cannam@127
|
1 /*
|
cannam@127
|
2 * Copyright (c) 2003, 2007-14 Matteo Frigo
|
cannam@127
|
3 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
|
cannam@127
|
4 *
|
cannam@127
|
5 * This program is free software; you can redistribute it and/or modify
|
cannam@127
|
6 * it under the terms of the GNU General Public License as published by
|
cannam@127
|
7 * the Free Software Foundation; either version 2 of the License, or
|
cannam@127
|
8 * (at your option) any later version.
|
cannam@127
|
9 *
|
cannam@127
|
10 * This program is distributed in the hope that it will be useful,
|
cannam@127
|
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
cannam@127
|
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
cannam@127
|
13 * GNU General Public License for more details.
|
cannam@127
|
14 *
|
cannam@127
|
15 * You should have received a copy of the GNU General Public License
|
cannam@127
|
16 * along with this program; if not, write to the Free Software
|
cannam@127
|
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
cannam@127
|
18 *
|
cannam@127
|
19 */
|
cannam@127
|
20
|
cannam@127
|
21 /* This file was automatically generated --- DO NOT EDIT */
|
cannam@127
|
22 /* Generated on Sat Jul 30 16:49:31 EDT 2016 */
|
cannam@127
|
23
|
cannam@127
|
24 #include "codelet-rdft.h"
|
cannam@127
|
25
|
cannam@127
|
26 #ifdef HAVE_FMA
|
cannam@127
|
27
|
cannam@127
|
28 /* Generated by: ../../../genfft/gen_r2cb.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -sign 1 -n 64 -name r2cb_64 -include r2cb.h */
|
cannam@127
|
29
|
cannam@127
|
30 /*
|
cannam@127
|
31 * This function contains 394 FP additions, 216 FP multiplications,
|
cannam@127
|
32 * (or, 178 additions, 0 multiplications, 216 fused multiply/add),
|
cannam@127
|
33 * 143 stack variables, 18 constants, and 128 memory accesses
|
cannam@127
|
34 */
|
cannam@127
|
35 #include "r2cb.h"
|
cannam@127
|
36
|
cannam@127
|
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@127
|
38 {
|
cannam@127
|
39 DK(KP1_990369453, +1.990369453344393772489673906218959843150949737);
|
cannam@127
|
40 DK(KP1_546020906, +1.546020906725473921621813219516939601942082586);
|
cannam@127
|
41 DK(KP098491403, +0.098491403357164253077197521291327432293052451);
|
cannam@127
|
42 DK(KP820678790, +0.820678790828660330972281985331011598767386482);
|
cannam@127
|
43 DK(KP1_913880671, +1.913880671464417729871595773960539938965698411);
|
cannam@127
|
44 DK(KP1_763842528, +1.763842528696710059425513727320776699016885241);
|
cannam@127
|
45 DK(KP303346683, +0.303346683607342391675883946941299872384187453);
|
cannam@127
|
46 DK(KP534511135, +0.534511135950791641089685961295362908582039528);
|
cannam@127
|
47 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
cannam@127
|
48 DK(KP1_662939224, +1.662939224605090474157576755235811513477121624);
|
cannam@127
|
49 DK(KP1_961570560, +1.961570560806460898252364472268478073947867462);
|
cannam@127
|
50 DK(KP668178637, +0.668178637919298919997757686523080761552472251);
|
cannam@127
|
51 DK(KP198912367, +0.198912367379658006911597622644676228597850501);
|
cannam@127
|
52 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
cannam@127
|
53 DK(KP1_847759065, +1.847759065022573512256366378793576573644833252);
|
cannam@127
|
54 DK(KP414213562, +0.414213562373095048801688724209698078569671875);
|
cannam@127
|
55 DK(KP1_414213562, +1.414213562373095048801688724209698078569671875);
|
cannam@127
|
56 DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
|
cannam@127
|
57 {
|
cannam@127
|
58 INT i;
|
cannam@127
|
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@127
|
60 E T3d, T32, T37, T2Z, T3f, T3b, T3c, T35;
|
cannam@127
|
61 {
|
cannam@127
|
62 E T5H, T9, T5j, T4p, T2T, T1b, T3Z, T3j, Tg, T5I, T5k, T4u, T40, T3m, T2U;
|
cannam@127
|
63 E T1m, T3o, T1s, T1J, T3r, T5K, Tw, T5N, T6c, T4A, T5n, T3s, T1D, T5m, T4F;
|
cannam@127
|
64 E T3p, T1M, T3w, T1U, T2z, T3H, T5Q, TM, T6f, T5Y, T5q, T4M, T3I, T25, T5t;
|
cannam@127
|
65 E T53, T3x, T2C, T3A, T5V, T11, T6g, T5T, T55, T4W, T3z, T2E, T2h, T2F, T2s;
|
cannam@127
|
66 E T3L, T3E, T54, T4R;
|
cannam@127
|
67 {
|
cannam@127
|
68 E Td, T1c, Tc, T4r, T1k, Te, T1d, T1e;
|
cannam@127
|
69 {
|
cannam@127
|
70 E T3h, T15, T1a, T3i;
|
cannam@127
|
71 {
|
cannam@127
|
72 E T4, T14, T17, T13, T3, T16, T8, T18;
|
cannam@127
|
73 T4 = Cr[WS(csr, 16)];
|
cannam@127
|
74 T14 = Ci[WS(csi, 16)];
|
cannam@127
|
75 {
|
cannam@127
|
76 E T1, T2, T6, T7;
|
cannam@127
|
77 T1 = Cr[0];
|
cannam@127
|
78 T2 = Cr[WS(csr, 32)];
|
cannam@127
|
79 T6 = Cr[WS(csr, 8)];
|
cannam@127
|
80 T7 = Cr[WS(csr, 24)];
|
cannam@127
|
81 T17 = Ci[WS(csi, 8)];
|
cannam@127
|
82 T13 = T1 - T2;
|
cannam@127
|
83 T3 = T1 + T2;
|
cannam@127
|
84 T16 = T6 - T7;
|
cannam@127
|
85 T8 = T6 + T7;
|
cannam@127
|
86 T18 = Ci[WS(csi, 24)];
|
cannam@127
|
87 }
|
cannam@127
|
88 {
|
cannam@127
|
89 E T4n, T5, T4o, T19;
|
cannam@127
|
90 T4n = FNMS(KP2_000000000, T4, T3);
|
cannam@127
|
91 T5 = FMA(KP2_000000000, T4, T3);
|
cannam@127
|
92 T3h = FMA(KP2_000000000, T14, T13);
|
cannam@127
|
93 T15 = FNMS(KP2_000000000, T14, T13);
|
cannam@127
|
94 T4o = T17 - T18;
|
cannam@127
|
95 T19 = T17 + T18;
|
cannam@127
|
96 T5H = FNMS(KP2_000000000, T8, T5);
|
cannam@127
|
97 T9 = FMA(KP2_000000000, T8, T5);
|
cannam@127
|
98 T5j = FMA(KP2_000000000, T4o, T4n);
|
cannam@127
|
99 T4p = FNMS(KP2_000000000, T4o, T4n);
|
cannam@127
|
100 T1a = T16 - T19;
|
cannam@127
|
101 T3i = T16 + T19;
|
cannam@127
|
102 }
|
cannam@127
|
103 }
|
cannam@127
|
104 {
|
cannam@127
|
105 E Ta, Tb, T1i, T1j;
|
cannam@127
|
106 Ta = Cr[WS(csr, 4)];
|
cannam@127
|
107 T2T = FNMS(KP1_414213562, T1a, T15);
|
cannam@127
|
108 T1b = FMA(KP1_414213562, T1a, T15);
|
cannam@127
|
109 T3Z = FMA(KP1_414213562, T3i, T3h);
|
cannam@127
|
110 T3j = FNMS(KP1_414213562, T3i, T3h);
|
cannam@127
|
111 Tb = Cr[WS(csr, 28)];
|
cannam@127
|
112 T1i = Ci[WS(csi, 4)];
|
cannam@127
|
113 T1j = Ci[WS(csi, 28)];
|
cannam@127
|
114 Td = Cr[WS(csr, 20)];
|
cannam@127
|
115 T1c = Ta - Tb;
|
cannam@127
|
116 Tc = Ta + Tb;
|
cannam@127
|
117 T4r = T1i - T1j;
|
cannam@127
|
118 T1k = T1i + T1j;
|
cannam@127
|
119 Te = Cr[WS(csr, 12)];
|
cannam@127
|
120 T1d = Ci[WS(csi, 20)];
|
cannam@127
|
121 T1e = Ci[WS(csi, 12)];
|
cannam@127
|
122 }
|
cannam@127
|
123 }
|
cannam@127
|
124 {
|
cannam@127
|
125 E T4B, T4E, T1K, T1L;
|
cannam@127
|
126 {
|
cannam@127
|
127 E T1o, Tk, T4C, T1I, T1F, Tn, T4D, T1r, Ts, T1t, Tr, T4y, T1w, Tt, T1z;
|
cannam@127
|
128 E T1A;
|
cannam@127
|
129 {
|
cannam@127
|
130 E Tl, Tm, T1p, T1q;
|
cannam@127
|
131 {
|
cannam@127
|
132 E Ti, Tj, T1G, T1H, T1h, Tf;
|
cannam@127
|
133 Ti = Cr[WS(csr, 2)];
|
cannam@127
|
134 T1h = Td - Te;
|
cannam@127
|
135 Tf = Td + Te;
|
cannam@127
|
136 {
|
cannam@127
|
137 E T4s, T1f, T3k, T1l;
|
cannam@127
|
138 T4s = T1d - T1e;
|
cannam@127
|
139 T1f = T1d + T1e;
|
cannam@127
|
140 T3k = T1k - T1h;
|
cannam@127
|
141 T1l = T1h + T1k;
|
cannam@127
|
142 {
|
cannam@127
|
143 E T4q, T4t, T3l, T1g;
|
cannam@127
|
144 T4q = Tc - Tf;
|
cannam@127
|
145 Tg = Tc + Tf;
|
cannam@127
|
146 T4t = T4r - T4s;
|
cannam@127
|
147 T5I = T4s + T4r;
|
cannam@127
|
148 T3l = T1c + T1f;
|
cannam@127
|
149 T1g = T1c - T1f;
|
cannam@127
|
150 T5k = T4q + T4t;
|
cannam@127
|
151 T4u = T4q - T4t;
|
cannam@127
|
152 T40 = FMA(KP414213562, T3k, T3l);
|
cannam@127
|
153 T3m = FNMS(KP414213562, T3l, T3k);
|
cannam@127
|
154 T2U = FMA(KP414213562, T1g, T1l);
|
cannam@127
