cannam@167
|
1 /*
|
cannam@167
|
2 * Copyright (c) 2003, 2007-14 Matteo Frigo
|
cannam@167
|
3 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
|
cannam@167
|
4 *
|
cannam@167
|
5 * This program is free software; you can redistribute it and/or modify
|
cannam@167
|
6 * it under the terms of the GNU General Public License as published by
|
cannam@167
|
7 * the Free Software Foundation; either version 2 of the License, or
|
cannam@167
|
8 * (at your option) any later version.
|
cannam@167
|
9 *
|
cannam@167
|
10 * This program is distributed in the hope that it will be useful,
|
cannam@167
|
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
cannam@167
|
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
cannam@167
|
13 * GNU General Public License for more details.
|
cannam@167
|
14 *
|
cannam@167
|
15 * You should have received a copy of the GNU General Public License
|
cannam@167
|
16 * along with this program; if not, write to the Free Software
|
cannam@167
|
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
cannam@167
|
18 *
|
cannam@167
|
19 */
|
cannam@167
|
20
|
cannam@167
|
21 /* This file was automatically generated --- DO NOT EDIT */
|
cannam@167
|
22 /* Generated on Thu May 24 08:07:45 EDT 2018 */
|
cannam@167
|
23
|
cannam@167
|
24 #include "rdft/codelet-rdft.h"
|
cannam@167
|
25
|
cannam@167
|
26 #if defined(ARCH_PREFERS_FMA) || defined(ISA_EXTENSION_PREFERS_FMA)
|
cannam@167
|
27
|
cannam@167
|
28 /* Generated by: ../../../genfft/gen_r2cb.native -fma -compact -variables 4 -pipeline-latency 4 -sign 1 -n 32 -name r2cbIII_32 -dft-III -include rdft/scalar/r2cbIII.h */
|
cannam@167
|
29
|
cannam@167
|
30 /*
|
cannam@167
|
31 * This function contains 174 FP additions, 100 FP multiplications,
|
cannam@167
|
32 * (or, 106 additions, 32 multiplications, 68 fused multiply/add),
|
cannam@167
|
33 * 65 stack variables, 18 constants, and 64 memory accesses
|
cannam@167
|
34 */
|
cannam@167
|
35 #include "rdft/scalar/r2cbIII.h"
|
cannam@167
|
36
|
cannam@167
|
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)
|
cannam@167
|
38 {
|
cannam@167
|
39 DK(KP303346683, +0.303346683607342391675883946941299872384187453);
|
cannam@167
|
40 DK(KP1_913880671, +1.913880671464417729871595773960539938965698411);
|
cannam@167
|
41 DK(KP534511135, +0.534511135950791641089685961295362908582039528);
|
cannam@167
|
42 DK(KP1_763842528, +1.763842528696710059425513727320776699016885241);
|
cannam@167
|
43 DK(KP820678790, +0.820678790828660330972281985331011598767386482);
|
cannam@167
|
44 DK(KP1_546020906, +1.546020906725473921621813219516939601942082586);
|
cannam@167
|
45 DK(KP098491403, +0.098491403357164253077197521291327432293052451);
|
cannam@167
|
46 DK(KP1_990369453, +1.990369453344393772489673906218959843150949737);
|
cannam@167
|
47 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
cannam@167
|
48 DK(KP1_847759065, +1.847759065022573512256366378793576573644833252);
|
cannam@167
|
49 DK(KP198912367, +0.198912367379658006911597622644676228597850501);
|
cannam@167
|
50 DK(KP1_961570560, +1.961570560806460898252364472268478073947867462);
|
cannam@167
|
51 DK(KP668178637, +0.668178637919298919997757686523080761552472251);
|
cannam@167
|
52 DK(KP1_662939224, +1.662939224605090474157576755235811513477121624);
|
cannam@167
|
53 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
cannam@167
|
54 DK(KP1_414213562, +1.414213562373095048801688724209698078569671875);
|
cannam@167
|
55 DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
|
cannam@167
|
56 DK(KP414213562, +0.414213562373095048801688724209698078569671875);
|
cannam@167
|
57 {
|
cannam@167
|
58 INT i;
|
cannam@167
|
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)) {
|
cannam@167
|
60 E T7, T2i, T2E, Tz, T1e, T1I, T1Z, T1x, Te, T22, T2F, T2j, T1h, T1y, TK;
|
cannam@167
|
61 E T1J, Tm, T2B, TW, T1k, T1C, T1M, T28, T2m, Tt, T2A, T17, T1j, T1F, T1L;
