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 07:42:02 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_hc2c.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -sign 1 -twiddle-log3 -precompute-twiddles -n 20 -dif -name hc2cb2_20 -include hc2cb.h */
|
Chris@10
|
29
|
Chris@10
|
30 /*
|
Chris@10
|
31 * This function contains 276 FP additions, 198 FP multiplications,
|
Chris@10
|
32 * (or, 136 additions, 58 multiplications, 140 fused multiply/add),
|
Chris@10
|
33 * 160 stack variables, 4 constants, and 80 memory accesses
|
Chris@10
|
34 */
|
Chris@10
|
35 #include "hc2cb.h"
|
Chris@10
|
36
|
Chris@10
|
37 static void hc2cb2_20(R *Rp, R *Ip, R *Rm, R *Im, 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(KP618033988, +0.618033988749894848204586834365638117720309180);
|
Chris@10
|
42 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
|
Chris@10
|
43 {
|
Chris@10
|
44 INT m;
|
Chris@10
|
45 for (m = mb, W = W + ((mb - 1) * 8); m < me; m = m + 1, Rp = Rp + ms, Ip = Ip + ms, Rm = Rm - ms, Im = Im - ms, W = W + 8, MAKE_VOLATILE_STRIDE(80, rs)) {
|
Chris@10
|
46 E T1S, T1O, T1s, TI, T24, T1Y, T2g, T2k, TS, TR, T1I, T26, T1o, T20, T1F;
|
Chris@10
|
47 E T25, TT, T1Z;
|
Chris@10
|
48 {
|
Chris@10
|
49 E TD, TH, TE, T1L, T1N, T1X, TG, T1V, T2Y, T2b, T29, T2s, T36, T3e, T31;
|
Chris@10
|
50 E T2o, T3b, T5b, T2c, T2U, T4y, T4u, T2f, T5g, T47, T5p, T4b, T5l;
|
Chris@10
|
51 {
|
Chris@10
|
52 E T1r, TF, T2T, T1M, T1R, T2X, T2r, T4x;
|
Chris@10
|
53 TD = W[0];
|
Chris@10
|
54 TH = W[3];
|
Chris@10
|
55 TE = W[2];
|
Chris@10
|
56 T1L = W[6];
|
Chris@10
|
57 T1N = W[7];
|
Chris@10
|
58 T1r = TD * TH;
|
Chris@10
|
59 TF = TD * TE;
|
Chris@10
|
60 T2T = TE * T1L;
|
Chris@10
|
61 T1M = TD * T1L;
|
Chris@10
|
62 T1R = TD * T1N;
|
Chris@10
|
63 T2X = TE * T1N;
|
Chris@10
|
64 T1X = W[5];
|
Chris@10
|
65 TG = W[1];
|
Chris@10
|
66 T1V = W[4];
|
Chris@10
|
67 T2Y = FNMS(TH, T1L, T2X);
|
Chris@10
|
68 T2r = TD * T1X;
|
Chris@10
|
69 {
|
Chris@10
|
70 E T23, T2n, T1W, T2a;
|
Chris@10
|
71 T23 = TE * T1X;
|
Chris@10
|
72 T1S = FNMS(TG, T1L, T1R);
|
Chris@10
|
73 T1O = FMA(TG, T1N, T1M);
|
Chris@10
|
74 T2b = FMA(TG, TE, T1r);
|
Chris@10
|
75 T1s = FNMS(TG, TE, T1r);
|
Chris@10
|
76 T29 = FNMS(TG, TH, TF);
|
Chris@10
|
77 TI = FMA(TG, TH, TF);
|
Chris@10
|
78 T2n = TD * T1V;
|
Chris@10
|
79 T1W = TE * T1V;
|
Chris@10
|
80 T2s = FMA(TG, T1V, T2r);
|
Chris@10
|
81 T36 = FNMS(TG, T1V, T2r);
|
Chris@10
|
82 T3e = FMA(TH, T1V, T23);
|
Chris@10
|
83 T24 = FNMS(TH, T1V, T23);
|
Chris@10
|
84 T2a = T29 * T1V;
|
Chris@10
|
85 T31 = FMA(TG, T1X, T2n);
|
Chris@10
|
86 T2o = FNMS(TG, T1X, T2n);
|
Chris@10
|
87 T3b = FNMS(TH, T1X, T1W);
|
Chris@10
|
88 T1Y = FMA(TH, T1X, T1W);
|
Chris@10
|
89 T5b = FNMS(T2b, T1X, T2a);
|
Chris@10
|
90 T2c = FMA(T2b, T1X, T2a);
|
Chris@10
|
91 T2U = FMA(TH, T1N, T2T);
|
Chris@10
|
92 }
|
Chris@10
|
93 T4x = T29 * T1N;
|
Chris@10
|
94 {
|
Chris@10
|
95 E T4t, T2d, T2j, T2e;
|
Chris@10
|
96 T4t = T29 * T1L;
|
Chris@10
|
97 T2e = T29 * T1X;
|
Chris@10
|
98 T4y = FNMS(T2b, T1L, T4x);
|
Chris@10
|
99 T4u = FMA(T2b, T1N, T4t);
|
Chris@10
|
100 T2f = FNMS(T2b, T1V, T2e);
|
Chris@10
|
101 T5g = FMA(T2b, T1V, T2e);
|
Chris@10
|
102 T2d = T2c * T1L;
|
Chris@10
|
103 T2j = T2c * T1N;
|
Chris@10
|
104 T47 = TI * T1V;
|
Chris@10
|
105 T2g = FMA(T2f, T1N, T2d);
|
Chris@10
|
106 T2k = FNMS(T2f, T1L, T2j);
|
Chris@10
|
107 T5p = TI * T1N;
|
Chris@10
|
108 T4b = TI * T1X;
|
Chris@10
|
109 T5l = TI * T1L;
|
Chris@10
|
110 }
|
Chris@10
|
111 }
|
Chris@10
|
112 {
|
Chris@10
|
113 E T4f, T48, T4c, T4k, T5m, T5q, T3V, T4V, TJ, T7, T3j, T4B, T2H, T1z, T3q;
|
Chris@10
|
114 E T43, T1n, T52, T42, T3x, T53, T2D, T18, T2A, T1H, T4R, T4X, T4W, T4O, T1G;
|
Chris@10
|
115 E T2O, T3I, T2P, T3P, T2K, T2M, T1C, T1E, TC, T2w, T40, T3Y, T4K, T4I, TQ;
|
Chris@10
|
116 {
|
Chris@10
|
117 E T3h, T3, T1w, T3T, T1v, T3U, T6, T1x;
|
Chris@10
|
118 {
|
Chris@10
|
119 E T1t, T1u, T1, T2, T4, T5;
|
Chris@10
|
120 T1 = Rp[0];
|
Chris@10
|
121 T2 = Rm[WS(rs, 9)];
|
Chris@10
|
122 T1t = Ip[0];
|
Chris@10
|
123 T4f = FNMS(T1s, T1X, T47);
|
Chris@10
|
124 T48 = FMA(T1s, T1X, T47);
|
Chris@10
|
125 T4c = FNMS(T1s, T1V, T4b);
|
Chris@10
|
126 T4k = FMA(T1s, T1V, T4b);
|
Chris@10
|
127 T5m = FMA(T1s, T1N, T5l);
|
Chris@10
|
128 T5q = FNMS(T1s, T1L, T5p);
|
Chris@10
|
129 T3h = T1 - T2;
|
Chris@10
|
130 T3 = T1 + T2;
|
Chris@10
|
131 T1u = Im[WS(rs, 9)];
|
Chris@10
|
132 T4 = Rp[WS(rs, 5)];
|
Chris@10
|
133 T5 = Rm[WS(rs, 4)];
|
Chris@10
|
134 T1w = Ip[WS(rs, 5)];
|
Chris@10
|
135 T3T = T1t + T1u;
|
Chris@10
|
136 T1v = T1t - T1u;
|
Chris@10
|
137 T3U = T4 - T5;
|
Chris@10
|
138 T6 = T4 + T5;
|
Chris@10
|
139 T1x = Im[WS(rs, 4)];
|
Chris@10
|
140 }
|
Chris@10
|
141 {
|
Chris@10
|
142 E T3L, T4M, TK, Te, T3m, T4C, T2y, T1f, T3H, T4Q, TO, TA, T3w, T4G, T2C;
|
Chris@10
|
143 E T17, T3O, T4N, TL, Tl, T3p, T4D, T2z, T1m, T3r, Tp, TX, T3C, TW, T3D;
|
Chris@10
|
144 E Ts, TY;
|
Chris@10
|
145 {
|
Chris@10
|
146 E T3u, Tw, T14, T3G, T13, T3F, Tz, T15;
|
Chris@10
|
147 {
|
Chris@10
|
