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 07:40:00 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 20 -dit -name hf_20 -include hf.h */
|
Chris@10
|
29
|
Chris@10
|
30 /*
|
Chris@10
|
31 * This function contains 246 FP additions, 148 FP multiplications,
|
Chris@10
|
32 * (or, 136 additions, 38 multiplications, 110 fused multiply/add),
|
Chris@10
|
33 * 100 stack variables, 4 constants, and 80 memory accesses
|
Chris@10
|
34 */
|
Chris@10
|
35 #include "hf.h"
|
Chris@10
|
36
|
Chris@10
|
37 static void hf_20(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 {
|
Chris@10
|
44 INT m;
|
Chris@10
|
45 for (m = mb, W = W + ((mb - 1) * 38); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 38, MAKE_VOLATILE_STRIDE(40, rs)) {
|
Chris@10
|
46 E T54, T5a, T5c, T56, T53, T55, T5b, T57;
|
Chris@10
|
47 {
|
Chris@10
|
48 E T4N, T4q, T8, T2i, T4r, T2n, T4O, Tl, T2v, T3v, T43, T4b, TN, T2b, T3F;
|
Chris@10
|
49 E T3a, T2R, T3z, T3T, T4f, T27, T2f, T3J, T3i, T2K, T3y, T3W, T4e, T1G, T2e;
|
Chris@10
|
50 E T3I, T3p, T2C, T3w, T40, T4c, T1e, T2c, T3G, T33;
|
Chris@10
|
51 {
|
Chris@10
|
52 E T1, T4p, T3, T6, T2, T5;
|
Chris@10
|
53 T1 = cr[0];
|
Chris@10
|
54 T4p = ci[0];
|
Chris@10
|
55 T3 = cr[WS(rs, 10)];
|
Chris@10
|
56 T6 = ci[WS(rs, 10)];
|
Chris@10
|
57 T2 = W[18];
|
Chris@10
|
58 T5 = W[19];
|
Chris@10
|
59 {
|
Chris@10
|
60 E Ta, Td, Tg, T2j, Tb, Tj, Tf, Tc, Ti;
|
Chris@10
|
61 {
|
Chris@10
|
62 E T4n, T4, T9, T4o, T7;
|
Chris@10
|
63 Ta = cr[WS(rs, 5)];
|
Chris@10
|
64 Td = ci[WS(rs, 5)];
|
Chris@10
|
65 T4n = T2 * T6;
|
Chris@10
|
66 T4 = T2 * T3;
|
Chris@10
|
67 T9 = W[8];
|
Chris@10
|
68 Tg = cr[WS(rs, 15)];
|
Chris@10
|
69 T4o = FNMS(T5, T3, T4n);
|
Chris@10
|
70 T7 = FMA(T5, T6, T4);
|
Chris@10
|
71 T2j = T9 * Td;
|
Chris@10
|
72 Tb = T9 * Ta;
|
Chris@10
|
73 T4N = T4p - T4o;
|
Chris@10
|
74 T4q = T4o + T4p;
|
Chris@10
|
75 T8 = T1 + T7;
|
Chris@10
|
76 T2i = T1 - T7;
|
Chris@10
|
77 Tj = ci[WS(rs, 15)];
|
Chris@10
|
78 Tf = W[28];
|
Chris@10
|
79 }
|
Chris@10
|
80 Tc = W[9];
|
Chris@10
|
81 Ti = W[29];
|
Chris@10
|
82 {
|
Chris@10
|
83 E T36, Ts, T2t, TL, TB, TE, TD, T38, Ty, T2q, TC;
|
Chris@10
|
84 {
|
Chris@10
|
85 E TH, TK, TJ, T2s, TI;
|
Chris@10
|
86 {
|
Chris@10
|
87 E To, Tr, Tp, T35, Tq, TG;
|
Chris@10
|
88 {
|
Chris@10
|
89 E T2k, Te, T2m, Tk, T2l, Th, Tn;
|
Chris@10
|
90 To = cr[WS(rs, 4)];
|
Chris@10
|
91 T2l = Tf * Tj;
|
Chris@10
|
92 Th = Tf * Tg;
|
Chris@10
|
93 T2k = FNMS(Tc, Ta, T2j);
|
Chris@10
|
94 Te = FMA(Tc, Td, Tb);
|
Chris@10
|
95 T2m = FNMS(Ti, Tg, T2l);
|
Chris@10
|
96 Tk = FMA(Ti, Tj, Th);
|
Chris@10
|
97 Tr = ci[WS(rs, 4)];
|
Chris@10
|
98 Tn = W[6];
|
Chris@10
|
99 T4r = T2k + T2m;
|
Chris@10
|
100 T2n = T2k - T2m;
|
Chris@10
|
101 T4O = Te - Tk;
|
Chris@10
|
102 Tl = Te + Tk;
|
Chris@10
|
103 Tp = Tn * To;
|
Chris@10
|
104 T35 = Tn * Tr;
|
Chris@10
|
105 }
|
Chris@10
|
106 Tq = W[7];
|
Chris@10
|
107 TH = cr[WS(rs, 19)];
|
Chris@10
|
108 TK = ci[WS(rs, 19)];
|
Chris@10
|
109 TG = W[36];
|
Chris@10
|
110 T36 = FNMS(Tq, To, T35);
|
Chris@10
|
111 Ts = FMA(Tq, Tr, Tp);
|
Chris@10
|
112 TJ = W[37];
|
Chris@10
|
113 T2s = TG * TK;
|
Chris@10
|
114 TI = TG * TH;
|
Chris@10
|
115 }
|
Chris@10
|
116 {
|
Chris@10
|
117 E Tu, Tx, Tt, Tw, T37, Tv, TA;
|
Chris@10
|
118 Tu = cr[WS(rs, 14)];
|
Chris@10
|
119 Tx = ci[WS(rs, 14)];
|
Chris@10
|
120 T2t = FNMS(TJ, TH, T2s);
|
Chris@10
|
121 TL = FMA(TJ, TK, TI);
|
Chris@10
|
122 Tt = W[26];
|
Chris@10
|
123 Tw = W[27];
|
Chris@10
|
124 TB = cr[WS(rs, 9)];
|
Chris@10
|
125 TE = ci[WS(rs, 9)];
|
Chris@10
|
126 T37 = Tt * Tx;
|
Chris@10
|
127 Tv = Tt * Tu;
|
Chris@10
|
128 TA = W[16];
|
Chris@10
|
129 TD = W[17];
|
Chris@10
|
130 T38 = FNMS(Tw, Tu, T37);
|
Chris@10
|
131 Ty = FMA(Tw, Tx, Tv);
|
Chris@10
|
132 T2q = TA * TE;
|
Chris@10
|
133 TC = TA * TB;
|
Chris@10
|
134 }
|
Chris@10
|
135 }
|
Chris@10
|
136 {
|
Chris@10
|
137 E T39, T42, Tz, T2p, T2r, TF;
|
Chris@10
|
138 T39 = T36 - T38;
|
Chris@10
|
139 T42 = T36 + T38;
|
Chris@10
|
140 Tz = Ts + Ty;
|
Chris@10
|
141 T2p = Ts - Ty;
|
Chris@10
|
142 T2r = FNMS(TD, TB, T2q);
|
Chris@10
|
143 TF = FMA(TD, TE, TC);
|
Chris@10
|
144 {
|
Chris@10
|
145 E T41, T2u, TM, T34;
|
Chris@10
|
146 T41 = T2r + T2t;
|
Chris@10
|
147 T2u = T2r - T2t;
|
Chris@10
|
148 TM = TF + TL;
|
Chris@10
|
149 T34 = TL - TF;
|
Chris@10
|
150 T2v = T2p - T2u;
|
Chris@10
|
151 T3v = T2p + T2u;
|
Chris@10
|
152 T43 = T41 - T42;
|
Chris@10
|
153 T4b = T42 + T41;
|
Chris@10
|
154 TN = Tz - TM;
|
Chris@10
|
155 T2b = Tz + TM;
|
Chris@10
|
156 T3F = T39 + T34;
|
Chris@10
|
157 T3a = T34 - T39;
|
Chris@10
|
158 }
|
Chris@10
|
159 }
|
Chris@10
|
160 }
|
Chris@10
|
161 }
|
