cannam@167
|
1 /*
|
cannam@167
|
2 * Copyright (c) 2003, 2007-14 Matteo Frigo
|
cannam@167
|
3 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
|
cannam@167
|
4 *
|
cannam@167
|
5 * This program is free software; you can redistribute it and/or modify
|
cannam@167
|
6 * it under the terms of the GNU General Public License as published by
|
cannam@167
|
7 * the Free Software Foundation; either version 2 of the License, or
|
cannam@167
|
8 * (at your option) any later version.
|
cannam@167
|
9 *
|
cannam@167
|
10 * This program is distributed in the hope that it will be useful,
|
cannam@167
|
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
cannam@167
|
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
cannam@167
|
13 * GNU General Public License for more details.
|
cannam@167
|
14 *
|
cannam@167
|
15 * You should have received a copy of the GNU General Public License
|
cannam@167
|
16 * along with this program; if not, write to the Free Software
|
cannam@167
|
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
cannam@167
|
18 *
|
cannam@167
|
19 */
|
cannam@167
|
20
|
cannam@167
|
21 /* This file was automatically generated --- DO NOT EDIT */
|
cannam@167
|
22 /* Generated on Thu May 24 08:06:27 EDT 2018 */
|
cannam@167
|
23
|
cannam@167
|
24 #include "rdft/codelet-rdft.h"
|
cannam@167
|
25
|
cannam@167
|
26 #if defined(ARCH_PREFERS_FMA) || defined(ISA_EXTENSION_PREFERS_FMA)
|
cannam@167
|
27
|
cannam@167
|
28 /* Generated by: ../../../genfft/gen_r2cf.native -fma -compact -variables 4 -pipeline-latency 4 -n 64 -name r2cf_64 -include rdft/scalar/r2cf.h */
|
cannam@167
|
29
|
cannam@167
|
30 /*
|
cannam@167
|
31 * This function contains 394 FP additions, 196 FP multiplications,
|
cannam@167
|
32 * (or, 198 additions, 0 multiplications, 196 fused multiply/add),
|
cannam@167
|
33 * 106 stack variables, 15 constants, and 128 memory accesses
|
cannam@167
|
34 */
|
cannam@167
|
35 #include "rdft/scalar/r2cf.h"
|
cannam@167
|
36
|
cannam@167
|
37 static void r2cf_64(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
|
cannam@167
|
38 {
|
cannam@167
|
39 DK(KP773010453, +0.773010453362736960810906609758469800971041293);
|
cannam@167
|
40 DK(KP995184726, +0.995184726672196886244836953109479921575474869);
|
cannam@167
|
41 DK(KP098491403, +0.098491403357164253077197521291327432293052451);
|
cannam@167
|
42 DK(KP820678790, +0.820678790828660330972281985331011598767386482);
|
cannam@167
|
43 DK(KP956940335, +0.956940335732208864935797886980269969482849206);
|
cannam@167
|
44 DK(KP881921264, +0.881921264348355029712756863660388349508442621);
|
cannam@167
|
45 DK(KP534511135, +0.534511135950791641089685961295362908582039528);
|
cannam@167
|
46 DK(KP303346683, +0.303346683607342391675883946941299872384187453);
|
cannam@167
|
47 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
cannam@167
|
48 DK(KP198912367, +0.198912367379658006911597622644676228597850501);
|
cannam@167
|
49 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
cannam@167
|
50 DK(KP668178637, +0.668178637919298919997757686523080761552472251);
|
cannam@167
|
51 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
cannam@167
|
52 DK(KP414213562, +0.414213562373095048801688724209698078569671875);
|
cannam@167
|
53 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
cannam@167
|
54 {
|
cannam@167
|
55 INT i;
|
cannam@167
|
56 for (i = v; i > 0; i = i - 1, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(256, rs), MAKE_VOLATILE_STRIDE(256, csr), MAKE_VOLATILE_STRIDE(256, csi)) {
|
cannam@167
|
57 E T11, T2j, T3D, T5p, T4P, T5P, T7, Te, Tf, T1k, T1H, T4a, T5A, T4l, T5D;
|
cannam@167
|
58 E T2U, T3i, T1R, T2e, T4v, T5H, T4G, T5K, T31, T3l, T3Z, T5t, T42, T5s, TZ;
|
cannam@167
|
59 E T3f, T1b, T2n, T3Q, T5w, T3T, T5v, TK, T3e, T18, T2m, Tm, Tt, Tu, T4S;
|
cannam@167
|
60 E T5q, T14, T2k, T3K, T5Q, T1z, T1I, T4o, T5B, T2X, T3j, T4h, T5E, T26, T2f;
|
cannam@167
|
61 E T4J, T5I, T34, T3m, T4C, T5L;
|
cannam@167
|
62 {
|
cannam@167
|
63 E T3, T3z, Td, T3B, T6, T4N, Ta, T3A, T3C, T4O;
|
cannam@167
|
64 {
|
cannam@167
|
65 E T1, T2, Tb, Tc;
|
cannam@167
|
66 T1 = R0[0];
|
cannam@167
|
67 T2 = R0[WS(rs, 16)];
|
cannam@167
|
68 T3 = T1 + T2;
|
cannam@167
|
69 T3z = T1 - T2;
|
cannam@167
|
70 Tb = R0[WS(rs, 28)];
|
cannam@167
|
71 Tc = R0[WS(rs, 12)];
|
cannam@167
|
72 Td = Tb + Tc;
|
cannam@167
|
73 T3B = Tb - Tc;
|
cannam@167
|
74 }
|
cannam@167
|
75 {
|
cannam@167
|
76 E T4, T5, T8, T9;
|
cannam@167
|
77 T4 = R0[WS(rs, 8)];
|
cannam@167
|
78 T5 = R0[WS(rs, 24)];
|
cannam@167
|
79 T6 = T4 + T5;
|
cannam@167
|
80 T4N = T4 - T5;
|
cannam@167
|
81 T8 = R0[WS(rs, 4)];
|
cannam@167
|
82 T9 = R0[WS(rs, 20)];
|
cannam@167
|
83 Ta = T8 + T9;
|
cannam@167
|
84 T3A = T8 - T9;
|
cannam@167
|
85 }
|
cannam@167
|
86 T11 = T3 - T6;
|
cannam@167
|
87 T2j = Td - Ta;
|
cannam@167
|
88 T3C = T3A + T3B;
|
cannam@167
|
89 T3D = FMA(KP707106781, T3C, T3z);
|
cannam@167
|
90 T5p = FNMS(KP707106781, T3C, T3z);
|
cannam@167
|
91 T4O = T3B - T3A;
|
cannam@167
|
92 T4P = FNMS(KP707106781, T4O, T4N);
|
cannam@167
|
93 T5P = FMA(KP707106781, T4O, T4N);
|
cannam@167
|
94 T7 = T3 + T6;
|
cannam@167
|
95 Te = Ta + Td;
|
cannam@167
|
96 Tf = T7 + Te;
|
cannam@167
|
97 }
|
cannam@167
|
98 {
|
cannam@167
|
99 E T1g, T46, T1G, T47, T1j, T4j, T1D, T48;
|
cannam@167
|
100 {
|
cannam@167
|
101 E T1e, T1f, T1E, T1F;
|
cannam@167
|
102 T1e = R1[0];
|
cannam@167
|
103 T1f = R1[WS(rs, 16)];
|
cannam@167
|
104 T1g = T1e + T1f;
|
cannam@167
|
105 T46 = T1e - T1f;
|
cannam@167
|
106 T1E = R1[WS(rs, 4)];
|
cannam@167
|
107 T1F = R1[WS(rs, 20)];
|
cannam@167
|
108 T1G = T1E + T1F;
|
cannam@167
|
109 T47 = T1E - T1F;
|
cannam@167
|
110 }
|
cannam@167
|
111 {
|
cannam@167
|
112 E T1h, T1i, T1B, T1C;
|
cannam@167
|
113 T1h = R1[WS(rs, 8)];
|
cannam@167
|
114 T1i = R1[WS(rs, 24)];
|
cannam@167
|
115 T1j = T1h + T1i;
|
cannam@167
|
116 T4j = T1h - T1i;
|
cannam@167
|
117 T1B = R1[WS(rs, 28)];
|
cannam@167
|
118 T1C = R1[WS(rs, 12)];
|
cannam@167
|
119 T1D = T1B + T1C;
|
cannam@167
|
120 T48 = T1B - T1C;
|
cannam@167
|
121 }
|
cannam@167
|
122 T1k = T1g - T1j;
|
cannam@167
|
123 T1H = T1D - T1G;
|
cannam@167
|
124 {
|
cannam@167
|
125 E T49, T4k, T2S, T2T;
|
cannam@167
|
126 T49 = T47 + T48;
|
cannam@167
|
127 T4a = FMA(KP707106781, T49, T46);
|
cannam@167
|
128 T5A = FNMS(KP707106781, T49, T46);
|
cannam@167
|
129 T4k = T47 - T48;
|
cannam@167
|
130 T4l = FMA(KP707106781, T4k, T4j);
|
cannam@167
|
131 T5D = FNMS(KP707106781, T4k, T4j);
|
cannam@167
|
132 T2S = T1g + T1j;
|
cannam@167
|
133 T2T = T1G + T1D;
|
cannam@167
|
134 T2U = T2S + T2T;
|
cannam@167
|
135 T3i = T2S - T2T;
|
cannam@167
|
136 }
|
cannam@167
|
137 }
|
cannam@167
|
138 {
|
cannam@167
|
139 E T1N, T4r, T2d, T4s, T1Q, T4E, T2a, T4t;
|
cannam@167
|
140 {
|
cannam@167
|
141 E T1L, T1M, T2b, T2c;
|
cannam@167
|
142 T1L = R1[WS(rs, 31)];
|
cannam@167
|
143 T1M = R1[WS(rs, 15)];
|
cannam@167
|
144 T1N = T1L + T1M;
|
cannam@167
|
145 T4r = T1L - T1M;
|
cannam@167
|
146 T2b = R1[WS(rs, 3)];
|
cannam@167
|
147 T2c = R1[WS(rs, 19)];
|
cannam@167
|
148 T2d = T2b + T2c;
|
cannam@167
|
149 T4s = T2b - T2c;
|
cannam@167
|
150 }
|
cannam@167
|
151 {
|
cannam@167
|
152 E T1O, T1P, T28, T29;
|
cannam@167
|
153 T1O = R1[WS(rs, 7)];
|
cannam@167
|
154 T1P = R1[WS(rs, 23)];
|
cannam@167
|
155 T1Q = T1O + T1P;
|
cannam@167
|
156 T4E = T1P - T1O;
