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 08:07:29 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 64 -name r2cb_64 -include rdft/scalar/r2cb.h */
|
cannam@167
|
29
|
cannam@167
|
30 /*
|
cannam@167
|
31 * This function contains 394 FP additions, 216 FP multiplications,
|
cannam@167
|
32 * (or, 178 additions, 0 multiplications, 216 fused multiply/add),
|
cannam@167
|
33 * 109 stack variables, 18 constants, and 128 memory accesses
|
cannam@167
|
34 */
|
cannam@167
|
35 #include "rdft/scalar/r2cb.h"
|
cannam@167
|
36
|
cannam@167
|
37 static void r2cb_64(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(KP1_546020906, +1.546020906725473921621813219516939601942082586);
|
cannam@167
|
40 DK(KP820678790, +0.820678790828660330972281985331011598767386482);
|
cannam@167
|
41 DK(KP1_990369453, +1.990369453344393772489673906218959843150949737);
|
cannam@167
|
42 DK(KP098491403, +0.098491403357164253077197521291327432293052451);
|
cannam@167
|
43 DK(KP1_763842528, +1.763842528696710059425513727320776699016885241);
|
cannam@167
|
44 DK(KP534511135, +0.534511135950791641089685961295362908582039528);
|
cannam@167
|
45 DK(KP1_913880671, +1.913880671464417729871595773960539938965698411);
|
cannam@167
|
46 DK(KP303346683, +0.303346683607342391675883946941299872384187453);
|
cannam@167
|
47 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
cannam@167
|
48 DK(KP1_662939224, +1.662939224605090474157576755235811513477121624);
|
cannam@167
|
49 DK(KP668178637, +0.668178637919298919997757686523080761552472251);
|
cannam@167
|
50 DK(KP1_961570560, +1.961570560806460898252364472268478073947867462);
|
cannam@167
|
51 DK(KP198912367, +0.198912367379658006911597622644676228597850501);
|
cannam@167
|
52 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
cannam@167
|
53 DK(KP1_847759065, +1.847759065022573512256366378793576573644833252);
|
cannam@167
|
54 DK(KP414213562, +0.414213562373095048801688724209698078569671875);
|
cannam@167
|
55 DK(KP1_414213562, +1.414213562373095048801688724209698078569671875);
|
cannam@167
|
56 DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
|
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(256, rs), MAKE_VOLATILE_STRIDE(256, csr), MAKE_VOLATILE_STRIDE(256, csi)) {
|
cannam@167
|
60 E T9, T5H, T4p, T5j, T1b, T2T, T3j, T3Z, Tg, T5I, T1m, T2U, T3m, T40, T4u;
|
cannam@167
|
61 E T5k, T1s, T3o, T3r, T1J, To, Tv, T5K, T5L, T5M, T5N, T4A, T5n, T1D, T3s;
|
cannam@167
|
62 E T4F, T5m, T1M, T3p, T1U, T3w, T3H, T2z, TM, T5Q, T5Y, T6f, T4M, T5q, T25;
|
cannam@167
|
63 E T3I, T53, T5t, T2C, T3x, T11, T5V, T4W, T55, T5T, T6g, T2h, T2E, T2s, T2F;
|
cannam@167
|
64 E T3E, T3L, T4R, T54, T3B, T3K;
|
cannam@167
|
65 {
|
cannam@167
|
66 E T4, T14, T3, T13, T8, T16, T19, T4o, T1, T2, T5, T4n;
|
cannam@167
|
67 T4 = Cr[WS(csr, 16)];
|
cannam@167
|
68 T14 = Ci[WS(csi, 16)];
|
cannam@167
|
69 T1 = Cr[0];
|
cannam@167
|
70 T2 = Cr[WS(csr, 32)];
|
cannam@167
|
71 T3 = T1 + T2;
|
cannam@167
|
72 T13 = T1 - T2;
|
cannam@167
|
73 {
|
cannam@167
|
74 E T6, T7, T17, T18;
|
cannam@167
|
75 T6 = Cr[WS(csr, 8)];
|
cannam@167
|
76 T7 = Cr[WS(csr, 24)];
|
cannam@167
|
77 T8 = T6 + T7;
|
cannam@167
|
78 T16 = T6 - T7;
|
cannam@167
|
79 T17 = Ci[WS(csi, 8)];
|
cannam@167
|
80 T18 = Ci[WS(csi, 24)];
|
cannam@167
|
81 T19 = T17 + T18;
|
cannam@167
|
82 T4o = T17 - T18;
|
cannam@167
|
83 }
|
cannam@167
|
84 T5 = FMA(KP2_000000000, T4, T3);
|
cannam@167
|
85 T9 = FMA(KP2_000000000, T8, T5);
|
cannam@167
|
86 T5H = FNMS(KP2_000000000, T8, T5);
|
cannam@167
|
87 T4n = FNMS(KP2_000000000, T4, T3);
|
cannam@167
|
88 T4p = FNMS(KP2_000000000, T4o, T4n);
|
cannam@167
|
89 T5j = FMA(KP2_000000000, T4o, T4n);
|
cannam@167
|
90 {
|
cannam@167
|
91 E T15, T1a, T3h, T3i;
|
cannam@167
|
92 T15 = FNMS(KP2_000000000, T14, T13);
|
cannam@167
|
93 T1a = T16 - T19;
|
cannam@167
|
94 T1b = FMA(KP1_414213562, T1a, T15);
|
cannam@167
|
95 T2T = FNMS(KP1_414213562, T1a, T15);
|
cannam@167
|
96 T3h = FMA(KP2_000000000, T14, T13);
|
cannam@167
|
97 T3i = T16 + T19;
|
cannam@167
|
98 T3j = FNMS(KP1_414213562, T3i, T3h);
|
cannam@167
|
99 T3Z = FMA(KP1_414213562, T3i, T3h);
|
cannam@167
|
100 }
|
cannam@167
|
101 }
|
cannam@167
|
102 {
|
cannam@167
|
103 E Tc, T1c, T1k, T4r, Tf, T1h, T1f, T4s, T1g, T1l;
|
cannam@167
|
104 {
|
cannam@167
|
105 E Ta, Tb, T1i, T1j;
|
cannam@167
|
106 Ta = Cr[WS(csr, 4)];
|
cannam@167
|
107 Tb = Cr[WS(csr, 28)];
|
cannam@167
|
108 Tc = Ta + Tb;
|
cannam@167
|
109 T1c = Ta - Tb;
|
cannam@167
|
110 T1i = Ci[WS(csi, 4)];
|
cannam@167
|
111 T1j = Ci[WS(csi, 28)];
|
cannam@167
|
112 T1k = T1i + T1j;
|
cannam@167
|
113 T4r = T1i - T1j;
|
cannam@167
|
114 }
|
cannam@167
|
115 {
|
cannam@167
|
116 E Td, Te, T1d, T1e;
|
cannam@167
|
117 Td = Cr[WS(csr, 20)];
|
cannam@167
|
118 Te = Cr[WS(csr, 12)];
|
cannam@167
|
119 Tf = Td + Te;
|
cannam@167
|
120 T1h = Td - Te;
|
cannam@167
|
121 T1d = Ci[WS(csi, 20)];
|
cannam@167
|
122 T1e = Ci[WS(csi, 12)];
|
cannam@167
|
123 T1f = T1d + T1e;
|
cannam@167
|
124 T4s = T1d - T1e;
|
cannam@167
|
125 }
|
cannam@167
|
126 Tg = Tc + Tf;
|
cannam@167
|
127 T5I = T4s + T4r;
|
cannam@167
|
128 T1g = T1c - T1f;
|
cannam@167
|
129 T1l = T1h + T1k;
|
cannam@167
|
130 T1m = FNMS(KP414213562, T1l, T1g);
|
cannam@167
|
131 T2U = FMA(KP414213562, T1g, T1l);
|
cannam@167
|
132 {
|
cannam@167
|
133 E T3k, T3l, T4q, T4t;
|
cannam@167
|
134 T3k = T1k - T1h;
|
cannam@167
|
135 T3l = T1c + T1f;
|
cannam@167
|
136 T3m = FNMS(KP414213562, T3l, T3k);
|
cannam@167
|
137 T40 = FMA(KP414213562, T3k, T3l);
|
cannam@167
|
138 T4q = Tc - Tf;
|
cannam@167
|
139 T4t = T4r - T4s;
|
cannam@167
|
140 T4u = T4q - T4t;
|
cannam@167
|
141 T5k = T4q + T4t;
|
cannam@167
|
142 }
|
cannam@167
|
143 }
|
cannam@167
|
144 {
|
cannam@167
|
145 E Tk, T1o, T1I, T4C, Tn, T1F, T1r, T4D, Tr, T1t, T1w, T4y, Tu, T1y, T1B;
|
cannam@167
|
146 E T4x;
|
cannam@167
|
147 {
|
cannam@167
|
148 E Ti, Tj, T1G, T1H;
|
cannam@167
|
149 Ti = Cr[WS(csr, 2)];
|
cannam@167
|
150 Tj = Cr[WS(csr, 30)];
|
cannam@167
|
151 Tk = Ti + Tj;
|
cannam@167
|
152 T1o = Ti - Tj;
|
cannam@167
|
153 T1G = Ci[WS(csi, 2)];
|
cannam@167
|
154 T1H = Ci[WS(csi, 30)];
|
cannam@167
|
155 T1I = T1G + T1H;
|
cannam@167
|
156 T4C = T1G - T1H;
|
cannam@167
|
157 }
|
cannam@167
|
158 {
|
cannam@167
|
159 E Tl, Tm, T1p, T1q;
|
cannam@167
|
160 Tl = Cr[WS(csr, 18)];
|
cannam@167
|
161 Tm = Cr[WS(csr, 14)];
|
cannam@167
|
162 Tn = Tl + Tm;
|
cannam@167
|
163 T1F = Tl - Tm;
|
cannam@167
|
164 T1p = Ci[WS(csi, 18)];
|
cannam@167
|
165 T1q = Ci[WS(csi, 14)];
|
cannam@167
|
166 T1r = T1p + T1q;
|
cannam@167
|
167 T4D = T1p - T1q;
|
cannam@167
|
168 }
|
cannam@167
|
169 {
|
cannam@167
|
170 E Tp, Tq, T1u, T1v;
|
cannam@167
|
171 Tp = Cr[WS(csr, 10)];
|
cannam@167
|
172 Tq = Cr[WS(csr, 22)];
|
cannam@167
|
173 Tr = Tp + Tq;
|
cannam@167
|
174 T1t = Tp - Tq;
|
cannam@167
|
175 T1u = Ci[WS(csi, 10)];
|
cannam@167
|
176 T1v = Ci[WS(csi, 22)];
|
cannam@167
|
177 T1w = T1u + T1v;
|
cannam@167
|
178 T4y = T1u - T1v;
|
cannam@167
|
179 }
|
cannam@167
|
180 {
|
cannam@167
|
181 E Ts, Tt, T1z, T1A;
|
cannam@167
|
182 Ts = Cr[WS(csr, 6)];
|
cannam@167
|
183 Tt = Cr[WS(csr, 26)];
|
cannam@167
|
184 Tu = Ts + Tt;
|
cannam@167
|
185 T1y = Ts - Tt;
|
cannam@167
|
186 T1z = Ci[WS(csi, 6)];
|
cannam@167
|
187 T1A = Ci[WS(csi, 26)];
|
cannam@167
|
188 T1B = T1z + T1A;
|
cannam@167
|
