T2N, T2P, T2Q, T2S, T2l, T2R, T2D, T2a, T2I, T2s, T2A;
|
cannam@167
|
271 {
|
cannam@167
|
272 E T2m, T2n, T2O, T2k, T2i, T2j;
|
cannam@167
|
273 T2m = T1F - T1I;
|
cannam@167
|
274 T2n = T1M - T1P;
|
cannam@167
|
275 T2o = FMA(KP618033988, T2n, T2m);
|
cannam@167
|
276 T2E = FNMS(KP618033988, T2m, T2n);
|
cannam@167
|
277 T2O = T1C + T1R;
|
cannam@167
|
278 T2N = W[18];
|
cannam@167
|
279 T2P = T2N * T2O;
|
cannam@167
|
280 T2Q = W[19];
|
cannam@167
|
281 T2S = T2Q * T2O;
|
cannam@167
|
282 T2k = T2g - T2h;
|
cannam@167
|
283 T2i = T2g + T2h;
|
cannam@167
|
284 T2j = FNMS(KP250000000, T2i, T2f);
|
cannam@167
|
285 T2l = FMA(KP559016994, T2k, T2j);
|
cannam@167
|
286 T2R = T2f + T2i;
|
cannam@167
|
287 T2D = FNMS(KP559016994, T2k, T2j);
|
cannam@167
|
288 {
|
cannam@167
|
289 E T29, T2z, T1U, T2y, T1S, T1T;
|
cannam@167
|
290 T29 = FMA(KP618033988, T28, T21);
|
cannam@167
|
291 T2z = FNMS(KP618033988, T21, T28);
|
cannam@167
|
292 T1S = FNMS(KP250000000, T1R, T1C);
|
cannam@167
|
293 T1T = T1J - T1Q;
|
cannam@167
|
294 T1U = FMA(KP559016994, T1T, T1S);
|
cannam@167
|
295 T2y = FNMS(KP559016994, T1T, T1S);
|
cannam@167
|
296 T2a = FNMS(KP951056516, T29, T1U);
|
cannam@167
|
297 T2I = FNMS(KP951056516, T2z, T2y);
|
cannam@167
|
298 T2s = FMA(KP951056516, T29, T1U);
|
cannam@167
|
299 T2A = FMA(KP951056516, T2z, T2y);
|
cannam@167
|
300 }
|
cannam@167
|
301 }
|
cannam@167
|
302 cr[WS(rs, 10)] = FNMS(T2Q, T2R, T2P);
|
cannam@167
|
303 ci[WS(rs, 10)] = FMA(T2N, T2R, T2S);
|
cannam@167
|
304 {
|
cannam@167
|
305 E T2p, T2c, T2q, T1z, T2b;
|
cannam@167
|
306 T2p = FMA(KP951056516, T2o, T2l);
|
cannam@167
|
307 T2c = W[1];
|
cannam@167
|
308 T2q = T2c * T2a;
|
cannam@167
|
309 T1z = W[0];
|
cannam@167
|
310 T2b = T1z * T2a;
|
cannam@167
|
311 cr[WS(rs, 1)] = FNMS(T2c, T2p, T2b);
|
cannam@167
|
312 ci[WS(rs, 1)] = FMA(T1z, T2p, T2q);
|
cannam@167
|
313 }
|
cannam@167
|
314 {
|
cannam@167
|
315 E T2L, T2K, T2M, T2H, T2J;
|
cannam@167
|
316 T2L = FMA(KP951056516, T2E, T2D);
|
cannam@167
|
317 T2K = W[25];
|
cannam@167
|
318 T2M = T2K * T2I;
|
cannam@167
|
319 T2H = W[24];
|
cannam@167
|
320 T2J = T2H * T2I;
|
cannam@167
|
321 cr[WS(rs, 13)] = FNMS(T2K, T2L, T2J);
|
cannam@167
|
322 ci[WS(rs, 13)] = FMA(T2H, T2L, T2M);
|
cannam@167
|
323 }
|
cannam@167
|
324 {
|
cannam@167
|
325 E T2F, T2C, T2G, T2x, T2B;
|
cannam@167
|
326 T2F = FNMS(KP951056516, T2E, T2D);
|
cannam@167
|
327 T2C = W[13];
|
cannam@167
|
328 T2G = T2C * T2A;
|
cannam@167
|
329 T2x = W[12];
|
cannam@167
|
330 T2B = T2x * T2A;
|
cannam@167
|
331 cr[WS(rs, 7)] = FNMS(T2C, T2F, T2B);
|
cannam@167
|
332 ci[WS(rs, 7)] = FMA(T2x, T2F, T2G);
|
cannam@167
|
333 }
|
cannam@167
|
334 {
|
cannam@167
|
335 E T2v, T2u, T2w, T2r, T2t;
