cannam@127
|
1 /*
|
cannam@127
|
2 * Copyright (c) 2003, 2007-14 Matteo Frigo
|
cannam@127
|
3 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
|
cannam@127
|
4 *
|
cannam@127
|
5 * This program is free software; you can redistribute it and/or modify
|
cannam@127
|
6 * it under the terms of the GNU General Public License as published by
|
cannam@127
|
7 * the Free Software Foundation; either version 2 of the License, or
|
cannam@127
|
8 * (at your option) any later version.
|
cannam@127
|
9 *
|
cannam@127
|
10 * This program is distributed in the hope that it will be useful,
|
cannam@127
|
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
cannam@127
|
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
cannam@127
|
13 * GNU General Public License for more details.
|
cannam@127
|
14 *
|
cannam@127
|
15 * You should have received a copy of the GNU General Public License
|
cannam@127
|
16 * along with this program; if not, write to the Free Software
|
cannam@127
|
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
cannam@127
|
18 *
|
cannam@127
|
19 */
|
cannam@127
|
20
|
cannam@127
|
21 /* This file was automatically generated --- DO NOT EDIT */
|
cannam@127
|
22 /* Generated on Sat Jul 30 16:46:08 EDT 2016 */
|
cannam@127
|
23
|
cannam@127
|
24 #include "codelet-rdft.h"
|
cannam@127
|
25
|
cannam@127
|
26 #ifdef HAVE_FMA
|
cannam@127
|
27
|
cannam@127
|
28 /* Generated by: ../../../genfft/gen_r2cf.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -n 64 -name r2cf_64 -include r2cf.h */
|
cannam@127
|
29
|
cannam@127
|
30 /*
|
cannam@127
|
31 * This function contains 394 FP additions, 196 FP multiplications,
|
cannam@127
|
32 * (or, 198 additions, 0 multiplications, 196 fused multiply/add),
|
cannam@127
|
33 * 133 stack variables, 15 constants, and 128 memory accesses
|
cannam@127
|
34 */
|
cannam@127
|
35 #include "r2cf.h"
|
cannam@127
|
36
|
cannam@127
|
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@127
|
38 {
|
cannam@127
|
39 DK(KP773010453, +0.773010453362736960810906609758469800971041293);
|
cannam@127
|
40 DK(KP995184726, +0.995184726672196886244836953109479921575474869);
|
cannam@127
|
41 DK(KP098491403, +0.098491403357164253077197521291327432293052451);
|
cannam@127
|
42 DK(KP820678790, +0.820678790828660330972281985331011598767386482);
|
cannam@127
|
43 DK(KP956940335, +0.956940335732208864935797886980269969482849206);
|
cannam@127
|
44 DK(KP881921264, +0.881921264348355029712756863660388349508442621);
|
cannam@127
|
45 DK(KP534511135, +0.534511135950791641089685961295362908582039528);
|
cannam@127
|
46 DK(KP303346683, +0.303346683607342391675883946941299872384187453);
|
cannam@127
|
47 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
cannam@127
|
48 DK(KP198912367, +0.198912367379658006911597622644676228597850501);
|
cannam@127
|
49 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
cannam@127
|
50 DK(KP668178637, +0.668178637919298919997757686523080761552472251);
|
cannam@127
|
51 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
cannam@127
|
52 DK(KP414213562, +0.414213562373095048801688724209698078569671875);
|
cannam@127
|
53 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
cannam@127
|
54 {
|
cannam@127
|
55 INT i;
|
cannam@127
|
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@127
|
57 E T5n, T5o;
|
cannam@127
|
58 {
|
cannam@127
|
59 E T11, T2j, T4P, T5P, T3D, T5p, T3d, Tf, T1k, T1H, T5D, T4l, T5A, T4a, T3i;
|
cannam@127
|
60 E T2U, T1R, T2e, T5K, T4G, T5H, T4v, T3l, T31, T5s, T42, T5t, T3Z, T2n, T1b;
|
cannam@127
|
61 E T3f, TZ, T5v, T3T, T5w, T3Q, T2m, T18, T3e, TK, T3K, T5Q, T4S, T5q, T14;
|
cannam@127
|
62 E T2k, T3p, Tu, T4w, T1U, T5E, T4h, T5B, T4o, T3j, T2X, T1I, T1z, T1Z, T4A;
|
cannam@127
|
63 E T24, T4x, T1X, T20;
|
cannam@127
|
64 {
|
cannam@127
|
65 E TN, T3V, TS, TX, T3X, TQ, T40, TT;
|
cannam@127
|
66 {
|
cannam@127
|
67 E T1g, T46, T1B, T1G, T47, T1j, T4j, T1C;
|
cannam@127
|
68 {
|
cannam@127
|
69 E T4, T3z, T3, T3B, Td, T5, T8, T9;
|
cannam@127
|
70 {
|
cannam@127
|
71 E T1, T2, Tb, Tc;
|
cannam@127
|
72 T1 = R0[0];
|
cannam@127
|
73 T2 = R0[WS(rs, 16)];
|
cannam@127
|
74 Tb = R0[WS(rs, 28)];
|
cannam@127
|
75 Tc = R0[WS(rs, 12)];
|
cannam@127
|
76 T4 = R0[WS(rs, 8)];
|
cannam@127
|
77 T3z = T1 - T2;
|
cannam@127
|
78 T3 = T1 + T2;
|
cannam@127
|
79 T3B = Tb - Tc;
|
cannam@127
|
80 Td = Tb + Tc;
|
cannam@127
|
81 T5 = R0[WS(rs, 24)];
|
cannam@127
|
82 T8 = R0[WS(rs, 4)];
|
cannam@127
|
83 T9 = R0[WS(rs, 20)];
|
cannam@127
|
84 }
|
cannam@127
|
85 {
|
cannam@127
|
86 E T1E, T1F, T1h, T1i;
|
cannam@127
|
87 {
|
cannam@127
|
88 E T1e, T4N, T6, T3A, Ta, T1f;
|
cannam@127
|
89 T1e = R1[0];
|
cannam@127
|
90 T4N = T4 - T5;
|
cannam@127
|
91 T6 = T4 + T5;
|
cannam@127
|
92 T3A = T8 - T9;
|
cannam@127
|
93 Ta = T8 + T9;
|
cannam@127
|
94 T1f = R1[WS(rs, 16)];
|
cannam@127
|
95 {
|
cannam@127
|
96 E T7, T3C, T4O, Te;
|
cannam@127
|
97 T11 = T3 - T6;
|
cannam@127
|
98 T7 = T3 + T6;
|
cannam@127
|
99 T3C = T3A + T3B;
|
cannam@127
|
100 T4O = T3B - T3A;
|
cannam@127
|
101 T2j = Td - Ta;
|
cannam@127
|
102 Te = Ta + Td;
|
cannam@127
|
103 T4P = FNMS(KP707106781, T4O, T4N);
|
cannam@127
|
104 T5P = FMA(KP707106781, T4O, T4N);
|
cannam@127
|
105 T3D = FMA(KP707106781, T3C, T3z);
|
cannam@127
|
106 T5p = FNMS(KP707106781, T3C, T3z);
|
cannam@127
|
107 T3d = T7 - Te;
|
cannam@127
|
108 Tf = T7 + Te;
|
cannam@127
|
109 T1g = T1e + T1f;
|
cannam@127
|
110 T46 = T1e - T1f;
|
cannam@127
|
111 }
|
cannam@127
|
112 }
|
cannam@127
|
113 T1E = R1[WS(rs, 4)];
|
cannam@127
|
114 T1F = R1[WS(rs, 20)];
|
cannam@127
|
115 T1h = R1[WS(rs, 8)];
|
cannam@127
|
116 T1i = R1[WS(rs, 24)];
|
cannam@127
|
117 T1B = R1[WS(rs, 28)];
|
cannam@127
|
118 T1G = T1E + T1F;
|
cannam@127
|
119 T47 = T1E - T1F;
|
cannam@127
|
120 T1j = T1h + T1i;
|
cannam@127
|
121 T4j = T1h - T1i;
|
cannam@127
|
122 T1C = R1[WS(rs, 12)];
|
cannam@127
|
123 }
|
cannam@127
|
124 }
|
cannam@127
|
125 {
|
cannam@127
|
126 E T1N, T4r, T28, T2d, T4s, T1Q, T4E, T29;
|
cannam@127
|
127 {
|
cannam@127
|
128 E T2b, T2c, T1O, T1P;
|
cannam@127
|
129 {
|
cannam@127
|
130 E T2S, T48, T1D, T1L, T1M, T4k, T49, T2T;
|
cannam@127
|
131 T1L = R1[WS(rs, 31)];
|
cannam@127
|
132 T1M = R1[WS(rs, 15)];
|
cannam@127
|
133 T2S = T1g + T1j;
|
cannam@127
|
134 T1k = T1g - T1j;
|
cannam@127
|
135 T48 = T1B - T1C;
|
cannam@127
|
136 T1D = T1B + T1C;
|
cannam@127
|
137 T1N = T1L + T1M;
|
cannam@127
|
138 T4r = T1L - T1M;
|
cannam@127
|
139 T4k = T47 - T48;
|
cannam@127
|
140 T49 = T47 + T48;
|
cannam@127
|
141 T2T = T1G + T1D;
|
cannam@127
|
142 T1H = T1D - T1G;
|
cannam@127
|
143 T5D = FNMS(KP707106781, T4k, T4j);
|
cannam@127
|
144 T4l = FMA(KP707106781, T4k, T4j);
|
cannam@127
|
145 T5A = FNMS(KP707106781, T49, T46);
|
cannam@127
|
146 T4a = FMA(KP707106781, T49, T46);
|
cannam@127
|
147 T3i = T2S - T2T;
|
cannam@127
|
148 T2U = T2S + T2T;
|
cannam@127
|
149 T2b = R1[WS(rs, 3)];
|
cannam@127
|
150 T2c = R1[WS(rs, 19)];
|
cannam@127
|
151 }
|
cannam@127
|
152 T1O = R1[WS(rs, 7)];
|
cannam@127
|
153 T1P = R1[WS(rs, 23)];
|
cannam@127
|
154 T28 = R1[WS(rs, 27)];
|
cannam@127
|
155 T2d = T2b + T2c;
|
cannam@127
|
156 T4s = T2b - T2c;
|
cannam@127
|
157 T1Q = T1O + T1P;
|
cannam@127
