cannam@167
|
1 /*
|
cannam@167
|
2 * Copyright (c) 2003, 2007-14 Matteo Frigo
|
cannam@167
|
3 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
|
cannam@167
|
4 *
|
cannam@167
|
5 * This program is free software; you can redistribute it and/or modify
|
cannam@167
|
6 * it under the terms of the GNU General Public License as published by
|
cannam@167
|
7 * the Free Software Foundation; either version 2 of the License, or
|
cannam@167
|
8 * (at your option) any later version.
|
cannam@167
|
9 *
|
cannam@167
|
10 * This program is distributed in the hope that it will be useful,
|
cannam@167
|
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
cannam@167
|
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
cannam@167
|
13 * GNU General Public License for more details.
|
cannam@167
|
14 *
|
cannam@167
|
15 * You should have received a copy of the GNU General Public License
|
cannam@167
|
16 * along with this program; if not, write to the Free Software
|
cannam@167
|
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
cannam@167
|
18 *
|
cannam@167
|
19 */
|
cannam@167
|
20
|
cannam@167
|
21 /* This file was automatically generated --- DO NOT EDIT */
|
cannam@167
|
22 /* Generated on Thu May 24 08:07:17 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_hc2cdft.native -fma -compact -variables 4 -pipeline-latency 4 -twiddle-log3 -precompute-twiddles -n 32 -dit -name hc2cfdft2_32 -include rdft/scalar/hc2cf.h */
|
cannam@167
|
29
|
cannam@167
|
30 /*
|
cannam@167
|
31 * This function contains 552 FP additions, 414 FP multiplications,
|
cannam@167
|
32 * (or, 300 additions, 162 multiplications, 252 fused multiply/add),
|
cannam@167
|
33 * 175 stack variables, 8 constants, and 128 memory accesses
|
cannam@167
|
34 */
|
cannam@167
|
35 #include "rdft/scalar/hc2cf.h"
|
cannam@167
|
36
|
cannam@167
|
37 static void hc2cfdft2_32(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms)
|
cannam@167
|
38 {
|
cannam@167
|
39 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
cannam@167
|
40 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
cannam@167
|
41 DK(KP198912367, +0.198912367379658006911597622644676228597850501);
|
cannam@167
|
42 DK(KP668178637, +0.668178637919298919997757686523080761552472251);
|
cannam@167
|
43 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
cannam@167
|
44 DK(KP414213562, +0.414213562373095048801688724209698078569671875);
|
cannam@167
|
45 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
cannam@167
|
46 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
|
cannam@167
|
47 {
|
cannam@167
|
48 INT m;
|
cannam@167
|
49 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(128, rs)) {
|
cannam@167
|
50 E T1, Th, T2, T5, Ti, Tl, T4, T6, T1a, Tc, T1c, Tk, Tz, T2H, T2v;
|
cannam@167
|
51 E T1u, Tm, Ts, T15, T2W, TZ, T2l, T2q, T2R, TR, TL, T3B, T3S, T3F, T3V;
|
cannam@167
|
52 E T4E, T4Y, T4I, T51, TF, T40, T44, T2A, T4M, T4Q, T1A, T3s, T3w, T2M, T4l;
|
cannam@167
|
53 E T4p, T1g, T1H, T1F, T1d, T1h, T1O, T1n, T1I, T28, T34, T32, T25, T29, T3b;
|
cannam@167
|
54 E T2f, T35;
|
cannam@167
|
55 {
|
cannam@167
|
56 E Tj, TY, TK, Tr, T14, TQ, T1b, T24, TE, T1z;
|
cannam@167
|
57 {
|
cannam@167
|
58 E T3, T1t, Tb, Ty;
|
cannam@167
|
59 T1 = W[0];
|
cannam@167
|
60 Th = W[4];
|
cannam@167
|
61 T2 = W[2];
|
cannam@167
|
62 T5 = W[3];
|
cannam@167
|
63 T3 = T1 * T2;
|
cannam@167
|
64 T1t = T2 * Th;
|
cannam@167
|
65 Tb = T1 * T5;
|
cannam@167
|
66 Ty = T1 * Th;
|
cannam@167
|
67 Ti = W[6];
|
cannam@167
|
68 Tj = Th * Ti;
|
cannam@167
|
69 TY = T2 * Ti;
|
cannam@167
|
70 TK = T1 * Ti;
|
cannam@167
|
71 Tl = W[7];
|
cannam@167
|
72 Tr = Th * Tl;
|
cannam@167
|
73 T14 = T2 * Tl;
|
cannam@167
|
74 TQ = T1 * Tl;
|
cannam@167
|
75 T4 = W[1];
|
cannam@167
|
76 T6 = FMA(T4, T5, T3);
|
cannam@167
|
77 T1a = FNMS(T4, T5, T3);
|
cannam@167
|
78 T1b = T1a * Th;
|
cannam@167
|
79 T24 = T6 * Th;
|
cannam@167
|
80 Tc = FNMS(T4, T2, Tb);
|
cannam@167
|
81 T1c = FMA(T4, T2, Tb);
|
cannam@167
|
82 Tk = W[5];
|
cannam@167
|
83 TE = T1 * Tk;
|
cannam@167
|
84 T1z = T2 * Tk;
|
cannam@167
|
85 Tz = FNMS(T4, Tk, Ty);
|
cannam@167
|
86 T2H = FMA(T4, Tk, Ty);
|
cannam@167
|
87 T2v = FNMS(T5, Tk, T1t);
|
cannam@167
|
88 T1u = FMA(T5, Tk, T1t);
|
cannam@167
|
89 }
|
cannam@167
|
90 Tm = FMA(Tk, Tl, Tj);
|
cannam@167
|
91 Ts = FNMS(Tk, Ti, Tr);
|
cannam@167
|
92 T15 = FMA(T5, Ti, T14);
|
cannam@167
|
93 T2W = FNMS(T5, Ti, T14);
|
cannam@167
|
94 TZ = FNMS(T5, Tl, TY);
|
cannam@167
|
95 T2l = FNMS(T4, Tl, TK);
|
cannam@167
|
96 T2q = FMA(T4, Ti, TQ);
|
cannam@167
|
97 T2R = FMA(T5, Tl, TY);
|
cannam@167
|
98 TR = FNMS(T4, Ti, TQ);
|
cannam@167
|
99 TL = FMA(T4, Tl, TK);
|
cannam@167
|
100 {
|
cannam@167
|
101 E T3A, T3E, T4k, T4o;
|
cannam@167
|
102 T3A = T6 * Ti;
|
cannam@167
|
103 T3B = FNMS(Tc, Tl, T3A);
|
cannam@167
|
104 T3S = FMA(Tc, Tl, T3A);
|
cannam@167
|
105 T3E = T6 * Tl;
|
cannam@167
|
106 T3F = FMA(Tc, Ti, T3E);
|
cannam@167
|
107 T3V = FNMS(Tc, Ti, T3E);
|
cannam@167
|
108 {
|
cannam@167
|
109 E T4D, T4H, T3Z, T43;
|
cannam@167
|
110 T4D = T1a * Ti;
|
cannam@167
|
111 T4E = FNMS(T1c, Tl, T4D);
|
cannam@167
|
112 T4Y = FMA(T1c, Tl, T4D);
|
cannam@167
|
113 T4H = T1a * Tl;
|
cannam@167
|
114 T4I = FMA(T1c, Ti, T4H);
|
cannam@167
|
115 T51 = FNMS(T1c, Ti, T4H);
|
cannam@167
|
116 T3Z = Tz * Ti;
|
cannam@167
|
117 T43 = Tz * Tl;
|
cannam@167
|
118 TF = FMA(T4, Th, TE);
|
cannam@167
|
119 T40 = FMA(TF, Tl, T3Z);
|
cannam@167
|
120 T44 = FNMS(TF, Ti, T43);
|
cannam@167
|
121 }
|
cannam@167
|
122 {
|
cannam@167
|
123 E T4L, T4P, T3r, T3v;
|
cannam@167
|
124 T4L = T2v * Ti;
|
cannam@167
|
125 T4P = T2v * Tl;
|
cannam@167
|
126 T2A = FMA(T5, Th, T1z);
|
cannam@167
|
127 T4M = FMA(T2A, Tl, T4L);
|
cannam@167
|
128 T4Q = FNMS(T2A, Ti, T4P);
|
cannam@167
|
129 T3r = T1u * Ti;
|
cannam@167
|
130 T3v = T1u * Tl;
|
cannam@167
|
131 T1A = FNMS(T5, Th, T1z);
|
cannam@167
|
132 T3s = FMA(T1A, Tl, T3r);
|
cannam@167
|
133 T3w = FNMS(T1A, Ti, T3v);
|
cannam@167
|
134 }
|
cannam@167
|
135 T4k = T2H * Ti;
|
cannam@167
|
136 T4o = T2H * Tl;
|
cannam@167
|
137 T2M = FNMS(T4, Th, TE);
|
cannam@167
|
138 T4l = FMA(T2M, Tl, T4k);
|
cannam@167
|
139 T4p = FNMS(T2M, Ti, T4o);
|
cannam@167
|
140 {
|
cannam@167
|
141 E T1G, T1N, T1e, T1m, T1f;
|
cannam@167
|
142 T1f = T1a * Tk;
|
cannam@167
|
143 T1g = FMA(T1c, Th, T1f);
|
cannam@167
|
144 T1H = FNMS(T1c, Th, T1f);
|
cannam@167
|
145 T1F = FMA(T1c, Tk, T1b);
|
cannam@167
|
146 T1G = T1F * Ti;
|
cannam@167
|
147 T1N = T1F * Tl;
|
cannam@167
|
148 T1d = FNMS(T1c, Tk, T1b);
|
cannam@167
|
149 T1e = T1d * Ti;
|
cannam@167
|
150 T1m = T1d * Tl;
|
cannam@167
|
151 T1h = FMA(T1g, Tl, T1e);
|
cannam@167
|
152 T1O = FNMS(T1H, Ti, T1N);
|
cannam@167
|
153 T1n = FNMS(T1g, Ti, T1m);
|
cannam@167
|
154 T1I = FMA(T1H, Tl, T1G);
|
cannam@167
|
155 }
|
cannam@167
|
156 {
|
cannam@167
|
157 E T33, T3a, T26, T2e, T27;
|
cannam@167
|
158 T27 = T6 * Tk;
|
cannam@167
|
159 T28 = FNMS(Tc, Th, T27);
|
cannam@167
|
160 T34 = FMA(Tc, Th, T27);
|
cannam@167
|
161 T32 = FNMS(Tc, Tk, T24);
|
cannam@167
|
162 T33 = T32 * Ti;
|
cannam@167
|
163 T3a = T32 * Tl;
|
cannam@167
|
164 T25 = FMA(Tc, Tk, T24);
|
cannam@167
|
165 T26 = T25 * Ti;
|
cannam@167
|
166 T2e = T25 * Tl;
|
cannam@167
|
167 T29 = FMA(T28, Tl, T26);
|
cannam@167
|
168 T3b = FNMS(T34, Ti, T3a);
|
cannam@167
|
169 T2f = FNMS(T28, Ti, T2e);
|
cannam@167
|
170 T35 = FMA(T34, Tl, T33);
|
cannam@167
|
171 }
|
cannam@167
|
172 }
|
cannam@167
|
173 }
|
cannam@167
|
174 {
|
cannam@167
|
175 E T3j, T7Z, T5b, T93, T4d, T8J, T6B, T8V, T1T, T8l, T6e, T8r, T54, T8C, T5O;
|
cannam@167
|
176 E T8i, T31, T94, T6w, T8K, T3Y, T8U, T5g, T80, T1s, T8h, T69, T8B, T4T, T8q;
|
cannam@167
|
177 E T5J, T8k, Tx, T8a, T5y, T8d, T4s, T8E, T5Y, T8v, T2k, T82, T5m, T83, T3z;
|
cannam@167
|
178 E T8X, T6l, T8O, T2F, T86, T5r, T85, T3M, T8Y, T6q, T8R, TW, T8e, T5D, T8b;
|
cannam@167
|
179 E T4B, T8F, T63, T8y;
|
cannam@167
|
180 {
|
cannam@167
|
181 E T3i, T4b, T38, T39, T45, T4a, T6z, T58, T3e, T42, T6x, T59, T3f, T5a;
|
cannam@167
|
182 {
|
cannam@167
|
183 E T3g, T3h, T36, T37;
|
cannam@167
|
184 T3g = Ip[0];
|
cannam@167
|
185 T3h = Im[0];
|
cannam@167
|
186 T3i = T3g - T3h;
|
cannam@167
|
187 T4b = T3g + T3h;
|
cannam@167
|
188 T36 = Ip[WS(rs, 8)];
|
cannam@167
|
189 T37 = Im[WS(rs, 8)];
|
cannam@167
|
190 T38 = T36 - T37;
|
cannam@167
|
191 T39 = T35 * T38;
|
cannam@167
|
192 T45 = T36 + T37;
|
cannam@167
|
193 }
|
cannam@167
|
194 {
|
cannam@167
|
195 E T47, T48, T49, T41, T3c, T3d;
|
cannam@167
|
196 T47 = Rm[0];
|
cannam@167
|
197 T48 = Rp[0];
|
cannam@167
|
198 T49 = T47 - T48;
|
cannam@167
|
199 T4a = T1 * T49;
|
cannam@167
|
200 T6z = T4 * T49;
|
cannam@167
|
201 T58 = T48 + T47;
|
cannam@167
|
202 T3c = Rp[WS(rs, 8)];
|
cannam@167
|
203 T3d = Rm[WS(rs, 8)];
|
cannam@167
|
204 T3e = T3c + T3d;
|
cannam@167
|
205 T41 = T3d - T3c;
|
cannam@167
|
206 T42 = T40 * T41;
|
cannam@167
|
207 T6x = T44 * T41;
|
cannam@167
|
208 T59 = T35 * T3e;
|
cannam@167
|
209 }
|
cannam@167
|
210 T3f = FNMS(T3b, T3e, T39);
|
cannam@167
|
211 T3j = T3f + T3i;
|
cannam@167
|
212 T7Z = T3i - T3f;
|
cannam@167
|
213 T5a = FMA(T3b, T38, T59);
|
cannam@167
|
214 T5b = T58 + T5a;
|
cannam@167
|
215 T93 = T58 - T5a;
|
cannam@167
|
216 {
|
cannam@167
|
217 E T46, T4c, T6y, T6A;
|
cannam@167
|
218 T46 = FNMS(T44, T45, T42);
|
cannam@167
|
219 T4c = FNMS(T4, T4b, T4a);
|
cannam@167
|
220 T4d = T46 + T4c;
|
cannam@167
|
221 T8J = T4c - T46;
|
cannam@167
|
222 T6y = FMA(T40, T45, T6x);
|
cannam@167
|
223 T6A = FMA(T1, T4b, T6z);
|
cannam@167
|
224 T6B = T6y + T6A;
|
cannam@167
|
225 T8V = T6A - T6y;
|
cannam@167
|
226 }
|
cannam@167
|
227 }
|
cannam@167
|
228 {
|
cannam@167
|
229 E T1x, T4W, T1y, T6a, T1D, T4U, T4V, T5K, T1L, T52, T1M, T6c, T1R, T4Z, T50;
|
cannam@167
|
230 E T5M;
|
cannam@167
|
231 {
|
cannam@167
|
232 E T1v, T1w, T1B, T1C;
|
cannam@167
|
233 T1v = Ip[WS(rs, 3)];
|
cannam@167
|
234 T1w = Im[WS(rs, 3)];
|
cannam@167
|
235 T1x = T1v - T1w;
|
cannam@167
|
236 T4W = T1v + T1w;
|
cannam@167
|
237 T1y = T1u * T1x;
|
cannam@167
|
238 T6a = T25 * T4W;
|
cannam@167
|
239 T1B = Rp[WS(rs, 3)];
|
cannam@167
|
240 T1C = Rm[WS(rs, 3)];
|
cannam@167
|
241 T1D = T1B + T1C;
|
cannam@167
|
242 T4U = T1B - T1C;
|
cannam@167
|
243 T4V = T25 * T4U;
|
cannam@167
|
244 T5K = T1u * T1D;
|
cannam@167
|
245 }
|
cannam@167
|
246 {
|
cannam@167
|
247 E T1J, T1K, T1P, T1Q;
