Chris@10
|
1 /*
|
Chris@10
|
2 * Copyright (c) 2003, 2007-11 Matteo Frigo
|
Chris@10
|
3 * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology
|
Chris@10
|
4 *
|
Chris@10
|
5 * This program is free software; you can redistribute it and/or modify
|
Chris@10
|
6 * it under the terms of the GNU General Public License as published by
|
Chris@10
|
7 * the Free Software Foundation; either version 2 of the License, or
|
Chris@10
|
8 * (at your option) any later version.
|
Chris@10
|
9 *
|
Chris@10
|
10 * This program is distributed in the hope that it will be useful,
|
Chris@10
|
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
Chris@10
|
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
Chris@10
|
13 * GNU General Public License for more details.
|
Chris@10
|
14 *
|
Chris@10
|
15 * You should have received a copy of the GNU General Public License
|
Chris@10
|
16 * along with this program; if not, write to the Free Software
|
Chris@10
|
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
Chris@10
|
18 *
|
Chris@10
|
19 */
|
Chris@10
|
20
|
Chris@10
|
21 /* This file was automatically generated --- DO NOT EDIT */
|
Chris@10
|
22 /* Generated on Sun Nov 25 07:40:16 EST 2012 */
|
Chris@10
|
23
|
Chris@10
|
24 #include "codelet-rdft.h"
|
Chris@10
|
25
|
Chris@10
|
26 #ifdef HAVE_FMA
|
Chris@10
|
27
|
Chris@10
|
28 /* Generated by: ../../../genfft/gen_r2cf.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -n 32 -name r2cfII_32 -dft-II -include r2cfII.h */
|
Chris@10
|
29
|
Chris@10
|
30 /*
|
Chris@10
|
31 * This function contains 174 FP additions, 128 FP multiplications,
|
Chris@10
|
32 * (or, 46 additions, 0 multiplications, 128 fused multiply/add),
|
Chris@10
|
33 * 96 stack variables, 15 constants, and 64 memory accesses
|
Chris@10
|
34 */
|
Chris@10
|
35 #include "r2cfII.h"
|
Chris@10
|
36
|
Chris@10
|
37 static void r2cfII_32(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
|
Chris@10
|
38 {
|
Chris@10
|
39 DK(KP773010453, +0.773010453362736960810906609758469800971041293);
|
Chris@10
|
40 DK(KP820678790, +0.820678790828660330972281985331011598767386482);
|
Chris@10
|
41 DK(KP956940335, +0.956940335732208864935797886980269969482849206);
|
Chris@10
|
42 DK(KP303346683, +0.303346683607342391675883946941299872384187453);
|
Chris@10
|
43 DK(KP995184726, +0.995184726672196886244836953109479921575474869);
|
Chris@10
|
44 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
Chris@10
|
45 DK(KP098491403, +0.098491403357164253077197521291327432293052451);
|
Chris@10
|
46 DK(KP881921264, +0.881921264348355029712756863660388349508442621);
|
Chris@10
|
47 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
Chris@10
|
48 DK(KP534511135, +0.534511135950791641089685961295362908582039528);
|
Chris@10
|
49 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
Chris@10
|
50 DK(KP668178637, +0.668178637919298919997757686523080761552472251);
|
Chris@10
|
51 DK(KP198912367, +0.198912367379658006911597622644676228597850501);
|
Chris@10
|
52 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
Chris@10
|
53 DK(KP414213562, +0.414213562373095048801688724209698078569671875);
|
Chris@10
|
54 {
|
Chris@10
|
55 INT i;
|
Chris@10
|
56 for (i = v; i > 0; i = i - 1, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(128, rs), MAKE_VOLATILE_STRIDE(128, csr), MAKE_VOLATILE_STRIDE(128, csi)) {
|
Chris@10
|
57 E T23, T1S, T21, T1L, T2z, T2x, T1Z, T22;
|
Chris@10
|
58 {
|
Chris@10
|
59 E T2n, T2B, T1z, T5, T1C, T2C, T2o, Tc, T27, T1J, T1l, Tm, T26, T1G, T1k;
|
Chris@10
|
60 E Tv, T1s, T1c, T2e, T1Y, T1r, T15, T2d, T1V, TP, TF, T1M, TC, T1P, TN;