|
cannam@167
|
336 T2v = FNMS(KP951056516, T2o, T2l);
|
cannam@167
|
337 T2u = W[7];
|
cannam@167
|
338 T2w = T2u * T2s;
|
cannam@167
|
339 T2r = W[6];
|
cannam@167
|
340 T2t = T2r * T2s;
|
cannam@167
|
341 cr[WS(rs, 4)] = FNMS(T2u, T2v, T2t);
|
cannam@167
|
342 ci[WS(rs, 4)] = FMA(T2r, T2v, T2w);
|
cannam@167
|
343 }
|
cannam@167
|
344 }
|
cannam@167
|
345 {
|
cannam@167
|
346 E T3o, T3E, T3N, T3P, T3Q, T3S, T3l, T3R, T3D, T3c, T3I, T3s, T3A;
|
cannam@167
|
347 {
|
cannam@167
|
348 E T3m, T3n, T3O, T3k, T3i, T3j;
|
cannam@167
|
349 T3m = T2Y - T2Z;
|
cannam@167
|
350 T3n = T2V - T2W;
|
cannam@167
|
351 T3o = FNMS(KP618033988, T3n, T3m);
|
cannam@167
|
352 T3E = FMA(KP618033988, T3m, T3n);
|
cannam@167
|
353 T3O = T2U + T31;
|
cannam@167
|
354 T3N = W[8];
|
cannam@167
|
355 T3P = T3N * T3O;
|
cannam@167
|
356 T3Q = W[9];
|
cannam@167
|
357 T3S = T3Q * T3O;
|
cannam@167
|
358 T3k = T3g - T3h;
|
cannam@167
|
359 T3i = T3g + T3h;
|
cannam@167
|
360 T3j = FNMS(KP250000000, T3i, T3f);
|
cannam@167
|
361 T3l = FNMS(KP559016994, T3k, T3j);
|
cannam@167
|
362 T3R = T3f + T3i;
|
cannam@167
|
363 T3D = FMA(KP559016994, T3k, T3j);
|
cannam@167
|
364 {
|
cannam@167
|
365 E T3b, T3z, T34, T3y, T32, T33;
|
cannam@167
|
366 T3b = FNMS(KP618033988, T3a, T37);
|
cannam@167
|
367 T3z = FMA(KP618033988, T37, T3a);
|
cannam@167
|
368 T32 = FNMS(KP250000000, T31, T2U);
|
cannam@167
|
369 T33 = T2X - T30;
|
cannam@167
|
370 T34 = FNMS(KP559016994, T33, T32);
|
cannam@167
|
371 T3y = FMA(KP559016994, T33, T32);
|
cannam@167
|
372 T3c = FMA(KP951056516, T3b, T34);
|
cannam@167
|
373 T3I = FMA(KP951056516, T3z, T3y);
|
cannam@167
|
374 T3s = FNMS(KP951056516, T3b, T34);
|
cannam@167
|
375 T3A = FNMS(KP951056516, T3z, T3y);
|
cannam@167
|
376 }
|
cannam@167
|
377 }
|
cannam@167
|
378 cr[WS(rs, 5)] = FNMS(T3Q, T3R, T3P);
|
cannam@167
|
379 ci[WS(rs, 5)] = FMA(T3N, T3R, T3S);
|
cannam@167
|
380 {
|
cannam@167
|
381 E T3p, T3e, T3q, T2T, T3d;
|
cannam@167
|
382 T3p = FNMS(KP951056516, T3o, T3l);
|
cannam@167
|
383 T3e = W[3];
|
cannam@167
|
384 T3q = T3e * T3c;
|
cannam@167
|
385 T2T = W[2];
|
cannam@167
|
386 T3d = T2T * T3c;
|
cannam@167
|
387 cr[WS(rs, 2)] = FNMS(T3e, T3p, T3d);
|
cannam@167
|
388 ci[WS(rs, 2)] = FMA(T2T, T3p, T3q);
|
cannam@167
|
389 }
|
cannam@167
|
390 {
|
cannam@167
|
391 E T3L, T3K, T3M, T3H, T3J;
|
cannam@167
|
392 T3L = FNMS(KP951056516, T3E, T3D);
|
cannam@167
|
393 T3K = W[27];
|
cannam@167
|
394 T3M = T3K * T3I;
|
cannam@167
|
395 T3H = W[26];
|
cannam@167
|
396 T3J = T3H * T3I;
|
cannam@167
|
397 cr[WS(rs, 14)] = FNMS(T3K, T3L, T3J);
|
cannam@167
|
398 ci[WS(rs, 14)] = FMA(T3H, T3L, T3M);
|
cannam@167
|
399 }
|
cannam@167
|
400 {
|
cannam@167
|
