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 08:06:33 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 -n 20 -dit -name hf_20 -include rdft/scalar/hf.h */
|
cannam@167
|
29
|
cannam@167
|
30 /*
|
cannam@167
|
31 * This function contains 246 FP additions, 148 FP multiplications,
|
cannam@167
|
32 * (or, 136 additions, 38 multiplications, 110 fused multiply/add),
|
cannam@167
|
33 * 61 stack variables, 4 constants, and 80 memory accesses
|
cannam@167
|
34 */
|
cannam@167
|
35 #include "rdft/scalar/hf.h"
|
cannam@167
|
36
|
cannam@167
|
37 static void hf_20(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 {
|
cannam@167
|
44 INT m;
|
cannam@167
|
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)) {
|
cannam@167
|
46 E T8, T4N, T2i, T4q, Tl, T4O, T2n, T4r, TN, T2b, T43, T4b, T2v, T3v, T3a;
|
cannam@167
|
47 E T3F, T27, T2f, T3T, T4f, T2R, T3z, T3i, T3J, T1G, T2e, T3W, T4e, T2K, T3y;
|
cannam@167
|
48 E T3p, T3I, T1e, T2c, T40, T4c, T2C, T3w, T33, T3G;
|
cannam@167
|
49 {
|
cannam@167
|
50 E T1, T4p, T3, T6, T4, T4n, T2, T7, T4o, T5;
|
cannam@167
|
51 T1 = cr[0];
|
cannam@167
|
52 T4p = ci[0];
|
cannam@167
|
53 T3 = cr[WS(rs, 10)];
|
cannam@167
|
54 T6 = ci[WS(rs, 10)];
|
cannam@167
|
55 T2 = W[18];
|
cannam@167
|
56 T4 = T2 * T3;
|
cannam@167
|
57 T4n = T2 * T6;
|
cannam@167
|
58 T5 = W[19];
|
cannam@167
|
59 T7 = FMA(T5, T6, T4);
|
cannam@167
|
60 T4o = FNMS(T5, T3, T4n);
|
cannam@167
|
61 T8 = T1 + T7;
|
cannam@167
|
62 T4N = T4p - T4o;
|
cannam@167
|
63 T2i = T1 - T7;
|
cannam@167
|
64 T4q = T4o + T4p;
|
cannam@167
|
65 }
|
cannam@167
|
66 {
|
cannam@167
|
67 E Ta, Td, Tb, T2j, Tg, Tj, Th, T2l, T9, Tf;
|
cannam@167
|
68 Ta = cr[WS(rs, 5)];
|
cannam@167
|
69 Td = ci[WS(rs, 5)];
|
cannam@167
|
70 T9 = W[8];
|
cannam@167
|
71 Tb = T9 * Ta;
|
cannam@167
|
72 T2j = T9 * Td;
|
cannam@167
|
73 Tg = cr[WS(rs, 15)];
|
cannam@167
|
74 Tj = ci[WS(rs, 15)];
|
cannam@167
|
75 Tf = W[28];
|
cannam@167
|
76 Th = Tf * Tg;
|
cannam@167
|
77 T2l = Tf * Tj;
|
cannam@167
|
78 {
|
cannam@167
|
79 E Te, T2k, Tk, T2m, Tc, Ti;
|
cannam@167
|
80 Tc = W[9];
|
cannam@167
|
81 Te = FMA(Tc, Td, Tb);
|
cannam@167
|
82 T2k = FNMS(Tc, Ta, T2j);
|
cannam@167
|
83 Ti = W[29];
|
cannam@167
|
84 Tk = FMA(Ti, Tj, Th);
|
cannam@167
|
85 T2m = FNMS(Ti, Tg, T2l);
|
cannam@167
|
86 Tl = Te + Tk;
|
cannam@167
|
87 T4O = Te - Tk;
|
cannam@167
|
88 T2n = T2k - T2m;
|
cannam@167
|
89 T4r = T2k + T2m;
|
cannam@167
|
90 }
|
cannam@167
|
91 }
|
cannam@167
|
92 {
|
cannam@167
|
93 E Ts, T36, TL, T2t, Ty, T38, TF, T2r;
|
cannam@167
|
94 {
|
cannam@167
|
95 E To, Tr, Tp, T35, Tn, Tq;
|
cannam@167
|
96 To = cr[WS(rs, 4)];
|
cannam@167
|
97 Tr = ci[WS(rs, 4)];
|
cannam@167
|
98 Tn = W[6];
|
cannam@167
|
99 Tp = Tn * To;
|
cannam@167
|
100 T35 = Tn * Tr;
|
cannam@167
|
101 Tq = W[7];
|
cannam@167
|
102 Ts = FMA(Tq, Tr, Tp);
|
cannam@167
|
103 T36 = FNMS(Tq, To, T35);
|
cannam@167
|
104 }
|
cannam@167
|
105 {
|
cannam@167
|
106 E TH, TK, TI, T2s, TG, TJ;
|
cannam@167
|
107 TH = cr[WS(rs, 19)];
|
cannam@167
|
108 TK = ci[WS(rs, 19)];
|
cannam@167
|
109 TG = W[36];
|
cannam@167
|
110 TI = TG * TH;
|
cannam@167
|
111 T2s = TG * TK;
|
cannam@167
|
112 TJ = W[37];
|
cannam@167
|
113 TL = FMA(TJ, TK, TI);
|
cannam@167
|
114 T2t = FNMS(TJ, TH, T2s);
|
cannam@167
|
115 }
|
cannam@167
|
116 {
|
cannam@167
|
117 E Tu, Tx, Tv, T37, Tt, Tw;
|
cannam@167
|
118 Tu = cr[WS(rs, 14)];
|
cannam@167
|
119 Tx = ci[WS(rs, 14)];
|
cannam@167
|
120 Tt = W[26];
|
cannam@167
|
121 Tv = Tt * Tu;
|
cannam@167
|
122 T37 = Tt * Tx;
|
cannam@167
|
123 Tw = W[27];
|
cannam@167
|
124 Ty = FMA(Tw, Tx, Tv);
|
cannam@167
|
125 T38 = FNMS(Tw, Tu, T37);
|
cannam@167
|
126 }
|
cannam@167
|
127 {
|
cannam@167
|
128 E TB, TE, TC, T2q, TA, TD;
|
cannam@167
|
129 TB = cr[WS(rs, 9)];
|
cannam@167
|
130 TE = ci[WS(rs, 9)];
|
cannam@167
|
131 TA = W[16];
|
cannam@167
|
132 TC = TA * TB;
|
cannam@167
|
133 T2q = TA * TE;
|
cannam@167
|
134 TD = W[17];
|
cannam@167
|
135 TF = FMA(TD, TE, TC);
|
cannam@167
|
136 T2r = FNMS(TD, TB, T2q);
|
cannam@167
|
137 }
|
cannam@167
|
138 {
|
cannam@167
|
139 E Tz, TM, T41, T42;
|
cannam@167
|
140 Tz = Ts + Ty;
|
cannam@167
|
141 TM = TF + TL;
|
cannam@167
|
142 TN = Tz - TM;
|
cannam@167
|
143 T2b = Tz + TM;
|
cannam@167
|
144 T41 = T2r + T2t;
|
cannam@167
|
145 T42 = T36 + T38;
|
cannam@167
|
146 T43 = T41 - T42;
|
cannam@167
|
147 T4b = T42 + T41;
|
cannam@167
|
148 }
|
cannam@167
|
149 {
|
cannam@167
|
150 E T2p, T2u, T34, T39;
|
cannam@167
|
151 T2p = Ts - Ty;
|
cannam@167
|
152 T2u = T2r - T2t;
|
cannam@167
|
153 T2v = T2p - T2u;
|
cannam@167
|
154 T3v = T2p + T2u;
|
cannam@167
|
155 T34 = TL - TF;
|
cannam@167
|
156 T39 = T36 - T38;
|
cannam@167
|
157 T3a = T34 - T39;
|
cannam@167
|
158 T3F = T39 + T34;
|
cannam@167
|
159 }
|
cannam@167
|
160 }
|
cannam@167
|
161 {
|
cannam@167
|
162 E T1M, T3e, T25, T2P, T1S, T3g, T1Z, T2N;
|
cannam@167
|
163 {
|
cannam@167
|
164 E T1I, T1L, T1J, T3d, T1H, T1K;
