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 07:35:43 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_notw.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -n 13 -name n1_13 -include n.h */
|
Chris@10
|
29
|
Chris@10
|
30 /*
|
Chris@10
|
31 * This function contains 176 FP additions, 114 FP multiplications,
|
Chris@10
|
32 * (or, 62 additions, 0 multiplications, 114 fused multiply/add),
|
Chris@10
|
33 * 87 stack variables, 25 constants, and 52 memory accesses
|
Chris@10
|
34 */
|
Chris@10
|
35 #include "n.h"
|
Chris@10
|
36
|
Chris@10
|
37 static void n1_13(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
|
Chris@10
|
38 {
|
Chris@10
|
39 DK(KP875502302, +0.875502302409147941146295545768755143177842006);
|
Chris@10
|
40 DK(KP520028571, +0.520028571888864619117130500499232802493238139);
|
Chris@10
|
41 DK(KP575140729, +0.575140729474003121368385547455453388461001608);
|
Chris@10
|
42 DK(KP600477271, +0.600477271932665282925769253334763009352012849);
|
Chris@10
|
43 DK(KP300462606, +0.300462606288665774426601772289207995520941381);
|
Chris@10
|
44 DK(KP516520780, +0.516520780623489722840901288569017135705033622);
|
Chris@10
|
45 DK(KP968287244, +0.968287244361984016049539446938120421179794516);
|
Chris@10
|
46 DK(KP503537032, +0.503537032863766627246873853868466977093348562);
|
Chris@10
|
47 DK(KP251768516, +0.251768516431883313623436926934233488546674281);
|
Chris@10
|
48 DK(KP581704778, +0.581704778510515730456870384989698884939833902);
|
Chris@10
|
49 DK(KP859542535, +0.859542535098774820163672132761689612766401925);
|
Chris@10
|
50 DK(KP083333333, +0.083333333333333333333333333333333333333333333);
|
Chris@10
|
51 DK(KP957805992, +0.957805992594665126462521754605754580515587217);
|
Chris@10
|
52 DK(KP522026385, +0.522026385161275033714027226654165028300441940);
|
Chris@10
|
53 DK(KP853480001, +0.853480001859823990758994934970528322872359049);
|
Chris@10
|
54 DK(KP769338817, +0.769338817572980603471413688209101117038278899);
|
Chris@10
|
55 DK(KP612264650, +0.612264650376756543746494474777125408779395514);
|
Chris@10
|
56 DK(KP038632954, +0.038632954644348171955506895830342264440241080);
|
Chris@10
|
57 DK(KP302775637, +0.302775637731994646559610633735247973125648287);
|
Chris@10
|
58 DK(KP514918778, +0.514918778086315755491789696138117261566051239);
|
Chris@10
|
59 DK(KP686558370, +0.686558370781754340655719594850823015421401653);
|
Chris@10
|
60 DK(KP226109445, +0.226109445035782405468510155372505010481906348);
|
Chris@10
|
61 DK(KP301479260, +0.301479260047709873958013540496673347309208464);
|
Chris@10
|
62 DK(KP866025403, +0.866025403784438646763723170752936183471402627);
|
Chris@10
|
63 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
|
Chris@10
|
64 {
|
Chris@10
|
65 INT i;
|
Chris@10
|
66 for (i = v; i > 0; i = i - 1, ri = ri + ivs, ii = ii + ivs, ro = ro + ovs, io = io + ovs, MAKE_VOLATILE_STRIDE(52, is), MAKE_VOLATILE_STRIDE(52, os)) {
|
Chris@10
|
67 E T2B, T2H, T2I, T2G;
|
Chris@10
|
68 {
|
Chris@10
|
69 E T1, T1P, T2n, T2o, To, TH, T2h, T2k, TE, TB, TF, Tw, T2j, T2c, T1m;
|
Chris@10
|
70 E T1W, T1X, T1c, T19, T1j, T12, T1f, T21, T24, T27, T1U;
|
Chris@10
|
71 T1 = ri[0];
|
Chris@10
|
72 T1P = ii[0];
|
Chris@10
|
73 {
|
Chris@10
|
74 E T2b, Tv, Ts, T2a;
|
Chris@10
|
75 {
|
Chris@10
|
76 E T2d, Tf, Tq, Ty, Tb, Tr, T6, Tx, Ti, Tt, Tu, Tl;
|
Chris@10
|
77 {
|
Chris@10
|
78 E T7, T8, T9, Td, Te;
|
Chris@10
|
79 Td = ri[WS(is, 8)];
|
Chris@10
|
80 Te = ri[WS(is, 5)];
|
Chris@10
|