|
Chris@10
|
61 E TO, TI;
|
Chris@10
|
62 {
|
Chris@10
|
63 E T1A, T8, Te, Tj, Tf, T1B, Tb, Tg;
|
Chris@10
|
64 {
|
Chris@10
|
65 E T1, T2l, T2, T3, T9, Ta;
|
Chris@10
|
66 T1 = R0[0];
|
Chris@10
|
67 T2l = R0[WS(rs, 8)];
|
Chris@10
|
68 T2 = R0[WS(rs, 4)];
|
Chris@10
|
69 T3 = R0[WS(rs, 12)];
|
Chris@10
|
70 {
|
Chris@10
|
71 E T6, T7, T2m, T4;
|
Chris@10
|
72 T6 = R0[WS(rs, 10)];
|
Chris@10
|
73 T7 = R0[WS(rs, 2)];
|
Chris@10
|
74 T9 = R0[WS(rs, 6)];
|
Chris@10
|
75 T2m = T2 + T3;
|
Chris@10
|
76 T4 = T2 - T3;
|
Chris@10
|
77 T1A = FNMS(KP414213562, T6, T7);
|
Chris@10
|
78 T8 = FMA(KP414213562, T7, T6);
|
Chris@10
|
79 T2n = FMA(KP707106781, T2m, T2l);
|
Chris@10
|
80 T2B = FNMS(KP707106781, T2m, T2l);
|
Chris@10
|
81 T1z = FMA(KP707106781, T4, T1);
|
Chris@10
|
82 T5 = FNMS(KP707106781, T4, T1);
|
Chris@10
|
83 Ta = R0[WS(rs, 14)];
|
Chris@10
|
84 }
|
Chris@10
|
85 Te = R0[WS(rs, 7)];
|
Chris@10
|
86 Tj = R0[WS(rs, 15)];
|
Chris@10
|
87 Tf = R0[WS(rs, 3)];
|
Chris@10
|
88 T1B = FMS(KP414213562, T9, Ta);
|
Chris@10
|
89 Tb = FMA(KP414213562, Ta, T9);
|
Chris@10
|
90 Tg = R0[WS(rs, 11)];
|
Chris@10
|
91 }
|
Chris@10
|
92 {
|
Chris@10
|
93 E Tn, Ts, To, T1I, Tl, T1H, Ti, Tp, Tk, Th, T1T, T1U;
|
Chris@10
|
94 Tn = R0[WS(rs, 9)];
|
Chris@10
|
95 T1C = T1A + T1B;
|
Chris@10
|
96 T2C = T1B - T1A;
|
Chris@10
|
97 T2o = T8 + Tb;
|
Chris@10
|
98 Tc = T8 - Tb;
|
Chris@10
|
99 Tk = Tg - Tf;
|
Chris@10
|
100 Th = Tf + Tg;
|
Chris@10
|
101 Ts = R0[WS(rs, 1)];
|
Chris@10
|
102 To = R0[WS(rs, 5)];
|
Chris@10
|
103 T1I = FMA(KP707106781, Tk, Tj);
|
Chris@10
|
104 Tl = FNMS(KP707106781, Tk, Tj);
|
Chris@10
|
105 T1H = FMA(KP707106781, Th, Te);
|
Chris@10
|
106 Ti = FNMS(KP707106781, Th, Te);
|
Chris@10
|
107 Tp = R0[WS(rs, 13)];
|
Chris@10
|
108 {
|
Chris@10
|
109 E TT, T16, TY, T17, TW, TZ, T11, T12, Tt, Tq;
|
Chris@10
|
110 TT = R1[WS(rs, 15)];
|
Chris@10
|
111 T27 = FNMS(KP198912367, T1H, T1I);
|
Chris@10
|
112 T1J = FMA(KP198912367, T1I, T1H);
|
Chris@10
|
113 T1l = FMA(KP668178637, Ti, Tl);
|
Chris@10
|
114 Tm = FNMS(KP668178637, Tl, Ti);
|
Chris@10
|
115 Tt = To - Tp;
|
Chris@10
|
116 Tq = To + Tp;
|
Chris@10
|
117 T16 = R1[WS(rs, 7)];
|
Chris@10
|
118 {
|
Chris@10
|
119 E TU, T1F, Tu, T1E, Tr, TV;
|
Chris@10
|
120 TU = R1[WS(rs, 3)];
|
Chris@10
|
121 T1F = FMA(KP707106781, Tt, Ts);
|
Chris@10
|
122 Tu = FNMS(KP707106781, Tt, Ts);
|
Chris@10
|
123 T1E = FMA(KP707106781, Tq, Tn);
|
Chris@10
|
124 Tr = FNMS(KP707106781, Tq, Tn);
|
Chris@10
|
125 TV = R1[WS(rs, 11)];
|
Chris@10
|
126 TY = R1[WS(rs, 9)];
|
Chris@10
|
127 T26 = FNMS(KP198912367, T1E, T1F);
|
Chris@10
|
128 T1G = FMA(KP198912367, T1F, T1E);
|
Chris@10
|
129 T1k = FMA(KP668178637, Tr, Tu);
|
Chris@10
|
130 Tv = FNMS(KP668178637, Tu, Tr);
|
Chris@10
|
131 T17 = TU + TV;
|
Chris@10
|
132 TW = TU - TV;
|
Chris@10
|
133 TZ = R1[WS(rs, 1)];
|
Chris@10
|
134 T11 = R1[WS(rs, 5)];
|
Chris@10
|
135 T12 = R1[WS(rs, 13)];
|
Chris@10
|
136 }
|
Chris@10
|
137 {
|
Chris@10
|
138 E TX, T1a, T10, T19, T13, T1W, T18, T1b, T14, T1X;
|
Chris@10
|
139 T1T = FMS(KP707106781, TW, TT);
|
Chris@10
|
140 TX = FMA(KP707106781, TW, TT);
|
Chris@10
|
141 T1a = FNMS(KP414213562, TY, TZ);
|
Chris@10
|
142 T10 = FMA(KP414213562, TZ, TY);
|
Chris@10
|
143 T19 = FMS(KP414213562, T11, T12);
|
Chris@10
|
144 T13 = FMA(KP414213562, T12, T11);
|
Chris@10
|
145 T1W = FMA(KP707106781, T17, T16);
|
Chris@10
|
146 T18 = FNMS(KP707106781, T17, T16);
|
Chris@10
|
147 T1b = T19 - T1a;
|
Chris@10
|
148 T1U = T1a + T19;
|
Chris@10
|
149 T14 = T10 - T13;
|
Chris@10
|
150 T1X = T10 + T13;
|
Chris@10