|
cannam@167
|
248 T1J = Ip[WS(rs, 11)];
|
cannam@167
|
249 T1K = Im[WS(rs, 11)];
|
cannam@167
|
250 T1L = T1J - T1K;
|
cannam@167
|
251 T52 = T1J + T1K;
|
cannam@167
|
252 T1M = T1I * T1L;
|
cannam@167
|
253 T6c = T4Y * T52;
|
cannam@167
|
254 T1P = Rp[WS(rs, 11)];
|
cannam@167
|
255 T1Q = Rm[WS(rs, 11)];
|
cannam@167
|
256 T1R = T1P + T1Q;
|
cannam@167
|
257 T4Z = T1P - T1Q;
|
cannam@167
|
258 T50 = T4Y * T4Z;
|
cannam@167
|
259 T5M = T1I * T1R;
|
cannam@167
|
260 }
|
cannam@167
|
261 {
|
cannam@167
|
262 E T1E, T1S, T6b, T6d;
|
cannam@167
|
263 T1E = FNMS(T1A, T1D, T1y);
|
cannam@167
|
264 T1S = FNMS(T1O, T1R, T1M);
|
cannam@167
|
265 T1T = T1E + T1S;
|
cannam@167
|
266 T8l = T1E - T1S;
|
cannam@167
|
267 T6b = FNMS(T28, T4U, T6a);
|
cannam@167
|
268 T6d = FNMS(T51, T4Z, T6c);
|
cannam@167
|
269 T6e = T6b + T6d;
|
cannam@167
|
270 T8r = T6d - T6b;
|
cannam@167
|
271 }
|
cannam@167
|
272 {
|
cannam@167
|
273 E T4X, T53, T5L, T5N;
|
cannam@167
|
274 T4X = FMA(T28, T4W, T4V);
|
cannam@167
|
275 T53 = FMA(T51, T52, T50);
|
cannam@167
|
276 T54 = T4X + T53;
|
cannam@167
|
277 T8C = T53 - T4X;
|
cannam@167
|
278 T5L = FMA(T1A, T1x, T5K);
|
cannam@167
|
279 T5N = FMA(T1O, T1L, T5M);
|
cannam@167
|
280 T5O = T5L + T5N;
|
cannam@167
|
281 T8i = T5L - T5N;
|
cannam@167
|
282 }
|
cannam@167
|
283 }
|
cannam@167
|
284 {
|
cannam@167
|
285 E T2K, T2L, T3Q, T2P, T3P, T6s, T5c, T2U, T2V, T3W, T2Z, T3U, T6u, T5e;
|
cannam@167
|
286 {
|
cannam@167
|
287 E T2I, T2J, T3O, T2N, T2O;
|
cannam@167
|
288 T2I = Ip[WS(rs, 4)];
|
cannam@167
|
289 T2J = Im[WS(rs, 4)];
|
cannam@167
|
290 T2K = T2I - T2J;
|
cannam@167
|
291 T2L = T2H * T2K;
|
cannam@167
|
292 T3Q = T2I + T2J;
|
cannam@167
|
293 T2N = Rp[WS(rs, 4)];
|
cannam@167
|
294 T2O = Rm[WS(rs, 4)];
|
cannam@167
|
295 T2P = T2N + T2O;
|
cannam@167
|
296 T3O = T2O - T2N;
|
cannam@167
|
297 T3P = Th * T3O;
|
cannam@167
|
298 T6s = Tk * T3O;
|
cannam@167
|
299 T5c = T2H * T2P;
|
cannam@167
|
300 }
|
cannam@167
|
301 {
|
cannam@167
|
302 E T2S, T2T, T3T, T2X, T2Y;
|
cannam@167
|
303 T2S = Ip[WS(rs, 12)];
|
cannam@167
|
304 T2T = Im[WS(rs, 12)];
|
cannam@167
|
305 T2U = T2S - T2T;
|
cannam@167
|
306 T2V = T2R * T2U;
|
cannam@167
|
307 T3W = T2S + T2T;
|
cannam@167
|
308 T2X = Rp[WS(rs, 12)];
|
cannam@167
|
309 T2Y = Rm[WS(rs, 12)];
|
cannam@167
|
310 T2Z = T2X + T2Y;
|
cannam@167
|
311 T3T = T2Y - T2X;
|
cannam@167
|
312 T3U = T3S * T3T;
|
cannam@167
|
313 T6u = T3V * T3T;
|
cannam@167
|
314 T5e = T2R * T2Z;
|
cannam@167
|
315 }
|
cannam@167
|
316 {
|
cannam@167
|
317 E T2Q, T30, T6t, T6v;
|
cannam@167
|
318 T2Q = FNMS(T2M, T2P, T2L);
|
cannam@167
|
319 T30 = FNMS(T2W, T2Z, T2V);
|
cannam@167
|
320 T31 = T2Q + T30;
|
cannam@167
|
321 T94 = T2Q - T30;
|
cannam@167
|
322 T6t = FMA(Th, T3Q, T6s);
|
cannam@167
|
323 T6v = FMA(T3S, T3W, T6u);
|
cannam@167
|
324 T6w = T6t + T6v;
|
cannam@167
|
325 T8K = T6t - T6v;
|
cannam@167
|
326 }
|
cannam@167
|
327 {
|
cannam@167
|
328 E T3R, T3X, T5d, T5f;
|
cannam@167
|
329 T3R = FNMS(Tk, T3Q, T3P);
|
cannam@167
|
330 T3X = FNMS(T3V, T3W, T3U);
|
cannam@167
|
331 T3Y = T3R + T3X;
|
cannam@167
|
332 T8U = T3R - T3X;
|
cannam@167
|
333 T5d = FMA(T2M, T2K, T5c);
|
cannam@167
|
334 T5f = FMA(T2W, T2U, T5e);
|
cannam@167
|
335 T5g = T5d + T5f;
|
cannam@167
|
336 T80 = T5d - T5f;
|
cannam@167
|
337 }
|
cannam@167
|
338 }
|
cannam@167
|
339 {
|
cannam@167
|
340 E T12, T4J, T13, T65, T18, T4F, T4G, T5F, T1k, T4R, T1l, T67, T1q, T4N, T4O;
|
cannam@167
|
341 E T5H;
|
cannam@167
|
342 {
|
cannam@167
|
343 E T10, T11, T16, T17;
|
cannam@167
|
344 T10 = Ip[WS(rs, 15)];
|
cannam@167
|
345 T11 = Im[WS(rs, 15)];
|
cannam@167
|
346 T12 = T10 - T11;
|
cannam@167
|
347 T4J = T10 + T11;
|
cannam@167
|
348 T13 = TZ * T12;
|
cannam@167
|
349 T65 = T4E * T4J;
|
cannam@167
|
350 T16 = Rp[WS(rs, 15)];
|
cannam@167
|
351 T17 = Rm[WS(rs, 15)];
|
cannam@167
|
352 T18 = T16 + T17;
|
cannam@167
|
353 T4F = T16 - T17;
|
cannam@167
|
354 T4G = T4E * T4F;
|
cannam@167
|
355 T5F = TZ * T18;
|
cannam@167
|
356 }
|
cannam@167
|
357 {
|
cannam@167
|
358 E T1i, T1j, T1o, T1p;
|
cannam@167
|
359 T1i = Ip[WS(rs, 7)];
|
cannam@167
|
360 T1j = Im[WS(rs, 7)];
|
cannam@167
|
361 T1k = T1i - T1j;
|
cannam@167
|
362 T4R = T1i + T1j;
|
cannam@167
|
363 T1l = T1h * T1k;
|
cannam@167
|
364 T67 = T4M * T4R;
|
cannam@167
|
365 T1o = Rp[WS(rs, 7)];
|
cannam@167
|
366 T1p = Rm[WS(rs, 7)];
|
cannam@167
|
367 T1q = T1o + T1p;
|
cannam@167
|
368 T4N = T1o - T1p;
|
cannam@167
|
369 T4O = T4M * T4N;
|
cannam@167
|
370 T5H = T1h * T1q;
|
cannam@167
|
371 }
|
cannam@167
|
372 {
|
cannam@167
|
373 E T19, T1r, T66, T68;
|
cannam@167
|
374 T19 = FNMS(T15, T18, T13);
|
cannam@167
|
375 T1r = FNMS(T1n, T1q, T1l);
|
cannam@167
|
376 T1s = T19 + T1r;
|
cannam@167
|
377 T8h = T19 - T1r;
|
cannam@167
|
378 T66 = FNMS(T4I, T4F, T65);
|
cannam@167
|
379 T68 = FNMS(T4Q, T4N, T67);
|
cannam@167
|
380 T69 = T66 + T68;
|
cannam@167
|
381 T8B = T66 - T68;
|
cannam@167
|
382 }
|
cannam@167
|
383 {
|
cannam@167
|
384 E T4K, T4S, T5G, T5I;
|
cannam@167
|
385 T4K = FMA(T4I, T4J, T4G);
|
cannam@167
|
386 T4S = FMA(T4Q, T4R, T4O);
|
cannam@167
|
387 T4T = T4K + T4S;
|
cannam@167
|
388 T8q = T4S - T4K;
|
cannam@167
|
389 T5G = FMA(T15, T12, T5F);
|
cannam@167
|
390 T5I = FMA(T1n, T1k, T5H);
|
cannam@167
|
391 T5J = T5G + T5I;
|
cannam@167
|
392 T8k = T5G - T5I;
|
cannam@167
|
393 }
|
cannam@167
|
394 }
|
cannam@167
|
395 {
|
cannam@167
|
396 E T9, T4i, Ta, T5U, Tf, T4g, T4h, T5u, Tp, T4q, Tq, T5W, Tv, T4m, T4n;
|
cannam@167
|
397 E T5w;
|
cannam@167
|
398 {
|
cannam@167
|
399 E T7, T8, Td, Te;
|
cannam@167
|
400 T7 = Ip[WS(rs, 1)];
|
cannam@167
|
401 T8 = Im[WS(rs, 1)];
|
cannam@167
|
402 T9 = T7 - T8;
|
cannam@167
|
403 T4i = T7 + T8;
|
cannam@167
|
404 Ta = T6 * T9;
|
cannam@167
|
405 T5U = T2 * T4i;
|
cannam@167
|
406 Td = Rp[WS(rs, 1)];
|
cannam@167
|
407 Te = Rm[WS(rs, 1)];
|
cannam@167
|
408 Tf = Td + Te;
|
cannam@167
|
409 T4g = Td - Te;
|
cannam@167
|
410 T4h = T2 * T4g;
|
cannam@167
|
411 T5u = T6 * Tf;
|
cannam@167
|
412 }
|
cannam@167
|
413 {
|
cannam@167
|
414 E Tn, To, Tt, Tu;
|
cannam@167
|
415 Tn = Ip[WS(rs, 9)];
|
cannam@167
|
416 To = Im[WS(rs, 9)];
|
cannam@167
|
417 Tp = Tn - To;
|
cannam@167
|
418 T4q = Tn + To;
|
cannam@167
|
419 Tq = Tm * Tp;
|
cannam@167
|
420 T5W = T4l * T4q;
|
cannam@167
|
421 Tt = Rp[WS(rs, 9)];
|
cannam@167
|
422 Tu = Rm[WS(rs, 9)];
|
cannam@167
|
423 Tv = Tt + Tu;
|
cannam@167
|
424 T4m = Tt - Tu;
|
cannam@167
|
425 T4n = T4l * T4m;
|
cannam@167
|
426 T5w = Tm * Tv;
|
cannam@167
|
427 }
|
cannam@167
|
428 {
|
cannam@167
|
429 E Tg, Tw, T5v, T5x;
|
cannam@167
|
430 Tg = FNMS(Tc, Tf, Ta);
|
cannam@167
|
431 Tw = FNMS(Ts, Tv, Tq);
|
cannam@167
|
432 Tx = Tg + Tw;
|
cannam@167
|
433 T8a = Tg - Tw;
|
cannam@167
|
434 T5v = FMA(Tc, T9, T5u);
|
cannam@167
|
435 T5x = FMA(Ts, Tp, T5w);
|
cannam@167
|
436 T5y = T5v + T5x;
|
cannam@167
|
437 T8d = T5v - T5x;
|
cannam@167
|
438 {
|
cannam@167
|
439 E T4j, T4r, T8t, T5V, T5X, T8u;
|
cannam@167
|
440 T4j = FMA(T5, T4i, T4h);
|
cannam@167
|
441 T4r = FMA(T4p, T4q, T4n);
|
cannam@167
|
442 T8t = T4r - T4j;
|
cannam@167
|
443 T5V = FNMS(T5, T4g, T5U);
|
cannam@167
|
444 T5X = FNMS(T4p, T4m, T5W);
|
cannam@167
|
445 T8u = T5V - T5X;
|
cannam@167
|
446 T4s = T4j + T4r;
|
cannam@167
|
447 T8E = T8u + T8t;
|
cannam@167
|
448 T5Y = T5V + T5X;
|
cannam@167
|
449 T8v = T8t - T8u;
|
cannam@167
|
450 }
|
cannam@167
|
451 }
|
cannam@167
|
452 }
|
cannam@167
|
453 {
|
cannam@167
|
454 E T1Y, T1Z, T3p, T22, T3o, T6h, T5i, T2c, T2d, T3x, T2i, T3u, T6j, T5k;
|
cannam@167
|
455 {
|
cannam@167
|
456 E T1W, T1X, T3n, T20, T21;
|
cannam@167
|
457 T1W = Ip[WS(rs, 2)];
|
cannam@167
|
458 T1X = Im[WS(rs, 2)];
|
cannam@167
|
459 T1Y = T1W - T1X;
|
cannam@167
|
460 T1Z = T1a * T1Y;
|
cannam@167
|
461 T3p = T1W + T1X;
|
cannam@167
|
462 T20 = Rp[WS(rs, 2)];
|
cannam@167
|
463 T21 = Rm[WS(rs, 2)];
|
cannam@167
|
464 T22 = T20 + T21;
|
cannam@167
|
465 T3n = T21 - T20;
|
cannam@167
|
466 T3o = T1F * T3n;
|
cannam@167
|
467 T6h = T1H * T3n;
|
cannam@167
|
468 T5i = T1a * T22;
|
cannam@167
|
469 }
|
cannam@167
|
470 {
|
cannam@167
|
471 E T2a, T2b, T3t, T2g, T2h;
|
cannam@167
|
472 T2a = Ip[WS(rs, 10)];
|
cannam@167
|
473 T2b = Im[WS(rs, 10)];
|
cannam@167
|
474 T2c = T2a - T2b;
|
cannam@167
|
475 T2d = T29 * T2c;
|
cannam@167
|
476 T3x = T2a + T2b;
|
cannam@167
|
477 T2g = Rp[WS(rs, 10)];
|
cannam@167
|
478 T2h = Rm[WS(rs, 10)];
|
cannam@167
|
479 T2i = T2g + T2h;
|
cannam@167
|
480 T3t = T2h - T2g;
|
cannam@167
|
481 T3u = T3s * T3t;
|
cannam@167
|
482 T6j = T3w * T3t;
|
cannam@167
|
483 T5k = T29 * T2i;
|
cannam@167
|
484 }
|
cannam@167
|
485 {
|
cannam@167
|
486 E T23, T2j, T5j, T5l;
|
cannam@167
|
487 T23 = FNMS(T1c, T22, T1Z);
|
cannam@167
|
488 T2j = FNMS(T2f, T2i, T2d);
|
cannam@167
|
489 T2k = T23 + T2j;
|
cannam@167
|
490 T82 = T23 - T2j;
|
cannam@167
|
491 T5j = FMA(T1c, T1Y, T5i);
|
cannam@167
|
492 T5l = FMA(T2f, T2c, T5k);
|
cannam@167
|
493 T5m = T5j + T5l;
|
cannam@167
|
494 T83 = T5j - T5l;
|
cannam@167
|
495 {
|
cannam@167
|
496 E T3q, T3y, T8M, T6i, T6k, T8N;
|
cannam@167
|
497 T3q = FNMS(T1H, T3p, T3o);
|
cannam@167
|
498 T3y = FNMS(T3w, T3x, T3u);
|
cannam@167
|
499 T8M = T3q - T3y;
|
cannam@167
|
500 T6i = FMA(T1F, T3p, T6h);
|
cannam@167
|
501 T6k = FMA(T3s, T3x, T6j);
|
cannam@167
|
502 T8N = T6i - T6k;
|
cannam@167
|
503 T3z = T3q + T3y;
|
cannam@167
|
504 T8X = T8M + T8N;
|
cannam@167
|
505 T6l = T6i + T6k;
|
cannam@167
|
506 T8O = T8M - T8N;
|
cannam@167
|
507 }
|
cannam@167
|
508 }
|
cannam@167
|
509 }
|
cannam@167
|
510 {
|
cannam@167
|
511 E T2o, T2p, T3G, T2t, T3D, T6m, T5n, T2y, T2z, T3K, T2D, T3J, T6o, T5p;
|
cannam@167
|
512 {
|
cannam@167
|
513 E T2m, T2n, T3C, T2r, T2s;
|
cannam@167
|
514 T2m = Ip[WS(rs, 14)];
|
cannam@167
|
515 T2n = Im[WS(rs, 14)];
|
cannam@167
|
516 T2o = T2m - T2n;
|
cannam@167
|
517 T2p = T2l * T2o;
|
cannam@167
|
518 T3G = T2m + T2n;
|
cannam@167
|
519 T2r = Rp[WS(rs, 14)];
|
cannam@167
|
520 T2s = Rm[WS(rs, 14)];
|
cannam@167
|
521 T2t = T2r + T2s;
|
cannam@167
|
522 T3C = T2s - T2r;
|
cannam@167
|
523 T3D = T3B * T3C;
|
cannam@167
|
524 T6m = T3F * T3C;
|
cannam@167
|
525 T5n = T2l * T2t;
|
cannam@167
|
526 }
|
cannam@167
|
527 {
|
cannam@167
|
528 E T2w, T2x, T3I, T2B, T2C;
|
cannam@167
|
529 T2w = Ip[WS(rs, 6)];
|
cannam@167
|
530 T2x = Im[WS(rs, 6)];
|
cannam@167
|