|
cannam@167
|
165 T1I = cr[WS(rs, 12)];
|
cannam@167
|
166 T1L = ci[WS(rs, 12)];
|
cannam@167
|
167 T1H = W[22];
|
cannam@167
|
168 T1J = T1H * T1I;
|
cannam@167
|
169 T3d = T1H * T1L;
|
cannam@167
|
170 T1K = W[23];
|
cannam@167
|
171 T1M = FMA(T1K, T1L, T1J);
|
cannam@167
|
172 T3e = FNMS(T1K, T1I, T3d);
|
cannam@167
|
173 }
|
cannam@167
|
174 {
|
cannam@167
|
175 E T21, T24, T22, T2O, T20, T23;
|
cannam@167
|
176 T21 = cr[WS(rs, 7)];
|
cannam@167
|
177 T24 = ci[WS(rs, 7)];
|
cannam@167
|
178 T20 = W[12];
|
cannam@167
|
179 T22 = T20 * T21;
|
cannam@167
|
180 T2O = T20 * T24;
|
cannam@167
|
181 T23 = W[13];
|
cannam@167
|
182 T25 = FMA(T23, T24, T22);
|
cannam@167
|
183 T2P = FNMS(T23, T21, T2O);
|
cannam@167
|
184 }
|
cannam@167
|
185 {
|
cannam@167
|
186 E T1O, T1R, T1P, T3f, T1N, T1Q;
|
cannam@167
|
187 T1O = cr[WS(rs, 2)];
|
cannam@167
|
188 T1R = ci[WS(rs, 2)];
|
cannam@167
|
189 T1N = W[2];
|
cannam@167
|
190 T1P = T1N * T1O;
|
cannam@167
|
191 T3f = T1N * T1R;
|
cannam@167
|
192 T1Q = W[3];
|
cannam@167
|
193 T1S = FMA(T1Q, T1R, T1P);
|
cannam@167
|
194 T3g = FNMS(T1Q, T1O, T3f);
|
cannam@167
|
195 }
|
cannam@167
|
196 {
|
cannam@167
|
197 E T1V, T1Y, T1W, T2M, T1U, T1X;
|
cannam@167
|
198 T1V = cr[WS(rs, 17)];
|
cannam@167
|
199 T1Y = ci[WS(rs, 17)];
|
cannam@167
|
200 T1U = W[32];
|
cannam@167
|
201 T1W = T1U * T1V;
|
cannam@167
|
202 T2M = T1U * T1Y;
|
cannam@167
|
203 T1X = W[33];
|
cannam@167
|
204 T1Z = FMA(T1X, T1Y, T1W);
|
cannam@167
|
205 T2N = FNMS(T1X, T1V, T2M);
|
cannam@167
|
206 }
|
cannam@167
|
207 {
|
cannam@167
|
208 E T1T, T26, T3R, T3S;
|
cannam@167
|
209 T1T = T1M + T1S;
|
cannam@167
|
210 T26 = T1Z + T25;
|
cannam@167
|
211 T27 = T1T - T26;
|
cannam@167
|
212 T2f = T1T + T26;
|
cannam@167
|
213 T3R = T2N + T2P;
|
cannam@167
|
214 T3S = T3e + T3g;
|
cannam@167
|
215 T3T = T3R - T3S;
|
cannam@167
|
216 T4f = T3S + T3R;
|
cannam@167
|
217 }
|
cannam@167
|
218 {
|
cannam@167
|
219 E T2L, T2Q, T3c, T3h;
|
cannam@167
|
220 T2L = T1M - T1S;
|
cannam@167
|
221 T2Q = T2N - T2P;
|
cannam@167
|
222 T2R = T2L - T2Q;
|
cannam@167
|
223 T3z = T2L + T2Q;
|
cannam@167
|
224 T3c = T25 - T1Z;
|
cannam@167
|
225 T3h = T3e - T3g;
|
cannam@167
|
226 T3i = T3c - T3h;
|
cannam@167
|
227 T3J = T3h + T3c;
|
cannam@167
|
228 }
|
cannam@167
|
229 }
|
cannam@167
|
230 {
|
cannam@167
|
231 E T1l, T3l, T1E, T2I, T1r, T3n, T1y, T2G;
|
cannam@167
|
232 {
|
cannam@167
|
233 E T1h, T1k, T1i, T3k, T1g, T1j;
|
cannam@167
|
234 T1h = cr[WS(rs, 8)];
|
cannam@167
|
235 T1k = ci[WS(rs, 8)];
|
cannam@167
|
236 T1g = W[14];
|
cannam@167
|
237 T1i = T1g * T1h;
|
cannam@167
|
238 T3k = T1g * T1k;
|
cannam@167
|
239 T1j = W[15];
|
cannam@167
|
240 T1l = FMA(T1j, T1k, T1i);
|
cannam@167
|
241 T3l = FNMS(T1j, T1h, T3k);
|
cannam@167
|
242 }
|
cannam@167
|
243 {
|
cannam@167
|
244 E T1A, T1D, T1B, T2H, T1z, T1C;
|
cannam@167
|
245 T1A = cr[WS(rs, 3)];
|
cannam@167
|
246 T1D = ci[WS(rs, 3)];
|
cannam@167
|
247 T1z = W[4];
|
cannam@167
|
248 T1B = T1z * T1A;
|
cannam@167
|
249 T2H = T1z * T1D;
|
cannam@167
|
250 T1C = W[5];
|
cannam@167
|
251 T1E = FMA(T1C, T1D, T1B);
|
cannam@167
|
252 T2I = FNMS(T1C, T1A, T2H);
|
cannam@167
|
253 }
|
cannam@167
|
254 {
|
cannam@167
|
255 E T1n, T1q, T1o, T3m, T1m, T1p;
|
cannam@167
|
256 T1n = cr[WS(rs, 18)];
|
cannam@167
|
257 T1q = ci[WS(rs, 18)];
|
cannam@167
|
258 T1m = W[34];
|
cannam@167
|
259 T1o = T1m * T1n;
|
cannam@167
|
260 T3m = T1m * T1q;
|
cannam@167
|
261 T1p = W[35];
|
cannam@167
|
262 T1r = FMA(T1p, T1q, T1o);
|
cannam@167
|
263 T3n = FNMS(T1p, T1n, T3m);
|
cannam@167
|
264 }
|
cannam@167
|
265 {
|
cannam@167
|
266 E T1u, T1x, T1v, T2F, T1t, T1w;
|
cannam@167
|
267 T1u = cr[WS(rs, 13)];
|
cannam@167
|
268 T1x = ci[WS(rs, 13)];
|
cannam@167
|
269 T1t = W[24];
|
cannam@167
|
270 T1v = T1t * T1u;
|
cannam@167
|
271 T2F = T1t * T1x;
|
cannam@167
|
272 T1w = W[25];
|
cannam@167
|
273 T1y = FMA(T1w, T1x, T1v);
|
cannam@167
|
274 T2G = FNMS(T1w, T1u, T2F);
|
cannam@167
|
275 }
|
cannam@167
|
276 {
|
cannam@167
|
277 E T1s, T1F, T3U, T3V;
|
cannam@167
|
278 T1s = T1l + T1r;
|
cannam@167
|
279 T1F = T1y + T1E;
|
cannam@167
|
280 T1G = T1s - T1F;
|
cannam@167
|
281 T2e = T1s + T1F;
|
cannam@167
|
282 T3U = T2G + T2I;
|
cannam@167
|
283 T3V = T3l + T3n;
|
cannam@167
|
284 T3W = T3U - T3V;
|
cannam@167
|
285 T4e = T3V + T3U;
|
cannam@167
|
286 }
|
cannam@167
|
287 {
|
cannam@167
|
288 E T2E, T2J, T3j, T3o;
|
cannam@167
|
289 T2E = T1l - T1r;
|
cannam@167
|
290 T2J = T2G - T2I;
|
cannam@167
|
291 T2K = T2E - T2J;
|
cannam@167
|
292 T3y = T2E + T2J;
|
cannam@167
|
293 T3j = T1E - T1y;
|
cannam@167
|
294 T3o = T3l - T3n;
|
cannam@167
|
295 T3p = T3j - T3o;
|
cannam@167
|
296 T3I = T3o + T3j;
|
cannam@167
|
297 }
|
cannam@167
|
298 }
|
cannam@167
|
299 {
|
cannam@167
|
300 E TT, T2Z, T1c, T2A, TZ, T31, T16, T2y;
|
cannam@167
|
301 {
|
cannam@167
|
302 E TP, TS, TQ, T2Y, TO, TR;
|
cannam@167
|
303 TP = cr[WS(rs, 16)];
|
cannam@167
|