|
cannam@167
|
157 T28 = R1[WS(rs, 27)];
|
cannam@167
|
158 T29 = R1[WS(rs, 11)];
|
cannam@167
|
159 T2a = T28 + T29;
|
cannam@167
|
160 T4t = T28 - T29;
|
cannam@167
|
161 }
|
cannam@167
|
162 T1R = T1N - T1Q;
|
cannam@167
|
163 T2e = T2a - T2d;
|
cannam@167
|
164 {
|
cannam@167
|
165 E T4u, T4F, T2Z, T30;
|
cannam@167
|
166 T4u = T4s + T4t;
|
cannam@167
|
167 T4v = FMA(KP707106781, T4u, T4r);
|
cannam@167
|
168 T5H = FNMS(KP707106781, T4u, T4r);
|
cannam@167
|
169 T4F = T4t - T4s;
|
cannam@167
|
170 T4G = FMA(KP707106781, T4F, T4E);
|
cannam@167
|
171 T5K = FNMS(KP707106781, T4F, T4E);
|
cannam@167
|
172 T2Z = T1N + T1Q;
|
cannam@167
|
173 T30 = T2d + T2a;
|
cannam@167
|
174 T31 = T2Z + T30;
|
cannam@167
|
175 T3l = T2Z - T30;
|
cannam@167
|
176 }
|
cannam@167
|
177 }
|
cannam@167
|
178 {
|
cannam@167
|
179 E TN, T3V, TX, T3X, TQ, T40, TU, T3W, T3Y, T41;
|
cannam@167
|
180 {
|
cannam@167
|
181 E TL, TM, TV, TW;
|
cannam@167
|
182 TL = R0[WS(rs, 31)];
|
cannam@167
|
183 TM = R0[WS(rs, 15)];
|
cannam@167
|
184 TN = TL + TM;
|
cannam@167
|
185 T3V = TL - TM;
|
cannam@167
|
186 TV = R0[WS(rs, 27)];
|
cannam@167
|
187 TW = R0[WS(rs, 11)];
|
cannam@167
|
188 TX = TV + TW;
|
cannam@167
|
189 T3X = TV - TW;
|
cannam@167
|
190 }
|
cannam@167
|
191 {
|
cannam@167
|
192 E TO, TP, TS, TT;
|
cannam@167
|
193 TO = R0[WS(rs, 7)];
|
cannam@167
|
194 TP = R0[WS(rs, 23)];
|
cannam@167
|
195 TQ = TO + TP;
|
cannam@167
|
196 T40 = TO - TP;
|
cannam@167
|
197 TS = R0[WS(rs, 3)];
|
cannam@167
|
198 TT = R0[WS(rs, 19)];
|
cannam@167
|
199 TU = TS + TT;
|
cannam@167
|
200 T3W = TS - TT;
|
cannam@167
|
201 }
|
cannam@167
|
202 T3Y = T3W + T3X;
|
cannam@167
|
203 T3Z = FMA(KP707106781, T3Y, T3V);
|
cannam@167
|
204 T5t = FNMS(KP707106781, T3Y, T3V);
|
cannam@167
|
205 T41 = T3W - T3X;
|
cannam@167
|
206 T42 = FMA(KP707106781, T41, T40);
|
cannam@167
|
207 T5s = FNMS(KP707106781, T41, T40);
|
cannam@167
|
208 {
|
cannam@167
|
209 E TR, TY, T19, T1a;
|
cannam@167
|
210 TR = TN + TQ;
|
cannam@167
|
211 TY = TU + TX;
|
cannam@167
|
212 TZ = TR + TY;
|
cannam@167
|
213 T3f = TR - TY;
|
cannam@167
|
214 T19 = TN - TQ;
|
cannam@167
|
215 T1a = TX - TU;
|
cannam@167
|
216 T1b = FNMS(KP414213562, T1a, T19);
|
cannam@167
|
217 T2n = FMA(KP414213562, T19, T1a);
|
cannam@167
|
218 }
|
cannam@167
|
219 }
|
cannam@167
|
220 {
|
cannam@167
|
221 E Ty, T3M, TI, T3O, TB, T3R, TF, T3N, T3P, T3S;
|
cannam@167
|
222 {
|
cannam@167
|
223 E Tw, Tx, TG, TH;
|
cannam@167
|
224 Tw = R0[WS(rs, 1)];
|
cannam@167
|
225 Tx = R0[WS(rs, 17)];
|
cannam@167
|
226 Ty = Tw + Tx;
|
cannam@167
|
227 T3M = Tw - Tx;
|
cannam@167
|
228 TG = R0[WS(rs, 29)];
|
cannam@167
|
229 TH = R0[WS(rs, 13)];
|
cannam@167
|
230 TI = TG + TH;
|
cannam@167
|
231 T3O = TG - TH;
|
cannam@167
|
232 }
|
cannam@167
|
233 {
|
cannam@167
|
234 E Tz, TA, TD, TE;
|
cannam@167
|
235 Tz = R0[WS(rs, 9)];
|
cannam@167
|
236 TA = R0[WS(rs, 25)];
|
cannam@167
|
237 TB = Tz + TA;
|
cannam@167
|
238 T3R = Tz - TA;
|
cannam@167
|
239 TD = R0[WS(rs, 5)];
|
cannam@167
|
240 TE = R0[WS(rs, 21)];
|
cannam@167
|
241 TF = TD + TE;
|
cannam@167
|
242 T3N = TD - TE;
|
cannam@167
|
243 }
|
cannam@167
|
244 T3P = T3N + T3O;
|
cannam@167
|
245 T3Q = FMA(KP707106781, T3P, T3M);
|
cannam@167
|
246 T5w = FNMS(KP707106781, T3P, T3M);
|
cannam@167
|
247 T3S = T3N - T3O;
|
cannam@167
|
248 T3T = FMA(KP707106781, T3S, T3R);
|
cannam@167
|
249 T5v = FNMS(KP707106781, T3S, T3R);
|
cannam@167
|
250 {
|
cannam@167
|
251 E TC, TJ, T16, T17;
|
cannam@167
|
252 TC = Ty + TB;
|
cannam@167
|
253 TJ = TF + TI;
|
cannam@167
|
254 TK = TC + TJ;
|
cannam@167
|
255 T3e = TC - TJ;
|
cannam@167
|
256 T16 = Ty - TB;
|
cannam@167
|
257 T17 = TI - TF;
|
cannam@167
|
258 T18 = FMA(KP414213562, T17, T16);
|
cannam@167
|
259 T2m = FNMS(KP414213562, T16, T17);
|
cannam@167
|
260 }
|
cannam@167
|
261 }
|
cannam@167
|
262 {
|
cannam@167
|
263 E Ti, T3E, Ts, T3I, Tl, T3F, Tp, T3H, T4Q, T4R;
|
cannam@167
|
264 {
|
cannam@167
|
265 E Tg, Th, Tq, Tr;
|
cannam@167
|
266 Tg = R0[WS(rs, 2)];
|
cannam@167
|
267 Th = R0[WS(rs, 18)];
|
cannam@167
|
268 Ti = Tg + Th;
|
cannam@167
|
269 T3E = Tg - Th;
|
cannam@167
|
270 Tq = R0[WS(rs, 6)];
|
cannam@167
|
271 Tr = R0[WS(rs, 22)];
|
cannam@167
|
272 Ts = Tq + Tr;
|
cannam@167
|
273 T3I = Tq - Tr;
|
cannam@167
|
274 }
|
cannam@167
|
275 {
|
cannam@167
|
276 E Tj, Tk, Tn, To;
|
cannam@167
|
277 Tj = R0[WS(rs, 10)];
|
cannam@167
|
278 Tk = R0[WS(rs, 26)];
|
cannam@167
|
279 Tl = Tj + Tk;
|
cannam@167
|
280 T3F = Tj - Tk;
|
cannam@167
|
281 Tn = R0[WS(rs, 30)];
|
cannam@167
|
282 To = R0[WS(rs, 14)];
|
cannam@167
|
283 Tp = Tn + To;
|
cannam@167
|
284 T3H = Tn - To;
|
cannam@167
|
285 }
|
cannam@167
|
286 Tm = Ti + Tl;
|
cannam@167
|
287 Tt = Tp + Ts;
|
cannam@167
|
288 Tu = Tm + Tt;
|
cannam@167
|
289 T4Q = FMA(KP414213562, T3E, T3F);
|
cannam@167
|
290 T4R = FNMS(KP414213562, T3H, T3I);
|
cannam@167
|
291 T4S = T4Q + T4R;
|
cannam@167
|
292 T5q = T4Q - T4R;
|
cannam@167
|
293 {
|
cannam@167
|
294 E T12, T13, T3G, T3J;
|
cannam@167
|
295 T12 = Ti - Tl;
|
cannam@167
|
296 T13 = Tp - Ts;
|
cannam@167
|
297 T14 = T12 + T13;
|
cannam@167
|
298 T2k = T13 - T12;
|
cannam@167
|
299 T3G = FNMS(KP414213562, T3F, T3E);
|
cannam@167
|
300 T3J = FMA(KP414213562, T3I, T3H);
|
cannam@167
|
301 T3K = T3G + T3J;
|
cannam@167
|
302 T5Q = T3J - T3G;
|
cannam@167
|
303 }
|
cannam@167
|
304 }
|
cannam@167
|
305 {
|
cannam@167
|
306 E T1n, T4b, T1x, T4f, T1q, T4c, T1u, T4e;
|
cannam@167
|
307 {
|
cannam@167
|
308 E T1l, T1m, T1v, T1w;
|
cannam@167
|
309 T1l = R1[WS(rs, 2)];
|
cannam@167
|
310 T1m = R1[WS(rs, 18)];
|
cannam@167
|
311 T1n = T1l + T1m;
|
cannam@167
|
312 T4b = T1l - T1m;
|
cannam@167
|
313 T1v = R1[WS(rs, 6)];
|
cannam@167
|
314 T1w = R1[WS(rs, 22)];
|
cannam@167
|
315 T1x = T1v + T1w;
|
cannam@167
|
316 T4f = T1v - T1w;
|
cannam@167
|
317 }
|
cannam@167
|
318 {
|
cannam@167
|
319 E T1o, T1p, T1s, T1t;
|
cannam@167
|
320 T1o = R1[WS(rs, 10)];
|
cannam@167
|
321 T1p = R1[WS(rs, 26)];
|
cannam@167
|
322 T1q = T1o + T1p;
|
cannam@167
|
323 T4c = T1o - T1p;
|
cannam@167
|
324 T1s = R1[WS(rs, 30)];
|
cannam@167
|
325 T1t = R1[WS(rs, 14)];
|
cannam@167
|
326 T1u = T1s + T1t;
|
cannam@167
|
327 T4e = T1s - T1t;
|
cannam@167
|
328 }
|
cannam@167
|
329 {
|
cannam@167
|
330 E T1r, T1y, T4m, T4n;
|
cannam@167
|
331 T1r = T1n - T1q;
|
cannam@167
|
332 T1y = T1u - T1x;
|
cannam@167
|
333 T1z = T1r + T1y;
|
cannam@167
|
334 T1I = T1y - T1r;
|
cannam@167
|
335 T4m = FMA(KP414213562, T4b, T4c);
|
cannam@167
|
336 T4n = FNMS(KP414213562, T4e, T4f);
|
cannam@167
|
337 T4o = T4m + T4n;
|
cannam@167
|
338 T5B = T4m - T4n;
|
cannam@167
|
339 }
|
cannam@167
|
340 {
|
cannam@167
|
341 E T2V, T2W, T4d, T4g;
|
cannam@167
|
342 T2V = T1n + T1q;
|
cannam@167
|
343 T2W = T1u + T1x;
|
cannam@167
|
344 T2X = T2V + T2W;
|
cannam@167
|
345 T3j = T2W - T2V;
|
cannam@167
|
346 T4d = FNMS(KP414213562, T4c, T4b);
|
cannam@167
|
347 T4g = FMA(KP414213562, T4f, T4e);
|
cannam@167
|
348 T4h = T4d + T4g;
|
cannam@167
|
349 T5E = T4g - T4d;
|
cannam@167
|
350 }
|
cannam@167
|
351 }
|
cannam@167
|
352 {
|
cannam@167
|
353 E T1U, T4w, T24, T4A, T1X, T4x, T21, T4z;
|
cannam@167
|
354 {
|
cannam@167
|
355 E T1S, T1T, T22, T23;
|
cannam@167
|
356 T1S = R1[WS(rs, 1)];
|
cannam@167
|
357 T1T = R1[WS(rs, 17)];
|
cannam@167
|
358 T1U = T1S + T1T;
|
cannam@167