Chris@10
|
162 }
|
Chris@10
|
163 {
|
Chris@10
|
164 E T3e, T1M, T2P, T25, T1V, T1Y, T1X, T3g, T1S, T2M, T1W;
|
Chris@10
|
165 {
|
Chris@10
|
166 E T21, T24, T23, T2O, T22;
|
Chris@10
|
167 {
|
Chris@10
|
168 E T1I, T1L, T1H, T1K, T3d, T1J, T20;
|
Chris@10
|
169 T1I = cr[WS(rs, 12)];
|
Chris@10
|
170 T1L = ci[WS(rs, 12)];
|
Chris@10
|
171 T1H = W[22];
|
Chris@10
|
172 T1K = W[23];
|
Chris@10
|
173 T21 = cr[WS(rs, 7)];
|
Chris@10
|
174 T24 = ci[WS(rs, 7)];
|
Chris@10
|
175 T3d = T1H * T1L;
|
Chris@10
|
176 T1J = T1H * T1I;
|
Chris@10
|
177 T20 = W[12];
|
Chris@10
|
178 T23 = W[13];
|
Chris@10
|
179 T3e = FNMS(T1K, T1I, T3d);
|
Chris@10
|
180 T1M = FMA(T1K, T1L, T1J);
|
Chris@10
|
181 T2O = T20 * T24;
|
Chris@10
|
182 T22 = T20 * T21;
|
Chris@10
|
183 }
|
Chris@10
|
184 {
|
Chris@10
|
185 E T1O, T1R, T1N, T1Q, T3f, T1P, T1U;
|
Chris@10
|
186 T1O = cr[WS(rs, 2)];
|
Chris@10
|
187 T1R = ci[WS(rs, 2)];
|
Chris@10
|
188 T2P = FNMS(T23, T21, T2O);
|
Chris@10
|
189 T25 = FMA(T23, T24, T22);
|
Chris@10
|
190 T1N = W[2];
|
Chris@10
|
191 T1Q = W[3];
|
Chris@10
|
192 T1V = cr[WS(rs, 17)];
|
Chris@10
|
193 T1Y = ci[WS(rs, 17)];
|
Chris@10
|
194 T3f = T1N * T1R;
|
Chris@10
|
195 T1P = T1N * T1O;
|
Chris@10
|
196 T1U = W[32];
|
Chris@10
|
197 T1X = W[33];
|
Chris@10
|
198 T3g = FNMS(T1Q, T1O, T3f);
|
Chris@10
|
199 T1S = FMA(T1Q, T1R, T1P);
|
Chris@10
|
200 T2M = T1U * T1Y;
|
Chris@10
|
201 T1W = T1U * T1V;
|
Chris@10
|
202 }
|
Chris@10
|
203 }
|
Chris@10
|
204 {
|
Chris@10
|
205 E T3h, T3S, T1T, T2L, T2N, T1Z;
|
Chris@10
|
206 T3h = T3e - T3g;
|
Chris@10
|
207 T3S = T3e + T3g;
|
Chris@10
|
208 T1T = T1M + T1S;
|
Chris@10
|
209 T2L = T1M - T1S;
|
Chris@10
|
210 T2N = FNMS(T1X, T1V, T2M);
|
Chris@10
|
211 T1Z = FMA(T1X, T1Y, T1W);
|
Chris@10
|
212 {
|
Chris@10
|
213 E T3R, T2Q, T26, T3c;
|
Chris@10
|
214 T3R = T2N + T2P;
|
Chris@10
|
215 T2Q = T2N - T2P;
|
Chris@10
|
216 T26 = T1Z + T25;
|
Chris@10
|
217 T3c = T25 - T1Z;
|
Chris@10
|
218 T2R = T2L - T2Q;
|
Chris@10
|
219 T3z = T2L + T2Q;
|
Chris@10
|
220 T3T = T3R - T3S;
|
Chris@10
|
221 T4f = T3S + T3R;
|
Chris@10
|
222 T27 = T1T - T26;
|
Chris@10
|
223 T2f = T1T + T26;
|
Chris@10
|
224 T3J = T3h + T3c;
|
Chris@10
|
225 T3i = T3c - T3h;
|
Chris@10
|
226 }
|
Chris@10
|
227 }
|
Chris@10
|
228 }
|
Chris@10
|
229 {
|
Chris@10
|
230 E T3l, T1l, T2I, T1E, T1u, T1x, T1w, T3n, T1r, T2F, T1v;
|
Chris@10
|
231 {
|
Chris@10
|
232 E T1A, T1D, T1C, T2H, T1B;
|
Chris@10
|
233 {
|
Chris@10
|
234 E T1h, T1k, T1g, T1j, T3k, T1i, T1z;
|
Chris@10
|
235 T1h = cr[WS(rs, 8)];
|
Chris@10
|
236 T1k = ci[WS(rs, 8)];
|
Chris@10
|
237 T1g = W[14];
|
Chris@10
|
238 T1j = W[15];
|
Chris@10
|
239 T1A = cr[WS(rs, 3)];
|
Chris@10
|
240 T1D = ci[WS(rs, 3)];
|
Chris@10
|
241 T3k = T1g * T1k;
|
Chris@10
|
242 T1i = T1g * T1h;
|
Chris@10
|
243 T1z = W[4];
|
Chris@10
|
244 T1C = W[5];
|
Chris@10
|
245 T3l = FNMS(T1j, T1h, T3k);
|
Chris@10
|
246 T1l = FMA(T1j, T1k, T1i);
|
Chris@10
|
247 T2H = T1z * T1D;
|
Chris@10
|
248 T1B = T1z * T1A;
|
Chris@10
|
249 }
|
Chris@10
|
250 {
|
Chris@10
|
251 E T1n, T1q, T1m, T1p, T3m, T1o, T1t;
|
Chris@10
|
252 T1n = cr[WS(rs, 18)];
|
Chris@10
|
253 T1q = ci[WS(rs, 18)];
|
Chris@10
|
254 T2I = FNMS(T1C, T1A, T2H);
|
Chris@10
|
255 T1E = FMA(T1C, T1D, T1B);
|
Chris@10
|
256 T1m = W[34];
|
Chris@10
|
257 T1p = W[35];
|
Chris@10
|
258 T1u = cr[WS(rs, 13)];
|
Chris@10
|
259 T1x = ci[WS(rs, 13)];
|
Chris@10
|
260 T3m = T1m * T1q;
|
Chris@10
|
261 T1o = T1m * T1n;
|
Chris@10
|
262 T1t = W[24];
|
Chris@10
|
263 T1w = W[25];
|
Chris@10
|
264 T3n = FNMS(T1p, T1n, T3m);
|
Chris@10
|
265 T1r = FMA(T1p, T1q, T1o);
|
Chris@10
|
266 T2F = T1t * T1x;
|
Chris@10
|
267 T1v = T1t * T1u;
|
Chris@10
|
268 }
|
Chris@10
|
269 }
|
Chris@10
|
270 {
|
Chris@10
|
271 E T3o, T3V, T1s, T2E, T2G, T1y;
|
Chris@10
|
272 T3o = T3l - T3n;
|
Chris@10
|
273 T3V = T3l + T3n;
|
Chris@10
|
274 T1s = T1l + T1r;
|
Chris@10
|
275 T2E = T1l - T1r;
|
Chris@10
|
276 T2G = FNMS(T1w, T1u, T2F);
|
Chris@10
|
277 T1y = FMA(T1w, T1x, T1v);
|
Chris@10
|
278 {
|
Chris@10
|
279 E T3U, T2J, T1F, T3j;
|
Chris@10
|
280 T3U = T2G + T2I;
|
Chris@10
|
281 T2J = T2G - T2I;
|
Chris@10
|
282 T1F = T1y + T1E;
|
Chris@10
|
283 T3j = T1E - T1y;
|
Chris@10
|
284 T2K = T2E - T2J;
|
Chris@10
|
285 T3y = T2E + T2J;
|
Chris@10
|
286 T3W = T3U - T3V;
|
Chris@10
|
287 T4e = T3V + T3U;
|
Chris@10
|
288 T1G = T1s - T1F;
|
Chris@10
|
289 T2e = T1s + T1F;
|
Chris@10
|
290 T3I = T3o + T3j;
|
Chris@10
|
291 T3p = T3j - T3o;
|
Chris@10
|
292 }
|
Chris@10
|
293 }
|
Chris@10
|
294 }
|
Chris@10
|
295 {
|
Chris@10
|
296 E T2Z, TT, T2A, T1c, T12, T15, T14, T31, TZ, T2x, T13;
|
Chris@10
|
297 {
|
Chris@10
|