401 E T3F, T3C, T3G, T3x, T3B;
|
cannam@167
|
402 T3F = FMA(KP951056516, T3E, T3D);
|
cannam@167
|
403 T3C = W[21];
|
cannam@167
|
404 T3G = T3C * T3A;
|
cannam@167
|
405 T3x = W[20];
|
cannam@167
|
406 T3B = T3x * T3A;
|
cannam@167
|
407 cr[WS(rs, 11)] = FNMS(T3C, T3F, T3B);
|
cannam@167
|
408 ci[WS(rs, 11)] = FMA(T3x, T3F, T3G);
|
cannam@167
|
409 }
|
cannam@167
|
410 {
|
cannam@167
|
411 E T3v, T3u, T3w, T3r, T3t;
|
cannam@167
|
412 T3v = FMA(KP951056516, T3o, T3l);
|
cannam@167
|
413 T3u = W[15];
|
cannam@167
|
414 T3w = T3u * T3s;
|
cannam@167
|
415 T3r = W[14];
|
cannam@167
|
416 T3t = T3r * T3s;
|
cannam@167
|
417 cr[WS(rs, 8)] = FNMS(T3u, T3v, T3t);
|
cannam@167
|
418 ci[WS(rs, 8)] = FMA(T3r, T3v, T3w);
|
cannam@167
|
419 }
|
cannam@167
|
420 }
|
cannam@167
|
421 }
|
cannam@167
|
422 }
|
cannam@167
|
423 }
|
cannam@167
|
424
|
cannam@167
|
425 static const tw_instr twinstr[] = {
|
cannam@167
|
426 {TW_FULL, 1, 15},
|
cannam@167
|
427 {TW_NEXT, 1, 0}
|
cannam@167
|
428 };
|
cannam@167
|
429
|
cannam@167
|
430 static const hc2hc_desc desc = { 15, "hb_15", twinstr, &GENUS, {72, 28, 112, 0} };
|
cannam@167
|
431
|
cannam@167
|
432 void X(codelet_hb_15) (planner *p) {
|
cannam@167
|
433 X(khc2hc_register) (p, hb_15, &desc);
|
cannam@167
|
434 }
|
cannam@167
|
435 #else
|
cannam@167
|
436
|
cannam@167
|
437 /* Generated by: ../../../genfft/gen_hc2hc.native -compact -variables 4 -pipeline-latency 4 -sign 1 -n 15 -dif -name hb_15 -include rdft/scalar/hb.h */
|
cannam@167
|
438
|
cannam@167
|
439 /*
|
cannam@167
|
440 * This function contains 184 FP additions, 112 FP multiplications,
|
cannam@167
|
441 * (or, 128 additions, 56 multiplications, 56 fused multiply/add),
|
cannam@167
|
442 * 75 stack variables, 6 constants, and 60 memory accesses
|
cannam@167
|
443 */
|
cannam@167
|
444 #include "rdft/scalar/hb.h"
|
cannam@167
|
445
|
cannam@167
|
446 static void hb_15(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
|
cannam@167
|
447 {
|
cannam@167
|
448 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
|
cannam@167
|
449 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
|
cannam@167
|
450 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
|
cannam@167
|
451 DK(KP587785252, +0.587785252292473129168705954639072768597652438);
|
cannam@167
|
452 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
|
cannam@167
|
453 DK(KP866025403, +0.866025403784438646763723170752936183471402627);
|
cannam@167
|
454 {
|
cannam@167
|
455 INT m;
|
cannam@167
|
456 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)) {
|
cannam@167
|
457 E T5, T10, T1J, T2C, T2c, T2M, TH, T18, T17, TS, T2Q, T2R, T2S, Tg, Tr;
|
cannam@167
|
458 E Ts, T11, T12, T13, T2N, T2O, T2P, T1u, T1x, T1y, T1W, T1Z, T28, T1P, T1S;
