Chris@10
|
1 /*
|
Chris@10
|
2 * Copyright (c) 2003, 2007-11 Matteo Frigo
|
Chris@10
|
3 * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology
|
Chris@10
|
4 *
|
Chris@10
|
5 * This program is free software; you can redistribute it and/or modify
|
Chris@10
|
6 * it under the terms of the GNU General Public License as published by
|
Chris@10
|
7 * the Free Software Foundation; either version 2 of the License, or
|
Chris@10
|
8 * (at your option) any later version.
|
Chris@10
|
9 *
|
Chris@10
|
10 * This program is distributed in the hope that it will be useful,
|
Chris@10
|
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
Chris@10
|
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
Chris@10
|
13 * GNU General Public License for more details.
|
Chris@10
|
14 *
|
Chris@10
|
15 * You should have received a copy of the GNU General Public License
|
Chris@10
|
16 * along with this program; if not, write to the Free Software
|
Chris@10
|
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
Chris@10
|
18 *
|
Chris@10
|
19 */
|
Chris@10
|
20
|
Chris@10
|
21 /* This file was automatically generated --- DO NOT EDIT */
|
Chris@10
|
22 /* Generated on Sun Nov 25 07:41:45 EST 2012 */
|
Chris@10
|
23
|
Chris@10
|
24 #include "codelet-rdft.h"
|
Chris@10
|
25
|
Chris@10
|
26 #ifdef HAVE_FMA
|
Chris@10
|
27
|
Chris@10
|
28 /* Generated by: ../../../genfft/gen_r2cb.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -sign 1 -n 25 -name r2cbIII_25 -dft-III -include r2cbIII.h */
|
Chris@10
|
29
|
Chris@10
|
30 /*
|
Chris@10
|
31 * This function contains 152 FP additions, 120 FP multiplications,
|
Chris@10
|
32 * (or, 32 additions, 0 multiplications, 120 fused multiply/add),
|
Chris@10
|
33 * 115 stack variables, 44 constants, and 50 memory accesses
|
Chris@10
|
34 */
|
Chris@10
|
35 #include "r2cbIII.h"
|
Chris@10
|
36
|
Chris@10
|
37 static void r2cbIII_25(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(KP979740652, +0.979740652857618686258237536568998933733477632);
|
Chris@10
|
40 DK(KP438153340, +0.438153340021931793654057951961031291699532119);
|
Chris@10
|
41 DK(KP1_752613360, +1.752613360087727174616231807844125166798128477);
|
Chris@10
|
42 DK(KP963507348, +0.963507348203430549974383005744259307057084020);
|
Chris@10
|
43 DK(KP1_721083328, +1.721083328735889354196523361841037632825608373);
|
Chris@10
|
44 DK(KP1_606007150, +1.606007150877320829666881187140752009270929701);
|
Chris@10
|
45 DK(KP1_011627398, +1.011627398597394192215998921771049272931807941);
|
Chris@10
|
46 DK(KP641441904, +0.641441904830606407298806329068862424939687989);
|
Chris@10
|
47 DK(KP595480289, +0.595480289600000014706716770488118292997907308);
|
Chris@10
|
48 DK(KP452413526, +0.452413526233009763856834323966348796985206956);
|
Chris@10
|
49 DK(KP1_809654104, +1.809654104932039055427337295865395187940827822);
|
Chris@10
|
50 DK(KP933137358, +0.933137358350283770603023973254446451924190884);
|
Chris@10
|
51 DK(KP1_666834356, +1.666834356657377354817925100486477686277992119);
|
Chris@10
|
52 DK(KP1_842354653, +1.842354653930286640500894870830132058718564461);
|
Chris@10
|
53 DK(KP1_082908895, +1.082908895072625554092571180165639018104066379);
|
Chris@10
|
54 DK(KP576710603, +0.576710603632765877371579268136471017090111488);
|
Chris@10
|
55 DK(KP662318342, +0.662318342759882818626911127577439236802190210);
|
Chris@10
|
56 DK(KP484291580, +0.484291580564315559745084187732367906918006201);
|
Chris@10
|
57 DK(KP1_937166322, +1.937166322257262238980336750929471627672024806);
|
Chris@10
|
58 DK(KP1_898359647, +1.898359647016882523151110931686726543423167685);
|
Chris@10
|