|
cannam@167
|
62 E T2d, T2l;
|
cannam@167
|
63 {
|
cannam@167
|
64 E T3, Tv, T1d, T2g, T6, T1a, Ty, T2h;
|
cannam@167
|
65 {
|
cannam@167
|
66 E T1, T2, T1b, T1c;
|
cannam@167
|
67 T1 = Cr[0];
|
cannam@167
|
68 T2 = Cr[WS(csr, 15)];
|
cannam@167
|
69 T3 = T1 + T2;
|
cannam@167
|
70 Tv = T1 - T2;
|
cannam@167
|
71 T1b = Ci[0];
|
cannam@167
|
72 T1c = Ci[WS(csi, 15)];
|
cannam@167
|
73 T1d = T1b + T1c;
|
cannam@167
|
74 T2g = T1c - T1b;
|
cannam@167
|
75 }
|
cannam@167
|
76 {
|
cannam@167
|
77 E T4, T5, Tw, Tx;
|
cannam@167
|
78 T4 = Cr[WS(csr, 8)];
|
cannam@167
|
79 T5 = Cr[WS(csr, 7)];
|
cannam@167
|
80 T6 = T4 + T5;
|
cannam@167
|
81 T1a = T4 - T5;
|
cannam@167
|
82 Tw = Ci[WS(csi, 8)];
|
cannam@167
|
83 Tx = Ci[WS(csi, 7)];
|
cannam@167
|
84 Ty = Tw + Tx;
|
cannam@167
|
85 T2h = Tx - Tw;
|
cannam@167
|
86 }
|
cannam@167
|
87 T7 = T3 + T6;
|
cannam@167
|
88 T2i = T2g - T2h;
|
cannam@167
|
89 T2E = T2h + T2g;
|
cannam@167
|
90 Tz = Tv - Ty;
|
cannam@167
|
91 T1e = T1a + T1d;
|
cannam@167
|
92 T1I = T1a - T1d;
|
cannam@167
|
93 T1Z = T3 - T6;
|
cannam@167
|
94 T1x = Tv + Ty;
|
cannam@167
|
95 }
|
cannam@167
|
96 {
|
cannam@167
|
97 E Ta, TA, TD, T20, Td, TF, TI, T21;
|
cannam@167
|
98 {
|
cannam@167
|
99 E T8, T9, TB, TC;
|
cannam@167
|
100 T8 = Cr[WS(csr, 4)];
|
cannam@167
|
101 T9 = Cr[WS(csr, 11)];
|
cannam@167
|
102 Ta = T8 + T9;
|
cannam@167
|
103 TA = T8 - T9;
|
cannam@167
|
104 TB = Ci[WS(csi, 4)];
|
cannam@167
|
105 TC = Ci[WS(csi, 11)];
|
cannam@167
|
106 TD = TB + TC;
|
cannam@167
|
107 T20 = TC - TB;
|
cannam@167
|
108 }
|
cannam@167
|
109 {
|
cannam@167
|
110 E Tb, Tc, TG, TH;
|
cannam@167
|
111 Tb = Cr[WS(csr, 3)];
|
cannam@167
|
112 Tc = Cr[WS(csr, 12)];
|
cannam@167
|
113 Td = Tb + Tc;
|
cannam@167
|
114 TF = Tb - Tc;
|
cannam@167
|
115 TG = Ci[WS(csi, 3)];
|
cannam@167
|
116 TH = Ci[WS(csi, 12)];
|
cannam@167
|
117 TI = TG + TH;
|
cannam@167
|
118 T21 = TG - TH;
|
cannam@167
|
119 }
|
cannam@167
|
120 Te = Ta + Td;
|
cannam@167
|
121 T22 = T20 - T21;
|
cannam@167
|
122 T2F = T20 + T21;
|
cannam@167
|
123 T2j = Ta - Td;
|
cannam@167
|
124 {
|
cannam@167
|
125 E T1f, T1g, TE, TJ;
|
cannam@167
|
126 T1f = TF + TI;
|
cannam@167
|
127 T1g = TA + TD;
|
cannam@167
|
128 T1h = T1f - T1g;
|
cannam@167
|
129 T1y = T1g + T1f;
|
cannam@167
|
130 TE = TA - TD;
|
cannam@167
|
131 TJ = TF - TI;
|
cannam@167
|
132 TK = TE + TJ;
|
cannam@167
|
133 T1J = TE - TJ;
|
cannam@167
|
134 }
|
cannam@167
|
135 }
|
cannam@167
|
136 {
|
cannam@167
|
137 E Ti, TM, TU, T25, Tl, TR, TP, T26, TQ, TV;
|
cannam@167
|
138 {
|
cannam@167
|
139 E Tg, Th, TS, TT;
|
cannam@167
|
140 Tg = Cr[WS(csr, 2)];
|
cannam@167
|
141 Th = Cr[WS(csr, 13)];
|
cannam@167
|
142 Ti = Tg + Th;
|
cannam@167
|
143 TM = Tg - Th;
|
cannam@167
|
144 TS = Ci[WS(csi, 2)];
|
cannam@167
|
145 TT = Ci[WS(csi, 13)];
|
cannam@167
|
146 TU = TS + TT;
|
cannam@167
|
147 T25 = TS - TT;
|
cannam@167
|
148 }
|
cannam@167
|
149 {
|
cannam@167
|
150 E Tj, Tk, TN, TO;
|
cannam@167
|
151 Tj = Cr[WS(csr, 10)];
|
cannam@167
|
152 Tk = Cr[WS(csr, 5)];
|
cannam@167
|
153 Tl = Tj + Tk;
|
cannam@167
|
154 TR = Tj - Tk;
|
cannam@167
|
155 TN = Ci[WS(csi, 10)];
|
cannam@167
|
156 TO = Ci[WS(csi, 5)];
|
cannam@167
|
157 TP = TN + TO;
|
cannam@167
|
158 T26 = TN - TO;
|
cannam@167
|
159 }
|
cannam@167
|
160 Tm = Ti + Tl;
|
cannam@167
|
161 T2B = T26 + T25;
|
cannam@167
|
162 TQ = TM - TP;
|
cannam@167
|
163 TV = TR + TU;
|
cannam@167
|
164 TW = FNMS(KP414213562, TV, TQ);
|
cannam@167
|
165 T1k = FMA(KP414213562, TQ, TV);
|
cannam@167
|
166 {
|
cannam@167
|
167 E T1A, T1B, T24, T27;
|
cannam@167
|
168 T1A = TR - TU;
|
cannam@167
|
169 T1B = TM + TP;
|
cannam@167
|
170 T1C = FMA(KP414213562, T1B, T1A);
|
cannam@167
|
171 T1M = FNMS(KP414213562, T1A, T1B);
|
cannam@167
|
172 T24 = Ti - Tl;
|
cannam@167
|
173 T27 = T25 - T26;
|
cannam@167
|
174 T28 = T24 - T27;
|
