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 07:36:24 EST 2012 */
|
Chris@10
|
23
|
Chris@10
|
24 #include "codelet-dft.h"
|
Chris@10
|
25
|
Chris@10
|
26 #ifdef HAVE_FMA
|
Chris@10
|
27
|
Chris@10
|
28 /* Generated by: ../../../genfft/gen_twidsq.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -reload-twiddle -dif -n 6 -name q1_6 -include q.h */
|
Chris@10
|
29
|
Chris@10
|
30 /*
|
Chris@10
|
31 * This function contains 276 FP additions, 192 FP multiplications,
|
Chris@10
|
32 * (or, 144 additions, 60 multiplications, 132 fused multiply/add),
|
Chris@10
|
33 * 129 stack variables, 2 constants, and 144 memory accesses
|
Chris@10
|
34 */
|
Chris@10
|
35 #include "q.h"
|
Chris@10
|
36
|
Chris@10
|
37 static void q1_6(R *rio, R *iio, const R *W, stride rs, stride vs, INT mb, INT me, INT ms)
|
Chris@10
|
38 {
|
Chris@10
|
39 DK(KP866025403, +0.866025403784438646763723170752936183471402627);
|
Chris@10
|
40 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
|
Chris@10
|
41 {
|
Chris@10
|
42 INT m;
|
Chris@10
|
43 for (m = mb, W = W + (mb * 10); m < me; m = m + 1, rio = rio + ms, iio = iio + ms, W = W + 10, MAKE_VOLATILE_STRIDE(12, rs), MAKE_VOLATILE_STRIDE(0, vs)) {
|
Chris@10
|
44 E T4c, T4f, T4e, T4g, T4d;
|
Chris@10
|
45 {
|
Chris@10
|
46 E T3, Tw, Ta, TW, Tg, TG, TM, TT, TU, TP, Tn, T17, TV, TJ, Tv;
|
Chris@10
|
47 E T1A, T1e, T20, T1k, T1K, T1Q, T1X, T1Y, T1T, T1r, T1Z, T1N, T1z, T31, T32;
|
Chris@10
|
48 E T2X, T2v, T2b, T33, T2R, T2D, T2E, T2i, T34, T3f, T2o, T2O, T2U, T3I, T3m;
|
Chris@10
|
49 E T48, T3s, T3S, T3Y, T45, T46, T41, T3z, T4j, T47, T3V, T3H, T4M, T4q, T5c;
|
Chris@10
|
50 E T4w, T4W, T52, T59, T5a, T55, T4D, T5b, T4Z, T4L, T6d, T5r, T6e, T69, T5H;
|
Chris@10
|
51 E T5w, T5n, T6f, T63, T5P, T5s, T5o, T5p;
|
Chris@10
|
52 {
|
Chris@10
|
53 E T2f, T2k, T2g, T2c, T2d;
|
Chris@10
|
54 {
|
Chris@10
|
55 E T1b, T1g, T1c, T18, T19;
|
Chris@10
|
56 {
|
Chris@10
|
57 E T4, Tc, Te, T9, T5;
|
Chris@10
|
58 {
|
Chris@10
|
59 E T1, T2, T7, T8;
|
Chris@10
|
60 T1 = rio[0];
|
Chris@10
|
61 T2 = rio[WS(rs, 3)];
|
Chris@10
|
62 T7 = rio[WS(rs, 4)];
|
Chris@10
|
63 T8 = rio[WS(rs, 1)];
|
Chris@10
|
64 T4 = rio[WS(rs, 2)];
|
Chris@10
|
65 Tc = T1 - T2;
|
Chris@10
|
66 T3 = T1 + T2;
|
Chris@10
|
67 Te = T7 - T8;
|
Chris@10
|
68 T9 = T7 + T8;
|
Chris@10
|
69 T5 = rio[WS(rs, 5)];
|
Chris@10
|
70 }
|
Chris@10
|
71 {
|
Chris@10
|
72 E TN, Tj, Tk, Tl, Tt, Th, Ti;
|
Chris@10
|
73 Th = iio[WS(rs, 2)];
|
Chris@10
|
74 Ti = iio[WS(rs, 5)];
|
Chris@10
|
75 {
|
Chris@10
|
76 E Tr, Ts, Td, T6, Tf;
|
Chris@10
|
77 Tr = iio[0];
|
Chris@10
|
78 Td = T4 - T5;
|
Chris@10
|
79 T6 = T4 + T5;
|
Chris@10
|
80 TN = Th + Ti;
|
Chris@10
|
81 Tj = Th - Ti;
|
Chris@10
|
82 Tf = Td + Te;
|
Chris@10
|
83 Tw = Te - Td;
|
Chris@10
|
84 Ta = T6 + T9;
|
Chris@10
|
85 TW = T9 - T6;
|
Chris@10
|
86 Tg = FNMS(KP500000000, Tf, Tc);
|
Chris@10
|
87 TG = Tc + Tf;
|
Chris@10
|
88 Ts = iio[WS(rs, 3)];
|
Chris@10
|
89 TM = FNMS(KP500000000, Ta, T3);
|
Chris@10
|
90 Tk = iio[WS(rs, 4)];
|
Chris@10
|
91 Tl = iio[WS(rs, 1)];
|
Chris@10
|
92 Tt = Tr - Ts;
|
Chris@10
|
93 TT = Tr + Ts;
|
Chris@10
|
94 }
|
Chris@10
|
95 {
|
Chris@10
|
96 E T15, TO, Tm, T16, Tu;
|
Chris@10
|
97 T15 = rio[WS(vs, 1)];
|
Chris@10
|
98 TO = Tk + Tl;
|
Chris@10
|
99 Tm = Tk - Tl;
|
Chris@10
|
100 T16 = rio[WS(vs, 1) + WS(rs, 3)];
|
Chris@10
|
101 T1b = rio[WS(vs, 1) + WS(rs, 4)];
|
Chris@10
|
102 TU = TN + TO;
|
Chris@10
|
103 TP = TN - TO;
|
Chris@10
|
104 Tu = Tj + Tm;
|
Chris@10
|
105 Tn = Tj - Tm;
|
Chris@10
|
106 T1g = T15 - T16;
|
Chris@10
|
107 T17 = T15 + T16;
|
Chris@10
|
108 TV = FNMS(KP500000000, TU, TT);
|
Chris@10
|
109 TJ = Tt + Tu;
|
Chris@10
|
110 Tv = FNMS(KP500000000, Tu, Tt);
|
Chris@10
|
111 T1c = rio[WS(vs, 1) + WS(rs, 1)];
|
Chris@10
|
112 T18 = rio[WS(vs, 1) + WS(rs, 2)];
|
Chris@10
|
113 T19 = rio[WS(vs, 1) + WS(rs, 5)];
|
Chris@10
|
114 }
|
Chris@10
|
115 }
|
Chris@10
|
116 }
|
Chris@10
|
117 {
|
Chris@10
|
118 E T1v, T1R, T1n, T1w, T1o, T1p;
|
Chris@10
|
119 {
|
Chris@10
|
120 E T1l, T1i, T1d, T1h, T1a, T1m, T1j;
|
Chris@10
|
121 T1l = iio[WS(vs, 1) + WS(rs, 2)];
|
Chris@10
|
122 T1i = T1b - T1c;
|
Chris@10
|
123 T1d = T1b + T1c;
|
Chris@10
|
124 T1h = T18 - T19;
|
Chris@10
|
125 T1a = T18 + T19;
|
Chris@10
|
126 T1m = iio[WS(vs, 1) + WS(rs, 5)];
|
Chris@10
|
127 T1v = iio[WS(vs, 1)];
|
Chris@10
|
128 T1j = T1h + T1i;
|
Chris@10
|
129 T1A = T1i - T1h;
|
Chris@10
|
130 T1e = T1a + T1d;
|
Chris@10
|
131 T20 = T1d - T1a;
|
Chris@10
|
132 T1R = T1l + T1m;
|
Chris@10
|
133 T1n = T1l - T1m;
|
Chris@10
|
134 T1k = FNMS(KP500000000, T1j, T1g);
|
Chris@10
|
135 T1K = T1g + T1j;
|
Chris@10
|
136 T1Q = FNMS(KP500000000, T1e, T17);
|
Chris@10
|
137 T1w = iio[WS(vs, 1) + WS(rs, 3)];
|
Chris@10
|
138 T1o = iio[WS(vs, 1) + WS(rs, 4)];
|
Chris@10
|
139 T1p = iio[WS(vs, 1) + WS(rs, 1)];
|
Chris@10
|
140 }
|
Chris@10
|
141 {
|
Chris@10
|
142 E T2z, T2V, T2r, T2A, T2s, T2t;
|
Chris@10
|
143 {
|
Chris@10
|
144 E T2p, T1x, T1S, T1q, T2q, T1y;
|
Chris@10
|
145 T2p = iio[WS(vs, 2) + WS(rs, 2)];
|
Chris@10
|
146 T1X = T1v + T1w;
|
Chris@10
|
147 T1x = T1v - T1w;
|
Chris@10
|
148 T1S = T1o + T1p;
|
Chris@10
|
149 T1q = T1o - T1p;
|
Chris@10
|
150 T2q = iio[WS(vs, 2) + WS(rs, 5)];
|
Chris@10
|
151 T2z = iio[WS(vs, 2)];
|
Chris@10
|
152 T1Y = T1R + T1S;
|
Chris@10
|
153 T1T = T1R - T1S;
|
Chris@10
|
154 T1y = T1n + T1q;
|
Chris@10
|
155 T1r = T1n - T1q;
|
Chris@10
|
156 T2V = T2p + T2q;
|
Chris@10
|
157 T2r = T2p - T2q;
|
Chris@10
|
158 T1Z = FNMS(KP500000000, T1Y, T1X);
|
Chris@10
|
159 T1N = T1x + T1y;
|
Chris@10
|
160 T1z = FNMS(KP500000000, T1y, T1x);
|
Chris@10
|
161 T2A = iio[WS(vs, 2) + WS(rs, 3)];
|
Chris@10
|
162 T2s = iio[WS(vs, 2) + WS(rs, 4)];
|
Chris@10
|
163 T2t = iio[WS(vs, 2) + WS(rs, 1)];
|
Chris@10
|
164 }
|
Chris@10
|
165 {
|
Chris@10
|
166 E T29, T2B, T2W, T2u, T2a, T2C;
|
Chris@10
|
167 T29 = rio[WS(vs, 2)];
|
Chris@10
|
168 T31 = T2z + T2A;
|
Chris@10
|
169 T2B = T2z - T2A;
|
Chris@10
|
170 T2W = T2s + T2t;
|
Chris@10
|
171 T2u = T2s - T2t;
|
Chris@10
|
172 T2a = rio[WS(vs, 2) + WS(rs, 3)];
|
Chris@10
|
173 T2f = rio[WS(vs, 2) + WS(rs, 4)];
|
Chris@10
|
174 T32 = T2V + T2W;
|
Chris@10
|
175 T2X = T2V - T2W;
|
Chris@10
|
176 T2C = T2r + T2u;
|
Chris@10
|
177 T2v = T2r - T2u;
|
Chris@10
|
178 T2k = T29 - T2a;
|
Chris@10
|
179 T2b = T29 + T2a;
|
Chris@10
|