531 T2y = T2w - T2x;
|
cannam@167
|
532 T2z = T2v * T2y;
|
cannam@167
|
533 T3K = T2w + T2x;
|
cannam@167
|
534 T2B = Rp[WS(rs, 6)];
|
cannam@167
|
535 T2C = Rm[WS(rs, 6)];
|
cannam@167
|
536 T2D = T2B + T2C;
|
cannam@167
|
537 T3I = T2C - T2B;
|
cannam@167
|
538 T3J = T1d * T3I;
|
cannam@167
|
539 T6o = T1g * T3I;
|
cannam@167
|
540 T5p = T2v * T2D;
|
cannam@167
|
541 }
|
cannam@167
|
542 {
|
cannam@167
|
543 E T2u, T2E, T5o, T5q;
|
cannam@167
|
544 T2u = FNMS(T2q, T2t, T2p);
|
cannam@167
|
545 T2E = FNMS(T2A, T2D, T2z);
|
cannam@167
|
546 T2F = T2u + T2E;
|
cannam@167
|
547 T86 = T2u - T2E;
|
cannam@167
|
548 T5o = FMA(T2q, T2o, T5n);
|
cannam@167
|
549 T5q = FMA(T2A, T2y, T5p);
|
cannam@167
|
550 T5r = T5o + T5q;
|
cannam@167
|
551 T85 = T5o - T5q;
|
cannam@167
|
552 {
|
cannam@167
|
553 E T3H, T3L, T8P, T6n, T6p, T8Q;
|
cannam@167
|
554 T3H = FNMS(T3F, T3G, T3D);
|
cannam@167
|
555 T3L = FNMS(T1g, T3K, T3J);
|
cannam@167
|
556 T8P = T3H - T3L;
|
cannam@167
|
557 T6n = FMA(T3B, T3G, T6m);
|
cannam@167
|
558 T6p = FMA(T1d, T3K, T6o);
|
cannam@167
|
559 T8Q = T6n - T6p;
|
cannam@167
|
560 T3M = T3H + T3L;
|
cannam@167
|
561 T8Y = T8Q - T8P;
|
cannam@167
|
562 T6q = T6n + T6p;
|
cannam@167
|
563 T8R = T8P + T8Q;
|
cannam@167
|
564 }
|
cannam@167
|
565 }
|
cannam@167
|
566 }
|
cannam@167
|
567 {
|
cannam@167
|
568 E TC, T4v, TD, T5Z, TI, T4t, T4u, T5z, TO, T4z, TP, T61, TU, T4x, T4y;
|
cannam@167
|
569 E T5B;
|
cannam@167
|
570 {
|
cannam@167
|
571 E TA, TB, TG, TH;
|
cannam@167
|
572 TA = Ip[WS(rs, 5)];
|
cannam@167
|
573 TB = Im[WS(rs, 5)];
|
cannam@167
|
574 TC = TA - TB;
|
cannam@167
|
575 T4v = TA + TB;
|
cannam@167
|
576 TD = Tz * TC;
|
cannam@167
|
577 T5Z = T32 * T4v;
|
cannam@167
|
578 TG = Rp[WS(rs, 5)];
|
cannam@167
|
579 TH = Rm[WS(rs, 5)];
|
cannam@167
|
580 TI = TG + TH;
|
cannam@167
|
581 T4t = TG - TH;
|
cannam@167
|
582 T4u = T32 * T4t;
|
cannam@167
|
583 T5z = Tz * TI;
|
cannam@167
|
584 }
|
cannam@167
|
585 {
|
cannam@167
|
586 E TM, TN, TS, TT;
|
cannam@167
|
587 TM = Ip[WS(rs, 13)];
|
cannam@167
|
588 TN = Im[WS(rs, 13)];
|
cannam@167
|
589 TO = TM - TN;
|
cannam@167
|
590 T4z = TM + TN;
|
cannam@167
|
591 TP = TL * TO;
|
cannam@167
|
592 T61 = Ti * T4z;
|
cannam@167
|
593 TS = Rp[WS(rs, 13)];
|
cannam@167
|
594 TT = Rm[WS(rs, 13)];
|
cannam@167
|
595 TU = TS + TT;
|
cannam@167
|
596 T4x = TS - TT;
|
cannam@167
|
597 T4y = Ti * T4x;
|
cannam@167
|
598 T5B = TL * TU;
|
cannam@167
|
599 }
|
cannam@167
|
600 {
|
cannam@167
|
601 E TJ, TV, T5A, T5C;
|
cannam@167
|
602 TJ = FNMS(TF, TI, TD);
|
cannam@167
|
603 TV = FNMS(TR, TU, TP);
|
cannam@167
|
604 TW = TJ + TV;
|
cannam@167
|
605 T8e = TJ - TV;
|
cannam@167
|
606 T5A = FMA(TF, TC, T5z);
|
cannam@167
|
607 T5C = FMA(TR, TO, T5B);
|
cannam@167
|
608 T5D = T5A + T5C;
|
cannam@167
|
609 T8b = T5A - T5C;
|
cannam@167
|
610 {
|
cannam@167
|
611 E T4w, T4A, T8x, T60, T62, T8w;
|
cannam@167
|
612 T4w = FMA(T34, T4v, T4u);
|
cannam@167
|
613 T4A = FMA(Tl, T4z, T4y);
|
cannam@167
|
614 T8x = T4w - T4A;
|
cannam@167
|
615 T60 = FNMS(T34, T4t, T5Z);
|
cannam@167
|
616 T62 = FNMS(Tl, T4x, T61);
|
cannam@167
|
617 T8w = T62 - T60;
|
cannam@167
|
618 T4B = T4w + T4A;
|
cannam@167
|
619 T8F = T8w - T8x;
|
cannam@167
|
620 T63 = T60 + T62;
|
cannam@167
|
621 T8y = T8w + T8x;
|
cannam@167
|
622 }
|
cannam@167
|
623 }
|
cannam@167
|
624 }
|
cannam@167
|
625 {
|
cannam@167
|
626 E T1V, T6S, T3l, T6I, T5Q, T6H, T5t, T6R, T56, T6W, T6g, T6M, T4f, T6X, T6D;
|
cannam@167
|
627 E T6P;
|
cannam@167
|
628 {
|
cannam@167
|
629 E TX, T1U, T5h, T5s;
|
cannam@167
|
630 TX = Tx + TW;
|
cannam@167
|
631 T1U = T1s + T1T;
|
cannam@167
|
632 T1V = TX + T1U;
|
cannam@167
|
633 T6S = TX - T1U;
|
cannam@167
|
634 {
|
cannam@167
|
635 E T2G, T3k, T5E, T5P;
|
cannam@167
|
636 T2G = T2k + T2F;
|
cannam@167
|
637 T3k = T31 + T3j;
|
cannam@167
|
638 T3l = T2G + T3k;
|
cannam@167
|
639 T6I = T3k - T2G;
|
cannam@167
|
640 T5E = T5y + T5D;
|
cannam@167
|
641 T5P = T5J + T5O;
|
cannam@167
|
642 T5Q = T5E + T5P;
|
cannam@167
|
643 T6H = T5P - T5E;
|
cannam@167
|
644 }
|
cannam@167
|
645 T5h = T5b + T5g;
|
cannam@167
|
646 T5s = T5m + T5r;
|
cannam@167
|
647 T5t = T5h + T5s;
|
cannam@167
|
648 T6R = T5h - T5s;
|
cannam@167
|
649 {
|
cannam@167
|
650 E T4C, T55, T6L, T64, T6f, T6K;
|
cannam@167
|
651 T4C = T4s + T4B;
|
cannam@167
|
652 T55 = T4T + T54;
|
cannam@167
|
653 T6L = T4C - T55;
|
cannam@167
|
654 T64 = T5Y + T63;
|
cannam@167
|
655 T6f = T69 + T6e;
|
cannam@167
|
656 T6K = T6f - T64;
|
cannam@167
|
657 T56 = T4C + T55;
|
cannam@167
|
658 T6W = T6K - T6L;
|
cannam@167
|
659 T6g = T64 + T6f;
|
cannam@167
|
660 T6M = T6K + T6L;
|
cannam@167
|
661 }
|
cannam@167
|
662 {
|
cannam@167
|
663 E T3N, T4e, T6N, T6r, T6C, T6O;
|
cannam@167
|
664 T3N = T3z + T3M;
|
cannam@167
|
665 T4e = T3Y + T4d;
|
cannam@167
|
666 T6N = T4e - T3N;
|
cannam@167
|
667 T6r = T6l + T6q;
|
cannam@167
|
668 T6C = T6w + T6B;
|
cannam@167
|
669 T6O = T6C - T6r;
|
cannam@167
|
670 T4f = T3N + T4e;
|
cannam@167
|
671 T6X = T6N + T6O;
|
cannam@167
|
672 T6D = T6r + T6C;
|
cannam@167
|
673 T6P = T6N - T6O;
|
cannam@167
|
674 }
|
cannam@167
|
675 }
|
cannam@167
|
676 {
|
cannam@167
|
677 E T3m, T57, T6F, T6G;
|
cannam@167
|
678 T3m = T1V + T3l;
|
cannam@167
|
679 T57 = T4f - T56;
|
cannam@167
|
680 Ip[0] = KP500000000 * (T3m + T57);
|
cannam@167
|
681 Im[WS(rs, 15)] = KP500000000 * (T57 - T3m);
|
cannam@167
|
682 T6F = T5t + T5Q;
|
cannam@167
|
683 T6G = T6g + T6D;
|
cannam@167
|
684 Rm[WS(rs, 15)] = KP500000000 * (T6F - T6G);
|
cannam@167
|
685 Rp[0] = KP500000000 * (T6F + T6G);
|
cannam@167
|
686 }
|
cannam@167
|
687 {
|
cannam@167
|
688 E T5R, T5S, T5T, T6E;
|
cannam@167
|
689 T5R = T5t - T5Q;
|
cannam@167
|
690 T5S = T56 + T4f;
|
cannam@167
|
691 Rm[WS(rs, 7)] = KP500000000 * (T5R - T5S);
|
cannam@167
|
692 Rp[WS(rs, 8)] = KP500000000 * (T5R + T5S);
|
cannam@167
|
693 T5T = T3l - T1V;
|
cannam@167
|
694 T6E = T6g - T6D;
|
cannam@167
|
695 Ip[WS(rs, 8)] = KP500000000 * (T5T + T6E);
|
cannam@167
|
696 Im[WS(rs, 7)] = KP500000000 * (T6E - T5T);
|
cannam@167
|
697 }
|
cannam@167
|
698 {
|
cannam@167
|
699 E T6J, T6Q, T6Z, T70;
|
cannam@167
|
700 T6J = T6H + T6I;
|
cannam@167
|
701 T6Q = T6M + T6P;
|
cannam@167
|
702 Ip[WS(rs, 4)] = KP500000000 * (FMA(KP707106781, T6Q, T6J));
|
cannam@167
|
703 Im[WS(rs, 11)] = -(KP500000000 * (FNMS(KP707106781, T6Q, T6J)));
|
cannam@167
|
704 T6Z = T6R + T6S;
|
cannam@167
|
705 T70 = T6W + T6X;
|
cannam@167
|
706 Rm[WS(rs, 11)] = KP500000000 * (FNMS(KP707106781, T70, T6Z));
|
cannam@167
|
707 Rp[WS(rs, 4)] = KP500000000 * (FMA(KP707106781, T70, T6Z));
|
cannam@167
|
708 }
|
cannam@167
|
709 {
|
cannam@167
|
710 E T6T, T6U, T6V, T6Y;
|
cannam@167
|
711 T6T = T6R - T6S;
|
cannam@167
|
712 T6U = T6P - T6M;
|
cannam@167
|
713 Rm[WS(rs, 3)] = KP500000000 * (FNMS(KP707106781, T6U, T6T));
|
cannam@167
|
714 Rp[WS(rs, 12)] = KP500000000 * (FMA(KP707106781, T6U, T6T));
|
cannam@167
|
715 T6V = T6I - T6H;
|
cannam@167
|
716 T6Y = T6W - T6X;
|
cannam@167
|
717 Ip[WS(rs, 12)] = KP500000000 * (FMA(KP707106781, T6Y, T6V));
|
cannam@167
|
718 Im[WS(rs, 3)] = -(KP500000000 * (FNMS(KP707106781, T6Y, T6V)));
|
cannam@167
|
719 }
|
cannam@167
|
720 }
|
cannam@167
|
721 {
|
cannam@167
|
722 E T73, T7F, T7t, T7P, T7a, T7Q, T7w, T7G, T7i, T7U, T7A, T7K, T7p, T7V, T7B;
|
cannam@167
|
723 E T7N;
|
cannam@167
|
724 {
|
cannam@167
|
725 E T71, T72, T7r, T7s;
|
cannam@167
|
726 T71 = T5r - T5m;
|
cannam@167
|
727 T72 = T3j - T31;
|
cannam@167
|
728 T73 = T71 + T72;
|
cannam@167
|
729 T7F = T72 - T71;
|
cannam@167
|
730 T7r = T5b - T5g;
|
cannam@167
|
731 T7s = T2k - T2F;
|
cannam@167
|
732 T7t = T7r + T7s;
|
cannam@167
|
733 T7P = T7r - T7s;
|
cannam@167
|
734 }
|
cannam@167
|
735 {
|
cannam@167
|
736 E T76, T7u, T79, T7v;
|
cannam@167
|
737 {
|
cannam@167
|
738 E T74, T75, T77, T78;
|
cannam@167
|
739 T74 = Tx - TW;
|
cannam@167
|
740 T75 = T5y - T5D;
|
cannam@167
|
741 T76 = T74 - T75;
|
cannam@167
|
742 T7u = T75 + T74;
|
cannam@167
|
743 T77 = T5J - T5O;
|
cannam@167
|
744 T78 = T1s - T1T;
|
cannam@167
|
745 T79 = T77 + T78;
|
cannam@167
|
746 T7v = T77 - T78;
|
cannam@167
|
747 }
|
cannam@167
|
748 T7a = T76 + T79;
|
cannam@167
|
749 T7Q = T76 - T79;
|
cannam@167
|
750 T7w = T7u + T7v;
|
cannam@167
|
751 T7G = T7v - T7u;
|
cannam@167
|
752 }
|
cannam@167
|
753 {
|
cannam@167
|
754 E T7e, T7I, T7h, T7J;
|
cannam@167
|
755 {
|
cannam@167
|
756 E T7c, T7d, T7f, T7g;
|
cannam@167
|
757 T7c = T63 - T5Y;
|
cannam@167
|
758 T7d = T54 - T4T;
|
cannam@167
|
759 T7e = T7c + T7d;
|
cannam@167
|
760 T7I = T7c - T7d;
|
cannam@167
|
761 T7f = T4B - T4s;
|
cannam@167
|
762 T7g = T69 - T6e;
|
cannam@167
|
763 T7h = T7f + T7g;
|
cannam@167
|
764 T7J = T7g - T7f;
|
cannam@167
|
765 }
|
cannam@167
|
766 T7i = FMA(KP414213562, T7h, T7e);
|
cannam@167
|
767 T7U = FNMS(KP414213562, T7I, T7J);
|
cannam@167
|
768 T7A = FNMS(KP414213562, T7e, T7h);
|
cannam@167
|
769 T7K = FMA(KP414213562, T7J, T7I);
|
cannam@167
|
770 }
|
cannam@167
|
771 {
|
cannam@167
|
772 E T7l, T7L, T7o, T7M;
|
cannam@167
|
773 {
|
cannam@167
|
774 E T7j, T7k, T7m, T7n;
|
cannam@167
|
775 T7j = T6q - T6l;
|
cannam@167
|
776 T7k = T4d - T3Y;
|
cannam@167
|
777 T7l = T7j + T7k;
|
cannam@167
|
778 T7L = T7k - T7j;
|
cannam@167
|
779 T7m = T3z - T3M;
|
cannam@167
|
780 T7n = T6B - T6w;
|
cannam@167
|
781 T7o = T7m + T7n;
|
cannam@167
|
782 T7M = T7n - T7m;
|
cannam@167
|
783 }
|
cannam@167
|
784 T7p = FNMS(KP414213562, T7o, T7l);
|
cannam@167
|
785 T7V = FNMS(KP414213562, T7L, T7M);
|
cannam@167
|
786 T7B = FMA(KP414213562, T7l, T7o);
|
cannam@167
|
787 T7N = FMA(KP414213562, T7M, T7L);
|
cannam@167
|
788 }
|
cannam@167
|
789 {
|
cannam@167
|
790 E T7b, T7q, T7D, T7E;
|
cannam@167
|
791 T7b = FMA(KP707106781, T7a, T73);
|
cannam@167
|
792 T7q = T7i + T7p;
|
cannam@167
|
793 Ip[WS(rs, 2)] = KP500000000 * (FMA(KP923879532, T7q, T7b));
|
cannam@167
|
794 Im[WS(rs, 13)] = -(KP500000000 * (FNMS(KP923879532, T7q, T7b)));
|
cannam@167
|
795 T7D = FMA(KP707106781, T7w, T7t);
|
cannam@167
|
796 T7E = T7A + T7B;
|
cannam@167
|
797 Rm[WS(rs, 13)] = KP500000000 * (FNMS(KP923879532, T7E, T7D));
|
cannam@167
|
798 Rp[WS(rs, 2)] = KP500000000 * (FMA(KP923879532, T7E, T7D));
|
cannam@167
|
799 }
|
cannam@167
|
800 {
|
cannam@167
|
801 E T7x, T7y, T7z, T7C;
|
cannam@167
|
802 T7x = FNMS(KP707106781, T7w, T7t);
|
cannam@167
|
803 T7y = T7p - T7i;