cannam@167
|
175 T2m = T24 + T27;
|
cannam@167
|
176 }
|
cannam@167
|
177 }
|
cannam@167
|
178 {
|
cannam@167
|
179 E Tp, TX, T14, T2a, Ts, T15, T10, T2b, T11, T16;
|
cannam@167
|
180 {
|
cannam@167
|
181 E Tn, To, T12, T13;
|
cannam@167
|
182 Tn = Cr[WS(csr, 1)];
|
cannam@167
|
183 To = Cr[WS(csr, 14)];
|
cannam@167
|
184 Tp = Tn + To;
|
cannam@167
|
185 TX = Tn - To;
|
cannam@167
|
186 T12 = Ci[WS(csi, 1)];
|
cannam@167
|
187 T13 = Ci[WS(csi, 14)];
|
cannam@167
|
188 T14 = T12 + T13;
|
cannam@167
|
189 T2a = T13 - T12;
|
cannam@167
|
190 }
|
cannam@167
|
191 {
|
cannam@167
|
192 E Tq, Tr, TY, TZ;
|
cannam@167
|
193 Tq = Cr[WS(csr, 6)];
|
cannam@167
|
194 Tr = Cr[WS(csr, 9)];
|
cannam@167
|
195 Ts = Tq + Tr;
|
cannam@167
|
196 T15 = Tq - Tr;
|
cannam@167
|
197 TY = Ci[WS(csi, 6)];
|
cannam@167
|
198 TZ = Ci[WS(csi, 9)];
|
cannam@167
|
199 T10 = TY + TZ;
|
cannam@167
|
200 T2b = TY - TZ;
|
cannam@167
|
201 }
|
cannam@167
|
202 Tt = Tp + Ts;
|
cannam@167
|
203 T2A = T2b + T2a;
|
cannam@167
|
204 T11 = TX - T10;
|
cannam@167
|
205 T16 = T14 - T15;
|
cannam@167
|
206 T17 = FNMS(KP414213562, T16, T11);
|
cannam@167
|
207 T1j = FMA(KP414213562, T11, T16);
|
cannam@167
|
208 {
|
cannam@167
|
209 E T1D, T1E, T29, T2c;
|
cannam@167
|
210 T1D = T15 + T14;
|
cannam@167
|
211 T1E = TX + T10;
|
cannam@167
|
212 T1F = FNMS(KP414213562, T1E, T1D);
|
cannam@167
|
213 T1L = FMA(KP414213562, T1D, T1E);
|
cannam@167
|
214 T29 = Tp - Ts;
|
cannam@167
|
215 T2c = T2a - T2b;
|
cannam@167
|
216 T2d = T29 + T2c;
|
cannam@167
|
217 T2l = T29 - T2c;
|
cannam@167
|
218 }
|
cannam@167
|
219 }
|
cannam@167
|
220 {
|
cannam@167
|
221 E Tf, Tu, T2L, T2M, T2N, T2O;
|
cannam@167
|
222 Tf = T7 + Te;
|
cannam@167
|
223 Tu = Tm + Tt;
|
cannam@167
|
224 T2L = Tf - Tu;
|
cannam@167
|
225 T2M = T2B + T2A;
|
cannam@167
|
226 T2N = T2F + T2E;
|
cannam@167
|
227 T2O = T2M + T2N;
|
cannam@167
|
228 R0[0] = KP2_000000000 * (Tf + Tu);
|
cannam@167
|
229 R0[WS(rs, 8)] = KP2_000000000 * (T2N - T2M);
|
cannam@167
|
230 R0[WS(rs, 4)] = KP1_414213562 * (T2L + T2O);
|
cannam@167
|
231 R0[WS(rs, 12)] = KP1_414213562 * (T2O - T2L);
|
cannam@167
|
232 }
|
cannam@167
|
233 {
|
cannam@167
|
234 E T2t, T2y, T2w, T2x;
|
cannam@167
|
235 {
|
cannam@167
|
236 E T2r, T2s, T2u, T2v;
|
cannam@167
|
237 T2r = T1Z - T22;
|
cannam@167
|
238 T2s = T2m + T2l;
|
cannam@167
|
239 T2t = FNMS(KP707106781, T2s, T2r);
|
cannam@167
|
240 T2y = FMA(KP707106781, T2s, T2r);
|
cannam@167
|
241 T2u = T2j + T2i;
|
cannam@167
|
242 T2v = T28 - T2d;
|
cannam@167
|
243 T2w = FNMS(KP707106781, T2v, T2u);
|
cannam@167
|
244 T2x = FMA(KP707106781, T2v, T2u);
|
cannam@167
|
245 }
|
cannam@167
|
246 R0[WS(rs, 3)] = KP1_662939224 * (FMA(KP668178637, T2w, T2t));
|
cannam@167
|
247 R0[WS(rs, 15)] = -(KP1_961570560 * (FNMS(KP198912367, T2x, T2y)));
|
cannam@167
|
248 R0[WS(rs, 11)] = KP1_662939224 * (FNMS(KP668178637, T2t, T2w));
|
cannam@167
|
249 R0[WS(rs, 7)] = KP1_961570560 * (FMA(KP198912367, T2y, T2x));
|
cannam@167
|
250 }
|
cannam@167
|
251 {
|
cannam@167
|
252 E T2D, T2K, T2I, T2J;
|
cannam@167
|
253 {
|
cannam@167
|
254 E T2z, T2C, T2G, T2H;
|
cannam@167
|
255 T2z = T7 - Te;
|
cannam@167
|
256 T2C = T2A - T2B;
|
cannam@167
|
257 T2D = T2z + T2C;
|
cannam@167
|
258 T2K = T2z - T2C;
|
cannam@167
|
259 T2G = T2E - T2F;
|
cannam@167
|
260 T2H = Tm - Tt;
|
cannam@167
|
261 T2I = T2G - T2H;
|
cannam@167
|
262 T2J = T2H + T2G;
|
cannam@167
|
263 }
|
cannam@167
|
264 R0[WS(rs, 2)] = KP1_847759065 * (FMA(KP414213562, T2I, T2D));
|
cannam@167
|
265 R0[WS(rs, 14)] = -(KP1_847759065 * (FNMS(KP414213562, T2J, T2K)));
|
cannam@167
|
266 R0[WS(rs, 10)] = KP1_847759065 * (FNMS(KP414213562, T2D, T2I));
|
cannam@167
|
267 R0[WS(rs, 6)] = KP1_847759065 * (FMA(KP414213562, T2K, T2J));
|
cannam@167
|
268 }
|
cannam@167
|
269 {
|
cannam@167
|
270 E T19, T1o, T1m, T1n;
|
cannam@167
|
271 {
|
cannam@167
|