59 DK(KP1_386580726, +1.386580726567734802700860150804827247498955921);
|
Chris@10
|
60 DK(KP904730450, +0.904730450839922351881287709692877908104763647);
|
Chris@10
|
61 DK(KP1_115827804, +1.115827804063668528375399296931134075984874304);
|
Chris@10
|
62 DK(KP470564281, +0.470564281212251493087595091036643380879947982);
|
Chris@10
|
63 DK(KP634619297, +0.634619297544148100711287640319130485732531031);
|
Chris@10
|
64 DK(KP499013364, +0.499013364214135780976168403431725276668452610);
|
Chris@10
|
65 DK(KP1_996053456, +1.996053456856543123904673613726901106673810439);
|
Chris@10
|
66 DK(KP559154169, +0.559154169276087864842202529084232643714075927);
|
Chris@10
|
67 DK(KP683113946, +0.683113946453479238701949862233725244439656928);
|
Chris@10
|
68 DK(KP730409924, +0.730409924561256563751459444999838399157094302);
|
Chris@10
|
69 DK(KP549754652, +0.549754652192770074288023275540779861653779767);
|
Chris@10
|
70 DK(KP256756360, +0.256756360367726783319498520922669048172391148);
|
Chris@10
|
71 DK(KP451418159, +0.451418159099103183892477933432151804893354132);
|
Chris@10
|
72 DK(KP846146756, +0.846146756728608505452954290121135880883743802);
|
Chris@10
|
73 DK(KP1_902113032, +1.902113032590307144232878666758764286811397268);
|
Chris@10
|
74 DK(KP062914667, +0.062914667253649757225485955897349402364686947);
|
Chris@10
|
75 DK(KP939062505, +0.939062505817492352556001843133229685779824606);
|
Chris@10
|
76 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
|
Chris@10
|
77 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
|
Chris@10
|
78 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
|
Chris@10
|
79 DK(KP1_118033988, +1.118033988749894848204586834365638117720309180);
|
Chris@10
|
80 DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
|
Chris@10
|
81 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
|
Chris@10
|
82 DK(KP618033988, +0.618033988749894848204586834365638117720309180);
|
Chris@10
|
83 {
|
Chris@10
|
84 INT i;
|
Chris@10
|
85 for (i = v; i > 0; i = i - 1, R0 = R0 + ovs, R1 = R1 + ovs, Cr = Cr + ivs, Ci = Ci + ivs, MAKE_VOLATILE_STRIDE(100, rs), MAKE_VOLATILE_STRIDE(100, csr), MAKE_VOLATILE_STRIDE(100, csi)) {
|
Chris@10
|
86 E T1P, T2c, T2a, T24, T26, T25, T27, T2b;
|
Chris@10
|
87 {
|
Chris@10
|
88 E T1O, TS, T5, T1N, TP, Te, TA, T2i, T1V, T17, T1B, T2h, T1S, T10, T1C;
|
Chris@10
|
89 E T1a, T19, Tn, T1h, T1l, T1Y, T1e, T21, TJ, T1g;
|
Chris@10
|
90 {
|
Chris@10
|
91 E T1, T2, T3, TQ, TR;
|
Chris@10
|
92 TQ = Ci[WS(csi, 7)];
|
Chris@10
|
93 TR = Ci[WS(csi, 2)];
|
Chris@10
|
94 T1 = Cr[WS(csr, 12)];
|
Chris@10
|
95 T2 = Cr[WS(csr, 7)];
|
Chris@10
|
96 T3 = Cr[WS(csr, 2)];
|
Chris@10
|
97 T1O = FNMS(KP618033988, TQ, TR);
|
Chris@10
|
98 TS = FMA(KP618033988, TR, TQ);
|
Chris@10
|
99 {
|
Chris@10
|
100 E TV, TU, T1U, T16, T12, T1R, TZ, T11;
|
Chris@10
|
101 {
|
Chris@10
|
102 E T6, Tz, T14, T15, TX, Tu, Td, Tx, TY, T4, TO, Ty;
|
Chris@10
|
103 T6 = Cr[WS(csr, 11)];
|
Chris@10
|
104 T4 = T2 + T3;
|
Chris@10
|
105 TO = T3 - T2;
|
Chris@10
|
106 Tz = Ci[WS(csi, 11)];
|
Chris@10
|
107 {
|
Chris@10
|
108 E Ta, T9, Tb, T7, T8, TN;
|
Chris@10
|
109 T7 = Cr[WS(csr, 6)];
|
Chris@10
|
110 T8 = Cr[WS(csr, 8)];
|
Chris@10
|
111 TN = FNMS(KP500000000, T4, T1);
|
Chris@10
|
112 T5 = FMA(KP2_000000000, T4, T1);
|
Chris@10
|
113 Ta = Cr[WS(csr, 1)];
|
Chris@10
|
114 T14 = T8 - T7;
|
Chris@10
|
115 T9 = T7 + T8;
|
Chris@10
|
116 T1N = FMA(KP1_118033988, TO, TN);
|
Chris@10
|
117 TP = FNMS(KP1_118033988, TO, TN);
|
Chris@10
|
118 Tb = Cr[WS(csr, 3)];
|
Chris@10
|
119 {
|
Chris@10
|
120 E Tv, Tw, Ts, Tt, Tc;
|
Chris@10
|
121 Ts = Ci[WS(csi, 8)];
|
Chris@10
|
122 Tt = Ci[WS(csi, 6)];