189 T4x = T1A - T1z;
|
cannam@167
|
190 }
|
cannam@167
|
191 T1s = T1o - T1r;
|
cannam@167
|
192 T3o = T1o + T1r;
|
cannam@167
|
193 T3r = T1I - T1F;
|
cannam@167
|
194 T1J = T1F + T1I;
|
cannam@167
|
195 To = Tk + Tn;
|
cannam@167
|
196 Tv = Tr + Tu;
|
cannam@167
|
197 T5K = To - Tv;
|
cannam@167
|
198 {
|
cannam@167
|
199 E T4w, T4z, T1x, T1C;
|
cannam@167
|
200 T5L = T4D + T4C;
|
cannam@167
|
201 T5M = T4y + T4x;
|
cannam@167
|
202 T5N = T5L - T5M;
|
cannam@167
|
203 T4w = Tk - Tn;
|
cannam@167
|
204 T4z = T4x - T4y;
|
cannam@167
|
205 T4A = T4w + T4z;
|
cannam@167
|
206 T5n = T4w - T4z;
|
cannam@167
|
207 T1x = T1t - T1w;
|
cannam@167
|
208 T1C = T1y - T1B;
|
cannam@167
|
209 T1D = T1x + T1C;
|
cannam@167
|
210 T3s = T1C - T1x;
|
cannam@167
|
211 {
|
cannam@167
|
212 E T4B, T4E, T1K, T1L;
|
cannam@167
|
213 T4B = Tr - Tu;
|
cannam@167
|
214 T4E = T4C - T4D;
|
cannam@167
|
215 T4F = T4B + T4E;
|
cannam@167
|
216 T5m = T4E - T4B;
|
cannam@167
|
217 T1K = T1t + T1w;
|
cannam@167
|
218 T1L = T1y + T1B;
|
cannam@167
|
219 T1M = T1K - T1L;
|
cannam@167
|
220 T3p = T1K + T1L;
|
cannam@167
|
221 }
|
cannam@167
|
222 }
|
cannam@167
|
223 }
|
cannam@167
|
224 {
|
cannam@167
|
225 E TA, T1Q, T2y, T50, TD, T2v, T1T, T51, TH, T1V, T1Y, T4K, TK, T20, T23;
|
cannam@167
|
226 E T4J;
|
cannam@167
|
227 {
|
cannam@167
|
228 E Ty, Tz, T2w, T2x;
|
cannam@167
|
229 Ty = Cr[WS(csr, 1)];
|
cannam@167
|
230 Tz = Cr[WS(csr, 31)];
|
cannam@167
|
231 TA = Ty + Tz;
|
cannam@167
|
232 T1Q = Ty - Tz;
|
cannam@167
|
233 T2w = Ci[WS(csi, 1)];
|
cannam@167
|
234 T2x = Ci[WS(csi, 31)];
|
cannam@167
|
235 T2y = T2w + T2x;
|
cannam@167
|
236 T50 = T2w - T2x;
|
cannam@167
|
237 }
|
cannam@167
|
238 {
|
cannam@167
|
239 E TB, TC, T1R, T1S;
|
cannam@167
|
240 TB = Cr[WS(csr, 17)];
|
cannam@167
|
241 TC = Cr[WS(csr, 15)];
|
cannam@167
|
242 TD = TB + TC;
|
cannam@167
|
243 T2v = TB - TC;
|
cannam@167
|
244 T1R = Ci[WS(csi, 17)];
|
cannam@167
|
245 T1S = Ci[WS(csi, 15)];
|
cannam@167
|
246 T1T = T1R + T1S;
|
cannam@167
|
247 T51 = T1R - T1S;
|
cannam@167
|
248 }
|
cannam@167
|
249 {
|
cannam@167
|
250 E TF, TG, T1W, T1X;
|
cannam@167
|
251 TF = Cr[WS(csr, 9)];
|
cannam@167
|
252 TG = Cr[WS(csr, 23)];
|
cannam@167
|
253 TH = TF + TG;
|
cannam@167
|
254 T1V = TF - TG;
|
cannam@167
|
255 T1W = Ci[WS(csi, 9)];
|
cannam@167
|
256 T1X = Ci[WS(csi, 23)];
|
cannam@167
|
257 T1Y = T1W + T1X;
|
cannam@167
|
258 T4K = T1W - T1X;
|
cannam@167
|
259 }
|
cannam@167
|
260 {
|
cannam@167
|
261 E TI, TJ, T21, T22;
|
cannam@167
|
262 TI = Cr[WS(csr, 7)];
|
cannam@167
|
263 TJ = Cr[WS(csr, 25)];
|
cannam@167
|
264 TK = TI + TJ;
|
cannam@167
|
265 T20 = TI - TJ;
|
cannam@167
|
266 T21 = Ci[WS(csi, 7)];
|
cannam@167
|
267 T22 = Ci[WS(csi, 25)];
|
cannam@167
|
268 T23 = T21 + T22;
|
cannam@167
|
269 T4J = T22 - T21;
|
cannam@167
|
270 }
|
cannam@167
|
271 {
|
cannam@167
|
272 E TE, TL, T1Z, T24;
|
cannam@167
|
273 T1U = T1Q - T1T;
|
cannam@167
|
274 T3w = T1Q + T1T;
|
cannam@167
|
275 T3H = T2y - T2v;
|
cannam@167
|
276 T2z = T2v + T2y;
|
cannam@167
|
277 TE = TA + TD;
|
cannam@167
|
278 TL = TH + TK;
|
cannam@167
|
279 TM = TE + TL;
|
cannam@167
|
280 T5Q = TE - TL;
|
cannam@167
|
281 {
|
cannam@167
|
282 E T5W, T5X, T4I, T4L;
|
cannam@167
|
283 T5W = T51 + T50;
|
cannam@167
|
284 T5X = T4K + T4J;
|
cannam@167
|
285 T5Y = T5W - T5X;
|
cannam@167
|
286 T6f = T5X + T5W;
|
cannam@167
|
287 T4I = TA - TD;
|
cannam@167
|
288 T4L = T4J - T4K;
|
cannam@167
|
289 T4M = T4I + T4L;
|
cannam@167
|
290 T5q = T4I - T4L;
|
cannam@167
|
291 }
|
cannam@167
|
292 T1Z = T1V - T1Y;
|
cannam@167
|
293 T24 = T20 - T23;
|
cannam@167
|
294 T25 = T1Z + T24;
|
cannam@167
|
295 T3I = T24 - T1Z;
|
cannam@167
|
296 {
|
cannam@167
|
297 E T4Z, T52, T2A, T2B;
|
cannam@167
|
298 T4Z = TH - TK;
|
cannam@167
|
299 T52 = T50 - T51;
|
cannam@167
|
300 T53 = T4Z + T52;
|
cannam@167
|
301 T5t = T52 - T4Z;
|
cannam@167
|
302 T2A = T1V + T1Y;
|
cannam@167
|
303 T2B = T20 + T23;
|
cannam@167
|
304 T2C = T2A - T2B;
|
cannam@167
|
305 T3x = T2A + T2B;
|
cannam@167
|
306 }
|
cannam@167
|
307 }
|
cannam@167
|
308 }
|
cannam@167
|
309 {
|
cannam@167
|
310 E TP, T27, T2f, T4O, TS, T2c, T2a, T4P, TW, T2i, T2q, T4T, TZ, T2n, T2l;
|
cannam@167
|
311 E T4U;
|
cannam@167
|
312 {
|
cannam@167
|
313 E TN, TO, T2d, T2e;
|
cannam@167
|
314 TN = Cr[WS(csr, 5)];
|
cannam@167
|
315 TO = Cr[WS(csr, 27)];
|
cannam@167
|
316 TP = TN + TO;
|
cannam@167
|
317 T27 = TN - TO;
|
cannam@167
|
318 T2d = Ci[WS(csi, 5)];
|
cannam@167
|
319 T2e = Ci[WS(csi, 27)];
|
cannam@167
|
320 T2f = T2d + T2e;
|
cannam@167
|
321 T4O = T2d - T2e;
|
cannam@167
|
322 }
|
cannam@167
|
323 {
|
cannam@167
|
324 E TQ, TR, T28, T29;
|
cannam@167
|
325 TQ = Cr[WS(csr, 21)];
|
cannam@167
|
326 TR = Cr[WS(csr, 11)];
|
cannam@167
|
327 TS = TQ + TR;
|
cannam@167
|
328 T2c = TQ - TR;
|
cannam@167
|
329 T28 = Ci[WS(csi, 21)];
|
cannam@167
|
330 T29 = Ci[WS(csi, 11)];
|
cannam@167
|
331 T2a = T28 + T29;
|
cannam@167
|
332 T4P = T28 - T29;
|
cannam@167
|
333 }
|
cannam@167
|
334 {
|
cannam@167
|
335 E TU, TV, T2o, T2p;
|
cannam@167
|
336 TU = Cr[WS(csr, 3)];
|
cannam@167
|
337 TV = Cr[WS(csr, 29)];
|
cannam@167
|
338 TW = TU + TV;
|
cannam@167
|
339 T2i = TU - TV;
|
cannam@167
|
340 T2o = Ci[WS(csi, 3)];
|
cannam@167
|
341 T2p = Ci[WS(csi, 29)];
|
cannam@167
|
342 T2q = T2o + T2p;
|
cannam@167
|
343 T4T = T2p - T2o;
|
cannam@167
|
344 }
|
cannam@167
|
345 {
|
cannam@167
|
346 E TX, TY, T2j, T2k;
|
cannam@167
|
347 TX = Cr[WS(csr, 13)];
|
cannam@167
|
348 TY = Cr[WS(csr, 19)];
|
cannam@167
|
349 TZ = TX + TY;
|
cannam@167
|
350 T2n = TX - TY;
|
cannam@167
|
351 T2j = Ci[WS(csi, 13)];
|
cannam@167
|
352 T2k = Ci[WS(csi, 19)];
|
cannam@167
|
353 T2l = T2j + T2k;
|
cannam@167
|
354 T4U = T2j - T2k;
|
cannam@167
|
355 }
|
cannam@167
|
356 {
|
cannam@167
|
357 E TT, T10, T4S, T4V;
|
cannam@167
|
358 TT = TP + TS;
|
cannam@167
|
359 T10 = TW + TZ;
|
cannam@167
|
360 T11 = TT + T10;
|
cannam@167
|
361 T5V = TT - T10;
|
cannam@167
|
362 T4S = TW - TZ;
|
cannam@167
|
363 T4V = T4T - T4U;
|
cannam@167
|
364 T4W = T4S + T4V;
|
cannam@167
|
365 T55 = T4V - T4S;
|
cannam@167
|
366 }
|
cannam@167
|
367 {
|
cannam@167
|
368 E T5R, T5S, T2b, T2g;
|
cannam@167
|
369 T5R = T4U + T4T;
|
cannam@167
|
370 T5S = T4P + T4O;
|
cannam@167
|
371 T5T = T5R - T5S;
|
cannam@167
|
372 T6g = T5S + T5R;
|
cannam@167
|
373 T2b = T27 - T2a;
|
cannam@167
|
374 T2g = T2c + T2f;
|
cannam@167
|
375 T2h = FNMS(KP414213562, T2g, T2b);
|
cannam@167
|
376 T2E = FMA(KP414213562, T2b, T2g);
|
cannam@167
|
377 }
|
cannam@167
|
378 {
|
cannam@167
|
379 E T2m, T2r, T3C, T3D;
|
cannam@167
|
380 T2m = T2i - T2l;
|
cannam@167
|
381 T2r = T2n - T2q;
|
cannam@167
|
382 T2s = FMA(KP414213562, T2r, T2m);
|
cannam@167
|
383 T2F = FNMS(KP414213562, T2m, T2r);
|
cannam@167
|
384 T3C = T2n + T2q;
|
cannam@167
|
385 T3D = T2i + T2l;
|
cannam@167
|
386 T3E = FNMS(KP414213562, T3D, T3C);
|
cannam@167
|
387 T3L = FMA(KP414213562, T3C, T3D);
|
cannam@167
|
388 }
|
cannam@167
|
389 {
|
cannam@167
|
390 E T4N, T4Q, T3z, T3A;
|
cannam@167
|
391 T4N = TP - TS;
|
cannam@167
|
392 T4Q = T4O - T4P;
|
cannam@167
|
393 T4R = T4N - T4Q;
|
cannam@167
|
394 T54 = T4N + T4Q;
|
cannam@167
|
395 T3z = T2f - T2c;
|
cannam@167
|
396 T3A = T27 + T2a;
|
cannam@167
|
397 T3B = FNMS(KP414213562, T3A, T3z);
|
cannam@167
|