272 E TL, T18, T1i, T1l;
|
cannam@167
|
273 TL = FMA(KP707106781, TK, Tz);
|
cannam@167
|
274 T18 = TW + T17;
|
cannam@167
|
275 T19 = FMA(KP923879532, T18, TL);
|
cannam@167
|
276 T1o = FNMS(KP923879532, T18, TL);
|
cannam@167
|
277 T1i = FNMS(KP707106781, T1h, T1e);
|
cannam@167
|
278 T1l = T1j - T1k;
|
cannam@167
|
279 T1m = FNMS(KP923879532, T1l, T1i);
|
cannam@167
|
280 T1n = FMA(KP923879532, T1l, T1i);
|
cannam@167
|
281 }
|
cannam@167
|
282 R1[0] = KP1_990369453 * (FNMS(KP098491403, T1m, T19));
|
cannam@167
|
283 R1[WS(rs, 12)] = -(KP1_546020906 * (FMA(KP820678790, T1n, T1o)));
|
cannam@167
|
284 R1[WS(rs, 8)] = -(KP1_990369453 * (FMA(KP098491403, T19, T1m)));
|
cannam@167
|
285 R1[WS(rs, 4)] = -(KP1_546020906 * (FNMS(KP820678790, T1o, T1n)));
|
cannam@167
|
286 }
|
cannam@167
|
287 {
|
cannam@167
|
288 E T1r, T1w, T1u, T1v;
|
cannam@167
|
289 {
|
cannam@167
|
290 E T1p, T1q, T1s, T1t;
|
cannam@167
|
291 T1p = FNMS(KP707106781, TK, Tz);
|
cannam@167
|
292 T1q = T1k + T1j;
|
cannam@167
|
293 T1r = FNMS(KP923879532, T1q, T1p);
|
cannam@167
|
294 T1w = FMA(KP923879532, T1q, T1p);
|
cannam@167
|
295 T1s = FMA(KP707106781, T1h, T1e);
|
cannam@167
|
296 T1t = TW - T17;
|
cannam@167
|
297 T1u = FMA(KP923879532, T1t, T1s);
|
cannam@167
|
298 T1v = FNMS(KP923879532, T1t, T1s);
|
cannam@167
|
299 }
|
cannam@167
|
300 R1[WS(rs, 2)] = KP1_763842528 * (FNMS(KP534511135, T1u, T1r));
|
cannam@167
|
301 R1[WS(rs, 14)] = -(KP1_913880671 * (FMA(KP303346683, T1v, T1w)));
|
cannam@167
|
302 R1[WS(rs, 10)] = -(KP1_763842528 * (FMA(KP534511135, T1r, T1u)));
|
cannam@167
|
303 R1[WS(rs, 6)] = -(KP1_913880671 * (FNMS(KP303346683, T1w, T1v)));
|
cannam@167
|
304 }
|
cannam@167
|
305 {
|
cannam@167
|
306 E T1T, T1Y, T1W, T1X;
|
cannam@167
|
307 {
|
cannam@167
|
308 E T1R, T1S, T1U, T1V;
|
cannam@167
|
309 T1R = FMA(KP707106781, T1y, T1x);
|
cannam@167
|
310 T1S = T1M + T1L;
|
cannam@167
|
311 T1T = FNMS(KP923879532, T1S, T1R);
|
cannam@167
|
312 T1Y = FMA(KP923879532, T1S, T1R);
|
cannam@167
|
313 T1U = FMA(KP707106781, T1J, T1I);
|
cannam@167
|
314 T1V = T1C + T1F;
|
cannam@167
|
315 T1W = FNMS(KP923879532, T1V, T1U);
|
cannam@167
|
316 T1X = FMA(KP923879532, T1V, T1U);
|
cannam@167
|
317 }
|
cannam@167
|
318 R1[WS(rs, 3)] = KP1_546020906 * (FMA(KP820678790, T1W, T1T));
|
cannam@167
|
319 R1[WS(rs, 15)] = -(KP1_990369453 * (FNMS(KP098491403, T1X, T1Y)));
|
cannam@167
|
320 R1[WS(rs, 11)] = KP1_546020906 * (FNMS(KP820678790, T1T, T1W));
|
cannam@167
|
321 R1[WS(rs, 7)] = KP1_990369453 * (FMA(KP098491403, T1Y, T1X));
|
cannam@167
|
322 }
|
cannam@167
|
323 {
|
cannam@167
|
324 E T2f, T2q, T2o, T2p;
|
cannam@167
|
325 {
|
cannam@167
|
326 E T23, T2e, T2k, T2n;
|
cannam@167
|
327 T23 = T1Z + T22;
|
cannam@167
|
328 T2e = T28 + T2d;
|
cannam@167
|
329 T2f = FMA(KP707106781, T2e, T23);
|
cannam@167
|
330 T2q = FNMS(KP707106781, T2e, T23);
|
cannam@167
|
331 T2k = T2i - T2j;
|
cannam@167
|
332 T2n = T2l - T2m;
|
cannam@167
|
333 T2o = FMA(KP707106781, T2n, T2k);
|
cannam@167
|
334 T2p = FNMS(KP707106781, T2n, T2k);
|
cannam@167
|
335 }
|
cannam@167
|
336 R0[WS(rs, 1)] = KP1_961570560 * (FMA(KP198912367, T2o, T2f));
|
cannam@167
|
337 R0[WS(rs, 13)] = -(KP1_662939224 * (FNMS(KP668178637, T2p, T2q)));
|
cannam@167
|
338 R0[WS(rs, 9)] = KP1_961570560 * (FNMS(KP198912367, T2f, T2o));
|
cannam@167
|
339 R0[WS(rs, 5)] = KP1_662939224 * (FMA(KP668178637, T2q, T2p));
|
cannam@167
|
340 }
|
cannam@167
|
341 {
|
cannam@167
|
342 E T1H, T1Q, T1O, T1P;
|
cannam@167
|
343 {
|
cannam@167
|
344 E T1z, T1G, T1K, T1N;
|
cannam@167
|
345 T1z = FNMS(KP707106781, T1y, T1x);
|
cannam@167
|
346 T1G = T1C - T1F;
|
cannam@167
|
347 T1H = FMA(KP923879532, T1G, T1z);
|
cannam@167
|
348 T1Q = FNMS(KP923879532, T1G, T1z);
|
cannam@167
|
349 T1K = FNMS(KP707106781, T1J, T1I);
|
cannam@167
|
350 T1N = T1L - T1M;
|
cannam@167
|
351 T1O = FMA(KP923879532, T1N, T1K);
|
