Chris@10
|
1 /*
|
Chris@10
|
2 * Copyright (c) 2003, 2007-11 Matteo Frigo
|
Chris@10
|
3 * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology
|
Chris@10
|
4 *
|
Chris@10
|
5 * This program is free software; you can redistribute it and/or modify
|
Chris@10
|
6 * it under the terms of the GNU General Public License as published by
|
Chris@10
|
7 * the Free Software Foundation; either version 2 of the License, or
|
Chris@10
|
8 * (at your option) any later version.
|
Chris@10
|
9 *
|
Chris@10
|
10 * This program is distributed in the hope that it will be useful,
|
Chris@10
|
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
Chris@10
|
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
Chris@10
|
13 * GNU General Public License for more details.
|
Chris@10
|
14 *
|
Chris@10
|
15 * You should have received a copy of the GNU General Public License
|
Chris@10
|
16 * along with this program; if not, write to the Free Software
|
Chris@10
|
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
Chris@10
|
18 *
|
Chris@10
|
19 */
|
Chris@10
|
20
|
Chris@10
|
21 /* This file was automatically generated --- DO NOT EDIT */
|
Chris@10
|
22 /* Generated on Sun Nov 25 07:41:38 EST 2012 */
|
Chris@10
|
23
|
Chris@10
|
24 #include "codelet-rdft.h"
|
Chris@10
|
25
|
Chris@10
|
26 #ifdef HAVE_FMA
|
Chris@10
|
27
|
Chris@10
|
28 /* Generated by: ../../../genfft/gen_r2cb.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -sign 1 -n 64 -name r2cbIII_64 -dft-III -include r2cbIII.h */
|
Chris@10
|
29
|
Chris@10
|
30 /*
|
Chris@10
|
31 * This function contains 434 FP additions, 260 FP multiplications,
|
Chris@10
|
32 * (or, 238 additions, 64 multiplications, 196 fused multiply/add),
|
Chris@10
|
33 * 165 stack variables, 36 constants, and 128 memory accesses
|
Chris@10
|
34 */
|
Chris@10
|
35 #include "r2cbIII.h"
|
Chris@10
|
36
|
Chris@10
|
37 static void r2cbIII_64(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
|
Chris@10
|
38 {
|
Chris@10
|
39 DK(KP357805721, +0.357805721314524104672487743774474392487532769);
|
Chris@10
|
40 DK(KP1_883088130, +1.883088130366041556825018805199004714371179592);
|
Chris@10
|
41 DK(KP472964775, +0.472964775891319928124438237972992463904131113);
|
Chris@10
|
42 DK(KP1_807978586, +1.807978586246886663172400594461074097420264050);
|
Chris@10
|
43 DK(KP049126849, +0.049126849769467254105343321271313617079695752);
|
Chris@10
|
44 DK(KP1_997590912, +1.997590912410344785429543209518201388886407229);
|
Chris@10
|
45 DK(KP906347169, +0.906347169019147157946142717268914412664134293);
|
Chris@10
|
46 DK(KP1_481902250, +1.481902250709918182351233794990325459457910619);
|
Chris@10
|
47 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
Chris@10
|
48 DK(KP250486960, +0.250486960191305461595702160124721208578685568);
|
Chris@10
|
49 DK(KP1_940062506, +1.940062506389087985207968414572200502913731924);
|
Chris@10
|
50 DK(KP599376933, +0.599376933681923766271389869014404232837890546);
|
Chris@10
|
51 DK(KP1_715457220, +1.715457220000544139804539968569540274084981599);
|
Chris@10
|
52 DK(KP148335987, +0.148335987538347428753676511486911367000625355);
|
Chris@10
|
53 DK(KP1_978353019, +1.978353019929561946903347476032486127967379067);
|
Chris@10
|
54 DK(KP741650546, +0.741650546272035369581266691172079863842265220);
|
Chris@10
|
55 DK(KP1_606415062, +1.606415062961289819613353025926283847759138854);
|
Chris@10
|
56 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
Chris@10
|
57 DK(KP303346683, +0.303346683607342391675883946941299872384187453);
|
Chris@10
|
58 DK(KP1_913880671, +1.913880671464417729871595773960539938965698411);
|
Chris@10
|
59 DK(KP534511135, +0.534511135950791641089685961295362908582039528);
|
Chris@10
|
60 DK(KP1_763842528, +1.763842528696710059425513727320776699016885241);
|
Chris@10
|
61 DK(KP098491403, +0.098491403357164253077197521291327432293052451);
|
Chris@10
|
62 DK(KP1_990369453, +1.990369453344393772489673906218959843150949737);
|
Chris@10
|
63 DK(KP820678790, +0.820678790828660330972281985331011598767386482);
|
Chris@10
|
64 DK(KP1_546020906, +1.546020906725473921621813219516939601942082586);
|
Chris@10
|
65 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
Chris@10
|
66 DK(KP198912367, +0.198912367379658006911597622644676228597850501);
|
Chris@10
|
67 DK(KP1_961570560, +1.961570560806460898252364472268478073947867462);
|
Chris@10
|
68 DK(KP668178637, +0.668178637919298919997757686523080761552472251);
|
Chris@10
|
69 DK(KP1_662939224, +1.662939224605090474157576755235811513477121624);
|
Chris@10
|
70 DK(KP1_847759065, +1.847759065022573512256366378793576573644833252);
|
Chris@10
|
71 DK(KP1_414213562, +1.414213562373095048801688724209698078569671875);
|
Chris@10
|
72 DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
|
Chris@10
|
73 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
Chris@10
|
74 DK(KP414213562, +0.414213562373095048801688724209698078569671875);
|
Chris@10
|
75 {
|
Chris@10
|
76 INT i;
|
Chris@10
|
77 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)) {
|
Chris@10
|
78 E T43, T4b, T49, T4e, T3T, T46, T40, T4a;
|
Chris@10
|
79 {
|
Chris@10
|
80 E T3t, T15, T2E, T3U, T6b, Tf, T6Q, T6u, T5J, T4L, T3V, T1g, T5U, T5q, T3u;
|
Chris@10
|
81 E T2H, T6v, Tu, T5r, T4V, T6R, T6e, T2K, T1s, T2J, T1D, T3X, T3B, T5s, T4Q;
|
Chris@10
|
82 E T3Y, T3y, T6g, TK, T5M, T57, T6N, T6j, T35, T1W, T34, T25, T4i, T3J, T5N;
|
Chris@10
|
83 E T52, T4j, T3G, T6l, TZ, T3L, T5P, T5i, T6M, T6o, T3M, T38, T2n, T37, T2w;
|
Chris@10
|
84 E T4l, T3Q, T5Q, T5d;
|
Chris@10
|
85 {
|
Chris@10
|
86 E T3x, T3w, T3E, T3F;
|
Chris@10
|
87 {
|
Chris@10
|
88 E T5p, T5o, T2G, T2F;
|
Chris@10
|
89 {
|
Chris@10
|
90 E T11, T3, T5m, T2D, T2A, T6, T5n, T14, Tb, T16, Ta, T4I, T19, Tc, T1c;
|
Chris@10
|
91 E T1d;
|
Chris@10
|
92 {
|
Chris@10
|
93 E T4, T5, T12, T13;
|
Chris@10
|
94 {
|
Chris@10
|
95 E T1, T2, T2B, T2C;
|
Chris@10
|
96 T1 = Cr[0];
|
Chris@10
|
97 T2 = Cr[WS(csr, 31)];
|
Chris@10
|
98 T2B = Ci[0];
|
Chris@10
|
99 T2C = Ci[WS(csi, 31)];
|
Chris@10
|
100 T4 = Cr[WS(csr, 16)];
|
Chris@10
|
101 T11 = T1 - T2;
|
Chris@10
|
102 T3 = T1 + T2;
|
Chris@10
|
103 T5m = T2C - T2B;
|
Chris@10
|
104 T2D = T2B + T2C;
|
Chris@10
|
105 T5 = Cr[WS(csr, 15)];
|
Chris@10
|
106 T12 = Ci[WS(csi, 16)];
|
Chris@10
|
107 T13 = Ci[WS(csi, 15)];
|
Chris@10
|
108 }
|
Chris@10
|
109 {
|
Chris@10
|
110 E T8, T9, T17, T18;
|
Chris@10
|
111 T8 = Cr[WS(csr, 8)];
|
Chris@10
|
112 T2A = T4 - T5;
|
Chris@10
|
113 T6 = T4 + T5;
|
Chris@10
|
114 T5n = T13 - T12;
|
Chris@10
|
115 T14 = T12 + T13;
|
Chris@10
|
116 T9 = Cr[WS(csr, 23)];
|
Chris@10
|
117 T17 = Ci[WS(csi, 8)];
|
Chris@10
|
118 T18 = Ci[WS(csi, 23)];
|
Chris@10
|
119 Tb = Cr[WS(csr, 7)];
|
Chris@10
|
120 T16 = T8 - T9;
|
Chris@10
|
121 Ta = T8 + T9;
|
Chris@10
|
122 T4I = T18 - T17;
|
Chris@10
|
123 T19 = T17 + T18;
|
Chris@10
|
124 Tc = Cr[WS(csr, 24)];
|
Chris@10
|
125 T1c = Ci[WS(csi, 7)];
|
Chris@10
|
126 T1d = Ci[WS(csi, 24)];
|
Chris@10
|
127 }
|
Chris@10
|
128 }
|
Chris@10
|
129 {
|
Chris@10
|
130 E T1b, T4J, T1e, T4H, T7, Te, Td;
|
Chris@10
|
131 T3t = T11 + T14;
|
Chris@10
|
132 T15 = T11 - T14;
|
Chris@10
|
133 T1b = Tb - Tc;
|
Chris@10
|
134 Td = Tb + Tc;
|
Chris@10
|
135 T4J = T1c - T1d;
|
Chris@10
|
136 T1e = T1c + T1d;
|
Chris@10
|
137 T2E = T2A + T2D;
|
Chris@10
|
138 T3U = T2A - T2D;
|
Chris@10
|
139 T4H = T3 - T6;
|
Chris@10
|
140 T7 = T3 + T6;
|
Chris@10
|
141 Te = Ta + Td;
|
Chris@10
|
142 T5p = Ta - Td;
|
Chris@10
|
143 {
|
Chris@10
|
144 E T4K, T6s, T6t, T1a, T1f;
|
Chris@10
|
145 T5o = T5m - T5n;
|
Chris@10
|
146 T6s = T5n + T5m;
|
Chris@10
|
147 T6t = T4I + T4J;
|
Chris@10
|
148 T4K = T4I - T4J;
|
Chris@10
|
149 T6b = T7 - Te;
|
Chris@10
|
150 Tf = T7 + Te;
|
Chris@10
|
151 T6Q = T6t + T6s;
|
Chris@10
|
152 T6u = T6s - T6t;
|
Chris@10
|
153 T2G = T16 + T19;
|
Chris@10
|
154 T1a = T16 - T19;
|
Chris@10
|
155 T1f = T1b - T1e;
|
Chris@10
|
156 T2F = T1b + T1e;
|
Chris@10
|
157 T5J = T4H - T4K;
|
Chris@10
|
158 T4L = T4H + T4K;
|
Chris@10
|
159 T3V = T1a - T1f;
|
Chris@10
|
160 T1g = T1a + T1f;
|
Chris@10
|
161 }
|
Chris@10
|
162 }
|
Chris@10
|
163 }
|
Chris@10
|
164 {
|
Chris@10
|
165 E T1i, Ti, T4O, T1q, T1n, Tl, T4N, T1l, Tq, T1t, Tp, T4T, T1A, Tr, T1u;
|
Chris@10
|