|
cannam@167
|
804 Rm[WS(rs, 5)] = KP500000000 * (FNMS(KP923879532, T7y, T7x));
|
cannam@167
|
805 Rp[WS(rs, 10)] = KP500000000 * (FMA(KP923879532, T7y, T7x));
|
cannam@167
|
806 T7z = FNMS(KP707106781, T7a, T73);
|
cannam@167
|
807 T7C = T7A - T7B;
|
cannam@167
|
808 Ip[WS(rs, 10)] = KP500000000 * (FMA(KP923879532, T7C, T7z));
|
cannam@167
|
809 Im[WS(rs, 5)] = -(KP500000000 * (FNMS(KP923879532, T7C, T7z)));
|
cannam@167
|
810 }
|
cannam@167
|
811 {
|
cannam@167
|
812 E T7H, T7O, T7X, T7Y;
|
cannam@167
|
813 T7H = FNMS(KP707106781, T7G, T7F);
|
cannam@167
|
814 T7O = T7K - T7N;
|
cannam@167
|
815 Ip[WS(rs, 14)] = KP500000000 * (FMA(KP923879532, T7O, T7H));
|
cannam@167
|
816 Im[WS(rs, 1)] = -(KP500000000 * (FNMS(KP923879532, T7O, T7H)));
|
cannam@167
|
817 T7X = FNMS(KP707106781, T7Q, T7P);
|
cannam@167
|
818 T7Y = T7U + T7V;
|
cannam@167
|
819 Rp[WS(rs, 14)] = KP500000000 * (FNMS(KP923879532, T7Y, T7X));
|
cannam@167
|
820 Rm[WS(rs, 1)] = KP500000000 * (FMA(KP923879532, T7Y, T7X));
|
cannam@167
|
821 }
|
cannam@167
|
822 {
|
cannam@167
|
823 E T7R, T7S, T7T, T7W;
|
cannam@167
|
824 T7R = FMA(KP707106781, T7Q, T7P);
|
cannam@167
|
825 T7S = T7K + T7N;
|
cannam@167
|
826 Rm[WS(rs, 9)] = KP500000000 * (FNMS(KP923879532, T7S, T7R));
|
cannam@167
|
827 Rp[WS(rs, 6)] = KP500000000 * (FMA(KP923879532, T7S, T7R));
|
cannam@167
|
828 T7T = FMA(KP707106781, T7G, T7F);
|
cannam@167
|
829 T7W = T7U - T7V;
|
cannam@167
|
830 Ip[WS(rs, 6)] = KP500000000 * (FMA(KP923879532, T7W, T7T));
|
cannam@167
|
831 Im[WS(rs, 9)] = -(KP500000000 * (FNMS(KP923879532, T7W, T7T)));
|
cannam@167
|
832 }
|
cannam@167
|
833 }
|
cannam@167
|
834 {
|
cannam@167
|
835 E T89, Tat, T9l, Ta7, T99, Taj, T9v, T9H, T8o, T9w, T9c, T9m, Ta3, Tay, Tae;
|
cannam@167
|
836 E Tao, T8I, T9A, T9g, T9q, T9O, Tau, Taa, Tak, T9W, Taz, Taf, Tar, T91, T9B;
|
cannam@167
|
837 E T9h, T9t;
|
cannam@167
|
838 {
|
cannam@167
|
839 E T81, Ta5, T88, Ta6, T84, T87;
|
cannam@167
|
840 T81 = T7Z - T80;
|
cannam@167
|
841 Ta5 = T93 - T94;
|
cannam@167
|
842 T84 = T82 - T83;
|
cannam@167
|
843 T87 = T85 + T86;
|
cannam@167
|
844 T88 = T84 + T87;
|
cannam@167
|
845 Ta6 = T84 - T87;
|
cannam@167
|
846 T89 = FMA(KP707106781, T88, T81);
|
cannam@167
|
847 Tat = FNMS(KP707106781, Ta6, Ta5);
|
cannam@167
|
848 T9l = FNMS(KP707106781, T88, T81);
|
cannam@167
|
849 Ta7 = FMA(KP707106781, Ta6, Ta5);
|
cannam@167
|
850 }
|
cannam@167
|
851 {
|
cannam@167
|
852 E T95, T9F, T98, T9G, T96, T97;
|
cannam@167
|
853 T95 = T93 + T94;
|
cannam@167
|
854 T9F = T80 + T7Z;
|
cannam@167
|
855 T96 = T83 + T82;
|
cannam@167
|
856 T97 = T85 - T86;
|
cannam@167
|
857 T98 = T96 + T97;
|
cannam@167
|
858 T9G = T97 - T96;
|
cannam@167
|
859 T99 = FMA(KP707106781, T98, T95);
|
cannam@167
|
860 Taj = FNMS(KP707106781, T9G, T9F);
|
cannam@167
|
861 T9v = FNMS(KP707106781, T98, T95);
|
cannam@167
|
862 T9H = FMA(KP707106781, T9G, T9F);
|
cannam@167
|
863 }
|
cannam@167
|
864 {
|
cannam@167
|
865 E T8g, T9a, T8n, T9b;
|
cannam@167
|
866 {
|
cannam@167
|
867 E T8c, T8f, T8j, T8m;
|
cannam@167
|
868 T8c = T8a - T8b;
|
cannam@167
|
869 T8f = T8d + T8e;
|
cannam@167
|
870 T8g = FNMS(KP414213562, T8f, T8c);
|
cannam@167
|
871 T9a = FMA(KP414213562, T8c, T8f);
|
cannam@167
|
872 T8j = T8h - T8i;
|
cannam@167
|
873 T8m = T8k + T8l;
|
cannam@167
|
874 T8n = FMA(KP414213562, T8m, T8j);
|
cannam@167
|
875 T9b = FNMS(KP414213562, T8j, T8m);
|
cannam@167
|
876 }
|
cannam@167
|
877 T8o = T8g + T8n;
|
cannam@167
|
878 T9w = T8g - T8n;
|
cannam@167
|
879 T9c = T9a + T9b;
|
cannam@167
|
880 T9m = T9b - T9a;
|
cannam@167
|
881 }
|
cannam@167
|
882 {
|
cannam@167
|
883 E T9Z, Tam, Ta2, Tan;
|
cannam@167
|
884 {
|
cannam@167
|
885 E T9X, T9Y, Ta0, Ta1;
|
cannam@167
|
886 T9X = T8r - T8q;
|
cannam@167
|
887 T9Y = T8F - T8E;
|
cannam@167
|
888 T9Z = FNMS(KP707106781, T9Y, T9X);
|
cannam@167
|
889 Tam = FMA(KP707106781, T9Y, T9X);
|
cannam@167
|
890 Ta0 = T8B - T8C;
|
cannam@167
|
891 Ta1 = T8y - T8v;
|
cannam@167
|
892 Ta2 = FNMS(KP707106781, Ta1, Ta0);
|
cannam@167
|
893 Tan = FMA(KP707106781, Ta1, Ta0);
|
cannam@167
|
894 }
|
cannam@167
|
895 Ta3 = FNMS(KP668178637, Ta2, T9Z);
|
cannam@167
|
896 Tay = FNMS(KP198912367, Tam, Tan);
|
cannam@167
|
897 Tae = FMA(KP668178637, T9Z, Ta2);
|
cannam@167
|
898 Tao = FMA(KP198912367, Tan, Tam);
|
cannam@167
|
899 }
|
cannam@167
|
900 {
|
cannam@167
|
901 E T8A, T9o, T8H, T9p;
|
cannam@167
|
902 {
|
cannam@167
|
903 E T8s, T8z, T8D, T8G;
|
cannam@167
|
904 T8s = T8q + T8r;
|
cannam@167
|
905 T8z = T8v + T8y;
|
cannam@167
|
906 T8A = FMA(KP707106781, T8z, T8s);
|
cannam@167
|
907 T9o = FNMS(KP707106781, T8z, T8s);
|
cannam@167
|
908 T8D = T8B + T8C;
|
cannam@167
|
909 T8G = T8E + T8F;
|
cannam@167
|
910 T8H = FMA(KP707106781, T8G, T8D);
|
cannam@167
|
911 T9p = FNMS(KP707106781, T8G, T8D);
|
cannam@167
|
912 }
|
cannam@167
|
913 T8I = FMA(KP198912367, T8H, T8A);
|
cannam@167
|
914 T9A = FMA(KP668178637, T9o, T9p);
|
cannam@167
|
915 T9g = FNMS(KP198912367, T8A, T8H);
|
cannam@167
|
916 T9q = FNMS(KP668178637, T9p, T9o);
|
cannam@167
|
917 }
|
cannam@167
|
918 {
|
cannam@167
|
919 E T9K, Ta9, T9N, Ta8;
|
cannam@167
|
920 {
|
cannam@167
|
921 E T9I, T9J, T9L, T9M;
|
cannam@167
|
922 T9I = T8k - T8l;
|
cannam@167
|
923 T9J = T8h + T8i;
|
cannam@167
|
924 T9K = FMA(KP414213562, T9J, T9I);
|
cannam@167
|
925 Ta9 = FNMS(KP414213562, T9I, T9J);
|
cannam@167
|
926 T9L = T8d - T8e;
|
cannam@167
|
927 T9M = T8a + T8b;
|
cannam@167
|
928 T9N = FNMS(KP414213562, T9M, T9L);
|
cannam@167
|
929 Ta8 = FMA(KP414213562, T9L, T9M);
|
cannam@167
|
930 }
|
cannam@167
|
931 T9O = T9K - T9N;
|
cannam@167
|
932 Tau = T9N + T9K;
|
cannam@167
|
933 Taa = Ta8 - Ta9;
|
cannam@167
|
934 Tak = Ta8 + Ta9;
|
cannam@167
|
935 }
|
cannam@167
|
936 {
|
cannam@167
|
937 E T9S, Tap, T9V, Taq;
|
cannam@167
|
938 {
|
cannam@167
|
939 E T9Q, T9R, T9T, T9U;
|
cannam@167
|
940 T9Q = T8K + T8J;
|
cannam@167
|
941 T9R = T8X - T8Y;
|
cannam@167
|
942 T9S = FNMS(KP707106781, T9R, T9Q);
|
cannam@167
|
943 Tap = FMA(KP707106781, T9R, T9Q);
|
cannam@167
|
944 T9T = T8V - T8U;
|
cannam@167
|
945 T9U = T8R - T8O;
|
cannam@167
|
946 T9V = FNMS(KP707106781, T9U, T9T);
|
cannam@167
|
947 Taq = FMA(KP707106781, T9U, T9T);
|
cannam@167
|
948 }
|
cannam@167
|
949 T9W = FNMS(KP668178637, T9V, T9S);
|
cannam@167
|
950 Taz = FNMS(KP198912367, Tap, Taq);
|
cannam@167
|
951 Taf = FMA(KP668178637, T9S, T9V);
|
cannam@167
|
952 Tar = FMA(KP198912367, Taq, Tap);
|
cannam@167
|
953 }
|
cannam@167
|
954 {
|
cannam@167
|
955 E T8T, T9r, T90, T9s;
|
cannam@167
|
956 {
|
cannam@167
|
957 E T8L, T8S, T8W, T8Z;
|
cannam@167
|
958 T8L = T8J - T8K;
|
cannam@167
|
959 T8S = T8O + T8R;
|
cannam@167
|
960 T8T = FMA(KP707106781, T8S, T8L);
|
cannam@167
|
961 T9r = FNMS(KP707106781, T8S, T8L);
|
cannam@167
|
962 T8W = T8U + T8V;
|
cannam@167
|
963 T8Z = T8X + T8Y;
|
cannam@167
|
964 T90 = FMA(KP707106781, T8Z, T8W);
|
cannam@167
|
965 T9s = FNMS(KP707106781, T8Z, T8W);
|
cannam@167
|
966 }
|
cannam@167
|
967 T91 = FNMS(KP198912367, T90, T8T);
|
cannam@167
|
968 T9B = FNMS(KP668178637, T9r, T9s);
|
cannam@167
|
969 T9h = FMA(KP198912367, T8T, T90);
|
cannam@167
|
970 T9t = FMA(KP668178637, T9s, T9r);
|
cannam@167
|
971 }
|
cannam@167
|
972 {
|
cannam@167
|
973 E T8p, T92, T9j, T9k;
|
cannam@167
|
974 T8p = FMA(KP923879532, T8o, T89);
|
cannam@167
|
975 T92 = T8I + T91;
|
cannam@167
|
976 Ip[WS(rs, 1)] = KP500000000 * (FMA(KP980785280, T92, T8p));
|
cannam@167
|
977 Im[WS(rs, 14)] = -(KP500000000 * (FNMS(KP980785280, T92, T8p)));
|
cannam@167
|
978 T9j = FMA(KP923879532, T9c, T99);
|
cannam@167
|
979 T9k = T9g + T9h;
|
cannam@167
|
980 Rm[WS(rs, 14)] = KP500000000 * (FNMS(KP980785280, T9k, T9j));
|
cannam@167
|
981 Rp[WS(rs, 1)] = KP500000000 * (FMA(KP980785280, T9k, T9j));
|
cannam@167
|
982 }
|
cannam@167
|
983 {
|
cannam@167
|
984 E T9d, T9e, T9f, T9i;
|
cannam@167
|
985 T9d = FNMS(KP923879532, T9c, T99);
|
cannam@167
|
986 T9e = T91 - T8I;
|
cannam@167
|
987 Rm[WS(rs, 6)] = KP500000000 * (FNMS(KP980785280, T9e, T9d));
|
cannam@167
|
988 Rp[WS(rs, 9)] = KP500000000 * (FMA(KP980785280, T9e, T9d));
|
cannam@167
|
989 T9f = FNMS(KP923879532, T8o, T89);
|
cannam@167
|
990 T9i = T9g - T9h;
|
cannam@167
|
991 Ip[WS(rs, 9)] = KP500000000 * (FMA(KP980785280, T9i, T9f));
|
cannam@167
|
992 Im[WS(rs, 6)] = -(KP500000000 * (FNMS(KP980785280, T9i, T9f)));
|
cannam@167
|
993 }
|
cannam@167
|
994 {
|
cannam@167
|
995 E T9n, T9u, T9D, T9E;
|
cannam@167
|
996 T9n = FNMS(KP923879532, T9m, T9l);
|
cannam@167
|
997 T9u = T9q + T9t;
|
cannam@167
|
998 Ip[WS(rs, 13)] = KP500000000 * (FNMS(KP831469612, T9u, T9n));
|
cannam@167
|
999 Im[WS(rs, 2)] = -(KP500000000 * (FMA(KP831469612, T9u, T9n)));
|
cannam@167
|
1000 T9D = FNMS(KP923879532, T9w, T9v);
|
cannam@167
|
1001 T9E = T9A + T9B;
|
cannam@167
|
1002 Rp[WS(rs, 13)] = KP500000000 * (FNMS(KP831469612, T9E, T9D));
|
cannam@167
|
1003 Rm[WS(rs, 2)] = KP500000000 * (FMA(KP831469612, T9E, T9D));
|
cannam@167
|
1004 }
|
cannam@167
|
1005 {
|
cannam@167
|
1006 E T9x, T9y, T9z, T9C;
|
cannam@167
|
1007 T9x = FMA(KP923879532, T9w, T9v);
|
cannam@167
|
1008 T9y = T9t - T9q;
|
cannam@167
|
1009 Rm[WS(rs, 10)] = KP500000000 * (FNMS(KP831469612, T9y, T9x));
|
cannam@167
|
1010 Rp[WS(rs, 5)] = KP500000000 * (FMA(KP831469612, T9y, T9x));
|
cannam@167
|
1011 T9z = FMA(KP923879532, T9m, T9l);
|
cannam@167
|
1012 T9C = T9A - T9B;
|
cannam@167
|
1013 Ip[WS(rs, 5)] = KP500000000 * (FMA(KP831469612, T9C, T9z));
|
cannam@167
|
1014 Im[WS(rs, 10)] = -(KP500000000 * (FNMS(KP831469612, T9C, T9z)));
|
cannam@167
|
1015 }
|
cannam@167
|
1016 {
|
cannam@167
|
1017 E T9P, Ta4, Tah, Tai;
|
cannam@167
|
1018 T9P = FMA(KP923879532, T9O, T9H);
|
cannam@167
|
1019 Ta4 = T9W - Ta3;
|
cannam@167
|
1020 Ip[WS(rs, 3)] = KP500000000 * (FMA(KP831469612, Ta4, T9P));
|
cannam@167
|
1021 Im[WS(rs, 12)] = -(KP500000000 * (FNMS(KP831469612, Ta4, T9P)));
|
cannam@167
|
1022 Tah = FMA(KP923879532, Taa, Ta7);
|
cannam@167
|
1023 Tai = Tae + Taf;
|
cannam@167
|
1024 Rm[WS(rs, 12)] = KP500000000 * (FNMS(KP831469612, Tai, Tah));
|
cannam@167
|
1025 Rp[WS(rs, 3)] = KP500000000 * (FMA(KP831469612, Tai, Tah));
|
cannam@167
|
1026 }
|
cannam@167
|
1027 {
|
cannam@167
|
1028 E Tab, Tac, Tad, Tag;
|
cannam@167
|
1029 Tab = FNMS(KP923879532, Taa, Ta7);
|
cannam@167
|
1030 Tac = Ta3 + T9W;
|
cannam@167
|
1031 Rm[WS(rs, 4)] = KP500000000 * (FNMS(KP831469612, Tac, Tab));
