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 07:41:37 EST 2012 */
|
Chris@10
|
23
|
Chris@10
|
24 #include "codelet-rdft.h"
|
Chris@10
|
25
|
Chris@10
|
26 #ifdef HAVE_FMA
|
Chris@10
|
27
|
Chris@10
|
28 /* Generated by: ../../../genfft/gen_r2cb.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -sign 1 -n 32 -name r2cbIII_32 -dft-III -include r2cbIII.h */
|
Chris@10
|
29
|
Chris@10
|
30 /*
|
Chris@10
|
31 * This function contains 174 FP additions, 100 FP multiplications,
|
Chris@10
|
32 * (or, 106 additions, 32 multiplications, 68 fused multiply/add),
|
Chris@10
|
33 * 101 stack variables, 18 constants, and 64 memory accesses
|
Chris@10
|
34 */
|
Chris@10
|
35 #include "r2cbIII.h"
|
Chris@10
|
36
|
Chris@10
|
37 static void r2cbIII_32(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
|
Chris@10
|
38 {
|
Chris@10
|
39 DK(KP534511135, +0.534511135950791641089685961295362908582039528);
|
Chris@10
|
40 DK(KP1_763842528, +1.763842528696710059425513727320776699016885241);
|
Chris@10
|
41 DK(KP303346683, +0.303346683607342391675883946941299872384187453);
|
Chris@10
|
42 DK(KP1_913880671, +1.913880671464417729871595773960539938965698411);
|
Chris@10
|
43 DK(KP098491403, +0.098491403357164253077197521291327432293052451);
|
Chris@10
|
44 DK(KP1_990369453, +1.990369453344393772489673906218959843150949737);
|
Chris@10
|
45 DK(KP820678790, +0.820678790828660330972281985331011598767386482);
|
Chris@10
|
46 DK(KP1_546020906, +1.546020906725473921621813219516939601942082586);
|
Chris@10
|
47 DK(KP1_847759065, +1.847759065022573512256366378793576573644833252);
|
Chris@10
|
48 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
Chris@10
|
49 DK(KP668178637, +0.668178637919298919997757686523080761552472251);
|
Chris@10
|
50 DK(KP1_662939224, +1.662939224605090474157576755235811513477121624);
|
Chris@10
|
51 DK(KP198912367, +0.198912367379658006911597622644676228597850501);
|
Chris@10
|
52 DK(KP1_961570560, +1.961570560806460898252364472268478073947867462);
|
Chris@10
|
53 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
Chris@10
|
54 DK(KP1_414213562, +1.414213562373095048801688724209698078569671875);
|
Chris@10
|
55 DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
|
Chris@10
|
56 DK(KP414213562, +0.414213562373095048801688724209698078569671875);
|
Chris@10
|
57 {
|
Chris@10
|
58 INT i;
|
Chris@10
|
59 for (i = v; i > 0; i = i - 1, R0 = R0 + ovs, R1 = R1 + ovs, Cr = Cr + ivs, Ci = Ci + ivs, MAKE_VOLATILE_STRIDE(128, rs), MAKE_VOLATILE_STRIDE(128, csr), MAKE_VOLATILE_STRIDE(128, csi)) {
|
Chris@10
|
60 E T1N, T1K, T1Q, T1H, T1O, T1P;
|
Chris@10
|
61 {
|
Chris@10
|
62 E T1I, T1e, T1Z, T7, T2E, T2i, T1x, Tz, Te, T2j, T22, T2F, T1h, T1y, TK;
|
Chris@10
|
63 E T1J, Tm, T2B, TX, Tp, T2m, T28, T1M, T1C, T1k, TW, TY, T2a, T14, T15;
|
Chris@10
|
64 E Ts, TZ;
|
Chris@10
|
65 {
|
Chris@10
|
66 E TE, T1g, TJ, T1f;
|
Chris@10
|
67 {
|
Chris@10
|
68 E T4, Tv, T3, T2g, T1d, T5, Tw, Tx;
|
Chris@10
|
69 {
|
Chris@10
|
70 E T1, T2, T1b, T1c;
|
Chris@10
|
71 T1 = Cr[0];
|
Chris@10
|
72 T2 = Cr[WS(csr, 15)];
|
Chris@10
|
73 T1b = Ci[0];
|
Chris@10
|
74 T1c = Ci[WS(csi, 15)];
|
Chris@10
|
75 T4 = Cr[WS(csr, 8)];
|
Chris@10
|
76 Tv = T1 - T2;
|
Chris@10
|
77 T3 = T1 + T2;
|
Chris@10
|
78 T2g = T1c - T1b;
|
Chris@10
|
79 T1d = T1b + T1c;
|
Chris@10
|
80 T5 = Cr[WS(csr, 7)];
|
Chris@10
|
81 Tw = Ci[WS(csi, 8)];
|
Chris@10
|
82 Tx = Ci[WS(csi, 7)];
|
Chris@10
|
83 }
|
Chris@10
|
84 {
|
Chris@10
|
85 E Tb, TA, Ta, T20, TD, Tc, TG, TH;
|
Chris@10
|
86 {
|
Chris@10
|
87 E T8, T9, TB, TC;
|