|
359 T4w = T1S - T1T;
|
cannam@167
|
360 T22 = R1[WS(rs, 5)];
|
cannam@167
|
361 T23 = R1[WS(rs, 21)];
|
cannam@167
|
362 T24 = T22 + T23;
|
cannam@167
|
363 T4A = T23 - T22;
|
cannam@167
|
364 }
|
cannam@167
|
365 {
|
cannam@167
|
366 E T1V, T1W, T1Z, T20;
|
cannam@167
|
367 T1V = R1[WS(rs, 9)];
|
cannam@167
|
368 T1W = R1[WS(rs, 25)];
|
cannam@167
|
369 T1X = T1V + T1W;
|
cannam@167
|
370 T4x = T1W - T1V;
|
cannam@167
|
371 T1Z = R1[WS(rs, 29)];
|
cannam@167
|
372 T20 = R1[WS(rs, 13)];
|
cannam@167
|
373 T21 = T1Z + T20;
|
cannam@167
|
374 T4z = T1Z - T20;
|
cannam@167
|
375 }
|
cannam@167
|
376 {
|
cannam@167
|
377 E T1Y, T25, T4H, T4I;
|
cannam@167
|
378 T1Y = T1U - T1X;
|
cannam@167
|
379 T25 = T21 - T24;
|
cannam@167
|
380 T26 = T1Y + T25;
|
cannam@167
|
381 T2f = T25 - T1Y;
|
cannam@167
|
382 T4H = FNMS(KP414213562, T4w, T4x);
|
cannam@167
|
383 T4I = FMA(KP414213562, T4z, T4A);
|
cannam@167
|
384 T4J = T4H + T4I;
|
cannam@167
|
385 T5I = T4I - T4H;
|
cannam@167
|
386 }
|
cannam@167
|
387 {
|
cannam@167
|
388 E T32, T33, T4y, T4B;
|
cannam@167
|
389 T32 = T1U + T1X;
|
cannam@167
|
390 T33 = T21 + T24;
|
cannam@167
|
391 T34 = T32 + T33;
|
cannam@167
|
392 T3m = T33 - T32;
|
cannam@167
|
393 T4y = FMA(KP414213562, T4x, T4w);
|
cannam@167
|
394 T4B = FNMS(KP414213562, T4A, T4z);
|
cannam@167
|
395 T4C = T4y + T4B;
|
cannam@167
|
396 T5L = T4B - T4y;
|
cannam@167
|
397 }
|
cannam@167
|
398 }
|
cannam@167
|
399 {
|
cannam@167
|
400 E Tv, T10, T39, T3a, T3b, T3c;
|
cannam@167
|
401 Tv = Tf + Tu;
|
cannam@167
|
402 T10 = TK + TZ;
|
cannam@167
|
403 T39 = Tv + T10;
|
cannam@167
|
404 T3a = T2U + T2X;
|
cannam@167
|
405 T3b = T31 + T34;
|
cannam@167
|
406 T3c = T3a + T3b;
|
cannam@167
|
407 Cr[WS(csr, 16)] = Tv - T10;
|
cannam@167
|
408 Ci[WS(csi, 16)] = T3b - T3a;
|
cannam@167
|
409 Cr[WS(csr, 32)] = T39 - T3c;
|
cannam@167
|
410 Cr[0] = T39 + T3c;
|
cannam@167
|
411 }
|
cannam@167
|
412 {
|
cannam@167
|
413 E T2R, T37, T36, T38, T2Y, T35;
|
cannam@167
|
414 T2R = Tf - Tu;
|
cannam@167
|
415 T37 = TZ - TK;
|
cannam@167
|
416 T2Y = T2U - T2X;
|
cannam@167
|
417 T35 = T31 - T34;
|
cannam@167
|
418 T36 = T2Y + T35;
|
cannam@167
|
419 T38 = T35 - T2Y;
|
cannam@167
|
420 Cr[WS(csr, 24)] = FNMS(KP707106781, T36, T2R);
|
cannam@167
|
421 Ci[WS(csi, 24)] = FMS(KP707106781, T38, T37);
|
cannam@167
|
422 Cr[WS(csr, 8)] = FMA(KP707106781, T36, T2R);
|
cannam@167
|
423 Ci[WS(csi, 8)] = FMA(KP707106781, T38, T37);
|
cannam@167
|
424 }
|
cannam@167
|
425 {
|
cannam@167
|
426 E T3h, T3x, T3w, T3y, T3o, T3s, T3r, T3t;
|
cannam@167
|
427 {
|
cannam@167
|
428 E T3d, T3g, T3u, T3v;
|
cannam@167
|
429 T3d = T7 - Te;
|
cannam@167
|
430 T3g = T3e + T3f;
|
cannam@167
|
431 T3h = FMA(KP707106781, T3g, T3d);
|
cannam@167
|
432 T3x = FNMS(KP707106781, T3g, T3d);
|
cannam@167
|
433 T3u = FNMS(KP414213562, T3i, T3j);
|
cannam@167
|
434 T3v = FMA(KP414213562, T3l, T3m);
|
cannam@167
|
435 T3w = T3u + T3v;
|
cannam@167
|
436 T3y = T3v - T3u;
|
cannam@167
|
437 }
|
cannam@167
|
438 {
|
cannam@167
|
439 E T3k, T3n, T3p, T3q;
|
cannam@167
|
440 T3k = FMA(KP414213562, T3j, T3i);
|
cannam@167
|
441 T3n = FNMS(KP414213562, T3m, T3l);
|
cannam@167
|
442 T3o = T3k + T3n;
|
cannam@167
|
443 T3s = T3n - T3k;
|
cannam@167
|
444 T3p = Tt - Tm;
|
cannam@167
|
445 T3q = T3f - T3e;
|
cannam@167
|
446 T3r = FNMS(KP707106781, T3q, T3p);
|
cannam@167
|
447 T3t = FMA(KP707106781, T3q, T3p);
|
cannam@167
|
448 }
|
cannam@167
|
449 Cr[WS(csr, 28)] = FNMS(KP923879532, T3o, T3h);
|
cannam@167
|
450 Ci[WS(csi, 28)] = FMS(KP923879532, T3w, T3t);
|
cannam@167
|
451 Cr[WS(csr, 4)] = FMA(KP923879532, T3o, T3h);
|
cannam@167
|
452 Ci[WS(csi, 4)] = FMA(KP923879532, T3w, T3t);
|
cannam@167
|
453 Ci[WS(csi, 12)] = FMS(KP923879532, T3s, T3r);
|
cannam@167
|
454 Cr[WS(csr, 12)] = FMA(KP923879532, T3y, T3x);
|
cannam@167
|
455 Ci[WS(csi, 20)] = FMA(KP923879532, T3s, T3r);
|
cannam@167
|
456 Cr[WS(csr, 20)] = FNMS(KP923879532, T3y, T3x);
|
cannam@167
|
457 }
|
cannam@167
|
458 {
|
cannam@167
|
459 E T2z, T2P, T2J, T2L, T2C, T2M, T2F, T2N;
|
cannam@167
|
460 {
|
cannam@167
|
461 E T2x, T2y, T2H, T2I;
|
cannam@167
|
462 T2x = FNMS(KP707106781, T14, T11);
|
cannam@167
|
463 T2y = T2n - T2m;
|
cannam@167
|
464 T2z = FMA(KP923879532, T2y, T2x);
|
cannam@167
|
465 T2P = FNMS(KP923879532, T2y, T2x);
|
cannam@167
|
466 T2H = FNMS(KP707106781, T2k, T2j);
|
cannam@167
|
467 T2I = T1b - T18;
|
cannam@167
|
468 T2J = FMA(KP923879532, T2I, T2H);
|
cannam@167
|
469 T2L = FNMS(KP923879532, T2I, T2H);
|
cannam@167
|
470 }
|
cannam@167
|
471 {
|
cannam@167
|
472 E T2A, T2B, T2D, T2E;
|
cannam@167
|
473 T2A = FNMS(KP707106781, T1z, T1k);
|
cannam@167
|
474 T2B = FNMS(KP707106781, T1I, T1H);
|
cannam@167
|
475 T2C = FNMS(KP668178637, T2B, T2A);
|
cannam@167
|
476 T2M = FMA(KP668178637, T2A, T2B);
|
cannam@167
|
477 T2D = FNMS(KP707106781, T26, T1R);
|
cannam@167
|
478 T2E = FNMS(KP707106781, T2f, T2e);
|
cannam@167
|
479 T2F = FMA(KP668178637, T2E, T2D);
|
cannam@167
|
480 T2N = FNMS(KP668178637, T2D, T2E);
|
cannam@167
|
481 }
|
cannam@167
|
482 {
|
cannam@167
|
483 E T2G, T2O, T2K, T2Q;
|
cannam@167
|
484 T2G = T2C + T2F;
|
cannam@167
|
485 Cr[WS(csr, 26)] = FNMS(KP831469612, T2G, T2z);
|
cannam@167
|
486 Cr[WS(csr, 6)] = FMA(KP831469612, T2G, T2z);
|
cannam@167
|
487 T2O = T2M + T2N;
|
cannam@167
|
488 Ci[WS(csi, 6)] = -(FMA(KP831469612, T2O, T2L));
|
cannam@167
|
489 Ci[WS(csi, 26)] = FNMS(KP831469612, T2O, T2L);
|
cannam@167
|
490 T2K = T2F - T2C;
|
cannam@167
|
491 Ci[WS(csi, 10)] = FMA(KP831469612, T2K, T2J);
|
cannam@167
|
492 Ci[WS(csi, 22)] = FMS(KP831469612, T2K, T2J);
|
cannam@167
|
493 T2Q = T2M - T2N;
|
cannam@167
|
494 Cr[WS(csr, 22)] = FNMS(KP831469612, T2Q, T2P);
|
cannam@167
|
495 Cr[WS(csr, 10)] = FMA(KP831469612, T2Q, T2P);
|
cannam@167
|
496 }
|
cannam@167
|
497 }
|
cannam@167
|
498 {
|
cannam@167
|
499 E T1d, T2v, T2p, T2r, T1K, T2s, T2h, T2t;
|
cannam@167
|
500 {
|
cannam@167
|
501 E T15, T1c, T2l, T2o;
|
cannam@167
|
502 T15 = FMA(KP707106781, T14, T11);
|
cannam@167
|
503 T1c = T18 + T1b;
|
cannam@167
|
504 T1d = FMA(KP923879532, T1c, T15);
|
cannam@167
|
505 T2v = FNMS(KP923879532, T1c, T15);
|
cannam@167
|
506 T2l = FMA(KP707106781, T2k, T2j);
|
cannam@167
|
507 T2o = T2m + T2n;
|
cannam@167
|
508 T2p = FNMS(KP923879532, T2o, T2l);
|
cannam@167
|
509 T2r = FMA(KP923879532, T2o, T2l);
|
cannam@167
|
510 }
|
cannam@167
|
511 {
|
cannam@167
|
512 E T1A, T1J, T27, T2g;
|
cannam@167
|
513 T1A = FMA(KP707106781, T1z, T1k);
|
cannam@167
|
514 T1J = FMA(KP707106781, T1I, T1H);
|
cannam@167
|
515 T1K = FMA(KP198912367, T1J, T1A);
|
cannam@167
|
516 T2s = FNMS(KP198912367, T1A, T1J);
|
cannam@167
|
517 T27 = FMA(KP707106781, T26, T1R);
|
cannam@167
|
518 T2g = FMA(KP707106781, T2f, T2e);
|
cannam@167
|
519 T2h = FNMS(KP198912367, T2g, T27);
|
cannam@167
|
520 T2t = FMA(KP198912367, T27, T2g);
|
cannam@167
|
521 }
|
cannam@167
|
522 {
|
cannam@167
|
523 E T2i, T2u, T2q, T2w;
|
cannam@167
|
524 T2i = T1K + T2h;
|
cannam@167
|
525 Cr[WS(csr, 30)] = FNMS(KP980785280, T2i, T1d);
|
cannam@167
|
526 Cr[WS(csr, 2)] = FMA(KP980785280, T2i, T1d);
|
cannam@167
|
527 T2u = T2s + T2t;
|
cannam@167
|
528 Ci[WS(csi, 2)] = FMA(KP980785280, T2u, T2r);
|
cannam@167
|
529 Ci[WS(csi, 30)] = FMS(KP980785280, T2u, T2r);
|
cannam@167
|
530 T2q = T2h - T1K;
|
cannam@167
|