|
151 T1s = FMA(KP923879532, T1b, T18);
|
Chris@10
|
152 T1c = FNMS(KP923879532, T1b, T18);
|
Chris@10
|
153 T2e = FMA(KP923879532, T1X, T1W);
|
Chris@10
|
154 T1Y = FNMS(KP923879532, T1X, T1W);
|
Chris@10
|
155 T1r = FNMS(KP923879532, T14, TX);
|
Chris@10
|
156 T15 = FMA(KP923879532, T14, TX);
|
Chris@10
|
157 }
|
Chris@10
|
158 }
|
Chris@10
|
159 {
|
Chris@10
|
160 E Ty, TL, TG, TM, TB, TH;
|
Chris@10
|
161 Ty = R1[0];
|
Chris@10
|
162 TL = R1[WS(rs, 8)];
|
Chris@10
|
163 {
|
Chris@10
|
164 E Tz, TA, TD, TE;
|
Chris@10
|
165 Tz = R1[WS(rs, 4)];
|
Chris@10
|
166 T2d = FMA(KP923879532, T1U, T1T);
|
Chris@10
|
167 T1V = FNMS(KP923879532, T1U, T1T);
|
Chris@10
|
168 TA = R1[WS(rs, 12)];
|
Chris@10
|
169 TD = R1[WS(rs, 10)];
|
Chris@10
|
170 TE = R1[WS(rs, 2)];
|
Chris@10
|
171 TG = R1[WS(rs, 6)];
|
Chris@10
|
172 TM = Tz + TA;
|
Chris@10
|
173 TB = Tz - TA;
|
Chris@10
|
174 TP = FNMS(KP414213562, TD, TE);
|
Chris@10
|
175 TF = FMA(KP414213562, TE, TD);
|
Chris@10
|
176 TH = R1[WS(rs, 14)];
|
Chris@10
|
177 }
|
Chris@10
|
178 T1M = FMA(KP707106781, TB, Ty);
|
Chris@10
|
179 TC = FNMS(KP707106781, TB, Ty);
|
Chris@10
|
180 T1P = FMA(KP707106781, TM, TL);
|
Chris@10
|
181 TN = FNMS(KP707106781, TM, TL);
|
Chris@10
|
182 TO = FMS(KP414213562, TG, TH);
|
Chris@10
|
183 TI = FMA(KP414213562, TH, TG);
|
Chris@10
|
184 }
|
Chris@10
|
185 }
|
Chris@10
|
186 }
|
Chris@10
|
187 {
|
Chris@10
|
188 E T1j, T1O, T1p, T1R, T1o, T2E, T2D, T1m, T1D, T2w, T2v, T1K, T2i, T2c, T2h;
|
Chris@10
|
189 E T29, T2t, T2r, T2f, T2j;
|
Chris@10
|
190 {
|
Chris@10
|
191 E T2a, T2b, T1g, TS, T1f, Tx, T2N, T2L, T1d, T1h;
|
Chris@10
|
192 {
|
Chris@10
|
193 E Td, TR, TK, Tw, T2J, T2K;
|
Chris@10
|
194 T1j = FMA(KP923879532, Tc, T5);
|
Chris@10
|
195 Td = FNMS(KP923879532, Tc, T5);
|
Chris@10
|
196 {
|
Chris@10
|
197 E T1N, TQ, T1Q, TJ;
|
Chris@10
|
198 T1N = TP + TO;
|
Chris@10
|
199 TQ = TO - TP;
|
Chris@10
|
200 T1Q = TF + TI;
|
Chris@10
|
201 TJ = TF - TI;
|
Chris@10
|
202 T2a = FMA(KP923879532, T1N, T1M);
|
Chris@10
|
203 T1O = FNMS(KP923879532, T1N, T1M);
|
Chris@10
|
204 T1p = FMA(KP923879532, TQ, TN);
|
Chris@10
|
205 TR = FNMS(KP923879532, TQ, TN);
|
Chris@10
|
206 T2b = FMA(KP923879532, T1Q, T1P);
|
Chris@10
|
207 T1R = FNMS(KP923879532, T1Q, T1P);
|
Chris@10
|
208 T1o = FMA(KP923879532, TJ, TC);
|
Chris@10
|
209 TK = FNMS(KP923879532, TJ, TC);
|
Chris@10
|
210 Tw = Tm - Tv;
|
Chris@10
|
211 T2E = Tv + Tm;
|
Chris@10
|
212 }
|
Chris@10
|
213 T2D = FMA(KP923879532, T2C, T2B);
|
Chris@10
|
214 T2J = FNMS(KP923879532, T2C, T2B);
|
Chris@10
|
215 T2K = T1k + T1l;
|
Chris@10
|
216 T1m = T1k - T1l;
|
Chris@10
|
217 T1g = FMA(KP534511135, TK, TR);
|
Chris@10
|
218 TS = FNMS(KP534511135, TR, TK);
|
Chris@10
|
219 T1f = FNMS(KP831469612, Tw, Td);
|
Chris@10
|
220 Tx = FMA(KP831469612, Tw, Td);
|
Chris@10
|
221 T2N = FNMS(KP831469612, T2K, T2J);
|
Chris@10
|
222 T2L = FMA(KP831469612, T2K, T2J);
|
Chris@10
|
223 T1d = FNMS(KP534511135, T1c, T15);
|
Chris@10
|
224 T1h = FMA(KP534511135, T15, T1c);
|
Chris@10
|
225 }
|
Chris@10
|
226 {
|
Chris@10
|
227 E T25, T28, T2p, T2q;
|
Chris@10
|
228 T1D = FNMS(KP923879532, T1C, T1z);
|
Chris@10
|
229 T25 = FMA(KP923879532, T1C, T1z);
|
Chris@10
|
230 {
|
Chris@10
|
231 E T2O, T1e, T2M, T1i;
|
Chris@10
|
232 T2O = TS + T1d;
|
Chris@10
|
233 T1e = TS - T1d;
|
Chris@10
|
234 T2M = T1g + T1h;
|
Chris@10
|
235 T1i = T1g - T1h;
|
Chris@10
|
236 Ci[WS(csi, 5)] = FNMS(KP881921264, T2O, T2N);
|
Chris@10
|
237 Ci[WS(csi, 10)] = -(FMA(KP881921264, T2O, T2N));
|
Chris@10
|
238 Cr[WS(csr, 2)] = FMA(KP881921264, T1e, Tx);
|