148 E T3k, Ta, T1c, T3J, T1b, T3K, Td, T1d;
|
Chris@10
|
149 {
|
Chris@10
|
150 E T19, T1a, Tb, Tc;
|
Chris@10
|
151 {
|
Chris@10
|
152 E T8, T3i, T1y, T9;
|
Chris@10
|
153 T8 = Rp[WS(rs, 4)];
|
Chris@10
|
154 T3V = T3T - T3U;
|
Chris@10
|
155 T4V = T3U + T3T;
|
Chris@10
|
156 TJ = T3 - T6;
|
Chris@10
|
157 T7 = T3 + T6;
|
Chris@10
|
158 T3i = T1w + T1x;
|
Chris@10
|
159 T1y = T1w - T1x;
|
Chris@10
|
160 T9 = Rm[WS(rs, 5)];
|
Chris@10
|
161 T19 = Ip[WS(rs, 4)];
|
Chris@10
|
162 T3j = T3h + T3i;
|
Chris@10
|
163 T4B = T3h - T3i;
|
Chris@10
|
164 T2H = T1v + T1y;
|
Chris@10
|
165 T1z = T1v - T1y;
|
Chris@10
|
166 T3k = T8 - T9;
|
Chris@10
|
167 Ta = T8 + T9;
|
Chris@10
|
168 T1a = Im[WS(rs, 5)];
|
Chris@10
|
169 }
|
Chris@10
|
170 Tb = Rp[WS(rs, 9)];
|
Chris@10
|
171 Tc = Rm[0];
|
Chris@10
|
172 T1c = Ip[WS(rs, 9)];
|
Chris@10
|
173 T3J = T19 + T1a;
|
Chris@10
|
174 T1b = T19 - T1a;
|
Chris@10
|
175 T3K = Tb - Tc;
|
Chris@10
|
176 Td = Tb + Tc;
|
Chris@10
|
177 T1d = Im[0];
|
Chris@10
|
178 }
|
Chris@10
|
179 {
|
Chris@10
|
180 E T11, T12, Tx, Ty;
|
Chris@10
|
181 {
|
Chris@10
|
182 E Tu, T3l, T1e, Tv;
|
Chris@10
|
183 Tu = Rm[WS(rs, 7)];
|
Chris@10
|
184 T3L = T3J - T3K;
|
Chris@10
|
185 T4M = T3K + T3J;
|
Chris@10
|
186 TK = Ta - Td;
|
Chris@10
|
187 Te = Ta + Td;
|
Chris@10
|
188 T3l = T1c + T1d;
|
Chris@10
|
189 T1e = T1c - T1d;
|
Chris@10
|
190 Tv = Rp[WS(rs, 2)];
|
Chris@10
|
191 T11 = Ip[WS(rs, 2)];
|
Chris@10
|
192 T3m = T3k + T3l;
|
Chris@10
|
193 T4C = T3k - T3l;
|
Chris@10
|
194 T2y = T1b + T1e;
|
Chris@10
|
195 T1f = T1b - T1e;
|
Chris@10
|
196 T3u = Tu - Tv;
|
Chris@10
|
197 Tw = Tu + Tv;
|
Chris@10
|
198 T12 = Im[WS(rs, 7)];
|
Chris@10
|
199 }
|
Chris@10
|
200 Tx = Rm[WS(rs, 2)];
|
Chris@10
|
201 Ty = Rp[WS(rs, 7)];
|
Chris@10
|
202 T14 = Ip[WS(rs, 7)];
|
Chris@10
|
203 T3G = T11 + T12;
|
Chris@10
|
204 T13 = T11 - T12;
|
Chris@10
|
205 T3F = Tx - Ty;
|
Chris@10
|
206 Tz = Tx + Ty;
|
Chris@10
|
207 T15 = Im[WS(rs, 2)];
|
Chris@10
|
208 }
|
Chris@10
|
209 }
|
Chris@10
|
210 {
|
Chris@10
|
211 E T3n, Th, T1j, T3N, T1i, T3M, Tk, T1k;
|
Chris@10
|
212 {
|
Chris@10
|
213 E T1g, T1h, Ti, Tj;
|
Chris@10
|
214 {
|
Chris@10
|
215 E Tf, T3v, T16, Tg;
|
Chris@10
|
216 Tf = Rm[WS(rs, 3)];
|
Chris@10
|
217 T3H = T3F + T3G;
|
Chris@10
|
218 T4Q = T3F - T3G;
|
Chris@10
|
219 TO = Tw - Tz;
|
Chris@10
|
220 TA = Tw + Tz;
|
Chris@10
|
221 T3v = T14 + T15;
|
Chris@10
|
222 T16 = T14 - T15;
|
Chris@10
|
223 Tg = Rp[WS(rs, 6)];
|
Chris@10
|
224 T1g = Ip[WS(rs, 6)];
|
Chris@10
|
225 T3w = T3u - T3v;
|
Chris@10
|
226 T4G = T3u + T3v;
|
Chris@10
|
227 T2C = T13 + T16;
|
Chris@10
|
228 T17 = T13 - T16;
|
Chris@10
|
229 T3n = Tf - Tg;
|
Chris@10
|
230 Th = Tf + Tg;
|
Chris@10
|
231 T1h = Im[WS(rs, 3)];
|
Chris@10
|
232 }
|
Chris@10
|
233 Ti = Rp[WS(rs, 1)];
|
Chris@10
|
234 Tj = Rm[WS(rs, 8)];
|
Chris@10
|
235 T1j = Ip[WS(rs, 1)];
|
Chris@10
|
236 T3N = T1g + T1h;
|
Chris@10
|
237 T1i = T1g - T1h;
|
Chris@10
|
238 T3M = Ti - Tj;
|
Chris@10
|
239 Tk = Ti + Tj;
|
Chris@10
|
240 T1k = Im[WS(rs, 8)];
|
Chris@10
|
241 }
|
Chris@10
|
242 {
|
Chris@10
|
243 E TU, TV, Tq, Tr;
|
Chris@10
|
244 {
|
Chris@10
|
245 E Tn, T3o, T1l, To;
|
Chris@10
|
246 Tn = Rp[WS(rs, 8)];
|
Chris@10
|
247 T3O = T3M + T3N;
|
Chris@10
|
248 T4N = T3M - T3N;
|
Chris@10
|
249 TL = Th - Tk;
|
Chris@10
|
250 Tl = Th + Tk;
|
Chris@10
|
251 T3o = T1j + T1k;
|
Chris@10
|
252 T1l = T1j - T1k;
|
Chris@10
|
253 To = Rm[WS(rs, 1)];
|
Chris@10
|
254 TU = Ip[WS(rs, 8)];
|
Chris@10
|
255 T3p = T3n + T3o;
|
Chris@10
|
256 T4D = T3n - T3o;
|
Chris@10
|
257 T2z = T1i + T1l;
|
Chris@10
|
258 T1m = T1i - T1l;
|
Chris@10
|
259 T3r = Tn - To;
|
Chris@10
|
260 Tp = Tn + To;
|
Chris@10
|
261 TV = Im[WS(rs, 1)];
|
Chris@10
|
262 }
|
Chris@10
|
263 Tq = Rm[WS(rs, 6)];
|
Chris@10
|
264 Tr = Rp[WS(rs, 3)];
|
Chris@10
|
265 TX = Ip[WS(rs, 3)];
|
Chris@10
|
266 T3C = TU + TV;
|
Chris@10
|
267 TW = TU - TV;
|
Chris@10
|
268 T3D = Tq - Tr;
|
Chris@10
|
269 Ts = Tq + Tr;
|
Chris@10
|
270 TY = Im[WS(rs, 6)];
|
Chris@10
|
271 }
|
Chris@10
|
272 }
|
Chris@10
|
273 }
|
Chris@10
|
274 {
|
Chris@10
|
275 E T3E, Tt, T1A, T4E, T4H, T2J, T1B, T2I, TM, TP;
|
Chris@10
|
276 {
|
Chris@10
|
277 E T4P, TN, T3s, TZ;
|
Chris@10
|
278 T3q = T3m + T3p;
|
Chris@10
|
279 T43 = T3m - T3p;
|
Chris@10
|
280 T3E = T3C - T3D;
|
Chris@10
|
281 T4P = T3D + T3C;
|
Chris@10
|
282 TN = Tp - Ts;
|
Chris@10
|
283 Tt = Tp + Ts;
|
Chris@10
|
284 T3s = TX + TY;
|
Chris@10
|
285 TZ = TX - TY;
|
Chris@10
|
286 T1n = T1f - T1m;
|
Chris@10
|
287 T1A = T1f + T1m;
|
Chris@10
|
288 T4E = T4C + T4D;
|
Chris@10
|
289 T52 = T4C - T4D;
|
Chris@10
|
290 {
|
Chris@10
|
291 E T3t, T4F, T2B, T10;
|
Chris@10
|
292 T3t = T3r - T3s;
|
Chris@10
|
293 T4F = T3r + T3s;
|
Chris@10
|
294 T2B = TW + TZ;
|
Chris@10
|
295 T10 = TW - TZ;
|
Chris@10
|
296 T42 = T3t - T3w;
|
Chris@10
|
297 T3x = T3t + T3w;
|
Chris@10
|
298 T4H = T4F + T4G;
|
Chris@10
|
299 T53 = T4F - T4G;
|
Chris@10
|
300 T2D = T2B - T2C;
|
Chris@10
|
301 T2J = T2B + T2C;
|
Chris@10
|
302 T1B = T10 + T17;
|