398 T3K = FMA(KP414213562, T3z, T3A);
|
cannam@167
|
399 }
|
cannam@167
|
400 }
|
cannam@167
|
401 {
|
cannam@167
|
402 E T12, T6m, Tx, T6l, Th, Tw;
|
cannam@167
|
403 T12 = TM + T11;
|
cannam@167
|
404 T6m = T6g + T6f;
|
cannam@167
|
405 Th = FMA(KP2_000000000, Tg, T9);
|
cannam@167
|
406 Tw = To + Tv;
|
cannam@167
|
407 Tx = FMA(KP2_000000000, Tw, Th);
|
cannam@167
|
408 T6l = FNMS(KP2_000000000, Tw, Th);
|
cannam@167
|
409 R0[WS(rs, 16)] = FNMS(KP2_000000000, T12, Tx);
|
cannam@167
|
410 R0[WS(rs, 24)] = FMA(KP2_000000000, T6m, T6l);
|
cannam@167
|
411 R0[0] = FMA(KP2_000000000, T12, Tx);
|
cannam@167
|
412 R0[WS(rs, 8)] = FNMS(KP2_000000000, T6m, T6l);
|
cannam@167
|
413 }
|
cannam@167
|
414 {
|
cannam@167
|
415 E T65, T69, T68, T6a;
|
cannam@167
|
416 {
|
cannam@167
|
417 E T63, T64, T66, T67;
|
cannam@167
|
418 T63 = FMA(KP2_000000000, T5I, T5H);
|
cannam@167
|
419 T64 = T5K + T5N;
|
cannam@167
|
420 T65 = FNMS(KP1_414213562, T64, T63);
|
cannam@167
|
421 T69 = FMA(KP1_414213562, T64, T63);
|
cannam@167
|
422 T66 = T5Y - T5V;
|
cannam@167
|
423 T67 = T5Q - T5T;
|
cannam@167
|
424 T68 = FNMS(KP414213562, T67, T66);
|
cannam@167
|
425 T6a = FMA(KP414213562, T66, T67);
|
cannam@167
|
426 }
|
cannam@167
|
427 R0[WS(rs, 6)] = FNMS(KP1_847759065, T68, T65);
|
cannam@167
|
428 R0[WS(rs, 30)] = FMA(KP1_847759065, T6a, T69);
|
cannam@167
|
429 R0[WS(rs, 22)] = FMA(KP1_847759065, T68, T65);
|
cannam@167
|
430 R0[WS(rs, 14)] = FNMS(KP1_847759065, T6a, T69);
|
cannam@167
|
431 }
|
cannam@167
|
432 {
|
cannam@167
|
433 E T6d, T6j, T6i, T6k;
|
cannam@167
|
434 {
|
cannam@167
|
435 E T6b, T6c, T6e, T6h;
|
cannam@167
|
436 T6b = FNMS(KP2_000000000, Tg, T9);
|
cannam@167
|
437 T6c = T5M + T5L;
|
cannam@167
|
438 T6d = FNMS(KP2_000000000, T6c, T6b);
|
cannam@167
|
439 T6j = FMA(KP2_000000000, T6c, T6b);
|
cannam@167
|
440 T6e = TM - T11;
|
cannam@167
|
441 T6h = T6f - T6g;
|
cannam@167
|
442 T6i = T6e - T6h;
|
cannam@167
|
443 T6k = T6e + T6h;
|
cannam@167
|
444 }
|
cannam@167
|
445 R0[WS(rs, 20)] = FNMS(KP1_414213562, T6i, T6d);
|
cannam@167
|
446 R0[WS(rs, 28)] = FMA(KP1_414213562, T6k, T6j);
|
cannam@167
|
447 R0[WS(rs, 4)] = FMA(KP1_414213562, T6i, T6d);
|
cannam@167
|
448 R0[WS(rs, 12)] = FNMS(KP1_414213562, T6k, T6j);
|
cannam@167
|
449 }
|
cannam@167
|
450 {
|
cannam@167
|
451 E T5P, T61, T60, T62;
|
cannam@167
|
452 {
|
cannam@167
|
453 E T5J, T5O, T5U, T5Z;
|
cannam@167
|
454 T5J = FNMS(KP2_000000000, T5I, T5H);
|
cannam@167
|
455 T5O = T5K - T5N;
|
cannam@167
|
456 T5P = FMA(KP1_414213562, T5O, T5J);
|
cannam@167
|
457 T61 = FNMS(KP1_414213562, T5O, T5J);
|
cannam@167
|
458 T5U = T5Q + T5T;
|
cannam@167
|
459 T5Z = T5V + T5Y;
|
cannam@167
|
460 T60 = FNMS(KP414213562, T5Z, T5U);
|
cannam@167
|
461 T62 = FMA(KP414213562, T5U, T5Z);
|
cannam@167
|
462 }
|
cannam@167
|
463 R0[WS(rs, 18)] = FNMS(KP1_847759065, T60, T5P);
|
cannam@167
|
464 R0[WS(rs, 26)] = FMA(KP1_847759065, T62, T61);
|
cannam@167
|
465 R0[WS(rs, 2)] = FMA(KP1_847759065, T60, T5P);
|
cannam@167
|
466 R0[WS(rs, 10)] = FNMS(KP1_847759065, T62, T61);
|
cannam@167
|
467 }
|
cannam@167
|
468 {
|
cannam@167
|
469 E T4Y, T5f, T57, T5e, T4H, T59, T5d, T5h, T4X, T56;
|
cannam@167
|
470 T4X = T4R + T4W;
|
cannam@167
|
471 T4Y = FMA(KP707106781, T4X, T4M);
|
cannam@167
|
472 T5f = FNMS(KP707106781, T4X, T4M);
|
cannam@167
|
473 T56 = T54 + T55;
|
cannam@167
|
474 T57 = FMA(KP707106781, T56, T53);
|
cannam@167
|
475 T5e = FNMS(KP707106781, T56, T53);
|
cannam@167
|
476 {
|
cannam@167
|
477 E T4v, T4G, T5b, T5c;
|
cannam@167
|
478 T4v = FMA(KP1_414213562, T4u, T4p);
|
cannam@167
|
479 T4G = FNMS(KP414213562, T4F, T4A);
|
cannam@167
|
480 T4H = FMA(KP1_847759065, T4G, T4v);
|
cannam@167
|
481 T59 = FNMS(KP1_847759065, T4G, T4v);
|
cannam@167
|
482 T5b = FNMS(KP1_414213562, T4u, T4p);
|
cannam@167
|
483 T5c = FMA(KP414213562, T4A, T4F);
|
cannam@167
|
484 T5d = FNMS(KP1_847759065, T5c, T5b);
|
cannam@167
|
485 T5h = FMA(KP1_847759065, T5c, T5b);
|
cannam@167
|
486 }
|
cannam@167
|
487 {
|
cannam@167
|
488 E T58, T5i, T5a, T5g;
|
cannam@167
|
489 T58 = FNMS(KP198912367, T57, T4Y);
|
cannam@167
|
490 R0[WS(rs, 17)] = FNMS(KP1_961570560, T58, T4H);
|
cannam@167
|
491 R0[WS(rs, 1)] = FMA(KP1_961570560, T58, T4H);
|
cannam@167
|
492 T5i = FMA(KP668178637, T5e, T5f);
|
cannam@167
|
493 R0[WS(rs, 13)] = FNMS(KP1_662939224, T5i, T5h);
|
cannam@167
|
494 R0[WS(rs, 29)] = FMA(KP1_662939224, T5i, T5h);
|
cannam@167
|
495 T5a = FMA(KP198912367, T4Y, T57);
|
cannam@167
|
496 R0[WS(rs, 9)] = FNMS(KP1_961570560, T5a, T59);
|
cannam@167
|
497 R0[WS(rs, 25)] = FMA(KP1_961570560, T5a, T59);
|
cannam@167
|
498 T5g = FNMS(KP668178637, T5f, T5e);
|
cannam@167
|
499 R0[WS(rs, 5)] = FNMS(KP1_662939224, T5g, T5d);
|
cannam@167
|
500 R0[WS(rs, 21)] = FMA(KP1_662939224, T5g, T5d);
|
cannam@167
|
501 }
|
cannam@167
|
502 }
|
cannam@167
|
503 {
|
cannam@167
|
504 E T5s, T5D, T5v, T5C, T5p, T5x, T5B, T5F, T5r, T5u;
|
cannam@167
|
505 T5r = T54 - T55;
|
cannam@167
|
506 T5s = FNMS(KP707106781, T5r, T5q);
|
cannam@167
|
507 T5D = FMA(KP707106781, T5r, T5q);
|
cannam@167
|
508 T5u = T4W - T4R;
|
cannam@167
|
509 T5v = FNMS(KP707106781, T5u, T5t);
|
cannam@167
|
510 T5C = FMA(KP707106781, T5u, T5t);
|
cannam@167
|
511 {
|
cannam@167
|
512 E T5l, T5o, T5z, T5A;
|
cannam@167
|
513 T5l = FNMS(KP1_414213562, T5k, T5j);
|
cannam@167
|
514 T5o = FNMS(KP414213562, T5n, T5m);
|
cannam@167
|
515 T5p = FNMS(KP1_847759065, T5o, T5l);
|
cannam@167
|
516 T5x = FMA(KP1_847759065, T5o, T5l);
|
cannam@167
|
517 T5z = FMA(KP1_414213562, T5k, T5j);
|
cannam@167
|
518 T5A = FMA(KP414213562, T5m, T5n);
|
cannam@167
|
519 T5B = FNMS(KP1_847759065, T5A, T5z);
|
cannam@167
|
520 T5F = FMA(KP1_847759065, T5A, T5z);
|
cannam@167
|
521 }
|
cannam@167
|
522 {
|
cannam@167
|
523 E T5w, T5G, T5y, T5E;
|
cannam@167
|
524 T5w = FNMS(KP668178637, T5v, T5s);
|
cannam@167
|
525 R0[WS(rs, 19)] = FNMS(KP1_662939224, T5w, T5p);
|
cannam@167
|
526 R0[WS(rs, 3)] = FMA(KP1_662939224, T5w, T5p);
|
cannam@167
|
527 T5G = FMA(KP198912367, T5C, T5D);
|
cannam@167
|
528 R0[WS(rs, 15)] = FNMS(KP1_961570560, T5G, T5F);
|
cannam@167
|
529 R0[WS(rs, 31)] = FMA(KP1_961570560, T5G, T5F);
|
cannam@167
|
530 T5y = FMA(KP668178637, T5s, T5v);
|
cannam@167
|
531 R0[WS(rs, 11)] = FNMS(KP1_662939224, T5y, T5x);
|
cannam@167
|
532 R0[WS(rs, 27)] = FMA(KP1_662939224, T5y, T5x);
|
cannam@167
|
533 T5E = FNMS(KP198912367, T5D, T5C);
|
cannam@167
|
534 R0[WS(rs, 7)] = FNMS(KP1_961570560, T5E, T5B);
|
cannam@167
|
535 R0[WS(rs, 23)] = FMA(KP1_961570560, T5E, T5B);
|
cannam@167
|
536 }
|
cannam@167
|
537 }
|
cannam@167
|
538 {
|
cannam@167
|
539 E T3n, T3R, T3u, T3S, T3G, T3V, T3N, T3U, T3q, T3t;
|
cannam@167
|
540 T3n = FNMS(KP1_847759065, T3m, T3j);
|
cannam@167
|
541 T3R = FMA(KP1_847759065, T3m, T3j);
|
cannam@167
|
542 T3q = FNMS(KP707106781, T3p, T3o);
|
cannam@167
|
543 T3t = FNMS(KP707106781, T3s, T3r);
|
cannam@167
|
544 T3u = FNMS(KP668178637, T3t, T3q);
|
cannam@167
|
545 T3S = FMA(KP668178637, T3q, T3t);
|
cannam@167
|
546 {
|
cannam@167
|
547 E T3y, T3F, T3J, T3M;
|
cannam@167
|
548 T3y = FNMS(KP707106781, T3x, T3w);
|
cannam@167
|
549 T3F = T3B + T3E;
|
cannam@167
|
550 T3G = FNMS(KP923879532, T3F, T3y);
|
cannam@167
|
551 T3V = FMA(KP923879532, T3F, T3y);
|
cannam@167
|
552 T3J = FNMS(KP707106781, T3I, T3H);
|
cannam@167
|
553 T3M = T3K - T3L;
|