|
158 T4E = T1P - T1O;
|
cannam@127
|
159 T29 = R1[WS(rs, 11)];
|
cannam@127
|
160 }
|
cannam@127
|
161 {
|
cannam@127
|
162 E TV, TW, TO, TP;
|
cannam@127
|
163 {
|
cannam@127
|
164 E T2Z, T4t, T2a, TL, TM, T4F, T4u, T30;
|
cannam@127
|
165 TL = R0[WS(rs, 31)];
|
cannam@127
|
166 TM = R0[WS(rs, 15)];
|
cannam@127
|
167 T2Z = T1N + T1Q;
|
cannam@127
|
168 T1R = T1N - T1Q;
|
cannam@127
|
169 T4t = T28 - T29;
|
cannam@127
|
170 T2a = T28 + T29;
|
cannam@127
|
171 TN = TL + TM;
|
cannam@127
|
172 T3V = TL - TM;
|
cannam@127
|
173 T4F = T4t - T4s;
|
cannam@127
|
174 T4u = T4s + T4t;
|
cannam@127
|
175 T30 = T2d + T2a;
|
cannam@127
|
176 T2e = T2a - T2d;
|
cannam@127
|
177 T5K = FNMS(KP707106781, T4F, T4E);
|
cannam@127
|
178 T4G = FMA(KP707106781, T4F, T4E);
|
cannam@127
|
179 T5H = FNMS(KP707106781, T4u, T4r);
|
cannam@127
|
180 T4v = FMA(KP707106781, T4u, T4r);
|
cannam@127
|
181 T3l = T2Z - T30;
|
cannam@127
|
182 T31 = T2Z + T30;
|
cannam@127
|
183 TV = R0[WS(rs, 27)];
|
cannam@127
|
184 TW = R0[WS(rs, 11)];
|
cannam@127
|
185 }
|
cannam@127
|
186 TO = R0[WS(rs, 7)];
|
cannam@127
|
187 TP = R0[WS(rs, 23)];
|
cannam@127
|
188 TS = R0[WS(rs, 3)];
|
cannam@127
|
189 TX = TV + TW;
|
cannam@127
|
190 T3X = TV - TW;
|
cannam@127
|
191 TQ = TO + TP;
|
cannam@127
|
192 T40 = TO - TP;
|
cannam@127
|
193 TT = R0[WS(rs, 19)];
|
cannam@127
|
194 }
|
cannam@127
|
195 }
|
cannam@127
|
196 }
|
cannam@127
|
197 {
|
cannam@127
|
198 E Ti, T3E, Tn, Ts, T3I, Tl, T3F, To;
|
cannam@127
|
199 {
|
cannam@127
|
200 E Ty, T3M, TD, TI, T3O, TB, T3R, TE;
|
cannam@127
|
201 {
|
cannam@127
|
202 E TG, TH, Tz, TA;
|
cannam@127
|
203 {
|
cannam@127
|
204 E T19, TR, T3W, TU, Tw, Tx;
|
cannam@127
|
205 Tw = R0[WS(rs, 1)];
|
cannam@127
|
206 Tx = R0[WS(rs, 17)];
|
cannam@127
|
207 T19 = TN - TQ;
|
cannam@127
|
208 TR = TN + TQ;
|
cannam@127
|
209 T3W = TS - TT;
|
cannam@127
|
210 TU = TS + TT;
|
cannam@127
|
211 Ty = Tw + Tx;
|
cannam@127
|
212 T3M = Tw - Tx;
|
cannam@127
|
213 {
|
cannam@127
|
214 E T41, T3Y, T1a, TY;
|
cannam@127
|
215 T41 = T3W - T3X;
|
cannam@127
|
216 T3Y = T3W + T3X;
|
cannam@127
|
217 T1a = TX - TU;
|
cannam@127
|
218 TY = TU + TX;
|
cannam@127
|
219 T5s = FNMS(KP707106781, T41, T40);
|
cannam@127
|
220 T42 = FMA(KP707106781, T41, T40);
|
cannam@127
|
221 T5t = FNMS(KP707106781, T3Y, T3V);
|
cannam@127
|
222 T3Z = FMA(KP707106781, T3Y, T3V);
|
cannam@127
|
223 T2n = FMA(KP414213562, T19, T1a);
|
cannam@127
|
224 T1b = FNMS(KP414213562, T1a, T19);
|
cannam@127
|
225 T3f = TR - TY;
|
cannam@127
|
226 TZ = TR + TY;
|
cannam@127
|
227 TG = R0[WS(rs, 29)];
|
cannam@127
|
228 TH = R0[WS(rs, 13)];
|
cannam@127
|
229 }
|
cannam@127
|
230 }
|
cannam@127
|
231 Tz = R0[WS(rs, 9)];
|
cannam@127
|
232 TA = R0[WS(rs, 25)];
|
cannam@127
|
233 TD = R0[WS(rs, 5)];
|
cannam@127
|
234 TI = TG + TH;
|
cannam@127
|
235 T3O = TG - TH;
|
cannam@127
|
236 TB = Tz + TA;
|
cannam@127
|
237 T3R = Tz - TA;
|
cannam@127
|
238 TE = R0[WS(rs, 21)];
|
cannam@127
|
239 }
|
cannam@127
|
240 {
|
cannam@127
|
241 E Tq, Tr, Tj, Tk;
|
cannam@127
|
242 {
|
cannam@127
|
243 E T16, TC, T3N, TF, Tg, Th;
|
cannam@127
|
244 Tg = R0[WS(rs, 2)];
|
cannam@127
|
245 Th = R0[WS(rs, 18)];
|
cannam@127
|
246 T16 = Ty - TB;
|
cannam@127
|
247 TC = Ty + TB;
|
cannam@127
|
248 T3N = TD - TE;
|
cannam@127
|
249 TF = TD + TE;
|
cannam@127
|
250 Ti = Tg + Th;
|
cannam@127
|
251 T3E = Tg - Th;
|
cannam@127
|
252 {
|
cannam@127
|
253 E T3S, T3P, T17, TJ;
|
cannam@127
|
254 T3S = T3N - T3O;
|
cannam@127
|
255 T3P = T3N + T3O;
|
cannam@127
|
256 T17 = TI - TF;
|
cannam@127
|
257 TJ = TF + TI;
|
cannam@127
|
258 T5v = FNMS(KP707106781, T3S, T3R);
|
cannam@127
|
259 T3T = FMA(KP707106781, T3S, T3R);
|
cannam@127
|
260 T5w = FNMS(KP707106781, T3P, T3M);
|
cannam@127
|
261 T3Q = FMA(KP707106781, T3P, T3M);
|
cannam@127
|
262 T2m = FNMS(KP414213562, T16, T17);
|
cannam@127
|
263 T18 = FMA(KP414213562, T17, T16);
|
cannam@127
|
264 T3e = TC - TJ;
|
cannam@127
|
265 TK = TC + TJ;
|
cannam@127
|
266 Tq = R0[WS(rs, 6)];
|
cannam@127
|
267 Tr = R0[WS(rs, 22)];
|
cannam@127
|
268 }
|
cannam@127
|
269 }
|
cannam@127
|
270 Tj = R0[WS(rs, 10)];
|
cannam@127
|
271 Tk = R0[WS(rs, 26)];
|
cannam@127
|
272 Tn = R0[WS(rs, 30)];
|
cannam@127
|
273 Ts = Tq + Tr;
|
cannam@127
|
274 T3I = Tq - Tr;
|
cannam@127
|
275 Tl = Tj + Tk;
|
cannam@127
|
276 T3F = Tj - Tk;
|
cannam@127
|
277 To = R0[WS(rs, 14)];
|
cannam@127
|
278 }
|
cannam@127
|
279 }
|
cannam@127
|
280 {
|
cannam@127
|
281 E T1n, T4b, T1s, T4f, T1x, T4c, T1q, T1t;
|
cannam@127
|
282 {
|
cannam@127
|
283 E T1v, T1w, T1o, T1p;
|
cannam@127
|
284 {
|
cannam@127
|
285 E T1l, T4Q, T3G, Tm, T12, Tp, T3H, T1m;
|
cannam@127
|
286 T1l = R1[WS(rs, 2)];
|
cannam@127
|
287 T4Q = FMA(KP414213562, T3E, T3F);
|
cannam@127
|
288 T3G = FNMS(KP414213562, T3F, T3E);
|
cannam@127
|
289 Tm = Ti + Tl;
|
cannam@127
|
290 T12 = Ti - Tl;
|
cannam@127
|
291 Tp = Tn + To;
|
cannam@127
|
292 T3H = Tn - To;
|
cannam@127
|
293 T1m = R1[WS(rs, 18)];
|
cannam@127
|
294 T1v = R1[WS(rs, 6)];
|
cannam@127
|
295 {
|
cannam@127
|
296 E T4R, T3J, Tt, T13;
|
cannam@127
|
297 T4R = FNMS(KP414213562, T3H, T3I);
|
cannam@127
|
298 T3J = FMA(KP414213562, T3I, T3H);
|
cannam@127
|
299 Tt = Tp + Ts;
|
cannam@127
|
300 T13 = Tp - Ts;
|
cannam@127
|
301 T1n = T1l + T1m;
|
cannam@127
|
302 T4b = T1l - T1m;
|
cannam@127
|
303 T3K = T3G + T3J;
|
cannam@127
|
304 T5Q = T3J - T3G;
|
cannam@127
|
305 T4S = T4Q + T4R;
|
cannam@127
|
306 T5q = T4Q - T4R;
|
cannam@127
|
307 T14 = T12 + T13;
|
cannam@127
|
308 T2k = T13 - T12;
|
cannam@127
|
309 T3p = Tt - Tm;
|
cannam@127
|
310 Tu = Tm + Tt;
|
cannam@127
|
311 T1w = R1[WS(rs, 22)];
|
cannam@127
|
312 }
|
cannam@127
|
313 }
|
cannam@127
|
314 T1o = R1[WS(rs, 10)];
|
cannam@127
|
315 T1p = R1[WS(rs, 26)];
|
cannam@127
|
316 T1s = R1[WS(rs, 30)];
|
cannam@127
|
317 T4f = T1v - T1w;
|
cannam@127
|
318 T1x = T1v + T1w;
|
cannam@127
|
319 T4c = T1o - T1p;
|
cannam@127
|
320 T1q = T1o + T1p;
|
cannam@127
|
321 T1t = R1[WS(rs, 14)];
|
cannam@127
|
322 }
|
cannam@127
|
323 {
|
cannam@127
|
324 E T22, T23, T1V, T1W;
|
cannam@127
|
325 {
|
cannam@127
|
326 E T1S, T4d, T4m, T2V, T1r, T4e, T1u, T1T;
|
cannam@127
|
327 T1S = R1[WS(rs, 1)];
|
cannam@127
|
328 T4d = FNMS(KP414213562, T4c, T4b);
|
cannam@127
|
329 T4m = FMA(KP414213562, T4b, T4c);
|
cannam@127
|
330 T2V = T1n + T1q;
|
cannam@127
|
331 T1r = T1n - T1q;
|
cannam@127
|
332 T4e = T1s - T1t;
|
cannam@127
|
333 T1u = T1s + T1t;
|
cannam@127
|
334 T1T = R1[WS(rs, 17)];
|
cannam@127
|
335 T22 = R1[WS(rs, 5)];
|
cannam@127
|
336 {
|
cannam@127
|
337 E T4g, T4n, T2W, T1y;
|
cannam@127
|
338 T4g = FMA(KP414213562, T4f, T4e);
|
cannam@127
|
339 T4n = FNMS(KP414213562, T4e, T4f);
|
cannam@127
|
340 T2W = T1u + T1x;
|
cannam@127
|
341 T1y = T1u - T1x;
|
cannam@127
|
342 T4w = T1S - T1T;
|
cannam@127
|
343 T1U = T1S + T1T;
|
cannam@127
|
344 T5E = T4g - T4d;
|
cannam@127
|
345 T4h = T4d + T4g;
|
cannam@127
|
346 T5B = T4m - T4n;
|
cannam@127
|
347 T4o = T4m + T4n;
|
cannam@127
|
348 T3j = T2W - T2V;
|
cannam@127
|
349 T2X = T2V + T2W;
|
cannam@127
|
350 T1I = T1y - T1r;
|
cannam@127
|
351 T1z = T1r + T1y;
|
cannam@127
|
352 T23 = R1[WS(rs, 21)];
|
cannam@127