Chris@10
|
303 T18 = T10 - T17;
|
Chris@10
|
304 T2A = T2y - T2z;
|
Chris@10
|
305 T2I = T2y + T2z;
|
Chris@10
|
306 TM = TK + TL;
|
Chris@10
|
307 T1H = TK - TL;
|
Chris@10
|
308 }
|
Chris@10
|
309 T4R = T4P - T4Q;
|
Chris@10
|
310 T4X = T4P + T4Q;
|
Chris@10
|
311 T4W = T4M + T4N;
|
Chris@10
|
312 T4O = T4M - T4N;
|
Chris@10
|
313 T1G = TN - TO;
|
Chris@10
|
314 TP = TN + TO;
|
Chris@10
|
315 }
|
Chris@10
|
316 {
|
Chris@10
|
317 E Tm, T3X, TB, T3W;
|
Chris@10
|
318 Tm = Te + Tl;
|
Chris@10
|
319 T2O = Te - Tl;
|
Chris@10
|
320 T3I = T3E + T3H;
|
Chris@10
|
321 T3X = T3E - T3H;
|
Chris@10
|
322 TB = Tt + TA;
|
Chris@10
|
323 T2P = Tt - TA;
|
Chris@10
|
324 T3P = T3L + T3O;
|
Chris@10
|
325 T3W = T3L - T3O;
|
Chris@10
|
326 T2K = T2I + T2J;
|
Chris@10
|
327 T2M = T2I - T2J;
|
Chris@10
|
328 T1C = T1A + T1B;
|
Chris@10
|
329 T1E = T1A - T1B;
|
Chris@10
|
330 TC = Tm + TB;
|
Chris@10
|
331 T2w = Tm - TB;
|
Chris@10
|
332 T40 = T3W - T3X;
|
Chris@10
|
333 T3Y = T3W + T3X;
|
Chris@10
|
334 T4K = T4E - T4H;
|
Chris@10
|
335 T4I = T4E + T4H;
|
Chris@10
|
336 TS = TM - TP;
|
Chris@10
|
337 TQ = TM + TP;
|
Chris@10
|
338 }
|
Chris@10
|
339 }
|
Chris@10
|
340 }
|
Chris@10
|
341 }
|
Chris@10
|
342 {
|
Chris@10
|
343 E T3A, T3y, T50, T1D, T2t, T2p, T4J, T5t, T5v, T4Z, T4Y;
|
Chris@10
|
344 Rp[0] = T7 + TC;
|
Chris@10
|
345 T3A = T3q - T3x;
|
Chris@10
|
346 T3y = T3q + T3x;
|
Chris@10
|
347 T50 = T4W - T4X;
|
Chris@10
|
348 T4Y = T4W + T4X;
|
Chris@10
|
349 Rm[0] = T2H + T2K;
|
Chris@10
|
350 T1D = FNMS(KP250000000, T1C, T1z);
|
Chris@10
|
351 T2t = T1z + T1C;
|
Chris@10
|
352 T2p = TJ + TQ;
|
Chris@10
|
353 TR = FNMS(KP250000000, TQ, TJ);
|
Chris@10
|
354 T4J = FNMS(KP250000000, T4I, T4B);
|
Chris@10
|
355 T5t = T4B + T4I;
|
Chris@10
|
356 T5v = T4V + T4Y;
|
Chris@10
|
357 T4Z = FNMS(KP250000000, T4Y, T4V);
|
Chris@10
|
358 {
|
Chris@10
|
359 E T4m, T44, T4i, T4p, T49, T3R, T4j, T4a, T3S, T4l, T41, T4q;
|
Chris@10
|
360 {
|
Chris@10
|
361 E T3z, T4v, T4w, T3Z, T4z;
|
Chris@10
|
362 T3z = FNMS(KP250000000, T3y, T3j);
|
Chris@10
|
363 T4v = T3j + T3y;
|
Chris@10
|
364 {
|
Chris@10
|
365 E T2u, T2q, T5u, T5w;
|
Chris@10
|
366 T2u = T2s * T2p;
|
Chris@10
|
367 T2q = T2o * T2p;
|
Chris@10
|
368 T5u = T2c * T5t;
|
Chris@10
|
369 T5w = T2c * T5v;
|
Chris@10
|
370 Rm[WS(rs, 5)] = FMA(T2o, T2t, T2u);
|
Chris@10
|
371 Rp[WS(rs, 5)] = FNMS(T2s, T2t, T2q);
|
Chris@10
|
372 Ip[WS(rs, 2)] = FNMS(T2f, T5v, T5u);
|
Chris@10
|
373 Im[WS(rs, 2)] = FMA(T2f, T5t, T5w);
|
Chris@10
|
374 T4w = T4u * T4v;
|
Chris@10
|
375 }
|
Chris@10
|
376 T3Z = FNMS(KP250000000, T3Y, T3V);
|
Chris@10
|
377 T4z = T3V + T3Y;
|
Chris@10
|
378 {
|
Chris@10
|
379 E T3Q, T4h, T4A, T4g, T3B;
|
Chris@10
|
380 T3Q = FNMS(KP618033988, T3P, T3I);
|
Chris@10
|
381 T4h = FMA(KP618033988, T3I, T3P);
|
Chris@10
|
382 Ip[WS(rs, 7)] = FNMS(T4y, T4z, T4w);
|
Chris@10
|
383 T4A = T4u * T4z;
|
Chris@10
|
384 T4m = FMA(KP618033988, T42, T43);
|
Chris@10
|
385 T44 = FNMS(KP618033988, T43, T42);
|
Chris@10
|
386 T4g = FMA(KP559016994, T3A, T3z);
|
Chris@10
|
387 T3B = FNMS(KP559016994, T3A, T3z);
|
Chris@10
|
388 Im[WS(rs, 7)] = FMA(T4y, T4v, T4A);
|
Chris@10
|
389 T4i = FNMS(KP951056516, T4h, T4g);
|
Chris@10
|
390 T4p = FMA(KP951056516, T4h, T4g);
|
Chris@10
|
391 T49 = FMA(KP951056516, T3Q, T3B);
|
Chris@10
|
392 T3R = FNMS(KP951056516, T3Q, T3B);
|
Chris@10
|
393 }
|
Chris@10
|
394 T4j = T4f * T4i;
|
Chris@10
|
395 T4a = T48 * T49;
|
Chris@10
|
396 T3S = TE * T3R;
|
Chris@10
|
397 T4l = FMA(KP559016994, T40, T3Z);
|
Chris@10
|
398 T41 = FNMS(KP559016994, T40, T3Z);
|
Chris@10
|
399 T4q = T1L * T4p;
|
Chris@10
|
400 }
|
Chris@10
|
401 {
|
Chris@10
|
402 E T5d, T4S, T54, T5i, T4L, T5c;
|
Chris@10
|
403 T5d = FNMS(KP618033988, T4O, T4R);
|
Chris@10
|
404 T4S = FMA(KP618033988, T4R, T4O);
|
Chris@10
|
405 {
|
Chris@10
|
406 E T4n, T4r, T4d, T45;
|
Chris@10
|
407 T4n = FMA(KP951056516, T4m, T4l);
|
Chris@10
|
408 T4r = FNMS(KP951056516, T4m, T4l);
|
Chris@10
|
409 T4d = FNMS(KP951056516, T44, T41);
|
Chris@10
|
410 T45 = FMA(KP951056516, T44, T41);
|
Chris@10
|
411 {
|
Chris@10
|
412 E T4o, T4s, T4e, T46;
|
Chris@10
|
413 T4o = T4f * T4n;
|
Chris@10
|
414 Ip[WS(rs, 5)] = FNMS(T4k, T4n, T4j);
|
Chris@10
|
415 T4s = T1L * T4r;
|
Chris@10
|
416 Ip[WS(rs, 9)] = FNMS(T1N, T4r, T4q);
|
Chris@10
|
417 T4e = T48 * T4d;
|
Chris@10
|
418 Ip[WS(rs, 3)] = FNMS(T4c, T4d, T4a);
|
Chris@10
|
419 T46 = TE * T45;
|
Chris@10
|
420 Ip[WS(rs, 1)] = FNMS(TH, T45, T3S);
|
Chris@10
|
421 Im[WS(rs, 5)] = FMA(T4k, T4i, T4o);
|
Chris@10
|
422 Im[WS(rs, 9)] = FMA(T1N, T4p, T4s);
|
Chris@10
|
423 Im[WS(rs, 3)] = FMA(T4c, T49, T4e);
|
Chris@10
|
424 Im[WS(rs, 1)] = FMA(TH, T3R, T46);
|
Chris@10
|
425 }
|
Chris@10
|
426 }
|
Chris@10
|
427 T54 = FMA(KP618033988, T53, T52);
|
Chris@10
|
428 T5i = FNMS(KP618033988, T52, T53);
|
Chris@10
|
429 T4L = FMA(KP559016994, T4K, T4J);
|
Chris@10
|
430 T5c = FNMS(KP559016994, T4K, T4J);
|
Chris@10
|
431 {
|
Chris@10
|
432 E T38, T2Q, T33, T2E, T2v, T37, T2N, T5h, T51, T2L, T2x, T32;
|
Chris@10
|