|
Chris@10
|
123 T15 = Tb - Ta;
|
Chris@10
|
124 Tc = Ta + Tb;
|
Chris@10
|
125 Tv = Ci[WS(csi, 3)];
|
Chris@10
|
126 TX = Tt + Ts;
|
Chris@10
|
127 Tu = Ts - Tt;
|
Chris@10
|
128 Tw = Ci[WS(csi, 1)];
|
Chris@10
|
129 Td = T9 + Tc;
|
Chris@10
|
130 TV = Tc - T9;
|
Chris@10
|
131 Tx = Tv - Tw;
|
Chris@10
|
132 TY = Tw + Tv;
|
Chris@10
|
133 }
|
Chris@10
|
134 }
|
Chris@10
|
135 Te = T6 + Td;
|
Chris@10
|
136 TU = FMS(KP250000000, Td, T6);
|
Chris@10
|
137 T1U = FNMS(KP618033988, T14, T15);
|
Chris@10
|
138 T16 = FMA(KP618033988, T15, T14);
|
Chris@10
|
139 T12 = Tx - Tu;
|
Chris@10
|
140 Ty = Tu + Tx;
|
Chris@10
|
141 T1R = FNMS(KP618033988, TX, TY);
|
Chris@10
|
142 TZ = FMA(KP618033988, TY, TX);
|
Chris@10
|
143 TA = Ty - Tz;
|
Chris@10
|
144 T11 = FMA(KP250000000, Ty, Tz);
|
Chris@10
|
145 }
|
Chris@10
|
146 {
|
Chris@10
|
147 E Tf, TI, T1j, T1k, Tm, T1c, TD, TG, T1d, TH;
|
Chris@10
|
148 Tf = Cr[WS(csr, 10)];
|
Chris@10
|
149 TI = Ci[WS(csi, 10)];
|
Chris@10
|
150 {
|
Chris@10
|
151 E T13, T1T, TW, T1Q;
|
Chris@10
|
152 T13 = FMA(KP559016994, T12, T11);
|
Chris@10
|
153 T1T = FNMS(KP559016994, T12, T11);
|
Chris@10
|
154 TW = FMA(KP559016994, TV, TU);
|
Chris@10
|
155 T1Q = FNMS(KP559016994, TV, TU);
|
Chris@10
|
156 T2i = FMA(KP951056516, T1U, T1T);
|
Chris@10
|
157 T1V = FNMS(KP951056516, T1U, T1T);
|
Chris@10
|
158 T17 = FMA(KP951056516, T16, T13);
|
Chris@10
|
159 T1B = FNMS(KP951056516, T16, T13);
|
Chris@10
|
160 T2h = FNMS(KP951056516, T1R, T1Q);
|
Chris@10
|
161 T1S = FMA(KP951056516, T1R, T1Q);
|
Chris@10
|
162 T10 = FNMS(KP951056516, TZ, TW);
|
Chris@10
|
163 T1C = FMA(KP951056516, TZ, TW);
|
Chris@10
|
164 {
|
Chris@10
|
165 E Tg, Th, Tj, Tk;
|
Chris@10
|
166 Tg = Cr[WS(csr, 5)];
|
Chris@10
|
167 Th = Cr[WS(csr, 9)];
|
Chris@10
|
168 Tj = Cr[0];
|
Chris@10
|
169 Tk = Cr[WS(csr, 4)];
|
Chris@10
|
170 {
|
Chris@10
|
171 E TB, Ti, Tl, TC, TE, TF;
|
Chris@10
|
172 TB = Ci[WS(csi, 9)];
|
Chris@10
|
173 T1j = Tg - Th;
|
Chris@10
|
174 Ti = Tg + Th;
|
Chris@10
|
175 T1k = Tk - Tj;
|
Chris@10
|
176 Tl = Tj + Tk;
|
Chris@10
|
177 TC = Ci[WS(csi, 5)];
|
Chris@10
|
178 TE = Ci[WS(csi, 4)];
|
Chris@10
|
179 TF = Ci[0];
|
Chris@10
|
180 Tm = Ti + Tl;
|
Chris@10
|
181 T1a = Ti - Tl;
|
Chris@10
|
182 T1c = TC + TB;
|
Chris@10
|
183 TD = TB - TC;
|
Chris@10
|
184 TG = TE - TF;
|
Chris@10
|
185 T1d = TF + TE;
|
Chris@10
|
186 }
|
Chris@10
|
187 }
|
Chris@10
|
188 }
|
Chris@10
|
189 T19 = FMS(KP250000000, Tm, Tf);
|
Chris@10
|
190 Tn = Tf + Tm;
|
Chris@10
|
191 T1h = TD - TG;
|
Chris@10
|
192 TH = TD + TG;
|
Chris@10
|
193 T1l = FNMS(KP618033988, T1k, T1j);
|
Chris@10
|
194 T1Y = FMA(KP618033988, T1j, T1k);
|
Chris@10
|
195 T1e = FMA(KP618033988, T1d, T1c);
|
Chris@10
|
196 T21 = FNMS(KP618033988, T1c, T1d);
|
Chris@10
|
197 TJ = TH - TI;
|
Chris@10
|
198 T1g = FMA(KP250000000, TH, TI);
|
Chris@10
|
199 }
|
Chris@10
|
200 }
|
Chris@10
|
201 }
|
Chris@10
|
202 {
|
Chris@10
|
203 E T1Z, T1m, T1y, T22, T1f, T1z, T2j, T2g, T2d, T2q, T2s;
|
Chris@10
|
204 {
|
Chris@10
|
205 E Tq, To, T2e, T2f;
|
Chris@10
|
206 Tq = Tn - Te;
|
Chris@10
|
207 To = Te + Tn;
|
Chris@10
|
208 {
|
Chris@10
|
209 E T1i, T1X, T1b, T20;
|
Chris@10
|
210 T1i = FNMS(KP559016994, T1h, T1g);
|
Chris@10
|
211 T1X = FMA(KP559016994, T1h, T1g);
|
Chris@10
|
212 T1b = FNMS(KP559016994, T1a, T19);
|
Chris@10
|
213 T20 = FMA(KP559016994, T1a, T19);
|
Chris@10
|
214 T2e = FMA(KP951056516, T1Y, T1X);
|
Chris@10
|
215 T1Z = FNMS(KP951056516, T1Y, T1X);
|
Chris@10
|
216 T1m = FNMS(KP951056516, T1l, T1i);
|
Chris@10
|
217 T1y = FMA(KP951056516, T1l, T1i);
|
Chris@10
|
218 T2f = FNMS(KP951056516, T21, T20);