166 E T1v;
|
Chris@10
|
167 {
|
Chris@10
|
168 E Tj, Tk, T1j, T1k;
|
Chris@10
|
169 {
|
Chris@10
|
170 E Tg, Th, T1o, T1p;
|
Chris@10
|
171 Tg = Cr[WS(csr, 4)];
|
Chris@10
|
172 T5U = T5p + T5o;
|
Chris@10
|
173 T5q = T5o - T5p;
|
Chris@10
|
174 T3u = T2G + T2F;
|
Chris@10
|
175 T2H = T2F - T2G;
|
Chris@10
|
176 Th = Cr[WS(csr, 27)];
|
Chris@10
|
177 T1o = Ci[WS(csi, 4)];
|
Chris@10
|
178 T1p = Ci[WS(csi, 27)];
|
Chris@10
|
179 Tj = Cr[WS(csr, 20)];
|
Chris@10
|
180 T1i = Tg - Th;
|
Chris@10
|
181 Ti = Tg + Th;
|
Chris@10
|
182 T4O = T1p - T1o;
|
Chris@10
|
183 T1q = T1o + T1p;
|
Chris@10
|
184 Tk = Cr[WS(csr, 11)];
|
Chris@10
|
185 T1j = Ci[WS(csi, 20)];
|
Chris@10
|
186 T1k = Ci[WS(csi, 11)];
|
Chris@10
|
187 }
|
Chris@10
|
188 {
|
Chris@10
|
189 E Tn, To, T1y, T1z;
|
Chris@10
|
190 Tn = Cr[WS(csr, 3)];
|
Chris@10
|
191 T1n = Tj - Tk;
|
Chris@10
|
192 Tl = Tj + Tk;
|
Chris@10
|
193 T4N = T1k - T1j;
|
Chris@10
|
194 T1l = T1j + T1k;
|
Chris@10
|
195 To = Cr[WS(csr, 28)];
|
Chris@10
|
196 T1y = Ci[WS(csi, 3)];
|
Chris@10
|
197 T1z = Ci[WS(csi, 28)];
|
Chris@10
|
198 Tq = Cr[WS(csr, 12)];
|
Chris@10
|
199 T1t = Tn - To;
|
Chris@10
|
200 Tp = Tn + To;
|
Chris@10
|
201 T4T = T1y - T1z;
|
Chris@10
|
202 T1A = T1y + T1z;
|
Chris@10
|
203 Tr = Cr[WS(csr, 19)];
|
Chris@10
|
204 T1u = Ci[WS(csi, 12)];
|
Chris@10
|
205 T1v = Ci[WS(csi, 19)];
|
Chris@10
|
206 }
|
Chris@10
|
207 }
|
Chris@10
|
208 {
|
Chris@10
|
209 E T4M, T1B, T1w, T4P, T1m, T1r, Tm, Ts, T4S;
|
Chris@10
|
210 T4M = Ti - Tl;
|
Chris@10
|
211 Tm = Ti + Tl;
|
Chris@10
|
212 T1B = Tq - Tr;
|
Chris@10
|
213 Ts = Tq + Tr;
|
Chris@10
|
214 T4S = T1v - T1u;
|
Chris@10
|
215 T1w = T1u + T1v;
|
Chris@10
|
216 {
|
Chris@10
|
217 E T6c, Tt, T4R, T6d, T4U;
|
Chris@10
|
218 T6c = T4N + T4O;
|
Chris@10
|
219 T4P = T4N - T4O;
|
Chris@10
|
220 Tt = Tp + Ts;
|
Chris@10
|
221 T4R = Tp - Ts;
|
Chris@10
|
222 T6d = T4S + T4T;
|
Chris@10
|
223 T4U = T4S - T4T;
|
Chris@10
|
224 T3x = T1i + T1l;
|
Chris@10
|
225 T1m = T1i - T1l;
|
Chris@10
|
226 T6v = Tm - Tt;
|
Chris@10
|
227 Tu = Tm + Tt;
|
Chris@10
|
228 T5r = T4R - T4U;
|
Chris@10
|
229 T4V = T4R + T4U;
|
Chris@10
|
230 T6R = T6c + T6d;
|
Chris@10
|
231 T6e = T6c - T6d;
|
Chris@10
|
232 T1r = T1n + T1q;
|
Chris@10
|
233 T3w = T1n - T1q;
|
Chris@10
|
234 }
|
Chris@10
|
235 {
|
Chris@10
|
236 E T3A, T3z, T1x, T1C;
|
Chris@10
|
237 T3A = T1t + T1w;
|
Chris@10
|
238 T1x = T1t - T1w;
|
Chris@10
|
239 T1C = T1A - T1B;
|
Chris@10
|
240 T3z = T1B + T1A;
|
Chris@10
|
241 T2K = FMA(KP414213562, T1m, T1r);
|
Chris@10
|
242 T1s = FNMS(KP414213562, T1r, T1m);
|
Chris@10
|
243 T2J = FMA(KP414213562, T1x, T1C);
|
Chris@10
|
244 T1D = FNMS(KP414213562, T1C, T1x);
|
Chris@10
|
245 T3X = FMA(KP414213562, T3z, T3A);
|
Chris@10
|
246 T3B = FNMS(KP414213562, T3A, T3z);
|
Chris@10
|
247 T5s = T4M + T4P;
|
Chris@10
|
248 T4Q = T4M - T4P;
|
Chris@10
|
249 }
|
Chris@10
|
250 }
|
Chris@10
|
251 }
|
Chris@10
|
252 }
|
Chris@10
|
253 {
|
Chris@10
|
254 E T1G, Ty, T54, T20, T1X, TB, T53, T1J, TI, T4Z, T1L, TF, T22, T1U, T50;
|
Chris@10
|
255 E T1O;
|
Chris@10
|
256 {
|
Chris@10
|
257 E T1Y, T1Z, Tz, TA, Tw, Tx, T1H, T1I;
|
Chris@10
|
258 Tw = Cr[WS(csr, 2)];
|
Chris@10
|
259 Tx = Cr[WS(csr, 29)];
|
Chris@10
|
260 T1Y = Ci[WS(csi, 2)];
|
Chris@10
|
261 T3Y = FNMS(KP414213562, T3w, T3x);
|
Chris@10
|
262 T3y = FMA(KP414213562, T3x, T3w);
|
Chris@10
|
263 T1G = Tw - Tx;
|
Chris@10
|
264 Ty = Tw + Tx;
|
Chris@10
|
265 T1Z = Ci[WS(csi, 29)];
|
Chris@10
|
266 Tz = Cr[WS(csr, 18)];
|
Chris@10
|
267 TA = Cr[WS(csr, 13)];
|
Chris@10
|
268 T1H = Ci[WS(csi, 18)];
|
Chris@10
|
269 T54 = T1Y - T1Z;
|
Chris@10
|
270 T20 = T1Y + T1Z;
|
Chris@10
|
271 T1X = Tz - TA;
|
Chris@10
|
272 TB = Tz + TA;
|
Chris@10
|
273 T1I = Ci[WS(csi, 13)];
|
Chris@10
|
274 {
|
Chris@10
|
275 E T1R, T1Q, T1S, TG, TH;
|
Chris@10
|
276 TG = Cr[WS(csr, 5)];
|
Chris@10
|
277 TH = Cr[WS(csr, 26)];
|
Chris@10
|
278 T1R = Ci[WS(csi, 5)];
|
Chris@10
|
279 T53 = T1H - T1I;
|
Chris@10
|
280 T1J = T1H + T1I;
|
Chris@10
|
281 T1Q = TG - TH;
|
Chris@10
|
282 TI = TG + TH;
|
Chris@10
|
283 T1S = Ci[WS(csi, 26)];
|
Chris@10
|
284 {
|
Chris@10
|
285 E T1M, T1N, TD, TE, T1T;
|
Chris@10
|
286 TD = Cr[WS(csr, 10)];
|
Chris@10
|
287 TE = Cr[WS(csr, 21)];
|
Chris@10
|
288 T1T = T1R + T1S;
|
Chris@10
|
289 T4Z = T1S - T1R;
|
Chris@10
|
290 T1M = Ci[WS(csi, 10)];
|
Chris@10
|
291 T1L = TD - TE;
|
Chris@10
|
292 TF = TD + TE;
|
Chris@10
|
293 T1N = Ci[WS(csi, 21)];
|
Chris@10
|
294 T22 = T1Q + T1T;
|
Chris@10
|
295 T1U = T1Q - T1T;
|
Chris@10
|
296 T50 = T1M - T1N;
|
Chris@10
|
297 T1O = T1M + T1N;
|
Chris@10
|
298 }
|
Chris@10
|
299 }
|
Chris@10
|
300 }
|
Chris@10
|
301 {
|
Chris@10
|
302 E T4Y, T23, T51, T1K, T1V, T3I, T3H, T21, T24;
|
Chris@10
|
303 {
|
Chris@10
|
304 E T56, T1P, T6h, T55, TC, TJ, T6i;
|
Chris@10
|
305 T4Y = Ty - TB;
|
Chris@10
|
306 TC = Ty + TB;
|
Chris@10
|
307 TJ = TF + TI;
|
Chris@10
|
308 T56 = TF - TI;
|
Chris@10
|
309 T1P = T1L - T1O;
|
Chris@10
|
310 T23 = T1L + T1O;
|
Chris@10
|
311 T6h = T53 + T54;
|
Chris@10
|
312 T55 = T53 - T54;
|
Chris@10
|
313 T6g = TC - TJ;
|
Chris@10
|
314 TK = TC + TJ;
|
Chris@10
|
315 T6i = T50 + T4Z;
|
Chris@10
|
316 T51 = T4Z - T50;
|
Chris@10
|
317 T3E = T1G + T1J;
|
Chris@10
|
318 T1K = T1G - T1J;
|
Chris@10
|
319 T5M = T56 + T55;
|
Chris@10
|
320 T57 = T55 - T56;
|
Chris@10
|
321 T6N = T6i + T6h;
|
Chris@10
|
322 T6j = T6h - T6i;
|
Chris@10
|
323 T1V = T1P + T1U;
|
Chris@10
|
324 T3I = T1P - T1U;
|
Chris@10
|
325 }
|
Chris@10
|
326 T3H = T1X - T20;
|
Chris@10
|
327 T21 = T1X + T20;
|
Chris@10
|
328 T24 = T22 - T23;
|
Chris@10
|
329 T3F = T23 + T22;
|
Chris@10
|
330 T35 = FNMS(KP707106781, T1V, T1K);
|
Chris@10
|
331 T1W = FMA(KP707106781, T1V, T1K);
|
Chris@10
|
332 T34 = FMA(KP707106781, T24, T21);
|
Chris@10
|
333 T25 = FNMS(KP707106781, T24, T21);
|
Chris@10
|
334 T4i = FMA(KP707106781, T3I, T3H);
|
Chris@10
|
335 T3J = FNMS(KP707106781, T3I, T3H);
|
Chris@10
|
336 T5N = T4Y - T51;
|
Chris@10
|
337 T52 = T4Y + T51;
|
Chris@10
|
338 }
|
Chris@10
|
339 }
|
Chris@10
|
340 {
|
Chris@10
|
341 E T27, TN, T5f, T2q, T2r, TQ, T5e, T2a, TX, T5a, T2c, TU, T2t, T2l, T5b;
|
Chris@10
|
342 E T2f;
|
Chris@10
|
343 {
|
Chris@10
|
344 E T2o, T2p, TO, TP, TL, TM, T28, T29;
|
Chris@10
|
345 TL = Cr[WS(csr, 1)];
|
Chris@10
|
346 TM = Cr[WS(csr, 30)];
|
Chris@10
|
347 T2o = Ci[WS(csi, 1)];
|
Chris@10
|
348 T4j = FMA(KP707106781, T3F, T3E);
|
Chris@10
|
349 T3G = FNMS(KP707106781, T3F, T3E);
|
Chris@10
|
350 T27 = TL - TM;
|
Chris@10
|
351 TN = TL + TM;
|
Chris@10
|
352 T2p = Ci[WS(csi, 30)];
|
Chris@10
|
353 TO = Cr[WS(csr, 14)];
|
Chris@10
|
354 TP = Cr[WS(csr, 17)];
|
Chris@10
|
355 T28 = Ci[WS(csi, 14)];
|
Chris@10
|
356 T5f = T2p - T2o;
|
Chris@10
|
357 T2q = T2o + T2p;
|
Chris@10
|
358 T2r = TO - TP;
|
Chris@10
|
359 TQ = TO + TP;
|
Chris@10
|
360 T29 = Ci[WS(csi, 17)];
|
Chris@10
|
361 {
|
Chris@10
|
362 E T2i, T2h, T2j, TV, TW;
|
Chris@10
|
363 TV = Cr[WS(csr, 9)];
|
Chris@10
|
364 TW = Cr[WS(csr, 22)];
|
Chris@10
|
365 T2i = Ci[WS(csi, 9)];
|
Chris@10
|
366 T5e = T28 - T29;
|
Chris@10
|
367 T2a = T28 + T29;
|
Chris@10
|
368 T2h = TV - TW;
|
Chris@10
|
369 TX = TV + TW;
|
Chris@10
|
370 T2j = Ci[WS(csi, 22)];
|
Chris@10
|
371 {
|
Chris@10
|
372 E T2d, T2e, TS, TT, T2k;
|
Chris@10
|
373 TS = Cr[WS(csr, 6)];
|
Chris@10
|
374 TT = Cr[WS(csr, 25)];
|
Chris@10
|
375 T2k = T2i + T2j;
|
Chris@10
|
376 T5a = T2j - T2i;
|
Chris@10
|
377 T2d = Ci[WS(csi, 6)];
|
Chris@10
|
378 T2c = TS - TT;
|
Chris@10
|
379 TU = TS + TT;
|
Chris@10
|
380 T2e = Ci[WS(csi, 25)];
|
Chris@10
|
381 T2t = T2h + T2k;
|
Chris@10
|
382 T2l = T2h - T2k;
|
Chris@10
|
383 T5b = T2d - T2e;
|
Chris@10
|
384 T2f = T2d + T2e;
|
Chris@10
|
385 }
|
Chris@10
|
386 }
|
Chris@10
|
387 }
|
Chris@10
|
388 {
|
Chris@10
|
389 E T59, T2u, T5c, T2b, T2m, T3P, T3O, T2s, T2v;
|
Chris@10
|
390 {
|
Chris@10
|
391 E T5h, T2g, T6m, T5g, TR, TY, T6n;
|
Chris@10
|
392 T59 = TN - TQ;
|
Chris@10
|
393 TR = TN + TQ;
|
Chris@10
|
394 TY = TU + TX;
|
Chris@10
|
395 T5h = TU - TX;
|
Chris@10
|
396 T2g = T2c - T2f;
|
Chris@10
|
397 T2u = T2c + T2f;
|
Chris@10
|
398 T6m = T5e + T5f;
|
Chris@10
|
399 T5g = T5e - T5f;
|
Chris@10
|
400 T6l = TR - TY;
|
Chris@10
|
401 TZ = TR + TY;
|