cannam@167
|
554 T3N = FMA(KP923879532, T3M, T3J);
|
cannam@167
|
555 T3U = FNMS(KP923879532, T3M, T3J);
|
cannam@167
|
556 }
|
cannam@167
|
557 {
|
cannam@167
|
558 E T3v, T3O, T3X, T3Y;
|
cannam@167
|
559 T3v = FMA(KP1_662939224, T3u, T3n);
|
cannam@167
|
560 T3O = FNMS(KP303346683, T3N, T3G);
|
cannam@167
|
561 R1[WS(rs, 17)] = FNMS(KP1_913880671, T3O, T3v);
|
cannam@167
|
562 R1[WS(rs, 1)] = FMA(KP1_913880671, T3O, T3v);
|
cannam@167
|
563 T3X = FMA(KP1_662939224, T3S, T3R);
|
cannam@167
|
564 T3Y = FMA(KP534511135, T3U, T3V);
|
cannam@167
|
565 R1[WS(rs, 13)] = FNMS(KP1_763842528, T3Y, T3X);
|
cannam@167
|
566 R1[WS(rs, 29)] = FMA(KP1_763842528, T3Y, T3X);
|
cannam@167
|
567 }
|
cannam@167
|
568 {
|
cannam@167
|
569 E T3P, T3Q, T3T, T3W;
|
cannam@167
|
570 T3P = FNMS(KP1_662939224, T3u, T3n);
|
cannam@167
|
571 T3Q = FMA(KP303346683, T3G, T3N);
|
cannam@167
|
572 R1[WS(rs, 9)] = FNMS(KP1_913880671, T3Q, T3P);
|
cannam@167
|
573 R1[WS(rs, 25)] = FMA(KP1_913880671, T3Q, T3P);
|
cannam@167
|
574 T3T = FNMS(KP1_662939224, T3S, T3R);
|
cannam@167
|
575 T3W = FNMS(KP534511135, T3V, T3U);
|
cannam@167
|
576 R1[WS(rs, 5)] = FNMS(KP1_763842528, T3W, T3T);
|
cannam@167
|
577 R1[WS(rs, 21)] = FMA(KP1_763842528, T3W, T3T);
|
cannam@167
|
578 }
|
cannam@167
|
579 }
|
cannam@167
|
580 {
|
cannam@167
|
581 E T1n, T2L, T1O, T2M, T2u, T2P, T2H, T2O, T1E, T1N;
|
cannam@167
|
582 T1n = FMA(KP1_847759065, T1m, T1b);
|
cannam@167
|
583 T2L = FNMS(KP1_847759065, T1m, T1b);
|
cannam@167
|
584 T1E = FMA(KP707106781, T1D, T1s);
|
cannam@167
|
585 T1N = FMA(KP707106781, T1M, T1J);
|
cannam@167
|
586 T1O = FNMS(KP198912367, T1N, T1E);
|
cannam@167
|
587 T2M = FMA(KP198912367, T1E, T1N);
|
cannam@167
|
588 {
|
cannam@167
|
589 E T26, T2t, T2D, T2G;
|
cannam@167
|
590 T26 = FMA(KP707106781, T25, T1U);
|
cannam@167
|
591 T2t = T2h + T2s;
|
cannam@167
|
592 T2u = FMA(KP923879532, T2t, T26);
|
cannam@167
|
593 T2P = FNMS(KP923879532, T2t, T26);
|
cannam@167
|
594 T2D = FMA(KP707106781, T2C, T2z);
|
cannam@167
|
595 T2G = T2E + T2F;
|
cannam@167
|
596 T2H = FMA(KP923879532, T2G, T2D);
|
cannam@167
|
597 T2O = FNMS(KP923879532, T2G, T2D);
|
cannam@167
|
598 }
|
cannam@167
|
599 {
|
cannam@167
|
600 E T1P, T2I, T2R, T2S;
|
cannam@167
|
601 T1P = FMA(KP1_961570560, T1O, T1n);
|
cannam@167
|
602 T2I = FNMS(KP098491403, T2H, T2u);
|
cannam@167
|
603 R1[WS(rs, 16)] = FNMS(KP1_990369453, T2I, T1P);
|
cannam@167
|
604 R1[0] = FMA(KP1_990369453, T2I, T1P);
|
cannam@167
|
605 T2R = FMA(KP1_961570560, T2M, T2L);
|
cannam@167
|
606 T2S = FMA(KP820678790, T2O, T2P);
|
cannam@167
|
607 R1[WS(rs, 12)] = FNMS(KP1_546020906, T2S, T2R);
|
cannam@167
|
608 R1[WS(rs, 28)] = FMA(KP1_546020906, T2S, T2R);
|
cannam@167
|
609 }
|
cannam@167
|
610 {
|
cannam@167
|
611 E T2J, T2K, T2N, T2Q;
|
cannam@167
|
612 T2J = FNMS(KP1_961570560, T1O, T1n);
|
cannam@167
|
613 T2K = FMA(KP098491403, T2u, T2H);
|
cannam@167
|
614 R1[WS(rs, 8)] = FNMS(KP1_990369453, T2K, T2J);
|
cannam@167
|
615 R1[WS(rs, 24)] = FMA(KP1_990369453, T2K, T2J);
|
cannam@167
|
616 T2N = FNMS(KP1_961570560, T2M, T2L);
|
cannam@167
|
617 T2Q = FNMS(KP820678790, T2P, T2O);
|
cannam@167
|
618 R1[WS(rs, 4)] = FNMS(KP1_546020906, T2Q, T2N);
|
cannam@167
|
619 R1[WS(rs, 20)] = FMA(KP1_546020906, T2Q, T2N);
|
cannam@167
|
620 }
|
cannam@167
|
621 }
|
cannam@167
|
622 {
|
cannam@167
|
623 E T41, T4f, T44, T4g, T48, T4j, T4b, T4i, T42, T43;
|
cannam@167
|
624 T41 = FNMS(KP1_847759065, T40, T3Z);
|
cannam@167
|
625 T4f = FMA(KP1_847759065, T40, T3Z);
|
cannam@167
|
626 T42 = FMA(KP707106781, T3s, T3r);
|
cannam@167
|
627 T43 = FMA(KP707106781, T3p, T3o);
|
cannam@167
|
628 T44 = FNMS(KP198912367, T43, T42);
|
cannam@167
|
629 T4g = FMA(KP198912367, T42, T43);
|
cannam@167
|
630 {
|
cannam@167
|
631 E T46, T47, T49, T4a;
|
cannam@167
|
632 T46 = FMA(KP707106781, T3x, T3w);
|
cannam@167
|
633 T47 = T3K + T3L;
|
cannam@167
|
634 T48 = FNMS(KP923879532, T47, T46);
|
cannam@167
|
635 T4j = FMA(KP923879532, T47, T46);
|
cannam@167
|
636 T49 = FMA(KP707106781, T3I, T3H);
|
cannam@167
|
637 T4a = T3B - T3E;
|
cannam@167
|
638 T4b = FNMS(KP923879532, T4a, T49);
|
cannam@167
|
639 T4i = FMA(KP923879532, T4a, T49);
|
cannam@167
|
640 }
|
cannam@167
|
641 {
|
cannam@167
|
642 E T45, T4c, T4l, T4m;
|
cannam@167
|
643 T45 = FNMS(KP1_961570560, T44, T41);
|
cannam@167
|
644 T4c = FNMS(KP820678790, T4b, T48);
|
cannam@167
|
645 R1[WS(rs, 19)] = FNMS(KP1_546020906, T4c, T45);
|
cannam@167
|
646 R1[WS(rs, 3)] = FMA(KP1_546020906, T4c, T45);
|
cannam@167
|
647 T4l = FMA(KP1_961570560, T4g, T4f);
|
cannam@167
|
648 T4m = FMA(KP098491403, T4i, T4j);
|
cannam@167
|
649 R1[WS(rs, 15)] = FNMS(KP1_990369453, T4m, T4l);
|
cannam@167
|
650 R1[WS(rs, 31)] = FMA(KP1_990369453, T4m, T4l);
|
cannam@167
|
651 }
|
cannam@167
|
652 {
|
cannam@167
|
653 E T4d, T4e, T4h, T4k;
|
cannam@167
|
654 T4d = FMA(KP1_961570560, T44, T41);
|
cannam@167
|
655 T4e = FMA(KP820678790, T48, T4b);
|
cannam@167
|
656 R1[WS(rs, 11)] = FNMS(KP1_546020906, T4e, T4d);
|
cannam@167
|
657 R1[WS(rs, 27)] = FMA(KP1_546020906, T4e, T4d);
|
cannam@167
|
658 T4h = FNMS(KP1_961570560, T4g, T4f);
|
cannam@167
|
659 T4k = FNMS(KP098491403, T4j, T4i);
|
cannam@167
|
660 R1[WS(rs, 7)] = FNMS(KP1_990369453, T4k, T4h);
|
cannam@167
|
661 R1[WS(rs, 23)] = FMA(KP1_990369453, T4k, T4h);
|
cannam@167
|
662 }
|
cannam@167
|
663 }
|
cannam@167
|
664 {
|
cannam@167
|
665 E T2V, T39, T2Y, T3a, T32, T3d, T35, T3c, T2W, T2X;
|
cannam@167
|
666 T2V = FNMS(KP1_847759065, T2U, T2T);
|
cannam@167
|
667 T39 = FMA(KP1_847759065, T2U, T2T);
|
cannam@167
|
668 T2W = FNMS(KP707106781, T1M, T1J);
|
cannam@167
|
669 T2X = FNMS(KP707106781, T1D, T1s);
|
cannam@167
|
670 T2Y = FNMS(KP668178637, T2X, T2W);
|
cannam@167
|
671 T3a = FMA(KP668178637, T2W, T2X);
|
cannam@167
|
672 {
|
cannam@167
|
673 E T30, T31, T33, T34;
|
cannam@167
|
674 T30 = FNMS(KP707106781, T25, T1U);
|
cannam@167
|
675 T31 = T2E - T2F;
|
cannam@167
|
676 T32 = FNMS(KP923879532, T31, T30);
|
cannam@167
|
677 T3d = FMA(KP923879532, T31, T30);
|
cannam@167
|
678 T33 = FNMS(KP707106781, T2C, T2z);
|
cannam@167
|
679 T34 = T2s - T2h;
|
cannam@167
|
680 T35 = FNMS(KP923879532, T34, T33);
|
cannam@167
|
681 T3c = FMA(KP923879532, T34, T33);
|
cannam@167
|
682 }
|
cannam@167
|
683 {
|
cannam@167
|
684 E T2Z, T36, T3f, T3g;
|
cannam@167
|
685 T2Z = FNMS(KP1_662939224, T2Y, T2V);
|
cannam@167
|
686 T36 = FNMS(KP534511135, T35, T32);
|
cannam@167
|
687 R1[WS(rs, 18)] = FNMS(KP1_763842528, T36, T2Z);
|
cannam@167
|
688 R1[WS(rs, 2)] = FMA(KP1_763842528, T36, T2Z);
|
cannam@167
|
689 T3f = FMA(KP1_662939224, T3a, T39);
|
cannam@167
|
690 T3g = FMA(KP303346683, T3c, T3d);
|
cannam@167
|
691 R1[WS(rs, 14)] = FNMS(KP1_913880671, T3g, T3f);
|
cannam@167
|
692 R1[WS(rs, 30)] = FMA(KP1_913880671, T3g, T3f);
|
cannam@167
|
693 }
|
cannam@167
|
694 {
|
cannam@167
|
695 E T37, T38, T3b, T3e;
|
cannam@167
|
696 T37 = FMA(KP1_662939224, T2Y, T2V);
|
cannam@167
|
697 T38 = FMA(KP534511135, T32, T35);
|
cannam@167
|
698 R1[WS(rs, 10)] = FNMS(KP1_763842528, T38, T37);
|
cannam@167
|
699 R1[WS(rs, 26)] = FMA(KP1_763842528, T38, T37);
|
cannam@167
|
700 T3b = FNMS(KP1_662939224, T3a, T39);
|
cannam@167
|
701 T3e = FNMS(KP303346683, T3d, T3c);
|
cannam@167
|
702 R1[WS(rs, 6)] = FNMS(KP1_913880671, T3e, T3b);
|
cannam@167
|
703 R1[WS(rs, 22)] = FMA(KP1_913880671, T3e, T3b);
