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 07:40:53 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_hc2cdft.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -twiddle-log3 -precompute-twiddles -n 20 -dit -name hc2cfdft2_20 -include hc2cf.h */
|
Chris@10
|
29
|
Chris@10
|
30 /*
|
Chris@10
|
31 * This function contains 316 FP additions, 238 FP multiplications,
|
Chris@10
|
32 * (or, 176 additions, 98 multiplications, 140 fused multiply/add),
|
Chris@10
|
33 * 180 stack variables, 5 constants, and 80 memory accesses
|
Chris@10
|
34 */
|
Chris@10
|
35 #include "hc2cf.h"
|
Chris@10
|
36
|
Chris@10
|
37 static void hc2cfdft2_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(KP250000000, +0.250000000000000000000000000000000000000000000);
|
Chris@10
|
42 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
|
Chris@10
|
43 DK(KP618033988, +0.618033988749894848204586834365638117720309180);
|
Chris@10
|
44 {
|
Chris@10
|
45 INT m;
|
Chris@10
|
46 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
|
47 E T5h, T5C, T5E, T5y, T5w, T5x, T5D, T5z;
|
Chris@10
|
48 {
|
Chris@10
|
49 E Tm, Tq, Tn, T1, T6, Tg, Tp, Tb, T1i, TU, Tr, TW, Tx, T2B, T1A;
|
Chris@10
|
50 E T1u, T2y, T33, T26, T1o, T30, T22, TD, T1Q, T2a, T2e, T2V, T2R, TG, T1V;
|
Chris@10
|
51 E TV, TH, TN, T2t, T12, T2p;
|
Chris@10
|
52 {
|
Chris@10
|
53 E Tw, To, T29, T1h, T1n, T2d, TC, T2U;
|
Chris@10
|
54 Tm = W[0];
|
Chris@10
|
55 Tq = W[3];
|
Chris@10
|
56 Tn = W[2];
|
Chris@10
|
57 T1 = W[6];
|
Chris@10
|
58 T6 = W[7];
|
Chris@10
|
59 Tw = Tm * Tq;
|
Chris@10
|
60 To = Tm * Tn;
|
Chris@10
|
61 T29 = Tm * T1;
|
Chris@10
|
62 T1h = Tn * T1;
|
Chris@10
|
63 T1n = Tn * T6;
|
Chris@10
|
64 T2d = Tm * T6;
|
Chris@10
|
65 Tg = W[5];
|
Chris@10
|
66 Tp = W[1];
|
Chris@10
|
67 Tb = W[4];
|
Chris@10
|
68 {
|
Chris@10
|
69 E T21, T25, T1t, T1z;
|
Chris@10
|
70 T1i = FMA(Tq, T6, T1h);
|
Chris@10
|
71 T25 = Tm * Tg;
|
Chris@10
|
72 T1z = Tn * Tg;
|
Chris@10
|
73 TU = FMA(Tp, Tq, To);
|
Chris@10
|
74 Tr = FNMS(Tp, Tq, To);
|
Chris@10
|
75 TW = FNMS(Tp, Tn, Tw);
|
Chris@10
|
76 Tx = FMA(Tp, Tn, Tw);
|
Chris@10
|
77 T1t = Tn * Tb;
|
Chris@10
|
78 T21 = Tm * Tb;
|
Chris@10
|
79 T2B = FMA(Tq, Tb, T1z);
|
Chris@10
|
80 T1A = FNMS(Tq, Tb, T1z);
|
Chris@10
|
81 TC = Tr * Tb;
|
Chris@10
|
82 T1u = FMA(Tq, Tg, T1t);
|
Chris@10
|
83 T2y = FNMS(Tq, Tg, T1t);
|
Chris@10
|
84 T33 = FMA(Tp, Tb, T25);
|
Chris@10
|
85 T26 = FNMS(Tp, Tb, T25);
|
Chris@10
|
86 T1o = FNMS(Tq, T1, T1n);
|
Chris@10
|
87 T30 = FNMS(Tp, Tg, T21);
|
Chris@10
|
88 T22 = FMA(Tp, Tg, T21);
|
Chris@10
|
89 }
|
Chris@10
|
90 TD = FMA(Tx, Tg, TC);
|
Chris@10
|
91 T1Q = FNMS(Tx, Tg, TC);
|
Chris@10
|
92 T2a = FMA(Tp, T6, T29);
|
Chris@10
|
93 T2e = FNMS(Tp, T1, T2d);
|
Chris@10
|
94 T2U = Tr * T6;
|
Chris@10
|
95 {
|
Chris@10
|
96 E T2Q, TE, TM, TF;
|
Chris@10
|
97 T2Q = Tr * T1;
|
Chris@10
|
98 TF = Tr * Tg;
|
Chris@10
|
99 T2V = FNMS(Tx, T1, T2U);
|
Chris@10
|
100 T2R = FMA(Tx, T6, T2Q);
|
Chris@10
|
101 TG = FNMS(Tx, Tb, TF);
|
Chris@10
|
102 T1V = FMA(Tx, Tb, TF);
|
Chris@10
|
103 TE = TD * T1;
|
Chris@10
|
104 TM = TD * T6;
|
Chris@10
|
105 TV = TU * Tb;
|
Chris@10
|
106 TH = FMA(TG, T6, TE);
|
Chris@10
|
107 TN = FNMS(TG, T1, TM);
|
Chris@10
|
108 T2t = TU * T1;
|
Chris@10
|
109 T12 = TU * Tg;
|
Chris@10
|
110 T2p = TU * T6;
|
Chris@10
|
111 }
|
Chris@10
|
112 }
|
Chris@10
|
113 {
|
Chris@10
|
114 E T36, T3Q, T5f, T4D, T5g, T2Y, T4E, T3P, T5R, T5k, T39, TT, T3T, T3m, T49;
|
Chris@10
|
115 E T4X, T5T, T5r, T3c, T2i, T3W, T3B, T4o, T4U, T5U, T5u, T3d, T2J, T3X, T3I;
|
Chris@10
|
116 E T4v, T4V, T5Q, T5n, T3a, T1G, T3U, T3t, T4g, T4Y;
|
Chris@10
|
117 {
|
Chris@10
|
118 E T13, T2m, T2q, T2u, T2f, T9, T2O, TA, T2c, T4k, T3i, T5, T2Z, T1e, T2G;
|
Chris@10
|
119 E T1O, T2W, TQ, T2C, T1Y, T3v, T27, Tj, T1l, T2v, T3g, T1m, T1D, T2n, T1x;
|
Chris@10
|
120 E T2k, T3E, T4c, T2l, T1y, T10, T31, T16, T34, T32, T11, T4B, T3p, T4A, T1T;
|
Chris@10
|
121 E T3n, T1b, T2A, T4q, T1U, Te, Tf, T24, T4i, T1r, T4a, T3C, T2s, T43, Tv;
|
Chris@10
|
122 E T3L, T2N, T45, TL, T3N, T2T, T2E, T1K;
|
Chris@10
|
123 {
|
Chris@10
|
124 E T2j, TX, T1B, T1C;
|
Chris@10
|
125 {
|
Chris@10
|
126 E T1c, T1d, T1M, T1N;
|
Chris@10
|
127 {
|
Chris@10
|
128 E T2, T3, T7, T8;
|
Chris@10
|
129 T7 = Rp[WS(rs, 9)];
|
Chris@10
|
130 T8 = Rm[WS(rs, 9)];
|
Chris@10
|
131 T2 = Ip[WS(rs, 9)];
|
Chris@10
|
132 T2j = FMA(TW, Tg, TV);
|
Chris@10
|
133 TX = FNMS(TW, Tg, TV);
|
Chris@10
|
134 T13 = FMA(TW, Tb, T12);
|
Chris@10
|
135 T2m = FNMS(TW, Tb, T12);
|
Chris@10
|
136 T2q = FNMS(TW, T1, T2p);
|
Chris@10
|
137 T2u = FMA(TW, T6, T2t);
|
Chris@10
|
138 T2f = T7 + T8;
|
Chris@10
|
139 T9 = T7 - T8;
|
Chris@10
|
140 T3 = Im[WS(rs, 9)];
|
Chris@10
|
141 {
|
Chris@10
|
142 E Ty, Tz, T2b, T4;
|
Chris@10
|
143 Ty = Rp[WS(rs, 2)];
|
Chris@10
|
144 Tz = Rm[WS(rs, 2)];
|
Chris@10
|
145 T1c = Ip[0];
|
Chris@10
|
146 T2b = T2 - T3;
|
Chris@10
|
147 T4 = T2 + T3;
|
Chris@10
|
148 T2O = Ty - Tz;
|
Chris@10
|
149 TA = Ty + Tz;
|
Chris@10
|
150 T2c = T2a * T2b;
|
Chris@10
|
151 T4k = T2e * T2b;
|
Chris@10
|
152 T3i = T6 * T4;
|
Chris@10
|
153 T5 = T1 * T4;
|
Chris@10
|
154 T1d = Im[0];
|
Chris@10
|
155 T1M = Rp[WS(rs, 1)];
|
Chris@10
|
156 T1N = Rm[WS(rs, 1)];
|
Chris@10
|
157 }
|
Chris@10
|
158 }
|
Chris@10
|
159 {
|
Chris@10
|
160 E TO, TP, T1W, T1X;
|
Chris@10
|
161 TO = Rp[WS(rs, 7)];
|
Chris@10
|
162 T2Z = T1c - T1d;
|
Chris@10
|
163 T1e = T1c + T1d;
|
Chris@10
|
164 T2G = T1M + T1N;
|
Chris@10
|
165 T1O = T1M - T1N;
|
Chris@10
|
166 TP = Rm[WS(rs, 7)];
|
Chris@10
|
167 T1W = Rm[WS(rs, 6)];
|
Chris@10
|
168 T1X = Rp[WS(rs, 6)];