304 TS = ci[WS(rs, 16)];
|
cannam@167
|
305 TO = W[30];
|
cannam@167
|
306 TQ = TO * TP;
|
cannam@167
|
307 T2Y = TO * TS;
|
cannam@167
|
308 TR = W[31];
|
cannam@167
|
309 TT = FMA(TR, TS, TQ);
|
cannam@167
|
310 T2Z = FNMS(TR, TP, T2Y);
|
cannam@167
|
311 }
|
cannam@167
|
312 {
|
cannam@167
|
313 E T18, T1b, T19, T2z, T17, T1a;
|
cannam@167
|
314 T18 = cr[WS(rs, 11)];
|
cannam@167
|
315 T1b = ci[WS(rs, 11)];
|
cannam@167
|
316 T17 = W[20];
|
cannam@167
|
317 T19 = T17 * T18;
|
cannam@167
|
318 T2z = T17 * T1b;
|
cannam@167
|
319 T1a = W[21];
|
cannam@167
|
320 T1c = FMA(T1a, T1b, T19);
|
cannam@167
|
321 T2A = FNMS(T1a, T18, T2z);
|
cannam@167
|
322 }
|
cannam@167
|
323 {
|
cannam@167
|
324 E TV, TY, TW, T30, TU, TX;
|
cannam@167
|
325 TV = cr[WS(rs, 6)];
|
cannam@167
|
326 TY = ci[WS(rs, 6)];
|
cannam@167
|
327 TU = W[10];
|
cannam@167
|
328 TW = TU * TV;
|
cannam@167
|
329 T30 = TU * TY;
|
cannam@167
|
330 TX = W[11];
|
cannam@167
|
331 TZ = FMA(TX, TY, TW);
|
cannam@167
|
332 T31 = FNMS(TX, TV, T30);
|
cannam@167
|
333 }
|
cannam@167
|
334 {
|
cannam@167
|
335 E T12, T15, T13, T2x, T11, T14;
|
cannam@167
|
336 T12 = cr[WS(rs, 1)];
|
cannam@167
|
337 T15 = ci[WS(rs, 1)];
|
cannam@167
|
338 T11 = W[0];
|
cannam@167
|
339 T13 = T11 * T12;
|
cannam@167
|
340 T2x = T11 * T15;
|
cannam@167
|
341 T14 = W[1];
|
cannam@167
|
342 T16 = FMA(T14, T15, T13);
|
cannam@167
|
343 T2y = FNMS(T14, T12, T2x);
|
cannam@167
|
344 }
|
cannam@167
|
345 {
|
cannam@167
|
346 E T10, T1d, T3Y, T3Z;
|
cannam@167
|
347 T10 = TT + TZ;
|
cannam@167
|
348 T1d = T16 + T1c;
|
cannam@167
|
349 T1e = T10 - T1d;
|
cannam@167
|
350 T2c = T10 + T1d;
|
cannam@167
|
351 T3Y = T2y + T2A;
|
cannam@167
|
352 T3Z = T2Z + T31;
|
cannam@167
|
353 T40 = T3Y - T3Z;
|
cannam@167
|
354 T4c = T3Z + T3Y;
|
cannam@167
|
355 }
|
cannam@167
|
356 {
|
cannam@167
|
357 E T2w, T2B, T2X, T32;
|
cannam@167
|
358 T2w = TT - TZ;
|
cannam@167
|
359 T2B = T2y - T2A;
|
cannam@167
|
360 T2C = T2w - T2B;
|
cannam@167
|
361 T3w = T2w + T2B;
|
cannam@167
|
362 T2X = T1c - T16;
|
cannam@167
|
363 T32 = T2Z - T31;
|
cannam@167
|
364 T33 = T2X - T32;
|
cannam@167
|
365 T3G = T32 + T2X;
|
cannam@167
|
366 }
|
cannam@167
|
367 }
|
cannam@167
|
368 {
|
cannam@167
|
369 E T45, T47, Tm, T29, T3O, T3P, T46, T3Q;
|
cannam@167
|
370 {
|
cannam@167
|
371 E T3X, T44, T1f, T28;
|
cannam@167
|
372 T3X = T3T - T3W;
|
cannam@167
|
373 T44 = T40 - T43;
|
cannam@167
|
374 T45 = FNMS(KP618033988, T44, T3X);
|
cannam@167
|
375 T47 = FMA(KP618033988, T3X, T44);
|
cannam@167
|
376 Tm = T8 - Tl;
|
cannam@167
|
377 T1f = TN + T1e;
|
cannam@167
|
378 T28 = T1G + T27;
|
cannam@167
|
379 T29 = T1f + T28;
|
cannam@167
|
380 T3O = FNMS(KP250000000, T29, Tm);
|
cannam@167
|
381 T3P = T1f - T28;
|
cannam@167
|
382 }
|
cannam@167
|
383 ci[WS(rs, 9)] = Tm + T29;
|
cannam@167
|
384 T46 = FMA(KP559016994, T3P, T3O);
|
cannam@167
|
385 ci[WS(rs, 5)] = FNMS(KP951056516, T47, T46);
|
cannam@167
|
386 cr[WS(rs, 6)] = FMA(KP951056516, T47, T46);
|
cannam@167
|
387 T3Q = FNMS(KP559016994, T3P, T3O);
|
cannam@167
|
388 cr[WS(rs, 2)] = FNMS(KP951056516, T45, T3Q);
|
cannam@167
|
389 ci[WS(rs, 1)] = FMA(KP951056516, T45, T3Q);
|
cannam@167
|
390 }
|
cannam@167
|
391 {
|
cannam@167
|
392 E T3L, T3N, T3u, T3B, T3C, T3D, T3M, T3E;
|
cannam@167
|
393 {
|
cannam@167
|
394 E T3H, T3K, T3x, T3A;
|
cannam@167
|
395 T3H = T3F - T3G;
|
cannam@167
|
396 T3K = T3I - T3J;
|
cannam@167
|
397 T3L = FMA(KP618033988, T3K, T3H);
|
cannam@167
|
398 T3N = FNMS(KP618033988, T3H, T3K);
|
cannam@167
|
399 T3u = T2i + T2n;
|
cannam@167
|
400 T3x = T3v + T3w;
|
cannam@167
|
401 T3A = T3y + T3z;
|
cannam@167
|
402 T3B = T3x + T3A;
|
cannam@167
|
403 T3C = FNMS(KP250000000, T3B, T3u);
|
cannam@167
|
404 T3D = T3x - T3A;
|
cannam@167
|
405 }
|
cannam@167
|
406 cr[WS(rs, 5)] = T3u + T3B;
|
cannam@167
|
407 T3M = FNMS(KP559016994, T3D, T3C);
|
cannam@167
|
408 ci[WS(rs, 2)] = FNMS(KP951056516, T3N, T3M);
|
cannam@167
|
409 ci[WS(rs, 6)] = FMA(KP951056516, T3N, T3M);
|
cannam@167
|
410 T3E = FMA(KP559016994, T3D, T3C);
|
cannam@167
|
411 cr[WS(rs, 1)] = FMA(KP951056516, T3L, T3E);
|
cannam@167
|
412 cr[WS(rs, 9)] = FNMS(KP951056516, T3L, T3E);
|
cannam@167
|
413 }
|
cannam@167
|
414 {
|
cannam@167
|
415 E T4h, T4j, T2a, T2h, T48, T49, T4i, T4a;
|
cannam@167
|
416 {
|
cannam@167
|
417 E T4d, T4g, T2d, T2g;
|
cannam@167
|
418 T4d = T4b - T4c;
|
cannam@167
|
419 T4g = T4e - T4f;
|
cannam@167
|
420 T4h = FMA(KP618033988, T4g, T4d);
|
cannam@167
|
421 T4j = FNMS(KP618033988, T4d, T4g);
|
cannam@167
|
422 T2a = T8 + Tl;
|
cannam@167
|
423 T2d = T2b + T2c;
|
cannam@167
|
424 T2g = T2e + T2f;
|
cannam@167
|
425 T2h = T2d + T2g;
|
cannam@167
|
426 T48 = FNMS(KP250000000, T2h, T2a);
|
cannam@167
|
427 T49 = T2d - T2g;
|
cannam@167
|
428 }
|
cannam@167
|
429 cr[0] = T2a + T2h;
|
cannam@167
|
430 T4i = FNMS(KP559016994, T49, T48);
|
cannam@167
|
431 ci[WS(rs, 7)] = FNMS(KP951056516, T4j, T4i);
|
cannam@167
|