cannam@167
|
352 T1P = FNMS(KP923879532, T1N, T1K);
|
cannam@167
|
353 }
|
cannam@167
|
354 R1[WS(rs, 1)] = KP1_913880671 * (FMA(KP303346683, T1O, T1H));
|
cannam@167
|
355 R1[WS(rs, 13)] = -(KP1_763842528 * (FNMS(KP534511135, T1P, T1Q)));
|
cannam@167
|
356 R1[WS(rs, 9)] = KP1_913880671 * (FNMS(KP303346683, T1H, T1O));
|
cannam@167
|
357 R1[WS(rs, 5)] = KP1_763842528 * (FMA(KP534511135, T1Q, T1P));
|
cannam@167
|
358 }
|
cannam@167
|
359 }
|
cannam@167
|
360 }
|
cannam@167
|
361 }
|
cannam@167
|
362
|
cannam@167
|
363 static const kr2c_desc desc = { 32, "r2cbIII_32", {106, 32, 68, 0}, &GENUS };
|
cannam@167
|
364
|
cannam@167
|
365 void X(codelet_r2cbIII_32) (planner *p) {
|
cannam@167
|
366 X(kr2c_register) (p, r2cbIII_32, &desc);
|
cannam@167
|
367 }
|
cannam@167
|
368
|
cannam@167
|
369 #else
|
cannam@167
|
370
|
cannam@167
|
371 /* Generated by: ../../../genfft/gen_r2cb.native -compact -variables 4 -pipeline-latency 4 -sign 1 -n 32 -name r2cbIII_32 -dft-III -include rdft/scalar/r2cbIII.h */
|
cannam@167
|
372
|
cannam@167
|
373 /*
|
cannam@167
|
374 * This function contains 174 FP additions, 84 FP multiplications,
|
cannam@167
|
375 * (or, 138 additions, 48 multiplications, 36 fused multiply/add),
|
cannam@167
|
376 * 66 stack variables, 19 constants, and 64 memory accesses
|
cannam@167
|
377 */
|
cannam@167
|
378 #include "rdft/scalar/r2cbIII.h"
|
cannam@167
|
379
|
cannam@167
|
380 static void r2cbIII_32(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
|
cannam@167
|
381 {
|
cannam@167
|
382 DK(KP1_913880671, +1.913880671464417729871595773960539938965698411);
|
cannam@167
|
383 DK(KP580569354, +0.580569354508924735272384751634790549382952557);
|
cannam@167
|
384 DK(KP942793473, +0.942793473651995297112775251810508755314920638);
|
cannam@167
|
385 DK(KP1_763842528, +1.763842528696710059425513727320776699016885241);
|
cannam@167
|
386 DK(KP1_546020906, +1.546020906725473921621813219516939601942082586);
|
cannam@167
|
387 DK(KP1_268786568, +1.268786568327290996430343226450986741351374190);
|
cannam@167
|
388 DK(KP196034280, +0.196034280659121203988391127777283691722273346);
|
cannam@167
|
389 DK(KP1_990369453, +1.990369453344393772489673906218959843150949737);
|
cannam@167
|
390 DK(KP765366864, +0.765366864730179543456919968060797733522689125);
|
cannam@167
|
391 DK(KP1_847759065, +1.847759065022573512256366378793576573644833252);
|
cannam@167
|
392 DK(KP1_961570560, +1.961570560806460898252364472268478073947867462);
|
cannam@167
|
393 DK(KP390180644, +0.390180644032256535696569736954044481855383236);
|
cannam@167
|
394 DK(KP1_111140466, +1.111140466039204449485661627897065748749874382);
|
cannam@167
|
395 DK(KP1_662939224, +1.662939224605090474157576755235811513477121624);
|
cannam@167
|
396 DK(KP1_414213562, +1.414213562373095048801688724209698078569671875);
|
cannam@167
|
397 DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
|
cannam@167
|
398 DK(KP382683432, +0.382683432365089771728459984030398866761344562);
|
cannam@167
|
399 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
cannam@167
|
400 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
cannam@167
|
401 {
|
cannam@167
|
402 INT i;
|
cannam@167
|
403 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)) {
|
cannam@167
|
404 E T7, T2i, T2F, Tz, T1k, T1I, T1Z, T1x, Te, T22, T2E, T2j, T1f, T1y, TK;
|
cannam@167
|
405 E T1J, Tm, T2B, TW, T1a, T1C, T1L, T28, T2l, Tt, T2A, T17, T1b, T1F, T1M;
|
cannam@167
|
406 E T2d, T2m;
|
cannam@167
|
407 {
|
cannam@167
|
408 E T3, Tv, T1j, T2h, T6, T1g, Ty, T2g;
|
cannam@167
|
409 {
|
cannam@167
|
410 E T1, T2, T1h, T1i;
|
cannam@167
|
411 T1 = Cr[0];
|
cannam@167
|
412 T2 = Cr[WS(csr, 15)];
|
cannam@167
|
413 T3 = T1 + T2;
|
cannam@167
|
414 Tv = T1 - T2;
|
cannam@167
|
415 T1h = Ci[0];
|
cannam@167
|
416 T1i = Ci[WS(csi, 15)];
|
cannam@167
|
417 T1j = T1h + T1i;
|
cannam@167
|
418 T2h = T1i - T1h;