|
Chris@10
|
169 {
|
Chris@10
|
170 E Th, Ti, T1j, T1k;
|
Chris@10
|
171 Th = Rm[WS(rs, 4)];
|
Chris@10
|
172 T2W = TO - TP;
|
Chris@10
|
173 TQ = TO + TP;
|
Chris@10
|
174 T2C = T1X + T1W;
|
Chris@10
|
175 T1Y = T1W - T1X;
|
Chris@10
|
176 Ti = Rp[WS(rs, 4)];
|
Chris@10
|
177 T1j = Ip[WS(rs, 8)];
|
Chris@10
|
178 T1k = Im[WS(rs, 8)];
|
Chris@10
|
179 T3v = T1Q * T1Y;
|
Chris@10
|
180 T27 = Ti + Th;
|
Chris@10
|
181 Tj = Th - Ti;
|
Chris@10
|
182 T1l = T1j - T1k;
|
Chris@10
|
183 T2v = T1j + T1k;
|
Chris@10
|
184 T1B = Rp[WS(rs, 3)];
|
Chris@10
|
185 T3g = Tb * Tj;
|
Chris@10
|
186 T1m = T1i * T1l;
|
Chris@10
|
187 T1C = Rm[WS(rs, 3)];
|
Chris@10
|
188 }
|
Chris@10
|
189 }
|
Chris@10
|
190 }
|
Chris@10
|
191 {
|
Chris@10
|
192 E T18, T19, T1R, T1S;
|
Chris@10
|
193 {
|
Chris@10
|
194 E TY, TZ, T1v, T1w, T14, T15;
|
Chris@10
|
195 T1v = Ip[WS(rs, 3)];
|
Chris@10
|
196 T1w = Im[WS(rs, 3)];
|
Chris@10
|
197 TY = Ip[WS(rs, 5)];
|
Chris@10
|
198 T1D = T1B + T1C;
|
Chris@10
|
199 T2n = T1B - T1C;
|
Chris@10
|
200 T1x = T1v - T1w;
|
Chris@10
|
201 T2k = T1v + T1w;
|
Chris@10
|
202 T3E = T2j * T2n;
|
Chris@10
|
203 T4c = T1u * T1D;
|
Chris@10
|
204 T2l = T2j * T2k;
|
Chris@10
|
205 T1y = T1u * T1x;
|
Chris@10
|
206 TZ = Im[WS(rs, 5)];
|
Chris@10
|
207 T14 = Rp[WS(rs, 5)];
|
Chris@10
|
208 T15 = Rm[WS(rs, 5)];
|
Chris@10
|
209 T18 = Rm[0];
|
Chris@10
|
210 T10 = TY + TZ;
|
Chris@10
|
211 T31 = TY - TZ;
|
Chris@10
|
212 T16 = T14 - T15;
|
Chris@10
|
213 T34 = T14 + T15;
|
Chris@10
|
214 T32 = T30 * T31;
|
Chris@10
|
215 T11 = TX * T10;
|
Chris@10
|
216 T4B = T30 * T34;
|
Chris@10
|
217 T3p = TX * T16;
|
Chris@10
|
218 T19 = Rp[0];
|
Chris@10
|
219 T1R = Ip[WS(rs, 6)];
|
Chris@10
|
220 T1S = Im[WS(rs, 6)];
|
Chris@10
|
221 }
|
Chris@10
|
222 {
|
Chris@10
|
223 E T2r, T23, T1p, T1q;
|
Chris@10
|
224 {
|
Chris@10
|
225 E Tc, T1a, T2z, Td;
|
Chris@10
|
226 Tc = Ip[WS(rs, 4)];
|
Chris@10
|
227 T1a = T18 - T19;
|
Chris@10
|
228 T4A = T19 + T18;
|
Chris@10
|
229 T1T = T1R + T1S;
|
Chris@10
|
230 T2z = T1R - T1S;
|
Chris@10
|
231 Td = Im[WS(rs, 4)];
|
Chris@10
|
232 T3n = Tm * T1a;
|
Chris@10
|
233 T1b = Tp * T1a;
|
Chris@10
|
234 T2A = T2y * T2z;
|
Chris@10
|
235 T4q = T2B * T2z;
|
Chris@10
|
236 T1U = T1Q * T1T;
|
Chris@10
|
237 T23 = Tc - Td;
|
Chris@10
|
238 Te = Tc + Td;
|
Chris@10
|
239 }
|
Chris@10
|
240 T1p = Rp[WS(rs, 8)];
|
Chris@10
|
241 T1q = Rm[WS(rs, 8)];
|
Chris@10
|
242 Tf = Tb * Te;
|
Chris@10
|
243 T24 = T22 * T23;
|
Chris@10
|
244 T4i = T26 * T23;
|
Chris@10
|
245 T1r = T1p + T1q;
|
Chris@10
|
246 T2r = T1q - T1p;
|
Chris@10
|
247 {
|
Chris@10
|
248 E T2M, Tu, Ts, Tt;
|
Chris@10
|
249 Ts = Ip[WS(rs, 2)];
|
Chris@10
|
250 Tt = Im[WS(rs, 2)];
|
Chris@10
|
251 T4a = T1i * T1r;
|
Chris@10
|
252 T3C = T2u * T2r;
|
Chris@10
|
253 T2s = T2q * T2r;
|
Chris@10
|
254 T2M = Ts + Tt;
|
Chris@10
|
255 Tu = Ts - Tt;
|
Chris@10
|
256 {
|
Chris@10
|
257 E T2S, TK, TI, TJ, T1I, T1J;
|
Chris@10
|
258 TI = Ip[WS(rs, 7)];
|
Chris@10
|
259 TJ = Im[WS(rs, 7)];
|
Chris@10
|
260 T43 = Tx * Tu;
|
Chris@10
|
261 Tv = Tr * Tu;
|
Chris@10
|
262 T3L = TG * T2M;
|
Chris@10
|
263 T2N = TD * T2M;
|
Chris@10
|
264 T2S = TI + TJ;
|
Chris@10
|
265 TK = TI - TJ;
|
Chris@10
|
266 T1I = Ip[WS(rs, 1)];
|
Chris@10
|
267 T1J = Im[WS(rs, 1)];
|
Chris@10
|
268 T45 = TN * TK;
|
Chris@10
|
269 TL = TH * TK;
|
Chris@10
|
270 T3N = T2V * T2S;
|
Chris@10
|
271 T2T = T2R * T2S;
|
Chris@10
|
272 T2E = T1I - T1J;
|
Chris@10
|
273 T1K = T1I + T1J;
|
Chris@10
|
274 }
|
Chris@10
|
275 }
|
Chris@10
|
276 }
|
Chris@10
|
277 }
|
Chris@10
|
278 }
|
Chris@10
|
279 {
|
Chris@10
|
280 E T3x, T1L, T2F, T4s, T2P, T2X, T3M, T3O, T35, T4C;
|
Chris@10
|
281 T35 = FNMS(T33, T34, T32);
|
Chris@10
|
282 T4C = FMA(T33, T31, T4B);
|
Chris@10
|
283 T3x = Tq * T1K;
|
Chris@10
|
284 T1L = Tn * T1K;
|
Chris@10
|
285 T2F = TU * T2E;
|
Chris@10
|
286 T4s = TW * T2E;
|
Chris@10
|
287 T36 = T2Z - T35;
|
Chris@10
|
288 T3Q = T35 + T2Z;
|
Chris@10
|
289 T5f = T4A + T4C;
|
Chris@10
|
290 T4D = T4A - T4C;
|
Chris@10
|
291 T2P = FNMS(TG, T2O, T2N);
|
Chris@10
|
292 T2X = FNMS(T2V, T2W, T2T);
|
Chris@10
|
293 T3M = FMA(TD, T2O, T3L);
|
Chris@10
|
294 T3O = FMA(T2R, T2W, T3N);
|
Chris@10
|
295 {
|
Chris@10
|
296 E TB, T5j, Tl, T5i, T47, TR, T3h, T3j;
|
Chris@10
|
297 {
|
Chris@10
|
298 E Ta, Tk, T44, T46;
|
Chris@10
|
299 Ta = FNMS(T6, T9, T5);
|
Chris@10
|
300 T5g = T2P + T2X;
|
Chris@10
|
301 T2Y = T2P - T2X;
|
Chris@10
|
302 T4E = T3O - T3M;
|
Chris@10
|
303 T3P = T3M + T3O;
|
Chris@10
|
304 Tk = FMA(Tg, Tj, Tf);
|
Chris@10
|
305 T44 = FMA(Tr, TA, T43);
|
Chris@10
|
306 T46 = FMA(TH, TQ, T45);
|
Chris@10
|
307 TB = FNMS(Tx, TA, Tv);
|
Chris@10
|
308 T5j = Tk + Ta;
|
Chris@10
|
309 Tl = Ta - Tk;
|
Chris@10
|
310 T5i = T44 + T46;
|
Chris@10
|
311 T47 = T44 - T46;
|
Chris@10
|
312 TR = FNMS(TN, TQ, TL);
|
Chris@10
|
313 T3h = FNMS(Tg, Te, T3g);
|
Chris@10
|
314 T3j = FMA(T1, T9, T3i);
|
Chris@10
|
315 }
|
Chris@10
|
316 {
|
Chris@10
|
317 E T3l, T48, T3k, TS;
|
Chris@10
|
318 T5R = T5i - T5j;
|
Chris@10
|
319 T5k = T5i + T5j;
|
Chris@10
|
320 T3l = TB + TR;
|
Chris@10
|
321 TS = TB - TR;
|
Chris@10
|
322 T48 = T3h + T3j;
|
Chris@10
|
323 T3k = T3h - T3j;
|
Chris@10
|
324 T39 = TS + Tl;
|
Chris@10
|
325 TT = Tl - TS;
|
Chris@10
|
326 T3T = T3l + T3k;
|
Chris@10
|
327 T3m = T3k - T3l;
|
Chris@10
|
328 T49 = T47 + T48;
|
Chris@10
|
329 T4X = T47 - T48;
|
Chris@10
|
330 }
|
Chris@10
|
331 }
|
Chris@10
|
332 {
|
Chris@10
|
333 E T28, T5q, T20, T5p, T4m, T2g, T3w, T3y;
|
Chris@10
|
334 {
|
Chris@10
|
335 E T1P, T1Z, T4j, T4l;
|
Chris@10
|
336 T1P = FNMS(Tq, T1O, T1L);