|
155 T1m = FNMS(KP414213562, T1l, T1g);
|
cannam@127
|
156 Tj = Cr[WS(csr, 30)];
|
cannam@127
|
157 }
|
cannam@127
|
158 }
|
cannam@127
|
159 T1G = Ci[WS(csi, 2)];
|
cannam@127
|
160 T1H = Ci[WS(csi, 30)];
|
cannam@127
|
161 Tl = Cr[WS(csr, 18)];
|
cannam@127
|
162 T1o = Ti - Tj;
|
cannam@127
|
163 Tk = Ti + Tj;
|
cannam@127
|
164 T4C = T1G - T1H;
|
cannam@127
|
165 T1I = T1G + T1H;
|
cannam@127
|
166 Tm = Cr[WS(csr, 14)];
|
cannam@127
|
167 T1p = Ci[WS(csi, 18)];
|
cannam@127
|
168 T1q = Ci[WS(csi, 14)];
|
cannam@127
|
169 }
|
cannam@127
|
170 {
|
cannam@127
|
171 E Tp, Tq, T1u, T1v;
|
cannam@127
|
172 Tp = Cr[WS(csr, 10)];
|
cannam@127
|
173 T1F = Tl - Tm;
|
cannam@127
|
174 Tn = Tl + Tm;
|
cannam@127
|
175 T4D = T1p - T1q;
|
cannam@127
|
176 T1r = T1p + T1q;
|
cannam@127
|
177 Tq = Cr[WS(csr, 22)];
|
cannam@127
|
178 T1u = Ci[WS(csi, 10)];
|
cannam@127
|
179 T1v = Ci[WS(csi, 22)];
|
cannam@127
|
180 Ts = Cr[WS(csr, 6)];
|
cannam@127
|
181 T1t = Tp - Tq;
|
cannam@127
|
182 Tr = Tp + Tq;
|
cannam@127
|
183 T4y = T1u - T1v;
|
cannam@127
|
184 T1w = T1u + T1v;
|
cannam@127
|
185 Tt = Cr[WS(csr, 26)];
|
cannam@127
|
186 T1z = Ci[WS(csi, 6)];
|
cannam@127
|
187 T1A = Ci[WS(csi, 26)];
|
cannam@127
|
188 }
|
cannam@127
|
189 }
|
cannam@127
|
190 {
|
cannam@127
|
191 E T1y, T4x, T1B, T4w, To, Tv, Tu;
|
cannam@127
|
192 T3o = T1o + T1r;
|
cannam@127
|
193 T1s = T1o - T1r;
|
cannam@127
|
194 T1y = Ts - Tt;
|
cannam@127
|
195 Tu = Ts + Tt;
|
cannam@127
|
196 T4x = T1A - T1z;
|
cannam@127
|
197 T1B = T1z + T1A;
|
cannam@127
|
198 T1J = T1F + T1I;
|
cannam@127
|
199 T3r = T1I - T1F;
|
cannam@127
|
200 T4w = Tk - Tn;
|
cannam@127
|
201 To = Tk + Tn;
|
cannam@127
|
202 Tv = Tr + Tu;
|
cannam@127
|
203 T4B = Tr - Tu;
|
cannam@127
|
204 {
|
cannam@127
|
205 E T4z, T5L, T5M, T1x, T1C;
|
cannam@127
|
206 T4E = T4C - T4D;
|
cannam@127
|
207 T5L = T4D + T4C;
|
cannam@127
|
208 T5M = T4y + T4x;
|
cannam@127
|
209 T4z = T4x - T4y;
|
cannam@127
|
210 T5K = To - Tv;
|
cannam@127
|
211 Tw = To + Tv;
|
cannam@127
|
212 T5N = T5L - T5M;
|
cannam@127
|
213 T6c = T5M + T5L;
|
cannam@127
|
214 T1K = T1t + T1w;
|
cannam@127
|
215 T1x = T1t - T1w;
|
cannam@127
|
216 T1C = T1y - T1B;
|
cannam@127
|
217 T1L = T1y + T1B;
|
cannam@127
|
218 T4A = T4w + T4z;
|
cannam@127
|
219 T5n = T4w - T4z;
|
cannam@127
|
220 T3s = T1C - T1x;
|
cannam@127
|
221 T1D = T1x + T1C;
|
cannam@127
|
222 }
|
cannam@127
|
223 }
|
cannam@127
|
224 }
|
cannam@127
|
225 {
|
cannam@127
|
226 E T4Z, T52, T2A, T2B;
|
cannam@127
|
227 {
|
cannam@127
|
228 E T1Q, TA, T50, T2y, T2v, TD, T51, T1T, TI, T1V, TH, T4K, T1Y, TJ, T21;
|
cannam@127
|
229 E T22;
|
cannam@127
|
230 {
|
cannam@127
|
231 E TB, TC, T1R, T1S;
|
cannam@127
|
232 {
|
cannam@127
|
233 E Ty, Tz, T2w, T2x;
|
cannam@127
|
234 Ty = Cr[WS(csr, 1)];
|
cannam@127
|
235 T5m = T4E - T4B;
|
cannam@127
|
236 T4F = T4B + T4E;
|
cannam@127
|
237 T3p = T1K + T1L;
|
cannam@127
|
238 T1M = T1K - T1L;
|
cannam@127
|
239 Tz = Cr[WS(csr, 31)];
|
cannam@127
|
240 T2w = Ci[WS(csi, 1)];
|
cannam@127
|
241 T2x = Ci[WS(csi, 31)];
|
cannam@127
|
242 TB = Cr[WS(csr, 17)];
|
cannam@127
|
243 T1Q = Ty - Tz;
|
cannam@127
|
244 TA = Ty + Tz;
|
cannam@127
|
245 T50 = T2w - T2x;
|
cannam@127
|
246 T2y = T2w + T2x;
|
cannam@127
|
247 TC = Cr[WS(csr, 15)];
|
cannam@127
|
248 T1R = Ci[WS(csi, 17)];
|
cannam@127
|
249 T1S = Ci[WS(csi, 15)];
|
cannam@127
|
250 }
|
cannam@127
|
251 {
|
cannam@127
|
252 E TF, TG, T1W, T1X;
|
cannam@127
|
253 TF = Cr[WS(csr, 9)];
|
cannam@127
|
254 T2v = TB - TC;
|
cannam@127
|
255 TD = TB + TC;
|
cannam@127
|
256 T51 = T1R - T1S;
|
cannam@127
|
257 T1T = T1R + T1S;
|
cannam@127
|
258 TG = Cr[WS(csr, 23)];
|
cannam@127
|
259 T1W = Ci[WS(csi, 9)];
|
cannam@127
|
260 T1X = Ci[WS(csi, 23)];
|
cannam@127
|
261 TI = Cr[WS(csr, 7)];
|
cannam@127
|
262 T1V = TF - TG;
|
cannam@127
|
263 TH = TF + TG;
|
cannam@127
|
264 T4K = T1W - T1X;
|
cannam@127
|
265 T1Y = T1W + T1X;
|
cannam@127
|
266 TJ = Cr[WS(csr, 25)];
|
cannam@127
|
267 T21 = Ci[WS(csi, 7)];
|
cannam@127
|
268 T22 = Ci[WS(csi, 25)];
|
cannam@127
|
269 }
|
cannam@127
|
270 }
|
cannam@127
|
271 {
|
cannam@127
|
272 E T20, T4J, T23, T4I, TE, TL, TK;
|
cannam@127
|
273 T3w = T1Q + T1T;
|
cannam@127
|
274 T1U = T1Q - T1T;
|
cannam@127
|
275 T20 = TI - TJ;
|
cannam@127
|
276 TK = TI + TJ;
|
cannam@127
|
277 T4J = T22 - T21;
|
cannam@127
|
278 T23 = T21 + T22;
|
cannam@127
|
279 T2z = T2v + T2y;
|
cannam@127
|
280 T3H = T2y - T2v;
|
cannam@127
|
281 T4I = TA - TD;
|
cannam@127
|
282 TE = TA + TD;
|
cannam@127
|
283 TL = TH + TK;
|
cannam@127
|
284 T4Z = TH - TK;
|
cannam@127
|
285 {
|
cannam@127
|
286 E T4L, T5W, T5X, T1Z, T24;
|
cannam@127
|
287 T52 = T50 - T51;
|
cannam@127
|
288 T5W = T51 + T50;
|
cannam@127
|
289 T5X = T4K + T4J;
|
cannam@127
|
290 T4L = T4J - T4K;
|
cannam@127
|
291 T5Q = TE - TL;
|
cannam@127
|
292 TM = TE + TL;
|
cannam@127
|
293 T6f = T5X + T5W;
|
cannam@127
|
294 T5Y = T5W - T5X;
|
cannam@127
|
295 T2A = T1V + T1Y;
|
cannam@127
|
296 T1Z = T1V - T1Y;
|
cannam@127
|
297 T24 = T20 - T23;
|
cannam@127
|
298 T2B = T20 + T23;
|
cannam@127
|
299 T5q = T4I - T4L;
|
cannam@127
|
300 T4M = T4I + T4L;
|
cannam@127
|
301 T3I = T24 - T1Z;
|
cannam@127
|
302 T25 = T1Z + T24;
|
cannam@127
|
303 }
|
cannam@127
|
304 }
|
cannam@127
|
305 }
|
cannam@127
|
306 {
|
cannam@127
|
307 E T27, TP, T4O, T2f, T2c, TS, T4P, T2a, TX, T2i, TW, T4T, T2q, TY, T2j;
|
cannam@127
|
308 E T2k;
|
cannam@127
|
309 {
|
cannam@127
|
310 E TQ, TR, T28, T29;
|
cannam@127
|
311 {
|
cannam@127
|
312 E TN, TO, T2d, T2e;
|
cannam@127
|
313 TN = Cr[WS(csr, 5)];
|
cannam@127
|
314 T5t = T52 - T4Z;
|
cannam@127
|
315 T53 = T4Z + T52;
|
cannam@127
|
316 T3x = T2A + T2B;
|
cannam@127
|
317 T2C = T2A - T2B;
|
cannam@127
|
318 TO = Cr[WS(csr, 27)];
|
cannam@127
|
319 T2d = Ci[WS(csi, 5)];
|
cannam@127
|
320 T2e = Ci[WS(csi, 27)];
|
cannam@127
|
321 TQ = Cr[WS(csr, 21)];
|
cannam@127
|
322 T27 = TN - TO;
|
cannam@127
|
323 TP = TN + TO;
|
cannam@127
|
324 T4O = T2d - T2e;
|
cannam@127
|
325 T2f = T2d + T2e;
|
cannam@127
|
326 TR = Cr[WS(csr, 11)];
|
cannam@127
|
327 T28 = Ci[WS(csi, 21)];
|
cannam@127
|
328 T29 = Ci[WS(csi, 11)];
|
cannam@127
|
329 }
|
cannam@127
|
330 {
|
cannam@127
|
331 E TU, TV, T2o, T2p;
|
cannam@127
|
332 TU = Cr[WS(csr, 3)];
|
cannam@127
|
333 T2c = TQ - TR;
|
cannam@127
|
334 TS = TQ + TR;
|
cannam@127
|
335 T4P = T28 - T29;
|
cannam@127
|
336 T2a = T28 + T29;
|
cannam@127
|
337 TV = Cr[WS(csr, 29)];
|
cannam@127
|
338 T2o = Ci[WS(csi, 3)];
|
cannam@127
|
339 T2p = Ci[WS(csi, 29)];
|
cannam@127
|
340 TX = Cr[WS(csr, 13)];
|
cannam@127
|
341 T2i = TU - TV;
|
cannam@127
|
342 TW = TU + TV;
|
cannam@127
|
343 T4T = T2p - T2o;
|
cannam@127
|
344 T2q = T2o + T2p;
|
cannam@127
|
345 TY = Cr[WS(csr, 19)];
|
cannam@127
|
346 T2j = Ci[WS(csi, 13)];
|
cannam@127
|
347 T2k = Ci[WS(csi, 19)];
|
cannam@127
|
348 }
|
cannam@127
|
349 }
|
cannam@127
|
350 {
|
cannam@127
|
351 E T4N, T2n, T2l, T4Q, T2b, T2g, TT, TZ, T4U;
|
cannam@127
|
352 T4N = TP - TS;
|
cannam@127
|
353 TT = TP + TS;
|
cannam@127
|
354 T2n = TX - TY;
|
cannam@127
|
355 TZ = TX + TY;
|
cannam@127
|
356 T4U = T2j - T2k;
|
cannam@127
|
357 T2l = T2j + T2k;
|
cannam@127
|
358 {
|
cannam@127
|
359 E T5S, T10, T4S, T4V, T5R;
|
cannam@127
|
360 T5S = T4P + T4O;
|
cannam@127
|
361 T4Q = T4O - T4P;
|
cannam@127
|
362 T10 = TW + TZ;
|
cannam@127
|
363 T4S = TW - TZ;
|
cannam@127
|
364 T4V = T4T - T4U;
|