432 cr[WS(rs, 8)] = FMA(KP951056516, T4j, T4i);
|
cannam@167
|
433 T4a = FMA(KP559016994, T49, T48);
|
cannam@167
|
434 cr[WS(rs, 4)] = FNMS(KP951056516, T4h, T4a);
|
cannam@167
|
435 ci[WS(rs, 3)] = FMA(KP951056516, T4h, T4a);
|
cannam@167
|
436 }
|
cannam@167
|
437 {
|
cannam@167
|
438 E T3r, T3t, T2o, T2T, T2U, T2V, T3s, T2W;
|
cannam@167
|
439 {
|
cannam@167
|
440 E T3b, T3q, T2D, T2S;
|
cannam@167
|
441 T3b = T33 - T3a;
|
cannam@167
|
442 T3q = T3i - T3p;
|
cannam@167
|
443 T3r = FMA(KP618033988, T3q, T3b);
|
cannam@167
|
444 T3t = FNMS(KP618033988, T3b, T3q);
|
cannam@167
|
445 T2o = T2i - T2n;
|
cannam@167
|
446 T2D = T2v + T2C;
|
cannam@167
|
447 T2S = T2K + T2R;
|
cannam@167
|
448 T2T = T2D + T2S;
|
cannam@167
|
449 T2U = FNMS(KP250000000, T2T, T2o);
|
cannam@167
|
450 T2V = T2D - T2S;
|
cannam@167
|
451 }
|
cannam@167
|
452 ci[WS(rs, 4)] = T2o + T2T;
|
cannam@167
|
453 T3s = FNMS(KP559016994, T2V, T2U);
|
cannam@167
|
454 cr[WS(rs, 3)] = FMA(KP951056516, T3t, T3s);
|
cannam@167
|
455 cr[WS(rs, 7)] = FNMS(KP951056516, T3t, T3s);
|
cannam@167
|
456 T2W = FMA(KP559016994, T2V, T2U);
|
cannam@167
|
457 ci[0] = FNMS(KP951056516, T3r, T2W);
|
cannam@167
|
458 ci[WS(rs, 8)] = FMA(KP951056516, T3r, T2W);
|
cannam@167
|
459 }
|
cannam@167
|
460 {
|
cannam@167
|
461 E T4y, T4A, T4s, T4m, T4t, T4u, T4z, T4v;
|
cannam@167
|
462 {
|
cannam@167
|
463 E T4w, T4x, T4k, T4l;
|
cannam@167
|
464 T4w = T1e - TN;
|
cannam@167
|
465 T4x = T1G - T27;
|
cannam@167
|
466 T4y = FNMS(KP618033988, T4x, T4w);
|
cannam@167
|
467 T4A = FMA(KP618033988, T4w, T4x);
|
cannam@167
|
468 T4s = T4q - T4r;
|
cannam@167
|
469 T4k = T43 + T40;
|
cannam@167
|
470 T4l = T3W + T3T;
|
cannam@167
|
471 T4m = T4k + T4l;
|
cannam@167
|
472 T4t = FMA(KP250000000, T4m, T4s);
|
cannam@167
|
473 T4u = T4l - T4k;
|
cannam@167
|
474 }
|
cannam@167
|
475 cr[WS(rs, 10)] = T4m - T4s;
|
cannam@167
|
476 T4z = FNMS(KP559016994, T4u, T4t);
|
cannam@167
|
477 cr[WS(rs, 18)] = FMS(KP951056516, T4A, T4z);
|
cannam@167
|
478 ci[WS(rs, 17)] = FMA(KP951056516, T4A, T4z);
|
cannam@167
|
479 T4v = FMA(KP559016994, T4u, T4t);
|
cannam@167
|
480 cr[WS(rs, 14)] = FMS(KP951056516, T4y, T4v);
|
cannam@167
|
481 ci[WS(rs, 13)] = FMA(KP951056516, T4y, T4v);
|
cannam@167
|
482 }
|
cannam@167
|
483 {
|
cannam@167
|
484 E T4Y, T50, T4P, T4S, T4T, T4U, T4Z, T4V;
|
cannam@167
|
485 {
|
cannam@167
|
486 E T4W, T4X, T4Q, T4R;
|
cannam@167
|
487 T4W = T3y - T3z;
|
cannam@167
|
488 T4X = T3v - T3w;
|
cannam@167
|
489 T4Y = FNMS(KP618033988, T4X, T4W);
|
cannam@167
|
490 T50 = FMA(KP618033988, T4W, T4X);
|
cannam@167
|
491 T4P = T4N - T4O;
|
cannam@167
|
492 T4Q = T3F + T3G;
|
cannam@167
|
493 T4R = T3I + T3J;
|
cannam@167
|
494 T4S = T4Q + T4R;
|
cannam@167
|
495 T4T = FNMS(KP250000000, T4S, T4P);
|
cannam@167
|
496 T4U = T4Q - T4R;
|
cannam@167
|
497 }
|
cannam@167
|
498 ci[WS(rs, 14)] = T4S + T4P;
|
cannam@167
|
499 T4Z = FMA(KP559016994, T4U, T4T);
|
cannam@167
|
500 ci[WS(rs, 10)] = FMA(KP951056516, T50, T4Z);
|
cannam@167
|
501 ci[WS(rs, 18)] = FNMS(KP951056516, T50, T4Z);
|
cannam@167
|
502 T4V = FNMS(KP559016994, T4U, T4T);
|
cannam@167
|
503 cr[WS(rs, 13)] = FMS(KP951056516, T4Y, T4V);
|
cannam@167
|
504 cr[WS(rs, 17)] = -(FMA(KP951056516, T4Y, T4V));
|
cannam@167
|
505 }
|
cannam@167
|
506 {
|
cannam@167
|
507 E T4K, T4M, T4B, T4E, T4F, T4G, T4L, T4H;
|
cannam@167
|
508 {
|
cannam@167
|
509 E T4I, T4J, T4C, T4D;
|
cannam@167
|
510 T4I = T2f - T2e;
|
cannam@167
|
511 T4J = T2b - T2c;
|
cannam@167
|
512 T4K = FMA(KP618033988, T4J, T4I);
|
cannam@167
|
513 T4M = FNMS(KP618033988, T4I, T4J);
|
cannam@167
|
514 T4B = T4r + T4q;
|
cannam@167
|
515 T4C = T4b + T4c;
|
cannam@167
|
516 T4D = T4e + T4f;
|
cannam@167
|
517 T4E = T4C + T4D;
|
cannam@167
|
518 T4F = FNMS(KP250000000, T4E, T4B);
|
cannam@167
|
519 T4G = T4C - T4D;
|
cannam@167
|
520 }
|
cannam@167
|
521 ci[WS(rs, 19)] = T4E + T4B;
|
cannam@167
|
522 T4L = FMA(KP559016994, T4G, T4F);
|
cannam@167
|
523 cr[WS(rs, 16)] = FMS(KP951056516, T4M, T4L);
|
cannam@167
|
524 ci[WS(rs, 15)] = FMA(KP951056516, T4M, T4L);
|
cannam@167
|
525 T4H = FNMS(KP559016994, T4G, T4F);
|
cannam@167
|
526 cr[WS(rs, 12)] = FMS(KP951056516, T4K, T4H);
|
cannam@167
|
527 ci[WS(rs, 11)] = FMA(KP951056516, T4K, T4H);
|
cannam@167
|
528 }
|
cannam@167
|
529 {
|
cannam@167
|
530 E T5a, T5c, T54, T53, T55, T56, T5b, T57;
|
cannam@167
|
531 {
|
cannam@167
|
532 E T58, T59, T51, T52;
|
cannam@167
|
533 T58 = T2v - T2C;
|
cannam@167
|
534 T59 = T2K - T2R;
|
cannam@167
|
535 T5a = FMA(KP618033988, T59, T58);
|
cannam@167
|
536 T5c = FNMS(KP618033988, T58, T59);
|
cannam@167
|
537 T54 = T4O + T4N;
|
cannam@167
|
538 T51 = T3a + T33;
|
cannam@167
|
539 T52 = T3p + T3i;
|
cannam@167
|
540 T53 = T51 + T52;
|
cannam@167
|
541 T55 = FMA(KP250000000, T53, T54);
|
cannam@167
|
542 T56 = T51 - T52;
|
cannam@167
|
543 }
|
cannam@167
|
544 cr[WS(rs, 15)] = T53 - T54;
|
cannam@167
|
545 T5b = FMA(KP559016994, T56, T55);
|