|
cannam@167
|
419 }
|
cannam@167
|
420 {
|
cannam@167
|
421 E T4, T5, Tw, Tx;
|
cannam@167
|
422 T4 = Cr[WS(csr, 8)];
|
cannam@167
|
423 T5 = Cr[WS(csr, 7)];
|
cannam@167
|
424 T6 = T4 + T5;
|
cannam@167
|
425 T1g = T4 - T5;
|
cannam@167
|
426 Tw = Ci[WS(csi, 8)];
|
cannam@167
|
427 Tx = Ci[WS(csi, 7)];
|
cannam@167
|
428 Ty = Tw + Tx;
|
cannam@167
|
429 T2g = Tw - Tx;
|
cannam@167
|
430 }
|
cannam@167
|
431 T7 = T3 + T6;
|
cannam@167
|
432 T2i = T2g + T2h;
|
cannam@167
|
433 T2F = T2h - T2g;
|
cannam@167
|
434 Tz = Tv - Ty;
|
cannam@167
|
435 T1k = T1g + T1j;
|
cannam@167
|
436 T1I = T1g - T1j;
|
cannam@167
|
437 T1Z = T3 - T6;
|
cannam@167
|
438 T1x = Tv + Ty;
|
cannam@167
|
439 }
|
cannam@167
|
440 {
|
cannam@167
|
441 E Ta, TA, TD, T21, Td, TF, TI, T20;
|
cannam@167
|
442 {
|
cannam@167
|
443 E T8, T9, TB, TC;
|
cannam@167
|
444 T8 = Cr[WS(csr, 4)];
|
cannam@167
|
445 T9 = Cr[WS(csr, 11)];
|
cannam@167
|
446 Ta = T8 + T9;
|
cannam@167
|
447 TA = T8 - T9;
|
cannam@167
|
448 TB = Ci[WS(csi, 4)];
|
cannam@167
|
449 TC = Ci[WS(csi, 11)];
|
cannam@167
|
450 TD = TB + TC;
|
cannam@167
|
451 T21 = TB - TC;
|
cannam@167
|
452 }
|
cannam@167
|
453 {
|
cannam@167
|
454 E Tb, Tc, TG, TH;
|
cannam@167
|
455 Tb = Cr[WS(csr, 3)];
|
cannam@167
|
456 Tc = Cr[WS(csr, 12)];
|
cannam@167
|
457 Td = Tb + Tc;
|
cannam@167
|
458 TF = Tb - Tc;
|
cannam@167
|
459 TG = Ci[WS(csi, 3)];
|
cannam@167
|
460 TH = Ci[WS(csi, 12)];
|
cannam@167
|
461 TI = TG + TH;
|
cannam@167
|
462 T20 = TH - TG;
|
cannam@167
|
463 }
|
cannam@167
|
464 Te = Ta + Td;
|
cannam@167
|
465 T22 = T20 - T21;
|
cannam@167
|
466 T2E = T21 + T20;
|
cannam@167
|
467 T2j = Ta - Td;
|
cannam@167
|
468 {
|
cannam@167
|
469 E T1d, T1e, TE, TJ;
|
cannam@167
|
470 T1d = TA + TD;
|
cannam@167
|
471 T1e = TF + TI;
|
cannam@167
|
472 T1f = KP707106781 * (T1d - T1e);
|
cannam@167
|
473 T1y = KP707106781 * (T1d + T1e);
|
cannam@167
|
474 TE = TA - TD;
|
cannam@167
|
475 TJ = TF - TI;
|
cannam@167
|
476 TK = KP707106781 * (TE + TJ);
|
cannam@167
|
477 T1J = KP707106781 * (TE - TJ);
|
cannam@167
|
478 }
|
cannam@167
|
479 }
|
cannam@167
|
480 {
|
cannam@167
|
481 E Ti, TM, TU, T25, Tl, TR, TP, T26, TQ, TV;
|
cannam@167
|
482 {
|
cannam@167
|
483 E Tg, Th, TS, TT;
|
cannam@167
|
484 Tg = Cr[WS(csr, 2)];
|
cannam@167
|
485 Th = Cr[WS(csr, 13)];
|
cannam@167
|
486 Ti = Tg + Th;
|
cannam@167
|
487 TM = Tg - Th;
|
cannam@167
|
488 TS = Ci[WS(csi, 2)];
|
cannam@167
|
489 TT = Ci[WS(csi, 13)];
|
cannam@167
|
490 TU = TS + TT;
|
cannam@167
|
491 T25 = TS - TT;
|
cannam@167
|
492 }
|
cannam@167
|
493 {
|
cannam@167
|
494 E Tj, Tk, TN, TO;
|
cannam@167
|
495 Tj = Cr[WS(csr, 10)];
|
cannam@167
|
496 Tk = Cr[WS(csr, 5)];
|
cannam@167
|
497 Tl = Tj + Tk;
|
cannam@167
|
498 TR = Tj - Tk;
|
cannam@167
|
499 TN = Ci[WS(csi, 10)];
|
cannam@167
|
500 TO = Ci[WS(csi, 5)];
|
cannam@167
|
501 TP = TN + TO;
|
cannam@167
|
502 T26 = TN - TO;
|
cannam@167
|
503 }
|
cannam@167
|
504 Tm = Ti + Tl;
|
cannam@167
|
505 T2B = T26 + T25;
|
cannam@167
|
506 TQ = TM - TP;
|
cannam@167
|
507 TV = TR + TU;
|
cannam@167
|
508 TW = FNMS(KP382683432, TV, KP923879532 * TQ);
|
cannam@167
|
509 T1a = FMA(KP382683432, TQ, KP923879532 * TV);
|
cannam@167
|
510 {
|
cannam@167
|
511 E T1A, T1B, T24, T27;
|
cannam@167
|
512 T1A = TM + TP;
|
cannam@167
|
513 T1B = TU - TR;
|
cannam@167
|
514 T1C = FNMS(KP923879532, T1B, KP382683432 * T1A);
|
cannam@167
|
515 T1L = FMA(KP923879532, T1A, KP382683432 * T1B);
|
cannam@167
|
516 T24 = Ti - Tl;
|
cannam@167
|
517 T27 = T25 - T26;
|
cannam@167
|
518 T28 = T24 - T27;
|
cannam@167
|
519 T2l = T24 + T27;
|
cannam@167
|
520 }
|
cannam@167
|
521 }
|
cannam@167
|
522 {
|
cannam@167
|
523 E Tp, TX, T15, T2a, Ts, T12, T10, T2b, T11, T16;
|
cannam@167
|
524 {
|
cannam@167
|
525 E Tn, To, T13, T14;
|
cannam@167
|
526 Tn = Cr[WS(csr, 1)];
|
cannam@167
|
527 To = Cr[WS(csr, 14)];