|
cannam@167
|
459 E T27, T1B, T1E, T1F, T2G, T2H, T2I, T2D, T2E, T2F;
|
cannam@167
|
460 {
|
cannam@167
|
461 E T1, TW, T4, T2a, TZ, T1I, T1H, T2b;
|
cannam@167
|
462 T1 = cr[0];
|
cannam@167
|
463 TW = ci[WS(rs, 14)];
|
cannam@167
|
464 {
|
cannam@167
|
465 E T2, T3, TX, TY;
|
cannam@167
|
466 T2 = cr[WS(rs, 5)];
|
cannam@167
|
467 T3 = ci[WS(rs, 4)];
|
cannam@167
|
468 T4 = T2 + T3;
|
cannam@167
|
469 T2a = KP866025403 * (T2 - T3);
|
cannam@167
|
470 TX = ci[WS(rs, 9)];
|
cannam@167
|
471 TY = cr[WS(rs, 10)];
|
cannam@167
|
472 TZ = TX - TY;
|
cannam@167
|
473 T1I = KP866025403 * (TX + TY);
|
cannam@167
|
474 }
|
cannam@167
|
475 T5 = T1 + T4;
|
cannam@167
|
476 T10 = TW + TZ;
|
cannam@167
|
477 T1H = FNMS(KP500000000, T4, T1);
|
cannam@167
|
478 T1J = T1H - T1I;
|
cannam@167
|
479 T2C = T1H + T1I;
|
cannam@167
|
480 T2b = FNMS(KP500000000, TZ, TW);
|
cannam@167
|
481 T2c = T2a + T2b;
|
cannam@167
|
482 T2M = T2b - T2a;
|
cannam@167
|
483 }
|
cannam@167
|
484 {
|
cannam@167
|
485 E Ta, T1N, T1s, Tl, T1U, T1z, Tf, T1Q, T1v, TG, T1R, T1w, Tq, T1X, T1C;
|
cannam@167
|
486 E TM, T1V, T1A, TB, T1O, T1t, TR, T1Y, T1D;
|
cannam@167
|
487 {
|
cannam@167
|
488 E T6, T7, T8, T9;
|
cannam@167
|
489 T6 = cr[WS(rs, 3)];
|
cannam@167
|
490 T7 = ci[WS(rs, 6)];
|
cannam@167
|
491 T8 = ci[WS(rs, 1)];
|
cannam@167
|
492 T9 = T7 + T8;
|
cannam@167
|
493 Ta = T6 + T9;
|
cannam@167
|
494 T1N = KP866025403 * (T7 - T8);
|
cannam@167
|
495 T1s = FNMS(KP500000000, T9, T6);
|
cannam@167
|
496 }
|
cannam@167
|
497 {
|
cannam@167
|
498 E Th, Ti, Tj, Tk;
|
cannam@167
|
499 Th = cr[WS(rs, 6)];
|
cannam@167
|
500 Ti = ci[WS(rs, 3)];
|
cannam@167
|
501 Tj = cr[WS(rs, 1)];
|
cannam@167
|
502 Tk = Ti + Tj;
|
cannam@167
|
503 Tl = Th + Tk;
|
cannam@167
|
504 T1U = KP866025403 * (Ti - Tj);
|
cannam@167
|
505 T1z = FNMS(KP500000000, Tk, Th);
|
cannam@167
|
506 }
|
cannam@167
|
507 {
|
cannam@167
|
508 E Tb, Tc, Td, Te;
|
cannam@167
|
509 Tb = ci[WS(rs, 2)];
|
cannam@167
|
510 Tc = cr[WS(rs, 2)];
|
cannam@167
|
511 Td = cr[WS(rs, 7)];
|
cannam@167
|
512 Te = Tc + Td;
|
cannam@167
|
513 Tf = Tb + Te;
|
cannam@167
|
514 T1Q = KP866025403 * (Tc - Td);
|
cannam@167
|
515 T1v = FNMS(KP500000000, Te, Tb);
|
cannam@167
|
516 }
|
cannam@167
|
517 {
|
cannam@167
|
518 E TF, TC, TD, TE;
|
cannam@167
|
519 TF = cr[WS(rs, 12)];
|
cannam@167
|
520 TC = ci[WS(rs, 12)];
|
cannam@167
|
521 TD = ci[WS(rs, 7)];
|
cannam@167
|
522 TE = TC + TD;
|
cannam@167
|
523 TG = TE - TF;
|
cannam@167
|
524 T1R = FMA(KP500000000, TE, TF);
|
cannam@167
|
525 T1w = KP866025403 * (TD - TC);
|
cannam@167
|
526 }
|
cannam@167
|
527 {
|
cannam@167
|
528 E Tm, Tn, To, Tp;
|
cannam@167
|
529 Tm = ci[WS(rs, 5)];
|