180 T33 = FNMS(KP500000000, T32, T31);
|
Chris@10
|
181 T2R = T2B + T2C;
|
Chris@10
|
182 T2D = FNMS(KP500000000, T2C, T2B);
|
Chris@10
|
183 T2g = rio[WS(vs, 2) + WS(rs, 1)];
|
Chris@10
|
184 T2c = rio[WS(vs, 2) + WS(rs, 2)];
|
Chris@10
|
185 T2d = rio[WS(vs, 2) + WS(rs, 5)];
|
Chris@10
|
186 }
|
Chris@10
|
187 }
|
Chris@10
|
188 }
|
Chris@10
|
189 }
|
Chris@10
|
190 {
|
Chris@10
|
191 E T4n, T4s, T4o, T4k, T4l;
|
Chris@10
|
192 {
|
Chris@10
|
193 E T3j, T3o, T3k, T3g, T3h;
|
Chris@10
|
194 {
|
Chris@10
|
195 E T3d, T2m, T2h, T2l, T2e, T3e, T2n;
|
Chris@10
|
196 T3d = rio[WS(vs, 3)];
|
Chris@10
|
197 T2m = T2f - T2g;
|
Chris@10
|
198 T2h = T2f + T2g;
|
Chris@10
|
199 T2l = T2c - T2d;
|
Chris@10
|
200 T2e = T2c + T2d;
|
Chris@10
|
201 T3e = rio[WS(vs, 3) + WS(rs, 3)];
|
Chris@10
|
202 T3j = rio[WS(vs, 3) + WS(rs, 4)];
|
Chris@10
|
203 T2n = T2l + T2m;
|
Chris@10
|
204 T2E = T2m - T2l;
|
Chris@10
|
205 T2i = T2e + T2h;
|
Chris@10
|
206 T34 = T2h - T2e;
|
Chris@10
|
207 T3o = T3d - T3e;
|
Chris@10
|
208 T3f = T3d + T3e;
|
Chris@10
|
209 T2o = FNMS(KP500000000, T2n, T2k);
|
Chris@10
|
210 T2O = T2k + T2n;
|
Chris@10
|
211 T2U = FNMS(KP500000000, T2i, T2b);
|
Chris@10
|
212 T3k = rio[WS(vs, 3) + WS(rs, 1)];
|
Chris@10
|
213 T3g = rio[WS(vs, 3) + WS(rs, 2)];
|
Chris@10
|
214 T3h = rio[WS(vs, 3) + WS(rs, 5)];
|
Chris@10
|
215 }
|
Chris@10
|
216 {
|
Chris@10
|
217 E T3D, T3Z, T3v, T3E, T3w, T3x;
|
Chris@10
|
218 {
|
Chris@10
|
219 E T3t, T3q, T3l, T3p, T3i, T3u, T3r;
|
Chris@10
|
220 T3t = iio[WS(vs, 3) + WS(rs, 2)];
|
Chris@10
|
221 T3q = T3j - T3k;
|
Chris@10
|
222 T3l = T3j + T3k;
|
Chris@10
|
223 T3p = T3g - T3h;
|
Chris@10
|
224 T3i = T3g + T3h;
|
Chris@10
|
225 T3u = iio[WS(vs, 3) + WS(rs, 5)];
|
Chris@10
|
226 T3D = iio[WS(vs, 3)];
|
Chris@10
|
227 T3r = T3p + T3q;
|
Chris@10
|
228 T3I = T3q - T3p;
|
Chris@10
|
229 T3m = T3i + T3l;
|
Chris@10
|
230 T48 = T3l - T3i;
|
Chris@10
|
231 T3Z = T3t + T3u;
|
Chris@10
|
232 T3v = T3t - T3u;
|
Chris@10
|
233 T3s = FNMS(KP500000000, T3r, T3o);
|
Chris@10
|
234 T3S = T3o + T3r;
|
Chris@10
|
235 T3Y = FNMS(KP500000000, T3m, T3f);
|
Chris@10
|
236 T3E = iio[WS(vs, 3) + WS(rs, 3)];
|
Chris@10
|
237 T3w = iio[WS(vs, 3) + WS(rs, 4)];
|
Chris@10
|
238 T3x = iio[WS(vs, 3) + WS(rs, 1)];
|
Chris@10
|
239 }
|
Chris@10
|
240 {
|
Chris@10
|
241 E T4h, T3F, T40, T3y, T4i, T3G;
|
Chris@10
|
242 T4h = rio[WS(vs, 4)];
|
Chris@10
|
243 T45 = T3D + T3E;
|
Chris@10
|
244 T3F = T3D - T3E;
|
Chris@10
|
245 T40 = T3w + T3x;
|
Chris@10
|
246 T3y = T3w - T3x;
|
Chris@10
|
247 T4i = rio[WS(vs, 4) + WS(rs, 3)];
|
Chris@10
|
248 T4n = rio[WS(vs, 4) + WS(rs, 4)];
|
Chris@10
|
249 T46 = T3Z + T40;
|
Chris@10
|
250 T41 = T3Z - T40;
|
Chris@10
|
251 T3G = T3v + T3y;
|
Chris@10
|
252 T3z = T3v - T3y;
|
Chris@10
|
253 T4s = T4h - T4i;
|
Chris@10
|
254 T4j = T4h + T4i;
|
Chris@10
|
255 T47 = FNMS(KP500000000, T46, T45);
|
Chris@10
|
256 T3V = T3F + T3G;
|
Chris@10
|
257 T3H = FNMS(KP500000000, T3G, T3F);
|
Chris@10
|
258 T4o = rio[WS(vs, 4) + WS(rs, 1)];
|
Chris@10
|
259 T4k = rio[WS(vs, 4) + WS(rs, 2)];
|
Chris@10
|
260 T4l = rio[WS(vs, 4) + WS(rs, 5)];
|
Chris@10
|
261 }
|
Chris@10
|
262 }
|
Chris@10
|
263 }
|
Chris@10
|
264 {
|
Chris@10
|
265 E T4H, T53, T4z, T4I, T4A, T4B;
|
Chris@10
|
266 {
|
Chris@10
|
267 E T4x, T4u, T4p, T4t, T4m, T4y, T4v;
|
Chris@10
|
268 T4x = iio[WS(vs, 4) + WS(rs, 2)];
|
Chris@10
|
269 T4u = T4n - T4o;
|
Chris@10
|
270 T4p = T4n + T4o;
|
Chris@10
|
271 T4t = T4k - T4l;
|
Chris@10
|
272 T4m = T4k + T4l;
|
Chris@10
|
273 T4y = iio[WS(vs, 4) + WS(rs, 5)];
|
Chris@10
|
274 T4H = iio[WS(vs, 4)];
|
Chris@10
|
275 T4v = T4t + T4u;
|
Chris@10
|
276 T4M = T4u - T4t;
|
Chris@10
|
277 T4q = T4m + T4p;
|
Chris@10
|
278 T5c = T4p - T4m;
|
Chris@10
|
279 T53 = T4x + T4y;
|
Chris@10
|
280 T4z = T4x - T4y;
|
Chris@10
|
281 T4w = FNMS(KP500000000, T4v, T4s);
|
Chris@10
|
282 T4W = T4s + T4v;
|
Chris@10
|
283 T52 = FNMS(KP500000000, T4q, T4j);
|
Chris@10
|
284 T4I = iio[WS(vs, 4) + WS(rs, 3)];
|
Chris@10
|
285 T4A = iio[WS(vs, 4) + WS(rs, 4)];
|
Chris@10
|
286 T4B = iio[WS(vs, 4) + WS(rs, 1)];
|
Chris@10
|
287 }
|
Chris@10
|
288 {
|
Chris@10
|
289 E T5L, T67, T5D, T5M, T5E, T5F;
|
Chris@10
|
290 {
|
Chris@10
|
291 E T5B, T4J, T54, T4C, T5C, T4K;
|
Chris@10
|
292 T5B = iio[WS(vs, 5) + WS(rs, 2)];
|
Chris@10
|
293 T59 = T4H + T4I;
|
Chris@10
|
294 T4J = T4H - T4I;
|
Chris@10
|
295 T54 = T4A + T4B;
|
Chris@10
|
296 T4C = T4A - T4B;
|
Chris@10
|
297 T5C = iio[WS(vs, 5) + WS(rs, 5)];
|
Chris@10
|
298 T5L = iio[WS(vs, 5)];
|
Chris@10
|
299 T5a = T53 + T54;
|
Chris@10
|
300 T55 = T53 - T54;
|
Chris@10
|
301 T4K = T4z + T4C;
|
Chris@10
|
302 T4D = T4z - T4C;
|
Chris@10
|
303 T67 = T5B + T5C;
|
Chris@10
|
304 T5D = T5B - T5C;
|
Chris@10
|
305 T5b = FNMS(KP500000000, T5a, T59);
|
Chris@10
|
306 T4Z = T4J + T4K;
|
Chris@10
|
307 T4L = FNMS(KP500000000, T4K, T4J);
|
Chris@10
|
308 T5M = iio[WS(vs, 5) + WS(rs, 3)];
|
Chris@10
|
309 T5E = iio[WS(vs, 5) + WS(rs, 4)];
|
Chris@10
|
310 T5F = iio[WS(vs, 5) + WS(rs, 1)];
|
Chris@10
|
311 }
|
Chris@10
|
312 {
|
Chris@10
|
313 E T5l, T5N, T68, T5G, T5m, T5O;
|
Chris@10
|
314 T5l = rio[WS(vs, 5)];
|
Chris@10
|
315 T6d = T5L + T5M;
|
Chris@10
|
316 T5N = T5L - T5M;
|
Chris@10
|
317 T68 = T5E + T5F;
|
Chris@10
|
318 T5G = T5E - T5F;
|
Chris@10
|
319 T5m = rio[WS(vs, 5) + WS(rs, 3)];
|
Chris@10
|
320 T5r = rio[WS(vs, 5) + WS(rs, 4)];
|
Chris@10
|
321 T6e = T67 + T68;
|
Chris@10
|
322 T69 = T67 - T68;
|
Chris@10
|
323 T5O = T5D + T5G;
|
Chris@10
|
324 T5H = T5D - T5G;
|
Chris@10
|
325 T5w = T5l - T5m;
|
Chris@10
|
326 T5n = T5l + T5m;
|
Chris@10
|
327 T6f = FNMS(KP500000000, T6e, T6d);
|
Chris@10
|
328 T63 = T5N + T5O;
|
Chris@10
|
329 T5P = FNMS(KP500000000, T5O, T5N);
|
Chris@10
|
330 T5s = rio[WS(vs, 5) + WS(rs, 1)];
|
Chris@10
|
331 T5o = rio[WS(vs, 5) + WS(rs, 2)];
|
Chris@10
|
332 T5p = rio[WS(vs, 5) + WS(rs, 5)];
|
Chris@10
|
333 }
|
Chris@10
|
334 }
|
Chris@10
|
335 }
|
Chris@10
|
336 }
|
Chris@10
|
337 }
|
Chris@10
|
338 {
|
Chris@10
|
339 E T6a, T6h, T5I, T5R, T65, T6c;
|
Chris@10
|
340 {
|
Chris@10
|
341 E T5Q, T5u, T6g, T5A, T60, T66;
|
Chris@10
|
342 {
|
Chris@10
|
343 E T5y, T5t, T5x, T5q, T5z;
|
Chris@10
|
344 rio[0] = T3 + Ta;
|
Chris@10
|
345 T5y = T5r - T5s;
|
Chris@10
|
346 T5t = T5r + T5s;
|
Chris@10
|
347 T5x = T5o - T5p;
|
Chris@10
|
348 T5q = T5o + T5p;
|