81 T7 = ri[WS(is, 12)];
|
Chris@10
|
82 T8 = ri[WS(is, 10)];
|
Chris@10
|
83 T9 = ri[WS(is, 4)];
|
Chris@10
|
84 T2d = Td - Te;
|
Chris@10
|
85 Tf = Td + Te;
|
Chris@10
|
86 {
|
Chris@10
|
87 E T2, Ta, T3, T4;
|
Chris@10
|
88 T2 = ri[WS(is, 1)];
|
Chris@10
|
89 Ta = T8 + T9;
|
Chris@10
|
90 Tq = T8 - T9;
|
Chris@10
|
91 T3 = ri[WS(is, 3)];
|
Chris@10
|
92 T4 = ri[WS(is, 9)];
|
Chris@10
|
93 {
|
Chris@10
|
94 E Tg, T5, Th, Tj, Tk;
|
Chris@10
|
95 Tg = ri[WS(is, 11)];
|
Chris@10
|
96 Ty = FMS(KP500000000, Ta, T7);
|
Chris@10
|
97 Tb = T7 + Ta;
|
Chris@10
|
98 Tr = T4 - T3;
|
Chris@10
|
99 T5 = T3 + T4;
|
Chris@10
|
100 Th = ri[WS(is, 6)];
|
Chris@10
|
101 Tj = ri[WS(is, 7)];
|
Chris@10
|
102 Tk = ri[WS(is, 2)];
|
Chris@10
|
103 T6 = T2 + T5;
|
Chris@10
|
104 Tx = FNMS(KP500000000, T5, T2);
|
Chris@10
|
105 Ti = Tg + Th;
|
Chris@10
|
106 Tt = Tg - Th;
|
Chris@10
|
107 Tu = Tj - Tk;
|
Chris@10
|
108 Tl = Tj + Tk;
|
Chris@10
|
109 }
|
Chris@10
|
110 }
|
Chris@10
|
111 }
|
Chris@10
|
112 {
|
Chris@10
|
113 E Tc, Tm, T2e, T2g;
|
Chris@10
|
114 Tc = T6 + Tb;
|
Chris@10
|
115 T2n = T6 - Tb;
|
Chris@10
|
116 T2b = Ti - Tl;
|
Chris@10
|
117 Tm = Ti + Tl;
|
Chris@10
|
118 T2e = Tt + Tu;
|
Chris@10
|
119 Tv = Tt - Tu;
|
Chris@10
|
120 Ts = Tq - Tr;
|
Chris@10
|
121 T2g = Tr + Tq;
|
Chris@10
|
122 {
|
Chris@10
|
123 E Tz, TA, Tn, T2f;
|
Chris@10
|
124 Tz = Tx - Ty;
|
Chris@10
|
125 T2a = Tx + Ty;
|
Chris@10
|
126 TA = FNMS(KP500000000, Tm, Tf);
|
Chris@10
|
127 Tn = Tf + Tm;
|
Chris@10
|
128 T2f = FNMS(KP500000000, T2e, T2d);
|
Chris@10
|
129 T2o = T2d + T2e;
|
Chris@10
|
130 To = Tc + Tn;
|
Chris@10
|
131 TH = Tc - Tn;
|
Chris@10
|
132 T2h = FMA(KP866025403, T2g, T2f);
|
Chris@10
|
133 T2k = FNMS(KP866025403, T2g, T2f);
|
Chris@10
|
134 TE = Tz - TA;
|
Chris@10
|
135 TB = Tz + TA;
|
Chris@10
|
136 }
|
Chris@10
|
137 }
|
Chris@10
|
138 }
|
Chris@10
|
139 {
|
Chris@10
|
140 E T1R, TM, T10, T18, T1l, TX, T1k, T15, TP, T1a, T1b, TS;
|
Chris@10
|
141 {
|
Chris@10
|
142 E T16, TY, TZ, TK, TL;
|
Chris@10
|
143 TK = ii[WS(is, 8)];
|
Chris@10
|
144 TF = Ts - Tv;
|
Chris@10
|
145 Tw = Ts + Tv;
|
Chris@10
|
146 T2j = FNMS(KP866025403, T2b, T2a);
|
Chris@10
|
147 T2c = FMA(KP866025403, T2b, T2a);
|
Chris@10
|
148 TL = ii[WS(is, 5)];
|
Chris@10
|
149 T16 = ii[WS(is, 12)];
|
Chris@10
|
150 TY = ii[WS(is, 10)];
|
Chris@10
|
151 TZ = ii[WS(is, 4)];
|
Chris@10
|
152 T1R = TK + TL;
|
Chris@10
|
153 TM = TK - TL;
|
Chris@10
|
154 {
|
Chris@10
|
155 E T13, T17, TV, TW;
|
Chris@10
|
156 T13 = ii[WS(is, 1)];
|
Chris@10
|
157 T17 = TY + TZ;
|
Chris@10
|
158 T10 = TY - TZ;
|
Chris@10
|
159 TV = ii[WS(is, 9)];
|
Chris@10
|
160 TW = ii[WS(is, 3)];
|
Chris@10
|
161 {
|
Chris@10
|
162 E TN, T14, TO, TQ, TR;
|
Chris@10
|
163 TN = ii[WS(is, 11)];
|
Chris@10
|
164 T18 = FMS(KP500000000, T17, T16);
|
Chris@10
|
165 T1l = T16 + T17;
|
Chris@10
|
166 TX = TV - TW;
|
Chris@10
|
167 T14 = TW + TV;
|
Chris@10
|
168 TO = ii[WS(is, 6)];
|
Chris@10
|
169 TQ = ii[WS(is, 7)];
|
Chris@10
|
170 TR = ii[WS(is, 2)];
|
Chris@10
|
171 T1k = T13 + T14;
|
Chris@10
|
172 T15 = FNMS(KP500000000, T14, T13);
|
Chris@10
|
173 TP = TN - TO;
|
Chris@10
|
174 T1a = TN + TO;
|
Chris@10
|
175 T1b = TQ + TR;
|
Chris@10
|
176 TS = TQ - TR;
|
Chris@10
|
177 }
|
Chris@10
|
178 }
|
Chris@10
|
179 }
|
Chris@10
|
180 {
|
Chris@10
|
181 E T1Q, T11, TT, T1S;
|
Chris@10
|
182 T1Q = T1k + T1l;
|
Chris@10
|
183 T1m = T1k - T1l;
|
Chris@10
|
184 T11 = TX + T10;
|
Chris@10
|
185 T1W = T10 - TX;
|
Chris@10