cannam@167
|
546 ci[WS(rs, 12)] = FMA(KP951056516, T5c, T5b);
|
cannam@167
|
547 ci[WS(rs, 16)] = FNMS(KP951056516, T5c, T5b);
|
cannam@167
|
548 T57 = FNMS(KP559016994, T56, T55);
|
cannam@167
|
549 cr[WS(rs, 11)] = FMS(KP951056516, T5a, T57);
|
cannam@167
|
550 cr[WS(rs, 19)] = -(FMA(KP951056516, T5a, T57));
|
cannam@167
|
551 }
|
cannam@167
|
552 }
|
cannam@167
|
553 }
|
cannam@167
|
554 }
|
cannam@167
|
555
|
cannam@167
|
556 static const tw_instr twinstr[] = {
|
cannam@167
|
557 {TW_FULL, 1, 20},
|
cannam@167
|
558 {TW_NEXT, 1, 0}
|
cannam@167
|
559 };
|
cannam@167
|
560
|
cannam@167
|
561 static const hc2hc_desc desc = { 20, "hf_20", twinstr, &GENUS, {136, 38, 110, 0} };
|
cannam@167
|
562
|
cannam@167
|
563 void X(codelet_hf_20) (planner *p) {
|
cannam@167
|
564 X(khc2hc_register) (p, hf_20, &desc);
|
cannam@167
|
565 }
|
cannam@167
|
566 #else
|
cannam@167
|
567
|
cannam@167
|
568 /* Generated by: ../../../genfft/gen_hc2hc.native -compact -variables 4 -pipeline-latency 4 -n 20 -dit -name hf_20 -include rdft/scalar/hf.h */
|
cannam@167
|
569
|
cannam@167
|
570 /*
|
cannam@167
|
571 * This function contains 246 FP additions, 124 FP multiplications,
|
cannam@167
|
572 * (or, 184 additions, 62 multiplications, 62 fused multiply/add),
|
cannam@167
|
573 * 85 stack variables, 4 constants, and 80 memory accesses
|
cannam@167
|
574 */
|
cannam@167
|
575 #include "rdft/scalar/hf.h"
|
cannam@167
|
576
|
cannam@167
|
577 static void hf_20(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
|
cannam@167
|
578 {
|
cannam@167
|
579 DK(KP587785252, +0.587785252292473129168705954639072768597652438);
|
cannam@167
|
580 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
|
cannam@167
|
581 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
|
cannam@167
|
582 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
|
cannam@167
|
583 {
|
cannam@167
|
584 INT m;
|
cannam@167
|
585 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)) {
|
cannam@167
|
586 E Tj, T1R, T4j, T4s, T2q, T37, T3Q, T42, T1r, T1O, T1P, T3i, T3l, T3J, T3D;
|
cannam@167
|
587 E T3E, T44, T1V, T1W, T1X, T2e, T2j, T2k, T2W, T2X, T4f, T33, T34, T35, T2J;
|
cannam@167
|
588 E T2O, T4q, TG, T13, T14, T3p, T3s, T3K, T3A, T3B, T43, T1S, T1T, T1U, T23;
|
cannam@167
|
589 E T28, T29, T2T, T2U, T4e, T30, T31, T32, T2y, T2D, T4p;
|
cannam@167
|
590 {
|
cannam@167
|
591 E T1, T3N, T6, T3M, Tc, T2n, Th, T2o;
|
cannam@167
|
592 T1 = cr[0];
|
cannam@167
|
593 T3N = ci[0];
|
cannam@167
|
594 {
|
cannam@167
|
595 E T3, T5, T2, T4;
|
cannam@167
|
596 T3 = cr[WS(rs, 10)];
|
cannam@167
|
597 T5 = ci[WS(rs, 10)];
|
cannam@167
|
598 T2 = W[18];
|
cannam@167
|
599 T4 = W[19];
|
cannam@167
|
600 T6 = FMA(T2, T3, T4 * T5);
|
cannam@167
|
601 T3M = FNMS(T4, T3, T2 * T5);
|
cannam@167
|
602 }
|
cannam@167
|
603 {
|
cannam@167
|
604 E T9, Tb, T8, Ta;
|
cannam@167
|
605 T9 = cr[WS(rs, 5)];
|
cannam@167
|
606 Tb = ci[WS(rs, 5)];
|
cannam@167
|
607 T8 = W[8];
|
cannam@167
|
608 Ta = W[9];
|
cannam@167
|
609 Tc = FMA(T8, T9, Ta * Tb);
|
cannam@167
|
610 T2n = FNMS(Ta, T9, T8 * Tb);
|
cannam@167
|
611 }
|
cannam@167
|
612 {
|
cannam@167
|
613 E Te, Tg, Td, Tf;
|
cannam@167
|
614 Te = cr[WS(rs, 15)];
|
cannam@167
|
615 Tg = ci[WS(rs, 15)];
|
cannam@167
|
616 Td = W[28];
|
cannam@167
|
617 Tf = W[29];
|
cannam@167
|
618 Th = FMA(Td, Te, Tf * Tg);
|
cannam@167
|
619 T2o = FNMS(Tf, Te, Td * Tg);
|
cannam@167
|
620 }
|
cannam@167
|
621 {
|
cannam@167
|
622 E T7, Ti, T4h, T4i;
|
cannam@167
|
623 T7 = T1 + T6;
|
cannam@167
|
624 Ti = Tc + Th;
|
cannam@167
|
625 Tj = T7 - Ti;
|
cannam@167
|
626 T1R = T7 + Ti;
|
cannam@167
|
627 T4h = T3N - T3M;
|
cannam@167
|
628 T4i = Tc - Th;
|
cannam@167
|
629 T4j = T4h - T4i;
|
cannam@167
|
630 T4s = T4i + T4h;
|
cannam@167
|
631 }
|
cannam@167
|
632 {
|
cannam@167
|
633 E T2m, T2p, T3O, T3P;
|
cannam@167
|
634 T2m = T1 - T6;
|
cannam@167
|
635 T2p = T2n - T2o;
|
cannam@167
|
636 T2q = T2m - T2p;
|
cannam@167
|
637 T37 = T2m + T2p;
|
cannam@167
|
638 T3O = T3M + T3N;
|
cannam@167
|
639 T3P = T2n + T2o;
|
cannam@167
|
640 T3Q = T3O - T3P;
|
cannam@167
|
641 T42 = T3P + T3O;
|
cannam@167
|
642 }
|
cannam@167
|
643 }
|
cannam@167
|
644 {
|
cannam@167
|
645 E T1f, T3g, T2a, T2H, T1N, T3j, T2i, T2N, T1q, T3h, T2d, T2I, T1C, T3k, T2f;
|
cannam@167
|
646 E T2M;
|
cannam@167
|
647 {
|
cannam@167
|
648 E T19, T2F, T1e, T2G;
|
cannam@167
|
649 {
|
cannam@167
|
650 E T16, T18, T15, T17;
|
cannam@167
|
651 T16 = cr[WS(rs, 8)];
|
cannam@167
|
652 T18 = ci[WS(rs, 8)];
|
cannam@167
|
653 T15 = W[14];
|
cannam@167
|
654 T17 = W[15];
|
cannam@167
|
655 T19 = FMA(T15, T16, T17 * T18);
|
cannam@167
|
656 T2F = FNMS(T17, T16, T15 * T18);
|
cannam@167
|
657 }
|
cannam@167
|
658 {
|
cannam@167
|
659 E T1b, T1d, T1a, T1c;
|
cannam@167
|
660 T1b = cr[WS(rs, 18)];
|
cannam@167
|
661 T1d = ci[WS(rs, 18)];
|
cannam@167
|
662 T1a = W[34];
|
cannam@167
|
663 T1c = W[35];
|
cannam@167
|
664 T1e = FMA(T1a, T1b, T1c * T1d);
|