|
Chris@10
|
219 T22 = FMA(KP951056516, T21, T20);
|
Chris@10
|
220 T1f = FNMS(KP951056516, T1e, T1b);
|
Chris@10
|
221 T1z = FMA(KP951056516, T1e, T1b);
|
Chris@10
|
222 }
|
Chris@10
|
223 {
|
Chris@10
|
224 E T2o, TK, TM, T2p, Tr, TL, Tp;
|
Chris@10
|
225 T2o = FMA(KP939062505, T2h, T2i);
|
Chris@10
|
226 T2j = FNMS(KP939062505, T2i, T2h);
|
Chris@10
|
227 R0[0] = FMA(KP2_000000000, To, T5);
|
Chris@10
|
228 Tp = FNMS(KP500000000, To, T5);
|
Chris@10
|
229 TK = FMA(KP618033988, TJ, TA);
|
Chris@10
|
230 TM = FNMS(KP618033988, TA, TJ);
|
Chris@10
|
231 T2g = FNMS(KP062914667, T2f, T2e);
|
Chris@10
|
232 T2p = FMA(KP062914667, T2e, T2f);
|
Chris@10
|
233 Tr = FNMS(KP1_118033988, Tq, Tp);
|
Chris@10
|
234 TL = FMA(KP1_118033988, Tq, Tp);
|
Chris@10
|
235 T2d = FMA(KP1_902113032, T1O, T1N);
|
Chris@10
|
236 T1P = FNMS(KP1_902113032, T1O, T1N);
|
Chris@10
|
237 T2q = FMA(KP846146756, T2p, T2o);
|
Chris@10
|
238 T2s = FNMS(KP451418159, T2o, T2p);
|
Chris@10
|
239 R0[WS(rs, 10)] = FMA(KP1_902113032, TK, Tr);
|
Chris@10
|
240 R1[WS(rs, 2)] = FMS(KP1_902113032, TK, Tr);
|
Chris@10
|
241 R1[WS(rs, 7)] = FMS(KP1_902113032, TM, TL);
|
Chris@10
|
242 R0[WS(rs, 5)] = FMA(KP1_902113032, TM, TL);
|
Chris@10
|
243 }
|
Chris@10
|
244 }
|
Chris@10
|
245 {
|
Chris@10
|
246 E T18, T1n, T1x, TT, T2m, T1w, T1u, T2l, T1s, T1t, T2k;
|
Chris@10
|
247 T18 = FNMS(KP256756360, T17, T10);
|
Chris@10
|
248 T1s = FMA(KP256756360, T10, T17);
|
Chris@10
|
249 T1t = FMA(KP549754652, T1f, T1m);
|
Chris@10
|
250 T1n = FNMS(KP549754652, T1m, T1f);
|
Chris@10
|
251 T1x = FNMS(KP1_902113032, TS, TP);
|
Chris@10
|
252 TT = FMA(KP1_902113032, TS, TP);
|
Chris@10
|
253 T2m = FMA(KP730409924, T2j, T2g);
|
Chris@10
|
254 T2k = FNMS(KP730409924, T2j, T2g);
|
Chris@10
|
255 T1w = FNMS(KP683113946, T1s, T1t);
|
Chris@10
|
256 T1u = FMA(KP559154169, T1t, T1s);
|
Chris@10
|
257 R1[WS(rs, 1)] = -(FMA(KP1_996053456, T2k, T2d));
|
Chris@10
|
258 T2l = FNMS(KP499013364, T2k, T2d);
|
Chris@10
|
259 {
|
Chris@10
|
260 E T1K, T1M, T1G, T1E;
|
Chris@10
|
261 {
|
Chris@10
|
262 E T1D, T1A, T1q, T1p, T1v, T1r;
|
Chris@10
|
263 {
|
Chris@10
|
264 E T1I, T1J, T2n, T2r, T1o;
|
Chris@10
|
265 T1I = FMA(KP634619297, T1B, T1C);
|
Chris@10
|
266 T1D = FNMS(KP634619297, T1C, T1B);
|
Chris@10
|
267 T1A = FMA(KP470564281, T1z, T1y);
|
Chris@10
|
268 T1J = FNMS(KP470564281, T1y, T1z);
|
Chris@10
|
269 T2n = FNMS(KP1_115827804, T2m, T2l);
|
Chris@10
|
270 T2r = FMA(KP1_115827804, T2m, T2l);
|
Chris@10
|
271 T1q = FNMS(KP904730450, T1n, T18);
|
Chris@10
|
272 T1o = FMA(KP904730450, T1n, T18);
|
Chris@10
|
273 R1[WS(rs, 11)] = FMS(KP1_386580726, T2q, T2n);
|
Chris@10
|
274 R0[WS(rs, 4)] = FMA(KP1_386580726, T2q, T2n);
|
Chris@10
|
275 R0[WS(rs, 9)] = FMA(KP1_898359647, T2s, T2r);
|
Chris@10
|
276 R1[WS(rs, 6)] = FMS(KP1_898359647, T2s, T2r);
|
Chris@10
|
277 R1[0] = FMS(KP1_937166322, T1o, TT);
|
Chris@10
|
278 T1p = FMA(KP484291580, T1o, TT);
|
Chris@10
|
279 T1K = FMA(KP662318342, T1J, T1I);
|
Chris@10
|
280 T1M = FNMS(KP576710603, T1I, T1J);
|
Chris@10
|
281 }
|
Chris@10
|
282 T1v = FMA(KP1_082908895, T1q, T1p);
|
Chris@10
|
283 T1r = FNMS(KP1_082908895, T1q, T1p);
|
Chris@10
|
284 R1[WS(rs, 10)] = FMS(KP1_842354653, T1u, T1r);
|
Chris@10
|
285 R0[WS(rs, 3)] = FMA(KP1_842354653, T1u, T1r);
|
Chris@10
|
286 R0[WS(rs, 8)] = FMA(KP1_666834356, T1w, T1v);
|
Chris@10
|
287 R1[WS(rs, 5)] = FMS(KP1_666834356, T1w, T1v);
|
Chris@10
|
288 T1G = FNMS(KP933137358, T1D, T1A);
|
Chris@10
|
289 T1E = FMA(KP933137358, T1D, T1A);
|
Chris@10
|
290 }
|
Chris@10
|
291 {
|
Chris@10
|
292 E T23, T28, T29, T1W, T1F, T1H, T1L;
|
Chris@10
|