|
186 T1X = TP - TS;
|
Chris@10
|
187 TT = TP + TS;
|
Chris@10
|
188 T1S = T1a + T1b;
|
Chris@10
|
189 T1c = T1a - T1b;
|
Chris@10
|
190 {
|
Chris@10
|
191 E T1Z, TU, T1T, T20;
|
Chris@10
|
192 T19 = T15 + T18;
|
Chris@10
|
193 T1Z = T15 - T18;
|
Chris@10
|
194 T1j = TM + TT;
|
Chris@10
|
195 TU = FNMS(KP500000000, TT, TM);
|
Chris@10
|
196 T1T = T1R + T1S;
|
Chris@10
|
197 T20 = FNMS(KP500000000, T1S, T1R);
|
Chris@10
|
198 T12 = FMA(KP866025403, T11, TU);
|
Chris@10
|
199 T1f = FNMS(KP866025403, T11, TU);
|
Chris@10
|
200 T21 = T1Z + T20;
|
Chris@10
|
201 T24 = T1Z - T20;
|
Chris@10
|
202 T27 = T1Q - T1T;
|
Chris@10
|
203 T1U = T1Q + T1T;
|
Chris@10
|
204 }
|
Chris@10
|
205 }
|
Chris@10
|
206 }
|
Chris@10
|
207 }
|
Chris@10
|
208 {
|
Chris@10
|
209 E T1g, T1d, T25, T1Y;
|
Chris@10
|
210 ro[0] = T1 + To;
|
Chris@10
|
211 T1g = FNMS(KP866025403, T1c, T19);
|
Chris@10
|
212 T1d = FMA(KP866025403, T1c, T19);
|
Chris@10
|
213 T25 = T1W - T1X;
|
Chris@10
|
214 T1Y = T1W + T1X;
|
Chris@10
|
215 io[0] = T1P + T1U;
|
Chris@10
|
216 {
|
Chris@10
|
217 E T1C, T1B, T1F, T1K;
|
Chris@10
|
218 {
|
Chris@10
|
219 E TC, T1J, T1z, T1w, T1I, T1O, Tp, T1E, T1q, TI, T1o, T1s;
|
Chris@10
|
220 {
|
Chris@10
|
221 E TG, T1n, T1G, T1u, T1e, T1h, T1v, T1x, T1y, T1H, T1i;
|
Chris@10
|
222 TC = FMA(KP301479260, TB, Tw);
|
Chris@10
|
223 T1x = FNMS(KP226109445, Tw, TB);
|
Chris@10
|
224 T1y = FMA(KP686558370, TE, TF);
|
Chris@10
|
225 TG = FNMS(KP514918778, TF, TE);
|
Chris@10
|
226 T1n = FNMS(KP302775637, T1m, T1j);
|
Chris@10
|
227 T1G = FMA(KP302775637, T1j, T1m);
|
Chris@10
|
228 T1u = FNMS(KP038632954, T12, T1d);
|
Chris@10
|
229 T1e = FMA(KP038632954, T1d, T12);
|
Chris@10
|
230 T1h = FMA(KP612264650, T1g, T1f);
|
Chris@10
|
231 T1v = FNMS(KP612264650, T1f, T1g);
|
Chris@10
|
232 T1J = FMA(KP769338817, T1y, T1x);
|
Chris@10
|
233 T1z = FNMS(KP769338817, T1y, T1x);
|
Chris@10
|
234 T1H = FNMS(KP853480001, T1v, T1u);
|
Chris@10
|
235 T1w = FMA(KP853480001, T1v, T1u);
|
Chris@10
|
236 T1I = FNMS(KP522026385, T1H, T1G);
|
Chris@10
|
237 T1O = FMA(KP957805992, T1G, T1H);
|
Chris@10
|
238 Tp = FNMS(KP083333333, To, T1);
|
Chris@10
|
239 T1E = FMA(KP853480001, T1h, T1e);
|
Chris@10
|
240 T1i = FNMS(KP853480001, T1h, T1e);
|
Chris@10
|
241 T1q = FNMS(KP859542535, TG, TH);
|
Chris@10
|
242 TI = FMA(KP581704778, TH, TG);
|
Chris@10
|
243 T1o = FMA(KP957805992, T1n, T1i);
|
Chris@10
|
244 T1s = FNMS(KP522026385, T1i, T1n);
|
Chris@10
|
245 }
|
Chris@10
|
246 {
|
Chris@10
|
247 E T1A, T1D, T1t, T1L, T1M;
|
Chris@10
|
248 {
|
Chris@10
|
249 E T1p, TD, TJ, T1N, T1r;
|
Chris@10
|
250 T1p = FNMS(KP251768516, TC, Tp);
|
Chris@10
|
251 TD = FMA(KP503537032, TC, Tp);
|
Chris@10
|
252 T1C = FNMS(KP968287244, T1z, T1w);
|
Chris@10
|
253 T1A = FMA(KP968287244, T1z, T1w);
|
Chris@10
|
254 TJ = FMA(KP516520780, TI, TD);
|
Chris@10
|
255 T1N = FNMS(KP516520780, TI, TD);
|
Chris@10
|
256 T1D = FNMS(KP300462606, T1q, T1p);
|
Chris@10
|
257 T1r = FMA(KP300462606, T1q, T1p);
|
Chris@10
|
258 ro[WS(os, 8)] = FNMS(KP600477271, T1O, T1N);
|
Chris@10
|
259 ro[WS(os, 12)] = FMA(KP600477271, T1o, TJ);
|
Chris@10
|
260 ro[WS(os, 1)] = FNMS(KP600477271, T1o, TJ);
|
Chris@10
|
261 T1t = FNMS(KP575140729, T1s, T1r);
|
Chris@10
|
262 T1B = FMA(KP575140729, T1s, T1r);
|
Chris@10
|
263 ro[WS(os, 5)] = FMA(KP600477271, T1O, T1N);
|
Chris@10
|
264 }
|
Chris@10
|
265 T1L = FNMS(KP520028571, T1E, T1D);
|
Chris@10
|
266 T1F = FMA(KP520028571, T1E, T1D);
|
Chris@10
|