|
cannam@167
|
704 }
|
cannam@167
|
705 }
|
cannam@167
|
706 }
|
cannam@167
|
707 }
|
cannam@167
|
708 }
|
cannam@167
|
709
|
cannam@167
|
710 static const kr2c_desc desc = { 64, "r2cb_64", {178, 0, 216, 0}, &GENUS };
|
cannam@167
|
711
|
cannam@167
|
712 void X(codelet_r2cb_64) (planner *p) {
|
cannam@167
|
713 X(kr2c_register) (p, r2cb_64, &desc);
|
cannam@167
|
714 }
|
cannam@167
|
715
|
cannam@167
|
716 #else
|
cannam@167
|
717
|
cannam@167
|
718 /* Generated by: ../../../genfft/gen_r2cb.native -compact -variables 4 -pipeline-latency 4 -sign 1 -n 64 -name r2cb_64 -include rdft/scalar/r2cb.h */
|
cannam@167
|
719
|
cannam@167
|
720 /*
|
cannam@167
|
721 * This function contains 394 FP additions, 134 FP multiplications,
|
cannam@167
|
722 * (or, 342 additions, 82 multiplications, 52 fused multiply/add),
|
cannam@167
|
723 * 110 stack variables, 19 constants, and 128 memory accesses
|
cannam@167
|
724 */
|
cannam@167
|
725 #include "rdft/scalar/r2cb.h"
|
cannam@167
|
726
|
cannam@167
|
727 static void r2cb_64(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
|
cannam@167
|
728 {
|
cannam@167
|
729 DK(KP1_268786568, +1.268786568327290996430343226450986741351374190);
|
cannam@167
|
730 DK(KP1_546020906, +1.546020906725473921621813219516939601942082586);
|
cannam@167
|
731 DK(KP196034280, +0.196034280659121203988391127777283691722273346);
|
cannam@167
|
732 DK(KP1_990369453, +1.990369453344393772489673906218959843150949737);
|
cannam@167
|
733 DK(KP942793473, +0.942793473651995297112775251810508755314920638);
|
cannam@167
|
734 DK(KP1_763842528, +1.763842528696710059425513727320776699016885241);
|
cannam@167
|
735 DK(KP580569354, +0.580569354508924735272384751634790549382952557);
|
cannam@167
|
736 DK(KP1_913880671, +1.913880671464417729871595773960539938965698411);
|
cannam@167
|
737 DK(KP1_111140466, +1.111140466039204449485661627897065748749874382);
|
cannam@167
|
738 DK(KP1_662939224, +1.662939224605090474157576755235811513477121624);
|
cannam@167
|
739 DK(KP390180644, +0.390180644032256535696569736954044481855383236);
|
cannam@167
|
740 DK(KP1_961570560, +1.961570560806460898252364472268478073947867462);
|
cannam@167
|
741 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
cannam@167
|
742 DK(KP382683432, +0.382683432365089771728459984030398866761344562);
|
cannam@167
|
743 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
cannam@167
|
744 DK(KP765366864, +0.765366864730179543456919968060797733522689125);
|
cannam@167
|
745 DK(KP1_847759065, +1.847759065022573512256366378793576573644833252);
|
cannam@167
|
746 DK(KP1_414213562, +1.414213562373095048801688724209698078569671875);
|
cannam@167
|
747 DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
|
cannam@167
|
748 {
|
cannam@167
|
749 INT i;
|
cannam@167
|
750 for (i = v; i > 0; i = i - 1, R0 = R0 + ovs, R1 = R1 + ovs, Cr = Cr + ivs, Ci = Ci + ivs, MAKE_VOLATILE_STRIDE(256, rs), MAKE_VOLATILE_STRIDE(256, csr), MAKE_VOLATILE_STRIDE(256, csi)) {
|
cannam@167
|
751 E Ta, T2S, T18, T2u, T3F, T4V, T5l, T61, Th, T2T, T1h, T2v, T3M, T4W, T5o;
|
cannam@167
|
752 E T62, T3Q, T5q, T5u, T44, Tp, Tw, T2V, T2W, T2X, T2Y, T3X, T5t, T1r, T2x;
|
cannam@167
|
753 E T41, T5r, T1A, T2y, T4a, T5y, T5N, T4H, TN, T31, T4E, T5z, T39, T3q, T1L;
|
cannam@167
|
754 E T2B, T4h, T5M, T2h, T2F, T12, T36, T5D, T5J, T5G, T5K, T1U, T26, T23, T27;
|
cannam@167
|
755 E T4p, T4z, T4w, T4A, T34, T3r;
|
cannam@167
|
756 {
|
cannam@167
|
757 E T5, T3A, T3, T3y, T9, T3C, T17, T3D, T6, T14;
|
cannam@167
|
758 {
|
cannam@167
|
759 E T4, T3z, T1, T2;
|
cannam@167
|
760 T4 = Cr[WS(csr, 16)];
|
cannam@167
|
761 T5 = KP2_000000000 * T4;
|
cannam@167
|
762 T3z = Ci[WS(csi, 16)];
|
cannam@167
|
763 T3A = KP2_000000000 * T3z;
|
cannam@167
|
764 T1 = Cr[0];
|
cannam@167
|
765 T2 = Cr[WS(csr, 32)];
|
cannam@167
|
766 T3 = T1 + T2;
|
cannam@167
|
767 T3y = T1 - T2;
|
cannam@167
|
768 {
|
cannam@167
|
769 E T7, T8, T15, T16;
|
cannam@167
|
770 T7 = Cr[WS(csr, 8)];
|
cannam@167
|
771 T8 = Cr[WS(csr, 24)];
|
cannam@167
|
772 T9 = KP2_000000000 * (T7 + T8);
|
cannam@167
|
773 T3C = T7 - T8;
|
cannam@167
|
774 T15 = Ci[WS(csi, 8)];
|
cannam@167
|
775 T16 = Ci[WS(csi, 24)];
|
cannam@167
|
776 T17 = KP2_000000000 * (T15 - T16);
|
cannam@167
|
777 T3D = T15 + T16;
|
cannam@167
|
778 }
|
cannam@167
|
779 }
|
cannam@167
|
780 T6 = T3 + T5;
|
cannam@167
|
781 Ta = T6 + T9;
|
cannam@167
|
782 T2S = T6 - T9;
|
cannam@167
|
783 T14 = T3 - T5;
|
cannam@167
|
784 T18 = T14 - T17;
|
cannam@167
|
785 T2u = T14 + T17;
|
cannam@167
|
786 {
|
cannam@167
|
787 E T3B, T3E, T5j, T5k;
|
cannam@167
|
788 T3B = T3y - T3A;
|
cannam@167
|
789 T3E = KP1_414213562 * (T3C - T3D);
|
cannam@167
|
790 T3F = T3B + T3E;
|
cannam@167
|
791 T4V = T3B - T3E;
|
cannam@167
|
792 T5j = T3y + T3A;
|
cannam@167
|
793 T5k = KP1_414213562 * (T3C + T3D);
|
cannam@167
|
794 T5l = T5j - T5k;
|
cannam@167
|
795 T61 = T5j + T5k;
|
cannam@167
|
796 }
|
cannam@167
|
797 }
|
cannam@167
|
798 {
|
cannam@167
|
799 E Td, T3G, T1c, T3K, Tg, T3J, T1f, T3H, T19, T1g;
|
cannam@167
|
800 {
|
cannam@167
|
801 E Tb, Tc, T1a, T1b;
|
cannam@167
|
802 Tb = Cr[WS(csr, 4)];
|
cannam@167
|
803 Tc = Cr[WS(csr, 28)];
|
cannam@167
|
804 Td = Tb + Tc;
|
cannam@167
|
805 T3G = Tb - Tc;
|
cannam@167
|
806 T1a = Ci[WS(csi, 4)];
|
cannam@167
|
807 T1b = Ci[WS(csi, 28)];
|
cannam@167
|
808 T1c = T1a - T1b;
|
cannam@167
|
809 T3K = T1a + T1b;
|
cannam@167
|
810 }
|
cannam@167
|
811 {
|
cannam@167
|
812 E Te, Tf, T1d, T1e;
|
cannam@167
|
813 Te = Cr[WS(csr, 20)];
|
cannam@167
|
814 Tf = Cr[WS(csr, 12)];
|
cannam@167
|
815 Tg = Te + Tf;
|
cannam@167
|
816 T3J = Te - Tf;
|
cannam@167
|
817 T1d = Ci[WS(csi, 20)];
|
cannam@167
|
818 T1e = Ci[WS(csi, 12)];
|
cannam@167
|
819 T1f = T1d - T1e;
|
cannam@167
|
820 T3H = T1d + T1e;
|
cannam@167
|
821 }
|
cannam@167
|
822 Th = KP2_000000000 * (Td + Tg);
|
cannam@167
|
823 T2T = KP2_000000000 * (T1f + T1c);
|
cannam@167
|
824 T19 = Td - Tg;
|
cannam@167
|
825 T1g = T1c - T1f;
|
cannam@167
|
826 T1h = KP1_414213562 * (T19 - T1g);
|
cannam@167
|
827 T2v = KP1_414213562 * (T19 + T1g);
|
cannam@167
|
828 {
|
cannam@167
|
829 E T3I, T3L, T5m, T5n;
|
cannam@167
|
830 T3I = T3G - T3H;
|
cannam@167
|
831 T3L = T3J + T3K;
|
cannam@167
|
832 T3M = FNMS(KP765366864, T3L, KP1_847759065 * T3I);
|
cannam@167
|
833 T4W = FMA(KP765366864, T3I, KP1_847759065 * T3L);
|
cannam@167
|
834 T5m = T3G + T3H;
|
cannam@167
|
835 T5n = T3K - T3J;
|
cannam@167
|
836 T5o = FNMS(KP1_847759065, T5n, KP765366864 * T5m);
|
cannam@167
|
837 T62 = FMA(KP1_847759065, T5m, KP765366864 * T5n);
|
cannam@167
|
838 }
|
cannam@167
|
839 }
|
cannam@167
|
840 {
|
cannam@167
|
841 E Tl, T3O, T1v, T43, To, T42, T1y, T3P, Ts, T3R, T1p, T3S, Tv, T3U, T1m;
|
cannam@167
|
842 E T3V;
|
cannam@167
|
843 {
|
cannam@167
|
844 E Tj, Tk, T1t, T1u;
|
cannam@167
|
845 Tj = Cr[WS(csr, 2)];
|
cannam@167
|
846 Tk = Cr[WS(csr, 30)];
|
cannam@167
|
847 Tl = Tj + Tk;
|
cannam@167
|
848 T3O = Tj - Tk;
|
cannam@167
|
849 T1t = Ci[WS(csi, 2)];
|
cannam@167
|
850 T1u = Ci[WS(csi, 30)];
|
cannam@167
|
851 T1v = T1t - T1u;
|
cannam@167
|
852 T43 = T1t + T1u;
|
cannam@167
|
853 }
|
cannam@167
|
854 {
|
cannam@167
|
855 E Tm, Tn, T1w, T1x;
|
cannam@167
|
856 Tm = Cr[WS(csr, 18)];
|
cannam@167
|
857 Tn = Cr[WS(csr, 14)];
|
cannam@167
|
858 To = Tm + Tn;
|
cannam@167
|