Chris@10
|
349 iio[0] = TT + TU;
|
Chris@10
|
350 rio[WS(rs, 1)] = T17 + T1e;
|
Chris@10
|
351 T5z = T5x + T5y;
|
Chris@10
|
352 T5Q = T5y - T5x;
|
Chris@10
|
353 T5u = T5q + T5t;
|
Chris@10
|
354 T6g = T5t - T5q;
|
Chris@10
|
355 T5A = FNMS(KP500000000, T5z, T5w);
|
Chris@10
|
356 T60 = T5w + T5z;
|
Chris@10
|
357 iio[WS(rs, 1)] = T1X + T1Y;
|
Chris@10
|
358 T66 = FNMS(KP500000000, T5u, T5n);
|
Chris@10
|
359 rio[WS(rs, 2)] = T2b + T2i;
|
Chris@10
|
360 }
|
Chris@10
|
361 iio[WS(rs, 2)] = T31 + T32;
|
Chris@10
|
362 iio[WS(rs, 4)] = T59 + T5a;
|
Chris@10
|
363 rio[WS(rs, 4)] = T4j + T4q;
|
Chris@10
|
364 rio[WS(rs, 3)] = T3f + T3m;
|
Chris@10
|
365 iio[WS(rs, 3)] = T45 + T46;
|
Chris@10
|
366 {
|
Chris@10
|
367 E TA, TD, TQ, T10, T13, TX, TZ, T12;
|
Chris@10
|
368 rio[WS(rs, 5)] = T5n + T5u;
|
Chris@10
|
369 iio[WS(rs, 5)] = T6d + T6e;
|
Chris@10
|
370 {
|
Chris@10
|
371 E To, Tx, Tb, Tq;
|
Chris@10
|
372 TA = FNMS(KP866025403, Tn, Tg);
|
Chris@10
|
373 To = FMA(KP866025403, Tn, Tg);
|
Chris@10
|
374 Tx = FMA(KP866025403, Tw, Tv);
|
Chris@10
|
375 TD = FNMS(KP866025403, Tw, Tv);
|
Chris@10
|
376 Tb = W[0];
|
Chris@10
|
377 Tq = W[1];
|
Chris@10
|
378 {
|
Chris@10
|
379 E TI, TK, TH, Ty, Tp, TF;
|
Chris@10
|
380 Ty = Tb * Tx;
|
Chris@10
|
381 Tp = Tb * To;
|
Chris@10
|
382 TF = W[4];
|
Chris@10
|
383 TI = W[5];
|
Chris@10
|
384 iio[WS(vs, 1)] = FNMS(Tq, To, Ty);
|
Chris@10
|
385 rio[WS(vs, 1)] = FMA(Tq, Tx, Tp);
|
Chris@10
|
386 TK = TF * TJ;
|
Chris@10
|
387 TH = TF * TG;
|
Chris@10
|
388 TQ = FNMS(KP866025403, TP, TM);
|
Chris@10
|
389 T10 = FMA(KP866025403, TP, TM);
|
Chris@10
|
390 T13 = FMA(KP866025403, TW, TV);
|
Chris@10
|
391 TX = FNMS(KP866025403, TW, TV);
|
Chris@10
|
392 iio[WS(vs, 3)] = FNMS(TI, TG, TK);
|
Chris@10
|
393 rio[WS(vs, 3)] = FMA(TI, TJ, TH);
|
Chris@10
|
394 TZ = W[6];
|
Chris@10
|
395 T12 = W[7];
|
Chris@10
|
396 }
|
Chris@10
|
397 }
|
Chris@10
|
398 {
|
Chris@10
|
399 E TC, TE, TB, TL, TS;
|
Chris@10
|
400 {
|
Chris@10
|
401 E T62, T64, T61, T14, T11, T5Z;
|
Chris@10
|
402 T14 = TZ * T13;
|
Chris@10
|
403 T11 = TZ * T10;
|
Chris@10
|
404 T5Z = W[4];
|
Chris@10
|
405 T62 = W[5];
|
Chris@10
|
406 iio[WS(vs, 4)] = FNMS(T12, T10, T14);
|
Chris@10
|
407 rio[WS(vs, 4)] = FMA(T12, T13, T11);
|
Chris@10
|
408 T64 = T5Z * T63;
|
Chris@10
|
409 T61 = T5Z * T60;
|
Chris@10
|
410 {
|
Chris@10
|
411 E T6k, T6n, T6j, T6m, T6o, T6l, Tz;
|
Chris@10
|
412 T6a = FNMS(KP866025403, T69, T66);
|
Chris@10
|
413 T6k = FMA(KP866025403, T69, T66);
|
Chris@10
|
414 T6n = FMA(KP866025403, T6g, T6f);
|
Chris@10
|
415 T6h = FNMS(KP866025403, T6g, T6f);
|
Chris@10
|
416 iio[WS(vs, 3) + WS(rs, 5)] = FNMS(T62, T60, T64);
|
Chris@10
|
417 rio[WS(vs, 3) + WS(rs, 5)] = FMA(T62, T63, T61);
|
Chris@10
|
418 T6j = W[6];
|
Chris@10
|
419 T6m = W[7];
|
Chris@10
|
420 T6o = T6j * T6n;
|
Chris@10
|
421 T6l = T6j * T6k;
|
Chris@10
|
422 Tz = W[8];
|
Chris@10
|
423 TC = W[9];
|
Chris@10
|
424 iio[WS(vs, 4) + WS(rs, 5)] = FNMS(T6m, T6k, T6o);
|
Chris@10
|
425 rio[WS(vs, 4) + WS(rs, 5)] = FMA(T6m, T6n, T6l);
|
Chris@10
|
426 TE = Tz * TD;
|
Chris@10
|
427 TB = Tz * TA;
|
Chris@10
|
428 }
|
Chris@10
|
429 }
|
Chris@10
|
430 iio[WS(vs, 5)] = FNMS(TC, TA, TE);
|
Chris@10
|
431 rio[WS(vs, 5)] = FMA(TC, TD, TB);
|
Chris@10
|
432 TL = W[2];
|
Chris@10
|
433 TS = W[3];
|
Chris@10
|
434 {
|
Chris@10
|
435 E T5U, T5X, T5W, T5Y, T5V, TY, TR, T5T;
|
Chris@10
|
436 T5I = FMA(KP866025403, T5H, T5A);
|
Chris@10
|
437 T5U = FNMS(KP866025403, T5H, T5A);
|
Chris@10
|
438 T5X = FNMS(KP866025403, T5Q, T5P);
|
Chris@10
|
439 T5R = FMA(KP866025403, T5Q, T5P);
|
Chris@10
|
440 TY = TL * TX;
|
Chris@10
|
441 TR = TL * TQ;
|
Chris@10
|
442 T5T = W[8];
|
Chris@10
|
443 T5W = W[9];
|
Chris@10
|
444 iio[WS(vs, 2)] = FNMS(TS, TQ, TY);
|
Chris@10
|
445 rio[WS(vs, 2)] = FMA(TS, TX, TR);
|
Chris@10
|
446 T5Y = T5T * T5X;
|
Chris@10
|
447 T5V = T5T * T5U;
|
Chris@10
|
448 iio[WS(vs, 5) + WS(rs, 5)] = FNMS(T5W, T5U, T5Y);
|
Chris@10
|
449 rio[WS(vs, 5) + WS(rs, 5)] = FMA(T5W, T5X, T5V);
|
Chris@10
|
450 T65 = W[2];
|
Chris@10
|
451 T6c = W[3];
|
Chris@10
|
452 }
|
Chris@10
|
453 }
|
Chris@10
|
454 }
|
Chris@10
|
455 }
|
Chris@10
|
456 {
|
Chris@10
|
457 E T5g, T5j, T5f, T5i;
|
Chris@10
|
458 {
|
Chris@10
|
459 E T1E, T1H, T3M, T3P, T56, T5d, T58, T5e, T57;
|
Chris@10
|
460 {
|
Chris@10
|
461 E T1s, T1B, T1f, T1u;
|
Chris@10
|
462 {
|
Chris@10
|
463 E T5K, T5S, T5J, T6i, T6b, T5v;
|
Chris@10
|
464 T6i = T65 * T6h;
|
Chris@10
|
465 T6b = T65 * T6a;
|
Chris@10
|
466 T5v = W[0];
|
Chris@10
|
467 T5K = W[1];
|
Chris@10
|
468 iio[WS(vs, 2) + WS(rs, 5)] = FNMS(T6c, T6a, T6i);
|
Chris@10
|
469 rio[WS(vs, 2) + WS(rs, 5)] = FMA(T6c, T6h, T6b);
|
Chris@10
|
470 T5S = T5v * T5R;
|
Chris@10
|
471 T5J = T5v * T5I;
|
Chris@10
|
472 T1E = FNMS(KP866025403, T1r, T1k);
|
Chris@10
|
473 T1s = FMA(KP866025403, T1r, T1k);
|
Chris@10
|
474 T1B = FMA(KP866025403, T1A, T1z);
|
Chris@10
|
475 T1H = FNMS(KP866025403, T1A, T1z);
|
Chris@10
|
476 iio[WS(vs, 1) + WS(rs, 5)] = FNMS(T5K, T5I, T5S);
|
Chris@10
|
477 rio[WS(vs, 1) + WS(rs, 5)] = FMA(T5K, T5R, T5J);
|
Chris@10
|
478 T1f = W[0];
|
Chris@10
|
479 T1u = W[1];
|
Chris@10
|
480 }
|
Chris@10
|
481 {
|
Chris@10
|
482 E T3U, T3W, T3T, T1C, T1t, T3R;
|
Chris@10
|
483 T1C = T1f * T1B;
|
Chris@10
|
484 T1t = T1f * T1s;
|
Chris@10
|
485 T3R = W[4];
|
Chris@10
|
486 T3U = W[5];
|
Chris@10
|
487 iio[WS(vs, 1) + WS(rs, 1)] = FNMS(T1u, T1s, T1C);
|
Chris@10
|
488 rio[WS(vs, 1) + WS(rs, 1)] = FMA(T1u, T1B, T1t);
|
Chris@10
|
489 T3W = T3R * T3V;
|
Chris@10
|
490 T3T = T3R * T3S;
|
Chris@10
|
491 {
|
Chris@10
|
492 E T3A, T3J, T3n, T3C, T3K, T3B, T51;
|
Chris@10
|
493 T3M = FNMS(KP866025403, T3z, T3s);
|
Chris@10
|
494 T3A = FMA(KP866025403, T3z, T3s);
|
Chris@10
|
495 T3J = FMA(KP866025403, T3I, T3H);
|
Chris@10
|
496 T3P = FNMS(KP866025403, T3I, T3H);
|
Chris@10
|
497 iio[WS(vs, 3) + WS(rs, 3)] = FNMS(T3U, T3S, T3W);
|
Chris@10
|
498 rio[WS(vs, 3) + WS(rs, 3)] = FMA(T3U, T3V, T3T);
|
Chris@10
|
499 T3n = W[0];
|
Chris@10
|
500 T3C = W[1];
|
Chris@10
|
501 T5g = FMA(KP866025403, T55, T52);
|
Chris@10
|
502 T56 = FNMS(KP866025403, T55, T52);
|
Chris@10
|
503 T5d = FNMS(KP866025403, T5c, T5b);
|
Chris@10
|
504 T5j = FMA(KP866025403, T5c, T5b);
|
Chris@10
|
505 T3K = T3n * T3J;
|
Chris@10
|
506 T3B = T3n * T3A;
|
Chris@10
|
507 T51 = W[2];
|
Chris@10
|