|
Chris@10
|
337 T1Z = FMA(T1V, T1Y, T1U);
|
Chris@10
|
338 T4j = FMA(T22, T27, T4i);
|
Chris@10
|
339 T4l = FMA(T2a, T2f, T4k);
|
Chris@10
|
340 T28 = FNMS(T26, T27, T24);
|
Chris@10
|
341 T5q = T1Z + T1P;
|
Chris@10
|
342 T20 = T1P - T1Z;
|
Chris@10
|
343 T5p = T4j + T4l;
|
Chris@10
|
344 T4m = T4j - T4l;
|
Chris@10
|
345 T2g = FNMS(T2e, T2f, T2c);
|
Chris@10
|
346 T3w = FNMS(T1V, T1T, T3v);
|
Chris@10
|
347 T3y = FMA(Tn, T1O, T3x);
|
Chris@10
|
348 }
|
Chris@10
|
349 {
|
Chris@10
|
350 E T3A, T4n, T3z, T2h;
|
Chris@10
|
351 T5T = T5p - T5q;
|
Chris@10
|
352 T5r = T5p + T5q;
|
Chris@10
|
353 T3A = T28 + T2g;
|
Chris@10
|
354 T2h = T28 - T2g;
|
Chris@10
|
355 T4n = T3w + T3y;
|
Chris@10
|
356 T3z = T3w - T3y;
|
Chris@10
|
357 T3c = T2h + T20;
|
Chris@10
|
358 T2i = T20 - T2h;
|
Chris@10
|
359 T3W = T3A + T3z;
|
Chris@10
|
360 T3B = T3z - T3A;
|
Chris@10
|
361 T4o = T4m + T4n;
|
Chris@10
|
362 T4U = T4m - T4n;
|
Chris@10
|
363 }
|
Chris@10
|
364 }
|
Chris@10
|
365 {
|
Chris@10
|
366 E T2D, T5s, T2x, T5t, T4u, T2H, T3D, T3F;
|
Chris@10
|
367 {
|
Chris@10
|
368 E T2o, T2w, T4r, T4t;
|
Chris@10
|
369 T2o = FNMS(T2m, T2n, T2l);
|
Chris@10
|
370 T2w = FMA(T2u, T2v, T2s);
|
Chris@10
|
371 T4r = FMA(T2y, T2C, T4q);
|
Chris@10
|
372 T4t = FMA(TU, T2G, T4s);
|
Chris@10
|
373 T2D = FNMS(T2B, T2C, T2A);
|
Chris@10
|
374 T5s = T2w + T2o;
|
Chris@10
|
375 T2x = T2o - T2w;
|
Chris@10
|
376 T5t = T4r + T4t;
|
Chris@10
|
377 T4u = T4r - T4t;
|
Chris@10
|
378 T2H = FNMS(TW, T2G, T2F);
|
Chris@10
|
379 T3D = FNMS(T2q, T2v, T3C);
|
Chris@10
|
380 T3F = FMA(T2m, T2k, T3E);
|
Chris@10
|
381 }
|
Chris@10
|
382 {
|
Chris@10
|
383 E T3H, T4p, T3G, T2I;
|
Chris@10
|
384 T5U = T5t - T5s;
|
Chris@10
|
385 T5u = T5s + T5t;
|
Chris@10
|
386 T3H = T2D + T2H;
|
Chris@10
|
387 T2I = T2D - T2H;
|
Chris@10
|
388 T4p = T3D + T3F;
|
Chris@10
|
389 T3G = T3D - T3F;
|
Chris@10
|
390 T3d = T2x + T2I;
|
Chris@10
|
391 T2J = T2x - T2I;
|
Chris@10
|
392 T3X = T3G + T3H;
|
Chris@10
|
393 T3I = T3G - T3H;
|
Chris@10
|
394 T4v = T4p + T4u;
|
Chris@10
|
395 T4V = T4u - T4p;
|
Chris@10
|
396 }
|
Chris@10
|
397 }
|
Chris@10
|
398 {
|
Chris@10
|
399 E T1s, T5m, T1g, T5l, T4e, T1E, T3o, T3q;
|
Chris@10
|
400 {
|
Chris@10
|
401 E T17, T1f, T4b, T4d;
|
Chris@10
|
402 T17 = FNMS(T13, T16, T11);
|
Chris@10
|
403 T1f = FMA(Tm, T1e, T1b);
|
Chris@10
|
404 T4b = FMA(T1o, T1l, T4a);
|
Chris@10
|
405 T4d = FMA(T1A, T1x, T4c);
|
Chris@10
|
406 T1s = FNMS(T1o, T1r, T1m);
|
Chris@10
|
407 T5m = T17 + T1f;
|
Chris@10
|
408 T1g = T17 - T1f;
|
Chris@10
|
409 T5l = T4b + T4d;
|
Chris@10
|
410 T4e = T4b - T4d;
|
Chris@10
|
411 T1E = FNMS(T1A, T1D, T1y);
|
Chris@10
|
412 T3o = FNMS(Tp, T1e, T3n);
|
Chris@10
|
413 T3q = FMA(T13, T10, T3p);
|
Chris@10
|
414 }
|
Chris@10
|
415 {
|
Chris@10
|
416 E T3s, T4f, T3r, T1F;
|
Chris@10
|
417 T5Q = T5l - T5m;
|
Chris@10
|
418 T5n = T5l + T5m;
|
Chris@10
|
419 T3s = T1s + T1E;
|
Chris@10
|
420 T1F = T1s - T1E;
|
Chris@10
|
421 T4f = T3q + T3o;
|
Chris@10
|
422 T3r = T3o - T3q;
|
Chris@10
|
423 T3a = T1F + T1g;
|
Chris@10
|
424 T1G = T1g - T1F;
|
Chris@10
|
425 T3U = T3s + T3r;
|
Chris@10
|
426 T3t = T3r - T3s;
|
Chris@10
|
427 T4g = T4e + T4f;
|
Chris@10
|
428 T4Y = T4e - T4f;
|
Chris@10
|
429 }
|
Chris@10
|
430 }
|
Chris@10
|
431 }
|
Chris@10
|
432 }
|
Chris@10
|
433 {
|
Chris@10
|
434 E T4F, T4G, T4H, T4x, T4z, T41, T4O, T4Q, T40;
|
Chris@10
|
435 {
|
Chris@10
|
436 E T55, T38, T54, T50, T52, T53, T5e, T5c, T51, T4T;
|
Chris@10
|
437 {
|
Chris@10
|
438 E T4W, T37, T4Z, T1H, T5b, T5a, T2K, T2L, T4S, T4R;
|
Chris@10
|
439 T55 = T4U + T4V;
|
Chris@10
|
440 T4W = T4U - T4V;
|
Chris@10
|
441 T37 = T2Y + T36;
|
Chris@10
|
442 T38 = T36 - T2Y;
|
Chris@10
|
443 T54 = T4X + T4Y;
|
Chris@10
|
444 T4Z = T4X - T4Y;
|
Chris@10
|
445 T1H = TT + T1G;
|
Chris@10
|
446 T5b = T1G - TT;
|
Chris@10
|
447 T5a = T2J - T2i;
|
Chris@10
|
448 T2K = T2i + T2J;
|
Chris@10
|
449 T50 = FNMS(KP618033988, T4Z, T4W);
|
Chris@10
|
450 T52 = FMA(KP618033988, T4W, T4Z);
|
Chris@10
|
451 T2L = T1H + T2K;
|
Chris@10
|
452 T4S = T1H - T2K;
|
Chris@10
|
453 T53 = T4D - T4E;
|
Chris@10
|
454 T4F = T4D + T4E;
|
Chris@10
|
455 Im[WS(rs, 4)] = KP500000000 * (T2L - T37);
|
Chris@10
|
456 T4R = FMA(KP250000000, T2L, T37);
|
Chris@10
|
457 T5e = FMA(KP618033988, T5a, T5b);
|
Chris@10
|
458 T5c = FNMS(KP618033988, T5b, T5a);
|
Chris@10
|
459 T51 = FNMS(KP559016994, T4S, T4R);
|
Chris@10
|
460 T4T = FMA(KP559016994, T4S, T4R);
|
Chris@10
|
461 }
|
Chris@10
|
462 {
|
Chris@10
|
463 E T3b, T4M, T4N, T3e, T3f;
|
Chris@10
|
464 {
|
Chris@10
|
465 E T4h, T58, T57, T4w, T56, T5d, T59;
|
Chris@10
|
466 T4G = T49 + T4g;
|
Chris@10
|
467 T4h = T49 - T4g;
|
Chris@10
|
468 T58 = T54 - T55;
|
Chris@10
|
469 T56 = T54 + T55;
|
Chris@10
|
470 Ip[WS(rs, 7)] = KP500000000 * (FMA(KP951056516, T50, T4T));
|
Chris@10
|
471 Ip[WS(rs, 3)] = KP500000000 * (FNMS(KP951056516, T50, T4T));
|
Chris@10
|
472 Im[WS(rs, 8)] = -(KP500000000 * (FNMS(KP951056516, T52, T51)));
|
Chris@10
|
473 Im[0] = -(KP500000000 * (FMA(KP951056516, T52, T51)));
|
Chris@10
|
474 Rm[WS(rs, 4)] = KP500000000 * (T53 + T56);
|
Chris@10
|
475 T57 = FNMS(KP250000000, T56, T53);
|
Chris@10
|
476 T4w = T4o - T4v;
|
Chris@10
|
477 T4H = T4o + T4v;
|
Chris@10
|
478 T3b = T39 + T3a;
|
Chris@10
|
479 T4M = T39 - T3a;
|
Chris@10
|
480 T5d = FMA(KP559016994, T58, T57);
|
Chris@10
|
481 T59 = FNMS(KP559016994, T58, T57);
|
Chris@10
|
482 T4x = FMA(KP618033988, T4w, T4h);
|
Chris@10
|
483 T4z = FNMS(KP618033988, T4h, T4w);
|
Chris@10
|
484 Rp[WS(rs, 7)] = KP500000000 * (FNMS(KP951056516, T5c, T59));
|