433 T38 = FNMS(KP618033988, T2O, T2P);
|
Chris@10
|
434 T2Q = FMA(KP618033988, T2P, T2O);
|
Chris@10
|
435 T5h = FNMS(KP559016994, T50, T4Z);
|
Chris@10
|
436 T51 = FMA(KP559016994, T50, T4Z);
|
Chris@10
|
437 {
|
Chris@10
|
438 E T5e, T5n, T57, T4T;
|
Chris@10
|
439 T5e = FNMS(KP951056516, T5d, T5c);
|
Chris@10
|
440 T5n = FMA(KP951056516, T5d, T5c);
|
Chris@10
|
441 T57 = FMA(KP951056516, T4S, T4L);
|
Chris@10
|
442 T4T = FNMS(KP951056516, T4S, T4L);
|
Chris@10
|
443 {
|
Chris@10
|
444 E T5j, T5r, T59, T55;
|
Chris@10
|
445 T5j = FMA(KP951056516, T5i, T5h);
|
Chris@10
|
446 T5r = FNMS(KP951056516, T5i, T5h);
|
Chris@10
|
447 T59 = FNMS(KP951056516, T54, T51);
|
Chris@10
|
448 T55 = FMA(KP951056516, T54, T51);
|
Chris@10
|
449 {
|
Chris@10
|
450 E T5f, T5o, T58, T4U;
|
Chris@10
|
451 T5f = T5b * T5e;
|
Chris@10
|
452 T5o = T5m * T5n;
|
Chris@10
|
453 T58 = T1V * T57;
|
Chris@10
|
454 T4U = TD * T4T;
|
Chris@10
|
455 {
|
Chris@10
|
456 E T5k, T5s, T5a, T56;
|
Chris@10
|
457 T5k = T5b * T5j;
|
Chris@10
|
458 T5s = T5m * T5r;
|
Chris@10
|
459 T5a = T1V * T59;
|
Chris@10
|
460 T56 = TD * T55;
|
Chris@10
|
461 Ip[WS(rs, 6)] = FNMS(T5g, T5j, T5f);
|
Chris@10
|
462 Ip[WS(rs, 8)] = FNMS(T5q, T5r, T5o);
|
Chris@10
|
463 Ip[WS(rs, 4)] = FNMS(T1X, T59, T58);
|
Chris@10
|
464 Ip[0] = FNMS(TG, T55, T4U);
|
Chris@10
|
465 Im[WS(rs, 6)] = FMA(T5g, T5e, T5k);
|
Chris@10
|
466 Im[WS(rs, 8)] = FMA(T5q, T5n, T5s);
|
Chris@10
|
467 Im[WS(rs, 4)] = FMA(T1X, T57, T5a);
|
Chris@10
|
468 Im[0] = FMA(TG, T4T, T56);
|
Chris@10
|
469 }
|
Chris@10
|
470 }
|
Chris@10
|
471 }
|
Chris@10
|
472 }
|
Chris@10
|
473 T2L = FNMS(KP250000000, T2K, T2H);
|
Chris@10
|
474 T33 = FNMS(KP618033988, T2A, T2D);
|
Chris@10
|
475 T2E = FMA(KP618033988, T2D, T2A);
|
Chris@10
|
476 T2v = FNMS(KP250000000, TC, T7);
|
Chris@10
|
477 T37 = FNMS(KP559016994, T2M, T2L);
|
Chris@10
|
478 T2N = FMA(KP559016994, T2M, T2L);
|
Chris@10
|
479 T1I = FNMS(KP618033988, T1H, T1G);
|
Chris@10
|
480 T26 = FMA(KP618033988, T1G, T1H);
|
Chris@10
|
481 T2x = FMA(KP559016994, T2w, T2v);
|
Chris@10
|
482 T32 = FNMS(KP559016994, T2w, T2v);
|
Chris@10
|
483 {
|
Chris@10
|
484 E T3f, T39, T2R, T2Z;
|
Chris@10
|
485 T3f = FNMS(KP951056516, T38, T37);
|
Chris@10
|
486 T39 = FMA(KP951056516, T38, T37);
|
Chris@10
|
487 T2R = FNMS(KP951056516, T2Q, T2N);
|
Chris@10
|
488 T2Z = FMA(KP951056516, T2Q, T2N);
|
Chris@10
|
489 {
|
Chris@10
|
490 E T3c, T34, T2F, T2V;
|
Chris@10
|
491 T3c = FMA(KP951056516, T33, T32);
|
Chris@10
|
492 T34 = FNMS(KP951056516, T33, T32);
|
Chris@10
|
493 T2F = FMA(KP951056516, T2E, T2x);
|
Chris@10
|
494 T2V = FNMS(KP951056516, T2E, T2x);
|
Chris@10
|
495 {
|
Chris@10
|
496 E T3a, T35, T3g, T3d;
|
Chris@10
|
497 T3a = T36 * T34;
|
Chris@10
|
498 T35 = T31 * T34;
|
Chris@10
|
499 T3g = T3e * T3c;
|
Chris@10
|
500 T3d = T3b * T3c;
|
Chris@10
|
501 {
|
Chris@10
|
502 E T30, T2W, T2S, T2G;
|
Chris@10
|
503 T30 = T2Y * T2V;
|
Chris@10
|
504 T2W = T2U * T2V;
|
Chris@10
|
505 T2S = T2b * T2F;
|
Chris@10
|
506 T2G = T29 * T2F;
|
Chris@10
|
507 Rm[WS(rs, 4)] = FMA(T31, T39, T3a);
|
Chris@10
|
508 Rp[WS(rs, 4)] = FNMS(T36, T39, T35);
|
Chris@10
|
509 Rm[WS(rs, 6)] = FMA(T3b, T3f, T3g);
|
Chris@10
|
510 Rp[WS(rs, 6)] = FNMS(T3e, T3f, T3d);
|
Chris@10
|
511 Rm[WS(rs, 8)] = FMA(T2U, T2Z, T30);
|
Chris@10
|
512 Rp[WS(rs, 8)] = FNMS(T2Y, T2Z, T2W);
|
Chris@10
|
513 Rm[WS(rs, 2)] = FMA(T29, T2R, T2S);
|
Chris@10
|
514 Rp[WS(rs, 2)] = FNMS(T2b, T2R, T2G);
|
Chris@10
|
515 }
|
Chris@10
|
516 }
|
Chris@10
|
517 }
|
Chris@10
|
518 }
|
Chris@10
|
519 T1o = FNMS(KP618033988, T1n, T18);
|
Chris@10
|
520 T20 = FMA(KP618033988, T18, T1n);
|
Chris@10
|
521 T1F = FNMS(KP559016994, T1E, T1D);
|
Chris@10
|
522 T25 = FMA(KP559016994, T1E, T1D);
|
Chris@10
|
523 }
|
Chris@10
|
524 }
|
Chris@10
|
525 }
|
Chris@10
|
526 }
|
Chris@10
|
527 }
|
Chris@10
|
528 }
|
Chris@10
|
529 TT = FNMS(KP559016994, TS, TR);
|
Chris@10
|
530 T1Z = FMA(KP559016994, TS, TR);
|
Chris@10
|
531 {
|
Chris@10
|
532 E T2l, T27, T1J, T1T;
|
Chris@10
|
533 T2l = FNMS(KP951056516, T26, T25);
|
Chris@10
|
534 T27 = FMA(KP951056516, T26, T25);
|
Chris@10
|
535 T1J = FNMS(KP951056516, T1I, T1F);
|
Chris@10
|
536 T1T = FMA(KP951056516, T1I, T1F);
|
Chris@10
|
537 {
|
Chris@10
|
538 E T2h, T21, T1p, T1P;
|
Chris@10
|
539 T2h = FMA(KP951056516, T20, T1Z);
|
Chris@10
|
540 T21 = FNMS(KP951056516, T20, T1Z);
|
Chris@10
|
541 T1p = FMA(KP951056516, T1o, TT);
|
Chris@10
|
542 T1P = FNMS(KP951056516, T1o, TT);
|
Chris@10
|
543 {
|
Chris@10
|
544 E T28, T22, T2m, T2i;
|
Chris@10
|
545 T28 = T24 * T21;
|
Chris@10
|
546 T22 = T1Y * T21;
|
Chris@10
|
547 T2m = T2k * T2h;
|
Chris@10
|
548 T2i = T2g * T2h;
|
Chris@10
|
549 {
|
Chris@10
|
550 E T1U, T1Q, T1K, T1q;
|
Chris@10
|
551 T1U = T1S * T1P;
|
Chris@10
|
552 T1Q = T1O * T1P;
|
Chris@10
|
553 T1K = T1s * T1p;
|
Chris@10
|
554 T1q = TI * T1p;
|
Chris@10
|
555 Rm[WS(rs, 3)] = FMA(T1Y, T27, T28);
|
Chris@10
|
556 Rp[WS(rs, 3)] = FNMS(T24, T27, T22);
|
Chris@10
|
557 Rm[WS(rs, 7)] = FMA(T2g, T2l, T2m);
|
Chris@10
|