cannam@167
|
530 Tn = ci[0];
|
cannam@167
|
531 To = cr[WS(rs, 4)];
|
cannam@167
|
532 Tp = Tn + To;
|
cannam@167
|
533 Tq = Tm + Tp;
|
cannam@167
|
534 T1X = KP866025403 * (Tn - To);
|
cannam@167
|
535 T1C = FNMS(KP500000000, Tp, Tm);
|
cannam@167
|
536 }
|
cannam@167
|
537 {
|
cannam@167
|
538 E TI, TJ, TK, TL;
|
cannam@167
|
539 TI = ci[WS(rs, 8)];
|
cannam@167
|
540 TJ = ci[WS(rs, 13)];
|
cannam@167
|
541 TK = cr[WS(rs, 11)];
|
cannam@167
|
542 TL = TJ - TK;
|
cannam@167
|
543 TM = TI + TL;
|
cannam@167
|
544 T1V = FNMS(KP500000000, TL, TI);
|
cannam@167
|
545 T1A = KP866025403 * (TJ + TK);
|
cannam@167
|
546 }
|
cannam@167
|
547 {
|
cannam@167
|
548 E Tx, Ty, Tz, TA;
|
cannam@167
|
549 Tx = ci[WS(rs, 11)];
|
cannam@167
|
550 Ty = cr[WS(rs, 8)];
|
cannam@167
|
551 Tz = cr[WS(rs, 13)];
|
cannam@167
|
552 TA = Ty + Tz;
|
cannam@167
|
553 TB = Tx - TA;
|
cannam@167
|
554 T1O = FMA(KP500000000, TA, Tx);
|
cannam@167
|
555 T1t = KP866025403 * (Ty - Tz);
|
cannam@167
|
556 }
|
cannam@167
|
557 {
|
cannam@167
|
558 E TQ, TN, TO, TP;
|
cannam@167
|
559 TQ = cr[WS(rs, 9)];
|
cannam@167
|
560 TN = ci[WS(rs, 10)];
|
cannam@167
|
561 TO = cr[WS(rs, 14)];
|
cannam@167
|
562 TP = TN - TO;
|
cannam@167
|
563 TR = TP - TQ;
|
cannam@167
|
564 T1Y = FMA(KP500000000, TP, TQ);
|
cannam@167
|
565 T1D = KP866025403 * (TN + TO);
|
cannam@167
|
566 }
|
cannam@167
|
567 TH = TB - TG;
|
cannam@167
|
568 T18 = Tl - Tq;
|
cannam@167
|
569 T17 = Ta - Tf;
|
cannam@167
|
570 TS = TM - TR;
|
cannam@167
|
571 T2Q = T1V - T1U;
|
cannam@167
|
572 T2R = T1X + T1Y;
|
cannam@167
|
573 T2S = T2Q - T2R;
|
cannam@167
|
574 Tg = Ta + Tf;
|
cannam@167
|
575 Tr = Tl + Tq;
|
cannam@167
|
576 Ts = Tg + Tr;
|
cannam@167
|
577 T11 = TB + TG;
|
cannam@167
|
578 T12 = TM + TR;
|
cannam@167
|
579 T13 = T11 + T12;
|
cannam@167
|
580 T2N = T1O - T1N;
|
cannam@167
|
581 T2O = T1Q + T1R;
|
cannam@167
|
582 T2P = T2N - T2O;
|
cannam@167
|
583 T1u = T1s + T1t;
|
cannam@167
|
584 T1x = T1v + T1w;
|
cannam@167
|
585 T1y = T1u + T1x;
|
cannam@167
|
586 T1W = T1U + T1V;
|
cannam@167
|
587 T1Z = T1X - T1Y;
|
cannam@167
|
588 T28 = T1W + T1Z;
|
cannam@167
|
589 T1P = T1N + T1O;
|
cannam@167
|
590 T1S = T1Q - T1R;
|
cannam@167
|
591 T27 = T1P + T1S;
|
cannam@167
|
592 T1B = T1z + T1A;
|
cannam@167
|
593 T1E = T1C + T1D;
|
cannam@167
|
594 T1F = T1B + T1E;
|
cannam@167
|
595 T2G = T1z - T1A;
|
cannam@167
|
596 T2H = T1C - T1D;
|
cannam@167
|
597 T2I = T2G + T2H;
|
cannam@167
|
598 T2D = T1s - T1t;
|
cannam@167
|
599 T2E = T1v - T1w;
|
cannam@167
|
600 T2F = T2D + T2E;
|
cannam@167
|
601 }
|
cannam@167
|
602 cr[0] = T5 + Ts;
|
cannam@167
|
603 ci[0] = T10 + T13;
|
cannam@167
|
604 {
|
cannam@167
|
605 E TT, T19, T1k, T1h, T16, T1l, Tw, T1g;