267 T1K = FMA(KP875502302, T1J, T1I);
|
Chris@10
|
268 T1M = FNMS(KP875502302, T1J, T1I);
|
Chris@10
|
269 ro[WS(os, 3)] = FMA(KP520028571, T1A, T1t);
|
Chris@10
|
270 ro[WS(os, 9)] = FNMS(KP520028571, T1A, T1t);
|
Chris@10
|
271 ro[WS(os, 6)] = FMA(KP575140729, T1M, T1L);
|
Chris@10
|
272 ro[WS(os, 11)] = FNMS(KP575140729, T1M, T1L);
|
Chris@10
|
273 }
|
Chris@10
|
274 }
|
Chris@10
|
275 {
|
Chris@10
|
276 E T22, T2F, T2N, T2K, T2w, T2A, T1V, T2C, T28, T2y, T2M, T2q;
|
Chris@10
|
277 {
|
Chris@10
|
278 E T26, T2v, T2p, T2i, T2s, T2t, T2l, T2D, T2E, T2u, T2m;
|
Chris@10
|
279 T2D = FNMS(KP226109445, T1Y, T21);
|
Chris@10
|
280 T22 = FMA(KP301479260, T21, T1Y);
|
Chris@10
|
281 ro[WS(os, 2)] = FMA(KP575140729, T1K, T1F);
|
Chris@10
|
282 ro[WS(os, 7)] = FNMS(KP575140729, T1K, T1F);
|
Chris@10
|
283 ro[WS(os, 4)] = FMA(KP520028571, T1C, T1B);
|
Chris@10
|
284 ro[WS(os, 10)] = FNMS(KP520028571, T1C, T1B);
|
Chris@10
|
285 T26 = FNMS(KP514918778, T25, T24);
|
Chris@10
|
286 T2E = FMA(KP686558370, T24, T25);
|
Chris@10
|
287 T2v = FNMS(KP302775637, T2n, T2o);
|
Chris@10
|
288 T2p = FMA(KP302775637, T2o, T2n);
|
Chris@10
|
289 T2i = FNMS(KP038632954, T2h, T2c);
|
Chris@10
|
290 T2s = FMA(KP038632954, T2c, T2h);
|
Chris@10
|
291 T2t = FMA(KP612264650, T2j, T2k);
|
Chris@10
|
292 T2l = FNMS(KP612264650, T2k, T2j);
|
Chris@10
|
293 T2F = FNMS(KP769338817, T2E, T2D);
|
Chris@10
|
294 T2N = FMA(KP769338817, T2E, T2D);
|
Chris@10
|
295 T2K = FMA(KP853480001, T2t, T2s);
|
Chris@10
|
296 T2u = FNMS(KP853480001, T2t, T2s);
|
Chris@10
|
297 T2w = FMA(KP957805992, T2v, T2u);
|
Chris@10
|
298 T2A = FNMS(KP522026385, T2u, T2v);
|
Chris@10
|
299 T1V = FNMS(KP083333333, T1U, T1P);
|
Chris@10
|
300 T2m = FNMS(KP853480001, T2l, T2i);
|
Chris@10
|
301 T2C = FMA(KP853480001, T2l, T2i);
|
Chris@10
|
302 T28 = FMA(KP581704778, T27, T26);
|
Chris@10
|
303 T2y = FNMS(KP859542535, T26, T27);
|
Chris@10
|
304 T2M = FNMS(KP522026385, T2m, T2p);
|
Chris@10
|
305 T2q = FMA(KP957805992, T2p, T2m);
|
Chris@10
|
306 }
|
Chris@10
|
307 {
|
Chris@10
|
308 E T2O, T2Q, T2z, T2P, T2L;
|
Chris@10
|
309 {
|
Chris@10
|
310 E T23, T2x, T2r, T29, T2J;
|
Chris@10
|
311 T23 = FMA(KP503537032, T22, T1V);
|
Chris@10
|
312 T2x = FNMS(KP251768516, T22, T1V);
|
Chris@10
|
313 T2O = FNMS(KP875502302, T2N, T2M);
|
Chris@10
|
314 T2Q = FMA(KP875502302, T2N, T2M);
|
Chris@10
|
315 T2r = FMA(KP516520780, T28, T23);
|
Chris@10
|
316 T29 = FNMS(KP516520780, T28, T23);
|
Chris@10
|
317 T2z = FMA(KP300462606, T2y, T2x);
|
Chris@10
|
318 T2J = FNMS(KP300462606, T2y, T2x);
|
Chris@10
|
319 io[WS(os, 12)] = FNMS(KP600477271, T2w, T2r);
|
Chris@10
|
320 io[WS(os, 1)] = FMA(KP600477271, T2w, T2r);
|
Chris@10
|
321 io[WS(os, 8)] = FMA(KP600477271, T2q, T29);
|
Chris@10
|
322 io[WS(os, 5)] = FNMS(KP600477271, T2q, T29);
|
Chris@10
|
323 T2P = FMA(KP520028571, T2K, T2J);
|
Chris@10
|
324 T2L = FNMS(KP520028571, T2K, T2J);
|
Chris@10
|
325 }
|
Chris@10
|
326 T2B = FMA(KP575140729, T2A, T2z);
|
Chris@10
|
327 T2H = FNMS(KP575140729, T2A, T2z);
|
Chris@10
|
328 io[WS(os, 11)] = FMA(KP575140729, T2Q, T2P);
|
Chris@10
|
329 io[WS(os, 6)] = FNMS(KP575140729, T2Q, T2P);
|
Chris@10
|
330 io[WS(os, 7)] = FMA(KP575140729, T2O, T2L);
|
Chris@10
|
331 io[WS(os, 2)] = FNMS(KP575140729, T2O, T2L);
|
Chris@10
|
332 T2I = FMA(KP968287244, T2F, T2C);
|
Chris@10
|
333 T2G = FNMS(KP968287244, T2F, T2C);
|
Chris@10
|
334 }
|
Chris@10
|
335 }
|
Chris@10
|
336 }
|
Chris@10
|
337 }
|
Chris@10
|
338 }
|
Chris@10
|