558 Rp[WS(rs, 7)] = FNMS(T2k, T2l, T2i);
|
Chris@10
|
559 Rm[WS(rs, 9)] = FMA(T1O, T1T, T1U);
|
Chris@10
|
560 Rp[WS(rs, 9)] = FNMS(T1S, T1T, T1Q);
|
Chris@10
|
561 Rm[WS(rs, 1)] = FMA(TI, T1J, T1K);
|
Chris@10
|
562 Rp[WS(rs, 1)] = FNMS(T1s, T1J, T1q);
|
Chris@10
|
563 }
|
Chris@10
|
564 }
|
Chris@10
|
565 }
|
Chris@10
|
566 }
|
Chris@10
|
567 }
|
Chris@10
|
568 }
|
Chris@10
|
569 }
|
Chris@10
|
570
|
Chris@10
|
571 static const tw_instr twinstr[] = {
|
Chris@10
|
572 {TW_CEXP, 1, 1},
|
Chris@10
|
573 {TW_CEXP, 1, 3},
|
Chris@10
|
574 {TW_CEXP, 1, 9},
|
Chris@10
|
575 {TW_CEXP, 1, 19},
|
Chris@10
|
576 {TW_NEXT, 1, 0}
|
Chris@10
|
577 };
|
Chris@10
|
578
|
Chris@10
|
579 static const hc2c_desc desc = { 20, "hc2cb2_20", twinstr, &GENUS, {136, 58, 140, 0} };
|
Chris@10
|
580
|
Chris@10
|
581 void X(codelet_hc2cb2_20) (planner *p) {
|
Chris@10
|
582 X(khc2c_register) (p, hc2cb2_20, &desc, HC2C_VIA_RDFT);
|
Chris@10
|
583 }
|
Chris@10
|
584 #else /* HAVE_FMA */
|
Chris@10
|
585
|
Chris@10
|
586 /* Generated by: ../../../genfft/gen_hc2c.native -compact -variables 4 -pipeline-latency 4 -sign 1 -twiddle-log3 -precompute-twiddles -n 20 -dif -name hc2cb2_20 -include hc2cb.h */
|
Chris@10
|
587
|
Chris@10
|
588 /*
|
Chris@10
|
589 * This function contains 276 FP additions, 164 FP multiplications,
|
Chris@10
|
590 * (or, 204 additions, 92 multiplications, 72 fused multiply/add),
|
Chris@10
|
591 * 137 stack variables, 4 constants, and 80 memory accesses
|
Chris@10
|
592 */
|
Chris@10
|
593 #include "hc2cb.h"
|
Chris@10
|
594
|
Chris@10
|
595 static void hc2cb2_20(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms)
|
Chris@10
|
596 {
|
Chris@10
|
597 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
|
Chris@10
|
598 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
|
Chris@10
|
599 DK(KP587785252, +0.587785252292473129168705954639072768597652438);
|
Chris@10
|
600 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
|
Chris@10
|
601 {
|
Chris@10
|
602 INT m;
|
Chris@10
|
603 for (m = mb, W = W + ((mb - 1) * 8); m < me; m = m + 1, Rp = Rp + ms, Ip = Ip + ms, Rm = Rm - ms, Im = Im - ms, W = W + 8, MAKE_VOLATILE_STRIDE(80, rs)) {
|
Chris@10
|
604 E TD, TG, TE, TH, TJ, T1t, T27, T25, T1T, T1R, T1V, T2j, T2Z, T21, T2X;
|
Chris@10
|
605 E T2T, T2n, T2P, T3V, T41, T3R, T3X, T29, T2c, T4H, T4L, T1L, T1M, T1N, T2d;
|
Chris@10
|
606 E T4R, T1P, T4P, T49, T2N, T2f, T47, T2L;
|
Chris@10
|
607 {
|
Chris@10
|
608 E T1U, T2l, T1Z, T2i, T1S, T2m, T20, T2h;
|
Chris@10
|
609 {
|
Chris@10
|
610 E TF, T1s, TI, T1r;
|
Chris@10
|
611 TD = W[0];
|
Chris@10
|
612 TG = W[1];
|
Chris@10
|
613 TE = W[2];
|
Chris@10
|
614 TH = W[3];
|
Chris@10
|
615 TF = TD * TE;
|
Chris@10
|
616 T1s = TG * TE;
|
Chris@10
|
617 TI = TG * TH;
|
Chris@10
|
618 T1r = TD * TH;
|
Chris@10
|
619 TJ = TF + TI;
|
Chris@10
|
620 T1t = T1r - T1s;
|
Chris@10
|
621 T27 = T1r + T1s;
|
Chris@10
|
622 T25 = TF - TI;
|
Chris@10
|
623 T1T = W[5];
|
Chris@10
|
624 T1U = TH * T1T;
|
Chris@10
|
625 T2l = TD * T1T;
|
Chris@10
|
626 T1Z = TE * T1T;
|
Chris@10
|
627 T2i = TG * T1T;
|
Chris@10
|
628 T1R = W[4];
|
Chris@10
|
629 T1S = TE * T1R;
|
Chris@10
|
630 T2m = TG * T1R;
|
Chris@10
|
631 T20 = TH * T1R;
|
Chris@10
|
632 T2h = TD * T1R;
|
Chris@10
|
633 }
|
Chris@10
|
634 T1V = T1S + T1U;
|
Chris@10
|
635 T2j = T2h - T2i;
|
Chris@10
|
636 T2Z = T1Z + T20;
|
Chris@10
|
637 T21 = T1Z - T20;
|
Chris@10
|
638 T2X = T1S - T1U;
|
Chris@10
|
639 T2T = T2l - T2m;
|
Chris@10
|
640 T2n = T2l + T2m;
|
Chris@10
|
641 T2P = T2h + T2i;
|
Chris@10
|
642 {
|
Chris@10
|
643 E T3T, T3U, T3P, T3Q;
|
Chris@10
|
644 T3T = TJ * T1T;
|
Chris@10
|
645 T3U = T1t * T1R;
|
Chris@10
|
646 T3V = T3T - T3U;
|
Chris@10
|
647 T41 = T3T + T3U;
|
Chris@10
|
648 T3P = TJ * T1R;
|
Chris@10
|
649 T3Q = T1t * T1T;
|
Chris@10
|
650 T3R = T3P + T3Q;
|
Chris@10
|
651 T3X = T3P - T3Q;
|
Chris@10
|
652 {
|
Chris@10
|
653 E T26, T28, T2a, T2b;
|
Chris@10
|
654 T26 = T25 * T1R;
|
Chris@10
|
655 T28 = T27 * T1T;
|
Chris@10
|
656 T29 = T26 + T28;
|
Chris@10
|
657 T2a = T25 * T1T;
|
Chris@10
|
658 T2b = T27 * T1R;
|
Chris@10
|
659 T2c = T2a - T2b;
|
Chris@10
|
660 T4H = T26 - T28;
|
Chris@10
|
661 T4L = T2a + T2b;
|
Chris@10
|
662 T1L = W[6];
|
Chris@10
|
663 T1M = W[7];
|
Chris@10
|
664 T1N = FMA(TD, T1L, TG * T1M);
|
Chris@10
|
665 T2d = FMA(T29, T1L, T2c * T1M);
|
Chris@10
|
666 T4R = FNMS(T1t, T1L, TJ * T1M);
|
Chris@10
|
667 T1P = FNMS(TG, T1L, TD * T1M);
|
Chris@10
|
668 T4P = FMA(TJ, T1L, T1t * T1M);
|
Chris@10
|
669 T49 = FNMS(T27, T1L, T25 * T1M);
|
Chris@10
|
670 T2N = FNMS(TH, T1L, TE * T1M);
|
Chris@10
|
671 T2f = FNMS(T2c, T1L, T29 * T1M);
|
Chris@10
|
672 T47 = FMA(T25, T1L, T27 * T1M);
|
Chris@10
|
673 T2L = FMA(TE, T1L, TH * T1M);
|
Chris@10
|
674 }
|
Chris@10
|
675 }
|
Chris@10
|
676 }
|
Chris@10
|
677 {
|
Chris@10
|
678 E T7, T4i, T4x, TK, T1D, T3i, T3E, T2D, T19, T3L, T3M, T1o, T2x, T4C, T4B;
|
Chris@10
|
679 E T2u, T1v, T4r, T4o, T1u, T2H, T37, T2I, T3e, T3p, T3w, T3x, Tm, TB, TC;
|
Chris@10
|
680 E T4u, T4v, T4y, T2A, T2B, T2E, T1E, T1F, T1G, T4d, T4g, T4j, T3F, T3G, T3H;