|
353 }
|
cannam@127
|
354 }
|
cannam@127
|
355 T1V = R1[WS(rs, 9)];
|
cannam@127
|
356 T1W = R1[WS(rs, 25)];
|
cannam@127
|
357 T1Z = R1[WS(rs, 29)];
|
cannam@127
|
358 T4A = T23 - T22;
|
cannam@127
|
359 T24 = T22 + T23;
|
cannam@127
|
360 T4x = T1W - T1V;
|
cannam@127
|
361 T1X = T1V + T1W;
|
cannam@127
|
362 T20 = R1[WS(rs, 13)];
|
cannam@127
|
363 }
|
cannam@127
|
364 }
|
cannam@127
|
365 }
|
cannam@127
|
366 }
|
cannam@127
|
367 {
|
cannam@127
|
368 E T4C, T5L, T4J, T5I, T26, T2f, T3q, T3h, T3w, T3s, T3o, T3r, T3t;
|
cannam@127
|
369 {
|
cannam@127
|
370 E T2R, T37, T2Y, T3a, T39, T3m, T3b, T35, Tv, T10, T34, T3c, T3x, T3y;
|
cannam@127
|
371 {
|
cannam@127
|
372 E T4y, T4H, T32, T1Y, T4z, T21;
|
cannam@127
|
373 T2R = Tf - Tu;
|
cannam@127
|
374 Tv = Tf + Tu;
|
cannam@127
|
375 T4y = FMA(KP414213562, T4x, T4w);
|
cannam@127
|
376 T4H = FNMS(KP414213562, T4w, T4x);
|
cannam@127
|
377 T32 = T1U + T1X;
|
cannam@127
|
378 T1Y = T1U - T1X;
|
cannam@127
|
379 T4z = T1Z - T20;
|
cannam@127
|
380 T21 = T1Z + T20;
|
cannam@127
|
381 T10 = TK + TZ;
|
cannam@127
|
382 T37 = TZ - TK;
|
cannam@127
|
383 T2Y = T2U - T2X;
|
cannam@127
|
384 T3a = T2U + T2X;
|
cannam@127
|
385 {
|
cannam@127
|
386 E T4B, T4I, T33, T25;
|
cannam@127
|
387 T4B = FNMS(KP414213562, T4A, T4z);
|
cannam@127
|
388 T4I = FMA(KP414213562, T4z, T4A);
|
cannam@127
|
389 T33 = T21 + T24;
|
cannam@127
|
390 T25 = T21 - T24;
|
cannam@127
|
391 T39 = Tv + T10;
|
cannam@127
|
392 T4C = T4y + T4B;
|
cannam@127
|
393 T5L = T4B - T4y;
|
cannam@127
|
394 T4J = T4H + T4I;
|
cannam@127
|
395 T5I = T4I - T4H;
|
cannam@127
|
396 T34 = T32 + T33;
|
cannam@127
|
397 T3m = T33 - T32;
|
cannam@127
|
398 T26 = T1Y + T25;
|
cannam@127
|
399 T2f = T25 - T1Y;
|
cannam@127
|
400 }
|
cannam@127
|
401 }
|
cannam@127
|
402 Cr[WS(csr, 16)] = Tv - T10;
|
cannam@127
|
403 T3b = T31 + T34;
|
cannam@127
|
404 T35 = T31 - T34;
|
cannam@127
|
405 Ci[WS(csi, 16)] = T3b - T3a;
|
cannam@127
|
406 T3c = T3a + T3b;
|
cannam@127
|
407 {
|
cannam@127
|
408 E T3k, T3u, T3v, T3n, T36, T38, T3g;
|
cannam@127
|
409 T3g = T3e + T3f;
|
cannam@127
|
410 T3q = T3f - T3e;
|
cannam@127
|
411 Cr[0] = T39 + T3c;
|
cannam@127
|
412 Cr[WS(csr, 32)] = T39 - T3c;
|
cannam@127
|
413 T36 = T2Y + T35;
|
cannam@127
|
414 T38 = T35 - T2Y;
|
cannam@127
|
415 T3x = FNMS(KP707106781, T3g, T3d);
|
cannam@127
|
416 T3h = FMA(KP707106781, T3g, T3d);
|
cannam@127
|
417 Ci[WS(csi, 8)] = FMA(KP707106781, T38, T37);
|
cannam@127
|
418 Ci[WS(csi, 24)] = FMS(KP707106781, T38, T37);
|
cannam@127
|
419 Cr[WS(csr, 8)] = FMA(KP707106781, T36, T2R);
|
cannam@127
|
420 Cr[WS(csr, 24)] = FNMS(KP707106781, T36, T2R);
|
cannam@127
|
421 T3k = FMA(KP414213562, T3j, T3i);
|
cannam@127
|
422 T3u = FNMS(KP414213562, T3i, T3j);
|
cannam@127
|
423 T3v = FMA(KP414213562, T3l, T3m);
|
cannam@127
|
424 T3n = FNMS(KP414213562, T3m, T3l);
|
cannam@127
|
425 T3y = T3v - T3u;
|
cannam@127
|
426 T3w = T3u + T3v;
|
cannam@127
|
427 T3s = T3n - T3k;
|
cannam@127
|
428 T3o = T3k + T3n;
|
cannam@127
|
429 }
|
cannam@127
|
430 Cr[WS(csr, 12)] = FMA(KP923879532, T3y, T3x);
|
cannam@127
|
431 Cr[WS(csr, 20)] = FNMS(KP923879532, T3y, T3x);
|
cannam@127
|
432 }
|
cannam@127
|
433 Cr[WS(csr, 4)] = FMA(KP923879532, T3o, T3h);
|
cannam@127
|
434 Cr[WS(csr, 28)] = FNMS(KP923879532, T3o, T3h);
|
cannam@127
|
435 T3r = FNMS(KP707106781, T3q, T3p);
|
cannam@127
|
436 T3t = FMA(KP707106781, T3q, T3p);
|
cannam@127
|
437 {
|
cannam@127
|
438 E T27, T2g, T2v, T1d, T2r, T2p, T2s, T1K, T6l, T6m;
|
cannam@127
|
439 {
|
cannam@127
|
440 E T15, T2o, T2P, T2z, T2l, T1c, T1A, T1J, T2D, T2L, T2J, T2M, T2C, T2E, T2N;
|
cannam@127
|
441 E T2F;
|
cannam@127
|
442 {
|
cannam@127
|
443 E T2H, T2I, T2x, T2y, T2A, T2B;
|
cannam@127
|
444 T15 = FMA(KP707106781, T14, T11);
|
cannam@127
|
445 T2x = FNMS(KP707106781, T14, T11);
|
cannam@127
|
446 T2y = T2n - T2m;
|
cannam@127
|
447 T2o = T2m + T2n;
|
cannam@127
|
448 Ci[WS(csi, 4)] = FMA(KP923879532, T3w, T3t);
|
cannam@127
|
449 Ci[WS(csi, 28)] = FMS(KP923879532, T3w, T3t);
|
cannam@127
|
450 Ci[WS(csi, 20)] = FMA(KP923879532, T3s, T3r);
|
cannam@127
|
451 Ci[WS(csi, 12)] = FMS(KP923879532, T3s, T3r);
|
cannam@127
|
452 T2P = FNMS(KP923879532, T2y, T2x);
|
cannam@127
|
453 T2z = FMA(KP923879532, T2y, T2x);
|
cannam@127
|
454 T2l = FMA(KP707106781, T2k, T2j);
|
cannam@127
|
455 T2H = FNMS(KP707106781, T2k, T2j);
|
cannam@127
|
456 T2I = T1b - T18;
|
cannam@127
|
457 T1c = T18 + T1b;
|
cannam@127
|
458 T1A = FMA(KP707106781, T1z, T1k);
|
cannam@127
|
459 T2A = FNMS(KP707106781, T1z, T1k);
|
cannam@127
|
460 T2B = FNMS(KP707106781, T1I, T1H);
|
cannam@127
|
461 T1J = FMA(KP707106781, T1I, T1H);
|
cannam@127
|
462 T27 = FMA(KP707106781, T26, T1R);
|
cannam@127
|
463 T2D = FNMS(KP707106781, T26, T1R);
|
cannam@127
|
464 T2L = FNMS(KP923879532, T2I, T2H);
|
cannam@127
|
465 T2J = FMA(KP923879532, T2I, T2H);
|
cannam@127
|
466 T2M = FMA(KP668178637, T2A, T2B);
|
cannam@127
|
467 T2C = FNMS(KP668178637, T2B, T2A);
|
cannam@127
|
468 T2E = FNMS(KP707106781, T2f, T2e);
|
cannam@127
|
469 T2g = FMA(KP707106781, T2f, T2e);
|
cannam@127
|
470 }
|
cannam@127
|
471 T2N = FNMS(KP668178637, T2D, T2E);
|
cannam@127
|
472 T2F = FMA(KP668178637, T2E, T2D);
|
cannam@127
|
473 T2v = FNMS(KP923879532, T1c, T15);
|
cannam@127
|
474 T1d = FMA(KP923879532, T1c, T15);
|
cannam@127
|
475 {
|
cannam@127
|
476 E T2Q, T2O, T2K, T2G;
|
cannam@127
|
477 T2Q = T2M - T2N;
|
cannam@127
|
478 T2O = T2M + T2N;
|
cannam@127
|
479 T2K = T2F - T2C;
|
cannam@127
|
480 T2G = T2C + T2F;
|
cannam@127
|
481 Cr[WS(csr, 10)] = FMA(KP831469612, T2Q, T2P);
|
cannam@127
|
482 Cr[WS(csr, 22)] = FNMS(KP831469612, T2Q, T2P);
|
cannam@127
|
483 Ci[WS(csi, 26)] = FNMS(KP831469612, T2O, T2L);
|
cannam@127
|
484 Ci[WS(csi, 6)] = -(FMA(KP831469612, T2O, T2L));
|
cannam@127
|
485 Ci[WS(csi, 22)] = FMS(KP831469612, T2K, T2J);
|
cannam@127
|
486 Ci[WS(csi, 10)] = FMA(KP831469612, T2K, T2J);
|
cannam@127
|
487 Cr[WS(csr, 6)] = FMA(KP831469612, T2G, T2z);
|
cannam@127
|
488 Cr[WS(csr, 26)] = FNMS(KP831469612, T2G, T2z);
|
cannam@127
|
489 }
|
cannam@127
|
490 T2r = FMA(KP923879532, T2o, T2l);
|
cannam@127
|
491 T2p = FNMS(KP923879532, T2o, T2l);
|
cannam@127
|
492 T2s = FNMS(KP198912367, T1A, T1J);
|
cannam@127
|
493 T1K = FMA(KP198912367, T1J, T1A);
|
cannam@127
|
494 }
|
cannam@127
|
495 {
|
cannam@127
|
496 E T63, T5r, T5R, T6d, T5J, T5M, T6e, T5y, T6j, T6b, T66, T67, T64, T5U, T5Z;
|
cannam@127
|
497 E T5G;
|
cannam@127
|
498 {
|
cannam@127
|
499 E T5S, T5u, T5x, T5T, T2t, T2h;
|
cannam@127
|
500 T63 = FMA(KP923879532, T5q, T5p);
|
cannam@127
|
501 T5r = FNMS(KP923879532, T5q, T5p);
|
cannam@127
|
502 T5R = FNMS(KP923879532, T5Q, T5P);
|
cannam@127
|
503 T6d = FMA(KP923879532, T5Q, T5P);
|
cannam@127
|
504 T2t = FMA(KP198912367, T27, T2g);
|
cannam@127
|
505 T2h = FNMS(KP198912367, T2g, T27);
|
cannam@127
|
506 T5S = FNMS(KP668178637, T5s, T5t);
|
cannam@127
|
507 T5u = FMA(KP668178637, T5t, T5s);
|
cannam@127
|
508 {
|
cannam@127
|
509 E T2w, T2u, T2q, T2i;
|
cannam@127
|
510 T2w = T2t - T2s;
|
cannam@127
|
511 T2u = T2s + T2t;
|
cannam@127
|
512 T2q = T2h - T1K;
|
cannam@127
|
513 T2i = T1K + T2h;
|
cannam@127
|