Chris@10
|
88 T8 = Cr[WS(csr, 4)];
|
Chris@10
|
89 {
|
Chris@10
|
90 E T1a, T6, T2h, Ty;
|
Chris@10
|
91 T1a = T4 - T5;
|
Chris@10
|
92 T6 = T4 + T5;
|
Chris@10
|
93 T2h = Tx - Tw;
|
Chris@10
|
94 Ty = Tw + Tx;
|
Chris@10
|
95 T1I = T1a - T1d;
|
Chris@10
|
96 T1e = T1a + T1d;
|
Chris@10
|
97 T1Z = T3 - T6;
|
Chris@10
|
98 T7 = T3 + T6;
|
Chris@10
|
99 T2E = T2h + T2g;
|
Chris@10
|
100 T2i = T2g - T2h;
|
Chris@10
|
101 T1x = Tv + Ty;
|
Chris@10
|
102 Tz = Tv - Ty;
|
Chris@10
|
103 T9 = Cr[WS(csr, 11)];
|
Chris@10
|
104 }
|
Chris@10
|
105 TB = Ci[WS(csi, 4)];
|
Chris@10
|
106 TC = Ci[WS(csi, 11)];
|
Chris@10
|
107 Tb = Cr[WS(csr, 3)];
|
Chris@10
|
108 TA = T8 - T9;
|
Chris@10
|
109 Ta = T8 + T9;
|
Chris@10
|
110 T20 = TC - TB;
|
Chris@10
|
111 TD = TB + TC;
|
Chris@10
|
112 Tc = Cr[WS(csr, 12)];
|
Chris@10
|
113 TG = Ci[WS(csi, 3)];
|
Chris@10
|
114 TH = Ci[WS(csi, 12)];
|
Chris@10
|
115 }
|
Chris@10
|
116 {
|
Chris@10
|
117 E TF, Td, T21, TI;
|
Chris@10
|
118 TE = TA - TD;
|
Chris@10
|
119 T1g = TA + TD;
|
Chris@10
|
120 TF = Tb - Tc;
|
Chris@10
|
121 Td = Tb + Tc;
|
Chris@10
|
122 T21 = TG - TH;
|
Chris@10
|
123 TI = TG + TH;
|
Chris@10
|
124 Te = Ta + Td;
|
Chris@10
|
125 T2j = Ta - Td;
|
Chris@10
|
126 T22 = T20 - T21;
|
Chris@10
|
127 T2F = T20 + T21;
|
Chris@10
|
128 TJ = TF - TI;
|
Chris@10
|
129 T1f = TF + TI;
|
Chris@10
|
130 }
|
Chris@10
|
131 }
|
Chris@10
|
132 }
|
Chris@10
|
133 {
|
Chris@10
|
134 E TM, Ti, TN, T25, TU, TR, Tl, TO;
|
Chris@10
|
135 {
|
Chris@10
|
136 E TS, TT, Tg, Th, Tj, Tk;
|
Chris@10
|
137 Tg = Cr[WS(csr, 2)];
|
Chris@10
|
138 Th = Cr[WS(csr, 13)];
|
Chris@10
|
139 T1h = T1f - T1g;
|
Chris@10
|
140 T1y = T1g + T1f;
|
Chris@10
|
141 TK = TE + TJ;
|
Chris@10
|
142 T1J = TE - TJ;
|
Chris@10
|
143 TM = Tg - Th;
|
Chris@10
|
144 Ti = Tg + Th;
|
Chris@10
|
145 TS = Ci[WS(csi, 2)];
|
Chris@10
|
146 TT = Ci[WS(csi, 13)];
|
Chris@10
|
147 Tj = Cr[WS(csr, 10)];
|
Chris@10
|
148 Tk = Cr[WS(csr, 5)];
|
Chris@10
|
149 TN = Ci[WS(csi, 10)];
|
Chris@10
|
150 T25 = TS - TT;
|
Chris@10
|
151 TU = TS + TT;
|
Chris@10
|
152 TR = Tj - Tk;
|
Chris@10
|
153 Tl = Tj + Tk;
|
Chris@10
|
154 TO = Ci[WS(csi, 5)];
|
Chris@10
|
155 }
|
Chris@10
|
156 {
|
Chris@10
|
157 E T12, T13, Tq, Tr;
|
Chris@10
|
158 {
|
Chris@10
|
159 E Tn, T1A, TV, T24, T26, TP, To, T27, T1B, TQ;
|
Chris@10
|
160 Tn = Cr[WS(csr, 1)];
|
Chris@10
|
161 T1A = TR - TU;
|
Chris@10
|
162 TV = TR + TU;
|
Chris@10
|
163 T24 = Ti - Tl;
|
Chris@10
|
164 Tm = Ti + Tl;
|
Chris@10
|
165 T26 = TN - TO;
|
Chris@10
|
166 TP = TN + TO;
|
Chris@10
|
167 To = Cr[WS(csr, 14)];
|
Chris@10
|
168 T12 = Ci[WS(csi, 1)];
|
Chris@10
|
169 T27 = T25 - T26;
|
Chris@10
|
170 T2B = T26 + T25;
|
Chris@10
|
171 T1B = TM + TP;
|
Chris@10
|
172 TQ = TM - TP;
|
Chris@10
|
173 TX = Tn - To;
|
Chris@10
|
174 Tp = Tn + To;
|
Chris@10
|
175 T2m = T24 + T27;
|
Chris@10
|
176 T28 = T24 - T27;
|
Chris@10
|
177 T1M = FNMS(KP414213562, T1A, T1B);
|
Chris@10
|
178 T1C = FMA(KP414213562, T1B, T1A);
|
Chris@10
|
179 T1k = FMA(KP414213562, TQ, TV);
|
Chris@10
|
180 TW = FNMS(KP414213562, TV, TQ);
|
Chris@10
|
181 T13 = Ci[WS(csi, 14)];
|
Chris@10
|
182 }
|
Chris@10
|
183 Tq = Cr[WS(csr, 6)];
|
Chris@10
|
184 Tr = Cr[WS(csr, 9)];
|
Chris@10
|
185 TY = Ci[WS(csi, 6)];
|
Chris@10
|
186 T2a = T13 - T12;
|
Chris@10
|
187 T14 = T12 + T13;
|
Chris@10
|
188 T15 = Tq - Tr;
|
Chris@10
|
189 Ts = Tq + Tr;
|
Chris@10
|
190 TZ = Ci[WS(csi, 9)];
|
Chris@10
|
191 }
|
Chris@10
|
192 }
|
Chris@10
|
193 }
|
Chris@10
|
194 {
|
Chris@10
|