531 Ci[WS(csi, 14)] = FMS(KP980785280, T2q, T2p);
|
cannam@167
|
532 Ci[WS(csi, 18)] = FMA(KP980785280, T2q, T2p);
|
cannam@167
|
533 T2w = T2t - T2s;
|
cannam@167
|
534 Cr[WS(csr, 18)] = FNMS(KP980785280, T2w, T2v);
|
cannam@167
|
535 Cr[WS(csr, 14)] = FMA(KP980785280, T2w, T2v);
|
cannam@167
|
536 }
|
cannam@167
|
537 }
|
cannam@167
|
538 {
|
cannam@167
|
539 E T5r, T63, T6d, T5R, T5y, T6e, T6b, T6j, T5U, T64, T5G, T5Z, T68, T6i, T5N;
|
cannam@167
|
540 E T5Y;
|
cannam@167
|
541 {
|
cannam@167
|
542 E T5u, T5x, T5C, T5F;
|
cannam@167
|
543 T5r = FNMS(KP923879532, T5q, T5p);
|
cannam@167
|
544 T63 = FMA(KP923879532, T5q, T5p);
|
cannam@167
|
545 T6d = FMA(KP923879532, T5Q, T5P);
|
cannam@167
|
546 T5R = FNMS(KP923879532, T5Q, T5P);
|
cannam@167
|
547 T5u = FMA(KP668178637, T5t, T5s);
|
cannam@167
|
548 T5x = FNMS(KP668178637, T5w, T5v);
|
cannam@167
|
549 T5y = T5u - T5x;
|
cannam@167
|
550 T6e = T5x + T5u;
|
cannam@167
|
551 {
|
cannam@167
|
552 E T69, T6a, T5S, T5T;
|
cannam@167
|
553 T69 = FMA(KP923879532, T5I, T5H);
|
cannam@167
|
554 T6a = FNMS(KP923879532, T5L, T5K);
|
cannam@167
|
555 T6b = FMA(KP303346683, T6a, T69);
|
cannam@167
|
556 T6j = FNMS(KP303346683, T69, T6a);
|
cannam@167
|
557 T5S = FNMS(KP668178637, T5s, T5t);
|
cannam@167
|
558 T5T = FMA(KP668178637, T5v, T5w);
|
cannam@167
|
559 T5U = T5S - T5T;
|
cannam@167
|
560 T64 = T5T + T5S;
|
cannam@167
|
561 }
|
cannam@167
|
562 T5C = FNMS(KP923879532, T5B, T5A);
|
cannam@167
|
563 T5F = FNMS(KP923879532, T5E, T5D);
|
cannam@167
|
564 T5G = FNMS(KP534511135, T5F, T5C);
|
cannam@167
|
565 T5Z = FMA(KP534511135, T5C, T5F);
|
cannam@167
|
566 {
|
cannam@167
|
567 E T66, T67, T5J, T5M;
|
cannam@167
|
568 T66 = FMA(KP923879532, T5B, T5A);
|
cannam@167
|
569 T67 = FMA(KP923879532, T5E, T5D);
|
cannam@167
|
570 T68 = FMA(KP303346683, T67, T66);
|
cannam@167
|
571 T6i = FNMS(KP303346683, T66, T67);
|
cannam@167
|
572 T5J = FNMS(KP923879532, T5I, T5H);
|
cannam@167
|
573 T5M = FMA(KP923879532, T5L, T5K);
|
cannam@167
|
574 T5N = FNMS(KP534511135, T5M, T5J);
|
cannam@167
|
575 T5Y = FMA(KP534511135, T5J, T5M);
|
cannam@167
|
576 }
|
cannam@167
|
577 }
|
cannam@167
|
578 {
|
cannam@167
|
579 E T5z, T5O, T5X, T60;
|
cannam@167
|
580 T5z = FMA(KP831469612, T5y, T5r);
|
cannam@167
|
581 T5O = T5G + T5N;
|
cannam@167
|
582 Cr[WS(csr, 27)] = FNMS(KP881921264, T5O, T5z);
|
cannam@167
|
583 Cr[WS(csr, 5)] = FMA(KP881921264, T5O, T5z);
|
cannam@167
|
584 T5X = FNMS(KP831469612, T5U, T5R);
|
cannam@167
|
585 T60 = T5Y - T5Z;
|
cannam@167
|
586 Ci[WS(csi, 5)] = FMS(KP881921264, T60, T5X);
|
cannam@167
|
587 Ci[WS(csi, 27)] = FMA(KP881921264, T60, T5X);
|
cannam@167
|
588 }
|
cannam@167
|
589 {
|
cannam@167
|
590 E T5V, T5W, T61, T62;
|
cannam@167
|
591 T5V = FMA(KP831469612, T5U, T5R);
|
cannam@167
|
592 T5W = T5N - T5G;
|
cannam@167
|
593 Ci[WS(csi, 11)] = FMA(KP881921264, T5W, T5V);
|
cannam@167
|
594 Ci[WS(csi, 21)] = FMS(KP881921264, T5W, T5V);
|
cannam@167
|
595 T61 = FNMS(KP831469612, T5y, T5r);
|
cannam@167
|
596 T62 = T5Z + T5Y;
|
cannam@167
|
597 Cr[WS(csr, 21)] = FNMS(KP881921264, T62, T61);
|
cannam@167
|
598 Cr[WS(csr, 11)] = FMA(KP881921264, T62, T61);
|
cannam@167
|
599 }
|
cannam@167
|
600 {
|
cannam@167
|
601 E T65, T6c, T6h, T6k;
|
cannam@167
|
602 T65 = FMA(KP831469612, T64, T63);
|
cannam@167
|
603 T6c = T68 + T6b;
|
cannam@167
|
604 Cr[WS(csr, 29)] = FNMS(KP956940335, T6c, T65);
|
cannam@167
|
605 Cr[WS(csr, 3)] = FMA(KP956940335, T6c, T65);
|
cannam@167
|
606 T6h = FMA(KP831469612, T6e, T6d);
|
cannam@167
|
607 T6k = T6i - T6j;
|
cannam@167
|
608 Ci[WS(csi, 3)] = FMA(KP956940335, T6k, T6h);
|
cannam@167
|
609 Ci[WS(csi, 29)] = FMS(KP956940335, T6k, T6h);
|
cannam@167
|
610 }
|
cannam@167
|
611 {
|
cannam@167
|
612 E T6f, T6g, T6l, T6m;
|
cannam@167
|
613 T6f = FNMS(KP831469612, T6e, T6d);
|
cannam@167
|
614 T6g = T6b - T68;
|
cannam@167
|
615 Ci[WS(csi, 13)] = FMS(KP956940335, T6g, T6f);
|
cannam@167
|
616 Ci[WS(csi, 19)] = FMA(KP956940335, T6g, T6f);
|
cannam@167
|
617 T6l = FNMS(KP831469612, T64, T63);
|
cannam@167
|
618 T6m = T6i + T6j;
|
cannam@167
|
619 Cr[WS(csr, 19)] = FMA(KP956940335, T6m, T6l);
|
cannam@167
|
620 Cr[WS(csr, 13)] = FNMS(KP956940335, T6m, T6l);
|
cannam@167
|
621 }
|
cannam@167
|
622 }
|
cannam@167
|
623 {
|
cannam@167
|
624 E T3L, T55, T5f, T4T, T44, T5g, T5d, T5l, T4W, T56, T4q, T51, T5a, T5k, T4L;
|
cannam@167
|
625 E T50;
|
cannam@167
|
626 {
|
cannam@167
|
627 E T3U, T43, T4i, T4p;
|
cannam@167
|
628 T3L = FMA(KP923879532, T3K, T3D);
|
cannam@167
|
629 T55 = FNMS(KP923879532, T3K, T3D);
|
cannam@167
|
630 T5f = FNMS(KP923879532, T4S, T4P);
|
cannam@167
|
631 T4T = FMA(KP923879532, T4S, T4P);
|
cannam@167
|
632 T3U = FNMS(KP198912367, T3T, T3Q);
|
cannam@167
|
633 T43 = FMA(KP198912367, T42, T3Z);
|
cannam@167
|
634 T44 = T3U + T43;
|
cannam@167
|
635 T5g = T43 - T3U;
|
cannam@167
|
636 {
|
cannam@167
|
637 E T5b, T5c, T4U, T4V;
|
cannam@167
|
638 T5b = FNMS(KP923879532, T4C, T4v);
|
cannam@167
|
639 T5c = FNMS(KP923879532, T4J, T4G);
|
cannam@167
|
640 T5d = FMA(KP820678790, T5c, T5b);
|
cannam@167
|
641 T5l = FNMS(KP820678790, T5b, T5c);
|
cannam@167
|
642 T4U = FMA(KP198912367, T3Q, T3T);
|
cannam@167
|
643 T4V = FNMS(KP198912367, T3Z, T42);
|
cannam@167
|
644 T4W = T4U + T4V;
|
cannam@167
|
645 T56 = T4U - T4V;
|
cannam@167
|
646 }
|
cannam@167
|
647 T4i = FMA(KP923879532, T4h, T4a);
|
cannam@167
|
648 T4p = FMA(KP923879532, T4o, T4l);
|
cannam@167
|
649 T4q = FNMS(KP098491403, T4p, T4i);
|
cannam@167
|
650 T51 = FMA(KP098491403, T4i, T4p);
|
cannam@167
|
651 {
|
cannam@167
|
652 E T58, T59, T4D, T4K;
|
cannam@167
|
653 T58 = FNMS(KP923879532, T4h, T4a);
|
cannam@167
|
654 T59 = FNMS(KP923879532, T4o, T4l);
|
cannam@167
|
655 T5a = FMA(KP820678790, T59, T58);
|
cannam@167
|
656 T5k = FNMS(KP820678790, T58, T59);
|
cannam@167
|
657 T4D = FMA(KP923879532, T4C, T4v);
|
cannam@167
|
658 T4K = FMA(KP923879532, T4J, T4G);
|
cannam@167
|
659 T4L = FNMS(KP098491403, T4K, T4D);
|
cannam@167
|
660 T50 = FMA(KP098491403, T4D, T4K);
|
cannam@167
|
661 }
|
cannam@167
|
662 }
|
cannam@167
|
663 {
|
cannam@167
|
664 E T45, T4M, T4Z, T52;
|
cannam@167
|
665 T45 = FMA(KP980785280, T44, T3L);
|
cannam@167
|
666 T4M = T4q + T4L;
|
cannam@167
|
667 Cr[WS(csr, 31)] = FNMS(KP995184726, T4M, T45);
|
cannam@167
|
668 Cr[WS(csr, 1)] = FMA(KP995184726, T4M, T45);
|
cannam@167
|
669 T4Z = FMA(KP980785280, T4W, T4T);
|
cannam@167
|
670 T52 = T50 - T51;
|
cannam@167
|
671 Ci[WS(csi, 1)] = FMS(KP995184726, T52, T4Z);
|
cannam@167
|
672 Ci[WS(csi, 31)] = FMA(KP995184726, T52, T4Z);
|
cannam@167
|
673 }
|
cannam@167
|
674 {
|
cannam@167
|
675 E T4X, T4Y, T53, T54;
|
cannam@167
|
676 T4X = FNMS(KP980785280, T4W, T4T);
|
cannam@167
|
677 T4Y = T4L - T4q;
|
cannam@167
|
678 Ci[WS(csi, 15)] = FMA(KP995184726, T4Y, T4X);
|
cannam@167
|
679 Ci[WS(csi, 17)] = FMS(KP995184726, T4Y, T4X);
|
cannam@167
|
680 T53 = FNMS(KP980785280, T44, T3L);
|
cannam@167
|
681 T54 = T51 + T50;
|
cannam@167
|
682 Cr[WS(csr, 17)] = FNMS(KP995184726, T54, T53);
|
cannam@167
|
683 Cr[WS(csr, 15)] = FMA(KP995184726, T54, T53);
|
cannam@167
|
684 }
|
cannam@167
|
685 {
|
cannam@167
|
686 E T57, T5e, T5j, T5m;
|
cannam@167
|
687 T57 = FMA(KP980785280, T56, T55);
|
cannam@167
|
688 T5e = T5a + T5d;
|
cannam@167
|
689 Cr[WS(csr, 25)] = FNMS(KP773010453, T5e, T57);
|
cannam@167
|