508 T58 = W[3];
|
Chris@10
|
509 iio[WS(vs, 1) + WS(rs, 3)] = FNMS(T3C, T3A, T3K);
|
Chris@10
|
510 rio[WS(vs, 1) + WS(rs, 3)] = FMA(T3C, T3J, T3B);
|
Chris@10
|
511 T5e = T51 * T5d;
|
Chris@10
|
512 T57 = T51 * T56;
|
Chris@10
|
513 }
|
Chris@10
|
514 }
|
Chris@10
|
515 }
|
Chris@10
|
516 {
|
Chris@10
|
517 E T38, T3b, T3O, T3Q, T3N, T37, T3a;
|
Chris@10
|
518 {
|
Chris@10
|
519 E T2Y, T35, T2T, T30, T36, T2Z, T3L;
|
Chris@10
|
520 T38 = FMA(KP866025403, T2X, T2U);
|
Chris@10
|
521 T2Y = FNMS(KP866025403, T2X, T2U);
|
Chris@10
|
522 T35 = FNMS(KP866025403, T34, T33);
|
Chris@10
|
523 T3b = FMA(KP866025403, T34, T33);
|
Chris@10
|
524 iio[WS(vs, 2) + WS(rs, 4)] = FNMS(T58, T56, T5e);
|
Chris@10
|
525 rio[WS(vs, 2) + WS(rs, 4)] = FMA(T58, T5d, T57);
|
Chris@10
|
526 T2T = W[2];
|
Chris@10
|
527 T30 = W[3];
|
Chris@10
|
528 T36 = T2T * T35;
|
Chris@10
|
529 T2Z = T2T * T2Y;
|
Chris@10
|
530 T3L = W[8];
|
Chris@10
|
531 T3O = W[9];
|
Chris@10
|
532 iio[WS(vs, 2) + WS(rs, 2)] = FNMS(T30, T2Y, T36);
|
Chris@10
|
533 rio[WS(vs, 2) + WS(rs, 2)] = FMA(T30, T35, T2Z);
|
Chris@10
|
534 T3Q = T3L * T3P;
|
Chris@10
|
535 T3N = T3L * T3M;
|
Chris@10
|
536 }
|
Chris@10
|
537 iio[WS(vs, 5) + WS(rs, 3)] = FNMS(T3O, T3M, T3Q);
|
Chris@10
|
538 rio[WS(vs, 5) + WS(rs, 3)] = FMA(T3O, T3P, T3N);
|
Chris@10
|
539 T37 = W[6];
|
Chris@10
|
540 T3a = W[7];
|
Chris@10
|
541 {
|
Chris@10
|
542 E T1G, T1I, T1F, T3c, T39, T1D;
|
Chris@10
|
543 T3c = T37 * T3b;
|
Chris@10
|
544 T39 = T37 * T38;
|
Chris@10
|
545 T1D = W[8];
|
Chris@10
|
546 T1G = W[9];
|
Chris@10
|
547 iio[WS(vs, 4) + WS(rs, 2)] = FNMS(T3a, T38, T3c);
|
Chris@10
|
548 rio[WS(vs, 4) + WS(rs, 2)] = FMA(T3a, T3b, T39);
|
Chris@10
|
549 T1I = T1D * T1H;
|
Chris@10
|
550 T1F = T1D * T1E;
|
Chris@10
|
551 iio[WS(vs, 5) + WS(rs, 1)] = FNMS(T1G, T1E, T1I);
|
Chris@10
|
552 rio[WS(vs, 5) + WS(rs, 1)] = FMA(T1G, T1H, T1F);
|
Chris@10
|
553 T5f = W[6];
|
Chris@10
|
554 T5i = W[7];
|
Chris@10
|
555 }
|
Chris@10
|
556 }
|
Chris@10
|
557 }
|
Chris@10
|
558 {
|
Chris@10
|
559 E T4Q, T4T, T2I, T2w, T2F, T2L, T2y, T2G, T2x, T4V, T4Y;
|
Chris@10
|
560 {
|
Chris@10
|
561 E T1M, T1O, T1L, T5k, T5h, T1J;
|
Chris@10
|
562 T5k = T5f * T5j;
|
Chris@10
|
563 T5h = T5f * T5g;
|
Chris@10
|
564 T1J = W[4];
|
Chris@10
|
565 T1M = W[5];
|
Chris@10
|
566 iio[WS(vs, 4) + WS(rs, 4)] = FNMS(T5i, T5g, T5k);
|
Chris@10
|
567 rio[WS(vs, 4) + WS(rs, 4)] = FMA(T5i, T5j, T5h);
|
Chris@10
|
568 T1O = T1J * T1N;
|
Chris@10
|
569 T1L = T1J * T1K;
|
Chris@10
|
570 iio[WS(vs, 3) + WS(rs, 1)] = FNMS(T1M, T1K, T1O);
|
Chris@10
|
571 rio[WS(vs, 3) + WS(rs, 1)] = FMA(T1M, T1N, T1L);
|
Chris@10
|
572 T4V = W[4];
|
Chris@10
|
573 T4Y = W[5];
|
Chris@10
|
574 }
|
Chris@10
|
575 {
|
Chris@10
|
576 E T4E, T4N, T4G, T4O, T4F, T50, T4X, T4r;
|
Chris@10
|
577 T4Q = FNMS(KP866025403, T4D, T4w);
|
Chris@10
|
578 T4E = FMA(KP866025403, T4D, T4w);
|
Chris@10
|
579 T4N = FMA(KP866025403, T4M, T4L);
|
Chris@10
|
580 T4T = FNMS(KP866025403, T4M, T4L);
|
Chris@10
|
581 T50 = T4V * T4Z;
|
Chris@10
|
582 T4X = T4V * T4W;
|
Chris@10
|
583 T4r = W[0];
|
Chris@10
|
584 T4G = W[1];
|
Chris@10
|
585 iio[WS(vs, 3) + WS(rs, 4)] = FNMS(T4Y, T4W, T50);
|
Chris@10
|
586 rio[WS(vs, 3) + WS(rs, 4)] = FMA(T4Y, T4Z, T4X);
|
Chris@10
|
587 T4O = T4r * T4N;
|
Chris@10
|
588 T4F = T4r * T4E;
|
Chris@10
|
589 {
|
Chris@10
|
590 E T2N, T2Q, T2S, T2P, T2j;
|
Chris@10
|
591 iio[WS(vs, 1) + WS(rs, 4)] = FNMS(T4G, T4E, T4O);
|
Chris@10
|
592 rio[WS(vs, 1) + WS(rs, 4)] = FMA(T4G, T4N, T4F);
|
Chris@10
|
593 T2N = W[4];
|
Chris@10
|
594 T2Q = W[5];
|
Chris@10
|
595 T2I = FNMS(KP866025403, T2v, T2o);
|
Chris@10
|
596 T2w = FMA(KP866025403, T2v, T2o);
|
Chris@10
|
597 T2F = FMA(KP866025403, T2E, T2D);
|
Chris@10
|
598 T2L = FNMS(KP866025403, T2E, T2D);
|
Chris@10
|
599 T2S = T2N * T2R;
|
Chris@10
|
600 T2P = T2N * T2O;
|
Chris@10
|
601 T2j = W[0];
|
Chris@10
|
602 T2y = W[1];
|
Chris@10
|
603 iio[WS(vs, 3) + WS(rs, 2)] = FNMS(T2Q, T2O, T2S);
|
Chris@10
|
604 rio[WS(vs, 3) + WS(rs, 2)] = FMA(T2Q, T2R, T2P);
|
Chris@10
|
605 T2G = T2j * T2F;
|
Chris@10
|
606 T2x = T2j * T2w;
|
Chris@10
|
607 }
|
Chris@10
|
608 }
|
Chris@10
|
609 {
|
Chris@10
|
610 E T1U, T21, T2H, T2K;
|
Chris@10
|
611 {
|
Chris@10
|
612 E T24, T27, T23, T26;
|
Chris@10
|
613 T1U = FNMS(KP866025403, T1T, T1Q);
|
Chris@10
|
614 T24 = FMA(KP866025403, T1T, T1Q);
|
Chris@10
|
615 T27 = FMA(KP866025403, T20, T1Z);
|
Chris@10
|
616 T21 = FNMS(KP866025403, T20, T1Z);
|
Chris@10
|
617 iio[WS(vs, 1) + WS(rs, 2)] = FNMS(T2y, T2w, T2G);
|
Chris@10
|
618 rio[WS(vs, 1) + WS(rs, 2)] = FMA(T2y, T2F, T2x);
|
Chris@10
|
619 T23 = W[6];
|
Chris@10
|
620 T26 = W[7];
|
Chris@10
|
621 {
|
Chris@10
|
622 E T42, T49, T44, T4a, T43, T28, T25, T3X;
|
Chris@10
|
623 T4c = FMA(KP866025403, T41, T3Y);
|
Chris@10
|
624 T42 = FNMS(KP866025403, T41, T3Y);
|
Chris@10
|
625 T49 = FNMS(KP866025403, T48, T47);
|
Chris@10
|
626 T4f = FMA(KP866025403, T48, T47);
|
Chris@10
|
627 T28 = T23 * T27;
|
Chris@10
|
628 T25 = T23 * T24;
|
Chris@10
|
629 T3X = W[2];
|
Chris@10
|
630 T44 = W[3];
|
Chris@10
|
631 iio[WS(vs, 4) + WS(rs, 1)] = FNMS(T26, T24, T28);
|
Chris@10
|
632 rio[WS(vs, 4) + WS(rs, 1)] = FMA(T26, T27, T25);
|
Chris@10
|
633 T4a = T3X * T49;
|
Chris@10
|
634 T43 = T3X * T42;
|
Chris@10
|
635 iio[WS(vs, 2) + WS(rs, 3)] = FNMS(T44, T42, T4a);
|
Chris@10
|
636 rio[WS(vs, 2) + WS(rs, 3)] = FMA(T44, T49, T43);
|
Chris@10
|
637 T2H = W[8];
|
Chris@10
|
638 T2K = W[9];
|
Chris@10
|
639 }
|
Chris@10
|
640 }
|
Chris@10
|
641 {
|
Chris@10
|
642 E T4S, T4U, T4R, T2M, T2J, T4P;
|
Chris@10
|
643 T2M = T2H * T2L;
|
Chris@10
|
644 T2J = T2H * T2I;
|
Chris@10
|
645 T4P = W[8];
|
Chris@10
|
646 T4S = W[9];
|
Chris@10
|
647 iio[WS(vs, 5) + WS(rs, 2)] = FNMS(T2K, T2I, T2M);
|
Chris@10
|
648 rio[WS(vs, 5) + WS(rs, 2)] = FMA(T2K, T2L, T2J);
|
Chris@10
|
649 T4U = T4P * T4T;
|
Chris@10
|
650 T4R = T4P * T4Q;
|
Chris@10
|
651 {
|
Chris@10
|
652 E T1P, T1W, T22, T1V, T4b;
|
Chris@10
|
653 iio[WS(vs, 5) + WS(rs, 4)] = FNMS(T4S, T4Q, T4U);
|
Chris@10
|
654 rio[WS(vs, 5) + WS(rs, 4)] = FMA(T4S, T4T, T4R);
|
Chris@10
|
655 T1P = W[2];
|
Chris@10
|
656 T1W = W[3];
|
Chris@10
|
657 T22 = T1P * T21;
|
Chris@10
|
658 T1V = T1P * T1U;
|
Chris@10
|
659 T4b = W[6];
|
Chris@10
|
660 T4e = W[7];
|
Chris@10
|
661 iio[WS(vs, 2) + WS(rs, 1)] = FNMS(T1W, T1U, T22);
|
Chris@10
|