|
cannam@167
|
528 Tp = Tn + To;
|
cannam@167
|
529 TX = Tn - To;
|
cannam@167
|
530 T13 = Ci[WS(csi, 1)];
|
cannam@167
|
531 T14 = Ci[WS(csi, 14)];
|
cannam@167
|
532 T15 = T13 + T14;
|
cannam@167
|
533 T2a = T14 - T13;
|
cannam@167
|
534 }
|
cannam@167
|
535 {
|
cannam@167
|
536 E Tq, Tr, TY, TZ;
|
cannam@167
|
537 Tq = Cr[WS(csr, 6)];
|
cannam@167
|
538 Tr = Cr[WS(csr, 9)];
|
cannam@167
|
539 Ts = Tq + Tr;
|
cannam@167
|
540 T12 = Tq - Tr;
|
cannam@167
|
541 TY = Ci[WS(csi, 6)];
|
cannam@167
|
542 TZ = Ci[WS(csi, 9)];
|
cannam@167
|
543 T10 = TY + TZ;
|
cannam@167
|
544 T2b = TY - TZ;
|
cannam@167
|
545 }
|
cannam@167
|
546 Tt = Tp + Ts;
|
cannam@167
|
547 T2A = T2b + T2a;
|
cannam@167
|
548 T11 = TX - T10;
|
cannam@167
|
549 T16 = T12 - T15;
|
cannam@167
|
550 T17 = FMA(KP923879532, T11, KP382683432 * T16);
|
cannam@167
|
551 T1b = FNMS(KP382683432, T11, KP923879532 * T16);
|
cannam@167
|
552 {
|
cannam@167
|
553 E T1D, T1E, T29, T2c;
|
cannam@167
|
554 T1D = TX + T10;
|
cannam@167
|
555 T1E = T12 + T15;
|
cannam@167
|
556 T1F = FNMS(KP923879532, T1E, KP382683432 * T1D);
|
cannam@167
|
557 T1M = FMA(KP923879532, T1D, KP382683432 * T1E);
|
cannam@167
|
558 T29 = Tp - Ts;
|
cannam@167
|
559 T2c = T2a - T2b;
|
cannam@167
|
560 T2d = T29 + T2c;
|
cannam@167
|
561 T2m = T2c - T29;
|
cannam@167
|
562 }
|
cannam@167
|
563 }
|
cannam@167
|
564 {
|
cannam@167
|
565 E Tf, Tu, T2L, T2M, T2N, T2O;
|
cannam@167
|
566 Tf = T7 + Te;
|
cannam@167
|
567 Tu = Tm + Tt;
|
cannam@167
|
568 T2L = Tf - Tu;
|
cannam@167
|
569 T2M = T2B + T2A;
|
cannam@167
|
570 T2N = T2F - T2E;
|
cannam@167
|
571 T2O = T2M + T2N;
|
cannam@167
|
572 R0[0] = KP2_000000000 * (Tf + Tu);
|
cannam@167
|
573 R0[WS(rs, 8)] = KP2_000000000 * (T2N - T2M);
|
cannam@167
|
574 R0[WS(rs, 4)] = KP1_414213562 * (T2L + T2O);
|
cannam@167
|
575 R0[WS(rs, 12)] = KP1_414213562 * (T2O - T2L);
|
cannam@167
|
576 }
|
cannam@167
|
577 {
|
cannam@167
|
578 E T2t, T2x, T2w, T2y;
|
cannam@167
|
579 {
|
cannam@167
|
580 E T2r, T2s, T2u, T2v;
|
cannam@167
|
581 T2r = T1Z - T22;
|
cannam@167
|
582 T2s = KP707106781 * (T2m - T2l);
|
cannam@167
|
583 T2t = T2r + T2s;
|
cannam@167
|
584 T2x = T2r - T2s;
|
cannam@167
|
585 T2u = T2j + T2i;
|
cannam@167
|
586 T2v = KP707106781 * (T28 - T2d);
|
cannam@167
|
587 T2w = T2u - T2v;
|
cannam@167
|
588 T2y = T2v + T2u;
|
cannam@167
|
589 }
|
cannam@167
|
590 R0[WS(rs, 3)] = FMA(KP1_662939224, T2t, KP1_111140466 * T2w);
|
cannam@167
|
591 R0[WS(rs, 15)] = FNMS(KP1_961570560, T2x, KP390180644 * T2y);
|
cannam@167
|
592 R0[WS(rs, 11)] = FNMS(KP1_111140466, T2t, KP1_662939224 * T2w);
|
cannam@167
|
593 R0[WS(rs, 7)] = FMA(KP390180644, T2x, KP1_961570560 * T2y);
|
cannam@167
|
594 }
|
cannam@167
|
595 {
|
cannam@167
|
596 E T2D, T2J, T2I, T2K;
|
cannam@167
|
597 {
|
cannam@167
|
598 E T2z, T2C, T2G, T2H;
|
cannam@167
|
599 T2z = T7 - Te;
|
cannam@167
|
600 T2C = T2A - T2B;
|
cannam@167
|
601 T2D = T2z + T2C;
|
cannam@167
|
602 T2J = T2z - T2C;
|
cannam@167
|
603 T2G = T2E + T2F;
|
cannam@167
|
604 T2H = Tm - Tt;
|
cannam@167
|
605 T2I = T2G - T2H;
|
cannam@167
|
606 T2K = T2H + T2G;
|
cannam@167
|
607 }
|
cannam@167
|
608 R0[WS(rs, 2)] = FMA(KP1_847759065, T2D, KP765366864 * T2I);
|
cannam@167
|
609 R0[WS(rs, 14)] = FNMS(KP1_847759065, T2J, KP765366864 * T2K);
|
cannam@167
|
610 R0[WS(rs, 10)] = FNMS(KP765366864, T2D, KP1_847759065 * T2I);
|
cannam@167
|
611 R0[WS(rs, 6)] = FMA(KP765366864, T2J, KP1_847759065 * T2K);
|
cannam@167
|
612 }
|
cannam@167
|
613 {
|
cannam@167
|
614 E T19, T1n, T1m, T1o;
|
cannam@167
|
615 {
|
cannam@167
|
616 E TL, T18, T1c, T1l;
|
cannam@167
|
617 TL = Tz + TK;
|
cannam@167
|
618 T18 = TW + T17;
|
cannam@167
|
619 T19 = TL + T18;
|
cannam@167
|
620 T1n = TL - T18;
|
cannam@167
|
621 T1c = T1a + T1b;
|
cannam@167
|
622 T1l = T1f + T1k;
|
cannam@167
|
623 T1m = T1c + T1l;
|
cannam@167
|
624 T1o = T1c - T1l;