Chris@10
|
402 T6n = T5b + T5a;
|
Chris@10
|
403 T5c = T5a - T5b;
|
Chris@10
|
404 T3L = T27 + T2a;
|
Chris@10
|
405 T2b = T27 - T2a;
|
Chris@10
|
406 T5P = T5h + T5g;
|
Chris@10
|
407 T5i = T5g - T5h;
|
Chris@10
|
408 T6M = T6n + T6m;
|
Chris@10
|
409 T6o = T6m - T6n;
|
Chris@10
|
410 T2m = T2g + T2l;
|
Chris@10
|
411 T3P = T2g - T2l;
|
Chris@10
|
412 }
|
Chris@10
|
413 T3O = T2r + T2q;
|
Chris@10
|
414 T2s = T2q - T2r;
|
Chris@10
|
415 T2v = T2t - T2u;
|
Chris@10
|
416 T3M = T2u + T2t;
|
Chris@10
|
417 T38 = FNMS(KP707106781, T2m, T2b);
|
Chris@10
|
418 T2n = FMA(KP707106781, T2m, T2b);
|
Chris@10
|
419 T37 = FNMS(KP707106781, T2v, T2s);
|
Chris@10
|
420 T2w = FMA(KP707106781, T2v, T2s);
|
Chris@10
|
421 T4l = FMA(KP707106781, T3P, T3O);
|
Chris@10
|
422 T3Q = FNMS(KP707106781, T3P, T3O);
|
Chris@10
|
423 T5Q = T59 - T5c;
|
Chris@10
|
424 T5d = T59 + T5c;
|
Chris@10
|
425 }
|
Chris@10
|
426 }
|
Chris@10
|
427 }
|
Chris@10
|
428 {
|
Chris@10
|
429 E T4m, T3N, T5t, T5L, T63, T4W, T5Y, T5X, T66, T5W, T67, T5S;
|
Chris@10
|
430 {
|
Chris@10
|
431 E T6T, T6S, T6W, T6P;
|
Chris@10
|
432 {
|
Chris@10
|
433 E T6L, T6O, T6Y, T6X, T6Z, Tv, T10, T70;
|
Chris@10
|
434 T6L = Tf - Tu;
|
Chris@10
|
435 Tv = Tf + Tu;
|
Chris@10
|
436 T10 = TK + TZ;
|
Chris@10
|
437 T6T = TK - TZ;
|
Chris@10
|
438 T6O = T6M - T6N;
|
Chris@10
|
439 T6Y = T6N + T6M;
|
Chris@10
|
440 T4m = FMA(KP707106781, T3M, T3L);
|
Chris@10
|
441 T3N = FNMS(KP707106781, T3M, T3L);
|
Chris@10
|
442 T6X = Tv - T10;
|
Chris@10
|
443 T6S = T6Q - T6R;
|
Chris@10
|
444 T6Z = T6R + T6Q;
|
Chris@10
|
445 R0[0] = KP2_000000000 * (Tv + T10);
|
Chris@10
|
446 R0[WS(rs, 16)] = KP2_000000000 * (T6Z - T6Y);
|
Chris@10
|
447 T70 = T6Y + T6Z;
|
Chris@10
|
448 T6W = T6L - T6O;
|
Chris@10
|
449 T6P = T6L + T6O;
|
Chris@10
|
450 R0[WS(rs, 24)] = KP1_414213562 * (T70 - T6X);
|
Chris@10
|
451 R0[WS(rs, 8)] = KP1_414213562 * (T6X + T70);
|
Chris@10
|
452 }
|
Chris@10
|
453 {
|
Chris@10
|
454 E T6D, T6f, T6w, T6G, T6p, T6x, T6y, T6k, T6V, T6U;
|
Chris@10
|
455 T6D = T6b - T6e;
|
Chris@10
|
456 T6f = T6b + T6e;
|
Chris@10
|
457 T6w = T6u - T6v;
|
Chris@10
|
458 T6G = T6v + T6u;
|
Chris@10
|
459 T6V = T6T + T6S;
|
Chris@10
|
460 T6U = T6S - T6T;
|
Chris@10
|
461 T6p = T6l + T6o;
|
Chris@10
|
462 T6x = T6l - T6o;
|
Chris@10
|
463 R0[WS(rs, 12)] = KP1_847759065 * (FMA(KP414213562, T6W, T6V));
|
Chris@10
|
464 R0[WS(rs, 28)] = -(KP1_847759065 * (FNMS(KP414213562, T6V, T6W)));
|
Chris@10
|
465 R0[WS(rs, 20)] = KP1_847759065 * (FNMS(KP414213562, T6P, T6U));
|
Chris@10
|
466 R0[WS(rs, 4)] = KP1_847759065 * (FMA(KP414213562, T6U, T6P));
|
Chris@10
|
467 T6y = T6g + T6j;
|
Chris@10
|
468 T6k = T6g - T6j;
|
Chris@10
|
469 {
|
Chris@10
|
470 E T5V, T5K, T5O, T5R;
|
Chris@10
|
471 T5t = T5r - T5s;
|
Chris@10
|
472 T5K = T5s + T5r;
|
Chris@10
|
473 {
|
Chris@10
|
474 E T6E, T6z, T6H, T6q;
|
Chris@10
|
475 T6E = T6y + T6x;
|
Chris@10
|
476 T6z = T6x - T6y;
|
Chris@10
|
477 T6H = T6k - T6p;
|
Chris@10
|
478 T6q = T6k + T6p;
|
Chris@10
|
479 {
|
Chris@10
|
480 E T6F, T6K, T6B, T6A;
|
Chris@10
|
481 T6F = FNMS(KP707106781, T6E, T6D);
|
Chris@10
|
482 T6K = FMA(KP707106781, T6E, T6D);
|
Chris@10
|
483 T6B = FNMS(KP707106781, T6z, T6w);
|
Chris@10
|
484 T6A = FMA(KP707106781, T6z, T6w);
|
Chris@10
|
485 {
|
Chris@10
|
486 E T6I, T6J, T6C, T6r;
|
Chris@10
|
487 T6I = FNMS(KP707106781, T6H, T6G);
|
Chris@10
|
488 T6J = FMA(KP707106781, T6H, T6G);
|
Chris@10
|
489 T6C = FNMS(KP707106781, T6q, T6f);
|
Chris@10
|
490 T6r = FMA(KP707106781, T6q, T6f);
|
Chris@10
|
491 R0[WS(rs, 22)] = KP1_662939224 * (FNMS(KP668178637, T6F, T6I));
|
Chris@10
|
492 R0[WS(rs, 6)] = KP1_662939224 * (FMA(KP668178637, T6I, T6F));
|
Chris@10
|
493 R0[WS(rs, 30)] = -(KP1_961570560 * (FNMS(KP198912367, T6J, T6K)));
|
Chris@10
|
494 R0[WS(rs, 14)] = KP1_961570560 * (FMA(KP198912367, T6K, T6J));
|
Chris@10
|
495 R0[WS(rs, 26)] = -(KP1_662939224 * (FNMS(KP668178637, T6B, T6C)));
|
Chris@10
|
496 R0[WS(rs, 10)] = KP1_662939224 * (FMA(KP668178637, T6C, T6B));
|
Chris@10
|
497 R0[WS(rs, 18)] = KP1_961570560 * (FNMS(KP198912367, T6r, T6A));
|
Chris@10
|
498 R0[WS(rs, 2)] = KP1_961570560 * (FMA(KP198912367, T6A, T6r));
|
Chris@10
|
499 T5L = FNMS(KP707106781, T5K, T5J);
|
Chris@10
|
500 T63 = FMA(KP707106781, T5K, T5J);
|
Chris@10
|
501 }
|
Chris@10
|
502 }
|
Chris@10
|
503 }
|
Chris@10
|
504 T5V = T4Q - T4V;
|
Chris@10
|
505 T4W = T4Q + T4V;
|
Chris@10
|
506 T5Y = FNMS(KP414213562, T5M, T5N);
|
Chris@10
|
507 T5O = FMA(KP414213562, T5N, T5M);
|
Chris@10
|
508 T5R = FNMS(KP414213562, T5Q, T5P);
|
Chris@10
|
509 T5X = FMA(KP414213562, T5P, T5Q);
|
Chris@10
|
510 T66 = FMA(KP707106781, T5V, T5U);
|
Chris@10
|
511 T5W = FNMS(KP707106781, T5V, T5U);
|
Chris@10
|
512 T67 = T5O + T5R;
|
Chris@10
|
513 T5S = T5O - T5R;
|
Chris@10
|
514 }
|
Chris@10
|
515 }
|
Chris@10
|
516 }
|
Chris@10
|
517 {
|
Chris@10
|
518 E T1h, T2L, T2I, T3h, T3p, T1E, T3n, T3s, T3b, T3k, T3e, T3o;
|
Chris@10
|
519 {
|
Chris@10
|
520 E T4X, T5B, T5v, T5w, T5E, T5u, T5F, T5k, T58, T5j;
|
Chris@10
|
521 {
|
Chris@10
|
522 E T68, T69, T62, T5T, T64, T5Z;
|
Chris@10
|
523 T68 = FNMS(KP923879532, T67, T66);
|
Chris@10
|
524 T69 = FMA(KP923879532, T67, T66);
|
Chris@10
|
525 T62 = FNMS(KP923879532, T5S, T5L);
|
Chris@10
|
526 T5T = FMA(KP923879532, T5S, T5L);
|
Chris@10
|
527 T64 = T5Y + T5X;
|
Chris@10
|
528 T5Z = T5X - T5Y;
|
Chris@10
|
529 T4X = FMA(KP707106781, T4W, T4L);
|
Chris@10
|
530 T5B = FNMS(KP707106781, T4W, T4L);
|
Chris@10
|
531 {
|
Chris@10
|
532 E T65, T6a, T61, T60;
|
Chris@10
|
533 T65 = FNMS(KP923879532, T64, T63);
|
Chris@10
|
534 T6a = FMA(KP923879532, T64, T63);
|
Chris@10
|
535 T61 = FNMS(KP923879532, T5Z, T5W);
|
Chris@10
|
536 T60 = FMA(KP923879532, T5Z, T5W);
|
Chris@10
|
537 R0[WS(rs, 23)] = KP1_546020906 * (FNMS(KP820678790, T65, T68));
|
Chris@10
|
538 R0[WS(rs, 7)] = KP1_546020906 * (FMA(KP820678790, T68, T65));
|
Chris@10
|
539 R0[WS(rs, 31)] = -(KP1_990369453 * (FNMS(KP098491403, T69, T6a)));
|
Chris@10
|
540 R0[WS(rs, 15)] = KP1_990369453 * (FMA(KP098491403, T6a, T69));
|
Chris@10
|
541 R0[WS(rs, 27)] = -(KP1_763842528 * (FNMS(KP534511135, T61, T62)));
|
Chris@10
|
542 R0[WS(rs, 11)] = KP1_763842528 * (FMA(KP534511135, T62, T61));
|
Chris@10
|
543 R0[WS(rs, 19)] = KP1_913880671 * (FNMS(KP303346683, T5T, T60));
|
Chris@10
|
544 R0[WS(rs, 3)] = KP1_913880671 * (FMA(KP303346683, T60, T5T));
|
Chris@10
|
545 }
|
Chris@10
|
546 }
|
Chris@10
|
547 T5v = FNMS(KP414213562, T52, T57);
|
Chris@10
|
548 T58 = FMA(KP414213562, T57, T52);
|
Chris@10
|
549 T5j = FNMS(KP414213562, T5i, T5d);
|
Chris@10
|
550 T5w = FMA(KP414213562, T5d, T5i);
|
Chris@10
|
551 T5E = FNMS(KP707106781, T5t, T5q);
|
Chris@10
|
552 T5u = FMA(KP707106781, T5t, T5q);
|
Chris@10
|
553 T5F = T58 - T5j;
|
Chris@10
|
554 T5k = T58 + T5j;
|
Chris@10
|
555 {
|
Chris@10
|
556 E T3l, T33, T3c, T3m, T3a, T3d;
|
Chris@10
|
557 {
|
Chris@10
|
558 E T39, T3f, T3g, T36;
|
Chris@10
|
559 {
|
Chris@10
|
560 E T31, T5G, T5H, T5A, T5l, T5C, T5x, T32;
|
Chris@10
|
561 T1h = FMA(KP707106781, T1g, T15);
|
Chris@10
|
562 T31 = FNMS(KP707106781, T1g, T15);
|
Chris@10
|
563 T5G = FNMS(KP923879532, T5F, T5E);
|
Chris@10
|
564 T5H = FMA(KP923879532, T5F, T5E);
|
Chris@10
|
565 T5A = FNMS(KP923879532, T5k, T4X);
|
Chris@10
|
566 T5l = FMA(KP923879532, T5k, T4X);
|
Chris@10
|
567 T5C = T5w - T5v;
|
Chris@10
|
568 T5x = T5v + T5w;
|
Chris@10
|
569 T32 = T2K + T2J;
|
Chris@10
|
570 T2L = T2J - T2K;
|
Chris@10
|
571 T39 = FNMS(KP668178637, T38, T37);
|
Chris@10
|
572 T3f = FMA(KP668178637, T37, T38);
|
Chris@10
|
573 {
|
Chris@10
|
574 E T5D, T5I, T5z, T5y;
|
Chris@10
|
575 T5D = FNMS(KP923879532, T5C, T5B);
|
Chris@10
|
576 T5I = FMA(KP923879532, T5C, T5B);
|
Chris@10
|
577 T5z = FNMS(KP923879532, T5x, T5u);
|
Chris@10
|
578 T5y = FMA(KP923879532, T5x, T5u);
|
Chris@10
|
579 T3l = FMA(KP923879532, T32, T31);
|
Chris@10
|
580 T33 = FNMS(KP923879532, T32, T31);
|
Chris@10
|
581 R0[WS(rs, 21)] = KP1_763842528 * (FNMS(KP534511135, T5D, T5G));
|
Chris@10
|
582 R0[WS(rs, 5)] = KP1_763842528 * (FMA(KP534511135, T5G, T5D));
|
Chris@10
|
583 R0[WS(rs, 29)] = -(KP1_913880671 * (FNMS(KP303346683, T5H, T5I)));
|
Chris@10
|
584 R0[WS(rs, 13)] = KP1_913880671 * (FMA(KP303346683, T5I, T5H));