cannam@127
|
365 T5R = T4U + T4T;
|
cannam@127
|
366 T3A = T27 + T2a;
|
cannam@127
|
367 T2b = T27 - T2a;
|
cannam@127
|
368 T5V = TT - T10;
|
cannam@127
|
369 T11 = TT + T10;
|
cannam@127
|
370 T6g = T5S + T5R;
|
cannam@127
|
371 T5T = T5R - T5S;
|
cannam@127
|
372 T55 = T4V - T4S;
|
cannam@127
|
373 T4W = T4S + T4V;
|
cannam@127
|
374 T2g = T2c + T2f;
|
cannam@127
|
375 T3z = T2f - T2c;
|
cannam@127
|
376 }
|
cannam@127
|
377 {
|
cannam@127
|
378 E T3D, T3C, T2m, T2r;
|
cannam@127
|
379 T3D = T2i + T2l;
|
cannam@127
|
380 T2m = T2i - T2l;
|
cannam@127
|
381 T2r = T2n - T2q;
|
cannam@127
|
382 T3C = T2n + T2q;
|
cannam@127
|
383 T2E = FMA(KP414213562, T2b, T2g);
|
cannam@127
|
384 T2h = FNMS(KP414213562, T2g, T2b);
|
cannam@127
|
385 T2F = FNMS(KP414213562, T2m, T2r);
|
cannam@127
|
386 T2s = FMA(KP414213562, T2r, T2m);
|
cannam@127
|
387 T3L = FMA(KP414213562, T3C, T3D);
|
cannam@127
|
388 T3E = FNMS(KP414213562, T3D, T3C);
|
cannam@127
|
389 T54 = T4N + T4Q;
|
cannam@127
|
390 T4R = T4N - T4Q;
|
cannam@127
|
391 }
|
cannam@127
|
392 }
|
cannam@127
|
393 }
|
cannam@127
|
394 }
|
cannam@127
|
395 }
|
cannam@127
|
396 }
|
cannam@127
|
397 {
|
cannam@127
|
398 E T3K, T3B, T5u, T5r, T5d, T5g;
|
cannam@127
|
399 {
|
cannam@127
|
400 E T6e, T6h, T6b, T5J, T5O, T5Z, T66, T69, T65, T67, T5U, T12, T6m, Th;
|
cannam@127
|
401 T6e = TM - T11;
|
cannam@127
|
402 T12 = TM + T11;
|
cannam@127
|
403 T6m = T6g + T6f;
|
cannam@127
|
404 T6h = T6f - T6g;
|
cannam@127
|
405 T6b = FNMS(KP2_000000000, Tg, T9);
|
cannam@127
|
406 Th = FMA(KP2_000000000, Tg, T9);
|
cannam@127
|
407 T3K = FMA(KP414213562, T3z, T3A);
|
cannam@127
|
408 T3B = FNMS(KP414213562, T3A, T3z);
|
cannam@127
|
409 {
|
cannam@127
|
410 E T63, T64, T6l, Tx;
|
cannam@127
|
411 T5J = FNMS(KP2_000000000, T5I, T5H);
|
cannam@127
|
412 T63 = FMA(KP2_000000000, T5I, T5H);
|
cannam@127
|
413 T64 = T5K + T5N;
|
cannam@127
|
414 T5O = T5K - T5N;
|
cannam@127
|
415 T5Z = T5V + T5Y;
|
cannam@127
|
416 T66 = T5Y - T5V;
|
cannam@127
|
417 T6l = FNMS(KP2_000000000, Tw, Th);
|
cannam@127
|
418 Tx = FMA(KP2_000000000, Tw, Th);
|
cannam@127
|
419 T69 = FMA(KP1_414213562, T64, T63);
|
cannam@127
|
420 T65 = FNMS(KP1_414213562, T64, T63);
|
cannam@127
|
421 R0[WS(rs, 8)] = FNMS(KP2_000000000, T6m, T6l);
|
cannam@127
|
422 R0[WS(rs, 24)] = FMA(KP2_000000000, T6m, T6l);
|
cannam@127
|
423 R0[0] = FMA(KP2_000000000, T12, Tx);
|
cannam@127
|
424 R0[WS(rs, 16)] = FNMS(KP2_000000000, T12, Tx);
|
cannam@127
|
425 T67 = T5Q - T5T;
|
cannam@127
|
426 T5U = T5Q + T5T;
|
cannam@127
|
427 }
|
cannam@127
|
428 {
|
cannam@127
|
429 E T6j, T6d, T6a, T68;
|
cannam@127
|
430 T6a = FMA(KP414213562, T66, T67);
|
cannam@127
|
431 T68 = FNMS(KP414213562, T67, T66);
|
cannam@127
|
432 T6j = FMA(KP2_000000000, T6c, T6b);
|
cannam@127
|
433 T6d = FNMS(KP2_000000000, T6c, T6b);
|
cannam@127
|
434 R0[WS(rs, 14)] = FNMS(KP1_847759065, T6a, T69);
|
cannam@127
|
435 R0[WS(rs, 30)] = FMA(KP1_847759065, T6a, T69);
|
cannam@127
|
436 R0[WS(rs, 22)] = FMA(KP1_847759065, T68, T65);
|
cannam@127
|
437 R0[WS(rs, 6)] = FNMS(KP1_847759065, T68, T65);
|
cannam@127
|
438 {
|
cannam@127
|
439 E T61, T5P, T6k, T6i;
|
cannam@127
|
440 T6k = T6e + T6h;
|
cannam@127
|
441 T6i = T6e - T6h;
|
cannam@127
|
442 T61 = FNMS(KP1_414213562, T5O, T5J);
|
cannam@127
|
443 T5P = FMA(KP1_414213562, T5O, T5J);
|
cannam@127
|
444 R0[WS(rs, 12)] = FNMS(KP1_414213562, T6k, T6j);
|
cannam@127
|
445 R0[WS(rs, 28)] = FMA(KP1_414213562, T6k, T6j);
|
cannam@127
|
446 R0[WS(rs, 4)] = FMA(KP1_414213562, T6i, T6d);
|
cannam@127
|
447 R0[WS(rs, 20)] = FNMS(KP1_414213562, T6i, T6d);
|
cannam@127
|
448 {
|
cannam@127
|
449 E T5b, T4v, T5f, T4Y, T5e, T57, T4G, T5c;
|
cannam@127
|
450 {
|
cannam@127
|
451 E T4X, T56, T62, T60;
|
cannam@127
|
452 T5u = T4W - T4R;
|
cannam@127
|
453 T4X = T4R + T4W;
|
cannam@127
|
454 T56 = T54 + T55;
|
cannam@127
|
455 T5r = T54 - T55;
|
cannam@127
|
456 T5b = FNMS(KP1_414213562, T4u, T4p);
|
cannam@127
|
457 T4v = FMA(KP1_414213562, T4u, T4p);
|
cannam@127
|
458 T62 = FMA(KP414213562, T5U, T5Z);
|
cannam@127
|
459 T60 = FNMS(KP414213562, T5Z, T5U);
|
cannam@127
|
460 T5f = FNMS(KP707106781, T4X, T4M);
|
cannam@127
|
461 T4Y = FMA(KP707106781, T4X, T4M);
|
cannam@127
|
462 T5e = FNMS(KP707106781, T56, T53);
|
cannam@127
|
463 T57 = FMA(KP707106781, T56, T53);
|
cannam@127
|
464 R0[WS(rs, 10)] = FNMS(KP1_847759065, T62, T61);
|
cannam@127
|
465 R0[WS(rs, 26)] = FMA(KP1_847759065, T62, T61);
|
cannam@127
|
466 R0[WS(rs, 2)] = FMA(KP1_847759065, T60, T5P);
|
cannam@127
|
467 R0[WS(rs, 18)] = FNMS(KP1_847759065, T60, T5P);
|
cannam@127
|
468 T4G = FNMS(KP414213562, T4F, T4A);
|
cannam@127
|
469 T5c = FMA(KP414213562, T4A, T4F);
|
cannam@127
|
470 }
|
cannam@127
|
471 {
|
cannam@127
|
472 E T5a, T59, T5h, T5i, T58, T4H;
|
cannam@127
|
473 T5a = FMA(KP198912367, T4Y, T57);
|
cannam@127
|
474 T58 = FNMS(KP198912367, T57, T4Y);
|
cannam@127
|
475 T59 = FNMS(KP1_847759065, T4G, T4v);
|
cannam@127
|
476 T4H = FMA(KP1_847759065, T4G, T4v);
|
cannam@127
|
477 T5h = FMA(KP1_847759065, T5c, T5b);
|
cannam@127
|
478 T5d = FNMS(KP1_847759065, T5c, T5b);
|
cannam@127
|
479 T5i = FMA(KP668178637, T5e, T5f);
|
cannam@127
|
480 T5g = FNMS(KP668178637, T5f, T5e);
|
cannam@127
|
481 R0[WS(rs, 1)] = FMA(KP1_961570560, T58, T4H);
|
cannam@127
|
482 R0[WS(rs, 17)] = FNMS(KP1_961570560, T58, T4H);
|
cannam@127
|
483 R0[WS(rs, 29)] = FMA(KP1_662939224, T5i, T5h);
|
cannam@127
|
484 R0[WS(rs, 13)] = FNMS(KP1_662939224, T5i, T5h);
|
cannam@127
|
485 R0[WS(rs, 25)] = FMA(KP1_961570560, T5a, T59);
|
cannam@127
|
486 R0[WS(rs, 9)] = FNMS(KP1_961570560, T5a, T59);
|
cannam@127
|
487 }
|
cannam@127
|
488 }
|
cannam@127
|
489 }
|
cannam@127
|
490 }
|
cannam@127
|
491 }
|
cannam@127
|
492 {
|
cannam@127
|
493 E T43, T42, T46, T4a, T49, T3V, T3G, T47, T3P, T3v, T3X, T3T, T3U, T3N, T5B;
|
cannam@127
|
494 E T5E;
|
cannam@127
|
495 {
|
cannam@127
|
496 E T5s, T5D, T5z, T5l, T5C, T5v, T5o, T5A;
|
cannam@127
|
497 R0[WS(rs, 21)] = FMA(KP1_662939224, T5g, T5d);
|
cannam@127
|
498 R0[WS(rs, 5)] = FNMS(KP1_662939224, T5g, T5d);
|
cannam@127
|
499 T5s = FNMS(KP707106781, T5r, T5q);
|
cannam@127
|
500 T5D = FMA(KP707106781, T5r, T5q);
|
cannam@127
|
501 T5z = FMA(KP1_414213562, T5k, T5j);
|
cannam@127
|
502 T5l = FNMS(KP1_414213562, T5k, T5j);
|
cannam@127
|
503 T5C = FMA(KP707106781, T5u, T5t);
|
cannam@127
|
504 T5v = FNMS(KP707106781, T5u, T5t);
|
cannam@127
|
505 T5o = FNMS(KP414213562, T5n, T5m);
|
cannam@127
|
506 T5A = FMA(KP414213562, T5m, T5n);
|
cannam@127
|
507 {
|
cannam@127
|
508 E T5y, T5x, T5F, T5G, T5w, T5p;
|
cannam@127
|
509 T5y = FMA(KP668178637, T5s, T5v);
|
cannam@127
|
510 T5w = FNMS(KP668178637, T5v, T5s);
|
cannam@127
|
511 T5x = FMA(KP1_847759065, T5o, T5l);
|
cannam@127
|
512 T5p = FNMS(KP1_847759065, T5o, T5l);
|
cannam@127
|
513 T5F = FMA(KP1_847759065, T5A, T5z);
|
cannam@127
|
514 T5B = FNMS(KP1_847759065, T5A, T5z);
|
cannam@127
|
515 T5G = FMA(KP198912367, T5C, T5D);
|
cannam@127
|
516 T5E = FNMS(KP198912367, T5D, T5C);
|
cannam@127
|
517 R0[WS(rs, 3)] = FMA(KP1_662939224, T5w, T5p);
|
cannam@127
|
518 R0[WS(rs, 19)] = FNMS(KP1_662939224, T5w, T5p);
|
cannam@127
|
519 R0[WS(rs, 31)] = FMA(KP1_961570560, T5G, T5F);
|
cannam@127
|
520 R0[WS(rs, 15)] = FNMS(KP1_961570560, T5G, T5F);
|
cannam@127
|
521 R0[WS(rs, 27)] = FMA(KP1_662939224, T5y, T5x);
|
cannam@127
|
522 R0[WS(rs, 11)] = FNMS(KP1_662939224, T5y, T5x);
|