|
cannam@167
|
1032 Rp[WS(rs, 11)] = KP500000000 * (FMA(KP831469612, Tac, Tab));
|
cannam@167
|
1033 Tad = FNMS(KP923879532, T9O, T9H);
|
cannam@167
|
1034 Tag = Tae - Taf;
|
cannam@167
|
1035 Ip[WS(rs, 11)] = KP500000000 * (FMA(KP831469612, Tag, Tad));
|
cannam@167
|
1036 Im[WS(rs, 4)] = -(KP500000000 * (FNMS(KP831469612, Tag, Tad)));
|
cannam@167
|
1037 }
|
cannam@167
|
1038 {
|
cannam@167
|
1039 E Tal, Tas, TaB, TaC;
|
cannam@167
|
1040 Tal = FMA(KP923879532, Tak, Taj);
|
cannam@167
|
1041 Tas = Tao - Tar;
|
cannam@167
|
1042 Ip[WS(rs, 15)] = KP500000000 * (FMA(KP980785280, Tas, Tal));
|
cannam@167
|
1043 Im[0] = -(KP500000000 * (FNMS(KP980785280, Tas, Tal)));
|
cannam@167
|
1044 TaB = FMA(KP923879532, Tau, Tat);
|
cannam@167
|
1045 TaC = Tay + Taz;
|
cannam@167
|
1046 Rp[WS(rs, 15)] = KP500000000 * (FNMS(KP980785280, TaC, TaB));
|
cannam@167
|
1047 Rm[0] = KP500000000 * (FMA(KP980785280, TaC, TaB));
|
cannam@167
|
1048 }
|
cannam@167
|
1049 {
|
cannam@167
|
1050 E Tav, Taw, Tax, TaA;
|
cannam@167
|
1051 Tav = FNMS(KP923879532, Tau, Tat);
|
cannam@167
|
1052 Taw = Tao + Tar;
|
cannam@167
|
1053 Rm[WS(rs, 8)] = KP500000000 * (FNMS(KP980785280, Taw, Tav));
|
cannam@167
|
1054 Rp[WS(rs, 7)] = KP500000000 * (FMA(KP980785280, Taw, Tav));
|
cannam@167
|
1055 Tax = FNMS(KP923879532, Tak, Taj);
|
cannam@167
|
1056 TaA = Tay - Taz;
|
cannam@167
|
1057 Ip[WS(rs, 7)] = KP500000000 * (FMA(KP980785280, TaA, Tax));
|
cannam@167
|
1058 Im[WS(rs, 8)] = -(KP500000000 * (FNMS(KP980785280, TaA, Tax)));
|
cannam@167
|
1059 }
|
cannam@167
|
1060 }
|
cannam@167
|
1061 }
|
cannam@167
|
1062 }
|
cannam@167
|
1063 }
|
cannam@167
|
1064 }
|
cannam@167
|
1065
|
cannam@167
|
1066 static const tw_instr twinstr[] = {
|
cannam@167
|
1067 {TW_CEXP, 1, 1},
|
cannam@167
|
1068 {TW_CEXP, 1, 3},
|
cannam@167
|
1069 {TW_CEXP, 1, 9},
|
cannam@167
|
1070 {TW_CEXP, 1, 27},
|
cannam@167
|
1071 {TW_NEXT, 1, 0}
|
cannam@167
|
1072 };
|
cannam@167
|
1073
|
cannam@167
|
1074 static const hc2c_desc desc = { 32, "hc2cfdft2_32", twinstr, &GENUS, {300, 162, 252, 0} };
|
cannam@167
|
1075
|
cannam@167
|
1076 void X(codelet_hc2cfdft2_32) (planner *p) {
|
cannam@167
|
1077 X(khc2c_register) (p, hc2cfdft2_32, &desc, HC2C_VIA_DFT);
|
cannam@167
|
1078 }
|
cannam@167
|
1079 #else
|
cannam@167
|
1080
|
cannam@167
|
1081 /* Generated by: ../../../genfft/gen_hc2cdft.native -compact -variables 4 -pipeline-latency 4 -twiddle-log3 -precompute-twiddles -n 32 -dit -name hc2cfdft2_32 -include rdft/scalar/hc2cf.h */
|
cannam@167
|
1082
|
cannam@167
|
1083 /*
|
cannam@167
|
1084 * This function contains 552 FP additions, 300 FP multiplications,
|
cannam@167
|
1085 * (or, 440 additions, 188 multiplications, 112 fused multiply/add),
|
cannam@167
|
1086 * 166 stack variables, 9 constants, and 128 memory accesses
|
cannam@167
|
1087 */
|
cannam@167
|
1088 #include "rdft/scalar/hc2cf.h"
|
cannam@167
|
1089
|
cannam@167
|
1090 static void hc2cfdft2_32(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms)
|
cannam@167
|
1091 {
|
cannam@167
|
1092 DK(KP277785116, +0.277785116509801112371415406974266437187468595);
|
cannam@167
|
1093 DK(KP415734806, +0.415734806151272618539394188808952878369280406);
|
cannam@167
|
1094 DK(KP097545161, +0.097545161008064133924142434238511120463845809);
|
cannam@167
|
1095 DK(KP490392640, +0.490392640201615224563091118067119518486966865);
|
cannam@167
|
1096 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
cannam@167
|
1097 DK(KP191341716, +0.191341716182544885864229992015199433380672281);
|
cannam@167
|
1098 DK(KP461939766, +0.461939766255643378064091594698394143411208313);
|
cannam@167
|
1099 DK(KP353553390, +0.353553390593273762200422181052424519642417969);
|
cannam@167
|
1100 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
|
cannam@167
|
1101 {
|
cannam@167
|
1102 INT m;
|
cannam@167
|
1103 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(128, rs)) {
|
cannam@167
|
1104 E T1, T4, T2, T5, T7, T1b, T1d, Td, Ti, Tk, Tj, Tl, TL, TR, T2h;
|
cannam@167
|
1105 E T2O, T16, T2l, T10, T2K, Tm, Tq, T3s, T3K, T3w, T3M, T4e, T4u, T4i, T4w;
|
cannam@167
|
1106 E Ty, TE, T3h, T3j, T2q, T2u, T4l, T4n, T1v, T1B, T3E, T3G, T2B, T2F, T3Y;
|
cannam@167
|
1107 E T40, T1f, T1G, T1i, T1H, T1j, T1M, T1n, T1I, T23, T2U, T26, T2V, T27, T30;
|
cannam@167
|
1108 E T2b, T2W;
|
cannam@167
|
1109 {
|
cannam@167
|
1110 E Tw, T1A, TD, T1t, Tx, T1z, TC, T1u, TJ, T15, TQ, TY, TK, T14, TP;
|
cannam@167
|
1111 E TZ;
|
cannam@167
|
1112 {
|
cannam@167
|
1113 E T3, Tc, T6, Tb;
|
cannam@167
|
1114 T1 = W[0];
|
cannam@167
|
1115 T4 = W[1];
|
cannam@167
|
1116 T2 = W[2];
|
cannam@167
|
1117 T5 = W[3];
|
cannam@167
|
1118 T3 = T1 * T2;
|
cannam@167
|
1119 Tc = T4 * T2;
|
cannam@167
|
1120 T6 = T4 * T5;
|
cannam@167
|
1121 Tb = T1 * T5;
|
cannam@167
|
1122 T7 = T3 + T6;
|
cannam@167
|
1123 T1b = T3 - T6;
|
cannam@167
|
1124 T1d = Tb + Tc;
|
cannam@167
|
1125 Td = Tb - Tc;
|
cannam@167
|
1126 Ti = W[4];
|
cannam@167
|
1127 Tw = T1 * Ti;
|
cannam@167
|
1128 T1A = T5 * Ti;
|
cannam@167
|
1129 TD = T4 * Ti;
|
cannam@167
|
1130 T1t = T2 * Ti;
|
cannam@167
|
1131 Tk = W[5];
|
cannam@167
|
1132 Tx = T4 * Tk;
|
cannam@167
|
1133 T1z = T2 * Tk;
|
cannam@167
|
1134 TC = T1 * Tk;
|
cannam@167
|
1135 T1u = T5 * Tk;
|
cannam@167
|
1136 Tj = W[6];
|
cannam@167
|
1137 TJ = T1 * Tj;
|
cannam@167
|
1138 T15 = T5 * Tj;
|
cannam@167
|
1139 TQ = T4 * Tj;
|
cannam@167
|
1140 TY = T2 * Tj;
|
cannam@167
|
1141 Tl = W[7];
|
cannam@167
|
1142 TK = T4 * Tl;
|
cannam@167
|
1143 T14 = T2 * Tl;
|
cannam@167
|
1144 TP = T1 * Tl;
|
cannam@167
|
1145 TZ = T5 * Tl;
|
cannam@167
|
1146 }
|
cannam@167
|
1147 TL = TJ + TK;
|
cannam@167
|
1148 TR = TP - TQ;
|
cannam@167
|
1149 T2h = TJ - TK;
|
cannam@167
|
1150 T2O = T14 - T15;
|
cannam@167
|
1151 T16 = T14 + T15;
|
cannam@167
|
1152 T2l = TP + TQ;
|
cannam@167
|
1153 T10 = TY - TZ;
|
cannam@167
|
1154 T2K = TY + TZ;
|
cannam@167
|
1155 Tm = FMA(Ti, Tj, Tk * Tl);
|
cannam@167
|
1156 Tq = FNMS(Tk, Tj, Ti * Tl);
|
cannam@167
|
1157 {
|
cannam@167
|
1158 E T3q, T3r, T3u, T3v;
|
cannam@167
|
1159 T3q = T7 * Tj;
|
cannam@167
|
1160 T3r = Td * Tl;
|
cannam@167
|
1161 T3s = T3q + T3r;
|
cannam@167
|
1162 T3K = T3q - T3r;
|
cannam@167
|
1163 T3u = T7 * Tl;
|
cannam@167
|
1164 T3v = Td * Tj;
|
cannam@167
|
1165 T3w = T3u - T3v;
|
cannam@167
|
1166 T3M = T3u + T3v;
|
cannam@167
|
1167 }
|
cannam@167
|
1168 {
|
cannam@167
|
1169 E T4c, T4d, T4g, T4h;
|
cannam@167
|
1170 T4c = T1b * Tj;
|
cannam@167
|
1171 T4d = T1d * Tl;
|
cannam@167
|
1172 T4e = T4c - T4d;
|
cannam@167
|
1173 T4u = T4c + T4d;
|
cannam@167
|
1174 T4g = T1b * Tl;
|
cannam@167
|
1175 T4h = T1d * Tj;
|
cannam@167
|
1176 T4i = T4g + T4h;
|
cannam@167
|
1177 T4w = T4g - T4h;
|
cannam@167
|
1178 Ty = Tw - Tx;
|
cannam@167
|
1179 TE = TC + TD;
|
cannam@167
|
1180 T3h = FMA(Ty, Tj, TE * Tl);
|
cannam@167
|
1181 T3j = FNMS(TE, Tj, Ty * Tl);
|
cannam@167
|
1182 }
|
cannam@167
|
1183 T2q = T1t - T1u;
|
cannam@167
|
1184 T2u = T1z + T1A;
|
cannam@167
|
1185 T4l = FMA(T2q, Tj, T2u * Tl);
|
cannam@167
|
1186 T4n = FNMS(T2u, Tj, T2q * Tl);
|
cannam@167
|
1187 T1v = T1t + T1u;
|
cannam@167
|
1188 T1B = T1z - T1A;
|
cannam@167
|
1189 T3E = FMA(T1v, Tj, T1B * Tl);
|
cannam@167
|
1190 T3G = FNMS(T1B, Tj, T1v * Tl);
|
cannam@167
|
1191 T2B = Tw + Tx;
|
cannam@167
|
1192 T2F = TC - TD;
|
cannam@167
|
1193 T3Y = FMA(T2B, Tj, T2F * Tl);
|
cannam@167
|
1194 T40 = FNMS(T2F, Tj, T2B * Tl);
|
cannam@167
|
1195 {
|
cannam@167
|
1196 E T1c, T1e, T1g, T1h;
|
cannam@167
|
1197 T1c = T1b * Ti;
|
cannam@167
|
1198 T1e = T1d * Tk;
|
cannam@167
|
1199 T1f = T1c - T1e;
|
cannam@167
|
1200 T1G = T1c + T1e;
|
cannam@167
|
1201 T1g = T1b * Tk;
|
cannam@167
|
1202 T1h = T1d * Ti;
|
cannam@167
|
1203 T1i = T1g + T1h;
|
cannam@167
|
1204 T1H = T1g - T1h;
|
cannam@167
|
1205 }
|
cannam@167
|
1206 T1j = FMA(T1f, Tj, T1i * Tl);
|
cannam@167
|
1207 T1M = FNMS(T1H, Tj, T1G * Tl);
|
cannam@167
|
1208 T1n = FNMS(T1i, Tj, T1f * Tl);
|
cannam@167
|
1209 T1I = FMA(T1G, Tj, T1H * Tl);
|
cannam@167
|
1210 {
|
cannam@167
|
1211 E T21, T22, T24, T25;
|
cannam@167
|
1212 T21 = T7 * Ti;
|
cannam@167
|
1213 T22 = Td * Tk;
|
cannam@167
|
1214 T23 = T21 + T22;
|
cannam@167
|
1215 T2U = T21 - T22;
|
cannam@167
|
1216 T24 = T7 * Tk;
|
cannam@167
|
1217 T25 = Td * Ti;
|
cannam@167
|
1218 T26 = T24 - T25;
|
cannam@167
|
1219 T2V = T24 + T25;
|
cannam@167
|
1220 }
|
cannam@167
|
1221 T27 = FMA(T23, Tj, T26 * Tl);
|
cannam@167
|
1222 T30 = FNMS(T2V, Tj, T2U * Tl);
|
cannam@167
|
1223 T2b = FNMS(T26, Tj, T23 * Tl);
|
cannam@167
|
1224 T2W = FMA(T2U, Tj, T2V * Tl);
|
cannam@167
|
1225 }
|
cannam@167
|
1226 {
|
cannam@167
|
1227 E T38, T7l, T7S, T8Y, T7Z, T91, T3A, T6k, T4F, T83, T5C, T6n, T2T, T84, T4I;
|
cannam@167
|
1228 E T7m, T2g, T4M, T4P, T2z, T3T, T6m, T7O, T7V, T7j, T87, T5v, T6j, T7L, T7U;
|
cannam@167
|
1229 E T7g, T86, Tv, TW, T61, T4U, T4X, T62, T4b, T6c, T7v, T7C, T5g, T6f, T74;
|
cannam@167
|
1230 E T8G, T7s, T7B, T71, T8F, T1s, T1R, T65, T51, T54, T64, T4A, T6g, T7G, T8U;
|
cannam@167
|
1231 E T5n, T6d, T7b, T8J, T7z, T8R, T78, T8I;
|
cannam@167
|
1232 {
|
cannam@167
|
1233 E T2E, T2I, T3p, T5w, T37, T4D, T3g, T5A, T2N, T2R, T3y, T5x, T2Z, T33, T3l;
|
cannam@167
|
1234 E T5z;
|
cannam@167
|
1235 {
|
cannam@167
|
1236 E T2C, T2D, T3o, T2G, T2H, T3n;
|
cannam@167
|
1237 T2C = Ip[WS(rs, 4)];
|
cannam@167
|
1238 T2D = Im[WS(rs, 4)];
|
cannam@167
|
1239 T3o = T2C + T2D;
|
cannam@167
|
1240 T2G = Rp[WS(rs, 4)];
|
cannam@167
|
1241 T2H = Rm[WS(rs, 4)];
|
cannam@167
|
1242 T3n = T2G - T2H;
|
cannam@167
|
1243 T2E = T2C - T2D;
|
cannam@167
|
1244 T2I = T2G + T2H;
|
cannam@167
|
1245 T3p = FMA(Ti, T3n, Tk * T3o);
|
cannam@167
|
1246 T5w = FNMS(Tk, T3n, Ti * T3o);
|
cannam@167
|
1247 }
|
cannam@167
|
1248 {
|
cannam@167
|
1249 E T35, T36, T3f, T3c, T3d, T3e;
|
cannam@167
|
1250 T35 = Ip[0];
|
cannam@167
|
1251 T36 = Im[0];
|
cannam@167
|
1252 T3f = T35 + T36;
|
cannam@167
|
1253 T3c = Rm[0];
|
cannam@167
|
1254 T3d = Rp[0];
|
cannam@167
|
1255 T3e = T3c - T3d;
|
cannam@167
|
1256 T37 = T35 - T36;
|
cannam@167
|
1257 T4D = T3d + T3c;
|
cannam@167
|
1258 T3g = FNMS(T4, T3f, T1 * T3e);
|
cannam@167
|
1259 T5A = FMA(T4, T3e, T1 * T3f);
|
cannam@167
|
1260 }
|
cannam@167
|
1261 {
|
cannam@167
|
1262 E T2L, T2M, T3x, T2P, T2Q, T3t;
|
cannam@167
|
1263 T2L = Ip[WS(rs, 12)];
|
cannam@167
|
1264 T2M = Im[WS(rs, 12)];