298 E T18, T1b, T1a, T2z, T19;
|
Chris@10
|
299 {
|
Chris@10
|
300 E TP, TS, TO, TR, T2Y, TQ, T17;
|
Chris@10
|
301 TP = cr[WS(rs, 16)];
|
Chris@10
|
302 TS = ci[WS(rs, 16)];
|
Chris@10
|
303 TO = W[30];
|
Chris@10
|
304 TR = W[31];
|
Chris@10
|
305 T18 = cr[WS(rs, 11)];
|
Chris@10
|
306 T1b = ci[WS(rs, 11)];
|
Chris@10
|
307 T2Y = TO * TS;
|
Chris@10
|
308 TQ = TO * TP;
|
Chris@10
|
309 T17 = W[20];
|
Chris@10
|
310 T1a = W[21];
|
Chris@10
|
311 T2Z = FNMS(TR, TP, T2Y);
|
Chris@10
|
312 TT = FMA(TR, TS, TQ);
|
Chris@10
|
313 T2z = T17 * T1b;
|
Chris@10
|
314 T19 = T17 * T18;
|
Chris@10
|
315 }
|
Chris@10
|
316 {
|
Chris@10
|
317 E TV, TY, TU, TX, T30, TW, T11;
|
Chris@10
|
318 TV = cr[WS(rs, 6)];
|
Chris@10
|
319 TY = ci[WS(rs, 6)];
|
Chris@10
|
320 T2A = FNMS(T1a, T18, T2z);
|
Chris@10
|
321 T1c = FMA(T1a, T1b, T19);
|
Chris@10
|
322 TU = W[10];
|
Chris@10
|
323 TX = W[11];
|
Chris@10
|
324 T12 = cr[WS(rs, 1)];
|
Chris@10
|
325 T15 = ci[WS(rs, 1)];
|
Chris@10
|
326 T30 = TU * TY;
|
Chris@10
|
327 TW = TU * TV;
|
Chris@10
|
328 T11 = W[0];
|
Chris@10
|
329 T14 = W[1];
|
Chris@10
|
330 T31 = FNMS(TX, TV, T30);
|
Chris@10
|
331 TZ = FMA(TX, TY, TW);
|
Chris@10
|
332 T2x = T11 * T15;
|
Chris@10
|
333 T13 = T11 * T12;
|
Chris@10
|
334 }
|
Chris@10
|
335 }
|
Chris@10
|
336 {
|
Chris@10
|
337 E T32, T3Z, T10, T2w, T2y, T16;
|
Chris@10
|
338 T32 = T2Z - T31;
|
Chris@10
|
339 T3Z = T2Z + T31;
|
Chris@10
|
340 T10 = TT + TZ;
|
Chris@10
|
341 T2w = TT - TZ;
|
Chris@10
|
342 T2y = FNMS(T14, T12, T2x);
|
Chris@10
|
343 T16 = FMA(T14, T15, T13);
|
Chris@10
|
344 {
|
Chris@10
|
345 E T3Y, T2B, T1d, T2X;
|
Chris@10
|
346 T3Y = T2y + T2A;
|
Chris@10
|
347 T2B = T2y - T2A;
|
Chris@10
|
348 T1d = T16 + T1c;
|
Chris@10
|
349 T2X = T1c - T16;
|
Chris@10
|
350 T2C = T2w - T2B;
|
Chris@10
|
351 T3w = T2w + T2B;
|
Chris@10
|
352 T40 = T3Y - T3Z;
|
Chris@10
|
353 T4c = T3Z + T3Y;
|
Chris@10
|
354 T1e = T10 - T1d;
|
Chris@10
|
355 T2c = T10 + T1d;
|
Chris@10
|
356 T3G = T32 + T2X;
|
Chris@10
|
357 T33 = T2X - T32;
|
Chris@10
|
358 }
|
Chris@10
|
359 }
|
Chris@10
|
360 }
|
Chris@10
|
361 {
|
Chris@10
|
362 E T4l, T4k, T4w, T4x, T4Q, T4R, T2o, T4X, T4W, T4C, T4D, T4J, T4h, T4j, T4I;
|
Chris@10
|
363 E T51, T52, T49, T3r, T3t, T58, T2D, T48, T2S, T59;
|
Chris@10
|
364 {
|
Chris@10
|
365 E T2a, T47, T45, T3u, T3x, T3N, T3L, T3A, T46, T3Q;
|
Chris@10
|
366 {
|
Chris@10
|
367 E Tm, T1f, T28, T3X, T44;
|
Chris@10
|
368 T4l = T3W + T3T;
|
Chris@10
|
369 T3X = T3T - T3W;
|
Chris@10
|
370 T44 = T40 - T43;
|
Chris@10
|
371 T4k = T43 + T40;
|
Chris@10
|
372 T2a = T8 + Tl;
|
Chris@10
|
373 Tm = T8 - Tl;
|
Chris@10
|
374 T1f = TN + T1e;
|
Chris@10
|
375 T4w = T1e - TN;
|
Chris@10
|
376 T4x = T1G - T27;
|
Chris@10
|
377 T28 = T1G + T27;
|
Chris@10
|
378 T47 = FMA(KP618033988, T3X, T44);
|
Chris@10
|
379 T45 = FNMS(KP618033988, T44, T3X);
|
Chris@10
|
380 {
|
Chris@10
|
381 E T3H, T29, T3P, T3K, T3O;
|
Chris@10
|
382 T3H = T3F - T3G;
|
Chris@10
|
383 T4Q = T3F + T3G;
|
Chris@10
|
384 T29 = T1f + T28;
|
Chris@10
|
385 T3P = T1f - T28;
|
Chris@10
|
386 T4R = T3I + T3J;
|
Chris@10
|
387 T3K = T3I - T3J;
|
Chris@10
|
388 T3u = T2i + T2n;
|
Chris@10
|
389 T2o = T2i - T2n;
|
Chris@10
|
390 T4X = T3v - T3w;
|
Chris@10
|
391 T3x = T3v + T3w;
|
Chris@10
|
392 ci[WS(rs, 9)] = Tm + T29;
|
Chris@10
|
393 T3O = FNMS(KP250000000, T29, Tm);
|
Chris@10
|
394 T3N = FNMS(KP618033988, T3H, T3K);
|
Chris@10
|
395 T3L = FMA(KP618033988, T3K, T3H);
|
Chris@10
|
396 T3A = T3y + T3z;
|
Chris@10
|
397 T4W = T3y - T3z;
|
Chris@10
|
398 T46 = FMA(KP559016994, T3P, T3O);
|
Chris@10
|
399 T3Q = FNMS(KP559016994, T3P, T3O);
|
Chris@10
|
400 }
|
Chris@10
|
401 }
|
Chris@10
|
402 {
|
Chris@10
|
403 E T2d, T2g, T3b, T3q, T2h;
|
Chris@10
|
404 {
|
Chris@10
|
405 E T4d, T3D, T3C, T4g, T3B, T3M, T3E;
|
Chris@10
|
406 T4C = T4b + T4c;
|
Chris@10
|
407 T4d = T4b - T4c;
|
Chris@10
|
408 T3D = T3x - T3A;
|
Chris@10
|
409 T3B = T3x + T3A;
|
Chris@10
|
410 ci[WS(rs, 1)] = FMA(KP951056516, T45, T3Q);
|
Chris@10
|
411 cr[WS(rs, 2)] = FNMS(KP951056516, T45, T3Q);
|
Chris@10
|
412 cr[WS(rs, 6)] = FMA(KP951056516, T47, T46);
|
Chris@10
|
413 ci[WS(rs, 5)] = FNMS(KP951056516, T47, T46);
|
Chris@10
|
414 cr[WS(rs, 5)] = T3u + T3B;
|
Chris@10
|
415 T3C = FNMS(KP250000000, T3B, T3u);
|
Chris@10
|
416 T4g = T4e - T4f;
|
Chris@10
|
417 T4D = T4e + T4f;
|
Chris@10
|
418 T2d = T2b + T2c;
|
Chris@10
|
419 T4J = T2b - T2c;
|
Chris@10
|
420 T3M = FNMS(KP559016994, T3D, T3C);
|
Chris@10
|
421 T3E = FMA(KP559016994, T3D, T3C);
|
Chris@10
|
422 T4h = FMA(KP618033988, T4g, T4d);
|
Chris@10
|
423 T4j = FNMS(KP618033988, T4d, T4g);
|
Chris@10
|
424 cr[WS(rs, 9)] = FNMS(KP951056516, T3L, T3E);