195 E T1L, T1F, T23, T2n, T2k, T2e, T1p, T1t, T1s, T1i, T1o, T19, T1l, T1q;
|
Chris@10
|
196 {
|
Chris@10
|
197 E T2z, T2G, T2H, T2C, T1j, T17, T2r, T2s, T2u, T2v, T2K, T2D;
|
Chris@10
|
198 {
|
Chris@10
|
199 E T2L, T2d, T2l, T2O;
|
Chris@10
|
200 {
|
Chris@10
|
201 E Tf, T2N, Tu, T2M;
|
Chris@10
|
202 {
|
Chris@10
|
203 E T1D, T16, T29, Tt, T2b, T10;
|
Chris@10
|
204 T2z = T7 - Te;
|
Chris@10
|
205 Tf = T7 + Te;
|
Chris@10
|
206 T1D = T15 + T14;
|
Chris@10
|
207 T16 = T14 - T15;
|
Chris@10
|
208 T29 = Tp - Ts;
|
Chris@10
|
209 Tt = Tp + Ts;
|
Chris@10
|
210 T2b = TY - TZ;
|
Chris@10
|
211 T10 = TY + TZ;
|
Chris@10
|
212 T2N = T2F + T2E;
|
Chris@10
|
213 T2G = T2E - T2F;
|
Chris@10
|
214 T2H = Tm - Tt;
|
Chris@10
|
215 Tu = Tm + Tt;
|
Chris@10
|
216 {
|
Chris@10
|
217 E T2c, T2A, T1E, T11;
|
Chris@10
|
218 T2c = T2a - T2b;
|
Chris@10
|
219 T2A = T2b + T2a;
|
Chris@10
|
220 T1E = TX + T10;
|
Chris@10
|
221 T11 = TX - T10;
|
Chris@10
|
222 T2L = Tf - Tu;
|
Chris@10
|
223 T2d = T29 + T2c;
|
Chris@10
|
224 T2l = T29 - T2c;
|
Chris@10
|
225 T2C = T2A - T2B;
|
Chris@10
|
226 T2M = T2B + T2A;
|
Chris@10
|
227 T1L = FMA(KP414213562, T1D, T1E);
|
Chris@10
|
228 T1F = FNMS(KP414213562, T1E, T1D);
|
Chris@10
|
229 T1j = FMA(KP414213562, T11, T16);
|
Chris@10
|
230 T17 = FNMS(KP414213562, T16, T11);
|
Chris@10
|
231 T2O = T2M + T2N;
|
Chris@10
|
232 }
|
Chris@10
|
233 }
|
Chris@10
|
234 R0[0] = KP2_000000000 * (Tf + Tu);
|
Chris@10
|
235 R0[WS(rs, 8)] = KP2_000000000 * (T2N - T2M);
|
Chris@10
|
236 }
|
Chris@10
|
237 T23 = T1Z + T22;
|
Chris@10
|
238 T2r = T1Z - T22;
|
Chris@10
|
239 R0[WS(rs, 12)] = KP1_414213562 * (T2O - T2L);
|
Chris@10
|
240 R0[WS(rs, 4)] = KP1_414213562 * (T2L + T2O);
|
Chris@10
|
241 T2s = T2m + T2l;
|
Chris@10
|
242 T2n = T2l - T2m;
|
Chris@10
|
243 T2k = T2i - T2j;
|
Chris@10
|
244 T2u = T2j + T2i;
|
Chris@10
|
245 T2v = T28 - T2d;
|
Chris@10
|
246 T2e = T28 + T2d;
|
Chris@10
|
247 }
|
Chris@10
|
248 {
|
Chris@10
|
249 E T2y, T2t, T2x, T2w;
|
Chris@10
|
250 T2y = FMA(KP707106781, T2s, T2r);
|
Chris@10
|
251 T2t = FNMS(KP707106781, T2s, T2r);
|
Chris@10
|
252 T2x = FMA(KP707106781, T2v, T2u);
|
Chris@10
|
253 T2w = FNMS(KP707106781, T2v, T2u);
|
Chris@10
|
254 R0[WS(rs, 7)] = KP1_961570560 * (FMA(KP198912367, T2y, T2x));
|
Chris@10
|
255 R0[WS(rs, 15)] = -(KP1_961570560 * (FNMS(KP198912367, T2x, T2y)));
|
Chris@10
|
256 R0[WS(rs, 11)] = KP1_662939224 * (FNMS(KP668178637, T2t, T2w));
|
Chris@10
|
257 R0[WS(rs, 3)] = KP1_662939224 * (FMA(KP668178637, T2w, T2t));
|
Chris@10
|
258 T2K = T2z - T2C;
|
Chris@10
|
259 T2D = T2z + T2C;
|
Chris@10
|
260 }
|
Chris@10
|
261 {
|
Chris@10
|
262 E TL, T18, T2J, T2I;
|
Chris@10
|
263 T1p = FNMS(KP707106781, TK, Tz);
|
Chris@10
|
264 TL = FMA(KP707106781, TK, Tz);
|
Chris@10
|
265 T18 = TW + T17;
|
Chris@10
|
266 T1t = TW - T17;
|
Chris@10
|
267 T1s = FMA(KP707106781, T1h, T1e);
|
Chris@10
|
268 T1i = FNMS(KP707106781, T1h, T1e);
|
Chris@10
|
269 T2J = T2H + T2G;
|
Chris@10
|
270 T2I = T2G - T2H;
|
Chris@10
|
271 T1o = FNMS(KP923879532, T18, TL);
|
Chris@10
|
272 T19 = FMA(KP923879532, T18, TL);
|
Chris@10
|
273 R0[WS(rs, 6)] = KP1_847759065 * (FMA(KP414213562, T2K, T2J));
|
Chris@10
|
274 R0[WS(rs, 14)] = -(KP1_847759065 * (FNMS(KP414213562, T2J, T2K)));
|
Chris@10
|
275 R0[WS(rs, 10)] = KP1_847759065 * (FNMS(KP414213562, T2D, T2I));
|
Chris@10
|
276 R0[WS(rs, 2)] = KP1_847759065 * (FMA(KP414213562, T2I, T2D));
|
Chris@10
|
277 T1l = T1j - T1k;
|
Chris@10
|
278 T1q = T1k + T1j;
|
Chris@10
|
279 }
|
Chris@10
|
280 }
|
Chris@10
|
281 {
|
Chris@10
|
282 E T1z, T1U, T1Y, T1T, T1V, T1G;