Chris@10
|
485 Rp[WS(rs, 3)] = KP500000000 * (FMA(KP951056516, T5c, T59));
|
Chris@10
|
486 Rm[0] = KP500000000 * (FNMS(KP951056516, T5e, T5d));
|
Chris@10
|
487 Rm[WS(rs, 8)] = KP500000000 * (FMA(KP951056516, T5e, T5d));
|
Chris@10
|
488 T4N = T3c - T3d;
|
Chris@10
|
489 T3e = T3c + T3d;
|
Chris@10
|
490 }
|
Chris@10
|
491 T3f = T3b + T3e;
|
Chris@10
|
492 T41 = T3b - T3e;
|
Chris@10
|
493 T4O = FMA(KP618033988, T4N, T4M);
|
Chris@10
|
494 T4Q = FNMS(KP618033988, T4M, T4N);
|
Chris@10
|
495 Ip[WS(rs, 5)] = KP500000000 * (T38 + T3f);
|
Chris@10
|
496 T40 = FNMS(KP250000000, T3f, T38);
|
Chris@10
|
497 }
|
Chris@10
|
498 }
|
Chris@10
|
499 {
|
Chris@10
|
500 E T3S, T5Z, T68, T6a, T64, T62;
|
Chris@10
|
501 {
|
Chris@10
|
502 E T60, T61, T5Y, T5W, T3R, T67, T66, T3K, T5O, T4K, T4J, T5N, T5X, T5P;
|
Chris@10
|
503 {
|
Chris@10
|
504 E T5S, T5V, T4y, T42, T4I;
|
Chris@10
|
505 T60 = T5R + T5Q;
|
Chris@10
|
506 T5S = T5Q - T5R;
|
Chris@10
|
507 T5V = T5T - T5U;
|
Chris@10
|
508 T61 = T5T + T5U;
|
Chris@10
|
509 T4y = FNMS(KP559016994, T41, T40);
|
Chris@10
|
510 T42 = FMA(KP559016994, T41, T40);
|
Chris@10
|
511 T4I = T4G + T4H;
|
Chris@10
|
512 T4K = T4G - T4H;
|
Chris@10
|
513 Ip[WS(rs, 9)] = KP500000000 * (FMA(KP951056516, T4x, T42));
|
Chris@10
|
514 Ip[WS(rs, 1)] = KP500000000 * (FNMS(KP951056516, T4x, T42));
|
Chris@10
|
515 Im[WS(rs, 6)] = -(KP500000000 * (FNMS(KP951056516, T4z, T4y)));
|
Chris@10
|
516 Im[WS(rs, 2)] = -(KP500000000 * (FMA(KP951056516, T4z, T4y)));
|
Chris@10
|
517 Rp[WS(rs, 5)] = KP500000000 * (T4F + T4I);
|
Chris@10
|
518 T4J = FNMS(KP250000000, T4I, T4F);
|
Chris@10
|
519 T5Y = FMA(KP618033988, T5S, T5V);
|
Chris@10
|
520 T5W = FNMS(KP618033988, T5V, T5S);
|
Chris@10
|
521 }
|
Chris@10
|
522 T3S = T3Q - T3P;
|
Chris@10
|
523 T3R = T3P + T3Q;
|
Chris@10
|
524 {
|
Chris@10
|
525 E T4L, T4P, T3u, T3J;
|
Chris@10
|
526 T4L = FMA(KP559016994, T4K, T4J);
|
Chris@10
|
527 T4P = FNMS(KP559016994, T4K, T4J);
|
Chris@10
|
528 T3u = T3m + T3t;
|
Chris@10
|
529 T67 = T3t - T3m;
|
Chris@10
|
530 T66 = T3I - T3B;
|
Chris@10
|
531 T3J = T3B + T3I;
|
Chris@10
|
532 Rp[WS(rs, 9)] = KP500000000 * (FNMS(KP951056516, T4O, T4L));
|
Chris@10
|
533 Rp[WS(rs, 1)] = KP500000000 * (FMA(KP951056516, T4O, T4L));
|
Chris@10
|
534 Rm[WS(rs, 2)] = KP500000000 * (FNMS(KP951056516, T4Q, T4P));
|
Chris@10
|
535 Rm[WS(rs, 6)] = KP500000000 * (FMA(KP951056516, T4Q, T4P));
|
Chris@10
|
536 T3K = T3u + T3J;
|
Chris@10
|
537 T5O = T3J - T3u;
|
Chris@10
|
538 }
|
Chris@10
|
539 Im[WS(rs, 9)] = KP500000000 * (T3K - T3R);
|
Chris@10
|
540 T5N = FMA(KP250000000, T3K, T3R);
|
Chris@10
|
541 T5Z = T5f - T5g;
|
Chris@10
|
542 T5h = T5f + T5g;
|
Chris@10
|
543 T68 = FNMS(KP618033988, T67, T66);
|
Chris@10
|
544 T6a = FMA(KP618033988, T66, T67);
|
Chris@10
|
545 T5X = FNMS(KP559016994, T5O, T5N);
|
Chris@10
|
546 T5P = FMA(KP559016994, T5O, T5N);
|
Chris@10
|
547 Im[WS(rs, 5)] = -(KP500000000 * (FNMS(KP951056516, T5W, T5P)));
|
Chris@10
|
548 Ip[WS(rs, 6)] = KP500000000 * (FMA(KP951056516, T5W, T5P));
|
Chris@10
|
549 Im[WS(rs, 1)] = -(KP500000000 * (FNMS(KP951056516, T5Y, T5X)));
|
Chris@10
|
550 Ip[WS(rs, 2)] = KP500000000 * (FMA(KP951056516, T5Y, T5X));
|
Chris@10
|
551 T64 = T60 - T61;
|
Chris@10
|
552 T62 = T60 + T61;
|
Chris@10
|
553 }
|
Chris@10
|
554 {
|
Chris@10
|
555 E T5o, T5v, T5M, T5K, T5A, T5B, T3Z, T5G, T5I, T5J, T63, T5F, T5L, T5H;
|
Chris@10
|
556 T5o = T5k + T5n;
|
Chris@10
|
557 T5I = T5k - T5n;
|
Chris@10
|
558 T5J = T5u - T5r;
|
Chris@10
|
559 T5v = T5r + T5u;
|
Chris@10
|
560 Rm[WS(rs, 9)] = KP500000000 * (T5Z + T62);
|
Chris@10
|
561 T63 = FNMS(KP250000000, T62, T5Z);
|
Chris@10
|
562 T5M = FMA(KP618033988, T5I, T5J);
|
Chris@10
|
563 T5K = FNMS(KP618033988, T5J, T5I);
|
Chris@10
|
564 {
|
Chris@10
|
565 E T65, T69, T3V, T3Y;
|
Chris@10
|
566 T65 = FNMS(KP559016994, T64, T63);
|
Chris@10
|
567 T69 = FMA(KP559016994, T64, T63);
|
Chris@10
|
568 T3V = T3T + T3U;
|
Chris@10
|
569 T5A = T3T - T3U;
|
Chris@10
|
570 T5B = T3W - T3X;
|
Chris@10
|
571 T3Y = T3W + T3X;
|
Chris@10
|
572 Rm[WS(rs, 1)] = KP500000000 * (FMA(KP951056516, T68, T65));
|
Chris@10
|
573 Rp[WS(rs, 2)] = KP500000000 * (FNMS(KP951056516, T68, T65));
|
Chris@10
|
574 Rm[WS(rs, 5)] = KP500000000 * (FNMS(KP951056516, T6a, T69));
|
Chris@10
|
575 Rp[WS(rs, 6)] = KP500000000 * (FMA(KP951056516, T6a, T69));
|
Chris@10
|
576 T3Z = T3V + T3Y;
|
Chris@10
|
577 T5G = T3V - T3Y;
|
Chris@10
|
578 }
|
Chris@10
|
579 Ip[0] = KP500000000 * (T3S + T3Z);
|
Chris@10
|
580 T5F = FNMS(KP250000000, T3Z, T3S);
|
Chris@10
|
581 T5C = FMA(KP618033988, T5B, T5A);
|
Chris@10
|
582 T5E = FNMS(KP618033988, T5A, T5B);
|
Chris@10
|
583 T5L = FNMS(KP559016994, T5G, T5F);
|
Chris@10
|
584 T5H = FMA(KP559016994, T5G, T5F);
|
Chris@10
|
585 Im[WS(rs, 3)] = -(KP500000000 * (FNMS(KP951056516, T5K, T5H)));
|
Chris@10
|
586 Ip[WS(rs, 4)] = KP500000000 * (FMA(KP951056516, T5K, T5H));
|
Chris@10
|
587 Im[WS(rs, 7)] = -(KP500000000 * (FNMS(KP951056516, T5M, T5L)));
|
Chris@10
|
588 Ip[WS(rs, 8)] = KP500000000 * (FMA(KP951056516, T5M, T5L));
|
Chris@10
|
589 T5y = T5o - T5v;
|
Chris@10
|
590 T5w = T5o + T5v;
|
Chris@10
|
591 }
|
Chris@10
|
592 }
|
Chris@10
|
593 }
|
Chris@10
|
594 }
|
Chris@10
|
595 }
|
Chris@10
|
596 Rp[0] = KP500000000 * (T5h + T5w);
|
Chris@10
|
597 T5x = FNMS(KP250000000, T5w, T5h);
|
Chris@10
|
598 T5D = FNMS(KP559016994, T5y, T5x);
|
Chris@10
|
599 T5z = FMA(KP559016994, T5y, T5x);
|
Chris@10
|
600 Rm[WS(rs, 3)] = KP500000000 * (FMA(KP951056516, T5C, T5z));
|
Chris@10
|
601 Rp[WS(rs, 4)] = KP500000000 * (FNMS(KP951056516, T5C, T5z));
|