cannam@127
|
523 }
|
cannam@127
|
524 }
|
cannam@127
|
525 {
|
cannam@127
|
526 E T3R, T3n, T3J, T3S, T3u, T3M;
|
cannam@127
|
527 T3R = FMA(KP1_847759065, T3m, T3j);
|
cannam@127
|
528 T3n = FNMS(KP1_847759065, T3m, T3j);
|
cannam@127
|
529 R0[WS(rs, 23)] = FMA(KP1_961570560, T5E, T5B);
|
cannam@127
|
530 R0[WS(rs, 7)] = FNMS(KP1_961570560, T5E, T5B);
|
cannam@127
|
531 {
|
cannam@127
|
532 E T3q, T3t, T3y, T3F;
|
cannam@127
|
533 T43 = FMA(KP707106781, T3p, T3o);
|
cannam@127
|
534 T3q = FNMS(KP707106781, T3p, T3o);
|
cannam@127
|
535 T3t = FNMS(KP707106781, T3s, T3r);
|
cannam@127
|
536 T42 = FMA(KP707106781, T3s, T3r);
|
cannam@127
|
537 T46 = FMA(KP707106781, T3x, T3w);
|
cannam@127
|
538 T3y = FNMS(KP707106781, T3x, T3w);
|
cannam@127
|
539 T3F = T3B + T3E;
|
cannam@127
|
540 T4a = T3B - T3E;
|
cannam@127
|
541 T49 = FMA(KP707106781, T3I, T3H);
|
cannam@127
|
542 T3J = FNMS(KP707106781, T3I, T3H);
|
cannam@127
|
543 T3S = FMA(KP668178637, T3q, T3t);
|
cannam@127
|
544 T3u = FNMS(KP668178637, T3t, T3q);
|
cannam@127
|
545 T3V = FMA(KP923879532, T3F, T3y);
|
cannam@127
|
546 T3G = FNMS(KP923879532, T3F, T3y);
|
cannam@127
|
547 T3M = T3K - T3L;
|
cannam@127
|
548 T47 = T3K + T3L;
|
cannam@127
|
549 }
|
cannam@127
|
550 T3P = FNMS(KP1_662939224, T3u, T3n);
|
cannam@127
|
551 T3v = FMA(KP1_662939224, T3u, T3n);
|
cannam@127
|
552 T3X = FMA(KP1_662939224, T3S, T3R);
|
cannam@127
|
553 T3T = FNMS(KP1_662939224, T3S, T3R);
|
cannam@127
|
554 T3U = FNMS(KP923879532, T3M, T3J);
|
cannam@127
|
555 T3N = FMA(KP923879532, T3M, T3J);
|
cannam@127
|
556 }
|
cannam@127
|
557 {
|
cannam@127
|
558 E T2X, T2W, T30, T34, T33, T2P, T2u, T31, T2J, T1P, T2R, T2N, T2O, T2H;
|
cannam@127
|
559 {
|
cannam@127
|
560 E T2L, T1n, T2D, T2M, T1O, T2G;
|
cannam@127
|
561 T2L = FNMS(KP1_847759065, T1m, T1b);
|
cannam@127
|
562 T1n = FMA(KP1_847759065, T1m, T1b);
|
cannam@127
|
563 {
|
cannam@127
|
564 E T3W, T3Y, T3Q, T3O;
|
cannam@127
|
565 T3W = FNMS(KP534511135, T3V, T3U);
|
cannam@127
|
566 T3Y = FMA(KP534511135, T3U, T3V);
|
cannam@127
|
567 T3Q = FMA(KP303346683, T3G, T3N);
|
cannam@127
|
568 T3O = FNMS(KP303346683, T3N, T3G);
|
cannam@127
|
569 R1[WS(rs, 21)] = FMA(KP1_763842528, T3W, T3T);
|
cannam@127
|
570 R1[WS(rs, 5)] = FNMS(KP1_763842528, T3W, T3T);
|
cannam@127
|
571 R1[WS(rs, 29)] = FMA(KP1_763842528, T3Y, T3X);
|
cannam@127
|
572 R1[WS(rs, 13)] = FNMS(KP1_763842528, T3Y, T3X);
|
cannam@127
|
573 R1[WS(rs, 25)] = FMA(KP1_913880671, T3Q, T3P);
|
cannam@127
|
574 R1[WS(rs, 9)] = FNMS(KP1_913880671, T3Q, T3P);
|
cannam@127
|
575 R1[WS(rs, 1)] = FMA(KP1_913880671, T3O, T3v);
|
cannam@127
|
576 R1[WS(rs, 17)] = FNMS(KP1_913880671, T3O, T3v);
|
cannam@127
|
577 }
|
cannam@127
|
578 {
|
cannam@127
|
579 E T1E, T1N, T26, T2t;
|
cannam@127
|
580 T2X = FNMS(KP707106781, T1D, T1s);
|
cannam@127
|
581 T1E = FMA(KP707106781, T1D, T1s);
|
cannam@127
|
582 T1N = FMA(KP707106781, T1M, T1J);
|
cannam@127
|
583 T2W = FNMS(KP707106781, T1M, T1J);
|
cannam@127
|
584 T30 = FNMS(KP707106781, T25, T1U);
|
cannam@127
|
585 T26 = FMA(KP707106781, T25, T1U);
|
cannam@127
|
586 T2t = T2h + T2s;
|
cannam@127
|
587 T34 = T2s - T2h;
|
cannam@127
|
588 T33 = FNMS(KP707106781, T2C, T2z);
|
cannam@127
|
589 T2D = FMA(KP707106781, T2C, T2z);
|
cannam@127
|
590 T2M = FMA(KP198912367, T1E, T1N);
|
cannam@127
|
591 T1O = FNMS(KP198912367, T1N, T1E);
|
cannam@127
|
592 T2P = FNMS(KP923879532, T2t, T26);
|
cannam@127
|
593 T2u = FMA(KP923879532, T2t, T26);
|
cannam@127
|
594 T2G = T2E + T2F;
|
cannam@127
|
595 T31 = T2E - T2F;
|
cannam@127
|
596 }
|
cannam@127
|
597 T2J = FNMS(KP1_961570560, T1O, T1n);
|
cannam@127
|
598 T1P = FMA(KP1_961570560, T1O, T1n);
|
cannam@127
|
599 T2R = FMA(KP1_961570560, T2M, T2L);
|
cannam@127
|
600 T2N = FNMS(KP1_961570560, T2M, T2L);
|
cannam@127
|
601 T2O = FNMS(KP923879532, T2G, T2D);
|
cannam@127
|
602 T2H = FMA(KP923879532, T2G, T2D);
|
cannam@127
|
603 }
|
cannam@127
|
604 {
|
cannam@127
|
605 E T4j, T48, T4d, T45, T4l, T4h, T4i, T4b;
|
cannam@127
|
606 {
|
cannam@127
|
607 E T4f, T41, T4g, T44;
|
cannam@127
|
608 T4f = FMA(KP1_847759065, T40, T3Z);
|
cannam@127
|
609 T41 = FNMS(KP1_847759065, T40, T3Z);
|
cannam@127
|
610 {
|
cannam@127
|
611 E T2Q, T2S, T2K, T2I;
|
cannam@127
|
612 T2Q = FNMS(KP820678790, T2P, T2O);
|
cannam@127
|
613 T2S = FMA(KP820678790, T2O, T2P);
|
cannam@127
|
614 T2K = FMA(KP098491403, T2u, T2H);
|
cannam@127
|
615 T2I = FNMS(KP098491403, T2H, T2u);
|
cannam@127
|
616 R1[WS(rs, 20)] = FMA(KP1_546020906, T2Q, T2N);
|
cannam@127
|
617 R1[WS(rs, 4)] = FNMS(KP1_546020906, T2Q, T2N);
|
cannam@127
|
618 R1[WS(rs, 28)] = FMA(KP1_546020906, T2S, T2R);
|
cannam@127
|
619 R1[WS(rs, 12)] = FNMS(KP1_546020906, T2S, T2R);
|
cannam@127
|
620 R1[WS(rs, 24)] = FMA(KP1_990369453, T2K, T2J);
|
cannam@127
|
621 R1[WS(rs, 8)] = FNMS(KP1_990369453, T2K, T2J);
|
cannam@127
|
622 R1[0] = FMA(KP1_990369453, T2I, T1P);
|
cannam@127
|
623 R1[WS(rs, 16)] = FNMS(KP1_990369453, T2I, T1P);
|
cannam@127
|
624 }
|
cannam@127
|
625 T4g = FMA(KP198912367, T42, T43);
|
cannam@127
|
626 T44 = FNMS(KP198912367, T43, T42);
|
cannam@127
|
627 T4j = FMA(KP923879532, T47, T46);
|
cannam@127
|
628 T48 = FNMS(KP923879532, T47, T46);
|
cannam@127
|
629 T4d = FMA(KP1_961570560, T44, T41);
|
cannam@127
|
630 T45 = FNMS(KP1_961570560, T44, T41);
|
cannam@127
|
631 T4l = FMA(KP1_961570560, T4g, T4f);
|
cannam@127
|
632 T4h = FNMS(KP1_961570560, T4g, T4f);
|
cannam@127
|
633 T4i = FMA(KP923879532, T4a, T49);
|
cannam@127
|
634 T4b = FNMS(KP923879532, T4a, T49);
|
cannam@127
|
635 }
|
cannam@127
|
636 {
|
cannam@127
|
637 E T39, T2V, T3a, T2Y;
|
cannam@127
|
638 T39 = FMA(KP1_847759065, T2U, T2T);
|
cannam@127
|
639 T2V = FNMS(KP1_847759065, T2U, T2T);
|
cannam@127
|
640 {
|
cannam@127
|
641 E T4k, T4m, T4e, T4c;
|
cannam@127
|
642 T4k = FNMS(KP098491403, T4j, T4i);
|
cannam@127
|
643 T4m = FMA(KP098491403, T4i, T4j);
|
cannam@127
|
644 T4e = FMA(KP820678790, T48, T4b);
|
cannam@127
|
645 T4c = FNMS(KP820678790, T4b, T48);
|
cannam@127
|
646 R1[WS(rs, 23)] = FMA(KP1_990369453, T4k, T4h);
|
cannam@127
|
647 R1[WS(rs, 7)] = FNMS(KP1_990369453, T4k, T4h);
|
cannam@127
|
648 R1[WS(rs, 31)] = FMA(KP1_990369453, T4m, T4l);
|
cannam@127
|
649 R1[WS(rs, 15)] = FNMS(KP1_990369453, T4m, T4l);
|
cannam@127
|
650 R1[WS(rs, 27)] = FMA(KP1_546020906, T4e, T4d);
|
cannam@127
|
651 R1[WS(rs, 11)] = FNMS(KP1_546020906, T4e, T4d);
|
cannam@127
|
652 R1[WS(rs, 3)] = FMA(KP1_546020906, T4c, T45);
|
cannam@127
|
653 R1[WS(rs, 19)] = FNMS(KP1_546020906, T4c, T45);
|
cannam@127
|
654 }
|
cannam@127
|
655 T3a = FMA(KP668178637, T2W, T2X);
|
cannam@127
|
656 T2Y = FNMS(KP668178637, T2X, T2W);
|
cannam@127
|
657 T3d = FMA(KP923879532, T31, T30);
|
cannam@127
|
658 T32 = FNMS(KP923879532, T31, T30);
|
cannam@127
|
659 T37 = FMA(KP1_662939224, T2Y, T2V);
|
cannam@127
|
660 T2Z = FNMS(KP1_662939224, T2Y, T2V);
|
cannam@127
|
661 T3f = FMA(KP1_662939224, T3a, T39);
|
cannam@127
|
662 T3b = FNMS(KP1_662939224, T3a, T39);
|
cannam@127
|
663 T3c = FMA(KP923879532, T34, T33);
|
cannam@127
|
664 T35 = FNMS(KP923879532, T34, T33);
|
cannam@127
|
665 }
|
cannam@127
|
666 }
|
cannam@127
|
667 }
|
cannam@127
|
668 }
|
cannam@127
|
669 }
|
cannam@127
|
670 }
|
cannam@127
|
671 {
|
cannam@127
|
672 E T3g, T3e, T36, T38;
|
cannam@127
|
673 T3g = FMA(KP303346683, T3c, T3d);
|
cannam@127
|
674 T3e = FNMS(KP303346683, T3d, T3c);
|
cannam@127
|
675 T36 = FNMS(KP534511135, T35, T32);