|
Chris@10
|
283 {
|
Chris@10
|
284 E T1w, T1r, T1n, T1m;
|
Chris@10
|
285 T1n = FMA(KP923879532, T1l, T1i);
|
Chris@10
|
286 T1m = FNMS(KP923879532, T1l, T1i);
|
Chris@10
|
287 T1w = FMA(KP923879532, T1q, T1p);
|
Chris@10
|
288 T1r = FNMS(KP923879532, T1q, T1p);
|
Chris@10
|
289 R1[WS(rs, 4)] = -(KP1_546020906 * (FNMS(KP820678790, T1o, T1n)));
|
Chris@10
|
290 R1[WS(rs, 12)] = -(KP1_546020906 * (FMA(KP820678790, T1n, T1o)));
|
Chris@10
|
291 R1[WS(rs, 8)] = -(KP1_990369453 * (FMA(KP098491403, T19, T1m)));
|
Chris@10
|
292 R1[0] = KP1_990369453 * (FNMS(KP098491403, T1m, T19));
|
Chris@10
|
293 {
|
Chris@10
|
294 E T1R, T1S, T1v, T1u;
|
Chris@10
|
295 T1z = FNMS(KP707106781, T1y, T1x);
|
Chris@10
|
296 T1R = FMA(KP707106781, T1y, T1x);
|
Chris@10
|
297 T1S = T1M + T1L;
|
Chris@10
|
298 T1N = T1L - T1M;
|
Chris@10
|
299 T1K = FNMS(KP707106781, T1J, T1I);
|
Chris@10
|
300 T1U = FMA(KP707106781, T1J, T1I);
|
Chris@10
|
301 T1v = FNMS(KP923879532, T1t, T1s);
|
Chris@10
|
302 T1u = FMA(KP923879532, T1t, T1s);
|
Chris@10
|
303 T1Y = FMA(KP923879532, T1S, T1R);
|
Chris@10
|
304 T1T = FNMS(KP923879532, T1S, T1R);
|
Chris@10
|
305 R1[WS(rs, 6)] = -(KP1_913880671 * (FNMS(KP303346683, T1w, T1v)));
|
Chris@10
|
306 R1[WS(rs, 14)] = -(KP1_913880671 * (FMA(KP303346683, T1v, T1w)));
|
Chris@10
|
307 R1[WS(rs, 10)] = -(KP1_763842528 * (FMA(KP534511135, T1r, T1u)));
|
Chris@10
|
308 R1[WS(rs, 2)] = KP1_763842528 * (FNMS(KP534511135, T1u, T1r));
|
Chris@10
|
309 T1V = T1C + T1F;
|
Chris@10
|
310 T1G = T1C - T1F;
|
Chris@10
|
311 }
|
Chris@10
|
312 }
|
Chris@10
|
313 {
|
Chris@10
|
314 E T2q, T2f, T1X, T1W, T2p, T2o;
|
Chris@10
|
315 T1X = FMA(KP923879532, T1V, T1U);
|
Chris@10
|
316 T1W = FNMS(KP923879532, T1V, T1U);
|
Chris@10
|
317 T2q = FNMS(KP707106781, T2e, T23);
|
Chris@10
|
318 T2f = FMA(KP707106781, T2e, T23);
|
Chris@10
|
319 R1[WS(rs, 7)] = KP1_990369453 * (FMA(KP098491403, T1Y, T1X));
|
Chris@10
|
320 R1[WS(rs, 15)] = -(KP1_990369453 * (FNMS(KP098491403, T1X, T1Y)));
|
Chris@10
|
321 R1[WS(rs, 11)] = KP1_546020906 * (FNMS(KP820678790, T1T, T1W));
|
Chris@10
|
322 R1[WS(rs, 3)] = KP1_546020906 * (FMA(KP820678790, T1W, T1T));
|
Chris@10
|
323 T2p = FNMS(KP707106781, T2n, T2k);
|
Chris@10
|
324 T2o = FMA(KP707106781, T2n, T2k);
|
Chris@10
|
325 T1Q = FNMS(KP923879532, T1G, T1z);
|
Chris@10
|
326 T1H = FMA(KP923879532, T1G, T1z);
|
Chris@10
|
327 R0[WS(rs, 5)] = KP1_662939224 * (FMA(KP668178637, T2q, T2p));
|
Chris@10
|
328 R0[WS(rs, 13)] = -(KP1_662939224 * (FNMS(KP668178637, T2p, T2q)));
|
Chris@10
|
329 R0[WS(rs, 9)] = KP1_961570560 * (FNMS(KP198912367, T2f, T2o));
|
Chris@10
|
330 R0[WS(rs, 1)] = KP1_961570560 * (FMA(KP198912367, T2o, T2f));
|
Chris@10
|
331 }
|
Chris@10
|
332 }
|
Chris@10
|
333 }
|
Chris@10
|
334 }
|
Chris@10
|
335 T1O = FMA(KP923879532, T1N, T1K);
|
Chris@10
|
336 T1P = FNMS(KP923879532, T1N, T1K);
|
Chris@10
|
337 R1[WS(rs, 5)] = KP1_763842528 * (FMA(KP534511135, T1Q, T1P));
|
Chris@10
|
338 R1[WS(rs, 13)] = -(KP1_763842528 * (FNMS(KP534511135, T1P, T1Q)));
|
Chris@10
|
339 R1[WS(rs, 9)] = KP1_913880671 * (FNMS(KP303346683, T1H, T1O));
|
Chris@10
|
340 R1[WS(rs, 1)] = KP1_913880671 * (FMA(KP303346683, T1O, T1H));
|
Chris@10
|
341 }
|
Chris@10
|
342 }
|
Chris@10
|
343 }
|
Chris@10
|
344
|
Chris@10
|
345 static const kr2c_desc desc = { 32, "r2cbIII_32", {106, 32, 68, 0}, &GENUS };
|
Chris@10
|
346
|
Chris@10
|
347 void X(codelet_r2cbIII_32) (planner *p) {
|
Chris@10
|
348 X(kr2c_register) (p, r2cbIII_32, &desc);
|
Chris@10
|
349 }
|
Chris@10
|
350
|