Chris@10
|
602 Rm[WS(rs, 7)] = KP500000000 * (FNMS(KP951056516, T5E, T5D));
|
Chris@10
|
603 Rp[WS(rs, 8)] = KP500000000 * (FMA(KP951056516, T5E, T5D));
|
Chris@10
|
604 }
|
Chris@10
|
605 }
|
Chris@10
|
606 }
|
Chris@10
|
607
|
Chris@10
|
608 static const tw_instr twinstr[] = {
|
Chris@10
|
609 {TW_CEXP, 1, 1},
|
Chris@10
|
610 {TW_CEXP, 1, 3},
|
Chris@10
|
611 {TW_CEXP, 1, 9},
|
Chris@10
|
612 {TW_CEXP, 1, 19},
|
Chris@10
|
613 {TW_NEXT, 1, 0}
|
Chris@10
|
614 };
|
Chris@10
|
615
|
Chris@10
|
616 static const hc2c_desc desc = { 20, "hc2cfdft2_20", twinstr, &GENUS, {176, 98, 140, 0} };
|
Chris@10
|
617
|
Chris@10
|
618 void X(codelet_hc2cfdft2_20) (planner *p) {
|
Chris@10
|
619 X(khc2c_register) (p, hc2cfdft2_20, &desc, HC2C_VIA_DFT);
|
Chris@10
|
620 }
|
Chris@10
|
621 #else /* HAVE_FMA */
|
Chris@10
|
622
|
Chris@10
|
623 /* Generated by: ../../../genfft/gen_hc2cdft.native -compact -variables 4 -pipeline-latency 4 -twiddle-log3 -precompute-twiddles -n 20 -dit -name hc2cfdft2_20 -include hc2cf.h */
|
Chris@10
|
624
|
Chris@10
|
625 /*
|
Chris@10
|
626 * This function contains 316 FP additions, 180 FP multiplications,
|
Chris@10
|
627 * (or, 244 additions, 108 multiplications, 72 fused multiply/add),
|
Chris@10
|
628 * 134 stack variables, 5 constants, and 80 memory accesses
|
Chris@10
|
629 */
|
Chris@10
|
630 #include "hc2cf.h"
|
Chris@10
|
631
|
Chris@10
|
632 static void hc2cfdft2_20(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms)
|
Chris@10
|
633 {
|
Chris@10
|
634 DK(KP125000000, +0.125000000000000000000000000000000000000000000);
|
Chris@10
|
635 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
|
Chris@10
|
636 DK(KP279508497, +0.279508497187473712051146708591409529430077295);
|
Chris@10
|
637 DK(KP293892626, +0.293892626146236564584352977319536384298826219);
|
Chris@10
|
638 DK(KP475528258, +0.475528258147576786058219666689691071702849317);
|
Chris@10
|
639 {
|
Chris@10
|
640 INT m;
|
Chris@10
|
641 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
|
642 E T4, T7, Tm, To, Tq, Tu, T1I, T1G, T8, T5, Ta, T1u, T2u, Tg, T2s;
|
Chris@10
|
643 E T21, T1A, T1Z, T1O, T2I, T1K, T2G, Tw, TC, T2a, T2e, TH, TI, TJ, TX;
|
Chris@10
|
644 E T2D, TN, T2B, T26, T1n, TZ, T24, T1j;
|
Chris@10
|
645 {
|
Chris@10
|
646 E T9, T1y, Te, T1t, T6, T1z, Tf, T1s;
|
Chris@10
|
647 {
|
Chris@10
|
648 E Tn, Tt, Tp, Ts;
|
Chris@10
|
649 T4 = W[0];
|
Chris@10
|
650 T7 = W[1];
|
Chris@10
|
651 Tm = W[2];
|
Chris@10
|
652 To = W[3];
|
Chris@10
|
653 Tn = T4 * Tm;
|
Chris@10
|
654 Tt = T7 * Tm;
|
Chris@10
|
655 Tp = T7 * To;
|
Chris@10
|
656 Ts = T4 * To;
|
Chris@10
|
657 Tq = Tn - Tp;
|
Chris@10
|
658 Tu = Ts + Tt;
|
Chris@10
|
659 T1I = Ts - Tt;
|
Chris@10
|
660 T1G = Tn + Tp;
|
Chris@10
|
661 T8 = W[5];
|
Chris@10
|
662 T9 = T7 * T8;
|
Chris@10
|
663 T1y = Tm * T8;
|
Chris@10
|
664 Te = T4 * T8;
|
Chris@10
|
665 T1t = To * T8;
|
Chris@10
|
666 T5 = W[4];
|
Chris@10
|
667 T6 = T4 * T5;
|
Chris@10
|
668 T1z = To * T5;
|
Chris@10
|
669 Tf = T7 * T5;
|
Chris@10
|
670 T1s = Tm * T5;
|
Chris@10
|
671 }
|
Chris@10
|
672 Ta = T6 - T9;
|
Chris@10
|
673 T1u = T1s + T1t;
|
Chris@10
|
674 T2u = T1y + T1z;
|
Chris@10
|
675 Tg = Te + Tf;
|
Chris@10
|
676 T2s = T1s - T1t;
|
Chris@10
|
677 T21 = Te - Tf;
|
Chris@10
|
678 T1A = T1y - T1z;
|
Chris@10
|
679 T1Z = T6 + T9;
|
Chris@10
|
680 {
|
Chris@10
|
681 E T1M, T1N, T1H, T1J;
|
Chris@10
|
682 T1M = T1G * T8;
|
Chris@10
|
683 T1N = T1I * T5;
|
Chris@10
|
684 T1O = T1M + T1N;
|
Chris@10
|
685 T2I = T1M - T1N;
|
Chris@10
|
686 T1H = T1G * T5;
|
Chris@10
|
687 T1J = T1I * T8;
|
Chris@10
|
688 T1K = T1H - T1J;
|
Chris@10
|
689 T2G = T1H + T1J;
|
Chris@10
|
690 {
|
Chris@10
|
691 E Tr, Tv, TA, TB;
|
Chris@10
|
692 Tr = Tq * T5;
|
Chris@10
|
693 Tv = Tu * T8;
|
Chris@10
|
694 Tw = Tr + Tv;
|
Chris@10
|
695 TA = Tq * T8;
|
Chris@10
|
696 TB = Tu * T5;
|
Chris@10
|
697 TC = TA - TB;
|
Chris@10
|
698 T2a = Tr - Tv;
|
Chris@10
|
699 T2e = TA + TB;
|
Chris@10
|
700 TH = W[6];
|
Chris@10
|
701 TI = W[7];
|
Chris@10
|
702 TJ = FMA(Tq, TH, Tu * TI);
|
Chris@10
|
703 TX = FMA(Tw, TH, TC * TI);
|
Chris@10
|
704 T2D = FMA(T1G, TH, T1I * TI);
|
Chris@10
|
705 TN = FNMS(Tu, TH, Tq * TI);
|
Chris@10
|
706 T2B = FNMS(T1I, TH, T1G * TI);
|
Chris@10
|
707 T26 = FNMS(T7, TH, T4 * TI);
|
Chris@10
|
708 T1n = FNMS(To, TH, Tm * TI);
|
Chris@10
|
709 TZ = FNMS(TC, TH, Tw * TI);
|
Chris@10
|
710 T24 = FMA(T4, TH, T7 * TI);
|
Chris@10
|
711 T1j = FMA(Tm, TH, To * TI);
|
Chris@10
|
712 }
|
Chris@10
|
713 }
|
Chris@10
|
714 }
|
Chris@10
|
715 {
|
Chris@10
|
716 E Tl, T3n, T1i, T2Q, T47, T50, T4S, T5i, T2M, T2T, T4I, T5f, T4L, T5e, T4P;
|
Chris@10
|
717 E T5h, T2r, T2S, T1X, T2P, T31, T3u, T36, T3t, T3E, T4l, T3U, T4j, T3h, T3r;
|
Chris@10
|
718 E T3J, T4m, T3c, T3q, T3P, T4i, TS, T51, T3m, T48;
|
Chris@10
|
719 {
|
Chris@10
|
720 E T3, T45, T1V, T3f, Tz, TF, TW, T3A, TM, TQ, T11, T3B, Td, Tj, T1Q;
|
Chris@10
|
721 E T3e, T19, T3L, T23, T39, T2p, T3S, T2z, T34, T1E, T3G, T2K, T2Y, T1g, T3M;
|
Chris@10
|
722 E T28, T3a, T2i, T3R, T2w, T33, T1r, T3F, T2F, T2X, T4N, T4O;
|
Chris@10
|
723 {
|
Chris@10
|
724 E T1, T2, T1R, T1S, T1T, T1U;
|
Chris@10
|
725 T1 = Ip[0];
|
Chris@10
|
726 T2 = Im[0];
|
Chris@10
|
727 T1R = T1 + T2;
|
Chris@10
|
728 T1S = Rp[0];
|
Chris@10
|
729 T1T = Rm[0];
|
Chris@10
|
730 T1U = T1S - T1T;
|
Chris@10
|
731 T3 = T1 - T2;
|
Chris@10
|
732 T45 = T1S + T1T;
|
Chris@10
|
733 T1V = FNMS(T7, T1U, T4 * T1R);
|
Chris@10
|
734 T3f = FMA(T4, T1U, T7 * T1R);
|
Chris@10
|
735 }
|
Chris@10
|
736 {
|
Chris@10
|
737 E Tx, Ty, TU, TD, TE, TV;
|
Chris@10
|
738 Tx = Ip[WS(rs, 2)];