|
Chris@10
|
681 E TN, TQ, TR, T48, T4a;
|
Chris@10
|
682 {
|
Chris@10
|
683 E T3, T3g, T1z, T3C, T6, T3D, T1C, T3h;
|
Chris@10
|
684 {
|
Chris@10
|
685 E T1, T2, T1x, T1y;
|
Chris@10
|
686 T1 = Rp[0];
|
Chris@10
|
687 T2 = Rm[WS(rs, 9)];
|
Chris@10
|
688 T3 = T1 + T2;
|
Chris@10
|
689 T3g = T1 - T2;
|
Chris@10
|
690 T1x = Ip[0];
|
Chris@10
|
691 T1y = Im[WS(rs, 9)];
|
Chris@10
|
692 T1z = T1x - T1y;
|
Chris@10
|
693 T3C = T1x + T1y;
|
Chris@10
|
694 }
|
Chris@10
|
695 {
|
Chris@10
|
696 E T4, T5, T1A, T1B;
|
Chris@10
|
697 T4 = Rp[WS(rs, 5)];
|
Chris@10
|
698 T5 = Rm[WS(rs, 4)];
|
Chris@10
|
699 T6 = T4 + T5;
|
Chris@10
|
700 T3D = T4 - T5;
|
Chris@10
|
701 T1A = Ip[WS(rs, 5)];
|
Chris@10
|
702 T1B = Im[WS(rs, 4)];
|
Chris@10
|
703 T1C = T1A - T1B;
|
Chris@10
|
704 T3h = T1A + T1B;
|
Chris@10
|
705 }
|
Chris@10
|
706 T7 = T3 + T6;
|
Chris@10
|
707 T4i = T3g - T3h;
|
Chris@10
|
708 T4x = T3D + T3C;
|
Chris@10
|
709 TK = T3 - T6;
|
Chris@10
|
710 T1D = T1z - T1C;
|
Chris@10
|
711 T3i = T3g + T3h;
|
Chris@10
|
712 T3E = T3C - T3D;
|
Chris@10
|
713 T2D = T1z + T1C;
|
Chris@10
|
714 }
|
Chris@10
|
715 {
|
Chris@10
|
716 E Te, T4b, T4m, TL, T11, T33, T3l, T2s, TA, T4f, T4q, TP, T1n, T3d, T3v;
|
Chris@10
|
717 E T2w, Tl, T4c, T4n, TM, T18, T36, T3o, T2t, Tt, T4e, T4p, TO, T1g, T3a;
|
Chris@10
|
718 E T3s, T2v;
|
Chris@10
|
719 {
|
Chris@10
|
720 E Ta, T3j, TX, T31, Td, T32, T10, T3k;
|
Chris@10
|
721 {
|
Chris@10
|
722 E T8, T9, TV, TW;
|
Chris@10
|
723 T8 = Rp[WS(rs, 4)];
|
Chris@10
|
724 T9 = Rm[WS(rs, 5)];
|
Chris@10
|
725 Ta = T8 + T9;
|
Chris@10
|
726 T3j = T8 - T9;
|
Chris@10
|
727 TV = Ip[WS(rs, 4)];
|
Chris@10
|
728 TW = Im[WS(rs, 5)];
|
Chris@10
|
729 TX = TV - TW;
|
Chris@10
|
730 T31 = TV + TW;
|
Chris@10
|
731 }
|
Chris@10
|
732 {
|
Chris@10
|
733 E Tb, Tc, TY, TZ;
|
Chris@10
|
734 Tb = Rp[WS(rs, 9)];
|
Chris@10
|
735 Tc = Rm[0];
|
Chris@10
|
736 Td = Tb + Tc;
|
Chris@10
|
737 T32 = Tb - Tc;
|
Chris@10
|
738 TY = Ip[WS(rs, 9)];
|
Chris@10
|
739 TZ = Im[0];
|
Chris@10
|
740 T10 = TY - TZ;
|
Chris@10
|
741 T3k = TY + TZ;
|
Chris@10
|
742 }
|
Chris@10
|
743 Te = Ta + Td;
|
Chris@10
|
744 T4b = T3j - T3k;
|
Chris@10
|
745 T4m = T32 + T31;
|
Chris@10
|
746 TL = Ta - Td;
|
Chris@10
|
747 T11 = TX - T10;
|
Chris@10
|
748 T33 = T31 - T32;
|
Chris@10
|
749 T3l = T3j + T3k;
|
Chris@10
|
750 T2s = TX + T10;
|
Chris@10
|
751 }
|
Chris@10
|
752 {
|
Chris@10
|
753 E Tw, T3t, T1j, T3c, Tz, T3b, T1m, T3u;
|
Chris@10
|
754 {
|
Chris@10
|
755 E Tu, Tv, T1h, T1i;
|
Chris@10
|
756 Tu = Rm[WS(rs, 7)];
|
Chris@10
|
757 Tv = Rp[WS(rs, 2)];
|
Chris@10
|
758 Tw = Tu + Tv;
|
Chris@10
|
759 T3t = Tu - Tv;
|
Chris@10
|
760 T1h = Ip[WS(rs, 2)];
|
Chris@10
|
761 T1i = Im[WS(rs, 7)];
|
Chris@10
|
762 T1j = T1h - T1i;
|
Chris@10
|
763 T3c = T1h + T1i;
|
Chris@10
|
764 }
|
Chris@10
|
765 {
|
Chris@10
|
766 E Tx, Ty, T1k, T1l;
|
Chris@10
|
767 Tx = Rm[WS(rs, 2)];
|
Chris@10
|
768 Ty = Rp[WS(rs, 7)];
|
Chris@10
|
769 Tz = Tx + Ty;
|
Chris@10
|
770 T3b = Tx - Ty;
|
Chris@10
|
771 T1k = Ip[WS(rs, 7)];
|
Chris@10
|
772 T1l = Im[WS(rs, 2)];
|
Chris@10
|
773 T1m = T1k - T1l;
|
Chris@10
|
774 T3u = T1k + T1l;
|
Chris@10
|
775 }
|
Chris@10
|
776 TA = Tw + Tz;
|
Chris@10
|
777 T4f = T3t + T3u;
|
Chris@10
|
778 T4q = T3b - T3c;
|
Chris@10
|
779 TP = Tw - Tz;
|
Chris@10
|
780 T1n = T1j - T1m;
|
Chris@10
|
781 T3d = T3b + T3c;
|
Chris@10
|
782 T3v = T3t - T3u;
|
Chris@10
|
783 T2w = T1j + T1m;
|
Chris@10
|
784 }
|
Chris@10
|
785 {
|
Chris@10
|
786 E Th, T3m, T14, T35, Tk, T34, T17, T3n;
|
Chris@10
|
787 {
|
Chris@10
|
788 E Tf, Tg, T12, T13;
|
Chris@10
|
789 Tf = Rm[WS(rs, 3)];
|
Chris@10
|
790 Tg = Rp[WS(rs, 6)];
|
Chris@10
|
791 Th = Tf + Tg;
|
Chris@10
|
792 T3m = Tf - Tg;
|
Chris@10
|
793 T12 = Ip[WS(rs, 6)];
|
Chris@10
|
794 T13 = Im[WS(rs, 3)];
|
Chris@10
|
795 T14 = T12 - T13;
|
Chris@10
|
796 T35 = T12 + T13;
|
Chris@10
|
797 }
|
Chris@10
|
798 {
|
Chris@10
|
799 E Ti, Tj, T15, T16;
|
Chris@10
|
800 Ti = Rp[WS(rs, 1)];
|
Chris@10
|
801 Tj = Rm[WS(rs, 8)];
|
Chris@10
|
802 Tk = Ti + Tj;
|
Chris@10
|
803 T34 = Ti - Tj;
|
Chris@10
|
804 T15 = Ip[WS(rs, 1)];
|
Chris@10
|
805 T16 = Im[WS(rs, 8)];
|
Chris@10
|
806 T17 = T15 - T16;
|
Chris@10
|
807 T3n = T15 + T16;
|
Chris@10
|
808 }
|
Chris@10
|
809 Tl = Th + Tk;
|
Chris@10
|
810 T4c = T3m - T3n;
|
Chris@10
|
811 T4n = T34 - T35;
|
Chris@10
|
812 TM = Th - Tk;
|
Chris@10
|
813 T18 = T14 - T17;
|
Chris@10
|
814 T36 = T34 + T35;
|
Chris@10
|
815 T3o = T3m + T3n;
|
Chris@10
|
816 T2t = T14 + T17;
|
Chris@10
|
817 }
|
Chris@10
|
818 {
|
Chris@10
|
819 E Tp, T3q, T1c, T38, Ts, T39, T1f, T3r;
|
Chris@10
|
820 {
|
Chris@10
|
821 E Tn, To, T1a, T1b;
|
Chris@10
|
822 Tn = Rp[WS(rs, 8)];
|
Chris@10
|
823 To = Rm[WS(rs, 1)];
|
Chris@10
|
824 Tp = Tn + To;
|
Chris@10
|
825 T3q = Tn - To;
|
Chris@10
|
826 T1a = Ip[WS(rs, 8)];
|
Chris@10
|
827 T1b = Im[WS(rs, 1)];
|
Chris@10
|
828 T1c = T1a - T1b;
|
Chris@10
|
829 T38 = T1a + T1b;
|
Chris@10
|
830 }
|
Chris@10
|
831 {
|
Chris@10
|
832 E Tq, Tr, T1d, T1e;