Chris@10
|
351 #else /* HAVE_FMA */
|
Chris@10
|
352
|
Chris@10
|
353 /* Generated by: ../../../genfft/gen_r2cb.native -compact -variables 4 -pipeline-latency 4 -sign 1 -n 32 -name r2cbIII_32 -dft-III -include r2cbIII.h */
|
Chris@10
|
354
|
Chris@10
|
355 /*
|
Chris@10
|
356 * This function contains 174 FP additions, 84 FP multiplications,
|
Chris@10
|
357 * (or, 138 additions, 48 multiplications, 36 fused multiply/add),
|
Chris@10
|
358 * 66 stack variables, 19 constants, and 64 memory accesses
|
Chris@10
|
359 */
|
Chris@10
|
360 #include "r2cbIII.h"
|
Chris@10
|
361
|
Chris@10
|
362 static void r2cbIII_32(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
|
Chris@10
|
363 {
|
Chris@10
|
364 DK(KP1_913880671, +1.913880671464417729871595773960539938965698411);
|
Chris@10
|
365 DK(KP580569354, +0.580569354508924735272384751634790549382952557);
|
Chris@10
|
366 DK(KP942793473, +0.942793473651995297112775251810508755314920638);
|
Chris@10
|
367 DK(KP1_763842528, +1.763842528696710059425513727320776699016885241);
|
Chris@10
|
368 DK(KP1_546020906, +1.546020906725473921621813219516939601942082586);
|
Chris@10
|
369 DK(KP1_268786568, +1.268786568327290996430343226450986741351374190);
|
Chris@10
|
370 DK(KP196034280, +0.196034280659121203988391127777283691722273346);
|
Chris@10
|
371 DK(KP1_990369453, +1.990369453344393772489673906218959843150949737);
|
Chris@10
|
372 DK(KP765366864, +0.765366864730179543456919968060797733522689125);
|
Chris@10
|
373 DK(KP1_847759065, +1.847759065022573512256366378793576573644833252);
|
Chris@10
|
374 DK(KP1_961570560, +1.961570560806460898252364472268478073947867462);
|
Chris@10
|
375 DK(KP390180644, +0.390180644032256535696569736954044481855383236);
|
Chris@10
|
376 DK(KP1_111140466, +1.111140466039204449485661627897065748749874382);
|
Chris@10
|
377 DK(KP1_662939224, +1.662939224605090474157576755235811513477121624);
|
Chris@10
|
378 DK(KP1_414213562, +1.414213562373095048801688724209698078569671875);
|
Chris@10
|
379 DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
|
Chris@10
|
380 DK(KP382683432, +0.382683432365089771728459984030398866761344562);
|
Chris@10
|
381 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
Chris@10
|
382 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
Chris@10
|
383 {
|
Chris@10
|
384 INT i;
|
Chris@10
|
385 for (i = v; i > 0; i = i - 1, R0 = R0 + ovs, R1 = R1 + ovs, Cr = Cr + ivs, Ci = Ci + ivs, MAKE_VOLATILE_STRIDE(128, rs), MAKE_VOLATILE_STRIDE(128, csr), MAKE_VOLATILE_STRIDE(128, csi)) {
|
Chris@10
|
386 E T7, T2i, T2F, Tz, T1k, T1I, T1Z, T1x, Te, T22, T2E, T2j, T1f, T1y, TK;
|
Chris@10
|
387 E T1J, Tm, T2B, TW, T1a, T1C, T1L, T28, T2l, Tt, T2A, T17, T1b, T1F, T1M;
|
Chris@10
|
388 E T2d, T2m;
|
Chris@10
|
389 {
|
Chris@10
|
390 E T3, Tv, T1j, T2h, T6, T1g, Ty, T2g;
|
Chris@10
|
391 {
|
Chris@10
|
392 E T1, T2, T1h, T1i;
|
Chris@10
|
393 T1 = Cr[0];
|
Chris@10
|
394 T2 = Cr[WS(csr, 15)];
|
Chris@10
|
395 T3 = T1 + T2;
|
Chris@10
|
396 Tv = T1 - T2;
|
Chris@10
|
397 T1h = Ci[0];
|
Chris@10
|
398 T1i = Ci[WS(csi, 15)];
|
Chris@10
|
399 T1j = T1h + T1i;
|
Chris@10
|
400 T2h = T1i - T1h;
|
Chris@10
|
401 }
|
Chris@10
|
402 {
|
Chris@10
|
403 E T4, T5, Tw, Tx;
|
Chris@10
|
404 T4 = Cr[WS(csr, 8)];
|
Chris@10
|
405 T5 = Cr[WS(csr, 7)];
|
Chris@10
|
406 T6 = T4 + T5;
|
Chris@10
|
407 T1g = T4 - T5;
|
Chris@10
|
408 Tw = Ci[WS(csi, 8)];
|
Chris@10
|
409 Tx = Ci[WS(csi, 7)];
|
Chris@10
|
410 Ty = Tw + Tx;
|
Chris@10
|
411 T2g = Tw - Tx;
|
Chris@10
|
412 }
|
Chris@10
|
413 T7 = T3 + T6;
|
Chris@10
|