690 Cr[WS(csr, 7)] = FMA(KP773010453, T5e, T57);
|
cannam@167
|
691 T5j = FMA(KP980785280, T5g, T5f);
|
cannam@167
|
692 T5m = T5k - T5l;
|
cannam@167
|
693 Ci[WS(csi, 7)] = FMA(KP773010453, T5m, T5j);
|
cannam@167
|
694 Ci[WS(csi, 25)] = FMS(KP773010453, T5m, T5j);
|
cannam@167
|
695 }
|
cannam@167
|
696 {
|
cannam@167
|
697 E T5h, T5i, T5n, T5o;
|
cannam@167
|
698 T5h = FNMS(KP980785280, T5g, T5f);
|
cannam@167
|
699 T5i = T5d - T5a;
|
cannam@167
|
700 Ci[WS(csi, 9)] = FMS(KP773010453, T5i, T5h);
|
cannam@167
|
701 Ci[WS(csi, 23)] = FMA(KP773010453, T5i, T5h);
|
cannam@167
|
702 T5n = FNMS(KP980785280, T56, T55);
|
cannam@167
|
703 T5o = T5k + T5l;
|
cannam@167
|
704 Cr[WS(csr, 23)] = FMA(KP773010453, T5o, T5n);
|
cannam@167
|
705 Cr[WS(csr, 9)] = FNMS(KP773010453, T5o, T5n);
|
cannam@167
|
706 }
|
cannam@167
|
707 }
|
cannam@167
|
708 }
|
cannam@167
|
709 }
|
cannam@167
|
710 }
|
cannam@167
|
711
|
cannam@167
|
712 static const kr2c_desc desc = { 64, "r2cf_64", {198, 0, 196, 0}, &GENUS };
|
cannam@167
|
713
|
cannam@167
|
714 void X(codelet_r2cf_64) (planner *p) {
|
cannam@167
|
715 X(kr2c_register) (p, r2cf_64, &desc);
|
cannam@167
|
716 }
|
cannam@167
|
717
|
cannam@167
|
718 #else
|
cannam@167
|
719
|
cannam@167
|
720 /* Generated by: ../../../genfft/gen_r2cf.native -compact -variables 4 -pipeline-latency 4 -n 64 -name r2cf_64 -include rdft/scalar/r2cf.h */
|
cannam@167
|
721
|
cannam@167
|
722 /*
|
cannam@167
|
723 * This function contains 394 FP additions, 124 FP multiplications,
|
cannam@167
|
724 * (or, 342 additions, 72 multiplications, 52 fused multiply/add),
|
cannam@167
|
725 * 106 stack variables, 15 constants, and 128 memory accesses
|
cannam@167
|
726 */
|
cannam@167
|
727 #include "rdft/scalar/r2cf.h"
|
cannam@167
|
728
|
cannam@167
|
729 static void r2cf_64(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
|
cannam@167
|
730 {
|
cannam@167
|
731 DK(KP773010453, +0.773010453362736960810906609758469800971041293);
|
cannam@167
|
732 DK(KP634393284, +0.634393284163645498215171613225493370675687095);
|
cannam@167
|
733 DK(KP098017140, +0.098017140329560601994195563888641845861136673);
|
cannam@167
|
734 DK(KP995184726, +0.995184726672196886244836953109479921575474869);
|
cannam@167
|
735 DK(KP290284677, +0.290284677254462367636192375817395274691476278);
|
cannam@167
|
736 DK(KP956940335, +0.956940335732208864935797886980269969482849206);
|
cannam@167
|
737 DK(KP471396736, +0.471396736825997648556387625905254377657460319);
|
cannam@167
|
738 DK(KP881921264, +0.881921264348355029712756863660388349508442621);
|
cannam@167
|
739 DK(KP195090322, +0.195090322016128267848284868477022240927691618);
|
cannam@167
|
740 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
cannam@167
|
741 DK(KP555570233, +0.555570233019602224742830813948532874374937191);
|
cannam@167
|
742 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
cannam@167
|
743 DK(KP382683432, +0.382683432365089771728459984030398866761344562);
|
cannam@167
|
744 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
cannam@167
|
745 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
cannam@167
|
746 {
|
cannam@167
|
747 INT i;
|
cannam@167
|
748 for (i = v; i > 0; i = i - 1, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(256, rs), MAKE_VOLATILE_STRIDE(256, csr), MAKE_VOLATILE_STRIDE(256, csi)) {
|
cannam@167
|
749 E T4l, T5a, T15, T3n, T2T, T3Q, T7, Te, Tf, T4A, T4L, T1X, T3B, T23, T3y;
|
cannam@167
|
750 E T5I, T66, T4R, T52, T2j, T3F, T2H, T3I, T5P, T69, T1i, T3t, T1l, T3u, TZ;
|
cannam@167
|
751 E T63, T4v, T58, T1r, T3r, T1u, T3q, TK, T62, T4s, T57, Tm, Tt, Tu, T4o;
|
cannam@167
|
752 E T5b, T1c, T3R, T2Q, T3o, T1M, T3z, T5L, T67, T26, T3C, T4H, T4M, T2y, T3J;
|
cannam@167
|
753 E T5S, T6a, T2C, T3G, T4Y, T53;
|
cannam@167
|
754 {
|
cannam@167
|
755 E T3, T11, Td, T13, T6, T2S, Ta, T12, T14, T2R;
|
cannam@167
|
756 {
|
cannam@167
|
757 E T1, T2, Tb, Tc;
|
cannam@167
|
758 T1 = R0[0];
|
cannam@167
|
759 T2 = R0[WS(rs, 16)];
|
cannam@167
|
760 T3 = T1 + T2;
|
cannam@167
|
761 T11 = T1 - T2;
|
cannam@167
|
762 Tb = R0[WS(rs, 28)];
|
cannam@167
|
763 Tc = R0[WS(rs, 12)];
|
cannam@167
|
764 Td = Tb + Tc;
|
cannam@167
|
765 T13 = Tb - Tc;
|
cannam@167
|
766 }
|
cannam@167
|
767 {
|
cannam@167
|
768 E T4, T5, T8, T9;
|
cannam@167
|
769 T4 = R0[WS(rs, 8)];
|
cannam@167
|
770 T5 = R0[WS(rs, 24)];
|
cannam@167
|
771 T6 = T4 + T5;
|
cannam@167
|
772 T2S = T4 - T5;
|
cannam@167
|
773 T8 = R0[WS(rs, 4)];
|
cannam@167
|
774 T9 = R0[WS(rs, 20)];
|
cannam@167
|
775 Ta = T8 + T9;
|
cannam@167
|
776 T12 = T8 - T9;
|
cannam@167
|
777 }
|
cannam@167
|
778 T4l = T3 - T6;
|
cannam@167
|
779 T5a = Td - Ta;
|
cannam@167
|
780 T14 = KP707106781 * (T12 + T13);
|
cannam@167
|
781 T15 = T11 + T14;
|
cannam@167
|
782 T3n = T11 - T14;
|
cannam@167
|
783 T2R = KP707106781 * (T13 - T12);
|
cannam@167
|
784 T2T = T2R - T2S;
|
cannam@167
|
785 T3Q = T2S + T2R;
|
cannam@167
|
786 T7 = T3 + T6;
|
cannam@167
|
787 Te = Ta + Td;
|
cannam@167
|
788 Tf = T7 + Te;
|
cannam@167
|
789 }
|
cannam@167
|
790 {
|
cannam@167
|
791 E T1P, T4J, T21, T4y, T1S, T4K, T1W, T4z;
|
cannam@167
|
792 {
|
cannam@167
|
793 E T1N, T1O, T1Z, T20;
|
cannam@167
|
794 T1N = R1[WS(rs, 28)];
|
cannam@167
|
795 T1O = R1[WS(rs, 12)];
|
cannam@167
|
796 T1P = T1N - T1O;
|
cannam@167
|
797 T4J = T1N + T1O;
|
cannam@167
|
798 T1Z = R1[0];
|
cannam@167
|
799 T20 = R1[WS(rs, 16)];
|
cannam@167
|
800 T21 = T1Z - T20;
|
cannam@167
|
801 T4y = T1Z + T20;
|
cannam@167
|
802 }
|
cannam@167
|
803 {
|
cannam@167
|
804 E T1Q, T1R, T1U, T1V;
|
cannam@167
|
805 T1Q = R1[WS(rs, 4)];
|
cannam@167
|
806 T1R = R1[WS(rs, 20)];
|
cannam@167
|
807 T1S = T1Q - T1R;
|
cannam@167
|
808 T4K = T1Q + T1R;
|
cannam@167
|
809 T1U = R1[WS(rs, 8)];
|
cannam@167
|
810 T1V = R1[WS(rs, 24)];
|
cannam@167
|
811 T1W = T1U - T1V;
|
cannam@167
|
812 T4z = T1U + T1V;
|
cannam@167
|
813 }
|
cannam@167
|
814 T4A = T4y - T4z;
|
cannam@167
|
815 T4L = T4J - T4K;
|
cannam@167
|
816 {
|
cannam@167
|
817 E T1T, T22, T5G, T5H;
|
cannam@167
|
818 T1T = KP707106781 * (T1P - T1S);
|
cannam@167
|
819 T1X = T1T - T1W;
|
cannam@167
|
820 T3B = T1W + T1T;
|
cannam@167
|
821 T22 = KP707106781 * (T1S + T1P);
|
cannam@167
|
822 T23 = T21 + T22;
|
cannam@167
|
823 T3y = T21 - T22;
|
cannam@167
|
824 T5G = T4y + T4z;
|
cannam@167
|
825 T5H = T4K + T4J;
|
cannam@167
|
826 T5I = T5G + T5H;
|
cannam@167
|
827 T66 = T5G - T5H;
|
cannam@167
|
828 }
|
cannam@167
|
829 }
|
cannam@167
|
830 {
|
cannam@167
|
831 E T2b, T4P, T2G, T4Q, T2e, T51, T2h, T50;
|
cannam@167
|
832 {
|
cannam@167
|
833 E T29, T2a, T2E, T2F;
|
cannam@167
|
834 T29 = R1[WS(rs, 31)];
|
cannam@167
|
835 T2a = R1[WS(rs, 15)];
|
cannam@167
|
836 T2b = T29 - T2a;
|
cannam@167
|
837 T4P = T29 + T2a;
|
cannam@167
|
838 T2E = R1[WS(rs, 7)];
|
cannam@167
|
839 T2F = R1[WS(rs, 23)];
|
cannam@167
|
840 T2G = T2E - T2F;
|
cannam@167
|
841 T4Q = T2E + T2F;
|
cannam@167
|
842 }
|
cannam@167
|
843 {
|
cannam@167
|
844 E T2c, T2d, T2f, T2g;
|
cannam@167
|
845 T2c = R1[WS(rs, 3)];
|
cannam@167
|
846 T2d = R1[WS(rs, 19)];
|
cannam@167
|
847 T2e = T2c - T2d;
|
cannam@167
|
848 T51 = T2c + T2d;
|
cannam@167
|
849 T2f = R1[WS(rs, 27)];
|
cannam@167
|
850 T2g = R1[WS(rs, 11)];
|
cannam@167
|
851 T2h = T2f - T2g;
|
cannam@167
|
852 T50 = T2f + T2g;
|
cannam@167
|
853 }
|
cannam@167
|
854 T4R = T4P - T4Q;
|
cannam@167