|
cannam@127
|
676 T38 = FMA(KP534511135, T32, T35);
|
cannam@127
|
677 R1[WS(rs, 22)] = FMA(KP1_913880671, T3e, T3b);
|
cannam@127
|
678 R1[WS(rs, 6)] = FNMS(KP1_913880671, T3e, T3b);
|
cannam@127
|
679 R1[WS(rs, 30)] = FMA(KP1_913880671, T3g, T3f);
|
cannam@127
|
680 R1[WS(rs, 14)] = FNMS(KP1_913880671, T3g, T3f);
|
cannam@127
|
681 R1[WS(rs, 26)] = FMA(KP1_763842528, T38, T37);
|
cannam@127
|
682 R1[WS(rs, 10)] = FNMS(KP1_763842528, T38, T37);
|
cannam@127
|
683 R1[WS(rs, 2)] = FMA(KP1_763842528, T36, T2Z);
|
cannam@127
|
684 R1[WS(rs, 18)] = FNMS(KP1_763842528, T36, T2Z);
|
cannam@127
|
685 }
|
cannam@127
|
686 }
|
cannam@127
|
687 }
|
cannam@127
|
688 }
|
cannam@127
|
689
|
cannam@127
|
690 static const kr2c_desc desc = { 64, "r2cb_64", {178, 0, 216, 0}, &GENUS };
|
cannam@127
|
691
|
cannam@127
|
692 void X(codelet_r2cb_64) (planner *p) {
|
cannam@127
|
693 X(kr2c_register) (p, r2cb_64, &desc);
|
cannam@127
|
694 }
|
cannam@127
|
695
|
cannam@127
|
696 #else /* HAVE_FMA */
|
cannam@127
|
697
|
cannam@127
|
698 /* Generated by: ../../../genfft/gen_r2cb.native -compact -variables 4 -pipeline-latency 4 -sign 1 -n 64 -name r2cb_64 -include r2cb.h */
|
cannam@127
|
699
|
cannam@127
|
700 /*
|
cannam@127
|
701 * This function contains 394 FP additions, 134 FP multiplications,
|
cannam@127
|
702 * (or, 342 additions, 82 multiplications, 52 fused multiply/add),
|
cannam@127
|
703 * 110 stack variables, 19 constants, and 128 memory accesses
|
cannam@127
|
704 */
|
cannam@127
|
705 #include "r2cb.h"
|
cannam@127
|
706
|
cannam@127
|
707 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@127
|
708 {
|
cannam@127
|
709 DK(KP1_268786568, +1.268786568327290996430343226450986741351374190);
|
cannam@127
|
710 DK(KP1_546020906, +1.546020906725473921621813219516939601942082586);
|
cannam@127
|
711 DK(KP196034280, +0.196034280659121203988391127777283691722273346);
|
cannam@127
|
712 DK(KP1_990369453, +1.990369453344393772489673906218959843150949737);
|
cannam@127
|
713 DK(KP942793473, +0.942793473651995297112775251810508755314920638);
|
cannam@127
|
714 DK(KP1_763842528, +1.763842528696710059425513727320776699016885241);
|
cannam@127
|
715 DK(KP580569354, +0.580569354508924735272384751634790549382952557);
|
cannam@127
|
716 DK(KP1_913880671, +1.913880671464417729871595773960539938965698411);
|
cannam@127
|
717 DK(KP1_111140466, +1.111140466039204449485661627897065748749874382);
|
cannam@127
|
718 DK(KP1_662939224, +1.662939224605090474157576755235811513477121624);
|
cannam@127
|
719 DK(KP390180644, +0.390180644032256535696569736954044481855383236);
|
cannam@127
|
720 DK(KP1_961570560, +1.961570560806460898252364472268478073947867462);
|
cannam@127
|
721 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
cannam@127
|
722 DK(KP382683432, +0.382683432365089771728459984030398866761344562);
|
cannam@127
|
723 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
cannam@127
|
724 DK(KP765366864, +0.765366864730179543456919968060797733522689125);
|
cannam@127
|
725 DK(KP1_847759065, +1.847759065022573512256366378793576573644833252);
|
cannam@127
|
726 DK(KP1_414213562, +1.414213562373095048801688724209698078569671875);
|
cannam@127
|
727 DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
|
cannam@127
|
728 {
|
cannam@127
|
729 INT i;
|
cannam@127
|
730 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@127
|
731 E Ta, T2S, T18, T2u, T3F, T4V, T5l, T61, Th, T2T, T1h, T2v, T3M, T4W, T5o;
|
cannam@127
|
732 E T62, T3Q, T5q, T5u, T44, Tp, Tw, T2V, T2W, T2X, T2Y, T3X, T5t, T1r, T2x;
|
cannam@127
|
733 E T41, T5r, T1A, T2y, T4a, T5y, T5N, T4H, TN, T31, T4E, T5z, T39, T3q, T1L;
|
cannam@127
|
734 E T2B, T4h, T5M, T2h, T2F, T12, T36, T5D, T5J, T5G, T5K, T1U, T26, T23, T27;
|
cannam@127
|
735 E T4p, T4z, T4w, T4A, T34, T3r;
|
cannam@127
|
736 {
|
cannam@127
|
737 E T5, T3A, T3, T3y, T9, T3C, T17, T3D, T6, T14;
|
cannam@127
|
738 {
|
cannam@127
|
739 E T4, T3z, T1, T2;
|
cannam@127
|
740 T4 = Cr[WS(csr, 16)];
|
cannam@127
|
741 T5 = KP2_000000000 * T4;
|
cannam@127
|
742 T3z = Ci[WS(csi, 16)];
|
cannam@127
|
743 T3A = KP2_000000000 * T3z;
|
cannam@127
|
744 T1 = Cr[0];
|
cannam@127
|
745 T2 = Cr[WS(csr, 32)];
|
cannam@127
|
746 T3 = T1 + T2;
|
cannam@127
|
747 T3y = T1 - T2;
|
cannam@127
|
748 {
|
cannam@127
|
749 E T7, T8, T15, T16;
|
cannam@127
|
750 T7 = Cr[WS(csr, 8)];
|
cannam@127
|
751 T8 = Cr[WS(csr, 24)];
|
cannam@127
|
752 T9 = KP2_000000000 * (T7 + T8);
|
cannam@127
|
753 T3C = T7 - T8;
|
cannam@127
|
754 T15 = Ci[WS(csi, 8)];
|
cannam@127
|
755 T16 = Ci[WS(csi, 24)];
|
cannam@127
|
756 T17 = KP2_000000000 * (T15 - T16);
|
cannam@127
|
757 T3D = T15 + T16;
|
cannam@127
|
758 }
|
cannam@127
|
759 }
|
cannam@127
|
760 T6 = T3 + T5;
|
cannam@127
|
761 Ta = T6 + T9;
|
cannam@127
|
762 T2S = T6 - T9;
|
cannam@127
|
763 T14 = T3 - T5;
|
cannam@127
|
764 T18 = T14 - T17;
|
cannam@127
|
765 T2u = T14 + T17;
|
cannam@127
|
766 {
|
cannam@127
|
767 E T3B, T3E, T5j, T5k;
|
cannam@127
|
768 T3B = T3y - T3A;
|
cannam@127
|
769 T3E = KP1_414213562 * (T3C - T3D);
|
cannam@127
|
770 T3F = T3B + T3E;
|
cannam@127
|
771 T4V = T3B - T3E;
|
cannam@127
|
772 T5j = T3y + T3A;
|
cannam@127
|
773 T5k = KP1_414213562 * (T3C + T3D);
|
cannam@127
|
774 T5l = T5j - T5k;
|
cannam@127
|
775 T61 = T5j + T5k;
|
cannam@127
|
776 }
|
cannam@127
|
777 }
|
cannam@127
|
778 {
|
cannam@127
|
779 E Td, T3G, T1c, T3K, Tg, T3J, T1f, T3H, T19, T1g;
|
cannam@127
|
780 {
|
cannam@127
|
781 E Tb, Tc, T1a, T1b;
|
cannam@127
|
782 Tb = Cr[WS(csr, 4)];
|
cannam@127
|
783 Tc = Cr[WS(csr, 28)];
|
cannam@127
|
784 Td = Tb + Tc;
|
cannam@127
|
785 T3G = Tb - Tc;
|
cannam@127
|
786 T1a = Ci[WS(csi, 4)];
|
cannam@127
|
787 T1b = Ci[WS(csi, 28)];
|
cannam@127
|
788 T1c = T1a - T1b;
|
cannam@127
|
789 T3K = T1a + T1b;
|
cannam@127
|
790 }
|
cannam@127
|
791 {
|
cannam@127
|
792 E Te, Tf, T1d, T1e;
|
cannam@127
|
793 Te = Cr[WS(csr, 20)];
|
cannam@127
|
794 Tf = Cr[WS(csr, 12)];
|
cannam@127
|
795 Tg = Te + Tf;
|
cannam@127
|
796 T3J = Te - Tf;
|
cannam@127
|
797 T1d = Ci[WS(csi, 20)];
|
cannam@127
|
798 T1e = Ci[WS(csi, 12)];
|
cannam@127
|
799 T1f = T1d - T1e;
|
cannam@127
|
800 T3H = T1d + T1e;
|
cannam@127
|
801 }
|
cannam@127
|
802 Th = KP2_000000000 * (Td + Tg);
|
cannam@127
|
803 T2T = KP2_000000000 * (T1f + T1c);
|
cannam@127
|
804 T19 = Td - Tg;
|
cannam@127
|
805 T1g = T1c - T1f;
|
cannam@127
|
806 T1h = KP1_414213562 * (T19 - T1g);
|
cannam@127
|
807 T2v = KP1_414213562 * (T19 + T1g);
|
cannam@127
|
808 {
|
cannam@127
|
809 E T3I, T3L, T5m, T5n;
|
cannam@127
|
810 T3I = T3G - T3H;
|
cannam@127
|
811 T3L = T3J + T3K;
|
cannam@127
|
812 T3M = FNMS(KP765366864, T3L, KP1_847759065 * T3I);
|
cannam@127
|
813 T4W = FMA(KP765366864, T3I, KP1_847759065 * T3L);
|
cannam@127
|
814 T5m = T3G + T3H;
|
cannam@127
|
815 T5n = T3K - T3J;
|
cannam@127
|
816 T5o = FNMS(KP1_847759065, T5n, KP765366864 * T5m);
|
cannam@127
|
817 T62 = FMA(KP1_847759065, T5m, KP765366864 * T5n);
|
cannam@127
|
818 }
|
cannam@127
|
819 }
|
cannam@127
|
820 {
|
cannam@127
|
821 E Tl, T3O, T1v, T43, To, T42, T1y, T3P, Ts, T3R, T1p, T3S, Tv, T3U, T1m;
|
cannam@127
|
822 E T3V;
|
cannam@127
|
823 {
|
cannam@127
|
824 E Tj, Tk, T1t, T1u;
|
cannam@127
|
825 Tj = Cr[WS(csr, 2)];
|
cannam@127
|
826 Tk = Cr[WS(csr, 30)];
|
cannam@127
|
827 Tl = Tj + Tk;
|
cannam@127
|
828 T3O = Tj - Tk;
|
cannam@127
|
829 T1t = Ci[WS(csi, 2)];
|
cannam@127
|
830 T1u = Ci[WS(csi, 30)];
|
cannam@127