414 T2i = T2g + T2h;
|
Chris@10
|
415 T2F = T2h - T2g;
|
Chris@10
|
416 Tz = Tv - Ty;
|
Chris@10
|
417 T1k = T1g + T1j;
|
Chris@10
|
418 T1I = T1g - T1j;
|
Chris@10
|
419 T1Z = T3 - T6;
|
Chris@10
|
420 T1x = Tv + Ty;
|
Chris@10
|
421 }
|
Chris@10
|
422 {
|
Chris@10
|
423 E Ta, TA, TD, T21, Td, TF, TI, T20;
|
Chris@10
|
424 {
|
Chris@10
|
425 E T8, T9, TB, TC;
|
Chris@10
|
426 T8 = Cr[WS(csr, 4)];
|
Chris@10
|
427 T9 = Cr[WS(csr, 11)];
|
Chris@10
|
428 Ta = T8 + T9;
|
Chris@10
|
429 TA = T8 - T9;
|
Chris@10
|
430 TB = Ci[WS(csi, 4)];
|
Chris@10
|
431 TC = Ci[WS(csi, 11)];
|
Chris@10
|
432 TD = TB + TC;
|
Chris@10
|
433 T21 = TB - TC;
|
Chris@10
|
434 }
|
Chris@10
|
435 {
|
Chris@10
|
436 E Tb, Tc, TG, TH;
|
Chris@10
|
437 Tb = Cr[WS(csr, 3)];
|
Chris@10
|
438 Tc = Cr[WS(csr, 12)];
|
Chris@10
|
439 Td = Tb + Tc;
|
Chris@10
|
440 TF = Tb - Tc;
|
Chris@10
|
441 TG = Ci[WS(csi, 3)];
|
Chris@10
|
442 TH = Ci[WS(csi, 12)];
|
Chris@10
|
443 TI = TG + TH;
|
Chris@10
|
444 T20 = TH - TG;
|
Chris@10
|
445 }
|
Chris@10
|
446 Te = Ta + Td;
|
Chris@10
|
447 T22 = T20 - T21;
|
Chris@10
|
448 T2E = T21 + T20;
|
Chris@10
|
449 T2j = Ta - Td;
|
Chris@10
|
450 {
|
Chris@10
|
451 E T1d, T1e, TE, TJ;
|
Chris@10
|
452 T1d = TA + TD;
|
Chris@10
|
453 T1e = TF + TI;
|
Chris@10
|
454 T1f = KP707106781 * (T1d - T1e);
|
Chris@10
|
455 T1y = KP707106781 * (T1d + T1e);
|
Chris@10
|
456 TE = TA - TD;
|
Chris@10
|
457 TJ = TF - TI;
|
Chris@10
|
458 TK = KP707106781 * (TE + TJ);
|
Chris@10
|
459 T1J = KP707106781 * (TE - TJ);
|
Chris@10
|
460 }
|
Chris@10
|
461 }
|
Chris@10
|
462 {
|
Chris@10
|
463 E Ti, TM, TU, T25, Tl, TR, TP, T26, TQ, TV;
|
Chris@10
|
464 {
|
Chris@10
|
465 E Tg, Th, TS, TT;
|
Chris@10
|
466 Tg = Cr[WS(csr, 2)];
|
Chris@10
|
467 Th = Cr[WS(csr, 13)];
|
Chris@10
|
468 Ti = Tg + Th;
|
Chris@10
|
469 TM = Tg - Th;
|
Chris@10
|
470 TS = Ci[WS(csi, 2)];
|
Chris@10
|
471 TT = Ci[WS(csi, 13)];
|
Chris@10
|
472 TU = TS + TT;
|
Chris@10
|
473 T25 = TS - TT;
|
Chris@10
|
474 }
|
Chris@10
|
475 {
|
Chris@10
|
476 E Tj, Tk, TN, TO;
|
Chris@10
|
477 Tj = Cr[WS(csr, 10)];
|
Chris@10
|
478 Tk = Cr[WS(csr, 5)];
|
Chris@10
|
479 Tl = Tj + Tk;
|
Chris@10
|
480 TR = Tj - Tk;
|
Chris@10
|
481 TN = Ci[WS(csi, 10)];
|
Chris@10
|
482 TO = Ci[WS(csi, 5)];
|
Chris@10
|
483 TP = TN + TO;
|
Chris@10
|
484 T26 = TN - TO;
|
Chris@10
|
485 }
|
Chris@10
|
486 Tm = Ti + Tl;
|
Chris@10
|
487 T2B = T26 + T25;
|
Chris@10
|
488 TQ = TM - TP;
|
Chris@10
|
489 TV = TR + TU;
|
Chris@10
|
490 TW = FNMS(KP382683432, TV, KP923879532 * TQ);
|
Chris@10
|
491 T1a = FMA(KP382683432, TQ, KP923879532 * TV);
|
Chris@10
|
492 {
|
Chris@10
|
493 E T1A, T1B, T24, T27;
|
Chris@10
|
494 T1A = TM + TP;
|
Chris@10
|
495 T1B = TU - TR;
|
Chris@10
|
496 T1C = FNMS(KP923879532, T1B, KP382683432 * T1A);
|
Chris@10
|
497 T1L = FMA(KP923879532, T1A, KP382683432 * T1B);
|
Chris@10
|
498 T24 = Ti - Tl;
|
Chris@10
|
499 T27 = T25 - T26;
|
Chris@10
|
500 T28 = T24 - T27;
|
Chris@10
|
501 T2l = T24 + T27;
|
Chris@10
|
502 }
|
Chris@10
|
503 }
|
Chris@10
|
504 {
|
Chris@10
|
505 E Tp, TX, T15, T2a, Ts, T12, T10, T2b, T11, T16;
|
Chris@10
|
506 {
|
Chris@10
|
507 E Tn, To, T13, T14;
|
Chris@10
|
508 Tn = Cr[WS(csr, 1)];
|
Chris@10
|
509 To = Cr[WS(csr, 14)];
|
Chris@10
|
510 Tp = Tn + To;
|
Chris@10
|
511 TX = Tn - To;
|
Chris@10
|
512 T13 = Ci[WS(csi, 1)];
|
Chris@10
|
513 T14 = Ci[WS(csi, 14)];
|
Chris@10
|
514 T15 = T13 + T14;
|
Chris@10
|
515 T2a = T14 - T13;
|
Chris@10
|
516 }
|
Chris@10
|
517 {
|
Chris@10
|
518 E Tq, Tr, TY, TZ;
|
Chris@10
|
519 Tq = Cr[WS(csr, 6)];