|
855 T52 = T50 - T51;
|
cannam@167
|
856 {
|
cannam@167
|
857 E T2i, T2D, T5N, T5O;
|
cannam@167
|
858 T2i = KP707106781 * (T2e + T2h);
|
cannam@167
|
859 T2j = T2b + T2i;
|
cannam@167
|
860 T3F = T2b - T2i;
|
cannam@167
|
861 T2D = KP707106781 * (T2h - T2e);
|
cannam@167
|
862 T2H = T2D - T2G;
|
cannam@167
|
863 T3I = T2G + T2D;
|
cannam@167
|
864 T5N = T4P + T4Q;
|
cannam@167
|
865 T5O = T51 + T50;
|
cannam@167
|
866 T5P = T5N + T5O;
|
cannam@167
|
867 T69 = T5N - T5O;
|
cannam@167
|
868 }
|
cannam@167
|
869 }
|
cannam@167
|
870 {
|
cannam@167
|
871 E TN, T1e, TX, T1g, TQ, T1k, TU, T1f, T1h, T1j;
|
cannam@167
|
872 {
|
cannam@167
|
873 E TL, TM, TV, TW;
|
cannam@167
|
874 TL = R0[WS(rs, 31)];
|
cannam@167
|
875 TM = R0[WS(rs, 15)];
|
cannam@167
|
876 TN = TL + TM;
|
cannam@167
|
877 T1e = TL - TM;
|
cannam@167
|
878 TV = R0[WS(rs, 27)];
|
cannam@167
|
879 TW = R0[WS(rs, 11)];
|
cannam@167
|
880 TX = TV + TW;
|
cannam@167
|
881 T1g = TV - TW;
|
cannam@167
|
882 }
|
cannam@167
|
883 {
|
cannam@167
|
884 E TO, TP, TS, TT;
|
cannam@167
|
885 TO = R0[WS(rs, 7)];
|
cannam@167
|
886 TP = R0[WS(rs, 23)];
|
cannam@167
|
887 TQ = TO + TP;
|
cannam@167
|
888 T1k = TO - TP;
|
cannam@167
|
889 TS = R0[WS(rs, 3)];
|
cannam@167
|
890 TT = R0[WS(rs, 19)];
|
cannam@167
|
891 TU = TS + TT;
|
cannam@167
|
892 T1f = TS - TT;
|
cannam@167
|
893 }
|
cannam@167
|
894 T1h = KP707106781 * (T1f + T1g);
|
cannam@167
|
895 T1i = T1e + T1h;
|
cannam@167
|
896 T3t = T1e - T1h;
|
cannam@167
|
897 T1j = KP707106781 * (T1g - T1f);
|
cannam@167
|
898 T1l = T1j - T1k;
|
cannam@167
|
899 T3u = T1k + T1j;
|
cannam@167
|
900 {
|
cannam@167
|
901 E TR, TY, T4t, T4u;
|
cannam@167
|
902 TR = TN + TQ;
|
cannam@167
|
903 TY = TU + TX;
|
cannam@167
|
904 TZ = TR + TY;
|
cannam@167
|
905 T63 = TR - TY;
|
cannam@167
|
906 T4t = TN - TQ;
|
cannam@167
|
907 T4u = TX - TU;
|
cannam@167
|
908 T4v = FNMS(KP382683432, T4u, KP923879532 * T4t);
|
cannam@167
|
909 T58 = FMA(KP382683432, T4t, KP923879532 * T4u);
|
cannam@167
|
910 }
|
cannam@167
|
911 }
|
cannam@167
|
912 {
|
cannam@167
|
913 E Ty, T1s, TI, T1n, TB, T1q, TF, T1o, T1p, T1t;
|
cannam@167
|
914 {
|
cannam@167
|
915 E Tw, Tx, TG, TH;
|
cannam@167
|
916 Tw = R0[WS(rs, 1)];
|
cannam@167
|
917 Tx = R0[WS(rs, 17)];
|
cannam@167
|
918 Ty = Tw + Tx;
|
cannam@167
|
919 T1s = Tw - Tx;
|
cannam@167
|
920 TG = R0[WS(rs, 29)];
|
cannam@167
|
921 TH = R0[WS(rs, 13)];
|
cannam@167
|
922 TI = TG + TH;
|
cannam@167
|
923 T1n = TG - TH;
|
cannam@167
|
924 }
|
cannam@167
|
925 {
|
cannam@167
|
926 E Tz, TA, TD, TE;
|
cannam@167
|
927 Tz = R0[WS(rs, 9)];
|
cannam@167
|
928 TA = R0[WS(rs, 25)];
|
cannam@167
|
929 TB = Tz + TA;
|
cannam@167
|
930 T1q = Tz - TA;
|
cannam@167
|
931 TD = R0[WS(rs, 5)];
|
cannam@167
|
932 TE = R0[WS(rs, 21)];
|
cannam@167
|
933 TF = TD + TE;
|
cannam@167
|
934 T1o = TD - TE;
|
cannam@167
|
935 }
|
cannam@167
|
936 T1p = KP707106781 * (T1n - T1o);
|
cannam@167
|
937 T1r = T1p - T1q;
|
cannam@167
|
938 T3r = T1q + T1p;
|
cannam@167
|
939 T1t = KP707106781 * (T1o + T1n);
|
cannam@167
|
940 T1u = T1s + T1t;
|
cannam@167
|
941 T3q = T1s - T1t;
|
cannam@167
|
942 {
|
cannam@167
|
943 E TC, TJ, T4q, T4r;
|
cannam@167
|
944 TC = Ty + TB;
|
cannam@167
|
945 TJ = TF + TI;
|
cannam@167
|
946 TK = TC + TJ;
|
cannam@167
|
947 T62 = TC - TJ;
|
cannam@167
|
948 T4q = Ty - TB;
|
cannam@167
|
949 T4r = TI - TF;
|
cannam@167
|
950 T4s = FMA(KP923879532, T4q, KP382683432 * T4r);
|
cannam@167
|
951 T57 = FNMS(KP382683432, T4q, KP923879532 * T4r);
|
cannam@167
|
952 }
|
cannam@167
|
953 }
|
cannam@167
|
954 {
|
cannam@167
|
955 E Ti, T16, Ts, T1a, Tl, T17, Tp, T19, T4m, T4n;
|
cannam@167
|
956 {
|
cannam@167
|
957 E Tg, Th, Tq, Tr;
|
cannam@167
|
958 Tg = R0[WS(rs, 2)];
|
cannam@167
|
959 Th = R0[WS(rs, 18)];
|
cannam@167
|
960 Ti = Tg + Th;
|
cannam@167
|
961 T16 = Tg - Th;
|
cannam@167
|
962 Tq = R0[WS(rs, 6)];
|
cannam@167
|
963 Tr = R0[WS(rs, 22)];
|
cannam@167
|
964 Ts = Tq + Tr;
|
cannam@167
|
965 T1a = Tq - Tr;
|
cannam@167
|
966 }
|
cannam@167
|
967 {
|
cannam@167
|
968 E Tj, Tk, Tn, To;
|
cannam@167
|
969 Tj = R0[WS(rs, 10)];
|
cannam@167
|
970 Tk = R0[WS(rs, 26)];
|
cannam@167
|
971 Tl = Tj + Tk;
|
cannam@167
|
972 T17 = Tj - Tk;
|
cannam@167
|
973 Tn = R0[WS(rs, 30)];
|
cannam@167
|
974 To = R0[WS(rs, 14)];
|
cannam@167
|
975 Tp = Tn + To;
|
cannam@167
|
976 T19 = Tn - To;
|
cannam@167
|
977 }
|
cannam@167
|
978 Tm = Ti + Tl;
|
cannam@167
|
979 Tt = Tp + Ts;
|
cannam@167
|
980 Tu = Tm + Tt;
|
cannam@167
|
981 T4m = Ti - Tl;
|
cannam@167
|
982 T4n = Tp - Ts;
|
cannam@167
|
983 T4o = KP707106781 * (T4m + T4n);
|
cannam@167
|
984 T5b = KP707106781 * (T4n - T4m);
|
cannam@167
|
985 {
|
cannam@167
|
986 E T18, T1b, T2O, T2P;
|
cannam@167
|
987 T18 = FNMS(KP382683432, T17, KP923879532 * T16);
|
cannam@167
|
988 T1b = FMA(KP923879532, T19, KP382683432 * T1a);
|
cannam@167
|
989 T1c = T18 + T1b;
|
cannam@167
|
990 T3R = T1b - T18;
|
cannam@167
|
991 T2O = FNMS(KP923879532, T1a, KP382683432 * T19);
|
cannam@167
|
992 T2P = FMA(KP382683432, T16, KP923879532 * T17);
|
cannam@167
|
993 T2Q = T2O - T2P;
|
cannam@167
|
994 T3o = T2P + T2O;
|
cannam@167
|
995 }
|
cannam@167
|
996 }
|
cannam@167
|
997 {
|
cannam@167
|
998 E T1A, T4E, T1K, T4C, T1D, T4F, T1H, T4B;
|
cannam@167
|
999 {
|
cannam@167
|
1000 E T1y, T1z, T1I, T1J;
|
cannam@167
|
1001 T1y = R1[WS(rs, 30)];
|
cannam@167
|
1002 T1z = R1[WS(rs, 14)];
|
cannam@167
|
1003 T1A = T1y - T1z;
|
cannam@167
|
1004 T4E = T1y + T1z;
|
cannam@167
|
1005 T1I = R1[WS(rs, 10)];
|
cannam@167
|
1006 T1J = R1[WS(rs, 26)];
|
cannam@167
|
1007 T1K = T1I - T1J;
|
cannam@167
|
1008 T4C = T1I + T1J;
|
cannam@167
|
1009 }
|
cannam@167
|
1010 {
|
cannam@167
|
1011 E T1B, T1C, T1F, T1G;
|
cannam@167
|
1012 T1B = R1[WS(rs, 6)];
|
cannam@167
|
1013 T1C = R1[WS(rs, 22)];
|
cannam@167
|
1014 T1D = T1B - T1C;
|
cannam@167
|
1015 T4F = T1B + T1C;
|
cannam@167
|
1016 T1F = R1[WS(rs, 2)];
|
cannam@167
|
1017 T1G = R1[WS(rs, 18)];
|
cannam@167
|
1018 T1H = T1F - T1G;
|
cannam@167
|
1019 T4B = T1F + T1G;
|
cannam@167
|
1020 }
|
cannam@167
|
1021 {
|
cannam@167
|
1022 E T1E, T1L, T5J, T5K;
|
cannam@167
|
1023 T1E = FNMS(KP923879532, T1D, KP382683432 * T1A);
|
cannam@167
|
1024 T1L = FMA(KP382683432, T1H, KP923879532 * T1K);
|
cannam@167
|
1025 T1M = T1E - T1L;
|
cannam@167
|
1026 T3z = T1L + T1E;
|
cannam@167
|
1027 T5J = T4B + T4C;
|
cannam@167
|
1028 T5K = T4E + T4F;
|
cannam@167
|
1029 T5L = T5J + T5K;
|
cannam@167
|
1030 T67 = T5K - T5J;
|
cannam@167
|
1031 }
|
cannam@167
|
1032 {
|
cannam@167
|
1033 E T24, T25, T4D, T4G;
|
cannam@167
|
1034 T24 = FNMS(KP382683432, T1K, KP923879532 * T1H);
|
cannam@167
|
1035 T25 = FMA(KP923879532, T1A, KP382683432 * T1D);
|
cannam@167
|
1036 T26 = T24 + T25;
|
cannam@167
|
1037 T3C = T25 - T24;
|
cannam@167
|
1038 T4D = T4B - T4C;
|
cannam@167
|
1039 T4G = T4E - T4F;
|
cannam@167
|
1040 T4H = KP707106781 * (T4D + T4G);
|
cannam@167
|
1041 T4M = KP707106781 * (T4G - T4D);
|
cannam@167
|
1042 }
|
cannam@167
|
1043 }
|
cannam@167
|
1044 {
|
cannam@167
|
1045 E T2m, T4S, T2w, T4W, T2p, T4T, T2t, T4V;
|
cannam@167
|
1046 {
|
cannam@167
|
1047 E T2k, T2l, T2u, T2v;
|
cannam@167
|
1048 T2k = R1[WS(rs, 1)];
|
cannam@167
|
1049 T2l = R1[WS(rs, 17)];
|
cannam@167
|
1050 T2m = T2k - T2l;
|
cannam@167
|
1051 T4S = T2k + T2l;
|
cannam@167
|
1052 T2u = R1[WS(rs, 5)];