|
Chris@10
|
425 cr[WS(rs, 1)] = FMA(KP951056516, T3L, T3E);
|
Chris@10
|
426 ci[WS(rs, 6)] = FMA(KP951056516, T3N, T3M);
|
Chris@10
|
427 ci[WS(rs, 2)] = FNMS(KP951056516, T3N, T3M);
|
Chris@10
|
428 T4I = T2f - T2e;
|
Chris@10
|
429 T2g = T2e + T2f;
|
Chris@10
|
430 }
|
Chris@10
|
431 T3b = T33 - T3a;
|
Chris@10
|
432 T51 = T3a + T33;
|
Chris@10
|
433 T52 = T3p + T3i;
|
Chris@10
|
434 T3q = T3i - T3p;
|
Chris@10
|
435 T2h = T2d + T2g;
|
Chris@10
|
436 T49 = T2d - T2g;
|
Chris@10
|
437 T3r = FMA(KP618033988, T3q, T3b);
|
Chris@10
|
438 T3t = FNMS(KP618033988, T3b, T3q);
|
Chris@10
|
439 T58 = T2v - T2C;
|
Chris@10
|
440 T2D = T2v + T2C;
|
Chris@10
|
441 cr[0] = T2a + T2h;
|
Chris@10
|
442 T48 = FNMS(KP250000000, T2h, T2a);
|
Chris@10
|
443 T2S = T2K + T2R;
|
Chris@10
|
444 T59 = T2K - T2R;
|
Chris@10
|
445 }
|
Chris@10
|
446 }
|
Chris@10
|
447 {
|
Chris@10
|
448 E T4B, T4P, T4Y, T50, T4U, T4S;
|
Chris@10
|
449 {
|
Chris@10
|
450 E T4A, T4y, T4s, T4m, T4u, T4t, T4z, T4v;
|
Chris@10
|
451 {
|
Chris@10
|
452 E T2V, T2U, T4i, T4a, T2T, T2W, T3s;
|
Chris@10
|
453 T4i = FNMS(KP559016994, T49, T48);
|
Chris@10
|
454 T4a = FMA(KP559016994, T49, T48);
|
Chris@10
|
455 T2T = T2D + T2S;
|
Chris@10
|
456 T2V = T2D - T2S;
|
Chris@10
|
457 ci[WS(rs, 3)] = FMA(KP951056516, T4h, T4a);
|
Chris@10
|
458 cr[WS(rs, 4)] = FNMS(KP951056516, T4h, T4a);
|
Chris@10
|
459 cr[WS(rs, 8)] = FMA(KP951056516, T4j, T4i);
|
Chris@10
|
460 ci[WS(rs, 7)] = FNMS(KP951056516, T4j, T4i);
|
Chris@10
|
461 ci[WS(rs, 4)] = T2o + T2T;
|
Chris@10
|
462 T2U = FNMS(KP250000000, T2T, T2o);
|
Chris@10
|
463 T4A = FMA(KP618033988, T4w, T4x);
|
Chris@10
|
464 T4y = FNMS(KP618033988, T4x, T4w);
|
Chris@10
|
465 T4B = T4r + T4q;
|
Chris@10
|
466 T4s = T4q - T4r;
|
Chris@10
|
467 T2W = FMA(KP559016994, T2V, T2U);
|
Chris@10
|
468 T3s = FNMS(KP559016994, T2V, T2U);
|
Chris@10
|
469 ci[WS(rs, 8)] = FMA(KP951056516, T3r, T2W);
|
Chris@10
|
470 ci[0] = FNMS(KP951056516, T3r, T2W);
|
Chris@10
|
471 cr[WS(rs, 7)] = FNMS(KP951056516, T3t, T3s);
|
Chris@10
|
472 cr[WS(rs, 3)] = FMA(KP951056516, T3t, T3s);
|
Chris@10
|
473 T4m = T4k + T4l;
|
Chris@10
|
474 T4u = T4l - T4k;
|
Chris@10
|
475 }
|
Chris@10
|
476 cr[WS(rs, 10)] = T4m - T4s;
|
Chris@10
|
477 T4t = FMA(KP250000000, T4m, T4s);
|
Chris@10
|
478 T4P = T4N - T4O;
|
Chris@10
|
479 T54 = T4O + T4N;
|
Chris@10
|
480 T4Y = FNMS(KP618033988, T4X, T4W);
|
Chris@10
|
481 T50 = FMA(KP618033988, T4W, T4X);
|
Chris@10
|
482 T4z = FNMS(KP559016994, T4u, T4t);
|
Chris@10
|
483 T4v = FMA(KP559016994, T4u, T4t);
|
Chris@10
|
484 ci[WS(rs, 13)] = FMA(KP951056516, T4y, T4v);
|
Chris@10
|
485 cr[WS(rs, 14)] = FMS(KP951056516, T4y, T4v);
|
Chris@10
|
486 ci[WS(rs, 17)] = FMA(KP951056516, T4A, T4z);
|
Chris@10
|
487 cr[WS(rs, 18)] = FMS(KP951056516, T4A, T4z);
|
Chris@10
|
488 T4U = T4Q - T4R;
|
Chris@10
|
489 T4S = T4Q + T4R;
|
Chris@10
|
490 }
|
Chris@10
|
491 {
|
Chris@10
|
492 E T4M, T4K, T4E, T4G, T4T, T4V, T4Z, T4F, T4L, T4H;
|
Chris@10
|
493 ci[WS(rs, 14)] = T4S + T4P;
|
Chris@10
|
494 T4T = FNMS(KP250000000, T4S, T4P);
|
Chris@10
|
495 T4M = FNMS(KP618033988, T4I, T4J);
|
Chris@10
|
496 T4K = FMA(KP618033988, T4J, T4I);
|
Chris@10
|
497 T4V = FNMS(KP559016994, T4U, T4T);
|
Chris@10
|
498 T4Z = FMA(KP559016994, T4U, T4T);
|
Chris@10
|
499 cr[WS(rs, 17)] = -(FMA(KP951056516, T4Y, T4V));
|
Chris@10
|
500 cr[WS(rs, 13)] = FMS(KP951056516, T4Y, T4V);
|
Chris@10
|
501 ci[WS(rs, 18)] = FNMS(KP951056516, T50, T4Z);
|
Chris@10
|
502 ci[WS(rs, 10)] = FMA(KP951056516, T50, T4Z);
|
Chris@10
|
503 T4E = T4C + T4D;
|
Chris@10
|
504 T4G = T4C - T4D;
|
Chris@10
|
505 ci[WS(rs, 19)] = T4E + T4B;
|
Chris@10
|
506 T4F = FNMS(KP250000000, T4E, T4B);
|
Chris@10
|
507 T5a = FMA(KP618033988, T59, T58);
|
Chris@10
|
508 T5c = FNMS(KP618033988, T58, T59);
|
Chris@10
|
509 T4L = FMA(KP559016994, T4G, T4F);
|
Chris@10
|
510 T4H = FNMS(KP559016994, T4G, T4F);
|
Chris@10
|
511 ci[WS(rs, 11)] = FMA(KP951056516, T4K, T4H);
|
Chris@10
|
512 cr[WS(rs, 12)] = FMS(KP951056516, T4K, T4H);
|
Chris@10
|
513 ci[WS(rs, 15)] = FMA(KP951056516, T4M, T4L);
|
Chris@10
|
514 cr[WS(rs, 16)] = FMS(KP951056516, T4M, T4L);
|
Chris@10
|
515 T56 = T51 - T52;
|
Chris@10
|
516 T53 = T51 + T52;
|
Chris@10
|
517 }
|
Chris@10
|
518 }
|
Chris@10
|
519 }
|
Chris@10
|
520 }
|
Chris@10
|
521 cr[WS(rs, 15)] = T53 - T54;
|
Chris@10
|
522 T55 = FMA(KP250000000, T53, T54);
|
Chris@10
|
523 T5b = FMA(KP559016994, T56, T55);
|
Chris@10
|
524 T57 = FNMS(KP559016994, T56, T55);
|
Chris@10
|
525 cr[WS(rs, 19)] = -(FMA(KP951056516, T5a, T57));
|
Chris@10
|
526 cr[WS(rs, 11)] = FMS(KP951056516, T5a, T57);
|
Chris@10
|