339 io[WS(os, 10)] = FMA(KP520028571, T2I, T2H);
|
Chris@10
|
340 io[WS(os, 4)] = FNMS(KP520028571, T2I, T2H);
|
Chris@10
|
341 io[WS(os, 9)] = FMA(KP520028571, T2G, T2B);
|
Chris@10
|
342 io[WS(os, 3)] = FNMS(KP520028571, T2G, T2B);
|
Chris@10
|
343 }
|
Chris@10
|
344 }
|
Chris@10
|
345 }
|
Chris@10
|
346
|
Chris@10
|
347 static const kdft_desc desc = { 13, "n1_13", {62, 0, 114, 0}, &GENUS, 0, 0, 0, 0 };
|
Chris@10
|
348
|
Chris@10
|
349 void X(codelet_n1_13) (planner *p) {
|
Chris@10
|
350 X(kdft_register) (p, n1_13, &desc);
|
Chris@10
|
351 }
|
Chris@10
|
352
|
Chris@10
|
353 #else /* HAVE_FMA */
|
Chris@10
|
354
|
Chris@10
|
355 /* Generated by: ../../../genfft/gen_notw.native -compact -variables 4 -pipeline-latency 4 -n 13 -name n1_13 -include n.h */
|
Chris@10
|
356
|
Chris@10
|
357 /*
|
Chris@10
|
358 * This function contains 176 FP additions, 68 FP multiplications,
|
Chris@10
|
359 * (or, 138 additions, 30 multiplications, 38 fused multiply/add),
|
Chris@10
|
360 * 71 stack variables, 20 constants, and 52 memory accesses
|
Chris@10
|
361 */
|
Chris@10
|
362 #include "n.h"
|
Chris@10
|
363
|
Chris@10
|
364 static void n1_13(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
|
Chris@10
|
365 {
|
Chris@10
|
366 DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
|
Chris@10
|
367 DK(KP083333333, +0.083333333333333333333333333333333333333333333);
|
Chris@10
|
368 DK(KP251768516, +0.251768516431883313623436926934233488546674281);
|
Chris@10
|
369 DK(KP075902986, +0.075902986037193865983102897245103540356428373);
|
Chris@10
|
370 DK(KP132983124, +0.132983124607418643793760531921092974399165133);
|
Chris@10
|
371 DK(KP258260390, +0.258260390311744861420450644284508567852516811);
|
Chris@10
|
372 DK(KP1_732050807, +1.732050807568877293527446341505872366942805254);
|
Chris@10
|
373 DK(KP300238635, +0.300238635966332641462884626667381504676006424);
|
Chris@10
|
374 DK(KP011599105, +0.011599105605768290721655456654083252189827041);
|
Chris@10
|
375 DK(KP156891391, +0.156891391051584611046832726756003269660212636);
|
Chris@10
|
376 DK(KP256247671, +0.256247671582936600958684654061725059144125175);
|
Chris@10
|
377 DK(KP174138601, +0.174138601152135905005660794929264742616964676);
|
Chris@10
|
378 DK(KP575140729, +0.575140729474003121368385547455453388461001608);
|
Chris@10
|
379 DK(KP503537032, +0.503537032863766627246873853868466977093348562);
|
Chris@10
|
380 DK(KP113854479, +0.113854479055790798974654345867655310534642560);
|
Chris@10
|
381 DK(KP265966249, +0.265966249214837287587521063842185948798330267);
|
Chris@10
|
382 DK(KP387390585, +0.387390585467617292130675966426762851778775217);
|
Chris@10
|
383 DK(KP866025403, +0.866025403784438646763723170752936183471402627);
|
Chris@10
|
384 DK(KP300462606, +0.300462606288665774426601772289207995520941381);
|
Chris@10
|
385 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
|
Chris@10
|
386 {
|
Chris@10
|
387 INT i;
|
Chris@10
|
388 for (i = v; i > 0; i = i - 1, ri = ri + ivs, ii = ii + ivs, ro = ro + ovs, io = io + ovs, MAKE_VOLATILE_STRIDE(52, is), MAKE_VOLATILE_STRIDE(52, os)) {
|
Chris@10
|
389 E T1, T1q, Tt, Tu, To, T22, T20, T24, TF, TH, TA, TI, T1X, T25, T2a;
|
Chris@10
|
390 E T2d, T18, T1n, T2k, T2n, T1l, T1r, T1f, T1o, T2h, T2m;
|
Chris@10
|
391 T1 = ri[0];
|
Chris@10
|
392 T1q = ii[0];
|
Chris@10
|
393 {
|
Chris@10
|
394 E Tf, Tp, Tb, TC, Tx, T6, TB, Tw, Ti, Tq, Tl, Tr, Tm, Ts, Td;
|
Chris@10
|
395 E Te, Tc, Tn;
|
Chris@10
|
396 Td = ri[WS(is, 8)];
|
Chris@10