662 rio[WS(vs, 2) + WS(rs, 1)] = FMA(T1W, T21, T1V);
|
Chris@10
|
663 T4g = T4b * T4f;
|
Chris@10
|
664 T4d = T4b * T4c;
|
Chris@10
|
665 }
|
Chris@10
|
666 }
|
Chris@10
|
667 }
|
Chris@10
|
668 }
|
Chris@10
|
669 }
|
Chris@10
|
670 }
|
Chris@10
|
671 }
|
Chris@10
|
672 iio[WS(vs, 4) + WS(rs, 3)] = FNMS(T4e, T4c, T4g);
|
Chris@10
|
673 rio[WS(vs, 4) + WS(rs, 3)] = FMA(T4e, T4f, T4d);
|
Chris@10
|
674 }
|
Chris@10
|
675 }
|
Chris@10
|
676 }
|
Chris@10
|
677
|
Chris@10
|
678 static const tw_instr twinstr[] = {
|
Chris@10
|
679 {TW_FULL, 0, 6},
|
Chris@10
|
680 {TW_NEXT, 1, 0}
|
Chris@10
|
681 };
|
Chris@10
|
682
|
Chris@10
|
683 static const ct_desc desc = { 6, "q1_6", twinstr, &GENUS, {144, 60, 132, 0}, 0, 0, 0 };
|
Chris@10
|
684
|
Chris@10
|
685 void X(codelet_q1_6) (planner *p) {
|
Chris@10
|
686 X(kdft_difsq_register) (p, q1_6, &desc);
|
Chris@10
|
687 }
|
Chris@10
|
688 #else /* HAVE_FMA */
|
Chris@10
|
689
|
Chris@10
|
690 /* Generated by: ../../../genfft/gen_twidsq.native -compact -variables 4 -pipeline-latency 4 -reload-twiddle -dif -n 6 -name q1_6 -include q.h */
|
Chris@10
|
691
|
Chris@10
|
692 /*
|
Chris@10
|
693 * This function contains 276 FP additions, 168 FP multiplications,
|
Chris@10
|
694 * (or, 192 additions, 84 multiplications, 84 fused multiply/add),
|
Chris@10
|
695 * 85 stack variables, 2 constants, and 144 memory accesses
|
Chris@10
|
696 */
|
Chris@10
|
697 #include "q.h"
|
Chris@10
|
698
|
Chris@10
|
699 static void q1_6(R *rio, R *iio, const R *W, stride rs, stride vs, INT mb, INT me, INT ms)
|
Chris@10
|
700 {
|
Chris@10
|
701 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
|
Chris@10
|
702 DK(KP866025403, +0.866025403784438646763723170752936183471402627);
|
Chris@10
|
703 {
|
Chris@10
|
704 INT m;
|
Chris@10
|
705 for (m = mb, W = W + (mb * 10); m < me; m = m + 1, rio = rio + ms, iio = iio + ms, W = W + 10, MAKE_VOLATILE_STRIDE(12, rs), MAKE_VOLATILE_STRIDE(0, vs)) {
|
Chris@10
|
706 E T3, Tc, Tt, TM, TX, T16, T1n, T1G, T2h, T2A, T1R, T20, T2L, T2U, T3b;
|
Chris@10
|
707 E T3u, T3F, T3O, T45, T4o, T4Z, T5i, T4z, T4I, Ta, TP, Tf, Tq, Tn, TN;
|
Chris@10
|
708 E Tu, TJ, T14, T1J, T19, T1k, T1h, T1H, T1o, T1D, T2b, T2B, T2i, T2x, T1Y;
|
Chris@10
|
709 E T2D, T23, T2e, T2S, T3x, T2X, T38, T35, T3v, T3c, T3r, T3M, T4r, T3R, T42;
|
Chris@10
|
710 E T3Z, T4p, T46, T4l, T4T, T5j, T50, T5f, T4G, T5l, T4L, T4W;
|
Chris@10
|
711 {
|
Chris@10
|
712 E T1, T2, T1l, T1m;
|
Chris@10
|
713 T1 = rio[0];
|
Chris@10
|
714 T2 = rio[WS(rs, 3)];
|
Chris@10
|
715 T3 = T1 + T2;
|
Chris@10
|
716 Tc = T1 - T2;
|
Chris@10
|
717 {
|
Chris@10
|
718 E Tr, Ts, TV, TW;
|
Chris@10
|
719 Tr = iio[0];
|
Chris@10
|
720 Ts = iio[WS(rs, 3)];
|
Chris@10
|
721 Tt = Tr - Ts;
|
Chris@10
|
722 TM = Tr + Ts;
|
Chris@10
|
723 TV = rio[WS(vs, 1)];
|
Chris@10
|
724 TW = rio[WS(vs, 1) + WS(rs, 3)];
|
Chris@10
|
725 TX = TV + TW;
|
Chris@10
|
726 T16 = TV - TW;
|
Chris@10
|
727 }
|
Chris@10
|
728 T1l = iio[WS(vs, 1)];
|
Chris@10
|
729 T1m = iio[WS(vs, 1) + WS(rs, 3)];
|
Chris@10
|
730 T1n = T1l - T1m;
|
Chris@10
|
731 T1G = T1l + T1m;
|
Chris@10
|
732 {
|
Chris@10
|
733 E T2f, T2g, T1P, T1Q;
|
Chris@10
|
734 T2f = iio[WS(vs, 2)];
|
Chris@10
|
735 T2g = iio[WS(vs, 2) + WS(rs, 3)];
|
Chris@10
|
736 T2h = T2f - T2g;
|
Chris@10
|
737 T2A = T2f + T2g;
|
Chris@10
|
738 T1P = rio[WS(vs, 2)];
|
Chris@10
|
739 T1Q = rio[WS(vs, 2) + WS(rs, 3)];
|
Chris@10
|
740 T1R = T1P + T1Q;
|
Chris@10
|
741 T20 = T1P - T1Q;
|
Chris@10
|
742 }
|
Chris@10
|
743 }
|
Chris@10
|
744 {
|
Chris@10
|
745 E T2J, T2K, T43, T44;
|
Chris@10
|
746 T2J = rio[WS(vs, 3)];
|
Chris@10
|
747 T2K = rio[WS(vs, 3) + WS(rs, 3)];
|
Chris@10
|
748 T2L = T2J + T2K;
|
Chris@10
|
749 T2U = T2J - T2K;
|
Chris@10
|
750 {
|
Chris@10
|
751 E T39, T3a, T3D, T3E;
|
Chris@10
|
752 T39 = iio[WS(vs, 3)];
|
Chris@10
|
753 T3a = iio[WS(vs, 3) + WS(rs, 3)];
|
Chris@10
|
754 T3b = T39 - T3a;
|
Chris@10
|
755 T3u = T39 + T3a;
|
Chris@10
|
756 T3D = rio[WS(vs, 4)];
|
Chris@10
|
757 T3E = rio[WS(vs, 4) + WS(rs, 3)];
|
Chris@10
|
758 T3F = T3D + T3E;
|
Chris@10
|
759 T3O = T3D - T3E;
|
Chris@10
|
760 }
|
Chris@10
|
761 T43 = iio[WS(vs, 4)];
|
Chris@10
|
762 T44 = iio[WS(vs, 4) + WS(rs, 3)];
|
Chris@10
|
763 T45 = T43 - T44;
|
Chris@10
|
764 T4o = T43 + T44;
|
Chris@10
|
765 {
|
Chris@10
|
766 E T4X, T4Y, T4x, T4y;
|
Chris@10
|
767 T4X = iio[WS(vs, 5)];
|
Chris@10
|
768 T4Y = iio[WS(vs, 5) + WS(rs, 3)];
|
Chris@10
|
769 T4Z = T4X - T4Y;
|
Chris@10
|
770 T5i = T4X + T4Y;
|
Chris@10
|
771 T4x = rio[WS(vs, 5)];
|
Chris@10
|
772 T4y = rio[WS(vs, 5) + WS(rs, 3)];
|
Chris@10
|
773 T4z = T4x + T4y;
|
Chris@10
|
774 T4I = T4x - T4y;
|
Chris@10
|
775 }
|
Chris@10
|
776 }
|
Chris@10
|
777 {
|
Chris@10
|
778 E T6, Td, T9, Te;
|
Chris@10
|
779 {
|
Chris@10
|
780 E T4, T5, T7, T8;
|
Chris@10
|
781 T4 = rio[WS(rs, 2)];
|
Chris@10
|
782 T5 = rio[WS(rs, 5)];
|
Chris@10
|
783 T6 = T4 + T5;
|
Chris@10
|
784 Td = T4 - T5;
|
Chris@10
|
785 T7 = rio[WS(rs, 4)];
|
Chris@10
|
786 T8 = rio[WS(rs, 1)];
|
Chris@10
|
787 T9 = T7 + T8;
|
Chris@10
|
788 Te = T7 - T8;
|
Chris@10
|
789 }
|
Chris@10
|
790 Ta = T6 + T9;
|
Chris@10
|
791 TP = KP866025403 * (T9 - T6);
|
Chris@10
|
792 Tf = Td + Te;
|
Chris@10
|
793 Tq = KP866025403 * (Te - Td);
|
Chris@10
|
794 }
|
Chris@10
|
795 {
|
Chris@10
|
796 E Tj, TH, Tm, TI;
|
Chris@10
|
797 {
|
Chris@10
|
798 E Th, Ti, Tk, Tl;
|
Chris@10
|
799 Th = iio[WS(rs, 2)];
|
Chris@10
|
800 Ti = iio[WS(rs, 5)];
|
Chris@10
|
801 Tj = Th - Ti;
|
Chris@10
|
802 TH = Th + Ti;
|
Chris@10
|
803 Tk = iio[WS(rs, 4)];
|
Chris@10
|
804 Tl = iio[WS(rs, 1)];
|
Chris@10
|
805 Tm = Tk - Tl;
|
Chris@10
|
806 TI = Tk + Tl;
|
Chris@10
|
807 }
|
Chris@10
|
808 Tn = KP866025403 * (Tj - Tm);
|
Chris@10
|
809 TN = TH + TI;
|
Chris@10
|
810 Tu = Tj + Tm;
|
Chris@10
|
811 TJ = KP866025403 * (TH - TI);
|
Chris@10
|
812 }
|
Chris@10
|
813 {
|
Chris@10
|
814 E T10, T17, T13, T18;
|
Chris@10
|
815 {
|
Chris@10
|
816 E TY, TZ, T11, T12;
|
Chris@10
|
817 TY = rio[WS(vs, 1) + WS(rs, 2)];
|
Chris@10
|
818 TZ = rio[WS(vs, 1) + WS(rs, 5)];
|
Chris@10
|
819 T10 = TY + TZ;
|
Chris@10
|
820 T17 = TY - TZ;
|
Chris@10
|
821 T11 = rio[WS(vs, 1) + WS(rs, 4)];
|
Chris@10
|
822 T12 = rio[WS(vs, 1) + WS(rs, 1)];
|
Chris@10
|
823 T13 = T11 + T12;
|
Chris@10
|
824 T18 = T11 - T12;
|
Chris@10
|
825 }
|
Chris@10
|
826 T14 = T10 + T13;
|
Chris@10
|