|
831 T1v = T1t - T1u;
|
cannam@127
|
832 T43 = T1t + T1u;
|
cannam@127
|
833 }
|
cannam@127
|
834 {
|
cannam@127
|
835 E Tm, Tn, T1w, T1x;
|
cannam@127
|
836 Tm = Cr[WS(csr, 18)];
|
cannam@127
|
837 Tn = Cr[WS(csr, 14)];
|
cannam@127
|
838 To = Tm + Tn;
|
cannam@127
|
839 T42 = Tm - Tn;
|
cannam@127
|
840 T1w = Ci[WS(csi, 18)];
|
cannam@127
|
841 T1x = Ci[WS(csi, 14)];
|
cannam@127
|
842 T1y = T1w - T1x;
|
cannam@127
|
843 T3P = T1w + T1x;
|
cannam@127
|
844 }
|
cannam@127
|
845 {
|
cannam@127
|
846 E Tq, Tr, T1n, T1o;
|
cannam@127
|
847 Tq = Cr[WS(csr, 10)];
|
cannam@127
|
848 Tr = Cr[WS(csr, 22)];
|
cannam@127
|
849 Ts = Tq + Tr;
|
cannam@127
|
850 T3R = Tq - Tr;
|
cannam@127
|
851 T1n = Ci[WS(csi, 10)];
|
cannam@127
|
852 T1o = Ci[WS(csi, 22)];
|
cannam@127
|
853 T1p = T1n - T1o;
|
cannam@127
|
854 T3S = T1n + T1o;
|
cannam@127
|
855 }
|
cannam@127
|
856 {
|
cannam@127
|
857 E Tt, Tu, T1k, T1l;
|
cannam@127
|
858 Tt = Cr[WS(csr, 6)];
|
cannam@127
|
859 Tu = Cr[WS(csr, 26)];
|
cannam@127
|
860 Tv = Tt + Tu;
|
cannam@127
|
861 T3U = Tt - Tu;
|
cannam@127
|
862 T1k = Ci[WS(csi, 26)];
|
cannam@127
|
863 T1l = Ci[WS(csi, 6)];
|
cannam@127
|
864 T1m = T1k - T1l;
|
cannam@127
|
865 T3V = T1l + T1k;
|
cannam@127
|
866 }
|
cannam@127
|
867 T3Q = T3O - T3P;
|
cannam@127
|
868 T5q = T3O + T3P;
|
cannam@127
|
869 T5u = T43 - T42;
|
cannam@127
|
870 T44 = T42 + T43;
|
cannam@127
|
871 Tp = Tl + To;
|
cannam@127
|
872 Tw = Ts + Tv;
|
cannam@127
|
873 T2V = Tp - Tw;
|
cannam@127
|
874 {
|
cannam@127
|
875 E T3T, T3W, T1j, T1q;
|
cannam@127
|
876 T2W = T1y + T1v;
|
cannam@127
|
877 T2X = T1p + T1m;
|
cannam@127
|
878 T2Y = T2W - T2X;
|
cannam@127
|
879 T3T = T3R - T3S;
|
cannam@127
|
880 T3W = T3U - T3V;
|
cannam@127
|
881 T3X = KP707106781 * (T3T + T3W);
|
cannam@127
|
882 T5t = KP707106781 * (T3T - T3W);
|
cannam@127
|
883 T1j = Tl - To;
|
cannam@127
|
884 T1q = T1m - T1p;
|
cannam@127
|
885 T1r = T1j + T1q;
|
cannam@127
|
886 T2x = T1j - T1q;
|
cannam@127
|
887 {
|
cannam@127
|
888 E T3Z, T40, T1s, T1z;
|
cannam@127
|
889 T3Z = T3R + T3S;
|
cannam@127
|
890 T40 = T3U + T3V;
|
cannam@127
|
891 T41 = KP707106781 * (T3Z - T40);
|
cannam@127
|
892 T5r = KP707106781 * (T3Z + T40);
|
cannam@127
|
893 T1s = Ts - Tv;
|
cannam@127
|
894 T1z = T1v - T1y;
|
cannam@127
|
895 T1A = T1s + T1z;
|
cannam@127
|
896 T2y = T1z - T1s;
|
cannam@127
|
897 }
|
cannam@127
|
898 }
|
cannam@127
|
899 }
|
cannam@127
|
900 {
|
cannam@127
|
901 E TB, T48, T2c, T4G, TE, T4F, T2f, T49, TI, T4b, T1J, T4c, TL, T4e, T1G;
|
cannam@127
|
902 E T4f;
|
cannam@127
|
903 {
|
cannam@127
|
904 E Tz, TA, T2a, T2b;
|
cannam@127
|
905 Tz = Cr[WS(csr, 1)];
|
cannam@127
|
906 TA = Cr[WS(csr, 31)];
|
cannam@127
|
907 TB = Tz + TA;
|
cannam@127
|
908 T48 = Tz - TA;
|
cannam@127
|
909 T2a = Ci[WS(csi, 1)];
|
cannam@127
|
910 T2b = Ci[WS(csi, 31)];
|
cannam@127
|
911 T2c = T2a - T2b;
|
cannam@127
|
912 T4G = T2a + T2b;
|
cannam@127
|
913 }
|
cannam@127
|
914 {
|
cannam@127
|
915 E TC, TD, T2d, T2e;
|
cannam@127
|
916 TC = Cr[WS(csr, 17)];
|
cannam@127
|
917 TD = Cr[WS(csr, 15)];
|
cannam@127
|
918 TE = TC + TD;
|
cannam@127
|
919 T4F = TC - TD;
|
cannam@127
|
920 T2d = Ci[WS(csi, 17)];
|
cannam@127
|
921 T2e = Ci[WS(csi, 15)];
|
cannam@127
|
922 T2f = T2d - T2e;
|
cannam@127
|
923 T49 = T2d + T2e;
|
cannam@127
|
924 }
|
cannam@127
|
925 {
|
cannam@127
|
926 E TG, TH, T1H, T1I;
|
cannam@127
|
927 TG = Cr[WS(csr, 9)];
|
cannam@127
|
928 TH = Cr[WS(csr, 23)];
|
cannam@127
|
929 TI = TG + TH;
|
cannam@127
|
930 T4b = TG - TH;
|
cannam@127
|
931 T1H = Ci[WS(csi, 9)];
|
cannam@127
|
932 T1I = Ci[WS(csi, 23)];
|
cannam@127
|
933 T1J = T1H - T1I;
|
cannam@127
|
934 T4c = T1H + T1I;
|
cannam@127
|
935 }
|
cannam@127
|
936 {
|
cannam@127
|
937 E TJ, TK, T1E, T1F;
|
cannam@127
|
938 TJ = Cr[WS(csr, 7)];
|
cannam@127
|
939 TK = Cr[WS(csr, 25)];
|
cannam@127
|
940 TL = TJ + TK;
|
cannam@127
|
941 T4e = TJ - TK;
|
cannam@127
|
942 T1E = Ci[WS(csi, 25)];
|
cannam@127
|
943 T1F = Ci[WS(csi, 7)];
|
cannam@127
|
944 T1G = T1E - T1F;
|
cannam@127
|
945 T4f = T1F + T1E;
|
cannam@127
|
946 }
|
cannam@127
|
947 {
|
cannam@127
|
948 E TF, TM, T1D, T1K;
|
cannam@127
|
949 T4a = T48 - T49;
|
cannam@127
|
950 T5y = T48 + T49;
|
cannam@127
|
951 T5N = T4G - T4F;
|
cannam@127
|
952 T4H = T4F + T4G;
|
cannam@127
|
953 TF = TB + TE;
|
cannam@127
|
954 TM = TI + TL;
|
cannam@127
|
955 TN = TF + TM;
|
cannam@127
|
956 T31 = TF - TM;
|
cannam@127
|
957 {
|
cannam@127
|
958 E T4C, T4D, T37, T38;
|
cannam@127
|
959 T4C = T4b + T4c;
|
cannam@127
|
960 T4D = T4e + T4f;
|
cannam@127
|
961 T4E = KP707106781 * (T4C - T4D);
|
cannam@127
|
962 T5z = KP707106781 * (T4C + T4D);
|
cannam@127
|
963 T37 = T2f + T2c;
|
cannam@127
|
964 T38 = T1J + T1G;
|
cannam@127
|
965 T39 = T37 - T38;
|
cannam@127
|
966 T3q = T38 + T37;
|
cannam@127
|
967 }
|
cannam@127
|
968 T1D = TB - TE;
|
cannam@127
|
969 T1K = T1G - T1J;
|
cannam@127
|
970 T1L = T1D + T1K;
|
cannam@127
|
971 T2B = T1D - T1K;
|
cannam@127
|
972 {
|
cannam@127
|
973 E T4d, T4g, T29, T2g;
|
cannam@127
|
974 T4d = T4b - T4c;
|
cannam@127
|
975 T4g = T4e - T4f;
|
cannam@127
|
976 T4h = KP707106781 * (T4d + T4g);
|
cannam@127
|
977 T5M = KP707106781 * (T4d - T4g);
|
cannam@127
|
978 T29 = TI - TL;
|
cannam@127
|
979 T2g = T2c - T2f;
|
cannam@127
|
980 T2h = T29 + T2g;
|
cannam@127
|
981 T2F = T2g - T29;
|
cannam@127
|
982 }
|
cannam@127
|
983 }
|
cannam@127
|
984 }
|
cannam@127
|
985 {
|
cannam@127
|
986 E TQ, T4j, T1P, T4n, TT, T4m, T1S, T4k, TX, T4q, T1Y, T4u, T10, T4t, T21;
|
cannam@127
|
987 E T4r;
|
cannam@127
|
988 {
|
cannam@127
|
989 E TO, TP, T1N, T1O;
|
cannam@127
|
990 TO = Cr[WS(csr, 5)];
|
cannam@127
|
991 TP = Cr[WS(csr, 27)];
|
cannam@127
|
992 TQ = TO + TP;
|
cannam@127
|
993 T4j = TO - TP;
|
cannam@127
|
994 T1N = Ci[WS(csi, 5)];
|
cannam@127
|
995 T1O = Ci[WS(csi, 27)];
|
cannam@127
|
996 T1P = T1N - T1O;
|
cannam@127
|
997 T4n = T1N + T1O;
|
cannam@127
|
998 }
|
cannam@127
|
999 {
|
cannam@127
|
1000 E TR, TS, T1Q, T1R;
|
cannam@127
|
1001 TR = Cr[WS(csr, 21)];
|
cannam@127
|
1002 TS = Cr[WS(csr, 11)];
|
cannam@127
|
1003 TT = TR + TS;
|
cannam@127
|
1004 T4m = TR - TS;
|
cannam@127
|
1005 T1Q = Ci[WS(csi, 21)];
|
cannam@127
|
1006 T1R = Ci[WS(csi, 11)];
|
cannam@127
|
1007 T1S = T1Q - T1R;
|
cannam@127
|
1008 T4k = T1Q + T1R;
|
cannam@127
|
1009 }
|
cannam@127
|
1010 {
|
cannam@127
|
1011 E TV, TW, T1W, T1X;
|
cannam@127
|
1012 TV = Cr[WS(csr, 3)];
|
cannam@127
|
1013 TW = Cr[WS(csr, 29)];
|
cannam@127
|
1014 TX = TV + TW;
|
cannam@127
|
1015 T4q = TV - TW;
|
cannam@127
|
1016 T1W = Ci[WS(csi, 29)];
|
cannam@127
|
1017 T1X = Ci[WS(csi, 3)];
|
cannam@127
|
1018 T1Y = T1W - T1X;
|
cannam@127
|
1019 T4u = T1X + T1W;
|
cannam@127
|
1020 }
|
cannam@127
|
1021 {
|
cannam@127
|
1022 E TY, TZ, T1Z, T20;
|
cannam@127
|
1023 TY = Cr[WS(csr, 13)];
|
cannam@127
|
1024 TZ = Cr[WS(csr, 19)];
|
cannam@127
|
1025 T10 = TY + TZ;
|
cannam@127
|
1026 T4t = TY - TZ;
|
cannam@127
|
1027 T1Z = Ci[WS(csi, 13)];
|
cannam@127
|
1028 T20 = Ci[WS(csi, 19)];
|
cannam@127
|
1029 T21 = T1Z - T20;
|
cannam@127
|
1030 T4r = T1Z + T20;
|
cannam@127
|
1031 }
|
cannam@127
|
1032 {
|
cannam@127
|
1033 E TU, T11, T5B, T5C;
|
cannam@127
|
1034 TU = TQ + TT;
|
cannam@127
|
1035 T11 = TX + T10;
|
cannam@127
|
1036 T12 = TU + T11;
|
cannam@127
|
1037 T36 = TU - T11;
|
cannam@127
|
1038 T5B = T4j + T4k;
|
cannam@127
|
1039 T5C = T4n - T4m;
|
cannam@127
|