859 T42 = Tm - Tn;
|
cannam@167
|
860 T1w = Ci[WS(csi, 18)];
|
cannam@167
|
861 T1x = Ci[WS(csi, 14)];
|
cannam@167
|
862 T1y = T1w - T1x;
|
cannam@167
|
863 T3P = T1w + T1x;
|
cannam@167
|
864 }
|
cannam@167
|
865 {
|
cannam@167
|
866 E Tq, Tr, T1n, T1o;
|
cannam@167
|
867 Tq = Cr[WS(csr, 10)];
|
cannam@167
|
868 Tr = Cr[WS(csr, 22)];
|
cannam@167
|
869 Ts = Tq + Tr;
|
cannam@167
|
870 T3R = Tq - Tr;
|
cannam@167
|
871 T1n = Ci[WS(csi, 10)];
|
cannam@167
|
872 T1o = Ci[WS(csi, 22)];
|
cannam@167
|
873 T1p = T1n - T1o;
|
cannam@167
|
874 T3S = T1n + T1o;
|
cannam@167
|
875 }
|
cannam@167
|
876 {
|
cannam@167
|
877 E Tt, Tu, T1k, T1l;
|
cannam@167
|
878 Tt = Cr[WS(csr, 6)];
|
cannam@167
|
879 Tu = Cr[WS(csr, 26)];
|
cannam@167
|
880 Tv = Tt + Tu;
|
cannam@167
|
881 T3U = Tt - Tu;
|
cannam@167
|
882 T1k = Ci[WS(csi, 26)];
|
cannam@167
|
883 T1l = Ci[WS(csi, 6)];
|
cannam@167
|
884 T1m = T1k - T1l;
|
cannam@167
|
885 T3V = T1l + T1k;
|
cannam@167
|
886 }
|
cannam@167
|
887 T3Q = T3O - T3P;
|
cannam@167
|
888 T5q = T3O + T3P;
|
cannam@167
|
889 T5u = T43 - T42;
|
cannam@167
|
890 T44 = T42 + T43;
|
cannam@167
|
891 Tp = Tl + To;
|
cannam@167
|
892 Tw = Ts + Tv;
|
cannam@167
|
893 T2V = Tp - Tw;
|
cannam@167
|
894 {
|
cannam@167
|
895 E T3T, T3W, T1j, T1q;
|
cannam@167
|
896 T2W = T1y + T1v;
|
cannam@167
|
897 T2X = T1p + T1m;
|
cannam@167
|
898 T2Y = T2W - T2X;
|
cannam@167
|
899 T3T = T3R - T3S;
|
cannam@167
|
900 T3W = T3U - T3V;
|
cannam@167
|
901 T3X = KP707106781 * (T3T + T3W);
|
cannam@167
|
902 T5t = KP707106781 * (T3T - T3W);
|
cannam@167
|
903 T1j = Tl - To;
|
cannam@167
|
904 T1q = T1m - T1p;
|
cannam@167
|
905 T1r = T1j + T1q;
|
cannam@167
|
906 T2x = T1j - T1q;
|
cannam@167
|
907 {
|
cannam@167
|
908 E T3Z, T40, T1s, T1z;
|
cannam@167
|
909 T3Z = T3R + T3S;
|
cannam@167
|
910 T40 = T3U + T3V;
|
cannam@167
|
911 T41 = KP707106781 * (T3Z - T40);
|
cannam@167
|
912 T5r = KP707106781 * (T3Z + T40);
|
cannam@167
|
913 T1s = Ts - Tv;
|
cannam@167
|
914 T1z = T1v - T1y;
|
cannam@167
|
915 T1A = T1s + T1z;
|
cannam@167
|
916 T2y = T1z - T1s;
|
cannam@167
|
917 }
|
cannam@167
|
918 }
|
cannam@167
|
919 }
|
cannam@167
|
920 {
|
cannam@167
|
921 E TB, T48, T2c, T4G, TE, T4F, T2f, T49, TI, T4b, T1J, T4c, TL, T4e, T1G;
|
cannam@167
|
922 E T4f;
|
cannam@167
|
923 {
|
cannam@167
|
924 E Tz, TA, T2a, T2b;
|
cannam@167
|
925 Tz = Cr[WS(csr, 1)];
|
cannam@167
|
926 TA = Cr[WS(csr, 31)];
|
cannam@167
|
927 TB = Tz + TA;
|
cannam@167
|
928 T48 = Tz - TA;
|
cannam@167
|
929 T2a = Ci[WS(csi, 1)];
|
cannam@167
|
930 T2b = Ci[WS(csi, 31)];
|
cannam@167
|
931 T2c = T2a - T2b;
|
cannam@167
|
932 T4G = T2a + T2b;
|
cannam@167
|
933 }
|
cannam@167
|
934 {
|
cannam@167
|
935 E TC, TD, T2d, T2e;
|
cannam@167
|
936 TC = Cr[WS(csr, 17)];
|
cannam@167
|
937 TD = Cr[WS(csr, 15)];
|
cannam@167
|
938 TE = TC + TD;
|
cannam@167
|
939 T4F = TC - TD;
|
cannam@167
|
940 T2d = Ci[WS(csi, 17)];
|
cannam@167
|
941 T2e = Ci[WS(csi, 15)];
|
cannam@167
|
942 T2f = T2d - T2e;
|
cannam@167
|
943 T49 = T2d + T2e;
|
cannam@167
|
944 }
|
cannam@167
|
945 {
|
cannam@167
|
946 E TG, TH, T1H, T1I;
|
cannam@167
|
947 TG = Cr[WS(csr, 9)];
|
cannam@167
|
948 TH = Cr[WS(csr, 23)];
|
cannam@167
|
949 TI = TG + TH;
|
cannam@167
|
950 T4b = TG - TH;
|
cannam@167
|
951 T1H = Ci[WS(csi, 9)];
|
cannam@167
|
952 T1I = Ci[WS(csi, 23)];
|
cannam@167
|
953 T1J = T1H - T1I;
|
cannam@167
|
954 T4c = T1H + T1I;
|
cannam@167
|
955 }
|
cannam@167
|
956 {
|
cannam@167
|
957 E TJ, TK, T1E, T1F;
|
cannam@167
|
958 TJ = Cr[WS(csr, 7)];
|
cannam@167
|
959 TK = Cr[WS(csr, 25)];
|
cannam@167
|
960 TL = TJ + TK;
|
cannam@167
|
961 T4e = TJ - TK;
|
cannam@167
|
962 T1E = Ci[WS(csi, 25)];
|
cannam@167
|
963 T1F = Ci[WS(csi, 7)];
|
cannam@167
|
964 T1G = T1E - T1F;
|
cannam@167
|
965 T4f = T1F + T1E;
|
cannam@167
|
966 }
|
cannam@167
|
967 {
|
cannam@167
|
968 E TF, TM, T1D, T1K;
|
cannam@167
|
969 T4a = T48 - T49;
|
cannam@167
|
970 T5y = T48 + T49;
|
cannam@167
|
971 T5N = T4G - T4F;
|
cannam@167
|
972 T4H = T4F + T4G;
|
cannam@167
|
973 TF = TB + TE;
|
cannam@167
|
974 TM = TI + TL;
|
cannam@167
|
975 TN = TF + TM;
|
cannam@167
|
976 T31 = TF - TM;
|
cannam@167
|
977 {
|
cannam@167
|
978 E T4C, T4D, T37, T38;
|
cannam@167
|
979 T4C = T4b + T4c;
|
cannam@167
|
980 T4D = T4e + T4f;
|
cannam@167
|
981 T4E = KP707106781 * (T4C - T4D);
|
cannam@167
|
982 T5z = KP707106781 * (T4C + T4D);
|
cannam@167
|
983 T37 = T2f + T2c;
|
cannam@167
|
984 T38 = T1J + T1G;
|
cannam@167
|
985 T39 = T37 - T38;
|
cannam@167
|
986 T3q = T38 + T37;
|
cannam@167
|
987 }
|
cannam@167
|
988 T1D = TB - TE;
|
cannam@167
|
989 T1K = T1G - T1J;
|
cannam@167
|
990 T1L = T1D + T1K;
|
cannam@167
|
991 T2B = T1D - T1K;
|
cannam@167
|
992 {
|
cannam@167
|
993 E T4d, T4g, T29, T2g;
|
cannam@167
|
994 T4d = T4b - T4c;
|
cannam@167
|
995 T4g = T4e - T4f;
|
cannam@167
|
996 T4h = KP707106781 * (T4d + T4g);
|
cannam@167
|
997 T5M = KP707106781 * (T4d - T4g);
|
cannam@167
|
998 T29 = TI - TL;
|
cannam@167
|
999 T2g = T2c - T2f;
|
cannam@167
|
1000 T2h = T29 + T2g;
|
cannam@167
|
1001 T2F = T2g - T29;
|
cannam@167
|
1002 }
|
cannam@167
|
1003 }
|
cannam@167
|
1004 }
|
cannam@167
|
1005 {
|
cannam@167
|
1006 E TQ, T4j, T1P, T4n, TT, T4m, T1S, T4k, TX, T4q, T1Y, T4u, T10, T4t, T21;
|
cannam@167
|
1007 E T4r;
|
cannam@167
|
1008 {
|
cannam@167
|
1009 E TO, TP, T1N, T1O;
|
cannam@167
|
1010 TO = Cr[WS(csr, 5)];
|
cannam@167
|
1011 TP = Cr[WS(csr, 27)];
|
cannam@167
|
1012 TQ = TO + TP;
|
cannam@167
|
1013 T4j = TO - TP;
|
cannam@167
|
1014 T1N = Ci[WS(csi, 5)];
|
cannam@167
|
1015 T1O = Ci[WS(csi, 27)];
|
cannam@167
|
1016 T1P = T1N - T1O;
|
cannam@167
|
1017 T4n = T1N + T1O;
|
cannam@167
|
1018 }
|
cannam@167
|
1019 {
|
cannam@167
|
1020 E TR, TS, T1Q, T1R;
|
cannam@167
|
1021 TR = Cr[WS(csr, 21)];
|
cannam@167
|
1022 TS = Cr[WS(csr, 11)];
|
cannam@167
|
1023 TT = TR + TS;
|
cannam@167
|
1024 T4m = TR - TS;
|
cannam@167
|
1025 T1Q = Ci[WS(csi, 21)];
|
cannam@167
|
1026 T1R = Ci[WS(csi, 11)];
|
cannam@167
|
1027 T1S = T1Q - T1R;
|
cannam@167
|
1028 T4k = T1Q + T1R;
|
cannam@167
|
1029 }
|
cannam@167
|
1030 {
|
cannam@167
|
1031 E TV, TW, T1W, T1X;
|
cannam@167
|
1032 TV = Cr[WS(csr, 3)];
|
cannam@167
|
1033 TW = Cr[WS(csr, 29)];
|
cannam@167
|
1034 TX = TV + TW;
|
cannam@167
|
1035 T4q = TV - TW;
|
cannam@167
|
1036 T1W = Ci[WS(csi, 29)];
|
cannam@167
|
1037 T1X = Ci[WS(csi, 3)];
|
cannam@167
|
1038 T1Y = T1W - T1X;
|
cannam@167
|
1039 T4u = T1X + T1W;
|
cannam@167
|
1040 }
|
cannam@167
|
1041 {
|
cannam@167
|
1042 E TY, TZ, T1Z, T20;
|
cannam@167
|
1043 TY = Cr[WS(csr, 13)];
|
cannam@167
|
1044 TZ = Cr[WS(csr, 19)];
|
cannam@167
|
1045 T10 = TY + TZ;
|
cannam@167
|
1046 T4t = TY - TZ;
|
cannam@167
|
1047 T1Z = Ci[WS(csi, 13)];
|
cannam@167
|
1048 T20 = Ci[WS(csi, 19)];
|
cannam@167
|
1049 T21 = T1Z - T20;
|
cannam@167
|
1050 T4r = T1Z + T20;
|
cannam@167
|
1051 }
|
cannam@167
|
1052 {
|
cannam@167
|
1053 E TU, T11, T5B, T5C;
|
cannam@167
|
1054 TU = TQ + TT;
|
cannam@167
|
1055 T11 = TX + T10;
|
cannam@167
|
1056 T12 = TU + T11;
|
cannam@167
|
1057 T36 = TU - T11;
|
cannam@167
|
1058 T5B = T4j + T4k;
|
cannam@167
|
1059 T5C = T4n - T4m;
|
cannam@167
|
1060 T5D = FNMS(KP923879532, T5C, KP382683432 * T5B);
|
cannam@167
|
1061 T5J = FMA(KP923879532, T5B, KP382683432 * T5C);
|
cannam@167
|
1062 }
|
cannam@167
|
1063 {
|
cannam@167