cannam@167
|
665 T2G = FNMS(T1c, T1b, T1a * T1d);
|
cannam@167
|
666 }
|
cannam@167
|
667 T1f = T19 + T1e;
|
cannam@167
|
668 T3g = T2F + T2G;
|
cannam@167
|
669 T2a = T19 - T1e;
|
cannam@167
|
670 T2H = T2F - T2G;
|
cannam@167
|
671 }
|
cannam@167
|
672 {
|
cannam@167
|
673 E T1H, T2g, T1M, T2h;
|
cannam@167
|
674 {
|
cannam@167
|
675 E T1E, T1G, T1D, T1F;
|
cannam@167
|
676 T1E = cr[WS(rs, 17)];
|
cannam@167
|
677 T1G = ci[WS(rs, 17)];
|
cannam@167
|
678 T1D = W[32];
|
cannam@167
|
679 T1F = W[33];
|
cannam@167
|
680 T1H = FMA(T1D, T1E, T1F * T1G);
|
cannam@167
|
681 T2g = FNMS(T1F, T1E, T1D * T1G);
|
cannam@167
|
682 }
|
cannam@167
|
683 {
|
cannam@167
|
684 E T1J, T1L, T1I, T1K;
|
cannam@167
|
685 T1J = cr[WS(rs, 7)];
|
cannam@167
|
686 T1L = ci[WS(rs, 7)];
|
cannam@167
|
687 T1I = W[12];
|
cannam@167
|
688 T1K = W[13];
|
cannam@167
|
689 T1M = FMA(T1I, T1J, T1K * T1L);
|
cannam@167
|
690 T2h = FNMS(T1K, T1J, T1I * T1L);
|
cannam@167
|
691 }
|
cannam@167
|
692 T1N = T1H + T1M;
|
cannam@167
|
693 T3j = T2g + T2h;
|
cannam@167
|
694 T2i = T2g - T2h;
|
cannam@167
|
695 T2N = T1H - T1M;
|
cannam@167
|
696 }
|
cannam@167
|
697 {
|
cannam@167
|
698 E T1k, T2b, T1p, T2c;
|
cannam@167
|
699 {
|
cannam@167
|
700 E T1h, T1j, T1g, T1i;
|
cannam@167
|
701 T1h = cr[WS(rs, 13)];
|
cannam@167
|
702 T1j = ci[WS(rs, 13)];
|
cannam@167
|
703 T1g = W[24];
|
cannam@167
|
704 T1i = W[25];
|
cannam@167
|
705 T1k = FMA(T1g, T1h, T1i * T1j);
|
cannam@167
|
706 T2b = FNMS(T1i, T1h, T1g * T1j);
|
cannam@167
|
707 }
|
cannam@167
|
708 {
|
cannam@167
|
709 E T1m, T1o, T1l, T1n;
|
cannam@167
|
710 T1m = cr[WS(rs, 3)];
|
cannam@167
|
711 T1o = ci[WS(rs, 3)];
|
cannam@167
|
712 T1l = W[4];
|
cannam@167
|
713 T1n = W[5];
|
cannam@167
|
714 T1p = FMA(T1l, T1m, T1n * T1o);
|
cannam@167
|
715 T2c = FNMS(T1n, T1m, T1l * T1o);
|
cannam@167
|
716 }
|
cannam@167
|
717 T1q = T1k + T1p;
|
cannam@167
|
718 T3h = T2b + T2c;
|
cannam@167
|
719 T2d = T2b - T2c;
|
cannam@167
|
720 T2I = T1k - T1p;
|
cannam@167
|
721 }
|
cannam@167
|
722 {
|
cannam@167
|
723 E T1w, T2K, T1B, T2L;
|
cannam@167
|
724 {
|
cannam@167
|
725 E T1t, T1v, T1s, T1u;
|
cannam@167
|
726 T1t = cr[WS(rs, 12)];
|
cannam@167
|
727 T1v = ci[WS(rs, 12)];
|
cannam@167
|
728 T1s = W[22];
|
cannam@167
|
729 T1u = W[23];
|
cannam@167
|
730 T1w = FMA(T1s, T1t, T1u * T1v);
|
cannam@167
|
731 T2K = FNMS(T1u, T1t, T1s * T1v);
|
cannam@167
|
732 }
|
cannam@167
|
733 {
|
cannam@167
|
734 E T1y, T1A, T1x, T1z;
|
cannam@167
|
735 T1y = cr[WS(rs, 2)];
|
cannam@167
|
736 T1A = ci[WS(rs, 2)];
|
cannam@167
|
737 T1x = W[2];
|
cannam@167
|
738 T1z = W[3];
|
cannam@167
|
739 T1B = FMA(T1x, T1y, T1z * T1A);
|
cannam@167
|
740 T2L = FNMS(T1z, T1y, T1x * T1A);
|
cannam@167
|
741 }
|
cannam@167
|
742 T1C = T1w + T1B;
|
cannam@167
|
743 T3k = T2K + T2L;
|
cannam@167
|
744 T2f = T1w - T1B;
|
cannam@167
|
745 T2M = T2K - T2L;
|
cannam@167
|
746 }
|
cannam@167
|
747 T1r = T1f - T1q;
|
cannam@167
|
748 T1O = T1C - T1N;
|
cannam@167
|
749 T1P = T1r + T1O;
|
cannam@167
|
750 T3i = T3g - T3h;
|
cannam@167
|
751 T3l = T3j - T3k;
|
cannam@167
|
752 T3J = T3l - T3i;
|
cannam@167
|
753 T3D = T3g + T3h;
|
cannam@167
|
754 T3E = T3k + T3j;
|
cannam@167
|
755 T44 = T3D + T3E;
|
cannam@167
|
756 T1V = T1f + T1q;
|
cannam@167
|
757 T1W = T1C + T1N;
|
cannam@167
|
758 T1X = T1V + T1W;
|
cannam@167
|
759 T2e = T2a - T2d;
|
cannam@167
|
760 T2j = T2f - T2i;
|
cannam@167
|
761 T2k = T2e + T2j;
|
cannam@167
|
762 T2W = T2H - T2I;
|
cannam@167
|
763 T2X = T2M - T2N;
|
cannam@167
|
764 T4f = T2W + T2X;
|
cannam@167
|
765 T33 = T2a + T2d;
|
cannam@167
|
766 T34 = T2f + T2i;
|
cannam@167
|
767 T35 = T33 + T34;
|
cannam@167
|
768 T2J = T2H + T2I;
|
cannam@167
|
769 T2O = T2M + T2N;
|
cannam@167
|
770 T4q = T2J + T2O;
|
cannam@167
|
771 }
|
cannam@167
|
772 {
|
cannam@167
|
773 E Tu, T3n, T1Z, T2w, T12, T3r, T27, T2z, TF, T3o, T22, T2x, TR, T3q, T24;
|
cannam@167
|
774 E T2C;
|
cannam@167
|
775 {
|
cannam@167
|
776 E To, T2u, Tt, T2v;
|
cannam@167
|
777 {
|
cannam@167
|
778 E Tl, Tn, Tk, Tm;
|
cannam@167
|
779 Tl = cr[WS(rs, 4)];
|
cannam@167
|
780 Tn = ci[WS(rs, 4)];
|
cannam@167
|
781 Tk = W[6];
|
cannam@167
|
782 Tm = W[7];
|
cannam@167
|
783 To = FMA(Tk, Tl, Tm * Tn);
|
cannam@167
|
784 T2u = FNMS(Tm, Tl, Tk * Tn);
|
cannam@167
|
785 }
|
cannam@167
|
786 {
|
cannam@167
|
787 E Tq, Ts, Tp, Tr;
|
cannam@167
|
788 Tq = cr[WS(rs, 14)];
|
cannam@167
|
789 Ts = ci[WS(rs, 14)];
|
cannam@167
|
790 Tp = W[26];
|
cannam@167
|
791 Tr = W[27];
|
cannam@167
|
792 Tt = FMA(Tp, Tq, Tr * Ts);
|
cannam@167
|
793 T2v = FNMS(Tr, Tq, Tp * Ts);
|
cannam@167
|
794 }
|
cannam@167
|
795 Tu = To + Tt;
|
cannam@167
|
796 T3n = T2u + T2v;
|
cannam@167
|
797 T1Z = To - Tt;
|
cannam@167
|
798 T2w = T2u - T2v;
|
cannam@167
|
799 }
|
cannam@167
|
800 {
|
cannam@167
|
801 E TW, T25, T11, T26;
|
cannam@167
|
802 {
|
cannam@167
|
803 E TT, TV, TS, TU;
|
cannam@167
|
804 TT = cr[WS(rs, 1)];
|
cannam@167
|