Chris@10
|
239 Cr[WS(csr, 13)] = FNMS(KP881921264, T1e, Tx);
|
Chris@10
|
240 Ci[WS(csi, 2)] = -(FMA(KP881921264, T2M, T2L));
|
Chris@10
|
241 Ci[WS(csi, 13)] = FNMS(KP881921264, T2M, T2L);
|
Chris@10
|
242 Cr[WS(csr, 5)] = FMA(KP881921264, T1i, T1f);
|
Chris@10
|
243 Cr[WS(csr, 10)] = FNMS(KP881921264, T1i, T1f);
|
Chris@10
|
244 T28 = T26 - T27;
|
Chris@10
|
245 T2w = T26 + T27;
|
Chris@10
|
246 }
|
Chris@10
|
247 T2v = FNMS(KP923879532, T2o, T2n);
|
Chris@10
|
248 T2p = FMA(KP923879532, T2o, T2n);
|
Chris@10
|
249 T2q = T1G + T1J;
|
Chris@10
|
250 T1K = T1G - T1J;
|
Chris@10
|
251 T2i = FMA(KP098491403, T2a, T2b);
|
Chris@10
|
252 T2c = FNMS(KP098491403, T2b, T2a);
|
Chris@10
|
253 T2h = FNMS(KP980785280, T28, T25);
|
Chris@10
|
254 T29 = FMA(KP980785280, T28, T25);
|
Chris@10
|
255 T2t = FNMS(KP980785280, T2q, T2p);
|
Chris@10
|
256 T2r = FMA(KP980785280, T2q, T2p);
|
Chris@10
|
257 T2f = FMA(KP098491403, T2e, T2d);
|
Chris@10
|
258 T2j = FNMS(KP098491403, T2d, T2e);
|
Chris@10
|
259 }
|
Chris@10
|
260 }
|
Chris@10
|
261 {
|
Chris@10
|
262 E T1x, T1q, T1v, T1n, T2H, T2F, T1t, T1w;
|
Chris@10
|
263 {
|
Chris@10
|
264 E T2u, T2g, T2s, T2k;
|
Chris@10
|
265 T2u = T2f - T2c;
|
Chris@10
|
266 T2g = T2c + T2f;
|
Chris@10
|
267 T2s = T2i + T2j;
|
Chris@10
|
268 T2k = T2i - T2j;
|
Chris@10
|
269 Ci[WS(csi, 7)] = FMA(KP995184726, T2u, T2t);
|
Chris@10
|
270 Ci[WS(csi, 8)] = FMS(KP995184726, T2u, T2t);
|
Chris@10
|
271 Cr[0] = FMA(KP995184726, T2g, T29);
|
Chris@10
|
272 Cr[WS(csr, 15)] = FNMS(KP995184726, T2g, T29);
|
Chris@10
|
273 Ci[0] = -(FMA(KP995184726, T2s, T2r));
|
Chris@10
|
274 Ci[WS(csi, 15)] = FNMS(KP995184726, T2s, T2r);
|
Chris@10
|
275 Cr[WS(csr, 7)] = FMA(KP995184726, T2k, T2h);
|
Chris@10
|
276 Cr[WS(csr, 8)] = FNMS(KP995184726, T2k, T2h);
|
Chris@10
|
277 }
|
Chris@10
|
278 T1x = FNMS(KP303346683, T1o, T1p);
|
Chris@10
|
279 T1q = FMA(KP303346683, T1p, T1o);
|
Chris@10
|
280 T1v = FNMS(KP831469612, T1m, T1j);
|
Chris@10
|
281 T1n = FMA(KP831469612, T1m, T1j);
|
Chris@10
|
282 T2H = FNMS(KP831469612, T2E, T2D);
|
Chris@10
|
283 T2F = FMA(KP831469612, T2E, T2D);
|
Chris@10
|
284 T1t = FMA(KP303346683, T1s, T1r);
|
Chris@10
|
285 T1w = FNMS(KP303346683, T1r, T1s);
|
Chris@10
|
286 {
|
Chris@10
|
287 E T2I, T1u, T2G, T1y;
|
Chris@10
|
288 T2I = T1q + T1t;
|
Chris@10
|
289 T1u = T1q - T1t;
|
Chris@10
|
290 T2G = T1x + T1w;
|
Chris@10
|
291 T1y = T1w - T1x;
|
Chris@10
|
292 Ci[WS(csi, 6)] = -(FMA(KP956940335, T2I, T2H));
|
Chris@10
|
293 Ci[WS(csi, 9)] = FNMS(KP956940335, T2I, T2H);
|
Chris@10
|
294 Cr[WS(csr, 1)] = FMA(KP956940335, T1u, T1n);
|
Chris@10
|
295 Cr[WS(csr, 14)] = FNMS(KP956940335, T1u, T1n);
|
Chris@10
|
296 Ci[WS(csi, 1)] = FMA(KP956940335, T2G, T2F);
|
Chris@10
|
297 Ci[WS(csi, 14)] = FMS(KP956940335, T2G, T2F);
|
Chris@10
|
298 Cr[WS(csr, 6)] = FMA(KP956940335, T1y, T1v);
|
Chris@10
|
299 Cr[WS(csr, 9)] = FNMS(KP956940335, T1y, T1v);
|
Chris@10
|
300 }
|
Chris@10
|
301 T23 = FNMS(KP820678790, T1O, T1R);
|
Chris@10
|
302 T1S = FMA(KP820678790, T1R, T1O);
|
Chris@10
|
303 T21 = FNMS(KP980785280, T1K, T1D);
|
Chris@10
|
304 T1L = FMA(KP980785280, T1K, T1D);
|
Chris@10
|
305 T2z = FMA(KP980785280, T2w, T2v);
|
Chris@10
|
306 T2x = FNMS(KP980785280, T2w, T2v);
|
Chris@10
|
307 T1Z = FNMS(KP820678790, T1Y, T1V);
|
Chris@10
|
308 T22 = FMA(KP820678790, T1V, T1Y);
|
Chris@10
|
309 }
|
Chris@10
|
310 }
|
Chris@10
|
311 }
|
Chris@10
|
312 {
|
Chris@10
|
313 E T20, T2A, T24, T2y;
|
Chris@10
|
314 T20 = T1S + T1Z;
|
Chris@10
|
315 T2A = T1Z - T1S;
|
Chris@10
|
316 T24 = T22 - T23;
|
Chris@10
|
317 T2y = T23 + T22;