|
cannam@167
|
625 }
|
cannam@167
|
626 R1[0] = FNMS(KP196034280, T1m, KP1_990369453 * T19);
|
cannam@167
|
627 R1[WS(rs, 12)] = FNMS(KP1_546020906, T1n, KP1_268786568 * T1o);
|
cannam@167
|
628 R1[WS(rs, 8)] = -(FMA(KP196034280, T19, KP1_990369453 * T1m));
|
cannam@167
|
629 R1[WS(rs, 4)] = FMA(KP1_268786568, T1n, KP1_546020906 * T1o);
|
cannam@167
|
630 }
|
cannam@167
|
631 {
|
cannam@167
|
632 E T1r, T1v, T1u, T1w;
|
cannam@167
|
633 {
|
cannam@167
|
634 E T1p, T1q, T1s, T1t;
|
cannam@167
|
635 T1p = Tz - TK;
|
cannam@167
|
636 T1q = T1b - T1a;
|
cannam@167
|
637 T1r = T1p + T1q;
|
cannam@167
|
638 T1v = T1p - T1q;
|
cannam@167
|
639 T1s = T1f - T1k;
|
cannam@167
|
640 T1t = TW - T17;
|
cannam@167
|
641 T1u = T1s - T1t;
|
cannam@167
|
642 T1w = T1t + T1s;
|
cannam@167
|
643 }
|
cannam@167
|
644 R1[WS(rs, 2)] = FMA(KP1_763842528, T1r, KP942793473 * T1u);
|
cannam@167
|
645 R1[WS(rs, 14)] = FNMS(KP1_913880671, T1v, KP580569354 * T1w);
|
cannam@167
|
646 R1[WS(rs, 10)] = FNMS(KP942793473, T1r, KP1_763842528 * T1u);
|
cannam@167
|
647 R1[WS(rs, 6)] = FMA(KP580569354, T1v, KP1_913880671 * T1w);
|
cannam@167
|
648 }
|
cannam@167
|
649 {
|
cannam@167
|
650 E T1T, T1X, T1W, T1Y;
|
cannam@167
|
651 {
|
cannam@167
|
652 E T1R, T1S, T1U, T1V;
|
cannam@167
|
653 T1R = T1x + T1y;
|
cannam@167
|
654 T1S = T1L + T1M;
|
cannam@167
|
655 T1T = T1R - T1S;
|
cannam@167
|
656 T1X = T1R + T1S;
|
cannam@167
|
657 T1U = T1J + T1I;
|
cannam@167
|
658 T1V = T1C - T1F;
|
cannam@167
|
659 T1W = T1U - T1V;
|
cannam@167
|
660 T1Y = T1V + T1U;
|
cannam@167
|
661 }
|
cannam@167
|
662 R1[WS(rs, 3)] = FMA(KP1_546020906, T1T, KP1_268786568 * T1W);
|
cannam@167
|
663 R1[WS(rs, 15)] = FNMS(KP1_990369453, T1X, KP196034280 * T1Y);
|
cannam@167
|
664 R1[WS(rs, 11)] = FNMS(KP1_268786568, T1T, KP1_546020906 * T1W);
|
cannam@167
|
665 R1[WS(rs, 7)] = FMA(KP196034280, T1X, KP1_990369453 * T1Y);
|
cannam@167
|
666 }
|
cannam@167
|
667 {
|
cannam@167
|
668 E T2f, T2p, T2o, T2q;
|
cannam@167
|
669 {
|
cannam@167
|
670 E T23, T2e, T2k, T2n;
|
cannam@167
|
671 T23 = T1Z + T22;
|
cannam@167
|
672 T2e = KP707106781 * (T28 + T2d);
|
cannam@167
|
673 T2f = T23 + T2e;
|
cannam@167
|
674 T2p = T23 - T2e;
|
cannam@167
|
675 T2k = T2i - T2j;
|
cannam@167
|
676 T2n = KP707106781 * (T2l + T2m);
|
cannam@167
|
677 T2o = T2k - T2n;
|
cannam@167
|
678 T2q = T2n + T2k;
|
cannam@167
|
679 }
|
cannam@167
|
680 R0[WS(rs, 1)] = FMA(KP1_961570560, T2f, KP390180644 * T2o);
|
cannam@167
|
681 R0[WS(rs, 13)] = FNMS(KP1_662939224, T2p, KP1_111140466 * T2q);
|
cannam@167
|
682 R0[WS(rs, 9)] = FNMS(KP390180644, T2f, KP1_961570560 * T2o);
|
cannam@167
|
683 R0[WS(rs, 5)] = FMA(KP1_111140466, T2p, KP1_662939224 * T2q);
|
cannam@167
|
684 }
|
cannam@167
|
685 {
|
cannam@167
|
686 E T1H, T1P, T1O, T1Q;
|
cannam@167
|
687 {
|
cannam@167
|
688 E T1z, T1G, T1K, T1N;
|
cannam@167
|
689 T1z = T1x - T1y;
|
cannam@167
|
690 T1G = T1C + T1F;
|
cannam@167
|
691 T1H = T1z + T1G;
|
cannam@167
|
692 T1P = T1z - T1G;
|
cannam@167
|
693 T1K = T1I - T1J;
|
cannam@167
|
694 T1N = T1L - T1M;
|
cannam@167
|
695 T1O = T1K - T1N;
|
cannam@167
|
696 T1Q = T1N + T1K;
|
cannam@167
|
697 }
|
cannam@167
|
698 R1[WS(rs, 1)] = FMA(KP1_913880671, T1H, KP580569354 * T1O);
|
cannam@167
|
699 R1[WS(rs, 13)] = FNMS(KP1_763842528, T1P, KP942793473 * T1Q);
|
cannam@167
|
700 R1[WS(rs, 9)] = FNMS(KP580569354, T1H, KP1_913880671 * T1O);
|
cannam@167
|
701 R1[WS(rs, 5)] = FMA(KP942793473, T1P, KP1_763842528 * T1Q);
|
cannam@167
|
702 }
|
cannam@167
|
703 }
|
cannam@167
|
704 }
|
cannam@167
|
705 }
|
cannam@167
|
706
|
cannam@167
|
707 static const kr2c_desc desc = { 32, "r2cbIII_32", {138, 48, 36, 0}, &GENUS };
|
cannam@167
|
708
|
cannam@167
|
709 void X(codelet_r2cbIII_32) (planner *p) {
|
cannam@167
|
710 X(kr2c_register) (p, r2cbIII_32, &desc);
|
cannam@167
|
711 }
|
cannam@167
|
712
|
cannam@167
|
713 #endif
|