1040 T5D = FNMS(KP923879532, T5C, KP382683432 * T5B);
|
cannam@127
|
1041 T5J = FMA(KP923879532, T5B, KP382683432 * T5C);
|
cannam@127
|
1042 }
|
cannam@127
|
1043 {
|
cannam@127
|
1044 E T5E, T5F, T1M, T1T;
|
cannam@127
|
1045 T5E = T4q + T4r;
|
cannam@127
|
1046 T5F = T4t + T4u;
|
cannam@127
|
1047 T5G = FNMS(KP923879532, T5F, KP382683432 * T5E);
|
cannam@127
|
1048 T5K = FMA(KP923879532, T5E, KP382683432 * T5F);
|
cannam@127
|
1049 T1M = TQ - TT;
|
cannam@127
|
1050 T1T = T1P - T1S;
|
cannam@127
|
1051 T1U = T1M - T1T;
|
cannam@127
|
1052 T26 = T1M + T1T;
|
cannam@127
|
1053 }
|
cannam@127
|
1054 {
|
cannam@127
|
1055 E T1V, T22, T4l, T4o;
|
cannam@127
|
1056 T1V = TX - T10;
|
cannam@127
|
1057 T22 = T1Y - T21;
|
cannam@127
|
1058 T23 = T1V + T22;
|
cannam@127
|
1059 T27 = T22 - T1V;
|
cannam@127
|
1060 T4l = T4j - T4k;
|
cannam@127
|
1061 T4o = T4m + T4n;
|
cannam@127
|
1062 T4p = FNMS(KP382683432, T4o, KP923879532 * T4l);
|
cannam@127
|
1063 T4z = FMA(KP382683432, T4l, KP923879532 * T4o);
|
cannam@127
|
1064 }
|
cannam@127
|
1065 {
|
cannam@127
|
1066 E T4s, T4v, T32, T33;
|
cannam@127
|
1067 T4s = T4q - T4r;
|
cannam@127
|
1068 T4v = T4t - T4u;
|
cannam@127
|
1069 T4w = FMA(KP923879532, T4s, KP382683432 * T4v);
|
cannam@127
|
1070 T4A = FNMS(KP382683432, T4s, KP923879532 * T4v);
|
cannam@127
|
1071 T32 = T21 + T1Y;
|
cannam@127
|
1072 T33 = T1S + T1P;
|
cannam@127
|
1073 T34 = T32 - T33;
|
cannam@127
|
1074 T3r = T33 + T32;
|
cannam@127
|
1075 }
|
cannam@127
|
1076 }
|
cannam@127
|
1077 {
|
cannam@127
|
1078 E T13, T3x, Ty, T3w, Ti, Tx;
|
cannam@127
|
1079 T13 = KP2_000000000 * (TN + T12);
|
cannam@127
|
1080 T3x = KP2_000000000 * (T3r + T3q);
|
cannam@127
|
1081 Ti = Ta + Th;
|
cannam@127
|
1082 Tx = KP2_000000000 * (Tp + Tw);
|
cannam@127
|
1083 Ty = Ti + Tx;
|
cannam@127
|
1084 T3w = Ti - Tx;
|
cannam@127
|
1085 R0[WS(rs, 16)] = Ty - T13;
|
cannam@127
|
1086 R0[WS(rs, 24)] = T3w + T3x;
|
cannam@127
|
1087 R0[0] = Ty + T13;
|
cannam@127
|
1088 R0[WS(rs, 8)] = T3w - T3x;
|
cannam@127
|
1089 }
|
cannam@127
|
1090 {
|
cannam@127
|
1091 E T3g, T3k, T3j, T3l;
|
cannam@127
|
1092 {
|
cannam@127
|
1093 E T3e, T3f, T3h, T3i;
|
cannam@127
|
1094 T3e = T2S + T2T;
|
cannam@127
|
1095 T3f = KP1_414213562 * (T2V + T2Y);
|
cannam@127
|
1096 T3g = T3e - T3f;
|
cannam@127
|
1097 T3k = T3e + T3f;
|
cannam@127
|
1098 T3h = T31 - T34;
|
cannam@127
|
1099 T3i = T39 - T36;
|
cannam@127
|
1100 T3j = FNMS(KP1_847759065, T3i, KP765366864 * T3h);
|
cannam@127
|
1101 T3l = FMA(KP1_847759065, T3h, KP765366864 * T3i);
|
cannam@127
|
1102 }
|
cannam@127
|
1103 R0[WS(rs, 22)] = T3g - T3j;
|
cannam@127
|
1104 R0[WS(rs, 30)] = T3k + T3l;
|
cannam@127
|
1105 R0[WS(rs, 6)] = T3g + T3j;
|
cannam@127
|
1106 R0[WS(rs, 14)] = T3k - T3l;
|
cannam@127
|
1107 }
|
cannam@127
|
1108 {
|
cannam@127
|
1109 E T3o, T3u, T3t, T3v;
|
cannam@127
|
1110 {
|
cannam@127
|
1111 E T3m, T3n, T3p, T3s;
|
cannam@127
|
1112 T3m = Ta - Th;
|
cannam@127
|
1113 T3n = KP2_000000000 * (T2X + T2W);
|
cannam@127
|
1114 T3o = T3m - T3n;
|
cannam@127
|
1115 T3u = T3m + T3n;
|
cannam@127
|
1116 T3p = TN - T12;
|
cannam@127
|
1117 T3s = T3q - T3r;
|
cannam@127
|
1118 T3t = KP1_414213562 * (T3p - T3s);
|
cannam@127
|
1119 T3v = KP1_414213562 * (T3p + T3s);
|
cannam@127
|
1120 }
|
cannam@127
|
1121 R0[WS(rs, 20)] = T3o - T3t;
|
cannam@127
|
1122 R0[WS(rs, 28)] = T3u + T3v;
|
cannam@127
|
1123 R0[WS(rs, 4)] = T3o + T3t;
|
cannam@127
|
1124 R0[WS(rs, 12)] = T3u - T3v;
|
cannam@127
|
1125 }
|
cannam@127
|
1126 {
|
cannam@127
|
1127 E T30, T3c, T3b, T3d;
|
cannam@127
|
1128 {
|
cannam@127
|
1129 E T2U, T2Z, T35, T3a;
|
cannam@127
|
1130 T2U = T2S - T2T;
|
cannam@127
|
1131 T2Z = KP1_414213562 * (T2V - T2Y);
|
cannam@127
|
1132 T30 = T2U + T2Z;
|
cannam@127
|
1133 T3c = T2U - T2Z;
|
cannam@127
|
1134 T35 = T31 + T34;
|
cannam@127
|
1135 T3a = T36 + T39;
|
cannam@127
|
1136 T3b = FNMS(KP765366864, T3a, KP1_847759065 * T35);
|
cannam@127
|
1137 T3d = FMA(KP765366864, T35, KP1_847759065 * T3a);
|
cannam@127
|
1138 }
|
cannam@127
|
1139 R0[WS(rs, 18)] = T30 - T3b;
|
cannam@127
|
1140 R0[WS(rs, 26)] = T3c + T3d;
|
cannam@127
|
1141 R0[WS(rs, 2)] = T30 + T3b;
|
cannam@127
|
1142 R0[WS(rs, 10)] = T3c - T3d;
|
cannam@127
|
1143 }
|
cannam@127
|
1144 {
|
cannam@127
|
1145 E T25, T2p, T2i, T2q, T1C, T2k, T2o, T2s, T24, T28;
|
cannam@127
|
1146 T24 = KP707106781 * (T1U + T23);
|
cannam@127
|
1147 T25 = T1L + T24;
|
cannam@127
|
1148 T2p = T1L - T24;
|
cannam@127
|
1149 T28 = KP707106781 * (T26 + T27);
|
cannam@127
|
1150 T2i = T28 + T2h;
|
cannam@127
|
1151 T2q = T2h - T28;
|
cannam@127
|
1152 {
|
cannam@127
|
1153 E T1i, T1B, T2m, T2n;
|
cannam@127
|
1154 T1i = T18 + T1h;
|
cannam@127
|
1155 T1B = FNMS(KP765366864, T1A, KP1_847759065 * T1r);
|
cannam@127
|
1156 T1C = T1i + T1B;
|
cannam@127
|
1157 T2k = T1i - T1B;
|
cannam@127
|
1158 T2m = T18 - T1h;
|
cannam@127
|
1159 T2n = FMA(KP765366864, T1r, KP1_847759065 * T1A);
|
cannam@127
|
1160 T2o = T2m - T2n;
|
cannam@127
|
1161 T2s = T2m + T2n;
|
cannam@127
|
1162 }
|
cannam@127
|
1163 {
|
cannam@127
|
1164 E T2j, T2t, T2l, T2r;
|
cannam@127
|
1165 T2j = FNMS(KP390180644, T2i, KP1_961570560 * T25);
|
cannam@127
|
1166 R0[WS(rs, 17)] = T1C - T2j;
|
cannam@127
|
1167 R0[WS(rs, 1)] = T1C + T2j;
|
cannam@127
|
1168 T2t = FMA(KP1_662939224, T2p, KP1_111140466 * T2q);
|
cannam@127
|
1169 R0[WS(rs, 13)] = T2s - T2t;
|
cannam@127
|
1170 R0[WS(rs, 29)] = T2s + T2t;
|
cannam@127
|
1171 T2l = FMA(KP390180644, T25, KP1_961570560 * T2i);
|
cannam@127
|
1172 R0[WS(rs, 9)] = T2k - T2l;
|
cannam@127
|
1173 R0[WS(rs, 25)] = T2k + T2l;
|
cannam@127
|
1174 T2r = FNMS(KP1_662939224, T2q, KP1_111140466 * T2p);
|
cannam@127
|
1175 R0[WS(rs, 21)] = T2o - T2r;
|
cannam@127
|
1176 R0[WS(rs, 5)] = T2o + T2r;
|
cannam@127
|
1177 }
|
cannam@127
|
1178 }
|
cannam@127
|
1179 {
|
cannam@127
|
1180 E T2D, T2N, T2G, T2O, T2A, T2I, T2M, T2Q, T2C, T2E;
|
cannam@127
|
1181 T2C = KP707106781 * (T27 - T26);
|
cannam@127
|
1182 T2D = T2B + T2C;
|
cannam@127
|
1183 T2N = T2B - T2C;
|
cannam@127
|
1184 T2E = KP707106781 * (T1U - T23);
|
cannam@127
|
1185 T2G = T2E + T2F;
|
cannam@127
|
1186 T2O = T2F - T2E;
|
cannam@127
|
1187 {
|
cannam@127
|
1188 E T2w, T2z, T2K, T2L;
|
cannam@127
|
1189 T2w = T2u - T2v;
|
cannam@127
|
1190 T2z = FNMS(KP1_847759065, T2y, KP765366864 * T2x);
|
cannam@127
|
1191 T2A = T2w + T2z;
|
cannam@127
|
1192 T2I = T2w - T2z;
|
cannam@127
|
1193 T2K = T2u + T2v;
|
cannam@127
|
1194 T2L = FMA(KP1_847759065, T2x, KP765366864 * T2y);
|
cannam@127
|
1195 T2M = T2K - T2L;
|
cannam@127
|
1196 T2Q = T2K + T2L;
|
cannam@127
|
1197 }
|
cannam@127
|
1198 {
|
cannam@127
|
1199 E T2H, T2R, T2J, T2P;
|
cannam@127
|
1200 T2H = FNMS(KP1_111140466, T2G, KP1_662939224 * T2D);
|
cannam@127
|
1201 R0[WS(rs, 19)] = T2A - T2H;
|
cannam@127
|
1202 R0[WS(rs, 3)] = T2A + T2H;
|
cannam@127
|
1203 T2R = FMA(KP1_961570560, T2N, KP390180644 * T2O);
|
cannam@127
|
1204 R0[WS(rs, 15)] = T2Q - T2R;
|
cannam@127
|
1205 R0[WS(rs, 31)] = T2Q + T2R;
|
cannam@127
|
1206 T2J = FMA(KP1_111140466, T2D, KP1_662939224 * T2G);
|
cannam@127
|
1207 R0[WS(rs, 11)] = T2I - T2J;
|
cannam@127
|
1208 R0[WS(rs, 27)] = T2I + T2J;
|
cannam@127
|
1209 T2P = FNMS(KP1_961570560, T2O, KP390180644 * T2N);
|
cannam@127
|
1210 R0[WS(rs, 23)] = T2M - T2P;
|
cannam@127
|
1211 R0[WS(rs, 7)] = T2M + T2P;
|
cannam@127
|
1212 }
|
cannam@127
|
1213 }
|
cannam@127
|
1214 {
|
cannam@127
|