805 TV = ci[WS(rs, 1)];
|
cannam@167
|
806 TS = W[0];
|
cannam@167
|
807 TU = W[1];
|
cannam@167
|
808 TW = FMA(TS, TT, TU * TV);
|
cannam@167
|
809 T25 = FNMS(TU, TT, TS * TV);
|
cannam@167
|
810 }
|
cannam@167
|
811 {
|
cannam@167
|
812 E TY, T10, TX, TZ;
|
cannam@167
|
813 TY = cr[WS(rs, 11)];
|
cannam@167
|
814 T10 = ci[WS(rs, 11)];
|
cannam@167
|
815 TX = W[20];
|
cannam@167
|
816 TZ = W[21];
|
cannam@167
|
817 T11 = FMA(TX, TY, TZ * T10);
|
cannam@167
|
818 T26 = FNMS(TZ, TY, TX * T10);
|
cannam@167
|
819 }
|
cannam@167
|
820 T12 = TW + T11;
|
cannam@167
|
821 T3r = T25 + T26;
|
cannam@167
|
822 T27 = T25 - T26;
|
cannam@167
|
823 T2z = T11 - TW;
|
cannam@167
|
824 }
|
cannam@167
|
825 {
|
cannam@167
|
826 E Tz, T20, TE, T21;
|
cannam@167
|
827 {
|
cannam@167
|
828 E Tw, Ty, Tv, Tx;
|
cannam@167
|
829 Tw = cr[WS(rs, 9)];
|
cannam@167
|
830 Ty = ci[WS(rs, 9)];
|
cannam@167
|
831 Tv = W[16];
|
cannam@167
|
832 Tx = W[17];
|
cannam@167
|
833 Tz = FMA(Tv, Tw, Tx * Ty);
|
cannam@167
|
834 T20 = FNMS(Tx, Tw, Tv * Ty);
|
cannam@167
|
835 }
|
cannam@167
|
836 {
|
cannam@167
|
837 E TB, TD, TA, TC;
|
cannam@167
|
838 TB = cr[WS(rs, 19)];
|
cannam@167
|
839 TD = ci[WS(rs, 19)];
|
cannam@167
|
840 TA = W[36];
|
cannam@167
|
841 TC = W[37];
|
cannam@167
|
842 TE = FMA(TA, TB, TC * TD);
|
cannam@167
|
843 T21 = FNMS(TC, TB, TA * TD);
|
cannam@167
|
844 }
|
cannam@167
|
845 TF = Tz + TE;
|
cannam@167
|
846 T3o = T20 + T21;
|
cannam@167
|
847 T22 = T20 - T21;
|
cannam@167
|
848 T2x = Tz - TE;
|
cannam@167
|
849 }
|
cannam@167
|
850 {
|
cannam@167
|
851 E TL, T2A, TQ, T2B;
|
cannam@167
|
852 {
|
cannam@167
|
853 E TI, TK, TH, TJ;
|
cannam@167
|
854 TI = cr[WS(rs, 16)];
|
cannam@167
|
855 TK = ci[WS(rs, 16)];
|
cannam@167
|
856 TH = W[30];
|
cannam@167
|
857 TJ = W[31];
|
cannam@167
|
858 TL = FMA(TH, TI, TJ * TK);
|
cannam@167
|
859 T2A = FNMS(TJ, TI, TH * TK);
|
cannam@167
|
860 }
|
cannam@167
|
861 {
|
cannam@167
|
862 E TN, TP, TM, TO;
|
cannam@167
|
863 TN = cr[WS(rs, 6)];
|
cannam@167
|
864 TP = ci[WS(rs, 6)];
|
cannam@167
|
865 TM = W[10];
|
cannam@167
|
866 TO = W[11];
|
cannam@167
|
867 TQ = FMA(TM, TN, TO * TP);
|
cannam@167
|
868 T2B = FNMS(TO, TN, TM * TP);
|
cannam@167
|
869 }
|
cannam@167
|
870 TR = TL + TQ;
|
cannam@167
|
871 T3q = T2A + T2B;
|
cannam@167
|
872 T24 = TL - TQ;
|
cannam@167
|
873 T2C = T2A - T2B;
|
cannam@167
|
874 }
|
cannam@167
|
875 TG = Tu - TF;
|
cannam@167
|
876 T13 = TR - T12;
|
cannam@167
|
877 T14 = TG + T13;
|
cannam@167
|
878 T3p = T3n - T3o;
|
cannam@167
|
879 T3s = T3q - T3r;
|
cannam@167
|
880 T3K = T3p + T3s;
|
cannam@167
|
881 T3A = T3n + T3o;
|
cannam@167
|
882 T3B = T3q + T3r;
|
cannam@167
|
883 T43 = T3A + T3B;
|
cannam@167
|
884 T1S = Tu + TF;
|
cannam@167
|
885 T1T = TR + T12;
|
cannam@167
|
886 T1U = T1S + T1T;
|
cannam@167
|
887 T23 = T1Z - T22;
|
cannam@167
|
888 T28 = T24 - T27;
|
cannam@167
|
889 T29 = T23 + T28;
|
cannam@167
|
890 T2T = T2w - T2x;
|
cannam@167
|
891 T2U = T2C + T2z;
|
cannam@167
|
892 T4e = T2T + T2U;
|
cannam@167
|
893 T30 = T1Z + T22;
|
cannam@167
|
894 T31 = T24 + T27;
|
cannam@167
|
895 T32 = T30 + T31;
|
cannam@167
|
896 T2y = T2w + T2x;
|
cannam@167
|
897 T2D = T2z - T2C;
|
cannam@167
|
898 T4p = T2D - T2y;
|
cannam@167
|
899 }
|
cannam@167
|
900 {
|
cannam@167
|
901 E T3e, T1Q, T3d, T3u, T3w, T3m, T3t, T3v, T3f;
|
cannam@167
|
902 T3e = KP559016994 * (T14 - T1P);
|
cannam@167
|
903 T1Q = T14 + T1P;
|
cannam@167
|
904 T3d = FNMS(KP250000000, T1Q, Tj);
|
cannam@167
|
905 T3m = T3i + T3l;
|
cannam@167
|
906 T3t = T3p - T3s;
|
cannam@167
|
907 T3u = FNMS(KP587785252, T3t, KP951056516 * T3m);
|
cannam@167
|
908 T3w = FMA(KP951056516, T3t, KP587785252 * T3m);
|
cannam@167
|
909 ci[WS(rs, 9)] = Tj + T1Q;
|
cannam@167
|
910 T3v = T3e + T3d;
|
cannam@167
|
911 ci[WS(rs, 5)] = T3v - T3w;
|
cannam@167
|
912 cr[WS(rs, 6)] = T3v + T3w;
|
cannam@167
|
913 T3f = T3d - T3e;
|
cannam@167
|
914 cr[WS(rs, 2)] = T3f - T3u;
|
cannam@167
|
915 ci[WS(rs, 1)] = T3f + T3u;
|
cannam@167
|
916 }
|
cannam@167
|
917 {
|
cannam@167
|
918 E T36, T38, T39, T2Z, T3c, T2V, T2Y, T3b, T3a;
|
cannam@167
|
919 T36 = KP559016994 * (T32 - T35);
|
cannam@167
|
920 T38 = T32 + T35;
|
cannam@167
|
921 T39 = FNMS(KP250000000, T38, T37);
|
cannam@167
|
922 T2V = T2T - T2U;
|
cannam@167
|
923 T2Y = T2W - T2X;
|
cannam@167
|
924 T2Z = FMA(KP951056516, T2V, KP587785252 * T2Y);
|
cannam@167
|
925 T3c = FNMS(KP587785252, T2V, KP951056516 * T2Y);
|
cannam@167
|
926 cr[WS(rs, 5)] = T37 + T38;
|
cannam@167
|
927 T3b = T39 - T36;
|
cannam@167
|
928 ci[WS(rs, 2)] = T3b - T3c;
|
cannam@167
|
929 ci[WS(rs, 6)] = T3c + T3b;
|
cannam@167
|
930 T3a = T36 + T39;
|
cannam@167
|
931 cr[WS(rs, 1)] = T2Z + T3a;
|
cannam@167
|
932 cr[WS(rs, 9)] = T3a - T2Z;
|
cannam@167
|
933 }
|
cannam@167
|
934 {
|
cannam@167
|
935 E T3x, T1Y, T3y, T3G, T3I, T3C, T3F, T3H, T3z;
|
cannam@167
|
936 T3x = KP559016994 * (T1U - T1X);
|
cannam@167
|
937 T1Y = T1U + T1X;
|
cannam@167