514 Cr[WS(csr, 14)] = FMA(KP980785280, T2w, T2v);
|
cannam@127
|
515 Cr[WS(csr, 18)] = FNMS(KP980785280, T2w, T2v);
|
cannam@127
|
516 Ci[WS(csi, 30)] = FMS(KP980785280, T2u, T2r);
|
cannam@127
|
517 Ci[WS(csi, 2)] = FMA(KP980785280, T2u, T2r);
|
cannam@127
|
518 Ci[WS(csi, 18)] = FMA(KP980785280, T2q, T2p);
|
cannam@127
|
519 Ci[WS(csi, 14)] = FMS(KP980785280, T2q, T2p);
|
cannam@127
|
520 Cr[WS(csr, 2)] = FMA(KP980785280, T2i, T1d);
|
cannam@127
|
521 Cr[WS(csr, 30)] = FNMS(KP980785280, T2i, T1d);
|
cannam@127
|
522 T5x = FNMS(KP668178637, T5w, T5v);
|
cannam@127
|
523 T5T = FMA(KP668178637, T5v, T5w);
|
cannam@127
|
524 }
|
cannam@127
|
525 {
|
cannam@127
|
526 E T69, T6a, T5C, T5F;
|
cannam@127
|
527 T5J = FNMS(KP923879532, T5I, T5H);
|
cannam@127
|
528 T69 = FMA(KP923879532, T5I, T5H);
|
cannam@127
|
529 T6a = FNMS(KP923879532, T5L, T5K);
|
cannam@127
|
530 T5M = FMA(KP923879532, T5L, T5K);
|
cannam@127
|
531 T6e = T5x + T5u;
|
cannam@127
|
532 T5y = T5u - T5x;
|
cannam@127
|
533 T6j = FNMS(KP303346683, T69, T6a);
|
cannam@127
|
534 T6b = FMA(KP303346683, T6a, T69);
|
cannam@127
|
535 T66 = FMA(KP923879532, T5B, T5A);
|
cannam@127
|
536 T5C = FNMS(KP923879532, T5B, T5A);
|
cannam@127
|
537 T5F = FNMS(KP923879532, T5E, T5D);
|
cannam@127
|
538 T67 = FMA(KP923879532, T5E, T5D);
|
cannam@127
|
539 T64 = T5T + T5S;
|
cannam@127
|
540 T5U = T5S - T5T;
|
cannam@127
|
541 T5Z = FMA(KP534511135, T5C, T5F);
|
cannam@127
|
542 T5G = FNMS(KP534511135, T5F, T5C);
|
cannam@127
|
543 }
|
cannam@127
|
544 }
|
cannam@127
|
545 {
|
cannam@127
|
546 E T61, T6i, T68, T62;
|
cannam@127
|
547 {
|
cannam@127
|
548 E T5z, T5Y, T5N, T5X, T5V, T60, T5W, T5O;
|
cannam@127
|
549 T61 = FNMS(KP831469612, T5y, T5r);
|
cannam@127
|
550 T5z = FMA(KP831469612, T5y, T5r);
|
cannam@127
|
551 T6i = FNMS(KP303346683, T66, T67);
|
cannam@127
|
552 T68 = FMA(KP303346683, T67, T66);
|
cannam@127
|
553 T5Y = FMA(KP534511135, T5J, T5M);
|
cannam@127
|
554 T5N = FNMS(KP534511135, T5M, T5J);
|
cannam@127
|
555 T5X = FNMS(KP831469612, T5U, T5R);
|
cannam@127
|
556 T5V = FMA(KP831469612, T5U, T5R);
|
cannam@127
|
557 T60 = T5Y - T5Z;
|
cannam@127
|
558 T62 = T5Z + T5Y;
|
cannam@127
|
559 T5W = T5N - T5G;
|
cannam@127
|
560 T5O = T5G + T5N;
|
cannam@127
|
561 Ci[WS(csi, 27)] = FMA(KP881921264, T60, T5X);
|
cannam@127
|
562 Ci[WS(csi, 5)] = FMS(KP881921264, T60, T5X);
|
cannam@127
|
563 Cr[WS(csr, 5)] = FMA(KP881921264, T5O, T5z);
|
cannam@127
|
564 Cr[WS(csr, 27)] = FNMS(KP881921264, T5O, T5z);
|
cannam@127
|
565 Ci[WS(csi, 21)] = FMS(KP881921264, T5W, T5V);
|
cannam@127
|
566 Ci[WS(csi, 11)] = FMA(KP881921264, T5W, T5V);
|
cannam@127
|
567 }
|
cannam@127
|
568 {
|
cannam@127
|
569 E T6g, T6f, T6h, T6k, T65, T6c;
|
cannam@127
|
570 T6l = FNMS(KP831469612, T64, T63);
|
cannam@127
|
571 T65 = FMA(KP831469612, T64, T63);
|
cannam@127
|
572 T6c = T68 + T6b;
|
cannam@127
|
573 T6g = T6b - T68;
|
cannam@127
|
574 T6f = FNMS(KP831469612, T6e, T6d);
|
cannam@127
|
575 T6h = FMA(KP831469612, T6e, T6d);
|
cannam@127
|
576 Cr[WS(csr, 11)] = FMA(KP881921264, T62, T61);
|
cannam@127
|
577 Cr[WS(csr, 21)] = FNMS(KP881921264, T62, T61);
|
cannam@127
|
578 Cr[WS(csr, 3)] = FMA(KP956940335, T6c, T65);
|
cannam@127
|
579 Cr[WS(csr, 29)] = FNMS(KP956940335, T6c, T65);
|
cannam@127
|
580 T6k = T6i - T6j;
|
cannam@127
|
581 T6m = T6i + T6j;
|
cannam@127
|
582 Ci[WS(csi, 29)] = FMS(KP956940335, T6k, T6h);
|
cannam@127
|
583 Ci[WS(csi, 3)] = FMA(KP956940335, T6k, T6h);
|
cannam@127
|
584 Ci[WS(csi, 19)] = FMA(KP956940335, T6g, T6f);
|
cannam@127
|
585 Ci[WS(csi, 13)] = FMS(KP956940335, T6g, T6f);
|
cannam@127
|
586 }
|
cannam@127
|
587 }
|
cannam@127
|
588 }
|
cannam@127
|
589 {
|
cannam@127
|
590 E T55, T3L, T4T, T5f, T4D, T4K, T5g, T44, T5l, T5d, T58, T59, T56, T4W, T51;
|
cannam@127
|
591 E T4q;
|
cannam@127
|
592 {
|
cannam@127
|
593 E T4U, T3U, T43, T4V;
|
cannam@127
|
594 T55 = FNMS(KP923879532, T3K, T3D);
|
cannam@127
|
595 T3L = FMA(KP923879532, T3K, T3D);
|
cannam@127
|
596 T4T = FMA(KP923879532, T4S, T4P);
|
cannam@127
|
597 T5f = FNMS(KP923879532, T4S, T4P);
|
cannam@127
|
598 Cr[WS(csr, 13)] = FNMS(KP956940335, T6m, T6l);
|
cannam@127
|
599 Cr[WS(csr, 19)] = FMA(KP956940335, T6m, T6l);
|
cannam@127
|
600 T4U = FMA(KP198912367, T3Q, T3T);
|
cannam@127
|
601 T3U = FNMS(KP198912367, T3T, T3Q);
|
cannam@127
|
602 T43 = FMA(KP198912367, T42, T3Z);
|
cannam@127
|
603 T4V = FNMS(KP198912367, T3Z, T42);
|
cannam@127
|
604 {
|
cannam@127
|
605 E T5b, T5c, T4i, T4p;
|
cannam@127
|
606 T4D = FMA(KP923879532, T4C, T4v);
|
cannam@127
|
607 T5b = FNMS(KP923879532, T4C, T4v);
|
cannam@127
|
608 T5c = FNMS(KP923879532, T4J, T4G);
|
cannam@127
|
609 T4K = FMA(KP923879532, T4J, T4G);
|
cannam@127
|
610 T5g = T43 - T3U;
|
cannam@127
|
611 T44 = T3U + T43;
|
cannam@127
|
612 T5l = FNMS(KP820678790, T5b, T5c);
|
cannam@127
|
613 T5d = FMA(KP820678790, T5c, T5b);
|
cannam@127
|
614 T58 = FNMS(KP923879532, T4h, T4a);
|
cannam@127
|
615 T4i = FMA(KP923879532, T4h, T4a);
|
cannam@127
|
616 T4p = FMA(KP923879532, T4o, T4l);
|
cannam@127
|
617 T59 = FNMS(KP923879532, T4o, T4l);
|
cannam@127
|
618 T56 = T4U - T4V;
|
cannam@127
|
619 T4W = T4U + T4V;
|
cannam@127
|
620 T51 = FMA(KP098491403, T4i, T4p);
|
cannam@127
|
621 T4q = FNMS(KP098491403, T4p, T4i);
|
cannam@127
|
622 }
|
cannam@127
|
623 }
|
cannam@127
|
624 {
|
cannam@127
|
625 E T53, T5k, T5a, T54;
|
cannam@127
|
626 {
|
cannam@127
|
627 E T45, T50, T4L, T4Z, T4X, T52, T4Y, T4M;
|
cannam@127
|
628 T53 = FNMS(KP980785280, T44, T3L);
|
cannam@127
|
629 T45 = FMA(KP980785280, T44, T3L);
|
cannam@127
|
630 T5k = FNMS(KP820678790, T58, T59);
|
cannam@127
|
631 T5a = FMA(KP820678790, T59, T58);
|
cannam@127
|
632 T50 = FMA(KP098491403, T4D, T4K);
|
cannam@127
|
633 T4L = FNMS(KP098491403, T4K, T4D);
|
cannam@127
|
634 T4Z = FMA(KP980785280, T4W, T4T);
|
cannam@127
|
635 T4X = FNMS(KP980785280, T4W, T4T);
|
cannam@127
|
636 T52 = T50 - T51;
|
cannam@127
|
637 T54 = T51 + T50;
|
cannam@127
|
638 T4Y = T4L - T4q;
|
cannam@127
|
639 T4M = T4q + T4L;
|
cannam@127
|
640 Ci[WS(csi, 31)] = FMA(KP995184726, T52, T4Z);
|
cannam@127
|
641 Ci[WS(csi, 1)] = FMS(KP995184726, T52, T4Z);
|
cannam@127
|
642 Cr[WS(csr, 1)] = FMA(KP995184726, T4M, T45);
|
cannam@127
|
643 Cr[WS(csr, 31)] = FNMS(KP995184726, T4M, T45);
|
cannam@127
|
644 Ci[WS(csi, 17)] = FMS(KP995184726, T4Y, T4X);
|
cannam@127
|
645 Ci[WS(csi, 15)] = FMA(KP995184726, T4Y, T4X);
|
cannam@127
|
646 }
|
cannam@127
|
647 {
|
cannam@127
|
648 E T5i, T5h, T5j, T5m, T57, T5e;
|
cannam@127
|
649 T5n = FNMS(KP980785280, T56, T55);
|
cannam@127
|
650 T57 = FMA(KP980785280, T56, T55);
|
cannam@127
|
651 T5e = T5a + T5d;
|
cannam@127
|
652 T5i = T5d - T5a;
|
cannam@127
|
653 T5h = FNMS(KP980785280, T5g, T5f);
|
cannam@127
|
654 T5j = FMA(KP980785280, T5g, T5f);
|
cannam@127
|
655 Cr[WS(csr, 15)] = FMA(KP995184726, T54, T53);
|
cannam@127
|
656 Cr[WS(csr, 17)] = FNMS(KP995184726, T54, T53);
|
cannam@127
|
657 Cr[WS(csr, 7)] = FMA(KP773010453, T5e, T57);
|
cannam@127
|
658 Cr[WS(csr, 25)] = FNMS(KP773010453, T5e, T57);
|
cannam@127
|
659 T5m = T5k - T5l;
|
cannam@127
|
660 T5o = T5k + T5l;
|
cannam@127
|
661 Ci[WS(csi, 25)] = FMS(KP773010453, T5m, T5j);
|
cannam@127
|