|
Chris@10
|
585 R0[WS(rs, 25)] = -(KP1_546020906 * (FNMS(KP820678790, T5z, T5A)));
|
Chris@10
|
586 R0[WS(rs, 9)] = KP1_546020906 * (FMA(KP820678790, T5A, T5z));
|
Chris@10
|
587 R0[WS(rs, 17)] = KP1_990369453 * (FNMS(KP098491403, T5l, T5y));
|
Chris@10
|
588 R0[WS(rs, 1)] = KP1_990369453 * (FMA(KP098491403, T5y, T5l));
|
Chris@10
|
589 T3g = FMA(KP668178637, T34, T35);
|
Chris@10
|
590 T36 = FNMS(KP668178637, T35, T34);
|
Chris@10
|
591 }
|
Chris@10
|
592 }
|
Chris@10
|
593 T2I = FNMS(KP707106781, T2H, T2E);
|
Chris@10
|
594 T3c = FMA(KP707106781, T2H, T2E);
|
Chris@10
|
595 T3m = T3g + T3f;
|
Chris@10
|
596 T3h = T3f - T3g;
|
Chris@10
|
597 T3p = T39 - T36;
|
Chris@10
|
598 T3a = T36 + T39;
|
Chris@10
|
599 T3d = T1s - T1D;
|
Chris@10
|
600 T1E = T1s + T1D;
|
Chris@10
|
601 }
|
Chris@10
|
602 T3n = FNMS(KP831469612, T3m, T3l);
|
Chris@10
|
603 T3s = FMA(KP831469612, T3m, T3l);
|
Chris@10
|
604 T3b = FNMS(KP831469612, T3a, T33);
|
Chris@10
|
605 T3k = FMA(KP831469612, T3a, T33);
|
Chris@10
|
606 T3e = FMA(KP923879532, T3d, T3c);
|
Chris@10
|
607 T3o = FNMS(KP923879532, T3d, T3c);
|
Chris@10
|
608 }
|
Chris@10
|
609 }
|
Chris@10
|
610 {
|
Chris@10
|
611 E T3v, T3Z, T3W, T4v, T4D, T3C, T4B, T4G, T4p, T4y, T4s, T4C;
|
Chris@10
|
612 {
|
Chris@10
|
613 E T4z, T4h, T4q, T4A, T4o, T4r;
|
Chris@10
|
614 {
|
Chris@10
|
615 E T4n, T4t, T4u, T4k, T4f, T4g;
|
Chris@10
|
616 T3v = FNMS(KP707106781, T3u, T3t);
|
Chris@10
|
617 T4f = FMA(KP707106781, T3u, T3t);
|
Chris@10
|
618 T4g = T3Y + T3X;
|
Chris@10
|
619 T3Z = T3X - T3Y;
|
Chris@10
|
620 {
|
Chris@10
|
621 E T3r, T3q, T3i, T3j;
|
Chris@10
|
622 T3r = FNMS(KP831469612, T3p, T3o);
|
Chris@10
|
623 T3q = FMA(KP831469612, T3p, T3o);
|
Chris@10
|
624 T3i = FNMS(KP831469612, T3h, T3e);
|
Chris@10
|
625 T3j = FMA(KP831469612, T3h, T3e);
|
Chris@10
|
626 R1[WS(rs, 22)] = -(KP1_606415062 * (FMA(KP741650546, T3n, T3q)));
|
Chris@10
|
627 R1[WS(rs, 6)] = KP1_606415062 * (FNMS(KP741650546, T3q, T3n));
|
Chris@10
|
628 R1[WS(rs, 30)] = -(KP1_978353019 * (FMA(KP148335987, T3r, T3s)));
|
Chris@10
|
629 R1[WS(rs, 14)] = -(KP1_978353019 * (FNMS(KP148335987, T3s, T3r)));
|
Chris@10
|
630 R1[WS(rs, 26)] = -(KP1_715457220 * (FMA(KP599376933, T3j, T3k)));
|
Chris@10
|
631 R1[WS(rs, 10)] = -(KP1_715457220 * (FNMS(KP599376933, T3k, T3j)));
|
Chris@10
|
632 R1[WS(rs, 18)] = -(KP1_940062506 * (FMA(KP250486960, T3b, T3i)));
|
Chris@10
|
633 R1[WS(rs, 2)] = KP1_940062506 * (FNMS(KP250486960, T3i, T3b));
|
Chris@10
|
634 T4z = FMA(KP923879532, T4g, T4f);
|
Chris@10
|
635 T4h = FNMS(KP923879532, T4g, T4f);
|
Chris@10
|
636 }
|
Chris@10
|
637 T4n = FNMS(KP198912367, T4m, T4l);
|
Chris@10
|
638 T4t = FMA(KP198912367, T4l, T4m);
|
Chris@10
|
639 T4u = FNMS(KP198912367, T4i, T4j);
|
Chris@10
|
640 T4k = FMA(KP198912367, T4j, T4i);
|
Chris@10
|
641 T3W = FNMS(KP707106781, T3V, T3U);
|
Chris@10
|
642 T4q = FMA(KP707106781, T3V, T3U);
|
Chris@10
|
643 T4A = T4u + T4t;
|
Chris@10
|
644 T4v = T4t - T4u;
|
Chris@10
|
645 T4D = T4k + T4n;
|
Chris@10
|
646 T4o = T4k - T4n;
|
Chris@10
|
647 T4r = T3y + T3B;
|
Chris@10
|
648 T3C = T3y - T3B;
|
Chris@10
|
649 }
|
Chris@10
|
650 T4B = FNMS(KP980785280, T4A, T4z);
|
Chris@10
|
651 T4G = FMA(KP980785280, T4A, T4z);
|
Chris@10
|
652 T4p = FMA(KP980785280, T4o, T4h);
|
Chris@10
|
653 T4y = FNMS(KP980785280, T4o, T4h);
|
Chris@10
|
654 T4s = FNMS(KP923879532, T4r, T4q);
|
Chris@10
|
655 T4C = FMA(KP923879532, T4r, T4q);
|
Chris@10
|
656 }
|
Chris@10
|
657 {
|
Chris@10
|
658 E T2P, T2X, T2V, T30, T2z, T2S, T2M, T2W;
|
Chris@10
|
659 {
|
Chris@10
|
660 E T2T, T1F, T2U, T2y;
|
Chris@10
|
661 {
|
Chris@10
|
662 E T2x, T2N, T2O, T26;
|
Chris@10
|
663 {
|
Chris@10
|
664 E T4F, T4E, T4w, T4x;
|
Chris@10
|
665 T4F = FMA(KP980785280, T4D, T4C);
|
Chris@10
|
666 T4E = FNMS(KP980785280, T4D, T4C);
|
Chris@10
|
667 T4w = FMA(KP980785280, T4v, T4s);
|
Chris@10
|
668 T4x = FNMS(KP980785280, T4v, T4s);
|
Chris@10
|
669 R1[WS(rs, 23)] = KP1_481902250 * (FNMS(KP906347169, T4B, T4E));
|
Chris@10
|
670 R1[WS(rs, 7)] = KP1_481902250 * (FMA(KP906347169, T4E, T4B));
|
Chris@10
|
671 R1[WS(rs, 31)] = -(KP1_997590912 * (FNMS(KP049126849, T4F, T4G)));
|
Chris@10
|
672 R1[WS(rs, 15)] = KP1_997590912 * (FMA(KP049126849, T4G, T4F));
|
Chris@10
|
673 R1[WS(rs, 27)] = -(KP1_807978586 * (FNMS(KP472964775, T4x, T4y)));
|
Chris@10
|
674 R1[WS(rs, 11)] = KP1_807978586 * (FMA(KP472964775, T4y, T4x));
|
Chris@10
|
675 R1[WS(rs, 19)] = KP1_883088130 * (FNMS(KP357805721, T4p, T4w));
|
Chris@10
|
676 R1[WS(rs, 3)] = KP1_883088130 * (FMA(KP357805721, T4w, T4p));
|
Chris@10
|
677 T2T = FNMS(KP923879532, T1E, T1h);
|
Chris@10
|
678 T1F = FMA(KP923879532, T1E, T1h);
|
Chris@10
|
679 }
|
Chris@10
|
680 T2x = FNMS(KP198912367, T2w, T2n);
|
Chris@10
|
681 T2N = FMA(KP198912367, T2n, T2w);
|
Chris@10
|
682 T2O = FMA(KP198912367, T1W, T25);
|
Chris@10
|
683 T26 = FNMS(KP198912367, T25, T1W);
|
Chris@10
|
684 T2U = T2O + T2N;
|
Chris@10
|
685 T2P = T2N - T2O;
|
Chris@10
|
686 T2X = T26 - T2x;
|
Chris@10
|
687 T2y = T26 + T2x;
|
Chris@10
|
688 }
|
Chris@10
|
689 T2V = FNMS(KP980785280, T2U, T2T);
|
Chris@10
|
690 T30 = FMA(KP980785280, T2U, T2T);
|
Chris@10
|
691 T2z = FMA(KP980785280, T2y, T1F);
|
Chris@10
|
692 T2S = FNMS(KP980785280, T2y, T1F);
|
Chris@10
|
693 T2M = FNMS(KP923879532, T2L, T2I);
|
Chris@10
|
694 T2W = FMA(KP923879532, T2L, T2I);
|
Chris@10
|
695 }
|
Chris@10
|
696 {
|
Chris@10
|
697 E T47, T3D, T48, T3S;
|
Chris@10
|
698 {
|
Chris@10
|
699 E T3K, T41, T42, T3R;
|
Chris@10
|
700 {
|
Chris@10
|
701 E T2Z, T2Y, T2Q, T2R;
|
Chris@10
|
702 T2Z = FNMS(KP980785280, T2X, T2W);
|
Chris@10
|
703 T2Y = FMA(KP980785280, T2X, T2W);
|
Chris@10
|
704 T2Q = FNMS(KP980785280, T2P, T2M);
|
Chris@10
|
705 T2R = FMA(KP980785280, T2P, T2M);
|
Chris@10
|
706 R1[WS(rs, 20)] = -(KP1_807978586 * (FMA(KP472964775, T2V, T2Y)));
|
Chris@10
|
707 R1[WS(rs, 4)] = KP1_807978586 * (FNMS(KP472964775, T2Y, T2V));
|
Chris@10
|
708 R1[WS(rs, 28)] = -(KP1_883088130 * (FMA(KP357805721, T2Z, T30)));
|
Chris@10
|
709 R1[WS(rs, 12)] = -(KP1_883088130 * (FNMS(KP357805721, T30, T2Z)));
|
Chris@10
|
710 R1[WS(rs, 24)] = -(KP1_481902250 * (FMA(KP906347169, T2R, T2S)));
|
Chris@10
|
711 R1[WS(rs, 8)] = -(KP1_481902250 * (FNMS(KP906347169, T2S, T2R)));
|
Chris@10
|
712 R1[WS(rs, 16)] = -(KP1_997590912 * (FMA(KP049126849, T2z, T2Q)));
|
Chris@10
|
713 R1[0] = KP1_997590912 * (FNMS(KP049126849, T2Q, T2z));
|
Chris@10
|
714 T47 = FNMS(KP923879532, T3C, T3v);
|
Chris@10
|
715 T3D = FMA(KP923879532, T3C, T3v);
|
Chris@10
|
716 }
|
Chris@10
|
717 T3K = FMA(KP668178637, T3J, T3G);
|
Chris@10
|
718 T41 = FNMS(KP668178637, T3G, T3J);
|
Chris@10
|
719 T42 = FMA(KP668178637, T3N, T3Q);
|
Chris@10
|
720 T3R = FNMS(KP668178637, T3Q, T3N);
|
Chris@10
|
721 T48 = T42 - T41;
|
Chris@10
|
722 T43 = T41 + T42;
|
Chris@10
|
723 T4b = T3K - T3R;
|
Chris@10
|
724 T3S = T3K + T3R;
|
Chris@10
|
725 }
|
Chris@10
|
726 T49 = FNMS(KP831469612, T48, T47);
|
Chris@10
|
727 T4e = FMA(KP831469612, T48, T47);
|
Chris@10
|
728 T3T = FMA(KP831469612, T3S, T3D);
|
Chris@10
|
729 T46 = FNMS(KP831469612, T3S, T3D);
|
Chris@10
|
730 T40 = FMA(KP923879532, T3Z, T3W);
|
Chris@10
|
731 T4a = FNMS(KP923879532, T3Z, T3W);
|
Chris@10
|
732 }
|
Chris@10
|
733 }
|
Chris@10
|
734 }
|
Chris@10
|
735 }
|
Chris@10
|
736 }
|
Chris@10
|
737 }
|
Chris@10
|
738 {
|
Chris@10
|
739 E T4d, T4c, T44, T45;
|
Chris@10
|
740 T4d = FMA(KP831469612, T4b, T4a);
|
Chris@10
|
741 T4c = FNMS(KP831469612, T4b, T4a);
|
Chris@10
|
742 T44 = FMA(KP831469612, T43, T40);
|
Chris@10
|
743 T45 = FNMS(KP831469612, T43, T40);
|
Chris@10
|
744 R1[WS(rs, 21)] = KP1_715457220 * (FNMS(KP599376933, T49, T4c));
|
Chris@10
|
745 R1[WS(rs, 5)] = KP1_715457220 * (FMA(KP599376933, T4c, T49));
|
Chris@10
|
746 R1[WS(rs, 29)] = -(KP1_940062506 * (FNMS(KP250486960, T4d, T4e)));
|
Chris@10
|
747 R1[WS(rs, 13)] = KP1_940062506 * (FMA(KP250486960, T4e, T4d));
|
Chris@10
|
748 R1[WS(rs, 25)] = -(KP1_606415062 * (FNMS(KP741650546, T45, T46)));
|
Chris@10
|
749 R1[WS(rs, 9)] = KP1_606415062 * (FMA(KP741650546, T46, T45));
|
Chris@10
|
750 R1[WS(rs, 17)] = KP1_978353019 * (FNMS(KP148335987, T3T, T44));
|
Chris@10
|
751 R1[WS(rs, 1)] = KP1_978353019 * (FMA(KP148335987, T44, T3T));