827 T1J = KP866025403 * (T13 - T10);
|
Chris@10
|
828 T19 = T17 + T18;
|
Chris@10
|
829 T1k = KP866025403 * (T18 - T17);
|
Chris@10
|
830 }
|
Chris@10
|
831 {
|
Chris@10
|
832 E T1d, T1B, T1g, T1C;
|
Chris@10
|
833 {
|
Chris@10
|
834 E T1b, T1c, T1e, T1f;
|
Chris@10
|
835 T1b = iio[WS(vs, 1) + WS(rs, 2)];
|
Chris@10
|
836 T1c = iio[WS(vs, 1) + WS(rs, 5)];
|
Chris@10
|
837 T1d = T1b - T1c;
|
Chris@10
|
838 T1B = T1b + T1c;
|
Chris@10
|
839 T1e = iio[WS(vs, 1) + WS(rs, 4)];
|
Chris@10
|
840 T1f = iio[WS(vs, 1) + WS(rs, 1)];
|
Chris@10
|
841 T1g = T1e - T1f;
|
Chris@10
|
842 T1C = T1e + T1f;
|
Chris@10
|
843 }
|
Chris@10
|
844 T1h = KP866025403 * (T1d - T1g);
|
Chris@10
|
845 T1H = T1B + T1C;
|
Chris@10
|
846 T1o = T1d + T1g;
|
Chris@10
|
847 T1D = KP866025403 * (T1B - T1C);
|
Chris@10
|
848 }
|
Chris@10
|
849 {
|
Chris@10
|
850 E T27, T2v, T2a, T2w;
|
Chris@10
|
851 {
|
Chris@10
|
852 E T25, T26, T28, T29;
|
Chris@10
|
853 T25 = iio[WS(vs, 2) + WS(rs, 2)];
|
Chris@10
|
854 T26 = iio[WS(vs, 2) + WS(rs, 5)];
|
Chris@10
|
855 T27 = T25 - T26;
|
Chris@10
|
856 T2v = T25 + T26;
|
Chris@10
|
857 T28 = iio[WS(vs, 2) + WS(rs, 4)];
|
Chris@10
|
858 T29 = iio[WS(vs, 2) + WS(rs, 1)];
|
Chris@10
|
859 T2a = T28 - T29;
|
Chris@10
|
860 T2w = T28 + T29;
|
Chris@10
|
861 }
|
Chris@10
|
862 T2b = KP866025403 * (T27 - T2a);
|
Chris@10
|
863 T2B = T2v + T2w;
|
Chris@10
|
864 T2i = T27 + T2a;
|
Chris@10
|
865 T2x = KP866025403 * (T2v - T2w);
|
Chris@10
|
866 }
|
Chris@10
|
867 {
|
Chris@10
|
868 E T1U, T21, T1X, T22;
|
Chris@10
|
869 {
|
Chris@10
|
870 E T1S, T1T, T1V, T1W;
|
Chris@10
|
871 T1S = rio[WS(vs, 2) + WS(rs, 2)];
|
Chris@10
|
872 T1T = rio[WS(vs, 2) + WS(rs, 5)];
|
Chris@10
|
873 T1U = T1S + T1T;
|
Chris@10
|
874 T21 = T1S - T1T;
|
Chris@10
|
875 T1V = rio[WS(vs, 2) + WS(rs, 4)];
|
Chris@10
|
876 T1W = rio[WS(vs, 2) + WS(rs, 1)];
|
Chris@10
|
877 T1X = T1V + T1W;
|
Chris@10
|
878 T22 = T1V - T1W;
|
Chris@10
|
879 }
|
Chris@10
|
880 T1Y = T1U + T1X;
|
Chris@10
|
881 T2D = KP866025403 * (T1X - T1U);
|
Chris@10
|
882 T23 = T21 + T22;
|
Chris@10
|
883 T2e = KP866025403 * (T22 - T21);
|
Chris@10
|
884 }
|
Chris@10
|
885 {
|
Chris@10
|
886 E T2O, T2V, T2R, T2W;
|
Chris@10
|
887 {
|
Chris@10
|
888 E T2M, T2N, T2P, T2Q;
|
Chris@10
|
889 T2M = rio[WS(vs, 3) + WS(rs, 2)];
|
Chris@10
|
890 T2N = rio[WS(vs, 3) + WS(rs, 5)];
|
Chris@10
|
891 T2O = T2M + T2N;
|
Chris@10
|
892 T2V = T2M - T2N;
|
Chris@10
|
893 T2P = rio[WS(vs, 3) + WS(rs, 4)];
|
Chris@10
|
894 T2Q = rio[WS(vs, 3) + WS(rs, 1)];
|
Chris@10
|
895 T2R = T2P + T2Q;
|
Chris@10
|
896 T2W = T2P - T2Q;
|
Chris@10
|
897 }
|
Chris@10
|
898 T2S = T2O + T2R;
|
Chris@10
|
899 T3x = KP866025403 * (T2R - T2O);
|
Chris@10
|
900 T2X = T2V + T2W;
|
Chris@10
|
901 T38 = KP866025403 * (T2W - T2V);
|
Chris@10
|
902 }
|
Chris@10
|
903 {
|
Chris@10
|
904 E T31, T3p, T34, T3q;
|
Chris@10
|
905 {
|
Chris@10
|
906 E T2Z, T30, T32, T33;
|
Chris@10
|
907 T2Z = iio[WS(vs, 3) + WS(rs, 2)];
|
Chris@10
|
908 T30 = iio[WS(vs, 3) + WS(rs, 5)];
|
Chris@10
|
909 T31 = T2Z - T30;
|
Chris@10
|
910 T3p = T2Z + T30;
|
Chris@10
|
911 T32 = iio[WS(vs, 3) + WS(rs, 4)];
|
Chris@10
|
912 T33 = iio[WS(vs, 3) + WS(rs, 1)];
|
Chris@10
|
913 T34 = T32 - T33;
|
Chris@10
|
914 T3q = T32 + T33;
|
Chris@10
|
915 }
|
Chris@10
|
916 T35 = KP866025403 * (T31 - T34);
|
Chris@10
|
917 T3v = T3p + T3q;
|
Chris@10
|
918 T3c = T31 + T34;
|
Chris@10
|
919 T3r = KP866025403 * (T3p - T3q);
|
Chris@10
|
920 }
|
Chris@10
|
921 {
|
Chris@10
|
922 E T3I, T3P, T3L, T3Q;
|
Chris@10
|
923 {
|
Chris@10
|
924 E T3G, T3H, T3J, T3K;
|
Chris@10
|
925 T3G = rio[WS(vs, 4) + WS(rs, 2)];
|
Chris@10
|
926 T3H = rio[WS(vs, 4) + WS(rs, 5)];
|
Chris@10
|
927 T3I = T3G + T3H;
|
Chris@10
|
928 T3P = T3G - T3H;
|
Chris@10
|
929 T3J = rio[WS(vs, 4) + WS(rs, 4)];
|
Chris@10
|
930 T3K = rio[WS(vs, 4) + WS(rs, 1)];
|
Chris@10
|
931 T3L = T3J + T3K;
|
Chris@10
|
932 T3Q = T3J - T3K;
|
Chris@10
|
933 }
|
Chris@10
|
934 T3M = T3I + T3L;
|
Chris@10
|
935 T4r = KP866025403 * (T3L - T3I);
|
Chris@10
|
936 T3R = T3P + T3Q;
|
Chris@10
|
937 T42 = KP866025403 * (T3Q - T3P);
|
Chris@10
|
938 }
|
Chris@10
|
939 {
|
Chris@10
|
940 E T3V, T4j, T3Y, T4k;
|
Chris@10
|
941 {
|
Chris@10
|
942 E T3T, T3U, T3W, T3X;
|
Chris@10
|
943 T3T = iio[WS(vs, 4) + WS(rs, 2)];
|
Chris@10
|
944 T3U = iio[WS(vs, 4) + WS(rs, 5)];
|
Chris@10
|
945 T3V = T3T - T3U;
|
Chris@10
|
946 T4j = T3T + T3U;
|
Chris@10
|
947 T3W = iio[WS(vs, 4) + WS(rs, 4)];
|
Chris@10
|
948 T3X = iio[WS(vs, 4) + WS(rs, 1)];
|
Chris@10
|
949 T3Y = T3W - T3X;
|
Chris@10
|
950 T4k = T3W + T3X;
|
Chris@10
|
951 }
|
Chris@10
|
952 T3Z = KP866025403 * (T3V - T3Y);
|
Chris@10
|
953 T4p = T4j + T4k;
|
Chris@10
|
954 T46 = T3V + T3Y;
|
Chris@10
|
955 T4l = KP866025403 * (T4j - T4k);
|
Chris@10
|
956 }
|
Chris@10
|
957 {
|
Chris@10
|
958 E T4P, T5d, T4S, T5e;
|
Chris@10
|
959 {
|
Chris@10
|
960 E T4N, T4O, T4Q, T4R;
|
Chris@10
|
961 T4N = iio[WS(vs, 5) + WS(rs, 2)];
|
Chris@10
|
962 T4O = iio[WS(vs, 5) + WS(rs, 5)];
|
Chris@10
|
963 T4P = T4N - T4O;
|
Chris@10
|
964 T5d = T4N + T4O;
|
Chris@10
|
965 T4Q = iio[WS(vs, 5) + WS(rs, 4)];
|
Chris@10
|
966 T4R = iio[WS(vs, 5) + WS(rs, 1)];
|
Chris@10
|
967 T4S = T4Q - T4R;
|
Chris@10
|
968 T5e = T4Q + T4R;
|
Chris@10
|
969 }
|
Chris@10
|
970 T4T = KP866025403 * (T4P - T4S);
|
Chris@10
|
971 T5j = T5d + T5e;
|
Chris@10
|
972 T50 = T4P + T4S;
|
Chris@10
|
973 T5f = KP866025403 * (T5d - T5e);
|
Chris@10
|
974 }
|
Chris@10
|
975 {
|
Chris@10
|
976 E T4C, T4J, T4F, T4K;
|
Chris@10
|
977 {
|
Chris@10
|
978 E T4A, T4B, T4D, T4E;
|
Chris@10
|
979 T4A = rio[WS(vs, 5) + WS(rs, 2)];
|
Chris@10
|
980 T4B = rio[WS(vs, 5) + WS(rs, 5)];
|
Chris@10
|
981 T4C = T4A + T4B;
|
Chris@10
|
982 T4J = T4A - T4B;
|
Chris@10
|
983 T4D = rio[WS(vs, 5) + WS(rs, 4)];
|
Chris@10
|
984 T4E = rio[WS(vs, 5) + WS(rs, 1)];
|
Chris@10
|
985 T4F = T4D + T4E;
|
Chris@10
|
986 T4K = T4D - T4E;
|
Chris@10
|
987 }
|
Chris@10
|
988 T4G = T4C + T4F;
|
Chris@10
|
989 T5l = KP866025403 * (T4F - T4C);
|
Chris@10
|
990 T4L = T4J + T4K;
|
Chris@10
|
991 T4W = KP866025403 * (T4K - T4J);
|
Chris@10
|
992 }
|
Chris@10
|
993 rio[0] = T3 + Ta;
|
Chris@10
|
994 iio[0] = TM + TN;
|
Chris@10
|
995 rio[WS(rs, 1)] = TX + T14;
|
Chris@10
|
996 iio[WS(rs, 1)] = T1G + T1H;
|
Chris@10
|
997 rio[WS(rs, 3)] = T2L + T2S;
|
Chris@10
|
998 rio[WS(rs, 2)] = T1R + T1Y;