|
397 Te = ri[WS(is, 5)];
|
Chris@10
|
398 Tf = Td + Te;
|
Chris@10
|
399 Tp = Td - Te;
|
Chris@10
|
400 {
|
Chris@10
|
401 E T7, T8, T9, Ta;
|
Chris@10
|
402 T7 = ri[WS(is, 12)];
|
Chris@10
|
403 T8 = ri[WS(is, 10)];
|
Chris@10
|
404 T9 = ri[WS(is, 4)];
|
Chris@10
|
405 Ta = T8 + T9;
|
Chris@10
|
406 Tb = T7 + Ta;
|
Chris@10
|
407 TC = T8 - T9;
|
Chris@10
|
408 Tx = FNMS(KP500000000, Ta, T7);
|
Chris@10
|
409 }
|
Chris@10
|
410 {
|
Chris@10
|
411 E T2, T3, T4, T5;
|
Chris@10
|
412 T2 = ri[WS(is, 1)];
|
Chris@10
|
413 T3 = ri[WS(is, 3)];
|
Chris@10
|
414 T4 = ri[WS(is, 9)];
|
Chris@10
|
415 T5 = T3 + T4;
|
Chris@10
|
416 T6 = T2 + T5;
|
Chris@10
|
417 TB = T3 - T4;
|
Chris@10
|
418 Tw = FNMS(KP500000000, T5, T2);
|
Chris@10
|
419 }
|
Chris@10
|
420 {
|
Chris@10
|
421 E Tg, Th, Tj, Tk;
|
Chris@10
|
422 Tg = ri[WS(is, 11)];
|
Chris@10
|
423 Th = ri[WS(is, 6)];
|
Chris@10
|
424 Ti = Tg + Th;
|
Chris@10
|
425 Tq = Tg - Th;
|
Chris@10
|
426 Tj = ri[WS(is, 7)];
|
Chris@10
|
427 Tk = ri[WS(is, 2)];
|
Chris@10
|
428 Tl = Tj + Tk;
|
Chris@10
|
429 Tr = Tj - Tk;
|
Chris@10
|
430 }
|
Chris@10
|
431 Tm = Ti + Tl;
|
Chris@10
|
432 Ts = Tq + Tr;
|
Chris@10
|
433 Tt = Tp + Ts;
|
Chris@10
|
434 Tu = T6 - Tb;
|
Chris@10
|
435 Tc = T6 + Tb;
|
Chris@10
|
436 Tn = Tf + Tm;
|
Chris@10
|
437 To = Tc + Tn;
|
Chris@10
|
438 T22 = KP300462606 * (Tc - Tn);
|
Chris@10
|
439 {
|
Chris@10
|
440 E T1Y, T1Z, TD, TE;
|
Chris@10
|
441 T1Y = TB + TC;
|
Chris@10
|
442 T1Z = Tq - Tr;
|
Chris@10
|
443 T20 = T1Y - T1Z;
|
Chris@10
|
444 T24 = T1Y + T1Z;
|
Chris@10
|
445 TD = KP866025403 * (TB - TC);
|
Chris@10
|
446 TE = FNMS(KP500000000, Ts, Tp);
|
Chris@10
|
447 TF = TD - TE;
|
Chris@10
|
448 TH = TD + TE;
|
Chris@10
|
449 }
|
Chris@10
|
450 {
|
Chris@10
|
451 E Ty, Tz, T1V, T1W;
|
Chris@10
|
452 Ty = Tw - Tx;
|
Chris@10
|
453 Tz = KP866025403 * (Ti - Tl);
|
Chris@10
|
454 TA = Ty + Tz;
|
Chris@10
|
455 TI = Ty - Tz;
|
Chris@10
|
456 T1V = Tw + Tx;
|
Chris@10
|
457 T1W = FNMS(KP500000000, Tm, Tf);
|
Chris@10
|
458 T1X = T1V - T1W;
|
Chris@10
|
459 T25 = T1V + T1W;
|
Chris@10
|
460 }
|
Chris@10
|
461 }
|
Chris@10
|
462 {
|
Chris@10
|
463 E TZ, T2b, TV, T1i, T1a, TQ, T1h, T19, T12, T1d, T15, T1c, T16, T2c, TX;
|
Chris@10
|
464 E TY, TW, T17;
|
Chris@10
|
465 TX = ii[WS(is, 8)];
|
Chris@10
|
466 TY = ii[WS(is, 5)];
|
Chris@10
|
467 TZ = TX + TY;
|
Chris@10
|
468 T2b = TX - TY;
|
Chris@10
|
469 {
|
Chris@10
|
470 E TR, TS, TT, TU;
|
Chris@10
|
471 TR = ii[WS(is, 12)];
|
Chris@10
|
472 TS = ii[WS(is, 10)];
|
Chris@10
|
473 TT = ii[WS(is, 4)];
|
Chris@10
|
474 TU = TS + TT;
|
Chris@10
|
475 TV = FNMS(KP500000000, TU, TR);
|
Chris@10
|
476 T1i = TR + TU;
|
Chris@10
|
477 T1a = TS - TT;
|
Chris@10
|
478 }
|
Chris@10
|
479 {
|
Chris@10
|
480 E TM, TN, TO, TP;
|
Chris@10
|
481 TM = ii[WS(is, 1)];
|
Chris@10
|
482 TN = ii[WS(is, 3)];
|
Chris@10
|
483 TO = ii[WS(is, 9)];
|
Chris@10
|
484 TP = TN + TO;
|
Chris@10
|
485 TQ = FNMS(KP500000000, TP, TM);
|
Chris@10
|
486 T1h = TM + TP;
|
Chris@10
|
487 T19 = TN - TO;
|
Chris@10
|
488 }
|
Chris@10
|
489 {
|
Chris@10
|
490 E T10, T11, T13, T14;
|
Chris@10
|
491 T10 = ii[WS(is, 11)];
|
Chris@10
|
492 T11 = ii[WS(is, 6)];
|
Chris@10
|
493 T12 = T10 + T11;
|
Chris@10
|
494 T1d = T10 - T11;
|
Chris@10
|
495 T13 = ii[WS(is, 7)];
|
Chris@10
|
496 T14 = ii[WS(is, 2)];
|
Chris@10
|
497 T15 = T13 + T14;
|
Chris@10
|
498 T1c = T13 - T14;
|
Chris@10
|
499 }
|
Chris@10
|
500 T16 = T12 + T15;
|
Chris@10
|
501 T2c = T1d + T1c;
|
Chris@10