662 Ci[WS(csi, 7)] = FMA(KP773010453, T5m, T5j);
|
cannam@127
|
663 Ci[WS(csi, 23)] = FMA(KP773010453, T5i, T5h);
|
cannam@127
|
664 Ci[WS(csi, 9)] = FMS(KP773010453, T5i, T5h);
|
cannam@127
|
665 }
|
cannam@127
|
666 }
|
cannam@127
|
667 }
|
cannam@127
|
668 }
|
cannam@127
|
669 }
|
cannam@127
|
670 }
|
cannam@127
|
671 Cr[WS(csr, 9)] = FNMS(KP773010453, T5o, T5n);
|
cannam@127
|
672 Cr[WS(csr, 23)] = FMA(KP773010453, T5o, T5n);
|
cannam@127
|
673 }
|
cannam@127
|
674 }
|
cannam@127
|
675 }
|
cannam@127
|
676
|
cannam@127
|
677 static const kr2c_desc desc = { 64, "r2cf_64", {198, 0, 196, 0}, &GENUS };
|
cannam@127
|
678
|
cannam@127
|
679 void X(codelet_r2cf_64) (planner *p) {
|
cannam@127
|
680 X(kr2c_register) (p, r2cf_64, &desc);
|
cannam@127
|
681 }
|
cannam@127
|
682
|
cannam@127
|
683 #else /* HAVE_FMA */
|
cannam@127
|
684
|
cannam@127
|
685 /* Generated by: ../../../genfft/gen_r2cf.native -compact -variables 4 -pipeline-latency 4 -n 64 -name r2cf_64 -include r2cf.h */
|
cannam@127
|
686
|
cannam@127
|
687 /*
|
cannam@127
|
688 * This function contains 394 FP additions, 124 FP multiplications,
|
cannam@127
|
689 * (or, 342 additions, 72 multiplications, 52 fused multiply/add),
|
cannam@127
|
690 * 106 stack variables, 15 constants, and 128 memory accesses
|
cannam@127
|
691 */
|
cannam@127
|
692 #include "r2cf.h"
|
cannam@127
|
693
|
cannam@127
|
694 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@127
|
695 {
|
cannam@127
|
696 DK(KP773010453, +0.773010453362736960810906609758469800971041293);
|
cannam@127
|
697 DK(KP634393284, +0.634393284163645498215171613225493370675687095);
|
cannam@127
|
698 DK(KP098017140, +0.098017140329560601994195563888641845861136673);
|
cannam@127
|
699 DK(KP995184726, +0.995184726672196886244836953109479921575474869);
|
cannam@127
|
700 DK(KP290284677, +0.290284677254462367636192375817395274691476278);
|
cannam@127
|
701 DK(KP956940335, +0.956940335732208864935797886980269969482849206);
|
cannam@127
|
702 DK(KP471396736, +0.471396736825997648556387625905254377657460319);
|
cannam@127
|
703 DK(KP881921264, +0.881921264348355029712756863660388349508442621);
|
cannam@127
|
704 DK(KP195090322, +0.195090322016128267848284868477022240927691618);
|
cannam@127
|
705 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
cannam@127
|
706 DK(KP555570233, +0.555570233019602224742830813948532874374937191);
|
cannam@127
|
707 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
cannam@127
|
708 DK(KP382683432, +0.382683432365089771728459984030398866761344562);
|
cannam@127
|
709 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
cannam@127
|
710 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
cannam@127
|
711 {
|
cannam@127
|
712 INT i;
|
cannam@127
|
713 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@127
|
714 E T4l, T5a, T15, T3n, T2T, T3Q, T7, Te, Tf, T4A, T4L, T1X, T3B, T23, T3y;
|
cannam@127
|
715 E T5I, T66, T4R, T52, T2j, T3F, T2H, T3I, T5P, T69, T1i, T3t, T1l, T3u, TZ;
|
cannam@127
|
716 E T63, T4v, T58, T1r, T3r, T1u, T3q, TK, T62, T4s, T57, Tm, Tt, Tu, T4o;
|
cannam@127
|
717 E T5b, T1c, T3R, T2Q, T3o, T1M, T3z, T5L, T67, T26, T3C, T4H, T4M, T2y, T3J;
|
cannam@127
|
718 E T5S, T6a, T2C, T3G, T4Y, T53;
|
cannam@127
|
719 {
|
cannam@127
|
720 E T3, T11, Td, T13, T6, T2S, Ta, T12, T14, T2R;
|
cannam@127
|
721 {
|
cannam@127
|
722 E T1, T2, Tb, Tc;
|
cannam@127
|
723 T1 = R0[0];
|
cannam@127
|
724 T2 = R0[WS(rs, 16)];
|
cannam@127
|
725 T3 = T1 + T2;
|
cannam@127
|
726 T11 = T1 - T2;
|
cannam@127
|
727 Tb = R0[WS(rs, 28)];
|
cannam@127
|
728 Tc = R0[WS(rs, 12)];
|
cannam@127
|
729 Td = Tb + Tc;
|
cannam@127
|
730 T13 = Tb - Tc;
|
cannam@127
|
731 }
|
cannam@127
|
732 {
|
cannam@127
|
733 E T4, T5, T8, T9;
|
cannam@127
|
734 T4 = R0[WS(rs, 8)];
|
cannam@127
|
735 T5 = R0[WS(rs, 24)];
|
cannam@127
|
736 T6 = T4 + T5;
|
cannam@127
|
737 T2S = T4 - T5;
|
cannam@127
|
738 T8 = R0[WS(rs, 4)];
|
cannam@127
|
739 T9 = R0[WS(rs, 20)];
|
cannam@127
|
740 Ta = T8 + T9;
|
cannam@127
|
741 T12 = T8 - T9;
|
cannam@127
|
742 }
|
cannam@127
|
743 T4l = T3 - T6;
|
cannam@127
|
744 T5a = Td - Ta;
|
cannam@127
|
745 T14 = KP707106781 * (T12 + T13);
|
cannam@127
|
746 T15 = T11 + T14;
|
cannam@127
|
747 T3n = T11 - T14;
|
cannam@127
|
748 T2R = KP707106781 * (T13 - T12);
|
cannam@127
|
749 T2T = T2R - T2S;
|
cannam@127
|
750 T3Q = T2S + T2R;
|
cannam@127
|
751 T7 = T3 + T6;
|
cannam@127
|
752 Te = Ta + Td;
|
cannam@127
|
753 Tf = T7 + Te;
|
cannam@127
|
754 }
|
cannam@127
|
755 {
|
cannam@127
|
756 E T1P, T4J, T21, T4y, T1S, T4K, T1W, T4z;
|
cannam@127
|
757 {
|
cannam@127
|
758 E T1N, T1O, T1Z, T20;
|
cannam@127
|
759 T1N = R1[WS(rs, 28)];
|
cannam@127
|
760 T1O = R1[WS(rs, 12)];
|
cannam@127
|
761 T1P = T1N - T1O;
|
cannam@127
|
762 T4J = T1N + T1O;
|
cannam@127
|
763 T1Z = R1[0];
|
cannam@127
|
764 T20 = R1[WS(rs, 16)];
|
cannam@127
|
765 T21 = T1Z - T20;
|
cannam@127
|
766 T4y = T1Z + T20;
|
cannam@127
|
767 }
|
cannam@127
|
768 {
|
cannam@127
|
769 E T1Q, T1R, T1U, T1V;
|
cannam@127
|
770 T1Q = R1[WS(rs, 4)];
|
cannam@127
|
771 T1R = R1[WS(rs, 20)];
|
cannam@127
|
772 T1S = T1Q - T1R;
|
cannam@127
|
773 T4K = T1Q + T1R;
|
cannam@127
|
774 T1U = R1[WS(rs, 8)];
|
cannam@127
|
775 T1V = R1[WS(rs, 24)];
|
cannam@127
|
776 T1W = T1U - T1V;
|
cannam@127
|
777 T4z = T1U + T1V;
|
cannam@127
|
778 }
|
cannam@127
|
779 T4A = T4y - T4z;
|
cannam@127
|
780 T4L = T4J - T4K;
|
cannam@127
|
781 {
|
cannam@127
|
782 E T1T, T22, T5G, T5H;
|
cannam@127
|
783 T1T = KP707106781 * (T1P - T1S);
|
cannam@127
|
784 T1X = T1T - T1W;
|
cannam@127
|
785 T3B = T1W + T1T;
|
cannam@127
|
786 T22 = KP707106781 * (T1S + T1P);
|
cannam@127
|
787 T23 = T21 + T22;
|
cannam@127
|
788 T3y = T21 - T22;
|
cannam@127
|
789 T5G = T4y + T4z;
|
cannam@127
|
790 T5H = T4K + T4J;
|
cannam@127
|
791 T5I = T5G + T5H;
|
cannam@127
|
792 T66 = T5G - T5H;
|
cannam@127
|
793 }
|
cannam@127
|
794 }
|
cannam@127
|
795 {
|
cannam@127
|
796 E T2b, T4P, T2G, T4Q, T2e, T51, T2h, T50;
|
cannam@127
|
797 {
|
cannam@127
|
798 E T29, T2a, T2E, T2F;
|
cannam@127
|
799 T29 = R1[WS(rs, 31)];
|
cannam@127
|
800 T2a = R1[WS(rs, 15)];
|
cannam@127
|
801 T2b = T29 - T2a;
|
cannam@127
|
802 T4P = T29 + T2a;
|
cannam@127
|
803 T2E = R1[WS(rs, 7)];
|
cannam@127
|
804 T2F = R1[WS(rs, 23)];
|
cannam@127
|
805 T2G = T2E - T2F;
|
cannam@127
|
806 T4Q = T2E + T2F;
|
cannam@127
|
807 }
|
cannam@127
|
808 {
|
cannam@127
|
809 E T2c, T2d, T2f, T2g;
|
cannam@127
|
810 T2c = R1[WS(rs, 3)];
|
cannam@127
|
811 T2d = R1[WS(rs, 19)];
|
cannam@127
|
812 T2e = T2c - T2d;
|
cannam@127
|
813 T51 = T2c + T2d;
|
cannam@127
|
814 T2f = R1[WS(rs, 27)];
|
cannam@127
|
815 T2g = R1[WS(rs, 11)];
|
cannam@127
|
816 T2h = T2f - T2g;
|
cannam@127
|
817 T50 = T2f + T2g;
|
cannam@127
|
818 }
|
cannam@127
|
819 T4R = T4P - T4Q;
|
cannam@127
|
820 T52 = T50 - T51;
|
cannam@127
|
821 {
|
cannam@127
|
822 E T2i, T2D, T5N, T5O;
|
cannam@127
|
823 T2i = KP707106781 * (T2e + T2h);
|
cannam@127
|
824 T2j = T2b + T2i;
|
cannam@127
|
825 T3F = T2b - T2i;
|
cannam@127
|
826 T2D = KP707106781 * (T2h - T2e);
|
cannam@127
|
827 T2H = T2D - T2G;
|
cannam@127
|
828 T3I = T2G + T2D;
|
cannam@127
|
829 T5N = T4P + T4Q;
|
cannam@127
|
830 T5O = T51 + T50;