|
cannam@167
|
606 TT = FNMS(KP951056516, TS, KP587785252 * TH);
|
cannam@167
|
607 T19 = FNMS(KP951056516, T18, KP587785252 * T17);
|
cannam@167
|
608 T1k = FMA(KP951056516, T17, KP587785252 * T18);
|
cannam@167
|
609 T1h = FMA(KP951056516, TH, KP587785252 * TS);
|
cannam@167
|
610 {
|
cannam@167
|
611 E T14, T15, Tu, Tv;
|
cannam@167
|
612 T14 = FNMS(KP250000000, T13, T10);
|
cannam@167
|
613 T15 = KP559016994 * (T11 - T12);
|
cannam@167
|
614 T16 = T14 - T15;
|
cannam@167
|
615 T1l = T15 + T14;
|
cannam@167
|
616 Tu = FNMS(KP250000000, Ts, T5);
|
cannam@167
|
617 Tv = KP559016994 * (Tg - Tr);
|
cannam@167
|
618 Tw = Tu - Tv;
|
cannam@167
|
619 T1g = Tv + Tu;
|
cannam@167
|
620 }
|
cannam@167
|
621 {
|
cannam@167
|
622 E TU, T1a, Tt, TV;
|
cannam@167
|
623 TU = Tw + TT;
|
cannam@167
|
624 T1a = T16 - T19;
|
cannam@167
|
625 Tt = W[4];
|
cannam@167
|
626 TV = W[5];
|
cannam@167
|
627 cr[WS(rs, 3)] = FNMS(TV, T1a, Tt * TU);
|
cannam@167
|
628 ci[WS(rs, 3)] = FMA(TV, TU, Tt * T1a);
|
cannam@167
|
629 }
|
cannam@167
|
630 {
|
cannam@167
|
631 E T1o, T1q, T1n, T1p;
|
cannam@167
|
632 T1o = T1g + T1h;
|
cannam@167
|
633 T1q = T1l - T1k;
|
cannam@167
|
634 T1n = W[16];
|
cannam@167
|
635 T1p = W[17];
|
cannam@167
|
636 cr[WS(rs, 9)] = FNMS(T1p, T1q, T1n * T1o);
|
cannam@167
|
637 ci[WS(rs, 9)] = FMA(T1p, T1o, T1n * T1q);
|
cannam@167
|
638 }
|
cannam@167
|
639 {
|
cannam@167
|
640 E T1c, T1e, T1b, T1d;
|
cannam@167
|
641 T1c = Tw - TT;
|
cannam@167
|
642 T1e = T19 + T16;
|
cannam@167
|
643 T1b = W[22];
|
cannam@167
|
644 T1d = W[23];
|
cannam@167
|
645 cr[WS(rs, 12)] = FNMS(T1d, T1e, T1b * T1c);
|
cannam@167
|
646 ci[WS(rs, 12)] = FMA(T1d, T1c, T1b * T1e);
|
cannam@167
|
647 }
|
cannam@167
|
648 {
|
cannam@167
|
649 E T1i, T1m, T1f, T1j;
|
cannam@167
|
650 T1i = T1g - T1h;
|
cannam@167
|
651 T1m = T1k + T1l;
|
cannam@167
|
652 T1f = W[10];
|
cannam@167
|
653 T1j = W[11];
|
cannam@167
|
654 cr[WS(rs, 6)] = FNMS(T1j, T1m, T1f * T1i);
|
cannam@167
|
655 ci[WS(rs, 6)] = FMA(T1j, T1i, T1f * T1m);
|
cannam@167
|
656 }
|
cannam@167
|
657 }
|
cannam@167
|
658 {
|
cannam@167
|
659 E T21, T2n, T26, T2q, T1M, T2y, T2m, T2f, T2A, T2r, T2x, T2z;
|
cannam@167
|
660 {
|
cannam@167
|
661 E T1T, T20, T24, T25;
|
cannam@167
|
662 T1T = T1P - T1S;
|
cannam@167
|
663 T20 = T1W - T1Z;
|
cannam@167
|
664 T21 = FMA(KP951056516, T1T, KP587785252 * T20);
|
cannam@167
|
665 T2n = FNMS(KP951056516, T20, KP587785252 * T1T);
|
cannam@167
|
666 T24 = T1u - T1x;
|
cannam@167
|
667 T25 = T1B - T1E;
|
cannam@167
|
668 T26 = FMA(KP951056516, T24, KP587785252 * T25);
|
cannam@167
|
669 T2q = FNMS(KP951056516, T25, KP587785252 * T24);
|
cannam@167
|
670 }
|
cannam@167
|
671 {
|
cannam@167
|