|
Chris@10
|
520 Tr = Cr[WS(csr, 9)];
|
Chris@10
|
521 Ts = Tq + Tr;
|
Chris@10
|
522 T12 = Tq - Tr;
|
Chris@10
|
523 TY = Ci[WS(csi, 6)];
|
Chris@10
|
524 TZ = Ci[WS(csi, 9)];
|
Chris@10
|
525 T10 = TY + TZ;
|
Chris@10
|
526 T2b = TY - TZ;
|
Chris@10
|
527 }
|
Chris@10
|
528 Tt = Tp + Ts;
|
Chris@10
|
529 T2A = T2b + T2a;
|
Chris@10
|
530 T11 = TX - T10;
|
Chris@10
|
531 T16 = T12 - T15;
|
Chris@10
|
532 T17 = FMA(KP923879532, T11, KP382683432 * T16);
|
Chris@10
|
533 T1b = FNMS(KP382683432, T11, KP923879532 * T16);
|
Chris@10
|
534 {
|
Chris@10
|
535 E T1D, T1E, T29, T2c;
|
Chris@10
|
536 T1D = TX + T10;
|
Chris@10
|
537 T1E = T12 + T15;
|
Chris@10
|
538 T1F = FNMS(KP923879532, T1E, KP382683432 * T1D);
|
Chris@10
|
539 T1M = FMA(KP923879532, T1D, KP382683432 * T1E);
|
Chris@10
|
540 T29 = Tp - Ts;
|
Chris@10
|
541 T2c = T2a - T2b;
|
Chris@10
|
542 T2d = T29 + T2c;
|
Chris@10
|
543 T2m = T2c - T29;
|
Chris@10
|
544 }
|
Chris@10
|
545 }
|
Chris@10
|
546 {
|
Chris@10
|
547 E Tf, Tu, T2L, T2M, T2N, T2O;
|
Chris@10
|
548 Tf = T7 + Te;
|
Chris@10
|
549 Tu = Tm + Tt;
|
Chris@10
|
550 T2L = Tf - Tu;
|
Chris@10
|
551 T2M = T2B + T2A;
|
Chris@10
|
552 T2N = T2F - T2E;
|
Chris@10
|
553 T2O = T2M + T2N;
|
Chris@10
|
554 R0[0] = KP2_000000000 * (Tf + Tu);
|
Chris@10
|
555 R0[WS(rs, 8)] = KP2_000000000 * (T2N - T2M);
|
Chris@10
|
556 R0[WS(rs, 4)] = KP1_414213562 * (T2L + T2O);
|
Chris@10
|
557 R0[WS(rs, 12)] = KP1_414213562 * (T2O - T2L);
|
Chris@10
|
558 }
|
Chris@10
|
559 {
|
Chris@10
|
560 E T2t, T2x, T2w, T2y;
|
Chris@10
|
561 {
|
Chris@10
|
562 E T2r, T2s, T2u, T2v;
|
Chris@10
|
563 T2r = T1Z - T22;
|
Chris@10
|
564 T2s = KP707106781 * (T2m - T2l);
|
Chris@10
|
565 T2t = T2r + T2s;
|
Chris@10
|
566 T2x = T2r - T2s;
|
Chris@10
|
567 T2u = T2j + T2i;
|
Chris@10
|
568 T2v = KP707106781 * (T28 - T2d);
|
Chris@10
|
569 T2w = T2u - T2v;
|
Chris@10
|
570 T2y = T2v + T2u;
|
Chris@10
|
571 }
|
Chris@10
|
572 R0[WS(rs, 3)] = FMA(KP1_662939224, T2t, KP1_111140466 * T2w);
|
Chris@10
|
573 R0[WS(rs, 15)] = FNMS(KP1_961570560, T2x, KP390180644 * T2y);
|
Chris@10
|
574 R0[WS(rs, 11)] = FNMS(KP1_111140466, T2t, KP1_662939224 * T2w);
|
Chris@10
|
575 R0[WS(rs, 7)] = FMA(KP390180644, T2x, KP1_961570560 * T2y);
|
Chris@10
|
576 }
|
Chris@10
|
577 {
|
Chris@10
|
578 E T2D, T2J, T2I, T2K;
|
Chris@10
|
579 {
|
Chris@10
|
580 E T2z, T2C, T2G, T2H;
|
Chris@10
|
581 T2z = T7 - Te;
|
Chris@10
|
582 T2C = T2A - T2B;
|
Chris@10
|
583 T2D = T2z + T2C;
|
Chris@10
|
584 T2J = T2z - T2C;
|
Chris@10
|
585 T2G = T2E + T2F;
|
Chris@10
|
586 T2H = Tm - Tt;
|
Chris@10
|
587 T2I = T2G - T2H;
|
Chris@10
|
588 T2K = T2H + T2G;
|
Chris@10
|
589 }
|
Chris@10
|
590 R0[WS(rs, 2)] = FMA(KP1_847759065, T2D, KP765366864 * T2I);
|
Chris@10
|
591 R0[WS(rs, 14)] = FNMS(KP1_847759065, T2J, KP765366864 * T2K);
|
Chris@10
|
592 R0[WS(rs, 10)] = FNMS(KP765366864, T2D, KP1_847759065 * T2I);
|
Chris@10
|
593 R0[WS(rs, 6)] = FMA(KP765366864, T2J, KP1_847759065 * T2K);
|
Chris@10
|
594 }
|
Chris@10
|
595 {
|
Chris@10
|
596 E T19, T1n, T1m, T1o;
|
Chris@10
|
597 {
|
Chris@10
|
598 E TL, T18, T1c, T1l;
|
Chris@10
|
599 TL = Tz + TK;
|
Chris@10
|
600 T18 = TW + T17;
|
Chris@10
|
601 T19 = TL + T18;
|
Chris@10
|
602 T1n = TL - T18;
|
Chris@10
|
603 T1c = T1a + T1b;
|
Chris@10
|
604 T1l = T1f + T1k;
|
Chris@10
|
605 T1m = T1c + T1l;
|
Chris@10
|
606 T1o = T1c - T1l;
|
Chris@10
|
607 }
|
Chris@10
|
608 R1[0] = FNMS(KP196034280, T1m, KP1_990369453 * T19);
|
Chris@10
|
609 R1[WS(rs, 12)] = FNMS(KP1_546020906, T1n, KP1_268786568 * T1o);
|