|
Chris@10
|
833 Tq = Rm[WS(rs, 6)];
|
Chris@10
|
834 Tr = Rp[WS(rs, 3)];
|
Chris@10
|
835 Ts = Tq + Tr;
|
Chris@10
|
836 T39 = Tq - Tr;
|
Chris@10
|
837 T1d = Ip[WS(rs, 3)];
|
Chris@10
|
838 T1e = Im[WS(rs, 6)];
|
Chris@10
|
839 T1f = T1d - T1e;
|
Chris@10
|
840 T3r = T1d + T1e;
|
Chris@10
|
841 }
|
Chris@10
|
842 Tt = Tp + Ts;
|
Chris@10
|
843 T4e = T3q + T3r;
|
Chris@10
|
844 T4p = T39 + T38;
|
Chris@10
|
845 TO = Tp - Ts;
|
Chris@10
|
846 T1g = T1c - T1f;
|
Chris@10
|
847 T3a = T38 - T39;
|
Chris@10
|
848 T3s = T3q - T3r;
|
Chris@10
|
849 T2v = T1c + T1f;
|
Chris@10
|
850 }
|
Chris@10
|
851 T19 = T11 - T18;
|
Chris@10
|
852 T3L = T3l - T3o;
|
Chris@10
|
853 T3M = T3s - T3v;
|
Chris@10
|
854 T1o = T1g - T1n;
|
Chris@10
|
855 T2x = T2v - T2w;
|
Chris@10
|
856 T4C = T4e - T4f;
|
Chris@10
|
857 T4B = T4b - T4c;
|
Chris@10
|
858 T2u = T2s - T2t;
|
Chris@10
|
859 T1v = TO - TP;
|
Chris@10
|
860 T4r = T4p - T4q;
|
Chris@10
|
861 T4o = T4m - T4n;
|
Chris@10
|
862 T1u = TL - TM;
|
Chris@10
|
863 T2H = Te - Tl;
|
Chris@10
|
864 T37 = T33 + T36;
|
Chris@10
|
865 T2I = Tt - TA;
|
Chris@10
|
866 T3e = T3a + T3d;
|
Chris@10
|
867 T3p = T3l + T3o;
|
Chris@10
|
868 T3w = T3s + T3v;
|
Chris@10
|
869 T3x = T3p + T3w;
|
Chris@10
|
870 Tm = Te + Tl;
|
Chris@10
|
871 TB = Tt + TA;
|
Chris@10
|
872 TC = Tm + TB;
|
Chris@10
|
873 T4u = T4m + T4n;
|
Chris@10
|
874 T4v = T4p + T4q;
|
Chris@10
|
875 T4y = T4u + T4v;
|
Chris@10
|
876 T2A = T2s + T2t;
|
Chris@10
|
877 T2B = T2v + T2w;
|
Chris@10
|
878 T2E = T2A + T2B;
|
Chris@10
|
879 T1E = T11 + T18;
|
Chris@10
|
880 T1F = T1g + T1n;
|
Chris@10
|
881 T1G = T1E + T1F;
|
Chris@10
|
882 T4d = T4b + T4c;
|
Chris@10
|
883 T4g = T4e + T4f;
|
Chris@10
|
884 T4j = T4d + T4g;
|
Chris@10
|
885 T3F = T33 - T36;
|
Chris@10
|
886 T3G = T3a - T3d;
|
Chris@10
|
887 T3H = T3F + T3G;
|
Chris@10
|
888 TN = TL + TM;
|
Chris@10
|
889 TQ = TO + TP;
|
Chris@10
|
890 TR = TN + TQ;
|
Chris@10
|
891 }
|
Chris@10
|
892 Rp[0] = T7 + TC;
|
Chris@10
|
893 Rm[0] = T2D + T2E;
|
Chris@10
|
894 {
|
Chris@10
|
895 E T2k, T2o, T4T, T4U;
|
Chris@10
|
896 T2k = TK + TR;
|
Chris@10
|
897 T2o = T1D + T1G;
|
Chris@10
|
898 Rp[WS(rs, 5)] = FNMS(T2n, T2o, T2j * T2k);
|
Chris@10
|
899 Rm[WS(rs, 5)] = FMA(T2n, T2k, T2j * T2o);
|
Chris@10
|
900 T4T = T4i + T4j;
|
Chris@10
|
901 T4U = T4x + T4y;
|
Chris@10
|
902 Ip[WS(rs, 2)] = FNMS(T2c, T4U, T29 * T4T);
|
Chris@10
|
903 Im[WS(rs, 2)] = FMA(T29, T4U, T2c * T4T);
|
Chris@10
|
904 }
|
Chris@10
|
905 T48 = T3i + T3x;
|
Chris@10
|
906 T4a = T3E + T3H;
|
Chris@10
|
907 Ip[WS(rs, 7)] = FNMS(T49, T4a, T47 * T48);
|
Chris@10
|
908 Im[WS(rs, 7)] = FMA(T47, T4a, T49 * T48);
|
Chris@10
|
909 {
|
Chris@10
|
910 E T2y, T2J, T2V, T2R, T2G, T2U, T2r, T2Q;
|
Chris@10
|
911 T2y = FMA(KP951056516, T2u, KP587785252 * T2x);
|
Chris@10
|
912 T2J = FMA(KP951056516, T2H, KP587785252 * T2I);
|
Chris@10
|
913 T2V = FNMS(KP951056516, T2I, KP587785252 * T2H);
|
Chris@10
|
914 T2R = FNMS(KP951056516, T2x, KP587785252 * T2u);
|
Chris@10
|
915 {
|
Chris@10
|
916 E T2C, T2F, T2p, T2q;
|
Chris@10
|
917 T2C = KP559016994 * (T2A - T2B);
|
Chris@10
|
918 T2F = FNMS(KP250000000, T2E, T2D);
|
Chris@10
|
919 T2G = T2C + T2F;
|
Chris@10
|
920 T2U = T2F - T2C;
|
Chris@10
|
921 T2p = KP559016994 * (Tm - TB);
|
Chris@10
|
922 T2q = FNMS(KP250000000, TC, T7);
|
Chris@10
|
923 T2r = T2p + T2q;
|
Chris@10
|
924 T2Q = T2q - T2p;
|
Chris@10
|
925 }
|
Chris@10
|
926 {
|
Chris@10
|
927 E T2z, T2K, T2Y, T30;
|
Chris@10
|
928 T2z = T2r + T2y;
|
Chris@10
|
929 T2K = T2G - T2J;
|
Chris@10
|
930 Rp[WS(rs, 2)] = FNMS(T27, T2K, T25 * T2z);
|
Chris@10
|
931 Rm[WS(rs, 2)] = FMA(T27, T2z, T25 * T2K);
|
Chris@10
|
932 T2Y = T2Q - T2R;
|
Chris@10
|
933 T30 = T2V + T2U;
|
Chris@10
|
934 Rp[WS(rs, 6)] = FNMS(T2Z, T30, T2X * T2Y);
|
Chris@10
|
935 Rm[WS(rs, 6)] = FMA(T2Z, T2Y, T2X * T30);
|
Chris@10
|
936 }
|
Chris@10
|
937 {
|
Chris@10
|
938 E T2M, T2O, T2S, T2W;
|
Chris@10
|
939 T2M = T2r - T2y;
|
Chris@10
|
940 T2O = T2J + T2G;
|
Chris@10
|
941 Rp[WS(rs, 8)] = FNMS(T2N, T2O, T2L * T2M);
|
Chris@10
|
942 Rm[WS(rs, 8)] = FMA(T2N, T2M, T2L * T2O);
|
Chris@10
|
943 T2S = T2Q + T2R;
|
Chris@10
|
944 T2W = T2U - T2V;
|
Chris@10
|
945 Rp[WS(rs, 4)] = FNMS(T2T, T2W, T2P * T2S);
|
Chris@10
|
946 Rm[WS(rs, 4)] = FMA(T2T, T2S, T2P * T2W);
|
Chris@10
|
947 }
|
Chris@10
|
948 }
|
Chris@10
|
949 {
|
Chris@10
|
950 E T4s, T4D, T4N, T4I, T4A, T4M, T4l, T4J;
|
Chris@10
|
951 T4s = FMA(KP951056516, T4o, KP587785252 * T4r);
|
Chris@10
|
952 T4D = FMA(KP951056516, T4B, KP587785252 * T4C);
|
Chris@10
|
953 T4N = FNMS(KP951056516, T4C, KP587785252 * T4B);
|
Chris@10
|
954 T4I = FNMS(KP951056516, T4r, KP587785252 * T4o);
|
Chris@10
|
955 {
|
Chris@10
|
956 E T4w, T4z, T4h, T4k;
|
Chris@10
|
957 T4w = KP559016994 * (T4u - T4v);
|
Chris@10
|
958 T4z = FNMS(KP250000000, T4y, T4x);
|
Chris@10
|
959 T4A = T4w + T4z;
|
Chris@10
|
960 T4M = T4z - T4w;
|
Chris@10
|
961 T4h = KP559016994 * (T4d - T4g);
|
Chris@10
|
962 T4k = FNMS(KP250000000, T4j, T4i);
|
Chris@10
|
963 T4l = T4h + T4k;
|
Chris@10