293 T23 = FNMS(KP634619297, T22, T1Z);
|
Chris@10
|
294 T28 = FMA(KP634619297, T1Z, T22);
|
Chris@10
|
295 T29 = FMA(KP549754652, T1S, T1V);
|
Chris@10
|
296 T1W = FNMS(KP549754652, T1V, T1S);
|
Chris@10
|
297 R0[WS(rs, 2)] = FMA(KP1_809654104, T1E, T1x);
|
Chris@10
|
298 T1F = FNMS(KP452413526, T1E, T1x);
|
Chris@10
|
299 T2c = FMA(KP595480289, T28, T29);
|
Chris@10
|
300 T2a = FNMS(KP641441904, T29, T28);
|
Chris@10
|
301 T1H = FNMS(KP1_011627398, T1G, T1F);
|
Chris@10
|
302 T1L = FMA(KP1_011627398, T1G, T1F);
|
Chris@10
|
303 R0[WS(rs, 12)] = FNMS(KP1_606007150, T1K, T1H);
|
Chris@10
|
304 R1[WS(rs, 4)] = -(FMA(KP1_606007150, T1K, T1H));
|
Chris@10
|
305 R1[WS(rs, 9)] = -(FMA(KP1_721083328, T1M, T1L));
|
Chris@10
|
306 R0[WS(rs, 7)] = FNMS(KP1_721083328, T1M, T1L);
|
Chris@10
|
307 T24 = FNMS(KP963507348, T23, T1W);
|
Chris@10
|
308 T26 = FMA(KP963507348, T23, T1W);
|
Chris@10
|
309 }
|
Chris@10
|
310 }
|
Chris@10
|
311 }
|
Chris@10
|
312 }
|
Chris@10
|
313 }
|
Chris@10
|
314 R0[WS(rs, 1)] = FNMS(KP1_752613360, T24, T1P);
|
Chris@10
|
315 T25 = FMA(KP438153340, T24, T1P);
|
Chris@10
|
316 T27 = FMA(KP979740652, T26, T25);
|
Chris@10
|
317 T2b = FNMS(KP979740652, T26, T25);
|
Chris@10
|
318 R1[WS(rs, 8)] = -(FMA(KP1_606007150, T2a, T27));
|
Chris@10
|
319 R0[WS(rs, 6)] = FNMS(KP1_606007150, T2a, T27);
|
Chris@10
|
320 R1[WS(rs, 3)] = -(FMA(KP1_666834356, T2c, T2b));
|
Chris@10
|
321 R0[WS(rs, 11)] = FNMS(KP1_666834356, T2c, T2b);
|
Chris@10
|
322 }
|
Chris@10
|
323 }
|
Chris@10
|
324 }
|
Chris@10
|
325
|
Chris@10
|
326 static const kr2c_desc desc = { 25, "r2cbIII_25", {32, 0, 120, 0}, &GENUS };
|
Chris@10
|
327
|
Chris@10
|
328 void X(codelet_r2cbIII_25) (planner *p) {
|
Chris@10
|
329 X(kr2c_register) (p, r2cbIII_25, &desc);
|
Chris@10
|
330 }
|
Chris@10
|
331
|
Chris@10
|
332 #else /* HAVE_FMA */
|
Chris@10
|
333
|
Chris@10
|
334 /* Generated by: ../../../genfft/gen_r2cb.native -compact -variables 4 -pipeline-latency 4 -sign 1 -n 25 -name r2cbIII_25 -dft-III -include r2cbIII.h */
|
Chris@10
|
335
|
Chris@10
|
336 /*
|
Chris@10
|
337 * This function contains 152 FP additions, 98 FP multiplications,
|
Chris@10
|
338 * (or, 100 additions, 46 multiplications, 52 fused multiply/add),
|
Chris@10
|
339 * 65 stack variables, 21 constants, and 50 memory accesses
|
Chris@10
|
340 */
|
Chris@10
|
341 #include "r2cbIII.h"
|
Chris@10
|
342
|
Chris@10
|
343 static void r2cbIII_25(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
|
Chris@10
|
344 {
|
Chris@10
|
345 DK(KP968583161, +0.968583161128631119490168375464735813836012403);
|
Chris@10
|
346 DK(KP248689887, +0.248689887164854788242283746006447968417567406);
|
Chris@10
|
347 DK(KP684547105, +0.684547105928688673732283357621209269889519233);
|
Chris@10
|
348 DK(KP728968627, +0.728968627421411523146730319055259111372571664);
|
Chris@10
|
349 DK(KP062790519, +0.062790519529313376076178224565631133122484832);
|
Chris@10
|
350 DK(KP998026728, +0.998026728428271561952336806863450553336905220);
|
Chris@10
|
351 DK(KP876306680, +0.876306680043863587308115903922062583399064238);
|
Chris@10
|
352 DK(KP481753674, +0.481753674101715274987191502872129653528542010);
|
Chris@10
|
353 DK(KP535826794, +0.535826794978996618271308767867639978063575346);
|
Chris@10
|
354 DK(KP844327925, +0.844327925502015078548558063966681505381659241);
|
Chris@10
|
355 DK(KP904827052, +0.904827052466019527713668647932697593970413911);
|
Chris@10
|
356 DK(KP425779291, +0.425779291565072648862502445744251703979973042);
|
Chris@10
|
357 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
|
Chris@10
|
358 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