|
Chris@10
|
752 }
|
Chris@10
|
753 }
|
Chris@10
|
754 }
|
Chris@10
|
755 }
|
Chris@10
|
756
|
Chris@10
|
757 static const kr2c_desc desc = { 64, "r2cbIII_64", {238, 64, 196, 0}, &GENUS };
|
Chris@10
|
758
|
Chris@10
|
759 void X(codelet_r2cbIII_64) (planner *p) {
|
Chris@10
|
760 X(kr2c_register) (p, r2cbIII_64, &desc);
|
Chris@10
|
761 }
|
Chris@10
|
762
|
Chris@10
|
763 #else /* HAVE_FMA */
|
Chris@10
|
764
|
Chris@10
|
765 /* Generated by: ../../../genfft/gen_r2cb.native -compact -variables 4 -pipeline-latency 4 -sign 1 -n 64 -name r2cbIII_64 -dft-III -include r2cbIII.h */
|
Chris@10
|
766
|
Chris@10
|
767 /*
|
Chris@10
|
768 * This function contains 434 FP additions, 208 FP multiplications,
|
Chris@10
|
769 * (or, 342 additions, 116 multiplications, 92 fused multiply/add),
|
Chris@10
|
770 * 130 stack variables, 39 constants, and 128 memory accesses
|
Chris@10
|
771 */
|
Chris@10
|
772 #include "r2cbIII.h"
|
Chris@10
|
773
|
Chris@10
|
774 static void r2cbIII_64(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
|
Chris@10
|
775 {
|
Chris@10
|
776 DK(KP1_343117909, +1.343117909694036801250753700854843606457501264);
|
Chris@10
|
777 DK(KP1_481902250, +1.481902250709918182351233794990325459457910619);
|
Chris@10
|
778 DK(KP1_807978586, +1.807978586246886663172400594461074097420264050);
|
Chris@10
|
779 DK(KP855110186, +0.855110186860564188641933713777597068609157259);
|
Chris@10
|
780 DK(KP1_997590912, +1.997590912410344785429543209518201388886407229);
|
Chris@10
|
781 DK(KP098135348, +0.098135348654836028509909953885365316629490726);
|
Chris@10
|
782 DK(KP673779706, +0.673779706784440101378506425238295140955533559);
|
Chris@10
|
783 DK(KP1_883088130, +1.883088130366041556825018805199004714371179592);
|
Chris@10
|
784 DK(KP195090322, +0.195090322016128267848284868477022240927691618);
|
Chris@10
|
785 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
Chris@10
|
786 DK(KP1_191398608, +1.191398608984866686934073057659939779023852677);
|
Chris@10
|
787 DK(KP1_606415062, +1.606415062961289819613353025926283847759138854);
|
Chris@10
|
788 DK(KP1_715457220, +1.715457220000544139804539968569540274084981599);
|
Chris@10
|
789 DK(KP1_028205488, +1.028205488386443453187387677937631545216098241);
|
Chris@10
|
790 DK(KP1_978353019, +1.978353019929561946903347476032486127967379067);
|
Chris@10
|
791 DK(KP293460948, +0.293460948910723503317700259293435639412430633);
|
Chris@10
|
792 DK(KP485960359, +0.485960359806527779896548324154942236641981567);
|
Chris@10
|
793 DK(KP1_940062506, +1.940062506389087985207968414572200502913731924);
|
Chris@10
|
794 DK(KP555570233, +0.555570233019602224742830813948532874374937191);
|
Chris@10
|
795 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
Chris@10
|
796 DK(KP1_268786568, +1.268786568327290996430343226450986741351374190);
|
Chris@10
|
797 DK(KP1_546020906, +1.546020906725473921621813219516939601942082586);
|
Chris@10
|
798 DK(KP1_763842528, +1.763842528696710059425513727320776699016885241);
|
Chris@10
|
799 DK(KP942793473, +0.942793473651995297112775251810508755314920638);
|
Chris@10
|
800 DK(KP1_990369453, +1.990369453344393772489673906218959843150949737);
|
Chris@10
|
801 DK(KP196034280, +0.196034280659121203988391127777283691722273346);
|
Chris@10
|
802 DK(KP580569354, +0.580569354508924735272384751634790549382952557);
|
Chris@10
|
803 DK(KP1_913880671, +1.913880671464417729871595773960539938965698411);
|
Chris@10
|
804 DK(KP1_662939224, +1.662939224605090474157576755235811513477121624);
|
Chris@10
|
805 DK(KP1_111140466, +1.111140466039204449485661627897065748749874382);
|
Chris@10
|
806 DK(KP390180644, +0.390180644032256535696569736954044481855383236);
|
Chris@10
|
807 DK(KP1_961570560, +1.961570560806460898252364472268478073947867462);
|
Chris@10
|
808 DK(KP765366864, +0.765366864730179543456919968060797733522689125);
|
Chris@10
|
809 DK(KP1_847759065, +1.847759065022573512256366378793576573644833252);
|
Chris@10
|
810 DK(KP1_414213562, +1.414213562373095048801688724209698078569671875);
|
Chris@10
|
811 DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
|
Chris@10
|
812 DK(KP382683432, +0.382683432365089771728459984030398866761344562);
|
Chris@10
|
813 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
Chris@10
|
814 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
Chris@10
|
815 {
|
Chris@10
|
816 INT i;
|
Chris@10
|
817 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)) {
|
Chris@10
|
818 E T15, T3t, T3U, T2N, Tf, T6b, T6u, T6R, T4L, T5J, T1g, T3V, T5q, T5U, T2I;
|
Chris@10
|
819 E T3u, Tu, T6v, T4V, T5s, T6e, T6Q, T1s, T2D, T1D, T2E, T3B, T3Y, T4Q, T5r;
|
Chris@10
|
820 E T3y, T3X, TK, T6g, T57, T5N, T6j, T6N, T1W, T34, T25, T35, T3J, T4j, T52;
|
Chris@10
|
821 E T5M, T3G, T4i, TZ, T6l, T5i, T5Q, T6o, T6M, T2n, T37, T2w, T38, T3Q, T4m;
|
Chris@10
|
822 E T5d, T5P, T3N, T4l;
|
Chris@10
|
823 {
|
Chris@10
|
824 E T3, T11, T2M, T5n, T6, T2J, T14, T5m, Ta, T16, T19, T4J, Td, T1b, T1e;
|
Chris@10
|
825 E T4I;
|
Chris@10
|
826 {
|
Chris@10
|
827 E T1, T2, T2K, T2L;
|
Chris@10
|
828 T1 = Cr[0];
|
Chris@10
|
829 T2 = Cr[WS(csr, 31)];
|
Chris@10
|
830 T3 = T1 + T2;
|
Chris@10
|
831 T11 = T1 - T2;
|
Chris@10
|
832 T2K = Ci[0];
|
Chris@10
|
833 T2L = Ci[WS(csi, 31)];
|
Chris@10
|
834 T2M = T2K + T2L;
|
Chris@10
|
835 T5n = T2L - T2K;
|
Chris@10
|
836 }
|
Chris@10
|
837 {
|
Chris@10
|
838 E T4, T5, T12, T13;
|
Chris@10
|
839 T4 = Cr[WS(csr, 16)];
|
Chris@10
|
840 T5 = Cr[WS(csr, 15)];
|
Chris@10
|
841 T6 = T4 + T5;
|
Chris@10
|
842 T2J = T4 - T5;
|
Chris@10
|
843 T12 = Ci[WS(csi, 16)];
|
Chris@10
|
844 T13 = Ci[WS(csi, 15)];
|
Chris@10
|
845 T14 = T12 + T13;
|
Chris@10
|
846 T5m = T12 - T13;
|
Chris@10
|
847 }
|
Chris@10
|
848 {
|
Chris@10
|
849 E T8, T9, T17, T18;
|
Chris@10
|
850 T8 = Cr[WS(csr, 8)];
|
Chris@10
|
851 T9 = Cr[WS(csr, 23)];
|
Chris@10
|
852 Ta = T8 + T9;
|
Chris@10
|
853 T16 = T8 - T9;
|
Chris@10
|
854 T17 = Ci[WS(csi, 8)];
|
Chris@10
|
855 T18 = Ci[WS(csi, 23)];
|
Chris@10
|
856 T19 = T17 + T18;
|
Chris@10
|
857 T4J = T17 - T18;
|
Chris@10
|
858 }
|
Chris@10
|
859 {
|
Chris@10
|
860 E Tb, Tc, T1c, T1d;
|
Chris@10
|
861 Tb = Cr[WS(csr, 7)];
|
Chris@10
|
862 Tc = Cr[WS(csr, 24)];
|
Chris@10
|
863 Td = Tb + Tc;
|
Chris@10
|
864 T1b = Tb - Tc;
|
Chris@10
|
865 T1c = Ci[WS(csi, 7)];
|
Chris@10
|
866 T1d = Ci[WS(csi, 24)];
|
Chris@10
|
867 T1e = T1c + T1d;
|
Chris@10
|
868 T4I = T1d - T1c;
|
Chris@10
|
869 }
|
Chris@10
|
870 {
|
Chris@10
|
871 E T7, Te, T1a, T1f;
|
Chris@10
|
872 T15 = T11 - T14;
|
Chris@10
|
873 T3t = T11 + T14;
|
Chris@10
|
874 T3U = T2J - T2M;
|
Chris@10
|
875 T2N = T2J + T2M;
|
Chris@10
|
876 T7 = T3 + T6;
|
Chris@10
|
877 Te = Ta + Td;
|
Chris@10
|
878 Tf = T7 + Te;
|
Chris@10
|
879 T6b = T7 - Te;
|
Chris@10
|
880 {
|
Chris@10
|
881 E T6s, T6t, T4H, T4K;
|
Chris@10
|
882 T6s = T4J + T4I;
|
Chris@10
|
883 T6t = T5n - T5m;
|
Chris@10
|
884 T6u = T6s + T6t;
|
Chris@10
|
885 T6R = T6t - T6s;
|
Chris@10
|
886 T4H = T3 - T6;
|
Chris@10
|
887 T4K = T4I - T4J;
|
Chris@10
|
888 T4L = T4H + T4K;
|
Chris@10
|
889 T5J = T4H - T4K;
|
Chris@10
|
890 }
|
Chris@10
|
891 T1a = T16 - T19;
|
Chris@10
|
892 T1f = T1b - T1e;
|
Chris@10
|
893 T1g = KP707106781 * (T1a + T1f);
|
Chris@10
|
894 T3V = KP707106781 * (T1a - T1f);
|
Chris@10
|
895 {
|
Chris@10
|
896 E T5o, T5p, T2G, T2H;
|
Chris@10
|
897 T5o = T5m + T5n;
|
Chris@10
|
898 T5p = Ta - Td;
|
Chris@10
|
899 T5q = T5o - T5p;
|
Chris@10
|
900 T5U = T5p + T5o;
|
Chris@10
|
901 T2G = T16 + T19;
|
Chris@10
|
902 T2H = T1b + T1e;
|
Chris@10
|
903 T2I = KP707106781 * (T2G - T2H);
|
Chris@10
|
904 T3u = KP707106781 * (T2G + T2H);
|
Chris@10
|
905 }
|
Chris@10
|
906 }
|
Chris@10
|
907 }
|
Chris@10
|
908 {
|
Chris@10
|
909 E Ti, T1i, T1q, T4N, Tl, T1n, T1l, T4O, Tp, T1t, T1B, T4S, Ts, T1y, T1w;
|
Chris@10
|
910 E T4T;
|
Chris@10
|
911 {
|
Chris@10
|
912 E Tg, Th, T1o, T1p;
|
Chris@10
|
913 Tg = Cr[WS(csr, 4)];
|
Chris@10
|
914 Th = Cr[WS(csr, 27)];
|
Chris@10
|
915 Ti = Tg + Th;
|
Chris@10
|
916 T1i = Tg - Th;
|
Chris@10
|
917 T1o = Ci[WS(csi, 4)];
|
Chris@10
|
918 T1p = Ci[WS(csi, 27)];
|
Chris@10
|
919 T1q = T1o + T1p;
|
Chris@10
|
920 T4N = T1o - T1p;
|
Chris@10
|
921 }
|
Chris@10
|
922 {
|
Chris@10
|
923 E Tj, Tk, T1j, T1k;