|
1064 E T5E, T5F, T1M, T1T;
|
cannam@167
|
1065 T5E = T4q + T4r;
|
cannam@167
|
1066 T5F = T4t + T4u;
|
cannam@167
|
1067 T5G = FNMS(KP923879532, T5F, KP382683432 * T5E);
|
cannam@167
|
1068 T5K = FMA(KP923879532, T5E, KP382683432 * T5F);
|
cannam@167
|
1069 T1M = TQ - TT;
|
cannam@167
|
1070 T1T = T1P - T1S;
|
cannam@167
|
1071 T1U = T1M - T1T;
|
cannam@167
|
1072 T26 = T1M + T1T;
|
cannam@167
|
1073 }
|
cannam@167
|
1074 {
|
cannam@167
|
1075 E T1V, T22, T4l, T4o;
|
cannam@167
|
1076 T1V = TX - T10;
|
cannam@167
|
1077 T22 = T1Y - T21;
|
cannam@167
|
1078 T23 = T1V + T22;
|
cannam@167
|
1079 T27 = T22 - T1V;
|
cannam@167
|
1080 T4l = T4j - T4k;
|
cannam@167
|
1081 T4o = T4m + T4n;
|
cannam@167
|
1082 T4p = FNMS(KP382683432, T4o, KP923879532 * T4l);
|
cannam@167
|
1083 T4z = FMA(KP382683432, T4l, KP923879532 * T4o);
|
cannam@167
|
1084 }
|
cannam@167
|
1085 {
|
cannam@167
|
1086 E T4s, T4v, T32, T33;
|
cannam@167
|
1087 T4s = T4q - T4r;
|
cannam@167
|
1088 T4v = T4t - T4u;
|
cannam@167
|
1089 T4w = FMA(KP923879532, T4s, KP382683432 * T4v);
|
cannam@167
|
1090 T4A = FNMS(KP382683432, T4s, KP923879532 * T4v);
|
cannam@167
|
1091 T32 = T21 + T1Y;
|
cannam@167
|
1092 T33 = T1S + T1P;
|
cannam@167
|
1093 T34 = T32 - T33;
|
cannam@167
|
1094 T3r = T33 + T32;
|
cannam@167
|
1095 }
|
cannam@167
|
1096 }
|
cannam@167
|
1097 {
|
cannam@167
|
1098 E T13, T3x, Ty, T3w, Ti, Tx;
|
cannam@167
|
1099 T13 = KP2_000000000 * (TN + T12);
|
cannam@167
|
1100 T3x = KP2_000000000 * (T3r + T3q);
|
cannam@167
|
1101 Ti = Ta + Th;
|
cannam@167
|
1102 Tx = KP2_000000000 * (Tp + Tw);
|
cannam@167
|
1103 Ty = Ti + Tx;
|
cannam@167
|
1104 T3w = Ti - Tx;
|
cannam@167
|
1105 R0[WS(rs, 16)] = Ty - T13;
|
cannam@167
|
1106 R0[WS(rs, 24)] = T3w + T3x;
|
cannam@167
|
1107 R0[0] = Ty + T13;
|
cannam@167
|
1108 R0[WS(rs, 8)] = T3w - T3x;
|
cannam@167
|
1109 }
|
cannam@167
|
1110 {
|
cannam@167
|
1111 E T3g, T3k, T3j, T3l;
|
cannam@167
|
1112 {
|
cannam@167
|
1113 E T3e, T3f, T3h, T3i;
|
cannam@167
|
1114 T3e = T2S + T2T;
|
cannam@167
|
1115 T3f = KP1_414213562 * (T2V + T2Y);
|
cannam@167
|
1116 T3g = T3e - T3f;
|
cannam@167
|
1117 T3k = T3e + T3f;
|
cannam@167
|
1118 T3h = T31 - T34;
|
cannam@167
|
1119 T3i = T39 - T36;
|
cannam@167
|
1120 T3j = FNMS(KP1_847759065, T3i, KP765366864 * T3h);
|
cannam@167
|
1121 T3l = FMA(KP1_847759065, T3h, KP765366864 * T3i);
|
cannam@167
|
1122 }
|
cannam@167
|
1123 R0[WS(rs, 22)] = T3g - T3j;
|
cannam@167
|
1124 R0[WS(rs, 30)] = T3k + T3l;
|
cannam@167
|
1125 R0[WS(rs, 6)] = T3g + T3j;
|
cannam@167
|
1126 R0[WS(rs, 14)] = T3k - T3l;
|
cannam@167
|
1127 }
|
cannam@167
|
1128 {
|
cannam@167
|
1129 E T3o, T3u, T3t, T3v;
|
cannam@167
|
1130 {
|
cannam@167
|
1131 E T3m, T3n, T3p, T3s;
|
cannam@167
|
1132 T3m = Ta - Th;
|
cannam@167
|
1133 T3n = KP2_000000000 * (T2X + T2W);
|
cannam@167
|
1134 T3o = T3m - T3n;
|
cannam@167
|
1135 T3u = T3m + T3n;
|
cannam@167
|
1136 T3p = TN - T12;
|
cannam@167
|
1137 T3s = T3q - T3r;
|
cannam@167
|
1138 T3t = KP1_414213562 * (T3p - T3s);
|
cannam@167
|
1139 T3v = KP1_414213562 * (T3p + T3s);
|
cannam@167
|
1140 }
|
cannam@167
|
1141 R0[WS(rs, 20)] = T3o - T3t;
|
cannam@167
|
1142 R0[WS(rs, 28)] = T3u + T3v;
|
cannam@167
|
1143 R0[WS(rs, 4)] = T3o + T3t;
|
cannam@167
|
1144 R0[WS(rs, 12)] = T3u - T3v;
|
cannam@167
|
1145 }
|
cannam@167
|
1146 {
|
cannam@167
|
1147 E T30, T3c, T3b, T3d;
|
cannam@167
|
1148 {
|
cannam@167
|
1149 E T2U, T2Z, T35, T3a;
|
cannam@167
|
1150 T2U = T2S - T2T;
|
cannam@167
|
1151 T2Z = KP1_414213562 * (T2V - T2Y);
|
cannam@167
|
1152 T30 = T2U + T2Z;
|
cannam@167
|
1153 T3c = T2U - T2Z;
|
cannam@167
|
1154 T35 = T31 + T34;
|
cannam@167
|
1155 T3a = T36 + T39;
|
cannam@167
|
1156 T3b = FNMS(KP765366864, T3a, KP1_847759065 * T35);
|
cannam@167
|
1157 T3d = FMA(KP765366864, T35, KP1_847759065 * T3a);
|
cannam@167
|
1158 }
|
cannam@167
|
1159 R0[WS(rs, 18)] = T30 - T3b;
|
cannam@167
|
1160 R0[WS(rs, 26)] = T3c + T3d;
|
cannam@167
|
1161 R0[WS(rs, 2)] = T30 + T3b;
|
cannam@167
|
1162 R0[WS(rs, 10)] = T3c - T3d;
|
cannam@167
|
1163 }
|
cannam@167
|
1164 {
|
cannam@167
|
1165 E T25, T2p, T2i, T2q, T1C, T2k, T2o, T2s, T24, T28;
|
cannam@167
|
1166 T24 = KP707106781 * (T1U + T23);
|
cannam@167
|
1167 T25 = T1L + T24;
|
cannam@167
|
1168 T2p = T1L - T24;
|
cannam@167
|
1169 T28 = KP707106781 * (T26 + T27);
|
cannam@167
|
1170 T2i = T28 + T2h;
|
cannam@167
|
1171 T2q = T2h - T28;
|
cannam@167
|
1172 {
|
cannam@167
|
1173 E T1i, T1B, T2m, T2n;
|
cannam@167
|
1174 T1i = T18 + T1h;
|
cannam@167
|
1175 T1B = FNMS(KP765366864, T1A, KP1_847759065 * T1r);
|
cannam@167
|
1176 T1C = T1i + T1B;
|
cannam@167
|
1177 T2k = T1i - T1B;
|
cannam@167
|
1178 T2m = T18 - T1h;
|
cannam@167
|
1179 T2n = FMA(KP765366864, T1r, KP1_847759065 * T1A);
|
cannam@167
|
1180 T2o = T2m - T2n;
|
cannam@167
|
1181 T2s = T2m + T2n;
|
cannam@167
|
1182 }
|
cannam@167
|
1183 {
|
cannam@167
|
1184 E T2j, T2t, T2l, T2r;
|
cannam@167
|
1185 T2j = FNMS(KP390180644, T2i, KP1_961570560 * T25);
|
cannam@167
|
1186 R0[WS(rs, 17)] = T1C - T2j;
|
cannam@167
|
1187 R0[WS(rs, 1)] = T1C + T2j;
|
cannam@167
|
1188 T2t = FMA(KP1_662939224, T2p, KP1_111140466 * T2q);
|
cannam@167
|
1189 R0[WS(rs, 13)] = T2s - T2t;
|
cannam@167
|
1190 R0[WS(rs, 29)] = T2s + T2t;
|
cannam@167
|
1191 T2l = FMA(KP390180644, T25, KP1_961570560 * T2i);
|
cannam@167
|
1192 R0[WS(rs, 9)] = T2k - T2l;
|
cannam@167
|
1193 R0[WS(rs, 25)] = T2k + T2l;
|
cannam@167
|
1194 T2r = FNMS(KP1_662939224, T2q, KP1_111140466 * T2p);
|
cannam@167
|
1195 R0[WS(rs, 21)] = T2o - T2r;
|
cannam@167
|
1196 R0[WS(rs, 5)] = T2o + T2r;
|
cannam@167
|
1197 }
|
cannam@167
|
1198 }
|
cannam@167
|
1199 {
|
cannam@167
|
1200 E T2D, T2N, T2G, T2O, T2A, T2I, T2M, T2Q, T2C, T2E;
|
cannam@167
|
1201 T2C = KP707106781 * (T27 - T26);
|
cannam@167
|
1202 T2D = T2B + T2C;
|
cannam@167
|
1203 T2N = T2B - T2C;
|
cannam@167
|
1204 T2E = KP707106781 * (T1U - T23);
|
cannam@167
|
1205 T2G = T2E + T2F;
|
cannam@167
|
1206 T2O = T2F - T2E;
|
cannam@167
|
1207 {
|
cannam@167
|
1208 E T2w, T2z, T2K, T2L;
|
cannam@167
|
1209 T2w = T2u - T2v;
|
cannam@167
|
1210 T2z = FNMS(KP1_847759065, T2y, KP765366864 * T2x);
|
cannam@167
|
1211 T2A = T2w + T2z;
|
cannam@167
|
1212 T2I = T2w - T2z;
|
cannam@167
|
1213 T2K = T2u + T2v;
|
cannam@167
|
1214 T2L = FMA(KP1_847759065, T2x, KP765366864 * T2y);
|
cannam@167
|
1215 T2M = T2K - T2L;
|
cannam@167
|
1216 T2Q = T2K + T2L;
|
cannam@167
|
1217 }
|
cannam@167
|
1218 {
|
cannam@167
|
1219 E T2H, T2R, T2J, T2P;
|
cannam@167
|
1220 T2H = FNMS(KP1_111140466, T2G, KP1_662939224 * T2D);
|
cannam@167
|
1221 R0[WS(rs, 19)] = T2A - T2H;
|
cannam@167
|
1222 R0[WS(rs, 3)] = T2A + T2H;
|
cannam@167
|
1223 T2R = FMA(KP1_961570560, T2N, KP390180644 * T2O);
|
cannam@167
|
1224 R0[WS(rs, 15)] = T2Q - T2R;
|
cannam@167
|
1225 R0[WS(rs, 31)] = T2Q + T2R;
|
cannam@167
|
1226 T2J = FMA(KP1_111140466, T2D, KP1_662939224 * T2G);
|
cannam@167
|
1227 R0[WS(rs, 11)] = T2I - T2J;
|
cannam@167
|
1228 R0[WS(rs, 27)] = T2I + T2J;
|
cannam@167
|
1229 T2P = FNMS(KP1_961570560, T2O, KP390180644 * T2N);
|
cannam@167
|
1230 R0[WS(rs, 23)] = T2M - T2P;
|
cannam@167
|
1231 R0[WS(rs, 7)] = T2M + T2P;
|
cannam@167
|
1232 }
|
cannam@167
|
1233 }
|
cannam@167
|
1234 {
|
cannam@167
|
1235 E T5p, T5T, T5w, T5U, T5I, T5W, T5P, T5X, T5s, T5v;
|
cannam@167
|
1236 T5p = T5l + T5o;
|