|
502 T2a = T1h - T1i;
|
Chris@10
|
503 T2d = T2b + T2c;
|
Chris@10
|
504 TW = TQ + TV;
|
Chris@10
|
505 T17 = FNMS(KP500000000, T16, TZ);
|
Chris@10
|
506 T18 = TW - T17;
|
Chris@10
|
507 T1n = TW + T17;
|
Chris@10
|
508 {
|
Chris@10
|
509 E T2i, T2j, T1j, T1k;
|
Chris@10
|
510 T2i = TQ - TV;
|
Chris@10
|
511 T2j = KP866025403 * (T15 - T12);
|
Chris@10
|
512 T2k = T2i + T2j;
|
Chris@10
|
513 T2n = T2i - T2j;
|
Chris@10
|
514 T1j = T1h + T1i;
|
Chris@10
|
515 T1k = TZ + T16;
|
Chris@10
|
516 T1l = KP300462606 * (T1j - T1k);
|
Chris@10
|
517 T1r = T1j + T1k;
|
Chris@10
|
518 }
|
Chris@10
|
519 {
|
Chris@10
|
520 E T1b, T1e, T2f, T2g;
|
Chris@10
|
521 T1b = T19 + T1a;
|
Chris@10
|
522 T1e = T1c - T1d;
|
Chris@10
|
523 T1f = T1b + T1e;
|
Chris@10
|
524 T1o = T1e - T1b;
|
Chris@10
|
525 T2f = FNMS(KP500000000, T2c, T2b);
|
Chris@10
|
526 T2g = KP866025403 * (T1a - T19);
|
Chris@10
|
527 T2h = T2f - T2g;
|
Chris@10
|
528 T2m = T2g + T2f;
|
Chris@10
|
529 }
|
Chris@10
|
530 }
|
Chris@10
|
531 ro[0] = T1 + To;
|
Chris@10
|
532 io[0] = T1q + T1r;
|
Chris@10
|
533 {
|
Chris@10
|
534 E T1D, T1N, T1y, T1x, T1E, T1O, Tv, TK, T1J, T1Q, T1m, T1R, T1t, T1I, TG;
|
Chris@10
|
535 E TJ;
|
Chris@10
|
536 {
|
Chris@10
|
537 E T1B, T1C, T1v, T1w;
|
Chris@10
|
538 T1B = FMA(KP387390585, T1f, KP265966249 * T18);
|
Chris@10
|
539 T1C = FMA(KP113854479, T1o, KP503537032 * T1n);
|
Chris@10
|
540 T1D = T1B + T1C;
|
Chris@10
|
541 T1N = T1C - T1B;
|
Chris@10
|
542 T1y = FMA(KP575140729, Tu, KP174138601 * Tt);
|
Chris@10
|
543 T1v = FNMS(KP156891391, TH, KP256247671 * TI);
|
Chris@10
|
544 T1w = FMA(KP011599105, TF, KP300238635 * TA);
|
Chris@10
|
545 T1x = T1v - T1w;
|
Chris@10
|
546 T1E = T1y + T1x;
|
Chris@10
|
547 T1O = KP1_732050807 * (T1v + T1w);
|
Chris@10
|
548 }
|
Chris@10
|
549 Tv = FNMS(KP174138601, Tu, KP575140729 * Tt);
|
Chris@10
|
550 TG = FNMS(KP300238635, TF, KP011599105 * TA);
|
Chris@10
|
551 TJ = FMA(KP256247671, TH, KP156891391 * TI);
|
Chris@10
|
552 TK = TG - TJ;
|
Chris@10
|
553 T1J = KP1_732050807 * (TJ + TG);
|
Chris@10
|
554 T1Q = Tv - TK;
|
Chris@10
|
555 {
|
Chris@10
|
556 E T1g, T1H, T1p, T1s, T1G;
|
Chris@10
|
557 T1g = FNMS(KP132983124, T1f, KP258260390 * T18);
|
Chris@10
|
558 T1H = T1l - T1g;
|
Chris@10
|
559 T1p = FNMS(KP251768516, T1o, KP075902986 * T1n);
|
Chris@10
|
560 T1s = FNMS(KP083333333, T1r, T1q);
|
Chris@10
|
561 T1G = T1s - T1p;
|
Chris@10
|
562 T1m = FMA(KP2_000000000, T1g, T1l);
|
Chris@10
|
563 T1R = T1H + T1G;
|
Chris@10
|
564 T1t = FMA(KP2_000000000, T1p, T1s);
|
Chris@10
|
565 T1I = T1G - T1H;
|
Chris@10
|
566 }
|
Chris@10
|
567 {
|
Chris@10
|
568 E TL, T1u, T1P, T1S;
|
Chris@10
|
569 TL = FMA(KP2_000000000, TK, Tv);
|
Chris@10
|
570 T1u = T1m + T1t;
|
Chris@10
|
571 io[WS(os, 1)] = TL + T1u;
|
Chris@10
|
572 io[WS(os, 12)] = T1u - TL;
|
Chris@10
|
573 {
|
Chris@10
|
574 E T1z, T1A, T1T, T1U;
|
Chris@10
|
575 T1z = FMS(KP2_000000000, T1x, T1y);
|
Chris@10
|
576 T1A = T1t - T1m;
|
Chris@10
|
577 io[WS(os, 5)] = T1z + T1A;
|
Chris@10
|
578 io[WS(os, 8)] = T1A - T1z;
|
Chris@10
|
579 T1T = T1R - T1Q;
|
Chris@10
|
580 T1U = T1O + T1N;
|
Chris@10
|
581 io[WS(os, 4)] = T1T - T1U;
|
Chris@10
|
582 io[WS(os, 10)] = T1U + T1T;
|
Chris@10
|
583 }
|
Chris@10
|
584 T1P = T1N - T1O;
|
Chris@10
|
585 T1S = T1Q + T1R;
|
Chris@10
|
586 io[WS(os, 3)] = T1P + T1S;
|
Chris@10
|
587 io[WS(os, 9)] = T1S - T1P;
|
Chris@10
|
588 {
|
Chris@10
|
589 E T1L, T1M, T1F, T1K;
|
Chris@10
|
590 T1L = T1J + T1I;
|
Chris@10