|
Chris@10
|
924 Tj = Cr[WS(csr, 20)];
|
Chris@10
|
925 Tk = Cr[WS(csr, 11)];
|
Chris@10
|
926 Tl = Tj + Tk;
|
Chris@10
|
927 T1n = Tj - Tk;
|
Chris@10
|
928 T1j = Ci[WS(csi, 20)];
|
Chris@10
|
929 T1k = Ci[WS(csi, 11)];
|
Chris@10
|
930 T1l = T1j + T1k;
|
Chris@10
|
931 T4O = T1j - T1k;
|
Chris@10
|
932 }
|
Chris@10
|
933 {
|
Chris@10
|
934 E Tn, To, T1z, T1A;
|
Chris@10
|
935 Tn = Cr[WS(csr, 3)];
|
Chris@10
|
936 To = Cr[WS(csr, 28)];
|
Chris@10
|
937 Tp = Tn + To;
|
Chris@10
|
938 T1t = Tn - To;
|
Chris@10
|
939 T1z = Ci[WS(csi, 3)];
|
Chris@10
|
940 T1A = Ci[WS(csi, 28)];
|
Chris@10
|
941 T1B = T1z + T1A;
|
Chris@10
|
942 T4S = T1A - T1z;
|
Chris@10
|
943 }
|
Chris@10
|
944 {
|
Chris@10
|
945 E Tq, Tr, T1u, T1v;
|
Chris@10
|
946 Tq = Cr[WS(csr, 12)];
|
Chris@10
|
947 Tr = Cr[WS(csr, 19)];
|
Chris@10
|
948 Ts = Tq + Tr;
|
Chris@10
|
949 T1y = Tq - Tr;
|
Chris@10
|
950 T1u = Ci[WS(csi, 12)];
|
Chris@10
|
951 T1v = Ci[WS(csi, 19)];
|
Chris@10
|
952 T1w = T1u + T1v;
|
Chris@10
|
953 T4T = T1u - T1v;
|
Chris@10
|
954 }
|
Chris@10
|
955 {
|
Chris@10
|
956 E Tm, Tt, T4R, T4U;
|
Chris@10
|
957 Tm = Ti + Tl;
|
Chris@10
|
958 Tt = Tp + Ts;
|
Chris@10
|
959 Tu = Tm + Tt;
|
Chris@10
|
960 T6v = Tm - Tt;
|
Chris@10
|
961 T4R = Tp - Ts;
|
Chris@10
|
962 T4U = T4S - T4T;
|
Chris@10
|
963 T4V = T4R + T4U;
|
Chris@10
|
964 T5s = T4U - T4R;
|
Chris@10
|
965 }
|
Chris@10
|
966 {
|
Chris@10
|
967 E T6c, T6d, T1m, T1r;
|
Chris@10
|
968 T6c = T4T + T4S;
|
Chris@10
|
969 T6d = T4O + T4N;
|
Chris@10
|
970 T6e = T6c - T6d;
|
Chris@10
|
971 T6Q = T6d + T6c;
|
Chris@10
|
972 T1m = T1i - T1l;
|
Chris@10
|
973 T1r = T1n + T1q;
|
Chris@10
|
974 T1s = FNMS(KP382683432, T1r, KP923879532 * T1m);
|
Chris@10
|
975 T2D = FMA(KP382683432, T1m, KP923879532 * T1r);
|
Chris@10
|
976 }
|
Chris@10
|
977 {
|
Chris@10
|
978 E T1x, T1C, T3z, T3A;
|
Chris@10
|
979 T1x = T1t - T1w;
|
Chris@10
|
980 T1C = T1y - T1B;
|
Chris@10
|
981 T1D = FMA(KP923879532, T1x, KP382683432 * T1C);
|
Chris@10
|
982 T2E = FNMS(KP382683432, T1x, KP923879532 * T1C);
|
Chris@10
|
983 T3z = T1t + T1w;
|
Chris@10
|
984 T3A = T1y + T1B;
|
Chris@10
|
985 T3B = FNMS(KP923879532, T3A, KP382683432 * T3z);
|
Chris@10
|
986 T3Y = FMA(KP923879532, T3z, KP382683432 * T3A);
|
Chris@10
|
987 }
|
Chris@10
|
988 {
|
Chris@10
|
989 E T4M, T4P, T3w, T3x;
|
Chris@10
|
990 T4M = Ti - Tl;
|
Chris@10
|
991 T4P = T4N - T4O;
|
Chris@10
|
992 T4Q = T4M - T4P;
|
Chris@10
|
993 T5r = T4M + T4P;
|
Chris@10
|
994 T3w = T1i + T1l;
|
Chris@10
|
995 T3x = T1q - T1n;
|
Chris@10
|
996 T3y = FNMS(KP923879532, T3x, KP382683432 * T3w);
|
Chris@10
|
997 T3X = FMA(KP923879532, T3w, KP382683432 * T3x);
|
Chris@10
|
998 }
|
Chris@10
|
999 }
|
Chris@10
|
1000 {
|
Chris@10
|
1001 E Ty, T1G, T23, T54, TB, T20, T1J, T55, TI, T4Z, T1U, T1Y, TF, T50, T1P;
|
Chris@10
|
1002 E T1X;
|
Chris@10
|
1003 {
|
Chris@10
|
1004 E Tw, Tx, T1H, T1I;
|
Chris@10
|
1005 Tw = Cr[WS(csr, 2)];
|
Chris@10
|
1006 Tx = Cr[WS(csr, 29)];
|
Chris@10
|
1007 Ty = Tw + Tx;
|
Chris@10
|
1008 T1G = Tw - Tx;
|
Chris@10
|
1009 {
|
Chris@10
|
1010 E T21, T22, Tz, TA;
|
Chris@10
|
1011 T21 = Ci[WS(csi, 2)];
|
Chris@10
|
1012 T22 = Ci[WS(csi, 29)];
|
Chris@10
|
1013 T23 = T21 + T22;
|
Chris@10
|
1014 T54 = T21 - T22;
|
Chris@10
|
1015 Tz = Cr[WS(csr, 18)];
|
Chris@10
|
1016 TA = Cr[WS(csr, 13)];
|
Chris@10
|
1017 TB = Tz + TA;
|
Chris@10
|
1018 T20 = Tz - TA;
|
Chris@10
|
1019 }
|
Chris@10
|
1020 T1H = Ci[WS(csi, 18)];
|
Chris@10
|
1021 T1I = Ci[WS(csi, 13)];
|
Chris@10
|
1022 T1J = T1H + T1I;
|
Chris@10
|
1023 T55 = T1H - T1I;
|
Chris@10
|
1024 {
|
Chris@10
|
1025 E TG, TH, T1Q, T1R, T1S, T1T;
|
Chris@10
|
1026 TG = Cr[WS(csr, 5)];
|
Chris@10
|
1027 TH = Cr[WS(csr, 26)];
|
Chris@10
|
1028 T1Q = TG - TH;
|
Chris@10
|
1029 T1R = Ci[WS(csi, 5)];
|
Chris@10
|
1030 T1S = Ci[WS(csi, 26)];
|
Chris@10
|
1031 T1T = T1R + T1S;
|
Chris@10
|
1032 TI = TG + TH;
|
Chris@10
|
1033 T4Z = T1S - T1R;
|
Chris@10
|
1034 T1U = T1Q - T1T;
|
Chris@10
|
1035 T1Y = T1Q + T1T;
|
Chris@10
|
1036 }
|
Chris@10
|
1037 {
|
Chris@10
|
1038 E TD, TE, T1L, T1M, T1N, T1O;
|
Chris@10
|
1039 TD = Cr[WS(csr, 10)];
|
Chris@10
|
1040 TE = Cr[WS(csr, 21)];
|
Chris@10
|
1041 T1L = TD - TE;
|
Chris@10
|
1042 T1M = Ci[WS(csi, 10)];
|
Chris@10
|
1043 T1N = Ci[WS(csi, 21)];
|
Chris@10
|
1044 T1O = T1M + T1N;
|
Chris@10
|
1045 TF = TD + TE;
|
Chris@10
|
1046 T50 = T1M - T1N;
|
Chris@10
|
1047 T1P = T1L - T1O;
|
Chris@10
|
1048 T1X = T1L + T1O;
|
Chris@10
|
1049 }
|
Chris@10
|
1050 }
|
Chris@10
|
1051 {
|
Chris@10
|
1052 E TC, TJ, T53, T56;
|
Chris@10
|
1053 TC = Ty + TB;
|
Chris@10
|
1054 TJ = TF + TI;
|
Chris@10
|
1055 TK = TC + TJ;
|
Chris@10
|
1056 T6g = TC - TJ;
|
Chris@10
|
1057 T53 = TF - TI;
|
Chris@10
|
1058 T56 = T54 - T55;
|
Chris@10
|
1059 T57 = T53 + T56;
|
Chris@10
|
1060 T5N = T56 - T53;
|
Chris@10
|
1061 }
|
Chris@10
|
1062 {
|
Chris@10
|
1063 E T6h, T6i, T1K, T1V;
|
Chris@10
|
1064 T6h = T55 + T54;
|
Chris@10
|
1065 T6i = T50 + T4Z;
|
Chris@10
|
1066 T6j = T6h - T6i;
|
Chris@10
|
1067 T6N = T6i + T6h;
|
Chris@10
|
1068 T1K = T1G - T1J;
|
Chris@10
|
1069 T1V = KP707106781 * (T1P + T1U);
|
Chris@10
|
1070 T1W = T1K + T1V;
|
Chris@10
|
1071 T34 = T1K - T1V;
|
Chris@10
|
1072 }
|
Chris@10
|
1073 {
|
Chris@10
|
1074 E T1Z, T24, T3H, T3I;
|
Chris@10
|
1075 T1Z = KP707106781 * (T1X - T1Y);
|
Chris@10
|
1076 T24 = T20 + T23;
|
Chris@10
|
1077 T25 = T1Z + T24;
|
Chris@10
|
1078 T35 = T24 - T1Z;
|
Chris@10
|
1079 T3H = KP707106781 * (T1P - T1U);
|
Chris@10
|
1080 T3I = T23 - T20;
|
Chris@10
|
1081 T3J = T3H + T3I;
|
Chris@10
|
1082 T4j = T3I - T3H;
|
Chris@10
|
1083 }
|
Chris@10
|
1084 {
|
Chris@10
|
1085 E T4Y, T51, T3E, T3F;
|
Chris@10
|
1086 T4Y = Ty - TB;
|
Chris@10
|
1087 T51 = T4Z - T50;
|
Chris@10
|
1088 T52 = T4Y + T51;
|
Chris@10
|
1089 T5M = T4Y - T51;
|
Chris@10
|
1090 T3E = T1G + T1J;
|
Chris@10
|
1091 T3F = KP707106781 * (T1X + T1Y);
|
Chris@10
|
1092 T3G = T3E - T3F;
|
Chris@10
|
1093 T4i = T3E + T3F;
|
Chris@10
|
1094 }
|
Chris@10
|
1095 }
|
Chris@10
|
1096 {
|
Chris@10
|
1097 E TN, T27, T2u, T5f, TQ, T2r, T2a, T5g, TX, T5a, T2l, T2p, TU, T5b, T2g;
|
Chris@10
|
1098 E T2o;
|
Chris@10
|
1099 {
|
Chris@10
|
1100 E TL, TM, T28, T29;
|
Chris@10
|
1101 TL = Cr[WS(csr, 1)];
|
Chris@10
|
1102 TM = Cr[WS(csr, 30)];
|
Chris@10
|
1103 TN = TL + TM;
|
Chris@10
|
1104 T27 = TL - TM;
|
Chris@10
|
1105 {
|
Chris@10
|
1106 E T2s, T2t, TO, TP;
|
Chris@10
|
1107 T2s = Ci[WS(csi, 1)];
|
Chris@10
|
1108 T2t = Ci[WS(csi, 30)];
|
Chris@10
|
1109 T2u = T2s + T2t;
|
Chris@10
|
1110 T5f = T2t - T2s;
|
Chris@10
|
1111 TO = Cr[WS(csr, 14)];
|
Chris@10
|
1112 TP = Cr[WS(csr, 17)];
|
Chris@10
|
1113 TQ = TO + TP;
|
Chris@10
|
1114 T2r = TO - TP;
|
Chris@10
|
1115 }
|
Chris@10
|
1116 T28 = Ci[WS(csi, 14)];
|
Chris@10
|
1117 T29 = Ci[WS(csi, 17)];
|
Chris@10
|
1118 T2a = T28 + T29;
|
Chris@10
|
1119 T5g = T28 - T29;
|
Chris@10
|
1120 {
|
Chris@10
|
1121 E TV, TW, T2h, T2i, T2j, T2k;
|
Chris@10
|
1122 TV = Cr[WS(csr, 9)];
|
Chris@10
|
1123 TW = Cr[WS(csr, 22)];
|
Chris@10
|
1124 T2h = TV - TW;
|
Chris@10
|
1125 T2i = Ci[WS(csi, 9)];
|
Chris@10
|
1126 T2j = Ci[WS(csi, 22)];
|
Chris@10
|
1127 T2k = T2i + T2j;
|
Chris@10
|
1128 TX = TV + TW;
|
Chris@10
|
1129 T5a = T2j - T2i;
|
Chris@10
|
1130 T2l = T2h - T2k;
|
Chris@10
|
1131 T2p = T2h + T2k;
|
Chris@10
|
1132 }
|
Chris@10
|
1133 {
|
Chris@10
|
1134 E TS, TT, T2c, T2d, T2e, T2f;
|
Chris@10
|
1135 TS = Cr[WS(csr, 6)];
|
Chris@10
|
1136 TT = Cr[WS(csr, 25)];
|
Chris@10
|
1137 T2c = TS - TT;
|
Chris@10
|
1138 T2d = Ci[WS(csi, 6)];
|
Chris@10
|
1139 T2e = Ci[WS(csi, 25)];
|
Chris@10
|
1140 T2f = T2d + T2e;
|
Chris@10
|
1141 TU = TS + TT;
|
Chris@10
|
1142 T5b = T2d - T2e;
|
Chris@10
|
1143 T2g = T2c - T2f;
|
Chris@10
|
1144 T2o = T2c + T2f;
|
Chris@10
|
1145 }
|
Chris@10
|
1146 }
|
Chris@10
|
1147 {
|
Chris@10
|
1148 E TR, TY, T5e, T5h;
|
Chris@10
|
1149 TR = TN + TQ;
|
Chris@10
|
1150 TY = TU + TX;
|
Chris@10
|
1151 TZ = TR + TY;
|
Chris@10
|
1152 T6l = TR - TY;
|
Chris@10
|
1153 T5e = TU - TX;
|
Chris@10
|
1154 T5h = T5f - T5g;
|
Chris@10
|
1155 T5i = T5e + T5h;
|
Chris@10
|
1156 T5Q = T5h - T5e;
|
Chris@10
|
1157 }
|
Chris@10
|
1158 {
|
Chris@10
|
1159 E T6m, T6n, T2b, T2m;
|
Chris@10
|
1160 T6m = T5g + T5f;