|
Chris@10
|
999 iio[WS(rs, 2)] = T2A + T2B;
|
Chris@10
|
1000 iio[WS(rs, 3)] = T3u + T3v;
|
Chris@10
|
1001 iio[WS(rs, 4)] = T4o + T4p;
|
Chris@10
|
1002 iio[WS(rs, 5)] = T5i + T5j;
|
Chris@10
|
1003 rio[WS(rs, 5)] = T4z + T4G;
|
Chris@10
|
1004 rio[WS(rs, 4)] = T3F + T3M;
|
Chris@10
|
1005 {
|
Chris@10
|
1006 E T1w, T1y, T1v, T1x;
|
Chris@10
|
1007 T1w = T16 + T19;
|
Chris@10
|
1008 T1y = T1n + T1o;
|
Chris@10
|
1009 T1v = W[4];
|
Chris@10
|
1010 T1x = W[5];
|
Chris@10
|
1011 rio[WS(vs, 3) + WS(rs, 1)] = FMA(T1v, T1w, T1x * T1y);
|
Chris@10
|
1012 iio[WS(vs, 3) + WS(rs, 1)] = FNMS(T1x, T1w, T1v * T1y);
|
Chris@10
|
1013 }
|
Chris@10
|
1014 {
|
Chris@10
|
1015 E T58, T5a, T57, T59;
|
Chris@10
|
1016 T58 = T4I + T4L;
|
Chris@10
|
1017 T5a = T4Z + T50;
|
Chris@10
|
1018 T57 = W[4];
|
Chris@10
|
1019 T59 = W[5];
|
Chris@10
|
1020 rio[WS(vs, 3) + WS(rs, 5)] = FMA(T57, T58, T59 * T5a);
|
Chris@10
|
1021 iio[WS(vs, 3) + WS(rs, 5)] = FNMS(T59, T58, T57 * T5a);
|
Chris@10
|
1022 }
|
Chris@10
|
1023 {
|
Chris@10
|
1024 E TC, TE, TB, TD;
|
Chris@10
|
1025 TC = Tc + Tf;
|
Chris@10
|
1026 TE = Tt + Tu;
|
Chris@10
|
1027 TB = W[4];
|
Chris@10
|
1028 TD = W[5];
|
Chris@10
|
1029 rio[WS(vs, 3)] = FMA(TB, TC, TD * TE);
|
Chris@10
|
1030 iio[WS(vs, 3)] = FNMS(TD, TC, TB * TE);
|
Chris@10
|
1031 }
|
Chris@10
|
1032 {
|
Chris@10
|
1033 E T4e, T4g, T4d, T4f;
|
Chris@10
|
1034 T4e = T3O + T3R;
|
Chris@10
|
1035 T4g = T45 + T46;
|
Chris@10
|
1036 T4d = W[4];
|
Chris@10
|
1037 T4f = W[5];
|
Chris@10
|
1038 rio[WS(vs, 3) + WS(rs, 4)] = FMA(T4d, T4e, T4f * T4g);
|
Chris@10
|
1039 iio[WS(vs, 3) + WS(rs, 4)] = FNMS(T4f, T4e, T4d * T4g);
|
Chris@10
|
1040 }
|
Chris@10
|
1041 {
|
Chris@10
|
1042 E T3k, T3m, T3j, T3l;
|
Chris@10
|
1043 T3k = T2U + T2X;
|
Chris@10
|
1044 T3m = T3b + T3c;
|
Chris@10
|
1045 T3j = W[4];
|
Chris@10
|
1046 T3l = W[5];
|
Chris@10
|
1047 rio[WS(vs, 3) + WS(rs, 3)] = FMA(T3j, T3k, T3l * T3m);
|
Chris@10
|
1048 iio[WS(vs, 3) + WS(rs, 3)] = FNMS(T3l, T3k, T3j * T3m);
|
Chris@10
|
1049 }
|
Chris@10
|
1050 {
|
Chris@10
|
1051 E T2q, T2s, T2p, T2r;
|
Chris@10
|
1052 T2q = T20 + T23;
|
Chris@10
|
1053 T2s = T2h + T2i;
|
Chris@10
|
1054 T2p = W[4];
|
Chris@10
|
1055 T2r = W[5];
|
Chris@10
|
1056 rio[WS(vs, 3) + WS(rs, 2)] = FMA(T2p, T2q, T2r * T2s);
|
Chris@10
|
1057 iio[WS(vs, 3) + WS(rs, 2)] = FNMS(T2r, T2q, T2p * T2s);
|
Chris@10
|
1058 }
|
Chris@10
|
1059 {
|
Chris@10
|
1060 E T5g, T5o, T5m, T5q, T5c, T5k;
|
Chris@10
|
1061 T5c = FNMS(KP500000000, T4G, T4z);
|
Chris@10
|
1062 T5g = T5c - T5f;
|
Chris@10
|
1063 T5o = T5c + T5f;
|
Chris@10
|
1064 T5k = FNMS(KP500000000, T5j, T5i);
|
Chris@10
|
1065 T5m = T5k - T5l;
|
Chris@10
|
1066 T5q = T5l + T5k;
|
Chris@10
|
1067 {
|
Chris@10
|
1068 E T5b, T5h, T5n, T5p;
|
Chris@10
|
1069 T5b = W[2];
|
Chris@10
|
1070 T5h = W[3];
|
Chris@10
|
1071 rio[WS(vs, 2) + WS(rs, 5)] = FMA(T5b, T5g, T5h * T5m);
|
Chris@10
|
1072 iio[WS(vs, 2) + WS(rs, 5)] = FNMS(T5h, T5g, T5b * T5m);
|
Chris@10
|
1073 T5n = W[6];
|
Chris@10
|
1074 T5p = W[7];
|
Chris@10
|
1075 rio[WS(vs, 4) + WS(rs, 5)] = FMA(T5n, T5o, T5p * T5q);
|
Chris@10
|
1076 iio[WS(vs, 4) + WS(rs, 5)] = FNMS(T5p, T5o, T5n * T5q);
|
Chris@10
|
1077 }
|
Chris@10
|
1078 }
|
Chris@10
|
1079 {
|
Chris@10
|
1080 E To, Ty, Tw, TA, Tg, Tv;
|
Chris@10
|
1081 Tg = FNMS(KP500000000, Tf, Tc);
|
Chris@10
|
1082 To = Tg + Tn;
|
Chris@10
|
1083 Ty = Tg - Tn;
|
Chris@10
|
1084 Tv = FNMS(KP500000000, Tu, Tt);
|
Chris@10
|
1085 Tw = Tq + Tv;
|
Chris@10
|
1086 TA = Tv - Tq;
|
Chris@10
|
1087 {
|
Chris@10
|
1088 E Tb, Tp, Tx, Tz;
|
Chris@10
|
1089 Tb = W[0];
|
Chris@10
|
1090 Tp = W[1];
|
Chris@10
|
1091 rio[WS(vs, 1)] = FMA(Tb, To, Tp * Tw);
|
Chris@10
|
1092 iio[WS(vs, 1)] = FNMS(Tp, To, Tb * Tw);
|
Chris@10
|
1093 Tx = W[8];
|
Chris@10
|
1094 Tz = W[9];
|
Chris@10
|
1095 rio[WS(vs, 5)] = FMA(Tx, Ty, Tz * TA);
|
Chris@10
|
1096 iio[WS(vs, 5)] = FNMS(Tz, Ty, Tx * TA);
|
Chris@10
|
1097 }
|
Chris@10
|
1098 }
|
Chris@10
|
1099 {
|
Chris@10
|
1100 E T36, T3g, T3e, T3i, T2Y, T3d;
|
Chris@10
|
1101 T2Y = FNMS(KP500000000, T2X, T2U);
|
Chris@10
|
1102 T36 = T2Y + T35;
|
Chris@10
|
1103 T3g = T2Y - T35;
|
Chris@10
|
1104 T3d = FNMS(KP500000000, T3c, T3b);
|
Chris@10
|
1105 T3e = T38 + T3d;
|
Chris@10
|
1106 T3i = T3d - T38;
|
Chris@10
|
1107 {
|
Chris@10
|
1108 E T2T, T37, T3f, T3h;
|
Chris@10
|
1109 T2T = W[0];
|
Chris@10
|
1110 T37 = W[1];
|
Chris@10
|
1111 rio[WS(vs, 1) + WS(rs, 3)] = FMA(T2T, T36, T37 * T3e);
|
Chris@10
|
1112 iio[WS(vs, 1) + WS(rs, 3)] = FNMS(T37, T36, T2T * T3e);
|
Chris@10
|
1113 T3f = W[8];
|
Chris@10
|
1114 T3h = W[9];
|
Chris@10
|
1115 rio[WS(vs, 5) + WS(rs, 3)] = FMA(T3f, T3g, T3h * T3i);
|
Chris@10
|
1116 iio[WS(vs, 5) + WS(rs, 3)] = FNMS(T3h, T3g, T3f * T3i);
|
Chris@10
|
1117 }
|
Chris@10
|
1118 }
|
Chris@10
|
1119 {
|
Chris@10
|
1120 E T2y, T2G, T2E, T2I, T2u, T2C;
|
Chris@10
|
1121 T2u = FNMS(KP500000000, T1Y, T1R);
|
Chris@10
|
1122 T2y = T2u - T2x;
|
Chris@10
|
1123 T2G = T2u + T2x;
|
Chris@10
|
1124 T2C = FNMS(KP500000000, T2B, T2A);
|
Chris@10
|
1125 T2E = T2C - T2D;
|
Chris@10
|
1126 T2I = T2D + T2C;
|
Chris@10
|
1127 {
|
Chris@10
|
1128 E T2t, T2z, T2F, T2H;
|
Chris@10
|
1129 T2t = W[2];
|
Chris@10
|
1130 T2z = W[3];
|
Chris@10
|
1131 rio[WS(vs, 2) + WS(rs, 2)] = FMA(T2t, T2y, T2z * T2E);
|
Chris@10
|
1132 iio[WS(vs, 2) + WS(rs, 2)] = FNMS(T2z, T2y, T2t * T2E);
|
Chris@10
|
1133 T2F = W[6];
|
Chris@10
|
1134 T2H = W[7];
|
Chris@10
|
1135 rio[WS(vs, 4) + WS(rs, 2)] = FMA(T2F, T2G, T2H * T2I);
|
Chris@10
|
1136 iio[WS(vs, 4) + WS(rs, 2)] = FNMS(T2H, T2G, T2F * T2I);
|
Chris@10
|
1137 }
|
Chris@10
|
1138 }
|
Chris@10
|
1139 {
|
Chris@10
|
1140 E T3s, T3A, T3y, T3C, T3o, T3w;
|
Chris@10
|
1141 T3o = FNMS(KP500000000, T2S, T2L);
|
Chris@10
|
1142 T3s = T3o - T3r;
|
Chris@10
|
1143 T3A = T3o + T3r;
|
Chris@10
|
1144 T3w = FNMS(KP500000000, T3v, T3u);
|
Chris@10
|
1145 T3y = T3w - T3x;
|
Chris@10
|
1146 T3C = T3x + T3w;
|
Chris@10
|
1147 {
|
Chris@10
|
1148 E T3n, T3t, T3z, T3B;
|
Chris@10
|
1149 T3n = W[2];
|
Chris@10
|
1150 T3t = W[3];
|
Chris@10
|
1151 rio[WS(vs, 2) + WS(rs, 3)] = FMA(T3n, T3s, T3t * T3y);
|
Chris@10
|
1152 iio[WS(vs, 2) + WS(rs, 3)] = FNMS(T3t, T3s, T3n * T3y);
|
Chris@10
|
1153 T3z = W[6];
|
Chris@10
|
1154 T3B = W[7];
|
Chris@10
|
1155 rio[WS(vs, 4) + WS(rs, 3)] = FMA(T3z, T3A, T3B * T3C);
|
Chris@10
|