|
cannam@167
|
1053 T2v = R1[WS(rs, 21)];
|
cannam@167
|
1054 T2w = T2u - T2v;
|
cannam@167
|
1055 T4W = T2u + T2v;
|
cannam@167
|
1056 }
|
cannam@167
|
1057 {
|
cannam@167
|
1058 E T2n, T2o, T2r, T2s;
|
cannam@167
|
1059 T2n = R1[WS(rs, 9)];
|
cannam@167
|
1060 T2o = R1[WS(rs, 25)];
|
cannam@167
|
1061 T2p = T2n - T2o;
|
cannam@167
|
1062 T4T = T2n + T2o;
|
cannam@167
|
1063 T2r = R1[WS(rs, 29)];
|
cannam@167
|
1064 T2s = R1[WS(rs, 13)];
|
cannam@167
|
1065 T2t = T2r - T2s;
|
cannam@167
|
1066 T4V = T2r + T2s;
|
cannam@167
|
1067 }
|
cannam@167
|
1068 {
|
cannam@167
|
1069 E T2q, T2x, T5Q, T5R;
|
cannam@167
|
1070 T2q = FNMS(KP382683432, T2p, KP923879532 * T2m);
|
cannam@167
|
1071 T2x = FMA(KP923879532, T2t, KP382683432 * T2w);
|
cannam@167
|
1072 T2y = T2q + T2x;
|
cannam@167
|
1073 T3J = T2x - T2q;
|
cannam@167
|
1074 T5Q = T4S + T4T;
|
cannam@167
|
1075 T5R = T4V + T4W;
|
cannam@167
|
1076 T5S = T5Q + T5R;
|
cannam@167
|
1077 T6a = T5R - T5Q;
|
cannam@167
|
1078 }
|
cannam@167
|
1079 {
|
cannam@167
|
1080 E T2A, T2B, T4U, T4X;
|
cannam@167
|
1081 T2A = FNMS(KP923879532, T2w, KP382683432 * T2t);
|
cannam@167
|
1082 T2B = FMA(KP382683432, T2m, KP923879532 * T2p);
|
cannam@167
|
1083 T2C = T2A - T2B;
|
cannam@167
|
1084 T3G = T2B + T2A;
|
cannam@167
|
1085 T4U = T4S - T4T;
|
cannam@167
|
1086 T4X = T4V - T4W;
|
cannam@167
|
1087 T4Y = KP707106781 * (T4U + T4X);
|
cannam@167
|
1088 T53 = KP707106781 * (T4X - T4U);
|
cannam@167
|
1089 }
|
cannam@167
|
1090 }
|
cannam@167
|
1091 {
|
cannam@167
|
1092 E Tv, T10, T5X, T5Y, T5Z, T60;
|
cannam@167
|
1093 Tv = Tf + Tu;
|
cannam@167
|
1094 T10 = TK + TZ;
|
cannam@167
|
1095 T5X = Tv + T10;
|
cannam@167
|
1096 T5Y = T5I + T5L;
|
cannam@167
|
1097 T5Z = T5P + T5S;
|
cannam@167
|
1098 T60 = T5Y + T5Z;
|
cannam@167
|
1099 Cr[WS(csr, 16)] = Tv - T10;
|
cannam@167
|
1100 Ci[WS(csi, 16)] = T5Z - T5Y;
|
cannam@167
|
1101 Cr[WS(csr, 32)] = T5X - T60;
|
cannam@167
|
1102 Cr[0] = T5X + T60;
|
cannam@167
|
1103 }
|
cannam@167
|
1104 {
|
cannam@167
|
1105 E T5F, T5V, T5U, T5W, T5M, T5T;
|
cannam@167
|
1106 T5F = Tf - Tu;
|
cannam@167
|
1107 T5V = TZ - TK;
|
cannam@167
|
1108 T5M = T5I - T5L;
|
cannam@167
|
1109 T5T = T5P - T5S;
|
cannam@167
|
1110 T5U = KP707106781 * (T5M + T5T);
|
cannam@167
|
1111 T5W = KP707106781 * (T5T - T5M);
|
cannam@167
|
1112 Cr[WS(csr, 24)] = T5F - T5U;
|
cannam@167
|
1113 Ci[WS(csi, 24)] = T5W - T5V;
|
cannam@167
|
1114 Cr[WS(csr, 8)] = T5F + T5U;
|
cannam@167
|
1115 Ci[WS(csi, 8)] = T5V + T5W;
|
cannam@167
|
1116 }
|
cannam@167
|
1117 {
|
cannam@167
|
1118 E T65, T6l, T6k, T6m, T6c, T6g, T6f, T6h;
|
cannam@167
|
1119 {
|
cannam@167
|
1120 E T61, T64, T6i, T6j;
|
cannam@167
|
1121 T61 = T7 - Te;
|
cannam@167
|
1122 T64 = KP707106781 * (T62 + T63);
|
cannam@167
|
1123 T65 = T61 + T64;
|
cannam@167
|
1124 T6l = T61 - T64;
|
cannam@167
|
1125 T6i = FNMS(KP382683432, T66, KP923879532 * T67);
|
cannam@167
|
1126 T6j = FMA(KP382683432, T69, KP923879532 * T6a);
|
cannam@167
|
1127 T6k = T6i + T6j;
|
cannam@167
|
1128 T6m = T6j - T6i;
|
cannam@167
|
1129 }
|
cannam@167
|
1130 {
|
cannam@167
|
1131 E T68, T6b, T6d, T6e;
|
cannam@167
|
1132 T68 = FMA(KP923879532, T66, KP382683432 * T67);
|
cannam@167
|
1133 T6b = FNMS(KP382683432, T6a, KP923879532 * T69);
|
cannam@167
|
1134 T6c = T68 + T6b;
|
cannam@167
|
1135 T6g = T6b - T68;
|
cannam@167
|
1136 T6d = KP707106781 * (T63 - T62);
|
cannam@167
|
1137 T6e = Tt - Tm;
|
cannam@167
|
1138 T6f = T6d - T6e;
|
cannam@167
|
1139 T6h = T6e + T6d;
|
cannam@167
|
1140 }
|
cannam@167
|
1141 Cr[WS(csr, 28)] = T65 - T6c;
|
cannam@167
|
1142 Ci[WS(csi, 28)] = T6k - T6h;
|
cannam@167
|
1143 Cr[WS(csr, 4)] = T65 + T6c;
|
cannam@167
|
1144 Ci[WS(csi, 4)] = T6h + T6k;
|
cannam@167
|
1145 Ci[WS(csi, 12)] = T6f + T6g;
|
cannam@167
|
1146 Cr[WS(csr, 12)] = T6l + T6m;
|
cannam@167
|
1147 Ci[WS(csi, 20)] = T6g - T6f;
|
cannam@167
|
1148 Cr[WS(csr, 20)] = T6l - T6m;
|
cannam@167
|
1149 }
|
cannam@167
|
1150 {
|
cannam@167
|
1151 E T5n, T5D, T5x, T5z, T5q, T5A, T5t, T5B;
|
cannam@167
|
1152 {
|
cannam@167
|
1153 E T5l, T5m, T5v, T5w;
|
cannam@167
|
1154 T5l = T4l - T4o;
|
cannam@167
|
1155 T5m = T58 - T57;
|
cannam@167
|
1156 T5n = T5l + T5m;
|
cannam@167
|
1157 T5D = T5l - T5m;
|
cannam@167
|
1158 T5v = T4v - T4s;
|
cannam@167
|
1159 T5w = T5b - T5a;
|
cannam@167
|
1160 T5x = T5v - T5w;
|
cannam@167
|
1161 T5z = T5w + T5v;
|
cannam@167
|
1162 }
|
cannam@167
|
1163 {
|
cannam@167
|
1164 E T5o, T5p, T5r, T5s;
|
cannam@167
|
1165 T5o = T4A - T4H;
|
cannam@167
|
1166 T5p = T4M - T4L;
|
cannam@167
|
1167 T5q = FMA(KP831469612, T5o, KP555570233 * T5p);
|
cannam@167
|
1168 T5A = FNMS(KP555570233, T5o, KP831469612 * T5p);
|
cannam@167
|
1169 T5r = T4R - T4Y;
|
cannam@167
|
1170 T5s = T53 - T52;
|
cannam@167
|
1171 T5t = FNMS(KP555570233, T5s, KP831469612 * T5r);
|
cannam@167
|
1172 T5B = FMA(KP555570233, T5r, KP831469612 * T5s);
|
cannam@167
|
1173 }
|
cannam@167
|
1174 {
|
cannam@167
|
1175 E T5u, T5C, T5y, T5E;
|
cannam@167
|
1176 T5u = T5q + T5t;
|
cannam@167
|
1177 Cr[WS(csr, 26)] = T5n - T5u;
|
cannam@167
|
1178 Cr[WS(csr, 6)] = T5n + T5u;
|
cannam@167
|
1179 T5C = T5A + T5B;
|
cannam@167
|
1180 Ci[WS(csi, 6)] = T5z + T5C;
|
cannam@167
|
1181 Ci[WS(csi, 26)] = T5C - T5z;
|
cannam@167
|
1182 T5y = T5t - T5q;
|
cannam@167
|
1183 Ci[WS(csi, 10)] = T5x + T5y;
|
cannam@167
|
1184 Ci[WS(csi, 22)] = T5y - T5x;
|
cannam@167
|
1185 T5E = T5B - T5A;
|
cannam@167
|
1186 Cr[WS(csr, 22)] = T5D - T5E;
|
cannam@167
|
1187 Cr[WS(csr, 10)] = T5D + T5E;
|
cannam@167
|
1188 }
|
cannam@167
|
1189 }
|
cannam@167
|
1190 {
|
cannam@167
|
1191 E T4x, T5j, T5d, T5f, T4O, T5g, T55, T5h;
|
cannam@167
|
1192 {
|
cannam@167
|
1193 E T4p, T4w, T59, T5c;
|
cannam@167
|
1194 T4p = T4l + T4o;
|
cannam@167
|
1195 T4w = T4s + T4v;
|
cannam@167
|
1196 T4x = T4p + T4w;
|
cannam@167
|
1197 T5j = T4p - T4w;
|
cannam@167
|
1198 T59 = T57 + T58;
|
cannam@167
|
1199 T5c = T5a + T5b;
|
cannam@167
|
1200 T5d = T59 - T5c;
|
cannam@167
|
1201 T5f = T5c + T59;
|
cannam@167
|
1202 }
|
cannam@167
|
1203 {
|
cannam@167
|
1204 E T4I, T4N, T4Z, T54;
|
cannam@167
|
1205 T4I = T4A + T4H;
|
cannam@167
|
1206 T4N = T4L + T4M;
|
cannam@167
|
1207 T4O = FMA(KP980785280, T4I, KP195090322 * T4N);
|
cannam@167
|
1208 T5g = FNMS(KP195090322, T4I, KP980785280 * T4N);
|
cannam@167
|
1209 T4Z = T4R + T4Y;
|
cannam@167
|
1210 T54 = T52 + T53;
|
cannam@167
|
1211 T55 = FNMS(KP195090322, T54, KP980785280 * T4Z);
|
cannam@167
|
1212 T5h = FMA(KP195090322, T4Z, KP980785280 * T54);
|
cannam@167
|
1213 }
|
cannam@167
|
1214 {
|
cannam@167
|
1215 E T56, T5i, T5e, T5k;
|
cannam@167
|
1216 T56 = T4O + T55;
|
cannam@167
|
1217 Cr[WS(csr, 30)] = T4x - T56;
|
cannam@167
|
1218 Cr[WS(csr, 2)] = T4x + T56;
|
cannam@167
|
1219 T5i = T5g + T5h;
|
cannam@167
|
1220 Ci[WS(csi, 2)] = T5f + T5i;
|
cannam@167
|
1221 Ci[WS(csi, 30)] = T5i - T5f;
|
cannam@167
|
1222 T5e = T55 - T4O;
|
cannam@167
|
1223 Ci[WS(csi, 14)] = T5d + T5e;
|
cannam@167
|
1224 Ci[WS(csi, 18)] = T5e - T5d;
|
cannam@167
|
1225 T5k = T5h - T5g;
|
cannam@167
|
1226 Cr[WS(csr, 18)] = T5j - T5k;
|
cannam@167
|
1227 Cr[WS(csr, 14)] = T5j + T5k;
|
cannam@167
|
1228 }
|
cannam@167
|
1229 }
|
cannam@167
|
1230 {
|
cannam@167
|
1231 E T3p, T41, T4c, T3S, T3w, T4b, T49, T4h, T3P, T42, T3E, T3W, T46, T4g, T3L;
|
cannam@167
|
1232 E T3X;
|
cannam@167