527 ci[WS(rs, 16)] = FNMS(KP951056516, T5c, T5b);
|
Chris@10
|
528 ci[WS(rs, 12)] = FMA(KP951056516, T5c, T5b);
|
Chris@10
|
529 }
|
Chris@10
|
530 }
|
Chris@10
|
531 }
|
Chris@10
|
532
|
Chris@10
|
533 static const tw_instr twinstr[] = {
|
Chris@10
|
534 {TW_FULL, 1, 20},
|
Chris@10
|
535 {TW_NEXT, 1, 0}
|
Chris@10
|
536 };
|
Chris@10
|
537
|
Chris@10
|
538 static const hc2hc_desc desc = { 20, "hf_20", twinstr, &GENUS, {136, 38, 110, 0} };
|
Chris@10
|
539
|
Chris@10
|
540 void X(codelet_hf_20) (planner *p) {
|
Chris@10
|
541 X(khc2hc_register) (p, hf_20, &desc);
|
Chris@10
|
542 }
|
Chris@10
|
543 #else /* HAVE_FMA */
|
Chris@10
|
544
|
Chris@10
|
545 /* Generated by: ../../../genfft/gen_hc2hc.native -compact -variables 4 -pipeline-latency 4 -n 20 -dit -name hf_20 -include hf.h */
|
Chris@10
|
546
|
Chris@10
|
547 /*
|
Chris@10
|
548 * This function contains 246 FP additions, 124 FP multiplications,
|
Chris@10
|
549 * (or, 184 additions, 62 multiplications, 62 fused multiply/add),
|
Chris@10
|
550 * 85 stack variables, 4 constants, and 80 memory accesses
|
Chris@10
|
551 */
|
Chris@10
|
552 #include "hf.h"
|
Chris@10
|
553
|
Chris@10
|
554 static void hf_20(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
|
Chris@10
|
555 {
|
Chris@10
|
556 DK(KP587785252, +0.587785252292473129168705954639072768597652438);
|
Chris@10
|
557 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
|
Chris@10
|
558 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
|
Chris@10
|
559 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
|
Chris@10
|
560 {
|
Chris@10
|
561 INT m;
|
Chris@10
|
562 for (m = mb, W = W + ((mb - 1) * 38); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 38, MAKE_VOLATILE_STRIDE(40, rs)) {
|
Chris@10
|
563 E Tj, T1R, T4j, T4s, T2q, T37, T3Q, T42, T1r, T1O, T1P, T3i, T3l, T3J, T3D;
|
Chris@10
|
564 E T3E, T44, T1V, T1W, T1X, T2e, T2j, T2k, T2W, T2X, T4f, T33, T34, T35, T2J;
|
Chris@10
|
565 E T2O, T4q, TG, T13, T14, T3p, T3s, T3K, T3A, T3B, T43, T1S, T1T, T1U, T23;
|
Chris@10
|
566 E T28, T29, T2T, T2U, T4e, T30, T31, T32, T2y, T2D, T4p;
|
Chris@10
|
567 {
|
Chris@10
|
568 E T1, T3N, T6, T3M, Tc, T2n, Th, T2o;
|
Chris@10
|
569 T1 = cr[0];
|
Chris@10
|
570 T3N = ci[0];
|
Chris@10
|
571 {
|
Chris@10
|
572 E T3, T5, T2, T4;
|
Chris@10
|
573 T3 = cr[WS(rs, 10)];
|
Chris@10
|
574 T5 = ci[WS(rs, 10)];
|
Chris@10
|
575 T2 = W[18];
|
Chris@10
|
576 T4 = W[19];
|
Chris@10
|
577 T6 = FMA(T2, T3, T4 * T5);
|
Chris@10
|
578 T3M = FNMS(T4, T3, T2 * T5);
|
Chris@10
|
579 }
|
Chris@10
|
580 {
|
Chris@10
|
581 E T9, Tb, T8, Ta;
|
Chris@10
|
582 T9 = cr[WS(rs, 5)];
|
Chris@10
|
583 Tb = ci[WS(rs, 5)];
|
Chris@10
|
584 T8 = W[8];
|
Chris@10
|
585 Ta = W[9];
|
Chris@10
|
586 Tc = FMA(T8, T9, Ta * Tb);
|
Chris@10
|
587 T2n = FNMS(Ta, T9, T8 * Tb);
|
Chris@10
|
588 }
|
Chris@10
|
589 {
|
Chris@10
|
590 E Te, Tg, Td, Tf;
|
Chris@10
|
591 Te = cr[WS(rs, 15)];
|
Chris@10
|
592 Tg = ci[WS(rs, 15)];
|
Chris@10
|
593 Td = W[28];
|
Chris@10
|
594 Tf = W[29];
|
Chris@10
|
595 Th = FMA(Td, Te, Tf * Tg);
|
Chris@10
|
596 T2o = FNMS(Tf, Te, Td * Tg);
|
Chris@10
|
597 }
|
Chris@10
|
598 {
|
Chris@10
|
599 E T7, Ti, T4h, T4i;
|
Chris@10
|
600 T7 = T1 + T6;
|
Chris@10
|
601 Ti = Tc + Th;
|
Chris@10
|
602 Tj = T7 - Ti;
|
Chris@10
|
603 T1R = T7 + Ti;
|
Chris@10
|
604 T4h = T3N - T3M;
|
Chris@10
|
605 T4i = Tc - Th;
|
Chris@10
|
606 T4j = T4h - T4i;
|
Chris@10
|
607 T4s = T4i + T4h;
|
Chris@10
|
608 }
|
Chris@10
|
609 {
|
Chris@10
|
610 E T2m, T2p, T3O, T3P;
|
Chris@10
|
611 T2m = T1 - T6;
|
Chris@10
|
612 T2p = T2n - T2o;
|
Chris@10
|
613 T2q = T2m - T2p;
|
Chris@10
|
614 T37 = T2m + T2p;
|
Chris@10
|
615 T3O = T3M + T3N;
|
Chris@10
|
616 T3P = T2n + T2o;
|
Chris@10
|
617 T3Q = T3O - T3P;
|
Chris@10
|
618 T42 = T3P + T3O;
|
Chris@10
|
619 }
|
Chris@10
|
620 }
|
Chris@10
|
621 {
|
Chris@10
|
622 E T1f, T3g, T2a, T2H, T1N, T3j, T2i, T2N, T1q, T3h, T2d, T2I, T1C, T3k, T2f;
|
Chris@10
|
623 E T2M;
|
Chris@10
|
624 {
|
Chris@10
|
625 E T19, T2F, T1e, T2G;
|
Chris@10
|
626 {
|
Chris@10
|
627 E T16, T18, T15, T17;
|
Chris@10
|
628 T16 = cr[WS(rs, 8)];
|
Chris@10
|
629 T18 = ci[WS(rs, 8)];
|
Chris@10
|
630 T15 = W[14];
|
Chris@10
|
631 T17 = W[15];
|
Chris@10
|
632 T19 = FMA(T15, T16, T17 * T18);
|
Chris@10
|
633 T2F = FNMS(T17, T16, T15 * T18);
|
Chris@10
|
634 }
|
Chris@10
|
635 {
|
Chris@10
|
636 E T1b, T1d, T1a, T1c;
|
Chris@10
|
637 T1b = cr[WS(rs, 18)];
|
Chris@10
|
638 T1d = ci[WS(rs, 18)];
|
Chris@10
|
639 T1a = W[34];
|
Chris@10
|
640 T1c = W[35];
|
Chris@10
|
641 T1e = FMA(T1a, T1b, T1c * T1d);
|
Chris@10
|
642 T2G = FNMS(T1c, T1b, T1a * T1d);
|
Chris@10
|
643 }
|
Chris@10
|
644 T1f = T19 + T1e;
|
Chris@10