|
Chris@10
|
359 DK(KP587785252, +0.587785252292473129168705954639072768597652438);
|
Chris@10
|
360 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
|
Chris@10
|
361 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
|
Chris@10
|
362 DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
|
Chris@10
|
363 DK(KP1_118033988, +1.118033988749894848204586834365638117720309180);
|
Chris@10
|
364 DK(KP1_175570504, +1.175570504584946258337411909278145537195304875);
|
Chris@10
|
365 DK(KP1_902113032, +1.902113032590307144232878666758764286811397268);
|
Chris@10
|
366 {
|
Chris@10
|
367 INT i;
|
Chris@10
|
368 for (i = v; i > 0; i = i - 1, R0 = R0 + ovs, R1 = R1 + ovs, Cr = Cr + ivs, Ci = Ci + ivs, MAKE_VOLATILE_STRIDE(100, rs), MAKE_VOLATILE_STRIDE(100, csr), MAKE_VOLATILE_STRIDE(100, csi)) {
|
Chris@10
|
369 E TS, T1O, T5, TP, T1N, TI, TH, Te, T17, T2h, T1y, T1V, T10, T2g, T1x;
|
Chris@10
|
370 E T1S, Tz, Ty, Tn, T1m, T2e, T1B, T22, T1f, T2d, T1A, T1Z, TQ, TR;
|
Chris@10
|
371 TQ = Ci[WS(csi, 2)];
|
Chris@10
|
372 TR = Ci[WS(csi, 7)];
|
Chris@10
|
373 TS = FNMS(KP1_175570504, TR, KP1_902113032 * TQ);
|
Chris@10
|
374 T1O = FMA(KP1_902113032, TR, KP1_175570504 * TQ);
|
Chris@10
|
375 {
|
Chris@10
|
376 E T1, T4, TN, T2, T3, TO;
|
Chris@10
|
377 T1 = Cr[WS(csr, 12)];
|
Chris@10
|
378 T2 = Cr[WS(csr, 7)];
|
Chris@10
|
379 T3 = Cr[WS(csr, 2)];
|
Chris@10
|
380 T4 = T2 + T3;
|
Chris@10
|
381 TN = KP1_118033988 * (T3 - T2);
|
Chris@10
|
382 T5 = FMA(KP2_000000000, T4, T1);
|
Chris@10
|
383 TO = FMS(KP500000000, T4, T1);
|
Chris@10
|
384 TP = TN - TO;
|
Chris@10
|
385 T1N = TO + TN;
|
Chris@10
|
386 }
|
Chris@10
|
387 {
|
Chris@10
|
388 E T6, Td, T15, TU, T14, T11, TX, TY;
|
Chris@10
|
389 T6 = Cr[WS(csr, 11)];
|
Chris@10
|
390 TI = Ci[WS(csi, 11)];
|
Chris@10
|
391 {
|
Chris@10
|
392 E T7, T8, T9, Ta, Tb, Tc;
|
Chris@10
|
393 T7 = Cr[WS(csr, 6)];
|
Chris@10
|
394 T8 = Cr[WS(csr, 8)];
|
Chris@10
|
395 T9 = T7 + T8;
|
Chris@10
|
396 Ta = Cr[WS(csr, 1)];
|
Chris@10
|
397 Tb = Cr[WS(csr, 3)];
|
Chris@10
|
398 Tc = Ta + Tb;
|
Chris@10
|
399 Td = T9 + Tc;
|
Chris@10
|
400 T15 = Ta - Tb;
|
Chris@10
|
401 TU = KP559016994 * (Tc - T9);
|
Chris@10
|
402 T14 = T8 - T7;
|
Chris@10
|
403 }
|
Chris@10
|
404 {
|
Chris@10
|
405 E TB, TC, TD, TE, TF, TG;
|
Chris@10
|
406 TB = Ci[WS(csi, 6)];
|
Chris@10
|
407 TC = Ci[WS(csi, 8)];
|
Chris@10
|
408 TD = TB - TC;
|
Chris@10
|
409 TE = Ci[WS(csi, 1)];
|
Chris@10
|
410 TF = Ci[WS(csi, 3)];
|
Chris@10
|
411 TG = TE - TF;
|
Chris@10
|
412 TH = TD + TG;
|
Chris@10
|
413 T11 = KP559016994 * (TD - TG);
|
Chris@10
|
414 TX = TB + TC;
|
Chris@10
|
415 TY = TE + TF;
|
Chris@10
|
416 }
|
Chris@10
|
417 Te = T6 + Td;
|
Chris@10
|
418 {
|
Chris@10
|
419 E T16, T1T, T13, T1U, T12;
|
Chris@10
|
420 T16 = FMA(KP587785252, T14, KP951056516 * T15);
|
Chris@10
|
421 T1T = FNMS(KP587785252, T15, KP951056516 * T14);
|
Chris@10
|
422 T12 = FNMS(KP250000000, TH, TI);
|
Chris@10
|
423 T13 = T11 - T12;
|
Chris@10
|
424 T1U = T11 + T12;
|
Chris@10
|
425 T17 = T13 - T16;
|
Chris@10
|
426 T2h = T1T - T1U;
|
Chris@10
|
427 T1y = T16 + T13;
|
Chris@10
|
428 T1V = T1T + T1U;
|
Chris@10
|
429 }
|
Chris@10
|
430 {
|
Chris@10
|
431 E TZ, T1R, TW, T1Q, TV;
|
Chris@10
|
432 TZ = FNMS(KP951056516, TY, KP587785252 * TX);
|
Chris@10
|
433 T1R = FMA(KP951056516, TX, KP587785252 * TY);
|
Chris@10
|
434 TV = FMS(KP250000000, Td, T6);
|
Chris@10
|
435 TW = TU - TV;
|
Chris@10
|
436 T1Q = TV + TU;
|
Chris@10
|
437 T10 = TW + TZ;
|
Chris@10
|
438 T2g = T1Q + T1R;
|
Chris@10
|
439 T1x = TZ - TW;
|
Chris@10
|
440 T1S = T1Q - T1R;
|
Chris@10
|
441 }
|
Chris@10
|
442 }
|
Chris@10
|
443 {