672 E T1G, T1K, T1L, T29, T2d, T2e;
|
cannam@167
|
673 T1G = KP559016994 * (T1y - T1F);
|
cannam@167
|
674 T1K = T1y + T1F;
|
cannam@167
|
675 T1L = FNMS(KP250000000, T1K, T1J);
|
cannam@167
|
676 T1M = T1G + T1L;
|
cannam@167
|
677 T2y = T1J + T1K;
|
cannam@167
|
678 T2m = T1L - T1G;
|
cannam@167
|
679 T29 = KP559016994 * (T27 - T28);
|
cannam@167
|
680 T2d = T27 + T28;
|
cannam@167
|
681 T2e = FNMS(KP250000000, T2d, T2c);
|
cannam@167
|
682 T2f = T29 + T2e;
|
cannam@167
|
683 T2A = T2c + T2d;
|
cannam@167
|
684 T2r = T2e - T29;
|
cannam@167
|
685 }
|
cannam@167
|
686 T2x = W[18];
|
cannam@167
|
687 T2z = W[19];
|
cannam@167
|
688 cr[WS(rs, 10)] = FNMS(T2z, T2A, T2x * T2y);
|
cannam@167
|
689 ci[WS(rs, 10)] = FMA(T2z, T2y, T2x * T2A);
|
cannam@167
|
690 {
|
cannam@167
|
691 E T2u, T2w, T2t, T2v;
|
cannam@167
|
692 T2u = T2m + T2n;
|
cannam@167
|
693 T2w = T2r - T2q;
|
cannam@167
|
694 T2t = W[24];
|
cannam@167
|
695 T2v = W[25];
|
cannam@167
|
696 cr[WS(rs, 13)] = FNMS(T2v, T2w, T2t * T2u);
|
cannam@167
|
697 ci[WS(rs, 13)] = FMA(T2v, T2u, T2t * T2w);
|
cannam@167
|
698 }
|
cannam@167
|
699 {
|
cannam@167
|
700 E T22, T2g, T1r, T23;
|
cannam@167
|
701 T22 = T1M - T21;
|
cannam@167
|
702 T2g = T26 + T2f;
|
cannam@167
|
703 T1r = W[0];
|
cannam@167
|
704 T23 = W[1];
|
cannam@167
|
705 cr[WS(rs, 1)] = FNMS(T23, T2g, T1r * T22);
|
cannam@167
|
706 ci[WS(rs, 1)] = FMA(T23, T22, T1r * T2g);
|
cannam@167
|
707 }
|
cannam@167
|
708 {
|
cannam@167
|
709 E T2i, T2k, T2h, T2j;
|
cannam@167
|
710 T2i = T1M + T21;
|
cannam@167
|
711 T2k = T2f - T26;
|
cannam@167
|
712 T2h = W[6];
|
cannam@167
|
713 T2j = W[7];
|
cannam@167
|
714 cr[WS(rs, 4)] = FNMS(T2j, T2k, T2h * T2i);
|
cannam@167
|
715 ci[WS(rs, 4)] = FMA(T2j, T2i, T2h * T2k);
|
cannam@167
|
716 }
|
cannam@167
|
717 {
|
cannam@167
|
718 E T2o, T2s, T2l, T2p;
|
cannam@167
|
719 T2o = T2m - T2n;
|
cannam@167
|
720 T2s = T2q + T2r;
|
cannam@167
|
721 T2l = W[12];
|
cannam@167
|
722 T2p = W[13];
|
cannam@167
|
723 cr[WS(rs, 7)] = FNMS(T2p, T2s, T2l * T2o);
|
cannam@167
|
724 ci[WS(rs, 7)] = FMA(T2p, T2o, T2l * T2s);
|
cannam@167
|
725 }
|
cannam@167
|
726 }
|
cannam@167
|
727 {
|
cannam@167
|
728 E T31, T3h, T36, T3k, T2K, T3g, T2Y, T2U, T3l, T39, T2B, T2L;
|
cannam@167
|
729 {
|
cannam@167
|
730 E T2Z, T30, T34, T35;
|
cannam@167
|
731 T2Z = T2N + T2O;
|
cannam@167
|
732 T30 = T2Q + T2R;
|
cannam@167
|
733 T31 = FNMS(KP951056516, T30, KP587785252 * T2Z);
|
cannam@167
|
734 T3h = FMA(KP951056516, T2Z, KP587785252 * T30);
|
cannam@167
|
735 T34 = T2D - T2E;
|
cannam@167
|
736 T35 = T2G - T2H;
|
cannam@167
|
737 T36 = FNMS(KP951056516, T35, KP587785252 * T34);
|
cannam@167
|
738 T3k = FMA(KP951056516, T34, KP587785252 * T35);
|
cannam@167