|
Chris@10
|
318 Ci[WS(csi, 4)] = FMS(KP773010453, T2A, T2z);
|
Chris@10
|
319 Ci[WS(csi, 11)] = FMA(KP773010453, T2A, T2z);
|
Chris@10
|
320 Cr[WS(csr, 3)] = FMA(KP773010453, T20, T1L);
|
Chris@10
|
321 Cr[WS(csr, 12)] = FNMS(KP773010453, T20, T1L);
|
Chris@10
|
322 Ci[WS(csi, 3)] = FMA(KP773010453, T2y, T2x);
|
Chris@10
|
323 Ci[WS(csi, 12)] = FMS(KP773010453, T2y, T2x);
|
Chris@10
|
324 Cr[WS(csr, 4)] = FMA(KP773010453, T24, T21);
|
Chris@10
|
325 Cr[WS(csr, 11)] = FNMS(KP773010453, T24, T21);
|
Chris@10
|
326 }
|
Chris@10
|
327 }
|
Chris@10
|
328 }
|
Chris@10
|
329 }
|
Chris@10
|
330
|
Chris@10
|
331 static const kr2c_desc desc = { 32, "r2cfII_32", {46, 0, 128, 0}, &GENUS };
|
Chris@10
|
332
|
Chris@10
|
333 void X(codelet_r2cfII_32) (planner *p) {
|
Chris@10
|
334 X(kr2c_register) (p, r2cfII_32, &desc);
|
Chris@10
|
335 }
|
Chris@10
|
336
|
Chris@10
|
337 #else /* HAVE_FMA */
|
Chris@10
|
338
|
Chris@10
|
339 /* Generated by: ../../../genfft/gen_r2cf.native -compact -variables 4 -pipeline-latency 4 -n 32 -name r2cfII_32 -dft-II -include r2cfII.h */
|
Chris@10
|
340
|
Chris@10
|
341 /*
|
Chris@10
|
342 * This function contains 174 FP additions, 82 FP multiplications,
|
Chris@10
|
343 * (or, 138 additions, 46 multiplications, 36 fused multiply/add),
|
Chris@10
|
344 * 62 stack variables, 15 constants, and 64 memory accesses
|
Chris@10
|
345 */
|
Chris@10
|
346 #include "r2cfII.h"
|
Chris@10
|
347
|
Chris@10
|
348 static void r2cfII_32(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
|
Chris@10
|
349 {
|
Chris@10
|
350 DK(KP471396736, +0.471396736825997648556387625905254377657460319);
|
Chris@10
|
351 DK(KP881921264, +0.881921264348355029712756863660388349508442621);
|
Chris@10
|
352 DK(KP634393284, +0.634393284163645498215171613225493370675687095);
|
Chris@10
|
353 DK(KP773010453, +0.773010453362736960810906609758469800971041293);
|
Chris@10
|
354 DK(KP290284677, +0.290284677254462367636192375817395274691476278);
|
Chris@10
|
355 DK(KP956940335, +0.956940335732208864935797886980269969482849206);
|
Chris@10
|
356 DK(KP995184726, +0.995184726672196886244836953109479921575474869);
|
Chris@10
|
357 DK(KP098017140, +0.098017140329560601994195563888641845861136673);
|
Chris@10
|
358 DK(KP555570233, +0.555570233019602224742830813948532874374937191);
|
Chris@10
|
359 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
Chris@10
|
360 DK(KP195090322, +0.195090322016128267848284868477022240927691618);
|
Chris@10
|
361 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
Chris@10
|
362 DK(KP382683432, +0.382683432365089771728459984030398866761344562);
|
Chris@10
|
363 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
Chris@10
|
364 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
Chris@10
|
365 {
|
Chris@10
|
366 INT i;
|
Chris@10
|
367 for (i = v; i > 0; i = i - 1, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(128, rs), MAKE_VOLATILE_STRIDE(128, csr), MAKE_VOLATILE_STRIDE(128, csi)) {
|
Chris@10
|
368 E T5, T2D, T1z, T2q, Tc, T2C, T1C, T2n, Tm, T1k, T1J, T26, Tv, T1l, T1G;
|
Chris@10
|
369 E T27, T15, T1r, T1Y, T2e, T1c, T1s, T1V, T2d, TK, T1o, T1R, T2b, TR, T1p;
|
Chris@10
|
370 E T1O, T2a;
|
Chris@10
|
371 {
|
Chris@10
|
372 E T1, T2p, T4, T2o, T2, T3;
|
Chris@10
|
373 T1 = R0[0];
|
Chris@10
|
374 T2p = R0[WS(rs, 8)];
|
Chris@10
|
375 T2 = R0[WS(rs, 4)];
|
Chris@10
|
376 T3 = R0[WS(rs, 12)];
|
Chris@10
|
377 T4 = KP707106781 * (T2 - T3);
|
Chris@10
|
378 T2o = KP707106781 * (T2 + T3);
|
Chris@10
|
379 T5 = T1 + T4;
|
Chris@10
|
380 T2D = T2p - T2o;
|
Chris@10
|
381 T1z = T1 - T4;
|
Chris@10