|
cannam@127
|
831 T5P = T5N + T5O;
|
cannam@127
|
832 T69 = T5N - T5O;
|
cannam@127
|
833 }
|
cannam@127
|
834 }
|
cannam@127
|
835 {
|
cannam@127
|
836 E TN, T1e, TX, T1g, TQ, T1k, TU, T1f, T1h, T1j;
|
cannam@127
|
837 {
|
cannam@127
|
838 E TL, TM, TV, TW;
|
cannam@127
|
839 TL = R0[WS(rs, 31)];
|
cannam@127
|
840 TM = R0[WS(rs, 15)];
|
cannam@127
|
841 TN = TL + TM;
|
cannam@127
|
842 T1e = TL - TM;
|
cannam@127
|
843 TV = R0[WS(rs, 27)];
|
cannam@127
|
844 TW = R0[WS(rs, 11)];
|
cannam@127
|
845 TX = TV + TW;
|
cannam@127
|
846 T1g = TV - TW;
|
cannam@127
|
847 }
|
cannam@127
|
848 {
|
cannam@127
|
849 E TO, TP, TS, TT;
|
cannam@127
|
850 TO = R0[WS(rs, 7)];
|
cannam@127
|
851 TP = R0[WS(rs, 23)];
|
cannam@127
|
852 TQ = TO + TP;
|
cannam@127
|
853 T1k = TO - TP;
|
cannam@127
|
854 TS = R0[WS(rs, 3)];
|
cannam@127
|
855 TT = R0[WS(rs, 19)];
|
cannam@127
|
856 TU = TS + TT;
|
cannam@127
|
857 T1f = TS - TT;
|
cannam@127
|
858 }
|
cannam@127
|
859 T1h = KP707106781 * (T1f + T1g);
|
cannam@127
|
860 T1i = T1e + T1h;
|
cannam@127
|
861 T3t = T1e - T1h;
|
cannam@127
|
862 T1j = KP707106781 * (T1g - T1f);
|
cannam@127
|
863 T1l = T1j - T1k;
|
cannam@127
|
864 T3u = T1k + T1j;
|
cannam@127
|
865 {
|
cannam@127
|
866 E TR, TY, T4t, T4u;
|
cannam@127
|
867 TR = TN + TQ;
|
cannam@127
|
868 TY = TU + TX;
|
cannam@127
|
869 TZ = TR + TY;
|
cannam@127
|
870 T63 = TR - TY;
|
cannam@127
|
871 T4t = TN - TQ;
|
cannam@127
|
872 T4u = TX - TU;
|
cannam@127
|
873 T4v = FNMS(KP382683432, T4u, KP923879532 * T4t);
|
cannam@127
|
874 T58 = FMA(KP382683432, T4t, KP923879532 * T4u);
|
cannam@127
|
875 }
|
cannam@127
|
876 }
|
cannam@127
|
877 {
|
cannam@127
|
878 E Ty, T1s, TI, T1n, TB, T1q, TF, T1o, T1p, T1t;
|
cannam@127
|
879 {
|
cannam@127
|
880 E Tw, Tx, TG, TH;
|
cannam@127
|
881 Tw = R0[WS(rs, 1)];
|
cannam@127
|
882 Tx = R0[WS(rs, 17)];
|
cannam@127
|
883 Ty = Tw + Tx;
|
cannam@127
|
884 T1s = Tw - Tx;
|
cannam@127
|
885 TG = R0[WS(rs, 29)];
|
cannam@127
|
886 TH = R0[WS(rs, 13)];
|
cannam@127
|
887 TI = TG + TH;
|
cannam@127
|
888 T1n = TG - TH;
|
cannam@127
|
889 }
|
cannam@127
|
890 {
|
cannam@127
|
891 E Tz, TA, TD, TE;
|
cannam@127
|
892 Tz = R0[WS(rs, 9)];
|
cannam@127
|
893 TA = R0[WS(rs, 25)];
|
cannam@127
|
894 TB = Tz + TA;
|
cannam@127
|
895 T1q = Tz - TA;
|
cannam@127
|
896 TD = R0[WS(rs, 5)];
|
cannam@127
|
897 TE = R0[WS(rs, 21)];
|
cannam@127
|
898 TF = TD + TE;
|
cannam@127
|
899 T1o = TD - TE;
|
cannam@127
|
900 }
|
cannam@127
|
901 T1p = KP707106781 * (T1n - T1o);
|
cannam@127
|
902 T1r = T1p - T1q;
|
cannam@127
|
903 T3r = T1q + T1p;
|
cannam@127
|
904 T1t = KP707106781 * (T1o + T1n);
|
cannam@127
|
905 T1u = T1s + T1t;
|
cannam@127
|
906 T3q = T1s - T1t;
|
cannam@127
|
907 {
|
cannam@127
|
908 E TC, TJ, T4q, T4r;
|
cannam@127
|
909 TC = Ty + TB;
|
cannam@127
|
910 TJ = TF + TI;
|
cannam@127
|
911 TK = TC + TJ;
|
cannam@127
|
912 T62 = TC - TJ;
|
cannam@127
|
913 T4q = Ty - TB;
|
cannam@127
|
914 T4r = TI - TF;
|
cannam@127
|
915 T4s = FMA(KP923879532, T4q, KP382683432 * T4r);
|
cannam@127
|
916 T57 = FNMS(KP382683432, T4q, KP923879532 * T4r);
|
cannam@127
|
917 }
|
cannam@127
|
918 }
|
cannam@127
|
919 {
|
cannam@127
|
920 E Ti, T16, Ts, T1a, Tl, T17, Tp, T19, T4m, T4n;
|
cannam@127
|
921 {
|
cannam@127
|
922 E Tg, Th, Tq, Tr;
|
cannam@127
|
923 Tg = R0[WS(rs, 2)];
|
cannam@127
|
924 Th = R0[WS(rs, 18)];
|
cannam@127
|
925 Ti = Tg + Th;
|
cannam@127
|
926 T16 = Tg - Th;
|
cannam@127
|
927 Tq = R0[WS(rs, 6)];
|
cannam@127
|
928 Tr = R0[WS(rs, 22)];
|
cannam@127
|
929 Ts = Tq + Tr;
|
cannam@127
|
930 T1a = Tq - Tr;
|
cannam@127
|
931 }
|
cannam@127
|
932 {
|
cannam@127
|
933 E Tj, Tk, Tn, To;
|
cannam@127
|
934 Tj = R0[WS(rs, 10)];
|
cannam@127
|
935 Tk = R0[WS(rs, 26)];
|
cannam@127
|
936 Tl = Tj + Tk;
|
cannam@127
|
937 T17 = Tj - Tk;
|
cannam@127
|
938 Tn = R0[WS(rs, 30)];
|
cannam@127
|
939 To = R0[WS(rs, 14)];
|
cannam@127
|
940 Tp = Tn + To;
|
cannam@127
|
941 T19 = Tn - To;
|
cannam@127
|
942 }
|
cannam@127
|
943 Tm = Ti + Tl;
|
cannam@127
|
944 Tt = Tp + Ts;
|
cannam@127
|
945 Tu = Tm + Tt;
|
cannam@127
|
946 T4m = Ti - Tl;
|
cannam@127
|
947 T4n = Tp - Ts;
|
cannam@127
|
948 T4o = KP707106781 * (T4m + T4n);
|
cannam@127
|
949 T5b = KP707106781 * (T4n - T4m);
|
cannam@127
|
950 {
|
cannam@127
|
951 E T18, T1b, T2O, T2P;
|
cannam@127
|
952 T18 = FNMS(KP382683432, T17, KP923879532 * T16);
|
cannam@127
|
953 T1b = FMA(KP923879532, T19, KP382683432 * T1a);
|
cannam@127
|
954 T1c = T18 + T1b;
|
cannam@127
|
955 T3R = T1b - T18;
|
cannam@127
|
956 T2O = FNMS(KP923879532, T1a, KP382683432 * T19);
|
cannam@127
|
957 T2P = FMA(KP382683432, T16, KP923879532 * T17);
|
cannam@127
|
958 T2Q = T2O - T2P;
|
cannam@127
|
959 T3o = T2P + T2O;
|
cannam@127
|
960 }
|
cannam@127
|
961 }
|
cannam@127
|
962 {
|
cannam@127
|
963 E T1A, T4E, T1K, T4C, T1D, T4F, T1H, T4B;
|
cannam@127
|
964 {
|
cannam@127
|
965 E T1y, T1z, T1I, T1J;
|
cannam@127
|
966 T1y = R1[WS(rs, 30)];
|
cannam@127
|
967 T1z = R1[WS(rs, 14)];
|
cannam@127
|
968 T1A = T1y - T1z;
|
cannam@127
|
969 T4E = T1y + T1z;
|
cannam@127
|
970 T1I = R1[WS(rs, 10)];
|
cannam@127
|
971 T1J = R1[WS(rs, 26)];
|
cannam@127
|
972 T1K = T1I - T1J;
|
cannam@127
|
973 T4C = T1I + T1J;
|
cannam@127
|
974 }
|
cannam@127
|
975 {
|
cannam@127
|
976 E T1B, T1C, T1F, T1G;
|
cannam@127
|
977 T1B = R1[WS(rs, 6)];
|
cannam@127
|
978 T1C = R1[WS(rs, 22)];
|
cannam@127
|
979 T1D = T1B - T1C;
|
cannam@127
|
980 T4F = T1B + T1C;
|
cannam@127
|
981 T1F = R1[WS(rs, 2)];
|
cannam@127
|
982 T1G = R1[WS(rs, 18)];
|
cannam@127
|
983 T1H = T1F - T1G;
|
cannam@127
|
984 T4B = T1F + T1G;
|
cannam@127
|
985 }
|
cannam@127
|
986 {
|
cannam@127
|
987 E T1E, T1L, T5J, T5K;
|
cannam@127
|
988 T1E = FNMS(KP923879532, T1D, KP382683432 * T1A);
|
cannam@127
|
989 T1L = FMA(KP382683432, T1H, KP923879532 * T1K);
|
cannam@127
|
990 T1M = T1E - T1L;
|
cannam@127
|
991 T3z = T1L + T1E;
|
cannam@127
|
992 T5J = T4B + T4C;
|
cannam@127
|
993 T5K = T4E + T4F;
|
cannam@127
|
994 T5L = T5J + T5K;
|
cannam@127
|
995 T67 = T5K - T5J;
|
cannam@127
|
996 }
|
cannam@127
|
997 {
|
cannam@127
|
998 E T24, T25, T4D, T4G;
|
cannam@127
|
999 T24 = FNMS(KP382683432, T1K, KP923879532 * T1H);
|
cannam@127
|
1000 T25 = FMA(KP923879532, T1A, KP382683432 * T1D);
|
cannam@127
|
1001 T26 = T24 + T25;
|
cannam@127
|
1002 T3C = T25 - T24;
|
cannam@127
|
1003 T4D = T4B - T4C;
|
cannam@127
|
1004 T4G = T4E - T4F;
|
cannam@127
|
1005 T4H = KP707106781 * (T4D + T4G);
|
cannam@127
|
1006 T4M = KP707106781 * (T4G - T4D);
|
cannam@127
|
1007 }
|
cannam@127
|
1008 }
|
cannam@127
|
1009 {
|
cannam@127
|
1010 E T2m, T4S, T2w, T4W, T2p, T4T, T2t, T4V;
|
cannam@127
|
1011 {
|
cannam@127
|
1012 E T2k, T2l, T2u, T2v;
|
cannam@127
|
1013 T2k = R1[WS(rs, 1)];
|
cannam@127
|
1014 T2l = R1[WS(rs, 17)];
|
cannam@127
|
1015 T2m = T2k - T2l;
|
cannam@127
|
1016 T4S = T2k + T2l;
|
cannam@127
|
1017 T2u = R1[WS(rs, 5)];
|
cannam@127
|
1018 T2v = R1[WS(rs, 21)];
|
cannam@127
|
1019 T2w = T2u - T2v;
|
cannam@127
|
1020 T4W = T2u + T2v;
|
cannam@127
|
1021 }
|
cannam@127
|
1022 {
|
cannam@127
|
1023 E T2n, T2o, T2r, T2s;
|
cannam@127
|
1024 T2n = R1[WS(rs, 9)];
|
cannam@127
|
1025 T2o = R1[WS(rs, 25)];
|