1156 iio[WS(vs, 4) + WS(rs, 3)] = FNMS(T3B, T3A, T3z * T3C);
|
Chris@10
|
1157 }
|
Chris@10
|
1158 }
|
Chris@10
|
1159 {
|
Chris@10
|
1160 E T1E, T1M, T1K, T1O, T1A, T1I;
|
Chris@10
|
1161 T1A = FNMS(KP500000000, T14, TX);
|
Chris@10
|
1162 T1E = T1A - T1D;
|
Chris@10
|
1163 T1M = T1A + T1D;
|
Chris@10
|
1164 T1I = FNMS(KP500000000, T1H, T1G);
|
Chris@10
|
1165 T1K = T1I - T1J;
|
Chris@10
|
1166 T1O = T1J + T1I;
|
Chris@10
|
1167 {
|
Chris@10
|
1168 E T1z, T1F, T1L, T1N;
|
Chris@10
|
1169 T1z = W[2];
|
Chris@10
|
1170 T1F = W[3];
|
Chris@10
|
1171 rio[WS(vs, 2) + WS(rs, 1)] = FMA(T1z, T1E, T1F * T1K);
|
Chris@10
|
1172 iio[WS(vs, 2) + WS(rs, 1)] = FNMS(T1F, T1E, T1z * T1K);
|
Chris@10
|
1173 T1L = W[6];
|
Chris@10
|
1174 T1N = W[7];
|
Chris@10
|
1175 rio[WS(vs, 4) + WS(rs, 1)] = FMA(T1L, T1M, T1N * T1O);
|
Chris@10
|
1176 iio[WS(vs, 4) + WS(rs, 1)] = FNMS(T1N, T1M, T1L * T1O);
|
Chris@10
|
1177 }
|
Chris@10
|
1178 }
|
Chris@10
|
1179 {
|
Chris@10
|
1180 E T4m, T4u, T4s, T4w, T4i, T4q;
|
Chris@10
|
1181 T4i = FNMS(KP500000000, T3M, T3F);
|
Chris@10
|
1182 T4m = T4i - T4l;
|
Chris@10
|
1183 T4u = T4i + T4l;
|
Chris@10
|
1184 T4q = FNMS(KP500000000, T4p, T4o);
|
Chris@10
|
1185 T4s = T4q - T4r;
|
Chris@10
|
1186 T4w = T4r + T4q;
|
Chris@10
|
1187 {
|
Chris@10
|
1188 E T4h, T4n, T4t, T4v;
|
Chris@10
|
1189 T4h = W[2];
|
Chris@10
|
1190 T4n = W[3];
|
Chris@10
|
1191 rio[WS(vs, 2) + WS(rs, 4)] = FMA(T4h, T4m, T4n * T4s);
|
Chris@10
|
1192 iio[WS(vs, 2) + WS(rs, 4)] = FNMS(T4n, T4m, T4h * T4s);
|
Chris@10
|
1193 T4t = W[6];
|
Chris@10
|
1194 T4v = W[7];
|
Chris@10
|
1195 rio[WS(vs, 4) + WS(rs, 4)] = FMA(T4t, T4u, T4v * T4w);
|
Chris@10
|
1196 iio[WS(vs, 4) + WS(rs, 4)] = FNMS(T4v, T4u, T4t * T4w);
|
Chris@10
|
1197 }
|
Chris@10
|
1198 }
|
Chris@10
|
1199 {
|
Chris@10
|
1200 E TK, TS, TQ, TU, TG, TO;
|
Chris@10
|
1201 TG = FNMS(KP500000000, Ta, T3);
|
Chris@10
|
1202 TK = TG - TJ;
|
Chris@10
|
1203 TS = TG + TJ;
|
Chris@10
|
1204 TO = FNMS(KP500000000, TN, TM);
|
Chris@10
|
1205 TQ = TO - TP;
|
Chris@10
|
1206 TU = TP + TO;
|
Chris@10
|
1207 {
|
Chris@10
|
1208 E TF, TL, TR, TT;
|
Chris@10
|
1209 TF = W[2];
|
Chris@10
|
1210 TL = W[3];
|
Chris@10
|
1211 rio[WS(vs, 2)] = FMA(TF, TK, TL * TQ);
|
Chris@10
|
1212 iio[WS(vs, 2)] = FNMS(TL, TK, TF * TQ);
|
Chris@10
|
1213 TR = W[6];
|
Chris@10
|
1214 TT = W[7];
|
Chris@10
|
1215 rio[WS(vs, 4)] = FMA(TR, TS, TT * TU);
|
Chris@10
|
1216 iio[WS(vs, 4)] = FNMS(TT, TS, TR * TU);
|
Chris@10
|
1217 }
|
Chris@10
|
1218 }
|
Chris@10
|
1219 {
|
Chris@10
|
1220 E T2c, T2m, T2k, T2o, T24, T2j;
|
Chris@10
|
1221 T24 = FNMS(KP500000000, T23, T20);
|
Chris@10
|
1222 T2c = T24 + T2b;
|
Chris@10
|
1223 T2m = T24 - T2b;
|
Chris@10
|
1224 T2j = FNMS(KP500000000, T2i, T2h);
|
Chris@10
|
1225 T2k = T2e + T2j;
|
Chris@10
|
1226 T2o = T2j - T2e;
|
Chris@10
|
1227 {
|
Chris@10
|
1228 E T1Z, T2d, T2l, T2n;
|
Chris@10
|
1229 T1Z = W[0];
|
Chris@10
|
1230 T2d = W[1];
|
Chris@10
|
1231 rio[WS(vs, 1) + WS(rs, 2)] = FMA(T1Z, T2c, T2d * T2k);
|
Chris@10
|
1232 iio[WS(vs, 1) + WS(rs, 2)] = FNMS(T2d, T2c, T1Z * T2k);
|
Chris@10
|
1233 T2l = W[8];
|
Chris@10
|
1234 T2n = W[9];
|
Chris@10
|
1235 rio[WS(vs, 5) + WS(rs, 2)] = FMA(T2l, T2m, T2n * T2o);
|
Chris@10
|
1236 iio[WS(vs, 5) + WS(rs, 2)] = FNMS(T2n, T2m, T2l * T2o);
|
Chris@10
|
1237 }
|
Chris@10
|
1238 }
|
Chris@10
|
1239 {
|
Chris@10
|
1240 E T40, T4a, T48, T4c, T3S, T47;
|
Chris@10
|
1241 T3S = FNMS(KP500000000, T3R, T3O);
|
Chris@10
|
1242 T40 = T3S + T3Z;
|
Chris@10
|
1243 T4a = T3S - T3Z;
|
Chris@10
|
1244 T47 = FNMS(KP500000000, T46, T45);
|
Chris@10
|
1245 T48 = T42 + T47;
|
Chris@10
|
1246 T4c = T47 - T42;
|
Chris@10
|
1247 {
|
Chris@10
|
1248 E T3N, T41, T49, T4b;
|
Chris@10
|
1249 T3N = W[0];
|
Chris@10
|
1250 T41 = W[1];
|
Chris@10
|
1251 rio[WS(vs, 1) + WS(rs, 4)] = FMA(T3N, T40, T41 * T48);
|
Chris@10
|
1252 iio[WS(vs, 1) + WS(rs, 4)] = FNMS(T41, T40, T3N * T48);
|
Chris@10
|
1253 T49 = W[8];
|
Chris@10
|
1254 T4b = W[9];
|
Chris@10
|
1255 rio[WS(vs, 5) + WS(rs, 4)] = FMA(T49, T4a, T4b * T4c);
|
Chris@10
|
1256 iio[WS(vs, 5) + WS(rs, 4)] = FNMS(T4b, T4a, T49 * T4c);
|
Chris@10
|
1257 }
|
Chris@10
|
1258 }
|
Chris@10
|
1259 {
|
Chris@10
|
1260 E T1i, T1s, T1q, T1u, T1a, T1p;
|
Chris@10
|
1261 T1a = FNMS(KP500000000, T19, T16);
|
Chris@10
|
1262 T1i = T1a + T1h;
|
Chris@10
|
1263 T1s = T1a - T1h;
|
Chris@10
|
1264 T1p = FNMS(KP500000000, T1o, T1n);
|
Chris@10
|
1265 T1q = T1k + T1p;
|
Chris@10
|
1266 T1u = T1p - T1k;
|
Chris@10
|
1267 {
|
Chris@10
|
1268 E T15, T1j, T1r, T1t;
|
Chris@10
|
1269 T15 = W[0];
|
Chris@10
|
1270 T1j = W[1];
|
Chris@10
|
1271 rio[WS(vs, 1) + WS(rs, 1)] = FMA(T15, T1i, T1j * T1q);
|
Chris@10
|
1272 iio[WS(vs, 1) + WS(rs, 1)] = FNMS(T1j, T1i, T15 * T1q);
|
Chris@10
|
1273 T1r = W[8];
|
Chris@10
|
1274 T1t = W[9];
|
Chris@10
|
1275 rio[WS(vs, 5) + WS(rs, 1)] = FMA(T1r, T1s, T1t * T1u);
|
Chris@10
|
1276 iio[WS(vs, 5) + WS(rs, 1)] = FNMS(T1t, T1s, T1r * T1u);
|
Chris@10
|
1277 }
|
Chris@10
|
1278 }
|
Chris@10
|
1279 {
|
Chris@10
|
1280 E T4U, T54, T52, T56, T4M, T51;
|
Chris@10
|
1281 T4M = FNMS(KP500000000, T4L, T4I);
|
Chris@10
|
1282 T4U = T4M + T4T;
|
Chris@10
|
1283 T54 = T4M - T4T;
|
Chris@10
|
1284 T51 = FNMS(KP500000000, T50, T4Z);
|
Chris@10
|
1285 T52 = T4W + T51;
|
Chris@10
|
1286 T56 = T51 - T4W;
|
Chris@10
|
1287 {
|
Chris@10
|
1288 E T4H, T4V, T53, T55;
|
Chris@10
|
1289 T4H = W[0];
|
Chris@10
|
1290 T4V = W[1];
|
Chris@10
|
1291 rio[WS(vs, 1) + WS(rs, 5)] = FMA(T4H, T4U, T4V * T52);
|
Chris@10
|
1292 iio[WS(vs, 1) + WS(rs, 5)] = FNMS(T4V, T4U, T4H * T52);
|
Chris@10
|
1293 T53 = W[8];
|
Chris@10
|
1294 T55 = W[9];
|
Chris@10
|
1295 rio[WS(vs, 5) + WS(rs, 5)] = FMA(T53, T54, T55 * T56);
|
Chris@10
|
1296 iio[WS(vs, 5) + WS(rs, 5)] = FNMS(T55, T54, T53 * T56);
|
Chris@10
|
1297 }
|
Chris@10
|
1298 }
|
Chris@10
|
1299 }
|
Chris@10
|
1300 }
|
Chris@10
|
1301 }
|
Chris@10
|
1302
|
Chris@10
|
1303 static const tw_instr twinstr[] = {
|
Chris@10
|
1304 {TW_FULL, 0, 6},
|
Chris@10
|
1305 {TW_NEXT, 1, 0}
|
Chris@10
|
1306 };
|
Chris@10
|
1307
|
Chris@10
|
1308 static const ct_desc desc = { 6, "q1_6", twinstr, &GENUS, {192, 84, 84, 0}, 0, 0, 0 };
|
Chris@10
|
1309
|
Chris@10
|
1310 void X(codelet_q1_6) (planner *p) {
|
Chris@10
|
1311 X(kdft_difsq_register) (p, q1_6, &desc);
|
Chris@10
|
1312 }
|
Chris@10
|
1313 #endif /* HAVE_FMA */
|