|
1233 {
|
cannam@167
|
1234 E T3s, T3v, T3A, T3D;
|
cannam@167
|
1235 T3p = T3n + T3o;
|
cannam@167
|
1236 T41 = T3n - T3o;
|
cannam@167
|
1237 T4c = T3R - T3Q;
|
cannam@167
|
1238 T3S = T3Q + T3R;
|
cannam@167
|
1239 T3s = FMA(KP831469612, T3q, KP555570233 * T3r);
|
cannam@167
|
1240 T3v = FNMS(KP555570233, T3u, KP831469612 * T3t);
|
cannam@167
|
1241 T3w = T3s + T3v;
|
cannam@167
|
1242 T4b = T3v - T3s;
|
cannam@167
|
1243 {
|
cannam@167
|
1244 E T47, T48, T3N, T3O;
|
cannam@167
|
1245 T47 = T3F - T3G;
|
cannam@167
|
1246 T48 = T3J - T3I;
|
cannam@167
|
1247 T49 = FNMS(KP471396736, T48, KP881921264 * T47);
|
cannam@167
|
1248 T4h = FMA(KP471396736, T47, KP881921264 * T48);
|
cannam@167
|
1249 T3N = FNMS(KP555570233, T3q, KP831469612 * T3r);
|
cannam@167
|
1250 T3O = FMA(KP555570233, T3t, KP831469612 * T3u);
|
cannam@167
|
1251 T3P = T3N + T3O;
|
cannam@167
|
1252 T42 = T3O - T3N;
|
cannam@167
|
1253 }
|
cannam@167
|
1254 T3A = T3y + T3z;
|
cannam@167
|
1255 T3D = T3B + T3C;
|
cannam@167
|
1256 T3E = FMA(KP956940335, T3A, KP290284677 * T3D);
|
cannam@167
|
1257 T3W = FNMS(KP290284677, T3A, KP956940335 * T3D);
|
cannam@167
|
1258 {
|
cannam@167
|
1259 E T44, T45, T3H, T3K;
|
cannam@167
|
1260 T44 = T3y - T3z;
|
cannam@167
|
1261 T45 = T3C - T3B;
|
cannam@167
|
1262 T46 = FMA(KP881921264, T44, KP471396736 * T45);
|
cannam@167
|
1263 T4g = FNMS(KP471396736, T44, KP881921264 * T45);
|
cannam@167
|
1264 T3H = T3F + T3G;
|
cannam@167
|
1265 T3K = T3I + T3J;
|
cannam@167
|
1266 T3L = FNMS(KP290284677, T3K, KP956940335 * T3H);
|
cannam@167
|
1267 T3X = FMA(KP290284677, T3H, KP956940335 * T3K);
|
cannam@167
|
1268 }
|
cannam@167
|
1269 }
|
cannam@167
|
1270 {
|
cannam@167
|
1271 E T3x, T3M, T3V, T3Y;
|
cannam@167
|
1272 T3x = T3p + T3w;
|
cannam@167
|
1273 T3M = T3E + T3L;
|
cannam@167
|
1274 Cr[WS(csr, 29)] = T3x - T3M;
|
cannam@167
|
1275 Cr[WS(csr, 3)] = T3x + T3M;
|
cannam@167
|
1276 T3V = T3S + T3P;
|
cannam@167
|
1277 T3Y = T3W + T3X;
|
cannam@167
|
1278 Ci[WS(csi, 3)] = T3V + T3Y;
|
cannam@167
|
1279 Ci[WS(csi, 29)] = T3Y - T3V;
|
cannam@167
|
1280 }
|
cannam@167
|
1281 {
|
cannam@167
|
1282 E T3T, T3U, T3Z, T40;
|
cannam@167
|
1283 T3T = T3P - T3S;
|
cannam@167
|
1284 T3U = T3L - T3E;
|
cannam@167
|
1285 Ci[WS(csi, 13)] = T3T + T3U;
|
cannam@167
|
1286 Ci[WS(csi, 19)] = T3U - T3T;
|
cannam@167
|
1287 T3Z = T3p - T3w;
|
cannam@167
|
1288 T40 = T3X - T3W;
|
cannam@167
|
1289 Cr[WS(csr, 19)] = T3Z - T40;
|
cannam@167
|
1290 Cr[WS(csr, 13)] = T3Z + T40;
|
cannam@167
|
1291 }
|
cannam@167
|
1292 {
|
cannam@167
|
1293 E T43, T4a, T4f, T4i;
|
cannam@167
|
1294 T43 = T41 + T42;
|
cannam@167
|
1295 T4a = T46 + T49;
|
cannam@167
|
1296 Cr[WS(csr, 27)] = T43 - T4a;
|
cannam@167
|
1297 Cr[WS(csr, 5)] = T43 + T4a;
|
cannam@167
|
1298 T4f = T4c + T4b;
|
cannam@167
|
1299 T4i = T4g + T4h;
|
cannam@167
|
1300 Ci[WS(csi, 5)] = T4f + T4i;
|
cannam@167
|
1301 Ci[WS(csi, 27)] = T4i - T4f;
|
cannam@167
|
1302 }
|
cannam@167
|
1303 {
|
cannam@167
|
1304 E T4d, T4e, T4j, T4k;
|
cannam@167
|
1305 T4d = T4b - T4c;
|
cannam@167
|
1306 T4e = T49 - T46;
|
cannam@167
|
1307 Ci[WS(csi, 11)] = T4d + T4e;
|
cannam@167
|
1308 Ci[WS(csi, 21)] = T4e - T4d;
|
cannam@167
|
1309 T4j = T41 - T42;
|
cannam@167
|
1310 T4k = T4h - T4g;
|
cannam@167
|
1311 Cr[WS(csr, 21)] = T4j - T4k;
|
cannam@167
|
1312 Cr[WS(csr, 11)] = T4j + T4k;
|
cannam@167
|
1313 }
|
cannam@167
|
1314 }
|
cannam@167
|
1315 {
|
cannam@167
|
1316 E T1d, T33, T3e, T2U, T1w, T3d, T3b, T3j, T2N, T34, T28, T2Y, T38, T3i, T2J;
|
cannam@167
|
1317 E T2Z;
|
cannam@167
|
1318 {
|
cannam@167
|
1319 E T1m, T1v, T1Y, T27;
|
cannam@167
|
1320 T1d = T15 - T1c;
|
cannam@167
|
1321 T33 = T15 + T1c;
|
cannam@167
|
1322 T3e = T2T + T2Q;
|
cannam@167
|
1323 T2U = T2Q - T2T;
|
cannam@167
|
1324 T1m = FMA(KP195090322, T1i, KP980785280 * T1l);
|
cannam@167
|
1325 T1v = FNMS(KP195090322, T1u, KP980785280 * T1r);
|
cannam@167
|
1326 T1w = T1m - T1v;
|
cannam@167
|
1327 T3d = T1v + T1m;
|
cannam@167
|
1328 {
|
cannam@167
|
1329 E T39, T3a, T2L, T2M;
|
cannam@167
|
1330 T39 = T2j + T2y;
|
cannam@167
|
1331 T3a = T2H + T2C;
|
cannam@167
|
1332 T3b = FNMS(KP098017140, T3a, KP995184726 * T39);
|
cannam@167
|
1333 T3j = FMA(KP995184726, T3a, KP098017140 * T39);
|
cannam@167
|
1334 T2L = FNMS(KP195090322, T1l, KP980785280 * T1i);
|
cannam@167
|
1335 T2M = FMA(KP980785280, T1u, KP195090322 * T1r);
|
cannam@167
|
1336 T2N = T2L - T2M;
|
cannam@167
|
1337 T34 = T2M + T2L;
|
cannam@167
|
1338 }
|
cannam@167
|
1339 T1Y = T1M - T1X;
|
cannam@167
|
1340 T27 = T23 - T26;
|
cannam@167
|
1341 T28 = FMA(KP634393284, T1Y, KP773010453 * T27);
|
cannam@167
|
1342 T2Y = FNMS(KP634393284, T27, KP773010453 * T1Y);
|
cannam@167
|
1343 {
|
cannam@167
|
1344 E T36, T37, T2z, T2I;
|
cannam@167
|
1345 T36 = T1X + T1M;
|
cannam@167
|
1346 T37 = T23 + T26;
|
cannam@167
|
1347 T38 = FMA(KP098017140, T36, KP995184726 * T37);
|
cannam@167
|
1348 T3i = FNMS(KP098017140, T37, KP995184726 * T36);
|
cannam@167
|
1349 T2z = T2j - T2y;
|
cannam@167
|
1350 T2I = T2C - T2H;
|
cannam@167
|
1351 T2J = FNMS(KP634393284, T2I, KP773010453 * T2z);
|
cannam@167
|
1352 T2Z = FMA(KP773010453, T2I, KP634393284 * T2z);
|
cannam@167
|
1353 }
|
cannam@167
|
1354 }
|
cannam@167
|
1355 {
|
cannam@167
|
1356 E T1x, T2K, T2X, T30;
|
cannam@167
|
1357 T1x = T1d + T1w;
|
cannam@167
|
1358 T2K = T28 + T2J;
|
cannam@167
|
1359 Cr[WS(csr, 25)] = T1x - T2K;
|
cannam@167
|
1360 Cr[WS(csr, 7)] = T1x + T2K;
|
cannam@167
|
1361 T2X = T2U + T2N;
|
cannam@167
|
1362 T30 = T2Y + T2Z;
|
cannam@167
|
1363 Ci[WS(csi, 7)] = T2X + T30;
|
cannam@167
|
1364 Ci[WS(csi, 25)] = T30 - T2X;
|
cannam@167
|
1365 }
|
cannam@167
|
1366 {
|
cannam@167
|
1367 E T2V, T2W, T31, T32;
|
cannam@167
|
1368 T2V = T2N - T2U;
|
cannam@167
|
1369 T2W = T2J - T28;
|
cannam@167
|
1370 Ci[WS(csi, 9)] = T2V + T2W;
|
cannam@167
|
1371 Ci[WS(csi, 23)] = T2W - T2V;
|
cannam@167
|
1372 T31 = T1d - T1w;
|
cannam@167
|
1373 T32 = T2Z - T2Y;
|
cannam@167
|
1374 Cr[WS(csr, 23)] = T31 - T32;
|
cannam@167
|
1375 Cr[WS(csr, 9)] = T31 + T32;
|
cannam@167
|
1376 }
|
cannam@167
|
1377 {
|
cannam@167
|
1378 E T35, T3c, T3h, T3k;
|
cannam@167
|
1379 T35 = T33 + T34;
|
cannam@167
|
1380 T3c = T38 + T3b;
|
cannam@167
|
1381 Cr[WS(csr, 31)] = T35 - T3c;
|
cannam@167
|
1382 Cr[WS(csr, 1)] = T35 + T3c;
|
cannam@167
|
1383 T3h = T3e + T3d;
|
cannam@167
|
1384 T3k = T3i + T3j;
|
cannam@167
|
1385 Ci[WS(csi, 1)] = T3h + T3k;
|
cannam@167
|
1386 Ci[WS(csi, 31)] = T3k - T3h;
|
cannam@167
|
1387 }
|
cannam@167
|
1388 {
|
cannam@167
|
1389 E T3f, T3g, T3l, T3m;
|
cannam@167
|
1390 T3f = T3d - T3e;
|
cannam@167
|
1391 T3g = T3b - T38;
|
cannam@167
|
1392 Ci[WS(csi, 15)] = T3f + T3g;
|
cannam@167
|
1393 Ci[WS(csi, 17)] = T3g - T3f;
|
cannam@167
|
1394 T3l = T33 - T34;
|
cannam@167
|
1395 T3m = T3j - T3i;
|
cannam@167
|
1396 Cr[WS(csr, 17)] = T3l - T3m;
|
cannam@167
|
1397 Cr[WS(csr, 15)] = T3l + T3m;
|
cannam@167
|
1398 }
|
cannam@167
|
1399 }
|
cannam@167
|
1400 }
|
cannam@167
|
1401 }
|
cannam@167
|
1402 }
|
cannam@167
|
1403
|
cannam@167
|
1404 static const kr2c_desc desc = { 64, "r2cf_64", {342, 72, 52, 0}, &GENUS };
|
cannam@167
|
1405
|
cannam@167
|
1406 void X(codelet_r2cf_64) (planner *p) {
|
cannam@167
|
1407 X(kr2c_register) (p, r2cf_64, &desc);
|
cannam@167
|
1408 }
|
cannam@167
|
1409
|
cannam@167
|
1410 #endif
|