|
739 }
|
cannam@167
|
740 {
|
cannam@167
|
741 E T2X, T2J, T2W, T38, T2T, T37;
|
cannam@167
|
742 T2X = KP559016994 * (T2F - T2I);
|
cannam@167
|
743 T2J = T2F + T2I;
|
cannam@167
|
744 T2W = FNMS(KP250000000, T2J, T2C);
|
cannam@167
|
745 T2K = T2C + T2J;
|
cannam@167
|
746 T3g = T2X + T2W;
|
cannam@167
|
747 T2Y = T2W - T2X;
|
cannam@167
|
748 T38 = KP559016994 * (T2P - T2S);
|
cannam@167
|
749 T2T = T2P + T2S;
|
cannam@167
|
750 T37 = FNMS(KP250000000, T2T, T2M);
|
cannam@167
|
751 T2U = T2M + T2T;
|
cannam@167
|
752 T3l = T38 + T37;
|
cannam@167
|
753 T39 = T37 - T38;
|
cannam@167
|
754 }
|
cannam@167
|
755 T2B = W[8];
|
cannam@167
|
756 T2L = W[9];
|
cannam@167
|
757 cr[WS(rs, 5)] = FNMS(T2L, T2U, T2B * T2K);
|
cannam@167
|
758 ci[WS(rs, 5)] = FMA(T2L, T2K, T2B * T2U);
|
cannam@167
|
759 {
|
cannam@167
|
760 E T3o, T3q, T3n, T3p;
|
cannam@167
|
761 T3o = T3g + T3h;
|
cannam@167
|
762 T3q = T3l - T3k;
|
cannam@167
|
763 T3n = W[26];
|
cannam@167
|
764 T3p = W[27];
|
cannam@167
|
765 cr[WS(rs, 14)] = FNMS(T3p, T3q, T3n * T3o);
|
cannam@167
|
766 ci[WS(rs, 14)] = FMA(T3n, T3q, T3p * T3o);
|
cannam@167
|
767 }
|
cannam@167
|
768 {
|
cannam@167
|
769 E T32, T3a, T2V, T33;
|
cannam@167
|
770 T32 = T2Y - T31;
|
cannam@167
|
771 T3a = T36 + T39;
|
cannam@167
|
772 T2V = W[2];
|
cannam@167
|
773 T33 = W[3];
|
cannam@167
|
774 cr[WS(rs, 2)] = FNMS(T33, T3a, T2V * T32);
|
cannam@167
|
775 ci[WS(rs, 2)] = FMA(T2V, T3a, T33 * T32);
|
cannam@167
|
776 }
|
cannam@167
|
777 {
|
cannam@167
|
778 E T3c, T3e, T3b, T3d;
|
cannam@167
|
779 T3c = T2Y + T31;
|
cannam@167
|
780 T3e = T39 - T36;
|
cannam@167
|
781 T3b = W[14];
|
cannam@167
|
782 T3d = W[15];
|
cannam@167
|
783 cr[WS(rs, 8)] = FNMS(T3d, T3e, T3b * T3c);
|
cannam@167
|
784 ci[WS(rs, 8)] = FMA(T3b, T3e, T3d * T3c);
|
cannam@167
|
785 }
|
cannam@167
|
786 {
|
cannam@167
|
787 E T3i, T3m, T3f, T3j;
|
cannam@167
|
788 T3i = T3g - T3h;
|
cannam@167
|
789 T3m = T3k + T3l;
|
cannam@167
|
790 T3f = W[20];
|
cannam@167
|
791 T3j = W[21];
|
cannam@167
|
792 cr[WS(rs, 11)] = FNMS(T3j, T3m, T3f * T3i);
|
cannam@167
|
793 ci[WS(rs, 11)] = FMA(T3f, T3m, T3j * T3i);
|
cannam@167
|
794 }
|
cannam@167
|
795 }
|
cannam@167
|
796 }
|
cannam@167
|
797 }
|
cannam@167
|
798 }
|
cannam@167
|
799
|
cannam@167
|
800 static const tw_instr twinstr[] = {
|
cannam@167
|
801 {TW_FULL, 1, 15},
|
cannam@167
|
802 {TW_NEXT, 1, 0}
|
cannam@167
|
803 };
|
cannam@167
|
804
|
cannam@167
|
805 static const hc2hc_desc desc = { 15, "hb_15", twinstr, &GENUS, {128, 56, 56, 0} };
|
cannam@167
|
806
|
cannam@167
|
807 void X(codelet_hb_15) (planner *p) {
|
cannam@167
|
808 X(khc2hc_register) (p, hb_15, &desc);
|
cannam@167
|
809 }
|
cannam@167
|
810 #endif
|