|
Chris@10
|
1161 T6n = T5b + T5a;
|
Chris@10
|
1162 T6o = T6m - T6n;
|
Chris@10
|
1163 T6M = T6n + T6m;
|
Chris@10
|
1164 T2b = T27 - T2a;
|
Chris@10
|
1165 T2m = KP707106781 * (T2g + T2l);
|
Chris@10
|
1166 T2n = T2b + T2m;
|
Chris@10
|
1167 T37 = T2b - T2m;
|
Chris@10
|
1168 }
|
Chris@10
|
1169 {
|
Chris@10
|
1170 E T2q, T2v, T3O, T3P;
|
Chris@10
|
1171 T2q = KP707106781 * (T2o - T2p);
|
Chris@10
|
1172 T2v = T2r - T2u;
|
Chris@10
|
1173 T2w = T2q + T2v;
|
Chris@10
|
1174 T38 = T2v - T2q;
|
Chris@10
|
1175 T3O = KP707106781 * (T2g - T2l);
|
Chris@10
|
1176 T3P = T2r + T2u;
|
Chris@10
|
1177 T3Q = T3O - T3P;
|
Chris@10
|
1178 T4m = T3O + T3P;
|
Chris@10
|
1179 }
|
Chris@10
|
1180 {
|
Chris@10
|
1181 E T59, T5c, T3L, T3M;
|
Chris@10
|
1182 T59 = TN - TQ;
|
Chris@10
|
1183 T5c = T5a - T5b;
|
Chris@10
|
1184 T5d = T59 + T5c;
|
Chris@10
|
1185 T5P = T59 - T5c;
|
Chris@10
|
1186 T3L = T27 + T2a;
|
Chris@10
|
1187 T3M = KP707106781 * (T2o + T2p);
|
Chris@10
|
1188 T3N = T3L - T3M;
|
Chris@10
|
1189 T4l = T3L + T3M;
|
Chris@10
|
1190 }
|
Chris@10
|
1191 }
|
Chris@10
|
1192 {
|
Chris@10
|
1193 E Tv, T10, T6X, T6Y, T6Z, T70;
|
Chris@10
|
1194 Tv = Tf + Tu;
|
Chris@10
|
1195 T10 = TK + TZ;
|
Chris@10
|
1196 T6X = Tv - T10;
|
Chris@10
|
1197 T6Y = T6N + T6M;
|
Chris@10
|
1198 T6Z = T6R - T6Q;
|
Chris@10
|
1199 T70 = T6Y + T6Z;
|
Chris@10
|
1200 R0[0] = KP2_000000000 * (Tv + T10);
|
Chris@10
|
1201 R0[WS(rs, 16)] = KP2_000000000 * (T6Z - T6Y);
|
Chris@10
|
1202 R0[WS(rs, 8)] = KP1_414213562 * (T6X + T70);
|
Chris@10
|
1203 R0[WS(rs, 24)] = KP1_414213562 * (T70 - T6X);
|
Chris@10
|
1204 }
|
Chris@10
|
1205 {
|
Chris@10
|
1206 E T6P, T6V, T6U, T6W;
|
Chris@10
|
1207 {
|
Chris@10
|
1208 E T6L, T6O, T6S, T6T;
|
Chris@10
|
1209 T6L = Tf - Tu;
|
Chris@10
|
1210 T6O = T6M - T6N;
|
Chris@10
|
1211 T6P = T6L + T6O;
|
Chris@10
|
1212 T6V = T6L - T6O;
|
Chris@10
|
1213 T6S = T6Q + T6R;
|
Chris@10
|
1214 T6T = TK - TZ;
|
Chris@10
|
1215 T6U = T6S - T6T;
|
Chris@10
|
1216 T6W = T6T + T6S;
|
Chris@10
|
1217 }
|
Chris@10
|
1218 R0[WS(rs, 4)] = FMA(KP1_847759065, T6P, KP765366864 * T6U);
|
Chris@10
|
1219 R0[WS(rs, 28)] = FNMS(KP1_847759065, T6V, KP765366864 * T6W);
|
Chris@10
|
1220 R0[WS(rs, 20)] = FNMS(KP765366864, T6P, KP1_847759065 * T6U);
|
Chris@10
|
1221 R0[WS(rs, 12)] = FMA(KP765366864, T6V, KP1_847759065 * T6W);
|
Chris@10
|
1222 }
|
Chris@10
|
1223 {
|
Chris@10
|
1224 E T6f, T6w, T6G, T6D, T6z, T6E, T6q, T6H;
|
Chris@10
|
1225 T6f = T6b + T6e;
|
Chris@10
|
1226 T6w = T6u - T6v;
|
Chris@10
|
1227 T6G = T6v + T6u;
|
Chris@10
|
1228 T6D = T6b - T6e;
|
Chris@10
|
1229 {
|
Chris@10
|
1230 E T6x, T6y, T6k, T6p;
|
Chris@10
|
1231 T6x = T6g + T6j;
|
Chris@10
|
1232 T6y = T6o - T6l;
|
Chris@10
|
1233 T6z = KP707106781 * (T6x + T6y);
|
Chris@10
|
1234 T6E = KP707106781 * (T6y - T6x);
|
Chris@10
|
1235 T6k = T6g - T6j;
|
Chris@10
|
1236 T6p = T6l + T6o;
|
Chris@10
|
1237 T6q = KP707106781 * (T6k + T6p);
|
Chris@10
|
1238 T6H = KP707106781 * (T6k - T6p);
|
Chris@10
|
1239 }
|
Chris@10
|
1240 {
|
Chris@10
|
1241 E T6r, T6A, T6J, T6K;
|
Chris@10
|
1242 T6r = T6f + T6q;
|
Chris@10
|
1243 T6A = T6w - T6z;
|
Chris@10
|
1244 R0[WS(rs, 2)] = FMA(KP1_961570560, T6r, KP390180644 * T6A);
|
Chris@10
|
1245 R0[WS(rs, 18)] = FNMS(KP390180644, T6r, KP1_961570560 * T6A);
|
Chris@10
|
1246 T6J = T6D - T6E;
|
Chris@10
|
1247 T6K = T6H + T6G;
|
Chris@10
|
1248 R0[WS(rs, 14)] = FMA(KP390180644, T6J, KP1_961570560 * T6K);
|
Chris@10
|
1249 R0[WS(rs, 30)] = FNMS(KP1_961570560, T6J, KP390180644 * T6K);
|
Chris@10
|
1250 }
|
Chris@10
|
1251 {
|
Chris@10
|
1252 E T6B, T6C, T6F, T6I;
|
Chris@10
|
1253 T6B = T6f - T6q;
|
Chris@10
|
1254 T6C = T6z + T6w;
|
Chris@10
|
1255 R0[WS(rs, 10)] = FMA(KP1_111140466, T6B, KP1_662939224 * T6C);
|
Chris@10
|
1256 R0[WS(rs, 26)] = FNMS(KP1_662939224, T6B, KP1_111140466 * T6C);
|
Chris@10
|
1257 T6F = T6D + T6E;
|
Chris@10
|
1258 T6I = T6G - T6H;
|
Chris@10
|
1259 R0[WS(rs, 6)] = FMA(KP1_662939224, T6F, KP1_111140466 * T6I);
|
Chris@10
|
1260 R0[WS(rs, 22)] = FNMS(KP1_111140466, T6F, KP1_662939224 * T6I);
|
Chris@10
|
1261 }
|
Chris@10
|
1262 }
|
Chris@10
|
1263 {
|
Chris@10
|
1264 E T5L, T63, T5W, T66, T5S, T67, T5Z, T64, T5K, T5V;
|
Chris@10
|
1265 T5K = KP707106781 * (T5s - T5r);
|
Chris@10
|
1266 T5L = T5J + T5K;
|
Chris@10
|
1267 T63 = T5J - T5K;
|
Chris@10
|
1268 T5V = KP707106781 * (T4Q - T4V);
|
Chris@10
|
1269 T5W = T5U - T5V;
|
Chris@10
|
1270 T66 = T5V + T5U;
|
Chris@10
|
1271 {
|
Chris@10
|
1272 E T5O, T5R, T5X, T5Y;
|
Chris@10
|
1273 T5O = FNMS(KP923879532, T5N, KP382683432 * T5M);
|
Chris@10
|
1274 T5R = FMA(KP382683432, T5P, KP923879532 * T5Q);
|
Chris@10
|
1275 T5S = T5O + T5R;
|
Chris@10
|
1276 T67 = T5O - T5R;
|
Chris@10
|
1277 T5X = FMA(KP923879532, T5M, KP382683432 * T5N);
|
Chris@10
|
1278 T5Y = FNMS(KP923879532, T5P, KP382683432 * T5Q);
|
Chris@10
|
1279 T5Z = T5X + T5Y;
|
Chris@10
|
1280 T64 = T5Y - T5X;
|
Chris@10
|
1281 }
|
Chris@10
|
1282 {
|
Chris@10
|
1283 E T5T, T60, T69, T6a;
|
Chris@10
|
1284 T5T = T5L + T5S;
|
Chris@10
|
1285 T60 = T5W - T5Z;
|
Chris@10
|
1286 R0[WS(rs, 3)] = FMA(KP1_913880671, T5T, KP580569354 * T60);
|
Chris@10
|
1287 R0[WS(rs, 19)] = FNMS(KP580569354, T5T, KP1_913880671 * T60);
|
Chris@10
|
1288 T69 = T63 - T64;
|
Chris@10
|
1289 T6a = T67 + T66;
|
Chris@10
|
1290 R0[WS(rs, 15)] = FMA(KP196034280, T69, KP1_990369453 * T6a);
|
Chris@10
|
1291 R0[WS(rs, 31)] = FNMS(KP1_990369453, T69, KP196034280 * T6a);
|
Chris@10
|
1292 }
|
Chris@10
|
1293 {
|
Chris@10
|
1294 E T61, T62, T65, T68;
|
Chris@10
|
1295 T61 = T5L - T5S;
|
Chris@10
|
1296 T62 = T5Z + T5W;
|
Chris@10
|
1297 R0[WS(rs, 11)] = FMA(KP942793473, T61, KP1_763842528 * T62);
|
Chris@10
|
1298 R0[WS(rs, 27)] = FNMS(KP1_763842528, T61, KP942793473 * T62);
|
Chris@10
|
1299 T65 = T63 + T64;
|
Chris@10
|
1300 T68 = T66 - T67;
|
Chris@10
|
1301 R0[WS(rs, 7)] = FMA(KP1_546020906, T65, KP1_268786568 * T68);
|
Chris@10
|
1302 R0[WS(rs, 23)] = FNMS(KP1_268786568, T65, KP1_546020906 * T68);
|
Chris@10
|
1303 }
|
Chris@10
|
1304 }
|
Chris@10
|
1305 {
|
Chris@10
|
1306 E T4X, T5B, T5u, T5E, T5k, T5F, T5x, T5C, T4W, T5t;
|
Chris@10
|
1307 T4W = KP707106781 * (T4Q + T4V);
|
Chris@10
|
1308 T4X = T4L + T4W;
|
Chris@10
|
1309 T5B = T4L - T4W;
|
Chris@10
|
1310 T5t = KP707106781 * (T5r + T5s);
|
Chris@10
|
1311 T5u = T5q - T5t;
|
Chris@10
|
1312 T5E = T5t + T5q;
|
Chris@10
|
1313 {
|
Chris@10
|
1314 E T58, T5j, T5v, T5w;
|
Chris@10
|
1315 T58 = FNMS(KP382683432, T57, KP923879532 * T52);
|
Chris@10
|
1316 T5j = FMA(KP923879532, T5d, KP382683432 * T5i);
|
Chris@10
|
1317 T5k = T58 + T5j;
|
Chris@10
|
1318 T5F = T58 - T5j;
|
Chris@10
|
1319 T5v = FMA(KP382683432, T52, KP923879532 * T57);
|
Chris@10
|
1320 T5w = FNMS(KP382683432, T5d, KP923879532 * T5i);
|
Chris@10
|
1321 T5x = T5v + T5w;
|
Chris@10
|
1322 T5C = T5w - T5v;
|
Chris@10
|
1323 }
|
Chris@10
|
1324 {
|
Chris@10
|
1325 E T5l, T5y, T5H, T5I;
|
Chris@10
|
1326 T5l = T4X + T5k;
|
Chris@10
|
1327 T5y = T5u - T5x;
|
Chris@10
|
1328 R0[WS(rs, 1)] = FMA(KP1_990369453, T5l, KP196034280 * T5y);
|
Chris@10
|
1329 R0[WS(rs, 17)] = FNMS(KP196034280, T5l, KP1_990369453 * T5y);
|
Chris@10
|
1330 T5H = T5B - T5C;
|
Chris@10
|
1331 T5I = T5F + T5E;
|
Chris@10
|
1332 R0[WS(rs, 13)] = FMA(KP580569354, T5H, KP1_913880671 * T5I);
|
Chris@10
|
1333 R0[WS(rs, 29)] = FNMS(KP1_913880671, T5H, KP580569354 * T5I);
|
Chris@10
|
1334 }
|
Chris@10
|
1335 {
|
Chris@10
|
1336 E T5z, T5A, T5D, T5G;
|
Chris@10
|
1337 T5z = T4X - T5k;
|
Chris@10
|
1338 T5A = T5x + T5u;
|
Chris@10
|
1339 R0[WS(rs, 9)] = FMA(KP1_268786568, T5z, KP1_546020906 * T5A);
|
Chris@10
|
1340 R0[WS(rs, 25)] = FNMS(KP1_546020906, T5z, KP1_268786568 * T5A);
|
Chris@10
|
1341 T5D = T5B + T5C;
|
Chris@10
|
1342 T5G = T5E - T5F;
|
Chris@10
|
1343 R0[WS(rs, 5)] = FMA(KP1_763842528, T5D, KP942793473 * T5G);
|
Chris@10
|
1344 R0[WS(rs, 21)] = FNMS(KP942793473, T5D, KP1_763842528 * T5G);
|
Chris@10
|
1345 }
|
Chris@10
|
1346 }
|
Chris@10
|
1347 {
|
Chris@10
|
1348 E T33, T3l, T3h, T3m, T3a, T3p, T3e, T3o;
|
Chris@10
|
1349 {
|
Chris@10
|
1350 E T31, T32, T3f, T3g;
|
Chris@10
|
1351 T31 = T15 - T1g;
|
Chris@10
|
1352 T32 = T2E - T2D;
|
Chris@10
|
1353 T33 = T31 + T32;
|
Chris@10
|
1354 T3l = T31 - T32;