|
382 T2q = T2o + T2p;
|
Chris@10
|
383 }
|
Chris@10
|
384 {
|
Chris@10
|
385 E T8, T1A, Tb, T1B;
|
Chris@10
|
386 {
|
Chris@10
|
387 E T6, T7, T9, Ta;
|
Chris@10
|
388 T6 = R0[WS(rs, 2)];
|
Chris@10
|
389 T7 = R0[WS(rs, 10)];
|
Chris@10
|
390 T8 = FNMS(KP382683432, T7, KP923879532 * T6);
|
Chris@10
|
391 T1A = FMA(KP382683432, T6, KP923879532 * T7);
|
Chris@10
|
392 T9 = R0[WS(rs, 6)];
|
Chris@10
|
393 Ta = R0[WS(rs, 14)];
|
Chris@10
|
394 Tb = FNMS(KP923879532, Ta, KP382683432 * T9);
|
Chris@10
|
395 T1B = FMA(KP923879532, T9, KP382683432 * Ta);
|
Chris@10
|
396 }
|
Chris@10
|
397 Tc = T8 + Tb;
|
Chris@10
|
398 T2C = Tb - T8;
|
Chris@10
|
399 T1C = T1A - T1B;
|
Chris@10
|
400 T2n = T1A + T1B;
|
Chris@10
|
401 }
|
Chris@10
|
402 {
|
Chris@10
|
403 E Te, Tk, Th, Tj, Tf, Tg;
|
Chris@10
|
404 Te = R0[WS(rs, 1)];
|
Chris@10
|
405 Tk = R0[WS(rs, 9)];
|
Chris@10
|
406 Tf = R0[WS(rs, 5)];
|
Chris@10
|
407 Tg = R0[WS(rs, 13)];
|
Chris@10
|
408 Th = KP707106781 * (Tf - Tg);
|
Chris@10
|
409 Tj = KP707106781 * (Tf + Tg);
|
Chris@10
|
410 {
|
Chris@10
|
411 E Ti, Tl, T1H, T1I;
|
Chris@10
|
412 Ti = Te + Th;
|
Chris@10
|
413 Tl = Tj + Tk;
|
Chris@10
|
414 Tm = FNMS(KP195090322, Tl, KP980785280 * Ti);
|
Chris@10
|
415 T1k = FMA(KP195090322, Ti, KP980785280 * Tl);
|
Chris@10
|
416 T1H = Tk - Tj;
|
Chris@10
|
417 T1I = Te - Th;
|
Chris@10
|
418 T1J = FNMS(KP555570233, T1I, KP831469612 * T1H);
|
Chris@10
|
419 T26 = FMA(KP831469612, T1I, KP555570233 * T1H);
|
Chris@10
|
420 }
|
Chris@10
|
421 }
|
Chris@10
|
422 {
|
Chris@10
|
423 E Tq, Tt, Tp, Ts, Tn, To;
|
Chris@10
|
424 Tq = R0[WS(rs, 15)];
|
Chris@10
|
425 Tt = R0[WS(rs, 7)];
|
Chris@10
|
426 Tn = R0[WS(rs, 3)];
|
Chris@10
|
427 To = R0[WS(rs, 11)];
|
Chris@10
|
428 Tp = KP707106781 * (Tn - To);
|
Chris@10
|
429 Ts = KP707106781 * (Tn + To);
|
Chris@10
|
430 {
|
Chris@10
|
431 E Tr, Tu, T1E, T1F;
|
Chris@10
|
432 Tr = Tp - Tq;
|
Chris@10
|
433 Tu = Ts + Tt;
|
Chris@10
|
434 Tv = FMA(KP980785280, Tr, KP195090322 * Tu);
|
Chris@10
|
435 T1l = FNMS(KP980785280, Tu, KP195090322 * Tr);
|
Chris@10
|
436 T1E = Tt - Ts;
|
Chris@10
|
437 T1F = Tp + Tq;
|
Chris@10
|
438 T1G = FNMS(KP555570233, T1F, KP831469612 * T1E);
|
Chris@10
|
439 T27 = FMA(KP831469612, T1F, KP555570233 * T1E);
|
Chris@10
|
440 }
|
Chris@10
|
441 }
|
Chris@10
|
442 {
|
Chris@10
|
443 E TW, T1a, TV, T19, T10, T16, T13, T17, TT, TU;
|
Chris@10
|
444 TW = R1[WS(rs, 15)];
|
Chris@10
|
445 T1a = R1[WS(rs, 7)];
|
Chris@10
|
446 TT = R1[WS(rs, 3)];
|
Chris@10
|
447 TU = R1[WS(rs, 11)];
|
Chris@10
|
448 TV = KP707106781 * (TT - TU);
|
Chris@10
|
449 T19 = KP707106781 * (TT + TU);
|
Chris@10
|
450 {
|
Chris@10
|
451 E TY, TZ, T11, T12;
|
Chris@10
|
452 TY = R1[WS(rs, 1)];
|
Chris@10
|
453 TZ = R1[WS(rs, 9)];
|
Chris@10
|
454 T10 = FNMS(KP382683432, TZ, KP923879532 * TY);
|
Chris@10
|
455 T16 = FMA(KP382683432, TY, KP923879532 * TZ);
|
Chris@10
|
456 T11 = R1[WS(rs, 5)];
|
Chris@10
|
457 T12 = R1[WS(rs, 13)];
|
Chris@10
|
458 T13 = FNMS(KP923879532, T12, KP382683432 * T11);
|
Chris@10
|
459 T17 = FMA(KP923879532, T11, KP382683432 * T12);
|
Chris@10
|
460 }
|
Chris@10
|
461 {
|
Chris@10
|
462 E TX, T14, T1W, T1X;
|
Chris@10
|
463 TX = TV - TW;
|
Chris@10
|
464 T14 = T10 + T13;
|
Chris@10
|
465 T15 = TX + T14;
|
Chris@10
|
466 T1r = TX - T14;
|
Chris@10
|
467 T1W = T13 - T10;
|
Chris@10
|
468 T1X = T1a - T19;
|
Chris@10
|
469 T1Y = T1W - T1X;
|
Chris@10
|
470 T2e = T1W + T1X;
|
Chris@10
|
471 }
|
Chris@10
|
472 {
|
Chris@10
|
473 E T18, T1b, T1T, T1U;
|
Chris@10
|
474 T18 = T16 + T17;
|
Chris@10
|
475 T1b = T19 + T1a;
|
Chris@10
|