|
Chris@10
|
739 Ty = Im[WS(rs, 2)];
|
Chris@10
|
740 TU = Tx - Ty;
|
Chris@10
|
741 TD = Rp[WS(rs, 2)];
|
Chris@10
|
742 TE = Rm[WS(rs, 2)];
|
Chris@10
|
743 TV = TD + TE;
|
Chris@10
|
744 Tz = Tx + Ty;
|
Chris@10
|
745 TF = TD - TE;
|
Chris@10
|
746 TW = FNMS(Tu, TV, Tq * TU);
|
Chris@10
|
747 T3A = FMA(Tu, TU, Tq * TV);
|
Chris@10
|
748 }
|
Chris@10
|
749 {
|
Chris@10
|
750 E TK, TL, TY, TO, TP, T10;
|
Chris@10
|
751 TK = Ip[WS(rs, 7)];
|
Chris@10
|
752 TL = Im[WS(rs, 7)];
|
Chris@10
|
753 TY = TK - TL;
|
Chris@10
|
754 TO = Rp[WS(rs, 7)];
|
Chris@10
|
755 TP = Rm[WS(rs, 7)];
|
Chris@10
|
756 T10 = TO + TP;
|
Chris@10
|
757 TM = TK + TL;
|
Chris@10
|
758 TQ = TO - TP;
|
Chris@10
|
759 T11 = FNMS(TZ, T10, TX * TY);
|
Chris@10
|
760 T3B = FMA(TZ, TY, TX * T10);
|
Chris@10
|
761 }
|
Chris@10
|
762 {
|
Chris@10
|
763 E Tb, Tc, T1L, Th, Ti, T1P;
|
Chris@10
|
764 Tb = Ip[WS(rs, 5)];
|
Chris@10
|
765 Tc = Im[WS(rs, 5)];
|
Chris@10
|
766 T1L = Tb + Tc;
|
Chris@10
|
767 Th = Rp[WS(rs, 5)];
|
Chris@10
|
768 Ti = Rm[WS(rs, 5)];
|
Chris@10
|
769 T1P = Th - Ti;
|
Chris@10
|
770 Td = Tb - Tc;
|
Chris@10
|
771 Tj = Th + Ti;
|
Chris@10
|
772 T1Q = FNMS(T1O, T1P, T1K * T1L);
|
Chris@10
|
773 T3e = FMA(T1K, T1P, T1O * T1L);
|
Chris@10
|
774 }
|
Chris@10
|
775 {
|
Chris@10
|
776 E T15, T20, T18, T22;
|
Chris@10
|
777 {
|
Chris@10
|
778 E T13, T14, T16, T17;
|
Chris@10
|
779 T13 = Ip[WS(rs, 4)];
|
Chris@10
|
780 T14 = Im[WS(rs, 4)];
|
Chris@10
|
781 T15 = T13 + T14;
|
Chris@10
|
782 T20 = T13 - T14;
|
Chris@10
|
783 T16 = Rp[WS(rs, 4)];
|
Chris@10
|
784 T17 = Rm[WS(rs, 4)];
|
Chris@10
|
785 T18 = T16 - T17;
|
Chris@10
|
786 T22 = T16 + T17;
|
Chris@10
|
787 }
|
Chris@10
|
788 T19 = FNMS(T8, T18, T5 * T15);
|
Chris@10
|
789 T3L = FMA(T21, T20, T1Z * T22);
|
Chris@10
|
790 T23 = FNMS(T21, T22, T1Z * T20);
|
Chris@10
|
791 T39 = FMA(T8, T15, T5 * T18);
|
Chris@10
|
792 }
|
Chris@10
|
793 {
|
Chris@10
|
794 E T2l, T2x, T2o, T2y;
|
Chris@10
|
795 {
|
Chris@10
|
796 E T2j, T2k, T2m, T2n;
|
Chris@10
|
797 T2j = Ip[WS(rs, 1)];
|
Chris@10
|
798 T2k = Im[WS(rs, 1)];
|
Chris@10
|
799 T2l = T2j + T2k;
|
Chris@10
|
800 T2x = T2j - T2k;
|
Chris@10
|
801 T2m = Rp[WS(rs, 1)];
|
Chris@10
|
802 T2n = Rm[WS(rs, 1)];
|
Chris@10
|
803 T2o = T2m - T2n;
|
Chris@10
|
804 T2y = T2m + T2n;
|
Chris@10
|
805 }
|
Chris@10
|
806 T2p = FNMS(To, T2o, Tm * T2l);
|
Chris@10
|
807 T3S = FMA(T1I, T2x, T1G * T2y);
|
Chris@10
|
808 T2z = FNMS(T1I, T2y, T1G * T2x);
|
Chris@10
|
809 T34 = FMA(To, T2l, Tm * T2o);
|
Chris@10
|
810 }
|
Chris@10
|
811 {
|
Chris@10
|
812 E T1x, T2H, T1D, T2J;
|
Chris@10
|
813 {
|
Chris@10
|
814 E T1v, T1w, T1B, T1C;
|
Chris@10
|
815 T1v = Ip[WS(rs, 3)];
|
Chris@10
|
816 T1w = Im[WS(rs, 3)];
|
Chris@10
|
817 T1x = T1v - T1w;
|
Chris@10
|
818 T2H = T1v + T1w;
|
Chris@10
|
819 T1B = Rp[WS(rs, 3)];
|
Chris@10
|
820 T1C = Rm[WS(rs, 3)];
|
Chris@10
|
821 T1D = T1B + T1C;
|
Chris@10
|
822 T2J = T1B - T1C;
|
Chris@10
|
823 }
|
Chris@10
|
824 T1E = FNMS(T1A, T1D, T1u * T1x);
|
Chris@10
|
825 T3G = FMA(T1u, T1D, T1A * T1x);
|
Chris@10
|
826 T2K = FNMS(T2I, T2J, T2G * T2H);
|
Chris@10
|
827 T2Y = FMA(T2G, T2J, T2I * T2H);
|
Chris@10
|
828 }
|
Chris@10
|
829 {
|
Chris@10
|
830 E T1c, T25, T1f, T27;
|
Chris@10
|
831 {
|
Chris@10
|
832 E T1a, T1b, T1d, T1e;
|
Chris@10
|
833 T1a = Ip[WS(rs, 9)];
|
Chris@10
|
834 T1b = Im[WS(rs, 9)];
|
Chris@10
|
835 T1c = T1a + T1b;
|
Chris@10
|
836 T25 = T1a - T1b;
|
Chris@10
|
837 T1d = Rp[WS(rs, 9)];
|
Chris@10
|
838 T1e = Rm[WS(rs, 9)];
|
Chris@10
|
839 T1f = T1d - T1e;
|
Chris@10
|
840 T27 = T1d + T1e;
|
Chris@10
|
841 }
|
Chris@10
|
842 T1g = FNMS(TI, T1f, TH * T1c);
|
Chris@10
|
843 T3M = FMA(T26, T25, T24 * T27);
|
Chris@10
|
844 T28 = FNMS(T26, T27, T24 * T25);
|
Chris@10
|
845 T3a = FMA(TI, T1c, TH * T1f);
|
Chris@10
|
846 }
|
Chris@10
|
847 {
|
Chris@10
|
848 E T2d, T2t, T2h, T2v;
|
Chris@10
|
849 {
|
Chris@10
|
850 E T2b, T2c, T2f, T2g;
|
Chris@10
|
851 T2b = Ip[WS(rs, 6)];
|
Chris@10
|
852 T2c = Im[WS(rs, 6)];
|
Chris@10
|
853 T2d = T2b + T2c;
|
Chris@10
|
854 T2t = T2b - T2c;
|
Chris@10
|
855 T2f = Rp[WS(rs, 6)];
|
Chris@10
|
856 T2g = Rm[WS(rs, 6)];
|
Chris@10
|
857 T2h = T2f - T2g;
|
Chris@10
|
858 T2v = T2f + T2g;
|
Chris@10
|
859 }
|
Chris@10
|
860 T2i = FNMS(T2e, T2h, T2a * T2d);
|
Chris@10
|
861 T3R = FMA(T2u, T2t, T2s * T2v);
|
Chris@10
|
862 T2w = FNMS(T2u, T2v, T2s * T2t);
|
Chris@10
|
863 T33 = FMA(T2e, T2d, T2a * T2h);
|
Chris@10
|
864 }
|
Chris@10
|
865 {
|
Chris@10
|
866 E T1m, T2E, T1q, T2C;
|
Chris@10
|
867 {
|
Chris@10
|
868 E T1k, T1l, T1o, T1p;
|
Chris@10
|
869 T1k = Ip[WS(rs, 8)];
|
Chris@10
|
870 T1l = Im[WS(rs, 8)];
|
Chris@10
|
871 T1m = T1k - T1l;
|
Chris@10
|
872 T2E = T1k + T1l;
|
Chris@10
|
873 T1o = Rp[WS(rs, 8)];
|
Chris@10
|
874 T1p = Rm[WS(rs, 8)];
|
Chris@10
|
875 T1q = T1o + T1p;
|
Chris@10
|
876 T2C = T1p - T1o;
|
Chris@10
|
877 }
|
Chris@10
|
878 T1r = FNMS(T1n, T1q, T1j * T1m);
|
Chris@10
|
879 T3F = FMA(T1j, T1q, T1n * T1m);
|
Chris@10
|
880 T2F = FMA(T2B, T2C, T2D * T2E);
|
Chris@10
|
881 T2X = FNMS(T2B, T2E, T2D * T2C);
|
Chris@10
|
882 }
|
Chris@10
|
883 {
|
Chris@10
|
884 E Tk, T12, T1h, T46;
|
Chris@10
|
885 Tk = FNMS(Tg, Tj, Ta * Td);
|
Chris@10
|
886 Tl = T3 - Tk;
|
Chris@10
|
887 T3n = Tk + T3;
|
Chris@10
|
888 T12 = TW - T11;
|
Chris@10
|
889 T1h = T19 - T1g;
|
Chris@10
|
890 T1i = T12 - T1h;
|
Chris@10
|
891 T2Q = T12 + T1h;
|
Chris@10
|
892 T46 = FMA(Ta, Tj, Tg * Td);
|
Chris@10
|
893 T47 = T45 - T46;
|
Chris@10
|
894 T50 = T45 + T46;
|
Chris@10
|
895 {
|
Chris@10
|
896 E T4Q, T4R, T2A, T2L;
|
Chris@10
|
897 T4Q = T2F + T2K;