cannam@127
|
1026 T2p = T2n - T2o;
|
cannam@127
|
1027 T4T = T2n + T2o;
|
cannam@127
|
1028 T2r = R1[WS(rs, 29)];
|
cannam@127
|
1029 T2s = R1[WS(rs, 13)];
|
cannam@127
|
1030 T2t = T2r - T2s;
|
cannam@127
|
1031 T4V = T2r + T2s;
|
cannam@127
|
1032 }
|
cannam@127
|
1033 {
|
cannam@127
|
1034 E T2q, T2x, T5Q, T5R;
|
cannam@127
|
1035 T2q = FNMS(KP382683432, T2p, KP923879532 * T2m);
|
cannam@127
|
1036 T2x = FMA(KP923879532, T2t, KP382683432 * T2w);
|
cannam@127
|
1037 T2y = T2q + T2x;
|
cannam@127
|
1038 T3J = T2x - T2q;
|
cannam@127
|
1039 T5Q = T4S + T4T;
|
cannam@127
|
1040 T5R = T4V + T4W;
|
cannam@127
|
1041 T5S = T5Q + T5R;
|
cannam@127
|
1042 T6a = T5R - T5Q;
|
cannam@127
|
1043 }
|
cannam@127
|
1044 {
|
cannam@127
|
1045 E T2A, T2B, T4U, T4X;
|
cannam@127
|
1046 T2A = FNMS(KP923879532, T2w, KP382683432 * T2t);
|
cannam@127
|
1047 T2B = FMA(KP382683432, T2m, KP923879532 * T2p);
|
cannam@127
|
1048 T2C = T2A - T2B;
|
cannam@127
|
1049 T3G = T2B + T2A;
|
cannam@127
|
1050 T4U = T4S - T4T;
|
cannam@127
|
1051 T4X = T4V - T4W;
|
cannam@127
|
1052 T4Y = KP707106781 * (T4U + T4X);
|
cannam@127
|
1053 T53 = KP707106781 * (T4X - T4U);
|
cannam@127
|
1054 }
|
cannam@127
|
1055 }
|
cannam@127
|
1056 {
|
cannam@127
|
1057 E Tv, T10, T5X, T5Y, T5Z, T60;
|
cannam@127
|
1058 Tv = Tf + Tu;
|
cannam@127
|
1059 T10 = TK + TZ;
|
cannam@127
|
1060 T5X = Tv + T10;
|
cannam@127
|
1061 T5Y = T5I + T5L;
|
cannam@127
|
1062 T5Z = T5P + T5S;
|
cannam@127
|
1063 T60 = T5Y + T5Z;
|
cannam@127
|
1064 Cr[WS(csr, 16)] = Tv - T10;
|
cannam@127
|
1065 Ci[WS(csi, 16)] = T5Z - T5Y;
|
cannam@127
|
1066 Cr[WS(csr, 32)] = T5X - T60;
|
cannam@127
|
1067 Cr[0] = T5X + T60;
|
cannam@127
|
1068 }
|
cannam@127
|
1069 {
|
cannam@127
|
1070 E T5F, T5V, T5U, T5W, T5M, T5T;
|
cannam@127
|
1071 T5F = Tf - Tu;
|
cannam@127
|
1072 T5V = TZ - TK;
|
cannam@127
|
1073 T5M = T5I - T5L;
|
cannam@127
|
1074 T5T = T5P - T5S;
|
cannam@127
|
1075 T5U = KP707106781 * (T5M + T5T);
|
cannam@127
|
1076 T5W = KP707106781 * (T5T - T5M);
|
cannam@127
|
1077 Cr[WS(csr, 24)] = T5F - T5U;
|
cannam@127
|
1078 Ci[WS(csi, 24)] = T5W - T5V;
|
cannam@127
|
1079 Cr[WS(csr, 8)] = T5F + T5U;
|
cannam@127
|
1080 Ci[WS(csi, 8)] = T5V + T5W;
|
cannam@127
|
1081 }
|
cannam@127
|
1082 {
|
cannam@127
|
1083 E T65, T6l, T6k, T6m, T6c, T6g, T6f, T6h;
|
cannam@127
|
1084 {
|
cannam@127
|
1085 E T61, T64, T6i, T6j;
|
cannam@127
|
1086 T61 = T7 - Te;
|
cannam@127
|
1087 T64 = KP707106781 * (T62 + T63);
|
cannam@127
|
1088 T65 = T61 + T64;
|
cannam@127
|
1089 T6l = T61 - T64;
|
cannam@127
|
1090 T6i = FNMS(KP382683432, T66, KP923879532 * T67);
|
cannam@127
|
1091 T6j = FMA(KP382683432, T69, KP923879532 * T6a);
|
cannam@127
|
1092 T6k = T6i + T6j;
|
cannam@127
|
1093 T6m = T6j - T6i;
|
cannam@127
|
1094 }
|
cannam@127
|
1095 {
|
cannam@127
|
1096 E T68, T6b, T6d, T6e;
|
cannam@127
|
1097 T68 = FMA(KP923879532, T66, KP382683432 * T67);
|
cannam@127
|
1098 T6b = FNMS(KP382683432, T6a, KP923879532 * T69);
|
cannam@127
|
1099 T6c = T68 + T6b;
|
cannam@127
|
1100 T6g = T6b - T68;
|
cannam@127
|
1101 T6d = KP707106781 * (T63 - T62);
|
cannam@127
|
1102 T6e = Tt - Tm;
|
cannam@127
|
1103 T6f = T6d - T6e;
|
cannam@127
|
1104 T6h = T6e + T6d;
|
cannam@127
|
1105 }
|
cannam@127
|
1106 Cr[WS(csr, 28)] = T65 - T6c;
|
cannam@127
|
1107 Ci[WS(csi, 28)] = T6k - T6h;
|
cannam@127
|
1108 Cr[WS(csr, 4)] = T65 + T6c;
|
cannam@127
|
1109 Ci[WS(csi, 4)] = T6h + T6k;
|
cannam@127
|
1110 Ci[WS(csi, 12)] = T6f + T6g;
|
cannam@127
|
1111 Cr[WS(csr, 12)] = T6l + T6m;
|
cannam@127
|
1112 Ci[WS(csi, 20)] = T6g - T6f;
|
cannam@127
|
1113 Cr[WS(csr, 20)] = T6l - T6m;
|
cannam@127
|
1114 }
|
cannam@127
|
1115 {
|
cannam@127
|
1116 E T5n, T5D, T5x, T5z, T5q, T5A, T5t, T5B;
|
cannam@127
|
1117 {
|
cannam@127
|
1118 E T5l, T5m, T5v, T5w;
|
cannam@127
|
1119 T5l = T4l - T4o;
|
cannam@127
|
1120 T5m = T58 - T57;
|
cannam@127
|
1121 T5n = T5l + T5m;
|
cannam@127
|
1122 T5D = T5l - T5m;
|
cannam@127
|
1123 T5v = T4v - T4s;
|
cannam@127
|
1124 T5w = T5b - T5a;
|
cannam@127
|
1125 T5x = T5v - T5w;
|
cannam@127
|
1126 T5z = T5w + T5v;
|
cannam@127
|
1127 }
|
cannam@127
|
1128 {
|
cannam@127
|
1129 E T5o, T5p, T5r, T5s;
|
cannam@127
|
1130 T5o = T4A - T4H;
|
cannam@127
|
1131 T5p = T4M - T4L;
|
cannam@127
|
1132 T5q = FMA(KP831469612, T5o, KP555570233 * T5p);
|
cannam@127
|
1133 T5A = FNMS(KP555570233, T5o, KP831469612 * T5p);
|
cannam@127
|
1134 T5r = T4R - T4Y;
|
cannam@127
|
1135 T5s = T53 - T52;
|
cannam@127
|
1136 T5t = FNMS(KP555570233, T5s, KP831469612 * T5r);
|
cannam@127
|
1137 T5B = FMA(KP555570233, T5r, KP831469612 * T5s);
|
cannam@127
|
1138 }
|
cannam@127
|
1139 {
|
cannam@127
|
1140 E T5u, T5C, T5y, T5E;
|
cannam@127
|
1141 T5u = T5q + T5t;
|
cannam@127
|
1142 Cr[WS(csr, 26)] = T5n - T5u;
|
cannam@127
|
1143 Cr[WS(csr, 6)] = T5n + T5u;
|
cannam@127
|
1144 T5C = T5A + T5B;
|
cannam@127
|
1145 Ci[WS(csi, 6)] = T5z + T5C;
|
cannam@127
|
1146 Ci[WS(csi, 26)] = T5C - T5z;
|
cannam@127
|
1147 T5y = T5t - T5q;
|
cannam@127
|
1148 Ci[WS(csi, 10)] = T5x + T5y;
|
cannam@127
|
1149 Ci[WS(csi, 22)] = T5y - T5x;
|
cannam@127
|
1150 T5E = T5B - T5A;
|
cannam@127
|
1151 Cr[WS(csr, 22)] = T5D - T5E;
|
cannam@127
|
1152 Cr[WS(csr, 10)] = T5D + T5E;
|
cannam@127
|
1153 }
|
cannam@127
|
1154 }
|
cannam@127
|
1155 {
|
cannam@127
|
1156 E T4x, T5j, T5d, T5f, T4O, T5g, T55, T5h;
|
cannam@127
|
1157 {
|
cannam@127
|
1158 E T4p, T4w, T59, T5c;
|
cannam@127
|
1159 T4p = T4l + T4o;
|
cannam@127
|
1160 T4w = T4s + T4v;
|
cannam@127
|
1161 T4x = T4p + T4w;
|
cannam@127
|
1162 T5j = T4p - T4w;
|
cannam@127
|
1163 T59 = T57 + T58;
|
cannam@127
|
1164 T5c = T5a + T5b;
|
cannam@127
|
1165 T5d = T59 - T5c;
|
cannam@127
|
1166 T5f = T5c + T59;
|
cannam@127
|
1167 }
|
cannam@127
|
1168 {
|
cannam@127
|
1169 E T4I, T4N, T4Z, T54;
|
cannam@127
|
1170 T4I = T4A + T4H;
|
cannam@127
|
1171 T4N = T4L + T4M;
|
cannam@127
|
1172 T4O = FMA(KP980785280, T4I, KP195090322 * T4N);
|
cannam@127
|
1173 T5g = FNMS(KP195090322, T4I, KP980785280 * T4N);
|
cannam@127
|
1174 T4Z = T4R + T4Y;
|
cannam@127
|
1175 T54 = T52 + T53;
|
cannam@127
|
1176 T55 = FNMS(KP195090322, T54, KP980785280 * T4Z);
|
cannam@127
|
1177 T5h = FMA(KP195090322, T4Z, KP980785280 * T54);
|
cannam@127
|
1178 }
|
cannam@127
|
1179 {
|
cannam@127
|
1180 E T56, T5i, T5e, T5k;
|
cannam@127
|
1181 T56 = T4O + T55;
|
cannam@127
|
1182 Cr[WS(csr, 30)] = T4x - T56;
|
cannam@127
|
1183 Cr[WS(csr, 2)] = T4x + T56;
|
cannam@127
|
1184 T5i = T5g + T5h;
|
cannam@127
|
1185 Ci[WS(csi, 2)] = T5f + T5i;
|
cannam@127
|
1186 Ci[WS(csi, 30)] = T5i - T5f;
|
cannam@127
|
1187 T5e = T55 - T4O;
|
cannam@127
|
1188 Ci[WS(csi, 14)] = T5d + T5e;
|
cannam@127
|
1189 Ci[WS(csi, 18)] = T5e - T5d;
|
cannam@127
|
1190 T5k = T5h - T5g;
|
cannam@127
|
1191 Cr[WS(csr, 18)] = T5j - T5k;
|
cannam@127
|
1192 Cr[WS(csr, 14)] = T5j + T5k;
|
cannam@127
|
1193 }
|
cannam@127
|
1194 }
|
cannam@127
|
1195 {
|
cannam@127
|
1196 E T3p, T41, T4c, T3S, T3w, T4b, T49, T4h, T3P, T42, T3E, T3W, T46, T4g, T3L;
|
cannam@127
|
1197 E T3X;
|
cannam@127
|
1198 {
|
cannam@127
|
1199 E T3s, T3v, T3A, T3D;
|
cannam@127
|
1200 T3p = T3n + T3o;
|
cannam@127
|
1201 T41 = T3n - T3o;
|
cannam@127
|
1202 T4c = T3R - T3Q;