|
cannam@167
|
1265 T3x = T2L + T2M;
|
cannam@167
|
1266 T2P = Rp[WS(rs, 12)];
|
cannam@167
|
1267 T2Q = Rm[WS(rs, 12)];
|
cannam@167
|
1268 T3t = T2P - T2Q;
|
cannam@167
|
1269 T2N = T2L - T2M;
|
cannam@167
|
1270 T2R = T2P + T2Q;
|
cannam@167
|
1271 T3y = FMA(T3s, T3t, T3w * T3x);
|
cannam@167
|
1272 T5x = FNMS(T3w, T3t, T3s * T3x);
|
cannam@167
|
1273 }
|
cannam@167
|
1274 {
|
cannam@167
|
1275 E T2X, T2Y, T3k, T31, T32, T3i;
|
cannam@167
|
1276 T2X = Ip[WS(rs, 8)];
|
cannam@167
|
1277 T2Y = Im[WS(rs, 8)];
|
cannam@167
|
1278 T3k = T2X + T2Y;
|
cannam@167
|
1279 T31 = Rp[WS(rs, 8)];
|
cannam@167
|
1280 T32 = Rm[WS(rs, 8)];
|
cannam@167
|
1281 T3i = T31 - T32;
|
cannam@167
|
1282 T2Z = T2X - T2Y;
|
cannam@167
|
1283 T33 = T31 + T32;
|
cannam@167
|
1284 T3l = FMA(T3h, T3i, T3j * T3k);
|
cannam@167
|
1285 T5z = FNMS(T3j, T3i, T3h * T3k);
|
cannam@167
|
1286 }
|
cannam@167
|
1287 {
|
cannam@167
|
1288 E T34, T7Q, T7R, T4E, T5y, T5B;
|
cannam@167
|
1289 T34 = FNMS(T30, T33, T2W * T2Z);
|
cannam@167
|
1290 T38 = T34 + T37;
|
cannam@167
|
1291 T7l = T37 - T34;
|
cannam@167
|
1292 T7Q = T3l + T3g;
|
cannam@167
|
1293 T7R = T5w - T5x;
|
cannam@167
|
1294 T7S = T7Q - T7R;
|
cannam@167
|
1295 T8Y = T7R + T7Q;
|
cannam@167
|
1296 {
|
cannam@167
|
1297 E T7X, T7Y, T3m, T3z;
|
cannam@167
|
1298 T7X = T3y - T3p;
|
cannam@167
|
1299 T7Y = T5A - T5z;
|
cannam@167
|
1300 T7Z = T7X + T7Y;
|
cannam@167
|
1301 T91 = T7Y - T7X;
|
cannam@167
|
1302 T3m = T3g - T3l;
|
cannam@167
|
1303 T3z = T3p + T3y;
|
cannam@167
|
1304 T3A = T3m - T3z;
|
cannam@167
|
1305 T6k = T3z + T3m;
|
cannam@167
|
1306 }
|
cannam@167
|
1307 T4E = FMA(T2W, T33, T30 * T2Z);
|
cannam@167
|
1308 T4F = T4D + T4E;
|
cannam@167
|
1309 T83 = T4D - T4E;
|
cannam@167
|
1310 T5y = T5w + T5x;
|
cannam@167
|
1311 T5B = T5z + T5A;
|
cannam@167
|
1312 T5C = T5y + T5B;
|
cannam@167
|
1313 T6n = T5B - T5y;
|
cannam@167
|
1314 {
|
cannam@167
|
1315 E T2J, T2S, T4G, T4H;
|
cannam@167
|
1316 T2J = FNMS(T2F, T2I, T2B * T2E);
|
cannam@167
|
1317 T2S = FNMS(T2O, T2R, T2K * T2N);
|
cannam@167
|
1318 T2T = T2J + T2S;
|
cannam@167
|
1319 T84 = T2J - T2S;
|
cannam@167
|
1320 T4G = FMA(T2B, T2I, T2F * T2E);
|
cannam@167
|
1321 T4H = FMA(T2K, T2R, T2O * T2N);
|
cannam@167
|
1322 T4I = T4G + T4H;
|
cannam@167
|
1323 T7m = T4G - T4H;
|
cannam@167
|
1324 }
|
cannam@167
|
1325 }
|
cannam@167
|
1326 }
|
cannam@167
|
1327 {
|
cannam@167
|
1328 E T20, T5p, T3D, T4K, T2y, T5t, T3R, T4O, T2f, T5q, T3I, T4L, T2p, T5s, T3O;
|
cannam@167
|
1329 E T4N;
|
cannam@167
|
1330 {
|
cannam@167
|
1331 E T1W, T3C, T1Z, T3B;
|
cannam@167
|
1332 {
|
cannam@167
|
1333 E T1U, T1V, T1X, T1Y;
|
cannam@167
|
1334 T1U = Ip[WS(rs, 2)];
|
cannam@167
|
1335 T1V = Im[WS(rs, 2)];
|
cannam@167
|
1336 T1W = T1U - T1V;
|
cannam@167
|
1337 T3C = T1U + T1V;
|
cannam@167
|
1338 T1X = Rp[WS(rs, 2)];
|
cannam@167
|
1339 T1Y = Rm[WS(rs, 2)];
|
cannam@167
|
1340 T1Z = T1X + T1Y;
|
cannam@167
|
1341 T3B = T1X - T1Y;
|
cannam@167
|
1342 }
|
cannam@167
|
1343 T20 = FNMS(T1d, T1Z, T1b * T1W);
|
cannam@167
|
1344 T5p = FNMS(T1H, T3B, T1G * T3C);
|
cannam@167
|
1345 T3D = FMA(T1G, T3B, T1H * T3C);
|
cannam@167
|
1346 T4K = FMA(T1b, T1Z, T1d * T1W);
|
cannam@167
|
1347 }
|
cannam@167
|
1348 {
|
cannam@167
|
1349 E T2t, T3Q, T2x, T3P;
|
cannam@167
|
1350 {
|
cannam@167
|
1351 E T2r, T2s, T2v, T2w;
|
cannam@167
|
1352 T2r = Ip[WS(rs, 6)];
|
cannam@167
|
1353 T2s = Im[WS(rs, 6)];
|
cannam@167
|
1354 T2t = T2r - T2s;
|
cannam@167
|
1355 T3Q = T2r + T2s;
|
cannam@167
|
1356 T2v = Rp[WS(rs, 6)];
|
cannam@167
|
1357 T2w = Rm[WS(rs, 6)];
|
cannam@167
|
1358 T2x = T2v + T2w;
|
cannam@167
|
1359 T3P = T2v - T2w;
|
cannam@167
|
1360 }
|
cannam@167
|
1361 T2y = FNMS(T2u, T2x, T2q * T2t);
|
cannam@167
|
1362 T5t = FNMS(T1i, T3P, T1f * T3Q);
|
cannam@167
|
1363 T3R = FMA(T1f, T3P, T1i * T3Q);
|
cannam@167
|
1364 T4O = FMA(T2q, T2x, T2u * T2t);
|
cannam@167
|
1365 }
|
cannam@167
|
1366 {
|
cannam@167
|
1367 E T2a, T3H, T2e, T3F;
|
cannam@167
|
1368 {
|
cannam@167
|
1369 E T28, T29, T2c, T2d;
|
cannam@167
|
1370 T28 = Ip[WS(rs, 10)];
|
cannam@167
|
1371 T29 = Im[WS(rs, 10)];
|
cannam@167
|
1372 T2a = T28 - T29;
|
cannam@167
|
1373 T3H = T28 + T29;
|
cannam@167
|
1374 T2c = Rp[WS(rs, 10)];
|
cannam@167
|
1375 T2d = Rm[WS(rs, 10)];
|
cannam@167
|
1376 T2e = T2c + T2d;
|
cannam@167
|
1377 T3F = T2c - T2d;
|
cannam@167
|
1378 }
|
cannam@167
|
1379 T2f = FNMS(T2b, T2e, T27 * T2a);
|
cannam@167
|
1380 T5q = FNMS(T3G, T3F, T3E * T3H);
|
cannam@167
|
1381 T3I = FMA(T3E, T3F, T3G * T3H);
|
cannam@167
|
1382 T4L = FMA(T27, T2e, T2b * T2a);
|
cannam@167
|
1383 }
|
cannam@167
|
1384 {
|
cannam@167
|
1385 E T2k, T3N, T2o, T3L;
|
cannam@167
|
1386 {
|
cannam@167
|
1387 E T2i, T2j, T2m, T2n;
|
cannam@167
|
1388 T2i = Ip[WS(rs, 14)];
|
cannam@167
|
1389 T2j = Im[WS(rs, 14)];
|
cannam@167
|
1390 T2k = T2i - T2j;
|
cannam@167
|
1391 T3N = T2i + T2j;
|
cannam@167
|
1392 T2m = Rp[WS(rs, 14)];
|
cannam@167
|
1393 T2n = Rm[WS(rs, 14)];
|
cannam@167
|
1394 T2o = T2m + T2n;
|
cannam@167
|
1395 T3L = T2m - T2n;
|
cannam@167
|
1396 }
|
cannam@167
|
1397 T2p = FNMS(T2l, T2o, T2h * T2k);
|
cannam@167
|
1398 T5s = FNMS(T3M, T3L, T3K * T3N);
|
cannam@167
|
1399 T3O = FMA(T3K, T3L, T3M * T3N);
|
cannam@167
|
1400 T4N = FMA(T2h, T2o, T2l * T2k);
|
cannam@167
|
1401 }
|
cannam@167
|
1402 {
|
cannam@167
|
1403 E T3J, T3S, T5r, T5u;
|
cannam@167
|
1404 T2g = T20 + T2f;
|
cannam@167
|
1405 T4M = T4K + T4L;
|
cannam@167
|
1406 T4P = T4N + T4O;
|
cannam@167
|
1407 T2z = T2p + T2y;
|
cannam@167
|
1408 T3J = T3D + T3I;
|
cannam@167
|
1409 T3S = T3O + T3R;
|
cannam@167
|
1410 T3T = T3J + T3S;
|
cannam@167
|
1411 T6m = T3S - T3J;
|
cannam@167
|
1412 {
|
cannam@167
|
1413 E T7M, T7N, T7h, T7i;
|
cannam@167
|
1414 T7M = T5s - T5t;
|
cannam@167
|
1415 T7N = T3R - T3O;
|
cannam@167
|
1416 T7O = T7M + T7N;
|
cannam@167
|
1417 T7V = T7M - T7N;
|
cannam@167
|
1418 T7h = T4N - T4O;
|
cannam@167
|
1419 T7i = T2p - T2y;
|
cannam@167
|
1420 T7j = T7h + T7i;
|
cannam@167
|
1421 T87 = T7h - T7i;
|
cannam@167
|
1422 }
|
cannam@167
|
1423 T5r = T5p + T5q;
|
cannam@167
|
1424 T5u = T5s + T5t;
|
cannam@167
|
1425 T5v = T5r + T5u;
|
cannam@167
|
1426 T6j = T5u - T5r;
|
cannam@167
|
1427 {
|
cannam@167
|
1428 E T7J, T7K, T7e, T7f;
|
cannam@167
|
1429 T7J = T3I - T3D;
|
cannam@167
|
1430 T7K = T5p - T5q;
|
cannam@167
|
1431 T7L = T7J - T7K;
|
cannam@167
|
1432 T7U = T7K + T7J;
|
cannam@167
|
1433 T7e = T20 - T2f;
|
cannam@167
|
1434 T7f = T4K - T4L;
|
cannam@167
|
1435 T7g = T7e - T7f;
|
cannam@167
|
1436 T86 = T7f + T7e;
|
cannam@167
|
1437 }
|
cannam@167
|
1438 }
|
cannam@167
|
1439 }
|
cannam@167
|
1440 {
|
cannam@167
|
1441 E Th, T5a, T3X, T4S, TV, T5e, T49, T4W, Tu, T5b, T42, T4T, TI, T5d, T46;
|
cannam@167
|
1442 E T4V;
|
cannam@167
|
1443 {
|
cannam@167
|
1444 E Ta, T3W, Tg, T3V;
|
cannam@167
|
1445 {
|
cannam@167
|
1446 E T8, T9, Te, Tf;
|
cannam@167
|
1447 T8 = Ip[WS(rs, 1)];
|
cannam@167
|
1448 T9 = Im[WS(rs, 1)];
|
cannam@167
|
1449 Ta = T8 - T9;
|
cannam@167
|
1450 T3W = T8 + T9;
|
cannam@167
|
1451 Te = Rp[WS(rs, 1)];
|
cannam@167
|
1452 Tf = Rm[WS(rs, 1)];
|
cannam@167
|
1453 Tg = Te + Tf;
|
cannam@167
|
1454 T3V = Te - Tf;
|
cannam@167
|
1455 }
|
cannam@167
|
1456 Th = FNMS(Td, Tg, T7 * Ta);
|
cannam@167
|
1457 T5a = FNMS(T5, T3V, T2 * T3W);
|
cannam@167
|
1458 T3X = FMA(T2, T3V, T5 * T3W);
|
cannam@167
|
1459 T4S = FMA(T7, Tg, Td * Ta);
|
cannam@167
|
1460 }
|
cannam@167
|
1461 {
|
cannam@167
|
1462 E TO, T48, TU, T47;
|
cannam@167
|
1463 {
|
cannam@167
|
1464 E TM, TN, TS, TT;
|
cannam@167
|
1465 TM = Ip[WS(rs, 13)];
|
cannam@167
|
1466 TN = Im[WS(rs, 13)];
|
cannam@167
|
1467 TO = TM - TN;
|
cannam@167
|
1468 T48 = TM + TN;
|
cannam@167
|
1469 TS = Rp[WS(rs, 13)];
|
cannam@167
|
1470 TT = Rm[WS(rs, 13)];
|
cannam@167
|
1471 TU = TS + TT;
|
cannam@167
|
1472 T47 = TS - TT;
|
cannam@167
|
1473 }
|
cannam@167
|
1474 TV = FNMS(TR, TU, TL * TO);
|
cannam@167
|
1475 T5e = FNMS(Tl, T47, Tj * T48);
|
cannam@167
|
1476 T49 = FMA(Tj, T47, Tl * T48);
|
cannam@167
|
1477 T4W = FMA(TL, TU, TR * TO);
|
cannam@167
|
1478 }
|
cannam@167
|
1479 {
|
cannam@167
|
1480 E Tp, T41, Tt, T3Z;
|
cannam@167
|
1481 {
|
cannam@167
|
1482 E Tn, To, Tr, Ts;
|
cannam@167
|
1483 Tn = Ip[WS(rs, 9)];
|
cannam@167
|
1484 To = Im[WS(rs, 9)];
|
cannam@167
|
1485 Tp = Tn - To;
|
cannam@167
|
1486 T41 = Tn + To;
|
cannam@167
|
1487 Tr = Rp[WS(rs, 9)];
|
cannam@167
|
1488 Ts = Rm[WS(rs, 9)];
|
cannam@167
|
1489 Tt = Tr + Ts;
|
cannam@167
|
1490 T3Z = Tr - Ts;
|
cannam@167
|
1491 }
|
cannam@167
|
1492 Tu = FNMS(Tq, Tt, Tm * Tp);
|
cannam@167
|
1493 T5b = FNMS(T40, T3Z, T3Y * T41);
|
cannam@167
|
1494 T42 = FMA(T3Y, T3Z, T40 * T41);
|
cannam@167
|
1495 T4T = FMA(Tm, Tt, Tq * Tp);
|
cannam@167
|
1496 }
|
cannam@167
|
1497 {
|
cannam@167
|
1498 E TB, T45, TH, T44;
|
cannam@167
|
1499 {
|
cannam@167
|
1500 E Tz, TA, TF, TG;
|
cannam@167
|
1501 Tz = Ip[WS(rs, 5)];
|
cannam@167
|
1502 TA = Im[WS(rs, 5)];
|
cannam@167
|
1503 TB = Tz - TA;
|
cannam@167
|
1504 T45 = Tz + TA;
|
cannam@167
|
1505 TF = Rp[WS(rs, 5)];
|
cannam@167
|
1506 TG = Rm[WS(rs, 5)];
|
cannam@167
|
1507 TH = TF + TG;
|
cannam@167
|
1508 T44 = TF - TG;
|
cannam@167
|
1509 }
|
cannam@167
|
1510 TI = FNMS(TE, TH, Ty * TB);
|
cannam@167
|
1511 T5d = FNMS(T2V, T44, T2U * T45);
|
cannam@167
|
1512 T46 = FMA(T2U, T44, T2V * T45);
|
cannam@167
|
1513 T4V = FMA(Ty, TH, TE * TB);
|
cannam@167
|
1514 }
|
cannam@167
|
1515 Tv = Th + Tu;
|
cannam@167
|
1516 TW = TI + TV;
|
cannam@167
|
1517 T61 = Tv - TW;
|
cannam@167
|
1518 T4U = T4S + T4T;
|
cannam@167
|
1519 T4X = T4V + T4W;
|
cannam@167
|
1520 T62 = T4U - T4X;
|
cannam@167
|
1521 {
|
cannam@167
|
1522 E T43, T4a, T7t, T7u;
|
cannam@167
|
1523 T43 = T3X + T42;
|
cannam@167
|
1524 T4a = T46 + T49;
|
cannam@167
|
1525 T4b = T43 + T4a;
|
cannam@167
|
1526 T6c = T4a - T43;
|
cannam@167
|
1527 T7t = T5e - T5d;
|
cannam@167
|
1528 T7u = T46 - T49;
|
cannam@167
|
1529 T7v = T7t + T7u;
|
cannam@167
|
1530 T7C = T7t - T7u;
|
cannam@167
|
1531 }
|
cannam@167
|
1532 {
|
cannam@167
|
1533 E T5c, T5f, T72, T73;
|
cannam@167
|
1534 T5c = T5a + T5b;
|
cannam@167
|
1535 T5f = T5d + T5e;
|
cannam@167
|
1536 T5g = T5c + T5f;
|
cannam@167
|
1537 T6f = T5f - T5c;
|
cannam@167
|
1538 T72 = T4S - T4T;
|
cannam@167
|
1539 T73 = TI - TV;
|
cannam@167
|
1540 T74 = T72 + T73;
|
cannam@167
|
1541 T8G = T72 - T73;