1215 E T5p, T5T, T5w, T5U, T5I, T5W, T5P, T5X, T5s, T5v;
|
cannam@127
|
1216 T5p = T5l + T5o;
|
cannam@127
|
1217 T5T = T5l - T5o;
|
cannam@127
|
1218 T5s = T5q - T5r;
|
cannam@127
|
1219 T5v = T5t + T5u;
|
cannam@127
|
1220 T5w = FNMS(KP1_111140466, T5v, KP1_662939224 * T5s);
|
cannam@127
|
1221 T5U = FMA(KP1_111140466, T5s, KP1_662939224 * T5v);
|
cannam@127
|
1222 {
|
cannam@127
|
1223 E T5A, T5H, T5L, T5O;
|
cannam@127
|
1224 T5A = T5y - T5z;
|
cannam@127
|
1225 T5H = T5D + T5G;
|
cannam@127
|
1226 T5I = T5A + T5H;
|
cannam@127
|
1227 T5W = T5A - T5H;
|
cannam@127
|
1228 T5L = T5J - T5K;
|
cannam@127
|
1229 T5O = T5M + T5N;
|
cannam@127
|
1230 T5P = T5L + T5O;
|
cannam@127
|
1231 T5X = T5O - T5L;
|
cannam@127
|
1232 }
|
cannam@127
|
1233 {
|
cannam@127
|
1234 E T5x, T5Q, T5Z, T60;
|
cannam@127
|
1235 T5x = T5p + T5w;
|
cannam@127
|
1236 T5Q = FNMS(KP580569354, T5P, KP1_913880671 * T5I);
|
cannam@127
|
1237 R1[WS(rs, 17)] = T5x - T5Q;
|
cannam@127
|
1238 R1[WS(rs, 1)] = T5x + T5Q;
|
cannam@127
|
1239 T5Z = T5T + T5U;
|
cannam@127
|
1240 T60 = FMA(KP1_763842528, T5W, KP942793473 * T5X);
|
cannam@127
|
1241 R1[WS(rs, 13)] = T5Z - T60;
|
cannam@127
|
1242 R1[WS(rs, 29)] = T5Z + T60;
|
cannam@127
|
1243 }
|
cannam@127
|
1244 {
|
cannam@127
|
1245 E T5R, T5S, T5V, T5Y;
|
cannam@127
|
1246 T5R = T5p - T5w;
|
cannam@127
|
1247 T5S = FMA(KP580569354, T5I, KP1_913880671 * T5P);
|
cannam@127
|
1248 R1[WS(rs, 9)] = T5R - T5S;
|
cannam@127
|
1249 R1[WS(rs, 25)] = T5R + T5S;
|
cannam@127
|
1250 T5V = T5T - T5U;
|
cannam@127
|
1251 T5Y = FNMS(KP1_763842528, T5X, KP942793473 * T5W);
|
cannam@127
|
1252 R1[WS(rs, 21)] = T5V - T5Y;
|
cannam@127
|
1253 R1[WS(rs, 5)] = T5V + T5Y;
|
cannam@127
|
1254 }
|
cannam@127
|
1255 }
|
cannam@127
|
1256 {
|
cannam@127
|
1257 E T3N, T4N, T46, T4O, T4y, T4Q, T4J, T4R, T3Y, T45;
|
cannam@127
|
1258 T3N = T3F + T3M;
|
cannam@127
|
1259 T4N = T3F - T3M;
|
cannam@127
|
1260 T3Y = T3Q + T3X;
|
cannam@127
|
1261 T45 = T41 + T44;
|
cannam@127
|
1262 T46 = FNMS(KP390180644, T45, KP1_961570560 * T3Y);
|
cannam@127
|
1263 T4O = FMA(KP390180644, T3Y, KP1_961570560 * T45);
|
cannam@127
|
1264 {
|
cannam@127
|
1265 E T4i, T4x, T4B, T4I;
|
cannam@127
|
1266 T4i = T4a + T4h;
|
cannam@127
|
1267 T4x = T4p + T4w;
|
cannam@127
|
1268 T4y = T4i + T4x;
|
cannam@127
|
1269 T4Q = T4i - T4x;
|
cannam@127
|
1270 T4B = T4z + T4A;
|
cannam@127
|
1271 T4I = T4E + T4H;
|
cannam@127
|
1272 T4J = T4B + T4I;
|
cannam@127
|
1273 T4R = T4I - T4B;
|
cannam@127
|
1274 }
|
cannam@127
|
1275 {
|
cannam@127
|
1276 E T47, T4K, T4T, T4U;
|
cannam@127
|
1277 T47 = T3N + T46;
|
cannam@127
|
1278 T4K = FNMS(KP196034280, T4J, KP1_990369453 * T4y);
|
cannam@127
|
1279 R1[WS(rs, 16)] = T47 - T4K;
|
cannam@127
|
1280 R1[0] = T47 + T4K;
|
cannam@127
|
1281 T4T = T4N + T4O;
|
cannam@127
|
1282 T4U = FMA(KP1_546020906, T4Q, KP1_268786568 * T4R);
|
cannam@127
|
1283 R1[WS(rs, 12)] = T4T - T4U;
|
cannam@127
|
1284 R1[WS(rs, 28)] = T4T + T4U;
|
cannam@127
|
1285 }
|
cannam@127
|
1286 {
|
cannam@127
|
1287 E T4L, T4M, T4P, T4S;
|
cannam@127
|
1288 T4L = T3N - T46;
|
cannam@127
|
1289 T4M = FMA(KP196034280, T4y, KP1_990369453 * T4J);
|
cannam@127
|
1290 R1[WS(rs, 8)] = T4L - T4M;
|
cannam@127
|
1291 R1[WS(rs, 24)] = T4L + T4M;
|
cannam@127
|
1292 T4P = T4N - T4O;
|
cannam@127
|
1293 T4S = FNMS(KP1_546020906, T4R, KP1_268786568 * T4Q);
|
cannam@127
|
1294 R1[WS(rs, 20)] = T4P - T4S;
|
cannam@127
|
1295 R1[WS(rs, 4)] = T4P + T4S;
|
cannam@127
|
1296 }
|
cannam@127
|
1297 }
|
cannam@127
|
1298 {
|
cannam@127
|
1299 E T63, T6h, T66, T6i, T6a, T6k, T6d, T6l, T64, T65;
|
cannam@127
|
1300 T63 = T61 - T62;
|
cannam@127
|
1301 T6h = T61 + T62;
|
cannam@127
|
1302 T64 = T5q + T5r;
|
cannam@127
|
1303 T65 = T5u - T5t;
|
cannam@127
|
1304 T66 = FNMS(KP1_961570560, T65, KP390180644 * T64);
|
cannam@127
|
1305 T6i = FMA(KP1_961570560, T64, KP390180644 * T65);
|
cannam@127
|
1306 {
|
cannam@127
|
1307 E T68, T69, T6b, T6c;
|
cannam@127
|
1308 T68 = T5y + T5z;
|
cannam@127
|
1309 T69 = T5J + T5K;
|
cannam@127
|
1310 T6a = T68 - T69;
|
cannam@127
|
1311 T6k = T68 + T69;
|
cannam@127
|
1312 T6b = T5D - T5G;
|
cannam@127
|
1313 T6c = T5N - T5M;
|
cannam@127
|
1314 T6d = T6b + T6c;
|
cannam@127
|
1315 T6l = T6c - T6b;
|
cannam@127
|
1316 }
|
cannam@127
|
1317 {
|
cannam@127
|
1318 E T67, T6e, T6n, T6o;
|
cannam@127
|
1319 T67 = T63 + T66;
|
cannam@127
|
1320 T6e = FNMS(KP1_268786568, T6d, KP1_546020906 * T6a);
|
cannam@127
|
1321 R1[WS(rs, 19)] = T67 - T6e;
|
cannam@127
|
1322 R1[WS(rs, 3)] = T67 + T6e;
|
cannam@127
|
1323 T6n = T6h + T6i;
|
cannam@127
|
1324 T6o = FMA(KP1_990369453, T6k, KP196034280 * T6l);
|
cannam@127
|
1325 R1[WS(rs, 15)] = T6n - T6o;
|
cannam@127
|
1326 R1[WS(rs, 31)] = T6n + T6o;
|
cannam@127
|
1327 }
|
cannam@127
|
1328 {
|
cannam@127
|
1329 E T6f, T6g, T6j, T6m;
|
cannam@127
|
1330 T6f = T63 - T66;
|
cannam@127
|
1331 T6g = FMA(KP1_268786568, T6a, KP1_546020906 * T6d);
|
cannam@127
|
1332 R1[WS(rs, 11)] = T6f - T6g;
|
cannam@127
|
1333 R1[WS(rs, 27)] = T6f + T6g;
|
cannam@127
|
1334 T6j = T6h - T6i;
|
cannam@127
|
1335 T6m = FNMS(KP1_990369453, T6l, KP196034280 * T6k);
|
cannam@127
|
1336 R1[WS(rs, 23)] = T6j - T6m;
|
cannam@127
|
1337 R1[WS(rs, 7)] = T6j + T6m;
|
cannam@127
|
1338 }
|
cannam@127
|
1339 }
|
cannam@127
|
1340 {
|
cannam@127
|
1341 E T4X, T5b, T50, T5c, T54, T5e, T57, T5f, T4Y, T4Z;
|
cannam@127
|
1342 T4X = T4V - T4W;
|
cannam@127
|
1343 T5b = T4V + T4W;
|
cannam@127
|
1344 T4Y = T3Q - T3X;
|
cannam@127
|
1345 T4Z = T44 - T41;
|
cannam@127
|
1346 T50 = FNMS(KP1_662939224, T4Z, KP1_111140466 * T4Y);
|
cannam@127
|
1347 T5c = FMA(KP1_662939224, T4Y, KP1_111140466 * T4Z);
|
cannam@127
|
1348 {
|
cannam@127
|
1349 E T52, T53, T55, T56;
|
cannam@127
|
1350 T52 = T4a - T4h;
|
cannam@127
|
1351 T53 = T4A - T4z;
|
cannam@127
|
1352 T54 = T52 + T53;
|
cannam@127
|
1353 T5e = T52 - T53;
|
cannam@127
|
1354 T55 = T4p - T4w;
|
cannam@127
|
1355 T56 = T4H - T4E;
|
cannam@127
|
1356 T57 = T55 + T56;
|
cannam@127
|
1357 T5f = T56 - T55;
|
cannam@127
|
1358 }
|
cannam@127
|
1359 {
|
cannam@127
|
1360 E T51, T58, T5h, T5i;
|
cannam@127
|
1361 T51 = T4X + T50;
|
cannam@127
|
1362 T58 = FNMS(KP942793473, T57, KP1_763842528 * T54);
|
cannam@127
|
1363 R1[WS(rs, 18)] = T51 - T58;
|
cannam@127
|
1364 R1[WS(rs, 2)] = T51 + T58;
|
cannam@127
|
1365 T5h = T5b + T5c;
|
cannam@127
|
1366 T5i = FMA(KP1_913880671, T5e, KP580569354 * T5f);
|
cannam@127
|
1367 R1[WS(rs, 14)] = T5h - T5i;
|
cannam@127
|
1368 R1[WS(rs, 30)] = T5h + T5i;
|
cannam@127
|
1369 }
|
cannam@127
|
1370 {
|
cannam@127
|
1371 E T59, T5a, T5d, T5g;
|
cannam@127
|
1372 T59 = T4X - T50;
|
cannam@127
|
1373 T5a = FMA(KP942793473, T54, KP1_763842528 * T57);
|
cannam@127
|
1374 R1[WS(rs, 10)] = T59 - T5a;
|
cannam@127
|
1375 R1[WS(rs, 26)] = T59 + T5a;
|
cannam@127
|
1376 T5d = T5b - T5c;
|
cannam@127
|
1377 T5g = FNMS(KP1_913880671, T5f, KP580569354 * T5e);
|
cannam@127
|
1378 R1[WS(rs, 22)] = T5d - T5g;
|
cannam@127
|
1379 R1[WS(rs, 6)] = T5d + T5g;
|
cannam@127
|
1380 }
|
cannam@127
|
1381 }
|
cannam@127
|
1382 }
|
cannam@127
|
1383 }
|
cannam@127
|
1384 }
|
cannam@127
|
1385
|
cannam@127
|
1386 static const kr2c_desc desc = { 64, "r2cb_64", {342, 82, 52, 0}, &GENUS };
|
cannam@127
|
1387
|
cannam@127
|
1388 void X(codelet_r2cb_64) (planner *p) {
|
cannam@127
|
1389 X(kr2c_register) (p, r2cb_64, &desc);
|
cannam@127
|
1390 }
|
cannam@127
|
1391
|
cannam@127
|
1392 #endif /* HAVE_FMA */
|