|
645 T3g = T2F + T2G;
|
Chris@10
|
646 T2a = T19 - T1e;
|
Chris@10
|
647 T2H = T2F - T2G;
|
Chris@10
|
648 }
|
Chris@10
|
649 {
|
Chris@10
|
650 E T1H, T2g, T1M, T2h;
|
Chris@10
|
651 {
|
Chris@10
|
652 E T1E, T1G, T1D, T1F;
|
Chris@10
|
653 T1E = cr[WS(rs, 17)];
|
Chris@10
|
654 T1G = ci[WS(rs, 17)];
|
Chris@10
|
655 T1D = W[32];
|
Chris@10
|
656 T1F = W[33];
|
Chris@10
|
657 T1H = FMA(T1D, T1E, T1F * T1G);
|
Chris@10
|
658 T2g = FNMS(T1F, T1E, T1D * T1G);
|
Chris@10
|
659 }
|
Chris@10
|
660 {
|
Chris@10
|
661 E T1J, T1L, T1I, T1K;
|
Chris@10
|
662 T1J = cr[WS(rs, 7)];
|
Chris@10
|
663 T1L = ci[WS(rs, 7)];
|
Chris@10
|
664 T1I = W[12];
|
Chris@10
|
665 T1K = W[13];
|
Chris@10
|
666 T1M = FMA(T1I, T1J, T1K * T1L);
|
Chris@10
|
667 T2h = FNMS(T1K, T1J, T1I * T1L);
|
Chris@10
|
668 }
|
Chris@10
|
669 T1N = T1H + T1M;
|
Chris@10
|
670 T3j = T2g + T2h;
|
Chris@10
|
671 T2i = T2g - T2h;
|
Chris@10
|
672 T2N = T1H - T1M;
|
Chris@10
|
673 }
|
Chris@10
|
674 {
|
Chris@10
|
675 E T1k, T2b, T1p, T2c;
|
Chris@10
|
676 {
|
Chris@10
|
677 E T1h, T1j, T1g, T1i;
|
Chris@10
|
678 T1h = cr[WS(rs, 13)];
|
Chris@10
|
679 T1j = ci[WS(rs, 13)];
|
Chris@10
|
680 T1g = W[24];
|
Chris@10
|
681 T1i = W[25];
|
Chris@10
|
682 T1k = FMA(T1g, T1h, T1i * T1j);
|
Chris@10
|
683 T2b = FNMS(T1i, T1h, T1g * T1j);
|
Chris@10
|
684 }
|
Chris@10
|
685 {
|
Chris@10
|
686 E T1m, T1o, T1l, T1n;
|
Chris@10
|
687 T1m = cr[WS(rs, 3)];
|
Chris@10
|
688 T1o = ci[WS(rs, 3)];
|
Chris@10
|
689 T1l = W[4];
|
Chris@10
|
690 T1n = W[5];
|
Chris@10
|
691 T1p = FMA(T1l, T1m, T1n * T1o);
|
Chris@10
|
692 T2c = FNMS(T1n, T1m, T1l * T1o);
|
Chris@10
|
693 }
|
Chris@10
|
694 T1q = T1k + T1p;
|
Chris@10
|
695 T3h = T2b + T2c;
|
Chris@10
|
696 T2d = T2b - T2c;
|
Chris@10
|
697 T2I = T1k - T1p;
|
Chris@10
|
698 }
|
Chris@10
|
699 {
|
Chris@10
|
700 E T1w, T2K, T1B, T2L;
|
Chris@10
|
701 {
|
Chris@10
|
702 E T1t, T1v, T1s, T1u;
|
Chris@10
|
703 T1t = cr[WS(rs, 12)];
|
Chris@10
|
704 T1v = ci[WS(rs, 12)];
|
Chris@10
|
705 T1s = W[22];
|
Chris@10
|
706 T1u = W[23];
|
Chris@10
|
707 T1w = FMA(T1s, T1t, T1u * T1v);
|
Chris@10
|
708 T2K = FNMS(T1u, T1t, T1s * T1v);
|
Chris@10
|
709 }
|
Chris@10
|
710 {
|
Chris@10
|
711 E T1y, T1A, T1x, T1z;
|
Chris@10
|
712 T1y = cr[WS(rs, 2)];
|
Chris@10
|
713 T1A = ci[WS(rs, 2)];
|
Chris@10
|
714 T1x = W[2];
|
Chris@10
|
715 T1z = W[3];
|
Chris@10
|
716 T1B = FMA(T1x, T1y, T1z * T1A);
|
Chris@10
|
717 T2L = FNMS(T1z, T1y, T1x * T1A);
|
Chris@10
|
718 }
|
Chris@10
|
719 T1C = T1w + T1B;
|
Chris@10
|
720 T3k = T2K + T2L;
|
Chris@10
|
721 T2f = T1w - T1B;
|
Chris@10
|
722 T2M = T2K - T2L;
|
Chris@10
|
723 }
|
Chris@10
|
724 T1r = T1f - T1q;
|
Chris@10
|
725 T1O = T1C - T1N;
|
Chris@10
|
726 T1P = T1r + T1O;
|
Chris@10
|
727 T3i = T3g - T3h;
|
Chris@10
|
728 T3l = T3j - T3k;
|
Chris@10
|
729 T3J = T3l - T3i;
|
Chris@10
|
730 T3D = T3g + T3h;
|
Chris@10
|
731 T3E = T3k + T3j;
|
Chris@10
|
732 T44 = T3D + T3E;
|
Chris@10
|
733 T1V = T1f + T1q;
|
Chris@10
|
734 T1W = T1C + T1N;
|
Chris@10
|
735 T1X = T1V + T1W;
|
Chris@10
|
736 T2e = T2a - T2d;
|
Chris@10
|
737 T2j = T2f - T2i;
|
Chris@10
|
738 T2k = T2e + T2j;
|
Chris@10
|
739 T2W = T2H - T2I;
|
Chris@10
|
740 T2X = T2M - T2N;
|
Chris@10
|
741 T4f = T2W + T2X;
|
Chris@10
|
742 T33 = T2a + T2d;
|
Chris@10
|
743 T34 = T2f + T2i;
|
Chris@10
|
744 T35 = T33 + T34;
|
Chris@10
|
745 T2J = T2H + T2I;
|
Chris@10
|
746 T2O = T2M + T2N;
|
Chris@10
|
747 T4q = T2J + T2O;
|
Chris@10
|
748 }
|
Chris@10
|
749 {
|
Chris@10
|
750 E Tu, T3n, T1Z, T2w, T12, T3r, T27, T2z, TF, T3o, T22, T2x, TR, T3q, T24;
|
Chris@10
|
751 E T2C;
|
Chris@10
|
752 {
|
Chris@10
|
753 E To, T2u, Tt, T2v;
|
Chris@10
|
754 {
|
Chris@10
|
755 E Tl, Tn, Tk, Tm;
|
Chris@10
|
756 Tl = cr[WS(rs, 4)];
|
Chris@10
|
757 Tn = ci[WS(rs, 4)];
|
Chris@10
|
758 Tk = W[6];
|
Chris@10
|
759 Tm = W[7];
|
Chris@10
|
760 To = FMA(Tk, Tl, Tm * Tn);
|
Chris@10
|
761 T2u = FNMS(Tm, Tl, Tk * Tn);
|
Chris@10
|
762 }
|
Chris@10
|
763 {
|
Chris@10
|
764 E Tq, Ts, Tp, Tr;
|
Chris@10
|
765 Tq = cr[WS(rs, 14)];
|
Chris@10
|
766 Ts = ci[WS(rs, 14)];
|
Chris@10
|
767 Tp = W[26];
|
Chris@10
|
768 Tr = W[27];
|
Chris@10
|
769 Tt = FMA(Tp, Tq, Tr * Ts);
|
Chris@10
|
770 T2v = FNMS(Tr, Tq, Tp * Ts);
|
Chris@10
|
771 }
|
Chris@10
|
772 Tu = To + Tt;
|
Chris@10
|
773 T3n = T2u + T2v;
|
Chris@10
|
774 T1Z = To - Tt;
|
Chris@10
|
775 T2w = T2u - T2v;
|
Chris@10
|
776 }
|
Chris@10
|
777 {
|
Chris@10
|
778 E TW, T25, T11, T26;
|
Chris@10
|
779 {
|
Chris@10
|
780 E TT, TV, TS, TU;
|
Chris@10
|
781 TT = cr[WS(rs, 1)];
|
Chris@10
|
782 TV = ci[WS(rs, 1)];
|
Chris@10
|
783 TS = W[0];
|
Chris@10
|
784 TU = W[1];
|
Chris@10
|