Chris@10
|
610 R1[WS(rs, 8)] = -(FMA(KP196034280, T19, KP1_990369453 * T1m));
|
Chris@10
|
611 R1[WS(rs, 4)] = FMA(KP1_268786568, T1n, KP1_546020906 * T1o);
|
Chris@10
|
612 }
|
Chris@10
|
613 {
|
Chris@10
|
614 E T1r, T1v, T1u, T1w;
|
Chris@10
|
615 {
|
Chris@10
|
616 E T1p, T1q, T1s, T1t;
|
Chris@10
|
617 T1p = Tz - TK;
|
Chris@10
|
618 T1q = T1b - T1a;
|
Chris@10
|
619 T1r = T1p + T1q;
|
Chris@10
|
620 T1v = T1p - T1q;
|
Chris@10
|
621 T1s = T1f - T1k;
|
Chris@10
|
622 T1t = TW - T17;
|
Chris@10
|
623 T1u = T1s - T1t;
|
Chris@10
|
624 T1w = T1t + T1s;
|
Chris@10
|
625 }
|
Chris@10
|
626 R1[WS(rs, 2)] = FMA(KP1_763842528, T1r, KP942793473 * T1u);
|
Chris@10
|
627 R1[WS(rs, 14)] = FNMS(KP1_913880671, T1v, KP580569354 * T1w);
|
Chris@10
|
628 R1[WS(rs, 10)] = FNMS(KP942793473, T1r, KP1_763842528 * T1u);
|
Chris@10
|
629 R1[WS(rs, 6)] = FMA(KP580569354, T1v, KP1_913880671 * T1w);
|
Chris@10
|
630 }
|
Chris@10
|
631 {
|
Chris@10
|
632 E T1T, T1X, T1W, T1Y;
|
Chris@10
|
633 {
|
Chris@10
|
634 E T1R, T1S, T1U, T1V;
|
Chris@10
|
635 T1R = T1x + T1y;
|
Chris@10
|
636 T1S = T1L + T1M;
|
Chris@10
|
637 T1T = T1R - T1S;
|
Chris@10
|
638 T1X = T1R + T1S;
|
Chris@10
|
639 T1U = T1J + T1I;
|
Chris@10
|
640 T1V = T1C - T1F;
|
Chris@10
|
641 T1W = T1U - T1V;
|
Chris@10
|
642 T1Y = T1V + T1U;
|
Chris@10
|
643 }
|
Chris@10
|
644 R1[WS(rs, 3)] = FMA(KP1_546020906, T1T, KP1_268786568 * T1W);
|
Chris@10
|
645 R1[WS(rs, 15)] = FNMS(KP1_990369453, T1X, KP196034280 * T1Y);
|
Chris@10
|
646 R1[WS(rs, 11)] = FNMS(KP1_268786568, T1T, KP1_546020906 * T1W);
|
Chris@10
|
647 R1[WS(rs, 7)] = FMA(KP196034280, T1X, KP1_990369453 * T1Y);
|
Chris@10
|
648 }
|
Chris@10
|
649 {
|
Chris@10
|
650 E T2f, T2p, T2o, T2q;
|
Chris@10
|
651 {
|
Chris@10
|
652 E T23, T2e, T2k, T2n;
|
Chris@10
|
653 T23 = T1Z + T22;
|
Chris@10
|
654 T2e = KP707106781 * (T28 + T2d);
|
Chris@10
|
655 T2f = T23 + T2e;
|
Chris@10
|
656 T2p = T23 - T2e;
|
Chris@10
|
657 T2k = T2i - T2j;
|
Chris@10
|
658 T2n = KP707106781 * (T2l + T2m);
|
Chris@10
|
659 T2o = T2k - T2n;
|
Chris@10
|
660 T2q = T2n + T2k;
|
Chris@10
|
661 }
|
Chris@10
|
662 R0[WS(rs, 1)] = FMA(KP1_961570560, T2f, KP390180644 * T2o);
|
Chris@10
|
663 R0[WS(rs, 13)] = FNMS(KP1_662939224, T2p, KP1_111140466 * T2q);
|
Chris@10
|
664 R0[WS(rs, 9)] = FNMS(KP390180644, T2f, KP1_961570560 * T2o);
|
Chris@10
|
665 R0[WS(rs, 5)] = FMA(KP1_111140466, T2p, KP1_662939224 * T2q);
|
Chris@10
|
666 }
|
Chris@10
|
667 {
|
Chris@10
|
668 E T1H, T1P, T1O, T1Q;
|
Chris@10
|
669 {
|
Chris@10
|
670 E T1z, T1G, T1K, T1N;
|
Chris@10
|
671 T1z = T1x - T1y;
|
Chris@10
|
672 T1G = T1C + T1F;
|
Chris@10
|
673 T1H = T1z + T1G;
|
Chris@10
|
674 T1P = T1z - T1G;
|
Chris@10
|
675 T1K = T1I - T1J;
|
Chris@10
|
676 T1N = T1L - T1M;
|
Chris@10
|
677 T1O = T1K - T1N;
|
Chris@10
|
678 T1Q = T1N + T1K;
|
Chris@10
|
679 }
|
Chris@10
|
680 R1[WS(rs, 1)] = FMA(KP1_913880671, T1H, KP580569354 * T1O);
|
Chris@10
|
681 R1[WS(rs, 13)] = FNMS(KP1_763842528, T1P, KP942793473 * T1Q);
|
Chris@10
|
682 R1[WS(rs, 9)] = FNMS(KP580569354, T1H, KP1_913880671 * T1O);
|
Chris@10
|
683 R1[WS(rs, 5)] = FMA(KP942793473, T1P, KP1_763842528 * T1Q);
|
Chris@10
|
684 }
|
Chris@10
|
685 }
|
Chris@10
|
686 }
|
Chris@10
|
687 }
|
Chris@10
|
688
|
Chris@10
|
689 static const kr2c_desc desc = { 32, "r2cbIII_32", {138, 48, 36, 0}, &GENUS };
|
Chris@10
|
690
|
Chris@10
|
691 void X(codelet_r2cbIII_32) (planner *p) {
|
Chris@10
|
692 X(kr2c_register) (p, r2cbIII_32, &desc);
|
Chris@10
|
693 }
|
Chris@10
|
694
|
Chris@10
|
695 #endif /* HAVE_FMA */
|