|
Chris@10
|
444 E Tf, Tm, T1k, T19, T1j, T1g, T1c, T1d;
|
Chris@10
|
445 Tf = Cr[WS(csr, 10)];
|
Chris@10
|
446 Tz = Ci[WS(csi, 10)];
|
Chris@10
|
447 {
|
Chris@10
|
448 E Tg, Th, Ti, Tj, Tk, Tl;
|
Chris@10
|
449 Tg = Cr[WS(csr, 5)];
|
Chris@10
|
450 Th = Cr[WS(csr, 9)];
|
Chris@10
|
451 Ti = Tg + Th;
|
Chris@10
|
452 Tj = Cr[0];
|
Chris@10
|
453 Tk = Cr[WS(csr, 4)];
|
Chris@10
|
454 Tl = Tj + Tk;
|
Chris@10
|
455 Tm = Ti + Tl;
|
Chris@10
|
456 T1k = Tj - Tk;
|
Chris@10
|
457 T19 = KP559016994 * (Tl - Ti);
|
Chris@10
|
458 T1j = Th - Tg;
|
Chris@10
|
459 }
|
Chris@10
|
460 {
|
Chris@10
|
461 E Ts, Tt, Tu, Tv, Tw, Tx;
|
Chris@10
|
462 Ts = Ci[WS(csi, 4)];
|
Chris@10
|
463 Tt = Ci[0];
|
Chris@10
|
464 Tu = Ts - Tt;
|
Chris@10
|
465 Tv = Ci[WS(csi, 5)];
|
Chris@10
|
466 Tw = Ci[WS(csi, 9)];
|
Chris@10
|
467 Tx = Tv - Tw;
|
Chris@10
|
468 Ty = Tu - Tx;
|
Chris@10
|
469 T1g = KP559016994 * (Tx + Tu);
|
Chris@10
|
470 T1c = Tv + Tw;
|
Chris@10
|
471 T1d = Tt + Ts;
|
Chris@10
|
472 }
|
Chris@10
|
473 Tn = Tf + Tm;
|
Chris@10
|
474 {
|
Chris@10
|
475 E T1l, T20, T1i, T21, T1h;
|
Chris@10
|
476 T1l = FMA(KP587785252, T1j, KP951056516 * T1k);
|
Chris@10
|
477 T20 = FNMS(KP587785252, T1k, KP951056516 * T1j);
|
Chris@10
|
478 T1h = FMA(KP250000000, Ty, Tz);
|
Chris@10
|
479 T1i = T1g - T1h;
|
Chris@10
|
480 T21 = T1g + T1h;
|
Chris@10
|
481 T1m = T1i - T1l;
|
Chris@10
|
482 T2e = T21 - T20;
|
Chris@10
|
483 T1B = T1l + T1i;
|
Chris@10
|
484 T22 = T20 + T21;
|
Chris@10
|
485 }
|
Chris@10
|
486 {
|
Chris@10
|
487 E T1e, T1Y, T1b, T1X, T1a;
|
Chris@10
|
488 T1e = FNMS(KP951056516, T1d, KP587785252 * T1c);
|
Chris@10
|
489 T1Y = FMA(KP951056516, T1c, KP587785252 * T1d);
|
Chris@10
|
490 T1a = FMS(KP250000000, Tm, Tf);
|
Chris@10
|
491 T1b = T19 - T1a;
|
Chris@10
|
492 T1X = T1a + T19;
|
Chris@10
|
493 T1f = T1b + T1e;
|
Chris@10
|
494 T2d = T1X + T1Y;
|
Chris@10
|
495 T1A = T1e - T1b;
|
Chris@10
|
496 T1Z = T1X - T1Y;
|
Chris@10
|
497 }
|
Chris@10
|
498 }
|
Chris@10
|
499 {
|
Chris@10
|
500 E Tq, To, Tp, TK, TM, TA, TJ, TL, Tr;
|
Chris@10
|
501 Tq = KP1_118033988 * (Tn - Te);
|
Chris@10
|
502 To = Te + Tn;
|
Chris@10
|
503 Tp = FMS(KP500000000, To, T5);
|
Chris@10
|
504 TA = Ty - Tz;
|
Chris@10
|
505 TJ = TH + TI;
|
Chris@10
|
506 TK = FNMS(KP1_902113032, TJ, KP1_175570504 * TA);
|
Chris@10
|
507 TM = FMA(KP1_175570504, TJ, KP1_902113032 * TA);
|
Chris@10
|
508 R0[0] = FMA(KP2_000000000, To, T5);
|
Chris@10
|
509 TL = Tq - Tp;
|
Chris@10
|
510 R0[WS(rs, 5)] = TL + TM;
|
Chris@10
|
511 R1[WS(rs, 7)] = TM - TL;
|
Chris@10
|
512 Tr = Tp + Tq;
|
Chris@10
|
513 R1[WS(rs, 2)] = Tr + TK;
|
Chris@10
|
514 R0[WS(rs, 10)] = TK - Tr;
|
Chris@10
|
515 }
|
Chris@10
|
516 {
|
Chris@10
|
517 E T2q, T2s, T2k, T2j, T2l, T2m, T2r, T2n;
|
Chris@10
|
518 {
|
Chris@10
|
519 E T2o, T2p, T2f, T2i;
|
Chris@10
|
520 T2o = FNMS(KP904827052, T2d, KP425779291 * T2e);
|
Chris@10
|
521 T2p = FNMS(KP535826794, T2h, KP844327925 * T2g);
|
Chris@10
|
522 T2q = FNMS(KP1_902113032, T2p, KP1_175570504 * T2o);
|
Chris@10
|
523 T2s = FMA(KP1_175570504, T2p, KP1_902113032 * T2o);
|
Chris@10
|
524 T2k = T1N + T1O;
|
Chris@10
|
525 T2f = FMA(KP425779291, T2d, KP904827052 * T2e);
|
Chris@10
|
526 T2i = FMA(KP535826794, T2g, KP844327925 * T2h);
|
Chris@10
|
527 T2j = T2f - T2i;
|
Chris@10
|
528 T2l = FMA(KP500000000, T2j, T2k);
|
Chris@10
|
529 T2m = KP1_118033988 * (T2i + T2f);
|
Chris@10
|
530 }
|
Chris@10
|
531 R0[WS(rs, 2)] = FMS(KP2_000000000, T2j, T2k);
|
Chris@10
|
532 T2r = T2m - T2l;
|
Chris@10
|
533 R0[WS(rs, 7)] = T2r + T2s;
|
Chris@10
|
534 R1[WS(rs, 9)] = T2s - T2r;
|
Chris@10
|
535 T2n = T2l + T2m;
|