|
cannam@167
|
1542 }
|
cannam@167
|
1543 {
|
cannam@167
|
1544 E T7q, T7r, T6Z, T70;
|
cannam@167
|
1545 T7q = T42 - T3X;
|
cannam@167
|
1546 T7r = T5a - T5b;
|
cannam@167
|
1547 T7s = T7q - T7r;
|
cannam@167
|
1548 T7B = T7r + T7q;
|
cannam@167
|
1549 T6Z = Th - Tu;
|
cannam@167
|
1550 T70 = T4V - T4W;
|
cannam@167
|
1551 T71 = T6Z - T70;
|
cannam@167
|
1552 T8F = T6Z + T70;
|
cannam@167
|
1553 }
|
cannam@167
|
1554 }
|
cannam@167
|
1555 {
|
cannam@167
|
1556 E T1a, T5h, T4k, T4Z, T1Q, T5l, T4y, T53, T1r, T5i, T4p, T50, T1F, T5k, T4t;
|
cannam@167
|
1557 E T52;
|
cannam@167
|
1558 {
|
cannam@167
|
1559 E T13, T4j, T19, T4f;
|
cannam@167
|
1560 {
|
cannam@167
|
1561 E T11, T12, T17, T18;
|
cannam@167
|
1562 T11 = Ip[WS(rs, 15)];
|
cannam@167
|
1563 T12 = Im[WS(rs, 15)];
|
cannam@167
|
1564 T13 = T11 - T12;
|
cannam@167
|
1565 T4j = T11 + T12;
|
cannam@167
|
1566 T17 = Rp[WS(rs, 15)];
|
cannam@167
|
1567 T18 = Rm[WS(rs, 15)];
|
cannam@167
|
1568 T19 = T17 + T18;
|
cannam@167
|
1569 T4f = T17 - T18;
|
cannam@167
|
1570 }
|
cannam@167
|
1571 T1a = FNMS(T16, T19, T10 * T13);
|
cannam@167
|
1572 T5h = FNMS(T4i, T4f, T4e * T4j);
|
cannam@167
|
1573 T4k = FMA(T4e, T4f, T4i * T4j);
|
cannam@167
|
1574 T4Z = FMA(T10, T19, T16 * T13);
|
cannam@167
|
1575 }
|
cannam@167
|
1576 {
|
cannam@167
|
1577 E T1L, T4x, T1P, T4v;
|
cannam@167
|
1578 {
|
cannam@167
|
1579 E T1J, T1K, T1N, T1O;
|
cannam@167
|
1580 T1J = Ip[WS(rs, 11)];
|
cannam@167
|
1581 T1K = Im[WS(rs, 11)];
|
cannam@167
|
1582 T1L = T1J - T1K;
|
cannam@167
|
1583 T4x = T1J + T1K;
|
cannam@167
|
1584 T1N = Rp[WS(rs, 11)];
|
cannam@167
|
1585 T1O = Rm[WS(rs, 11)];
|
cannam@167
|
1586 T1P = T1N + T1O;
|
cannam@167
|
1587 T4v = T1N - T1O;
|
cannam@167
|
1588 }
|
cannam@167
|
1589 T1Q = FNMS(T1M, T1P, T1I * T1L);
|
cannam@167
|
1590 T5l = FNMS(T4w, T4v, T4u * T4x);
|
cannam@167
|
1591 T4y = FMA(T4u, T4v, T4w * T4x);
|
cannam@167
|
1592 T53 = FMA(T1I, T1P, T1M * T1L);
|
cannam@167
|
1593 }
|
cannam@167
|
1594 {
|
cannam@167
|
1595 E T1m, T4o, T1q, T4m;
|
cannam@167
|
1596 {
|
cannam@167
|
1597 E T1k, T1l, T1o, T1p;
|
cannam@167
|
1598 T1k = Ip[WS(rs, 7)];
|
cannam@167
|
1599 T1l = Im[WS(rs, 7)];
|
cannam@167
|
1600 T1m = T1k - T1l;
|
cannam@167
|
1601 T4o = T1k + T1l;
|
cannam@167
|
1602 T1o = Rp[WS(rs, 7)];
|
cannam@167
|
1603 T1p = Rm[WS(rs, 7)];
|
cannam@167
|
1604 T1q = T1o + T1p;
|
cannam@167
|
1605 T4m = T1o - T1p;
|
cannam@167
|
1606 }
|
cannam@167
|
1607 T1r = FNMS(T1n, T1q, T1j * T1m);
|
cannam@167
|
1608 T5i = FNMS(T4n, T4m, T4l * T4o);
|
cannam@167
|
1609 T4p = FMA(T4l, T4m, T4n * T4o);
|
cannam@167
|
1610 T50 = FMA(T1j, T1q, T1n * T1m);
|
cannam@167
|
1611 }
|
cannam@167
|
1612 {
|
cannam@167
|
1613 E T1y, T4s, T1E, T4r;
|
cannam@167
|
1614 {
|
cannam@167
|
1615 E T1w, T1x, T1C, T1D;
|
cannam@167
|
1616 T1w = Ip[WS(rs, 3)];
|
cannam@167
|
1617 T1x = Im[WS(rs, 3)];
|
cannam@167
|
1618 T1y = T1w - T1x;
|
cannam@167
|
1619 T4s = T1w + T1x;
|
cannam@167
|
1620 T1C = Rp[WS(rs, 3)];
|
cannam@167
|
1621 T1D = Rm[WS(rs, 3)];
|
cannam@167
|
1622 T1E = T1C + T1D;
|
cannam@167
|
1623 T4r = T1C - T1D;
|
cannam@167
|
1624 }
|
cannam@167
|
1625 T1F = FNMS(T1B, T1E, T1v * T1y);
|
cannam@167
|
1626 T5k = FNMS(T26, T4r, T23 * T4s);
|
cannam@167
|
1627 T4t = FMA(T23, T4r, T26 * T4s);
|
cannam@167
|
1628 T52 = FMA(T1v, T1E, T1B * T1y);
|
cannam@167
|
1629 }
|
cannam@167
|
1630 T1s = T1a + T1r;
|
cannam@167
|
1631 T1R = T1F + T1Q;
|
cannam@167
|
1632 T65 = T1s - T1R;
|
cannam@167
|
1633 T51 = T4Z + T50;
|
cannam@167
|
1634 T54 = T52 + T53;
|
cannam@167
|
1635 T64 = T51 - T54;
|
cannam@167
|
1636 {
|
cannam@167
|
1637 E T4q, T4z, T7E, T7F;
|
cannam@167
|
1638 T4q = T4k + T4p;
|
cannam@167
|
1639 T4z = T4t + T4y;
|
cannam@167
|
1640 T4A = T4q + T4z;
|
cannam@167
|
1641 T6g = T4z - T4q;
|
cannam@167
|
1642 T7E = T5h - T5i;
|
cannam@167
|
1643 T7F = T4y - T4t;
|
cannam@167
|
1644 T7G = T7E + T7F;
|
cannam@167
|
1645 T8U = T7E - T7F;
|
cannam@167
|
1646 }
|
cannam@167
|
1647 {
|
cannam@167
|
1648 E T5j, T5m, T79, T7a;
|
cannam@167
|
1649 T5j = T5h + T5i;
|
cannam@167
|
1650 T5m = T5k + T5l;
|
cannam@167
|
1651 T5n = T5j + T5m;
|
cannam@167
|
1652 T6d = T5j - T5m;
|
cannam@167
|
1653 T79 = T4Z - T50;
|
cannam@167
|
1654 T7a = T1F - T1Q;
|
cannam@167
|
1655 T7b = T79 + T7a;
|
cannam@167
|
1656 T8J = T79 - T7a;
|
cannam@167
|
1657 }
|
cannam@167
|
1658 {
|
cannam@167
|
1659 E T7x, T7y, T76, T77;
|
cannam@167
|
1660 T7x = T4p - T4k;
|
cannam@167
|
1661 T7y = T5k - T5l;
|
cannam@167
|
1662 T7z = T7x - T7y;
|
cannam@167
|
1663 T8R = T7x + T7y;
|
cannam@167
|
1664 T76 = T1a - T1r;
|
cannam@167
|
1665 T77 = T52 - T53;
|
cannam@167
|
1666 T78 = T76 - T77;
|
cannam@167
|
1667 T8I = T76 + T77;
|
cannam@167
|
1668 }
|
cannam@167
|
1669 }
|
cannam@167
|
1670 {
|
cannam@167
|
1671 E T1T, T5S, T5M, T5W, T5P, T5X, T3a, T5I, T4C, T58, T56, T5H, T5E, T5G, T4R;
|
cannam@167
|
1672 E T5R;
|
cannam@167
|
1673 {
|
cannam@167
|
1674 E TX, T1S, T5K, T5L;
|
cannam@167
|
1675 TX = Tv + TW;
|
cannam@167
|
1676 T1S = T1s + T1R;
|
cannam@167
|
1677 T1T = TX + T1S;
|
cannam@167
|
1678 T5S = TX - T1S;
|
cannam@167
|
1679 T5K = T5n - T5g;
|
cannam@167
|
1680 T5L = T4b - T4A;
|
cannam@167
|
1681 T5M = T5K + T5L;
|
cannam@167
|
1682 T5W = T5K - T5L;
|
cannam@167
|
1683 }
|
cannam@167
|
1684 {
|
cannam@167
|
1685 E T5N, T5O, T2A, T39;
|
cannam@167
|
1686 T5N = T3T + T3A;
|
cannam@167
|
1687 T5O = T5C - T5v;
|
cannam@167
|
1688 T5P = T5N - T5O;
|
cannam@167
|
1689 T5X = T5N + T5O;
|
cannam@167
|
1690 T2A = T2g + T2z;
|
cannam@167
|
1691 T39 = T2T + T38;
|
cannam@167
|
1692 T3a = T2A + T39;
|
cannam@167
|
1693 T5I = T39 - T2A;
|
cannam@167
|
1694 }
|
cannam@167
|
1695 {
|
cannam@167
|
1696 E T3U, T4B, T4Y, T55;
|
cannam@167
|
1697 T3U = T3A - T3T;
|
cannam@167
|
1698 T4B = T4b + T4A;
|
cannam@167
|
1699 T4C = T3U - T4B;
|
cannam@167
|
1700 T58 = T4B + T3U;
|
cannam@167
|
1701 T4Y = T4U + T4X;
|
cannam@167
|
1702 T55 = T51 + T54;
|
cannam@167
|
1703 T56 = T4Y + T55;
|
cannam@167
|
1704 T5H = T55 - T4Y;
|
cannam@167
|
1705 }
|
cannam@167
|
1706 {
|
cannam@167
|
1707 E T5o, T5D, T4J, T4Q;
|
cannam@167
|
1708 T5o = T5g + T5n;
|
cannam@167
|
1709 T5D = T5v + T5C;
|
cannam@167
|
1710 T5E = T5o - T5D;
|
cannam@167
|
1711 T5G = T5o + T5D;
|
cannam@167
|
1712 T4J = T4F + T4I;
|
cannam@167
|
1713 T4Q = T4M + T4P;
|
cannam@167
|
1714 T4R = T4J + T4Q;
|
cannam@167
|
1715 T5R = T4J - T4Q;
|
cannam@167
|
1716 }
|
cannam@167
|
1717 {
|
cannam@167
|
1718 E T3b, T5F, T57, T59;
|
cannam@167
|
1719 T3b = T1T + T3a;
|
cannam@167
|
1720 Ip[0] = KP500000000 * (T3b + T4C);
|
cannam@167
|
1721 Im[WS(rs, 15)] = KP500000000 * (T4C - T3b);
|
cannam@167
|
1722 T5F = T4R + T56;
|
cannam@167
|
1723 Rm[WS(rs, 15)] = KP500000000 * (T5F - T5G);
|
cannam@167
|
1724 Rp[0] = KP500000000 * (T5F + T5G);
|
cannam@167
|
1725 T57 = T4R - T56;
|
cannam@167
|
1726 Rm[WS(rs, 7)] = KP500000000 * (T57 - T58);
|
cannam@167
|
1727 Rp[WS(rs, 8)] = KP500000000 * (T57 + T58);
|
cannam@167
|
1728 T59 = T3a - T1T;
|
cannam@167
|
1729 Ip[WS(rs, 8)] = KP500000000 * (T59 + T5E);
|
cannam@167
|
1730 Im[WS(rs, 7)] = KP500000000 * (T5E - T59);
|
cannam@167
|
1731 }
|
cannam@167
|
1732 {
|
cannam@167
|
1733 E T5J, T5Q, T5Z, T60;
|
cannam@167
|
1734 T5J = KP500000000 * (T5H + T5I);
|
cannam@167
|
1735 T5Q = KP353553390 * (T5M + T5P);
|
cannam@167
|
1736 Ip[WS(rs, 4)] = T5J + T5Q;
|
cannam@167
|
1737 Im[WS(rs, 11)] = T5Q - T5J;
|
cannam@167
|
1738 T5Z = KP500000000 * (T5R + T5S);
|
cannam@167
|
1739 T60 = KP353553390 * (T5W + T5X);
|
cannam@167
|
1740 Rm[WS(rs, 11)] = T5Z - T60;
|
cannam@167
|
1741 Rp[WS(rs, 4)] = T5Z + T60;
|
cannam@167
|
1742 }
|
cannam@167
|
1743 {
|
cannam@167
|
1744 E T5T, T5U, T5V, T5Y;
|
cannam@167
|
1745 T5T = KP500000000 * (T5R - T5S);
|
cannam@167
|
1746 T5U = KP353553390 * (T5P - T5M);
|
cannam@167
|
1747 Rm[WS(rs, 3)] = T5T - T5U;
|
cannam@167
|
1748 Rp[WS(rs, 12)] = T5T + T5U;
|
cannam@167
|
1749 T5V = KP500000000 * (T5I - T5H);
|
cannam@167
|
1750 T5Y = KP353553390 * (T5W - T5X);
|
cannam@167
|
1751 Ip[WS(rs, 12)] = T5V + T5Y;
|
cannam@167
|
1752 Im[WS(rs, 3)] = T5Y - T5V;
|
cannam@167
|
1753 }
|
cannam@167
|
1754 }
|
cannam@167
|
1755 {
|
cannam@167
|
1756 E T67, T6Q, T6K, T6U, T6N, T6V, T6a, T6G, T6i, T6A, T6t, T6P, T6w, T6F, T6p;
|
cannam@167
|
1757 E T6B;
|
cannam@167
|
1758 {
|
cannam@167
|
1759 E T63, T66, T6I, T6J;
|
cannam@167
|
1760 T63 = T61 - T62;
|
cannam@167
|
1761 T66 = T64 + T65;
|
cannam@167
|
1762 T67 = KP353553390 * (T63 + T66);
|
cannam@167
|
1763 T6Q = KP353553390 * (T63 - T66);
|
cannam@167
|
1764 T6I = T6d - T6c;
|
cannam@167
|
1765 T6J = T6g - T6f;
|
cannam@167
|
1766 T6K = FMA(KP461939766, T6I, KP191341716 * T6J);
|
cannam@167
|
1767 T6U = FNMS(KP461939766, T6J, KP191341716 * T6I);
|
cannam@167
|
1768 }
|
cannam@167
|
1769 {
|
cannam@167
|
1770 E T6L, T6M, T68, T69;
|
cannam@167
|
1771 T6L = T6k - T6j;
|
cannam@167
|
1772 T6M = T6n - T6m;
|
cannam@167
|
1773 T6N = FNMS(KP461939766, T6M, KP191341716 * T6L);
|
cannam@167
|
1774 T6V = FMA(KP461939766, T6L, KP191341716 * T6M);
|
cannam@167
|
1775 T68 = T4P - T4M;
|
cannam@167
|
1776 T69 = T38 - T2T;
|
cannam@167
|
1777 T6a = KP500000000 * (T68 + T69);
|
cannam@167
|
1778 T6G = KP500000000 * (T69 - T68);
|
cannam@167
|
1779 }
|
cannam@167
|
1780 {
|
cannam@167
|
1781 E T6e, T6h, T6r, T6s;
|
cannam@167
|
1782 T6e = T6c + T6d;
|
cannam@167
|
1783 T6h = T6f + T6g;
|
cannam@167
|
1784 T6i = FMA(KP191341716, T6e, KP461939766 * T6h);
|
cannam@167
|
1785 T6A = FNMS(KP191341716, T6h, KP461939766 * T6e);
|
cannam@167
|
1786 T6r = T4F - T4I;
|
cannam@167
|
1787 T6s = T2g - T2z;
|
cannam@167
|
1788 T6t = KP500000000 * (T6r + T6s);
|
cannam@167
|
1789 T6P = KP500000000 * (T6r - T6s);
|
cannam@167
|
1790 }
|
cannam@167
|
1791 {
|
cannam@167
|
1792 E T6u, T6v, T6l, T6o;
|
cannam@167
|
1793 T6u = T62 + T61;
|
cannam@167
|
1794 T6v = T64 - T65;
|
cannam@167
|
1795 T6w = KP353553390 * (T6u + T6v);
|
cannam@167
|
1796 T6F = KP353553390 * (T6v - T6u);
|
cannam@167
|
1797 T6l = T6j + T6k;
|
cannam@167
|
1798 T6o = T6m + T6n;
|
cannam@167
|
1799 T6p = FNMS(KP191341716, T6o, KP461939766 * T6l);
|
cannam@167
|
1800 T6B = FMA(KP191341716, T6l, KP461939766 * T6o);
|
cannam@167
|
1801 }
|
cannam@167
|
1802 {
|
cannam@167
|
1803 E T6b, T6q, T6D, T6E;
|
cannam@167
|
1804 T6b = T67 + T6a;
|
cannam@167
|
1805 T6q = T6i + T6p;
|
cannam@167
|
1806 Ip[WS(rs, 2)] = T6b + T6q;