|
591 T1M = T1E + T1D;
|
Chris@10
|
592 io[WS(os, 6)] = T1L - T1M;
|
Chris@10
|
593 io[WS(os, 11)] = T1M + T1L;
|
Chris@10
|
594 T1F = T1D - T1E;
|
Chris@10
|
595 T1K = T1I - T1J;
|
Chris@10
|
596 io[WS(os, 2)] = T1F + T1K;
|
Chris@10
|
597 io[WS(os, 7)] = T1K - T1F;
|
Chris@10
|
598 }
|
Chris@10
|
599 }
|
Chris@10
|
600 }
|
Chris@10
|
601 {
|
Chris@10
|
602 E T2y, T2I, T2J, T2K, T2B, T2L, T2e, T2p, T2u, T2G, T23, T2F, T28, T2t, T2l;
|
Chris@10
|
603 E T2o;
|
Chris@10
|
604 {
|
Chris@10
|
605 E T2w, T2x, T2z, T2A;
|
Chris@10
|
606 T2w = FMA(KP387390585, T20, KP265966249 * T1X);
|
Chris@10
|
607 T2x = FNMS(KP503537032, T25, KP113854479 * T24);
|
Chris@10
|
608 T2y = T2w + T2x;
|
Chris@10
|
609 T2I = T2w - T2x;
|
Chris@10
|
610 T2J = FMA(KP575140729, T2a, KP174138601 * T2d);
|
Chris@10
|
611 T2z = FNMS(KP300238635, T2n, KP011599105 * T2m);
|
Chris@10
|
612 T2A = FNMS(KP156891391, T2h, KP256247671 * T2k);
|
Chris@10
|
613 T2K = T2z + T2A;
|
Chris@10
|
614 T2B = KP1_732050807 * (T2z - T2A);
|
Chris@10
|
615 T2L = T2J + T2K;
|
Chris@10
|
616 }
|
Chris@10
|
617 T2e = FNMS(KP575140729, T2d, KP174138601 * T2a);
|
Chris@10
|
618 T2l = FMA(KP256247671, T2h, KP156891391 * T2k);
|
Chris@10
|
619 T2o = FMA(KP300238635, T2m, KP011599105 * T2n);
|
Chris@10
|
620 T2p = T2l - T2o;
|
Chris@10
|
621 T2u = T2e - T2p;
|
Chris@10
|
622 T2G = KP1_732050807 * (T2o + T2l);
|
Chris@10
|
623 {
|
Chris@10
|
624 E T21, T2r, T26, T27, T2s;
|
Chris@10
|
625 T21 = FNMS(KP132983124, T20, KP258260390 * T1X);
|
Chris@10
|
626 T2r = T22 - T21;
|
Chris@10
|
627 T26 = FMA(KP251768516, T24, KP075902986 * T25);
|
Chris@10
|
628 T27 = FNMS(KP083333333, To, T1);
|
Chris@10
|
629 T2s = T27 - T26;
|
Chris@10
|
630 T23 = FMA(KP2_000000000, T21, T22);
|
Chris@10
|
631 T2F = T2s - T2r;
|
Chris@10
|
632 T28 = FMA(KP2_000000000, T26, T27);
|
Chris@10
|
633 T2t = T2r + T2s;
|
Chris@10
|
634 }
|
Chris@10
|
635 {
|
Chris@10
|
636 E T29, T2q, T2N, T2O;
|
Chris@10
|
637 T29 = T23 + T28;
|
Chris@10
|
638 T2q = FMA(KP2_000000000, T2p, T2e);
|
Chris@10
|
639 ro[WS(os, 12)] = T29 - T2q;
|
Chris@10
|
640 ro[WS(os, 1)] = T29 + T2q;
|
Chris@10
|
641 {
|
Chris@10
|
642 E T2v, T2C, T2P, T2Q;
|
Chris@10
|
643 T2v = T2t - T2u;
|
Chris@10
|
644 T2C = T2y - T2B;
|
Chris@10
|
645 ro[WS(os, 10)] = T2v - T2C;
|
Chris@10
|
646 ro[WS(os, 4)] = T2v + T2C;
|
Chris@10
|
647 T2P = T28 - T23;
|
Chris@10
|
648 T2Q = FMS(KP2_000000000, T2K, T2J);
|
Chris@10
|
649 ro[WS(os, 5)] = T2P - T2Q;
|
Chris@10
|
650 ro[WS(os, 8)] = T2P + T2Q;
|
Chris@10
|
651 }
|
Chris@10
|
652 T2N = T2F - T2G;
|
Chris@10
|
653 T2O = T2L - T2I;
|
Chris@10
|
654 ro[WS(os, 11)] = T2N - T2O;
|
Chris@10
|
655 ro[WS(os, 6)] = T2N + T2O;
|
Chris@10
|
656 {
|
Chris@10
|
657 E T2H, T2M, T2D, T2E;
|
Chris@10
|
658 T2H = T2F + T2G;
|
Chris@10
|
659 T2M = T2I + T2L;
|
Chris@10
|
660 ro[WS(os, 7)] = T2H - T2M;
|
Chris@10
|
661 ro[WS(os, 2)] = T2H + T2M;
|
Chris@10
|
662 T2D = T2t + T2u;
|
Chris@10
|
663 T2E = T2y + T2B;
|
Chris@10
|
664 ro[WS(os, 3)] = T2D - T2E;
|
Chris@10
|
665 ro[WS(os, 9)] = T2D + T2E;
|
Chris@10
|
666 }
|
Chris@10
|
667 }
|
Chris@10
|
668 }
|
Chris@10
|
669 }
|
Chris@10
|
670 }
|
Chris@10
|
671 }
|
Chris@10
|
672
|
Chris@10
|
673 static const kdft_desc desc = { 13, "n1_13", {138, 30, 38, 0}, &GENUS, 0, 0, 0, 0 };
|
Chris@10
|
674
|
Chris@10
|
675 void X(codelet_n1_13) (planner *p) {
|
Chris@10
|
676 X(kdft_register) (p, n1_13, &desc);
|
Chris@10
|
677 }
|
Chris@10
|
678
|
Chris@10
|
679 #endif /* HAVE_FMA */
|