|
Chris@10
|
898 T4R = T3R + T3S;
|
Chris@10
|
899 T4S = T4Q + T4R;
|
Chris@10
|
900 T5i = T4R - T4Q;
|
Chris@10
|
901 T2A = T2w - T2z;
|
Chris@10
|
902 T2L = T2F - T2K;
|
Chris@10
|
903 T2M = T2A - T2L;
|
Chris@10
|
904 T2T = T2L + T2A;
|
Chris@10
|
905 }
|
Chris@10
|
906 }
|
Chris@10
|
907 {
|
Chris@10
|
908 E T4G, T4H, T4J, T4K;
|
Chris@10
|
909 T4G = T3A + T3B;
|
Chris@10
|
910 T4H = T19 + T1g;
|
Chris@10
|
911 T4I = T4G + T4H;
|
Chris@10
|
912 T5f = T4G - T4H;
|
Chris@10
|
913 T4J = T3F + T3G;
|
Chris@10
|
914 T4K = T1Q + T1V;
|
Chris@10
|
915 T4L = T4J + T4K;
|
Chris@10
|
916 T5e = T4J - T4K;
|
Chris@10
|
917 }
|
Chris@10
|
918 T4N = T3L + T3M;
|
Chris@10
|
919 T4O = T2i + T2p;
|
Chris@10
|
920 T4P = T4N + T4O;
|
Chris@10
|
921 T5h = T4N - T4O;
|
Chris@10
|
922 {
|
Chris@10
|
923 E T29, T2q, T1F, T1W;
|
Chris@10
|
924 T29 = T23 - T28;
|
Chris@10
|
925 T2q = T2i - T2p;
|
Chris@10
|
926 T2r = T29 - T2q;
|
Chris@10
|
927 T2S = T29 + T2q;
|
Chris@10
|
928 T1F = T1r - T1E;
|
Chris@10
|
929 T1W = T1Q - T1V;
|
Chris@10
|
930 T1X = T1F + T1W;
|
Chris@10
|
931 T2P = T1W - T1F;
|
Chris@10
|
932 }
|
Chris@10
|
933 {
|
Chris@10
|
934 E T3C, T3D, T3N, T3O;
|
Chris@10
|
935 {
|
Chris@10
|
936 E T2Z, T30, T32, T35;
|
Chris@10
|
937 T2Z = T2X - T2Y;
|
Chris@10
|
938 T30 = T2w + T2z;
|
Chris@10
|
939 T31 = T2Z - T30;
|
Chris@10
|
940 T3u = T2Z + T30;
|
Chris@10
|
941 T32 = T23 + T28;
|
Chris@10
|
942 T35 = T33 + T34;
|
Chris@10
|
943 T36 = T32 + T35;
|
Chris@10
|
944 T3t = T32 - T35;
|
Chris@10
|
945 }
|
Chris@10
|
946 T3C = T3A - T3B;
|
Chris@10
|
947 T3D = T3a - T39;
|
Chris@10
|
948 T3E = T3C + T3D;
|
Chris@10
|
949 T4l = T3C - T3D;
|
Chris@10
|
950 {
|
Chris@10
|
951 E T3Q, T3T, T3d, T3g;
|
Chris@10
|
952 T3Q = T2X + T2Y;
|
Chris@10
|
953 T3T = T3R - T3S;
|
Chris@10
|
954 T3U = T3Q + T3T;
|
Chris@10
|
955 T4j = T3T - T3Q;
|
Chris@10
|
956 T3d = T1r + T1E;
|
Chris@10
|
957 T3g = T3e + T3f;
|
Chris@10
|
958 T3h = T3d + T3g;
|
Chris@10
|
959 T3r = T3d - T3g;
|
Chris@10
|
960 }
|
Chris@10
|
961 {
|
Chris@10
|
962 E T3H, T3I, T38, T3b;
|
Chris@10
|
963 T3H = T3F - T3G;
|
Chris@10
|
964 T3I = T3e - T3f;
|
Chris@10
|
965 T3J = T3H + T3I;
|
Chris@10
|
966 T4m = T3H - T3I;
|
Chris@10
|
967 T38 = TW + T11;
|
Chris@10
|
968 T3b = T39 + T3a;
|
Chris@10
|
969 T3c = T38 + T3b;
|
Chris@10
|
970 T3q = T38 - T3b;
|
Chris@10
|
971 }
|
Chris@10
|
972 T3N = T3L - T3M;
|
Chris@10
|
973 T3O = T34 - T33;
|
Chris@10
|
974 T3P = T3N + T3O;
|
Chris@10
|
975 T4i = T3N - T3O;
|
Chris@10
|
976 {
|
Chris@10
|
977 E TG, TR, T3k, T3l;
|
Chris@10
|
978 TG = FNMS(TC, TF, Tw * Tz);
|
Chris@10
|
979 TR = FNMS(TN, TQ, TJ * TM);
|
Chris@10
|
980 TS = TG - TR;
|
Chris@10
|
981 T51 = TG + TR;
|
Chris@10
|
982 T3k = FMA(TC, Tz, Tw * TF);
|
Chris@10
|
983 T3l = FMA(TN, TM, TJ * TQ);
|
Chris@10
|
984 T3m = T3k + T3l;
|
Chris@10
|
985 T48 = T3l - T3k;
|
Chris@10
|
986 }
|
Chris@10
|
987 }
|
Chris@10
|
988 }
|
Chris@10
|
989 {
|
Chris@10
|
990 E T3W, T3Y, TT, T2O, T3x, T3y, T3X, T3z;
|
Chris@10
|
991 {
|
Chris@10
|
992 E T3K, T3V, T1Y, T2N;
|
Chris@10
|
993 T3K = T3E - T3J;
|
Chris@10
|
994 T3V = T3P - T3U;
|
Chris@10
|
995 T3W = FMA(KP475528258, T3K, KP293892626 * T3V);
|
Chris@10
|
996 T3Y = FNMS(KP293892626, T3K, KP475528258 * T3V);
|
Chris@10
|
997 TT = Tl - TS;
|
Chris@10
|
998 T1Y = T1i + T1X;
|
Chris@10
|
999 T2N = T2r + T2M;
|
Chris@10
|
1000 T2O = T1Y + T2N;
|
Chris@10
|
1001 T3x = KP279508497 * (T1Y - T2N);
|
Chris@10
|
1002 T3y = FNMS(KP125000000, T2O, KP500000000 * TT);
|
Chris@10
|
1003 }
|
Chris@10
|
1004 Ip[WS(rs, 5)] = KP500000000 * (TT + T2O);
|
Chris@10
|
1005 T3X = T3x - T3y;
|
Chris@10
|
1006 Im[WS(rs, 2)] = T3X - T3Y;
|
Chris@10
|
1007 Im[WS(rs, 6)] = T3X + T3Y;
|
Chris@10
|
1008 T3z = T3x + T3y;
|
Chris@10
|
1009 Ip[WS(rs, 1)] = T3z - T3W;
|
Chris@10
|
1010 Ip[WS(rs, 9)] = T3z + T3W;
|
Chris@10
|
1011 }
|
Chris@10
|
1012 {
|
Chris@10
|
1013 E T41, T4d, T49, T4a, T44, T4b, T4e, T4c;
|
Chris@10
|
1014 {
|
Chris@10
|
1015 E T3Z, T40, T42, T43;
|
Chris@10
|
1016 T3Z = T1i - T1X;
|
Chris@10
|
1017 T40 = T2r - T2M;
|
Chris@10
|
1018 T41 = FMA(KP475528258, T3Z, KP293892626 * T40);
|
Chris@10
|
1019 T4d = FNMS(KP293892626, T3Z, KP475528258 * T40);
|
Chris@10
|
1020 T49 = T47 + T48;
|
Chris@10
|
1021 T42 = T3E + T3J;
|
Chris@10
|
1022 T43 = T3P + T3U;
|
Chris@10
|
1023 T4a = T42 + T43;
|
Chris@10
|
1024 T44 = KP279508497 * (T42 - T43);
|
Chris@10
|
1025 T4b = FNMS(KP125000000, T4a, KP500000000 * T49);
|
Chris@10
|
1026 }
|
Chris@10
|
1027 Rp[WS(rs, 5)] = KP500000000 * (T49 + T4a);
|
Chris@10
|
1028 T4e = T4b - T44;
|
Chris@10
|
1029 Rm[WS(rs, 6)] = T4d + T4e;
|
Chris@10
|
1030 Rm[WS(rs, 2)] = T4e - T4d;
|
Chris@10
|
1031 T4c = T44 + T4b;
|
Chris@10
|
1032 Rp[WS(rs, 1)] = T41 + T4c;
|
Chris@10
|
1033 Rp[WS(rs, 9)] = T4c - T41;
|
Chris@10
|
1034 }
|
Chris@10
|
1035 {
|
Chris@10
|
1036 E T4o, T4q, T2W, T2V, T4f, T4g, T4p, T4h;
|
Chris@10
|
1037 {
|
Chris@10
|
1038 E T4k, T4n, T2R, T2U;
|
Chris@10
|
1039 T4k = T4i - T4j;
|
Chris@10
|
1040 T4n = T4l - T4m;
|
Chris@10
|
1041 T4o = FNMS(KP293892626, T4n, KP475528258 * T4k);
|
Chris@10
|
1042 T4q = FMA(KP475528258, T4n, KP293892626 * T4k);
|
Chris@10
|
1043 T2W = TS + Tl;
|
Chris@10
|
1044 T2R = T2P - T2Q;
|
Chris@10
|
1045 T2U = T2S + T2T;
|
Chris@10
|
1046 T2V = T2R - T2U;
|
Chris@10
|
1047 T4f = FMA(KP500000000, T2W, KP125000000 * T2V);
|
Chris@10
|
1048 T4g = KP279508497 * (T2R + T2U);
|
Chris@10
|
1049 }
|
Chris@10
|
1050 Im[WS(rs, 4)] = KP500000000 * (T2V - T2W);
|
Chris@10
|
1051 T4p = T4g - T4f;
|
Chris@10
|
1052 Im[0] = T4p - T4q;
|
Chris@10
|
1053 Im[WS(rs, 8)] = T4p + T4q;