|
938 T3y = FNMS(KP250000000, T1Y, T1R);
|
cannam@167
|
939 T3C = T3A - T3B;
|
cannam@167
|
940 T3F = T3D - T3E;
|
cannam@167
|
941 T3G = FMA(KP951056516, T3C, KP587785252 * T3F);
|
cannam@167
|
942 T3I = FNMS(KP587785252, T3C, KP951056516 * T3F);
|
cannam@167
|
943 cr[0] = T1R + T1Y;
|
cannam@167
|
944 T3H = T3y - T3x;
|
cannam@167
|
945 ci[WS(rs, 7)] = T3H - T3I;
|
cannam@167
|
946 cr[WS(rs, 8)] = T3H + T3I;
|
cannam@167
|
947 T3z = T3x + T3y;
|
cannam@167
|
948 cr[WS(rs, 4)] = T3z - T3G;
|
cannam@167
|
949 ci[WS(rs, 3)] = T3z + T3G;
|
cannam@167
|
950 }
|
cannam@167
|
951 {
|
cannam@167
|
952 E T2l, T2r, T2s, T2Q, T2R, T2E, T2P, T2S, T2t;
|
cannam@167
|
953 T2l = KP559016994 * (T29 - T2k);
|
cannam@167
|
954 T2r = T29 + T2k;
|
cannam@167
|
955 T2s = FNMS(KP250000000, T2r, T2q);
|
cannam@167
|
956 T2E = T2y + T2D;
|
cannam@167
|
957 T2P = T2J - T2O;
|
cannam@167
|
958 T2Q = FMA(KP951056516, T2E, KP587785252 * T2P);
|
cannam@167
|
959 T2R = FNMS(KP587785252, T2E, KP951056516 * T2P);
|
cannam@167
|
960 ci[WS(rs, 4)] = T2q + T2r;
|
cannam@167
|
961 T2S = T2s - T2l;
|
cannam@167
|
962 cr[WS(rs, 3)] = T2R + T2S;
|
cannam@167
|
963 cr[WS(rs, 7)] = T2S - T2R;
|
cannam@167
|
964 T2t = T2l + T2s;
|
cannam@167
|
965 ci[0] = T2t - T2Q;
|
cannam@167
|
966 ci[WS(rs, 8)] = T2Q + T2t;
|
cannam@167
|
967 }
|
cannam@167
|
968 {
|
cannam@167
|
969 E T3U, T3L, T3V, T3T, T3X, T3R, T3S, T3Y, T3W;
|
cannam@167
|
970 T3U = KP559016994 * (T3K + T3J);
|
cannam@167
|
971 T3L = T3J - T3K;
|
cannam@167
|
972 T3V = FMA(KP250000000, T3L, T3Q);
|
cannam@167
|
973 T3R = T13 - TG;
|
cannam@167
|
974 T3S = T1r - T1O;
|
cannam@167
|
975 T3T = FNMS(KP587785252, T3S, KP951056516 * T3R);
|
cannam@167
|
976 T3X = FMA(KP587785252, T3R, KP951056516 * T3S);
|
cannam@167
|
977 cr[WS(rs, 10)] = T3L - T3Q;
|
cannam@167
|
978 T3Y = T3V - T3U;
|
cannam@167
|
979 cr[WS(rs, 18)] = T3X - T3Y;
|
cannam@167
|
980 ci[WS(rs, 17)] = T3X + T3Y;
|
cannam@167
|
981 T3W = T3U + T3V;
|
cannam@167
|
982 cr[WS(rs, 14)] = T3T - T3W;
|
cannam@167
|
983 ci[WS(rs, 13)] = T3T + T3W;
|
cannam@167
|
984 }
|
cannam@167
|
985 {
|
cannam@167
|
986 E T4g, T4k, T4l, T4d, T4n, T4b, T4c, T4o, T4m;
|
cannam@167
|
987 T4g = KP559016994 * (T4e - T4f);
|
cannam@167
|
988 T4k = T4e + T4f;
|
cannam@167
|
989 T4l = FNMS(KP250000000, T4k, T4j);
|
cannam@167
|
990 T4b = T33 - T34;
|
cannam@167
|
991 T4c = T30 - T31;
|
cannam@167
|
992 T4d = FNMS(KP587785252, T4c, KP951056516 * T4b);
|
cannam@167
|
993 T4n = FMA(KP951056516, T4c, KP587785252 * T4b);
|
cannam@167
|
994 ci[WS(rs, 14)] = T4k + T4j;
|
cannam@167
|
995 T4o = T4g + T4l;
|
cannam@167
|
996 ci[WS(rs, 10)] = T4n + T4o;
|
cannam@167
|
997 ci[WS(rs, 18)] = T4o - T4n;
|
cannam@167
|
998 T4m = T4g - T4l;
|
cannam@167
|
999 cr[WS(rs, 13)] = T4d + T4m;
|
cannam@167
|
1000 cr[WS(rs, 17)] = T4m - T4d;
|
cannam@167
|
1001 }
|
cannam@167
|
1002 {
|
cannam@167
|
1003 E T47, T45, T46, T41, T49, T3Z, T40, T4a, T48;
|
cannam@167
|
1004 T47 = KP559016994 * (T43 - T44);
|
cannam@167
|
1005 T45 = T43 + T44;
|
cannam@167
|
1006 T46 = FNMS(KP250000000, T45, T42);
|
cannam@167
|
1007 T3Z = T1S - T1T;
|
cannam@167
|
1008 T40 = T1V - T1W;
|
cannam@167
|
1009 T41 = FNMS(KP951056516, T40, KP587785252 * T3Z);
|
cannam@167
|
1010 T49 = FMA(KP951056516, T3Z, KP587785252 * T40);
|
cannam@167
|
1011 ci[WS(rs, 19)] = T45 + T42;
|
cannam@167
|
1012 T4a = T47 + T46;
|
cannam@167
|
1013 cr[WS(rs, 16)] = T49 - T4a;
|
cannam@167
|
1014 ci[WS(rs, 15)] = T49 + T4a;
|
cannam@167
|
1015 T48 = T46 - T47;
|
cannam@167
|
1016 cr[WS(rs, 12)] = T41 - T48;
|
cannam@167
|
1017 ci[WS(rs, 11)] = T41 + T48;
|
cannam@167
|
1018 }
|
cannam@167
|
1019 {
|
cannam@167
|
1020 E T4w, T4r, T4x, T4v, T4z, T4t, T4u, T4A, T4y;
|
cannam@167
|
1021 T4w = KP559016994 * (T4p + T4q);
|
cannam@167
|
1022 T4r = T4p - T4q;
|
cannam@167
|
1023 T4x = FMA(KP250000000, T4r, T4s);
|
cannam@167
|
1024 T4t = T23 - T28;
|
cannam@167
|
1025 T4u = T2e - T2j;
|
cannam@167
|
1026 T4v = FMA(KP951056516, T4t, KP587785252 * T4u);
|
cannam@167
|
1027 T4z = FNMS(KP587785252, T4t, KP951056516 * T4u);
|
cannam@167
|
1028 cr[WS(rs, 15)] = T4r - T4s;
|
cannam@167
|
1029 T4A = T4w + T4x;
|
cannam@167
|
1030 ci[WS(rs, 12)] = T4z + T4A;
|
cannam@167
|
1031 ci[WS(rs, 16)] = T4A - T4z;
|
cannam@167
|
1032 T4y = T4w - T4x;
|
cannam@167
|
1033 cr[WS(rs, 11)] = T4v + T4y;
|
cannam@167
|
1034 cr[WS(rs, 19)] = T4y - T4v;
|
cannam@167
|
1035 }
|
cannam@167
|
1036 }
|
cannam@167
|
1037 }
|
cannam@167
|
1038 }
|
cannam@167
|
1039
|
cannam@167
|
1040 static const tw_instr twinstr[] = {
|
cannam@167
|
1041 {TW_FULL, 1, 20},
|
cannam@167
|
1042 {TW_NEXT, 1, 0}
|
cannam@167
|
1043 };
|
cannam@167
|
1044
|
cannam@167
|
1045 static const hc2hc_desc desc = { 20, "hf_20", twinstr, &GENUS, {184, 62, 62, 0} };
|
cannam@167
|
1046
|
cannam@167
|
1047 void X(codelet_hf_20) (planner *p) {
|
cannam@167
|
1048 X(khc2hc_register) (p, hf_20, &desc);
|
cannam@167
|
1049 }
|
cannam@167
|
1050 #endif
|