785 TW = FMA(TS, TT, TU * TV);
|
Chris@10
|
786 T25 = FNMS(TU, TT, TS * TV);
|
Chris@10
|
787 }
|
Chris@10
|
788 {
|
Chris@10
|
789 E TY, T10, TX, TZ;
|
Chris@10
|
790 TY = cr[WS(rs, 11)];
|
Chris@10
|
791 T10 = ci[WS(rs, 11)];
|
Chris@10
|
792 TX = W[20];
|
Chris@10
|
793 TZ = W[21];
|
Chris@10
|
794 T11 = FMA(TX, TY, TZ * T10);
|
Chris@10
|
795 T26 = FNMS(TZ, TY, TX * T10);
|
Chris@10
|
796 }
|
Chris@10
|
797 T12 = TW + T11;
|
Chris@10
|
798 T3r = T25 + T26;
|
Chris@10
|
799 T27 = T25 - T26;
|
Chris@10
|
800 T2z = T11 - TW;
|
Chris@10
|
801 }
|
Chris@10
|
802 {
|
Chris@10
|
803 E Tz, T20, TE, T21;
|
Chris@10
|
804 {
|
Chris@10
|
805 E Tw, Ty, Tv, Tx;
|
Chris@10
|
806 Tw = cr[WS(rs, 9)];
|
Chris@10
|
807 Ty = ci[WS(rs, 9)];
|
Chris@10
|
808 Tv = W[16];
|
Chris@10
|
809 Tx = W[17];
|
Chris@10
|
810 Tz = FMA(Tv, Tw, Tx * Ty);
|
Chris@10
|
811 T20 = FNMS(Tx, Tw, Tv * Ty);
|
Chris@10
|
812 }
|
Chris@10
|
813 {
|
Chris@10
|
814 E TB, TD, TA, TC;
|
Chris@10
|
815 TB = cr[WS(rs, 19)];
|
Chris@10
|
816 TD = ci[WS(rs, 19)];
|
Chris@10
|
817 TA = W[36];
|
Chris@10
|
818 TC = W[37];
|
Chris@10
|
819 TE = FMA(TA, TB, TC * TD);
|
Chris@10
|
820 T21 = FNMS(TC, TB, TA * TD);
|
Chris@10
|
821 }
|
Chris@10
|
822 TF = Tz + TE;
|
Chris@10
|
823 T3o = T20 + T21;
|
Chris@10
|
824 T22 = T20 - T21;
|
Chris@10
|
825 T2x = Tz - TE;
|
Chris@10
|
826 }
|
Chris@10
|
827 {
|
Chris@10
|
828 E TL, T2A, TQ, T2B;
|
Chris@10
|
829 {
|
Chris@10
|
830 E TI, TK, TH, TJ;
|
Chris@10
|
831 TI = cr[WS(rs, 16)];
|
Chris@10
|
832 TK = ci[WS(rs, 16)];
|
Chris@10
|
833 TH = W[30];
|
Chris@10
|
834 TJ = W[31];
|
Chris@10
|
835 TL = FMA(TH, TI, TJ * TK);
|
Chris@10
|
836 T2A = FNMS(TJ, TI, TH * TK);
|
Chris@10
|
837 }
|
Chris@10
|
838 {
|
Chris@10
|
839 E TN, TP, TM, TO;
|
Chris@10
|
840 TN = cr[WS(rs, 6)];
|
Chris@10
|
841 TP = ci[WS(rs, 6)];
|
Chris@10
|
842 TM = W[10];
|
Chris@10
|
843 TO = W[11];
|
Chris@10
|
844 TQ = FMA(TM, TN, TO * TP);
|
Chris@10
|
845 T2B = FNMS(TO, TN, TM * TP);
|
Chris@10
|
846 }
|
Chris@10
|
847 TR = TL + TQ;
|
Chris@10
|
848 T3q = T2A + T2B;
|
Chris@10
|
849 T24 = TL - TQ;
|
Chris@10
|
850 T2C = T2A - T2B;
|
Chris@10
|
851 }
|
Chris@10
|
852 TG = Tu - TF;
|
Chris@10
|
853 T13 = TR - T12;
|
Chris@10
|
854 T14 = TG + T13;
|
Chris@10
|
855 T3p = T3n - T3o;
|
Chris@10
|
856 T3s = T3q - T3r;
|
Chris@10
|
857 T3K = T3p + T3s;
|
Chris@10
|
858 T3A = T3n + T3o;
|
Chris@10
|
859 T3B = T3q + T3r;
|
Chris@10
|
860 T43 = T3A + T3B;
|
Chris@10
|
861 T1S = Tu + TF;
|
Chris@10
|
862 T1T = TR + T12;
|
Chris@10
|
863 T1U = T1S + T1T;
|
Chris@10
|
864 T23 = T1Z - T22;
|
Chris@10
|
865 T28 = T24 - T27;
|
Chris@10
|
866 T29 = T23 + T28;
|
Chris@10
|
867 T2T = T2w - T2x;
|
Chris@10
|
868 T2U = T2C + T2z;
|
Chris@10
|
869 T4e = T2T + T2U;
|
Chris@10
|
870 T30 = T1Z + T22;
|
Chris@10
|
871 T31 = T24 + T27;
|
Chris@10
|
872 T32 = T30 + T31;
|
Chris@10
|
873 T2y = T2w + T2x;
|
Chris@10
|
874 T2D = T2z - T2C;
|
Chris@10
|
875 T4p = T2D - T2y;
|
Chris@10
|
876 }
|
Chris@10
|
877 {
|
Chris@10
|
878 E T3e, T1Q, T3d, T3u, T3w, T3m, T3t, T3v, T3f;
|
Chris@10
|
879 T3e = KP559016994 * (T14 - T1P);
|
Chris@10
|
880 T1Q = T14 + T1P;
|
Chris@10
|
881 T3d = FNMS(KP250000000, T1Q, Tj);
|
Chris@10
|
882 T3m = T3i + T3l;
|
Chris@10
|
883 T3t = T3p - T3s;
|
Chris@10
|
884 T3u = FNMS(KP587785252, T3t, KP951056516 * T3m);
|
Chris@10
|
885 T3w = FMA(KP951056516, T3t, KP587785252 * T3m);
|
Chris@10
|
886 ci[WS(rs, 9)] = Tj + T1Q;
|
Chris@10
|
887 T3v = T3e + T3d;
|
Chris@10
|
888 ci[WS(rs, 5)] = T3v - T3w;
|
Chris@10
|
889 cr[WS(rs, 6)] = T3v + T3w;
|
Chris@10
|
890 T3f = T3d - T3e;
|
Chris@10
|
891 cr[WS(rs, 2)] = T3f - T3u;
|
Chris@10
|
892 ci[WS(rs, 1)] = T3f + T3u;
|
Chris@10
|
893 }
|
Chris@10
|
894 {
|
Chris@10
|
895 E T36, T38, T39, T2Z, T3c, T2V, T2Y, T3b, T3a;
|
Chris@10
|
896 T36 = KP559016994 * (T32 - T35);
|
Chris@10
|
897 T38 = T32 + T35;
|
Chris@10
|
898 T39 = FNMS(KP250000000, T38, T37);
|
Chris@10
|
899 T2V = T2T - T2U;
|
Chris@10
|
900 T2Y = T2W - T2X;
|
Chris@10
|
901 T2Z = FMA(KP951056516, T2V, KP587785252 * T2Y);
|
Chris@10
|
902 T3c = FNMS(KP587785252, T2V, KP951056516 * T2Y);
|
Chris@10
|
903 cr[WS(rs, 5)] = T37 + T38;
|
Chris@10
|
904 T3b = T39 - T36;
|
Chris@10
|
905 ci[WS(rs, 2)] = T3b - T3c;
|
Chris@10
|
906 ci[WS(rs, 6)] = T3c + T3b;
|
Chris@10
|
907 T3a = T36 + T39;
|
Chris@10
|
908 cr[WS(rs, 1)] = T2Z + T3a;
|
Chris@10
|
909 cr[WS(rs, 9)] = T3a - T2Z;
|
Chris@10
|
910 }
|
Chris@10
|
911 {
|
Chris@10
|
912 E T3x, T1Y, T3y, T3G, T3I, T3C, T3F, T3H, T3z;
|
Chris@10
|
913 T3x = KP559016994 * (T1U - T1X);
|
Chris@10
|
914 T1Y = T1U + T1X;
|
Chris@10
|
915 T3y = FNMS(KP250000000, T1Y, T1R);
|
Chris@10
|