476 T1c = T18 + T1b;
|
Chris@10
|
477 T1s = T1b - T18;
|
Chris@10
|
478 T1T = TV + TW;
|
Chris@10
|
479 T1U = T16 - T17;
|
Chris@10
|
480 T1V = T1T + T1U;
|
Chris@10
|
481 T2d = T1U - T1T;
|
Chris@10
|
482 }
|
Chris@10
|
483 }
|
Chris@10
|
484 {
|
Chris@10
|
485 E Ty, TP, TB, TO, TF, TL, TI, TM, Tz, TA;
|
Chris@10
|
486 Ty = R1[0];
|
Chris@10
|
487 TP = R1[WS(rs, 8)];
|
Chris@10
|
488 Tz = R1[WS(rs, 4)];
|
Chris@10
|
489 TA = R1[WS(rs, 12)];
|
Chris@10
|
490 TB = KP707106781 * (Tz - TA);
|
Chris@10
|
491 TO = KP707106781 * (Tz + TA);
|
Chris@10
|
492 {
|
Chris@10
|
493 E TD, TE, TG, TH;
|
Chris@10
|
494 TD = R1[WS(rs, 2)];
|
Chris@10
|
495 TE = R1[WS(rs, 10)];
|
Chris@10
|
496 TF = FNMS(KP382683432, TE, KP923879532 * TD);
|
Chris@10
|
497 TL = FMA(KP382683432, TD, KP923879532 * TE);
|
Chris@10
|
498 TG = R1[WS(rs, 6)];
|
Chris@10
|
499 TH = R1[WS(rs, 14)];
|
Chris@10
|
500 TI = FNMS(KP923879532, TH, KP382683432 * TG);
|
Chris@10
|
501 TM = FMA(KP923879532, TG, KP382683432 * TH);
|
Chris@10
|
502 }
|
Chris@10
|
503 {
|
Chris@10
|
504 E TC, TJ, T1P, T1Q;
|
Chris@10
|
505 TC = Ty + TB;
|
Chris@10
|
506 TJ = TF + TI;
|
Chris@10
|
507 TK = TC + TJ;
|
Chris@10
|
508 T1o = TC - TJ;
|
Chris@10
|
509 T1P = TI - TF;
|
Chris@10
|
510 T1Q = TP - TO;
|
Chris@10
|
511 T1R = T1P - T1Q;
|
Chris@10
|
512 T2b = T1P + T1Q;
|
Chris@10
|
513 }
|
Chris@10
|
514 {
|
Chris@10
|
515 E TN, TQ, T1M, T1N;
|
Chris@10
|
516 TN = TL + TM;
|
Chris@10
|
517 TQ = TO + TP;
|
Chris@10
|
518 TR = TN + TQ;
|
Chris@10
|
519 T1p = TQ - TN;
|
Chris@10
|
520 T1M = Ty - TB;
|
Chris@10
|
521 T1N = TL - TM;
|
Chris@10
|
522 T1O = T1M - T1N;
|
Chris@10
|
523 T2a = T1M + T1N;
|
Chris@10
|
524 }
|
Chris@10
|
525 }
|
Chris@10
|
526 {
|
Chris@10
|
527 E Tx, T1f, T2s, T2u, T1e, T2l, T1i, T2t;
|
Chris@10
|
528 {
|
Chris@10
|
529 E Td, Tw, T2m, T2r;
|
Chris@10
|
530 Td = T5 + Tc;
|
Chris@10
|
531 Tw = Tm + Tv;
|
Chris@10
|
532 Tx = Td - Tw;
|
Chris@10
|
533 T1f = Td + Tw;
|
Chris@10
|
534 T2m = T1l - T1k;
|
Chris@10
|
535 T2r = T2n + T2q;
|
Chris@10
|
536 T2s = T2m - T2r;
|
Chris@10
|
537 T2u = T2m + T2r;
|
Chris@10
|
538 }
|
Chris@10
|
539 {
|
Chris@10
|
540 E TS, T1d, T1g, T1h;
|
Chris@10
|
541 TS = FMA(KP098017140, TK, KP995184726 * TR);
|
Chris@10
|
542 T1d = FNMS(KP995184726, T1c, KP098017140 * T15);
|
Chris@10
|
543 T1e = TS + T1d;
|
Chris@10
|
544 T2l = T1d - TS;
|
Chris@10
|
545 T1g = FNMS(KP098017140, TR, KP995184726 * TK);
|
Chris@10
|
546 T1h = FMA(KP995184726, T15, KP098017140 * T1c);
|
Chris@10
|
547 T1i = T1g + T1h;
|
Chris@10
|
548 T2t = T1h - T1g;
|
Chris@10
|
549 }
|
Chris@10
|
550 Cr[WS(csr, 8)] = Tx - T1e;
|
Chris@10
|
551 Ci[WS(csi, 8)] = T2t - T2u;
|
Chris@10
|
552 Cr[WS(csr, 7)] = Tx + T1e;
|
Chris@10
|
553 Ci[WS(csi, 7)] = T2t + T2u;
|
Chris@10
|
554 Cr[WS(csr, 15)] = T1f - T1i;
|
Chris@10
|
555 Ci[WS(csi, 15)] = T2l - T2s;
|
Chris@10
|
556 Cr[0] = T1f + T1i;
|
Chris@10
|
557 Ci[0] = T2l + T2s;
|
Chris@10
|
558 }
|
Chris@10
|
559 {
|
Chris@10
|
560 E T29, T2h, T2M, T2O, T2g, T2J, T2k, T2N;
|
Chris@10
|
561 {
|
Chris@10
|
562 E T25, T28, T2K, T2L;
|
Chris@10
|
563 T25 = T1z + T1C;
|
Chris@10
|
564 T28 = T26 - T27;
|
Chris@10
|
565 T29 = T25 + T28;
|
Chris@10
|
566 T2h = T25 - T28;
|
Chris@10
|
567 T2K = T1J + T1G;
|
Chris@10
|
568 T2L = T2C + T2D;
|
Chris@10
|
569 T2M = T2K - T2L;
|
Chris@10
|
570 T2O = T2K + T2L;
|
Chris@10
|
571 }
|
Chris@10
|
572 {
|
Chris@10
|
573 E T2c, T2f, T2i, T2j;
|
Chris@10
|
574 T2c = FMA(KP956940335, T2a, KP290284677 * T2b);
|
Chris@10
|
575 T2f = FNMS(KP290284677, T2e, KP956940335 * T2d);
|
Chris@10
|
576 T2g = T2c + T2f;