|
cannam@127
|
1203 T3S = T3Q + T3R;
|
cannam@127
|
1204 T3s = FMA(KP831469612, T3q, KP555570233 * T3r);
|
cannam@127
|
1205 T3v = FNMS(KP555570233, T3u, KP831469612 * T3t);
|
cannam@127
|
1206 T3w = T3s + T3v;
|
cannam@127
|
1207 T4b = T3v - T3s;
|
cannam@127
|
1208 {
|
cannam@127
|
1209 E T47, T48, T3N, T3O;
|
cannam@127
|
1210 T47 = T3F - T3G;
|
cannam@127
|
1211 T48 = T3J - T3I;
|
cannam@127
|
1212 T49 = FNMS(KP471396736, T48, KP881921264 * T47);
|
cannam@127
|
1213 T4h = FMA(KP471396736, T47, KP881921264 * T48);
|
cannam@127
|
1214 T3N = FNMS(KP555570233, T3q, KP831469612 * T3r);
|
cannam@127
|
1215 T3O = FMA(KP555570233, T3t, KP831469612 * T3u);
|
cannam@127
|
1216 T3P = T3N + T3O;
|
cannam@127
|
1217 T42 = T3O - T3N;
|
cannam@127
|
1218 }
|
cannam@127
|
1219 T3A = T3y + T3z;
|
cannam@127
|
1220 T3D = T3B + T3C;
|
cannam@127
|
1221 T3E = FMA(KP956940335, T3A, KP290284677 * T3D);
|
cannam@127
|
1222 T3W = FNMS(KP290284677, T3A, KP956940335 * T3D);
|
cannam@127
|
1223 {
|
cannam@127
|
1224 E T44, T45, T3H, T3K;
|
cannam@127
|
1225 T44 = T3y - T3z;
|
cannam@127
|
1226 T45 = T3C - T3B;
|
cannam@127
|
1227 T46 = FMA(KP881921264, T44, KP471396736 * T45);
|
cannam@127
|
1228 T4g = FNMS(KP471396736, T44, KP881921264 * T45);
|
cannam@127
|
1229 T3H = T3F + T3G;
|
cannam@127
|
1230 T3K = T3I + T3J;
|
cannam@127
|
1231 T3L = FNMS(KP290284677, T3K, KP956940335 * T3H);
|
cannam@127
|
1232 T3X = FMA(KP290284677, T3H, KP956940335 * T3K);
|
cannam@127
|
1233 }
|
cannam@127
|
1234 }
|
cannam@127
|
1235 {
|
cannam@127
|
1236 E T3x, T3M, T3V, T3Y;
|
cannam@127
|
1237 T3x = T3p + T3w;
|
cannam@127
|
1238 T3M = T3E + T3L;
|
cannam@127
|
1239 Cr[WS(csr, 29)] = T3x - T3M;
|
cannam@127
|
1240 Cr[WS(csr, 3)] = T3x + T3M;
|
cannam@127
|
1241 T3V = T3S + T3P;
|
cannam@127
|
1242 T3Y = T3W + T3X;
|
cannam@127
|
1243 Ci[WS(csi, 3)] = T3V + T3Y;
|
cannam@127
|
1244 Ci[WS(csi, 29)] = T3Y - T3V;
|
cannam@127
|
1245 }
|
cannam@127
|
1246 {
|
cannam@127
|
1247 E T3T, T3U, T3Z, T40;
|
cannam@127
|
1248 T3T = T3P - T3S;
|
cannam@127
|
1249 T3U = T3L - T3E;
|
cannam@127
|
1250 Ci[WS(csi, 13)] = T3T + T3U;
|
cannam@127
|
1251 Ci[WS(csi, 19)] = T3U - T3T;
|
cannam@127
|
1252 T3Z = T3p - T3w;
|
cannam@127
|
1253 T40 = T3X - T3W;
|
cannam@127
|
1254 Cr[WS(csr, 19)] = T3Z - T40;
|
cannam@127
|
1255 Cr[WS(csr, 13)] = T3Z + T40;
|
cannam@127
|
1256 }
|
cannam@127
|
1257 {
|
cannam@127
|
1258 E T43, T4a, T4f, T4i;
|
cannam@127
|
1259 T43 = T41 + T42;
|
cannam@127
|
1260 T4a = T46 + T49;
|
cannam@127
|
1261 Cr[WS(csr, 27)] = T43 - T4a;
|
cannam@127
|
1262 Cr[WS(csr, 5)] = T43 + T4a;
|
cannam@127
|
1263 T4f = T4c + T4b;
|
cannam@127
|
1264 T4i = T4g + T4h;
|
cannam@127
|
1265 Ci[WS(csi, 5)] = T4f + T4i;
|
cannam@127
|
1266 Ci[WS(csi, 27)] = T4i - T4f;
|
cannam@127
|
1267 }
|
cannam@127
|
1268 {
|
cannam@127
|
1269 E T4d, T4e, T4j, T4k;
|
cannam@127
|
1270 T4d = T4b - T4c;
|
cannam@127
|
1271 T4e = T49 - T46;
|
cannam@127
|
1272 Ci[WS(csi, 11)] = T4d + T4e;
|
cannam@127
|
1273 Ci[WS(csi, 21)] = T4e - T4d;
|
cannam@127
|
1274 T4j = T41 - T42;
|
cannam@127
|
1275 T4k = T4h - T4g;
|
cannam@127
|
1276 Cr[WS(csr, 21)] = T4j - T4k;
|
cannam@127
|
1277 Cr[WS(csr, 11)] = T4j + T4k;
|
cannam@127
|
1278 }
|
cannam@127
|
1279 }
|
cannam@127
|
1280 {
|
cannam@127
|
1281 E T1d, T33, T3e, T2U, T1w, T3d, T3b, T3j, T2N, T34, T28, T2Y, T38, T3i, T2J;
|
cannam@127
|
1282 E T2Z;
|
cannam@127
|
1283 {
|
cannam@127
|
1284 E T1m, T1v, T1Y, T27;
|
cannam@127
|
1285 T1d = T15 - T1c;
|
cannam@127
|
1286 T33 = T15 + T1c;
|
cannam@127
|
1287 T3e = T2T + T2Q;
|
cannam@127
|
1288 T2U = T2Q - T2T;
|
cannam@127
|
1289 T1m = FMA(KP195090322, T1i, KP980785280 * T1l);
|
cannam@127
|
1290 T1v = FNMS(KP195090322, T1u, KP980785280 * T1r);
|
cannam@127
|
1291 T1w = T1m - T1v;
|
cannam@127
|
1292 T3d = T1v + T1m;
|
cannam@127
|
1293 {
|
cannam@127
|
1294 E T39, T3a, T2L, T2M;
|
cannam@127
|
1295 T39 = T2j + T2y;
|
cannam@127
|
1296 T3a = T2H + T2C;
|
cannam@127
|
1297 T3b = FNMS(KP098017140, T3a, KP995184726 * T39);
|
cannam@127
|
1298 T3j = FMA(KP995184726, T3a, KP098017140 * T39);
|
cannam@127
|
1299 T2L = FNMS(KP195090322, T1l, KP980785280 * T1i);
|
cannam@127
|
1300 T2M = FMA(KP980785280, T1u, KP195090322 * T1r);
|
cannam@127
|
1301 T2N = T2L - T2M;
|
cannam@127
|
1302 T34 = T2M + T2L;
|
cannam@127
|
1303 }
|
cannam@127
|
1304 T1Y = T1M - T1X;
|
cannam@127
|
1305 T27 = T23 - T26;
|
cannam@127
|
1306 T28 = FMA(KP634393284, T1Y, KP773010453 * T27);
|
cannam@127
|
1307 T2Y = FNMS(KP634393284, T27, KP773010453 * T1Y);
|
cannam@127
|
1308 {
|
cannam@127
|
1309 E T36, T37, T2z, T2I;
|
cannam@127
|
1310 T36 = T1X + T1M;
|
cannam@127
|
1311 T37 = T23 + T26;
|
cannam@127
|
1312 T38 = FMA(KP098017140, T36, KP995184726 * T37);
|
cannam@127
|
1313 T3i = FNMS(KP098017140, T37, KP995184726 * T36);
|
cannam@127
|
1314 T2z = T2j - T2y;
|
cannam@127
|
1315 T2I = T2C - T2H;
|
cannam@127
|
1316 T2J = FNMS(KP634393284, T2I, KP773010453 * T2z);
|
cannam@127
|
1317 T2Z = FMA(KP773010453, T2I, KP634393284 * T2z);
|
cannam@127
|
1318 }
|
cannam@127
|
1319 }
|
cannam@127
|
1320 {
|
cannam@127
|
1321 E T1x, T2K, T2X, T30;
|
cannam@127
|
1322 T1x = T1d + T1w;
|
cannam@127
|
1323 T2K = T28 + T2J;
|
cannam@127
|
1324 Cr[WS(csr, 25)] = T1x - T2K;
|
cannam@127
|
1325 Cr[WS(csr, 7)] = T1x + T2K;
|
cannam@127
|
1326 T2X = T2U + T2N;
|
cannam@127
|
1327 T30 = T2Y + T2Z;
|
cannam@127
|
1328 Ci[WS(csi, 7)] = T2X + T30;
|
cannam@127
|
1329 Ci[WS(csi, 25)] = T30 - T2X;
|
cannam@127
|
1330 }
|
cannam@127
|
1331 {
|
cannam@127
|
1332 E T2V, T2W, T31, T32;
|
cannam@127
|
1333 T2V = T2N - T2U;
|
cannam@127
|
1334 T2W = T2J - T28;
|
cannam@127
|
1335 Ci[WS(csi, 9)] = T2V + T2W;
|
cannam@127
|
1336 Ci[WS(csi, 23)] = T2W - T2V;
|
cannam@127
|
1337 T31 = T1d - T1w;
|
cannam@127
|
1338 T32 = T2Z - T2Y;
|
cannam@127
|
1339 Cr[WS(csr, 23)] = T31 - T32;
|
cannam@127
|
1340 Cr[WS(csr, 9)] = T31 + T32;
|
cannam@127
|
1341 }
|
cannam@127
|
1342 {
|
cannam@127
|
1343 E T35, T3c, T3h, T3k;
|
cannam@127
|
1344 T35 = T33 + T34;
|
cannam@127
|
1345 T3c = T38 + T3b;
|
cannam@127
|
1346 Cr[WS(csr, 31)] = T35 - T3c;
|
cannam@127
|
1347 Cr[WS(csr, 1)] = T35 + T3c;
|
cannam@127
|
1348 T3h = T3e + T3d;
|
cannam@127
|
1349 T3k = T3i + T3j;
|
cannam@127
|
1350 Ci[WS(csi, 1)] = T3h + T3k;
|
cannam@127
|
1351 Ci[WS(csi, 31)] = T3k - T3h;
|
cannam@127
|
1352 }
|
cannam@127
|
1353 {
|
cannam@127
|
1354 E T3f, T3g, T3l, T3m;
|
cannam@127
|
1355 T3f = T3d - T3e;
|
cannam@127
|
1356 T3g = T3b - T38;
|
cannam@127
|
1357 Ci[WS(csi, 15)] = T3f + T3g;
|
cannam@127
|
1358 Ci[WS(csi, 17)] = T3g - T3f;
|
cannam@127
|
1359 T3l = T33 - T34;
|
cannam@127
|
1360 T3m = T3j - T3i;
|
cannam@127
|
1361 Cr[WS(csr, 17)] = T3l - T3m;
|
cannam@127
|
1362 Cr[WS(csr, 15)] = T3l + T3m;
|
cannam@127
|
1363 }
|
cannam@127
|
1364 }
|
cannam@127
|
1365 }
|
cannam@127
|
1366 }
|
cannam@127
|
1367 }
|
cannam@127
|
1368
|
cannam@127
|
1369 static const kr2c_desc desc = { 64, "r2cf_64", {342, 72, 52, 0}, &GENUS };
|
cannam@127
|
1370
|
cannam@127
|
1371 void X(codelet_r2cf_64) (planner *p) {
|
cannam@127
|
1372 X(kr2c_register) (p, r2cf_64, &desc);
|
cannam@127
|
1373 }
|
cannam@127
|
1374
|
cannam@127
|
1375 #endif /* HAVE_FMA */
|