916 T3C = T3A - T3B;
|
Chris@10
|
917 T3F = T3D - T3E;
|
Chris@10
|
918 T3G = FMA(KP951056516, T3C, KP587785252 * T3F);
|
Chris@10
|
919 T3I = FNMS(KP587785252, T3C, KP951056516 * T3F);
|
Chris@10
|
920 cr[0] = T1R + T1Y;
|
Chris@10
|
921 T3H = T3y - T3x;
|
Chris@10
|
922 ci[WS(rs, 7)] = T3H - T3I;
|
Chris@10
|
923 cr[WS(rs, 8)] = T3H + T3I;
|
Chris@10
|
924 T3z = T3x + T3y;
|
Chris@10
|
925 cr[WS(rs, 4)] = T3z - T3G;
|
Chris@10
|
926 ci[WS(rs, 3)] = T3z + T3G;
|
Chris@10
|
927 }
|
Chris@10
|
928 {
|
Chris@10
|
929 E T2l, T2r, T2s, T2Q, T2R, T2E, T2P, T2S, T2t;
|
Chris@10
|
930 T2l = KP559016994 * (T29 - T2k);
|
Chris@10
|
931 T2r = T29 + T2k;
|
Chris@10
|
932 T2s = FNMS(KP250000000, T2r, T2q);
|
Chris@10
|
933 T2E = T2y + T2D;
|
Chris@10
|
934 T2P = T2J - T2O;
|
Chris@10
|
935 T2Q = FMA(KP951056516, T2E, KP587785252 * T2P);
|
Chris@10
|
936 T2R = FNMS(KP587785252, T2E, KP951056516 * T2P);
|
Chris@10
|
937 ci[WS(rs, 4)] = T2q + T2r;
|
Chris@10
|
938 T2S = T2s - T2l;
|
Chris@10
|
939 cr[WS(rs, 3)] = T2R + T2S;
|
Chris@10
|
940 cr[WS(rs, 7)] = T2S - T2R;
|
Chris@10
|
941 T2t = T2l + T2s;
|
Chris@10
|
942 ci[0] = T2t - T2Q;
|
Chris@10
|
943 ci[WS(rs, 8)] = T2Q + T2t;
|
Chris@10
|
944 }
|
Chris@10
|
945 {
|
Chris@10
|
946 E T3U, T3L, T3V, T3T, T3X, T3R, T3S, T3Y, T3W;
|
Chris@10
|
947 T3U = KP559016994 * (T3K + T3J);
|
Chris@10
|
948 T3L = T3J - T3K;
|
Chris@10
|
949 T3V = FMA(KP250000000, T3L, T3Q);
|
Chris@10
|
950 T3R = T13 - TG;
|
Chris@10
|
951 T3S = T1r - T1O;
|
Chris@10
|
952 T3T = FNMS(KP587785252, T3S, KP951056516 * T3R);
|
Chris@10
|
953 T3X = FMA(KP587785252, T3R, KP951056516 * T3S);
|
Chris@10
|
954 cr[WS(rs, 10)] = T3L - T3Q;
|
Chris@10
|
955 T3Y = T3V - T3U;
|
Chris@10
|
956 cr[WS(rs, 18)] = T3X - T3Y;
|
Chris@10
|
957 ci[WS(rs, 17)] = T3X + T3Y;
|
Chris@10
|
958 T3W = T3U + T3V;
|
Chris@10
|
959 cr[WS(rs, 14)] = T3T - T3W;
|
Chris@10
|
960 ci[WS(rs, 13)] = T3T + T3W;
|
Chris@10
|
961 }
|
Chris@10
|
962 {
|
Chris@10
|
963 E T4g, T4k, T4l, T4d, T4n, T4b, T4c, T4o, T4m;
|
Chris@10
|
964 T4g = KP559016994 * (T4e - T4f);
|
Chris@10
|
965 T4k = T4e + T4f;
|
Chris@10
|
966 T4l = FNMS(KP250000000, T4k, T4j);
|
Chris@10
|
967 T4b = T33 - T34;
|
Chris@10
|
968 T4c = T30 - T31;
|
Chris@10
|
969 T4d = FNMS(KP587785252, T4c, KP951056516 * T4b);
|
Chris@10
|
970 T4n = FMA(KP951056516, T4c, KP587785252 * T4b);
|
Chris@10
|
971 ci[WS(rs, 14)] = T4k + T4j;
|
Chris@10
|
972 T4o = T4g + T4l;
|
Chris@10
|
973 ci[WS(rs, 10)] = T4n + T4o;
|
Chris@10
|
974 ci[WS(rs, 18)] = T4o - T4n;
|
Chris@10
|
975 T4m = T4g - T4l;
|
Chris@10
|
976 cr[WS(rs, 13)] = T4d + T4m;
|
Chris@10
|
977 cr[WS(rs, 17)] = T4m - T4d;
|
Chris@10
|
978 }
|
Chris@10
|
979 {
|
Chris@10
|
980 E T47, T45, T46, T41, T49, T3Z, T40, T4a, T48;
|
Chris@10
|
981 T47 = KP559016994 * (T43 - T44);
|
Chris@10
|
982 T45 = T43 + T44;
|
Chris@10
|
983 T46 = FNMS(KP250000000, T45, T42);
|
Chris@10
|
984 T3Z = T1S - T1T;
|
Chris@10
|
985 T40 = T1V - T1W;
|
Chris@10
|
986 T41 = FNMS(KP951056516, T40, KP587785252 * T3Z);
|
Chris@10
|
987 T49 = FMA(KP951056516, T3Z, KP587785252 * T40);
|
Chris@10
|
988 ci[WS(rs, 19)] = T45 + T42;
|
Chris@10
|
989 T4a = T47 + T46;
|
Chris@10
|
990 cr[WS(rs, 16)] = T49 - T4a;
|
Chris@10
|
991 ci[WS(rs, 15)] = T49 + T4a;
|
Chris@10
|
992 T48 = T46 - T47;
|
Chris@10
|
993 cr[WS(rs, 12)] = T41 - T48;
|
Chris@10
|
994 ci[WS(rs, 11)] = T41 + T48;
|
Chris@10
|
995 }
|
Chris@10
|
996 {
|
Chris@10
|
997 E T4w, T4r, T4x, T4v, T4z, T4t, T4u, T4A, T4y;
|
Chris@10
|
998 T4w = KP559016994 * (T4p + T4q);
|
Chris@10
|
999 T4r = T4p - T4q;
|
Chris@10
|
1000 T4x = FMA(KP250000000, T4r, T4s);
|
Chris@10
|
1001 T4t = T23 - T28;
|
Chris@10
|
1002 T4u = T2e - T2j;
|
Chris@10
|
1003 T4v = FMA(KP951056516, T4t, KP587785252 * T4u);
|
Chris@10
|
1004 T4z = FNMS(KP587785252, T4t, KP951056516 * T4u);
|
Chris@10
|
1005 cr[WS(rs, 15)] = T4r - T4s;
|
Chris@10
|
1006 T4A = T4w + T4x;
|
Chris@10
|
1007 ci[WS(rs, 12)] = T4z + T4A;
|
Chris@10
|
1008 ci[WS(rs, 16)] = T4A - T4z;
|
Chris@10
|
1009 T4y = T4w - T4x;
|
Chris@10
|
1010 cr[WS(rs, 11)] = T4v + T4y;
|
Chris@10
|
1011 cr[WS(rs, 19)] = T4y - T4v;
|
Chris@10
|
1012 }
|
Chris@10
|
1013 }
|
Chris@10
|
1014 }
|
Chris@10
|
1015 }
|
Chris@10
|
1016
|
Chris@10
|
1017 static const tw_instr twinstr[] = {
|
Chris@10
|
1018 {TW_FULL, 1, 20},
|
Chris@10
|
1019 {TW_NEXT, 1, 0}
|
Chris@10
|
1020 };
|
Chris@10
|
1021
|
Chris@10
|
1022 static const hc2hc_desc desc = { 20, "hf_20", twinstr, &GENUS, {184, 62, 62, 0} };
|
Chris@10
|
1023
|
Chris@10
|
1024 void X(codelet_hf_20) (planner *p) {
|
Chris@10
|
1025 X(khc2hc_register) (p, hf_20, &desc);
|
Chris@10
|
1026 }
|
Chris@10
|
1027 #endif /* HAVE_FMA */
|