|
964 T4J = T4k - T4h;
|
Chris@10
|
965 }
|
Chris@10
|
966 {
|
Chris@10
|
967 E T4t, T4E, T4Q, T4S;
|
Chris@10
|
968 T4t = T4l - T4s;
|
Chris@10
|
969 T4E = T4A + T4D;
|
Chris@10
|
970 Ip[0] = FNMS(TG, T4E, TD * T4t);
|
Chris@10
|
971 Im[0] = FMA(TD, T4E, TG * T4t);
|
Chris@10
|
972 T4Q = T4J - T4I;
|
Chris@10
|
973 T4S = T4M + T4N;
|
Chris@10
|
974 Ip[WS(rs, 8)] = FNMS(T4R, T4S, T4P * T4Q);
|
Chris@10
|
975 Im[WS(rs, 8)] = FMA(T4P, T4S, T4R * T4Q);
|
Chris@10
|
976 }
|
Chris@10
|
977 {
|
Chris@10
|
978 E T4F, T4G, T4K, T4O;
|
Chris@10
|
979 T4F = T4s + T4l;
|
Chris@10
|
980 T4G = T4A - T4D;
|
Chris@10
|
981 Ip[WS(rs, 4)] = FNMS(T1T, T4G, T1R * T4F);
|
Chris@10
|
982 Im[WS(rs, 4)] = FMA(T1R, T4G, T1T * T4F);
|
Chris@10
|
983 T4K = T4I + T4J;
|
Chris@10
|
984 T4O = T4M - T4N;
|
Chris@10
|
985 Ip[WS(rs, 6)] = FNMS(T4L, T4O, T4H * T4K);
|
Chris@10
|
986 Im[WS(rs, 6)] = FMA(T4H, T4O, T4L * T4K);
|
Chris@10
|
987 }
|
Chris@10
|
988 }
|
Chris@10
|
989 {
|
Chris@10
|
990 E T1p, T1w, T22, T1X, T1J, T23, TU, T1W;
|
Chris@10
|
991 T1p = FNMS(KP951056516, T1o, KP587785252 * T19);
|
Chris@10
|
992 T1w = FNMS(KP951056516, T1v, KP587785252 * T1u);
|
Chris@10
|
993 T22 = FMA(KP951056516, T1u, KP587785252 * T1v);
|
Chris@10
|
994 T1X = FMA(KP951056516, T19, KP587785252 * T1o);
|
Chris@10
|
995 {
|
Chris@10
|
996 E T1H, T1I, TS, TT;
|
Chris@10
|
997 T1H = FNMS(KP250000000, T1G, T1D);
|
Chris@10
|
998 T1I = KP559016994 * (T1E - T1F);
|
Chris@10
|
999 T1J = T1H - T1I;
|
Chris@10
|
1000 T23 = T1I + T1H;
|
Chris@10
|
1001 TS = FNMS(KP250000000, TR, TK);
|
Chris@10
|
1002 TT = KP559016994 * (TN - TQ);
|
Chris@10
|
1003 TU = TS - TT;
|
Chris@10
|
1004 T1W = TT + TS;
|
Chris@10
|
1005 }
|
Chris@10
|
1006 {
|
Chris@10
|
1007 E T1q, T1K, T2e, T2g;
|
Chris@10
|
1008 T1q = TU - T1p;
|
Chris@10
|
1009 T1K = T1w + T1J;
|
Chris@10
|
1010 Rp[WS(rs, 1)] = FNMS(T1t, T1K, TJ * T1q);
|
Chris@10
|
1011 Rm[WS(rs, 1)] = FMA(T1t, T1q, TJ * T1K);
|
Chris@10
|
1012 T2e = T1W + T1X;
|
Chris@10
|
1013 T2g = T23 - T22;
|
Chris@10
|
1014 Rp[WS(rs, 7)] = FNMS(T2f, T2g, T2d * T2e);
|
Chris@10
|
1015 Rm[WS(rs, 7)] = FMA(T2f, T2e, T2d * T2g);
|
Chris@10
|
1016 }
|
Chris@10
|
1017 {
|
Chris@10
|
1018 E T1O, T1Q, T1Y, T24;
|
Chris@10
|
1019 T1O = TU + T1p;
|
Chris@10
|
1020 T1Q = T1J - T1w;
|
Chris@10
|
1021 Rp[WS(rs, 9)] = FNMS(T1P, T1Q, T1N * T1O);
|
Chris@10
|
1022 Rm[WS(rs, 9)] = FMA(T1P, T1O, T1N * T1Q);
|
Chris@10
|
1023 T1Y = T1W - T1X;
|
Chris@10
|
1024 T24 = T22 + T23;
|
Chris@10
|
1025 Rp[WS(rs, 3)] = FNMS(T21, T24, T1V * T1Y);
|
Chris@10
|
1026 Rm[WS(rs, 3)] = FMA(T21, T1Y, T1V * T24);
|
Chris@10
|
1027 }
|
Chris@10
|
1028 }
|
Chris@10
|
1029 {
|
Chris@10
|
1030 E T3f, T3N, T43, T3Z, T3K, T42, T3A, T3Y;
|
Chris@10
|
1031 T3f = FNMS(KP951056516, T3e, KP587785252 * T37);
|
Chris@10
|
1032 T3N = FNMS(KP951056516, T3M, KP587785252 * T3L);
|
Chris@10
|
1033 T43 = FMA(KP951056516, T3L, KP587785252 * T3M);
|
Chris@10
|
1034 T3Z = FMA(KP951056516, T37, KP587785252 * T3e);
|
Chris@10
|
1035 {
|
Chris@10
|
1036 E T3I, T3J, T3y, T3z;
|
Chris@10
|
1037 T3I = FNMS(KP250000000, T3H, T3E);
|
Chris@10
|
1038 T3J = KP559016994 * (T3F - T3G);
|
Chris@10
|
1039 T3K = T3I - T3J;
|
Chris@10
|
1040 T42 = T3J + T3I;
|
Chris@10
|
1041 T3y = FNMS(KP250000000, T3x, T3i);
|
Chris@10
|
1042 T3z = KP559016994 * (T3p - T3w);
|
Chris@10
|
1043 T3A = T3y - T3z;
|
Chris@10
|
1044 T3Y = T3z + T3y;
|
Chris@10
|
1045 }
|
Chris@10
|
1046 {
|
Chris@10
|
1047 E T3B, T3O, T45, T46;
|
Chris@10
|
1048 T3B = T3f + T3A;
|
Chris@10
|
1049 T3O = T3K - T3N;
|
Chris@10
|
1050 Ip[WS(rs, 1)] = FNMS(TH, T3O, TE * T3B);
|
Chris@10
|
1051 Im[WS(rs, 1)] = FMA(TE, T3O, TH * T3B);
|
Chris@10
|
1052 T45 = T3Z + T3Y;
|
Chris@10
|
1053 T46 = T42 - T43;
|
Chris@10
|
1054 Ip[WS(rs, 9)] = FNMS(T1M, T46, T1L * T45);
|
Chris@10
|
1055 Im[WS(rs, 9)] = FMA(T1L, T46, T1M * T45);
|
Chris@10
|
1056 }
|
Chris@10
|
1057 {
|
Chris@10
|
1058 E T3S, T3W, T40, T44;
|
Chris@10
|
1059 T3S = T3A - T3f;
|
Chris@10
|
1060 T3W = T3K + T3N;
|
Chris@10
|
1061 Ip[WS(rs, 3)] = FNMS(T3V, T3W, T3R * T3S);
|
Chris@10
|
1062 Im[WS(rs, 3)] = FMA(T3R, T3W, T3V * T3S);
|
Chris@10
|
1063 T40 = T3Y - T3Z;
|
Chris@10
|
1064 T44 = T42 + T43;
|
Chris@10
|
1065 Ip[WS(rs, 5)] = FNMS(T41, T44, T3X * T40);
|
Chris@10
|
1066 Im[WS(rs, 5)] = FMA(T3X, T44, T41 * T40);
|
Chris@10
|
1067 }
|
Chris@10
|
1068 }
|
Chris@10
|
1069 }
|
Chris@10
|
1070 }
|
Chris@10
|
1071 }
|
Chris@10
|
1072 }
|
Chris@10
|
1073
|
Chris@10
|
1074 static const tw_instr twinstr[] = {
|
Chris@10
|
1075 {TW_CEXP, 1, 1},
|
Chris@10
|
1076 {TW_CEXP, 1, 3},
|
Chris@10
|
1077 {TW_CEXP, 1, 9},
|
Chris@10
|
1078 {TW_CEXP, 1, 19},
|
Chris@10
|
1079 {TW_NEXT, 1, 0}
|
Chris@10
|
1080 };
|
Chris@10
|
1081
|
Chris@10
|
1082 static const hc2c_desc desc = { 20, "hc2cb2_20", twinstr, &GENUS, {204, 92, 72, 0} };
|
Chris@10
|
1083
|
Chris@10
|
1084 void X(codelet_hc2cb2_20) (planner *p) {
|
Chris@10
|
1085 X(khc2c_register) (p, hc2cb2_20, &desc, HC2C_VIA_RDFT);
|
Chris@10
|
1086 }
|
Chris@10
|
1087 #endif /* HAVE_FMA */
|