|
cannam@167
|
1807 Im[WS(rs, 13)] = T6q - T6b;
|
cannam@167
|
1808 T6D = T6t + T6w;
|
cannam@167
|
1809 T6E = T6A + T6B;
|
cannam@167
|
1810 Rm[WS(rs, 13)] = T6D - T6E;
|
cannam@167
|
1811 Rp[WS(rs, 2)] = T6D + T6E;
|
cannam@167
|
1812 }
|
cannam@167
|
1813 {
|
cannam@167
|
1814 E T6x, T6y, T6z, T6C;
|
cannam@167
|
1815 T6x = T6t - T6w;
|
cannam@167
|
1816 T6y = T6p - T6i;
|
cannam@167
|
1817 Rm[WS(rs, 5)] = T6x - T6y;
|
cannam@167
|
1818 Rp[WS(rs, 10)] = T6x + T6y;
|
cannam@167
|
1819 T6z = T6a - T67;
|
cannam@167
|
1820 T6C = T6A - T6B;
|
cannam@167
|
1821 Ip[WS(rs, 10)] = T6z + T6C;
|
cannam@167
|
1822 Im[WS(rs, 5)] = T6C - T6z;
|
cannam@167
|
1823 }
|
cannam@167
|
1824 {
|
cannam@167
|
1825 E T6H, T6O, T6X, T6Y;
|
cannam@167
|
1826 T6H = T6F + T6G;
|
cannam@167
|
1827 T6O = T6K + T6N;
|
cannam@167
|
1828 Ip[WS(rs, 6)] = T6H + T6O;
|
cannam@167
|
1829 Im[WS(rs, 9)] = T6O - T6H;
|
cannam@167
|
1830 T6X = T6P + T6Q;
|
cannam@167
|
1831 T6Y = T6U + T6V;
|
cannam@167
|
1832 Rm[WS(rs, 9)] = T6X - T6Y;
|
cannam@167
|
1833 Rp[WS(rs, 6)] = T6X + T6Y;
|
cannam@167
|
1834 }
|
cannam@167
|
1835 {
|
cannam@167
|
1836 E T6R, T6S, T6T, T6W;
|
cannam@167
|
1837 T6R = T6P - T6Q;
|
cannam@167
|
1838 T6S = T6N - T6K;
|
cannam@167
|
1839 Rm[WS(rs, 1)] = T6R - T6S;
|
cannam@167
|
1840 Rp[WS(rs, 14)] = T6R + T6S;
|
cannam@167
|
1841 T6T = T6G - T6F;
|
cannam@167
|
1842 T6W = T6U - T6V;
|
cannam@167
|
1843 Ip[WS(rs, 14)] = T6T + T6W;
|
cannam@167
|
1844 Im[WS(rs, 1)] = T6W - T6T;
|
cannam@167
|
1845 }
|
cannam@167
|
1846 }
|
cannam@167
|
1847 {
|
cannam@167
|
1848 E T7d, T8w, T7o, T8m, T8c, T8l, T89, T8v, T81, T8B, T8h, T8t, T7I, T8A, T8g;
|
cannam@167
|
1849 E T8q;
|
cannam@167
|
1850 {
|
cannam@167
|
1851 E T75, T7c, T85, T88;
|
cannam@167
|
1852 T75 = FNMS(KP191341716, T74, KP461939766 * T71);
|
cannam@167
|
1853 T7c = FMA(KP461939766, T78, KP191341716 * T7b);
|
cannam@167
|
1854 T7d = T75 + T7c;
|
cannam@167
|
1855 T8w = T75 - T7c;
|
cannam@167
|
1856 {
|
cannam@167
|
1857 E T7k, T7n, T8a, T8b;
|
cannam@167
|
1858 T7k = KP353553390 * (T7g + T7j);
|
cannam@167
|
1859 T7n = KP500000000 * (T7l - T7m);
|
cannam@167
|
1860 T7o = T7k + T7n;
|
cannam@167
|
1861 T8m = T7n - T7k;
|
cannam@167
|
1862 T8a = FMA(KP191341716, T71, KP461939766 * T74);
|
cannam@167
|
1863 T8b = FNMS(KP191341716, T78, KP461939766 * T7b);
|
cannam@167
|
1864 T8c = T8a + T8b;
|
cannam@167
|
1865 T8l = T8b - T8a;
|
cannam@167
|
1866 }
|
cannam@167
|
1867 T85 = KP500000000 * (T83 + T84);
|
cannam@167
|
1868 T88 = KP353553390 * (T86 + T87);
|
cannam@167
|
1869 T89 = T85 + T88;
|
cannam@167
|
1870 T8v = T85 - T88;
|
cannam@167
|
1871 {
|
cannam@167
|
1872 E T7T, T8r, T80, T8s, T7P, T7W;
|
cannam@167
|
1873 T7P = KP707106781 * (T7L + T7O);
|
cannam@167
|
1874 T7T = T7P + T7S;
|
cannam@167
|
1875 T8r = T7S - T7P;
|
cannam@167
|
1876 T7W = KP707106781 * (T7U + T7V);
|
cannam@167
|
1877 T80 = T7W + T7Z;
|
cannam@167
|
1878 T8s = T7Z - T7W;
|
cannam@167
|
1879 T81 = FNMS(KP097545161, T80, KP490392640 * T7T);
|
cannam@167
|
1880 T8B = FMA(KP415734806, T8r, KP277785116 * T8s);
|
cannam@167
|
1881 T8h = FMA(KP097545161, T7T, KP490392640 * T80);
|
cannam@167
|
1882 T8t = FNMS(KP415734806, T8s, KP277785116 * T8r);
|
cannam@167
|
1883 }
|
cannam@167
|
1884 {
|
cannam@167
|
1885 E T7A, T8o, T7H, T8p, T7w, T7D;
|
cannam@167
|
1886 T7w = KP707106781 * (T7s + T7v);
|
cannam@167
|
1887 T7A = T7w + T7z;
|
cannam@167
|
1888 T8o = T7z - T7w;
|
cannam@167
|
1889 T7D = KP707106781 * (T7B + T7C);
|
cannam@167
|
1890 T7H = T7D + T7G;
|
cannam@167
|
1891 T8p = T7G - T7D;
|
cannam@167
|
1892 T7I = FMA(KP490392640, T7A, KP097545161 * T7H);
|
cannam@167
|
1893 T8A = FNMS(KP415734806, T8o, KP277785116 * T8p);
|
cannam@167
|
1894 T8g = FNMS(KP097545161, T7A, KP490392640 * T7H);
|
cannam@167
|
1895 T8q = FMA(KP277785116, T8o, KP415734806 * T8p);
|
cannam@167
|
1896 }
|
cannam@167
|
1897 }
|
cannam@167
|
1898 {
|
cannam@167
|
1899 E T7p, T82, T8j, T8k;
|
cannam@167
|
1900 T7p = T7d + T7o;
|
cannam@167
|
1901 T82 = T7I + T81;
|
cannam@167
|
1902 Ip[WS(rs, 1)] = T7p + T82;
|
cannam@167
|
1903 Im[WS(rs, 14)] = T82 - T7p;
|
cannam@167
|
1904 T8j = T89 + T8c;
|
cannam@167
|
1905 T8k = T8g + T8h;
|
cannam@167
|
1906 Rm[WS(rs, 14)] = T8j - T8k;
|
cannam@167
|
1907 Rp[WS(rs, 1)] = T8j + T8k;
|
cannam@167
|
1908 }
|
cannam@167
|
1909 {
|
cannam@167
|
1910 E T8d, T8e, T8f, T8i;
|
cannam@167
|
1911 T8d = T89 - T8c;
|
cannam@167
|
1912 T8e = T81 - T7I;
|
cannam@167
|
1913 Rm[WS(rs, 6)] = T8d - T8e;
|
cannam@167
|
1914 Rp[WS(rs, 9)] = T8d + T8e;
|
cannam@167
|
1915 T8f = T7o - T7d;
|
cannam@167
|
1916 T8i = T8g - T8h;
|
cannam@167
|
1917 Ip[WS(rs, 9)] = T8f + T8i;
|
cannam@167
|
1918 Im[WS(rs, 6)] = T8i - T8f;
|
cannam@167
|
1919 }
|
cannam@167
|
1920 {
|
cannam@167
|
1921 E T8n, T8u, T8D, T8E;
|
cannam@167
|
1922 T8n = T8l + T8m;
|
cannam@167
|
1923 T8u = T8q + T8t;
|
cannam@167
|
1924 Ip[WS(rs, 5)] = T8n + T8u;
|
cannam@167
|
1925 Im[WS(rs, 10)] = T8u - T8n;
|
cannam@167
|
1926 T8D = T8v + T8w;
|
cannam@167
|
1927 T8E = T8A + T8B;
|
cannam@167
|
1928 Rm[WS(rs, 10)] = T8D - T8E;
|
cannam@167
|
1929 Rp[WS(rs, 5)] = T8D + T8E;
|
cannam@167
|
1930 }
|
cannam@167
|
1931 {
|
cannam@167
|
1932 E T8x, T8y, T8z, T8C;
|
cannam@167
|
1933 T8x = T8v - T8w;
|
cannam@167
|
1934 T8y = T8t - T8q;
|
cannam@167
|
1935 Rm[WS(rs, 2)] = T8x - T8y;
|
cannam@167
|
1936 Rp[WS(rs, 13)] = T8x + T8y;
|
cannam@167
|
1937 T8z = T8m - T8l;
|
cannam@167
|
1938 T8C = T8A - T8B;
|
cannam@167
|
1939 Ip[WS(rs, 13)] = T8z + T8C;
|
cannam@167
|
1940 Im[WS(rs, 2)] = T8C - T8z;
|
cannam@167
|
1941 }
|
cannam@167
|
1942 }
|
cannam@167
|
1943 {
|
cannam@167
|
1944 E T8L, T9u, T8O, T9k, T9a, T9j, T97, T9t, T93, T9z, T9f, T9r, T8W, T9y, T9e;
|
cannam@167
|
1945 E T9o;
|
cannam@167
|
1946 {
|
cannam@167
|
1947 E T8H, T8K, T95, T96;
|
cannam@167
|
1948 T8H = FNMS(KP461939766, T8G, KP191341716 * T8F);
|
cannam@167
|
1949 T8K = FMA(KP191341716, T8I, KP461939766 * T8J);
|
cannam@167
|
1950 T8L = T8H + T8K;
|
cannam@167
|
1951 T9u = T8H - T8K;
|
cannam@167
|
1952 {
|
cannam@167
|
1953 E T8M, T8N, T98, T99;
|
cannam@167
|
1954 T8M = KP353553390 * (T87 - T86);
|
cannam@167
|
1955 T8N = KP500000000 * (T7m + T7l);
|
cannam@167
|
1956 T8O = T8M + T8N;
|
cannam@167
|
1957 T9k = T8N - T8M;
|
cannam@167
|
1958 T98 = FMA(KP461939766, T8F, KP191341716 * T8G);
|
cannam@167
|
1959 T99 = FNMS(KP461939766, T8I, KP191341716 * T8J);
|
cannam@167
|
1960 T9a = T98 + T99;
|
cannam@167
|
1961 T9j = T99 - T98;
|
cannam@167
|
1962 }
|
cannam@167
|
1963 T95 = KP500000000 * (T83 - T84);
|
cannam@167
|
1964 T96 = KP353553390 * (T7g - T7j);
|
cannam@167
|
1965 T97 = T95 + T96;
|
cannam@167
|
1966 T9t = T95 - T96;
|
cannam@167
|
1967 {
|
cannam@167
|
1968 E T8Z, T9p, T92, T9q, T8X, T90;
|
cannam@167
|
1969 T8X = KP707106781 * (T7V - T7U);
|
cannam@167
|
1970 T8Z = T8X + T8Y;
|
cannam@167
|
1971 T9p = T8Y - T8X;
|
cannam@167
|
1972 T90 = KP707106781 * (T7L - T7O);
|
cannam@167
|
1973 T92 = T90 + T91;
|
cannam@167
|
1974 T9q = T91 - T90;
|
cannam@167
|
1975 T93 = FNMS(KP277785116, T92, KP415734806 * T8Z);
|
cannam@167
|
1976 T9z = FMA(KP490392640, T9p, KP097545161 * T9q);
|
cannam@167
|
1977 T9f = FMA(KP277785116, T8Z, KP415734806 * T92);
|
cannam@167
|
1978 T9r = FNMS(KP490392640, T9q, KP097545161 * T9p);
|
cannam@167
|
1979 }
|
cannam@167
|
1980 {
|
cannam@167
|
1981 E T8S, T9m, T8V, T9n, T8Q, T8T;
|
cannam@167
|
1982 T8Q = KP707106781 * (T7C - T7B);
|
cannam@167
|
1983 T8S = T8Q + T8R;
|
cannam@167
|
1984 T9m = T8R - T8Q;
|
cannam@167
|
1985 T8T = KP707106781 * (T7s - T7v);
|
cannam@167
|
1986 T8V = T8T + T8U;
|
cannam@167
|
1987 T9n = T8U - T8T;
|
cannam@167
|
1988 T8W = FMA(KP415734806, T8S, KP277785116 * T8V);
|
cannam@167
|
1989 T9y = FNMS(KP490392640, T9m, KP097545161 * T9n);
|
cannam@167
|
1990 T9e = FNMS(KP277785116, T8S, KP415734806 * T8V);
|
cannam@167
|
1991 T9o = FMA(KP097545161, T9m, KP490392640 * T9n);
|
cannam@167
|
1992 }
|
cannam@167
|
1993 }
|
cannam@167
|
1994 {
|
cannam@167
|
1995 E T8P, T94, T9h, T9i;
|
cannam@167
|
1996 T8P = T8L + T8O;
|
cannam@167
|
1997 T94 = T8W + T93;
|
cannam@167
|
1998 Ip[WS(rs, 3)] = T8P + T94;
|
cannam@167
|
1999 Im[WS(rs, 12)] = T94 - T8P;
|
cannam@167
|
2000 T9h = T97 + T9a;
|
cannam@167
|
2001 T9i = T9e + T9f;
|
cannam@167
|
2002 Rm[WS(rs, 12)] = T9h - T9i;
|
cannam@167
|
2003 Rp[WS(rs, 3)] = T9h + T9i;
|
cannam@167
|
2004 }
|
cannam@167
|
2005 {
|
cannam@167
|
2006 E T9b, T9c, T9d, T9g;
|
cannam@167
|
2007 T9b = T97 - T9a;
|
cannam@167
|
2008 T9c = T93 - T8W;
|
cannam@167
|
2009 Rm[WS(rs, 4)] = T9b - T9c;
|
cannam@167
|
2010 Rp[WS(rs, 11)] = T9b + T9c;
|
cannam@167
|
2011 T9d = T8O - T8L;
|
cannam@167
|
2012 T9g = T9e - T9f;
|
cannam@167
|
2013 Ip[WS(rs, 11)] = T9d + T9g;
|
cannam@167
|
2014 Im[WS(rs, 4)] = T9g - T9d;
|
cannam@167
|
2015 }
|
cannam@167
|
2016 {
|
cannam@167
|
2017 E T9l, T9s, T9B, T9C;
|
cannam@167
|
2018 T9l = T9j + T9k;
|
cannam@167
|
2019 T9s = T9o + T9r;
|
cannam@167
|
2020 Ip[WS(rs, 7)] = T9l + T9s;
|
cannam@167
|
2021 Im[WS(rs, 8)] = T9s - T9l;
|
cannam@167
|
2022 T9B = T9t + T9u;
|
cannam@167
|
2023 T9C = T9y + T9z;
|
cannam@167
|
2024 Rm[WS(rs, 8)] = T9B - T9C;
|
cannam@167
|
2025 Rp[WS(rs, 7)] = T9B + T9C;
|
cannam@167
|
2026 }
|
cannam@167
|
2027 {
|
cannam@167
|
2028 E T9v, T9w, T9x, T9A;
|
cannam@167
|
2029 T9v = T9t - T9u;
|
cannam@167
|
2030 T9w = T9r - T9o;
|
cannam@167
|
2031 Rm[0] = T9v - T9w;
|
cannam@167
|
2032 Rp[WS(rs, 15)] = T9v + T9w;
|
cannam@167
|
2033 T9x = T9k - T9j;
|
cannam@167
|
2034 T9A = T9y - T9z;
|
cannam@167
|
2035 Ip[WS(rs, 15)] = T9x + T9A;
|
cannam@167
|
2036 Im[0] = T9A - T9x;
|
cannam@167
|
2037 }
|
cannam@167
|
2038 }
|
cannam@167
|
2039 }
|
cannam@167
|
2040 }
|
cannam@167
|
2041 }
|
cannam@167
|
2042 }
|
cannam@167
|
2043
|
cannam@167
|
2044 static const tw_instr twinstr[] = {
|
cannam@167
|
2045 {TW_CEXP, 1, 1},
|
cannam@167
|
2046 {TW_CEXP, 1, 3},
|
cannam@167
|
2047 {TW_CEXP, 1, 9},
|
cannam@167
|
2048 {TW_CEXP, 1, 27},
|
cannam@167
|
2049 {TW_NEXT, 1, 0}
|
cannam@167
|
2050 };
|
cannam@167
|
2051
|
cannam@167
|
2052 static const hc2c_desc desc = { 32, "hc2cfdft2_32", twinstr, &GENUS, {440, 188, 112, 0} };
|
cannam@167
|
2053
|
cannam@167
|
2054 void X(codelet_hc2cfdft2_32) (planner *p) {
|
cannam@167
|
2055 X(khc2c_register) (p, hc2cfdft2_32, &desc, HC2C_VIA_DFT);
|
cannam@167
|
2056 }
|
cannam@167
|
2057 #endif
|