|
Chris@10
|
1054 T4h = T4f + T4g;
|
Chris@10
|
1055 Ip[WS(rs, 3)] = T4h - T4o;
|
Chris@10
|
1056 Ip[WS(rs, 7)] = T4h + T4o;
|
Chris@10
|
1057 }
|
Chris@10
|
1058 {
|
Chris@10
|
1059 E T4t, T4B, T4u, T4x, T4y, T4z, T4C, T4A;
|
Chris@10
|
1060 {
|
Chris@10
|
1061 E T4r, T4s, T4v, T4w;
|
Chris@10
|
1062 T4r = T2S - T2T;
|
Chris@10
|
1063 T4s = T2Q + T2P;
|
Chris@10
|
1064 T4t = FNMS(KP293892626, T4s, KP475528258 * T4r);
|
Chris@10
|
1065 T4B = FMA(KP475528258, T4s, KP293892626 * T4r);
|
Chris@10
|
1066 T4u = T47 - T48;
|
Chris@10
|
1067 T4v = T4l + T4m;
|
Chris@10
|
1068 T4w = T4i + T4j;
|
Chris@10
|
1069 T4x = T4v + T4w;
|
Chris@10
|
1070 T4y = FNMS(KP125000000, T4x, KP500000000 * T4u);
|
Chris@10
|
1071 T4z = KP279508497 * (T4v - T4w);
|
Chris@10
|
1072 }
|
Chris@10
|
1073 Rm[WS(rs, 4)] = KP500000000 * (T4u + T4x);
|
Chris@10
|
1074 T4C = T4z + T4y;
|
Chris@10
|
1075 Rm[WS(rs, 8)] = T4B + T4C;
|
Chris@10
|
1076 Rm[0] = T4C - T4B;
|
Chris@10
|
1077 T4A = T4y - T4z;
|
Chris@10
|
1078 Rp[WS(rs, 3)] = T4t + T4A;
|
Chris@10
|
1079 Rp[WS(rs, 7)] = T4A - T4t;
|
Chris@10
|
1080 }
|
Chris@10
|
1081 {
|
Chris@10
|
1082 E T5k, T5m, T3o, T3j, T5b, T5c, T5l, T5d;
|
Chris@10
|
1083 {
|
Chris@10
|
1084 E T5g, T5j, T37, T3i;
|
Chris@10
|
1085 T5g = T5e - T5f;
|
Chris@10
|
1086 T5j = T5h - T5i;
|
Chris@10
|
1087 T5k = FNMS(KP293892626, T5j, KP475528258 * T5g);
|
Chris@10
|
1088 T5m = FMA(KP293892626, T5g, KP475528258 * T5j);
|
Chris@10
|
1089 T3o = T3m + T3n;
|
Chris@10
|
1090 T37 = T31 - T36;
|
Chris@10
|
1091 T3i = T3c + T3h;
|
Chris@10
|
1092 T3j = T37 - T3i;
|
Chris@10
|
1093 T5b = FMA(KP500000000, T3o, KP125000000 * T3j);
|
Chris@10
|
1094 T5c = KP279508497 * (T3i + T37);
|
Chris@10
|
1095 }
|
Chris@10
|
1096 Im[WS(rs, 9)] = KP500000000 * (T3j - T3o);
|
Chris@10
|
1097 T5l = T5b - T5c;
|
Chris@10
|
1098 Ip[WS(rs, 2)] = T5l + T5m;
|
Chris@10
|
1099 Im[WS(rs, 1)] = T5m - T5l;
|
Chris@10
|
1100 T5d = T5b + T5c;
|
Chris@10
|
1101 Ip[WS(rs, 6)] = T5d + T5k;
|
Chris@10
|
1102 Im[WS(rs, 5)] = T5k - T5d;
|
Chris@10
|
1103 }
|
Chris@10
|
1104 {
|
Chris@10
|
1105 E T5w, T5x, T5n, T5q, T5r, T5s, T5y, T5t;
|
Chris@10
|
1106 {
|
Chris@10
|
1107 E T5u, T5v, T5o, T5p;
|
Chris@10
|
1108 T5u = T36 + T31;
|
Chris@10
|
1109 T5v = T3c - T3h;
|
Chris@10
|
1110 T5w = FNMS(KP293892626, T5v, KP475528258 * T5u);
|
Chris@10
|
1111 T5x = FMA(KP475528258, T5v, KP293892626 * T5u);
|
Chris@10
|
1112 T5n = T50 - T51;
|
Chris@10
|
1113 T5o = T5f + T5e;
|
Chris@10
|
1114 T5p = T5h + T5i;
|
Chris@10
|
1115 T5q = T5o + T5p;
|
Chris@10
|
1116 T5r = FNMS(KP125000000, T5q, KP500000000 * T5n);
|
Chris@10
|
1117 T5s = KP279508497 * (T5o - T5p);
|
Chris@10
|
1118 }
|
Chris@10
|
1119 Rm[WS(rs, 9)] = KP500000000 * (T5n + T5q);
|
Chris@10
|
1120 T5y = T5s + T5r;
|
Chris@10
|
1121 Rp[WS(rs, 6)] = T5x + T5y;
|
Chris@10
|
1122 Rm[WS(rs, 5)] = T5y - T5x;
|
Chris@10
|
1123 T5t = T5r - T5s;
|
Chris@10
|
1124 Rp[WS(rs, 2)] = T5t - T5w;
|
Chris@10
|
1125 Rm[WS(rs, 1)] = T5w + T5t;
|
Chris@10
|
1126 }
|
Chris@10
|
1127 {
|
Chris@10
|
1128 E T4U, T4W, T3p, T3w, T4D, T4E, T4V, T4F;
|
Chris@10
|
1129 {
|
Chris@10
|
1130 E T4M, T4T, T3s, T3v;
|
Chris@10
|
1131 T4M = T4I - T4L;
|
Chris@10
|
1132 T4T = T4P - T4S;
|
Chris@10
|
1133 T4U = FNMS(KP475528258, T4T, KP293892626 * T4M);
|
Chris@10
|
1134 T4W = FMA(KP475528258, T4M, KP293892626 * T4T);
|
Chris@10
|
1135 T3p = T3n - T3m;
|
Chris@10
|
1136 T3s = T3q + T3r;
|
Chris@10
|
1137 T3v = T3t + T3u;
|
Chris@10
|
1138 T3w = T3s + T3v;
|
Chris@10
|
1139 T4D = FNMS(KP125000000, T3w, KP500000000 * T3p);
|
Chris@10
|
1140 T4E = KP279508497 * (T3s - T3v);
|
Chris@10
|
1141 }
|
Chris@10
|
1142 Ip[0] = KP500000000 * (T3p + T3w);
|
Chris@10
|
1143 T4V = T4E + T4D;
|
Chris@10
|
1144 Ip[WS(rs, 4)] = T4V + T4W;
|
Chris@10
|
1145 Im[WS(rs, 3)] = T4W - T4V;
|
Chris@10
|
1146 T4F = T4D - T4E;
|
Chris@10
|
1147 Ip[WS(rs, 8)] = T4F + T4U;
|
Chris@10
|
1148 Im[WS(rs, 7)] = T4U - T4F;
|
Chris@10
|
1149 }
|
Chris@10
|
1150 {
|
Chris@10
|
1151 E T58, T59, T52, T53, T4Z, T54, T5a, T55;
|
Chris@10
|
1152 {
|
Chris@10
|
1153 E T56, T57, T4X, T4Y;
|
Chris@10
|
1154 T56 = T3q - T3r;
|
Chris@10
|
1155 T57 = T3t - T3u;
|
Chris@10
|
1156 T58 = FMA(KP475528258, T56, KP293892626 * T57);
|
Chris@10
|
1157 T59 = FNMS(KP293892626, T56, KP475528258 * T57);
|
Chris@10
|
1158 T52 = T50 + T51;
|
Chris@10
|
1159 T4X = T4I + T4L;
|
Chris@10
|
1160 T4Y = T4P + T4S;
|
Chris@10
|
1161 T53 = T4X + T4Y;
|
Chris@10
|
1162 T4Z = KP279508497 * (T4X - T4Y);
|
Chris@10
|
1163 T54 = FNMS(KP125000000, T53, KP500000000 * T52);
|
Chris@10
|
1164 }
|
Chris@10
|
1165 Rp[0] = KP500000000 * (T52 + T53);
|
Chris@10
|
1166 T5a = T54 - T4Z;
|
Chris@10
|
1167 Rp[WS(rs, 8)] = T59 + T5a;
|
Chris@10
|
1168 Rm[WS(rs, 7)] = T5a - T59;
|
Chris@10
|
1169 T55 = T4Z + T54;
|
Chris@10
|
1170 Rp[WS(rs, 4)] = T55 - T58;
|
Chris@10
|
1171 Rm[WS(rs, 3)] = T58 + T55;
|
Chris@10
|
1172 }
|
Chris@10
|
1173 }
|
Chris@10
|
1174 }
|
Chris@10
|
1175 }
|
Chris@10
|
1176 }
|
Chris@10
|
1177
|
Chris@10
|
1178 static const tw_instr twinstr[] = {
|
Chris@10
|
1179 {TW_CEXP, 1, 1},
|
Chris@10
|
1180 {TW_CEXP, 1, 3},
|
Chris@10
|
1181 {TW_CEXP, 1, 9},
|
Chris@10
|
1182 {TW_CEXP, 1, 19},
|
Chris@10
|
1183 {TW_NEXT, 1, 0}
|
Chris@10
|
1184 };
|
Chris@10
|
1185
|
Chris@10
|
1186 static const hc2c_desc desc = { 20, "hc2cfdft2_20", twinstr, &GENUS, {244, 108, 72, 0} };
|
Chris@10
|
1187
|
Chris@10
|
1188 void X(codelet_hc2cfdft2_20) (planner *p) {
|
Chris@10
|
1189 X(khc2c_register) (p, hc2cfdft2_20, &desc, HC2C_VIA_DFT);
|
Chris@10
|
1190 }
|
Chris@10
|
1191 #endif /* HAVE_FMA */
|