Chris@10
|
536 R1[WS(rs, 4)] = T2n + T2q;
|
Chris@10
|
537 R0[WS(rs, 12)] = T2q - T2n;
|
Chris@10
|
538 }
|
Chris@10
|
539 {
|
Chris@10
|
540 E T1u, T1w, TT, T1o, T1p, T1q, T1v, T1r;
|
Chris@10
|
541 {
|
Chris@10
|
542 E T1s, T1t, T18, T1n;
|
Chris@10
|
543 T1s = FMA(KP481753674, T10, KP876306680 * T17);
|
Chris@10
|
544 T1t = FMA(KP844327925, T1f, KP535826794 * T1m);
|
Chris@10
|
545 T1u = FMA(KP1_902113032, T1s, KP1_175570504 * T1t);
|
Chris@10
|
546 T1w = FNMS(KP1_175570504, T1s, KP1_902113032 * T1t);
|
Chris@10
|
547 TT = TP - TS;
|
Chris@10
|
548 T18 = FNMS(KP481753674, T17, KP876306680 * T10);
|
Chris@10
|
549 T1n = FNMS(KP844327925, T1m, KP535826794 * T1f);
|
Chris@10
|
550 T1o = T18 + T1n;
|
Chris@10
|
551 T1p = FMS(KP500000000, T1o, TT);
|
Chris@10
|
552 T1q = KP1_118033988 * (T1n - T18);
|
Chris@10
|
553 }
|
Chris@10
|
554 R0[WS(rs, 1)] = FMA(KP2_000000000, T1o, TT);
|
Chris@10
|
555 T1v = T1q - T1p;
|
Chris@10
|
556 R0[WS(rs, 6)] = T1v + T1w;
|
Chris@10
|
557 R1[WS(rs, 8)] = T1w - T1v;
|
Chris@10
|
558 T1r = T1p + T1q;
|
Chris@10
|
559 R1[WS(rs, 3)] = T1r + T1u;
|
Chris@10
|
560 R0[WS(rs, 11)] = T1u - T1r;
|
Chris@10
|
561 }
|
Chris@10
|
562 {
|
Chris@10
|
563 E T1H, T1L, T1E, T1D, T1I, T1J, T1M, T1K;
|
Chris@10
|
564 {
|
Chris@10
|
565 E T1F, T1G, T1z, T1C;
|
Chris@10
|
566 T1F = FNMS(KP062790519, T1B, KP998026728 * T1A);
|
Chris@10
|
567 T1G = FNMS(KP684547105, T1x, KP728968627 * T1y);
|
Chris@10
|
568 T1H = FNMS(KP1_902113032, T1G, KP1_175570504 * T1F);
|
Chris@10
|
569 T1L = FMA(KP1_175570504, T1G, KP1_902113032 * T1F);
|
Chris@10
|
570 T1E = TP + TS;
|
Chris@10
|
571 T1z = FMA(KP728968627, T1x, KP684547105 * T1y);
|
Chris@10
|
572 T1C = FMA(KP062790519, T1A, KP998026728 * T1B);
|
Chris@10
|
573 T1D = T1z + T1C;
|
Chris@10
|
574 T1I = FMA(KP500000000, T1D, T1E);
|
Chris@10
|
575 T1J = KP1_118033988 * (T1C - T1z);
|
Chris@10
|
576 }
|
Chris@10
|
577 R1[WS(rs, 1)] = FMS(KP2_000000000, T1D, T1E);
|
Chris@10
|
578 T1M = T1J - T1I;
|
Chris@10
|
579 R0[WS(rs, 9)] = T1L - T1M;
|
Chris@10
|
580 R1[WS(rs, 6)] = T1L + T1M;
|
Chris@10
|
581 T1K = T1I + T1J;
|
Chris@10
|
582 R1[WS(rs, 11)] = T1H - T1K;
|
Chris@10
|
583 R0[WS(rs, 4)] = T1H + T1K;
|
Chris@10
|
584 }
|
Chris@10
|
585 {
|
Chris@10
|
586 E T2a, T2c, T1P, T24, T25, T26, T2b, T27;
|
Chris@10
|
587 {
|
Chris@10
|
588 E T28, T29, T1W, T23;
|
Chris@10
|
589 T28 = FMA(KP248689887, T1S, KP968583161 * T1V);
|
Chris@10
|
590 T29 = FMA(KP481753674, T1Z, KP876306680 * T22);
|
Chris@10
|
591 T2a = FMA(KP1_902113032, T28, KP1_175570504 * T29);
|
Chris@10
|
592 T2c = FNMS(KP1_175570504, T28, KP1_902113032 * T29);
|
Chris@10
|
593 T1P = T1N - T1O;
|
Chris@10
|
594 T1W = FNMS(KP248689887, T1V, KP968583161 * T1S);
|
Chris@10
|
595 T23 = FNMS(KP481753674, T22, KP876306680 * T1Z);
|
Chris@10
|
596 T24 = T1W + T23;
|
Chris@10
|
597 T25 = FMS(KP500000000, T24, T1P);
|
Chris@10
|
598 T26 = KP1_118033988 * (T23 - T1W);
|
Chris@10
|
599 }
|
Chris@10
|
600 R1[0] = FMA(KP2_000000000, T24, T1P);
|
Chris@10
|
601 T2b = T26 - T25;
|
Chris@10
|
602 R1[WS(rs, 5)] = T2b + T2c;
|
Chris@10
|
603 R0[WS(rs, 8)] = T2c - T2b;
|
Chris@10
|
604 T27 = T25 + T26;
|
Chris@10
|
605 R0[WS(rs, 3)] = T27 + T2a;
|
Chris@10
|
606 R1[WS(rs, 10)] = T2a - T27;
|
Chris@10
|
607 }
|
Chris@10
|
608 }
|
Chris@10
|
609 }
|
Chris@10
|
610 }
|
Chris@10
|
611
|
Chris@10
|
612 static const kr2c_desc desc = { 25, "r2cbIII_25", {100, 46, 52, 0}, &GENUS };
|
Chris@10
|
613
|
Chris@10
|
614 void X(codelet_r2cbIII_25) (planner *p) {
|
Chris@10
|
615 X(kr2c_register) (p, r2cbIII_25, &desc);
|
Chris@10
|
616 }
|
Chris@10
|
617
|
Chris@10
|
618 #endif /* HAVE_FMA */
|