|
Chris@10
|
1355 T3f = FMA(KP831469612, T34, KP555570233 * T35);
|
Chris@10
|
1356 T3g = FNMS(KP831469612, T37, KP555570233 * T38);
|
Chris@10
|
1357 T3h = T3f + T3g;
|
Chris@10
|
1358 T3m = T3g - T3f;
|
Chris@10
|
1359 }
|
Chris@10
|
1360 {
|
Chris@10
|
1361 E T36, T39, T3c, T3d;
|
Chris@10
|
1362 T36 = FNMS(KP831469612, T35, KP555570233 * T34);
|
Chris@10
|
1363 T39 = FMA(KP555570233, T37, KP831469612 * T38);
|
Chris@10
|
1364 T3a = T36 + T39;
|
Chris@10
|
1365 T3p = T36 - T39;
|
Chris@10
|
1366 T3c = T2I - T2N;
|
Chris@10
|
1367 T3d = T1s - T1D;
|
Chris@10
|
1368 T3e = T3c - T3d;
|
Chris@10
|
1369 T3o = T3d + T3c;
|
Chris@10
|
1370 }
|
Chris@10
|
1371 {
|
Chris@10
|
1372 E T3b, T3i, T3r, T3s;
|
Chris@10
|
1373 T3b = T33 + T3a;
|
Chris@10
|
1374 T3i = T3e - T3h;
|
Chris@10
|
1375 R1[WS(rs, 2)] = FMA(KP1_940062506, T3b, KP485960359 * T3i);
|
Chris@10
|
1376 R1[WS(rs, 18)] = FNMS(KP485960359, T3b, KP1_940062506 * T3i);
|
Chris@10
|
1377 T3r = T3l - T3m;
|
Chris@10
|
1378 T3s = T3p + T3o;
|
Chris@10
|
1379 R1[WS(rs, 14)] = FMA(KP293460948, T3r, KP1_978353019 * T3s);
|
Chris@10
|
1380 R1[WS(rs, 30)] = FNMS(KP1_978353019, T3r, KP293460948 * T3s);
|
Chris@10
|
1381 }
|
Chris@10
|
1382 {
|
Chris@10
|
1383 E T3j, T3k, T3n, T3q;
|
Chris@10
|
1384 T3j = T33 - T3a;
|
Chris@10
|
1385 T3k = T3h + T3e;
|
Chris@10
|
1386 R1[WS(rs, 10)] = FMA(KP1_028205488, T3j, KP1_715457220 * T3k);
|
Chris@10
|
1387 R1[WS(rs, 26)] = FNMS(KP1_715457220, T3j, KP1_028205488 * T3k);
|
Chris@10
|
1388 T3n = T3l + T3m;
|
Chris@10
|
1389 T3q = T3o - T3p;
|
Chris@10
|
1390 R1[WS(rs, 6)] = FMA(KP1_606415062, T3n, KP1_191398608 * T3q);
|
Chris@10
|
1391 R1[WS(rs, 22)] = FNMS(KP1_191398608, T3n, KP1_606415062 * T3q);
|
Chris@10
|
1392 }
|
Chris@10
|
1393 }
|
Chris@10
|
1394 {
|
Chris@10
|
1395 E T4h, T4z, T4v, T4A, T4o, T4D, T4s, T4C;
|
Chris@10
|
1396 {
|
Chris@10
|
1397 E T4f, T4g, T4t, T4u;
|
Chris@10
|
1398 T4f = T3t + T3u;
|
Chris@10
|
1399 T4g = T3X + T3Y;
|
Chris@10
|
1400 T4h = T4f - T4g;
|
Chris@10
|
1401 T4z = T4f + T4g;
|
Chris@10
|
1402 T4t = FMA(KP980785280, T4i, KP195090322 * T4j);
|
Chris@10
|
1403 T4u = FMA(KP980785280, T4l, KP195090322 * T4m);
|
Chris@10
|
1404 T4v = T4t - T4u;
|
Chris@10
|
1405 T4A = T4t + T4u;
|
Chris@10
|
1406 }
|
Chris@10
|
1407 {
|
Chris@10
|
1408 E T4k, T4n, T4q, T4r;
|
Chris@10
|
1409 T4k = FNMS(KP980785280, T4j, KP195090322 * T4i);
|
Chris@10
|
1410 T4n = FNMS(KP980785280, T4m, KP195090322 * T4l);
|
Chris@10
|
1411 T4o = T4k + T4n;
|
Chris@10
|
1412 T4D = T4k - T4n;
|
Chris@10
|
1413 T4q = T3V + T3U;
|
Chris@10
|
1414 T4r = T3y - T3B;
|
Chris@10
|
1415 T4s = T4q - T4r;
|
Chris@10
|
1416 T4C = T4r + T4q;
|
Chris@10
|
1417 }
|
Chris@10
|
1418 {
|
Chris@10
|
1419 E T4p, T4w, T4F, T4G;
|
Chris@10
|
1420 T4p = T4h + T4o;
|
Chris@10
|
1421 T4w = T4s - T4v;
|
Chris@10
|
1422 R1[WS(rs, 3)] = FMA(KP1_883088130, T4p, KP673779706 * T4w);
|
Chris@10
|
1423 R1[WS(rs, 19)] = FNMS(KP673779706, T4p, KP1_883088130 * T4w);
|
Chris@10
|
1424 T4F = T4z + T4A;
|
Chris@10
|
1425 T4G = T4D + T4C;
|
Chris@10
|
1426 R1[WS(rs, 15)] = FMA(KP098135348, T4F, KP1_997590912 * T4G);
|
Chris@10
|
1427 R1[WS(rs, 31)] = FNMS(KP1_997590912, T4F, KP098135348 * T4G);
|
Chris@10
|
1428 }
|
Chris@10
|
1429 {
|
Chris@10
|
1430 E T4x, T4y, T4B, T4E;
|
Chris@10
|
1431 T4x = T4h - T4o;
|
Chris@10
|
1432 T4y = T4v + T4s;
|
Chris@10
|
1433 R1[WS(rs, 11)] = FMA(KP855110186, T4x, KP1_807978586 * T4y);
|
Chris@10
|
1434 R1[WS(rs, 27)] = FNMS(KP1_807978586, T4x, KP855110186 * T4y);
|
Chris@10
|
1435 T4B = T4z - T4A;
|
Chris@10
|
1436 T4E = T4C - T4D;
|
Chris@10
|
1437 R1[WS(rs, 7)] = FMA(KP1_481902250, T4B, KP1_343117909 * T4E);
|
Chris@10
|
1438 R1[WS(rs, 23)] = FNMS(KP1_343117909, T4B, KP1_481902250 * T4E);
|
Chris@10
|
1439 }
|
Chris@10
|
1440 }
|
Chris@10
|
1441 {
|
Chris@10
|
1442 E T1F, T2T, T2P, T2W, T2y, T2X, T2C, T2U;
|
Chris@10
|
1443 {
|
Chris@10
|
1444 E T1h, T1E, T2F, T2O;
|
Chris@10
|
1445 T1h = T15 + T1g;
|
Chris@10
|
1446 T1E = T1s + T1D;
|
Chris@10
|
1447 T1F = T1h + T1E;
|
Chris@10
|
1448 T2T = T1h - T1E;
|
Chris@10
|
1449 T2F = T2D + T2E;
|
Chris@10
|
1450 T2O = T2I + T2N;
|
Chris@10
|
1451 T2P = T2F + T2O;
|
Chris@10
|
1452 T2W = T2F - T2O;
|
Chris@10
|
1453 }
|
Chris@10
|
1454 {
|
Chris@10
|
1455 E T26, T2x, T2A, T2B;
|
Chris@10
|
1456 T26 = FNMS(KP195090322, T25, KP980785280 * T1W);
|
Chris@10
|
1457 T2x = FMA(KP980785280, T2n, KP195090322 * T2w);
|
Chris@10
|
1458 T2y = T26 + T2x;
|
Chris@10
|
1459 T2X = T26 - T2x;
|
Chris@10
|
1460 T2A = FMA(KP195090322, T1W, KP980785280 * T25);
|
Chris@10
|
1461 T2B = FNMS(KP195090322, T2n, KP980785280 * T2w);
|
Chris@10
|
1462 T2C = T2A + T2B;
|
Chris@10
|
1463 T2U = T2B - T2A;
|
Chris@10
|
1464 }
|
Chris@10
|
1465 {
|
Chris@10
|
1466 E T2z, T2Q, T2Z, T30;
|
Chris@10
|
1467 T2z = T1F + T2y;
|
Chris@10
|
1468 T2Q = T2C + T2P;
|
Chris@10
|
1469 R1[0] = FNMS(KP098135348, T2Q, KP1_997590912 * T2z);
|
Chris@10
|
1470 R1[WS(rs, 16)] = -(FMA(KP098135348, T2z, KP1_997590912 * T2Q));
|
Chris@10
|
1471 T2Z = T2T - T2U;
|
Chris@10
|
1472 T30 = T2X + T2W;
|
Chris@10
|
1473 R1[WS(rs, 12)] = FMA(KP673779706, T2Z, KP1_883088130 * T30);
|
Chris@10
|
1474 R1[WS(rs, 28)] = FNMS(KP1_883088130, T2Z, KP673779706 * T30);
|
Chris@10
|
1475 }
|
Chris@10
|
1476 {
|
Chris@10
|
1477 E T2R, T2S, T2V, T2Y;
|
Chris@10
|
1478 T2R = T1F - T2y;
|
Chris@10
|
1479 T2S = T2C - T2P;
|
Chris@10
|
1480 R1[WS(rs, 8)] = FMA(KP1_343117909, T2R, KP1_481902250 * T2S);
|
Chris@10
|
1481 R1[WS(rs, 24)] = FNMS(KP1_481902250, T2R, KP1_343117909 * T2S);
|
Chris@10
|
1482 T2V = T2T + T2U;
|
Chris@10
|
1483 T2Y = T2W - T2X;
|
Chris@10
|
1484 R1[WS(rs, 4)] = FMA(KP1_807978586, T2V, KP855110186 * T2Y);
|
Chris@10
|
1485 R1[WS(rs, 20)] = FNMS(KP855110186, T2V, KP1_807978586 * T2Y);
|
Chris@10
|
1486 }
|
Chris@10
|
1487 }
|
Chris@10
|
1488 {
|
Chris@10
|
1489 E T3D, T47, T43, T48, T3S, T4b, T40, T4a;
|
Chris@10
|
1490 {
|
Chris@10
|
1491 E T3v, T3C, T41, T42;
|
Chris@10
|
1492 T3v = T3t - T3u;
|
Chris@10
|
1493 T3C = T3y + T3B;
|
Chris@10
|
1494 T3D = T3v + T3C;
|
Chris@10
|
1495 T47 = T3v - T3C;
|
Chris@10
|
1496 T41 = FMA(KP555570233, T3G, KP831469612 * T3J);
|
Chris@10
|
1497 T42 = FNMS(KP555570233, T3N, KP831469612 * T3Q);
|
Chris@10
|
1498 T43 = T41 + T42;
|
Chris@10
|
1499 T48 = T42 - T41;
|
Chris@10
|
1500 }
|
Chris@10
|
1501 {
|
Chris@10
|
1502 E T3K, T3R, T3W, T3Z;
|
Chris@10
|
1503 T3K = FNMS(KP555570233, T3J, KP831469612 * T3G);
|
Chris@10
|
1504 T3R = FMA(KP831469612, T3N, KP555570233 * T3Q);
|
Chris@10
|
1505 T3S = T3K + T3R;
|
Chris@10
|
1506 T4b = T3K - T3R;
|
Chris@10
|
1507 T3W = T3U - T3V;
|
Chris@10
|
1508 T3Z = T3X - T3Y;
|
Chris@10
|
1509 T40 = T3W - T3Z;
|
Chris@10
|
1510 T4a = T3Z + T3W;
|
Chris@10
|
1511 }
|
Chris@10
|
1512 {
|
Chris@10
|
1513 E T3T, T44, T4d, T4e;
|
Chris@10
|
1514 T3T = T3D + T3S;
|
Chris@10
|
1515 T44 = T40 - T43;
|
Chris@10
|
1516 R1[WS(rs, 1)] = FMA(KP1_978353019, T3T, KP293460948 * T44);
|
Chris@10
|
1517 R1[WS(rs, 17)] = FNMS(KP293460948, T3T, KP1_978353019 * T44);
|
Chris@10
|
1518 T4d = T47 - T48;
|
Chris@10
|
1519 T4e = T4b + T4a;
|
Chris@10
|
1520 R1[WS(rs, 13)] = FMA(KP485960359, T4d, KP1_940062506 * T4e);
|
Chris@10
|
1521 R1[WS(rs, 29)] = FNMS(KP1_940062506, T4d, KP485960359 * T4e);
|
Chris@10
|
1522 }
|
Chris@10
|
1523 {
|
Chris@10
|
1524 E T45, T46, T49, T4c;
|
Chris@10
|
1525 T45 = T3D - T3S;
|
Chris@10
|
1526 T46 = T43 + T40;
|
Chris@10
|
1527 R1[WS(rs, 9)] = FMA(KP1_191398608, T45, KP1_606415062 * T46);
|
Chris@10
|
1528 R1[WS(rs, 25)] = FNMS(KP1_606415062, T45, KP1_191398608 * T46);
|
Chris@10
|
1529 T49 = T47 + T48;
|
Chris@10
|
1530 T4c = T4a - T4b;
|
Chris@10
|
1531 R1[WS(rs, 5)] = FMA(KP1_715457220, T49, KP1_028205488 * T4c);
|
Chris@10
|
1532 R1[WS(rs, 21)] = FNMS(KP1_028205488, T49, KP1_715457220 * T4c);
|
Chris@10
|
1533 }
|
Chris@10
|
1534 }
|
Chris@10
|
1535 }
|
Chris@10
|
1536 }
|
Chris@10
|
1537 }
|
Chris@10
|
1538
|
Chris@10
|
1539 static const kr2c_desc desc = { 64, "r2cbIII_64", {342, 116, 92, 0}, &GENUS };
|
Chris@10
|
1540
|
Chris@10
|
1541 void X(codelet_r2cbIII_64) (planner *p) {
|
Chris@10
|
1542 X(kr2c_register) (p, r2cbIII_64, &desc);
|
Chris@10
|
1543 }
|
Chris@10
|
1544
|
Chris@10
|
1545 #endif /* HAVE_FMA */
|