cannam@167
|
1237 T5T = T5l - T5o;
|
cannam@167
|
1238 T5s = T5q - T5r;
|
cannam@167
|
1239 T5v = T5t + T5u;
|
cannam@167
|
1240 T5w = FNMS(KP1_111140466, T5v, KP1_662939224 * T5s);
|
cannam@167
|
1241 T5U = FMA(KP1_111140466, T5s, KP1_662939224 * T5v);
|
cannam@167
|
1242 {
|
cannam@167
|
1243 E T5A, T5H, T5L, T5O;
|
cannam@167
|
1244 T5A = T5y - T5z;
|
cannam@167
|
1245 T5H = T5D + T5G;
|
cannam@167
|
1246 T5I = T5A + T5H;
|
cannam@167
|
1247 T5W = T5A - T5H;
|
cannam@167
|
1248 T5L = T5J - T5K;
|
cannam@167
|
1249 T5O = T5M + T5N;
|
cannam@167
|
1250 T5P = T5L + T5O;
|
cannam@167
|
1251 T5X = T5O - T5L;
|
cannam@167
|
1252 }
|
cannam@167
|
1253 {
|
cannam@167
|
1254 E T5x, T5Q, T5Z, T60;
|
cannam@167
|
1255 T5x = T5p + T5w;
|
cannam@167
|
1256 T5Q = FNMS(KP580569354, T5P, KP1_913880671 * T5I);
|
cannam@167
|
1257 R1[WS(rs, 17)] = T5x - T5Q;
|
cannam@167
|
1258 R1[WS(rs, 1)] = T5x + T5Q;
|
cannam@167
|
1259 T5Z = T5T + T5U;
|
cannam@167
|
1260 T60 = FMA(KP1_763842528, T5W, KP942793473 * T5X);
|
cannam@167
|
1261 R1[WS(rs, 13)] = T5Z - T60;
|
cannam@167
|
1262 R1[WS(rs, 29)] = T5Z + T60;
|
cannam@167
|
1263 }
|
cannam@167
|
1264 {
|
cannam@167
|
1265 E T5R, T5S, T5V, T5Y;
|
cannam@167
|
1266 T5R = T5p - T5w;
|
cannam@167
|
1267 T5S = FMA(KP580569354, T5I, KP1_913880671 * T5P);
|
cannam@167
|
1268 R1[WS(rs, 9)] = T5R - T5S;
|
cannam@167
|
1269 R1[WS(rs, 25)] = T5R + T5S;
|
cannam@167
|
1270 T5V = T5T - T5U;
|
cannam@167
|
1271 T5Y = FNMS(KP1_763842528, T5X, KP942793473 * T5W);
|
cannam@167
|
1272 R1[WS(rs, 21)] = T5V - T5Y;
|
cannam@167
|
1273 R1[WS(rs, 5)] = T5V + T5Y;
|
cannam@167
|
1274 }
|
cannam@167
|
1275 }
|
cannam@167
|
1276 {
|
cannam@167
|
1277 E T3N, T4N, T46, T4O, T4y, T4Q, T4J, T4R, T3Y, T45;
|
cannam@167
|
1278 T3N = T3F + T3M;
|
cannam@167
|
1279 T4N = T3F - T3M;
|
cannam@167
|
1280 T3Y = T3Q + T3X;
|
cannam@167
|
1281 T45 = T41 + T44;
|
cannam@167
|
1282 T46 = FNMS(KP390180644, T45, KP1_961570560 * T3Y);
|
cannam@167
|
1283 T4O = FMA(KP390180644, T3Y, KP1_961570560 * T45);
|
cannam@167
|
1284 {
|
cannam@167
|
1285 E T4i, T4x, T4B, T4I;
|
cannam@167
|
1286 T4i = T4a + T4h;
|
cannam@167
|
1287 T4x = T4p + T4w;
|
cannam@167
|
1288 T4y = T4i + T4x;
|
cannam@167
|
1289 T4Q = T4i - T4x;
|
cannam@167
|
1290 T4B = T4z + T4A;
|
cannam@167
|
1291 T4I = T4E + T4H;
|
cannam@167
|
1292 T4J = T4B + T4I;
|
cannam@167
|
1293 T4R = T4I - T4B;
|
cannam@167
|
1294 }
|
cannam@167
|
1295 {
|
cannam@167
|
1296 E T47, T4K, T4T, T4U;
|
cannam@167
|
1297 T47 = T3N + T46;
|
cannam@167
|
1298 T4K = FNMS(KP196034280, T4J, KP1_990369453 * T4y);
|
cannam@167
|
1299 R1[WS(rs, 16)] = T47 - T4K;
|
cannam@167
|
1300 R1[0] = T47 + T4K;
|
cannam@167
|
1301 T4T = T4N + T4O;
|
cannam@167
|
1302 T4U = FMA(KP1_546020906, T4Q, KP1_268786568 * T4R);
|
cannam@167
|
1303 R1[WS(rs, 12)] = T4T - T4U;
|
cannam@167
|
1304 R1[WS(rs, 28)] = T4T + T4U;
|
cannam@167
|
1305 }
|
cannam@167
|
1306 {
|
cannam@167
|
1307 E T4L, T4M, T4P, T4S;
|
cannam@167
|
1308 T4L = T3N - T46;
|
cannam@167
|
1309 T4M = FMA(KP196034280, T4y, KP1_990369453 * T4J);
|
cannam@167
|
1310 R1[WS(rs, 8)] = T4L - T4M;
|
cannam@167
|
1311 R1[WS(rs, 24)] = T4L + T4M;
|
cannam@167
|
1312 T4P = T4N - T4O;
|
cannam@167
|
1313 T4S = FNMS(KP1_546020906, T4R, KP1_268786568 * T4Q);
|
cannam@167
|
1314 R1[WS(rs, 20)] = T4P - T4S;
|
cannam@167
|
1315 R1[WS(rs, 4)] = T4P + T4S;
|
cannam@167
|
1316 }
|
cannam@167
|
1317 }
|
cannam@167
|
1318 {
|
cannam@167
|
1319 E T63, T6h, T66, T6i, T6a, T6k, T6d, T6l, T64, T65;
|
cannam@167
|
1320 T63 = T61 - T62;
|
cannam@167
|
1321 T6h = T61 + T62;
|
cannam@167
|
1322 T64 = T5q + T5r;
|
cannam@167
|
1323 T65 = T5u - T5t;
|
cannam@167
|
1324 T66 = FNMS(KP1_961570560, T65, KP390180644 * T64);
|
cannam@167
|
1325 T6i = FMA(KP1_961570560, T64, KP390180644 * T65);
|
cannam@167
|
1326 {
|
cannam@167
|
1327 E T68, T69, T6b, T6c;
|
cannam@167
|
1328 T68 = T5y + T5z;
|
cannam@167
|
1329 T69 = T5J + T5K;
|
cannam@167
|
1330 T6a = T68 - T69;
|
cannam@167
|
1331 T6k = T68 + T69;
|
cannam@167
|
1332 T6b = T5D - T5G;
|
cannam@167
|
1333 T6c = T5N - T5M;
|
cannam@167
|
1334 T6d = T6b + T6c;
|
cannam@167
|
1335 T6l = T6c - T6b;
|
cannam@167
|
1336 }
|
cannam@167
|
1337 {
|
cannam@167
|
1338 E T67, T6e, T6n, T6o;
|
cannam@167
|
1339 T67 = T63 + T66;
|
cannam@167
|
1340 T6e = FNMS(KP1_268786568, T6d, KP1_546020906 * T6a);
|
cannam@167
|
1341 R1[WS(rs, 19)] = T67 - T6e;
|
cannam@167
|
1342 R1[WS(rs, 3)] = T67 + T6e;
|
cannam@167
|
1343 T6n = T6h + T6i;
|
cannam@167
|
1344 T6o = FMA(KP1_990369453, T6k, KP196034280 * T6l);
|
cannam@167
|
1345 R1[WS(rs, 15)] = T6n - T6o;
|
cannam@167
|
1346 R1[WS(rs, 31)] = T6n + T6o;
|
cannam@167
|
1347 }
|
cannam@167
|
1348 {
|
cannam@167
|
1349 E T6f, T6g, T6j, T6m;
|
cannam@167
|
1350 T6f = T63 - T66;
|
cannam@167
|
1351 T6g = FMA(KP1_268786568, T6a, KP1_546020906 * T6d);
|
cannam@167
|
1352 R1[WS(rs, 11)] = T6f - T6g;
|
cannam@167
|
1353 R1[WS(rs, 27)] = T6f + T6g;
|
cannam@167
|
1354 T6j = T6h - T6i;
|
cannam@167
|
1355 T6m = FNMS(KP1_990369453, T6l, KP196034280 * T6k);
|
cannam@167
|
1356 R1[WS(rs, 23)] = T6j - T6m;
|
cannam@167
|
1357 R1[WS(rs, 7)] = T6j + T6m;
|
cannam@167
|
1358 }
|
cannam@167
|
1359 }
|
cannam@167
|
1360 {
|
cannam@167
|
1361 E T4X, T5b, T50, T5c, T54, T5e, T57, T5f, T4Y, T4Z;
|
cannam@167
|
1362 T4X = T4V - T4W;
|
cannam@167
|
1363 T5b = T4V + T4W;
|
cannam@167
|
1364 T4Y = T3Q - T3X;
|
cannam@167
|
1365 T4Z = T44 - T41;
|
cannam@167
|
1366 T50 = FNMS(KP1_662939224, T4Z, KP1_111140466 * T4Y);
|
cannam@167
|
1367 T5c = FMA(KP1_662939224, T4Y, KP1_111140466 * T4Z);
|
cannam@167
|
1368 {
|
cannam@167
|
1369 E T52, T53, T55, T56;
|
cannam@167
|
1370 T52 = T4a - T4h;
|
cannam@167
|
1371 T53 = T4A - T4z;
|
cannam@167
|
1372 T54 = T52 + T53;
|
cannam@167
|
1373 T5e = T52 - T53;
|
cannam@167
|
1374 T55 = T4p - T4w;
|
cannam@167
|
1375 T56 = T4H - T4E;
|
cannam@167
|
1376 T57 = T55 + T56;
|
cannam@167
|
1377 T5f = T56 - T55;
|
cannam@167
|
1378 }
|
cannam@167
|
1379 {
|
cannam@167
|
1380 E T51, T58, T5h, T5i;
|
cannam@167
|
1381 T51 = T4X + T50;
|
cannam@167
|
1382 T58 = FNMS(KP942793473, T57, KP1_763842528 * T54);
|
cannam@167
|
1383 R1[WS(rs, 18)] = T51 - T58;
|
cannam@167
|
1384 R1[WS(rs, 2)] = T51 + T58;
|
cannam@167
|
1385 T5h = T5b + T5c;
|
cannam@167
|
1386 T5i = FMA(KP1_913880671, T5e, KP580569354 * T5f);
|
cannam@167
|
1387 R1[WS(rs, 14)] = T5h - T5i;
|
cannam@167
|
1388 R1[WS(rs, 30)] = T5h + T5i;
|
cannam@167
|
1389 }
|
cannam@167
|
1390 {
|
cannam@167
|
1391 E T59, T5a, T5d, T5g;
|
cannam@167
|
1392 T59 = T4X - T50;
|
cannam@167
|
1393 T5a = FMA(KP942793473, T54, KP1_763842528 * T57);
|
cannam@167
|
1394 R1[WS(rs, 10)] = T59 - T5a;
|
cannam@167
|
1395 R1[WS(rs, 26)] = T59 + T5a;
|
cannam@167
|
1396 T5d = T5b - T5c;
|
cannam@167
|
1397 T5g = FNMS(KP1_913880671, T5f, KP580569354 * T5e);
|
cannam@167
|
1398 R1[WS(rs, 22)] = T5d - T5g;
|
cannam@167
|
1399 R1[WS(rs, 6)] = T5d + T5g;
|
cannam@167
|
1400 }
|
cannam@167
|
1401 }
|
cannam@167
|
1402 }
|
cannam@167
|
1403 }
|
cannam@167
|
1404 }
|
cannam@167
|
1405
|
cannam@167
|
1406 static const kr2c_desc desc = { 64, "r2cb_64", {342, 82, 52, 0}, &GENUS };
|
cannam@167
|
1407
|
cannam@167
|
1408 void X(codelet_r2cb_64) (planner *p) {
|
cannam@167
|
1409 X(kr2c_register) (p, r2cb_64, &desc);
|
cannam@167
|
1410 }
|
cannam@167
|
1411
|
cannam@167
|
1412 #endif
|