|
Chris@10
|
577 T2J = T2f - T2c;
|
Chris@10
|
578 T2i = FMA(KP290284677, T2d, KP956940335 * T2e);
|
Chris@10
|
579 T2j = FNMS(KP290284677, T2a, KP956940335 * T2b);
|
Chris@10
|
580 T2k = T2i - T2j;
|
Chris@10
|
581 T2N = T2j + T2i;
|
Chris@10
|
582 }
|
Chris@10
|
583 Cr[WS(csr, 14)] = T29 - T2g;
|
Chris@10
|
584 Ci[WS(csi, 14)] = T2N - T2O;
|
Chris@10
|
585 Cr[WS(csr, 1)] = T29 + T2g;
|
Chris@10
|
586 Ci[WS(csi, 1)] = T2N + T2O;
|
Chris@10
|
587 Cr[WS(csr, 9)] = T2h - T2k;
|
Chris@10
|
588 Ci[WS(csi, 9)] = T2J - T2M;
|
Chris@10
|
589 Cr[WS(csr, 6)] = T2h + T2k;
|
Chris@10
|
590 Ci[WS(csi, 6)] = T2J + T2M;
|
Chris@10
|
591 }
|
Chris@10
|
592 {
|
Chris@10
|
593 E T1n, T1v, T2y, T2A, T1u, T2v, T1y, T2z;
|
Chris@10
|
594 {
|
Chris@10
|
595 E T1j, T1m, T2w, T2x;
|
Chris@10
|
596 T1j = T5 - Tc;
|
Chris@10
|
597 T1m = T1k + T1l;
|
Chris@10
|
598 T1n = T1j + T1m;
|
Chris@10
|
599 T1v = T1j - T1m;
|
Chris@10
|
600 T2w = Tv - Tm;
|
Chris@10
|
601 T2x = T2q - T2n;
|
Chris@10
|
602 T2y = T2w - T2x;
|
Chris@10
|
603 T2A = T2w + T2x;
|
Chris@10
|
604 }
|
Chris@10
|
605 {
|
Chris@10
|
606 E T1q, T1t, T1w, T1x;
|
Chris@10
|
607 T1q = FMA(KP773010453, T1o, KP634393284 * T1p);
|
Chris@10
|
608 T1t = FNMS(KP634393284, T1s, KP773010453 * T1r);
|
Chris@10
|
609 T1u = T1q + T1t;
|
Chris@10
|
610 T2v = T1t - T1q;
|
Chris@10
|
611 T1w = FMA(KP634393284, T1r, KP773010453 * T1s);
|
Chris@10
|
612 T1x = FNMS(KP634393284, T1o, KP773010453 * T1p);
|
Chris@10
|
613 T1y = T1w - T1x;
|
Chris@10
|
614 T2z = T1x + T1w;
|
Chris@10
|
615 }
|
Chris@10
|
616 Cr[WS(csr, 12)] = T1n - T1u;
|
Chris@10
|
617 Ci[WS(csi, 12)] = T2z - T2A;
|
Chris@10
|
618 Cr[WS(csr, 3)] = T1n + T1u;
|
Chris@10
|
619 Ci[WS(csi, 3)] = T2z + T2A;
|
Chris@10
|
620 Cr[WS(csr, 11)] = T1v - T1y;
|
Chris@10
|
621 Ci[WS(csi, 11)] = T2v - T2y;
|
Chris@10
|
622 Cr[WS(csr, 4)] = T1v + T1y;
|
Chris@10
|
623 Ci[WS(csi, 4)] = T2v + T2y;
|
Chris@10
|
624 }
|
Chris@10
|
625 {
|
Chris@10
|
626 E T1L, T21, T2G, T2I, T20, T2H, T24, T2B;
|
Chris@10
|
627 {
|
Chris@10
|
628 E T1D, T1K, T2E, T2F;
|
Chris@10
|
629 T1D = T1z - T1C;
|
Chris@10
|
630 T1K = T1G - T1J;
|
Chris@10
|
631 T1L = T1D + T1K;
|
Chris@10
|
632 T21 = T1D - T1K;
|
Chris@10
|
633 T2E = T2C - T2D;
|
Chris@10
|
634 T2F = T26 + T27;
|
Chris@10
|
635 T2G = T2E - T2F;
|
Chris@10
|
636 T2I = T2F + T2E;
|
Chris@10
|
637 }
|
Chris@10
|
638 {
|
Chris@10
|
639 E T1S, T1Z, T22, T23;
|
Chris@10
|
640 T1S = FMA(KP881921264, T1O, KP471396736 * T1R);
|
Chris@10
|
641 T1Z = FMA(KP881921264, T1V, KP471396736 * T1Y);
|
Chris@10
|
642 T20 = T1S - T1Z;
|
Chris@10
|
643 T2H = T1S + T1Z;
|
Chris@10
|
644 T22 = FNMS(KP471396736, T1V, KP881921264 * T1Y);
|
Chris@10
|
645 T23 = FNMS(KP471396736, T1O, KP881921264 * T1R);
|
Chris@10
|
646 T24 = T22 - T23;
|
Chris@10
|
647 T2B = T23 + T22;
|
Chris@10
|
648 }
|
Chris@10
|
649 Cr[WS(csr, 13)] = T1L - T20;
|
Chris@10
|
650 Ci[WS(csi, 13)] = T2B - T2G;
|
Chris@10
|
651 Cr[WS(csr, 2)] = T1L + T20;
|
Chris@10
|
652 Ci[WS(csi, 2)] = T2B + T2G;
|
Chris@10
|
653 Cr[WS(csr, 10)] = T21 - T24;
|
Chris@10
|
654 Ci[WS(csi, 10)] = T2I - T2H;
|
Chris@10
|
655 Cr[WS(csr, 5)] = T21 + T24;
|
Chris@10
|
656 Ci[WS(csi, 5)] = -(T2H + T2I);
|
Chris@10
|
657 }
|
Chris@10
|
658 }
|
Chris@10
|
659 }
|
Chris@10
|
660 }
|
Chris@10
|
661
|
Chris@10
|
662 static const kr2c_desc desc = { 32, "r2cfII_32", {138, 46, 36, 0}, &GENUS };
|
Chris@10
|
663
|
Chris@10
|
664 void X(codelet_r2cfII_32) (planner *p) {
|
Chris@10
|
665 X(kr2c_register) (p, r2cfII_32, &desc);
|
Chris@10
|
666 }
|
Chris@10
|
667
|
Chris@10
|
668 #endif /* HAVE_FMA */
|